xref: /dpdk/drivers/net/ngbe/ngbe_rxtx.c (revision 24cd85f7e5cff0b64dc76cbec3fd9948cd0738e1)
1 /* SPDX-License-Identifier: BSD-3-Clause
2  * Copyright(c) 2018-2021 Beijing WangXun Technology Co., Ltd.
3  * Copyright(c) 2010-2017 Intel Corporation
4  */
5 
6 #include <sys/queue.h>
7 
8 #include <stdint.h>
9 #include <rte_ethdev.h>
10 #include <ethdev_driver.h>
11 #include <rte_malloc.h>
12 #include <rte_net.h>
13 
14 #include "ngbe_logs.h"
15 #include "base/ngbe.h"
16 #include "ngbe_ethdev.h"
17 #include "ngbe_rxtx.h"
18 
19 #ifdef RTE_LIBRTE_IEEE1588
20 #define NGBE_TX_IEEE1588_TMST PKT_TX_IEEE1588_TMST
21 #else
22 #define NGBE_TX_IEEE1588_TMST 0
23 #endif
24 
25 /* Bit Mask to indicate what bits required for building Tx context */
26 static const u64 NGBE_TX_OFFLOAD_MASK = (RTE_MBUF_F_TX_IP_CKSUM |
27 		RTE_MBUF_F_TX_OUTER_IPV6 |
28 		RTE_MBUF_F_TX_OUTER_IPV4 |
29 		RTE_MBUF_F_TX_IPV6 |
30 		RTE_MBUF_F_TX_IPV4 |
31 		RTE_MBUF_F_TX_VLAN |
32 		RTE_MBUF_F_TX_L4_MASK |
33 		RTE_MBUF_F_TX_TCP_SEG |
34 		RTE_MBUF_F_TX_TUNNEL_MASK |
35 		RTE_MBUF_F_TX_OUTER_IP_CKSUM |
36 		NGBE_TX_IEEE1588_TMST);
37 
38 #define NGBE_TX_OFFLOAD_NOTSUP_MASK \
39 		(RTE_MBUF_F_TX_OFFLOAD_MASK ^ NGBE_TX_OFFLOAD_MASK)
40 
41 /*
42  * Prefetch a cache line into all cache levels.
43  */
44 #define rte_ngbe_prefetch(p)   rte_prefetch0(p)
45 
46 /*********************************************************************
47  *
48  *  Tx functions
49  *
50  **********************************************************************/
51 
52 /*
53  * Check for descriptors with their DD bit set and free mbufs.
54  * Return the total number of buffers freed.
55  */
56 static __rte_always_inline int
57 ngbe_tx_free_bufs(struct ngbe_tx_queue *txq)
58 {
59 	struct ngbe_tx_entry *txep;
60 	uint32_t status;
61 	int i, nb_free = 0;
62 	struct rte_mbuf *m, *free[RTE_NGBE_TX_MAX_FREE_BUF_SZ];
63 
64 	/* check DD bit on threshold descriptor */
65 	status = txq->tx_ring[txq->tx_next_dd].dw3;
66 	if (!(status & rte_cpu_to_le_32(NGBE_TXD_DD))) {
67 		if (txq->nb_tx_free >> 1 < txq->tx_free_thresh)
68 			ngbe_set32_masked(txq->tdc_reg_addr,
69 				NGBE_TXCFG_FLUSH, NGBE_TXCFG_FLUSH);
70 		return 0;
71 	}
72 
73 	/*
74 	 * first buffer to free from S/W ring is at index
75 	 * tx_next_dd - (tx_free_thresh-1)
76 	 */
77 	txep = &txq->sw_ring[txq->tx_next_dd - (txq->tx_free_thresh - 1)];
78 	for (i = 0; i < txq->tx_free_thresh; ++i, ++txep) {
79 		/* free buffers one at a time */
80 		m = rte_pktmbuf_prefree_seg(txep->mbuf);
81 		txep->mbuf = NULL;
82 
83 		if (unlikely(m == NULL))
84 			continue;
85 
86 		if (nb_free >= RTE_NGBE_TX_MAX_FREE_BUF_SZ ||
87 		    (nb_free > 0 && m->pool != free[0]->pool)) {
88 			rte_mempool_put_bulk(free[0]->pool,
89 					     (void **)free, nb_free);
90 			nb_free = 0;
91 		}
92 
93 		free[nb_free++] = m;
94 	}
95 
96 	if (nb_free > 0)
97 		rte_mempool_put_bulk(free[0]->pool, (void **)free, nb_free);
98 
99 	/* buffers were freed, update counters */
100 	txq->nb_tx_free = (uint16_t)(txq->nb_tx_free + txq->tx_free_thresh);
101 	txq->tx_next_dd = (uint16_t)(txq->tx_next_dd + txq->tx_free_thresh);
102 	if (txq->tx_next_dd >= txq->nb_tx_desc)
103 		txq->tx_next_dd = (uint16_t)(txq->tx_free_thresh - 1);
104 
105 	return txq->tx_free_thresh;
106 }
107 
108 /* Populate 4 descriptors with data from 4 mbufs */
109 static inline void
110 tx4(volatile struct ngbe_tx_desc *txdp, struct rte_mbuf **pkts)
111 {
112 	uint64_t buf_dma_addr;
113 	uint32_t pkt_len;
114 	int i;
115 
116 	for (i = 0; i < 4; ++i, ++txdp, ++pkts) {
117 		buf_dma_addr = rte_mbuf_data_iova(*pkts);
118 		pkt_len = (*pkts)->data_len;
119 
120 		/* write data to descriptor */
121 		txdp->qw0 = rte_cpu_to_le_64(buf_dma_addr);
122 		txdp->dw2 = cpu_to_le32(NGBE_TXD_FLAGS |
123 					NGBE_TXD_DATLEN(pkt_len));
124 		txdp->dw3 = cpu_to_le32(NGBE_TXD_PAYLEN(pkt_len));
125 
126 		rte_prefetch0(&(*pkts)->pool);
127 	}
128 }
129 
130 /* Populate 1 descriptor with data from 1 mbuf */
131 static inline void
132 tx1(volatile struct ngbe_tx_desc *txdp, struct rte_mbuf **pkts)
133 {
134 	uint64_t buf_dma_addr;
135 	uint32_t pkt_len;
136 
137 	buf_dma_addr = rte_mbuf_data_iova(*pkts);
138 	pkt_len = (*pkts)->data_len;
139 
140 	/* write data to descriptor */
141 	txdp->qw0 = cpu_to_le64(buf_dma_addr);
142 	txdp->dw2 = cpu_to_le32(NGBE_TXD_FLAGS |
143 				NGBE_TXD_DATLEN(pkt_len));
144 	txdp->dw3 = cpu_to_le32(NGBE_TXD_PAYLEN(pkt_len));
145 
146 	rte_prefetch0(&(*pkts)->pool);
147 }
148 
149 /*
150  * Fill H/W descriptor ring with mbuf data.
151  * Copy mbuf pointers to the S/W ring.
152  */
153 static inline void
154 ngbe_tx_fill_hw_ring(struct ngbe_tx_queue *txq, struct rte_mbuf **pkts,
155 		      uint16_t nb_pkts)
156 {
157 	volatile struct ngbe_tx_desc *txdp = &txq->tx_ring[txq->tx_tail];
158 	struct ngbe_tx_entry *txep = &txq->sw_ring[txq->tx_tail];
159 	const int N_PER_LOOP = 4;
160 	const int N_PER_LOOP_MASK = N_PER_LOOP - 1;
161 	int mainpart, leftover;
162 	int i, j;
163 
164 	/*
165 	 * Process most of the packets in chunks of N pkts.  Any
166 	 * leftover packets will get processed one at a time.
167 	 */
168 	mainpart = (nb_pkts & ((uint32_t)~N_PER_LOOP_MASK));
169 	leftover = (nb_pkts & ((uint32_t)N_PER_LOOP_MASK));
170 	for (i = 0; i < mainpart; i += N_PER_LOOP) {
171 		/* Copy N mbuf pointers to the S/W ring */
172 		for (j = 0; j < N_PER_LOOP; ++j)
173 			(txep + i + j)->mbuf = *(pkts + i + j);
174 		tx4(txdp + i, pkts + i);
175 	}
176 
177 	if (unlikely(leftover > 0)) {
178 		for (i = 0; i < leftover; ++i) {
179 			(txep + mainpart + i)->mbuf = *(pkts + mainpart + i);
180 			tx1(txdp + mainpart + i, pkts + mainpart + i);
181 		}
182 	}
183 }
184 
185 static inline uint16_t
186 tx_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts,
187 	     uint16_t nb_pkts)
188 {
189 	struct ngbe_tx_queue *txq = (struct ngbe_tx_queue *)tx_queue;
190 	uint16_t n = 0;
191 
192 	/*
193 	 * Begin scanning the H/W ring for done descriptors when the
194 	 * number of available descriptors drops below tx_free_thresh.
195 	 * For each done descriptor, free the associated buffer.
196 	 */
197 	if (txq->nb_tx_free < txq->tx_free_thresh)
198 		ngbe_tx_free_bufs(txq);
199 
200 	/* Only use descriptors that are available */
201 	nb_pkts = (uint16_t)RTE_MIN(txq->nb_tx_free, nb_pkts);
202 	if (unlikely(nb_pkts == 0))
203 		return 0;
204 
205 	/* Use exactly nb_pkts descriptors */
206 	txq->nb_tx_free = (uint16_t)(txq->nb_tx_free - nb_pkts);
207 
208 	/*
209 	 * At this point, we know there are enough descriptors in the
210 	 * ring to transmit all the packets.  This assumes that each
211 	 * mbuf contains a single segment, and that no new offloads
212 	 * are expected, which would require a new context descriptor.
213 	 */
214 
215 	/*
216 	 * See if we're going to wrap-around. If so, handle the top
217 	 * of the descriptor ring first, then do the bottom.  If not,
218 	 * the processing looks just like the "bottom" part anyway...
219 	 */
220 	if ((txq->tx_tail + nb_pkts) > txq->nb_tx_desc) {
221 		n = (uint16_t)(txq->nb_tx_desc - txq->tx_tail);
222 		ngbe_tx_fill_hw_ring(txq, tx_pkts, n);
223 		txq->tx_tail = 0;
224 	}
225 
226 	/* Fill H/W descriptor ring with mbuf data */
227 	ngbe_tx_fill_hw_ring(txq, tx_pkts + n, (uint16_t)(nb_pkts - n));
228 	txq->tx_tail = (uint16_t)(txq->tx_tail + (nb_pkts - n));
229 
230 	/*
231 	 * Check for wrap-around. This would only happen if we used
232 	 * up to the last descriptor in the ring, no more, no less.
233 	 */
234 	if (txq->tx_tail >= txq->nb_tx_desc)
235 		txq->tx_tail = 0;
236 
237 	PMD_TX_LOG(DEBUG, "port_id=%u queue_id=%u tx_tail=%u nb_tx=%u",
238 		   (uint16_t)txq->port_id, (uint16_t)txq->queue_id,
239 		   (uint16_t)txq->tx_tail, (uint16_t)nb_pkts);
240 
241 	/* update tail pointer */
242 	rte_wmb();
243 	ngbe_set32_relaxed(txq->tdt_reg_addr, txq->tx_tail);
244 
245 	return nb_pkts;
246 }
247 
248 uint16_t
249 ngbe_xmit_pkts_simple(void *tx_queue, struct rte_mbuf **tx_pkts,
250 		       uint16_t nb_pkts)
251 {
252 	uint16_t nb_tx;
253 
254 	/* Try to transmit at least chunks of TX_MAX_BURST pkts */
255 	if (likely(nb_pkts <= RTE_PMD_NGBE_TX_MAX_BURST))
256 		return tx_xmit_pkts(tx_queue, tx_pkts, nb_pkts);
257 
258 	/* transmit more than the max burst, in chunks of TX_MAX_BURST */
259 	nb_tx = 0;
260 	while (nb_pkts != 0) {
261 		uint16_t ret, n;
262 
263 		n = (uint16_t)RTE_MIN(nb_pkts, RTE_PMD_NGBE_TX_MAX_BURST);
264 		ret = tx_xmit_pkts(tx_queue, &tx_pkts[nb_tx], n);
265 		nb_tx = (uint16_t)(nb_tx + ret);
266 		nb_pkts = (uint16_t)(nb_pkts - ret);
267 		if (ret < n)
268 			break;
269 	}
270 
271 	return nb_tx;
272 }
273 
274 static inline void
275 ngbe_set_xmit_ctx(struct ngbe_tx_queue *txq,
276 		volatile struct ngbe_tx_ctx_desc *ctx_txd,
277 		uint64_t ol_flags, union ngbe_tx_offload tx_offload)
278 {
279 	union ngbe_tx_offload tx_offload_mask;
280 	uint32_t type_tucmd_mlhl;
281 	uint32_t mss_l4len_idx;
282 	uint32_t ctx_idx;
283 	uint32_t vlan_macip_lens;
284 	uint32_t tunnel_seed;
285 
286 	ctx_idx = txq->ctx_curr;
287 	tx_offload_mask.data[0] = 0;
288 	tx_offload_mask.data[1] = 0;
289 
290 	/* Specify which HW CTX to upload. */
291 	mss_l4len_idx = NGBE_TXD_IDX(ctx_idx);
292 	type_tucmd_mlhl = NGBE_TXD_CTXT;
293 
294 	tx_offload_mask.ptid |= ~0;
295 	type_tucmd_mlhl |= NGBE_TXD_PTID(tx_offload.ptid);
296 
297 	/* check if TCP segmentation required for this packet */
298 	if (ol_flags & RTE_MBUF_F_TX_TCP_SEG) {
299 		tx_offload_mask.l2_len |= ~0;
300 		tx_offload_mask.l3_len |= ~0;
301 		tx_offload_mask.l4_len |= ~0;
302 		tx_offload_mask.tso_segsz |= ~0;
303 		mss_l4len_idx |= NGBE_TXD_MSS(tx_offload.tso_segsz);
304 		mss_l4len_idx |= NGBE_TXD_L4LEN(tx_offload.l4_len);
305 	} else { /* no TSO, check if hardware checksum is needed */
306 		if (ol_flags & RTE_MBUF_F_TX_IP_CKSUM) {
307 			tx_offload_mask.l2_len |= ~0;
308 			tx_offload_mask.l3_len |= ~0;
309 		}
310 
311 		switch (ol_flags & RTE_MBUF_F_TX_L4_MASK) {
312 		case RTE_MBUF_F_TX_UDP_CKSUM:
313 			mss_l4len_idx |=
314 				NGBE_TXD_L4LEN(sizeof(struct rte_udp_hdr));
315 			tx_offload_mask.l2_len |= ~0;
316 			tx_offload_mask.l3_len |= ~0;
317 			break;
318 		case RTE_MBUF_F_TX_TCP_CKSUM:
319 			mss_l4len_idx |=
320 				NGBE_TXD_L4LEN(sizeof(struct rte_tcp_hdr));
321 			tx_offload_mask.l2_len |= ~0;
322 			tx_offload_mask.l3_len |= ~0;
323 			break;
324 		case RTE_MBUF_F_TX_SCTP_CKSUM:
325 			mss_l4len_idx |=
326 				NGBE_TXD_L4LEN(sizeof(struct rte_sctp_hdr));
327 			tx_offload_mask.l2_len |= ~0;
328 			tx_offload_mask.l3_len |= ~0;
329 			break;
330 		default:
331 			break;
332 		}
333 	}
334 
335 	vlan_macip_lens = NGBE_TXD_IPLEN(tx_offload.l3_len >> 1);
336 
337 	if (ol_flags & RTE_MBUF_F_TX_TUNNEL_MASK) {
338 		tx_offload_mask.outer_tun_len |= ~0;
339 		tx_offload_mask.outer_l2_len |= ~0;
340 		tx_offload_mask.outer_l3_len |= ~0;
341 		tx_offload_mask.l2_len |= ~0;
342 		tunnel_seed = NGBE_TXD_ETUNLEN(tx_offload.outer_tun_len >> 1);
343 		tunnel_seed |= NGBE_TXD_EIPLEN(tx_offload.outer_l3_len >> 2);
344 
345 		switch (ol_flags & RTE_MBUF_F_TX_TUNNEL_MASK) {
346 		case RTE_MBUF_F_TX_TUNNEL_IPIP:
347 			/* for non UDP / GRE tunneling, set to 0b */
348 			break;
349 		default:
350 			PMD_TX_LOG(ERR, "Tunnel type not supported");
351 			return;
352 		}
353 		vlan_macip_lens |= NGBE_TXD_MACLEN(tx_offload.outer_l2_len);
354 	} else {
355 		tunnel_seed = 0;
356 		vlan_macip_lens |= NGBE_TXD_MACLEN(tx_offload.l2_len);
357 	}
358 
359 	if (ol_flags & RTE_MBUF_F_TX_VLAN) {
360 		tx_offload_mask.vlan_tci |= ~0;
361 		vlan_macip_lens |= NGBE_TXD_VLAN(tx_offload.vlan_tci);
362 	}
363 
364 	txq->ctx_cache[ctx_idx].flags = ol_flags;
365 	txq->ctx_cache[ctx_idx].tx_offload.data[0] =
366 		tx_offload_mask.data[0] & tx_offload.data[0];
367 	txq->ctx_cache[ctx_idx].tx_offload.data[1] =
368 		tx_offload_mask.data[1] & tx_offload.data[1];
369 	txq->ctx_cache[ctx_idx].tx_offload_mask = tx_offload_mask;
370 
371 	ctx_txd->dw0 = rte_cpu_to_le_32(vlan_macip_lens);
372 	ctx_txd->dw1 = rte_cpu_to_le_32(tunnel_seed);
373 	ctx_txd->dw2 = rte_cpu_to_le_32(type_tucmd_mlhl);
374 	ctx_txd->dw3 = rte_cpu_to_le_32(mss_l4len_idx);
375 }
376 
377 /*
378  * Check which hardware context can be used. Use the existing match
379  * or create a new context descriptor.
380  */
381 static inline uint32_t
382 what_ctx_update(struct ngbe_tx_queue *txq, uint64_t flags,
383 		   union ngbe_tx_offload tx_offload)
384 {
385 	/* If match with the current used context */
386 	if (likely(txq->ctx_cache[txq->ctx_curr].flags == flags &&
387 		   (txq->ctx_cache[txq->ctx_curr].tx_offload.data[0] ==
388 		    (txq->ctx_cache[txq->ctx_curr].tx_offload_mask.data[0]
389 		     & tx_offload.data[0])) &&
390 		   (txq->ctx_cache[txq->ctx_curr].tx_offload.data[1] ==
391 		    (txq->ctx_cache[txq->ctx_curr].tx_offload_mask.data[1]
392 		     & tx_offload.data[1]))))
393 		return txq->ctx_curr;
394 
395 	/* What if match with the next context  */
396 	txq->ctx_curr ^= 1;
397 	if (likely(txq->ctx_cache[txq->ctx_curr].flags == flags &&
398 		   (txq->ctx_cache[txq->ctx_curr].tx_offload.data[0] ==
399 		    (txq->ctx_cache[txq->ctx_curr].tx_offload_mask.data[0]
400 		     & tx_offload.data[0])) &&
401 		   (txq->ctx_cache[txq->ctx_curr].tx_offload.data[1] ==
402 		    (txq->ctx_cache[txq->ctx_curr].tx_offload_mask.data[1]
403 		     & tx_offload.data[1]))))
404 		return txq->ctx_curr;
405 
406 	/* Mismatch, use the previous context */
407 	return NGBE_CTX_NUM;
408 }
409 
410 static inline uint32_t
411 tx_desc_cksum_flags_to_olinfo(uint64_t ol_flags)
412 {
413 	uint32_t tmp = 0;
414 
415 	if ((ol_flags & RTE_MBUF_F_TX_L4_MASK) != RTE_MBUF_F_TX_L4_NO_CKSUM) {
416 		tmp |= NGBE_TXD_CC;
417 		tmp |= NGBE_TXD_L4CS;
418 	}
419 	if (ol_flags & RTE_MBUF_F_TX_IP_CKSUM) {
420 		tmp |= NGBE_TXD_CC;
421 		tmp |= NGBE_TXD_IPCS;
422 	}
423 	if (ol_flags & RTE_MBUF_F_TX_OUTER_IP_CKSUM) {
424 		tmp |= NGBE_TXD_CC;
425 		tmp |= NGBE_TXD_EIPCS;
426 	}
427 	if (ol_flags & RTE_MBUF_F_TX_TCP_SEG) {
428 		tmp |= NGBE_TXD_CC;
429 		/* implies IPv4 cksum */
430 		if (ol_flags & RTE_MBUF_F_TX_IPV4)
431 			tmp |= NGBE_TXD_IPCS;
432 		tmp |= NGBE_TXD_L4CS;
433 	}
434 	if (ol_flags & RTE_MBUF_F_TX_VLAN)
435 		tmp |= NGBE_TXD_CC;
436 
437 	return tmp;
438 }
439 
440 static inline uint32_t
441 tx_desc_ol_flags_to_cmdtype(uint64_t ol_flags)
442 {
443 	uint32_t cmdtype = 0;
444 
445 	if (ol_flags & RTE_MBUF_F_TX_VLAN)
446 		cmdtype |= NGBE_TXD_VLE;
447 	if (ol_flags & RTE_MBUF_F_TX_TCP_SEG)
448 		cmdtype |= NGBE_TXD_TSE;
449 	return cmdtype;
450 }
451 
452 static inline uint8_t
453 tx_desc_ol_flags_to_ptid(uint64_t oflags, uint32_t ptype)
454 {
455 	bool tun;
456 
457 	if (ptype)
458 		return ngbe_encode_ptype(ptype);
459 
460 	/* Only support flags in NGBE_TX_OFFLOAD_MASK */
461 	tun = !!(oflags & RTE_MBUF_F_TX_TUNNEL_MASK);
462 
463 	/* L2 level */
464 	ptype = RTE_PTYPE_L2_ETHER;
465 	if (oflags & RTE_MBUF_F_TX_VLAN)
466 		ptype |= RTE_PTYPE_L2_ETHER_VLAN;
467 
468 	/* L3 level */
469 	if (oflags & (RTE_MBUF_F_TX_OUTER_IPV4 | RTE_MBUF_F_TX_OUTER_IP_CKSUM))
470 		ptype |= RTE_PTYPE_L3_IPV4;
471 	else if (oflags & (RTE_MBUF_F_TX_OUTER_IPV6))
472 		ptype |= RTE_PTYPE_L3_IPV6;
473 
474 	if (oflags & (RTE_MBUF_F_TX_IPV4 | RTE_MBUF_F_TX_IP_CKSUM))
475 		ptype |= (tun ? RTE_PTYPE_INNER_L3_IPV4 : RTE_PTYPE_L3_IPV4);
476 	else if (oflags & (RTE_MBUF_F_TX_IPV6))
477 		ptype |= (tun ? RTE_PTYPE_INNER_L3_IPV6 : RTE_PTYPE_L3_IPV6);
478 
479 	/* L4 level */
480 	switch (oflags & (RTE_MBUF_F_TX_L4_MASK)) {
481 	case RTE_MBUF_F_TX_TCP_CKSUM:
482 		ptype |= (tun ? RTE_PTYPE_INNER_L4_TCP : RTE_PTYPE_L4_TCP);
483 		break;
484 	case RTE_MBUF_F_TX_UDP_CKSUM:
485 		ptype |= (tun ? RTE_PTYPE_INNER_L4_UDP : RTE_PTYPE_L4_UDP);
486 		break;
487 	case RTE_MBUF_F_TX_SCTP_CKSUM:
488 		ptype |= (tun ? RTE_PTYPE_INNER_L4_SCTP : RTE_PTYPE_L4_SCTP);
489 		break;
490 	}
491 
492 	if (oflags & RTE_MBUF_F_TX_TCP_SEG)
493 		ptype |= (tun ? RTE_PTYPE_INNER_L4_TCP : RTE_PTYPE_L4_TCP);
494 
495 	/* Tunnel */
496 	switch (oflags & RTE_MBUF_F_TX_TUNNEL_MASK) {
497 	case RTE_MBUF_F_TX_TUNNEL_IPIP:
498 	case RTE_MBUF_F_TX_TUNNEL_IP:
499 		ptype |= RTE_PTYPE_L2_ETHER |
500 			 RTE_PTYPE_L3_IPV4 |
501 			 RTE_PTYPE_TUNNEL_IP;
502 		break;
503 	}
504 
505 	return ngbe_encode_ptype(ptype);
506 }
507 
508 /* Reset transmit descriptors after they have been used */
509 static inline int
510 ngbe_xmit_cleanup(struct ngbe_tx_queue *txq)
511 {
512 	struct ngbe_tx_entry *sw_ring = txq->sw_ring;
513 	volatile struct ngbe_tx_desc *txr = txq->tx_ring;
514 	uint16_t last_desc_cleaned = txq->last_desc_cleaned;
515 	uint16_t nb_tx_desc = txq->nb_tx_desc;
516 	uint16_t desc_to_clean_to;
517 	uint16_t nb_tx_to_clean;
518 	uint32_t status;
519 
520 	/* Determine the last descriptor needing to be cleaned */
521 	desc_to_clean_to = (uint16_t)(last_desc_cleaned + txq->tx_free_thresh);
522 	if (desc_to_clean_to >= nb_tx_desc)
523 		desc_to_clean_to = (uint16_t)(desc_to_clean_to - nb_tx_desc);
524 
525 	/* Check to make sure the last descriptor to clean is done */
526 	desc_to_clean_to = sw_ring[desc_to_clean_to].last_id;
527 	status = txr[desc_to_clean_to].dw3;
528 	if (!(status & rte_cpu_to_le_32(NGBE_TXD_DD))) {
529 		PMD_TX_LOG(DEBUG,
530 			"Tx descriptor %4u is not done"
531 			"(port=%d queue=%d)",
532 			desc_to_clean_to,
533 			txq->port_id, txq->queue_id);
534 		if (txq->nb_tx_free >> 1 < txq->tx_free_thresh)
535 			ngbe_set32_masked(txq->tdc_reg_addr,
536 				NGBE_TXCFG_FLUSH, NGBE_TXCFG_FLUSH);
537 		/* Failed to clean any descriptors, better luck next time */
538 		return -(1);
539 	}
540 
541 	/* Figure out how many descriptors will be cleaned */
542 	if (last_desc_cleaned > desc_to_clean_to)
543 		nb_tx_to_clean = (uint16_t)((nb_tx_desc - last_desc_cleaned) +
544 							desc_to_clean_to);
545 	else
546 		nb_tx_to_clean = (uint16_t)(desc_to_clean_to -
547 						last_desc_cleaned);
548 
549 	PMD_TX_LOG(DEBUG,
550 		"Cleaning %4u Tx descriptors: %4u to %4u (port=%d queue=%d)",
551 		nb_tx_to_clean, last_desc_cleaned, desc_to_clean_to,
552 		txq->port_id, txq->queue_id);
553 
554 	/*
555 	 * The last descriptor to clean is done, so that means all the
556 	 * descriptors from the last descriptor that was cleaned
557 	 * up to the last descriptor with the RS bit set
558 	 * are done. Only reset the threshold descriptor.
559 	 */
560 	txr[desc_to_clean_to].dw3 = 0;
561 
562 	/* Update the txq to reflect the last descriptor that was cleaned */
563 	txq->last_desc_cleaned = desc_to_clean_to;
564 	txq->nb_tx_free = (uint16_t)(txq->nb_tx_free + nb_tx_to_clean);
565 
566 	/* No Error */
567 	return 0;
568 }
569 
570 uint16_t
571 ngbe_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts,
572 		uint16_t nb_pkts)
573 {
574 	struct ngbe_tx_queue *txq;
575 	struct ngbe_tx_entry *sw_ring;
576 	struct ngbe_tx_entry *txe, *txn;
577 	volatile struct ngbe_tx_desc *txr;
578 	volatile struct ngbe_tx_desc *txd;
579 	struct rte_mbuf     *tx_pkt;
580 	struct rte_mbuf     *m_seg;
581 	uint64_t buf_dma_addr;
582 	uint32_t olinfo_status;
583 	uint32_t cmd_type_len;
584 	uint32_t pkt_len;
585 	uint16_t slen;
586 	uint64_t ol_flags;
587 	uint16_t tx_id;
588 	uint16_t tx_last;
589 	uint16_t nb_tx;
590 	uint16_t nb_used;
591 	uint64_t tx_ol_req;
592 	uint32_t ctx = 0;
593 	uint32_t new_ctx;
594 	union ngbe_tx_offload tx_offload;
595 
596 	tx_offload.data[0] = 0;
597 	tx_offload.data[1] = 0;
598 	txq = tx_queue;
599 	sw_ring = txq->sw_ring;
600 	txr     = txq->tx_ring;
601 	tx_id   = txq->tx_tail;
602 	txe = &sw_ring[tx_id];
603 
604 	/* Determine if the descriptor ring needs to be cleaned. */
605 	if (txq->nb_tx_free < txq->tx_free_thresh)
606 		ngbe_xmit_cleanup(txq);
607 
608 	rte_prefetch0(&txe->mbuf->pool);
609 
610 	/* Tx loop */
611 	for (nb_tx = 0; nb_tx < nb_pkts; nb_tx++) {
612 		new_ctx = 0;
613 		tx_pkt = *tx_pkts++;
614 		pkt_len = tx_pkt->pkt_len;
615 
616 		/*
617 		 * Determine how many (if any) context descriptors
618 		 * are needed for offload functionality.
619 		 */
620 		ol_flags = tx_pkt->ol_flags;
621 
622 		/* If hardware offload required */
623 		tx_ol_req = ol_flags & NGBE_TX_OFFLOAD_MASK;
624 		if (tx_ol_req) {
625 			tx_offload.ptid = tx_desc_ol_flags_to_ptid(tx_ol_req,
626 					tx_pkt->packet_type);
627 			tx_offload.l2_len = tx_pkt->l2_len;
628 			tx_offload.l3_len = tx_pkt->l3_len;
629 			tx_offload.l4_len = tx_pkt->l4_len;
630 			tx_offload.vlan_tci = tx_pkt->vlan_tci;
631 			tx_offload.tso_segsz = tx_pkt->tso_segsz;
632 			tx_offload.outer_l2_len = tx_pkt->outer_l2_len;
633 			tx_offload.outer_l3_len = tx_pkt->outer_l3_len;
634 			tx_offload.outer_tun_len = 0;
635 
636 			/* If new context need be built or reuse the exist ctx*/
637 			ctx = what_ctx_update(txq, tx_ol_req, tx_offload);
638 			/* Only allocate context descriptor if required */
639 			new_ctx = (ctx == NGBE_CTX_NUM);
640 			ctx = txq->ctx_curr;
641 		}
642 
643 		/*
644 		 * Keep track of how many descriptors are used this loop
645 		 * This will always be the number of segments + the number of
646 		 * Context descriptors required to transmit the packet
647 		 */
648 		nb_used = (uint16_t)(tx_pkt->nb_segs + new_ctx);
649 
650 		/*
651 		 * The number of descriptors that must be allocated for a
652 		 * packet is the number of segments of that packet, plus 1
653 		 * Context Descriptor for the hardware offload, if any.
654 		 * Determine the last Tx descriptor to allocate in the Tx ring
655 		 * for the packet, starting from the current position (tx_id)
656 		 * in the ring.
657 		 */
658 		tx_last = (uint16_t)(tx_id + nb_used - 1);
659 
660 		/* Circular ring */
661 		if (tx_last >= txq->nb_tx_desc)
662 			tx_last = (uint16_t)(tx_last - txq->nb_tx_desc);
663 
664 		PMD_TX_LOG(DEBUG, "port_id=%u queue_id=%u pktlen=%u"
665 			   " tx_first=%u tx_last=%u",
666 			   (uint16_t)txq->port_id,
667 			   (uint16_t)txq->queue_id,
668 			   (uint32_t)pkt_len,
669 			   (uint16_t)tx_id,
670 			   (uint16_t)tx_last);
671 
672 		/*
673 		 * Make sure there are enough Tx descriptors available to
674 		 * transmit the entire packet.
675 		 * nb_used better be less than or equal to txq->tx_free_thresh
676 		 */
677 		if (nb_used > txq->nb_tx_free) {
678 			PMD_TX_LOG(DEBUG,
679 				"Not enough free Tx descriptors "
680 				"nb_used=%4u nb_free=%4u "
681 				"(port=%d queue=%d)",
682 				nb_used, txq->nb_tx_free,
683 				txq->port_id, txq->queue_id);
684 
685 			if (ngbe_xmit_cleanup(txq) != 0) {
686 				/* Could not clean any descriptors */
687 				if (nb_tx == 0)
688 					return 0;
689 				goto end_of_tx;
690 			}
691 
692 			/* nb_used better be <= txq->tx_free_thresh */
693 			if (unlikely(nb_used > txq->tx_free_thresh)) {
694 				PMD_TX_LOG(DEBUG,
695 					"The number of descriptors needed to "
696 					"transmit the packet exceeds the "
697 					"RS bit threshold. This will impact "
698 					"performance."
699 					"nb_used=%4u nb_free=%4u "
700 					"tx_free_thresh=%4u. "
701 					"(port=%d queue=%d)",
702 					nb_used, txq->nb_tx_free,
703 					txq->tx_free_thresh,
704 					txq->port_id, txq->queue_id);
705 				/*
706 				 * Loop here until there are enough Tx
707 				 * descriptors or until the ring cannot be
708 				 * cleaned.
709 				 */
710 				while (nb_used > txq->nb_tx_free) {
711 					if (ngbe_xmit_cleanup(txq) != 0) {
712 						/*
713 						 * Could not clean any
714 						 * descriptors
715 						 */
716 						if (nb_tx == 0)
717 							return 0;
718 						goto end_of_tx;
719 					}
720 				}
721 			}
722 		}
723 
724 		/*
725 		 * By now there are enough free Tx descriptors to transmit
726 		 * the packet.
727 		 */
728 
729 		/*
730 		 * Set common flags of all Tx Data Descriptors.
731 		 *
732 		 * The following bits must be set in the first Data Descriptor
733 		 * and are ignored in the other ones:
734 		 *   - NGBE_TXD_FCS
735 		 *
736 		 * The following bits must only be set in the last Data
737 		 * Descriptor:
738 		 *   - NGBE_TXD_EOP
739 		 */
740 		cmd_type_len = NGBE_TXD_FCS;
741 
742 #ifdef RTE_LIBRTE_IEEE1588
743 		if (ol_flags & PKT_TX_IEEE1588_TMST)
744 			cmd_type_len |= NGBE_TXD_1588;
745 #endif
746 
747 		olinfo_status = 0;
748 		if (tx_ol_req) {
749 			if (ol_flags & RTE_MBUF_F_TX_TCP_SEG) {
750 				/* when TSO is on, paylen in descriptor is the
751 				 * not the packet len but the tcp payload len
752 				 */
753 				pkt_len -= (tx_offload.l2_len +
754 					tx_offload.l3_len + tx_offload.l4_len);
755 				pkt_len -=
756 					(tx_pkt->ol_flags & RTE_MBUF_F_TX_TUNNEL_MASK)
757 					? tx_offload.outer_l2_len +
758 					  tx_offload.outer_l3_len : 0;
759 			}
760 
761 			/*
762 			 * Setup the Tx Context Descriptor if required
763 			 */
764 			if (new_ctx) {
765 				volatile struct ngbe_tx_ctx_desc *ctx_txd;
766 
767 				ctx_txd = (volatile struct ngbe_tx_ctx_desc *)
768 				    &txr[tx_id];
769 
770 				txn = &sw_ring[txe->next_id];
771 				rte_prefetch0(&txn->mbuf->pool);
772 
773 				if (txe->mbuf != NULL) {
774 					rte_pktmbuf_free_seg(txe->mbuf);
775 					txe->mbuf = NULL;
776 				}
777 
778 				ngbe_set_xmit_ctx(txq, ctx_txd, tx_ol_req,
779 					tx_offload);
780 
781 				txe->last_id = tx_last;
782 				tx_id = txe->next_id;
783 				txe = txn;
784 			}
785 
786 			/*
787 			 * Setup the Tx Data Descriptor,
788 			 * This path will go through
789 			 * whatever new/reuse the context descriptor
790 			 */
791 			cmd_type_len  |= tx_desc_ol_flags_to_cmdtype(ol_flags);
792 			olinfo_status |=
793 				tx_desc_cksum_flags_to_olinfo(ol_flags);
794 			olinfo_status |= NGBE_TXD_IDX(ctx);
795 		}
796 
797 		olinfo_status |= NGBE_TXD_PAYLEN(pkt_len);
798 
799 		m_seg = tx_pkt;
800 		do {
801 			txd = &txr[tx_id];
802 			txn = &sw_ring[txe->next_id];
803 			rte_prefetch0(&txn->mbuf->pool);
804 
805 			if (txe->mbuf != NULL)
806 				rte_pktmbuf_free_seg(txe->mbuf);
807 			txe->mbuf = m_seg;
808 
809 			/*
810 			 * Set up Transmit Data Descriptor.
811 			 */
812 			slen = m_seg->data_len;
813 			buf_dma_addr = rte_mbuf_data_iova(m_seg);
814 			txd->qw0 = rte_cpu_to_le_64(buf_dma_addr);
815 			txd->dw2 = rte_cpu_to_le_32(cmd_type_len | slen);
816 			txd->dw3 = rte_cpu_to_le_32(olinfo_status);
817 			txe->last_id = tx_last;
818 			tx_id = txe->next_id;
819 			txe = txn;
820 			m_seg = m_seg->next;
821 		} while (m_seg != NULL);
822 
823 		/*
824 		 * The last packet data descriptor needs End Of Packet (EOP)
825 		 */
826 		cmd_type_len |= NGBE_TXD_EOP;
827 		txq->nb_tx_free = (uint16_t)(txq->nb_tx_free - nb_used);
828 
829 		txd->dw2 |= rte_cpu_to_le_32(cmd_type_len);
830 	}
831 
832 end_of_tx:
833 
834 	rte_wmb();
835 
836 	/*
837 	 * Set the Transmit Descriptor Tail (TDT)
838 	 */
839 	PMD_TX_LOG(DEBUG, "port_id=%u queue_id=%u tx_tail=%u nb_tx=%u",
840 		   (uint16_t)txq->port_id, (uint16_t)txq->queue_id,
841 		   (uint16_t)tx_id, (uint16_t)nb_tx);
842 	ngbe_set32_relaxed(txq->tdt_reg_addr, tx_id);
843 	txq->tx_tail = tx_id;
844 
845 	return nb_tx;
846 }
847 
848 /*********************************************************************
849  *
850  *  Tx prep functions
851  *
852  **********************************************************************/
853 uint16_t
854 ngbe_prep_pkts(void *tx_queue, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
855 {
856 	int i, ret;
857 	uint64_t ol_flags;
858 	struct rte_mbuf *m;
859 	struct ngbe_tx_queue *txq = (struct ngbe_tx_queue *)tx_queue;
860 
861 	for (i = 0; i < nb_pkts; i++) {
862 		m = tx_pkts[i];
863 		ol_flags = m->ol_flags;
864 
865 		/**
866 		 * Check if packet meets requirements for number of segments
867 		 *
868 		 * NOTE: for ngbe it's always (40 - WTHRESH) for both TSO and
869 		 *       non-TSO
870 		 */
871 
872 		if (m->nb_segs > NGBE_TX_MAX_SEG - txq->wthresh) {
873 			rte_errno = -EINVAL;
874 			return i;
875 		}
876 
877 		if (ol_flags & NGBE_TX_OFFLOAD_NOTSUP_MASK) {
878 			rte_errno = -ENOTSUP;
879 			return i;
880 		}
881 
882 #ifdef RTE_ETHDEV_DEBUG_TX
883 		ret = rte_validate_tx_offload(m);
884 		if (ret != 0) {
885 			rte_errno = ret;
886 			return i;
887 		}
888 #endif
889 		ret = rte_net_intel_cksum_prepare(m);
890 		if (ret != 0) {
891 			rte_errno = ret;
892 			return i;
893 		}
894 	}
895 
896 	return i;
897 }
898 
899 /*********************************************************************
900  *
901  *  Rx functions
902  *
903  **********************************************************************/
904 static inline uint32_t
905 ngbe_rxd_pkt_info_to_pkt_type(uint32_t pkt_info, uint16_t ptid_mask)
906 {
907 	uint16_t ptid = NGBE_RXD_PTID(pkt_info);
908 
909 	ptid &= ptid_mask;
910 
911 	return ngbe_decode_ptype(ptid);
912 }
913 
914 static inline uint64_t
915 ngbe_rxd_pkt_info_to_pkt_flags(uint32_t pkt_info)
916 {
917 	static uint64_t ip_rss_types_map[16] __rte_cache_aligned = {
918 		0, RTE_MBUF_F_RX_RSS_HASH, RTE_MBUF_F_RX_RSS_HASH, RTE_MBUF_F_RX_RSS_HASH,
919 		0, RTE_MBUF_F_RX_RSS_HASH, 0, RTE_MBUF_F_RX_RSS_HASH,
920 		RTE_MBUF_F_RX_RSS_HASH, 0, 0, 0,
921 		0, 0, 0,  RTE_MBUF_F_RX_FDIR,
922 	};
923 #ifdef RTE_LIBRTE_IEEE1588
924 	static uint64_t ip_pkt_etqf_map[8] = {
925 		0, 0, 0, PKT_RX_IEEE1588_PTP,
926 		0, 0, 0, 0,
927 	};
928 	int etfid = ngbe_etflt_id(NGBE_RXD_PTID(pkt_info));
929 	if (likely(-1 != etfid))
930 		return ip_pkt_etqf_map[etfid] |
931 		       ip_rss_types_map[NGBE_RXD_RSSTYPE(pkt_info)];
932 	else
933 		return ip_rss_types_map[NGBE_RXD_RSSTYPE(pkt_info)];
934 #else
935 	return ip_rss_types_map[NGBE_RXD_RSSTYPE(pkt_info)];
936 #endif
937 }
938 
939 static inline uint64_t
940 rx_desc_status_to_pkt_flags(uint32_t rx_status, uint64_t vlan_flags)
941 {
942 	uint64_t pkt_flags;
943 
944 	/*
945 	 * Check if VLAN present only.
946 	 * Do not check whether L3/L4 rx checksum done by NIC or not,
947 	 * That can be found from rte_eth_rxmode.offloads flag
948 	 */
949 	pkt_flags = (rx_status & NGBE_RXD_STAT_VLAN &&
950 		     vlan_flags & RTE_MBUF_F_RX_VLAN_STRIPPED)
951 		    ? vlan_flags : 0;
952 
953 #ifdef RTE_LIBRTE_IEEE1588
954 	if (rx_status & NGBE_RXD_STAT_1588)
955 		pkt_flags = pkt_flags | PKT_RX_IEEE1588_TMST;
956 #endif
957 	return pkt_flags;
958 }
959 
960 static inline uint64_t
961 rx_desc_error_to_pkt_flags(uint32_t rx_status)
962 {
963 	uint64_t pkt_flags = 0;
964 
965 	/* checksum offload can't be disabled */
966 	if (rx_status & NGBE_RXD_STAT_IPCS)
967 		pkt_flags |= (rx_status & NGBE_RXD_ERR_IPCS
968 				? RTE_MBUF_F_RX_IP_CKSUM_BAD : RTE_MBUF_F_RX_IP_CKSUM_GOOD);
969 
970 	if (rx_status & NGBE_RXD_STAT_L4CS)
971 		pkt_flags |= (rx_status & NGBE_RXD_ERR_L4CS
972 				? RTE_MBUF_F_RX_L4_CKSUM_BAD : RTE_MBUF_F_RX_L4_CKSUM_GOOD);
973 
974 	if (rx_status & NGBE_RXD_STAT_EIPCS &&
975 	    rx_status & NGBE_RXD_ERR_EIPCS)
976 		pkt_flags |= RTE_MBUF_F_RX_OUTER_IP_CKSUM_BAD;
977 
978 	return pkt_flags;
979 }
980 
981 /*
982  * LOOK_AHEAD defines how many desc statuses to check beyond the
983  * current descriptor.
984  * It must be a pound define for optimal performance.
985  * Do not change the value of LOOK_AHEAD, as the ngbe_rx_scan_hw_ring
986  * function only works with LOOK_AHEAD=8.
987  */
988 #define LOOK_AHEAD 8
989 #if (LOOK_AHEAD != 8)
990 #error "PMD NGBE: LOOK_AHEAD must be 8\n"
991 #endif
992 static inline int
993 ngbe_rx_scan_hw_ring(struct ngbe_rx_queue *rxq)
994 {
995 	volatile struct ngbe_rx_desc *rxdp;
996 	struct ngbe_rx_entry *rxep;
997 	struct rte_mbuf *mb;
998 	uint16_t pkt_len;
999 	uint64_t pkt_flags;
1000 	int nb_dd;
1001 	uint32_t s[LOOK_AHEAD];
1002 	uint32_t pkt_info[LOOK_AHEAD];
1003 	int i, j, nb_rx = 0;
1004 	uint32_t status;
1005 
1006 	/* get references to current descriptor and S/W ring entry */
1007 	rxdp = &rxq->rx_ring[rxq->rx_tail];
1008 	rxep = &rxq->sw_ring[rxq->rx_tail];
1009 
1010 	status = rxdp->qw1.lo.status;
1011 	/* check to make sure there is at least 1 packet to receive */
1012 	if (!(status & rte_cpu_to_le_32(NGBE_RXD_STAT_DD)))
1013 		return 0;
1014 
1015 	/*
1016 	 * Scan LOOK_AHEAD descriptors at a time to determine which descriptors
1017 	 * reference packets that are ready to be received.
1018 	 */
1019 	for (i = 0; i < RTE_PMD_NGBE_RX_MAX_BURST;
1020 	     i += LOOK_AHEAD, rxdp += LOOK_AHEAD, rxep += LOOK_AHEAD) {
1021 		/* Read desc statuses backwards to avoid race condition */
1022 		for (j = 0; j < LOOK_AHEAD; j++)
1023 			s[j] = rte_le_to_cpu_32(rxdp[j].qw1.lo.status);
1024 
1025 		rte_atomic_thread_fence(__ATOMIC_ACQUIRE);
1026 
1027 		/* Compute how many status bits were set */
1028 		for (nb_dd = 0; nb_dd < LOOK_AHEAD &&
1029 				(s[nb_dd] & NGBE_RXD_STAT_DD); nb_dd++)
1030 			;
1031 
1032 		for (j = 0; j < nb_dd; j++)
1033 			pkt_info[j] = rte_le_to_cpu_32(rxdp[j].qw0.dw0);
1034 
1035 		nb_rx += nb_dd;
1036 
1037 		/* Translate descriptor info to mbuf format */
1038 		for (j = 0; j < nb_dd; ++j) {
1039 			mb = rxep[j].mbuf;
1040 			pkt_len = rte_le_to_cpu_16(rxdp[j].qw1.hi.len) -
1041 				  rxq->crc_len;
1042 			mb->data_len = pkt_len;
1043 			mb->pkt_len = pkt_len;
1044 			mb->vlan_tci = rte_le_to_cpu_16(rxdp[j].qw1.hi.tag);
1045 
1046 			/* convert descriptor fields to rte mbuf flags */
1047 			pkt_flags = rx_desc_status_to_pkt_flags(s[j],
1048 					rxq->vlan_flags);
1049 			pkt_flags |= rx_desc_error_to_pkt_flags(s[j]);
1050 			pkt_flags |=
1051 				ngbe_rxd_pkt_info_to_pkt_flags(pkt_info[j]);
1052 			mb->ol_flags = pkt_flags;
1053 			mb->packet_type =
1054 				ngbe_rxd_pkt_info_to_pkt_type(pkt_info[j],
1055 				NGBE_PTID_MASK);
1056 
1057 			if (likely(pkt_flags & RTE_MBUF_F_RX_RSS_HASH))
1058 				mb->hash.rss =
1059 					rte_le_to_cpu_32(rxdp[j].qw0.dw1);
1060 		}
1061 
1062 		/* Move mbuf pointers from the S/W ring to the stage */
1063 		for (j = 0; j < LOOK_AHEAD; ++j)
1064 			rxq->rx_stage[i + j] = rxep[j].mbuf;
1065 
1066 		/* stop if all requested packets could not be received */
1067 		if (nb_dd != LOOK_AHEAD)
1068 			break;
1069 	}
1070 
1071 	/* clear software ring entries so we can cleanup correctly */
1072 	for (i = 0; i < nb_rx; ++i)
1073 		rxq->sw_ring[rxq->rx_tail + i].mbuf = NULL;
1074 
1075 	return nb_rx;
1076 }
1077 
1078 static inline int
1079 ngbe_rx_alloc_bufs(struct ngbe_rx_queue *rxq, bool reset_mbuf)
1080 {
1081 	volatile struct ngbe_rx_desc *rxdp;
1082 	struct ngbe_rx_entry *rxep;
1083 	struct rte_mbuf *mb;
1084 	uint16_t alloc_idx;
1085 	__le64 dma_addr;
1086 	int diag, i;
1087 
1088 	/* allocate buffers in bulk directly into the S/W ring */
1089 	alloc_idx = rxq->rx_free_trigger - (rxq->rx_free_thresh - 1);
1090 	rxep = &rxq->sw_ring[alloc_idx];
1091 	diag = rte_mempool_get_bulk(rxq->mb_pool, (void *)rxep,
1092 				    rxq->rx_free_thresh);
1093 	if (unlikely(diag != 0))
1094 		return -ENOMEM;
1095 
1096 	rxdp = &rxq->rx_ring[alloc_idx];
1097 	for (i = 0; i < rxq->rx_free_thresh; ++i) {
1098 		/* populate the static rte mbuf fields */
1099 		mb = rxep[i].mbuf;
1100 		if (reset_mbuf)
1101 			mb->port = rxq->port_id;
1102 
1103 		rte_mbuf_refcnt_set(mb, 1);
1104 		mb->data_off = RTE_PKTMBUF_HEADROOM;
1105 
1106 		/* populate the descriptors */
1107 		dma_addr = rte_cpu_to_le_64(rte_mbuf_data_iova_default(mb));
1108 		NGBE_RXD_HDRADDR(&rxdp[i], 0);
1109 		NGBE_RXD_PKTADDR(&rxdp[i], dma_addr);
1110 	}
1111 
1112 	/* update state of internal queue structure */
1113 	rxq->rx_free_trigger = rxq->rx_free_trigger + rxq->rx_free_thresh;
1114 	if (rxq->rx_free_trigger >= rxq->nb_rx_desc)
1115 		rxq->rx_free_trigger = rxq->rx_free_thresh - 1;
1116 
1117 	/* no errors */
1118 	return 0;
1119 }
1120 
1121 static inline uint16_t
1122 ngbe_rx_fill_from_stage(struct ngbe_rx_queue *rxq, struct rte_mbuf **rx_pkts,
1123 			 uint16_t nb_pkts)
1124 {
1125 	struct rte_mbuf **stage = &rxq->rx_stage[rxq->rx_next_avail];
1126 	int i;
1127 
1128 	/* how many packets are ready to return? */
1129 	nb_pkts = (uint16_t)RTE_MIN(nb_pkts, rxq->rx_nb_avail);
1130 
1131 	/* copy mbuf pointers to the application's packet list */
1132 	for (i = 0; i < nb_pkts; ++i)
1133 		rx_pkts[i] = stage[i];
1134 
1135 	/* update internal queue state */
1136 	rxq->rx_nb_avail = (uint16_t)(rxq->rx_nb_avail - nb_pkts);
1137 	rxq->rx_next_avail = (uint16_t)(rxq->rx_next_avail + nb_pkts);
1138 
1139 	return nb_pkts;
1140 }
1141 
1142 static inline uint16_t
1143 ngbe_rx_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts,
1144 	     uint16_t nb_pkts)
1145 {
1146 	struct ngbe_rx_queue *rxq = (struct ngbe_rx_queue *)rx_queue;
1147 	struct rte_eth_dev *dev = &rte_eth_devices[rxq->port_id];
1148 	uint16_t nb_rx = 0;
1149 
1150 	/* Any previously recv'd pkts will be returned from the Rx stage */
1151 	if (rxq->rx_nb_avail)
1152 		return ngbe_rx_fill_from_stage(rxq, rx_pkts, nb_pkts);
1153 
1154 	/* Scan the H/W ring for packets to receive */
1155 	nb_rx = (uint16_t)ngbe_rx_scan_hw_ring(rxq);
1156 
1157 	/* update internal queue state */
1158 	rxq->rx_next_avail = 0;
1159 	rxq->rx_nb_avail = nb_rx;
1160 	rxq->rx_tail = (uint16_t)(rxq->rx_tail + nb_rx);
1161 
1162 	/* if required, allocate new buffers to replenish descriptors */
1163 	if (rxq->rx_tail > rxq->rx_free_trigger) {
1164 		uint16_t cur_free_trigger = rxq->rx_free_trigger;
1165 
1166 		if (ngbe_rx_alloc_bufs(rxq, true) != 0) {
1167 			int i, j;
1168 
1169 			PMD_RX_LOG(DEBUG, "RX mbuf alloc failed port_id=%u "
1170 				   "queue_id=%u", (uint16_t)rxq->port_id,
1171 				   (uint16_t)rxq->queue_id);
1172 
1173 			dev->data->rx_mbuf_alloc_failed +=
1174 				rxq->rx_free_thresh;
1175 
1176 			/*
1177 			 * Need to rewind any previous receives if we cannot
1178 			 * allocate new buffers to replenish the old ones.
1179 			 */
1180 			rxq->rx_nb_avail = 0;
1181 			rxq->rx_tail = (uint16_t)(rxq->rx_tail - nb_rx);
1182 			for (i = 0, j = rxq->rx_tail; i < nb_rx; ++i, ++j)
1183 				rxq->sw_ring[j].mbuf = rxq->rx_stage[i];
1184 
1185 			return 0;
1186 		}
1187 
1188 		/* update tail pointer */
1189 		rte_wmb();
1190 		ngbe_set32_relaxed(rxq->rdt_reg_addr, cur_free_trigger);
1191 	}
1192 
1193 	if (rxq->rx_tail >= rxq->nb_rx_desc)
1194 		rxq->rx_tail = 0;
1195 
1196 	/* received any packets this loop? */
1197 	if (rxq->rx_nb_avail)
1198 		return ngbe_rx_fill_from_stage(rxq, rx_pkts, nb_pkts);
1199 
1200 	return 0;
1201 }
1202 
1203 /* split requests into chunks of size RTE_PMD_NGBE_RX_MAX_BURST */
1204 uint16_t
1205 ngbe_recv_pkts_bulk_alloc(void *rx_queue, struct rte_mbuf **rx_pkts,
1206 			   uint16_t nb_pkts)
1207 {
1208 	uint16_t nb_rx;
1209 
1210 	if (unlikely(nb_pkts == 0))
1211 		return 0;
1212 
1213 	if (likely(nb_pkts <= RTE_PMD_NGBE_RX_MAX_BURST))
1214 		return ngbe_rx_recv_pkts(rx_queue, rx_pkts, nb_pkts);
1215 
1216 	/* request is relatively large, chunk it up */
1217 	nb_rx = 0;
1218 	while (nb_pkts) {
1219 		uint16_t ret, n;
1220 
1221 		n = (uint16_t)RTE_MIN(nb_pkts, RTE_PMD_NGBE_RX_MAX_BURST);
1222 		ret = ngbe_rx_recv_pkts(rx_queue, &rx_pkts[nb_rx], n);
1223 		nb_rx = (uint16_t)(nb_rx + ret);
1224 		nb_pkts = (uint16_t)(nb_pkts - ret);
1225 		if (ret < n)
1226 			break;
1227 	}
1228 
1229 	return nb_rx;
1230 }
1231 
1232 uint16_t
1233 ngbe_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts,
1234 		uint16_t nb_pkts)
1235 {
1236 	struct ngbe_rx_queue *rxq;
1237 	volatile struct ngbe_rx_desc *rx_ring;
1238 	volatile struct ngbe_rx_desc *rxdp;
1239 	struct ngbe_rx_entry *sw_ring;
1240 	struct ngbe_rx_entry *rxe;
1241 	struct rte_mbuf *rxm;
1242 	struct rte_mbuf *nmb;
1243 	struct ngbe_rx_desc rxd;
1244 	uint64_t dma_addr;
1245 	uint32_t staterr;
1246 	uint32_t pkt_info;
1247 	uint16_t pkt_len;
1248 	uint16_t rx_id;
1249 	uint16_t nb_rx;
1250 	uint16_t nb_hold;
1251 	uint64_t pkt_flags;
1252 
1253 	nb_rx = 0;
1254 	nb_hold = 0;
1255 	rxq = rx_queue;
1256 	rx_id = rxq->rx_tail;
1257 	rx_ring = rxq->rx_ring;
1258 	sw_ring = rxq->sw_ring;
1259 	struct rte_eth_dev *dev = &rte_eth_devices[rxq->port_id];
1260 	while (nb_rx < nb_pkts) {
1261 		/*
1262 		 * The order of operations here is important as the DD status
1263 		 * bit must not be read after any other descriptor fields.
1264 		 * rx_ring and rxdp are pointing to volatile data so the order
1265 		 * of accesses cannot be reordered by the compiler. If they were
1266 		 * not volatile, they could be reordered which could lead to
1267 		 * using invalid descriptor fields when read from rxd.
1268 		 */
1269 		rxdp = &rx_ring[rx_id];
1270 		staterr = rxdp->qw1.lo.status;
1271 		if (!(staterr & rte_cpu_to_le_32(NGBE_RXD_STAT_DD)))
1272 			break;
1273 		rxd = *rxdp;
1274 
1275 		/*
1276 		 * End of packet.
1277 		 *
1278 		 * If the NGBE_RXD_STAT_EOP flag is not set, the Rx packet
1279 		 * is likely to be invalid and to be dropped by the various
1280 		 * validation checks performed by the network stack.
1281 		 *
1282 		 * Allocate a new mbuf to replenish the RX ring descriptor.
1283 		 * If the allocation fails:
1284 		 *    - arrange for that Rx descriptor to be the first one
1285 		 *      being parsed the next time the receive function is
1286 		 *      invoked [on the same queue].
1287 		 *
1288 		 *    - Stop parsing the Rx ring and return immediately.
1289 		 *
1290 		 * This policy do not drop the packet received in the Rx
1291 		 * descriptor for which the allocation of a new mbuf failed.
1292 		 * Thus, it allows that packet to be later retrieved if
1293 		 * mbuf have been freed in the mean time.
1294 		 * As a side effect, holding Rx descriptors instead of
1295 		 * systematically giving them back to the NIC may lead to
1296 		 * Rx ring exhaustion situations.
1297 		 * However, the NIC can gracefully prevent such situations
1298 		 * to happen by sending specific "back-pressure" flow control
1299 		 * frames to its peer(s).
1300 		 */
1301 		PMD_RX_LOG(DEBUG,
1302 			   "port_id=%u queue_id=%u rx_id=%u ext_err_stat=0x%08x pkt_len=%u",
1303 			   (uint16_t)rxq->port_id, (uint16_t)rxq->queue_id,
1304 			   (uint16_t)rx_id, (uint32_t)staterr,
1305 			   (uint16_t)rte_le_to_cpu_16(rxd.qw1.hi.len));
1306 
1307 		nmb = rte_mbuf_raw_alloc(rxq->mb_pool);
1308 		if (nmb == NULL) {
1309 			PMD_RX_LOG(DEBUG,
1310 				   "Rx mbuf alloc failed port_id=%u queue_id=%u",
1311 				   (uint16_t)rxq->port_id,
1312 				   (uint16_t)rxq->queue_id);
1313 			dev->data->rx_mbuf_alloc_failed++;
1314 			break;
1315 		}
1316 
1317 		nb_hold++;
1318 		rxe = &sw_ring[rx_id];
1319 		rx_id++;
1320 		if (rx_id == rxq->nb_rx_desc)
1321 			rx_id = 0;
1322 
1323 		/* Prefetch next mbuf while processing current one. */
1324 		rte_ngbe_prefetch(sw_ring[rx_id].mbuf);
1325 
1326 		/*
1327 		 * When next Rx descriptor is on a cache-line boundary,
1328 		 * prefetch the next 4 Rx descriptors and the next 8 pointers
1329 		 * to mbufs.
1330 		 */
1331 		if ((rx_id & 0x3) == 0) {
1332 			rte_ngbe_prefetch(&rx_ring[rx_id]);
1333 			rte_ngbe_prefetch(&sw_ring[rx_id]);
1334 		}
1335 
1336 		rxm = rxe->mbuf;
1337 		rxe->mbuf = nmb;
1338 		dma_addr = rte_cpu_to_le_64(rte_mbuf_data_iova_default(nmb));
1339 		NGBE_RXD_HDRADDR(rxdp, 0);
1340 		NGBE_RXD_PKTADDR(rxdp, dma_addr);
1341 
1342 		/*
1343 		 * Initialize the returned mbuf.
1344 		 * 1) setup generic mbuf fields:
1345 		 *    - number of segments,
1346 		 *    - next segment,
1347 		 *    - packet length,
1348 		 *    - Rx port identifier.
1349 		 * 2) integrate hardware offload data, if any:
1350 		 *    - RSS flag & hash,
1351 		 *    - IP checksum flag,
1352 		 *    - VLAN TCI, if any,
1353 		 *    - error flags.
1354 		 */
1355 		pkt_len = (uint16_t)(rte_le_to_cpu_16(rxd.qw1.hi.len) -
1356 				      rxq->crc_len);
1357 		rxm->data_off = RTE_PKTMBUF_HEADROOM;
1358 		rte_packet_prefetch((char *)rxm->buf_addr + rxm->data_off);
1359 		rxm->nb_segs = 1;
1360 		rxm->next = NULL;
1361 		rxm->pkt_len = pkt_len;
1362 		rxm->data_len = pkt_len;
1363 		rxm->port = rxq->port_id;
1364 
1365 		pkt_info = rte_le_to_cpu_32(rxd.qw0.dw0);
1366 		/* Only valid if RTE_MBUF_F_RX_VLAN set in pkt_flags */
1367 		rxm->vlan_tci = rte_le_to_cpu_16(rxd.qw1.hi.tag);
1368 
1369 		pkt_flags = rx_desc_status_to_pkt_flags(staterr,
1370 					rxq->vlan_flags);
1371 		pkt_flags |= rx_desc_error_to_pkt_flags(staterr);
1372 		pkt_flags |= ngbe_rxd_pkt_info_to_pkt_flags(pkt_info);
1373 		rxm->ol_flags = pkt_flags;
1374 		rxm->packet_type = ngbe_rxd_pkt_info_to_pkt_type(pkt_info,
1375 						       NGBE_PTID_MASK);
1376 
1377 		if (likely(pkt_flags & RTE_MBUF_F_RX_RSS_HASH))
1378 			rxm->hash.rss = rte_le_to_cpu_32(rxd.qw0.dw1);
1379 
1380 		/*
1381 		 * Store the mbuf address into the next entry of the array
1382 		 * of returned packets.
1383 		 */
1384 		rx_pkts[nb_rx++] = rxm;
1385 	}
1386 	rxq->rx_tail = rx_id;
1387 
1388 	/*
1389 	 * If the number of free Rx descriptors is greater than the Rx free
1390 	 * threshold of the queue, advance the Receive Descriptor Tail (RDT)
1391 	 * register.
1392 	 * Update the RDT with the value of the last processed Rx descriptor
1393 	 * minus 1, to guarantee that the RDT register is never equal to the
1394 	 * RDH register, which creates a "full" ring situation from the
1395 	 * hardware point of view...
1396 	 */
1397 	nb_hold = (uint16_t)(nb_hold + rxq->nb_rx_hold);
1398 	if (nb_hold > rxq->rx_free_thresh) {
1399 		PMD_RX_LOG(DEBUG,
1400 			   "port_id=%u queue_id=%u rx_tail=%u nb_hold=%u nb_rx=%u",
1401 			   (uint16_t)rxq->port_id, (uint16_t)rxq->queue_id,
1402 			   (uint16_t)rx_id, (uint16_t)nb_hold,
1403 			   (uint16_t)nb_rx);
1404 		rx_id = (uint16_t)((rx_id == 0) ?
1405 				(rxq->nb_rx_desc - 1) : (rx_id - 1));
1406 		ngbe_set32(rxq->rdt_reg_addr, rx_id);
1407 		nb_hold = 0;
1408 	}
1409 	rxq->nb_rx_hold = nb_hold;
1410 	return nb_rx;
1411 }
1412 
1413 /**
1414  * ngbe_fill_cluster_head_buf - fill the first mbuf of the returned packet
1415  *
1416  * Fill the following info in the HEAD buffer of the Rx cluster:
1417  *    - RX port identifier
1418  *    - hardware offload data, if any:
1419  *      - RSS flag & hash
1420  *      - IP checksum flag
1421  *      - VLAN TCI, if any
1422  *      - error flags
1423  * @head HEAD of the packet cluster
1424  * @desc HW descriptor to get data from
1425  * @rxq Pointer to the Rx queue
1426  */
1427 static inline void
1428 ngbe_fill_cluster_head_buf(struct rte_mbuf *head, struct ngbe_rx_desc *desc,
1429 		struct ngbe_rx_queue *rxq, uint32_t staterr)
1430 {
1431 	uint32_t pkt_info;
1432 	uint64_t pkt_flags;
1433 
1434 	head->port = rxq->port_id;
1435 
1436 	/* The vlan_tci field is only valid when RTE_MBUF_F_RX_VLAN is
1437 	 * set in the pkt_flags field.
1438 	 */
1439 	head->vlan_tci = rte_le_to_cpu_16(desc->qw1.hi.tag);
1440 	pkt_info = rte_le_to_cpu_32(desc->qw0.dw0);
1441 	pkt_flags = rx_desc_status_to_pkt_flags(staterr, rxq->vlan_flags);
1442 	pkt_flags |= rx_desc_error_to_pkt_flags(staterr);
1443 	pkt_flags |= ngbe_rxd_pkt_info_to_pkt_flags(pkt_info);
1444 	head->ol_flags = pkt_flags;
1445 	head->packet_type = ngbe_rxd_pkt_info_to_pkt_type(pkt_info,
1446 						NGBE_PTID_MASK);
1447 
1448 	if (likely(pkt_flags & RTE_MBUF_F_RX_RSS_HASH))
1449 		head->hash.rss = rte_le_to_cpu_32(desc->qw0.dw1);
1450 }
1451 
1452 /**
1453  * ngbe_recv_pkts_sc - receive handler for scatter case.
1454  *
1455  * @rx_queue Rx queue handle
1456  * @rx_pkts table of received packets
1457  * @nb_pkts size of rx_pkts table
1458  * @bulk_alloc if TRUE bulk allocation is used for a HW ring refilling
1459  *
1460  * Returns the number of received packets/clusters (according to the "bulk
1461  * receive" interface).
1462  */
1463 static inline uint16_t
1464 ngbe_recv_pkts_sc(void *rx_queue, struct rte_mbuf **rx_pkts, uint16_t nb_pkts,
1465 		    bool bulk_alloc)
1466 {
1467 	struct ngbe_rx_queue *rxq = rx_queue;
1468 	struct rte_eth_dev *dev = &rte_eth_devices[rxq->port_id];
1469 	volatile struct ngbe_rx_desc *rx_ring = rxq->rx_ring;
1470 	struct ngbe_rx_entry *sw_ring = rxq->sw_ring;
1471 	struct ngbe_scattered_rx_entry *sw_sc_ring = rxq->sw_sc_ring;
1472 	uint16_t rx_id = rxq->rx_tail;
1473 	uint16_t nb_rx = 0;
1474 	uint16_t nb_hold = rxq->nb_rx_hold;
1475 	uint16_t prev_id = rxq->rx_tail;
1476 
1477 	while (nb_rx < nb_pkts) {
1478 		bool eop;
1479 		struct ngbe_rx_entry *rxe;
1480 		struct ngbe_scattered_rx_entry *sc_entry;
1481 		struct ngbe_scattered_rx_entry *next_sc_entry = NULL;
1482 		struct ngbe_rx_entry *next_rxe = NULL;
1483 		struct rte_mbuf *first_seg;
1484 		struct rte_mbuf *rxm;
1485 		struct rte_mbuf *nmb = NULL;
1486 		struct ngbe_rx_desc rxd;
1487 		uint16_t data_len;
1488 		uint16_t next_id;
1489 		volatile struct ngbe_rx_desc *rxdp;
1490 		uint32_t staterr;
1491 
1492 next_desc:
1493 		rxdp = &rx_ring[rx_id];
1494 		staterr = rte_le_to_cpu_32(rxdp->qw1.lo.status);
1495 
1496 		if (!(staterr & NGBE_RXD_STAT_DD))
1497 			break;
1498 
1499 		rxd = *rxdp;
1500 
1501 		PMD_RX_LOG(DEBUG, "port_id=%u queue_id=%u rx_id=%u "
1502 				  "staterr=0x%x data_len=%u",
1503 			   rxq->port_id, rxq->queue_id, rx_id, staterr,
1504 			   rte_le_to_cpu_16(rxd.qw1.hi.len));
1505 
1506 		if (!bulk_alloc) {
1507 			nmb = rte_mbuf_raw_alloc(rxq->mb_pool);
1508 			if (nmb == NULL) {
1509 				PMD_RX_LOG(DEBUG, "Rx mbuf alloc failed "
1510 						  "port_id=%u queue_id=%u",
1511 					   rxq->port_id, rxq->queue_id);
1512 
1513 				dev->data->rx_mbuf_alloc_failed++;
1514 				break;
1515 			}
1516 		} else if (nb_hold > rxq->rx_free_thresh) {
1517 			uint16_t next_rdt = rxq->rx_free_trigger;
1518 
1519 			if (!ngbe_rx_alloc_bufs(rxq, false)) {
1520 				rte_wmb();
1521 				ngbe_set32_relaxed(rxq->rdt_reg_addr,
1522 							    next_rdt);
1523 				nb_hold -= rxq->rx_free_thresh;
1524 			} else {
1525 				PMD_RX_LOG(DEBUG, "Rx bulk alloc failed "
1526 						  "port_id=%u queue_id=%u",
1527 					   rxq->port_id, rxq->queue_id);
1528 
1529 				dev->data->rx_mbuf_alloc_failed++;
1530 				break;
1531 			}
1532 		}
1533 
1534 		nb_hold++;
1535 		rxe = &sw_ring[rx_id];
1536 		eop = staterr & NGBE_RXD_STAT_EOP;
1537 
1538 		next_id = rx_id + 1;
1539 		if (next_id == rxq->nb_rx_desc)
1540 			next_id = 0;
1541 
1542 		/* Prefetch next mbuf while processing current one. */
1543 		rte_ngbe_prefetch(sw_ring[next_id].mbuf);
1544 
1545 		/*
1546 		 * When next Rx descriptor is on a cache-line boundary,
1547 		 * prefetch the next 4 RX descriptors and the next 4 pointers
1548 		 * to mbufs.
1549 		 */
1550 		if ((next_id & 0x3) == 0) {
1551 			rte_ngbe_prefetch(&rx_ring[next_id]);
1552 			rte_ngbe_prefetch(&sw_ring[next_id]);
1553 		}
1554 
1555 		rxm = rxe->mbuf;
1556 
1557 		if (!bulk_alloc) {
1558 			__le64 dma =
1559 			  rte_cpu_to_le_64(rte_mbuf_data_iova_default(nmb));
1560 			/*
1561 			 * Update Rx descriptor with the physical address of the
1562 			 * new data buffer of the new allocated mbuf.
1563 			 */
1564 			rxe->mbuf = nmb;
1565 
1566 			rxm->data_off = RTE_PKTMBUF_HEADROOM;
1567 			NGBE_RXD_HDRADDR(rxdp, 0);
1568 			NGBE_RXD_PKTADDR(rxdp, dma);
1569 		} else {
1570 			rxe->mbuf = NULL;
1571 		}
1572 
1573 		/*
1574 		 * Set data length & data buffer address of mbuf.
1575 		 */
1576 		data_len = rte_le_to_cpu_16(rxd.qw1.hi.len);
1577 		rxm->data_len = data_len;
1578 
1579 		if (!eop) {
1580 			uint16_t nextp_id;
1581 
1582 			nextp_id = next_id;
1583 			next_sc_entry = &sw_sc_ring[nextp_id];
1584 			next_rxe = &sw_ring[nextp_id];
1585 			rte_ngbe_prefetch(next_rxe);
1586 		}
1587 
1588 		sc_entry = &sw_sc_ring[rx_id];
1589 		first_seg = sc_entry->fbuf;
1590 		sc_entry->fbuf = NULL;
1591 
1592 		/*
1593 		 * If this is the first buffer of the received packet,
1594 		 * set the pointer to the first mbuf of the packet and
1595 		 * initialize its context.
1596 		 * Otherwise, update the total length and the number of segments
1597 		 * of the current scattered packet, and update the pointer to
1598 		 * the last mbuf of the current packet.
1599 		 */
1600 		if (first_seg == NULL) {
1601 			first_seg = rxm;
1602 			first_seg->pkt_len = data_len;
1603 			first_seg->nb_segs = 1;
1604 		} else {
1605 			first_seg->pkt_len += data_len;
1606 			first_seg->nb_segs++;
1607 		}
1608 
1609 		prev_id = rx_id;
1610 		rx_id = next_id;
1611 
1612 		/*
1613 		 * If this is not the last buffer of the received packet, update
1614 		 * the pointer to the first mbuf at the NEXTP entry in the
1615 		 * sw_sc_ring and continue to parse the Rx ring.
1616 		 */
1617 		if (!eop && next_rxe) {
1618 			rxm->next = next_rxe->mbuf;
1619 			next_sc_entry->fbuf = first_seg;
1620 			goto next_desc;
1621 		}
1622 
1623 		/* Initialize the first mbuf of the returned packet */
1624 		ngbe_fill_cluster_head_buf(first_seg, &rxd, rxq, staterr);
1625 
1626 		/* Deal with the case, when HW CRC srip is disabled. */
1627 		first_seg->pkt_len -= rxq->crc_len;
1628 		if (unlikely(rxm->data_len <= rxq->crc_len)) {
1629 			struct rte_mbuf *lp;
1630 
1631 			for (lp = first_seg; lp->next != rxm; lp = lp->next)
1632 				;
1633 
1634 			first_seg->nb_segs--;
1635 			lp->data_len -= rxq->crc_len - rxm->data_len;
1636 			lp->next = NULL;
1637 			rte_pktmbuf_free_seg(rxm);
1638 		} else {
1639 			rxm->data_len -= rxq->crc_len;
1640 		}
1641 
1642 		/* Prefetch data of first segment, if configured to do so. */
1643 		rte_packet_prefetch((char *)first_seg->buf_addr +
1644 			first_seg->data_off);
1645 
1646 		/*
1647 		 * Store the mbuf address into the next entry of the array
1648 		 * of returned packets.
1649 		 */
1650 		rx_pkts[nb_rx++] = first_seg;
1651 	}
1652 
1653 	/*
1654 	 * Record index of the next Rx descriptor to probe.
1655 	 */
1656 	rxq->rx_tail = rx_id;
1657 
1658 	/*
1659 	 * If the number of free Rx descriptors is greater than the Rx free
1660 	 * threshold of the queue, advance the Receive Descriptor Tail (RDT)
1661 	 * register.
1662 	 * Update the RDT with the value of the last processed Rx descriptor
1663 	 * minus 1, to guarantee that the RDT register is never equal to the
1664 	 * RDH register, which creates a "full" ring situation from the
1665 	 * hardware point of view...
1666 	 */
1667 	if (!bulk_alloc && nb_hold > rxq->rx_free_thresh) {
1668 		PMD_RX_LOG(DEBUG, "port_id=%u queue_id=%u rx_tail=%u "
1669 			   "nb_hold=%u nb_rx=%u",
1670 			   rxq->port_id, rxq->queue_id, rx_id, nb_hold, nb_rx);
1671 
1672 		rte_wmb();
1673 		ngbe_set32_relaxed(rxq->rdt_reg_addr, prev_id);
1674 		nb_hold = 0;
1675 	}
1676 
1677 	rxq->nb_rx_hold = nb_hold;
1678 	return nb_rx;
1679 }
1680 
1681 uint16_t
1682 ngbe_recv_pkts_sc_single_alloc(void *rx_queue, struct rte_mbuf **rx_pkts,
1683 				 uint16_t nb_pkts)
1684 {
1685 	return ngbe_recv_pkts_sc(rx_queue, rx_pkts, nb_pkts, false);
1686 }
1687 
1688 uint16_t
1689 ngbe_recv_pkts_sc_bulk_alloc(void *rx_queue, struct rte_mbuf **rx_pkts,
1690 			       uint16_t nb_pkts)
1691 {
1692 	return ngbe_recv_pkts_sc(rx_queue, rx_pkts, nb_pkts, true);
1693 }
1694 
1695 /*********************************************************************
1696  *
1697  *  Queue management functions
1698  *
1699  **********************************************************************/
1700 
1701 static void
1702 ngbe_tx_queue_release_mbufs(struct ngbe_tx_queue *txq)
1703 {
1704 	unsigned int i;
1705 
1706 	if (txq->sw_ring != NULL) {
1707 		for (i = 0; i < txq->nb_tx_desc; i++) {
1708 			if (txq->sw_ring[i].mbuf != NULL) {
1709 				rte_pktmbuf_free_seg(txq->sw_ring[i].mbuf);
1710 				txq->sw_ring[i].mbuf = NULL;
1711 			}
1712 		}
1713 	}
1714 }
1715 
1716 static void
1717 ngbe_tx_free_swring(struct ngbe_tx_queue *txq)
1718 {
1719 	if (txq != NULL)
1720 		rte_free(txq->sw_ring);
1721 }
1722 
1723 static void
1724 ngbe_tx_queue_release(struct ngbe_tx_queue *txq)
1725 {
1726 	if (txq != NULL) {
1727 		if (txq->ops != NULL) {
1728 			txq->ops->release_mbufs(txq);
1729 			txq->ops->free_swring(txq);
1730 		}
1731 		rte_free(txq);
1732 	}
1733 }
1734 
1735 void
1736 ngbe_dev_tx_queue_release(struct rte_eth_dev *dev, uint16_t qid)
1737 {
1738 	ngbe_tx_queue_release(dev->data->tx_queues[qid]);
1739 }
1740 
1741 /* (Re)set dynamic ngbe_tx_queue fields to defaults */
1742 static void
1743 ngbe_reset_tx_queue(struct ngbe_tx_queue *txq)
1744 {
1745 	static const struct ngbe_tx_desc zeroed_desc = {0};
1746 	struct ngbe_tx_entry *txe = txq->sw_ring;
1747 	uint16_t prev, i;
1748 
1749 	/* Zero out HW ring memory */
1750 	for (i = 0; i < txq->nb_tx_desc; i++)
1751 		txq->tx_ring[i] = zeroed_desc;
1752 
1753 	/* Initialize SW ring entries */
1754 	prev = (uint16_t)(txq->nb_tx_desc - 1);
1755 	for (i = 0; i < txq->nb_tx_desc; i++) {
1756 		/* the ring can also be modified by hardware */
1757 		volatile struct ngbe_tx_desc *txd = &txq->tx_ring[i];
1758 
1759 		txd->dw3 = rte_cpu_to_le_32(NGBE_TXD_DD);
1760 		txe[i].mbuf = NULL;
1761 		txe[i].last_id = i;
1762 		txe[prev].next_id = i;
1763 		prev = i;
1764 	}
1765 
1766 	txq->tx_next_dd = (uint16_t)(txq->tx_free_thresh - 1);
1767 	txq->tx_tail = 0;
1768 
1769 	/*
1770 	 * Always allow 1 descriptor to be un-allocated to avoid
1771 	 * a H/W race condition
1772 	 */
1773 	txq->last_desc_cleaned = (uint16_t)(txq->nb_tx_desc - 1);
1774 	txq->nb_tx_free = (uint16_t)(txq->nb_tx_desc - 1);
1775 	txq->ctx_curr = 0;
1776 	memset((void *)&txq->ctx_cache, 0,
1777 		NGBE_CTX_NUM * sizeof(struct ngbe_ctx_info));
1778 }
1779 
1780 static const struct ngbe_txq_ops def_txq_ops = {
1781 	.release_mbufs = ngbe_tx_queue_release_mbufs,
1782 	.free_swring = ngbe_tx_free_swring,
1783 	.reset = ngbe_reset_tx_queue,
1784 };
1785 
1786 /* Takes an ethdev and a queue and sets up the tx function to be used based on
1787  * the queue parameters. Used in tx_queue_setup by primary process and then
1788  * in dev_init by secondary process when attaching to an existing ethdev.
1789  */
1790 void
1791 ngbe_set_tx_function(struct rte_eth_dev *dev, struct ngbe_tx_queue *txq)
1792 {
1793 	/* Use a simple Tx queue (no offloads, no multi segs) if possible */
1794 	if (txq->offloads == 0 &&
1795 			txq->tx_free_thresh >= RTE_PMD_NGBE_TX_MAX_BURST) {
1796 		PMD_INIT_LOG(DEBUG, "Using simple tx code path");
1797 		dev->tx_pkt_burst = ngbe_xmit_pkts_simple;
1798 		dev->tx_pkt_prepare = NULL;
1799 	} else {
1800 		PMD_INIT_LOG(DEBUG, "Using full-featured tx code path");
1801 		PMD_INIT_LOG(DEBUG,
1802 				" - offloads = 0x%" PRIx64,
1803 				txq->offloads);
1804 		PMD_INIT_LOG(DEBUG,
1805 				" - tx_free_thresh = %lu [RTE_PMD_NGBE_TX_MAX_BURST=%lu]",
1806 				(unsigned long)txq->tx_free_thresh,
1807 				(unsigned long)RTE_PMD_NGBE_TX_MAX_BURST);
1808 		dev->tx_pkt_burst = ngbe_xmit_pkts;
1809 		dev->tx_pkt_prepare = ngbe_prep_pkts;
1810 	}
1811 }
1812 
1813 static const struct {
1814 	eth_tx_burst_t pkt_burst;
1815 	const char *info;
1816 } ngbe_tx_burst_infos[] = {
1817 	{ ngbe_xmit_pkts_simple,   "Scalar Simple"},
1818 	{ ngbe_xmit_pkts,          "Scalar"},
1819 };
1820 
1821 int
1822 ngbe_tx_burst_mode_get(struct rte_eth_dev *dev, __rte_unused uint16_t queue_id,
1823 		      struct rte_eth_burst_mode *mode)
1824 {
1825 	eth_tx_burst_t pkt_burst = dev->tx_pkt_burst;
1826 	int ret = -EINVAL;
1827 	unsigned int i;
1828 
1829 	for (i = 0; i < RTE_DIM(ngbe_tx_burst_infos); ++i) {
1830 		if (pkt_burst == ngbe_tx_burst_infos[i].pkt_burst) {
1831 			snprintf(mode->info, sizeof(mode->info), "%s",
1832 				 ngbe_tx_burst_infos[i].info);
1833 			ret = 0;
1834 			break;
1835 		}
1836 	}
1837 
1838 	return ret;
1839 }
1840 
1841 uint64_t
1842 ngbe_get_tx_port_offloads(struct rte_eth_dev *dev)
1843 {
1844 	uint64_t tx_offload_capa;
1845 	struct ngbe_hw *hw = ngbe_dev_hw(dev);
1846 
1847 	tx_offload_capa =
1848 		RTE_ETH_TX_OFFLOAD_VLAN_INSERT |
1849 		RTE_ETH_TX_OFFLOAD_IPV4_CKSUM  |
1850 		RTE_ETH_TX_OFFLOAD_UDP_CKSUM   |
1851 		RTE_ETH_TX_OFFLOAD_TCP_CKSUM   |
1852 		RTE_ETH_TX_OFFLOAD_SCTP_CKSUM  |
1853 		RTE_ETH_TX_OFFLOAD_OUTER_IPV4_CKSUM |
1854 		RTE_ETH_TX_OFFLOAD_TCP_TSO     |
1855 		RTE_ETH_TX_OFFLOAD_UDP_TSO	   |
1856 		RTE_ETH_TX_OFFLOAD_UDP_TNL_TSO	|
1857 		RTE_ETH_TX_OFFLOAD_IP_TNL_TSO	|
1858 		RTE_ETH_TX_OFFLOAD_IPIP_TNL_TSO	|
1859 		RTE_ETH_TX_OFFLOAD_MULTI_SEGS;
1860 
1861 	if (hw->is_pf)
1862 		tx_offload_capa |= RTE_ETH_TX_OFFLOAD_QINQ_INSERT;
1863 
1864 	return tx_offload_capa;
1865 }
1866 
1867 int
1868 ngbe_dev_tx_queue_setup(struct rte_eth_dev *dev,
1869 			 uint16_t queue_idx,
1870 			 uint16_t nb_desc,
1871 			 unsigned int socket_id,
1872 			 const struct rte_eth_txconf *tx_conf)
1873 {
1874 	const struct rte_memzone *tz;
1875 	struct ngbe_tx_queue *txq;
1876 	struct ngbe_hw     *hw;
1877 	uint16_t tx_free_thresh;
1878 	uint64_t offloads;
1879 
1880 	PMD_INIT_FUNC_TRACE();
1881 	hw = ngbe_dev_hw(dev);
1882 
1883 	offloads = tx_conf->offloads | dev->data->dev_conf.txmode.offloads;
1884 
1885 	/*
1886 	 * The Tx descriptor ring will be cleaned after txq->tx_free_thresh
1887 	 * descriptors are used or if the number of descriptors required
1888 	 * to transmit a packet is greater than the number of free Tx
1889 	 * descriptors.
1890 	 * One descriptor in the Tx ring is used as a sentinel to avoid a
1891 	 * H/W race condition, hence the maximum threshold constraints.
1892 	 * When set to zero use default values.
1893 	 */
1894 	tx_free_thresh = (uint16_t)((tx_conf->tx_free_thresh) ?
1895 			tx_conf->tx_free_thresh : DEFAULT_TX_FREE_THRESH);
1896 	if (tx_free_thresh >= (nb_desc - 3)) {
1897 		PMD_INIT_LOG(ERR,
1898 			     "tx_free_thresh must be less than the number of TX descriptors minus 3. (tx_free_thresh=%u port=%d queue=%d)",
1899 			     (unsigned int)tx_free_thresh,
1900 			     (int)dev->data->port_id, (int)queue_idx);
1901 		return -(EINVAL);
1902 	}
1903 
1904 	if (nb_desc % tx_free_thresh != 0) {
1905 		PMD_INIT_LOG(ERR,
1906 			     "tx_free_thresh must be a divisor of the number of Tx descriptors. (tx_free_thresh=%u port=%d queue=%d)",
1907 			     (unsigned int)tx_free_thresh,
1908 			     (int)dev->data->port_id, (int)queue_idx);
1909 		return -(EINVAL);
1910 	}
1911 
1912 	/* Free memory prior to re-allocation if needed... */
1913 	if (dev->data->tx_queues[queue_idx] != NULL) {
1914 		ngbe_tx_queue_release(dev->data->tx_queues[queue_idx]);
1915 		dev->data->tx_queues[queue_idx] = NULL;
1916 	}
1917 
1918 	/* First allocate the Tx queue data structure */
1919 	txq = rte_zmalloc_socket("ethdev Tx queue",
1920 				 sizeof(struct ngbe_tx_queue),
1921 				 RTE_CACHE_LINE_SIZE, socket_id);
1922 	if (txq == NULL)
1923 		return -ENOMEM;
1924 
1925 	/*
1926 	 * Allocate Tx ring hardware descriptors. A memzone large enough to
1927 	 * handle the maximum ring size is allocated in order to allow for
1928 	 * resizing in later calls to the queue setup function.
1929 	 */
1930 	tz = rte_eth_dma_zone_reserve(dev, "tx_ring", queue_idx,
1931 			sizeof(struct ngbe_tx_desc) * NGBE_RING_DESC_MAX,
1932 			NGBE_ALIGN, socket_id);
1933 	if (tz == NULL) {
1934 		ngbe_tx_queue_release(txq);
1935 		return -ENOMEM;
1936 	}
1937 
1938 	txq->nb_tx_desc = nb_desc;
1939 	txq->tx_free_thresh = tx_free_thresh;
1940 	txq->pthresh = tx_conf->tx_thresh.pthresh;
1941 	txq->hthresh = tx_conf->tx_thresh.hthresh;
1942 	txq->wthresh = tx_conf->tx_thresh.wthresh;
1943 	txq->queue_id = queue_idx;
1944 	txq->reg_idx = (uint16_t)((RTE_ETH_DEV_SRIOV(dev).active == 0) ?
1945 		queue_idx : RTE_ETH_DEV_SRIOV(dev).def_pool_q_idx + queue_idx);
1946 	txq->port_id = dev->data->port_id;
1947 	txq->offloads = offloads;
1948 	txq->ops = &def_txq_ops;
1949 	txq->tx_deferred_start = tx_conf->tx_deferred_start;
1950 
1951 	txq->tdt_reg_addr = NGBE_REG_ADDR(hw, NGBE_TXWP(txq->reg_idx));
1952 	txq->tdc_reg_addr = NGBE_REG_ADDR(hw, NGBE_TXCFG(txq->reg_idx));
1953 
1954 	txq->tx_ring_phys_addr = TMZ_PADDR(tz);
1955 	txq->tx_ring = (struct ngbe_tx_desc *)TMZ_VADDR(tz);
1956 
1957 	/* Allocate software ring */
1958 	txq->sw_ring = rte_zmalloc_socket("txq->sw_ring",
1959 				sizeof(struct ngbe_tx_entry) * nb_desc,
1960 				RTE_CACHE_LINE_SIZE, socket_id);
1961 	if (txq->sw_ring == NULL) {
1962 		ngbe_tx_queue_release(txq);
1963 		return -ENOMEM;
1964 	}
1965 	PMD_INIT_LOG(DEBUG,
1966 		     "sw_ring=%p hw_ring=%p dma_addr=0x%" PRIx64,
1967 		     txq->sw_ring, txq->tx_ring, txq->tx_ring_phys_addr);
1968 
1969 	/* set up scalar Tx function as appropriate */
1970 	ngbe_set_tx_function(dev, txq);
1971 
1972 	txq->ops->reset(txq);
1973 
1974 	dev->data->tx_queues[queue_idx] = txq;
1975 
1976 	return 0;
1977 }
1978 
1979 /**
1980  * ngbe_free_sc_cluster - free the not-yet-completed scattered cluster
1981  *
1982  * The "next" pointer of the last segment of (not-yet-completed) RSC clusters
1983  * in the sw_sc_ring is not set to NULL but rather points to the next
1984  * mbuf of this RSC aggregation (that has not been completed yet and still
1985  * resides on the HW ring). So, instead of calling for rte_pktmbuf_free() we
1986  * will just free first "nb_segs" segments of the cluster explicitly by calling
1987  * an rte_pktmbuf_free_seg().
1988  *
1989  * @m scattered cluster head
1990  */
1991 static void
1992 ngbe_free_sc_cluster(struct rte_mbuf *m)
1993 {
1994 	uint16_t i, nb_segs = m->nb_segs;
1995 	struct rte_mbuf *next_seg;
1996 
1997 	for (i = 0; i < nb_segs; i++) {
1998 		next_seg = m->next;
1999 		rte_pktmbuf_free_seg(m);
2000 		m = next_seg;
2001 	}
2002 }
2003 
2004 static void
2005 ngbe_rx_queue_release_mbufs(struct ngbe_rx_queue *rxq)
2006 {
2007 	unsigned int i;
2008 
2009 	if (rxq->sw_ring != NULL) {
2010 		for (i = 0; i < rxq->nb_rx_desc; i++) {
2011 			if (rxq->sw_ring[i].mbuf != NULL) {
2012 				rte_pktmbuf_free_seg(rxq->sw_ring[i].mbuf);
2013 				rxq->sw_ring[i].mbuf = NULL;
2014 			}
2015 		}
2016 		for (i = 0; i < rxq->rx_nb_avail; ++i) {
2017 			struct rte_mbuf *mb;
2018 
2019 			mb = rxq->rx_stage[rxq->rx_next_avail + i];
2020 			rte_pktmbuf_free_seg(mb);
2021 		}
2022 		rxq->rx_nb_avail = 0;
2023 	}
2024 
2025 	if (rxq->sw_sc_ring != NULL)
2026 		for (i = 0; i < rxq->nb_rx_desc; i++)
2027 			if (rxq->sw_sc_ring[i].fbuf != NULL) {
2028 				ngbe_free_sc_cluster(rxq->sw_sc_ring[i].fbuf);
2029 				rxq->sw_sc_ring[i].fbuf = NULL;
2030 			}
2031 }
2032 
2033 static void
2034 ngbe_rx_queue_release(struct ngbe_rx_queue *rxq)
2035 {
2036 	if (rxq != NULL) {
2037 		ngbe_rx_queue_release_mbufs(rxq);
2038 		rte_free(rxq->sw_ring);
2039 		rte_free(rxq->sw_sc_ring);
2040 		rte_free(rxq);
2041 	}
2042 }
2043 
2044 void
2045 ngbe_dev_rx_queue_release(struct rte_eth_dev *dev, uint16_t qid)
2046 {
2047 	ngbe_rx_queue_release(dev->data->rx_queues[qid]);
2048 }
2049 
2050 /*
2051  * Check if Rx Burst Bulk Alloc function can be used.
2052  * Return
2053  *        0: the preconditions are satisfied and the bulk allocation function
2054  *           can be used.
2055  *  -EINVAL: the preconditions are NOT satisfied and the default Rx burst
2056  *           function must be used.
2057  */
2058 static inline int
2059 check_rx_burst_bulk_alloc_preconditions(struct ngbe_rx_queue *rxq)
2060 {
2061 	int ret = 0;
2062 
2063 	/*
2064 	 * Make sure the following pre-conditions are satisfied:
2065 	 *   rxq->rx_free_thresh >= RTE_PMD_NGBE_RX_MAX_BURST
2066 	 *   rxq->rx_free_thresh < rxq->nb_rx_desc
2067 	 *   (rxq->nb_rx_desc % rxq->rx_free_thresh) == 0
2068 	 * Scattered packets are not supported.  This should be checked
2069 	 * outside of this function.
2070 	 */
2071 	if (rxq->rx_free_thresh < RTE_PMD_NGBE_RX_MAX_BURST) {
2072 		PMD_INIT_LOG(DEBUG,
2073 			     "Rx Burst Bulk Alloc Preconditions: rxq->rx_free_thresh=%d, RTE_PMD_NGBE_RX_MAX_BURST=%d",
2074 			     rxq->rx_free_thresh, RTE_PMD_NGBE_RX_MAX_BURST);
2075 		ret = -EINVAL;
2076 	} else if (rxq->rx_free_thresh >= rxq->nb_rx_desc) {
2077 		PMD_INIT_LOG(DEBUG,
2078 			     "Rx Burst Bulk Alloc Preconditions: rxq->rx_free_thresh=%d, rxq->nb_rx_desc=%d",
2079 			     rxq->rx_free_thresh, rxq->nb_rx_desc);
2080 		ret = -EINVAL;
2081 	} else if ((rxq->nb_rx_desc % rxq->rx_free_thresh) != 0) {
2082 		PMD_INIT_LOG(DEBUG,
2083 			     "Rx Burst Bulk Alloc Preconditions: rxq->nb_rx_desc=%d, rxq->rx_free_thresh=%d",
2084 			     rxq->nb_rx_desc, rxq->rx_free_thresh);
2085 		ret = -EINVAL;
2086 	}
2087 
2088 	return ret;
2089 }
2090 
2091 /* Reset dynamic ngbe_rx_queue fields back to defaults */
2092 static void
2093 ngbe_reset_rx_queue(struct ngbe_adapter *adapter, struct ngbe_rx_queue *rxq)
2094 {
2095 	static const struct ngbe_rx_desc zeroed_desc = {
2096 						{{0}, {0} }, {{0}, {0} } };
2097 	unsigned int i;
2098 	uint16_t len = rxq->nb_rx_desc;
2099 
2100 	/*
2101 	 * By default, the Rx queue setup function allocates enough memory for
2102 	 * NGBE_RING_DESC_MAX.  The Rx Burst bulk allocation function requires
2103 	 * extra memory at the end of the descriptor ring to be zero'd out.
2104 	 */
2105 	if (adapter->rx_bulk_alloc_allowed)
2106 		/* zero out extra memory */
2107 		len += RTE_PMD_NGBE_RX_MAX_BURST;
2108 
2109 	/*
2110 	 * Zero out HW ring memory. Zero out extra memory at the end of
2111 	 * the H/W ring so look-ahead logic in Rx Burst bulk alloc function
2112 	 * reads extra memory as zeros.
2113 	 */
2114 	for (i = 0; i < len; i++)
2115 		rxq->rx_ring[i] = zeroed_desc;
2116 
2117 	/*
2118 	 * initialize extra software ring entries. Space for these extra
2119 	 * entries is always allocated
2120 	 */
2121 	memset(&rxq->fake_mbuf, 0x0, sizeof(rxq->fake_mbuf));
2122 	for (i = rxq->nb_rx_desc; i < len; ++i)
2123 		rxq->sw_ring[i].mbuf = &rxq->fake_mbuf;
2124 
2125 	rxq->rx_nb_avail = 0;
2126 	rxq->rx_next_avail = 0;
2127 	rxq->rx_free_trigger = (uint16_t)(rxq->rx_free_thresh - 1);
2128 	rxq->rx_tail = 0;
2129 	rxq->nb_rx_hold = 0;
2130 	rxq->pkt_first_seg = NULL;
2131 	rxq->pkt_last_seg = NULL;
2132 }
2133 
2134 uint64_t
2135 ngbe_get_rx_queue_offloads(struct rte_eth_dev *dev __rte_unused)
2136 {
2137 	return RTE_ETH_RX_OFFLOAD_VLAN_STRIP;
2138 }
2139 
2140 uint64_t
2141 ngbe_get_rx_port_offloads(struct rte_eth_dev *dev)
2142 {
2143 	uint64_t offloads;
2144 	struct ngbe_hw *hw = ngbe_dev_hw(dev);
2145 
2146 	offloads = RTE_ETH_RX_OFFLOAD_IPV4_CKSUM  |
2147 		   RTE_ETH_RX_OFFLOAD_UDP_CKSUM   |
2148 		   RTE_ETH_RX_OFFLOAD_TCP_CKSUM   |
2149 		   RTE_ETH_RX_OFFLOAD_KEEP_CRC    |
2150 		   RTE_ETH_RX_OFFLOAD_VLAN_FILTER |
2151 		   RTE_ETH_RX_OFFLOAD_SCATTER;
2152 
2153 	if (hw->is_pf)
2154 		offloads |= (RTE_ETH_RX_OFFLOAD_QINQ_STRIP |
2155 			     RTE_ETH_RX_OFFLOAD_VLAN_EXTEND);
2156 
2157 	return offloads;
2158 }
2159 
2160 int
2161 ngbe_dev_rx_queue_setup(struct rte_eth_dev *dev,
2162 			 uint16_t queue_idx,
2163 			 uint16_t nb_desc,
2164 			 unsigned int socket_id,
2165 			 const struct rte_eth_rxconf *rx_conf,
2166 			 struct rte_mempool *mp)
2167 {
2168 	const struct rte_memzone *rz;
2169 	struct ngbe_rx_queue *rxq;
2170 	struct ngbe_hw     *hw;
2171 	uint16_t len;
2172 	struct ngbe_adapter *adapter = ngbe_dev_adapter(dev);
2173 	uint64_t offloads;
2174 
2175 	PMD_INIT_FUNC_TRACE();
2176 	hw = ngbe_dev_hw(dev);
2177 
2178 	offloads = rx_conf->offloads | dev->data->dev_conf.rxmode.offloads;
2179 
2180 	/* Free memory prior to re-allocation if needed... */
2181 	if (dev->data->rx_queues[queue_idx] != NULL) {
2182 		ngbe_rx_queue_release(dev->data->rx_queues[queue_idx]);
2183 		dev->data->rx_queues[queue_idx] = NULL;
2184 	}
2185 
2186 	/* First allocate the Rx queue data structure */
2187 	rxq = rte_zmalloc_socket("ethdev RX queue",
2188 				 sizeof(struct ngbe_rx_queue),
2189 				 RTE_CACHE_LINE_SIZE, socket_id);
2190 	if (rxq == NULL)
2191 		return -ENOMEM;
2192 	rxq->mb_pool = mp;
2193 	rxq->nb_rx_desc = nb_desc;
2194 	rxq->rx_free_thresh = rx_conf->rx_free_thresh;
2195 	rxq->queue_id = queue_idx;
2196 	rxq->reg_idx = (uint16_t)((RTE_ETH_DEV_SRIOV(dev).active == 0) ?
2197 		queue_idx : RTE_ETH_DEV_SRIOV(dev).def_pool_q_idx + queue_idx);
2198 	rxq->port_id = dev->data->port_id;
2199 	if (dev->data->dev_conf.rxmode.offloads & RTE_ETH_RX_OFFLOAD_KEEP_CRC)
2200 		rxq->crc_len = RTE_ETHER_CRC_LEN;
2201 	else
2202 		rxq->crc_len = 0;
2203 	rxq->drop_en = rx_conf->rx_drop_en;
2204 	rxq->rx_deferred_start = rx_conf->rx_deferred_start;
2205 	rxq->offloads = offloads;
2206 
2207 	/*
2208 	 * Allocate Rx ring hardware descriptors. A memzone large enough to
2209 	 * handle the maximum ring size is allocated in order to allow for
2210 	 * resizing in later calls to the queue setup function.
2211 	 */
2212 	rz = rte_eth_dma_zone_reserve(dev, "rx_ring", queue_idx,
2213 				      RX_RING_SZ, NGBE_ALIGN, socket_id);
2214 	if (rz == NULL) {
2215 		ngbe_rx_queue_release(rxq);
2216 		return -ENOMEM;
2217 	}
2218 
2219 	/*
2220 	 * Zero init all the descriptors in the ring.
2221 	 */
2222 	memset(rz->addr, 0, RX_RING_SZ);
2223 
2224 	rxq->rdt_reg_addr = NGBE_REG_ADDR(hw, NGBE_RXWP(rxq->reg_idx));
2225 	rxq->rdh_reg_addr = NGBE_REG_ADDR(hw, NGBE_RXRP(rxq->reg_idx));
2226 
2227 	rxq->rx_ring_phys_addr = TMZ_PADDR(rz);
2228 	rxq->rx_ring = (struct ngbe_rx_desc *)TMZ_VADDR(rz);
2229 
2230 	/*
2231 	 * Certain constraints must be met in order to use the bulk buffer
2232 	 * allocation Rx burst function. If any of Rx queues doesn't meet them
2233 	 * the feature should be disabled for the whole port.
2234 	 */
2235 	if (check_rx_burst_bulk_alloc_preconditions(rxq)) {
2236 		PMD_INIT_LOG(DEBUG,
2237 			     "queue[%d] doesn't meet Rx Bulk Alloc preconditions - canceling the feature for the whole port[%d]",
2238 			     rxq->queue_id, rxq->port_id);
2239 		adapter->rx_bulk_alloc_allowed = false;
2240 	}
2241 
2242 	/*
2243 	 * Allocate software ring. Allow for space at the end of the
2244 	 * S/W ring to make sure look-ahead logic in bulk alloc Rx burst
2245 	 * function does not access an invalid memory region.
2246 	 */
2247 	len = nb_desc;
2248 	if (adapter->rx_bulk_alloc_allowed)
2249 		len += RTE_PMD_NGBE_RX_MAX_BURST;
2250 
2251 	rxq->sw_ring = rte_zmalloc_socket("rxq->sw_ring",
2252 					  sizeof(struct ngbe_rx_entry) * len,
2253 					  RTE_CACHE_LINE_SIZE, socket_id);
2254 	if (rxq->sw_ring == NULL) {
2255 		ngbe_rx_queue_release(rxq);
2256 		return -ENOMEM;
2257 	}
2258 
2259 	/*
2260 	 * Always allocate even if it's not going to be needed in order to
2261 	 * simplify the code.
2262 	 *
2263 	 * This ring is used in Scattered Rx cases and Scattered Rx may
2264 	 * be requested in ngbe_dev_rx_init(), which is called later from
2265 	 * dev_start() flow.
2266 	 */
2267 	rxq->sw_sc_ring =
2268 		rte_zmalloc_socket("rxq->sw_sc_ring",
2269 				  sizeof(struct ngbe_scattered_rx_entry) * len,
2270 				  RTE_CACHE_LINE_SIZE, socket_id);
2271 	if (rxq->sw_sc_ring == NULL) {
2272 		ngbe_rx_queue_release(rxq);
2273 		return -ENOMEM;
2274 	}
2275 
2276 	PMD_INIT_LOG(DEBUG,
2277 		     "sw_ring=%p sw_sc_ring=%p hw_ring=%p dma_addr=0x%" PRIx64,
2278 		     rxq->sw_ring, rxq->sw_sc_ring, rxq->rx_ring,
2279 		     rxq->rx_ring_phys_addr);
2280 
2281 	dev->data->rx_queues[queue_idx] = rxq;
2282 
2283 	ngbe_reset_rx_queue(adapter, rxq);
2284 
2285 	return 0;
2286 }
2287 
2288 void
2289 ngbe_dev_clear_queues(struct rte_eth_dev *dev)
2290 {
2291 	unsigned int i;
2292 	struct ngbe_adapter *adapter = ngbe_dev_adapter(dev);
2293 
2294 	PMD_INIT_FUNC_TRACE();
2295 
2296 	for (i = 0; i < dev->data->nb_tx_queues; i++) {
2297 		struct ngbe_tx_queue *txq = dev->data->tx_queues[i];
2298 
2299 		if (txq != NULL) {
2300 			txq->ops->release_mbufs(txq);
2301 			txq->ops->reset(txq);
2302 		}
2303 	}
2304 
2305 	for (i = 0; i < dev->data->nb_rx_queues; i++) {
2306 		struct ngbe_rx_queue *rxq = dev->data->rx_queues[i];
2307 
2308 		if (rxq != NULL) {
2309 			ngbe_rx_queue_release_mbufs(rxq);
2310 			ngbe_reset_rx_queue(adapter, rxq);
2311 		}
2312 	}
2313 }
2314 
2315 void
2316 ngbe_dev_free_queues(struct rte_eth_dev *dev)
2317 {
2318 	unsigned int i;
2319 
2320 	PMD_INIT_FUNC_TRACE();
2321 
2322 	for (i = 0; i < dev->data->nb_rx_queues; i++) {
2323 		ngbe_dev_rx_queue_release(dev, i);
2324 		dev->data->rx_queues[i] = NULL;
2325 	}
2326 	dev->data->nb_rx_queues = 0;
2327 
2328 	for (i = 0; i < dev->data->nb_tx_queues; i++) {
2329 		ngbe_dev_tx_queue_release(dev, i);
2330 		dev->data->tx_queues[i] = NULL;
2331 	}
2332 	dev->data->nb_tx_queues = 0;
2333 }
2334 
2335 /**
2336  * Receive Side Scaling (RSS)
2337  *
2338  * Principles:
2339  * The source and destination IP addresses of the IP header and the source
2340  * and destination ports of TCP/UDP headers, if any, of received packets are
2341  * hashed against a configurable random key to compute a 32-bit RSS hash result.
2342  * The seven (7) LSBs of the 32-bit hash result are used as an index into a
2343  * 128-entry redirection table (RETA).  Each entry of the RETA provides a 3-bit
2344  * RSS output index which is used as the Rx queue index where to store the
2345  * received packets.
2346  * The following output is supplied in the Rx write-back descriptor:
2347  *     - 32-bit result of the Microsoft RSS hash function,
2348  *     - 4-bit RSS type field.
2349  */
2350 
2351 /*
2352  * Used as the default key.
2353  */
2354 static uint8_t rss_intel_key[40] = {
2355 	0x6D, 0x5A, 0x56, 0xDA, 0x25, 0x5B, 0x0E, 0xC2,
2356 	0x41, 0x67, 0x25, 0x3D, 0x43, 0xA3, 0x8F, 0xB0,
2357 	0xD0, 0xCA, 0x2B, 0xCB, 0xAE, 0x7B, 0x30, 0xB4,
2358 	0x77, 0xCB, 0x2D, 0xA3, 0x80, 0x30, 0xF2, 0x0C,
2359 	0x6A, 0x42, 0xB7, 0x3B, 0xBE, 0xAC, 0x01, 0xFA,
2360 };
2361 
2362 static void
2363 ngbe_rss_disable(struct rte_eth_dev *dev)
2364 {
2365 	struct ngbe_hw *hw = ngbe_dev_hw(dev);
2366 
2367 	wr32m(hw, NGBE_RACTL, NGBE_RACTL_RSSENA, 0);
2368 }
2369 
2370 int
2371 ngbe_dev_rss_hash_update(struct rte_eth_dev *dev,
2372 			  struct rte_eth_rss_conf *rss_conf)
2373 {
2374 	struct ngbe_hw *hw = ngbe_dev_hw(dev);
2375 	uint8_t  *hash_key;
2376 	uint32_t mrqc;
2377 	uint32_t rss_key;
2378 	uint64_t rss_hf;
2379 	uint16_t i;
2380 
2381 	if (!hw->is_pf) {
2382 		PMD_DRV_LOG(ERR, "RSS hash update is not supported on this "
2383 			"NIC.");
2384 		return -ENOTSUP;
2385 	}
2386 
2387 	hash_key = rss_conf->rss_key;
2388 	if (hash_key) {
2389 		/* Fill in RSS hash key */
2390 		for (i = 0; i < 10; i++) {
2391 			rss_key  = LS32(hash_key[(i * 4) + 0], 0, 0xFF);
2392 			rss_key |= LS32(hash_key[(i * 4) + 1], 8, 0xFF);
2393 			rss_key |= LS32(hash_key[(i * 4) + 2], 16, 0xFF);
2394 			rss_key |= LS32(hash_key[(i * 4) + 3], 24, 0xFF);
2395 			wr32a(hw, NGBE_REG_RSSKEY, i, rss_key);
2396 		}
2397 	}
2398 
2399 	/* Set configured hashing protocols */
2400 	rss_hf = rss_conf->rss_hf & NGBE_RSS_OFFLOAD_ALL;
2401 
2402 	mrqc = rd32(hw, NGBE_RACTL);
2403 	mrqc &= ~NGBE_RACTL_RSSMASK;
2404 	if (rss_hf & RTE_ETH_RSS_IPV4)
2405 		mrqc |= NGBE_RACTL_RSSIPV4;
2406 	if (rss_hf & RTE_ETH_RSS_NONFRAG_IPV4_TCP)
2407 		mrqc |= NGBE_RACTL_RSSIPV4TCP;
2408 	if (rss_hf & RTE_ETH_RSS_IPV6 ||
2409 	    rss_hf & RTE_ETH_RSS_IPV6_EX)
2410 		mrqc |= NGBE_RACTL_RSSIPV6;
2411 	if (rss_hf & RTE_ETH_RSS_NONFRAG_IPV6_TCP ||
2412 	    rss_hf & RTE_ETH_RSS_IPV6_TCP_EX)
2413 		mrqc |= NGBE_RACTL_RSSIPV6TCP;
2414 	if (rss_hf & RTE_ETH_RSS_NONFRAG_IPV4_UDP)
2415 		mrqc |= NGBE_RACTL_RSSIPV4UDP;
2416 	if (rss_hf & RTE_ETH_RSS_NONFRAG_IPV6_UDP ||
2417 	    rss_hf & RTE_ETH_RSS_IPV6_UDP_EX)
2418 		mrqc |= NGBE_RACTL_RSSIPV6UDP;
2419 
2420 	if (rss_hf)
2421 		mrqc |= NGBE_RACTL_RSSENA;
2422 	else
2423 		mrqc &= ~NGBE_RACTL_RSSENA;
2424 
2425 	wr32(hw, NGBE_RACTL, mrqc);
2426 
2427 	return 0;
2428 }
2429 
2430 int
2431 ngbe_dev_rss_hash_conf_get(struct rte_eth_dev *dev,
2432 			    struct rte_eth_rss_conf *rss_conf)
2433 {
2434 	struct ngbe_hw *hw = ngbe_dev_hw(dev);
2435 	uint8_t *hash_key;
2436 	uint32_t mrqc;
2437 	uint32_t rss_key;
2438 	uint64_t rss_hf;
2439 	uint16_t i;
2440 
2441 	hash_key = rss_conf->rss_key;
2442 	if (hash_key) {
2443 		/* Return RSS hash key */
2444 		for (i = 0; i < 10; i++) {
2445 			rss_key = rd32a(hw, NGBE_REG_RSSKEY, i);
2446 			hash_key[(i * 4) + 0] = RS32(rss_key, 0, 0xFF);
2447 			hash_key[(i * 4) + 1] = RS32(rss_key, 8, 0xFF);
2448 			hash_key[(i * 4) + 2] = RS32(rss_key, 16, 0xFF);
2449 			hash_key[(i * 4) + 3] = RS32(rss_key, 24, 0xFF);
2450 		}
2451 	}
2452 
2453 	rss_hf = 0;
2454 
2455 	mrqc = rd32(hw, NGBE_RACTL);
2456 	if (mrqc & NGBE_RACTL_RSSIPV4)
2457 		rss_hf |= RTE_ETH_RSS_IPV4;
2458 	if (mrqc & NGBE_RACTL_RSSIPV4TCP)
2459 		rss_hf |= RTE_ETH_RSS_NONFRAG_IPV4_TCP;
2460 	if (mrqc & NGBE_RACTL_RSSIPV6)
2461 		rss_hf |= RTE_ETH_RSS_IPV6 |
2462 			  RTE_ETH_RSS_IPV6_EX;
2463 	if (mrqc & NGBE_RACTL_RSSIPV6TCP)
2464 		rss_hf |= RTE_ETH_RSS_NONFRAG_IPV6_TCP |
2465 			  RTE_ETH_RSS_IPV6_TCP_EX;
2466 	if (mrqc & NGBE_RACTL_RSSIPV4UDP)
2467 		rss_hf |= RTE_ETH_RSS_NONFRAG_IPV4_UDP;
2468 	if (mrqc & NGBE_RACTL_RSSIPV6UDP)
2469 		rss_hf |= RTE_ETH_RSS_NONFRAG_IPV6_UDP |
2470 			  RTE_ETH_RSS_IPV6_UDP_EX;
2471 	if (!(mrqc & NGBE_RACTL_RSSENA))
2472 		rss_hf = 0;
2473 
2474 	rss_hf &= NGBE_RSS_OFFLOAD_ALL;
2475 
2476 	rss_conf->rss_hf = rss_hf;
2477 	return 0;
2478 }
2479 
2480 static void
2481 ngbe_rss_configure(struct rte_eth_dev *dev)
2482 {
2483 	struct rte_eth_rss_conf rss_conf;
2484 	struct ngbe_adapter *adapter = ngbe_dev_adapter(dev);
2485 	struct ngbe_hw *hw = ngbe_dev_hw(dev);
2486 	uint32_t reta;
2487 	uint16_t i;
2488 	uint16_t j;
2489 
2490 	PMD_INIT_FUNC_TRACE();
2491 
2492 	/*
2493 	 * Fill in redirection table
2494 	 * The byte-swap is needed because NIC registers are in
2495 	 * little-endian order.
2496 	 */
2497 	if (adapter->rss_reta_updated == 0) {
2498 		reta = 0;
2499 		for (i = 0, j = 0; i < RTE_ETH_RSS_RETA_SIZE_128; i++, j++) {
2500 			if (j == dev->data->nb_rx_queues)
2501 				j = 0;
2502 			reta = (reta >> 8) | LS32(j, 24, 0xFF);
2503 			if ((i & 3) == 3)
2504 				wr32a(hw, NGBE_REG_RSSTBL, i >> 2, reta);
2505 		}
2506 	}
2507 	/*
2508 	 * Configure the RSS key and the RSS protocols used to compute
2509 	 * the RSS hash of input packets.
2510 	 */
2511 	rss_conf = dev->data->dev_conf.rx_adv_conf.rss_conf;
2512 	if (rss_conf.rss_key == NULL)
2513 		rss_conf.rss_key = rss_intel_key; /* Default hash key */
2514 	ngbe_dev_rss_hash_update(dev, &rss_conf);
2515 }
2516 
2517 void ngbe_configure_port(struct rte_eth_dev *dev)
2518 {
2519 	struct ngbe_hw *hw = ngbe_dev_hw(dev);
2520 	int i = 0;
2521 	uint16_t tpids[8] = {RTE_ETHER_TYPE_VLAN, RTE_ETHER_TYPE_QINQ,
2522 				0x9100, 0x9200,
2523 				0x0000, 0x0000,
2524 				0x0000, 0x0000};
2525 
2526 	PMD_INIT_FUNC_TRACE();
2527 
2528 	/* default outer vlan tpid */
2529 	wr32(hw, NGBE_EXTAG,
2530 		NGBE_EXTAG_ETAG(RTE_ETHER_TYPE_ETAG) |
2531 		NGBE_EXTAG_VLAN(RTE_ETHER_TYPE_QINQ));
2532 
2533 	/* default inner vlan tpid */
2534 	wr32m(hw, NGBE_VLANCTL,
2535 		NGBE_VLANCTL_TPID_MASK,
2536 		NGBE_VLANCTL_TPID(RTE_ETHER_TYPE_VLAN));
2537 	wr32m(hw, NGBE_DMATXCTRL,
2538 		NGBE_DMATXCTRL_TPID_MASK,
2539 		NGBE_DMATXCTRL_TPID(RTE_ETHER_TYPE_VLAN));
2540 
2541 	/* default vlan tpid filters */
2542 	for (i = 0; i < 8; i++) {
2543 		wr32m(hw, NGBE_TAGTPID(i / 2),
2544 			(i % 2 ? NGBE_TAGTPID_MSB_MASK
2545 			       : NGBE_TAGTPID_LSB_MASK),
2546 			(i % 2 ? NGBE_TAGTPID_MSB(tpids[i])
2547 			       : NGBE_TAGTPID_LSB(tpids[i])));
2548 	}
2549 }
2550 
2551 static int
2552 ngbe_alloc_rx_queue_mbufs(struct ngbe_rx_queue *rxq)
2553 {
2554 	struct ngbe_rx_entry *rxe = rxq->sw_ring;
2555 	uint64_t dma_addr;
2556 	unsigned int i;
2557 
2558 	/* Initialize software ring entries */
2559 	for (i = 0; i < rxq->nb_rx_desc; i++) {
2560 		/* the ring can also be modified by hardware */
2561 		volatile struct ngbe_rx_desc *rxd;
2562 		struct rte_mbuf *mbuf = rte_mbuf_raw_alloc(rxq->mb_pool);
2563 
2564 		if (mbuf == NULL) {
2565 			PMD_INIT_LOG(ERR, "Rx mbuf alloc failed queue_id=%u port_id=%u",
2566 				     (unsigned int)rxq->queue_id,
2567 				     (unsigned int)rxq->port_id);
2568 			return -ENOMEM;
2569 		}
2570 
2571 		mbuf->data_off = RTE_PKTMBUF_HEADROOM;
2572 		mbuf->port = rxq->port_id;
2573 
2574 		dma_addr =
2575 			rte_cpu_to_le_64(rte_mbuf_data_iova_default(mbuf));
2576 		rxd = &rxq->rx_ring[i];
2577 		NGBE_RXD_HDRADDR(rxd, 0);
2578 		NGBE_RXD_PKTADDR(rxd, dma_addr);
2579 		rxe[i].mbuf = mbuf;
2580 	}
2581 
2582 	return 0;
2583 }
2584 
2585 static int
2586 ngbe_dev_mq_rx_configure(struct rte_eth_dev *dev)
2587 {
2588 	if (RTE_ETH_DEV_SRIOV(dev).active == 0) {
2589 		switch (dev->data->dev_conf.rxmode.mq_mode) {
2590 		case RTE_ETH_MQ_RX_RSS:
2591 			ngbe_rss_configure(dev);
2592 			break;
2593 
2594 		case RTE_ETH_MQ_RX_NONE:
2595 		default:
2596 			/* if mq_mode is none, disable rss mode.*/
2597 			ngbe_rss_disable(dev);
2598 			break;
2599 		}
2600 	}
2601 
2602 	return 0;
2603 }
2604 
2605 void
2606 ngbe_set_rx_function(struct rte_eth_dev *dev)
2607 {
2608 	struct ngbe_adapter *adapter = ngbe_dev_adapter(dev);
2609 
2610 	if (dev->data->scattered_rx) {
2611 		/*
2612 		 * Set the scattered callback: there are bulk and
2613 		 * single allocation versions.
2614 		 */
2615 		if (adapter->rx_bulk_alloc_allowed) {
2616 			PMD_INIT_LOG(DEBUG, "Using a Scattered with bulk "
2617 					   "allocation callback (port=%d).",
2618 				     dev->data->port_id);
2619 			dev->rx_pkt_burst = ngbe_recv_pkts_sc_bulk_alloc;
2620 		} else {
2621 			PMD_INIT_LOG(DEBUG, "Using Regular (non-vector, "
2622 					    "single allocation) "
2623 					    "Scattered Rx callback "
2624 					    "(port=%d).",
2625 				     dev->data->port_id);
2626 
2627 			dev->rx_pkt_burst = ngbe_recv_pkts_sc_single_alloc;
2628 		}
2629 	/*
2630 	 * Below we set "simple" callbacks according to port/queues parameters.
2631 	 * If parameters allow we are going to choose between the following
2632 	 * callbacks:
2633 	 *    - Bulk Allocation
2634 	 *    - Single buffer allocation (the simplest one)
2635 	 */
2636 	} else if (adapter->rx_bulk_alloc_allowed) {
2637 		PMD_INIT_LOG(DEBUG, "Rx Burst Bulk Alloc Preconditions are "
2638 				    "satisfied. Rx Burst Bulk Alloc function "
2639 				    "will be used on port=%d.",
2640 			     dev->data->port_id);
2641 
2642 		dev->rx_pkt_burst = ngbe_recv_pkts_bulk_alloc;
2643 	} else {
2644 		PMD_INIT_LOG(DEBUG, "Rx Burst Bulk Alloc Preconditions are not "
2645 				    "satisfied, or Scattered Rx is requested "
2646 				    "(port=%d).",
2647 			     dev->data->port_id);
2648 
2649 		dev->rx_pkt_burst = ngbe_recv_pkts;
2650 	}
2651 }
2652 
2653 static const struct {
2654 	eth_rx_burst_t pkt_burst;
2655 	const char *info;
2656 } ngbe_rx_burst_infos[] = {
2657 	{ ngbe_recv_pkts_sc_single_alloc,    "Scalar Scattered"},
2658 	{ ngbe_recv_pkts_sc_bulk_alloc,      "Scalar Scattered Bulk Alloc"},
2659 	{ ngbe_recv_pkts_bulk_alloc,         "Scalar Bulk Alloc"},
2660 	{ ngbe_recv_pkts,                    "Scalar"},
2661 };
2662 
2663 int
2664 ngbe_rx_burst_mode_get(struct rte_eth_dev *dev, __rte_unused uint16_t queue_id,
2665 		      struct rte_eth_burst_mode *mode)
2666 {
2667 	eth_rx_burst_t pkt_burst = dev->rx_pkt_burst;
2668 	int ret = -EINVAL;
2669 	unsigned int i;
2670 
2671 	for (i = 0; i < RTE_DIM(ngbe_rx_burst_infos); ++i) {
2672 		if (pkt_burst == ngbe_rx_burst_infos[i].pkt_burst) {
2673 			snprintf(mode->info, sizeof(mode->info), "%s",
2674 				 ngbe_rx_burst_infos[i].info);
2675 			ret = 0;
2676 			break;
2677 		}
2678 	}
2679 
2680 	return ret;
2681 }
2682 
2683 /*
2684  * Initializes Receive Unit.
2685  */
2686 int
2687 ngbe_dev_rx_init(struct rte_eth_dev *dev)
2688 {
2689 	struct ngbe_hw *hw;
2690 	struct ngbe_rx_queue *rxq;
2691 	uint64_t bus_addr;
2692 	uint32_t fctrl;
2693 	uint32_t hlreg0;
2694 	uint32_t srrctl;
2695 	uint32_t rdrxctl;
2696 	uint32_t rxcsum;
2697 	uint16_t buf_size;
2698 	uint16_t i;
2699 	struct rte_eth_rxmode *rx_conf = &dev->data->dev_conf.rxmode;
2700 
2701 	PMD_INIT_FUNC_TRACE();
2702 	hw = ngbe_dev_hw(dev);
2703 
2704 	/*
2705 	 * Make sure receives are disabled while setting
2706 	 * up the Rx context (registers, descriptor rings, etc.).
2707 	 */
2708 	wr32m(hw, NGBE_MACRXCFG, NGBE_MACRXCFG_ENA, 0);
2709 	wr32m(hw, NGBE_PBRXCTL, NGBE_PBRXCTL_ENA, 0);
2710 
2711 	/* Enable receipt of broadcasted frames */
2712 	fctrl = rd32(hw, NGBE_PSRCTL);
2713 	fctrl |= NGBE_PSRCTL_BCA;
2714 	wr32(hw, NGBE_PSRCTL, fctrl);
2715 
2716 	/*
2717 	 * Configure CRC stripping, if any.
2718 	 */
2719 	hlreg0 = rd32(hw, NGBE_SECRXCTL);
2720 	if (rx_conf->offloads & RTE_ETH_RX_OFFLOAD_KEEP_CRC)
2721 		hlreg0 &= ~NGBE_SECRXCTL_CRCSTRIP;
2722 	else
2723 		hlreg0 |= NGBE_SECRXCTL_CRCSTRIP;
2724 	hlreg0 &= ~NGBE_SECRXCTL_XDSA;
2725 	wr32(hw, NGBE_SECRXCTL, hlreg0);
2726 
2727 	/*
2728 	 * Configure jumbo frame support, if any.
2729 	 */
2730 	wr32m(hw, NGBE_FRMSZ, NGBE_FRMSZ_MAX_MASK,
2731 		NGBE_FRMSZ_MAX(dev->data->mtu + NGBE_ETH_OVERHEAD));
2732 
2733 	/*
2734 	 * If loopback mode is configured, set LPBK bit.
2735 	 */
2736 	hlreg0 = rd32(hw, NGBE_PSRCTL);
2737 	if (hw->is_pf && dev->data->dev_conf.lpbk_mode)
2738 		hlreg0 |= NGBE_PSRCTL_LBENA;
2739 	else
2740 		hlreg0 &= ~NGBE_PSRCTL_LBENA;
2741 
2742 	wr32(hw, NGBE_PSRCTL, hlreg0);
2743 
2744 	/*
2745 	 * Assume no header split and no VLAN strip support
2746 	 * on any Rx queue first .
2747 	 */
2748 	rx_conf->offloads &= ~RTE_ETH_RX_OFFLOAD_VLAN_STRIP;
2749 
2750 	/* Setup Rx queues */
2751 	for (i = 0; i < dev->data->nb_rx_queues; i++) {
2752 		rxq = dev->data->rx_queues[i];
2753 
2754 		/*
2755 		 * Reset crc_len in case it was changed after queue setup by a
2756 		 * call to configure.
2757 		 */
2758 		if (rx_conf->offloads & RTE_ETH_RX_OFFLOAD_KEEP_CRC)
2759 			rxq->crc_len = RTE_ETHER_CRC_LEN;
2760 		else
2761 			rxq->crc_len = 0;
2762 
2763 		/* Setup the Base and Length of the Rx Descriptor Rings */
2764 		bus_addr = rxq->rx_ring_phys_addr;
2765 		wr32(hw, NGBE_RXBAL(rxq->reg_idx),
2766 				(uint32_t)(bus_addr & BIT_MASK32));
2767 		wr32(hw, NGBE_RXBAH(rxq->reg_idx),
2768 				(uint32_t)(bus_addr >> 32));
2769 		wr32(hw, NGBE_RXRP(rxq->reg_idx), 0);
2770 		wr32(hw, NGBE_RXWP(rxq->reg_idx), 0);
2771 
2772 		srrctl = NGBE_RXCFG_RNGLEN(rxq->nb_rx_desc);
2773 
2774 		/* Set if packets are dropped when no descriptors available */
2775 		if (rxq->drop_en)
2776 			srrctl |= NGBE_RXCFG_DROP;
2777 
2778 		/*
2779 		 * Configure the Rx buffer size in the PKTLEN field of
2780 		 * the RXCFG register of the queue.
2781 		 * The value is in 1 KB resolution. Valid values can be from
2782 		 * 1 KB to 16 KB.
2783 		 */
2784 		buf_size = (uint16_t)(rte_pktmbuf_data_room_size(rxq->mb_pool) -
2785 			RTE_PKTMBUF_HEADROOM);
2786 		buf_size = ROUND_DOWN(buf_size, 0x1 << 10);
2787 		srrctl |= NGBE_RXCFG_PKTLEN(buf_size);
2788 
2789 		wr32(hw, NGBE_RXCFG(rxq->reg_idx), srrctl);
2790 
2791 		/* It adds dual VLAN length for supporting dual VLAN */
2792 		if (dev->data->mtu + NGBE_ETH_OVERHEAD +
2793 				2 * NGBE_VLAN_TAG_SIZE > buf_size)
2794 			dev->data->scattered_rx = 1;
2795 		if (rxq->offloads & RTE_ETH_RX_OFFLOAD_VLAN_STRIP)
2796 			rx_conf->offloads |= RTE_ETH_RX_OFFLOAD_VLAN_STRIP;
2797 	}
2798 
2799 	if (rx_conf->offloads & RTE_ETH_RX_OFFLOAD_SCATTER)
2800 		dev->data->scattered_rx = 1;
2801 
2802 	/*
2803 	 * Device configured with multiple RX queues.
2804 	 */
2805 	ngbe_dev_mq_rx_configure(dev);
2806 
2807 	/*
2808 	 * Setup the Checksum Register.
2809 	 * Disable Full-Packet Checksum which is mutually exclusive with RSS.
2810 	 * Enable IP/L4 checksum computation by hardware if requested to do so.
2811 	 */
2812 	rxcsum = rd32(hw, NGBE_PSRCTL);
2813 	rxcsum |= NGBE_PSRCTL_PCSD;
2814 	if (rx_conf->offloads & RTE_ETH_RX_OFFLOAD_CHECKSUM)
2815 		rxcsum |= NGBE_PSRCTL_L4CSUM;
2816 	else
2817 		rxcsum &= ~NGBE_PSRCTL_L4CSUM;
2818 
2819 	wr32(hw, NGBE_PSRCTL, rxcsum);
2820 
2821 	if (hw->is_pf) {
2822 		rdrxctl = rd32(hw, NGBE_SECRXCTL);
2823 		if (rx_conf->offloads & RTE_ETH_RX_OFFLOAD_KEEP_CRC)
2824 			rdrxctl &= ~NGBE_SECRXCTL_CRCSTRIP;
2825 		else
2826 			rdrxctl |= NGBE_SECRXCTL_CRCSTRIP;
2827 		wr32(hw, NGBE_SECRXCTL, rdrxctl);
2828 	}
2829 
2830 	ngbe_set_rx_function(dev);
2831 
2832 	return 0;
2833 }
2834 
2835 /*
2836  * Initializes Transmit Unit.
2837  */
2838 void
2839 ngbe_dev_tx_init(struct rte_eth_dev *dev)
2840 {
2841 	struct ngbe_hw     *hw;
2842 	struct ngbe_tx_queue *txq;
2843 	uint64_t bus_addr;
2844 	uint16_t i;
2845 
2846 	PMD_INIT_FUNC_TRACE();
2847 	hw = ngbe_dev_hw(dev);
2848 
2849 	wr32m(hw, NGBE_SECTXCTL, NGBE_SECTXCTL_ODSA, NGBE_SECTXCTL_ODSA);
2850 	wr32m(hw, NGBE_SECTXCTL, NGBE_SECTXCTL_XDSA, 0);
2851 
2852 	/* Setup the Base and Length of the Tx Descriptor Rings */
2853 	for (i = 0; i < dev->data->nb_tx_queues; i++) {
2854 		txq = dev->data->tx_queues[i];
2855 
2856 		bus_addr = txq->tx_ring_phys_addr;
2857 		wr32(hw, NGBE_TXBAL(txq->reg_idx),
2858 				(uint32_t)(bus_addr & BIT_MASK32));
2859 		wr32(hw, NGBE_TXBAH(txq->reg_idx),
2860 				(uint32_t)(bus_addr >> 32));
2861 		wr32m(hw, NGBE_TXCFG(txq->reg_idx), NGBE_TXCFG_BUFLEN_MASK,
2862 			NGBE_TXCFG_BUFLEN(txq->nb_tx_desc));
2863 		/* Setup the HW Tx Head and TX Tail descriptor pointers */
2864 		wr32(hw, NGBE_TXRP(txq->reg_idx), 0);
2865 		wr32(hw, NGBE_TXWP(txq->reg_idx), 0);
2866 	}
2867 }
2868 
2869 /*
2870  * Set up link loopback mode Tx->Rx.
2871  */
2872 static inline void
2873 ngbe_setup_loopback_link(struct ngbe_hw *hw)
2874 {
2875 	PMD_INIT_FUNC_TRACE();
2876 
2877 	wr32m(hw, NGBE_MACRXCFG, NGBE_MACRXCFG_LB, NGBE_MACRXCFG_LB);
2878 
2879 	msec_delay(50);
2880 }
2881 
2882 /*
2883  * Start Transmit and Receive Units.
2884  */
2885 int
2886 ngbe_dev_rxtx_start(struct rte_eth_dev *dev)
2887 {
2888 	struct ngbe_hw     *hw;
2889 	struct ngbe_tx_queue *txq;
2890 	struct ngbe_rx_queue *rxq;
2891 	uint32_t dmatxctl;
2892 	uint32_t rxctrl;
2893 	uint16_t i;
2894 	int ret = 0;
2895 
2896 	PMD_INIT_FUNC_TRACE();
2897 	hw = ngbe_dev_hw(dev);
2898 
2899 	for (i = 0; i < dev->data->nb_tx_queues; i++) {
2900 		txq = dev->data->tx_queues[i];
2901 		/* Setup Transmit Threshold Registers */
2902 		wr32m(hw, NGBE_TXCFG(txq->reg_idx),
2903 		      NGBE_TXCFG_HTHRESH_MASK |
2904 		      NGBE_TXCFG_WTHRESH_MASK,
2905 		      NGBE_TXCFG_HTHRESH(txq->hthresh) |
2906 		      NGBE_TXCFG_WTHRESH(txq->wthresh));
2907 	}
2908 
2909 	dmatxctl = rd32(hw, NGBE_DMATXCTRL);
2910 	dmatxctl |= NGBE_DMATXCTRL_ENA;
2911 	wr32(hw, NGBE_DMATXCTRL, dmatxctl);
2912 
2913 	for (i = 0; i < dev->data->nb_tx_queues; i++) {
2914 		txq = dev->data->tx_queues[i];
2915 		if (txq->tx_deferred_start == 0) {
2916 			ret = ngbe_dev_tx_queue_start(dev, i);
2917 			if (ret < 0)
2918 				return ret;
2919 		}
2920 	}
2921 
2922 	for (i = 0; i < dev->data->nb_rx_queues; i++) {
2923 		rxq = dev->data->rx_queues[i];
2924 		if (rxq->rx_deferred_start == 0) {
2925 			ret = ngbe_dev_rx_queue_start(dev, i);
2926 			if (ret < 0)
2927 				return ret;
2928 		}
2929 	}
2930 
2931 	/* Enable Receive engine */
2932 	rxctrl = rd32(hw, NGBE_PBRXCTL);
2933 	rxctrl |= NGBE_PBRXCTL_ENA;
2934 	hw->mac.enable_rx_dma(hw, rxctrl);
2935 
2936 	/* If loopback mode is enabled, set up the link accordingly */
2937 	if (hw->is_pf && dev->data->dev_conf.lpbk_mode)
2938 		ngbe_setup_loopback_link(hw);
2939 
2940 	return 0;
2941 }
2942 
2943 void
2944 ngbe_dev_save_rx_queue(struct ngbe_hw *hw, uint16_t rx_queue_id)
2945 {
2946 	u32 *reg = &hw->q_rx_regs[rx_queue_id * 8];
2947 	*(reg++) = rd32(hw, NGBE_RXBAL(rx_queue_id));
2948 	*(reg++) = rd32(hw, NGBE_RXBAH(rx_queue_id));
2949 	*(reg++) = rd32(hw, NGBE_RXCFG(rx_queue_id));
2950 }
2951 
2952 void
2953 ngbe_dev_store_rx_queue(struct ngbe_hw *hw, uint16_t rx_queue_id)
2954 {
2955 	u32 *reg = &hw->q_rx_regs[rx_queue_id * 8];
2956 	wr32(hw, NGBE_RXBAL(rx_queue_id), *(reg++));
2957 	wr32(hw, NGBE_RXBAH(rx_queue_id), *(reg++));
2958 	wr32(hw, NGBE_RXCFG(rx_queue_id), *(reg++) & ~NGBE_RXCFG_ENA);
2959 }
2960 
2961 void
2962 ngbe_dev_save_tx_queue(struct ngbe_hw *hw, uint16_t tx_queue_id)
2963 {
2964 	u32 *reg = &hw->q_tx_regs[tx_queue_id * 8];
2965 	*(reg++) = rd32(hw, NGBE_TXBAL(tx_queue_id));
2966 	*(reg++) = rd32(hw, NGBE_TXBAH(tx_queue_id));
2967 	*(reg++) = rd32(hw, NGBE_TXCFG(tx_queue_id));
2968 }
2969 
2970 void
2971 ngbe_dev_store_tx_queue(struct ngbe_hw *hw, uint16_t tx_queue_id)
2972 {
2973 	u32 *reg = &hw->q_tx_regs[tx_queue_id * 8];
2974 	wr32(hw, NGBE_TXBAL(tx_queue_id), *(reg++));
2975 	wr32(hw, NGBE_TXBAH(tx_queue_id), *(reg++));
2976 	wr32(hw, NGBE_TXCFG(tx_queue_id), *(reg++) & ~NGBE_TXCFG_ENA);
2977 }
2978 
2979 /*
2980  * Start Receive Units for specified queue.
2981  */
2982 int
2983 ngbe_dev_rx_queue_start(struct rte_eth_dev *dev, uint16_t rx_queue_id)
2984 {
2985 	struct ngbe_hw *hw = ngbe_dev_hw(dev);
2986 	struct ngbe_rx_queue *rxq;
2987 	uint32_t rxdctl;
2988 	int poll_ms;
2989 
2990 	PMD_INIT_FUNC_TRACE();
2991 
2992 	rxq = dev->data->rx_queues[rx_queue_id];
2993 
2994 	/* Allocate buffers for descriptor rings */
2995 	if (ngbe_alloc_rx_queue_mbufs(rxq) != 0) {
2996 		PMD_INIT_LOG(ERR, "Could not alloc mbuf for queue:%d",
2997 			     rx_queue_id);
2998 		return -1;
2999 	}
3000 	rxdctl = rd32(hw, NGBE_RXCFG(rxq->reg_idx));
3001 	rxdctl |= NGBE_RXCFG_ENA;
3002 	wr32(hw, NGBE_RXCFG(rxq->reg_idx), rxdctl);
3003 
3004 	/* Wait until Rx Enable ready */
3005 	poll_ms = RTE_NGBE_REGISTER_POLL_WAIT_10_MS;
3006 	do {
3007 		rte_delay_ms(1);
3008 		rxdctl = rd32(hw, NGBE_RXCFG(rxq->reg_idx));
3009 	} while (--poll_ms && !(rxdctl & NGBE_RXCFG_ENA));
3010 	if (poll_ms == 0)
3011 		PMD_INIT_LOG(ERR, "Could not enable Rx Queue %d", rx_queue_id);
3012 	rte_wmb();
3013 	wr32(hw, NGBE_RXRP(rxq->reg_idx), 0);
3014 	wr32(hw, NGBE_RXWP(rxq->reg_idx), rxq->nb_rx_desc - 1);
3015 	dev->data->rx_queue_state[rx_queue_id] = RTE_ETH_QUEUE_STATE_STARTED;
3016 
3017 	return 0;
3018 }
3019 
3020 /*
3021  * Stop Receive Units for specified queue.
3022  */
3023 int
3024 ngbe_dev_rx_queue_stop(struct rte_eth_dev *dev, uint16_t rx_queue_id)
3025 {
3026 	struct ngbe_hw *hw = ngbe_dev_hw(dev);
3027 	struct ngbe_adapter *adapter = ngbe_dev_adapter(dev);
3028 	struct ngbe_rx_queue *rxq;
3029 	uint32_t rxdctl;
3030 	int poll_ms;
3031 
3032 	PMD_INIT_FUNC_TRACE();
3033 
3034 	rxq = dev->data->rx_queues[rx_queue_id];
3035 
3036 	ngbe_dev_save_rx_queue(hw, rxq->reg_idx);
3037 	wr32m(hw, NGBE_RXCFG(rxq->reg_idx), NGBE_RXCFG_ENA, 0);
3038 
3039 	/* Wait until Rx Enable bit clear */
3040 	poll_ms = RTE_NGBE_REGISTER_POLL_WAIT_10_MS;
3041 	do {
3042 		rte_delay_ms(1);
3043 		rxdctl = rd32(hw, NGBE_RXCFG(rxq->reg_idx));
3044 	} while (--poll_ms && (rxdctl & NGBE_RXCFG_ENA));
3045 	if (poll_ms == 0)
3046 		PMD_INIT_LOG(ERR, "Could not disable Rx Queue %d", rx_queue_id);
3047 
3048 	rte_delay_us(RTE_NGBE_WAIT_100_US);
3049 	ngbe_dev_store_rx_queue(hw, rxq->reg_idx);
3050 
3051 	ngbe_rx_queue_release_mbufs(rxq);
3052 	ngbe_reset_rx_queue(adapter, rxq);
3053 	dev->data->rx_queue_state[rx_queue_id] = RTE_ETH_QUEUE_STATE_STOPPED;
3054 
3055 	return 0;
3056 }
3057 
3058 /*
3059  * Start Transmit Units for specified queue.
3060  */
3061 int
3062 ngbe_dev_tx_queue_start(struct rte_eth_dev *dev, uint16_t tx_queue_id)
3063 {
3064 	struct ngbe_hw *hw = ngbe_dev_hw(dev);
3065 	struct ngbe_tx_queue *txq;
3066 	uint32_t txdctl;
3067 	int poll_ms;
3068 
3069 	PMD_INIT_FUNC_TRACE();
3070 
3071 	txq = dev->data->tx_queues[tx_queue_id];
3072 	wr32m(hw, NGBE_TXCFG(txq->reg_idx), NGBE_TXCFG_ENA, NGBE_TXCFG_ENA);
3073 
3074 	/* Wait until Tx Enable ready */
3075 	poll_ms = RTE_NGBE_REGISTER_POLL_WAIT_10_MS;
3076 	do {
3077 		rte_delay_ms(1);
3078 		txdctl = rd32(hw, NGBE_TXCFG(txq->reg_idx));
3079 	} while (--poll_ms && !(txdctl & NGBE_TXCFG_ENA));
3080 	if (poll_ms == 0)
3081 		PMD_INIT_LOG(ERR, "Could not enable Tx Queue %d",
3082 			     tx_queue_id);
3083 
3084 	rte_wmb();
3085 	wr32(hw, NGBE_TXWP(txq->reg_idx), txq->tx_tail);
3086 	dev->data->tx_queue_state[tx_queue_id] = RTE_ETH_QUEUE_STATE_STARTED;
3087 
3088 	return 0;
3089 }
3090 
3091 /*
3092  * Stop Transmit Units for specified queue.
3093  */
3094 int
3095 ngbe_dev_tx_queue_stop(struct rte_eth_dev *dev, uint16_t tx_queue_id)
3096 {
3097 	struct ngbe_hw *hw = ngbe_dev_hw(dev);
3098 	struct ngbe_tx_queue *txq;
3099 	uint32_t txdctl;
3100 	uint32_t txtdh, txtdt;
3101 	int poll_ms;
3102 
3103 	PMD_INIT_FUNC_TRACE();
3104 
3105 	txq = dev->data->tx_queues[tx_queue_id];
3106 
3107 	/* Wait until Tx queue is empty */
3108 	poll_ms = RTE_NGBE_REGISTER_POLL_WAIT_10_MS;
3109 	do {
3110 		rte_delay_us(RTE_NGBE_WAIT_100_US);
3111 		txtdh = rd32(hw, NGBE_TXRP(txq->reg_idx));
3112 		txtdt = rd32(hw, NGBE_TXWP(txq->reg_idx));
3113 	} while (--poll_ms && (txtdh != txtdt));
3114 	if (poll_ms == 0)
3115 		PMD_INIT_LOG(ERR, "Tx Queue %d is not empty when stopping.",
3116 			     tx_queue_id);
3117 
3118 	ngbe_dev_save_tx_queue(hw, txq->reg_idx);
3119 	wr32m(hw, NGBE_TXCFG(txq->reg_idx), NGBE_TXCFG_ENA, 0);
3120 
3121 	/* Wait until Tx Enable bit clear */
3122 	poll_ms = RTE_NGBE_REGISTER_POLL_WAIT_10_MS;
3123 	do {
3124 		rte_delay_ms(1);
3125 		txdctl = rd32(hw, NGBE_TXCFG(txq->reg_idx));
3126 	} while (--poll_ms && (txdctl & NGBE_TXCFG_ENA));
3127 	if (poll_ms == 0)
3128 		PMD_INIT_LOG(ERR, "Could not disable Tx Queue %d",
3129 			     tx_queue_id);
3130 
3131 	rte_delay_us(RTE_NGBE_WAIT_100_US);
3132 	ngbe_dev_store_tx_queue(hw, txq->reg_idx);
3133 
3134 	if (txq->ops != NULL) {
3135 		txq->ops->release_mbufs(txq);
3136 		txq->ops->reset(txq);
3137 	}
3138 	dev->data->tx_queue_state[tx_queue_id] = RTE_ETH_QUEUE_STATE_STOPPED;
3139 
3140 	return 0;
3141 }
3142