xref: /dpdk/drivers/net/mlx5/mlx5_rxtx.c (revision 7adf992fb9bf7162a7edc45b50d10fbb1d57824d)
1 /* SPDX-License-Identifier: BSD-3-Clause
2  * Copyright 2015 6WIND S.A.
3  * Copyright 2015-2019 Mellanox Technologies, Ltd
4  */
5 
6 #include <stdint.h>
7 #include <string.h>
8 #include <stdlib.h>
9 
10 /* Verbs header. */
11 /* ISO C doesn't support unnamed structs/unions, disabling -pedantic. */
12 #ifdef PEDANTIC
13 #pragma GCC diagnostic ignored "-Wpedantic"
14 #endif
15 #include <infiniband/verbs.h>
16 #include <infiniband/mlx5dv.h>
17 #ifdef PEDANTIC
18 #pragma GCC diagnostic error "-Wpedantic"
19 #endif
20 
21 #include <rte_mbuf.h>
22 #include <rte_mempool.h>
23 #include <rte_prefetch.h>
24 #include <rte_common.h>
25 #include <rte_branch_prediction.h>
26 #include <rte_ether.h>
27 #include <rte_cycles.h>
28 #include <rte_flow.h>
29 
30 #include <mlx5_devx_cmds.h>
31 #include <mlx5_prm.h>
32 #include <mlx5_common.h>
33 
34 #include "mlx5_defs.h"
35 #include "mlx5.h"
36 #include "mlx5_utils.h"
37 #include "mlx5_rxtx.h"
38 #include "mlx5_autoconf.h"
39 
40 /* TX burst subroutines return codes. */
41 enum mlx5_txcmp_code {
42 	MLX5_TXCMP_CODE_EXIT = 0,
43 	MLX5_TXCMP_CODE_ERROR,
44 	MLX5_TXCMP_CODE_SINGLE,
45 	MLX5_TXCMP_CODE_MULTI,
46 	MLX5_TXCMP_CODE_TSO,
47 	MLX5_TXCMP_CODE_EMPW,
48 };
49 
50 /*
51  * These defines are used to configure Tx burst routine option set
52  * supported at compile time. The not specified options are optimized out
53  * out due to if conditions can be explicitly calculated at compile time.
54  * The offloads with bigger runtime check (require more CPU cycles to
55  * skip) overhead should have the bigger index - this is needed to
56  * select the better matching routine function if no exact match and
57  * some offloads are not actually requested.
58  */
59 #define MLX5_TXOFF_CONFIG_MULTI (1u << 0) /* Multi-segment packets.*/
60 #define MLX5_TXOFF_CONFIG_TSO (1u << 1) /* TCP send offload supported.*/
61 #define MLX5_TXOFF_CONFIG_SWP (1u << 2) /* Tunnels/SW Parser offloads.*/
62 #define MLX5_TXOFF_CONFIG_CSUM (1u << 3) /* Check Sums offloaded. */
63 #define MLX5_TXOFF_CONFIG_INLINE (1u << 4) /* Data inlining supported. */
64 #define MLX5_TXOFF_CONFIG_VLAN (1u << 5) /* VLAN insertion supported.*/
65 #define MLX5_TXOFF_CONFIG_METADATA (1u << 6) /* Flow metadata. */
66 #define MLX5_TXOFF_CONFIG_EMPW (1u << 8) /* Enhanced MPW supported.*/
67 #define MLX5_TXOFF_CONFIG_MPW (1u << 9) /* Legacy MPW supported.*/
68 
69 /* The most common offloads groups. */
70 #define MLX5_TXOFF_CONFIG_NONE 0
71 #define MLX5_TXOFF_CONFIG_FULL (MLX5_TXOFF_CONFIG_MULTI | \
72 				MLX5_TXOFF_CONFIG_TSO | \
73 				MLX5_TXOFF_CONFIG_SWP | \
74 				MLX5_TXOFF_CONFIG_CSUM | \
75 				MLX5_TXOFF_CONFIG_INLINE | \
76 				MLX5_TXOFF_CONFIG_VLAN | \
77 				MLX5_TXOFF_CONFIG_METADATA)
78 
79 #define MLX5_TXOFF_CONFIG(mask) (olx & MLX5_TXOFF_CONFIG_##mask)
80 
81 #define MLX5_TXOFF_DECL(func, olx) \
82 static uint16_t mlx5_tx_burst_##func(void *txq, \
83 				     struct rte_mbuf **pkts, \
84 				    uint16_t pkts_n) \
85 { \
86 	return mlx5_tx_burst_tmpl((struct mlx5_txq_data *)txq, \
87 		    pkts, pkts_n, (olx)); \
88 }
89 
90 #define MLX5_TXOFF_INFO(func, olx) {mlx5_tx_burst_##func, olx},
91 
92 static __rte_always_inline uint32_t
93 rxq_cq_to_pkt_type(struct mlx5_rxq_data *rxq, volatile struct mlx5_cqe *cqe);
94 
95 static __rte_always_inline int
96 mlx5_rx_poll_len(struct mlx5_rxq_data *rxq, volatile struct mlx5_cqe *cqe,
97 		 uint16_t cqe_cnt, volatile struct mlx5_mini_cqe8 **mcqe);
98 
99 static __rte_always_inline uint32_t
100 rxq_cq_to_ol_flags(volatile struct mlx5_cqe *cqe);
101 
102 static __rte_always_inline void
103 rxq_cq_to_mbuf(struct mlx5_rxq_data *rxq, struct rte_mbuf *pkt,
104 	       volatile struct mlx5_cqe *cqe, uint32_t rss_hash_res);
105 
106 static __rte_always_inline void
107 mprq_buf_replace(struct mlx5_rxq_data *rxq, uint16_t rq_idx,
108 		 const unsigned int strd_n);
109 
110 static int
111 mlx5_queue_state_modify(struct rte_eth_dev *dev,
112 			struct mlx5_mp_arg_queue_state_modify *sm);
113 
114 static inline void
115 mlx5_lro_update_tcp_hdr(struct rte_tcp_hdr *restrict tcp,
116 			volatile struct mlx5_cqe *restrict cqe,
117 			uint32_t phcsum);
118 
119 static inline void
120 mlx5_lro_update_hdr(uint8_t *restrict padd,
121 		    volatile struct mlx5_cqe *restrict cqe,
122 		    uint32_t len);
123 
124 uint32_t mlx5_ptype_table[] __rte_cache_aligned = {
125 	[0xff] = RTE_PTYPE_ALL_MASK, /* Last entry for errored packet. */
126 };
127 
128 uint8_t mlx5_cksum_table[1 << 10] __rte_cache_aligned;
129 uint8_t mlx5_swp_types_table[1 << 10] __rte_cache_aligned;
130 
131 uint64_t rte_net_mlx5_dynf_inline_mask;
132 #define PKT_TX_DYNF_NOINLINE rte_net_mlx5_dynf_inline_mask
133 
134 /**
135  * Build a table to translate Rx completion flags to packet type.
136  *
137  * @note: fix mlx5_dev_supported_ptypes_get() if any change here.
138  */
139 void
140 mlx5_set_ptype_table(void)
141 {
142 	unsigned int i;
143 	uint32_t (*p)[RTE_DIM(mlx5_ptype_table)] = &mlx5_ptype_table;
144 
145 	/* Last entry must not be overwritten, reserved for errored packet. */
146 	for (i = 0; i < RTE_DIM(mlx5_ptype_table) - 1; ++i)
147 		(*p)[i] = RTE_PTYPE_UNKNOWN;
148 	/*
149 	 * The index to the array should have:
150 	 * bit[1:0] = l3_hdr_type
151 	 * bit[4:2] = l4_hdr_type
152 	 * bit[5] = ip_frag
153 	 * bit[6] = tunneled
154 	 * bit[7] = outer_l3_type
155 	 */
156 	/* L2 */
157 	(*p)[0x00] = RTE_PTYPE_L2_ETHER;
158 	/* L3 */
159 	(*p)[0x01] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
160 		     RTE_PTYPE_L4_NONFRAG;
161 	(*p)[0x02] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
162 		     RTE_PTYPE_L4_NONFRAG;
163 	/* Fragmented */
164 	(*p)[0x21] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
165 		     RTE_PTYPE_L4_FRAG;
166 	(*p)[0x22] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
167 		     RTE_PTYPE_L4_FRAG;
168 	/* TCP */
169 	(*p)[0x05] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
170 		     RTE_PTYPE_L4_TCP;
171 	(*p)[0x06] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
172 		     RTE_PTYPE_L4_TCP;
173 	(*p)[0x0d] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
174 		     RTE_PTYPE_L4_TCP;
175 	(*p)[0x0e] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
176 		     RTE_PTYPE_L4_TCP;
177 	(*p)[0x11] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
178 		     RTE_PTYPE_L4_TCP;
179 	(*p)[0x12] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
180 		     RTE_PTYPE_L4_TCP;
181 	/* UDP */
182 	(*p)[0x09] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
183 		     RTE_PTYPE_L4_UDP;
184 	(*p)[0x0a] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
185 		     RTE_PTYPE_L4_UDP;
186 	/* Repeat with outer_l3_type being set. Just in case. */
187 	(*p)[0x81] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
188 		     RTE_PTYPE_L4_NONFRAG;
189 	(*p)[0x82] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
190 		     RTE_PTYPE_L4_NONFRAG;
191 	(*p)[0xa1] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
192 		     RTE_PTYPE_L4_FRAG;
193 	(*p)[0xa2] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
194 		     RTE_PTYPE_L4_FRAG;
195 	(*p)[0x85] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
196 		     RTE_PTYPE_L4_TCP;
197 	(*p)[0x86] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
198 		     RTE_PTYPE_L4_TCP;
199 	(*p)[0x8d] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
200 		     RTE_PTYPE_L4_TCP;
201 	(*p)[0x8e] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
202 		     RTE_PTYPE_L4_TCP;
203 	(*p)[0x91] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
204 		     RTE_PTYPE_L4_TCP;
205 	(*p)[0x92] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
206 		     RTE_PTYPE_L4_TCP;
207 	(*p)[0x89] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
208 		     RTE_PTYPE_L4_UDP;
209 	(*p)[0x8a] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
210 		     RTE_PTYPE_L4_UDP;
211 	/* Tunneled - L3 */
212 	(*p)[0x40] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN;
213 	(*p)[0x41] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
214 		     RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN |
215 		     RTE_PTYPE_INNER_L4_NONFRAG;
216 	(*p)[0x42] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
217 		     RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN |
218 		     RTE_PTYPE_INNER_L4_NONFRAG;
219 	(*p)[0xc0] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN;
220 	(*p)[0xc1] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
221 		     RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN |
222 		     RTE_PTYPE_INNER_L4_NONFRAG;
223 	(*p)[0xc2] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
224 		     RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN |
225 		     RTE_PTYPE_INNER_L4_NONFRAG;
226 	/* Tunneled - Fragmented */
227 	(*p)[0x61] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
228 		     RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN |
229 		     RTE_PTYPE_INNER_L4_FRAG;
230 	(*p)[0x62] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
231 		     RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN |
232 		     RTE_PTYPE_INNER_L4_FRAG;
233 	(*p)[0xe1] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
234 		     RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN |
235 		     RTE_PTYPE_INNER_L4_FRAG;
236 	(*p)[0xe2] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
237 		     RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN |
238 		     RTE_PTYPE_INNER_L4_FRAG;
239 	/* Tunneled - TCP */
240 	(*p)[0x45] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
241 		     RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN |
242 		     RTE_PTYPE_INNER_L4_TCP;
243 	(*p)[0x46] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
244 		     RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN |
245 		     RTE_PTYPE_INNER_L4_TCP;
246 	(*p)[0x4d] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
247 		     RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN |
248 		     RTE_PTYPE_INNER_L4_TCP;
249 	(*p)[0x4e] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
250 		     RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN |
251 		     RTE_PTYPE_INNER_L4_TCP;
252 	(*p)[0x51] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
253 		     RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN |
254 		     RTE_PTYPE_INNER_L4_TCP;
255 	(*p)[0x52] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
256 		     RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN |
257 		     RTE_PTYPE_INNER_L4_TCP;
258 	(*p)[0xc5] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
259 		     RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN |
260 		     RTE_PTYPE_INNER_L4_TCP;
261 	(*p)[0xc6] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
262 		     RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN |
263 		     RTE_PTYPE_INNER_L4_TCP;
264 	(*p)[0xcd] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
265 		     RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN |
266 		     RTE_PTYPE_INNER_L4_TCP;
267 	(*p)[0xce] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
268 		     RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN |
269 		     RTE_PTYPE_INNER_L4_TCP;
270 	(*p)[0xd1] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
271 		     RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN |
272 		     RTE_PTYPE_INNER_L4_TCP;
273 	(*p)[0xd2] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
274 		     RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN |
275 		     RTE_PTYPE_INNER_L4_TCP;
276 	/* Tunneled - UDP */
277 	(*p)[0x49] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
278 		     RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN |
279 		     RTE_PTYPE_INNER_L4_UDP;
280 	(*p)[0x4a] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
281 		     RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN |
282 		     RTE_PTYPE_INNER_L4_UDP;
283 	(*p)[0xc9] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
284 		     RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN |
285 		     RTE_PTYPE_INNER_L4_UDP;
286 	(*p)[0xca] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
287 		     RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN |
288 		     RTE_PTYPE_INNER_L4_UDP;
289 }
290 
291 /**
292  * Build a table to translate packet to checksum type of Verbs.
293  */
294 void
295 mlx5_set_cksum_table(void)
296 {
297 	unsigned int i;
298 	uint8_t v;
299 
300 	/*
301 	 * The index should have:
302 	 * bit[0] = PKT_TX_TCP_SEG
303 	 * bit[2:3] = PKT_TX_UDP_CKSUM, PKT_TX_TCP_CKSUM
304 	 * bit[4] = PKT_TX_IP_CKSUM
305 	 * bit[8] = PKT_TX_OUTER_IP_CKSUM
306 	 * bit[9] = tunnel
307 	 */
308 	for (i = 0; i < RTE_DIM(mlx5_cksum_table); ++i) {
309 		v = 0;
310 		if (i & (1 << 9)) {
311 			/* Tunneled packet. */
312 			if (i & (1 << 8)) /* Outer IP. */
313 				v |= MLX5_ETH_WQE_L3_CSUM;
314 			if (i & (1 << 4)) /* Inner IP. */
315 				v |= MLX5_ETH_WQE_L3_INNER_CSUM;
316 			if (i & (3 << 2 | 1 << 0)) /* L4 or TSO. */
317 				v |= MLX5_ETH_WQE_L4_INNER_CSUM;
318 		} else {
319 			/* No tunnel. */
320 			if (i & (1 << 4)) /* IP. */
321 				v |= MLX5_ETH_WQE_L3_CSUM;
322 			if (i & (3 << 2 | 1 << 0)) /* L4 or TSO. */
323 				v |= MLX5_ETH_WQE_L4_CSUM;
324 		}
325 		mlx5_cksum_table[i] = v;
326 	}
327 }
328 
329 /**
330  * Build a table to translate packet type of mbuf to SWP type of Verbs.
331  */
332 void
333 mlx5_set_swp_types_table(void)
334 {
335 	unsigned int i;
336 	uint8_t v;
337 
338 	/*
339 	 * The index should have:
340 	 * bit[0:1] = PKT_TX_L4_MASK
341 	 * bit[4] = PKT_TX_IPV6
342 	 * bit[8] = PKT_TX_OUTER_IPV6
343 	 * bit[9] = PKT_TX_OUTER_UDP
344 	 */
345 	for (i = 0; i < RTE_DIM(mlx5_swp_types_table); ++i) {
346 		v = 0;
347 		if (i & (1 << 8))
348 			v |= MLX5_ETH_WQE_L3_OUTER_IPV6;
349 		if (i & (1 << 9))
350 			v |= MLX5_ETH_WQE_L4_OUTER_UDP;
351 		if (i & (1 << 4))
352 			v |= MLX5_ETH_WQE_L3_INNER_IPV6;
353 		if ((i & 3) == (PKT_TX_UDP_CKSUM >> 52))
354 			v |= MLX5_ETH_WQE_L4_INNER_UDP;
355 		mlx5_swp_types_table[i] = v;
356 	}
357 }
358 
359 /**
360  * Set Software Parser flags and offsets in Ethernet Segment of WQE.
361  * Flags must be preliminary initialized to zero.
362  *
363  * @param loc
364  *   Pointer to burst routine local context.
365  * @param swp_flags
366  *   Pointer to store Software Parser flags
367  * @param olx
368  *   Configured Tx offloads mask. It is fully defined at
369  *   compile time and may be used for optimization.
370  *
371  * @return
372  *   Software Parser offsets packed in dword.
373  *   Software Parser flags are set by pointer.
374  */
375 static __rte_always_inline uint32_t
376 txq_mbuf_to_swp(struct mlx5_txq_local *restrict loc,
377 		uint8_t *swp_flags,
378 		unsigned int olx)
379 {
380 	uint64_t ol, tunnel;
381 	unsigned int idx, off;
382 	uint32_t set;
383 
384 	if (!MLX5_TXOFF_CONFIG(SWP))
385 		return 0;
386 	ol = loc->mbuf->ol_flags;
387 	tunnel = ol & PKT_TX_TUNNEL_MASK;
388 	/*
389 	 * Check whether Software Parser is required.
390 	 * Only customized tunnels may ask for.
391 	 */
392 	if (likely(tunnel != PKT_TX_TUNNEL_UDP && tunnel != PKT_TX_TUNNEL_IP))
393 		return 0;
394 	/*
395 	 * The index should have:
396 	 * bit[0:1] = PKT_TX_L4_MASK
397 	 * bit[4] = PKT_TX_IPV6
398 	 * bit[8] = PKT_TX_OUTER_IPV6
399 	 * bit[9] = PKT_TX_OUTER_UDP
400 	 */
401 	idx = (ol & (PKT_TX_L4_MASK | PKT_TX_IPV6 | PKT_TX_OUTER_IPV6)) >> 52;
402 	idx |= (tunnel == PKT_TX_TUNNEL_UDP) ? (1 << 9) : 0;
403 	*swp_flags = mlx5_swp_types_table[idx];
404 	/*
405 	 * Set offsets for SW parser. Since ConnectX-5, SW parser just
406 	 * complements HW parser. SW parser starts to engage only if HW parser
407 	 * can't reach a header. For the older devices, HW parser will not kick
408 	 * in if any of SWP offsets is set. Therefore, all of the L3 offsets
409 	 * should be set regardless of HW offload.
410 	 */
411 	off = loc->mbuf->outer_l2_len;
412 	if (MLX5_TXOFF_CONFIG(VLAN) && ol & PKT_TX_VLAN_PKT)
413 		off += sizeof(struct rte_vlan_hdr);
414 	set = (off >> 1) << 8; /* Outer L3 offset. */
415 	off += loc->mbuf->outer_l3_len;
416 	if (tunnel == PKT_TX_TUNNEL_UDP)
417 		set |= off >> 1; /* Outer L4 offset. */
418 	if (ol & (PKT_TX_IPV4 | PKT_TX_IPV6)) { /* Inner IP. */
419 		const uint64_t csum = ol & PKT_TX_L4_MASK;
420 			off += loc->mbuf->l2_len;
421 		set |= (off >> 1) << 24; /* Inner L3 offset. */
422 		if (csum == PKT_TX_TCP_CKSUM ||
423 		    csum == PKT_TX_UDP_CKSUM ||
424 		    (MLX5_TXOFF_CONFIG(TSO) && ol & PKT_TX_TCP_SEG)) {
425 			off += loc->mbuf->l3_len;
426 			set |= (off >> 1) << 16; /* Inner L4 offset. */
427 		}
428 	}
429 	set = rte_cpu_to_le_32(set);
430 	return set;
431 }
432 
433 /**
434  * Convert the Checksum offloads to Verbs.
435  *
436  * @param buf
437  *   Pointer to the mbuf.
438  *
439  * @return
440  *   Converted checksum flags.
441  */
442 static __rte_always_inline uint8_t
443 txq_ol_cksum_to_cs(struct rte_mbuf *buf)
444 {
445 	uint32_t idx;
446 	uint8_t is_tunnel = !!(buf->ol_flags & PKT_TX_TUNNEL_MASK);
447 	const uint64_t ol_flags_mask = PKT_TX_TCP_SEG | PKT_TX_L4_MASK |
448 				       PKT_TX_IP_CKSUM | PKT_TX_OUTER_IP_CKSUM;
449 
450 	/*
451 	 * The index should have:
452 	 * bit[0] = PKT_TX_TCP_SEG
453 	 * bit[2:3] = PKT_TX_UDP_CKSUM, PKT_TX_TCP_CKSUM
454 	 * bit[4] = PKT_TX_IP_CKSUM
455 	 * bit[8] = PKT_TX_OUTER_IP_CKSUM
456 	 * bit[9] = tunnel
457 	 */
458 	idx = ((buf->ol_flags & ol_flags_mask) >> 50) | (!!is_tunnel << 9);
459 	return mlx5_cksum_table[idx];
460 }
461 
462 /**
463  * Internal function to compute the number of used descriptors in an RX queue
464  *
465  * @param rxq
466  *   The Rx queue.
467  *
468  * @return
469  *   The number of used rx descriptor.
470  */
471 static uint32_t
472 rx_queue_count(struct mlx5_rxq_data *rxq)
473 {
474 	struct rxq_zip *zip = &rxq->zip;
475 	volatile struct mlx5_cqe *cqe;
476 	const unsigned int cqe_n = (1 << rxq->cqe_n);
477 	const unsigned int cqe_cnt = cqe_n - 1;
478 	unsigned int cq_ci;
479 	unsigned int used;
480 
481 	/* if we are processing a compressed cqe */
482 	if (zip->ai) {
483 		used = zip->cqe_cnt - zip->ca;
484 		cq_ci = zip->cq_ci;
485 	} else {
486 		used = 0;
487 		cq_ci = rxq->cq_ci;
488 	}
489 	cqe = &(*rxq->cqes)[cq_ci & cqe_cnt];
490 	while (check_cqe(cqe, cqe_n, cq_ci) != MLX5_CQE_STATUS_HW_OWN) {
491 		int8_t op_own;
492 		unsigned int n;
493 
494 		op_own = cqe->op_own;
495 		if (MLX5_CQE_FORMAT(op_own) == MLX5_COMPRESSED)
496 			n = rte_be_to_cpu_32(cqe->byte_cnt);
497 		else
498 			n = 1;
499 		cq_ci += n;
500 		used += n;
501 		cqe = &(*rxq->cqes)[cq_ci & cqe_cnt];
502 	}
503 	used = RTE_MIN(used, (1U << rxq->elts_n) - 1);
504 	return used;
505 }
506 
507 /**
508  * DPDK callback to check the status of a rx descriptor.
509  *
510  * @param rx_queue
511  *   The Rx queue.
512  * @param[in] offset
513  *   The index of the descriptor in the ring.
514  *
515  * @return
516  *   The status of the tx descriptor.
517  */
518 int
519 mlx5_rx_descriptor_status(void *rx_queue, uint16_t offset)
520 {
521 	struct mlx5_rxq_data *rxq = rx_queue;
522 	struct mlx5_rxq_ctrl *rxq_ctrl =
523 			container_of(rxq, struct mlx5_rxq_ctrl, rxq);
524 	struct rte_eth_dev *dev = ETH_DEV(rxq_ctrl->priv);
525 
526 	if (dev->rx_pkt_burst != mlx5_rx_burst) {
527 		rte_errno = ENOTSUP;
528 		return -rte_errno;
529 	}
530 	if (offset >= (1 << rxq->elts_n)) {
531 		rte_errno = EINVAL;
532 		return -rte_errno;
533 	}
534 	if (offset < rx_queue_count(rxq))
535 		return RTE_ETH_RX_DESC_DONE;
536 	return RTE_ETH_RX_DESC_AVAIL;
537 }
538 
539 /**
540  * DPDK callback to get the RX queue information
541  *
542  * @param dev
543  *   Pointer to the device structure.
544  *
545  * @param rx_queue_id
546  *   Rx queue identificator.
547  *
548  * @param qinfo
549  *   Pointer to the RX queue information structure.
550  *
551  * @return
552  *   None.
553  */
554 
555 void
556 mlx5_rxq_info_get(struct rte_eth_dev *dev, uint16_t rx_queue_id,
557 		  struct rte_eth_rxq_info *qinfo)
558 {
559 	struct mlx5_priv *priv = dev->data->dev_private;
560 	struct mlx5_rxq_data *rxq = (*priv->rxqs)[rx_queue_id];
561 	struct mlx5_rxq_ctrl *rxq_ctrl =
562 		container_of(rxq, struct mlx5_rxq_ctrl, rxq);
563 
564 	if (!rxq)
565 		return;
566 	qinfo->mp = mlx5_rxq_mprq_enabled(&rxq_ctrl->rxq) ?
567 					rxq->mprq_mp : rxq->mp;
568 	qinfo->conf.rx_thresh.pthresh = 0;
569 	qinfo->conf.rx_thresh.hthresh = 0;
570 	qinfo->conf.rx_thresh.wthresh = 0;
571 	qinfo->conf.rx_free_thresh = rxq->rq_repl_thresh;
572 	qinfo->conf.rx_drop_en = 1;
573 	qinfo->conf.rx_deferred_start = rxq_ctrl ? 0 : 1;
574 	qinfo->conf.offloads = dev->data->dev_conf.rxmode.offloads;
575 	qinfo->scattered_rx = dev->data->scattered_rx;
576 	qinfo->nb_desc = 1 << rxq->elts_n;
577 }
578 
579 /**
580  * DPDK callback to get the RX packet burst mode information
581  *
582  * @param dev
583  *   Pointer to the device structure.
584  *
585  * @param rx_queue_id
586  *   Rx queue identificatior.
587  *
588  * @param mode
589  *   Pointer to the burts mode information.
590  *
591  * @return
592  *   0 as success, -EINVAL as failure.
593  */
594 
595 int
596 mlx5_rx_burst_mode_get(struct rte_eth_dev *dev,
597 		       uint16_t rx_queue_id __rte_unused,
598 		       struct rte_eth_burst_mode *mode)
599 {
600 	eth_rx_burst_t pkt_burst = dev->rx_pkt_burst;
601 
602 	if (pkt_burst == mlx5_rx_burst) {
603 		snprintf(mode->info, sizeof(mode->info), "%s", "Scalar");
604 	} else if (pkt_burst == mlx5_rx_burst_mprq) {
605 		snprintf(mode->info, sizeof(mode->info), "%s", "Multi-Packet RQ");
606 	} else if (pkt_burst == mlx5_rx_burst_vec) {
607 #if defined RTE_ARCH_X86_64
608 		snprintf(mode->info, sizeof(mode->info), "%s", "Vector SSE");
609 #elif defined RTE_ARCH_ARM64
610 		snprintf(mode->info, sizeof(mode->info), "%s", "Vector Neon");
611 #elif defined RTE_ARCH_PPC_64
612 		snprintf(mode->info, sizeof(mode->info), "%s", "Vector AltiVec");
613 #else
614 		return -EINVAL;
615 #endif
616 	} else {
617 		return -EINVAL;
618 	}
619 	return 0;
620 }
621 
622 /**
623  * DPDK callback to get the number of used descriptors in a RX queue
624  *
625  * @param dev
626  *   Pointer to the device structure.
627  *
628  * @param rx_queue_id
629  *   The Rx queue.
630  *
631  * @return
632  *   The number of used rx descriptor.
633  *   -EINVAL if the queue is invalid
634  */
635 uint32_t
636 mlx5_rx_queue_count(struct rte_eth_dev *dev, uint16_t rx_queue_id)
637 {
638 	struct mlx5_priv *priv = dev->data->dev_private;
639 	struct mlx5_rxq_data *rxq;
640 
641 	if (dev->rx_pkt_burst != mlx5_rx_burst) {
642 		rte_errno = ENOTSUP;
643 		return -rte_errno;
644 	}
645 	rxq = (*priv->rxqs)[rx_queue_id];
646 	if (!rxq) {
647 		rte_errno = EINVAL;
648 		return -rte_errno;
649 	}
650 	return rx_queue_count(rxq);
651 }
652 
653 #define MLX5_SYSTEM_LOG_DIR "/var/log"
654 /**
655  * Dump debug information to log file.
656  *
657  * @param fname
658  *   The file name.
659  * @param hex_title
660  *   If not NULL this string is printed as a header to the output
661  *   and the output will be in hexadecimal view.
662  * @param buf
663  *   This is the buffer address to print out.
664  * @param len
665  *   The number of bytes to dump out.
666  */
667 void
668 mlx5_dump_debug_information(const char *fname, const char *hex_title,
669 			    const void *buf, unsigned int hex_len)
670 {
671 	FILE *fd;
672 
673 	MKSTR(path, "%s/%s", MLX5_SYSTEM_LOG_DIR, fname);
674 	fd = fopen(path, "a+");
675 	if (!fd) {
676 		DRV_LOG(WARNING, "cannot open %s for debug dump", path);
677 		MKSTR(path2, "./%s", fname);
678 		fd = fopen(path2, "a+");
679 		if (!fd) {
680 			DRV_LOG(ERR, "cannot open %s for debug dump", path2);
681 			return;
682 		}
683 		DRV_LOG(INFO, "New debug dump in file %s", path2);
684 	} else {
685 		DRV_LOG(INFO, "New debug dump in file %s", path);
686 	}
687 	if (hex_title)
688 		rte_hexdump(fd, hex_title, buf, hex_len);
689 	else
690 		fprintf(fd, "%s", (const char *)buf);
691 	fprintf(fd, "\n\n\n");
692 	fclose(fd);
693 }
694 
695 /**
696  * Move QP from error state to running state and initialize indexes.
697  *
698  * @param txq_ctrl
699  *   Pointer to TX queue control structure.
700  *
701  * @return
702  *   0 on success, else -1.
703  */
704 static int
705 tx_recover_qp(struct mlx5_txq_ctrl *txq_ctrl)
706 {
707 	struct mlx5_mp_arg_queue_state_modify sm = {
708 			.is_wq = 0,
709 			.queue_id = txq_ctrl->txq.idx,
710 	};
711 
712 	if (mlx5_queue_state_modify(ETH_DEV(txq_ctrl->priv), &sm))
713 		return -1;
714 	txq_ctrl->txq.wqe_ci = 0;
715 	txq_ctrl->txq.wqe_pi = 0;
716 	txq_ctrl->txq.elts_comp = 0;
717 	return 0;
718 }
719 
720 /* Return 1 if the error CQE is signed otherwise, sign it and return 0. */
721 static int
722 check_err_cqe_seen(volatile struct mlx5_err_cqe *err_cqe)
723 {
724 	static const uint8_t magic[] = "seen";
725 	int ret = 1;
726 	unsigned int i;
727 
728 	for (i = 0; i < sizeof(magic); ++i)
729 		if (!ret || err_cqe->rsvd1[i] != magic[i]) {
730 			ret = 0;
731 			err_cqe->rsvd1[i] = magic[i];
732 		}
733 	return ret;
734 }
735 
736 /**
737  * Handle error CQE.
738  *
739  * @param txq
740  *   Pointer to TX queue structure.
741  * @param error_cqe
742  *   Pointer to the error CQE.
743  *
744  * @return
745  *   Negative value if queue recovery failed, otherwise
746  *   the error completion entry is handled successfully.
747  */
748 static int
749 mlx5_tx_error_cqe_handle(struct mlx5_txq_data *restrict txq,
750 			 volatile struct mlx5_err_cqe *err_cqe)
751 {
752 	if (err_cqe->syndrome != MLX5_CQE_SYNDROME_WR_FLUSH_ERR) {
753 		const uint16_t wqe_m = ((1 << txq->wqe_n) - 1);
754 		struct mlx5_txq_ctrl *txq_ctrl =
755 				container_of(txq, struct mlx5_txq_ctrl, txq);
756 		uint16_t new_wqe_pi = rte_be_to_cpu_16(err_cqe->wqe_counter);
757 		int seen = check_err_cqe_seen(err_cqe);
758 
759 		if (!seen && txq_ctrl->dump_file_n <
760 		    txq_ctrl->priv->config.max_dump_files_num) {
761 			MKSTR(err_str, "Unexpected CQE error syndrome "
762 			      "0x%02x CQN = %u SQN = %u wqe_counter = %u "
763 			      "wq_ci = %u cq_ci = %u", err_cqe->syndrome,
764 			      txq->cqe_s, txq->qp_num_8s >> 8,
765 			      rte_be_to_cpu_16(err_cqe->wqe_counter),
766 			      txq->wqe_ci, txq->cq_ci);
767 			MKSTR(name, "dpdk_mlx5_port_%u_txq_%u_index_%u_%u",
768 			      PORT_ID(txq_ctrl->priv), txq->idx,
769 			      txq_ctrl->dump_file_n, (uint32_t)rte_rdtsc());
770 			mlx5_dump_debug_information(name, NULL, err_str, 0);
771 			mlx5_dump_debug_information(name, "MLX5 Error CQ:",
772 						    (const void *)((uintptr_t)
773 						    txq->cqes),
774 						    sizeof(*err_cqe) *
775 						    (1 << txq->cqe_n));
776 			mlx5_dump_debug_information(name, "MLX5 Error SQ:",
777 						    (const void *)((uintptr_t)
778 						    txq->wqes),
779 						    MLX5_WQE_SIZE *
780 						    (1 << txq->wqe_n));
781 			txq_ctrl->dump_file_n++;
782 		}
783 		if (!seen)
784 			/*
785 			 * Count errors in WQEs units.
786 			 * Later it can be improved to count error packets,
787 			 * for example, by SQ parsing to find how much packets
788 			 * should be counted for each WQE.
789 			 */
790 			txq->stats.oerrors += ((txq->wqe_ci & wqe_m) -
791 						new_wqe_pi) & wqe_m;
792 		if (tx_recover_qp(txq_ctrl)) {
793 			/* Recovering failed - retry later on the same WQE. */
794 			return -1;
795 		}
796 		/* Release all the remaining buffers. */
797 		txq_free_elts(txq_ctrl);
798 	}
799 	return 0;
800 }
801 
802 /**
803  * Translate RX completion flags to packet type.
804  *
805  * @param[in] rxq
806  *   Pointer to RX queue structure.
807  * @param[in] cqe
808  *   Pointer to CQE.
809  *
810  * @note: fix mlx5_dev_supported_ptypes_get() if any change here.
811  *
812  * @return
813  *   Packet type for struct rte_mbuf.
814  */
815 static inline uint32_t
816 rxq_cq_to_pkt_type(struct mlx5_rxq_data *rxq, volatile struct mlx5_cqe *cqe)
817 {
818 	uint8_t idx;
819 	uint8_t pinfo = cqe->pkt_info;
820 	uint16_t ptype = cqe->hdr_type_etc;
821 
822 	/*
823 	 * The index to the array should have:
824 	 * bit[1:0] = l3_hdr_type
825 	 * bit[4:2] = l4_hdr_type
826 	 * bit[5] = ip_frag
827 	 * bit[6] = tunneled
828 	 * bit[7] = outer_l3_type
829 	 */
830 	idx = ((pinfo & 0x3) << 6) | ((ptype & 0xfc00) >> 10);
831 	return mlx5_ptype_table[idx] | rxq->tunnel * !!(idx & (1 << 6));
832 }
833 
834 /**
835  * Initialize Rx WQ and indexes.
836  *
837  * @param[in] rxq
838  *   Pointer to RX queue structure.
839  */
840 void
841 mlx5_rxq_initialize(struct mlx5_rxq_data *rxq)
842 {
843 	const unsigned int wqe_n = 1 << rxq->elts_n;
844 	unsigned int i;
845 
846 	for (i = 0; (i != wqe_n); ++i) {
847 		volatile struct mlx5_wqe_data_seg *scat;
848 		uintptr_t addr;
849 		uint32_t byte_count;
850 
851 		if (mlx5_rxq_mprq_enabled(rxq)) {
852 			struct mlx5_mprq_buf *buf = (*rxq->mprq_bufs)[i];
853 
854 			scat = &((volatile struct mlx5_wqe_mprq *)
855 				rxq->wqes)[i].dseg;
856 			addr = (uintptr_t)mlx5_mprq_buf_addr(buf,
857 							 1 << rxq->strd_num_n);
858 			byte_count = (1 << rxq->strd_sz_n) *
859 					(1 << rxq->strd_num_n);
860 		} else {
861 			struct rte_mbuf *buf = (*rxq->elts)[i];
862 
863 			scat = &((volatile struct mlx5_wqe_data_seg *)
864 					rxq->wqes)[i];
865 			addr = rte_pktmbuf_mtod(buf, uintptr_t);
866 			byte_count = DATA_LEN(buf);
867 		}
868 		/* scat->addr must be able to store a pointer. */
869 		MLX5_ASSERT(sizeof(scat->addr) >= sizeof(uintptr_t));
870 		*scat = (struct mlx5_wqe_data_seg){
871 			.addr = rte_cpu_to_be_64(addr),
872 			.byte_count = rte_cpu_to_be_32(byte_count),
873 			.lkey = mlx5_rx_addr2mr(rxq, addr),
874 		};
875 	}
876 	rxq->consumed_strd = 0;
877 	rxq->decompressed = 0;
878 	rxq->rq_pi = 0;
879 	rxq->zip = (struct rxq_zip){
880 		.ai = 0,
881 	};
882 	/* Update doorbell counter. */
883 	rxq->rq_ci = wqe_n >> rxq->sges_n;
884 	rte_cio_wmb();
885 	*rxq->rq_db = rte_cpu_to_be_32(rxq->rq_ci);
886 }
887 
888 /**
889  * Modify a Verbs/DevX queue state.
890  * This must be called from the primary process.
891  *
892  * @param dev
893  *   Pointer to Ethernet device.
894  * @param sm
895  *   State modify request parameters.
896  *
897  * @return
898  *   0 in case of success else non-zero value and rte_errno is set.
899  */
900 int
901 mlx5_queue_state_modify_primary(struct rte_eth_dev *dev,
902 			const struct mlx5_mp_arg_queue_state_modify *sm)
903 {
904 	int ret;
905 	struct mlx5_priv *priv = dev->data->dev_private;
906 
907 	if (sm->is_wq) {
908 		struct mlx5_rxq_data *rxq = (*priv->rxqs)[sm->queue_id];
909 		struct mlx5_rxq_ctrl *rxq_ctrl =
910 			container_of(rxq, struct mlx5_rxq_ctrl, rxq);
911 
912 		if (rxq_ctrl->obj->type == MLX5_RXQ_OBJ_TYPE_IBV) {
913 			struct ibv_wq_attr mod = {
914 				.attr_mask = IBV_WQ_ATTR_STATE,
915 				.wq_state = sm->state,
916 			};
917 
918 			ret = mlx5_glue->modify_wq(rxq_ctrl->obj->wq, &mod);
919 		} else { /* rxq_ctrl->obj->type == MLX5_RXQ_OBJ_TYPE_DEVX_RQ. */
920 			struct mlx5_devx_modify_rq_attr rq_attr;
921 
922 			memset(&rq_attr, 0, sizeof(rq_attr));
923 			if (sm->state == IBV_WQS_RESET) {
924 				rq_attr.rq_state = MLX5_RQC_STATE_ERR;
925 				rq_attr.state = MLX5_RQC_STATE_RST;
926 			} else if (sm->state == IBV_WQS_RDY) {
927 				rq_attr.rq_state = MLX5_RQC_STATE_RST;
928 				rq_attr.state = MLX5_RQC_STATE_RDY;
929 			} else if (sm->state == IBV_WQS_ERR) {
930 				rq_attr.rq_state = MLX5_RQC_STATE_RDY;
931 				rq_attr.state = MLX5_RQC_STATE_ERR;
932 			}
933 			ret = mlx5_devx_cmd_modify_rq(rxq_ctrl->obj->rq,
934 						      &rq_attr);
935 		}
936 		if (ret) {
937 			DRV_LOG(ERR, "Cannot change Rx WQ state to %u  - %s",
938 					sm->state, strerror(errno));
939 			rte_errno = errno;
940 			return ret;
941 		}
942 	} else {
943 		struct mlx5_txq_data *txq = (*priv->txqs)[sm->queue_id];
944 		struct mlx5_txq_ctrl *txq_ctrl =
945 			container_of(txq, struct mlx5_txq_ctrl, txq);
946 		struct ibv_qp_attr mod = {
947 			.qp_state = IBV_QPS_RESET,
948 			.port_num = (uint8_t)priv->ibv_port,
949 		};
950 		struct ibv_qp *qp = txq_ctrl->obj->qp;
951 
952 		ret = mlx5_glue->modify_qp(qp, &mod, IBV_QP_STATE);
953 		if (ret) {
954 			DRV_LOG(ERR, "Cannot change the Tx QP state to RESET "
955 				"%s", strerror(errno));
956 			rte_errno = errno;
957 			return ret;
958 		}
959 		mod.qp_state = IBV_QPS_INIT;
960 		ret = mlx5_glue->modify_qp(qp, &mod,
961 					   (IBV_QP_STATE | IBV_QP_PORT));
962 		if (ret) {
963 			DRV_LOG(ERR, "Cannot change Tx QP state to INIT %s",
964 				strerror(errno));
965 			rte_errno = errno;
966 			return ret;
967 		}
968 		mod.qp_state = IBV_QPS_RTR;
969 		ret = mlx5_glue->modify_qp(qp, &mod, IBV_QP_STATE);
970 		if (ret) {
971 			DRV_LOG(ERR, "Cannot change Tx QP state to RTR %s",
972 				strerror(errno));
973 			rte_errno = errno;
974 			return ret;
975 		}
976 		mod.qp_state = IBV_QPS_RTS;
977 		ret = mlx5_glue->modify_qp(qp, &mod, IBV_QP_STATE);
978 		if (ret) {
979 			DRV_LOG(ERR, "Cannot change Tx QP state to RTS %s",
980 				strerror(errno));
981 			rte_errno = errno;
982 			return ret;
983 		}
984 	}
985 	return 0;
986 }
987 
988 /**
989  * Modify a Verbs queue state.
990  *
991  * @param dev
992  *   Pointer to Ethernet device.
993  * @param sm
994  *   State modify request parameters.
995  *
996  * @return
997  *   0 in case of success else non-zero value.
998  */
999 static int
1000 mlx5_queue_state_modify(struct rte_eth_dev *dev,
1001 			struct mlx5_mp_arg_queue_state_modify *sm)
1002 {
1003 	int ret = 0;
1004 
1005 	switch (rte_eal_process_type()) {
1006 	case RTE_PROC_PRIMARY:
1007 		ret = mlx5_queue_state_modify_primary(dev, sm);
1008 		break;
1009 	case RTE_PROC_SECONDARY:
1010 		ret = mlx5_mp_req_queue_state_modify(dev, sm);
1011 		break;
1012 	default:
1013 		break;
1014 	}
1015 	return ret;
1016 }
1017 
1018 /**
1019  * Handle a Rx error.
1020  * The function inserts the RQ state to reset when the first error CQE is
1021  * shown, then drains the CQ by the caller function loop. When the CQ is empty,
1022  * it moves the RQ state to ready and initializes the RQ.
1023  * Next CQE identification and error counting are in the caller responsibility.
1024  *
1025  * @param[in] rxq
1026  *   Pointer to RX queue structure.
1027  * @param[in] vec
1028  *   1 when called from vectorized Rx burst, need to prepare mbufs for the RQ.
1029  *   0 when called from non-vectorized Rx burst.
1030  *
1031  * @return
1032  *   -1 in case of recovery error, otherwise the CQE status.
1033  */
1034 int
1035 mlx5_rx_err_handle(struct mlx5_rxq_data *rxq, uint8_t vec)
1036 {
1037 	const uint16_t cqe_n = 1 << rxq->cqe_n;
1038 	const uint16_t cqe_mask = cqe_n - 1;
1039 	const unsigned int wqe_n = 1 << rxq->elts_n;
1040 	struct mlx5_rxq_ctrl *rxq_ctrl =
1041 			container_of(rxq, struct mlx5_rxq_ctrl, rxq);
1042 	union {
1043 		volatile struct mlx5_cqe *cqe;
1044 		volatile struct mlx5_err_cqe *err_cqe;
1045 	} u = {
1046 		.cqe = &(*rxq->cqes)[rxq->cq_ci & cqe_mask],
1047 	};
1048 	struct mlx5_mp_arg_queue_state_modify sm;
1049 	int ret;
1050 
1051 	switch (rxq->err_state) {
1052 	case MLX5_RXQ_ERR_STATE_NO_ERROR:
1053 		rxq->err_state = MLX5_RXQ_ERR_STATE_NEED_RESET;
1054 		/* Fall-through */
1055 	case MLX5_RXQ_ERR_STATE_NEED_RESET:
1056 		sm.is_wq = 1;
1057 		sm.queue_id = rxq->idx;
1058 		sm.state = IBV_WQS_RESET;
1059 		if (mlx5_queue_state_modify(ETH_DEV(rxq_ctrl->priv), &sm))
1060 			return -1;
1061 		if (rxq_ctrl->dump_file_n <
1062 		    rxq_ctrl->priv->config.max_dump_files_num) {
1063 			MKSTR(err_str, "Unexpected CQE error syndrome "
1064 			      "0x%02x CQN = %u RQN = %u wqe_counter = %u"
1065 			      " rq_ci = %u cq_ci = %u", u.err_cqe->syndrome,
1066 			      rxq->cqn, rxq_ctrl->wqn,
1067 			      rte_be_to_cpu_16(u.err_cqe->wqe_counter),
1068 			      rxq->rq_ci << rxq->sges_n, rxq->cq_ci);
1069 			MKSTR(name, "dpdk_mlx5_port_%u_rxq_%u_%u",
1070 			      rxq->port_id, rxq->idx, (uint32_t)rte_rdtsc());
1071 			mlx5_dump_debug_information(name, NULL, err_str, 0);
1072 			mlx5_dump_debug_information(name, "MLX5 Error CQ:",
1073 						    (const void *)((uintptr_t)
1074 								    rxq->cqes),
1075 						    sizeof(*u.cqe) * cqe_n);
1076 			mlx5_dump_debug_information(name, "MLX5 Error RQ:",
1077 						    (const void *)((uintptr_t)
1078 								    rxq->wqes),
1079 						    16 * wqe_n);
1080 			rxq_ctrl->dump_file_n++;
1081 		}
1082 		rxq->err_state = MLX5_RXQ_ERR_STATE_NEED_READY;
1083 		/* Fall-through */
1084 	case MLX5_RXQ_ERR_STATE_NEED_READY:
1085 		ret = check_cqe(u.cqe, cqe_n, rxq->cq_ci);
1086 		if (ret == MLX5_CQE_STATUS_HW_OWN) {
1087 			rte_cio_wmb();
1088 			*rxq->cq_db = rte_cpu_to_be_32(rxq->cq_ci);
1089 			rte_cio_wmb();
1090 			/*
1091 			 * The RQ consumer index must be zeroed while moving
1092 			 * from RESET state to RDY state.
1093 			 */
1094 			*rxq->rq_db = rte_cpu_to_be_32(0);
1095 			rte_cio_wmb();
1096 			sm.is_wq = 1;
1097 			sm.queue_id = rxq->idx;
1098 			sm.state = IBV_WQS_RDY;
1099 			if (mlx5_queue_state_modify(ETH_DEV(rxq_ctrl->priv),
1100 						    &sm))
1101 				return -1;
1102 			if (vec) {
1103 				const uint16_t q_mask = wqe_n - 1;
1104 				uint16_t elt_idx;
1105 				struct rte_mbuf **elt;
1106 				int i;
1107 				unsigned int n = wqe_n - (rxq->rq_ci -
1108 							  rxq->rq_pi);
1109 
1110 				for (i = 0; i < (int)n; ++i) {
1111 					elt_idx = (rxq->rq_ci + i) & q_mask;
1112 					elt = &(*rxq->elts)[elt_idx];
1113 					*elt = rte_mbuf_raw_alloc(rxq->mp);
1114 					if (!*elt) {
1115 						for (i--; i >= 0; --i) {
1116 							elt_idx = (rxq->rq_ci +
1117 								   i) & q_mask;
1118 							elt = &(*rxq->elts)
1119 								[elt_idx];
1120 							rte_pktmbuf_free_seg
1121 								(*elt);
1122 						}
1123 						return -1;
1124 					}
1125 				}
1126 				for (i = 0; i < (int)wqe_n; ++i) {
1127 					elt = &(*rxq->elts)[i];
1128 					DATA_LEN(*elt) =
1129 						(uint16_t)((*elt)->buf_len -
1130 						rte_pktmbuf_headroom(*elt));
1131 				}
1132 				/* Padding with a fake mbuf for vec Rx. */
1133 				for (i = 0; i < MLX5_VPMD_DESCS_PER_LOOP; ++i)
1134 					(*rxq->elts)[wqe_n + i] =
1135 								&rxq->fake_mbuf;
1136 			}
1137 			mlx5_rxq_initialize(rxq);
1138 			rxq->err_state = MLX5_RXQ_ERR_STATE_NO_ERROR;
1139 		}
1140 		return ret;
1141 	default:
1142 		return -1;
1143 	}
1144 }
1145 
1146 /**
1147  * Get size of the next packet for a given CQE. For compressed CQEs, the
1148  * consumer index is updated only once all packets of the current one have
1149  * been processed.
1150  *
1151  * @param rxq
1152  *   Pointer to RX queue.
1153  * @param cqe
1154  *   CQE to process.
1155  * @param[out] mcqe
1156  *   Store pointer to mini-CQE if compressed. Otherwise, the pointer is not
1157  *   written.
1158  *
1159  * @return
1160  *   0 in case of empty CQE, otherwise the packet size in bytes.
1161  */
1162 static inline int
1163 mlx5_rx_poll_len(struct mlx5_rxq_data *rxq, volatile struct mlx5_cqe *cqe,
1164 		 uint16_t cqe_cnt, volatile struct mlx5_mini_cqe8 **mcqe)
1165 {
1166 	struct rxq_zip *zip = &rxq->zip;
1167 	uint16_t cqe_n = cqe_cnt + 1;
1168 	int len;
1169 	uint16_t idx, end;
1170 
1171 	do {
1172 		len = 0;
1173 		/* Process compressed data in the CQE and mini arrays. */
1174 		if (zip->ai) {
1175 			volatile struct mlx5_mini_cqe8 (*mc)[8] =
1176 				(volatile struct mlx5_mini_cqe8 (*)[8])
1177 				(uintptr_t)(&(*rxq->cqes)[zip->ca &
1178 							  cqe_cnt].pkt_info);
1179 
1180 			len = rte_be_to_cpu_32((*mc)[zip->ai & 7].byte_cnt);
1181 			*mcqe = &(*mc)[zip->ai & 7];
1182 			if ((++zip->ai & 7) == 0) {
1183 				/* Invalidate consumed CQEs */
1184 				idx = zip->ca;
1185 				end = zip->na;
1186 				while (idx != end) {
1187 					(*rxq->cqes)[idx & cqe_cnt].op_own =
1188 						MLX5_CQE_INVALIDATE;
1189 					++idx;
1190 				}
1191 				/*
1192 				 * Increment consumer index to skip the number
1193 				 * of CQEs consumed. Hardware leaves holes in
1194 				 * the CQ ring for software use.
1195 				 */
1196 				zip->ca = zip->na;
1197 				zip->na += 8;
1198 			}
1199 			if (unlikely(rxq->zip.ai == rxq->zip.cqe_cnt)) {
1200 				/* Invalidate the rest */
1201 				idx = zip->ca;
1202 				end = zip->cq_ci;
1203 
1204 				while (idx != end) {
1205 					(*rxq->cqes)[idx & cqe_cnt].op_own =
1206 						MLX5_CQE_INVALIDATE;
1207 					++idx;
1208 				}
1209 				rxq->cq_ci = zip->cq_ci;
1210 				zip->ai = 0;
1211 			}
1212 		/*
1213 		 * No compressed data, get next CQE and verify if it is
1214 		 * compressed.
1215 		 */
1216 		} else {
1217 			int ret;
1218 			int8_t op_own;
1219 
1220 			ret = check_cqe(cqe, cqe_n, rxq->cq_ci);
1221 			if (unlikely(ret != MLX5_CQE_STATUS_SW_OWN)) {
1222 				if (unlikely(ret == MLX5_CQE_STATUS_ERR ||
1223 					     rxq->err_state)) {
1224 					ret = mlx5_rx_err_handle(rxq, 0);
1225 					if (ret == MLX5_CQE_STATUS_HW_OWN ||
1226 					    ret == -1)
1227 						return 0;
1228 				} else {
1229 					return 0;
1230 				}
1231 			}
1232 			++rxq->cq_ci;
1233 			op_own = cqe->op_own;
1234 			if (MLX5_CQE_FORMAT(op_own) == MLX5_COMPRESSED) {
1235 				volatile struct mlx5_mini_cqe8 (*mc)[8] =
1236 					(volatile struct mlx5_mini_cqe8 (*)[8])
1237 					(uintptr_t)(&(*rxq->cqes)
1238 						[rxq->cq_ci &
1239 						 cqe_cnt].pkt_info);
1240 
1241 				/* Fix endianness. */
1242 				zip->cqe_cnt = rte_be_to_cpu_32(cqe->byte_cnt);
1243 				/*
1244 				 * Current mini array position is the one
1245 				 * returned by check_cqe64().
1246 				 *
1247 				 * If completion comprises several mini arrays,
1248 				 * as a special case the second one is located
1249 				 * 7 CQEs after the initial CQE instead of 8
1250 				 * for subsequent ones.
1251 				 */
1252 				zip->ca = rxq->cq_ci;
1253 				zip->na = zip->ca + 7;
1254 				/* Compute the next non compressed CQE. */
1255 				--rxq->cq_ci;
1256 				zip->cq_ci = rxq->cq_ci + zip->cqe_cnt;
1257 				/* Get packet size to return. */
1258 				len = rte_be_to_cpu_32((*mc)[0].byte_cnt);
1259 				*mcqe = &(*mc)[0];
1260 				zip->ai = 1;
1261 				/* Prefetch all to be invalidated */
1262 				idx = zip->ca;
1263 				end = zip->cq_ci;
1264 				while (idx != end) {
1265 					rte_prefetch0(&(*rxq->cqes)[(idx) &
1266 								    cqe_cnt]);
1267 					++idx;
1268 				}
1269 			} else {
1270 				len = rte_be_to_cpu_32(cqe->byte_cnt);
1271 			}
1272 		}
1273 		if (unlikely(rxq->err_state)) {
1274 			cqe = &(*rxq->cqes)[rxq->cq_ci & cqe_cnt];
1275 			++rxq->stats.idropped;
1276 		} else {
1277 			return len;
1278 		}
1279 	} while (1);
1280 }
1281 
1282 /**
1283  * Translate RX completion flags to offload flags.
1284  *
1285  * @param[in] cqe
1286  *   Pointer to CQE.
1287  *
1288  * @return
1289  *   Offload flags (ol_flags) for struct rte_mbuf.
1290  */
1291 static inline uint32_t
1292 rxq_cq_to_ol_flags(volatile struct mlx5_cqe *cqe)
1293 {
1294 	uint32_t ol_flags = 0;
1295 	uint16_t flags = rte_be_to_cpu_16(cqe->hdr_type_etc);
1296 
1297 	ol_flags =
1298 		TRANSPOSE(flags,
1299 			  MLX5_CQE_RX_L3_HDR_VALID,
1300 			  PKT_RX_IP_CKSUM_GOOD) |
1301 		TRANSPOSE(flags,
1302 			  MLX5_CQE_RX_L4_HDR_VALID,
1303 			  PKT_RX_L4_CKSUM_GOOD);
1304 	return ol_flags;
1305 }
1306 
1307 /**
1308  * Fill in mbuf fields from RX completion flags.
1309  * Note that pkt->ol_flags should be initialized outside of this function.
1310  *
1311  * @param rxq
1312  *   Pointer to RX queue.
1313  * @param pkt
1314  *   mbuf to fill.
1315  * @param cqe
1316  *   CQE to process.
1317  * @param rss_hash_res
1318  *   Packet RSS Hash result.
1319  */
1320 static inline void
1321 rxq_cq_to_mbuf(struct mlx5_rxq_data *rxq, struct rte_mbuf *pkt,
1322 	       volatile struct mlx5_cqe *cqe, uint32_t rss_hash_res)
1323 {
1324 	/* Update packet information. */
1325 	pkt->packet_type = rxq_cq_to_pkt_type(rxq, cqe);
1326 	if (rss_hash_res && rxq->rss_hash) {
1327 		pkt->hash.rss = rss_hash_res;
1328 		pkt->ol_flags |= PKT_RX_RSS_HASH;
1329 	}
1330 	if (rxq->mark && MLX5_FLOW_MARK_IS_VALID(cqe->sop_drop_qpn)) {
1331 		pkt->ol_flags |= PKT_RX_FDIR;
1332 		if (cqe->sop_drop_qpn !=
1333 		    rte_cpu_to_be_32(MLX5_FLOW_MARK_DEFAULT)) {
1334 			uint32_t mark = cqe->sop_drop_qpn;
1335 
1336 			pkt->ol_flags |= PKT_RX_FDIR_ID;
1337 			pkt->hash.fdir.hi = mlx5_flow_mark_get(mark);
1338 		}
1339 	}
1340 	if (rte_flow_dynf_metadata_avail() && cqe->flow_table_metadata) {
1341 		pkt->ol_flags |= PKT_RX_DYNF_METADATA;
1342 		*RTE_FLOW_DYNF_METADATA(pkt) = cqe->flow_table_metadata;
1343 	}
1344 	if (rxq->csum)
1345 		pkt->ol_flags |= rxq_cq_to_ol_flags(cqe);
1346 	if (rxq->vlan_strip &&
1347 	    (cqe->hdr_type_etc & rte_cpu_to_be_16(MLX5_CQE_VLAN_STRIPPED))) {
1348 		pkt->ol_flags |= PKT_RX_VLAN | PKT_RX_VLAN_STRIPPED;
1349 		pkt->vlan_tci = rte_be_to_cpu_16(cqe->vlan_info);
1350 	}
1351 	if (rxq->hw_timestamp) {
1352 		pkt->timestamp = rte_be_to_cpu_64(cqe->timestamp);
1353 		pkt->ol_flags |= PKT_RX_TIMESTAMP;
1354 	}
1355 }
1356 
1357 /**
1358  * DPDK callback for RX.
1359  *
1360  * @param dpdk_rxq
1361  *   Generic pointer to RX queue structure.
1362  * @param[out] pkts
1363  *   Array to store received packets.
1364  * @param pkts_n
1365  *   Maximum number of packets in array.
1366  *
1367  * @return
1368  *   Number of packets successfully received (<= pkts_n).
1369  */
1370 uint16_t
1371 mlx5_rx_burst(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n)
1372 {
1373 	struct mlx5_rxq_data *rxq = dpdk_rxq;
1374 	const unsigned int wqe_cnt = (1 << rxq->elts_n) - 1;
1375 	const unsigned int cqe_cnt = (1 << rxq->cqe_n) - 1;
1376 	const unsigned int sges_n = rxq->sges_n;
1377 	struct rte_mbuf *pkt = NULL;
1378 	struct rte_mbuf *seg = NULL;
1379 	volatile struct mlx5_cqe *cqe =
1380 		&(*rxq->cqes)[rxq->cq_ci & cqe_cnt];
1381 	unsigned int i = 0;
1382 	unsigned int rq_ci = rxq->rq_ci << sges_n;
1383 	int len = 0; /* keep its value across iterations. */
1384 
1385 	while (pkts_n) {
1386 		unsigned int idx = rq_ci & wqe_cnt;
1387 		volatile struct mlx5_wqe_data_seg *wqe =
1388 			&((volatile struct mlx5_wqe_data_seg *)rxq->wqes)[idx];
1389 		struct rte_mbuf *rep = (*rxq->elts)[idx];
1390 		volatile struct mlx5_mini_cqe8 *mcqe = NULL;
1391 		uint32_t rss_hash_res;
1392 
1393 		if (pkt)
1394 			NEXT(seg) = rep;
1395 		seg = rep;
1396 		rte_prefetch0(seg);
1397 		rte_prefetch0(cqe);
1398 		rte_prefetch0(wqe);
1399 		rep = rte_mbuf_raw_alloc(rxq->mp);
1400 		if (unlikely(rep == NULL)) {
1401 			++rxq->stats.rx_nombuf;
1402 			if (!pkt) {
1403 				/*
1404 				 * no buffers before we even started,
1405 				 * bail out silently.
1406 				 */
1407 				break;
1408 			}
1409 			while (pkt != seg) {
1410 				MLX5_ASSERT(pkt != (*rxq->elts)[idx]);
1411 				rep = NEXT(pkt);
1412 				NEXT(pkt) = NULL;
1413 				NB_SEGS(pkt) = 1;
1414 				rte_mbuf_raw_free(pkt);
1415 				pkt = rep;
1416 			}
1417 			break;
1418 		}
1419 		if (!pkt) {
1420 			cqe = &(*rxq->cqes)[rxq->cq_ci & cqe_cnt];
1421 			len = mlx5_rx_poll_len(rxq, cqe, cqe_cnt, &mcqe);
1422 			if (!len) {
1423 				rte_mbuf_raw_free(rep);
1424 				break;
1425 			}
1426 			pkt = seg;
1427 			MLX5_ASSERT(len >= (rxq->crc_present << 2));
1428 			pkt->ol_flags &= EXT_ATTACHED_MBUF;
1429 			/* If compressed, take hash result from mini-CQE. */
1430 			rss_hash_res = rte_be_to_cpu_32(mcqe == NULL ?
1431 							cqe->rx_hash_res :
1432 							mcqe->rx_hash_result);
1433 			rxq_cq_to_mbuf(rxq, pkt, cqe, rss_hash_res);
1434 			if (rxq->crc_present)
1435 				len -= RTE_ETHER_CRC_LEN;
1436 			PKT_LEN(pkt) = len;
1437 			if (cqe->lro_num_seg > 1) {
1438 				mlx5_lro_update_hdr
1439 					(rte_pktmbuf_mtod(pkt, uint8_t *), cqe,
1440 					 len);
1441 				pkt->ol_flags |= PKT_RX_LRO;
1442 				pkt->tso_segsz = len / cqe->lro_num_seg;
1443 			}
1444 		}
1445 		DATA_LEN(rep) = DATA_LEN(seg);
1446 		PKT_LEN(rep) = PKT_LEN(seg);
1447 		SET_DATA_OFF(rep, DATA_OFF(seg));
1448 		PORT(rep) = PORT(seg);
1449 		(*rxq->elts)[idx] = rep;
1450 		/*
1451 		 * Fill NIC descriptor with the new buffer.  The lkey and size
1452 		 * of the buffers are already known, only the buffer address
1453 		 * changes.
1454 		 */
1455 		wqe->addr = rte_cpu_to_be_64(rte_pktmbuf_mtod(rep, uintptr_t));
1456 		/* If there's only one MR, no need to replace LKey in WQE. */
1457 		if (unlikely(mlx5_mr_btree_len(&rxq->mr_ctrl.cache_bh) > 1))
1458 			wqe->lkey = mlx5_rx_mb2mr(rxq, rep);
1459 		if (len > DATA_LEN(seg)) {
1460 			len -= DATA_LEN(seg);
1461 			++NB_SEGS(pkt);
1462 			++rq_ci;
1463 			continue;
1464 		}
1465 		DATA_LEN(seg) = len;
1466 #ifdef MLX5_PMD_SOFT_COUNTERS
1467 		/* Increment bytes counter. */
1468 		rxq->stats.ibytes += PKT_LEN(pkt);
1469 #endif
1470 		/* Return packet. */
1471 		*(pkts++) = pkt;
1472 		pkt = NULL;
1473 		--pkts_n;
1474 		++i;
1475 		/* Align consumer index to the next stride. */
1476 		rq_ci >>= sges_n;
1477 		++rq_ci;
1478 		rq_ci <<= sges_n;
1479 	}
1480 	if (unlikely((i == 0) && ((rq_ci >> sges_n) == rxq->rq_ci)))
1481 		return 0;
1482 	/* Update the consumer index. */
1483 	rxq->rq_ci = rq_ci >> sges_n;
1484 	rte_cio_wmb();
1485 	*rxq->cq_db = rte_cpu_to_be_32(rxq->cq_ci);
1486 	rte_cio_wmb();
1487 	*rxq->rq_db = rte_cpu_to_be_32(rxq->rq_ci);
1488 #ifdef MLX5_PMD_SOFT_COUNTERS
1489 	/* Increment packets counter. */
1490 	rxq->stats.ipackets += i;
1491 #endif
1492 	return i;
1493 }
1494 
1495 /**
1496  * Update LRO packet TCP header.
1497  * The HW LRO feature doesn't update the TCP header after coalescing the
1498  * TCP segments but supplies information in CQE to fill it by SW.
1499  *
1500  * @param tcp
1501  *   Pointer to the TCP header.
1502  * @param cqe
1503  *   Pointer to the completion entry..
1504  * @param phcsum
1505  *   The L3 pseudo-header checksum.
1506  */
1507 static inline void
1508 mlx5_lro_update_tcp_hdr(struct rte_tcp_hdr *restrict tcp,
1509 			volatile struct mlx5_cqe *restrict cqe,
1510 			uint32_t phcsum)
1511 {
1512 	uint8_t l4_type = (rte_be_to_cpu_16(cqe->hdr_type_etc) &
1513 			   MLX5_CQE_L4_TYPE_MASK) >> MLX5_CQE_L4_TYPE_SHIFT;
1514 	/*
1515 	 * The HW calculates only the TCP payload checksum, need to complete
1516 	 * the TCP header checksum and the L3 pseudo-header checksum.
1517 	 */
1518 	uint32_t csum = phcsum + cqe->csum;
1519 
1520 	if (l4_type == MLX5_L4_HDR_TYPE_TCP_EMPTY_ACK ||
1521 	    l4_type == MLX5_L4_HDR_TYPE_TCP_WITH_ACL) {
1522 		tcp->tcp_flags |= RTE_TCP_ACK_FLAG;
1523 		tcp->recv_ack = cqe->lro_ack_seq_num;
1524 		tcp->rx_win = cqe->lro_tcp_win;
1525 	}
1526 	if (cqe->lro_tcppsh_abort_dupack & MLX5_CQE_LRO_PUSH_MASK)
1527 		tcp->tcp_flags |= RTE_TCP_PSH_FLAG;
1528 	tcp->cksum = 0;
1529 	csum += rte_raw_cksum(tcp, (tcp->data_off & 0xF) * 4);
1530 	csum = ((csum & 0xffff0000) >> 16) + (csum & 0xffff);
1531 	csum = (~csum) & 0xffff;
1532 	if (csum == 0)
1533 		csum = 0xffff;
1534 	tcp->cksum = csum;
1535 }
1536 
1537 /**
1538  * Update LRO packet headers.
1539  * The HW LRO feature doesn't update the L3/TCP headers after coalescing the
1540  * TCP segments but supply information in CQE to fill it by SW.
1541  *
1542  * @param padd
1543  *   The packet address.
1544  * @param cqe
1545  *   Pointer to the completion entry..
1546  * @param len
1547  *   The packet length.
1548  */
1549 static inline void
1550 mlx5_lro_update_hdr(uint8_t *restrict padd,
1551 		    volatile struct mlx5_cqe *restrict cqe,
1552 		    uint32_t len)
1553 {
1554 	union {
1555 		struct rte_ether_hdr *eth;
1556 		struct rte_vlan_hdr *vlan;
1557 		struct rte_ipv4_hdr *ipv4;
1558 		struct rte_ipv6_hdr *ipv6;
1559 		struct rte_tcp_hdr *tcp;
1560 		uint8_t *hdr;
1561 	} h = {
1562 			.hdr = padd,
1563 	};
1564 	uint16_t proto = h.eth->ether_type;
1565 	uint32_t phcsum;
1566 
1567 	h.eth++;
1568 	while (proto == RTE_BE16(RTE_ETHER_TYPE_VLAN) ||
1569 	       proto == RTE_BE16(RTE_ETHER_TYPE_QINQ)) {
1570 		proto = h.vlan->eth_proto;
1571 		h.vlan++;
1572 	}
1573 	if (proto == RTE_BE16(RTE_ETHER_TYPE_IPV4)) {
1574 		h.ipv4->time_to_live = cqe->lro_min_ttl;
1575 		h.ipv4->total_length = rte_cpu_to_be_16(len - (h.hdr - padd));
1576 		h.ipv4->hdr_checksum = 0;
1577 		h.ipv4->hdr_checksum = rte_ipv4_cksum(h.ipv4);
1578 		phcsum = rte_ipv4_phdr_cksum(h.ipv4, 0);
1579 		h.ipv4++;
1580 	} else {
1581 		h.ipv6->hop_limits = cqe->lro_min_ttl;
1582 		h.ipv6->payload_len = rte_cpu_to_be_16(len - (h.hdr - padd) -
1583 						       sizeof(*h.ipv6));
1584 		phcsum = rte_ipv6_phdr_cksum(h.ipv6, 0);
1585 		h.ipv6++;
1586 	}
1587 	mlx5_lro_update_tcp_hdr(h.tcp, cqe, phcsum);
1588 }
1589 
1590 void
1591 mlx5_mprq_buf_free_cb(void *addr __rte_unused, void *opaque)
1592 {
1593 	struct mlx5_mprq_buf *buf = opaque;
1594 
1595 	if (rte_atomic16_read(&buf->refcnt) == 1) {
1596 		rte_mempool_put(buf->mp, buf);
1597 	} else if (rte_atomic16_add_return(&buf->refcnt, -1) == 0) {
1598 		rte_atomic16_set(&buf->refcnt, 1);
1599 		rte_mempool_put(buf->mp, buf);
1600 	}
1601 }
1602 
1603 void
1604 mlx5_mprq_buf_free(struct mlx5_mprq_buf *buf)
1605 {
1606 	mlx5_mprq_buf_free_cb(NULL, buf);
1607 }
1608 
1609 static inline void
1610 mprq_buf_replace(struct mlx5_rxq_data *rxq, uint16_t rq_idx,
1611 		 const unsigned int strd_n)
1612 {
1613 	struct mlx5_mprq_buf *rep = rxq->mprq_repl;
1614 	volatile struct mlx5_wqe_data_seg *wqe =
1615 		&((volatile struct mlx5_wqe_mprq *)rxq->wqes)[rq_idx].dseg;
1616 	void *addr;
1617 
1618 	MLX5_ASSERT(rep != NULL);
1619 	/* Replace MPRQ buf. */
1620 	(*rxq->mprq_bufs)[rq_idx] = rep;
1621 	/* Replace WQE. */
1622 	addr = mlx5_mprq_buf_addr(rep, strd_n);
1623 	wqe->addr = rte_cpu_to_be_64((uintptr_t)addr);
1624 	/* If there's only one MR, no need to replace LKey in WQE. */
1625 	if (unlikely(mlx5_mr_btree_len(&rxq->mr_ctrl.cache_bh) > 1))
1626 		wqe->lkey = mlx5_rx_addr2mr(rxq, (uintptr_t)addr);
1627 	/* Stash a mbuf for next replacement. */
1628 	if (likely(!rte_mempool_get(rxq->mprq_mp, (void **)&rep)))
1629 		rxq->mprq_repl = rep;
1630 	else
1631 		rxq->mprq_repl = NULL;
1632 }
1633 
1634 /**
1635  * DPDK callback for RX with Multi-Packet RQ support.
1636  *
1637  * @param dpdk_rxq
1638  *   Generic pointer to RX queue structure.
1639  * @param[out] pkts
1640  *   Array to store received packets.
1641  * @param pkts_n
1642  *   Maximum number of packets in array.
1643  *
1644  * @return
1645  *   Number of packets successfully received (<= pkts_n).
1646  */
1647 uint16_t
1648 mlx5_rx_burst_mprq(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n)
1649 {
1650 	struct mlx5_rxq_data *rxq = dpdk_rxq;
1651 	const unsigned int strd_n = 1 << rxq->strd_num_n;
1652 	const unsigned int strd_sz = 1 << rxq->strd_sz_n;
1653 	const unsigned int strd_shift =
1654 		MLX5_MPRQ_STRIDE_SHIFT_BYTE * rxq->strd_shift_en;
1655 	const unsigned int cq_mask = (1 << rxq->cqe_n) - 1;
1656 	const unsigned int wq_mask = (1 << rxq->elts_n) - 1;
1657 	volatile struct mlx5_cqe *cqe = &(*rxq->cqes)[rxq->cq_ci & cq_mask];
1658 	unsigned int i = 0;
1659 	uint32_t rq_ci = rxq->rq_ci;
1660 	uint16_t consumed_strd = rxq->consumed_strd;
1661 	uint16_t headroom_sz = rxq->strd_headroom_en * RTE_PKTMBUF_HEADROOM;
1662 	struct mlx5_mprq_buf *buf = (*rxq->mprq_bufs)[rq_ci & wq_mask];
1663 
1664 	while (i < pkts_n) {
1665 		struct rte_mbuf *pkt;
1666 		void *addr;
1667 		int ret;
1668 		unsigned int len;
1669 		uint16_t strd_cnt;
1670 		uint16_t strd_idx;
1671 		uint32_t offset;
1672 		uint32_t byte_cnt;
1673 		volatile struct mlx5_mini_cqe8 *mcqe = NULL;
1674 		uint32_t rss_hash_res = 0;
1675 		uint8_t lro_num_seg;
1676 
1677 		if (consumed_strd == strd_n) {
1678 			/* Replace WQE only if the buffer is still in use. */
1679 			if (rte_atomic16_read(&buf->refcnt) > 1) {
1680 				mprq_buf_replace(rxq, rq_ci & wq_mask, strd_n);
1681 				/* Release the old buffer. */
1682 				mlx5_mprq_buf_free(buf);
1683 			} else if (unlikely(rxq->mprq_repl == NULL)) {
1684 				struct mlx5_mprq_buf *rep;
1685 
1686 				/*
1687 				 * Currently, the MPRQ mempool is out of buffer
1688 				 * and doing memcpy regardless of the size of Rx
1689 				 * packet. Retry allocation to get back to
1690 				 * normal.
1691 				 */
1692 				if (!rte_mempool_get(rxq->mprq_mp,
1693 						     (void **)&rep))
1694 					rxq->mprq_repl = rep;
1695 			}
1696 			/* Advance to the next WQE. */
1697 			consumed_strd = 0;
1698 			++rq_ci;
1699 			buf = (*rxq->mprq_bufs)[rq_ci & wq_mask];
1700 		}
1701 		cqe = &(*rxq->cqes)[rxq->cq_ci & cq_mask];
1702 		ret = mlx5_rx_poll_len(rxq, cqe, cq_mask, &mcqe);
1703 		if (!ret)
1704 			break;
1705 		byte_cnt = ret;
1706 		strd_cnt = (byte_cnt & MLX5_MPRQ_STRIDE_NUM_MASK) >>
1707 			   MLX5_MPRQ_STRIDE_NUM_SHIFT;
1708 		MLX5_ASSERT(strd_cnt);
1709 		consumed_strd += strd_cnt;
1710 		if (byte_cnt & MLX5_MPRQ_FILLER_MASK)
1711 			continue;
1712 		if (mcqe == NULL) {
1713 			rss_hash_res = rte_be_to_cpu_32(cqe->rx_hash_res);
1714 			strd_idx = rte_be_to_cpu_16(cqe->wqe_counter);
1715 		} else {
1716 			/* mini-CQE for MPRQ doesn't have hash result. */
1717 			strd_idx = rte_be_to_cpu_16(mcqe->stride_idx);
1718 		}
1719 		MLX5_ASSERT(strd_idx < strd_n);
1720 		MLX5_ASSERT(!((rte_be_to_cpu_16(cqe->wqe_id) ^ rq_ci) &
1721 			    wq_mask));
1722 		lro_num_seg = cqe->lro_num_seg;
1723 		/*
1724 		 * Currently configured to receive a packet per a stride. But if
1725 		 * MTU is adjusted through kernel interface, device could
1726 		 * consume multiple strides without raising an error. In this
1727 		 * case, the packet should be dropped because it is bigger than
1728 		 * the max_rx_pkt_len.
1729 		 */
1730 		if (unlikely(!lro_num_seg && strd_cnt > 1)) {
1731 			++rxq->stats.idropped;
1732 			continue;
1733 		}
1734 		pkt = rte_pktmbuf_alloc(rxq->mp);
1735 		if (unlikely(pkt == NULL)) {
1736 			++rxq->stats.rx_nombuf;
1737 			break;
1738 		}
1739 		len = (byte_cnt & MLX5_MPRQ_LEN_MASK) >> MLX5_MPRQ_LEN_SHIFT;
1740 		MLX5_ASSERT((int)len >= (rxq->crc_present << 2));
1741 		if (rxq->crc_present)
1742 			len -= RTE_ETHER_CRC_LEN;
1743 		offset = strd_idx * strd_sz + strd_shift;
1744 		addr = RTE_PTR_ADD(mlx5_mprq_buf_addr(buf, strd_n), offset);
1745 		/*
1746 		 * Memcpy packets to the target mbuf if:
1747 		 * - The size of packet is smaller than mprq_max_memcpy_len.
1748 		 * - Out of buffer in the Mempool for Multi-Packet RQ.
1749 		 */
1750 		if (len <= rxq->mprq_max_memcpy_len || rxq->mprq_repl == NULL) {
1751 			/*
1752 			 * When memcpy'ing packet due to out-of-buffer, the
1753 			 * packet must be smaller than the target mbuf.
1754 			 */
1755 			if (unlikely(rte_pktmbuf_tailroom(pkt) < len)) {
1756 				rte_pktmbuf_free_seg(pkt);
1757 				++rxq->stats.idropped;
1758 				continue;
1759 			}
1760 			rte_memcpy(rte_pktmbuf_mtod(pkt, void *), addr, len);
1761 			DATA_LEN(pkt) = len;
1762 		} else {
1763 			rte_iova_t buf_iova;
1764 			struct rte_mbuf_ext_shared_info *shinfo;
1765 			uint16_t buf_len = strd_cnt * strd_sz;
1766 			void *buf_addr;
1767 
1768 			/* Increment the refcnt of the whole chunk. */
1769 			rte_atomic16_add_return(&buf->refcnt, 1);
1770 			MLX5_ASSERT((uint16_t)rte_atomic16_read(&buf->refcnt) <=
1771 				    strd_n + 1);
1772 			buf_addr = RTE_PTR_SUB(addr, headroom_sz);
1773 			/*
1774 			 * MLX5 device doesn't use iova but it is necessary in a
1775 			 * case where the Rx packet is transmitted via a
1776 			 * different PMD.
1777 			 */
1778 			buf_iova = rte_mempool_virt2iova(buf) +
1779 				   RTE_PTR_DIFF(buf_addr, buf);
1780 			shinfo = &buf->shinfos[strd_idx];
1781 			rte_mbuf_ext_refcnt_set(shinfo, 1);
1782 			/*
1783 			 * EXT_ATTACHED_MBUF will be set to pkt->ol_flags when
1784 			 * attaching the stride to mbuf and more offload flags
1785 			 * will be added below by calling rxq_cq_to_mbuf().
1786 			 * Other fields will be overwritten.
1787 			 */
1788 			rte_pktmbuf_attach_extbuf(pkt, buf_addr, buf_iova,
1789 						  buf_len, shinfo);
1790 			/* Set mbuf head-room. */
1791 			pkt->data_off = headroom_sz;
1792 			MLX5_ASSERT(pkt->ol_flags == EXT_ATTACHED_MBUF);
1793 			/*
1794 			 * Prevent potential overflow due to MTU change through
1795 			 * kernel interface.
1796 			 */
1797 			if (unlikely(rte_pktmbuf_tailroom(pkt) < len)) {
1798 				rte_pktmbuf_free_seg(pkt);
1799 				++rxq->stats.idropped;
1800 				continue;
1801 			}
1802 			DATA_LEN(pkt) = len;
1803 			/*
1804 			 * LRO packet may consume all the stride memory, in this
1805 			 * case packet head-room space is not guaranteed so must
1806 			 * to add an empty mbuf for the head-room.
1807 			 */
1808 			if (!rxq->strd_headroom_en) {
1809 				struct rte_mbuf *headroom_mbuf =
1810 						rte_pktmbuf_alloc(rxq->mp);
1811 
1812 				if (unlikely(headroom_mbuf == NULL)) {
1813 					rte_pktmbuf_free_seg(pkt);
1814 					++rxq->stats.rx_nombuf;
1815 					break;
1816 				}
1817 				PORT(pkt) = rxq->port_id;
1818 				NEXT(headroom_mbuf) = pkt;
1819 				pkt = headroom_mbuf;
1820 				NB_SEGS(pkt) = 2;
1821 			}
1822 		}
1823 		rxq_cq_to_mbuf(rxq, pkt, cqe, rss_hash_res);
1824 		if (lro_num_seg > 1) {
1825 			mlx5_lro_update_hdr(addr, cqe, len);
1826 			pkt->ol_flags |= PKT_RX_LRO;
1827 			pkt->tso_segsz = strd_sz;
1828 		}
1829 		PKT_LEN(pkt) = len;
1830 		PORT(pkt) = rxq->port_id;
1831 #ifdef MLX5_PMD_SOFT_COUNTERS
1832 		/* Increment bytes counter. */
1833 		rxq->stats.ibytes += PKT_LEN(pkt);
1834 #endif
1835 		/* Return packet. */
1836 		*(pkts++) = pkt;
1837 		++i;
1838 	}
1839 	/* Update the consumer indexes. */
1840 	rxq->consumed_strd = consumed_strd;
1841 	rte_cio_wmb();
1842 	*rxq->cq_db = rte_cpu_to_be_32(rxq->cq_ci);
1843 	if (rq_ci != rxq->rq_ci) {
1844 		rxq->rq_ci = rq_ci;
1845 		rte_cio_wmb();
1846 		*rxq->rq_db = rte_cpu_to_be_32(rxq->rq_ci);
1847 	}
1848 #ifdef MLX5_PMD_SOFT_COUNTERS
1849 	/* Increment packets counter. */
1850 	rxq->stats.ipackets += i;
1851 #endif
1852 	return i;
1853 }
1854 
1855 /**
1856  * Dummy DPDK callback for TX.
1857  *
1858  * This function is used to temporarily replace the real callback during
1859  * unsafe control operations on the queue, or in case of error.
1860  *
1861  * @param dpdk_txq
1862  *   Generic pointer to TX queue structure.
1863  * @param[in] pkts
1864  *   Packets to transmit.
1865  * @param pkts_n
1866  *   Number of packets in array.
1867  *
1868  * @return
1869  *   Number of packets successfully transmitted (<= pkts_n).
1870  */
1871 uint16_t
1872 removed_tx_burst(void *dpdk_txq __rte_unused,
1873 		 struct rte_mbuf **pkts __rte_unused,
1874 		 uint16_t pkts_n __rte_unused)
1875 {
1876 	rte_mb();
1877 	return 0;
1878 }
1879 
1880 /**
1881  * Dummy DPDK callback for RX.
1882  *
1883  * This function is used to temporarily replace the real callback during
1884  * unsafe control operations on the queue, or in case of error.
1885  *
1886  * @param dpdk_rxq
1887  *   Generic pointer to RX queue structure.
1888  * @param[out] pkts
1889  *   Array to store received packets.
1890  * @param pkts_n
1891  *   Maximum number of packets in array.
1892  *
1893  * @return
1894  *   Number of packets successfully received (<= pkts_n).
1895  */
1896 uint16_t
1897 removed_rx_burst(void *dpdk_txq __rte_unused,
1898 		 struct rte_mbuf **pkts __rte_unused,
1899 		 uint16_t pkts_n __rte_unused)
1900 {
1901 	rte_mb();
1902 	return 0;
1903 }
1904 
1905 /*
1906  * Vectorized Rx/Tx routines are not compiled in when required vector
1907  * instructions are not supported on a target architecture. The following null
1908  * stubs are needed for linkage when those are not included outside of this file
1909  * (e.g.  mlx5_rxtx_vec_sse.c for x86).
1910  */
1911 
1912 __rte_weak uint16_t
1913 mlx5_rx_burst_vec(void *dpdk_txq __rte_unused,
1914 		  struct rte_mbuf **pkts __rte_unused,
1915 		  uint16_t pkts_n __rte_unused)
1916 {
1917 	return 0;
1918 }
1919 
1920 __rte_weak int
1921 mlx5_rxq_check_vec_support(struct mlx5_rxq_data *rxq __rte_unused)
1922 {
1923 	return -ENOTSUP;
1924 }
1925 
1926 __rte_weak int
1927 mlx5_check_vec_rx_support(struct rte_eth_dev *dev __rte_unused)
1928 {
1929 	return -ENOTSUP;
1930 }
1931 
1932 /**
1933  * Free the mbufs from the linear array of pointers.
1934  *
1935  * @param pkts
1936  *   Pointer to array of packets to be free.
1937  * @param pkts_n
1938  *   Number of packets to be freed.
1939  * @param olx
1940  *   Configured Tx offloads mask. It is fully defined at
1941  *   compile time and may be used for optimization.
1942  */
1943 static __rte_always_inline void
1944 mlx5_tx_free_mbuf(struct rte_mbuf **restrict pkts,
1945 		  unsigned int pkts_n,
1946 		  unsigned int olx __rte_unused)
1947 {
1948 	struct rte_mempool *pool = NULL;
1949 	struct rte_mbuf **p_free = NULL;
1950 	struct rte_mbuf *mbuf;
1951 	unsigned int n_free = 0;
1952 
1953 	/*
1954 	 * The implemented algorithm eliminates
1955 	 * copying pointers to temporary array
1956 	 * for rte_mempool_put_bulk() calls.
1957 	 */
1958 	MLX5_ASSERT(pkts);
1959 	MLX5_ASSERT(pkts_n);
1960 	for (;;) {
1961 		for (;;) {
1962 			/*
1963 			 * Decrement mbuf reference counter, detach
1964 			 * indirect and external buffers if needed.
1965 			 */
1966 			mbuf = rte_pktmbuf_prefree_seg(*pkts);
1967 			if (likely(mbuf != NULL)) {
1968 				MLX5_ASSERT(mbuf == *pkts);
1969 				if (likely(n_free != 0)) {
1970 					if (unlikely(pool != mbuf->pool))
1971 						/* From different pool. */
1972 						break;
1973 				} else {
1974 					/* Start new scan array. */
1975 					pool = mbuf->pool;
1976 					p_free = pkts;
1977 				}
1978 				++n_free;
1979 				++pkts;
1980 				--pkts_n;
1981 				if (unlikely(pkts_n == 0)) {
1982 					mbuf = NULL;
1983 					break;
1984 				}
1985 			} else {
1986 				/*
1987 				 * This happens if mbuf is still referenced.
1988 				 * We can't put it back to the pool, skip.
1989 				 */
1990 				++pkts;
1991 				--pkts_n;
1992 				if (unlikely(n_free != 0))
1993 					/* There is some array to free.*/
1994 					break;
1995 				if (unlikely(pkts_n == 0))
1996 					/* Last mbuf, nothing to free. */
1997 					return;
1998 			}
1999 		}
2000 		for (;;) {
2001 			/*
2002 			 * This loop is implemented to avoid multiple
2003 			 * inlining of rte_mempool_put_bulk().
2004 			 */
2005 			MLX5_ASSERT(pool);
2006 			MLX5_ASSERT(p_free);
2007 			MLX5_ASSERT(n_free);
2008 			/*
2009 			 * Free the array of pre-freed mbufs
2010 			 * belonging to the same memory pool.
2011 			 */
2012 			rte_mempool_put_bulk(pool, (void *)p_free, n_free);
2013 			if (unlikely(mbuf != NULL)) {
2014 				/* There is the request to start new scan. */
2015 				pool = mbuf->pool;
2016 				p_free = pkts++;
2017 				n_free = 1;
2018 				--pkts_n;
2019 				if (likely(pkts_n != 0))
2020 					break;
2021 				/*
2022 				 * This is the last mbuf to be freed.
2023 				 * Do one more loop iteration to complete.
2024 				 * This is rare case of the last unique mbuf.
2025 				 */
2026 				mbuf = NULL;
2027 				continue;
2028 			}
2029 			if (likely(pkts_n == 0))
2030 				return;
2031 			n_free = 0;
2032 			break;
2033 		}
2034 	}
2035 }
2036 
2037 /**
2038  * Free the mbuf from the elts ring buffer till new tail.
2039  *
2040  * @param txq
2041  *   Pointer to Tx queue structure.
2042  * @param tail
2043  *   Index in elts to free up to, becomes new elts tail.
2044  * @param olx
2045  *   Configured Tx offloads mask. It is fully defined at
2046  *   compile time and may be used for optimization.
2047  */
2048 static __rte_always_inline void
2049 mlx5_tx_free_elts(struct mlx5_txq_data *restrict txq,
2050 		  uint16_t tail,
2051 		  unsigned int olx __rte_unused)
2052 {
2053 	uint16_t n_elts = tail - txq->elts_tail;
2054 
2055 	MLX5_ASSERT(n_elts);
2056 	MLX5_ASSERT(n_elts <= txq->elts_s);
2057 	/*
2058 	 * Implement a loop to support ring buffer wraparound
2059 	 * with single inlining of mlx5_tx_free_mbuf().
2060 	 */
2061 	do {
2062 		unsigned int part;
2063 
2064 		part = txq->elts_s - (txq->elts_tail & txq->elts_m);
2065 		part = RTE_MIN(part, n_elts);
2066 		MLX5_ASSERT(part);
2067 		MLX5_ASSERT(part <= txq->elts_s);
2068 		mlx5_tx_free_mbuf(&txq->elts[txq->elts_tail & txq->elts_m],
2069 				  part, olx);
2070 		txq->elts_tail += part;
2071 		n_elts -= part;
2072 	} while (n_elts);
2073 }
2074 
2075 /**
2076  * Store the mbuf being sent into elts ring buffer.
2077  * On Tx completion these mbufs will be freed.
2078  *
2079  * @param txq
2080  *   Pointer to Tx queue structure.
2081  * @param pkts
2082  *   Pointer to array of packets to be stored.
2083  * @param pkts_n
2084  *   Number of packets to be stored.
2085  * @param olx
2086  *   Configured Tx offloads mask. It is fully defined at
2087  *   compile time and may be used for optimization.
2088  */
2089 static __rte_always_inline void
2090 mlx5_tx_copy_elts(struct mlx5_txq_data *restrict txq,
2091 		  struct rte_mbuf **restrict pkts,
2092 		  unsigned int pkts_n,
2093 		  unsigned int olx __rte_unused)
2094 {
2095 	unsigned int part;
2096 	struct rte_mbuf **elts = (struct rte_mbuf **)txq->elts;
2097 
2098 	MLX5_ASSERT(pkts);
2099 	MLX5_ASSERT(pkts_n);
2100 	part = txq->elts_s - (txq->elts_head & txq->elts_m);
2101 	MLX5_ASSERT(part);
2102 	MLX5_ASSERT(part <= txq->elts_s);
2103 	/* This code is a good candidate for vectorizing with SIMD. */
2104 	rte_memcpy((void *)(elts + (txq->elts_head & txq->elts_m)),
2105 		   (void *)pkts,
2106 		   RTE_MIN(part, pkts_n) * sizeof(struct rte_mbuf *));
2107 	txq->elts_head += pkts_n;
2108 	if (unlikely(part < pkts_n))
2109 		/* The copy is wrapping around the elts array. */
2110 		rte_memcpy((void *)elts, (void *)(pkts + part),
2111 			   (pkts_n - part) * sizeof(struct rte_mbuf *));
2112 }
2113 
2114 /**
2115  * Update completion queue consuming index via doorbell
2116  * and flush the completed data buffers.
2117  *
2118  * @param txq
2119  *   Pointer to TX queue structure.
2120  * @param valid CQE pointer
2121  *   if not NULL update txq->wqe_pi and flush the buffers
2122  * @param olx
2123  *   Configured Tx offloads mask. It is fully defined at
2124  *   compile time and may be used for optimization.
2125  */
2126 static __rte_always_inline void
2127 mlx5_tx_comp_flush(struct mlx5_txq_data *restrict txq,
2128 		   volatile struct mlx5_cqe *last_cqe,
2129 		   unsigned int olx __rte_unused)
2130 {
2131 	if (likely(last_cqe != NULL)) {
2132 		uint16_t tail;
2133 
2134 		txq->wqe_pi = rte_be_to_cpu_16(last_cqe->wqe_counter);
2135 		tail = txq->fcqs[(txq->cq_ci - 1) & txq->cqe_m];
2136 		if (likely(tail != txq->elts_tail)) {
2137 			mlx5_tx_free_elts(txq, tail, olx);
2138 			MLX5_ASSERT(tail == txq->elts_tail);
2139 		}
2140 	}
2141 }
2142 
2143 /**
2144  * Manage TX completions. This routine checks the CQ for
2145  * arrived CQEs, deduces the last accomplished WQE in SQ,
2146  * updates SQ producing index and frees all completed mbufs.
2147  *
2148  * @param txq
2149  *   Pointer to TX queue structure.
2150  * @param olx
2151  *   Configured Tx offloads mask. It is fully defined at
2152  *   compile time and may be used for optimization.
2153  *
2154  * NOTE: not inlined intentionally, it makes tx_burst
2155  * routine smaller, simple and faster - from experiments.
2156  */
2157 static void
2158 mlx5_tx_handle_completion(struct mlx5_txq_data *restrict txq,
2159 			  unsigned int olx __rte_unused)
2160 {
2161 	unsigned int count = MLX5_TX_COMP_MAX_CQE;
2162 	volatile struct mlx5_cqe *last_cqe = NULL;
2163 	uint16_t ci = txq->cq_ci;
2164 	int ret;
2165 
2166 	static_assert(MLX5_CQE_STATUS_HW_OWN < 0, "Must be negative value");
2167 	static_assert(MLX5_CQE_STATUS_SW_OWN < 0, "Must be negative value");
2168 	do {
2169 		volatile struct mlx5_cqe *cqe;
2170 
2171 		cqe = &txq->cqes[ci & txq->cqe_m];
2172 		ret = check_cqe(cqe, txq->cqe_s, ci);
2173 		if (unlikely(ret != MLX5_CQE_STATUS_SW_OWN)) {
2174 			if (likely(ret != MLX5_CQE_STATUS_ERR)) {
2175 				/* No new CQEs in completion queue. */
2176 				MLX5_ASSERT(ret == MLX5_CQE_STATUS_HW_OWN);
2177 				break;
2178 			}
2179 			/*
2180 			 * Some error occurred, try to restart.
2181 			 * We have no barrier after WQE related Doorbell
2182 			 * written, make sure all writes are completed
2183 			 * here, before we might perform SQ reset.
2184 			 */
2185 			rte_wmb();
2186 			txq->cq_ci = ci;
2187 			ret = mlx5_tx_error_cqe_handle
2188 				(txq, (volatile struct mlx5_err_cqe *)cqe);
2189 			if (unlikely(ret < 0)) {
2190 				/*
2191 				 * Some error occurred on queue error
2192 				 * handling, we do not advance the index
2193 				 * here, allowing to retry on next call.
2194 				 */
2195 				return;
2196 			}
2197 			/*
2198 			 * We are going to fetch all entries with
2199 			 * MLX5_CQE_SYNDROME_WR_FLUSH_ERR status.
2200 			 * The send queue is supposed to be empty.
2201 			 */
2202 			++ci;
2203 			txq->cq_pi = ci;
2204 			last_cqe = NULL;
2205 			continue;
2206 		}
2207 		/* Normal transmit completion. */
2208 		MLX5_ASSERT(ci != txq->cq_pi);
2209 		MLX5_ASSERT((txq->fcqs[ci & txq->cqe_m] >> 16) ==
2210 			    cqe->wqe_counter);
2211 		++ci;
2212 		last_cqe = cqe;
2213 		/*
2214 		 * We have to restrict the amount of processed CQEs
2215 		 * in one tx_burst routine call. The CQ may be large
2216 		 * and many CQEs may be updated by the NIC in one
2217 		 * transaction. Buffers freeing is time consuming,
2218 		 * multiple iterations may introduce significant
2219 		 * latency.
2220 		 */
2221 		if (likely(--count == 0))
2222 			break;
2223 	} while (true);
2224 	if (likely(ci != txq->cq_ci)) {
2225 		/*
2226 		 * Update completion queue consuming index
2227 		 * and ring doorbell to notify hardware.
2228 		 */
2229 		rte_compiler_barrier();
2230 		txq->cq_ci = ci;
2231 		*txq->cq_db = rte_cpu_to_be_32(ci);
2232 		mlx5_tx_comp_flush(txq, last_cqe, olx);
2233 	}
2234 }
2235 
2236 /**
2237  * Check if the completion request flag should be set in the last WQE.
2238  * Both pushed mbufs and WQEs are monitored and the completion request
2239  * flag is set if any of thresholds is reached.
2240  *
2241  * @param txq
2242  *   Pointer to TX queue structure.
2243  * @param loc
2244  *   Pointer to burst routine local context.
2245  * @param olx
2246  *   Configured Tx offloads mask. It is fully defined at
2247  *   compile time and may be used for optimization.
2248  */
2249 static __rte_always_inline void
2250 mlx5_tx_request_completion(struct mlx5_txq_data *restrict txq,
2251 			   struct mlx5_txq_local *restrict loc,
2252 			   unsigned int olx)
2253 {
2254 	uint16_t head = txq->elts_head;
2255 	unsigned int part;
2256 
2257 	part = MLX5_TXOFF_CONFIG(INLINE) ?
2258 	       0 : loc->pkts_sent - loc->pkts_copy;
2259 	head += part;
2260 	if ((uint16_t)(head - txq->elts_comp) >= MLX5_TX_COMP_THRESH ||
2261 	     (MLX5_TXOFF_CONFIG(INLINE) &&
2262 	     (uint16_t)(txq->wqe_ci - txq->wqe_comp) >= txq->wqe_thres)) {
2263 		volatile struct mlx5_wqe *last = loc->wqe_last;
2264 
2265 		txq->elts_comp = head;
2266 		if (MLX5_TXOFF_CONFIG(INLINE))
2267 			txq->wqe_comp = txq->wqe_ci;
2268 		/* Request unconditional completion on last WQE. */
2269 		last->cseg.flags = RTE_BE32(MLX5_COMP_ALWAYS <<
2270 					    MLX5_COMP_MODE_OFFSET);
2271 		/* Save elts_head in dedicated free on completion queue. */
2272 #ifdef RTE_LIBRTE_MLX5_DEBUG
2273 		txq->fcqs[txq->cq_pi++ & txq->cqe_m] = head |
2274 			  (last->cseg.opcode >> 8) << 16;
2275 #else
2276 		txq->fcqs[txq->cq_pi++ & txq->cqe_m] = head;
2277 #endif
2278 		/* A CQE slot must always be available. */
2279 		MLX5_ASSERT((txq->cq_pi - txq->cq_ci) <= txq->cqe_s);
2280 	}
2281 }
2282 
2283 /**
2284  * DPDK callback to check the status of a tx descriptor.
2285  *
2286  * @param tx_queue
2287  *   The tx queue.
2288  * @param[in] offset
2289  *   The index of the descriptor in the ring.
2290  *
2291  * @return
2292  *   The status of the tx descriptor.
2293  */
2294 int
2295 mlx5_tx_descriptor_status(void *tx_queue, uint16_t offset)
2296 {
2297 	struct mlx5_txq_data *restrict txq = tx_queue;
2298 	uint16_t used;
2299 
2300 	mlx5_tx_handle_completion(txq, 0);
2301 	used = txq->elts_head - txq->elts_tail;
2302 	if (offset < used)
2303 		return RTE_ETH_TX_DESC_FULL;
2304 	return RTE_ETH_TX_DESC_DONE;
2305 }
2306 
2307 /**
2308  * Build the Control Segment with specified opcode:
2309  * - MLX5_OPCODE_SEND
2310  * - MLX5_OPCODE_ENHANCED_MPSW
2311  * - MLX5_OPCODE_TSO
2312  *
2313  * @param txq
2314  *   Pointer to TX queue structure.
2315  * @param loc
2316  *   Pointer to burst routine local context.
2317  * @param wqe
2318  *   Pointer to WQE to fill with built Control Segment.
2319  * @param ds
2320  *   Supposed length of WQE in segments.
2321  * @param opcode
2322  *   SQ WQE opcode to put into Control Segment.
2323  * @param olx
2324  *   Configured Tx offloads mask. It is fully defined at
2325  *   compile time and may be used for optimization.
2326  */
2327 static __rte_always_inline void
2328 mlx5_tx_cseg_init(struct mlx5_txq_data *restrict txq,
2329 		  struct mlx5_txq_local *restrict loc __rte_unused,
2330 		  struct mlx5_wqe *restrict wqe,
2331 		  unsigned int ds,
2332 		  unsigned int opcode,
2333 		  unsigned int olx __rte_unused)
2334 {
2335 	struct mlx5_wqe_cseg *restrict cs = &wqe->cseg;
2336 
2337 	/* For legacy MPW replace the EMPW by TSO with modifier. */
2338 	if (MLX5_TXOFF_CONFIG(MPW) && opcode == MLX5_OPCODE_ENHANCED_MPSW)
2339 		opcode = MLX5_OPCODE_TSO | MLX5_OPC_MOD_MPW << 24;
2340 	cs->opcode = rte_cpu_to_be_32((txq->wqe_ci << 8) | opcode);
2341 	cs->sq_ds = rte_cpu_to_be_32(txq->qp_num_8s | ds);
2342 	cs->flags = RTE_BE32(MLX5_COMP_ONLY_FIRST_ERR <<
2343 			     MLX5_COMP_MODE_OFFSET);
2344 	cs->misc = RTE_BE32(0);
2345 }
2346 
2347 /**
2348  * Build the Ethernet Segment without inlined data.
2349  * Supports Software Parser, Checksums and VLAN
2350  * insertion Tx offload features.
2351  *
2352  * @param txq
2353  *   Pointer to TX queue structure.
2354  * @param loc
2355  *   Pointer to burst routine local context.
2356  * @param wqe
2357  *   Pointer to WQE to fill with built Ethernet Segment.
2358  * @param olx
2359  *   Configured Tx offloads mask. It is fully defined at
2360  *   compile time and may be used for optimization.
2361  */
2362 static __rte_always_inline void
2363 mlx5_tx_eseg_none(struct mlx5_txq_data *restrict txq __rte_unused,
2364 		  struct mlx5_txq_local *restrict loc,
2365 		  struct mlx5_wqe *restrict wqe,
2366 		  unsigned int olx)
2367 {
2368 	struct mlx5_wqe_eseg *restrict es = &wqe->eseg;
2369 	uint32_t csum;
2370 
2371 	/*
2372 	 * Calculate and set check sum flags first, dword field
2373 	 * in segment may be shared with Software Parser flags.
2374 	 */
2375 	csum = MLX5_TXOFF_CONFIG(CSUM) ? txq_ol_cksum_to_cs(loc->mbuf) : 0;
2376 	es->flags = rte_cpu_to_le_32(csum);
2377 	/*
2378 	 * Calculate and set Software Parser offsets and flags.
2379 	 * These flags a set for custom UDP and IP tunnel packets.
2380 	 */
2381 	es->swp_offs = txq_mbuf_to_swp(loc, &es->swp_flags, olx);
2382 	/* Fill metadata field if needed. */
2383 	es->metadata = MLX5_TXOFF_CONFIG(METADATA) ?
2384 		       loc->mbuf->ol_flags & PKT_TX_DYNF_METADATA ?
2385 		       *RTE_FLOW_DYNF_METADATA(loc->mbuf) : 0 : 0;
2386 	/* Engage VLAN tag insertion feature if requested. */
2387 	if (MLX5_TXOFF_CONFIG(VLAN) &&
2388 	    loc->mbuf->ol_flags & PKT_TX_VLAN_PKT) {
2389 		/*
2390 		 * We should get here only if device support
2391 		 * this feature correctly.
2392 		 */
2393 		MLX5_ASSERT(txq->vlan_en);
2394 		es->inline_hdr = rte_cpu_to_be_32(MLX5_ETH_WQE_VLAN_INSERT |
2395 						  loc->mbuf->vlan_tci);
2396 	} else {
2397 		es->inline_hdr = RTE_BE32(0);
2398 	}
2399 }
2400 
2401 /**
2402  * Build the Ethernet Segment with minimal inlined data
2403  * of MLX5_ESEG_MIN_INLINE_SIZE bytes length. This is
2404  * used to fill the gap in single WQEBB WQEs.
2405  * Supports Software Parser, Checksums and VLAN
2406  * insertion Tx offload features.
2407  *
2408  * @param txq
2409  *   Pointer to TX queue structure.
2410  * @param loc
2411  *   Pointer to burst routine local context.
2412  * @param wqe
2413  *   Pointer to WQE to fill with built Ethernet Segment.
2414  * @param vlan
2415  *   Length of VLAN tag insertion if any.
2416  * @param olx
2417  *   Configured Tx offloads mask. It is fully defined at
2418  *   compile time and may be used for optimization.
2419  */
2420 static __rte_always_inline void
2421 mlx5_tx_eseg_dmin(struct mlx5_txq_data *restrict txq __rte_unused,
2422 		  struct mlx5_txq_local *restrict loc,
2423 		  struct mlx5_wqe *restrict wqe,
2424 		  unsigned int vlan,
2425 		  unsigned int olx)
2426 {
2427 	struct mlx5_wqe_eseg *restrict es = &wqe->eseg;
2428 	uint32_t csum;
2429 	uint8_t *psrc, *pdst;
2430 
2431 	/*
2432 	 * Calculate and set check sum flags first, dword field
2433 	 * in segment may be shared with Software Parser flags.
2434 	 */
2435 	csum = MLX5_TXOFF_CONFIG(CSUM) ? txq_ol_cksum_to_cs(loc->mbuf) : 0;
2436 	es->flags = rte_cpu_to_le_32(csum);
2437 	/*
2438 	 * Calculate and set Software Parser offsets and flags.
2439 	 * These flags a set for custom UDP and IP tunnel packets.
2440 	 */
2441 	es->swp_offs = txq_mbuf_to_swp(loc, &es->swp_flags, olx);
2442 	/* Fill metadata field if needed. */
2443 	es->metadata = MLX5_TXOFF_CONFIG(METADATA) ?
2444 		       loc->mbuf->ol_flags & PKT_TX_DYNF_METADATA ?
2445 		       *RTE_FLOW_DYNF_METADATA(loc->mbuf) : 0 : 0;
2446 	static_assert(MLX5_ESEG_MIN_INLINE_SIZE ==
2447 				(sizeof(uint16_t) +
2448 				 sizeof(rte_v128u32_t)),
2449 		      "invalid Ethernet Segment data size");
2450 	static_assert(MLX5_ESEG_MIN_INLINE_SIZE ==
2451 				(sizeof(uint16_t) +
2452 				 sizeof(struct rte_vlan_hdr) +
2453 				 2 * RTE_ETHER_ADDR_LEN),
2454 		      "invalid Ethernet Segment data size");
2455 	psrc = rte_pktmbuf_mtod(loc->mbuf, uint8_t *);
2456 	es->inline_hdr_sz = RTE_BE16(MLX5_ESEG_MIN_INLINE_SIZE);
2457 	es->inline_data = *(unaligned_uint16_t *)psrc;
2458 	psrc +=	sizeof(uint16_t);
2459 	pdst = (uint8_t *)(es + 1);
2460 	if (MLX5_TXOFF_CONFIG(VLAN) && vlan) {
2461 		/* Implement VLAN tag insertion as part inline data. */
2462 		memcpy(pdst, psrc, 2 * RTE_ETHER_ADDR_LEN - sizeof(uint16_t));
2463 		pdst += 2 * RTE_ETHER_ADDR_LEN - sizeof(uint16_t);
2464 		psrc +=	2 * RTE_ETHER_ADDR_LEN - sizeof(uint16_t);
2465 		/* Insert VLAN ethertype + VLAN tag. */
2466 		*(unaligned_uint32_t *)pdst = rte_cpu_to_be_32
2467 						((RTE_ETHER_TYPE_VLAN << 16) |
2468 						 loc->mbuf->vlan_tci);
2469 		pdst += sizeof(struct rte_vlan_hdr);
2470 		/* Copy the rest two bytes from packet data. */
2471 		MLX5_ASSERT(pdst == RTE_PTR_ALIGN(pdst, sizeof(uint16_t)));
2472 		*(uint16_t *)pdst = *(unaligned_uint16_t *)psrc;
2473 	} else {
2474 		/* Fill the gap in the title WQEBB with inline data. */
2475 		rte_mov16(pdst, psrc);
2476 	}
2477 }
2478 
2479 /**
2480  * Build the Ethernet Segment with entire packet
2481  * data inlining. Checks the boundary of WQEBB and
2482  * ring buffer wrapping, supports Software Parser,
2483  * Checksums and VLAN insertion Tx offload features.
2484  *
2485  * @param txq
2486  *   Pointer to TX queue structure.
2487  * @param loc
2488  *   Pointer to burst routine local context.
2489  * @param wqe
2490  *   Pointer to WQE to fill with built Ethernet Segment.
2491  * @param vlan
2492  *   Length of VLAN tag insertion if any.
2493  * @param inlen
2494  *   Length of data to inline (VLAN included, if any).
2495  * @param tso
2496  *   TSO flag, set mss field from the packet.
2497  * @param olx
2498  *   Configured Tx offloads mask. It is fully defined at
2499  *   compile time and may be used for optimization.
2500  *
2501  * @return
2502  *   Pointer to the next Data Segment (aligned and wrapped around).
2503  */
2504 static __rte_always_inline struct mlx5_wqe_dseg *
2505 mlx5_tx_eseg_data(struct mlx5_txq_data *restrict txq,
2506 		  struct mlx5_txq_local *restrict loc,
2507 		  struct mlx5_wqe *restrict wqe,
2508 		  unsigned int vlan,
2509 		  unsigned int inlen,
2510 		  unsigned int tso,
2511 		  unsigned int olx)
2512 {
2513 	struct mlx5_wqe_eseg *restrict es = &wqe->eseg;
2514 	uint32_t csum;
2515 	uint8_t *psrc, *pdst;
2516 	unsigned int part;
2517 
2518 	/*
2519 	 * Calculate and set check sum flags first, dword field
2520 	 * in segment may be shared with Software Parser flags.
2521 	 */
2522 	csum = MLX5_TXOFF_CONFIG(CSUM) ? txq_ol_cksum_to_cs(loc->mbuf) : 0;
2523 	if (tso) {
2524 		csum <<= 24;
2525 		csum |= loc->mbuf->tso_segsz;
2526 		es->flags = rte_cpu_to_be_32(csum);
2527 	} else {
2528 		es->flags = rte_cpu_to_le_32(csum);
2529 	}
2530 	/*
2531 	 * Calculate and set Software Parser offsets and flags.
2532 	 * These flags a set for custom UDP and IP tunnel packets.
2533 	 */
2534 	es->swp_offs = txq_mbuf_to_swp(loc, &es->swp_flags, olx);
2535 	/* Fill metadata field if needed. */
2536 	es->metadata = MLX5_TXOFF_CONFIG(METADATA) ?
2537 		       loc->mbuf->ol_flags & PKT_TX_DYNF_METADATA ?
2538 		       *RTE_FLOW_DYNF_METADATA(loc->mbuf) : 0 : 0;
2539 	static_assert(MLX5_ESEG_MIN_INLINE_SIZE ==
2540 				(sizeof(uint16_t) +
2541 				 sizeof(rte_v128u32_t)),
2542 		      "invalid Ethernet Segment data size");
2543 	static_assert(MLX5_ESEG_MIN_INLINE_SIZE ==
2544 				(sizeof(uint16_t) +
2545 				 sizeof(struct rte_vlan_hdr) +
2546 				 2 * RTE_ETHER_ADDR_LEN),
2547 		      "invalid Ethernet Segment data size");
2548 	psrc = rte_pktmbuf_mtod(loc->mbuf, uint8_t *);
2549 	es->inline_hdr_sz = rte_cpu_to_be_16(inlen);
2550 	es->inline_data = *(unaligned_uint16_t *)psrc;
2551 	psrc +=	sizeof(uint16_t);
2552 	pdst = (uint8_t *)(es + 1);
2553 	if (MLX5_TXOFF_CONFIG(VLAN) && vlan) {
2554 		/* Implement VLAN tag insertion as part inline data. */
2555 		memcpy(pdst, psrc, 2 * RTE_ETHER_ADDR_LEN - sizeof(uint16_t));
2556 		pdst += 2 * RTE_ETHER_ADDR_LEN - sizeof(uint16_t);
2557 		psrc +=	2 * RTE_ETHER_ADDR_LEN - sizeof(uint16_t);
2558 		/* Insert VLAN ethertype + VLAN tag. */
2559 		*(unaligned_uint32_t *)pdst = rte_cpu_to_be_32
2560 						((RTE_ETHER_TYPE_VLAN << 16) |
2561 						 loc->mbuf->vlan_tci);
2562 		pdst += sizeof(struct rte_vlan_hdr);
2563 		/* Copy the rest two bytes from packet data. */
2564 		MLX5_ASSERT(pdst == RTE_PTR_ALIGN(pdst, sizeof(uint16_t)));
2565 		*(uint16_t *)pdst = *(unaligned_uint16_t *)psrc;
2566 		psrc += sizeof(uint16_t);
2567 	} else {
2568 		/* Fill the gap in the title WQEBB with inline data. */
2569 		rte_mov16(pdst, psrc);
2570 		psrc += sizeof(rte_v128u32_t);
2571 	}
2572 	pdst = (uint8_t *)(es + 2);
2573 	MLX5_ASSERT(inlen >= MLX5_ESEG_MIN_INLINE_SIZE);
2574 	MLX5_ASSERT(pdst < (uint8_t *)txq->wqes_end);
2575 	inlen -= MLX5_ESEG_MIN_INLINE_SIZE;
2576 	if (!inlen) {
2577 		MLX5_ASSERT(pdst == RTE_PTR_ALIGN(pdst, MLX5_WSEG_SIZE));
2578 		return (struct mlx5_wqe_dseg *)pdst;
2579 	}
2580 	/*
2581 	 * The WQEBB space availability is checked by caller.
2582 	 * Here we should be aware of WQE ring buffer wraparound only.
2583 	 */
2584 	part = (uint8_t *)txq->wqes_end - pdst;
2585 	part = RTE_MIN(part, inlen);
2586 	do {
2587 		rte_memcpy(pdst, psrc, part);
2588 		inlen -= part;
2589 		if (likely(!inlen)) {
2590 			/*
2591 			 * If return value is not used by the caller
2592 			 * the code below will be optimized out.
2593 			 */
2594 			pdst += part;
2595 			pdst = RTE_PTR_ALIGN(pdst, MLX5_WSEG_SIZE);
2596 			if (unlikely(pdst >= (uint8_t *)txq->wqes_end))
2597 				pdst = (uint8_t *)txq->wqes;
2598 			return (struct mlx5_wqe_dseg *)pdst;
2599 		}
2600 		pdst = (uint8_t *)txq->wqes;
2601 		psrc += part;
2602 		part = inlen;
2603 	} while (true);
2604 }
2605 
2606 /**
2607  * Copy data from chain of mbuf to the specified linear buffer.
2608  * Checksums and VLAN insertion Tx offload features. If data
2609  * from some mbuf copied completely this mbuf is freed. Local
2610  * structure is used to keep the byte stream state.
2611  *
2612  * @param pdst
2613  *   Pointer to the destination linear buffer.
2614  * @param loc
2615  *   Pointer to burst routine local context.
2616  * @param len
2617  *   Length of data to be copied.
2618  * @param must
2619  *   Length of data to be copied ignoring no inline hint.
2620  * @param olx
2621  *   Configured Tx offloads mask. It is fully defined at
2622  *   compile time and may be used for optimization.
2623  *
2624  * @return
2625  *   Number of actual copied data bytes. This is always greater than or
2626  *   equal to must parameter and might be lesser than len in no inline
2627  *   hint flag is encountered.
2628  */
2629 static __rte_always_inline unsigned int
2630 mlx5_tx_mseg_memcpy(uint8_t *pdst,
2631 		    struct mlx5_txq_local *restrict loc,
2632 		    unsigned int len,
2633 		    unsigned int must,
2634 		    unsigned int olx __rte_unused)
2635 {
2636 	struct rte_mbuf *mbuf;
2637 	unsigned int part, dlen, copy = 0;
2638 	uint8_t *psrc;
2639 
2640 	MLX5_ASSERT(len);
2641 	MLX5_ASSERT(must <= len);
2642 	do {
2643 		/* Allow zero length packets, must check first. */
2644 		dlen = rte_pktmbuf_data_len(loc->mbuf);
2645 		if (dlen <= loc->mbuf_off) {
2646 			/* Exhausted packet, just free. */
2647 			mbuf = loc->mbuf;
2648 			loc->mbuf = mbuf->next;
2649 			rte_pktmbuf_free_seg(mbuf);
2650 			loc->mbuf_off = 0;
2651 			MLX5_ASSERT(loc->mbuf_nseg > 1);
2652 			MLX5_ASSERT(loc->mbuf);
2653 			--loc->mbuf_nseg;
2654 			if (loc->mbuf->ol_flags & PKT_TX_DYNF_NOINLINE) {
2655 				unsigned int diff;
2656 
2657 				if (copy >= must) {
2658 					/*
2659 					 * We already copied the minimal
2660 					 * requested amount of data.
2661 					 */
2662 					return copy;
2663 				}
2664 				diff = must - copy;
2665 				if (diff <= rte_pktmbuf_data_len(loc->mbuf)) {
2666 					/*
2667 					 * Copy only the minimal required
2668 					 * part of the data buffer.
2669 					 */
2670 					len = diff;
2671 				}
2672 			}
2673 			continue;
2674 		}
2675 		dlen -= loc->mbuf_off;
2676 		psrc = rte_pktmbuf_mtod_offset(loc->mbuf, uint8_t *,
2677 					       loc->mbuf_off);
2678 		part = RTE_MIN(len, dlen);
2679 		rte_memcpy(pdst, psrc, part);
2680 		copy += part;
2681 		loc->mbuf_off += part;
2682 		len -= part;
2683 		if (!len) {
2684 			if (loc->mbuf_off >= rte_pktmbuf_data_len(loc->mbuf)) {
2685 				loc->mbuf_off = 0;
2686 				/* Exhausted packet, just free. */
2687 				mbuf = loc->mbuf;
2688 				loc->mbuf = mbuf->next;
2689 				rte_pktmbuf_free_seg(mbuf);
2690 				loc->mbuf_off = 0;
2691 				MLX5_ASSERT(loc->mbuf_nseg >= 1);
2692 				--loc->mbuf_nseg;
2693 			}
2694 			return copy;
2695 		}
2696 		pdst += part;
2697 	} while (true);
2698 }
2699 
2700 /**
2701  * Build the Ethernet Segment with inlined data from
2702  * multi-segment packet. Checks the boundary of WQEBB
2703  * and ring buffer wrapping, supports Software Parser,
2704  * Checksums and VLAN insertion Tx offload features.
2705  *
2706  * @param txq
2707  *   Pointer to TX queue structure.
2708  * @param loc
2709  *   Pointer to burst routine local context.
2710  * @param wqe
2711  *   Pointer to WQE to fill with built Ethernet Segment.
2712  * @param vlan
2713  *   Length of VLAN tag insertion if any.
2714  * @param inlen
2715  *   Length of data to inline (VLAN included, if any).
2716  * @param tso
2717  *   TSO flag, set mss field from the packet.
2718  * @param olx
2719  *   Configured Tx offloads mask. It is fully defined at
2720  *   compile time and may be used for optimization.
2721  *
2722  * @return
2723  *   Pointer to the next Data Segment (aligned and
2724  *   possible NOT wrapped around - caller should do
2725  *   wrapping check on its own).
2726  */
2727 static __rte_always_inline struct mlx5_wqe_dseg *
2728 mlx5_tx_eseg_mdat(struct mlx5_txq_data *restrict txq,
2729 		  struct mlx5_txq_local *restrict loc,
2730 		  struct mlx5_wqe *restrict wqe,
2731 		  unsigned int vlan,
2732 		  unsigned int inlen,
2733 		  unsigned int tso,
2734 		  unsigned int olx)
2735 {
2736 	struct mlx5_wqe_eseg *restrict es = &wqe->eseg;
2737 	uint32_t csum;
2738 	uint8_t *pdst;
2739 	unsigned int part, tlen = 0;
2740 
2741 	/*
2742 	 * Calculate and set check sum flags first, uint32_t field
2743 	 * in segment may be shared with Software Parser flags.
2744 	 */
2745 	csum = MLX5_TXOFF_CONFIG(CSUM) ? txq_ol_cksum_to_cs(loc->mbuf) : 0;
2746 	if (tso) {
2747 		csum <<= 24;
2748 		csum |= loc->mbuf->tso_segsz;
2749 		es->flags = rte_cpu_to_be_32(csum);
2750 	} else {
2751 		es->flags = rte_cpu_to_le_32(csum);
2752 	}
2753 	/*
2754 	 * Calculate and set Software Parser offsets and flags.
2755 	 * These flags a set for custom UDP and IP tunnel packets.
2756 	 */
2757 	es->swp_offs = txq_mbuf_to_swp(loc, &es->swp_flags, olx);
2758 	/* Fill metadata field if needed. */
2759 	es->metadata = MLX5_TXOFF_CONFIG(METADATA) ?
2760 		       loc->mbuf->ol_flags & PKT_TX_DYNF_METADATA ?
2761 		       *RTE_FLOW_DYNF_METADATA(loc->mbuf) : 0 : 0;
2762 	static_assert(MLX5_ESEG_MIN_INLINE_SIZE ==
2763 				(sizeof(uint16_t) +
2764 				 sizeof(rte_v128u32_t)),
2765 		      "invalid Ethernet Segment data size");
2766 	static_assert(MLX5_ESEG_MIN_INLINE_SIZE ==
2767 				(sizeof(uint16_t) +
2768 				 sizeof(struct rte_vlan_hdr) +
2769 				 2 * RTE_ETHER_ADDR_LEN),
2770 		      "invalid Ethernet Segment data size");
2771 	MLX5_ASSERT(inlen >= MLX5_ESEG_MIN_INLINE_SIZE);
2772 	pdst = (uint8_t *)&es->inline_data;
2773 	if (MLX5_TXOFF_CONFIG(VLAN) && vlan) {
2774 		/* Implement VLAN tag insertion as part inline data. */
2775 		mlx5_tx_mseg_memcpy(pdst, loc,
2776 				    2 * RTE_ETHER_ADDR_LEN,
2777 				    2 * RTE_ETHER_ADDR_LEN, olx);
2778 		pdst += 2 * RTE_ETHER_ADDR_LEN;
2779 		*(unaligned_uint32_t *)pdst = rte_cpu_to_be_32
2780 						((RTE_ETHER_TYPE_VLAN << 16) |
2781 						 loc->mbuf->vlan_tci);
2782 		pdst += sizeof(struct rte_vlan_hdr);
2783 		tlen += 2 * RTE_ETHER_ADDR_LEN + sizeof(struct rte_vlan_hdr);
2784 	}
2785 	MLX5_ASSERT(pdst < (uint8_t *)txq->wqes_end);
2786 	/*
2787 	 * The WQEBB space availability is checked by caller.
2788 	 * Here we should be aware of WQE ring buffer wraparound only.
2789 	 */
2790 	part = (uint8_t *)txq->wqes_end - pdst;
2791 	part = RTE_MIN(part, inlen - tlen);
2792 	MLX5_ASSERT(part);
2793 	do {
2794 		unsigned int copy;
2795 
2796 		/*
2797 		 * Copying may be interrupted inside the routine
2798 		 * if run into no inline hint flag.
2799 		 */
2800 		copy = tlen >= txq->inlen_mode ? 0 : (txq->inlen_mode - tlen);
2801 		copy = mlx5_tx_mseg_memcpy(pdst, loc, part, copy, olx);
2802 		tlen += copy;
2803 		if (likely(inlen <= tlen) || copy < part) {
2804 			es->inline_hdr_sz = rte_cpu_to_be_16(tlen);
2805 			pdst += copy;
2806 			pdst = RTE_PTR_ALIGN(pdst, MLX5_WSEG_SIZE);
2807 			return (struct mlx5_wqe_dseg *)pdst;
2808 		}
2809 		pdst = (uint8_t *)txq->wqes;
2810 		part = inlen - tlen;
2811 	} while (true);
2812 }
2813 
2814 /**
2815  * Build the Data Segment of pointer type.
2816  *
2817  * @param txq
2818  *   Pointer to TX queue structure.
2819  * @param loc
2820  *   Pointer to burst routine local context.
2821  * @param dseg
2822  *   Pointer to WQE to fill with built Data Segment.
2823  * @param buf
2824  *   Data buffer to point.
2825  * @param len
2826  *   Data buffer length.
2827  * @param olx
2828  *   Configured Tx offloads mask. It is fully defined at
2829  *   compile time and may be used for optimization.
2830  */
2831 static __rte_always_inline void
2832 mlx5_tx_dseg_ptr(struct mlx5_txq_data *restrict txq,
2833 		 struct mlx5_txq_local *restrict loc,
2834 		 struct mlx5_wqe_dseg *restrict dseg,
2835 		 uint8_t *buf,
2836 		 unsigned int len,
2837 		 unsigned int olx __rte_unused)
2838 
2839 {
2840 	MLX5_ASSERT(len);
2841 	dseg->bcount = rte_cpu_to_be_32(len);
2842 	dseg->lkey = mlx5_tx_mb2mr(txq, loc->mbuf);
2843 	dseg->pbuf = rte_cpu_to_be_64((uintptr_t)buf);
2844 }
2845 
2846 /**
2847  * Build the Data Segment of pointer type or inline
2848  * if data length is less than buffer in minimal
2849  * Data Segment size.
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 WQE to 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 static __rte_always_inline void
2866 mlx5_tx_dseg_iptr(struct mlx5_txq_data *restrict txq,
2867 		  struct mlx5_txq_local *restrict loc,
2868 		  struct mlx5_wqe_dseg *restrict dseg,
2869 		  uint8_t *buf,
2870 		  unsigned int len,
2871 		  unsigned int olx __rte_unused)
2872 
2873 {
2874 	uintptr_t dst, src;
2875 
2876 	MLX5_ASSERT(len);
2877 	if (len > MLX5_DSEG_MIN_INLINE_SIZE) {
2878 		dseg->bcount = rte_cpu_to_be_32(len);
2879 		dseg->lkey = mlx5_tx_mb2mr(txq, loc->mbuf);
2880 		dseg->pbuf = rte_cpu_to_be_64((uintptr_t)buf);
2881 
2882 		return;
2883 	}
2884 	dseg->bcount = rte_cpu_to_be_32(len | MLX5_ETH_WQE_DATA_INLINE);
2885 	/* Unrolled implementation of generic rte_memcpy. */
2886 	dst = (uintptr_t)&dseg->inline_data[0];
2887 	src = (uintptr_t)buf;
2888 	if (len & 0x08) {
2889 #ifdef RTE_ARCH_STRICT_ALIGN
2890 		MLX5_ASSERT(dst == RTE_PTR_ALIGN(dst, sizeof(uint32_t)));
2891 		*(uint32_t *)dst = *(unaligned_uint32_t *)src;
2892 		dst += sizeof(uint32_t);
2893 		src += sizeof(uint32_t);
2894 		*(uint32_t *)dst = *(unaligned_uint32_t *)src;
2895 		dst += sizeof(uint32_t);
2896 		src += sizeof(uint32_t);
2897 #else
2898 		*(uint64_t *)dst = *(unaligned_uint64_t *)src;
2899 		dst += sizeof(uint64_t);
2900 		src += sizeof(uint64_t);
2901 #endif
2902 	}
2903 	if (len & 0x04) {
2904 		*(uint32_t *)dst = *(unaligned_uint32_t *)src;
2905 		dst += sizeof(uint32_t);
2906 		src += sizeof(uint32_t);
2907 	}
2908 	if (len & 0x02) {
2909 		*(uint16_t *)dst = *(unaligned_uint16_t *)src;
2910 		dst += sizeof(uint16_t);
2911 		src += sizeof(uint16_t);
2912 	}
2913 	if (len & 0x01)
2914 		*(uint8_t *)dst = *(uint8_t *)src;
2915 }
2916 
2917 /**
2918  * Build the Data Segment of inlined data from single
2919  * segment packet, no VLAN insertion.
2920  *
2921  * @param txq
2922  *   Pointer to TX queue structure.
2923  * @param loc
2924  *   Pointer to burst routine local context.
2925  * @param dseg
2926  *   Pointer to WQE to fill with built Data Segment.
2927  * @param buf
2928  *   Data buffer to point.
2929  * @param len
2930  *   Data buffer length.
2931  * @param olx
2932  *   Configured Tx offloads mask. It is fully defined at
2933  *   compile time and may be used for optimization.
2934  *
2935  * @return
2936  *   Pointer to the next Data Segment after inlined data.
2937  *   Ring buffer wraparound check is needed. We do not
2938  *   do it here because it may not be needed for the
2939  *   last packet in the eMPW session.
2940  */
2941 static __rte_always_inline struct mlx5_wqe_dseg *
2942 mlx5_tx_dseg_empw(struct mlx5_txq_data *restrict txq,
2943 		  struct mlx5_txq_local *restrict loc __rte_unused,
2944 		  struct mlx5_wqe_dseg *restrict dseg,
2945 		  uint8_t *buf,
2946 		  unsigned int len,
2947 		  unsigned int olx __rte_unused)
2948 {
2949 	unsigned int part;
2950 	uint8_t *pdst;
2951 
2952 	dseg->bcount = rte_cpu_to_be_32(len | MLX5_ETH_WQE_DATA_INLINE);
2953 	pdst = &dseg->inline_data[0];
2954 	/*
2955 	 * The WQEBB space availability is checked by caller.
2956 	 * Here we should be aware of WQE ring buffer wraparound only.
2957 	 */
2958 	part = (uint8_t *)txq->wqes_end - pdst;
2959 	part = RTE_MIN(part, len);
2960 	do {
2961 		rte_memcpy(pdst, buf, part);
2962 		len -= part;
2963 		if (likely(!len)) {
2964 			pdst += part;
2965 			pdst = RTE_PTR_ALIGN(pdst, MLX5_WSEG_SIZE);
2966 			/* Note: no final wraparound check here. */
2967 			return (struct mlx5_wqe_dseg *)pdst;
2968 		}
2969 		pdst = (uint8_t *)txq->wqes;
2970 		buf += part;
2971 		part = len;
2972 	} while (true);
2973 }
2974 
2975 /**
2976  * Build the Data Segment of inlined data from single
2977  * segment packet with VLAN insertion.
2978  *
2979  * @param txq
2980  *   Pointer to TX queue structure.
2981  * @param loc
2982  *   Pointer to burst routine local context.
2983  * @param dseg
2984  *   Pointer to the dseg fill with built Data Segment.
2985  * @param buf
2986  *   Data buffer to point.
2987  * @param len
2988  *   Data buffer length.
2989  * @param olx
2990  *   Configured Tx offloads mask. It is fully defined at
2991  *   compile time and may be used for optimization.
2992  *
2993  * @return
2994  *   Pointer to the next Data Segment after inlined data.
2995  *   Ring buffer wraparound check is needed.
2996  */
2997 static __rte_always_inline struct mlx5_wqe_dseg *
2998 mlx5_tx_dseg_vlan(struct mlx5_txq_data *restrict txq,
2999 		  struct mlx5_txq_local *restrict loc __rte_unused,
3000 		  struct mlx5_wqe_dseg *restrict dseg,
3001 		  uint8_t *buf,
3002 		  unsigned int len,
3003 		  unsigned int olx __rte_unused)
3004 
3005 {
3006 	unsigned int part;
3007 	uint8_t *pdst;
3008 
3009 	MLX5_ASSERT(len > MLX5_ESEG_MIN_INLINE_SIZE);
3010 	static_assert(MLX5_DSEG_MIN_INLINE_SIZE ==
3011 				 (2 * RTE_ETHER_ADDR_LEN),
3012 		      "invalid Data Segment data size");
3013 	dseg->bcount = rte_cpu_to_be_32((len + sizeof(struct rte_vlan_hdr)) |
3014 					MLX5_ETH_WQE_DATA_INLINE);
3015 	pdst = &dseg->inline_data[0];
3016 	memcpy(pdst, buf, MLX5_DSEG_MIN_INLINE_SIZE);
3017 	buf += MLX5_DSEG_MIN_INLINE_SIZE;
3018 	pdst += MLX5_DSEG_MIN_INLINE_SIZE;
3019 	len -= MLX5_DSEG_MIN_INLINE_SIZE;
3020 	/* Insert VLAN ethertype + VLAN tag. Pointer is aligned. */
3021 	MLX5_ASSERT(pdst == RTE_PTR_ALIGN(pdst, MLX5_WSEG_SIZE));
3022 	if (unlikely(pdst >= (uint8_t *)txq->wqes_end))
3023 		pdst = (uint8_t *)txq->wqes;
3024 	*(uint32_t *)pdst = rte_cpu_to_be_32((RTE_ETHER_TYPE_VLAN << 16) |
3025 					      loc->mbuf->vlan_tci);
3026 	pdst += sizeof(struct rte_vlan_hdr);
3027 	/*
3028 	 * The WQEBB space availability is checked by caller.
3029 	 * Here we should be aware of WQE ring buffer wraparound only.
3030 	 */
3031 	part = (uint8_t *)txq->wqes_end - pdst;
3032 	part = RTE_MIN(part, len);
3033 	do {
3034 		rte_memcpy(pdst, buf, part);
3035 		len -= part;
3036 		if (likely(!len)) {
3037 			pdst += part;
3038 			pdst = RTE_PTR_ALIGN(pdst, MLX5_WSEG_SIZE);
3039 			/* Note: no final wraparound check here. */
3040 			return (struct mlx5_wqe_dseg *)pdst;
3041 		}
3042 		pdst = (uint8_t *)txq->wqes;
3043 		buf += part;
3044 		part = len;
3045 	} while (true);
3046 }
3047 
3048 /**
3049  * Build the Ethernet Segment with optionally inlined data with
3050  * VLAN insertion and following Data Segments (if any) from
3051  * multi-segment packet. Used by ordinary send and TSO.
3052  *
3053  * @param txq
3054  *   Pointer to TX queue structure.
3055  * @param loc
3056  *   Pointer to burst routine local context.
3057  * @param wqe
3058  *   Pointer to WQE to fill with built Ethernet/Data Segments.
3059  * @param vlan
3060  *   Length of VLAN header to insert, 0 means no VLAN insertion.
3061  * @param inlen
3062  *   Data length to inline. For TSO this parameter specifies
3063  *   exact value, for ordinary send routine can be aligned by
3064  *   caller to provide better WQE space saving and data buffer
3065  *   start address alignment. This length includes VLAN header
3066  *   being inserted.
3067  * @param tso
3068  *   Zero means ordinary send, inlined data can be extended,
3069  *   otherwise this is TSO, inlined data length is fixed.
3070  * @param olx
3071  *   Configured Tx offloads mask. It is fully defined at
3072  *   compile time and may be used for optimization.
3073  *
3074  * @return
3075  *   Actual size of built WQE in segments.
3076  */
3077 static __rte_always_inline unsigned int
3078 mlx5_tx_mseg_build(struct mlx5_txq_data *restrict txq,
3079 		   struct mlx5_txq_local *restrict loc,
3080 		   struct mlx5_wqe *restrict wqe,
3081 		   unsigned int vlan,
3082 		   unsigned int inlen,
3083 		   unsigned int tso,
3084 		   unsigned int olx __rte_unused)
3085 {
3086 	struct mlx5_wqe_dseg *restrict dseg;
3087 	unsigned int ds;
3088 
3089 	MLX5_ASSERT((rte_pktmbuf_pkt_len(loc->mbuf) + vlan) >= inlen);
3090 	loc->mbuf_nseg = NB_SEGS(loc->mbuf);
3091 	loc->mbuf_off = 0;
3092 
3093 	dseg = mlx5_tx_eseg_mdat(txq, loc, wqe, vlan, inlen, tso, olx);
3094 	if (!loc->mbuf_nseg)
3095 		goto dseg_done;
3096 	/*
3097 	 * There are still some mbuf remaining, not inlined.
3098 	 * The first mbuf may be partially inlined and we
3099 	 * must process the possible non-zero data offset.
3100 	 */
3101 	if (loc->mbuf_off) {
3102 		unsigned int dlen;
3103 		uint8_t *dptr;
3104 
3105 		/*
3106 		 * Exhausted packets must be dropped before.
3107 		 * Non-zero offset means there are some data
3108 		 * remained in the packet.
3109 		 */
3110 		MLX5_ASSERT(loc->mbuf_off < rte_pktmbuf_data_len(loc->mbuf));
3111 		MLX5_ASSERT(rte_pktmbuf_data_len(loc->mbuf));
3112 		dptr = rte_pktmbuf_mtod_offset(loc->mbuf, uint8_t *,
3113 					       loc->mbuf_off);
3114 		dlen = rte_pktmbuf_data_len(loc->mbuf) - loc->mbuf_off;
3115 		/*
3116 		 * Build the pointer/minimal data Data Segment.
3117 		 * Do ring buffer wrapping check in advance.
3118 		 */
3119 		if ((uintptr_t)dseg >= (uintptr_t)txq->wqes_end)
3120 			dseg = (struct mlx5_wqe_dseg *)txq->wqes;
3121 		mlx5_tx_dseg_iptr(txq, loc, dseg, dptr, dlen, olx);
3122 		/* Store the mbuf to be freed on completion. */
3123 		MLX5_ASSERT(loc->elts_free);
3124 		txq->elts[txq->elts_head++ & txq->elts_m] = loc->mbuf;
3125 		--loc->elts_free;
3126 		++dseg;
3127 		if (--loc->mbuf_nseg == 0)
3128 			goto dseg_done;
3129 		loc->mbuf = loc->mbuf->next;
3130 		loc->mbuf_off = 0;
3131 	}
3132 	do {
3133 		if (unlikely(!rte_pktmbuf_data_len(loc->mbuf))) {
3134 			struct rte_mbuf *mbuf;
3135 
3136 			/* Zero length segment found, just skip. */
3137 			mbuf = loc->mbuf;
3138 			loc->mbuf = loc->mbuf->next;
3139 			rte_pktmbuf_free_seg(mbuf);
3140 			if (--loc->mbuf_nseg == 0)
3141 				break;
3142 		} else {
3143 			if ((uintptr_t)dseg >= (uintptr_t)txq->wqes_end)
3144 				dseg = (struct mlx5_wqe_dseg *)txq->wqes;
3145 			mlx5_tx_dseg_iptr
3146 				(txq, loc, dseg,
3147 				 rte_pktmbuf_mtod(loc->mbuf, uint8_t *),
3148 				 rte_pktmbuf_data_len(loc->mbuf), olx);
3149 			MLX5_ASSERT(loc->elts_free);
3150 			txq->elts[txq->elts_head++ & txq->elts_m] = loc->mbuf;
3151 			--loc->elts_free;
3152 			++dseg;
3153 			if (--loc->mbuf_nseg == 0)
3154 				break;
3155 			loc->mbuf = loc->mbuf->next;
3156 		}
3157 	} while (true);
3158 
3159 dseg_done:
3160 	/* Calculate actual segments used from the dseg pointer. */
3161 	if ((uintptr_t)wqe < (uintptr_t)dseg)
3162 		ds = ((uintptr_t)dseg - (uintptr_t)wqe) / MLX5_WSEG_SIZE;
3163 	else
3164 		ds = (((uintptr_t)dseg - (uintptr_t)wqe) +
3165 		      txq->wqe_s * MLX5_WQE_SIZE) / MLX5_WSEG_SIZE;
3166 	return ds;
3167 }
3168 
3169 /**
3170  * Tx one packet function for multi-segment TSO. Supports all
3171  * types of Tx offloads, uses MLX5_OPCODE_TSO to build WQEs,
3172  * sends one packet per WQE.
3173  *
3174  * This routine is responsible for storing processed mbuf
3175  * into elts ring buffer and update elts_head.
3176  *
3177  * @param txq
3178  *   Pointer to TX queue structure.
3179  * @param loc
3180  *   Pointer to burst routine local context.
3181  * @param olx
3182  *   Configured Tx offloads mask. It is fully defined at
3183  *   compile time and may be used for optimization.
3184  *
3185  * @return
3186  *   MLX5_TXCMP_CODE_EXIT - sending is done or impossible.
3187  *   MLX5_TXCMP_CODE_ERROR - some unrecoverable error occurred.
3188  * Local context variables partially updated.
3189  */
3190 static __rte_always_inline enum mlx5_txcmp_code
3191 mlx5_tx_packet_multi_tso(struct mlx5_txq_data *restrict txq,
3192 			struct mlx5_txq_local *restrict loc,
3193 			unsigned int olx)
3194 {
3195 	struct mlx5_wqe *restrict wqe;
3196 	unsigned int ds, dlen, inlen, ntcp, vlan = 0;
3197 
3198 	/*
3199 	 * Calculate data length to be inlined to estimate
3200 	 * the required space in WQE ring buffer.
3201 	 */
3202 	dlen = rte_pktmbuf_pkt_len(loc->mbuf);
3203 	if (MLX5_TXOFF_CONFIG(VLAN) && loc->mbuf->ol_flags & PKT_TX_VLAN_PKT)
3204 		vlan = sizeof(struct rte_vlan_hdr);
3205 	inlen = loc->mbuf->l2_len + vlan +
3206 		loc->mbuf->l3_len + loc->mbuf->l4_len;
3207 	if (unlikely((!inlen || !loc->mbuf->tso_segsz)))
3208 		return MLX5_TXCMP_CODE_ERROR;
3209 	if (loc->mbuf->ol_flags & PKT_TX_TUNNEL_MASK)
3210 		inlen += loc->mbuf->outer_l2_len + loc->mbuf->outer_l3_len;
3211 	/* Packet must contain all TSO headers. */
3212 	if (unlikely(inlen > MLX5_MAX_TSO_HEADER ||
3213 		     inlen <= MLX5_ESEG_MIN_INLINE_SIZE ||
3214 		     inlen > (dlen + vlan)))
3215 		return MLX5_TXCMP_CODE_ERROR;
3216 	MLX5_ASSERT(inlen >= txq->inlen_mode);
3217 	/*
3218 	 * Check whether there are enough free WQEBBs:
3219 	 * - Control Segment
3220 	 * - Ethernet Segment
3221 	 * - First Segment of inlined Ethernet data
3222 	 * - ... data continued ...
3223 	 * - Data Segments of pointer/min inline type
3224 	 */
3225 	ds = NB_SEGS(loc->mbuf) + 2 + (inlen -
3226 				       MLX5_ESEG_MIN_INLINE_SIZE +
3227 				       MLX5_WSEG_SIZE +
3228 				       MLX5_WSEG_SIZE - 1) / MLX5_WSEG_SIZE;
3229 	if (unlikely(loc->wqe_free < ((ds + 3) / 4)))
3230 		return MLX5_TXCMP_CODE_EXIT;
3231 	/* Check for maximal WQE size. */
3232 	if (unlikely((MLX5_WQE_SIZE_MAX / MLX5_WSEG_SIZE) < ((ds + 3) / 4)))
3233 		return MLX5_TXCMP_CODE_ERROR;
3234 #ifdef MLX5_PMD_SOFT_COUNTERS
3235 	/* Update sent data bytes/packets counters. */
3236 	ntcp = (dlen - (inlen - vlan) + loc->mbuf->tso_segsz - 1) /
3237 		loc->mbuf->tso_segsz;
3238 	/*
3239 	 * One will be added for mbuf itself
3240 	 * at the end of the mlx5_tx_burst from
3241 	 * loc->pkts_sent field.
3242 	 */
3243 	--ntcp;
3244 	txq->stats.opackets += ntcp;
3245 	txq->stats.obytes += dlen + vlan + ntcp * inlen;
3246 #endif
3247 	wqe = txq->wqes + (txq->wqe_ci & txq->wqe_m);
3248 	loc->wqe_last = wqe;
3249 	mlx5_tx_cseg_init(txq, loc, wqe, 0, MLX5_OPCODE_TSO, olx);
3250 	ds = mlx5_tx_mseg_build(txq, loc, wqe, vlan, inlen, 1, olx);
3251 	wqe->cseg.sq_ds = rte_cpu_to_be_32(txq->qp_num_8s | ds);
3252 	txq->wqe_ci += (ds + 3) / 4;
3253 	loc->wqe_free -= (ds + 3) / 4;
3254 	return MLX5_TXCMP_CODE_MULTI;
3255 }
3256 
3257 /**
3258  * Tx one packet function for multi-segment SEND. Supports all
3259  * types of Tx offloads, uses MLX5_OPCODE_SEND to build WQEs,
3260  * sends one packet per WQE, without any data inlining in
3261  * Ethernet Segment.
3262  *
3263  * This routine is responsible for storing processed mbuf
3264  * into elts ring buffer and update elts_head.
3265  *
3266  * @param txq
3267  *   Pointer to TX queue structure.
3268  * @param loc
3269  *   Pointer to burst routine local context.
3270  * @param olx
3271  *   Configured Tx offloads mask. It is fully defined at
3272  *   compile time and may be used for optimization.
3273  *
3274  * @return
3275  *   MLX5_TXCMP_CODE_EXIT - sending is done or impossible.
3276  *   MLX5_TXCMP_CODE_ERROR - some unrecoverable error occurred.
3277  * Local context variables partially updated.
3278  */
3279 static __rte_always_inline enum mlx5_txcmp_code
3280 mlx5_tx_packet_multi_send(struct mlx5_txq_data *restrict txq,
3281 			  struct mlx5_txq_local *restrict loc,
3282 			  unsigned int olx)
3283 {
3284 	struct mlx5_wqe_dseg *restrict dseg;
3285 	struct mlx5_wqe *restrict wqe;
3286 	unsigned int ds, nseg;
3287 
3288 	MLX5_ASSERT(NB_SEGS(loc->mbuf) > 1);
3289 	/*
3290 	 * No inline at all, it means the CPU cycles saving
3291 	 * is prioritized at configuration, we should not
3292 	 * copy any packet data to WQE.
3293 	 */
3294 	nseg = NB_SEGS(loc->mbuf);
3295 	ds = 2 + nseg;
3296 	if (unlikely(loc->wqe_free < ((ds + 3) / 4)))
3297 		return MLX5_TXCMP_CODE_EXIT;
3298 	/* Check for maximal WQE size. */
3299 	if (unlikely((MLX5_WQE_SIZE_MAX / MLX5_WSEG_SIZE) < ((ds + 3) / 4)))
3300 		return MLX5_TXCMP_CODE_ERROR;
3301 	/*
3302 	 * Some Tx offloads may cause an error if
3303 	 * packet is not long enough, check against
3304 	 * assumed minimal length.
3305 	 */
3306 	if (rte_pktmbuf_pkt_len(loc->mbuf) <= MLX5_ESEG_MIN_INLINE_SIZE)
3307 		return MLX5_TXCMP_CODE_ERROR;
3308 #ifdef MLX5_PMD_SOFT_COUNTERS
3309 	/* Update sent data bytes counter. */
3310 	txq->stats.obytes += rte_pktmbuf_pkt_len(loc->mbuf);
3311 	if (MLX5_TXOFF_CONFIG(VLAN) &&
3312 	    loc->mbuf->ol_flags & PKT_TX_VLAN_PKT)
3313 		txq->stats.obytes += sizeof(struct rte_vlan_hdr);
3314 #endif
3315 	/*
3316 	 * SEND WQE, one WQEBB:
3317 	 * - Control Segment, SEND opcode
3318 	 * - Ethernet Segment, optional VLAN, no inline
3319 	 * - Data Segments, pointer only type
3320 	 */
3321 	wqe = txq->wqes + (txq->wqe_ci & txq->wqe_m);
3322 	loc->wqe_last = wqe;
3323 	mlx5_tx_cseg_init(txq, loc, wqe, ds, MLX5_OPCODE_SEND, olx);
3324 	mlx5_tx_eseg_none(txq, loc, wqe, olx);
3325 	dseg = &wqe->dseg[0];
3326 	do {
3327 		if (unlikely(!rte_pktmbuf_data_len(loc->mbuf))) {
3328 			struct rte_mbuf *mbuf;
3329 
3330 			/*
3331 			 * Zero length segment found, have to
3332 			 * correct total size of WQE in segments.
3333 			 * It is supposed to be rare occasion, so
3334 			 * in normal case (no zero length segments)
3335 			 * we avoid extra writing to the Control
3336 			 * Segment.
3337 			 */
3338 			--ds;
3339 			wqe->cseg.sq_ds -= RTE_BE32(1);
3340 			mbuf = loc->mbuf;
3341 			loc->mbuf = mbuf->next;
3342 			rte_pktmbuf_free_seg(mbuf);
3343 			if (--nseg == 0)
3344 				break;
3345 		} else {
3346 			mlx5_tx_dseg_ptr
3347 				(txq, loc, dseg,
3348 				 rte_pktmbuf_mtod(loc->mbuf, uint8_t *),
3349 				 rte_pktmbuf_data_len(loc->mbuf), olx);
3350 			txq->elts[txq->elts_head++ & txq->elts_m] = loc->mbuf;
3351 			--loc->elts_free;
3352 			if (--nseg == 0)
3353 				break;
3354 			++dseg;
3355 			if ((uintptr_t)dseg >= (uintptr_t)txq->wqes_end)
3356 				dseg = (struct mlx5_wqe_dseg *)txq->wqes;
3357 			loc->mbuf = loc->mbuf->next;
3358 		}
3359 	} while (true);
3360 	txq->wqe_ci += (ds + 3) / 4;
3361 	loc->wqe_free -= (ds + 3) / 4;
3362 	return MLX5_TXCMP_CODE_MULTI;
3363 }
3364 
3365 /**
3366  * Tx one packet function for multi-segment SEND. Supports all
3367  * types of Tx offloads, uses MLX5_OPCODE_SEND to build WQEs,
3368  * sends one packet per WQE, with data inlining in
3369  * Ethernet Segment and minimal Data Segments.
3370  *
3371  * This routine is responsible for storing processed mbuf
3372  * into elts ring buffer and update elts_head.
3373  *
3374  * @param txq
3375  *   Pointer to TX queue structure.
3376  * @param loc
3377  *   Pointer to burst routine local context.
3378  * @param olx
3379  *   Configured Tx offloads mask. It is fully defined at
3380  *   compile time and may be used for optimization.
3381  *
3382  * @return
3383  *   MLX5_TXCMP_CODE_EXIT - sending is done or impossible.
3384  *   MLX5_TXCMP_CODE_ERROR - some unrecoverable error occurred.
3385  * Local context variables partially updated.
3386  */
3387 static __rte_always_inline enum mlx5_txcmp_code
3388 mlx5_tx_packet_multi_inline(struct mlx5_txq_data *restrict txq,
3389 			    struct mlx5_txq_local *restrict loc,
3390 			    unsigned int olx)
3391 {
3392 	struct mlx5_wqe *restrict wqe;
3393 	unsigned int ds, inlen, dlen, vlan = 0;
3394 
3395 	MLX5_ASSERT(MLX5_TXOFF_CONFIG(INLINE));
3396 	MLX5_ASSERT(NB_SEGS(loc->mbuf) > 1);
3397 	/*
3398 	 * First calculate data length to be inlined
3399 	 * to estimate the required space for WQE.
3400 	 */
3401 	dlen = rte_pktmbuf_pkt_len(loc->mbuf);
3402 	if (MLX5_TXOFF_CONFIG(VLAN) && loc->mbuf->ol_flags & PKT_TX_VLAN_PKT)
3403 		vlan = sizeof(struct rte_vlan_hdr);
3404 	inlen = dlen + vlan;
3405 	/* Check against minimal length. */
3406 	if (inlen <= MLX5_ESEG_MIN_INLINE_SIZE)
3407 		return MLX5_TXCMP_CODE_ERROR;
3408 	MLX5_ASSERT(txq->inlen_send >= MLX5_ESEG_MIN_INLINE_SIZE);
3409 	if (inlen > txq->inlen_send ||
3410 	    loc->mbuf->ol_flags & PKT_TX_DYNF_NOINLINE) {
3411 		struct rte_mbuf *mbuf;
3412 		unsigned int nxlen;
3413 		uintptr_t start;
3414 
3415 		/*
3416 		 * Packet length exceeds the allowed inline
3417 		 * data length, check whether the minimal
3418 		 * inlining is required.
3419 		 */
3420 		if (txq->inlen_mode) {
3421 			MLX5_ASSERT(txq->inlen_mode >=
3422 				    MLX5_ESEG_MIN_INLINE_SIZE);
3423 			MLX5_ASSERT(txq->inlen_mode <= txq->inlen_send);
3424 			inlen = txq->inlen_mode;
3425 		} else {
3426 			if (loc->mbuf->ol_flags & PKT_TX_DYNF_NOINLINE ||
3427 			    !vlan || txq->vlan_en) {
3428 				/*
3429 				 * VLAN insertion will be done inside by HW.
3430 				 * It is not utmost effective - VLAN flag is
3431 				 * checked twice, but we should proceed the
3432 				 * inlining length correctly and take into
3433 				 * account the VLAN header being inserted.
3434 				 */
3435 				return mlx5_tx_packet_multi_send
3436 							(txq, loc, olx);
3437 			}
3438 			inlen = MLX5_ESEG_MIN_INLINE_SIZE;
3439 		}
3440 		/*
3441 		 * Now we know the minimal amount of data is requested
3442 		 * to inline. Check whether we should inline the buffers
3443 		 * from the chain beginning to eliminate some mbufs.
3444 		 */
3445 		mbuf = loc->mbuf;
3446 		nxlen = rte_pktmbuf_data_len(mbuf);
3447 		if (unlikely(nxlen <= txq->inlen_send)) {
3448 			/* We can inline first mbuf at least. */
3449 			if (nxlen < inlen) {
3450 				unsigned int smlen;
3451 
3452 				/* Scan mbufs till inlen filled. */
3453 				do {
3454 					smlen = nxlen;
3455 					mbuf = NEXT(mbuf);
3456 					MLX5_ASSERT(mbuf);
3457 					nxlen = rte_pktmbuf_data_len(mbuf);
3458 					nxlen += smlen;
3459 				} while (unlikely(nxlen < inlen));
3460 				if (unlikely(nxlen > txq->inlen_send)) {
3461 					/* We cannot inline entire mbuf. */
3462 					smlen = inlen - smlen;
3463 					start = rte_pktmbuf_mtod_offset
3464 						    (mbuf, uintptr_t, smlen);
3465 					goto do_align;
3466 				}
3467 			}
3468 			do {
3469 				inlen = nxlen;
3470 				mbuf = NEXT(mbuf);
3471 				/* There should be not end of packet. */
3472 				MLX5_ASSERT(mbuf);
3473 				nxlen = inlen + rte_pktmbuf_data_len(mbuf);
3474 			} while (unlikely(nxlen < txq->inlen_send));
3475 		}
3476 		start = rte_pktmbuf_mtod(mbuf, uintptr_t);
3477 		/*
3478 		 * Check whether we can do inline to align start
3479 		 * address of data buffer to cacheline.
3480 		 */
3481 do_align:
3482 		start = (~start + 1) & (RTE_CACHE_LINE_SIZE - 1);
3483 		if (unlikely(start)) {
3484 			start += inlen;
3485 			if (start <= txq->inlen_send)
3486 				inlen = start;
3487 		}
3488 	}
3489 	/*
3490 	 * Check whether there are enough free WQEBBs:
3491 	 * - Control Segment
3492 	 * - Ethernet Segment
3493 	 * - First Segment of inlined Ethernet data
3494 	 * - ... data continued ...
3495 	 * - Data Segments of pointer/min inline type
3496 	 *
3497 	 * Estimate the number of Data Segments conservatively,
3498 	 * supposing no any mbufs is being freed during inlining.
3499 	 */
3500 	MLX5_ASSERT(inlen <= txq->inlen_send);
3501 	ds = NB_SEGS(loc->mbuf) + 2 + (inlen -
3502 				       MLX5_ESEG_MIN_INLINE_SIZE +
3503 				       MLX5_WSEG_SIZE +
3504 				       MLX5_WSEG_SIZE - 1) / MLX5_WSEG_SIZE;
3505 	if (unlikely(loc->wqe_free < ((ds + 3) / 4)))
3506 		return MLX5_TXCMP_CODE_EXIT;
3507 	/* Check for maximal WQE size. */
3508 	if (unlikely((MLX5_WQE_SIZE_MAX / MLX5_WSEG_SIZE) < ((ds + 3) / 4)))
3509 		return MLX5_TXCMP_CODE_ERROR;
3510 #ifdef MLX5_PMD_SOFT_COUNTERS
3511 	/* Update sent data bytes/packets counters. */
3512 	txq->stats.obytes += dlen + vlan;
3513 #endif
3514 	wqe = txq->wqes + (txq->wqe_ci & txq->wqe_m);
3515 	loc->wqe_last = wqe;
3516 	mlx5_tx_cseg_init(txq, loc, wqe, 0, MLX5_OPCODE_SEND, olx);
3517 	ds = mlx5_tx_mseg_build(txq, loc, wqe, vlan, inlen, 0, olx);
3518 	wqe->cseg.sq_ds = rte_cpu_to_be_32(txq->qp_num_8s | ds);
3519 	txq->wqe_ci += (ds + 3) / 4;
3520 	loc->wqe_free -= (ds + 3) / 4;
3521 	return MLX5_TXCMP_CODE_MULTI;
3522 }
3523 
3524 /**
3525  * Tx burst function for multi-segment packets. Supports all
3526  * types of Tx offloads, uses MLX5_OPCODE_SEND/TSO to build WQEs,
3527  * sends one packet per WQE. Function stops sending if it
3528  * encounters the single-segment packet.
3529  *
3530  * This routine is responsible for storing processed mbuf
3531  * into elts ring buffer and update elts_head.
3532  *
3533  * @param txq
3534  *   Pointer to TX queue structure.
3535  * @param[in] pkts
3536  *   Packets to transmit.
3537  * @param pkts_n
3538  *   Number of packets in array.
3539  * @param loc
3540  *   Pointer to burst routine local context.
3541  * @param olx
3542  *   Configured Tx offloads mask. It is fully defined at
3543  *   compile time and may be used for optimization.
3544  *
3545  * @return
3546  *   MLX5_TXCMP_CODE_EXIT - sending is done or impossible.
3547  *   MLX5_TXCMP_CODE_ERROR - some unrecoverable error occurred.
3548  *   MLX5_TXCMP_CODE_SINGLE - single-segment packet encountered.
3549  *   MLX5_TXCMP_CODE_TSO - TSO single-segment packet encountered.
3550  * Local context variables updated.
3551  */
3552 static __rte_always_inline enum mlx5_txcmp_code
3553 mlx5_tx_burst_mseg(struct mlx5_txq_data *restrict txq,
3554 		   struct rte_mbuf **restrict pkts,
3555 		   unsigned int pkts_n,
3556 		   struct mlx5_txq_local *restrict loc,
3557 		   unsigned int olx)
3558 {
3559 	MLX5_ASSERT(loc->elts_free && loc->wqe_free);
3560 	MLX5_ASSERT(pkts_n > loc->pkts_sent);
3561 	pkts += loc->pkts_sent + 1;
3562 	pkts_n -= loc->pkts_sent;
3563 	for (;;) {
3564 		enum mlx5_txcmp_code ret;
3565 
3566 		MLX5_ASSERT(NB_SEGS(loc->mbuf) > 1);
3567 		/*
3568 		 * Estimate the number of free elts quickly but
3569 		 * conservatively. Some segment may be fully inlined
3570 		 * and freed, ignore this here - precise estimation
3571 		 * is costly.
3572 		 */
3573 		if (loc->elts_free < NB_SEGS(loc->mbuf))
3574 			return MLX5_TXCMP_CODE_EXIT;
3575 		if (MLX5_TXOFF_CONFIG(TSO) &&
3576 		    unlikely(loc->mbuf->ol_flags & PKT_TX_TCP_SEG)) {
3577 			/* Proceed with multi-segment TSO. */
3578 			ret = mlx5_tx_packet_multi_tso(txq, loc, olx);
3579 		} else if (MLX5_TXOFF_CONFIG(INLINE)) {
3580 			/* Proceed with multi-segment SEND with inlining. */
3581 			ret = mlx5_tx_packet_multi_inline(txq, loc, olx);
3582 		} else {
3583 			/* Proceed with multi-segment SEND w/o inlining. */
3584 			ret = mlx5_tx_packet_multi_send(txq, loc, olx);
3585 		}
3586 		if (ret == MLX5_TXCMP_CODE_EXIT)
3587 			return MLX5_TXCMP_CODE_EXIT;
3588 		if (ret == MLX5_TXCMP_CODE_ERROR)
3589 			return MLX5_TXCMP_CODE_ERROR;
3590 		/* WQE is built, go to the next packet. */
3591 		++loc->pkts_sent;
3592 		--pkts_n;
3593 		if (unlikely(!pkts_n || !loc->elts_free || !loc->wqe_free))
3594 			return MLX5_TXCMP_CODE_EXIT;
3595 		loc->mbuf = *pkts++;
3596 		if (pkts_n > 1)
3597 			rte_prefetch0(*pkts);
3598 		if (likely(NB_SEGS(loc->mbuf) > 1))
3599 			continue;
3600 		/* Here ends the series of multi-segment packets. */
3601 		if (MLX5_TXOFF_CONFIG(TSO) &&
3602 		    unlikely(loc->mbuf->ol_flags & PKT_TX_TCP_SEG))
3603 			return MLX5_TXCMP_CODE_TSO;
3604 		return MLX5_TXCMP_CODE_SINGLE;
3605 	}
3606 	MLX5_ASSERT(false);
3607 }
3608 
3609 /**
3610  * Tx burst function for single-segment packets with TSO.
3611  * Supports all types of Tx offloads, except multi-packets.
3612  * Uses MLX5_OPCODE_TSO to build WQEs, sends one packet per WQE.
3613  * Function stops sending if it encounters the multi-segment
3614  * packet or packet without TSO requested.
3615  *
3616  * The routine is responsible for storing processed mbuf
3617  * into elts ring buffer and update elts_head if inline
3618  * offloads is requested due to possible early freeing
3619  * of the inlined mbufs (can not store pkts array in elts
3620  * as a batch).
3621  *
3622  * @param txq
3623  *   Pointer to TX queue structure.
3624  * @param[in] pkts
3625  *   Packets to transmit.
3626  * @param pkts_n
3627  *   Number of packets in array.
3628  * @param loc
3629  *   Pointer to burst routine local context.
3630  * @param olx
3631  *   Configured Tx offloads mask. It is fully defined at
3632  *   compile time and may be used for optimization.
3633  *
3634  * @return
3635  *   MLX5_TXCMP_CODE_EXIT - sending is done or impossible.
3636  *   MLX5_TXCMP_CODE_ERROR - some unrecoverable error occurred.
3637  *   MLX5_TXCMP_CODE_SINGLE - single-segment packet encountered.
3638  *   MLX5_TXCMP_CODE_MULTI - multi-segment packet encountered.
3639  * Local context variables updated.
3640  */
3641 static __rte_always_inline enum mlx5_txcmp_code
3642 mlx5_tx_burst_tso(struct mlx5_txq_data *restrict txq,
3643 		  struct rte_mbuf **restrict pkts,
3644 		  unsigned int pkts_n,
3645 		  struct mlx5_txq_local *restrict loc,
3646 		  unsigned int olx)
3647 {
3648 	MLX5_ASSERT(loc->elts_free && loc->wqe_free);
3649 	MLX5_ASSERT(pkts_n > loc->pkts_sent);
3650 	pkts += loc->pkts_sent + 1;
3651 	pkts_n -= loc->pkts_sent;
3652 	for (;;) {
3653 		struct mlx5_wqe_dseg *restrict dseg;
3654 		struct mlx5_wqe *restrict wqe;
3655 		unsigned int ds, dlen, hlen, ntcp, vlan = 0;
3656 		uint8_t *dptr;
3657 
3658 		MLX5_ASSERT(NB_SEGS(loc->mbuf) == 1);
3659 		dlen = rte_pktmbuf_data_len(loc->mbuf);
3660 		if (MLX5_TXOFF_CONFIG(VLAN) &&
3661 		    loc->mbuf->ol_flags & PKT_TX_VLAN_PKT) {
3662 			vlan = sizeof(struct rte_vlan_hdr);
3663 		}
3664 		/*
3665 		 * First calculate the WQE size to check
3666 		 * whether we have enough space in ring buffer.
3667 		 */
3668 		hlen = loc->mbuf->l2_len + vlan +
3669 		       loc->mbuf->l3_len + loc->mbuf->l4_len;
3670 		if (unlikely((!hlen || !loc->mbuf->tso_segsz)))
3671 			return MLX5_TXCMP_CODE_ERROR;
3672 		if (loc->mbuf->ol_flags & PKT_TX_TUNNEL_MASK)
3673 			hlen += loc->mbuf->outer_l2_len +
3674 				loc->mbuf->outer_l3_len;
3675 		/* Segment must contain all TSO headers. */
3676 		if (unlikely(hlen > MLX5_MAX_TSO_HEADER ||
3677 			     hlen <= MLX5_ESEG_MIN_INLINE_SIZE ||
3678 			     hlen > (dlen + vlan)))
3679 			return MLX5_TXCMP_CODE_ERROR;
3680 		/*
3681 		 * Check whether there are enough free WQEBBs:
3682 		 * - Control Segment
3683 		 * - Ethernet Segment
3684 		 * - First Segment of inlined Ethernet data
3685 		 * - ... data continued ...
3686 		 * - Finishing Data Segment of pointer type
3687 		 */
3688 		ds = 4 + (hlen - MLX5_ESEG_MIN_INLINE_SIZE +
3689 			  MLX5_WSEG_SIZE - 1) / MLX5_WSEG_SIZE;
3690 		if (loc->wqe_free < ((ds + 3) / 4))
3691 			return MLX5_TXCMP_CODE_EXIT;
3692 #ifdef MLX5_PMD_SOFT_COUNTERS
3693 		/* Update sent data bytes/packets counters. */
3694 		ntcp = (dlen + vlan - hlen +
3695 			loc->mbuf->tso_segsz - 1) /
3696 			loc->mbuf->tso_segsz;
3697 		/*
3698 		 * One will be added for mbuf itself at the end
3699 		 * of the mlx5_tx_burst from loc->pkts_sent field.
3700 		 */
3701 		--ntcp;
3702 		txq->stats.opackets += ntcp;
3703 		txq->stats.obytes += dlen + vlan + ntcp * hlen;
3704 #endif
3705 		/*
3706 		 * Build the TSO WQE:
3707 		 * - Control Segment
3708 		 * - Ethernet Segment with hlen bytes inlined
3709 		 * - Data Segment of pointer type
3710 		 */
3711 		wqe = txq->wqes + (txq->wqe_ci & txq->wqe_m);
3712 		loc->wqe_last = wqe;
3713 		mlx5_tx_cseg_init(txq, loc, wqe, ds,
3714 				  MLX5_OPCODE_TSO, olx);
3715 		dseg = mlx5_tx_eseg_data(txq, loc, wqe, vlan, hlen, 1, olx);
3716 		dptr = rte_pktmbuf_mtod(loc->mbuf, uint8_t *) + hlen - vlan;
3717 		dlen -= hlen - vlan;
3718 		mlx5_tx_dseg_ptr(txq, loc, dseg, dptr, dlen, olx);
3719 		/*
3720 		 * WQE is built, update the loop parameters
3721 		 * and go to the next packet.
3722 		 */
3723 		txq->wqe_ci += (ds + 3) / 4;
3724 		loc->wqe_free -= (ds + 3) / 4;
3725 		if (MLX5_TXOFF_CONFIG(INLINE))
3726 			txq->elts[txq->elts_head++ & txq->elts_m] = loc->mbuf;
3727 		--loc->elts_free;
3728 		++loc->pkts_sent;
3729 		--pkts_n;
3730 		if (unlikely(!pkts_n || !loc->elts_free || !loc->wqe_free))
3731 			return MLX5_TXCMP_CODE_EXIT;
3732 		loc->mbuf = *pkts++;
3733 		if (pkts_n > 1)
3734 			rte_prefetch0(*pkts);
3735 		if (MLX5_TXOFF_CONFIG(MULTI) &&
3736 		    unlikely(NB_SEGS(loc->mbuf) > 1))
3737 			return MLX5_TXCMP_CODE_MULTI;
3738 		if (likely(!(loc->mbuf->ol_flags & PKT_TX_TCP_SEG)))
3739 			return MLX5_TXCMP_CODE_SINGLE;
3740 		/* Continue with the next TSO packet. */
3741 	}
3742 	MLX5_ASSERT(false);
3743 }
3744 
3745 /**
3746  * Analyze the packet and select the best method to send.
3747  *
3748  * @param txq
3749  *   Pointer to TX queue structure.
3750  * @param loc
3751  *   Pointer to burst routine local context.
3752  * @param olx
3753  *   Configured Tx offloads mask. It is fully defined at
3754  *   compile time and may be used for optimization.
3755  * @param newp
3756  *   The predefined flag whether do complete check for
3757  *   multi-segment packets and TSO.
3758  *
3759  * @return
3760  *  MLX5_TXCMP_CODE_MULTI - multi-segment packet encountered.
3761  *  MLX5_TXCMP_CODE_TSO - TSO required, use TSO/LSO.
3762  *  MLX5_TXCMP_CODE_SINGLE - single-segment packet, use SEND.
3763  *  MLX5_TXCMP_CODE_EMPW - single-segment packet, use MPW.
3764  */
3765 static __rte_always_inline enum mlx5_txcmp_code
3766 mlx5_tx_able_to_empw(struct mlx5_txq_data *restrict txq,
3767 		     struct mlx5_txq_local *restrict loc,
3768 		     unsigned int olx,
3769 		     bool newp)
3770 {
3771 	/* Check for multi-segment packet. */
3772 	if (newp &&
3773 	    MLX5_TXOFF_CONFIG(MULTI) &&
3774 	    unlikely(NB_SEGS(loc->mbuf) > 1))
3775 		return MLX5_TXCMP_CODE_MULTI;
3776 	/* Check for TSO packet. */
3777 	if (newp &&
3778 	    MLX5_TXOFF_CONFIG(TSO) &&
3779 	    unlikely(loc->mbuf->ol_flags & PKT_TX_TCP_SEG))
3780 		return MLX5_TXCMP_CODE_TSO;
3781 	/* Check if eMPW is enabled at all. */
3782 	if (!MLX5_TXOFF_CONFIG(EMPW))
3783 		return MLX5_TXCMP_CODE_SINGLE;
3784 	/* Check if eMPW can be engaged. */
3785 	if (MLX5_TXOFF_CONFIG(VLAN) &&
3786 	    unlikely(loc->mbuf->ol_flags & PKT_TX_VLAN_PKT) &&
3787 		(!MLX5_TXOFF_CONFIG(INLINE) ||
3788 		 unlikely((rte_pktmbuf_data_len(loc->mbuf) +
3789 			   sizeof(struct rte_vlan_hdr)) > txq->inlen_empw))) {
3790 		/*
3791 		 * eMPW does not support VLAN insertion offload,
3792 		 * we have to inline the entire packet but
3793 		 * packet is too long for inlining.
3794 		 */
3795 		return MLX5_TXCMP_CODE_SINGLE;
3796 	}
3797 	return MLX5_TXCMP_CODE_EMPW;
3798 }
3799 
3800 /**
3801  * Check the next packet attributes to match with the eMPW batch ones.
3802  * In addition, for legacy MPW the packet length is checked either.
3803  *
3804  * @param txq
3805  *   Pointer to TX queue structure.
3806  * @param es
3807  *   Pointer to Ethernet Segment of eMPW batch.
3808  * @param loc
3809  *   Pointer to burst routine local context.
3810  * @param dlen
3811  *   Length of previous packet in MPW descriptor.
3812  * @param olx
3813  *   Configured Tx offloads mask. It is fully defined at
3814  *   compile time and may be used for optimization.
3815  *
3816  * @return
3817  *  true - packet match with eMPW batch attributes.
3818  *  false - no match, eMPW should be restarted.
3819  */
3820 static __rte_always_inline bool
3821 mlx5_tx_match_empw(struct mlx5_txq_data *restrict txq __rte_unused,
3822 		   struct mlx5_wqe_eseg *restrict es,
3823 		   struct mlx5_txq_local *restrict loc,
3824 		   uint32_t dlen,
3825 		   unsigned int olx)
3826 {
3827 	uint8_t swp_flags = 0;
3828 
3829 	/* Compare the checksum flags, if any. */
3830 	if (MLX5_TXOFF_CONFIG(CSUM) &&
3831 	    txq_ol_cksum_to_cs(loc->mbuf) != es->cs_flags)
3832 		return false;
3833 	/* Compare the Software Parser offsets and flags. */
3834 	if (MLX5_TXOFF_CONFIG(SWP) &&
3835 	    (es->swp_offs != txq_mbuf_to_swp(loc, &swp_flags, olx) ||
3836 	     es->swp_flags != swp_flags))
3837 		return false;
3838 	/* Fill metadata field if needed. */
3839 	if (MLX5_TXOFF_CONFIG(METADATA) &&
3840 		es->metadata != (loc->mbuf->ol_flags & PKT_TX_DYNF_METADATA ?
3841 				 *RTE_FLOW_DYNF_METADATA(loc->mbuf) : 0))
3842 		return false;
3843 	/* Legacy MPW can send packets with the same lengt only. */
3844 	if (MLX5_TXOFF_CONFIG(MPW) &&
3845 	    dlen != rte_pktmbuf_data_len(loc->mbuf))
3846 		return false;
3847 	/* There must be no VLAN packets in eMPW loop. */
3848 	if (MLX5_TXOFF_CONFIG(VLAN))
3849 		MLX5_ASSERT(!(loc->mbuf->ol_flags & PKT_TX_VLAN_PKT));
3850 	return true;
3851 }
3852 
3853 /*
3854  * Update send loop variables and WQE for eMPW loop
3855  * without data inlining. Number of Data Segments is
3856  * equal to the number of sent packets.
3857  *
3858  * @param txq
3859  *   Pointer to TX queue structure.
3860  * @param loc
3861  *   Pointer to burst routine local context.
3862  * @param ds
3863  *   Number of packets/Data Segments/Packets.
3864  * @param slen
3865  *   Accumulated statistics, bytes sent
3866  * @param olx
3867  *   Configured Tx offloads mask. It is fully defined at
3868  *   compile time and may be used for optimization.
3869  *
3870  * @return
3871  *  true - packet match with eMPW batch attributes.
3872  *  false - no match, eMPW should be restarted.
3873  */
3874 static __rte_always_inline void
3875 mlx5_tx_sdone_empw(struct mlx5_txq_data *restrict txq,
3876 		   struct mlx5_txq_local *restrict loc,
3877 		   unsigned int ds,
3878 		   unsigned int slen,
3879 		   unsigned int olx __rte_unused)
3880 {
3881 	MLX5_ASSERT(!MLX5_TXOFF_CONFIG(INLINE));
3882 #ifdef MLX5_PMD_SOFT_COUNTERS
3883 	/* Update sent data bytes counter. */
3884 	 txq->stats.obytes += slen;
3885 #else
3886 	(void)slen;
3887 #endif
3888 	loc->elts_free -= ds;
3889 	loc->pkts_sent += ds;
3890 	ds += 2;
3891 	loc->wqe_last->cseg.sq_ds = rte_cpu_to_be_32(txq->qp_num_8s | ds);
3892 	txq->wqe_ci += (ds + 3) / 4;
3893 	loc->wqe_free -= (ds + 3) / 4;
3894 }
3895 
3896 /*
3897  * Update send loop variables and WQE for eMPW loop
3898  * with data inlining. Gets the size of pushed descriptors
3899  * and data to the WQE.
3900  *
3901  * @param txq
3902  *   Pointer to TX queue structure.
3903  * @param loc
3904  *   Pointer to burst routine local context.
3905  * @param len
3906  *   Total size of descriptor/data in bytes.
3907  * @param slen
3908  *   Accumulated statistics, data bytes sent.
3909  * @param olx
3910  *   Configured Tx offloads mask. It is fully defined at
3911  *   compile time and may be used for optimization.
3912  *
3913  * @return
3914  *  true - packet match with eMPW batch attributes.
3915  *  false - no match, eMPW should be restarted.
3916  */
3917 static __rte_always_inline void
3918 mlx5_tx_idone_empw(struct mlx5_txq_data *restrict txq,
3919 		   struct mlx5_txq_local *restrict loc,
3920 		   unsigned int len,
3921 		   unsigned int slen,
3922 		   unsigned int olx __rte_unused)
3923 {
3924 	MLX5_ASSERT(MLX5_TXOFF_CONFIG(INLINE));
3925 	MLX5_ASSERT((len % MLX5_WSEG_SIZE) == 0);
3926 #ifdef MLX5_PMD_SOFT_COUNTERS
3927 	/* Update sent data bytes counter. */
3928 	 txq->stats.obytes += slen;
3929 #else
3930 	(void)slen;
3931 #endif
3932 	len = len / MLX5_WSEG_SIZE + 2;
3933 	loc->wqe_last->cseg.sq_ds = rte_cpu_to_be_32(txq->qp_num_8s | len);
3934 	txq->wqe_ci += (len + 3) / 4;
3935 	loc->wqe_free -= (len + 3) / 4;
3936 }
3937 
3938 /**
3939  * The set of Tx burst functions for single-segment packets
3940  * without TSO and with Multi-Packet Writing feature support.
3941  * Supports all types of Tx offloads, except multi-packets
3942  * and TSO.
3943  *
3944  * Uses MLX5_OPCODE_EMPW to build WQEs if possible and sends
3945  * as many packet per WQE as it can. If eMPW is not configured
3946  * or packet can not be sent with eMPW (VLAN insertion) the
3947  * ordinary SEND opcode is used and only one packet placed
3948  * in WQE.
3949  *
3950  * Functions stop sending if it encounters the multi-segment
3951  * packet or packet with TSO requested.
3952  *
3953  * The routines are responsible for storing processed mbuf
3954  * into elts ring buffer and update elts_head if inlining
3955  * offload is requested. Otherwise the copying mbufs to elts
3956  * can be postponed and completed at the end of burst routine.
3957  *
3958  * @param txq
3959  *   Pointer to TX queue structure.
3960  * @param[in] pkts
3961  *   Packets to transmit.
3962  * @param pkts_n
3963  *   Number of packets in array.
3964  * @param loc
3965  *   Pointer to burst routine local context.
3966  * @param olx
3967  *   Configured Tx offloads mask. It is fully defined at
3968  *   compile time and may be used for optimization.
3969  *
3970  * @return
3971  *   MLX5_TXCMP_CODE_EXIT - sending is done or impossible.
3972  *   MLX5_TXCMP_CODE_ERROR - some unrecoverable error occurred.
3973  *   MLX5_TXCMP_CODE_MULTI - multi-segment packet encountered.
3974  *   MLX5_TXCMP_CODE_TSO - TSO packet encountered.
3975  *   MLX5_TXCMP_CODE_SINGLE - used inside functions set.
3976  *   MLX5_TXCMP_CODE_EMPW - used inside functions set.
3977  *
3978  * Local context variables updated.
3979  *
3980  *
3981  * The routine sends packets with MLX5_OPCODE_EMPW
3982  * without inlining, this is dedicated optimized branch.
3983  * No VLAN insertion is supported.
3984  */
3985 static __rte_always_inline enum mlx5_txcmp_code
3986 mlx5_tx_burst_empw_simple(struct mlx5_txq_data *restrict txq,
3987 			  struct rte_mbuf **restrict pkts,
3988 			  unsigned int pkts_n,
3989 			  struct mlx5_txq_local *restrict loc,
3990 			  unsigned int olx)
3991 {
3992 	/*
3993 	 * Subroutine is the part of mlx5_tx_burst_single()
3994 	 * and sends single-segment packet with eMPW opcode
3995 	 * without data inlining.
3996 	 */
3997 	MLX5_ASSERT(!MLX5_TXOFF_CONFIG(INLINE));
3998 	MLX5_ASSERT(MLX5_TXOFF_CONFIG(EMPW));
3999 	MLX5_ASSERT(loc->elts_free && loc->wqe_free);
4000 	MLX5_ASSERT(pkts_n > loc->pkts_sent);
4001 	static_assert(MLX5_EMPW_MIN_PACKETS >= 2, "invalid min size");
4002 	pkts += loc->pkts_sent + 1;
4003 	pkts_n -= loc->pkts_sent;
4004 	for (;;) {
4005 		struct mlx5_wqe_dseg *restrict dseg;
4006 		struct mlx5_wqe_eseg *restrict eseg;
4007 		enum mlx5_txcmp_code ret;
4008 		unsigned int part, loop;
4009 		unsigned int slen = 0;
4010 
4011 next_empw:
4012 		MLX5_ASSERT(NB_SEGS(loc->mbuf) == 1);
4013 		part = RTE_MIN(pkts_n, MLX5_TXOFF_CONFIG(MPW) ?
4014 				       MLX5_MPW_MAX_PACKETS :
4015 				       MLX5_EMPW_MAX_PACKETS);
4016 		if (unlikely(loc->elts_free < part)) {
4017 			/* We have no enough elts to save all mbufs. */
4018 			if (unlikely(loc->elts_free < MLX5_EMPW_MIN_PACKETS))
4019 				return MLX5_TXCMP_CODE_EXIT;
4020 			/* But we still able to send at least minimal eMPW. */
4021 			part = loc->elts_free;
4022 		}
4023 		/* Check whether we have enough WQEs */
4024 		if (unlikely(loc->wqe_free < ((2 + part + 3) / 4))) {
4025 			if (unlikely(loc->wqe_free <
4026 				((2 + MLX5_EMPW_MIN_PACKETS + 3) / 4)))
4027 				return MLX5_TXCMP_CODE_EXIT;
4028 			part = (loc->wqe_free * 4) - 2;
4029 		}
4030 		if (likely(part > 1))
4031 			rte_prefetch0(*pkts);
4032 		loc->wqe_last = txq->wqes + (txq->wqe_ci & txq->wqe_m);
4033 		/*
4034 		 * Build eMPW title WQEBB:
4035 		 * - Control Segment, eMPW opcode
4036 		 * - Ethernet Segment, no inline
4037 		 */
4038 		mlx5_tx_cseg_init(txq, loc, loc->wqe_last, part + 2,
4039 				  MLX5_OPCODE_ENHANCED_MPSW, olx);
4040 		mlx5_tx_eseg_none(txq, loc, loc->wqe_last,
4041 				  olx & ~MLX5_TXOFF_CONFIG_VLAN);
4042 		eseg = &loc->wqe_last->eseg;
4043 		dseg = &loc->wqe_last->dseg[0];
4044 		loop = part;
4045 		/* Store the packet length for legacy MPW. */
4046 		if (MLX5_TXOFF_CONFIG(MPW))
4047 			eseg->mss = rte_cpu_to_be_16
4048 					(rte_pktmbuf_data_len(loc->mbuf));
4049 		for (;;) {
4050 			uint32_t dlen = rte_pktmbuf_data_len(loc->mbuf);
4051 #ifdef MLX5_PMD_SOFT_COUNTERS
4052 			/* Update sent data bytes counter. */
4053 			slen += dlen;
4054 #endif
4055 			mlx5_tx_dseg_ptr
4056 				(txq, loc, dseg,
4057 				 rte_pktmbuf_mtod(loc->mbuf, uint8_t *),
4058 				 dlen, olx);
4059 			if (unlikely(--loop == 0))
4060 				break;
4061 			loc->mbuf = *pkts++;
4062 			if (likely(loop > 1))
4063 				rte_prefetch0(*pkts);
4064 			ret = mlx5_tx_able_to_empw(txq, loc, olx, true);
4065 			/*
4066 			 * Unroll the completion code to avoid
4067 			 * returning variable value - it results in
4068 			 * unoptimized sequent checking in caller.
4069 			 */
4070 			if (ret == MLX5_TXCMP_CODE_MULTI) {
4071 				part -= loop;
4072 				mlx5_tx_sdone_empw(txq, loc, part, slen, olx);
4073 				if (unlikely(!loc->elts_free ||
4074 					     !loc->wqe_free))
4075 					return MLX5_TXCMP_CODE_EXIT;
4076 				return MLX5_TXCMP_CODE_MULTI;
4077 			}
4078 			MLX5_ASSERT(NB_SEGS(loc->mbuf) == 1);
4079 			if (ret == MLX5_TXCMP_CODE_TSO) {
4080 				part -= loop;
4081 				mlx5_tx_sdone_empw(txq, loc, part, slen, olx);
4082 				if (unlikely(!loc->elts_free ||
4083 					     !loc->wqe_free))
4084 					return MLX5_TXCMP_CODE_EXIT;
4085 				return MLX5_TXCMP_CODE_TSO;
4086 			}
4087 			if (ret == MLX5_TXCMP_CODE_SINGLE) {
4088 				part -= loop;
4089 				mlx5_tx_sdone_empw(txq, loc, part, slen, olx);
4090 				if (unlikely(!loc->elts_free ||
4091 					     !loc->wqe_free))
4092 					return MLX5_TXCMP_CODE_EXIT;
4093 				return MLX5_TXCMP_CODE_SINGLE;
4094 			}
4095 			if (ret != MLX5_TXCMP_CODE_EMPW) {
4096 				MLX5_ASSERT(false);
4097 				part -= loop;
4098 				mlx5_tx_sdone_empw(txq, loc, part, slen, olx);
4099 				return MLX5_TXCMP_CODE_ERROR;
4100 			}
4101 			/*
4102 			 * Check whether packet parameters coincide
4103 			 * within assumed eMPW batch:
4104 			 * - check sum settings
4105 			 * - metadata value
4106 			 * - software parser settings
4107 			 * - packets length (legacy MPW only)
4108 			 */
4109 			if (!mlx5_tx_match_empw(txq, eseg, loc, dlen, olx)) {
4110 				MLX5_ASSERT(loop);
4111 				part -= loop;
4112 				mlx5_tx_sdone_empw(txq, loc, part, slen, olx);
4113 				if (unlikely(!loc->elts_free ||
4114 					     !loc->wqe_free))
4115 					return MLX5_TXCMP_CODE_EXIT;
4116 				pkts_n -= part;
4117 				goto next_empw;
4118 			}
4119 			/* Packet attributes match, continue the same eMPW. */
4120 			++dseg;
4121 			if ((uintptr_t)dseg >= (uintptr_t)txq->wqes_end)
4122 				dseg = (struct mlx5_wqe_dseg *)txq->wqes;
4123 		}
4124 		/* eMPW is built successfully, update loop parameters. */
4125 		MLX5_ASSERT(!loop);
4126 		MLX5_ASSERT(pkts_n >= part);
4127 #ifdef MLX5_PMD_SOFT_COUNTERS
4128 		/* Update sent data bytes counter. */
4129 		txq->stats.obytes += slen;
4130 #endif
4131 		loc->elts_free -= part;
4132 		loc->pkts_sent += part;
4133 		txq->wqe_ci += (2 + part + 3) / 4;
4134 		loc->wqe_free -= (2 + part + 3) / 4;
4135 		pkts_n -= part;
4136 		if (unlikely(!pkts_n || !loc->elts_free || !loc->wqe_free))
4137 			return MLX5_TXCMP_CODE_EXIT;
4138 		loc->mbuf = *pkts++;
4139 		ret = mlx5_tx_able_to_empw(txq, loc, olx, true);
4140 		if (unlikely(ret != MLX5_TXCMP_CODE_EMPW))
4141 			return ret;
4142 		/* Continue sending eMPW batches. */
4143 	}
4144 	MLX5_ASSERT(false);
4145 }
4146 
4147 /**
4148  * The routine sends packets with MLX5_OPCODE_EMPW
4149  * with inlining, optionally supports VLAN insertion.
4150  */
4151 static __rte_always_inline enum mlx5_txcmp_code
4152 mlx5_tx_burst_empw_inline(struct mlx5_txq_data *restrict txq,
4153 			  struct rte_mbuf **restrict pkts,
4154 			  unsigned int pkts_n,
4155 			  struct mlx5_txq_local *restrict loc,
4156 			  unsigned int olx)
4157 {
4158 	/*
4159 	 * Subroutine is the part of mlx5_tx_burst_single()
4160 	 * and sends single-segment packet with eMPW opcode
4161 	 * with data inlining.
4162 	 */
4163 	MLX5_ASSERT(MLX5_TXOFF_CONFIG(INLINE));
4164 	MLX5_ASSERT(MLX5_TXOFF_CONFIG(EMPW));
4165 	MLX5_ASSERT(loc->elts_free && loc->wqe_free);
4166 	MLX5_ASSERT(pkts_n > loc->pkts_sent);
4167 	static_assert(MLX5_EMPW_MIN_PACKETS >= 2, "invalid min size");
4168 	pkts += loc->pkts_sent + 1;
4169 	pkts_n -= loc->pkts_sent;
4170 	for (;;) {
4171 		struct mlx5_wqe_dseg *restrict dseg;
4172 		struct mlx5_wqe_eseg *restrict eseg;
4173 		enum mlx5_txcmp_code ret;
4174 		unsigned int room, part, nlim;
4175 		unsigned int slen = 0;
4176 
4177 		MLX5_ASSERT(NB_SEGS(loc->mbuf) == 1);
4178 		/*
4179 		 * Limits the amount of packets in one WQE
4180 		 * to improve CQE latency generation.
4181 		 */
4182 		nlim = RTE_MIN(pkts_n, MLX5_TXOFF_CONFIG(MPW) ?
4183 				       MLX5_MPW_INLINE_MAX_PACKETS :
4184 				       MLX5_EMPW_MAX_PACKETS);
4185 		/* Check whether we have minimal amount WQEs */
4186 		if (unlikely(loc->wqe_free <
4187 			    ((2 + MLX5_EMPW_MIN_PACKETS + 3) / 4)))
4188 			return MLX5_TXCMP_CODE_EXIT;
4189 		if (likely(pkts_n > 1))
4190 			rte_prefetch0(*pkts);
4191 		loc->wqe_last = txq->wqes + (txq->wqe_ci & txq->wqe_m);
4192 		/*
4193 		 * Build eMPW title WQEBB:
4194 		 * - Control Segment, eMPW opcode, zero DS
4195 		 * - Ethernet Segment, no inline
4196 		 */
4197 		mlx5_tx_cseg_init(txq, loc, loc->wqe_last, 0,
4198 				  MLX5_OPCODE_ENHANCED_MPSW, olx);
4199 		mlx5_tx_eseg_none(txq, loc, loc->wqe_last,
4200 				  olx & ~MLX5_TXOFF_CONFIG_VLAN);
4201 		eseg = &loc->wqe_last->eseg;
4202 		dseg = &loc->wqe_last->dseg[0];
4203 		/* Store the packet length for legacy MPW. */
4204 		if (MLX5_TXOFF_CONFIG(MPW))
4205 			eseg->mss = rte_cpu_to_be_16
4206 					(rte_pktmbuf_data_len(loc->mbuf));
4207 		room = RTE_MIN(MLX5_WQE_SIZE_MAX / MLX5_WQE_SIZE,
4208 			       loc->wqe_free) * MLX5_WQE_SIZE -
4209 					MLX5_WQE_CSEG_SIZE -
4210 					MLX5_WQE_ESEG_SIZE;
4211 		/* Build WQE till we have space, packets and resources. */
4212 		part = room;
4213 		for (;;) {
4214 			uint32_t dlen = rte_pktmbuf_data_len(loc->mbuf);
4215 			uint8_t *dptr = rte_pktmbuf_mtod(loc->mbuf, uint8_t *);
4216 			unsigned int tlen;
4217 
4218 			MLX5_ASSERT(room >= MLX5_WQE_DSEG_SIZE);
4219 			MLX5_ASSERT((room % MLX5_WQE_DSEG_SIZE) == 0);
4220 			MLX5_ASSERT((uintptr_t)dseg < (uintptr_t)txq->wqes_end);
4221 			/*
4222 			 * Some Tx offloads may cause an error if
4223 			 * packet is not long enough, check against
4224 			 * assumed minimal length.
4225 			 */
4226 			if (unlikely(dlen <= MLX5_ESEG_MIN_INLINE_SIZE)) {
4227 				part -= room;
4228 				if (unlikely(!part))
4229 					return MLX5_TXCMP_CODE_ERROR;
4230 				/*
4231 				 * We have some successfully built
4232 				 * packet Data Segments to send.
4233 				 */
4234 				mlx5_tx_idone_empw(txq, loc, part, slen, olx);
4235 				return MLX5_TXCMP_CODE_ERROR;
4236 			}
4237 			/* Inline or not inline - that's the Question. */
4238 			if (dlen > txq->inlen_empw ||
4239 			    loc->mbuf->ol_flags & PKT_TX_DYNF_NOINLINE)
4240 				goto pointer_empw;
4241 			/* Inline entire packet, optional VLAN insertion. */
4242 			tlen = sizeof(dseg->bcount) + dlen;
4243 			if (MLX5_TXOFF_CONFIG(VLAN) &&
4244 			    loc->mbuf->ol_flags & PKT_TX_VLAN_PKT) {
4245 				/*
4246 				 * The packet length must be checked in
4247 				 * mlx5_tx_able_to_empw() and packet
4248 				 * fits into inline length guaranteed.
4249 				 */
4250 				MLX5_ASSERT((dlen +
4251 					     sizeof(struct rte_vlan_hdr)) <=
4252 					    txq->inlen_empw);
4253 				tlen += sizeof(struct rte_vlan_hdr);
4254 				if (room < tlen)
4255 					break;
4256 				dseg = mlx5_tx_dseg_vlan(txq, loc, dseg,
4257 							 dptr, dlen, olx);
4258 #ifdef MLX5_PMD_SOFT_COUNTERS
4259 				/* Update sent data bytes counter. */
4260 				slen +=	sizeof(struct rte_vlan_hdr);
4261 #endif
4262 			} else {
4263 				if (room < tlen)
4264 					break;
4265 				dseg = mlx5_tx_dseg_empw(txq, loc, dseg,
4266 							 dptr, dlen, olx);
4267 			}
4268 			tlen = RTE_ALIGN(tlen, MLX5_WSEG_SIZE);
4269 			MLX5_ASSERT(room >= tlen);
4270 			room -= tlen;
4271 			/*
4272 			 * Packet data are completely inlined,
4273 			 * free the packet immediately.
4274 			 */
4275 			rte_pktmbuf_free_seg(loc->mbuf);
4276 			goto next_mbuf;
4277 pointer_empw:
4278 			/*
4279 			 * Not inlinable VLAN packets are
4280 			 * proceeded outside of this routine.
4281 			 */
4282 			MLX5_ASSERT(room >= MLX5_WQE_DSEG_SIZE);
4283 			if (MLX5_TXOFF_CONFIG(VLAN))
4284 				MLX5_ASSERT(!(loc->mbuf->ol_flags &
4285 					    PKT_TX_VLAN_PKT));
4286 			mlx5_tx_dseg_ptr(txq, loc, dseg, dptr, dlen, olx);
4287 			/* We have to store mbuf in elts.*/
4288 			txq->elts[txq->elts_head++ & txq->elts_m] = loc->mbuf;
4289 			room -= MLX5_WQE_DSEG_SIZE;
4290 			/* Ring buffer wraparound is checked at the loop end.*/
4291 			++dseg;
4292 next_mbuf:
4293 #ifdef MLX5_PMD_SOFT_COUNTERS
4294 			/* Update sent data bytes counter. */
4295 			slen += dlen;
4296 #endif
4297 			loc->pkts_sent++;
4298 			loc->elts_free--;
4299 			pkts_n--;
4300 			if (unlikely(!pkts_n || !loc->elts_free)) {
4301 				/*
4302 				 * We have no resources/packets to
4303 				 * continue build descriptors.
4304 				 */
4305 				part -= room;
4306 				mlx5_tx_idone_empw(txq, loc, part, slen, olx);
4307 				return MLX5_TXCMP_CODE_EXIT;
4308 			}
4309 			loc->mbuf = *pkts++;
4310 			if (likely(pkts_n > 1))
4311 				rte_prefetch0(*pkts);
4312 			ret = mlx5_tx_able_to_empw(txq, loc, olx, true);
4313 			/*
4314 			 * Unroll the completion code to avoid
4315 			 * returning variable value - it results in
4316 			 * unoptimized sequent checking in caller.
4317 			 */
4318 			if (ret == MLX5_TXCMP_CODE_MULTI) {
4319 				part -= room;
4320 				mlx5_tx_idone_empw(txq, loc, part, slen, olx);
4321 				if (unlikely(!loc->elts_free ||
4322 					     !loc->wqe_free))
4323 					return MLX5_TXCMP_CODE_EXIT;
4324 				return MLX5_TXCMP_CODE_MULTI;
4325 			}
4326 			MLX5_ASSERT(NB_SEGS(loc->mbuf) == 1);
4327 			if (ret == MLX5_TXCMP_CODE_TSO) {
4328 				part -= room;
4329 				mlx5_tx_idone_empw(txq, loc, part, slen, olx);
4330 				if (unlikely(!loc->elts_free ||
4331 					     !loc->wqe_free))
4332 					return MLX5_TXCMP_CODE_EXIT;
4333 				return MLX5_TXCMP_CODE_TSO;
4334 			}
4335 			if (ret == MLX5_TXCMP_CODE_SINGLE) {
4336 				part -= room;
4337 				mlx5_tx_idone_empw(txq, loc, part, slen, olx);
4338 				if (unlikely(!loc->elts_free ||
4339 					     !loc->wqe_free))
4340 					return MLX5_TXCMP_CODE_EXIT;
4341 				return MLX5_TXCMP_CODE_SINGLE;
4342 			}
4343 			if (ret != MLX5_TXCMP_CODE_EMPW) {
4344 				MLX5_ASSERT(false);
4345 				part -= room;
4346 				mlx5_tx_idone_empw(txq, loc, part, slen, olx);
4347 				return MLX5_TXCMP_CODE_ERROR;
4348 			}
4349 			/* Check if we have minimal room left. */
4350 			nlim--;
4351 			if (unlikely(!nlim || room < MLX5_WQE_DSEG_SIZE))
4352 				break;
4353 			/*
4354 			 * Check whether packet parameters coincide
4355 			 * within assumed eMPW batch:
4356 			 * - check sum settings
4357 			 * - metadata value
4358 			 * - software parser settings
4359 			 * - packets length (legacy MPW only)
4360 			 */
4361 			if (!mlx5_tx_match_empw(txq, eseg, loc, dlen, olx))
4362 				break;
4363 			/* Packet attributes match, continue the same eMPW. */
4364 			if ((uintptr_t)dseg >= (uintptr_t)txq->wqes_end)
4365 				dseg = (struct mlx5_wqe_dseg *)txq->wqes;
4366 		}
4367 		/*
4368 		 * We get here to close an existing eMPW
4369 		 * session and start the new one.
4370 		 */
4371 		MLX5_ASSERT(pkts_n);
4372 		part -= room;
4373 		if (unlikely(!part))
4374 			return MLX5_TXCMP_CODE_EXIT;
4375 		mlx5_tx_idone_empw(txq, loc, part, slen, olx);
4376 		if (unlikely(!loc->elts_free ||
4377 			     !loc->wqe_free))
4378 			return MLX5_TXCMP_CODE_EXIT;
4379 		/* Continue the loop with new eMPW session. */
4380 	}
4381 	MLX5_ASSERT(false);
4382 }
4383 
4384 /**
4385  * The routine sends packets with ordinary MLX5_OPCODE_SEND.
4386  * Data inlining and VLAN insertion are supported.
4387  */
4388 static __rte_always_inline enum mlx5_txcmp_code
4389 mlx5_tx_burst_single_send(struct mlx5_txq_data *restrict txq,
4390 			  struct rte_mbuf **restrict pkts,
4391 			  unsigned int pkts_n,
4392 			  struct mlx5_txq_local *restrict loc,
4393 			  unsigned int olx)
4394 {
4395 	/*
4396 	 * Subroutine is the part of mlx5_tx_burst_single()
4397 	 * and sends single-segment packet with SEND opcode.
4398 	 */
4399 	MLX5_ASSERT(loc->elts_free && loc->wqe_free);
4400 	MLX5_ASSERT(pkts_n > loc->pkts_sent);
4401 	pkts += loc->pkts_sent + 1;
4402 	pkts_n -= loc->pkts_sent;
4403 	for (;;) {
4404 		struct mlx5_wqe *restrict wqe;
4405 		enum mlx5_txcmp_code ret;
4406 
4407 		MLX5_ASSERT(NB_SEGS(loc->mbuf) == 1);
4408 		if (MLX5_TXOFF_CONFIG(INLINE)) {
4409 			unsigned int inlen, vlan = 0;
4410 
4411 			inlen = rte_pktmbuf_data_len(loc->mbuf);
4412 			if (MLX5_TXOFF_CONFIG(VLAN) &&
4413 			    loc->mbuf->ol_flags & PKT_TX_VLAN_PKT) {
4414 				vlan = sizeof(struct rte_vlan_hdr);
4415 				inlen += vlan;
4416 				static_assert((sizeof(struct rte_vlan_hdr) +
4417 					       sizeof(struct rte_ether_hdr)) ==
4418 					       MLX5_ESEG_MIN_INLINE_SIZE,
4419 					       "invalid min inline data size");
4420 			}
4421 			/*
4422 			 * If inlining is enabled at configuration time
4423 			 * the limit must be not less than minimal size.
4424 			 * Otherwise we would do extra check for data
4425 			 * size to avoid crashes due to length overflow.
4426 			 */
4427 			MLX5_ASSERT(txq->inlen_send >=
4428 				    MLX5_ESEG_MIN_INLINE_SIZE);
4429 			if (inlen <= txq->inlen_send) {
4430 				unsigned int seg_n, wqe_n;
4431 
4432 				rte_prefetch0(rte_pktmbuf_mtod
4433 						(loc->mbuf, uint8_t *));
4434 				/* Check against minimal length. */
4435 				if (inlen <= MLX5_ESEG_MIN_INLINE_SIZE)
4436 					return MLX5_TXCMP_CODE_ERROR;
4437 				if (loc->mbuf->ol_flags &
4438 				    PKT_TX_DYNF_NOINLINE) {
4439 					/*
4440 					 * The hint flag not to inline packet
4441 					 * data is set. Check whether we can
4442 					 * follow the hint.
4443 					 */
4444 					if ((!MLX5_TXOFF_CONFIG(EMPW) &&
4445 					      txq->inlen_mode) ||
4446 					    (MLX5_TXOFF_CONFIG(MPW) &&
4447 					     txq->inlen_mode)) {
4448 						/*
4449 						 * The hardware requires the
4450 						 * minimal inline data header.
4451 						 */
4452 						goto single_min_inline;
4453 					}
4454 					if (MLX5_TXOFF_CONFIG(VLAN) &&
4455 					    vlan && !txq->vlan_en) {
4456 						/*
4457 						 * We must insert VLAN tag
4458 						 * by software means.
4459 						 */
4460 						goto single_part_inline;
4461 					}
4462 					goto single_no_inline;
4463 				}
4464 				/*
4465 				 * Completely inlined packet data WQE:
4466 				 * - Control Segment, SEND opcode
4467 				 * - Ethernet Segment, no VLAN insertion
4468 				 * - Data inlined, VLAN optionally inserted
4469 				 * - Alignment to MLX5_WSEG_SIZE
4470 				 * Have to estimate amount of WQEBBs
4471 				 */
4472 				seg_n = (inlen + 3 * MLX5_WSEG_SIZE -
4473 					 MLX5_ESEG_MIN_INLINE_SIZE +
4474 					 MLX5_WSEG_SIZE - 1) / MLX5_WSEG_SIZE;
4475 				/* Check if there are enough WQEBBs. */
4476 				wqe_n = (seg_n + 3) / 4;
4477 				if (wqe_n > loc->wqe_free)
4478 					return MLX5_TXCMP_CODE_EXIT;
4479 				wqe = txq->wqes + (txq->wqe_ci & txq->wqe_m);
4480 				loc->wqe_last = wqe;
4481 				mlx5_tx_cseg_init(txq, loc, wqe, seg_n,
4482 						  MLX5_OPCODE_SEND, olx);
4483 				mlx5_tx_eseg_data(txq, loc, wqe,
4484 						  vlan, inlen, 0, olx);
4485 				txq->wqe_ci += wqe_n;
4486 				loc->wqe_free -= wqe_n;
4487 				/*
4488 				 * Packet data are completely inlined,
4489 				 * free the packet immediately.
4490 				 */
4491 				rte_pktmbuf_free_seg(loc->mbuf);
4492 			} else if ((!MLX5_TXOFF_CONFIG(EMPW) ||
4493 				     MLX5_TXOFF_CONFIG(MPW)) &&
4494 					txq->inlen_mode) {
4495 				/*
4496 				 * If minimal inlining is requested the eMPW
4497 				 * feature should be disabled due to data is
4498 				 * inlined into Ethernet Segment, which can
4499 				 * not contain inlined data for eMPW due to
4500 				 * segment shared for all packets.
4501 				 */
4502 				struct mlx5_wqe_dseg *restrict dseg;
4503 				unsigned int ds;
4504 				uint8_t *dptr;
4505 
4506 				/*
4507 				 * The inline-mode settings require
4508 				 * to inline the specified amount of
4509 				 * data bytes to the Ethernet Segment.
4510 				 * We should check the free space in
4511 				 * WQE ring buffer to inline partially.
4512 				 */
4513 single_min_inline:
4514 				MLX5_ASSERT(txq->inlen_send >= txq->inlen_mode);
4515 				MLX5_ASSERT(inlen > txq->inlen_mode);
4516 				MLX5_ASSERT(txq->inlen_mode >=
4517 					    MLX5_ESEG_MIN_INLINE_SIZE);
4518 				/*
4519 				 * Check whether there are enough free WQEBBs:
4520 				 * - Control Segment
4521 				 * - Ethernet Segment
4522 				 * - First Segment of inlined Ethernet data
4523 				 * - ... data continued ...
4524 				 * - Finishing Data Segment of pointer type
4525 				 */
4526 				ds = (MLX5_WQE_CSEG_SIZE +
4527 				      MLX5_WQE_ESEG_SIZE +
4528 				      MLX5_WQE_DSEG_SIZE +
4529 				      txq->inlen_mode -
4530 				      MLX5_ESEG_MIN_INLINE_SIZE +
4531 				      MLX5_WQE_DSEG_SIZE +
4532 				      MLX5_WSEG_SIZE - 1) / MLX5_WSEG_SIZE;
4533 				if (loc->wqe_free < ((ds + 3) / 4))
4534 					return MLX5_TXCMP_CODE_EXIT;
4535 				/*
4536 				 * Build the ordinary SEND WQE:
4537 				 * - Control Segment
4538 				 * - Ethernet Segment, inline inlen_mode bytes
4539 				 * - Data Segment of pointer type
4540 				 */
4541 				wqe = txq->wqes + (txq->wqe_ci & txq->wqe_m);
4542 				loc->wqe_last = wqe;
4543 				mlx5_tx_cseg_init(txq, loc, wqe, ds,
4544 						  MLX5_OPCODE_SEND, olx);
4545 				dseg = mlx5_tx_eseg_data(txq, loc, wqe, vlan,
4546 							 txq->inlen_mode,
4547 							 0, olx);
4548 				dptr = rte_pktmbuf_mtod(loc->mbuf, uint8_t *) +
4549 				       txq->inlen_mode - vlan;
4550 				inlen -= txq->inlen_mode;
4551 				mlx5_tx_dseg_ptr(txq, loc, dseg,
4552 						 dptr, inlen, olx);
4553 				/*
4554 				 * WQE is built, update the loop parameters
4555 				 * and got to the next packet.
4556 				 */
4557 				txq->wqe_ci += (ds + 3) / 4;
4558 				loc->wqe_free -= (ds + 3) / 4;
4559 				/* We have to store mbuf in elts.*/
4560 				MLX5_ASSERT(MLX5_TXOFF_CONFIG(INLINE));
4561 				txq->elts[txq->elts_head++ & txq->elts_m] =
4562 						loc->mbuf;
4563 				--loc->elts_free;
4564 			} else {
4565 				uint8_t *dptr;
4566 				unsigned int dlen;
4567 
4568 				/*
4569 				 * Partially inlined packet data WQE, we have
4570 				 * some space in title WQEBB, we can fill it
4571 				 * with some packet data. It takes one WQEBB,
4572 				 * it is available, no extra space check:
4573 				 * - Control Segment, SEND opcode
4574 				 * - Ethernet Segment, no VLAN insertion
4575 				 * - MLX5_ESEG_MIN_INLINE_SIZE bytes of Data
4576 				 * - Data Segment, pointer type
4577 				 *
4578 				 * We also get here if VLAN insertion is not
4579 				 * supported by HW, the inline is enabled.
4580 				 */
4581 single_part_inline:
4582 				wqe = txq->wqes + (txq->wqe_ci & txq->wqe_m);
4583 				loc->wqe_last = wqe;
4584 				mlx5_tx_cseg_init(txq, loc, wqe, 4,
4585 						  MLX5_OPCODE_SEND, olx);
4586 				mlx5_tx_eseg_dmin(txq, loc, wqe, vlan, olx);
4587 				dptr = rte_pktmbuf_mtod(loc->mbuf, uint8_t *) +
4588 				       MLX5_ESEG_MIN_INLINE_SIZE - vlan;
4589 				/*
4590 				 * The length check is performed above, by
4591 				 * comparing with txq->inlen_send. We should
4592 				 * not get overflow here.
4593 				 */
4594 				MLX5_ASSERT(inlen > MLX5_ESEG_MIN_INLINE_SIZE);
4595 				dlen = inlen - MLX5_ESEG_MIN_INLINE_SIZE;
4596 				mlx5_tx_dseg_ptr(txq, loc, &wqe->dseg[1],
4597 						 dptr, dlen, olx);
4598 				++txq->wqe_ci;
4599 				--loc->wqe_free;
4600 				/* We have to store mbuf in elts.*/
4601 				MLX5_ASSERT(MLX5_TXOFF_CONFIG(INLINE));
4602 				txq->elts[txq->elts_head++ & txq->elts_m] =
4603 						loc->mbuf;
4604 				--loc->elts_free;
4605 			}
4606 #ifdef MLX5_PMD_SOFT_COUNTERS
4607 			/* Update sent data bytes counter. */
4608 			txq->stats.obytes += vlan +
4609 					rte_pktmbuf_data_len(loc->mbuf);
4610 #endif
4611 		} else {
4612 			/*
4613 			 * No inline at all, it means the CPU cycles saving
4614 			 * is prioritized at configuration, we should not
4615 			 * copy any packet data to WQE.
4616 			 *
4617 			 * SEND WQE, one WQEBB:
4618 			 * - Control Segment, SEND opcode
4619 			 * - Ethernet Segment, optional VLAN, no inline
4620 			 * - Data Segment, pointer type
4621 			 */
4622 single_no_inline:
4623 			wqe = txq->wqes + (txq->wqe_ci & txq->wqe_m);
4624 			loc->wqe_last = wqe;
4625 			mlx5_tx_cseg_init(txq, loc, wqe, 3,
4626 					  MLX5_OPCODE_SEND, olx);
4627 			mlx5_tx_eseg_none(txq, loc, wqe, olx);
4628 			mlx5_tx_dseg_ptr
4629 				(txq, loc, &wqe->dseg[0],
4630 				 rte_pktmbuf_mtod(loc->mbuf, uint8_t *),
4631 				 rte_pktmbuf_data_len(loc->mbuf), olx);
4632 			++txq->wqe_ci;
4633 			--loc->wqe_free;
4634 			/*
4635 			 * We should not store mbuf pointer in elts
4636 			 * if no inlining is configured, this is done
4637 			 * by calling routine in a batch copy.
4638 			 */
4639 			MLX5_ASSERT(!MLX5_TXOFF_CONFIG(INLINE));
4640 			--loc->elts_free;
4641 #ifdef MLX5_PMD_SOFT_COUNTERS
4642 			/* Update sent data bytes counter. */
4643 			txq->stats.obytes += rte_pktmbuf_data_len(loc->mbuf);
4644 			if (MLX5_TXOFF_CONFIG(VLAN) &&
4645 			    loc->mbuf->ol_flags & PKT_TX_VLAN_PKT)
4646 				txq->stats.obytes +=
4647 					sizeof(struct rte_vlan_hdr);
4648 #endif
4649 		}
4650 		++loc->pkts_sent;
4651 		--pkts_n;
4652 		if (unlikely(!pkts_n || !loc->elts_free || !loc->wqe_free))
4653 			return MLX5_TXCMP_CODE_EXIT;
4654 		loc->mbuf = *pkts++;
4655 		if (pkts_n > 1)
4656 			rte_prefetch0(*pkts);
4657 		ret = mlx5_tx_able_to_empw(txq, loc, olx, true);
4658 		if (unlikely(ret != MLX5_TXCMP_CODE_SINGLE))
4659 			return ret;
4660 	}
4661 	MLX5_ASSERT(false);
4662 }
4663 
4664 static __rte_always_inline enum mlx5_txcmp_code
4665 mlx5_tx_burst_single(struct mlx5_txq_data *restrict txq,
4666 		     struct rte_mbuf **restrict pkts,
4667 		     unsigned int pkts_n,
4668 		     struct mlx5_txq_local *restrict loc,
4669 		     unsigned int olx)
4670 {
4671 	enum mlx5_txcmp_code ret;
4672 
4673 	ret = mlx5_tx_able_to_empw(txq, loc, olx, false);
4674 	if (ret == MLX5_TXCMP_CODE_SINGLE)
4675 		goto ordinary_send;
4676 	MLX5_ASSERT(ret == MLX5_TXCMP_CODE_EMPW);
4677 	for (;;) {
4678 		/* Optimize for inline/no inline eMPW send. */
4679 		ret = (MLX5_TXOFF_CONFIG(INLINE)) ?
4680 			mlx5_tx_burst_empw_inline
4681 				(txq, pkts, pkts_n, loc, olx) :
4682 			mlx5_tx_burst_empw_simple
4683 				(txq, pkts, pkts_n, loc, olx);
4684 		if (ret != MLX5_TXCMP_CODE_SINGLE)
4685 			return ret;
4686 		/* The resources to send one packet should remain. */
4687 		MLX5_ASSERT(loc->elts_free && loc->wqe_free);
4688 ordinary_send:
4689 		ret = mlx5_tx_burst_single_send(txq, pkts, pkts_n, loc, olx);
4690 		MLX5_ASSERT(ret != MLX5_TXCMP_CODE_SINGLE);
4691 		if (ret != MLX5_TXCMP_CODE_EMPW)
4692 			return ret;
4693 		/* The resources to send one packet should remain. */
4694 		MLX5_ASSERT(loc->elts_free && loc->wqe_free);
4695 	}
4696 }
4697 
4698 /**
4699  * DPDK Tx callback template. This is configured template
4700  * used to generate routines optimized for specified offload setup.
4701  * One of this generated functions is chosen at SQ configuration
4702  * time.
4703  *
4704  * @param txq
4705  *   Generic pointer to TX queue structure.
4706  * @param[in] pkts
4707  *   Packets to transmit.
4708  * @param pkts_n
4709  *   Number of packets in array.
4710  * @param olx
4711  *   Configured offloads mask, presents the bits of MLX5_TXOFF_CONFIG_xxx
4712  *   values. Should be static to take compile time static configuration
4713  *   advantages.
4714  *
4715  * @return
4716  *   Number of packets successfully transmitted (<= pkts_n).
4717  */
4718 static __rte_always_inline uint16_t
4719 mlx5_tx_burst_tmpl(struct mlx5_txq_data *restrict txq,
4720 		   struct rte_mbuf **restrict pkts,
4721 		   uint16_t pkts_n,
4722 		   unsigned int olx)
4723 {
4724 	struct mlx5_txq_local loc;
4725 	enum mlx5_txcmp_code ret;
4726 	unsigned int part;
4727 
4728 	MLX5_ASSERT(txq->elts_s >= (uint16_t)(txq->elts_head - txq->elts_tail));
4729 	MLX5_ASSERT(txq->wqe_s >= (uint16_t)(txq->wqe_ci - txq->wqe_pi));
4730 	if (unlikely(!pkts_n))
4731 		return 0;
4732 	loc.pkts_sent = 0;
4733 	loc.pkts_copy = 0;
4734 	loc.wqe_last = NULL;
4735 
4736 send_loop:
4737 	loc.pkts_loop = loc.pkts_sent;
4738 	/*
4739 	 * Check if there are some CQEs, if any:
4740 	 * - process an encountered errors
4741 	 * - process the completed WQEs
4742 	 * - free related mbufs
4743 	 * - doorbell the NIC about processed CQEs
4744 	 */
4745 	rte_prefetch0(*(pkts + loc.pkts_sent));
4746 	mlx5_tx_handle_completion(txq, olx);
4747 	/*
4748 	 * Calculate the number of available resources - elts and WQEs.
4749 	 * There are two possible different scenarios:
4750 	 * - no data inlining into WQEs, one WQEBB may contains upto
4751 	 *   four packets, in this case elts become scarce resource
4752 	 * - data inlining into WQEs, one packet may require multiple
4753 	 *   WQEBBs, the WQEs become the limiting factor.
4754 	 */
4755 	MLX5_ASSERT(txq->elts_s >= (uint16_t)(txq->elts_head - txq->elts_tail));
4756 	loc.elts_free = txq->elts_s -
4757 				(uint16_t)(txq->elts_head - txq->elts_tail);
4758 	MLX5_ASSERT(txq->wqe_s >= (uint16_t)(txq->wqe_ci - txq->wqe_pi));
4759 	loc.wqe_free = txq->wqe_s -
4760 				(uint16_t)(txq->wqe_ci - txq->wqe_pi);
4761 	if (unlikely(!loc.elts_free || !loc.wqe_free))
4762 		goto burst_exit;
4763 	for (;;) {
4764 		/*
4765 		 * Fetch the packet from array. Usually this is
4766 		 * the first packet in series of multi/single
4767 		 * segment packets.
4768 		 */
4769 		loc.mbuf = *(pkts + loc.pkts_sent);
4770 		/* Dedicated branch for multi-segment packets. */
4771 		if (MLX5_TXOFF_CONFIG(MULTI) &&
4772 		    unlikely(NB_SEGS(loc.mbuf) > 1)) {
4773 			/*
4774 			 * Multi-segment packet encountered.
4775 			 * Hardware is able to process it only
4776 			 * with SEND/TSO opcodes, one packet
4777 			 * per WQE, do it in dedicated routine.
4778 			 */
4779 enter_send_multi:
4780 			MLX5_ASSERT(loc.pkts_sent >= loc.pkts_copy);
4781 			part = loc.pkts_sent - loc.pkts_copy;
4782 			if (!MLX5_TXOFF_CONFIG(INLINE) && part) {
4783 				/*
4784 				 * There are some single-segment mbufs not
4785 				 * stored in elts. The mbufs must be in the
4786 				 * same order as WQEs, so we must copy the
4787 				 * mbufs to elts here, before the coming
4788 				 * multi-segment packet mbufs is appended.
4789 				 */
4790 				mlx5_tx_copy_elts(txq, pkts + loc.pkts_copy,
4791 						  part, olx);
4792 				loc.pkts_copy = loc.pkts_sent;
4793 			}
4794 			MLX5_ASSERT(pkts_n > loc.pkts_sent);
4795 			ret = mlx5_tx_burst_mseg(txq, pkts, pkts_n, &loc, olx);
4796 			if (!MLX5_TXOFF_CONFIG(INLINE))
4797 				loc.pkts_copy = loc.pkts_sent;
4798 			/*
4799 			 * These returned code checks are supposed
4800 			 * to be optimized out due to routine inlining.
4801 			 */
4802 			if (ret == MLX5_TXCMP_CODE_EXIT) {
4803 				/*
4804 				 * The routine returns this code when
4805 				 * all packets are sent or there is no
4806 				 * enough resources to complete request.
4807 				 */
4808 				break;
4809 			}
4810 			if (ret == MLX5_TXCMP_CODE_ERROR) {
4811 				/*
4812 				 * The routine returns this code when
4813 				 * some error in the incoming packets
4814 				 * format occurred.
4815 				 */
4816 				txq->stats.oerrors++;
4817 				break;
4818 			}
4819 			if (ret == MLX5_TXCMP_CODE_SINGLE) {
4820 				/*
4821 				 * The single-segment packet was encountered
4822 				 * in the array, try to send it with the
4823 				 * best optimized way, possible engaging eMPW.
4824 				 */
4825 				goto enter_send_single;
4826 			}
4827 			if (MLX5_TXOFF_CONFIG(TSO) &&
4828 			    ret == MLX5_TXCMP_CODE_TSO) {
4829 				/*
4830 				 * The single-segment TSO packet was
4831 				 * encountered in the array.
4832 				 */
4833 				goto enter_send_tso;
4834 			}
4835 			/* We must not get here. Something is going wrong. */
4836 			MLX5_ASSERT(false);
4837 			txq->stats.oerrors++;
4838 			break;
4839 		}
4840 		/* Dedicated branch for single-segment TSO packets. */
4841 		if (MLX5_TXOFF_CONFIG(TSO) &&
4842 		    unlikely(loc.mbuf->ol_flags & PKT_TX_TCP_SEG)) {
4843 			/*
4844 			 * TSO might require special way for inlining
4845 			 * (dedicated parameters) and is sent with
4846 			 * MLX5_OPCODE_TSO opcode only, provide this
4847 			 * in dedicated branch.
4848 			 */
4849 enter_send_tso:
4850 			MLX5_ASSERT(NB_SEGS(loc.mbuf) == 1);
4851 			MLX5_ASSERT(pkts_n > loc.pkts_sent);
4852 			ret = mlx5_tx_burst_tso(txq, pkts, pkts_n, &loc, olx);
4853 			/*
4854 			 * These returned code checks are supposed
4855 			 * to be optimized out due to routine inlining.
4856 			 */
4857 			if (ret == MLX5_TXCMP_CODE_EXIT)
4858 				break;
4859 			if (ret == MLX5_TXCMP_CODE_ERROR) {
4860 				txq->stats.oerrors++;
4861 				break;
4862 			}
4863 			if (ret == MLX5_TXCMP_CODE_SINGLE)
4864 				goto enter_send_single;
4865 			if (MLX5_TXOFF_CONFIG(MULTI) &&
4866 			    ret == MLX5_TXCMP_CODE_MULTI) {
4867 				/*
4868 				 * The multi-segment packet was
4869 				 * encountered in the array.
4870 				 */
4871 				goto enter_send_multi;
4872 			}
4873 			/* We must not get here. Something is going wrong. */
4874 			MLX5_ASSERT(false);
4875 			txq->stats.oerrors++;
4876 			break;
4877 		}
4878 		/*
4879 		 * The dedicated branch for the single-segment packets
4880 		 * without TSO. Often these ones can be sent using
4881 		 * MLX5_OPCODE_EMPW with multiple packets in one WQE.
4882 		 * The routine builds the WQEs till it encounters
4883 		 * the TSO or multi-segment packet (in case if these
4884 		 * offloads are requested at SQ configuration time).
4885 		 */
4886 enter_send_single:
4887 		MLX5_ASSERT(pkts_n > loc.pkts_sent);
4888 		ret = mlx5_tx_burst_single(txq, pkts, pkts_n, &loc, olx);
4889 		/*
4890 		 * These returned code checks are supposed
4891 		 * to be optimized out due to routine inlining.
4892 		 */
4893 		if (ret == MLX5_TXCMP_CODE_EXIT)
4894 			break;
4895 		if (ret == MLX5_TXCMP_CODE_ERROR) {
4896 			txq->stats.oerrors++;
4897 			break;
4898 		}
4899 		if (MLX5_TXOFF_CONFIG(MULTI) &&
4900 		    ret == MLX5_TXCMP_CODE_MULTI) {
4901 			/*
4902 			 * The multi-segment packet was
4903 			 * encountered in the array.
4904 			 */
4905 			goto enter_send_multi;
4906 		}
4907 		if (MLX5_TXOFF_CONFIG(TSO) &&
4908 		    ret == MLX5_TXCMP_CODE_TSO) {
4909 			/*
4910 			 * The single-segment TSO packet was
4911 			 * encountered in the array.
4912 			 */
4913 			goto enter_send_tso;
4914 		}
4915 		/* We must not get here. Something is going wrong. */
4916 		MLX5_ASSERT(false);
4917 		txq->stats.oerrors++;
4918 		break;
4919 	}
4920 	/*
4921 	 * Main Tx loop is completed, do the rest:
4922 	 * - set completion request if thresholds are reached
4923 	 * - doorbell the hardware
4924 	 * - copy the rest of mbufs to elts (if any)
4925 	 */
4926 	MLX5_ASSERT(MLX5_TXOFF_CONFIG(INLINE) ||
4927 		    loc.pkts_sent >= loc.pkts_copy);
4928 	/* Take a shortcut if nothing is sent. */
4929 	if (unlikely(loc.pkts_sent == loc.pkts_loop))
4930 		goto burst_exit;
4931 	/* Request CQE generation if limits are reached. */
4932 	mlx5_tx_request_completion(txq, &loc, olx);
4933 	/*
4934 	 * Ring QP doorbell immediately after WQE building completion
4935 	 * to improve latencies. The pure software related data treatment
4936 	 * can be completed after doorbell. Tx CQEs for this SQ are
4937 	 * processed in this thread only by the polling.
4938 	 *
4939 	 * The rdma core library can map doorbell register in two ways,
4940 	 * depending on the environment variable "MLX5_SHUT_UP_BF":
4941 	 *
4942 	 * - as regular cached memory, the variable is either missing or
4943 	 *   set to zero. This type of mapping may cause the significant
4944 	 *   doorbell register writing latency and requires explicit
4945 	 *   memory write barrier to mitigate this issue and prevent
4946 	 *   write combining.
4947 	 *
4948 	 * - as non-cached memory, the variable is present and set to
4949 	 *   not "0" value. This type of mapping may cause performance
4950 	 *   impact under heavy loading conditions but the explicit write
4951 	 *   memory barrier is not required and it may improve core
4952 	 *   performance.
4953 	 *
4954 	 * - the legacy behaviour (prior 19.08 release) was to use some
4955 	 *   heuristics to decide whether write memory barrier should
4956 	 *   be performed. This behavior is supported with specifying
4957 	 *   tx_db_nc=2, write barrier is skipped if application
4958 	 *   provides the full recommended burst of packets, it
4959 	 *   supposes the next packets are coming and the write barrier
4960 	 *   will be issued on the next burst (after descriptor writing,
4961 	 *   at least).
4962 	 */
4963 	mlx5_tx_dbrec_cond_wmb(txq, loc.wqe_last, !txq->db_nc &&
4964 			(!txq->db_heu || pkts_n % MLX5_TX_DEFAULT_BURST));
4965 	/* Not all of the mbufs may be stored into elts yet. */
4966 	part = MLX5_TXOFF_CONFIG(INLINE) ? 0 : loc.pkts_sent - loc.pkts_copy;
4967 	if (!MLX5_TXOFF_CONFIG(INLINE) && part) {
4968 		/*
4969 		 * There are some single-segment mbufs not stored in elts.
4970 		 * It can be only if the last packet was single-segment.
4971 		 * The copying is gathered into one place due to it is
4972 		 * a good opportunity to optimize that with SIMD.
4973 		 * Unfortunately if inlining is enabled the gaps in
4974 		 * pointer array may happen due to early freeing of the
4975 		 * inlined mbufs.
4976 		 */
4977 		mlx5_tx_copy_elts(txq, pkts + loc.pkts_copy, part, olx);
4978 		loc.pkts_copy = loc.pkts_sent;
4979 	}
4980 	MLX5_ASSERT(txq->elts_s >= (uint16_t)(txq->elts_head - txq->elts_tail));
4981 	MLX5_ASSERT(txq->wqe_s >= (uint16_t)(txq->wqe_ci - txq->wqe_pi));
4982 	if (pkts_n > loc.pkts_sent) {
4983 		/*
4984 		 * If burst size is large there might be no enough CQE
4985 		 * fetched from completion queue and no enough resources
4986 		 * freed to send all the packets.
4987 		 */
4988 		goto send_loop;
4989 	}
4990 burst_exit:
4991 #ifdef MLX5_PMD_SOFT_COUNTERS
4992 	/* Increment sent packets counter. */
4993 	txq->stats.opackets += loc.pkts_sent;
4994 #endif
4995 	return loc.pkts_sent;
4996 }
4997 
4998 /* Generate routines with Enhanced Multi-Packet Write support. */
4999 MLX5_TXOFF_DECL(full_empw,
5000 		MLX5_TXOFF_CONFIG_FULL | MLX5_TXOFF_CONFIG_EMPW)
5001 
5002 MLX5_TXOFF_DECL(none_empw,
5003 		MLX5_TXOFF_CONFIG_NONE | MLX5_TXOFF_CONFIG_EMPW)
5004 
5005 MLX5_TXOFF_DECL(md_empw,
5006 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
5007 
5008 MLX5_TXOFF_DECL(mt_empw,
5009 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
5010 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
5011 
5012 MLX5_TXOFF_DECL(mtsc_empw,
5013 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
5014 		MLX5_TXOFF_CONFIG_SWP |	MLX5_TXOFF_CONFIG_CSUM |
5015 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
5016 
5017 MLX5_TXOFF_DECL(mti_empw,
5018 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
5019 		MLX5_TXOFF_CONFIG_INLINE |
5020 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
5021 
5022 MLX5_TXOFF_DECL(mtv_empw,
5023 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
5024 		MLX5_TXOFF_CONFIG_VLAN |
5025 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
5026 
5027 MLX5_TXOFF_DECL(mtiv_empw,
5028 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
5029 		MLX5_TXOFF_CONFIG_INLINE | MLX5_TXOFF_CONFIG_VLAN |
5030 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
5031 
5032 MLX5_TXOFF_DECL(sc_empw,
5033 		MLX5_TXOFF_CONFIG_SWP |	MLX5_TXOFF_CONFIG_CSUM |
5034 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
5035 
5036 MLX5_TXOFF_DECL(sci_empw,
5037 		MLX5_TXOFF_CONFIG_SWP |	MLX5_TXOFF_CONFIG_CSUM |
5038 		MLX5_TXOFF_CONFIG_INLINE |
5039 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
5040 
5041 MLX5_TXOFF_DECL(scv_empw,
5042 		MLX5_TXOFF_CONFIG_SWP |	MLX5_TXOFF_CONFIG_CSUM |
5043 		MLX5_TXOFF_CONFIG_VLAN |
5044 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
5045 
5046 MLX5_TXOFF_DECL(sciv_empw,
5047 		MLX5_TXOFF_CONFIG_SWP |	MLX5_TXOFF_CONFIG_CSUM |
5048 		MLX5_TXOFF_CONFIG_INLINE | MLX5_TXOFF_CONFIG_VLAN |
5049 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
5050 
5051 MLX5_TXOFF_DECL(i_empw,
5052 		MLX5_TXOFF_CONFIG_INLINE |
5053 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
5054 
5055 MLX5_TXOFF_DECL(v_empw,
5056 		MLX5_TXOFF_CONFIG_VLAN |
5057 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
5058 
5059 MLX5_TXOFF_DECL(iv_empw,
5060 		MLX5_TXOFF_CONFIG_INLINE | MLX5_TXOFF_CONFIG_VLAN |
5061 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
5062 
5063 /* Generate routines without Enhanced Multi-Packet Write support. */
5064 MLX5_TXOFF_DECL(full,
5065 		MLX5_TXOFF_CONFIG_FULL)
5066 
5067 MLX5_TXOFF_DECL(none,
5068 		MLX5_TXOFF_CONFIG_NONE)
5069 
5070 MLX5_TXOFF_DECL(md,
5071 		MLX5_TXOFF_CONFIG_METADATA)
5072 
5073 MLX5_TXOFF_DECL(mt,
5074 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
5075 		MLX5_TXOFF_CONFIG_METADATA)
5076 
5077 MLX5_TXOFF_DECL(mtsc,
5078 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
5079 		MLX5_TXOFF_CONFIG_SWP |	MLX5_TXOFF_CONFIG_CSUM |
5080 		MLX5_TXOFF_CONFIG_METADATA)
5081 
5082 MLX5_TXOFF_DECL(mti,
5083 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
5084 		MLX5_TXOFF_CONFIG_INLINE |
5085 		MLX5_TXOFF_CONFIG_METADATA)
5086 
5087 
5088 MLX5_TXOFF_DECL(mtv,
5089 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
5090 		MLX5_TXOFF_CONFIG_VLAN |
5091 		MLX5_TXOFF_CONFIG_METADATA)
5092 
5093 
5094 MLX5_TXOFF_DECL(mtiv,
5095 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
5096 		MLX5_TXOFF_CONFIG_INLINE | MLX5_TXOFF_CONFIG_VLAN |
5097 		MLX5_TXOFF_CONFIG_METADATA)
5098 
5099 MLX5_TXOFF_DECL(sc,
5100 		MLX5_TXOFF_CONFIG_SWP |	MLX5_TXOFF_CONFIG_CSUM |
5101 		MLX5_TXOFF_CONFIG_METADATA)
5102 
5103 MLX5_TXOFF_DECL(sci,
5104 		MLX5_TXOFF_CONFIG_SWP |	MLX5_TXOFF_CONFIG_CSUM |
5105 		MLX5_TXOFF_CONFIG_INLINE |
5106 		MLX5_TXOFF_CONFIG_METADATA)
5107 
5108 
5109 MLX5_TXOFF_DECL(scv,
5110 		MLX5_TXOFF_CONFIG_SWP |	MLX5_TXOFF_CONFIG_CSUM |
5111 		MLX5_TXOFF_CONFIG_VLAN |
5112 		MLX5_TXOFF_CONFIG_METADATA)
5113 
5114 
5115 MLX5_TXOFF_DECL(sciv,
5116 		MLX5_TXOFF_CONFIG_SWP |	MLX5_TXOFF_CONFIG_CSUM |
5117 		MLX5_TXOFF_CONFIG_INLINE | MLX5_TXOFF_CONFIG_VLAN |
5118 		MLX5_TXOFF_CONFIG_METADATA)
5119 
5120 MLX5_TXOFF_DECL(i,
5121 		MLX5_TXOFF_CONFIG_INLINE |
5122 		MLX5_TXOFF_CONFIG_METADATA)
5123 
5124 MLX5_TXOFF_DECL(v,
5125 		MLX5_TXOFF_CONFIG_VLAN |
5126 		MLX5_TXOFF_CONFIG_METADATA)
5127 
5128 MLX5_TXOFF_DECL(iv,
5129 		MLX5_TXOFF_CONFIG_INLINE | MLX5_TXOFF_CONFIG_VLAN |
5130 		MLX5_TXOFF_CONFIG_METADATA)
5131 
5132 /*
5133  * Generate routines with Legacy Multi-Packet Write support.
5134  * This mode is supported by ConnectX-4LX only and imposes
5135  * offload limitations, not supported:
5136  *   - ACL/Flows (metadata are becoming meaningless)
5137  *   - WQE Inline headers
5138  *   - SRIOV (E-Switch offloads)
5139  *   - VLAN insertion
5140  *   - tunnel encapsulation/decapsulation
5141  *   - TSO
5142  */
5143 MLX5_TXOFF_DECL(none_mpw,
5144 		MLX5_TXOFF_CONFIG_NONE | MLX5_TXOFF_CONFIG_EMPW |
5145 		MLX5_TXOFF_CONFIG_MPW)
5146 
5147 MLX5_TXOFF_DECL(mci_mpw,
5148 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_CSUM |
5149 		MLX5_TXOFF_CONFIG_INLINE | MLX5_TXOFF_CONFIG_EMPW |
5150 		MLX5_TXOFF_CONFIG_MPW)
5151 
5152 MLX5_TXOFF_DECL(mc_mpw,
5153 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_CSUM |
5154 		MLX5_TXOFF_CONFIG_EMPW | MLX5_TXOFF_CONFIG_MPW)
5155 
5156 MLX5_TXOFF_DECL(i_mpw,
5157 		MLX5_TXOFF_CONFIG_INLINE | MLX5_TXOFF_CONFIG_EMPW |
5158 		MLX5_TXOFF_CONFIG_MPW)
5159 
5160 /*
5161  * Array of declared and compiled Tx burst function and corresponding
5162  * supported offloads set. The array is used to select the Tx burst
5163  * function for specified offloads set at Tx queue configuration time.
5164  */
5165 const struct {
5166 	eth_tx_burst_t func;
5167 	unsigned int olx;
5168 } txoff_func[] = {
5169 MLX5_TXOFF_INFO(full_empw,
5170 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
5171 		MLX5_TXOFF_CONFIG_SWP |	MLX5_TXOFF_CONFIG_CSUM |
5172 		MLX5_TXOFF_CONFIG_INLINE | MLX5_TXOFF_CONFIG_VLAN |
5173 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
5174 
5175 MLX5_TXOFF_INFO(none_empw,
5176 		MLX5_TXOFF_CONFIG_NONE | MLX5_TXOFF_CONFIG_EMPW)
5177 
5178 MLX5_TXOFF_INFO(md_empw,
5179 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
5180 
5181 MLX5_TXOFF_INFO(mt_empw,
5182 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
5183 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
5184 
5185 MLX5_TXOFF_INFO(mtsc_empw,
5186 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
5187 		MLX5_TXOFF_CONFIG_SWP |	MLX5_TXOFF_CONFIG_CSUM |
5188 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
5189 
5190 MLX5_TXOFF_INFO(mti_empw,
5191 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
5192 		MLX5_TXOFF_CONFIG_INLINE |
5193 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
5194 
5195 MLX5_TXOFF_INFO(mtv_empw,
5196 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
5197 		MLX5_TXOFF_CONFIG_VLAN |
5198 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
5199 
5200 MLX5_TXOFF_INFO(mtiv_empw,
5201 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
5202 		MLX5_TXOFF_CONFIG_INLINE | MLX5_TXOFF_CONFIG_VLAN |
5203 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
5204 
5205 MLX5_TXOFF_INFO(sc_empw,
5206 		MLX5_TXOFF_CONFIG_SWP |	MLX5_TXOFF_CONFIG_CSUM |
5207 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
5208 
5209 MLX5_TXOFF_INFO(sci_empw,
5210 		MLX5_TXOFF_CONFIG_SWP |	MLX5_TXOFF_CONFIG_CSUM |
5211 		MLX5_TXOFF_CONFIG_INLINE |
5212 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
5213 
5214 MLX5_TXOFF_INFO(scv_empw,
5215 		MLX5_TXOFF_CONFIG_SWP |	MLX5_TXOFF_CONFIG_CSUM |
5216 		MLX5_TXOFF_CONFIG_VLAN |
5217 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
5218 
5219 MLX5_TXOFF_INFO(sciv_empw,
5220 		MLX5_TXOFF_CONFIG_SWP |	MLX5_TXOFF_CONFIG_CSUM |
5221 		MLX5_TXOFF_CONFIG_INLINE | MLX5_TXOFF_CONFIG_VLAN |
5222 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
5223 
5224 MLX5_TXOFF_INFO(i_empw,
5225 		MLX5_TXOFF_CONFIG_INLINE |
5226 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
5227 
5228 MLX5_TXOFF_INFO(v_empw,
5229 		MLX5_TXOFF_CONFIG_VLAN |
5230 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
5231 
5232 MLX5_TXOFF_INFO(iv_empw,
5233 		MLX5_TXOFF_CONFIG_INLINE | MLX5_TXOFF_CONFIG_VLAN |
5234 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
5235 
5236 MLX5_TXOFF_INFO(full,
5237 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
5238 		MLX5_TXOFF_CONFIG_SWP |	MLX5_TXOFF_CONFIG_CSUM |
5239 		MLX5_TXOFF_CONFIG_INLINE | MLX5_TXOFF_CONFIG_VLAN |
5240 		MLX5_TXOFF_CONFIG_METADATA)
5241 
5242 MLX5_TXOFF_INFO(none,
5243 		MLX5_TXOFF_CONFIG_NONE)
5244 
5245 MLX5_TXOFF_INFO(md,
5246 		MLX5_TXOFF_CONFIG_METADATA)
5247 
5248 MLX5_TXOFF_INFO(mt,
5249 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
5250 		MLX5_TXOFF_CONFIG_METADATA)
5251 
5252 MLX5_TXOFF_INFO(mtsc,
5253 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
5254 		MLX5_TXOFF_CONFIG_SWP |	MLX5_TXOFF_CONFIG_CSUM |
5255 		MLX5_TXOFF_CONFIG_METADATA)
5256 
5257 MLX5_TXOFF_INFO(mti,
5258 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
5259 		MLX5_TXOFF_CONFIG_INLINE |
5260 		MLX5_TXOFF_CONFIG_METADATA)
5261 
5262 MLX5_TXOFF_INFO(mtv,
5263 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
5264 		MLX5_TXOFF_CONFIG_VLAN |
5265 		MLX5_TXOFF_CONFIG_METADATA)
5266 
5267 MLX5_TXOFF_INFO(mtiv,
5268 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
5269 		MLX5_TXOFF_CONFIG_INLINE | MLX5_TXOFF_CONFIG_VLAN |
5270 		MLX5_TXOFF_CONFIG_METADATA)
5271 
5272 MLX5_TXOFF_INFO(sc,
5273 		MLX5_TXOFF_CONFIG_SWP |	MLX5_TXOFF_CONFIG_CSUM |
5274 		MLX5_TXOFF_CONFIG_METADATA)
5275 
5276 MLX5_TXOFF_INFO(sci,
5277 		MLX5_TXOFF_CONFIG_SWP |	MLX5_TXOFF_CONFIG_CSUM |
5278 		MLX5_TXOFF_CONFIG_INLINE |
5279 		MLX5_TXOFF_CONFIG_METADATA)
5280 
5281 MLX5_TXOFF_INFO(scv,
5282 		MLX5_TXOFF_CONFIG_SWP |	MLX5_TXOFF_CONFIG_CSUM |
5283 		MLX5_TXOFF_CONFIG_VLAN |
5284 		MLX5_TXOFF_CONFIG_METADATA)
5285 
5286 MLX5_TXOFF_INFO(sciv,
5287 		MLX5_TXOFF_CONFIG_SWP |	MLX5_TXOFF_CONFIG_CSUM |
5288 		MLX5_TXOFF_CONFIG_INLINE | MLX5_TXOFF_CONFIG_VLAN |
5289 		MLX5_TXOFF_CONFIG_METADATA)
5290 
5291 MLX5_TXOFF_INFO(i,
5292 		MLX5_TXOFF_CONFIG_INLINE |
5293 		MLX5_TXOFF_CONFIG_METADATA)
5294 
5295 MLX5_TXOFF_INFO(v,
5296 		MLX5_TXOFF_CONFIG_VLAN |
5297 		MLX5_TXOFF_CONFIG_METADATA)
5298 
5299 MLX5_TXOFF_INFO(iv,
5300 		MLX5_TXOFF_CONFIG_INLINE | MLX5_TXOFF_CONFIG_VLAN |
5301 		MLX5_TXOFF_CONFIG_METADATA)
5302 
5303 MLX5_TXOFF_INFO(none_mpw,
5304 		MLX5_TXOFF_CONFIG_NONE | MLX5_TXOFF_CONFIG_EMPW |
5305 		MLX5_TXOFF_CONFIG_MPW)
5306 
5307 MLX5_TXOFF_INFO(mci_mpw,
5308 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_CSUM |
5309 		MLX5_TXOFF_CONFIG_INLINE | MLX5_TXOFF_CONFIG_EMPW |
5310 		MLX5_TXOFF_CONFIG_MPW)
5311 
5312 MLX5_TXOFF_INFO(mc_mpw,
5313 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_CSUM |
5314 		MLX5_TXOFF_CONFIG_EMPW | MLX5_TXOFF_CONFIG_MPW)
5315 
5316 MLX5_TXOFF_INFO(i_mpw,
5317 		MLX5_TXOFF_CONFIG_INLINE | MLX5_TXOFF_CONFIG_EMPW |
5318 		MLX5_TXOFF_CONFIG_MPW)
5319 };
5320 
5321 /**
5322  * Configure the Tx function to use. The routine checks configured
5323  * Tx offloads for the device and selects appropriate Tx burst
5324  * routine. There are multiple Tx burst routines compiled from
5325  * the same template in the most optimal way for the dedicated
5326  * Tx offloads set.
5327  *
5328  * @param dev
5329  *   Pointer to private data structure.
5330  *
5331  * @return
5332  *   Pointer to selected Tx burst function.
5333  */
5334 eth_tx_burst_t
5335 mlx5_select_tx_function(struct rte_eth_dev *dev)
5336 {
5337 	struct mlx5_priv *priv = dev->data->dev_private;
5338 	struct mlx5_dev_config *config = &priv->config;
5339 	uint64_t tx_offloads = dev->data->dev_conf.txmode.offloads;
5340 	unsigned int diff = 0, olx = 0, i, m;
5341 
5342 	static_assert(MLX5_WQE_SIZE_MAX / MLX5_WSEG_SIZE <=
5343 		      MLX5_DSEG_MAX, "invalid WQE max size");
5344 	static_assert(MLX5_WQE_CSEG_SIZE == MLX5_WSEG_SIZE,
5345 		      "invalid WQE Control Segment size");
5346 	static_assert(MLX5_WQE_ESEG_SIZE == MLX5_WSEG_SIZE,
5347 		      "invalid WQE Ethernet Segment size");
5348 	static_assert(MLX5_WQE_DSEG_SIZE == MLX5_WSEG_SIZE,
5349 		      "invalid WQE Data Segment size");
5350 	static_assert(MLX5_WQE_SIZE == 4 * MLX5_WSEG_SIZE,
5351 		      "invalid WQE size");
5352 	MLX5_ASSERT(priv);
5353 	if (tx_offloads & DEV_TX_OFFLOAD_MULTI_SEGS) {
5354 		/* We should support Multi-Segment Packets. */
5355 		olx |= MLX5_TXOFF_CONFIG_MULTI;
5356 	}
5357 	if (tx_offloads & (DEV_TX_OFFLOAD_TCP_TSO |
5358 			   DEV_TX_OFFLOAD_VXLAN_TNL_TSO |
5359 			   DEV_TX_OFFLOAD_GRE_TNL_TSO |
5360 			   DEV_TX_OFFLOAD_IP_TNL_TSO |
5361 			   DEV_TX_OFFLOAD_UDP_TNL_TSO)) {
5362 		/* We should support TCP Send Offload. */
5363 		olx |= MLX5_TXOFF_CONFIG_TSO;
5364 	}
5365 	if (tx_offloads & (DEV_TX_OFFLOAD_IP_TNL_TSO |
5366 			   DEV_TX_OFFLOAD_UDP_TNL_TSO |
5367 			   DEV_TX_OFFLOAD_OUTER_IPV4_CKSUM)) {
5368 		/* We should support Software Parser for Tunnels. */
5369 		olx |= MLX5_TXOFF_CONFIG_SWP;
5370 	}
5371 	if (tx_offloads & (DEV_TX_OFFLOAD_IPV4_CKSUM |
5372 			   DEV_TX_OFFLOAD_UDP_CKSUM |
5373 			   DEV_TX_OFFLOAD_TCP_CKSUM |
5374 			   DEV_TX_OFFLOAD_OUTER_IPV4_CKSUM)) {
5375 		/* We should support IP/TCP/UDP Checksums. */
5376 		olx |= MLX5_TXOFF_CONFIG_CSUM;
5377 	}
5378 	if (tx_offloads & DEV_TX_OFFLOAD_VLAN_INSERT) {
5379 		/* We should support VLAN insertion. */
5380 		olx |= MLX5_TXOFF_CONFIG_VLAN;
5381 	}
5382 	if (priv->txqs_n && (*priv->txqs)[0]) {
5383 		struct mlx5_txq_data *txd = (*priv->txqs)[0];
5384 
5385 		if (txd->inlen_send) {
5386 			/*
5387 			 * Check the data inline requirements. Data inline
5388 			 * is enabled on per device basis, we can check
5389 			 * the first Tx queue only.
5390 			 *
5391 			 * If device does not support VLAN insertion in WQE
5392 			 * and some queues are requested to perform VLAN
5393 			 * insertion offload than inline must be enabled.
5394 			 */
5395 			olx |= MLX5_TXOFF_CONFIG_INLINE;
5396 		}
5397 	}
5398 	if (config->mps == MLX5_MPW_ENHANCED &&
5399 	    config->txq_inline_min <= 0) {
5400 		/*
5401 		 * The NIC supports Enhanced Multi-Packet Write
5402 		 * and does not require minimal inline data.
5403 		 */
5404 		olx |= MLX5_TXOFF_CONFIG_EMPW;
5405 	}
5406 	if (rte_flow_dynf_metadata_avail()) {
5407 		/* We should support Flow metadata. */
5408 		olx |= MLX5_TXOFF_CONFIG_METADATA;
5409 	}
5410 	if (config->mps == MLX5_MPW) {
5411 		/*
5412 		 * The NIC supports Legacy Multi-Packet Write.
5413 		 * The MLX5_TXOFF_CONFIG_MPW controls the
5414 		 * descriptor building method in combination
5415 		 * with MLX5_TXOFF_CONFIG_EMPW.
5416 		 */
5417 		if (!(olx & (MLX5_TXOFF_CONFIG_TSO |
5418 			     MLX5_TXOFF_CONFIG_SWP |
5419 			     MLX5_TXOFF_CONFIG_VLAN |
5420 			     MLX5_TXOFF_CONFIG_METADATA)))
5421 			olx |= MLX5_TXOFF_CONFIG_EMPW |
5422 			       MLX5_TXOFF_CONFIG_MPW;
5423 	}
5424 	/*
5425 	 * Scan the routines table to find the minimal
5426 	 * satisfying routine with requested offloads.
5427 	 */
5428 	m = RTE_DIM(txoff_func);
5429 	for (i = 0; i < RTE_DIM(txoff_func); i++) {
5430 		unsigned int tmp;
5431 
5432 		tmp = txoff_func[i].olx;
5433 		if (tmp == olx) {
5434 			/* Meets requested offloads exactly.*/
5435 			m = i;
5436 			break;
5437 		}
5438 		if ((tmp & olx) != olx) {
5439 			/* Does not meet requested offloads at all. */
5440 			continue;
5441 		}
5442 		if ((olx ^ tmp) & MLX5_TXOFF_CONFIG_EMPW)
5443 			/* Do not enable eMPW if not configured. */
5444 			continue;
5445 		if ((olx ^ tmp) & MLX5_TXOFF_CONFIG_INLINE)
5446 			/* Do not enable inlining if not configured. */
5447 			continue;
5448 		/*
5449 		 * Some routine meets the requirements.
5450 		 * Check whether it has minimal amount
5451 		 * of not requested offloads.
5452 		 */
5453 		tmp = __builtin_popcountl(tmp & ~olx);
5454 		if (m >= RTE_DIM(txoff_func) || tmp < diff) {
5455 			/* First or better match, save and continue. */
5456 			m = i;
5457 			diff = tmp;
5458 			continue;
5459 		}
5460 		if (tmp == diff) {
5461 			tmp = txoff_func[i].olx ^ txoff_func[m].olx;
5462 			if (__builtin_ffsl(txoff_func[i].olx & ~tmp) <
5463 			    __builtin_ffsl(txoff_func[m].olx & ~tmp)) {
5464 				/* Lighter not requested offload. */
5465 				m = i;
5466 			}
5467 		}
5468 	}
5469 	if (m >= RTE_DIM(txoff_func)) {
5470 		DRV_LOG(DEBUG, "port %u has no selected Tx function"
5471 			       " for requested offloads %04X",
5472 				dev->data->port_id, olx);
5473 		return NULL;
5474 	}
5475 	DRV_LOG(DEBUG, "port %u has selected Tx function"
5476 		       " supporting offloads %04X/%04X",
5477 			dev->data->port_id, olx, txoff_func[m].olx);
5478 	if (txoff_func[m].olx & MLX5_TXOFF_CONFIG_MULTI)
5479 		DRV_LOG(DEBUG, "\tMULTI (multi segment)");
5480 	if (txoff_func[m].olx & MLX5_TXOFF_CONFIG_TSO)
5481 		DRV_LOG(DEBUG, "\tTSO   (TCP send offload)");
5482 	if (txoff_func[m].olx & MLX5_TXOFF_CONFIG_SWP)
5483 		DRV_LOG(DEBUG, "\tSWP   (software parser)");
5484 	if (txoff_func[m].olx & MLX5_TXOFF_CONFIG_CSUM)
5485 		DRV_LOG(DEBUG, "\tCSUM  (checksum offload)");
5486 	if (txoff_func[m].olx & MLX5_TXOFF_CONFIG_INLINE)
5487 		DRV_LOG(DEBUG, "\tINLIN (inline data)");
5488 	if (txoff_func[m].olx & MLX5_TXOFF_CONFIG_VLAN)
5489 		DRV_LOG(DEBUG, "\tVLANI (VLAN insertion)");
5490 	if (txoff_func[m].olx & MLX5_TXOFF_CONFIG_METADATA)
5491 		DRV_LOG(DEBUG, "\tMETAD (tx Flow metadata)");
5492 	if (txoff_func[m].olx & MLX5_TXOFF_CONFIG_EMPW) {
5493 		if (txoff_func[m].olx & MLX5_TXOFF_CONFIG_MPW)
5494 			DRV_LOG(DEBUG, "\tMPW   (Legacy MPW)");
5495 		else
5496 			DRV_LOG(DEBUG, "\tEMPW  (Enhanced MPW)");
5497 	}
5498 	return txoff_func[m].func;
5499 }
5500 
5501 /**
5502  * DPDK callback to get the TX queue information
5503  *
5504  * @param dev
5505  *   Pointer to the device structure.
5506  *
5507  * @param tx_queue_id
5508  *   Tx queue identificator.
5509  *
5510  * @param qinfo
5511  *   Pointer to the TX queue information structure.
5512  *
5513  * @return
5514  *   None.
5515  */
5516 
5517 void
5518 mlx5_txq_info_get(struct rte_eth_dev *dev, uint16_t tx_queue_id,
5519 		  struct rte_eth_txq_info *qinfo)
5520 {
5521 	struct mlx5_priv *priv = dev->data->dev_private;
5522 	struct mlx5_txq_data *txq = (*priv->txqs)[tx_queue_id];
5523 	struct mlx5_txq_ctrl *txq_ctrl =
5524 			container_of(txq, struct mlx5_txq_ctrl, txq);
5525 
5526 	if (!txq)
5527 		return;
5528 	qinfo->nb_desc = txq->elts_s;
5529 	qinfo->conf.tx_thresh.pthresh = 0;
5530 	qinfo->conf.tx_thresh.hthresh = 0;
5531 	qinfo->conf.tx_thresh.wthresh = 0;
5532 	qinfo->conf.tx_rs_thresh = 0;
5533 	qinfo->conf.tx_free_thresh = 0;
5534 	qinfo->conf.tx_deferred_start = txq_ctrl ? 0 : 1;
5535 	qinfo->conf.offloads = dev->data->dev_conf.txmode.offloads;
5536 }
5537 
5538 /**
5539  * DPDK callback to get the TX packet burst mode information
5540  *
5541  * @param dev
5542  *   Pointer to the device structure.
5543  *
5544  * @param tx_queue_id
5545  *   Tx queue identificatior.
5546  *
5547  * @param mode
5548  *   Pointer to the burts mode information.
5549  *
5550  * @return
5551  *   0 as success, -EINVAL as failure.
5552  */
5553 
5554 int
5555 mlx5_tx_burst_mode_get(struct rte_eth_dev *dev,
5556 		       uint16_t tx_queue_id __rte_unused,
5557 		       struct rte_eth_burst_mode *mode)
5558 {
5559 	eth_tx_burst_t pkt_burst = dev->tx_pkt_burst;
5560 	unsigned int i, olx;
5561 
5562 	for (i = 0; i < RTE_DIM(txoff_func); i++) {
5563 		if (pkt_burst == txoff_func[i].func) {
5564 			olx = txoff_func[i].olx;
5565 			snprintf(mode->info, sizeof(mode->info),
5566 				 "%s%s%s%s%s%s%s%s",
5567 				 (olx & MLX5_TXOFF_CONFIG_EMPW) ?
5568 				 ((olx & MLX5_TXOFF_CONFIG_MPW) ?
5569 				 "Legacy MPW" : "Enhanced MPW") : "No MPW",
5570 				 (olx & MLX5_TXOFF_CONFIG_MULTI) ?
5571 				 " + MULTI" : "",
5572 				 (olx & MLX5_TXOFF_CONFIG_TSO) ?
5573 				 " + TSO" : "",
5574 				 (olx & MLX5_TXOFF_CONFIG_SWP) ?
5575 				 " + SWP" : "",
5576 				 (olx & MLX5_TXOFF_CONFIG_CSUM) ?
5577 				 "  + CSUM" : "",
5578 				 (olx & MLX5_TXOFF_CONFIG_INLINE) ?
5579 				 " + INLINE" : "",
5580 				 (olx & MLX5_TXOFF_CONFIG_VLAN) ?
5581 				 " + VLAN" : "",
5582 				 (olx & MLX5_TXOFF_CONFIG_METADATA) ?
5583 				 " + METADATA" : "");
5584 			return 0;
5585 		}
5586 	}
5587 	return -EINVAL;
5588 }
5589