xref: /dpdk/drivers/net/mlx5/mlx5_rxtx.c (revision f5862ae99e058c0cee36a08dfd51f8a3b766999a)
1 /* SPDX-License-Identifier: BSD-3-Clause
2  * Copyright 2015 6WIND S.A.
3  * Copyright 2015-2019 Mellanox Technologies, Ltd
4  */
5 
6 #include <assert.h>
7 #include <stdint.h>
8 #include <string.h>
9 #include <stdlib.h>
10 
11 /* Verbs header. */
12 /* ISO C doesn't support unnamed structs/unions, disabling -pedantic. */
13 #ifdef PEDANTIC
14 #pragma GCC diagnostic ignored "-Wpedantic"
15 #endif
16 #include <infiniband/verbs.h>
17 #include <infiniband/mlx5dv.h>
18 #ifdef PEDANTIC
19 #pragma GCC diagnostic error "-Wpedantic"
20 #endif
21 
22 #include <rte_mbuf.h>
23 #include <rte_mempool.h>
24 #include <rte_prefetch.h>
25 #include <rte_common.h>
26 #include <rte_branch_prediction.h>
27 #include <rte_ether.h>
28 #include <rte_cycles.h>
29 #include <rte_flow.h>
30 
31 #include "mlx5.h"
32 #include "mlx5_utils.h"
33 #include "mlx5_rxtx.h"
34 #include "mlx5_autoconf.h"
35 #include "mlx5_defs.h"
36 #include "mlx5_prm.h"
37 
38 /* TX burst subroutines return codes. */
39 enum mlx5_txcmp_code {
40 	MLX5_TXCMP_CODE_EXIT = 0,
41 	MLX5_TXCMP_CODE_ERROR,
42 	MLX5_TXCMP_CODE_SINGLE,
43 	MLX5_TXCMP_CODE_MULTI,
44 	MLX5_TXCMP_CODE_TSO,
45 	MLX5_TXCMP_CODE_EMPW,
46 };
47 
48 /*
49  * These defines are used to configure Tx burst routine option set
50  * supported at compile time. The not specified options are optimized out
51  * out due to if conditions can be explicitly calculated at compile time.
52  * The offloads with bigger runtime check (require more CPU cycles to
53  * skip) overhead should have the bigger index - this is needed to
54  * select the better matching routine function if no exact match and
55  * some offloads are not actually requested.
56  */
57 #define MLX5_TXOFF_CONFIG_MULTI (1u << 0) /* Multi-segment packets.*/
58 #define MLX5_TXOFF_CONFIG_TSO (1u << 1) /* TCP send offload supported.*/
59 #define MLX5_TXOFF_CONFIG_SWP (1u << 2) /* Tunnels/SW Parser offloads.*/
60 #define MLX5_TXOFF_CONFIG_CSUM (1u << 3) /* Check Sums offloaded. */
61 #define MLX5_TXOFF_CONFIG_INLINE (1u << 4) /* Data inlining supported. */
62 #define MLX5_TXOFF_CONFIG_VLAN (1u << 5) /* VLAN insertion supported.*/
63 #define MLX5_TXOFF_CONFIG_METADATA (1u << 6) /* Flow metadata. */
64 #define MLX5_TXOFF_CONFIG_EMPW (1u << 8) /* Enhanced MPW supported.*/
65 #define MLX5_TXOFF_CONFIG_MPW (1u << 9) /* Legacy MPW supported.*/
66 
67 /* The most common offloads groups. */
68 #define MLX5_TXOFF_CONFIG_NONE 0
69 #define MLX5_TXOFF_CONFIG_FULL (MLX5_TXOFF_CONFIG_MULTI | \
70 				MLX5_TXOFF_CONFIG_TSO | \
71 				MLX5_TXOFF_CONFIG_SWP | \
72 				MLX5_TXOFF_CONFIG_CSUM | \
73 				MLX5_TXOFF_CONFIG_INLINE | \
74 				MLX5_TXOFF_CONFIG_VLAN | \
75 				MLX5_TXOFF_CONFIG_METADATA)
76 
77 #define MLX5_TXOFF_CONFIG(mask) (olx & MLX5_TXOFF_CONFIG_##mask)
78 
79 #define MLX5_TXOFF_DECL(func, olx) \
80 static uint16_t mlx5_tx_burst_##func(void *txq, \
81 				     struct rte_mbuf **pkts, \
82 				    uint16_t pkts_n) \
83 { \
84 	return mlx5_tx_burst_tmpl((struct mlx5_txq_data *)txq, \
85 		    pkts, pkts_n, (olx)); \
86 }
87 
88 #define MLX5_TXOFF_INFO(func, olx) {mlx5_tx_burst_##func, olx},
89 
90 static __rte_always_inline uint32_t
91 rxq_cq_to_pkt_type(struct mlx5_rxq_data *rxq, volatile struct mlx5_cqe *cqe);
92 
93 static __rte_always_inline int
94 mlx5_rx_poll_len(struct mlx5_rxq_data *rxq, volatile struct mlx5_cqe *cqe,
95 		 uint16_t cqe_cnt, volatile struct mlx5_mini_cqe8 **mcqe);
96 
97 static __rte_always_inline uint32_t
98 rxq_cq_to_ol_flags(volatile struct mlx5_cqe *cqe);
99 
100 static __rte_always_inline void
101 rxq_cq_to_mbuf(struct mlx5_rxq_data *rxq, struct rte_mbuf *pkt,
102 	       volatile struct mlx5_cqe *cqe, uint32_t rss_hash_res);
103 
104 static __rte_always_inline void
105 mprq_buf_replace(struct mlx5_rxq_data *rxq, uint16_t rq_idx,
106 		 const unsigned int strd_n);
107 
108 static int
109 mlx5_queue_state_modify(struct rte_eth_dev *dev,
110 			struct mlx5_mp_arg_queue_state_modify *sm);
111 
112 static inline void
113 mlx5_lro_update_tcp_hdr(struct rte_tcp_hdr *restrict tcp,
114 			volatile struct mlx5_cqe *restrict cqe,
115 			uint32_t phcsum);
116 
117 static inline void
118 mlx5_lro_update_hdr(uint8_t *restrict padd,
119 		    volatile struct mlx5_cqe *restrict cqe,
120 		    uint32_t len);
121 
122 uint32_t mlx5_ptype_table[] __rte_cache_aligned = {
123 	[0xff] = RTE_PTYPE_ALL_MASK, /* Last entry for errored packet. */
124 };
125 
126 uint8_t mlx5_cksum_table[1 << 10] __rte_cache_aligned;
127 uint8_t mlx5_swp_types_table[1 << 10] __rte_cache_aligned;
128 
129 /**
130  * Build a table to translate Rx completion flags to packet type.
131  *
132  * @note: fix mlx5_dev_supported_ptypes_get() if any change here.
133  */
134 void
135 mlx5_set_ptype_table(void)
136 {
137 	unsigned int i;
138 	uint32_t (*p)[RTE_DIM(mlx5_ptype_table)] = &mlx5_ptype_table;
139 
140 	/* Last entry must not be overwritten, reserved for errored packet. */
141 	for (i = 0; i < RTE_DIM(mlx5_ptype_table) - 1; ++i)
142 		(*p)[i] = RTE_PTYPE_UNKNOWN;
143 	/*
144 	 * The index to the array should have:
145 	 * bit[1:0] = l3_hdr_type
146 	 * bit[4:2] = l4_hdr_type
147 	 * bit[5] = ip_frag
148 	 * bit[6] = tunneled
149 	 * bit[7] = outer_l3_type
150 	 */
151 	/* L2 */
152 	(*p)[0x00] = RTE_PTYPE_L2_ETHER;
153 	/* L3 */
154 	(*p)[0x01] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
155 		     RTE_PTYPE_L4_NONFRAG;
156 	(*p)[0x02] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
157 		     RTE_PTYPE_L4_NONFRAG;
158 	/* Fragmented */
159 	(*p)[0x21] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
160 		     RTE_PTYPE_L4_FRAG;
161 	(*p)[0x22] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
162 		     RTE_PTYPE_L4_FRAG;
163 	/* TCP */
164 	(*p)[0x05] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
165 		     RTE_PTYPE_L4_TCP;
166 	(*p)[0x06] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
167 		     RTE_PTYPE_L4_TCP;
168 	(*p)[0x0d] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
169 		     RTE_PTYPE_L4_TCP;
170 	(*p)[0x0e] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
171 		     RTE_PTYPE_L4_TCP;
172 	(*p)[0x11] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
173 		     RTE_PTYPE_L4_TCP;
174 	(*p)[0x12] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
175 		     RTE_PTYPE_L4_TCP;
176 	/* UDP */
177 	(*p)[0x09] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
178 		     RTE_PTYPE_L4_UDP;
179 	(*p)[0x0a] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
180 		     RTE_PTYPE_L4_UDP;
181 	/* Repeat with outer_l3_type being set. Just in case. */
182 	(*p)[0x81] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
183 		     RTE_PTYPE_L4_NONFRAG;
184 	(*p)[0x82] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
185 		     RTE_PTYPE_L4_NONFRAG;
186 	(*p)[0xa1] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
187 		     RTE_PTYPE_L4_FRAG;
188 	(*p)[0xa2] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
189 		     RTE_PTYPE_L4_FRAG;
190 	(*p)[0x85] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
191 		     RTE_PTYPE_L4_TCP;
192 	(*p)[0x86] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
193 		     RTE_PTYPE_L4_TCP;
194 	(*p)[0x8d] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
195 		     RTE_PTYPE_L4_TCP;
196 	(*p)[0x8e] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
197 		     RTE_PTYPE_L4_TCP;
198 	(*p)[0x91] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
199 		     RTE_PTYPE_L4_TCP;
200 	(*p)[0x92] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
201 		     RTE_PTYPE_L4_TCP;
202 	(*p)[0x89] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
203 		     RTE_PTYPE_L4_UDP;
204 	(*p)[0x8a] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
205 		     RTE_PTYPE_L4_UDP;
206 	/* Tunneled - L3 */
207 	(*p)[0x40] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN;
208 	(*p)[0x41] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
209 		     RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN |
210 		     RTE_PTYPE_INNER_L4_NONFRAG;
211 	(*p)[0x42] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
212 		     RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN |
213 		     RTE_PTYPE_INNER_L4_NONFRAG;
214 	(*p)[0xc0] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN;
215 	(*p)[0xc1] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
216 		     RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN |
217 		     RTE_PTYPE_INNER_L4_NONFRAG;
218 	(*p)[0xc2] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
219 		     RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN |
220 		     RTE_PTYPE_INNER_L4_NONFRAG;
221 	/* Tunneled - Fragmented */
222 	(*p)[0x61] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
223 		     RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN |
224 		     RTE_PTYPE_INNER_L4_FRAG;
225 	(*p)[0x62] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
226 		     RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN |
227 		     RTE_PTYPE_INNER_L4_FRAG;
228 	(*p)[0xe1] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
229 		     RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN |
230 		     RTE_PTYPE_INNER_L4_FRAG;
231 	(*p)[0xe2] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
232 		     RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN |
233 		     RTE_PTYPE_INNER_L4_FRAG;
234 	/* Tunneled - TCP */
235 	(*p)[0x45] = 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)[0x46] = 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)[0x4d] = 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)[0x4e] = 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)[0x51] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
248 		     RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN |
249 		     RTE_PTYPE_INNER_L4_TCP;
250 	(*p)[0x52] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
251 		     RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN |
252 		     RTE_PTYPE_INNER_L4_TCP;
253 	(*p)[0xc5] = 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)[0xc6] = 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)[0xcd] = 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)[0xce] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
263 		     RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN |
264 		     RTE_PTYPE_INNER_L4_TCP;
265 	(*p)[0xd1] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
266 		     RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN |
267 		     RTE_PTYPE_INNER_L4_TCP;
268 	(*p)[0xd2] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
269 		     RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN |
270 		     RTE_PTYPE_INNER_L4_TCP;
271 	/* Tunneled - UDP */
272 	(*p)[0x49] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
273 		     RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN |
274 		     RTE_PTYPE_INNER_L4_UDP;
275 	(*p)[0x4a] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
276 		     RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN |
277 		     RTE_PTYPE_INNER_L4_UDP;
278 	(*p)[0xc9] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
279 		     RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN |
280 		     RTE_PTYPE_INNER_L4_UDP;
281 	(*p)[0xca] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
282 		     RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN |
283 		     RTE_PTYPE_INNER_L4_UDP;
284 }
285 
286 /**
287  * Build a table to translate packet to checksum type of Verbs.
288  */
289 void
290 mlx5_set_cksum_table(void)
291 {
292 	unsigned int i;
293 	uint8_t v;
294 
295 	/*
296 	 * The index should have:
297 	 * bit[0] = PKT_TX_TCP_SEG
298 	 * bit[2:3] = PKT_TX_UDP_CKSUM, PKT_TX_TCP_CKSUM
299 	 * bit[4] = PKT_TX_IP_CKSUM
300 	 * bit[8] = PKT_TX_OUTER_IP_CKSUM
301 	 * bit[9] = tunnel
302 	 */
303 	for (i = 0; i < RTE_DIM(mlx5_cksum_table); ++i) {
304 		v = 0;
305 		if (i & (1 << 9)) {
306 			/* Tunneled packet. */
307 			if (i & (1 << 8)) /* Outer IP. */
308 				v |= MLX5_ETH_WQE_L3_CSUM;
309 			if (i & (1 << 4)) /* Inner IP. */
310 				v |= MLX5_ETH_WQE_L3_INNER_CSUM;
311 			if (i & (3 << 2 | 1 << 0)) /* L4 or TSO. */
312 				v |= MLX5_ETH_WQE_L4_INNER_CSUM;
313 		} else {
314 			/* No tunnel. */
315 			if (i & (1 << 4)) /* IP. */
316 				v |= MLX5_ETH_WQE_L3_CSUM;
317 			if (i & (3 << 2 | 1 << 0)) /* L4 or TSO. */
318 				v |= MLX5_ETH_WQE_L4_CSUM;
319 		}
320 		mlx5_cksum_table[i] = v;
321 	}
322 }
323 
324 /**
325  * Build a table to translate packet type of mbuf to SWP type of Verbs.
326  */
327 void
328 mlx5_set_swp_types_table(void)
329 {
330 	unsigned int i;
331 	uint8_t v;
332 
333 	/*
334 	 * The index should have:
335 	 * bit[0:1] = PKT_TX_L4_MASK
336 	 * bit[4] = PKT_TX_IPV6
337 	 * bit[8] = PKT_TX_OUTER_IPV6
338 	 * bit[9] = PKT_TX_OUTER_UDP
339 	 */
340 	for (i = 0; i < RTE_DIM(mlx5_swp_types_table); ++i) {
341 		v = 0;
342 		if (i & (1 << 8))
343 			v |= MLX5_ETH_WQE_L3_OUTER_IPV6;
344 		if (i & (1 << 9))
345 			v |= MLX5_ETH_WQE_L4_OUTER_UDP;
346 		if (i & (1 << 4))
347 			v |= MLX5_ETH_WQE_L3_INNER_IPV6;
348 		if ((i & 3) == (PKT_TX_UDP_CKSUM >> 52))
349 			v |= MLX5_ETH_WQE_L4_INNER_UDP;
350 		mlx5_swp_types_table[i] = v;
351 	}
352 }
353 
354 /**
355  * Set Software Parser flags and offsets in Ethernet Segment of WQE.
356  * Flags must be preliminary initialized to zero.
357  *
358  * @param loc
359  *   Pointer to burst routine local context.
360  * @param swp_flags
361  *   Pointer to store Software Parser flags
362  * @param olx
363  *   Configured Tx offloads mask. It is fully defined at
364  *   compile time and may be used for optimization.
365  *
366  * @return
367  *   Software Parser offsets packed in dword.
368  *   Software Parser flags are set by pointer.
369  */
370 static __rte_always_inline uint32_t
371 txq_mbuf_to_swp(struct mlx5_txq_local *restrict loc,
372 		uint8_t *swp_flags,
373 		unsigned int olx)
374 {
375 	uint64_t ol, tunnel;
376 	unsigned int idx, off;
377 	uint32_t set;
378 
379 	if (!MLX5_TXOFF_CONFIG(SWP))
380 		return 0;
381 	ol = loc->mbuf->ol_flags;
382 	tunnel = ol & PKT_TX_TUNNEL_MASK;
383 	/*
384 	 * Check whether Software Parser is required.
385 	 * Only customized tunnels may ask for.
386 	 */
387 	if (likely(tunnel != PKT_TX_TUNNEL_UDP && tunnel != PKT_TX_TUNNEL_IP))
388 		return 0;
389 	/*
390 	 * The index should have:
391 	 * bit[0:1] = PKT_TX_L4_MASK
392 	 * bit[4] = PKT_TX_IPV6
393 	 * bit[8] = PKT_TX_OUTER_IPV6
394 	 * bit[9] = PKT_TX_OUTER_UDP
395 	 */
396 	idx = (ol & (PKT_TX_L4_MASK | PKT_TX_IPV6 | PKT_TX_OUTER_IPV6)) >> 52;
397 	idx |= (tunnel == PKT_TX_TUNNEL_UDP) ? (1 << 9) : 0;
398 	*swp_flags = mlx5_swp_types_table[idx];
399 	/*
400 	 * Set offsets for SW parser. Since ConnectX-5, SW parser just
401 	 * complements HW parser. SW parser starts to engage only if HW parser
402 	 * can't reach a header. For the older devices, HW parser will not kick
403 	 * in if any of SWP offsets is set. Therefore, all of the L3 offsets
404 	 * should be set regardless of HW offload.
405 	 */
406 	off = loc->mbuf->outer_l2_len;
407 	if (MLX5_TXOFF_CONFIG(VLAN) && ol & PKT_TX_VLAN_PKT)
408 		off += sizeof(struct rte_vlan_hdr);
409 	set = (off >> 1) << 8; /* Outer L3 offset. */
410 	off += loc->mbuf->outer_l3_len;
411 	if (tunnel == PKT_TX_TUNNEL_UDP)
412 		set |= off >> 1; /* Outer L4 offset. */
413 	if (ol & (PKT_TX_IPV4 | PKT_TX_IPV6)) { /* Inner IP. */
414 		const uint64_t csum = ol & PKT_TX_L4_MASK;
415 			off += loc->mbuf->l2_len;
416 		set |= (off >> 1) << 24; /* Inner L3 offset. */
417 		if (csum == PKT_TX_TCP_CKSUM ||
418 		    csum == PKT_TX_UDP_CKSUM ||
419 		    (MLX5_TXOFF_CONFIG(TSO) && ol & PKT_TX_TCP_SEG)) {
420 			off += loc->mbuf->l3_len;
421 			set |= (off >> 1) << 16; /* Inner L4 offset. */
422 		}
423 	}
424 	set = rte_cpu_to_le_32(set);
425 	return set;
426 }
427 
428 /**
429  * Convert the Checksum offloads to Verbs.
430  *
431  * @param buf
432  *   Pointer to the mbuf.
433  *
434  * @return
435  *   Converted checksum flags.
436  */
437 static __rte_always_inline uint8_t
438 txq_ol_cksum_to_cs(struct rte_mbuf *buf)
439 {
440 	uint32_t idx;
441 	uint8_t is_tunnel = !!(buf->ol_flags & PKT_TX_TUNNEL_MASK);
442 	const uint64_t ol_flags_mask = PKT_TX_TCP_SEG | PKT_TX_L4_MASK |
443 				       PKT_TX_IP_CKSUM | PKT_TX_OUTER_IP_CKSUM;
444 
445 	/*
446 	 * The index should have:
447 	 * bit[0] = PKT_TX_TCP_SEG
448 	 * bit[2:3] = PKT_TX_UDP_CKSUM, PKT_TX_TCP_CKSUM
449 	 * bit[4] = PKT_TX_IP_CKSUM
450 	 * bit[8] = PKT_TX_OUTER_IP_CKSUM
451 	 * bit[9] = tunnel
452 	 */
453 	idx = ((buf->ol_flags & ol_flags_mask) >> 50) | (!!is_tunnel << 9);
454 	return mlx5_cksum_table[idx];
455 }
456 
457 /**
458  * Internal function to compute the number of used descriptors in an RX queue
459  *
460  * @param rxq
461  *   The Rx queue.
462  *
463  * @return
464  *   The number of used rx descriptor.
465  */
466 static uint32_t
467 rx_queue_count(struct mlx5_rxq_data *rxq)
468 {
469 	struct rxq_zip *zip = &rxq->zip;
470 	volatile struct mlx5_cqe *cqe;
471 	const unsigned int cqe_n = (1 << rxq->cqe_n);
472 	const unsigned int cqe_cnt = cqe_n - 1;
473 	unsigned int cq_ci;
474 	unsigned int used;
475 
476 	/* if we are processing a compressed cqe */
477 	if (zip->ai) {
478 		used = zip->cqe_cnt - zip->ca;
479 		cq_ci = zip->cq_ci;
480 	} else {
481 		used = 0;
482 		cq_ci = rxq->cq_ci;
483 	}
484 	cqe = &(*rxq->cqes)[cq_ci & cqe_cnt];
485 	while (check_cqe(cqe, cqe_n, cq_ci) != MLX5_CQE_STATUS_HW_OWN) {
486 		int8_t op_own;
487 		unsigned int n;
488 
489 		op_own = cqe->op_own;
490 		if (MLX5_CQE_FORMAT(op_own) == MLX5_COMPRESSED)
491 			n = rte_be_to_cpu_32(cqe->byte_cnt);
492 		else
493 			n = 1;
494 		cq_ci += n;
495 		used += n;
496 		cqe = &(*rxq->cqes)[cq_ci & cqe_cnt];
497 	}
498 	used = RTE_MIN(used, (1U << rxq->elts_n) - 1);
499 	return used;
500 }
501 
502 /**
503  * DPDK callback to check the status of a rx descriptor.
504  *
505  * @param rx_queue
506  *   The Rx queue.
507  * @param[in] offset
508  *   The index of the descriptor in the ring.
509  *
510  * @return
511  *   The status of the tx descriptor.
512  */
513 int
514 mlx5_rx_descriptor_status(void *rx_queue, uint16_t offset)
515 {
516 	struct mlx5_rxq_data *rxq = rx_queue;
517 	struct mlx5_rxq_ctrl *rxq_ctrl =
518 			container_of(rxq, struct mlx5_rxq_ctrl, rxq);
519 	struct rte_eth_dev *dev = ETH_DEV(rxq_ctrl->priv);
520 
521 	if (dev->rx_pkt_burst != mlx5_rx_burst) {
522 		rte_errno = ENOTSUP;
523 		return -rte_errno;
524 	}
525 	if (offset >= (1 << rxq->elts_n)) {
526 		rte_errno = EINVAL;
527 		return -rte_errno;
528 	}
529 	if (offset < rx_queue_count(rxq))
530 		return RTE_ETH_RX_DESC_DONE;
531 	return RTE_ETH_RX_DESC_AVAIL;
532 }
533 
534 /**
535  * DPDK callback to get the number of used descriptors in a RX queue
536  *
537  * @param dev
538  *   Pointer to the device structure.
539  *
540  * @param rx_queue_id
541  *   The Rx queue.
542  *
543  * @return
544  *   The number of used rx descriptor.
545  *   -EINVAL if the queue is invalid
546  */
547 uint32_t
548 mlx5_rx_queue_count(struct rte_eth_dev *dev, uint16_t rx_queue_id)
549 {
550 	struct mlx5_priv *priv = dev->data->dev_private;
551 	struct mlx5_rxq_data *rxq;
552 
553 	if (dev->rx_pkt_burst != mlx5_rx_burst) {
554 		rte_errno = ENOTSUP;
555 		return -rte_errno;
556 	}
557 	rxq = (*priv->rxqs)[rx_queue_id];
558 	if (!rxq) {
559 		rte_errno = EINVAL;
560 		return -rte_errno;
561 	}
562 	return rx_queue_count(rxq);
563 }
564 
565 #define MLX5_SYSTEM_LOG_DIR "/var/log"
566 /**
567  * Dump debug information to log file.
568  *
569  * @param fname
570  *   The file name.
571  * @param hex_title
572  *   If not NULL this string is printed as a header to the output
573  *   and the output will be in hexadecimal view.
574  * @param buf
575  *   This is the buffer address to print out.
576  * @param len
577  *   The number of bytes to dump out.
578  */
579 void
580 mlx5_dump_debug_information(const char *fname, const char *hex_title,
581 			    const void *buf, unsigned int hex_len)
582 {
583 	FILE *fd;
584 
585 	MKSTR(path, "%s/%s", MLX5_SYSTEM_LOG_DIR, fname);
586 	fd = fopen(path, "a+");
587 	if (!fd) {
588 		DRV_LOG(WARNING, "cannot open %s for debug dump", path);
589 		MKSTR(path2, "./%s", fname);
590 		fd = fopen(path2, "a+");
591 		if (!fd) {
592 			DRV_LOG(ERR, "cannot open %s for debug dump", path2);
593 			return;
594 		}
595 		DRV_LOG(INFO, "New debug dump in file %s", path2);
596 	} else {
597 		DRV_LOG(INFO, "New debug dump in file %s", path);
598 	}
599 	if (hex_title)
600 		rte_hexdump(fd, hex_title, buf, hex_len);
601 	else
602 		fprintf(fd, "%s", (const char *)buf);
603 	fprintf(fd, "\n\n\n");
604 	fclose(fd);
605 }
606 
607 /**
608  * Move QP from error state to running state and initialize indexes.
609  *
610  * @param txq_ctrl
611  *   Pointer to TX queue control structure.
612  *
613  * @return
614  *   0 on success, else -1.
615  */
616 static int
617 tx_recover_qp(struct mlx5_txq_ctrl *txq_ctrl)
618 {
619 	struct mlx5_mp_arg_queue_state_modify sm = {
620 			.is_wq = 0,
621 			.queue_id = txq_ctrl->txq.idx,
622 	};
623 
624 	if (mlx5_queue_state_modify(ETH_DEV(txq_ctrl->priv), &sm))
625 		return -1;
626 	txq_ctrl->txq.wqe_ci = 0;
627 	txq_ctrl->txq.wqe_pi = 0;
628 	txq_ctrl->txq.elts_comp = 0;
629 	return 0;
630 }
631 
632 /* Return 1 if the error CQE is signed otherwise, sign it and return 0. */
633 static int
634 check_err_cqe_seen(volatile struct mlx5_err_cqe *err_cqe)
635 {
636 	static const uint8_t magic[] = "seen";
637 	int ret = 1;
638 	unsigned int i;
639 
640 	for (i = 0; i < sizeof(magic); ++i)
641 		if (!ret || err_cqe->rsvd1[i] != magic[i]) {
642 			ret = 0;
643 			err_cqe->rsvd1[i] = magic[i];
644 		}
645 	return ret;
646 }
647 
648 /**
649  * Handle error CQE.
650  *
651  * @param txq
652  *   Pointer to TX queue structure.
653  * @param error_cqe
654  *   Pointer to the error CQE.
655  *
656  * @return
657  *   Negative value if queue recovery failed, otherwise
658  *   the error completion entry is handled successfully.
659  */
660 static int
661 mlx5_tx_error_cqe_handle(struct mlx5_txq_data *restrict txq,
662 			 volatile struct mlx5_err_cqe *err_cqe)
663 {
664 	if (err_cqe->syndrome != MLX5_CQE_SYNDROME_WR_FLUSH_ERR) {
665 		const uint16_t wqe_m = ((1 << txq->wqe_n) - 1);
666 		struct mlx5_txq_ctrl *txq_ctrl =
667 				container_of(txq, struct mlx5_txq_ctrl, txq);
668 		uint16_t new_wqe_pi = rte_be_to_cpu_16(err_cqe->wqe_counter);
669 		int seen = check_err_cqe_seen(err_cqe);
670 
671 		if (!seen && txq_ctrl->dump_file_n <
672 		    txq_ctrl->priv->config.max_dump_files_num) {
673 			MKSTR(err_str, "Unexpected CQE error syndrome "
674 			      "0x%02x CQN = %u SQN = %u wqe_counter = %u "
675 			      "wq_ci = %u cq_ci = %u", err_cqe->syndrome,
676 			      txq->cqe_s, txq->qp_num_8s >> 8,
677 			      rte_be_to_cpu_16(err_cqe->wqe_counter),
678 			      txq->wqe_ci, txq->cq_ci);
679 			MKSTR(name, "dpdk_mlx5_port_%u_txq_%u_index_%u_%u",
680 			      PORT_ID(txq_ctrl->priv), txq->idx,
681 			      txq_ctrl->dump_file_n, (uint32_t)rte_rdtsc());
682 			mlx5_dump_debug_information(name, NULL, err_str, 0);
683 			mlx5_dump_debug_information(name, "MLX5 Error CQ:",
684 						    (const void *)((uintptr_t)
685 						    txq->cqes),
686 						    sizeof(*err_cqe) *
687 						    (1 << txq->cqe_n));
688 			mlx5_dump_debug_information(name, "MLX5 Error SQ:",
689 						    (const void *)((uintptr_t)
690 						    txq->wqes),
691 						    MLX5_WQE_SIZE *
692 						    (1 << txq->wqe_n));
693 			txq_ctrl->dump_file_n++;
694 		}
695 		if (!seen)
696 			/*
697 			 * Count errors in WQEs units.
698 			 * Later it can be improved to count error packets,
699 			 * for example, by SQ parsing to find how much packets
700 			 * should be counted for each WQE.
701 			 */
702 			txq->stats.oerrors += ((txq->wqe_ci & wqe_m) -
703 						new_wqe_pi) & wqe_m;
704 		if (tx_recover_qp(txq_ctrl)) {
705 			/* Recovering failed - retry later on the same WQE. */
706 			return -1;
707 		}
708 		/* Release all the remaining buffers. */
709 		txq_free_elts(txq_ctrl);
710 	}
711 	return 0;
712 }
713 
714 /**
715  * Translate RX completion flags to packet type.
716  *
717  * @param[in] rxq
718  *   Pointer to RX queue structure.
719  * @param[in] cqe
720  *   Pointer to CQE.
721  *
722  * @note: fix mlx5_dev_supported_ptypes_get() if any change here.
723  *
724  * @return
725  *   Packet type for struct rte_mbuf.
726  */
727 static inline uint32_t
728 rxq_cq_to_pkt_type(struct mlx5_rxq_data *rxq, volatile struct mlx5_cqe *cqe)
729 {
730 	uint8_t idx;
731 	uint8_t pinfo = cqe->pkt_info;
732 	uint16_t ptype = cqe->hdr_type_etc;
733 
734 	/*
735 	 * The index to the array should have:
736 	 * bit[1:0] = l3_hdr_type
737 	 * bit[4:2] = l4_hdr_type
738 	 * bit[5] = ip_frag
739 	 * bit[6] = tunneled
740 	 * bit[7] = outer_l3_type
741 	 */
742 	idx = ((pinfo & 0x3) << 6) | ((ptype & 0xfc00) >> 10);
743 	return mlx5_ptype_table[idx] | rxq->tunnel * !!(idx & (1 << 6));
744 }
745 
746 /**
747  * Initialize Rx WQ and indexes.
748  *
749  * @param[in] rxq
750  *   Pointer to RX queue structure.
751  */
752 void
753 mlx5_rxq_initialize(struct mlx5_rxq_data *rxq)
754 {
755 	const unsigned int wqe_n = 1 << rxq->elts_n;
756 	unsigned int i;
757 
758 	for (i = 0; (i != wqe_n); ++i) {
759 		volatile struct mlx5_wqe_data_seg *scat;
760 		uintptr_t addr;
761 		uint32_t byte_count;
762 
763 		if (mlx5_rxq_mprq_enabled(rxq)) {
764 			struct mlx5_mprq_buf *buf = (*rxq->mprq_bufs)[i];
765 
766 			scat = &((volatile struct mlx5_wqe_mprq *)
767 				rxq->wqes)[i].dseg;
768 			addr = (uintptr_t)mlx5_mprq_buf_addr(buf,
769 							 1 << rxq->strd_num_n);
770 			byte_count = (1 << rxq->strd_sz_n) *
771 					(1 << rxq->strd_num_n);
772 		} else {
773 			struct rte_mbuf *buf = (*rxq->elts)[i];
774 
775 			scat = &((volatile struct mlx5_wqe_data_seg *)
776 					rxq->wqes)[i];
777 			addr = rte_pktmbuf_mtod(buf, uintptr_t);
778 			byte_count = DATA_LEN(buf);
779 		}
780 		/* scat->addr must be able to store a pointer. */
781 		assert(sizeof(scat->addr) >= sizeof(uintptr_t));
782 		*scat = (struct mlx5_wqe_data_seg){
783 			.addr = rte_cpu_to_be_64(addr),
784 			.byte_count = rte_cpu_to_be_32(byte_count),
785 			.lkey = mlx5_rx_addr2mr(rxq, addr),
786 		};
787 	}
788 	rxq->consumed_strd = 0;
789 	rxq->decompressed = 0;
790 	rxq->rq_pi = 0;
791 	rxq->zip = (struct rxq_zip){
792 		.ai = 0,
793 	};
794 	/* Update doorbell counter. */
795 	rxq->rq_ci = wqe_n >> rxq->sges_n;
796 	rte_cio_wmb();
797 	*rxq->rq_db = rte_cpu_to_be_32(rxq->rq_ci);
798 }
799 
800 /**
801  * Modify a Verbs/DevX queue state.
802  * This must be called from the primary process.
803  *
804  * @param dev
805  *   Pointer to Ethernet device.
806  * @param sm
807  *   State modify request parameters.
808  *
809  * @return
810  *   0 in case of success else non-zero value and rte_errno is set.
811  */
812 int
813 mlx5_queue_state_modify_primary(struct rte_eth_dev *dev,
814 			const struct mlx5_mp_arg_queue_state_modify *sm)
815 {
816 	int ret;
817 	struct mlx5_priv *priv = dev->data->dev_private;
818 
819 	if (sm->is_wq) {
820 		struct mlx5_rxq_data *rxq = (*priv->rxqs)[sm->queue_id];
821 		struct mlx5_rxq_ctrl *rxq_ctrl =
822 			container_of(rxq, struct mlx5_rxq_ctrl, rxq);
823 
824 		if (rxq_ctrl->obj->type == MLX5_RXQ_OBJ_TYPE_IBV) {
825 			struct ibv_wq_attr mod = {
826 				.attr_mask = IBV_WQ_ATTR_STATE,
827 				.wq_state = sm->state,
828 			};
829 
830 			ret = mlx5_glue->modify_wq(rxq_ctrl->obj->wq, &mod);
831 		} else { /* rxq_ctrl->obj->type == MLX5_RXQ_OBJ_TYPE_DEVX_RQ. */
832 			struct mlx5_devx_modify_rq_attr rq_attr;
833 
834 			memset(&rq_attr, 0, sizeof(rq_attr));
835 			if (sm->state == IBV_WQS_RESET) {
836 				rq_attr.rq_state = MLX5_RQC_STATE_ERR;
837 				rq_attr.state = MLX5_RQC_STATE_RST;
838 			} else if (sm->state == IBV_WQS_RDY) {
839 				rq_attr.rq_state = MLX5_RQC_STATE_RST;
840 				rq_attr.state = MLX5_RQC_STATE_RDY;
841 			} else if (sm->state == IBV_WQS_ERR) {
842 				rq_attr.rq_state = MLX5_RQC_STATE_RDY;
843 				rq_attr.state = MLX5_RQC_STATE_ERR;
844 			}
845 			ret = mlx5_devx_cmd_modify_rq(rxq_ctrl->obj->rq,
846 						      &rq_attr);
847 		}
848 		if (ret) {
849 			DRV_LOG(ERR, "Cannot change Rx WQ state to %u  - %s",
850 					sm->state, strerror(errno));
851 			rte_errno = errno;
852 			return ret;
853 		}
854 	} else {
855 		struct mlx5_txq_data *txq = (*priv->txqs)[sm->queue_id];
856 		struct mlx5_txq_ctrl *txq_ctrl =
857 			container_of(txq, struct mlx5_txq_ctrl, txq);
858 		struct ibv_qp_attr mod = {
859 			.qp_state = IBV_QPS_RESET,
860 			.port_num = (uint8_t)priv->ibv_port,
861 		};
862 		struct ibv_qp *qp = txq_ctrl->obj->qp;
863 
864 		ret = mlx5_glue->modify_qp(qp, &mod, IBV_QP_STATE);
865 		if (ret) {
866 			DRV_LOG(ERR, "Cannot change the Tx QP state to RESET "
867 				"%s", strerror(errno));
868 			rte_errno = errno;
869 			return ret;
870 		}
871 		mod.qp_state = IBV_QPS_INIT;
872 		ret = mlx5_glue->modify_qp(qp, &mod,
873 					   (IBV_QP_STATE | IBV_QP_PORT));
874 		if (ret) {
875 			DRV_LOG(ERR, "Cannot change Tx QP state to INIT %s",
876 				strerror(errno));
877 			rte_errno = errno;
878 			return ret;
879 		}
880 		mod.qp_state = IBV_QPS_RTR;
881 		ret = mlx5_glue->modify_qp(qp, &mod, IBV_QP_STATE);
882 		if (ret) {
883 			DRV_LOG(ERR, "Cannot change Tx QP state to RTR %s",
884 				strerror(errno));
885 			rte_errno = errno;
886 			return ret;
887 		}
888 		mod.qp_state = IBV_QPS_RTS;
889 		ret = mlx5_glue->modify_qp(qp, &mod, IBV_QP_STATE);
890 		if (ret) {
891 			DRV_LOG(ERR, "Cannot change Tx QP state to RTS %s",
892 				strerror(errno));
893 			rte_errno = errno;
894 			return ret;
895 		}
896 	}
897 	return 0;
898 }
899 
900 /**
901  * Modify a Verbs queue state.
902  *
903  * @param dev
904  *   Pointer to Ethernet device.
905  * @param sm
906  *   State modify request parameters.
907  *
908  * @return
909  *   0 in case of success else non-zero value.
910  */
911 static int
912 mlx5_queue_state_modify(struct rte_eth_dev *dev,
913 			struct mlx5_mp_arg_queue_state_modify *sm)
914 {
915 	int ret = 0;
916 
917 	switch (rte_eal_process_type()) {
918 	case RTE_PROC_PRIMARY:
919 		ret = mlx5_queue_state_modify_primary(dev, sm);
920 		break;
921 	case RTE_PROC_SECONDARY:
922 		ret = mlx5_mp_req_queue_state_modify(dev, sm);
923 		break;
924 	default:
925 		break;
926 	}
927 	return ret;
928 }
929 
930 /**
931  * Handle a Rx error.
932  * The function inserts the RQ state to reset when the first error CQE is
933  * shown, then drains the CQ by the caller function loop. When the CQ is empty,
934  * it moves the RQ state to ready and initializes the RQ.
935  * Next CQE identification and error counting are in the caller responsibility.
936  *
937  * @param[in] rxq
938  *   Pointer to RX queue structure.
939  * @param[in] vec
940  *   1 when called from vectorized Rx burst, need to prepare mbufs for the RQ.
941  *   0 when called from non-vectorized Rx burst.
942  *
943  * @return
944  *   -1 in case of recovery error, otherwise the CQE status.
945  */
946 int
947 mlx5_rx_err_handle(struct mlx5_rxq_data *rxq, uint8_t vec)
948 {
949 	const uint16_t cqe_n = 1 << rxq->cqe_n;
950 	const uint16_t cqe_mask = cqe_n - 1;
951 	const unsigned int wqe_n = 1 << rxq->elts_n;
952 	struct mlx5_rxq_ctrl *rxq_ctrl =
953 			container_of(rxq, struct mlx5_rxq_ctrl, rxq);
954 	union {
955 		volatile struct mlx5_cqe *cqe;
956 		volatile struct mlx5_err_cqe *err_cqe;
957 	} u = {
958 		.cqe = &(*rxq->cqes)[rxq->cq_ci & cqe_mask],
959 	};
960 	struct mlx5_mp_arg_queue_state_modify sm;
961 	int ret;
962 
963 	switch (rxq->err_state) {
964 	case MLX5_RXQ_ERR_STATE_NO_ERROR:
965 		rxq->err_state = MLX5_RXQ_ERR_STATE_NEED_RESET;
966 		/* Fall-through */
967 	case MLX5_RXQ_ERR_STATE_NEED_RESET:
968 		sm.is_wq = 1;
969 		sm.queue_id = rxq->idx;
970 		sm.state = IBV_WQS_RESET;
971 		if (mlx5_queue_state_modify(ETH_DEV(rxq_ctrl->priv), &sm))
972 			return -1;
973 		if (rxq_ctrl->dump_file_n <
974 		    rxq_ctrl->priv->config.max_dump_files_num) {
975 			MKSTR(err_str, "Unexpected CQE error syndrome "
976 			      "0x%02x CQN = %u RQN = %u wqe_counter = %u"
977 			      " rq_ci = %u cq_ci = %u", u.err_cqe->syndrome,
978 			      rxq->cqn, rxq_ctrl->wqn,
979 			      rte_be_to_cpu_16(u.err_cqe->wqe_counter),
980 			      rxq->rq_ci << rxq->sges_n, rxq->cq_ci);
981 			MKSTR(name, "dpdk_mlx5_port_%u_rxq_%u_%u",
982 			      rxq->port_id, rxq->idx, (uint32_t)rte_rdtsc());
983 			mlx5_dump_debug_information(name, NULL, err_str, 0);
984 			mlx5_dump_debug_information(name, "MLX5 Error CQ:",
985 						    (const void *)((uintptr_t)
986 								    rxq->cqes),
987 						    sizeof(*u.cqe) * cqe_n);
988 			mlx5_dump_debug_information(name, "MLX5 Error RQ:",
989 						    (const void *)((uintptr_t)
990 								    rxq->wqes),
991 						    16 * wqe_n);
992 			rxq_ctrl->dump_file_n++;
993 		}
994 		rxq->err_state = MLX5_RXQ_ERR_STATE_NEED_READY;
995 		/* Fall-through */
996 	case MLX5_RXQ_ERR_STATE_NEED_READY:
997 		ret = check_cqe(u.cqe, cqe_n, rxq->cq_ci);
998 		if (ret == MLX5_CQE_STATUS_HW_OWN) {
999 			rte_cio_wmb();
1000 			*rxq->cq_db = rte_cpu_to_be_32(rxq->cq_ci);
1001 			rte_cio_wmb();
1002 			/*
1003 			 * The RQ consumer index must be zeroed while moving
1004 			 * from RESET state to RDY state.
1005 			 */
1006 			*rxq->rq_db = rte_cpu_to_be_32(0);
1007 			rte_cio_wmb();
1008 			sm.is_wq = 1;
1009 			sm.queue_id = rxq->idx;
1010 			sm.state = IBV_WQS_RDY;
1011 			if (mlx5_queue_state_modify(ETH_DEV(rxq_ctrl->priv),
1012 						    &sm))
1013 				return -1;
1014 			if (vec) {
1015 				const uint16_t q_mask = wqe_n - 1;
1016 				uint16_t elt_idx;
1017 				struct rte_mbuf **elt;
1018 				int i;
1019 				unsigned int n = wqe_n - (rxq->rq_ci -
1020 							  rxq->rq_pi);
1021 
1022 				for (i = 0; i < (int)n; ++i) {
1023 					elt_idx = (rxq->rq_ci + i) & q_mask;
1024 					elt = &(*rxq->elts)[elt_idx];
1025 					*elt = rte_mbuf_raw_alloc(rxq->mp);
1026 					if (!*elt) {
1027 						for (i--; i >= 0; --i) {
1028 							elt_idx = (rxq->rq_ci +
1029 								   i) & q_mask;
1030 							elt = &(*rxq->elts)
1031 								[elt_idx];
1032 							rte_pktmbuf_free_seg
1033 								(*elt);
1034 						}
1035 						return -1;
1036 					}
1037 				}
1038 				for (i = 0; i < (int)wqe_n; ++i) {
1039 					elt = &(*rxq->elts)[i];
1040 					DATA_LEN(*elt) =
1041 						(uint16_t)((*elt)->buf_len -
1042 						rte_pktmbuf_headroom(*elt));
1043 				}
1044 				/* Padding with a fake mbuf for vec Rx. */
1045 				for (i = 0; i < MLX5_VPMD_DESCS_PER_LOOP; ++i)
1046 					(*rxq->elts)[wqe_n + i] =
1047 								&rxq->fake_mbuf;
1048 			}
1049 			mlx5_rxq_initialize(rxq);
1050 			rxq->err_state = MLX5_RXQ_ERR_STATE_NO_ERROR;
1051 		}
1052 		return ret;
1053 	default:
1054 		return -1;
1055 	}
1056 }
1057 
1058 /**
1059  * Get size of the next packet for a given CQE. For compressed CQEs, the
1060  * consumer index is updated only once all packets of the current one have
1061  * been processed.
1062  *
1063  * @param rxq
1064  *   Pointer to RX queue.
1065  * @param cqe
1066  *   CQE to process.
1067  * @param[out] mcqe
1068  *   Store pointer to mini-CQE if compressed. Otherwise, the pointer is not
1069  *   written.
1070  *
1071  * @return
1072  *   0 in case of empty CQE, otherwise the packet size in bytes.
1073  */
1074 static inline int
1075 mlx5_rx_poll_len(struct mlx5_rxq_data *rxq, volatile struct mlx5_cqe *cqe,
1076 		 uint16_t cqe_cnt, volatile struct mlx5_mini_cqe8 **mcqe)
1077 {
1078 	struct rxq_zip *zip = &rxq->zip;
1079 	uint16_t cqe_n = cqe_cnt + 1;
1080 	int len;
1081 	uint16_t idx, end;
1082 
1083 	do {
1084 		len = 0;
1085 		/* Process compressed data in the CQE and mini arrays. */
1086 		if (zip->ai) {
1087 			volatile struct mlx5_mini_cqe8 (*mc)[8] =
1088 				(volatile struct mlx5_mini_cqe8 (*)[8])
1089 				(uintptr_t)(&(*rxq->cqes)[zip->ca &
1090 							  cqe_cnt].pkt_info);
1091 
1092 			len = rte_be_to_cpu_32((*mc)[zip->ai & 7].byte_cnt);
1093 			*mcqe = &(*mc)[zip->ai & 7];
1094 			if ((++zip->ai & 7) == 0) {
1095 				/* Invalidate consumed CQEs */
1096 				idx = zip->ca;
1097 				end = zip->na;
1098 				while (idx != end) {
1099 					(*rxq->cqes)[idx & cqe_cnt].op_own =
1100 						MLX5_CQE_INVALIDATE;
1101 					++idx;
1102 				}
1103 				/*
1104 				 * Increment consumer index to skip the number
1105 				 * of CQEs consumed. Hardware leaves holes in
1106 				 * the CQ ring for software use.
1107 				 */
1108 				zip->ca = zip->na;
1109 				zip->na += 8;
1110 			}
1111 			if (unlikely(rxq->zip.ai == rxq->zip.cqe_cnt)) {
1112 				/* Invalidate the rest */
1113 				idx = zip->ca;
1114 				end = zip->cq_ci;
1115 
1116 				while (idx != end) {
1117 					(*rxq->cqes)[idx & cqe_cnt].op_own =
1118 						MLX5_CQE_INVALIDATE;
1119 					++idx;
1120 				}
1121 				rxq->cq_ci = zip->cq_ci;
1122 				zip->ai = 0;
1123 			}
1124 		/*
1125 		 * No compressed data, get next CQE and verify if it is
1126 		 * compressed.
1127 		 */
1128 		} else {
1129 			int ret;
1130 			int8_t op_own;
1131 
1132 			ret = check_cqe(cqe, cqe_n, rxq->cq_ci);
1133 			if (unlikely(ret != MLX5_CQE_STATUS_SW_OWN)) {
1134 				if (unlikely(ret == MLX5_CQE_STATUS_ERR ||
1135 					     rxq->err_state)) {
1136 					ret = mlx5_rx_err_handle(rxq, 0);
1137 					if (ret == MLX5_CQE_STATUS_HW_OWN ||
1138 					    ret == -1)
1139 						return 0;
1140 				} else {
1141 					return 0;
1142 				}
1143 			}
1144 			++rxq->cq_ci;
1145 			op_own = cqe->op_own;
1146 			if (MLX5_CQE_FORMAT(op_own) == MLX5_COMPRESSED) {
1147 				volatile struct mlx5_mini_cqe8 (*mc)[8] =
1148 					(volatile struct mlx5_mini_cqe8 (*)[8])
1149 					(uintptr_t)(&(*rxq->cqes)
1150 						[rxq->cq_ci &
1151 						 cqe_cnt].pkt_info);
1152 
1153 				/* Fix endianness. */
1154 				zip->cqe_cnt = rte_be_to_cpu_32(cqe->byte_cnt);
1155 				/*
1156 				 * Current mini array position is the one
1157 				 * returned by check_cqe64().
1158 				 *
1159 				 * If completion comprises several mini arrays,
1160 				 * as a special case the second one is located
1161 				 * 7 CQEs after the initial CQE instead of 8
1162 				 * for subsequent ones.
1163 				 */
1164 				zip->ca = rxq->cq_ci;
1165 				zip->na = zip->ca + 7;
1166 				/* Compute the next non compressed CQE. */
1167 				--rxq->cq_ci;
1168 				zip->cq_ci = rxq->cq_ci + zip->cqe_cnt;
1169 				/* Get packet size to return. */
1170 				len = rte_be_to_cpu_32((*mc)[0].byte_cnt);
1171 				*mcqe = &(*mc)[0];
1172 				zip->ai = 1;
1173 				/* Prefetch all to be invalidated */
1174 				idx = zip->ca;
1175 				end = zip->cq_ci;
1176 				while (idx != end) {
1177 					rte_prefetch0(&(*rxq->cqes)[(idx) &
1178 								    cqe_cnt]);
1179 					++idx;
1180 				}
1181 			} else {
1182 				len = rte_be_to_cpu_32(cqe->byte_cnt);
1183 			}
1184 		}
1185 		if (unlikely(rxq->err_state)) {
1186 			cqe = &(*rxq->cqes)[rxq->cq_ci & cqe_cnt];
1187 			++rxq->stats.idropped;
1188 		} else {
1189 			return len;
1190 		}
1191 	} while (1);
1192 }
1193 
1194 /**
1195  * Translate RX completion flags to offload flags.
1196  *
1197  * @param[in] cqe
1198  *   Pointer to CQE.
1199  *
1200  * @return
1201  *   Offload flags (ol_flags) for struct rte_mbuf.
1202  */
1203 static inline uint32_t
1204 rxq_cq_to_ol_flags(volatile struct mlx5_cqe *cqe)
1205 {
1206 	uint32_t ol_flags = 0;
1207 	uint16_t flags = rte_be_to_cpu_16(cqe->hdr_type_etc);
1208 
1209 	ol_flags =
1210 		TRANSPOSE(flags,
1211 			  MLX5_CQE_RX_L3_HDR_VALID,
1212 			  PKT_RX_IP_CKSUM_GOOD) |
1213 		TRANSPOSE(flags,
1214 			  MLX5_CQE_RX_L4_HDR_VALID,
1215 			  PKT_RX_L4_CKSUM_GOOD);
1216 	return ol_flags;
1217 }
1218 
1219 /**
1220  * Fill in mbuf fields from RX completion flags.
1221  * Note that pkt->ol_flags should be initialized outside of this function.
1222  *
1223  * @param rxq
1224  *   Pointer to RX queue.
1225  * @param pkt
1226  *   mbuf to fill.
1227  * @param cqe
1228  *   CQE to process.
1229  * @param rss_hash_res
1230  *   Packet RSS Hash result.
1231  */
1232 static inline void
1233 rxq_cq_to_mbuf(struct mlx5_rxq_data *rxq, struct rte_mbuf *pkt,
1234 	       volatile struct mlx5_cqe *cqe, uint32_t rss_hash_res)
1235 {
1236 	/* Update packet information. */
1237 	pkt->packet_type = rxq_cq_to_pkt_type(rxq, cqe);
1238 	if (rss_hash_res && rxq->rss_hash) {
1239 		pkt->hash.rss = rss_hash_res;
1240 		pkt->ol_flags |= PKT_RX_RSS_HASH;
1241 	}
1242 	if (rxq->mark && MLX5_FLOW_MARK_IS_VALID(cqe->sop_drop_qpn)) {
1243 		pkt->ol_flags |= PKT_RX_FDIR;
1244 		if (cqe->sop_drop_qpn !=
1245 		    rte_cpu_to_be_32(MLX5_FLOW_MARK_DEFAULT)) {
1246 			uint32_t mark = cqe->sop_drop_qpn;
1247 
1248 			pkt->ol_flags |= PKT_RX_FDIR_ID;
1249 			pkt->hash.fdir.hi = mlx5_flow_mark_get(mark);
1250 		}
1251 	}
1252 	if (rte_flow_dynf_metadata_avail() && cqe->flow_table_metadata) {
1253 		pkt->ol_flags |= PKT_RX_DYNF_METADATA;
1254 		*RTE_FLOW_DYNF_METADATA(pkt) = cqe->flow_table_metadata;
1255 	}
1256 	if (rxq->csum)
1257 		pkt->ol_flags |= rxq_cq_to_ol_flags(cqe);
1258 	if (rxq->vlan_strip &&
1259 	    (cqe->hdr_type_etc & rte_cpu_to_be_16(MLX5_CQE_VLAN_STRIPPED))) {
1260 		pkt->ol_flags |= PKT_RX_VLAN | PKT_RX_VLAN_STRIPPED;
1261 		pkt->vlan_tci = rte_be_to_cpu_16(cqe->vlan_info);
1262 	}
1263 	if (rxq->hw_timestamp) {
1264 		pkt->timestamp = rte_be_to_cpu_64(cqe->timestamp);
1265 		pkt->ol_flags |= PKT_RX_TIMESTAMP;
1266 	}
1267 }
1268 
1269 /**
1270  * DPDK callback for RX.
1271  *
1272  * @param dpdk_rxq
1273  *   Generic pointer to RX queue structure.
1274  * @param[out] pkts
1275  *   Array to store received packets.
1276  * @param pkts_n
1277  *   Maximum number of packets in array.
1278  *
1279  * @return
1280  *   Number of packets successfully received (<= pkts_n).
1281  */
1282 uint16_t
1283 mlx5_rx_burst(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n)
1284 {
1285 	struct mlx5_rxq_data *rxq = dpdk_rxq;
1286 	const unsigned int wqe_cnt = (1 << rxq->elts_n) - 1;
1287 	const unsigned int cqe_cnt = (1 << rxq->cqe_n) - 1;
1288 	const unsigned int sges_n = rxq->sges_n;
1289 	struct rte_mbuf *pkt = NULL;
1290 	struct rte_mbuf *seg = NULL;
1291 	volatile struct mlx5_cqe *cqe =
1292 		&(*rxq->cqes)[rxq->cq_ci & cqe_cnt];
1293 	unsigned int i = 0;
1294 	unsigned int rq_ci = rxq->rq_ci << sges_n;
1295 	int len = 0; /* keep its value across iterations. */
1296 
1297 	while (pkts_n) {
1298 		unsigned int idx = rq_ci & wqe_cnt;
1299 		volatile struct mlx5_wqe_data_seg *wqe =
1300 			&((volatile struct mlx5_wqe_data_seg *)rxq->wqes)[idx];
1301 		struct rte_mbuf *rep = (*rxq->elts)[idx];
1302 		volatile struct mlx5_mini_cqe8 *mcqe = NULL;
1303 		uint32_t rss_hash_res;
1304 
1305 		if (pkt)
1306 			NEXT(seg) = rep;
1307 		seg = rep;
1308 		rte_prefetch0(seg);
1309 		rte_prefetch0(cqe);
1310 		rte_prefetch0(wqe);
1311 		rep = rte_mbuf_raw_alloc(rxq->mp);
1312 		if (unlikely(rep == NULL)) {
1313 			++rxq->stats.rx_nombuf;
1314 			if (!pkt) {
1315 				/*
1316 				 * no buffers before we even started,
1317 				 * bail out silently.
1318 				 */
1319 				break;
1320 			}
1321 			while (pkt != seg) {
1322 				assert(pkt != (*rxq->elts)[idx]);
1323 				rep = NEXT(pkt);
1324 				NEXT(pkt) = NULL;
1325 				NB_SEGS(pkt) = 1;
1326 				rte_mbuf_raw_free(pkt);
1327 				pkt = rep;
1328 			}
1329 			break;
1330 		}
1331 		if (!pkt) {
1332 			cqe = &(*rxq->cqes)[rxq->cq_ci & cqe_cnt];
1333 			len = mlx5_rx_poll_len(rxq, cqe, cqe_cnt, &mcqe);
1334 			if (!len) {
1335 				rte_mbuf_raw_free(rep);
1336 				break;
1337 			}
1338 			pkt = seg;
1339 			assert(len >= (rxq->crc_present << 2));
1340 			pkt->ol_flags &= EXT_ATTACHED_MBUF;
1341 			/* If compressed, take hash result from mini-CQE. */
1342 			rss_hash_res = rte_be_to_cpu_32(mcqe == NULL ?
1343 							cqe->rx_hash_res :
1344 							mcqe->rx_hash_result);
1345 			rxq_cq_to_mbuf(rxq, pkt, cqe, rss_hash_res);
1346 			if (rxq->crc_present)
1347 				len -= RTE_ETHER_CRC_LEN;
1348 			PKT_LEN(pkt) = len;
1349 			if (cqe->lro_num_seg > 1) {
1350 				mlx5_lro_update_hdr
1351 					(rte_pktmbuf_mtod(pkt, uint8_t *), cqe,
1352 					 len);
1353 				pkt->ol_flags |= PKT_RX_LRO;
1354 				pkt->tso_segsz = len / cqe->lro_num_seg;
1355 			}
1356 		}
1357 		DATA_LEN(rep) = DATA_LEN(seg);
1358 		PKT_LEN(rep) = PKT_LEN(seg);
1359 		SET_DATA_OFF(rep, DATA_OFF(seg));
1360 		PORT(rep) = PORT(seg);
1361 		(*rxq->elts)[idx] = rep;
1362 		/*
1363 		 * Fill NIC descriptor with the new buffer.  The lkey and size
1364 		 * of the buffers are already known, only the buffer address
1365 		 * changes.
1366 		 */
1367 		wqe->addr = rte_cpu_to_be_64(rte_pktmbuf_mtod(rep, uintptr_t));
1368 		/* If there's only one MR, no need to replace LKey in WQE. */
1369 		if (unlikely(mlx5_mr_btree_len(&rxq->mr_ctrl.cache_bh) > 1))
1370 			wqe->lkey = mlx5_rx_mb2mr(rxq, rep);
1371 		if (len > DATA_LEN(seg)) {
1372 			len -= DATA_LEN(seg);
1373 			++NB_SEGS(pkt);
1374 			++rq_ci;
1375 			continue;
1376 		}
1377 		DATA_LEN(seg) = len;
1378 #ifdef MLX5_PMD_SOFT_COUNTERS
1379 		/* Increment bytes counter. */
1380 		rxq->stats.ibytes += PKT_LEN(pkt);
1381 #endif
1382 		/* Return packet. */
1383 		*(pkts++) = pkt;
1384 		pkt = NULL;
1385 		--pkts_n;
1386 		++i;
1387 		/* Align consumer index to the next stride. */
1388 		rq_ci >>= sges_n;
1389 		++rq_ci;
1390 		rq_ci <<= sges_n;
1391 	}
1392 	if (unlikely((i == 0) && ((rq_ci >> sges_n) == rxq->rq_ci)))
1393 		return 0;
1394 	/* Update the consumer index. */
1395 	rxq->rq_ci = rq_ci >> sges_n;
1396 	rte_cio_wmb();
1397 	*rxq->cq_db = rte_cpu_to_be_32(rxq->cq_ci);
1398 	rte_cio_wmb();
1399 	*rxq->rq_db = rte_cpu_to_be_32(rxq->rq_ci);
1400 #ifdef MLX5_PMD_SOFT_COUNTERS
1401 	/* Increment packets counter. */
1402 	rxq->stats.ipackets += i;
1403 #endif
1404 	return i;
1405 }
1406 
1407 /**
1408  * Update LRO packet TCP header.
1409  * The HW LRO feature doesn't update the TCP header after coalescing the
1410  * TCP segments but supplies information in CQE to fill it by SW.
1411  *
1412  * @param tcp
1413  *   Pointer to the TCP header.
1414  * @param cqe
1415  *   Pointer to the completion entry..
1416  * @param phcsum
1417  *   The L3 pseudo-header checksum.
1418  */
1419 static inline void
1420 mlx5_lro_update_tcp_hdr(struct rte_tcp_hdr *restrict tcp,
1421 			volatile struct mlx5_cqe *restrict cqe,
1422 			uint32_t phcsum)
1423 {
1424 	uint8_t l4_type = (rte_be_to_cpu_16(cqe->hdr_type_etc) &
1425 			   MLX5_CQE_L4_TYPE_MASK) >> MLX5_CQE_L4_TYPE_SHIFT;
1426 	/*
1427 	 * The HW calculates only the TCP payload checksum, need to complete
1428 	 * the TCP header checksum and the L3 pseudo-header checksum.
1429 	 */
1430 	uint32_t csum = phcsum + cqe->csum;
1431 
1432 	if (l4_type == MLX5_L4_HDR_TYPE_TCP_EMPTY_ACK ||
1433 	    l4_type == MLX5_L4_HDR_TYPE_TCP_WITH_ACL) {
1434 		tcp->tcp_flags |= RTE_TCP_ACK_FLAG;
1435 		tcp->recv_ack = cqe->lro_ack_seq_num;
1436 		tcp->rx_win = cqe->lro_tcp_win;
1437 	}
1438 	if (cqe->lro_tcppsh_abort_dupack & MLX5_CQE_LRO_PUSH_MASK)
1439 		tcp->tcp_flags |= RTE_TCP_PSH_FLAG;
1440 	tcp->cksum = 0;
1441 	csum += rte_raw_cksum(tcp, (tcp->data_off & 0xF) * 4);
1442 	csum = ((csum & 0xffff0000) >> 16) + (csum & 0xffff);
1443 	csum = (~csum) & 0xffff;
1444 	if (csum == 0)
1445 		csum = 0xffff;
1446 	tcp->cksum = csum;
1447 }
1448 
1449 /**
1450  * Update LRO packet headers.
1451  * The HW LRO feature doesn't update the L3/TCP headers after coalescing the
1452  * TCP segments but supply information in CQE to fill it by SW.
1453  *
1454  * @param padd
1455  *   The packet address.
1456  * @param cqe
1457  *   Pointer to the completion entry..
1458  * @param len
1459  *   The packet length.
1460  */
1461 static inline void
1462 mlx5_lro_update_hdr(uint8_t *restrict padd,
1463 		    volatile struct mlx5_cqe *restrict cqe,
1464 		    uint32_t len)
1465 {
1466 	union {
1467 		struct rte_ether_hdr *eth;
1468 		struct rte_vlan_hdr *vlan;
1469 		struct rte_ipv4_hdr *ipv4;
1470 		struct rte_ipv6_hdr *ipv6;
1471 		struct rte_tcp_hdr *tcp;
1472 		uint8_t *hdr;
1473 	} h = {
1474 			.hdr = padd,
1475 	};
1476 	uint16_t proto = h.eth->ether_type;
1477 	uint32_t phcsum;
1478 
1479 	h.eth++;
1480 	while (proto == RTE_BE16(RTE_ETHER_TYPE_VLAN) ||
1481 	       proto == RTE_BE16(RTE_ETHER_TYPE_QINQ)) {
1482 		proto = h.vlan->eth_proto;
1483 		h.vlan++;
1484 	}
1485 	if (proto == RTE_BE16(RTE_ETHER_TYPE_IPV4)) {
1486 		h.ipv4->time_to_live = cqe->lro_min_ttl;
1487 		h.ipv4->total_length = rte_cpu_to_be_16(len - (h.hdr - padd));
1488 		h.ipv4->hdr_checksum = 0;
1489 		h.ipv4->hdr_checksum = rte_ipv4_cksum(h.ipv4);
1490 		phcsum = rte_ipv4_phdr_cksum(h.ipv4, 0);
1491 		h.ipv4++;
1492 	} else {
1493 		h.ipv6->hop_limits = cqe->lro_min_ttl;
1494 		h.ipv6->payload_len = rte_cpu_to_be_16(len - (h.hdr - padd) -
1495 						       sizeof(*h.ipv6));
1496 		phcsum = rte_ipv6_phdr_cksum(h.ipv6, 0);
1497 		h.ipv6++;
1498 	}
1499 	mlx5_lro_update_tcp_hdr(h.tcp, cqe, phcsum);
1500 }
1501 
1502 void
1503 mlx5_mprq_buf_free_cb(void *addr __rte_unused, void *opaque)
1504 {
1505 	struct mlx5_mprq_buf *buf = opaque;
1506 
1507 	if (rte_atomic16_read(&buf->refcnt) == 1) {
1508 		rte_mempool_put(buf->mp, buf);
1509 	} else if (rte_atomic16_add_return(&buf->refcnt, -1) == 0) {
1510 		rte_atomic16_set(&buf->refcnt, 1);
1511 		rte_mempool_put(buf->mp, buf);
1512 	}
1513 }
1514 
1515 void
1516 mlx5_mprq_buf_free(struct mlx5_mprq_buf *buf)
1517 {
1518 	mlx5_mprq_buf_free_cb(NULL, buf);
1519 }
1520 
1521 static inline void
1522 mprq_buf_replace(struct mlx5_rxq_data *rxq, uint16_t rq_idx,
1523 		 const unsigned int strd_n)
1524 {
1525 	struct mlx5_mprq_buf *rep = rxq->mprq_repl;
1526 	volatile struct mlx5_wqe_data_seg *wqe =
1527 		&((volatile struct mlx5_wqe_mprq *)rxq->wqes)[rq_idx].dseg;
1528 	void *addr;
1529 
1530 	assert(rep != NULL);
1531 	/* Replace MPRQ buf. */
1532 	(*rxq->mprq_bufs)[rq_idx] = rep;
1533 	/* Replace WQE. */
1534 	addr = mlx5_mprq_buf_addr(rep, strd_n);
1535 	wqe->addr = rte_cpu_to_be_64((uintptr_t)addr);
1536 	/* If there's only one MR, no need to replace LKey in WQE. */
1537 	if (unlikely(mlx5_mr_btree_len(&rxq->mr_ctrl.cache_bh) > 1))
1538 		wqe->lkey = mlx5_rx_addr2mr(rxq, (uintptr_t)addr);
1539 	/* Stash a mbuf for next replacement. */
1540 	if (likely(!rte_mempool_get(rxq->mprq_mp, (void **)&rep)))
1541 		rxq->mprq_repl = rep;
1542 	else
1543 		rxq->mprq_repl = NULL;
1544 }
1545 
1546 /**
1547  * DPDK callback for RX with Multi-Packet RQ support.
1548  *
1549  * @param dpdk_rxq
1550  *   Generic pointer to RX queue structure.
1551  * @param[out] pkts
1552  *   Array to store received packets.
1553  * @param pkts_n
1554  *   Maximum number of packets in array.
1555  *
1556  * @return
1557  *   Number of packets successfully received (<= pkts_n).
1558  */
1559 uint16_t
1560 mlx5_rx_burst_mprq(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n)
1561 {
1562 	struct mlx5_rxq_data *rxq = dpdk_rxq;
1563 	const unsigned int strd_n = 1 << rxq->strd_num_n;
1564 	const unsigned int strd_sz = 1 << rxq->strd_sz_n;
1565 	const unsigned int strd_shift =
1566 		MLX5_MPRQ_STRIDE_SHIFT_BYTE * rxq->strd_shift_en;
1567 	const unsigned int cq_mask = (1 << rxq->cqe_n) - 1;
1568 	const unsigned int wq_mask = (1 << rxq->elts_n) - 1;
1569 	volatile struct mlx5_cqe *cqe = &(*rxq->cqes)[rxq->cq_ci & cq_mask];
1570 	unsigned int i = 0;
1571 	uint32_t rq_ci = rxq->rq_ci;
1572 	uint16_t consumed_strd = rxq->consumed_strd;
1573 	uint16_t headroom_sz = rxq->strd_headroom_en * RTE_PKTMBUF_HEADROOM;
1574 	struct mlx5_mprq_buf *buf = (*rxq->mprq_bufs)[rq_ci & wq_mask];
1575 
1576 	while (i < pkts_n) {
1577 		struct rte_mbuf *pkt;
1578 		void *addr;
1579 		int ret;
1580 		unsigned int len;
1581 		uint16_t strd_cnt;
1582 		uint16_t strd_idx;
1583 		uint32_t offset;
1584 		uint32_t byte_cnt;
1585 		volatile struct mlx5_mini_cqe8 *mcqe = NULL;
1586 		uint32_t rss_hash_res = 0;
1587 		uint8_t lro_num_seg;
1588 
1589 		if (consumed_strd == strd_n) {
1590 			/* Replace WQE only if the buffer is still in use. */
1591 			if (rte_atomic16_read(&buf->refcnt) > 1) {
1592 				mprq_buf_replace(rxq, rq_ci & wq_mask, strd_n);
1593 				/* Release the old buffer. */
1594 				mlx5_mprq_buf_free(buf);
1595 			} else if (unlikely(rxq->mprq_repl == NULL)) {
1596 				struct mlx5_mprq_buf *rep;
1597 
1598 				/*
1599 				 * Currently, the MPRQ mempool is out of buffer
1600 				 * and doing memcpy regardless of the size of Rx
1601 				 * packet. Retry allocation to get back to
1602 				 * normal.
1603 				 */
1604 				if (!rte_mempool_get(rxq->mprq_mp,
1605 						     (void **)&rep))
1606 					rxq->mprq_repl = rep;
1607 			}
1608 			/* Advance to the next WQE. */
1609 			consumed_strd = 0;
1610 			++rq_ci;
1611 			buf = (*rxq->mprq_bufs)[rq_ci & wq_mask];
1612 		}
1613 		cqe = &(*rxq->cqes)[rxq->cq_ci & cq_mask];
1614 		ret = mlx5_rx_poll_len(rxq, cqe, cq_mask, &mcqe);
1615 		if (!ret)
1616 			break;
1617 		byte_cnt = ret;
1618 		strd_cnt = (byte_cnt & MLX5_MPRQ_STRIDE_NUM_MASK) >>
1619 			   MLX5_MPRQ_STRIDE_NUM_SHIFT;
1620 		assert(strd_cnt);
1621 		consumed_strd += strd_cnt;
1622 		if (byte_cnt & MLX5_MPRQ_FILLER_MASK)
1623 			continue;
1624 		if (mcqe == NULL) {
1625 			rss_hash_res = rte_be_to_cpu_32(cqe->rx_hash_res);
1626 			strd_idx = rte_be_to_cpu_16(cqe->wqe_counter);
1627 		} else {
1628 			/* mini-CQE for MPRQ doesn't have hash result. */
1629 			strd_idx = rte_be_to_cpu_16(mcqe->stride_idx);
1630 		}
1631 		assert(strd_idx < strd_n);
1632 		assert(!((rte_be_to_cpu_16(cqe->wqe_id) ^ rq_ci) & wq_mask));
1633 		lro_num_seg = cqe->lro_num_seg;
1634 		/*
1635 		 * Currently configured to receive a packet per a stride. But if
1636 		 * MTU is adjusted through kernel interface, device could
1637 		 * consume multiple strides without raising an error. In this
1638 		 * case, the packet should be dropped because it is bigger than
1639 		 * the max_rx_pkt_len.
1640 		 */
1641 		if (unlikely(!lro_num_seg && strd_cnt > 1)) {
1642 			++rxq->stats.idropped;
1643 			continue;
1644 		}
1645 		pkt = rte_pktmbuf_alloc(rxq->mp);
1646 		if (unlikely(pkt == NULL)) {
1647 			++rxq->stats.rx_nombuf;
1648 			break;
1649 		}
1650 		len = (byte_cnt & MLX5_MPRQ_LEN_MASK) >> MLX5_MPRQ_LEN_SHIFT;
1651 		assert((int)len >= (rxq->crc_present << 2));
1652 		if (rxq->crc_present)
1653 			len -= RTE_ETHER_CRC_LEN;
1654 		offset = strd_idx * strd_sz + strd_shift;
1655 		addr = RTE_PTR_ADD(mlx5_mprq_buf_addr(buf, strd_n), offset);
1656 		/*
1657 		 * Memcpy packets to the target mbuf if:
1658 		 * - The size of packet is smaller than mprq_max_memcpy_len.
1659 		 * - Out of buffer in the Mempool for Multi-Packet RQ.
1660 		 */
1661 		if (len <= rxq->mprq_max_memcpy_len || rxq->mprq_repl == NULL) {
1662 			/*
1663 			 * When memcpy'ing packet due to out-of-buffer, the
1664 			 * packet must be smaller than the target mbuf.
1665 			 */
1666 			if (unlikely(rte_pktmbuf_tailroom(pkt) < len)) {
1667 				rte_pktmbuf_free_seg(pkt);
1668 				++rxq->stats.idropped;
1669 				continue;
1670 			}
1671 			rte_memcpy(rte_pktmbuf_mtod(pkt, void *), addr, len);
1672 			DATA_LEN(pkt) = len;
1673 		} else {
1674 			rte_iova_t buf_iova;
1675 			struct rte_mbuf_ext_shared_info *shinfo;
1676 			uint16_t buf_len = strd_cnt * strd_sz;
1677 			void *buf_addr;
1678 
1679 			/* Increment the refcnt of the whole chunk. */
1680 			rte_atomic16_add_return(&buf->refcnt, 1);
1681 			assert((uint16_t)rte_atomic16_read(&buf->refcnt) <=
1682 			       strd_n + 1);
1683 			buf_addr = RTE_PTR_SUB(addr, headroom_sz);
1684 			/*
1685 			 * MLX5 device doesn't use iova but it is necessary in a
1686 			 * case where the Rx packet is transmitted via a
1687 			 * different PMD.
1688 			 */
1689 			buf_iova = rte_mempool_virt2iova(buf) +
1690 				   RTE_PTR_DIFF(buf_addr, buf);
1691 			shinfo = &buf->shinfos[strd_idx];
1692 			rte_mbuf_ext_refcnt_set(shinfo, 1);
1693 			/*
1694 			 * EXT_ATTACHED_MBUF will be set to pkt->ol_flags when
1695 			 * attaching the stride to mbuf and more offload flags
1696 			 * will be added below by calling rxq_cq_to_mbuf().
1697 			 * Other fields will be overwritten.
1698 			 */
1699 			rte_pktmbuf_attach_extbuf(pkt, buf_addr, buf_iova,
1700 						  buf_len, shinfo);
1701 			/* Set mbuf head-room. */
1702 			pkt->data_off = headroom_sz;
1703 			assert(pkt->ol_flags == EXT_ATTACHED_MBUF);
1704 			/*
1705 			 * Prevent potential overflow due to MTU change through
1706 			 * kernel interface.
1707 			 */
1708 			if (unlikely(rte_pktmbuf_tailroom(pkt) < len)) {
1709 				rte_pktmbuf_free_seg(pkt);
1710 				++rxq->stats.idropped;
1711 				continue;
1712 			}
1713 			DATA_LEN(pkt) = len;
1714 			/*
1715 			 * LRO packet may consume all the stride memory, in this
1716 			 * case packet head-room space is not guaranteed so must
1717 			 * to add an empty mbuf for the head-room.
1718 			 */
1719 			if (!rxq->strd_headroom_en) {
1720 				struct rte_mbuf *headroom_mbuf =
1721 						rte_pktmbuf_alloc(rxq->mp);
1722 
1723 				if (unlikely(headroom_mbuf == NULL)) {
1724 					rte_pktmbuf_free_seg(pkt);
1725 					++rxq->stats.rx_nombuf;
1726 					break;
1727 				}
1728 				PORT(pkt) = rxq->port_id;
1729 				NEXT(headroom_mbuf) = pkt;
1730 				pkt = headroom_mbuf;
1731 				NB_SEGS(pkt) = 2;
1732 			}
1733 		}
1734 		rxq_cq_to_mbuf(rxq, pkt, cqe, rss_hash_res);
1735 		if (lro_num_seg > 1) {
1736 			mlx5_lro_update_hdr(addr, cqe, len);
1737 			pkt->ol_flags |= PKT_RX_LRO;
1738 			pkt->tso_segsz = strd_sz;
1739 		}
1740 		PKT_LEN(pkt) = len;
1741 		PORT(pkt) = rxq->port_id;
1742 #ifdef MLX5_PMD_SOFT_COUNTERS
1743 		/* Increment bytes counter. */
1744 		rxq->stats.ibytes += PKT_LEN(pkt);
1745 #endif
1746 		/* Return packet. */
1747 		*(pkts++) = pkt;
1748 		++i;
1749 	}
1750 	/* Update the consumer indexes. */
1751 	rxq->consumed_strd = consumed_strd;
1752 	rte_cio_wmb();
1753 	*rxq->cq_db = rte_cpu_to_be_32(rxq->cq_ci);
1754 	if (rq_ci != rxq->rq_ci) {
1755 		rxq->rq_ci = rq_ci;
1756 		rte_cio_wmb();
1757 		*rxq->rq_db = rte_cpu_to_be_32(rxq->rq_ci);
1758 	}
1759 #ifdef MLX5_PMD_SOFT_COUNTERS
1760 	/* Increment packets counter. */
1761 	rxq->stats.ipackets += i;
1762 #endif
1763 	return i;
1764 }
1765 
1766 /**
1767  * Dummy DPDK callback for TX.
1768  *
1769  * This function is used to temporarily replace the real callback during
1770  * unsafe control operations on the queue, or in case of error.
1771  *
1772  * @param dpdk_txq
1773  *   Generic pointer to TX queue structure.
1774  * @param[in] pkts
1775  *   Packets to transmit.
1776  * @param pkts_n
1777  *   Number of packets in array.
1778  *
1779  * @return
1780  *   Number of packets successfully transmitted (<= pkts_n).
1781  */
1782 uint16_t
1783 removed_tx_burst(void *dpdk_txq __rte_unused,
1784 		 struct rte_mbuf **pkts __rte_unused,
1785 		 uint16_t pkts_n __rte_unused)
1786 {
1787 	rte_mb();
1788 	return 0;
1789 }
1790 
1791 /**
1792  * Dummy DPDK callback for RX.
1793  *
1794  * This function is used to temporarily replace the real callback during
1795  * unsafe control operations on the queue, or in case of error.
1796  *
1797  * @param dpdk_rxq
1798  *   Generic pointer to RX queue structure.
1799  * @param[out] pkts
1800  *   Array to store received packets.
1801  * @param pkts_n
1802  *   Maximum number of packets in array.
1803  *
1804  * @return
1805  *   Number of packets successfully received (<= pkts_n).
1806  */
1807 uint16_t
1808 removed_rx_burst(void *dpdk_txq __rte_unused,
1809 		 struct rte_mbuf **pkts __rte_unused,
1810 		 uint16_t pkts_n __rte_unused)
1811 {
1812 	rte_mb();
1813 	return 0;
1814 }
1815 
1816 /*
1817  * Vectorized Rx/Tx routines are not compiled in when required vector
1818  * instructions are not supported on a target architecture. The following null
1819  * stubs are needed for linkage when those are not included outside of this file
1820  * (e.g.  mlx5_rxtx_vec_sse.c for x86).
1821  */
1822 
1823 __rte_weak uint16_t
1824 mlx5_rx_burst_vec(void *dpdk_txq __rte_unused,
1825 		  struct rte_mbuf **pkts __rte_unused,
1826 		  uint16_t pkts_n __rte_unused)
1827 {
1828 	return 0;
1829 }
1830 
1831 __rte_weak int
1832 mlx5_rxq_check_vec_support(struct mlx5_rxq_data *rxq __rte_unused)
1833 {
1834 	return -ENOTSUP;
1835 }
1836 
1837 __rte_weak int
1838 mlx5_check_vec_rx_support(struct rte_eth_dev *dev __rte_unused)
1839 {
1840 	return -ENOTSUP;
1841 }
1842 
1843 /**
1844  * Free the mbufs from the linear array of pointers.
1845  *
1846  * @param pkts
1847  *   Pointer to array of packets to be free.
1848  * @param pkts_n
1849  *   Number of packets to be freed.
1850  * @param olx
1851  *   Configured Tx offloads mask. It is fully defined at
1852  *   compile time and may be used for optimization.
1853  */
1854 static __rte_always_inline void
1855 mlx5_tx_free_mbuf(struct rte_mbuf **restrict pkts,
1856 		  unsigned int pkts_n,
1857 		  unsigned int olx __rte_unused)
1858 {
1859 	struct rte_mempool *pool = NULL;
1860 	struct rte_mbuf **p_free = NULL;
1861 	struct rte_mbuf *mbuf;
1862 	unsigned int n_free = 0;
1863 
1864 	/*
1865 	 * The implemented algorithm eliminates
1866 	 * copying pointers to temporary array
1867 	 * for rte_mempool_put_bulk() calls.
1868 	 */
1869 	assert(pkts);
1870 	assert(pkts_n);
1871 	for (;;) {
1872 		for (;;) {
1873 			/*
1874 			 * Decrement mbuf reference counter, detach
1875 			 * indirect and external buffers if needed.
1876 			 */
1877 			mbuf = rte_pktmbuf_prefree_seg(*pkts);
1878 			if (likely(mbuf != NULL)) {
1879 				assert(mbuf == *pkts);
1880 				if (likely(n_free != 0)) {
1881 					if (unlikely(pool != mbuf->pool))
1882 						/* From different pool. */
1883 						break;
1884 				} else {
1885 					/* Start new scan array. */
1886 					pool = mbuf->pool;
1887 					p_free = pkts;
1888 				}
1889 				++n_free;
1890 				++pkts;
1891 				--pkts_n;
1892 				if (unlikely(pkts_n == 0)) {
1893 					mbuf = NULL;
1894 					break;
1895 				}
1896 			} else {
1897 				/*
1898 				 * This happens if mbuf is still referenced.
1899 				 * We can't put it back to the pool, skip.
1900 				 */
1901 				++pkts;
1902 				--pkts_n;
1903 				if (unlikely(n_free != 0))
1904 					/* There is some array to free.*/
1905 					break;
1906 				if (unlikely(pkts_n == 0))
1907 					/* Last mbuf, nothing to free. */
1908 					return;
1909 			}
1910 		}
1911 		for (;;) {
1912 			/*
1913 			 * This loop is implemented to avoid multiple
1914 			 * inlining of rte_mempool_put_bulk().
1915 			 */
1916 			assert(pool);
1917 			assert(p_free);
1918 			assert(n_free);
1919 			/*
1920 			 * Free the array of pre-freed mbufs
1921 			 * belonging to the same memory pool.
1922 			 */
1923 			rte_mempool_put_bulk(pool, (void *)p_free, n_free);
1924 			if (unlikely(mbuf != NULL)) {
1925 				/* There is the request to start new scan. */
1926 				pool = mbuf->pool;
1927 				p_free = pkts++;
1928 				n_free = 1;
1929 				--pkts_n;
1930 				if (likely(pkts_n != 0))
1931 					break;
1932 				/*
1933 				 * This is the last mbuf to be freed.
1934 				 * Do one more loop iteration to complete.
1935 				 * This is rare case of the last unique mbuf.
1936 				 */
1937 				mbuf = NULL;
1938 				continue;
1939 			}
1940 			if (likely(pkts_n == 0))
1941 				return;
1942 			n_free = 0;
1943 			break;
1944 		}
1945 	}
1946 }
1947 
1948 /**
1949  * Free the mbuf from the elts ring buffer till new tail.
1950  *
1951  * @param txq
1952  *   Pointer to Tx queue structure.
1953  * @param tail
1954  *   Index in elts to free up to, becomes new elts tail.
1955  * @param olx
1956  *   Configured Tx offloads mask. It is fully defined at
1957  *   compile time and may be used for optimization.
1958  */
1959 static __rte_always_inline void
1960 mlx5_tx_free_elts(struct mlx5_txq_data *restrict txq,
1961 		  uint16_t tail,
1962 		  unsigned int olx __rte_unused)
1963 {
1964 	uint16_t n_elts = tail - txq->elts_tail;
1965 
1966 	assert(n_elts);
1967 	assert(n_elts <= txq->elts_s);
1968 	/*
1969 	 * Implement a loop to support ring buffer wraparound
1970 	 * with single inlining of mlx5_tx_free_mbuf().
1971 	 */
1972 	do {
1973 		unsigned int part;
1974 
1975 		part = txq->elts_s - (txq->elts_tail & txq->elts_m);
1976 		part = RTE_MIN(part, n_elts);
1977 		assert(part);
1978 		assert(part <= txq->elts_s);
1979 		mlx5_tx_free_mbuf(&txq->elts[txq->elts_tail & txq->elts_m],
1980 				  part, olx);
1981 		txq->elts_tail += part;
1982 		n_elts -= part;
1983 	} while (n_elts);
1984 }
1985 
1986 /**
1987  * Store the mbuf being sent into elts ring buffer.
1988  * On Tx completion these mbufs will be freed.
1989  *
1990  * @param txq
1991  *   Pointer to Tx queue structure.
1992  * @param pkts
1993  *   Pointer to array of packets to be stored.
1994  * @param pkts_n
1995  *   Number of packets to be stored.
1996  * @param olx
1997  *   Configured Tx offloads mask. It is fully defined at
1998  *   compile time and may be used for optimization.
1999  */
2000 static __rte_always_inline void
2001 mlx5_tx_copy_elts(struct mlx5_txq_data *restrict txq,
2002 		  struct rte_mbuf **restrict pkts,
2003 		  unsigned int pkts_n,
2004 		  unsigned int olx __rte_unused)
2005 {
2006 	unsigned int part;
2007 	struct rte_mbuf **elts = (struct rte_mbuf **)txq->elts;
2008 
2009 	assert(pkts);
2010 	assert(pkts_n);
2011 	part = txq->elts_s - (txq->elts_head & txq->elts_m);
2012 	assert(part);
2013 	assert(part <= txq->elts_s);
2014 	/* This code is a good candidate for vectorizing with SIMD. */
2015 	rte_memcpy((void *)(elts + (txq->elts_head & txq->elts_m)),
2016 		   (void *)pkts,
2017 		   RTE_MIN(part, pkts_n) * sizeof(struct rte_mbuf *));
2018 	txq->elts_head += pkts_n;
2019 	if (unlikely(part < pkts_n))
2020 		/* The copy is wrapping around the elts array. */
2021 		rte_memcpy((void *)elts, (void *)(pkts + part),
2022 			   (pkts_n - part) * sizeof(struct rte_mbuf *));
2023 }
2024 
2025 /**
2026  * Update completion queue consuming index via doorbell
2027  * and flush the completed data buffers.
2028  *
2029  * @param txq
2030  *   Pointer to TX queue structure.
2031  * @param valid CQE pointer
2032  *   if not NULL update txq->wqe_pi and flush the buffers
2033  * @param olx
2034  *   Configured Tx offloads mask. It is fully defined at
2035  *   compile time and may be used for optimization.
2036  */
2037 static __rte_always_inline void
2038 mlx5_tx_comp_flush(struct mlx5_txq_data *restrict txq,
2039 		   volatile struct mlx5_cqe *last_cqe,
2040 		   unsigned int olx __rte_unused)
2041 {
2042 	if (likely(last_cqe != NULL)) {
2043 		uint16_t tail;
2044 
2045 		txq->wqe_pi = rte_be_to_cpu_16(last_cqe->wqe_counter);
2046 		tail = txq->fcqs[(txq->cq_ci - 1) & txq->cqe_m];
2047 		if (likely(tail != txq->elts_tail)) {
2048 			mlx5_tx_free_elts(txq, tail, olx);
2049 			assert(tail == txq->elts_tail);
2050 		}
2051 	}
2052 }
2053 
2054 /**
2055  * Manage TX completions. This routine checks the CQ for
2056  * arrived CQEs, deduces the last accomplished WQE in SQ,
2057  * updates SQ producing index and frees all completed mbufs.
2058  *
2059  * @param txq
2060  *   Pointer to TX queue structure.
2061  * @param olx
2062  *   Configured Tx offloads mask. It is fully defined at
2063  *   compile time and may be used for optimization.
2064  *
2065  * NOTE: not inlined intentionally, it makes tx_burst
2066  * routine smaller, simple and faster - from experiments.
2067  */
2068 static void
2069 mlx5_tx_handle_completion(struct mlx5_txq_data *restrict txq,
2070 			  unsigned int olx __rte_unused)
2071 {
2072 	unsigned int count = MLX5_TX_COMP_MAX_CQE;
2073 	volatile struct mlx5_cqe *last_cqe = NULL;
2074 	uint16_t ci = txq->cq_ci;
2075 	int ret;
2076 
2077 	static_assert(MLX5_CQE_STATUS_HW_OWN < 0, "Must be negative value");
2078 	static_assert(MLX5_CQE_STATUS_SW_OWN < 0, "Must be negative value");
2079 	do {
2080 		volatile struct mlx5_cqe *cqe;
2081 
2082 		cqe = &txq->cqes[ci & txq->cqe_m];
2083 		ret = check_cqe(cqe, txq->cqe_s, ci);
2084 		if (unlikely(ret != MLX5_CQE_STATUS_SW_OWN)) {
2085 			if (likely(ret != MLX5_CQE_STATUS_ERR)) {
2086 				/* No new CQEs in completion queue. */
2087 				assert(ret == MLX5_CQE_STATUS_HW_OWN);
2088 				break;
2089 			}
2090 			/*
2091 			 * Some error occurred, try to restart.
2092 			 * We have no barrier after WQE related Doorbell
2093 			 * written, make sure all writes are completed
2094 			 * here, before we might perform SQ reset.
2095 			 */
2096 			rte_wmb();
2097 			txq->cq_ci = ci;
2098 			ret = mlx5_tx_error_cqe_handle
2099 				(txq, (volatile struct mlx5_err_cqe *)cqe);
2100 			if (unlikely(ret < 0)) {
2101 				/*
2102 				 * Some error occurred on queue error
2103 				 * handling, we do not advance the index
2104 				 * here, allowing to retry on next call.
2105 				 */
2106 				return;
2107 			}
2108 			/*
2109 			 * We are going to fetch all entries with
2110 			 * MLX5_CQE_SYNDROME_WR_FLUSH_ERR status.
2111 			 * The send queue is supposed to be empty.
2112 			 */
2113 			++ci;
2114 			txq->cq_pi = ci;
2115 			last_cqe = NULL;
2116 			continue;
2117 		}
2118 		/* Normal transmit completion. */
2119 		assert(ci != txq->cq_pi);
2120 		assert((txq->fcqs[ci & txq->cqe_m] >> 16) == cqe->wqe_counter);
2121 		++ci;
2122 		last_cqe = cqe;
2123 		/*
2124 		 * We have to restrict the amount of processed CQEs
2125 		 * in one tx_burst routine call. The CQ may be large
2126 		 * and many CQEs may be updated by the NIC in one
2127 		 * transaction. Buffers freeing is time consuming,
2128 		 * multiple iterations may introduce significant
2129 		 * latency.
2130 		 */
2131 		if (likely(--count == 0))
2132 			break;
2133 	} while (true);
2134 	if (likely(ci != txq->cq_ci)) {
2135 		/*
2136 		 * Update completion queue consuming index
2137 		 * and ring doorbell to notify hardware.
2138 		 */
2139 		rte_compiler_barrier();
2140 		txq->cq_ci = ci;
2141 		*txq->cq_db = rte_cpu_to_be_32(ci);
2142 		mlx5_tx_comp_flush(txq, last_cqe, olx);
2143 	}
2144 }
2145 
2146 /**
2147  * Check if the completion request flag should be set in the last WQE.
2148  * Both pushed mbufs and WQEs are monitored and the completion request
2149  * flag is set if any of thresholds is reached.
2150  *
2151  * @param txq
2152  *   Pointer to TX queue structure.
2153  * @param loc
2154  *   Pointer to burst routine local context.
2155  * @param olx
2156  *   Configured Tx offloads mask. It is fully defined at
2157  *   compile time and may be used for optimization.
2158  */
2159 static __rte_always_inline void
2160 mlx5_tx_request_completion(struct mlx5_txq_data *restrict txq,
2161 			   struct mlx5_txq_local *restrict loc,
2162 			   unsigned int olx)
2163 {
2164 	uint16_t head = txq->elts_head;
2165 	unsigned int part;
2166 
2167 	part = MLX5_TXOFF_CONFIG(INLINE) ?
2168 	       0 : loc->pkts_sent - loc->pkts_copy;
2169 	head += part;
2170 	if ((uint16_t)(head - txq->elts_comp) >= MLX5_TX_COMP_THRESH ||
2171 	     (MLX5_TXOFF_CONFIG(INLINE) &&
2172 	     (uint16_t)(txq->wqe_ci - txq->wqe_comp) >= txq->wqe_thres)) {
2173 		volatile struct mlx5_wqe *last = loc->wqe_last;
2174 
2175 		txq->elts_comp = head;
2176 		if (MLX5_TXOFF_CONFIG(INLINE))
2177 			txq->wqe_comp = txq->wqe_ci;
2178 		/* Request unconditional completion on last WQE. */
2179 		last->cseg.flags = RTE_BE32(MLX5_COMP_ALWAYS <<
2180 					    MLX5_COMP_MODE_OFFSET);
2181 		/* Save elts_head in dedicated free on completion queue. */
2182 #ifdef NDEBUG
2183 		txq->fcqs[txq->cq_pi++ & txq->cqe_m] = head;
2184 #else
2185 		txq->fcqs[txq->cq_pi++ & txq->cqe_m] = head |
2186 					(last->cseg.opcode >> 8) << 16;
2187 #endif
2188 		/* A CQE slot must always be available. */
2189 		assert((txq->cq_pi - txq->cq_ci) <= txq->cqe_s);
2190 	}
2191 }
2192 
2193 /**
2194  * DPDK callback to check the status of a tx descriptor.
2195  *
2196  * @param tx_queue
2197  *   The tx queue.
2198  * @param[in] offset
2199  *   The index of the descriptor in the ring.
2200  *
2201  * @return
2202  *   The status of the tx descriptor.
2203  */
2204 int
2205 mlx5_tx_descriptor_status(void *tx_queue, uint16_t offset)
2206 {
2207 	struct mlx5_txq_data *restrict txq = tx_queue;
2208 	uint16_t used;
2209 
2210 	mlx5_tx_handle_completion(txq, 0);
2211 	used = txq->elts_head - txq->elts_tail;
2212 	if (offset < used)
2213 		return RTE_ETH_TX_DESC_FULL;
2214 	return RTE_ETH_TX_DESC_DONE;
2215 }
2216 
2217 /**
2218  * Build the Control Segment with specified opcode:
2219  * - MLX5_OPCODE_SEND
2220  * - MLX5_OPCODE_ENHANCED_MPSW
2221  * - MLX5_OPCODE_TSO
2222  *
2223  * @param txq
2224  *   Pointer to TX queue structure.
2225  * @param loc
2226  *   Pointer to burst routine local context.
2227  * @param wqe
2228  *   Pointer to WQE to fill with built Control Segment.
2229  * @param ds
2230  *   Supposed length of WQE in segments.
2231  * @param opcode
2232  *   SQ WQE opcode to put into Control Segment.
2233  * @param olx
2234  *   Configured Tx offloads mask. It is fully defined at
2235  *   compile time and may be used for optimization.
2236  */
2237 static __rte_always_inline void
2238 mlx5_tx_cseg_init(struct mlx5_txq_data *restrict txq,
2239 		  struct mlx5_txq_local *restrict loc __rte_unused,
2240 		  struct mlx5_wqe *restrict wqe,
2241 		  unsigned int ds,
2242 		  unsigned int opcode,
2243 		  unsigned int olx __rte_unused)
2244 {
2245 	struct mlx5_wqe_cseg *restrict cs = &wqe->cseg;
2246 
2247 	/* For legacy MPW replace the EMPW by TSO with modifier. */
2248 	if (MLX5_TXOFF_CONFIG(MPW) && opcode == MLX5_OPCODE_ENHANCED_MPSW)
2249 		opcode = MLX5_OPCODE_TSO | MLX5_OPC_MOD_MPW << 24;
2250 	cs->opcode = rte_cpu_to_be_32((txq->wqe_ci << 8) | opcode);
2251 	cs->sq_ds = rte_cpu_to_be_32(txq->qp_num_8s | ds);
2252 	cs->flags = RTE_BE32(MLX5_COMP_ONLY_FIRST_ERR <<
2253 			     MLX5_COMP_MODE_OFFSET);
2254 	cs->misc = RTE_BE32(0);
2255 }
2256 
2257 /**
2258  * Build the Ethernet Segment without inlined data.
2259  * Supports Software Parser, Checksums and VLAN
2260  * insertion Tx offload features.
2261  *
2262  * @param txq
2263  *   Pointer to TX queue structure.
2264  * @param loc
2265  *   Pointer to burst routine local context.
2266  * @param wqe
2267  *   Pointer to WQE to fill with built Ethernet Segment.
2268  * @param olx
2269  *   Configured Tx offloads mask. It is fully defined at
2270  *   compile time and may be used for optimization.
2271  */
2272 static __rte_always_inline void
2273 mlx5_tx_eseg_none(struct mlx5_txq_data *restrict txq __rte_unused,
2274 		  struct mlx5_txq_local *restrict loc,
2275 		  struct mlx5_wqe *restrict wqe,
2276 		  unsigned int olx)
2277 {
2278 	struct mlx5_wqe_eseg *restrict es = &wqe->eseg;
2279 	uint32_t csum;
2280 
2281 	/*
2282 	 * Calculate and set check sum flags first, dword field
2283 	 * in segment may be shared with Software Parser flags.
2284 	 */
2285 	csum = MLX5_TXOFF_CONFIG(CSUM) ? txq_ol_cksum_to_cs(loc->mbuf) : 0;
2286 	es->flags = rte_cpu_to_le_32(csum);
2287 	/*
2288 	 * Calculate and set Software Parser offsets and flags.
2289 	 * These flags a set for custom UDP and IP tunnel packets.
2290 	 */
2291 	es->swp_offs = txq_mbuf_to_swp(loc, &es->swp_flags, olx);
2292 	/* Fill metadata field if needed. */
2293 	es->metadata = MLX5_TXOFF_CONFIG(METADATA) ?
2294 		       loc->mbuf->ol_flags & PKT_TX_DYNF_METADATA ?
2295 		       *RTE_FLOW_DYNF_METADATA(loc->mbuf) : 0 : 0;
2296 	/* Engage VLAN tag insertion feature if requested. */
2297 	if (MLX5_TXOFF_CONFIG(VLAN) &&
2298 	    loc->mbuf->ol_flags & PKT_TX_VLAN_PKT) {
2299 		/*
2300 		 * We should get here only if device support
2301 		 * this feature correctly.
2302 		 */
2303 		assert(txq->vlan_en);
2304 		es->inline_hdr = rte_cpu_to_be_32(MLX5_ETH_WQE_VLAN_INSERT |
2305 						  loc->mbuf->vlan_tci);
2306 	} else {
2307 		es->inline_hdr = RTE_BE32(0);
2308 	}
2309 }
2310 
2311 /**
2312  * Build the Ethernet Segment with minimal inlined data
2313  * of MLX5_ESEG_MIN_INLINE_SIZE bytes length. This is
2314  * used to fill the gap in single WQEBB WQEs.
2315  * Supports Software Parser, Checksums and VLAN
2316  * insertion Tx offload features.
2317  *
2318  * @param txq
2319  *   Pointer to TX queue structure.
2320  * @param loc
2321  *   Pointer to burst routine local context.
2322  * @param wqe
2323  *   Pointer to WQE to fill with built Ethernet Segment.
2324  * @param vlan
2325  *   Length of VLAN tag insertion if any.
2326  * @param olx
2327  *   Configured Tx offloads mask. It is fully defined at
2328  *   compile time and may be used for optimization.
2329  */
2330 static __rte_always_inline void
2331 mlx5_tx_eseg_dmin(struct mlx5_txq_data *restrict txq __rte_unused,
2332 		  struct mlx5_txq_local *restrict loc,
2333 		  struct mlx5_wqe *restrict wqe,
2334 		  unsigned int vlan,
2335 		  unsigned int olx)
2336 {
2337 	struct mlx5_wqe_eseg *restrict es = &wqe->eseg;
2338 	uint32_t csum;
2339 	uint8_t *psrc, *pdst;
2340 
2341 	/*
2342 	 * Calculate and set check sum flags first, dword field
2343 	 * in segment may be shared with Software Parser flags.
2344 	 */
2345 	csum = MLX5_TXOFF_CONFIG(CSUM) ? txq_ol_cksum_to_cs(loc->mbuf) : 0;
2346 	es->flags = rte_cpu_to_le_32(csum);
2347 	/*
2348 	 * Calculate and set Software Parser offsets and flags.
2349 	 * These flags a set for custom UDP and IP tunnel packets.
2350 	 */
2351 	es->swp_offs = txq_mbuf_to_swp(loc, &es->swp_flags, olx);
2352 	/* Fill metadata field if needed. */
2353 	es->metadata = MLX5_TXOFF_CONFIG(METADATA) ?
2354 		       loc->mbuf->ol_flags & PKT_TX_DYNF_METADATA ?
2355 		       *RTE_FLOW_DYNF_METADATA(loc->mbuf) : 0 : 0;
2356 	static_assert(MLX5_ESEG_MIN_INLINE_SIZE ==
2357 				(sizeof(uint16_t) +
2358 				 sizeof(rte_v128u32_t)),
2359 		      "invalid Ethernet Segment data size");
2360 	static_assert(MLX5_ESEG_MIN_INLINE_SIZE ==
2361 				(sizeof(uint16_t) +
2362 				 sizeof(struct rte_vlan_hdr) +
2363 				 2 * RTE_ETHER_ADDR_LEN),
2364 		      "invalid Ethernet Segment data size");
2365 	psrc = rte_pktmbuf_mtod(loc->mbuf, uint8_t *);
2366 	es->inline_hdr_sz = RTE_BE16(MLX5_ESEG_MIN_INLINE_SIZE);
2367 	es->inline_data = *(unaligned_uint16_t *)psrc;
2368 	psrc +=	sizeof(uint16_t);
2369 	pdst = (uint8_t *)(es + 1);
2370 	if (MLX5_TXOFF_CONFIG(VLAN) && vlan) {
2371 		/* Implement VLAN tag insertion as part inline data. */
2372 		memcpy(pdst, psrc, 2 * RTE_ETHER_ADDR_LEN - sizeof(uint16_t));
2373 		pdst += 2 * RTE_ETHER_ADDR_LEN - sizeof(uint16_t);
2374 		psrc +=	2 * RTE_ETHER_ADDR_LEN - sizeof(uint16_t);
2375 		/* Insert VLAN ethertype + VLAN tag. */
2376 		*(unaligned_uint32_t *)pdst = rte_cpu_to_be_32
2377 						((RTE_ETHER_TYPE_VLAN << 16) |
2378 						 loc->mbuf->vlan_tci);
2379 		pdst += sizeof(struct rte_vlan_hdr);
2380 		/* Copy the rest two bytes from packet data. */
2381 		assert(pdst == RTE_PTR_ALIGN(pdst, sizeof(uint16_t)));
2382 		*(uint16_t *)pdst = *(unaligned_uint16_t *)psrc;
2383 	} else {
2384 		/* Fill the gap in the title WQEBB with inline data. */
2385 		rte_mov16(pdst, psrc);
2386 	}
2387 }
2388 
2389 /**
2390  * Build the Ethernet Segment with entire packet
2391  * data inlining. Checks the boundary of WQEBB and
2392  * ring buffer wrapping, supports Software Parser,
2393  * Checksums and VLAN insertion Tx offload features.
2394  *
2395  * @param txq
2396  *   Pointer to TX queue structure.
2397  * @param loc
2398  *   Pointer to burst routine local context.
2399  * @param wqe
2400  *   Pointer to WQE to fill with built Ethernet Segment.
2401  * @param vlan
2402  *   Length of VLAN tag insertion if any.
2403  * @param inlen
2404  *   Length of data to inline (VLAN included, if any).
2405  * @param tso
2406  *   TSO flag, set mss field from the packet.
2407  * @param olx
2408  *   Configured Tx offloads mask. It is fully defined at
2409  *   compile time and may be used for optimization.
2410  *
2411  * @return
2412  *   Pointer to the next Data Segment (aligned and wrapped around).
2413  */
2414 static __rte_always_inline struct mlx5_wqe_dseg *
2415 mlx5_tx_eseg_data(struct mlx5_txq_data *restrict txq,
2416 		  struct mlx5_txq_local *restrict loc,
2417 		  struct mlx5_wqe *restrict wqe,
2418 		  unsigned int vlan,
2419 		  unsigned int inlen,
2420 		  unsigned int tso,
2421 		  unsigned int olx)
2422 {
2423 	struct mlx5_wqe_eseg *restrict es = &wqe->eseg;
2424 	uint32_t csum;
2425 	uint8_t *psrc, *pdst;
2426 	unsigned int part;
2427 
2428 	/*
2429 	 * Calculate and set check sum flags first, dword field
2430 	 * in segment may be shared with Software Parser flags.
2431 	 */
2432 	csum = MLX5_TXOFF_CONFIG(CSUM) ? txq_ol_cksum_to_cs(loc->mbuf) : 0;
2433 	if (tso) {
2434 		csum <<= 24;
2435 		csum |= loc->mbuf->tso_segsz;
2436 		es->flags = rte_cpu_to_be_32(csum);
2437 	} else {
2438 		es->flags = rte_cpu_to_le_32(csum);
2439 	}
2440 	/*
2441 	 * Calculate and set Software Parser offsets and flags.
2442 	 * These flags a set for custom UDP and IP tunnel packets.
2443 	 */
2444 	es->swp_offs = txq_mbuf_to_swp(loc, &es->swp_flags, olx);
2445 	/* Fill metadata field if needed. */
2446 	es->metadata = MLX5_TXOFF_CONFIG(METADATA) ?
2447 		       loc->mbuf->ol_flags & PKT_TX_DYNF_METADATA ?
2448 		       *RTE_FLOW_DYNF_METADATA(loc->mbuf) : 0 : 0;
2449 	static_assert(MLX5_ESEG_MIN_INLINE_SIZE ==
2450 				(sizeof(uint16_t) +
2451 				 sizeof(rte_v128u32_t)),
2452 		      "invalid Ethernet Segment data size");
2453 	static_assert(MLX5_ESEG_MIN_INLINE_SIZE ==
2454 				(sizeof(uint16_t) +
2455 				 sizeof(struct rte_vlan_hdr) +
2456 				 2 * RTE_ETHER_ADDR_LEN),
2457 		      "invalid Ethernet Segment data size");
2458 	psrc = rte_pktmbuf_mtod(loc->mbuf, uint8_t *);
2459 	es->inline_hdr_sz = rte_cpu_to_be_16(inlen);
2460 	es->inline_data = *(unaligned_uint16_t *)psrc;
2461 	psrc +=	sizeof(uint16_t);
2462 	pdst = (uint8_t *)(es + 1);
2463 	if (MLX5_TXOFF_CONFIG(VLAN) && vlan) {
2464 		/* Implement VLAN tag insertion as part inline data. */
2465 		memcpy(pdst, psrc, 2 * RTE_ETHER_ADDR_LEN - sizeof(uint16_t));
2466 		pdst += 2 * RTE_ETHER_ADDR_LEN - sizeof(uint16_t);
2467 		psrc +=	2 * RTE_ETHER_ADDR_LEN - sizeof(uint16_t);
2468 		/* Insert VLAN ethertype + VLAN tag. */
2469 		*(unaligned_uint32_t *)pdst = rte_cpu_to_be_32
2470 						((RTE_ETHER_TYPE_VLAN << 16) |
2471 						 loc->mbuf->vlan_tci);
2472 		pdst += sizeof(struct rte_vlan_hdr);
2473 		/* Copy the rest two bytes from packet data. */
2474 		assert(pdst == RTE_PTR_ALIGN(pdst, sizeof(uint16_t)));
2475 		*(uint16_t *)pdst = *(unaligned_uint16_t *)psrc;
2476 		psrc += sizeof(uint16_t);
2477 	} else {
2478 		/* Fill the gap in the title WQEBB with inline data. */
2479 		rte_mov16(pdst, psrc);
2480 		psrc += sizeof(rte_v128u32_t);
2481 	}
2482 	pdst = (uint8_t *)(es + 2);
2483 	assert(inlen >= MLX5_ESEG_MIN_INLINE_SIZE);
2484 	assert(pdst < (uint8_t *)txq->wqes_end);
2485 	inlen -= MLX5_ESEG_MIN_INLINE_SIZE;
2486 	if (!inlen) {
2487 		assert(pdst == RTE_PTR_ALIGN(pdst, MLX5_WSEG_SIZE));
2488 		return (struct mlx5_wqe_dseg *)pdst;
2489 	}
2490 	/*
2491 	 * The WQEBB space availability is checked by caller.
2492 	 * Here we should be aware of WQE ring buffer wraparound only.
2493 	 */
2494 	part = (uint8_t *)txq->wqes_end - pdst;
2495 	part = RTE_MIN(part, inlen);
2496 	do {
2497 		rte_memcpy(pdst, psrc, part);
2498 		inlen -= part;
2499 		if (likely(!inlen)) {
2500 			/*
2501 			 * If return value is not used by the caller
2502 			 * the code below will be optimized out.
2503 			 */
2504 			pdst += part;
2505 			pdst = RTE_PTR_ALIGN(pdst, MLX5_WSEG_SIZE);
2506 			if (unlikely(pdst >= (uint8_t *)txq->wqes_end))
2507 				pdst = (uint8_t *)txq->wqes;
2508 			return (struct mlx5_wqe_dseg *)pdst;
2509 		}
2510 		pdst = (uint8_t *)txq->wqes;
2511 		psrc += part;
2512 		part = inlen;
2513 	} while (true);
2514 }
2515 
2516 /**
2517  * Copy data from chain of mbuf to the specified linear buffer.
2518  * Checksums and VLAN insertion Tx offload features. If data
2519  * from some mbuf copied completely this mbuf is freed. Local
2520  * structure is used to keep the byte stream state.
2521  *
2522  * @param pdst
2523  *   Pointer to the destination linear buffer.
2524  * @param loc
2525  *   Pointer to burst routine local context.
2526  * @param len
2527  *   Length of data to be copied.
2528  * @param olx
2529  *   Configured Tx offloads mask. It is fully defined at
2530  *   compile time and may be used for optimization.
2531  */
2532 static __rte_always_inline void
2533 mlx5_tx_mseg_memcpy(uint8_t *pdst,
2534 		    struct mlx5_txq_local *restrict loc,
2535 		    unsigned int len,
2536 		    unsigned int olx __rte_unused)
2537 {
2538 	struct rte_mbuf *mbuf;
2539 	unsigned int part, dlen;
2540 	uint8_t *psrc;
2541 
2542 	assert(len);
2543 	do {
2544 		/* Allow zero length packets, must check first. */
2545 		dlen = rte_pktmbuf_data_len(loc->mbuf);
2546 		if (dlen <= loc->mbuf_off) {
2547 			/* Exhausted packet, just free. */
2548 			mbuf = loc->mbuf;
2549 			loc->mbuf = mbuf->next;
2550 			rte_pktmbuf_free_seg(mbuf);
2551 			loc->mbuf_off = 0;
2552 			assert(loc->mbuf_nseg > 1);
2553 			assert(loc->mbuf);
2554 			--loc->mbuf_nseg;
2555 			continue;
2556 		}
2557 		dlen -= loc->mbuf_off;
2558 		psrc = rte_pktmbuf_mtod_offset(loc->mbuf, uint8_t *,
2559 					       loc->mbuf_off);
2560 		part = RTE_MIN(len, dlen);
2561 		rte_memcpy(pdst, psrc, part);
2562 		loc->mbuf_off += part;
2563 		len -= part;
2564 		if (!len) {
2565 			if (loc->mbuf_off >= rte_pktmbuf_data_len(loc->mbuf)) {
2566 				loc->mbuf_off = 0;
2567 				/* Exhausted packet, just free. */
2568 				mbuf = loc->mbuf;
2569 				loc->mbuf = mbuf->next;
2570 				rte_pktmbuf_free_seg(mbuf);
2571 				loc->mbuf_off = 0;
2572 				assert(loc->mbuf_nseg >= 1);
2573 				--loc->mbuf_nseg;
2574 			}
2575 			return;
2576 		}
2577 		pdst += part;
2578 	} while (true);
2579 }
2580 
2581 /**
2582  * Build the Ethernet Segment with inlined data from
2583  * multi-segment packet. Checks the boundary of WQEBB
2584  * and ring buffer wrapping, supports Software Parser,
2585  * Checksums and VLAN insertion Tx offload features.
2586  *
2587  * @param txq
2588  *   Pointer to TX queue structure.
2589  * @param loc
2590  *   Pointer to burst routine local context.
2591  * @param wqe
2592  *   Pointer to WQE to fill with built Ethernet Segment.
2593  * @param vlan
2594  *   Length of VLAN tag insertion if any.
2595  * @param inlen
2596  *   Length of data to inline (VLAN included, if any).
2597  * @param tso
2598  *   TSO flag, set mss field from the packet.
2599  * @param olx
2600  *   Configured Tx offloads mask. It is fully defined at
2601  *   compile time and may be used for optimization.
2602  *
2603  * @return
2604  *   Pointer to the next Data Segment (aligned and
2605  *   possible NOT wrapped around - caller should do
2606  *   wrapping check on its own).
2607  */
2608 static __rte_always_inline struct mlx5_wqe_dseg *
2609 mlx5_tx_eseg_mdat(struct mlx5_txq_data *restrict txq,
2610 		  struct mlx5_txq_local *restrict loc,
2611 		  struct mlx5_wqe *restrict wqe,
2612 		  unsigned int vlan,
2613 		  unsigned int inlen,
2614 		  unsigned int tso,
2615 		  unsigned int olx)
2616 {
2617 	struct mlx5_wqe_eseg *restrict es = &wqe->eseg;
2618 	uint32_t csum;
2619 	uint8_t *pdst;
2620 	unsigned int part;
2621 
2622 	/*
2623 	 * Calculate and set check sum flags first, uint32_t field
2624 	 * in segment may be shared with Software Parser flags.
2625 	 */
2626 	csum = MLX5_TXOFF_CONFIG(CSUM) ? txq_ol_cksum_to_cs(loc->mbuf) : 0;
2627 	if (tso) {
2628 		csum <<= 24;
2629 		csum |= loc->mbuf->tso_segsz;
2630 		es->flags = rte_cpu_to_be_32(csum);
2631 	} else {
2632 		es->flags = rte_cpu_to_le_32(csum);
2633 	}
2634 	/*
2635 	 * Calculate and set Software Parser offsets and flags.
2636 	 * These flags a set for custom UDP and IP tunnel packets.
2637 	 */
2638 	es->swp_offs = txq_mbuf_to_swp(loc, &es->swp_flags, olx);
2639 	/* Fill metadata field if needed. */
2640 	es->metadata = MLX5_TXOFF_CONFIG(METADATA) ?
2641 		       loc->mbuf->ol_flags & PKT_TX_DYNF_METADATA ?
2642 		       *RTE_FLOW_DYNF_METADATA(loc->mbuf) : 0 : 0;
2643 	static_assert(MLX5_ESEG_MIN_INLINE_SIZE ==
2644 				(sizeof(uint16_t) +
2645 				 sizeof(rte_v128u32_t)),
2646 		      "invalid Ethernet Segment data size");
2647 	static_assert(MLX5_ESEG_MIN_INLINE_SIZE ==
2648 				(sizeof(uint16_t) +
2649 				 sizeof(struct rte_vlan_hdr) +
2650 				 2 * RTE_ETHER_ADDR_LEN),
2651 		      "invalid Ethernet Segment data size");
2652 	assert(inlen >= MLX5_ESEG_MIN_INLINE_SIZE);
2653 	es->inline_hdr_sz = rte_cpu_to_be_16(inlen);
2654 	pdst = (uint8_t *)&es->inline_data;
2655 	if (MLX5_TXOFF_CONFIG(VLAN) && vlan) {
2656 		/* Implement VLAN tag insertion as part inline data. */
2657 		mlx5_tx_mseg_memcpy(pdst, loc, 2 * RTE_ETHER_ADDR_LEN, olx);
2658 		pdst += 2 * RTE_ETHER_ADDR_LEN;
2659 		*(unaligned_uint32_t *)pdst = rte_cpu_to_be_32
2660 						((RTE_ETHER_TYPE_VLAN << 16) |
2661 						 loc->mbuf->vlan_tci);
2662 		pdst += sizeof(struct rte_vlan_hdr);
2663 		inlen -= 2 * RTE_ETHER_ADDR_LEN + sizeof(struct rte_vlan_hdr);
2664 	}
2665 	assert(pdst < (uint8_t *)txq->wqes_end);
2666 	/*
2667 	 * The WQEBB space availability is checked by caller.
2668 	 * Here we should be aware of WQE ring buffer wraparound only.
2669 	 */
2670 	part = (uint8_t *)txq->wqes_end - pdst;
2671 	part = RTE_MIN(part, inlen);
2672 	assert(part);
2673 	do {
2674 		mlx5_tx_mseg_memcpy(pdst, loc, part, olx);
2675 		inlen -= part;
2676 		if (likely(!inlen)) {
2677 			pdst += part;
2678 			pdst = RTE_PTR_ALIGN(pdst, MLX5_WSEG_SIZE);
2679 			return (struct mlx5_wqe_dseg *)pdst;
2680 		}
2681 		pdst = (uint8_t *)txq->wqes;
2682 		part = inlen;
2683 	} while (true);
2684 }
2685 
2686 /**
2687  * Build the Data Segment of pointer type.
2688  *
2689  * @param txq
2690  *   Pointer to TX queue structure.
2691  * @param loc
2692  *   Pointer to burst routine local context.
2693  * @param dseg
2694  *   Pointer to WQE to fill with built Data Segment.
2695  * @param buf
2696  *   Data buffer to point.
2697  * @param len
2698  *   Data buffer length.
2699  * @param olx
2700  *   Configured Tx offloads mask. It is fully defined at
2701  *   compile time and may be used for optimization.
2702  */
2703 static __rte_always_inline void
2704 mlx5_tx_dseg_ptr(struct mlx5_txq_data *restrict txq,
2705 		 struct mlx5_txq_local *restrict loc,
2706 		 struct mlx5_wqe_dseg *restrict dseg,
2707 		 uint8_t *buf,
2708 		 unsigned int len,
2709 		 unsigned int olx __rte_unused)
2710 
2711 {
2712 	assert(len);
2713 	dseg->bcount = rte_cpu_to_be_32(len);
2714 	dseg->lkey = mlx5_tx_mb2mr(txq, loc->mbuf);
2715 	dseg->pbuf = rte_cpu_to_be_64((uintptr_t)buf);
2716 }
2717 
2718 /**
2719  * Build the Data Segment of pointer type or inline
2720  * if data length is less than buffer in minimal
2721  * Data Segment size.
2722  *
2723  * @param txq
2724  *   Pointer to TX queue structure.
2725  * @param loc
2726  *   Pointer to burst routine local context.
2727  * @param dseg
2728  *   Pointer to WQE to fill with built Data Segment.
2729  * @param buf
2730  *   Data buffer to point.
2731  * @param len
2732  *   Data buffer length.
2733  * @param olx
2734  *   Configured Tx offloads mask. It is fully defined at
2735  *   compile time and may be used for optimization.
2736  */
2737 static __rte_always_inline void
2738 mlx5_tx_dseg_iptr(struct mlx5_txq_data *restrict txq,
2739 		  struct mlx5_txq_local *restrict loc,
2740 		  struct mlx5_wqe_dseg *restrict dseg,
2741 		  uint8_t *buf,
2742 		  unsigned int len,
2743 		  unsigned int olx __rte_unused)
2744 
2745 {
2746 	uintptr_t dst, src;
2747 
2748 	assert(len);
2749 	if (len > MLX5_DSEG_MIN_INLINE_SIZE) {
2750 		dseg->bcount = rte_cpu_to_be_32(len);
2751 		dseg->lkey = mlx5_tx_mb2mr(txq, loc->mbuf);
2752 		dseg->pbuf = rte_cpu_to_be_64((uintptr_t)buf);
2753 
2754 		return;
2755 	}
2756 	dseg->bcount = rte_cpu_to_be_32(len | MLX5_ETH_WQE_DATA_INLINE);
2757 	/* Unrolled implementation of generic rte_memcpy. */
2758 	dst = (uintptr_t)&dseg->inline_data[0];
2759 	src = (uintptr_t)buf;
2760 	if (len & 0x08) {
2761 #ifdef RTE_ARCH_STRICT_ALIGN
2762 		assert(dst == RTE_PTR_ALIGN(dst, sizeof(uint32_t)));
2763 		*(uint32_t *)dst = *(unaligned_uint32_t *)src;
2764 		dst += sizeof(uint32_t);
2765 		src += sizeof(uint32_t);
2766 		*(uint32_t *)dst = *(unaligned_uint32_t *)src;
2767 		dst += sizeof(uint32_t);
2768 		src += sizeof(uint32_t);
2769 #else
2770 		*(uint64_t *)dst = *(unaligned_uint64_t *)src;
2771 		dst += sizeof(uint64_t);
2772 		src += sizeof(uint64_t);
2773 #endif
2774 	}
2775 	if (len & 0x04) {
2776 		*(uint32_t *)dst = *(unaligned_uint32_t *)src;
2777 		dst += sizeof(uint32_t);
2778 		src += sizeof(uint32_t);
2779 	}
2780 	if (len & 0x02) {
2781 		*(uint16_t *)dst = *(unaligned_uint16_t *)src;
2782 		dst += sizeof(uint16_t);
2783 		src += sizeof(uint16_t);
2784 	}
2785 	if (len & 0x01)
2786 		*(uint8_t *)dst = *(uint8_t *)src;
2787 }
2788 
2789 /**
2790  * Build the Data Segment of inlined data from single
2791  * segment packet, no VLAN insertion.
2792  *
2793  * @param txq
2794  *   Pointer to TX queue structure.
2795  * @param loc
2796  *   Pointer to burst routine local context.
2797  * @param dseg
2798  *   Pointer to WQE to fill with built Data Segment.
2799  * @param buf
2800  *   Data buffer to point.
2801  * @param len
2802  *   Data buffer length.
2803  * @param olx
2804  *   Configured Tx offloads mask. It is fully defined at
2805  *   compile time and may be used for optimization.
2806  *
2807  * @return
2808  *   Pointer to the next Data Segment after inlined data.
2809  *   Ring buffer wraparound check is needed. We do not
2810  *   do it here because it may not be needed for the
2811  *   last packet in the eMPW session.
2812  */
2813 static __rte_always_inline struct mlx5_wqe_dseg *
2814 mlx5_tx_dseg_empw(struct mlx5_txq_data *restrict txq,
2815 		  struct mlx5_txq_local *restrict loc __rte_unused,
2816 		  struct mlx5_wqe_dseg *restrict dseg,
2817 		  uint8_t *buf,
2818 		  unsigned int len,
2819 		  unsigned int olx __rte_unused)
2820 {
2821 	unsigned int part;
2822 	uint8_t *pdst;
2823 
2824 	dseg->bcount = rte_cpu_to_be_32(len | MLX5_ETH_WQE_DATA_INLINE);
2825 	pdst = &dseg->inline_data[0];
2826 	/*
2827 	 * The WQEBB space availability is checked by caller.
2828 	 * Here we should be aware of WQE ring buffer wraparound only.
2829 	 */
2830 	part = (uint8_t *)txq->wqes_end - pdst;
2831 	part = RTE_MIN(part, len);
2832 	do {
2833 		rte_memcpy(pdst, buf, part);
2834 		len -= part;
2835 		if (likely(!len)) {
2836 			pdst += part;
2837 			pdst = RTE_PTR_ALIGN(pdst, MLX5_WSEG_SIZE);
2838 			/* Note: no final wraparound check here. */
2839 			return (struct mlx5_wqe_dseg *)pdst;
2840 		}
2841 		pdst = (uint8_t *)txq->wqes;
2842 		buf += part;
2843 		part = len;
2844 	} while (true);
2845 }
2846 
2847 /**
2848  * Build the Data Segment of inlined data from single
2849  * segment packet with VLAN insertion.
2850  *
2851  * @param txq
2852  *   Pointer to TX queue structure.
2853  * @param loc
2854  *   Pointer to burst routine local context.
2855  * @param dseg
2856  *   Pointer to the dseg fill with built Data Segment.
2857  * @param buf
2858  *   Data buffer to point.
2859  * @param len
2860  *   Data buffer length.
2861  * @param olx
2862  *   Configured Tx offloads mask. It is fully defined at
2863  *   compile time and may be used for optimization.
2864  *
2865  * @return
2866  *   Pointer to the next Data Segment after inlined data.
2867  *   Ring buffer wraparound check is needed.
2868  */
2869 static __rte_always_inline struct mlx5_wqe_dseg *
2870 mlx5_tx_dseg_vlan(struct mlx5_txq_data *restrict txq,
2871 		  struct mlx5_txq_local *restrict loc __rte_unused,
2872 		  struct mlx5_wqe_dseg *restrict dseg,
2873 		  uint8_t *buf,
2874 		  unsigned int len,
2875 		  unsigned int olx __rte_unused)
2876 
2877 {
2878 	unsigned int part;
2879 	uint8_t *pdst;
2880 
2881 	assert(len > MLX5_ESEG_MIN_INLINE_SIZE);
2882 	static_assert(MLX5_DSEG_MIN_INLINE_SIZE ==
2883 				 (2 * RTE_ETHER_ADDR_LEN),
2884 		      "invalid Data Segment data size");
2885 	dseg->bcount = rte_cpu_to_be_32((len + sizeof(struct rte_vlan_hdr)) |
2886 					MLX5_ETH_WQE_DATA_INLINE);
2887 	pdst = &dseg->inline_data[0];
2888 	memcpy(pdst, buf, MLX5_DSEG_MIN_INLINE_SIZE);
2889 	buf += MLX5_DSEG_MIN_INLINE_SIZE;
2890 	pdst += MLX5_DSEG_MIN_INLINE_SIZE;
2891 	len -= MLX5_DSEG_MIN_INLINE_SIZE;
2892 	/* Insert VLAN ethertype + VLAN tag. Pointer is aligned. */
2893 	assert(pdst == RTE_PTR_ALIGN(pdst, MLX5_WSEG_SIZE));
2894 	if (unlikely(pdst >= (uint8_t *)txq->wqes_end))
2895 		pdst = (uint8_t *)txq->wqes;
2896 	*(uint32_t *)pdst = rte_cpu_to_be_32((RTE_ETHER_TYPE_VLAN << 16) |
2897 					      loc->mbuf->vlan_tci);
2898 	pdst += sizeof(struct rte_vlan_hdr);
2899 	/*
2900 	 * The WQEBB space availability is checked by caller.
2901 	 * Here we should be aware of WQE ring buffer wraparound only.
2902 	 */
2903 	part = (uint8_t *)txq->wqes_end - pdst;
2904 	part = RTE_MIN(part, len);
2905 	do {
2906 		rte_memcpy(pdst, buf, part);
2907 		len -= part;
2908 		if (likely(!len)) {
2909 			pdst += part;
2910 			pdst = RTE_PTR_ALIGN(pdst, MLX5_WSEG_SIZE);
2911 			/* Note: no final wraparound check here. */
2912 			return (struct mlx5_wqe_dseg *)pdst;
2913 		}
2914 		pdst = (uint8_t *)txq->wqes;
2915 		buf += part;
2916 		part = len;
2917 	} while (true);
2918 }
2919 
2920 /**
2921  * Build the Ethernet Segment with optionally inlined data with
2922  * VLAN insertion and following Data Segments (if any) from
2923  * multi-segment packet. Used by ordinary send and TSO.
2924  *
2925  * @param txq
2926  *   Pointer to TX queue structure.
2927  * @param loc
2928  *   Pointer to burst routine local context.
2929  * @param wqe
2930  *   Pointer to WQE to fill with built Ethernet/Data Segments.
2931  * @param vlan
2932  *   Length of VLAN header to insert, 0 means no VLAN insertion.
2933  * @param inlen
2934  *   Data length to inline. For TSO this parameter specifies
2935  *   exact value, for ordinary send routine can be aligned by
2936  *   caller to provide better WQE space saving and data buffer
2937  *   start address alignment. This length includes VLAN header
2938  *   being inserted.
2939  * @param tso
2940  *   Zero means ordinary send, inlined data can be extended,
2941  *   otherwise this is TSO, inlined data length is fixed.
2942  * @param olx
2943  *   Configured Tx offloads mask. It is fully defined at
2944  *   compile time and may be used for optimization.
2945  *
2946  * @return
2947  *   Actual size of built WQE in segments.
2948  */
2949 static __rte_always_inline unsigned int
2950 mlx5_tx_mseg_build(struct mlx5_txq_data *restrict txq,
2951 		   struct mlx5_txq_local *restrict loc,
2952 		   struct mlx5_wqe *restrict wqe,
2953 		   unsigned int vlan,
2954 		   unsigned int inlen,
2955 		   unsigned int tso,
2956 		   unsigned int olx __rte_unused)
2957 {
2958 	struct mlx5_wqe_dseg *restrict dseg;
2959 	unsigned int ds;
2960 
2961 	assert((rte_pktmbuf_pkt_len(loc->mbuf) + vlan) >= inlen);
2962 	loc->mbuf_nseg = NB_SEGS(loc->mbuf);
2963 	loc->mbuf_off = 0;
2964 
2965 	dseg = mlx5_tx_eseg_mdat(txq, loc, wqe, vlan, inlen, tso, olx);
2966 	if (!loc->mbuf_nseg)
2967 		goto dseg_done;
2968 	/*
2969 	 * There are still some mbuf remaining, not inlined.
2970 	 * The first mbuf may be partially inlined and we
2971 	 * must process the possible non-zero data offset.
2972 	 */
2973 	if (loc->mbuf_off) {
2974 		unsigned int dlen;
2975 		uint8_t *dptr;
2976 
2977 		/*
2978 		 * Exhausted packets must be dropped before.
2979 		 * Non-zero offset means there are some data
2980 		 * remained in the packet.
2981 		 */
2982 		assert(loc->mbuf_off < rte_pktmbuf_data_len(loc->mbuf));
2983 		assert(rte_pktmbuf_data_len(loc->mbuf));
2984 		dptr = rte_pktmbuf_mtod_offset(loc->mbuf, uint8_t *,
2985 					       loc->mbuf_off);
2986 		dlen = rte_pktmbuf_data_len(loc->mbuf) - loc->mbuf_off;
2987 		/*
2988 		 * Build the pointer/minimal data Data Segment.
2989 		 * Do ring buffer wrapping check in advance.
2990 		 */
2991 		if ((uintptr_t)dseg >= (uintptr_t)txq->wqes_end)
2992 			dseg = (struct mlx5_wqe_dseg *)txq->wqes;
2993 		mlx5_tx_dseg_iptr(txq, loc, dseg, dptr, dlen, olx);
2994 		/* Store the mbuf to be freed on completion. */
2995 		assert(loc->elts_free);
2996 		txq->elts[txq->elts_head++ & txq->elts_m] = loc->mbuf;
2997 		--loc->elts_free;
2998 		++dseg;
2999 		if (--loc->mbuf_nseg == 0)
3000 			goto dseg_done;
3001 		loc->mbuf = loc->mbuf->next;
3002 		loc->mbuf_off = 0;
3003 	}
3004 	do {
3005 		if (unlikely(!rte_pktmbuf_data_len(loc->mbuf))) {
3006 			struct rte_mbuf *mbuf;
3007 
3008 			/* Zero length segment found, just skip. */
3009 			mbuf = loc->mbuf;
3010 			loc->mbuf = loc->mbuf->next;
3011 			rte_pktmbuf_free_seg(mbuf);
3012 			if (--loc->mbuf_nseg == 0)
3013 				break;
3014 		} else {
3015 			if ((uintptr_t)dseg >= (uintptr_t)txq->wqes_end)
3016 				dseg = (struct mlx5_wqe_dseg *)txq->wqes;
3017 			mlx5_tx_dseg_iptr
3018 				(txq, loc, dseg,
3019 				 rte_pktmbuf_mtod(loc->mbuf, uint8_t *),
3020 				 rte_pktmbuf_data_len(loc->mbuf), olx);
3021 			assert(loc->elts_free);
3022 			txq->elts[txq->elts_head++ & txq->elts_m] = loc->mbuf;
3023 			--loc->elts_free;
3024 			++dseg;
3025 			if (--loc->mbuf_nseg == 0)
3026 				break;
3027 			loc->mbuf = loc->mbuf->next;
3028 		}
3029 	} while (true);
3030 
3031 dseg_done:
3032 	/* Calculate actual segments used from the dseg pointer. */
3033 	if ((uintptr_t)wqe < (uintptr_t)dseg)
3034 		ds = ((uintptr_t)dseg - (uintptr_t)wqe) / MLX5_WSEG_SIZE;
3035 	else
3036 		ds = (((uintptr_t)dseg - (uintptr_t)wqe) +
3037 		      txq->wqe_s * MLX5_WQE_SIZE) / MLX5_WSEG_SIZE;
3038 	return ds;
3039 }
3040 
3041 /**
3042  * Tx one packet function for multi-segment TSO. Supports all
3043  * types of Tx offloads, uses MLX5_OPCODE_TSO to build WQEs,
3044  * sends one packet per WQE.
3045  *
3046  * This routine is responsible for storing processed mbuf
3047  * into elts ring buffer and update elts_head.
3048  *
3049  * @param txq
3050  *   Pointer to TX queue structure.
3051  * @param loc
3052  *   Pointer to burst routine local context.
3053  * @param olx
3054  *   Configured Tx offloads mask. It is fully defined at
3055  *   compile time and may be used for optimization.
3056  *
3057  * @return
3058  *   MLX5_TXCMP_CODE_EXIT - sending is done or impossible.
3059  *   MLX5_TXCMP_CODE_ERROR - some unrecoverable error occurred.
3060  * Local context variables partially updated.
3061  */
3062 static __rte_always_inline enum mlx5_txcmp_code
3063 mlx5_tx_packet_multi_tso(struct mlx5_txq_data *restrict txq,
3064 			struct mlx5_txq_local *restrict loc,
3065 			unsigned int olx)
3066 {
3067 	struct mlx5_wqe *restrict wqe;
3068 	unsigned int ds, dlen, inlen, ntcp, vlan = 0;
3069 
3070 	/*
3071 	 * Calculate data length to be inlined to estimate
3072 	 * the required space in WQE ring buffer.
3073 	 */
3074 	dlen = rte_pktmbuf_pkt_len(loc->mbuf);
3075 	if (MLX5_TXOFF_CONFIG(VLAN) && loc->mbuf->ol_flags & PKT_TX_VLAN_PKT)
3076 		vlan = sizeof(struct rte_vlan_hdr);
3077 	inlen = loc->mbuf->l2_len + vlan +
3078 		loc->mbuf->l3_len + loc->mbuf->l4_len;
3079 	if (unlikely((!inlen || !loc->mbuf->tso_segsz)))
3080 		return MLX5_TXCMP_CODE_ERROR;
3081 	if (loc->mbuf->ol_flags & PKT_TX_TUNNEL_MASK)
3082 		inlen += loc->mbuf->outer_l2_len + loc->mbuf->outer_l3_len;
3083 	/* Packet must contain all TSO headers. */
3084 	if (unlikely(inlen > MLX5_MAX_TSO_HEADER ||
3085 		     inlen <= MLX5_ESEG_MIN_INLINE_SIZE ||
3086 		     inlen > (dlen + vlan)))
3087 		return MLX5_TXCMP_CODE_ERROR;
3088 	assert(inlen >= txq->inlen_mode);
3089 	/*
3090 	 * Check whether there are enough free WQEBBs:
3091 	 * - Control Segment
3092 	 * - Ethernet Segment
3093 	 * - First Segment of inlined Ethernet data
3094 	 * - ... data continued ...
3095 	 * - Data Segments of pointer/min inline type
3096 	 */
3097 	ds = NB_SEGS(loc->mbuf) + 2 + (inlen -
3098 				       MLX5_ESEG_MIN_INLINE_SIZE +
3099 				       MLX5_WSEG_SIZE +
3100 				       MLX5_WSEG_SIZE - 1) / MLX5_WSEG_SIZE;
3101 	if (unlikely(loc->wqe_free < ((ds + 3) / 4)))
3102 		return MLX5_TXCMP_CODE_EXIT;
3103 	/* Check for maximal WQE size. */
3104 	if (unlikely((MLX5_WQE_SIZE_MAX / MLX5_WSEG_SIZE) < ((ds + 3) / 4)))
3105 		return MLX5_TXCMP_CODE_ERROR;
3106 #ifdef MLX5_PMD_SOFT_COUNTERS
3107 	/* Update sent data bytes/packets counters. */
3108 	ntcp = (dlen - (inlen - vlan) + loc->mbuf->tso_segsz - 1) /
3109 		loc->mbuf->tso_segsz;
3110 	/*
3111 	 * One will be added for mbuf itself
3112 	 * at the end of the mlx5_tx_burst from
3113 	 * loc->pkts_sent field.
3114 	 */
3115 	--ntcp;
3116 	txq->stats.opackets += ntcp;
3117 	txq->stats.obytes += dlen + vlan + ntcp * inlen;
3118 #endif
3119 	wqe = txq->wqes + (txq->wqe_ci & txq->wqe_m);
3120 	loc->wqe_last = wqe;
3121 	mlx5_tx_cseg_init(txq, loc, wqe, 0, MLX5_OPCODE_TSO, olx);
3122 	ds = mlx5_tx_mseg_build(txq, loc, wqe, vlan, inlen, 1, olx);
3123 	wqe->cseg.sq_ds = rte_cpu_to_be_32(txq->qp_num_8s | ds);
3124 	txq->wqe_ci += (ds + 3) / 4;
3125 	loc->wqe_free -= (ds + 3) / 4;
3126 	return MLX5_TXCMP_CODE_MULTI;
3127 }
3128 
3129 /**
3130  * Tx one packet function for multi-segment SEND. Supports all
3131  * types of Tx offloads, uses MLX5_OPCODE_SEND to build WQEs,
3132  * sends one packet per WQE, without any data inlining in
3133  * Ethernet Segment.
3134  *
3135  * This routine is responsible for storing processed mbuf
3136  * into elts ring buffer and update elts_head.
3137  *
3138  * @param txq
3139  *   Pointer to TX queue structure.
3140  * @param loc
3141  *   Pointer to burst routine local context.
3142  * @param olx
3143  *   Configured Tx offloads mask. It is fully defined at
3144  *   compile time and may be used for optimization.
3145  *
3146  * @return
3147  *   MLX5_TXCMP_CODE_EXIT - sending is done or impossible.
3148  *   MLX5_TXCMP_CODE_ERROR - some unrecoverable error occurred.
3149  * Local context variables partially updated.
3150  */
3151 static __rte_always_inline enum mlx5_txcmp_code
3152 mlx5_tx_packet_multi_send(struct mlx5_txq_data *restrict txq,
3153 			  struct mlx5_txq_local *restrict loc,
3154 			  unsigned int olx)
3155 {
3156 	struct mlx5_wqe_dseg *restrict dseg;
3157 	struct mlx5_wqe *restrict wqe;
3158 	unsigned int ds, nseg;
3159 
3160 	assert(NB_SEGS(loc->mbuf) > 1);
3161 	/*
3162 	 * No inline at all, it means the CPU cycles saving
3163 	 * is prioritized at configuration, we should not
3164 	 * copy any packet data to WQE.
3165 	 */
3166 	nseg = NB_SEGS(loc->mbuf);
3167 	ds = 2 + nseg;
3168 	if (unlikely(loc->wqe_free < ((ds + 3) / 4)))
3169 		return MLX5_TXCMP_CODE_EXIT;
3170 	/* Check for maximal WQE size. */
3171 	if (unlikely((MLX5_WQE_SIZE_MAX / MLX5_WSEG_SIZE) < ((ds + 3) / 4)))
3172 		return MLX5_TXCMP_CODE_ERROR;
3173 	/*
3174 	 * Some Tx offloads may cause an error if
3175 	 * packet is not long enough, check against
3176 	 * assumed minimal length.
3177 	 */
3178 	if (rte_pktmbuf_pkt_len(loc->mbuf) <= MLX5_ESEG_MIN_INLINE_SIZE)
3179 		return MLX5_TXCMP_CODE_ERROR;
3180 #ifdef MLX5_PMD_SOFT_COUNTERS
3181 	/* Update sent data bytes counter. */
3182 	txq->stats.obytes += rte_pktmbuf_pkt_len(loc->mbuf);
3183 	if (MLX5_TXOFF_CONFIG(VLAN) &&
3184 	    loc->mbuf->ol_flags & PKT_TX_VLAN_PKT)
3185 		txq->stats.obytes += sizeof(struct rte_vlan_hdr);
3186 #endif
3187 	/*
3188 	 * SEND WQE, one WQEBB:
3189 	 * - Control Segment, SEND opcode
3190 	 * - Ethernet Segment, optional VLAN, no inline
3191 	 * - Data Segments, pointer only type
3192 	 */
3193 	wqe = txq->wqes + (txq->wqe_ci & txq->wqe_m);
3194 	loc->wqe_last = wqe;
3195 	mlx5_tx_cseg_init(txq, loc, wqe, ds, MLX5_OPCODE_SEND, olx);
3196 	mlx5_tx_eseg_none(txq, loc, wqe, olx);
3197 	dseg = &wqe->dseg[0];
3198 	do {
3199 		if (unlikely(!rte_pktmbuf_data_len(loc->mbuf))) {
3200 			struct rte_mbuf *mbuf;
3201 
3202 			/*
3203 			 * Zero length segment found, have to
3204 			 * correct total size of WQE in segments.
3205 			 * It is supposed to be rare occasion, so
3206 			 * in normal case (no zero length segments)
3207 			 * we avoid extra writing to the Control
3208 			 * Segment.
3209 			 */
3210 			--ds;
3211 			wqe->cseg.sq_ds -= RTE_BE32(1);
3212 			mbuf = loc->mbuf;
3213 			loc->mbuf = mbuf->next;
3214 			rte_pktmbuf_free_seg(mbuf);
3215 			if (--nseg == 0)
3216 				break;
3217 		} else {
3218 			mlx5_tx_dseg_ptr
3219 				(txq, loc, dseg,
3220 				 rte_pktmbuf_mtod(loc->mbuf, uint8_t *),
3221 				 rte_pktmbuf_data_len(loc->mbuf), olx);
3222 			txq->elts[txq->elts_head++ & txq->elts_m] = loc->mbuf;
3223 			--loc->elts_free;
3224 			if (--nseg == 0)
3225 				break;
3226 			++dseg;
3227 			if ((uintptr_t)dseg >= (uintptr_t)txq->wqes_end)
3228 				dseg = (struct mlx5_wqe_dseg *)txq->wqes;
3229 			loc->mbuf = loc->mbuf->next;
3230 		}
3231 	} while (true);
3232 	txq->wqe_ci += (ds + 3) / 4;
3233 	loc->wqe_free -= (ds + 3) / 4;
3234 	return MLX5_TXCMP_CODE_MULTI;
3235 }
3236 
3237 /**
3238  * Tx one packet function for multi-segment SEND. Supports all
3239  * types of Tx offloads, uses MLX5_OPCODE_SEND to build WQEs,
3240  * sends one packet per WQE, with data inlining in
3241  * Ethernet Segment and minimal Data Segments.
3242  *
3243  * This routine is responsible for storing processed mbuf
3244  * into elts ring buffer and update elts_head.
3245  *
3246  * @param txq
3247  *   Pointer to TX queue structure.
3248  * @param loc
3249  *   Pointer to burst routine local context.
3250  * @param olx
3251  *   Configured Tx offloads mask. It is fully defined at
3252  *   compile time and may be used for optimization.
3253  *
3254  * @return
3255  *   MLX5_TXCMP_CODE_EXIT - sending is done or impossible.
3256  *   MLX5_TXCMP_CODE_ERROR - some unrecoverable error occurred.
3257  * Local context variables partially updated.
3258  */
3259 static __rte_always_inline enum mlx5_txcmp_code
3260 mlx5_tx_packet_multi_inline(struct mlx5_txq_data *restrict txq,
3261 			    struct mlx5_txq_local *restrict loc,
3262 			    unsigned int olx)
3263 {
3264 	struct mlx5_wqe *restrict wqe;
3265 	unsigned int ds, inlen, dlen, vlan = 0;
3266 
3267 	assert(MLX5_TXOFF_CONFIG(INLINE));
3268 	assert(NB_SEGS(loc->mbuf) > 1);
3269 	/*
3270 	 * First calculate data length to be inlined
3271 	 * to estimate the required space for WQE.
3272 	 */
3273 	dlen = rte_pktmbuf_pkt_len(loc->mbuf);
3274 	if (MLX5_TXOFF_CONFIG(VLAN) && loc->mbuf->ol_flags & PKT_TX_VLAN_PKT)
3275 		vlan = sizeof(struct rte_vlan_hdr);
3276 	inlen = dlen + vlan;
3277 	/* Check against minimal length. */
3278 	if (inlen <= MLX5_ESEG_MIN_INLINE_SIZE)
3279 		return MLX5_TXCMP_CODE_ERROR;
3280 	assert(txq->inlen_send >= MLX5_ESEG_MIN_INLINE_SIZE);
3281 	if (inlen > txq->inlen_send) {
3282 		struct rte_mbuf *mbuf;
3283 		unsigned int nxlen;
3284 		uintptr_t start;
3285 
3286 		/*
3287 		 * Packet length exceeds the allowed inline
3288 		 * data length, check whether the minimal
3289 		 * inlining is required.
3290 		 */
3291 		if (txq->inlen_mode) {
3292 			assert(txq->inlen_mode >= MLX5_ESEG_MIN_INLINE_SIZE);
3293 			assert(txq->inlen_mode <= txq->inlen_send);
3294 			inlen = txq->inlen_mode;
3295 		} else {
3296 			if (!vlan || txq->vlan_en) {
3297 				/*
3298 				 * VLAN insertion will be done inside by HW.
3299 				 * It is not utmost effective - VLAN flag is
3300 				 * checked twice, but we should proceed the
3301 				 * inlining length correctly and take into
3302 				 * account the VLAN header being inserted.
3303 				 */
3304 				return mlx5_tx_packet_multi_send
3305 							(txq, loc, olx);
3306 			}
3307 			inlen = MLX5_ESEG_MIN_INLINE_SIZE;
3308 		}
3309 		/*
3310 		 * Now we know the minimal amount of data is requested
3311 		 * to inline. Check whether we should inline the buffers
3312 		 * from the chain beginning to eliminate some mbufs.
3313 		 */
3314 		mbuf = loc->mbuf;
3315 		nxlen = rte_pktmbuf_data_len(mbuf);
3316 		if (unlikely(nxlen <= txq->inlen_send)) {
3317 			/* We can inline first mbuf at least. */
3318 			if (nxlen < inlen) {
3319 				unsigned int smlen;
3320 
3321 				/* Scan mbufs till inlen filled. */
3322 				do {
3323 					smlen = nxlen;
3324 					mbuf = NEXT(mbuf);
3325 					assert(mbuf);
3326 					nxlen = rte_pktmbuf_data_len(mbuf);
3327 					nxlen += smlen;
3328 				} while (unlikely(nxlen < inlen));
3329 				if (unlikely(nxlen > txq->inlen_send)) {
3330 					/* We cannot inline entire mbuf. */
3331 					smlen = inlen - smlen;
3332 					start = rte_pktmbuf_mtod_offset
3333 						    (mbuf, uintptr_t, smlen);
3334 					goto do_align;
3335 				}
3336 			}
3337 			do {
3338 				inlen = nxlen;
3339 				mbuf = NEXT(mbuf);
3340 				/* There should be not end of packet. */
3341 				assert(mbuf);
3342 				nxlen = inlen + rte_pktmbuf_data_len(mbuf);
3343 			} while (unlikely(nxlen < txq->inlen_send));
3344 		}
3345 		start = rte_pktmbuf_mtod(mbuf, uintptr_t);
3346 		/*
3347 		 * Check whether we can do inline to align start
3348 		 * address of data buffer to cacheline.
3349 		 */
3350 do_align:
3351 		start = (~start + 1) & (RTE_CACHE_LINE_SIZE - 1);
3352 		if (unlikely(start)) {
3353 			start += inlen;
3354 			if (start <= txq->inlen_send)
3355 				inlen = start;
3356 		}
3357 	}
3358 	/*
3359 	 * Check whether there are enough free WQEBBs:
3360 	 * - Control Segment
3361 	 * - Ethernet Segment
3362 	 * - First Segment of inlined Ethernet data
3363 	 * - ... data continued ...
3364 	 * - Data Segments of pointer/min inline type
3365 	 *
3366 	 * Estimate the number of Data Segments conservatively,
3367 	 * supposing no any mbufs is being freed during inlining.
3368 	 */
3369 	assert(inlen <= txq->inlen_send);
3370 	ds = NB_SEGS(loc->mbuf) + 2 + (inlen -
3371 				       MLX5_ESEG_MIN_INLINE_SIZE +
3372 				       MLX5_WSEG_SIZE +
3373 				       MLX5_WSEG_SIZE - 1) / MLX5_WSEG_SIZE;
3374 	if (unlikely(loc->wqe_free < ((ds + 3) / 4)))
3375 		return MLX5_TXCMP_CODE_EXIT;
3376 	/* Check for maximal WQE size. */
3377 	if (unlikely((MLX5_WQE_SIZE_MAX / MLX5_WSEG_SIZE) < ((ds + 3) / 4)))
3378 		return MLX5_TXCMP_CODE_ERROR;
3379 #ifdef MLX5_PMD_SOFT_COUNTERS
3380 	/* Update sent data bytes/packets counters. */
3381 	txq->stats.obytes += dlen + vlan;
3382 #endif
3383 	wqe = txq->wqes + (txq->wqe_ci & txq->wqe_m);
3384 	loc->wqe_last = wqe;
3385 	mlx5_tx_cseg_init(txq, loc, wqe, 0, MLX5_OPCODE_SEND, olx);
3386 	ds = mlx5_tx_mseg_build(txq, loc, wqe, vlan, inlen, 0, olx);
3387 	wqe->cseg.sq_ds = rte_cpu_to_be_32(txq->qp_num_8s | ds);
3388 	txq->wqe_ci += (ds + 3) / 4;
3389 	loc->wqe_free -= (ds + 3) / 4;
3390 	return MLX5_TXCMP_CODE_MULTI;
3391 }
3392 
3393 /**
3394  * Tx burst function for multi-segment packets. Supports all
3395  * types of Tx offloads, uses MLX5_OPCODE_SEND/TSO to build WQEs,
3396  * sends one packet per WQE. Function stops sending if it
3397  * encounters the single-segment packet.
3398  *
3399  * This routine is responsible for storing processed mbuf
3400  * into elts ring buffer and update elts_head.
3401  *
3402  * @param txq
3403  *   Pointer to TX queue structure.
3404  * @param[in] pkts
3405  *   Packets to transmit.
3406  * @param pkts_n
3407  *   Number of packets in array.
3408  * @param loc
3409  *   Pointer to burst routine local context.
3410  * @param olx
3411  *   Configured Tx offloads mask. It is fully defined at
3412  *   compile time and may be used for optimization.
3413  *
3414  * @return
3415  *   MLX5_TXCMP_CODE_EXIT - sending is done or impossible.
3416  *   MLX5_TXCMP_CODE_ERROR - some unrecoverable error occurred.
3417  *   MLX5_TXCMP_CODE_SINGLE - single-segment packet encountered.
3418  *   MLX5_TXCMP_CODE_TSO - TSO single-segment packet encountered.
3419  * Local context variables updated.
3420  */
3421 static __rte_always_inline enum mlx5_txcmp_code
3422 mlx5_tx_burst_mseg(struct mlx5_txq_data *restrict txq,
3423 		   struct rte_mbuf **restrict pkts,
3424 		   unsigned int pkts_n,
3425 		   struct mlx5_txq_local *restrict loc,
3426 		   unsigned int olx)
3427 {
3428 	assert(loc->elts_free && loc->wqe_free);
3429 	assert(pkts_n > loc->pkts_sent);
3430 	pkts += loc->pkts_sent + 1;
3431 	pkts_n -= loc->pkts_sent;
3432 	for (;;) {
3433 		enum mlx5_txcmp_code ret;
3434 
3435 		assert(NB_SEGS(loc->mbuf) > 1);
3436 		/*
3437 		 * Estimate the number of free elts quickly but
3438 		 * conservatively. Some segment may be fully inlined
3439 		 * and freed, ignore this here - precise estimation
3440 		 * is costly.
3441 		 */
3442 		if (loc->elts_free < NB_SEGS(loc->mbuf))
3443 			return MLX5_TXCMP_CODE_EXIT;
3444 		if (MLX5_TXOFF_CONFIG(TSO) &&
3445 		    unlikely(loc->mbuf->ol_flags & PKT_TX_TCP_SEG)) {
3446 			/* Proceed with multi-segment TSO. */
3447 			ret = mlx5_tx_packet_multi_tso(txq, loc, olx);
3448 		} else if (MLX5_TXOFF_CONFIG(INLINE)) {
3449 			/* Proceed with multi-segment SEND with inlining. */
3450 			ret = mlx5_tx_packet_multi_inline(txq, loc, olx);
3451 		} else {
3452 			/* Proceed with multi-segment SEND w/o inlining. */
3453 			ret = mlx5_tx_packet_multi_send(txq, loc, olx);
3454 		}
3455 		if (ret == MLX5_TXCMP_CODE_EXIT)
3456 			return MLX5_TXCMP_CODE_EXIT;
3457 		if (ret == MLX5_TXCMP_CODE_ERROR)
3458 			return MLX5_TXCMP_CODE_ERROR;
3459 		/* WQE is built, go to the next packet. */
3460 		++loc->pkts_sent;
3461 		--pkts_n;
3462 		if (unlikely(!pkts_n || !loc->elts_free || !loc->wqe_free))
3463 			return MLX5_TXCMP_CODE_EXIT;
3464 		loc->mbuf = *pkts++;
3465 		if (pkts_n > 1)
3466 			rte_prefetch0(*pkts);
3467 		if (likely(NB_SEGS(loc->mbuf) > 1))
3468 			continue;
3469 		/* Here ends the series of multi-segment packets. */
3470 		if (MLX5_TXOFF_CONFIG(TSO) &&
3471 		    unlikely(loc->mbuf->ol_flags & PKT_TX_TCP_SEG))
3472 			return MLX5_TXCMP_CODE_TSO;
3473 		return MLX5_TXCMP_CODE_SINGLE;
3474 	}
3475 	assert(false);
3476 }
3477 
3478 /**
3479  * Tx burst function for single-segment packets with TSO.
3480  * Supports all types of Tx offloads, except multi-packets.
3481  * Uses MLX5_OPCODE_TSO to build WQEs, sends one packet per WQE.
3482  * Function stops sending if it encounters the multi-segment
3483  * packet or packet without TSO requested.
3484  *
3485  * The routine is responsible for storing processed mbuf
3486  * into elts ring buffer and update elts_head if inline
3487  * offloads is requested due to possible early freeing
3488  * of the inlined mbufs (can not store pkts array in elts
3489  * as a batch).
3490  *
3491  * @param txq
3492  *   Pointer to TX queue structure.
3493  * @param[in] pkts
3494  *   Packets to transmit.
3495  * @param pkts_n
3496  *   Number of packets in array.
3497  * @param loc
3498  *   Pointer to burst routine local context.
3499  * @param olx
3500  *   Configured Tx offloads mask. It is fully defined at
3501  *   compile time and may be used for optimization.
3502  *
3503  * @return
3504  *   MLX5_TXCMP_CODE_EXIT - sending is done or impossible.
3505  *   MLX5_TXCMP_CODE_ERROR - some unrecoverable error occurred.
3506  *   MLX5_TXCMP_CODE_SINGLE - single-segment packet encountered.
3507  *   MLX5_TXCMP_CODE_MULTI - multi-segment packet encountered.
3508  * Local context variables updated.
3509  */
3510 static __rte_always_inline enum mlx5_txcmp_code
3511 mlx5_tx_burst_tso(struct mlx5_txq_data *restrict txq,
3512 		  struct rte_mbuf **restrict pkts,
3513 		  unsigned int pkts_n,
3514 		  struct mlx5_txq_local *restrict loc,
3515 		  unsigned int olx)
3516 {
3517 	assert(loc->elts_free && loc->wqe_free);
3518 	assert(pkts_n > loc->pkts_sent);
3519 	pkts += loc->pkts_sent + 1;
3520 	pkts_n -= loc->pkts_sent;
3521 	for (;;) {
3522 		struct mlx5_wqe_dseg *restrict dseg;
3523 		struct mlx5_wqe *restrict wqe;
3524 		unsigned int ds, dlen, hlen, ntcp, vlan = 0;
3525 		uint8_t *dptr;
3526 
3527 		assert(NB_SEGS(loc->mbuf) == 1);
3528 		dlen = rte_pktmbuf_data_len(loc->mbuf);
3529 		if (MLX5_TXOFF_CONFIG(VLAN) &&
3530 		    loc->mbuf->ol_flags & PKT_TX_VLAN_PKT) {
3531 			vlan = sizeof(struct rte_vlan_hdr);
3532 		}
3533 		/*
3534 		 * First calculate the WQE size to check
3535 		 * whether we have enough space in ring buffer.
3536 		 */
3537 		hlen = loc->mbuf->l2_len + vlan +
3538 		       loc->mbuf->l3_len + loc->mbuf->l4_len;
3539 		if (unlikely((!hlen || !loc->mbuf->tso_segsz)))
3540 			return MLX5_TXCMP_CODE_ERROR;
3541 		if (loc->mbuf->ol_flags & PKT_TX_TUNNEL_MASK)
3542 			hlen += loc->mbuf->outer_l2_len +
3543 				loc->mbuf->outer_l3_len;
3544 		/* Segment must contain all TSO headers. */
3545 		if (unlikely(hlen > MLX5_MAX_TSO_HEADER ||
3546 			     hlen <= MLX5_ESEG_MIN_INLINE_SIZE ||
3547 			     hlen > (dlen + vlan)))
3548 			return MLX5_TXCMP_CODE_ERROR;
3549 		/*
3550 		 * Check whether there are enough free WQEBBs:
3551 		 * - Control Segment
3552 		 * - Ethernet Segment
3553 		 * - First Segment of inlined Ethernet data
3554 		 * - ... data continued ...
3555 		 * - Finishing Data Segment of pointer type
3556 		 */
3557 		ds = 4 + (hlen - MLX5_ESEG_MIN_INLINE_SIZE +
3558 			  MLX5_WSEG_SIZE - 1) / MLX5_WSEG_SIZE;
3559 		if (loc->wqe_free < ((ds + 3) / 4))
3560 			return MLX5_TXCMP_CODE_EXIT;
3561 #ifdef MLX5_PMD_SOFT_COUNTERS
3562 		/* Update sent data bytes/packets counters. */
3563 		ntcp = (dlen + vlan - hlen +
3564 			loc->mbuf->tso_segsz - 1) /
3565 			loc->mbuf->tso_segsz;
3566 		/*
3567 		 * One will be added for mbuf itself at the end
3568 		 * of the mlx5_tx_burst from loc->pkts_sent field.
3569 		 */
3570 		--ntcp;
3571 		txq->stats.opackets += ntcp;
3572 		txq->stats.obytes += dlen + vlan + ntcp * hlen;
3573 #endif
3574 		/*
3575 		 * Build the TSO WQE:
3576 		 * - Control Segment
3577 		 * - Ethernet Segment with hlen bytes inlined
3578 		 * - Data Segment of pointer type
3579 		 */
3580 		wqe = txq->wqes + (txq->wqe_ci & txq->wqe_m);
3581 		loc->wqe_last = wqe;
3582 		mlx5_tx_cseg_init(txq, loc, wqe, ds,
3583 				  MLX5_OPCODE_TSO, olx);
3584 		dseg = mlx5_tx_eseg_data(txq, loc, wqe, vlan, hlen, 1, olx);
3585 		dptr = rte_pktmbuf_mtod(loc->mbuf, uint8_t *) + hlen - vlan;
3586 		dlen -= hlen - vlan;
3587 		mlx5_tx_dseg_ptr(txq, loc, dseg, dptr, dlen, olx);
3588 		/*
3589 		 * WQE is built, update the loop parameters
3590 		 * and go to the next packet.
3591 		 */
3592 		txq->wqe_ci += (ds + 3) / 4;
3593 		loc->wqe_free -= (ds + 3) / 4;
3594 		if (MLX5_TXOFF_CONFIG(INLINE))
3595 			txq->elts[txq->elts_head++ & txq->elts_m] = loc->mbuf;
3596 		--loc->elts_free;
3597 		++loc->pkts_sent;
3598 		--pkts_n;
3599 		if (unlikely(!pkts_n || !loc->elts_free || !loc->wqe_free))
3600 			return MLX5_TXCMP_CODE_EXIT;
3601 		loc->mbuf = *pkts++;
3602 		if (pkts_n > 1)
3603 			rte_prefetch0(*pkts);
3604 		if (MLX5_TXOFF_CONFIG(MULTI) &&
3605 		    unlikely(NB_SEGS(loc->mbuf) > 1))
3606 			return MLX5_TXCMP_CODE_MULTI;
3607 		if (likely(!(loc->mbuf->ol_flags & PKT_TX_TCP_SEG)))
3608 			return MLX5_TXCMP_CODE_SINGLE;
3609 		/* Continue with the next TSO packet. */
3610 	}
3611 	assert(false);
3612 }
3613 
3614 /**
3615  * Analyze the packet and select the best method to send.
3616  *
3617  * @param txq
3618  *   Pointer to TX queue structure.
3619  * @param loc
3620  *   Pointer to burst routine local context.
3621  * @param olx
3622  *   Configured Tx offloads mask. It is fully defined at
3623  *   compile time and may be used for optimization.
3624  * @param newp
3625  *   The predefined flag whether do complete check for
3626  *   multi-segment packets and TSO.
3627  *
3628  * @return
3629  *  MLX5_TXCMP_CODE_MULTI - multi-segment packet encountered.
3630  *  MLX5_TXCMP_CODE_TSO - TSO required, use TSO/LSO.
3631  *  MLX5_TXCMP_CODE_SINGLE - single-segment packet, use SEND.
3632  *  MLX5_TXCMP_CODE_EMPW - single-segment packet, use MPW.
3633  */
3634 static __rte_always_inline enum mlx5_txcmp_code
3635 mlx5_tx_able_to_empw(struct mlx5_txq_data *restrict txq,
3636 		     struct mlx5_txq_local *restrict loc,
3637 		     unsigned int olx,
3638 		     bool newp)
3639 {
3640 	/* Check for multi-segment packet. */
3641 	if (newp &&
3642 	    MLX5_TXOFF_CONFIG(MULTI) &&
3643 	    unlikely(NB_SEGS(loc->mbuf) > 1))
3644 		return MLX5_TXCMP_CODE_MULTI;
3645 	/* Check for TSO packet. */
3646 	if (newp &&
3647 	    MLX5_TXOFF_CONFIG(TSO) &&
3648 	    unlikely(loc->mbuf->ol_flags & PKT_TX_TCP_SEG))
3649 		return MLX5_TXCMP_CODE_TSO;
3650 	/* Check if eMPW is enabled at all. */
3651 	if (!MLX5_TXOFF_CONFIG(EMPW))
3652 		return MLX5_TXCMP_CODE_SINGLE;
3653 	/* Check if eMPW can be engaged. */
3654 	if (MLX5_TXOFF_CONFIG(VLAN) &&
3655 	    unlikely(loc->mbuf->ol_flags & PKT_TX_VLAN_PKT) &&
3656 		(!MLX5_TXOFF_CONFIG(INLINE) ||
3657 		 unlikely((rte_pktmbuf_data_len(loc->mbuf) +
3658 			   sizeof(struct rte_vlan_hdr)) > txq->inlen_empw))) {
3659 		/*
3660 		 * eMPW does not support VLAN insertion offload,
3661 		 * we have to inline the entire packet but
3662 		 * packet is too long for inlining.
3663 		 */
3664 		return MLX5_TXCMP_CODE_SINGLE;
3665 	}
3666 	return MLX5_TXCMP_CODE_EMPW;
3667 }
3668 
3669 /**
3670  * Check the next packet attributes to match with the eMPW batch ones.
3671  * In addition, for legacy MPW the packet length is checked either.
3672  *
3673  * @param txq
3674  *   Pointer to TX queue structure.
3675  * @param es
3676  *   Pointer to Ethernet Segment of eMPW batch.
3677  * @param loc
3678  *   Pointer to burst routine local context.
3679  * @param dlen
3680  *   Length of previous packet in MPW descriptor.
3681  * @param olx
3682  *   Configured Tx offloads mask. It is fully defined at
3683  *   compile time and may be used for optimization.
3684  *
3685  * @return
3686  *  true - packet match with eMPW batch attributes.
3687  *  false - no match, eMPW should be restarted.
3688  */
3689 static __rte_always_inline bool
3690 mlx5_tx_match_empw(struct mlx5_txq_data *restrict txq __rte_unused,
3691 		   struct mlx5_wqe_eseg *restrict es,
3692 		   struct mlx5_txq_local *restrict loc,
3693 		   uint32_t dlen,
3694 		   unsigned int olx)
3695 {
3696 	uint8_t swp_flags = 0;
3697 
3698 	/* Compare the checksum flags, if any. */
3699 	if (MLX5_TXOFF_CONFIG(CSUM) &&
3700 	    txq_ol_cksum_to_cs(loc->mbuf) != es->cs_flags)
3701 		return false;
3702 	/* Compare the Software Parser offsets and flags. */
3703 	if (MLX5_TXOFF_CONFIG(SWP) &&
3704 	    (es->swp_offs != txq_mbuf_to_swp(loc, &swp_flags, olx) ||
3705 	     es->swp_flags != swp_flags))
3706 		return false;
3707 	/* Fill metadata field if needed. */
3708 	if (MLX5_TXOFF_CONFIG(METADATA) &&
3709 		es->metadata != (loc->mbuf->ol_flags & PKT_TX_DYNF_METADATA ?
3710 				 *RTE_FLOW_DYNF_METADATA(loc->mbuf) : 0))
3711 		return false;
3712 	/* Legacy MPW can send packets with the same lengt only. */
3713 	if (MLX5_TXOFF_CONFIG(MPW) &&
3714 	    dlen != rte_pktmbuf_data_len(loc->mbuf))
3715 		return false;
3716 	/* There must be no VLAN packets in eMPW loop. */
3717 	if (MLX5_TXOFF_CONFIG(VLAN))
3718 		assert(!(loc->mbuf->ol_flags & PKT_TX_VLAN_PKT));
3719 	return true;
3720 }
3721 
3722 /*
3723  * Update send loop variables and WQE for eMPW loop
3724  * without data inlining. Number of Data Segments is
3725  * equal to the number of sent packets.
3726  *
3727  * @param txq
3728  *   Pointer to TX queue structure.
3729  * @param loc
3730  *   Pointer to burst routine local context.
3731  * @param ds
3732  *   Number of packets/Data Segments/Packets.
3733  * @param slen
3734  *   Accumulated statistics, bytes sent
3735  * @param olx
3736  *   Configured Tx offloads mask. It is fully defined at
3737  *   compile time and may be used for optimization.
3738  *
3739  * @return
3740  *  true - packet match with eMPW batch attributes.
3741  *  false - no match, eMPW should be restarted.
3742  */
3743 static __rte_always_inline void
3744 mlx5_tx_sdone_empw(struct mlx5_txq_data *restrict txq,
3745 		   struct mlx5_txq_local *restrict loc,
3746 		   unsigned int ds,
3747 		   unsigned int slen,
3748 		   unsigned int olx __rte_unused)
3749 {
3750 	assert(!MLX5_TXOFF_CONFIG(INLINE));
3751 #ifdef MLX5_PMD_SOFT_COUNTERS
3752 	/* Update sent data bytes counter. */
3753 	 txq->stats.obytes += slen;
3754 #else
3755 	(void)slen;
3756 #endif
3757 	loc->elts_free -= ds;
3758 	loc->pkts_sent += ds;
3759 	ds += 2;
3760 	loc->wqe_last->cseg.sq_ds = rte_cpu_to_be_32(txq->qp_num_8s | ds);
3761 	txq->wqe_ci += (ds + 3) / 4;
3762 	loc->wqe_free -= (ds + 3) / 4;
3763 }
3764 
3765 /*
3766  * Update send loop variables and WQE for eMPW loop
3767  * with data inlining. Gets the size of pushed descriptors
3768  * and data to the WQE.
3769  *
3770  * @param txq
3771  *   Pointer to TX queue structure.
3772  * @param loc
3773  *   Pointer to burst routine local context.
3774  * @param len
3775  *   Total size of descriptor/data in bytes.
3776  * @param slen
3777  *   Accumulated statistics, data bytes sent.
3778  * @param olx
3779  *   Configured Tx offloads mask. It is fully defined at
3780  *   compile time and may be used for optimization.
3781  *
3782  * @return
3783  *  true - packet match with eMPW batch attributes.
3784  *  false - no match, eMPW should be restarted.
3785  */
3786 static __rte_always_inline void
3787 mlx5_tx_idone_empw(struct mlx5_txq_data *restrict txq,
3788 		   struct mlx5_txq_local *restrict loc,
3789 		   unsigned int len,
3790 		   unsigned int slen,
3791 		   unsigned int olx __rte_unused)
3792 {
3793 	assert(MLX5_TXOFF_CONFIG(INLINE));
3794 	assert((len % MLX5_WSEG_SIZE) == 0);
3795 #ifdef MLX5_PMD_SOFT_COUNTERS
3796 	/* Update sent data bytes counter. */
3797 	 txq->stats.obytes += slen;
3798 #else
3799 	(void)slen;
3800 #endif
3801 	len = len / MLX5_WSEG_SIZE + 2;
3802 	loc->wqe_last->cseg.sq_ds = rte_cpu_to_be_32(txq->qp_num_8s | len);
3803 	txq->wqe_ci += (len + 3) / 4;
3804 	loc->wqe_free -= (len + 3) / 4;
3805 }
3806 
3807 /**
3808  * The set of Tx burst functions for single-segment packets
3809  * without TSO and with Multi-Packet Writing feature support.
3810  * Supports all types of Tx offloads, except multi-packets
3811  * and TSO.
3812  *
3813  * Uses MLX5_OPCODE_EMPW to build WQEs if possible and sends
3814  * as many packet per WQE as it can. If eMPW is not configured
3815  * or packet can not be sent with eMPW (VLAN insertion) the
3816  * ordinary SEND opcode is used and only one packet placed
3817  * in WQE.
3818  *
3819  * Functions stop sending if it encounters the multi-segment
3820  * packet or packet with TSO requested.
3821  *
3822  * The routines are responsible for storing processed mbuf
3823  * into elts ring buffer and update elts_head if inlining
3824  * offload is requested. Otherwise the copying mbufs to elts
3825  * can be postponed and completed at the end of burst routine.
3826  *
3827  * @param txq
3828  *   Pointer to TX queue structure.
3829  * @param[in] pkts
3830  *   Packets to transmit.
3831  * @param pkts_n
3832  *   Number of packets in array.
3833  * @param loc
3834  *   Pointer to burst routine local context.
3835  * @param olx
3836  *   Configured Tx offloads mask. It is fully defined at
3837  *   compile time and may be used for optimization.
3838  *
3839  * @return
3840  *   MLX5_TXCMP_CODE_EXIT - sending is done or impossible.
3841  *   MLX5_TXCMP_CODE_ERROR - some unrecoverable error occurred.
3842  *   MLX5_TXCMP_CODE_MULTI - multi-segment packet encountered.
3843  *   MLX5_TXCMP_CODE_TSO - TSO packet encountered.
3844  *   MLX5_TXCMP_CODE_SINGLE - used inside functions set.
3845  *   MLX5_TXCMP_CODE_EMPW - used inside functions set.
3846  *
3847  * Local context variables updated.
3848  *
3849  *
3850  * The routine sends packets with MLX5_OPCODE_EMPW
3851  * without inlining, this is dedicated optimized branch.
3852  * No VLAN insertion is supported.
3853  */
3854 static __rte_always_inline enum mlx5_txcmp_code
3855 mlx5_tx_burst_empw_simple(struct mlx5_txq_data *restrict txq,
3856 			  struct rte_mbuf **restrict pkts,
3857 			  unsigned int pkts_n,
3858 			  struct mlx5_txq_local *restrict loc,
3859 			  unsigned int olx)
3860 {
3861 	/*
3862 	 * Subroutine is the part of mlx5_tx_burst_single()
3863 	 * and sends single-segment packet with eMPW opcode
3864 	 * without data inlining.
3865 	 */
3866 	assert(!MLX5_TXOFF_CONFIG(INLINE));
3867 	assert(MLX5_TXOFF_CONFIG(EMPW));
3868 	assert(loc->elts_free && loc->wqe_free);
3869 	assert(pkts_n > loc->pkts_sent);
3870 	static_assert(MLX5_EMPW_MIN_PACKETS >= 2, "invalid min size");
3871 	pkts += loc->pkts_sent + 1;
3872 	pkts_n -= loc->pkts_sent;
3873 	for (;;) {
3874 		struct mlx5_wqe_dseg *restrict dseg;
3875 		struct mlx5_wqe_eseg *restrict eseg;
3876 		enum mlx5_txcmp_code ret;
3877 		unsigned int part, loop;
3878 		unsigned int slen = 0;
3879 
3880 next_empw:
3881 		assert(NB_SEGS(loc->mbuf) == 1);
3882 		part = RTE_MIN(pkts_n, MLX5_TXOFF_CONFIG(MPW) ?
3883 				       MLX5_MPW_MAX_PACKETS :
3884 				       MLX5_EMPW_MAX_PACKETS);
3885 		if (unlikely(loc->elts_free < part)) {
3886 			/* We have no enough elts to save all mbufs. */
3887 			if (unlikely(loc->elts_free < MLX5_EMPW_MIN_PACKETS))
3888 				return MLX5_TXCMP_CODE_EXIT;
3889 			/* But we still able to send at least minimal eMPW. */
3890 			part = loc->elts_free;
3891 		}
3892 		/* Check whether we have enough WQEs */
3893 		if (unlikely(loc->wqe_free < ((2 + part + 3) / 4))) {
3894 			if (unlikely(loc->wqe_free <
3895 				((2 + MLX5_EMPW_MIN_PACKETS + 3) / 4)))
3896 				return MLX5_TXCMP_CODE_EXIT;
3897 			part = (loc->wqe_free * 4) - 2;
3898 		}
3899 		if (likely(part > 1))
3900 			rte_prefetch0(*pkts);
3901 		loc->wqe_last = txq->wqes + (txq->wqe_ci & txq->wqe_m);
3902 		/*
3903 		 * Build eMPW title WQEBB:
3904 		 * - Control Segment, eMPW opcode
3905 		 * - Ethernet Segment, no inline
3906 		 */
3907 		mlx5_tx_cseg_init(txq, loc, loc->wqe_last, part + 2,
3908 				  MLX5_OPCODE_ENHANCED_MPSW, olx);
3909 		mlx5_tx_eseg_none(txq, loc, loc->wqe_last,
3910 				  olx & ~MLX5_TXOFF_CONFIG_VLAN);
3911 		eseg = &loc->wqe_last->eseg;
3912 		dseg = &loc->wqe_last->dseg[0];
3913 		loop = part;
3914 		/* Store the packet length for legacy MPW. */
3915 		if (MLX5_TXOFF_CONFIG(MPW))
3916 			eseg->mss = rte_cpu_to_be_16
3917 					(rte_pktmbuf_data_len(loc->mbuf));
3918 		for (;;) {
3919 			uint32_t dlen = rte_pktmbuf_data_len(loc->mbuf);
3920 #ifdef MLX5_PMD_SOFT_COUNTERS
3921 			/* Update sent data bytes counter. */
3922 			slen += dlen;
3923 #endif
3924 			mlx5_tx_dseg_ptr
3925 				(txq, loc, dseg,
3926 				 rte_pktmbuf_mtod(loc->mbuf, uint8_t *),
3927 				 dlen, olx);
3928 			if (unlikely(--loop == 0))
3929 				break;
3930 			loc->mbuf = *pkts++;
3931 			if (likely(loop > 1))
3932 				rte_prefetch0(*pkts);
3933 			ret = mlx5_tx_able_to_empw(txq, loc, olx, true);
3934 			/*
3935 			 * Unroll the completion code to avoid
3936 			 * returning variable value - it results in
3937 			 * unoptimized sequent checking in caller.
3938 			 */
3939 			if (ret == MLX5_TXCMP_CODE_MULTI) {
3940 				part -= loop;
3941 				mlx5_tx_sdone_empw(txq, loc, part, slen, olx);
3942 				if (unlikely(!loc->elts_free ||
3943 					     !loc->wqe_free))
3944 					return MLX5_TXCMP_CODE_EXIT;
3945 				return MLX5_TXCMP_CODE_MULTI;
3946 			}
3947 			assert(NB_SEGS(loc->mbuf) == 1);
3948 			if (ret == MLX5_TXCMP_CODE_TSO) {
3949 				part -= loop;
3950 				mlx5_tx_sdone_empw(txq, loc, part, slen, olx);
3951 				if (unlikely(!loc->elts_free ||
3952 					     !loc->wqe_free))
3953 					return MLX5_TXCMP_CODE_EXIT;
3954 				return MLX5_TXCMP_CODE_TSO;
3955 			}
3956 			if (ret == MLX5_TXCMP_CODE_SINGLE) {
3957 				part -= loop;
3958 				mlx5_tx_sdone_empw(txq, loc, part, slen, olx);
3959 				if (unlikely(!loc->elts_free ||
3960 					     !loc->wqe_free))
3961 					return MLX5_TXCMP_CODE_EXIT;
3962 				return MLX5_TXCMP_CODE_SINGLE;
3963 			}
3964 			if (ret != MLX5_TXCMP_CODE_EMPW) {
3965 				assert(false);
3966 				part -= loop;
3967 				mlx5_tx_sdone_empw(txq, loc, part, slen, olx);
3968 				return MLX5_TXCMP_CODE_ERROR;
3969 			}
3970 			/*
3971 			 * Check whether packet parameters coincide
3972 			 * within assumed eMPW batch:
3973 			 * - check sum settings
3974 			 * - metadata value
3975 			 * - software parser settings
3976 			 * - packets length (legacy MPW only)
3977 			 */
3978 			if (!mlx5_tx_match_empw(txq, eseg, loc, dlen, olx)) {
3979 				assert(loop);
3980 				part -= loop;
3981 				mlx5_tx_sdone_empw(txq, loc, part, slen, olx);
3982 				if (unlikely(!loc->elts_free ||
3983 					     !loc->wqe_free))
3984 					return MLX5_TXCMP_CODE_EXIT;
3985 				pkts_n -= part;
3986 				goto next_empw;
3987 			}
3988 			/* Packet attributes match, continue the same eMPW. */
3989 			++dseg;
3990 			if ((uintptr_t)dseg >= (uintptr_t)txq->wqes_end)
3991 				dseg = (struct mlx5_wqe_dseg *)txq->wqes;
3992 		}
3993 		/* eMPW is built successfully, update loop parameters. */
3994 		assert(!loop);
3995 		assert(pkts_n >= part);
3996 #ifdef MLX5_PMD_SOFT_COUNTERS
3997 		/* Update sent data bytes counter. */
3998 		txq->stats.obytes += slen;
3999 #endif
4000 		loc->elts_free -= part;
4001 		loc->pkts_sent += part;
4002 		txq->wqe_ci += (2 + part + 3) / 4;
4003 		loc->wqe_free -= (2 + part + 3) / 4;
4004 		pkts_n -= part;
4005 		if (unlikely(!pkts_n || !loc->elts_free || !loc->wqe_free))
4006 			return MLX5_TXCMP_CODE_EXIT;
4007 		loc->mbuf = *pkts++;
4008 		ret = mlx5_tx_able_to_empw(txq, loc, olx, true);
4009 		if (unlikely(ret != MLX5_TXCMP_CODE_EMPW))
4010 			return ret;
4011 		/* Continue sending eMPW batches. */
4012 	}
4013 	assert(false);
4014 }
4015 
4016 /**
4017  * The routine sends packets with MLX5_OPCODE_EMPW
4018  * with inlining, optionally supports VLAN insertion.
4019  */
4020 static __rte_always_inline enum mlx5_txcmp_code
4021 mlx5_tx_burst_empw_inline(struct mlx5_txq_data *restrict txq,
4022 			  struct rte_mbuf **restrict pkts,
4023 			  unsigned int pkts_n,
4024 			  struct mlx5_txq_local *restrict loc,
4025 			  unsigned int olx)
4026 {
4027 	/*
4028 	 * Subroutine is the part of mlx5_tx_burst_single()
4029 	 * and sends single-segment packet with eMPW opcode
4030 	 * with data inlining.
4031 	 */
4032 	assert(MLX5_TXOFF_CONFIG(INLINE));
4033 	assert(MLX5_TXOFF_CONFIG(EMPW));
4034 	assert(loc->elts_free && loc->wqe_free);
4035 	assert(pkts_n > loc->pkts_sent);
4036 	static_assert(MLX5_EMPW_MIN_PACKETS >= 2, "invalid min size");
4037 	pkts += loc->pkts_sent + 1;
4038 	pkts_n -= loc->pkts_sent;
4039 	for (;;) {
4040 		struct mlx5_wqe_dseg *restrict dseg;
4041 		struct mlx5_wqe_eseg *restrict eseg;
4042 		enum mlx5_txcmp_code ret;
4043 		unsigned int room, part, nlim;
4044 		unsigned int slen = 0;
4045 
4046 		assert(NB_SEGS(loc->mbuf) == 1);
4047 		/*
4048 		 * Limits the amount of packets in one WQE
4049 		 * to improve CQE latency generation.
4050 		 */
4051 		nlim = RTE_MIN(pkts_n, MLX5_TXOFF_CONFIG(MPW) ?
4052 				       MLX5_MPW_INLINE_MAX_PACKETS :
4053 				       MLX5_EMPW_MAX_PACKETS);
4054 		/* Check whether we have minimal amount WQEs */
4055 		if (unlikely(loc->wqe_free <
4056 			    ((2 + MLX5_EMPW_MIN_PACKETS + 3) / 4)))
4057 			return MLX5_TXCMP_CODE_EXIT;
4058 		if (likely(pkts_n > 1))
4059 			rte_prefetch0(*pkts);
4060 		loc->wqe_last = txq->wqes + (txq->wqe_ci & txq->wqe_m);
4061 		/*
4062 		 * Build eMPW title WQEBB:
4063 		 * - Control Segment, eMPW opcode, zero DS
4064 		 * - Ethernet Segment, no inline
4065 		 */
4066 		mlx5_tx_cseg_init(txq, loc, loc->wqe_last, 0,
4067 				  MLX5_OPCODE_ENHANCED_MPSW, olx);
4068 		mlx5_tx_eseg_none(txq, loc, loc->wqe_last,
4069 				  olx & ~MLX5_TXOFF_CONFIG_VLAN);
4070 		eseg = &loc->wqe_last->eseg;
4071 		dseg = &loc->wqe_last->dseg[0];
4072 		/* Store the packet length for legacy MPW. */
4073 		if (MLX5_TXOFF_CONFIG(MPW))
4074 			eseg->mss = rte_cpu_to_be_16
4075 					(rte_pktmbuf_data_len(loc->mbuf));
4076 		room = RTE_MIN(MLX5_WQE_SIZE_MAX / MLX5_WQE_SIZE,
4077 			       loc->wqe_free) * MLX5_WQE_SIZE -
4078 					MLX5_WQE_CSEG_SIZE -
4079 					MLX5_WQE_ESEG_SIZE;
4080 		/* Build WQE till we have space, packets and resources. */
4081 		part = room;
4082 		for (;;) {
4083 			uint32_t dlen = rte_pktmbuf_data_len(loc->mbuf);
4084 			uint8_t *dptr = rte_pktmbuf_mtod(loc->mbuf, uint8_t *);
4085 			unsigned int tlen;
4086 
4087 			assert(room >= MLX5_WQE_DSEG_SIZE);
4088 			assert((room % MLX5_WQE_DSEG_SIZE) == 0);
4089 			assert((uintptr_t)dseg < (uintptr_t)txq->wqes_end);
4090 			/*
4091 			 * Some Tx offloads may cause an error if
4092 			 * packet is not long enough, check against
4093 			 * assumed minimal length.
4094 			 */
4095 			if (unlikely(dlen <= MLX5_ESEG_MIN_INLINE_SIZE)) {
4096 				part -= room;
4097 				if (unlikely(!part))
4098 					return MLX5_TXCMP_CODE_ERROR;
4099 				/*
4100 				 * We have some successfully built
4101 				 * packet Data Segments to send.
4102 				 */
4103 				mlx5_tx_idone_empw(txq, loc, part, slen, olx);
4104 				return MLX5_TXCMP_CODE_ERROR;
4105 			}
4106 			/* Inline or not inline - that's the Question. */
4107 			if (dlen > txq->inlen_empw)
4108 				goto pointer_empw;
4109 			/* Inline entire packet, optional VLAN insertion. */
4110 			tlen = sizeof(dseg->bcount) + dlen;
4111 			if (MLX5_TXOFF_CONFIG(VLAN) &&
4112 			    loc->mbuf->ol_flags & PKT_TX_VLAN_PKT) {
4113 				/*
4114 				 * The packet length must be checked in
4115 				 * mlx5_tx_able_to_empw() and packet
4116 				 * fits into inline length guaranteed.
4117 				 */
4118 				assert((dlen + sizeof(struct rte_vlan_hdr)) <=
4119 					txq->inlen_empw);
4120 				tlen += sizeof(struct rte_vlan_hdr);
4121 				if (room < tlen)
4122 					break;
4123 				dseg = mlx5_tx_dseg_vlan(txq, loc, dseg,
4124 							 dptr, dlen, olx);
4125 #ifdef MLX5_PMD_SOFT_COUNTERS
4126 				/* Update sent data bytes counter. */
4127 				slen +=	sizeof(struct rte_vlan_hdr);
4128 #endif
4129 			} else {
4130 				if (room < tlen)
4131 					break;
4132 				dseg = mlx5_tx_dseg_empw(txq, loc, dseg,
4133 							 dptr, dlen, olx);
4134 			}
4135 			tlen = RTE_ALIGN(tlen, MLX5_WSEG_SIZE);
4136 			assert(room >= tlen);
4137 			room -= tlen;
4138 			/*
4139 			 * Packet data are completely inlined,
4140 			 * free the packet immediately.
4141 			 */
4142 			rte_pktmbuf_free_seg(loc->mbuf);
4143 			goto next_mbuf;
4144 pointer_empw:
4145 			/*
4146 			 * Not inlinable VLAN packets are
4147 			 * proceeded outside of this routine.
4148 			 */
4149 			assert(room >= MLX5_WQE_DSEG_SIZE);
4150 			if (MLX5_TXOFF_CONFIG(VLAN))
4151 				assert(!(loc->mbuf->ol_flags &
4152 					 PKT_TX_VLAN_PKT));
4153 			mlx5_tx_dseg_ptr(txq, loc, dseg, dptr, dlen, olx);
4154 			/* We have to store mbuf in elts.*/
4155 			txq->elts[txq->elts_head++ & txq->elts_m] = loc->mbuf;
4156 			room -= MLX5_WQE_DSEG_SIZE;
4157 			/* Ring buffer wraparound is checked at the loop end.*/
4158 			++dseg;
4159 next_mbuf:
4160 #ifdef MLX5_PMD_SOFT_COUNTERS
4161 			/* Update sent data bytes counter. */
4162 			slen += dlen;
4163 #endif
4164 			loc->pkts_sent++;
4165 			loc->elts_free--;
4166 			pkts_n--;
4167 			if (unlikely(!pkts_n || !loc->elts_free)) {
4168 				/*
4169 				 * We have no resources/packets to
4170 				 * continue build descriptors.
4171 				 */
4172 				part -= room;
4173 				mlx5_tx_idone_empw(txq, loc, part, slen, olx);
4174 				return MLX5_TXCMP_CODE_EXIT;
4175 			}
4176 			loc->mbuf = *pkts++;
4177 			if (likely(pkts_n > 1))
4178 				rte_prefetch0(*pkts);
4179 			ret = mlx5_tx_able_to_empw(txq, loc, olx, true);
4180 			/*
4181 			 * Unroll the completion code to avoid
4182 			 * returning variable value - it results in
4183 			 * unoptimized sequent checking in caller.
4184 			 */
4185 			if (ret == MLX5_TXCMP_CODE_MULTI) {
4186 				part -= room;
4187 				mlx5_tx_idone_empw(txq, loc, part, slen, olx);
4188 				if (unlikely(!loc->elts_free ||
4189 					     !loc->wqe_free))
4190 					return MLX5_TXCMP_CODE_EXIT;
4191 				return MLX5_TXCMP_CODE_MULTI;
4192 			}
4193 			assert(NB_SEGS(loc->mbuf) == 1);
4194 			if (ret == MLX5_TXCMP_CODE_TSO) {
4195 				part -= room;
4196 				mlx5_tx_idone_empw(txq, loc, part, slen, olx);
4197 				if (unlikely(!loc->elts_free ||
4198 					     !loc->wqe_free))
4199 					return MLX5_TXCMP_CODE_EXIT;
4200 				return MLX5_TXCMP_CODE_TSO;
4201 			}
4202 			if (ret == MLX5_TXCMP_CODE_SINGLE) {
4203 				part -= room;
4204 				mlx5_tx_idone_empw(txq, loc, part, slen, olx);
4205 				if (unlikely(!loc->elts_free ||
4206 					     !loc->wqe_free))
4207 					return MLX5_TXCMP_CODE_EXIT;
4208 				return MLX5_TXCMP_CODE_SINGLE;
4209 			}
4210 			if (ret != MLX5_TXCMP_CODE_EMPW) {
4211 				assert(false);
4212 				part -= room;
4213 				mlx5_tx_idone_empw(txq, loc, part, slen, olx);
4214 				return MLX5_TXCMP_CODE_ERROR;
4215 			}
4216 			/* Check if we have minimal room left. */
4217 			nlim--;
4218 			if (unlikely(!nlim || room < MLX5_WQE_DSEG_SIZE))
4219 				break;
4220 			/*
4221 			 * Check whether packet parameters coincide
4222 			 * within assumed eMPW batch:
4223 			 * - check sum settings
4224 			 * - metadata value
4225 			 * - software parser settings
4226 			 * - packets length (legacy MPW only)
4227 			 */
4228 			if (!mlx5_tx_match_empw(txq, eseg, loc, dlen, olx))
4229 				break;
4230 			/* Packet attributes match, continue the same eMPW. */
4231 			if ((uintptr_t)dseg >= (uintptr_t)txq->wqes_end)
4232 				dseg = (struct mlx5_wqe_dseg *)txq->wqes;
4233 		}
4234 		/*
4235 		 * We get here to close an existing eMPW
4236 		 * session and start the new one.
4237 		 */
4238 		assert(pkts_n);
4239 		part -= room;
4240 		if (unlikely(!part))
4241 			return MLX5_TXCMP_CODE_EXIT;
4242 		mlx5_tx_idone_empw(txq, loc, part, slen, olx);
4243 		if (unlikely(!loc->elts_free ||
4244 			     !loc->wqe_free))
4245 			return MLX5_TXCMP_CODE_EXIT;
4246 		/* Continue the loop with new eMPW session. */
4247 	}
4248 	assert(false);
4249 }
4250 
4251 /**
4252  * The routine sends packets with ordinary MLX5_OPCODE_SEND.
4253  * Data inlining and VLAN insertion are supported.
4254  */
4255 static __rte_always_inline enum mlx5_txcmp_code
4256 mlx5_tx_burst_single_send(struct mlx5_txq_data *restrict txq,
4257 			  struct rte_mbuf **restrict pkts,
4258 			  unsigned int pkts_n,
4259 			  struct mlx5_txq_local *restrict loc,
4260 			  unsigned int olx)
4261 {
4262 	/*
4263 	 * Subroutine is the part of mlx5_tx_burst_single()
4264 	 * and sends single-segment packet with SEND opcode.
4265 	 */
4266 	assert(loc->elts_free && loc->wqe_free);
4267 	assert(pkts_n > loc->pkts_sent);
4268 	pkts += loc->pkts_sent + 1;
4269 	pkts_n -= loc->pkts_sent;
4270 	for (;;) {
4271 		struct mlx5_wqe *restrict wqe;
4272 		enum mlx5_txcmp_code ret;
4273 
4274 		assert(NB_SEGS(loc->mbuf) == 1);
4275 		if (MLX5_TXOFF_CONFIG(INLINE)) {
4276 			unsigned int inlen, vlan = 0;
4277 
4278 			inlen = rte_pktmbuf_data_len(loc->mbuf);
4279 			if (MLX5_TXOFF_CONFIG(VLAN) &&
4280 			    loc->mbuf->ol_flags & PKT_TX_VLAN_PKT) {
4281 				vlan = sizeof(struct rte_vlan_hdr);
4282 				inlen += vlan;
4283 				static_assert((sizeof(struct rte_vlan_hdr) +
4284 					       sizeof(struct rte_ether_hdr)) ==
4285 					       MLX5_ESEG_MIN_INLINE_SIZE,
4286 					       "invalid min inline data size");
4287 			}
4288 			/*
4289 			 * If inlining is enabled at configuration time
4290 			 * the limit must be not less than minimal size.
4291 			 * Otherwise we would do extra check for data
4292 			 * size to avoid crashes due to length overflow.
4293 			 */
4294 			assert(txq->inlen_send >= MLX5_ESEG_MIN_INLINE_SIZE);
4295 			if (inlen <= txq->inlen_send) {
4296 				unsigned int seg_n, wqe_n;
4297 
4298 				rte_prefetch0(rte_pktmbuf_mtod
4299 						(loc->mbuf, uint8_t *));
4300 				/* Check against minimal length. */
4301 				if (inlen <= MLX5_ESEG_MIN_INLINE_SIZE)
4302 					return MLX5_TXCMP_CODE_ERROR;
4303 				/*
4304 				 * Completely inlined packet data WQE:
4305 				 * - Control Segment, SEND opcode
4306 				 * - Ethernet Segment, no VLAN insertion
4307 				 * - Data inlined, VLAN optionally inserted
4308 				 * - Alignment to MLX5_WSEG_SIZE
4309 				 * Have to estimate amount of WQEBBs
4310 				 */
4311 				seg_n = (inlen + 3 * MLX5_WSEG_SIZE -
4312 					 MLX5_ESEG_MIN_INLINE_SIZE +
4313 					 MLX5_WSEG_SIZE - 1) / MLX5_WSEG_SIZE;
4314 				/* Check if there are enough WQEBBs. */
4315 				wqe_n = (seg_n + 3) / 4;
4316 				if (wqe_n > loc->wqe_free)
4317 					return MLX5_TXCMP_CODE_EXIT;
4318 				wqe = txq->wqes + (txq->wqe_ci & txq->wqe_m);
4319 				loc->wqe_last = wqe;
4320 				mlx5_tx_cseg_init(txq, loc, wqe, seg_n,
4321 						  MLX5_OPCODE_SEND, olx);
4322 				mlx5_tx_eseg_data(txq, loc, wqe,
4323 						  vlan, inlen, 0, olx);
4324 				txq->wqe_ci += wqe_n;
4325 				loc->wqe_free -= wqe_n;
4326 				/*
4327 				 * Packet data are completely inlined,
4328 				 * free the packet immediately.
4329 				 */
4330 				rte_pktmbuf_free_seg(loc->mbuf);
4331 			} else if ((!MLX5_TXOFF_CONFIG(EMPW) ||
4332 				     MLX5_TXOFF_CONFIG(MPW)) &&
4333 					txq->inlen_mode) {
4334 				/*
4335 				 * If minimal inlining is requested the eMPW
4336 				 * feature should be disabled due to data is
4337 				 * inlined into Ethernet Segment, which can
4338 				 * not contain inlined data for eMPW due to
4339 				 * segment shared for all packets.
4340 				 */
4341 				struct mlx5_wqe_dseg *restrict dseg;
4342 				unsigned int ds;
4343 				uint8_t *dptr;
4344 
4345 				/*
4346 				 * The inline-mode settings require
4347 				 * to inline the specified amount of
4348 				 * data bytes to the Ethernet Segment.
4349 				 * We should check the free space in
4350 				 * WQE ring buffer to inline partially.
4351 				 */
4352 				assert(txq->inlen_send >= txq->inlen_mode);
4353 				assert(inlen > txq->inlen_mode);
4354 				assert(txq->inlen_mode >=
4355 						MLX5_ESEG_MIN_INLINE_SIZE);
4356 				/*
4357 				 * Check whether there are enough free WQEBBs:
4358 				 * - Control Segment
4359 				 * - Ethernet Segment
4360 				 * - First Segment of inlined Ethernet data
4361 				 * - ... data continued ...
4362 				 * - Finishing Data Segment of pointer type
4363 				 */
4364 				ds = (MLX5_WQE_CSEG_SIZE +
4365 				      MLX5_WQE_ESEG_SIZE +
4366 				      MLX5_WQE_DSEG_SIZE +
4367 				      txq->inlen_mode -
4368 				      MLX5_ESEG_MIN_INLINE_SIZE +
4369 				      MLX5_WQE_DSEG_SIZE +
4370 				      MLX5_WSEG_SIZE - 1) / MLX5_WSEG_SIZE;
4371 				if (loc->wqe_free < ((ds + 3) / 4))
4372 					return MLX5_TXCMP_CODE_EXIT;
4373 				/*
4374 				 * Build the ordinary SEND WQE:
4375 				 * - Control Segment
4376 				 * - Ethernet Segment, inline inlen_mode bytes
4377 				 * - Data Segment of pointer type
4378 				 */
4379 				wqe = txq->wqes + (txq->wqe_ci & txq->wqe_m);
4380 				loc->wqe_last = wqe;
4381 				mlx5_tx_cseg_init(txq, loc, wqe, ds,
4382 						  MLX5_OPCODE_SEND, olx);
4383 				dseg = mlx5_tx_eseg_data(txq, loc, wqe, vlan,
4384 							 txq->inlen_mode,
4385 							 0, olx);
4386 				dptr = rte_pktmbuf_mtod(loc->mbuf, uint8_t *) +
4387 				       txq->inlen_mode - vlan;
4388 				inlen -= txq->inlen_mode;
4389 				mlx5_tx_dseg_ptr(txq, loc, dseg,
4390 						 dptr, inlen, olx);
4391 				/*
4392 				 * WQE is built, update the loop parameters
4393 				 * and got to the next packet.
4394 				 */
4395 				txq->wqe_ci += (ds + 3) / 4;
4396 				loc->wqe_free -= (ds + 3) / 4;
4397 				/* We have to store mbuf in elts.*/
4398 				assert(MLX5_TXOFF_CONFIG(INLINE));
4399 				txq->elts[txq->elts_head++ & txq->elts_m] =
4400 						loc->mbuf;
4401 				--loc->elts_free;
4402 			} else {
4403 				uint8_t *dptr;
4404 				unsigned int dlen;
4405 
4406 				/*
4407 				 * Partially inlined packet data WQE, we have
4408 				 * some space in title WQEBB, we can fill it
4409 				 * with some packet data. It takes one WQEBB,
4410 				 * it is available, no extra space check:
4411 				 * - Control Segment, SEND opcode
4412 				 * - Ethernet Segment, no VLAN insertion
4413 				 * - MLX5_ESEG_MIN_INLINE_SIZE bytes of Data
4414 				 * - Data Segment, pointer type
4415 				 *
4416 				 * We also get here if VLAN insertion is not
4417 				 * supported by HW, the inline is enabled.
4418 				 */
4419 				wqe = txq->wqes + (txq->wqe_ci & txq->wqe_m);
4420 				loc->wqe_last = wqe;
4421 				mlx5_tx_cseg_init(txq, loc, wqe, 4,
4422 						  MLX5_OPCODE_SEND, olx);
4423 				mlx5_tx_eseg_dmin(txq, loc, wqe, vlan, olx);
4424 				dptr = rte_pktmbuf_mtod(loc->mbuf, uint8_t *) +
4425 				       MLX5_ESEG_MIN_INLINE_SIZE - vlan;
4426 				/*
4427 				 * The length check is performed above, by
4428 				 * comparing with txq->inlen_send. We should
4429 				 * not get overflow here.
4430 				 */
4431 				assert(inlen > MLX5_ESEG_MIN_INLINE_SIZE);
4432 				dlen = inlen - MLX5_ESEG_MIN_INLINE_SIZE;
4433 				mlx5_tx_dseg_ptr(txq, loc, &wqe->dseg[1],
4434 						 dptr, dlen, olx);
4435 				++txq->wqe_ci;
4436 				--loc->wqe_free;
4437 				/* We have to store mbuf in elts.*/
4438 				assert(MLX5_TXOFF_CONFIG(INLINE));
4439 				txq->elts[txq->elts_head++ & txq->elts_m] =
4440 						loc->mbuf;
4441 				--loc->elts_free;
4442 			}
4443 #ifdef MLX5_PMD_SOFT_COUNTERS
4444 			/* Update sent data bytes counter. */
4445 			txq->stats.obytes += vlan +
4446 					rte_pktmbuf_data_len(loc->mbuf);
4447 #endif
4448 		} else {
4449 			/*
4450 			 * No inline at all, it means the CPU cycles saving
4451 			 * is prioritized at configuration, we should not
4452 			 * copy any packet data to WQE.
4453 			 *
4454 			 * SEND WQE, one WQEBB:
4455 			 * - Control Segment, SEND opcode
4456 			 * - Ethernet Segment, optional VLAN, no inline
4457 			 * - Data Segment, pointer type
4458 			 */
4459 			wqe = txq->wqes + (txq->wqe_ci & txq->wqe_m);
4460 			loc->wqe_last = wqe;
4461 			mlx5_tx_cseg_init(txq, loc, wqe, 3,
4462 					  MLX5_OPCODE_SEND, olx);
4463 			mlx5_tx_eseg_none(txq, loc, wqe, olx);
4464 			mlx5_tx_dseg_ptr
4465 				(txq, loc, &wqe->dseg[0],
4466 				 rte_pktmbuf_mtod(loc->mbuf, uint8_t *),
4467 				 rte_pktmbuf_data_len(loc->mbuf), olx);
4468 			++txq->wqe_ci;
4469 			--loc->wqe_free;
4470 			/*
4471 			 * We should not store mbuf pointer in elts
4472 			 * if no inlining is configured, this is done
4473 			 * by calling routine in a batch copy.
4474 			 */
4475 			assert(!MLX5_TXOFF_CONFIG(INLINE));
4476 			--loc->elts_free;
4477 #ifdef MLX5_PMD_SOFT_COUNTERS
4478 			/* Update sent data bytes counter. */
4479 			txq->stats.obytes += rte_pktmbuf_data_len(loc->mbuf);
4480 			if (MLX5_TXOFF_CONFIG(VLAN) &&
4481 			    loc->mbuf->ol_flags & PKT_TX_VLAN_PKT)
4482 				txq->stats.obytes +=
4483 					sizeof(struct rte_vlan_hdr);
4484 #endif
4485 		}
4486 		++loc->pkts_sent;
4487 		--pkts_n;
4488 		if (unlikely(!pkts_n || !loc->elts_free || !loc->wqe_free))
4489 			return MLX5_TXCMP_CODE_EXIT;
4490 		loc->mbuf = *pkts++;
4491 		if (pkts_n > 1)
4492 			rte_prefetch0(*pkts);
4493 		ret = mlx5_tx_able_to_empw(txq, loc, olx, true);
4494 		if (unlikely(ret != MLX5_TXCMP_CODE_SINGLE))
4495 			return ret;
4496 	}
4497 	assert(false);
4498 }
4499 
4500 static __rte_always_inline enum mlx5_txcmp_code
4501 mlx5_tx_burst_single(struct mlx5_txq_data *restrict txq,
4502 		     struct rte_mbuf **restrict pkts,
4503 		     unsigned int pkts_n,
4504 		     struct mlx5_txq_local *restrict loc,
4505 		     unsigned int olx)
4506 {
4507 	enum mlx5_txcmp_code ret;
4508 
4509 	ret = mlx5_tx_able_to_empw(txq, loc, olx, false);
4510 	if (ret == MLX5_TXCMP_CODE_SINGLE)
4511 		goto ordinary_send;
4512 	assert(ret == MLX5_TXCMP_CODE_EMPW);
4513 	for (;;) {
4514 		/* Optimize for inline/no inline eMPW send. */
4515 		ret = (MLX5_TXOFF_CONFIG(INLINE)) ?
4516 			mlx5_tx_burst_empw_inline
4517 				(txq, pkts, pkts_n, loc, olx) :
4518 			mlx5_tx_burst_empw_simple
4519 				(txq, pkts, pkts_n, loc, olx);
4520 		if (ret != MLX5_TXCMP_CODE_SINGLE)
4521 			return ret;
4522 		/* The resources to send one packet should remain. */
4523 		assert(loc->elts_free && loc->wqe_free);
4524 ordinary_send:
4525 		ret = mlx5_tx_burst_single_send(txq, pkts, pkts_n, loc, olx);
4526 		assert(ret != MLX5_TXCMP_CODE_SINGLE);
4527 		if (ret != MLX5_TXCMP_CODE_EMPW)
4528 			return ret;
4529 		/* The resources to send one packet should remain. */
4530 		assert(loc->elts_free && loc->wqe_free);
4531 	}
4532 }
4533 
4534 /**
4535  * DPDK Tx callback template. This is configured template
4536  * used to generate routines optimized for specified offload setup.
4537  * One of this generated functions is chosen at SQ configuration
4538  * time.
4539  *
4540  * @param txq
4541  *   Generic pointer to TX queue structure.
4542  * @param[in] pkts
4543  *   Packets to transmit.
4544  * @param pkts_n
4545  *   Number of packets in array.
4546  * @param olx
4547  *   Configured offloads mask, presents the bits of MLX5_TXOFF_CONFIG_xxx
4548  *   values. Should be static to take compile time static configuration
4549  *   advantages.
4550  *
4551  * @return
4552  *   Number of packets successfully transmitted (<= pkts_n).
4553  */
4554 static __rte_always_inline uint16_t
4555 mlx5_tx_burst_tmpl(struct mlx5_txq_data *restrict txq,
4556 		   struct rte_mbuf **restrict pkts,
4557 		   uint16_t pkts_n,
4558 		   unsigned int olx)
4559 {
4560 	struct mlx5_txq_local loc;
4561 	enum mlx5_txcmp_code ret;
4562 	unsigned int part;
4563 
4564 	assert(txq->elts_s >= (uint16_t)(txq->elts_head - txq->elts_tail));
4565 	assert(txq->wqe_s >= (uint16_t)(txq->wqe_ci - txq->wqe_pi));
4566 	if (unlikely(!pkts_n))
4567 		return 0;
4568 	loc.pkts_sent = 0;
4569 	loc.pkts_copy = 0;
4570 	loc.wqe_last = NULL;
4571 
4572 send_loop:
4573 	loc.pkts_loop = loc.pkts_sent;
4574 	/*
4575 	 * Check if there are some CQEs, if any:
4576 	 * - process an encountered errors
4577 	 * - process the completed WQEs
4578 	 * - free related mbufs
4579 	 * - doorbell the NIC about processed CQEs
4580 	 */
4581 	rte_prefetch0(*(pkts + loc.pkts_sent));
4582 	mlx5_tx_handle_completion(txq, olx);
4583 	/*
4584 	 * Calculate the number of available resources - elts and WQEs.
4585 	 * There are two possible different scenarios:
4586 	 * - no data inlining into WQEs, one WQEBB may contains upto
4587 	 *   four packets, in this case elts become scarce resource
4588 	 * - data inlining into WQEs, one packet may require multiple
4589 	 *   WQEBBs, the WQEs become the limiting factor.
4590 	 */
4591 	assert(txq->elts_s >= (uint16_t)(txq->elts_head - txq->elts_tail));
4592 	loc.elts_free = txq->elts_s -
4593 				(uint16_t)(txq->elts_head - txq->elts_tail);
4594 	assert(txq->wqe_s >= (uint16_t)(txq->wqe_ci - txq->wqe_pi));
4595 	loc.wqe_free = txq->wqe_s -
4596 				(uint16_t)(txq->wqe_ci - txq->wqe_pi);
4597 	if (unlikely(!loc.elts_free || !loc.wqe_free))
4598 		goto burst_exit;
4599 	for (;;) {
4600 		/*
4601 		 * Fetch the packet from array. Usually this is
4602 		 * the first packet in series of multi/single
4603 		 * segment packets.
4604 		 */
4605 		loc.mbuf = *(pkts + loc.pkts_sent);
4606 		/* Dedicated branch for multi-segment packets. */
4607 		if (MLX5_TXOFF_CONFIG(MULTI) &&
4608 		    unlikely(NB_SEGS(loc.mbuf) > 1)) {
4609 			/*
4610 			 * Multi-segment packet encountered.
4611 			 * Hardware is able to process it only
4612 			 * with SEND/TSO opcodes, one packet
4613 			 * per WQE, do it in dedicated routine.
4614 			 */
4615 enter_send_multi:
4616 			assert(loc.pkts_sent >= loc.pkts_copy);
4617 			part = loc.pkts_sent - loc.pkts_copy;
4618 			if (!MLX5_TXOFF_CONFIG(INLINE) && part) {
4619 				/*
4620 				 * There are some single-segment mbufs not
4621 				 * stored in elts. The mbufs must be in the
4622 				 * same order as WQEs, so we must copy the
4623 				 * mbufs to elts here, before the coming
4624 				 * multi-segment packet mbufs is appended.
4625 				 */
4626 				mlx5_tx_copy_elts(txq, pkts + loc.pkts_copy,
4627 						  part, olx);
4628 				loc.pkts_copy = loc.pkts_sent;
4629 			}
4630 			assert(pkts_n > loc.pkts_sent);
4631 			ret = mlx5_tx_burst_mseg(txq, pkts, pkts_n, &loc, olx);
4632 			if (!MLX5_TXOFF_CONFIG(INLINE))
4633 				loc.pkts_copy = loc.pkts_sent;
4634 			/*
4635 			 * These returned code checks are supposed
4636 			 * to be optimized out due to routine inlining.
4637 			 */
4638 			if (ret == MLX5_TXCMP_CODE_EXIT) {
4639 				/*
4640 				 * The routine returns this code when
4641 				 * all packets are sent or there is no
4642 				 * enough resources to complete request.
4643 				 */
4644 				break;
4645 			}
4646 			if (ret == MLX5_TXCMP_CODE_ERROR) {
4647 				/*
4648 				 * The routine returns this code when
4649 				 * some error in the incoming packets
4650 				 * format occurred.
4651 				 */
4652 				txq->stats.oerrors++;
4653 				break;
4654 			}
4655 			if (ret == MLX5_TXCMP_CODE_SINGLE) {
4656 				/*
4657 				 * The single-segment packet was encountered
4658 				 * in the array, try to send it with the
4659 				 * best optimized way, possible engaging eMPW.
4660 				 */
4661 				goto enter_send_single;
4662 			}
4663 			if (MLX5_TXOFF_CONFIG(TSO) &&
4664 			    ret == MLX5_TXCMP_CODE_TSO) {
4665 				/*
4666 				 * The single-segment TSO packet was
4667 				 * encountered in the array.
4668 				 */
4669 				goto enter_send_tso;
4670 			}
4671 			/* We must not get here. Something is going wrong. */
4672 			assert(false);
4673 			txq->stats.oerrors++;
4674 			break;
4675 		}
4676 		/* Dedicated branch for single-segment TSO packets. */
4677 		if (MLX5_TXOFF_CONFIG(TSO) &&
4678 		    unlikely(loc.mbuf->ol_flags & PKT_TX_TCP_SEG)) {
4679 			/*
4680 			 * TSO might require special way for inlining
4681 			 * (dedicated parameters) and is sent with
4682 			 * MLX5_OPCODE_TSO opcode only, provide this
4683 			 * in dedicated branch.
4684 			 */
4685 enter_send_tso:
4686 			assert(NB_SEGS(loc.mbuf) == 1);
4687 			assert(pkts_n > loc.pkts_sent);
4688 			ret = mlx5_tx_burst_tso(txq, pkts, pkts_n, &loc, olx);
4689 			/*
4690 			 * These returned code checks are supposed
4691 			 * to be optimized out due to routine inlining.
4692 			 */
4693 			if (ret == MLX5_TXCMP_CODE_EXIT)
4694 				break;
4695 			if (ret == MLX5_TXCMP_CODE_ERROR) {
4696 				txq->stats.oerrors++;
4697 				break;
4698 			}
4699 			if (ret == MLX5_TXCMP_CODE_SINGLE)
4700 				goto enter_send_single;
4701 			if (MLX5_TXOFF_CONFIG(MULTI) &&
4702 			    ret == MLX5_TXCMP_CODE_MULTI) {
4703 				/*
4704 				 * The multi-segment packet was
4705 				 * encountered in the array.
4706 				 */
4707 				goto enter_send_multi;
4708 			}
4709 			/* We must not get here. Something is going wrong. */
4710 			assert(false);
4711 			txq->stats.oerrors++;
4712 			break;
4713 		}
4714 		/*
4715 		 * The dedicated branch for the single-segment packets
4716 		 * without TSO. Often these ones can be sent using
4717 		 * MLX5_OPCODE_EMPW with multiple packets in one WQE.
4718 		 * The routine builds the WQEs till it encounters
4719 		 * the TSO or multi-segment packet (in case if these
4720 		 * offloads are requested at SQ configuration time).
4721 		 */
4722 enter_send_single:
4723 		assert(pkts_n > loc.pkts_sent);
4724 		ret = mlx5_tx_burst_single(txq, pkts, pkts_n, &loc, olx);
4725 		/*
4726 		 * These returned code checks are supposed
4727 		 * to be optimized out due to routine inlining.
4728 		 */
4729 		if (ret == MLX5_TXCMP_CODE_EXIT)
4730 			break;
4731 		if (ret == MLX5_TXCMP_CODE_ERROR) {
4732 			txq->stats.oerrors++;
4733 			break;
4734 		}
4735 		if (MLX5_TXOFF_CONFIG(MULTI) &&
4736 		    ret == MLX5_TXCMP_CODE_MULTI) {
4737 			/*
4738 			 * The multi-segment packet was
4739 			 * encountered in the array.
4740 			 */
4741 			goto enter_send_multi;
4742 		}
4743 		if (MLX5_TXOFF_CONFIG(TSO) &&
4744 		    ret == MLX5_TXCMP_CODE_TSO) {
4745 			/*
4746 			 * The single-segment TSO packet was
4747 			 * encountered in the array.
4748 			 */
4749 			goto enter_send_tso;
4750 		}
4751 		/* We must not get here. Something is going wrong. */
4752 		assert(false);
4753 		txq->stats.oerrors++;
4754 		break;
4755 	}
4756 	/*
4757 	 * Main Tx loop is completed, do the rest:
4758 	 * - set completion request if thresholds are reached
4759 	 * - doorbell the hardware
4760 	 * - copy the rest of mbufs to elts (if any)
4761 	 */
4762 	assert(MLX5_TXOFF_CONFIG(INLINE) || loc.pkts_sent >= loc.pkts_copy);
4763 	/* Take a shortcut if nothing is sent. */
4764 	if (unlikely(loc.pkts_sent == loc.pkts_loop))
4765 		goto burst_exit;
4766 	/* Request CQE generation if limits are reached. */
4767 	mlx5_tx_request_completion(txq, &loc, olx);
4768 	/*
4769 	 * Ring QP doorbell immediately after WQE building completion
4770 	 * to improve latencies. The pure software related data treatment
4771 	 * can be completed after doorbell. Tx CQEs for this SQ are
4772 	 * processed in this thread only by the polling.
4773 	 *
4774 	 * The rdma core library can map doorbell register in two ways,
4775 	 * depending on the environment variable "MLX5_SHUT_UP_BF":
4776 	 *
4777 	 * - as regular cached memory, the variable is either missing or
4778 	 *   set to zero. This type of mapping may cause the significant
4779 	 *   doorbell register writing latency and requires explicit
4780 	 *   memory write barrier to mitigate this issue and prevent
4781 	 *   write combining.
4782 	 *
4783 	 * - as non-cached memory, the variable is present and set to
4784 	 *   not "0" value. This type of mapping may cause performance
4785 	 *   impact under heavy loading conditions but the explicit write
4786 	 *   memory barrier is not required and it may improve core
4787 	 *   performance.
4788 	 *
4789 	 * - the legacy behaviour (prior 19.08 release) was to use some
4790 	 *   heuristics to decide whether write memory barrier should
4791 	 *   be performed. This behavior is supported with specifying
4792 	 *   tx_db_nc=2, write barrier is skipped if application
4793 	 *   provides the full recommended burst of packets, it
4794 	 *   supposes the next packets are coming and the write barrier
4795 	 *   will be issued on the next burst (after descriptor writing,
4796 	 *   at least).
4797 	 */
4798 	mlx5_tx_dbrec_cond_wmb(txq, loc.wqe_last, !txq->db_nc &&
4799 			(!txq->db_heu || pkts_n % MLX5_TX_DEFAULT_BURST));
4800 	/* Not all of the mbufs may be stored into elts yet. */
4801 	part = MLX5_TXOFF_CONFIG(INLINE) ? 0 : loc.pkts_sent - loc.pkts_copy;
4802 	if (!MLX5_TXOFF_CONFIG(INLINE) && part) {
4803 		/*
4804 		 * There are some single-segment mbufs not stored in elts.
4805 		 * It can be only if the last packet was single-segment.
4806 		 * The copying is gathered into one place due to it is
4807 		 * a good opportunity to optimize that with SIMD.
4808 		 * Unfortunately if inlining is enabled the gaps in
4809 		 * pointer array may happen due to early freeing of the
4810 		 * inlined mbufs.
4811 		 */
4812 		mlx5_tx_copy_elts(txq, pkts + loc.pkts_copy, part, olx);
4813 		loc.pkts_copy = loc.pkts_sent;
4814 	}
4815 	assert(txq->elts_s >= (uint16_t)(txq->elts_head - txq->elts_tail));
4816 	assert(txq->wqe_s >= (uint16_t)(txq->wqe_ci - txq->wqe_pi));
4817 	if (pkts_n > loc.pkts_sent) {
4818 		/*
4819 		 * If burst size is large there might be no enough CQE
4820 		 * fetched from completion queue and no enough resources
4821 		 * freed to send all the packets.
4822 		 */
4823 		goto send_loop;
4824 	}
4825 burst_exit:
4826 #ifdef MLX5_PMD_SOFT_COUNTERS
4827 	/* Increment sent packets counter. */
4828 	txq->stats.opackets += loc.pkts_sent;
4829 #endif
4830 	return loc.pkts_sent;
4831 }
4832 
4833 /* Generate routines with Enhanced Multi-Packet Write support. */
4834 MLX5_TXOFF_DECL(full_empw,
4835 		MLX5_TXOFF_CONFIG_FULL | MLX5_TXOFF_CONFIG_EMPW)
4836 
4837 MLX5_TXOFF_DECL(none_empw,
4838 		MLX5_TXOFF_CONFIG_NONE | MLX5_TXOFF_CONFIG_EMPW)
4839 
4840 MLX5_TXOFF_DECL(md_empw,
4841 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
4842 
4843 MLX5_TXOFF_DECL(mt_empw,
4844 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
4845 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
4846 
4847 MLX5_TXOFF_DECL(mtsc_empw,
4848 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
4849 		MLX5_TXOFF_CONFIG_SWP |	MLX5_TXOFF_CONFIG_CSUM |
4850 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
4851 
4852 MLX5_TXOFF_DECL(mti_empw,
4853 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
4854 		MLX5_TXOFF_CONFIG_INLINE |
4855 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
4856 
4857 MLX5_TXOFF_DECL(mtv_empw,
4858 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
4859 		MLX5_TXOFF_CONFIG_VLAN |
4860 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
4861 
4862 MLX5_TXOFF_DECL(mtiv_empw,
4863 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
4864 		MLX5_TXOFF_CONFIG_INLINE | MLX5_TXOFF_CONFIG_VLAN |
4865 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
4866 
4867 MLX5_TXOFF_DECL(sc_empw,
4868 		MLX5_TXOFF_CONFIG_SWP |	MLX5_TXOFF_CONFIG_CSUM |
4869 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
4870 
4871 MLX5_TXOFF_DECL(sci_empw,
4872 		MLX5_TXOFF_CONFIG_SWP |	MLX5_TXOFF_CONFIG_CSUM |
4873 		MLX5_TXOFF_CONFIG_INLINE |
4874 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
4875 
4876 MLX5_TXOFF_DECL(scv_empw,
4877 		MLX5_TXOFF_CONFIG_SWP |	MLX5_TXOFF_CONFIG_CSUM |
4878 		MLX5_TXOFF_CONFIG_VLAN |
4879 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
4880 
4881 MLX5_TXOFF_DECL(sciv_empw,
4882 		MLX5_TXOFF_CONFIG_SWP |	MLX5_TXOFF_CONFIG_CSUM |
4883 		MLX5_TXOFF_CONFIG_INLINE | MLX5_TXOFF_CONFIG_VLAN |
4884 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
4885 
4886 MLX5_TXOFF_DECL(i_empw,
4887 		MLX5_TXOFF_CONFIG_INLINE |
4888 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
4889 
4890 MLX5_TXOFF_DECL(v_empw,
4891 		MLX5_TXOFF_CONFIG_VLAN |
4892 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
4893 
4894 MLX5_TXOFF_DECL(iv_empw,
4895 		MLX5_TXOFF_CONFIG_INLINE | MLX5_TXOFF_CONFIG_VLAN |
4896 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
4897 
4898 /* Generate routines without Enhanced Multi-Packet Write support. */
4899 MLX5_TXOFF_DECL(full,
4900 		MLX5_TXOFF_CONFIG_FULL)
4901 
4902 MLX5_TXOFF_DECL(none,
4903 		MLX5_TXOFF_CONFIG_NONE)
4904 
4905 MLX5_TXOFF_DECL(md,
4906 		MLX5_TXOFF_CONFIG_METADATA)
4907 
4908 MLX5_TXOFF_DECL(mt,
4909 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
4910 		MLX5_TXOFF_CONFIG_METADATA)
4911 
4912 MLX5_TXOFF_DECL(mtsc,
4913 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
4914 		MLX5_TXOFF_CONFIG_SWP |	MLX5_TXOFF_CONFIG_CSUM |
4915 		MLX5_TXOFF_CONFIG_METADATA)
4916 
4917 MLX5_TXOFF_DECL(mti,
4918 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
4919 		MLX5_TXOFF_CONFIG_INLINE |
4920 		MLX5_TXOFF_CONFIG_METADATA)
4921 
4922 
4923 MLX5_TXOFF_DECL(mtv,
4924 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
4925 		MLX5_TXOFF_CONFIG_VLAN |
4926 		MLX5_TXOFF_CONFIG_METADATA)
4927 
4928 
4929 MLX5_TXOFF_DECL(mtiv,
4930 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
4931 		MLX5_TXOFF_CONFIG_INLINE | MLX5_TXOFF_CONFIG_VLAN |
4932 		MLX5_TXOFF_CONFIG_METADATA)
4933 
4934 MLX5_TXOFF_DECL(sc,
4935 		MLX5_TXOFF_CONFIG_SWP |	MLX5_TXOFF_CONFIG_CSUM |
4936 		MLX5_TXOFF_CONFIG_METADATA)
4937 
4938 MLX5_TXOFF_DECL(sci,
4939 		MLX5_TXOFF_CONFIG_SWP |	MLX5_TXOFF_CONFIG_CSUM |
4940 		MLX5_TXOFF_CONFIG_INLINE |
4941 		MLX5_TXOFF_CONFIG_METADATA)
4942 
4943 
4944 MLX5_TXOFF_DECL(scv,
4945 		MLX5_TXOFF_CONFIG_SWP |	MLX5_TXOFF_CONFIG_CSUM |
4946 		MLX5_TXOFF_CONFIG_VLAN |
4947 		MLX5_TXOFF_CONFIG_METADATA)
4948 
4949 
4950 MLX5_TXOFF_DECL(sciv,
4951 		MLX5_TXOFF_CONFIG_SWP |	MLX5_TXOFF_CONFIG_CSUM |
4952 		MLX5_TXOFF_CONFIG_INLINE | MLX5_TXOFF_CONFIG_VLAN |
4953 		MLX5_TXOFF_CONFIG_METADATA)
4954 
4955 MLX5_TXOFF_DECL(i,
4956 		MLX5_TXOFF_CONFIG_INLINE |
4957 		MLX5_TXOFF_CONFIG_METADATA)
4958 
4959 MLX5_TXOFF_DECL(v,
4960 		MLX5_TXOFF_CONFIG_VLAN |
4961 		MLX5_TXOFF_CONFIG_METADATA)
4962 
4963 MLX5_TXOFF_DECL(iv,
4964 		MLX5_TXOFF_CONFIG_INLINE | MLX5_TXOFF_CONFIG_VLAN |
4965 		MLX5_TXOFF_CONFIG_METADATA)
4966 
4967 /*
4968  * Generate routines with Legacy Multi-Packet Write support.
4969  * This mode is supported by ConnectX-4LX only and imposes
4970  * offload limitations, not supported:
4971  *   - ACL/Flows (metadata are becoming meaningless)
4972  *   - WQE Inline headers
4973  *   - SRIOV (E-Switch offloads)
4974  *   - VLAN insertion
4975  *   - tunnel encapsulation/decapsulation
4976  *   - TSO
4977  */
4978 MLX5_TXOFF_DECL(none_mpw,
4979 		MLX5_TXOFF_CONFIG_NONE | MLX5_TXOFF_CONFIG_EMPW |
4980 		MLX5_TXOFF_CONFIG_MPW)
4981 
4982 MLX5_TXOFF_DECL(mci_mpw,
4983 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_CSUM |
4984 		MLX5_TXOFF_CONFIG_INLINE | MLX5_TXOFF_CONFIG_EMPW |
4985 		MLX5_TXOFF_CONFIG_MPW)
4986 
4987 MLX5_TXOFF_DECL(mc_mpw,
4988 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_CSUM |
4989 		MLX5_TXOFF_CONFIG_EMPW | MLX5_TXOFF_CONFIG_MPW)
4990 
4991 MLX5_TXOFF_DECL(i_mpw,
4992 		MLX5_TXOFF_CONFIG_INLINE | MLX5_TXOFF_CONFIG_EMPW |
4993 		MLX5_TXOFF_CONFIG_MPW)
4994 
4995 /*
4996  * Array of declared and compiled Tx burst function and corresponding
4997  * supported offloads set. The array is used to select the Tx burst
4998  * function for specified offloads set at Tx queue configuration time.
4999  */
5000 const struct {
5001 	eth_tx_burst_t func;
5002 	unsigned int olx;
5003 } txoff_func[] = {
5004 MLX5_TXOFF_INFO(full_empw,
5005 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
5006 		MLX5_TXOFF_CONFIG_SWP |	MLX5_TXOFF_CONFIG_CSUM |
5007 		MLX5_TXOFF_CONFIG_INLINE | MLX5_TXOFF_CONFIG_VLAN |
5008 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
5009 
5010 MLX5_TXOFF_INFO(none_empw,
5011 		MLX5_TXOFF_CONFIG_NONE | MLX5_TXOFF_CONFIG_EMPW)
5012 
5013 MLX5_TXOFF_INFO(md_empw,
5014 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
5015 
5016 MLX5_TXOFF_INFO(mt_empw,
5017 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
5018 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
5019 
5020 MLX5_TXOFF_INFO(mtsc_empw,
5021 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
5022 		MLX5_TXOFF_CONFIG_SWP |	MLX5_TXOFF_CONFIG_CSUM |
5023 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
5024 
5025 MLX5_TXOFF_INFO(mti_empw,
5026 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
5027 		MLX5_TXOFF_CONFIG_INLINE |
5028 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
5029 
5030 MLX5_TXOFF_INFO(mtv_empw,
5031 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
5032 		MLX5_TXOFF_CONFIG_VLAN |
5033 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
5034 
5035 MLX5_TXOFF_INFO(mtiv_empw,
5036 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
5037 		MLX5_TXOFF_CONFIG_INLINE | MLX5_TXOFF_CONFIG_VLAN |
5038 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
5039 
5040 MLX5_TXOFF_INFO(sc_empw,
5041 		MLX5_TXOFF_CONFIG_SWP |	MLX5_TXOFF_CONFIG_CSUM |
5042 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
5043 
5044 MLX5_TXOFF_INFO(sci_empw,
5045 		MLX5_TXOFF_CONFIG_SWP |	MLX5_TXOFF_CONFIG_CSUM |
5046 		MLX5_TXOFF_CONFIG_INLINE |
5047 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
5048 
5049 MLX5_TXOFF_INFO(scv_empw,
5050 		MLX5_TXOFF_CONFIG_SWP |	MLX5_TXOFF_CONFIG_CSUM |
5051 		MLX5_TXOFF_CONFIG_VLAN |
5052 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
5053 
5054 MLX5_TXOFF_INFO(sciv_empw,
5055 		MLX5_TXOFF_CONFIG_SWP |	MLX5_TXOFF_CONFIG_CSUM |
5056 		MLX5_TXOFF_CONFIG_INLINE | MLX5_TXOFF_CONFIG_VLAN |
5057 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
5058 
5059 MLX5_TXOFF_INFO(i_empw,
5060 		MLX5_TXOFF_CONFIG_INLINE |
5061 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
5062 
5063 MLX5_TXOFF_INFO(v_empw,
5064 		MLX5_TXOFF_CONFIG_VLAN |
5065 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
5066 
5067 MLX5_TXOFF_INFO(iv_empw,
5068 		MLX5_TXOFF_CONFIG_INLINE | MLX5_TXOFF_CONFIG_VLAN |
5069 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
5070 
5071 MLX5_TXOFF_INFO(full,
5072 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
5073 		MLX5_TXOFF_CONFIG_SWP |	MLX5_TXOFF_CONFIG_CSUM |
5074 		MLX5_TXOFF_CONFIG_INLINE | MLX5_TXOFF_CONFIG_VLAN |
5075 		MLX5_TXOFF_CONFIG_METADATA)
5076 
5077 MLX5_TXOFF_INFO(none,
5078 		MLX5_TXOFF_CONFIG_NONE)
5079 
5080 MLX5_TXOFF_INFO(md,
5081 		MLX5_TXOFF_CONFIG_METADATA)
5082 
5083 MLX5_TXOFF_INFO(mt,
5084 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
5085 		MLX5_TXOFF_CONFIG_METADATA)
5086 
5087 MLX5_TXOFF_INFO(mtsc,
5088 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
5089 		MLX5_TXOFF_CONFIG_SWP |	MLX5_TXOFF_CONFIG_CSUM |
5090 		MLX5_TXOFF_CONFIG_METADATA)
5091 
5092 MLX5_TXOFF_INFO(mti,
5093 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
5094 		MLX5_TXOFF_CONFIG_INLINE |
5095 		MLX5_TXOFF_CONFIG_METADATA)
5096 
5097 MLX5_TXOFF_INFO(mtv,
5098 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
5099 		MLX5_TXOFF_CONFIG_VLAN |
5100 		MLX5_TXOFF_CONFIG_METADATA)
5101 
5102 MLX5_TXOFF_INFO(mtiv,
5103 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
5104 		MLX5_TXOFF_CONFIG_INLINE | MLX5_TXOFF_CONFIG_VLAN |
5105 		MLX5_TXOFF_CONFIG_METADATA)
5106 
5107 MLX5_TXOFF_INFO(sc,
5108 		MLX5_TXOFF_CONFIG_SWP |	MLX5_TXOFF_CONFIG_CSUM |
5109 		MLX5_TXOFF_CONFIG_METADATA)
5110 
5111 MLX5_TXOFF_INFO(sci,
5112 		MLX5_TXOFF_CONFIG_SWP |	MLX5_TXOFF_CONFIG_CSUM |
5113 		MLX5_TXOFF_CONFIG_INLINE |
5114 		MLX5_TXOFF_CONFIG_METADATA)
5115 
5116 MLX5_TXOFF_INFO(scv,
5117 		MLX5_TXOFF_CONFIG_SWP |	MLX5_TXOFF_CONFIG_CSUM |
5118 		MLX5_TXOFF_CONFIG_VLAN |
5119 		MLX5_TXOFF_CONFIG_METADATA)
5120 
5121 MLX5_TXOFF_INFO(sciv,
5122 		MLX5_TXOFF_CONFIG_SWP |	MLX5_TXOFF_CONFIG_CSUM |
5123 		MLX5_TXOFF_CONFIG_INLINE | MLX5_TXOFF_CONFIG_VLAN |
5124 		MLX5_TXOFF_CONFIG_METADATA)
5125 
5126 MLX5_TXOFF_INFO(i,
5127 		MLX5_TXOFF_CONFIG_INLINE |
5128 		MLX5_TXOFF_CONFIG_METADATA)
5129 
5130 MLX5_TXOFF_INFO(v,
5131 		MLX5_TXOFF_CONFIG_VLAN |
5132 		MLX5_TXOFF_CONFIG_METADATA)
5133 
5134 MLX5_TXOFF_INFO(iv,
5135 		MLX5_TXOFF_CONFIG_INLINE | MLX5_TXOFF_CONFIG_VLAN |
5136 		MLX5_TXOFF_CONFIG_METADATA)
5137 
5138 MLX5_TXOFF_INFO(none_mpw,
5139 		MLX5_TXOFF_CONFIG_NONE | MLX5_TXOFF_CONFIG_EMPW |
5140 		MLX5_TXOFF_CONFIG_MPW)
5141 
5142 MLX5_TXOFF_INFO(mci_mpw,
5143 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_CSUM |
5144 		MLX5_TXOFF_CONFIG_INLINE | MLX5_TXOFF_CONFIG_EMPW |
5145 		MLX5_TXOFF_CONFIG_MPW)
5146 
5147 MLX5_TXOFF_INFO(mc_mpw,
5148 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_CSUM |
5149 		MLX5_TXOFF_CONFIG_EMPW | MLX5_TXOFF_CONFIG_MPW)
5150 
5151 MLX5_TXOFF_INFO(i_mpw,
5152 		MLX5_TXOFF_CONFIG_INLINE | MLX5_TXOFF_CONFIG_EMPW |
5153 		MLX5_TXOFF_CONFIG_MPW)
5154 };
5155 
5156 /**
5157  * Configure the Tx function to use. The routine checks configured
5158  * Tx offloads for the device and selects appropriate Tx burst
5159  * routine. There are multiple Tx burst routines compiled from
5160  * the same template in the most optimal way for the dedicated
5161  * Tx offloads set.
5162  *
5163  * @param dev
5164  *   Pointer to private data structure.
5165  *
5166  * @return
5167  *   Pointer to selected Tx burst function.
5168  */
5169 eth_tx_burst_t
5170 mlx5_select_tx_function(struct rte_eth_dev *dev)
5171 {
5172 	struct mlx5_priv *priv = dev->data->dev_private;
5173 	struct mlx5_dev_config *config = &priv->config;
5174 	uint64_t tx_offloads = dev->data->dev_conf.txmode.offloads;
5175 	unsigned int diff = 0, olx = 0, i, m;
5176 
5177 	static_assert(MLX5_WQE_SIZE_MAX / MLX5_WSEG_SIZE <=
5178 		      MLX5_DSEG_MAX, "invalid WQE max size");
5179 	static_assert(MLX5_WQE_CSEG_SIZE == MLX5_WSEG_SIZE,
5180 		      "invalid WQE Control Segment size");
5181 	static_assert(MLX5_WQE_ESEG_SIZE == MLX5_WSEG_SIZE,
5182 		      "invalid WQE Ethernet Segment size");
5183 	static_assert(MLX5_WQE_DSEG_SIZE == MLX5_WSEG_SIZE,
5184 		      "invalid WQE Data Segment size");
5185 	static_assert(MLX5_WQE_SIZE == 4 * MLX5_WSEG_SIZE,
5186 		      "invalid WQE size");
5187 	assert(priv);
5188 	if (tx_offloads & DEV_TX_OFFLOAD_MULTI_SEGS) {
5189 		/* We should support Multi-Segment Packets. */
5190 		olx |= MLX5_TXOFF_CONFIG_MULTI;
5191 	}
5192 	if (tx_offloads & (DEV_TX_OFFLOAD_TCP_TSO |
5193 			   DEV_TX_OFFLOAD_VXLAN_TNL_TSO |
5194 			   DEV_TX_OFFLOAD_GRE_TNL_TSO |
5195 			   DEV_TX_OFFLOAD_IP_TNL_TSO |
5196 			   DEV_TX_OFFLOAD_UDP_TNL_TSO)) {
5197 		/* We should support TCP Send Offload. */
5198 		olx |= MLX5_TXOFF_CONFIG_TSO;
5199 	}
5200 	if (tx_offloads & (DEV_TX_OFFLOAD_IP_TNL_TSO |
5201 			   DEV_TX_OFFLOAD_UDP_TNL_TSO |
5202 			   DEV_TX_OFFLOAD_OUTER_IPV4_CKSUM)) {
5203 		/* We should support Software Parser for Tunnels. */
5204 		olx |= MLX5_TXOFF_CONFIG_SWP;
5205 	}
5206 	if (tx_offloads & (DEV_TX_OFFLOAD_IPV4_CKSUM |
5207 			   DEV_TX_OFFLOAD_UDP_CKSUM |
5208 			   DEV_TX_OFFLOAD_TCP_CKSUM |
5209 			   DEV_TX_OFFLOAD_OUTER_IPV4_CKSUM)) {
5210 		/* We should support IP/TCP/UDP Checksums. */
5211 		olx |= MLX5_TXOFF_CONFIG_CSUM;
5212 	}
5213 	if (tx_offloads & DEV_TX_OFFLOAD_VLAN_INSERT) {
5214 		/* We should support VLAN insertion. */
5215 		olx |= MLX5_TXOFF_CONFIG_VLAN;
5216 	}
5217 	if (priv->txqs_n && (*priv->txqs)[0]) {
5218 		struct mlx5_txq_data *txd = (*priv->txqs)[0];
5219 
5220 		if (txd->inlen_send) {
5221 			/*
5222 			 * Check the data inline requirements. Data inline
5223 			 * is enabled on per device basis, we can check
5224 			 * the first Tx queue only.
5225 			 *
5226 			 * If device does not support VLAN insertion in WQE
5227 			 * and some queues are requested to perform VLAN
5228 			 * insertion offload than inline must be enabled.
5229 			 */
5230 			olx |= MLX5_TXOFF_CONFIG_INLINE;
5231 		}
5232 	}
5233 	if (config->mps == MLX5_MPW_ENHANCED &&
5234 	    config->txq_inline_min <= 0) {
5235 		/*
5236 		 * The NIC supports Enhanced Multi-Packet Write
5237 		 * and does not require minimal inline data.
5238 		 */
5239 		olx |= MLX5_TXOFF_CONFIG_EMPW;
5240 	}
5241 	if (rte_flow_dynf_metadata_avail()) {
5242 		/* We should support Flow metadata. */
5243 		olx |= MLX5_TXOFF_CONFIG_METADATA;
5244 	}
5245 	if (config->mps == MLX5_MPW) {
5246 		/*
5247 		 * The NIC supports Legacy Multi-Packet Write.
5248 		 * The MLX5_TXOFF_CONFIG_MPW controls the
5249 		 * descriptor building method in combination
5250 		 * with MLX5_TXOFF_CONFIG_EMPW.
5251 		 */
5252 		if (!(olx & (MLX5_TXOFF_CONFIG_TSO |
5253 			     MLX5_TXOFF_CONFIG_SWP |
5254 			     MLX5_TXOFF_CONFIG_VLAN |
5255 			     MLX5_TXOFF_CONFIG_METADATA)))
5256 			olx |= MLX5_TXOFF_CONFIG_EMPW |
5257 			       MLX5_TXOFF_CONFIG_MPW;
5258 	}
5259 	/*
5260 	 * Scan the routines table to find the minimal
5261 	 * satisfying routine with requested offloads.
5262 	 */
5263 	m = RTE_DIM(txoff_func);
5264 	for (i = 0; i < RTE_DIM(txoff_func); i++) {
5265 		unsigned int tmp;
5266 
5267 		tmp = txoff_func[i].olx;
5268 		if (tmp == olx) {
5269 			/* Meets requested offloads exactly.*/
5270 			m = i;
5271 			break;
5272 		}
5273 		if ((tmp & olx) != olx) {
5274 			/* Does not meet requested offloads at all. */
5275 			continue;
5276 		}
5277 		if ((olx ^ tmp) & MLX5_TXOFF_CONFIG_EMPW)
5278 			/* Do not enable eMPW if not configured. */
5279 			continue;
5280 		if ((olx ^ tmp) & MLX5_TXOFF_CONFIG_INLINE)
5281 			/* Do not enable inlining if not configured. */
5282 			continue;
5283 		/*
5284 		 * Some routine meets the requirements.
5285 		 * Check whether it has minimal amount
5286 		 * of not requested offloads.
5287 		 */
5288 		tmp = __builtin_popcountl(tmp & ~olx);
5289 		if (m >= RTE_DIM(txoff_func) || tmp < diff) {
5290 			/* First or better match, save and continue. */
5291 			m = i;
5292 			diff = tmp;
5293 			continue;
5294 		}
5295 		if (tmp == diff) {
5296 			tmp = txoff_func[i].olx ^ txoff_func[m].olx;
5297 			if (__builtin_ffsl(txoff_func[i].olx & ~tmp) <
5298 			    __builtin_ffsl(txoff_func[m].olx & ~tmp)) {
5299 				/* Lighter not requested offload. */
5300 				m = i;
5301 			}
5302 		}
5303 	}
5304 	if (m >= RTE_DIM(txoff_func)) {
5305 		DRV_LOG(DEBUG, "port %u has no selected Tx function"
5306 			       " for requested offloads %04X",
5307 				dev->data->port_id, olx);
5308 		return NULL;
5309 	}
5310 	DRV_LOG(DEBUG, "port %u has selected Tx function"
5311 		       " supporting offloads %04X/%04X",
5312 			dev->data->port_id, olx, txoff_func[m].olx);
5313 	if (txoff_func[m].olx & MLX5_TXOFF_CONFIG_MULTI)
5314 		DRV_LOG(DEBUG, "\tMULTI (multi segment)");
5315 	if (txoff_func[m].olx & MLX5_TXOFF_CONFIG_TSO)
5316 		DRV_LOG(DEBUG, "\tTSO   (TCP send offload)");
5317 	if (txoff_func[m].olx & MLX5_TXOFF_CONFIG_SWP)
5318 		DRV_LOG(DEBUG, "\tSWP   (software parser)");
5319 	if (txoff_func[m].olx & MLX5_TXOFF_CONFIG_CSUM)
5320 		DRV_LOG(DEBUG, "\tCSUM  (checksum offload)");
5321 	if (txoff_func[m].olx & MLX5_TXOFF_CONFIG_INLINE)
5322 		DRV_LOG(DEBUG, "\tINLIN (inline data)");
5323 	if (txoff_func[m].olx & MLX5_TXOFF_CONFIG_VLAN)
5324 		DRV_LOG(DEBUG, "\tVLANI (VLAN insertion)");
5325 	if (txoff_func[m].olx & MLX5_TXOFF_CONFIG_METADATA)
5326 		DRV_LOG(DEBUG, "\tMETAD (tx Flow metadata)");
5327 	if (txoff_func[m].olx & MLX5_TXOFF_CONFIG_EMPW) {
5328 		if (txoff_func[m].olx & MLX5_TXOFF_CONFIG_MPW)
5329 			DRV_LOG(DEBUG, "\tMPW   (Legacy MPW)");
5330 		else
5331 			DRV_LOG(DEBUG, "\tEMPW  (Enhanced MPW)");
5332 	}
5333 	return txoff_func[m].func;
5334 }
5335