1 // SPDX-License-Identifier: (GPL-2.0-only OR BSD-3-Clause)
2 /* QLogic qede NIC Driver
3  * Copyright (c) 2015-2017  QLogic Corporation
4  * Copyright (c) 2019-2020 Marvell International Ltd.
5  */
6 
7 #include <linux/netdevice.h>
8 #include <linux/etherdevice.h>
9 #include <linux/skbuff.h>
10 #include <linux/bpf_trace.h>
11 #include <net/udp_tunnel.h>
12 #include <linux/ip.h>
13 #include <net/ipv6.h>
14 #include <net/tcp.h>
15 #include <linux/if_ether.h>
16 #include <linux/if_vlan.h>
17 #include <net/ip6_checksum.h>
18 #include "qede_ptp.h"
19 
20 #include <linux/qed/qed_if.h>
21 #include "qede.h"
22 /*********************************
23  * Content also used by slowpath *
24  *********************************/
25 
26 int qede_alloc_rx_buffer(struct qede_rx_queue *rxq, bool allow_lazy)
27 {
28 	struct sw_rx_data *sw_rx_data;
29 	struct eth_rx_bd *rx_bd;
30 	dma_addr_t mapping;
31 	struct page *data;
32 
33 	/* In case lazy-allocation is allowed, postpone allocation until the
34 	 * end of the NAPI run. We'd still need to make sure the Rx ring has
35 	 * sufficient buffers to guarantee an additional Rx interrupt.
36 	 */
37 	if (allow_lazy && likely(rxq->filled_buffers > 12)) {
38 		rxq->filled_buffers--;
39 		return 0;
40 	}
41 
42 	data = alloc_pages(GFP_ATOMIC, 0);
43 	if (unlikely(!data))
44 		return -ENOMEM;
45 
46 	/* Map the entire page as it would be used
47 	 * for multiple RX buffer segment size mapping.
48 	 */
49 	mapping = dma_map_page(rxq->dev, data, 0,
50 			       PAGE_SIZE, rxq->data_direction);
51 	if (unlikely(dma_mapping_error(rxq->dev, mapping))) {
52 		__free_page(data);
53 		return -ENOMEM;
54 	}
55 
56 	sw_rx_data = &rxq->sw_rx_ring[rxq->sw_rx_prod & NUM_RX_BDS_MAX];
57 	sw_rx_data->page_offset = 0;
58 	sw_rx_data->data = data;
59 	sw_rx_data->mapping = mapping;
60 
61 	/* Advance PROD and get BD pointer */
62 	rx_bd = (struct eth_rx_bd *)qed_chain_produce(&rxq->rx_bd_ring);
63 	WARN_ON(!rx_bd);
64 	rx_bd->addr.hi = cpu_to_le32(upper_32_bits(mapping));
65 	rx_bd->addr.lo = cpu_to_le32(lower_32_bits(mapping) +
66 				     rxq->rx_headroom);
67 
68 	rxq->sw_rx_prod++;
69 	rxq->filled_buffers++;
70 
71 	return 0;
72 }
73 
74 /* Unmap the data and free skb */
75 int qede_free_tx_pkt(struct qede_dev *edev, struct qede_tx_queue *txq, int *len)
76 {
77 	u16 idx = txq->sw_tx_cons;
78 	struct sk_buff *skb = txq->sw_tx_ring.skbs[idx].skb;
79 	struct eth_tx_1st_bd *first_bd;
80 	struct eth_tx_bd *tx_data_bd;
81 	int bds_consumed = 0;
82 	int nbds;
83 	bool data_split = txq->sw_tx_ring.skbs[idx].flags & QEDE_TSO_SPLIT_BD;
84 	int i, split_bd_len = 0;
85 
86 	if (unlikely(!skb)) {
87 		DP_ERR(edev,
88 		       "skb is null for txq idx=%d txq->sw_tx_cons=%d txq->sw_tx_prod=%d\n",
89 		       idx, txq->sw_tx_cons, txq->sw_tx_prod);
90 		return -1;
91 	}
92 
93 	*len = skb->len;
94 
95 	first_bd = (struct eth_tx_1st_bd *)qed_chain_consume(&txq->tx_pbl);
96 
97 	bds_consumed++;
98 
99 	nbds = first_bd->data.nbds;
100 
101 	if (data_split) {
102 		struct eth_tx_bd *split = (struct eth_tx_bd *)
103 			qed_chain_consume(&txq->tx_pbl);
104 		split_bd_len = BD_UNMAP_LEN(split);
105 		bds_consumed++;
106 	}
107 	dma_unmap_single(&edev->pdev->dev, BD_UNMAP_ADDR(first_bd),
108 			 BD_UNMAP_LEN(first_bd) + split_bd_len, DMA_TO_DEVICE);
109 
110 	/* Unmap the data of the skb frags */
111 	for (i = 0; i < skb_shinfo(skb)->nr_frags; i++, bds_consumed++) {
112 		tx_data_bd = (struct eth_tx_bd *)
113 			qed_chain_consume(&txq->tx_pbl);
114 		dma_unmap_page(&edev->pdev->dev, BD_UNMAP_ADDR(tx_data_bd),
115 			       BD_UNMAP_LEN(tx_data_bd), DMA_TO_DEVICE);
116 	}
117 
118 	while (bds_consumed++ < nbds)
119 		qed_chain_consume(&txq->tx_pbl);
120 
121 	/* Free skb */
122 	dev_kfree_skb_any(skb);
123 	txq->sw_tx_ring.skbs[idx].skb = NULL;
124 	txq->sw_tx_ring.skbs[idx].flags = 0;
125 
126 	return 0;
127 }
128 
129 /* Unmap the data and free skb when mapping failed during start_xmit */
130 static void qede_free_failed_tx_pkt(struct qede_tx_queue *txq,
131 				    struct eth_tx_1st_bd *first_bd,
132 				    int nbd, bool data_split)
133 {
134 	u16 idx = txq->sw_tx_prod;
135 	struct sk_buff *skb = txq->sw_tx_ring.skbs[idx].skb;
136 	struct eth_tx_bd *tx_data_bd;
137 	int i, split_bd_len = 0;
138 
139 	/* Return prod to its position before this skb was handled */
140 	qed_chain_set_prod(&txq->tx_pbl,
141 			   le16_to_cpu(txq->tx_db.data.bd_prod), first_bd);
142 
143 	first_bd = (struct eth_tx_1st_bd *)qed_chain_produce(&txq->tx_pbl);
144 
145 	if (data_split) {
146 		struct eth_tx_bd *split = (struct eth_tx_bd *)
147 					  qed_chain_produce(&txq->tx_pbl);
148 		split_bd_len = BD_UNMAP_LEN(split);
149 		nbd--;
150 	}
151 
152 	dma_unmap_single(txq->dev, BD_UNMAP_ADDR(first_bd),
153 			 BD_UNMAP_LEN(first_bd) + split_bd_len, DMA_TO_DEVICE);
154 
155 	/* Unmap the data of the skb frags */
156 	for (i = 0; i < nbd; i++) {
157 		tx_data_bd = (struct eth_tx_bd *)
158 			qed_chain_produce(&txq->tx_pbl);
159 		if (tx_data_bd->nbytes)
160 			dma_unmap_page(txq->dev,
161 				       BD_UNMAP_ADDR(tx_data_bd),
162 				       BD_UNMAP_LEN(tx_data_bd), DMA_TO_DEVICE);
163 	}
164 
165 	/* Return again prod to its position before this skb was handled */
166 	qed_chain_set_prod(&txq->tx_pbl,
167 			   le16_to_cpu(txq->tx_db.data.bd_prod), first_bd);
168 
169 	/* Free skb */
170 	dev_kfree_skb_any(skb);
171 	txq->sw_tx_ring.skbs[idx].skb = NULL;
172 	txq->sw_tx_ring.skbs[idx].flags = 0;
173 }
174 
175 static u32 qede_xmit_type(struct sk_buff *skb, int *ipv6_ext)
176 {
177 	u32 rc = XMIT_L4_CSUM;
178 	__be16 l3_proto;
179 
180 	if (skb->ip_summed != CHECKSUM_PARTIAL)
181 		return XMIT_PLAIN;
182 
183 	l3_proto = vlan_get_protocol(skb);
184 	if (l3_proto == htons(ETH_P_IPV6) &&
185 	    (ipv6_hdr(skb)->nexthdr == NEXTHDR_IPV6))
186 		*ipv6_ext = 1;
187 
188 	if (skb->encapsulation) {
189 		rc |= XMIT_ENC;
190 		if (skb_is_gso(skb)) {
191 			unsigned short gso_type = skb_shinfo(skb)->gso_type;
192 
193 			if ((gso_type & SKB_GSO_UDP_TUNNEL_CSUM) ||
194 			    (gso_type & SKB_GSO_GRE_CSUM))
195 				rc |= XMIT_ENC_GSO_L4_CSUM;
196 
197 			rc |= XMIT_LSO;
198 			return rc;
199 		}
200 	}
201 
202 	if (skb_is_gso(skb))
203 		rc |= XMIT_LSO;
204 
205 	return rc;
206 }
207 
208 static void qede_set_params_for_ipv6_ext(struct sk_buff *skb,
209 					 struct eth_tx_2nd_bd *second_bd,
210 					 struct eth_tx_3rd_bd *third_bd)
211 {
212 	u8 l4_proto;
213 	u16 bd2_bits1 = 0, bd2_bits2 = 0;
214 
215 	bd2_bits1 |= (1 << ETH_TX_DATA_2ND_BD_IPV6_EXT_SHIFT);
216 
217 	bd2_bits2 |= ((((u8 *)skb_transport_header(skb) - skb->data) >> 1) &
218 		     ETH_TX_DATA_2ND_BD_L4_HDR_START_OFFSET_W_MASK)
219 		    << ETH_TX_DATA_2ND_BD_L4_HDR_START_OFFSET_W_SHIFT;
220 
221 	bd2_bits1 |= (ETH_L4_PSEUDO_CSUM_CORRECT_LENGTH <<
222 		      ETH_TX_DATA_2ND_BD_L4_PSEUDO_CSUM_MODE_SHIFT);
223 
224 	if (vlan_get_protocol(skb) == htons(ETH_P_IPV6))
225 		l4_proto = ipv6_hdr(skb)->nexthdr;
226 	else
227 		l4_proto = ip_hdr(skb)->protocol;
228 
229 	if (l4_proto == IPPROTO_UDP)
230 		bd2_bits1 |= 1 << ETH_TX_DATA_2ND_BD_L4_UDP_SHIFT;
231 
232 	if (third_bd)
233 		third_bd->data.bitfields |=
234 			cpu_to_le16(((tcp_hdrlen(skb) / 4) &
235 				ETH_TX_DATA_3RD_BD_TCP_HDR_LEN_DW_MASK) <<
236 				ETH_TX_DATA_3RD_BD_TCP_HDR_LEN_DW_SHIFT);
237 
238 	second_bd->data.bitfields1 = cpu_to_le16(bd2_bits1);
239 	second_bd->data.bitfields2 = cpu_to_le16(bd2_bits2);
240 }
241 
242 static int map_frag_to_bd(struct qede_tx_queue *txq,
243 			  skb_frag_t *frag, struct eth_tx_bd *bd)
244 {
245 	dma_addr_t mapping;
246 
247 	/* Map skb non-linear frag data for DMA */
248 	mapping = skb_frag_dma_map(txq->dev, frag, 0,
249 				   skb_frag_size(frag), DMA_TO_DEVICE);
250 	if (unlikely(dma_mapping_error(txq->dev, mapping)))
251 		return -ENOMEM;
252 
253 	/* Setup the data pointer of the frag data */
254 	BD_SET_UNMAP_ADDR_LEN(bd, mapping, skb_frag_size(frag));
255 
256 	return 0;
257 }
258 
259 static u16 qede_get_skb_hlen(struct sk_buff *skb, bool is_encap_pkt)
260 {
261 	if (is_encap_pkt)
262 		return (skb_inner_transport_header(skb) +
263 			inner_tcp_hdrlen(skb) - skb->data);
264 	else
265 		return (skb_transport_header(skb) +
266 			tcp_hdrlen(skb) - skb->data);
267 }
268 
269 /* +2 for 1st BD for headers and 2nd BD for headlen (if required) */
270 #if ((MAX_SKB_FRAGS + 2) > ETH_TX_MAX_BDS_PER_NON_LSO_PACKET)
271 static bool qede_pkt_req_lin(struct sk_buff *skb, u8 xmit_type)
272 {
273 	int allowed_frags = ETH_TX_MAX_BDS_PER_NON_LSO_PACKET - 1;
274 
275 	if (xmit_type & XMIT_LSO) {
276 		int hlen;
277 
278 		hlen = qede_get_skb_hlen(skb, xmit_type & XMIT_ENC);
279 
280 		/* linear payload would require its own BD */
281 		if (skb_headlen(skb) > hlen)
282 			allowed_frags--;
283 	}
284 
285 	return (skb_shinfo(skb)->nr_frags > allowed_frags);
286 }
287 #endif
288 
289 static inline void qede_update_tx_producer(struct qede_tx_queue *txq)
290 {
291 	/* wmb makes sure that the BDs data is updated before updating the
292 	 * producer, otherwise FW may read old data from the BDs.
293 	 */
294 	wmb();
295 	barrier();
296 	writel(txq->tx_db.raw, txq->doorbell_addr);
297 
298 	/* Fence required to flush the write combined buffer, since another
299 	 * CPU may write to the same doorbell address and data may be lost
300 	 * due to relaxed order nature of write combined bar.
301 	 */
302 	wmb();
303 }
304 
305 static int qede_xdp_xmit(struct qede_tx_queue *txq, dma_addr_t dma, u16 pad,
306 			 u16 len, struct page *page, struct xdp_frame *xdpf)
307 {
308 	struct eth_tx_1st_bd *bd;
309 	struct sw_tx_xdp *xdp;
310 	u16 val;
311 
312 	if (unlikely(qed_chain_get_elem_used(&txq->tx_pbl) >=
313 		     txq->num_tx_buffers)) {
314 		txq->stopped_cnt++;
315 		return -ENOMEM;
316 	}
317 
318 	bd = qed_chain_produce(&txq->tx_pbl);
319 	bd->data.nbds = 1;
320 	bd->data.bd_flags.bitfields = BIT(ETH_TX_1ST_BD_FLAGS_START_BD_SHIFT);
321 
322 	val = (len & ETH_TX_DATA_1ST_BD_PKT_LEN_MASK) <<
323 	       ETH_TX_DATA_1ST_BD_PKT_LEN_SHIFT;
324 
325 	bd->data.bitfields = cpu_to_le16(val);
326 
327 	/* We can safely ignore the offset, as it's 0 for XDP */
328 	BD_SET_UNMAP_ADDR_LEN(bd, dma + pad, len);
329 
330 	xdp = txq->sw_tx_ring.xdp + txq->sw_tx_prod;
331 	xdp->mapping = dma;
332 	xdp->page = page;
333 	xdp->xdpf = xdpf;
334 
335 	txq->sw_tx_prod = (txq->sw_tx_prod + 1) % txq->num_tx_buffers;
336 
337 	return 0;
338 }
339 
340 int qede_xdp_transmit(struct net_device *dev, int n_frames,
341 		      struct xdp_frame **frames, u32 flags)
342 {
343 	struct qede_dev *edev = netdev_priv(dev);
344 	struct device *dmadev = &edev->pdev->dev;
345 	struct qede_tx_queue *xdp_tx;
346 	struct xdp_frame *xdpf;
347 	dma_addr_t mapping;
348 	int i, nxmit = 0;
349 	u16 xdp_prod;
350 
351 	if (unlikely(flags & ~XDP_XMIT_FLAGS_MASK))
352 		return -EINVAL;
353 
354 	if (unlikely(!netif_running(dev)))
355 		return -ENETDOWN;
356 
357 	i = smp_processor_id() % edev->total_xdp_queues;
358 	xdp_tx = edev->fp_array[i].xdp_tx;
359 
360 	spin_lock(&xdp_tx->xdp_tx_lock);
361 
362 	for (i = 0; i < n_frames; i++) {
363 		xdpf = frames[i];
364 
365 		mapping = dma_map_single(dmadev, xdpf->data, xdpf->len,
366 					 DMA_TO_DEVICE);
367 		if (unlikely(dma_mapping_error(dmadev, mapping)))
368 			break;
369 
370 		if (unlikely(qede_xdp_xmit(xdp_tx, mapping, 0, xdpf->len,
371 					   NULL, xdpf)))
372 			break;
373 		nxmit++;
374 	}
375 
376 	if (flags & XDP_XMIT_FLUSH) {
377 		xdp_prod = qed_chain_get_prod_idx(&xdp_tx->tx_pbl);
378 
379 		xdp_tx->tx_db.data.bd_prod = cpu_to_le16(xdp_prod);
380 		qede_update_tx_producer(xdp_tx);
381 	}
382 
383 	spin_unlock(&xdp_tx->xdp_tx_lock);
384 
385 	return nxmit;
386 }
387 
388 int qede_txq_has_work(struct qede_tx_queue *txq)
389 {
390 	u16 hw_bd_cons;
391 
392 	/* Tell compiler that consumer and producer can change */
393 	barrier();
394 	hw_bd_cons = le16_to_cpu(*txq->hw_cons_ptr);
395 	if (qed_chain_get_cons_idx(&txq->tx_pbl) == hw_bd_cons + 1)
396 		return 0;
397 
398 	return hw_bd_cons != qed_chain_get_cons_idx(&txq->tx_pbl);
399 }
400 
401 static void qede_xdp_tx_int(struct qede_dev *edev, struct qede_tx_queue *txq)
402 {
403 	struct sw_tx_xdp *xdp_info, *xdp_arr = txq->sw_tx_ring.xdp;
404 	struct device *dev = &edev->pdev->dev;
405 	struct xdp_frame *xdpf;
406 	u16 hw_bd_cons;
407 
408 	hw_bd_cons = le16_to_cpu(*txq->hw_cons_ptr);
409 	barrier();
410 
411 	while (hw_bd_cons != qed_chain_get_cons_idx(&txq->tx_pbl)) {
412 		xdp_info = xdp_arr + txq->sw_tx_cons;
413 		xdpf = xdp_info->xdpf;
414 
415 		if (xdpf) {
416 			dma_unmap_single(dev, xdp_info->mapping, xdpf->len,
417 					 DMA_TO_DEVICE);
418 			xdp_return_frame(xdpf);
419 
420 			xdp_info->xdpf = NULL;
421 		} else {
422 			dma_unmap_page(dev, xdp_info->mapping, PAGE_SIZE,
423 				       DMA_BIDIRECTIONAL);
424 			__free_page(xdp_info->page);
425 		}
426 
427 		qed_chain_consume(&txq->tx_pbl);
428 		txq->sw_tx_cons = (txq->sw_tx_cons + 1) % txq->num_tx_buffers;
429 		txq->xmit_pkts++;
430 	}
431 }
432 
433 static int qede_tx_int(struct qede_dev *edev, struct qede_tx_queue *txq)
434 {
435 	unsigned int pkts_compl = 0, bytes_compl = 0;
436 	struct netdev_queue *netdev_txq;
437 	u16 hw_bd_cons;
438 	int rc;
439 
440 	netdev_txq = netdev_get_tx_queue(edev->ndev, txq->ndev_txq_id);
441 
442 	hw_bd_cons = le16_to_cpu(*txq->hw_cons_ptr);
443 	barrier();
444 
445 	while (hw_bd_cons != qed_chain_get_cons_idx(&txq->tx_pbl)) {
446 		int len = 0;
447 
448 		rc = qede_free_tx_pkt(edev, txq, &len);
449 		if (rc) {
450 			DP_NOTICE(edev, "hw_bd_cons = %d, chain_cons=%d\n",
451 				  hw_bd_cons,
452 				  qed_chain_get_cons_idx(&txq->tx_pbl));
453 			break;
454 		}
455 
456 		bytes_compl += len;
457 		pkts_compl++;
458 		txq->sw_tx_cons = (txq->sw_tx_cons + 1) % txq->num_tx_buffers;
459 		txq->xmit_pkts++;
460 	}
461 
462 	netdev_tx_completed_queue(netdev_txq, pkts_compl, bytes_compl);
463 
464 	/* Need to make the tx_bd_cons update visible to start_xmit()
465 	 * before checking for netif_tx_queue_stopped().  Without the
466 	 * memory barrier, there is a small possibility that
467 	 * start_xmit() will miss it and cause the queue to be stopped
468 	 * forever.
469 	 * On the other hand we need an rmb() here to ensure the proper
470 	 * ordering of bit testing in the following
471 	 * netif_tx_queue_stopped(txq) call.
472 	 */
473 	smp_mb();
474 
475 	if (unlikely(netif_tx_queue_stopped(netdev_txq))) {
476 		/* Taking tx_lock is needed to prevent reenabling the queue
477 		 * while it's empty. This could have happen if rx_action() gets
478 		 * suspended in qede_tx_int() after the condition before
479 		 * netif_tx_wake_queue(), while tx_action (qede_start_xmit()):
480 		 *
481 		 * stops the queue->sees fresh tx_bd_cons->releases the queue->
482 		 * sends some packets consuming the whole queue again->
483 		 * stops the queue
484 		 */
485 
486 		__netif_tx_lock(netdev_txq, smp_processor_id());
487 
488 		if ((netif_tx_queue_stopped(netdev_txq)) &&
489 		    (edev->state == QEDE_STATE_OPEN) &&
490 		    (qed_chain_get_elem_left(&txq->tx_pbl)
491 		      >= (MAX_SKB_FRAGS + 1))) {
492 			netif_tx_wake_queue(netdev_txq);
493 			DP_VERBOSE(edev, NETIF_MSG_TX_DONE,
494 				   "Wake queue was called\n");
495 		}
496 
497 		__netif_tx_unlock(netdev_txq);
498 	}
499 
500 	return 0;
501 }
502 
503 bool qede_has_rx_work(struct qede_rx_queue *rxq)
504 {
505 	u16 hw_comp_cons, sw_comp_cons;
506 
507 	/* Tell compiler that status block fields can change */
508 	barrier();
509 
510 	hw_comp_cons = le16_to_cpu(*rxq->hw_cons_ptr);
511 	sw_comp_cons = qed_chain_get_cons_idx(&rxq->rx_comp_ring);
512 
513 	return hw_comp_cons != sw_comp_cons;
514 }
515 
516 static inline void qede_rx_bd_ring_consume(struct qede_rx_queue *rxq)
517 {
518 	qed_chain_consume(&rxq->rx_bd_ring);
519 	rxq->sw_rx_cons++;
520 }
521 
522 /* This function reuses the buffer(from an offset) from
523  * consumer index to producer index in the bd ring
524  */
525 static inline void qede_reuse_page(struct qede_rx_queue *rxq,
526 				   struct sw_rx_data *curr_cons)
527 {
528 	struct eth_rx_bd *rx_bd_prod = qed_chain_produce(&rxq->rx_bd_ring);
529 	struct sw_rx_data *curr_prod;
530 	dma_addr_t new_mapping;
531 
532 	curr_prod = &rxq->sw_rx_ring[rxq->sw_rx_prod & NUM_RX_BDS_MAX];
533 	*curr_prod = *curr_cons;
534 
535 	new_mapping = curr_prod->mapping + curr_prod->page_offset;
536 
537 	rx_bd_prod->addr.hi = cpu_to_le32(upper_32_bits(new_mapping));
538 	rx_bd_prod->addr.lo = cpu_to_le32(lower_32_bits(new_mapping) +
539 					  rxq->rx_headroom);
540 
541 	rxq->sw_rx_prod++;
542 	curr_cons->data = NULL;
543 }
544 
545 /* In case of allocation failures reuse buffers
546  * from consumer index to produce buffers for firmware
547  */
548 void qede_recycle_rx_bd_ring(struct qede_rx_queue *rxq, u8 count)
549 {
550 	struct sw_rx_data *curr_cons;
551 
552 	for (; count > 0; count--) {
553 		curr_cons = &rxq->sw_rx_ring[rxq->sw_rx_cons & NUM_RX_BDS_MAX];
554 		qede_reuse_page(rxq, curr_cons);
555 		qede_rx_bd_ring_consume(rxq);
556 	}
557 }
558 
559 static inline int qede_realloc_rx_buffer(struct qede_rx_queue *rxq,
560 					 struct sw_rx_data *curr_cons)
561 {
562 	/* Move to the next segment in the page */
563 	curr_cons->page_offset += rxq->rx_buf_seg_size;
564 
565 	if (curr_cons->page_offset == PAGE_SIZE) {
566 		if (unlikely(qede_alloc_rx_buffer(rxq, true))) {
567 			/* Since we failed to allocate new buffer
568 			 * current buffer can be used again.
569 			 */
570 			curr_cons->page_offset -= rxq->rx_buf_seg_size;
571 
572 			return -ENOMEM;
573 		}
574 
575 		dma_unmap_page(rxq->dev, curr_cons->mapping,
576 			       PAGE_SIZE, rxq->data_direction);
577 	} else {
578 		/* Increment refcount of the page as we don't want
579 		 * network stack to take the ownership of the page
580 		 * which can be recycled multiple times by the driver.
581 		 */
582 		page_ref_inc(curr_cons->data);
583 		qede_reuse_page(rxq, curr_cons);
584 	}
585 
586 	return 0;
587 }
588 
589 void qede_update_rx_prod(struct qede_dev *edev, struct qede_rx_queue *rxq)
590 {
591 	u16 bd_prod = qed_chain_get_prod_idx(&rxq->rx_bd_ring);
592 	u16 cqe_prod = qed_chain_get_prod_idx(&rxq->rx_comp_ring);
593 	struct eth_rx_prod_data rx_prods = {0};
594 
595 	/* Update producers */
596 	rx_prods.bd_prod = cpu_to_le16(bd_prod);
597 	rx_prods.cqe_prod = cpu_to_le16(cqe_prod);
598 
599 	/* Make sure that the BD and SGE data is updated before updating the
600 	 * producers since FW might read the BD/SGE right after the producer
601 	 * is updated.
602 	 */
603 	wmb();
604 
605 	internal_ram_wr(rxq->hw_rxq_prod_addr, sizeof(rx_prods),
606 			(u32 *)&rx_prods);
607 }
608 
609 static void qede_get_rxhash(struct sk_buff *skb, u8 bitfields, __le32 rss_hash)
610 {
611 	enum pkt_hash_types hash_type = PKT_HASH_TYPE_NONE;
612 	enum rss_hash_type htype;
613 	u32 hash = 0;
614 
615 	htype = GET_FIELD(bitfields, ETH_FAST_PATH_RX_REG_CQE_RSS_HASH_TYPE);
616 	if (htype) {
617 		hash_type = ((htype == RSS_HASH_TYPE_IPV4) ||
618 			     (htype == RSS_HASH_TYPE_IPV6)) ?
619 			    PKT_HASH_TYPE_L3 : PKT_HASH_TYPE_L4;
620 		hash = le32_to_cpu(rss_hash);
621 	}
622 	skb_set_hash(skb, hash, hash_type);
623 }
624 
625 static void qede_set_skb_csum(struct sk_buff *skb, u8 csum_flag)
626 {
627 	skb_checksum_none_assert(skb);
628 
629 	if (csum_flag & QEDE_CSUM_UNNECESSARY)
630 		skb->ip_summed = CHECKSUM_UNNECESSARY;
631 
632 	if (csum_flag & QEDE_TUNN_CSUM_UNNECESSARY) {
633 		skb->csum_level = 1;
634 		skb->encapsulation = 1;
635 	}
636 }
637 
638 static inline void qede_skb_receive(struct qede_dev *edev,
639 				    struct qede_fastpath *fp,
640 				    struct qede_rx_queue *rxq,
641 				    struct sk_buff *skb, u16 vlan_tag)
642 {
643 	if (vlan_tag)
644 		__vlan_hwaccel_put_tag(skb, htons(ETH_P_8021Q), vlan_tag);
645 
646 	napi_gro_receive(&fp->napi, skb);
647 }
648 
649 static void qede_set_gro_params(struct qede_dev *edev,
650 				struct sk_buff *skb,
651 				struct eth_fast_path_rx_tpa_start_cqe *cqe)
652 {
653 	u16 parsing_flags = le16_to_cpu(cqe->pars_flags.flags);
654 
655 	if (((parsing_flags >> PARSING_AND_ERR_FLAGS_L3TYPE_SHIFT) &
656 	    PARSING_AND_ERR_FLAGS_L3TYPE_MASK) == 2)
657 		skb_shinfo(skb)->gso_type = SKB_GSO_TCPV6;
658 	else
659 		skb_shinfo(skb)->gso_type = SKB_GSO_TCPV4;
660 
661 	skb_shinfo(skb)->gso_size = __le16_to_cpu(cqe->len_on_first_bd) -
662 				    cqe->header_len;
663 }
664 
665 static int qede_fill_frag_skb(struct qede_dev *edev,
666 			      struct qede_rx_queue *rxq,
667 			      u8 tpa_agg_index, u16 len_on_bd)
668 {
669 	struct sw_rx_data *current_bd = &rxq->sw_rx_ring[rxq->sw_rx_cons &
670 							 NUM_RX_BDS_MAX];
671 	struct qede_agg_info *tpa_info = &rxq->tpa_info[tpa_agg_index];
672 	struct sk_buff *skb = tpa_info->skb;
673 
674 	if (unlikely(tpa_info->state != QEDE_AGG_STATE_START))
675 		goto out;
676 
677 	/* Add one frag and update the appropriate fields in the skb */
678 	skb_fill_page_desc(skb, tpa_info->frag_id++,
679 			   current_bd->data,
680 			   current_bd->page_offset + rxq->rx_headroom,
681 			   len_on_bd);
682 
683 	if (unlikely(qede_realloc_rx_buffer(rxq, current_bd))) {
684 		/* Incr page ref count to reuse on allocation failure
685 		 * so that it doesn't get freed while freeing SKB.
686 		 */
687 		page_ref_inc(current_bd->data);
688 		goto out;
689 	}
690 
691 	qede_rx_bd_ring_consume(rxq);
692 
693 	skb->data_len += len_on_bd;
694 	skb->truesize += rxq->rx_buf_seg_size;
695 	skb->len += len_on_bd;
696 
697 	return 0;
698 
699 out:
700 	tpa_info->state = QEDE_AGG_STATE_ERROR;
701 	qede_recycle_rx_bd_ring(rxq, 1);
702 
703 	return -ENOMEM;
704 }
705 
706 static bool qede_tunn_exist(u16 flag)
707 {
708 	return !!(flag & (PARSING_AND_ERR_FLAGS_TUNNELEXIST_MASK <<
709 			  PARSING_AND_ERR_FLAGS_TUNNELEXIST_SHIFT));
710 }
711 
712 static u8 qede_check_tunn_csum(u16 flag)
713 {
714 	u16 csum_flag = 0;
715 	u8 tcsum = 0;
716 
717 	if (flag & (PARSING_AND_ERR_FLAGS_TUNNELL4CHKSMWASCALCULATED_MASK <<
718 		    PARSING_AND_ERR_FLAGS_TUNNELL4CHKSMWASCALCULATED_SHIFT))
719 		csum_flag |= PARSING_AND_ERR_FLAGS_TUNNELL4CHKSMERROR_MASK <<
720 			     PARSING_AND_ERR_FLAGS_TUNNELL4CHKSMERROR_SHIFT;
721 
722 	if (flag & (PARSING_AND_ERR_FLAGS_L4CHKSMWASCALCULATED_MASK <<
723 		    PARSING_AND_ERR_FLAGS_L4CHKSMWASCALCULATED_SHIFT)) {
724 		csum_flag |= PARSING_AND_ERR_FLAGS_L4CHKSMERROR_MASK <<
725 			     PARSING_AND_ERR_FLAGS_L4CHKSMERROR_SHIFT;
726 		tcsum = QEDE_TUNN_CSUM_UNNECESSARY;
727 	}
728 
729 	csum_flag |= PARSING_AND_ERR_FLAGS_TUNNELIPHDRERROR_MASK <<
730 		     PARSING_AND_ERR_FLAGS_TUNNELIPHDRERROR_SHIFT |
731 		     PARSING_AND_ERR_FLAGS_IPHDRERROR_MASK <<
732 		     PARSING_AND_ERR_FLAGS_IPHDRERROR_SHIFT;
733 
734 	if (csum_flag & flag)
735 		return QEDE_CSUM_ERROR;
736 
737 	return QEDE_CSUM_UNNECESSARY | tcsum;
738 }
739 
740 static inline struct sk_buff *
741 qede_build_skb(struct qede_rx_queue *rxq,
742 	       struct sw_rx_data *bd, u16 len, u16 pad)
743 {
744 	struct sk_buff *skb;
745 	void *buf;
746 
747 	buf = page_address(bd->data) + bd->page_offset;
748 	skb = build_skb(buf, rxq->rx_buf_seg_size);
749 
750 	skb_reserve(skb, pad);
751 	skb_put(skb, len);
752 
753 	return skb;
754 }
755 
756 static struct sk_buff *
757 qede_tpa_rx_build_skb(struct qede_dev *edev,
758 		      struct qede_rx_queue *rxq,
759 		      struct sw_rx_data *bd, u16 len, u16 pad,
760 		      bool alloc_skb)
761 {
762 	struct sk_buff *skb;
763 
764 	skb = qede_build_skb(rxq, bd, len, pad);
765 	bd->page_offset += rxq->rx_buf_seg_size;
766 
767 	if (bd->page_offset == PAGE_SIZE) {
768 		if (unlikely(qede_alloc_rx_buffer(rxq, true))) {
769 			DP_NOTICE(edev,
770 				  "Failed to allocate RX buffer for tpa start\n");
771 			bd->page_offset -= rxq->rx_buf_seg_size;
772 			page_ref_inc(bd->data);
773 			dev_kfree_skb_any(skb);
774 			return NULL;
775 		}
776 	} else {
777 		page_ref_inc(bd->data);
778 		qede_reuse_page(rxq, bd);
779 	}
780 
781 	/* We've consumed the first BD and prepared an SKB */
782 	qede_rx_bd_ring_consume(rxq);
783 
784 	return skb;
785 }
786 
787 static struct sk_buff *
788 qede_rx_build_skb(struct qede_dev *edev,
789 		  struct qede_rx_queue *rxq,
790 		  struct sw_rx_data *bd, u16 len, u16 pad)
791 {
792 	struct sk_buff *skb = NULL;
793 
794 	/* For smaller frames still need to allocate skb, memcpy
795 	 * data and benefit in reusing the page segment instead of
796 	 * un-mapping it.
797 	 */
798 	if ((len + pad <= edev->rx_copybreak)) {
799 		unsigned int offset = bd->page_offset + pad;
800 
801 		skb = netdev_alloc_skb(edev->ndev, QEDE_RX_HDR_SIZE);
802 		if (unlikely(!skb))
803 			return NULL;
804 
805 		skb_reserve(skb, pad);
806 		skb_put_data(skb, page_address(bd->data) + offset, len);
807 		qede_reuse_page(rxq, bd);
808 		goto out;
809 	}
810 
811 	skb = qede_build_skb(rxq, bd, len, pad);
812 
813 	if (unlikely(qede_realloc_rx_buffer(rxq, bd))) {
814 		/* Incr page ref count to reuse on allocation failure so
815 		 * that it doesn't get freed while freeing SKB [as its
816 		 * already mapped there].
817 		 */
818 		page_ref_inc(bd->data);
819 		dev_kfree_skb_any(skb);
820 		return NULL;
821 	}
822 out:
823 	/* We've consumed the first BD and prepared an SKB */
824 	qede_rx_bd_ring_consume(rxq);
825 
826 	return skb;
827 }
828 
829 static void qede_tpa_start(struct qede_dev *edev,
830 			   struct qede_rx_queue *rxq,
831 			   struct eth_fast_path_rx_tpa_start_cqe *cqe)
832 {
833 	struct qede_agg_info *tpa_info = &rxq->tpa_info[cqe->tpa_agg_index];
834 	struct sw_rx_data *sw_rx_data_cons;
835 	u16 pad;
836 
837 	sw_rx_data_cons = &rxq->sw_rx_ring[rxq->sw_rx_cons & NUM_RX_BDS_MAX];
838 	pad = cqe->placement_offset + rxq->rx_headroom;
839 
840 	tpa_info->skb = qede_tpa_rx_build_skb(edev, rxq, sw_rx_data_cons,
841 					      le16_to_cpu(cqe->len_on_first_bd),
842 					      pad, false);
843 	tpa_info->buffer.page_offset = sw_rx_data_cons->page_offset;
844 	tpa_info->buffer.mapping = sw_rx_data_cons->mapping;
845 
846 	if (unlikely(!tpa_info->skb)) {
847 		DP_NOTICE(edev, "Failed to allocate SKB for gro\n");
848 
849 		/* Consume from ring but do not produce since
850 		 * this might be used by FW still, it will be re-used
851 		 * at TPA end.
852 		 */
853 		tpa_info->tpa_start_fail = true;
854 		qede_rx_bd_ring_consume(rxq);
855 		tpa_info->state = QEDE_AGG_STATE_ERROR;
856 		goto cons_buf;
857 	}
858 
859 	tpa_info->frag_id = 0;
860 	tpa_info->state = QEDE_AGG_STATE_START;
861 
862 	if ((le16_to_cpu(cqe->pars_flags.flags) >>
863 	     PARSING_AND_ERR_FLAGS_TAG8021QEXIST_SHIFT) &
864 	    PARSING_AND_ERR_FLAGS_TAG8021QEXIST_MASK)
865 		tpa_info->vlan_tag = le16_to_cpu(cqe->vlan_tag);
866 	else
867 		tpa_info->vlan_tag = 0;
868 
869 	qede_get_rxhash(tpa_info->skb, cqe->bitfields, cqe->rss_hash);
870 
871 	/* This is needed in order to enable forwarding support */
872 	qede_set_gro_params(edev, tpa_info->skb, cqe);
873 
874 cons_buf: /* We still need to handle bd_len_list to consume buffers */
875 	if (likely(cqe->bw_ext_bd_len_list[0]))
876 		qede_fill_frag_skb(edev, rxq, cqe->tpa_agg_index,
877 				   le16_to_cpu(cqe->bw_ext_bd_len_list[0]));
878 
879 	if (unlikely(cqe->bw_ext_bd_len_list[1])) {
880 		DP_ERR(edev,
881 		       "Unlikely - got a TPA aggregation with more than one bw_ext_bd_len_list entry in the TPA start\n");
882 		tpa_info->state = QEDE_AGG_STATE_ERROR;
883 	}
884 }
885 
886 #ifdef CONFIG_INET
887 static void qede_gro_ip_csum(struct sk_buff *skb)
888 {
889 	const struct iphdr *iph = ip_hdr(skb);
890 	struct tcphdr *th;
891 
892 	skb_set_transport_header(skb, sizeof(struct iphdr));
893 	th = tcp_hdr(skb);
894 
895 	th->check = ~tcp_v4_check(skb->len - skb_transport_offset(skb),
896 				  iph->saddr, iph->daddr, 0);
897 
898 	tcp_gro_complete(skb);
899 }
900 
901 static void qede_gro_ipv6_csum(struct sk_buff *skb)
902 {
903 	struct ipv6hdr *iph = ipv6_hdr(skb);
904 	struct tcphdr *th;
905 
906 	skb_set_transport_header(skb, sizeof(struct ipv6hdr));
907 	th = tcp_hdr(skb);
908 
909 	th->check = ~tcp_v6_check(skb->len - skb_transport_offset(skb),
910 				  &iph->saddr, &iph->daddr, 0);
911 	tcp_gro_complete(skb);
912 }
913 #endif
914 
915 static void qede_gro_receive(struct qede_dev *edev,
916 			     struct qede_fastpath *fp,
917 			     struct sk_buff *skb,
918 			     u16 vlan_tag)
919 {
920 	/* FW can send a single MTU sized packet from gro flow
921 	 * due to aggregation timeout/last segment etc. which
922 	 * is not expected to be a gro packet. If a skb has zero
923 	 * frags then simply push it in the stack as non gso skb.
924 	 */
925 	if (unlikely(!skb->data_len)) {
926 		skb_shinfo(skb)->gso_type = 0;
927 		skb_shinfo(skb)->gso_size = 0;
928 		goto send_skb;
929 	}
930 
931 #ifdef CONFIG_INET
932 	if (skb_shinfo(skb)->gso_size) {
933 		skb_reset_network_header(skb);
934 
935 		switch (skb->protocol) {
936 		case htons(ETH_P_IP):
937 			qede_gro_ip_csum(skb);
938 			break;
939 		case htons(ETH_P_IPV6):
940 			qede_gro_ipv6_csum(skb);
941 			break;
942 		default:
943 			DP_ERR(edev,
944 			       "Error: FW GRO supports only IPv4/IPv6, not 0x%04x\n",
945 			       ntohs(skb->protocol));
946 		}
947 	}
948 #endif
949 
950 send_skb:
951 	skb_record_rx_queue(skb, fp->rxq->rxq_id);
952 	qede_skb_receive(edev, fp, fp->rxq, skb, vlan_tag);
953 }
954 
955 static inline void qede_tpa_cont(struct qede_dev *edev,
956 				 struct qede_rx_queue *rxq,
957 				 struct eth_fast_path_rx_tpa_cont_cqe *cqe)
958 {
959 	int i;
960 
961 	for (i = 0; cqe->len_list[i]; i++)
962 		qede_fill_frag_skb(edev, rxq, cqe->tpa_agg_index,
963 				   le16_to_cpu(cqe->len_list[i]));
964 
965 	if (unlikely(i > 1))
966 		DP_ERR(edev,
967 		       "Strange - TPA cont with more than a single len_list entry\n");
968 }
969 
970 static int qede_tpa_end(struct qede_dev *edev,
971 			struct qede_fastpath *fp,
972 			struct eth_fast_path_rx_tpa_end_cqe *cqe)
973 {
974 	struct qede_rx_queue *rxq = fp->rxq;
975 	struct qede_agg_info *tpa_info;
976 	struct sk_buff *skb;
977 	int i;
978 
979 	tpa_info = &rxq->tpa_info[cqe->tpa_agg_index];
980 	skb = tpa_info->skb;
981 
982 	if (tpa_info->buffer.page_offset == PAGE_SIZE)
983 		dma_unmap_page(rxq->dev, tpa_info->buffer.mapping,
984 			       PAGE_SIZE, rxq->data_direction);
985 
986 	for (i = 0; cqe->len_list[i]; i++)
987 		qede_fill_frag_skb(edev, rxq, cqe->tpa_agg_index,
988 				   le16_to_cpu(cqe->len_list[i]));
989 	if (unlikely(i > 1))
990 		DP_ERR(edev,
991 		       "Strange - TPA emd with more than a single len_list entry\n");
992 
993 	if (unlikely(tpa_info->state != QEDE_AGG_STATE_START))
994 		goto err;
995 
996 	/* Sanity */
997 	if (unlikely(cqe->num_of_bds != tpa_info->frag_id + 1))
998 		DP_ERR(edev,
999 		       "Strange - TPA had %02x BDs, but SKB has only %d frags\n",
1000 		       cqe->num_of_bds, tpa_info->frag_id);
1001 	if (unlikely(skb->len != le16_to_cpu(cqe->total_packet_len)))
1002 		DP_ERR(edev,
1003 		       "Strange - total packet len [cqe] is %4x but SKB has len %04x\n",
1004 		       le16_to_cpu(cqe->total_packet_len), skb->len);
1005 
1006 	/* Finalize the SKB */
1007 	skb->protocol = eth_type_trans(skb, edev->ndev);
1008 	skb->ip_summed = CHECKSUM_UNNECESSARY;
1009 
1010 	/* tcp_gro_complete() will copy NAPI_GRO_CB(skb)->count
1011 	 * to skb_shinfo(skb)->gso_segs
1012 	 */
1013 	NAPI_GRO_CB(skb)->count = le16_to_cpu(cqe->num_of_coalesced_segs);
1014 
1015 	qede_gro_receive(edev, fp, skb, tpa_info->vlan_tag);
1016 
1017 	tpa_info->state = QEDE_AGG_STATE_NONE;
1018 
1019 	return 1;
1020 err:
1021 	tpa_info->state = QEDE_AGG_STATE_NONE;
1022 
1023 	if (tpa_info->tpa_start_fail) {
1024 		qede_reuse_page(rxq, &tpa_info->buffer);
1025 		tpa_info->tpa_start_fail = false;
1026 	}
1027 
1028 	dev_kfree_skb_any(tpa_info->skb);
1029 	tpa_info->skb = NULL;
1030 	return 0;
1031 }
1032 
1033 static u8 qede_check_notunn_csum(u16 flag)
1034 {
1035 	u16 csum_flag = 0;
1036 	u8 csum = 0;
1037 
1038 	if (flag & (PARSING_AND_ERR_FLAGS_L4CHKSMWASCALCULATED_MASK <<
1039 		    PARSING_AND_ERR_FLAGS_L4CHKSMWASCALCULATED_SHIFT)) {
1040 		csum_flag |= PARSING_AND_ERR_FLAGS_L4CHKSMERROR_MASK <<
1041 			     PARSING_AND_ERR_FLAGS_L4CHKSMERROR_SHIFT;
1042 		csum = QEDE_CSUM_UNNECESSARY;
1043 	}
1044 
1045 	csum_flag |= PARSING_AND_ERR_FLAGS_IPHDRERROR_MASK <<
1046 		     PARSING_AND_ERR_FLAGS_IPHDRERROR_SHIFT;
1047 
1048 	if (csum_flag & flag)
1049 		return QEDE_CSUM_ERROR;
1050 
1051 	return csum;
1052 }
1053 
1054 static u8 qede_check_csum(u16 flag)
1055 {
1056 	if (!qede_tunn_exist(flag))
1057 		return qede_check_notunn_csum(flag);
1058 	else
1059 		return qede_check_tunn_csum(flag);
1060 }
1061 
1062 static bool qede_pkt_is_ip_fragmented(struct eth_fast_path_rx_reg_cqe *cqe,
1063 				      u16 flag)
1064 {
1065 	u8 tun_pars_flg = cqe->tunnel_pars_flags.flags;
1066 
1067 	if ((tun_pars_flg & (ETH_TUNNEL_PARSING_FLAGS_IPV4_FRAGMENT_MASK <<
1068 			     ETH_TUNNEL_PARSING_FLAGS_IPV4_FRAGMENT_SHIFT)) ||
1069 	    (flag & (PARSING_AND_ERR_FLAGS_IPV4FRAG_MASK <<
1070 		     PARSING_AND_ERR_FLAGS_IPV4FRAG_SHIFT)))
1071 		return true;
1072 
1073 	return false;
1074 }
1075 
1076 /* Return true iff packet is to be passed to stack */
1077 static bool qede_rx_xdp(struct qede_dev *edev,
1078 			struct qede_fastpath *fp,
1079 			struct qede_rx_queue *rxq,
1080 			struct bpf_prog *prog,
1081 			struct sw_rx_data *bd,
1082 			struct eth_fast_path_rx_reg_cqe *cqe,
1083 			u16 *data_offset, u16 *len)
1084 {
1085 	struct xdp_buff xdp;
1086 	enum xdp_action act;
1087 
1088 	xdp_init_buff(&xdp, rxq->rx_buf_seg_size, &rxq->xdp_rxq);
1089 	xdp_prepare_buff(&xdp, page_address(bd->data), *data_offset,
1090 			 *len, false);
1091 
1092 	act = bpf_prog_run_xdp(prog, &xdp);
1093 
1094 	/* Recalculate, as XDP might have changed the headers */
1095 	*data_offset = xdp.data - xdp.data_hard_start;
1096 	*len = xdp.data_end - xdp.data;
1097 
1098 	if (act == XDP_PASS)
1099 		return true;
1100 
1101 	/* Count number of packets not to be passed to stack */
1102 	rxq->xdp_no_pass++;
1103 
1104 	switch (act) {
1105 	case XDP_TX:
1106 		/* We need the replacement buffer before transmit. */
1107 		if (unlikely(qede_alloc_rx_buffer(rxq, true))) {
1108 			qede_recycle_rx_bd_ring(rxq, 1);
1109 
1110 			trace_xdp_exception(edev->ndev, prog, act);
1111 			break;
1112 		}
1113 
1114 		/* Now if there's a transmission problem, we'd still have to
1115 		 * throw current buffer, as replacement was already allocated.
1116 		 */
1117 		if (unlikely(qede_xdp_xmit(fp->xdp_tx, bd->mapping,
1118 					   *data_offset, *len, bd->data,
1119 					   NULL))) {
1120 			dma_unmap_page(rxq->dev, bd->mapping, PAGE_SIZE,
1121 				       rxq->data_direction);
1122 			__free_page(bd->data);
1123 
1124 			trace_xdp_exception(edev->ndev, prog, act);
1125 		} else {
1126 			dma_sync_single_for_device(rxq->dev,
1127 						   bd->mapping + *data_offset,
1128 						   *len, rxq->data_direction);
1129 			fp->xdp_xmit |= QEDE_XDP_TX;
1130 		}
1131 
1132 		/* Regardless, we've consumed an Rx BD */
1133 		qede_rx_bd_ring_consume(rxq);
1134 		break;
1135 	case XDP_REDIRECT:
1136 		/* We need the replacement buffer before transmit. */
1137 		if (unlikely(qede_alloc_rx_buffer(rxq, true))) {
1138 			qede_recycle_rx_bd_ring(rxq, 1);
1139 
1140 			trace_xdp_exception(edev->ndev, prog, act);
1141 			break;
1142 		}
1143 
1144 		dma_unmap_page(rxq->dev, bd->mapping, PAGE_SIZE,
1145 			       rxq->data_direction);
1146 
1147 		if (unlikely(xdp_do_redirect(edev->ndev, &xdp, prog)))
1148 			DP_NOTICE(edev, "Failed to redirect the packet\n");
1149 		else
1150 			fp->xdp_xmit |= QEDE_XDP_REDIRECT;
1151 
1152 		qede_rx_bd_ring_consume(rxq);
1153 		break;
1154 	default:
1155 		bpf_warn_invalid_xdp_action(act);
1156 		fallthrough;
1157 	case XDP_ABORTED:
1158 		trace_xdp_exception(edev->ndev, prog, act);
1159 		fallthrough;
1160 	case XDP_DROP:
1161 		qede_recycle_rx_bd_ring(rxq, cqe->bd_num);
1162 	}
1163 
1164 	return false;
1165 }
1166 
1167 static int qede_rx_build_jumbo(struct qede_dev *edev,
1168 			       struct qede_rx_queue *rxq,
1169 			       struct sk_buff *skb,
1170 			       struct eth_fast_path_rx_reg_cqe *cqe,
1171 			       u16 first_bd_len)
1172 {
1173 	u16 pkt_len = le16_to_cpu(cqe->pkt_len);
1174 	struct sw_rx_data *bd;
1175 	u16 bd_cons_idx;
1176 	u8 num_frags;
1177 
1178 	pkt_len -= first_bd_len;
1179 
1180 	/* We've already used one BD for the SKB. Now take care of the rest */
1181 	for (num_frags = cqe->bd_num - 1; num_frags > 0; num_frags--) {
1182 		u16 cur_size = pkt_len > rxq->rx_buf_size ? rxq->rx_buf_size :
1183 		    pkt_len;
1184 
1185 		if (unlikely(!cur_size)) {
1186 			DP_ERR(edev,
1187 			       "Still got %d BDs for mapping jumbo, but length became 0\n",
1188 			       num_frags);
1189 			goto out;
1190 		}
1191 
1192 		/* We need a replacement buffer for each BD */
1193 		if (unlikely(qede_alloc_rx_buffer(rxq, true)))
1194 			goto out;
1195 
1196 		/* Now that we've allocated the replacement buffer,
1197 		 * we can safely consume the next BD and map it to the SKB.
1198 		 */
1199 		bd_cons_idx = rxq->sw_rx_cons & NUM_RX_BDS_MAX;
1200 		bd = &rxq->sw_rx_ring[bd_cons_idx];
1201 		qede_rx_bd_ring_consume(rxq);
1202 
1203 		dma_unmap_page(rxq->dev, bd->mapping,
1204 			       PAGE_SIZE, DMA_FROM_DEVICE);
1205 
1206 		skb_add_rx_frag(skb, skb_shinfo(skb)->nr_frags, bd->data,
1207 				rxq->rx_headroom, cur_size, PAGE_SIZE);
1208 
1209 		pkt_len -= cur_size;
1210 	}
1211 
1212 	if (unlikely(pkt_len))
1213 		DP_ERR(edev,
1214 		       "Mapped all BDs of jumbo, but still have %d bytes\n",
1215 		       pkt_len);
1216 
1217 out:
1218 	return num_frags;
1219 }
1220 
1221 static int qede_rx_process_tpa_cqe(struct qede_dev *edev,
1222 				   struct qede_fastpath *fp,
1223 				   struct qede_rx_queue *rxq,
1224 				   union eth_rx_cqe *cqe,
1225 				   enum eth_rx_cqe_type type)
1226 {
1227 	switch (type) {
1228 	case ETH_RX_CQE_TYPE_TPA_START:
1229 		qede_tpa_start(edev, rxq, &cqe->fast_path_tpa_start);
1230 		return 0;
1231 	case ETH_RX_CQE_TYPE_TPA_CONT:
1232 		qede_tpa_cont(edev, rxq, &cqe->fast_path_tpa_cont);
1233 		return 0;
1234 	case ETH_RX_CQE_TYPE_TPA_END:
1235 		return qede_tpa_end(edev, fp, &cqe->fast_path_tpa_end);
1236 	default:
1237 		return 0;
1238 	}
1239 }
1240 
1241 static int qede_rx_process_cqe(struct qede_dev *edev,
1242 			       struct qede_fastpath *fp,
1243 			       struct qede_rx_queue *rxq)
1244 {
1245 	struct bpf_prog *xdp_prog = READ_ONCE(rxq->xdp_prog);
1246 	struct eth_fast_path_rx_reg_cqe *fp_cqe;
1247 	u16 len, pad, bd_cons_idx, parse_flag;
1248 	enum eth_rx_cqe_type cqe_type;
1249 	union eth_rx_cqe *cqe;
1250 	struct sw_rx_data *bd;
1251 	struct sk_buff *skb;
1252 	__le16 flags;
1253 	u8 csum_flag;
1254 
1255 	/* Get the CQE from the completion ring */
1256 	cqe = (union eth_rx_cqe *)qed_chain_consume(&rxq->rx_comp_ring);
1257 	cqe_type = cqe->fast_path_regular.type;
1258 
1259 	/* Process an unlikely slowpath event */
1260 	if (unlikely(cqe_type == ETH_RX_CQE_TYPE_SLOW_PATH)) {
1261 		struct eth_slow_path_rx_cqe *sp_cqe;
1262 
1263 		sp_cqe = (struct eth_slow_path_rx_cqe *)cqe;
1264 		edev->ops->eth_cqe_completion(edev->cdev, fp->id, sp_cqe);
1265 		return 0;
1266 	}
1267 
1268 	/* Handle TPA cqes */
1269 	if (cqe_type != ETH_RX_CQE_TYPE_REGULAR)
1270 		return qede_rx_process_tpa_cqe(edev, fp, rxq, cqe, cqe_type);
1271 
1272 	/* Get the data from the SW ring; Consume it only after it's evident
1273 	 * we wouldn't recycle it.
1274 	 */
1275 	bd_cons_idx = rxq->sw_rx_cons & NUM_RX_BDS_MAX;
1276 	bd = &rxq->sw_rx_ring[bd_cons_idx];
1277 
1278 	fp_cqe = &cqe->fast_path_regular;
1279 	len = le16_to_cpu(fp_cqe->len_on_first_bd);
1280 	pad = fp_cqe->placement_offset + rxq->rx_headroom;
1281 
1282 	/* Run eBPF program if one is attached */
1283 	if (xdp_prog)
1284 		if (!qede_rx_xdp(edev, fp, rxq, xdp_prog, bd, fp_cqe,
1285 				 &pad, &len))
1286 			return 0;
1287 
1288 	/* If this is an error packet then drop it */
1289 	flags = cqe->fast_path_regular.pars_flags.flags;
1290 	parse_flag = le16_to_cpu(flags);
1291 
1292 	csum_flag = qede_check_csum(parse_flag);
1293 	if (unlikely(csum_flag == QEDE_CSUM_ERROR)) {
1294 		if (qede_pkt_is_ip_fragmented(fp_cqe, parse_flag))
1295 			rxq->rx_ip_frags++;
1296 		else
1297 			rxq->rx_hw_errors++;
1298 	}
1299 
1300 	/* Basic validation passed; Need to prepare an SKB. This would also
1301 	 * guarantee to finally consume the first BD upon success.
1302 	 */
1303 	skb = qede_rx_build_skb(edev, rxq, bd, len, pad);
1304 	if (!skb) {
1305 		rxq->rx_alloc_errors++;
1306 		qede_recycle_rx_bd_ring(rxq, fp_cqe->bd_num);
1307 		return 0;
1308 	}
1309 
1310 	/* In case of Jumbo packet, several PAGE_SIZEd buffers will be pointed
1311 	 * by a single cqe.
1312 	 */
1313 	if (fp_cqe->bd_num > 1) {
1314 		u16 unmapped_frags = qede_rx_build_jumbo(edev, rxq, skb,
1315 							 fp_cqe, len);
1316 
1317 		if (unlikely(unmapped_frags > 0)) {
1318 			qede_recycle_rx_bd_ring(rxq, unmapped_frags);
1319 			dev_kfree_skb_any(skb);
1320 			return 0;
1321 		}
1322 	}
1323 
1324 	/* The SKB contains all the data. Now prepare meta-magic */
1325 	skb->protocol = eth_type_trans(skb, edev->ndev);
1326 	qede_get_rxhash(skb, fp_cqe->bitfields, fp_cqe->rss_hash);
1327 	qede_set_skb_csum(skb, csum_flag);
1328 	skb_record_rx_queue(skb, rxq->rxq_id);
1329 	qede_ptp_record_rx_ts(edev, cqe, skb);
1330 
1331 	/* SKB is prepared - pass it to stack */
1332 	qede_skb_receive(edev, fp, rxq, skb, le16_to_cpu(fp_cqe->vlan_tag));
1333 
1334 	return 1;
1335 }
1336 
1337 static int qede_rx_int(struct qede_fastpath *fp, int budget)
1338 {
1339 	struct qede_rx_queue *rxq = fp->rxq;
1340 	struct qede_dev *edev = fp->edev;
1341 	int work_done = 0, rcv_pkts = 0;
1342 	u16 hw_comp_cons, sw_comp_cons;
1343 
1344 	hw_comp_cons = le16_to_cpu(*rxq->hw_cons_ptr);
1345 	sw_comp_cons = qed_chain_get_cons_idx(&rxq->rx_comp_ring);
1346 
1347 	/* Memory barrier to prevent the CPU from doing speculative reads of CQE
1348 	 * / BD in the while-loop before reading hw_comp_cons. If the CQE is
1349 	 * read before it is written by FW, then FW writes CQE and SB, and then
1350 	 * the CPU reads the hw_comp_cons, it will use an old CQE.
1351 	 */
1352 	rmb();
1353 
1354 	/* Loop to complete all indicated BDs */
1355 	while ((sw_comp_cons != hw_comp_cons) && (work_done < budget)) {
1356 		rcv_pkts += qede_rx_process_cqe(edev, fp, rxq);
1357 		qed_chain_recycle_consumed(&rxq->rx_comp_ring);
1358 		sw_comp_cons = qed_chain_get_cons_idx(&rxq->rx_comp_ring);
1359 		work_done++;
1360 	}
1361 
1362 	rxq->rcv_pkts += rcv_pkts;
1363 
1364 	/* Allocate replacement buffers */
1365 	while (rxq->num_rx_buffers - rxq->filled_buffers)
1366 		if (qede_alloc_rx_buffer(rxq, false))
1367 			break;
1368 
1369 	/* Update producers */
1370 	qede_update_rx_prod(edev, rxq);
1371 
1372 	return work_done;
1373 }
1374 
1375 static bool qede_poll_is_more_work(struct qede_fastpath *fp)
1376 {
1377 	qed_sb_update_sb_idx(fp->sb_info);
1378 
1379 	/* *_has_*_work() reads the status block, thus we need to ensure that
1380 	 * status block indices have been actually read (qed_sb_update_sb_idx)
1381 	 * prior to this check (*_has_*_work) so that we won't write the
1382 	 * "newer" value of the status block to HW (if there was a DMA right
1383 	 * after qede_has_rx_work and if there is no rmb, the memory reading
1384 	 * (qed_sb_update_sb_idx) may be postponed to right before *_ack_sb).
1385 	 * In this case there will never be another interrupt until there is
1386 	 * another update of the status block, while there is still unhandled
1387 	 * work.
1388 	 */
1389 	rmb();
1390 
1391 	if (likely(fp->type & QEDE_FASTPATH_RX))
1392 		if (qede_has_rx_work(fp->rxq))
1393 			return true;
1394 
1395 	if (fp->type & QEDE_FASTPATH_XDP)
1396 		if (qede_txq_has_work(fp->xdp_tx))
1397 			return true;
1398 
1399 	if (likely(fp->type & QEDE_FASTPATH_TX)) {
1400 		int cos;
1401 
1402 		for_each_cos_in_txq(fp->edev, cos) {
1403 			if (qede_txq_has_work(&fp->txq[cos]))
1404 				return true;
1405 		}
1406 	}
1407 
1408 	return false;
1409 }
1410 
1411 /*********************
1412  * NDO & API related *
1413  *********************/
1414 int qede_poll(struct napi_struct *napi, int budget)
1415 {
1416 	struct qede_fastpath *fp = container_of(napi, struct qede_fastpath,
1417 						napi);
1418 	struct qede_dev *edev = fp->edev;
1419 	int rx_work_done = 0;
1420 	u16 xdp_prod;
1421 
1422 	fp->xdp_xmit = 0;
1423 
1424 	if (likely(fp->type & QEDE_FASTPATH_TX)) {
1425 		int cos;
1426 
1427 		for_each_cos_in_txq(fp->edev, cos) {
1428 			if (qede_txq_has_work(&fp->txq[cos]))
1429 				qede_tx_int(edev, &fp->txq[cos]);
1430 		}
1431 	}
1432 
1433 	if ((fp->type & QEDE_FASTPATH_XDP) && qede_txq_has_work(fp->xdp_tx))
1434 		qede_xdp_tx_int(edev, fp->xdp_tx);
1435 
1436 	rx_work_done = (likely(fp->type & QEDE_FASTPATH_RX) &&
1437 			qede_has_rx_work(fp->rxq)) ?
1438 			qede_rx_int(fp, budget) : 0;
1439 	/* Handle case where we are called by netpoll with a budget of 0 */
1440 	if (rx_work_done < budget || !budget) {
1441 		if (!qede_poll_is_more_work(fp)) {
1442 			napi_complete_done(napi, rx_work_done);
1443 
1444 			/* Update and reenable interrupts */
1445 			qed_sb_ack(fp->sb_info, IGU_INT_ENABLE, 1);
1446 		} else {
1447 			rx_work_done = budget;
1448 		}
1449 	}
1450 
1451 	if (fp->xdp_xmit & QEDE_XDP_TX) {
1452 		xdp_prod = qed_chain_get_prod_idx(&fp->xdp_tx->tx_pbl);
1453 
1454 		fp->xdp_tx->tx_db.data.bd_prod = cpu_to_le16(xdp_prod);
1455 		qede_update_tx_producer(fp->xdp_tx);
1456 	}
1457 
1458 	if (fp->xdp_xmit & QEDE_XDP_REDIRECT)
1459 		xdp_do_flush_map();
1460 
1461 	return rx_work_done;
1462 }
1463 
1464 irqreturn_t qede_msix_fp_int(int irq, void *fp_cookie)
1465 {
1466 	struct qede_fastpath *fp = fp_cookie;
1467 
1468 	qed_sb_ack(fp->sb_info, IGU_INT_DISABLE, 0 /*do not update*/);
1469 
1470 	napi_schedule_irqoff(&fp->napi);
1471 	return IRQ_HANDLED;
1472 }
1473 
1474 /* Main transmit function */
1475 netdev_tx_t qede_start_xmit(struct sk_buff *skb, struct net_device *ndev)
1476 {
1477 	struct qede_dev *edev = netdev_priv(ndev);
1478 	struct netdev_queue *netdev_txq;
1479 	struct qede_tx_queue *txq;
1480 	struct eth_tx_1st_bd *first_bd;
1481 	struct eth_tx_2nd_bd *second_bd = NULL;
1482 	struct eth_tx_3rd_bd *third_bd = NULL;
1483 	struct eth_tx_bd *tx_data_bd = NULL;
1484 	u16 txq_index, val = 0;
1485 	u8 nbd = 0;
1486 	dma_addr_t mapping;
1487 	int rc, frag_idx = 0, ipv6_ext = 0;
1488 	u8 xmit_type;
1489 	u16 idx;
1490 	u16 hlen;
1491 	bool data_split = false;
1492 
1493 	/* Get tx-queue context and netdev index */
1494 	txq_index = skb_get_queue_mapping(skb);
1495 	WARN_ON(txq_index >= QEDE_TSS_COUNT(edev) * edev->dev_info.num_tc);
1496 	txq = QEDE_NDEV_TXQ_ID_TO_TXQ(edev, txq_index);
1497 	netdev_txq = netdev_get_tx_queue(ndev, txq_index);
1498 
1499 	WARN_ON(qed_chain_get_elem_left(&txq->tx_pbl) < (MAX_SKB_FRAGS + 1));
1500 
1501 	xmit_type = qede_xmit_type(skb, &ipv6_ext);
1502 
1503 #if ((MAX_SKB_FRAGS + 2) > ETH_TX_MAX_BDS_PER_NON_LSO_PACKET)
1504 	if (qede_pkt_req_lin(skb, xmit_type)) {
1505 		if (skb_linearize(skb)) {
1506 			txq->tx_mem_alloc_err++;
1507 
1508 			dev_kfree_skb_any(skb);
1509 			return NETDEV_TX_OK;
1510 		}
1511 	}
1512 #endif
1513 
1514 	/* Fill the entry in the SW ring and the BDs in the FW ring */
1515 	idx = txq->sw_tx_prod;
1516 	txq->sw_tx_ring.skbs[idx].skb = skb;
1517 	first_bd = (struct eth_tx_1st_bd *)
1518 		   qed_chain_produce(&txq->tx_pbl);
1519 	memset(first_bd, 0, sizeof(*first_bd));
1520 	first_bd->data.bd_flags.bitfields =
1521 		1 << ETH_TX_1ST_BD_FLAGS_START_BD_SHIFT;
1522 
1523 	if (unlikely(skb_shinfo(skb)->tx_flags & SKBTX_HW_TSTAMP))
1524 		qede_ptp_tx_ts(edev, skb);
1525 
1526 	/* Map skb linear data for DMA and set in the first BD */
1527 	mapping = dma_map_single(txq->dev, skb->data,
1528 				 skb_headlen(skb), DMA_TO_DEVICE);
1529 	if (unlikely(dma_mapping_error(txq->dev, mapping))) {
1530 		DP_NOTICE(edev, "SKB mapping failed\n");
1531 		qede_free_failed_tx_pkt(txq, first_bd, 0, false);
1532 		qede_update_tx_producer(txq);
1533 		return NETDEV_TX_OK;
1534 	}
1535 	nbd++;
1536 	BD_SET_UNMAP_ADDR_LEN(first_bd, mapping, skb_headlen(skb));
1537 
1538 	/* In case there is IPv6 with extension headers or LSO we need 2nd and
1539 	 * 3rd BDs.
1540 	 */
1541 	if (unlikely((xmit_type & XMIT_LSO) | ipv6_ext)) {
1542 		second_bd = (struct eth_tx_2nd_bd *)
1543 			qed_chain_produce(&txq->tx_pbl);
1544 		memset(second_bd, 0, sizeof(*second_bd));
1545 
1546 		nbd++;
1547 		third_bd = (struct eth_tx_3rd_bd *)
1548 			qed_chain_produce(&txq->tx_pbl);
1549 		memset(third_bd, 0, sizeof(*third_bd));
1550 
1551 		nbd++;
1552 		/* We need to fill in additional data in second_bd... */
1553 		tx_data_bd = (struct eth_tx_bd *)second_bd;
1554 	}
1555 
1556 	if (skb_vlan_tag_present(skb)) {
1557 		first_bd->data.vlan = cpu_to_le16(skb_vlan_tag_get(skb));
1558 		first_bd->data.bd_flags.bitfields |=
1559 			1 << ETH_TX_1ST_BD_FLAGS_VLAN_INSERTION_SHIFT;
1560 	}
1561 
1562 	/* Fill the parsing flags & params according to the requested offload */
1563 	if (xmit_type & XMIT_L4_CSUM) {
1564 		/* We don't re-calculate IP checksum as it is already done by
1565 		 * the upper stack
1566 		 */
1567 		first_bd->data.bd_flags.bitfields |=
1568 			1 << ETH_TX_1ST_BD_FLAGS_L4_CSUM_SHIFT;
1569 
1570 		if (xmit_type & XMIT_ENC) {
1571 			first_bd->data.bd_flags.bitfields |=
1572 				1 << ETH_TX_1ST_BD_FLAGS_IP_CSUM_SHIFT;
1573 
1574 			val |= (1 << ETH_TX_DATA_1ST_BD_TUNN_FLAG_SHIFT);
1575 		}
1576 
1577 		/* Legacy FW had flipped behavior in regard to this bit -
1578 		 * I.e., needed to set to prevent FW from touching encapsulated
1579 		 * packets when it didn't need to.
1580 		 */
1581 		if (unlikely(txq->is_legacy))
1582 			val ^= (1 << ETH_TX_DATA_1ST_BD_TUNN_FLAG_SHIFT);
1583 
1584 		/* If the packet is IPv6 with extension header, indicate that
1585 		 * to FW and pass few params, since the device cracker doesn't
1586 		 * support parsing IPv6 with extension header/s.
1587 		 */
1588 		if (unlikely(ipv6_ext))
1589 			qede_set_params_for_ipv6_ext(skb, second_bd, third_bd);
1590 	}
1591 
1592 	if (xmit_type & XMIT_LSO) {
1593 		first_bd->data.bd_flags.bitfields |=
1594 			(1 << ETH_TX_1ST_BD_FLAGS_LSO_SHIFT);
1595 		third_bd->data.lso_mss =
1596 			cpu_to_le16(skb_shinfo(skb)->gso_size);
1597 
1598 		if (unlikely(xmit_type & XMIT_ENC)) {
1599 			first_bd->data.bd_flags.bitfields |=
1600 				1 << ETH_TX_1ST_BD_FLAGS_TUNN_IP_CSUM_SHIFT;
1601 
1602 			if (xmit_type & XMIT_ENC_GSO_L4_CSUM) {
1603 				u8 tmp = ETH_TX_1ST_BD_FLAGS_TUNN_L4_CSUM_SHIFT;
1604 
1605 				first_bd->data.bd_flags.bitfields |= 1 << tmp;
1606 			}
1607 			hlen = qede_get_skb_hlen(skb, true);
1608 		} else {
1609 			first_bd->data.bd_flags.bitfields |=
1610 				1 << ETH_TX_1ST_BD_FLAGS_IP_CSUM_SHIFT;
1611 			hlen = qede_get_skb_hlen(skb, false);
1612 		}
1613 
1614 		/* @@@TBD - if will not be removed need to check */
1615 		third_bd->data.bitfields |=
1616 			cpu_to_le16(1 << ETH_TX_DATA_3RD_BD_HDR_NBD_SHIFT);
1617 
1618 		/* Make life easier for FW guys who can't deal with header and
1619 		 * data on same BD. If we need to split, use the second bd...
1620 		 */
1621 		if (unlikely(skb_headlen(skb) > hlen)) {
1622 			DP_VERBOSE(edev, NETIF_MSG_TX_QUEUED,
1623 				   "TSO split header size is %d (%x:%x)\n",
1624 				   first_bd->nbytes, first_bd->addr.hi,
1625 				   first_bd->addr.lo);
1626 
1627 			mapping = HILO_U64(le32_to_cpu(first_bd->addr.hi),
1628 					   le32_to_cpu(first_bd->addr.lo)) +
1629 					   hlen;
1630 
1631 			BD_SET_UNMAP_ADDR_LEN(tx_data_bd, mapping,
1632 					      le16_to_cpu(first_bd->nbytes) -
1633 					      hlen);
1634 
1635 			/* this marks the BD as one that has no
1636 			 * individual mapping
1637 			 */
1638 			txq->sw_tx_ring.skbs[idx].flags |= QEDE_TSO_SPLIT_BD;
1639 
1640 			first_bd->nbytes = cpu_to_le16(hlen);
1641 
1642 			tx_data_bd = (struct eth_tx_bd *)third_bd;
1643 			data_split = true;
1644 		}
1645 	} else {
1646 		val |= ((skb->len & ETH_TX_DATA_1ST_BD_PKT_LEN_MASK) <<
1647 			 ETH_TX_DATA_1ST_BD_PKT_LEN_SHIFT);
1648 	}
1649 
1650 	first_bd->data.bitfields = cpu_to_le16(val);
1651 
1652 	/* Handle fragmented skb */
1653 	/* special handle for frags inside 2nd and 3rd bds.. */
1654 	while (tx_data_bd && frag_idx < skb_shinfo(skb)->nr_frags) {
1655 		rc = map_frag_to_bd(txq,
1656 				    &skb_shinfo(skb)->frags[frag_idx],
1657 				    tx_data_bd);
1658 		if (rc) {
1659 			qede_free_failed_tx_pkt(txq, first_bd, nbd, data_split);
1660 			qede_update_tx_producer(txq);
1661 			return NETDEV_TX_OK;
1662 		}
1663 
1664 		if (tx_data_bd == (struct eth_tx_bd *)second_bd)
1665 			tx_data_bd = (struct eth_tx_bd *)third_bd;
1666 		else
1667 			tx_data_bd = NULL;
1668 
1669 		frag_idx++;
1670 	}
1671 
1672 	/* map last frags into 4th, 5th .... */
1673 	for (; frag_idx < skb_shinfo(skb)->nr_frags; frag_idx++, nbd++) {
1674 		tx_data_bd = (struct eth_tx_bd *)
1675 			     qed_chain_produce(&txq->tx_pbl);
1676 
1677 		memset(tx_data_bd, 0, sizeof(*tx_data_bd));
1678 
1679 		rc = map_frag_to_bd(txq,
1680 				    &skb_shinfo(skb)->frags[frag_idx],
1681 				    tx_data_bd);
1682 		if (rc) {
1683 			qede_free_failed_tx_pkt(txq, first_bd, nbd, data_split);
1684 			qede_update_tx_producer(txq);
1685 			return NETDEV_TX_OK;
1686 		}
1687 	}
1688 
1689 	/* update the first BD with the actual num BDs */
1690 	first_bd->data.nbds = nbd;
1691 
1692 	netdev_tx_sent_queue(netdev_txq, skb->len);
1693 
1694 	skb_tx_timestamp(skb);
1695 
1696 	/* Advance packet producer only before sending the packet since mapping
1697 	 * of pages may fail.
1698 	 */
1699 	txq->sw_tx_prod = (txq->sw_tx_prod + 1) % txq->num_tx_buffers;
1700 
1701 	/* 'next page' entries are counted in the producer value */
1702 	txq->tx_db.data.bd_prod =
1703 		cpu_to_le16(qed_chain_get_prod_idx(&txq->tx_pbl));
1704 
1705 	if (!netdev_xmit_more() || netif_xmit_stopped(netdev_txq))
1706 		qede_update_tx_producer(txq);
1707 
1708 	if (unlikely(qed_chain_get_elem_left(&txq->tx_pbl)
1709 		      < (MAX_SKB_FRAGS + 1))) {
1710 		if (netdev_xmit_more())
1711 			qede_update_tx_producer(txq);
1712 
1713 		netif_tx_stop_queue(netdev_txq);
1714 		txq->stopped_cnt++;
1715 		DP_VERBOSE(edev, NETIF_MSG_TX_QUEUED,
1716 			   "Stop queue was called\n");
1717 		/* paired memory barrier is in qede_tx_int(), we have to keep
1718 		 * ordering of set_bit() in netif_tx_stop_queue() and read of
1719 		 * fp->bd_tx_cons
1720 		 */
1721 		smp_mb();
1722 
1723 		if ((qed_chain_get_elem_left(&txq->tx_pbl) >=
1724 		     (MAX_SKB_FRAGS + 1)) &&
1725 		    (edev->state == QEDE_STATE_OPEN)) {
1726 			netif_tx_wake_queue(netdev_txq);
1727 			DP_VERBOSE(edev, NETIF_MSG_TX_QUEUED,
1728 				   "Wake queue was called\n");
1729 		}
1730 	}
1731 
1732 	return NETDEV_TX_OK;
1733 }
1734 
1735 u16 qede_select_queue(struct net_device *dev, struct sk_buff *skb,
1736 		      struct net_device *sb_dev)
1737 {
1738 	struct qede_dev *edev = netdev_priv(dev);
1739 	int total_txq;
1740 
1741 	total_txq = QEDE_TSS_COUNT(edev) * edev->dev_info.num_tc;
1742 
1743 	return QEDE_TSS_COUNT(edev) ?
1744 		netdev_pick_tx(dev, skb, NULL) % total_txq :  0;
1745 }
1746 
1747 /* 8B udp header + 8B base tunnel header + 32B option length */
1748 #define QEDE_MAX_TUN_HDR_LEN 48
1749 
1750 netdev_features_t qede_features_check(struct sk_buff *skb,
1751 				      struct net_device *dev,
1752 				      netdev_features_t features)
1753 {
1754 	if (skb->encapsulation) {
1755 		u8 l4_proto = 0;
1756 
1757 		switch (vlan_get_protocol(skb)) {
1758 		case htons(ETH_P_IP):
1759 			l4_proto = ip_hdr(skb)->protocol;
1760 			break;
1761 		case htons(ETH_P_IPV6):
1762 			l4_proto = ipv6_hdr(skb)->nexthdr;
1763 			break;
1764 		default:
1765 			return features;
1766 		}
1767 
1768 		/* Disable offloads for geneve tunnels, as HW can't parse
1769 		 * the geneve header which has option length greater than 32b
1770 		 * and disable offloads for the ports which are not offloaded.
1771 		 */
1772 		if (l4_proto == IPPROTO_UDP) {
1773 			struct qede_dev *edev = netdev_priv(dev);
1774 			u16 hdrlen, vxln_port, gnv_port;
1775 
1776 			hdrlen = QEDE_MAX_TUN_HDR_LEN;
1777 			vxln_port = edev->vxlan_dst_port;
1778 			gnv_port = edev->geneve_dst_port;
1779 
1780 			if ((skb_inner_mac_header(skb) -
1781 			     skb_transport_header(skb)) > hdrlen ||
1782 			     (ntohs(udp_hdr(skb)->dest) != vxln_port &&
1783 			      ntohs(udp_hdr(skb)->dest) != gnv_port))
1784 				return features & ~(NETIF_F_CSUM_MASK |
1785 						    NETIF_F_GSO_MASK);
1786 		} else if (l4_proto == IPPROTO_IPIP) {
1787 			/* IPIP tunnels are unknown to the device or at least unsupported natively,
1788 			 * offloads for them can't be done trivially, so disable them for such skb.
1789 			 */
1790 			return features & ~(NETIF_F_CSUM_MASK | NETIF_F_GSO_MASK);
1791 		}
1792 	}
1793 
1794 	return features;
1795 }
1796