xref: /dpdk/drivers/net/mlx5/mlx5_tx.h (revision 2490bb897182f57de80fd924dd3ae48dda819b8c)
1 /* SPDX-License-Identifier: BSD-3-Clause
2  * Copyright 2021 6WIND S.A.
3  * Copyright 2021 Mellanox Technologies, Ltd
4  */
5 
6 #ifndef RTE_PMD_MLX5_TX_H_
7 #define RTE_PMD_MLX5_TX_H_
8 
9 #include <stdint.h>
10 #include <sys/queue.h>
11 
12 #include <rte_mbuf.h>
13 #include <rte_mempool.h>
14 #include <rte_common.h>
15 #include <rte_spinlock.h>
16 
17 #include <mlx5_common.h>
18 #include <mlx5_common_mr.h>
19 
20 #include "mlx5.h"
21 #include "mlx5_autoconf.h"
22 
23 /* TX burst subroutines return codes. */
24 enum mlx5_txcmp_code {
25 	MLX5_TXCMP_CODE_EXIT = 0,
26 	MLX5_TXCMP_CODE_ERROR,
27 	MLX5_TXCMP_CODE_SINGLE,
28 	MLX5_TXCMP_CODE_MULTI,
29 	MLX5_TXCMP_CODE_TSO,
30 	MLX5_TXCMP_CODE_EMPW,
31 };
32 
33 /*
34  * These defines are used to configure Tx burst routine option set supported
35  * at compile time. The not specified options are optimized out due to if
36  * conditions can be explicitly calculated at compile time.
37  * The offloads with bigger runtime check (require more CPU cycles toskip)
38  * overhead should have the bigger index - this is needed to select the better
39  * matching routine function if no exact match and some offloads are not
40  * actually requested.
41  */
42 #define MLX5_TXOFF_CONFIG_MULTI (1u << 0) /* Multi-segment packets.*/
43 #define MLX5_TXOFF_CONFIG_TSO (1u << 1) /* TCP send offload supported.*/
44 #define MLX5_TXOFF_CONFIG_SWP (1u << 2) /* Tunnels/SW Parser offloads.*/
45 #define MLX5_TXOFF_CONFIG_CSUM (1u << 3) /* Check Sums offloaded. */
46 #define MLX5_TXOFF_CONFIG_INLINE (1u << 4) /* Data inlining supported. */
47 #define MLX5_TXOFF_CONFIG_VLAN (1u << 5) /* VLAN insertion supported.*/
48 #define MLX5_TXOFF_CONFIG_METADATA (1u << 6) /* Flow metadata. */
49 #define MLX5_TXOFF_CONFIG_EMPW (1u << 8) /* Enhanced MPW supported.*/
50 #define MLX5_TXOFF_CONFIG_MPW (1u << 9) /* Legacy MPW supported.*/
51 #define MLX5_TXOFF_CONFIG_TXPP (1u << 10) /* Scheduling on timestamp.*/
52 
53 /* The most common offloads groups. */
54 #define MLX5_TXOFF_CONFIG_NONE 0
55 #define MLX5_TXOFF_CONFIG_FULL (MLX5_TXOFF_CONFIG_MULTI | \
56 				MLX5_TXOFF_CONFIG_TSO | \
57 				MLX5_TXOFF_CONFIG_SWP | \
58 				MLX5_TXOFF_CONFIG_CSUM | \
59 				MLX5_TXOFF_CONFIG_INLINE | \
60 				MLX5_TXOFF_CONFIG_VLAN | \
61 				MLX5_TXOFF_CONFIG_METADATA)
62 
63 #define MLX5_TXOFF_CONFIG(mask) (olx & MLX5_TXOFF_CONFIG_##mask)
64 
65 #define MLX5_TXOFF_PRE_DECL(func) \
66 uint16_t mlx5_tx_burst_##func(void *txq, \
67 			      struct rte_mbuf **pkts, \
68 			      uint16_t pkts_n)
69 
70 #define MLX5_TXOFF_DECL(func, olx) \
71 uint16_t mlx5_tx_burst_##func(void *txq, \
72 			      struct rte_mbuf **pkts, \
73 			      uint16_t pkts_n) \
74 { \
75 	return mlx5_tx_burst_tmpl((struct mlx5_txq_data *)txq, \
76 		    pkts, pkts_n, (olx)); \
77 }
78 
79 /* Mbuf dynamic flag offset for inline. */
80 extern uint64_t rte_net_mlx5_dynf_inline_mask;
81 #define RTE_MBUF_F_TX_DYNF_NOINLINE rte_net_mlx5_dynf_inline_mask
82 
83 extern uint32_t mlx5_ptype_table[] __rte_cache_aligned;
84 extern uint8_t mlx5_cksum_table[1 << 10] __rte_cache_aligned;
85 extern uint8_t mlx5_swp_types_table[1 << 10] __rte_cache_aligned;
86 
87 struct mlx5_txq_stats {
88 #ifdef MLX5_PMD_SOFT_COUNTERS
89 	uint64_t opackets; /**< Total of successfully sent packets. */
90 	uint64_t obytes; /**< Total of successfully sent bytes. */
91 #endif
92 	uint64_t oerrors; /**< Total number of failed transmitted packets. */
93 };
94 
95 /* TX queue send local data. */
96 __extension__
97 struct mlx5_txq_local {
98 	struct mlx5_wqe *wqe_last; /* last sent WQE pointer. */
99 	struct rte_mbuf *mbuf; /* first mbuf to process. */
100 	uint16_t pkts_copy; /* packets copied to elts. */
101 	uint16_t pkts_sent; /* packets sent. */
102 	uint16_t pkts_loop; /* packets sent on loop entry. */
103 	uint16_t elts_free; /* available elts remain. */
104 	uint16_t wqe_free; /* available wqe remain. */
105 	uint16_t mbuf_off; /* data offset in current mbuf. */
106 	uint16_t mbuf_nseg; /* number of remaining mbuf. */
107 	uint16_t mbuf_free; /* number of inline mbufs to free. */
108 };
109 
110 /* TX queue descriptor. */
111 __extension__
112 struct mlx5_txq_data {
113 	uint16_t elts_head; /* Current counter in (*elts)[]. */
114 	uint16_t elts_tail; /* Counter of first element awaiting completion. */
115 	uint16_t elts_comp; /* elts index since last completion request. */
116 	uint16_t elts_s; /* Number of mbuf elements. */
117 	uint16_t elts_m; /* Mask for mbuf elements indices. */
118 	/* Fields related to elts mbuf storage. */
119 	uint16_t wqe_ci; /* Consumer index for work queue. */
120 	uint16_t wqe_pi; /* Producer index for work queue. */
121 	uint16_t wqe_s; /* Number of WQ elements. */
122 	uint16_t wqe_m; /* Mask Number for WQ elements. */
123 	uint16_t wqe_comp; /* WQE index since last completion request. */
124 	uint16_t wqe_thres; /* WQE threshold to request completion in CQ. */
125 	/* WQ related fields. */
126 	uint16_t cq_ci; /* Consumer index for completion queue. */
127 	uint16_t cq_pi; /* Production index for completion queue. */
128 	uint16_t cqe_s; /* Number of CQ elements. */
129 	uint16_t cqe_m; /* Mask for CQ indices. */
130 	/* CQ related fields. */
131 	uint16_t elts_n:4; /* elts[] length (in log2). */
132 	uint16_t cqe_n:4; /* Number of CQ elements (in log2). */
133 	uint16_t wqe_n:4; /* Number of WQ elements (in log2). */
134 	uint16_t tso_en:1; /* When set hardware TSO is enabled. */
135 	uint16_t tunnel_en:1;
136 	/* When set TX offload for tunneled packets are supported. */
137 	uint16_t swp_en:1; /* Whether SW parser is enabled. */
138 	uint16_t vlan_en:1; /* VLAN insertion in WQE is supported. */
139 	uint16_t db_nc:1; /* Doorbell mapped to non-cached region. */
140 	uint16_t db_heu:1; /* Doorbell heuristic write barrier. */
141 	uint16_t fast_free:1; /* mbuf fast free on Tx is enabled. */
142 	uint16_t inlen_send; /* Ordinary send data inline size. */
143 	uint16_t inlen_empw; /* eMPW max packet size to inline. */
144 	uint16_t inlen_mode; /* Minimal data length to inline. */
145 	uint32_t qp_num_8s; /* QP number shifted by 8. */
146 	uint64_t offloads; /* Offloads for Tx Queue. */
147 	struct mlx5_mr_ctrl mr_ctrl; /* MR control descriptor. */
148 	struct mlx5_wqe *wqes; /* Work queue. */
149 	struct mlx5_wqe *wqes_end; /* Work queue array limit. */
150 #ifdef RTE_LIBRTE_MLX5_DEBUG
151 	uint32_t *fcqs; /* Free completion queue (debug extended). */
152 #else
153 	uint16_t *fcqs; /* Free completion queue. */
154 #endif
155 	volatile struct mlx5_cqe *cqes; /* Completion queue. */
156 	volatile uint32_t *qp_db; /* Work queue doorbell. */
157 	volatile uint32_t *cq_db; /* Completion queue doorbell. */
158 	uint16_t port_id; /* Port ID of device. */
159 	uint16_t idx; /* Queue index. */
160 	uint64_t ts_mask; /* Timestamp flag dynamic mask. */
161 	int32_t ts_offset; /* Timestamp field dynamic offset. */
162 	struct mlx5_dev_ctx_shared *sh; /* Shared context. */
163 	struct mlx5_txq_stats stats; /* TX queue counters. */
164 	struct mlx5_uar_data uar_data;
165 	struct rte_mbuf *elts[0];
166 	/* Storage for queued packets, must be the last field. */
167 } __rte_cache_aligned;
168 
169 enum mlx5_txq_type {
170 	MLX5_TXQ_TYPE_STANDARD, /* Standard Tx queue. */
171 	MLX5_TXQ_TYPE_HAIRPIN, /* Hairpin Tx queue. */
172 };
173 
174 /* TX queue control descriptor. */
175 struct mlx5_txq_ctrl {
176 	LIST_ENTRY(mlx5_txq_ctrl) next; /* Pointer to the next element. */
177 	uint32_t refcnt; /* Reference counter. */
178 	unsigned int socket; /* CPU socket ID for allocations. */
179 	enum mlx5_txq_type type; /* The txq ctrl type. */
180 	unsigned int max_inline_data; /* Max inline data. */
181 	unsigned int max_tso_header; /* Max TSO header size. */
182 	struct mlx5_txq_obj *obj; /* Verbs/DevX queue object. */
183 	struct mlx5_priv *priv; /* Back pointer to private data. */
184 	off_t uar_mmap_offset; /* UAR mmap offset for non-primary process. */
185 	uint16_t dump_file_n; /* Number of dump files. */
186 	struct rte_eth_hairpin_conf hairpin_conf; /* Hairpin configuration. */
187 	uint32_t hairpin_status; /* Hairpin binding status. */
188 	struct mlx5_txq_data txq; /* Data path structure. */
189 	/* Must be the last field in the structure, contains elts[]. */
190 };
191 
192 /* mlx5_txq.c */
193 
194 int mlx5_tx_queue_start(struct rte_eth_dev *dev, uint16_t queue_id);
195 int mlx5_tx_queue_stop(struct rte_eth_dev *dev, uint16_t queue_id);
196 int mlx5_tx_queue_start_primary(struct rte_eth_dev *dev, uint16_t queue_id);
197 int mlx5_tx_queue_stop_primary(struct rte_eth_dev *dev, uint16_t queue_id);
198 int mlx5_tx_queue_setup(struct rte_eth_dev *dev, uint16_t idx, uint16_t desc,
199 			unsigned int socket, const struct rte_eth_txconf *conf);
200 int mlx5_tx_hairpin_queue_setup
201 	(struct rte_eth_dev *dev, uint16_t idx, uint16_t desc,
202 	 const struct rte_eth_hairpin_conf *hairpin_conf);
203 void mlx5_tx_queue_release(struct rte_eth_dev *dev, uint16_t qid);
204 int mlx5_tx_uar_init_secondary(struct rte_eth_dev *dev, int fd);
205 void mlx5_tx_uar_uninit_secondary(struct rte_eth_dev *dev);
206 int mlx5_txq_obj_verify(struct rte_eth_dev *dev);
207 struct mlx5_txq_ctrl *mlx5_txq_new(struct rte_eth_dev *dev, uint16_t idx,
208 				   uint16_t desc, unsigned int socket,
209 				   const struct rte_eth_txconf *conf);
210 struct mlx5_txq_ctrl *mlx5_txq_hairpin_new
211 	(struct rte_eth_dev *dev, uint16_t idx, uint16_t desc,
212 	 const struct rte_eth_hairpin_conf *hairpin_conf);
213 struct mlx5_txq_ctrl *mlx5_txq_get(struct rte_eth_dev *dev, uint16_t idx);
214 int mlx5_txq_release(struct rte_eth_dev *dev, uint16_t idx);
215 int mlx5_txq_releasable(struct rte_eth_dev *dev, uint16_t idx);
216 int mlx5_txq_verify(struct rte_eth_dev *dev);
217 void txq_alloc_elts(struct mlx5_txq_ctrl *txq_ctrl);
218 void txq_free_elts(struct mlx5_txq_ctrl *txq_ctrl);
219 uint64_t mlx5_get_tx_port_offloads(struct rte_eth_dev *dev);
220 void mlx5_txq_dynf_timestamp_set(struct rte_eth_dev *dev);
221 
222 /* mlx5_tx.c */
223 
224 uint16_t removed_tx_burst(void *dpdk_txq, struct rte_mbuf **pkts,
225 			  uint16_t pkts_n);
226 void mlx5_tx_handle_completion(struct mlx5_txq_data *__rte_restrict txq,
227 			       unsigned int olx __rte_unused);
228 int mlx5_tx_descriptor_status(void *tx_queue, uint16_t offset);
229 void mlx5_txq_info_get(struct rte_eth_dev *dev, uint16_t queue_id,
230 		       struct rte_eth_txq_info *qinfo);
231 int mlx5_tx_burst_mode_get(struct rte_eth_dev *dev, uint16_t tx_queue_id,
232 			   struct rte_eth_burst_mode *mode);
233 
234 /* mlx5_tx_empw.c */
235 
236 MLX5_TXOFF_PRE_DECL(full_empw);
237 MLX5_TXOFF_PRE_DECL(none_empw);
238 MLX5_TXOFF_PRE_DECL(md_empw);
239 MLX5_TXOFF_PRE_DECL(mt_empw);
240 MLX5_TXOFF_PRE_DECL(mtsc_empw);
241 MLX5_TXOFF_PRE_DECL(mti_empw);
242 MLX5_TXOFF_PRE_DECL(mtv_empw);
243 MLX5_TXOFF_PRE_DECL(mtiv_empw);
244 MLX5_TXOFF_PRE_DECL(sc_empw);
245 MLX5_TXOFF_PRE_DECL(sci_empw);
246 MLX5_TXOFF_PRE_DECL(scv_empw);
247 MLX5_TXOFF_PRE_DECL(sciv_empw);
248 MLX5_TXOFF_PRE_DECL(i_empw);
249 MLX5_TXOFF_PRE_DECL(v_empw);
250 MLX5_TXOFF_PRE_DECL(iv_empw);
251 
252 /* mlx5_tx_nompw.c */
253 
254 MLX5_TXOFF_PRE_DECL(full);
255 MLX5_TXOFF_PRE_DECL(none);
256 MLX5_TXOFF_PRE_DECL(md);
257 MLX5_TXOFF_PRE_DECL(mt);
258 MLX5_TXOFF_PRE_DECL(mtsc);
259 MLX5_TXOFF_PRE_DECL(mti);
260 MLX5_TXOFF_PRE_DECL(mtv);
261 MLX5_TXOFF_PRE_DECL(mtiv);
262 MLX5_TXOFF_PRE_DECL(sc);
263 MLX5_TXOFF_PRE_DECL(sci);
264 MLX5_TXOFF_PRE_DECL(scv);
265 MLX5_TXOFF_PRE_DECL(sciv);
266 MLX5_TXOFF_PRE_DECL(i);
267 MLX5_TXOFF_PRE_DECL(v);
268 MLX5_TXOFF_PRE_DECL(iv);
269 
270 /* mlx5_tx_txpp.c */
271 
272 MLX5_TXOFF_PRE_DECL(full_ts_nompw);
273 MLX5_TXOFF_PRE_DECL(full_ts_nompwi);
274 MLX5_TXOFF_PRE_DECL(full_ts);
275 MLX5_TXOFF_PRE_DECL(full_ts_noi);
276 MLX5_TXOFF_PRE_DECL(none_ts);
277 MLX5_TXOFF_PRE_DECL(mdi_ts);
278 MLX5_TXOFF_PRE_DECL(mti_ts);
279 MLX5_TXOFF_PRE_DECL(mtiv_ts);
280 
281 /* mlx5_tx_mpw.c */
282 
283 MLX5_TXOFF_PRE_DECL(none_mpw);
284 MLX5_TXOFF_PRE_DECL(mci_mpw);
285 MLX5_TXOFF_PRE_DECL(mc_mpw);
286 MLX5_TXOFF_PRE_DECL(i_mpw);
287 
288 static __rte_always_inline struct mlx5_uar_data *
289 mlx5_tx_bfreg(struct mlx5_txq_data *txq)
290 {
291 	return &MLX5_PROC_PRIV(txq->port_id)->uar_table[txq->idx];
292 }
293 
294 /**
295  * Ring TX queue doorbell and flush the update by write memory barrier.
296  *
297  * @param txq
298  *   Pointer to TX queue structure.
299  * @param wqe
300  *   Pointer to the last WQE posted in the NIC.
301  */
302 static __rte_always_inline void
303 mlx5_tx_dbrec(struct mlx5_txq_data *txq, volatile struct mlx5_wqe *wqe)
304 {
305 	mlx5_doorbell_ring(mlx5_tx_bfreg(txq), *(volatile uint64_t *)wqe,
306 			   txq->wqe_ci, txq->qp_db, 1);
307 }
308 
309 /**
310  * Convert timestamp from mbuf format to linear counter
311  * of Clock Queue completions (24 bits).
312  *
313  * @param sh
314  *   Pointer to the device shared context to fetch Tx
315  *   packet pacing timestamp and parameters.
316  * @param ts
317  *   Timestamp from mbuf to convert.
318  * @return
319  *   positive or zero value - completion ID to wait.
320  *   negative value - conversion error.
321  */
322 static __rte_always_inline int32_t
323 mlx5_txpp_convert_tx_ts(struct mlx5_dev_ctx_shared *sh, uint64_t mts)
324 {
325 	uint64_t ts, ci;
326 	uint32_t tick;
327 
328 	do {
329 		/*
330 		 * Read atomically two uint64_t fields and compare lsb bits.
331 		 * It there is no match - the timestamp was updated in
332 		 * the service thread, data should be re-read.
333 		 */
334 		rte_compiler_barrier();
335 		ci = __atomic_load_n(&sh->txpp.ts.ci_ts, __ATOMIC_RELAXED);
336 		ts = __atomic_load_n(&sh->txpp.ts.ts, __ATOMIC_RELAXED);
337 		rte_compiler_barrier();
338 		if (!((ts ^ ci) << (64 - MLX5_CQ_INDEX_WIDTH)))
339 			break;
340 	} while (true);
341 	/* Perform the skew correction, positive value to send earlier. */
342 	mts -= sh->txpp.skew;
343 	mts -= ts;
344 	if (unlikely(mts >= UINT64_MAX / 2)) {
345 		/* We have negative integer, mts is in the past. */
346 		__atomic_fetch_add(&sh->txpp.err_ts_past,
347 				   1, __ATOMIC_RELAXED);
348 		return -1;
349 	}
350 	tick = sh->txpp.tick;
351 	MLX5_ASSERT(tick);
352 	/* Convert delta to completions, round up. */
353 	mts = (mts + tick - 1) / tick;
354 	if (unlikely(mts >= (1 << MLX5_CQ_INDEX_WIDTH) / 2 - 1)) {
355 		/* We have mts is too distant future. */
356 		__atomic_fetch_add(&sh->txpp.err_ts_future,
357 				   1, __ATOMIC_RELAXED);
358 		return -1;
359 	}
360 	mts <<= 64 - MLX5_CQ_INDEX_WIDTH;
361 	ci += mts;
362 	ci >>= 64 - MLX5_CQ_INDEX_WIDTH;
363 	return ci;
364 }
365 
366 /**
367  * Set Software Parser flags and offsets in Ethernet Segment of WQE.
368  * Flags must be preliminary initialized to zero.
369  *
370  * @param loc
371  *   Pointer to burst routine local context.
372  * @param swp_flags
373  *   Pointer to store Software Parser flags.
374  * @param olx
375  *   Configured Tx offloads mask. It is fully defined at
376  *   compile time and may be used for optimization.
377  *
378  * @return
379  *   Software Parser offsets packed in dword.
380  *   Software Parser flags are set by pointer.
381  */
382 static __rte_always_inline uint32_t
383 txq_mbuf_to_swp(struct mlx5_txq_local *__rte_restrict loc,
384 		uint8_t *swp_flags,
385 		unsigned int olx)
386 {
387 	uint64_t ol, tunnel;
388 	unsigned int idx, off;
389 	uint32_t set;
390 
391 	if (!MLX5_TXOFF_CONFIG(SWP))
392 		return 0;
393 	ol = loc->mbuf->ol_flags;
394 	tunnel = ol & RTE_MBUF_F_TX_TUNNEL_MASK;
395 	/*
396 	 * Check whether Software Parser is required.
397 	 * Only customized tunnels may ask for.
398 	 */
399 	if (likely(tunnel != RTE_MBUF_F_TX_TUNNEL_UDP && tunnel != RTE_MBUF_F_TX_TUNNEL_IP))
400 		return 0;
401 	/*
402 	 * The index should have:
403 	 * bit[0:1] = RTE_MBUF_F_TX_L4_MASK
404 	 * bit[4] = RTE_MBUF_F_TX_IPV6
405 	 * bit[8] = RTE_MBUF_F_TX_OUTER_IPV6
406 	 * bit[9] = RTE_MBUF_F_TX_OUTER_UDP
407 	 */
408 	idx = (ol & (RTE_MBUF_F_TX_L4_MASK | RTE_MBUF_F_TX_IPV6 | RTE_MBUF_F_TX_OUTER_IPV6)) >> 52;
409 	idx |= (tunnel == RTE_MBUF_F_TX_TUNNEL_UDP) ? (1 << 9) : 0;
410 	*swp_flags = mlx5_swp_types_table[idx];
411 	/*
412 	 * Set offsets for SW parser. Since ConnectX-5, SW parser just
413 	 * complements HW parser. SW parser starts to engage only if HW parser
414 	 * can't reach a header. For the older devices, HW parser will not kick
415 	 * in if any of SWP offsets is set. Therefore, all of the L3 offsets
416 	 * should be set regardless of HW offload.
417 	 */
418 	off = loc->mbuf->outer_l2_len;
419 	if (MLX5_TXOFF_CONFIG(VLAN) && ol & RTE_MBUF_F_TX_VLAN)
420 		off += sizeof(struct rte_vlan_hdr);
421 	set = (off >> 1) << 8; /* Outer L3 offset. */
422 	off += loc->mbuf->outer_l3_len;
423 	if (tunnel == RTE_MBUF_F_TX_TUNNEL_UDP)
424 		set |= off >> 1; /* Outer L4 offset. */
425 	if (ol & (RTE_MBUF_F_TX_IPV4 | RTE_MBUF_F_TX_IPV6)) { /* Inner IP. */
426 		const uint64_t csum = ol & RTE_MBUF_F_TX_L4_MASK;
427 			off += loc->mbuf->l2_len;
428 		set |= (off >> 1) << 24; /* Inner L3 offset. */
429 		if (csum == RTE_MBUF_F_TX_TCP_CKSUM ||
430 		    csum == RTE_MBUF_F_TX_UDP_CKSUM ||
431 		    (MLX5_TXOFF_CONFIG(TSO) && ol & RTE_MBUF_F_TX_TCP_SEG)) {
432 			off += loc->mbuf->l3_len;
433 			set |= (off >> 1) << 16; /* Inner L4 offset. */
434 		}
435 	}
436 	set = rte_cpu_to_le_32(set);
437 	return set;
438 }
439 
440 /**
441  * Convert the Checksum offloads to Verbs.
442  *
443  * @param buf
444  *   Pointer to the mbuf.
445  *
446  * @return
447  *   Converted checksum flags.
448  */
449 static __rte_always_inline uint8_t
450 txq_ol_cksum_to_cs(struct rte_mbuf *buf)
451 {
452 	uint32_t idx;
453 	uint8_t is_tunnel = !!(buf->ol_flags & RTE_MBUF_F_TX_TUNNEL_MASK);
454 	const uint64_t ol_flags_mask = RTE_MBUF_F_TX_TCP_SEG | RTE_MBUF_F_TX_L4_MASK |
455 				       RTE_MBUF_F_TX_IP_CKSUM | RTE_MBUF_F_TX_OUTER_IP_CKSUM;
456 
457 	/*
458 	 * The index should have:
459 	 * bit[0] = RTE_MBUF_F_TX_TCP_SEG
460 	 * bit[2:3] = RTE_MBUF_F_TX_UDP_CKSUM, RTE_MBUF_F_TX_TCP_CKSUM
461 	 * bit[4] = RTE_MBUF_F_TX_IP_CKSUM
462 	 * bit[8] = RTE_MBUF_F_TX_OUTER_IP_CKSUM
463 	 * bit[9] = tunnel
464 	 */
465 	idx = ((buf->ol_flags & ol_flags_mask) >> 50) | (!!is_tunnel << 9);
466 	return mlx5_cksum_table[idx];
467 }
468 
469 /**
470  * Free the mbufs from the linear array of pointers.
471  *
472  * @param txq
473  *   Pointer to Tx queue structure.
474  * @param pkts
475  *   Pointer to array of packets to be free.
476  * @param pkts_n
477  *   Number of packets to be freed.
478  * @param olx
479  *   Configured Tx offloads mask. It is fully defined at
480  *   compile time and may be used for optimization.
481  */
482 static __rte_always_inline void
483 mlx5_tx_free_mbuf(struct mlx5_txq_data *__rte_restrict txq,
484 		  struct rte_mbuf **__rte_restrict pkts,
485 		  unsigned int pkts_n,
486 		  unsigned int olx __rte_unused)
487 {
488 	struct rte_mempool *pool = NULL;
489 	struct rte_mbuf **p_free = NULL;
490 	struct rte_mbuf *mbuf;
491 	unsigned int n_free = 0;
492 
493 	/*
494 	 * The implemented algorithm eliminates
495 	 * copying pointers to temporary array
496 	 * for rte_mempool_put_bulk() calls.
497 	 */
498 	MLX5_ASSERT(pkts);
499 	MLX5_ASSERT(pkts_n);
500 	/*
501 	 * Free mbufs directly to the pool in bulk
502 	 * if fast free offload is engaged
503 	 */
504 	if (!MLX5_TXOFF_CONFIG(MULTI) && txq->fast_free) {
505 		mbuf = *pkts;
506 		pool = mbuf->pool;
507 		rte_mempool_put_bulk(pool, (void *)pkts, pkts_n);
508 		return;
509 	}
510 	for (;;) {
511 		for (;;) {
512 			/*
513 			 * Decrement mbuf reference counter, detach
514 			 * indirect and external buffers if needed.
515 			 */
516 			mbuf = rte_pktmbuf_prefree_seg(*pkts);
517 			if (likely(mbuf != NULL)) {
518 				MLX5_ASSERT(mbuf == *pkts);
519 				if (likely(n_free != 0)) {
520 					if (unlikely(pool != mbuf->pool))
521 						/* From different pool. */
522 						break;
523 				} else {
524 					/* Start new scan array. */
525 					pool = mbuf->pool;
526 					p_free = pkts;
527 				}
528 				++n_free;
529 				++pkts;
530 				--pkts_n;
531 				if (unlikely(pkts_n == 0)) {
532 					mbuf = NULL;
533 					break;
534 				}
535 			} else {
536 				/*
537 				 * This happens if mbuf is still referenced.
538 				 * We can't put it back to the pool, skip.
539 				 */
540 				++pkts;
541 				--pkts_n;
542 				if (unlikely(n_free != 0))
543 					/* There is some array to free.*/
544 					break;
545 				if (unlikely(pkts_n == 0))
546 					/* Last mbuf, nothing to free. */
547 					return;
548 			}
549 		}
550 		for (;;) {
551 			/*
552 			 * This loop is implemented to avoid multiple
553 			 * inlining of rte_mempool_put_bulk().
554 			 */
555 			MLX5_ASSERT(pool);
556 			MLX5_ASSERT(p_free);
557 			MLX5_ASSERT(n_free);
558 			/*
559 			 * Free the array of pre-freed mbufs
560 			 * belonging to the same memory pool.
561 			 */
562 			rte_mempool_put_bulk(pool, (void *)p_free, n_free);
563 			if (unlikely(mbuf != NULL)) {
564 				/* There is the request to start new scan. */
565 				pool = mbuf->pool;
566 				p_free = pkts++;
567 				n_free = 1;
568 				--pkts_n;
569 				if (likely(pkts_n != 0))
570 					break;
571 				/*
572 				 * This is the last mbuf to be freed.
573 				 * Do one more loop iteration to complete.
574 				 * This is rare case of the last unique mbuf.
575 				 */
576 				mbuf = NULL;
577 				continue;
578 			}
579 			if (likely(pkts_n == 0))
580 				return;
581 			n_free = 0;
582 			break;
583 		}
584 	}
585 }
586 
587 /**
588  * No inline version to free buffers for optimal call
589  * on the tx_burst completion.
590  */
591 static __rte_noinline void
592 __mlx5_tx_free_mbuf(struct mlx5_txq_data *__rte_restrict txq,
593 		    struct rte_mbuf **__rte_restrict pkts,
594 		    unsigned int pkts_n,
595 		    unsigned int olx __rte_unused)
596 {
597 	mlx5_tx_free_mbuf(txq, pkts, pkts_n, olx);
598 }
599 
600 /**
601  * Free the mbuf from the elts ring buffer till new tail.
602  *
603  * @param txq
604  *   Pointer to Tx queue structure.
605  * @param tail
606  *   Index in elts to free up to, becomes new elts tail.
607  * @param olx
608  *   Configured Tx offloads mask. It is fully defined at
609  *   compile time and may be used for optimization.
610  */
611 static __rte_always_inline void
612 mlx5_tx_free_elts(struct mlx5_txq_data *__rte_restrict txq,
613 		  uint16_t tail,
614 		  unsigned int olx __rte_unused)
615 {
616 	uint16_t n_elts = tail - txq->elts_tail;
617 
618 	MLX5_ASSERT(n_elts);
619 	MLX5_ASSERT(n_elts <= txq->elts_s);
620 	/*
621 	 * Implement a loop to support ring buffer wraparound
622 	 * with single inlining of mlx5_tx_free_mbuf().
623 	 */
624 	do {
625 		unsigned int part;
626 
627 		part = txq->elts_s - (txq->elts_tail & txq->elts_m);
628 		part = RTE_MIN(part, n_elts);
629 		MLX5_ASSERT(part);
630 		MLX5_ASSERT(part <= txq->elts_s);
631 		mlx5_tx_free_mbuf(txq,
632 				  &txq->elts[txq->elts_tail & txq->elts_m],
633 				  part, olx);
634 		txq->elts_tail += part;
635 		n_elts -= part;
636 	} while (n_elts);
637 }
638 
639 /**
640  * Store the mbuf being sent into elts ring buffer.
641  * On Tx completion these mbufs will be freed.
642  *
643  * @param txq
644  *   Pointer to Tx queue structure.
645  * @param pkts
646  *   Pointer to array of packets to be stored.
647  * @param pkts_n
648  *   Number of packets to be stored.
649  * @param olx
650  *   Configured Tx offloads mask. It is fully defined at
651  *   compile time and may be used for optimization.
652  */
653 static __rte_always_inline void
654 mlx5_tx_copy_elts(struct mlx5_txq_data *__rte_restrict txq,
655 		  struct rte_mbuf **__rte_restrict pkts,
656 		  unsigned int pkts_n,
657 		  unsigned int olx __rte_unused)
658 {
659 	unsigned int part;
660 	struct rte_mbuf **elts = (struct rte_mbuf **)txq->elts;
661 
662 	MLX5_ASSERT(pkts);
663 	MLX5_ASSERT(pkts_n);
664 	part = txq->elts_s - (txq->elts_head & txq->elts_m);
665 	MLX5_ASSERT(part);
666 	MLX5_ASSERT(part <= txq->elts_s);
667 	/* This code is a good candidate for vectorizing with SIMD. */
668 	rte_memcpy((void *)(elts + (txq->elts_head & txq->elts_m)),
669 		   (void *)pkts,
670 		   RTE_MIN(part, pkts_n) * sizeof(struct rte_mbuf *));
671 	txq->elts_head += pkts_n;
672 	if (unlikely(part < pkts_n))
673 		/* The copy is wrapping around the elts array. */
674 		rte_memcpy((void *)elts, (void *)(pkts + part),
675 			   (pkts_n - part) * sizeof(struct rte_mbuf *));
676 }
677 
678 /**
679  * Check if the completion request flag should be set in the last WQE.
680  * Both pushed mbufs and WQEs are monitored and the completion request
681  * flag is set if any of thresholds is reached.
682  *
683  * @param txq
684  *   Pointer to TX queue structure.
685  * @param loc
686  *   Pointer to burst routine local context.
687  * @param olx
688  *   Configured Tx offloads mask. It is fully defined at
689  *   compile time and may be used for optimization.
690  */
691 static __rte_always_inline void
692 mlx5_tx_request_completion(struct mlx5_txq_data *__rte_restrict txq,
693 			   struct mlx5_txq_local *__rte_restrict loc,
694 			   unsigned int olx)
695 {
696 	uint16_t head = txq->elts_head;
697 	unsigned int part;
698 
699 	part = MLX5_TXOFF_CONFIG(INLINE) ?
700 	       0 : loc->pkts_sent - loc->pkts_copy;
701 	head += part;
702 	if ((uint16_t)(head - txq->elts_comp) >= MLX5_TX_COMP_THRESH ||
703 	     (MLX5_TXOFF_CONFIG(INLINE) &&
704 	     (uint16_t)(txq->wqe_ci - txq->wqe_comp) >= txq->wqe_thres)) {
705 		volatile struct mlx5_wqe *last = loc->wqe_last;
706 
707 		MLX5_ASSERT(last);
708 		txq->elts_comp = head;
709 		if (MLX5_TXOFF_CONFIG(INLINE))
710 			txq->wqe_comp = txq->wqe_ci;
711 		/* Request unconditional completion on last WQE. */
712 		last->cseg.flags = RTE_BE32(MLX5_COMP_ALWAYS <<
713 					    MLX5_COMP_MODE_OFFSET);
714 		/* Save elts_head in dedicated free on completion queue. */
715 #ifdef RTE_LIBRTE_MLX5_DEBUG
716 		txq->fcqs[txq->cq_pi++ & txq->cqe_m] = head |
717 			  (last->cseg.opcode >> 8) << 16;
718 #else
719 		txq->fcqs[txq->cq_pi++ & txq->cqe_m] = head;
720 #endif
721 		/* A CQE slot must always be available. */
722 		MLX5_ASSERT((txq->cq_pi - txq->cq_ci) <= txq->cqe_s);
723 	}
724 }
725 
726 /**
727  * Build the Control Segment with specified opcode:
728  * - MLX5_OPCODE_SEND
729  * - MLX5_OPCODE_ENHANCED_MPSW
730  * - MLX5_OPCODE_TSO
731  *
732  * @param txq
733  *   Pointer to TX queue structure.
734  * @param loc
735  *   Pointer to burst routine local context.
736  * @param wqe
737  *   Pointer to WQE to fill with built Control Segment.
738  * @param ds
739  *   Supposed length of WQE in segments.
740  * @param opcode
741  *   SQ WQE opcode to put into Control Segment.
742  * @param olx
743  *   Configured Tx offloads mask. It is fully defined at
744  *   compile time and may be used for optimization.
745  */
746 static __rte_always_inline void
747 mlx5_tx_cseg_init(struct mlx5_txq_data *__rte_restrict txq,
748 		  struct mlx5_txq_local *__rte_restrict loc __rte_unused,
749 		  struct mlx5_wqe *__rte_restrict wqe,
750 		  unsigned int ds,
751 		  unsigned int opcode,
752 		  unsigned int olx __rte_unused)
753 {
754 	struct mlx5_wqe_cseg *__rte_restrict cs = &wqe->cseg;
755 
756 	/* For legacy MPW replace the EMPW by TSO with modifier. */
757 	if (MLX5_TXOFF_CONFIG(MPW) && opcode == MLX5_OPCODE_ENHANCED_MPSW)
758 		opcode = MLX5_OPCODE_TSO | MLX5_OPC_MOD_MPW << 24;
759 	cs->opcode = rte_cpu_to_be_32((txq->wqe_ci << 8) | opcode);
760 	cs->sq_ds = rte_cpu_to_be_32(txq->qp_num_8s | ds);
761 	cs->flags = RTE_BE32(MLX5_COMP_ONLY_FIRST_ERR <<
762 			     MLX5_COMP_MODE_OFFSET);
763 	cs->misc = RTE_BE32(0);
764 }
765 
766 /**
767  * Build the Synchronize Queue Segment with specified completion index.
768  *
769  * @param txq
770  *   Pointer to TX queue structure.
771  * @param loc
772  *   Pointer to burst routine local context.
773  * @param wqe
774  *   Pointer to WQE to fill with built Control Segment.
775  * @param wci
776  *   Completion index in Clock Queue to wait.
777  * @param olx
778  *   Configured Tx offloads mask. It is fully defined at
779  *   compile time and may be used for optimization.
780  */
781 static __rte_always_inline void
782 mlx5_tx_wseg_init(struct mlx5_txq_data *restrict txq,
783 		  struct mlx5_txq_local *restrict loc __rte_unused,
784 		  struct mlx5_wqe *restrict wqe,
785 		  unsigned int wci,
786 		  unsigned int olx __rte_unused)
787 {
788 	struct mlx5_wqe_qseg *qs;
789 
790 	qs = RTE_PTR_ADD(wqe, MLX5_WSEG_SIZE);
791 	qs->max_index = rte_cpu_to_be_32(wci);
792 	qs->qpn_cqn = rte_cpu_to_be_32(txq->sh->txpp.clock_queue.cq_obj.cq->id);
793 	qs->reserved0 = RTE_BE32(0);
794 	qs->reserved1 = RTE_BE32(0);
795 }
796 
797 /**
798  * Build the Ethernet Segment without inlined data.
799  * Supports Software Parser, Checksums and VLAN insertion Tx offload features.
800  *
801  * @param txq
802  *   Pointer to TX queue structure.
803  * @param loc
804  *   Pointer to burst routine local context.
805  * @param wqe
806  *   Pointer to WQE to fill with built Ethernet Segment.
807  * @param olx
808  *   Configured Tx offloads mask. It is fully defined at
809  *   compile time and may be used for optimization.
810  */
811 static __rte_always_inline void
812 mlx5_tx_eseg_none(struct mlx5_txq_data *__rte_restrict txq __rte_unused,
813 		  struct mlx5_txq_local *__rte_restrict loc,
814 		  struct mlx5_wqe *__rte_restrict wqe,
815 		  unsigned int olx)
816 {
817 	struct mlx5_wqe_eseg *__rte_restrict es = &wqe->eseg;
818 	uint32_t csum;
819 
820 	/*
821 	 * Calculate and set check sum flags first, dword field
822 	 * in segment may be shared with Software Parser flags.
823 	 */
824 	csum = MLX5_TXOFF_CONFIG(CSUM) ? txq_ol_cksum_to_cs(loc->mbuf) : 0;
825 	es->flags = rte_cpu_to_le_32(csum);
826 	/*
827 	 * Calculate and set Software Parser offsets and flags.
828 	 * These flags a set for custom UDP and IP tunnel packets.
829 	 */
830 	es->swp_offs = txq_mbuf_to_swp(loc, &es->swp_flags, olx);
831 	/* Fill metadata field if needed. */
832 	es->metadata = MLX5_TXOFF_CONFIG(METADATA) ?
833 		       loc->mbuf->ol_flags & RTE_MBUF_DYNFLAG_TX_METADATA ?
834 		       rte_cpu_to_be_32(*RTE_FLOW_DYNF_METADATA(loc->mbuf)) :
835 		       0 : 0;
836 	/* Engage VLAN tag insertion feature if requested. */
837 	if (MLX5_TXOFF_CONFIG(VLAN) &&
838 	    loc->mbuf->ol_flags & RTE_MBUF_F_TX_VLAN) {
839 		/*
840 		 * We should get here only if device support
841 		 * this feature correctly.
842 		 */
843 		MLX5_ASSERT(txq->vlan_en);
844 		es->inline_hdr = rte_cpu_to_be_32(MLX5_ETH_WQE_VLAN_INSERT |
845 						  loc->mbuf->vlan_tci);
846 	} else {
847 		es->inline_hdr = RTE_BE32(0);
848 	}
849 }
850 
851 /**
852  * Build the Ethernet Segment with minimal inlined data
853  * of MLX5_ESEG_MIN_INLINE_SIZE bytes length. This is
854  * used to fill the gap in single WQEBB WQEs.
855  * Supports Software Parser, Checksums and VLAN
856  * insertion Tx offload features.
857  *
858  * @param txq
859  *   Pointer to TX queue structure.
860  * @param loc
861  *   Pointer to burst routine local context.
862  * @param wqe
863  *   Pointer to WQE to fill with built Ethernet Segment.
864  * @param vlan
865  *   Length of VLAN tag insertion if any.
866  * @param olx
867  *   Configured Tx offloads mask. It is fully defined at
868  *   compile time and may be used for optimization.
869  */
870 static __rte_always_inline void
871 mlx5_tx_eseg_dmin(struct mlx5_txq_data *__rte_restrict txq __rte_unused,
872 		  struct mlx5_txq_local *__rte_restrict loc,
873 		  struct mlx5_wqe *__rte_restrict wqe,
874 		  unsigned int vlan,
875 		  unsigned int olx)
876 {
877 	struct mlx5_wqe_eseg *__rte_restrict es = &wqe->eseg;
878 	uint32_t csum;
879 	uint8_t *psrc, *pdst;
880 
881 	/*
882 	 * Calculate and set check sum flags first, dword field
883 	 * in segment may be shared with Software Parser flags.
884 	 */
885 	csum = MLX5_TXOFF_CONFIG(CSUM) ? txq_ol_cksum_to_cs(loc->mbuf) : 0;
886 	es->flags = rte_cpu_to_le_32(csum);
887 	/*
888 	 * Calculate and set Software Parser offsets and flags.
889 	 * These flags a set for custom UDP and IP tunnel packets.
890 	 */
891 	es->swp_offs = txq_mbuf_to_swp(loc, &es->swp_flags, olx);
892 	/* Fill metadata field if needed. */
893 	es->metadata = MLX5_TXOFF_CONFIG(METADATA) ?
894 		       loc->mbuf->ol_flags & RTE_MBUF_DYNFLAG_TX_METADATA ?
895 		       rte_cpu_to_be_32(*RTE_FLOW_DYNF_METADATA(loc->mbuf)) :
896 		       0 : 0;
897 	psrc = rte_pktmbuf_mtod(loc->mbuf, uint8_t *);
898 	es->inline_hdr_sz = RTE_BE16(MLX5_ESEG_MIN_INLINE_SIZE);
899 	es->inline_data = *(unaligned_uint16_t *)psrc;
900 	psrc +=	sizeof(uint16_t);
901 	pdst = (uint8_t *)(es + 1);
902 	if (MLX5_TXOFF_CONFIG(VLAN) && vlan) {
903 		/* Implement VLAN tag insertion as part inline data. */
904 		memcpy(pdst, psrc, 2 * RTE_ETHER_ADDR_LEN - sizeof(uint16_t));
905 		pdst += 2 * RTE_ETHER_ADDR_LEN - sizeof(uint16_t);
906 		psrc +=	2 * RTE_ETHER_ADDR_LEN - sizeof(uint16_t);
907 		/* Insert VLAN ethertype + VLAN tag. */
908 		*(unaligned_uint32_t *)pdst = rte_cpu_to_be_32
909 						((RTE_ETHER_TYPE_VLAN << 16) |
910 						 loc->mbuf->vlan_tci);
911 		pdst += sizeof(struct rte_vlan_hdr);
912 		/* Copy the rest two bytes from packet data. */
913 		MLX5_ASSERT(pdst == RTE_PTR_ALIGN(pdst, sizeof(uint16_t)));
914 		*(uint16_t *)pdst = *(unaligned_uint16_t *)psrc;
915 	} else {
916 		/* Fill the gap in the title WQEBB with inline data. */
917 		rte_mov16(pdst, psrc);
918 	}
919 }
920 
921 /**
922  * Build the Ethernet Segment with entire packet data inlining. Checks the
923  * boundary of WQEBB and ring buffer wrapping, supports Software Parser,
924  * Checksums and VLAN insertion Tx offload features.
925  *
926  * @param txq
927  *   Pointer to TX queue structure.
928  * @param loc
929  *   Pointer to burst routine local context.
930  * @param wqe
931  *   Pointer to WQE to fill with built Ethernet Segment.
932  * @param vlan
933  *   Length of VLAN tag insertion if any.
934  * @param inlen
935  *   Length of data to inline (VLAN included, if any).
936  * @param tso
937  *   TSO flag, set mss field from the packet.
938  * @param olx
939  *   Configured Tx offloads mask. It is fully defined at
940  *   compile time and may be used for optimization.
941  *
942  * @return
943  *   Pointer to the next Data Segment (aligned and wrapped around).
944  */
945 static __rte_always_inline struct mlx5_wqe_dseg *
946 mlx5_tx_eseg_data(struct mlx5_txq_data *__rte_restrict txq,
947 		  struct mlx5_txq_local *__rte_restrict loc,
948 		  struct mlx5_wqe *__rte_restrict wqe,
949 		  unsigned int vlan,
950 		  unsigned int inlen,
951 		  unsigned int tso,
952 		  unsigned int olx)
953 {
954 	struct mlx5_wqe_eseg *__rte_restrict es = &wqe->eseg;
955 	uint32_t csum;
956 	uint8_t *psrc, *pdst;
957 	unsigned int part;
958 
959 	/*
960 	 * Calculate and set check sum flags first, dword field
961 	 * in segment may be shared with Software Parser flags.
962 	 */
963 	csum = MLX5_TXOFF_CONFIG(CSUM) ? txq_ol_cksum_to_cs(loc->mbuf) : 0;
964 	if (tso) {
965 		csum <<= 24;
966 		csum |= loc->mbuf->tso_segsz;
967 		es->flags = rte_cpu_to_be_32(csum);
968 	} else {
969 		es->flags = rte_cpu_to_le_32(csum);
970 	}
971 	/*
972 	 * Calculate and set Software Parser offsets and flags.
973 	 * These flags a set for custom UDP and IP tunnel packets.
974 	 */
975 	es->swp_offs = txq_mbuf_to_swp(loc, &es->swp_flags, olx);
976 	/* Fill metadata field if needed. */
977 	es->metadata = MLX5_TXOFF_CONFIG(METADATA) ?
978 		       loc->mbuf->ol_flags & RTE_MBUF_DYNFLAG_TX_METADATA ?
979 		       rte_cpu_to_be_32(*RTE_FLOW_DYNF_METADATA(loc->mbuf)) :
980 		       0 : 0;
981 	psrc = rte_pktmbuf_mtod(loc->mbuf, uint8_t *);
982 	es->inline_hdr_sz = rte_cpu_to_be_16(inlen);
983 	es->inline_data = *(unaligned_uint16_t *)psrc;
984 	psrc +=	sizeof(uint16_t);
985 	pdst = (uint8_t *)(es + 1);
986 	if (MLX5_TXOFF_CONFIG(VLAN) && vlan) {
987 		/* Implement VLAN tag insertion as part inline data. */
988 		memcpy(pdst, psrc, 2 * RTE_ETHER_ADDR_LEN - sizeof(uint16_t));
989 		pdst += 2 * RTE_ETHER_ADDR_LEN - sizeof(uint16_t);
990 		psrc +=	2 * RTE_ETHER_ADDR_LEN - sizeof(uint16_t);
991 		/* Insert VLAN ethertype + VLAN tag. */
992 		*(unaligned_uint32_t *)pdst = rte_cpu_to_be_32
993 						((RTE_ETHER_TYPE_VLAN << 16) |
994 						 loc->mbuf->vlan_tci);
995 		pdst += sizeof(struct rte_vlan_hdr);
996 		/* Copy the rest two bytes from packet data. */
997 		MLX5_ASSERT(pdst == RTE_PTR_ALIGN(pdst, sizeof(uint16_t)));
998 		*(uint16_t *)pdst = *(unaligned_uint16_t *)psrc;
999 		psrc += sizeof(uint16_t);
1000 	} else {
1001 		/* Fill the gap in the title WQEBB with inline data. */
1002 		rte_mov16(pdst, psrc);
1003 		psrc += sizeof(rte_v128u32_t);
1004 	}
1005 	pdst = (uint8_t *)(es + 2);
1006 	MLX5_ASSERT(inlen >= MLX5_ESEG_MIN_INLINE_SIZE);
1007 	MLX5_ASSERT(pdst < (uint8_t *)txq->wqes_end);
1008 	inlen -= MLX5_ESEG_MIN_INLINE_SIZE;
1009 	if (!inlen) {
1010 		MLX5_ASSERT(pdst == RTE_PTR_ALIGN(pdst, MLX5_WSEG_SIZE));
1011 		return (struct mlx5_wqe_dseg *)pdst;
1012 	}
1013 	/*
1014 	 * The WQEBB space availability is checked by caller.
1015 	 * Here we should be aware of WQE ring buffer wraparound only.
1016 	 */
1017 	part = (uint8_t *)txq->wqes_end - pdst;
1018 	part = RTE_MIN(part, inlen);
1019 	do {
1020 		rte_memcpy(pdst, psrc, part);
1021 		inlen -= part;
1022 		if (likely(!inlen)) {
1023 			/*
1024 			 * If return value is not used by the caller
1025 			 * the code below will be optimized out.
1026 			 */
1027 			pdst += part;
1028 			pdst = RTE_PTR_ALIGN(pdst, MLX5_WSEG_SIZE);
1029 			if (unlikely(pdst >= (uint8_t *)txq->wqes_end))
1030 				pdst = (uint8_t *)txq->wqes;
1031 			return (struct mlx5_wqe_dseg *)pdst;
1032 		}
1033 		pdst = (uint8_t *)txq->wqes;
1034 		psrc += part;
1035 		part = inlen;
1036 	} while (true);
1037 }
1038 
1039 /**
1040  * Copy data from chain of mbuf to the specified linear buffer.
1041  * Checksums and VLAN insertion Tx offload features. If data
1042  * from some mbuf copied completely this mbuf is freed. Local
1043  * structure is used to keep the byte stream state.
1044  *
1045  * @param pdst
1046  *   Pointer to the destination linear buffer.
1047  * @param loc
1048  *   Pointer to burst routine local context.
1049  * @param len
1050  *   Length of data to be copied.
1051  * @param must
1052  *   Length of data to be copied ignoring no inline hint.
1053  * @param olx
1054  *   Configured Tx offloads mask. It is fully defined at
1055  *   compile time and may be used for optimization.
1056  *
1057  * @return
1058  *   Number of actual copied data bytes. This is always greater than or
1059  *   equal to must parameter and might be lesser than len in no inline
1060  *   hint flag is encountered.
1061  */
1062 static __rte_always_inline unsigned int
1063 mlx5_tx_mseg_memcpy(uint8_t *pdst,
1064 		    struct mlx5_txq_local *__rte_restrict loc,
1065 		    unsigned int len,
1066 		    unsigned int must,
1067 		    unsigned int olx __rte_unused)
1068 {
1069 	struct rte_mbuf *mbuf;
1070 	unsigned int part, dlen, copy = 0;
1071 	uint8_t *psrc;
1072 
1073 	MLX5_ASSERT(len);
1074 	MLX5_ASSERT(must <= len);
1075 	do {
1076 		/* Allow zero length packets, must check first. */
1077 		dlen = rte_pktmbuf_data_len(loc->mbuf);
1078 		if (dlen <= loc->mbuf_off) {
1079 			/* Exhausted packet, just free. */
1080 			mbuf = loc->mbuf;
1081 			loc->mbuf = mbuf->next;
1082 			rte_pktmbuf_free_seg(mbuf);
1083 			loc->mbuf_off = 0;
1084 			MLX5_ASSERT(loc->mbuf_nseg > 1);
1085 			MLX5_ASSERT(loc->mbuf);
1086 			--loc->mbuf_nseg;
1087 			if (loc->mbuf->ol_flags & RTE_MBUF_F_TX_DYNF_NOINLINE) {
1088 				unsigned int diff;
1089 
1090 				if (copy >= must) {
1091 					/*
1092 					 * We already copied the minimal
1093 					 * requested amount of data.
1094 					 */
1095 					return copy;
1096 				}
1097 				diff = must - copy;
1098 				if (diff <= rte_pktmbuf_data_len(loc->mbuf)) {
1099 					/*
1100 					 * Copy only the minimal required
1101 					 * part of the data buffer.
1102 					 */
1103 					len = diff;
1104 				}
1105 			}
1106 			continue;
1107 		}
1108 		dlen -= loc->mbuf_off;
1109 		psrc = rte_pktmbuf_mtod_offset(loc->mbuf, uint8_t *,
1110 					       loc->mbuf_off);
1111 		part = RTE_MIN(len, dlen);
1112 		rte_memcpy(pdst, psrc, part);
1113 		copy += part;
1114 		loc->mbuf_off += part;
1115 		len -= part;
1116 		if (!len) {
1117 			if (loc->mbuf_off >= rte_pktmbuf_data_len(loc->mbuf)) {
1118 				loc->mbuf_off = 0;
1119 				/* Exhausted packet, just free. */
1120 				mbuf = loc->mbuf;
1121 				loc->mbuf = mbuf->next;
1122 				rte_pktmbuf_free_seg(mbuf);
1123 				loc->mbuf_off = 0;
1124 				MLX5_ASSERT(loc->mbuf_nseg >= 1);
1125 				--loc->mbuf_nseg;
1126 			}
1127 			return copy;
1128 		}
1129 		pdst += part;
1130 	} while (true);
1131 }
1132 
1133 /**
1134  * Build the Ethernet Segment with inlined data from multi-segment packet.
1135  * Checks the boundary of WQEBB and ring buffer wrapping, supports Software
1136  * Parser, Checksums and VLAN insertion Tx offload features.
1137  *
1138  * @param txq
1139  *   Pointer to TX queue structure.
1140  * @param loc
1141  *   Pointer to burst routine local context.
1142  * @param wqe
1143  *   Pointer to WQE to fill with built Ethernet Segment.
1144  * @param vlan
1145  *   Length of VLAN tag insertion if any.
1146  * @param inlen
1147  *   Length of data to inline (VLAN included, if any).
1148  * @param tso
1149  *   TSO flag, set mss field from the packet.
1150  * @param olx
1151  *   Configured Tx offloads mask. It is fully defined at
1152  *   compile time and may be used for optimization.
1153  *
1154  * @return
1155  *   Pointer to the next Data Segment (aligned and possible NOT wrapped
1156  *   around - caller should do wrapping check on its own).
1157  */
1158 static __rte_always_inline struct mlx5_wqe_dseg *
1159 mlx5_tx_eseg_mdat(struct mlx5_txq_data *__rte_restrict txq,
1160 		  struct mlx5_txq_local *__rte_restrict loc,
1161 		  struct mlx5_wqe *__rte_restrict wqe,
1162 		  unsigned int vlan,
1163 		  unsigned int inlen,
1164 		  unsigned int tso,
1165 		  unsigned int olx)
1166 {
1167 	struct mlx5_wqe_eseg *__rte_restrict es = &wqe->eseg;
1168 	uint32_t csum;
1169 	uint8_t *pdst;
1170 	unsigned int part, tlen = 0;
1171 
1172 	/*
1173 	 * Calculate and set check sum flags first, uint32_t field
1174 	 * in segment may be shared with Software Parser flags.
1175 	 */
1176 	csum = MLX5_TXOFF_CONFIG(CSUM) ? txq_ol_cksum_to_cs(loc->mbuf) : 0;
1177 	if (tso) {
1178 		csum <<= 24;
1179 		csum |= loc->mbuf->tso_segsz;
1180 		es->flags = rte_cpu_to_be_32(csum);
1181 	} else {
1182 		es->flags = rte_cpu_to_le_32(csum);
1183 	}
1184 	/*
1185 	 * Calculate and set Software Parser offsets and flags.
1186 	 * These flags a set for custom UDP and IP tunnel packets.
1187 	 */
1188 	es->swp_offs = txq_mbuf_to_swp(loc, &es->swp_flags, olx);
1189 	/* Fill metadata field if needed. */
1190 	es->metadata = MLX5_TXOFF_CONFIG(METADATA) ?
1191 		       loc->mbuf->ol_flags & RTE_MBUF_DYNFLAG_TX_METADATA ?
1192 		       rte_cpu_to_be_32(*RTE_FLOW_DYNF_METADATA(loc->mbuf)) :
1193 		       0 : 0;
1194 	MLX5_ASSERT(inlen >= MLX5_ESEG_MIN_INLINE_SIZE);
1195 	pdst = (uint8_t *)&es->inline_data;
1196 	if (MLX5_TXOFF_CONFIG(VLAN) && vlan) {
1197 		/* Implement VLAN tag insertion as part inline data. */
1198 		mlx5_tx_mseg_memcpy(pdst, loc,
1199 				    2 * RTE_ETHER_ADDR_LEN,
1200 				    2 * RTE_ETHER_ADDR_LEN, olx);
1201 		pdst += 2 * RTE_ETHER_ADDR_LEN;
1202 		*(unaligned_uint32_t *)pdst = rte_cpu_to_be_32
1203 						((RTE_ETHER_TYPE_VLAN << 16) |
1204 						 loc->mbuf->vlan_tci);
1205 		pdst += sizeof(struct rte_vlan_hdr);
1206 		tlen += 2 * RTE_ETHER_ADDR_LEN + sizeof(struct rte_vlan_hdr);
1207 	}
1208 	MLX5_ASSERT(pdst < (uint8_t *)txq->wqes_end);
1209 	/*
1210 	 * The WQEBB space availability is checked by caller.
1211 	 * Here we should be aware of WQE ring buffer wraparound only.
1212 	 */
1213 	part = (uint8_t *)txq->wqes_end - pdst;
1214 	part = RTE_MIN(part, inlen - tlen);
1215 	MLX5_ASSERT(part);
1216 	do {
1217 		unsigned int copy;
1218 
1219 		/*
1220 		 * Copying may be interrupted inside the routine
1221 		 * if run into no inline hint flag.
1222 		 */
1223 		copy = tso ? inlen : txq->inlen_mode;
1224 		copy = tlen >= copy ? 0 : (copy - tlen);
1225 		copy = mlx5_tx_mseg_memcpy(pdst, loc, part, copy, olx);
1226 		tlen += copy;
1227 		if (likely(inlen <= tlen) || copy < part) {
1228 			es->inline_hdr_sz = rte_cpu_to_be_16(tlen);
1229 			pdst += copy;
1230 			pdst = RTE_PTR_ALIGN(pdst, MLX5_WSEG_SIZE);
1231 			return (struct mlx5_wqe_dseg *)pdst;
1232 		}
1233 		pdst = (uint8_t *)txq->wqes;
1234 		part = inlen - tlen;
1235 	} while (true);
1236 }
1237 
1238 /**
1239  * Build the Data Segment of pointer type.
1240  *
1241  * @param txq
1242  *   Pointer to TX queue structure.
1243  * @param loc
1244  *   Pointer to burst routine local context.
1245  * @param dseg
1246  *   Pointer to WQE to fill with built Data Segment.
1247  * @param buf
1248  *   Data buffer to point.
1249  * @param len
1250  *   Data buffer length.
1251  * @param olx
1252  *   Configured Tx offloads mask. It is fully defined at
1253  *   compile time and may be used for optimization.
1254  */
1255 static __rte_always_inline void
1256 mlx5_tx_dseg_ptr(struct mlx5_txq_data *__rte_restrict txq,
1257 		 struct mlx5_txq_local *__rte_restrict loc,
1258 		 struct mlx5_wqe_dseg *__rte_restrict dseg,
1259 		 uint8_t *buf,
1260 		 unsigned int len,
1261 		 unsigned int olx __rte_unused)
1262 
1263 {
1264 	MLX5_ASSERT(len);
1265 	dseg->bcount = rte_cpu_to_be_32(len);
1266 	dseg->lkey = mlx5_mr_mb2mr(&txq->mr_ctrl, loc->mbuf);
1267 	dseg->pbuf = rte_cpu_to_be_64((uintptr_t)buf);
1268 }
1269 
1270 /**
1271  * Build the Data Segment of pointer type or inline if data length is less than
1272  * buffer in minimal Data Segment size.
1273  *
1274  * @param txq
1275  *   Pointer to TX queue structure.
1276  * @param loc
1277  *   Pointer to burst routine local context.
1278  * @param dseg
1279  *   Pointer to WQE to fill with built Data Segment.
1280  * @param buf
1281  *   Data buffer to point.
1282  * @param len
1283  *   Data buffer length.
1284  * @param olx
1285  *   Configured Tx offloads mask. It is fully defined at
1286  *   compile time and may be used for optimization.
1287  */
1288 static __rte_always_inline void
1289 mlx5_tx_dseg_iptr(struct mlx5_txq_data *__rte_restrict txq,
1290 		  struct mlx5_txq_local *__rte_restrict loc,
1291 		  struct mlx5_wqe_dseg *__rte_restrict dseg,
1292 		  uint8_t *buf,
1293 		  unsigned int len,
1294 		  unsigned int olx __rte_unused)
1295 
1296 {
1297 	uintptr_t dst, src;
1298 
1299 	MLX5_ASSERT(len);
1300 	if (len > MLX5_DSEG_MIN_INLINE_SIZE) {
1301 		dseg->bcount = rte_cpu_to_be_32(len);
1302 		dseg->lkey = mlx5_mr_mb2mr(&txq->mr_ctrl, loc->mbuf);
1303 		dseg->pbuf = rte_cpu_to_be_64((uintptr_t)buf);
1304 
1305 		return;
1306 	}
1307 	dseg->bcount = rte_cpu_to_be_32(len | MLX5_ETH_WQE_DATA_INLINE);
1308 	/* Unrolled implementation of generic rte_memcpy. */
1309 	dst = (uintptr_t)&dseg->inline_data[0];
1310 	src = (uintptr_t)buf;
1311 	if (len & 0x08) {
1312 #ifdef RTE_ARCH_STRICT_ALIGN
1313 		MLX5_ASSERT(dst == RTE_PTR_ALIGN(dst, sizeof(uint32_t)));
1314 		*(uint32_t *)dst = *(unaligned_uint32_t *)src;
1315 		dst += sizeof(uint32_t);
1316 		src += sizeof(uint32_t);
1317 		*(uint32_t *)dst = *(unaligned_uint32_t *)src;
1318 		dst += sizeof(uint32_t);
1319 		src += sizeof(uint32_t);
1320 #else
1321 		*(uint64_t *)dst = *(unaligned_uint64_t *)src;
1322 		dst += sizeof(uint64_t);
1323 		src += sizeof(uint64_t);
1324 #endif
1325 	}
1326 	if (len & 0x04) {
1327 		*(uint32_t *)dst = *(unaligned_uint32_t *)src;
1328 		dst += sizeof(uint32_t);
1329 		src += sizeof(uint32_t);
1330 	}
1331 	if (len & 0x02) {
1332 		*(uint16_t *)dst = *(unaligned_uint16_t *)src;
1333 		dst += sizeof(uint16_t);
1334 		src += sizeof(uint16_t);
1335 	}
1336 	if (len & 0x01)
1337 		*(uint8_t *)dst = *(uint8_t *)src;
1338 }
1339 
1340 /**
1341  * Build the Data Segment of inlined data from single
1342  * segment packet, no VLAN insertion.
1343  *
1344  * @param txq
1345  *   Pointer to TX queue structure.
1346  * @param loc
1347  *   Pointer to burst routine local context.
1348  * @param dseg
1349  *   Pointer to WQE to fill with built Data Segment.
1350  * @param buf
1351  *   Data buffer to point.
1352  * @param len
1353  *   Data buffer length.
1354  * @param olx
1355  *   Configured Tx offloads mask. It is fully defined at
1356  *   compile time and may be used for optimization.
1357  *
1358  * @return
1359  *   Pointer to the next Data Segment after inlined data.
1360  *   Ring buffer wraparound check is needed. We do not do it here because it
1361  *   may not be needed for the last packet in the eMPW session.
1362  */
1363 static __rte_always_inline struct mlx5_wqe_dseg *
1364 mlx5_tx_dseg_empw(struct mlx5_txq_data *__rte_restrict txq,
1365 		  struct mlx5_txq_local *__rte_restrict loc __rte_unused,
1366 		  struct mlx5_wqe_dseg *__rte_restrict dseg,
1367 		  uint8_t *buf,
1368 		  unsigned int len,
1369 		  unsigned int olx __rte_unused)
1370 {
1371 	unsigned int part;
1372 	uint8_t *pdst;
1373 
1374 	if (!MLX5_TXOFF_CONFIG(MPW)) {
1375 		/* Store the descriptor byte counter for eMPW sessions. */
1376 		dseg->bcount = rte_cpu_to_be_32(len | MLX5_ETH_WQE_DATA_INLINE);
1377 		pdst = &dseg->inline_data[0];
1378 	} else {
1379 		/* The entire legacy MPW session counter is stored on close. */
1380 		pdst = (uint8_t *)dseg;
1381 	}
1382 	/*
1383 	 * The WQEBB space availability is checked by caller.
1384 	 * Here we should be aware of WQE ring buffer wraparound only.
1385 	 */
1386 	part = (uint8_t *)txq->wqes_end - pdst;
1387 	part = RTE_MIN(part, len);
1388 	do {
1389 		rte_memcpy(pdst, buf, part);
1390 		len -= part;
1391 		if (likely(!len)) {
1392 			pdst += part;
1393 			if (!MLX5_TXOFF_CONFIG(MPW))
1394 				pdst = RTE_PTR_ALIGN(pdst, MLX5_WSEG_SIZE);
1395 			/* Note: no final wraparound check here. */
1396 			return (struct mlx5_wqe_dseg *)pdst;
1397 		}
1398 		pdst = (uint8_t *)txq->wqes;
1399 		buf += part;
1400 		part = len;
1401 	} while (true);
1402 }
1403 
1404 /**
1405  * Build the Data Segment of inlined data from single
1406  * segment packet with VLAN insertion.
1407  *
1408  * @param txq
1409  *   Pointer to TX queue structure.
1410  * @param loc
1411  *   Pointer to burst routine local context.
1412  * @param dseg
1413  *   Pointer to the dseg fill with built Data Segment.
1414  * @param buf
1415  *   Data buffer to point.
1416  * @param len
1417  *   Data buffer length.
1418  * @param olx
1419  *   Configured Tx offloads mask. It is fully defined at
1420  *   compile time and may be used for optimization.
1421  *
1422  * @return
1423  *   Pointer to the next Data Segment after inlined data.
1424  *   Ring buffer wraparound check is needed.
1425  */
1426 static __rte_always_inline struct mlx5_wqe_dseg *
1427 mlx5_tx_dseg_vlan(struct mlx5_txq_data *__rte_restrict txq,
1428 		  struct mlx5_txq_local *__rte_restrict loc __rte_unused,
1429 		  struct mlx5_wqe_dseg *__rte_restrict dseg,
1430 		  uint8_t *buf,
1431 		  unsigned int len,
1432 		  unsigned int olx __rte_unused)
1433 
1434 {
1435 	unsigned int part;
1436 	uint8_t *pdst;
1437 
1438 	MLX5_ASSERT(len > MLX5_ESEG_MIN_INLINE_SIZE);
1439 	if (!MLX5_TXOFF_CONFIG(MPW)) {
1440 		/* Store the descriptor byte counter for eMPW sessions. */
1441 		dseg->bcount = rte_cpu_to_be_32
1442 				((len + sizeof(struct rte_vlan_hdr)) |
1443 				 MLX5_ETH_WQE_DATA_INLINE);
1444 		pdst = &dseg->inline_data[0];
1445 	} else {
1446 		/* The entire legacy MPW session counter is stored on close. */
1447 		pdst = (uint8_t *)dseg;
1448 	}
1449 	memcpy(pdst, buf, MLX5_DSEG_MIN_INLINE_SIZE);
1450 	buf += MLX5_DSEG_MIN_INLINE_SIZE;
1451 	pdst += MLX5_DSEG_MIN_INLINE_SIZE;
1452 	len -= MLX5_DSEG_MIN_INLINE_SIZE;
1453 	/* Insert VLAN ethertype + VLAN tag. Pointer is aligned. */
1454 	MLX5_ASSERT(pdst == RTE_PTR_ALIGN(pdst, MLX5_WSEG_SIZE));
1455 	if (unlikely(pdst >= (uint8_t *)txq->wqes_end))
1456 		pdst = (uint8_t *)txq->wqes;
1457 	*(uint32_t *)pdst = rte_cpu_to_be_32((RTE_ETHER_TYPE_VLAN << 16) |
1458 					      loc->mbuf->vlan_tci);
1459 	pdst += sizeof(struct rte_vlan_hdr);
1460 	/*
1461 	 * The WQEBB space availability is checked by caller.
1462 	 * Here we should be aware of WQE ring buffer wraparound only.
1463 	 */
1464 	part = (uint8_t *)txq->wqes_end - pdst;
1465 	part = RTE_MIN(part, len);
1466 	do {
1467 		rte_memcpy(pdst, buf, part);
1468 		len -= part;
1469 		if (likely(!len)) {
1470 			pdst += part;
1471 			if (!MLX5_TXOFF_CONFIG(MPW))
1472 				pdst = RTE_PTR_ALIGN(pdst, MLX5_WSEG_SIZE);
1473 			/* Note: no final wraparound check here. */
1474 			return (struct mlx5_wqe_dseg *)pdst;
1475 		}
1476 		pdst = (uint8_t *)txq->wqes;
1477 		buf += part;
1478 		part = len;
1479 	} while (true);
1480 }
1481 
1482 /**
1483  * Build the Ethernet Segment with optionally inlined data with
1484  * VLAN insertion and following Data Segments (if any) from
1485  * multi-segment packet. Used by ordinary send and TSO.
1486  *
1487  * @param txq
1488  *   Pointer to TX queue structure.
1489  * @param loc
1490  *   Pointer to burst routine local context.
1491  * @param wqe
1492  *   Pointer to WQE to fill with built Ethernet/Data Segments.
1493  * @param vlan
1494  *   Length of VLAN header to insert, 0 means no VLAN insertion.
1495  * @param inlen
1496  *   Data length to inline. For TSO this parameter specifies exact value,
1497  *   for ordinary send routine can be aligned by caller to provide better WQE
1498  *   space saving and data buffer start address alignment.
1499  *   This length includes VLAN header being inserted.
1500  * @param tso
1501  *   Zero means ordinary send, inlined data can be extended,
1502  *   otherwise this is TSO, inlined data length is fixed.
1503  * @param olx
1504  *   Configured Tx offloads mask. It is fully defined at
1505  *   compile time and may be used for optimization.
1506  *
1507  * @return
1508  *   Actual size of built WQE in segments.
1509  */
1510 static __rte_always_inline unsigned int
1511 mlx5_tx_mseg_build(struct mlx5_txq_data *__rte_restrict txq,
1512 		   struct mlx5_txq_local *__rte_restrict loc,
1513 		   struct mlx5_wqe *__rte_restrict wqe,
1514 		   unsigned int vlan,
1515 		   unsigned int inlen,
1516 		   unsigned int tso,
1517 		   unsigned int olx __rte_unused)
1518 {
1519 	struct mlx5_wqe_dseg *__rte_restrict dseg;
1520 	unsigned int ds;
1521 
1522 	MLX5_ASSERT((rte_pktmbuf_pkt_len(loc->mbuf) + vlan) >= inlen);
1523 	loc->mbuf_nseg = NB_SEGS(loc->mbuf);
1524 	loc->mbuf_off = 0;
1525 
1526 	dseg = mlx5_tx_eseg_mdat(txq, loc, wqe, vlan, inlen, tso, olx);
1527 	if (!loc->mbuf_nseg)
1528 		goto dseg_done;
1529 	/*
1530 	 * There are still some mbuf remaining, not inlined.
1531 	 * The first mbuf may be partially inlined and we
1532 	 * must process the possible non-zero data offset.
1533 	 */
1534 	if (loc->mbuf_off) {
1535 		unsigned int dlen;
1536 		uint8_t *dptr;
1537 
1538 		/*
1539 		 * Exhausted packets must be dropped before.
1540 		 * Non-zero offset means there are some data
1541 		 * remained in the packet.
1542 		 */
1543 		MLX5_ASSERT(loc->mbuf_off < rte_pktmbuf_data_len(loc->mbuf));
1544 		MLX5_ASSERT(rte_pktmbuf_data_len(loc->mbuf));
1545 		dptr = rte_pktmbuf_mtod_offset(loc->mbuf, uint8_t *,
1546 					       loc->mbuf_off);
1547 		dlen = rte_pktmbuf_data_len(loc->mbuf) - loc->mbuf_off;
1548 		/*
1549 		 * Build the pointer/minimal Data Segment.
1550 		 * Do ring buffer wrapping check in advance.
1551 		 */
1552 		if ((uintptr_t)dseg >= (uintptr_t)txq->wqes_end)
1553 			dseg = (struct mlx5_wqe_dseg *)txq->wqes;
1554 		mlx5_tx_dseg_iptr(txq, loc, dseg, dptr, dlen, olx);
1555 		/* Store the mbuf to be freed on completion. */
1556 		MLX5_ASSERT(loc->elts_free);
1557 		txq->elts[txq->elts_head++ & txq->elts_m] = loc->mbuf;
1558 		--loc->elts_free;
1559 		++dseg;
1560 		if (--loc->mbuf_nseg == 0)
1561 			goto dseg_done;
1562 		loc->mbuf = loc->mbuf->next;
1563 		loc->mbuf_off = 0;
1564 	}
1565 	do {
1566 		if (unlikely(!rte_pktmbuf_data_len(loc->mbuf))) {
1567 			struct rte_mbuf *mbuf;
1568 
1569 			/* Zero length segment found, just skip. */
1570 			mbuf = loc->mbuf;
1571 			loc->mbuf = loc->mbuf->next;
1572 			rte_pktmbuf_free_seg(mbuf);
1573 			if (--loc->mbuf_nseg == 0)
1574 				break;
1575 		} else {
1576 			if ((uintptr_t)dseg >= (uintptr_t)txq->wqes_end)
1577 				dseg = (struct mlx5_wqe_dseg *)txq->wqes;
1578 			mlx5_tx_dseg_iptr
1579 				(txq, loc, dseg,
1580 				 rte_pktmbuf_mtod(loc->mbuf, uint8_t *),
1581 				 rte_pktmbuf_data_len(loc->mbuf), olx);
1582 			MLX5_ASSERT(loc->elts_free);
1583 			txq->elts[txq->elts_head++ & txq->elts_m] = loc->mbuf;
1584 			--loc->elts_free;
1585 			++dseg;
1586 			if (--loc->mbuf_nseg == 0)
1587 				break;
1588 			loc->mbuf = loc->mbuf->next;
1589 		}
1590 	} while (true);
1591 
1592 dseg_done:
1593 	/* Calculate actual segments used from the dseg pointer. */
1594 	if ((uintptr_t)wqe < (uintptr_t)dseg)
1595 		ds = ((uintptr_t)dseg - (uintptr_t)wqe) / MLX5_WSEG_SIZE;
1596 	else
1597 		ds = (((uintptr_t)dseg - (uintptr_t)wqe) +
1598 		      txq->wqe_s * MLX5_WQE_SIZE) / MLX5_WSEG_SIZE;
1599 	return ds;
1600 }
1601 
1602 /**
1603  * The routine checks timestamp flag in the current packet,
1604  * and push WAIT WQE into the queue if scheduling is required.
1605  *
1606  * @param txq
1607  *   Pointer to TX queue structure.
1608  * @param loc
1609  *   Pointer to burst routine local context.
1610  * @param olx
1611  *   Configured Tx offloads mask. It is fully defined at
1612  *   compile time and may be used for optimization.
1613  *
1614  * @return
1615  *   MLX5_TXCMP_CODE_EXIT - sending is done or impossible.
1616  *   MLX5_TXCMP_CODE_SINGLE - continue processing with the packet.
1617  *   MLX5_TXCMP_CODE_MULTI - the WAIT inserted, continue processing.
1618  * Local context variables partially updated.
1619  */
1620 static __rte_always_inline enum mlx5_txcmp_code
1621 mlx5_tx_schedule_send(struct mlx5_txq_data *restrict txq,
1622 		      struct mlx5_txq_local *restrict loc,
1623 		      unsigned int olx)
1624 {
1625 	if (MLX5_TXOFF_CONFIG(TXPP) &&
1626 	    loc->mbuf->ol_flags & txq->ts_mask) {
1627 		struct mlx5_wqe *wqe;
1628 		uint64_t ts;
1629 		int32_t wci;
1630 
1631 		/*
1632 		 * Estimate the required space quickly and roughly.
1633 		 * We would like to ensure the packet can be pushed
1634 		 * to the queue and we won't get the orphan WAIT WQE.
1635 		 */
1636 		if (loc->wqe_free <= MLX5_WQE_SIZE_MAX / MLX5_WQE_SIZE ||
1637 		    loc->elts_free < NB_SEGS(loc->mbuf))
1638 			return MLX5_TXCMP_CODE_EXIT;
1639 		/* Convert the timestamp into completion to wait. */
1640 		ts = *RTE_MBUF_DYNFIELD(loc->mbuf, txq->ts_offset, uint64_t *);
1641 		wci = mlx5_txpp_convert_tx_ts(txq->sh, ts);
1642 		if (unlikely(wci < 0))
1643 			return MLX5_TXCMP_CODE_SINGLE;
1644 		/* Build the WAIT WQE with specified completion. */
1645 		wqe = txq->wqes + (txq->wqe_ci & txq->wqe_m);
1646 		mlx5_tx_cseg_init(txq, loc, wqe, 2, MLX5_OPCODE_WAIT, olx);
1647 		mlx5_tx_wseg_init(txq, loc, wqe, wci, olx);
1648 		++txq->wqe_ci;
1649 		--loc->wqe_free;
1650 		return MLX5_TXCMP_CODE_MULTI;
1651 	}
1652 	return MLX5_TXCMP_CODE_SINGLE;
1653 }
1654 
1655 /**
1656  * Tx one packet function for multi-segment TSO. Supports all
1657  * types of Tx offloads, uses MLX5_OPCODE_TSO to build WQEs,
1658  * sends one packet per WQE.
1659  *
1660  * This routine is responsible for storing processed mbuf
1661  * into elts ring buffer and update elts_head.
1662  *
1663  * @param txq
1664  *   Pointer to TX queue structure.
1665  * @param loc
1666  *   Pointer to burst routine local context.
1667  * @param olx
1668  *   Configured Tx offloads mask. It is fully defined at
1669  *   compile time and may be used for optimization.
1670  *
1671  * @return
1672  *   MLX5_TXCMP_CODE_EXIT - sending is done or impossible.
1673  *   MLX5_TXCMP_CODE_ERROR - some unrecoverable error occurred.
1674  * Local context variables partially updated.
1675  */
1676 static __rte_always_inline enum mlx5_txcmp_code
1677 mlx5_tx_packet_multi_tso(struct mlx5_txq_data *__rte_restrict txq,
1678 			struct mlx5_txq_local *__rte_restrict loc,
1679 			unsigned int olx)
1680 {
1681 	struct mlx5_wqe *__rte_restrict wqe;
1682 	unsigned int ds, dlen, inlen, ntcp, vlan = 0;
1683 
1684 	if (MLX5_TXOFF_CONFIG(TXPP)) {
1685 		enum mlx5_txcmp_code wret;
1686 
1687 		/* Generate WAIT for scheduling if requested. */
1688 		wret = mlx5_tx_schedule_send(txq, loc, olx);
1689 		if (wret == MLX5_TXCMP_CODE_EXIT)
1690 			return MLX5_TXCMP_CODE_EXIT;
1691 		if (wret == MLX5_TXCMP_CODE_ERROR)
1692 			return MLX5_TXCMP_CODE_ERROR;
1693 	}
1694 	/*
1695 	 * Calculate data length to be inlined to estimate
1696 	 * the required space in WQE ring buffer.
1697 	 */
1698 	dlen = rte_pktmbuf_pkt_len(loc->mbuf);
1699 	if (MLX5_TXOFF_CONFIG(VLAN) && loc->mbuf->ol_flags & RTE_MBUF_F_TX_VLAN)
1700 		vlan = sizeof(struct rte_vlan_hdr);
1701 	inlen = loc->mbuf->l2_len + vlan +
1702 		loc->mbuf->l3_len + loc->mbuf->l4_len;
1703 	if (unlikely((!inlen || !loc->mbuf->tso_segsz)))
1704 		return MLX5_TXCMP_CODE_ERROR;
1705 	if (loc->mbuf->ol_flags & RTE_MBUF_F_TX_TUNNEL_MASK)
1706 		inlen += loc->mbuf->outer_l2_len + loc->mbuf->outer_l3_len;
1707 	/* Packet must contain all TSO headers. */
1708 	if (unlikely(inlen > MLX5_MAX_TSO_HEADER ||
1709 		     inlen <= MLX5_ESEG_MIN_INLINE_SIZE ||
1710 		     inlen > (dlen + vlan)))
1711 		return MLX5_TXCMP_CODE_ERROR;
1712 	MLX5_ASSERT(inlen >= txq->inlen_mode);
1713 	/*
1714 	 * Check whether there are enough free WQEBBs:
1715 	 * - Control Segment
1716 	 * - Ethernet Segment
1717 	 * - First Segment of inlined Ethernet data
1718 	 * - ... data continued ...
1719 	 * - Data Segments of pointer/min inline type
1720 	 */
1721 	ds = NB_SEGS(loc->mbuf) + 2 + (inlen -
1722 				       MLX5_ESEG_MIN_INLINE_SIZE +
1723 				       MLX5_WSEG_SIZE +
1724 				       MLX5_WSEG_SIZE - 1) / MLX5_WSEG_SIZE;
1725 	if (unlikely(loc->wqe_free < ((ds + 3) / 4)))
1726 		return MLX5_TXCMP_CODE_EXIT;
1727 	/* Check for maximal WQE size. */
1728 	if (unlikely((MLX5_WQE_SIZE_MAX / MLX5_WSEG_SIZE) < ((ds + 3) / 4)))
1729 		return MLX5_TXCMP_CODE_ERROR;
1730 #ifdef MLX5_PMD_SOFT_COUNTERS
1731 	/* Update sent data bytes/packets counters. */
1732 	ntcp = (dlen - (inlen - vlan) + loc->mbuf->tso_segsz - 1) /
1733 		loc->mbuf->tso_segsz;
1734 	/*
1735 	 * One will be added for mbuf itself at the end of the mlx5_tx_burst
1736 	 * from loc->pkts_sent field.
1737 	 */
1738 	--ntcp;
1739 	txq->stats.opackets += ntcp;
1740 	txq->stats.obytes += dlen + vlan + ntcp * inlen;
1741 #endif
1742 	wqe = txq->wqes + (txq->wqe_ci & txq->wqe_m);
1743 	loc->wqe_last = wqe;
1744 	mlx5_tx_cseg_init(txq, loc, wqe, 0, MLX5_OPCODE_TSO, olx);
1745 	ds = mlx5_tx_mseg_build(txq, loc, wqe, vlan, inlen, 1, olx);
1746 	wqe->cseg.sq_ds = rte_cpu_to_be_32(txq->qp_num_8s | ds);
1747 	txq->wqe_ci += (ds + 3) / 4;
1748 	loc->wqe_free -= (ds + 3) / 4;
1749 	return MLX5_TXCMP_CODE_MULTI;
1750 }
1751 
1752 /**
1753  * Tx one packet function for multi-segment SEND. Supports all types of Tx
1754  * offloads, uses MLX5_OPCODE_SEND to build WQEs, sends one packet per WQE,
1755  * without any data inlining in Ethernet Segment.
1756  *
1757  * This routine is responsible for storing processed mbuf
1758  * into elts ring buffer and update elts_head.
1759  *
1760  * @param txq
1761  *   Pointer to TX queue structure.
1762  * @param loc
1763  *   Pointer to burst routine local context.
1764  * @param olx
1765  *   Configured Tx offloads mask. It is fully defined at
1766  *   compile time and may be used for optimization.
1767  *
1768  * @return
1769  *   MLX5_TXCMP_CODE_EXIT - sending is done or impossible.
1770  *   MLX5_TXCMP_CODE_ERROR - some unrecoverable error occurred.
1771  * Local context variables partially updated.
1772  */
1773 static __rte_always_inline enum mlx5_txcmp_code
1774 mlx5_tx_packet_multi_send(struct mlx5_txq_data *__rte_restrict txq,
1775 			  struct mlx5_txq_local *__rte_restrict loc,
1776 			  unsigned int olx)
1777 {
1778 	struct mlx5_wqe_dseg *__rte_restrict dseg;
1779 	struct mlx5_wqe *__rte_restrict wqe;
1780 	unsigned int ds, nseg;
1781 
1782 	MLX5_ASSERT(NB_SEGS(loc->mbuf) > 1);
1783 	if (MLX5_TXOFF_CONFIG(TXPP)) {
1784 		enum mlx5_txcmp_code wret;
1785 
1786 		/* Generate WAIT for scheduling if requested. */
1787 		wret = mlx5_tx_schedule_send(txq, loc, olx);
1788 		if (wret == MLX5_TXCMP_CODE_EXIT)
1789 			return MLX5_TXCMP_CODE_EXIT;
1790 		if (wret == MLX5_TXCMP_CODE_ERROR)
1791 			return MLX5_TXCMP_CODE_ERROR;
1792 	}
1793 	/*
1794 	 * No inline at all, it means the CPU cycles saving is prioritized at
1795 	 * configuration, we should not copy any packet data to WQE.
1796 	 */
1797 	nseg = NB_SEGS(loc->mbuf);
1798 	ds = 2 + nseg;
1799 	if (unlikely(loc->wqe_free < ((ds + 3) / 4)))
1800 		return MLX5_TXCMP_CODE_EXIT;
1801 	/* Check for maximal WQE size. */
1802 	if (unlikely((MLX5_WQE_SIZE_MAX / MLX5_WSEG_SIZE) < ((ds + 3) / 4)))
1803 		return MLX5_TXCMP_CODE_ERROR;
1804 	/*
1805 	 * Some Tx offloads may cause an error if packet is not long enough,
1806 	 * check against assumed minimal length.
1807 	 */
1808 	if (rte_pktmbuf_pkt_len(loc->mbuf) <= MLX5_ESEG_MIN_INLINE_SIZE)
1809 		return MLX5_TXCMP_CODE_ERROR;
1810 #ifdef MLX5_PMD_SOFT_COUNTERS
1811 	/* Update sent data bytes counter. */
1812 	txq->stats.obytes += rte_pktmbuf_pkt_len(loc->mbuf);
1813 	if (MLX5_TXOFF_CONFIG(VLAN) &&
1814 	    loc->mbuf->ol_flags & RTE_MBUF_F_TX_VLAN)
1815 		txq->stats.obytes += sizeof(struct rte_vlan_hdr);
1816 #endif
1817 	/*
1818 	 * SEND WQE, one WQEBB:
1819 	 * - Control Segment, SEND opcode
1820 	 * - Ethernet Segment, optional VLAN, no inline
1821 	 * - Data Segments, pointer only type
1822 	 */
1823 	wqe = txq->wqes + (txq->wqe_ci & txq->wqe_m);
1824 	loc->wqe_last = wqe;
1825 	mlx5_tx_cseg_init(txq, loc, wqe, ds, MLX5_OPCODE_SEND, olx);
1826 	mlx5_tx_eseg_none(txq, loc, wqe, olx);
1827 	dseg = &wqe->dseg[0];
1828 	do {
1829 		if (unlikely(!rte_pktmbuf_data_len(loc->mbuf))) {
1830 			struct rte_mbuf *mbuf;
1831 
1832 			/*
1833 			 * Zero length segment found, have to correct total
1834 			 * size of WQE in segments.
1835 			 * It is supposed to be rare occasion, so in normal
1836 			 * case (no zero length segments) we avoid extra
1837 			 * writing to the Control Segment.
1838 			 */
1839 			--ds;
1840 			wqe->cseg.sq_ds -= RTE_BE32(1);
1841 			mbuf = loc->mbuf;
1842 			loc->mbuf = mbuf->next;
1843 			rte_pktmbuf_free_seg(mbuf);
1844 			if (--nseg == 0)
1845 				break;
1846 		} else {
1847 			mlx5_tx_dseg_ptr
1848 				(txq, loc, dseg,
1849 				 rte_pktmbuf_mtod(loc->mbuf, uint8_t *),
1850 				 rte_pktmbuf_data_len(loc->mbuf), olx);
1851 			txq->elts[txq->elts_head++ & txq->elts_m] = loc->mbuf;
1852 			--loc->elts_free;
1853 			if (--nseg == 0)
1854 				break;
1855 			++dseg;
1856 			if ((uintptr_t)dseg >= (uintptr_t)txq->wqes_end)
1857 				dseg = (struct mlx5_wqe_dseg *)txq->wqes;
1858 			loc->mbuf = loc->mbuf->next;
1859 		}
1860 	} while (true);
1861 	txq->wqe_ci += (ds + 3) / 4;
1862 	loc->wqe_free -= (ds + 3) / 4;
1863 	return MLX5_TXCMP_CODE_MULTI;
1864 }
1865 
1866 /**
1867  * Tx one packet function for multi-segment SEND. Supports all
1868  * types of Tx offloads, uses MLX5_OPCODE_SEND to build WQEs,
1869  * sends one packet per WQE, with data inlining in
1870  * Ethernet Segment and minimal Data Segments.
1871  *
1872  * This routine is responsible for storing processed mbuf
1873  * into elts ring buffer and update elts_head.
1874  *
1875  * @param txq
1876  *   Pointer to TX queue structure.
1877  * @param loc
1878  *   Pointer to burst routine local context.
1879  * @param olx
1880  *   Configured Tx offloads mask. It is fully defined at
1881  *   compile time and may be used for optimization.
1882  *
1883  * @return
1884  *   MLX5_TXCMP_CODE_EXIT - sending is done or impossible.
1885  *   MLX5_TXCMP_CODE_ERROR - some unrecoverable error occurred.
1886  * Local context variables partially updated.
1887  */
1888 static __rte_always_inline enum mlx5_txcmp_code
1889 mlx5_tx_packet_multi_inline(struct mlx5_txq_data *__rte_restrict txq,
1890 			    struct mlx5_txq_local *__rte_restrict loc,
1891 			    unsigned int olx)
1892 {
1893 	struct mlx5_wqe *__rte_restrict wqe;
1894 	unsigned int ds, inlen, dlen, vlan = 0;
1895 
1896 	MLX5_ASSERT(MLX5_TXOFF_CONFIG(INLINE));
1897 	MLX5_ASSERT(NB_SEGS(loc->mbuf) > 1);
1898 	if (MLX5_TXOFF_CONFIG(TXPP)) {
1899 		enum mlx5_txcmp_code wret;
1900 
1901 		/* Generate WAIT for scheduling if requested. */
1902 		wret = mlx5_tx_schedule_send(txq, loc, olx);
1903 		if (wret == MLX5_TXCMP_CODE_EXIT)
1904 			return MLX5_TXCMP_CODE_EXIT;
1905 		if (wret == MLX5_TXCMP_CODE_ERROR)
1906 			return MLX5_TXCMP_CODE_ERROR;
1907 	}
1908 	/*
1909 	 * First calculate data length to be inlined
1910 	 * to estimate the required space for WQE.
1911 	 */
1912 	dlen = rte_pktmbuf_pkt_len(loc->mbuf);
1913 	if (MLX5_TXOFF_CONFIG(VLAN) && loc->mbuf->ol_flags & RTE_MBUF_F_TX_VLAN)
1914 		vlan = sizeof(struct rte_vlan_hdr);
1915 	inlen = dlen + vlan;
1916 	/* Check against minimal length. */
1917 	if (inlen <= MLX5_ESEG_MIN_INLINE_SIZE)
1918 		return MLX5_TXCMP_CODE_ERROR;
1919 	MLX5_ASSERT(txq->inlen_send >= MLX5_ESEG_MIN_INLINE_SIZE);
1920 	if (inlen > txq->inlen_send ||
1921 	    loc->mbuf->ol_flags & RTE_MBUF_F_TX_DYNF_NOINLINE) {
1922 		struct rte_mbuf *mbuf;
1923 		unsigned int nxlen;
1924 		uintptr_t start;
1925 
1926 		mbuf = loc->mbuf;
1927 		nxlen = rte_pktmbuf_data_len(mbuf);
1928 		/*
1929 		 * Packet length exceeds the allowed inline data length,
1930 		 * check whether the minimal inlining is required.
1931 		 */
1932 		if (txq->inlen_mode) {
1933 			MLX5_ASSERT(txq->inlen_mode >=
1934 				    MLX5_ESEG_MIN_INLINE_SIZE);
1935 			MLX5_ASSERT(txq->inlen_mode <= txq->inlen_send);
1936 			inlen = txq->inlen_mode;
1937 		} else if (vlan && !txq->vlan_en) {
1938 			/*
1939 			 * VLAN insertion is requested and hardware does not
1940 			 * support the offload, will do with software inline.
1941 			 */
1942 			inlen = MLX5_ESEG_MIN_INLINE_SIZE;
1943 		} else if (mbuf->ol_flags & RTE_MBUF_F_TX_DYNF_NOINLINE ||
1944 			   nxlen > txq->inlen_send) {
1945 			return mlx5_tx_packet_multi_send(txq, loc, olx);
1946 		} else {
1947 			goto do_first;
1948 		}
1949 		/*
1950 		 * Now we know the minimal amount of data is requested
1951 		 * to inline. Check whether we should inline the buffers
1952 		 * from the chain beginning to eliminate some mbufs.
1953 		 */
1954 		if (unlikely(nxlen <= txq->inlen_send)) {
1955 			/* We can inline first mbuf at least. */
1956 			if (nxlen < inlen) {
1957 				unsigned int smlen;
1958 
1959 				/* Scan mbufs till inlen filled. */
1960 				do {
1961 					smlen = nxlen;
1962 					mbuf = NEXT(mbuf);
1963 					MLX5_ASSERT(mbuf);
1964 					nxlen = rte_pktmbuf_data_len(mbuf);
1965 					nxlen += smlen;
1966 				} while (unlikely(nxlen < inlen));
1967 				if (unlikely(nxlen > txq->inlen_send)) {
1968 					/* We cannot inline entire mbuf. */
1969 					smlen = inlen - smlen;
1970 					start = rte_pktmbuf_mtod_offset
1971 						    (mbuf, uintptr_t, smlen);
1972 					goto do_align;
1973 				}
1974 			}
1975 do_first:
1976 			do {
1977 				inlen = nxlen;
1978 				mbuf = NEXT(mbuf);
1979 				/* There should be not end of packet. */
1980 				MLX5_ASSERT(mbuf);
1981 				nxlen = inlen + rte_pktmbuf_data_len(mbuf);
1982 			} while (unlikely(nxlen < txq->inlen_send));
1983 		}
1984 		start = rte_pktmbuf_mtod(mbuf, uintptr_t);
1985 		/*
1986 		 * Check whether we can do inline to align start
1987 		 * address of data buffer to cacheline.
1988 		 */
1989 do_align:
1990 		start = (~start + 1) & (RTE_CACHE_LINE_SIZE - 1);
1991 		if (unlikely(start)) {
1992 			start += inlen;
1993 			if (start <= txq->inlen_send)
1994 				inlen = start;
1995 		}
1996 	}
1997 	/*
1998 	 * Check whether there are enough free WQEBBs:
1999 	 * - Control Segment
2000 	 * - Ethernet Segment
2001 	 * - First Segment of inlined Ethernet data
2002 	 * - ... data continued ...
2003 	 * - Data Segments of pointer/min inline type
2004 	 *
2005 	 * Estimate the number of Data Segments conservatively,
2006 	 * supposing no any mbufs is being freed during inlining.
2007 	 */
2008 	MLX5_ASSERT(inlen <= txq->inlen_send);
2009 	ds = NB_SEGS(loc->mbuf) + 2 + (inlen -
2010 				       MLX5_ESEG_MIN_INLINE_SIZE +
2011 				       MLX5_WSEG_SIZE +
2012 				       MLX5_WSEG_SIZE - 1) / MLX5_WSEG_SIZE;
2013 	if (unlikely(loc->wqe_free < ((ds + 3) / 4)))
2014 		return MLX5_TXCMP_CODE_EXIT;
2015 	/* Check for maximal WQE size. */
2016 	if (unlikely((MLX5_WQE_SIZE_MAX / MLX5_WSEG_SIZE) < ((ds + 3) / 4)))
2017 		return MLX5_TXCMP_CODE_ERROR;
2018 #ifdef MLX5_PMD_SOFT_COUNTERS
2019 	/* Update sent data bytes/packets counters. */
2020 	txq->stats.obytes += dlen + vlan;
2021 #endif
2022 	wqe = txq->wqes + (txq->wqe_ci & txq->wqe_m);
2023 	loc->wqe_last = wqe;
2024 	mlx5_tx_cseg_init(txq, loc, wqe, 0, MLX5_OPCODE_SEND, olx);
2025 	ds = mlx5_tx_mseg_build(txq, loc, wqe, vlan, inlen, 0, olx);
2026 	wqe->cseg.sq_ds = rte_cpu_to_be_32(txq->qp_num_8s | ds);
2027 	txq->wqe_ci += (ds + 3) / 4;
2028 	loc->wqe_free -= (ds + 3) / 4;
2029 	return MLX5_TXCMP_CODE_MULTI;
2030 }
2031 
2032 /**
2033  * Tx burst function for multi-segment packets. Supports all
2034  * types of Tx offloads, uses MLX5_OPCODE_SEND/TSO to build WQEs,
2035  * sends one packet per WQE. Function stops sending if it
2036  * encounters the single-segment packet.
2037  *
2038  * This routine is responsible for storing processed mbuf
2039  * into elts ring buffer and update elts_head.
2040  *
2041  * @param txq
2042  *   Pointer to TX queue structure.
2043  * @param[in] pkts
2044  *   Packets to transmit.
2045  * @param pkts_n
2046  *   Number of packets in array.
2047  * @param loc
2048  *   Pointer to burst routine local context.
2049  * @param olx
2050  *   Configured Tx offloads mask. It is fully defined at
2051  *   compile time and may be used for optimization.
2052  *
2053  * @return
2054  *   MLX5_TXCMP_CODE_EXIT - sending is done or impossible.
2055  *   MLX5_TXCMP_CODE_ERROR - some unrecoverable error occurred.
2056  *   MLX5_TXCMP_CODE_SINGLE - single-segment packet encountered.
2057  *   MLX5_TXCMP_CODE_TSO - TSO single-segment packet encountered.
2058  * Local context variables updated.
2059  */
2060 static __rte_always_inline enum mlx5_txcmp_code
2061 mlx5_tx_burst_mseg(struct mlx5_txq_data *__rte_restrict txq,
2062 		   struct rte_mbuf **__rte_restrict pkts,
2063 		   unsigned int pkts_n,
2064 		   struct mlx5_txq_local *__rte_restrict loc,
2065 		   unsigned int olx)
2066 {
2067 	MLX5_ASSERT(loc->elts_free && loc->wqe_free);
2068 	MLX5_ASSERT(pkts_n > loc->pkts_sent);
2069 	pkts += loc->pkts_sent + 1;
2070 	pkts_n -= loc->pkts_sent;
2071 	for (;;) {
2072 		enum mlx5_txcmp_code ret;
2073 
2074 		MLX5_ASSERT(NB_SEGS(loc->mbuf) > 1);
2075 		/*
2076 		 * Estimate the number of free elts quickly but conservatively.
2077 		 * Some segment may be fully inlined and freed,
2078 		 * ignore this here - precise estimation is costly.
2079 		 */
2080 		if (loc->elts_free < NB_SEGS(loc->mbuf))
2081 			return MLX5_TXCMP_CODE_EXIT;
2082 		if (MLX5_TXOFF_CONFIG(TSO) &&
2083 		    unlikely(loc->mbuf->ol_flags & RTE_MBUF_F_TX_TCP_SEG)) {
2084 			/* Proceed with multi-segment TSO. */
2085 			ret = mlx5_tx_packet_multi_tso(txq, loc, olx);
2086 		} else if (MLX5_TXOFF_CONFIG(INLINE)) {
2087 			/* Proceed with multi-segment SEND with inlining. */
2088 			ret = mlx5_tx_packet_multi_inline(txq, loc, olx);
2089 		} else {
2090 			/* Proceed with multi-segment SEND w/o inlining. */
2091 			ret = mlx5_tx_packet_multi_send(txq, loc, olx);
2092 		}
2093 		if (ret == MLX5_TXCMP_CODE_EXIT)
2094 			return MLX5_TXCMP_CODE_EXIT;
2095 		if (ret == MLX5_TXCMP_CODE_ERROR)
2096 			return MLX5_TXCMP_CODE_ERROR;
2097 		/* WQE is built, go to the next packet. */
2098 		++loc->pkts_sent;
2099 		--pkts_n;
2100 		if (unlikely(!pkts_n || !loc->elts_free || !loc->wqe_free))
2101 			return MLX5_TXCMP_CODE_EXIT;
2102 		loc->mbuf = *pkts++;
2103 		if (pkts_n > 1)
2104 			rte_prefetch0(*pkts);
2105 		if (likely(NB_SEGS(loc->mbuf) > 1))
2106 			continue;
2107 		/* Here ends the series of multi-segment packets. */
2108 		if (MLX5_TXOFF_CONFIG(TSO) &&
2109 		    unlikely(loc->mbuf->ol_flags & RTE_MBUF_F_TX_TCP_SEG))
2110 			return MLX5_TXCMP_CODE_TSO;
2111 		return MLX5_TXCMP_CODE_SINGLE;
2112 	}
2113 	MLX5_ASSERT(false);
2114 }
2115 
2116 /**
2117  * Tx burst function for single-segment packets with TSO.
2118  * Supports all types of Tx offloads, except multi-packets.
2119  * Uses MLX5_OPCODE_TSO to build WQEs, sends one packet per WQE.
2120  * Function stops sending if it encounters the multi-segment
2121  * packet or packet without TSO requested.
2122  *
2123  * The routine is responsible for storing processed mbuf into elts ring buffer
2124  * and update elts_head if inline offloads is requested due to possible early
2125  * freeing of the inlined mbufs (can not store pkts array in elts as a batch).
2126  *
2127  * @param txq
2128  *   Pointer to TX queue structure.
2129  * @param[in] pkts
2130  *   Packets to transmit.
2131  * @param pkts_n
2132  *   Number of packets in array.
2133  * @param loc
2134  *   Pointer to burst routine local context.
2135  * @param olx
2136  *   Configured Tx offloads mask. It is fully defined at
2137  *   compile time and may be used for optimization.
2138  *
2139  * @return
2140  *   MLX5_TXCMP_CODE_EXIT - sending is done or impossible.
2141  *   MLX5_TXCMP_CODE_ERROR - some unrecoverable error occurred.
2142  *   MLX5_TXCMP_CODE_SINGLE - single-segment packet encountered.
2143  *   MLX5_TXCMP_CODE_MULTI - multi-segment packet encountered.
2144  * Local context variables updated.
2145  */
2146 static __rte_always_inline enum mlx5_txcmp_code
2147 mlx5_tx_burst_tso(struct mlx5_txq_data *__rte_restrict txq,
2148 		  struct rte_mbuf **__rte_restrict pkts,
2149 		  unsigned int pkts_n,
2150 		  struct mlx5_txq_local *__rte_restrict loc,
2151 		  unsigned int olx)
2152 {
2153 	MLX5_ASSERT(loc->elts_free && loc->wqe_free);
2154 	MLX5_ASSERT(pkts_n > loc->pkts_sent);
2155 	pkts += loc->pkts_sent + 1;
2156 	pkts_n -= loc->pkts_sent;
2157 	for (;;) {
2158 		struct mlx5_wqe_dseg *__rte_restrict dseg;
2159 		struct mlx5_wqe *__rte_restrict wqe;
2160 		unsigned int ds, dlen, hlen, ntcp, vlan = 0;
2161 		uint8_t *dptr;
2162 
2163 		MLX5_ASSERT(NB_SEGS(loc->mbuf) == 1);
2164 		if (MLX5_TXOFF_CONFIG(TXPP)) {
2165 			enum mlx5_txcmp_code wret;
2166 
2167 			/* Generate WAIT for scheduling if requested. */
2168 			wret = mlx5_tx_schedule_send(txq, loc, olx);
2169 			if (wret == MLX5_TXCMP_CODE_EXIT)
2170 				return MLX5_TXCMP_CODE_EXIT;
2171 			if (wret == MLX5_TXCMP_CODE_ERROR)
2172 				return MLX5_TXCMP_CODE_ERROR;
2173 		}
2174 		dlen = rte_pktmbuf_data_len(loc->mbuf);
2175 		if (MLX5_TXOFF_CONFIG(VLAN) &&
2176 		    loc->mbuf->ol_flags & RTE_MBUF_F_TX_VLAN) {
2177 			vlan = sizeof(struct rte_vlan_hdr);
2178 		}
2179 		/*
2180 		 * First calculate the WQE size to check
2181 		 * whether we have enough space in ring buffer.
2182 		 */
2183 		hlen = loc->mbuf->l2_len + vlan +
2184 		       loc->mbuf->l3_len + loc->mbuf->l4_len;
2185 		if (unlikely((!hlen || !loc->mbuf->tso_segsz)))
2186 			return MLX5_TXCMP_CODE_ERROR;
2187 		if (loc->mbuf->ol_flags & RTE_MBUF_F_TX_TUNNEL_MASK)
2188 			hlen += loc->mbuf->outer_l2_len +
2189 				loc->mbuf->outer_l3_len;
2190 		/* Segment must contain all TSO headers. */
2191 		if (unlikely(hlen > MLX5_MAX_TSO_HEADER ||
2192 			     hlen <= MLX5_ESEG_MIN_INLINE_SIZE ||
2193 			     hlen > (dlen + vlan)))
2194 			return MLX5_TXCMP_CODE_ERROR;
2195 		/*
2196 		 * Check whether there are enough free WQEBBs:
2197 		 * - Control Segment
2198 		 * - Ethernet Segment
2199 		 * - First Segment of inlined Ethernet data
2200 		 * - ... data continued ...
2201 		 * - Finishing Data Segment of pointer type
2202 		 */
2203 		ds = 4 + (hlen - MLX5_ESEG_MIN_INLINE_SIZE +
2204 			  MLX5_WSEG_SIZE - 1) / MLX5_WSEG_SIZE;
2205 		if (loc->wqe_free < ((ds + 3) / 4))
2206 			return MLX5_TXCMP_CODE_EXIT;
2207 #ifdef MLX5_PMD_SOFT_COUNTERS
2208 		/* Update sent data bytes/packets counters. */
2209 		ntcp = (dlen + vlan - hlen +
2210 			loc->mbuf->tso_segsz - 1) /
2211 			loc->mbuf->tso_segsz;
2212 		/*
2213 		 * One will be added for mbuf itself at the end
2214 		 * of the mlx5_tx_burst from loc->pkts_sent field.
2215 		 */
2216 		--ntcp;
2217 		txq->stats.opackets += ntcp;
2218 		txq->stats.obytes += dlen + vlan + ntcp * hlen;
2219 #endif
2220 		/*
2221 		 * Build the TSO WQE:
2222 		 * - Control Segment
2223 		 * - Ethernet Segment with hlen bytes inlined
2224 		 * - Data Segment of pointer type
2225 		 */
2226 		wqe = txq->wqes + (txq->wqe_ci & txq->wqe_m);
2227 		loc->wqe_last = wqe;
2228 		mlx5_tx_cseg_init(txq, loc, wqe, ds,
2229 				  MLX5_OPCODE_TSO, olx);
2230 		dseg = mlx5_tx_eseg_data(txq, loc, wqe, vlan, hlen, 1, olx);
2231 		dptr = rte_pktmbuf_mtod(loc->mbuf, uint8_t *) + hlen - vlan;
2232 		dlen -= hlen - vlan;
2233 		mlx5_tx_dseg_ptr(txq, loc, dseg, dptr, dlen, olx);
2234 		/*
2235 		 * WQE is built, update the loop parameters
2236 		 * and go to the next packet.
2237 		 */
2238 		txq->wqe_ci += (ds + 3) / 4;
2239 		loc->wqe_free -= (ds + 3) / 4;
2240 		if (MLX5_TXOFF_CONFIG(INLINE))
2241 			txq->elts[txq->elts_head++ & txq->elts_m] = loc->mbuf;
2242 		--loc->elts_free;
2243 		++loc->pkts_sent;
2244 		--pkts_n;
2245 		if (unlikely(!pkts_n || !loc->elts_free || !loc->wqe_free))
2246 			return MLX5_TXCMP_CODE_EXIT;
2247 		loc->mbuf = *pkts++;
2248 		if (pkts_n > 1)
2249 			rte_prefetch0(*pkts);
2250 		if (MLX5_TXOFF_CONFIG(MULTI) &&
2251 		    unlikely(NB_SEGS(loc->mbuf) > 1))
2252 			return MLX5_TXCMP_CODE_MULTI;
2253 		if (likely(!(loc->mbuf->ol_flags & RTE_MBUF_F_TX_TCP_SEG)))
2254 			return MLX5_TXCMP_CODE_SINGLE;
2255 		/* Continue with the next TSO packet. */
2256 	}
2257 	MLX5_ASSERT(false);
2258 }
2259 
2260 /**
2261  * Analyze the packet and select the best method to send.
2262  *
2263  * @param txq
2264  *   Pointer to TX queue structure.
2265  * @param loc
2266  *   Pointer to burst routine local context.
2267  * @param olx
2268  *   Configured Tx offloads mask. It is fully defined at
2269  *   compile time and may be used for optimization.
2270  * @param newp
2271  *   The predefined flag whether do complete check for
2272  *   multi-segment packets and TSO.
2273  *
2274  * @return
2275  *  MLX5_TXCMP_CODE_MULTI - multi-segment packet encountered.
2276  *  MLX5_TXCMP_CODE_TSO - TSO required, use TSO/LSO.
2277  *  MLX5_TXCMP_CODE_SINGLE - single-segment packet, use SEND.
2278  *  MLX5_TXCMP_CODE_EMPW - single-segment packet, use MPW.
2279  */
2280 static __rte_always_inline enum mlx5_txcmp_code
2281 mlx5_tx_able_to_empw(struct mlx5_txq_data *__rte_restrict txq,
2282 		     struct mlx5_txq_local *__rte_restrict loc,
2283 		     unsigned int olx,
2284 		     bool newp)
2285 {
2286 	/* Check for multi-segment packet. */
2287 	if (newp &&
2288 	    MLX5_TXOFF_CONFIG(MULTI) &&
2289 	    unlikely(NB_SEGS(loc->mbuf) > 1))
2290 		return MLX5_TXCMP_CODE_MULTI;
2291 	/* Check for TSO packet. */
2292 	if (newp &&
2293 	    MLX5_TXOFF_CONFIG(TSO) &&
2294 	    unlikely(loc->mbuf->ol_flags & RTE_MBUF_F_TX_TCP_SEG))
2295 		return MLX5_TXCMP_CODE_TSO;
2296 	/* Check if eMPW is enabled at all. */
2297 	if (!MLX5_TXOFF_CONFIG(EMPW))
2298 		return MLX5_TXCMP_CODE_SINGLE;
2299 	/* Check if eMPW can be engaged. */
2300 	if (MLX5_TXOFF_CONFIG(VLAN) &&
2301 	    unlikely(loc->mbuf->ol_flags & RTE_MBUF_F_TX_VLAN) &&
2302 		(!MLX5_TXOFF_CONFIG(INLINE) ||
2303 		 unlikely((rte_pktmbuf_data_len(loc->mbuf) +
2304 			   sizeof(struct rte_vlan_hdr)) > txq->inlen_empw))) {
2305 		/*
2306 		 * eMPW does not support VLAN insertion offload, we have to
2307 		 * inline the entire packet but packet is too long for inlining.
2308 		 */
2309 		return MLX5_TXCMP_CODE_SINGLE;
2310 	}
2311 	return MLX5_TXCMP_CODE_EMPW;
2312 }
2313 
2314 /**
2315  * Check the next packet attributes to match with the eMPW batch ones.
2316  * In addition, for legacy MPW the packet length is checked either.
2317  *
2318  * @param txq
2319  *   Pointer to TX queue structure.
2320  * @param es
2321  *   Pointer to Ethernet Segment of eMPW batch.
2322  * @param loc
2323  *   Pointer to burst routine local context.
2324  * @param dlen
2325  *   Length of previous packet in MPW descriptor.
2326  * @param olx
2327  *   Configured Tx offloads mask. It is fully defined at
2328  *   compile time and may be used for optimization.
2329  *
2330  * @return
2331  *  true - packet match with eMPW batch attributes.
2332  *  false - no match, eMPW should be restarted.
2333  */
2334 static __rte_always_inline bool
2335 mlx5_tx_match_empw(struct mlx5_txq_data *__rte_restrict txq,
2336 		   struct mlx5_wqe_eseg *__rte_restrict es,
2337 		   struct mlx5_txq_local *__rte_restrict loc,
2338 		   uint32_t dlen,
2339 		   unsigned int olx)
2340 {
2341 	uint8_t swp_flags = 0;
2342 
2343 	/* Compare the checksum flags, if any. */
2344 	if (MLX5_TXOFF_CONFIG(CSUM) &&
2345 	    txq_ol_cksum_to_cs(loc->mbuf) != es->cs_flags)
2346 		return false;
2347 	/* Compare the Software Parser offsets and flags. */
2348 	if (MLX5_TXOFF_CONFIG(SWP) &&
2349 	    (es->swp_offs != txq_mbuf_to_swp(loc, &swp_flags, olx) ||
2350 	     es->swp_flags != swp_flags))
2351 		return false;
2352 	/* Fill metadata field if needed. */
2353 	if (MLX5_TXOFF_CONFIG(METADATA) &&
2354 		es->metadata != (loc->mbuf->ol_flags & RTE_MBUF_DYNFLAG_TX_METADATA ?
2355 				 rte_cpu_to_be_32(*RTE_FLOW_DYNF_METADATA(loc->mbuf)) : 0))
2356 		return false;
2357 	/* Legacy MPW can send packets with the same length only. */
2358 	if (MLX5_TXOFF_CONFIG(MPW) &&
2359 	    dlen != rte_pktmbuf_data_len(loc->mbuf))
2360 		return false;
2361 	/* There must be no VLAN packets in eMPW loop. */
2362 	if (MLX5_TXOFF_CONFIG(VLAN))
2363 		MLX5_ASSERT(!(loc->mbuf->ol_flags & RTE_MBUF_F_TX_VLAN));
2364 	/* Check if the scheduling is requested. */
2365 	if (MLX5_TXOFF_CONFIG(TXPP) &&
2366 	    loc->mbuf->ol_flags & txq->ts_mask)
2367 		return false;
2368 	return true;
2369 }
2370 
2371 /**
2372  * Update send loop variables and WQE for eMPW loop without data inlining.
2373  * Number of Data Segments is equal to the number of sent packets.
2374  *
2375  * @param txq
2376  *   Pointer to TX queue structure.
2377  * @param loc
2378  *   Pointer to burst routine local context.
2379  * @param ds
2380  *   Number of packets/Data Segments/Packets.
2381  * @param slen
2382  *   Accumulated statistics, bytes sent.
2383  * @param olx
2384  *   Configured Tx offloads mask. It is fully defined at
2385  *   compile time and may be used for optimization.
2386  *
2387  * @return
2388  *  true - packet match with eMPW batch attributes.
2389  *  false - no match, eMPW should be restarted.
2390  */
2391 static __rte_always_inline void
2392 mlx5_tx_sdone_empw(struct mlx5_txq_data *__rte_restrict txq,
2393 		   struct mlx5_txq_local *__rte_restrict loc,
2394 		   unsigned int ds,
2395 		   unsigned int slen,
2396 		   unsigned int olx __rte_unused)
2397 {
2398 	MLX5_ASSERT(!MLX5_TXOFF_CONFIG(INLINE));
2399 #ifdef MLX5_PMD_SOFT_COUNTERS
2400 	/* Update sent data bytes counter. */
2401 	 txq->stats.obytes += slen;
2402 #else
2403 	(void)slen;
2404 #endif
2405 	loc->elts_free -= ds;
2406 	loc->pkts_sent += ds;
2407 	ds += 2;
2408 	loc->wqe_last->cseg.sq_ds = rte_cpu_to_be_32(txq->qp_num_8s | ds);
2409 	txq->wqe_ci += (ds + 3) / 4;
2410 	loc->wqe_free -= (ds + 3) / 4;
2411 }
2412 
2413 /**
2414  * Update send loop variables and WQE for eMPW loop with data inlining.
2415  * Gets the size of pushed descriptors and data to the WQE.
2416  *
2417  * @param txq
2418  *   Pointer to TX queue structure.
2419  * @param loc
2420  *   Pointer to burst routine local context.
2421  * @param len
2422  *   Total size of descriptor/data in bytes.
2423  * @param slen
2424  *   Accumulated statistics, data bytes sent.
2425  * @param wqem
2426  *   The base WQE for the eMPW/MPW descriptor.
2427  * @param olx
2428  *   Configured Tx offloads mask. It is fully defined at
2429  *   compile time and may be used for optimization.
2430  *
2431  * @return
2432  *  true - packet match with eMPW batch attributes.
2433  *  false - no match, eMPW should be restarted.
2434  */
2435 static __rte_always_inline void
2436 mlx5_tx_idone_empw(struct mlx5_txq_data *__rte_restrict txq,
2437 		   struct mlx5_txq_local *__rte_restrict loc,
2438 		   unsigned int len,
2439 		   unsigned int slen,
2440 		   struct mlx5_wqe *__rte_restrict wqem,
2441 		   unsigned int olx __rte_unused)
2442 {
2443 	struct mlx5_wqe_dseg *dseg = &wqem->dseg[0];
2444 
2445 	MLX5_ASSERT(MLX5_TXOFF_CONFIG(INLINE));
2446 #ifdef MLX5_PMD_SOFT_COUNTERS
2447 	/* Update sent data bytes counter. */
2448 	 txq->stats.obytes += slen;
2449 #else
2450 	(void)slen;
2451 #endif
2452 	if (MLX5_TXOFF_CONFIG(MPW) && dseg->bcount == RTE_BE32(0)) {
2453 		/*
2454 		 * If the legacy MPW session contains the inline packets
2455 		 * we should set the only inline data segment length
2456 		 * and align the total length to the segment size.
2457 		 */
2458 		MLX5_ASSERT(len > sizeof(dseg->bcount));
2459 		dseg->bcount = rte_cpu_to_be_32((len - sizeof(dseg->bcount)) |
2460 						MLX5_ETH_WQE_DATA_INLINE);
2461 		len = (len + MLX5_WSEG_SIZE - 1) / MLX5_WSEG_SIZE + 2;
2462 	} else {
2463 		/*
2464 		 * The session is not legacy MPW or contains the
2465 		 * data buffer pointer segments.
2466 		 */
2467 		MLX5_ASSERT((len % MLX5_WSEG_SIZE) == 0);
2468 		len = len / MLX5_WSEG_SIZE + 2;
2469 	}
2470 	wqem->cseg.sq_ds = rte_cpu_to_be_32(txq->qp_num_8s | len);
2471 	txq->wqe_ci += (len + 3) / 4;
2472 	loc->wqe_free -= (len + 3) / 4;
2473 	loc->wqe_last = wqem;
2474 }
2475 
2476 /**
2477  * The set of Tx burst functions for single-segment packets without TSO
2478  * and with Multi-Packet Writing feature support.
2479  * Supports all types of Tx offloads, except multi-packets and TSO.
2480  *
2481  * Uses MLX5_OPCODE_EMPW to build WQEs if possible and sends as many packet
2482  * per WQE as it can. If eMPW is not configured or packet can not be sent with
2483  * eMPW (VLAN insertion) the ordinary SEND opcode is used and only one packet
2484  * placed in WQE.
2485  *
2486  * Functions stop sending if it encounters the multi-segment packet or packet
2487  * with TSO requested.
2488  *
2489  * The routines are responsible for storing processed mbuf into elts ring buffer
2490  * and update elts_head if inlining offload is requested. Otherwise the copying
2491  * mbufs to elts can be postponed and completed at the end of burst routine.
2492  *
2493  * @param txq
2494  *   Pointer to TX queue structure.
2495  * @param[in] pkts
2496  *   Packets to transmit.
2497  * @param pkts_n
2498  *   Number of packets in array.
2499  * @param loc
2500  *   Pointer to burst routine local context.
2501  * @param olx
2502  *   Configured Tx offloads mask. It is fully defined at
2503  *   compile time and may be used for optimization.
2504  *
2505  * @return
2506  *   MLX5_TXCMP_CODE_EXIT - sending is done or impossible.
2507  *   MLX5_TXCMP_CODE_ERROR - some unrecoverable error occurred.
2508  *   MLX5_TXCMP_CODE_MULTI - multi-segment packet encountered.
2509  *   MLX5_TXCMP_CODE_TSO - TSO packet encountered.
2510  *   MLX5_TXCMP_CODE_SINGLE - used inside functions set.
2511  *   MLX5_TXCMP_CODE_EMPW - used inside functions set.
2512  *
2513  * Local context variables updated.
2514  *
2515  *
2516  * The routine sends packets with MLX5_OPCODE_EMPW
2517  * without inlining, this is dedicated optimized branch.
2518  * No VLAN insertion is supported.
2519  */
2520 static __rte_always_inline enum mlx5_txcmp_code
2521 mlx5_tx_burst_empw_simple(struct mlx5_txq_data *__rte_restrict txq,
2522 			  struct rte_mbuf **__rte_restrict pkts,
2523 			  unsigned int pkts_n,
2524 			  struct mlx5_txq_local *__rte_restrict loc,
2525 			  unsigned int olx)
2526 {
2527 	/*
2528 	 * Subroutine is the part of mlx5_tx_burst_single() and sends
2529 	 * single-segment packet with eMPW opcode without data inlining.
2530 	 */
2531 	MLX5_ASSERT(!MLX5_TXOFF_CONFIG(INLINE));
2532 	MLX5_ASSERT(MLX5_TXOFF_CONFIG(EMPW));
2533 	MLX5_ASSERT(loc->elts_free && loc->wqe_free);
2534 	MLX5_ASSERT(pkts_n > loc->pkts_sent);
2535 	pkts += loc->pkts_sent + 1;
2536 	pkts_n -= loc->pkts_sent;
2537 	for (;;) {
2538 		struct mlx5_wqe_dseg *__rte_restrict dseg;
2539 		struct mlx5_wqe_eseg *__rte_restrict eseg;
2540 		enum mlx5_txcmp_code ret;
2541 		unsigned int part, loop;
2542 		unsigned int slen = 0;
2543 
2544 next_empw:
2545 		MLX5_ASSERT(NB_SEGS(loc->mbuf) == 1);
2546 		if (MLX5_TXOFF_CONFIG(TXPP)) {
2547 			enum mlx5_txcmp_code wret;
2548 
2549 			/* Generate WAIT for scheduling if requested. */
2550 			wret = mlx5_tx_schedule_send(txq, loc, olx);
2551 			if (wret == MLX5_TXCMP_CODE_EXIT)
2552 				return MLX5_TXCMP_CODE_EXIT;
2553 			if (wret == MLX5_TXCMP_CODE_ERROR)
2554 				return MLX5_TXCMP_CODE_ERROR;
2555 		}
2556 		part = RTE_MIN(pkts_n, MLX5_TXOFF_CONFIG(MPW) ?
2557 				       MLX5_MPW_MAX_PACKETS :
2558 				       MLX5_EMPW_MAX_PACKETS);
2559 		if (unlikely(loc->elts_free < part)) {
2560 			/* We have no enough elts to save all mbufs. */
2561 			if (unlikely(loc->elts_free < MLX5_EMPW_MIN_PACKETS))
2562 				return MLX5_TXCMP_CODE_EXIT;
2563 			/* But we still able to send at least minimal eMPW. */
2564 			part = loc->elts_free;
2565 		}
2566 		/* Check whether we have enough WQEs */
2567 		if (unlikely(loc->wqe_free < ((2 + part + 3) / 4))) {
2568 			if (unlikely(loc->wqe_free <
2569 				((2 + MLX5_EMPW_MIN_PACKETS + 3) / 4)))
2570 				return MLX5_TXCMP_CODE_EXIT;
2571 			part = (loc->wqe_free * 4) - 2;
2572 		}
2573 		if (likely(part > 1))
2574 			rte_prefetch0(*pkts);
2575 		loc->wqe_last = txq->wqes + (txq->wqe_ci & txq->wqe_m);
2576 		/*
2577 		 * Build eMPW title WQEBB:
2578 		 * - Control Segment, eMPW opcode
2579 		 * - Ethernet Segment, no inline
2580 		 */
2581 		mlx5_tx_cseg_init(txq, loc, loc->wqe_last, part + 2,
2582 				  MLX5_OPCODE_ENHANCED_MPSW, olx);
2583 		mlx5_tx_eseg_none(txq, loc, loc->wqe_last,
2584 				  olx & ~MLX5_TXOFF_CONFIG_VLAN);
2585 		eseg = &loc->wqe_last->eseg;
2586 		dseg = &loc->wqe_last->dseg[0];
2587 		loop = part;
2588 		/* Store the packet length for legacy MPW. */
2589 		if (MLX5_TXOFF_CONFIG(MPW))
2590 			eseg->mss = rte_cpu_to_be_16
2591 					(rte_pktmbuf_data_len(loc->mbuf));
2592 		for (;;) {
2593 			uint32_t dlen = rte_pktmbuf_data_len(loc->mbuf);
2594 #ifdef MLX5_PMD_SOFT_COUNTERS
2595 			/* Update sent data bytes counter. */
2596 			slen += dlen;
2597 #endif
2598 			mlx5_tx_dseg_ptr
2599 				(txq, loc, dseg,
2600 				 rte_pktmbuf_mtod(loc->mbuf, uint8_t *),
2601 				 dlen, olx);
2602 			if (unlikely(--loop == 0))
2603 				break;
2604 			loc->mbuf = *pkts++;
2605 			if (likely(loop > 1))
2606 				rte_prefetch0(*pkts);
2607 			ret = mlx5_tx_able_to_empw(txq, loc, olx, true);
2608 			/*
2609 			 * Unroll the completion code to avoid
2610 			 * returning variable value - it results in
2611 			 * unoptimized sequent checking in caller.
2612 			 */
2613 			if (ret == MLX5_TXCMP_CODE_MULTI) {
2614 				part -= loop;
2615 				mlx5_tx_sdone_empw(txq, loc, part, slen, olx);
2616 				if (unlikely(!loc->elts_free ||
2617 					     !loc->wqe_free))
2618 					return MLX5_TXCMP_CODE_EXIT;
2619 				return MLX5_TXCMP_CODE_MULTI;
2620 			}
2621 			MLX5_ASSERT(NB_SEGS(loc->mbuf) == 1);
2622 			if (ret == MLX5_TXCMP_CODE_TSO) {
2623 				part -= loop;
2624 				mlx5_tx_sdone_empw(txq, loc, part, slen, olx);
2625 				if (unlikely(!loc->elts_free ||
2626 					     !loc->wqe_free))
2627 					return MLX5_TXCMP_CODE_EXIT;
2628 				return MLX5_TXCMP_CODE_TSO;
2629 			}
2630 			if (ret == MLX5_TXCMP_CODE_SINGLE) {
2631 				part -= loop;
2632 				mlx5_tx_sdone_empw(txq, loc, part, slen, olx);
2633 				if (unlikely(!loc->elts_free ||
2634 					     !loc->wqe_free))
2635 					return MLX5_TXCMP_CODE_EXIT;
2636 				return MLX5_TXCMP_CODE_SINGLE;
2637 			}
2638 			if (ret != MLX5_TXCMP_CODE_EMPW) {
2639 				MLX5_ASSERT(false);
2640 				part -= loop;
2641 				mlx5_tx_sdone_empw(txq, loc, part, slen, olx);
2642 				return MLX5_TXCMP_CODE_ERROR;
2643 			}
2644 			/*
2645 			 * Check whether packet parameters coincide
2646 			 * within assumed eMPW batch:
2647 			 * - check sum settings
2648 			 * - metadata value
2649 			 * - software parser settings
2650 			 * - packets length (legacy MPW only)
2651 			 * - scheduling is not required
2652 			 */
2653 			if (!mlx5_tx_match_empw(txq, eseg, loc, dlen, olx)) {
2654 				MLX5_ASSERT(loop);
2655 				part -= loop;
2656 				mlx5_tx_sdone_empw(txq, loc, part, slen, olx);
2657 				if (unlikely(!loc->elts_free ||
2658 					     !loc->wqe_free))
2659 					return MLX5_TXCMP_CODE_EXIT;
2660 				pkts_n -= part;
2661 				goto next_empw;
2662 			}
2663 			/* Packet attributes match, continue the same eMPW. */
2664 			++dseg;
2665 			if ((uintptr_t)dseg >= (uintptr_t)txq->wqes_end)
2666 				dseg = (struct mlx5_wqe_dseg *)txq->wqes;
2667 		}
2668 		/* eMPW is built successfully, update loop parameters. */
2669 		MLX5_ASSERT(!loop);
2670 		MLX5_ASSERT(pkts_n >= part);
2671 #ifdef MLX5_PMD_SOFT_COUNTERS
2672 		/* Update sent data bytes counter. */
2673 		txq->stats.obytes += slen;
2674 #endif
2675 		loc->elts_free -= part;
2676 		loc->pkts_sent += part;
2677 		txq->wqe_ci += (2 + part + 3) / 4;
2678 		loc->wqe_free -= (2 + part + 3) / 4;
2679 		pkts_n -= part;
2680 		if (unlikely(!pkts_n || !loc->elts_free || !loc->wqe_free))
2681 			return MLX5_TXCMP_CODE_EXIT;
2682 		loc->mbuf = *pkts++;
2683 		ret = mlx5_tx_able_to_empw(txq, loc, olx, true);
2684 		if (unlikely(ret != MLX5_TXCMP_CODE_EMPW))
2685 			return ret;
2686 		/* Continue sending eMPW batches. */
2687 	}
2688 	MLX5_ASSERT(false);
2689 }
2690 
2691 /**
2692  * The routine sends packets with MLX5_OPCODE_EMPW
2693  * with inlining, optionally supports VLAN insertion.
2694  */
2695 static __rte_always_inline enum mlx5_txcmp_code
2696 mlx5_tx_burst_empw_inline(struct mlx5_txq_data *__rte_restrict txq,
2697 			  struct rte_mbuf **__rte_restrict pkts,
2698 			  unsigned int pkts_n,
2699 			  struct mlx5_txq_local *__rte_restrict loc,
2700 			  unsigned int olx)
2701 {
2702 	/*
2703 	 * Subroutine is the part of mlx5_tx_burst_single() and sends
2704 	 * single-segment packet with eMPW opcode with data inlining.
2705 	 */
2706 	MLX5_ASSERT(MLX5_TXOFF_CONFIG(INLINE));
2707 	MLX5_ASSERT(MLX5_TXOFF_CONFIG(EMPW));
2708 	MLX5_ASSERT(loc->elts_free && loc->wqe_free);
2709 	MLX5_ASSERT(pkts_n > loc->pkts_sent);
2710 	pkts += loc->pkts_sent + 1;
2711 	pkts_n -= loc->pkts_sent;
2712 	for (;;) {
2713 		struct mlx5_wqe_dseg *__rte_restrict dseg;
2714 		struct mlx5_wqe *__rte_restrict wqem;
2715 		enum mlx5_txcmp_code ret;
2716 		unsigned int room, part, nlim;
2717 		unsigned int slen = 0;
2718 
2719 		MLX5_ASSERT(NB_SEGS(loc->mbuf) == 1);
2720 		if (MLX5_TXOFF_CONFIG(TXPP)) {
2721 			enum mlx5_txcmp_code wret;
2722 
2723 			/* Generate WAIT for scheduling if requested. */
2724 			wret = mlx5_tx_schedule_send(txq, loc, olx);
2725 			if (wret == MLX5_TXCMP_CODE_EXIT)
2726 				return MLX5_TXCMP_CODE_EXIT;
2727 			if (wret == MLX5_TXCMP_CODE_ERROR)
2728 				return MLX5_TXCMP_CODE_ERROR;
2729 		}
2730 		/*
2731 		 * Limits the amount of packets in one WQE
2732 		 * to improve CQE latency generation.
2733 		 */
2734 		nlim = RTE_MIN(pkts_n, MLX5_TXOFF_CONFIG(MPW) ?
2735 				       MLX5_MPW_INLINE_MAX_PACKETS :
2736 				       MLX5_EMPW_MAX_PACKETS);
2737 		/* Check whether we have minimal amount WQEs */
2738 		if (unlikely(loc->wqe_free <
2739 			    ((2 + MLX5_EMPW_MIN_PACKETS + 3) / 4)))
2740 			return MLX5_TXCMP_CODE_EXIT;
2741 		if (likely(pkts_n > 1))
2742 			rte_prefetch0(*pkts);
2743 		wqem = txq->wqes + (txq->wqe_ci & txq->wqe_m);
2744 		/*
2745 		 * Build eMPW title WQEBB:
2746 		 * - Control Segment, eMPW opcode, zero DS
2747 		 * - Ethernet Segment, no inline
2748 		 */
2749 		mlx5_tx_cseg_init(txq, loc, wqem, 0,
2750 				  MLX5_OPCODE_ENHANCED_MPSW, olx);
2751 		mlx5_tx_eseg_none(txq, loc, wqem,
2752 				  olx & ~MLX5_TXOFF_CONFIG_VLAN);
2753 		dseg = &wqem->dseg[0];
2754 		/* Store the packet length for legacy MPW. */
2755 		if (MLX5_TXOFF_CONFIG(MPW))
2756 			wqem->eseg.mss = rte_cpu_to_be_16
2757 					 (rte_pktmbuf_data_len(loc->mbuf));
2758 		room = RTE_MIN(MLX5_WQE_SIZE_MAX / MLX5_WQE_SIZE,
2759 			       loc->wqe_free) * MLX5_WQE_SIZE -
2760 					MLX5_WQE_CSEG_SIZE -
2761 					MLX5_WQE_ESEG_SIZE;
2762 		/* Limit the room for legacy MPW sessions for performance. */
2763 		if (MLX5_TXOFF_CONFIG(MPW))
2764 			room = RTE_MIN(room,
2765 				       RTE_MAX(txq->inlen_empw +
2766 					       sizeof(dseg->bcount) +
2767 					       (MLX5_TXOFF_CONFIG(VLAN) ?
2768 					       sizeof(struct rte_vlan_hdr) : 0),
2769 					       MLX5_MPW_INLINE_MAX_PACKETS *
2770 					       MLX5_WQE_DSEG_SIZE));
2771 		/* Build WQE till we have space, packets and resources. */
2772 		part = room;
2773 		for (;;) {
2774 			uint32_t dlen = rte_pktmbuf_data_len(loc->mbuf);
2775 			uint8_t *dptr = rte_pktmbuf_mtod(loc->mbuf, uint8_t *);
2776 			unsigned int tlen;
2777 
2778 			MLX5_ASSERT(room >= MLX5_WQE_DSEG_SIZE);
2779 			MLX5_ASSERT((room % MLX5_WQE_DSEG_SIZE) == 0);
2780 			MLX5_ASSERT((uintptr_t)dseg < (uintptr_t)txq->wqes_end);
2781 			/*
2782 			 * Some Tx offloads may cause an error if packet is not
2783 			 * long enough, check against assumed minimal length.
2784 			 */
2785 			if (unlikely(dlen <= MLX5_ESEG_MIN_INLINE_SIZE)) {
2786 				part -= room;
2787 				if (unlikely(!part))
2788 					return MLX5_TXCMP_CODE_ERROR;
2789 				/*
2790 				 * We have some successfully built
2791 				 * packet Data Segments to send.
2792 				 */
2793 				mlx5_tx_idone_empw(txq, loc, part,
2794 						   slen, wqem, olx);
2795 				return MLX5_TXCMP_CODE_ERROR;
2796 			}
2797 			/* Inline or not inline - that's the Question. */
2798 			if (dlen > txq->inlen_empw ||
2799 			    loc->mbuf->ol_flags & RTE_MBUF_F_TX_DYNF_NOINLINE)
2800 				goto pointer_empw;
2801 			if (MLX5_TXOFF_CONFIG(MPW)) {
2802 				if (dlen > txq->inlen_send)
2803 					goto pointer_empw;
2804 				tlen = dlen;
2805 				if (part == room) {
2806 					/* Open new inline MPW session. */
2807 					tlen += sizeof(dseg->bcount);
2808 					dseg->bcount = RTE_BE32(0);
2809 					dseg = RTE_PTR_ADD
2810 						(dseg, sizeof(dseg->bcount));
2811 				} else {
2812 					/*
2813 					 * No pointer and inline descriptor
2814 					 * intermix for legacy MPW sessions.
2815 					 */
2816 					if (wqem->dseg[0].bcount)
2817 						break;
2818 				}
2819 			} else {
2820 				tlen = sizeof(dseg->bcount) + dlen;
2821 			}
2822 			/* Inline entire packet, optional VLAN insertion. */
2823 			if (MLX5_TXOFF_CONFIG(VLAN) &&
2824 			    loc->mbuf->ol_flags & RTE_MBUF_F_TX_VLAN) {
2825 				/*
2826 				 * The packet length must be checked in
2827 				 * mlx5_tx_able_to_empw() and packet
2828 				 * fits into inline length guaranteed.
2829 				 */
2830 				MLX5_ASSERT((dlen +
2831 					     sizeof(struct rte_vlan_hdr)) <=
2832 					    txq->inlen_empw);
2833 				tlen += sizeof(struct rte_vlan_hdr);
2834 				if (room < tlen)
2835 					break;
2836 				dseg = mlx5_tx_dseg_vlan(txq, loc, dseg,
2837 							 dptr, dlen, olx);
2838 #ifdef MLX5_PMD_SOFT_COUNTERS
2839 				/* Update sent data bytes counter. */
2840 				slen +=	sizeof(struct rte_vlan_hdr);
2841 #endif
2842 			} else {
2843 				if (room < tlen)
2844 					break;
2845 				dseg = mlx5_tx_dseg_empw(txq, loc, dseg,
2846 							 dptr, dlen, olx);
2847 			}
2848 			if (!MLX5_TXOFF_CONFIG(MPW))
2849 				tlen = RTE_ALIGN(tlen, MLX5_WSEG_SIZE);
2850 			MLX5_ASSERT(room >= tlen);
2851 			room -= tlen;
2852 			/*
2853 			 * Packet data are completely inline,
2854 			 * we can try to free the packet.
2855 			 */
2856 			if (likely(loc->pkts_sent == loc->mbuf_free)) {
2857 				/*
2858 				 * All the packets from the burst beginning
2859 				 * are inline, we can free mbufs directly
2860 				 * from the origin array on tx_burst exit().
2861 				 */
2862 				loc->mbuf_free++;
2863 				goto next_mbuf;
2864 			}
2865 			/*
2866 			 * In order no to call rte_pktmbuf_free_seg() here,
2867 			 * in the most inner loop (that might be very
2868 			 * expensive) we just save the mbuf in elts.
2869 			 */
2870 			txq->elts[txq->elts_head++ & txq->elts_m] = loc->mbuf;
2871 			loc->elts_free--;
2872 			goto next_mbuf;
2873 pointer_empw:
2874 			/*
2875 			 * No pointer and inline descriptor
2876 			 * intermix for legacy MPW sessions.
2877 			 */
2878 			if (MLX5_TXOFF_CONFIG(MPW) &&
2879 			    part != room &&
2880 			    wqem->dseg[0].bcount == RTE_BE32(0))
2881 				break;
2882 			/*
2883 			 * Not inlinable VLAN packets are
2884 			 * proceeded outside of this routine.
2885 			 */
2886 			MLX5_ASSERT(room >= MLX5_WQE_DSEG_SIZE);
2887 			if (MLX5_TXOFF_CONFIG(VLAN))
2888 				MLX5_ASSERT(!(loc->mbuf->ol_flags &
2889 					    RTE_MBUF_F_TX_VLAN));
2890 			mlx5_tx_dseg_ptr(txq, loc, dseg, dptr, dlen, olx);
2891 			/* We have to store mbuf in elts.*/
2892 			txq->elts[txq->elts_head++ & txq->elts_m] = loc->mbuf;
2893 			loc->elts_free--;
2894 			room -= MLX5_WQE_DSEG_SIZE;
2895 			/* Ring buffer wraparound is checked at the loop end.*/
2896 			++dseg;
2897 next_mbuf:
2898 #ifdef MLX5_PMD_SOFT_COUNTERS
2899 			/* Update sent data bytes counter. */
2900 			slen += dlen;
2901 #endif
2902 			loc->pkts_sent++;
2903 			pkts_n--;
2904 			if (unlikely(!pkts_n || !loc->elts_free)) {
2905 				/*
2906 				 * We have no resources/packets to
2907 				 * continue build descriptors.
2908 				 */
2909 				part -= room;
2910 				mlx5_tx_idone_empw(txq, loc, part,
2911 						   slen, wqem, olx);
2912 				return MLX5_TXCMP_CODE_EXIT;
2913 			}
2914 			loc->mbuf = *pkts++;
2915 			if (likely(pkts_n > 1))
2916 				rte_prefetch0(*pkts);
2917 			ret = mlx5_tx_able_to_empw(txq, loc, olx, true);
2918 			/*
2919 			 * Unroll the completion code to avoid
2920 			 * returning variable value - it results in
2921 			 * unoptimized sequent checking in caller.
2922 			 */
2923 			if (ret == MLX5_TXCMP_CODE_MULTI) {
2924 				part -= room;
2925 				mlx5_tx_idone_empw(txq, loc, part,
2926 						   slen, wqem, olx);
2927 				if (unlikely(!loc->elts_free ||
2928 					     !loc->wqe_free))
2929 					return MLX5_TXCMP_CODE_EXIT;
2930 				return MLX5_TXCMP_CODE_MULTI;
2931 			}
2932 			MLX5_ASSERT(NB_SEGS(loc->mbuf) == 1);
2933 			if (ret == MLX5_TXCMP_CODE_TSO) {
2934 				part -= room;
2935 				mlx5_tx_idone_empw(txq, loc, part,
2936 						   slen, wqem, olx);
2937 				if (unlikely(!loc->elts_free ||
2938 					     !loc->wqe_free))
2939 					return MLX5_TXCMP_CODE_EXIT;
2940 				return MLX5_TXCMP_CODE_TSO;
2941 			}
2942 			if (ret == MLX5_TXCMP_CODE_SINGLE) {
2943 				part -= room;
2944 				mlx5_tx_idone_empw(txq, loc, part,
2945 						   slen, wqem, olx);
2946 				if (unlikely(!loc->elts_free ||
2947 					     !loc->wqe_free))
2948 					return MLX5_TXCMP_CODE_EXIT;
2949 				return MLX5_TXCMP_CODE_SINGLE;
2950 			}
2951 			if (ret != MLX5_TXCMP_CODE_EMPW) {
2952 				MLX5_ASSERT(false);
2953 				part -= room;
2954 				mlx5_tx_idone_empw(txq, loc, part,
2955 						   slen, wqem, olx);
2956 				return MLX5_TXCMP_CODE_ERROR;
2957 			}
2958 			/* Check if we have minimal room left. */
2959 			nlim--;
2960 			if (unlikely(!nlim || room < MLX5_WQE_DSEG_SIZE))
2961 				break;
2962 			/*
2963 			 * Check whether packet parameters coincide
2964 			 * within assumed eMPW batch:
2965 			 * - check sum settings
2966 			 * - metadata value
2967 			 * - software parser settings
2968 			 * - packets length (legacy MPW only)
2969 			 * - scheduling is not required
2970 			 */
2971 			if (!mlx5_tx_match_empw(txq, &wqem->eseg,
2972 						loc, dlen, olx))
2973 				break;
2974 			/* Packet attributes match, continue the same eMPW. */
2975 			if ((uintptr_t)dseg >= (uintptr_t)txq->wqes_end)
2976 				dseg = (struct mlx5_wqe_dseg *)txq->wqes;
2977 		}
2978 		/*
2979 		 * We get here to close an existing eMPW
2980 		 * session and start the new one.
2981 		 */
2982 		MLX5_ASSERT(pkts_n);
2983 		part -= room;
2984 		if (unlikely(!part))
2985 			return MLX5_TXCMP_CODE_EXIT;
2986 		mlx5_tx_idone_empw(txq, loc, part, slen, wqem, olx);
2987 		if (unlikely(!loc->elts_free ||
2988 			     !loc->wqe_free))
2989 			return MLX5_TXCMP_CODE_EXIT;
2990 		/* Continue the loop with new eMPW session. */
2991 	}
2992 	MLX5_ASSERT(false);
2993 }
2994 
2995 /**
2996  * The routine sends packets with ordinary MLX5_OPCODE_SEND.
2997  * Data inlining and VLAN insertion are supported.
2998  */
2999 static __rte_always_inline enum mlx5_txcmp_code
3000 mlx5_tx_burst_single_send(struct mlx5_txq_data *__rte_restrict txq,
3001 			  struct rte_mbuf **__rte_restrict pkts,
3002 			  unsigned int pkts_n,
3003 			  struct mlx5_txq_local *__rte_restrict loc,
3004 			  unsigned int olx)
3005 {
3006 	/*
3007 	 * Subroutine is the part of mlx5_tx_burst_single()
3008 	 * and sends single-segment packet with SEND opcode.
3009 	 */
3010 	MLX5_ASSERT(loc->elts_free && loc->wqe_free);
3011 	MLX5_ASSERT(pkts_n > loc->pkts_sent);
3012 	pkts += loc->pkts_sent + 1;
3013 	pkts_n -= loc->pkts_sent;
3014 	for (;;) {
3015 		struct mlx5_wqe *__rte_restrict wqe;
3016 		enum mlx5_txcmp_code ret;
3017 
3018 		MLX5_ASSERT(NB_SEGS(loc->mbuf) == 1);
3019 		if (MLX5_TXOFF_CONFIG(TXPP)) {
3020 			enum mlx5_txcmp_code wret;
3021 
3022 			/* Generate WAIT for scheduling if requested. */
3023 			wret = mlx5_tx_schedule_send(txq, loc, olx);
3024 			if (wret == MLX5_TXCMP_CODE_EXIT)
3025 				return MLX5_TXCMP_CODE_EXIT;
3026 			if (wret == MLX5_TXCMP_CODE_ERROR)
3027 				return MLX5_TXCMP_CODE_ERROR;
3028 		}
3029 		if (MLX5_TXOFF_CONFIG(INLINE)) {
3030 			unsigned int inlen, vlan = 0;
3031 
3032 			inlen = rte_pktmbuf_data_len(loc->mbuf);
3033 			if (MLX5_TXOFF_CONFIG(VLAN) &&
3034 			    loc->mbuf->ol_flags & RTE_MBUF_F_TX_VLAN) {
3035 				vlan = sizeof(struct rte_vlan_hdr);
3036 				inlen += vlan;
3037 			}
3038 			/*
3039 			 * If inlining is enabled at configuration time
3040 			 * the limit must be not less than minimal size.
3041 			 * Otherwise we would do extra check for data
3042 			 * size to avoid crashes due to length overflow.
3043 			 */
3044 			MLX5_ASSERT(txq->inlen_send >=
3045 				    MLX5_ESEG_MIN_INLINE_SIZE);
3046 			if (inlen <= txq->inlen_send) {
3047 				unsigned int seg_n, wqe_n;
3048 
3049 				rte_prefetch0(rte_pktmbuf_mtod
3050 						(loc->mbuf, uint8_t *));
3051 				/* Check against minimal length. */
3052 				if (inlen <= MLX5_ESEG_MIN_INLINE_SIZE)
3053 					return MLX5_TXCMP_CODE_ERROR;
3054 				if (loc->mbuf->ol_flags &
3055 				    RTE_MBUF_F_TX_DYNF_NOINLINE) {
3056 					/*
3057 					 * The hint flag not to inline packet
3058 					 * data is set. Check whether we can
3059 					 * follow the hint.
3060 					 */
3061 					if ((!MLX5_TXOFF_CONFIG(EMPW) &&
3062 					      txq->inlen_mode) ||
3063 					    (MLX5_TXOFF_CONFIG(MPW) &&
3064 					     txq->inlen_mode)) {
3065 						if (inlen <= txq->inlen_send)
3066 							goto single_inline;
3067 						/*
3068 						 * The hardware requires the
3069 						 * minimal inline data header.
3070 						 */
3071 						goto single_min_inline;
3072 					}
3073 					if (MLX5_TXOFF_CONFIG(VLAN) &&
3074 					    vlan && !txq->vlan_en) {
3075 						/*
3076 						 * We must insert VLAN tag
3077 						 * by software means.
3078 						 */
3079 						goto single_part_inline;
3080 					}
3081 					goto single_no_inline;
3082 				}
3083 single_inline:
3084 				/*
3085 				 * Completely inlined packet data WQE:
3086 				 * - Control Segment, SEND opcode
3087 				 * - Ethernet Segment, no VLAN insertion
3088 				 * - Data inlined, VLAN optionally inserted
3089 				 * - Alignment to MLX5_WSEG_SIZE
3090 				 * Have to estimate amount of WQEBBs
3091 				 */
3092 				seg_n = (inlen + 3 * MLX5_WSEG_SIZE -
3093 					 MLX5_ESEG_MIN_INLINE_SIZE +
3094 					 MLX5_WSEG_SIZE - 1) / MLX5_WSEG_SIZE;
3095 				/* Check if there are enough WQEBBs. */
3096 				wqe_n = (seg_n + 3) / 4;
3097 				if (wqe_n > loc->wqe_free)
3098 					return MLX5_TXCMP_CODE_EXIT;
3099 				wqe = txq->wqes + (txq->wqe_ci & txq->wqe_m);
3100 				loc->wqe_last = wqe;
3101 				mlx5_tx_cseg_init(txq, loc, wqe, seg_n,
3102 						  MLX5_OPCODE_SEND, olx);
3103 				mlx5_tx_eseg_data(txq, loc, wqe,
3104 						  vlan, inlen, 0, olx);
3105 				txq->wqe_ci += wqe_n;
3106 				loc->wqe_free -= wqe_n;
3107 				/*
3108 				 * Packet data are completely inlined,
3109 				 * free the packet immediately.
3110 				 */
3111 				rte_pktmbuf_free_seg(loc->mbuf);
3112 			} else if ((!MLX5_TXOFF_CONFIG(EMPW) ||
3113 				     MLX5_TXOFF_CONFIG(MPW)) &&
3114 					txq->inlen_mode) {
3115 				/*
3116 				 * If minimal inlining is requested the eMPW
3117 				 * feature should be disabled due to data is
3118 				 * inlined into Ethernet Segment, which can
3119 				 * not contain inlined data for eMPW due to
3120 				 * segment shared for all packets.
3121 				 */
3122 				struct mlx5_wqe_dseg *__rte_restrict dseg;
3123 				unsigned int ds;
3124 				uint8_t *dptr;
3125 
3126 				/*
3127 				 * The inline-mode settings require
3128 				 * to inline the specified amount of
3129 				 * data bytes to the Ethernet Segment.
3130 				 * We should check the free space in
3131 				 * WQE ring buffer to inline partially.
3132 				 */
3133 single_min_inline:
3134 				MLX5_ASSERT(txq->inlen_send >= txq->inlen_mode);
3135 				MLX5_ASSERT(inlen > txq->inlen_mode);
3136 				MLX5_ASSERT(txq->inlen_mode >=
3137 					    MLX5_ESEG_MIN_INLINE_SIZE);
3138 				/*
3139 				 * Check whether there are enough free WQEBBs:
3140 				 * - Control Segment
3141 				 * - Ethernet Segment
3142 				 * - First Segment of inlined Ethernet data
3143 				 * - ... data continued ...
3144 				 * - Finishing Data Segment of pointer type
3145 				 */
3146 				ds = (MLX5_WQE_CSEG_SIZE +
3147 				      MLX5_WQE_ESEG_SIZE +
3148 				      MLX5_WQE_DSEG_SIZE +
3149 				      txq->inlen_mode -
3150 				      MLX5_ESEG_MIN_INLINE_SIZE +
3151 				      MLX5_WQE_DSEG_SIZE +
3152 				      MLX5_WSEG_SIZE - 1) / MLX5_WSEG_SIZE;
3153 				if (loc->wqe_free < ((ds + 3) / 4))
3154 					return MLX5_TXCMP_CODE_EXIT;
3155 				/*
3156 				 * Build the ordinary SEND WQE:
3157 				 * - Control Segment
3158 				 * - Ethernet Segment, inline inlen_mode bytes
3159 				 * - Data Segment of pointer type
3160 				 */
3161 				wqe = txq->wqes + (txq->wqe_ci & txq->wqe_m);
3162 				loc->wqe_last = wqe;
3163 				mlx5_tx_cseg_init(txq, loc, wqe, ds,
3164 						  MLX5_OPCODE_SEND, olx);
3165 				dseg = mlx5_tx_eseg_data(txq, loc, wqe, vlan,
3166 							 txq->inlen_mode,
3167 							 0, olx);
3168 				dptr = rte_pktmbuf_mtod(loc->mbuf, uint8_t *) +
3169 				       txq->inlen_mode - vlan;
3170 				inlen -= txq->inlen_mode;
3171 				mlx5_tx_dseg_ptr(txq, loc, dseg,
3172 						 dptr, inlen, olx);
3173 				/*
3174 				 * WQE is built, update the loop parameters
3175 				 * and got to the next packet.
3176 				 */
3177 				txq->wqe_ci += (ds + 3) / 4;
3178 				loc->wqe_free -= (ds + 3) / 4;
3179 				/* We have to store mbuf in elts.*/
3180 				MLX5_ASSERT(MLX5_TXOFF_CONFIG(INLINE));
3181 				txq->elts[txq->elts_head++ & txq->elts_m] =
3182 						loc->mbuf;
3183 				--loc->elts_free;
3184 			} else {
3185 				uint8_t *dptr;
3186 				unsigned int dlen;
3187 
3188 				/*
3189 				 * Partially inlined packet data WQE, we have
3190 				 * some space in title WQEBB, we can fill it
3191 				 * with some packet data. It takes one WQEBB,
3192 				 * it is available, no extra space check:
3193 				 * - Control Segment, SEND opcode
3194 				 * - Ethernet Segment, no VLAN insertion
3195 				 * - MLX5_ESEG_MIN_INLINE_SIZE bytes of Data
3196 				 * - Data Segment, pointer type
3197 				 *
3198 				 * We also get here if VLAN insertion is not
3199 				 * supported by HW, the inline is enabled.
3200 				 */
3201 single_part_inline:
3202 				wqe = txq->wqes + (txq->wqe_ci & txq->wqe_m);
3203 				loc->wqe_last = wqe;
3204 				mlx5_tx_cseg_init(txq, loc, wqe, 4,
3205 						  MLX5_OPCODE_SEND, olx);
3206 				mlx5_tx_eseg_dmin(txq, loc, wqe, vlan, olx);
3207 				dptr = rte_pktmbuf_mtod(loc->mbuf, uint8_t *) +
3208 				       MLX5_ESEG_MIN_INLINE_SIZE - vlan;
3209 				/*
3210 				 * The length check is performed above, by
3211 				 * comparing with txq->inlen_send. We should
3212 				 * not get overflow here.
3213 				 */
3214 				MLX5_ASSERT(inlen > MLX5_ESEG_MIN_INLINE_SIZE);
3215 				dlen = inlen - MLX5_ESEG_MIN_INLINE_SIZE;
3216 				mlx5_tx_dseg_ptr(txq, loc, &wqe->dseg[1],
3217 						 dptr, dlen, olx);
3218 				++txq->wqe_ci;
3219 				--loc->wqe_free;
3220 				/* We have to store mbuf in elts.*/
3221 				MLX5_ASSERT(MLX5_TXOFF_CONFIG(INLINE));
3222 				txq->elts[txq->elts_head++ & txq->elts_m] =
3223 						loc->mbuf;
3224 				--loc->elts_free;
3225 			}
3226 #ifdef MLX5_PMD_SOFT_COUNTERS
3227 			/* Update sent data bytes counter. */
3228 			txq->stats.obytes += vlan +
3229 					rte_pktmbuf_data_len(loc->mbuf);
3230 #endif
3231 		} else {
3232 			/*
3233 			 * No inline at all, it means the CPU cycles saving
3234 			 * is prioritized at configuration, we should not
3235 			 * copy any packet data to WQE.
3236 			 *
3237 			 * SEND WQE, one WQEBB:
3238 			 * - Control Segment, SEND opcode
3239 			 * - Ethernet Segment, optional VLAN, no inline
3240 			 * - Data Segment, pointer type
3241 			 */
3242 single_no_inline:
3243 			wqe = txq->wqes + (txq->wqe_ci & txq->wqe_m);
3244 			loc->wqe_last = wqe;
3245 			mlx5_tx_cseg_init(txq, loc, wqe, 3,
3246 					  MLX5_OPCODE_SEND, olx);
3247 			mlx5_tx_eseg_none(txq, loc, wqe, olx);
3248 			mlx5_tx_dseg_ptr
3249 				(txq, loc, &wqe->dseg[0],
3250 				 rte_pktmbuf_mtod(loc->mbuf, uint8_t *),
3251 				 rte_pktmbuf_data_len(loc->mbuf), olx);
3252 			++txq->wqe_ci;
3253 			--loc->wqe_free;
3254 			/*
3255 			 * We should not store mbuf pointer in elts
3256 			 * if no inlining is configured, this is done
3257 			 * by calling routine in a batch copy.
3258 			 */
3259 			MLX5_ASSERT(!MLX5_TXOFF_CONFIG(INLINE));
3260 			--loc->elts_free;
3261 #ifdef MLX5_PMD_SOFT_COUNTERS
3262 			/* Update sent data bytes counter. */
3263 			txq->stats.obytes += rte_pktmbuf_data_len(loc->mbuf);
3264 			if (MLX5_TXOFF_CONFIG(VLAN) &&
3265 			    loc->mbuf->ol_flags & RTE_MBUF_F_TX_VLAN)
3266 				txq->stats.obytes +=
3267 					sizeof(struct rte_vlan_hdr);
3268 #endif
3269 		}
3270 		++loc->pkts_sent;
3271 		--pkts_n;
3272 		if (unlikely(!pkts_n || !loc->elts_free || !loc->wqe_free))
3273 			return MLX5_TXCMP_CODE_EXIT;
3274 		loc->mbuf = *pkts++;
3275 		if (pkts_n > 1)
3276 			rte_prefetch0(*pkts);
3277 		ret = mlx5_tx_able_to_empw(txq, loc, olx, true);
3278 		if (unlikely(ret != MLX5_TXCMP_CODE_SINGLE))
3279 			return ret;
3280 	}
3281 	MLX5_ASSERT(false);
3282 }
3283 
3284 static __rte_always_inline enum mlx5_txcmp_code
3285 mlx5_tx_burst_single(struct mlx5_txq_data *__rte_restrict txq,
3286 		     struct rte_mbuf **__rte_restrict pkts,
3287 		     unsigned int pkts_n,
3288 		     struct mlx5_txq_local *__rte_restrict loc,
3289 		     unsigned int olx)
3290 {
3291 	enum mlx5_txcmp_code ret;
3292 
3293 	ret = mlx5_tx_able_to_empw(txq, loc, olx, false);
3294 	if (ret == MLX5_TXCMP_CODE_SINGLE)
3295 		goto ordinary_send;
3296 	MLX5_ASSERT(ret == MLX5_TXCMP_CODE_EMPW);
3297 	for (;;) {
3298 		/* Optimize for inline/no inline eMPW send. */
3299 		ret = (MLX5_TXOFF_CONFIG(INLINE)) ?
3300 			mlx5_tx_burst_empw_inline
3301 				(txq, pkts, pkts_n, loc, olx) :
3302 			mlx5_tx_burst_empw_simple
3303 				(txq, pkts, pkts_n, loc, olx);
3304 		if (ret != MLX5_TXCMP_CODE_SINGLE)
3305 			return ret;
3306 		/* The resources to send one packet should remain. */
3307 		MLX5_ASSERT(loc->elts_free && loc->wqe_free);
3308 ordinary_send:
3309 		ret = mlx5_tx_burst_single_send(txq, pkts, pkts_n, loc, olx);
3310 		MLX5_ASSERT(ret != MLX5_TXCMP_CODE_SINGLE);
3311 		if (ret != MLX5_TXCMP_CODE_EMPW)
3312 			return ret;
3313 		/* The resources to send one packet should remain. */
3314 		MLX5_ASSERT(loc->elts_free && loc->wqe_free);
3315 	}
3316 }
3317 
3318 /**
3319  * DPDK Tx callback template. This is configured template used to generate
3320  * routines optimized for specified offload setup.
3321  * One of this generated functions is chosen at SQ configuration time.
3322  *
3323  * @param txq
3324  *   Generic pointer to TX queue structure.
3325  * @param[in] pkts
3326  *   Packets to transmit.
3327  * @param pkts_n
3328  *   Number of packets in array.
3329  * @param olx
3330  *   Configured offloads mask, presents the bits of MLX5_TXOFF_CONFIG_xxx
3331  *   values. Should be static to take compile time static configuration
3332  *   advantages.
3333  *
3334  * @return
3335  *   Number of packets successfully transmitted (<= pkts_n).
3336  */
3337 static __rte_always_inline uint16_t
3338 mlx5_tx_burst_tmpl(struct mlx5_txq_data *__rte_restrict txq,
3339 		   struct rte_mbuf **__rte_restrict pkts,
3340 		   uint16_t pkts_n,
3341 		   unsigned int olx)
3342 {
3343 	struct mlx5_txq_local loc;
3344 	enum mlx5_txcmp_code ret;
3345 	unsigned int part;
3346 
3347 	MLX5_ASSERT(txq->elts_s >= (uint16_t)(txq->elts_head - txq->elts_tail));
3348 	MLX5_ASSERT(txq->wqe_s >= (uint16_t)(txq->wqe_ci - txq->wqe_pi));
3349 	if (unlikely(!pkts_n))
3350 		return 0;
3351 	if (MLX5_TXOFF_CONFIG(INLINE))
3352 		loc.mbuf_free = 0;
3353 	loc.pkts_sent = 0;
3354 	loc.pkts_copy = 0;
3355 	loc.wqe_last = NULL;
3356 
3357 send_loop:
3358 	loc.pkts_loop = loc.pkts_sent;
3359 	/*
3360 	 * Check if there are some CQEs, if any:
3361 	 * - process an encountered errors
3362 	 * - process the completed WQEs
3363 	 * - free related mbufs
3364 	 * - doorbell the NIC about processed CQEs
3365 	 */
3366 	rte_prefetch0(*(pkts + loc.pkts_sent));
3367 	mlx5_tx_handle_completion(txq, olx);
3368 	/*
3369 	 * Calculate the number of available resources - elts and WQEs.
3370 	 * There are two possible different scenarios:
3371 	 * - no data inlining into WQEs, one WQEBB may contains up to
3372 	 *   four packets, in this case elts become scarce resource
3373 	 * - data inlining into WQEs, one packet may require multiple
3374 	 *   WQEBBs, the WQEs become the limiting factor.
3375 	 */
3376 	MLX5_ASSERT(txq->elts_s >= (uint16_t)(txq->elts_head - txq->elts_tail));
3377 	loc.elts_free = txq->elts_s -
3378 				(uint16_t)(txq->elts_head - txq->elts_tail);
3379 	MLX5_ASSERT(txq->wqe_s >= (uint16_t)(txq->wqe_ci - txq->wqe_pi));
3380 	loc.wqe_free = txq->wqe_s -
3381 				(uint16_t)(txq->wqe_ci - txq->wqe_pi);
3382 	if (unlikely(!loc.elts_free || !loc.wqe_free))
3383 		goto burst_exit;
3384 	for (;;) {
3385 		/*
3386 		 * Fetch the packet from array. Usually this is the first
3387 		 * packet in series of multi/single segment packets.
3388 		 */
3389 		loc.mbuf = *(pkts + loc.pkts_sent);
3390 		/* Dedicated branch for multi-segment packets. */
3391 		if (MLX5_TXOFF_CONFIG(MULTI) &&
3392 		    unlikely(NB_SEGS(loc.mbuf) > 1)) {
3393 			/*
3394 			 * Multi-segment packet encountered.
3395 			 * Hardware is able to process it only
3396 			 * with SEND/TSO opcodes, one packet
3397 			 * per WQE, do it in dedicated routine.
3398 			 */
3399 enter_send_multi:
3400 			MLX5_ASSERT(loc.pkts_sent >= loc.pkts_copy);
3401 			part = loc.pkts_sent - loc.pkts_copy;
3402 			if (!MLX5_TXOFF_CONFIG(INLINE) && part) {
3403 				/*
3404 				 * There are some single-segment mbufs not
3405 				 * stored in elts. The mbufs must be in the
3406 				 * same order as WQEs, so we must copy the
3407 				 * mbufs to elts here, before the coming
3408 				 * multi-segment packet mbufs is appended.
3409 				 */
3410 				mlx5_tx_copy_elts(txq, pkts + loc.pkts_copy,
3411 						  part, olx);
3412 				loc.pkts_copy = loc.pkts_sent;
3413 			}
3414 			MLX5_ASSERT(pkts_n > loc.pkts_sent);
3415 			ret = mlx5_tx_burst_mseg(txq, pkts, pkts_n, &loc, olx);
3416 			if (!MLX5_TXOFF_CONFIG(INLINE))
3417 				loc.pkts_copy = loc.pkts_sent;
3418 			/*
3419 			 * These returned code checks are supposed
3420 			 * to be optimized out due to routine inlining.
3421 			 */
3422 			if (ret == MLX5_TXCMP_CODE_EXIT) {
3423 				/*
3424 				 * The routine returns this code when
3425 				 * all packets are sent or there is no
3426 				 * enough resources to complete request.
3427 				 */
3428 				break;
3429 			}
3430 			if (ret == MLX5_TXCMP_CODE_ERROR) {
3431 				/*
3432 				 * The routine returns this code when some error
3433 				 * in the incoming packets format occurred.
3434 				 */
3435 				txq->stats.oerrors++;
3436 				break;
3437 			}
3438 			if (ret == MLX5_TXCMP_CODE_SINGLE) {
3439 				/*
3440 				 * The single-segment packet was encountered
3441 				 * in the array, try to send it with the
3442 				 * best optimized way, possible engaging eMPW.
3443 				 */
3444 				goto enter_send_single;
3445 			}
3446 			if (MLX5_TXOFF_CONFIG(TSO) &&
3447 			    ret == MLX5_TXCMP_CODE_TSO) {
3448 				/*
3449 				 * The single-segment TSO packet was
3450 				 * encountered in the array.
3451 				 */
3452 				goto enter_send_tso;
3453 			}
3454 			/* We must not get here. Something is going wrong. */
3455 			MLX5_ASSERT(false);
3456 			txq->stats.oerrors++;
3457 			break;
3458 		}
3459 		/* Dedicated branch for single-segment TSO packets. */
3460 		if (MLX5_TXOFF_CONFIG(TSO) &&
3461 		    unlikely(loc.mbuf->ol_flags & RTE_MBUF_F_TX_TCP_SEG)) {
3462 			/*
3463 			 * TSO might require special way for inlining
3464 			 * (dedicated parameters) and is sent with
3465 			 * MLX5_OPCODE_TSO opcode only, provide this
3466 			 * in dedicated branch.
3467 			 */
3468 enter_send_tso:
3469 			MLX5_ASSERT(NB_SEGS(loc.mbuf) == 1);
3470 			MLX5_ASSERT(pkts_n > loc.pkts_sent);
3471 			ret = mlx5_tx_burst_tso(txq, pkts, pkts_n, &loc, olx);
3472 			/*
3473 			 * These returned code checks are supposed
3474 			 * to be optimized out due to routine inlining.
3475 			 */
3476 			if (ret == MLX5_TXCMP_CODE_EXIT)
3477 				break;
3478 			if (ret == MLX5_TXCMP_CODE_ERROR) {
3479 				txq->stats.oerrors++;
3480 				break;
3481 			}
3482 			if (ret == MLX5_TXCMP_CODE_SINGLE)
3483 				goto enter_send_single;
3484 			if (MLX5_TXOFF_CONFIG(MULTI) &&
3485 			    ret == MLX5_TXCMP_CODE_MULTI) {
3486 				/*
3487 				 * The multi-segment packet was
3488 				 * encountered in the array.
3489 				 */
3490 				goto enter_send_multi;
3491 			}
3492 			/* We must not get here. Something is going wrong. */
3493 			MLX5_ASSERT(false);
3494 			txq->stats.oerrors++;
3495 			break;
3496 		}
3497 		/*
3498 		 * The dedicated branch for the single-segment packets
3499 		 * without TSO. Often these ones can be sent using
3500 		 * MLX5_OPCODE_EMPW with multiple packets in one WQE.
3501 		 * The routine builds the WQEs till it encounters
3502 		 * the TSO or multi-segment packet (in case if these
3503 		 * offloads are requested at SQ configuration time).
3504 		 */
3505 enter_send_single:
3506 		MLX5_ASSERT(pkts_n > loc.pkts_sent);
3507 		ret = mlx5_tx_burst_single(txq, pkts, pkts_n, &loc, olx);
3508 		/*
3509 		 * These returned code checks are supposed
3510 		 * to be optimized out due to routine inlining.
3511 		 */
3512 		if (ret == MLX5_TXCMP_CODE_EXIT)
3513 			break;
3514 		if (ret == MLX5_TXCMP_CODE_ERROR) {
3515 			txq->stats.oerrors++;
3516 			break;
3517 		}
3518 		if (MLX5_TXOFF_CONFIG(MULTI) &&
3519 		    ret == MLX5_TXCMP_CODE_MULTI) {
3520 			/*
3521 			 * The multi-segment packet was
3522 			 * encountered in the array.
3523 			 */
3524 			goto enter_send_multi;
3525 		}
3526 		if (MLX5_TXOFF_CONFIG(TSO) &&
3527 		    ret == MLX5_TXCMP_CODE_TSO) {
3528 			/*
3529 			 * The single-segment TSO packet was
3530 			 * encountered in the array.
3531 			 */
3532 			goto enter_send_tso;
3533 		}
3534 		/* We must not get here. Something is going wrong. */
3535 		MLX5_ASSERT(false);
3536 		txq->stats.oerrors++;
3537 		break;
3538 	}
3539 	/*
3540 	 * Main Tx loop is completed, do the rest:
3541 	 * - set completion request if thresholds are reached
3542 	 * - doorbell the hardware
3543 	 * - copy the rest of mbufs to elts (if any)
3544 	 */
3545 	MLX5_ASSERT(MLX5_TXOFF_CONFIG(INLINE) ||
3546 		    loc.pkts_sent >= loc.pkts_copy);
3547 	/* Take a shortcut if nothing is sent. */
3548 	if (unlikely(loc.pkts_sent == loc.pkts_loop))
3549 		goto burst_exit;
3550 	/* Request CQE generation if limits are reached. */
3551 	mlx5_tx_request_completion(txq, &loc, olx);
3552 	/*
3553 	 * Ring QP doorbell immediately after WQE building completion
3554 	 * to improve latencies. The pure software related data treatment
3555 	 * can be completed after doorbell. Tx CQEs for this SQ are
3556 	 * processed in this thread only by the polling.
3557 	 *
3558 	 * The rdma core library can map doorbell register in two ways,
3559 	 * depending on the environment variable "MLX5_SHUT_UP_BF":
3560 	 *
3561 	 * - as regular cached memory, the variable is either missing or
3562 	 *   set to zero. This type of mapping may cause the significant
3563 	 *   doorbell register writing latency and requires explicit memory
3564 	 *   write barrier to mitigate this issue and prevent write combining.
3565 	 *
3566 	 * - as non-cached memory, the variable is present and set to not "0"
3567 	 *   value. This type of mapping may cause performance impact under
3568 	 *   heavy loading conditions but the explicit write memory barrier is
3569 	 *   not required and it may improve core performance.
3570 	 *
3571 	 * - the legacy behaviour (prior 19.08 release) was to use some
3572 	 *   heuristics to decide whether write memory barrier should
3573 	 *   be performed. This behavior is supported with specifying
3574 	 *   tx_db_nc=2, write barrier is skipped if application provides
3575 	 *   the full recommended burst of packets, it supposes the next
3576 	 *   packets are coming and the write barrier will be issued on
3577 	 *   the next burst (after descriptor writing, at least).
3578 	 */
3579 	mlx5_doorbell_ring(mlx5_tx_bfreg(txq),
3580 			   *(volatile uint64_t *)loc.wqe_last, txq->wqe_ci,
3581 			   txq->qp_db, !txq->db_nc &&
3582 			   (!txq->db_heu || pkts_n % MLX5_TX_DEFAULT_BURST));
3583 	/* Not all of the mbufs may be stored into elts yet. */
3584 	part = MLX5_TXOFF_CONFIG(INLINE) ? 0 : loc.pkts_sent - loc.pkts_copy;
3585 	if (!MLX5_TXOFF_CONFIG(INLINE) && part) {
3586 		/*
3587 		 * There are some single-segment mbufs not stored in elts.
3588 		 * It can be only if the last packet was single-segment.
3589 		 * The copying is gathered into one place due to it is
3590 		 * a good opportunity to optimize that with SIMD.
3591 		 * Unfortunately if inlining is enabled the gaps in pointer
3592 		 * array may happen due to early freeing of the inlined mbufs.
3593 		 */
3594 		mlx5_tx_copy_elts(txq, pkts + loc.pkts_copy, part, olx);
3595 		loc.pkts_copy = loc.pkts_sent;
3596 	}
3597 	MLX5_ASSERT(txq->elts_s >= (uint16_t)(txq->elts_head - txq->elts_tail));
3598 	MLX5_ASSERT(txq->wqe_s >= (uint16_t)(txq->wqe_ci - txq->wqe_pi));
3599 	if (pkts_n > loc.pkts_sent) {
3600 		/*
3601 		 * If burst size is large there might be no enough CQE
3602 		 * fetched from completion queue and no enough resources
3603 		 * freed to send all the packets.
3604 		 */
3605 		goto send_loop;
3606 	}
3607 burst_exit:
3608 #ifdef MLX5_PMD_SOFT_COUNTERS
3609 	/* Increment sent packets counter. */
3610 	txq->stats.opackets += loc.pkts_sent;
3611 #endif
3612 	if (MLX5_TXOFF_CONFIG(INLINE) && loc.mbuf_free)
3613 		__mlx5_tx_free_mbuf(txq, pkts, loc.mbuf_free, olx);
3614 	return loc.pkts_sent;
3615 }
3616 
3617 #endif /* RTE_PMD_MLX5_TX_H_ */
3618