xref: /dpdk/drivers/net/sfc/sfc_tx.c (revision c7f5dba7d4bb7971fac51755aad09b71b10cef90)
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->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->dp_tx->features & SFC_DP_TX_FEAT_MULTI_SEG)
48 		caps |= DEV_TX_OFFLOAD_MULTI_SEGS;
49 
50 	if ((~sa->dp_tx->features & SFC_DP_TX_FEAT_MULTI_POOL) &&
51 	    (~sa->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 *txq)
114 {
115 	txq->state |= SFC_TXQ_FLUSHED;
116 	txq->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 
135 	sfc_log_init(sa, "TxQ = %u", sw_index);
136 
137 	rc = sa->dp_tx->qsize_up_rings(nb_tx_desc, &txq_entries, &evq_entries,
138 				       &txq_max_fill_level);
139 	if (rc != 0)
140 		goto fail_size_up_rings;
141 	SFC_ASSERT(txq_entries >= EFX_TXQ_MINNDESCS);
142 	SFC_ASSERT(txq_entries <= sa->txq_max_entries);
143 	SFC_ASSERT(txq_entries >= nb_tx_desc);
144 	SFC_ASSERT(txq_max_fill_level <= nb_tx_desc);
145 
146 	offloads = tx_conf->offloads |
147 		sa->eth_dev->data->dev_conf.txmode.offloads;
148 	rc = sfc_tx_qcheck_conf(sa, txq_max_fill_level, tx_conf, offloads);
149 	if (rc != 0)
150 		goto fail_bad_conf;
151 
152 	SFC_ASSERT(sw_index < sa->txq_count);
153 	txq_info = &sa->txq_info[sw_index];
154 
155 	txq_info->entries = txq_entries;
156 
157 	rc = sfc_ev_qinit(sa, SFC_EVQ_TYPE_TX, sw_index,
158 			  evq_entries, socket_id, &evq);
159 	if (rc != 0)
160 		goto fail_ev_qinit;
161 
162 	rc = ENOMEM;
163 	txq = rte_zmalloc_socket("sfc-txq", sizeof(*txq), 0, socket_id);
164 	if (txq == NULL)
165 		goto fail_txq_alloc;
166 
167 	txq_info->txq = txq;
168 
169 	txq->hw_index = sw_index;
170 	txq->evq = evq;
171 	txq->free_thresh =
172 		(tx_conf->tx_free_thresh) ? tx_conf->tx_free_thresh :
173 		SFC_TX_DEFAULT_FREE_THRESH;
174 	txq->offloads = offloads;
175 
176 	rc = sfc_dma_alloc(sa, "txq", sw_index, EFX_TXQ_SIZE(txq_info->entries),
177 			   socket_id, &txq->mem);
178 	if (rc != 0)
179 		goto fail_dma_alloc;
180 
181 	memset(&info, 0, sizeof(info));
182 	info.max_fill_level = txq_max_fill_level;
183 	info.free_thresh = txq->free_thresh;
184 	info.offloads = offloads;
185 	info.txq_entries = txq_info->entries;
186 	info.dma_desc_size_max = encp->enc_tx_dma_desc_size_max;
187 	info.txq_hw_ring = txq->mem.esm_base;
188 	info.evq_entries = evq_entries;
189 	info.evq_hw_ring = evq->mem.esm_base;
190 	info.hw_index = txq->hw_index;
191 	info.mem_bar = sa->mem_bar.esb_base;
192 	info.vi_window_shift = encp->enc_vi_window_shift;
193 
194 	rc = sa->dp_tx->qcreate(sa->eth_dev->data->port_id, sw_index,
195 				&RTE_ETH_DEV_TO_PCI(sa->eth_dev)->addr,
196 				socket_id, &info, &txq->dp);
197 	if (rc != 0)
198 		goto fail_dp_tx_qinit;
199 
200 	evq->dp_txq = txq->dp;
201 
202 	txq->state = SFC_TXQ_INITIALIZED;
203 
204 	txq_info->deferred_start = (tx_conf->tx_deferred_start != 0);
205 
206 	return 0;
207 
208 fail_dp_tx_qinit:
209 	sfc_dma_free(sa, &txq->mem);
210 
211 fail_dma_alloc:
212 	txq_info->txq = NULL;
213 	rte_free(txq);
214 
215 fail_txq_alloc:
216 	sfc_ev_qfini(evq);
217 
218 fail_ev_qinit:
219 	txq_info->entries = 0;
220 
221 fail_bad_conf:
222 fail_size_up_rings:
223 	sfc_log_init(sa, "failed (TxQ = %u, rc = %d)", sw_index, rc);
224 	return rc;
225 }
226 
227 void
228 sfc_tx_qfini(struct sfc_adapter *sa, unsigned int sw_index)
229 {
230 	struct sfc_txq_info *txq_info;
231 	struct sfc_txq *txq;
232 
233 	sfc_log_init(sa, "TxQ = %u", sw_index);
234 
235 	SFC_ASSERT(sw_index < sa->txq_count);
236 	sa->eth_dev->data->tx_queues[sw_index] = NULL;
237 
238 	txq_info = &sa->txq_info[sw_index];
239 
240 	txq = txq_info->txq;
241 	SFC_ASSERT(txq != NULL);
242 	SFC_ASSERT(txq->state == SFC_TXQ_INITIALIZED);
243 
244 	sa->dp_tx->qdestroy(txq->dp);
245 	txq->dp = NULL;
246 
247 	txq_info->txq = NULL;
248 	txq_info->entries = 0;
249 
250 	sfc_dma_free(sa, &txq->mem);
251 
252 	sfc_ev_qfini(txq->evq);
253 	txq->evq = NULL;
254 
255 	rte_free(txq);
256 }
257 
258 static int
259 sfc_tx_qinit_info(struct sfc_adapter *sa, unsigned int sw_index)
260 {
261 	sfc_log_init(sa, "TxQ = %u", sw_index);
262 
263 	return 0;
264 }
265 
266 static int
267 sfc_tx_check_mode(struct sfc_adapter *sa, const struct rte_eth_txmode *txmode)
268 {
269 	int rc = 0;
270 
271 	switch (txmode->mq_mode) {
272 	case ETH_MQ_TX_NONE:
273 		break;
274 	default:
275 		sfc_err(sa, "Tx multi-queue mode %u not supported",
276 			txmode->mq_mode);
277 		rc = EINVAL;
278 	}
279 
280 	/*
281 	 * These features are claimed to be i40e-specific,
282 	 * but it does make sense to double-check their absence
283 	 */
284 	if (txmode->hw_vlan_reject_tagged) {
285 		sfc_err(sa, "Rejecting tagged packets not supported");
286 		rc = EINVAL;
287 	}
288 
289 	if (txmode->hw_vlan_reject_untagged) {
290 		sfc_err(sa, "Rejecting untagged packets not supported");
291 		rc = EINVAL;
292 	}
293 
294 	if (txmode->hw_vlan_insert_pvid) {
295 		sfc_err(sa, "Port-based VLAN insertion not supported");
296 		rc = EINVAL;
297 	}
298 
299 	return rc;
300 }
301 
302 /**
303  * Destroy excess queues that are no longer needed after reconfiguration
304  * or complete close.
305  */
306 static void
307 sfc_tx_fini_queues(struct sfc_adapter *sa, unsigned int nb_tx_queues)
308 {
309 	int sw_index;
310 
311 	SFC_ASSERT(nb_tx_queues <= sa->txq_count);
312 
313 	sw_index = sa->txq_count;
314 	while (--sw_index >= (int)nb_tx_queues) {
315 		if (sa->txq_info[sw_index].txq != NULL)
316 			sfc_tx_qfini(sa, sw_index);
317 	}
318 
319 	sa->txq_count = nb_tx_queues;
320 }
321 
322 int
323 sfc_tx_configure(struct sfc_adapter *sa)
324 {
325 	const efx_nic_cfg_t *encp = efx_nic_cfg_get(sa->nic);
326 	const struct rte_eth_conf *dev_conf = &sa->eth_dev->data->dev_conf;
327 	const unsigned int nb_tx_queues = sa->eth_dev->data->nb_tx_queues;
328 	int rc = 0;
329 
330 	sfc_log_init(sa, "nb_tx_queues=%u (old %u)",
331 		     nb_tx_queues, sa->txq_count);
332 
333 	/*
334 	 * The datapath implementation assumes absence of boundary
335 	 * limits on Tx DMA descriptors. Addition of these checks on
336 	 * datapath would simply make the datapath slower.
337 	 */
338 	if (encp->enc_tx_dma_desc_boundary != 0) {
339 		rc = ENOTSUP;
340 		goto fail_tx_dma_desc_boundary;
341 	}
342 
343 	rc = sfc_tx_check_mode(sa, &dev_conf->txmode);
344 	if (rc != 0)
345 		goto fail_check_mode;
346 
347 	if (nb_tx_queues == sa->txq_count)
348 		goto done;
349 
350 	if (sa->txq_info == NULL) {
351 		sa->txq_info = rte_calloc_socket("sfc-txqs", nb_tx_queues,
352 						 sizeof(sa->txq_info[0]), 0,
353 						 sa->socket_id);
354 		if (sa->txq_info == NULL)
355 			goto fail_txqs_alloc;
356 	} else {
357 		struct sfc_txq_info *new_txq_info;
358 
359 		if (nb_tx_queues < sa->txq_count)
360 			sfc_tx_fini_queues(sa, nb_tx_queues);
361 
362 		new_txq_info =
363 			rte_realloc(sa->txq_info,
364 				    nb_tx_queues * sizeof(sa->txq_info[0]), 0);
365 		if (new_txq_info == NULL && nb_tx_queues > 0)
366 			goto fail_txqs_realloc;
367 
368 		sa->txq_info = new_txq_info;
369 		if (nb_tx_queues > sa->txq_count)
370 			memset(&sa->txq_info[sa->txq_count], 0,
371 			       (nb_tx_queues - sa->txq_count) *
372 			       sizeof(sa->txq_info[0]));
373 	}
374 
375 	while (sa->txq_count < nb_tx_queues) {
376 		rc = sfc_tx_qinit_info(sa, sa->txq_count);
377 		if (rc != 0)
378 			goto fail_tx_qinit_info;
379 
380 		sa->txq_count++;
381 	}
382 
383 done:
384 	return 0;
385 
386 fail_tx_qinit_info:
387 fail_txqs_realloc:
388 fail_txqs_alloc:
389 	sfc_tx_close(sa);
390 
391 fail_check_mode:
392 fail_tx_dma_desc_boundary:
393 	sfc_log_init(sa, "failed (rc = %d)", rc);
394 	return rc;
395 }
396 
397 void
398 sfc_tx_close(struct sfc_adapter *sa)
399 {
400 	sfc_tx_fini_queues(sa, 0);
401 
402 	rte_free(sa->txq_info);
403 	sa->txq_info = NULL;
404 }
405 
406 int
407 sfc_tx_qstart(struct sfc_adapter *sa, unsigned int sw_index)
408 {
409 	uint64_t offloads_supported = sfc_tx_get_dev_offload_caps(sa) |
410 				      sfc_tx_get_queue_offload_caps(sa);
411 	struct rte_eth_dev_data *dev_data;
412 	struct sfc_txq_info *txq_info;
413 	struct sfc_txq *txq;
414 	struct sfc_evq *evq;
415 	uint16_t flags = 0;
416 	unsigned int desc_index;
417 	int rc = 0;
418 
419 	sfc_log_init(sa, "TxQ = %u", sw_index);
420 
421 	SFC_ASSERT(sw_index < sa->txq_count);
422 	txq_info = &sa->txq_info[sw_index];
423 
424 	txq = txq_info->txq;
425 
426 	SFC_ASSERT(txq != NULL);
427 	SFC_ASSERT(txq->state == SFC_TXQ_INITIALIZED);
428 
429 	evq = txq->evq;
430 
431 	rc = sfc_ev_qstart(evq, sfc_evq_index_by_txq_sw_index(sa, sw_index));
432 	if (rc != 0)
433 		goto fail_ev_qstart;
434 
435 	if (txq->offloads & DEV_TX_OFFLOAD_IPV4_CKSUM)
436 		flags |= EFX_TXQ_CKSUM_IPV4;
437 
438 	if (txq->offloads & DEV_TX_OFFLOAD_OUTER_IPV4_CKSUM)
439 		flags |= EFX_TXQ_CKSUM_INNER_IPV4;
440 
441 	if ((txq->offloads & DEV_TX_OFFLOAD_TCP_CKSUM) ||
442 	    (txq->offloads & DEV_TX_OFFLOAD_UDP_CKSUM)) {
443 		flags |= EFX_TXQ_CKSUM_TCPUDP;
444 
445 		if (offloads_supported & DEV_TX_OFFLOAD_OUTER_IPV4_CKSUM)
446 			flags |= EFX_TXQ_CKSUM_INNER_TCPUDP;
447 	}
448 
449 	if (txq->offloads & DEV_TX_OFFLOAD_TCP_TSO)
450 		flags |= EFX_TXQ_FATSOV2;
451 
452 	rc = efx_tx_qcreate(sa->nic, sw_index, 0, &txq->mem,
453 			    txq_info->entries, 0 /* not used on EF10 */,
454 			    flags, evq->common,
455 			    &txq->common, &desc_index);
456 	if (rc != 0) {
457 		if (sa->tso && (rc == ENOSPC))
458 			sfc_err(sa, "ran out of TSO contexts");
459 
460 		goto fail_tx_qcreate;
461 	}
462 
463 	efx_tx_qenable(txq->common);
464 
465 	txq->state |= SFC_TXQ_STARTED;
466 
467 	rc = sa->dp_tx->qstart(txq->dp, evq->read_ptr, desc_index);
468 	if (rc != 0)
469 		goto fail_dp_qstart;
470 
471 	/*
472 	 * It seems to be used by DPDK for debug purposes only ('rte_ether')
473 	 */
474 	dev_data = sa->eth_dev->data;
475 	dev_data->tx_queue_state[sw_index] = RTE_ETH_QUEUE_STATE_STARTED;
476 
477 	return 0;
478 
479 fail_dp_qstart:
480 	txq->state = SFC_TXQ_INITIALIZED;
481 	efx_tx_qdestroy(txq->common);
482 
483 fail_tx_qcreate:
484 	sfc_ev_qstop(evq);
485 
486 fail_ev_qstart:
487 	return rc;
488 }
489 
490 void
491 sfc_tx_qstop(struct sfc_adapter *sa, unsigned int sw_index)
492 {
493 	struct rte_eth_dev_data *dev_data;
494 	struct sfc_txq_info *txq_info;
495 	struct sfc_txq *txq;
496 	unsigned int retry_count;
497 	unsigned int wait_count;
498 	int rc;
499 
500 	sfc_log_init(sa, "TxQ = %u", sw_index);
501 
502 	SFC_ASSERT(sw_index < sa->txq_count);
503 	txq_info = &sa->txq_info[sw_index];
504 
505 	txq = txq_info->txq;
506 
507 	if (txq == NULL || txq->state == SFC_TXQ_INITIALIZED)
508 		return;
509 
510 	SFC_ASSERT(txq->state & SFC_TXQ_STARTED);
511 
512 	sa->dp_tx->qstop(txq->dp, &txq->evq->read_ptr);
513 
514 	/*
515 	 * Retry TX queue flushing in case of flush failed or
516 	 * timeout; in the worst case it can delay for 6 seconds
517 	 */
518 	for (retry_count = 0;
519 	     ((txq->state & SFC_TXQ_FLUSHED) == 0) &&
520 	     (retry_count < SFC_TX_QFLUSH_ATTEMPTS);
521 	     ++retry_count) {
522 		rc = efx_tx_qflush(txq->common);
523 		if (rc != 0) {
524 			txq->state |= (rc == EALREADY) ?
525 				SFC_TXQ_FLUSHED : SFC_TXQ_FLUSH_FAILED;
526 			break;
527 		}
528 
529 		/*
530 		 * Wait for TX queue flush done or flush failed event at least
531 		 * SFC_TX_QFLUSH_POLL_WAIT_MS milliseconds and not more
532 		 * than 2 seconds (SFC_TX_QFLUSH_POLL_WAIT_MS multiplied
533 		 * by SFC_TX_QFLUSH_POLL_ATTEMPTS)
534 		 */
535 		wait_count = 0;
536 		do {
537 			rte_delay_ms(SFC_TX_QFLUSH_POLL_WAIT_MS);
538 			sfc_ev_qpoll(txq->evq);
539 		} while ((txq->state & SFC_TXQ_FLUSHING) &&
540 			 wait_count++ < SFC_TX_QFLUSH_POLL_ATTEMPTS);
541 
542 		if (txq->state & SFC_TXQ_FLUSHING)
543 			sfc_err(sa, "TxQ %u flush timed out", sw_index);
544 
545 		if (txq->state & SFC_TXQ_FLUSHED)
546 			sfc_notice(sa, "TxQ %u flushed", sw_index);
547 	}
548 
549 	sa->dp_tx->qreap(txq->dp);
550 
551 	txq->state = SFC_TXQ_INITIALIZED;
552 
553 	efx_tx_qdestroy(txq->common);
554 
555 	sfc_ev_qstop(txq->evq);
556 
557 	/*
558 	 * It seems to be used by DPDK for debug purposes only ('rte_ether')
559 	 */
560 	dev_data = sa->eth_dev->data;
561 	dev_data->tx_queue_state[sw_index] = RTE_ETH_QUEUE_STATE_STOPPED;
562 }
563 
564 int
565 sfc_tx_start(struct sfc_adapter *sa)
566 {
567 	unsigned int sw_index;
568 	int rc = 0;
569 
570 	sfc_log_init(sa, "txq_count = %u", sa->txq_count);
571 
572 	if (sa->tso) {
573 		if (!efx_nic_cfg_get(sa->nic)->enc_fw_assisted_tso_v2_enabled) {
574 			sfc_warn(sa, "TSO support was unable to be restored");
575 			sa->tso = B_FALSE;
576 		}
577 	}
578 
579 	rc = efx_tx_init(sa->nic);
580 	if (rc != 0)
581 		goto fail_efx_tx_init;
582 
583 	for (sw_index = 0; sw_index < sa->txq_count; ++sw_index) {
584 		if (sa->txq_info[sw_index].txq != NULL &&
585 		    (!(sa->txq_info[sw_index].deferred_start) ||
586 		     sa->txq_info[sw_index].deferred_started)) {
587 			rc = sfc_tx_qstart(sa, sw_index);
588 			if (rc != 0)
589 				goto fail_tx_qstart;
590 		}
591 	}
592 
593 	return 0;
594 
595 fail_tx_qstart:
596 	while (sw_index-- > 0)
597 		sfc_tx_qstop(sa, sw_index);
598 
599 	efx_tx_fini(sa->nic);
600 
601 fail_efx_tx_init:
602 	sfc_log_init(sa, "failed (rc = %d)", rc);
603 	return rc;
604 }
605 
606 void
607 sfc_tx_stop(struct sfc_adapter *sa)
608 {
609 	unsigned int sw_index;
610 
611 	sfc_log_init(sa, "txq_count = %u", sa->txq_count);
612 
613 	sw_index = sa->txq_count;
614 	while (sw_index-- > 0) {
615 		if (sa->txq_info[sw_index].txq != NULL)
616 			sfc_tx_qstop(sa, sw_index);
617 	}
618 
619 	efx_tx_fini(sa->nic);
620 }
621 
622 static void
623 sfc_efx_tx_reap(struct sfc_efx_txq *txq)
624 {
625 	unsigned int completed;
626 
627 	sfc_ev_qpoll(txq->evq);
628 
629 	for (completed = txq->completed;
630 	     completed != txq->pending; completed++) {
631 		struct sfc_efx_tx_sw_desc *txd;
632 
633 		txd = &txq->sw_ring[completed & txq->ptr_mask];
634 
635 		if (txd->mbuf != NULL) {
636 			rte_pktmbuf_free(txd->mbuf);
637 			txd->mbuf = NULL;
638 		}
639 	}
640 
641 	txq->completed = completed;
642 }
643 
644 /*
645  * The function is used to insert or update VLAN tag;
646  * the firmware has state of the firmware tag to insert per TxQ
647  * (controlled by option descriptors), hence, if the tag of the
648  * packet to be sent is different from one remembered by the firmware,
649  * the function will update it
650  */
651 static unsigned int
652 sfc_efx_tx_maybe_insert_tag(struct sfc_efx_txq *txq, struct rte_mbuf *m,
653 			    efx_desc_t **pend)
654 {
655 	uint16_t this_tag = ((m->ol_flags & PKT_TX_VLAN_PKT) ?
656 			     m->vlan_tci : 0);
657 
658 	if (this_tag == txq->hw_vlan_tci)
659 		return 0;
660 
661 	/*
662 	 * The expression inside SFC_ASSERT() is not desired to be checked in
663 	 * a non-debug build because it might be too expensive on the data path
664 	 */
665 	SFC_ASSERT(efx_nic_cfg_get(txq->evq->sa->nic)->enc_hw_tx_insert_vlan_enabled);
666 
667 	efx_tx_qdesc_vlantci_create(txq->common, rte_cpu_to_be_16(this_tag),
668 				    *pend);
669 	(*pend)++;
670 	txq->hw_vlan_tci = this_tag;
671 
672 	return 1;
673 }
674 
675 static uint16_t
676 sfc_efx_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
677 {
678 	struct sfc_dp_txq *dp_txq = (struct sfc_dp_txq *)tx_queue;
679 	struct sfc_efx_txq *txq = sfc_efx_txq_by_dp_txq(dp_txq);
680 	unsigned int added = txq->added;
681 	unsigned int pushed = added;
682 	unsigned int pkts_sent = 0;
683 	efx_desc_t *pend = &txq->pend_desc[0];
684 	const unsigned int hard_max_fill = txq->max_fill_level;
685 	const unsigned int soft_max_fill = hard_max_fill - txq->free_thresh;
686 	unsigned int fill_level = added - txq->completed;
687 	boolean_t reap_done;
688 	int rc __rte_unused;
689 	struct rte_mbuf **pktp;
690 
691 	if (unlikely((txq->flags & SFC_EFX_TXQ_FLAG_RUNNING) == 0))
692 		goto done;
693 
694 	/*
695 	 * If insufficient space for a single packet is present,
696 	 * we should reap; otherwise, we shouldn't do that all the time
697 	 * to avoid latency increase
698 	 */
699 	reap_done = (fill_level > soft_max_fill);
700 
701 	if (reap_done) {
702 		sfc_efx_tx_reap(txq);
703 		/*
704 		 * Recalculate fill level since 'txq->completed'
705 		 * might have changed on reap
706 		 */
707 		fill_level = added - txq->completed;
708 	}
709 
710 	for (pkts_sent = 0, pktp = &tx_pkts[0];
711 	     (pkts_sent < nb_pkts) && (fill_level <= soft_max_fill);
712 	     pkts_sent++, pktp++) {
713 		struct rte_mbuf		*m_seg = *pktp;
714 		size_t			pkt_len = m_seg->pkt_len;
715 		unsigned int		pkt_descs = 0;
716 		size_t			in_off = 0;
717 
718 		/*
719 		 * Here VLAN TCI is expected to be zero in case if no
720 		 * DEV_TX_OFFLOAD_VLAN_INSERT capability is advertised;
721 		 * if the calling app ignores the absence of
722 		 * DEV_TX_OFFLOAD_VLAN_INSERT and pushes VLAN TCI, then
723 		 * TX_ERROR will occur
724 		 */
725 		pkt_descs += sfc_efx_tx_maybe_insert_tag(txq, m_seg, &pend);
726 
727 		if (m_seg->ol_flags & PKT_TX_TCP_SEG) {
728 			/*
729 			 * We expect correct 'pkt->l[2, 3, 4]_len' values
730 			 * to be set correctly by the caller
731 			 */
732 			if (sfc_efx_tso_do(txq, added, &m_seg, &in_off, &pend,
733 					   &pkt_descs, &pkt_len) != 0) {
734 				/* We may have reached this place for
735 				 * one of the following reasons:
736 				 *
737 				 * 1) Packet header length is greater
738 				 *    than SFC_TSOH_STD_LEN
739 				 * 2) TCP header starts at more then
740 				 *    208 bytes into the frame
741 				 *
742 				 * We will deceive RTE saying that we have sent
743 				 * the packet, but we will actually drop it.
744 				 * Hence, we should revert 'pend' to the
745 				 * previous state (in case we have added
746 				 * VLAN descriptor) and start processing
747 				 * another one packet. But the original
748 				 * mbuf shouldn't be orphaned
749 				 */
750 				pend -= pkt_descs;
751 
752 				rte_pktmbuf_free(*pktp);
753 
754 				continue;
755 			}
756 
757 			/*
758 			 * We've only added 2 FATSOv2 option descriptors
759 			 * and 1 descriptor for the linearized packet header.
760 			 * The outstanding work will be done in the same manner
761 			 * as for the usual non-TSO path
762 			 */
763 		}
764 
765 		for (; m_seg != NULL; m_seg = m_seg->next) {
766 			efsys_dma_addr_t	next_frag;
767 			size_t			seg_len;
768 
769 			seg_len = m_seg->data_len;
770 			next_frag = rte_mbuf_data_iova(m_seg);
771 
772 			/*
773 			 * If we've started TSO transaction few steps earlier,
774 			 * we'll skip packet header using an offset in the
775 			 * current segment (which has been set to the
776 			 * first one containing payload)
777 			 */
778 			seg_len -= in_off;
779 			next_frag += in_off;
780 			in_off = 0;
781 
782 			do {
783 				efsys_dma_addr_t	frag_addr = next_frag;
784 				size_t			frag_len;
785 
786 				/*
787 				 * It is assumed here that there is no
788 				 * limitation on address boundary
789 				 * crossing by DMA descriptor.
790 				 */
791 				frag_len = MIN(seg_len, txq->dma_desc_size_max);
792 				next_frag += frag_len;
793 				seg_len -= frag_len;
794 				pkt_len -= frag_len;
795 
796 				efx_tx_qdesc_dma_create(txq->common,
797 							frag_addr, frag_len,
798 							(pkt_len == 0),
799 							pend++);
800 
801 				pkt_descs++;
802 			} while (seg_len != 0);
803 		}
804 
805 		added += pkt_descs;
806 
807 		fill_level += pkt_descs;
808 		if (unlikely(fill_level > hard_max_fill)) {
809 			/*
810 			 * Our estimation for maximum number of descriptors
811 			 * required to send a packet seems to be wrong.
812 			 * Try to reap (if we haven't yet).
813 			 */
814 			if (!reap_done) {
815 				sfc_efx_tx_reap(txq);
816 				reap_done = B_TRUE;
817 				fill_level = added - txq->completed;
818 				if (fill_level > hard_max_fill) {
819 					pend -= pkt_descs;
820 					break;
821 				}
822 			} else {
823 				pend -= pkt_descs;
824 				break;
825 			}
826 		}
827 
828 		/* Assign mbuf to the last used desc */
829 		txq->sw_ring[(added - 1) & txq->ptr_mask].mbuf = *pktp;
830 	}
831 
832 	if (likely(pkts_sent > 0)) {
833 		rc = efx_tx_qdesc_post(txq->common, txq->pend_desc,
834 				       pend - &txq->pend_desc[0],
835 				       txq->completed, &txq->added);
836 		SFC_ASSERT(rc == 0);
837 
838 		if (likely(pushed != txq->added))
839 			efx_tx_qpush(txq->common, txq->added, pushed);
840 	}
841 
842 #if SFC_TX_XMIT_PKTS_REAP_AT_LEAST_ONCE
843 	if (!reap_done)
844 		sfc_efx_tx_reap(txq);
845 #endif
846 
847 done:
848 	return pkts_sent;
849 }
850 
851 struct sfc_txq *
852 sfc_txq_by_dp_txq(const struct sfc_dp_txq *dp_txq)
853 {
854 	const struct sfc_dp_queue *dpq = &dp_txq->dpq;
855 	struct rte_eth_dev *eth_dev;
856 	struct sfc_adapter *sa;
857 	struct sfc_txq *txq;
858 
859 	SFC_ASSERT(rte_eth_dev_is_valid_port(dpq->port_id));
860 	eth_dev = &rte_eth_devices[dpq->port_id];
861 
862 	sa = eth_dev->data->dev_private;
863 
864 	SFC_ASSERT(dpq->queue_id < sa->txq_count);
865 	txq = sa->txq_info[dpq->queue_id].txq;
866 
867 	SFC_ASSERT(txq != NULL);
868 	return txq;
869 }
870 
871 static sfc_dp_tx_qsize_up_rings_t sfc_efx_tx_qsize_up_rings;
872 static int
873 sfc_efx_tx_qsize_up_rings(uint16_t nb_tx_desc,
874 			  unsigned int *txq_entries,
875 			  unsigned int *evq_entries,
876 			  unsigned int *txq_max_fill_level)
877 {
878 	*txq_entries = nb_tx_desc;
879 	*evq_entries = nb_tx_desc;
880 	*txq_max_fill_level = EFX_TXQ_LIMIT(*txq_entries);
881 	return 0;
882 }
883 
884 static sfc_dp_tx_qcreate_t sfc_efx_tx_qcreate;
885 static int
886 sfc_efx_tx_qcreate(uint16_t port_id, uint16_t queue_id,
887 		   const struct rte_pci_addr *pci_addr,
888 		   int socket_id,
889 		   const struct sfc_dp_tx_qcreate_info *info,
890 		   struct sfc_dp_txq **dp_txqp)
891 {
892 	struct sfc_efx_txq *txq;
893 	struct sfc_txq *ctrl_txq;
894 	int rc;
895 
896 	rc = ENOMEM;
897 	txq = rte_zmalloc_socket("sfc-efx-txq", sizeof(*txq),
898 				 RTE_CACHE_LINE_SIZE, socket_id);
899 	if (txq == NULL)
900 		goto fail_txq_alloc;
901 
902 	sfc_dp_queue_init(&txq->dp.dpq, port_id, queue_id, pci_addr);
903 
904 	rc = ENOMEM;
905 	txq->pend_desc = rte_calloc_socket("sfc-efx-txq-pend-desc",
906 					   EFX_TXQ_LIMIT(info->txq_entries),
907 					   sizeof(*txq->pend_desc), 0,
908 					   socket_id);
909 	if (txq->pend_desc == NULL)
910 		goto fail_pend_desc_alloc;
911 
912 	rc = ENOMEM;
913 	txq->sw_ring = rte_calloc_socket("sfc-efx-txq-sw_ring",
914 					 info->txq_entries,
915 					 sizeof(*txq->sw_ring),
916 					 RTE_CACHE_LINE_SIZE, socket_id);
917 	if (txq->sw_ring == NULL)
918 		goto fail_sw_ring_alloc;
919 
920 	ctrl_txq = sfc_txq_by_dp_txq(&txq->dp);
921 	if (ctrl_txq->evq->sa->tso) {
922 		rc = sfc_efx_tso_alloc_tsoh_objs(txq->sw_ring,
923 						 info->txq_entries, socket_id);
924 		if (rc != 0)
925 			goto fail_alloc_tsoh_objs;
926 	}
927 
928 	txq->evq = ctrl_txq->evq;
929 	txq->ptr_mask = info->txq_entries - 1;
930 	txq->max_fill_level = info->max_fill_level;
931 	txq->free_thresh = info->free_thresh;
932 	txq->dma_desc_size_max = info->dma_desc_size_max;
933 
934 	*dp_txqp = &txq->dp;
935 	return 0;
936 
937 fail_alloc_tsoh_objs:
938 	rte_free(txq->sw_ring);
939 
940 fail_sw_ring_alloc:
941 	rte_free(txq->pend_desc);
942 
943 fail_pend_desc_alloc:
944 	rte_free(txq);
945 
946 fail_txq_alloc:
947 	return rc;
948 }
949 
950 static sfc_dp_tx_qdestroy_t sfc_efx_tx_qdestroy;
951 static void
952 sfc_efx_tx_qdestroy(struct sfc_dp_txq *dp_txq)
953 {
954 	struct sfc_efx_txq *txq = sfc_efx_txq_by_dp_txq(dp_txq);
955 
956 	sfc_efx_tso_free_tsoh_objs(txq->sw_ring, txq->ptr_mask + 1);
957 	rte_free(txq->sw_ring);
958 	rte_free(txq->pend_desc);
959 	rte_free(txq);
960 }
961 
962 static sfc_dp_tx_qstart_t sfc_efx_tx_qstart;
963 static int
964 sfc_efx_tx_qstart(struct sfc_dp_txq *dp_txq,
965 		  __rte_unused unsigned int evq_read_ptr,
966 		  unsigned int txq_desc_index)
967 {
968 	/* libefx-based datapath is specific to libefx-based PMD */
969 	struct sfc_efx_txq *txq = sfc_efx_txq_by_dp_txq(dp_txq);
970 	struct sfc_txq *ctrl_txq = sfc_txq_by_dp_txq(dp_txq);
971 
972 	txq->common = ctrl_txq->common;
973 
974 	txq->pending = txq->completed = txq->added = txq_desc_index;
975 	txq->hw_vlan_tci = 0;
976 
977 	txq->flags |= (SFC_EFX_TXQ_FLAG_STARTED | SFC_EFX_TXQ_FLAG_RUNNING);
978 
979 	return 0;
980 }
981 
982 static sfc_dp_tx_qstop_t sfc_efx_tx_qstop;
983 static void
984 sfc_efx_tx_qstop(struct sfc_dp_txq *dp_txq,
985 		 __rte_unused unsigned int *evq_read_ptr)
986 {
987 	struct sfc_efx_txq *txq = sfc_efx_txq_by_dp_txq(dp_txq);
988 
989 	txq->flags &= ~SFC_EFX_TXQ_FLAG_RUNNING;
990 }
991 
992 static sfc_dp_tx_qreap_t sfc_efx_tx_qreap;
993 static void
994 sfc_efx_tx_qreap(struct sfc_dp_txq *dp_txq)
995 {
996 	struct sfc_efx_txq *txq = sfc_efx_txq_by_dp_txq(dp_txq);
997 	unsigned int txds;
998 
999 	sfc_efx_tx_reap(txq);
1000 
1001 	for (txds = 0; txds <= txq->ptr_mask; txds++) {
1002 		if (txq->sw_ring[txds].mbuf != NULL) {
1003 			rte_pktmbuf_free(txq->sw_ring[txds].mbuf);
1004 			txq->sw_ring[txds].mbuf = NULL;
1005 		}
1006 	}
1007 
1008 	txq->flags &= ~SFC_EFX_TXQ_FLAG_STARTED;
1009 }
1010 
1011 static sfc_dp_tx_qdesc_status_t sfc_efx_tx_qdesc_status;
1012 static int
1013 sfc_efx_tx_qdesc_status(struct sfc_dp_txq *dp_txq, uint16_t offset)
1014 {
1015 	struct sfc_efx_txq *txq = sfc_efx_txq_by_dp_txq(dp_txq);
1016 
1017 	if (unlikely(offset > txq->ptr_mask))
1018 		return -EINVAL;
1019 
1020 	if (unlikely(offset >= txq->max_fill_level))
1021 		return RTE_ETH_TX_DESC_UNAVAIL;
1022 
1023 	/*
1024 	 * Poll EvQ to derive up-to-date 'txq->pending' figure;
1025 	 * it is required for the queue to be running, but the
1026 	 * check is omitted because API design assumes that it
1027 	 * is the duty of the caller to satisfy all conditions
1028 	 */
1029 	SFC_ASSERT((txq->flags & SFC_EFX_TXQ_FLAG_RUNNING) ==
1030 		   SFC_EFX_TXQ_FLAG_RUNNING);
1031 	sfc_ev_qpoll(txq->evq);
1032 
1033 	/*
1034 	 * Ring tail is 'txq->pending', and although descriptors
1035 	 * between 'txq->completed' and 'txq->pending' are still
1036 	 * in use by the driver, they should be reported as DONE
1037 	 */
1038 	if (unlikely(offset < (txq->added - txq->pending)))
1039 		return RTE_ETH_TX_DESC_FULL;
1040 
1041 	/*
1042 	 * There is no separate return value for unused descriptors;
1043 	 * the latter will be reported as DONE because genuine DONE
1044 	 * descriptors will be freed anyway in SW on the next burst
1045 	 */
1046 	return RTE_ETH_TX_DESC_DONE;
1047 }
1048 
1049 struct sfc_dp_tx sfc_efx_tx = {
1050 	.dp = {
1051 		.name		= SFC_KVARG_DATAPATH_EFX,
1052 		.type		= SFC_DP_TX,
1053 		.hw_fw_caps	= 0,
1054 	},
1055 	.features		= SFC_DP_TX_FEAT_VLAN_INSERT |
1056 				  SFC_DP_TX_FEAT_TSO |
1057 				  SFC_DP_TX_FEAT_MULTI_POOL |
1058 				  SFC_DP_TX_FEAT_REFCNT |
1059 				  SFC_DP_TX_FEAT_MULTI_SEG,
1060 	.qsize_up_rings		= sfc_efx_tx_qsize_up_rings,
1061 	.qcreate		= sfc_efx_tx_qcreate,
1062 	.qdestroy		= sfc_efx_tx_qdestroy,
1063 	.qstart			= sfc_efx_tx_qstart,
1064 	.qstop			= sfc_efx_tx_qstop,
1065 	.qreap			= sfc_efx_tx_qreap,
1066 	.qdesc_status		= sfc_efx_tx_qdesc_status,
1067 	.pkt_burst		= sfc_efx_xmit_pkts,
1068 };
1069