1 // SPDX-License-Identifier: (GPL-2.0-only OR BSD-2-Clause)
2 /* Copyright (C) 2015-2019 Netronome Systems, Inc. */
3 
4 #include <linux/bpf_trace.h>
5 #include <linux/netdevice.h>
6 
7 #include "../nfp_app.h"
8 #include "../nfp_net.h"
9 #include "../nfp_net_dp.h"
10 #include "../nfp_net_xsk.h"
11 #include "../crypto/crypto.h"
12 #include "../crypto/fw.h"
13 #include "nfd3.h"
14 
15 /* Transmit processing
16  *
17  * One queue controller peripheral queue is used for transmit.  The
18  * driver en-queues packets for transmit by advancing the write
19  * pointer.  The device indicates that packets have transmitted by
20  * advancing the read pointer.  The driver maintains a local copy of
21  * the read and write pointer in @struct nfp_net_tx_ring.  The driver
22  * keeps @wr_p in sync with the queue controller write pointer and can
23  * determine how many packets have been transmitted by comparing its
24  * copy of the read pointer @rd_p with the read pointer maintained by
25  * the queue controller peripheral.
26  */
27 
28 /* Wrappers for deciding when to stop and restart TX queues */
29 static int nfp_nfd3_tx_ring_should_wake(struct nfp_net_tx_ring *tx_ring)
30 {
31 	return !nfp_net_tx_full(tx_ring, MAX_SKB_FRAGS * 4);
32 }
33 
34 static int nfp_nfd3_tx_ring_should_stop(struct nfp_net_tx_ring *tx_ring)
35 {
36 	return nfp_net_tx_full(tx_ring, MAX_SKB_FRAGS + 1);
37 }
38 
39 /**
40  * nfp_nfd3_tx_ring_stop() - stop tx ring
41  * @nd_q:    netdev queue
42  * @tx_ring: driver tx queue structure
43  *
44  * Safely stop TX ring.  Remember that while we are running .start_xmit()
45  * someone else may be cleaning the TX ring completions so we need to be
46  * extra careful here.
47  */
48 static void
49 nfp_nfd3_tx_ring_stop(struct netdev_queue *nd_q,
50 		      struct nfp_net_tx_ring *tx_ring)
51 {
52 	netif_tx_stop_queue(nd_q);
53 
54 	/* We can race with the TX completion out of NAPI so recheck */
55 	smp_mb();
56 	if (unlikely(nfp_nfd3_tx_ring_should_wake(tx_ring)))
57 		netif_tx_start_queue(nd_q);
58 }
59 
60 /**
61  * nfp_nfd3_tx_tso() - Set up Tx descriptor for LSO
62  * @r_vec: per-ring structure
63  * @txbuf: Pointer to driver soft TX descriptor
64  * @txd: Pointer to HW TX descriptor
65  * @skb: Pointer to SKB
66  * @md_bytes: Prepend length
67  *
68  * Set up Tx descriptor for LSO, do nothing for non-LSO skbs.
69  * Return error on packet header greater than maximum supported LSO header size.
70  */
71 static void
72 nfp_nfd3_tx_tso(struct nfp_net_r_vector *r_vec, struct nfp_nfd3_tx_buf *txbuf,
73 		struct nfp_nfd3_tx_desc *txd, struct sk_buff *skb, u32 md_bytes)
74 {
75 	u32 l3_offset, l4_offset, hdrlen;
76 	u16 mss;
77 
78 	if (!skb_is_gso(skb))
79 		return;
80 
81 	if (!skb->encapsulation) {
82 		l3_offset = skb_network_offset(skb);
83 		l4_offset = skb_transport_offset(skb);
84 		hdrlen = skb_transport_offset(skb) + tcp_hdrlen(skb);
85 	} else {
86 		l3_offset = skb_inner_network_offset(skb);
87 		l4_offset = skb_inner_transport_offset(skb);
88 		hdrlen = skb_inner_transport_header(skb) - skb->data +
89 			inner_tcp_hdrlen(skb);
90 	}
91 
92 	txbuf->pkt_cnt = skb_shinfo(skb)->gso_segs;
93 	txbuf->real_len += hdrlen * (txbuf->pkt_cnt - 1);
94 
95 	mss = skb_shinfo(skb)->gso_size & NFD3_DESC_TX_MSS_MASK;
96 	txd->l3_offset = l3_offset - md_bytes;
97 	txd->l4_offset = l4_offset - md_bytes;
98 	txd->lso_hdrlen = hdrlen - md_bytes;
99 	txd->mss = cpu_to_le16(mss);
100 	txd->flags |= NFD3_DESC_TX_LSO;
101 
102 	u64_stats_update_begin(&r_vec->tx_sync);
103 	r_vec->tx_lso++;
104 	u64_stats_update_end(&r_vec->tx_sync);
105 }
106 
107 /**
108  * nfp_nfd3_tx_csum() - Set TX CSUM offload flags in TX descriptor
109  * @dp:  NFP Net data path struct
110  * @r_vec: per-ring structure
111  * @txbuf: Pointer to driver soft TX descriptor
112  * @txd: Pointer to TX descriptor
113  * @skb: Pointer to SKB
114  *
115  * This function sets the TX checksum flags in the TX descriptor based
116  * on the configuration and the protocol of the packet to be transmitted.
117  */
118 static void
119 nfp_nfd3_tx_csum(struct nfp_net_dp *dp, struct nfp_net_r_vector *r_vec,
120 		 struct nfp_nfd3_tx_buf *txbuf, struct nfp_nfd3_tx_desc *txd,
121 		 struct sk_buff *skb)
122 {
123 	struct ipv6hdr *ipv6h;
124 	struct iphdr *iph;
125 	u8 l4_hdr;
126 
127 	if (!(dp->ctrl & NFP_NET_CFG_CTRL_TXCSUM))
128 		return;
129 
130 	if (skb->ip_summed != CHECKSUM_PARTIAL)
131 		return;
132 
133 	txd->flags |= NFD3_DESC_TX_CSUM;
134 	if (skb->encapsulation)
135 		txd->flags |= NFD3_DESC_TX_ENCAP;
136 
137 	iph = skb->encapsulation ? inner_ip_hdr(skb) : ip_hdr(skb);
138 	ipv6h = skb->encapsulation ? inner_ipv6_hdr(skb) : ipv6_hdr(skb);
139 
140 	if (iph->version == 4) {
141 		txd->flags |= NFD3_DESC_TX_IP4_CSUM;
142 		l4_hdr = iph->protocol;
143 	} else if (ipv6h->version == 6) {
144 		l4_hdr = ipv6h->nexthdr;
145 	} else {
146 		nn_dp_warn(dp, "partial checksum but ipv=%x!\n", iph->version);
147 		return;
148 	}
149 
150 	switch (l4_hdr) {
151 	case IPPROTO_TCP:
152 		txd->flags |= NFD3_DESC_TX_TCP_CSUM;
153 		break;
154 	case IPPROTO_UDP:
155 		txd->flags |= NFD3_DESC_TX_UDP_CSUM;
156 		break;
157 	default:
158 		nn_dp_warn(dp, "partial checksum but l4 proto=%x!\n", l4_hdr);
159 		return;
160 	}
161 
162 	u64_stats_update_begin(&r_vec->tx_sync);
163 	if (skb->encapsulation)
164 		r_vec->hw_csum_tx_inner += txbuf->pkt_cnt;
165 	else
166 		r_vec->hw_csum_tx += txbuf->pkt_cnt;
167 	u64_stats_update_end(&r_vec->tx_sync);
168 }
169 
170 static int nfp_nfd3_prep_tx_meta(struct sk_buff *skb, u64 tls_handle)
171 {
172 	struct metadata_dst *md_dst = skb_metadata_dst(skb);
173 	unsigned char *data;
174 	u32 meta_id = 0;
175 	int md_bytes;
176 
177 	if (likely(!md_dst && !tls_handle))
178 		return 0;
179 	if (unlikely(md_dst && md_dst->type != METADATA_HW_PORT_MUX)) {
180 		if (!tls_handle)
181 			return 0;
182 		md_dst = NULL;
183 	}
184 
185 	md_bytes = 4 + !!md_dst * 4 + !!tls_handle * 8;
186 
187 	if (unlikely(skb_cow_head(skb, md_bytes)))
188 		return -ENOMEM;
189 
190 	meta_id = 0;
191 	data = skb_push(skb, md_bytes) + md_bytes;
192 	if (md_dst) {
193 		data -= 4;
194 		put_unaligned_be32(md_dst->u.port_info.port_id, data);
195 		meta_id = NFP_NET_META_PORTID;
196 	}
197 	if (tls_handle) {
198 		/* conn handle is opaque, we just use u64 to be able to quickly
199 		 * compare it to zero
200 		 */
201 		data -= 8;
202 		memcpy(data, &tls_handle, sizeof(tls_handle));
203 		meta_id <<= NFP_NET_META_FIELD_SIZE;
204 		meta_id |= NFP_NET_META_CONN_HANDLE;
205 	}
206 
207 	data -= 4;
208 	put_unaligned_be32(meta_id, data);
209 
210 	return md_bytes;
211 }
212 
213 /**
214  * nfp_nfd3_tx() - Main transmit entry point
215  * @skb:    SKB to transmit
216  * @netdev: netdev structure
217  *
218  * Return: NETDEV_TX_OK on success.
219  */
220 netdev_tx_t nfp_nfd3_tx(struct sk_buff *skb, struct net_device *netdev)
221 {
222 	struct nfp_net *nn = netdev_priv(netdev);
223 	int f, nr_frags, wr_idx, md_bytes;
224 	struct nfp_net_tx_ring *tx_ring;
225 	struct nfp_net_r_vector *r_vec;
226 	struct nfp_nfd3_tx_buf *txbuf;
227 	struct nfp_nfd3_tx_desc *txd;
228 	struct netdev_queue *nd_q;
229 	const skb_frag_t *frag;
230 	struct nfp_net_dp *dp;
231 	dma_addr_t dma_addr;
232 	unsigned int fsize;
233 	u64 tls_handle = 0;
234 	u16 qidx;
235 
236 	dp = &nn->dp;
237 	qidx = skb_get_queue_mapping(skb);
238 	tx_ring = &dp->tx_rings[qidx];
239 	r_vec = tx_ring->r_vec;
240 
241 	nr_frags = skb_shinfo(skb)->nr_frags;
242 
243 	if (unlikely(nfp_net_tx_full(tx_ring, nr_frags + 1))) {
244 		nn_dp_warn(dp, "TX ring %d busy. wrp=%u rdp=%u\n",
245 			   qidx, tx_ring->wr_p, tx_ring->rd_p);
246 		nd_q = netdev_get_tx_queue(dp->netdev, qidx);
247 		netif_tx_stop_queue(nd_q);
248 		nfp_net_tx_xmit_more_flush(tx_ring);
249 		u64_stats_update_begin(&r_vec->tx_sync);
250 		r_vec->tx_busy++;
251 		u64_stats_update_end(&r_vec->tx_sync);
252 		return NETDEV_TX_BUSY;
253 	}
254 
255 	skb = nfp_net_tls_tx(dp, r_vec, skb, &tls_handle, &nr_frags);
256 	if (unlikely(!skb)) {
257 		nfp_net_tx_xmit_more_flush(tx_ring);
258 		return NETDEV_TX_OK;
259 	}
260 
261 	md_bytes = nfp_nfd3_prep_tx_meta(skb, tls_handle);
262 	if (unlikely(md_bytes < 0))
263 		goto err_flush;
264 
265 	/* Start with the head skbuf */
266 	dma_addr = dma_map_single(dp->dev, skb->data, skb_headlen(skb),
267 				  DMA_TO_DEVICE);
268 	if (dma_mapping_error(dp->dev, dma_addr))
269 		goto err_dma_err;
270 
271 	wr_idx = D_IDX(tx_ring, tx_ring->wr_p);
272 
273 	/* Stash the soft descriptor of the head then initialize it */
274 	txbuf = &tx_ring->txbufs[wr_idx];
275 	txbuf->skb = skb;
276 	txbuf->dma_addr = dma_addr;
277 	txbuf->fidx = -1;
278 	txbuf->pkt_cnt = 1;
279 	txbuf->real_len = skb->len;
280 
281 	/* Build TX descriptor */
282 	txd = &tx_ring->txds[wr_idx];
283 	txd->offset_eop = (nr_frags ? 0 : NFD3_DESC_TX_EOP) | md_bytes;
284 	txd->dma_len = cpu_to_le16(skb_headlen(skb));
285 	nfp_desc_set_dma_addr(txd, dma_addr);
286 	txd->data_len = cpu_to_le16(skb->len);
287 
288 	txd->flags = 0;
289 	txd->mss = 0;
290 	txd->lso_hdrlen = 0;
291 
292 	/* Do not reorder - tso may adjust pkt cnt, vlan may override fields */
293 	nfp_nfd3_tx_tso(r_vec, txbuf, txd, skb, md_bytes);
294 	nfp_nfd3_tx_csum(dp, r_vec, txbuf, txd, skb);
295 	if (skb_vlan_tag_present(skb) && dp->ctrl & NFP_NET_CFG_CTRL_TXVLAN) {
296 		txd->flags |= NFD3_DESC_TX_VLAN;
297 		txd->vlan = cpu_to_le16(skb_vlan_tag_get(skb));
298 	}
299 
300 	/* Gather DMA */
301 	if (nr_frags > 0) {
302 		__le64 second_half;
303 
304 		/* all descs must match except for in addr, length and eop */
305 		second_half = txd->vals8[1];
306 
307 		for (f = 0; f < nr_frags; f++) {
308 			frag = &skb_shinfo(skb)->frags[f];
309 			fsize = skb_frag_size(frag);
310 
311 			dma_addr = skb_frag_dma_map(dp->dev, frag, 0,
312 						    fsize, DMA_TO_DEVICE);
313 			if (dma_mapping_error(dp->dev, dma_addr))
314 				goto err_unmap;
315 
316 			wr_idx = D_IDX(tx_ring, wr_idx + 1);
317 			tx_ring->txbufs[wr_idx].skb = skb;
318 			tx_ring->txbufs[wr_idx].dma_addr = dma_addr;
319 			tx_ring->txbufs[wr_idx].fidx = f;
320 
321 			txd = &tx_ring->txds[wr_idx];
322 			txd->dma_len = cpu_to_le16(fsize);
323 			nfp_desc_set_dma_addr(txd, dma_addr);
324 			txd->offset_eop = md_bytes |
325 				((f == nr_frags - 1) ? NFD3_DESC_TX_EOP : 0);
326 			txd->vals8[1] = second_half;
327 		}
328 
329 		u64_stats_update_begin(&r_vec->tx_sync);
330 		r_vec->tx_gather++;
331 		u64_stats_update_end(&r_vec->tx_sync);
332 	}
333 
334 	skb_tx_timestamp(skb);
335 
336 	nd_q = netdev_get_tx_queue(dp->netdev, tx_ring->idx);
337 
338 	tx_ring->wr_p += nr_frags + 1;
339 	if (nfp_nfd3_tx_ring_should_stop(tx_ring))
340 		nfp_nfd3_tx_ring_stop(nd_q, tx_ring);
341 
342 	tx_ring->wr_ptr_add += nr_frags + 1;
343 	if (__netdev_tx_sent_queue(nd_q, txbuf->real_len, netdev_xmit_more()))
344 		nfp_net_tx_xmit_more_flush(tx_ring);
345 
346 	return NETDEV_TX_OK;
347 
348 err_unmap:
349 	while (--f >= 0) {
350 		frag = &skb_shinfo(skb)->frags[f];
351 		dma_unmap_page(dp->dev, tx_ring->txbufs[wr_idx].dma_addr,
352 			       skb_frag_size(frag), DMA_TO_DEVICE);
353 		tx_ring->txbufs[wr_idx].skb = NULL;
354 		tx_ring->txbufs[wr_idx].dma_addr = 0;
355 		tx_ring->txbufs[wr_idx].fidx = -2;
356 		wr_idx = wr_idx - 1;
357 		if (wr_idx < 0)
358 			wr_idx += tx_ring->cnt;
359 	}
360 	dma_unmap_single(dp->dev, tx_ring->txbufs[wr_idx].dma_addr,
361 			 skb_headlen(skb), DMA_TO_DEVICE);
362 	tx_ring->txbufs[wr_idx].skb = NULL;
363 	tx_ring->txbufs[wr_idx].dma_addr = 0;
364 	tx_ring->txbufs[wr_idx].fidx = -2;
365 err_dma_err:
366 	nn_dp_warn(dp, "Failed to map DMA TX buffer\n");
367 err_flush:
368 	nfp_net_tx_xmit_more_flush(tx_ring);
369 	u64_stats_update_begin(&r_vec->tx_sync);
370 	r_vec->tx_errors++;
371 	u64_stats_update_end(&r_vec->tx_sync);
372 	nfp_net_tls_tx_undo(skb, tls_handle);
373 	dev_kfree_skb_any(skb);
374 	return NETDEV_TX_OK;
375 }
376 
377 /**
378  * nfp_nfd3_tx_complete() - Handled completed TX packets
379  * @tx_ring:	TX ring structure
380  * @budget:	NAPI budget (only used as bool to determine if in NAPI context)
381  */
382 void nfp_nfd3_tx_complete(struct nfp_net_tx_ring *tx_ring, int budget)
383 {
384 	struct nfp_net_r_vector *r_vec = tx_ring->r_vec;
385 	struct nfp_net_dp *dp = &r_vec->nfp_net->dp;
386 	u32 done_pkts = 0, done_bytes = 0;
387 	struct netdev_queue *nd_q;
388 	u32 qcp_rd_p;
389 	int todo;
390 
391 	if (tx_ring->wr_p == tx_ring->rd_p)
392 		return;
393 
394 	/* Work out how many descriptors have been transmitted */
395 	qcp_rd_p = nfp_net_read_tx_cmpl(tx_ring, dp);
396 
397 	if (qcp_rd_p == tx_ring->qcp_rd_p)
398 		return;
399 
400 	todo = D_IDX(tx_ring, qcp_rd_p - tx_ring->qcp_rd_p);
401 
402 	while (todo--) {
403 		const skb_frag_t *frag;
404 		struct nfp_nfd3_tx_buf *tx_buf;
405 		struct sk_buff *skb;
406 		int fidx, nr_frags;
407 		int idx;
408 
409 		idx = D_IDX(tx_ring, tx_ring->rd_p++);
410 		tx_buf = &tx_ring->txbufs[idx];
411 
412 		skb = tx_buf->skb;
413 		if (!skb)
414 			continue;
415 
416 		nr_frags = skb_shinfo(skb)->nr_frags;
417 		fidx = tx_buf->fidx;
418 
419 		if (fidx == -1) {
420 			/* unmap head */
421 			dma_unmap_single(dp->dev, tx_buf->dma_addr,
422 					 skb_headlen(skb), DMA_TO_DEVICE);
423 
424 			done_pkts += tx_buf->pkt_cnt;
425 			done_bytes += tx_buf->real_len;
426 		} else {
427 			/* unmap fragment */
428 			frag = &skb_shinfo(skb)->frags[fidx];
429 			dma_unmap_page(dp->dev, tx_buf->dma_addr,
430 				       skb_frag_size(frag), DMA_TO_DEVICE);
431 		}
432 
433 		/* check for last gather fragment */
434 		if (fidx == nr_frags - 1)
435 			napi_consume_skb(skb, budget);
436 
437 		tx_buf->dma_addr = 0;
438 		tx_buf->skb = NULL;
439 		tx_buf->fidx = -2;
440 	}
441 
442 	tx_ring->qcp_rd_p = qcp_rd_p;
443 
444 	u64_stats_update_begin(&r_vec->tx_sync);
445 	r_vec->tx_bytes += done_bytes;
446 	r_vec->tx_pkts += done_pkts;
447 	u64_stats_update_end(&r_vec->tx_sync);
448 
449 	if (!dp->netdev)
450 		return;
451 
452 	nd_q = netdev_get_tx_queue(dp->netdev, tx_ring->idx);
453 	netdev_tx_completed_queue(nd_q, done_pkts, done_bytes);
454 	if (nfp_nfd3_tx_ring_should_wake(tx_ring)) {
455 		/* Make sure TX thread will see updated tx_ring->rd_p */
456 		smp_mb();
457 
458 		if (unlikely(netif_tx_queue_stopped(nd_q)))
459 			netif_tx_wake_queue(nd_q);
460 	}
461 
462 	WARN_ONCE(tx_ring->wr_p - tx_ring->rd_p > tx_ring->cnt,
463 		  "TX ring corruption rd_p=%u wr_p=%u cnt=%u\n",
464 		  tx_ring->rd_p, tx_ring->wr_p, tx_ring->cnt);
465 }
466 
467 static bool nfp_nfd3_xdp_complete(struct nfp_net_tx_ring *tx_ring)
468 {
469 	struct nfp_net_r_vector *r_vec = tx_ring->r_vec;
470 	struct nfp_net_dp *dp = &r_vec->nfp_net->dp;
471 	u32 done_pkts = 0, done_bytes = 0;
472 	bool done_all;
473 	int idx, todo;
474 	u32 qcp_rd_p;
475 
476 	/* Work out how many descriptors have been transmitted */
477 	qcp_rd_p = nfp_net_read_tx_cmpl(tx_ring, dp);
478 
479 	if (qcp_rd_p == tx_ring->qcp_rd_p)
480 		return true;
481 
482 	todo = D_IDX(tx_ring, qcp_rd_p - tx_ring->qcp_rd_p);
483 
484 	done_all = todo <= NFP_NET_XDP_MAX_COMPLETE;
485 	todo = min(todo, NFP_NET_XDP_MAX_COMPLETE);
486 
487 	tx_ring->qcp_rd_p = D_IDX(tx_ring, tx_ring->qcp_rd_p + todo);
488 
489 	done_pkts = todo;
490 	while (todo--) {
491 		idx = D_IDX(tx_ring, tx_ring->rd_p);
492 		tx_ring->rd_p++;
493 
494 		done_bytes += tx_ring->txbufs[idx].real_len;
495 	}
496 
497 	u64_stats_update_begin(&r_vec->tx_sync);
498 	r_vec->tx_bytes += done_bytes;
499 	r_vec->tx_pkts += done_pkts;
500 	u64_stats_update_end(&r_vec->tx_sync);
501 
502 	WARN_ONCE(tx_ring->wr_p - tx_ring->rd_p > tx_ring->cnt,
503 		  "XDP TX ring corruption rd_p=%u wr_p=%u cnt=%u\n",
504 		  tx_ring->rd_p, tx_ring->wr_p, tx_ring->cnt);
505 
506 	return done_all;
507 }
508 
509 /* Receive processing
510  */
511 
512 static void *
513 nfp_nfd3_napi_alloc_one(struct nfp_net_dp *dp, dma_addr_t *dma_addr)
514 {
515 	void *frag;
516 
517 	if (!dp->xdp_prog) {
518 		frag = napi_alloc_frag(dp->fl_bufsz);
519 		if (unlikely(!frag))
520 			return NULL;
521 	} else {
522 		struct page *page;
523 
524 		page = dev_alloc_page();
525 		if (unlikely(!page))
526 			return NULL;
527 		frag = page_address(page);
528 	}
529 
530 	*dma_addr = nfp_net_dma_map_rx(dp, frag);
531 	if (dma_mapping_error(dp->dev, *dma_addr)) {
532 		nfp_net_free_frag(frag, dp->xdp_prog);
533 		nn_dp_warn(dp, "Failed to map DMA RX buffer\n");
534 		return NULL;
535 	}
536 
537 	return frag;
538 }
539 
540 /**
541  * nfp_nfd3_rx_give_one() - Put mapped skb on the software and hardware rings
542  * @dp:		NFP Net data path struct
543  * @rx_ring:	RX ring structure
544  * @frag:	page fragment buffer
545  * @dma_addr:	DMA address of skb mapping
546  */
547 static void
548 nfp_nfd3_rx_give_one(const struct nfp_net_dp *dp,
549 		     struct nfp_net_rx_ring *rx_ring,
550 		     void *frag, dma_addr_t dma_addr)
551 {
552 	unsigned int wr_idx;
553 
554 	wr_idx = D_IDX(rx_ring, rx_ring->wr_p);
555 
556 	nfp_net_dma_sync_dev_rx(dp, dma_addr);
557 
558 	/* Stash SKB and DMA address away */
559 	rx_ring->rxbufs[wr_idx].frag = frag;
560 	rx_ring->rxbufs[wr_idx].dma_addr = dma_addr;
561 
562 	/* Fill freelist descriptor */
563 	rx_ring->rxds[wr_idx].fld.reserved = 0;
564 	rx_ring->rxds[wr_idx].fld.meta_len_dd = 0;
565 	nfp_desc_set_dma_addr(&rx_ring->rxds[wr_idx].fld,
566 			      dma_addr + dp->rx_dma_off);
567 
568 	rx_ring->wr_p++;
569 	if (!(rx_ring->wr_p % NFP_NET_FL_BATCH)) {
570 		/* Update write pointer of the freelist queue. Make
571 		 * sure all writes are flushed before telling the hardware.
572 		 */
573 		wmb();
574 		nfp_qcp_wr_ptr_add(rx_ring->qcp_fl, NFP_NET_FL_BATCH);
575 	}
576 }
577 
578 /**
579  * nfp_nfd3_rx_ring_fill_freelist() - Give buffers from the ring to FW
580  * @dp:	     NFP Net data path struct
581  * @rx_ring: RX ring to fill
582  */
583 void nfp_nfd3_rx_ring_fill_freelist(struct nfp_net_dp *dp,
584 				    struct nfp_net_rx_ring *rx_ring)
585 {
586 	unsigned int i;
587 
588 	if (nfp_net_has_xsk_pool_slow(dp, rx_ring->idx))
589 		return nfp_net_xsk_rx_ring_fill_freelist(rx_ring);
590 
591 	for (i = 0; i < rx_ring->cnt - 1; i++)
592 		nfp_nfd3_rx_give_one(dp, rx_ring, rx_ring->rxbufs[i].frag,
593 				     rx_ring->rxbufs[i].dma_addr);
594 }
595 
596 /**
597  * nfp_nfd3_rx_csum_has_errors() - group check if rxd has any csum errors
598  * @flags: RX descriptor flags field in CPU byte order
599  */
600 static int nfp_nfd3_rx_csum_has_errors(u16 flags)
601 {
602 	u16 csum_all_checked, csum_all_ok;
603 
604 	csum_all_checked = flags & __PCIE_DESC_RX_CSUM_ALL;
605 	csum_all_ok = flags & __PCIE_DESC_RX_CSUM_ALL_OK;
606 
607 	return csum_all_checked != (csum_all_ok << PCIE_DESC_RX_CSUM_OK_SHIFT);
608 }
609 
610 /**
611  * nfp_nfd3_rx_csum() - set SKB checksum field based on RX descriptor flags
612  * @dp:  NFP Net data path struct
613  * @r_vec: per-ring structure
614  * @rxd: Pointer to RX descriptor
615  * @meta: Parsed metadata prepend
616  * @skb: Pointer to SKB
617  */
618 void
619 nfp_nfd3_rx_csum(const struct nfp_net_dp *dp, struct nfp_net_r_vector *r_vec,
620 		 const struct nfp_net_rx_desc *rxd,
621 		 const struct nfp_meta_parsed *meta, struct sk_buff *skb)
622 {
623 	skb_checksum_none_assert(skb);
624 
625 	if (!(dp->netdev->features & NETIF_F_RXCSUM))
626 		return;
627 
628 	if (meta->csum_type) {
629 		skb->ip_summed = meta->csum_type;
630 		skb->csum = meta->csum;
631 		u64_stats_update_begin(&r_vec->rx_sync);
632 		r_vec->hw_csum_rx_complete++;
633 		u64_stats_update_end(&r_vec->rx_sync);
634 		return;
635 	}
636 
637 	if (nfp_nfd3_rx_csum_has_errors(le16_to_cpu(rxd->rxd.flags))) {
638 		u64_stats_update_begin(&r_vec->rx_sync);
639 		r_vec->hw_csum_rx_error++;
640 		u64_stats_update_end(&r_vec->rx_sync);
641 		return;
642 	}
643 
644 	/* Assume that the firmware will never report inner CSUM_OK unless outer
645 	 * L4 headers were successfully parsed. FW will always report zero UDP
646 	 * checksum as CSUM_OK.
647 	 */
648 	if (rxd->rxd.flags & PCIE_DESC_RX_TCP_CSUM_OK ||
649 	    rxd->rxd.flags & PCIE_DESC_RX_UDP_CSUM_OK) {
650 		__skb_incr_checksum_unnecessary(skb);
651 		u64_stats_update_begin(&r_vec->rx_sync);
652 		r_vec->hw_csum_rx_ok++;
653 		u64_stats_update_end(&r_vec->rx_sync);
654 	}
655 
656 	if (rxd->rxd.flags & PCIE_DESC_RX_I_TCP_CSUM_OK ||
657 	    rxd->rxd.flags & PCIE_DESC_RX_I_UDP_CSUM_OK) {
658 		__skb_incr_checksum_unnecessary(skb);
659 		u64_stats_update_begin(&r_vec->rx_sync);
660 		r_vec->hw_csum_rx_inner_ok++;
661 		u64_stats_update_end(&r_vec->rx_sync);
662 	}
663 }
664 
665 static void
666 nfp_nfd3_set_hash(struct net_device *netdev, struct nfp_meta_parsed *meta,
667 		  unsigned int type, __be32 *hash)
668 {
669 	if (!(netdev->features & NETIF_F_RXHASH))
670 		return;
671 
672 	switch (type) {
673 	case NFP_NET_RSS_IPV4:
674 	case NFP_NET_RSS_IPV6:
675 	case NFP_NET_RSS_IPV6_EX:
676 		meta->hash_type = PKT_HASH_TYPE_L3;
677 		break;
678 	default:
679 		meta->hash_type = PKT_HASH_TYPE_L4;
680 		break;
681 	}
682 
683 	meta->hash = get_unaligned_be32(hash);
684 }
685 
686 static void
687 nfp_nfd3_set_hash_desc(struct net_device *netdev, struct nfp_meta_parsed *meta,
688 		       void *data, struct nfp_net_rx_desc *rxd)
689 {
690 	struct nfp_net_rx_hash *rx_hash = data;
691 
692 	if (!(rxd->rxd.flags & PCIE_DESC_RX_RSS))
693 		return;
694 
695 	nfp_nfd3_set_hash(netdev, meta, get_unaligned_be32(&rx_hash->hash_type),
696 			  &rx_hash->hash);
697 }
698 
699 bool
700 nfp_nfd3_parse_meta(struct net_device *netdev, struct nfp_meta_parsed *meta,
701 		    void *data, void *pkt, unsigned int pkt_len, int meta_len)
702 {
703 	u32 meta_info;
704 
705 	meta_info = get_unaligned_be32(data);
706 	data += 4;
707 
708 	while (meta_info) {
709 		switch (meta_info & NFP_NET_META_FIELD_MASK) {
710 		case NFP_NET_META_HASH:
711 			meta_info >>= NFP_NET_META_FIELD_SIZE;
712 			nfp_nfd3_set_hash(netdev, meta,
713 					  meta_info & NFP_NET_META_FIELD_MASK,
714 					  (__be32 *)data);
715 			data += 4;
716 			break;
717 		case NFP_NET_META_MARK:
718 			meta->mark = get_unaligned_be32(data);
719 			data += 4;
720 			break;
721 		case NFP_NET_META_PORTID:
722 			meta->portid = get_unaligned_be32(data);
723 			data += 4;
724 			break;
725 		case NFP_NET_META_CSUM:
726 			meta->csum_type = CHECKSUM_COMPLETE;
727 			meta->csum =
728 				(__force __wsum)__get_unaligned_cpu32(data);
729 			data += 4;
730 			break;
731 		case NFP_NET_META_RESYNC_INFO:
732 			if (nfp_net_tls_rx_resync_req(netdev, data, pkt,
733 						      pkt_len))
734 				return false;
735 			data += sizeof(struct nfp_net_tls_resync_req);
736 			break;
737 		default:
738 			return true;
739 		}
740 
741 		meta_info >>= NFP_NET_META_FIELD_SIZE;
742 	}
743 
744 	return data != pkt;
745 }
746 
747 static void
748 nfp_nfd3_rx_drop(const struct nfp_net_dp *dp, struct nfp_net_r_vector *r_vec,
749 		 struct nfp_net_rx_ring *rx_ring, struct nfp_net_rx_buf *rxbuf,
750 		 struct sk_buff *skb)
751 {
752 	u64_stats_update_begin(&r_vec->rx_sync);
753 	r_vec->rx_drops++;
754 	/* If we have both skb and rxbuf the replacement buffer allocation
755 	 * must have failed, count this as an alloc failure.
756 	 */
757 	if (skb && rxbuf)
758 		r_vec->rx_replace_buf_alloc_fail++;
759 	u64_stats_update_end(&r_vec->rx_sync);
760 
761 	/* skb is build based on the frag, free_skb() would free the frag
762 	 * so to be able to reuse it we need an extra ref.
763 	 */
764 	if (skb && rxbuf && skb->head == rxbuf->frag)
765 		page_ref_inc(virt_to_head_page(rxbuf->frag));
766 	if (rxbuf)
767 		nfp_nfd3_rx_give_one(dp, rx_ring, rxbuf->frag, rxbuf->dma_addr);
768 	if (skb)
769 		dev_kfree_skb_any(skb);
770 }
771 
772 static bool
773 nfp_nfd3_tx_xdp_buf(struct nfp_net_dp *dp, struct nfp_net_rx_ring *rx_ring,
774 		    struct nfp_net_tx_ring *tx_ring,
775 		    struct nfp_net_rx_buf *rxbuf, unsigned int dma_off,
776 		    unsigned int pkt_len, bool *completed)
777 {
778 	unsigned int dma_map_sz = dp->fl_bufsz - NFP_NET_RX_BUF_NON_DATA;
779 	struct nfp_nfd3_tx_buf *txbuf;
780 	struct nfp_nfd3_tx_desc *txd;
781 	int wr_idx;
782 
783 	/* Reject if xdp_adjust_tail grow packet beyond DMA area */
784 	if (pkt_len + dma_off > dma_map_sz)
785 		return false;
786 
787 	if (unlikely(nfp_net_tx_full(tx_ring, 1))) {
788 		if (!*completed) {
789 			nfp_nfd3_xdp_complete(tx_ring);
790 			*completed = true;
791 		}
792 
793 		if (unlikely(nfp_net_tx_full(tx_ring, 1))) {
794 			nfp_nfd3_rx_drop(dp, rx_ring->r_vec, rx_ring, rxbuf,
795 					 NULL);
796 			return false;
797 		}
798 	}
799 
800 	wr_idx = D_IDX(tx_ring, tx_ring->wr_p);
801 
802 	/* Stash the soft descriptor of the head then initialize it */
803 	txbuf = &tx_ring->txbufs[wr_idx];
804 
805 	nfp_nfd3_rx_give_one(dp, rx_ring, txbuf->frag, txbuf->dma_addr);
806 
807 	txbuf->frag = rxbuf->frag;
808 	txbuf->dma_addr = rxbuf->dma_addr;
809 	txbuf->fidx = -1;
810 	txbuf->pkt_cnt = 1;
811 	txbuf->real_len = pkt_len;
812 
813 	dma_sync_single_for_device(dp->dev, rxbuf->dma_addr + dma_off,
814 				   pkt_len, DMA_BIDIRECTIONAL);
815 
816 	/* Build TX descriptor */
817 	txd = &tx_ring->txds[wr_idx];
818 	txd->offset_eop = NFD3_DESC_TX_EOP;
819 	txd->dma_len = cpu_to_le16(pkt_len);
820 	nfp_desc_set_dma_addr(txd, rxbuf->dma_addr + dma_off);
821 	txd->data_len = cpu_to_le16(pkt_len);
822 
823 	txd->flags = 0;
824 	txd->mss = 0;
825 	txd->lso_hdrlen = 0;
826 
827 	tx_ring->wr_p++;
828 	tx_ring->wr_ptr_add++;
829 	return true;
830 }
831 
832 /**
833  * nfp_nfd3_rx() - receive up to @budget packets on @rx_ring
834  * @rx_ring:   RX ring to receive from
835  * @budget:    NAPI budget
836  *
837  * Note, this function is separated out from the napi poll function to
838  * more cleanly separate packet receive code from other bookkeeping
839  * functions performed in the napi poll function.
840  *
841  * Return: Number of packets received.
842  */
843 static int nfp_nfd3_rx(struct nfp_net_rx_ring *rx_ring, int budget)
844 {
845 	struct nfp_net_r_vector *r_vec = rx_ring->r_vec;
846 	struct nfp_net_dp *dp = &r_vec->nfp_net->dp;
847 	struct nfp_net_tx_ring *tx_ring;
848 	struct bpf_prog *xdp_prog;
849 	bool xdp_tx_cmpl = false;
850 	unsigned int true_bufsz;
851 	struct sk_buff *skb;
852 	int pkts_polled = 0;
853 	struct xdp_buff xdp;
854 	int idx;
855 
856 	xdp_prog = READ_ONCE(dp->xdp_prog);
857 	true_bufsz = xdp_prog ? PAGE_SIZE : dp->fl_bufsz;
858 	xdp_init_buff(&xdp, PAGE_SIZE - NFP_NET_RX_BUF_HEADROOM,
859 		      &rx_ring->xdp_rxq);
860 	tx_ring = r_vec->xdp_ring;
861 
862 	while (pkts_polled < budget) {
863 		unsigned int meta_len, data_len, meta_off, pkt_len, pkt_off;
864 		struct nfp_net_rx_buf *rxbuf;
865 		struct nfp_net_rx_desc *rxd;
866 		struct nfp_meta_parsed meta;
867 		bool redir_egress = false;
868 		struct net_device *netdev;
869 		dma_addr_t new_dma_addr;
870 		u32 meta_len_xdp = 0;
871 		void *new_frag;
872 
873 		idx = D_IDX(rx_ring, rx_ring->rd_p);
874 
875 		rxd = &rx_ring->rxds[idx];
876 		if (!(rxd->rxd.meta_len_dd & PCIE_DESC_RX_DD))
877 			break;
878 
879 		/* Memory barrier to ensure that we won't do other reads
880 		 * before the DD bit.
881 		 */
882 		dma_rmb();
883 
884 		memset(&meta, 0, sizeof(meta));
885 
886 		rx_ring->rd_p++;
887 		pkts_polled++;
888 
889 		rxbuf =	&rx_ring->rxbufs[idx];
890 		/*         < meta_len >
891 		 *  <-- [rx_offset] -->
892 		 *  ---------------------------------------------------------
893 		 * | [XX] |  metadata  |             packet           | XXXX |
894 		 *  ---------------------------------------------------------
895 		 *         <---------------- data_len --------------->
896 		 *
897 		 * The rx_offset is fixed for all packets, the meta_len can vary
898 		 * on a packet by packet basis. If rx_offset is set to zero
899 		 * (_RX_OFFSET_DYNAMIC) metadata starts at the beginning of the
900 		 * buffer and is immediately followed by the packet (no [XX]).
901 		 */
902 		meta_len = rxd->rxd.meta_len_dd & PCIE_DESC_RX_META_LEN_MASK;
903 		data_len = le16_to_cpu(rxd->rxd.data_len);
904 		pkt_len = data_len - meta_len;
905 
906 		pkt_off = NFP_NET_RX_BUF_HEADROOM + dp->rx_dma_off;
907 		if (dp->rx_offset == NFP_NET_CFG_RX_OFFSET_DYNAMIC)
908 			pkt_off += meta_len;
909 		else
910 			pkt_off += dp->rx_offset;
911 		meta_off = pkt_off - meta_len;
912 
913 		/* Stats update */
914 		u64_stats_update_begin(&r_vec->rx_sync);
915 		r_vec->rx_pkts++;
916 		r_vec->rx_bytes += pkt_len;
917 		u64_stats_update_end(&r_vec->rx_sync);
918 
919 		if (unlikely(meta_len > NFP_NET_MAX_PREPEND ||
920 			     (dp->rx_offset && meta_len > dp->rx_offset))) {
921 			nn_dp_warn(dp, "oversized RX packet metadata %u\n",
922 				   meta_len);
923 			nfp_nfd3_rx_drop(dp, r_vec, rx_ring, rxbuf, NULL);
924 			continue;
925 		}
926 
927 		nfp_net_dma_sync_cpu_rx(dp, rxbuf->dma_addr + meta_off,
928 					data_len);
929 
930 		if (!dp->chained_metadata_format) {
931 			nfp_nfd3_set_hash_desc(dp->netdev, &meta,
932 					       rxbuf->frag + meta_off, rxd);
933 		} else if (meta_len) {
934 			if (unlikely(nfp_nfd3_parse_meta(dp->netdev, &meta,
935 							 rxbuf->frag + meta_off,
936 							 rxbuf->frag + pkt_off,
937 							 pkt_len, meta_len))) {
938 				nn_dp_warn(dp, "invalid RX packet metadata\n");
939 				nfp_nfd3_rx_drop(dp, r_vec, rx_ring, rxbuf,
940 						 NULL);
941 				continue;
942 			}
943 		}
944 
945 		if (xdp_prog && !meta.portid) {
946 			void *orig_data = rxbuf->frag + pkt_off;
947 			unsigned int dma_off;
948 			int act;
949 
950 			xdp_prepare_buff(&xdp,
951 					 rxbuf->frag + NFP_NET_RX_BUF_HEADROOM,
952 					 pkt_off - NFP_NET_RX_BUF_HEADROOM,
953 					 pkt_len, true);
954 
955 			act = bpf_prog_run_xdp(xdp_prog, &xdp);
956 
957 			pkt_len = xdp.data_end - xdp.data;
958 			pkt_off += xdp.data - orig_data;
959 
960 			switch (act) {
961 			case XDP_PASS:
962 				meta_len_xdp = xdp.data - xdp.data_meta;
963 				break;
964 			case XDP_TX:
965 				dma_off = pkt_off - NFP_NET_RX_BUF_HEADROOM;
966 				if (unlikely(!nfp_nfd3_tx_xdp_buf(dp, rx_ring,
967 								  tx_ring,
968 								  rxbuf,
969 								  dma_off,
970 								  pkt_len,
971 								  &xdp_tx_cmpl)))
972 					trace_xdp_exception(dp->netdev,
973 							    xdp_prog, act);
974 				continue;
975 			default:
976 				bpf_warn_invalid_xdp_action(dp->netdev, xdp_prog, act);
977 				fallthrough;
978 			case XDP_ABORTED:
979 				trace_xdp_exception(dp->netdev, xdp_prog, act);
980 				fallthrough;
981 			case XDP_DROP:
982 				nfp_nfd3_rx_give_one(dp, rx_ring, rxbuf->frag,
983 						     rxbuf->dma_addr);
984 				continue;
985 			}
986 		}
987 
988 		if (likely(!meta.portid)) {
989 			netdev = dp->netdev;
990 		} else if (meta.portid == NFP_META_PORT_ID_CTRL) {
991 			struct nfp_net *nn = netdev_priv(dp->netdev);
992 
993 			nfp_app_ctrl_rx_raw(nn->app, rxbuf->frag + pkt_off,
994 					    pkt_len);
995 			nfp_nfd3_rx_give_one(dp, rx_ring, rxbuf->frag,
996 					     rxbuf->dma_addr);
997 			continue;
998 		} else {
999 			struct nfp_net *nn;
1000 
1001 			nn = netdev_priv(dp->netdev);
1002 			netdev = nfp_app_dev_get(nn->app, meta.portid,
1003 						 &redir_egress);
1004 			if (unlikely(!netdev)) {
1005 				nfp_nfd3_rx_drop(dp, r_vec, rx_ring, rxbuf,
1006 						 NULL);
1007 				continue;
1008 			}
1009 
1010 			if (nfp_netdev_is_nfp_repr(netdev))
1011 				nfp_repr_inc_rx_stats(netdev, pkt_len);
1012 		}
1013 
1014 		skb = build_skb(rxbuf->frag, true_bufsz);
1015 		if (unlikely(!skb)) {
1016 			nfp_nfd3_rx_drop(dp, r_vec, rx_ring, rxbuf, NULL);
1017 			continue;
1018 		}
1019 		new_frag = nfp_nfd3_napi_alloc_one(dp, &new_dma_addr);
1020 		if (unlikely(!new_frag)) {
1021 			nfp_nfd3_rx_drop(dp, r_vec, rx_ring, rxbuf, skb);
1022 			continue;
1023 		}
1024 
1025 		nfp_net_dma_unmap_rx(dp, rxbuf->dma_addr);
1026 
1027 		nfp_nfd3_rx_give_one(dp, rx_ring, new_frag, new_dma_addr);
1028 
1029 		skb_reserve(skb, pkt_off);
1030 		skb_put(skb, pkt_len);
1031 
1032 		skb->mark = meta.mark;
1033 		skb_set_hash(skb, meta.hash, meta.hash_type);
1034 
1035 		skb_record_rx_queue(skb, rx_ring->idx);
1036 		skb->protocol = eth_type_trans(skb, netdev);
1037 
1038 		nfp_nfd3_rx_csum(dp, r_vec, rxd, &meta, skb);
1039 
1040 #ifdef CONFIG_TLS_DEVICE
1041 		if (rxd->rxd.flags & PCIE_DESC_RX_DECRYPTED) {
1042 			skb->decrypted = true;
1043 			u64_stats_update_begin(&r_vec->rx_sync);
1044 			r_vec->hw_tls_rx++;
1045 			u64_stats_update_end(&r_vec->rx_sync);
1046 		}
1047 #endif
1048 
1049 		if (rxd->rxd.flags & PCIE_DESC_RX_VLAN)
1050 			__vlan_hwaccel_put_tag(skb, htons(ETH_P_8021Q),
1051 					       le16_to_cpu(rxd->rxd.vlan));
1052 		if (meta_len_xdp)
1053 			skb_metadata_set(skb, meta_len_xdp);
1054 
1055 		if (likely(!redir_egress)) {
1056 			napi_gro_receive(&rx_ring->r_vec->napi, skb);
1057 		} else {
1058 			skb->dev = netdev;
1059 			skb_reset_network_header(skb);
1060 			__skb_push(skb, ETH_HLEN);
1061 			dev_queue_xmit(skb);
1062 		}
1063 	}
1064 
1065 	if (xdp_prog) {
1066 		if (tx_ring->wr_ptr_add)
1067 			nfp_net_tx_xmit_more_flush(tx_ring);
1068 		else if (unlikely(tx_ring->wr_p != tx_ring->rd_p) &&
1069 			 !xdp_tx_cmpl)
1070 			if (!nfp_nfd3_xdp_complete(tx_ring))
1071 				pkts_polled = budget;
1072 	}
1073 
1074 	return pkts_polled;
1075 }
1076 
1077 /**
1078  * nfp_nfd3_poll() - napi poll function
1079  * @napi:    NAPI structure
1080  * @budget:  NAPI budget
1081  *
1082  * Return: number of packets polled.
1083  */
1084 int nfp_nfd3_poll(struct napi_struct *napi, int budget)
1085 {
1086 	struct nfp_net_r_vector *r_vec =
1087 		container_of(napi, struct nfp_net_r_vector, napi);
1088 	unsigned int pkts_polled = 0;
1089 
1090 	if (r_vec->tx_ring)
1091 		nfp_nfd3_tx_complete(r_vec->tx_ring, budget);
1092 	if (r_vec->rx_ring)
1093 		pkts_polled = nfp_nfd3_rx(r_vec->rx_ring, budget);
1094 
1095 	if (pkts_polled < budget)
1096 		if (napi_complete_done(napi, pkts_polled))
1097 			nfp_net_irq_unmask(r_vec->nfp_net, r_vec->irq_entry);
1098 
1099 	if (r_vec->nfp_net->rx_coalesce_adapt_on && r_vec->rx_ring) {
1100 		struct dim_sample dim_sample = {};
1101 		unsigned int start;
1102 		u64 pkts, bytes;
1103 
1104 		do {
1105 			start = u64_stats_fetch_begin(&r_vec->rx_sync);
1106 			pkts = r_vec->rx_pkts;
1107 			bytes = r_vec->rx_bytes;
1108 		} while (u64_stats_fetch_retry(&r_vec->rx_sync, start));
1109 
1110 		dim_update_sample(r_vec->event_ctr, pkts, bytes, &dim_sample);
1111 		net_dim(&r_vec->rx_dim, dim_sample);
1112 	}
1113 
1114 	if (r_vec->nfp_net->tx_coalesce_adapt_on && r_vec->tx_ring) {
1115 		struct dim_sample dim_sample = {};
1116 		unsigned int start;
1117 		u64 pkts, bytes;
1118 
1119 		do {
1120 			start = u64_stats_fetch_begin(&r_vec->tx_sync);
1121 			pkts = r_vec->tx_pkts;
1122 			bytes = r_vec->tx_bytes;
1123 		} while (u64_stats_fetch_retry(&r_vec->tx_sync, start));
1124 
1125 		dim_update_sample(r_vec->event_ctr, pkts, bytes, &dim_sample);
1126 		net_dim(&r_vec->tx_dim, dim_sample);
1127 	}
1128 
1129 	return pkts_polled;
1130 }
1131 
1132 /* Control device data path
1133  */
1134 
1135 bool
1136 nfp_nfd3_ctrl_tx_one(struct nfp_net *nn, struct nfp_net_r_vector *r_vec,
1137 		     struct sk_buff *skb, bool old)
1138 {
1139 	unsigned int real_len = skb->len, meta_len = 0;
1140 	struct nfp_net_tx_ring *tx_ring;
1141 	struct nfp_nfd3_tx_buf *txbuf;
1142 	struct nfp_nfd3_tx_desc *txd;
1143 	struct nfp_net_dp *dp;
1144 	dma_addr_t dma_addr;
1145 	int wr_idx;
1146 
1147 	dp = &r_vec->nfp_net->dp;
1148 	tx_ring = r_vec->tx_ring;
1149 
1150 	if (WARN_ON_ONCE(skb_shinfo(skb)->nr_frags)) {
1151 		nn_dp_warn(dp, "Driver's CTRL TX does not implement gather\n");
1152 		goto err_free;
1153 	}
1154 
1155 	if (unlikely(nfp_net_tx_full(tx_ring, 1))) {
1156 		u64_stats_update_begin(&r_vec->tx_sync);
1157 		r_vec->tx_busy++;
1158 		u64_stats_update_end(&r_vec->tx_sync);
1159 		if (!old)
1160 			__skb_queue_tail(&r_vec->queue, skb);
1161 		else
1162 			__skb_queue_head(&r_vec->queue, skb);
1163 		return true;
1164 	}
1165 
1166 	if (nfp_app_ctrl_has_meta(nn->app)) {
1167 		if (unlikely(skb_headroom(skb) < 8)) {
1168 			nn_dp_warn(dp, "CTRL TX on skb without headroom\n");
1169 			goto err_free;
1170 		}
1171 		meta_len = 8;
1172 		put_unaligned_be32(NFP_META_PORT_ID_CTRL, skb_push(skb, 4));
1173 		put_unaligned_be32(NFP_NET_META_PORTID, skb_push(skb, 4));
1174 	}
1175 
1176 	/* Start with the head skbuf */
1177 	dma_addr = dma_map_single(dp->dev, skb->data, skb_headlen(skb),
1178 				  DMA_TO_DEVICE);
1179 	if (dma_mapping_error(dp->dev, dma_addr))
1180 		goto err_dma_warn;
1181 
1182 	wr_idx = D_IDX(tx_ring, tx_ring->wr_p);
1183 
1184 	/* Stash the soft descriptor of the head then initialize it */
1185 	txbuf = &tx_ring->txbufs[wr_idx];
1186 	txbuf->skb = skb;
1187 	txbuf->dma_addr = dma_addr;
1188 	txbuf->fidx = -1;
1189 	txbuf->pkt_cnt = 1;
1190 	txbuf->real_len = real_len;
1191 
1192 	/* Build TX descriptor */
1193 	txd = &tx_ring->txds[wr_idx];
1194 	txd->offset_eop = meta_len | NFD3_DESC_TX_EOP;
1195 	txd->dma_len = cpu_to_le16(skb_headlen(skb));
1196 	nfp_desc_set_dma_addr(txd, dma_addr);
1197 	txd->data_len = cpu_to_le16(skb->len);
1198 
1199 	txd->flags = 0;
1200 	txd->mss = 0;
1201 	txd->lso_hdrlen = 0;
1202 
1203 	tx_ring->wr_p++;
1204 	tx_ring->wr_ptr_add++;
1205 	nfp_net_tx_xmit_more_flush(tx_ring);
1206 
1207 	return false;
1208 
1209 err_dma_warn:
1210 	nn_dp_warn(dp, "Failed to DMA map TX CTRL buffer\n");
1211 err_free:
1212 	u64_stats_update_begin(&r_vec->tx_sync);
1213 	r_vec->tx_errors++;
1214 	u64_stats_update_end(&r_vec->tx_sync);
1215 	dev_kfree_skb_any(skb);
1216 	return false;
1217 }
1218 
1219 static void __nfp_ctrl_tx_queued(struct nfp_net_r_vector *r_vec)
1220 {
1221 	struct sk_buff *skb;
1222 
1223 	while ((skb = __skb_dequeue(&r_vec->queue)))
1224 		if (nfp_nfd3_ctrl_tx_one(r_vec->nfp_net, r_vec, skb, true))
1225 			return;
1226 }
1227 
1228 static bool
1229 nfp_ctrl_meta_ok(struct nfp_net *nn, void *data, unsigned int meta_len)
1230 {
1231 	u32 meta_type, meta_tag;
1232 
1233 	if (!nfp_app_ctrl_has_meta(nn->app))
1234 		return !meta_len;
1235 
1236 	if (meta_len != 8)
1237 		return false;
1238 
1239 	meta_type = get_unaligned_be32(data);
1240 	meta_tag = get_unaligned_be32(data + 4);
1241 
1242 	return (meta_type == NFP_NET_META_PORTID &&
1243 		meta_tag == NFP_META_PORT_ID_CTRL);
1244 }
1245 
1246 static bool
1247 nfp_ctrl_rx_one(struct nfp_net *nn, struct nfp_net_dp *dp,
1248 		struct nfp_net_r_vector *r_vec, struct nfp_net_rx_ring *rx_ring)
1249 {
1250 	unsigned int meta_len, data_len, meta_off, pkt_len, pkt_off;
1251 	struct nfp_net_rx_buf *rxbuf;
1252 	struct nfp_net_rx_desc *rxd;
1253 	dma_addr_t new_dma_addr;
1254 	struct sk_buff *skb;
1255 	void *new_frag;
1256 	int idx;
1257 
1258 	idx = D_IDX(rx_ring, rx_ring->rd_p);
1259 
1260 	rxd = &rx_ring->rxds[idx];
1261 	if (!(rxd->rxd.meta_len_dd & PCIE_DESC_RX_DD))
1262 		return false;
1263 
1264 	/* Memory barrier to ensure that we won't do other reads
1265 	 * before the DD bit.
1266 	 */
1267 	dma_rmb();
1268 
1269 	rx_ring->rd_p++;
1270 
1271 	rxbuf =	&rx_ring->rxbufs[idx];
1272 	meta_len = rxd->rxd.meta_len_dd & PCIE_DESC_RX_META_LEN_MASK;
1273 	data_len = le16_to_cpu(rxd->rxd.data_len);
1274 	pkt_len = data_len - meta_len;
1275 
1276 	pkt_off = NFP_NET_RX_BUF_HEADROOM + dp->rx_dma_off;
1277 	if (dp->rx_offset == NFP_NET_CFG_RX_OFFSET_DYNAMIC)
1278 		pkt_off += meta_len;
1279 	else
1280 		pkt_off += dp->rx_offset;
1281 	meta_off = pkt_off - meta_len;
1282 
1283 	/* Stats update */
1284 	u64_stats_update_begin(&r_vec->rx_sync);
1285 	r_vec->rx_pkts++;
1286 	r_vec->rx_bytes += pkt_len;
1287 	u64_stats_update_end(&r_vec->rx_sync);
1288 
1289 	nfp_net_dma_sync_cpu_rx(dp, rxbuf->dma_addr + meta_off,	data_len);
1290 
1291 	if (unlikely(!nfp_ctrl_meta_ok(nn, rxbuf->frag + meta_off, meta_len))) {
1292 		nn_dp_warn(dp, "incorrect metadata for ctrl packet (%d)\n",
1293 			   meta_len);
1294 		nfp_nfd3_rx_drop(dp, r_vec, rx_ring, rxbuf, NULL);
1295 		return true;
1296 	}
1297 
1298 	skb = build_skb(rxbuf->frag, dp->fl_bufsz);
1299 	if (unlikely(!skb)) {
1300 		nfp_nfd3_rx_drop(dp, r_vec, rx_ring, rxbuf, NULL);
1301 		return true;
1302 	}
1303 	new_frag = nfp_nfd3_napi_alloc_one(dp, &new_dma_addr);
1304 	if (unlikely(!new_frag)) {
1305 		nfp_nfd3_rx_drop(dp, r_vec, rx_ring, rxbuf, skb);
1306 		return true;
1307 	}
1308 
1309 	nfp_net_dma_unmap_rx(dp, rxbuf->dma_addr);
1310 
1311 	nfp_nfd3_rx_give_one(dp, rx_ring, new_frag, new_dma_addr);
1312 
1313 	skb_reserve(skb, pkt_off);
1314 	skb_put(skb, pkt_len);
1315 
1316 	nfp_app_ctrl_rx(nn->app, skb);
1317 
1318 	return true;
1319 }
1320 
1321 static bool nfp_ctrl_rx(struct nfp_net_r_vector *r_vec)
1322 {
1323 	struct nfp_net_rx_ring *rx_ring = r_vec->rx_ring;
1324 	struct nfp_net *nn = r_vec->nfp_net;
1325 	struct nfp_net_dp *dp = &nn->dp;
1326 	unsigned int budget = 512;
1327 
1328 	while (nfp_ctrl_rx_one(nn, dp, r_vec, rx_ring) && budget--)
1329 		continue;
1330 
1331 	return budget;
1332 }
1333 
1334 void nfp_nfd3_ctrl_poll(struct tasklet_struct *t)
1335 {
1336 	struct nfp_net_r_vector *r_vec = from_tasklet(r_vec, t, tasklet);
1337 
1338 	spin_lock(&r_vec->lock);
1339 	nfp_nfd3_tx_complete(r_vec->tx_ring, 0);
1340 	__nfp_ctrl_tx_queued(r_vec);
1341 	spin_unlock(&r_vec->lock);
1342 
1343 	if (nfp_ctrl_rx(r_vec)) {
1344 		nfp_net_irq_unmask(r_vec->nfp_net, r_vec->irq_entry);
1345 	} else {
1346 		tasklet_schedule(&r_vec->tasklet);
1347 		nn_dp_warn(&r_vec->nfp_net->dp,
1348 			   "control message budget exceeded!\n");
1349 	}
1350 }
1351