xref: /dpdk/drivers/net/sfc/sfc_tx.c (revision f8dbaebbf1c9efcbb2e2354b341ed62175466a57)
1 /* SPDX-License-Identifier: BSD-3-Clause
2  *
3  * Copyright(c) 2019-2021 Xilinx, Inc.
4  * Copyright(c) 2016-2019 Solarflare Communications Inc.
5  *
6  * This software was jointly developed between OKTET Labs (under contract
7  * for Solarflare) and Solarflare Communications, Inc.
8  */
9 
10 #include "sfc.h"
11 #include "sfc_debug.h"
12 #include "sfc_log.h"
13 #include "sfc_ev.h"
14 #include "sfc_tx.h"
15 #include "sfc_tweak.h"
16 #include "sfc_kvargs.h"
17 
18 /*
19  * Maximum number of TX queue flush attempts in case of
20  * failure or flush timeout
21  */
22 #define SFC_TX_QFLUSH_ATTEMPTS		(3)
23 
24 /*
25  * Time to wait between event queue polling attempts when waiting for TX
26  * queue flush done or flush failed events
27  */
28 #define SFC_TX_QFLUSH_POLL_WAIT_MS	(1)
29 
30 /*
31  * Maximum number of event queue polling attempts when waiting for TX queue
32  * flush done or flush failed events; it defines TX queue flush attempt timeout
33  * together with SFC_TX_QFLUSH_POLL_WAIT_MS
34  */
35 #define SFC_TX_QFLUSH_POLL_ATTEMPTS	(2000)
36 
37 struct sfc_txq_info *
38 sfc_txq_info_by_ethdev_qid(struct sfc_adapter_shared *sas,
39 			   sfc_ethdev_qid_t ethdev_qid)
40 {
41 	sfc_sw_index_t sw_index;
42 
43 	SFC_ASSERT((unsigned int)ethdev_qid < sas->ethdev_txq_count);
44 	SFC_ASSERT(ethdev_qid != SFC_ETHDEV_QID_INVALID);
45 
46 	sw_index = sfc_txq_sw_index_by_ethdev_tx_qid(sas, ethdev_qid);
47 	return &sas->txq_info[sw_index];
48 }
49 
50 static uint64_t
51 sfc_tx_get_offload_mask(struct sfc_adapter *sa)
52 {
53 	const efx_nic_cfg_t *encp = efx_nic_cfg_get(sa->nic);
54 	uint64_t no_caps = 0;
55 
56 	if (!encp->enc_hw_tx_insert_vlan_enabled)
57 		no_caps |= RTE_ETH_TX_OFFLOAD_VLAN_INSERT;
58 
59 	if (!encp->enc_tunnel_encapsulations_supported)
60 		no_caps |= RTE_ETH_TX_OFFLOAD_OUTER_IPV4_CKSUM;
61 
62 	if (!sa->tso)
63 		no_caps |= RTE_ETH_TX_OFFLOAD_TCP_TSO;
64 
65 	if (!sa->tso_encap ||
66 	    (encp->enc_tunnel_encapsulations_supported &
67 	     (1u << EFX_TUNNEL_PROTOCOL_VXLAN)) == 0)
68 		no_caps |= RTE_ETH_TX_OFFLOAD_VXLAN_TNL_TSO;
69 
70 	if (!sa->tso_encap ||
71 	    (encp->enc_tunnel_encapsulations_supported &
72 	     (1u << EFX_TUNNEL_PROTOCOL_GENEVE)) == 0)
73 		no_caps |= RTE_ETH_TX_OFFLOAD_GENEVE_TNL_TSO;
74 
75 	return ~no_caps;
76 }
77 
78 uint64_t
79 sfc_tx_get_dev_offload_caps(struct sfc_adapter *sa)
80 {
81 	return sa->priv.dp_tx->dev_offload_capa & sfc_tx_get_offload_mask(sa);
82 }
83 
84 uint64_t
85 sfc_tx_get_queue_offload_caps(struct sfc_adapter *sa)
86 {
87 	return sa->priv.dp_tx->queue_offload_capa & sfc_tx_get_offload_mask(sa);
88 }
89 
90 static int
91 sfc_tx_qcheck_conf(struct sfc_adapter *sa, unsigned int txq_max_fill_level,
92 		   const struct rte_eth_txconf *tx_conf,
93 		   uint64_t offloads)
94 {
95 	int rc = 0;
96 
97 	if (tx_conf->tx_rs_thresh != 0) {
98 		sfc_err(sa, "RS bit in transmit descriptor is not supported");
99 		rc = EINVAL;
100 	}
101 
102 	if (tx_conf->tx_free_thresh > txq_max_fill_level) {
103 		sfc_err(sa,
104 			"TxQ free threshold too large: %u vs maximum %u",
105 			tx_conf->tx_free_thresh, txq_max_fill_level);
106 		rc = EINVAL;
107 	}
108 
109 	if (tx_conf->tx_thresh.pthresh != 0 ||
110 	    tx_conf->tx_thresh.hthresh != 0 ||
111 	    tx_conf->tx_thresh.wthresh != 0) {
112 		sfc_warn(sa,
113 			"prefetch/host/writeback thresholds are not supported");
114 	}
115 
116 	/* We either perform both TCP and UDP offload, or no offload at all */
117 	if (((offloads & RTE_ETH_TX_OFFLOAD_TCP_CKSUM) == 0) !=
118 	    ((offloads & RTE_ETH_TX_OFFLOAD_UDP_CKSUM) == 0)) {
119 		sfc_err(sa, "TCP and UDP offloads can't be set independently");
120 		rc = EINVAL;
121 	}
122 
123 	return rc;
124 }
125 
126 void
127 sfc_tx_qflush_done(struct sfc_txq_info *txq_info)
128 {
129 	txq_info->state |= SFC_TXQ_FLUSHED;
130 	txq_info->state &= ~SFC_TXQ_FLUSHING;
131 }
132 
133 int
134 sfc_tx_qinit(struct sfc_adapter *sa, sfc_sw_index_t sw_index,
135 	     uint16_t nb_tx_desc, unsigned int socket_id,
136 	     const struct rte_eth_txconf *tx_conf)
137 {
138 	struct sfc_adapter_shared * const sas = sfc_sa2shared(sa);
139 	sfc_ethdev_qid_t ethdev_qid;
140 	const efx_nic_cfg_t *encp = efx_nic_cfg_get(sa->nic);
141 	unsigned int txq_entries;
142 	unsigned int evq_entries;
143 	unsigned int txq_max_fill_level;
144 	struct sfc_txq_info *txq_info;
145 	struct sfc_evq *evq;
146 	struct sfc_txq *txq;
147 	int rc = 0;
148 	struct sfc_dp_tx_qcreate_info info;
149 	uint64_t offloads;
150 	struct sfc_dp_tx_hw_limits hw_limits;
151 
152 	ethdev_qid = sfc_ethdev_tx_qid_by_txq_sw_index(sas, sw_index);
153 
154 	sfc_log_init(sa, "TxQ = %d (internal %u)", ethdev_qid, sw_index);
155 
156 	memset(&hw_limits, 0, sizeof(hw_limits));
157 	hw_limits.txq_max_entries = sa->txq_max_entries;
158 	hw_limits.txq_min_entries = sa->txq_min_entries;
159 
160 	rc = sa->priv.dp_tx->qsize_up_rings(nb_tx_desc, &hw_limits,
161 					    &txq_entries, &evq_entries,
162 					    &txq_max_fill_level);
163 	if (rc != 0)
164 		goto fail_size_up_rings;
165 	SFC_ASSERT(txq_entries >= sa->txq_min_entries);
166 	SFC_ASSERT(txq_entries <= sa->txq_max_entries);
167 	SFC_ASSERT(txq_entries >= nb_tx_desc);
168 	SFC_ASSERT(txq_max_fill_level <= nb_tx_desc);
169 
170 	offloads = tx_conf->offloads;
171 	/* Add device level Tx offloads if the queue is an ethdev Tx queue */
172 	if (ethdev_qid != SFC_ETHDEV_QID_INVALID)
173 		offloads |= sa->eth_dev->data->dev_conf.txmode.offloads;
174 
175 	rc = sfc_tx_qcheck_conf(sa, txq_max_fill_level, tx_conf, offloads);
176 	if (rc != 0)
177 		goto fail_bad_conf;
178 
179 	SFC_ASSERT(sw_index < sfc_sa2shared(sa)->txq_count);
180 	txq_info = &sfc_sa2shared(sa)->txq_info[sw_index];
181 
182 	txq_info->entries = txq_entries;
183 
184 	rc = sfc_ev_qinit(sa, SFC_EVQ_TYPE_TX, sw_index,
185 			  evq_entries, socket_id, &evq);
186 	if (rc != 0)
187 		goto fail_ev_qinit;
188 
189 	txq = &sa->txq_ctrl[sw_index];
190 	txq->hw_index = sw_index;
191 	txq->evq = evq;
192 	txq_info->free_thresh =
193 		(tx_conf->tx_free_thresh) ? tx_conf->tx_free_thresh :
194 		SFC_TX_DEFAULT_FREE_THRESH;
195 	txq_info->offloads = offloads;
196 
197 	rc = sfc_dma_alloc(sa, "txq", sw_index, EFX_NIC_DMA_ADDR_TX_RING,
198 			   efx_txq_size(sa->nic, txq_info->entries),
199 			   socket_id, &txq->mem);
200 	if (rc != 0)
201 		goto fail_dma_alloc;
202 
203 	memset(&info, 0, sizeof(info));
204 	info.max_fill_level = txq_max_fill_level;
205 	info.free_thresh = txq_info->free_thresh;
206 	info.offloads = offloads;
207 	info.txq_entries = txq_info->entries;
208 	info.dma_desc_size_max = encp->enc_tx_dma_desc_size_max;
209 	info.txq_hw_ring = txq->mem.esm_base;
210 	info.evq_entries = evq_entries;
211 	info.evq_hw_ring = evq->mem.esm_base;
212 	info.hw_index = txq->hw_index;
213 	info.mem_bar = sa->mem_bar.esb_base;
214 	info.vi_window_shift = encp->enc_vi_window_shift;
215 	info.tso_tcp_header_offset_limit =
216 		encp->enc_tx_tso_tcp_header_offset_limit;
217 	info.tso_max_nb_header_descs =
218 		RTE_MIN(encp->enc_tx_tso_max_header_ndescs,
219 			(uint32_t)UINT16_MAX);
220 	info.tso_max_header_len =
221 		RTE_MIN(encp->enc_tx_tso_max_header_length,
222 			(uint32_t)UINT16_MAX);
223 	info.tso_max_nb_payload_descs =
224 		RTE_MIN(encp->enc_tx_tso_max_payload_ndescs,
225 			(uint32_t)UINT16_MAX);
226 	info.tso_max_payload_len = encp->enc_tx_tso_max_payload_length;
227 	info.tso_max_nb_outgoing_frames = encp->enc_tx_tso_max_nframes;
228 
229 	info.nic_dma_info = &sas->nic_dma_info;
230 
231 	rc = sa->priv.dp_tx->qcreate(sa->eth_dev->data->port_id, sw_index,
232 				     &RTE_ETH_DEV_TO_PCI(sa->eth_dev)->addr,
233 				     socket_id, &info, &txq_info->dp);
234 	if (rc != 0)
235 		goto fail_dp_tx_qinit;
236 
237 	evq->dp_txq = txq_info->dp;
238 
239 	txq_info->state = SFC_TXQ_INITIALIZED;
240 
241 	txq_info->deferred_start = (tx_conf->tx_deferred_start != 0);
242 
243 	return 0;
244 
245 fail_dp_tx_qinit:
246 	sfc_dma_free(sa, &txq->mem);
247 
248 fail_dma_alloc:
249 	sfc_ev_qfini(evq);
250 
251 fail_ev_qinit:
252 	txq_info->entries = 0;
253 
254 fail_bad_conf:
255 fail_size_up_rings:
256 	sfc_log_init(sa, "failed (TxQ = %d (internal %u), rc = %d)", ethdev_qid,
257 		     sw_index, rc);
258 	return rc;
259 }
260 
261 void
262 sfc_tx_qfini(struct sfc_adapter *sa, sfc_sw_index_t sw_index)
263 {
264 	struct sfc_adapter_shared * const sas = sfc_sa2shared(sa);
265 	sfc_ethdev_qid_t ethdev_qid;
266 	struct sfc_txq_info *txq_info;
267 	struct sfc_txq *txq;
268 
269 	ethdev_qid = sfc_ethdev_tx_qid_by_txq_sw_index(sas, sw_index);
270 
271 	sfc_log_init(sa, "TxQ = %d (internal %u)", ethdev_qid, sw_index);
272 
273 	SFC_ASSERT(sw_index < sfc_sa2shared(sa)->txq_count);
274 	if (ethdev_qid != SFC_ETHDEV_QID_INVALID)
275 		sa->eth_dev->data->tx_queues[ethdev_qid] = NULL;
276 
277 	txq_info = &sfc_sa2shared(sa)->txq_info[sw_index];
278 
279 	SFC_ASSERT(txq_info->state == SFC_TXQ_INITIALIZED);
280 
281 	sa->priv.dp_tx->qdestroy(txq_info->dp);
282 	txq_info->dp = NULL;
283 
284 	txq_info->state &= ~SFC_TXQ_INITIALIZED;
285 	txq_info->entries = 0;
286 
287 	txq = &sa->txq_ctrl[sw_index];
288 
289 	sfc_dma_free(sa, &txq->mem);
290 
291 	sfc_ev_qfini(txq->evq);
292 	txq->evq = NULL;
293 }
294 
295 int
296 sfc_tx_qinit_info(struct sfc_adapter *sa, sfc_sw_index_t sw_index)
297 {
298 	struct sfc_adapter_shared * const sas = sfc_sa2shared(sa);
299 	sfc_ethdev_qid_t ethdev_qid;
300 
301 	ethdev_qid = sfc_ethdev_tx_qid_by_txq_sw_index(sas, sw_index);
302 
303 	sfc_log_init(sa, "TxQ = %d (internal %u)", ethdev_qid, sw_index);
304 
305 	return 0;
306 }
307 
308 static int
309 sfc_tx_check_mode(struct sfc_adapter *sa, const struct rte_eth_txmode *txmode)
310 {
311 	int rc = 0;
312 
313 	switch (txmode->mq_mode) {
314 	case RTE_ETH_MQ_TX_NONE:
315 		break;
316 	default:
317 		sfc_err(sa, "Tx multi-queue mode %u not supported",
318 			txmode->mq_mode);
319 		rc = EINVAL;
320 	}
321 
322 	/*
323 	 * These features are claimed to be i40e-specific,
324 	 * but it does make sense to double-check their absence
325 	 */
326 	if (txmode->hw_vlan_reject_tagged) {
327 		sfc_err(sa, "Rejecting tagged packets not supported");
328 		rc = EINVAL;
329 	}
330 
331 	if (txmode->hw_vlan_reject_untagged) {
332 		sfc_err(sa, "Rejecting untagged packets not supported");
333 		rc = EINVAL;
334 	}
335 
336 	if (txmode->hw_vlan_insert_pvid) {
337 		sfc_err(sa, "Port-based VLAN insertion not supported");
338 		rc = EINVAL;
339 	}
340 
341 	return rc;
342 }
343 
344 /**
345  * Destroy excess queues that are no longer needed after reconfiguration
346  * or complete close.
347  */
348 static void
349 sfc_tx_fini_queues(struct sfc_adapter *sa, unsigned int nb_tx_queues)
350 {
351 	struct sfc_adapter_shared * const sas = sfc_sa2shared(sa);
352 	sfc_sw_index_t sw_index;
353 	sfc_ethdev_qid_t ethdev_qid;
354 
355 	SFC_ASSERT(nb_tx_queues <= sas->ethdev_txq_count);
356 
357 	/*
358 	 * Finalize only ethdev queues since other ones are finalized only
359 	 * on device close and they may require additional deinitializaton.
360 	 */
361 	ethdev_qid = sas->ethdev_txq_count;
362 	while (--ethdev_qid >= (int)nb_tx_queues) {
363 		struct sfc_txq_info *txq_info;
364 
365 		sw_index = sfc_txq_sw_index_by_ethdev_tx_qid(sas, ethdev_qid);
366 		txq_info = sfc_txq_info_by_ethdev_qid(sas, ethdev_qid);
367 		if (txq_info->state & SFC_TXQ_INITIALIZED)
368 			sfc_tx_qfini(sa, sw_index);
369 	}
370 
371 	sas->ethdev_txq_count = nb_tx_queues;
372 }
373 
374 int
375 sfc_tx_configure(struct sfc_adapter *sa)
376 {
377 	struct sfc_adapter_shared * const sas = sfc_sa2shared(sa);
378 	const efx_nic_cfg_t *encp = efx_nic_cfg_get(sa->nic);
379 	const struct rte_eth_conf *dev_conf = &sa->eth_dev->data->dev_conf;
380 	const unsigned int nb_tx_queues = sa->eth_dev->data->nb_tx_queues;
381 	const unsigned int nb_rsvd_tx_queues = sfc_nb_txq_reserved(sas);
382 	const unsigned int nb_txq_total = nb_tx_queues + nb_rsvd_tx_queues;
383 	bool reconfigure;
384 	int rc = 0;
385 
386 	sfc_log_init(sa, "nb_tx_queues=%u (old %u)",
387 		     nb_tx_queues, sas->ethdev_txq_count);
388 
389 	/*
390 	 * The datapath implementation assumes absence of boundary
391 	 * limits on Tx DMA descriptors. Addition of these checks on
392 	 * datapath would simply make the datapath slower.
393 	 */
394 	if (encp->enc_tx_dma_desc_boundary != 0) {
395 		rc = ENOTSUP;
396 		goto fail_tx_dma_desc_boundary;
397 	}
398 
399 	rc = sfc_tx_check_mode(sa, &dev_conf->txmode);
400 	if (rc != 0)
401 		goto fail_check_mode;
402 
403 	if (nb_txq_total == sas->txq_count)
404 		goto done;
405 
406 	if (sas->txq_info == NULL) {
407 		reconfigure = false;
408 		sas->txq_info = rte_calloc_socket("sfc-txqs", nb_txq_total,
409 						  sizeof(sas->txq_info[0]), 0,
410 						  sa->socket_id);
411 		if (sas->txq_info == NULL)
412 			goto fail_txqs_alloc;
413 
414 		/*
415 		 * Allocate primary process only TxQ control from heap
416 		 * since it should not be shared.
417 		 */
418 		rc = ENOMEM;
419 		sa->txq_ctrl = calloc(nb_txq_total, sizeof(sa->txq_ctrl[0]));
420 		if (sa->txq_ctrl == NULL)
421 			goto fail_txqs_ctrl_alloc;
422 	} else {
423 		struct sfc_txq_info *new_txq_info;
424 		struct sfc_txq *new_txq_ctrl;
425 
426 		reconfigure = true;
427 
428 		if (nb_tx_queues < sas->ethdev_txq_count)
429 			sfc_tx_fini_queues(sa, nb_tx_queues);
430 
431 		new_txq_info =
432 			rte_realloc(sas->txq_info,
433 				    nb_txq_total * sizeof(sas->txq_info[0]), 0);
434 		if (new_txq_info == NULL && nb_txq_total > 0)
435 			goto fail_txqs_realloc;
436 
437 		new_txq_ctrl = realloc(sa->txq_ctrl,
438 				       nb_txq_total * sizeof(sa->txq_ctrl[0]));
439 		if (new_txq_ctrl == NULL && nb_txq_total > 0)
440 			goto fail_txqs_ctrl_realloc;
441 
442 		sas->txq_info = new_txq_info;
443 		sa->txq_ctrl = new_txq_ctrl;
444 		if (nb_txq_total > sas->txq_count) {
445 			memset(&sas->txq_info[sas->txq_count], 0,
446 			       (nb_txq_total - sas->txq_count) *
447 			       sizeof(sas->txq_info[0]));
448 			memset(&sa->txq_ctrl[sas->txq_count], 0,
449 			       (nb_txq_total - sas->txq_count) *
450 			       sizeof(sa->txq_ctrl[0]));
451 		}
452 	}
453 
454 	while (sas->ethdev_txq_count < nb_tx_queues) {
455 		sfc_sw_index_t sw_index;
456 
457 		sw_index = sfc_txq_sw_index_by_ethdev_tx_qid(sas,
458 				sas->ethdev_txq_count);
459 		rc = sfc_tx_qinit_info(sa, sw_index);
460 		if (rc != 0)
461 			goto fail_tx_qinit_info;
462 
463 		sas->ethdev_txq_count++;
464 	}
465 
466 	sas->txq_count = sas->ethdev_txq_count + nb_rsvd_tx_queues;
467 
468 	if (!reconfigure) {
469 		rc = sfc_repr_proxy_txq_init(sa);
470 		if (rc != 0)
471 			goto fail_repr_proxy_txq_init;
472 	}
473 
474 done:
475 	return 0;
476 
477 fail_repr_proxy_txq_init:
478 fail_tx_qinit_info:
479 fail_txqs_ctrl_realloc:
480 fail_txqs_realloc:
481 fail_txqs_ctrl_alloc:
482 fail_txqs_alloc:
483 	sfc_tx_close(sa);
484 
485 fail_check_mode:
486 fail_tx_dma_desc_boundary:
487 	sfc_log_init(sa, "failed (rc = %d)", rc);
488 	return rc;
489 }
490 
491 void
492 sfc_tx_close(struct sfc_adapter *sa)
493 {
494 	sfc_tx_fini_queues(sa, 0);
495 	sfc_repr_proxy_txq_fini(sa);
496 
497 	free(sa->txq_ctrl);
498 	sa->txq_ctrl = NULL;
499 
500 	rte_free(sfc_sa2shared(sa)->txq_info);
501 	sfc_sa2shared(sa)->txq_info = NULL;
502 }
503 
504 int
505 sfc_tx_qstart(struct sfc_adapter *sa, sfc_sw_index_t sw_index)
506 {
507 	struct sfc_adapter_shared * const sas = sfc_sa2shared(sa);
508 	sfc_ethdev_qid_t ethdev_qid;
509 	uint64_t offloads_supported = sfc_tx_get_dev_offload_caps(sa) |
510 				      sfc_tx_get_queue_offload_caps(sa);
511 	struct sfc_txq_info *txq_info;
512 	struct sfc_txq *txq;
513 	struct sfc_evq *evq;
514 	uint16_t flags = 0;
515 	unsigned int desc_index;
516 	int rc = 0;
517 
518 	ethdev_qid = sfc_ethdev_tx_qid_by_txq_sw_index(sas, sw_index);
519 
520 	sfc_log_init(sa, "TxQ = %d (internal %u)", ethdev_qid, sw_index);
521 
522 	SFC_ASSERT(sw_index < sas->txq_count);
523 	txq_info = &sas->txq_info[sw_index];
524 
525 	SFC_ASSERT(txq_info->state == SFC_TXQ_INITIALIZED);
526 
527 	txq = &sa->txq_ctrl[sw_index];
528 	evq = txq->evq;
529 
530 	rc = sfc_ev_qstart(evq, sfc_evq_sw_index_by_txq_sw_index(sa, sw_index));
531 	if (rc != 0)
532 		goto fail_ev_qstart;
533 
534 	if (txq_info->offloads & RTE_ETH_TX_OFFLOAD_IPV4_CKSUM)
535 		flags |= EFX_TXQ_CKSUM_IPV4;
536 
537 	if (txq_info->offloads & RTE_ETH_TX_OFFLOAD_OUTER_IPV4_CKSUM)
538 		flags |= EFX_TXQ_CKSUM_INNER_IPV4;
539 
540 	if ((txq_info->offloads & RTE_ETH_TX_OFFLOAD_TCP_CKSUM) ||
541 	    (txq_info->offloads & RTE_ETH_TX_OFFLOAD_UDP_CKSUM)) {
542 		flags |= EFX_TXQ_CKSUM_TCPUDP;
543 
544 		if (offloads_supported & RTE_ETH_TX_OFFLOAD_OUTER_IPV4_CKSUM)
545 			flags |= EFX_TXQ_CKSUM_INNER_TCPUDP;
546 	}
547 
548 	if (txq_info->offloads & (RTE_ETH_TX_OFFLOAD_TCP_TSO |
549 				  RTE_ETH_TX_OFFLOAD_VXLAN_TNL_TSO |
550 				  RTE_ETH_TX_OFFLOAD_GENEVE_TNL_TSO))
551 		flags |= EFX_TXQ_FATSOV2;
552 
553 	rc = efx_tx_qcreate(sa->nic, txq->hw_index, 0, &txq->mem,
554 			    txq_info->entries, 0 /* not used on EF10 */,
555 			    flags, evq->common,
556 			    &txq->common, &desc_index);
557 	if (rc != 0) {
558 		if (sa->tso && (rc == ENOSPC))
559 			sfc_err(sa, "ran out of TSO contexts");
560 
561 		goto fail_tx_qcreate;
562 	}
563 
564 	efx_tx_qenable(txq->common);
565 
566 	txq_info->state |= SFC_TXQ_STARTED;
567 
568 	rc = sa->priv.dp_tx->qstart(txq_info->dp, evq->read_ptr, desc_index);
569 	if (rc != 0)
570 		goto fail_dp_qstart;
571 
572 	if (ethdev_qid != SFC_ETHDEV_QID_INVALID) {
573 		struct rte_eth_dev_data *dev_data;
574 
575 		/*
576 		 * It sems to be used by DPDK for debug purposes only
577 		 * ('rte_ether').
578 		 */
579 		dev_data = sa->eth_dev->data;
580 		dev_data->tx_queue_state[ethdev_qid] =
581 			RTE_ETH_QUEUE_STATE_STARTED;
582 	}
583 
584 	return 0;
585 
586 fail_dp_qstart:
587 	txq_info->state = SFC_TXQ_INITIALIZED;
588 	efx_tx_qdestroy(txq->common);
589 
590 fail_tx_qcreate:
591 	sfc_ev_qstop(evq);
592 
593 fail_ev_qstart:
594 	return rc;
595 }
596 
597 void
598 sfc_tx_qstop(struct sfc_adapter *sa, sfc_sw_index_t sw_index)
599 {
600 	struct sfc_adapter_shared * const sas = sfc_sa2shared(sa);
601 	sfc_ethdev_qid_t ethdev_qid;
602 	struct sfc_txq_info *txq_info;
603 	struct sfc_txq *txq;
604 	unsigned int retry_count;
605 	unsigned int wait_count;
606 	int rc;
607 
608 	ethdev_qid = sfc_ethdev_tx_qid_by_txq_sw_index(sas, sw_index);
609 
610 	sfc_log_init(sa, "TxQ = %d (internal %u)", ethdev_qid, sw_index);
611 
612 	SFC_ASSERT(sw_index < sas->txq_count);
613 	txq_info = &sas->txq_info[sw_index];
614 
615 	if (txq_info->state == SFC_TXQ_INITIALIZED)
616 		return;
617 
618 	SFC_ASSERT(txq_info->state & SFC_TXQ_STARTED);
619 
620 	txq = &sa->txq_ctrl[sw_index];
621 	sa->priv.dp_tx->qstop(txq_info->dp, &txq->evq->read_ptr);
622 
623 	/*
624 	 * Retry TX queue flushing in case of flush failed or
625 	 * timeout; in the worst case it can delay for 6 seconds
626 	 */
627 	for (retry_count = 0;
628 	     ((txq_info->state & SFC_TXQ_FLUSHED) == 0) &&
629 	     (retry_count < SFC_TX_QFLUSH_ATTEMPTS);
630 	     ++retry_count) {
631 		rc = efx_tx_qflush(txq->common);
632 		if (rc != 0) {
633 			txq_info->state |= (rc == EALREADY) ?
634 				SFC_TXQ_FLUSHED : SFC_TXQ_FLUSH_FAILED;
635 			break;
636 		}
637 
638 		/*
639 		 * Wait for TX queue flush done or flush failed event at least
640 		 * SFC_TX_QFLUSH_POLL_WAIT_MS milliseconds and not more
641 		 * than 2 seconds (SFC_TX_QFLUSH_POLL_WAIT_MS multiplied
642 		 * by SFC_TX_QFLUSH_POLL_ATTEMPTS)
643 		 */
644 		wait_count = 0;
645 		do {
646 			rte_delay_ms(SFC_TX_QFLUSH_POLL_WAIT_MS);
647 			sfc_ev_qpoll(txq->evq);
648 		} while ((txq_info->state & SFC_TXQ_FLUSHING) &&
649 			 wait_count++ < SFC_TX_QFLUSH_POLL_ATTEMPTS);
650 
651 		if (txq_info->state & SFC_TXQ_FLUSHING)
652 			sfc_err(sa, "TxQ %d (internal %u) flush timed out",
653 				ethdev_qid, sw_index);
654 
655 		if (txq_info->state & SFC_TXQ_FLUSHED)
656 			sfc_notice(sa, "TxQ %d (internal %u) flushed",
657 				   ethdev_qid, sw_index);
658 	}
659 
660 	sa->priv.dp_tx->qreap(txq_info->dp);
661 
662 	txq_info->state = SFC_TXQ_INITIALIZED;
663 
664 	efx_tx_qdestroy(txq->common);
665 
666 	sfc_ev_qstop(txq->evq);
667 
668 	if (ethdev_qid != SFC_ETHDEV_QID_INVALID) {
669 		struct rte_eth_dev_data *dev_data;
670 
671 		/*
672 		 * It seems to be used by DPDK for debug purposes only
673 		 * ('rte_ether')
674 		 */
675 		dev_data = sa->eth_dev->data;
676 		dev_data->tx_queue_state[ethdev_qid] =
677 			RTE_ETH_QUEUE_STATE_STOPPED;
678 	}
679 }
680 
681 int
682 sfc_tx_start(struct sfc_adapter *sa)
683 {
684 	struct sfc_adapter_shared * const sas = sfc_sa2shared(sa);
685 	const efx_nic_cfg_t *encp = efx_nic_cfg_get(sa->nic);
686 	sfc_sw_index_t sw_index;
687 	int rc = 0;
688 
689 	sfc_log_init(sa, "txq_count = %u (internal %u)",
690 		     sas->ethdev_txq_count, sas->txq_count);
691 
692 	if (sa->tso) {
693 		if (!encp->enc_fw_assisted_tso_v2_enabled &&
694 		    !encp->enc_tso_v3_enabled) {
695 			sfc_warn(sa, "TSO support was unable to be restored");
696 			sa->tso = B_FALSE;
697 			sa->tso_encap = B_FALSE;
698 		}
699 	}
700 
701 	if (sa->tso_encap && !encp->enc_fw_assisted_tso_v2_encap_enabled &&
702 	    !encp->enc_tso_v3_enabled) {
703 		sfc_warn(sa, "Encapsulated TSO support was unable to be restored");
704 		sa->tso_encap = B_FALSE;
705 	}
706 
707 	rc = efx_tx_init(sa->nic);
708 	if (rc != 0)
709 		goto fail_efx_tx_init;
710 
711 	for (sw_index = 0; sw_index < sas->txq_count; ++sw_index) {
712 		if (sas->txq_info[sw_index].state == SFC_TXQ_INITIALIZED &&
713 		    (!(sas->txq_info[sw_index].deferred_start) ||
714 		     sas->txq_info[sw_index].deferred_started)) {
715 			rc = sfc_tx_qstart(sa, sw_index);
716 			if (rc != 0)
717 				goto fail_tx_qstart;
718 		}
719 	}
720 
721 	return 0;
722 
723 fail_tx_qstart:
724 	while (sw_index-- > 0)
725 		sfc_tx_qstop(sa, sw_index);
726 
727 	efx_tx_fini(sa->nic);
728 
729 fail_efx_tx_init:
730 	sfc_log_init(sa, "failed (rc = %d)", rc);
731 	return rc;
732 }
733 
734 void
735 sfc_tx_stop(struct sfc_adapter *sa)
736 {
737 	struct sfc_adapter_shared * const sas = sfc_sa2shared(sa);
738 	sfc_sw_index_t sw_index;
739 
740 	sfc_log_init(sa, "txq_count = %u (internal %u)",
741 		     sas->ethdev_txq_count, sas->txq_count);
742 
743 	sw_index = sas->txq_count;
744 	while (sw_index-- > 0) {
745 		if (sas->txq_info[sw_index].state & SFC_TXQ_STARTED)
746 			sfc_tx_qstop(sa, sw_index);
747 	}
748 
749 	efx_tx_fini(sa->nic);
750 }
751 
752 static void
753 sfc_efx_tx_reap(struct sfc_efx_txq *txq)
754 {
755 	unsigned int completed;
756 
757 	sfc_ev_qpoll(txq->evq);
758 
759 	for (completed = txq->completed;
760 	     completed != txq->pending; completed++) {
761 		struct sfc_efx_tx_sw_desc *txd;
762 
763 		txd = &txq->sw_ring[completed & txq->ptr_mask];
764 
765 		if (txd->mbuf != NULL) {
766 			rte_pktmbuf_free(txd->mbuf);
767 			txd->mbuf = NULL;
768 		}
769 	}
770 
771 	txq->completed = completed;
772 }
773 
774 /*
775  * The function is used to insert or update VLAN tag;
776  * the firmware has state of the firmware tag to insert per TxQ
777  * (controlled by option descriptors), hence, if the tag of the
778  * packet to be sent is different from one remembered by the firmware,
779  * the function will update it
780  */
781 static unsigned int
782 sfc_efx_tx_maybe_insert_tag(struct sfc_efx_txq *txq, struct rte_mbuf *m,
783 			    efx_desc_t **pend)
784 {
785 	uint16_t this_tag = ((m->ol_flags & RTE_MBUF_F_TX_VLAN) ?
786 			     m->vlan_tci : 0);
787 
788 	if (this_tag == txq->hw_vlan_tci)
789 		return 0;
790 
791 	/*
792 	 * The expression inside SFC_ASSERT() is not desired to be checked in
793 	 * a non-debug build because it might be too expensive on the data path
794 	 */
795 	SFC_ASSERT(efx_nic_cfg_get(txq->evq->sa->nic)->enc_hw_tx_insert_vlan_enabled);
796 
797 	efx_tx_qdesc_vlantci_create(txq->common, rte_cpu_to_be_16(this_tag),
798 				    *pend);
799 	(*pend)++;
800 	txq->hw_vlan_tci = this_tag;
801 
802 	return 1;
803 }
804 
805 static uint16_t
806 sfc_efx_prepare_pkts(void *tx_queue, struct rte_mbuf **tx_pkts,
807 		     uint16_t nb_pkts)
808 {
809 	struct sfc_dp_txq *dp_txq = tx_queue;
810 	struct sfc_efx_txq *txq = sfc_efx_txq_by_dp_txq(dp_txq);
811 	const efx_nic_cfg_t *encp = efx_nic_cfg_get(txq->evq->sa->nic);
812 	uint16_t i;
813 
814 	for (i = 0; i < nb_pkts; i++) {
815 		int ret;
816 
817 		/*
818 		 * EFX Tx datapath may require extra VLAN descriptor if VLAN
819 		 * insertion offload is requested regardless the offload
820 		 * requested/supported.
821 		 */
822 		ret = sfc_dp_tx_prepare_pkt(tx_pkts[i], 0, SFC_TSOH_STD_LEN,
823 				encp->enc_tx_tso_tcp_header_offset_limit,
824 				txq->max_fill_level, EFX_TX_FATSOV2_OPT_NDESCS,
825 				1);
826 		if (unlikely(ret != 0)) {
827 			rte_errno = ret;
828 			break;
829 		}
830 	}
831 
832 	return i;
833 }
834 
835 static uint16_t
836 sfc_efx_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
837 {
838 	struct sfc_dp_txq *dp_txq = (struct sfc_dp_txq *)tx_queue;
839 	struct sfc_efx_txq *txq = sfc_efx_txq_by_dp_txq(dp_txq);
840 	unsigned int added = txq->added;
841 	unsigned int pushed = added;
842 	unsigned int pkts_sent = 0;
843 	efx_desc_t *pend = &txq->pend_desc[0];
844 	const unsigned int hard_max_fill = txq->max_fill_level;
845 	const unsigned int soft_max_fill = hard_max_fill - txq->free_thresh;
846 	unsigned int fill_level = added - txq->completed;
847 	boolean_t reap_done;
848 	int rc __rte_unused;
849 	struct rte_mbuf **pktp;
850 
851 	if (unlikely((txq->flags & SFC_EFX_TXQ_FLAG_RUNNING) == 0))
852 		goto done;
853 
854 	/*
855 	 * If insufficient space for a single packet is present,
856 	 * we should reap; otherwise, we shouldn't do that all the time
857 	 * to avoid latency increase
858 	 */
859 	reap_done = (fill_level > soft_max_fill);
860 
861 	if (reap_done) {
862 		sfc_efx_tx_reap(txq);
863 		/*
864 		 * Recalculate fill level since 'txq->completed'
865 		 * might have changed on reap
866 		 */
867 		fill_level = added - txq->completed;
868 	}
869 
870 	for (pkts_sent = 0, pktp = &tx_pkts[0];
871 	     (pkts_sent < nb_pkts) && (fill_level <= soft_max_fill);
872 	     pkts_sent++, pktp++) {
873 		uint16_t		hw_vlan_tci_prev = txq->hw_vlan_tci;
874 		struct rte_mbuf		*m_seg = *pktp;
875 		size_t			pkt_len = m_seg->pkt_len;
876 		unsigned int		pkt_descs = 0;
877 		size_t			in_off = 0;
878 
879 		/*
880 		 * Here VLAN TCI is expected to be zero in case if no
881 		 * RTE_ETH_TX_OFFLOAD_VLAN_INSERT capability is advertised;
882 		 * if the calling app ignores the absence of
883 		 * RTE_ETH_TX_OFFLOAD_VLAN_INSERT and pushes VLAN TCI, then
884 		 * TX_ERROR will occur
885 		 */
886 		pkt_descs += sfc_efx_tx_maybe_insert_tag(txq, m_seg, &pend);
887 
888 		if (m_seg->ol_flags & RTE_MBUF_F_TX_TCP_SEG) {
889 			/*
890 			 * We expect correct 'pkt->l[2, 3, 4]_len' values
891 			 * to be set correctly by the caller
892 			 */
893 			if (sfc_efx_tso_do(txq, added, &m_seg, &in_off, &pend,
894 					   &pkt_descs, &pkt_len) != 0) {
895 				/* We may have reached this place if packet
896 				 * header linearization is needed but the
897 				 * header length is greater than
898 				 * SFC_TSOH_STD_LEN
899 				 *
900 				 * We will deceive RTE saying that we have sent
901 				 * the packet, but we will actually drop it.
902 				 * Hence, we should revert 'pend' to the
903 				 * previous state (in case we have added
904 				 * VLAN descriptor) and start processing
905 				 * another one packet. But the original
906 				 * mbuf shouldn't be orphaned
907 				 */
908 				pend -= pkt_descs;
909 				txq->hw_vlan_tci = hw_vlan_tci_prev;
910 
911 				rte_pktmbuf_free(*pktp);
912 
913 				continue;
914 			}
915 
916 			/*
917 			 * We've only added 2 FATSOv2 option descriptors
918 			 * and 1 descriptor for the linearized packet header.
919 			 * The outstanding work will be done in the same manner
920 			 * as for the usual non-TSO path
921 			 */
922 		}
923 
924 		for (; m_seg != NULL; m_seg = m_seg->next) {
925 			efsys_dma_addr_t	next_frag;
926 			size_t			seg_len;
927 
928 			seg_len = m_seg->data_len;
929 			next_frag = rte_mbuf_data_iova(m_seg);
930 
931 			/*
932 			 * If we've started TSO transaction few steps earlier,
933 			 * we'll skip packet header using an offset in the
934 			 * current segment (which has been set to the
935 			 * first one containing payload)
936 			 */
937 			seg_len -= in_off;
938 			next_frag += in_off;
939 			in_off = 0;
940 
941 			do {
942 				efsys_dma_addr_t	frag_addr = next_frag;
943 				size_t			frag_len;
944 
945 				/*
946 				 * It is assumed here that there is no
947 				 * limitation on address boundary
948 				 * crossing by DMA descriptor.
949 				 */
950 				frag_len = MIN(seg_len, txq->dma_desc_size_max);
951 				next_frag += frag_len;
952 				seg_len -= frag_len;
953 				pkt_len -= frag_len;
954 
955 				efx_tx_qdesc_dma_create(txq->common,
956 							frag_addr, frag_len,
957 							(pkt_len == 0),
958 							pend++);
959 
960 				pkt_descs++;
961 			} while (seg_len != 0);
962 		}
963 
964 		added += pkt_descs;
965 
966 		fill_level += pkt_descs;
967 		if (unlikely(fill_level > hard_max_fill)) {
968 			/*
969 			 * Our estimation for maximum number of descriptors
970 			 * required to send a packet seems to be wrong.
971 			 * Try to reap (if we haven't yet).
972 			 */
973 			if (!reap_done) {
974 				sfc_efx_tx_reap(txq);
975 				reap_done = B_TRUE;
976 				fill_level = added - txq->completed;
977 				if (fill_level > hard_max_fill) {
978 					pend -= pkt_descs;
979 					txq->hw_vlan_tci = hw_vlan_tci_prev;
980 					break;
981 				}
982 			} else {
983 				pend -= pkt_descs;
984 				txq->hw_vlan_tci = hw_vlan_tci_prev;
985 				break;
986 			}
987 		}
988 
989 		/* Assign mbuf to the last used desc */
990 		txq->sw_ring[(added - 1) & txq->ptr_mask].mbuf = *pktp;
991 	}
992 
993 	if (likely(pkts_sent > 0)) {
994 		rc = efx_tx_qdesc_post(txq->common, txq->pend_desc,
995 				       pend - &txq->pend_desc[0],
996 				       txq->completed, &txq->added);
997 		SFC_ASSERT(rc == 0);
998 
999 		if (likely(pushed != txq->added)) {
1000 			efx_tx_qpush(txq->common, txq->added, pushed);
1001 			txq->dp.dpq.dbells++;
1002 		}
1003 	}
1004 
1005 #if SFC_TX_XMIT_PKTS_REAP_AT_LEAST_ONCE
1006 	if (!reap_done)
1007 		sfc_efx_tx_reap(txq);
1008 #endif
1009 
1010 done:
1011 	return pkts_sent;
1012 }
1013 
1014 const struct sfc_dp_tx *
1015 sfc_dp_tx_by_dp_txq(const struct sfc_dp_txq *dp_txq)
1016 {
1017 	const struct sfc_dp_queue *dpq = &dp_txq->dpq;
1018 	struct rte_eth_dev *eth_dev;
1019 	struct sfc_adapter_priv *sap;
1020 
1021 	SFC_ASSERT(rte_eth_dev_is_valid_port(dpq->port_id));
1022 	eth_dev = &rte_eth_devices[dpq->port_id];
1023 
1024 	sap = sfc_adapter_priv_by_eth_dev(eth_dev);
1025 
1026 	return sap->dp_tx;
1027 }
1028 
1029 struct sfc_txq_info *
1030 sfc_txq_info_by_dp_txq(const struct sfc_dp_txq *dp_txq)
1031 {
1032 	const struct sfc_dp_queue *dpq = &dp_txq->dpq;
1033 	struct rte_eth_dev *eth_dev;
1034 	struct sfc_adapter_shared *sas;
1035 
1036 	SFC_ASSERT(rte_eth_dev_is_valid_port(dpq->port_id));
1037 	eth_dev = &rte_eth_devices[dpq->port_id];
1038 
1039 	sas = sfc_adapter_shared_by_eth_dev(eth_dev);
1040 
1041 	SFC_ASSERT(dpq->queue_id < sas->txq_count);
1042 	return &sas->txq_info[dpq->queue_id];
1043 }
1044 
1045 struct sfc_txq *
1046 sfc_txq_by_dp_txq(const struct sfc_dp_txq *dp_txq)
1047 {
1048 	const struct sfc_dp_queue *dpq = &dp_txq->dpq;
1049 	struct rte_eth_dev *eth_dev;
1050 	struct sfc_adapter *sa;
1051 
1052 	SFC_ASSERT(rte_eth_dev_is_valid_port(dpq->port_id));
1053 	eth_dev = &rte_eth_devices[dpq->port_id];
1054 
1055 	sa = sfc_adapter_by_eth_dev(eth_dev);
1056 
1057 	SFC_ASSERT(dpq->queue_id < sfc_sa2shared(sa)->txq_count);
1058 	return &sa->txq_ctrl[dpq->queue_id];
1059 }
1060 
1061 static sfc_dp_tx_qsize_up_rings_t sfc_efx_tx_qsize_up_rings;
1062 static int
1063 sfc_efx_tx_qsize_up_rings(uint16_t nb_tx_desc,
1064 			  __rte_unused struct sfc_dp_tx_hw_limits *limits,
1065 			  unsigned int *txq_entries,
1066 			  unsigned int *evq_entries,
1067 			  unsigned int *txq_max_fill_level)
1068 {
1069 	*txq_entries = nb_tx_desc;
1070 	*evq_entries = nb_tx_desc;
1071 	*txq_max_fill_level = EFX_TXQ_LIMIT(*txq_entries);
1072 	return 0;
1073 }
1074 
1075 static sfc_dp_tx_qcreate_t sfc_efx_tx_qcreate;
1076 static int
1077 sfc_efx_tx_qcreate(uint16_t port_id, uint16_t queue_id,
1078 		   const struct rte_pci_addr *pci_addr,
1079 		   int socket_id,
1080 		   const struct sfc_dp_tx_qcreate_info *info,
1081 		   struct sfc_dp_txq **dp_txqp)
1082 {
1083 	struct sfc_efx_txq *txq;
1084 	struct sfc_txq *ctrl_txq;
1085 	int rc;
1086 
1087 	rc = ENOTSUP;
1088 	if (info->nic_dma_info->nb_regions > 0)
1089 		goto fail_nic_dma;
1090 
1091 	rc = ENOMEM;
1092 	txq = rte_zmalloc_socket("sfc-efx-txq", sizeof(*txq),
1093 				 RTE_CACHE_LINE_SIZE, socket_id);
1094 	if (txq == NULL)
1095 		goto fail_txq_alloc;
1096 
1097 	sfc_dp_queue_init(&txq->dp.dpq, port_id, queue_id, pci_addr);
1098 
1099 	rc = ENOMEM;
1100 	txq->pend_desc = rte_calloc_socket("sfc-efx-txq-pend-desc",
1101 					   EFX_TXQ_LIMIT(info->txq_entries),
1102 					   sizeof(*txq->pend_desc), 0,
1103 					   socket_id);
1104 	if (txq->pend_desc == NULL)
1105 		goto fail_pend_desc_alloc;
1106 
1107 	rc = ENOMEM;
1108 	txq->sw_ring = rte_calloc_socket("sfc-efx-txq-sw_ring",
1109 					 info->txq_entries,
1110 					 sizeof(*txq->sw_ring),
1111 					 RTE_CACHE_LINE_SIZE, socket_id);
1112 	if (txq->sw_ring == NULL)
1113 		goto fail_sw_ring_alloc;
1114 
1115 	ctrl_txq = sfc_txq_by_dp_txq(&txq->dp);
1116 	if (ctrl_txq->evq->sa->tso) {
1117 		rc = sfc_efx_tso_alloc_tsoh_objs(txq->sw_ring,
1118 						 info->txq_entries, socket_id);
1119 		if (rc != 0)
1120 			goto fail_alloc_tsoh_objs;
1121 	}
1122 
1123 	txq->evq = ctrl_txq->evq;
1124 	txq->ptr_mask = info->txq_entries - 1;
1125 	txq->max_fill_level = info->max_fill_level;
1126 	txq->free_thresh = info->free_thresh;
1127 	txq->dma_desc_size_max = info->dma_desc_size_max;
1128 
1129 	*dp_txqp = &txq->dp;
1130 	return 0;
1131 
1132 fail_alloc_tsoh_objs:
1133 	rte_free(txq->sw_ring);
1134 
1135 fail_sw_ring_alloc:
1136 	rte_free(txq->pend_desc);
1137 
1138 fail_pend_desc_alloc:
1139 	rte_free(txq);
1140 
1141 fail_txq_alloc:
1142 fail_nic_dma:
1143 	return rc;
1144 }
1145 
1146 static sfc_dp_tx_qdestroy_t sfc_efx_tx_qdestroy;
1147 static void
1148 sfc_efx_tx_qdestroy(struct sfc_dp_txq *dp_txq)
1149 {
1150 	struct sfc_efx_txq *txq = sfc_efx_txq_by_dp_txq(dp_txq);
1151 
1152 	sfc_efx_tso_free_tsoh_objs(txq->sw_ring, txq->ptr_mask + 1);
1153 	rte_free(txq->sw_ring);
1154 	rte_free(txq->pend_desc);
1155 	rte_free(txq);
1156 }
1157 
1158 static sfc_dp_tx_qstart_t sfc_efx_tx_qstart;
1159 static int
1160 sfc_efx_tx_qstart(struct sfc_dp_txq *dp_txq,
1161 		  __rte_unused unsigned int evq_read_ptr,
1162 		  unsigned int txq_desc_index)
1163 {
1164 	/* libefx-based datapath is specific to libefx-based PMD */
1165 	struct sfc_efx_txq *txq = sfc_efx_txq_by_dp_txq(dp_txq);
1166 	struct sfc_txq *ctrl_txq = sfc_txq_by_dp_txq(dp_txq);
1167 
1168 	txq->common = ctrl_txq->common;
1169 
1170 	txq->pending = txq->completed = txq->added = txq_desc_index;
1171 	txq->hw_vlan_tci = 0;
1172 
1173 	txq->flags |= (SFC_EFX_TXQ_FLAG_STARTED | SFC_EFX_TXQ_FLAG_RUNNING);
1174 
1175 	return 0;
1176 }
1177 
1178 static sfc_dp_tx_qstop_t sfc_efx_tx_qstop;
1179 static void
1180 sfc_efx_tx_qstop(struct sfc_dp_txq *dp_txq,
1181 		 __rte_unused unsigned int *evq_read_ptr)
1182 {
1183 	struct sfc_efx_txq *txq = sfc_efx_txq_by_dp_txq(dp_txq);
1184 
1185 	txq->flags &= ~SFC_EFX_TXQ_FLAG_RUNNING;
1186 }
1187 
1188 static sfc_dp_tx_qreap_t sfc_efx_tx_qreap;
1189 static void
1190 sfc_efx_tx_qreap(struct sfc_dp_txq *dp_txq)
1191 {
1192 	struct sfc_efx_txq *txq = sfc_efx_txq_by_dp_txq(dp_txq);
1193 	unsigned int txds;
1194 
1195 	sfc_efx_tx_reap(txq);
1196 
1197 	for (txds = 0; txds <= txq->ptr_mask; txds++) {
1198 		if (txq->sw_ring[txds].mbuf != NULL) {
1199 			rte_pktmbuf_free(txq->sw_ring[txds].mbuf);
1200 			txq->sw_ring[txds].mbuf = NULL;
1201 		}
1202 	}
1203 
1204 	txq->flags &= ~SFC_EFX_TXQ_FLAG_STARTED;
1205 }
1206 
1207 static sfc_dp_tx_qdesc_status_t sfc_efx_tx_qdesc_status;
1208 static int
1209 sfc_efx_tx_qdesc_status(struct sfc_dp_txq *dp_txq, uint16_t offset)
1210 {
1211 	struct sfc_efx_txq *txq = sfc_efx_txq_by_dp_txq(dp_txq);
1212 
1213 	if (unlikely(offset > txq->ptr_mask))
1214 		return -EINVAL;
1215 
1216 	if (unlikely(offset >= txq->max_fill_level))
1217 		return RTE_ETH_TX_DESC_UNAVAIL;
1218 
1219 	/*
1220 	 * Poll EvQ to derive up-to-date 'txq->pending' figure;
1221 	 * it is required for the queue to be running, but the
1222 	 * check is omitted because API design assumes that it
1223 	 * is the duty of the caller to satisfy all conditions
1224 	 */
1225 	SFC_ASSERT((txq->flags & SFC_EFX_TXQ_FLAG_RUNNING) ==
1226 		   SFC_EFX_TXQ_FLAG_RUNNING);
1227 	sfc_ev_qpoll(txq->evq);
1228 
1229 	/*
1230 	 * Ring tail is 'txq->pending', and although descriptors
1231 	 * between 'txq->completed' and 'txq->pending' are still
1232 	 * in use by the driver, they should be reported as DONE
1233 	 */
1234 	if (unlikely(offset < (txq->added - txq->pending)))
1235 		return RTE_ETH_TX_DESC_FULL;
1236 
1237 	/*
1238 	 * There is no separate return value for unused descriptors;
1239 	 * the latter will be reported as DONE because genuine DONE
1240 	 * descriptors will be freed anyway in SW on the next burst
1241 	 */
1242 	return RTE_ETH_TX_DESC_DONE;
1243 }
1244 
1245 struct sfc_dp_tx sfc_efx_tx = {
1246 	.dp = {
1247 		.name		= SFC_KVARG_DATAPATH_EFX,
1248 		.type		= SFC_DP_TX,
1249 		.hw_fw_caps	= SFC_DP_HW_FW_CAP_TX_EFX,
1250 	},
1251 	.features		= 0,
1252 	.dev_offload_capa	= RTE_ETH_TX_OFFLOAD_VLAN_INSERT |
1253 				  RTE_ETH_TX_OFFLOAD_MULTI_SEGS,
1254 	.queue_offload_capa	= RTE_ETH_TX_OFFLOAD_IPV4_CKSUM |
1255 				  RTE_ETH_TX_OFFLOAD_UDP_CKSUM |
1256 				  RTE_ETH_TX_OFFLOAD_TCP_CKSUM |
1257 				  RTE_ETH_TX_OFFLOAD_OUTER_IPV4_CKSUM |
1258 				  RTE_ETH_TX_OFFLOAD_TCP_TSO,
1259 	.qsize_up_rings		= sfc_efx_tx_qsize_up_rings,
1260 	.qcreate		= sfc_efx_tx_qcreate,
1261 	.qdestroy		= sfc_efx_tx_qdestroy,
1262 	.qstart			= sfc_efx_tx_qstart,
1263 	.qstop			= sfc_efx_tx_qstop,
1264 	.qreap			= sfc_efx_tx_qreap,
1265 	.qdesc_status		= sfc_efx_tx_qdesc_status,
1266 	.pkt_prepare		= sfc_efx_prepare_pkts,
1267 	.pkt_burst		= sfc_efx_xmit_pkts,
1268 };
1269