xref: /dpdk/drivers/net/mlx5/mlx5_rxtx.c (revision e11bdd37745229bf26b557305c07d118c3dbaad7)
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 /* Verbs header. */
11 /* ISO C doesn't support unnamed structs/unions, disabling -pedantic. */
12 #ifdef PEDANTIC
13 #pragma GCC diagnostic ignored "-Wpedantic"
14 #endif
15 #include <infiniband/verbs.h>
16 #include <infiniband/mlx5dv.h>
17 #ifdef PEDANTIC
18 #pragma GCC diagnostic error "-Wpedantic"
19 #endif
20 
21 #include <rte_mbuf.h>
22 #include <rte_mempool.h>
23 #include <rte_prefetch.h>
24 #include <rte_common.h>
25 #include <rte_branch_prediction.h>
26 #include <rte_ether.h>
27 #include <rte_cycles.h>
28 #include <rte_flow.h>
29 
30 #include <mlx5_devx_cmds.h>
31 #include <mlx5_prm.h>
32 #include <mlx5_common.h>
33 
34 #include "mlx5_defs.h"
35 #include "mlx5.h"
36 #include "mlx5_mr.h"
37 #include "mlx5_utils.h"
38 #include "mlx5_rxtx.h"
39 #include "mlx5_autoconf.h"
40 
41 /* TX burst subroutines return codes. */
42 enum mlx5_txcmp_code {
43 	MLX5_TXCMP_CODE_EXIT = 0,
44 	MLX5_TXCMP_CODE_ERROR,
45 	MLX5_TXCMP_CODE_SINGLE,
46 	MLX5_TXCMP_CODE_MULTI,
47 	MLX5_TXCMP_CODE_TSO,
48 	MLX5_TXCMP_CODE_EMPW,
49 };
50 
51 /*
52  * These defines are used to configure Tx burst routine option set
53  * supported at compile time. The not specified options are optimized out
54  * out due to if conditions can be explicitly calculated at compile time.
55  * The offloads with bigger runtime check (require more CPU cycles to
56  * skip) overhead should have the bigger index - this is needed to
57  * select the better matching routine function if no exact match and
58  * some offloads are not actually requested.
59  */
60 #define MLX5_TXOFF_CONFIG_MULTI (1u << 0) /* Multi-segment packets.*/
61 #define MLX5_TXOFF_CONFIG_TSO (1u << 1) /* TCP send offload supported.*/
62 #define MLX5_TXOFF_CONFIG_SWP (1u << 2) /* Tunnels/SW Parser offloads.*/
63 #define MLX5_TXOFF_CONFIG_CSUM (1u << 3) /* Check Sums offloaded. */
64 #define MLX5_TXOFF_CONFIG_INLINE (1u << 4) /* Data inlining supported. */
65 #define MLX5_TXOFF_CONFIG_VLAN (1u << 5) /* VLAN insertion supported.*/
66 #define MLX5_TXOFF_CONFIG_METADATA (1u << 6) /* Flow metadata. */
67 #define MLX5_TXOFF_CONFIG_EMPW (1u << 8) /* Enhanced MPW supported.*/
68 #define MLX5_TXOFF_CONFIG_MPW (1u << 9) /* Legacy MPW supported.*/
69 
70 /* The most common offloads groups. */
71 #define MLX5_TXOFF_CONFIG_NONE 0
72 #define MLX5_TXOFF_CONFIG_FULL (MLX5_TXOFF_CONFIG_MULTI | \
73 				MLX5_TXOFF_CONFIG_TSO | \
74 				MLX5_TXOFF_CONFIG_SWP | \
75 				MLX5_TXOFF_CONFIG_CSUM | \
76 				MLX5_TXOFF_CONFIG_INLINE | \
77 				MLX5_TXOFF_CONFIG_VLAN | \
78 				MLX5_TXOFF_CONFIG_METADATA)
79 
80 #define MLX5_TXOFF_CONFIG(mask) (olx & MLX5_TXOFF_CONFIG_##mask)
81 
82 #define MLX5_TXOFF_DECL(func, olx) \
83 static uint16_t mlx5_tx_burst_##func(void *txq, \
84 				     struct rte_mbuf **pkts, \
85 				    uint16_t pkts_n) \
86 { \
87 	return mlx5_tx_burst_tmpl((struct mlx5_txq_data *)txq, \
88 		    pkts, pkts_n, (olx)); \
89 }
90 
91 #define MLX5_TXOFF_INFO(func, olx) {mlx5_tx_burst_##func, olx},
92 
93 static __rte_always_inline uint32_t
94 rxq_cq_to_pkt_type(struct mlx5_rxq_data *rxq, volatile struct mlx5_cqe *cqe);
95 
96 static __rte_always_inline int
97 mlx5_rx_poll_len(struct mlx5_rxq_data *rxq, volatile struct mlx5_cqe *cqe,
98 		 uint16_t cqe_cnt, volatile struct mlx5_mini_cqe8 **mcqe);
99 
100 static __rte_always_inline uint32_t
101 rxq_cq_to_ol_flags(volatile struct mlx5_cqe *cqe);
102 
103 static __rte_always_inline void
104 rxq_cq_to_mbuf(struct mlx5_rxq_data *rxq, struct rte_mbuf *pkt,
105 	       volatile struct mlx5_cqe *cqe, uint32_t rss_hash_res);
106 
107 static __rte_always_inline void
108 mprq_buf_replace(struct mlx5_rxq_data *rxq, uint16_t rq_idx,
109 		 const unsigned int strd_n);
110 
111 static int
112 mlx5_queue_state_modify(struct rte_eth_dev *dev,
113 			struct mlx5_mp_arg_queue_state_modify *sm);
114 
115 static inline void
116 mlx5_lro_update_tcp_hdr(struct rte_tcp_hdr *restrict tcp,
117 			volatile struct mlx5_cqe *restrict cqe,
118 			uint32_t phcsum);
119 
120 static inline void
121 mlx5_lro_update_hdr(uint8_t *restrict padd,
122 		    volatile struct mlx5_cqe *restrict cqe,
123 		    uint32_t len);
124 
125 uint32_t mlx5_ptype_table[] __rte_cache_aligned = {
126 	[0xff] = RTE_PTYPE_ALL_MASK, /* Last entry for errored packet. */
127 };
128 
129 uint8_t mlx5_cksum_table[1 << 10] __rte_cache_aligned;
130 uint8_t mlx5_swp_types_table[1 << 10] __rte_cache_aligned;
131 
132 uint64_t rte_net_mlx5_dynf_inline_mask;
133 #define PKT_TX_DYNF_NOINLINE rte_net_mlx5_dynf_inline_mask
134 
135 /**
136  * Build a table to translate Rx completion flags to packet type.
137  *
138  * @note: fix mlx5_dev_supported_ptypes_get() if any change here.
139  */
140 void
141 mlx5_set_ptype_table(void)
142 {
143 	unsigned int i;
144 	uint32_t (*p)[RTE_DIM(mlx5_ptype_table)] = &mlx5_ptype_table;
145 
146 	/* Last entry must not be overwritten, reserved for errored packet. */
147 	for (i = 0; i < RTE_DIM(mlx5_ptype_table) - 1; ++i)
148 		(*p)[i] = RTE_PTYPE_UNKNOWN;
149 	/*
150 	 * The index to the array should have:
151 	 * bit[1:0] = l3_hdr_type
152 	 * bit[4:2] = l4_hdr_type
153 	 * bit[5] = ip_frag
154 	 * bit[6] = tunneled
155 	 * bit[7] = outer_l3_type
156 	 */
157 	/* L2 */
158 	(*p)[0x00] = RTE_PTYPE_L2_ETHER;
159 	/* L3 */
160 	(*p)[0x01] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
161 		     RTE_PTYPE_L4_NONFRAG;
162 	(*p)[0x02] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
163 		     RTE_PTYPE_L4_NONFRAG;
164 	/* Fragmented */
165 	(*p)[0x21] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
166 		     RTE_PTYPE_L4_FRAG;
167 	(*p)[0x22] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
168 		     RTE_PTYPE_L4_FRAG;
169 	/* TCP */
170 	(*p)[0x05] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
171 		     RTE_PTYPE_L4_TCP;
172 	(*p)[0x06] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
173 		     RTE_PTYPE_L4_TCP;
174 	(*p)[0x0d] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
175 		     RTE_PTYPE_L4_TCP;
176 	(*p)[0x0e] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
177 		     RTE_PTYPE_L4_TCP;
178 	(*p)[0x11] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
179 		     RTE_PTYPE_L4_TCP;
180 	(*p)[0x12] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
181 		     RTE_PTYPE_L4_TCP;
182 	/* UDP */
183 	(*p)[0x09] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
184 		     RTE_PTYPE_L4_UDP;
185 	(*p)[0x0a] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
186 		     RTE_PTYPE_L4_UDP;
187 	/* Repeat with outer_l3_type being set. Just in case. */
188 	(*p)[0x81] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
189 		     RTE_PTYPE_L4_NONFRAG;
190 	(*p)[0x82] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
191 		     RTE_PTYPE_L4_NONFRAG;
192 	(*p)[0xa1] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
193 		     RTE_PTYPE_L4_FRAG;
194 	(*p)[0xa2] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
195 		     RTE_PTYPE_L4_FRAG;
196 	(*p)[0x85] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
197 		     RTE_PTYPE_L4_TCP;
198 	(*p)[0x86] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
199 		     RTE_PTYPE_L4_TCP;
200 	(*p)[0x8d] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
201 		     RTE_PTYPE_L4_TCP;
202 	(*p)[0x8e] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
203 		     RTE_PTYPE_L4_TCP;
204 	(*p)[0x91] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
205 		     RTE_PTYPE_L4_TCP;
206 	(*p)[0x92] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
207 		     RTE_PTYPE_L4_TCP;
208 	(*p)[0x89] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
209 		     RTE_PTYPE_L4_UDP;
210 	(*p)[0x8a] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
211 		     RTE_PTYPE_L4_UDP;
212 	/* Tunneled - L3 */
213 	(*p)[0x40] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN;
214 	(*p)[0x41] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
215 		     RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN |
216 		     RTE_PTYPE_INNER_L4_NONFRAG;
217 	(*p)[0x42] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
218 		     RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN |
219 		     RTE_PTYPE_INNER_L4_NONFRAG;
220 	(*p)[0xc0] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN;
221 	(*p)[0xc1] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
222 		     RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN |
223 		     RTE_PTYPE_INNER_L4_NONFRAG;
224 	(*p)[0xc2] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
225 		     RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN |
226 		     RTE_PTYPE_INNER_L4_NONFRAG;
227 	/* Tunneled - Fragmented */
228 	(*p)[0x61] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
229 		     RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN |
230 		     RTE_PTYPE_INNER_L4_FRAG;
231 	(*p)[0x62] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
232 		     RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN |
233 		     RTE_PTYPE_INNER_L4_FRAG;
234 	(*p)[0xe1] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
235 		     RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN |
236 		     RTE_PTYPE_INNER_L4_FRAG;
237 	(*p)[0xe2] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
238 		     RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN |
239 		     RTE_PTYPE_INNER_L4_FRAG;
240 	/* Tunneled - TCP */
241 	(*p)[0x45] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
242 		     RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN |
243 		     RTE_PTYPE_INNER_L4_TCP;
244 	(*p)[0x46] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
245 		     RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN |
246 		     RTE_PTYPE_INNER_L4_TCP;
247 	(*p)[0x4d] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
248 		     RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN |
249 		     RTE_PTYPE_INNER_L4_TCP;
250 	(*p)[0x4e] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
251 		     RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN |
252 		     RTE_PTYPE_INNER_L4_TCP;
253 	(*p)[0x51] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
254 		     RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN |
255 		     RTE_PTYPE_INNER_L4_TCP;
256 	(*p)[0x52] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
257 		     RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN |
258 		     RTE_PTYPE_INNER_L4_TCP;
259 	(*p)[0xc5] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
260 		     RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN |
261 		     RTE_PTYPE_INNER_L4_TCP;
262 	(*p)[0xc6] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
263 		     RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN |
264 		     RTE_PTYPE_INNER_L4_TCP;
265 	(*p)[0xcd] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
266 		     RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN |
267 		     RTE_PTYPE_INNER_L4_TCP;
268 	(*p)[0xce] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
269 		     RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN |
270 		     RTE_PTYPE_INNER_L4_TCP;
271 	(*p)[0xd1] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
272 		     RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN |
273 		     RTE_PTYPE_INNER_L4_TCP;
274 	(*p)[0xd2] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
275 		     RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN |
276 		     RTE_PTYPE_INNER_L4_TCP;
277 	/* Tunneled - UDP */
278 	(*p)[0x49] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
279 		     RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN |
280 		     RTE_PTYPE_INNER_L4_UDP;
281 	(*p)[0x4a] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
282 		     RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN |
283 		     RTE_PTYPE_INNER_L4_UDP;
284 	(*p)[0xc9] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
285 		     RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN |
286 		     RTE_PTYPE_INNER_L4_UDP;
287 	(*p)[0xca] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
288 		     RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN |
289 		     RTE_PTYPE_INNER_L4_UDP;
290 }
291 
292 /**
293  * Build a table to translate packet to checksum type of Verbs.
294  */
295 void
296 mlx5_set_cksum_table(void)
297 {
298 	unsigned int i;
299 	uint8_t v;
300 
301 	/*
302 	 * The index should have:
303 	 * bit[0] = PKT_TX_TCP_SEG
304 	 * bit[2:3] = PKT_TX_UDP_CKSUM, PKT_TX_TCP_CKSUM
305 	 * bit[4] = PKT_TX_IP_CKSUM
306 	 * bit[8] = PKT_TX_OUTER_IP_CKSUM
307 	 * bit[9] = tunnel
308 	 */
309 	for (i = 0; i < RTE_DIM(mlx5_cksum_table); ++i) {
310 		v = 0;
311 		if (i & (1 << 9)) {
312 			/* Tunneled packet. */
313 			if (i & (1 << 8)) /* Outer IP. */
314 				v |= MLX5_ETH_WQE_L3_CSUM;
315 			if (i & (1 << 4)) /* Inner IP. */
316 				v |= MLX5_ETH_WQE_L3_INNER_CSUM;
317 			if (i & (3 << 2 | 1 << 0)) /* L4 or TSO. */
318 				v |= MLX5_ETH_WQE_L4_INNER_CSUM;
319 		} else {
320 			/* No tunnel. */
321 			if (i & (1 << 4)) /* IP. */
322 				v |= MLX5_ETH_WQE_L3_CSUM;
323 			if (i & (3 << 2 | 1 << 0)) /* L4 or TSO. */
324 				v |= MLX5_ETH_WQE_L4_CSUM;
325 		}
326 		mlx5_cksum_table[i] = v;
327 	}
328 }
329 
330 /**
331  * Build a table to translate packet type of mbuf to SWP type of Verbs.
332  */
333 void
334 mlx5_set_swp_types_table(void)
335 {
336 	unsigned int i;
337 	uint8_t v;
338 
339 	/*
340 	 * The index should have:
341 	 * bit[0:1] = PKT_TX_L4_MASK
342 	 * bit[4] = PKT_TX_IPV6
343 	 * bit[8] = PKT_TX_OUTER_IPV6
344 	 * bit[9] = PKT_TX_OUTER_UDP
345 	 */
346 	for (i = 0; i < RTE_DIM(mlx5_swp_types_table); ++i) {
347 		v = 0;
348 		if (i & (1 << 8))
349 			v |= MLX5_ETH_WQE_L3_OUTER_IPV6;
350 		if (i & (1 << 9))
351 			v |= MLX5_ETH_WQE_L4_OUTER_UDP;
352 		if (i & (1 << 4))
353 			v |= MLX5_ETH_WQE_L3_INNER_IPV6;
354 		if ((i & 3) == (PKT_TX_UDP_CKSUM >> 52))
355 			v |= MLX5_ETH_WQE_L4_INNER_UDP;
356 		mlx5_swp_types_table[i] = v;
357 	}
358 }
359 
360 /**
361  * Set Software Parser flags and offsets in Ethernet Segment of WQE.
362  * Flags must be preliminary initialized to zero.
363  *
364  * @param loc
365  *   Pointer to burst routine local context.
366  * @param swp_flags
367  *   Pointer to store Software Parser flags
368  * @param olx
369  *   Configured Tx offloads mask. It is fully defined at
370  *   compile time and may be used for optimization.
371  *
372  * @return
373  *   Software Parser offsets packed in dword.
374  *   Software Parser flags are set by pointer.
375  */
376 static __rte_always_inline uint32_t
377 txq_mbuf_to_swp(struct mlx5_txq_local *restrict loc,
378 		uint8_t *swp_flags,
379 		unsigned int olx)
380 {
381 	uint64_t ol, tunnel;
382 	unsigned int idx, off;
383 	uint32_t set;
384 
385 	if (!MLX5_TXOFF_CONFIG(SWP))
386 		return 0;
387 	ol = loc->mbuf->ol_flags;
388 	tunnel = ol & PKT_TX_TUNNEL_MASK;
389 	/*
390 	 * Check whether Software Parser is required.
391 	 * Only customized tunnels may ask for.
392 	 */
393 	if (likely(tunnel != PKT_TX_TUNNEL_UDP && tunnel != PKT_TX_TUNNEL_IP))
394 		return 0;
395 	/*
396 	 * The index should have:
397 	 * bit[0:1] = PKT_TX_L4_MASK
398 	 * bit[4] = PKT_TX_IPV6
399 	 * bit[8] = PKT_TX_OUTER_IPV6
400 	 * bit[9] = PKT_TX_OUTER_UDP
401 	 */
402 	idx = (ol & (PKT_TX_L4_MASK | PKT_TX_IPV6 | PKT_TX_OUTER_IPV6)) >> 52;
403 	idx |= (tunnel == PKT_TX_TUNNEL_UDP) ? (1 << 9) : 0;
404 	*swp_flags = mlx5_swp_types_table[idx];
405 	/*
406 	 * Set offsets for SW parser. Since ConnectX-5, SW parser just
407 	 * complements HW parser. SW parser starts to engage only if HW parser
408 	 * can't reach a header. For the older devices, HW parser will not kick
409 	 * in if any of SWP offsets is set. Therefore, all of the L3 offsets
410 	 * should be set regardless of HW offload.
411 	 */
412 	off = loc->mbuf->outer_l2_len;
413 	if (MLX5_TXOFF_CONFIG(VLAN) && ol & PKT_TX_VLAN_PKT)
414 		off += sizeof(struct rte_vlan_hdr);
415 	set = (off >> 1) << 8; /* Outer L3 offset. */
416 	off += loc->mbuf->outer_l3_len;
417 	if (tunnel == PKT_TX_TUNNEL_UDP)
418 		set |= off >> 1; /* Outer L4 offset. */
419 	if (ol & (PKT_TX_IPV4 | PKT_TX_IPV6)) { /* Inner IP. */
420 		const uint64_t csum = ol & PKT_TX_L4_MASK;
421 			off += loc->mbuf->l2_len;
422 		set |= (off >> 1) << 24; /* Inner L3 offset. */
423 		if (csum == PKT_TX_TCP_CKSUM ||
424 		    csum == PKT_TX_UDP_CKSUM ||
425 		    (MLX5_TXOFF_CONFIG(TSO) && ol & PKT_TX_TCP_SEG)) {
426 			off += loc->mbuf->l3_len;
427 			set |= (off >> 1) << 16; /* Inner L4 offset. */
428 		}
429 	}
430 	set = rte_cpu_to_le_32(set);
431 	return set;
432 }
433 
434 /**
435  * Convert the Checksum offloads to Verbs.
436  *
437  * @param buf
438  *   Pointer to the mbuf.
439  *
440  * @return
441  *   Converted checksum flags.
442  */
443 static __rte_always_inline uint8_t
444 txq_ol_cksum_to_cs(struct rte_mbuf *buf)
445 {
446 	uint32_t idx;
447 	uint8_t is_tunnel = !!(buf->ol_flags & PKT_TX_TUNNEL_MASK);
448 	const uint64_t ol_flags_mask = PKT_TX_TCP_SEG | PKT_TX_L4_MASK |
449 				       PKT_TX_IP_CKSUM | PKT_TX_OUTER_IP_CKSUM;
450 
451 	/*
452 	 * The index should have:
453 	 * bit[0] = PKT_TX_TCP_SEG
454 	 * bit[2:3] = PKT_TX_UDP_CKSUM, PKT_TX_TCP_CKSUM
455 	 * bit[4] = PKT_TX_IP_CKSUM
456 	 * bit[8] = PKT_TX_OUTER_IP_CKSUM
457 	 * bit[9] = tunnel
458 	 */
459 	idx = ((buf->ol_flags & ol_flags_mask) >> 50) | (!!is_tunnel << 9);
460 	return mlx5_cksum_table[idx];
461 }
462 
463 /**
464  * Internal function to compute the number of used descriptors in an RX queue
465  *
466  * @param rxq
467  *   The Rx queue.
468  *
469  * @return
470  *   The number of used rx descriptor.
471  */
472 static uint32_t
473 rx_queue_count(struct mlx5_rxq_data *rxq)
474 {
475 	struct rxq_zip *zip = &rxq->zip;
476 	volatile struct mlx5_cqe *cqe;
477 	const unsigned int cqe_n = (1 << rxq->cqe_n);
478 	const unsigned int cqe_cnt = cqe_n - 1;
479 	unsigned int cq_ci;
480 	unsigned int used;
481 
482 	/* if we are processing a compressed cqe */
483 	if (zip->ai) {
484 		used = zip->cqe_cnt - zip->ca;
485 		cq_ci = zip->cq_ci;
486 	} else {
487 		used = 0;
488 		cq_ci = rxq->cq_ci;
489 	}
490 	cqe = &(*rxq->cqes)[cq_ci & cqe_cnt];
491 	while (check_cqe(cqe, cqe_n, cq_ci) != MLX5_CQE_STATUS_HW_OWN) {
492 		int8_t op_own;
493 		unsigned int n;
494 
495 		op_own = cqe->op_own;
496 		if (MLX5_CQE_FORMAT(op_own) == MLX5_COMPRESSED)
497 			n = rte_be_to_cpu_32(cqe->byte_cnt);
498 		else
499 			n = 1;
500 		cq_ci += n;
501 		used += n;
502 		cqe = &(*rxq->cqes)[cq_ci & cqe_cnt];
503 	}
504 	used = RTE_MIN(used, (1U << rxq->elts_n) - 1);
505 	return used;
506 }
507 
508 /**
509  * DPDK callback to check the status of a rx descriptor.
510  *
511  * @param rx_queue
512  *   The Rx queue.
513  * @param[in] offset
514  *   The index of the descriptor in the ring.
515  *
516  * @return
517  *   The status of the tx descriptor.
518  */
519 int
520 mlx5_rx_descriptor_status(void *rx_queue, uint16_t offset)
521 {
522 	struct mlx5_rxq_data *rxq = rx_queue;
523 	struct mlx5_rxq_ctrl *rxq_ctrl =
524 			container_of(rxq, struct mlx5_rxq_ctrl, rxq);
525 	struct rte_eth_dev *dev = ETH_DEV(rxq_ctrl->priv);
526 
527 	if (dev->rx_pkt_burst != mlx5_rx_burst) {
528 		rte_errno = ENOTSUP;
529 		return -rte_errno;
530 	}
531 	if (offset >= (1 << rxq->elts_n)) {
532 		rte_errno = EINVAL;
533 		return -rte_errno;
534 	}
535 	if (offset < rx_queue_count(rxq))
536 		return RTE_ETH_RX_DESC_DONE;
537 	return RTE_ETH_RX_DESC_AVAIL;
538 }
539 
540 /**
541  * DPDK callback to get the RX queue information
542  *
543  * @param dev
544  *   Pointer to the device structure.
545  *
546  * @param rx_queue_id
547  *   Rx queue identificator.
548  *
549  * @param qinfo
550  *   Pointer to the RX queue information structure.
551  *
552  * @return
553  *   None.
554  */
555 
556 void
557 mlx5_rxq_info_get(struct rte_eth_dev *dev, uint16_t rx_queue_id,
558 		  struct rte_eth_rxq_info *qinfo)
559 {
560 	struct mlx5_priv *priv = dev->data->dev_private;
561 	struct mlx5_rxq_data *rxq = (*priv->rxqs)[rx_queue_id];
562 	struct mlx5_rxq_ctrl *rxq_ctrl =
563 		container_of(rxq, struct mlx5_rxq_ctrl, rxq);
564 
565 	if (!rxq)
566 		return;
567 	qinfo->mp = mlx5_rxq_mprq_enabled(&rxq_ctrl->rxq) ?
568 					rxq->mprq_mp : rxq->mp;
569 	qinfo->conf.rx_thresh.pthresh = 0;
570 	qinfo->conf.rx_thresh.hthresh = 0;
571 	qinfo->conf.rx_thresh.wthresh = 0;
572 	qinfo->conf.rx_free_thresh = rxq->rq_repl_thresh;
573 	qinfo->conf.rx_drop_en = 1;
574 	qinfo->conf.rx_deferred_start = rxq_ctrl ? 0 : 1;
575 	qinfo->conf.offloads = dev->data->dev_conf.rxmode.offloads;
576 	qinfo->scattered_rx = dev->data->scattered_rx;
577 	qinfo->nb_desc = 1 << rxq->elts_n;
578 }
579 
580 /**
581  * DPDK callback to get the RX packet burst mode information
582  *
583  * @param dev
584  *   Pointer to the device structure.
585  *
586  * @param rx_queue_id
587  *   Rx queue identificatior.
588  *
589  * @param mode
590  *   Pointer to the burts mode information.
591  *
592  * @return
593  *   0 as success, -EINVAL as failure.
594  */
595 
596 int
597 mlx5_rx_burst_mode_get(struct rte_eth_dev *dev,
598 		       uint16_t rx_queue_id __rte_unused,
599 		       struct rte_eth_burst_mode *mode)
600 {
601 	eth_rx_burst_t pkt_burst = dev->rx_pkt_burst;
602 
603 	if (pkt_burst == mlx5_rx_burst) {
604 		snprintf(mode->info, sizeof(mode->info), "%s", "Scalar");
605 	} else if (pkt_burst == mlx5_rx_burst_mprq) {
606 		snprintf(mode->info, sizeof(mode->info), "%s", "Multi-Packet RQ");
607 	} else if (pkt_burst == mlx5_rx_burst_vec) {
608 #if defined RTE_ARCH_X86_64
609 		snprintf(mode->info, sizeof(mode->info), "%s", "Vector SSE");
610 #elif defined RTE_ARCH_ARM64
611 		snprintf(mode->info, sizeof(mode->info), "%s", "Vector Neon");
612 #elif defined RTE_ARCH_PPC_64
613 		snprintf(mode->info, sizeof(mode->info), "%s", "Vector AltiVec");
614 #else
615 		return -EINVAL;
616 #endif
617 	} else {
618 		return -EINVAL;
619 	}
620 	return 0;
621 }
622 
623 /**
624  * DPDK callback to get the number of used descriptors in a RX queue
625  *
626  * @param dev
627  *   Pointer to the device structure.
628  *
629  * @param rx_queue_id
630  *   The Rx queue.
631  *
632  * @return
633  *   The number of used rx descriptor.
634  *   -EINVAL if the queue is invalid
635  */
636 uint32_t
637 mlx5_rx_queue_count(struct rte_eth_dev *dev, uint16_t rx_queue_id)
638 {
639 	struct mlx5_priv *priv = dev->data->dev_private;
640 	struct mlx5_rxq_data *rxq;
641 
642 	if (dev->rx_pkt_burst != mlx5_rx_burst) {
643 		rte_errno = ENOTSUP;
644 		return -rte_errno;
645 	}
646 	rxq = (*priv->rxqs)[rx_queue_id];
647 	if (!rxq) {
648 		rte_errno = EINVAL;
649 		return -rte_errno;
650 	}
651 	return rx_queue_count(rxq);
652 }
653 
654 #define MLX5_SYSTEM_LOG_DIR "/var/log"
655 /**
656  * Dump debug information to log file.
657  *
658  * @param fname
659  *   The file name.
660  * @param hex_title
661  *   If not NULL this string is printed as a header to the output
662  *   and the output will be in hexadecimal view.
663  * @param buf
664  *   This is the buffer address to print out.
665  * @param len
666  *   The number of bytes to dump out.
667  */
668 void
669 mlx5_dump_debug_information(const char *fname, const char *hex_title,
670 			    const void *buf, unsigned int hex_len)
671 {
672 	FILE *fd;
673 
674 	MKSTR(path, "%s/%s", MLX5_SYSTEM_LOG_DIR, fname);
675 	fd = fopen(path, "a+");
676 	if (!fd) {
677 		DRV_LOG(WARNING, "cannot open %s for debug dump", path);
678 		MKSTR(path2, "./%s", fname);
679 		fd = fopen(path2, "a+");
680 		if (!fd) {
681 			DRV_LOG(ERR, "cannot open %s for debug dump", path2);
682 			return;
683 		}
684 		DRV_LOG(INFO, "New debug dump in file %s", path2);
685 	} else {
686 		DRV_LOG(INFO, "New debug dump in file %s", path);
687 	}
688 	if (hex_title)
689 		rte_hexdump(fd, hex_title, buf, hex_len);
690 	else
691 		fprintf(fd, "%s", (const char *)buf);
692 	fprintf(fd, "\n\n\n");
693 	fclose(fd);
694 }
695 
696 /**
697  * Move QP from error state to running state and initialize indexes.
698  *
699  * @param txq_ctrl
700  *   Pointer to TX queue control structure.
701  *
702  * @return
703  *   0 on success, else -1.
704  */
705 static int
706 tx_recover_qp(struct mlx5_txq_ctrl *txq_ctrl)
707 {
708 	struct mlx5_mp_arg_queue_state_modify sm = {
709 			.is_wq = 0,
710 			.queue_id = txq_ctrl->txq.idx,
711 	};
712 
713 	if (mlx5_queue_state_modify(ETH_DEV(txq_ctrl->priv), &sm))
714 		return -1;
715 	txq_ctrl->txq.wqe_ci = 0;
716 	txq_ctrl->txq.wqe_pi = 0;
717 	txq_ctrl->txq.elts_comp = 0;
718 	return 0;
719 }
720 
721 /* Return 1 if the error CQE is signed otherwise, sign it and return 0. */
722 static int
723 check_err_cqe_seen(volatile struct mlx5_err_cqe *err_cqe)
724 {
725 	static const uint8_t magic[] = "seen";
726 	int ret = 1;
727 	unsigned int i;
728 
729 	for (i = 0; i < sizeof(magic); ++i)
730 		if (!ret || err_cqe->rsvd1[i] != magic[i]) {
731 			ret = 0;
732 			err_cqe->rsvd1[i] = magic[i];
733 		}
734 	return ret;
735 }
736 
737 /**
738  * Handle error CQE.
739  *
740  * @param txq
741  *   Pointer to TX queue structure.
742  * @param error_cqe
743  *   Pointer to the error CQE.
744  *
745  * @return
746  *   Negative value if queue recovery failed, otherwise
747  *   the error completion entry is handled successfully.
748  */
749 static int
750 mlx5_tx_error_cqe_handle(struct mlx5_txq_data *restrict txq,
751 			 volatile struct mlx5_err_cqe *err_cqe)
752 {
753 	if (err_cqe->syndrome != MLX5_CQE_SYNDROME_WR_FLUSH_ERR) {
754 		const uint16_t wqe_m = ((1 << txq->wqe_n) - 1);
755 		struct mlx5_txq_ctrl *txq_ctrl =
756 				container_of(txq, struct mlx5_txq_ctrl, txq);
757 		uint16_t new_wqe_pi = rte_be_to_cpu_16(err_cqe->wqe_counter);
758 		int seen = check_err_cqe_seen(err_cqe);
759 
760 		if (!seen && txq_ctrl->dump_file_n <
761 		    txq_ctrl->priv->config.max_dump_files_num) {
762 			MKSTR(err_str, "Unexpected CQE error syndrome "
763 			      "0x%02x CQN = %u SQN = %u wqe_counter = %u "
764 			      "wq_ci = %u cq_ci = %u", err_cqe->syndrome,
765 			      txq->cqe_s, txq->qp_num_8s >> 8,
766 			      rte_be_to_cpu_16(err_cqe->wqe_counter),
767 			      txq->wqe_ci, txq->cq_ci);
768 			MKSTR(name, "dpdk_mlx5_port_%u_txq_%u_index_%u_%u",
769 			      PORT_ID(txq_ctrl->priv), txq->idx,
770 			      txq_ctrl->dump_file_n, (uint32_t)rte_rdtsc());
771 			mlx5_dump_debug_information(name, NULL, err_str, 0);
772 			mlx5_dump_debug_information(name, "MLX5 Error CQ:",
773 						    (const void *)((uintptr_t)
774 						    txq->cqes),
775 						    sizeof(*err_cqe) *
776 						    (1 << txq->cqe_n));
777 			mlx5_dump_debug_information(name, "MLX5 Error SQ:",
778 						    (const void *)((uintptr_t)
779 						    txq->wqes),
780 						    MLX5_WQE_SIZE *
781 						    (1 << txq->wqe_n));
782 			txq_ctrl->dump_file_n++;
783 		}
784 		if (!seen)
785 			/*
786 			 * Count errors in WQEs units.
787 			 * Later it can be improved to count error packets,
788 			 * for example, by SQ parsing to find how much packets
789 			 * should be counted for each WQE.
790 			 */
791 			txq->stats.oerrors += ((txq->wqe_ci & wqe_m) -
792 						new_wqe_pi) & wqe_m;
793 		if (tx_recover_qp(txq_ctrl)) {
794 			/* Recovering failed - retry later on the same WQE. */
795 			return -1;
796 		}
797 		/* Release all the remaining buffers. */
798 		txq_free_elts(txq_ctrl);
799 	}
800 	return 0;
801 }
802 
803 /**
804  * Translate RX completion flags to packet type.
805  *
806  * @param[in] rxq
807  *   Pointer to RX queue structure.
808  * @param[in] cqe
809  *   Pointer to CQE.
810  *
811  * @note: fix mlx5_dev_supported_ptypes_get() if any change here.
812  *
813  * @return
814  *   Packet type for struct rte_mbuf.
815  */
816 static inline uint32_t
817 rxq_cq_to_pkt_type(struct mlx5_rxq_data *rxq, volatile struct mlx5_cqe *cqe)
818 {
819 	uint8_t idx;
820 	uint8_t pinfo = cqe->pkt_info;
821 	uint16_t ptype = cqe->hdr_type_etc;
822 
823 	/*
824 	 * The index to the array should have:
825 	 * bit[1:0] = l3_hdr_type
826 	 * bit[4:2] = l4_hdr_type
827 	 * bit[5] = ip_frag
828 	 * bit[6] = tunneled
829 	 * bit[7] = outer_l3_type
830 	 */
831 	idx = ((pinfo & 0x3) << 6) | ((ptype & 0xfc00) >> 10);
832 	return mlx5_ptype_table[idx] | rxq->tunnel * !!(idx & (1 << 6));
833 }
834 
835 /**
836  * Initialize Rx WQ and indexes.
837  *
838  * @param[in] rxq
839  *   Pointer to RX queue structure.
840  */
841 void
842 mlx5_rxq_initialize(struct mlx5_rxq_data *rxq)
843 {
844 	const unsigned int wqe_n = 1 << rxq->elts_n;
845 	unsigned int i;
846 
847 	for (i = 0; (i != wqe_n); ++i) {
848 		volatile struct mlx5_wqe_data_seg *scat;
849 		uintptr_t addr;
850 		uint32_t byte_count;
851 
852 		if (mlx5_rxq_mprq_enabled(rxq)) {
853 			struct mlx5_mprq_buf *buf = (*rxq->mprq_bufs)[i];
854 
855 			scat = &((volatile struct mlx5_wqe_mprq *)
856 				rxq->wqes)[i].dseg;
857 			addr = (uintptr_t)mlx5_mprq_buf_addr(buf,
858 							 1 << rxq->strd_num_n);
859 			byte_count = (1 << rxq->strd_sz_n) *
860 					(1 << rxq->strd_num_n);
861 		} else {
862 			struct rte_mbuf *buf = (*rxq->elts)[i];
863 
864 			scat = &((volatile struct mlx5_wqe_data_seg *)
865 					rxq->wqes)[i];
866 			addr = rte_pktmbuf_mtod(buf, uintptr_t);
867 			byte_count = DATA_LEN(buf);
868 		}
869 		/* scat->addr must be able to store a pointer. */
870 		MLX5_ASSERT(sizeof(scat->addr) >= sizeof(uintptr_t));
871 		*scat = (struct mlx5_wqe_data_seg){
872 			.addr = rte_cpu_to_be_64(addr),
873 			.byte_count = rte_cpu_to_be_32(byte_count),
874 			.lkey = mlx5_rx_addr2mr(rxq, addr),
875 		};
876 	}
877 	rxq->consumed_strd = 0;
878 	rxq->decompressed = 0;
879 	rxq->rq_pi = 0;
880 	rxq->zip = (struct rxq_zip){
881 		.ai = 0,
882 	};
883 	/* Update doorbell counter. */
884 	rxq->rq_ci = wqe_n >> rxq->sges_n;
885 	rte_cio_wmb();
886 	*rxq->rq_db = rte_cpu_to_be_32(rxq->rq_ci);
887 }
888 
889 /**
890  * Modify a Verbs/DevX queue state.
891  * This must be called from the primary process.
892  *
893  * @param dev
894  *   Pointer to Ethernet device.
895  * @param sm
896  *   State modify request parameters.
897  *
898  * @return
899  *   0 in case of success else non-zero value and rte_errno is set.
900  */
901 int
902 mlx5_queue_state_modify_primary(struct rte_eth_dev *dev,
903 			const struct mlx5_mp_arg_queue_state_modify *sm)
904 {
905 	int ret;
906 	struct mlx5_priv *priv = dev->data->dev_private;
907 
908 	if (sm->is_wq) {
909 		struct mlx5_rxq_data *rxq = (*priv->rxqs)[sm->queue_id];
910 		struct mlx5_rxq_ctrl *rxq_ctrl =
911 			container_of(rxq, struct mlx5_rxq_ctrl, rxq);
912 
913 		if (rxq_ctrl->obj->type == MLX5_RXQ_OBJ_TYPE_IBV) {
914 			struct ibv_wq_attr mod = {
915 				.attr_mask = IBV_WQ_ATTR_STATE,
916 				.wq_state = sm->state,
917 			};
918 
919 			ret = mlx5_glue->modify_wq(rxq_ctrl->obj->wq, &mod);
920 		} else { /* rxq_ctrl->obj->type == MLX5_RXQ_OBJ_TYPE_DEVX_RQ. */
921 			struct mlx5_devx_modify_rq_attr rq_attr;
922 
923 			memset(&rq_attr, 0, sizeof(rq_attr));
924 			if (sm->state == IBV_WQS_RESET) {
925 				rq_attr.rq_state = MLX5_RQC_STATE_ERR;
926 				rq_attr.state = MLX5_RQC_STATE_RST;
927 			} else if (sm->state == IBV_WQS_RDY) {
928 				rq_attr.rq_state = MLX5_RQC_STATE_RST;
929 				rq_attr.state = MLX5_RQC_STATE_RDY;
930 			} else if (sm->state == IBV_WQS_ERR) {
931 				rq_attr.rq_state = MLX5_RQC_STATE_RDY;
932 				rq_attr.state = MLX5_RQC_STATE_ERR;
933 			}
934 			ret = mlx5_devx_cmd_modify_rq(rxq_ctrl->obj->rq,
935 						      &rq_attr);
936 		}
937 		if (ret) {
938 			DRV_LOG(ERR, "Cannot change Rx WQ state to %u  - %s",
939 					sm->state, strerror(errno));
940 			rte_errno = errno;
941 			return ret;
942 		}
943 	} else {
944 		struct mlx5_txq_data *txq = (*priv->txqs)[sm->queue_id];
945 		struct mlx5_txq_ctrl *txq_ctrl =
946 			container_of(txq, struct mlx5_txq_ctrl, txq);
947 		struct ibv_qp_attr mod = {
948 			.qp_state = IBV_QPS_RESET,
949 			.port_num = (uint8_t)priv->ibv_port,
950 		};
951 		struct ibv_qp *qp = txq_ctrl->obj->qp;
952 
953 		ret = mlx5_glue->modify_qp(qp, &mod, IBV_QP_STATE);
954 		if (ret) {
955 			DRV_LOG(ERR, "Cannot change the Tx QP state to RESET "
956 				"%s", strerror(errno));
957 			rte_errno = errno;
958 			return ret;
959 		}
960 		mod.qp_state = IBV_QPS_INIT;
961 		ret = mlx5_glue->modify_qp(qp, &mod,
962 					   (IBV_QP_STATE | IBV_QP_PORT));
963 		if (ret) {
964 			DRV_LOG(ERR, "Cannot change Tx QP state to INIT %s",
965 				strerror(errno));
966 			rte_errno = errno;
967 			return ret;
968 		}
969 		mod.qp_state = IBV_QPS_RTR;
970 		ret = mlx5_glue->modify_qp(qp, &mod, IBV_QP_STATE);
971 		if (ret) {
972 			DRV_LOG(ERR, "Cannot change Tx QP state to RTR %s",
973 				strerror(errno));
974 			rte_errno = errno;
975 			return ret;
976 		}
977 		mod.qp_state = IBV_QPS_RTS;
978 		ret = mlx5_glue->modify_qp(qp, &mod, IBV_QP_STATE);
979 		if (ret) {
980 			DRV_LOG(ERR, "Cannot change Tx QP state to RTS %s",
981 				strerror(errno));
982 			rte_errno = errno;
983 			return ret;
984 		}
985 	}
986 	return 0;
987 }
988 
989 /**
990  * Modify a Verbs queue state.
991  *
992  * @param dev
993  *   Pointer to Ethernet device.
994  * @param sm
995  *   State modify request parameters.
996  *
997  * @return
998  *   0 in case of success else non-zero value.
999  */
1000 static int
1001 mlx5_queue_state_modify(struct rte_eth_dev *dev,
1002 			struct mlx5_mp_arg_queue_state_modify *sm)
1003 {
1004 	struct mlx5_priv *priv = dev->data->dev_private;
1005 	int ret = 0;
1006 
1007 	switch (rte_eal_process_type()) {
1008 	case RTE_PROC_PRIMARY:
1009 		ret = mlx5_queue_state_modify_primary(dev, sm);
1010 		break;
1011 	case RTE_PROC_SECONDARY:
1012 		ret = mlx5_mp_req_queue_state_modify(&priv->mp_id, sm);
1013 		break;
1014 	default:
1015 		break;
1016 	}
1017 	return ret;
1018 }
1019 
1020 /**
1021  * Handle a Rx error.
1022  * The function inserts the RQ state to reset when the first error CQE is
1023  * shown, then drains the CQ by the caller function loop. When the CQ is empty,
1024  * it moves the RQ state to ready and initializes the RQ.
1025  * Next CQE identification and error counting are in the caller responsibility.
1026  *
1027  * @param[in] rxq
1028  *   Pointer to RX queue structure.
1029  * @param[in] vec
1030  *   1 when called from vectorized Rx burst, need to prepare mbufs for the RQ.
1031  *   0 when called from non-vectorized Rx burst.
1032  *
1033  * @return
1034  *   -1 in case of recovery error, otherwise the CQE status.
1035  */
1036 int
1037 mlx5_rx_err_handle(struct mlx5_rxq_data *rxq, uint8_t vec)
1038 {
1039 	const uint16_t cqe_n = 1 << rxq->cqe_n;
1040 	const uint16_t cqe_mask = cqe_n - 1;
1041 	const unsigned int wqe_n = 1 << rxq->elts_n;
1042 	struct mlx5_rxq_ctrl *rxq_ctrl =
1043 			container_of(rxq, struct mlx5_rxq_ctrl, rxq);
1044 	union {
1045 		volatile struct mlx5_cqe *cqe;
1046 		volatile struct mlx5_err_cqe *err_cqe;
1047 	} u = {
1048 		.cqe = &(*rxq->cqes)[rxq->cq_ci & cqe_mask],
1049 	};
1050 	struct mlx5_mp_arg_queue_state_modify sm;
1051 	int ret;
1052 
1053 	switch (rxq->err_state) {
1054 	case MLX5_RXQ_ERR_STATE_NO_ERROR:
1055 		rxq->err_state = MLX5_RXQ_ERR_STATE_NEED_RESET;
1056 		/* Fall-through */
1057 	case MLX5_RXQ_ERR_STATE_NEED_RESET:
1058 		sm.is_wq = 1;
1059 		sm.queue_id = rxq->idx;
1060 		sm.state = IBV_WQS_RESET;
1061 		if (mlx5_queue_state_modify(ETH_DEV(rxq_ctrl->priv), &sm))
1062 			return -1;
1063 		if (rxq_ctrl->dump_file_n <
1064 		    rxq_ctrl->priv->config.max_dump_files_num) {
1065 			MKSTR(err_str, "Unexpected CQE error syndrome "
1066 			      "0x%02x CQN = %u RQN = %u wqe_counter = %u"
1067 			      " rq_ci = %u cq_ci = %u", u.err_cqe->syndrome,
1068 			      rxq->cqn, rxq_ctrl->wqn,
1069 			      rte_be_to_cpu_16(u.err_cqe->wqe_counter),
1070 			      rxq->rq_ci << rxq->sges_n, rxq->cq_ci);
1071 			MKSTR(name, "dpdk_mlx5_port_%u_rxq_%u_%u",
1072 			      rxq->port_id, rxq->idx, (uint32_t)rte_rdtsc());
1073 			mlx5_dump_debug_information(name, NULL, err_str, 0);
1074 			mlx5_dump_debug_information(name, "MLX5 Error CQ:",
1075 						    (const void *)((uintptr_t)
1076 								    rxq->cqes),
1077 						    sizeof(*u.cqe) * cqe_n);
1078 			mlx5_dump_debug_information(name, "MLX5 Error RQ:",
1079 						    (const void *)((uintptr_t)
1080 								    rxq->wqes),
1081 						    16 * wqe_n);
1082 			rxq_ctrl->dump_file_n++;
1083 		}
1084 		rxq->err_state = MLX5_RXQ_ERR_STATE_NEED_READY;
1085 		/* Fall-through */
1086 	case MLX5_RXQ_ERR_STATE_NEED_READY:
1087 		ret = check_cqe(u.cqe, cqe_n, rxq->cq_ci);
1088 		if (ret == MLX5_CQE_STATUS_HW_OWN) {
1089 			rte_cio_wmb();
1090 			*rxq->cq_db = rte_cpu_to_be_32(rxq->cq_ci);
1091 			rte_cio_wmb();
1092 			/*
1093 			 * The RQ consumer index must be zeroed while moving
1094 			 * from RESET state to RDY state.
1095 			 */
1096 			*rxq->rq_db = rte_cpu_to_be_32(0);
1097 			rte_cio_wmb();
1098 			sm.is_wq = 1;
1099 			sm.queue_id = rxq->idx;
1100 			sm.state = IBV_WQS_RDY;
1101 			if (mlx5_queue_state_modify(ETH_DEV(rxq_ctrl->priv),
1102 						    &sm))
1103 				return -1;
1104 			if (vec) {
1105 				const uint16_t q_mask = wqe_n - 1;
1106 				uint16_t elt_idx;
1107 				struct rte_mbuf **elt;
1108 				int i;
1109 				unsigned int n = wqe_n - (rxq->rq_ci -
1110 							  rxq->rq_pi);
1111 
1112 				for (i = 0; i < (int)n; ++i) {
1113 					elt_idx = (rxq->rq_ci + i) & q_mask;
1114 					elt = &(*rxq->elts)[elt_idx];
1115 					*elt = rte_mbuf_raw_alloc(rxq->mp);
1116 					if (!*elt) {
1117 						for (i--; i >= 0; --i) {
1118 							elt_idx = (rxq->rq_ci +
1119 								   i) & q_mask;
1120 							elt = &(*rxq->elts)
1121 								[elt_idx];
1122 							rte_pktmbuf_free_seg
1123 								(*elt);
1124 						}
1125 						return -1;
1126 					}
1127 				}
1128 				for (i = 0; i < (int)wqe_n; ++i) {
1129 					elt = &(*rxq->elts)[i];
1130 					DATA_LEN(*elt) =
1131 						(uint16_t)((*elt)->buf_len -
1132 						rte_pktmbuf_headroom(*elt));
1133 				}
1134 				/* Padding with a fake mbuf for vec Rx. */
1135 				for (i = 0; i < MLX5_VPMD_DESCS_PER_LOOP; ++i)
1136 					(*rxq->elts)[wqe_n + i] =
1137 								&rxq->fake_mbuf;
1138 			}
1139 			mlx5_rxq_initialize(rxq);
1140 			rxq->err_state = MLX5_RXQ_ERR_STATE_NO_ERROR;
1141 		}
1142 		return ret;
1143 	default:
1144 		return -1;
1145 	}
1146 }
1147 
1148 /**
1149  * Get size of the next packet for a given CQE. For compressed CQEs, the
1150  * consumer index is updated only once all packets of the current one have
1151  * been processed.
1152  *
1153  * @param rxq
1154  *   Pointer to RX queue.
1155  * @param cqe
1156  *   CQE to process.
1157  * @param[out] mcqe
1158  *   Store pointer to mini-CQE if compressed. Otherwise, the pointer is not
1159  *   written.
1160  *
1161  * @return
1162  *   0 in case of empty CQE, otherwise the packet size in bytes.
1163  */
1164 static inline int
1165 mlx5_rx_poll_len(struct mlx5_rxq_data *rxq, volatile struct mlx5_cqe *cqe,
1166 		 uint16_t cqe_cnt, volatile struct mlx5_mini_cqe8 **mcqe)
1167 {
1168 	struct rxq_zip *zip = &rxq->zip;
1169 	uint16_t cqe_n = cqe_cnt + 1;
1170 	int len;
1171 	uint16_t idx, end;
1172 
1173 	do {
1174 		len = 0;
1175 		/* Process compressed data in the CQE and mini arrays. */
1176 		if (zip->ai) {
1177 			volatile struct mlx5_mini_cqe8 (*mc)[8] =
1178 				(volatile struct mlx5_mini_cqe8 (*)[8])
1179 				(uintptr_t)(&(*rxq->cqes)[zip->ca &
1180 							  cqe_cnt].pkt_info);
1181 
1182 			len = rte_be_to_cpu_32((*mc)[zip->ai & 7].byte_cnt);
1183 			*mcqe = &(*mc)[zip->ai & 7];
1184 			if ((++zip->ai & 7) == 0) {
1185 				/* Invalidate consumed CQEs */
1186 				idx = zip->ca;
1187 				end = zip->na;
1188 				while (idx != end) {
1189 					(*rxq->cqes)[idx & cqe_cnt].op_own =
1190 						MLX5_CQE_INVALIDATE;
1191 					++idx;
1192 				}
1193 				/*
1194 				 * Increment consumer index to skip the number
1195 				 * of CQEs consumed. Hardware leaves holes in
1196 				 * the CQ ring for software use.
1197 				 */
1198 				zip->ca = zip->na;
1199 				zip->na += 8;
1200 			}
1201 			if (unlikely(rxq->zip.ai == rxq->zip.cqe_cnt)) {
1202 				/* Invalidate the rest */
1203 				idx = zip->ca;
1204 				end = zip->cq_ci;
1205 
1206 				while (idx != end) {
1207 					(*rxq->cqes)[idx & cqe_cnt].op_own =
1208 						MLX5_CQE_INVALIDATE;
1209 					++idx;
1210 				}
1211 				rxq->cq_ci = zip->cq_ci;
1212 				zip->ai = 0;
1213 			}
1214 		/*
1215 		 * No compressed data, get next CQE and verify if it is
1216 		 * compressed.
1217 		 */
1218 		} else {
1219 			int ret;
1220 			int8_t op_own;
1221 
1222 			ret = check_cqe(cqe, cqe_n, rxq->cq_ci);
1223 			if (unlikely(ret != MLX5_CQE_STATUS_SW_OWN)) {
1224 				if (unlikely(ret == MLX5_CQE_STATUS_ERR ||
1225 					     rxq->err_state)) {
1226 					ret = mlx5_rx_err_handle(rxq, 0);
1227 					if (ret == MLX5_CQE_STATUS_HW_OWN ||
1228 					    ret == -1)
1229 						return 0;
1230 				} else {
1231 					return 0;
1232 				}
1233 			}
1234 			++rxq->cq_ci;
1235 			op_own = cqe->op_own;
1236 			if (MLX5_CQE_FORMAT(op_own) == MLX5_COMPRESSED) {
1237 				volatile struct mlx5_mini_cqe8 (*mc)[8] =
1238 					(volatile struct mlx5_mini_cqe8 (*)[8])
1239 					(uintptr_t)(&(*rxq->cqes)
1240 						[rxq->cq_ci &
1241 						 cqe_cnt].pkt_info);
1242 
1243 				/* Fix endianness. */
1244 				zip->cqe_cnt = rte_be_to_cpu_32(cqe->byte_cnt);
1245 				/*
1246 				 * Current mini array position is the one
1247 				 * returned by check_cqe64().
1248 				 *
1249 				 * If completion comprises several mini arrays,
1250 				 * as a special case the second one is located
1251 				 * 7 CQEs after the initial CQE instead of 8
1252 				 * for subsequent ones.
1253 				 */
1254 				zip->ca = rxq->cq_ci;
1255 				zip->na = zip->ca + 7;
1256 				/* Compute the next non compressed CQE. */
1257 				--rxq->cq_ci;
1258 				zip->cq_ci = rxq->cq_ci + zip->cqe_cnt;
1259 				/* Get packet size to return. */
1260 				len = rte_be_to_cpu_32((*mc)[0].byte_cnt);
1261 				*mcqe = &(*mc)[0];
1262 				zip->ai = 1;
1263 				/* Prefetch all to be invalidated */
1264 				idx = zip->ca;
1265 				end = zip->cq_ci;
1266 				while (idx != end) {
1267 					rte_prefetch0(&(*rxq->cqes)[(idx) &
1268 								    cqe_cnt]);
1269 					++idx;
1270 				}
1271 			} else {
1272 				len = rte_be_to_cpu_32(cqe->byte_cnt);
1273 			}
1274 		}
1275 		if (unlikely(rxq->err_state)) {
1276 			cqe = &(*rxq->cqes)[rxq->cq_ci & cqe_cnt];
1277 			++rxq->stats.idropped;
1278 		} else {
1279 			return len;
1280 		}
1281 	} while (1);
1282 }
1283 
1284 /**
1285  * Translate RX completion flags to offload flags.
1286  *
1287  * @param[in] cqe
1288  *   Pointer to CQE.
1289  *
1290  * @return
1291  *   Offload flags (ol_flags) for struct rte_mbuf.
1292  */
1293 static inline uint32_t
1294 rxq_cq_to_ol_flags(volatile struct mlx5_cqe *cqe)
1295 {
1296 	uint32_t ol_flags = 0;
1297 	uint16_t flags = rte_be_to_cpu_16(cqe->hdr_type_etc);
1298 
1299 	ol_flags =
1300 		TRANSPOSE(flags,
1301 			  MLX5_CQE_RX_L3_HDR_VALID,
1302 			  PKT_RX_IP_CKSUM_GOOD) |
1303 		TRANSPOSE(flags,
1304 			  MLX5_CQE_RX_L4_HDR_VALID,
1305 			  PKT_RX_L4_CKSUM_GOOD);
1306 	return ol_flags;
1307 }
1308 
1309 /**
1310  * Fill in mbuf fields from RX completion flags.
1311  * Note that pkt->ol_flags should be initialized outside of this function.
1312  *
1313  * @param rxq
1314  *   Pointer to RX queue.
1315  * @param pkt
1316  *   mbuf to fill.
1317  * @param cqe
1318  *   CQE to process.
1319  * @param rss_hash_res
1320  *   Packet RSS Hash result.
1321  */
1322 static inline void
1323 rxq_cq_to_mbuf(struct mlx5_rxq_data *rxq, struct rte_mbuf *pkt,
1324 	       volatile struct mlx5_cqe *cqe, uint32_t rss_hash_res)
1325 {
1326 	/* Update packet information. */
1327 	pkt->packet_type = rxq_cq_to_pkt_type(rxq, cqe);
1328 	if (rss_hash_res && rxq->rss_hash) {
1329 		pkt->hash.rss = rss_hash_res;
1330 		pkt->ol_flags |= PKT_RX_RSS_HASH;
1331 	}
1332 	if (rxq->mark && MLX5_FLOW_MARK_IS_VALID(cqe->sop_drop_qpn)) {
1333 		pkt->ol_flags |= PKT_RX_FDIR;
1334 		if (cqe->sop_drop_qpn !=
1335 		    rte_cpu_to_be_32(MLX5_FLOW_MARK_DEFAULT)) {
1336 			uint32_t mark = cqe->sop_drop_qpn;
1337 
1338 			pkt->ol_flags |= PKT_RX_FDIR_ID;
1339 			pkt->hash.fdir.hi = mlx5_flow_mark_get(mark);
1340 		}
1341 	}
1342 	if (rxq->dynf_meta && cqe->flow_table_metadata) {
1343 		pkt->ol_flags |= rxq->flow_meta_mask;
1344 		*RTE_MBUF_DYNFIELD(pkt, rxq->flow_meta_offset, uint32_t *) =
1345 			cqe->flow_table_metadata;
1346 	}
1347 	if (rxq->csum)
1348 		pkt->ol_flags |= rxq_cq_to_ol_flags(cqe);
1349 	if (rxq->vlan_strip &&
1350 	    (cqe->hdr_type_etc & rte_cpu_to_be_16(MLX5_CQE_VLAN_STRIPPED))) {
1351 		pkt->ol_flags |= PKT_RX_VLAN | PKT_RX_VLAN_STRIPPED;
1352 		pkt->vlan_tci = rte_be_to_cpu_16(cqe->vlan_info);
1353 	}
1354 	if (rxq->hw_timestamp) {
1355 		pkt->timestamp = rte_be_to_cpu_64(cqe->timestamp);
1356 		pkt->ol_flags |= PKT_RX_TIMESTAMP;
1357 	}
1358 }
1359 
1360 /**
1361  * DPDK callback for RX.
1362  *
1363  * @param dpdk_rxq
1364  *   Generic pointer to RX queue structure.
1365  * @param[out] pkts
1366  *   Array to store received packets.
1367  * @param pkts_n
1368  *   Maximum number of packets in array.
1369  *
1370  * @return
1371  *   Number of packets successfully received (<= pkts_n).
1372  */
1373 uint16_t
1374 mlx5_rx_burst(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n)
1375 {
1376 	struct mlx5_rxq_data *rxq = dpdk_rxq;
1377 	const unsigned int wqe_cnt = (1 << rxq->elts_n) - 1;
1378 	const unsigned int cqe_cnt = (1 << rxq->cqe_n) - 1;
1379 	const unsigned int sges_n = rxq->sges_n;
1380 	struct rte_mbuf *pkt = NULL;
1381 	struct rte_mbuf *seg = NULL;
1382 	volatile struct mlx5_cqe *cqe =
1383 		&(*rxq->cqes)[rxq->cq_ci & cqe_cnt];
1384 	unsigned int i = 0;
1385 	unsigned int rq_ci = rxq->rq_ci << sges_n;
1386 	int len = 0; /* keep its value across iterations. */
1387 
1388 	while (pkts_n) {
1389 		unsigned int idx = rq_ci & wqe_cnt;
1390 		volatile struct mlx5_wqe_data_seg *wqe =
1391 			&((volatile struct mlx5_wqe_data_seg *)rxq->wqes)[idx];
1392 		struct rte_mbuf *rep = (*rxq->elts)[idx];
1393 		volatile struct mlx5_mini_cqe8 *mcqe = NULL;
1394 		uint32_t rss_hash_res;
1395 
1396 		if (pkt)
1397 			NEXT(seg) = rep;
1398 		seg = rep;
1399 		rte_prefetch0(seg);
1400 		rte_prefetch0(cqe);
1401 		rte_prefetch0(wqe);
1402 		rep = rte_mbuf_raw_alloc(rxq->mp);
1403 		if (unlikely(rep == NULL)) {
1404 			++rxq->stats.rx_nombuf;
1405 			if (!pkt) {
1406 				/*
1407 				 * no buffers before we even started,
1408 				 * bail out silently.
1409 				 */
1410 				break;
1411 			}
1412 			while (pkt != seg) {
1413 				MLX5_ASSERT(pkt != (*rxq->elts)[idx]);
1414 				rep = NEXT(pkt);
1415 				NEXT(pkt) = NULL;
1416 				NB_SEGS(pkt) = 1;
1417 				rte_mbuf_raw_free(pkt);
1418 				pkt = rep;
1419 			}
1420 			break;
1421 		}
1422 		if (!pkt) {
1423 			cqe = &(*rxq->cqes)[rxq->cq_ci & cqe_cnt];
1424 			len = mlx5_rx_poll_len(rxq, cqe, cqe_cnt, &mcqe);
1425 			if (!len) {
1426 				rte_mbuf_raw_free(rep);
1427 				break;
1428 			}
1429 			pkt = seg;
1430 			MLX5_ASSERT(len >= (rxq->crc_present << 2));
1431 			pkt->ol_flags &= EXT_ATTACHED_MBUF;
1432 			/* If compressed, take hash result from mini-CQE. */
1433 			rss_hash_res = rte_be_to_cpu_32(mcqe == NULL ?
1434 							cqe->rx_hash_res :
1435 							mcqe->rx_hash_result);
1436 			rxq_cq_to_mbuf(rxq, pkt, cqe, rss_hash_res);
1437 			if (rxq->crc_present)
1438 				len -= RTE_ETHER_CRC_LEN;
1439 			PKT_LEN(pkt) = len;
1440 			if (cqe->lro_num_seg > 1) {
1441 				mlx5_lro_update_hdr
1442 					(rte_pktmbuf_mtod(pkt, uint8_t *), cqe,
1443 					 len);
1444 				pkt->ol_flags |= PKT_RX_LRO;
1445 				pkt->tso_segsz = len / cqe->lro_num_seg;
1446 			}
1447 		}
1448 		DATA_LEN(rep) = DATA_LEN(seg);
1449 		PKT_LEN(rep) = PKT_LEN(seg);
1450 		SET_DATA_OFF(rep, DATA_OFF(seg));
1451 		PORT(rep) = PORT(seg);
1452 		(*rxq->elts)[idx] = rep;
1453 		/*
1454 		 * Fill NIC descriptor with the new buffer.  The lkey and size
1455 		 * of the buffers are already known, only the buffer address
1456 		 * changes.
1457 		 */
1458 		wqe->addr = rte_cpu_to_be_64(rte_pktmbuf_mtod(rep, uintptr_t));
1459 		/* If there's only one MR, no need to replace LKey in WQE. */
1460 		if (unlikely(mlx5_mr_btree_len(&rxq->mr_ctrl.cache_bh) > 1))
1461 			wqe->lkey = mlx5_rx_mb2mr(rxq, rep);
1462 		if (len > DATA_LEN(seg)) {
1463 			len -= DATA_LEN(seg);
1464 			++NB_SEGS(pkt);
1465 			++rq_ci;
1466 			continue;
1467 		}
1468 		DATA_LEN(seg) = len;
1469 #ifdef MLX5_PMD_SOFT_COUNTERS
1470 		/* Increment bytes counter. */
1471 		rxq->stats.ibytes += PKT_LEN(pkt);
1472 #endif
1473 		/* Return packet. */
1474 		*(pkts++) = pkt;
1475 		pkt = NULL;
1476 		--pkts_n;
1477 		++i;
1478 		/* Align consumer index to the next stride. */
1479 		rq_ci >>= sges_n;
1480 		++rq_ci;
1481 		rq_ci <<= sges_n;
1482 	}
1483 	if (unlikely((i == 0) && ((rq_ci >> sges_n) == rxq->rq_ci)))
1484 		return 0;
1485 	/* Update the consumer index. */
1486 	rxq->rq_ci = rq_ci >> sges_n;
1487 	rte_cio_wmb();
1488 	*rxq->cq_db = rte_cpu_to_be_32(rxq->cq_ci);
1489 	rte_cio_wmb();
1490 	*rxq->rq_db = rte_cpu_to_be_32(rxq->rq_ci);
1491 #ifdef MLX5_PMD_SOFT_COUNTERS
1492 	/* Increment packets counter. */
1493 	rxq->stats.ipackets += i;
1494 #endif
1495 	return i;
1496 }
1497 
1498 /**
1499  * Update LRO packet TCP header.
1500  * The HW LRO feature doesn't update the TCP header after coalescing the
1501  * TCP segments but supplies information in CQE to fill it by SW.
1502  *
1503  * @param tcp
1504  *   Pointer to the TCP header.
1505  * @param cqe
1506  *   Pointer to the completion entry..
1507  * @param phcsum
1508  *   The L3 pseudo-header checksum.
1509  */
1510 static inline void
1511 mlx5_lro_update_tcp_hdr(struct rte_tcp_hdr *restrict tcp,
1512 			volatile struct mlx5_cqe *restrict cqe,
1513 			uint32_t phcsum)
1514 {
1515 	uint8_t l4_type = (rte_be_to_cpu_16(cqe->hdr_type_etc) &
1516 			   MLX5_CQE_L4_TYPE_MASK) >> MLX5_CQE_L4_TYPE_SHIFT;
1517 	/*
1518 	 * The HW calculates only the TCP payload checksum, need to complete
1519 	 * the TCP header checksum and the L3 pseudo-header checksum.
1520 	 */
1521 	uint32_t csum = phcsum + cqe->csum;
1522 
1523 	if (l4_type == MLX5_L4_HDR_TYPE_TCP_EMPTY_ACK ||
1524 	    l4_type == MLX5_L4_HDR_TYPE_TCP_WITH_ACL) {
1525 		tcp->tcp_flags |= RTE_TCP_ACK_FLAG;
1526 		tcp->recv_ack = cqe->lro_ack_seq_num;
1527 		tcp->rx_win = cqe->lro_tcp_win;
1528 	}
1529 	if (cqe->lro_tcppsh_abort_dupack & MLX5_CQE_LRO_PUSH_MASK)
1530 		tcp->tcp_flags |= RTE_TCP_PSH_FLAG;
1531 	tcp->cksum = 0;
1532 	csum += rte_raw_cksum(tcp, (tcp->data_off & 0xF) * 4);
1533 	csum = ((csum & 0xffff0000) >> 16) + (csum & 0xffff);
1534 	csum = (~csum) & 0xffff;
1535 	if (csum == 0)
1536 		csum = 0xffff;
1537 	tcp->cksum = csum;
1538 }
1539 
1540 /**
1541  * Update LRO packet headers.
1542  * The HW LRO feature doesn't update the L3/TCP headers after coalescing the
1543  * TCP segments but supply information in CQE to fill it by SW.
1544  *
1545  * @param padd
1546  *   The packet address.
1547  * @param cqe
1548  *   Pointer to the completion entry..
1549  * @param len
1550  *   The packet length.
1551  */
1552 static inline void
1553 mlx5_lro_update_hdr(uint8_t *restrict padd,
1554 		    volatile struct mlx5_cqe *restrict cqe,
1555 		    uint32_t len)
1556 {
1557 	union {
1558 		struct rte_ether_hdr *eth;
1559 		struct rte_vlan_hdr *vlan;
1560 		struct rte_ipv4_hdr *ipv4;
1561 		struct rte_ipv6_hdr *ipv6;
1562 		struct rte_tcp_hdr *tcp;
1563 		uint8_t *hdr;
1564 	} h = {
1565 			.hdr = padd,
1566 	};
1567 	uint16_t proto = h.eth->ether_type;
1568 	uint32_t phcsum;
1569 
1570 	h.eth++;
1571 	while (proto == RTE_BE16(RTE_ETHER_TYPE_VLAN) ||
1572 	       proto == RTE_BE16(RTE_ETHER_TYPE_QINQ)) {
1573 		proto = h.vlan->eth_proto;
1574 		h.vlan++;
1575 	}
1576 	if (proto == RTE_BE16(RTE_ETHER_TYPE_IPV4)) {
1577 		h.ipv4->time_to_live = cqe->lro_min_ttl;
1578 		h.ipv4->total_length = rte_cpu_to_be_16(len - (h.hdr - padd));
1579 		h.ipv4->hdr_checksum = 0;
1580 		h.ipv4->hdr_checksum = rte_ipv4_cksum(h.ipv4);
1581 		phcsum = rte_ipv4_phdr_cksum(h.ipv4, 0);
1582 		h.ipv4++;
1583 	} else {
1584 		h.ipv6->hop_limits = cqe->lro_min_ttl;
1585 		h.ipv6->payload_len = rte_cpu_to_be_16(len - (h.hdr - padd) -
1586 						       sizeof(*h.ipv6));
1587 		phcsum = rte_ipv6_phdr_cksum(h.ipv6, 0);
1588 		h.ipv6++;
1589 	}
1590 	mlx5_lro_update_tcp_hdr(h.tcp, cqe, phcsum);
1591 }
1592 
1593 void
1594 mlx5_mprq_buf_free_cb(void *addr __rte_unused, void *opaque)
1595 {
1596 	struct mlx5_mprq_buf *buf = opaque;
1597 
1598 	if (rte_atomic16_read(&buf->refcnt) == 1) {
1599 		rte_mempool_put(buf->mp, buf);
1600 	} else if (rte_atomic16_add_return(&buf->refcnt, -1) == 0) {
1601 		rte_atomic16_set(&buf->refcnt, 1);
1602 		rte_mempool_put(buf->mp, buf);
1603 	}
1604 }
1605 
1606 void
1607 mlx5_mprq_buf_free(struct mlx5_mprq_buf *buf)
1608 {
1609 	mlx5_mprq_buf_free_cb(NULL, buf);
1610 }
1611 
1612 static inline void
1613 mprq_buf_replace(struct mlx5_rxq_data *rxq, uint16_t rq_idx,
1614 		 const unsigned int strd_n)
1615 {
1616 	struct mlx5_mprq_buf *rep = rxq->mprq_repl;
1617 	volatile struct mlx5_wqe_data_seg *wqe =
1618 		&((volatile struct mlx5_wqe_mprq *)rxq->wqes)[rq_idx].dseg;
1619 	void *addr;
1620 
1621 	MLX5_ASSERT(rep != NULL);
1622 	/* Replace MPRQ buf. */
1623 	(*rxq->mprq_bufs)[rq_idx] = rep;
1624 	/* Replace WQE. */
1625 	addr = mlx5_mprq_buf_addr(rep, strd_n);
1626 	wqe->addr = rte_cpu_to_be_64((uintptr_t)addr);
1627 	/* If there's only one MR, no need to replace LKey in WQE. */
1628 	if (unlikely(mlx5_mr_btree_len(&rxq->mr_ctrl.cache_bh) > 1))
1629 		wqe->lkey = mlx5_rx_addr2mr(rxq, (uintptr_t)addr);
1630 	/* Stash a mbuf for next replacement. */
1631 	if (likely(!rte_mempool_get(rxq->mprq_mp, (void **)&rep)))
1632 		rxq->mprq_repl = rep;
1633 	else
1634 		rxq->mprq_repl = NULL;
1635 }
1636 
1637 /**
1638  * DPDK callback for RX with Multi-Packet RQ support.
1639  *
1640  * @param dpdk_rxq
1641  *   Generic pointer to RX queue structure.
1642  * @param[out] pkts
1643  *   Array to store received packets.
1644  * @param pkts_n
1645  *   Maximum number of packets in array.
1646  *
1647  * @return
1648  *   Number of packets successfully received (<= pkts_n).
1649  */
1650 uint16_t
1651 mlx5_rx_burst_mprq(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n)
1652 {
1653 	struct mlx5_rxq_data *rxq = dpdk_rxq;
1654 	const unsigned int strd_n = 1 << rxq->strd_num_n;
1655 	const unsigned int strd_sz = 1 << rxq->strd_sz_n;
1656 	const unsigned int strd_shift =
1657 		MLX5_MPRQ_STRIDE_SHIFT_BYTE * rxq->strd_shift_en;
1658 	const unsigned int cq_mask = (1 << rxq->cqe_n) - 1;
1659 	const unsigned int wq_mask = (1 << rxq->elts_n) - 1;
1660 	volatile struct mlx5_cqe *cqe = &(*rxq->cqes)[rxq->cq_ci & cq_mask];
1661 	unsigned int i = 0;
1662 	uint32_t rq_ci = rxq->rq_ci;
1663 	uint16_t consumed_strd = rxq->consumed_strd;
1664 	struct mlx5_mprq_buf *buf = (*rxq->mprq_bufs)[rq_ci & wq_mask];
1665 
1666 	while (i < pkts_n) {
1667 		struct rte_mbuf *pkt;
1668 		void *addr;
1669 		int ret;
1670 		uint32_t len;
1671 		uint16_t strd_cnt;
1672 		uint16_t strd_idx;
1673 		uint32_t offset;
1674 		uint32_t byte_cnt;
1675 		int32_t hdrm_overlap;
1676 		volatile struct mlx5_mini_cqe8 *mcqe = NULL;
1677 		uint32_t rss_hash_res = 0;
1678 
1679 		if (consumed_strd == strd_n) {
1680 			/* Replace WQE only if the buffer is still in use. */
1681 			if (rte_atomic16_read(&buf->refcnt) > 1) {
1682 				mprq_buf_replace(rxq, rq_ci & wq_mask, strd_n);
1683 				/* Release the old buffer. */
1684 				mlx5_mprq_buf_free(buf);
1685 			} else if (unlikely(rxq->mprq_repl == NULL)) {
1686 				struct mlx5_mprq_buf *rep;
1687 
1688 				/*
1689 				 * Currently, the MPRQ mempool is out of buffer
1690 				 * and doing memcpy regardless of the size of Rx
1691 				 * packet. Retry allocation to get back to
1692 				 * normal.
1693 				 */
1694 				if (!rte_mempool_get(rxq->mprq_mp,
1695 						     (void **)&rep))
1696 					rxq->mprq_repl = rep;
1697 			}
1698 			/* Advance to the next WQE. */
1699 			consumed_strd = 0;
1700 			++rq_ci;
1701 			buf = (*rxq->mprq_bufs)[rq_ci & wq_mask];
1702 		}
1703 		cqe = &(*rxq->cqes)[rxq->cq_ci & cq_mask];
1704 		ret = mlx5_rx_poll_len(rxq, cqe, cq_mask, &mcqe);
1705 		if (!ret)
1706 			break;
1707 		byte_cnt = ret;
1708 		strd_cnt = (byte_cnt & MLX5_MPRQ_STRIDE_NUM_MASK) >>
1709 			   MLX5_MPRQ_STRIDE_NUM_SHIFT;
1710 		MLX5_ASSERT(strd_cnt);
1711 		consumed_strd += strd_cnt;
1712 		if (byte_cnt & MLX5_MPRQ_FILLER_MASK)
1713 			continue;
1714 		if (mcqe == NULL) {
1715 			rss_hash_res = rte_be_to_cpu_32(cqe->rx_hash_res);
1716 			strd_idx = rte_be_to_cpu_16(cqe->wqe_counter);
1717 		} else {
1718 			/* mini-CQE for MPRQ doesn't have hash result. */
1719 			strd_idx = rte_be_to_cpu_16(mcqe->stride_idx);
1720 		}
1721 		MLX5_ASSERT(strd_idx < strd_n);
1722 		MLX5_ASSERT(!((rte_be_to_cpu_16(cqe->wqe_id) ^ rq_ci) &
1723 			    wq_mask));
1724 		pkt = rte_pktmbuf_alloc(rxq->mp);
1725 		if (unlikely(pkt == NULL)) {
1726 			++rxq->stats.rx_nombuf;
1727 			break;
1728 		}
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 		offset = strd_idx * strd_sz + strd_shift;
1734 		addr = RTE_PTR_ADD(mlx5_mprq_buf_addr(buf, strd_n), offset);
1735 		hdrm_overlap = len + RTE_PKTMBUF_HEADROOM - strd_cnt * strd_sz;
1736 		/*
1737 		 * Memcpy packets to the target mbuf if:
1738 		 * - The size of packet is smaller than mprq_max_memcpy_len.
1739 		 * - Out of buffer in the Mempool for Multi-Packet RQ.
1740 		 * - The packet's stride overlaps a headroom and scatter is off.
1741 		 */
1742 		if (len <= rxq->mprq_max_memcpy_len ||
1743 		    rxq->mprq_repl == NULL ||
1744 		    (hdrm_overlap > 0 && !rxq->strd_scatter_en)) {
1745 			if (likely(rte_pktmbuf_tailroom(pkt) >= len)) {
1746 				rte_memcpy(rte_pktmbuf_mtod(pkt, void *),
1747 					   addr, len);
1748 				DATA_LEN(pkt) = len;
1749 			} else if (rxq->strd_scatter_en) {
1750 				struct rte_mbuf *prev = pkt;
1751 				uint32_t seg_len =
1752 					RTE_MIN(rte_pktmbuf_tailroom(pkt), len);
1753 				uint32_t rem_len = len - seg_len;
1754 
1755 				rte_memcpy(rte_pktmbuf_mtod(pkt, void *),
1756 					   addr, seg_len);
1757 				DATA_LEN(pkt) = seg_len;
1758 				while (rem_len) {
1759 					struct rte_mbuf *next =
1760 						rte_pktmbuf_alloc(rxq->mp);
1761 
1762 					if (unlikely(next == NULL)) {
1763 						rte_pktmbuf_free(pkt);
1764 						++rxq->stats.rx_nombuf;
1765 						goto out;
1766 					}
1767 					NEXT(prev) = next;
1768 					SET_DATA_OFF(next, 0);
1769 					addr = RTE_PTR_ADD(addr, seg_len);
1770 					seg_len = RTE_MIN
1771 						(rte_pktmbuf_tailroom(next),
1772 						 rem_len);
1773 					rte_memcpy
1774 						(rte_pktmbuf_mtod(next, void *),
1775 						 addr, seg_len);
1776 					DATA_LEN(next) = seg_len;
1777 					rem_len -= seg_len;
1778 					prev = next;
1779 					++NB_SEGS(pkt);
1780 				}
1781 			} else {
1782 				rte_pktmbuf_free_seg(pkt);
1783 				++rxq->stats.idropped;
1784 				continue;
1785 			}
1786 		} else {
1787 			rte_iova_t buf_iova;
1788 			struct rte_mbuf_ext_shared_info *shinfo;
1789 			uint16_t buf_len = strd_cnt * strd_sz;
1790 			void *buf_addr;
1791 
1792 			/* Increment the refcnt of the whole chunk. */
1793 			rte_atomic16_add_return(&buf->refcnt, 1);
1794 			MLX5_ASSERT((uint16_t)rte_atomic16_read(&buf->refcnt) <=
1795 				    strd_n + 1);
1796 			buf_addr = RTE_PTR_SUB(addr, RTE_PKTMBUF_HEADROOM);
1797 			/*
1798 			 * MLX5 device doesn't use iova but it is necessary in a
1799 			 * case where the Rx packet is transmitted via a
1800 			 * different PMD.
1801 			 */
1802 			buf_iova = rte_mempool_virt2iova(buf) +
1803 				   RTE_PTR_DIFF(buf_addr, buf);
1804 			shinfo = &buf->shinfos[strd_idx];
1805 			rte_mbuf_ext_refcnt_set(shinfo, 1);
1806 			/*
1807 			 * EXT_ATTACHED_MBUF will be set to pkt->ol_flags when
1808 			 * attaching the stride to mbuf and more offload flags
1809 			 * will be added below by calling rxq_cq_to_mbuf().
1810 			 * Other fields will be overwritten.
1811 			 */
1812 			rte_pktmbuf_attach_extbuf(pkt, buf_addr, buf_iova,
1813 						  buf_len, shinfo);
1814 			/* Set mbuf head-room. */
1815 			SET_DATA_OFF(pkt, RTE_PKTMBUF_HEADROOM);
1816 			MLX5_ASSERT(pkt->ol_flags == EXT_ATTACHED_MBUF);
1817 			MLX5_ASSERT(rte_pktmbuf_tailroom(pkt) >=
1818 				len - (hdrm_overlap > 0 ? hdrm_overlap : 0));
1819 			DATA_LEN(pkt) = len;
1820 			/*
1821 			 * Copy the last fragment of a packet (up to headroom
1822 			 * size bytes) in case there is a stride overlap with
1823 			 * a next packet's headroom. Allocate a separate mbuf
1824 			 * to store this fragment and link it. Scatter is on.
1825 			 */
1826 			if (hdrm_overlap > 0) {
1827 				MLX5_ASSERT(rxq->strd_scatter_en);
1828 				struct rte_mbuf *seg =
1829 					rte_pktmbuf_alloc(rxq->mp);
1830 
1831 				if (unlikely(seg == NULL)) {
1832 					rte_pktmbuf_free_seg(pkt);
1833 					++rxq->stats.rx_nombuf;
1834 					break;
1835 				}
1836 				SET_DATA_OFF(seg, 0);
1837 				rte_memcpy(rte_pktmbuf_mtod(seg, void *),
1838 					RTE_PTR_ADD(addr, len - hdrm_overlap),
1839 					hdrm_overlap);
1840 				DATA_LEN(seg) = hdrm_overlap;
1841 				DATA_LEN(pkt) = len - hdrm_overlap;
1842 				NEXT(pkt) = seg;
1843 				NB_SEGS(pkt) = 2;
1844 			}
1845 		}
1846 		rxq_cq_to_mbuf(rxq, pkt, cqe, rss_hash_res);
1847 		if (cqe->lro_num_seg > 1) {
1848 			mlx5_lro_update_hdr(addr, cqe, len);
1849 			pkt->ol_flags |= PKT_RX_LRO;
1850 			pkt->tso_segsz = len / cqe->lro_num_seg;
1851 		}
1852 		PKT_LEN(pkt) = len;
1853 		PORT(pkt) = rxq->port_id;
1854 #ifdef MLX5_PMD_SOFT_COUNTERS
1855 		/* Increment bytes counter. */
1856 		rxq->stats.ibytes += PKT_LEN(pkt);
1857 #endif
1858 		/* Return packet. */
1859 		*(pkts++) = pkt;
1860 		++i;
1861 	}
1862 out:
1863 	/* Update the consumer indexes. */
1864 	rxq->consumed_strd = consumed_strd;
1865 	rte_cio_wmb();
1866 	*rxq->cq_db = rte_cpu_to_be_32(rxq->cq_ci);
1867 	if (rq_ci != rxq->rq_ci) {
1868 		rxq->rq_ci = rq_ci;
1869 		rte_cio_wmb();
1870 		*rxq->rq_db = rte_cpu_to_be_32(rxq->rq_ci);
1871 	}
1872 #ifdef MLX5_PMD_SOFT_COUNTERS
1873 	/* Increment packets counter. */
1874 	rxq->stats.ipackets += i;
1875 #endif
1876 	return i;
1877 }
1878 
1879 /**
1880  * Dummy DPDK callback for TX.
1881  *
1882  * This function is used to temporarily replace the real callback during
1883  * unsafe control operations on the queue, or in case of error.
1884  *
1885  * @param dpdk_txq
1886  *   Generic pointer to TX queue structure.
1887  * @param[in] pkts
1888  *   Packets to transmit.
1889  * @param pkts_n
1890  *   Number of packets in array.
1891  *
1892  * @return
1893  *   Number of packets successfully transmitted (<= pkts_n).
1894  */
1895 uint16_t
1896 removed_tx_burst(void *dpdk_txq __rte_unused,
1897 		 struct rte_mbuf **pkts __rte_unused,
1898 		 uint16_t pkts_n __rte_unused)
1899 {
1900 	rte_mb();
1901 	return 0;
1902 }
1903 
1904 /**
1905  * Dummy DPDK callback for RX.
1906  *
1907  * This function is used to temporarily replace the real callback during
1908  * unsafe control operations on the queue, or in case of error.
1909  *
1910  * @param dpdk_rxq
1911  *   Generic pointer to RX queue structure.
1912  * @param[out] pkts
1913  *   Array to store received packets.
1914  * @param pkts_n
1915  *   Maximum number of packets in array.
1916  *
1917  * @return
1918  *   Number of packets successfully received (<= pkts_n).
1919  */
1920 uint16_t
1921 removed_rx_burst(void *dpdk_txq __rte_unused,
1922 		 struct rte_mbuf **pkts __rte_unused,
1923 		 uint16_t pkts_n __rte_unused)
1924 {
1925 	rte_mb();
1926 	return 0;
1927 }
1928 
1929 /*
1930  * Vectorized Rx/Tx routines are not compiled in when required vector
1931  * instructions are not supported on a target architecture. The following null
1932  * stubs are needed for linkage when those are not included outside of this file
1933  * (e.g.  mlx5_rxtx_vec_sse.c for x86).
1934  */
1935 
1936 __rte_weak uint16_t
1937 mlx5_rx_burst_vec(void *dpdk_txq __rte_unused,
1938 		  struct rte_mbuf **pkts __rte_unused,
1939 		  uint16_t pkts_n __rte_unused)
1940 {
1941 	return 0;
1942 }
1943 
1944 __rte_weak int
1945 mlx5_rxq_check_vec_support(struct mlx5_rxq_data *rxq __rte_unused)
1946 {
1947 	return -ENOTSUP;
1948 }
1949 
1950 __rte_weak int
1951 mlx5_check_vec_rx_support(struct rte_eth_dev *dev __rte_unused)
1952 {
1953 	return -ENOTSUP;
1954 }
1955 
1956 /**
1957  * Free the mbufs from the linear array of pointers.
1958  *
1959  * @param pkts
1960  *   Pointer to array of packets to be free.
1961  * @param pkts_n
1962  *   Number of packets to be freed.
1963  * @param olx
1964  *   Configured Tx offloads mask. It is fully defined at
1965  *   compile time and may be used for optimization.
1966  */
1967 static __rte_always_inline void
1968 mlx5_tx_free_mbuf(struct rte_mbuf **restrict pkts,
1969 		  unsigned int pkts_n,
1970 		  unsigned int olx __rte_unused)
1971 {
1972 	struct rte_mempool *pool = NULL;
1973 	struct rte_mbuf **p_free = NULL;
1974 	struct rte_mbuf *mbuf;
1975 	unsigned int n_free = 0;
1976 
1977 	/*
1978 	 * The implemented algorithm eliminates
1979 	 * copying pointers to temporary array
1980 	 * for rte_mempool_put_bulk() calls.
1981 	 */
1982 	MLX5_ASSERT(pkts);
1983 	MLX5_ASSERT(pkts_n);
1984 	for (;;) {
1985 		for (;;) {
1986 			/*
1987 			 * Decrement mbuf reference counter, detach
1988 			 * indirect and external buffers if needed.
1989 			 */
1990 			mbuf = rte_pktmbuf_prefree_seg(*pkts);
1991 			if (likely(mbuf != NULL)) {
1992 				MLX5_ASSERT(mbuf == *pkts);
1993 				if (likely(n_free != 0)) {
1994 					if (unlikely(pool != mbuf->pool))
1995 						/* From different pool. */
1996 						break;
1997 				} else {
1998 					/* Start new scan array. */
1999 					pool = mbuf->pool;
2000 					p_free = pkts;
2001 				}
2002 				++n_free;
2003 				++pkts;
2004 				--pkts_n;
2005 				if (unlikely(pkts_n == 0)) {
2006 					mbuf = NULL;
2007 					break;
2008 				}
2009 			} else {
2010 				/*
2011 				 * This happens if mbuf is still referenced.
2012 				 * We can't put it back to the pool, skip.
2013 				 */
2014 				++pkts;
2015 				--pkts_n;
2016 				if (unlikely(n_free != 0))
2017 					/* There is some array to free.*/
2018 					break;
2019 				if (unlikely(pkts_n == 0))
2020 					/* Last mbuf, nothing to free. */
2021 					return;
2022 			}
2023 		}
2024 		for (;;) {
2025 			/*
2026 			 * This loop is implemented to avoid multiple
2027 			 * inlining of rte_mempool_put_bulk().
2028 			 */
2029 			MLX5_ASSERT(pool);
2030 			MLX5_ASSERT(p_free);
2031 			MLX5_ASSERT(n_free);
2032 			/*
2033 			 * Free the array of pre-freed mbufs
2034 			 * belonging to the same memory pool.
2035 			 */
2036 			rte_mempool_put_bulk(pool, (void *)p_free, n_free);
2037 			if (unlikely(mbuf != NULL)) {
2038 				/* There is the request to start new scan. */
2039 				pool = mbuf->pool;
2040 				p_free = pkts++;
2041 				n_free = 1;
2042 				--pkts_n;
2043 				if (likely(pkts_n != 0))
2044 					break;
2045 				/*
2046 				 * This is the last mbuf to be freed.
2047 				 * Do one more loop iteration to complete.
2048 				 * This is rare case of the last unique mbuf.
2049 				 */
2050 				mbuf = NULL;
2051 				continue;
2052 			}
2053 			if (likely(pkts_n == 0))
2054 				return;
2055 			n_free = 0;
2056 			break;
2057 		}
2058 	}
2059 }
2060 
2061 /**
2062  * Free the mbuf from the elts ring buffer till new tail.
2063  *
2064  * @param txq
2065  *   Pointer to Tx queue structure.
2066  * @param tail
2067  *   Index in elts to free up to, becomes new elts tail.
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_free_elts(struct mlx5_txq_data *restrict txq,
2074 		  uint16_t tail,
2075 		  unsigned int olx __rte_unused)
2076 {
2077 	uint16_t n_elts = tail - txq->elts_tail;
2078 
2079 	MLX5_ASSERT(n_elts);
2080 	MLX5_ASSERT(n_elts <= txq->elts_s);
2081 	/*
2082 	 * Implement a loop to support ring buffer wraparound
2083 	 * with single inlining of mlx5_tx_free_mbuf().
2084 	 */
2085 	do {
2086 		unsigned int part;
2087 
2088 		part = txq->elts_s - (txq->elts_tail & txq->elts_m);
2089 		part = RTE_MIN(part, n_elts);
2090 		MLX5_ASSERT(part);
2091 		MLX5_ASSERT(part <= txq->elts_s);
2092 		mlx5_tx_free_mbuf(&txq->elts[txq->elts_tail & txq->elts_m],
2093 				  part, olx);
2094 		txq->elts_tail += part;
2095 		n_elts -= part;
2096 	} while (n_elts);
2097 }
2098 
2099 /**
2100  * Store the mbuf being sent into elts ring buffer.
2101  * On Tx completion these mbufs will be freed.
2102  *
2103  * @param txq
2104  *   Pointer to Tx queue structure.
2105  * @param pkts
2106  *   Pointer to array of packets to be stored.
2107  * @param pkts_n
2108  *   Number of packets to be stored.
2109  * @param olx
2110  *   Configured Tx offloads mask. It is fully defined at
2111  *   compile time and may be used for optimization.
2112  */
2113 static __rte_always_inline void
2114 mlx5_tx_copy_elts(struct mlx5_txq_data *restrict txq,
2115 		  struct rte_mbuf **restrict pkts,
2116 		  unsigned int pkts_n,
2117 		  unsigned int olx __rte_unused)
2118 {
2119 	unsigned int part;
2120 	struct rte_mbuf **elts = (struct rte_mbuf **)txq->elts;
2121 
2122 	MLX5_ASSERT(pkts);
2123 	MLX5_ASSERT(pkts_n);
2124 	part = txq->elts_s - (txq->elts_head & txq->elts_m);
2125 	MLX5_ASSERT(part);
2126 	MLX5_ASSERT(part <= txq->elts_s);
2127 	/* This code is a good candidate for vectorizing with SIMD. */
2128 	rte_memcpy((void *)(elts + (txq->elts_head & txq->elts_m)),
2129 		   (void *)pkts,
2130 		   RTE_MIN(part, pkts_n) * sizeof(struct rte_mbuf *));
2131 	txq->elts_head += pkts_n;
2132 	if (unlikely(part < pkts_n))
2133 		/* The copy is wrapping around the elts array. */
2134 		rte_memcpy((void *)elts, (void *)(pkts + part),
2135 			   (pkts_n - part) * sizeof(struct rte_mbuf *));
2136 }
2137 
2138 /**
2139  * Update completion queue consuming index via doorbell
2140  * and flush the completed data buffers.
2141  *
2142  * @param txq
2143  *   Pointer to TX queue structure.
2144  * @param valid CQE pointer
2145  *   if not NULL update txq->wqe_pi and flush the buffers
2146  * @param olx
2147  *   Configured Tx offloads mask. It is fully defined at
2148  *   compile time and may be used for optimization.
2149  */
2150 static __rte_always_inline void
2151 mlx5_tx_comp_flush(struct mlx5_txq_data *restrict txq,
2152 		   volatile struct mlx5_cqe *last_cqe,
2153 		   unsigned int olx __rte_unused)
2154 {
2155 	if (likely(last_cqe != NULL)) {
2156 		uint16_t tail;
2157 
2158 		txq->wqe_pi = rte_be_to_cpu_16(last_cqe->wqe_counter);
2159 		tail = txq->fcqs[(txq->cq_ci - 1) & txq->cqe_m];
2160 		if (likely(tail != txq->elts_tail)) {
2161 			mlx5_tx_free_elts(txq, tail, olx);
2162 			MLX5_ASSERT(tail == txq->elts_tail);
2163 		}
2164 	}
2165 }
2166 
2167 /**
2168  * Manage TX completions. This routine checks the CQ for
2169  * arrived CQEs, deduces the last accomplished WQE in SQ,
2170  * updates SQ producing index and frees all completed mbufs.
2171  *
2172  * @param txq
2173  *   Pointer to TX queue structure.
2174  * @param olx
2175  *   Configured Tx offloads mask. It is fully defined at
2176  *   compile time and may be used for optimization.
2177  *
2178  * NOTE: not inlined intentionally, it makes tx_burst
2179  * routine smaller, simple and faster - from experiments.
2180  */
2181 static void
2182 mlx5_tx_handle_completion(struct mlx5_txq_data *restrict txq,
2183 			  unsigned int olx __rte_unused)
2184 {
2185 	unsigned int count = MLX5_TX_COMP_MAX_CQE;
2186 	volatile struct mlx5_cqe *last_cqe = NULL;
2187 	bool ring_doorbell = false;
2188 	int ret;
2189 
2190 	static_assert(MLX5_CQE_STATUS_HW_OWN < 0, "Must be negative value");
2191 	static_assert(MLX5_CQE_STATUS_SW_OWN < 0, "Must be negative value");
2192 	do {
2193 		volatile struct mlx5_cqe *cqe;
2194 
2195 		cqe = &txq->cqes[txq->cq_ci & txq->cqe_m];
2196 		ret = check_cqe(cqe, txq->cqe_s, txq->cq_ci);
2197 		if (unlikely(ret != MLX5_CQE_STATUS_SW_OWN)) {
2198 			if (likely(ret != MLX5_CQE_STATUS_ERR)) {
2199 				/* No new CQEs in completion queue. */
2200 				MLX5_ASSERT(ret == MLX5_CQE_STATUS_HW_OWN);
2201 				break;
2202 			}
2203 			/*
2204 			 * Some error occurred, try to restart.
2205 			 * We have no barrier after WQE related Doorbell
2206 			 * written, make sure all writes are completed
2207 			 * here, before we might perform SQ reset.
2208 			 */
2209 			rte_wmb();
2210 			ret = mlx5_tx_error_cqe_handle
2211 				(txq, (volatile struct mlx5_err_cqe *)cqe);
2212 			if (unlikely(ret < 0)) {
2213 				/*
2214 				 * Some error occurred on queue error
2215 				 * handling, we do not advance the index
2216 				 * here, allowing to retry on next call.
2217 				 */
2218 				return;
2219 			}
2220 			/*
2221 			 * We are going to fetch all entries with
2222 			 * MLX5_CQE_SYNDROME_WR_FLUSH_ERR status.
2223 			 * The send queue is supposed to be empty.
2224 			 */
2225 			ring_doorbell = true;
2226 			++txq->cq_ci;
2227 			txq->cq_pi = txq->cq_ci;
2228 			last_cqe = NULL;
2229 			continue;
2230 		}
2231 		/* Normal transmit completion. */
2232 		MLX5_ASSERT(txq->cq_ci != txq->cq_pi);
2233 		MLX5_ASSERT((txq->fcqs[txq->cq_ci & txq->cqe_m] >> 16) ==
2234 			    cqe->wqe_counter);
2235 		ring_doorbell = true;
2236 		++txq->cq_ci;
2237 		last_cqe = cqe;
2238 		/*
2239 		 * We have to restrict the amount of processed CQEs
2240 		 * in one tx_burst routine call. The CQ may be large
2241 		 * and many CQEs may be updated by the NIC in one
2242 		 * transaction. Buffers freeing is time consuming,
2243 		 * multiple iterations may introduce significant
2244 		 * latency.
2245 		 */
2246 		if (likely(--count == 0))
2247 			break;
2248 	} while (true);
2249 	if (likely(ring_doorbell)) {
2250 		/* Ring doorbell to notify hardware. */
2251 		rte_compiler_barrier();
2252 		*txq->cq_db = rte_cpu_to_be_32(txq->cq_ci);
2253 		mlx5_tx_comp_flush(txq, last_cqe, olx);
2254 	}
2255 }
2256 
2257 /**
2258  * Check if the completion request flag should be set in the last WQE.
2259  * Both pushed mbufs and WQEs are monitored and the completion request
2260  * flag is set if any of thresholds is reached.
2261  *
2262  * @param txq
2263  *   Pointer to TX queue structure.
2264  * @param loc
2265  *   Pointer to burst routine local context.
2266  * @param olx
2267  *   Configured Tx offloads mask. It is fully defined at
2268  *   compile time and may be used for optimization.
2269  */
2270 static __rte_always_inline void
2271 mlx5_tx_request_completion(struct mlx5_txq_data *restrict txq,
2272 			   struct mlx5_txq_local *restrict loc,
2273 			   unsigned int olx)
2274 {
2275 	uint16_t head = txq->elts_head;
2276 	unsigned int part;
2277 
2278 	part = MLX5_TXOFF_CONFIG(INLINE) ?
2279 	       0 : loc->pkts_sent - loc->pkts_copy;
2280 	head += part;
2281 	if ((uint16_t)(head - txq->elts_comp) >= MLX5_TX_COMP_THRESH ||
2282 	     (MLX5_TXOFF_CONFIG(INLINE) &&
2283 	     (uint16_t)(txq->wqe_ci - txq->wqe_comp) >= txq->wqe_thres)) {
2284 		volatile struct mlx5_wqe *last = loc->wqe_last;
2285 
2286 		MLX5_ASSERT(last);
2287 		txq->elts_comp = head;
2288 		if (MLX5_TXOFF_CONFIG(INLINE))
2289 			txq->wqe_comp = txq->wqe_ci;
2290 		/* Request unconditional completion on last WQE. */
2291 		last->cseg.flags = RTE_BE32(MLX5_COMP_ALWAYS <<
2292 					    MLX5_COMP_MODE_OFFSET);
2293 		/* Save elts_head in dedicated free on completion queue. */
2294 #ifdef RTE_LIBRTE_MLX5_DEBUG
2295 		txq->fcqs[txq->cq_pi++ & txq->cqe_m] = head |
2296 			  (last->cseg.opcode >> 8) << 16;
2297 #else
2298 		txq->fcqs[txq->cq_pi++ & txq->cqe_m] = head;
2299 #endif
2300 		/* A CQE slot must always be available. */
2301 		MLX5_ASSERT((txq->cq_pi - txq->cq_ci) <= txq->cqe_s);
2302 	}
2303 }
2304 
2305 /**
2306  * DPDK callback to check the status of a tx descriptor.
2307  *
2308  * @param tx_queue
2309  *   The tx queue.
2310  * @param[in] offset
2311  *   The index of the descriptor in the ring.
2312  *
2313  * @return
2314  *   The status of the tx descriptor.
2315  */
2316 int
2317 mlx5_tx_descriptor_status(void *tx_queue, uint16_t offset)
2318 {
2319 	struct mlx5_txq_data *restrict txq = tx_queue;
2320 	uint16_t used;
2321 
2322 	mlx5_tx_handle_completion(txq, 0);
2323 	used = txq->elts_head - txq->elts_tail;
2324 	if (offset < used)
2325 		return RTE_ETH_TX_DESC_FULL;
2326 	return RTE_ETH_TX_DESC_DONE;
2327 }
2328 
2329 /**
2330  * Build the Control Segment with specified opcode:
2331  * - MLX5_OPCODE_SEND
2332  * - MLX5_OPCODE_ENHANCED_MPSW
2333  * - MLX5_OPCODE_TSO
2334  *
2335  * @param txq
2336  *   Pointer to TX queue structure.
2337  * @param loc
2338  *   Pointer to burst routine local context.
2339  * @param wqe
2340  *   Pointer to WQE to fill with built Control Segment.
2341  * @param ds
2342  *   Supposed length of WQE in segments.
2343  * @param opcode
2344  *   SQ WQE opcode to put into Control Segment.
2345  * @param olx
2346  *   Configured Tx offloads mask. It is fully defined at
2347  *   compile time and may be used for optimization.
2348  */
2349 static __rte_always_inline void
2350 mlx5_tx_cseg_init(struct mlx5_txq_data *restrict txq,
2351 		  struct mlx5_txq_local *restrict loc __rte_unused,
2352 		  struct mlx5_wqe *restrict wqe,
2353 		  unsigned int ds,
2354 		  unsigned int opcode,
2355 		  unsigned int olx __rte_unused)
2356 {
2357 	struct mlx5_wqe_cseg *restrict cs = &wqe->cseg;
2358 
2359 	/* For legacy MPW replace the EMPW by TSO with modifier. */
2360 	if (MLX5_TXOFF_CONFIG(MPW) && opcode == MLX5_OPCODE_ENHANCED_MPSW)
2361 		opcode = MLX5_OPCODE_TSO | MLX5_OPC_MOD_MPW << 24;
2362 	cs->opcode = rte_cpu_to_be_32((txq->wqe_ci << 8) | opcode);
2363 	cs->sq_ds = rte_cpu_to_be_32(txq->qp_num_8s | ds);
2364 	cs->flags = RTE_BE32(MLX5_COMP_ONLY_FIRST_ERR <<
2365 			     MLX5_COMP_MODE_OFFSET);
2366 	cs->misc = RTE_BE32(0);
2367 }
2368 
2369 /**
2370  * Build the Ethernet Segment without inlined data.
2371  * Supports Software Parser, Checksums and VLAN
2372  * insertion Tx offload features.
2373  *
2374  * @param txq
2375  *   Pointer to TX queue structure.
2376  * @param loc
2377  *   Pointer to burst routine local context.
2378  * @param wqe
2379  *   Pointer to WQE to fill with built Ethernet Segment.
2380  * @param olx
2381  *   Configured Tx offloads mask. It is fully defined at
2382  *   compile time and may be used for optimization.
2383  */
2384 static __rte_always_inline void
2385 mlx5_tx_eseg_none(struct mlx5_txq_data *restrict txq __rte_unused,
2386 		  struct mlx5_txq_local *restrict loc,
2387 		  struct mlx5_wqe *restrict wqe,
2388 		  unsigned int olx)
2389 {
2390 	struct mlx5_wqe_eseg *restrict es = &wqe->eseg;
2391 	uint32_t csum;
2392 
2393 	/*
2394 	 * Calculate and set check sum flags first, dword field
2395 	 * in segment may be shared with Software Parser flags.
2396 	 */
2397 	csum = MLX5_TXOFF_CONFIG(CSUM) ? txq_ol_cksum_to_cs(loc->mbuf) : 0;
2398 	es->flags = rte_cpu_to_le_32(csum);
2399 	/*
2400 	 * Calculate and set Software Parser offsets and flags.
2401 	 * These flags a set for custom UDP and IP tunnel packets.
2402 	 */
2403 	es->swp_offs = txq_mbuf_to_swp(loc, &es->swp_flags, olx);
2404 	/* Fill metadata field if needed. */
2405 	es->metadata = MLX5_TXOFF_CONFIG(METADATA) ?
2406 		       loc->mbuf->ol_flags & PKT_TX_DYNF_METADATA ?
2407 		       *RTE_FLOW_DYNF_METADATA(loc->mbuf) : 0 : 0;
2408 	/* Engage VLAN tag insertion feature if requested. */
2409 	if (MLX5_TXOFF_CONFIG(VLAN) &&
2410 	    loc->mbuf->ol_flags & PKT_TX_VLAN_PKT) {
2411 		/*
2412 		 * We should get here only if device support
2413 		 * this feature correctly.
2414 		 */
2415 		MLX5_ASSERT(txq->vlan_en);
2416 		es->inline_hdr = rte_cpu_to_be_32(MLX5_ETH_WQE_VLAN_INSERT |
2417 						  loc->mbuf->vlan_tci);
2418 	} else {
2419 		es->inline_hdr = RTE_BE32(0);
2420 	}
2421 }
2422 
2423 /**
2424  * Build the Ethernet Segment with minimal inlined data
2425  * of MLX5_ESEG_MIN_INLINE_SIZE bytes length. This is
2426  * used to fill the gap in single WQEBB WQEs.
2427  * Supports Software Parser, Checksums and VLAN
2428  * insertion Tx offload features.
2429  *
2430  * @param txq
2431  *   Pointer to TX queue structure.
2432  * @param loc
2433  *   Pointer to burst routine local context.
2434  * @param wqe
2435  *   Pointer to WQE to fill with built Ethernet Segment.
2436  * @param vlan
2437  *   Length of VLAN tag insertion if any.
2438  * @param olx
2439  *   Configured Tx offloads mask. It is fully defined at
2440  *   compile time and may be used for optimization.
2441  */
2442 static __rte_always_inline void
2443 mlx5_tx_eseg_dmin(struct mlx5_txq_data *restrict txq __rte_unused,
2444 		  struct mlx5_txq_local *restrict loc,
2445 		  struct mlx5_wqe *restrict wqe,
2446 		  unsigned int vlan,
2447 		  unsigned int olx)
2448 {
2449 	struct mlx5_wqe_eseg *restrict es = &wqe->eseg;
2450 	uint32_t csum;
2451 	uint8_t *psrc, *pdst;
2452 
2453 	/*
2454 	 * Calculate and set check sum flags first, dword field
2455 	 * in segment may be shared with Software Parser flags.
2456 	 */
2457 	csum = MLX5_TXOFF_CONFIG(CSUM) ? txq_ol_cksum_to_cs(loc->mbuf) : 0;
2458 	es->flags = rte_cpu_to_le_32(csum);
2459 	/*
2460 	 * Calculate and set Software Parser offsets and flags.
2461 	 * These flags a set for custom UDP and IP tunnel packets.
2462 	 */
2463 	es->swp_offs = txq_mbuf_to_swp(loc, &es->swp_flags, olx);
2464 	/* Fill metadata field if needed. */
2465 	es->metadata = MLX5_TXOFF_CONFIG(METADATA) ?
2466 		       loc->mbuf->ol_flags & PKT_TX_DYNF_METADATA ?
2467 		       *RTE_FLOW_DYNF_METADATA(loc->mbuf) : 0 : 0;
2468 	static_assert(MLX5_ESEG_MIN_INLINE_SIZE ==
2469 				(sizeof(uint16_t) +
2470 				 sizeof(rte_v128u32_t)),
2471 		      "invalid Ethernet Segment data size");
2472 	static_assert(MLX5_ESEG_MIN_INLINE_SIZE ==
2473 				(sizeof(uint16_t) +
2474 				 sizeof(struct rte_vlan_hdr) +
2475 				 2 * RTE_ETHER_ADDR_LEN),
2476 		      "invalid Ethernet Segment data size");
2477 	psrc = rte_pktmbuf_mtod(loc->mbuf, uint8_t *);
2478 	es->inline_hdr_sz = RTE_BE16(MLX5_ESEG_MIN_INLINE_SIZE);
2479 	es->inline_data = *(unaligned_uint16_t *)psrc;
2480 	psrc +=	sizeof(uint16_t);
2481 	pdst = (uint8_t *)(es + 1);
2482 	if (MLX5_TXOFF_CONFIG(VLAN) && vlan) {
2483 		/* Implement VLAN tag insertion as part inline data. */
2484 		memcpy(pdst, psrc, 2 * RTE_ETHER_ADDR_LEN - sizeof(uint16_t));
2485 		pdst += 2 * RTE_ETHER_ADDR_LEN - sizeof(uint16_t);
2486 		psrc +=	2 * RTE_ETHER_ADDR_LEN - sizeof(uint16_t);
2487 		/* Insert VLAN ethertype + VLAN tag. */
2488 		*(unaligned_uint32_t *)pdst = rte_cpu_to_be_32
2489 						((RTE_ETHER_TYPE_VLAN << 16) |
2490 						 loc->mbuf->vlan_tci);
2491 		pdst += sizeof(struct rte_vlan_hdr);
2492 		/* Copy the rest two bytes from packet data. */
2493 		MLX5_ASSERT(pdst == RTE_PTR_ALIGN(pdst, sizeof(uint16_t)));
2494 		*(uint16_t *)pdst = *(unaligned_uint16_t *)psrc;
2495 	} else {
2496 		/* Fill the gap in the title WQEBB with inline data. */
2497 		rte_mov16(pdst, psrc);
2498 	}
2499 }
2500 
2501 /**
2502  * Build the Ethernet Segment with entire packet
2503  * data inlining. Checks the boundary of WQEBB and
2504  * ring buffer wrapping, supports Software Parser,
2505  * Checksums and VLAN insertion Tx offload features.
2506  *
2507  * @param txq
2508  *   Pointer to TX queue structure.
2509  * @param loc
2510  *   Pointer to burst routine local context.
2511  * @param wqe
2512  *   Pointer to WQE to fill with built Ethernet Segment.
2513  * @param vlan
2514  *   Length of VLAN tag insertion if any.
2515  * @param inlen
2516  *   Length of data to inline (VLAN included, if any).
2517  * @param tso
2518  *   TSO flag, set mss field from the packet.
2519  * @param olx
2520  *   Configured Tx offloads mask. It is fully defined at
2521  *   compile time and may be used for optimization.
2522  *
2523  * @return
2524  *   Pointer to the next Data Segment (aligned and wrapped around).
2525  */
2526 static __rte_always_inline struct mlx5_wqe_dseg *
2527 mlx5_tx_eseg_data(struct mlx5_txq_data *restrict txq,
2528 		  struct mlx5_txq_local *restrict loc,
2529 		  struct mlx5_wqe *restrict wqe,
2530 		  unsigned int vlan,
2531 		  unsigned int inlen,
2532 		  unsigned int tso,
2533 		  unsigned int olx)
2534 {
2535 	struct mlx5_wqe_eseg *restrict es = &wqe->eseg;
2536 	uint32_t csum;
2537 	uint8_t *psrc, *pdst;
2538 	unsigned int part;
2539 
2540 	/*
2541 	 * Calculate and set check sum flags first, dword field
2542 	 * in segment may be shared with Software Parser flags.
2543 	 */
2544 	csum = MLX5_TXOFF_CONFIG(CSUM) ? txq_ol_cksum_to_cs(loc->mbuf) : 0;
2545 	if (tso) {
2546 		csum <<= 24;
2547 		csum |= loc->mbuf->tso_segsz;
2548 		es->flags = rte_cpu_to_be_32(csum);
2549 	} else {
2550 		es->flags = rte_cpu_to_le_32(csum);
2551 	}
2552 	/*
2553 	 * Calculate and set Software Parser offsets and flags.
2554 	 * These flags a set for custom UDP and IP tunnel packets.
2555 	 */
2556 	es->swp_offs = txq_mbuf_to_swp(loc, &es->swp_flags, olx);
2557 	/* Fill metadata field if needed. */
2558 	es->metadata = MLX5_TXOFF_CONFIG(METADATA) ?
2559 		       loc->mbuf->ol_flags & PKT_TX_DYNF_METADATA ?
2560 		       *RTE_FLOW_DYNF_METADATA(loc->mbuf) : 0 : 0;
2561 	static_assert(MLX5_ESEG_MIN_INLINE_SIZE ==
2562 				(sizeof(uint16_t) +
2563 				 sizeof(rte_v128u32_t)),
2564 		      "invalid Ethernet Segment data size");
2565 	static_assert(MLX5_ESEG_MIN_INLINE_SIZE ==
2566 				(sizeof(uint16_t) +
2567 				 sizeof(struct rte_vlan_hdr) +
2568 				 2 * RTE_ETHER_ADDR_LEN),
2569 		      "invalid Ethernet Segment data size");
2570 	psrc = rte_pktmbuf_mtod(loc->mbuf, uint8_t *);
2571 	es->inline_hdr_sz = rte_cpu_to_be_16(inlen);
2572 	es->inline_data = *(unaligned_uint16_t *)psrc;
2573 	psrc +=	sizeof(uint16_t);
2574 	pdst = (uint8_t *)(es + 1);
2575 	if (MLX5_TXOFF_CONFIG(VLAN) && vlan) {
2576 		/* Implement VLAN tag insertion as part inline data. */
2577 		memcpy(pdst, psrc, 2 * RTE_ETHER_ADDR_LEN - sizeof(uint16_t));
2578 		pdst += 2 * RTE_ETHER_ADDR_LEN - sizeof(uint16_t);
2579 		psrc +=	2 * RTE_ETHER_ADDR_LEN - sizeof(uint16_t);
2580 		/* Insert VLAN ethertype + VLAN tag. */
2581 		*(unaligned_uint32_t *)pdst = rte_cpu_to_be_32
2582 						((RTE_ETHER_TYPE_VLAN << 16) |
2583 						 loc->mbuf->vlan_tci);
2584 		pdst += sizeof(struct rte_vlan_hdr);
2585 		/* Copy the rest two bytes from packet data. */
2586 		MLX5_ASSERT(pdst == RTE_PTR_ALIGN(pdst, sizeof(uint16_t)));
2587 		*(uint16_t *)pdst = *(unaligned_uint16_t *)psrc;
2588 		psrc += sizeof(uint16_t);
2589 	} else {
2590 		/* Fill the gap in the title WQEBB with inline data. */
2591 		rte_mov16(pdst, psrc);
2592 		psrc += sizeof(rte_v128u32_t);
2593 	}
2594 	pdst = (uint8_t *)(es + 2);
2595 	MLX5_ASSERT(inlen >= MLX5_ESEG_MIN_INLINE_SIZE);
2596 	MLX5_ASSERT(pdst < (uint8_t *)txq->wqes_end);
2597 	inlen -= MLX5_ESEG_MIN_INLINE_SIZE;
2598 	if (!inlen) {
2599 		MLX5_ASSERT(pdst == RTE_PTR_ALIGN(pdst, MLX5_WSEG_SIZE));
2600 		return (struct mlx5_wqe_dseg *)pdst;
2601 	}
2602 	/*
2603 	 * The WQEBB space availability is checked by caller.
2604 	 * Here we should be aware of WQE ring buffer wraparound only.
2605 	 */
2606 	part = (uint8_t *)txq->wqes_end - pdst;
2607 	part = RTE_MIN(part, inlen);
2608 	do {
2609 		rte_memcpy(pdst, psrc, part);
2610 		inlen -= part;
2611 		if (likely(!inlen)) {
2612 			/*
2613 			 * If return value is not used by the caller
2614 			 * the code below will be optimized out.
2615 			 */
2616 			pdst += part;
2617 			pdst = RTE_PTR_ALIGN(pdst, MLX5_WSEG_SIZE);
2618 			if (unlikely(pdst >= (uint8_t *)txq->wqes_end))
2619 				pdst = (uint8_t *)txq->wqes;
2620 			return (struct mlx5_wqe_dseg *)pdst;
2621 		}
2622 		pdst = (uint8_t *)txq->wqes;
2623 		psrc += part;
2624 		part = inlen;
2625 	} while (true);
2626 }
2627 
2628 /**
2629  * Copy data from chain of mbuf to the specified linear buffer.
2630  * Checksums and VLAN insertion Tx offload features. If data
2631  * from some mbuf copied completely this mbuf is freed. Local
2632  * structure is used to keep the byte stream state.
2633  *
2634  * @param pdst
2635  *   Pointer to the destination linear buffer.
2636  * @param loc
2637  *   Pointer to burst routine local context.
2638  * @param len
2639  *   Length of data to be copied.
2640  * @param must
2641  *   Length of data to be copied ignoring no inline hint.
2642  * @param olx
2643  *   Configured Tx offloads mask. It is fully defined at
2644  *   compile time and may be used for optimization.
2645  *
2646  * @return
2647  *   Number of actual copied data bytes. This is always greater than or
2648  *   equal to must parameter and might be lesser than len in no inline
2649  *   hint flag is encountered.
2650  */
2651 static __rte_always_inline unsigned int
2652 mlx5_tx_mseg_memcpy(uint8_t *pdst,
2653 		    struct mlx5_txq_local *restrict loc,
2654 		    unsigned int len,
2655 		    unsigned int must,
2656 		    unsigned int olx __rte_unused)
2657 {
2658 	struct rte_mbuf *mbuf;
2659 	unsigned int part, dlen, copy = 0;
2660 	uint8_t *psrc;
2661 
2662 	MLX5_ASSERT(len);
2663 	MLX5_ASSERT(must <= len);
2664 	do {
2665 		/* Allow zero length packets, must check first. */
2666 		dlen = rte_pktmbuf_data_len(loc->mbuf);
2667 		if (dlen <= loc->mbuf_off) {
2668 			/* Exhausted packet, just free. */
2669 			mbuf = loc->mbuf;
2670 			loc->mbuf = mbuf->next;
2671 			rte_pktmbuf_free_seg(mbuf);
2672 			loc->mbuf_off = 0;
2673 			MLX5_ASSERT(loc->mbuf_nseg > 1);
2674 			MLX5_ASSERT(loc->mbuf);
2675 			--loc->mbuf_nseg;
2676 			if (loc->mbuf->ol_flags & PKT_TX_DYNF_NOINLINE) {
2677 				unsigned int diff;
2678 
2679 				if (copy >= must) {
2680 					/*
2681 					 * We already copied the minimal
2682 					 * requested amount of data.
2683 					 */
2684 					return copy;
2685 				}
2686 				diff = must - copy;
2687 				if (diff <= rte_pktmbuf_data_len(loc->mbuf)) {
2688 					/*
2689 					 * Copy only the minimal required
2690 					 * part of the data buffer.
2691 					 */
2692 					len = diff;
2693 				}
2694 			}
2695 			continue;
2696 		}
2697 		dlen -= loc->mbuf_off;
2698 		psrc = rte_pktmbuf_mtod_offset(loc->mbuf, uint8_t *,
2699 					       loc->mbuf_off);
2700 		part = RTE_MIN(len, dlen);
2701 		rte_memcpy(pdst, psrc, part);
2702 		copy += part;
2703 		loc->mbuf_off += part;
2704 		len -= part;
2705 		if (!len) {
2706 			if (loc->mbuf_off >= rte_pktmbuf_data_len(loc->mbuf)) {
2707 				loc->mbuf_off = 0;
2708 				/* Exhausted packet, just free. */
2709 				mbuf = loc->mbuf;
2710 				loc->mbuf = mbuf->next;
2711 				rte_pktmbuf_free_seg(mbuf);
2712 				loc->mbuf_off = 0;
2713 				MLX5_ASSERT(loc->mbuf_nseg >= 1);
2714 				--loc->mbuf_nseg;
2715 			}
2716 			return copy;
2717 		}
2718 		pdst += part;
2719 	} while (true);
2720 }
2721 
2722 /**
2723  * Build the Ethernet Segment with inlined data from
2724  * multi-segment packet. Checks the boundary of WQEBB
2725  * and ring buffer wrapping, supports Software Parser,
2726  * Checksums and VLAN insertion Tx offload features.
2727  *
2728  * @param txq
2729  *   Pointer to TX queue structure.
2730  * @param loc
2731  *   Pointer to burst routine local context.
2732  * @param wqe
2733  *   Pointer to WQE to fill with built Ethernet Segment.
2734  * @param vlan
2735  *   Length of VLAN tag insertion if any.
2736  * @param inlen
2737  *   Length of data to inline (VLAN included, if any).
2738  * @param tso
2739  *   TSO flag, set mss field from the packet.
2740  * @param olx
2741  *   Configured Tx offloads mask. It is fully defined at
2742  *   compile time and may be used for optimization.
2743  *
2744  * @return
2745  *   Pointer to the next Data Segment (aligned and
2746  *   possible NOT wrapped around - caller should do
2747  *   wrapping check on its own).
2748  */
2749 static __rte_always_inline struct mlx5_wqe_dseg *
2750 mlx5_tx_eseg_mdat(struct mlx5_txq_data *restrict txq,
2751 		  struct mlx5_txq_local *restrict loc,
2752 		  struct mlx5_wqe *restrict wqe,
2753 		  unsigned int vlan,
2754 		  unsigned int inlen,
2755 		  unsigned int tso,
2756 		  unsigned int olx)
2757 {
2758 	struct mlx5_wqe_eseg *restrict es = &wqe->eseg;
2759 	uint32_t csum;
2760 	uint8_t *pdst;
2761 	unsigned int part, tlen = 0;
2762 
2763 	/*
2764 	 * Calculate and set check sum flags first, uint32_t field
2765 	 * in segment may be shared with Software Parser flags.
2766 	 */
2767 	csum = MLX5_TXOFF_CONFIG(CSUM) ? txq_ol_cksum_to_cs(loc->mbuf) : 0;
2768 	if (tso) {
2769 		csum <<= 24;
2770 		csum |= loc->mbuf->tso_segsz;
2771 		es->flags = rte_cpu_to_be_32(csum);
2772 	} else {
2773 		es->flags = rte_cpu_to_le_32(csum);
2774 	}
2775 	/*
2776 	 * Calculate and set Software Parser offsets and flags.
2777 	 * These flags a set for custom UDP and IP tunnel packets.
2778 	 */
2779 	es->swp_offs = txq_mbuf_to_swp(loc, &es->swp_flags, olx);
2780 	/* Fill metadata field if needed. */
2781 	es->metadata = MLX5_TXOFF_CONFIG(METADATA) ?
2782 		       loc->mbuf->ol_flags & PKT_TX_DYNF_METADATA ?
2783 		       *RTE_FLOW_DYNF_METADATA(loc->mbuf) : 0 : 0;
2784 	static_assert(MLX5_ESEG_MIN_INLINE_SIZE ==
2785 				(sizeof(uint16_t) +
2786 				 sizeof(rte_v128u32_t)),
2787 		      "invalid Ethernet Segment data size");
2788 	static_assert(MLX5_ESEG_MIN_INLINE_SIZE ==
2789 				(sizeof(uint16_t) +
2790 				 sizeof(struct rte_vlan_hdr) +
2791 				 2 * RTE_ETHER_ADDR_LEN),
2792 		      "invalid Ethernet Segment data size");
2793 	MLX5_ASSERT(inlen >= MLX5_ESEG_MIN_INLINE_SIZE);
2794 	pdst = (uint8_t *)&es->inline_data;
2795 	if (MLX5_TXOFF_CONFIG(VLAN) && vlan) {
2796 		/* Implement VLAN tag insertion as part inline data. */
2797 		mlx5_tx_mseg_memcpy(pdst, loc,
2798 				    2 * RTE_ETHER_ADDR_LEN,
2799 				    2 * RTE_ETHER_ADDR_LEN, olx);
2800 		pdst += 2 * RTE_ETHER_ADDR_LEN;
2801 		*(unaligned_uint32_t *)pdst = rte_cpu_to_be_32
2802 						((RTE_ETHER_TYPE_VLAN << 16) |
2803 						 loc->mbuf->vlan_tci);
2804 		pdst += sizeof(struct rte_vlan_hdr);
2805 		tlen += 2 * RTE_ETHER_ADDR_LEN + sizeof(struct rte_vlan_hdr);
2806 	}
2807 	MLX5_ASSERT(pdst < (uint8_t *)txq->wqes_end);
2808 	/*
2809 	 * The WQEBB space availability is checked by caller.
2810 	 * Here we should be aware of WQE ring buffer wraparound only.
2811 	 */
2812 	part = (uint8_t *)txq->wqes_end - pdst;
2813 	part = RTE_MIN(part, inlen - tlen);
2814 	MLX5_ASSERT(part);
2815 	do {
2816 		unsigned int copy;
2817 
2818 		/*
2819 		 * Copying may be interrupted inside the routine
2820 		 * if run into no inline hint flag.
2821 		 */
2822 		copy = tlen >= txq->inlen_mode ? 0 : (txq->inlen_mode - tlen);
2823 		copy = mlx5_tx_mseg_memcpy(pdst, loc, part, copy, olx);
2824 		tlen += copy;
2825 		if (likely(inlen <= tlen) || copy < part) {
2826 			es->inline_hdr_sz = rte_cpu_to_be_16(tlen);
2827 			pdst += copy;
2828 			pdst = RTE_PTR_ALIGN(pdst, MLX5_WSEG_SIZE);
2829 			return (struct mlx5_wqe_dseg *)pdst;
2830 		}
2831 		pdst = (uint8_t *)txq->wqes;
2832 		part = inlen - tlen;
2833 	} while (true);
2834 }
2835 
2836 /**
2837  * Build the Data Segment of pointer type.
2838  *
2839  * @param txq
2840  *   Pointer to TX queue structure.
2841  * @param loc
2842  *   Pointer to burst routine local context.
2843  * @param dseg
2844  *   Pointer to WQE to fill with built Data Segment.
2845  * @param buf
2846  *   Data buffer to point.
2847  * @param len
2848  *   Data buffer length.
2849  * @param olx
2850  *   Configured Tx offloads mask. It is fully defined at
2851  *   compile time and may be used for optimization.
2852  */
2853 static __rte_always_inline void
2854 mlx5_tx_dseg_ptr(struct mlx5_txq_data *restrict txq,
2855 		 struct mlx5_txq_local *restrict loc,
2856 		 struct mlx5_wqe_dseg *restrict dseg,
2857 		 uint8_t *buf,
2858 		 unsigned int len,
2859 		 unsigned int olx __rte_unused)
2860 
2861 {
2862 	MLX5_ASSERT(len);
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 
2868 /**
2869  * Build the Data Segment of pointer type or inline
2870  * if data length is less than buffer in minimal
2871  * Data Segment size.
2872  *
2873  * @param txq
2874  *   Pointer to TX queue structure.
2875  * @param loc
2876  *   Pointer to burst routine local context.
2877  * @param dseg
2878  *   Pointer to WQE to fill with built Data Segment.
2879  * @param buf
2880  *   Data buffer to point.
2881  * @param len
2882  *   Data buffer length.
2883  * @param olx
2884  *   Configured Tx offloads mask. It is fully defined at
2885  *   compile time and may be used for optimization.
2886  */
2887 static __rte_always_inline void
2888 mlx5_tx_dseg_iptr(struct mlx5_txq_data *restrict txq,
2889 		  struct mlx5_txq_local *restrict loc,
2890 		  struct mlx5_wqe_dseg *restrict dseg,
2891 		  uint8_t *buf,
2892 		  unsigned int len,
2893 		  unsigned int olx __rte_unused)
2894 
2895 {
2896 	uintptr_t dst, src;
2897 
2898 	MLX5_ASSERT(len);
2899 	if (len > MLX5_DSEG_MIN_INLINE_SIZE) {
2900 		dseg->bcount = rte_cpu_to_be_32(len);
2901 		dseg->lkey = mlx5_tx_mb2mr(txq, loc->mbuf);
2902 		dseg->pbuf = rte_cpu_to_be_64((uintptr_t)buf);
2903 
2904 		return;
2905 	}
2906 	dseg->bcount = rte_cpu_to_be_32(len | MLX5_ETH_WQE_DATA_INLINE);
2907 	/* Unrolled implementation of generic rte_memcpy. */
2908 	dst = (uintptr_t)&dseg->inline_data[0];
2909 	src = (uintptr_t)buf;
2910 	if (len & 0x08) {
2911 #ifdef RTE_ARCH_STRICT_ALIGN
2912 		MLX5_ASSERT(dst == RTE_PTR_ALIGN(dst, sizeof(uint32_t)));
2913 		*(uint32_t *)dst = *(unaligned_uint32_t *)src;
2914 		dst += sizeof(uint32_t);
2915 		src += sizeof(uint32_t);
2916 		*(uint32_t *)dst = *(unaligned_uint32_t *)src;
2917 		dst += sizeof(uint32_t);
2918 		src += sizeof(uint32_t);
2919 #else
2920 		*(uint64_t *)dst = *(unaligned_uint64_t *)src;
2921 		dst += sizeof(uint64_t);
2922 		src += sizeof(uint64_t);
2923 #endif
2924 	}
2925 	if (len & 0x04) {
2926 		*(uint32_t *)dst = *(unaligned_uint32_t *)src;
2927 		dst += sizeof(uint32_t);
2928 		src += sizeof(uint32_t);
2929 	}
2930 	if (len & 0x02) {
2931 		*(uint16_t *)dst = *(unaligned_uint16_t *)src;
2932 		dst += sizeof(uint16_t);
2933 		src += sizeof(uint16_t);
2934 	}
2935 	if (len & 0x01)
2936 		*(uint8_t *)dst = *(uint8_t *)src;
2937 }
2938 
2939 /**
2940  * Build the Data Segment of inlined data from single
2941  * segment packet, no VLAN insertion.
2942  *
2943  * @param txq
2944  *   Pointer to TX queue structure.
2945  * @param loc
2946  *   Pointer to burst routine local context.
2947  * @param dseg
2948  *   Pointer to WQE to fill with built Data Segment.
2949  * @param buf
2950  *   Data buffer to point.
2951  * @param len
2952  *   Data buffer length.
2953  * @param olx
2954  *   Configured Tx offloads mask. It is fully defined at
2955  *   compile time and may be used for optimization.
2956  *
2957  * @return
2958  *   Pointer to the next Data Segment after inlined data.
2959  *   Ring buffer wraparound check is needed. We do not
2960  *   do it here because it may not be needed for the
2961  *   last packet in the eMPW session.
2962  */
2963 static __rte_always_inline struct mlx5_wqe_dseg *
2964 mlx5_tx_dseg_empw(struct mlx5_txq_data *restrict txq,
2965 		  struct mlx5_txq_local *restrict loc __rte_unused,
2966 		  struct mlx5_wqe_dseg *restrict dseg,
2967 		  uint8_t *buf,
2968 		  unsigned int len,
2969 		  unsigned int olx __rte_unused)
2970 {
2971 	unsigned int part;
2972 	uint8_t *pdst;
2973 
2974 	if (!MLX5_TXOFF_CONFIG(MPW)) {
2975 		/* Store the descriptor byte counter for eMPW sessions. */
2976 		dseg->bcount = rte_cpu_to_be_32(len | MLX5_ETH_WQE_DATA_INLINE);
2977 		pdst = &dseg->inline_data[0];
2978 	} else {
2979 		/* The entire legacy MPW session counter is stored on close. */
2980 		pdst = (uint8_t *)dseg;
2981 	}
2982 	/*
2983 	 * The WQEBB space availability is checked by caller.
2984 	 * Here we should be aware of WQE ring buffer wraparound only.
2985 	 */
2986 	part = (uint8_t *)txq->wqes_end - pdst;
2987 	part = RTE_MIN(part, len);
2988 	do {
2989 		rte_memcpy(pdst, buf, part);
2990 		len -= part;
2991 		if (likely(!len)) {
2992 			pdst += part;
2993 			if (!MLX5_TXOFF_CONFIG(MPW))
2994 				pdst = RTE_PTR_ALIGN(pdst, MLX5_WSEG_SIZE);
2995 			/* Note: no final wraparound check here. */
2996 			return (struct mlx5_wqe_dseg *)pdst;
2997 		}
2998 		pdst = (uint8_t *)txq->wqes;
2999 		buf += part;
3000 		part = len;
3001 	} while (true);
3002 }
3003 
3004 /**
3005  * Build the Data Segment of inlined data from single
3006  * segment packet with VLAN insertion.
3007  *
3008  * @param txq
3009  *   Pointer to TX queue structure.
3010  * @param loc
3011  *   Pointer to burst routine local context.
3012  * @param dseg
3013  *   Pointer to the dseg fill with built Data Segment.
3014  * @param buf
3015  *   Data buffer to point.
3016  * @param len
3017  *   Data buffer length.
3018  * @param olx
3019  *   Configured Tx offloads mask. It is fully defined at
3020  *   compile time and may be used for optimization.
3021  *
3022  * @return
3023  *   Pointer to the next Data Segment after inlined data.
3024  *   Ring buffer wraparound check is needed.
3025  */
3026 static __rte_always_inline struct mlx5_wqe_dseg *
3027 mlx5_tx_dseg_vlan(struct mlx5_txq_data *restrict txq,
3028 		  struct mlx5_txq_local *restrict loc __rte_unused,
3029 		  struct mlx5_wqe_dseg *restrict dseg,
3030 		  uint8_t *buf,
3031 		  unsigned int len,
3032 		  unsigned int olx __rte_unused)
3033 
3034 {
3035 	unsigned int part;
3036 	uint8_t *pdst;
3037 
3038 	MLX5_ASSERT(len > MLX5_ESEG_MIN_INLINE_SIZE);
3039 	static_assert(MLX5_DSEG_MIN_INLINE_SIZE ==
3040 				 (2 * RTE_ETHER_ADDR_LEN),
3041 		      "invalid Data Segment data size");
3042 	if (!MLX5_TXOFF_CONFIG(MPW)) {
3043 		/* Store the descriptor byte counter for eMPW sessions. */
3044 		dseg->bcount = rte_cpu_to_be_32
3045 				((len + sizeof(struct rte_vlan_hdr)) |
3046 				 MLX5_ETH_WQE_DATA_INLINE);
3047 		pdst = &dseg->inline_data[0];
3048 	} else {
3049 		/* The entire legacy MPW session counter is stored on close. */
3050 		pdst = (uint8_t *)dseg;
3051 	}
3052 	memcpy(pdst, buf, MLX5_DSEG_MIN_INLINE_SIZE);
3053 	buf += MLX5_DSEG_MIN_INLINE_SIZE;
3054 	pdst += MLX5_DSEG_MIN_INLINE_SIZE;
3055 	len -= MLX5_DSEG_MIN_INLINE_SIZE;
3056 	/* Insert VLAN ethertype + VLAN tag. Pointer is aligned. */
3057 	MLX5_ASSERT(pdst == RTE_PTR_ALIGN(pdst, MLX5_WSEG_SIZE));
3058 	if (unlikely(pdst >= (uint8_t *)txq->wqes_end))
3059 		pdst = (uint8_t *)txq->wqes;
3060 	*(uint32_t *)pdst = rte_cpu_to_be_32((RTE_ETHER_TYPE_VLAN << 16) |
3061 					      loc->mbuf->vlan_tci);
3062 	pdst += sizeof(struct rte_vlan_hdr);
3063 	/*
3064 	 * The WQEBB space availability is checked by caller.
3065 	 * Here we should be aware of WQE ring buffer wraparound only.
3066 	 */
3067 	part = (uint8_t *)txq->wqes_end - pdst;
3068 	part = RTE_MIN(part, len);
3069 	do {
3070 		rte_memcpy(pdst, buf, part);
3071 		len -= part;
3072 		if (likely(!len)) {
3073 			pdst += part;
3074 			if (!MLX5_TXOFF_CONFIG(MPW))
3075 				pdst = RTE_PTR_ALIGN(pdst, MLX5_WSEG_SIZE);
3076 			/* Note: no final wraparound check here. */
3077 			return (struct mlx5_wqe_dseg *)pdst;
3078 		}
3079 		pdst = (uint8_t *)txq->wqes;
3080 		buf += part;
3081 		part = len;
3082 	} while (true);
3083 }
3084 
3085 /**
3086  * Build the Ethernet Segment with optionally inlined data with
3087  * VLAN insertion and following Data Segments (if any) from
3088  * multi-segment packet. Used by ordinary send and TSO.
3089  *
3090  * @param txq
3091  *   Pointer to TX queue structure.
3092  * @param loc
3093  *   Pointer to burst routine local context.
3094  * @param wqe
3095  *   Pointer to WQE to fill with built Ethernet/Data Segments.
3096  * @param vlan
3097  *   Length of VLAN header to insert, 0 means no VLAN insertion.
3098  * @param inlen
3099  *   Data length to inline. For TSO this parameter specifies
3100  *   exact value, for ordinary send routine can be aligned by
3101  *   caller to provide better WQE space saving and data buffer
3102  *   start address alignment. This length includes VLAN header
3103  *   being inserted.
3104  * @param tso
3105  *   Zero means ordinary send, inlined data can be extended,
3106  *   otherwise this is TSO, inlined data length is fixed.
3107  * @param olx
3108  *   Configured Tx offloads mask. It is fully defined at
3109  *   compile time and may be used for optimization.
3110  *
3111  * @return
3112  *   Actual size of built WQE in segments.
3113  */
3114 static __rte_always_inline unsigned int
3115 mlx5_tx_mseg_build(struct mlx5_txq_data *restrict txq,
3116 		   struct mlx5_txq_local *restrict loc,
3117 		   struct mlx5_wqe *restrict wqe,
3118 		   unsigned int vlan,
3119 		   unsigned int inlen,
3120 		   unsigned int tso,
3121 		   unsigned int olx __rte_unused)
3122 {
3123 	struct mlx5_wqe_dseg *restrict dseg;
3124 	unsigned int ds;
3125 
3126 	MLX5_ASSERT((rte_pktmbuf_pkt_len(loc->mbuf) + vlan) >= inlen);
3127 	loc->mbuf_nseg = NB_SEGS(loc->mbuf);
3128 	loc->mbuf_off = 0;
3129 
3130 	dseg = mlx5_tx_eseg_mdat(txq, loc, wqe, vlan, inlen, tso, olx);
3131 	if (!loc->mbuf_nseg)
3132 		goto dseg_done;
3133 	/*
3134 	 * There are still some mbuf remaining, not inlined.
3135 	 * The first mbuf may be partially inlined and we
3136 	 * must process the possible non-zero data offset.
3137 	 */
3138 	if (loc->mbuf_off) {
3139 		unsigned int dlen;
3140 		uint8_t *dptr;
3141 
3142 		/*
3143 		 * Exhausted packets must be dropped before.
3144 		 * Non-zero offset means there are some data
3145 		 * remained in the packet.
3146 		 */
3147 		MLX5_ASSERT(loc->mbuf_off < rte_pktmbuf_data_len(loc->mbuf));
3148 		MLX5_ASSERT(rte_pktmbuf_data_len(loc->mbuf));
3149 		dptr = rte_pktmbuf_mtod_offset(loc->mbuf, uint8_t *,
3150 					       loc->mbuf_off);
3151 		dlen = rte_pktmbuf_data_len(loc->mbuf) - loc->mbuf_off;
3152 		/*
3153 		 * Build the pointer/minimal data Data Segment.
3154 		 * Do ring buffer wrapping check in advance.
3155 		 */
3156 		if ((uintptr_t)dseg >= (uintptr_t)txq->wqes_end)
3157 			dseg = (struct mlx5_wqe_dseg *)txq->wqes;
3158 		mlx5_tx_dseg_iptr(txq, loc, dseg, dptr, dlen, olx);
3159 		/* Store the mbuf to be freed on completion. */
3160 		MLX5_ASSERT(loc->elts_free);
3161 		txq->elts[txq->elts_head++ & txq->elts_m] = loc->mbuf;
3162 		--loc->elts_free;
3163 		++dseg;
3164 		if (--loc->mbuf_nseg == 0)
3165 			goto dseg_done;
3166 		loc->mbuf = loc->mbuf->next;
3167 		loc->mbuf_off = 0;
3168 	}
3169 	do {
3170 		if (unlikely(!rte_pktmbuf_data_len(loc->mbuf))) {
3171 			struct rte_mbuf *mbuf;
3172 
3173 			/* Zero length segment found, just skip. */
3174 			mbuf = loc->mbuf;
3175 			loc->mbuf = loc->mbuf->next;
3176 			rte_pktmbuf_free_seg(mbuf);
3177 			if (--loc->mbuf_nseg == 0)
3178 				break;
3179 		} else {
3180 			if ((uintptr_t)dseg >= (uintptr_t)txq->wqes_end)
3181 				dseg = (struct mlx5_wqe_dseg *)txq->wqes;
3182 			mlx5_tx_dseg_iptr
3183 				(txq, loc, dseg,
3184 				 rte_pktmbuf_mtod(loc->mbuf, uint8_t *),
3185 				 rte_pktmbuf_data_len(loc->mbuf), olx);
3186 			MLX5_ASSERT(loc->elts_free);
3187 			txq->elts[txq->elts_head++ & txq->elts_m] = loc->mbuf;
3188 			--loc->elts_free;
3189 			++dseg;
3190 			if (--loc->mbuf_nseg == 0)
3191 				break;
3192 			loc->mbuf = loc->mbuf->next;
3193 		}
3194 	} while (true);
3195 
3196 dseg_done:
3197 	/* Calculate actual segments used from the dseg pointer. */
3198 	if ((uintptr_t)wqe < (uintptr_t)dseg)
3199 		ds = ((uintptr_t)dseg - (uintptr_t)wqe) / MLX5_WSEG_SIZE;
3200 	else
3201 		ds = (((uintptr_t)dseg - (uintptr_t)wqe) +
3202 		      txq->wqe_s * MLX5_WQE_SIZE) / MLX5_WSEG_SIZE;
3203 	return ds;
3204 }
3205 
3206 /**
3207  * Tx one packet function for multi-segment TSO. Supports all
3208  * types of Tx offloads, uses MLX5_OPCODE_TSO to build WQEs,
3209  * sends one packet per WQE.
3210  *
3211  * This routine is responsible for storing processed mbuf
3212  * into elts ring buffer and update elts_head.
3213  *
3214  * @param txq
3215  *   Pointer to TX queue structure.
3216  * @param loc
3217  *   Pointer to burst routine local context.
3218  * @param olx
3219  *   Configured Tx offloads mask. It is fully defined at
3220  *   compile time and may be used for optimization.
3221  *
3222  * @return
3223  *   MLX5_TXCMP_CODE_EXIT - sending is done or impossible.
3224  *   MLX5_TXCMP_CODE_ERROR - some unrecoverable error occurred.
3225  * Local context variables partially updated.
3226  */
3227 static __rte_always_inline enum mlx5_txcmp_code
3228 mlx5_tx_packet_multi_tso(struct mlx5_txq_data *restrict txq,
3229 			struct mlx5_txq_local *restrict loc,
3230 			unsigned int olx)
3231 {
3232 	struct mlx5_wqe *restrict wqe;
3233 	unsigned int ds, dlen, inlen, ntcp, vlan = 0;
3234 
3235 	/*
3236 	 * Calculate data length to be inlined to estimate
3237 	 * the required space in WQE ring buffer.
3238 	 */
3239 	dlen = rte_pktmbuf_pkt_len(loc->mbuf);
3240 	if (MLX5_TXOFF_CONFIG(VLAN) && loc->mbuf->ol_flags & PKT_TX_VLAN_PKT)
3241 		vlan = sizeof(struct rte_vlan_hdr);
3242 	inlen = loc->mbuf->l2_len + vlan +
3243 		loc->mbuf->l3_len + loc->mbuf->l4_len;
3244 	if (unlikely((!inlen || !loc->mbuf->tso_segsz)))
3245 		return MLX5_TXCMP_CODE_ERROR;
3246 	if (loc->mbuf->ol_flags & PKT_TX_TUNNEL_MASK)
3247 		inlen += loc->mbuf->outer_l2_len + loc->mbuf->outer_l3_len;
3248 	/* Packet must contain all TSO headers. */
3249 	if (unlikely(inlen > MLX5_MAX_TSO_HEADER ||
3250 		     inlen <= MLX5_ESEG_MIN_INLINE_SIZE ||
3251 		     inlen > (dlen + vlan)))
3252 		return MLX5_TXCMP_CODE_ERROR;
3253 	MLX5_ASSERT(inlen >= txq->inlen_mode);
3254 	/*
3255 	 * Check whether there are enough free WQEBBs:
3256 	 * - Control Segment
3257 	 * - Ethernet Segment
3258 	 * - First Segment of inlined Ethernet data
3259 	 * - ... data continued ...
3260 	 * - Data Segments of pointer/min inline type
3261 	 */
3262 	ds = NB_SEGS(loc->mbuf) + 2 + (inlen -
3263 				       MLX5_ESEG_MIN_INLINE_SIZE +
3264 				       MLX5_WSEG_SIZE +
3265 				       MLX5_WSEG_SIZE - 1) / MLX5_WSEG_SIZE;
3266 	if (unlikely(loc->wqe_free < ((ds + 3) / 4)))
3267 		return MLX5_TXCMP_CODE_EXIT;
3268 	/* Check for maximal WQE size. */
3269 	if (unlikely((MLX5_WQE_SIZE_MAX / MLX5_WSEG_SIZE) < ((ds + 3) / 4)))
3270 		return MLX5_TXCMP_CODE_ERROR;
3271 #ifdef MLX5_PMD_SOFT_COUNTERS
3272 	/* Update sent data bytes/packets counters. */
3273 	ntcp = (dlen - (inlen - vlan) + loc->mbuf->tso_segsz - 1) /
3274 		loc->mbuf->tso_segsz;
3275 	/*
3276 	 * One will be added for mbuf itself
3277 	 * at the end of the mlx5_tx_burst from
3278 	 * loc->pkts_sent field.
3279 	 */
3280 	--ntcp;
3281 	txq->stats.opackets += ntcp;
3282 	txq->stats.obytes += dlen + vlan + ntcp * inlen;
3283 #endif
3284 	wqe = txq->wqes + (txq->wqe_ci & txq->wqe_m);
3285 	loc->wqe_last = wqe;
3286 	mlx5_tx_cseg_init(txq, loc, wqe, 0, MLX5_OPCODE_TSO, olx);
3287 	ds = mlx5_tx_mseg_build(txq, loc, wqe, vlan, inlen, 1, olx);
3288 	wqe->cseg.sq_ds = rte_cpu_to_be_32(txq->qp_num_8s | ds);
3289 	txq->wqe_ci += (ds + 3) / 4;
3290 	loc->wqe_free -= (ds + 3) / 4;
3291 	return MLX5_TXCMP_CODE_MULTI;
3292 }
3293 
3294 /**
3295  * Tx one packet function for multi-segment SEND. Supports all
3296  * types of Tx offloads, uses MLX5_OPCODE_SEND to build WQEs,
3297  * sends one packet per WQE, without any data inlining in
3298  * Ethernet Segment.
3299  *
3300  * This routine is responsible for storing processed mbuf
3301  * into elts ring buffer and update elts_head.
3302  *
3303  * @param txq
3304  *   Pointer to TX queue structure.
3305  * @param loc
3306  *   Pointer to burst routine local context.
3307  * @param olx
3308  *   Configured Tx offloads mask. It is fully defined at
3309  *   compile time and may be used for optimization.
3310  *
3311  * @return
3312  *   MLX5_TXCMP_CODE_EXIT - sending is done or impossible.
3313  *   MLX5_TXCMP_CODE_ERROR - some unrecoverable error occurred.
3314  * Local context variables partially updated.
3315  */
3316 static __rte_always_inline enum mlx5_txcmp_code
3317 mlx5_tx_packet_multi_send(struct mlx5_txq_data *restrict txq,
3318 			  struct mlx5_txq_local *restrict loc,
3319 			  unsigned int olx)
3320 {
3321 	struct mlx5_wqe_dseg *restrict dseg;
3322 	struct mlx5_wqe *restrict wqe;
3323 	unsigned int ds, nseg;
3324 
3325 	MLX5_ASSERT(NB_SEGS(loc->mbuf) > 1);
3326 	/*
3327 	 * No inline at all, it means the CPU cycles saving
3328 	 * is prioritized at configuration, we should not
3329 	 * copy any packet data to WQE.
3330 	 */
3331 	nseg = NB_SEGS(loc->mbuf);
3332 	ds = 2 + nseg;
3333 	if (unlikely(loc->wqe_free < ((ds + 3) / 4)))
3334 		return MLX5_TXCMP_CODE_EXIT;
3335 	/* Check for maximal WQE size. */
3336 	if (unlikely((MLX5_WQE_SIZE_MAX / MLX5_WSEG_SIZE) < ((ds + 3) / 4)))
3337 		return MLX5_TXCMP_CODE_ERROR;
3338 	/*
3339 	 * Some Tx offloads may cause an error if
3340 	 * packet is not long enough, check against
3341 	 * assumed minimal length.
3342 	 */
3343 	if (rte_pktmbuf_pkt_len(loc->mbuf) <= MLX5_ESEG_MIN_INLINE_SIZE)
3344 		return MLX5_TXCMP_CODE_ERROR;
3345 #ifdef MLX5_PMD_SOFT_COUNTERS
3346 	/* Update sent data bytes counter. */
3347 	txq->stats.obytes += rte_pktmbuf_pkt_len(loc->mbuf);
3348 	if (MLX5_TXOFF_CONFIG(VLAN) &&
3349 	    loc->mbuf->ol_flags & PKT_TX_VLAN_PKT)
3350 		txq->stats.obytes += sizeof(struct rte_vlan_hdr);
3351 #endif
3352 	/*
3353 	 * SEND WQE, one WQEBB:
3354 	 * - Control Segment, SEND opcode
3355 	 * - Ethernet Segment, optional VLAN, no inline
3356 	 * - Data Segments, pointer only type
3357 	 */
3358 	wqe = txq->wqes + (txq->wqe_ci & txq->wqe_m);
3359 	loc->wqe_last = wqe;
3360 	mlx5_tx_cseg_init(txq, loc, wqe, ds, MLX5_OPCODE_SEND, olx);
3361 	mlx5_tx_eseg_none(txq, loc, wqe, olx);
3362 	dseg = &wqe->dseg[0];
3363 	do {
3364 		if (unlikely(!rte_pktmbuf_data_len(loc->mbuf))) {
3365 			struct rte_mbuf *mbuf;
3366 
3367 			/*
3368 			 * Zero length segment found, have to
3369 			 * correct total size of WQE in segments.
3370 			 * It is supposed to be rare occasion, so
3371 			 * in normal case (no zero length segments)
3372 			 * we avoid extra writing to the Control
3373 			 * Segment.
3374 			 */
3375 			--ds;
3376 			wqe->cseg.sq_ds -= RTE_BE32(1);
3377 			mbuf = loc->mbuf;
3378 			loc->mbuf = mbuf->next;
3379 			rte_pktmbuf_free_seg(mbuf);
3380 			if (--nseg == 0)
3381 				break;
3382 		} else {
3383 			mlx5_tx_dseg_ptr
3384 				(txq, loc, dseg,
3385 				 rte_pktmbuf_mtod(loc->mbuf, uint8_t *),
3386 				 rte_pktmbuf_data_len(loc->mbuf), olx);
3387 			txq->elts[txq->elts_head++ & txq->elts_m] = loc->mbuf;
3388 			--loc->elts_free;
3389 			if (--nseg == 0)
3390 				break;
3391 			++dseg;
3392 			if ((uintptr_t)dseg >= (uintptr_t)txq->wqes_end)
3393 				dseg = (struct mlx5_wqe_dseg *)txq->wqes;
3394 			loc->mbuf = loc->mbuf->next;
3395 		}
3396 	} while (true);
3397 	txq->wqe_ci += (ds + 3) / 4;
3398 	loc->wqe_free -= (ds + 3) / 4;
3399 	return MLX5_TXCMP_CODE_MULTI;
3400 }
3401 
3402 /**
3403  * Tx one packet function for multi-segment SEND. Supports all
3404  * types of Tx offloads, uses MLX5_OPCODE_SEND to build WQEs,
3405  * sends one packet per WQE, with data inlining in
3406  * Ethernet Segment and minimal Data Segments.
3407  *
3408  * This routine is responsible for storing processed mbuf
3409  * into elts ring buffer and update elts_head.
3410  *
3411  * @param txq
3412  *   Pointer to TX queue structure.
3413  * @param loc
3414  *   Pointer to burst routine local context.
3415  * @param olx
3416  *   Configured Tx offloads mask. It is fully defined at
3417  *   compile time and may be used for optimization.
3418  *
3419  * @return
3420  *   MLX5_TXCMP_CODE_EXIT - sending is done or impossible.
3421  *   MLX5_TXCMP_CODE_ERROR - some unrecoverable error occurred.
3422  * Local context variables partially updated.
3423  */
3424 static __rte_always_inline enum mlx5_txcmp_code
3425 mlx5_tx_packet_multi_inline(struct mlx5_txq_data *restrict txq,
3426 			    struct mlx5_txq_local *restrict loc,
3427 			    unsigned int olx)
3428 {
3429 	struct mlx5_wqe *restrict wqe;
3430 	unsigned int ds, inlen, dlen, vlan = 0;
3431 
3432 	MLX5_ASSERT(MLX5_TXOFF_CONFIG(INLINE));
3433 	MLX5_ASSERT(NB_SEGS(loc->mbuf) > 1);
3434 	/*
3435 	 * First calculate data length to be inlined
3436 	 * to estimate the required space for WQE.
3437 	 */
3438 	dlen = rte_pktmbuf_pkt_len(loc->mbuf);
3439 	if (MLX5_TXOFF_CONFIG(VLAN) && loc->mbuf->ol_flags & PKT_TX_VLAN_PKT)
3440 		vlan = sizeof(struct rte_vlan_hdr);
3441 	inlen = dlen + vlan;
3442 	/* Check against minimal length. */
3443 	if (inlen <= MLX5_ESEG_MIN_INLINE_SIZE)
3444 		return MLX5_TXCMP_CODE_ERROR;
3445 	MLX5_ASSERT(txq->inlen_send >= MLX5_ESEG_MIN_INLINE_SIZE);
3446 	if (inlen > txq->inlen_send ||
3447 	    loc->mbuf->ol_flags & PKT_TX_DYNF_NOINLINE) {
3448 		struct rte_mbuf *mbuf;
3449 		unsigned int nxlen;
3450 		uintptr_t start;
3451 
3452 		/*
3453 		 * Packet length exceeds the allowed inline
3454 		 * data length, check whether the minimal
3455 		 * inlining is required.
3456 		 */
3457 		if (txq->inlen_mode) {
3458 			MLX5_ASSERT(txq->inlen_mode >=
3459 				    MLX5_ESEG_MIN_INLINE_SIZE);
3460 			MLX5_ASSERT(txq->inlen_mode <= txq->inlen_send);
3461 			inlen = txq->inlen_mode;
3462 		} else {
3463 			if (loc->mbuf->ol_flags & PKT_TX_DYNF_NOINLINE ||
3464 			    !vlan || txq->vlan_en) {
3465 				/*
3466 				 * VLAN insertion will be done inside by HW.
3467 				 * It is not utmost effective - VLAN flag is
3468 				 * checked twice, but we should proceed the
3469 				 * inlining length correctly and take into
3470 				 * account the VLAN header being inserted.
3471 				 */
3472 				return mlx5_tx_packet_multi_send
3473 							(txq, loc, olx);
3474 			}
3475 			inlen = MLX5_ESEG_MIN_INLINE_SIZE;
3476 		}
3477 		/*
3478 		 * Now we know the minimal amount of data is requested
3479 		 * to inline. Check whether we should inline the buffers
3480 		 * from the chain beginning to eliminate some mbufs.
3481 		 */
3482 		mbuf = loc->mbuf;
3483 		nxlen = rte_pktmbuf_data_len(mbuf);
3484 		if (unlikely(nxlen <= txq->inlen_send)) {
3485 			/* We can inline first mbuf at least. */
3486 			if (nxlen < inlen) {
3487 				unsigned int smlen;
3488 
3489 				/* Scan mbufs till inlen filled. */
3490 				do {
3491 					smlen = nxlen;
3492 					mbuf = NEXT(mbuf);
3493 					MLX5_ASSERT(mbuf);
3494 					nxlen = rte_pktmbuf_data_len(mbuf);
3495 					nxlen += smlen;
3496 				} while (unlikely(nxlen < inlen));
3497 				if (unlikely(nxlen > txq->inlen_send)) {
3498 					/* We cannot inline entire mbuf. */
3499 					smlen = inlen - smlen;
3500 					start = rte_pktmbuf_mtod_offset
3501 						    (mbuf, uintptr_t, smlen);
3502 					goto do_align;
3503 				}
3504 			}
3505 			do {
3506 				inlen = nxlen;
3507 				mbuf = NEXT(mbuf);
3508 				/* There should be not end of packet. */
3509 				MLX5_ASSERT(mbuf);
3510 				nxlen = inlen + rte_pktmbuf_data_len(mbuf);
3511 			} while (unlikely(nxlen < txq->inlen_send));
3512 		}
3513 		start = rte_pktmbuf_mtod(mbuf, uintptr_t);
3514 		/*
3515 		 * Check whether we can do inline to align start
3516 		 * address of data buffer to cacheline.
3517 		 */
3518 do_align:
3519 		start = (~start + 1) & (RTE_CACHE_LINE_SIZE - 1);
3520 		if (unlikely(start)) {
3521 			start += inlen;
3522 			if (start <= txq->inlen_send)
3523 				inlen = start;
3524 		}
3525 	}
3526 	/*
3527 	 * Check whether there are enough free WQEBBs:
3528 	 * - Control Segment
3529 	 * - Ethernet Segment
3530 	 * - First Segment of inlined Ethernet data
3531 	 * - ... data continued ...
3532 	 * - Data Segments of pointer/min inline type
3533 	 *
3534 	 * Estimate the number of Data Segments conservatively,
3535 	 * supposing no any mbufs is being freed during inlining.
3536 	 */
3537 	MLX5_ASSERT(inlen <= txq->inlen_send);
3538 	ds = NB_SEGS(loc->mbuf) + 2 + (inlen -
3539 				       MLX5_ESEG_MIN_INLINE_SIZE +
3540 				       MLX5_WSEG_SIZE +
3541 				       MLX5_WSEG_SIZE - 1) / MLX5_WSEG_SIZE;
3542 	if (unlikely(loc->wqe_free < ((ds + 3) / 4)))
3543 		return MLX5_TXCMP_CODE_EXIT;
3544 	/* Check for maximal WQE size. */
3545 	if (unlikely((MLX5_WQE_SIZE_MAX / MLX5_WSEG_SIZE) < ((ds + 3) / 4)))
3546 		return MLX5_TXCMP_CODE_ERROR;
3547 #ifdef MLX5_PMD_SOFT_COUNTERS
3548 	/* Update sent data bytes/packets counters. */
3549 	txq->stats.obytes += dlen + vlan;
3550 #endif
3551 	wqe = txq->wqes + (txq->wqe_ci & txq->wqe_m);
3552 	loc->wqe_last = wqe;
3553 	mlx5_tx_cseg_init(txq, loc, wqe, 0, MLX5_OPCODE_SEND, olx);
3554 	ds = mlx5_tx_mseg_build(txq, loc, wqe, vlan, inlen, 0, olx);
3555 	wqe->cseg.sq_ds = rte_cpu_to_be_32(txq->qp_num_8s | ds);
3556 	txq->wqe_ci += (ds + 3) / 4;
3557 	loc->wqe_free -= (ds + 3) / 4;
3558 	return MLX5_TXCMP_CODE_MULTI;
3559 }
3560 
3561 /**
3562  * Tx burst function for multi-segment packets. Supports all
3563  * types of Tx offloads, uses MLX5_OPCODE_SEND/TSO to build WQEs,
3564  * sends one packet per WQE. Function stops sending if it
3565  * encounters the single-segment packet.
3566  *
3567  * This routine is responsible for storing processed mbuf
3568  * into elts ring buffer and update elts_head.
3569  *
3570  * @param txq
3571  *   Pointer to TX queue structure.
3572  * @param[in] pkts
3573  *   Packets to transmit.
3574  * @param pkts_n
3575  *   Number of packets in array.
3576  * @param loc
3577  *   Pointer to burst routine local context.
3578  * @param olx
3579  *   Configured Tx offloads mask. It is fully defined at
3580  *   compile time and may be used for optimization.
3581  *
3582  * @return
3583  *   MLX5_TXCMP_CODE_EXIT - sending is done or impossible.
3584  *   MLX5_TXCMP_CODE_ERROR - some unrecoverable error occurred.
3585  *   MLX5_TXCMP_CODE_SINGLE - single-segment packet encountered.
3586  *   MLX5_TXCMP_CODE_TSO - TSO single-segment packet encountered.
3587  * Local context variables updated.
3588  */
3589 static __rte_always_inline enum mlx5_txcmp_code
3590 mlx5_tx_burst_mseg(struct mlx5_txq_data *restrict txq,
3591 		   struct rte_mbuf **restrict pkts,
3592 		   unsigned int pkts_n,
3593 		   struct mlx5_txq_local *restrict loc,
3594 		   unsigned int olx)
3595 {
3596 	MLX5_ASSERT(loc->elts_free && loc->wqe_free);
3597 	MLX5_ASSERT(pkts_n > loc->pkts_sent);
3598 	pkts += loc->pkts_sent + 1;
3599 	pkts_n -= loc->pkts_sent;
3600 	for (;;) {
3601 		enum mlx5_txcmp_code ret;
3602 
3603 		MLX5_ASSERT(NB_SEGS(loc->mbuf) > 1);
3604 		/*
3605 		 * Estimate the number of free elts quickly but
3606 		 * conservatively. Some segment may be fully inlined
3607 		 * and freed, ignore this here - precise estimation
3608 		 * is costly.
3609 		 */
3610 		if (loc->elts_free < NB_SEGS(loc->mbuf))
3611 			return MLX5_TXCMP_CODE_EXIT;
3612 		if (MLX5_TXOFF_CONFIG(TSO) &&
3613 		    unlikely(loc->mbuf->ol_flags & PKT_TX_TCP_SEG)) {
3614 			/* Proceed with multi-segment TSO. */
3615 			ret = mlx5_tx_packet_multi_tso(txq, loc, olx);
3616 		} else if (MLX5_TXOFF_CONFIG(INLINE)) {
3617 			/* Proceed with multi-segment SEND with inlining. */
3618 			ret = mlx5_tx_packet_multi_inline(txq, loc, olx);
3619 		} else {
3620 			/* Proceed with multi-segment SEND w/o inlining. */
3621 			ret = mlx5_tx_packet_multi_send(txq, loc, olx);
3622 		}
3623 		if (ret == MLX5_TXCMP_CODE_EXIT)
3624 			return MLX5_TXCMP_CODE_EXIT;
3625 		if (ret == MLX5_TXCMP_CODE_ERROR)
3626 			return MLX5_TXCMP_CODE_ERROR;
3627 		/* WQE is built, go to the next packet. */
3628 		++loc->pkts_sent;
3629 		--pkts_n;
3630 		if (unlikely(!pkts_n || !loc->elts_free || !loc->wqe_free))
3631 			return MLX5_TXCMP_CODE_EXIT;
3632 		loc->mbuf = *pkts++;
3633 		if (pkts_n > 1)
3634 			rte_prefetch0(*pkts);
3635 		if (likely(NB_SEGS(loc->mbuf) > 1))
3636 			continue;
3637 		/* Here ends the series of multi-segment packets. */
3638 		if (MLX5_TXOFF_CONFIG(TSO) &&
3639 		    unlikely(loc->mbuf->ol_flags & PKT_TX_TCP_SEG))
3640 			return MLX5_TXCMP_CODE_TSO;
3641 		return MLX5_TXCMP_CODE_SINGLE;
3642 	}
3643 	MLX5_ASSERT(false);
3644 }
3645 
3646 /**
3647  * Tx burst function for single-segment packets with TSO.
3648  * Supports all types of Tx offloads, except multi-packets.
3649  * Uses MLX5_OPCODE_TSO to build WQEs, sends one packet per WQE.
3650  * Function stops sending if it encounters the multi-segment
3651  * packet or packet without TSO requested.
3652  *
3653  * The routine is responsible for storing processed mbuf
3654  * into elts ring buffer and update elts_head if inline
3655  * offloads is requested due to possible early freeing
3656  * of the inlined mbufs (can not store pkts array in elts
3657  * as a batch).
3658  *
3659  * @param txq
3660  *   Pointer to TX queue structure.
3661  * @param[in] pkts
3662  *   Packets to transmit.
3663  * @param pkts_n
3664  *   Number of packets in array.
3665  * @param loc
3666  *   Pointer to burst routine local context.
3667  * @param olx
3668  *   Configured Tx offloads mask. It is fully defined at
3669  *   compile time and may be used for optimization.
3670  *
3671  * @return
3672  *   MLX5_TXCMP_CODE_EXIT - sending is done or impossible.
3673  *   MLX5_TXCMP_CODE_ERROR - some unrecoverable error occurred.
3674  *   MLX5_TXCMP_CODE_SINGLE - single-segment packet encountered.
3675  *   MLX5_TXCMP_CODE_MULTI - multi-segment packet encountered.
3676  * Local context variables updated.
3677  */
3678 static __rte_always_inline enum mlx5_txcmp_code
3679 mlx5_tx_burst_tso(struct mlx5_txq_data *restrict txq,
3680 		  struct rte_mbuf **restrict pkts,
3681 		  unsigned int pkts_n,
3682 		  struct mlx5_txq_local *restrict loc,
3683 		  unsigned int olx)
3684 {
3685 	MLX5_ASSERT(loc->elts_free && loc->wqe_free);
3686 	MLX5_ASSERT(pkts_n > loc->pkts_sent);
3687 	pkts += loc->pkts_sent + 1;
3688 	pkts_n -= loc->pkts_sent;
3689 	for (;;) {
3690 		struct mlx5_wqe_dseg *restrict dseg;
3691 		struct mlx5_wqe *restrict wqe;
3692 		unsigned int ds, dlen, hlen, ntcp, vlan = 0;
3693 		uint8_t *dptr;
3694 
3695 		MLX5_ASSERT(NB_SEGS(loc->mbuf) == 1);
3696 		dlen = rte_pktmbuf_data_len(loc->mbuf);
3697 		if (MLX5_TXOFF_CONFIG(VLAN) &&
3698 		    loc->mbuf->ol_flags & PKT_TX_VLAN_PKT) {
3699 			vlan = sizeof(struct rte_vlan_hdr);
3700 		}
3701 		/*
3702 		 * First calculate the WQE size to check
3703 		 * whether we have enough space in ring buffer.
3704 		 */
3705 		hlen = loc->mbuf->l2_len + vlan +
3706 		       loc->mbuf->l3_len + loc->mbuf->l4_len;
3707 		if (unlikely((!hlen || !loc->mbuf->tso_segsz)))
3708 			return MLX5_TXCMP_CODE_ERROR;
3709 		if (loc->mbuf->ol_flags & PKT_TX_TUNNEL_MASK)
3710 			hlen += loc->mbuf->outer_l2_len +
3711 				loc->mbuf->outer_l3_len;
3712 		/* Segment must contain all TSO headers. */
3713 		if (unlikely(hlen > MLX5_MAX_TSO_HEADER ||
3714 			     hlen <= MLX5_ESEG_MIN_INLINE_SIZE ||
3715 			     hlen > (dlen + vlan)))
3716 			return MLX5_TXCMP_CODE_ERROR;
3717 		/*
3718 		 * Check whether there are enough free WQEBBs:
3719 		 * - Control Segment
3720 		 * - Ethernet Segment
3721 		 * - First Segment of inlined Ethernet data
3722 		 * - ... data continued ...
3723 		 * - Finishing Data Segment of pointer type
3724 		 */
3725 		ds = 4 + (hlen - MLX5_ESEG_MIN_INLINE_SIZE +
3726 			  MLX5_WSEG_SIZE - 1) / MLX5_WSEG_SIZE;
3727 		if (loc->wqe_free < ((ds + 3) / 4))
3728 			return MLX5_TXCMP_CODE_EXIT;
3729 #ifdef MLX5_PMD_SOFT_COUNTERS
3730 		/* Update sent data bytes/packets counters. */
3731 		ntcp = (dlen + vlan - hlen +
3732 			loc->mbuf->tso_segsz - 1) /
3733 			loc->mbuf->tso_segsz;
3734 		/*
3735 		 * One will be added for mbuf itself at the end
3736 		 * of the mlx5_tx_burst from loc->pkts_sent field.
3737 		 */
3738 		--ntcp;
3739 		txq->stats.opackets += ntcp;
3740 		txq->stats.obytes += dlen + vlan + ntcp * hlen;
3741 #endif
3742 		/*
3743 		 * Build the TSO WQE:
3744 		 * - Control Segment
3745 		 * - Ethernet Segment with hlen bytes inlined
3746 		 * - Data Segment of pointer type
3747 		 */
3748 		wqe = txq->wqes + (txq->wqe_ci & txq->wqe_m);
3749 		loc->wqe_last = wqe;
3750 		mlx5_tx_cseg_init(txq, loc, wqe, ds,
3751 				  MLX5_OPCODE_TSO, olx);
3752 		dseg = mlx5_tx_eseg_data(txq, loc, wqe, vlan, hlen, 1, olx);
3753 		dptr = rte_pktmbuf_mtod(loc->mbuf, uint8_t *) + hlen - vlan;
3754 		dlen -= hlen - vlan;
3755 		mlx5_tx_dseg_ptr(txq, loc, dseg, dptr, dlen, olx);
3756 		/*
3757 		 * WQE is built, update the loop parameters
3758 		 * and go to the next packet.
3759 		 */
3760 		txq->wqe_ci += (ds + 3) / 4;
3761 		loc->wqe_free -= (ds + 3) / 4;
3762 		if (MLX5_TXOFF_CONFIG(INLINE))
3763 			txq->elts[txq->elts_head++ & txq->elts_m] = loc->mbuf;
3764 		--loc->elts_free;
3765 		++loc->pkts_sent;
3766 		--pkts_n;
3767 		if (unlikely(!pkts_n || !loc->elts_free || !loc->wqe_free))
3768 			return MLX5_TXCMP_CODE_EXIT;
3769 		loc->mbuf = *pkts++;
3770 		if (pkts_n > 1)
3771 			rte_prefetch0(*pkts);
3772 		if (MLX5_TXOFF_CONFIG(MULTI) &&
3773 		    unlikely(NB_SEGS(loc->mbuf) > 1))
3774 			return MLX5_TXCMP_CODE_MULTI;
3775 		if (likely(!(loc->mbuf->ol_flags & PKT_TX_TCP_SEG)))
3776 			return MLX5_TXCMP_CODE_SINGLE;
3777 		/* Continue with the next TSO packet. */
3778 	}
3779 	MLX5_ASSERT(false);
3780 }
3781 
3782 /**
3783  * Analyze the packet and select the best method to send.
3784  *
3785  * @param txq
3786  *   Pointer to TX queue structure.
3787  * @param loc
3788  *   Pointer to burst routine local context.
3789  * @param olx
3790  *   Configured Tx offloads mask. It is fully defined at
3791  *   compile time and may be used for optimization.
3792  * @param newp
3793  *   The predefined flag whether do complete check for
3794  *   multi-segment packets and TSO.
3795  *
3796  * @return
3797  *  MLX5_TXCMP_CODE_MULTI - multi-segment packet encountered.
3798  *  MLX5_TXCMP_CODE_TSO - TSO required, use TSO/LSO.
3799  *  MLX5_TXCMP_CODE_SINGLE - single-segment packet, use SEND.
3800  *  MLX5_TXCMP_CODE_EMPW - single-segment packet, use MPW.
3801  */
3802 static __rte_always_inline enum mlx5_txcmp_code
3803 mlx5_tx_able_to_empw(struct mlx5_txq_data *restrict txq,
3804 		     struct mlx5_txq_local *restrict loc,
3805 		     unsigned int olx,
3806 		     bool newp)
3807 {
3808 	/* Check for multi-segment packet. */
3809 	if (newp &&
3810 	    MLX5_TXOFF_CONFIG(MULTI) &&
3811 	    unlikely(NB_SEGS(loc->mbuf) > 1))
3812 		return MLX5_TXCMP_CODE_MULTI;
3813 	/* Check for TSO packet. */
3814 	if (newp &&
3815 	    MLX5_TXOFF_CONFIG(TSO) &&
3816 	    unlikely(loc->mbuf->ol_flags & PKT_TX_TCP_SEG))
3817 		return MLX5_TXCMP_CODE_TSO;
3818 	/* Check if eMPW is enabled at all. */
3819 	if (!MLX5_TXOFF_CONFIG(EMPW))
3820 		return MLX5_TXCMP_CODE_SINGLE;
3821 	/* Check if eMPW can be engaged. */
3822 	if (MLX5_TXOFF_CONFIG(VLAN) &&
3823 	    unlikely(loc->mbuf->ol_flags & PKT_TX_VLAN_PKT) &&
3824 		(!MLX5_TXOFF_CONFIG(INLINE) ||
3825 		 unlikely((rte_pktmbuf_data_len(loc->mbuf) +
3826 			   sizeof(struct rte_vlan_hdr)) > txq->inlen_empw))) {
3827 		/*
3828 		 * eMPW does not support VLAN insertion offload,
3829 		 * we have to inline the entire packet but
3830 		 * packet is too long for inlining.
3831 		 */
3832 		return MLX5_TXCMP_CODE_SINGLE;
3833 	}
3834 	return MLX5_TXCMP_CODE_EMPW;
3835 }
3836 
3837 /**
3838  * Check the next packet attributes to match with the eMPW batch ones.
3839  * In addition, for legacy MPW the packet length is checked either.
3840  *
3841  * @param txq
3842  *   Pointer to TX queue structure.
3843  * @param es
3844  *   Pointer to Ethernet Segment of eMPW batch.
3845  * @param loc
3846  *   Pointer to burst routine local context.
3847  * @param dlen
3848  *   Length of previous packet in MPW descriptor.
3849  * @param olx
3850  *   Configured Tx offloads mask. It is fully defined at
3851  *   compile time and may be used for optimization.
3852  *
3853  * @return
3854  *  true - packet match with eMPW batch attributes.
3855  *  false - no match, eMPW should be restarted.
3856  */
3857 static __rte_always_inline bool
3858 mlx5_tx_match_empw(struct mlx5_txq_data *restrict txq __rte_unused,
3859 		   struct mlx5_wqe_eseg *restrict es,
3860 		   struct mlx5_txq_local *restrict loc,
3861 		   uint32_t dlen,
3862 		   unsigned int olx)
3863 {
3864 	uint8_t swp_flags = 0;
3865 
3866 	/* Compare the checksum flags, if any. */
3867 	if (MLX5_TXOFF_CONFIG(CSUM) &&
3868 	    txq_ol_cksum_to_cs(loc->mbuf) != es->cs_flags)
3869 		return false;
3870 	/* Compare the Software Parser offsets and flags. */
3871 	if (MLX5_TXOFF_CONFIG(SWP) &&
3872 	    (es->swp_offs != txq_mbuf_to_swp(loc, &swp_flags, olx) ||
3873 	     es->swp_flags != swp_flags))
3874 		return false;
3875 	/* Fill metadata field if needed. */
3876 	if (MLX5_TXOFF_CONFIG(METADATA) &&
3877 		es->metadata != (loc->mbuf->ol_flags & PKT_TX_DYNF_METADATA ?
3878 				 *RTE_FLOW_DYNF_METADATA(loc->mbuf) : 0))
3879 		return false;
3880 	/* Legacy MPW can send packets with the same lengt only. */
3881 	if (MLX5_TXOFF_CONFIG(MPW) &&
3882 	    dlen != rte_pktmbuf_data_len(loc->mbuf))
3883 		return false;
3884 	/* There must be no VLAN packets in eMPW loop. */
3885 	if (MLX5_TXOFF_CONFIG(VLAN))
3886 		MLX5_ASSERT(!(loc->mbuf->ol_flags & PKT_TX_VLAN_PKT));
3887 	return true;
3888 }
3889 
3890 /*
3891  * Update send loop variables and WQE for eMPW loop
3892  * without data inlining. Number of Data Segments is
3893  * equal to the number of sent packets.
3894  *
3895  * @param txq
3896  *   Pointer to TX queue structure.
3897  * @param loc
3898  *   Pointer to burst routine local context.
3899  * @param ds
3900  *   Number of packets/Data Segments/Packets.
3901  * @param slen
3902  *   Accumulated statistics, bytes sent
3903  * @param olx
3904  *   Configured Tx offloads mask. It is fully defined at
3905  *   compile time and may be used for optimization.
3906  *
3907  * @return
3908  *  true - packet match with eMPW batch attributes.
3909  *  false - no match, eMPW should be restarted.
3910  */
3911 static __rte_always_inline void
3912 mlx5_tx_sdone_empw(struct mlx5_txq_data *restrict txq,
3913 		   struct mlx5_txq_local *restrict loc,
3914 		   unsigned int ds,
3915 		   unsigned int slen,
3916 		   unsigned int olx __rte_unused)
3917 {
3918 	MLX5_ASSERT(!MLX5_TXOFF_CONFIG(INLINE));
3919 #ifdef MLX5_PMD_SOFT_COUNTERS
3920 	/* Update sent data bytes counter. */
3921 	 txq->stats.obytes += slen;
3922 #else
3923 	(void)slen;
3924 #endif
3925 	loc->elts_free -= ds;
3926 	loc->pkts_sent += ds;
3927 	ds += 2;
3928 	loc->wqe_last->cseg.sq_ds = rte_cpu_to_be_32(txq->qp_num_8s | ds);
3929 	txq->wqe_ci += (ds + 3) / 4;
3930 	loc->wqe_free -= (ds + 3) / 4;
3931 }
3932 
3933 /*
3934  * Update send loop variables and WQE for eMPW loop
3935  * with data inlining. Gets the size of pushed descriptors
3936  * and data to the WQE.
3937  *
3938  * @param txq
3939  *   Pointer to TX queue structure.
3940  * @param loc
3941  *   Pointer to burst routine local context.
3942  * @param len
3943  *   Total size of descriptor/data in bytes.
3944  * @param slen
3945  *   Accumulated statistics, data bytes sent.
3946  * @param wqem
3947  *   The base WQE for the eMPW/MPW descriptor.
3948  * @param olx
3949  *   Configured Tx offloads mask. It is fully defined at
3950  *   compile time and may be used for optimization.
3951  *
3952  * @return
3953  *  true - packet match with eMPW batch attributes.
3954  *  false - no match, eMPW should be restarted.
3955  */
3956 static __rte_always_inline void
3957 mlx5_tx_idone_empw(struct mlx5_txq_data *restrict txq,
3958 		   struct mlx5_txq_local *restrict loc,
3959 		   unsigned int len,
3960 		   unsigned int slen,
3961 		   struct mlx5_wqe *restrict wqem,
3962 		   unsigned int olx __rte_unused)
3963 {
3964 	struct mlx5_wqe_dseg *dseg = &wqem->dseg[0];
3965 
3966 	MLX5_ASSERT(MLX5_TXOFF_CONFIG(INLINE));
3967 #ifdef MLX5_PMD_SOFT_COUNTERS
3968 	/* Update sent data bytes counter. */
3969 	 txq->stats.obytes += slen;
3970 #else
3971 	(void)slen;
3972 #endif
3973 	if (MLX5_TXOFF_CONFIG(MPW) && dseg->bcount == RTE_BE32(0)) {
3974 		/*
3975 		 * If the legacy MPW session contains the inline packets
3976 		 * we should set the only inline data segment length
3977 		 * and align the total length to the segment size.
3978 		 */
3979 		MLX5_ASSERT(len > sizeof(dseg->bcount));
3980 		dseg->bcount = rte_cpu_to_be_32((len - sizeof(dseg->bcount)) |
3981 						MLX5_ETH_WQE_DATA_INLINE);
3982 		len = (len + MLX5_WSEG_SIZE - 1) / MLX5_WSEG_SIZE + 2;
3983 	} else {
3984 		/*
3985 		 * The session is not legacy MPW or contains the
3986 		 * data buffer pointer segments.
3987 		 */
3988 		MLX5_ASSERT((len % MLX5_WSEG_SIZE) == 0);
3989 		len = len / MLX5_WSEG_SIZE + 2;
3990 	}
3991 	wqem->cseg.sq_ds = rte_cpu_to_be_32(txq->qp_num_8s | len);
3992 	txq->wqe_ci += (len + 3) / 4;
3993 	loc->wqe_free -= (len + 3) / 4;
3994 	loc->wqe_last = wqem;
3995 }
3996 
3997 /**
3998  * The set of Tx burst functions for single-segment packets
3999  * without TSO and with Multi-Packet Writing feature support.
4000  * Supports all types of Tx offloads, except multi-packets
4001  * and TSO.
4002  *
4003  * Uses MLX5_OPCODE_EMPW to build WQEs if possible and sends
4004  * as many packet per WQE as it can. If eMPW is not configured
4005  * or packet can not be sent with eMPW (VLAN insertion) the
4006  * ordinary SEND opcode is used and only one packet placed
4007  * in WQE.
4008  *
4009  * Functions stop sending if it encounters the multi-segment
4010  * packet or packet with TSO requested.
4011  *
4012  * The routines are responsible for storing processed mbuf
4013  * into elts ring buffer and update elts_head if inlining
4014  * offload is requested. Otherwise the copying mbufs to elts
4015  * can be postponed and completed at the end of burst routine.
4016  *
4017  * @param txq
4018  *   Pointer to TX queue structure.
4019  * @param[in] pkts
4020  *   Packets to transmit.
4021  * @param pkts_n
4022  *   Number of packets in array.
4023  * @param loc
4024  *   Pointer to burst routine local context.
4025  * @param olx
4026  *   Configured Tx offloads mask. It is fully defined at
4027  *   compile time and may be used for optimization.
4028  *
4029  * @return
4030  *   MLX5_TXCMP_CODE_EXIT - sending is done or impossible.
4031  *   MLX5_TXCMP_CODE_ERROR - some unrecoverable error occurred.
4032  *   MLX5_TXCMP_CODE_MULTI - multi-segment packet encountered.
4033  *   MLX5_TXCMP_CODE_TSO - TSO packet encountered.
4034  *   MLX5_TXCMP_CODE_SINGLE - used inside functions set.
4035  *   MLX5_TXCMP_CODE_EMPW - used inside functions set.
4036  *
4037  * Local context variables updated.
4038  *
4039  *
4040  * The routine sends packets with MLX5_OPCODE_EMPW
4041  * without inlining, this is dedicated optimized branch.
4042  * No VLAN insertion is supported.
4043  */
4044 static __rte_always_inline enum mlx5_txcmp_code
4045 mlx5_tx_burst_empw_simple(struct mlx5_txq_data *restrict txq,
4046 			  struct rte_mbuf **restrict pkts,
4047 			  unsigned int pkts_n,
4048 			  struct mlx5_txq_local *restrict loc,
4049 			  unsigned int olx)
4050 {
4051 	/*
4052 	 * Subroutine is the part of mlx5_tx_burst_single()
4053 	 * and sends single-segment packet with eMPW opcode
4054 	 * without data inlining.
4055 	 */
4056 	MLX5_ASSERT(!MLX5_TXOFF_CONFIG(INLINE));
4057 	MLX5_ASSERT(MLX5_TXOFF_CONFIG(EMPW));
4058 	MLX5_ASSERT(loc->elts_free && loc->wqe_free);
4059 	MLX5_ASSERT(pkts_n > loc->pkts_sent);
4060 	static_assert(MLX5_EMPW_MIN_PACKETS >= 2, "invalid min size");
4061 	pkts += loc->pkts_sent + 1;
4062 	pkts_n -= loc->pkts_sent;
4063 	for (;;) {
4064 		struct mlx5_wqe_dseg *restrict dseg;
4065 		struct mlx5_wqe_eseg *restrict eseg;
4066 		enum mlx5_txcmp_code ret;
4067 		unsigned int part, loop;
4068 		unsigned int slen = 0;
4069 
4070 next_empw:
4071 		MLX5_ASSERT(NB_SEGS(loc->mbuf) == 1);
4072 		part = RTE_MIN(pkts_n, MLX5_TXOFF_CONFIG(MPW) ?
4073 				       MLX5_MPW_MAX_PACKETS :
4074 				       MLX5_EMPW_MAX_PACKETS);
4075 		if (unlikely(loc->elts_free < part)) {
4076 			/* We have no enough elts to save all mbufs. */
4077 			if (unlikely(loc->elts_free < MLX5_EMPW_MIN_PACKETS))
4078 				return MLX5_TXCMP_CODE_EXIT;
4079 			/* But we still able to send at least minimal eMPW. */
4080 			part = loc->elts_free;
4081 		}
4082 		/* Check whether we have enough WQEs */
4083 		if (unlikely(loc->wqe_free < ((2 + part + 3) / 4))) {
4084 			if (unlikely(loc->wqe_free <
4085 				((2 + MLX5_EMPW_MIN_PACKETS + 3) / 4)))
4086 				return MLX5_TXCMP_CODE_EXIT;
4087 			part = (loc->wqe_free * 4) - 2;
4088 		}
4089 		if (likely(part > 1))
4090 			rte_prefetch0(*pkts);
4091 		loc->wqe_last = txq->wqes + (txq->wqe_ci & txq->wqe_m);
4092 		/*
4093 		 * Build eMPW title WQEBB:
4094 		 * - Control Segment, eMPW opcode
4095 		 * - Ethernet Segment, no inline
4096 		 */
4097 		mlx5_tx_cseg_init(txq, loc, loc->wqe_last, part + 2,
4098 				  MLX5_OPCODE_ENHANCED_MPSW, olx);
4099 		mlx5_tx_eseg_none(txq, loc, loc->wqe_last,
4100 				  olx & ~MLX5_TXOFF_CONFIG_VLAN);
4101 		eseg = &loc->wqe_last->eseg;
4102 		dseg = &loc->wqe_last->dseg[0];
4103 		loop = part;
4104 		/* Store the packet length for legacy MPW. */
4105 		if (MLX5_TXOFF_CONFIG(MPW))
4106 			eseg->mss = rte_cpu_to_be_16
4107 					(rte_pktmbuf_data_len(loc->mbuf));
4108 		for (;;) {
4109 			uint32_t dlen = rte_pktmbuf_data_len(loc->mbuf);
4110 #ifdef MLX5_PMD_SOFT_COUNTERS
4111 			/* Update sent data bytes counter. */
4112 			slen += dlen;
4113 #endif
4114 			mlx5_tx_dseg_ptr
4115 				(txq, loc, dseg,
4116 				 rte_pktmbuf_mtod(loc->mbuf, uint8_t *),
4117 				 dlen, olx);
4118 			if (unlikely(--loop == 0))
4119 				break;
4120 			loc->mbuf = *pkts++;
4121 			if (likely(loop > 1))
4122 				rte_prefetch0(*pkts);
4123 			ret = mlx5_tx_able_to_empw(txq, loc, olx, true);
4124 			/*
4125 			 * Unroll the completion code to avoid
4126 			 * returning variable value - it results in
4127 			 * unoptimized sequent checking in caller.
4128 			 */
4129 			if (ret == MLX5_TXCMP_CODE_MULTI) {
4130 				part -= loop;
4131 				mlx5_tx_sdone_empw(txq, loc, part, slen, olx);
4132 				if (unlikely(!loc->elts_free ||
4133 					     !loc->wqe_free))
4134 					return MLX5_TXCMP_CODE_EXIT;
4135 				return MLX5_TXCMP_CODE_MULTI;
4136 			}
4137 			MLX5_ASSERT(NB_SEGS(loc->mbuf) == 1);
4138 			if (ret == MLX5_TXCMP_CODE_TSO) {
4139 				part -= loop;
4140 				mlx5_tx_sdone_empw(txq, loc, part, slen, olx);
4141 				if (unlikely(!loc->elts_free ||
4142 					     !loc->wqe_free))
4143 					return MLX5_TXCMP_CODE_EXIT;
4144 				return MLX5_TXCMP_CODE_TSO;
4145 			}
4146 			if (ret == MLX5_TXCMP_CODE_SINGLE) {
4147 				part -= loop;
4148 				mlx5_tx_sdone_empw(txq, loc, part, slen, olx);
4149 				if (unlikely(!loc->elts_free ||
4150 					     !loc->wqe_free))
4151 					return MLX5_TXCMP_CODE_EXIT;
4152 				return MLX5_TXCMP_CODE_SINGLE;
4153 			}
4154 			if (ret != MLX5_TXCMP_CODE_EMPW) {
4155 				MLX5_ASSERT(false);
4156 				part -= loop;
4157 				mlx5_tx_sdone_empw(txq, loc, part, slen, olx);
4158 				return MLX5_TXCMP_CODE_ERROR;
4159 			}
4160 			/*
4161 			 * Check whether packet parameters coincide
4162 			 * within assumed eMPW batch:
4163 			 * - check sum settings
4164 			 * - metadata value
4165 			 * - software parser settings
4166 			 * - packets length (legacy MPW only)
4167 			 */
4168 			if (!mlx5_tx_match_empw(txq, eseg, loc, dlen, olx)) {
4169 				MLX5_ASSERT(loop);
4170 				part -= loop;
4171 				mlx5_tx_sdone_empw(txq, loc, part, slen, olx);
4172 				if (unlikely(!loc->elts_free ||
4173 					     !loc->wqe_free))
4174 					return MLX5_TXCMP_CODE_EXIT;
4175 				pkts_n -= part;
4176 				goto next_empw;
4177 			}
4178 			/* Packet attributes match, continue the same eMPW. */
4179 			++dseg;
4180 			if ((uintptr_t)dseg >= (uintptr_t)txq->wqes_end)
4181 				dseg = (struct mlx5_wqe_dseg *)txq->wqes;
4182 		}
4183 		/* eMPW is built successfully, update loop parameters. */
4184 		MLX5_ASSERT(!loop);
4185 		MLX5_ASSERT(pkts_n >= part);
4186 #ifdef MLX5_PMD_SOFT_COUNTERS
4187 		/* Update sent data bytes counter. */
4188 		txq->stats.obytes += slen;
4189 #endif
4190 		loc->elts_free -= part;
4191 		loc->pkts_sent += part;
4192 		txq->wqe_ci += (2 + part + 3) / 4;
4193 		loc->wqe_free -= (2 + part + 3) / 4;
4194 		pkts_n -= part;
4195 		if (unlikely(!pkts_n || !loc->elts_free || !loc->wqe_free))
4196 			return MLX5_TXCMP_CODE_EXIT;
4197 		loc->mbuf = *pkts++;
4198 		ret = mlx5_tx_able_to_empw(txq, loc, olx, true);
4199 		if (unlikely(ret != MLX5_TXCMP_CODE_EMPW))
4200 			return ret;
4201 		/* Continue sending eMPW batches. */
4202 	}
4203 	MLX5_ASSERT(false);
4204 }
4205 
4206 /**
4207  * The routine sends packets with MLX5_OPCODE_EMPW
4208  * with inlining, optionally supports VLAN insertion.
4209  */
4210 static __rte_always_inline enum mlx5_txcmp_code
4211 mlx5_tx_burst_empw_inline(struct mlx5_txq_data *restrict txq,
4212 			  struct rte_mbuf **restrict pkts,
4213 			  unsigned int pkts_n,
4214 			  struct mlx5_txq_local *restrict loc,
4215 			  unsigned int olx)
4216 {
4217 	/*
4218 	 * Subroutine is the part of mlx5_tx_burst_single()
4219 	 * and sends single-segment packet with eMPW opcode
4220 	 * with data inlining.
4221 	 */
4222 	MLX5_ASSERT(MLX5_TXOFF_CONFIG(INLINE));
4223 	MLX5_ASSERT(MLX5_TXOFF_CONFIG(EMPW));
4224 	MLX5_ASSERT(loc->elts_free && loc->wqe_free);
4225 	MLX5_ASSERT(pkts_n > loc->pkts_sent);
4226 	static_assert(MLX5_EMPW_MIN_PACKETS >= 2, "invalid min size");
4227 	pkts += loc->pkts_sent + 1;
4228 	pkts_n -= loc->pkts_sent;
4229 	for (;;) {
4230 		struct mlx5_wqe_dseg *restrict dseg;
4231 		struct mlx5_wqe *restrict wqem;
4232 		enum mlx5_txcmp_code ret;
4233 		unsigned int room, part, nlim;
4234 		unsigned int slen = 0;
4235 
4236 		MLX5_ASSERT(NB_SEGS(loc->mbuf) == 1);
4237 		/*
4238 		 * Limits the amount of packets in one WQE
4239 		 * to improve CQE latency generation.
4240 		 */
4241 		nlim = RTE_MIN(pkts_n, MLX5_TXOFF_CONFIG(MPW) ?
4242 				       MLX5_MPW_INLINE_MAX_PACKETS :
4243 				       MLX5_EMPW_MAX_PACKETS);
4244 		/* Check whether we have minimal amount WQEs */
4245 		if (unlikely(loc->wqe_free <
4246 			    ((2 + MLX5_EMPW_MIN_PACKETS + 3) / 4)))
4247 			return MLX5_TXCMP_CODE_EXIT;
4248 		if (likely(pkts_n > 1))
4249 			rte_prefetch0(*pkts);
4250 		wqem = txq->wqes + (txq->wqe_ci & txq->wqe_m);
4251 		/*
4252 		 * Build eMPW title WQEBB:
4253 		 * - Control Segment, eMPW opcode, zero DS
4254 		 * - Ethernet Segment, no inline
4255 		 */
4256 		mlx5_tx_cseg_init(txq, loc, wqem, 0,
4257 				  MLX5_OPCODE_ENHANCED_MPSW, olx);
4258 		mlx5_tx_eseg_none(txq, loc, wqem,
4259 				  olx & ~MLX5_TXOFF_CONFIG_VLAN);
4260 		dseg = &wqem->dseg[0];
4261 		/* Store the packet length for legacy MPW. */
4262 		if (MLX5_TXOFF_CONFIG(MPW))
4263 			wqem->eseg.mss = rte_cpu_to_be_16
4264 					 (rte_pktmbuf_data_len(loc->mbuf));
4265 		room = RTE_MIN(MLX5_WQE_SIZE_MAX / MLX5_WQE_SIZE,
4266 			       loc->wqe_free) * MLX5_WQE_SIZE -
4267 					MLX5_WQE_CSEG_SIZE -
4268 					MLX5_WQE_ESEG_SIZE;
4269 		/* Limit the room for legacy MPW sessions for performance. */
4270 		if (MLX5_TXOFF_CONFIG(MPW))
4271 			room = RTE_MIN(room,
4272 				       RTE_MAX(txq->inlen_empw +
4273 					       sizeof(dseg->bcount) +
4274 					       (MLX5_TXOFF_CONFIG(VLAN) ?
4275 					       sizeof(struct rte_vlan_hdr) : 0),
4276 					       MLX5_MPW_INLINE_MAX_PACKETS *
4277 					       MLX5_WQE_DSEG_SIZE));
4278 		/* Build WQE till we have space, packets and resources. */
4279 		part = room;
4280 		for (;;) {
4281 			uint32_t dlen = rte_pktmbuf_data_len(loc->mbuf);
4282 			uint8_t *dptr = rte_pktmbuf_mtod(loc->mbuf, uint8_t *);
4283 			unsigned int tlen;
4284 
4285 			MLX5_ASSERT(room >= MLX5_WQE_DSEG_SIZE);
4286 			MLX5_ASSERT((room % MLX5_WQE_DSEG_SIZE) == 0);
4287 			MLX5_ASSERT((uintptr_t)dseg < (uintptr_t)txq->wqes_end);
4288 			/*
4289 			 * Some Tx offloads may cause an error if
4290 			 * packet is not long enough, check against
4291 			 * assumed minimal length.
4292 			 */
4293 			if (unlikely(dlen <= MLX5_ESEG_MIN_INLINE_SIZE)) {
4294 				part -= room;
4295 				if (unlikely(!part))
4296 					return MLX5_TXCMP_CODE_ERROR;
4297 				/*
4298 				 * We have some successfully built
4299 				 * packet Data Segments to send.
4300 				 */
4301 				mlx5_tx_idone_empw(txq, loc, part,
4302 						   slen, wqem, olx);
4303 				return MLX5_TXCMP_CODE_ERROR;
4304 			}
4305 			/* Inline or not inline - that's the Question. */
4306 			if (dlen > txq->inlen_empw ||
4307 			    loc->mbuf->ol_flags & PKT_TX_DYNF_NOINLINE)
4308 				goto pointer_empw;
4309 			if (MLX5_TXOFF_CONFIG(MPW)) {
4310 				if (dlen > txq->inlen_send)
4311 					goto pointer_empw;
4312 				tlen = dlen;
4313 				if (part == room) {
4314 					/* Open new inline MPW session. */
4315 					tlen += sizeof(dseg->bcount);
4316 					dseg->bcount = RTE_BE32(0);
4317 					dseg = RTE_PTR_ADD
4318 						(dseg, sizeof(dseg->bcount));
4319 				} else {
4320 					/*
4321 					 * No pointer and inline descriptor
4322 					 * intermix for legacy MPW sessions.
4323 					 */
4324 					if (wqem->dseg[0].bcount)
4325 						break;
4326 				}
4327 			} else {
4328 				tlen = sizeof(dseg->bcount) + dlen;
4329 			}
4330 			/* Inline entire packet, optional VLAN insertion. */
4331 			if (MLX5_TXOFF_CONFIG(VLAN) &&
4332 			    loc->mbuf->ol_flags & PKT_TX_VLAN_PKT) {
4333 				/*
4334 				 * The packet length must be checked in
4335 				 * mlx5_tx_able_to_empw() and packet
4336 				 * fits into inline length guaranteed.
4337 				 */
4338 				MLX5_ASSERT((dlen +
4339 					     sizeof(struct rte_vlan_hdr)) <=
4340 					    txq->inlen_empw);
4341 				tlen += sizeof(struct rte_vlan_hdr);
4342 				if (room < tlen)
4343 					break;
4344 				dseg = mlx5_tx_dseg_vlan(txq, loc, dseg,
4345 							 dptr, dlen, olx);
4346 #ifdef MLX5_PMD_SOFT_COUNTERS
4347 				/* Update sent data bytes counter. */
4348 				slen +=	sizeof(struct rte_vlan_hdr);
4349 #endif
4350 			} else {
4351 				if (room < tlen)
4352 					break;
4353 				dseg = mlx5_tx_dseg_empw(txq, loc, dseg,
4354 							 dptr, dlen, olx);
4355 			}
4356 			if (!MLX5_TXOFF_CONFIG(MPW))
4357 				tlen = RTE_ALIGN(tlen, MLX5_WSEG_SIZE);
4358 			MLX5_ASSERT(room >= tlen);
4359 			room -= tlen;
4360 			/*
4361 			 * Packet data are completely inlined,
4362 			 * free the packet immediately.
4363 			 */
4364 			rte_pktmbuf_free_seg(loc->mbuf);
4365 			goto next_mbuf;
4366 pointer_empw:
4367 			/*
4368 			 * No pointer and inline descriptor
4369 			 * intermix for legacy MPW sessions.
4370 			 */
4371 			if (MLX5_TXOFF_CONFIG(MPW) &&
4372 			    part != room &&
4373 			    wqem->dseg[0].bcount == RTE_BE32(0))
4374 				break;
4375 			/*
4376 			 * Not inlinable VLAN packets are
4377 			 * proceeded outside of this routine.
4378 			 */
4379 			MLX5_ASSERT(room >= MLX5_WQE_DSEG_SIZE);
4380 			if (MLX5_TXOFF_CONFIG(VLAN))
4381 				MLX5_ASSERT(!(loc->mbuf->ol_flags &
4382 					    PKT_TX_VLAN_PKT));
4383 			mlx5_tx_dseg_ptr(txq, loc, dseg, dptr, dlen, olx);
4384 			/* We have to store mbuf in elts.*/
4385 			txq->elts[txq->elts_head++ & txq->elts_m] = loc->mbuf;
4386 			room -= MLX5_WQE_DSEG_SIZE;
4387 			/* Ring buffer wraparound is checked at the loop end.*/
4388 			++dseg;
4389 next_mbuf:
4390 #ifdef MLX5_PMD_SOFT_COUNTERS
4391 			/* Update sent data bytes counter. */
4392 			slen += dlen;
4393 #endif
4394 			loc->pkts_sent++;
4395 			loc->elts_free--;
4396 			pkts_n--;
4397 			if (unlikely(!pkts_n || !loc->elts_free)) {
4398 				/*
4399 				 * We have no resources/packets to
4400 				 * continue build descriptors.
4401 				 */
4402 				part -= room;
4403 				mlx5_tx_idone_empw(txq, loc, part,
4404 						   slen, wqem, olx);
4405 				return MLX5_TXCMP_CODE_EXIT;
4406 			}
4407 			loc->mbuf = *pkts++;
4408 			if (likely(pkts_n > 1))
4409 				rte_prefetch0(*pkts);
4410 			ret = mlx5_tx_able_to_empw(txq, loc, olx, true);
4411 			/*
4412 			 * Unroll the completion code to avoid
4413 			 * returning variable value - it results in
4414 			 * unoptimized sequent checking in caller.
4415 			 */
4416 			if (ret == MLX5_TXCMP_CODE_MULTI) {
4417 				part -= room;
4418 				mlx5_tx_idone_empw(txq, loc, part,
4419 						   slen, wqem, olx);
4420 				if (unlikely(!loc->elts_free ||
4421 					     !loc->wqe_free))
4422 					return MLX5_TXCMP_CODE_EXIT;
4423 				return MLX5_TXCMP_CODE_MULTI;
4424 			}
4425 			MLX5_ASSERT(NB_SEGS(loc->mbuf) == 1);
4426 			if (ret == MLX5_TXCMP_CODE_TSO) {
4427 				part -= room;
4428 				mlx5_tx_idone_empw(txq, loc, part,
4429 						   slen, wqem, olx);
4430 				if (unlikely(!loc->elts_free ||
4431 					     !loc->wqe_free))
4432 					return MLX5_TXCMP_CODE_EXIT;
4433 				return MLX5_TXCMP_CODE_TSO;
4434 			}
4435 			if (ret == MLX5_TXCMP_CODE_SINGLE) {
4436 				part -= room;
4437 				mlx5_tx_idone_empw(txq, loc, part,
4438 						   slen, wqem, olx);
4439 				if (unlikely(!loc->elts_free ||
4440 					     !loc->wqe_free))
4441 					return MLX5_TXCMP_CODE_EXIT;
4442 				return MLX5_TXCMP_CODE_SINGLE;
4443 			}
4444 			if (ret != MLX5_TXCMP_CODE_EMPW) {
4445 				MLX5_ASSERT(false);
4446 				part -= room;
4447 				mlx5_tx_idone_empw(txq, loc, part,
4448 						   slen, wqem, olx);
4449 				return MLX5_TXCMP_CODE_ERROR;
4450 			}
4451 			/* Check if we have minimal room left. */
4452 			nlim--;
4453 			if (unlikely(!nlim || room < MLX5_WQE_DSEG_SIZE))
4454 				break;
4455 			/*
4456 			 * Check whether packet parameters coincide
4457 			 * within assumed eMPW batch:
4458 			 * - check sum settings
4459 			 * - metadata value
4460 			 * - software parser settings
4461 			 * - packets length (legacy MPW only)
4462 			 */
4463 			if (!mlx5_tx_match_empw(txq, &wqem->eseg,
4464 						loc, dlen, olx))
4465 				break;
4466 			/* Packet attributes match, continue the same eMPW. */
4467 			if ((uintptr_t)dseg >= (uintptr_t)txq->wqes_end)
4468 				dseg = (struct mlx5_wqe_dseg *)txq->wqes;
4469 		}
4470 		/*
4471 		 * We get here to close an existing eMPW
4472 		 * session and start the new one.
4473 		 */
4474 		MLX5_ASSERT(pkts_n);
4475 		part -= room;
4476 		if (unlikely(!part))
4477 			return MLX5_TXCMP_CODE_EXIT;
4478 		mlx5_tx_idone_empw(txq, loc, part, slen, wqem, olx);
4479 		if (unlikely(!loc->elts_free ||
4480 			     !loc->wqe_free))
4481 			return MLX5_TXCMP_CODE_EXIT;
4482 		/* Continue the loop with new eMPW session. */
4483 	}
4484 	MLX5_ASSERT(false);
4485 }
4486 
4487 /**
4488  * The routine sends packets with ordinary MLX5_OPCODE_SEND.
4489  * Data inlining and VLAN insertion are supported.
4490  */
4491 static __rte_always_inline enum mlx5_txcmp_code
4492 mlx5_tx_burst_single_send(struct mlx5_txq_data *restrict txq,
4493 			  struct rte_mbuf **restrict pkts,
4494 			  unsigned int pkts_n,
4495 			  struct mlx5_txq_local *restrict loc,
4496 			  unsigned int olx)
4497 {
4498 	/*
4499 	 * Subroutine is the part of mlx5_tx_burst_single()
4500 	 * and sends single-segment packet with SEND opcode.
4501 	 */
4502 	MLX5_ASSERT(loc->elts_free && loc->wqe_free);
4503 	MLX5_ASSERT(pkts_n > loc->pkts_sent);
4504 	pkts += loc->pkts_sent + 1;
4505 	pkts_n -= loc->pkts_sent;
4506 	for (;;) {
4507 		struct mlx5_wqe *restrict wqe;
4508 		enum mlx5_txcmp_code ret;
4509 
4510 		MLX5_ASSERT(NB_SEGS(loc->mbuf) == 1);
4511 		if (MLX5_TXOFF_CONFIG(INLINE)) {
4512 			unsigned int inlen, vlan = 0;
4513 
4514 			inlen = rte_pktmbuf_data_len(loc->mbuf);
4515 			if (MLX5_TXOFF_CONFIG(VLAN) &&
4516 			    loc->mbuf->ol_flags & PKT_TX_VLAN_PKT) {
4517 				vlan = sizeof(struct rte_vlan_hdr);
4518 				inlen += vlan;
4519 				static_assert((sizeof(struct rte_vlan_hdr) +
4520 					       sizeof(struct rte_ether_hdr)) ==
4521 					       MLX5_ESEG_MIN_INLINE_SIZE,
4522 					       "invalid min inline data size");
4523 			}
4524 			/*
4525 			 * If inlining is enabled at configuration time
4526 			 * the limit must be not less than minimal size.
4527 			 * Otherwise we would do extra check for data
4528 			 * size to avoid crashes due to length overflow.
4529 			 */
4530 			MLX5_ASSERT(txq->inlen_send >=
4531 				    MLX5_ESEG_MIN_INLINE_SIZE);
4532 			if (inlen <= txq->inlen_send) {
4533 				unsigned int seg_n, wqe_n;
4534 
4535 				rte_prefetch0(rte_pktmbuf_mtod
4536 						(loc->mbuf, uint8_t *));
4537 				/* Check against minimal length. */
4538 				if (inlen <= MLX5_ESEG_MIN_INLINE_SIZE)
4539 					return MLX5_TXCMP_CODE_ERROR;
4540 				if (loc->mbuf->ol_flags &
4541 				    PKT_TX_DYNF_NOINLINE) {
4542 					/*
4543 					 * The hint flag not to inline packet
4544 					 * data is set. Check whether we can
4545 					 * follow the hint.
4546 					 */
4547 					if ((!MLX5_TXOFF_CONFIG(EMPW) &&
4548 					      txq->inlen_mode) ||
4549 					    (MLX5_TXOFF_CONFIG(MPW) &&
4550 					     txq->inlen_mode)) {
4551 						/*
4552 						 * The hardware requires the
4553 						 * minimal inline data header.
4554 						 */
4555 						goto single_min_inline;
4556 					}
4557 					if (MLX5_TXOFF_CONFIG(VLAN) &&
4558 					    vlan && !txq->vlan_en) {
4559 						/*
4560 						 * We must insert VLAN tag
4561 						 * by software means.
4562 						 */
4563 						goto single_part_inline;
4564 					}
4565 					goto single_no_inline;
4566 				}
4567 				/*
4568 				 * Completely inlined packet data WQE:
4569 				 * - Control Segment, SEND opcode
4570 				 * - Ethernet Segment, no VLAN insertion
4571 				 * - Data inlined, VLAN optionally inserted
4572 				 * - Alignment to MLX5_WSEG_SIZE
4573 				 * Have to estimate amount of WQEBBs
4574 				 */
4575 				seg_n = (inlen + 3 * MLX5_WSEG_SIZE -
4576 					 MLX5_ESEG_MIN_INLINE_SIZE +
4577 					 MLX5_WSEG_SIZE - 1) / MLX5_WSEG_SIZE;
4578 				/* Check if there are enough WQEBBs. */
4579 				wqe_n = (seg_n + 3) / 4;
4580 				if (wqe_n > loc->wqe_free)
4581 					return MLX5_TXCMP_CODE_EXIT;
4582 				wqe = txq->wqes + (txq->wqe_ci & txq->wqe_m);
4583 				loc->wqe_last = wqe;
4584 				mlx5_tx_cseg_init(txq, loc, wqe, seg_n,
4585 						  MLX5_OPCODE_SEND, olx);
4586 				mlx5_tx_eseg_data(txq, loc, wqe,
4587 						  vlan, inlen, 0, olx);
4588 				txq->wqe_ci += wqe_n;
4589 				loc->wqe_free -= wqe_n;
4590 				/*
4591 				 * Packet data are completely inlined,
4592 				 * free the packet immediately.
4593 				 */
4594 				rte_pktmbuf_free_seg(loc->mbuf);
4595 			} else if ((!MLX5_TXOFF_CONFIG(EMPW) ||
4596 				     MLX5_TXOFF_CONFIG(MPW)) &&
4597 					txq->inlen_mode) {
4598 				/*
4599 				 * If minimal inlining is requested the eMPW
4600 				 * feature should be disabled due to data is
4601 				 * inlined into Ethernet Segment, which can
4602 				 * not contain inlined data for eMPW due to
4603 				 * segment shared for all packets.
4604 				 */
4605 				struct mlx5_wqe_dseg *restrict dseg;
4606 				unsigned int ds;
4607 				uint8_t *dptr;
4608 
4609 				/*
4610 				 * The inline-mode settings require
4611 				 * to inline the specified amount of
4612 				 * data bytes to the Ethernet Segment.
4613 				 * We should check the free space in
4614 				 * WQE ring buffer to inline partially.
4615 				 */
4616 single_min_inline:
4617 				MLX5_ASSERT(txq->inlen_send >= txq->inlen_mode);
4618 				MLX5_ASSERT(inlen > txq->inlen_mode);
4619 				MLX5_ASSERT(txq->inlen_mode >=
4620 					    MLX5_ESEG_MIN_INLINE_SIZE);
4621 				/*
4622 				 * Check whether there are enough free WQEBBs:
4623 				 * - Control Segment
4624 				 * - Ethernet Segment
4625 				 * - First Segment of inlined Ethernet data
4626 				 * - ... data continued ...
4627 				 * - Finishing Data Segment of pointer type
4628 				 */
4629 				ds = (MLX5_WQE_CSEG_SIZE +
4630 				      MLX5_WQE_ESEG_SIZE +
4631 				      MLX5_WQE_DSEG_SIZE +
4632 				      txq->inlen_mode -
4633 				      MLX5_ESEG_MIN_INLINE_SIZE +
4634 				      MLX5_WQE_DSEG_SIZE +
4635 				      MLX5_WSEG_SIZE - 1) / MLX5_WSEG_SIZE;
4636 				if (loc->wqe_free < ((ds + 3) / 4))
4637 					return MLX5_TXCMP_CODE_EXIT;
4638 				/*
4639 				 * Build the ordinary SEND WQE:
4640 				 * - Control Segment
4641 				 * - Ethernet Segment, inline inlen_mode bytes
4642 				 * - Data Segment of pointer type
4643 				 */
4644 				wqe = txq->wqes + (txq->wqe_ci & txq->wqe_m);
4645 				loc->wqe_last = wqe;
4646 				mlx5_tx_cseg_init(txq, loc, wqe, ds,
4647 						  MLX5_OPCODE_SEND, olx);
4648 				dseg = mlx5_tx_eseg_data(txq, loc, wqe, vlan,
4649 							 txq->inlen_mode,
4650 							 0, olx);
4651 				dptr = rte_pktmbuf_mtod(loc->mbuf, uint8_t *) +
4652 				       txq->inlen_mode - vlan;
4653 				inlen -= txq->inlen_mode;
4654 				mlx5_tx_dseg_ptr(txq, loc, dseg,
4655 						 dptr, inlen, olx);
4656 				/*
4657 				 * WQE is built, update the loop parameters
4658 				 * and got to the next packet.
4659 				 */
4660 				txq->wqe_ci += (ds + 3) / 4;
4661 				loc->wqe_free -= (ds + 3) / 4;
4662 				/* We have to store mbuf in elts.*/
4663 				MLX5_ASSERT(MLX5_TXOFF_CONFIG(INLINE));
4664 				txq->elts[txq->elts_head++ & txq->elts_m] =
4665 						loc->mbuf;
4666 				--loc->elts_free;
4667 			} else {
4668 				uint8_t *dptr;
4669 				unsigned int dlen;
4670 
4671 				/*
4672 				 * Partially inlined packet data WQE, we have
4673 				 * some space in title WQEBB, we can fill it
4674 				 * with some packet data. It takes one WQEBB,
4675 				 * it is available, no extra space check:
4676 				 * - Control Segment, SEND opcode
4677 				 * - Ethernet Segment, no VLAN insertion
4678 				 * - MLX5_ESEG_MIN_INLINE_SIZE bytes of Data
4679 				 * - Data Segment, pointer type
4680 				 *
4681 				 * We also get here if VLAN insertion is not
4682 				 * supported by HW, the inline is enabled.
4683 				 */
4684 single_part_inline:
4685 				wqe = txq->wqes + (txq->wqe_ci & txq->wqe_m);
4686 				loc->wqe_last = wqe;
4687 				mlx5_tx_cseg_init(txq, loc, wqe, 4,
4688 						  MLX5_OPCODE_SEND, olx);
4689 				mlx5_tx_eseg_dmin(txq, loc, wqe, vlan, olx);
4690 				dptr = rte_pktmbuf_mtod(loc->mbuf, uint8_t *) +
4691 				       MLX5_ESEG_MIN_INLINE_SIZE - vlan;
4692 				/*
4693 				 * The length check is performed above, by
4694 				 * comparing with txq->inlen_send. We should
4695 				 * not get overflow here.
4696 				 */
4697 				MLX5_ASSERT(inlen > MLX5_ESEG_MIN_INLINE_SIZE);
4698 				dlen = inlen - MLX5_ESEG_MIN_INLINE_SIZE;
4699 				mlx5_tx_dseg_ptr(txq, loc, &wqe->dseg[1],
4700 						 dptr, dlen, olx);
4701 				++txq->wqe_ci;
4702 				--loc->wqe_free;
4703 				/* We have to store mbuf in elts.*/
4704 				MLX5_ASSERT(MLX5_TXOFF_CONFIG(INLINE));
4705 				txq->elts[txq->elts_head++ & txq->elts_m] =
4706 						loc->mbuf;
4707 				--loc->elts_free;
4708 			}
4709 #ifdef MLX5_PMD_SOFT_COUNTERS
4710 			/* Update sent data bytes counter. */
4711 			txq->stats.obytes += vlan +
4712 					rte_pktmbuf_data_len(loc->mbuf);
4713 #endif
4714 		} else {
4715 			/*
4716 			 * No inline at all, it means the CPU cycles saving
4717 			 * is prioritized at configuration, we should not
4718 			 * copy any packet data to WQE.
4719 			 *
4720 			 * SEND WQE, one WQEBB:
4721 			 * - Control Segment, SEND opcode
4722 			 * - Ethernet Segment, optional VLAN, no inline
4723 			 * - Data Segment, pointer type
4724 			 */
4725 single_no_inline:
4726 			wqe = txq->wqes + (txq->wqe_ci & txq->wqe_m);
4727 			loc->wqe_last = wqe;
4728 			mlx5_tx_cseg_init(txq, loc, wqe, 3,
4729 					  MLX5_OPCODE_SEND, olx);
4730 			mlx5_tx_eseg_none(txq, loc, wqe, olx);
4731 			mlx5_tx_dseg_ptr
4732 				(txq, loc, &wqe->dseg[0],
4733 				 rte_pktmbuf_mtod(loc->mbuf, uint8_t *),
4734 				 rte_pktmbuf_data_len(loc->mbuf), olx);
4735 			++txq->wqe_ci;
4736 			--loc->wqe_free;
4737 			/*
4738 			 * We should not store mbuf pointer in elts
4739 			 * if no inlining is configured, this is done
4740 			 * by calling routine in a batch copy.
4741 			 */
4742 			MLX5_ASSERT(!MLX5_TXOFF_CONFIG(INLINE));
4743 			--loc->elts_free;
4744 #ifdef MLX5_PMD_SOFT_COUNTERS
4745 			/* Update sent data bytes counter. */
4746 			txq->stats.obytes += rte_pktmbuf_data_len(loc->mbuf);
4747 			if (MLX5_TXOFF_CONFIG(VLAN) &&
4748 			    loc->mbuf->ol_flags & PKT_TX_VLAN_PKT)
4749 				txq->stats.obytes +=
4750 					sizeof(struct rte_vlan_hdr);
4751 #endif
4752 		}
4753 		++loc->pkts_sent;
4754 		--pkts_n;
4755 		if (unlikely(!pkts_n || !loc->elts_free || !loc->wqe_free))
4756 			return MLX5_TXCMP_CODE_EXIT;
4757 		loc->mbuf = *pkts++;
4758 		if (pkts_n > 1)
4759 			rte_prefetch0(*pkts);
4760 		ret = mlx5_tx_able_to_empw(txq, loc, olx, true);
4761 		if (unlikely(ret != MLX5_TXCMP_CODE_SINGLE))
4762 			return ret;
4763 	}
4764 	MLX5_ASSERT(false);
4765 }
4766 
4767 static __rte_always_inline enum mlx5_txcmp_code
4768 mlx5_tx_burst_single(struct mlx5_txq_data *restrict txq,
4769 		     struct rte_mbuf **restrict pkts,
4770 		     unsigned int pkts_n,
4771 		     struct mlx5_txq_local *restrict loc,
4772 		     unsigned int olx)
4773 {
4774 	enum mlx5_txcmp_code ret;
4775 
4776 	ret = mlx5_tx_able_to_empw(txq, loc, olx, false);
4777 	if (ret == MLX5_TXCMP_CODE_SINGLE)
4778 		goto ordinary_send;
4779 	MLX5_ASSERT(ret == MLX5_TXCMP_CODE_EMPW);
4780 	for (;;) {
4781 		/* Optimize for inline/no inline eMPW send. */
4782 		ret = (MLX5_TXOFF_CONFIG(INLINE)) ?
4783 			mlx5_tx_burst_empw_inline
4784 				(txq, pkts, pkts_n, loc, olx) :
4785 			mlx5_tx_burst_empw_simple
4786 				(txq, pkts, pkts_n, loc, olx);
4787 		if (ret != MLX5_TXCMP_CODE_SINGLE)
4788 			return ret;
4789 		/* The resources to send one packet should remain. */
4790 		MLX5_ASSERT(loc->elts_free && loc->wqe_free);
4791 ordinary_send:
4792 		ret = mlx5_tx_burst_single_send(txq, pkts, pkts_n, loc, olx);
4793 		MLX5_ASSERT(ret != MLX5_TXCMP_CODE_SINGLE);
4794 		if (ret != MLX5_TXCMP_CODE_EMPW)
4795 			return ret;
4796 		/* The resources to send one packet should remain. */
4797 		MLX5_ASSERT(loc->elts_free && loc->wqe_free);
4798 	}
4799 }
4800 
4801 /**
4802  * DPDK Tx callback template. This is configured template
4803  * used to generate routines optimized for specified offload setup.
4804  * One of this generated functions is chosen at SQ configuration
4805  * time.
4806  *
4807  * @param txq
4808  *   Generic pointer to TX queue structure.
4809  * @param[in] pkts
4810  *   Packets to transmit.
4811  * @param pkts_n
4812  *   Number of packets in array.
4813  * @param olx
4814  *   Configured offloads mask, presents the bits of MLX5_TXOFF_CONFIG_xxx
4815  *   values. Should be static to take compile time static configuration
4816  *   advantages.
4817  *
4818  * @return
4819  *   Number of packets successfully transmitted (<= pkts_n).
4820  */
4821 static __rte_always_inline uint16_t
4822 mlx5_tx_burst_tmpl(struct mlx5_txq_data *restrict txq,
4823 		   struct rte_mbuf **restrict pkts,
4824 		   uint16_t pkts_n,
4825 		   unsigned int olx)
4826 {
4827 	struct mlx5_txq_local loc;
4828 	enum mlx5_txcmp_code ret;
4829 	unsigned int part;
4830 
4831 	MLX5_ASSERT(txq->elts_s >= (uint16_t)(txq->elts_head - txq->elts_tail));
4832 	MLX5_ASSERT(txq->wqe_s >= (uint16_t)(txq->wqe_ci - txq->wqe_pi));
4833 	if (unlikely(!pkts_n))
4834 		return 0;
4835 	loc.pkts_sent = 0;
4836 	loc.pkts_copy = 0;
4837 	loc.wqe_last = NULL;
4838 
4839 send_loop:
4840 	loc.pkts_loop = loc.pkts_sent;
4841 	/*
4842 	 * Check if there are some CQEs, if any:
4843 	 * - process an encountered errors
4844 	 * - process the completed WQEs
4845 	 * - free related mbufs
4846 	 * - doorbell the NIC about processed CQEs
4847 	 */
4848 	rte_prefetch0(*(pkts + loc.pkts_sent));
4849 	mlx5_tx_handle_completion(txq, olx);
4850 	/*
4851 	 * Calculate the number of available resources - elts and WQEs.
4852 	 * There are two possible different scenarios:
4853 	 * - no data inlining into WQEs, one WQEBB may contains up to
4854 	 *   four packets, in this case elts become scarce resource
4855 	 * - data inlining into WQEs, one packet may require multiple
4856 	 *   WQEBBs, the WQEs become the limiting factor.
4857 	 */
4858 	MLX5_ASSERT(txq->elts_s >= (uint16_t)(txq->elts_head - txq->elts_tail));
4859 	loc.elts_free = txq->elts_s -
4860 				(uint16_t)(txq->elts_head - txq->elts_tail);
4861 	MLX5_ASSERT(txq->wqe_s >= (uint16_t)(txq->wqe_ci - txq->wqe_pi));
4862 	loc.wqe_free = txq->wqe_s -
4863 				(uint16_t)(txq->wqe_ci - txq->wqe_pi);
4864 	if (unlikely(!loc.elts_free || !loc.wqe_free))
4865 		goto burst_exit;
4866 	for (;;) {
4867 		/*
4868 		 * Fetch the packet from array. Usually this is
4869 		 * the first packet in series of multi/single
4870 		 * segment packets.
4871 		 */
4872 		loc.mbuf = *(pkts + loc.pkts_sent);
4873 		/* Dedicated branch for multi-segment packets. */
4874 		if (MLX5_TXOFF_CONFIG(MULTI) &&
4875 		    unlikely(NB_SEGS(loc.mbuf) > 1)) {
4876 			/*
4877 			 * Multi-segment packet encountered.
4878 			 * Hardware is able to process it only
4879 			 * with SEND/TSO opcodes, one packet
4880 			 * per WQE, do it in dedicated routine.
4881 			 */
4882 enter_send_multi:
4883 			MLX5_ASSERT(loc.pkts_sent >= loc.pkts_copy);
4884 			part = loc.pkts_sent - loc.pkts_copy;
4885 			if (!MLX5_TXOFF_CONFIG(INLINE) && part) {
4886 				/*
4887 				 * There are some single-segment mbufs not
4888 				 * stored in elts. The mbufs must be in the
4889 				 * same order as WQEs, so we must copy the
4890 				 * mbufs to elts here, before the coming
4891 				 * multi-segment packet mbufs is appended.
4892 				 */
4893 				mlx5_tx_copy_elts(txq, pkts + loc.pkts_copy,
4894 						  part, olx);
4895 				loc.pkts_copy = loc.pkts_sent;
4896 			}
4897 			MLX5_ASSERT(pkts_n > loc.pkts_sent);
4898 			ret = mlx5_tx_burst_mseg(txq, pkts, pkts_n, &loc, olx);
4899 			if (!MLX5_TXOFF_CONFIG(INLINE))
4900 				loc.pkts_copy = loc.pkts_sent;
4901 			/*
4902 			 * These returned code checks are supposed
4903 			 * to be optimized out due to routine inlining.
4904 			 */
4905 			if (ret == MLX5_TXCMP_CODE_EXIT) {
4906 				/*
4907 				 * The routine returns this code when
4908 				 * all packets are sent or there is no
4909 				 * enough resources to complete request.
4910 				 */
4911 				break;
4912 			}
4913 			if (ret == MLX5_TXCMP_CODE_ERROR) {
4914 				/*
4915 				 * The routine returns this code when
4916 				 * some error in the incoming packets
4917 				 * format occurred.
4918 				 */
4919 				txq->stats.oerrors++;
4920 				break;
4921 			}
4922 			if (ret == MLX5_TXCMP_CODE_SINGLE) {
4923 				/*
4924 				 * The single-segment packet was encountered
4925 				 * in the array, try to send it with the
4926 				 * best optimized way, possible engaging eMPW.
4927 				 */
4928 				goto enter_send_single;
4929 			}
4930 			if (MLX5_TXOFF_CONFIG(TSO) &&
4931 			    ret == MLX5_TXCMP_CODE_TSO) {
4932 				/*
4933 				 * The single-segment TSO packet was
4934 				 * encountered in the array.
4935 				 */
4936 				goto enter_send_tso;
4937 			}
4938 			/* We must not get here. Something is going wrong. */
4939 			MLX5_ASSERT(false);
4940 			txq->stats.oerrors++;
4941 			break;
4942 		}
4943 		/* Dedicated branch for single-segment TSO packets. */
4944 		if (MLX5_TXOFF_CONFIG(TSO) &&
4945 		    unlikely(loc.mbuf->ol_flags & PKT_TX_TCP_SEG)) {
4946 			/*
4947 			 * TSO might require special way for inlining
4948 			 * (dedicated parameters) and is sent with
4949 			 * MLX5_OPCODE_TSO opcode only, provide this
4950 			 * in dedicated branch.
4951 			 */
4952 enter_send_tso:
4953 			MLX5_ASSERT(NB_SEGS(loc.mbuf) == 1);
4954 			MLX5_ASSERT(pkts_n > loc.pkts_sent);
4955 			ret = mlx5_tx_burst_tso(txq, pkts, pkts_n, &loc, olx);
4956 			/*
4957 			 * These returned code checks are supposed
4958 			 * to be optimized out due to routine inlining.
4959 			 */
4960 			if (ret == MLX5_TXCMP_CODE_EXIT)
4961 				break;
4962 			if (ret == MLX5_TXCMP_CODE_ERROR) {
4963 				txq->stats.oerrors++;
4964 				break;
4965 			}
4966 			if (ret == MLX5_TXCMP_CODE_SINGLE)
4967 				goto enter_send_single;
4968 			if (MLX5_TXOFF_CONFIG(MULTI) &&
4969 			    ret == MLX5_TXCMP_CODE_MULTI) {
4970 				/*
4971 				 * The multi-segment packet was
4972 				 * encountered in the array.
4973 				 */
4974 				goto enter_send_multi;
4975 			}
4976 			/* We must not get here. Something is going wrong. */
4977 			MLX5_ASSERT(false);
4978 			txq->stats.oerrors++;
4979 			break;
4980 		}
4981 		/*
4982 		 * The dedicated branch for the single-segment packets
4983 		 * without TSO. Often these ones can be sent using
4984 		 * MLX5_OPCODE_EMPW with multiple packets in one WQE.
4985 		 * The routine builds the WQEs till it encounters
4986 		 * the TSO or multi-segment packet (in case if these
4987 		 * offloads are requested at SQ configuration time).
4988 		 */
4989 enter_send_single:
4990 		MLX5_ASSERT(pkts_n > loc.pkts_sent);
4991 		ret = mlx5_tx_burst_single(txq, pkts, pkts_n, &loc, olx);
4992 		/*
4993 		 * These returned code checks are supposed
4994 		 * to be optimized out due to routine inlining.
4995 		 */
4996 		if (ret == MLX5_TXCMP_CODE_EXIT)
4997 			break;
4998 		if (ret == MLX5_TXCMP_CODE_ERROR) {
4999 			txq->stats.oerrors++;
5000 			break;
5001 		}
5002 		if (MLX5_TXOFF_CONFIG(MULTI) &&
5003 		    ret == MLX5_TXCMP_CODE_MULTI) {
5004 			/*
5005 			 * The multi-segment packet was
5006 			 * encountered in the array.
5007 			 */
5008 			goto enter_send_multi;
5009 		}
5010 		if (MLX5_TXOFF_CONFIG(TSO) &&
5011 		    ret == MLX5_TXCMP_CODE_TSO) {
5012 			/*
5013 			 * The single-segment TSO packet was
5014 			 * encountered in the array.
5015 			 */
5016 			goto enter_send_tso;
5017 		}
5018 		/* We must not get here. Something is going wrong. */
5019 		MLX5_ASSERT(false);
5020 		txq->stats.oerrors++;
5021 		break;
5022 	}
5023 	/*
5024 	 * Main Tx loop is completed, do the rest:
5025 	 * - set completion request if thresholds are reached
5026 	 * - doorbell the hardware
5027 	 * - copy the rest of mbufs to elts (if any)
5028 	 */
5029 	MLX5_ASSERT(MLX5_TXOFF_CONFIG(INLINE) ||
5030 		    loc.pkts_sent >= loc.pkts_copy);
5031 	/* Take a shortcut if nothing is sent. */
5032 	if (unlikely(loc.pkts_sent == loc.pkts_loop))
5033 		goto burst_exit;
5034 	/* Request CQE generation if limits are reached. */
5035 	mlx5_tx_request_completion(txq, &loc, olx);
5036 	/*
5037 	 * Ring QP doorbell immediately after WQE building completion
5038 	 * to improve latencies. The pure software related data treatment
5039 	 * can be completed after doorbell. Tx CQEs for this SQ are
5040 	 * processed in this thread only by the polling.
5041 	 *
5042 	 * The rdma core library can map doorbell register in two ways,
5043 	 * depending on the environment variable "MLX5_SHUT_UP_BF":
5044 	 *
5045 	 * - as regular cached memory, the variable is either missing or
5046 	 *   set to zero. This type of mapping may cause the significant
5047 	 *   doorbell register writing latency and requires explicit
5048 	 *   memory write barrier to mitigate this issue and prevent
5049 	 *   write combining.
5050 	 *
5051 	 * - as non-cached memory, the variable is present and set to
5052 	 *   not "0" value. This type of mapping may cause performance
5053 	 *   impact under heavy loading conditions but the explicit write
5054 	 *   memory barrier is not required and it may improve core
5055 	 *   performance.
5056 	 *
5057 	 * - the legacy behaviour (prior 19.08 release) was to use some
5058 	 *   heuristics to decide whether write memory barrier should
5059 	 *   be performed. This behavior is supported with specifying
5060 	 *   tx_db_nc=2, write barrier is skipped if application
5061 	 *   provides the full recommended burst of packets, it
5062 	 *   supposes the next packets are coming and the write barrier
5063 	 *   will be issued on the next burst (after descriptor writing,
5064 	 *   at least).
5065 	 */
5066 	mlx5_tx_dbrec_cond_wmb(txq, loc.wqe_last, !txq->db_nc &&
5067 			(!txq->db_heu || pkts_n % MLX5_TX_DEFAULT_BURST));
5068 	/* Not all of the mbufs may be stored into elts yet. */
5069 	part = MLX5_TXOFF_CONFIG(INLINE) ? 0 : loc.pkts_sent - loc.pkts_copy;
5070 	if (!MLX5_TXOFF_CONFIG(INLINE) && part) {
5071 		/*
5072 		 * There are some single-segment mbufs not stored in elts.
5073 		 * It can be only if the last packet was single-segment.
5074 		 * The copying is gathered into one place due to it is
5075 		 * a good opportunity to optimize that with SIMD.
5076 		 * Unfortunately if inlining is enabled the gaps in
5077 		 * pointer array may happen due to early freeing of the
5078 		 * inlined mbufs.
5079 		 */
5080 		mlx5_tx_copy_elts(txq, pkts + loc.pkts_copy, part, olx);
5081 		loc.pkts_copy = loc.pkts_sent;
5082 	}
5083 	MLX5_ASSERT(txq->elts_s >= (uint16_t)(txq->elts_head - txq->elts_tail));
5084 	MLX5_ASSERT(txq->wqe_s >= (uint16_t)(txq->wqe_ci - txq->wqe_pi));
5085 	if (pkts_n > loc.pkts_sent) {
5086 		/*
5087 		 * If burst size is large there might be no enough CQE
5088 		 * fetched from completion queue and no enough resources
5089 		 * freed to send all the packets.
5090 		 */
5091 		goto send_loop;
5092 	}
5093 burst_exit:
5094 #ifdef MLX5_PMD_SOFT_COUNTERS
5095 	/* Increment sent packets counter. */
5096 	txq->stats.opackets += loc.pkts_sent;
5097 #endif
5098 	return loc.pkts_sent;
5099 }
5100 
5101 /* Generate routines with Enhanced Multi-Packet Write support. */
5102 MLX5_TXOFF_DECL(full_empw,
5103 		MLX5_TXOFF_CONFIG_FULL | MLX5_TXOFF_CONFIG_EMPW)
5104 
5105 MLX5_TXOFF_DECL(none_empw,
5106 		MLX5_TXOFF_CONFIG_NONE | MLX5_TXOFF_CONFIG_EMPW)
5107 
5108 MLX5_TXOFF_DECL(md_empw,
5109 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
5110 
5111 MLX5_TXOFF_DECL(mt_empw,
5112 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
5113 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
5114 
5115 MLX5_TXOFF_DECL(mtsc_empw,
5116 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
5117 		MLX5_TXOFF_CONFIG_SWP |	MLX5_TXOFF_CONFIG_CSUM |
5118 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
5119 
5120 MLX5_TXOFF_DECL(mti_empw,
5121 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
5122 		MLX5_TXOFF_CONFIG_INLINE |
5123 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
5124 
5125 MLX5_TXOFF_DECL(mtv_empw,
5126 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
5127 		MLX5_TXOFF_CONFIG_VLAN |
5128 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
5129 
5130 MLX5_TXOFF_DECL(mtiv_empw,
5131 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
5132 		MLX5_TXOFF_CONFIG_INLINE | MLX5_TXOFF_CONFIG_VLAN |
5133 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
5134 
5135 MLX5_TXOFF_DECL(sc_empw,
5136 		MLX5_TXOFF_CONFIG_SWP |	MLX5_TXOFF_CONFIG_CSUM |
5137 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
5138 
5139 MLX5_TXOFF_DECL(sci_empw,
5140 		MLX5_TXOFF_CONFIG_SWP |	MLX5_TXOFF_CONFIG_CSUM |
5141 		MLX5_TXOFF_CONFIG_INLINE |
5142 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
5143 
5144 MLX5_TXOFF_DECL(scv_empw,
5145 		MLX5_TXOFF_CONFIG_SWP |	MLX5_TXOFF_CONFIG_CSUM |
5146 		MLX5_TXOFF_CONFIG_VLAN |
5147 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
5148 
5149 MLX5_TXOFF_DECL(sciv_empw,
5150 		MLX5_TXOFF_CONFIG_SWP |	MLX5_TXOFF_CONFIG_CSUM |
5151 		MLX5_TXOFF_CONFIG_INLINE | MLX5_TXOFF_CONFIG_VLAN |
5152 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
5153 
5154 MLX5_TXOFF_DECL(i_empw,
5155 		MLX5_TXOFF_CONFIG_INLINE |
5156 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
5157 
5158 MLX5_TXOFF_DECL(v_empw,
5159 		MLX5_TXOFF_CONFIG_VLAN |
5160 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
5161 
5162 MLX5_TXOFF_DECL(iv_empw,
5163 		MLX5_TXOFF_CONFIG_INLINE | MLX5_TXOFF_CONFIG_VLAN |
5164 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
5165 
5166 /* Generate routines without Enhanced Multi-Packet Write support. */
5167 MLX5_TXOFF_DECL(full,
5168 		MLX5_TXOFF_CONFIG_FULL)
5169 
5170 MLX5_TXOFF_DECL(none,
5171 		MLX5_TXOFF_CONFIG_NONE)
5172 
5173 MLX5_TXOFF_DECL(md,
5174 		MLX5_TXOFF_CONFIG_METADATA)
5175 
5176 MLX5_TXOFF_DECL(mt,
5177 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
5178 		MLX5_TXOFF_CONFIG_METADATA)
5179 
5180 MLX5_TXOFF_DECL(mtsc,
5181 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
5182 		MLX5_TXOFF_CONFIG_SWP |	MLX5_TXOFF_CONFIG_CSUM |
5183 		MLX5_TXOFF_CONFIG_METADATA)
5184 
5185 MLX5_TXOFF_DECL(mti,
5186 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
5187 		MLX5_TXOFF_CONFIG_INLINE |
5188 		MLX5_TXOFF_CONFIG_METADATA)
5189 
5190 
5191 MLX5_TXOFF_DECL(mtv,
5192 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
5193 		MLX5_TXOFF_CONFIG_VLAN |
5194 		MLX5_TXOFF_CONFIG_METADATA)
5195 
5196 
5197 MLX5_TXOFF_DECL(mtiv,
5198 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
5199 		MLX5_TXOFF_CONFIG_INLINE | MLX5_TXOFF_CONFIG_VLAN |
5200 		MLX5_TXOFF_CONFIG_METADATA)
5201 
5202 MLX5_TXOFF_DECL(sc,
5203 		MLX5_TXOFF_CONFIG_SWP |	MLX5_TXOFF_CONFIG_CSUM |
5204 		MLX5_TXOFF_CONFIG_METADATA)
5205 
5206 MLX5_TXOFF_DECL(sci,
5207 		MLX5_TXOFF_CONFIG_SWP |	MLX5_TXOFF_CONFIG_CSUM |
5208 		MLX5_TXOFF_CONFIG_INLINE |
5209 		MLX5_TXOFF_CONFIG_METADATA)
5210 
5211 
5212 MLX5_TXOFF_DECL(scv,
5213 		MLX5_TXOFF_CONFIG_SWP |	MLX5_TXOFF_CONFIG_CSUM |
5214 		MLX5_TXOFF_CONFIG_VLAN |
5215 		MLX5_TXOFF_CONFIG_METADATA)
5216 
5217 
5218 MLX5_TXOFF_DECL(sciv,
5219 		MLX5_TXOFF_CONFIG_SWP |	MLX5_TXOFF_CONFIG_CSUM |
5220 		MLX5_TXOFF_CONFIG_INLINE | MLX5_TXOFF_CONFIG_VLAN |
5221 		MLX5_TXOFF_CONFIG_METADATA)
5222 
5223 MLX5_TXOFF_DECL(i,
5224 		MLX5_TXOFF_CONFIG_INLINE |
5225 		MLX5_TXOFF_CONFIG_METADATA)
5226 
5227 MLX5_TXOFF_DECL(v,
5228 		MLX5_TXOFF_CONFIG_VLAN |
5229 		MLX5_TXOFF_CONFIG_METADATA)
5230 
5231 MLX5_TXOFF_DECL(iv,
5232 		MLX5_TXOFF_CONFIG_INLINE | MLX5_TXOFF_CONFIG_VLAN |
5233 		MLX5_TXOFF_CONFIG_METADATA)
5234 
5235 /*
5236  * Generate routines with Legacy Multi-Packet Write support.
5237  * This mode is supported by ConnectX-4 Lx only and imposes
5238  * offload limitations, not supported:
5239  *   - ACL/Flows (metadata are becoming meaningless)
5240  *   - WQE Inline headers
5241  *   - SRIOV (E-Switch offloads)
5242  *   - VLAN insertion
5243  *   - tunnel encapsulation/decapsulation
5244  *   - TSO
5245  */
5246 MLX5_TXOFF_DECL(none_mpw,
5247 		MLX5_TXOFF_CONFIG_NONE | MLX5_TXOFF_CONFIG_EMPW |
5248 		MLX5_TXOFF_CONFIG_MPW)
5249 
5250 MLX5_TXOFF_DECL(mci_mpw,
5251 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_CSUM |
5252 		MLX5_TXOFF_CONFIG_INLINE | MLX5_TXOFF_CONFIG_EMPW |
5253 		MLX5_TXOFF_CONFIG_MPW)
5254 
5255 MLX5_TXOFF_DECL(mc_mpw,
5256 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_CSUM |
5257 		MLX5_TXOFF_CONFIG_EMPW | MLX5_TXOFF_CONFIG_MPW)
5258 
5259 MLX5_TXOFF_DECL(i_mpw,
5260 		MLX5_TXOFF_CONFIG_INLINE | MLX5_TXOFF_CONFIG_EMPW |
5261 		MLX5_TXOFF_CONFIG_MPW)
5262 
5263 /*
5264  * Array of declared and compiled Tx burst function and corresponding
5265  * supported offloads set. The array is used to select the Tx burst
5266  * function for specified offloads set at Tx queue configuration time.
5267  */
5268 const struct {
5269 	eth_tx_burst_t func;
5270 	unsigned int olx;
5271 } txoff_func[] = {
5272 MLX5_TXOFF_INFO(full_empw,
5273 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
5274 		MLX5_TXOFF_CONFIG_SWP |	MLX5_TXOFF_CONFIG_CSUM |
5275 		MLX5_TXOFF_CONFIG_INLINE | MLX5_TXOFF_CONFIG_VLAN |
5276 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
5277 
5278 MLX5_TXOFF_INFO(none_empw,
5279 		MLX5_TXOFF_CONFIG_NONE | MLX5_TXOFF_CONFIG_EMPW)
5280 
5281 MLX5_TXOFF_INFO(md_empw,
5282 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
5283 
5284 MLX5_TXOFF_INFO(mt_empw,
5285 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
5286 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
5287 
5288 MLX5_TXOFF_INFO(mtsc_empw,
5289 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
5290 		MLX5_TXOFF_CONFIG_SWP |	MLX5_TXOFF_CONFIG_CSUM |
5291 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
5292 
5293 MLX5_TXOFF_INFO(mti_empw,
5294 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
5295 		MLX5_TXOFF_CONFIG_INLINE |
5296 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
5297 
5298 MLX5_TXOFF_INFO(mtv_empw,
5299 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
5300 		MLX5_TXOFF_CONFIG_VLAN |
5301 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
5302 
5303 MLX5_TXOFF_INFO(mtiv_empw,
5304 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
5305 		MLX5_TXOFF_CONFIG_INLINE | MLX5_TXOFF_CONFIG_VLAN |
5306 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
5307 
5308 MLX5_TXOFF_INFO(sc_empw,
5309 		MLX5_TXOFF_CONFIG_SWP |	MLX5_TXOFF_CONFIG_CSUM |
5310 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
5311 
5312 MLX5_TXOFF_INFO(sci_empw,
5313 		MLX5_TXOFF_CONFIG_SWP |	MLX5_TXOFF_CONFIG_CSUM |
5314 		MLX5_TXOFF_CONFIG_INLINE |
5315 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
5316 
5317 MLX5_TXOFF_INFO(scv_empw,
5318 		MLX5_TXOFF_CONFIG_SWP |	MLX5_TXOFF_CONFIG_CSUM |
5319 		MLX5_TXOFF_CONFIG_VLAN |
5320 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
5321 
5322 MLX5_TXOFF_INFO(sciv_empw,
5323 		MLX5_TXOFF_CONFIG_SWP |	MLX5_TXOFF_CONFIG_CSUM |
5324 		MLX5_TXOFF_CONFIG_INLINE | MLX5_TXOFF_CONFIG_VLAN |
5325 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
5326 
5327 MLX5_TXOFF_INFO(i_empw,
5328 		MLX5_TXOFF_CONFIG_INLINE |
5329 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
5330 
5331 MLX5_TXOFF_INFO(v_empw,
5332 		MLX5_TXOFF_CONFIG_VLAN |
5333 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
5334 
5335 MLX5_TXOFF_INFO(iv_empw,
5336 		MLX5_TXOFF_CONFIG_INLINE | MLX5_TXOFF_CONFIG_VLAN |
5337 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
5338 
5339 MLX5_TXOFF_INFO(full,
5340 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
5341 		MLX5_TXOFF_CONFIG_SWP |	MLX5_TXOFF_CONFIG_CSUM |
5342 		MLX5_TXOFF_CONFIG_INLINE | MLX5_TXOFF_CONFIG_VLAN |
5343 		MLX5_TXOFF_CONFIG_METADATA)
5344 
5345 MLX5_TXOFF_INFO(none,
5346 		MLX5_TXOFF_CONFIG_NONE)
5347 
5348 MLX5_TXOFF_INFO(md,
5349 		MLX5_TXOFF_CONFIG_METADATA)
5350 
5351 MLX5_TXOFF_INFO(mt,
5352 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
5353 		MLX5_TXOFF_CONFIG_METADATA)
5354 
5355 MLX5_TXOFF_INFO(mtsc,
5356 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
5357 		MLX5_TXOFF_CONFIG_SWP |	MLX5_TXOFF_CONFIG_CSUM |
5358 		MLX5_TXOFF_CONFIG_METADATA)
5359 
5360 MLX5_TXOFF_INFO(mti,
5361 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
5362 		MLX5_TXOFF_CONFIG_INLINE |
5363 		MLX5_TXOFF_CONFIG_METADATA)
5364 
5365 MLX5_TXOFF_INFO(mtv,
5366 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
5367 		MLX5_TXOFF_CONFIG_VLAN |
5368 		MLX5_TXOFF_CONFIG_METADATA)
5369 
5370 MLX5_TXOFF_INFO(mtiv,
5371 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
5372 		MLX5_TXOFF_CONFIG_INLINE | MLX5_TXOFF_CONFIG_VLAN |
5373 		MLX5_TXOFF_CONFIG_METADATA)
5374 
5375 MLX5_TXOFF_INFO(sc,
5376 		MLX5_TXOFF_CONFIG_SWP |	MLX5_TXOFF_CONFIG_CSUM |
5377 		MLX5_TXOFF_CONFIG_METADATA)
5378 
5379 MLX5_TXOFF_INFO(sci,
5380 		MLX5_TXOFF_CONFIG_SWP |	MLX5_TXOFF_CONFIG_CSUM |
5381 		MLX5_TXOFF_CONFIG_INLINE |
5382 		MLX5_TXOFF_CONFIG_METADATA)
5383 
5384 MLX5_TXOFF_INFO(scv,
5385 		MLX5_TXOFF_CONFIG_SWP |	MLX5_TXOFF_CONFIG_CSUM |
5386 		MLX5_TXOFF_CONFIG_VLAN |
5387 		MLX5_TXOFF_CONFIG_METADATA)
5388 
5389 MLX5_TXOFF_INFO(sciv,
5390 		MLX5_TXOFF_CONFIG_SWP |	MLX5_TXOFF_CONFIG_CSUM |
5391 		MLX5_TXOFF_CONFIG_INLINE | MLX5_TXOFF_CONFIG_VLAN |
5392 		MLX5_TXOFF_CONFIG_METADATA)
5393 
5394 MLX5_TXOFF_INFO(i,
5395 		MLX5_TXOFF_CONFIG_INLINE |
5396 		MLX5_TXOFF_CONFIG_METADATA)
5397 
5398 MLX5_TXOFF_INFO(v,
5399 		MLX5_TXOFF_CONFIG_VLAN |
5400 		MLX5_TXOFF_CONFIG_METADATA)
5401 
5402 MLX5_TXOFF_INFO(iv,
5403 		MLX5_TXOFF_CONFIG_INLINE | MLX5_TXOFF_CONFIG_VLAN |
5404 		MLX5_TXOFF_CONFIG_METADATA)
5405 
5406 MLX5_TXOFF_INFO(none_mpw,
5407 		MLX5_TXOFF_CONFIG_NONE | MLX5_TXOFF_CONFIG_EMPW |
5408 		MLX5_TXOFF_CONFIG_MPW)
5409 
5410 MLX5_TXOFF_INFO(mci_mpw,
5411 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_CSUM |
5412 		MLX5_TXOFF_CONFIG_INLINE | MLX5_TXOFF_CONFIG_EMPW |
5413 		MLX5_TXOFF_CONFIG_MPW)
5414 
5415 MLX5_TXOFF_INFO(mc_mpw,
5416 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_CSUM |
5417 		MLX5_TXOFF_CONFIG_EMPW | MLX5_TXOFF_CONFIG_MPW)
5418 
5419 MLX5_TXOFF_INFO(i_mpw,
5420 		MLX5_TXOFF_CONFIG_INLINE | MLX5_TXOFF_CONFIG_EMPW |
5421 		MLX5_TXOFF_CONFIG_MPW)
5422 };
5423 
5424 /**
5425  * Configure the Tx function to use. The routine checks configured
5426  * Tx offloads for the device and selects appropriate Tx burst
5427  * routine. There are multiple Tx burst routines compiled from
5428  * the same template in the most optimal way for the dedicated
5429  * Tx offloads set.
5430  *
5431  * @param dev
5432  *   Pointer to private data structure.
5433  *
5434  * @return
5435  *   Pointer to selected Tx burst function.
5436  */
5437 eth_tx_burst_t
5438 mlx5_select_tx_function(struct rte_eth_dev *dev)
5439 {
5440 	struct mlx5_priv *priv = dev->data->dev_private;
5441 	struct mlx5_dev_config *config = &priv->config;
5442 	uint64_t tx_offloads = dev->data->dev_conf.txmode.offloads;
5443 	unsigned int diff = 0, olx = 0, i, m;
5444 
5445 	static_assert(MLX5_WQE_SIZE_MAX / MLX5_WSEG_SIZE <=
5446 		      MLX5_DSEG_MAX, "invalid WQE max size");
5447 	static_assert(MLX5_WQE_CSEG_SIZE == MLX5_WSEG_SIZE,
5448 		      "invalid WQE Control Segment size");
5449 	static_assert(MLX5_WQE_ESEG_SIZE == MLX5_WSEG_SIZE,
5450 		      "invalid WQE Ethernet Segment size");
5451 	static_assert(MLX5_WQE_DSEG_SIZE == MLX5_WSEG_SIZE,
5452 		      "invalid WQE Data Segment size");
5453 	static_assert(MLX5_WQE_SIZE == 4 * MLX5_WSEG_SIZE,
5454 		      "invalid WQE size");
5455 	MLX5_ASSERT(priv);
5456 	if (tx_offloads & DEV_TX_OFFLOAD_MULTI_SEGS) {
5457 		/* We should support Multi-Segment Packets. */
5458 		olx |= MLX5_TXOFF_CONFIG_MULTI;
5459 	}
5460 	if (tx_offloads & (DEV_TX_OFFLOAD_TCP_TSO |
5461 			   DEV_TX_OFFLOAD_VXLAN_TNL_TSO |
5462 			   DEV_TX_OFFLOAD_GRE_TNL_TSO |
5463 			   DEV_TX_OFFLOAD_IP_TNL_TSO |
5464 			   DEV_TX_OFFLOAD_UDP_TNL_TSO)) {
5465 		/* We should support TCP Send Offload. */
5466 		olx |= MLX5_TXOFF_CONFIG_TSO;
5467 	}
5468 	if (tx_offloads & (DEV_TX_OFFLOAD_IP_TNL_TSO |
5469 			   DEV_TX_OFFLOAD_UDP_TNL_TSO |
5470 			   DEV_TX_OFFLOAD_OUTER_IPV4_CKSUM)) {
5471 		/* We should support Software Parser for Tunnels. */
5472 		olx |= MLX5_TXOFF_CONFIG_SWP;
5473 	}
5474 	if (tx_offloads & (DEV_TX_OFFLOAD_IPV4_CKSUM |
5475 			   DEV_TX_OFFLOAD_UDP_CKSUM |
5476 			   DEV_TX_OFFLOAD_TCP_CKSUM |
5477 			   DEV_TX_OFFLOAD_OUTER_IPV4_CKSUM)) {
5478 		/* We should support IP/TCP/UDP Checksums. */
5479 		olx |= MLX5_TXOFF_CONFIG_CSUM;
5480 	}
5481 	if (tx_offloads & DEV_TX_OFFLOAD_VLAN_INSERT) {
5482 		/* We should support VLAN insertion. */
5483 		olx |= MLX5_TXOFF_CONFIG_VLAN;
5484 	}
5485 	if (priv->txqs_n && (*priv->txqs)[0]) {
5486 		struct mlx5_txq_data *txd = (*priv->txqs)[0];
5487 
5488 		if (txd->inlen_send) {
5489 			/*
5490 			 * Check the data inline requirements. Data inline
5491 			 * is enabled on per device basis, we can check
5492 			 * the first Tx queue only.
5493 			 *
5494 			 * If device does not support VLAN insertion in WQE
5495 			 * and some queues are requested to perform VLAN
5496 			 * insertion offload than inline must be enabled.
5497 			 */
5498 			olx |= MLX5_TXOFF_CONFIG_INLINE;
5499 		}
5500 	}
5501 	if (config->mps == MLX5_MPW_ENHANCED &&
5502 	    config->txq_inline_min <= 0) {
5503 		/*
5504 		 * The NIC supports Enhanced Multi-Packet Write
5505 		 * and does not require minimal inline data.
5506 		 */
5507 		olx |= MLX5_TXOFF_CONFIG_EMPW;
5508 	}
5509 	if (rte_flow_dynf_metadata_avail()) {
5510 		/* We should support Flow metadata. */
5511 		olx |= MLX5_TXOFF_CONFIG_METADATA;
5512 	}
5513 	if (config->mps == MLX5_MPW) {
5514 		/*
5515 		 * The NIC supports Legacy Multi-Packet Write.
5516 		 * The MLX5_TXOFF_CONFIG_MPW controls the
5517 		 * descriptor building method in combination
5518 		 * with MLX5_TXOFF_CONFIG_EMPW.
5519 		 */
5520 		if (!(olx & (MLX5_TXOFF_CONFIG_TSO |
5521 			     MLX5_TXOFF_CONFIG_SWP |
5522 			     MLX5_TXOFF_CONFIG_VLAN |
5523 			     MLX5_TXOFF_CONFIG_METADATA)))
5524 			olx |= MLX5_TXOFF_CONFIG_EMPW |
5525 			       MLX5_TXOFF_CONFIG_MPW;
5526 	}
5527 	/*
5528 	 * Scan the routines table to find the minimal
5529 	 * satisfying routine with requested offloads.
5530 	 */
5531 	m = RTE_DIM(txoff_func);
5532 	for (i = 0; i < RTE_DIM(txoff_func); i++) {
5533 		unsigned int tmp;
5534 
5535 		tmp = txoff_func[i].olx;
5536 		if (tmp == olx) {
5537 			/* Meets requested offloads exactly.*/
5538 			m = i;
5539 			break;
5540 		}
5541 		if ((tmp & olx) != olx) {
5542 			/* Does not meet requested offloads at all. */
5543 			continue;
5544 		}
5545 		if ((olx ^ tmp) & MLX5_TXOFF_CONFIG_EMPW)
5546 			/* Do not enable eMPW if not configured. */
5547 			continue;
5548 		if ((olx ^ tmp) & MLX5_TXOFF_CONFIG_INLINE)
5549 			/* Do not enable inlining if not configured. */
5550 			continue;
5551 		/*
5552 		 * Some routine meets the requirements.
5553 		 * Check whether it has minimal amount
5554 		 * of not requested offloads.
5555 		 */
5556 		tmp = __builtin_popcountl(tmp & ~olx);
5557 		if (m >= RTE_DIM(txoff_func) || tmp < diff) {
5558 			/* First or better match, save and continue. */
5559 			m = i;
5560 			diff = tmp;
5561 			continue;
5562 		}
5563 		if (tmp == diff) {
5564 			tmp = txoff_func[i].olx ^ txoff_func[m].olx;
5565 			if (__builtin_ffsl(txoff_func[i].olx & ~tmp) <
5566 			    __builtin_ffsl(txoff_func[m].olx & ~tmp)) {
5567 				/* Lighter not requested offload. */
5568 				m = i;
5569 			}
5570 		}
5571 	}
5572 	if (m >= RTE_DIM(txoff_func)) {
5573 		DRV_LOG(DEBUG, "port %u has no selected Tx function"
5574 			       " for requested offloads %04X",
5575 				dev->data->port_id, olx);
5576 		return NULL;
5577 	}
5578 	DRV_LOG(DEBUG, "port %u has selected Tx function"
5579 		       " supporting offloads %04X/%04X",
5580 			dev->data->port_id, olx, txoff_func[m].olx);
5581 	if (txoff_func[m].olx & MLX5_TXOFF_CONFIG_MULTI)
5582 		DRV_LOG(DEBUG, "\tMULTI (multi segment)");
5583 	if (txoff_func[m].olx & MLX5_TXOFF_CONFIG_TSO)
5584 		DRV_LOG(DEBUG, "\tTSO   (TCP send offload)");
5585 	if (txoff_func[m].olx & MLX5_TXOFF_CONFIG_SWP)
5586 		DRV_LOG(DEBUG, "\tSWP   (software parser)");
5587 	if (txoff_func[m].olx & MLX5_TXOFF_CONFIG_CSUM)
5588 		DRV_LOG(DEBUG, "\tCSUM  (checksum offload)");
5589 	if (txoff_func[m].olx & MLX5_TXOFF_CONFIG_INLINE)
5590 		DRV_LOG(DEBUG, "\tINLIN (inline data)");
5591 	if (txoff_func[m].olx & MLX5_TXOFF_CONFIG_VLAN)
5592 		DRV_LOG(DEBUG, "\tVLANI (VLAN insertion)");
5593 	if (txoff_func[m].olx & MLX5_TXOFF_CONFIG_METADATA)
5594 		DRV_LOG(DEBUG, "\tMETAD (tx Flow metadata)");
5595 	if (txoff_func[m].olx & MLX5_TXOFF_CONFIG_EMPW) {
5596 		if (txoff_func[m].olx & MLX5_TXOFF_CONFIG_MPW)
5597 			DRV_LOG(DEBUG, "\tMPW   (Legacy MPW)");
5598 		else
5599 			DRV_LOG(DEBUG, "\tEMPW  (Enhanced MPW)");
5600 	}
5601 	return txoff_func[m].func;
5602 }
5603 
5604 /**
5605  * DPDK callback to get the TX queue information
5606  *
5607  * @param dev
5608  *   Pointer to the device structure.
5609  *
5610  * @param tx_queue_id
5611  *   Tx queue identificator.
5612  *
5613  * @param qinfo
5614  *   Pointer to the TX queue information structure.
5615  *
5616  * @return
5617  *   None.
5618  */
5619 
5620 void
5621 mlx5_txq_info_get(struct rte_eth_dev *dev, uint16_t tx_queue_id,
5622 		  struct rte_eth_txq_info *qinfo)
5623 {
5624 	struct mlx5_priv *priv = dev->data->dev_private;
5625 	struct mlx5_txq_data *txq = (*priv->txqs)[tx_queue_id];
5626 	struct mlx5_txq_ctrl *txq_ctrl =
5627 			container_of(txq, struct mlx5_txq_ctrl, txq);
5628 
5629 	if (!txq)
5630 		return;
5631 	qinfo->nb_desc = txq->elts_s;
5632 	qinfo->conf.tx_thresh.pthresh = 0;
5633 	qinfo->conf.tx_thresh.hthresh = 0;
5634 	qinfo->conf.tx_thresh.wthresh = 0;
5635 	qinfo->conf.tx_rs_thresh = 0;
5636 	qinfo->conf.tx_free_thresh = 0;
5637 	qinfo->conf.tx_deferred_start = txq_ctrl ? 0 : 1;
5638 	qinfo->conf.offloads = dev->data->dev_conf.txmode.offloads;
5639 }
5640 
5641 /**
5642  * DPDK callback to get the TX packet burst mode information
5643  *
5644  * @param dev
5645  *   Pointer to the device structure.
5646  *
5647  * @param tx_queue_id
5648  *   Tx queue identificatior.
5649  *
5650  * @param mode
5651  *   Pointer to the burts mode information.
5652  *
5653  * @return
5654  *   0 as success, -EINVAL as failure.
5655  */
5656 
5657 int
5658 mlx5_tx_burst_mode_get(struct rte_eth_dev *dev,
5659 		       uint16_t tx_queue_id __rte_unused,
5660 		       struct rte_eth_burst_mode *mode)
5661 {
5662 	eth_tx_burst_t pkt_burst = dev->tx_pkt_burst;
5663 	unsigned int i, olx;
5664 
5665 	for (i = 0; i < RTE_DIM(txoff_func); i++) {
5666 		if (pkt_burst == txoff_func[i].func) {
5667 			olx = txoff_func[i].olx;
5668 			snprintf(mode->info, sizeof(mode->info),
5669 				 "%s%s%s%s%s%s%s%s",
5670 				 (olx & MLX5_TXOFF_CONFIG_EMPW) ?
5671 				 ((olx & MLX5_TXOFF_CONFIG_MPW) ?
5672 				 "Legacy MPW" : "Enhanced MPW") : "No MPW",
5673 				 (olx & MLX5_TXOFF_CONFIG_MULTI) ?
5674 				 " + MULTI" : "",
5675 				 (olx & MLX5_TXOFF_CONFIG_TSO) ?
5676 				 " + TSO" : "",
5677 				 (olx & MLX5_TXOFF_CONFIG_SWP) ?
5678 				 " + SWP" : "",
5679 				 (olx & MLX5_TXOFF_CONFIG_CSUM) ?
5680 				 "  + CSUM" : "",
5681 				 (olx & MLX5_TXOFF_CONFIG_INLINE) ?
5682 				 " + INLINE" : "",
5683 				 (olx & MLX5_TXOFF_CONFIG_VLAN) ?
5684 				 " + VLAN" : "",
5685 				 (olx & MLX5_TXOFF_CONFIG_METADATA) ?
5686 				 " + METADATA" : "");
5687 			return 0;
5688 		}
5689 	}
5690 	return -EINVAL;
5691 }
5692