xref: /dpdk/drivers/net/mlx5/mlx5_rxtx.c (revision bc8e32473cc3978d763a1387eaa8244bcf75e77d)
1 /* SPDX-License-Identifier: BSD-3-Clause
2  * Copyright 2015 6WIND S.A.
3  * Copyright 2015-2019 Mellanox Technologies, Ltd
4  */
5 
6 #include <stdint.h>
7 #include <string.h>
8 #include <stdlib.h>
9 
10 #include <rte_mbuf.h>
11 #include <rte_mempool.h>
12 #include <rte_prefetch.h>
13 #include <rte_common.h>
14 #include <rte_branch_prediction.h>
15 #include <rte_ether.h>
16 #include <rte_cycles.h>
17 #include <rte_flow.h>
18 
19 #include <mlx5_prm.h>
20 #include <mlx5_common.h>
21 
22 #include "mlx5_autoconf.h"
23 #include "mlx5_defs.h"
24 #include "mlx5.h"
25 #include "mlx5_mr.h"
26 #include "mlx5_utils.h"
27 #include "mlx5_rxtx.h"
28 
29 /* TX burst subroutines return codes. */
30 enum mlx5_txcmp_code {
31 	MLX5_TXCMP_CODE_EXIT = 0,
32 	MLX5_TXCMP_CODE_ERROR,
33 	MLX5_TXCMP_CODE_SINGLE,
34 	MLX5_TXCMP_CODE_MULTI,
35 	MLX5_TXCMP_CODE_TSO,
36 	MLX5_TXCMP_CODE_EMPW,
37 };
38 
39 /*
40  * These defines are used to configure Tx burst routine option set
41  * supported at compile time. The not specified options are optimized out
42  * out due to if conditions can be explicitly calculated at compile time.
43  * The offloads with bigger runtime check (require more CPU cycles to
44  * skip) overhead should have the bigger index - this is needed to
45  * select the better matching routine function if no exact match and
46  * some offloads are not actually requested.
47  */
48 #define MLX5_TXOFF_CONFIG_MULTI (1u << 0) /* Multi-segment packets.*/
49 #define MLX5_TXOFF_CONFIG_TSO (1u << 1) /* TCP send offload supported.*/
50 #define MLX5_TXOFF_CONFIG_SWP (1u << 2) /* Tunnels/SW Parser offloads.*/
51 #define MLX5_TXOFF_CONFIG_CSUM (1u << 3) /* Check Sums offloaded. */
52 #define MLX5_TXOFF_CONFIG_INLINE (1u << 4) /* Data inlining supported. */
53 #define MLX5_TXOFF_CONFIG_VLAN (1u << 5) /* VLAN insertion supported.*/
54 #define MLX5_TXOFF_CONFIG_METADATA (1u << 6) /* Flow metadata. */
55 #define MLX5_TXOFF_CONFIG_EMPW (1u << 8) /* Enhanced MPW supported.*/
56 #define MLX5_TXOFF_CONFIG_MPW (1u << 9) /* Legacy MPW supported.*/
57 #define MLX5_TXOFF_CONFIG_TXPP (1u << 10) /* Scheduling on timestamp.*/
58 
59 /* The most common offloads groups. */
60 #define MLX5_TXOFF_CONFIG_NONE 0
61 #define MLX5_TXOFF_CONFIG_FULL (MLX5_TXOFF_CONFIG_MULTI | \
62 				MLX5_TXOFF_CONFIG_TSO | \
63 				MLX5_TXOFF_CONFIG_SWP | \
64 				MLX5_TXOFF_CONFIG_CSUM | \
65 				MLX5_TXOFF_CONFIG_INLINE | \
66 				MLX5_TXOFF_CONFIG_VLAN | \
67 				MLX5_TXOFF_CONFIG_METADATA)
68 
69 #define MLX5_TXOFF_CONFIG(mask) (olx & MLX5_TXOFF_CONFIG_##mask)
70 
71 #define MLX5_TXOFF_DECL(func, olx) \
72 static uint16_t mlx5_tx_burst_##func(void *txq, \
73 				     struct rte_mbuf **pkts, \
74 				    uint16_t pkts_n) \
75 { \
76 	return mlx5_tx_burst_tmpl((struct mlx5_txq_data *)txq, \
77 		    pkts, pkts_n, (olx)); \
78 }
79 
80 #define MLX5_TXOFF_INFO(func, olx) {mlx5_tx_burst_##func, olx},
81 
82 static __rte_always_inline uint32_t
83 rxq_cq_to_pkt_type(struct mlx5_rxq_data *rxq, volatile struct mlx5_cqe *cqe,
84 				   volatile struct mlx5_mini_cqe8 *mcqe);
85 
86 static __rte_always_inline int
87 mlx5_rx_poll_len(struct mlx5_rxq_data *rxq, volatile struct mlx5_cqe *cqe,
88 		 uint16_t cqe_cnt, volatile struct mlx5_mini_cqe8 **mcqe);
89 
90 static __rte_always_inline uint32_t
91 rxq_cq_to_ol_flags(volatile struct mlx5_cqe *cqe);
92 
93 static __rte_always_inline void
94 rxq_cq_to_mbuf(struct mlx5_rxq_data *rxq, struct rte_mbuf *pkt,
95 	       volatile struct mlx5_cqe *cqe,
96 	       volatile struct mlx5_mini_cqe8 *mcqe);
97 
98 static int
99 mlx5_queue_state_modify(struct rte_eth_dev *dev,
100 			struct mlx5_mp_arg_queue_state_modify *sm);
101 
102 static inline void
103 mlx5_lro_update_tcp_hdr(struct rte_tcp_hdr *__rte_restrict tcp,
104 			volatile struct mlx5_cqe *__rte_restrict cqe,
105 			uint32_t phcsum, uint8_t l4_type);
106 
107 static inline void
108 mlx5_lro_update_hdr(uint8_t *__rte_restrict padd,
109 		    volatile struct mlx5_cqe *__rte_restrict cqe,
110 			volatile struct mlx5_mini_cqe8 *mcqe,
111 		    struct mlx5_rxq_data *rxq, uint32_t len);
112 
113 uint32_t mlx5_ptype_table[] __rte_cache_aligned = {
114 	[0xff] = RTE_PTYPE_ALL_MASK, /* Last entry for errored packet. */
115 };
116 
117 uint8_t mlx5_cksum_table[1 << 10] __rte_cache_aligned;
118 uint8_t mlx5_swp_types_table[1 << 10] __rte_cache_aligned;
119 
120 uint64_t rte_net_mlx5_dynf_inline_mask;
121 #define PKT_TX_DYNF_NOINLINE rte_net_mlx5_dynf_inline_mask
122 
123 /**
124  * Build a table to translate Rx completion flags to packet type.
125  *
126  * @note: fix mlx5_dev_supported_ptypes_get() if any change here.
127  */
128 void
129 mlx5_set_ptype_table(void)
130 {
131 	unsigned int i;
132 	uint32_t (*p)[RTE_DIM(mlx5_ptype_table)] = &mlx5_ptype_table;
133 
134 	/* Last entry must not be overwritten, reserved for errored packet. */
135 	for (i = 0; i < RTE_DIM(mlx5_ptype_table) - 1; ++i)
136 		(*p)[i] = RTE_PTYPE_UNKNOWN;
137 	/*
138 	 * The index to the array should have:
139 	 * bit[1:0] = l3_hdr_type
140 	 * bit[4:2] = l4_hdr_type
141 	 * bit[5] = ip_frag
142 	 * bit[6] = tunneled
143 	 * bit[7] = outer_l3_type
144 	 */
145 	/* L2 */
146 	(*p)[0x00] = RTE_PTYPE_L2_ETHER;
147 	/* L3 */
148 	(*p)[0x01] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
149 		     RTE_PTYPE_L4_NONFRAG;
150 	(*p)[0x02] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
151 		     RTE_PTYPE_L4_NONFRAG;
152 	/* Fragmented */
153 	(*p)[0x21] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
154 		     RTE_PTYPE_L4_FRAG;
155 	(*p)[0x22] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
156 		     RTE_PTYPE_L4_FRAG;
157 	/* TCP */
158 	(*p)[0x05] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
159 		     RTE_PTYPE_L4_TCP;
160 	(*p)[0x06] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
161 		     RTE_PTYPE_L4_TCP;
162 	(*p)[0x0d] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
163 		     RTE_PTYPE_L4_TCP;
164 	(*p)[0x0e] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
165 		     RTE_PTYPE_L4_TCP;
166 	(*p)[0x11] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
167 		     RTE_PTYPE_L4_TCP;
168 	(*p)[0x12] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
169 		     RTE_PTYPE_L4_TCP;
170 	/* UDP */
171 	(*p)[0x09] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
172 		     RTE_PTYPE_L4_UDP;
173 	(*p)[0x0a] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
174 		     RTE_PTYPE_L4_UDP;
175 	/* Repeat with outer_l3_type being set. Just in case. */
176 	(*p)[0x81] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
177 		     RTE_PTYPE_L4_NONFRAG;
178 	(*p)[0x82] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
179 		     RTE_PTYPE_L4_NONFRAG;
180 	(*p)[0xa1] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
181 		     RTE_PTYPE_L4_FRAG;
182 	(*p)[0xa2] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
183 		     RTE_PTYPE_L4_FRAG;
184 	(*p)[0x85] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
185 		     RTE_PTYPE_L4_TCP;
186 	(*p)[0x86] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
187 		     RTE_PTYPE_L4_TCP;
188 	(*p)[0x8d] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
189 		     RTE_PTYPE_L4_TCP;
190 	(*p)[0x8e] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
191 		     RTE_PTYPE_L4_TCP;
192 	(*p)[0x91] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
193 		     RTE_PTYPE_L4_TCP;
194 	(*p)[0x92] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
195 		     RTE_PTYPE_L4_TCP;
196 	(*p)[0x89] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
197 		     RTE_PTYPE_L4_UDP;
198 	(*p)[0x8a] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
199 		     RTE_PTYPE_L4_UDP;
200 	/* Tunneled - L3 */
201 	(*p)[0x40] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN;
202 	(*p)[0x41] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
203 		     RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN |
204 		     RTE_PTYPE_INNER_L4_NONFRAG;
205 	(*p)[0x42] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
206 		     RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN |
207 		     RTE_PTYPE_INNER_L4_NONFRAG;
208 	(*p)[0xc0] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN;
209 	(*p)[0xc1] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
210 		     RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN |
211 		     RTE_PTYPE_INNER_L4_NONFRAG;
212 	(*p)[0xc2] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
213 		     RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN |
214 		     RTE_PTYPE_INNER_L4_NONFRAG;
215 	/* Tunneled - Fragmented */
216 	(*p)[0x61] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
217 		     RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN |
218 		     RTE_PTYPE_INNER_L4_FRAG;
219 	(*p)[0x62] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
220 		     RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN |
221 		     RTE_PTYPE_INNER_L4_FRAG;
222 	(*p)[0xe1] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
223 		     RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN |
224 		     RTE_PTYPE_INNER_L4_FRAG;
225 	(*p)[0xe2] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
226 		     RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN |
227 		     RTE_PTYPE_INNER_L4_FRAG;
228 	/* Tunneled - TCP */
229 	(*p)[0x45] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
230 		     RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN |
231 		     RTE_PTYPE_INNER_L4_TCP;
232 	(*p)[0x46] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
233 		     RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN |
234 		     RTE_PTYPE_INNER_L4_TCP;
235 	(*p)[0x4d] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
236 		     RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN |
237 		     RTE_PTYPE_INNER_L4_TCP;
238 	(*p)[0x4e] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
239 		     RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN |
240 		     RTE_PTYPE_INNER_L4_TCP;
241 	(*p)[0x51] = 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)[0x52] = 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)[0xc5] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
248 		     RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN |
249 		     RTE_PTYPE_INNER_L4_TCP;
250 	(*p)[0xc6] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
251 		     RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN |
252 		     RTE_PTYPE_INNER_L4_TCP;
253 	(*p)[0xcd] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
254 		     RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN |
255 		     RTE_PTYPE_INNER_L4_TCP;
256 	(*p)[0xce] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
257 		     RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN |
258 		     RTE_PTYPE_INNER_L4_TCP;
259 	(*p)[0xd1] = 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)[0xd2] = 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 	/* Tunneled - UDP */
266 	(*p)[0x49] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
267 		     RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN |
268 		     RTE_PTYPE_INNER_L4_UDP;
269 	(*p)[0x4a] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
270 		     RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN |
271 		     RTE_PTYPE_INNER_L4_UDP;
272 	(*p)[0xc9] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
273 		     RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN |
274 		     RTE_PTYPE_INNER_L4_UDP;
275 	(*p)[0xca] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
276 		     RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN |
277 		     RTE_PTYPE_INNER_L4_UDP;
278 }
279 
280 /**
281  * Build a table to translate packet to checksum type of Verbs.
282  */
283 void
284 mlx5_set_cksum_table(void)
285 {
286 	unsigned int i;
287 	uint8_t v;
288 
289 	/*
290 	 * The index should have:
291 	 * bit[0] = PKT_TX_TCP_SEG
292 	 * bit[2:3] = PKT_TX_UDP_CKSUM, PKT_TX_TCP_CKSUM
293 	 * bit[4] = PKT_TX_IP_CKSUM
294 	 * bit[8] = PKT_TX_OUTER_IP_CKSUM
295 	 * bit[9] = tunnel
296 	 */
297 	for (i = 0; i < RTE_DIM(mlx5_cksum_table); ++i) {
298 		v = 0;
299 		if (i & (1 << 9)) {
300 			/* Tunneled packet. */
301 			if (i & (1 << 8)) /* Outer IP. */
302 				v |= MLX5_ETH_WQE_L3_CSUM;
303 			if (i & (1 << 4)) /* Inner IP. */
304 				v |= MLX5_ETH_WQE_L3_INNER_CSUM;
305 			if (i & (3 << 2 | 1 << 0)) /* L4 or TSO. */
306 				v |= MLX5_ETH_WQE_L4_INNER_CSUM;
307 		} else {
308 			/* No tunnel. */
309 			if (i & (1 << 4)) /* IP. */
310 				v |= MLX5_ETH_WQE_L3_CSUM;
311 			if (i & (3 << 2 | 1 << 0)) /* L4 or TSO. */
312 				v |= MLX5_ETH_WQE_L4_CSUM;
313 		}
314 		mlx5_cksum_table[i] = v;
315 	}
316 }
317 
318 /**
319  * Build a table to translate packet type of mbuf to SWP type of Verbs.
320  */
321 void
322 mlx5_set_swp_types_table(void)
323 {
324 	unsigned int i;
325 	uint8_t v;
326 
327 	/*
328 	 * The index should have:
329 	 * bit[0:1] = PKT_TX_L4_MASK
330 	 * bit[4] = PKT_TX_IPV6
331 	 * bit[8] = PKT_TX_OUTER_IPV6
332 	 * bit[9] = PKT_TX_OUTER_UDP
333 	 */
334 	for (i = 0; i < RTE_DIM(mlx5_swp_types_table); ++i) {
335 		v = 0;
336 		if (i & (1 << 8))
337 			v |= MLX5_ETH_WQE_L3_OUTER_IPV6;
338 		if (i & (1 << 9))
339 			v |= MLX5_ETH_WQE_L4_OUTER_UDP;
340 		if (i & (1 << 4))
341 			v |= MLX5_ETH_WQE_L3_INNER_IPV6;
342 		if ((i & 3) == (PKT_TX_UDP_CKSUM >> 52))
343 			v |= MLX5_ETH_WQE_L4_INNER_UDP;
344 		mlx5_swp_types_table[i] = v;
345 	}
346 }
347 
348 /**
349  * Set Software Parser flags and offsets in Ethernet Segment of WQE.
350  * Flags must be preliminary initialized to zero.
351  *
352  * @param loc
353  *   Pointer to burst routine local context.
354  * @param swp_flags
355  *   Pointer to store Software Parser flags
356  * @param olx
357  *   Configured Tx offloads mask. It is fully defined at
358  *   compile time and may be used for optimization.
359  *
360  * @return
361  *   Software Parser offsets packed in dword.
362  *   Software Parser flags are set by pointer.
363  */
364 static __rte_always_inline uint32_t
365 txq_mbuf_to_swp(struct mlx5_txq_local *__rte_restrict loc,
366 		uint8_t *swp_flags,
367 		unsigned int olx)
368 {
369 	uint64_t ol, tunnel;
370 	unsigned int idx, off;
371 	uint32_t set;
372 
373 	if (!MLX5_TXOFF_CONFIG(SWP))
374 		return 0;
375 	ol = loc->mbuf->ol_flags;
376 	tunnel = ol & PKT_TX_TUNNEL_MASK;
377 	/*
378 	 * Check whether Software Parser is required.
379 	 * Only customized tunnels may ask for.
380 	 */
381 	if (likely(tunnel != PKT_TX_TUNNEL_UDP && tunnel != PKT_TX_TUNNEL_IP))
382 		return 0;
383 	/*
384 	 * The index should have:
385 	 * bit[0:1] = PKT_TX_L4_MASK
386 	 * bit[4] = PKT_TX_IPV6
387 	 * bit[8] = PKT_TX_OUTER_IPV6
388 	 * bit[9] = PKT_TX_OUTER_UDP
389 	 */
390 	idx = (ol & (PKT_TX_L4_MASK | PKT_TX_IPV6 | PKT_TX_OUTER_IPV6)) >> 52;
391 	idx |= (tunnel == PKT_TX_TUNNEL_UDP) ? (1 << 9) : 0;
392 	*swp_flags = mlx5_swp_types_table[idx];
393 	/*
394 	 * Set offsets for SW parser. Since ConnectX-5, SW parser just
395 	 * complements HW parser. SW parser starts to engage only if HW parser
396 	 * can't reach a header. For the older devices, HW parser will not kick
397 	 * in if any of SWP offsets is set. Therefore, all of the L3 offsets
398 	 * should be set regardless of HW offload.
399 	 */
400 	off = loc->mbuf->outer_l2_len;
401 	if (MLX5_TXOFF_CONFIG(VLAN) && ol & PKT_TX_VLAN_PKT)
402 		off += sizeof(struct rte_vlan_hdr);
403 	set = (off >> 1) << 8; /* Outer L3 offset. */
404 	off += loc->mbuf->outer_l3_len;
405 	if (tunnel == PKT_TX_TUNNEL_UDP)
406 		set |= off >> 1; /* Outer L4 offset. */
407 	if (ol & (PKT_TX_IPV4 | PKT_TX_IPV6)) { /* Inner IP. */
408 		const uint64_t csum = ol & PKT_TX_L4_MASK;
409 			off += loc->mbuf->l2_len;
410 		set |= (off >> 1) << 24; /* Inner L3 offset. */
411 		if (csum == PKT_TX_TCP_CKSUM ||
412 		    csum == PKT_TX_UDP_CKSUM ||
413 		    (MLX5_TXOFF_CONFIG(TSO) && ol & PKT_TX_TCP_SEG)) {
414 			off += loc->mbuf->l3_len;
415 			set |= (off >> 1) << 16; /* Inner L4 offset. */
416 		}
417 	}
418 	set = rte_cpu_to_le_32(set);
419 	return set;
420 }
421 
422 /**
423  * Convert the Checksum offloads to Verbs.
424  *
425  * @param buf
426  *   Pointer to the mbuf.
427  *
428  * @return
429  *   Converted checksum flags.
430  */
431 static __rte_always_inline uint8_t
432 txq_ol_cksum_to_cs(struct rte_mbuf *buf)
433 {
434 	uint32_t idx;
435 	uint8_t is_tunnel = !!(buf->ol_flags & PKT_TX_TUNNEL_MASK);
436 	const uint64_t ol_flags_mask = PKT_TX_TCP_SEG | PKT_TX_L4_MASK |
437 				       PKT_TX_IP_CKSUM | PKT_TX_OUTER_IP_CKSUM;
438 
439 	/*
440 	 * The index should have:
441 	 * bit[0] = PKT_TX_TCP_SEG
442 	 * bit[2:3] = PKT_TX_UDP_CKSUM, PKT_TX_TCP_CKSUM
443 	 * bit[4] = PKT_TX_IP_CKSUM
444 	 * bit[8] = PKT_TX_OUTER_IP_CKSUM
445 	 * bit[9] = tunnel
446 	 */
447 	idx = ((buf->ol_flags & ol_flags_mask) >> 50) | (!!is_tunnel << 9);
448 	return mlx5_cksum_table[idx];
449 }
450 
451 /**
452  * Internal function to compute the number of used descriptors in an RX queue
453  *
454  * @param rxq
455  *   The Rx queue.
456  *
457  * @return
458  *   The number of used rx descriptor.
459  */
460 static uint32_t
461 rx_queue_count(struct mlx5_rxq_data *rxq)
462 {
463 	struct rxq_zip *zip = &rxq->zip;
464 	volatile struct mlx5_cqe *cqe;
465 	unsigned int cq_ci = rxq->cq_ci;
466 	const unsigned int cqe_n = (1 << rxq->cqe_n);
467 	const unsigned int cqe_cnt = cqe_n - 1;
468 	unsigned int used = 0;
469 
470 	cqe = &(*rxq->cqes)[cq_ci & cqe_cnt];
471 	while (check_cqe(cqe, cqe_n, cq_ci) != MLX5_CQE_STATUS_HW_OWN) {
472 		int8_t op_own;
473 		unsigned int n;
474 
475 		op_own = cqe->op_own;
476 		if (MLX5_CQE_FORMAT(op_own) == MLX5_COMPRESSED)
477 			if (unlikely(zip->ai))
478 				n = zip->cqe_cnt - zip->ai;
479 			else
480 				n = rte_be_to_cpu_32(cqe->byte_cnt);
481 		else
482 			n = 1;
483 		cq_ci += n;
484 		used += n;
485 		cqe = &(*rxq->cqes)[cq_ci & cqe_cnt];
486 	}
487 	used = RTE_MIN(used, cqe_n);
488 	return used;
489 }
490 
491 /**
492  * DPDK callback to check the status of a rx descriptor.
493  *
494  * @param rx_queue
495  *   The Rx queue.
496  * @param[in] offset
497  *   The index of the descriptor in the ring.
498  *
499  * @return
500  *   The status of the tx descriptor.
501  */
502 int
503 mlx5_rx_descriptor_status(void *rx_queue, uint16_t offset)
504 {
505 	struct mlx5_rxq_data *rxq = rx_queue;
506 	struct mlx5_rxq_ctrl *rxq_ctrl =
507 			container_of(rxq, struct mlx5_rxq_ctrl, rxq);
508 	struct rte_eth_dev *dev = ETH_DEV(rxq_ctrl->priv);
509 
510 	if (dev->rx_pkt_burst == NULL ||
511 	    dev->rx_pkt_burst == removed_rx_burst) {
512 		rte_errno = ENOTSUP;
513 		return -rte_errno;
514 	}
515 	if (offset >= (1 << rxq->cqe_n)) {
516 		rte_errno = EINVAL;
517 		return -rte_errno;
518 	}
519 	if (offset < rx_queue_count(rxq))
520 		return RTE_ETH_RX_DESC_DONE;
521 	return RTE_ETH_RX_DESC_AVAIL;
522 }
523 
524 /**
525  * DPDK callback to get the RX queue information
526  *
527  * @param dev
528  *   Pointer to the device structure.
529  *
530  * @param rx_queue_id
531  *   Rx queue identificator.
532  *
533  * @param qinfo
534  *   Pointer to the RX queue information structure.
535  *
536  * @return
537  *   None.
538  */
539 
540 void
541 mlx5_rxq_info_get(struct rte_eth_dev *dev, uint16_t rx_queue_id,
542 		  struct rte_eth_rxq_info *qinfo)
543 {
544 	struct mlx5_priv *priv = dev->data->dev_private;
545 	struct mlx5_rxq_data *rxq = (*priv->rxqs)[rx_queue_id];
546 	struct mlx5_rxq_ctrl *rxq_ctrl =
547 		container_of(rxq, struct mlx5_rxq_ctrl, rxq);
548 
549 	if (!rxq)
550 		return;
551 	qinfo->mp = mlx5_rxq_mprq_enabled(rxq) ?
552 					rxq->mprq_mp : rxq->mp;
553 	qinfo->conf.rx_thresh.pthresh = 0;
554 	qinfo->conf.rx_thresh.hthresh = 0;
555 	qinfo->conf.rx_thresh.wthresh = 0;
556 	qinfo->conf.rx_free_thresh = rxq->rq_repl_thresh;
557 	qinfo->conf.rx_drop_en = 1;
558 	qinfo->conf.rx_deferred_start = rxq_ctrl ? 0 : 1;
559 	qinfo->conf.offloads = dev->data->dev_conf.rxmode.offloads;
560 	qinfo->scattered_rx = dev->data->scattered_rx;
561 	qinfo->nb_desc = mlx5_rxq_mprq_enabled(rxq) ?
562 		(1 << rxq->elts_n) * (1 << rxq->strd_num_n) :
563 		(1 << rxq->elts_n);
564 }
565 
566 /**
567  * DPDK callback to get the RX packet burst mode information
568  *
569  * @param dev
570  *   Pointer to the device structure.
571  *
572  * @param rx_queue_id
573  *   Rx queue identificatior.
574  *
575  * @param mode
576  *   Pointer to the burts mode information.
577  *
578  * @return
579  *   0 as success, -EINVAL as failure.
580  */
581 
582 int
583 mlx5_rx_burst_mode_get(struct rte_eth_dev *dev,
584 		       uint16_t rx_queue_id __rte_unused,
585 		       struct rte_eth_burst_mode *mode)
586 {
587 	eth_rx_burst_t pkt_burst = dev->rx_pkt_burst;
588 	struct mlx5_priv *priv = dev->data->dev_private;
589 	struct mlx5_rxq_data *rxq;
590 
591 	rxq = (*priv->rxqs)[rx_queue_id];
592 	if (!rxq) {
593 		rte_errno = EINVAL;
594 		return -rte_errno;
595 	}
596 	if (pkt_burst == mlx5_rx_burst) {
597 		snprintf(mode->info, sizeof(mode->info), "%s", "Scalar");
598 	} else if (pkt_burst == mlx5_rx_burst_mprq) {
599 		snprintf(mode->info, sizeof(mode->info), "%s", "Multi-Packet RQ");
600 	} else if (pkt_burst == mlx5_rx_burst_vec) {
601 #if defined RTE_ARCH_X86_64
602 		snprintf(mode->info, sizeof(mode->info), "%s", "Vector SSE");
603 #elif defined RTE_ARCH_ARM64
604 		snprintf(mode->info, sizeof(mode->info), "%s", "Vector Neon");
605 #elif defined RTE_ARCH_PPC_64
606 		snprintf(mode->info, sizeof(mode->info), "%s", "Vector AltiVec");
607 #else
608 		return -EINVAL;
609 #endif
610 	} else if (pkt_burst == mlx5_rx_burst_mprq_vec) {
611 #if defined RTE_ARCH_X86_64
612 		snprintf(mode->info, sizeof(mode->info), "%s", "MPRQ Vector SSE");
613 #elif defined RTE_ARCH_ARM64
614 		snprintf(mode->info, sizeof(mode->info), "%s", "MPRQ Vector Neon");
615 #elif defined RTE_ARCH_PPC_64
616 		snprintf(mode->info, sizeof(mode->info), "%s", "MPRQ Vector AltiVec");
617 #else
618 		return -EINVAL;
619 #endif
620 	} else {
621 		return -EINVAL;
622 	}
623 	return 0;
624 }
625 
626 /**
627  * DPDK callback to get the number of used descriptors in a RX queue
628  *
629  * @param dev
630  *   Pointer to the device structure.
631  *
632  * @param rx_queue_id
633  *   The Rx queue.
634  *
635  * @return
636  *   The number of used rx descriptor.
637  *   -EINVAL if the queue is invalid
638  */
639 uint32_t
640 mlx5_rx_queue_count(struct rte_eth_dev *dev, uint16_t rx_queue_id)
641 {
642 	struct mlx5_priv *priv = dev->data->dev_private;
643 	struct mlx5_rxq_data *rxq;
644 
645 	if (dev->rx_pkt_burst == NULL ||
646 	    dev->rx_pkt_burst == removed_rx_burst) {
647 		rte_errno = ENOTSUP;
648 		return -rte_errno;
649 	}
650 	rxq = (*priv->rxqs)[rx_queue_id];
651 	if (!rxq) {
652 		rte_errno = EINVAL;
653 		return -rte_errno;
654 	}
655 	return rx_queue_count(rxq);
656 }
657 
658 #define MLX5_SYSTEM_LOG_DIR "/var/log"
659 /**
660  * Dump debug information to log file.
661  *
662  * @param fname
663  *   The file name.
664  * @param hex_title
665  *   If not NULL this string is printed as a header to the output
666  *   and the output will be in hexadecimal view.
667  * @param buf
668  *   This is the buffer address to print out.
669  * @param len
670  *   The number of bytes to dump out.
671  */
672 void
673 mlx5_dump_debug_information(const char *fname, const char *hex_title,
674 			    const void *buf, unsigned int hex_len)
675 {
676 	FILE *fd;
677 
678 	MKSTR(path, "%s/%s", MLX5_SYSTEM_LOG_DIR, fname);
679 	fd = fopen(path, "a+");
680 	if (!fd) {
681 		DRV_LOG(WARNING, "cannot open %s for debug dump", path);
682 		MKSTR(path2, "./%s", fname);
683 		fd = fopen(path2, "a+");
684 		if (!fd) {
685 			DRV_LOG(ERR, "cannot open %s for debug dump", path2);
686 			return;
687 		}
688 		DRV_LOG(INFO, "New debug dump in file %s", path2);
689 	} else {
690 		DRV_LOG(INFO, "New debug dump in file %s", path);
691 	}
692 	if (hex_title)
693 		rte_hexdump(fd, hex_title, buf, hex_len);
694 	else
695 		fprintf(fd, "%s", (const char *)buf);
696 	fprintf(fd, "\n\n\n");
697 	fclose(fd);
698 }
699 
700 /**
701  * Move QP from error state to running state and initialize indexes.
702  *
703  * @param txq_ctrl
704  *   Pointer to TX queue control structure.
705  *
706  * @return
707  *   0 on success, else -1.
708  */
709 static int
710 tx_recover_qp(struct mlx5_txq_ctrl *txq_ctrl)
711 {
712 	struct mlx5_mp_arg_queue_state_modify sm = {
713 			.is_wq = 0,
714 			.queue_id = txq_ctrl->txq.idx,
715 	};
716 
717 	if (mlx5_queue_state_modify(ETH_DEV(txq_ctrl->priv), &sm))
718 		return -1;
719 	txq_ctrl->txq.wqe_ci = 0;
720 	txq_ctrl->txq.wqe_pi = 0;
721 	txq_ctrl->txq.elts_comp = 0;
722 	return 0;
723 }
724 
725 /* Return 1 if the error CQE is signed otherwise, sign it and return 0. */
726 static int
727 check_err_cqe_seen(volatile struct mlx5_err_cqe *err_cqe)
728 {
729 	static const uint8_t magic[] = "seen";
730 	int ret = 1;
731 	unsigned int i;
732 
733 	for (i = 0; i < sizeof(magic); ++i)
734 		if (!ret || err_cqe->rsvd1[i] != magic[i]) {
735 			ret = 0;
736 			err_cqe->rsvd1[i] = magic[i];
737 		}
738 	return ret;
739 }
740 
741 /**
742  * Handle error CQE.
743  *
744  * @param txq
745  *   Pointer to TX queue structure.
746  * @param error_cqe
747  *   Pointer to the error CQE.
748  *
749  * @return
750  *   Negative value if queue recovery failed, otherwise
751  *   the error completion entry is handled successfully.
752  */
753 static int
754 mlx5_tx_error_cqe_handle(struct mlx5_txq_data *__rte_restrict txq,
755 			 volatile struct mlx5_err_cqe *err_cqe)
756 {
757 	if (err_cqe->syndrome != MLX5_CQE_SYNDROME_WR_FLUSH_ERR) {
758 		const uint16_t wqe_m = ((1 << txq->wqe_n) - 1);
759 		struct mlx5_txq_ctrl *txq_ctrl =
760 				container_of(txq, struct mlx5_txq_ctrl, txq);
761 		uint16_t new_wqe_pi = rte_be_to_cpu_16(err_cqe->wqe_counter);
762 		int seen = check_err_cqe_seen(err_cqe);
763 
764 		if (!seen && txq_ctrl->dump_file_n <
765 		    txq_ctrl->priv->config.max_dump_files_num) {
766 			MKSTR(err_str, "Unexpected CQE error syndrome "
767 			      "0x%02x CQN = %u SQN = %u wqe_counter = %u "
768 			      "wq_ci = %u cq_ci = %u", err_cqe->syndrome,
769 			      txq->cqe_s, txq->qp_num_8s >> 8,
770 			      rte_be_to_cpu_16(err_cqe->wqe_counter),
771 			      txq->wqe_ci, txq->cq_ci);
772 			MKSTR(name, "dpdk_mlx5_port_%u_txq_%u_index_%u_%u",
773 			      PORT_ID(txq_ctrl->priv), txq->idx,
774 			      txq_ctrl->dump_file_n, (uint32_t)rte_rdtsc());
775 			mlx5_dump_debug_information(name, NULL, err_str, 0);
776 			mlx5_dump_debug_information(name, "MLX5 Error CQ:",
777 						    (const void *)((uintptr_t)
778 						    txq->cqes),
779 						    sizeof(*err_cqe) *
780 						    (1 << txq->cqe_n));
781 			mlx5_dump_debug_information(name, "MLX5 Error SQ:",
782 						    (const void *)((uintptr_t)
783 						    txq->wqes),
784 						    MLX5_WQE_SIZE *
785 						    (1 << txq->wqe_n));
786 			txq_ctrl->dump_file_n++;
787 		}
788 		if (!seen)
789 			/*
790 			 * Count errors in WQEs units.
791 			 * Later it can be improved to count error packets,
792 			 * for example, by SQ parsing to find how much packets
793 			 * should be counted for each WQE.
794 			 */
795 			txq->stats.oerrors += ((txq->wqe_ci & wqe_m) -
796 						new_wqe_pi) & wqe_m;
797 		if (tx_recover_qp(txq_ctrl)) {
798 			/* Recovering failed - retry later on the same WQE. */
799 			return -1;
800 		}
801 		/* Release all the remaining buffers. */
802 		txq_free_elts(txq_ctrl);
803 	}
804 	return 0;
805 }
806 
807 /**
808  * Translate RX completion flags to packet type.
809  *
810  * @param[in] rxq
811  *   Pointer to RX queue structure.
812  * @param[in] cqe
813  *   Pointer to CQE.
814  *
815  * @note: fix mlx5_dev_supported_ptypes_get() if any change here.
816  *
817  * @return
818  *   Packet type for struct rte_mbuf.
819  */
820 static inline uint32_t
821 rxq_cq_to_pkt_type(struct mlx5_rxq_data *rxq, volatile struct mlx5_cqe *cqe,
822 				   volatile struct mlx5_mini_cqe8 *mcqe)
823 {
824 	uint8_t idx;
825 	uint8_t ptype;
826 	uint8_t pinfo = (cqe->pkt_info & 0x3) << 6;
827 
828 	/* Get l3/l4 header from mini-CQE in case L3/L4 format*/
829 	if (mcqe == NULL ||
830 	    rxq->mcqe_format != MLX5_CQE_RESP_FORMAT_L34H_STRIDX)
831 		ptype = (cqe->hdr_type_etc & 0xfc00) >> 10;
832 	else
833 		ptype = mcqe->hdr_type >> 2;
834 	/*
835 	 * The index to the array should have:
836 	 * bit[1:0] = l3_hdr_type
837 	 * bit[4:2] = l4_hdr_type
838 	 * bit[5] = ip_frag
839 	 * bit[6] = tunneled
840 	 * bit[7] = outer_l3_type
841 	 */
842 	idx = pinfo | ptype;
843 	return mlx5_ptype_table[idx] | rxq->tunnel * !!(idx & (1 << 6));
844 }
845 
846 /**
847  * Initialize Rx WQ and indexes.
848  *
849  * @param[in] rxq
850  *   Pointer to RX queue structure.
851  */
852 void
853 mlx5_rxq_initialize(struct mlx5_rxq_data *rxq)
854 {
855 	const unsigned int wqe_n = 1 << rxq->elts_n;
856 	unsigned int i;
857 
858 	for (i = 0; (i != wqe_n); ++i) {
859 		volatile struct mlx5_wqe_data_seg *scat;
860 		uintptr_t addr;
861 		uint32_t byte_count;
862 
863 		if (mlx5_rxq_mprq_enabled(rxq)) {
864 			struct mlx5_mprq_buf *buf = (*rxq->mprq_bufs)[i];
865 
866 			scat = &((volatile struct mlx5_wqe_mprq *)
867 				rxq->wqes)[i].dseg;
868 			addr = (uintptr_t)mlx5_mprq_buf_addr(buf,
869 							 1 << rxq->strd_num_n);
870 			byte_count = (1 << rxq->strd_sz_n) *
871 					(1 << rxq->strd_num_n);
872 		} else {
873 			struct rte_mbuf *buf = (*rxq->elts)[i];
874 
875 			scat = &((volatile struct mlx5_wqe_data_seg *)
876 					rxq->wqes)[i];
877 			addr = rte_pktmbuf_mtod(buf, uintptr_t);
878 			byte_count = DATA_LEN(buf);
879 		}
880 		/* scat->addr must be able to store a pointer. */
881 		MLX5_ASSERT(sizeof(scat->addr) >= sizeof(uintptr_t));
882 		*scat = (struct mlx5_wqe_data_seg){
883 			.addr = rte_cpu_to_be_64(addr),
884 			.byte_count = rte_cpu_to_be_32(byte_count),
885 			.lkey = mlx5_rx_addr2mr(rxq, addr),
886 		};
887 	}
888 	rxq->consumed_strd = 0;
889 	rxq->decompressed = 0;
890 	rxq->rq_pi = 0;
891 	rxq->zip = (struct rxq_zip){
892 		.ai = 0,
893 	};
894 	rxq->elts_ci = mlx5_rxq_mprq_enabled(rxq) ?
895 		(wqe_n >> rxq->sges_n) * (1 << rxq->strd_num_n) : 0;
896 	/* Update doorbell counter. */
897 	rxq->rq_ci = wqe_n >> rxq->sges_n;
898 	rte_io_wmb();
899 	*rxq->rq_db = rte_cpu_to_be_32(rxq->rq_ci);
900 }
901 
902 /**
903  * Modify a Verbs/DevX queue state.
904  * This must be called from the primary process.
905  *
906  * @param dev
907  *   Pointer to Ethernet device.
908  * @param sm
909  *   State modify request parameters.
910  *
911  * @return
912  *   0 in case of success else non-zero value and rte_errno is set.
913  */
914 int
915 mlx5_queue_state_modify_primary(struct rte_eth_dev *dev,
916 			const struct mlx5_mp_arg_queue_state_modify *sm)
917 {
918 	int ret;
919 	struct mlx5_priv *priv = dev->data->dev_private;
920 
921 	if (sm->is_wq) {
922 		struct mlx5_rxq_data *rxq = (*priv->rxqs)[sm->queue_id];
923 		struct mlx5_rxq_ctrl *rxq_ctrl =
924 			container_of(rxq, struct mlx5_rxq_ctrl, rxq);
925 
926 		ret = priv->obj_ops.rxq_obj_modify(rxq_ctrl->obj, sm->state);
927 		if (ret) {
928 			DRV_LOG(ERR, "Cannot change Rx WQ state to %u  - %s",
929 					sm->state, strerror(errno));
930 			rte_errno = errno;
931 			return ret;
932 		}
933 	} else {
934 		struct mlx5_txq_data *txq = (*priv->txqs)[sm->queue_id];
935 		struct mlx5_txq_ctrl *txq_ctrl =
936 			container_of(txq, struct mlx5_txq_ctrl, txq);
937 
938 		ret = priv->obj_ops.txq_obj_modify(txq_ctrl->obj,
939 						   MLX5_TXQ_MOD_ERR2RDY,
940 						   (uint8_t)priv->dev_port);
941 		if (ret)
942 			return ret;
943 	}
944 	return 0;
945 }
946 
947 /**
948  * Modify a Verbs queue state.
949  *
950  * @param dev
951  *   Pointer to Ethernet device.
952  * @param sm
953  *   State modify request parameters.
954  *
955  * @return
956  *   0 in case of success else non-zero value.
957  */
958 static int
959 mlx5_queue_state_modify(struct rte_eth_dev *dev,
960 			struct mlx5_mp_arg_queue_state_modify *sm)
961 {
962 	struct mlx5_priv *priv = dev->data->dev_private;
963 	int ret = 0;
964 
965 	switch (rte_eal_process_type()) {
966 	case RTE_PROC_PRIMARY:
967 		ret = mlx5_queue_state_modify_primary(dev, sm);
968 		break;
969 	case RTE_PROC_SECONDARY:
970 		ret = mlx5_mp_req_queue_state_modify(&priv->mp_id, sm);
971 		break;
972 	default:
973 		break;
974 	}
975 	return ret;
976 }
977 
978 /**
979  * Handle a Rx error.
980  * The function inserts the RQ state to reset when the first error CQE is
981  * shown, then drains the CQ by the caller function loop. When the CQ is empty,
982  * it moves the RQ state to ready and initializes the RQ.
983  * Next CQE identification and error counting are in the caller responsibility.
984  *
985  * @param[in] rxq
986  *   Pointer to RX queue structure.
987  * @param[in] vec
988  *   1 when called from vectorized Rx burst, need to prepare mbufs for the RQ.
989  *   0 when called from non-vectorized Rx burst.
990  *
991  * @return
992  *   -1 in case of recovery error, otherwise the CQE status.
993  */
994 int
995 mlx5_rx_err_handle(struct mlx5_rxq_data *rxq, uint8_t vec)
996 {
997 	const uint16_t cqe_n = 1 << rxq->cqe_n;
998 	const uint16_t cqe_mask = cqe_n - 1;
999 	const uint16_t wqe_n = 1 << rxq->elts_n;
1000 	const uint16_t strd_n = 1 << rxq->strd_num_n;
1001 	struct mlx5_rxq_ctrl *rxq_ctrl =
1002 			container_of(rxq, struct mlx5_rxq_ctrl, rxq);
1003 	union {
1004 		volatile struct mlx5_cqe *cqe;
1005 		volatile struct mlx5_err_cqe *err_cqe;
1006 	} u = {
1007 		.cqe = &(*rxq->cqes)[rxq->cq_ci & cqe_mask],
1008 	};
1009 	struct mlx5_mp_arg_queue_state_modify sm;
1010 	int ret;
1011 
1012 	switch (rxq->err_state) {
1013 	case MLX5_RXQ_ERR_STATE_NO_ERROR:
1014 		rxq->err_state = MLX5_RXQ_ERR_STATE_NEED_RESET;
1015 		/* Fall-through */
1016 	case MLX5_RXQ_ERR_STATE_NEED_RESET:
1017 		sm.is_wq = 1;
1018 		sm.queue_id = rxq->idx;
1019 		sm.state = IBV_WQS_RESET;
1020 		if (mlx5_queue_state_modify(ETH_DEV(rxq_ctrl->priv), &sm))
1021 			return -1;
1022 		if (rxq_ctrl->dump_file_n <
1023 		    rxq_ctrl->priv->config.max_dump_files_num) {
1024 			MKSTR(err_str, "Unexpected CQE error syndrome "
1025 			      "0x%02x CQN = %u RQN = %u wqe_counter = %u"
1026 			      " rq_ci = %u cq_ci = %u", u.err_cqe->syndrome,
1027 			      rxq->cqn, rxq_ctrl->wqn,
1028 			      rte_be_to_cpu_16(u.err_cqe->wqe_counter),
1029 			      rxq->rq_ci << rxq->sges_n, rxq->cq_ci);
1030 			MKSTR(name, "dpdk_mlx5_port_%u_rxq_%u_%u",
1031 			      rxq->port_id, rxq->idx, (uint32_t)rte_rdtsc());
1032 			mlx5_dump_debug_information(name, NULL, err_str, 0);
1033 			mlx5_dump_debug_information(name, "MLX5 Error CQ:",
1034 						    (const void *)((uintptr_t)
1035 								    rxq->cqes),
1036 						    sizeof(*u.cqe) * cqe_n);
1037 			mlx5_dump_debug_information(name, "MLX5 Error RQ:",
1038 						    (const void *)((uintptr_t)
1039 								    rxq->wqes),
1040 						    16 * wqe_n);
1041 			rxq_ctrl->dump_file_n++;
1042 		}
1043 		rxq->err_state = MLX5_RXQ_ERR_STATE_NEED_READY;
1044 		/* Fall-through */
1045 	case MLX5_RXQ_ERR_STATE_NEED_READY:
1046 		ret = check_cqe(u.cqe, cqe_n, rxq->cq_ci);
1047 		if (ret == MLX5_CQE_STATUS_HW_OWN) {
1048 			rte_io_wmb();
1049 			*rxq->cq_db = rte_cpu_to_be_32(rxq->cq_ci);
1050 			rte_io_wmb();
1051 			/*
1052 			 * The RQ consumer index must be zeroed while moving
1053 			 * from RESET state to RDY state.
1054 			 */
1055 			*rxq->rq_db = rte_cpu_to_be_32(0);
1056 			rte_io_wmb();
1057 			sm.is_wq = 1;
1058 			sm.queue_id = rxq->idx;
1059 			sm.state = IBV_WQS_RDY;
1060 			if (mlx5_queue_state_modify(ETH_DEV(rxq_ctrl->priv),
1061 						    &sm))
1062 				return -1;
1063 			if (vec) {
1064 				const uint32_t elts_n =
1065 					mlx5_rxq_mprq_enabled(rxq) ?
1066 					wqe_n * strd_n : wqe_n;
1067 				const uint32_t e_mask = elts_n - 1;
1068 				uint32_t elts_ci =
1069 					mlx5_rxq_mprq_enabled(rxq) ?
1070 					rxq->elts_ci : rxq->rq_ci;
1071 				uint32_t elt_idx;
1072 				struct rte_mbuf **elt;
1073 				int i;
1074 				unsigned int n = elts_n - (elts_ci -
1075 							  rxq->rq_pi);
1076 
1077 				for (i = 0; i < (int)n; ++i) {
1078 					elt_idx = (elts_ci + i) & e_mask;
1079 					elt = &(*rxq->elts)[elt_idx];
1080 					*elt = rte_mbuf_raw_alloc(rxq->mp);
1081 					if (!*elt) {
1082 						for (i--; i >= 0; --i) {
1083 							elt_idx = (elts_ci +
1084 								   i) & elts_n;
1085 							elt = &(*rxq->elts)
1086 								[elt_idx];
1087 							rte_pktmbuf_free_seg
1088 								(*elt);
1089 						}
1090 						return -1;
1091 					}
1092 				}
1093 				for (i = 0; i < (int)elts_n; ++i) {
1094 					elt = &(*rxq->elts)[i];
1095 					DATA_LEN(*elt) =
1096 						(uint16_t)((*elt)->buf_len -
1097 						rte_pktmbuf_headroom(*elt));
1098 				}
1099 				/* Padding with a fake mbuf for vec Rx. */
1100 				for (i = 0; i < MLX5_VPMD_DESCS_PER_LOOP; ++i)
1101 					(*rxq->elts)[elts_n + i] =
1102 								&rxq->fake_mbuf;
1103 			}
1104 			mlx5_rxq_initialize(rxq);
1105 			rxq->err_state = MLX5_RXQ_ERR_STATE_NO_ERROR;
1106 		}
1107 		return ret;
1108 	default:
1109 		return -1;
1110 	}
1111 }
1112 
1113 /**
1114  * Get size of the next packet for a given CQE. For compressed CQEs, the
1115  * consumer index is updated only once all packets of the current one have
1116  * been processed.
1117  *
1118  * @param rxq
1119  *   Pointer to RX queue.
1120  * @param cqe
1121  *   CQE to process.
1122  * @param[out] mcqe
1123  *   Store pointer to mini-CQE if compressed. Otherwise, the pointer is not
1124  *   written.
1125  *
1126  * @return
1127  *   0 in case of empty CQE, otherwise the packet size in bytes.
1128  */
1129 static inline int
1130 mlx5_rx_poll_len(struct mlx5_rxq_data *rxq, volatile struct mlx5_cqe *cqe,
1131 		 uint16_t cqe_cnt, volatile struct mlx5_mini_cqe8 **mcqe)
1132 {
1133 	struct rxq_zip *zip = &rxq->zip;
1134 	uint16_t cqe_n = cqe_cnt + 1;
1135 	int len;
1136 	uint16_t idx, end;
1137 
1138 	do {
1139 		len = 0;
1140 		/* Process compressed data in the CQE and mini arrays. */
1141 		if (zip->ai) {
1142 			volatile struct mlx5_mini_cqe8 (*mc)[8] =
1143 				(volatile struct mlx5_mini_cqe8 (*)[8])
1144 				(uintptr_t)(&(*rxq->cqes)[zip->ca &
1145 							  cqe_cnt].pkt_info);
1146 			len = rte_be_to_cpu_32((*mc)[zip->ai & 7].byte_cnt &
1147 					       rxq->byte_mask);
1148 			*mcqe = &(*mc)[zip->ai & 7];
1149 			if ((++zip->ai & 7) == 0) {
1150 				/* Invalidate consumed CQEs */
1151 				idx = zip->ca;
1152 				end = zip->na;
1153 				while (idx != end) {
1154 					(*rxq->cqes)[idx & cqe_cnt].op_own =
1155 						MLX5_CQE_INVALIDATE;
1156 					++idx;
1157 				}
1158 				/*
1159 				 * Increment consumer index to skip the number
1160 				 * of CQEs consumed. Hardware leaves holes in
1161 				 * the CQ ring for software use.
1162 				 */
1163 				zip->ca = zip->na;
1164 				zip->na += 8;
1165 			}
1166 			if (unlikely(rxq->zip.ai == rxq->zip.cqe_cnt)) {
1167 				/* Invalidate the rest */
1168 				idx = zip->ca;
1169 				end = zip->cq_ci;
1170 
1171 				while (idx != end) {
1172 					(*rxq->cqes)[idx & cqe_cnt].op_own =
1173 						MLX5_CQE_INVALIDATE;
1174 					++idx;
1175 				}
1176 				rxq->cq_ci = zip->cq_ci;
1177 				zip->ai = 0;
1178 			}
1179 		/*
1180 		 * No compressed data, get next CQE and verify if it is
1181 		 * compressed.
1182 		 */
1183 		} else {
1184 			int ret;
1185 			int8_t op_own;
1186 			uint32_t cq_ci;
1187 
1188 			ret = check_cqe(cqe, cqe_n, rxq->cq_ci);
1189 			if (unlikely(ret != MLX5_CQE_STATUS_SW_OWN)) {
1190 				if (unlikely(ret == MLX5_CQE_STATUS_ERR ||
1191 					     rxq->err_state)) {
1192 					ret = mlx5_rx_err_handle(rxq, 0);
1193 					if (ret == MLX5_CQE_STATUS_HW_OWN ||
1194 					    ret == -1)
1195 						return 0;
1196 				} else {
1197 					return 0;
1198 				}
1199 			}
1200 			/*
1201 			 * Introduce the local variable to have queue cq_ci
1202 			 * index in queue structure always consistent with
1203 			 * actual CQE boundary (not pointing to the middle
1204 			 * of compressed CQE session).
1205 			 */
1206 			cq_ci = rxq->cq_ci + 1;
1207 			op_own = cqe->op_own;
1208 			if (MLX5_CQE_FORMAT(op_own) == MLX5_COMPRESSED) {
1209 				volatile struct mlx5_mini_cqe8 (*mc)[8] =
1210 					(volatile struct mlx5_mini_cqe8 (*)[8])
1211 					(uintptr_t)(&(*rxq->cqes)
1212 						[cq_ci & cqe_cnt].pkt_info);
1213 
1214 				/* Fix endianness. */
1215 				zip->cqe_cnt = rte_be_to_cpu_32(cqe->byte_cnt);
1216 				/*
1217 				 * Current mini array position is the one
1218 				 * returned by check_cqe64().
1219 				 *
1220 				 * If completion comprises several mini arrays,
1221 				 * as a special case the second one is located
1222 				 * 7 CQEs after the initial CQE instead of 8
1223 				 * for subsequent ones.
1224 				 */
1225 				zip->ca = cq_ci;
1226 				zip->na = zip->ca + 7;
1227 				/* Compute the next non compressed CQE. */
1228 				zip->cq_ci = rxq->cq_ci + zip->cqe_cnt;
1229 				/* Get packet size to return. */
1230 				len = rte_be_to_cpu_32((*mc)[0].byte_cnt &
1231 						       rxq->byte_mask);
1232 				*mcqe = &(*mc)[0];
1233 				zip->ai = 1;
1234 				/* Prefetch all to be invalidated */
1235 				idx = zip->ca;
1236 				end = zip->cq_ci;
1237 				while (idx != end) {
1238 					rte_prefetch0(&(*rxq->cqes)[(idx) &
1239 								    cqe_cnt]);
1240 					++idx;
1241 				}
1242 			} else {
1243 				rxq->cq_ci = cq_ci;
1244 				len = rte_be_to_cpu_32(cqe->byte_cnt);
1245 			}
1246 		}
1247 		if (unlikely(rxq->err_state)) {
1248 			cqe = &(*rxq->cqes)[rxq->cq_ci & cqe_cnt];
1249 			++rxq->stats.idropped;
1250 		} else {
1251 			return len;
1252 		}
1253 	} while (1);
1254 }
1255 
1256 /**
1257  * Translate RX completion flags to offload flags.
1258  *
1259  * @param[in] cqe
1260  *   Pointer to CQE.
1261  *
1262  * @return
1263  *   Offload flags (ol_flags) for struct rte_mbuf.
1264  */
1265 static inline uint32_t
1266 rxq_cq_to_ol_flags(volatile struct mlx5_cqe *cqe)
1267 {
1268 	uint32_t ol_flags = 0;
1269 	uint16_t flags = rte_be_to_cpu_16(cqe->hdr_type_etc);
1270 
1271 	ol_flags =
1272 		TRANSPOSE(flags,
1273 			  MLX5_CQE_RX_L3_HDR_VALID,
1274 			  PKT_RX_IP_CKSUM_GOOD) |
1275 		TRANSPOSE(flags,
1276 			  MLX5_CQE_RX_L4_HDR_VALID,
1277 			  PKT_RX_L4_CKSUM_GOOD);
1278 	return ol_flags;
1279 }
1280 
1281 /**
1282  * Fill in mbuf fields from RX completion flags.
1283  * Note that pkt->ol_flags should be initialized outside of this function.
1284  *
1285  * @param rxq
1286  *   Pointer to RX queue.
1287  * @param pkt
1288  *   mbuf to fill.
1289  * @param cqe
1290  *   CQE to process.
1291  * @param rss_hash_res
1292  *   Packet RSS Hash result.
1293  */
1294 static inline void
1295 rxq_cq_to_mbuf(struct mlx5_rxq_data *rxq, struct rte_mbuf *pkt,
1296 	       volatile struct mlx5_cqe *cqe,
1297 	       volatile struct mlx5_mini_cqe8 *mcqe)
1298 {
1299 	/* Update packet information. */
1300 	pkt->packet_type = rxq_cq_to_pkt_type(rxq, cqe, mcqe);
1301 
1302 	if (rxq->rss_hash) {
1303 		uint32_t rss_hash_res = 0;
1304 
1305 		/* If compressed, take hash result from mini-CQE. */
1306 		if (mcqe == NULL ||
1307 		    rxq->mcqe_format != MLX5_CQE_RESP_FORMAT_HASH)
1308 			rss_hash_res = rte_be_to_cpu_32(cqe->rx_hash_res);
1309 		else
1310 			rss_hash_res = rte_be_to_cpu_32(mcqe->rx_hash_result);
1311 		if (rss_hash_res) {
1312 			pkt->hash.rss = rss_hash_res;
1313 			pkt->ol_flags |= PKT_RX_RSS_HASH;
1314 		}
1315 	}
1316 	if (rxq->mark) {
1317 		uint32_t mark = 0;
1318 
1319 		/* If compressed, take flow tag from mini-CQE. */
1320 		if (mcqe == NULL ||
1321 		    rxq->mcqe_format != MLX5_CQE_RESP_FORMAT_FTAG_STRIDX)
1322 			mark = cqe->sop_drop_qpn;
1323 		else
1324 			mark = ((mcqe->byte_cnt_flow & 0xff) << 8) |
1325 				(mcqe->flow_tag_high << 16);
1326 		if (MLX5_FLOW_MARK_IS_VALID(mark)) {
1327 			pkt->ol_flags |= PKT_RX_FDIR;
1328 			if (mark != RTE_BE32(MLX5_FLOW_MARK_DEFAULT)) {
1329 				pkt->ol_flags |= PKT_RX_FDIR_ID;
1330 				pkt->hash.fdir.hi = mlx5_flow_mark_get(mark);
1331 			}
1332 		}
1333 	}
1334 	if (rxq->dynf_meta && cqe->flow_table_metadata) {
1335 		pkt->ol_flags |= rxq->flow_meta_mask;
1336 		*RTE_MBUF_DYNFIELD(pkt, rxq->flow_meta_offset, uint32_t *) =
1337 			cqe->flow_table_metadata;
1338 	}
1339 	if (rxq->csum)
1340 		pkt->ol_flags |= rxq_cq_to_ol_flags(cqe);
1341 	if (rxq->vlan_strip) {
1342 		bool vlan_strip;
1343 
1344 		if (mcqe == NULL ||
1345 		    rxq->mcqe_format != MLX5_CQE_RESP_FORMAT_L34H_STRIDX)
1346 			vlan_strip = cqe->hdr_type_etc &
1347 				     RTE_BE16(MLX5_CQE_VLAN_STRIPPED);
1348 		else
1349 			vlan_strip = mcqe->hdr_type &
1350 				     RTE_BE16(MLX5_CQE_VLAN_STRIPPED);
1351 		if (vlan_strip) {
1352 			pkt->ol_flags |= PKT_RX_VLAN | PKT_RX_VLAN_STRIPPED;
1353 			pkt->vlan_tci = rte_be_to_cpu_16(cqe->vlan_info);
1354 		}
1355 	}
1356 	if (rxq->hw_timestamp) {
1357 		uint64_t ts = rte_be_to_cpu_64(cqe->timestamp);
1358 
1359 		if (rxq->rt_timestamp)
1360 			ts = mlx5_txpp_convert_rx_ts(rxq->sh, ts);
1361 		mlx5_timestamp_set(pkt, rxq->timestamp_offset, ts);
1362 		pkt->ol_flags |= rxq->timestamp_rx_flag;
1363 	}
1364 }
1365 
1366 /**
1367  * DPDK callback for RX.
1368  *
1369  * @param dpdk_rxq
1370  *   Generic pointer to RX queue structure.
1371  * @param[out] pkts
1372  *   Array to store received packets.
1373  * @param pkts_n
1374  *   Maximum number of packets in array.
1375  *
1376  * @return
1377  *   Number of packets successfully received (<= pkts_n).
1378  */
1379 uint16_t
1380 mlx5_rx_burst(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n)
1381 {
1382 	struct mlx5_rxq_data *rxq = dpdk_rxq;
1383 	const unsigned int wqe_cnt = (1 << rxq->elts_n) - 1;
1384 	const unsigned int cqe_cnt = (1 << rxq->cqe_n) - 1;
1385 	const unsigned int sges_n = rxq->sges_n;
1386 	struct rte_mbuf *pkt = NULL;
1387 	struct rte_mbuf *seg = NULL;
1388 	volatile struct mlx5_cqe *cqe =
1389 		&(*rxq->cqes)[rxq->cq_ci & cqe_cnt];
1390 	unsigned int i = 0;
1391 	unsigned int rq_ci = rxq->rq_ci << sges_n;
1392 	int len = 0; /* keep its value across iterations. */
1393 
1394 	while (pkts_n) {
1395 		unsigned int idx = rq_ci & wqe_cnt;
1396 		volatile struct mlx5_wqe_data_seg *wqe =
1397 			&((volatile struct mlx5_wqe_data_seg *)rxq->wqes)[idx];
1398 		struct rte_mbuf *rep = (*rxq->elts)[idx];
1399 		volatile struct mlx5_mini_cqe8 *mcqe = NULL;
1400 
1401 		if (pkt)
1402 			NEXT(seg) = rep;
1403 		seg = rep;
1404 		rte_prefetch0(seg);
1405 		rte_prefetch0(cqe);
1406 		rte_prefetch0(wqe);
1407 		/* Allocate the buf from the same pool. */
1408 		rep = rte_mbuf_raw_alloc(seg->pool);
1409 		if (unlikely(rep == NULL)) {
1410 			++rxq->stats.rx_nombuf;
1411 			if (!pkt) {
1412 				/*
1413 				 * no buffers before we even started,
1414 				 * bail out silently.
1415 				 */
1416 				break;
1417 			}
1418 			while (pkt != seg) {
1419 				MLX5_ASSERT(pkt != (*rxq->elts)[idx]);
1420 				rep = NEXT(pkt);
1421 				NEXT(pkt) = NULL;
1422 				NB_SEGS(pkt) = 1;
1423 				rte_mbuf_raw_free(pkt);
1424 				pkt = rep;
1425 			}
1426 			break;
1427 		}
1428 		if (!pkt) {
1429 			cqe = &(*rxq->cqes)[rxq->cq_ci & cqe_cnt];
1430 			len = mlx5_rx_poll_len(rxq, cqe, cqe_cnt, &mcqe);
1431 			if (!len) {
1432 				rte_mbuf_raw_free(rep);
1433 				break;
1434 			}
1435 			pkt = seg;
1436 			MLX5_ASSERT(len >= (rxq->crc_present << 2));
1437 			pkt->ol_flags &= EXT_ATTACHED_MBUF;
1438 			rxq_cq_to_mbuf(rxq, pkt, cqe, mcqe);
1439 			if (rxq->crc_present)
1440 				len -= RTE_ETHER_CRC_LEN;
1441 			PKT_LEN(pkt) = len;
1442 			if (cqe->lro_num_seg > 1) {
1443 				mlx5_lro_update_hdr
1444 					(rte_pktmbuf_mtod(pkt, uint8_t *), cqe,
1445 					 mcqe, rxq, len);
1446 				pkt->ol_flags |= PKT_RX_LRO;
1447 				pkt->tso_segsz = len / cqe->lro_num_seg;
1448 			}
1449 		}
1450 		DATA_LEN(rep) = DATA_LEN(seg);
1451 		PKT_LEN(rep) = PKT_LEN(seg);
1452 		SET_DATA_OFF(rep, DATA_OFF(seg));
1453 		PORT(rep) = PORT(seg);
1454 		(*rxq->elts)[idx] = rep;
1455 		/*
1456 		 * Fill NIC descriptor with the new buffer.  The lkey and size
1457 		 * of the buffers are already known, only the buffer address
1458 		 * changes.
1459 		 */
1460 		wqe->addr = rte_cpu_to_be_64(rte_pktmbuf_mtod(rep, uintptr_t));
1461 		/* If there's only one MR, no need to replace LKey in WQE. */
1462 		if (unlikely(mlx5_mr_btree_len(&rxq->mr_ctrl.cache_bh) > 1))
1463 			wqe->lkey = mlx5_rx_mb2mr(rxq, rep);
1464 		if (len > DATA_LEN(seg)) {
1465 			len -= DATA_LEN(seg);
1466 			++NB_SEGS(pkt);
1467 			++rq_ci;
1468 			continue;
1469 		}
1470 		DATA_LEN(seg) = len;
1471 #ifdef MLX5_PMD_SOFT_COUNTERS
1472 		/* Increment bytes counter. */
1473 		rxq->stats.ibytes += PKT_LEN(pkt);
1474 #endif
1475 		/* Return packet. */
1476 		*(pkts++) = pkt;
1477 		pkt = NULL;
1478 		--pkts_n;
1479 		++i;
1480 		/* Align consumer index to the next stride. */
1481 		rq_ci >>= sges_n;
1482 		++rq_ci;
1483 		rq_ci <<= sges_n;
1484 	}
1485 	if (unlikely((i == 0) && ((rq_ci >> sges_n) == rxq->rq_ci)))
1486 		return 0;
1487 	/* Update the consumer index. */
1488 	rxq->rq_ci = rq_ci >> sges_n;
1489 	rte_io_wmb();
1490 	*rxq->cq_db = rte_cpu_to_be_32(rxq->cq_ci);
1491 	rte_io_wmb();
1492 	*rxq->rq_db = rte_cpu_to_be_32(rxq->rq_ci);
1493 #ifdef MLX5_PMD_SOFT_COUNTERS
1494 	/* Increment packets counter. */
1495 	rxq->stats.ipackets += i;
1496 #endif
1497 	return i;
1498 }
1499 
1500 /**
1501  * Update LRO packet TCP header.
1502  * The HW LRO feature doesn't update the TCP header after coalescing the
1503  * TCP segments but supplies information in CQE to fill it by SW.
1504  *
1505  * @param tcp
1506  *   Pointer to the TCP header.
1507  * @param cqe
1508  *   Pointer to the completion entry..
1509  * @param phcsum
1510  *   The L3 pseudo-header checksum.
1511  */
1512 static inline void
1513 mlx5_lro_update_tcp_hdr(struct rte_tcp_hdr *__rte_restrict tcp,
1514 			volatile struct mlx5_cqe *__rte_restrict cqe,
1515 			uint32_t phcsum, uint8_t l4_type)
1516 {
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 >> 4) * 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 *__rte_restrict padd,
1554 		    volatile struct mlx5_cqe *__rte_restrict cqe,
1555 		    volatile struct mlx5_mini_cqe8 *mcqe,
1556 		    struct mlx5_rxq_data *rxq, uint32_t len)
1557 {
1558 	union {
1559 		struct rte_ether_hdr *eth;
1560 		struct rte_vlan_hdr *vlan;
1561 		struct rte_ipv4_hdr *ipv4;
1562 		struct rte_ipv6_hdr *ipv6;
1563 		struct rte_tcp_hdr *tcp;
1564 		uint8_t *hdr;
1565 	} h = {
1566 			.hdr = padd,
1567 	};
1568 	uint16_t proto = h.eth->ether_type;
1569 	uint32_t phcsum;
1570 	uint8_t l4_type;
1571 
1572 	h.eth++;
1573 	while (proto == RTE_BE16(RTE_ETHER_TYPE_VLAN) ||
1574 	       proto == RTE_BE16(RTE_ETHER_TYPE_QINQ)) {
1575 		proto = h.vlan->eth_proto;
1576 		h.vlan++;
1577 	}
1578 	if (proto == RTE_BE16(RTE_ETHER_TYPE_IPV4)) {
1579 		h.ipv4->time_to_live = cqe->lro_min_ttl;
1580 		h.ipv4->total_length = rte_cpu_to_be_16(len - (h.hdr - padd));
1581 		h.ipv4->hdr_checksum = 0;
1582 		h.ipv4->hdr_checksum = rte_ipv4_cksum(h.ipv4);
1583 		phcsum = rte_ipv4_phdr_cksum(h.ipv4, 0);
1584 		h.ipv4++;
1585 	} else {
1586 		h.ipv6->hop_limits = cqe->lro_min_ttl;
1587 		h.ipv6->payload_len = rte_cpu_to_be_16(len - (h.hdr - padd) -
1588 						       sizeof(*h.ipv6));
1589 		phcsum = rte_ipv6_phdr_cksum(h.ipv6, 0);
1590 		h.ipv6++;
1591 	}
1592 	if (mcqe == NULL ||
1593 	    rxq->mcqe_format != MLX5_CQE_RESP_FORMAT_L34H_STRIDX)
1594 		l4_type = (rte_be_to_cpu_16(cqe->hdr_type_etc) &
1595 			   MLX5_CQE_L4_TYPE_MASK) >> MLX5_CQE_L4_TYPE_SHIFT;
1596 	else
1597 		l4_type = (rte_be_to_cpu_16(mcqe->hdr_type) &
1598 			   MLX5_CQE_L4_TYPE_MASK) >> MLX5_CQE_L4_TYPE_SHIFT;
1599 	mlx5_lro_update_tcp_hdr(h.tcp, cqe, phcsum, l4_type);
1600 }
1601 
1602 void
1603 mlx5_mprq_buf_free_cb(void *addr __rte_unused, void *opaque)
1604 {
1605 	struct mlx5_mprq_buf *buf = opaque;
1606 
1607 	if (__atomic_load_n(&buf->refcnt, __ATOMIC_RELAXED) == 1) {
1608 		rte_mempool_put(buf->mp, buf);
1609 	} else if (unlikely(__atomic_sub_fetch(&buf->refcnt, 1,
1610 					       __ATOMIC_RELAXED) == 0)) {
1611 		__atomic_store_n(&buf->refcnt, 1, __ATOMIC_RELAXED);
1612 		rte_mempool_put(buf->mp, buf);
1613 	}
1614 }
1615 
1616 void
1617 mlx5_mprq_buf_free(struct mlx5_mprq_buf *buf)
1618 {
1619 	mlx5_mprq_buf_free_cb(NULL, buf);
1620 }
1621 
1622 /**
1623  * DPDK callback for RX with Multi-Packet RQ support.
1624  *
1625  * @param dpdk_rxq
1626  *   Generic pointer to RX queue structure.
1627  * @param[out] pkts
1628  *   Array to store received packets.
1629  * @param pkts_n
1630  *   Maximum number of packets in array.
1631  *
1632  * @return
1633  *   Number of packets successfully received (<= pkts_n).
1634  */
1635 uint16_t
1636 mlx5_rx_burst_mprq(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n)
1637 {
1638 	struct mlx5_rxq_data *rxq = dpdk_rxq;
1639 	const uint32_t strd_n = 1 << rxq->strd_num_n;
1640 	const uint32_t strd_sz = 1 << rxq->strd_sz_n;
1641 	const uint32_t cq_mask = (1 << rxq->cqe_n) - 1;
1642 	const uint32_t wq_mask = (1 << rxq->elts_n) - 1;
1643 	volatile struct mlx5_cqe *cqe = &(*rxq->cqes)[rxq->cq_ci & cq_mask];
1644 	unsigned int i = 0;
1645 	uint32_t rq_ci = rxq->rq_ci;
1646 	uint16_t consumed_strd = rxq->consumed_strd;
1647 	struct mlx5_mprq_buf *buf = (*rxq->mprq_bufs)[rq_ci & wq_mask];
1648 
1649 	while (i < pkts_n) {
1650 		struct rte_mbuf *pkt;
1651 		int ret;
1652 		uint32_t len;
1653 		uint16_t strd_cnt;
1654 		uint16_t strd_idx;
1655 		uint32_t byte_cnt;
1656 		volatile struct mlx5_mini_cqe8 *mcqe = NULL;
1657 		enum mlx5_rqx_code rxq_code;
1658 
1659 		if (consumed_strd == strd_n) {
1660 			/* Replace WQE if the buffer is still in use. */
1661 			mprq_buf_replace(rxq, rq_ci & wq_mask);
1662 			/* Advance to the next WQE. */
1663 			consumed_strd = 0;
1664 			++rq_ci;
1665 			buf = (*rxq->mprq_bufs)[rq_ci & wq_mask];
1666 		}
1667 		cqe = &(*rxq->cqes)[rxq->cq_ci & cq_mask];
1668 		ret = mlx5_rx_poll_len(rxq, cqe, cq_mask, &mcqe);
1669 		if (!ret)
1670 			break;
1671 		byte_cnt = ret;
1672 		len = (byte_cnt & MLX5_MPRQ_LEN_MASK) >> MLX5_MPRQ_LEN_SHIFT;
1673 		MLX5_ASSERT((int)len >= (rxq->crc_present << 2));
1674 		if (rxq->crc_present)
1675 			len -= RTE_ETHER_CRC_LEN;
1676 		if (mcqe &&
1677 		    rxq->mcqe_format == MLX5_CQE_RESP_FORMAT_FTAG_STRIDX)
1678 			strd_cnt = (len / strd_sz) + !!(len % strd_sz);
1679 		else
1680 			strd_cnt = (byte_cnt & MLX5_MPRQ_STRIDE_NUM_MASK) >>
1681 				   MLX5_MPRQ_STRIDE_NUM_SHIFT;
1682 		MLX5_ASSERT(strd_cnt);
1683 		consumed_strd += strd_cnt;
1684 		if (byte_cnt & MLX5_MPRQ_FILLER_MASK)
1685 			continue;
1686 		strd_idx = rte_be_to_cpu_16(mcqe == NULL ?
1687 					cqe->wqe_counter :
1688 					mcqe->stride_idx);
1689 		MLX5_ASSERT(strd_idx < strd_n);
1690 		MLX5_ASSERT(!((rte_be_to_cpu_16(cqe->wqe_id) ^ rq_ci) &
1691 			    wq_mask));
1692 		pkt = rte_pktmbuf_alloc(rxq->mp);
1693 		if (unlikely(pkt == NULL)) {
1694 			++rxq->stats.rx_nombuf;
1695 			break;
1696 		}
1697 		len = (byte_cnt & MLX5_MPRQ_LEN_MASK) >> MLX5_MPRQ_LEN_SHIFT;
1698 		MLX5_ASSERT((int)len >= (rxq->crc_present << 2));
1699 		if (rxq->crc_present)
1700 			len -= RTE_ETHER_CRC_LEN;
1701 		rxq_code = mprq_buf_to_pkt(rxq, pkt, len, buf,
1702 					   strd_idx, strd_cnt);
1703 		if (unlikely(rxq_code != MLX5_RXQ_CODE_EXIT)) {
1704 			rte_pktmbuf_free_seg(pkt);
1705 			if (rxq_code == MLX5_RXQ_CODE_DROPPED) {
1706 				++rxq->stats.idropped;
1707 				continue;
1708 			}
1709 			if (rxq_code == MLX5_RXQ_CODE_NOMBUF) {
1710 				++rxq->stats.rx_nombuf;
1711 				break;
1712 			}
1713 		}
1714 		rxq_cq_to_mbuf(rxq, pkt, cqe, mcqe);
1715 		if (cqe->lro_num_seg > 1) {
1716 			mlx5_lro_update_hdr(rte_pktmbuf_mtod(pkt, uint8_t *),
1717 					    cqe, mcqe, rxq, len);
1718 			pkt->ol_flags |= PKT_RX_LRO;
1719 			pkt->tso_segsz = len / cqe->lro_num_seg;
1720 		}
1721 		PKT_LEN(pkt) = len;
1722 		PORT(pkt) = rxq->port_id;
1723 #ifdef MLX5_PMD_SOFT_COUNTERS
1724 		/* Increment bytes counter. */
1725 		rxq->stats.ibytes += PKT_LEN(pkt);
1726 #endif
1727 		/* Return packet. */
1728 		*(pkts++) = pkt;
1729 		++i;
1730 	}
1731 	/* Update the consumer indexes. */
1732 	rxq->consumed_strd = consumed_strd;
1733 	rte_io_wmb();
1734 	*rxq->cq_db = rte_cpu_to_be_32(rxq->cq_ci);
1735 	if (rq_ci != rxq->rq_ci) {
1736 		rxq->rq_ci = rq_ci;
1737 		rte_io_wmb();
1738 		*rxq->rq_db = rte_cpu_to_be_32(rxq->rq_ci);
1739 	}
1740 #ifdef MLX5_PMD_SOFT_COUNTERS
1741 	/* Increment packets counter. */
1742 	rxq->stats.ipackets += i;
1743 #endif
1744 	return i;
1745 }
1746 
1747 /**
1748  * Dummy DPDK callback for TX.
1749  *
1750  * This function is used to temporarily replace the real callback during
1751  * unsafe control operations on the queue, or in case of error.
1752  *
1753  * @param dpdk_txq
1754  *   Generic pointer to TX queue structure.
1755  * @param[in] pkts
1756  *   Packets to transmit.
1757  * @param pkts_n
1758  *   Number of packets in array.
1759  *
1760  * @return
1761  *   Number of packets successfully transmitted (<= pkts_n).
1762  */
1763 uint16_t
1764 removed_tx_burst(void *dpdk_txq __rte_unused,
1765 		 struct rte_mbuf **pkts __rte_unused,
1766 		 uint16_t pkts_n __rte_unused)
1767 {
1768 	rte_mb();
1769 	return 0;
1770 }
1771 
1772 /**
1773  * Dummy DPDK callback for RX.
1774  *
1775  * This function is used to temporarily replace the real callback during
1776  * unsafe control operations on the queue, or in case of error.
1777  *
1778  * @param dpdk_rxq
1779  *   Generic pointer to RX queue structure.
1780  * @param[out] pkts
1781  *   Array to store received packets.
1782  * @param pkts_n
1783  *   Maximum number of packets in array.
1784  *
1785  * @return
1786  *   Number of packets successfully received (<= pkts_n).
1787  */
1788 uint16_t
1789 removed_rx_burst(void *dpdk_txq __rte_unused,
1790 		 struct rte_mbuf **pkts __rte_unused,
1791 		 uint16_t pkts_n __rte_unused)
1792 {
1793 	rte_mb();
1794 	return 0;
1795 }
1796 
1797 /*
1798  * Vectorized Rx/Tx routines are not compiled in when required vector
1799  * instructions are not supported on a target architecture. The following null
1800  * stubs are needed for linkage when those are not included outside of this file
1801  * (e.g.  mlx5_rxtx_vec_sse.c for x86).
1802  */
1803 
1804 __rte_weak uint16_t
1805 mlx5_rx_burst_vec(void *dpdk_txq __rte_unused,
1806 		  struct rte_mbuf **pkts __rte_unused,
1807 		  uint16_t pkts_n __rte_unused)
1808 {
1809 	return 0;
1810 }
1811 
1812 __rte_weak uint16_t
1813 mlx5_rx_burst_mprq_vec(void *dpdk_txq __rte_unused,
1814 		       struct rte_mbuf **pkts __rte_unused,
1815 		       uint16_t pkts_n __rte_unused)
1816 {
1817 	return 0;
1818 }
1819 
1820 __rte_weak int
1821 mlx5_rxq_check_vec_support(struct mlx5_rxq_data *rxq __rte_unused)
1822 {
1823 	return -ENOTSUP;
1824 }
1825 
1826 __rte_weak int
1827 mlx5_check_vec_rx_support(struct rte_eth_dev *dev __rte_unused)
1828 {
1829 	return -ENOTSUP;
1830 }
1831 
1832 /**
1833  * Free the mbufs from the linear array of pointers.
1834  *
1835  * @param pkts
1836  *   Pointer to array of packets to be free.
1837  * @param pkts_n
1838  *   Number of packets to be freed.
1839  * @param olx
1840  *   Configured Tx offloads mask. It is fully defined at
1841  *   compile time and may be used for optimization.
1842  */
1843 static __rte_always_inline void
1844 mlx5_tx_free_mbuf(struct rte_mbuf **__rte_restrict pkts,
1845 		  unsigned int pkts_n,
1846 		  unsigned int olx __rte_unused)
1847 {
1848 	struct rte_mempool *pool = NULL;
1849 	struct rte_mbuf **p_free = NULL;
1850 	struct rte_mbuf *mbuf;
1851 	unsigned int n_free = 0;
1852 
1853 	/*
1854 	 * The implemented algorithm eliminates
1855 	 * copying pointers to temporary array
1856 	 * for rte_mempool_put_bulk() calls.
1857 	 */
1858 	MLX5_ASSERT(pkts);
1859 	MLX5_ASSERT(pkts_n);
1860 	for (;;) {
1861 		for (;;) {
1862 			/*
1863 			 * Decrement mbuf reference counter, detach
1864 			 * indirect and external buffers if needed.
1865 			 */
1866 			mbuf = rte_pktmbuf_prefree_seg(*pkts);
1867 			if (likely(mbuf != NULL)) {
1868 				MLX5_ASSERT(mbuf == *pkts);
1869 				if (likely(n_free != 0)) {
1870 					if (unlikely(pool != mbuf->pool))
1871 						/* From different pool. */
1872 						break;
1873 				} else {
1874 					/* Start new scan array. */
1875 					pool = mbuf->pool;
1876 					p_free = pkts;
1877 				}
1878 				++n_free;
1879 				++pkts;
1880 				--pkts_n;
1881 				if (unlikely(pkts_n == 0)) {
1882 					mbuf = NULL;
1883 					break;
1884 				}
1885 			} else {
1886 				/*
1887 				 * This happens if mbuf is still referenced.
1888 				 * We can't put it back to the pool, skip.
1889 				 */
1890 				++pkts;
1891 				--pkts_n;
1892 				if (unlikely(n_free != 0))
1893 					/* There is some array to free.*/
1894 					break;
1895 				if (unlikely(pkts_n == 0))
1896 					/* Last mbuf, nothing to free. */
1897 					return;
1898 			}
1899 		}
1900 		for (;;) {
1901 			/*
1902 			 * This loop is implemented to avoid multiple
1903 			 * inlining of rte_mempool_put_bulk().
1904 			 */
1905 			MLX5_ASSERT(pool);
1906 			MLX5_ASSERT(p_free);
1907 			MLX5_ASSERT(n_free);
1908 			/*
1909 			 * Free the array of pre-freed mbufs
1910 			 * belonging to the same memory pool.
1911 			 */
1912 			rte_mempool_put_bulk(pool, (void *)p_free, n_free);
1913 			if (unlikely(mbuf != NULL)) {
1914 				/* There is the request to start new scan. */
1915 				pool = mbuf->pool;
1916 				p_free = pkts++;
1917 				n_free = 1;
1918 				--pkts_n;
1919 				if (likely(pkts_n != 0))
1920 					break;
1921 				/*
1922 				 * This is the last mbuf to be freed.
1923 				 * Do one more loop iteration to complete.
1924 				 * This is rare case of the last unique mbuf.
1925 				 */
1926 				mbuf = NULL;
1927 				continue;
1928 			}
1929 			if (likely(pkts_n == 0))
1930 				return;
1931 			n_free = 0;
1932 			break;
1933 		}
1934 	}
1935 }
1936 
1937 /**
1938  * Free the mbuf from the elts ring buffer till new tail.
1939  *
1940  * @param txq
1941  *   Pointer to Tx queue structure.
1942  * @param tail
1943  *   Index in elts to free up to, becomes new elts tail.
1944  * @param olx
1945  *   Configured Tx offloads mask. It is fully defined at
1946  *   compile time and may be used for optimization.
1947  */
1948 static __rte_always_inline void
1949 mlx5_tx_free_elts(struct mlx5_txq_data *__rte_restrict txq,
1950 		  uint16_t tail,
1951 		  unsigned int olx __rte_unused)
1952 {
1953 	uint16_t n_elts = tail - txq->elts_tail;
1954 
1955 	MLX5_ASSERT(n_elts);
1956 	MLX5_ASSERT(n_elts <= txq->elts_s);
1957 	/*
1958 	 * Implement a loop to support ring buffer wraparound
1959 	 * with single inlining of mlx5_tx_free_mbuf().
1960 	 */
1961 	do {
1962 		unsigned int part;
1963 
1964 		part = txq->elts_s - (txq->elts_tail & txq->elts_m);
1965 		part = RTE_MIN(part, n_elts);
1966 		MLX5_ASSERT(part);
1967 		MLX5_ASSERT(part <= txq->elts_s);
1968 		mlx5_tx_free_mbuf(&txq->elts[txq->elts_tail & txq->elts_m],
1969 				  part, olx);
1970 		txq->elts_tail += part;
1971 		n_elts -= part;
1972 	} while (n_elts);
1973 }
1974 
1975 /**
1976  * Store the mbuf being sent into elts ring buffer.
1977  * On Tx completion these mbufs will be freed.
1978  *
1979  * @param txq
1980  *   Pointer to Tx queue structure.
1981  * @param pkts
1982  *   Pointer to array of packets to be stored.
1983  * @param pkts_n
1984  *   Number of packets to be stored.
1985  * @param olx
1986  *   Configured Tx offloads mask. It is fully defined at
1987  *   compile time and may be used for optimization.
1988  */
1989 static __rte_always_inline void
1990 mlx5_tx_copy_elts(struct mlx5_txq_data *__rte_restrict txq,
1991 		  struct rte_mbuf **__rte_restrict pkts,
1992 		  unsigned int pkts_n,
1993 		  unsigned int olx __rte_unused)
1994 {
1995 	unsigned int part;
1996 	struct rte_mbuf **elts = (struct rte_mbuf **)txq->elts;
1997 
1998 	MLX5_ASSERT(pkts);
1999 	MLX5_ASSERT(pkts_n);
2000 	part = txq->elts_s - (txq->elts_head & txq->elts_m);
2001 	MLX5_ASSERT(part);
2002 	MLX5_ASSERT(part <= txq->elts_s);
2003 	/* This code is a good candidate for vectorizing with SIMD. */
2004 	rte_memcpy((void *)(elts + (txq->elts_head & txq->elts_m)),
2005 		   (void *)pkts,
2006 		   RTE_MIN(part, pkts_n) * sizeof(struct rte_mbuf *));
2007 	txq->elts_head += pkts_n;
2008 	if (unlikely(part < pkts_n))
2009 		/* The copy is wrapping around the elts array. */
2010 		rte_memcpy((void *)elts, (void *)(pkts + part),
2011 			   (pkts_n - part) * sizeof(struct rte_mbuf *));
2012 }
2013 
2014 /**
2015  * Update completion queue consuming index via doorbell
2016  * and flush the completed data buffers.
2017  *
2018  * @param txq
2019  *   Pointer to TX queue structure.
2020  * @param valid CQE pointer
2021  *   if not NULL update txq->wqe_pi and flush the buffers
2022  * @param olx
2023  *   Configured Tx offloads mask. It is fully defined at
2024  *   compile time and may be used for optimization.
2025  */
2026 static __rte_always_inline void
2027 mlx5_tx_comp_flush(struct mlx5_txq_data *__rte_restrict txq,
2028 		   volatile struct mlx5_cqe *last_cqe,
2029 		   unsigned int olx __rte_unused)
2030 {
2031 	if (likely(last_cqe != NULL)) {
2032 		uint16_t tail;
2033 
2034 		txq->wqe_pi = rte_be_to_cpu_16(last_cqe->wqe_counter);
2035 		tail = txq->fcqs[(txq->cq_ci - 1) & txq->cqe_m];
2036 		if (likely(tail != txq->elts_tail)) {
2037 			mlx5_tx_free_elts(txq, tail, olx);
2038 			MLX5_ASSERT(tail == txq->elts_tail);
2039 		}
2040 	}
2041 }
2042 
2043 /**
2044  * Manage TX completions. This routine checks the CQ for
2045  * arrived CQEs, deduces the last accomplished WQE in SQ,
2046  * updates SQ producing index and frees all completed mbufs.
2047  *
2048  * @param txq
2049  *   Pointer to TX queue structure.
2050  * @param olx
2051  *   Configured Tx offloads mask. It is fully defined at
2052  *   compile time and may be used for optimization.
2053  *
2054  * NOTE: not inlined intentionally, it makes tx_burst
2055  * routine smaller, simple and faster - from experiments.
2056  */
2057 static void
2058 mlx5_tx_handle_completion(struct mlx5_txq_data *__rte_restrict txq,
2059 			  unsigned int olx __rte_unused)
2060 {
2061 	unsigned int count = MLX5_TX_COMP_MAX_CQE;
2062 	volatile struct mlx5_cqe *last_cqe = NULL;
2063 	bool ring_doorbell = false;
2064 	int ret;
2065 
2066 	static_assert(MLX5_CQE_STATUS_HW_OWN < 0, "Must be negative value");
2067 	static_assert(MLX5_CQE_STATUS_SW_OWN < 0, "Must be negative value");
2068 	do {
2069 		volatile struct mlx5_cqe *cqe;
2070 
2071 		cqe = &txq->cqes[txq->cq_ci & txq->cqe_m];
2072 		ret = check_cqe(cqe, txq->cqe_s, txq->cq_ci);
2073 		if (unlikely(ret != MLX5_CQE_STATUS_SW_OWN)) {
2074 			if (likely(ret != MLX5_CQE_STATUS_ERR)) {
2075 				/* No new CQEs in completion queue. */
2076 				MLX5_ASSERT(ret == MLX5_CQE_STATUS_HW_OWN);
2077 				break;
2078 			}
2079 			/*
2080 			 * Some error occurred, try to restart.
2081 			 * We have no barrier after WQE related Doorbell
2082 			 * written, make sure all writes are completed
2083 			 * here, before we might perform SQ reset.
2084 			 */
2085 			rte_wmb();
2086 			ret = mlx5_tx_error_cqe_handle
2087 				(txq, (volatile struct mlx5_err_cqe *)cqe);
2088 			if (unlikely(ret < 0)) {
2089 				/*
2090 				 * Some error occurred on queue error
2091 				 * handling, we do not advance the index
2092 				 * here, allowing to retry on next call.
2093 				 */
2094 				return;
2095 			}
2096 			/*
2097 			 * We are going to fetch all entries with
2098 			 * MLX5_CQE_SYNDROME_WR_FLUSH_ERR status.
2099 			 * The send queue is supposed to be empty.
2100 			 */
2101 			ring_doorbell = true;
2102 			++txq->cq_ci;
2103 			txq->cq_pi = txq->cq_ci;
2104 			last_cqe = NULL;
2105 			continue;
2106 		}
2107 		/* Normal transmit completion. */
2108 		MLX5_ASSERT(txq->cq_ci != txq->cq_pi);
2109 		MLX5_ASSERT((txq->fcqs[txq->cq_ci & txq->cqe_m] >> 16) ==
2110 			    cqe->wqe_counter);
2111 		ring_doorbell = true;
2112 		++txq->cq_ci;
2113 		last_cqe = cqe;
2114 		/*
2115 		 * We have to restrict the amount of processed CQEs
2116 		 * in one tx_burst routine call. The CQ may be large
2117 		 * and many CQEs may be updated by the NIC in one
2118 		 * transaction. Buffers freeing is time consuming,
2119 		 * multiple iterations may introduce significant
2120 		 * latency.
2121 		 */
2122 		if (likely(--count == 0))
2123 			break;
2124 	} while (true);
2125 	if (likely(ring_doorbell)) {
2126 		/* Ring doorbell to notify hardware. */
2127 		rte_compiler_barrier();
2128 		*txq->cq_db = rte_cpu_to_be_32(txq->cq_ci);
2129 		mlx5_tx_comp_flush(txq, last_cqe, olx);
2130 	}
2131 }
2132 
2133 /**
2134  * Check if the completion request flag should be set in the last WQE.
2135  * Both pushed mbufs and WQEs are monitored and the completion request
2136  * flag is set if any of thresholds is reached.
2137  *
2138  * @param txq
2139  *   Pointer to TX queue structure.
2140  * @param loc
2141  *   Pointer to burst routine local context.
2142  * @param olx
2143  *   Configured Tx offloads mask. It is fully defined at
2144  *   compile time and may be used for optimization.
2145  */
2146 static __rte_always_inline void
2147 mlx5_tx_request_completion(struct mlx5_txq_data *__rte_restrict txq,
2148 			   struct mlx5_txq_local *__rte_restrict loc,
2149 			   unsigned int olx)
2150 {
2151 	uint16_t head = txq->elts_head;
2152 	unsigned int part;
2153 
2154 	part = MLX5_TXOFF_CONFIG(INLINE) ?
2155 	       0 : loc->pkts_sent - loc->pkts_copy;
2156 	head += part;
2157 	if ((uint16_t)(head - txq->elts_comp) >= MLX5_TX_COMP_THRESH ||
2158 	     (MLX5_TXOFF_CONFIG(INLINE) &&
2159 	     (uint16_t)(txq->wqe_ci - txq->wqe_comp) >= txq->wqe_thres)) {
2160 		volatile struct mlx5_wqe *last = loc->wqe_last;
2161 
2162 		MLX5_ASSERT(last);
2163 		txq->elts_comp = head;
2164 		if (MLX5_TXOFF_CONFIG(INLINE))
2165 			txq->wqe_comp = txq->wqe_ci;
2166 		/* Request unconditional completion on last WQE. */
2167 		last->cseg.flags = RTE_BE32(MLX5_COMP_ALWAYS <<
2168 					    MLX5_COMP_MODE_OFFSET);
2169 		/* Save elts_head in dedicated free on completion queue. */
2170 #ifdef RTE_LIBRTE_MLX5_DEBUG
2171 		txq->fcqs[txq->cq_pi++ & txq->cqe_m] = head |
2172 			  (last->cseg.opcode >> 8) << 16;
2173 #else
2174 		txq->fcqs[txq->cq_pi++ & txq->cqe_m] = head;
2175 #endif
2176 		/* A CQE slot must always be available. */
2177 		MLX5_ASSERT((txq->cq_pi - txq->cq_ci) <= txq->cqe_s);
2178 	}
2179 }
2180 
2181 /**
2182  * DPDK callback to check the status of a tx descriptor.
2183  *
2184  * @param tx_queue
2185  *   The tx queue.
2186  * @param[in] offset
2187  *   The index of the descriptor in the ring.
2188  *
2189  * @return
2190  *   The status of the tx descriptor.
2191  */
2192 int
2193 mlx5_tx_descriptor_status(void *tx_queue, uint16_t offset)
2194 {
2195 	struct mlx5_txq_data *__rte_restrict txq = tx_queue;
2196 	uint16_t used;
2197 
2198 	mlx5_tx_handle_completion(txq, 0);
2199 	used = txq->elts_head - txq->elts_tail;
2200 	if (offset < used)
2201 		return RTE_ETH_TX_DESC_FULL;
2202 	return RTE_ETH_TX_DESC_DONE;
2203 }
2204 
2205 /**
2206  * Build the Control Segment with specified opcode:
2207  * - MLX5_OPCODE_SEND
2208  * - MLX5_OPCODE_ENHANCED_MPSW
2209  * - MLX5_OPCODE_TSO
2210  *
2211  * @param txq
2212  *   Pointer to TX queue structure.
2213  * @param loc
2214  *   Pointer to burst routine local context.
2215  * @param wqe
2216  *   Pointer to WQE to fill with built Control Segment.
2217  * @param ds
2218  *   Supposed length of WQE in segments.
2219  * @param opcode
2220  *   SQ WQE opcode to put into Control Segment.
2221  * @param olx
2222  *   Configured Tx offloads mask. It is fully defined at
2223  *   compile time and may be used for optimization.
2224  */
2225 static __rte_always_inline void
2226 mlx5_tx_cseg_init(struct mlx5_txq_data *__rte_restrict txq,
2227 		  struct mlx5_txq_local *__rte_restrict loc __rte_unused,
2228 		  struct mlx5_wqe *__rte_restrict wqe,
2229 		  unsigned int ds,
2230 		  unsigned int opcode,
2231 		  unsigned int olx __rte_unused)
2232 {
2233 	struct mlx5_wqe_cseg *__rte_restrict cs = &wqe->cseg;
2234 
2235 	/* For legacy MPW replace the EMPW by TSO with modifier. */
2236 	if (MLX5_TXOFF_CONFIG(MPW) && opcode == MLX5_OPCODE_ENHANCED_MPSW)
2237 		opcode = MLX5_OPCODE_TSO | MLX5_OPC_MOD_MPW << 24;
2238 	cs->opcode = rte_cpu_to_be_32((txq->wqe_ci << 8) | opcode);
2239 	cs->sq_ds = rte_cpu_to_be_32(txq->qp_num_8s | ds);
2240 	cs->flags = RTE_BE32(MLX5_COMP_ONLY_FIRST_ERR <<
2241 			     MLX5_COMP_MODE_OFFSET);
2242 	cs->misc = RTE_BE32(0);
2243 }
2244 
2245 /**
2246  * Build the Synchronize Queue Segment with specified completion index.
2247  *
2248  * @param txq
2249  *   Pointer to TX queue structure.
2250  * @param loc
2251  *   Pointer to burst routine local context.
2252  * @param wqe
2253  *   Pointer to WQE to fill with built Control Segment.
2254  * @param wci
2255  *   Completion index in Clock Queue to wait.
2256  * @param olx
2257  *   Configured Tx offloads mask. It is fully defined at
2258  *   compile time and may be used for optimization.
2259  */
2260 static __rte_always_inline void
2261 mlx5_tx_wseg_init(struct mlx5_txq_data *restrict txq,
2262 		  struct mlx5_txq_local *restrict loc __rte_unused,
2263 		  struct mlx5_wqe *restrict wqe,
2264 		  unsigned int wci,
2265 		  unsigned int olx __rte_unused)
2266 {
2267 	struct mlx5_wqe_qseg *qs;
2268 
2269 	qs = RTE_PTR_ADD(wqe, MLX5_WSEG_SIZE);
2270 	qs->max_index = rte_cpu_to_be_32(wci);
2271 	qs->qpn_cqn = rte_cpu_to_be_32(txq->sh->txpp.clock_queue.cq->id);
2272 	qs->reserved0 = RTE_BE32(0);
2273 	qs->reserved1 = RTE_BE32(0);
2274 }
2275 
2276 /**
2277  * Build the Ethernet Segment without inlined data.
2278  * Supports Software Parser, Checksums and VLAN
2279  * insertion Tx offload features.
2280  *
2281  * @param txq
2282  *   Pointer to TX queue structure.
2283  * @param loc
2284  *   Pointer to burst routine local context.
2285  * @param wqe
2286  *   Pointer to WQE to fill with built Ethernet Segment.
2287  * @param olx
2288  *   Configured Tx offloads mask. It is fully defined at
2289  *   compile time and may be used for optimization.
2290  */
2291 static __rte_always_inline void
2292 mlx5_tx_eseg_none(struct mlx5_txq_data *__rte_restrict txq __rte_unused,
2293 		  struct mlx5_txq_local *__rte_restrict loc,
2294 		  struct mlx5_wqe *__rte_restrict wqe,
2295 		  unsigned int olx)
2296 {
2297 	struct mlx5_wqe_eseg *__rte_restrict es = &wqe->eseg;
2298 	uint32_t csum;
2299 
2300 	/*
2301 	 * Calculate and set check sum flags first, dword field
2302 	 * in segment may be shared with Software Parser flags.
2303 	 */
2304 	csum = MLX5_TXOFF_CONFIG(CSUM) ? txq_ol_cksum_to_cs(loc->mbuf) : 0;
2305 	es->flags = rte_cpu_to_le_32(csum);
2306 	/*
2307 	 * Calculate and set Software Parser offsets and flags.
2308 	 * These flags a set for custom UDP and IP tunnel packets.
2309 	 */
2310 	es->swp_offs = txq_mbuf_to_swp(loc, &es->swp_flags, olx);
2311 	/* Fill metadata field if needed. */
2312 	es->metadata = MLX5_TXOFF_CONFIG(METADATA) ?
2313 		       loc->mbuf->ol_flags & PKT_TX_DYNF_METADATA ?
2314 		       *RTE_FLOW_DYNF_METADATA(loc->mbuf) : 0 : 0;
2315 	/* Engage VLAN tag insertion feature if requested. */
2316 	if (MLX5_TXOFF_CONFIG(VLAN) &&
2317 	    loc->mbuf->ol_flags & PKT_TX_VLAN_PKT) {
2318 		/*
2319 		 * We should get here only if device support
2320 		 * this feature correctly.
2321 		 */
2322 		MLX5_ASSERT(txq->vlan_en);
2323 		es->inline_hdr = rte_cpu_to_be_32(MLX5_ETH_WQE_VLAN_INSERT |
2324 						  loc->mbuf->vlan_tci);
2325 	} else {
2326 		es->inline_hdr = RTE_BE32(0);
2327 	}
2328 }
2329 
2330 /**
2331  * Build the Ethernet Segment with minimal inlined data
2332  * of MLX5_ESEG_MIN_INLINE_SIZE bytes length. This is
2333  * used to fill the gap in single WQEBB WQEs.
2334  * Supports Software Parser, Checksums and VLAN
2335  * insertion Tx offload features.
2336  *
2337  * @param txq
2338  *   Pointer to TX queue structure.
2339  * @param loc
2340  *   Pointer to burst routine local context.
2341  * @param wqe
2342  *   Pointer to WQE to fill with built Ethernet Segment.
2343  * @param vlan
2344  *   Length of VLAN tag insertion if any.
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_eseg_dmin(struct mlx5_txq_data *__rte_restrict txq __rte_unused,
2351 		  struct mlx5_txq_local *__rte_restrict loc,
2352 		  struct mlx5_wqe *__rte_restrict wqe,
2353 		  unsigned int vlan,
2354 		  unsigned int olx)
2355 {
2356 	struct mlx5_wqe_eseg *__rte_restrict es = &wqe->eseg;
2357 	uint32_t csum;
2358 	uint8_t *psrc, *pdst;
2359 
2360 	/*
2361 	 * Calculate and set check sum flags first, dword field
2362 	 * in segment may be shared with Software Parser flags.
2363 	 */
2364 	csum = MLX5_TXOFF_CONFIG(CSUM) ? txq_ol_cksum_to_cs(loc->mbuf) : 0;
2365 	es->flags = rte_cpu_to_le_32(csum);
2366 	/*
2367 	 * Calculate and set Software Parser offsets and flags.
2368 	 * These flags a set for custom UDP and IP tunnel packets.
2369 	 */
2370 	es->swp_offs = txq_mbuf_to_swp(loc, &es->swp_flags, olx);
2371 	/* Fill metadata field if needed. */
2372 	es->metadata = MLX5_TXOFF_CONFIG(METADATA) ?
2373 		       loc->mbuf->ol_flags & PKT_TX_DYNF_METADATA ?
2374 		       *RTE_FLOW_DYNF_METADATA(loc->mbuf) : 0 : 0;
2375 	static_assert(MLX5_ESEG_MIN_INLINE_SIZE ==
2376 				(sizeof(uint16_t) +
2377 				 sizeof(rte_v128u32_t)),
2378 		      "invalid Ethernet Segment data size");
2379 	static_assert(MLX5_ESEG_MIN_INLINE_SIZE ==
2380 				(sizeof(uint16_t) +
2381 				 sizeof(struct rte_vlan_hdr) +
2382 				 2 * RTE_ETHER_ADDR_LEN),
2383 		      "invalid Ethernet Segment data size");
2384 	psrc = rte_pktmbuf_mtod(loc->mbuf, uint8_t *);
2385 	es->inline_hdr_sz = RTE_BE16(MLX5_ESEG_MIN_INLINE_SIZE);
2386 	es->inline_data = *(unaligned_uint16_t *)psrc;
2387 	psrc +=	sizeof(uint16_t);
2388 	pdst = (uint8_t *)(es + 1);
2389 	if (MLX5_TXOFF_CONFIG(VLAN) && vlan) {
2390 		/* Implement VLAN tag insertion as part inline data. */
2391 		memcpy(pdst, psrc, 2 * RTE_ETHER_ADDR_LEN - sizeof(uint16_t));
2392 		pdst += 2 * RTE_ETHER_ADDR_LEN - sizeof(uint16_t);
2393 		psrc +=	2 * RTE_ETHER_ADDR_LEN - sizeof(uint16_t);
2394 		/* Insert VLAN ethertype + VLAN tag. */
2395 		*(unaligned_uint32_t *)pdst = rte_cpu_to_be_32
2396 						((RTE_ETHER_TYPE_VLAN << 16) |
2397 						 loc->mbuf->vlan_tci);
2398 		pdst += sizeof(struct rte_vlan_hdr);
2399 		/* Copy the rest two bytes from packet data. */
2400 		MLX5_ASSERT(pdst == RTE_PTR_ALIGN(pdst, sizeof(uint16_t)));
2401 		*(uint16_t *)pdst = *(unaligned_uint16_t *)psrc;
2402 	} else {
2403 		/* Fill the gap in the title WQEBB with inline data. */
2404 		rte_mov16(pdst, psrc);
2405 	}
2406 }
2407 
2408 /**
2409  * Build the Ethernet Segment with entire packet
2410  * data inlining. Checks the boundary of WQEBB and
2411  * ring buffer wrapping, supports Software Parser,
2412  * Checksums and VLAN insertion Tx offload features.
2413  *
2414  * @param txq
2415  *   Pointer to TX queue structure.
2416  * @param loc
2417  *   Pointer to burst routine local context.
2418  * @param wqe
2419  *   Pointer to WQE to fill with built Ethernet Segment.
2420  * @param vlan
2421  *   Length of VLAN tag insertion if any.
2422  * @param inlen
2423  *   Length of data to inline (VLAN included, if any).
2424  * @param tso
2425  *   TSO flag, set mss field from the packet.
2426  * @param olx
2427  *   Configured Tx offloads mask. It is fully defined at
2428  *   compile time and may be used for optimization.
2429  *
2430  * @return
2431  *   Pointer to the next Data Segment (aligned and wrapped around).
2432  */
2433 static __rte_always_inline struct mlx5_wqe_dseg *
2434 mlx5_tx_eseg_data(struct mlx5_txq_data *__rte_restrict txq,
2435 		  struct mlx5_txq_local *__rte_restrict loc,
2436 		  struct mlx5_wqe *__rte_restrict wqe,
2437 		  unsigned int vlan,
2438 		  unsigned int inlen,
2439 		  unsigned int tso,
2440 		  unsigned int olx)
2441 {
2442 	struct mlx5_wqe_eseg *__rte_restrict es = &wqe->eseg;
2443 	uint32_t csum;
2444 	uint8_t *psrc, *pdst;
2445 	unsigned int part;
2446 
2447 	/*
2448 	 * Calculate and set check sum flags first, dword field
2449 	 * in segment may be shared with Software Parser flags.
2450 	 */
2451 	csum = MLX5_TXOFF_CONFIG(CSUM) ? txq_ol_cksum_to_cs(loc->mbuf) : 0;
2452 	if (tso) {
2453 		csum <<= 24;
2454 		csum |= loc->mbuf->tso_segsz;
2455 		es->flags = rte_cpu_to_be_32(csum);
2456 	} else {
2457 		es->flags = rte_cpu_to_le_32(csum);
2458 	}
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_cpu_to_be_16(inlen);
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 		psrc += sizeof(uint16_t);
2496 	} else {
2497 		/* Fill the gap in the title WQEBB with inline data. */
2498 		rte_mov16(pdst, psrc);
2499 		psrc += sizeof(rte_v128u32_t);
2500 	}
2501 	pdst = (uint8_t *)(es + 2);
2502 	MLX5_ASSERT(inlen >= MLX5_ESEG_MIN_INLINE_SIZE);
2503 	MLX5_ASSERT(pdst < (uint8_t *)txq->wqes_end);
2504 	inlen -= MLX5_ESEG_MIN_INLINE_SIZE;
2505 	if (!inlen) {
2506 		MLX5_ASSERT(pdst == RTE_PTR_ALIGN(pdst, MLX5_WSEG_SIZE));
2507 		return (struct mlx5_wqe_dseg *)pdst;
2508 	}
2509 	/*
2510 	 * The WQEBB space availability is checked by caller.
2511 	 * Here we should be aware of WQE ring buffer wraparound only.
2512 	 */
2513 	part = (uint8_t *)txq->wqes_end - pdst;
2514 	part = RTE_MIN(part, inlen);
2515 	do {
2516 		rte_memcpy(pdst, psrc, part);
2517 		inlen -= part;
2518 		if (likely(!inlen)) {
2519 			/*
2520 			 * If return value is not used by the caller
2521 			 * the code below will be optimized out.
2522 			 */
2523 			pdst += part;
2524 			pdst = RTE_PTR_ALIGN(pdst, MLX5_WSEG_SIZE);
2525 			if (unlikely(pdst >= (uint8_t *)txq->wqes_end))
2526 				pdst = (uint8_t *)txq->wqes;
2527 			return (struct mlx5_wqe_dseg *)pdst;
2528 		}
2529 		pdst = (uint8_t *)txq->wqes;
2530 		psrc += part;
2531 		part = inlen;
2532 	} while (true);
2533 }
2534 
2535 /**
2536  * Copy data from chain of mbuf to the specified linear buffer.
2537  * Checksums and VLAN insertion Tx offload features. If data
2538  * from some mbuf copied completely this mbuf is freed. Local
2539  * structure is used to keep the byte stream state.
2540  *
2541  * @param pdst
2542  *   Pointer to the destination linear buffer.
2543  * @param loc
2544  *   Pointer to burst routine local context.
2545  * @param len
2546  *   Length of data to be copied.
2547  * @param must
2548  *   Length of data to be copied ignoring no inline hint.
2549  * @param olx
2550  *   Configured Tx offloads mask. It is fully defined at
2551  *   compile time and may be used for optimization.
2552  *
2553  * @return
2554  *   Number of actual copied data bytes. This is always greater than or
2555  *   equal to must parameter and might be lesser than len in no inline
2556  *   hint flag is encountered.
2557  */
2558 static __rte_always_inline unsigned int
2559 mlx5_tx_mseg_memcpy(uint8_t *pdst,
2560 		    struct mlx5_txq_local *__rte_restrict loc,
2561 		    unsigned int len,
2562 		    unsigned int must,
2563 		    unsigned int olx __rte_unused)
2564 {
2565 	struct rte_mbuf *mbuf;
2566 	unsigned int part, dlen, copy = 0;
2567 	uint8_t *psrc;
2568 
2569 	MLX5_ASSERT(len);
2570 	MLX5_ASSERT(must <= len);
2571 	do {
2572 		/* Allow zero length packets, must check first. */
2573 		dlen = rte_pktmbuf_data_len(loc->mbuf);
2574 		if (dlen <= loc->mbuf_off) {
2575 			/* Exhausted packet, just free. */
2576 			mbuf = loc->mbuf;
2577 			loc->mbuf = mbuf->next;
2578 			rte_pktmbuf_free_seg(mbuf);
2579 			loc->mbuf_off = 0;
2580 			MLX5_ASSERT(loc->mbuf_nseg > 1);
2581 			MLX5_ASSERT(loc->mbuf);
2582 			--loc->mbuf_nseg;
2583 			if (loc->mbuf->ol_flags & PKT_TX_DYNF_NOINLINE) {
2584 				unsigned int diff;
2585 
2586 				if (copy >= must) {
2587 					/*
2588 					 * We already copied the minimal
2589 					 * requested amount of data.
2590 					 */
2591 					return copy;
2592 				}
2593 				diff = must - copy;
2594 				if (diff <= rte_pktmbuf_data_len(loc->mbuf)) {
2595 					/*
2596 					 * Copy only the minimal required
2597 					 * part of the data buffer.
2598 					 */
2599 					len = diff;
2600 				}
2601 			}
2602 			continue;
2603 		}
2604 		dlen -= loc->mbuf_off;
2605 		psrc = rte_pktmbuf_mtod_offset(loc->mbuf, uint8_t *,
2606 					       loc->mbuf_off);
2607 		part = RTE_MIN(len, dlen);
2608 		rte_memcpy(pdst, psrc, part);
2609 		copy += part;
2610 		loc->mbuf_off += part;
2611 		len -= part;
2612 		if (!len) {
2613 			if (loc->mbuf_off >= rte_pktmbuf_data_len(loc->mbuf)) {
2614 				loc->mbuf_off = 0;
2615 				/* Exhausted packet, just free. */
2616 				mbuf = loc->mbuf;
2617 				loc->mbuf = mbuf->next;
2618 				rte_pktmbuf_free_seg(mbuf);
2619 				loc->mbuf_off = 0;
2620 				MLX5_ASSERT(loc->mbuf_nseg >= 1);
2621 				--loc->mbuf_nseg;
2622 			}
2623 			return copy;
2624 		}
2625 		pdst += part;
2626 	} while (true);
2627 }
2628 
2629 /**
2630  * Build the Ethernet Segment with inlined data from
2631  * multi-segment packet. Checks the boundary of WQEBB
2632  * and ring buffer wrapping, supports Software Parser,
2633  * Checksums and VLAN insertion Tx offload features.
2634  *
2635  * @param txq
2636  *   Pointer to TX queue structure.
2637  * @param loc
2638  *   Pointer to burst routine local context.
2639  * @param wqe
2640  *   Pointer to WQE to fill with built Ethernet Segment.
2641  * @param vlan
2642  *   Length of VLAN tag insertion if any.
2643  * @param inlen
2644  *   Length of data to inline (VLAN included, if any).
2645  * @param tso
2646  *   TSO flag, set mss field from the packet.
2647  * @param olx
2648  *   Configured Tx offloads mask. It is fully defined at
2649  *   compile time and may be used for optimization.
2650  *
2651  * @return
2652  *   Pointer to the next Data Segment (aligned and
2653  *   possible NOT wrapped around - caller should do
2654  *   wrapping check on its own).
2655  */
2656 static __rte_always_inline struct mlx5_wqe_dseg *
2657 mlx5_tx_eseg_mdat(struct mlx5_txq_data *__rte_restrict txq,
2658 		  struct mlx5_txq_local *__rte_restrict loc,
2659 		  struct mlx5_wqe *__rte_restrict wqe,
2660 		  unsigned int vlan,
2661 		  unsigned int inlen,
2662 		  unsigned int tso,
2663 		  unsigned int olx)
2664 {
2665 	struct mlx5_wqe_eseg *__rte_restrict es = &wqe->eseg;
2666 	uint32_t csum;
2667 	uint8_t *pdst;
2668 	unsigned int part, tlen = 0;
2669 
2670 	/*
2671 	 * Calculate and set check sum flags first, uint32_t field
2672 	 * in segment may be shared with Software Parser flags.
2673 	 */
2674 	csum = MLX5_TXOFF_CONFIG(CSUM) ? txq_ol_cksum_to_cs(loc->mbuf) : 0;
2675 	if (tso) {
2676 		csum <<= 24;
2677 		csum |= loc->mbuf->tso_segsz;
2678 		es->flags = rte_cpu_to_be_32(csum);
2679 	} else {
2680 		es->flags = rte_cpu_to_le_32(csum);
2681 	}
2682 	/*
2683 	 * Calculate and set Software Parser offsets and flags.
2684 	 * These flags a set for custom UDP and IP tunnel packets.
2685 	 */
2686 	es->swp_offs = txq_mbuf_to_swp(loc, &es->swp_flags, olx);
2687 	/* Fill metadata field if needed. */
2688 	es->metadata = MLX5_TXOFF_CONFIG(METADATA) ?
2689 		       loc->mbuf->ol_flags & PKT_TX_DYNF_METADATA ?
2690 		       *RTE_FLOW_DYNF_METADATA(loc->mbuf) : 0 : 0;
2691 	static_assert(MLX5_ESEG_MIN_INLINE_SIZE ==
2692 				(sizeof(uint16_t) +
2693 				 sizeof(rte_v128u32_t)),
2694 		      "invalid Ethernet Segment data size");
2695 	static_assert(MLX5_ESEG_MIN_INLINE_SIZE ==
2696 				(sizeof(uint16_t) +
2697 				 sizeof(struct rte_vlan_hdr) +
2698 				 2 * RTE_ETHER_ADDR_LEN),
2699 		      "invalid Ethernet Segment data size");
2700 	MLX5_ASSERT(inlen >= MLX5_ESEG_MIN_INLINE_SIZE);
2701 	pdst = (uint8_t *)&es->inline_data;
2702 	if (MLX5_TXOFF_CONFIG(VLAN) && vlan) {
2703 		/* Implement VLAN tag insertion as part inline data. */
2704 		mlx5_tx_mseg_memcpy(pdst, loc,
2705 				    2 * RTE_ETHER_ADDR_LEN,
2706 				    2 * RTE_ETHER_ADDR_LEN, olx);
2707 		pdst += 2 * RTE_ETHER_ADDR_LEN;
2708 		*(unaligned_uint32_t *)pdst = rte_cpu_to_be_32
2709 						((RTE_ETHER_TYPE_VLAN << 16) |
2710 						 loc->mbuf->vlan_tci);
2711 		pdst += sizeof(struct rte_vlan_hdr);
2712 		tlen += 2 * RTE_ETHER_ADDR_LEN + sizeof(struct rte_vlan_hdr);
2713 	}
2714 	MLX5_ASSERT(pdst < (uint8_t *)txq->wqes_end);
2715 	/*
2716 	 * The WQEBB space availability is checked by caller.
2717 	 * Here we should be aware of WQE ring buffer wraparound only.
2718 	 */
2719 	part = (uint8_t *)txq->wqes_end - pdst;
2720 	part = RTE_MIN(part, inlen - tlen);
2721 	MLX5_ASSERT(part);
2722 	do {
2723 		unsigned int copy;
2724 
2725 		/*
2726 		 * Copying may be interrupted inside the routine
2727 		 * if run into no inline hint flag.
2728 		 */
2729 		copy = tlen >= txq->inlen_mode ? 0 : (txq->inlen_mode - tlen);
2730 		copy = mlx5_tx_mseg_memcpy(pdst, loc, part, copy, olx);
2731 		tlen += copy;
2732 		if (likely(inlen <= tlen) || copy < part) {
2733 			es->inline_hdr_sz = rte_cpu_to_be_16(tlen);
2734 			pdst += copy;
2735 			pdst = RTE_PTR_ALIGN(pdst, MLX5_WSEG_SIZE);
2736 			return (struct mlx5_wqe_dseg *)pdst;
2737 		}
2738 		pdst = (uint8_t *)txq->wqes;
2739 		part = inlen - tlen;
2740 	} while (true);
2741 }
2742 
2743 /**
2744  * Build the Data Segment of pointer type.
2745  *
2746  * @param txq
2747  *   Pointer to TX queue structure.
2748  * @param loc
2749  *   Pointer to burst routine local context.
2750  * @param dseg
2751  *   Pointer to WQE to fill with built Data Segment.
2752  * @param buf
2753  *   Data buffer to point.
2754  * @param len
2755  *   Data buffer length.
2756  * @param olx
2757  *   Configured Tx offloads mask. It is fully defined at
2758  *   compile time and may be used for optimization.
2759  */
2760 static __rte_always_inline void
2761 mlx5_tx_dseg_ptr(struct mlx5_txq_data *__rte_restrict txq,
2762 		 struct mlx5_txq_local *__rte_restrict loc,
2763 		 struct mlx5_wqe_dseg *__rte_restrict dseg,
2764 		 uint8_t *buf,
2765 		 unsigned int len,
2766 		 unsigned int olx __rte_unused)
2767 
2768 {
2769 	MLX5_ASSERT(len);
2770 	dseg->bcount = rte_cpu_to_be_32(len);
2771 	dseg->lkey = mlx5_tx_mb2mr(txq, loc->mbuf);
2772 	dseg->pbuf = rte_cpu_to_be_64((uintptr_t)buf);
2773 }
2774 
2775 /**
2776  * Build the Data Segment of pointer type or inline
2777  * if data length is less than buffer in minimal
2778  * Data Segment size.
2779  *
2780  * @param txq
2781  *   Pointer to TX queue structure.
2782  * @param loc
2783  *   Pointer to burst routine local context.
2784  * @param dseg
2785  *   Pointer to WQE to fill with built Data Segment.
2786  * @param buf
2787  *   Data buffer to point.
2788  * @param len
2789  *   Data buffer length.
2790  * @param olx
2791  *   Configured Tx offloads mask. It is fully defined at
2792  *   compile time and may be used for optimization.
2793  */
2794 static __rte_always_inline void
2795 mlx5_tx_dseg_iptr(struct mlx5_txq_data *__rte_restrict txq,
2796 		  struct mlx5_txq_local *__rte_restrict loc,
2797 		  struct mlx5_wqe_dseg *__rte_restrict dseg,
2798 		  uint8_t *buf,
2799 		  unsigned int len,
2800 		  unsigned int olx __rte_unused)
2801 
2802 {
2803 	uintptr_t dst, src;
2804 
2805 	MLX5_ASSERT(len);
2806 	if (len > MLX5_DSEG_MIN_INLINE_SIZE) {
2807 		dseg->bcount = rte_cpu_to_be_32(len);
2808 		dseg->lkey = mlx5_tx_mb2mr(txq, loc->mbuf);
2809 		dseg->pbuf = rte_cpu_to_be_64((uintptr_t)buf);
2810 
2811 		return;
2812 	}
2813 	dseg->bcount = rte_cpu_to_be_32(len | MLX5_ETH_WQE_DATA_INLINE);
2814 	/* Unrolled implementation of generic rte_memcpy. */
2815 	dst = (uintptr_t)&dseg->inline_data[0];
2816 	src = (uintptr_t)buf;
2817 	if (len & 0x08) {
2818 #ifdef RTE_ARCH_STRICT_ALIGN
2819 		MLX5_ASSERT(dst == RTE_PTR_ALIGN(dst, sizeof(uint32_t)));
2820 		*(uint32_t *)dst = *(unaligned_uint32_t *)src;
2821 		dst += sizeof(uint32_t);
2822 		src += sizeof(uint32_t);
2823 		*(uint32_t *)dst = *(unaligned_uint32_t *)src;
2824 		dst += sizeof(uint32_t);
2825 		src += sizeof(uint32_t);
2826 #else
2827 		*(uint64_t *)dst = *(unaligned_uint64_t *)src;
2828 		dst += sizeof(uint64_t);
2829 		src += sizeof(uint64_t);
2830 #endif
2831 	}
2832 	if (len & 0x04) {
2833 		*(uint32_t *)dst = *(unaligned_uint32_t *)src;
2834 		dst += sizeof(uint32_t);
2835 		src += sizeof(uint32_t);
2836 	}
2837 	if (len & 0x02) {
2838 		*(uint16_t *)dst = *(unaligned_uint16_t *)src;
2839 		dst += sizeof(uint16_t);
2840 		src += sizeof(uint16_t);
2841 	}
2842 	if (len & 0x01)
2843 		*(uint8_t *)dst = *(uint8_t *)src;
2844 }
2845 
2846 /**
2847  * Build the Data Segment of inlined data from single
2848  * segment packet, no VLAN insertion.
2849  *
2850  * @param txq
2851  *   Pointer to TX queue structure.
2852  * @param loc
2853  *   Pointer to burst routine local context.
2854  * @param dseg
2855  *   Pointer to WQE to fill with built Data Segment.
2856  * @param buf
2857  *   Data buffer to point.
2858  * @param len
2859  *   Data buffer length.
2860  * @param olx
2861  *   Configured Tx offloads mask. It is fully defined at
2862  *   compile time and may be used for optimization.
2863  *
2864  * @return
2865  *   Pointer to the next Data Segment after inlined data.
2866  *   Ring buffer wraparound check is needed. We do not
2867  *   do it here because it may not be needed for the
2868  *   last packet in the eMPW session.
2869  */
2870 static __rte_always_inline struct mlx5_wqe_dseg *
2871 mlx5_tx_dseg_empw(struct mlx5_txq_data *__rte_restrict txq,
2872 		  struct mlx5_txq_local *__rte_restrict loc __rte_unused,
2873 		  struct mlx5_wqe_dseg *__rte_restrict dseg,
2874 		  uint8_t *buf,
2875 		  unsigned int len,
2876 		  unsigned int olx __rte_unused)
2877 {
2878 	unsigned int part;
2879 	uint8_t *pdst;
2880 
2881 	if (!MLX5_TXOFF_CONFIG(MPW)) {
2882 		/* Store the descriptor byte counter for eMPW sessions. */
2883 		dseg->bcount = rte_cpu_to_be_32(len | MLX5_ETH_WQE_DATA_INLINE);
2884 		pdst = &dseg->inline_data[0];
2885 	} else {
2886 		/* The entire legacy MPW session counter is stored on close. */
2887 		pdst = (uint8_t *)dseg;
2888 	}
2889 	/*
2890 	 * The WQEBB space availability is checked by caller.
2891 	 * Here we should be aware of WQE ring buffer wraparound only.
2892 	 */
2893 	part = (uint8_t *)txq->wqes_end - pdst;
2894 	part = RTE_MIN(part, len);
2895 	do {
2896 		rte_memcpy(pdst, buf, part);
2897 		len -= part;
2898 		if (likely(!len)) {
2899 			pdst += part;
2900 			if (!MLX5_TXOFF_CONFIG(MPW))
2901 				pdst = RTE_PTR_ALIGN(pdst, MLX5_WSEG_SIZE);
2902 			/* Note: no final wraparound check here. */
2903 			return (struct mlx5_wqe_dseg *)pdst;
2904 		}
2905 		pdst = (uint8_t *)txq->wqes;
2906 		buf += part;
2907 		part = len;
2908 	} while (true);
2909 }
2910 
2911 /**
2912  * Build the Data Segment of inlined data from single
2913  * segment packet with VLAN insertion.
2914  *
2915  * @param txq
2916  *   Pointer to TX queue structure.
2917  * @param loc
2918  *   Pointer to burst routine local context.
2919  * @param dseg
2920  *   Pointer to the dseg fill with built Data Segment.
2921  * @param buf
2922  *   Data buffer to point.
2923  * @param len
2924  *   Data buffer length.
2925  * @param olx
2926  *   Configured Tx offloads mask. It is fully defined at
2927  *   compile time and may be used for optimization.
2928  *
2929  * @return
2930  *   Pointer to the next Data Segment after inlined data.
2931  *   Ring buffer wraparound check is needed.
2932  */
2933 static __rte_always_inline struct mlx5_wqe_dseg *
2934 mlx5_tx_dseg_vlan(struct mlx5_txq_data *__rte_restrict txq,
2935 		  struct mlx5_txq_local *__rte_restrict loc __rte_unused,
2936 		  struct mlx5_wqe_dseg *__rte_restrict dseg,
2937 		  uint8_t *buf,
2938 		  unsigned int len,
2939 		  unsigned int olx __rte_unused)
2940 
2941 {
2942 	unsigned int part;
2943 	uint8_t *pdst;
2944 
2945 	MLX5_ASSERT(len > MLX5_ESEG_MIN_INLINE_SIZE);
2946 	static_assert(MLX5_DSEG_MIN_INLINE_SIZE ==
2947 				 (2 * RTE_ETHER_ADDR_LEN),
2948 		      "invalid Data Segment data size");
2949 	if (!MLX5_TXOFF_CONFIG(MPW)) {
2950 		/* Store the descriptor byte counter for eMPW sessions. */
2951 		dseg->bcount = rte_cpu_to_be_32
2952 				((len + sizeof(struct rte_vlan_hdr)) |
2953 				 MLX5_ETH_WQE_DATA_INLINE);
2954 		pdst = &dseg->inline_data[0];
2955 	} else {
2956 		/* The entire legacy MPW session counter is stored on close. */
2957 		pdst = (uint8_t *)dseg;
2958 	}
2959 	memcpy(pdst, buf, MLX5_DSEG_MIN_INLINE_SIZE);
2960 	buf += MLX5_DSEG_MIN_INLINE_SIZE;
2961 	pdst += MLX5_DSEG_MIN_INLINE_SIZE;
2962 	len -= MLX5_DSEG_MIN_INLINE_SIZE;
2963 	/* Insert VLAN ethertype + VLAN tag. Pointer is aligned. */
2964 	MLX5_ASSERT(pdst == RTE_PTR_ALIGN(pdst, MLX5_WSEG_SIZE));
2965 	if (unlikely(pdst >= (uint8_t *)txq->wqes_end))
2966 		pdst = (uint8_t *)txq->wqes;
2967 	*(uint32_t *)pdst = rte_cpu_to_be_32((RTE_ETHER_TYPE_VLAN << 16) |
2968 					      loc->mbuf->vlan_tci);
2969 	pdst += sizeof(struct rte_vlan_hdr);
2970 	/*
2971 	 * The WQEBB space availability is checked by caller.
2972 	 * Here we should be aware of WQE ring buffer wraparound only.
2973 	 */
2974 	part = (uint8_t *)txq->wqes_end - pdst;
2975 	part = RTE_MIN(part, len);
2976 	do {
2977 		rte_memcpy(pdst, buf, part);
2978 		len -= part;
2979 		if (likely(!len)) {
2980 			pdst += part;
2981 			if (!MLX5_TXOFF_CONFIG(MPW))
2982 				pdst = RTE_PTR_ALIGN(pdst, MLX5_WSEG_SIZE);
2983 			/* Note: no final wraparound check here. */
2984 			return (struct mlx5_wqe_dseg *)pdst;
2985 		}
2986 		pdst = (uint8_t *)txq->wqes;
2987 		buf += part;
2988 		part = len;
2989 	} while (true);
2990 }
2991 
2992 /**
2993  * Build the Ethernet Segment with optionally inlined data with
2994  * VLAN insertion and following Data Segments (if any) from
2995  * multi-segment packet. Used by ordinary send and TSO.
2996  *
2997  * @param txq
2998  *   Pointer to TX queue structure.
2999  * @param loc
3000  *   Pointer to burst routine local context.
3001  * @param wqe
3002  *   Pointer to WQE to fill with built Ethernet/Data Segments.
3003  * @param vlan
3004  *   Length of VLAN header to insert, 0 means no VLAN insertion.
3005  * @param inlen
3006  *   Data length to inline. For TSO this parameter specifies
3007  *   exact value, for ordinary send routine can be aligned by
3008  *   caller to provide better WQE space saving and data buffer
3009  *   start address alignment. This length includes VLAN header
3010  *   being inserted.
3011  * @param tso
3012  *   Zero means ordinary send, inlined data can be extended,
3013  *   otherwise this is TSO, inlined data length is fixed.
3014  * @param olx
3015  *   Configured Tx offloads mask. It is fully defined at
3016  *   compile time and may be used for optimization.
3017  *
3018  * @return
3019  *   Actual size of built WQE in segments.
3020  */
3021 static __rte_always_inline unsigned int
3022 mlx5_tx_mseg_build(struct mlx5_txq_data *__rte_restrict txq,
3023 		   struct mlx5_txq_local *__rte_restrict loc,
3024 		   struct mlx5_wqe *__rte_restrict wqe,
3025 		   unsigned int vlan,
3026 		   unsigned int inlen,
3027 		   unsigned int tso,
3028 		   unsigned int olx __rte_unused)
3029 {
3030 	struct mlx5_wqe_dseg *__rte_restrict dseg;
3031 	unsigned int ds;
3032 
3033 	MLX5_ASSERT((rte_pktmbuf_pkt_len(loc->mbuf) + vlan) >= inlen);
3034 	loc->mbuf_nseg = NB_SEGS(loc->mbuf);
3035 	loc->mbuf_off = 0;
3036 
3037 	dseg = mlx5_tx_eseg_mdat(txq, loc, wqe, vlan, inlen, tso, olx);
3038 	if (!loc->mbuf_nseg)
3039 		goto dseg_done;
3040 	/*
3041 	 * There are still some mbuf remaining, not inlined.
3042 	 * The first mbuf may be partially inlined and we
3043 	 * must process the possible non-zero data offset.
3044 	 */
3045 	if (loc->mbuf_off) {
3046 		unsigned int dlen;
3047 		uint8_t *dptr;
3048 
3049 		/*
3050 		 * Exhausted packets must be dropped before.
3051 		 * Non-zero offset means there are some data
3052 		 * remained in the packet.
3053 		 */
3054 		MLX5_ASSERT(loc->mbuf_off < rte_pktmbuf_data_len(loc->mbuf));
3055 		MLX5_ASSERT(rte_pktmbuf_data_len(loc->mbuf));
3056 		dptr = rte_pktmbuf_mtod_offset(loc->mbuf, uint8_t *,
3057 					       loc->mbuf_off);
3058 		dlen = rte_pktmbuf_data_len(loc->mbuf) - loc->mbuf_off;
3059 		/*
3060 		 * Build the pointer/minimal data Data Segment.
3061 		 * Do ring buffer wrapping check in advance.
3062 		 */
3063 		if ((uintptr_t)dseg >= (uintptr_t)txq->wqes_end)
3064 			dseg = (struct mlx5_wqe_dseg *)txq->wqes;
3065 		mlx5_tx_dseg_iptr(txq, loc, dseg, dptr, dlen, olx);
3066 		/* Store the mbuf to be freed on completion. */
3067 		MLX5_ASSERT(loc->elts_free);
3068 		txq->elts[txq->elts_head++ & txq->elts_m] = loc->mbuf;
3069 		--loc->elts_free;
3070 		++dseg;
3071 		if (--loc->mbuf_nseg == 0)
3072 			goto dseg_done;
3073 		loc->mbuf = loc->mbuf->next;
3074 		loc->mbuf_off = 0;
3075 	}
3076 	do {
3077 		if (unlikely(!rte_pktmbuf_data_len(loc->mbuf))) {
3078 			struct rte_mbuf *mbuf;
3079 
3080 			/* Zero length segment found, just skip. */
3081 			mbuf = loc->mbuf;
3082 			loc->mbuf = loc->mbuf->next;
3083 			rte_pktmbuf_free_seg(mbuf);
3084 			if (--loc->mbuf_nseg == 0)
3085 				break;
3086 		} else {
3087 			if ((uintptr_t)dseg >= (uintptr_t)txq->wqes_end)
3088 				dseg = (struct mlx5_wqe_dseg *)txq->wqes;
3089 			mlx5_tx_dseg_iptr
3090 				(txq, loc, dseg,
3091 				 rte_pktmbuf_mtod(loc->mbuf, uint8_t *),
3092 				 rte_pktmbuf_data_len(loc->mbuf), olx);
3093 			MLX5_ASSERT(loc->elts_free);
3094 			txq->elts[txq->elts_head++ & txq->elts_m] = loc->mbuf;
3095 			--loc->elts_free;
3096 			++dseg;
3097 			if (--loc->mbuf_nseg == 0)
3098 				break;
3099 			loc->mbuf = loc->mbuf->next;
3100 		}
3101 	} while (true);
3102 
3103 dseg_done:
3104 	/* Calculate actual segments used from the dseg pointer. */
3105 	if ((uintptr_t)wqe < (uintptr_t)dseg)
3106 		ds = ((uintptr_t)dseg - (uintptr_t)wqe) / MLX5_WSEG_SIZE;
3107 	else
3108 		ds = (((uintptr_t)dseg - (uintptr_t)wqe) +
3109 		      txq->wqe_s * MLX5_WQE_SIZE) / MLX5_WSEG_SIZE;
3110 	return ds;
3111 }
3112 
3113 /**
3114  * The routine checks timestamp flag in the current packet,
3115  * and push WAIT WQE into the queue if scheduling is required.
3116  *
3117  * @param txq
3118  *   Pointer to TX queue structure.
3119  * @param loc
3120  *   Pointer to burst routine local context.
3121  * @param olx
3122  *   Configured Tx offloads mask. It is fully defined at
3123  *   compile time and may be used for optimization.
3124  *
3125  * @return
3126  *   MLX5_TXCMP_CODE_EXIT - sending is done or impossible.
3127  *   MLX5_TXCMP_CODE_SINGLE - continue processing with the packet.
3128  *   MLX5_TXCMP_CODE_MULTI - the WAIT inserted, continue processing.
3129  * Local context variables partially updated.
3130  */
3131 static __rte_always_inline enum mlx5_txcmp_code
3132 mlx5_tx_schedule_send(struct mlx5_txq_data *restrict txq,
3133 		      struct mlx5_txq_local *restrict loc,
3134 		      unsigned int olx)
3135 {
3136 	if (MLX5_TXOFF_CONFIG(TXPP) &&
3137 	    loc->mbuf->ol_flags & txq->ts_mask) {
3138 		struct mlx5_wqe *wqe;
3139 		uint64_t ts;
3140 		int32_t wci;
3141 
3142 		/*
3143 		 * Estimate the required space quickly and roughly.
3144 		 * We would like to ensure the packet can be pushed
3145 		 * to the queue and we won't get the orphan WAIT WQE.
3146 		 */
3147 		if (loc->wqe_free <= MLX5_WQE_SIZE_MAX / MLX5_WQE_SIZE ||
3148 		    loc->elts_free < NB_SEGS(loc->mbuf))
3149 			return MLX5_TXCMP_CODE_EXIT;
3150 		/* Convert the timestamp into completion to wait. */
3151 		ts = *RTE_MBUF_DYNFIELD(loc->mbuf, txq->ts_offset, uint64_t *);
3152 		wci = mlx5_txpp_convert_tx_ts(txq->sh, ts);
3153 		if (unlikely(wci < 0))
3154 			return MLX5_TXCMP_CODE_SINGLE;
3155 		/* Build the WAIT WQE with specified completion. */
3156 		wqe = txq->wqes + (txq->wqe_ci & txq->wqe_m);
3157 		mlx5_tx_cseg_init(txq, loc, wqe, 2, MLX5_OPCODE_WAIT, olx);
3158 		mlx5_tx_wseg_init(txq, loc, wqe, wci, olx);
3159 		++txq->wqe_ci;
3160 		--loc->wqe_free;
3161 		return MLX5_TXCMP_CODE_MULTI;
3162 	}
3163 	return MLX5_TXCMP_CODE_SINGLE;
3164 }
3165 
3166 /**
3167  * Tx one packet function for multi-segment TSO. Supports all
3168  * types of Tx offloads, uses MLX5_OPCODE_TSO to build WQEs,
3169  * sends one packet per WQE.
3170  *
3171  * This routine is responsible for storing processed mbuf
3172  * into elts ring buffer and update elts_head.
3173  *
3174  * @param txq
3175  *   Pointer to TX queue structure.
3176  * @param loc
3177  *   Pointer to burst routine local context.
3178  * @param olx
3179  *   Configured Tx offloads mask. It is fully defined at
3180  *   compile time and may be used for optimization.
3181  *
3182  * @return
3183  *   MLX5_TXCMP_CODE_EXIT - sending is done or impossible.
3184  *   MLX5_TXCMP_CODE_ERROR - some unrecoverable error occurred.
3185  * Local context variables partially updated.
3186  */
3187 static __rte_always_inline enum mlx5_txcmp_code
3188 mlx5_tx_packet_multi_tso(struct mlx5_txq_data *__rte_restrict txq,
3189 			struct mlx5_txq_local *__rte_restrict loc,
3190 			unsigned int olx)
3191 {
3192 	struct mlx5_wqe *__rte_restrict wqe;
3193 	unsigned int ds, dlen, inlen, ntcp, vlan = 0;
3194 
3195 	if (MLX5_TXOFF_CONFIG(TXPP)) {
3196 		enum mlx5_txcmp_code wret;
3197 
3198 		/* Generate WAIT for scheduling if requested. */
3199 		wret = mlx5_tx_schedule_send(txq, loc, olx);
3200 		if (wret == MLX5_TXCMP_CODE_EXIT)
3201 			return MLX5_TXCMP_CODE_EXIT;
3202 		if (wret == MLX5_TXCMP_CODE_ERROR)
3203 			return MLX5_TXCMP_CODE_ERROR;
3204 	}
3205 	/*
3206 	 * Calculate data length to be inlined to estimate
3207 	 * the required space in WQE ring buffer.
3208 	 */
3209 	dlen = rte_pktmbuf_pkt_len(loc->mbuf);
3210 	if (MLX5_TXOFF_CONFIG(VLAN) && loc->mbuf->ol_flags & PKT_TX_VLAN_PKT)
3211 		vlan = sizeof(struct rte_vlan_hdr);
3212 	inlen = loc->mbuf->l2_len + vlan +
3213 		loc->mbuf->l3_len + loc->mbuf->l4_len;
3214 	if (unlikely((!inlen || !loc->mbuf->tso_segsz)))
3215 		return MLX5_TXCMP_CODE_ERROR;
3216 	if (loc->mbuf->ol_flags & PKT_TX_TUNNEL_MASK)
3217 		inlen += loc->mbuf->outer_l2_len + loc->mbuf->outer_l3_len;
3218 	/* Packet must contain all TSO headers. */
3219 	if (unlikely(inlen > MLX5_MAX_TSO_HEADER ||
3220 		     inlen <= MLX5_ESEG_MIN_INLINE_SIZE ||
3221 		     inlen > (dlen + vlan)))
3222 		return MLX5_TXCMP_CODE_ERROR;
3223 	MLX5_ASSERT(inlen >= txq->inlen_mode);
3224 	/*
3225 	 * Check whether there are enough free WQEBBs:
3226 	 * - Control Segment
3227 	 * - Ethernet Segment
3228 	 * - First Segment of inlined Ethernet data
3229 	 * - ... data continued ...
3230 	 * - Data Segments of pointer/min inline type
3231 	 */
3232 	ds = NB_SEGS(loc->mbuf) + 2 + (inlen -
3233 				       MLX5_ESEG_MIN_INLINE_SIZE +
3234 				       MLX5_WSEG_SIZE +
3235 				       MLX5_WSEG_SIZE - 1) / MLX5_WSEG_SIZE;
3236 	if (unlikely(loc->wqe_free < ((ds + 3) / 4)))
3237 		return MLX5_TXCMP_CODE_EXIT;
3238 	/* Check for maximal WQE size. */
3239 	if (unlikely((MLX5_WQE_SIZE_MAX / MLX5_WSEG_SIZE) < ((ds + 3) / 4)))
3240 		return MLX5_TXCMP_CODE_ERROR;
3241 #ifdef MLX5_PMD_SOFT_COUNTERS
3242 	/* Update sent data bytes/packets counters. */
3243 	ntcp = (dlen - (inlen - vlan) + loc->mbuf->tso_segsz - 1) /
3244 		loc->mbuf->tso_segsz;
3245 	/*
3246 	 * One will be added for mbuf itself
3247 	 * at the end of the mlx5_tx_burst from
3248 	 * loc->pkts_sent field.
3249 	 */
3250 	--ntcp;
3251 	txq->stats.opackets += ntcp;
3252 	txq->stats.obytes += dlen + vlan + ntcp * inlen;
3253 #endif
3254 	wqe = txq->wqes + (txq->wqe_ci & txq->wqe_m);
3255 	loc->wqe_last = wqe;
3256 	mlx5_tx_cseg_init(txq, loc, wqe, 0, MLX5_OPCODE_TSO, olx);
3257 	ds = mlx5_tx_mseg_build(txq, loc, wqe, vlan, inlen, 1, olx);
3258 	wqe->cseg.sq_ds = rte_cpu_to_be_32(txq->qp_num_8s | ds);
3259 	txq->wqe_ci += (ds + 3) / 4;
3260 	loc->wqe_free -= (ds + 3) / 4;
3261 	return MLX5_TXCMP_CODE_MULTI;
3262 }
3263 
3264 /**
3265  * Tx one packet function for multi-segment SEND. Supports all
3266  * types of Tx offloads, uses MLX5_OPCODE_SEND to build WQEs,
3267  * sends one packet per WQE, without any data inlining in
3268  * Ethernet Segment.
3269  *
3270  * This routine is responsible for storing processed mbuf
3271  * into elts ring buffer and update elts_head.
3272  *
3273  * @param txq
3274  *   Pointer to TX queue structure.
3275  * @param loc
3276  *   Pointer to burst routine local context.
3277  * @param olx
3278  *   Configured Tx offloads mask. It is fully defined at
3279  *   compile time and may be used for optimization.
3280  *
3281  * @return
3282  *   MLX5_TXCMP_CODE_EXIT - sending is done or impossible.
3283  *   MLX5_TXCMP_CODE_ERROR - some unrecoverable error occurred.
3284  * Local context variables partially updated.
3285  */
3286 static __rte_always_inline enum mlx5_txcmp_code
3287 mlx5_tx_packet_multi_send(struct mlx5_txq_data *__rte_restrict txq,
3288 			  struct mlx5_txq_local *__rte_restrict loc,
3289 			  unsigned int olx)
3290 {
3291 	struct mlx5_wqe_dseg *__rte_restrict dseg;
3292 	struct mlx5_wqe *__rte_restrict wqe;
3293 	unsigned int ds, nseg;
3294 
3295 	MLX5_ASSERT(NB_SEGS(loc->mbuf) > 1);
3296 	if (MLX5_TXOFF_CONFIG(TXPP)) {
3297 		enum mlx5_txcmp_code wret;
3298 
3299 		/* Generate WAIT for scheduling if requested. */
3300 		wret = mlx5_tx_schedule_send(txq, loc, olx);
3301 		if (wret == MLX5_TXCMP_CODE_EXIT)
3302 			return MLX5_TXCMP_CODE_EXIT;
3303 		if (wret == MLX5_TXCMP_CODE_ERROR)
3304 			return MLX5_TXCMP_CODE_ERROR;
3305 	}
3306 	/*
3307 	 * No inline at all, it means the CPU cycles saving
3308 	 * is prioritized at configuration, we should not
3309 	 * copy any packet data to WQE.
3310 	 */
3311 	nseg = NB_SEGS(loc->mbuf);
3312 	ds = 2 + nseg;
3313 	if (unlikely(loc->wqe_free < ((ds + 3) / 4)))
3314 		return MLX5_TXCMP_CODE_EXIT;
3315 	/* Check for maximal WQE size. */
3316 	if (unlikely((MLX5_WQE_SIZE_MAX / MLX5_WSEG_SIZE) < ((ds + 3) / 4)))
3317 		return MLX5_TXCMP_CODE_ERROR;
3318 	/*
3319 	 * Some Tx offloads may cause an error if
3320 	 * packet is not long enough, check against
3321 	 * assumed minimal length.
3322 	 */
3323 	if (rte_pktmbuf_pkt_len(loc->mbuf) <= MLX5_ESEG_MIN_INLINE_SIZE)
3324 		return MLX5_TXCMP_CODE_ERROR;
3325 #ifdef MLX5_PMD_SOFT_COUNTERS
3326 	/* Update sent data bytes counter. */
3327 	txq->stats.obytes += rte_pktmbuf_pkt_len(loc->mbuf);
3328 	if (MLX5_TXOFF_CONFIG(VLAN) &&
3329 	    loc->mbuf->ol_flags & PKT_TX_VLAN_PKT)
3330 		txq->stats.obytes += sizeof(struct rte_vlan_hdr);
3331 #endif
3332 	/*
3333 	 * SEND WQE, one WQEBB:
3334 	 * - Control Segment, SEND opcode
3335 	 * - Ethernet Segment, optional VLAN, no inline
3336 	 * - Data Segments, pointer only type
3337 	 */
3338 	wqe = txq->wqes + (txq->wqe_ci & txq->wqe_m);
3339 	loc->wqe_last = wqe;
3340 	mlx5_tx_cseg_init(txq, loc, wqe, ds, MLX5_OPCODE_SEND, olx);
3341 	mlx5_tx_eseg_none(txq, loc, wqe, olx);
3342 	dseg = &wqe->dseg[0];
3343 	do {
3344 		if (unlikely(!rte_pktmbuf_data_len(loc->mbuf))) {
3345 			struct rte_mbuf *mbuf;
3346 
3347 			/*
3348 			 * Zero length segment found, have to
3349 			 * correct total size of WQE in segments.
3350 			 * It is supposed to be rare occasion, so
3351 			 * in normal case (no zero length segments)
3352 			 * we avoid extra writing to the Control
3353 			 * Segment.
3354 			 */
3355 			--ds;
3356 			wqe->cseg.sq_ds -= RTE_BE32(1);
3357 			mbuf = loc->mbuf;
3358 			loc->mbuf = mbuf->next;
3359 			rte_pktmbuf_free_seg(mbuf);
3360 			if (--nseg == 0)
3361 				break;
3362 		} else {
3363 			mlx5_tx_dseg_ptr
3364 				(txq, loc, dseg,
3365 				 rte_pktmbuf_mtod(loc->mbuf, uint8_t *),
3366 				 rte_pktmbuf_data_len(loc->mbuf), olx);
3367 			txq->elts[txq->elts_head++ & txq->elts_m] = loc->mbuf;
3368 			--loc->elts_free;
3369 			if (--nseg == 0)
3370 				break;
3371 			++dseg;
3372 			if ((uintptr_t)dseg >= (uintptr_t)txq->wqes_end)
3373 				dseg = (struct mlx5_wqe_dseg *)txq->wqes;
3374 			loc->mbuf = loc->mbuf->next;
3375 		}
3376 	} while (true);
3377 	txq->wqe_ci += (ds + 3) / 4;
3378 	loc->wqe_free -= (ds + 3) / 4;
3379 	return MLX5_TXCMP_CODE_MULTI;
3380 }
3381 
3382 /**
3383  * Tx one packet function for multi-segment SEND. Supports all
3384  * types of Tx offloads, uses MLX5_OPCODE_SEND to build WQEs,
3385  * sends one packet per WQE, with data inlining in
3386  * Ethernet Segment and minimal Data Segments.
3387  *
3388  * This routine is responsible for storing processed mbuf
3389  * into elts ring buffer and update elts_head.
3390  *
3391  * @param txq
3392  *   Pointer to TX queue structure.
3393  * @param loc
3394  *   Pointer to burst routine local context.
3395  * @param olx
3396  *   Configured Tx offloads mask. It is fully defined at
3397  *   compile time and may be used for optimization.
3398  *
3399  * @return
3400  *   MLX5_TXCMP_CODE_EXIT - sending is done or impossible.
3401  *   MLX5_TXCMP_CODE_ERROR - some unrecoverable error occurred.
3402  * Local context variables partially updated.
3403  */
3404 static __rte_always_inline enum mlx5_txcmp_code
3405 mlx5_tx_packet_multi_inline(struct mlx5_txq_data *__rte_restrict txq,
3406 			    struct mlx5_txq_local *__rte_restrict loc,
3407 			    unsigned int olx)
3408 {
3409 	struct mlx5_wqe *__rte_restrict wqe;
3410 	unsigned int ds, inlen, dlen, vlan = 0;
3411 
3412 	MLX5_ASSERT(MLX5_TXOFF_CONFIG(INLINE));
3413 	MLX5_ASSERT(NB_SEGS(loc->mbuf) > 1);
3414 	if (MLX5_TXOFF_CONFIG(TXPP)) {
3415 		enum mlx5_txcmp_code wret;
3416 
3417 		/* Generate WAIT for scheduling if requested. */
3418 		wret = mlx5_tx_schedule_send(txq, loc, olx);
3419 		if (wret == MLX5_TXCMP_CODE_EXIT)
3420 			return MLX5_TXCMP_CODE_EXIT;
3421 		if (wret == MLX5_TXCMP_CODE_ERROR)
3422 			return MLX5_TXCMP_CODE_ERROR;
3423 	}
3424 	/*
3425 	 * First calculate data length to be inlined
3426 	 * to estimate the required space for WQE.
3427 	 */
3428 	dlen = rte_pktmbuf_pkt_len(loc->mbuf);
3429 	if (MLX5_TXOFF_CONFIG(VLAN) && loc->mbuf->ol_flags & PKT_TX_VLAN_PKT)
3430 		vlan = sizeof(struct rte_vlan_hdr);
3431 	inlen = dlen + vlan;
3432 	/* Check against minimal length. */
3433 	if (inlen <= MLX5_ESEG_MIN_INLINE_SIZE)
3434 		return MLX5_TXCMP_CODE_ERROR;
3435 	MLX5_ASSERT(txq->inlen_send >= MLX5_ESEG_MIN_INLINE_SIZE);
3436 	if (inlen > txq->inlen_send ||
3437 	    loc->mbuf->ol_flags & PKT_TX_DYNF_NOINLINE) {
3438 		struct rte_mbuf *mbuf;
3439 		unsigned int nxlen;
3440 		uintptr_t start;
3441 
3442 		/*
3443 		 * Packet length exceeds the allowed inline
3444 		 * data length, check whether the minimal
3445 		 * inlining is required.
3446 		 */
3447 		if (txq->inlen_mode) {
3448 			MLX5_ASSERT(txq->inlen_mode >=
3449 				    MLX5_ESEG_MIN_INLINE_SIZE);
3450 			MLX5_ASSERT(txq->inlen_mode <= txq->inlen_send);
3451 			inlen = txq->inlen_mode;
3452 		} else {
3453 			if (loc->mbuf->ol_flags & PKT_TX_DYNF_NOINLINE ||
3454 			    !vlan || txq->vlan_en) {
3455 				/*
3456 				 * VLAN insertion will be done inside by HW.
3457 				 * It is not utmost effective - VLAN flag is
3458 				 * checked twice, but we should proceed the
3459 				 * inlining length correctly and take into
3460 				 * account the VLAN header being inserted.
3461 				 */
3462 				return mlx5_tx_packet_multi_send
3463 							(txq, loc, olx);
3464 			}
3465 			inlen = MLX5_ESEG_MIN_INLINE_SIZE;
3466 		}
3467 		/*
3468 		 * Now we know the minimal amount of data is requested
3469 		 * to inline. Check whether we should inline the buffers
3470 		 * from the chain beginning to eliminate some mbufs.
3471 		 */
3472 		mbuf = loc->mbuf;
3473 		nxlen = rte_pktmbuf_data_len(mbuf);
3474 		if (unlikely(nxlen <= txq->inlen_send)) {
3475 			/* We can inline first mbuf at least. */
3476 			if (nxlen < inlen) {
3477 				unsigned int smlen;
3478 
3479 				/* Scan mbufs till inlen filled. */
3480 				do {
3481 					smlen = nxlen;
3482 					mbuf = NEXT(mbuf);
3483 					MLX5_ASSERT(mbuf);
3484 					nxlen = rte_pktmbuf_data_len(mbuf);
3485 					nxlen += smlen;
3486 				} while (unlikely(nxlen < inlen));
3487 				if (unlikely(nxlen > txq->inlen_send)) {
3488 					/* We cannot inline entire mbuf. */
3489 					smlen = inlen - smlen;
3490 					start = rte_pktmbuf_mtod_offset
3491 						    (mbuf, uintptr_t, smlen);
3492 					goto do_align;
3493 				}
3494 			}
3495 			do {
3496 				inlen = nxlen;
3497 				mbuf = NEXT(mbuf);
3498 				/* There should be not end of packet. */
3499 				MLX5_ASSERT(mbuf);
3500 				nxlen = inlen + rte_pktmbuf_data_len(mbuf);
3501 			} while (unlikely(nxlen < txq->inlen_send));
3502 		}
3503 		start = rte_pktmbuf_mtod(mbuf, uintptr_t);
3504 		/*
3505 		 * Check whether we can do inline to align start
3506 		 * address of data buffer to cacheline.
3507 		 */
3508 do_align:
3509 		start = (~start + 1) & (RTE_CACHE_LINE_SIZE - 1);
3510 		if (unlikely(start)) {
3511 			start += inlen;
3512 			if (start <= txq->inlen_send)
3513 				inlen = start;
3514 		}
3515 	}
3516 	/*
3517 	 * Check whether there are enough free WQEBBs:
3518 	 * - Control Segment
3519 	 * - Ethernet Segment
3520 	 * - First Segment of inlined Ethernet data
3521 	 * - ... data continued ...
3522 	 * - Data Segments of pointer/min inline type
3523 	 *
3524 	 * Estimate the number of Data Segments conservatively,
3525 	 * supposing no any mbufs is being freed during inlining.
3526 	 */
3527 	MLX5_ASSERT(inlen <= txq->inlen_send);
3528 	ds = NB_SEGS(loc->mbuf) + 2 + (inlen -
3529 				       MLX5_ESEG_MIN_INLINE_SIZE +
3530 				       MLX5_WSEG_SIZE +
3531 				       MLX5_WSEG_SIZE - 1) / MLX5_WSEG_SIZE;
3532 	if (unlikely(loc->wqe_free < ((ds + 3) / 4)))
3533 		return MLX5_TXCMP_CODE_EXIT;
3534 	/* Check for maximal WQE size. */
3535 	if (unlikely((MLX5_WQE_SIZE_MAX / MLX5_WSEG_SIZE) < ((ds + 3) / 4)))
3536 		return MLX5_TXCMP_CODE_ERROR;
3537 #ifdef MLX5_PMD_SOFT_COUNTERS
3538 	/* Update sent data bytes/packets counters. */
3539 	txq->stats.obytes += dlen + vlan;
3540 #endif
3541 	wqe = txq->wqes + (txq->wqe_ci & txq->wqe_m);
3542 	loc->wqe_last = wqe;
3543 	mlx5_tx_cseg_init(txq, loc, wqe, 0, MLX5_OPCODE_SEND, olx);
3544 	ds = mlx5_tx_mseg_build(txq, loc, wqe, vlan, inlen, 0, olx);
3545 	wqe->cseg.sq_ds = rte_cpu_to_be_32(txq->qp_num_8s | ds);
3546 	txq->wqe_ci += (ds + 3) / 4;
3547 	loc->wqe_free -= (ds + 3) / 4;
3548 	return MLX5_TXCMP_CODE_MULTI;
3549 }
3550 
3551 /**
3552  * Tx burst function for multi-segment packets. Supports all
3553  * types of Tx offloads, uses MLX5_OPCODE_SEND/TSO to build WQEs,
3554  * sends one packet per WQE. Function stops sending if it
3555  * encounters the single-segment packet.
3556  *
3557  * This routine is responsible for storing processed mbuf
3558  * into elts ring buffer and update elts_head.
3559  *
3560  * @param txq
3561  *   Pointer to TX queue structure.
3562  * @param[in] pkts
3563  *   Packets to transmit.
3564  * @param pkts_n
3565  *   Number of packets in array.
3566  * @param loc
3567  *   Pointer to burst routine local context.
3568  * @param olx
3569  *   Configured Tx offloads mask. It is fully defined at
3570  *   compile time and may be used for optimization.
3571  *
3572  * @return
3573  *   MLX5_TXCMP_CODE_EXIT - sending is done or impossible.
3574  *   MLX5_TXCMP_CODE_ERROR - some unrecoverable error occurred.
3575  *   MLX5_TXCMP_CODE_SINGLE - single-segment packet encountered.
3576  *   MLX5_TXCMP_CODE_TSO - TSO single-segment packet encountered.
3577  * Local context variables updated.
3578  */
3579 static __rte_always_inline enum mlx5_txcmp_code
3580 mlx5_tx_burst_mseg(struct mlx5_txq_data *__rte_restrict txq,
3581 		   struct rte_mbuf **__rte_restrict pkts,
3582 		   unsigned int pkts_n,
3583 		   struct mlx5_txq_local *__rte_restrict loc,
3584 		   unsigned int olx)
3585 {
3586 	MLX5_ASSERT(loc->elts_free && loc->wqe_free);
3587 	MLX5_ASSERT(pkts_n > loc->pkts_sent);
3588 	pkts += loc->pkts_sent + 1;
3589 	pkts_n -= loc->pkts_sent;
3590 	for (;;) {
3591 		enum mlx5_txcmp_code ret;
3592 
3593 		MLX5_ASSERT(NB_SEGS(loc->mbuf) > 1);
3594 		/*
3595 		 * Estimate the number of free elts quickly but
3596 		 * conservatively. Some segment may be fully inlined
3597 		 * and freed, ignore this here - precise estimation
3598 		 * is costly.
3599 		 */
3600 		if (loc->elts_free < NB_SEGS(loc->mbuf))
3601 			return MLX5_TXCMP_CODE_EXIT;
3602 		if (MLX5_TXOFF_CONFIG(TSO) &&
3603 		    unlikely(loc->mbuf->ol_flags & PKT_TX_TCP_SEG)) {
3604 			/* Proceed with multi-segment TSO. */
3605 			ret = mlx5_tx_packet_multi_tso(txq, loc, olx);
3606 		} else if (MLX5_TXOFF_CONFIG(INLINE)) {
3607 			/* Proceed with multi-segment SEND with inlining. */
3608 			ret = mlx5_tx_packet_multi_inline(txq, loc, olx);
3609 		} else {
3610 			/* Proceed with multi-segment SEND w/o inlining. */
3611 			ret = mlx5_tx_packet_multi_send(txq, loc, olx);
3612 		}
3613 		if (ret == MLX5_TXCMP_CODE_EXIT)
3614 			return MLX5_TXCMP_CODE_EXIT;
3615 		if (ret == MLX5_TXCMP_CODE_ERROR)
3616 			return MLX5_TXCMP_CODE_ERROR;
3617 		/* WQE is built, go to the next packet. */
3618 		++loc->pkts_sent;
3619 		--pkts_n;
3620 		if (unlikely(!pkts_n || !loc->elts_free || !loc->wqe_free))
3621 			return MLX5_TXCMP_CODE_EXIT;
3622 		loc->mbuf = *pkts++;
3623 		if (pkts_n > 1)
3624 			rte_prefetch0(*pkts);
3625 		if (likely(NB_SEGS(loc->mbuf) > 1))
3626 			continue;
3627 		/* Here ends the series of multi-segment packets. */
3628 		if (MLX5_TXOFF_CONFIG(TSO) &&
3629 		    unlikely(loc->mbuf->ol_flags & PKT_TX_TCP_SEG))
3630 			return MLX5_TXCMP_CODE_TSO;
3631 		return MLX5_TXCMP_CODE_SINGLE;
3632 	}
3633 	MLX5_ASSERT(false);
3634 }
3635 
3636 /**
3637  * Tx burst function for single-segment packets with TSO.
3638  * Supports all types of Tx offloads, except multi-packets.
3639  * Uses MLX5_OPCODE_TSO to build WQEs, sends one packet per WQE.
3640  * Function stops sending if it encounters the multi-segment
3641  * packet or packet without TSO requested.
3642  *
3643  * The routine is responsible for storing processed mbuf
3644  * into elts ring buffer and update elts_head if inline
3645  * offloads is requested due to possible early freeing
3646  * of the inlined mbufs (can not store pkts array in elts
3647  * as a batch).
3648  *
3649  * @param txq
3650  *   Pointer to TX queue structure.
3651  * @param[in] pkts
3652  *   Packets to transmit.
3653  * @param pkts_n
3654  *   Number of packets in array.
3655  * @param loc
3656  *   Pointer to burst routine local context.
3657  * @param olx
3658  *   Configured Tx offloads mask. It is fully defined at
3659  *   compile time and may be used for optimization.
3660  *
3661  * @return
3662  *   MLX5_TXCMP_CODE_EXIT - sending is done or impossible.
3663  *   MLX5_TXCMP_CODE_ERROR - some unrecoverable error occurred.
3664  *   MLX5_TXCMP_CODE_SINGLE - single-segment packet encountered.
3665  *   MLX5_TXCMP_CODE_MULTI - multi-segment packet encountered.
3666  * Local context variables updated.
3667  */
3668 static __rte_always_inline enum mlx5_txcmp_code
3669 mlx5_tx_burst_tso(struct mlx5_txq_data *__rte_restrict txq,
3670 		  struct rte_mbuf **__rte_restrict pkts,
3671 		  unsigned int pkts_n,
3672 		  struct mlx5_txq_local *__rte_restrict loc,
3673 		  unsigned int olx)
3674 {
3675 	MLX5_ASSERT(loc->elts_free && loc->wqe_free);
3676 	MLX5_ASSERT(pkts_n > loc->pkts_sent);
3677 	pkts += loc->pkts_sent + 1;
3678 	pkts_n -= loc->pkts_sent;
3679 	for (;;) {
3680 		struct mlx5_wqe_dseg *__rte_restrict dseg;
3681 		struct mlx5_wqe *__rte_restrict wqe;
3682 		unsigned int ds, dlen, hlen, ntcp, vlan = 0;
3683 		uint8_t *dptr;
3684 
3685 		MLX5_ASSERT(NB_SEGS(loc->mbuf) == 1);
3686 		if (MLX5_TXOFF_CONFIG(TXPP)) {
3687 			enum mlx5_txcmp_code wret;
3688 
3689 			/* Generate WAIT for scheduling if requested. */
3690 			wret = mlx5_tx_schedule_send(txq, loc, olx);
3691 			if (wret == MLX5_TXCMP_CODE_EXIT)
3692 				return MLX5_TXCMP_CODE_EXIT;
3693 			if (wret == MLX5_TXCMP_CODE_ERROR)
3694 				return MLX5_TXCMP_CODE_ERROR;
3695 		}
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 *__rte_restrict txq,
3804 		     struct mlx5_txq_local *__rte_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 *__rte_restrict txq,
3859 		   struct mlx5_wqe_eseg *__rte_restrict es,
3860 		   struct mlx5_txq_local *__rte_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 	/* Check if the scheduling is requested. */
3888 	if (MLX5_TXOFF_CONFIG(TXPP) &&
3889 	    loc->mbuf->ol_flags & txq->ts_mask)
3890 		return false;
3891 	return true;
3892 }
3893 
3894 /*
3895  * Update send loop variables and WQE for eMPW loop
3896  * without data inlining. Number of Data Segments is
3897  * equal to the number of sent packets.
3898  *
3899  * @param txq
3900  *   Pointer to TX queue structure.
3901  * @param loc
3902  *   Pointer to burst routine local context.
3903  * @param ds
3904  *   Number of packets/Data Segments/Packets.
3905  * @param slen
3906  *   Accumulated statistics, bytes sent
3907  * @param olx
3908  *   Configured Tx offloads mask. It is fully defined at
3909  *   compile time and may be used for optimization.
3910  *
3911  * @return
3912  *  true - packet match with eMPW batch attributes.
3913  *  false - no match, eMPW should be restarted.
3914  */
3915 static __rte_always_inline void
3916 mlx5_tx_sdone_empw(struct mlx5_txq_data *__rte_restrict txq,
3917 		   struct mlx5_txq_local *__rte_restrict loc,
3918 		   unsigned int ds,
3919 		   unsigned int slen,
3920 		   unsigned int olx __rte_unused)
3921 {
3922 	MLX5_ASSERT(!MLX5_TXOFF_CONFIG(INLINE));
3923 #ifdef MLX5_PMD_SOFT_COUNTERS
3924 	/* Update sent data bytes counter. */
3925 	 txq->stats.obytes += slen;
3926 #else
3927 	(void)slen;
3928 #endif
3929 	loc->elts_free -= ds;
3930 	loc->pkts_sent += ds;
3931 	ds += 2;
3932 	loc->wqe_last->cseg.sq_ds = rte_cpu_to_be_32(txq->qp_num_8s | ds);
3933 	txq->wqe_ci += (ds + 3) / 4;
3934 	loc->wqe_free -= (ds + 3) / 4;
3935 }
3936 
3937 /*
3938  * Update send loop variables and WQE for eMPW loop
3939  * with data inlining. Gets the size of pushed descriptors
3940  * and data to the WQE.
3941  *
3942  * @param txq
3943  *   Pointer to TX queue structure.
3944  * @param loc
3945  *   Pointer to burst routine local context.
3946  * @param len
3947  *   Total size of descriptor/data in bytes.
3948  * @param slen
3949  *   Accumulated statistics, data bytes sent.
3950  * @param wqem
3951  *   The base WQE for the eMPW/MPW descriptor.
3952  * @param olx
3953  *   Configured Tx offloads mask. It is fully defined at
3954  *   compile time and may be used for optimization.
3955  *
3956  * @return
3957  *  true - packet match with eMPW batch attributes.
3958  *  false - no match, eMPW should be restarted.
3959  */
3960 static __rte_always_inline void
3961 mlx5_tx_idone_empw(struct mlx5_txq_data *__rte_restrict txq,
3962 		   struct mlx5_txq_local *__rte_restrict loc,
3963 		   unsigned int len,
3964 		   unsigned int slen,
3965 		   struct mlx5_wqe *__rte_restrict wqem,
3966 		   unsigned int olx __rte_unused)
3967 {
3968 	struct mlx5_wqe_dseg *dseg = &wqem->dseg[0];
3969 
3970 	MLX5_ASSERT(MLX5_TXOFF_CONFIG(INLINE));
3971 #ifdef MLX5_PMD_SOFT_COUNTERS
3972 	/* Update sent data bytes counter. */
3973 	 txq->stats.obytes += slen;
3974 #else
3975 	(void)slen;
3976 #endif
3977 	if (MLX5_TXOFF_CONFIG(MPW) && dseg->bcount == RTE_BE32(0)) {
3978 		/*
3979 		 * If the legacy MPW session contains the inline packets
3980 		 * we should set the only inline data segment length
3981 		 * and align the total length to the segment size.
3982 		 */
3983 		MLX5_ASSERT(len > sizeof(dseg->bcount));
3984 		dseg->bcount = rte_cpu_to_be_32((len - sizeof(dseg->bcount)) |
3985 						MLX5_ETH_WQE_DATA_INLINE);
3986 		len = (len + MLX5_WSEG_SIZE - 1) / MLX5_WSEG_SIZE + 2;
3987 	} else {
3988 		/*
3989 		 * The session is not legacy MPW or contains the
3990 		 * data buffer pointer segments.
3991 		 */
3992 		MLX5_ASSERT((len % MLX5_WSEG_SIZE) == 0);
3993 		len = len / MLX5_WSEG_SIZE + 2;
3994 	}
3995 	wqem->cseg.sq_ds = rte_cpu_to_be_32(txq->qp_num_8s | len);
3996 	txq->wqe_ci += (len + 3) / 4;
3997 	loc->wqe_free -= (len + 3) / 4;
3998 	loc->wqe_last = wqem;
3999 }
4000 
4001 /**
4002  * The set of Tx burst functions for single-segment packets
4003  * without TSO and with Multi-Packet Writing feature support.
4004  * Supports all types of Tx offloads, except multi-packets
4005  * and TSO.
4006  *
4007  * Uses MLX5_OPCODE_EMPW to build WQEs if possible and sends
4008  * as many packet per WQE as it can. If eMPW is not configured
4009  * or packet can not be sent with eMPW (VLAN insertion) the
4010  * ordinary SEND opcode is used and only one packet placed
4011  * in WQE.
4012  *
4013  * Functions stop sending if it encounters the multi-segment
4014  * packet or packet with TSO requested.
4015  *
4016  * The routines are responsible for storing processed mbuf
4017  * into elts ring buffer and update elts_head if inlining
4018  * offload is requested. Otherwise the copying mbufs to elts
4019  * can be postponed and completed at the end of burst routine.
4020  *
4021  * @param txq
4022  *   Pointer to TX queue structure.
4023  * @param[in] pkts
4024  *   Packets to transmit.
4025  * @param pkts_n
4026  *   Number of packets in array.
4027  * @param loc
4028  *   Pointer to burst routine local context.
4029  * @param olx
4030  *   Configured Tx offloads mask. It is fully defined at
4031  *   compile time and may be used for optimization.
4032  *
4033  * @return
4034  *   MLX5_TXCMP_CODE_EXIT - sending is done or impossible.
4035  *   MLX5_TXCMP_CODE_ERROR - some unrecoverable error occurred.
4036  *   MLX5_TXCMP_CODE_MULTI - multi-segment packet encountered.
4037  *   MLX5_TXCMP_CODE_TSO - TSO packet encountered.
4038  *   MLX5_TXCMP_CODE_SINGLE - used inside functions set.
4039  *   MLX5_TXCMP_CODE_EMPW - used inside functions set.
4040  *
4041  * Local context variables updated.
4042  *
4043  *
4044  * The routine sends packets with MLX5_OPCODE_EMPW
4045  * without inlining, this is dedicated optimized branch.
4046  * No VLAN insertion is supported.
4047  */
4048 static __rte_always_inline enum mlx5_txcmp_code
4049 mlx5_tx_burst_empw_simple(struct mlx5_txq_data *__rte_restrict txq,
4050 			  struct rte_mbuf **__rte_restrict pkts,
4051 			  unsigned int pkts_n,
4052 			  struct mlx5_txq_local *__rte_restrict loc,
4053 			  unsigned int olx)
4054 {
4055 	/*
4056 	 * Subroutine is the part of mlx5_tx_burst_single()
4057 	 * and sends single-segment packet with eMPW opcode
4058 	 * without data inlining.
4059 	 */
4060 	MLX5_ASSERT(!MLX5_TXOFF_CONFIG(INLINE));
4061 	MLX5_ASSERT(MLX5_TXOFF_CONFIG(EMPW));
4062 	MLX5_ASSERT(loc->elts_free && loc->wqe_free);
4063 	MLX5_ASSERT(pkts_n > loc->pkts_sent);
4064 	static_assert(MLX5_EMPW_MIN_PACKETS >= 2, "invalid min size");
4065 	pkts += loc->pkts_sent + 1;
4066 	pkts_n -= loc->pkts_sent;
4067 	for (;;) {
4068 		struct mlx5_wqe_dseg *__rte_restrict dseg;
4069 		struct mlx5_wqe_eseg *__rte_restrict eseg;
4070 		enum mlx5_txcmp_code ret;
4071 		unsigned int part, loop;
4072 		unsigned int slen = 0;
4073 
4074 next_empw:
4075 		MLX5_ASSERT(NB_SEGS(loc->mbuf) == 1);
4076 		if (MLX5_TXOFF_CONFIG(TXPP)) {
4077 			enum mlx5_txcmp_code wret;
4078 
4079 			/* Generate WAIT for scheduling if requested. */
4080 			wret = mlx5_tx_schedule_send(txq, loc, olx);
4081 			if (wret == MLX5_TXCMP_CODE_EXIT)
4082 				return MLX5_TXCMP_CODE_EXIT;
4083 			if (wret == MLX5_TXCMP_CODE_ERROR)
4084 				return MLX5_TXCMP_CODE_ERROR;
4085 		}
4086 		part = RTE_MIN(pkts_n, MLX5_TXOFF_CONFIG(MPW) ?
4087 				       MLX5_MPW_MAX_PACKETS :
4088 				       MLX5_EMPW_MAX_PACKETS);
4089 		if (unlikely(loc->elts_free < part)) {
4090 			/* We have no enough elts to save all mbufs. */
4091 			if (unlikely(loc->elts_free < MLX5_EMPW_MIN_PACKETS))
4092 				return MLX5_TXCMP_CODE_EXIT;
4093 			/* But we still able to send at least minimal eMPW. */
4094 			part = loc->elts_free;
4095 		}
4096 		/* Check whether we have enough WQEs */
4097 		if (unlikely(loc->wqe_free < ((2 + part + 3) / 4))) {
4098 			if (unlikely(loc->wqe_free <
4099 				((2 + MLX5_EMPW_MIN_PACKETS + 3) / 4)))
4100 				return MLX5_TXCMP_CODE_EXIT;
4101 			part = (loc->wqe_free * 4) - 2;
4102 		}
4103 		if (likely(part > 1))
4104 			rte_prefetch0(*pkts);
4105 		loc->wqe_last = txq->wqes + (txq->wqe_ci & txq->wqe_m);
4106 		/*
4107 		 * Build eMPW title WQEBB:
4108 		 * - Control Segment, eMPW opcode
4109 		 * - Ethernet Segment, no inline
4110 		 */
4111 		mlx5_tx_cseg_init(txq, loc, loc->wqe_last, part + 2,
4112 				  MLX5_OPCODE_ENHANCED_MPSW, olx);
4113 		mlx5_tx_eseg_none(txq, loc, loc->wqe_last,
4114 				  olx & ~MLX5_TXOFF_CONFIG_VLAN);
4115 		eseg = &loc->wqe_last->eseg;
4116 		dseg = &loc->wqe_last->dseg[0];
4117 		loop = part;
4118 		/* Store the packet length for legacy MPW. */
4119 		if (MLX5_TXOFF_CONFIG(MPW))
4120 			eseg->mss = rte_cpu_to_be_16
4121 					(rte_pktmbuf_data_len(loc->mbuf));
4122 		for (;;) {
4123 			uint32_t dlen = rte_pktmbuf_data_len(loc->mbuf);
4124 #ifdef MLX5_PMD_SOFT_COUNTERS
4125 			/* Update sent data bytes counter. */
4126 			slen += dlen;
4127 #endif
4128 			mlx5_tx_dseg_ptr
4129 				(txq, loc, dseg,
4130 				 rte_pktmbuf_mtod(loc->mbuf, uint8_t *),
4131 				 dlen, olx);
4132 			if (unlikely(--loop == 0))
4133 				break;
4134 			loc->mbuf = *pkts++;
4135 			if (likely(loop > 1))
4136 				rte_prefetch0(*pkts);
4137 			ret = mlx5_tx_able_to_empw(txq, loc, olx, true);
4138 			/*
4139 			 * Unroll the completion code to avoid
4140 			 * returning variable value - it results in
4141 			 * unoptimized sequent checking in caller.
4142 			 */
4143 			if (ret == MLX5_TXCMP_CODE_MULTI) {
4144 				part -= loop;
4145 				mlx5_tx_sdone_empw(txq, loc, part, slen, olx);
4146 				if (unlikely(!loc->elts_free ||
4147 					     !loc->wqe_free))
4148 					return MLX5_TXCMP_CODE_EXIT;
4149 				return MLX5_TXCMP_CODE_MULTI;
4150 			}
4151 			MLX5_ASSERT(NB_SEGS(loc->mbuf) == 1);
4152 			if (ret == MLX5_TXCMP_CODE_TSO) {
4153 				part -= loop;
4154 				mlx5_tx_sdone_empw(txq, loc, part, slen, olx);
4155 				if (unlikely(!loc->elts_free ||
4156 					     !loc->wqe_free))
4157 					return MLX5_TXCMP_CODE_EXIT;
4158 				return MLX5_TXCMP_CODE_TSO;
4159 			}
4160 			if (ret == MLX5_TXCMP_CODE_SINGLE) {
4161 				part -= loop;
4162 				mlx5_tx_sdone_empw(txq, loc, part, slen, olx);
4163 				if (unlikely(!loc->elts_free ||
4164 					     !loc->wqe_free))
4165 					return MLX5_TXCMP_CODE_EXIT;
4166 				return MLX5_TXCMP_CODE_SINGLE;
4167 			}
4168 			if (ret != MLX5_TXCMP_CODE_EMPW) {
4169 				MLX5_ASSERT(false);
4170 				part -= loop;
4171 				mlx5_tx_sdone_empw(txq, loc, part, slen, olx);
4172 				return MLX5_TXCMP_CODE_ERROR;
4173 			}
4174 			/*
4175 			 * Check whether packet parameters coincide
4176 			 * within assumed eMPW batch:
4177 			 * - check sum settings
4178 			 * - metadata value
4179 			 * - software parser settings
4180 			 * - packets length (legacy MPW only)
4181 			 * - scheduling is not required
4182 			 */
4183 			if (!mlx5_tx_match_empw(txq, eseg, loc, dlen, olx)) {
4184 				MLX5_ASSERT(loop);
4185 				part -= loop;
4186 				mlx5_tx_sdone_empw(txq, loc, part, slen, olx);
4187 				if (unlikely(!loc->elts_free ||
4188 					     !loc->wqe_free))
4189 					return MLX5_TXCMP_CODE_EXIT;
4190 				pkts_n -= part;
4191 				goto next_empw;
4192 			}
4193 			/* Packet attributes match, continue the same eMPW. */
4194 			++dseg;
4195 			if ((uintptr_t)dseg >= (uintptr_t)txq->wqes_end)
4196 				dseg = (struct mlx5_wqe_dseg *)txq->wqes;
4197 		}
4198 		/* eMPW is built successfully, update loop parameters. */
4199 		MLX5_ASSERT(!loop);
4200 		MLX5_ASSERT(pkts_n >= part);
4201 #ifdef MLX5_PMD_SOFT_COUNTERS
4202 		/* Update sent data bytes counter. */
4203 		txq->stats.obytes += slen;
4204 #endif
4205 		loc->elts_free -= part;
4206 		loc->pkts_sent += part;
4207 		txq->wqe_ci += (2 + part + 3) / 4;
4208 		loc->wqe_free -= (2 + part + 3) / 4;
4209 		pkts_n -= part;
4210 		if (unlikely(!pkts_n || !loc->elts_free || !loc->wqe_free))
4211 			return MLX5_TXCMP_CODE_EXIT;
4212 		loc->mbuf = *pkts++;
4213 		ret = mlx5_tx_able_to_empw(txq, loc, olx, true);
4214 		if (unlikely(ret != MLX5_TXCMP_CODE_EMPW))
4215 			return ret;
4216 		/* Continue sending eMPW batches. */
4217 	}
4218 	MLX5_ASSERT(false);
4219 }
4220 
4221 /**
4222  * The routine sends packets with MLX5_OPCODE_EMPW
4223  * with inlining, optionally supports VLAN insertion.
4224  */
4225 static __rte_always_inline enum mlx5_txcmp_code
4226 mlx5_tx_burst_empw_inline(struct mlx5_txq_data *__rte_restrict txq,
4227 			  struct rte_mbuf **__rte_restrict pkts,
4228 			  unsigned int pkts_n,
4229 			  struct mlx5_txq_local *__rte_restrict loc,
4230 			  unsigned int olx)
4231 {
4232 	/*
4233 	 * Subroutine is the part of mlx5_tx_burst_single()
4234 	 * and sends single-segment packet with eMPW opcode
4235 	 * with data inlining.
4236 	 */
4237 	MLX5_ASSERT(MLX5_TXOFF_CONFIG(INLINE));
4238 	MLX5_ASSERT(MLX5_TXOFF_CONFIG(EMPW));
4239 	MLX5_ASSERT(loc->elts_free && loc->wqe_free);
4240 	MLX5_ASSERT(pkts_n > loc->pkts_sent);
4241 	static_assert(MLX5_EMPW_MIN_PACKETS >= 2, "invalid min size");
4242 	pkts += loc->pkts_sent + 1;
4243 	pkts_n -= loc->pkts_sent;
4244 	for (;;) {
4245 		struct mlx5_wqe_dseg *__rte_restrict dseg;
4246 		struct mlx5_wqe *__rte_restrict wqem;
4247 		enum mlx5_txcmp_code ret;
4248 		unsigned int room, part, nlim;
4249 		unsigned int slen = 0;
4250 
4251 		MLX5_ASSERT(NB_SEGS(loc->mbuf) == 1);
4252 		if (MLX5_TXOFF_CONFIG(TXPP)) {
4253 			enum mlx5_txcmp_code wret;
4254 
4255 			/* Generate WAIT for scheduling if requested. */
4256 			wret = mlx5_tx_schedule_send(txq, loc, olx);
4257 			if (wret == MLX5_TXCMP_CODE_EXIT)
4258 				return MLX5_TXCMP_CODE_EXIT;
4259 			if (wret == MLX5_TXCMP_CODE_ERROR)
4260 				return MLX5_TXCMP_CODE_ERROR;
4261 		}
4262 		/*
4263 		 * Limits the amount of packets in one WQE
4264 		 * to improve CQE latency generation.
4265 		 */
4266 		nlim = RTE_MIN(pkts_n, MLX5_TXOFF_CONFIG(MPW) ?
4267 				       MLX5_MPW_INLINE_MAX_PACKETS :
4268 				       MLX5_EMPW_MAX_PACKETS);
4269 		/* Check whether we have minimal amount WQEs */
4270 		if (unlikely(loc->wqe_free <
4271 			    ((2 + MLX5_EMPW_MIN_PACKETS + 3) / 4)))
4272 			return MLX5_TXCMP_CODE_EXIT;
4273 		if (likely(pkts_n > 1))
4274 			rte_prefetch0(*pkts);
4275 		wqem = txq->wqes + (txq->wqe_ci & txq->wqe_m);
4276 		/*
4277 		 * Build eMPW title WQEBB:
4278 		 * - Control Segment, eMPW opcode, zero DS
4279 		 * - Ethernet Segment, no inline
4280 		 */
4281 		mlx5_tx_cseg_init(txq, loc, wqem, 0,
4282 				  MLX5_OPCODE_ENHANCED_MPSW, olx);
4283 		mlx5_tx_eseg_none(txq, loc, wqem,
4284 				  olx & ~MLX5_TXOFF_CONFIG_VLAN);
4285 		dseg = &wqem->dseg[0];
4286 		/* Store the packet length for legacy MPW. */
4287 		if (MLX5_TXOFF_CONFIG(MPW))
4288 			wqem->eseg.mss = rte_cpu_to_be_16
4289 					 (rte_pktmbuf_data_len(loc->mbuf));
4290 		room = RTE_MIN(MLX5_WQE_SIZE_MAX / MLX5_WQE_SIZE,
4291 			       loc->wqe_free) * MLX5_WQE_SIZE -
4292 					MLX5_WQE_CSEG_SIZE -
4293 					MLX5_WQE_ESEG_SIZE;
4294 		/* Limit the room for legacy MPW sessions for performance. */
4295 		if (MLX5_TXOFF_CONFIG(MPW))
4296 			room = RTE_MIN(room,
4297 				       RTE_MAX(txq->inlen_empw +
4298 					       sizeof(dseg->bcount) +
4299 					       (MLX5_TXOFF_CONFIG(VLAN) ?
4300 					       sizeof(struct rte_vlan_hdr) : 0),
4301 					       MLX5_MPW_INLINE_MAX_PACKETS *
4302 					       MLX5_WQE_DSEG_SIZE));
4303 		/* Build WQE till we have space, packets and resources. */
4304 		part = room;
4305 		for (;;) {
4306 			uint32_t dlen = rte_pktmbuf_data_len(loc->mbuf);
4307 			uint8_t *dptr = rte_pktmbuf_mtod(loc->mbuf, uint8_t *);
4308 			unsigned int tlen;
4309 
4310 			MLX5_ASSERT(room >= MLX5_WQE_DSEG_SIZE);
4311 			MLX5_ASSERT((room % MLX5_WQE_DSEG_SIZE) == 0);
4312 			MLX5_ASSERT((uintptr_t)dseg < (uintptr_t)txq->wqes_end);
4313 			/*
4314 			 * Some Tx offloads may cause an error if
4315 			 * packet is not long enough, check against
4316 			 * assumed minimal length.
4317 			 */
4318 			if (unlikely(dlen <= MLX5_ESEG_MIN_INLINE_SIZE)) {
4319 				part -= room;
4320 				if (unlikely(!part))
4321 					return MLX5_TXCMP_CODE_ERROR;
4322 				/*
4323 				 * We have some successfully built
4324 				 * packet Data Segments to send.
4325 				 */
4326 				mlx5_tx_idone_empw(txq, loc, part,
4327 						   slen, wqem, olx);
4328 				return MLX5_TXCMP_CODE_ERROR;
4329 			}
4330 			/* Inline or not inline - that's the Question. */
4331 			if (dlen > txq->inlen_empw ||
4332 			    loc->mbuf->ol_flags & PKT_TX_DYNF_NOINLINE)
4333 				goto pointer_empw;
4334 			if (MLX5_TXOFF_CONFIG(MPW)) {
4335 				if (dlen > txq->inlen_send)
4336 					goto pointer_empw;
4337 				tlen = dlen;
4338 				if (part == room) {
4339 					/* Open new inline MPW session. */
4340 					tlen += sizeof(dseg->bcount);
4341 					dseg->bcount = RTE_BE32(0);
4342 					dseg = RTE_PTR_ADD
4343 						(dseg, sizeof(dseg->bcount));
4344 				} else {
4345 					/*
4346 					 * No pointer and inline descriptor
4347 					 * intermix for legacy MPW sessions.
4348 					 */
4349 					if (wqem->dseg[0].bcount)
4350 						break;
4351 				}
4352 			} else {
4353 				tlen = sizeof(dseg->bcount) + dlen;
4354 			}
4355 			/* Inline entire packet, optional VLAN insertion. */
4356 			if (MLX5_TXOFF_CONFIG(VLAN) &&
4357 			    loc->mbuf->ol_flags & PKT_TX_VLAN_PKT) {
4358 				/*
4359 				 * The packet length must be checked in
4360 				 * mlx5_tx_able_to_empw() and packet
4361 				 * fits into inline length guaranteed.
4362 				 */
4363 				MLX5_ASSERT((dlen +
4364 					     sizeof(struct rte_vlan_hdr)) <=
4365 					    txq->inlen_empw);
4366 				tlen += sizeof(struct rte_vlan_hdr);
4367 				if (room < tlen)
4368 					break;
4369 				dseg = mlx5_tx_dseg_vlan(txq, loc, dseg,
4370 							 dptr, dlen, olx);
4371 #ifdef MLX5_PMD_SOFT_COUNTERS
4372 				/* Update sent data bytes counter. */
4373 				slen +=	sizeof(struct rte_vlan_hdr);
4374 #endif
4375 			} else {
4376 				if (room < tlen)
4377 					break;
4378 				dseg = mlx5_tx_dseg_empw(txq, loc, dseg,
4379 							 dptr, dlen, olx);
4380 			}
4381 			if (!MLX5_TXOFF_CONFIG(MPW))
4382 				tlen = RTE_ALIGN(tlen, MLX5_WSEG_SIZE);
4383 			MLX5_ASSERT(room >= tlen);
4384 			room -= tlen;
4385 			/*
4386 			 * Packet data are completely inlined,
4387 			 * free the packet immediately.
4388 			 */
4389 			rte_pktmbuf_free_seg(loc->mbuf);
4390 			goto next_mbuf;
4391 pointer_empw:
4392 			/*
4393 			 * No pointer and inline descriptor
4394 			 * intermix for legacy MPW sessions.
4395 			 */
4396 			if (MLX5_TXOFF_CONFIG(MPW) &&
4397 			    part != room &&
4398 			    wqem->dseg[0].bcount == RTE_BE32(0))
4399 				break;
4400 			/*
4401 			 * Not inlinable VLAN packets are
4402 			 * proceeded outside of this routine.
4403 			 */
4404 			MLX5_ASSERT(room >= MLX5_WQE_DSEG_SIZE);
4405 			if (MLX5_TXOFF_CONFIG(VLAN))
4406 				MLX5_ASSERT(!(loc->mbuf->ol_flags &
4407 					    PKT_TX_VLAN_PKT));
4408 			mlx5_tx_dseg_ptr(txq, loc, dseg, dptr, dlen, olx);
4409 			/* We have to store mbuf in elts.*/
4410 			txq->elts[txq->elts_head++ & txq->elts_m] = loc->mbuf;
4411 			room -= MLX5_WQE_DSEG_SIZE;
4412 			/* Ring buffer wraparound is checked at the loop end.*/
4413 			++dseg;
4414 next_mbuf:
4415 #ifdef MLX5_PMD_SOFT_COUNTERS
4416 			/* Update sent data bytes counter. */
4417 			slen += dlen;
4418 #endif
4419 			loc->pkts_sent++;
4420 			loc->elts_free--;
4421 			pkts_n--;
4422 			if (unlikely(!pkts_n || !loc->elts_free)) {
4423 				/*
4424 				 * We have no resources/packets to
4425 				 * continue build descriptors.
4426 				 */
4427 				part -= room;
4428 				mlx5_tx_idone_empw(txq, loc, part,
4429 						   slen, wqem, olx);
4430 				return MLX5_TXCMP_CODE_EXIT;
4431 			}
4432 			loc->mbuf = *pkts++;
4433 			if (likely(pkts_n > 1))
4434 				rte_prefetch0(*pkts);
4435 			ret = mlx5_tx_able_to_empw(txq, loc, olx, true);
4436 			/*
4437 			 * Unroll the completion code to avoid
4438 			 * returning variable value - it results in
4439 			 * unoptimized sequent checking in caller.
4440 			 */
4441 			if (ret == MLX5_TXCMP_CODE_MULTI) {
4442 				part -= room;
4443 				mlx5_tx_idone_empw(txq, loc, part,
4444 						   slen, wqem, olx);
4445 				if (unlikely(!loc->elts_free ||
4446 					     !loc->wqe_free))
4447 					return MLX5_TXCMP_CODE_EXIT;
4448 				return MLX5_TXCMP_CODE_MULTI;
4449 			}
4450 			MLX5_ASSERT(NB_SEGS(loc->mbuf) == 1);
4451 			if (ret == MLX5_TXCMP_CODE_TSO) {
4452 				part -= room;
4453 				mlx5_tx_idone_empw(txq, loc, part,
4454 						   slen, wqem, olx);
4455 				if (unlikely(!loc->elts_free ||
4456 					     !loc->wqe_free))
4457 					return MLX5_TXCMP_CODE_EXIT;
4458 				return MLX5_TXCMP_CODE_TSO;
4459 			}
4460 			if (ret == MLX5_TXCMP_CODE_SINGLE) {
4461 				part -= room;
4462 				mlx5_tx_idone_empw(txq, loc, part,
4463 						   slen, wqem, olx);
4464 				if (unlikely(!loc->elts_free ||
4465 					     !loc->wqe_free))
4466 					return MLX5_TXCMP_CODE_EXIT;
4467 				return MLX5_TXCMP_CODE_SINGLE;
4468 			}
4469 			if (ret != MLX5_TXCMP_CODE_EMPW) {
4470 				MLX5_ASSERT(false);
4471 				part -= room;
4472 				mlx5_tx_idone_empw(txq, loc, part,
4473 						   slen, wqem, olx);
4474 				return MLX5_TXCMP_CODE_ERROR;
4475 			}
4476 			/* Check if we have minimal room left. */
4477 			nlim--;
4478 			if (unlikely(!nlim || room < MLX5_WQE_DSEG_SIZE))
4479 				break;
4480 			/*
4481 			 * Check whether packet parameters coincide
4482 			 * within assumed eMPW batch:
4483 			 * - check sum settings
4484 			 * - metadata value
4485 			 * - software parser settings
4486 			 * - packets length (legacy MPW only)
4487 			 * - scheduling is not required
4488 			 */
4489 			if (!mlx5_tx_match_empw(txq, &wqem->eseg,
4490 						loc, dlen, olx))
4491 				break;
4492 			/* Packet attributes match, continue the same eMPW. */
4493 			if ((uintptr_t)dseg >= (uintptr_t)txq->wqes_end)
4494 				dseg = (struct mlx5_wqe_dseg *)txq->wqes;
4495 		}
4496 		/*
4497 		 * We get here to close an existing eMPW
4498 		 * session and start the new one.
4499 		 */
4500 		MLX5_ASSERT(pkts_n);
4501 		part -= room;
4502 		if (unlikely(!part))
4503 			return MLX5_TXCMP_CODE_EXIT;
4504 		mlx5_tx_idone_empw(txq, loc, part, slen, wqem, olx);
4505 		if (unlikely(!loc->elts_free ||
4506 			     !loc->wqe_free))
4507 			return MLX5_TXCMP_CODE_EXIT;
4508 		/* Continue the loop with new eMPW session. */
4509 	}
4510 	MLX5_ASSERT(false);
4511 }
4512 
4513 /**
4514  * The routine sends packets with ordinary MLX5_OPCODE_SEND.
4515  * Data inlining and VLAN insertion are supported.
4516  */
4517 static __rte_always_inline enum mlx5_txcmp_code
4518 mlx5_tx_burst_single_send(struct mlx5_txq_data *__rte_restrict txq,
4519 			  struct rte_mbuf **__rte_restrict pkts,
4520 			  unsigned int pkts_n,
4521 			  struct mlx5_txq_local *__rte_restrict loc,
4522 			  unsigned int olx)
4523 {
4524 	/*
4525 	 * Subroutine is the part of mlx5_tx_burst_single()
4526 	 * and sends single-segment packet with SEND opcode.
4527 	 */
4528 	MLX5_ASSERT(loc->elts_free && loc->wqe_free);
4529 	MLX5_ASSERT(pkts_n > loc->pkts_sent);
4530 	pkts += loc->pkts_sent + 1;
4531 	pkts_n -= loc->pkts_sent;
4532 	for (;;) {
4533 		struct mlx5_wqe *__rte_restrict wqe;
4534 		enum mlx5_txcmp_code ret;
4535 
4536 		MLX5_ASSERT(NB_SEGS(loc->mbuf) == 1);
4537 		if (MLX5_TXOFF_CONFIG(TXPP)) {
4538 			enum mlx5_txcmp_code wret;
4539 
4540 			/* Generate WAIT for scheduling if requested. */
4541 			wret = mlx5_tx_schedule_send(txq, loc, olx);
4542 			if (wret == MLX5_TXCMP_CODE_EXIT)
4543 				return MLX5_TXCMP_CODE_EXIT;
4544 			if (wret == MLX5_TXCMP_CODE_ERROR)
4545 				return MLX5_TXCMP_CODE_ERROR;
4546 		}
4547 		if (MLX5_TXOFF_CONFIG(INLINE)) {
4548 			unsigned int inlen, vlan = 0;
4549 
4550 			inlen = rte_pktmbuf_data_len(loc->mbuf);
4551 			if (MLX5_TXOFF_CONFIG(VLAN) &&
4552 			    loc->mbuf->ol_flags & PKT_TX_VLAN_PKT) {
4553 				vlan = sizeof(struct rte_vlan_hdr);
4554 				inlen += vlan;
4555 				static_assert((sizeof(struct rte_vlan_hdr) +
4556 					       sizeof(struct rte_ether_hdr)) ==
4557 					       MLX5_ESEG_MIN_INLINE_SIZE,
4558 					       "invalid min inline data size");
4559 			}
4560 			/*
4561 			 * If inlining is enabled at configuration time
4562 			 * the limit must be not less than minimal size.
4563 			 * Otherwise we would do extra check for data
4564 			 * size to avoid crashes due to length overflow.
4565 			 */
4566 			MLX5_ASSERT(txq->inlen_send >=
4567 				    MLX5_ESEG_MIN_INLINE_SIZE);
4568 			if (inlen <= txq->inlen_send) {
4569 				unsigned int seg_n, wqe_n;
4570 
4571 				rte_prefetch0(rte_pktmbuf_mtod
4572 						(loc->mbuf, uint8_t *));
4573 				/* Check against minimal length. */
4574 				if (inlen <= MLX5_ESEG_MIN_INLINE_SIZE)
4575 					return MLX5_TXCMP_CODE_ERROR;
4576 				if (loc->mbuf->ol_flags &
4577 				    PKT_TX_DYNF_NOINLINE) {
4578 					/*
4579 					 * The hint flag not to inline packet
4580 					 * data is set. Check whether we can
4581 					 * follow the hint.
4582 					 */
4583 					if ((!MLX5_TXOFF_CONFIG(EMPW) &&
4584 					      txq->inlen_mode) ||
4585 					    (MLX5_TXOFF_CONFIG(MPW) &&
4586 					     txq->inlen_mode)) {
4587 						if (inlen <= txq->inlen_send)
4588 							goto single_inline;
4589 						/*
4590 						 * The hardware requires the
4591 						 * minimal inline data header.
4592 						 */
4593 						goto single_min_inline;
4594 					}
4595 					if (MLX5_TXOFF_CONFIG(VLAN) &&
4596 					    vlan && !txq->vlan_en) {
4597 						/*
4598 						 * We must insert VLAN tag
4599 						 * by software means.
4600 						 */
4601 						goto single_part_inline;
4602 					}
4603 					goto single_no_inline;
4604 				}
4605 single_inline:
4606 				/*
4607 				 * Completely inlined packet data WQE:
4608 				 * - Control Segment, SEND opcode
4609 				 * - Ethernet Segment, no VLAN insertion
4610 				 * - Data inlined, VLAN optionally inserted
4611 				 * - Alignment to MLX5_WSEG_SIZE
4612 				 * Have to estimate amount of WQEBBs
4613 				 */
4614 				seg_n = (inlen + 3 * MLX5_WSEG_SIZE -
4615 					 MLX5_ESEG_MIN_INLINE_SIZE +
4616 					 MLX5_WSEG_SIZE - 1) / MLX5_WSEG_SIZE;
4617 				/* Check if there are enough WQEBBs. */
4618 				wqe_n = (seg_n + 3) / 4;
4619 				if (wqe_n > loc->wqe_free)
4620 					return MLX5_TXCMP_CODE_EXIT;
4621 				wqe = txq->wqes + (txq->wqe_ci & txq->wqe_m);
4622 				loc->wqe_last = wqe;
4623 				mlx5_tx_cseg_init(txq, loc, wqe, seg_n,
4624 						  MLX5_OPCODE_SEND, olx);
4625 				mlx5_tx_eseg_data(txq, loc, wqe,
4626 						  vlan, inlen, 0, olx);
4627 				txq->wqe_ci += wqe_n;
4628 				loc->wqe_free -= wqe_n;
4629 				/*
4630 				 * Packet data are completely inlined,
4631 				 * free the packet immediately.
4632 				 */
4633 				rte_pktmbuf_free_seg(loc->mbuf);
4634 			} else if ((!MLX5_TXOFF_CONFIG(EMPW) ||
4635 				     MLX5_TXOFF_CONFIG(MPW)) &&
4636 					txq->inlen_mode) {
4637 				/*
4638 				 * If minimal inlining is requested the eMPW
4639 				 * feature should be disabled due to data is
4640 				 * inlined into Ethernet Segment, which can
4641 				 * not contain inlined data for eMPW due to
4642 				 * segment shared for all packets.
4643 				 */
4644 				struct mlx5_wqe_dseg *__rte_restrict dseg;
4645 				unsigned int ds;
4646 				uint8_t *dptr;
4647 
4648 				/*
4649 				 * The inline-mode settings require
4650 				 * to inline the specified amount of
4651 				 * data bytes to the Ethernet Segment.
4652 				 * We should check the free space in
4653 				 * WQE ring buffer to inline partially.
4654 				 */
4655 single_min_inline:
4656 				MLX5_ASSERT(txq->inlen_send >= txq->inlen_mode);
4657 				MLX5_ASSERT(inlen > txq->inlen_mode);
4658 				MLX5_ASSERT(txq->inlen_mode >=
4659 					    MLX5_ESEG_MIN_INLINE_SIZE);
4660 				/*
4661 				 * Check whether there are enough free WQEBBs:
4662 				 * - Control Segment
4663 				 * - Ethernet Segment
4664 				 * - First Segment of inlined Ethernet data
4665 				 * - ... data continued ...
4666 				 * - Finishing Data Segment of pointer type
4667 				 */
4668 				ds = (MLX5_WQE_CSEG_SIZE +
4669 				      MLX5_WQE_ESEG_SIZE +
4670 				      MLX5_WQE_DSEG_SIZE +
4671 				      txq->inlen_mode -
4672 				      MLX5_ESEG_MIN_INLINE_SIZE +
4673 				      MLX5_WQE_DSEG_SIZE +
4674 				      MLX5_WSEG_SIZE - 1) / MLX5_WSEG_SIZE;
4675 				if (loc->wqe_free < ((ds + 3) / 4))
4676 					return MLX5_TXCMP_CODE_EXIT;
4677 				/*
4678 				 * Build the ordinary SEND WQE:
4679 				 * - Control Segment
4680 				 * - Ethernet Segment, inline inlen_mode bytes
4681 				 * - Data Segment of pointer type
4682 				 */
4683 				wqe = txq->wqes + (txq->wqe_ci & txq->wqe_m);
4684 				loc->wqe_last = wqe;
4685 				mlx5_tx_cseg_init(txq, loc, wqe, ds,
4686 						  MLX5_OPCODE_SEND, olx);
4687 				dseg = mlx5_tx_eseg_data(txq, loc, wqe, vlan,
4688 							 txq->inlen_mode,
4689 							 0, olx);
4690 				dptr = rte_pktmbuf_mtod(loc->mbuf, uint8_t *) +
4691 				       txq->inlen_mode - vlan;
4692 				inlen -= txq->inlen_mode;
4693 				mlx5_tx_dseg_ptr(txq, loc, dseg,
4694 						 dptr, inlen, olx);
4695 				/*
4696 				 * WQE is built, update the loop parameters
4697 				 * and got to the next packet.
4698 				 */
4699 				txq->wqe_ci += (ds + 3) / 4;
4700 				loc->wqe_free -= (ds + 3) / 4;
4701 				/* We have to store mbuf in elts.*/
4702 				MLX5_ASSERT(MLX5_TXOFF_CONFIG(INLINE));
4703 				txq->elts[txq->elts_head++ & txq->elts_m] =
4704 						loc->mbuf;
4705 				--loc->elts_free;
4706 			} else {
4707 				uint8_t *dptr;
4708 				unsigned int dlen;
4709 
4710 				/*
4711 				 * Partially inlined packet data WQE, we have
4712 				 * some space in title WQEBB, we can fill it
4713 				 * with some packet data. It takes one WQEBB,
4714 				 * it is available, no extra space check:
4715 				 * - Control Segment, SEND opcode
4716 				 * - Ethernet Segment, no VLAN insertion
4717 				 * - MLX5_ESEG_MIN_INLINE_SIZE bytes of Data
4718 				 * - Data Segment, pointer type
4719 				 *
4720 				 * We also get here if VLAN insertion is not
4721 				 * supported by HW, the inline is enabled.
4722 				 */
4723 single_part_inline:
4724 				wqe = txq->wqes + (txq->wqe_ci & txq->wqe_m);
4725 				loc->wqe_last = wqe;
4726 				mlx5_tx_cseg_init(txq, loc, wqe, 4,
4727 						  MLX5_OPCODE_SEND, olx);
4728 				mlx5_tx_eseg_dmin(txq, loc, wqe, vlan, olx);
4729 				dptr = rte_pktmbuf_mtod(loc->mbuf, uint8_t *) +
4730 				       MLX5_ESEG_MIN_INLINE_SIZE - vlan;
4731 				/*
4732 				 * The length check is performed above, by
4733 				 * comparing with txq->inlen_send. We should
4734 				 * not get overflow here.
4735 				 */
4736 				MLX5_ASSERT(inlen > MLX5_ESEG_MIN_INLINE_SIZE);
4737 				dlen = inlen - MLX5_ESEG_MIN_INLINE_SIZE;
4738 				mlx5_tx_dseg_ptr(txq, loc, &wqe->dseg[1],
4739 						 dptr, dlen, olx);
4740 				++txq->wqe_ci;
4741 				--loc->wqe_free;
4742 				/* We have to store mbuf in elts.*/
4743 				MLX5_ASSERT(MLX5_TXOFF_CONFIG(INLINE));
4744 				txq->elts[txq->elts_head++ & txq->elts_m] =
4745 						loc->mbuf;
4746 				--loc->elts_free;
4747 			}
4748 #ifdef MLX5_PMD_SOFT_COUNTERS
4749 			/* Update sent data bytes counter. */
4750 			txq->stats.obytes += vlan +
4751 					rte_pktmbuf_data_len(loc->mbuf);
4752 #endif
4753 		} else {
4754 			/*
4755 			 * No inline at all, it means the CPU cycles saving
4756 			 * is prioritized at configuration, we should not
4757 			 * copy any packet data to WQE.
4758 			 *
4759 			 * SEND WQE, one WQEBB:
4760 			 * - Control Segment, SEND opcode
4761 			 * - Ethernet Segment, optional VLAN, no inline
4762 			 * - Data Segment, pointer type
4763 			 */
4764 single_no_inline:
4765 			wqe = txq->wqes + (txq->wqe_ci & txq->wqe_m);
4766 			loc->wqe_last = wqe;
4767 			mlx5_tx_cseg_init(txq, loc, wqe, 3,
4768 					  MLX5_OPCODE_SEND, olx);
4769 			mlx5_tx_eseg_none(txq, loc, wqe, olx);
4770 			mlx5_tx_dseg_ptr
4771 				(txq, loc, &wqe->dseg[0],
4772 				 rte_pktmbuf_mtod(loc->mbuf, uint8_t *),
4773 				 rte_pktmbuf_data_len(loc->mbuf), olx);
4774 			++txq->wqe_ci;
4775 			--loc->wqe_free;
4776 			/*
4777 			 * We should not store mbuf pointer in elts
4778 			 * if no inlining is configured, this is done
4779 			 * by calling routine in a batch copy.
4780 			 */
4781 			MLX5_ASSERT(!MLX5_TXOFF_CONFIG(INLINE));
4782 			--loc->elts_free;
4783 #ifdef MLX5_PMD_SOFT_COUNTERS
4784 			/* Update sent data bytes counter. */
4785 			txq->stats.obytes += rte_pktmbuf_data_len(loc->mbuf);
4786 			if (MLX5_TXOFF_CONFIG(VLAN) &&
4787 			    loc->mbuf->ol_flags & PKT_TX_VLAN_PKT)
4788 				txq->stats.obytes +=
4789 					sizeof(struct rte_vlan_hdr);
4790 #endif
4791 		}
4792 		++loc->pkts_sent;
4793 		--pkts_n;
4794 		if (unlikely(!pkts_n || !loc->elts_free || !loc->wqe_free))
4795 			return MLX5_TXCMP_CODE_EXIT;
4796 		loc->mbuf = *pkts++;
4797 		if (pkts_n > 1)
4798 			rte_prefetch0(*pkts);
4799 		ret = mlx5_tx_able_to_empw(txq, loc, olx, true);
4800 		if (unlikely(ret != MLX5_TXCMP_CODE_SINGLE))
4801 			return ret;
4802 	}
4803 	MLX5_ASSERT(false);
4804 }
4805 
4806 static __rte_always_inline enum mlx5_txcmp_code
4807 mlx5_tx_burst_single(struct mlx5_txq_data *__rte_restrict txq,
4808 		     struct rte_mbuf **__rte_restrict pkts,
4809 		     unsigned int pkts_n,
4810 		     struct mlx5_txq_local *__rte_restrict loc,
4811 		     unsigned int olx)
4812 {
4813 	enum mlx5_txcmp_code ret;
4814 
4815 	ret = mlx5_tx_able_to_empw(txq, loc, olx, false);
4816 	if (ret == MLX5_TXCMP_CODE_SINGLE)
4817 		goto ordinary_send;
4818 	MLX5_ASSERT(ret == MLX5_TXCMP_CODE_EMPW);
4819 	for (;;) {
4820 		/* Optimize for inline/no inline eMPW send. */
4821 		ret = (MLX5_TXOFF_CONFIG(INLINE)) ?
4822 			mlx5_tx_burst_empw_inline
4823 				(txq, pkts, pkts_n, loc, olx) :
4824 			mlx5_tx_burst_empw_simple
4825 				(txq, pkts, pkts_n, loc, olx);
4826 		if (ret != MLX5_TXCMP_CODE_SINGLE)
4827 			return ret;
4828 		/* The resources to send one packet should remain. */
4829 		MLX5_ASSERT(loc->elts_free && loc->wqe_free);
4830 ordinary_send:
4831 		ret = mlx5_tx_burst_single_send(txq, pkts, pkts_n, loc, olx);
4832 		MLX5_ASSERT(ret != MLX5_TXCMP_CODE_SINGLE);
4833 		if (ret != MLX5_TXCMP_CODE_EMPW)
4834 			return ret;
4835 		/* The resources to send one packet should remain. */
4836 		MLX5_ASSERT(loc->elts_free && loc->wqe_free);
4837 	}
4838 }
4839 
4840 /**
4841  * DPDK Tx callback template. This is configured template
4842  * used to generate routines optimized for specified offload setup.
4843  * One of this generated functions is chosen at SQ configuration
4844  * time.
4845  *
4846  * @param txq
4847  *   Generic pointer to TX queue structure.
4848  * @param[in] pkts
4849  *   Packets to transmit.
4850  * @param pkts_n
4851  *   Number of packets in array.
4852  * @param olx
4853  *   Configured offloads mask, presents the bits of MLX5_TXOFF_CONFIG_xxx
4854  *   values. Should be static to take compile time static configuration
4855  *   advantages.
4856  *
4857  * @return
4858  *   Number of packets successfully transmitted (<= pkts_n).
4859  */
4860 static __rte_always_inline uint16_t
4861 mlx5_tx_burst_tmpl(struct mlx5_txq_data *__rte_restrict txq,
4862 		   struct rte_mbuf **__rte_restrict pkts,
4863 		   uint16_t pkts_n,
4864 		   unsigned int olx)
4865 {
4866 	struct mlx5_txq_local loc;
4867 	enum mlx5_txcmp_code ret;
4868 	unsigned int part;
4869 
4870 	MLX5_ASSERT(txq->elts_s >= (uint16_t)(txq->elts_head - txq->elts_tail));
4871 	MLX5_ASSERT(txq->wqe_s >= (uint16_t)(txq->wqe_ci - txq->wqe_pi));
4872 	if (unlikely(!pkts_n))
4873 		return 0;
4874 	loc.pkts_sent = 0;
4875 	loc.pkts_copy = 0;
4876 	loc.wqe_last = NULL;
4877 
4878 send_loop:
4879 	loc.pkts_loop = loc.pkts_sent;
4880 	/*
4881 	 * Check if there are some CQEs, if any:
4882 	 * - process an encountered errors
4883 	 * - process the completed WQEs
4884 	 * - free related mbufs
4885 	 * - doorbell the NIC about processed CQEs
4886 	 */
4887 	rte_prefetch0(*(pkts + loc.pkts_sent));
4888 	mlx5_tx_handle_completion(txq, olx);
4889 	/*
4890 	 * Calculate the number of available resources - elts and WQEs.
4891 	 * There are two possible different scenarios:
4892 	 * - no data inlining into WQEs, one WQEBB may contains up to
4893 	 *   four packets, in this case elts become scarce resource
4894 	 * - data inlining into WQEs, one packet may require multiple
4895 	 *   WQEBBs, the WQEs become the limiting factor.
4896 	 */
4897 	MLX5_ASSERT(txq->elts_s >= (uint16_t)(txq->elts_head - txq->elts_tail));
4898 	loc.elts_free = txq->elts_s -
4899 				(uint16_t)(txq->elts_head - txq->elts_tail);
4900 	MLX5_ASSERT(txq->wqe_s >= (uint16_t)(txq->wqe_ci - txq->wqe_pi));
4901 	loc.wqe_free = txq->wqe_s -
4902 				(uint16_t)(txq->wqe_ci - txq->wqe_pi);
4903 	if (unlikely(!loc.elts_free || !loc.wqe_free))
4904 		goto burst_exit;
4905 	for (;;) {
4906 		/*
4907 		 * Fetch the packet from array. Usually this is
4908 		 * the first packet in series of multi/single
4909 		 * segment packets.
4910 		 */
4911 		loc.mbuf = *(pkts + loc.pkts_sent);
4912 		/* Dedicated branch for multi-segment packets. */
4913 		if (MLX5_TXOFF_CONFIG(MULTI) &&
4914 		    unlikely(NB_SEGS(loc.mbuf) > 1)) {
4915 			/*
4916 			 * Multi-segment packet encountered.
4917 			 * Hardware is able to process it only
4918 			 * with SEND/TSO opcodes, one packet
4919 			 * per WQE, do it in dedicated routine.
4920 			 */
4921 enter_send_multi:
4922 			MLX5_ASSERT(loc.pkts_sent >= loc.pkts_copy);
4923 			part = loc.pkts_sent - loc.pkts_copy;
4924 			if (!MLX5_TXOFF_CONFIG(INLINE) && part) {
4925 				/*
4926 				 * There are some single-segment mbufs not
4927 				 * stored in elts. The mbufs must be in the
4928 				 * same order as WQEs, so we must copy the
4929 				 * mbufs to elts here, before the coming
4930 				 * multi-segment packet mbufs is appended.
4931 				 */
4932 				mlx5_tx_copy_elts(txq, pkts + loc.pkts_copy,
4933 						  part, olx);
4934 				loc.pkts_copy = loc.pkts_sent;
4935 			}
4936 			MLX5_ASSERT(pkts_n > loc.pkts_sent);
4937 			ret = mlx5_tx_burst_mseg(txq, pkts, pkts_n, &loc, olx);
4938 			if (!MLX5_TXOFF_CONFIG(INLINE))
4939 				loc.pkts_copy = loc.pkts_sent;
4940 			/*
4941 			 * These returned code checks are supposed
4942 			 * to be optimized out due to routine inlining.
4943 			 */
4944 			if (ret == MLX5_TXCMP_CODE_EXIT) {
4945 				/*
4946 				 * The routine returns this code when
4947 				 * all packets are sent or there is no
4948 				 * enough resources to complete request.
4949 				 */
4950 				break;
4951 			}
4952 			if (ret == MLX5_TXCMP_CODE_ERROR) {
4953 				/*
4954 				 * The routine returns this code when
4955 				 * some error in the incoming packets
4956 				 * format occurred.
4957 				 */
4958 				txq->stats.oerrors++;
4959 				break;
4960 			}
4961 			if (ret == MLX5_TXCMP_CODE_SINGLE) {
4962 				/*
4963 				 * The single-segment packet was encountered
4964 				 * in the array, try to send it with the
4965 				 * best optimized way, possible engaging eMPW.
4966 				 */
4967 				goto enter_send_single;
4968 			}
4969 			if (MLX5_TXOFF_CONFIG(TSO) &&
4970 			    ret == MLX5_TXCMP_CODE_TSO) {
4971 				/*
4972 				 * The single-segment TSO packet was
4973 				 * encountered in the array.
4974 				 */
4975 				goto enter_send_tso;
4976 			}
4977 			/* We must not get here. Something is going wrong. */
4978 			MLX5_ASSERT(false);
4979 			txq->stats.oerrors++;
4980 			break;
4981 		}
4982 		/* Dedicated branch for single-segment TSO packets. */
4983 		if (MLX5_TXOFF_CONFIG(TSO) &&
4984 		    unlikely(loc.mbuf->ol_flags & PKT_TX_TCP_SEG)) {
4985 			/*
4986 			 * TSO might require special way for inlining
4987 			 * (dedicated parameters) and is sent with
4988 			 * MLX5_OPCODE_TSO opcode only, provide this
4989 			 * in dedicated branch.
4990 			 */
4991 enter_send_tso:
4992 			MLX5_ASSERT(NB_SEGS(loc.mbuf) == 1);
4993 			MLX5_ASSERT(pkts_n > loc.pkts_sent);
4994 			ret = mlx5_tx_burst_tso(txq, pkts, pkts_n, &loc, olx);
4995 			/*
4996 			 * These returned code checks are supposed
4997 			 * to be optimized out due to routine inlining.
4998 			 */
4999 			if (ret == MLX5_TXCMP_CODE_EXIT)
5000 				break;
5001 			if (ret == MLX5_TXCMP_CODE_ERROR) {
5002 				txq->stats.oerrors++;
5003 				break;
5004 			}
5005 			if (ret == MLX5_TXCMP_CODE_SINGLE)
5006 				goto enter_send_single;
5007 			if (MLX5_TXOFF_CONFIG(MULTI) &&
5008 			    ret == MLX5_TXCMP_CODE_MULTI) {
5009 				/*
5010 				 * The multi-segment packet was
5011 				 * encountered in the array.
5012 				 */
5013 				goto enter_send_multi;
5014 			}
5015 			/* We must not get here. Something is going wrong. */
5016 			MLX5_ASSERT(false);
5017 			txq->stats.oerrors++;
5018 			break;
5019 		}
5020 		/*
5021 		 * The dedicated branch for the single-segment packets
5022 		 * without TSO. Often these ones can be sent using
5023 		 * MLX5_OPCODE_EMPW with multiple packets in one WQE.
5024 		 * The routine builds the WQEs till it encounters
5025 		 * the TSO or multi-segment packet (in case if these
5026 		 * offloads are requested at SQ configuration time).
5027 		 */
5028 enter_send_single:
5029 		MLX5_ASSERT(pkts_n > loc.pkts_sent);
5030 		ret = mlx5_tx_burst_single(txq, pkts, pkts_n, &loc, olx);
5031 		/*
5032 		 * These returned code checks are supposed
5033 		 * to be optimized out due to routine inlining.
5034 		 */
5035 		if (ret == MLX5_TXCMP_CODE_EXIT)
5036 			break;
5037 		if (ret == MLX5_TXCMP_CODE_ERROR) {
5038 			txq->stats.oerrors++;
5039 			break;
5040 		}
5041 		if (MLX5_TXOFF_CONFIG(MULTI) &&
5042 		    ret == MLX5_TXCMP_CODE_MULTI) {
5043 			/*
5044 			 * The multi-segment packet was
5045 			 * encountered in the array.
5046 			 */
5047 			goto enter_send_multi;
5048 		}
5049 		if (MLX5_TXOFF_CONFIG(TSO) &&
5050 		    ret == MLX5_TXCMP_CODE_TSO) {
5051 			/*
5052 			 * The single-segment TSO packet was
5053 			 * encountered in the array.
5054 			 */
5055 			goto enter_send_tso;
5056 		}
5057 		/* We must not get here. Something is going wrong. */
5058 		MLX5_ASSERT(false);
5059 		txq->stats.oerrors++;
5060 		break;
5061 	}
5062 	/*
5063 	 * Main Tx loop is completed, do the rest:
5064 	 * - set completion request if thresholds are reached
5065 	 * - doorbell the hardware
5066 	 * - copy the rest of mbufs to elts (if any)
5067 	 */
5068 	MLX5_ASSERT(MLX5_TXOFF_CONFIG(INLINE) ||
5069 		    loc.pkts_sent >= loc.pkts_copy);
5070 	/* Take a shortcut if nothing is sent. */
5071 	if (unlikely(loc.pkts_sent == loc.pkts_loop))
5072 		goto burst_exit;
5073 	/* Request CQE generation if limits are reached. */
5074 	mlx5_tx_request_completion(txq, &loc, olx);
5075 	/*
5076 	 * Ring QP doorbell immediately after WQE building completion
5077 	 * to improve latencies. The pure software related data treatment
5078 	 * can be completed after doorbell. Tx CQEs for this SQ are
5079 	 * processed in this thread only by the polling.
5080 	 *
5081 	 * The rdma core library can map doorbell register in two ways,
5082 	 * depending on the environment variable "MLX5_SHUT_UP_BF":
5083 	 *
5084 	 * - as regular cached memory, the variable is either missing or
5085 	 *   set to zero. This type of mapping may cause the significant
5086 	 *   doorbell register writing latency and requires explicit
5087 	 *   memory write barrier to mitigate this issue and prevent
5088 	 *   write combining.
5089 	 *
5090 	 * - as non-cached memory, the variable is present and set to
5091 	 *   not "0" value. This type of mapping may cause performance
5092 	 *   impact under heavy loading conditions but the explicit write
5093 	 *   memory barrier is not required and it may improve core
5094 	 *   performance.
5095 	 *
5096 	 * - the legacy behaviour (prior 19.08 release) was to use some
5097 	 *   heuristics to decide whether write memory barrier should
5098 	 *   be performed. This behavior is supported with specifying
5099 	 *   tx_db_nc=2, write barrier is skipped if application
5100 	 *   provides the full recommended burst of packets, it
5101 	 *   supposes the next packets are coming and the write barrier
5102 	 *   will be issued on the next burst (after descriptor writing,
5103 	 *   at least).
5104 	 */
5105 	mlx5_tx_dbrec_cond_wmb(txq, loc.wqe_last, !txq->db_nc &&
5106 			(!txq->db_heu || pkts_n % MLX5_TX_DEFAULT_BURST));
5107 	/* Not all of the mbufs may be stored into elts yet. */
5108 	part = MLX5_TXOFF_CONFIG(INLINE) ? 0 : loc.pkts_sent - loc.pkts_copy;
5109 	if (!MLX5_TXOFF_CONFIG(INLINE) && part) {
5110 		/*
5111 		 * There are some single-segment mbufs not stored in elts.
5112 		 * It can be only if the last packet was single-segment.
5113 		 * The copying is gathered into one place due to it is
5114 		 * a good opportunity to optimize that with SIMD.
5115 		 * Unfortunately if inlining is enabled the gaps in
5116 		 * pointer array may happen due to early freeing of the
5117 		 * inlined mbufs.
5118 		 */
5119 		mlx5_tx_copy_elts(txq, pkts + loc.pkts_copy, part, olx);
5120 		loc.pkts_copy = loc.pkts_sent;
5121 	}
5122 	MLX5_ASSERT(txq->elts_s >= (uint16_t)(txq->elts_head - txq->elts_tail));
5123 	MLX5_ASSERT(txq->wqe_s >= (uint16_t)(txq->wqe_ci - txq->wqe_pi));
5124 	if (pkts_n > loc.pkts_sent) {
5125 		/*
5126 		 * If burst size is large there might be no enough CQE
5127 		 * fetched from completion queue and no enough resources
5128 		 * freed to send all the packets.
5129 		 */
5130 		goto send_loop;
5131 	}
5132 burst_exit:
5133 #ifdef MLX5_PMD_SOFT_COUNTERS
5134 	/* Increment sent packets counter. */
5135 	txq->stats.opackets += loc.pkts_sent;
5136 #endif
5137 	return loc.pkts_sent;
5138 }
5139 
5140 /* Generate routines with Enhanced Multi-Packet Write support. */
5141 MLX5_TXOFF_DECL(full_empw,
5142 		MLX5_TXOFF_CONFIG_FULL | MLX5_TXOFF_CONFIG_EMPW)
5143 
5144 MLX5_TXOFF_DECL(none_empw,
5145 		MLX5_TXOFF_CONFIG_NONE | MLX5_TXOFF_CONFIG_EMPW)
5146 
5147 MLX5_TXOFF_DECL(md_empw,
5148 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
5149 
5150 MLX5_TXOFF_DECL(mt_empw,
5151 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
5152 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
5153 
5154 MLX5_TXOFF_DECL(mtsc_empw,
5155 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
5156 		MLX5_TXOFF_CONFIG_SWP |	MLX5_TXOFF_CONFIG_CSUM |
5157 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
5158 
5159 MLX5_TXOFF_DECL(mti_empw,
5160 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
5161 		MLX5_TXOFF_CONFIG_INLINE |
5162 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
5163 
5164 MLX5_TXOFF_DECL(mtv_empw,
5165 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
5166 		MLX5_TXOFF_CONFIG_VLAN |
5167 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
5168 
5169 MLX5_TXOFF_DECL(mtiv_empw,
5170 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
5171 		MLX5_TXOFF_CONFIG_INLINE | MLX5_TXOFF_CONFIG_VLAN |
5172 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
5173 
5174 MLX5_TXOFF_DECL(sc_empw,
5175 		MLX5_TXOFF_CONFIG_SWP |	MLX5_TXOFF_CONFIG_CSUM |
5176 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
5177 
5178 MLX5_TXOFF_DECL(sci_empw,
5179 		MLX5_TXOFF_CONFIG_SWP |	MLX5_TXOFF_CONFIG_CSUM |
5180 		MLX5_TXOFF_CONFIG_INLINE |
5181 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
5182 
5183 MLX5_TXOFF_DECL(scv_empw,
5184 		MLX5_TXOFF_CONFIG_SWP |	MLX5_TXOFF_CONFIG_CSUM |
5185 		MLX5_TXOFF_CONFIG_VLAN |
5186 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
5187 
5188 MLX5_TXOFF_DECL(sciv_empw,
5189 		MLX5_TXOFF_CONFIG_SWP |	MLX5_TXOFF_CONFIG_CSUM |
5190 		MLX5_TXOFF_CONFIG_INLINE | MLX5_TXOFF_CONFIG_VLAN |
5191 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
5192 
5193 MLX5_TXOFF_DECL(i_empw,
5194 		MLX5_TXOFF_CONFIG_INLINE |
5195 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
5196 
5197 MLX5_TXOFF_DECL(v_empw,
5198 		MLX5_TXOFF_CONFIG_VLAN |
5199 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
5200 
5201 MLX5_TXOFF_DECL(iv_empw,
5202 		MLX5_TXOFF_CONFIG_INLINE | MLX5_TXOFF_CONFIG_VLAN |
5203 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
5204 
5205 /* Generate routines without Enhanced Multi-Packet Write support. */
5206 MLX5_TXOFF_DECL(full,
5207 		MLX5_TXOFF_CONFIG_FULL)
5208 
5209 MLX5_TXOFF_DECL(none,
5210 		MLX5_TXOFF_CONFIG_NONE)
5211 
5212 MLX5_TXOFF_DECL(md,
5213 		MLX5_TXOFF_CONFIG_METADATA)
5214 
5215 MLX5_TXOFF_DECL(mt,
5216 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
5217 		MLX5_TXOFF_CONFIG_METADATA)
5218 
5219 MLX5_TXOFF_DECL(mtsc,
5220 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
5221 		MLX5_TXOFF_CONFIG_SWP |	MLX5_TXOFF_CONFIG_CSUM |
5222 		MLX5_TXOFF_CONFIG_METADATA)
5223 
5224 MLX5_TXOFF_DECL(mti,
5225 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
5226 		MLX5_TXOFF_CONFIG_INLINE |
5227 		MLX5_TXOFF_CONFIG_METADATA)
5228 
5229 
5230 MLX5_TXOFF_DECL(mtv,
5231 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
5232 		MLX5_TXOFF_CONFIG_VLAN |
5233 		MLX5_TXOFF_CONFIG_METADATA)
5234 
5235 
5236 MLX5_TXOFF_DECL(mtiv,
5237 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
5238 		MLX5_TXOFF_CONFIG_INLINE | MLX5_TXOFF_CONFIG_VLAN |
5239 		MLX5_TXOFF_CONFIG_METADATA)
5240 
5241 MLX5_TXOFF_DECL(sc,
5242 		MLX5_TXOFF_CONFIG_SWP |	MLX5_TXOFF_CONFIG_CSUM |
5243 		MLX5_TXOFF_CONFIG_METADATA)
5244 
5245 MLX5_TXOFF_DECL(sci,
5246 		MLX5_TXOFF_CONFIG_SWP |	MLX5_TXOFF_CONFIG_CSUM |
5247 		MLX5_TXOFF_CONFIG_INLINE |
5248 		MLX5_TXOFF_CONFIG_METADATA)
5249 
5250 
5251 MLX5_TXOFF_DECL(scv,
5252 		MLX5_TXOFF_CONFIG_SWP |	MLX5_TXOFF_CONFIG_CSUM |
5253 		MLX5_TXOFF_CONFIG_VLAN |
5254 		MLX5_TXOFF_CONFIG_METADATA)
5255 
5256 
5257 MLX5_TXOFF_DECL(sciv,
5258 		MLX5_TXOFF_CONFIG_SWP |	MLX5_TXOFF_CONFIG_CSUM |
5259 		MLX5_TXOFF_CONFIG_INLINE | MLX5_TXOFF_CONFIG_VLAN |
5260 		MLX5_TXOFF_CONFIG_METADATA)
5261 
5262 MLX5_TXOFF_DECL(i,
5263 		MLX5_TXOFF_CONFIG_INLINE |
5264 		MLX5_TXOFF_CONFIG_METADATA)
5265 
5266 MLX5_TXOFF_DECL(v,
5267 		MLX5_TXOFF_CONFIG_VLAN |
5268 		MLX5_TXOFF_CONFIG_METADATA)
5269 
5270 MLX5_TXOFF_DECL(iv,
5271 		MLX5_TXOFF_CONFIG_INLINE | MLX5_TXOFF_CONFIG_VLAN |
5272 		MLX5_TXOFF_CONFIG_METADATA)
5273 
5274 /* Generate routines with timestamp scheduling. */
5275 MLX5_TXOFF_DECL(full_ts_nompw,
5276 		MLX5_TXOFF_CONFIG_FULL | MLX5_TXOFF_CONFIG_TXPP)
5277 
5278 MLX5_TXOFF_DECL(full_ts_nompwi,
5279 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
5280 		MLX5_TXOFF_CONFIG_SWP | MLX5_TXOFF_CONFIG_CSUM |
5281 		MLX5_TXOFF_CONFIG_VLAN | MLX5_TXOFF_CONFIG_METADATA |
5282 		MLX5_TXOFF_CONFIG_TXPP)
5283 
5284 MLX5_TXOFF_DECL(full_ts,
5285 		MLX5_TXOFF_CONFIG_FULL | MLX5_TXOFF_CONFIG_TXPP |
5286 		MLX5_TXOFF_CONFIG_EMPW)
5287 
5288 MLX5_TXOFF_DECL(full_ts_noi,
5289 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
5290 		MLX5_TXOFF_CONFIG_SWP | MLX5_TXOFF_CONFIG_CSUM |
5291 		MLX5_TXOFF_CONFIG_VLAN | MLX5_TXOFF_CONFIG_METADATA |
5292 		MLX5_TXOFF_CONFIG_TXPP | MLX5_TXOFF_CONFIG_EMPW)
5293 
5294 MLX5_TXOFF_DECL(none_ts,
5295 		MLX5_TXOFF_CONFIG_NONE | MLX5_TXOFF_CONFIG_TXPP |
5296 		MLX5_TXOFF_CONFIG_EMPW)
5297 
5298 MLX5_TXOFF_DECL(mdi_ts,
5299 		MLX5_TXOFF_CONFIG_INLINE | MLX5_TXOFF_CONFIG_METADATA |
5300 		MLX5_TXOFF_CONFIG_TXPP | MLX5_TXOFF_CONFIG_EMPW)
5301 
5302 MLX5_TXOFF_DECL(mti_ts,
5303 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
5304 		MLX5_TXOFF_CONFIG_INLINE | MLX5_TXOFF_CONFIG_METADATA |
5305 		MLX5_TXOFF_CONFIG_TXPP | MLX5_TXOFF_CONFIG_EMPW)
5306 
5307 MLX5_TXOFF_DECL(mtiv_ts,
5308 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
5309 		MLX5_TXOFF_CONFIG_INLINE | MLX5_TXOFF_CONFIG_VLAN |
5310 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_TXPP |
5311 		MLX5_TXOFF_CONFIG_EMPW)
5312 
5313 /*
5314  * Generate routines with Legacy Multi-Packet Write support.
5315  * This mode is supported by ConnectX-4 Lx only and imposes
5316  * offload limitations, not supported:
5317  *   - ACL/Flows (metadata are becoming meaningless)
5318  *   - WQE Inline headers
5319  *   - SRIOV (E-Switch offloads)
5320  *   - VLAN insertion
5321  *   - tunnel encapsulation/decapsulation
5322  *   - TSO
5323  */
5324 MLX5_TXOFF_DECL(none_mpw,
5325 		MLX5_TXOFF_CONFIG_NONE | MLX5_TXOFF_CONFIG_EMPW |
5326 		MLX5_TXOFF_CONFIG_MPW)
5327 
5328 MLX5_TXOFF_DECL(mci_mpw,
5329 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_CSUM |
5330 		MLX5_TXOFF_CONFIG_INLINE | MLX5_TXOFF_CONFIG_EMPW |
5331 		MLX5_TXOFF_CONFIG_MPW)
5332 
5333 MLX5_TXOFF_DECL(mc_mpw,
5334 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_CSUM |
5335 		MLX5_TXOFF_CONFIG_EMPW | MLX5_TXOFF_CONFIG_MPW)
5336 
5337 MLX5_TXOFF_DECL(i_mpw,
5338 		MLX5_TXOFF_CONFIG_INLINE | MLX5_TXOFF_CONFIG_EMPW |
5339 		MLX5_TXOFF_CONFIG_MPW)
5340 
5341 /*
5342  * Array of declared and compiled Tx burst function and corresponding
5343  * supported offloads set. The array is used to select the Tx burst
5344  * function for specified offloads set at Tx queue configuration time.
5345  */
5346 const struct {
5347 	eth_tx_burst_t func;
5348 	unsigned int olx;
5349 } txoff_func[] = {
5350 MLX5_TXOFF_INFO(full_empw,
5351 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
5352 		MLX5_TXOFF_CONFIG_SWP |	MLX5_TXOFF_CONFIG_CSUM |
5353 		MLX5_TXOFF_CONFIG_INLINE | MLX5_TXOFF_CONFIG_VLAN |
5354 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
5355 
5356 MLX5_TXOFF_INFO(none_empw,
5357 		MLX5_TXOFF_CONFIG_NONE | MLX5_TXOFF_CONFIG_EMPW)
5358 
5359 MLX5_TXOFF_INFO(md_empw,
5360 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
5361 
5362 MLX5_TXOFF_INFO(mt_empw,
5363 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
5364 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
5365 
5366 MLX5_TXOFF_INFO(mtsc_empw,
5367 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
5368 		MLX5_TXOFF_CONFIG_SWP |	MLX5_TXOFF_CONFIG_CSUM |
5369 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
5370 
5371 MLX5_TXOFF_INFO(mti_empw,
5372 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
5373 		MLX5_TXOFF_CONFIG_INLINE |
5374 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
5375 
5376 MLX5_TXOFF_INFO(mtv_empw,
5377 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
5378 		MLX5_TXOFF_CONFIG_VLAN |
5379 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
5380 
5381 MLX5_TXOFF_INFO(mtiv_empw,
5382 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
5383 		MLX5_TXOFF_CONFIG_INLINE | MLX5_TXOFF_CONFIG_VLAN |
5384 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
5385 
5386 MLX5_TXOFF_INFO(sc_empw,
5387 		MLX5_TXOFF_CONFIG_SWP |	MLX5_TXOFF_CONFIG_CSUM |
5388 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
5389 
5390 MLX5_TXOFF_INFO(sci_empw,
5391 		MLX5_TXOFF_CONFIG_SWP |	MLX5_TXOFF_CONFIG_CSUM |
5392 		MLX5_TXOFF_CONFIG_INLINE |
5393 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
5394 
5395 MLX5_TXOFF_INFO(scv_empw,
5396 		MLX5_TXOFF_CONFIG_SWP |	MLX5_TXOFF_CONFIG_CSUM |
5397 		MLX5_TXOFF_CONFIG_VLAN |
5398 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
5399 
5400 MLX5_TXOFF_INFO(sciv_empw,
5401 		MLX5_TXOFF_CONFIG_SWP |	MLX5_TXOFF_CONFIG_CSUM |
5402 		MLX5_TXOFF_CONFIG_INLINE | MLX5_TXOFF_CONFIG_VLAN |
5403 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
5404 
5405 MLX5_TXOFF_INFO(i_empw,
5406 		MLX5_TXOFF_CONFIG_INLINE |
5407 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
5408 
5409 MLX5_TXOFF_INFO(v_empw,
5410 		MLX5_TXOFF_CONFIG_VLAN |
5411 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
5412 
5413 MLX5_TXOFF_INFO(iv_empw,
5414 		MLX5_TXOFF_CONFIG_INLINE | MLX5_TXOFF_CONFIG_VLAN |
5415 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
5416 
5417 MLX5_TXOFF_INFO(full_ts_nompw,
5418 		MLX5_TXOFF_CONFIG_FULL | MLX5_TXOFF_CONFIG_TXPP)
5419 
5420 MLX5_TXOFF_INFO(full_ts_nompwi,
5421 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
5422 		MLX5_TXOFF_CONFIG_SWP | MLX5_TXOFF_CONFIG_CSUM |
5423 		MLX5_TXOFF_CONFIG_VLAN | MLX5_TXOFF_CONFIG_METADATA |
5424 		MLX5_TXOFF_CONFIG_TXPP)
5425 
5426 MLX5_TXOFF_INFO(full_ts,
5427 		MLX5_TXOFF_CONFIG_FULL | MLX5_TXOFF_CONFIG_TXPP |
5428 		MLX5_TXOFF_CONFIG_EMPW)
5429 
5430 MLX5_TXOFF_INFO(full_ts_noi,
5431 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
5432 		MLX5_TXOFF_CONFIG_SWP | MLX5_TXOFF_CONFIG_CSUM |
5433 		MLX5_TXOFF_CONFIG_VLAN | MLX5_TXOFF_CONFIG_METADATA |
5434 		MLX5_TXOFF_CONFIG_TXPP | MLX5_TXOFF_CONFIG_EMPW)
5435 
5436 MLX5_TXOFF_INFO(none_ts,
5437 		MLX5_TXOFF_CONFIG_NONE | MLX5_TXOFF_CONFIG_TXPP |
5438 		MLX5_TXOFF_CONFIG_EMPW)
5439 
5440 MLX5_TXOFF_INFO(mdi_ts,
5441 		MLX5_TXOFF_CONFIG_INLINE | MLX5_TXOFF_CONFIG_METADATA |
5442 		MLX5_TXOFF_CONFIG_TXPP | MLX5_TXOFF_CONFIG_EMPW)
5443 
5444 MLX5_TXOFF_INFO(mti_ts,
5445 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
5446 		MLX5_TXOFF_CONFIG_INLINE | MLX5_TXOFF_CONFIG_METADATA |
5447 		MLX5_TXOFF_CONFIG_TXPP | MLX5_TXOFF_CONFIG_EMPW)
5448 
5449 MLX5_TXOFF_INFO(mtiv_ts,
5450 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
5451 		MLX5_TXOFF_CONFIG_INLINE | MLX5_TXOFF_CONFIG_VLAN |
5452 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_TXPP |
5453 		MLX5_TXOFF_CONFIG_EMPW)
5454 
5455 MLX5_TXOFF_INFO(full,
5456 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
5457 		MLX5_TXOFF_CONFIG_SWP |	MLX5_TXOFF_CONFIG_CSUM |
5458 		MLX5_TXOFF_CONFIG_INLINE | MLX5_TXOFF_CONFIG_VLAN |
5459 		MLX5_TXOFF_CONFIG_METADATA)
5460 
5461 MLX5_TXOFF_INFO(none,
5462 		MLX5_TXOFF_CONFIG_NONE)
5463 
5464 MLX5_TXOFF_INFO(md,
5465 		MLX5_TXOFF_CONFIG_METADATA)
5466 
5467 MLX5_TXOFF_INFO(mt,
5468 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
5469 		MLX5_TXOFF_CONFIG_METADATA)
5470 
5471 MLX5_TXOFF_INFO(mtsc,
5472 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
5473 		MLX5_TXOFF_CONFIG_SWP |	MLX5_TXOFF_CONFIG_CSUM |
5474 		MLX5_TXOFF_CONFIG_METADATA)
5475 
5476 MLX5_TXOFF_INFO(mti,
5477 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
5478 		MLX5_TXOFF_CONFIG_INLINE |
5479 		MLX5_TXOFF_CONFIG_METADATA)
5480 
5481 MLX5_TXOFF_INFO(mtv,
5482 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
5483 		MLX5_TXOFF_CONFIG_VLAN |
5484 		MLX5_TXOFF_CONFIG_METADATA)
5485 
5486 MLX5_TXOFF_INFO(mtiv,
5487 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
5488 		MLX5_TXOFF_CONFIG_INLINE | MLX5_TXOFF_CONFIG_VLAN |
5489 		MLX5_TXOFF_CONFIG_METADATA)
5490 
5491 MLX5_TXOFF_INFO(sc,
5492 		MLX5_TXOFF_CONFIG_SWP |	MLX5_TXOFF_CONFIG_CSUM |
5493 		MLX5_TXOFF_CONFIG_METADATA)
5494 
5495 MLX5_TXOFF_INFO(sci,
5496 		MLX5_TXOFF_CONFIG_SWP |	MLX5_TXOFF_CONFIG_CSUM |
5497 		MLX5_TXOFF_CONFIG_INLINE |
5498 		MLX5_TXOFF_CONFIG_METADATA)
5499 
5500 MLX5_TXOFF_INFO(scv,
5501 		MLX5_TXOFF_CONFIG_SWP |	MLX5_TXOFF_CONFIG_CSUM |
5502 		MLX5_TXOFF_CONFIG_VLAN |
5503 		MLX5_TXOFF_CONFIG_METADATA)
5504 
5505 MLX5_TXOFF_INFO(sciv,
5506 		MLX5_TXOFF_CONFIG_SWP |	MLX5_TXOFF_CONFIG_CSUM |
5507 		MLX5_TXOFF_CONFIG_INLINE | MLX5_TXOFF_CONFIG_VLAN |
5508 		MLX5_TXOFF_CONFIG_METADATA)
5509 
5510 MLX5_TXOFF_INFO(i,
5511 		MLX5_TXOFF_CONFIG_INLINE |
5512 		MLX5_TXOFF_CONFIG_METADATA)
5513 
5514 MLX5_TXOFF_INFO(v,
5515 		MLX5_TXOFF_CONFIG_VLAN |
5516 		MLX5_TXOFF_CONFIG_METADATA)
5517 
5518 MLX5_TXOFF_INFO(iv,
5519 		MLX5_TXOFF_CONFIG_INLINE | MLX5_TXOFF_CONFIG_VLAN |
5520 		MLX5_TXOFF_CONFIG_METADATA)
5521 
5522 MLX5_TXOFF_INFO(none_mpw,
5523 		MLX5_TXOFF_CONFIG_NONE | MLX5_TXOFF_CONFIG_EMPW |
5524 		MLX5_TXOFF_CONFIG_MPW)
5525 
5526 MLX5_TXOFF_INFO(mci_mpw,
5527 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_CSUM |
5528 		MLX5_TXOFF_CONFIG_INLINE | MLX5_TXOFF_CONFIG_EMPW |
5529 		MLX5_TXOFF_CONFIG_MPW)
5530 
5531 MLX5_TXOFF_INFO(mc_mpw,
5532 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_CSUM |
5533 		MLX5_TXOFF_CONFIG_EMPW | MLX5_TXOFF_CONFIG_MPW)
5534 
5535 MLX5_TXOFF_INFO(i_mpw,
5536 		MLX5_TXOFF_CONFIG_INLINE | MLX5_TXOFF_CONFIG_EMPW |
5537 		MLX5_TXOFF_CONFIG_MPW)
5538 };
5539 
5540 /**
5541  * Configure the Tx function to use. The routine checks configured
5542  * Tx offloads for the device and selects appropriate Tx burst
5543  * routine. There are multiple Tx burst routines compiled from
5544  * the same template in the most optimal way for the dedicated
5545  * Tx offloads set.
5546  *
5547  * @param dev
5548  *   Pointer to private data structure.
5549  *
5550  * @return
5551  *   Pointer to selected Tx burst function.
5552  */
5553 eth_tx_burst_t
5554 mlx5_select_tx_function(struct rte_eth_dev *dev)
5555 {
5556 	struct mlx5_priv *priv = dev->data->dev_private;
5557 	struct mlx5_dev_config *config = &priv->config;
5558 	uint64_t tx_offloads = dev->data->dev_conf.txmode.offloads;
5559 	unsigned int diff = 0, olx = 0, i, m;
5560 
5561 	static_assert(MLX5_WQE_SIZE_MAX / MLX5_WSEG_SIZE <=
5562 		      MLX5_DSEG_MAX, "invalid WQE max size");
5563 	static_assert(MLX5_WQE_CSEG_SIZE == MLX5_WSEG_SIZE,
5564 		      "invalid WQE Control Segment size");
5565 	static_assert(MLX5_WQE_ESEG_SIZE == MLX5_WSEG_SIZE,
5566 		      "invalid WQE Ethernet Segment size");
5567 	static_assert(MLX5_WQE_DSEG_SIZE == MLX5_WSEG_SIZE,
5568 		      "invalid WQE Data Segment size");
5569 	static_assert(MLX5_WQE_SIZE == 4 * MLX5_WSEG_SIZE,
5570 		      "invalid WQE size");
5571 	MLX5_ASSERT(priv);
5572 	if (tx_offloads & DEV_TX_OFFLOAD_MULTI_SEGS) {
5573 		/* We should support Multi-Segment Packets. */
5574 		olx |= MLX5_TXOFF_CONFIG_MULTI;
5575 	}
5576 	if (tx_offloads & (DEV_TX_OFFLOAD_TCP_TSO |
5577 			   DEV_TX_OFFLOAD_VXLAN_TNL_TSO |
5578 			   DEV_TX_OFFLOAD_GRE_TNL_TSO |
5579 			   DEV_TX_OFFLOAD_IP_TNL_TSO |
5580 			   DEV_TX_OFFLOAD_UDP_TNL_TSO)) {
5581 		/* We should support TCP Send Offload. */
5582 		olx |= MLX5_TXOFF_CONFIG_TSO;
5583 	}
5584 	if (tx_offloads & (DEV_TX_OFFLOAD_IP_TNL_TSO |
5585 			   DEV_TX_OFFLOAD_UDP_TNL_TSO |
5586 			   DEV_TX_OFFLOAD_OUTER_IPV4_CKSUM)) {
5587 		/* We should support Software Parser for Tunnels. */
5588 		olx |= MLX5_TXOFF_CONFIG_SWP;
5589 	}
5590 	if (tx_offloads & (DEV_TX_OFFLOAD_IPV4_CKSUM |
5591 			   DEV_TX_OFFLOAD_UDP_CKSUM |
5592 			   DEV_TX_OFFLOAD_TCP_CKSUM |
5593 			   DEV_TX_OFFLOAD_OUTER_IPV4_CKSUM)) {
5594 		/* We should support IP/TCP/UDP Checksums. */
5595 		olx |= MLX5_TXOFF_CONFIG_CSUM;
5596 	}
5597 	if (tx_offloads & DEV_TX_OFFLOAD_VLAN_INSERT) {
5598 		/* We should support VLAN insertion. */
5599 		olx |= MLX5_TXOFF_CONFIG_VLAN;
5600 	}
5601 	if (tx_offloads & DEV_TX_OFFLOAD_SEND_ON_TIMESTAMP &&
5602 	    rte_mbuf_dynflag_lookup
5603 			(RTE_MBUF_DYNFLAG_TX_TIMESTAMP_NAME, NULL) >= 0 &&
5604 	    rte_mbuf_dynfield_lookup
5605 			(RTE_MBUF_DYNFIELD_TIMESTAMP_NAME, NULL) >= 0) {
5606 		/* Offload configured, dynamic entities registered. */
5607 		olx |= MLX5_TXOFF_CONFIG_TXPP;
5608 	}
5609 	if (priv->txqs_n && (*priv->txqs)[0]) {
5610 		struct mlx5_txq_data *txd = (*priv->txqs)[0];
5611 
5612 		if (txd->inlen_send) {
5613 			/*
5614 			 * Check the data inline requirements. Data inline
5615 			 * is enabled on per device basis, we can check
5616 			 * the first Tx queue only.
5617 			 *
5618 			 * If device does not support VLAN insertion in WQE
5619 			 * and some queues are requested to perform VLAN
5620 			 * insertion offload than inline must be enabled.
5621 			 */
5622 			olx |= MLX5_TXOFF_CONFIG_INLINE;
5623 		}
5624 	}
5625 	if (config->mps == MLX5_MPW_ENHANCED &&
5626 	    config->txq_inline_min <= 0) {
5627 		/*
5628 		 * The NIC supports Enhanced Multi-Packet Write
5629 		 * and does not require minimal inline data.
5630 		 */
5631 		olx |= MLX5_TXOFF_CONFIG_EMPW;
5632 	}
5633 	if (rte_flow_dynf_metadata_avail()) {
5634 		/* We should support Flow metadata. */
5635 		olx |= MLX5_TXOFF_CONFIG_METADATA;
5636 	}
5637 	if (config->mps == MLX5_MPW) {
5638 		/*
5639 		 * The NIC supports Legacy Multi-Packet Write.
5640 		 * The MLX5_TXOFF_CONFIG_MPW controls the
5641 		 * descriptor building method in combination
5642 		 * with MLX5_TXOFF_CONFIG_EMPW.
5643 		 */
5644 		if (!(olx & (MLX5_TXOFF_CONFIG_TSO |
5645 			     MLX5_TXOFF_CONFIG_SWP |
5646 			     MLX5_TXOFF_CONFIG_VLAN |
5647 			     MLX5_TXOFF_CONFIG_METADATA)))
5648 			olx |= MLX5_TXOFF_CONFIG_EMPW |
5649 			       MLX5_TXOFF_CONFIG_MPW;
5650 	}
5651 	/*
5652 	 * Scan the routines table to find the minimal
5653 	 * satisfying routine with requested offloads.
5654 	 */
5655 	m = RTE_DIM(txoff_func);
5656 	for (i = 0; i < RTE_DIM(txoff_func); i++) {
5657 		unsigned int tmp;
5658 
5659 		tmp = txoff_func[i].olx;
5660 		if (tmp == olx) {
5661 			/* Meets requested offloads exactly.*/
5662 			m = i;
5663 			break;
5664 		}
5665 		if ((tmp & olx) != olx) {
5666 			/* Does not meet requested offloads at all. */
5667 			continue;
5668 		}
5669 		if ((olx ^ tmp) & MLX5_TXOFF_CONFIG_MPW)
5670 			/* Do not enable legacy MPW if not configured. */
5671 			continue;
5672 		if ((olx ^ tmp) & MLX5_TXOFF_CONFIG_EMPW)
5673 			/* Do not enable eMPW if not configured. */
5674 			continue;
5675 		if ((olx ^ tmp) & MLX5_TXOFF_CONFIG_INLINE)
5676 			/* Do not enable inlining if not configured. */
5677 			continue;
5678 		if ((olx ^ tmp) & MLX5_TXOFF_CONFIG_TXPP)
5679 			/* Do not enable scheduling if not configured. */
5680 			continue;
5681 		/*
5682 		 * Some routine meets the requirements.
5683 		 * Check whether it has minimal amount
5684 		 * of not requested offloads.
5685 		 */
5686 		tmp = __builtin_popcountl(tmp & ~olx);
5687 		if (m >= RTE_DIM(txoff_func) || tmp < diff) {
5688 			/* First or better match, save and continue. */
5689 			m = i;
5690 			diff = tmp;
5691 			continue;
5692 		}
5693 		if (tmp == diff) {
5694 			tmp = txoff_func[i].olx ^ txoff_func[m].olx;
5695 			if (__builtin_ffsl(txoff_func[i].olx & ~tmp) <
5696 			    __builtin_ffsl(txoff_func[m].olx & ~tmp)) {
5697 				/* Lighter not requested offload. */
5698 				m = i;
5699 			}
5700 		}
5701 	}
5702 	if (m >= RTE_DIM(txoff_func)) {
5703 		DRV_LOG(DEBUG, "port %u has no selected Tx function"
5704 			       " for requested offloads %04X",
5705 				dev->data->port_id, olx);
5706 		return NULL;
5707 	}
5708 	DRV_LOG(DEBUG, "port %u has selected Tx function"
5709 		       " supporting offloads %04X/%04X",
5710 			dev->data->port_id, olx, txoff_func[m].olx);
5711 	if (txoff_func[m].olx & MLX5_TXOFF_CONFIG_MULTI)
5712 		DRV_LOG(DEBUG, "\tMULTI (multi segment)");
5713 	if (txoff_func[m].olx & MLX5_TXOFF_CONFIG_TSO)
5714 		DRV_LOG(DEBUG, "\tTSO   (TCP send offload)");
5715 	if (txoff_func[m].olx & MLX5_TXOFF_CONFIG_SWP)
5716 		DRV_LOG(DEBUG, "\tSWP   (software parser)");
5717 	if (txoff_func[m].olx & MLX5_TXOFF_CONFIG_CSUM)
5718 		DRV_LOG(DEBUG, "\tCSUM  (checksum offload)");
5719 	if (txoff_func[m].olx & MLX5_TXOFF_CONFIG_INLINE)
5720 		DRV_LOG(DEBUG, "\tINLIN (inline data)");
5721 	if (txoff_func[m].olx & MLX5_TXOFF_CONFIG_VLAN)
5722 		DRV_LOG(DEBUG, "\tVLANI (VLAN insertion)");
5723 	if (txoff_func[m].olx & MLX5_TXOFF_CONFIG_METADATA)
5724 		DRV_LOG(DEBUG, "\tMETAD (tx Flow metadata)");
5725 	if (txoff_func[m].olx & MLX5_TXOFF_CONFIG_TXPP)
5726 		DRV_LOG(DEBUG, "\tMETAD (tx Scheduling)");
5727 	if (txoff_func[m].olx & MLX5_TXOFF_CONFIG_EMPW) {
5728 		if (txoff_func[m].olx & MLX5_TXOFF_CONFIG_MPW)
5729 			DRV_LOG(DEBUG, "\tMPW   (Legacy MPW)");
5730 		else
5731 			DRV_LOG(DEBUG, "\tEMPW  (Enhanced MPW)");
5732 	}
5733 	return txoff_func[m].func;
5734 }
5735 
5736 /**
5737  * DPDK callback to get the TX queue information
5738  *
5739  * @param dev
5740  *   Pointer to the device structure.
5741  *
5742  * @param tx_queue_id
5743  *   Tx queue identificator.
5744  *
5745  * @param qinfo
5746  *   Pointer to the TX queue information structure.
5747  *
5748  * @return
5749  *   None.
5750  */
5751 
5752 void
5753 mlx5_txq_info_get(struct rte_eth_dev *dev, uint16_t tx_queue_id,
5754 		  struct rte_eth_txq_info *qinfo)
5755 {
5756 	struct mlx5_priv *priv = dev->data->dev_private;
5757 	struct mlx5_txq_data *txq = (*priv->txqs)[tx_queue_id];
5758 	struct mlx5_txq_ctrl *txq_ctrl =
5759 			container_of(txq, struct mlx5_txq_ctrl, txq);
5760 
5761 	if (!txq)
5762 		return;
5763 	qinfo->nb_desc = txq->elts_s;
5764 	qinfo->conf.tx_thresh.pthresh = 0;
5765 	qinfo->conf.tx_thresh.hthresh = 0;
5766 	qinfo->conf.tx_thresh.wthresh = 0;
5767 	qinfo->conf.tx_rs_thresh = 0;
5768 	qinfo->conf.tx_free_thresh = 0;
5769 	qinfo->conf.tx_deferred_start = txq_ctrl ? 0 : 1;
5770 	qinfo->conf.offloads = dev->data->dev_conf.txmode.offloads;
5771 }
5772 
5773 /**
5774  * DPDK callback to get the TX packet burst mode information
5775  *
5776  * @param dev
5777  *   Pointer to the device structure.
5778  *
5779  * @param tx_queue_id
5780  *   Tx queue identificatior.
5781  *
5782  * @param mode
5783  *   Pointer to the burts mode information.
5784  *
5785  * @return
5786  *   0 as success, -EINVAL as failure.
5787  */
5788 
5789 int
5790 mlx5_tx_burst_mode_get(struct rte_eth_dev *dev,
5791 		       uint16_t tx_queue_id __rte_unused,
5792 		       struct rte_eth_burst_mode *mode)
5793 {
5794 	eth_tx_burst_t pkt_burst = dev->tx_pkt_burst;
5795 	unsigned int i, olx;
5796 
5797 	for (i = 0; i < RTE_DIM(txoff_func); i++) {
5798 		if (pkt_burst == txoff_func[i].func) {
5799 			olx = txoff_func[i].olx;
5800 			snprintf(mode->info, sizeof(mode->info),
5801 				 "%s%s%s%s%s%s%s%s%s",
5802 				 (olx & MLX5_TXOFF_CONFIG_EMPW) ?
5803 				 ((olx & MLX5_TXOFF_CONFIG_MPW) ?
5804 				 "Legacy MPW" : "Enhanced MPW") : "No MPW",
5805 				 (olx & MLX5_TXOFF_CONFIG_MULTI) ?
5806 				 " + MULTI" : "",
5807 				 (olx & MLX5_TXOFF_CONFIG_TSO) ?
5808 				 " + TSO" : "",
5809 				 (olx & MLX5_TXOFF_CONFIG_SWP) ?
5810 				 " + SWP" : "",
5811 				 (olx & MLX5_TXOFF_CONFIG_CSUM) ?
5812 				 "  + CSUM" : "",
5813 				 (olx & MLX5_TXOFF_CONFIG_INLINE) ?
5814 				 " + INLINE" : "",
5815 				 (olx & MLX5_TXOFF_CONFIG_VLAN) ?
5816 				 " + VLAN" : "",
5817 				 (olx & MLX5_TXOFF_CONFIG_METADATA) ?
5818 				 " + METADATA" : "",
5819 				 (olx & MLX5_TXOFF_CONFIG_TXPP) ?
5820 				 " + TXPP" : "");
5821 			return 0;
5822 		}
5823 	}
5824 	return -EINVAL;
5825 }
5826