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/gro.h>
14 #include <net/ipv6.h>
15 #include <net/tcp.h>
16 #include <linux/if_ether.h>
17 #include <linux/if_vlan.h>
18 #include <net/ip6_checksum.h>
19 #include "qede_ptp.h"
20 
21 #include <linux/qed/qed_if.h>
22 #include "qede.h"
23 /*********************************
24  * Content also used by slowpath *
25  *********************************/
26 
27 int qede_alloc_rx_buffer(struct qede_rx_queue *rxq, bool allow_lazy)
28 {
29 	struct sw_rx_data *sw_rx_data;
30 	struct eth_rx_bd *rx_bd;
31 	dma_addr_t mapping;
32 	struct page *data;
33 
34 	/* In case lazy-allocation is allowed, postpone allocation until the
35 	 * end of the NAPI run. We'd still need to make sure the Rx ring has
36 	 * sufficient buffers to guarantee an additional Rx interrupt.
37 	 */
38 	if (allow_lazy && likely(rxq->filled_buffers > 12)) {
39 		rxq->filled_buffers--;
40 		return 0;
41 	}
42 
43 	data = alloc_pages(GFP_ATOMIC, 0);
44 	if (unlikely(!data))
45 		return -ENOMEM;
46 
47 	/* Map the entire page as it would be used
48 	 * for multiple RX buffer segment size mapping.
49 	 */
50 	mapping = dma_map_page(rxq->dev, data, 0,
51 			       PAGE_SIZE, rxq->data_direction);
52 	if (unlikely(dma_mapping_error(rxq->dev, mapping))) {
53 		__free_page(data);
54 		return -ENOMEM;
55 	}
56 
57 	sw_rx_data = &rxq->sw_rx_ring[rxq->sw_rx_prod & NUM_RX_BDS_MAX];
58 	sw_rx_data->page_offset = 0;
59 	sw_rx_data->data = data;
60 	sw_rx_data->mapping = mapping;
61 
62 	/* Advance PROD and get BD pointer */
63 	rx_bd = (struct eth_rx_bd *)qed_chain_produce(&rxq->rx_bd_ring);
64 	WARN_ON(!rx_bd);
65 	rx_bd->addr.hi = cpu_to_le32(upper_32_bits(mapping));
66 	rx_bd->addr.lo = cpu_to_le32(lower_32_bits(mapping) +
67 				     rxq->rx_headroom);
68 
69 	rxq->sw_rx_prod++;
70 	rxq->filled_buffers++;
71 
72 	return 0;
73 }
74 
75 /* Unmap the data and free skb */
76 int qede_free_tx_pkt(struct qede_dev *edev, struct qede_tx_queue *txq, int *len)
77 {
78 	u16 idx = txq->sw_tx_cons;
79 	struct sk_buff *skb = txq->sw_tx_ring.skbs[idx].skb;
80 	struct eth_tx_1st_bd *first_bd;
81 	struct eth_tx_bd *tx_data_bd;
82 	int bds_consumed = 0;
83 	int nbds;
84 	bool data_split = txq->sw_tx_ring.skbs[idx].flags & QEDE_TSO_SPLIT_BD;
85 	int i, split_bd_len = 0;
86 
87 	if (unlikely(!skb)) {
88 		DP_ERR(edev,
89 		       "skb is null for txq idx=%d txq->sw_tx_cons=%d txq->sw_tx_prod=%d\n",
90 		       idx, txq->sw_tx_cons, txq->sw_tx_prod);
91 		return -1;
92 	}
93 
94 	*len = skb->len;
95 
96 	first_bd = (struct eth_tx_1st_bd *)qed_chain_consume(&txq->tx_pbl);
97 
98 	bds_consumed++;
99 
100 	nbds = first_bd->data.nbds;
101 
102 	if (data_split) {
103 		struct eth_tx_bd *split = (struct eth_tx_bd *)
104 			qed_chain_consume(&txq->tx_pbl);
105 		split_bd_len = BD_UNMAP_LEN(split);
106 		bds_consumed++;
107 	}
108 	dma_unmap_single(&edev->pdev->dev, BD_UNMAP_ADDR(first_bd),
109 			 BD_UNMAP_LEN(first_bd) + split_bd_len, DMA_TO_DEVICE);
110 
111 	/* Unmap the data of the skb frags */
112 	for (i = 0; i < skb_shinfo(skb)->nr_frags; i++, bds_consumed++) {
113 		tx_data_bd = (struct eth_tx_bd *)
114 			qed_chain_consume(&txq->tx_pbl);
115 		dma_unmap_page(&edev->pdev->dev, BD_UNMAP_ADDR(tx_data_bd),
116 			       BD_UNMAP_LEN(tx_data_bd), DMA_TO_DEVICE);
117 	}
118 
119 	while (bds_consumed++ < nbds)
120 		qed_chain_consume(&txq->tx_pbl);
121 
122 	/* Free skb */
123 	dev_kfree_skb_any(skb);
124 	txq->sw_tx_ring.skbs[idx].skb = NULL;
125 	txq->sw_tx_ring.skbs[idx].flags = 0;
126 
127 	return 0;
128 }
129 
130 /* Unmap the data and free skb when mapping failed during start_xmit */
131 static void qede_free_failed_tx_pkt(struct qede_tx_queue *txq,
132 				    struct eth_tx_1st_bd *first_bd,
133 				    int nbd, bool data_split)
134 {
135 	u16 idx = txq->sw_tx_prod;
136 	struct sk_buff *skb = txq->sw_tx_ring.skbs[idx].skb;
137 	struct eth_tx_bd *tx_data_bd;
138 	int i, split_bd_len = 0;
139 
140 	/* Return prod to its position before this skb was handled */
141 	qed_chain_set_prod(&txq->tx_pbl,
142 			   le16_to_cpu(txq->tx_db.data.bd_prod), first_bd);
143 
144 	first_bd = (struct eth_tx_1st_bd *)qed_chain_produce(&txq->tx_pbl);
145 
146 	if (data_split) {
147 		struct eth_tx_bd *split = (struct eth_tx_bd *)
148 					  qed_chain_produce(&txq->tx_pbl);
149 		split_bd_len = BD_UNMAP_LEN(split);
150 		nbd--;
151 	}
152 
153 	dma_unmap_single(txq->dev, BD_UNMAP_ADDR(first_bd),
154 			 BD_UNMAP_LEN(first_bd) + split_bd_len, DMA_TO_DEVICE);
155 
156 	/* Unmap the data of the skb frags */
157 	for (i = 0; i < nbd; i++) {
158 		tx_data_bd = (struct eth_tx_bd *)
159 			qed_chain_produce(&txq->tx_pbl);
160 		if (tx_data_bd->nbytes)
161 			dma_unmap_page(txq->dev,
162 				       BD_UNMAP_ADDR(tx_data_bd),
163 				       BD_UNMAP_LEN(tx_data_bd), DMA_TO_DEVICE);
164 	}
165 
166 	/* Return again prod to its position before this skb was handled */
167 	qed_chain_set_prod(&txq->tx_pbl,
168 			   le16_to_cpu(txq->tx_db.data.bd_prod), first_bd);
169 
170 	/* Free skb */
171 	dev_kfree_skb_any(skb);
172 	txq->sw_tx_ring.skbs[idx].skb = NULL;
173 	txq->sw_tx_ring.skbs[idx].flags = 0;
174 }
175 
176 static u32 qede_xmit_type(struct sk_buff *skb, int *ipv6_ext)
177 {
178 	u32 rc = XMIT_L4_CSUM;
179 	__be16 l3_proto;
180 
181 	if (skb->ip_summed != CHECKSUM_PARTIAL)
182 		return XMIT_PLAIN;
183 
184 	l3_proto = vlan_get_protocol(skb);
185 	if (l3_proto == htons(ETH_P_IPV6) &&
186 	    (ipv6_hdr(skb)->nexthdr == NEXTHDR_IPV6))
187 		*ipv6_ext = 1;
188 
189 	if (skb->encapsulation) {
190 		rc |= XMIT_ENC;
191 		if (skb_is_gso(skb)) {
192 			unsigned short gso_type = skb_shinfo(skb)->gso_type;
193 
194 			if ((gso_type & SKB_GSO_UDP_TUNNEL_CSUM) ||
195 			    (gso_type & SKB_GSO_GRE_CSUM))
196 				rc |= XMIT_ENC_GSO_L4_CSUM;
197 
198 			rc |= XMIT_LSO;
199 			return rc;
200 		}
201 	}
202 
203 	if (skb_is_gso(skb))
204 		rc |= XMIT_LSO;
205 
206 	return rc;
207 }
208 
209 static void qede_set_params_for_ipv6_ext(struct sk_buff *skb,
210 					 struct eth_tx_2nd_bd *second_bd,
211 					 struct eth_tx_3rd_bd *third_bd)
212 {
213 	u8 l4_proto;
214 	u16 bd2_bits1 = 0, bd2_bits2 = 0;
215 
216 	bd2_bits1 |= (1 << ETH_TX_DATA_2ND_BD_IPV6_EXT_SHIFT);
217 
218 	bd2_bits2 |= ((((u8 *)skb_transport_header(skb) - skb->data) >> 1) &
219 		     ETH_TX_DATA_2ND_BD_L4_HDR_START_OFFSET_W_MASK)
220 		    << ETH_TX_DATA_2ND_BD_L4_HDR_START_OFFSET_W_SHIFT;
221 
222 	bd2_bits1 |= (ETH_L4_PSEUDO_CSUM_CORRECT_LENGTH <<
223 		      ETH_TX_DATA_2ND_BD_L4_PSEUDO_CSUM_MODE_SHIFT);
224 
225 	if (vlan_get_protocol(skb) == htons(ETH_P_IPV6))
226 		l4_proto = ipv6_hdr(skb)->nexthdr;
227 	else
228 		l4_proto = ip_hdr(skb)->protocol;
229 
230 	if (l4_proto == IPPROTO_UDP)
231 		bd2_bits1 |= 1 << ETH_TX_DATA_2ND_BD_L4_UDP_SHIFT;
232 
233 	if (third_bd)
234 		third_bd->data.bitfields |=
235 			cpu_to_le16(((tcp_hdrlen(skb) / 4) &
236 				ETH_TX_DATA_3RD_BD_TCP_HDR_LEN_DW_MASK) <<
237 				ETH_TX_DATA_3RD_BD_TCP_HDR_LEN_DW_SHIFT);
238 
239 	second_bd->data.bitfields1 = cpu_to_le16(bd2_bits1);
240 	second_bd->data.bitfields2 = cpu_to_le16(bd2_bits2);
241 }
242 
243 static int map_frag_to_bd(struct qede_tx_queue *txq,
244 			  skb_frag_t *frag, struct eth_tx_bd *bd)
245 {
246 	dma_addr_t mapping;
247 
248 	/* Map skb non-linear frag data for DMA */
249 	mapping = skb_frag_dma_map(txq->dev, frag, 0,
250 				   skb_frag_size(frag), DMA_TO_DEVICE);
251 	if (unlikely(dma_mapping_error(txq->dev, mapping)))
252 		return -ENOMEM;
253 
254 	/* Setup the data pointer of the frag data */
255 	BD_SET_UNMAP_ADDR_LEN(bd, mapping, skb_frag_size(frag));
256 
257 	return 0;
258 }
259 
260 static u16 qede_get_skb_hlen(struct sk_buff *skb, bool is_encap_pkt)
261 {
262 	if (is_encap_pkt)
263 		return (skb_inner_transport_header(skb) +
264 			inner_tcp_hdrlen(skb) - skb->data);
265 	else
266 		return (skb_transport_header(skb) +
267 			tcp_hdrlen(skb) - skb->data);
268 }
269 
270 /* +2 for 1st BD for headers and 2nd BD for headlen (if required) */
271 #if ((MAX_SKB_FRAGS + 2) > ETH_TX_MAX_BDS_PER_NON_LSO_PACKET)
272 static bool qede_pkt_req_lin(struct sk_buff *skb, u8 xmit_type)
273 {
274 	int allowed_frags = ETH_TX_MAX_BDS_PER_NON_LSO_PACKET - 1;
275 
276 	if (xmit_type & XMIT_LSO) {
277 		int hlen;
278 
279 		hlen = qede_get_skb_hlen(skb, xmit_type & XMIT_ENC);
280 
281 		/* linear payload would require its own BD */
282 		if (skb_headlen(skb) > hlen)
283 			allowed_frags--;
284 	}
285 
286 	return (skb_shinfo(skb)->nr_frags > allowed_frags);
287 }
288 #endif
289 
290 static inline void qede_update_tx_producer(struct qede_tx_queue *txq)
291 {
292 	/* wmb makes sure that the BDs data is updated before updating the
293 	 * producer, otherwise FW may read old data from the BDs.
294 	 */
295 	wmb();
296 	barrier();
297 	writel(txq->tx_db.raw, txq->doorbell_addr);
298 
299 	/* Fence required to flush the write combined buffer, since another
300 	 * CPU may write to the same doorbell address and data may be lost
301 	 * due to relaxed order nature of write combined bar.
302 	 */
303 	wmb();
304 }
305 
306 static int qede_xdp_xmit(struct qede_tx_queue *txq, dma_addr_t dma, u16 pad,
307 			 u16 len, struct page *page, struct xdp_frame *xdpf)
308 {
309 	struct eth_tx_1st_bd *bd;
310 	struct sw_tx_xdp *xdp;
311 	u16 val;
312 
313 	if (unlikely(qed_chain_get_elem_used(&txq->tx_pbl) >=
314 		     txq->num_tx_buffers)) {
315 		txq->stopped_cnt++;
316 		return -ENOMEM;
317 	}
318 
319 	bd = qed_chain_produce(&txq->tx_pbl);
320 	bd->data.nbds = 1;
321 	bd->data.bd_flags.bitfields = BIT(ETH_TX_1ST_BD_FLAGS_START_BD_SHIFT);
322 
323 	val = (len & ETH_TX_DATA_1ST_BD_PKT_LEN_MASK) <<
324 	       ETH_TX_DATA_1ST_BD_PKT_LEN_SHIFT;
325 
326 	bd->data.bitfields = cpu_to_le16(val);
327 
328 	/* We can safely ignore the offset, as it's 0 for XDP */
329 	BD_SET_UNMAP_ADDR_LEN(bd, dma + pad, len);
330 
331 	xdp = txq->sw_tx_ring.xdp + txq->sw_tx_prod;
332 	xdp->mapping = dma;
333 	xdp->page = page;
334 	xdp->xdpf = xdpf;
335 
336 	txq->sw_tx_prod = (txq->sw_tx_prod + 1) % txq->num_tx_buffers;
337 
338 	return 0;
339 }
340 
341 int qede_xdp_transmit(struct net_device *dev, int n_frames,
342 		      struct xdp_frame **frames, u32 flags)
343 {
344 	struct qede_dev *edev = netdev_priv(dev);
345 	struct device *dmadev = &edev->pdev->dev;
346 	struct qede_tx_queue *xdp_tx;
347 	struct xdp_frame *xdpf;
348 	dma_addr_t mapping;
349 	int i, nxmit = 0;
350 	u16 xdp_prod;
351 
352 	if (unlikely(flags & ~XDP_XMIT_FLAGS_MASK))
353 		return -EINVAL;
354 
355 	if (unlikely(!netif_running(dev)))
356 		return -ENETDOWN;
357 
358 	i = smp_processor_id() % edev->total_xdp_queues;
359 	xdp_tx = edev->fp_array[i].xdp_tx;
360 
361 	spin_lock(&xdp_tx->xdp_tx_lock);
362 
363 	for (i = 0; i < n_frames; i++) {
364 		xdpf = frames[i];
365 
366 		mapping = dma_map_single(dmadev, xdpf->data, xdpf->len,
367 					 DMA_TO_DEVICE);
368 		if (unlikely(dma_mapping_error(dmadev, mapping)))
369 			break;
370 
371 		if (unlikely(qede_xdp_xmit(xdp_tx, mapping, 0, xdpf->len,
372 					   NULL, xdpf)))
373 			break;
374 		nxmit++;
375 	}
376 
377 	if (flags & XDP_XMIT_FLUSH) {
378 		xdp_prod = qed_chain_get_prod_idx(&xdp_tx->tx_pbl);
379 
380 		xdp_tx->tx_db.data.bd_prod = cpu_to_le16(xdp_prod);
381 		qede_update_tx_producer(xdp_tx);
382 	}
383 
384 	spin_unlock(&xdp_tx->xdp_tx_lock);
385 
386 	return nxmit;
387 }
388 
389 int qede_txq_has_work(struct qede_tx_queue *txq)
390 {
391 	u16 hw_bd_cons;
392 
393 	/* Tell compiler that consumer and producer can change */
394 	barrier();
395 	hw_bd_cons = le16_to_cpu(*txq->hw_cons_ptr);
396 	if (qed_chain_get_cons_idx(&txq->tx_pbl) == hw_bd_cons + 1)
397 		return 0;
398 
399 	return hw_bd_cons != qed_chain_get_cons_idx(&txq->tx_pbl);
400 }
401 
402 static void qede_xdp_tx_int(struct qede_dev *edev, struct qede_tx_queue *txq)
403 {
404 	struct sw_tx_xdp *xdp_info, *xdp_arr = txq->sw_tx_ring.xdp;
405 	struct device *dev = &edev->pdev->dev;
406 	struct xdp_frame *xdpf;
407 	u16 hw_bd_cons;
408 
409 	hw_bd_cons = le16_to_cpu(*txq->hw_cons_ptr);
410 	barrier();
411 
412 	while (hw_bd_cons != qed_chain_get_cons_idx(&txq->tx_pbl)) {
413 		xdp_info = xdp_arr + txq->sw_tx_cons;
414 		xdpf = xdp_info->xdpf;
415 
416 		if (xdpf) {
417 			dma_unmap_single(dev, xdp_info->mapping, xdpf->len,
418 					 DMA_TO_DEVICE);
419 			xdp_return_frame(xdpf);
420 
421 			xdp_info->xdpf = NULL;
422 		} else {
423 			dma_unmap_page(dev, xdp_info->mapping, PAGE_SIZE,
424 				       DMA_BIDIRECTIONAL);
425 			__free_page(xdp_info->page);
426 		}
427 
428 		qed_chain_consume(&txq->tx_pbl);
429 		txq->sw_tx_cons = (txq->sw_tx_cons + 1) % txq->num_tx_buffers;
430 		txq->xmit_pkts++;
431 	}
432 }
433 
434 static int qede_tx_int(struct qede_dev *edev, struct qede_tx_queue *txq)
435 {
436 	unsigned int pkts_compl = 0, bytes_compl = 0;
437 	struct netdev_queue *netdev_txq;
438 	u16 hw_bd_cons;
439 	int rc;
440 
441 	netdev_txq = netdev_get_tx_queue(edev->ndev, txq->ndev_txq_id);
442 
443 	hw_bd_cons = le16_to_cpu(*txq->hw_cons_ptr);
444 	barrier();
445 
446 	while (hw_bd_cons != qed_chain_get_cons_idx(&txq->tx_pbl)) {
447 		int len = 0;
448 
449 		rc = qede_free_tx_pkt(edev, txq, &len);
450 		if (rc) {
451 			DP_NOTICE(edev, "hw_bd_cons = %d, chain_cons=%d\n",
452 				  hw_bd_cons,
453 				  qed_chain_get_cons_idx(&txq->tx_pbl));
454 			break;
455 		}
456 
457 		bytes_compl += len;
458 		pkts_compl++;
459 		txq->sw_tx_cons = (txq->sw_tx_cons + 1) % txq->num_tx_buffers;
460 		txq->xmit_pkts++;
461 	}
462 
463 	netdev_tx_completed_queue(netdev_txq, pkts_compl, bytes_compl);
464 
465 	/* Need to make the tx_bd_cons update visible to start_xmit()
466 	 * before checking for netif_tx_queue_stopped().  Without the
467 	 * memory barrier, there is a small possibility that
468 	 * start_xmit() will miss it and cause the queue to be stopped
469 	 * forever.
470 	 * On the other hand we need an rmb() here to ensure the proper
471 	 * ordering of bit testing in the following
472 	 * netif_tx_queue_stopped(txq) call.
473 	 */
474 	smp_mb();
475 
476 	if (unlikely(netif_tx_queue_stopped(netdev_txq))) {
477 		/* Taking tx_lock is needed to prevent reenabling the queue
478 		 * while it's empty. This could have happen if rx_action() gets
479 		 * suspended in qede_tx_int() after the condition before
480 		 * netif_tx_wake_queue(), while tx_action (qede_start_xmit()):
481 		 *
482 		 * stops the queue->sees fresh tx_bd_cons->releases the queue->
483 		 * sends some packets consuming the whole queue again->
484 		 * stops the queue
485 		 */
486 
487 		__netif_tx_lock(netdev_txq, smp_processor_id());
488 
489 		if ((netif_tx_queue_stopped(netdev_txq)) &&
490 		    (edev->state == QEDE_STATE_OPEN) &&
491 		    (qed_chain_get_elem_left(&txq->tx_pbl)
492 		      >= (MAX_SKB_FRAGS + 1))) {
493 			netif_tx_wake_queue(netdev_txq);
494 			DP_VERBOSE(edev, NETIF_MSG_TX_DONE,
495 				   "Wake queue was called\n");
496 		}
497 
498 		__netif_tx_unlock(netdev_txq);
499 	}
500 
501 	return 0;
502 }
503 
504 bool qede_has_rx_work(struct qede_rx_queue *rxq)
505 {
506 	u16 hw_comp_cons, sw_comp_cons;
507 
508 	/* Tell compiler that status block fields can change */
509 	barrier();
510 
511 	hw_comp_cons = le16_to_cpu(*rxq->hw_cons_ptr);
512 	sw_comp_cons = qed_chain_get_cons_idx(&rxq->rx_comp_ring);
513 
514 	return hw_comp_cons != sw_comp_cons;
515 }
516 
517 static inline void qede_rx_bd_ring_consume(struct qede_rx_queue *rxq)
518 {
519 	qed_chain_consume(&rxq->rx_bd_ring);
520 	rxq->sw_rx_cons++;
521 }
522 
523 /* This function reuses the buffer(from an offset) from
524  * consumer index to producer index in the bd ring
525  */
526 static inline void qede_reuse_page(struct qede_rx_queue *rxq,
527 				   struct sw_rx_data *curr_cons)
528 {
529 	struct eth_rx_bd *rx_bd_prod = qed_chain_produce(&rxq->rx_bd_ring);
530 	struct sw_rx_data *curr_prod;
531 	dma_addr_t new_mapping;
532 
533 	curr_prod = &rxq->sw_rx_ring[rxq->sw_rx_prod & NUM_RX_BDS_MAX];
534 	*curr_prod = *curr_cons;
535 
536 	new_mapping = curr_prod->mapping + curr_prod->page_offset;
537 
538 	rx_bd_prod->addr.hi = cpu_to_le32(upper_32_bits(new_mapping));
539 	rx_bd_prod->addr.lo = cpu_to_le32(lower_32_bits(new_mapping) +
540 					  rxq->rx_headroom);
541 
542 	rxq->sw_rx_prod++;
543 	curr_cons->data = NULL;
544 }
545 
546 /* In case of allocation failures reuse buffers
547  * from consumer index to produce buffers for firmware
548  */
549 void qede_recycle_rx_bd_ring(struct qede_rx_queue *rxq, u8 count)
550 {
551 	struct sw_rx_data *curr_cons;
552 
553 	for (; count > 0; count--) {
554 		curr_cons = &rxq->sw_rx_ring[rxq->sw_rx_cons & NUM_RX_BDS_MAX];
555 		qede_reuse_page(rxq, curr_cons);
556 		qede_rx_bd_ring_consume(rxq);
557 	}
558 }
559 
560 static inline int qede_realloc_rx_buffer(struct qede_rx_queue *rxq,
561 					 struct sw_rx_data *curr_cons)
562 {
563 	/* Move to the next segment in the page */
564 	curr_cons->page_offset += rxq->rx_buf_seg_size;
565 
566 	if (curr_cons->page_offset == PAGE_SIZE) {
567 		if (unlikely(qede_alloc_rx_buffer(rxq, true))) {
568 			/* Since we failed to allocate new buffer
569 			 * current buffer can be used again.
570 			 */
571 			curr_cons->page_offset -= rxq->rx_buf_seg_size;
572 
573 			return -ENOMEM;
574 		}
575 
576 		dma_unmap_page(rxq->dev, curr_cons->mapping,
577 			       PAGE_SIZE, rxq->data_direction);
578 	} else {
579 		/* Increment refcount of the page as we don't want
580 		 * network stack to take the ownership of the page
581 		 * which can be recycled multiple times by the driver.
582 		 */
583 		page_ref_inc(curr_cons->data);
584 		qede_reuse_page(rxq, curr_cons);
585 	}
586 
587 	return 0;
588 }
589 
590 void qede_update_rx_prod(struct qede_dev *edev, struct qede_rx_queue *rxq)
591 {
592 	u16 bd_prod = qed_chain_get_prod_idx(&rxq->rx_bd_ring);
593 	u16 cqe_prod = qed_chain_get_prod_idx(&rxq->rx_comp_ring);
594 	struct eth_rx_prod_data rx_prods = {0};
595 
596 	/* Update producers */
597 	rx_prods.bd_prod = cpu_to_le16(bd_prod);
598 	rx_prods.cqe_prod = cpu_to_le16(cqe_prod);
599 
600 	/* Make sure that the BD and SGE data is updated before updating the
601 	 * producers since FW might read the BD/SGE right after the producer
602 	 * is updated.
603 	 */
604 	wmb();
605 
606 	internal_ram_wr(rxq->hw_rxq_prod_addr, sizeof(rx_prods),
607 			(u32 *)&rx_prods);
608 }
609 
610 static void qede_get_rxhash(struct sk_buff *skb, u8 bitfields, __le32 rss_hash)
611 {
612 	enum pkt_hash_types hash_type = PKT_HASH_TYPE_NONE;
613 	enum rss_hash_type htype;
614 	u32 hash = 0;
615 
616 	htype = GET_FIELD(bitfields, ETH_FAST_PATH_RX_REG_CQE_RSS_HASH_TYPE);
617 	if (htype) {
618 		hash_type = ((htype == RSS_HASH_TYPE_IPV4) ||
619 			     (htype == RSS_HASH_TYPE_IPV6)) ?
620 			    PKT_HASH_TYPE_L3 : PKT_HASH_TYPE_L4;
621 		hash = le32_to_cpu(rss_hash);
622 	}
623 	skb_set_hash(skb, hash, hash_type);
624 }
625 
626 static void qede_set_skb_csum(struct sk_buff *skb, u8 csum_flag)
627 {
628 	skb_checksum_none_assert(skb);
629 
630 	if (csum_flag & QEDE_CSUM_UNNECESSARY)
631 		skb->ip_summed = CHECKSUM_UNNECESSARY;
632 
633 	if (csum_flag & QEDE_TUNN_CSUM_UNNECESSARY) {
634 		skb->csum_level = 1;
635 		skb->encapsulation = 1;
636 	}
637 }
638 
639 static inline void qede_skb_receive(struct qede_dev *edev,
640 				    struct qede_fastpath *fp,
641 				    struct qede_rx_queue *rxq,
642 				    struct sk_buff *skb, u16 vlan_tag)
643 {
644 	if (vlan_tag)
645 		__vlan_hwaccel_put_tag(skb, htons(ETH_P_8021Q), vlan_tag);
646 
647 	napi_gro_receive(&fp->napi, skb);
648 }
649 
650 static void qede_set_gro_params(struct qede_dev *edev,
651 				struct sk_buff *skb,
652 				struct eth_fast_path_rx_tpa_start_cqe *cqe)
653 {
654 	u16 parsing_flags = le16_to_cpu(cqe->pars_flags.flags);
655 
656 	if (((parsing_flags >> PARSING_AND_ERR_FLAGS_L3TYPE_SHIFT) &
657 	    PARSING_AND_ERR_FLAGS_L3TYPE_MASK) == 2)
658 		skb_shinfo(skb)->gso_type = SKB_GSO_TCPV6;
659 	else
660 		skb_shinfo(skb)->gso_type = SKB_GSO_TCPV4;
661 
662 	skb_shinfo(skb)->gso_size = __le16_to_cpu(cqe->len_on_first_bd) -
663 				    cqe->header_len;
664 }
665 
666 static int qede_fill_frag_skb(struct qede_dev *edev,
667 			      struct qede_rx_queue *rxq,
668 			      u8 tpa_agg_index, u16 len_on_bd)
669 {
670 	struct sw_rx_data *current_bd = &rxq->sw_rx_ring[rxq->sw_rx_cons &
671 							 NUM_RX_BDS_MAX];
672 	struct qede_agg_info *tpa_info = &rxq->tpa_info[tpa_agg_index];
673 	struct sk_buff *skb = tpa_info->skb;
674 
675 	if (unlikely(tpa_info->state != QEDE_AGG_STATE_START))
676 		goto out;
677 
678 	/* Add one frag and update the appropriate fields in the skb */
679 	skb_fill_page_desc(skb, tpa_info->frag_id++,
680 			   current_bd->data,
681 			   current_bd->page_offset + rxq->rx_headroom,
682 			   len_on_bd);
683 
684 	if (unlikely(qede_realloc_rx_buffer(rxq, current_bd))) {
685 		/* Incr page ref count to reuse on allocation failure
686 		 * so that it doesn't get freed while freeing SKB.
687 		 */
688 		page_ref_inc(current_bd->data);
689 		goto out;
690 	}
691 
692 	qede_rx_bd_ring_consume(rxq);
693 
694 	skb->data_len += len_on_bd;
695 	skb->truesize += rxq->rx_buf_seg_size;
696 	skb->len += len_on_bd;
697 
698 	return 0;
699 
700 out:
701 	tpa_info->state = QEDE_AGG_STATE_ERROR;
702 	qede_recycle_rx_bd_ring(rxq, 1);
703 
704 	return -ENOMEM;
705 }
706 
707 static bool qede_tunn_exist(u16 flag)
708 {
709 	return !!(flag & (PARSING_AND_ERR_FLAGS_TUNNELEXIST_MASK <<
710 			  PARSING_AND_ERR_FLAGS_TUNNELEXIST_SHIFT));
711 }
712 
713 static u8 qede_check_tunn_csum(u16 flag)
714 {
715 	u16 csum_flag = 0;
716 	u8 tcsum = 0;
717 
718 	if (flag & (PARSING_AND_ERR_FLAGS_TUNNELL4CHKSMWASCALCULATED_MASK <<
719 		    PARSING_AND_ERR_FLAGS_TUNNELL4CHKSMWASCALCULATED_SHIFT))
720 		csum_flag |= PARSING_AND_ERR_FLAGS_TUNNELL4CHKSMERROR_MASK <<
721 			     PARSING_AND_ERR_FLAGS_TUNNELL4CHKSMERROR_SHIFT;
722 
723 	if (flag & (PARSING_AND_ERR_FLAGS_L4CHKSMWASCALCULATED_MASK <<
724 		    PARSING_AND_ERR_FLAGS_L4CHKSMWASCALCULATED_SHIFT)) {
725 		csum_flag |= PARSING_AND_ERR_FLAGS_L4CHKSMERROR_MASK <<
726 			     PARSING_AND_ERR_FLAGS_L4CHKSMERROR_SHIFT;
727 		tcsum = QEDE_TUNN_CSUM_UNNECESSARY;
728 	}
729 
730 	csum_flag |= PARSING_AND_ERR_FLAGS_TUNNELIPHDRERROR_MASK <<
731 		     PARSING_AND_ERR_FLAGS_TUNNELIPHDRERROR_SHIFT |
732 		     PARSING_AND_ERR_FLAGS_IPHDRERROR_MASK <<
733 		     PARSING_AND_ERR_FLAGS_IPHDRERROR_SHIFT;
734 
735 	if (csum_flag & flag)
736 		return QEDE_CSUM_ERROR;
737 
738 	return QEDE_CSUM_UNNECESSARY | tcsum;
739 }
740 
741 static inline struct sk_buff *
742 qede_build_skb(struct qede_rx_queue *rxq,
743 	       struct sw_rx_data *bd, u16 len, u16 pad)
744 {
745 	struct sk_buff *skb;
746 	void *buf;
747 
748 	buf = page_address(bd->data) + bd->page_offset;
749 	skb = build_skb(buf, rxq->rx_buf_seg_size);
750 
751 	if (unlikely(!skb))
752 		return NULL;
753 
754 	skb_reserve(skb, pad);
755 	skb_put(skb, len);
756 
757 	return skb;
758 }
759 
760 static struct sk_buff *
761 qede_tpa_rx_build_skb(struct qede_dev *edev,
762 		      struct qede_rx_queue *rxq,
763 		      struct sw_rx_data *bd, u16 len, u16 pad,
764 		      bool alloc_skb)
765 {
766 	struct sk_buff *skb;
767 
768 	skb = qede_build_skb(rxq, bd, len, pad);
769 	bd->page_offset += rxq->rx_buf_seg_size;
770 
771 	if (bd->page_offset == PAGE_SIZE) {
772 		if (unlikely(qede_alloc_rx_buffer(rxq, true))) {
773 			DP_NOTICE(edev,
774 				  "Failed to allocate RX buffer for tpa start\n");
775 			bd->page_offset -= rxq->rx_buf_seg_size;
776 			page_ref_inc(bd->data);
777 			dev_kfree_skb_any(skb);
778 			return NULL;
779 		}
780 	} else {
781 		page_ref_inc(bd->data);
782 		qede_reuse_page(rxq, bd);
783 	}
784 
785 	/* We've consumed the first BD and prepared an SKB */
786 	qede_rx_bd_ring_consume(rxq);
787 
788 	return skb;
789 }
790 
791 static struct sk_buff *
792 qede_rx_build_skb(struct qede_dev *edev,
793 		  struct qede_rx_queue *rxq,
794 		  struct sw_rx_data *bd, u16 len, u16 pad)
795 {
796 	struct sk_buff *skb = NULL;
797 
798 	/* For smaller frames still need to allocate skb, memcpy
799 	 * data and benefit in reusing the page segment instead of
800 	 * un-mapping it.
801 	 */
802 	if ((len + pad <= edev->rx_copybreak)) {
803 		unsigned int offset = bd->page_offset + pad;
804 
805 		skb = netdev_alloc_skb(edev->ndev, QEDE_RX_HDR_SIZE);
806 		if (unlikely(!skb))
807 			return NULL;
808 
809 		skb_reserve(skb, pad);
810 		skb_put_data(skb, page_address(bd->data) + offset, len);
811 		qede_reuse_page(rxq, bd);
812 		goto out;
813 	}
814 
815 	skb = qede_build_skb(rxq, bd, len, pad);
816 
817 	if (unlikely(qede_realloc_rx_buffer(rxq, bd))) {
818 		/* Incr page ref count to reuse on allocation failure so
819 		 * that it doesn't get freed while freeing SKB [as its
820 		 * already mapped there].
821 		 */
822 		page_ref_inc(bd->data);
823 		dev_kfree_skb_any(skb);
824 		return NULL;
825 	}
826 out:
827 	/* We've consumed the first BD and prepared an SKB */
828 	qede_rx_bd_ring_consume(rxq);
829 
830 	return skb;
831 }
832 
833 static void qede_tpa_start(struct qede_dev *edev,
834 			   struct qede_rx_queue *rxq,
835 			   struct eth_fast_path_rx_tpa_start_cqe *cqe)
836 {
837 	struct qede_agg_info *tpa_info = &rxq->tpa_info[cqe->tpa_agg_index];
838 	struct sw_rx_data *sw_rx_data_cons;
839 	u16 pad;
840 
841 	sw_rx_data_cons = &rxq->sw_rx_ring[rxq->sw_rx_cons & NUM_RX_BDS_MAX];
842 	pad = cqe->placement_offset + rxq->rx_headroom;
843 
844 	tpa_info->skb = qede_tpa_rx_build_skb(edev, rxq, sw_rx_data_cons,
845 					      le16_to_cpu(cqe->len_on_first_bd),
846 					      pad, false);
847 	tpa_info->buffer.page_offset = sw_rx_data_cons->page_offset;
848 	tpa_info->buffer.mapping = sw_rx_data_cons->mapping;
849 
850 	if (unlikely(!tpa_info->skb)) {
851 		DP_NOTICE(edev, "Failed to allocate SKB for gro\n");
852 
853 		/* Consume from ring but do not produce since
854 		 * this might be used by FW still, it will be re-used
855 		 * at TPA end.
856 		 */
857 		tpa_info->tpa_start_fail = true;
858 		qede_rx_bd_ring_consume(rxq);
859 		tpa_info->state = QEDE_AGG_STATE_ERROR;
860 		goto cons_buf;
861 	}
862 
863 	tpa_info->frag_id = 0;
864 	tpa_info->state = QEDE_AGG_STATE_START;
865 
866 	if ((le16_to_cpu(cqe->pars_flags.flags) >>
867 	     PARSING_AND_ERR_FLAGS_TAG8021QEXIST_SHIFT) &
868 	    PARSING_AND_ERR_FLAGS_TAG8021QEXIST_MASK)
869 		tpa_info->vlan_tag = le16_to_cpu(cqe->vlan_tag);
870 	else
871 		tpa_info->vlan_tag = 0;
872 
873 	qede_get_rxhash(tpa_info->skb, cqe->bitfields, cqe->rss_hash);
874 
875 	/* This is needed in order to enable forwarding support */
876 	qede_set_gro_params(edev, tpa_info->skb, cqe);
877 
878 cons_buf: /* We still need to handle bd_len_list to consume buffers */
879 	if (likely(cqe->bw_ext_bd_len_list[0]))
880 		qede_fill_frag_skb(edev, rxq, cqe->tpa_agg_index,
881 				   le16_to_cpu(cqe->bw_ext_bd_len_list[0]));
882 
883 	if (unlikely(cqe->bw_ext_bd_len_list[1])) {
884 		DP_ERR(edev,
885 		       "Unlikely - got a TPA aggregation with more than one bw_ext_bd_len_list entry in the TPA start\n");
886 		tpa_info->state = QEDE_AGG_STATE_ERROR;
887 	}
888 }
889 
890 #ifdef CONFIG_INET
891 static void qede_gro_ip_csum(struct sk_buff *skb)
892 {
893 	const struct iphdr *iph = ip_hdr(skb);
894 	struct tcphdr *th;
895 
896 	skb_set_transport_header(skb, sizeof(struct iphdr));
897 	th = tcp_hdr(skb);
898 
899 	th->check = ~tcp_v4_check(skb->len - skb_transport_offset(skb),
900 				  iph->saddr, iph->daddr, 0);
901 
902 	tcp_gro_complete(skb);
903 }
904 
905 static void qede_gro_ipv6_csum(struct sk_buff *skb)
906 {
907 	struct ipv6hdr *iph = ipv6_hdr(skb);
908 	struct tcphdr *th;
909 
910 	skb_set_transport_header(skb, sizeof(struct ipv6hdr));
911 	th = tcp_hdr(skb);
912 
913 	th->check = ~tcp_v6_check(skb->len - skb_transport_offset(skb),
914 				  &iph->saddr, &iph->daddr, 0);
915 	tcp_gro_complete(skb);
916 }
917 #endif
918 
919 static void qede_gro_receive(struct qede_dev *edev,
920 			     struct qede_fastpath *fp,
921 			     struct sk_buff *skb,
922 			     u16 vlan_tag)
923 {
924 	/* FW can send a single MTU sized packet from gro flow
925 	 * due to aggregation timeout/last segment etc. which
926 	 * is not expected to be a gro packet. If a skb has zero
927 	 * frags then simply push it in the stack as non gso skb.
928 	 */
929 	if (unlikely(!skb->data_len)) {
930 		skb_shinfo(skb)->gso_type = 0;
931 		skb_shinfo(skb)->gso_size = 0;
932 		goto send_skb;
933 	}
934 
935 #ifdef CONFIG_INET
936 	if (skb_shinfo(skb)->gso_size) {
937 		skb_reset_network_header(skb);
938 
939 		switch (skb->protocol) {
940 		case htons(ETH_P_IP):
941 			qede_gro_ip_csum(skb);
942 			break;
943 		case htons(ETH_P_IPV6):
944 			qede_gro_ipv6_csum(skb);
945 			break;
946 		default:
947 			DP_ERR(edev,
948 			       "Error: FW GRO supports only IPv4/IPv6, not 0x%04x\n",
949 			       ntohs(skb->protocol));
950 		}
951 	}
952 #endif
953 
954 send_skb:
955 	skb_record_rx_queue(skb, fp->rxq->rxq_id);
956 	qede_skb_receive(edev, fp, fp->rxq, skb, vlan_tag);
957 }
958 
959 static inline void qede_tpa_cont(struct qede_dev *edev,
960 				 struct qede_rx_queue *rxq,
961 				 struct eth_fast_path_rx_tpa_cont_cqe *cqe)
962 {
963 	int i;
964 
965 	for (i = 0; cqe->len_list[i]; i++)
966 		qede_fill_frag_skb(edev, rxq, cqe->tpa_agg_index,
967 				   le16_to_cpu(cqe->len_list[i]));
968 
969 	if (unlikely(i > 1))
970 		DP_ERR(edev,
971 		       "Strange - TPA cont with more than a single len_list entry\n");
972 }
973 
974 static int qede_tpa_end(struct qede_dev *edev,
975 			struct qede_fastpath *fp,
976 			struct eth_fast_path_rx_tpa_end_cqe *cqe)
977 {
978 	struct qede_rx_queue *rxq = fp->rxq;
979 	struct qede_agg_info *tpa_info;
980 	struct sk_buff *skb;
981 	int i;
982 
983 	tpa_info = &rxq->tpa_info[cqe->tpa_agg_index];
984 	skb = tpa_info->skb;
985 
986 	if (tpa_info->buffer.page_offset == PAGE_SIZE)
987 		dma_unmap_page(rxq->dev, tpa_info->buffer.mapping,
988 			       PAGE_SIZE, rxq->data_direction);
989 
990 	for (i = 0; cqe->len_list[i]; i++)
991 		qede_fill_frag_skb(edev, rxq, cqe->tpa_agg_index,
992 				   le16_to_cpu(cqe->len_list[i]));
993 	if (unlikely(i > 1))
994 		DP_ERR(edev,
995 		       "Strange - TPA emd with more than a single len_list entry\n");
996 
997 	if (unlikely(tpa_info->state != QEDE_AGG_STATE_START))
998 		goto err;
999 
1000 	/* Sanity */
1001 	if (unlikely(cqe->num_of_bds != tpa_info->frag_id + 1))
1002 		DP_ERR(edev,
1003 		       "Strange - TPA had %02x BDs, but SKB has only %d frags\n",
1004 		       cqe->num_of_bds, tpa_info->frag_id);
1005 	if (unlikely(skb->len != le16_to_cpu(cqe->total_packet_len)))
1006 		DP_ERR(edev,
1007 		       "Strange - total packet len [cqe] is %4x but SKB has len %04x\n",
1008 		       le16_to_cpu(cqe->total_packet_len), skb->len);
1009 
1010 	/* Finalize the SKB */
1011 	skb->protocol = eth_type_trans(skb, edev->ndev);
1012 	skb->ip_summed = CHECKSUM_UNNECESSARY;
1013 
1014 	/* tcp_gro_complete() will copy NAPI_GRO_CB(skb)->count
1015 	 * to skb_shinfo(skb)->gso_segs
1016 	 */
1017 	NAPI_GRO_CB(skb)->count = le16_to_cpu(cqe->num_of_coalesced_segs);
1018 
1019 	qede_gro_receive(edev, fp, skb, tpa_info->vlan_tag);
1020 
1021 	tpa_info->state = QEDE_AGG_STATE_NONE;
1022 
1023 	return 1;
1024 err:
1025 	tpa_info->state = QEDE_AGG_STATE_NONE;
1026 
1027 	if (tpa_info->tpa_start_fail) {
1028 		qede_reuse_page(rxq, &tpa_info->buffer);
1029 		tpa_info->tpa_start_fail = false;
1030 	}
1031 
1032 	dev_kfree_skb_any(tpa_info->skb);
1033 	tpa_info->skb = NULL;
1034 	return 0;
1035 }
1036 
1037 static u8 qede_check_notunn_csum(u16 flag)
1038 {
1039 	u16 csum_flag = 0;
1040 	u8 csum = 0;
1041 
1042 	if (flag & (PARSING_AND_ERR_FLAGS_L4CHKSMWASCALCULATED_MASK <<
1043 		    PARSING_AND_ERR_FLAGS_L4CHKSMWASCALCULATED_SHIFT)) {
1044 		csum_flag |= PARSING_AND_ERR_FLAGS_L4CHKSMERROR_MASK <<
1045 			     PARSING_AND_ERR_FLAGS_L4CHKSMERROR_SHIFT;
1046 		csum = QEDE_CSUM_UNNECESSARY;
1047 	}
1048 
1049 	csum_flag |= PARSING_AND_ERR_FLAGS_IPHDRERROR_MASK <<
1050 		     PARSING_AND_ERR_FLAGS_IPHDRERROR_SHIFT;
1051 
1052 	if (csum_flag & flag)
1053 		return QEDE_CSUM_ERROR;
1054 
1055 	return csum;
1056 }
1057 
1058 static u8 qede_check_csum(u16 flag)
1059 {
1060 	if (!qede_tunn_exist(flag))
1061 		return qede_check_notunn_csum(flag);
1062 	else
1063 		return qede_check_tunn_csum(flag);
1064 }
1065 
1066 static bool qede_pkt_is_ip_fragmented(struct eth_fast_path_rx_reg_cqe *cqe,
1067 				      u16 flag)
1068 {
1069 	u8 tun_pars_flg = cqe->tunnel_pars_flags.flags;
1070 
1071 	if ((tun_pars_flg & (ETH_TUNNEL_PARSING_FLAGS_IPV4_FRAGMENT_MASK <<
1072 			     ETH_TUNNEL_PARSING_FLAGS_IPV4_FRAGMENT_SHIFT)) ||
1073 	    (flag & (PARSING_AND_ERR_FLAGS_IPV4FRAG_MASK <<
1074 		     PARSING_AND_ERR_FLAGS_IPV4FRAG_SHIFT)))
1075 		return true;
1076 
1077 	return false;
1078 }
1079 
1080 /* Return true iff packet is to be passed to stack */
1081 static bool qede_rx_xdp(struct qede_dev *edev,
1082 			struct qede_fastpath *fp,
1083 			struct qede_rx_queue *rxq,
1084 			struct bpf_prog *prog,
1085 			struct sw_rx_data *bd,
1086 			struct eth_fast_path_rx_reg_cqe *cqe,
1087 			u16 *data_offset, u16 *len)
1088 {
1089 	struct xdp_buff xdp;
1090 	enum xdp_action act;
1091 
1092 	xdp_init_buff(&xdp, rxq->rx_buf_seg_size, &rxq->xdp_rxq);
1093 	xdp_prepare_buff(&xdp, page_address(bd->data), *data_offset,
1094 			 *len, false);
1095 
1096 	act = bpf_prog_run_xdp(prog, &xdp);
1097 
1098 	/* Recalculate, as XDP might have changed the headers */
1099 	*data_offset = xdp.data - xdp.data_hard_start;
1100 	*len = xdp.data_end - xdp.data;
1101 
1102 	if (act == XDP_PASS)
1103 		return true;
1104 
1105 	/* Count number of packets not to be passed to stack */
1106 	rxq->xdp_no_pass++;
1107 
1108 	switch (act) {
1109 	case XDP_TX:
1110 		/* We need the replacement buffer before transmit. */
1111 		if (unlikely(qede_alloc_rx_buffer(rxq, true))) {
1112 			qede_recycle_rx_bd_ring(rxq, 1);
1113 
1114 			trace_xdp_exception(edev->ndev, prog, act);
1115 			break;
1116 		}
1117 
1118 		/* Now if there's a transmission problem, we'd still have to
1119 		 * throw current buffer, as replacement was already allocated.
1120 		 */
1121 		if (unlikely(qede_xdp_xmit(fp->xdp_tx, bd->mapping,
1122 					   *data_offset, *len, bd->data,
1123 					   NULL))) {
1124 			dma_unmap_page(rxq->dev, bd->mapping, PAGE_SIZE,
1125 				       rxq->data_direction);
1126 			__free_page(bd->data);
1127 
1128 			trace_xdp_exception(edev->ndev, prog, act);
1129 		} else {
1130 			dma_sync_single_for_device(rxq->dev,
1131 						   bd->mapping + *data_offset,
1132 						   *len, rxq->data_direction);
1133 			fp->xdp_xmit |= QEDE_XDP_TX;
1134 		}
1135 
1136 		/* Regardless, we've consumed an Rx BD */
1137 		qede_rx_bd_ring_consume(rxq);
1138 		break;
1139 	case XDP_REDIRECT:
1140 		/* We need the replacement buffer before transmit. */
1141 		if (unlikely(qede_alloc_rx_buffer(rxq, true))) {
1142 			qede_recycle_rx_bd_ring(rxq, 1);
1143 
1144 			trace_xdp_exception(edev->ndev, prog, act);
1145 			break;
1146 		}
1147 
1148 		dma_unmap_page(rxq->dev, bd->mapping, PAGE_SIZE,
1149 			       rxq->data_direction);
1150 
1151 		if (unlikely(xdp_do_redirect(edev->ndev, &xdp, prog)))
1152 			DP_NOTICE(edev, "Failed to redirect the packet\n");
1153 		else
1154 			fp->xdp_xmit |= QEDE_XDP_REDIRECT;
1155 
1156 		qede_rx_bd_ring_consume(rxq);
1157 		break;
1158 	default:
1159 		bpf_warn_invalid_xdp_action(edev->ndev, prog, act);
1160 		fallthrough;
1161 	case XDP_ABORTED:
1162 		trace_xdp_exception(edev->ndev, prog, act);
1163 		fallthrough;
1164 	case XDP_DROP:
1165 		qede_recycle_rx_bd_ring(rxq, cqe->bd_num);
1166 	}
1167 
1168 	return false;
1169 }
1170 
1171 static int qede_rx_build_jumbo(struct qede_dev *edev,
1172 			       struct qede_rx_queue *rxq,
1173 			       struct sk_buff *skb,
1174 			       struct eth_fast_path_rx_reg_cqe *cqe,
1175 			       u16 first_bd_len)
1176 {
1177 	u16 pkt_len = le16_to_cpu(cqe->pkt_len);
1178 	struct sw_rx_data *bd;
1179 	u16 bd_cons_idx;
1180 	u8 num_frags;
1181 
1182 	pkt_len -= first_bd_len;
1183 
1184 	/* We've already used one BD for the SKB. Now take care of the rest */
1185 	for (num_frags = cqe->bd_num - 1; num_frags > 0; num_frags--) {
1186 		u16 cur_size = pkt_len > rxq->rx_buf_size ? rxq->rx_buf_size :
1187 		    pkt_len;
1188 
1189 		if (unlikely(!cur_size)) {
1190 			DP_ERR(edev,
1191 			       "Still got %d BDs for mapping jumbo, but length became 0\n",
1192 			       num_frags);
1193 			goto out;
1194 		}
1195 
1196 		/* We need a replacement buffer for each BD */
1197 		if (unlikely(qede_alloc_rx_buffer(rxq, true)))
1198 			goto out;
1199 
1200 		/* Now that we've allocated the replacement buffer,
1201 		 * we can safely consume the next BD and map it to the SKB.
1202 		 */
1203 		bd_cons_idx = rxq->sw_rx_cons & NUM_RX_BDS_MAX;
1204 		bd = &rxq->sw_rx_ring[bd_cons_idx];
1205 		qede_rx_bd_ring_consume(rxq);
1206 
1207 		dma_unmap_page(rxq->dev, bd->mapping,
1208 			       PAGE_SIZE, DMA_FROM_DEVICE);
1209 
1210 		skb_add_rx_frag(skb, skb_shinfo(skb)->nr_frags, bd->data,
1211 				rxq->rx_headroom, cur_size, PAGE_SIZE);
1212 
1213 		pkt_len -= cur_size;
1214 	}
1215 
1216 	if (unlikely(pkt_len))
1217 		DP_ERR(edev,
1218 		       "Mapped all BDs of jumbo, but still have %d bytes\n",
1219 		       pkt_len);
1220 
1221 out:
1222 	return num_frags;
1223 }
1224 
1225 static int qede_rx_process_tpa_cqe(struct qede_dev *edev,
1226 				   struct qede_fastpath *fp,
1227 				   struct qede_rx_queue *rxq,
1228 				   union eth_rx_cqe *cqe,
1229 				   enum eth_rx_cqe_type type)
1230 {
1231 	switch (type) {
1232 	case ETH_RX_CQE_TYPE_TPA_START:
1233 		qede_tpa_start(edev, rxq, &cqe->fast_path_tpa_start);
1234 		return 0;
1235 	case ETH_RX_CQE_TYPE_TPA_CONT:
1236 		qede_tpa_cont(edev, rxq, &cqe->fast_path_tpa_cont);
1237 		return 0;
1238 	case ETH_RX_CQE_TYPE_TPA_END:
1239 		return qede_tpa_end(edev, fp, &cqe->fast_path_tpa_end);
1240 	default:
1241 		return 0;
1242 	}
1243 }
1244 
1245 static int qede_rx_process_cqe(struct qede_dev *edev,
1246 			       struct qede_fastpath *fp,
1247 			       struct qede_rx_queue *rxq)
1248 {
1249 	struct bpf_prog *xdp_prog = READ_ONCE(rxq->xdp_prog);
1250 	struct eth_fast_path_rx_reg_cqe *fp_cqe;
1251 	u16 len, pad, bd_cons_idx, parse_flag;
1252 	enum eth_rx_cqe_type cqe_type;
1253 	union eth_rx_cqe *cqe;
1254 	struct sw_rx_data *bd;
1255 	struct sk_buff *skb;
1256 	__le16 flags;
1257 	u8 csum_flag;
1258 
1259 	/* Get the CQE from the completion ring */
1260 	cqe = (union eth_rx_cqe *)qed_chain_consume(&rxq->rx_comp_ring);
1261 	cqe_type = cqe->fast_path_regular.type;
1262 
1263 	/* Process an unlikely slowpath event */
1264 	if (unlikely(cqe_type == ETH_RX_CQE_TYPE_SLOW_PATH)) {
1265 		struct eth_slow_path_rx_cqe *sp_cqe;
1266 
1267 		sp_cqe = (struct eth_slow_path_rx_cqe *)cqe;
1268 		edev->ops->eth_cqe_completion(edev->cdev, fp->id, sp_cqe);
1269 		return 0;
1270 	}
1271 
1272 	/* Handle TPA cqes */
1273 	if (cqe_type != ETH_RX_CQE_TYPE_REGULAR)
1274 		return qede_rx_process_tpa_cqe(edev, fp, rxq, cqe, cqe_type);
1275 
1276 	/* Get the data from the SW ring; Consume it only after it's evident
1277 	 * we wouldn't recycle it.
1278 	 */
1279 	bd_cons_idx = rxq->sw_rx_cons & NUM_RX_BDS_MAX;
1280 	bd = &rxq->sw_rx_ring[bd_cons_idx];
1281 
1282 	fp_cqe = &cqe->fast_path_regular;
1283 	len = le16_to_cpu(fp_cqe->len_on_first_bd);
1284 	pad = fp_cqe->placement_offset + rxq->rx_headroom;
1285 
1286 	/* Run eBPF program if one is attached */
1287 	if (xdp_prog)
1288 		if (!qede_rx_xdp(edev, fp, rxq, xdp_prog, bd, fp_cqe,
1289 				 &pad, &len))
1290 			return 0;
1291 
1292 	/* If this is an error packet then drop it */
1293 	flags = cqe->fast_path_regular.pars_flags.flags;
1294 	parse_flag = le16_to_cpu(flags);
1295 
1296 	csum_flag = qede_check_csum(parse_flag);
1297 	if (unlikely(csum_flag == QEDE_CSUM_ERROR)) {
1298 		if (qede_pkt_is_ip_fragmented(fp_cqe, parse_flag))
1299 			rxq->rx_ip_frags++;
1300 		else
1301 			rxq->rx_hw_errors++;
1302 	}
1303 
1304 	/* Basic validation passed; Need to prepare an SKB. This would also
1305 	 * guarantee to finally consume the first BD upon success.
1306 	 */
1307 	skb = qede_rx_build_skb(edev, rxq, bd, len, pad);
1308 	if (!skb) {
1309 		rxq->rx_alloc_errors++;
1310 		qede_recycle_rx_bd_ring(rxq, fp_cqe->bd_num);
1311 		return 0;
1312 	}
1313 
1314 	/* In case of Jumbo packet, several PAGE_SIZEd buffers will be pointed
1315 	 * by a single cqe.
1316 	 */
1317 	if (fp_cqe->bd_num > 1) {
1318 		u16 unmapped_frags = qede_rx_build_jumbo(edev, rxq, skb,
1319 							 fp_cqe, len);
1320 
1321 		if (unlikely(unmapped_frags > 0)) {
1322 			qede_recycle_rx_bd_ring(rxq, unmapped_frags);
1323 			dev_kfree_skb_any(skb);
1324 			return 0;
1325 		}
1326 	}
1327 
1328 	/* The SKB contains all the data. Now prepare meta-magic */
1329 	skb->protocol = eth_type_trans(skb, edev->ndev);
1330 	qede_get_rxhash(skb, fp_cqe->bitfields, fp_cqe->rss_hash);
1331 	qede_set_skb_csum(skb, csum_flag);
1332 	skb_record_rx_queue(skb, rxq->rxq_id);
1333 	qede_ptp_record_rx_ts(edev, cqe, skb);
1334 
1335 	/* SKB is prepared - pass it to stack */
1336 	qede_skb_receive(edev, fp, rxq, skb, le16_to_cpu(fp_cqe->vlan_tag));
1337 
1338 	return 1;
1339 }
1340 
1341 static int qede_rx_int(struct qede_fastpath *fp, int budget)
1342 {
1343 	struct qede_rx_queue *rxq = fp->rxq;
1344 	struct qede_dev *edev = fp->edev;
1345 	int work_done = 0, rcv_pkts = 0;
1346 	u16 hw_comp_cons, sw_comp_cons;
1347 
1348 	hw_comp_cons = le16_to_cpu(*rxq->hw_cons_ptr);
1349 	sw_comp_cons = qed_chain_get_cons_idx(&rxq->rx_comp_ring);
1350 
1351 	/* Memory barrier to prevent the CPU from doing speculative reads of CQE
1352 	 * / BD in the while-loop before reading hw_comp_cons. If the CQE is
1353 	 * read before it is written by FW, then FW writes CQE and SB, and then
1354 	 * the CPU reads the hw_comp_cons, it will use an old CQE.
1355 	 */
1356 	rmb();
1357 
1358 	/* Loop to complete all indicated BDs */
1359 	while ((sw_comp_cons != hw_comp_cons) && (work_done < budget)) {
1360 		rcv_pkts += qede_rx_process_cqe(edev, fp, rxq);
1361 		qed_chain_recycle_consumed(&rxq->rx_comp_ring);
1362 		sw_comp_cons = qed_chain_get_cons_idx(&rxq->rx_comp_ring);
1363 		work_done++;
1364 	}
1365 
1366 	rxq->rcv_pkts += rcv_pkts;
1367 
1368 	/* Allocate replacement buffers */
1369 	while (rxq->num_rx_buffers - rxq->filled_buffers)
1370 		if (qede_alloc_rx_buffer(rxq, false))
1371 			break;
1372 
1373 	/* Update producers */
1374 	qede_update_rx_prod(edev, rxq);
1375 
1376 	return work_done;
1377 }
1378 
1379 static bool qede_poll_is_more_work(struct qede_fastpath *fp)
1380 {
1381 	qed_sb_update_sb_idx(fp->sb_info);
1382 
1383 	/* *_has_*_work() reads the status block, thus we need to ensure that
1384 	 * status block indices have been actually read (qed_sb_update_sb_idx)
1385 	 * prior to this check (*_has_*_work) so that we won't write the
1386 	 * "newer" value of the status block to HW (if there was a DMA right
1387 	 * after qede_has_rx_work and if there is no rmb, the memory reading
1388 	 * (qed_sb_update_sb_idx) may be postponed to right before *_ack_sb).
1389 	 * In this case there will never be another interrupt until there is
1390 	 * another update of the status block, while there is still unhandled
1391 	 * work.
1392 	 */
1393 	rmb();
1394 
1395 	if (likely(fp->type & QEDE_FASTPATH_RX))
1396 		if (qede_has_rx_work(fp->rxq))
1397 			return true;
1398 
1399 	if (fp->type & QEDE_FASTPATH_XDP)
1400 		if (qede_txq_has_work(fp->xdp_tx))
1401 			return true;
1402 
1403 	if (likely(fp->type & QEDE_FASTPATH_TX)) {
1404 		int cos;
1405 
1406 		for_each_cos_in_txq(fp->edev, cos) {
1407 			if (qede_txq_has_work(&fp->txq[cos]))
1408 				return true;
1409 		}
1410 	}
1411 
1412 	return false;
1413 }
1414 
1415 /*********************
1416  * NDO & API related *
1417  *********************/
1418 int qede_poll(struct napi_struct *napi, int budget)
1419 {
1420 	struct qede_fastpath *fp = container_of(napi, struct qede_fastpath,
1421 						napi);
1422 	struct qede_dev *edev = fp->edev;
1423 	int rx_work_done = 0;
1424 	u16 xdp_prod;
1425 
1426 	fp->xdp_xmit = 0;
1427 
1428 	if (likely(fp->type & QEDE_FASTPATH_TX)) {
1429 		int cos;
1430 
1431 		for_each_cos_in_txq(fp->edev, cos) {
1432 			if (qede_txq_has_work(&fp->txq[cos]))
1433 				qede_tx_int(edev, &fp->txq[cos]);
1434 		}
1435 	}
1436 
1437 	if ((fp->type & QEDE_FASTPATH_XDP) && qede_txq_has_work(fp->xdp_tx))
1438 		qede_xdp_tx_int(edev, fp->xdp_tx);
1439 
1440 	rx_work_done = (likely(fp->type & QEDE_FASTPATH_RX) &&
1441 			qede_has_rx_work(fp->rxq)) ?
1442 			qede_rx_int(fp, budget) : 0;
1443 	/* Handle case where we are called by netpoll with a budget of 0 */
1444 	if (rx_work_done < budget || !budget) {
1445 		if (!qede_poll_is_more_work(fp)) {
1446 			napi_complete_done(napi, rx_work_done);
1447 
1448 			/* Update and reenable interrupts */
1449 			qed_sb_ack(fp->sb_info, IGU_INT_ENABLE, 1);
1450 		} else {
1451 			rx_work_done = budget;
1452 		}
1453 	}
1454 
1455 	if (fp->xdp_xmit & QEDE_XDP_TX) {
1456 		xdp_prod = qed_chain_get_prod_idx(&fp->xdp_tx->tx_pbl);
1457 
1458 		fp->xdp_tx->tx_db.data.bd_prod = cpu_to_le16(xdp_prod);
1459 		qede_update_tx_producer(fp->xdp_tx);
1460 	}
1461 
1462 	if (fp->xdp_xmit & QEDE_XDP_REDIRECT)
1463 		xdp_do_flush_map();
1464 
1465 	return rx_work_done;
1466 }
1467 
1468 irqreturn_t qede_msix_fp_int(int irq, void *fp_cookie)
1469 {
1470 	struct qede_fastpath *fp = fp_cookie;
1471 
1472 	qed_sb_ack(fp->sb_info, IGU_INT_DISABLE, 0 /*do not update*/);
1473 
1474 	napi_schedule_irqoff(&fp->napi);
1475 	return IRQ_HANDLED;
1476 }
1477 
1478 /* Main transmit function */
1479 netdev_tx_t qede_start_xmit(struct sk_buff *skb, struct net_device *ndev)
1480 {
1481 	struct qede_dev *edev = netdev_priv(ndev);
1482 	struct netdev_queue *netdev_txq;
1483 	struct qede_tx_queue *txq;
1484 	struct eth_tx_1st_bd *first_bd;
1485 	struct eth_tx_2nd_bd *second_bd = NULL;
1486 	struct eth_tx_3rd_bd *third_bd = NULL;
1487 	struct eth_tx_bd *tx_data_bd = NULL;
1488 	u16 txq_index, val = 0;
1489 	u8 nbd = 0;
1490 	dma_addr_t mapping;
1491 	int rc, frag_idx = 0, ipv6_ext = 0;
1492 	u8 xmit_type;
1493 	u16 idx;
1494 	u16 hlen;
1495 	bool data_split = false;
1496 
1497 	/* Get tx-queue context and netdev index */
1498 	txq_index = skb_get_queue_mapping(skb);
1499 	WARN_ON(txq_index >= QEDE_TSS_COUNT(edev) * edev->dev_info.num_tc);
1500 	txq = QEDE_NDEV_TXQ_ID_TO_TXQ(edev, txq_index);
1501 	netdev_txq = netdev_get_tx_queue(ndev, txq_index);
1502 
1503 	WARN_ON(qed_chain_get_elem_left(&txq->tx_pbl) < (MAX_SKB_FRAGS + 1));
1504 
1505 	xmit_type = qede_xmit_type(skb, &ipv6_ext);
1506 
1507 #if ((MAX_SKB_FRAGS + 2) > ETH_TX_MAX_BDS_PER_NON_LSO_PACKET)
1508 	if (qede_pkt_req_lin(skb, xmit_type)) {
1509 		if (skb_linearize(skb)) {
1510 			txq->tx_mem_alloc_err++;
1511 
1512 			dev_kfree_skb_any(skb);
1513 			return NETDEV_TX_OK;
1514 		}
1515 	}
1516 #endif
1517 
1518 	/* Fill the entry in the SW ring and the BDs in the FW ring */
1519 	idx = txq->sw_tx_prod;
1520 	txq->sw_tx_ring.skbs[idx].skb = skb;
1521 	first_bd = (struct eth_tx_1st_bd *)
1522 		   qed_chain_produce(&txq->tx_pbl);
1523 	memset(first_bd, 0, sizeof(*first_bd));
1524 	first_bd->data.bd_flags.bitfields =
1525 		1 << ETH_TX_1ST_BD_FLAGS_START_BD_SHIFT;
1526 
1527 	if (unlikely(skb_shinfo(skb)->tx_flags & SKBTX_HW_TSTAMP))
1528 		qede_ptp_tx_ts(edev, skb);
1529 
1530 	/* Map skb linear data for DMA and set in the first BD */
1531 	mapping = dma_map_single(txq->dev, skb->data,
1532 				 skb_headlen(skb), DMA_TO_DEVICE);
1533 	if (unlikely(dma_mapping_error(txq->dev, mapping))) {
1534 		DP_NOTICE(edev, "SKB mapping failed\n");
1535 		qede_free_failed_tx_pkt(txq, first_bd, 0, false);
1536 		qede_update_tx_producer(txq);
1537 		return NETDEV_TX_OK;
1538 	}
1539 	nbd++;
1540 	BD_SET_UNMAP_ADDR_LEN(first_bd, mapping, skb_headlen(skb));
1541 
1542 	/* In case there is IPv6 with extension headers or LSO we need 2nd and
1543 	 * 3rd BDs.
1544 	 */
1545 	if (unlikely((xmit_type & XMIT_LSO) | ipv6_ext)) {
1546 		second_bd = (struct eth_tx_2nd_bd *)
1547 			qed_chain_produce(&txq->tx_pbl);
1548 		memset(second_bd, 0, sizeof(*second_bd));
1549 
1550 		nbd++;
1551 		third_bd = (struct eth_tx_3rd_bd *)
1552 			qed_chain_produce(&txq->tx_pbl);
1553 		memset(third_bd, 0, sizeof(*third_bd));
1554 
1555 		nbd++;
1556 		/* We need to fill in additional data in second_bd... */
1557 		tx_data_bd = (struct eth_tx_bd *)second_bd;
1558 	}
1559 
1560 	if (skb_vlan_tag_present(skb)) {
1561 		first_bd->data.vlan = cpu_to_le16(skb_vlan_tag_get(skb));
1562 		first_bd->data.bd_flags.bitfields |=
1563 			1 << ETH_TX_1ST_BD_FLAGS_VLAN_INSERTION_SHIFT;
1564 	}
1565 
1566 	/* Fill the parsing flags & params according to the requested offload */
1567 	if (xmit_type & XMIT_L4_CSUM) {
1568 		/* We don't re-calculate IP checksum as it is already done by
1569 		 * the upper stack
1570 		 */
1571 		first_bd->data.bd_flags.bitfields |=
1572 			1 << ETH_TX_1ST_BD_FLAGS_L4_CSUM_SHIFT;
1573 
1574 		if (xmit_type & XMIT_ENC) {
1575 			first_bd->data.bd_flags.bitfields |=
1576 				1 << ETH_TX_1ST_BD_FLAGS_IP_CSUM_SHIFT;
1577 
1578 			val |= (1 << ETH_TX_DATA_1ST_BD_TUNN_FLAG_SHIFT);
1579 		}
1580 
1581 		/* Legacy FW had flipped behavior in regard to this bit -
1582 		 * I.e., needed to set to prevent FW from touching encapsulated
1583 		 * packets when it didn't need to.
1584 		 */
1585 		if (unlikely(txq->is_legacy))
1586 			val ^= (1 << ETH_TX_DATA_1ST_BD_TUNN_FLAG_SHIFT);
1587 
1588 		/* If the packet is IPv6 with extension header, indicate that
1589 		 * to FW and pass few params, since the device cracker doesn't
1590 		 * support parsing IPv6 with extension header/s.
1591 		 */
1592 		if (unlikely(ipv6_ext))
1593 			qede_set_params_for_ipv6_ext(skb, second_bd, third_bd);
1594 	}
1595 
1596 	if (xmit_type & XMIT_LSO) {
1597 		first_bd->data.bd_flags.bitfields |=
1598 			(1 << ETH_TX_1ST_BD_FLAGS_LSO_SHIFT);
1599 		third_bd->data.lso_mss =
1600 			cpu_to_le16(skb_shinfo(skb)->gso_size);
1601 
1602 		if (unlikely(xmit_type & XMIT_ENC)) {
1603 			first_bd->data.bd_flags.bitfields |=
1604 				1 << ETH_TX_1ST_BD_FLAGS_TUNN_IP_CSUM_SHIFT;
1605 
1606 			if (xmit_type & XMIT_ENC_GSO_L4_CSUM) {
1607 				u8 tmp = ETH_TX_1ST_BD_FLAGS_TUNN_L4_CSUM_SHIFT;
1608 
1609 				first_bd->data.bd_flags.bitfields |= 1 << tmp;
1610 			}
1611 			hlen = qede_get_skb_hlen(skb, true);
1612 		} else {
1613 			first_bd->data.bd_flags.bitfields |=
1614 				1 << ETH_TX_1ST_BD_FLAGS_IP_CSUM_SHIFT;
1615 			hlen = qede_get_skb_hlen(skb, false);
1616 		}
1617 
1618 		/* @@@TBD - if will not be removed need to check */
1619 		third_bd->data.bitfields |=
1620 			cpu_to_le16(1 << ETH_TX_DATA_3RD_BD_HDR_NBD_SHIFT);
1621 
1622 		/* Make life easier for FW guys who can't deal with header and
1623 		 * data on same BD. If we need to split, use the second bd...
1624 		 */
1625 		if (unlikely(skb_headlen(skb) > hlen)) {
1626 			DP_VERBOSE(edev, NETIF_MSG_TX_QUEUED,
1627 				   "TSO split header size is %d (%x:%x)\n",
1628 				   first_bd->nbytes, first_bd->addr.hi,
1629 				   first_bd->addr.lo);
1630 
1631 			mapping = HILO_U64(le32_to_cpu(first_bd->addr.hi),
1632 					   le32_to_cpu(first_bd->addr.lo)) +
1633 					   hlen;
1634 
1635 			BD_SET_UNMAP_ADDR_LEN(tx_data_bd, mapping,
1636 					      le16_to_cpu(first_bd->nbytes) -
1637 					      hlen);
1638 
1639 			/* this marks the BD as one that has no
1640 			 * individual mapping
1641 			 */
1642 			txq->sw_tx_ring.skbs[idx].flags |= QEDE_TSO_SPLIT_BD;
1643 
1644 			first_bd->nbytes = cpu_to_le16(hlen);
1645 
1646 			tx_data_bd = (struct eth_tx_bd *)third_bd;
1647 			data_split = true;
1648 		}
1649 	} else {
1650 		if (unlikely(skb->len > ETH_TX_MAX_NON_LSO_PKT_LEN)) {
1651 			DP_ERR(edev, "Unexpected non LSO skb length = 0x%x\n", skb->len);
1652 			qede_free_failed_tx_pkt(txq, first_bd, 0, false);
1653 			qede_update_tx_producer(txq);
1654 			return NETDEV_TX_OK;
1655 		}
1656 
1657 		val |= ((skb->len & ETH_TX_DATA_1ST_BD_PKT_LEN_MASK) <<
1658 			 ETH_TX_DATA_1ST_BD_PKT_LEN_SHIFT);
1659 	}
1660 
1661 	first_bd->data.bitfields = cpu_to_le16(val);
1662 
1663 	/* Handle fragmented skb */
1664 	/* special handle for frags inside 2nd and 3rd bds.. */
1665 	while (tx_data_bd && frag_idx < skb_shinfo(skb)->nr_frags) {
1666 		rc = map_frag_to_bd(txq,
1667 				    &skb_shinfo(skb)->frags[frag_idx],
1668 				    tx_data_bd);
1669 		if (rc) {
1670 			qede_free_failed_tx_pkt(txq, first_bd, nbd, data_split);
1671 			qede_update_tx_producer(txq);
1672 			return NETDEV_TX_OK;
1673 		}
1674 
1675 		if (tx_data_bd == (struct eth_tx_bd *)second_bd)
1676 			tx_data_bd = (struct eth_tx_bd *)third_bd;
1677 		else
1678 			tx_data_bd = NULL;
1679 
1680 		frag_idx++;
1681 	}
1682 
1683 	/* map last frags into 4th, 5th .... */
1684 	for (; frag_idx < skb_shinfo(skb)->nr_frags; frag_idx++, nbd++) {
1685 		tx_data_bd = (struct eth_tx_bd *)
1686 			     qed_chain_produce(&txq->tx_pbl);
1687 
1688 		memset(tx_data_bd, 0, sizeof(*tx_data_bd));
1689 
1690 		rc = map_frag_to_bd(txq,
1691 				    &skb_shinfo(skb)->frags[frag_idx],
1692 				    tx_data_bd);
1693 		if (rc) {
1694 			qede_free_failed_tx_pkt(txq, first_bd, nbd, data_split);
1695 			qede_update_tx_producer(txq);
1696 			return NETDEV_TX_OK;
1697 		}
1698 	}
1699 
1700 	/* update the first BD with the actual num BDs */
1701 	first_bd->data.nbds = nbd;
1702 
1703 	netdev_tx_sent_queue(netdev_txq, skb->len);
1704 
1705 	skb_tx_timestamp(skb);
1706 
1707 	/* Advance packet producer only before sending the packet since mapping
1708 	 * of pages may fail.
1709 	 */
1710 	txq->sw_tx_prod = (txq->sw_tx_prod + 1) % txq->num_tx_buffers;
1711 
1712 	/* 'next page' entries are counted in the producer value */
1713 	txq->tx_db.data.bd_prod =
1714 		cpu_to_le16(qed_chain_get_prod_idx(&txq->tx_pbl));
1715 
1716 	if (!netdev_xmit_more() || netif_xmit_stopped(netdev_txq))
1717 		qede_update_tx_producer(txq);
1718 
1719 	if (unlikely(qed_chain_get_elem_left(&txq->tx_pbl)
1720 		      < (MAX_SKB_FRAGS + 1))) {
1721 		if (netdev_xmit_more())
1722 			qede_update_tx_producer(txq);
1723 
1724 		netif_tx_stop_queue(netdev_txq);
1725 		txq->stopped_cnt++;
1726 		DP_VERBOSE(edev, NETIF_MSG_TX_QUEUED,
1727 			   "Stop queue was called\n");
1728 		/* paired memory barrier is in qede_tx_int(), we have to keep
1729 		 * ordering of set_bit() in netif_tx_stop_queue() and read of
1730 		 * fp->bd_tx_cons
1731 		 */
1732 		smp_mb();
1733 
1734 		if ((qed_chain_get_elem_left(&txq->tx_pbl) >=
1735 		     (MAX_SKB_FRAGS + 1)) &&
1736 		    (edev->state == QEDE_STATE_OPEN)) {
1737 			netif_tx_wake_queue(netdev_txq);
1738 			DP_VERBOSE(edev, NETIF_MSG_TX_QUEUED,
1739 				   "Wake queue was called\n");
1740 		}
1741 	}
1742 
1743 	return NETDEV_TX_OK;
1744 }
1745 
1746 u16 qede_select_queue(struct net_device *dev, struct sk_buff *skb,
1747 		      struct net_device *sb_dev)
1748 {
1749 	struct qede_dev *edev = netdev_priv(dev);
1750 	int total_txq;
1751 
1752 	total_txq = QEDE_TSS_COUNT(edev) * edev->dev_info.num_tc;
1753 
1754 	return QEDE_TSS_COUNT(edev) ?
1755 		netdev_pick_tx(dev, skb, NULL) % total_txq :  0;
1756 }
1757 
1758 /* 8B udp header + 8B base tunnel header + 32B option length */
1759 #define QEDE_MAX_TUN_HDR_LEN 48
1760 
1761 netdev_features_t qede_features_check(struct sk_buff *skb,
1762 				      struct net_device *dev,
1763 				      netdev_features_t features)
1764 {
1765 	if (skb->encapsulation) {
1766 		u8 l4_proto = 0;
1767 
1768 		switch (vlan_get_protocol(skb)) {
1769 		case htons(ETH_P_IP):
1770 			l4_proto = ip_hdr(skb)->protocol;
1771 			break;
1772 		case htons(ETH_P_IPV6):
1773 			l4_proto = ipv6_hdr(skb)->nexthdr;
1774 			break;
1775 		default:
1776 			return features;
1777 		}
1778 
1779 		/* Disable offloads for geneve tunnels, as HW can't parse
1780 		 * the geneve header which has option length greater than 32b
1781 		 * and disable offloads for the ports which are not offloaded.
1782 		 */
1783 		if (l4_proto == IPPROTO_UDP) {
1784 			struct qede_dev *edev = netdev_priv(dev);
1785 			u16 hdrlen, vxln_port, gnv_port;
1786 
1787 			hdrlen = QEDE_MAX_TUN_HDR_LEN;
1788 			vxln_port = edev->vxlan_dst_port;
1789 			gnv_port = edev->geneve_dst_port;
1790 
1791 			if ((skb_inner_mac_header(skb) -
1792 			     skb_transport_header(skb)) > hdrlen ||
1793 			     (ntohs(udp_hdr(skb)->dest) != vxln_port &&
1794 			      ntohs(udp_hdr(skb)->dest) != gnv_port))
1795 				return features & ~(NETIF_F_CSUM_MASK |
1796 						    NETIF_F_GSO_MASK);
1797 		} else if (l4_proto == IPPROTO_IPIP) {
1798 			/* IPIP tunnels are unknown to the device or at least unsupported natively,
1799 			 * offloads for them can't be done trivially, so disable them for such skb.
1800 			 */
1801 			return features & ~(NETIF_F_CSUM_MASK | NETIF_F_GSO_MASK);
1802 		}
1803 	}
1804 
1805 	return features;
1806 }
1807