xref: /dpdk/drivers/net/sfc/sfc_tx.c (revision bc8e32473cc3978d763a1387eaa8244bcf75e77d)
1 /* SPDX-License-Identifier: BSD-3-Clause
2  *
3  * Copyright(c) 2019-2020 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 static uint64_t
38 sfc_tx_get_offload_mask(struct sfc_adapter *sa)
39 {
40 	const efx_nic_cfg_t *encp = efx_nic_cfg_get(sa->nic);
41 	uint64_t no_caps = 0;
42 
43 	if (!encp->enc_hw_tx_insert_vlan_enabled)
44 		no_caps |= DEV_TX_OFFLOAD_VLAN_INSERT;
45 
46 	if (!encp->enc_tunnel_encapsulations_supported)
47 		no_caps |= DEV_TX_OFFLOAD_OUTER_IPV4_CKSUM;
48 
49 	if (!sa->tso)
50 		no_caps |= DEV_TX_OFFLOAD_TCP_TSO;
51 
52 	if (!sa->tso_encap)
53 		no_caps |= (DEV_TX_OFFLOAD_VXLAN_TNL_TSO |
54 			    DEV_TX_OFFLOAD_GENEVE_TNL_TSO);
55 
56 	return ~no_caps;
57 }
58 
59 uint64_t
60 sfc_tx_get_dev_offload_caps(struct sfc_adapter *sa)
61 {
62 	return sa->priv.dp_tx->dev_offload_capa & sfc_tx_get_offload_mask(sa);
63 }
64 
65 uint64_t
66 sfc_tx_get_queue_offload_caps(struct sfc_adapter *sa)
67 {
68 	return sa->priv.dp_tx->queue_offload_capa & sfc_tx_get_offload_mask(sa);
69 }
70 
71 static int
72 sfc_tx_qcheck_conf(struct sfc_adapter *sa, unsigned int txq_max_fill_level,
73 		   const struct rte_eth_txconf *tx_conf,
74 		   uint64_t offloads)
75 {
76 	int rc = 0;
77 
78 	if (tx_conf->tx_rs_thresh != 0) {
79 		sfc_err(sa, "RS bit in transmit descriptor is not supported");
80 		rc = EINVAL;
81 	}
82 
83 	if (tx_conf->tx_free_thresh > txq_max_fill_level) {
84 		sfc_err(sa,
85 			"TxQ free threshold too large: %u vs maximum %u",
86 			tx_conf->tx_free_thresh, txq_max_fill_level);
87 		rc = EINVAL;
88 	}
89 
90 	if (tx_conf->tx_thresh.pthresh != 0 ||
91 	    tx_conf->tx_thresh.hthresh != 0 ||
92 	    tx_conf->tx_thresh.wthresh != 0) {
93 		sfc_warn(sa,
94 			"prefetch/host/writeback thresholds are not supported");
95 	}
96 
97 	/* We either perform both TCP and UDP offload, or no offload at all */
98 	if (((offloads & DEV_TX_OFFLOAD_TCP_CKSUM) == 0) !=
99 	    ((offloads & DEV_TX_OFFLOAD_UDP_CKSUM) == 0)) {
100 		sfc_err(sa, "TCP and UDP offloads can't be set independently");
101 		rc = EINVAL;
102 	}
103 
104 	return rc;
105 }
106 
107 void
108 sfc_tx_qflush_done(struct sfc_txq_info *txq_info)
109 {
110 	txq_info->state |= SFC_TXQ_FLUSHED;
111 	txq_info->state &= ~SFC_TXQ_FLUSHING;
112 }
113 
114 int
115 sfc_tx_qinit(struct sfc_adapter *sa, unsigned int sw_index,
116 	     uint16_t nb_tx_desc, unsigned int socket_id,
117 	     const struct rte_eth_txconf *tx_conf)
118 {
119 	const efx_nic_cfg_t *encp = efx_nic_cfg_get(sa->nic);
120 	unsigned int txq_entries;
121 	unsigned int evq_entries;
122 	unsigned int txq_max_fill_level;
123 	struct sfc_txq_info *txq_info;
124 	struct sfc_evq *evq;
125 	struct sfc_txq *txq;
126 	int rc = 0;
127 	struct sfc_dp_tx_qcreate_info info;
128 	uint64_t offloads;
129 	struct sfc_dp_tx_hw_limits hw_limits;
130 
131 	sfc_log_init(sa, "TxQ = %u", sw_index);
132 
133 	memset(&hw_limits, 0, sizeof(hw_limits));
134 	hw_limits.txq_max_entries = sa->txq_max_entries;
135 	hw_limits.txq_min_entries = sa->txq_min_entries;
136 
137 	rc = sa->priv.dp_tx->qsize_up_rings(nb_tx_desc, &hw_limits,
138 					    &txq_entries, &evq_entries,
139 					    &txq_max_fill_level);
140 	if (rc != 0)
141 		goto fail_size_up_rings;
142 	SFC_ASSERT(txq_entries >= sa->txq_min_entries);
143 	SFC_ASSERT(txq_entries <= sa->txq_max_entries);
144 	SFC_ASSERT(txq_entries >= nb_tx_desc);
145 	SFC_ASSERT(txq_max_fill_level <= nb_tx_desc);
146 
147 	offloads = tx_conf->offloads |
148 		sa->eth_dev->data->dev_conf.txmode.offloads;
149 	rc = sfc_tx_qcheck_conf(sa, txq_max_fill_level, tx_conf, offloads);
150 	if (rc != 0)
151 		goto fail_bad_conf;
152 
153 	SFC_ASSERT(sw_index < sfc_sa2shared(sa)->txq_count);
154 	txq_info = &sfc_sa2shared(sa)->txq_info[sw_index];
155 
156 	txq_info->entries = txq_entries;
157 
158 	rc = sfc_ev_qinit(sa, SFC_EVQ_TYPE_TX, sw_index,
159 			  evq_entries, socket_id, &evq);
160 	if (rc != 0)
161 		goto fail_ev_qinit;
162 
163 	txq = &sa->txq_ctrl[sw_index];
164 	txq->hw_index = sw_index;
165 	txq->evq = evq;
166 	txq_info->free_thresh =
167 		(tx_conf->tx_free_thresh) ? tx_conf->tx_free_thresh :
168 		SFC_TX_DEFAULT_FREE_THRESH;
169 	txq_info->offloads = offloads;
170 
171 	rc = sfc_dma_alloc(sa, "txq", sw_index,
172 			   efx_txq_size(sa->nic, txq_info->entries),
173 			   socket_id, &txq->mem);
174 	if (rc != 0)
175 		goto fail_dma_alloc;
176 
177 	memset(&info, 0, sizeof(info));
178 	info.max_fill_level = txq_max_fill_level;
179 	info.free_thresh = txq_info->free_thresh;
180 	info.offloads = offloads;
181 	info.txq_entries = txq_info->entries;
182 	info.dma_desc_size_max = encp->enc_tx_dma_desc_size_max;
183 	info.txq_hw_ring = txq->mem.esm_base;
184 	info.evq_entries = evq_entries;
185 	info.evq_hw_ring = evq->mem.esm_base;
186 	info.hw_index = txq->hw_index;
187 	info.mem_bar = sa->mem_bar.esb_base;
188 	info.vi_window_shift = encp->enc_vi_window_shift;
189 	info.tso_tcp_header_offset_limit =
190 		encp->enc_tx_tso_tcp_header_offset_limit;
191 	info.tso_max_nb_header_descs =
192 		RTE_MIN(encp->enc_tx_tso_max_header_ndescs,
193 			(uint32_t)UINT16_MAX);
194 	info.tso_max_header_len =
195 		RTE_MIN(encp->enc_tx_tso_max_header_length,
196 			(uint32_t)UINT16_MAX);
197 	info.tso_max_nb_payload_descs =
198 		RTE_MIN(encp->enc_tx_tso_max_payload_ndescs,
199 			(uint32_t)UINT16_MAX);
200 	info.tso_max_payload_len = encp->enc_tx_tso_max_payload_length;
201 	info.tso_max_nb_outgoing_frames = encp->enc_tx_tso_max_nframes;
202 
203 	rc = sa->priv.dp_tx->qcreate(sa->eth_dev->data->port_id, sw_index,
204 				     &RTE_ETH_DEV_TO_PCI(sa->eth_dev)->addr,
205 				     socket_id, &info, &txq_info->dp);
206 	if (rc != 0)
207 		goto fail_dp_tx_qinit;
208 
209 	evq->dp_txq = txq_info->dp;
210 
211 	txq_info->state = SFC_TXQ_INITIALIZED;
212 
213 	txq_info->deferred_start = (tx_conf->tx_deferred_start != 0);
214 
215 	return 0;
216 
217 fail_dp_tx_qinit:
218 	sfc_dma_free(sa, &txq->mem);
219 
220 fail_dma_alloc:
221 	sfc_ev_qfini(evq);
222 
223 fail_ev_qinit:
224 	txq_info->entries = 0;
225 
226 fail_bad_conf:
227 fail_size_up_rings:
228 	sfc_log_init(sa, "failed (TxQ = %u, rc = %d)", sw_index, rc);
229 	return rc;
230 }
231 
232 void
233 sfc_tx_qfini(struct sfc_adapter *sa, unsigned int sw_index)
234 {
235 	struct sfc_txq_info *txq_info;
236 	struct sfc_txq *txq;
237 
238 	sfc_log_init(sa, "TxQ = %u", sw_index);
239 
240 	SFC_ASSERT(sw_index < sfc_sa2shared(sa)->txq_count);
241 	sa->eth_dev->data->tx_queues[sw_index] = NULL;
242 
243 	txq_info = &sfc_sa2shared(sa)->txq_info[sw_index];
244 
245 	SFC_ASSERT(txq_info->state == SFC_TXQ_INITIALIZED);
246 
247 	sa->priv.dp_tx->qdestroy(txq_info->dp);
248 	txq_info->dp = NULL;
249 
250 	txq_info->state &= ~SFC_TXQ_INITIALIZED;
251 	txq_info->entries = 0;
252 
253 	txq = &sa->txq_ctrl[sw_index];
254 
255 	sfc_dma_free(sa, &txq->mem);
256 
257 	sfc_ev_qfini(txq->evq);
258 	txq->evq = NULL;
259 }
260 
261 static int
262 sfc_tx_qinit_info(struct sfc_adapter *sa, unsigned int sw_index)
263 {
264 	sfc_log_init(sa, "TxQ = %u", sw_index);
265 
266 	return 0;
267 }
268 
269 static int
270 sfc_tx_check_mode(struct sfc_adapter *sa, const struct rte_eth_txmode *txmode)
271 {
272 	int rc = 0;
273 
274 	switch (txmode->mq_mode) {
275 	case ETH_MQ_TX_NONE:
276 		break;
277 	default:
278 		sfc_err(sa, "Tx multi-queue mode %u not supported",
279 			txmode->mq_mode);
280 		rc = EINVAL;
281 	}
282 
283 	/*
284 	 * These features are claimed to be i40e-specific,
285 	 * but it does make sense to double-check their absence
286 	 */
287 	if (txmode->hw_vlan_reject_tagged) {
288 		sfc_err(sa, "Rejecting tagged packets not supported");
289 		rc = EINVAL;
290 	}
291 
292 	if (txmode->hw_vlan_reject_untagged) {
293 		sfc_err(sa, "Rejecting untagged packets not supported");
294 		rc = EINVAL;
295 	}
296 
297 	if (txmode->hw_vlan_insert_pvid) {
298 		sfc_err(sa, "Port-based VLAN insertion not supported");
299 		rc = EINVAL;
300 	}
301 
302 	return rc;
303 }
304 
305 /**
306  * Destroy excess queues that are no longer needed after reconfiguration
307  * or complete close.
308  */
309 static void
310 sfc_tx_fini_queues(struct sfc_adapter *sa, unsigned int nb_tx_queues)
311 {
312 	struct sfc_adapter_shared * const sas = sfc_sa2shared(sa);
313 	int sw_index;
314 
315 	SFC_ASSERT(nb_tx_queues <= sas->txq_count);
316 
317 	sw_index = sas->txq_count;
318 	while (--sw_index >= (int)nb_tx_queues) {
319 		if (sas->txq_info[sw_index].state & SFC_TXQ_INITIALIZED)
320 			sfc_tx_qfini(sa, sw_index);
321 	}
322 
323 	sas->txq_count = nb_tx_queues;
324 }
325 
326 int
327 sfc_tx_configure(struct sfc_adapter *sa)
328 {
329 	struct sfc_adapter_shared * const sas = sfc_sa2shared(sa);
330 	const efx_nic_cfg_t *encp = efx_nic_cfg_get(sa->nic);
331 	const struct rte_eth_conf *dev_conf = &sa->eth_dev->data->dev_conf;
332 	const unsigned int nb_tx_queues = sa->eth_dev->data->nb_tx_queues;
333 	int rc = 0;
334 
335 	sfc_log_init(sa, "nb_tx_queues=%u (old %u)",
336 		     nb_tx_queues, sas->txq_count);
337 
338 	/*
339 	 * The datapath implementation assumes absence of boundary
340 	 * limits on Tx DMA descriptors. Addition of these checks on
341 	 * datapath would simply make the datapath slower.
342 	 */
343 	if (encp->enc_tx_dma_desc_boundary != 0) {
344 		rc = ENOTSUP;
345 		goto fail_tx_dma_desc_boundary;
346 	}
347 
348 	rc = sfc_tx_check_mode(sa, &dev_conf->txmode);
349 	if (rc != 0)
350 		goto fail_check_mode;
351 
352 	if (nb_tx_queues == sas->txq_count)
353 		goto done;
354 
355 	if (sas->txq_info == NULL) {
356 		sas->txq_info = rte_calloc_socket("sfc-txqs", nb_tx_queues,
357 						  sizeof(sas->txq_info[0]), 0,
358 						  sa->socket_id);
359 		if (sas->txq_info == NULL)
360 			goto fail_txqs_alloc;
361 
362 		/*
363 		 * Allocate primary process only TxQ control from heap
364 		 * since it should not be shared.
365 		 */
366 		rc = ENOMEM;
367 		sa->txq_ctrl = calloc(nb_tx_queues, sizeof(sa->txq_ctrl[0]));
368 		if (sa->txq_ctrl == NULL)
369 			goto fail_txqs_ctrl_alloc;
370 	} else {
371 		struct sfc_txq_info *new_txq_info;
372 		struct sfc_txq *new_txq_ctrl;
373 
374 		if (nb_tx_queues < sas->txq_count)
375 			sfc_tx_fini_queues(sa, nb_tx_queues);
376 
377 		new_txq_info =
378 			rte_realloc(sas->txq_info,
379 				    nb_tx_queues * sizeof(sas->txq_info[0]), 0);
380 		if (new_txq_info == NULL && nb_tx_queues > 0)
381 			goto fail_txqs_realloc;
382 
383 		new_txq_ctrl = realloc(sa->txq_ctrl,
384 				       nb_tx_queues * sizeof(sa->txq_ctrl[0]));
385 		if (new_txq_ctrl == NULL && nb_tx_queues > 0)
386 			goto fail_txqs_ctrl_realloc;
387 
388 		sas->txq_info = new_txq_info;
389 		sa->txq_ctrl = new_txq_ctrl;
390 		if (nb_tx_queues > sas->txq_count) {
391 			memset(&sas->txq_info[sas->txq_count], 0,
392 			       (nb_tx_queues - sas->txq_count) *
393 			       sizeof(sas->txq_info[0]));
394 			memset(&sa->txq_ctrl[sas->txq_count], 0,
395 			       (nb_tx_queues - sas->txq_count) *
396 			       sizeof(sa->txq_ctrl[0]));
397 		}
398 	}
399 
400 	while (sas->txq_count < nb_tx_queues) {
401 		rc = sfc_tx_qinit_info(sa, sas->txq_count);
402 		if (rc != 0)
403 			goto fail_tx_qinit_info;
404 
405 		sas->txq_count++;
406 	}
407 
408 done:
409 	return 0;
410 
411 fail_tx_qinit_info:
412 fail_txqs_ctrl_realloc:
413 fail_txqs_realloc:
414 fail_txqs_ctrl_alloc:
415 fail_txqs_alloc:
416 	sfc_tx_close(sa);
417 
418 fail_check_mode:
419 fail_tx_dma_desc_boundary:
420 	sfc_log_init(sa, "failed (rc = %d)", rc);
421 	return rc;
422 }
423 
424 void
425 sfc_tx_close(struct sfc_adapter *sa)
426 {
427 	sfc_tx_fini_queues(sa, 0);
428 
429 	free(sa->txq_ctrl);
430 	sa->txq_ctrl = NULL;
431 
432 	rte_free(sfc_sa2shared(sa)->txq_info);
433 	sfc_sa2shared(sa)->txq_info = NULL;
434 }
435 
436 int
437 sfc_tx_qstart(struct sfc_adapter *sa, unsigned int sw_index)
438 {
439 	struct sfc_adapter_shared * const sas = sfc_sa2shared(sa);
440 	uint64_t offloads_supported = sfc_tx_get_dev_offload_caps(sa) |
441 				      sfc_tx_get_queue_offload_caps(sa);
442 	struct rte_eth_dev_data *dev_data;
443 	struct sfc_txq_info *txq_info;
444 	struct sfc_txq *txq;
445 	struct sfc_evq *evq;
446 	uint16_t flags = 0;
447 	unsigned int desc_index;
448 	int rc = 0;
449 
450 	sfc_log_init(sa, "TxQ = %u", sw_index);
451 
452 	SFC_ASSERT(sw_index < sas->txq_count);
453 	txq_info = &sas->txq_info[sw_index];
454 
455 	SFC_ASSERT(txq_info->state == SFC_TXQ_INITIALIZED);
456 
457 	txq = &sa->txq_ctrl[sw_index];
458 	evq = txq->evq;
459 
460 	rc = sfc_ev_qstart(evq, sfc_evq_index_by_txq_sw_index(sa, sw_index));
461 	if (rc != 0)
462 		goto fail_ev_qstart;
463 
464 	if (txq_info->offloads & DEV_TX_OFFLOAD_IPV4_CKSUM)
465 		flags |= EFX_TXQ_CKSUM_IPV4;
466 
467 	if (txq_info->offloads & DEV_TX_OFFLOAD_OUTER_IPV4_CKSUM)
468 		flags |= EFX_TXQ_CKSUM_INNER_IPV4;
469 
470 	if ((txq_info->offloads & DEV_TX_OFFLOAD_TCP_CKSUM) ||
471 	    (txq_info->offloads & DEV_TX_OFFLOAD_UDP_CKSUM)) {
472 		flags |= EFX_TXQ_CKSUM_TCPUDP;
473 
474 		if (offloads_supported & DEV_TX_OFFLOAD_OUTER_IPV4_CKSUM)
475 			flags |= EFX_TXQ_CKSUM_INNER_TCPUDP;
476 	}
477 
478 	if (txq_info->offloads & (DEV_TX_OFFLOAD_TCP_TSO |
479 				  DEV_TX_OFFLOAD_VXLAN_TNL_TSO |
480 				  DEV_TX_OFFLOAD_GENEVE_TNL_TSO))
481 		flags |= EFX_TXQ_FATSOV2;
482 
483 	rc = efx_tx_qcreate(sa->nic, txq->hw_index, 0, &txq->mem,
484 			    txq_info->entries, 0 /* not used on EF10 */,
485 			    flags, evq->common,
486 			    &txq->common, &desc_index);
487 	if (rc != 0) {
488 		if (sa->tso && (rc == ENOSPC))
489 			sfc_err(sa, "ran out of TSO contexts");
490 
491 		goto fail_tx_qcreate;
492 	}
493 
494 	efx_tx_qenable(txq->common);
495 
496 	txq_info->state |= SFC_TXQ_STARTED;
497 
498 	rc = sa->priv.dp_tx->qstart(txq_info->dp, evq->read_ptr, desc_index);
499 	if (rc != 0)
500 		goto fail_dp_qstart;
501 
502 	/*
503 	 * It seems to be used by DPDK for debug purposes only ('rte_ether')
504 	 */
505 	dev_data = sa->eth_dev->data;
506 	dev_data->tx_queue_state[sw_index] = RTE_ETH_QUEUE_STATE_STARTED;
507 
508 	return 0;
509 
510 fail_dp_qstart:
511 	txq_info->state = SFC_TXQ_INITIALIZED;
512 	efx_tx_qdestroy(txq->common);
513 
514 fail_tx_qcreate:
515 	sfc_ev_qstop(evq);
516 
517 fail_ev_qstart:
518 	return rc;
519 }
520 
521 void
522 sfc_tx_qstop(struct sfc_adapter *sa, unsigned int sw_index)
523 {
524 	struct sfc_adapter_shared * const sas = sfc_sa2shared(sa);
525 	struct rte_eth_dev_data *dev_data;
526 	struct sfc_txq_info *txq_info;
527 	struct sfc_txq *txq;
528 	unsigned int retry_count;
529 	unsigned int wait_count;
530 	int rc;
531 
532 	sfc_log_init(sa, "TxQ = %u", sw_index);
533 
534 	SFC_ASSERT(sw_index < sas->txq_count);
535 	txq_info = &sas->txq_info[sw_index];
536 
537 	if (txq_info->state == SFC_TXQ_INITIALIZED)
538 		return;
539 
540 	SFC_ASSERT(txq_info->state & SFC_TXQ_STARTED);
541 
542 	txq = &sa->txq_ctrl[sw_index];
543 	sa->priv.dp_tx->qstop(txq_info->dp, &txq->evq->read_ptr);
544 
545 	/*
546 	 * Retry TX queue flushing in case of flush failed or
547 	 * timeout; in the worst case it can delay for 6 seconds
548 	 */
549 	for (retry_count = 0;
550 	     ((txq_info->state & SFC_TXQ_FLUSHED) == 0) &&
551 	     (retry_count < SFC_TX_QFLUSH_ATTEMPTS);
552 	     ++retry_count) {
553 		rc = efx_tx_qflush(txq->common);
554 		if (rc != 0) {
555 			txq_info->state |= (rc == EALREADY) ?
556 				SFC_TXQ_FLUSHED : SFC_TXQ_FLUSH_FAILED;
557 			break;
558 		}
559 
560 		/*
561 		 * Wait for TX queue flush done or flush failed event at least
562 		 * SFC_TX_QFLUSH_POLL_WAIT_MS milliseconds and not more
563 		 * than 2 seconds (SFC_TX_QFLUSH_POLL_WAIT_MS multiplied
564 		 * by SFC_TX_QFLUSH_POLL_ATTEMPTS)
565 		 */
566 		wait_count = 0;
567 		do {
568 			rte_delay_ms(SFC_TX_QFLUSH_POLL_WAIT_MS);
569 			sfc_ev_qpoll(txq->evq);
570 		} while ((txq_info->state & SFC_TXQ_FLUSHING) &&
571 			 wait_count++ < SFC_TX_QFLUSH_POLL_ATTEMPTS);
572 
573 		if (txq_info->state & SFC_TXQ_FLUSHING)
574 			sfc_err(sa, "TxQ %u flush timed out", sw_index);
575 
576 		if (txq_info->state & SFC_TXQ_FLUSHED)
577 			sfc_notice(sa, "TxQ %u flushed", sw_index);
578 	}
579 
580 	sa->priv.dp_tx->qreap(txq_info->dp);
581 
582 	txq_info->state = SFC_TXQ_INITIALIZED;
583 
584 	efx_tx_qdestroy(txq->common);
585 
586 	sfc_ev_qstop(txq->evq);
587 
588 	/*
589 	 * It seems to be used by DPDK for debug purposes only ('rte_ether')
590 	 */
591 	dev_data = sa->eth_dev->data;
592 	dev_data->tx_queue_state[sw_index] = RTE_ETH_QUEUE_STATE_STOPPED;
593 }
594 
595 int
596 sfc_tx_start(struct sfc_adapter *sa)
597 {
598 	struct sfc_adapter_shared * const sas = sfc_sa2shared(sa);
599 	const efx_nic_cfg_t *encp = efx_nic_cfg_get(sa->nic);
600 	unsigned int sw_index;
601 	int rc = 0;
602 
603 	sfc_log_init(sa, "txq_count = %u", sas->txq_count);
604 
605 	if (sa->tso) {
606 		if (!encp->enc_fw_assisted_tso_v2_enabled &&
607 		    !encp->enc_tso_v3_enabled) {
608 			sfc_warn(sa, "TSO support was unable to be restored");
609 			sa->tso = B_FALSE;
610 			sa->tso_encap = B_FALSE;
611 		}
612 	}
613 
614 	if (sa->tso_encap && !encp->enc_fw_assisted_tso_v2_encap_enabled &&
615 	    !encp->enc_tso_v3_enabled) {
616 		sfc_warn(sa, "Encapsulated TSO support was unable to be restored");
617 		sa->tso_encap = B_FALSE;
618 	}
619 
620 	rc = efx_tx_init(sa->nic);
621 	if (rc != 0)
622 		goto fail_efx_tx_init;
623 
624 	for (sw_index = 0; sw_index < sas->txq_count; ++sw_index) {
625 		if (sas->txq_info[sw_index].state == SFC_TXQ_INITIALIZED &&
626 		    (!(sas->txq_info[sw_index].deferred_start) ||
627 		     sas->txq_info[sw_index].deferred_started)) {
628 			rc = sfc_tx_qstart(sa, sw_index);
629 			if (rc != 0)
630 				goto fail_tx_qstart;
631 		}
632 	}
633 
634 	return 0;
635 
636 fail_tx_qstart:
637 	while (sw_index-- > 0)
638 		sfc_tx_qstop(sa, sw_index);
639 
640 	efx_tx_fini(sa->nic);
641 
642 fail_efx_tx_init:
643 	sfc_log_init(sa, "failed (rc = %d)", rc);
644 	return rc;
645 }
646 
647 void
648 sfc_tx_stop(struct sfc_adapter *sa)
649 {
650 	struct sfc_adapter_shared * const sas = sfc_sa2shared(sa);
651 	unsigned int sw_index;
652 
653 	sfc_log_init(sa, "txq_count = %u", sas->txq_count);
654 
655 	sw_index = sas->txq_count;
656 	while (sw_index-- > 0) {
657 		if (sas->txq_info[sw_index].state & SFC_TXQ_STARTED)
658 			sfc_tx_qstop(sa, sw_index);
659 	}
660 
661 	efx_tx_fini(sa->nic);
662 }
663 
664 static void
665 sfc_efx_tx_reap(struct sfc_efx_txq *txq)
666 {
667 	unsigned int completed;
668 
669 	sfc_ev_qpoll(txq->evq);
670 
671 	for (completed = txq->completed;
672 	     completed != txq->pending; completed++) {
673 		struct sfc_efx_tx_sw_desc *txd;
674 
675 		txd = &txq->sw_ring[completed & txq->ptr_mask];
676 
677 		if (txd->mbuf != NULL) {
678 			rte_pktmbuf_free(txd->mbuf);
679 			txd->mbuf = NULL;
680 		}
681 	}
682 
683 	txq->completed = completed;
684 }
685 
686 /*
687  * The function is used to insert or update VLAN tag;
688  * the firmware has state of the firmware tag to insert per TxQ
689  * (controlled by option descriptors), hence, if the tag of the
690  * packet to be sent is different from one remembered by the firmware,
691  * the function will update it
692  */
693 static unsigned int
694 sfc_efx_tx_maybe_insert_tag(struct sfc_efx_txq *txq, struct rte_mbuf *m,
695 			    efx_desc_t **pend)
696 {
697 	uint16_t this_tag = ((m->ol_flags & PKT_TX_VLAN_PKT) ?
698 			     m->vlan_tci : 0);
699 
700 	if (this_tag == txq->hw_vlan_tci)
701 		return 0;
702 
703 	/*
704 	 * The expression inside SFC_ASSERT() is not desired to be checked in
705 	 * a non-debug build because it might be too expensive on the data path
706 	 */
707 	SFC_ASSERT(efx_nic_cfg_get(txq->evq->sa->nic)->enc_hw_tx_insert_vlan_enabled);
708 
709 	efx_tx_qdesc_vlantci_create(txq->common, rte_cpu_to_be_16(this_tag),
710 				    *pend);
711 	(*pend)++;
712 	txq->hw_vlan_tci = this_tag;
713 
714 	return 1;
715 }
716 
717 static uint16_t
718 sfc_efx_prepare_pkts(void *tx_queue, struct rte_mbuf **tx_pkts,
719 		     uint16_t nb_pkts)
720 {
721 	struct sfc_dp_txq *dp_txq = tx_queue;
722 	struct sfc_efx_txq *txq = sfc_efx_txq_by_dp_txq(dp_txq);
723 	const efx_nic_cfg_t *encp = efx_nic_cfg_get(txq->evq->sa->nic);
724 	uint16_t i;
725 
726 	for (i = 0; i < nb_pkts; i++) {
727 		int ret;
728 
729 		/*
730 		 * EFX Tx datapath may require extra VLAN descriptor if VLAN
731 		 * insertion offload is requested regardless the offload
732 		 * requested/supported.
733 		 */
734 		ret = sfc_dp_tx_prepare_pkt(tx_pkts[i], 0, SFC_TSOH_STD_LEN,
735 				encp->enc_tx_tso_tcp_header_offset_limit,
736 				txq->max_fill_level, EFX_TX_FATSOV2_OPT_NDESCS,
737 				1);
738 		if (unlikely(ret != 0)) {
739 			rte_errno = ret;
740 			break;
741 		}
742 	}
743 
744 	return i;
745 }
746 
747 static uint16_t
748 sfc_efx_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
749 {
750 	struct sfc_dp_txq *dp_txq = (struct sfc_dp_txq *)tx_queue;
751 	struct sfc_efx_txq *txq = sfc_efx_txq_by_dp_txq(dp_txq);
752 	unsigned int added = txq->added;
753 	unsigned int pushed = added;
754 	unsigned int pkts_sent = 0;
755 	efx_desc_t *pend = &txq->pend_desc[0];
756 	const unsigned int hard_max_fill = txq->max_fill_level;
757 	const unsigned int soft_max_fill = hard_max_fill - txq->free_thresh;
758 	unsigned int fill_level = added - txq->completed;
759 	boolean_t reap_done;
760 	int rc __rte_unused;
761 	struct rte_mbuf **pktp;
762 
763 	if (unlikely((txq->flags & SFC_EFX_TXQ_FLAG_RUNNING) == 0))
764 		goto done;
765 
766 	/*
767 	 * If insufficient space for a single packet is present,
768 	 * we should reap; otherwise, we shouldn't do that all the time
769 	 * to avoid latency increase
770 	 */
771 	reap_done = (fill_level > soft_max_fill);
772 
773 	if (reap_done) {
774 		sfc_efx_tx_reap(txq);
775 		/*
776 		 * Recalculate fill level since 'txq->completed'
777 		 * might have changed on reap
778 		 */
779 		fill_level = added - txq->completed;
780 	}
781 
782 	for (pkts_sent = 0, pktp = &tx_pkts[0];
783 	     (pkts_sent < nb_pkts) && (fill_level <= soft_max_fill);
784 	     pkts_sent++, pktp++) {
785 		uint16_t		hw_vlan_tci_prev = txq->hw_vlan_tci;
786 		struct rte_mbuf		*m_seg = *pktp;
787 		size_t			pkt_len = m_seg->pkt_len;
788 		unsigned int		pkt_descs = 0;
789 		size_t			in_off = 0;
790 
791 		/*
792 		 * Here VLAN TCI is expected to be zero in case if no
793 		 * DEV_TX_OFFLOAD_VLAN_INSERT capability is advertised;
794 		 * if the calling app ignores the absence of
795 		 * DEV_TX_OFFLOAD_VLAN_INSERT and pushes VLAN TCI, then
796 		 * TX_ERROR will occur
797 		 */
798 		pkt_descs += sfc_efx_tx_maybe_insert_tag(txq, m_seg, &pend);
799 
800 		if (m_seg->ol_flags & PKT_TX_TCP_SEG) {
801 			/*
802 			 * We expect correct 'pkt->l[2, 3, 4]_len' values
803 			 * to be set correctly by the caller
804 			 */
805 			if (sfc_efx_tso_do(txq, added, &m_seg, &in_off, &pend,
806 					   &pkt_descs, &pkt_len) != 0) {
807 				/* We may have reached this place if packet
808 				 * header linearization is needed but the
809 				 * header length is greater than
810 				 * SFC_TSOH_STD_LEN
811 				 *
812 				 * We will deceive RTE saying that we have sent
813 				 * the packet, but we will actually drop it.
814 				 * Hence, we should revert 'pend' to the
815 				 * previous state (in case we have added
816 				 * VLAN descriptor) and start processing
817 				 * another one packet. But the original
818 				 * mbuf shouldn't be orphaned
819 				 */
820 				pend -= pkt_descs;
821 				txq->hw_vlan_tci = hw_vlan_tci_prev;
822 
823 				rte_pktmbuf_free(*pktp);
824 
825 				continue;
826 			}
827 
828 			/*
829 			 * We've only added 2 FATSOv2 option descriptors
830 			 * and 1 descriptor for the linearized packet header.
831 			 * The outstanding work will be done in the same manner
832 			 * as for the usual non-TSO path
833 			 */
834 		}
835 
836 		for (; m_seg != NULL; m_seg = m_seg->next) {
837 			efsys_dma_addr_t	next_frag;
838 			size_t			seg_len;
839 
840 			seg_len = m_seg->data_len;
841 			next_frag = rte_mbuf_data_iova(m_seg);
842 
843 			/*
844 			 * If we've started TSO transaction few steps earlier,
845 			 * we'll skip packet header using an offset in the
846 			 * current segment (which has been set to the
847 			 * first one containing payload)
848 			 */
849 			seg_len -= in_off;
850 			next_frag += in_off;
851 			in_off = 0;
852 
853 			do {
854 				efsys_dma_addr_t	frag_addr = next_frag;
855 				size_t			frag_len;
856 
857 				/*
858 				 * It is assumed here that there is no
859 				 * limitation on address boundary
860 				 * crossing by DMA descriptor.
861 				 */
862 				frag_len = MIN(seg_len, txq->dma_desc_size_max);
863 				next_frag += frag_len;
864 				seg_len -= frag_len;
865 				pkt_len -= frag_len;
866 
867 				efx_tx_qdesc_dma_create(txq->common,
868 							frag_addr, frag_len,
869 							(pkt_len == 0),
870 							pend++);
871 
872 				pkt_descs++;
873 			} while (seg_len != 0);
874 		}
875 
876 		added += pkt_descs;
877 
878 		fill_level += pkt_descs;
879 		if (unlikely(fill_level > hard_max_fill)) {
880 			/*
881 			 * Our estimation for maximum number of descriptors
882 			 * required to send a packet seems to be wrong.
883 			 * Try to reap (if we haven't yet).
884 			 */
885 			if (!reap_done) {
886 				sfc_efx_tx_reap(txq);
887 				reap_done = B_TRUE;
888 				fill_level = added - txq->completed;
889 				if (fill_level > hard_max_fill) {
890 					pend -= pkt_descs;
891 					txq->hw_vlan_tci = hw_vlan_tci_prev;
892 					break;
893 				}
894 			} else {
895 				pend -= pkt_descs;
896 				txq->hw_vlan_tci = hw_vlan_tci_prev;
897 				break;
898 			}
899 		}
900 
901 		/* Assign mbuf to the last used desc */
902 		txq->sw_ring[(added - 1) & txq->ptr_mask].mbuf = *pktp;
903 	}
904 
905 	if (likely(pkts_sent > 0)) {
906 		rc = efx_tx_qdesc_post(txq->common, txq->pend_desc,
907 				       pend - &txq->pend_desc[0],
908 				       txq->completed, &txq->added);
909 		SFC_ASSERT(rc == 0);
910 
911 		if (likely(pushed != txq->added))
912 			efx_tx_qpush(txq->common, txq->added, pushed);
913 	}
914 
915 #if SFC_TX_XMIT_PKTS_REAP_AT_LEAST_ONCE
916 	if (!reap_done)
917 		sfc_efx_tx_reap(txq);
918 #endif
919 
920 done:
921 	return pkts_sent;
922 }
923 
924 const struct sfc_dp_tx *
925 sfc_dp_tx_by_dp_txq(const struct sfc_dp_txq *dp_txq)
926 {
927 	const struct sfc_dp_queue *dpq = &dp_txq->dpq;
928 	struct rte_eth_dev *eth_dev;
929 	struct sfc_adapter_priv *sap;
930 
931 	SFC_ASSERT(rte_eth_dev_is_valid_port(dpq->port_id));
932 	eth_dev = &rte_eth_devices[dpq->port_id];
933 
934 	sap = sfc_adapter_priv_by_eth_dev(eth_dev);
935 
936 	return sap->dp_tx;
937 }
938 
939 struct sfc_txq_info *
940 sfc_txq_info_by_dp_txq(const struct sfc_dp_txq *dp_txq)
941 {
942 	const struct sfc_dp_queue *dpq = &dp_txq->dpq;
943 	struct rte_eth_dev *eth_dev;
944 	struct sfc_adapter_shared *sas;
945 
946 	SFC_ASSERT(rte_eth_dev_is_valid_port(dpq->port_id));
947 	eth_dev = &rte_eth_devices[dpq->port_id];
948 
949 	sas = sfc_adapter_shared_by_eth_dev(eth_dev);
950 
951 	SFC_ASSERT(dpq->queue_id < sas->txq_count);
952 	return &sas->txq_info[dpq->queue_id];
953 }
954 
955 struct sfc_txq *
956 sfc_txq_by_dp_txq(const struct sfc_dp_txq *dp_txq)
957 {
958 	const struct sfc_dp_queue *dpq = &dp_txq->dpq;
959 	struct rte_eth_dev *eth_dev;
960 	struct sfc_adapter *sa;
961 
962 	SFC_ASSERT(rte_eth_dev_is_valid_port(dpq->port_id));
963 	eth_dev = &rte_eth_devices[dpq->port_id];
964 
965 	sa = sfc_adapter_by_eth_dev(eth_dev);
966 
967 	SFC_ASSERT(dpq->queue_id < sfc_sa2shared(sa)->txq_count);
968 	return &sa->txq_ctrl[dpq->queue_id];
969 }
970 
971 static sfc_dp_tx_qsize_up_rings_t sfc_efx_tx_qsize_up_rings;
972 static int
973 sfc_efx_tx_qsize_up_rings(uint16_t nb_tx_desc,
974 			  __rte_unused struct sfc_dp_tx_hw_limits *limits,
975 			  unsigned int *txq_entries,
976 			  unsigned int *evq_entries,
977 			  unsigned int *txq_max_fill_level)
978 {
979 	*txq_entries = nb_tx_desc;
980 	*evq_entries = nb_tx_desc;
981 	*txq_max_fill_level = EFX_TXQ_LIMIT(*txq_entries);
982 	return 0;
983 }
984 
985 static sfc_dp_tx_qcreate_t sfc_efx_tx_qcreate;
986 static int
987 sfc_efx_tx_qcreate(uint16_t port_id, uint16_t queue_id,
988 		   const struct rte_pci_addr *pci_addr,
989 		   int socket_id,
990 		   const struct sfc_dp_tx_qcreate_info *info,
991 		   struct sfc_dp_txq **dp_txqp)
992 {
993 	struct sfc_efx_txq *txq;
994 	struct sfc_txq *ctrl_txq;
995 	int rc;
996 
997 	rc = ENOMEM;
998 	txq = rte_zmalloc_socket("sfc-efx-txq", sizeof(*txq),
999 				 RTE_CACHE_LINE_SIZE, socket_id);
1000 	if (txq == NULL)
1001 		goto fail_txq_alloc;
1002 
1003 	sfc_dp_queue_init(&txq->dp.dpq, port_id, queue_id, pci_addr);
1004 
1005 	rc = ENOMEM;
1006 	txq->pend_desc = rte_calloc_socket("sfc-efx-txq-pend-desc",
1007 					   EFX_TXQ_LIMIT(info->txq_entries),
1008 					   sizeof(*txq->pend_desc), 0,
1009 					   socket_id);
1010 	if (txq->pend_desc == NULL)
1011 		goto fail_pend_desc_alloc;
1012 
1013 	rc = ENOMEM;
1014 	txq->sw_ring = rte_calloc_socket("sfc-efx-txq-sw_ring",
1015 					 info->txq_entries,
1016 					 sizeof(*txq->sw_ring),
1017 					 RTE_CACHE_LINE_SIZE, socket_id);
1018 	if (txq->sw_ring == NULL)
1019 		goto fail_sw_ring_alloc;
1020 
1021 	ctrl_txq = sfc_txq_by_dp_txq(&txq->dp);
1022 	if (ctrl_txq->evq->sa->tso) {
1023 		rc = sfc_efx_tso_alloc_tsoh_objs(txq->sw_ring,
1024 						 info->txq_entries, socket_id);
1025 		if (rc != 0)
1026 			goto fail_alloc_tsoh_objs;
1027 	}
1028 
1029 	txq->evq = ctrl_txq->evq;
1030 	txq->ptr_mask = info->txq_entries - 1;
1031 	txq->max_fill_level = info->max_fill_level;
1032 	txq->free_thresh = info->free_thresh;
1033 	txq->dma_desc_size_max = info->dma_desc_size_max;
1034 
1035 	*dp_txqp = &txq->dp;
1036 	return 0;
1037 
1038 fail_alloc_tsoh_objs:
1039 	rte_free(txq->sw_ring);
1040 
1041 fail_sw_ring_alloc:
1042 	rte_free(txq->pend_desc);
1043 
1044 fail_pend_desc_alloc:
1045 	rte_free(txq);
1046 
1047 fail_txq_alloc:
1048 	return rc;
1049 }
1050 
1051 static sfc_dp_tx_qdestroy_t sfc_efx_tx_qdestroy;
1052 static void
1053 sfc_efx_tx_qdestroy(struct sfc_dp_txq *dp_txq)
1054 {
1055 	struct sfc_efx_txq *txq = sfc_efx_txq_by_dp_txq(dp_txq);
1056 
1057 	sfc_efx_tso_free_tsoh_objs(txq->sw_ring, txq->ptr_mask + 1);
1058 	rte_free(txq->sw_ring);
1059 	rte_free(txq->pend_desc);
1060 	rte_free(txq);
1061 }
1062 
1063 static sfc_dp_tx_qstart_t sfc_efx_tx_qstart;
1064 static int
1065 sfc_efx_tx_qstart(struct sfc_dp_txq *dp_txq,
1066 		  __rte_unused unsigned int evq_read_ptr,
1067 		  unsigned int txq_desc_index)
1068 {
1069 	/* libefx-based datapath is specific to libefx-based PMD */
1070 	struct sfc_efx_txq *txq = sfc_efx_txq_by_dp_txq(dp_txq);
1071 	struct sfc_txq *ctrl_txq = sfc_txq_by_dp_txq(dp_txq);
1072 
1073 	txq->common = ctrl_txq->common;
1074 
1075 	txq->pending = txq->completed = txq->added = txq_desc_index;
1076 	txq->hw_vlan_tci = 0;
1077 
1078 	txq->flags |= (SFC_EFX_TXQ_FLAG_STARTED | SFC_EFX_TXQ_FLAG_RUNNING);
1079 
1080 	return 0;
1081 }
1082 
1083 static sfc_dp_tx_qstop_t sfc_efx_tx_qstop;
1084 static void
1085 sfc_efx_tx_qstop(struct sfc_dp_txq *dp_txq,
1086 		 __rte_unused unsigned int *evq_read_ptr)
1087 {
1088 	struct sfc_efx_txq *txq = sfc_efx_txq_by_dp_txq(dp_txq);
1089 
1090 	txq->flags &= ~SFC_EFX_TXQ_FLAG_RUNNING;
1091 }
1092 
1093 static sfc_dp_tx_qreap_t sfc_efx_tx_qreap;
1094 static void
1095 sfc_efx_tx_qreap(struct sfc_dp_txq *dp_txq)
1096 {
1097 	struct sfc_efx_txq *txq = sfc_efx_txq_by_dp_txq(dp_txq);
1098 	unsigned int txds;
1099 
1100 	sfc_efx_tx_reap(txq);
1101 
1102 	for (txds = 0; txds <= txq->ptr_mask; txds++) {
1103 		if (txq->sw_ring[txds].mbuf != NULL) {
1104 			rte_pktmbuf_free(txq->sw_ring[txds].mbuf);
1105 			txq->sw_ring[txds].mbuf = NULL;
1106 		}
1107 	}
1108 
1109 	txq->flags &= ~SFC_EFX_TXQ_FLAG_STARTED;
1110 }
1111 
1112 static sfc_dp_tx_qdesc_status_t sfc_efx_tx_qdesc_status;
1113 static int
1114 sfc_efx_tx_qdesc_status(struct sfc_dp_txq *dp_txq, uint16_t offset)
1115 {
1116 	struct sfc_efx_txq *txq = sfc_efx_txq_by_dp_txq(dp_txq);
1117 
1118 	if (unlikely(offset > txq->ptr_mask))
1119 		return -EINVAL;
1120 
1121 	if (unlikely(offset >= txq->max_fill_level))
1122 		return RTE_ETH_TX_DESC_UNAVAIL;
1123 
1124 	/*
1125 	 * Poll EvQ to derive up-to-date 'txq->pending' figure;
1126 	 * it is required for the queue to be running, but the
1127 	 * check is omitted because API design assumes that it
1128 	 * is the duty of the caller to satisfy all conditions
1129 	 */
1130 	SFC_ASSERT((txq->flags & SFC_EFX_TXQ_FLAG_RUNNING) ==
1131 		   SFC_EFX_TXQ_FLAG_RUNNING);
1132 	sfc_ev_qpoll(txq->evq);
1133 
1134 	/*
1135 	 * Ring tail is 'txq->pending', and although descriptors
1136 	 * between 'txq->completed' and 'txq->pending' are still
1137 	 * in use by the driver, they should be reported as DONE
1138 	 */
1139 	if (unlikely(offset < (txq->added - txq->pending)))
1140 		return RTE_ETH_TX_DESC_FULL;
1141 
1142 	/*
1143 	 * There is no separate return value for unused descriptors;
1144 	 * the latter will be reported as DONE because genuine DONE
1145 	 * descriptors will be freed anyway in SW on the next burst
1146 	 */
1147 	return RTE_ETH_TX_DESC_DONE;
1148 }
1149 
1150 struct sfc_dp_tx sfc_efx_tx = {
1151 	.dp = {
1152 		.name		= SFC_KVARG_DATAPATH_EFX,
1153 		.type		= SFC_DP_TX,
1154 		.hw_fw_caps	= SFC_DP_HW_FW_CAP_TX_EFX,
1155 	},
1156 	.features		= 0,
1157 	.dev_offload_capa	= DEV_TX_OFFLOAD_VLAN_INSERT |
1158 				  DEV_TX_OFFLOAD_MULTI_SEGS,
1159 	.queue_offload_capa	= DEV_TX_OFFLOAD_IPV4_CKSUM |
1160 				  DEV_TX_OFFLOAD_UDP_CKSUM |
1161 				  DEV_TX_OFFLOAD_TCP_CKSUM |
1162 				  DEV_TX_OFFLOAD_OUTER_IPV4_CKSUM |
1163 				  DEV_TX_OFFLOAD_TCP_TSO,
1164 	.qsize_up_rings		= sfc_efx_tx_qsize_up_rings,
1165 	.qcreate		= sfc_efx_tx_qcreate,
1166 	.qdestroy		= sfc_efx_tx_qdestroy,
1167 	.qstart			= sfc_efx_tx_qstart,
1168 	.qstop			= sfc_efx_tx_qstop,
1169 	.qreap			= sfc_efx_tx_qreap,
1170 	.qdesc_status		= sfc_efx_tx_qdesc_status,
1171 	.pkt_prepare		= sfc_efx_prepare_pkts,
1172 	.pkt_burst		= sfc_efx_xmit_pkts,
1173 };
1174