xref: /dpdk/drivers/net/sfc/sfc_tx.c (revision 25d11a86c56d50947af33d0b79ede622809bd8b9)
1 /* SPDX-License-Identifier: BSD-3-Clause
2  *
3  * Copyright (c) 2016-2018 Solarflare Communications Inc.
4  * All rights reserved.
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 uint64_t
38 sfc_tx_get_dev_offload_caps(struct sfc_adapter *sa)
39 {
40 	const efx_nic_cfg_t *encp = efx_nic_cfg_get(sa->nic);
41 	uint64_t caps = 0;
42 
43 	if ((sa->priv.dp_tx->features & SFC_DP_TX_FEAT_VLAN_INSERT) &&
44 	    encp->enc_hw_tx_insert_vlan_enabled)
45 		caps |= DEV_TX_OFFLOAD_VLAN_INSERT;
46 
47 	if (sa->priv.dp_tx->features & SFC_DP_TX_FEAT_MULTI_SEG)
48 		caps |= DEV_TX_OFFLOAD_MULTI_SEGS;
49 
50 	if ((~sa->priv.dp_tx->features & SFC_DP_TX_FEAT_MULTI_POOL) &&
51 	    (~sa->priv.dp_tx->features & SFC_DP_TX_FEAT_REFCNT))
52 		caps |= DEV_TX_OFFLOAD_MBUF_FAST_FREE;
53 
54 	return caps;
55 }
56 
57 uint64_t
58 sfc_tx_get_queue_offload_caps(struct sfc_adapter *sa)
59 {
60 	const efx_nic_cfg_t *encp = efx_nic_cfg_get(sa->nic);
61 	uint64_t caps = 0;
62 
63 	caps |= DEV_TX_OFFLOAD_IPV4_CKSUM;
64 	caps |= DEV_TX_OFFLOAD_UDP_CKSUM;
65 	caps |= DEV_TX_OFFLOAD_TCP_CKSUM;
66 
67 	if (encp->enc_tunnel_encapsulations_supported)
68 		caps |= DEV_TX_OFFLOAD_OUTER_IPV4_CKSUM;
69 
70 	if (sa->tso)
71 		caps |= DEV_TX_OFFLOAD_TCP_TSO;
72 
73 	return caps;
74 }
75 
76 static int
77 sfc_tx_qcheck_conf(struct sfc_adapter *sa, unsigned int txq_max_fill_level,
78 		   const struct rte_eth_txconf *tx_conf,
79 		   uint64_t offloads)
80 {
81 	int rc = 0;
82 
83 	if (tx_conf->tx_rs_thresh != 0) {
84 		sfc_err(sa, "RS bit in transmit descriptor is not supported");
85 		rc = EINVAL;
86 	}
87 
88 	if (tx_conf->tx_free_thresh > txq_max_fill_level) {
89 		sfc_err(sa,
90 			"TxQ free threshold too large: %u vs maximum %u",
91 			tx_conf->tx_free_thresh, txq_max_fill_level);
92 		rc = EINVAL;
93 	}
94 
95 	if (tx_conf->tx_thresh.pthresh != 0 ||
96 	    tx_conf->tx_thresh.hthresh != 0 ||
97 	    tx_conf->tx_thresh.wthresh != 0) {
98 		sfc_warn(sa,
99 			"prefetch/host/writeback thresholds are not supported");
100 	}
101 
102 	/* We either perform both TCP and UDP offload, or no offload at all */
103 	if (((offloads & DEV_TX_OFFLOAD_TCP_CKSUM) == 0) !=
104 	    ((offloads & DEV_TX_OFFLOAD_UDP_CKSUM) == 0)) {
105 		sfc_err(sa, "TCP and UDP offloads can't be set independently");
106 		rc = EINVAL;
107 	}
108 
109 	return rc;
110 }
111 
112 void
113 sfc_tx_qflush_done(struct sfc_txq_info *txq_info)
114 {
115 	txq_info->state |= SFC_TXQ_FLUSHED;
116 	txq_info->state &= ~SFC_TXQ_FLUSHING;
117 }
118 
119 int
120 sfc_tx_qinit(struct sfc_adapter *sa, unsigned int sw_index,
121 	     uint16_t nb_tx_desc, unsigned int socket_id,
122 	     const struct rte_eth_txconf *tx_conf)
123 {
124 	const efx_nic_cfg_t *encp = efx_nic_cfg_get(sa->nic);
125 	unsigned int txq_entries;
126 	unsigned int evq_entries;
127 	unsigned int txq_max_fill_level;
128 	struct sfc_txq_info *txq_info;
129 	struct sfc_evq *evq;
130 	struct sfc_txq *txq;
131 	int rc = 0;
132 	struct sfc_dp_tx_qcreate_info info;
133 	uint64_t offloads;
134 	struct sfc_dp_tx_hw_limits hw_limits;
135 
136 	sfc_log_init(sa, "TxQ = %u", sw_index);
137 
138 	memset(&hw_limits, 0, sizeof(hw_limits));
139 	hw_limits.txq_max_entries = sa->txq_max_entries;
140 	hw_limits.txq_min_entries = sa->txq_min_entries;
141 
142 	rc = sa->priv.dp_tx->qsize_up_rings(nb_tx_desc, &hw_limits,
143 					    &txq_entries, &evq_entries,
144 					    &txq_max_fill_level);
145 	if (rc != 0)
146 		goto fail_size_up_rings;
147 	SFC_ASSERT(txq_entries >= sa->txq_min_entries);
148 	SFC_ASSERT(txq_entries <= sa->txq_max_entries);
149 	SFC_ASSERT(txq_entries >= nb_tx_desc);
150 	SFC_ASSERT(txq_max_fill_level <= nb_tx_desc);
151 
152 	offloads = tx_conf->offloads |
153 		sa->eth_dev->data->dev_conf.txmode.offloads;
154 	rc = sfc_tx_qcheck_conf(sa, txq_max_fill_level, tx_conf, offloads);
155 	if (rc != 0)
156 		goto fail_bad_conf;
157 
158 	SFC_ASSERT(sw_index < sfc_sa2shared(sa)->txq_count);
159 	txq_info = &sfc_sa2shared(sa)->txq_info[sw_index];
160 
161 	txq_info->entries = txq_entries;
162 
163 	rc = sfc_ev_qinit(sa, SFC_EVQ_TYPE_TX, sw_index,
164 			  evq_entries, socket_id, &evq);
165 	if (rc != 0)
166 		goto fail_ev_qinit;
167 
168 	txq = &sa->txq_ctrl[sw_index];
169 	txq->hw_index = sw_index;
170 	txq->evq = evq;
171 	txq_info->free_thresh =
172 		(tx_conf->tx_free_thresh) ? tx_conf->tx_free_thresh :
173 		SFC_TX_DEFAULT_FREE_THRESH;
174 	txq_info->offloads = offloads;
175 
176 	rc = sfc_dma_alloc(sa, "txq", sw_index,
177 			   efx_txq_size(sa->nic, txq_info->entries),
178 			   socket_id, &txq->mem);
179 	if (rc != 0)
180 		goto fail_dma_alloc;
181 
182 	memset(&info, 0, sizeof(info));
183 	info.max_fill_level = txq_max_fill_level;
184 	info.free_thresh = txq_info->free_thresh;
185 	info.offloads = offloads;
186 	info.txq_entries = txq_info->entries;
187 	info.dma_desc_size_max = encp->enc_tx_dma_desc_size_max;
188 	info.txq_hw_ring = txq->mem.esm_base;
189 	info.evq_entries = evq_entries;
190 	info.evq_hw_ring = evq->mem.esm_base;
191 	info.hw_index = txq->hw_index;
192 	info.mem_bar = sa->mem_bar.esb_base;
193 	info.vi_window_shift = encp->enc_vi_window_shift;
194 	info.tso_tcp_header_offset_limit =
195 		encp->enc_tx_tso_tcp_header_offset_limit;
196 
197 	rc = sa->priv.dp_tx->qcreate(sa->eth_dev->data->port_id, sw_index,
198 				     &RTE_ETH_DEV_TO_PCI(sa->eth_dev)->addr,
199 				     socket_id, &info, &txq_info->dp);
200 	if (rc != 0)
201 		goto fail_dp_tx_qinit;
202 
203 	evq->dp_txq = txq_info->dp;
204 
205 	txq_info->state = SFC_TXQ_INITIALIZED;
206 
207 	txq_info->deferred_start = (tx_conf->tx_deferred_start != 0);
208 
209 	return 0;
210 
211 fail_dp_tx_qinit:
212 	sfc_dma_free(sa, &txq->mem);
213 
214 fail_dma_alloc:
215 	sfc_ev_qfini(evq);
216 
217 fail_ev_qinit:
218 	txq_info->entries = 0;
219 
220 fail_bad_conf:
221 fail_size_up_rings:
222 	sfc_log_init(sa, "failed (TxQ = %u, rc = %d)", sw_index, rc);
223 	return rc;
224 }
225 
226 void
227 sfc_tx_qfini(struct sfc_adapter *sa, unsigned int sw_index)
228 {
229 	struct sfc_txq_info *txq_info;
230 	struct sfc_txq *txq;
231 
232 	sfc_log_init(sa, "TxQ = %u", sw_index);
233 
234 	SFC_ASSERT(sw_index < sfc_sa2shared(sa)->txq_count);
235 	sa->eth_dev->data->tx_queues[sw_index] = NULL;
236 
237 	txq_info = &sfc_sa2shared(sa)->txq_info[sw_index];
238 
239 	SFC_ASSERT(txq_info->state == SFC_TXQ_INITIALIZED);
240 
241 	sa->priv.dp_tx->qdestroy(txq_info->dp);
242 	txq_info->dp = NULL;
243 
244 	txq_info->state &= ~SFC_TXQ_INITIALIZED;
245 	txq_info->entries = 0;
246 
247 	txq = &sa->txq_ctrl[sw_index];
248 
249 	sfc_dma_free(sa, &txq->mem);
250 
251 	sfc_ev_qfini(txq->evq);
252 	txq->evq = NULL;
253 }
254 
255 static int
256 sfc_tx_qinit_info(struct sfc_adapter *sa, unsigned int sw_index)
257 {
258 	sfc_log_init(sa, "TxQ = %u", sw_index);
259 
260 	return 0;
261 }
262 
263 static int
264 sfc_tx_check_mode(struct sfc_adapter *sa, const struct rte_eth_txmode *txmode)
265 {
266 	int rc = 0;
267 
268 	switch (txmode->mq_mode) {
269 	case ETH_MQ_TX_NONE:
270 		break;
271 	default:
272 		sfc_err(sa, "Tx multi-queue mode %u not supported",
273 			txmode->mq_mode);
274 		rc = EINVAL;
275 	}
276 
277 	/*
278 	 * These features are claimed to be i40e-specific,
279 	 * but it does make sense to double-check their absence
280 	 */
281 	if (txmode->hw_vlan_reject_tagged) {
282 		sfc_err(sa, "Rejecting tagged packets not supported");
283 		rc = EINVAL;
284 	}
285 
286 	if (txmode->hw_vlan_reject_untagged) {
287 		sfc_err(sa, "Rejecting untagged packets not supported");
288 		rc = EINVAL;
289 	}
290 
291 	if (txmode->hw_vlan_insert_pvid) {
292 		sfc_err(sa, "Port-based VLAN insertion not supported");
293 		rc = EINVAL;
294 	}
295 
296 	return rc;
297 }
298 
299 /**
300  * Destroy excess queues that are no longer needed after reconfiguration
301  * or complete close.
302  */
303 static void
304 sfc_tx_fini_queues(struct sfc_adapter *sa, unsigned int nb_tx_queues)
305 {
306 	struct sfc_adapter_shared * const sas = sfc_sa2shared(sa);
307 	int sw_index;
308 
309 	SFC_ASSERT(nb_tx_queues <= sas->txq_count);
310 
311 	sw_index = sas->txq_count;
312 	while (--sw_index >= (int)nb_tx_queues) {
313 		if (sas->txq_info[sw_index].state & SFC_TXQ_INITIALIZED)
314 			sfc_tx_qfini(sa, sw_index);
315 	}
316 
317 	sas->txq_count = nb_tx_queues;
318 }
319 
320 int
321 sfc_tx_configure(struct sfc_adapter *sa)
322 {
323 	struct sfc_adapter_shared * const sas = sfc_sa2shared(sa);
324 	const efx_nic_cfg_t *encp = efx_nic_cfg_get(sa->nic);
325 	const struct rte_eth_conf *dev_conf = &sa->eth_dev->data->dev_conf;
326 	const unsigned int nb_tx_queues = sa->eth_dev->data->nb_tx_queues;
327 	int rc = 0;
328 
329 	sfc_log_init(sa, "nb_tx_queues=%u (old %u)",
330 		     nb_tx_queues, sas->txq_count);
331 
332 	/*
333 	 * The datapath implementation assumes absence of boundary
334 	 * limits on Tx DMA descriptors. Addition of these checks on
335 	 * datapath would simply make the datapath slower.
336 	 */
337 	if (encp->enc_tx_dma_desc_boundary != 0) {
338 		rc = ENOTSUP;
339 		goto fail_tx_dma_desc_boundary;
340 	}
341 
342 	rc = sfc_tx_check_mode(sa, &dev_conf->txmode);
343 	if (rc != 0)
344 		goto fail_check_mode;
345 
346 	if (nb_tx_queues == sas->txq_count)
347 		goto done;
348 
349 	if (sas->txq_info == NULL) {
350 		sas->txq_info = rte_calloc_socket("sfc-txqs", nb_tx_queues,
351 						  sizeof(sas->txq_info[0]), 0,
352 						  sa->socket_id);
353 		if (sas->txq_info == NULL)
354 			goto fail_txqs_alloc;
355 
356 		/*
357 		 * Allocate primary process only TxQ control from heap
358 		 * since it should not be shared.
359 		 */
360 		rc = ENOMEM;
361 		sa->txq_ctrl = calloc(nb_tx_queues, sizeof(sa->txq_ctrl[0]));
362 		if (sa->txq_ctrl == NULL)
363 			goto fail_txqs_ctrl_alloc;
364 	} else {
365 		struct sfc_txq_info *new_txq_info;
366 		struct sfc_txq *new_txq_ctrl;
367 
368 		if (nb_tx_queues < sas->txq_count)
369 			sfc_tx_fini_queues(sa, nb_tx_queues);
370 
371 		new_txq_info =
372 			rte_realloc(sas->txq_info,
373 				    nb_tx_queues * sizeof(sas->txq_info[0]), 0);
374 		if (new_txq_info == NULL && nb_tx_queues > 0)
375 			goto fail_txqs_realloc;
376 
377 		new_txq_ctrl = realloc(sa->txq_ctrl,
378 				       nb_tx_queues * sizeof(sa->txq_ctrl[0]));
379 		if (new_txq_ctrl == NULL && nb_tx_queues > 0)
380 			goto fail_txqs_ctrl_realloc;
381 
382 		sas->txq_info = new_txq_info;
383 		sa->txq_ctrl = new_txq_ctrl;
384 		if (nb_tx_queues > sas->txq_count) {
385 			memset(&sas->txq_info[sas->txq_count], 0,
386 			       (nb_tx_queues - sas->txq_count) *
387 			       sizeof(sas->txq_info[0]));
388 			memset(&sa->txq_ctrl[sas->txq_count], 0,
389 			       (nb_tx_queues - sas->txq_count) *
390 			       sizeof(sa->txq_ctrl[0]));
391 		}
392 	}
393 
394 	while (sas->txq_count < nb_tx_queues) {
395 		rc = sfc_tx_qinit_info(sa, sas->txq_count);
396 		if (rc != 0)
397 			goto fail_tx_qinit_info;
398 
399 		sas->txq_count++;
400 	}
401 
402 done:
403 	return 0;
404 
405 fail_tx_qinit_info:
406 fail_txqs_ctrl_realloc:
407 fail_txqs_realloc:
408 fail_txqs_ctrl_alloc:
409 fail_txqs_alloc:
410 	sfc_tx_close(sa);
411 
412 fail_check_mode:
413 fail_tx_dma_desc_boundary:
414 	sfc_log_init(sa, "failed (rc = %d)", rc);
415 	return rc;
416 }
417 
418 void
419 sfc_tx_close(struct sfc_adapter *sa)
420 {
421 	sfc_tx_fini_queues(sa, 0);
422 
423 	free(sa->txq_ctrl);
424 	sa->txq_ctrl = NULL;
425 
426 	rte_free(sfc_sa2shared(sa)->txq_info);
427 	sfc_sa2shared(sa)->txq_info = NULL;
428 }
429 
430 int
431 sfc_tx_qstart(struct sfc_adapter *sa, unsigned int sw_index)
432 {
433 	struct sfc_adapter_shared * const sas = sfc_sa2shared(sa);
434 	uint64_t offloads_supported = sfc_tx_get_dev_offload_caps(sa) |
435 				      sfc_tx_get_queue_offload_caps(sa);
436 	struct rte_eth_dev_data *dev_data;
437 	struct sfc_txq_info *txq_info;
438 	struct sfc_txq *txq;
439 	struct sfc_evq *evq;
440 	uint16_t flags = 0;
441 	unsigned int desc_index;
442 	int rc = 0;
443 
444 	sfc_log_init(sa, "TxQ = %u", sw_index);
445 
446 	SFC_ASSERT(sw_index < sas->txq_count);
447 	txq_info = &sas->txq_info[sw_index];
448 
449 	SFC_ASSERT(txq_info->state == SFC_TXQ_INITIALIZED);
450 
451 	txq = &sa->txq_ctrl[sw_index];
452 	evq = txq->evq;
453 
454 	rc = sfc_ev_qstart(evq, sfc_evq_index_by_txq_sw_index(sa, sw_index));
455 	if (rc != 0)
456 		goto fail_ev_qstart;
457 
458 	if (txq_info->offloads & DEV_TX_OFFLOAD_IPV4_CKSUM)
459 		flags |= EFX_TXQ_CKSUM_IPV4;
460 
461 	if (txq_info->offloads & DEV_TX_OFFLOAD_OUTER_IPV4_CKSUM)
462 		flags |= EFX_TXQ_CKSUM_INNER_IPV4;
463 
464 	if ((txq_info->offloads & DEV_TX_OFFLOAD_TCP_CKSUM) ||
465 	    (txq_info->offloads & DEV_TX_OFFLOAD_UDP_CKSUM)) {
466 		flags |= EFX_TXQ_CKSUM_TCPUDP;
467 
468 		if (offloads_supported & DEV_TX_OFFLOAD_OUTER_IPV4_CKSUM)
469 			flags |= EFX_TXQ_CKSUM_INNER_TCPUDP;
470 	}
471 
472 	if (txq_info->offloads & DEV_TX_OFFLOAD_TCP_TSO)
473 		flags |= EFX_TXQ_FATSOV2;
474 
475 	rc = efx_tx_qcreate(sa->nic, txq->hw_index, 0, &txq->mem,
476 			    txq_info->entries, 0 /* not used on EF10 */,
477 			    flags, evq->common,
478 			    &txq->common, &desc_index);
479 	if (rc != 0) {
480 		if (sa->tso && (rc == ENOSPC))
481 			sfc_err(sa, "ran out of TSO contexts");
482 
483 		goto fail_tx_qcreate;
484 	}
485 
486 	efx_tx_qenable(txq->common);
487 
488 	txq_info->state |= SFC_TXQ_STARTED;
489 
490 	rc = sa->priv.dp_tx->qstart(txq_info->dp, evq->read_ptr, desc_index);
491 	if (rc != 0)
492 		goto fail_dp_qstart;
493 
494 	/*
495 	 * It seems to be used by DPDK for debug purposes only ('rte_ether')
496 	 */
497 	dev_data = sa->eth_dev->data;
498 	dev_data->tx_queue_state[sw_index] = RTE_ETH_QUEUE_STATE_STARTED;
499 
500 	return 0;
501 
502 fail_dp_qstart:
503 	txq_info->state = SFC_TXQ_INITIALIZED;
504 	efx_tx_qdestroy(txq->common);
505 
506 fail_tx_qcreate:
507 	sfc_ev_qstop(evq);
508 
509 fail_ev_qstart:
510 	return rc;
511 }
512 
513 void
514 sfc_tx_qstop(struct sfc_adapter *sa, unsigned int sw_index)
515 {
516 	struct sfc_adapter_shared * const sas = sfc_sa2shared(sa);
517 	struct rte_eth_dev_data *dev_data;
518 	struct sfc_txq_info *txq_info;
519 	struct sfc_txq *txq;
520 	unsigned int retry_count;
521 	unsigned int wait_count;
522 	int rc;
523 
524 	sfc_log_init(sa, "TxQ = %u", sw_index);
525 
526 	SFC_ASSERT(sw_index < sas->txq_count);
527 	txq_info = &sas->txq_info[sw_index];
528 
529 	if (txq_info->state == SFC_TXQ_INITIALIZED)
530 		return;
531 
532 	SFC_ASSERT(txq_info->state & SFC_TXQ_STARTED);
533 
534 	txq = &sa->txq_ctrl[sw_index];
535 	sa->priv.dp_tx->qstop(txq_info->dp, &txq->evq->read_ptr);
536 
537 	/*
538 	 * Retry TX queue flushing in case of flush failed or
539 	 * timeout; in the worst case it can delay for 6 seconds
540 	 */
541 	for (retry_count = 0;
542 	     ((txq_info->state & SFC_TXQ_FLUSHED) == 0) &&
543 	     (retry_count < SFC_TX_QFLUSH_ATTEMPTS);
544 	     ++retry_count) {
545 		rc = efx_tx_qflush(txq->common);
546 		if (rc != 0) {
547 			txq_info->state |= (rc == EALREADY) ?
548 				SFC_TXQ_FLUSHED : SFC_TXQ_FLUSH_FAILED;
549 			break;
550 		}
551 
552 		/*
553 		 * Wait for TX queue flush done or flush failed event at least
554 		 * SFC_TX_QFLUSH_POLL_WAIT_MS milliseconds and not more
555 		 * than 2 seconds (SFC_TX_QFLUSH_POLL_WAIT_MS multiplied
556 		 * by SFC_TX_QFLUSH_POLL_ATTEMPTS)
557 		 */
558 		wait_count = 0;
559 		do {
560 			rte_delay_ms(SFC_TX_QFLUSH_POLL_WAIT_MS);
561 			sfc_ev_qpoll(txq->evq);
562 		} while ((txq_info->state & SFC_TXQ_FLUSHING) &&
563 			 wait_count++ < SFC_TX_QFLUSH_POLL_ATTEMPTS);
564 
565 		if (txq_info->state & SFC_TXQ_FLUSHING)
566 			sfc_err(sa, "TxQ %u flush timed out", sw_index);
567 
568 		if (txq_info->state & SFC_TXQ_FLUSHED)
569 			sfc_notice(sa, "TxQ %u flushed", sw_index);
570 	}
571 
572 	sa->priv.dp_tx->qreap(txq_info->dp);
573 
574 	txq_info->state = SFC_TXQ_INITIALIZED;
575 
576 	efx_tx_qdestroy(txq->common);
577 
578 	sfc_ev_qstop(txq->evq);
579 
580 	/*
581 	 * It seems to be used by DPDK for debug purposes only ('rte_ether')
582 	 */
583 	dev_data = sa->eth_dev->data;
584 	dev_data->tx_queue_state[sw_index] = RTE_ETH_QUEUE_STATE_STOPPED;
585 }
586 
587 int
588 sfc_tx_start(struct sfc_adapter *sa)
589 {
590 	struct sfc_adapter_shared * const sas = sfc_sa2shared(sa);
591 	unsigned int sw_index;
592 	int rc = 0;
593 
594 	sfc_log_init(sa, "txq_count = %u", sas->txq_count);
595 
596 	if (sa->tso) {
597 		if (!efx_nic_cfg_get(sa->nic)->enc_fw_assisted_tso_v2_enabled) {
598 			sfc_warn(sa, "TSO support was unable to be restored");
599 			sa->tso = B_FALSE;
600 		}
601 	}
602 
603 	rc = efx_tx_init(sa->nic);
604 	if (rc != 0)
605 		goto fail_efx_tx_init;
606 
607 	for (sw_index = 0; sw_index < sas->txq_count; ++sw_index) {
608 		if (sas->txq_info[sw_index].state == SFC_TXQ_INITIALIZED &&
609 		    (!(sas->txq_info[sw_index].deferred_start) ||
610 		     sas->txq_info[sw_index].deferred_started)) {
611 			rc = sfc_tx_qstart(sa, sw_index);
612 			if (rc != 0)
613 				goto fail_tx_qstart;
614 		}
615 	}
616 
617 	return 0;
618 
619 fail_tx_qstart:
620 	while (sw_index-- > 0)
621 		sfc_tx_qstop(sa, sw_index);
622 
623 	efx_tx_fini(sa->nic);
624 
625 fail_efx_tx_init:
626 	sfc_log_init(sa, "failed (rc = %d)", rc);
627 	return rc;
628 }
629 
630 void
631 sfc_tx_stop(struct sfc_adapter *sa)
632 {
633 	struct sfc_adapter_shared * const sas = sfc_sa2shared(sa);
634 	unsigned int sw_index;
635 
636 	sfc_log_init(sa, "txq_count = %u", sas->txq_count);
637 
638 	sw_index = sas->txq_count;
639 	while (sw_index-- > 0) {
640 		if (sas->txq_info[sw_index].state & SFC_TXQ_STARTED)
641 			sfc_tx_qstop(sa, sw_index);
642 	}
643 
644 	efx_tx_fini(sa->nic);
645 }
646 
647 static void
648 sfc_efx_tx_reap(struct sfc_efx_txq *txq)
649 {
650 	unsigned int completed;
651 
652 	sfc_ev_qpoll(txq->evq);
653 
654 	for (completed = txq->completed;
655 	     completed != txq->pending; completed++) {
656 		struct sfc_efx_tx_sw_desc *txd;
657 
658 		txd = &txq->sw_ring[completed & txq->ptr_mask];
659 
660 		if (txd->mbuf != NULL) {
661 			rte_pktmbuf_free(txd->mbuf);
662 			txd->mbuf = NULL;
663 		}
664 	}
665 
666 	txq->completed = completed;
667 }
668 
669 /*
670  * The function is used to insert or update VLAN tag;
671  * the firmware has state of the firmware tag to insert per TxQ
672  * (controlled by option descriptors), hence, if the tag of the
673  * packet to be sent is different from one remembered by the firmware,
674  * the function will update it
675  */
676 static unsigned int
677 sfc_efx_tx_maybe_insert_tag(struct sfc_efx_txq *txq, struct rte_mbuf *m,
678 			    efx_desc_t **pend)
679 {
680 	uint16_t this_tag = ((m->ol_flags & PKT_TX_VLAN_PKT) ?
681 			     m->vlan_tci : 0);
682 
683 	if (this_tag == txq->hw_vlan_tci)
684 		return 0;
685 
686 	/*
687 	 * The expression inside SFC_ASSERT() is not desired to be checked in
688 	 * a non-debug build because it might be too expensive on the data path
689 	 */
690 	SFC_ASSERT(efx_nic_cfg_get(txq->evq->sa->nic)->enc_hw_tx_insert_vlan_enabled);
691 
692 	efx_tx_qdesc_vlantci_create(txq->common, rte_cpu_to_be_16(this_tag),
693 				    *pend);
694 	(*pend)++;
695 	txq->hw_vlan_tci = this_tag;
696 
697 	return 1;
698 }
699 
700 static uint16_t
701 sfc_efx_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
702 {
703 	struct sfc_dp_txq *dp_txq = (struct sfc_dp_txq *)tx_queue;
704 	struct sfc_efx_txq *txq = sfc_efx_txq_by_dp_txq(dp_txq);
705 	unsigned int added = txq->added;
706 	unsigned int pushed = added;
707 	unsigned int pkts_sent = 0;
708 	efx_desc_t *pend = &txq->pend_desc[0];
709 	const unsigned int hard_max_fill = txq->max_fill_level;
710 	const unsigned int soft_max_fill = hard_max_fill - txq->free_thresh;
711 	unsigned int fill_level = added - txq->completed;
712 	boolean_t reap_done;
713 	int rc __rte_unused;
714 	struct rte_mbuf **pktp;
715 
716 	if (unlikely((txq->flags & SFC_EFX_TXQ_FLAG_RUNNING) == 0))
717 		goto done;
718 
719 	/*
720 	 * If insufficient space for a single packet is present,
721 	 * we should reap; otherwise, we shouldn't do that all the time
722 	 * to avoid latency increase
723 	 */
724 	reap_done = (fill_level > soft_max_fill);
725 
726 	if (reap_done) {
727 		sfc_efx_tx_reap(txq);
728 		/*
729 		 * Recalculate fill level since 'txq->completed'
730 		 * might have changed on reap
731 		 */
732 		fill_level = added - txq->completed;
733 	}
734 
735 	for (pkts_sent = 0, pktp = &tx_pkts[0];
736 	     (pkts_sent < nb_pkts) && (fill_level <= soft_max_fill);
737 	     pkts_sent++, pktp++) {
738 		uint16_t		hw_vlan_tci_prev = txq->hw_vlan_tci;
739 		struct rte_mbuf		*m_seg = *pktp;
740 		size_t			pkt_len = m_seg->pkt_len;
741 		unsigned int		pkt_descs = 0;
742 		size_t			in_off = 0;
743 
744 		/*
745 		 * Here VLAN TCI is expected to be zero in case if no
746 		 * DEV_TX_OFFLOAD_VLAN_INSERT capability is advertised;
747 		 * if the calling app ignores the absence of
748 		 * DEV_TX_OFFLOAD_VLAN_INSERT and pushes VLAN TCI, then
749 		 * TX_ERROR will occur
750 		 */
751 		pkt_descs += sfc_efx_tx_maybe_insert_tag(txq, m_seg, &pend);
752 
753 		if (m_seg->ol_flags & PKT_TX_TCP_SEG) {
754 			/*
755 			 * We expect correct 'pkt->l[2, 3, 4]_len' values
756 			 * to be set correctly by the caller
757 			 */
758 			if (sfc_efx_tso_do(txq, added, &m_seg, &in_off, &pend,
759 					   &pkt_descs, &pkt_len) != 0) {
760 				/* We may have reached this place for
761 				 * one of the following reasons:
762 				 *
763 				 * 1) Packet header length is greater
764 				 *    than SFC_TSOH_STD_LEN
765 				 * 2) TCP header starts at more then
766 				 *    208 bytes into the frame
767 				 *
768 				 * We will deceive RTE saying that we have sent
769 				 * the packet, but we will actually drop it.
770 				 * Hence, we should revert 'pend' to the
771 				 * previous state (in case we have added
772 				 * VLAN descriptor) and start processing
773 				 * another one packet. But the original
774 				 * mbuf shouldn't be orphaned
775 				 */
776 				pend -= pkt_descs;
777 				txq->hw_vlan_tci = hw_vlan_tci_prev;
778 
779 				rte_pktmbuf_free(*pktp);
780 
781 				continue;
782 			}
783 
784 			/*
785 			 * We've only added 2 FATSOv2 option descriptors
786 			 * and 1 descriptor for the linearized packet header.
787 			 * The outstanding work will be done in the same manner
788 			 * as for the usual non-TSO path
789 			 */
790 		}
791 
792 		for (; m_seg != NULL; m_seg = m_seg->next) {
793 			efsys_dma_addr_t	next_frag;
794 			size_t			seg_len;
795 
796 			seg_len = m_seg->data_len;
797 			next_frag = rte_mbuf_data_iova(m_seg);
798 
799 			/*
800 			 * If we've started TSO transaction few steps earlier,
801 			 * we'll skip packet header using an offset in the
802 			 * current segment (which has been set to the
803 			 * first one containing payload)
804 			 */
805 			seg_len -= in_off;
806 			next_frag += in_off;
807 			in_off = 0;
808 
809 			do {
810 				efsys_dma_addr_t	frag_addr = next_frag;
811 				size_t			frag_len;
812 
813 				/*
814 				 * It is assumed here that there is no
815 				 * limitation on address boundary
816 				 * crossing by DMA descriptor.
817 				 */
818 				frag_len = MIN(seg_len, txq->dma_desc_size_max);
819 				next_frag += frag_len;
820 				seg_len -= frag_len;
821 				pkt_len -= frag_len;
822 
823 				efx_tx_qdesc_dma_create(txq->common,
824 							frag_addr, frag_len,
825 							(pkt_len == 0),
826 							pend++);
827 
828 				pkt_descs++;
829 			} while (seg_len != 0);
830 		}
831 
832 		added += pkt_descs;
833 
834 		fill_level += pkt_descs;
835 		if (unlikely(fill_level > hard_max_fill)) {
836 			/*
837 			 * Our estimation for maximum number of descriptors
838 			 * required to send a packet seems to be wrong.
839 			 * Try to reap (if we haven't yet).
840 			 */
841 			if (!reap_done) {
842 				sfc_efx_tx_reap(txq);
843 				reap_done = B_TRUE;
844 				fill_level = added - txq->completed;
845 				if (fill_level > hard_max_fill) {
846 					pend -= pkt_descs;
847 					txq->hw_vlan_tci = hw_vlan_tci_prev;
848 					break;
849 				}
850 			} else {
851 				pend -= pkt_descs;
852 				txq->hw_vlan_tci = hw_vlan_tci_prev;
853 				break;
854 			}
855 		}
856 
857 		/* Assign mbuf to the last used desc */
858 		txq->sw_ring[(added - 1) & txq->ptr_mask].mbuf = *pktp;
859 	}
860 
861 	if (likely(pkts_sent > 0)) {
862 		rc = efx_tx_qdesc_post(txq->common, txq->pend_desc,
863 				       pend - &txq->pend_desc[0],
864 				       txq->completed, &txq->added);
865 		SFC_ASSERT(rc == 0);
866 
867 		if (likely(pushed != txq->added))
868 			efx_tx_qpush(txq->common, txq->added, pushed);
869 	}
870 
871 #if SFC_TX_XMIT_PKTS_REAP_AT_LEAST_ONCE
872 	if (!reap_done)
873 		sfc_efx_tx_reap(txq);
874 #endif
875 
876 done:
877 	return pkts_sent;
878 }
879 
880 const struct sfc_dp_tx *
881 sfc_dp_tx_by_dp_txq(const struct sfc_dp_txq *dp_txq)
882 {
883 	const struct sfc_dp_queue *dpq = &dp_txq->dpq;
884 	struct rte_eth_dev *eth_dev;
885 	struct sfc_adapter_priv *sap;
886 
887 	SFC_ASSERT(rte_eth_dev_is_valid_port(dpq->port_id));
888 	eth_dev = &rte_eth_devices[dpq->port_id];
889 
890 	sap = sfc_adapter_priv_by_eth_dev(eth_dev);
891 
892 	return sap->dp_tx;
893 }
894 
895 struct sfc_txq_info *
896 sfc_txq_info_by_dp_txq(const struct sfc_dp_txq *dp_txq)
897 {
898 	const struct sfc_dp_queue *dpq = &dp_txq->dpq;
899 	struct rte_eth_dev *eth_dev;
900 	struct sfc_adapter_shared *sas;
901 
902 	SFC_ASSERT(rte_eth_dev_is_valid_port(dpq->port_id));
903 	eth_dev = &rte_eth_devices[dpq->port_id];
904 
905 	sas = sfc_adapter_shared_by_eth_dev(eth_dev);
906 
907 	SFC_ASSERT(dpq->queue_id < sas->txq_count);
908 	return &sas->txq_info[dpq->queue_id];
909 }
910 
911 struct sfc_txq *
912 sfc_txq_by_dp_txq(const struct sfc_dp_txq *dp_txq)
913 {
914 	const struct sfc_dp_queue *dpq = &dp_txq->dpq;
915 	struct rte_eth_dev *eth_dev;
916 	struct sfc_adapter *sa;
917 
918 	SFC_ASSERT(rte_eth_dev_is_valid_port(dpq->port_id));
919 	eth_dev = &rte_eth_devices[dpq->port_id];
920 
921 	sa = sfc_adapter_by_eth_dev(eth_dev);
922 
923 	SFC_ASSERT(dpq->queue_id < sfc_sa2shared(sa)->txq_count);
924 	return &sa->txq_ctrl[dpq->queue_id];
925 }
926 
927 static sfc_dp_tx_qsize_up_rings_t sfc_efx_tx_qsize_up_rings;
928 static int
929 sfc_efx_tx_qsize_up_rings(uint16_t nb_tx_desc,
930 			  __rte_unused struct sfc_dp_tx_hw_limits *limits,
931 			  unsigned int *txq_entries,
932 			  unsigned int *evq_entries,
933 			  unsigned int *txq_max_fill_level)
934 {
935 	*txq_entries = nb_tx_desc;
936 	*evq_entries = nb_tx_desc;
937 	*txq_max_fill_level = EFX_TXQ_LIMIT(*txq_entries);
938 	return 0;
939 }
940 
941 static sfc_dp_tx_qcreate_t sfc_efx_tx_qcreate;
942 static int
943 sfc_efx_tx_qcreate(uint16_t port_id, uint16_t queue_id,
944 		   const struct rte_pci_addr *pci_addr,
945 		   int socket_id,
946 		   const struct sfc_dp_tx_qcreate_info *info,
947 		   struct sfc_dp_txq **dp_txqp)
948 {
949 	struct sfc_efx_txq *txq;
950 	struct sfc_txq *ctrl_txq;
951 	int rc;
952 
953 	rc = ENOMEM;
954 	txq = rte_zmalloc_socket("sfc-efx-txq", sizeof(*txq),
955 				 RTE_CACHE_LINE_SIZE, socket_id);
956 	if (txq == NULL)
957 		goto fail_txq_alloc;
958 
959 	sfc_dp_queue_init(&txq->dp.dpq, port_id, queue_id, pci_addr);
960 
961 	rc = ENOMEM;
962 	txq->pend_desc = rte_calloc_socket("sfc-efx-txq-pend-desc",
963 					   EFX_TXQ_LIMIT(info->txq_entries),
964 					   sizeof(*txq->pend_desc), 0,
965 					   socket_id);
966 	if (txq->pend_desc == NULL)
967 		goto fail_pend_desc_alloc;
968 
969 	rc = ENOMEM;
970 	txq->sw_ring = rte_calloc_socket("sfc-efx-txq-sw_ring",
971 					 info->txq_entries,
972 					 sizeof(*txq->sw_ring),
973 					 RTE_CACHE_LINE_SIZE, socket_id);
974 	if (txq->sw_ring == NULL)
975 		goto fail_sw_ring_alloc;
976 
977 	ctrl_txq = sfc_txq_by_dp_txq(&txq->dp);
978 	if (ctrl_txq->evq->sa->tso) {
979 		rc = sfc_efx_tso_alloc_tsoh_objs(txq->sw_ring,
980 						 info->txq_entries, socket_id);
981 		if (rc != 0)
982 			goto fail_alloc_tsoh_objs;
983 	}
984 
985 	txq->evq = ctrl_txq->evq;
986 	txq->ptr_mask = info->txq_entries - 1;
987 	txq->max_fill_level = info->max_fill_level;
988 	txq->free_thresh = info->free_thresh;
989 	txq->dma_desc_size_max = info->dma_desc_size_max;
990 
991 	*dp_txqp = &txq->dp;
992 	return 0;
993 
994 fail_alloc_tsoh_objs:
995 	rte_free(txq->sw_ring);
996 
997 fail_sw_ring_alloc:
998 	rte_free(txq->pend_desc);
999 
1000 fail_pend_desc_alloc:
1001 	rte_free(txq);
1002 
1003 fail_txq_alloc:
1004 	return rc;
1005 }
1006 
1007 static sfc_dp_tx_qdestroy_t sfc_efx_tx_qdestroy;
1008 static void
1009 sfc_efx_tx_qdestroy(struct sfc_dp_txq *dp_txq)
1010 {
1011 	struct sfc_efx_txq *txq = sfc_efx_txq_by_dp_txq(dp_txq);
1012 
1013 	sfc_efx_tso_free_tsoh_objs(txq->sw_ring, txq->ptr_mask + 1);
1014 	rte_free(txq->sw_ring);
1015 	rte_free(txq->pend_desc);
1016 	rte_free(txq);
1017 }
1018 
1019 static sfc_dp_tx_qstart_t sfc_efx_tx_qstart;
1020 static int
1021 sfc_efx_tx_qstart(struct sfc_dp_txq *dp_txq,
1022 		  __rte_unused unsigned int evq_read_ptr,
1023 		  unsigned int txq_desc_index)
1024 {
1025 	/* libefx-based datapath is specific to libefx-based PMD */
1026 	struct sfc_efx_txq *txq = sfc_efx_txq_by_dp_txq(dp_txq);
1027 	struct sfc_txq *ctrl_txq = sfc_txq_by_dp_txq(dp_txq);
1028 
1029 	txq->common = ctrl_txq->common;
1030 
1031 	txq->pending = txq->completed = txq->added = txq_desc_index;
1032 	txq->hw_vlan_tci = 0;
1033 
1034 	txq->flags |= (SFC_EFX_TXQ_FLAG_STARTED | SFC_EFX_TXQ_FLAG_RUNNING);
1035 
1036 	return 0;
1037 }
1038 
1039 static sfc_dp_tx_qstop_t sfc_efx_tx_qstop;
1040 static void
1041 sfc_efx_tx_qstop(struct sfc_dp_txq *dp_txq,
1042 		 __rte_unused unsigned int *evq_read_ptr)
1043 {
1044 	struct sfc_efx_txq *txq = sfc_efx_txq_by_dp_txq(dp_txq);
1045 
1046 	txq->flags &= ~SFC_EFX_TXQ_FLAG_RUNNING;
1047 }
1048 
1049 static sfc_dp_tx_qreap_t sfc_efx_tx_qreap;
1050 static void
1051 sfc_efx_tx_qreap(struct sfc_dp_txq *dp_txq)
1052 {
1053 	struct sfc_efx_txq *txq = sfc_efx_txq_by_dp_txq(dp_txq);
1054 	unsigned int txds;
1055 
1056 	sfc_efx_tx_reap(txq);
1057 
1058 	for (txds = 0; txds <= txq->ptr_mask; txds++) {
1059 		if (txq->sw_ring[txds].mbuf != NULL) {
1060 			rte_pktmbuf_free(txq->sw_ring[txds].mbuf);
1061 			txq->sw_ring[txds].mbuf = NULL;
1062 		}
1063 	}
1064 
1065 	txq->flags &= ~SFC_EFX_TXQ_FLAG_STARTED;
1066 }
1067 
1068 static sfc_dp_tx_qdesc_status_t sfc_efx_tx_qdesc_status;
1069 static int
1070 sfc_efx_tx_qdesc_status(struct sfc_dp_txq *dp_txq, uint16_t offset)
1071 {
1072 	struct sfc_efx_txq *txq = sfc_efx_txq_by_dp_txq(dp_txq);
1073 
1074 	if (unlikely(offset > txq->ptr_mask))
1075 		return -EINVAL;
1076 
1077 	if (unlikely(offset >= txq->max_fill_level))
1078 		return RTE_ETH_TX_DESC_UNAVAIL;
1079 
1080 	/*
1081 	 * Poll EvQ to derive up-to-date 'txq->pending' figure;
1082 	 * it is required for the queue to be running, but the
1083 	 * check is omitted because API design assumes that it
1084 	 * is the duty of the caller to satisfy all conditions
1085 	 */
1086 	SFC_ASSERT((txq->flags & SFC_EFX_TXQ_FLAG_RUNNING) ==
1087 		   SFC_EFX_TXQ_FLAG_RUNNING);
1088 	sfc_ev_qpoll(txq->evq);
1089 
1090 	/*
1091 	 * Ring tail is 'txq->pending', and although descriptors
1092 	 * between 'txq->completed' and 'txq->pending' are still
1093 	 * in use by the driver, they should be reported as DONE
1094 	 */
1095 	if (unlikely(offset < (txq->added - txq->pending)))
1096 		return RTE_ETH_TX_DESC_FULL;
1097 
1098 	/*
1099 	 * There is no separate return value for unused descriptors;
1100 	 * the latter will be reported as DONE because genuine DONE
1101 	 * descriptors will be freed anyway in SW on the next burst
1102 	 */
1103 	return RTE_ETH_TX_DESC_DONE;
1104 }
1105 
1106 struct sfc_dp_tx sfc_efx_tx = {
1107 	.dp = {
1108 		.name		= SFC_KVARG_DATAPATH_EFX,
1109 		.type		= SFC_DP_TX,
1110 		.hw_fw_caps	= 0,
1111 	},
1112 	.features		= SFC_DP_TX_FEAT_VLAN_INSERT |
1113 				  SFC_DP_TX_FEAT_TSO |
1114 				  SFC_DP_TX_FEAT_MULTI_POOL |
1115 				  SFC_DP_TX_FEAT_REFCNT |
1116 				  SFC_DP_TX_FEAT_MULTI_SEG,
1117 	.qsize_up_rings		= sfc_efx_tx_qsize_up_rings,
1118 	.qcreate		= sfc_efx_tx_qcreate,
1119 	.qdestroy		= sfc_efx_tx_qdestroy,
1120 	.qstart			= sfc_efx_tx_qstart,
1121 	.qstop			= sfc_efx_tx_qstop,
1122 	.qreap			= sfc_efx_tx_qreap,
1123 	.qdesc_status		= sfc_efx_tx_qdesc_status,
1124 	.pkt_burst		= sfc_efx_xmit_pkts,
1125 };
1126