xref: /dpdk/drivers/net/ionic/ionic_rxtx.c (revision 330a70b773f04060e3026004d6291182a5cbefaf)
1 /* SPDX-License-Identifier: (BSD-3-Clause OR GPL-2.0)
2  * Copyright(c) 2018-2019 Pensando Systems, Inc. All rights reserved.
3  */
4 
5 #include <sys/queue.h>
6 #include <stdio.h>
7 #include <stdlib.h>
8 #include <string.h>
9 #include <errno.h>
10 #include <stdint.h>
11 #include <stdarg.h>
12 #include <unistd.h>
13 #include <inttypes.h>
14 
15 #include <rte_byteorder.h>
16 #include <rte_common.h>
17 #include <rte_cycles.h>
18 #include <rte_log.h>
19 #include <rte_debug.h>
20 #include <rte_interrupts.h>
21 #include <rte_pci.h>
22 #include <rte_memory.h>
23 #include <rte_memzone.h>
24 #include <rte_launch.h>
25 #include <rte_eal.h>
26 #include <rte_per_lcore.h>
27 #include <rte_lcore.h>
28 #include <rte_atomic.h>
29 #include <rte_branch_prediction.h>
30 #include <rte_mempool.h>
31 #include <rte_malloc.h>
32 #include <rte_mbuf.h>
33 #include <rte_ether.h>
34 #include <ethdev_driver.h>
35 #include <rte_prefetch.h>
36 #include <rte_udp.h>
37 #include <rte_tcp.h>
38 #include <rte_sctp.h>
39 #include <rte_string_fns.h>
40 #include <rte_errno.h>
41 #include <rte_ip.h>
42 #include <rte_net.h>
43 
44 #include "ionic_logs.h"
45 #include "ionic_mac_api.h"
46 #include "ionic_ethdev.h"
47 #include "ionic_lif.h"
48 #include "ionic_rxtx.h"
49 
50 /*********************************************************************
51  *
52  *  TX functions
53  *
54  **********************************************************************/
55 
56 void
57 ionic_txq_info_get(struct rte_eth_dev *dev, uint16_t queue_id,
58 		struct rte_eth_txq_info *qinfo)
59 {
60 	struct ionic_tx_qcq *txq = dev->data->tx_queues[queue_id];
61 	struct ionic_queue *q = &txq->qcq.q;
62 
63 	qinfo->nb_desc = q->num_descs;
64 	qinfo->conf.offloads = dev->data->dev_conf.txmode.offloads;
65 	qinfo->conf.tx_deferred_start = txq->flags & IONIC_QCQ_F_DEFERRED;
66 }
67 
68 static __rte_always_inline void
69 ionic_tx_flush(struct ionic_tx_qcq *txq)
70 {
71 	struct ionic_cq *cq = &txq->qcq.cq;
72 	struct ionic_queue *q = &txq->qcq.q;
73 	struct rte_mbuf *txm, *next;
74 	struct ionic_txq_comp *cq_desc_base = cq->base;
75 	struct ionic_txq_comp *cq_desc;
76 	void **info;
77 	u_int32_t comp_index = (u_int32_t)-1;
78 
79 	cq_desc = &cq_desc_base[cq->tail_idx];
80 	while (color_match(cq_desc->color, cq->done_color)) {
81 		cq->tail_idx = Q_NEXT_TO_SRVC(cq, 1);
82 
83 		/* Prefetch the next 4 descriptors (not really useful here) */
84 		if ((cq->tail_idx & 0x3) == 0)
85 			rte_prefetch0(&cq_desc_base[cq->tail_idx]);
86 
87 		if (cq->tail_idx == 0)
88 			cq->done_color = !cq->done_color;
89 
90 		comp_index = cq_desc->comp_index;
91 
92 		cq_desc = &cq_desc_base[cq->tail_idx];
93 	}
94 
95 	if (comp_index != (u_int32_t)-1) {
96 		while (q->tail_idx != comp_index) {
97 			info = IONIC_INFO_PTR(q, q->tail_idx);
98 
99 			q->tail_idx = Q_NEXT_TO_SRVC(q, 1);
100 
101 			/* Prefetch the next 4 descriptors */
102 			if ((q->tail_idx & 0x3) == 0)
103 				/* q desc info */
104 				rte_prefetch0(&q->info[q->tail_idx]);
105 
106 			/*
107 			 * Note: you can just use rte_pktmbuf_free,
108 			 * but this loop is faster
109 			 */
110 			txm = info[0];
111 			while (txm != NULL) {
112 				next = txm->next;
113 				rte_pktmbuf_free_seg(txm);
114 				txm = next;
115 			}
116 		}
117 	}
118 }
119 
120 void __rte_cold
121 ionic_dev_tx_queue_release(void *tx_queue)
122 {
123 	struct ionic_tx_qcq *txq = tx_queue;
124 	struct ionic_tx_stats *stats = &txq->stats;
125 
126 	IONIC_PRINT_CALL();
127 
128 	IONIC_PRINT(DEBUG, "TX queue %u pkts %ju tso %ju",
129 		txq->qcq.q.index, stats->packets, stats->tso);
130 
131 	ionic_lif_txq_deinit(txq);
132 
133 	ionic_qcq_free(&txq->qcq);
134 }
135 
136 int __rte_cold
137 ionic_dev_tx_queue_stop(struct rte_eth_dev *eth_dev, uint16_t tx_queue_id)
138 {
139 	struct ionic_tx_qcq *txq;
140 
141 	IONIC_PRINT(DEBUG, "Stopping TX queue %u", tx_queue_id);
142 
143 	txq = eth_dev->data->tx_queues[tx_queue_id];
144 
145 	eth_dev->data->tx_queue_state[tx_queue_id] =
146 		RTE_ETH_QUEUE_STATE_STOPPED;
147 
148 	/*
149 	 * Note: we should better post NOP Tx desc and wait for its completion
150 	 * before disabling Tx queue
151 	 */
152 
153 	ionic_qcq_disable(&txq->qcq);
154 
155 	ionic_tx_flush(txq);
156 
157 	return 0;
158 }
159 
160 int __rte_cold
161 ionic_dev_tx_queue_setup(struct rte_eth_dev *eth_dev, uint16_t tx_queue_id,
162 		uint16_t nb_desc, uint32_t socket_id,
163 		const struct rte_eth_txconf *tx_conf)
164 {
165 	struct ionic_lif *lif = IONIC_ETH_DEV_TO_LIF(eth_dev);
166 	struct ionic_tx_qcq *txq;
167 	uint64_t offloads;
168 	int err;
169 
170 	if (tx_queue_id >= lif->ntxqcqs) {
171 		IONIC_PRINT(DEBUG, "Queue index %u not available "
172 			"(max %u queues)",
173 			tx_queue_id, lif->ntxqcqs);
174 		return -EINVAL;
175 	}
176 
177 	offloads = tx_conf->offloads | eth_dev->data->dev_conf.txmode.offloads;
178 	IONIC_PRINT(DEBUG,
179 		"Configuring skt %u TX queue %u with %u buffers, offloads %jx",
180 		socket_id, tx_queue_id, nb_desc, offloads);
181 
182 	/* Validate number of receive descriptors */
183 	if (!rte_is_power_of_2(nb_desc) || nb_desc < IONIC_MIN_RING_DESC)
184 		return -EINVAL; /* or use IONIC_DEFAULT_RING_DESC */
185 
186 	/* Free memory prior to re-allocation if needed... */
187 	if (eth_dev->data->tx_queues[tx_queue_id] != NULL) {
188 		void *tx_queue = eth_dev->data->tx_queues[tx_queue_id];
189 		ionic_dev_tx_queue_release(tx_queue);
190 		eth_dev->data->tx_queues[tx_queue_id] = NULL;
191 	}
192 
193 	eth_dev->data->tx_queue_state[tx_queue_id] =
194 		RTE_ETH_QUEUE_STATE_STOPPED;
195 
196 	err = ionic_tx_qcq_alloc(lif, socket_id, tx_queue_id, nb_desc, &txq);
197 	if (err) {
198 		IONIC_PRINT(DEBUG, "Queue allocation failure");
199 		return -EINVAL;
200 	}
201 
202 	/* Do not start queue with rte_eth_dev_start() */
203 	if (tx_conf->tx_deferred_start)
204 		txq->flags |= IONIC_QCQ_F_DEFERRED;
205 
206 	/* Convert the offload flags into queue flags */
207 	if (offloads & DEV_TX_OFFLOAD_IPV4_CKSUM)
208 		txq->flags |= IONIC_QCQ_F_CSUM_L3;
209 	if (offloads & DEV_TX_OFFLOAD_TCP_CKSUM)
210 		txq->flags |= IONIC_QCQ_F_CSUM_TCP;
211 	if (offloads & DEV_TX_OFFLOAD_UDP_CKSUM)
212 		txq->flags |= IONIC_QCQ_F_CSUM_UDP;
213 
214 	eth_dev->data->tx_queues[tx_queue_id] = txq;
215 
216 	return 0;
217 }
218 
219 /*
220  * Start Transmit Units for specified queue.
221  */
222 int __rte_cold
223 ionic_dev_tx_queue_start(struct rte_eth_dev *eth_dev, uint16_t tx_queue_id)
224 {
225 	uint8_t *tx_queue_state = eth_dev->data->tx_queue_state;
226 	struct ionic_tx_qcq *txq;
227 	int err;
228 
229 	if (tx_queue_state[tx_queue_id] == RTE_ETH_QUEUE_STATE_STARTED) {
230 		IONIC_PRINT(DEBUG, "TX queue %u already started",
231 			tx_queue_id);
232 		return 0;
233 	}
234 
235 	txq = eth_dev->data->tx_queues[tx_queue_id];
236 
237 	IONIC_PRINT(DEBUG, "Starting TX queue %u, %u descs",
238 		tx_queue_id, txq->qcq.q.num_descs);
239 
240 	if (!(txq->flags & IONIC_QCQ_F_INITED)) {
241 		err = ionic_lif_txq_init(txq);
242 		if (err)
243 			return err;
244 	} else {
245 		ionic_qcq_enable(&txq->qcq);
246 	}
247 
248 	tx_queue_state[tx_queue_id] = RTE_ETH_QUEUE_STATE_STARTED;
249 
250 	return 0;
251 }
252 
253 static void
254 ionic_tx_tcp_pseudo_csum(struct rte_mbuf *txm)
255 {
256 	struct ether_hdr *eth_hdr = rte_pktmbuf_mtod(txm, struct ether_hdr *);
257 	char *l3_hdr = ((char *)eth_hdr) + txm->l2_len;
258 	struct rte_tcp_hdr *tcp_hdr = (struct rte_tcp_hdr *)
259 		(l3_hdr + txm->l3_len);
260 
261 	if (txm->ol_flags & PKT_TX_IP_CKSUM) {
262 		struct rte_ipv4_hdr *ipv4_hdr = (struct rte_ipv4_hdr *)l3_hdr;
263 		ipv4_hdr->hdr_checksum = 0;
264 		tcp_hdr->cksum = 0;
265 		tcp_hdr->cksum = rte_ipv4_udptcp_cksum(ipv4_hdr, tcp_hdr);
266 	} else {
267 		struct rte_ipv6_hdr *ipv6_hdr = (struct rte_ipv6_hdr *)l3_hdr;
268 		tcp_hdr->cksum = 0;
269 		tcp_hdr->cksum = rte_ipv6_udptcp_cksum(ipv6_hdr, tcp_hdr);
270 	}
271 }
272 
273 static void
274 ionic_tx_tcp_inner_pseudo_csum(struct rte_mbuf *txm)
275 {
276 	struct ether_hdr *eth_hdr = rte_pktmbuf_mtod(txm, struct ether_hdr *);
277 	char *l3_hdr = ((char *)eth_hdr) + txm->outer_l2_len +
278 		txm->outer_l3_len + txm->l2_len;
279 	struct rte_tcp_hdr *tcp_hdr = (struct rte_tcp_hdr *)
280 		(l3_hdr + txm->l3_len);
281 
282 	if (txm->ol_flags & PKT_TX_IPV4) {
283 		struct rte_ipv4_hdr *ipv4_hdr = (struct rte_ipv4_hdr *)l3_hdr;
284 		ipv4_hdr->hdr_checksum = 0;
285 		tcp_hdr->cksum = 0;
286 		tcp_hdr->cksum = rte_ipv4_udptcp_cksum(ipv4_hdr, tcp_hdr);
287 	} else {
288 		struct rte_ipv6_hdr *ipv6_hdr = (struct rte_ipv6_hdr *)l3_hdr;
289 		tcp_hdr->cksum = 0;
290 		tcp_hdr->cksum = rte_ipv6_udptcp_cksum(ipv6_hdr, tcp_hdr);
291 	}
292 }
293 
294 static void
295 ionic_tx_tso_post(struct ionic_queue *q, struct ionic_txq_desc *desc,
296 		struct rte_mbuf *txm,
297 		rte_iova_t addr, uint8_t nsge, uint16_t len,
298 		uint32_t hdrlen, uint32_t mss,
299 		bool encap,
300 		uint16_t vlan_tci, bool has_vlan,
301 		bool start, bool done)
302 {
303 	void **info;
304 	uint8_t flags = 0;
305 	flags |= has_vlan ? IONIC_TXQ_DESC_FLAG_VLAN : 0;
306 	flags |= encap ? IONIC_TXQ_DESC_FLAG_ENCAP : 0;
307 	flags |= start ? IONIC_TXQ_DESC_FLAG_TSO_SOT : 0;
308 	flags |= done ? IONIC_TXQ_DESC_FLAG_TSO_EOT : 0;
309 
310 	desc->cmd = encode_txq_desc_cmd(IONIC_TXQ_DESC_OPCODE_TSO,
311 		flags, nsge, addr);
312 	desc->len = len;
313 	desc->vlan_tci = vlan_tci;
314 	desc->hdr_len = hdrlen;
315 	desc->mss = mss;
316 
317 	if (done) {
318 		info = IONIC_INFO_PTR(q, q->head_idx);
319 		info[0] = txm;
320 	}
321 
322 	q->head_idx = Q_NEXT_TO_POST(q, 1);
323 }
324 
325 static struct ionic_txq_desc *
326 ionic_tx_tso_next(struct ionic_tx_qcq *txq, struct ionic_txq_sg_elem **elem)
327 {
328 	struct ionic_queue *q = &txq->qcq.q;
329 	struct ionic_txq_desc *desc_base = q->base;
330 	struct ionic_txq_sg_desc_v1 *sg_desc_base = q->sg_base;
331 	struct ionic_txq_desc *desc = &desc_base[q->head_idx];
332 	struct ionic_txq_sg_desc_v1 *sg_desc = &sg_desc_base[q->head_idx];
333 
334 	*elem = sg_desc->elems;
335 	return desc;
336 }
337 
338 static int
339 ionic_tx_tso(struct ionic_tx_qcq *txq, struct rte_mbuf *txm)
340 {
341 	struct ionic_queue *q = &txq->qcq.q;
342 	struct ionic_tx_stats *stats = &txq->stats;
343 	struct ionic_txq_desc *desc;
344 	struct ionic_txq_sg_elem *elem;
345 	struct rte_mbuf *txm_seg;
346 	rte_iova_t data_iova;
347 	uint64_t desc_addr = 0, next_addr;
348 	uint16_t desc_len = 0;
349 	uint8_t desc_nsge;
350 	uint32_t hdrlen;
351 	uint32_t mss = txm->tso_segsz;
352 	uint32_t frag_left = 0;
353 	uint32_t left;
354 	uint32_t seglen;
355 	uint32_t len;
356 	uint32_t offset = 0;
357 	bool start, done;
358 	bool encap;
359 	bool has_vlan = !!(txm->ol_flags & PKT_TX_VLAN_PKT);
360 	uint16_t vlan_tci = txm->vlan_tci;
361 	uint64_t ol_flags = txm->ol_flags;
362 
363 	encap = ((ol_flags & PKT_TX_OUTER_IP_CKSUM) ||
364 		(ol_flags & PKT_TX_OUTER_UDP_CKSUM)) &&
365 		((ol_flags & PKT_TX_OUTER_IPV4) ||
366 		(ol_flags & PKT_TX_OUTER_IPV6));
367 
368 	/* Preload inner-most TCP csum field with IP pseudo hdr
369 	 * calculated with IP length set to zero.  HW will later
370 	 * add in length to each TCP segment resulting from the TSO.
371 	 */
372 
373 	if (encap) {
374 		ionic_tx_tcp_inner_pseudo_csum(txm);
375 		hdrlen = txm->outer_l2_len + txm->outer_l3_len +
376 			txm->l2_len + txm->l3_len + txm->l4_len;
377 	} else {
378 		ionic_tx_tcp_pseudo_csum(txm);
379 		hdrlen = txm->l2_len + txm->l3_len + txm->l4_len;
380 	}
381 
382 	seglen = hdrlen + mss;
383 	left = txm->data_len;
384 	data_iova = rte_mbuf_data_iova(txm);
385 
386 	desc = ionic_tx_tso_next(txq, &elem);
387 	start = true;
388 
389 	/* Chop data up into desc segments */
390 
391 	while (left > 0) {
392 		len = RTE_MIN(seglen, left);
393 		frag_left = seglen - len;
394 		desc_addr = rte_cpu_to_le_64(data_iova + offset);
395 		desc_len = len;
396 		desc_nsge = 0;
397 		left -= len;
398 		offset += len;
399 		if (txm->nb_segs > 1 && frag_left > 0)
400 			continue;
401 		done = (txm->nb_segs == 1 && left == 0);
402 		ionic_tx_tso_post(q, desc, txm,
403 			desc_addr, desc_nsge, desc_len,
404 			hdrlen, mss,
405 			encap,
406 			vlan_tci, has_vlan,
407 			start, done);
408 		desc = ionic_tx_tso_next(txq, &elem);
409 		start = false;
410 		seglen = mss;
411 	}
412 
413 	/* Chop frags into desc segments */
414 
415 	txm_seg = txm->next;
416 	while (txm_seg != NULL) {
417 		offset = 0;
418 		data_iova = rte_mbuf_data_iova(txm_seg);
419 		left = txm_seg->data_len;
420 
421 		while (left > 0) {
422 			next_addr = rte_cpu_to_le_64(data_iova + offset);
423 			if (frag_left > 0) {
424 				len = RTE_MIN(frag_left, left);
425 				frag_left -= len;
426 				elem->addr = next_addr;
427 				elem->len = len;
428 				elem++;
429 				desc_nsge++;
430 			} else {
431 				len = RTE_MIN(mss, left);
432 				frag_left = mss - len;
433 				desc_addr = next_addr;
434 				desc_len = len;
435 				desc_nsge = 0;
436 			}
437 			left -= len;
438 			offset += len;
439 			if (txm_seg->next != NULL && frag_left > 0)
440 				continue;
441 
442 			done = (txm_seg->next == NULL && left == 0);
443 			ionic_tx_tso_post(q, desc, txm_seg,
444 				desc_addr, desc_nsge, desc_len,
445 				hdrlen, mss,
446 				encap,
447 				vlan_tci, has_vlan,
448 				start, done);
449 			desc = ionic_tx_tso_next(txq, &elem);
450 			start = false;
451 		}
452 
453 		txm_seg = txm_seg->next;
454 	}
455 
456 	stats->tso++;
457 
458 	return 0;
459 }
460 
461 static __rte_always_inline int
462 ionic_tx(struct ionic_tx_qcq *txq, struct rte_mbuf *txm)
463 {
464 	struct ionic_queue *q = &txq->qcq.q;
465 	struct ionic_txq_desc *desc, *desc_base = q->base;
466 	struct ionic_txq_sg_desc_v1 *sg_desc_base = q->sg_base;
467 	struct ionic_txq_sg_elem *elem;
468 	struct ionic_tx_stats *stats = &txq->stats;
469 	struct rte_mbuf *txm_seg;
470 	void **info;
471 	bool encap;
472 	bool has_vlan;
473 	uint64_t ol_flags = txm->ol_flags;
474 	uint64_t addr;
475 	uint8_t opcode = IONIC_TXQ_DESC_OPCODE_CSUM_NONE;
476 	uint8_t flags = 0;
477 
478 	desc = &desc_base[q->head_idx];
479 	info = IONIC_INFO_PTR(q, q->head_idx);
480 
481 	if ((ol_flags & PKT_TX_IP_CKSUM) &&
482 	    (txq->flags & IONIC_QCQ_F_CSUM_L3)) {
483 		opcode = IONIC_TXQ_DESC_OPCODE_CSUM_HW;
484 		flags |= IONIC_TXQ_DESC_FLAG_CSUM_L3;
485 	}
486 
487 	if (((ol_flags & PKT_TX_TCP_CKSUM) &&
488 	     (txq->flags & IONIC_QCQ_F_CSUM_TCP)) ||
489 	    ((ol_flags & PKT_TX_UDP_CKSUM) &&
490 	     (txq->flags & IONIC_QCQ_F_CSUM_UDP))) {
491 		opcode = IONIC_TXQ_DESC_OPCODE_CSUM_HW;
492 		flags |= IONIC_TXQ_DESC_FLAG_CSUM_L4;
493 	}
494 
495 	if (opcode == IONIC_TXQ_DESC_OPCODE_CSUM_NONE)
496 		stats->no_csum++;
497 
498 	has_vlan = (ol_flags & PKT_TX_VLAN_PKT);
499 	encap = ((ol_flags & PKT_TX_OUTER_IP_CKSUM) ||
500 			(ol_flags & PKT_TX_OUTER_UDP_CKSUM)) &&
501 			((ol_flags & PKT_TX_OUTER_IPV4) ||
502 			(ol_flags & PKT_TX_OUTER_IPV6));
503 
504 	flags |= has_vlan ? IONIC_TXQ_DESC_FLAG_VLAN : 0;
505 	flags |= encap ? IONIC_TXQ_DESC_FLAG_ENCAP : 0;
506 
507 	addr = rte_cpu_to_le_64(rte_mbuf_data_iova(txm));
508 
509 	desc->cmd = encode_txq_desc_cmd(opcode, flags, txm->nb_segs - 1, addr);
510 	desc->len = txm->data_len;
511 	desc->vlan_tci = txm->vlan_tci;
512 
513 	info[0] = txm;
514 
515 	elem = sg_desc_base[q->head_idx].elems;
516 
517 	txm_seg = txm->next;
518 	while (txm_seg != NULL) {
519 		elem->len = txm_seg->data_len;
520 		elem->addr = rte_cpu_to_le_64(rte_mbuf_data_iova(txm_seg));
521 		elem++;
522 		txm_seg = txm_seg->next;
523 	}
524 
525 	q->head_idx = Q_NEXT_TO_POST(q, 1);
526 
527 	return 0;
528 }
529 
530 uint16_t
531 ionic_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts,
532 		uint16_t nb_pkts)
533 {
534 	struct ionic_tx_qcq *txq = tx_queue;
535 	struct ionic_queue *q = &txq->qcq.q;
536 	struct ionic_tx_stats *stats = &txq->stats;
537 	uint32_t next_q_head_idx;
538 	uint32_t bytes_tx = 0;
539 	uint16_t nb_avail, nb_tx = 0;
540 	int err;
541 
542 	/* Cleaning old buffers */
543 	ionic_tx_flush(txq);
544 
545 	nb_avail = ionic_q_space_avail(q);
546 	if (unlikely(nb_avail < nb_pkts)) {
547 		stats->stop += nb_pkts - nb_avail;
548 		nb_pkts = nb_avail;
549 	}
550 
551 	while (nb_tx < nb_pkts) {
552 		next_q_head_idx = Q_NEXT_TO_POST(q, 1);
553 		if ((next_q_head_idx & 0x3) == 0) {
554 			struct ionic_txq_desc *desc_base = q->base;
555 			rte_prefetch0(&desc_base[next_q_head_idx]);
556 			rte_prefetch0(&q->info[next_q_head_idx]);
557 		}
558 
559 		if (tx_pkts[nb_tx]->ol_flags & PKT_TX_TCP_SEG)
560 			err = ionic_tx_tso(txq, tx_pkts[nb_tx]);
561 		else
562 			err = ionic_tx(txq, tx_pkts[nb_tx]);
563 		if (err) {
564 			stats->drop += nb_pkts - nb_tx;
565 			break;
566 		}
567 
568 		bytes_tx += tx_pkts[nb_tx]->pkt_len;
569 		nb_tx++;
570 	}
571 
572 	if (nb_tx > 0) {
573 		rte_wmb();
574 		ionic_q_flush(q);
575 	}
576 
577 	stats->packets += nb_tx;
578 	stats->bytes += bytes_tx;
579 
580 	return nb_tx;
581 }
582 
583 /*********************************************************************
584  *
585  *  TX prep functions
586  *
587  **********************************************************************/
588 
589 #define IONIC_TX_OFFLOAD_MASK (	\
590 	PKT_TX_IPV4 |		\
591 	PKT_TX_IPV6 |		\
592 	PKT_TX_VLAN |		\
593 	PKT_TX_IP_CKSUM |	\
594 	PKT_TX_TCP_SEG |	\
595 	PKT_TX_L4_MASK)
596 
597 #define IONIC_TX_OFFLOAD_NOTSUP_MASK \
598 	(PKT_TX_OFFLOAD_MASK ^ IONIC_TX_OFFLOAD_MASK)
599 
600 uint16_t
601 ionic_prep_pkts(void *tx_queue, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
602 {
603 	struct ionic_tx_qcq *txq = tx_queue;
604 	struct rte_mbuf *txm;
605 	uint64_t offloads;
606 	int i = 0;
607 
608 	for (i = 0; i < nb_pkts; i++) {
609 		txm = tx_pkts[i];
610 
611 		if (txm->nb_segs > txq->num_segs_fw) {
612 			rte_errno = -EINVAL;
613 			break;
614 		}
615 
616 		offloads = txm->ol_flags;
617 
618 		if (offloads & IONIC_TX_OFFLOAD_NOTSUP_MASK) {
619 			rte_errno = -ENOTSUP;
620 			break;
621 		}
622 	}
623 
624 	return i;
625 }
626 
627 /*********************************************************************
628  *
629  *  RX functions
630  *
631  **********************************************************************/
632 
633 static void ionic_rx_recycle(struct ionic_queue *q, uint32_t q_desc_index,
634 		struct rte_mbuf *mbuf);
635 
636 void
637 ionic_rxq_info_get(struct rte_eth_dev *dev, uint16_t queue_id,
638 		struct rte_eth_rxq_info *qinfo)
639 {
640 	struct ionic_rx_qcq *rxq = dev->data->rx_queues[queue_id];
641 	struct ionic_queue *q = &rxq->qcq.q;
642 
643 	qinfo->mp = rxq->mb_pool;
644 	qinfo->scattered_rx = dev->data->scattered_rx;
645 	qinfo->nb_desc = q->num_descs;
646 	qinfo->conf.rx_deferred_start = rxq->flags & IONIC_QCQ_F_DEFERRED;
647 	qinfo->conf.offloads = dev->data->dev_conf.rxmode.offloads;
648 }
649 
650 static void __rte_cold
651 ionic_rx_empty(struct ionic_rx_qcq *rxq)
652 {
653 	struct ionic_queue *q = &rxq->qcq.q;
654 	struct rte_mbuf *mbuf;
655 	void **info;
656 
657 	while (q->tail_idx != q->head_idx) {
658 		info = IONIC_INFO_PTR(q, q->tail_idx);
659 		mbuf = info[0];
660 		rte_mempool_put(rxq->mb_pool, mbuf);
661 
662 		q->tail_idx = Q_NEXT_TO_SRVC(q, 1);
663 	}
664 }
665 
666 void __rte_cold
667 ionic_dev_rx_queue_release(void *rx_queue)
668 {
669 	struct ionic_rx_qcq *rxq = rx_queue;
670 	struct ionic_rx_stats *stats;
671 
672 	if (!rxq)
673 		return;
674 
675 	IONIC_PRINT_CALL();
676 
677 	stats = &rxq->stats;
678 
679 	IONIC_PRINT(DEBUG, "RX queue %u pkts %ju mtod %ju",
680 		rxq->qcq.q.index, stats->packets, stats->mtods);
681 
682 	ionic_rx_empty(rxq);
683 
684 	ionic_lif_rxq_deinit(rxq);
685 
686 	ionic_qcq_free(&rxq->qcq);
687 }
688 
689 int __rte_cold
690 ionic_dev_rx_queue_setup(struct rte_eth_dev *eth_dev,
691 		uint16_t rx_queue_id,
692 		uint16_t nb_desc,
693 		uint32_t socket_id,
694 		const struct rte_eth_rxconf *rx_conf,
695 		struct rte_mempool *mp)
696 {
697 	struct ionic_lif *lif = IONIC_ETH_DEV_TO_LIF(eth_dev);
698 	struct ionic_rx_qcq *rxq;
699 	uint64_t offloads;
700 	int err;
701 
702 	if (rx_queue_id >= lif->nrxqcqs) {
703 		IONIC_PRINT(ERR,
704 			"Queue index %u not available (max %u queues)",
705 			rx_queue_id, lif->nrxqcqs);
706 		return -EINVAL;
707 	}
708 
709 	offloads = rx_conf->offloads | eth_dev->data->dev_conf.rxmode.offloads;
710 	IONIC_PRINT(DEBUG,
711 		"Configuring skt %u RX queue %u with %u buffers, offloads %jx",
712 		socket_id, rx_queue_id, nb_desc, offloads);
713 
714 	if (!rx_conf->rx_drop_en)
715 		IONIC_PRINT(WARNING, "No-drop mode is not supported");
716 
717 	/* Validate number of receive descriptors */
718 	if (!rte_is_power_of_2(nb_desc) ||
719 			nb_desc < IONIC_MIN_RING_DESC ||
720 			nb_desc > IONIC_MAX_RING_DESC) {
721 		IONIC_PRINT(ERR,
722 			"Bad descriptor count (%u) for queue %u (min: %u)",
723 			nb_desc, rx_queue_id, IONIC_MIN_RING_DESC);
724 		return -EINVAL; /* or use IONIC_DEFAULT_RING_DESC */
725 	}
726 
727 	/* Free memory prior to re-allocation if needed... */
728 	if (eth_dev->data->rx_queues[rx_queue_id] != NULL) {
729 		void *rx_queue = eth_dev->data->rx_queues[rx_queue_id];
730 		ionic_dev_rx_queue_release(rx_queue);
731 		eth_dev->data->rx_queues[rx_queue_id] = NULL;
732 	}
733 
734 	eth_dev->data->rx_queue_state[rx_queue_id] =
735 		RTE_ETH_QUEUE_STATE_STOPPED;
736 
737 	err = ionic_rx_qcq_alloc(lif, socket_id, rx_queue_id, nb_desc,
738 			&rxq);
739 	if (err) {
740 		IONIC_PRINT(ERR, "Queue %d allocation failure", rx_queue_id);
741 		return -EINVAL;
742 	}
743 
744 	rxq->mb_pool = mp;
745 
746 	/*
747 	 * Note: the interface does not currently support
748 	 * DEV_RX_OFFLOAD_KEEP_CRC, please also consider ETHER_CRC_LEN
749 	 * when the adapter will be able to keep the CRC and subtract
750 	 * it to the length for all received packets:
751 	 * if (eth_dev->data->dev_conf.rxmode.offloads &
752 	 *     DEV_RX_OFFLOAD_KEEP_CRC)
753 	 *   rxq->crc_len = ETHER_CRC_LEN;
754 	 */
755 
756 	/* Do not start queue with rte_eth_dev_start() */
757 	if (rx_conf->rx_deferred_start)
758 		rxq->flags |= IONIC_QCQ_F_DEFERRED;
759 
760 	eth_dev->data->rx_queues[rx_queue_id] = rxq;
761 
762 	return 0;
763 }
764 
765 static __rte_always_inline void
766 ionic_rx_clean(struct ionic_rx_qcq *rxq,
767 		uint32_t q_desc_index, uint32_t cq_desc_index,
768 		void *service_cb_arg)
769 {
770 	struct ionic_queue *q = &rxq->qcq.q;
771 	struct ionic_cq *cq = &rxq->qcq.cq;
772 	struct ionic_rxq_comp *cq_desc_base = cq->base;
773 	struct ionic_rxq_comp *cq_desc = &cq_desc_base[cq_desc_index];
774 	struct rte_mbuf *rxm, *rxm_seg;
775 	uint32_t max_frame_size =
776 		rxq->qcq.lif->eth_dev->data->dev_conf.rxmode.max_rx_pkt_len;
777 	uint64_t pkt_flags = 0;
778 	uint32_t pkt_type;
779 	struct ionic_rx_stats *stats = &rxq->stats;
780 	struct ionic_rx_service *recv_args = (struct ionic_rx_service *)
781 		service_cb_arg;
782 	uint32_t buf_size = (uint16_t)
783 		(rte_pktmbuf_data_room_size(rxq->mb_pool) -
784 		RTE_PKTMBUF_HEADROOM);
785 	uint32_t left;
786 	void **info;
787 
788 	assert(q_desc_index == cq_desc->comp_index);
789 
790 	info = IONIC_INFO_PTR(q, cq_desc->comp_index);
791 
792 	rxm = info[0];
793 
794 	if (!recv_args) {
795 		stats->no_cb_arg++;
796 		/* Flush */
797 		rte_pktmbuf_free(rxm);
798 		/*
799 		 * Note: rte_mempool_put is faster with no segs
800 		 * rte_mempool_put(rxq->mb_pool, rxm);
801 		 */
802 		return;
803 	}
804 
805 	if (cq_desc->status) {
806 		stats->bad_cq_status++;
807 		ionic_rx_recycle(q, q_desc_index, rxm);
808 		return;
809 	}
810 
811 	if (recv_args->nb_rx >= recv_args->nb_pkts) {
812 		stats->no_room++;
813 		ionic_rx_recycle(q, q_desc_index, rxm);
814 		return;
815 	}
816 
817 	if (cq_desc->len > max_frame_size ||
818 			cq_desc->len == 0) {
819 		stats->bad_len++;
820 		ionic_rx_recycle(q, q_desc_index, rxm);
821 		return;
822 	}
823 
824 	rxm->data_off = RTE_PKTMBUF_HEADROOM;
825 	rte_prefetch1((char *)rxm->buf_addr + rxm->data_off);
826 	rxm->nb_segs = 1; /* cq_desc->num_sg_elems */
827 	rxm->pkt_len = cq_desc->len;
828 	rxm->port = rxq->qcq.lif->port_id;
829 
830 	left = cq_desc->len;
831 
832 	rxm->data_len = RTE_MIN(buf_size, left);
833 	left -= rxm->data_len;
834 
835 	rxm_seg = rxm->next;
836 	while (rxm_seg && left) {
837 		rxm_seg->data_len = RTE_MIN(buf_size, left);
838 		left -= rxm_seg->data_len;
839 
840 		rxm_seg = rxm_seg->next;
841 		rxm->nb_segs++;
842 	}
843 
844 	/* RSS */
845 	pkt_flags |= PKT_RX_RSS_HASH;
846 	rxm->hash.rss = cq_desc->rss_hash;
847 
848 	/* Vlan Strip */
849 	if (cq_desc->csum_flags & IONIC_RXQ_COMP_CSUM_F_VLAN) {
850 		pkt_flags |= PKT_RX_VLAN | PKT_RX_VLAN_STRIPPED;
851 		rxm->vlan_tci = cq_desc->vlan_tci;
852 	}
853 
854 	/* Checksum */
855 	if (cq_desc->csum_flags & IONIC_RXQ_COMP_CSUM_F_CALC) {
856 		if (cq_desc->csum_flags & IONIC_RXQ_COMP_CSUM_F_IP_OK)
857 			pkt_flags |= PKT_RX_IP_CKSUM_GOOD;
858 		else if (cq_desc->csum_flags & IONIC_RXQ_COMP_CSUM_F_IP_BAD)
859 			pkt_flags |= PKT_RX_IP_CKSUM_BAD;
860 
861 		if ((cq_desc->csum_flags & IONIC_RXQ_COMP_CSUM_F_TCP_OK) ||
862 			(cq_desc->csum_flags & IONIC_RXQ_COMP_CSUM_F_UDP_OK))
863 			pkt_flags |= PKT_RX_L4_CKSUM_GOOD;
864 		else if ((cq_desc->csum_flags &
865 				IONIC_RXQ_COMP_CSUM_F_TCP_BAD) ||
866 				(cq_desc->csum_flags &
867 				IONIC_RXQ_COMP_CSUM_F_UDP_BAD))
868 			pkt_flags |= PKT_RX_L4_CKSUM_BAD;
869 	}
870 
871 	rxm->ol_flags = pkt_flags;
872 
873 	/* Packet Type */
874 	switch (cq_desc->pkt_type_color & IONIC_RXQ_COMP_PKT_TYPE_MASK) {
875 	case IONIC_PKT_TYPE_IPV4:
876 		pkt_type = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4;
877 		break;
878 	case IONIC_PKT_TYPE_IPV6:
879 		pkt_type = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6;
880 		break;
881 	case IONIC_PKT_TYPE_IPV4_TCP:
882 		pkt_type = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4 |
883 			RTE_PTYPE_L4_TCP;
884 		break;
885 	case IONIC_PKT_TYPE_IPV6_TCP:
886 		pkt_type = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6 |
887 			RTE_PTYPE_L4_TCP;
888 		break;
889 	case IONIC_PKT_TYPE_IPV4_UDP:
890 		pkt_type = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4 |
891 			RTE_PTYPE_L4_UDP;
892 		break;
893 	case IONIC_PKT_TYPE_IPV6_UDP:
894 		pkt_type = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6 |
895 			RTE_PTYPE_L4_UDP;
896 		break;
897 	default:
898 		{
899 			struct rte_ether_hdr *eth_h = rte_pktmbuf_mtod(rxm,
900 				struct rte_ether_hdr *);
901 			uint16_t ether_type = eth_h->ether_type;
902 			if (ether_type == rte_cpu_to_be_16(RTE_ETHER_TYPE_ARP))
903 				pkt_type = RTE_PTYPE_L2_ETHER_ARP;
904 			else
905 				pkt_type = RTE_PTYPE_UNKNOWN;
906 			stats->mtods++;
907 			break;
908 		}
909 	}
910 
911 	rxm->packet_type = pkt_type;
912 
913 	recv_args->rx_pkts[recv_args->nb_rx] = rxm;
914 	recv_args->nb_rx++;
915 
916 	stats->packets++;
917 	stats->bytes += rxm->pkt_len;
918 }
919 
920 static void
921 ionic_rx_recycle(struct ionic_queue *q, uint32_t q_desc_index,
922 		 struct rte_mbuf *mbuf)
923 {
924 	struct ionic_rxq_desc *desc_base = q->base;
925 	struct ionic_rxq_desc *old = &desc_base[q_desc_index];
926 	struct ionic_rxq_desc *new = &desc_base[q->head_idx];
927 
928 	new->addr = old->addr;
929 	new->len = old->len;
930 
931 	q->info[q->head_idx] = mbuf;
932 
933 	q->head_idx = Q_NEXT_TO_POST(q, 1);
934 
935 	ionic_q_flush(q);
936 }
937 
938 static __rte_always_inline int
939 ionic_rx_fill(struct ionic_rx_qcq *rxq, uint32_t len)
940 {
941 	struct ionic_queue *q = &rxq->qcq.q;
942 	struct ionic_rxq_desc *desc, *desc_base = q->base;
943 	struct ionic_rxq_sg_desc *sg_desc, *sg_desc_base = q->sg_base;
944 	struct ionic_rxq_sg_elem *elem;
945 	void **info;
946 	rte_iova_t dma_addr;
947 	uint32_t i, j, nsegs, buf_size, size;
948 
949 	buf_size = (uint16_t)(rte_pktmbuf_data_room_size(rxq->mb_pool) -
950 		RTE_PKTMBUF_HEADROOM);
951 
952 	/* Initialize software ring entries */
953 	for (i = ionic_q_space_avail(q); i; i--) {
954 		struct rte_mbuf *rxm = rte_mbuf_raw_alloc(rxq->mb_pool);
955 		struct rte_mbuf *prev_rxm_seg;
956 
957 		if (rxm == NULL) {
958 			IONIC_PRINT(ERR, "RX mbuf alloc failed");
959 			return -ENOMEM;
960 		}
961 
962 		info = IONIC_INFO_PTR(q, q->head_idx);
963 
964 		nsegs = (len + buf_size - 1) / buf_size;
965 
966 		desc = &desc_base[q->head_idx];
967 		dma_addr = rte_cpu_to_le_64(rte_mbuf_data_iova_default(rxm));
968 		desc->addr = dma_addr;
969 		desc->len = buf_size;
970 		size = buf_size;
971 		desc->opcode = (nsegs > 1) ? IONIC_RXQ_DESC_OPCODE_SG :
972 			IONIC_RXQ_DESC_OPCODE_SIMPLE;
973 		rxm->next = NULL;
974 
975 		prev_rxm_seg = rxm;
976 		sg_desc = &sg_desc_base[q->head_idx];
977 		elem = sg_desc->elems;
978 		for (j = 0; j < nsegs - 1 && j < IONIC_RX_MAX_SG_ELEMS; j++) {
979 			struct rte_mbuf *rxm_seg;
980 			rte_iova_t data_iova;
981 
982 			rxm_seg = rte_mbuf_raw_alloc(rxq->mb_pool);
983 			if (rxm_seg == NULL) {
984 				IONIC_PRINT(ERR, "RX mbuf alloc failed");
985 				return -ENOMEM;
986 			}
987 
988 			data_iova = rte_mbuf_data_iova(rxm_seg);
989 			dma_addr = rte_cpu_to_le_64(data_iova);
990 			elem->addr = dma_addr;
991 			elem->len = buf_size;
992 			size += buf_size;
993 			elem++;
994 			rxm_seg->next = NULL;
995 			prev_rxm_seg->next = rxm_seg;
996 			prev_rxm_seg = rxm_seg;
997 		}
998 
999 		if (size < len)
1000 			IONIC_PRINT(ERR, "Rx SG size is not sufficient (%d < %d)",
1001 				size, len);
1002 
1003 		info[0] = rxm;
1004 
1005 		q->head_idx = Q_NEXT_TO_POST(q, 1);
1006 	}
1007 
1008 	ionic_q_flush(q);
1009 
1010 	return 0;
1011 }
1012 
1013 /*
1014  * Start Receive Units for specified queue.
1015  */
1016 int __rte_cold
1017 ionic_dev_rx_queue_start(struct rte_eth_dev *eth_dev, uint16_t rx_queue_id)
1018 {
1019 	uint32_t frame_size = eth_dev->data->dev_conf.rxmode.max_rx_pkt_len;
1020 	uint8_t *rx_queue_state = eth_dev->data->rx_queue_state;
1021 	struct ionic_rx_qcq *rxq;
1022 	int err;
1023 
1024 	if (rx_queue_state[rx_queue_id] == RTE_ETH_QUEUE_STATE_STARTED) {
1025 		IONIC_PRINT(DEBUG, "RX queue %u already started",
1026 			rx_queue_id);
1027 		return 0;
1028 	}
1029 
1030 	rxq = eth_dev->data->rx_queues[rx_queue_id];
1031 
1032 	IONIC_PRINT(DEBUG, "Starting RX queue %u, %u descs (size: %u)",
1033 		rx_queue_id, rxq->qcq.q.num_descs, frame_size);
1034 
1035 	if (!(rxq->flags & IONIC_QCQ_F_INITED)) {
1036 		err = ionic_lif_rxq_init(rxq);
1037 		if (err)
1038 			return err;
1039 	} else {
1040 		ionic_qcq_enable(&rxq->qcq);
1041 	}
1042 
1043 	/* Allocate buffers for descriptor rings */
1044 	if (ionic_rx_fill(rxq, frame_size) != 0) {
1045 		IONIC_PRINT(ERR, "Could not alloc mbuf for queue:%d",
1046 			rx_queue_id);
1047 		return -1;
1048 	}
1049 
1050 	rx_queue_state[rx_queue_id] = RTE_ETH_QUEUE_STATE_STARTED;
1051 
1052 	return 0;
1053 }
1054 
1055 static __rte_always_inline void
1056 ionic_rxq_service(struct ionic_rx_qcq *rxq, uint32_t work_to_do,
1057 		void *service_cb_arg)
1058 {
1059 	struct ionic_cq *cq = &rxq->qcq.cq;
1060 	struct ionic_queue *q = &rxq->qcq.q;
1061 	struct ionic_rxq_comp *cq_desc, *cq_desc_base = cq->base;
1062 	bool more;
1063 	uint32_t curr_q_tail_idx, curr_cq_tail_idx;
1064 	uint32_t work_done = 0;
1065 
1066 	if (work_to_do == 0)
1067 		return;
1068 
1069 	cq_desc = &cq_desc_base[cq->tail_idx];
1070 	while (color_match(cq_desc->pkt_type_color, cq->done_color)) {
1071 		curr_cq_tail_idx = cq->tail_idx;
1072 		cq->tail_idx = Q_NEXT_TO_SRVC(cq, 1);
1073 
1074 		if (cq->tail_idx == 0)
1075 			cq->done_color = !cq->done_color;
1076 
1077 		/* Prefetch the next 4 descriptors */
1078 		if ((cq->tail_idx & 0x3) == 0)
1079 			rte_prefetch0(&cq_desc_base[cq->tail_idx]);
1080 
1081 		do {
1082 			more = (q->tail_idx != cq_desc->comp_index);
1083 
1084 			curr_q_tail_idx = q->tail_idx;
1085 			q->tail_idx = Q_NEXT_TO_SRVC(q, 1);
1086 
1087 			/* Prefetch the next 4 descriptors */
1088 			if ((q->tail_idx & 0x3) == 0)
1089 				/* q desc info */
1090 				rte_prefetch0(&q->info[q->tail_idx]);
1091 
1092 			ionic_rx_clean(rxq, curr_q_tail_idx, curr_cq_tail_idx,
1093 				service_cb_arg);
1094 
1095 		} while (more);
1096 
1097 		if (++work_done == work_to_do)
1098 			break;
1099 
1100 		cq_desc = &cq_desc_base[cq->tail_idx];
1101 	}
1102 }
1103 
1104 /*
1105  * Stop Receive Units for specified queue.
1106  */
1107 int __rte_cold
1108 ionic_dev_rx_queue_stop(struct rte_eth_dev *eth_dev, uint16_t rx_queue_id)
1109 {
1110 	struct ionic_rx_qcq *rxq;
1111 
1112 	IONIC_PRINT(DEBUG, "Stopping RX queue %u", rx_queue_id);
1113 
1114 	rxq = eth_dev->data->rx_queues[rx_queue_id];
1115 
1116 	eth_dev->data->rx_queue_state[rx_queue_id] =
1117 		RTE_ETH_QUEUE_STATE_STOPPED;
1118 
1119 	ionic_qcq_disable(&rxq->qcq);
1120 
1121 	/* Flush */
1122 	ionic_rxq_service(rxq, -1, NULL);
1123 
1124 	return 0;
1125 }
1126 
1127 uint16_t
1128 ionic_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts,
1129 		uint16_t nb_pkts)
1130 {
1131 	struct ionic_rx_qcq *rxq = rx_queue;
1132 	uint32_t frame_size =
1133 		rxq->qcq.lif->eth_dev->data->dev_conf.rxmode.max_rx_pkt_len;
1134 	struct ionic_rx_service service_cb_arg;
1135 
1136 	service_cb_arg.rx_pkts = rx_pkts;
1137 	service_cb_arg.nb_pkts = nb_pkts;
1138 	service_cb_arg.nb_rx = 0;
1139 
1140 	ionic_rxq_service(rxq, nb_pkts, &service_cb_arg);
1141 
1142 	ionic_rx_fill(rxq, frame_size);
1143 
1144 	return service_cb_arg.nb_rx;
1145 }
1146