qede_fp.c

   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/array_size.h>
   8#include <linux/netdevice.h>
   9#include <linux/etherdevice.h>
  10#include <linux/skbuff.h>
  11#include <linux/bpf_trace.h>
  12#include <net/udp_tunnel.h>
  13#include <linux/ip.h>
  14#include <net/gro.h>
  15#include <net/ipv6.h>
  16#include <net/tcp.h>
  17#include <linux/if_ether.h>
  18#include <linux/if_vlan.h>
  19#include <net/ip6_checksum.h>
  20#include "qede_ptp.h"
  21
  22#include <linux/qed/qed_if.h>
  23#include "qede.h"
  24/*********************************
  25 * Content also used by slowpath *
  26 *********************************/
  27
  28int qede_alloc_rx_buffer(struct qede_rx_queue *rxq, bool allow_lazy)
  29{
  30	struct sw_rx_data *sw_rx_data;
  31	struct eth_rx_bd *rx_bd;
  32	dma_addr_t mapping;
  33	struct page *data;
  34
  35	/* In case lazy-allocation is allowed, postpone allocation until the
  36	 * end of the NAPI run. We'd still need to make sure the Rx ring has
  37	 * sufficient buffers to guarantee an additional Rx interrupt.
  38	 */
  39	if (allow_lazy && likely(rxq->filled_buffers > 12)) {
  40		rxq->filled_buffers--;
  41		return 0;
  42	}
  43
  44	data = alloc_pages(GFP_ATOMIC, 0);
  45	if (unlikely(!data))
  46		return -ENOMEM;
  47
  48	/* Map the entire page as it would be used
  49	 * for multiple RX buffer segment size mapping.
  50	 */
  51	mapping = dma_map_page(rxq->dev, data, 0,
  52			       PAGE_SIZE, rxq->data_direction);
  53	if (unlikely(dma_mapping_error(rxq->dev, mapping))) {
  54		__free_page(data);
  55		return -ENOMEM;
  56	}
  57
  58	sw_rx_data = &rxq->sw_rx_ring[rxq->sw_rx_prod & NUM_RX_BDS_MAX];
  59	sw_rx_data->page_offset = 0;
  60	sw_rx_data->data = data;
  61	sw_rx_data->mapping = mapping;
  62
  63	/* Advance PROD and get BD pointer */
  64	rx_bd = (struct eth_rx_bd *)qed_chain_produce(&rxq->rx_bd_ring);
  65	WARN_ON(!rx_bd);
  66	rx_bd->addr.hi = cpu_to_le32(upper_32_bits(mapping));
  67	rx_bd->addr.lo = cpu_to_le32(lower_32_bits(mapping) +
  68				     rxq->rx_headroom);
  69
  70	rxq->sw_rx_prod++;
  71	rxq->filled_buffers++;
  72
  73	return 0;
  74}
  75
  76/* Unmap the data and free skb */
  77int qede_free_tx_pkt(struct qede_dev *edev, struct qede_tx_queue *txq, int *len)
  78{
  79	u16 idx = txq->sw_tx_cons;
  80	struct sk_buff *skb = txq->sw_tx_ring.skbs[idx].skb;
  81	struct eth_tx_1st_bd *first_bd;
  82	struct eth_tx_bd *tx_data_bd;
  83	int bds_consumed = 0;
  84	int nbds;
  85	bool data_split = txq->sw_tx_ring.skbs[idx].flags & QEDE_TSO_SPLIT_BD;
  86	int i, split_bd_len = 0;
  87
  88	if (unlikely(!skb)) {
  89		DP_ERR(edev,
  90		       "skb is null for txq idx=%d txq->sw_tx_cons=%d txq->sw_tx_prod=%d\n",
  91		       idx, txq->sw_tx_cons, txq->sw_tx_prod);
  92		return -1;
  93	}
  94
  95	*len = skb->len;
  96
  97	first_bd = (struct eth_tx_1st_bd *)qed_chain_consume(&txq->tx_pbl);
  98
  99	bds_consumed++;
 100
 101	nbds = first_bd->data.nbds;
 102
 103	if (data_split) {
 104		struct eth_tx_bd *split = (struct eth_tx_bd *)
 105			qed_chain_consume(&txq->tx_pbl);
 106		split_bd_len = BD_UNMAP_LEN(split);
 107		bds_consumed++;
 108	}
 109	dma_unmap_single(&edev->pdev->dev, BD_UNMAP_ADDR(first_bd),
 110			 BD_UNMAP_LEN(first_bd) + split_bd_len, DMA_TO_DEVICE);
 111
 112	/* Unmap the data of the skb frags */
 113	for (i = 0; i < skb_shinfo(skb)->nr_frags; i++, bds_consumed++) {
 114		tx_data_bd = (struct eth_tx_bd *)
 115			qed_chain_consume(&txq->tx_pbl);
 116		dma_unmap_page(&edev->pdev->dev, BD_UNMAP_ADDR(tx_data_bd),
 117			       BD_UNMAP_LEN(tx_data_bd), DMA_TO_DEVICE);
 118	}
 119
 120	while (bds_consumed++ < nbds)
 121		qed_chain_consume(&txq->tx_pbl);
 122
 123	/* Free skb */
 124	dev_kfree_skb_any(skb);
 125	txq->sw_tx_ring.skbs[idx].skb = NULL;
 126	txq->sw_tx_ring.skbs[idx].flags = 0;
 127
 128	return 0;
 129}
 130
 131/* Unmap the data and free skb when mapping failed during start_xmit */
 132static void qede_free_failed_tx_pkt(struct qede_tx_queue *txq,
 133				    struct eth_tx_1st_bd *first_bd,
 134				    int nbd, bool data_split)
 135{
 136	u16 idx = txq->sw_tx_prod;
 137	struct sk_buff *skb = txq->sw_tx_ring.skbs[idx].skb;
 138	struct eth_tx_bd *tx_data_bd;
 139	int i, split_bd_len = 0;
 140
 141	/* Return prod to its position before this skb was handled */
 142	qed_chain_set_prod(&txq->tx_pbl,
 143			   le16_to_cpu(txq->tx_db.data.bd_prod), first_bd);
 144
 145	first_bd = (struct eth_tx_1st_bd *)qed_chain_produce(&txq->tx_pbl);
 146
 147	if (data_split) {
 148		struct eth_tx_bd *split = (struct eth_tx_bd *)
 149					  qed_chain_produce(&txq->tx_pbl);
 150		split_bd_len = BD_UNMAP_LEN(split);
 151		nbd--;
 152	}
 153
 154	dma_unmap_single(txq->dev, BD_UNMAP_ADDR(first_bd),
 155			 BD_UNMAP_LEN(first_bd) + split_bd_len, DMA_TO_DEVICE);
 156
 157	/* Unmap the data of the skb frags */
 158	for (i = 0; i < nbd; i++) {
 159		tx_data_bd = (struct eth_tx_bd *)
 160			qed_chain_produce(&txq->tx_pbl);
 161		if (tx_data_bd->nbytes)
 162			dma_unmap_page(txq->dev,
 163				       BD_UNMAP_ADDR(tx_data_bd),
 164				       BD_UNMAP_LEN(tx_data_bd), DMA_TO_DEVICE);
 165	}
 166
 167	/* Return again prod to its position before this skb was handled */
 168	qed_chain_set_prod(&txq->tx_pbl,
 169			   le16_to_cpu(txq->tx_db.data.bd_prod), first_bd);
 170
 171	/* Free skb */
 172	dev_kfree_skb_any(skb);
 173	txq->sw_tx_ring.skbs[idx].skb = NULL;
 174	txq->sw_tx_ring.skbs[idx].flags = 0;
 175}
 176
 177static u32 qede_xmit_type(struct sk_buff *skb, int *ipv6_ext)
 178{
 179	u32 rc = XMIT_L4_CSUM;
 180	__be16 l3_proto;
 181
 182	if (skb->ip_summed != CHECKSUM_PARTIAL)
 183		return XMIT_PLAIN;
 184
 185	l3_proto = vlan_get_protocol(skb);
 186	if (l3_proto == htons(ETH_P_IPV6) &&
 187	    (ipv6_hdr(skb)->nexthdr == NEXTHDR_IPV6))
 188		*ipv6_ext = 1;
 189
 190	if (skb->encapsulation) {
 191		rc |= XMIT_ENC;
 192		if (skb_is_gso(skb)) {
 193			unsigned short gso_type = skb_shinfo(skb)->gso_type;
 194
 195			if ((gso_type & SKB_GSO_UDP_TUNNEL_CSUM) ||
 196			    (gso_type & SKB_GSO_GRE_CSUM))
 197				rc |= XMIT_ENC_GSO_L4_CSUM;
 198
 199			rc |= XMIT_LSO;
 200			return rc;
 201		}
 202	}
 203
 204	if (skb_is_gso(skb))
 205		rc |= XMIT_LSO;
 206
 207	return rc;
 208}
 209
 210static void qede_set_params_for_ipv6_ext(struct sk_buff *skb,
 211					 struct eth_tx_2nd_bd *second_bd,
 212					 struct eth_tx_3rd_bd *third_bd)
 213{
 214	u8 l4_proto;
 215	u16 bd2_bits1 = 0, bd2_bits2 = 0;
 216
 217	bd2_bits1 |= (1 << ETH_TX_DATA_2ND_BD_IPV6_EXT_SHIFT);
 218
 219	bd2_bits2 |= ((skb_transport_offset(skb) >> 1) &
 220		     ETH_TX_DATA_2ND_BD_L4_HDR_START_OFFSET_W_MASK)
 221		    << ETH_TX_DATA_2ND_BD_L4_HDR_START_OFFSET_W_SHIFT;
 222
 223	bd2_bits1 |= (ETH_L4_PSEUDO_CSUM_CORRECT_LENGTH <<
 224		      ETH_TX_DATA_2ND_BD_L4_PSEUDO_CSUM_MODE_SHIFT);
 225
 226	if (vlan_get_protocol(skb) == htons(ETH_P_IPV6))
 227		l4_proto = ipv6_hdr(skb)->nexthdr;
 228	else
 229		l4_proto = ip_hdr(skb)->protocol;
 230
 231	if (l4_proto == IPPROTO_UDP)
 232		bd2_bits1 |= 1 << ETH_TX_DATA_2ND_BD_L4_UDP_SHIFT;
 233
 234	if (third_bd)
 235		third_bd->data.bitfields |=
 236			cpu_to_le16(((tcp_hdrlen(skb) / 4) &
 237				ETH_TX_DATA_3RD_BD_TCP_HDR_LEN_DW_MASK) <<
 238				ETH_TX_DATA_3RD_BD_TCP_HDR_LEN_DW_SHIFT);
 239
 240	second_bd->data.bitfields1 = cpu_to_le16(bd2_bits1);
 241	second_bd->data.bitfields2 = cpu_to_le16(bd2_bits2);
 242}
 243
 244static int map_frag_to_bd(struct qede_tx_queue *txq,
 245			  skb_frag_t *frag, struct eth_tx_bd *bd)
 246{
 247	dma_addr_t mapping;
 248
 249	/* Map skb non-linear frag data for DMA */
 250	mapping = skb_frag_dma_map(txq->dev, frag, 0,
 251				   skb_frag_size(frag), DMA_TO_DEVICE);
 252	if (unlikely(dma_mapping_error(txq->dev, mapping)))
 253		return -ENOMEM;
 254
 255	/* Setup the data pointer of the frag data */
 256	BD_SET_UNMAP_ADDR_LEN(bd, mapping, skb_frag_size(frag));
 257
 258	return 0;
 259}
 260
 261static u16 qede_get_skb_hlen(struct sk_buff *skb, bool is_encap_pkt)
 262{
 263	if (is_encap_pkt)
 264		return skb_inner_tcp_all_headers(skb);
 265
 266	return skb_tcp_all_headers(skb);
 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)
 271static 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
 289static 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
 305static 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
 340int 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
 388int 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
 401static 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
 433static 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
 503bool 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
 516static 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 */
 525static 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 */
 548void 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
 559static 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
 589void 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
 609static 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
 625static 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
 638static 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
 649static 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
 665static 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
 699out:
 700	tpa_info->state = QEDE_AGG_STATE_ERROR;
 701	qede_recycle_rx_bd_ring(rxq, 1);
 702
 703	return -ENOMEM;
 704}
 705
 706static 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
 712static 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
 740static inline struct sk_buff *
 741qede_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	if (unlikely(!skb))
 751		return NULL;
 752
 753	skb_reserve(skb, pad);
 754	skb_put(skb, len);
 755
 756	return skb;
 757}
 758
 759static struct sk_buff *
 760qede_tpa_rx_build_skb(struct qede_dev *edev,
 761		      struct qede_rx_queue *rxq,
 762		      struct sw_rx_data *bd, u16 len, u16 pad,
 763		      bool alloc_skb)
 764{
 765	struct sk_buff *skb;
 766
 767	skb = qede_build_skb(rxq, bd, len, pad);
 768	bd->page_offset += rxq->rx_buf_seg_size;
 769
 770	if (bd->page_offset == PAGE_SIZE) {
 771		if (unlikely(qede_alloc_rx_buffer(rxq, true))) {
 772			DP_NOTICE(edev,
 773				  "Failed to allocate RX buffer for tpa start\n");
 774			bd->page_offset -= rxq->rx_buf_seg_size;
 775			page_ref_inc(bd->data);
 776			dev_kfree_skb_any(skb);
 777			return NULL;
 778		}
 779	} else {
 780		page_ref_inc(bd->data);
 781		qede_reuse_page(rxq, bd);
 782	}
 783
 784	/* We've consumed the first BD and prepared an SKB */
 785	qede_rx_bd_ring_consume(rxq);
 786
 787	return skb;
 788}
 789
 790static struct sk_buff *
 791qede_rx_build_skb(struct qede_dev *edev,
 792		  struct qede_rx_queue *rxq,
 793		  struct sw_rx_data *bd, u16 len, u16 pad)
 794{
 795	struct sk_buff *skb = NULL;
 796
 797	/* For smaller frames still need to allocate skb, memcpy
 798	 * data and benefit in reusing the page segment instead of
 799	 * un-mapping it.
 800	 */
 801	if ((len + pad <= edev->rx_copybreak)) {
 802		unsigned int offset = bd->page_offset + pad;
 803
 804		skb = netdev_alloc_skb(edev->ndev, QEDE_RX_HDR_SIZE);
 805		if (unlikely(!skb))
 806			return NULL;
 807
 808		skb_reserve(skb, pad);
 809		skb_put_data(skb, page_address(bd->data) + offset, len);
 810		qede_reuse_page(rxq, bd);
 811		goto out;
 812	}
 813
 814	skb = qede_build_skb(rxq, bd, len, pad);
 815
 816	if (unlikely(qede_realloc_rx_buffer(rxq, bd))) {
 817		/* Incr page ref count to reuse on allocation failure so
 818		 * that it doesn't get freed while freeing SKB [as its
 819		 * already mapped there].
 820		 */
 821		page_ref_inc(bd->data);
 822		dev_kfree_skb_any(skb);
 823		return NULL;
 824	}
 825out:
 826	/* We've consumed the first BD and prepared an SKB */
 827	qede_rx_bd_ring_consume(rxq);
 828
 829	return skb;
 830}
 831
 832static void qede_tpa_start(struct qede_dev *edev,
 833			   struct qede_rx_queue *rxq,
 834			   struct eth_fast_path_rx_tpa_start_cqe *cqe)
 835{
 836	struct qede_agg_info *tpa_info = &rxq->tpa_info[cqe->tpa_agg_index];
 837	struct sw_rx_data *sw_rx_data_cons;
 838	u16 pad;
 839
 840	sw_rx_data_cons = &rxq->sw_rx_ring[rxq->sw_rx_cons & NUM_RX_BDS_MAX];
 841	pad = cqe->placement_offset + rxq->rx_headroom;
 842
 843	tpa_info->skb = qede_tpa_rx_build_skb(edev, rxq, sw_rx_data_cons,
 844					      le16_to_cpu(cqe->len_on_first_bd),
 845					      pad, false);
 846	tpa_info->buffer.page_offset = sw_rx_data_cons->page_offset;
 847	tpa_info->buffer.mapping = sw_rx_data_cons->mapping;
 848
 849	if (unlikely(!tpa_info->skb)) {
 850		DP_NOTICE(edev, "Failed to allocate SKB for gro\n");
 851
 852		/* Consume from ring but do not produce since
 853		 * this might be used by FW still, it will be re-used
 854		 * at TPA end.
 855		 */
 856		tpa_info->tpa_start_fail = true;
 857		qede_rx_bd_ring_consume(rxq);
 858		tpa_info->state = QEDE_AGG_STATE_ERROR;
 859		goto cons_buf;
 860	}
 861
 862	tpa_info->frag_id = 0;
 863	tpa_info->state = QEDE_AGG_STATE_START;
 864
 865	if ((le16_to_cpu(cqe->pars_flags.flags) >>
 866	     PARSING_AND_ERR_FLAGS_TAG8021QEXIST_SHIFT) &
 867	    PARSING_AND_ERR_FLAGS_TAG8021QEXIST_MASK)
 868		tpa_info->vlan_tag = le16_to_cpu(cqe->vlan_tag);
 869	else
 870		tpa_info->vlan_tag = 0;
 871
 872	qede_get_rxhash(tpa_info->skb, cqe->bitfields, cqe->rss_hash);
 873
 874	/* This is needed in order to enable forwarding support */
 875	qede_set_gro_params(edev, tpa_info->skb, cqe);
 876
 877cons_buf: /* We still need to handle bd_len_list to consume buffers */
 878	if (likely(cqe->bw_ext_bd_len_list[0]))
 879		qede_fill_frag_skb(edev, rxq, cqe->tpa_agg_index,
 880				   le16_to_cpu(cqe->bw_ext_bd_len_list[0]));
 881
 882	if (unlikely(cqe->bw_ext_bd_len_list[1])) {
 883		DP_ERR(edev,
 884		       "Unlikely - got a TPA aggregation with more than one bw_ext_bd_len_list entry in the TPA start\n");
 885		tpa_info->state = QEDE_AGG_STATE_ERROR;
 886	}
 887}
 888
 889#ifdef CONFIG_INET
 890static void qede_gro_ip_csum(struct sk_buff *skb)
 891{
 892	const struct iphdr *iph = ip_hdr(skb);
 893	struct tcphdr *th;
 894
 895	skb_set_transport_header(skb, sizeof(struct iphdr));
 896	th = tcp_hdr(skb);
 897
 898	th->check = ~tcp_v4_check(skb->len - skb_transport_offset(skb),
 899				  iph->saddr, iph->daddr, 0);
 900
 901	tcp_gro_complete(skb);
 902}
 903
 904static void qede_gro_ipv6_csum(struct sk_buff *skb)
 905{
 906	struct ipv6hdr *iph = ipv6_hdr(skb);
 907	struct tcphdr *th;
 908
 909	skb_set_transport_header(skb, sizeof(struct ipv6hdr));
 910	th = tcp_hdr(skb);
 911
 912	th->check = ~tcp_v6_check(skb->len - skb_transport_offset(skb),
 913				  &iph->saddr, &iph->daddr, 0);
 914	tcp_gro_complete(skb);
 915}
 916#endif
 917
 918static void qede_gro_receive(struct qede_dev *edev,
 919			     struct qede_fastpath *fp,
 920			     struct sk_buff *skb,
 921			     u16 vlan_tag)
 922{
 923	/* FW can send a single MTU sized packet from gro flow
 924	 * due to aggregation timeout/last segment etc. which
 925	 * is not expected to be a gro packet. If a skb has zero
 926	 * frags then simply push it in the stack as non gso skb.
 927	 */
 928	if (unlikely(!skb->data_len)) {
 929		skb_shinfo(skb)->gso_type = 0;
 930		skb_shinfo(skb)->gso_size = 0;
 931		goto send_skb;
 932	}
 933
 934#ifdef CONFIG_INET
 935	if (skb_shinfo(skb)->gso_size) {
 936		skb_reset_network_header(skb);
 937
 938		switch (skb->protocol) {
 939		case htons(ETH_P_IP):
 940			qede_gro_ip_csum(skb);
 941			break;
 942		case htons(ETH_P_IPV6):
 943			qede_gro_ipv6_csum(skb);
 944			break;
 945		default:
 946			DP_ERR(edev,
 947			       "Error: FW GRO supports only IPv4/IPv6, not 0x%04x\n",
 948			       ntohs(skb->protocol));
 949		}
 950	}
 951#endif
 952
 953send_skb:
 954	skb_record_rx_queue(skb, fp->rxq->rxq_id);
 955	qede_skb_receive(edev, fp, fp->rxq, skb, vlan_tag);
 956}
 957
 958static inline void qede_tpa_cont(struct qede_dev *edev,
 959				 struct qede_rx_queue *rxq,
 960				 struct eth_fast_path_rx_tpa_cont_cqe *cqe)
 961{
 962	int i;
 963
 964	for (i = 0; cqe->len_list[i] && i < ARRAY_SIZE(cqe->len_list); i++)
 965		qede_fill_frag_skb(edev, rxq, cqe->tpa_agg_index,
 966				   le16_to_cpu(cqe->len_list[i]));
 967
 968	if (unlikely(i > 1))
 969		DP_ERR(edev,
 970		       "Strange - TPA cont with more than a single len_list entry\n");
 971}
 972
 973static int qede_tpa_end(struct qede_dev *edev,
 974			struct qede_fastpath *fp,
 975			struct eth_fast_path_rx_tpa_end_cqe *cqe)
 976{
 977	struct qede_rx_queue *rxq = fp->rxq;
 978	struct qede_agg_info *tpa_info;
 979	struct sk_buff *skb;
 980	int i;
 981
 982	tpa_info = &rxq->tpa_info[cqe->tpa_agg_index];
 983	skb = tpa_info->skb;
 984
 985	if (tpa_info->buffer.page_offset == PAGE_SIZE)
 986		dma_unmap_page(rxq->dev, tpa_info->buffer.mapping,
 987			       PAGE_SIZE, rxq->data_direction);
 988
 989	for (i = 0; cqe->len_list[i] && i < ARRAY_SIZE(cqe->len_list); i++)
 990		qede_fill_frag_skb(edev, rxq, cqe->tpa_agg_index,
 991				   le16_to_cpu(cqe->len_list[i]));
 992	if (unlikely(i > 1))
 993		DP_ERR(edev,
 994		       "Strange - TPA emd with more than a single len_list entry\n");
 995
 996	if (unlikely(tpa_info->state != QEDE_AGG_STATE_START))
 997		goto err;
 998
 999	/* Sanity */
1000	if (unlikely(cqe->num_of_bds != tpa_info->frag_id + 1))
1001		DP_ERR(edev,
1002		       "Strange - TPA had %02x BDs, but SKB has only %d frags\n",
1003		       cqe->num_of_bds, tpa_info->frag_id);
1004	if (unlikely(skb->len != le16_to_cpu(cqe->total_packet_len)))
1005		DP_ERR(edev,
1006		       "Strange - total packet len [cqe] is %4x but SKB has len %04x\n",
1007		       le16_to_cpu(cqe->total_packet_len), skb->len);
1008
1009	/* Finalize the SKB */
1010	skb->protocol = eth_type_trans(skb, edev->ndev);
1011	skb->ip_summed = CHECKSUM_UNNECESSARY;
1012
1013	/* tcp_gro_complete() will copy NAPI_GRO_CB(skb)->count
1014	 * to skb_shinfo(skb)->gso_segs
1015	 */
1016	NAPI_GRO_CB(skb)->count = le16_to_cpu(cqe->num_of_coalesced_segs);
1017
1018	qede_gro_receive(edev, fp, skb, tpa_info->vlan_tag);
1019
1020	tpa_info->state = QEDE_AGG_STATE_NONE;
1021
1022	return 1;
1023err:
1024	tpa_info->state = QEDE_AGG_STATE_NONE;
1025
1026	if (tpa_info->tpa_start_fail) {
1027		qede_reuse_page(rxq, &tpa_info->buffer);
1028		tpa_info->tpa_start_fail = false;
1029	}
1030
1031	dev_kfree_skb_any(tpa_info->skb);
1032	tpa_info->skb = NULL;
1033	return 0;
1034}
1035
1036static u8 qede_check_notunn_csum(u16 flag)
1037{
1038	u16 csum_flag = 0;
1039	u8 csum = 0;
1040
1041	if (flag & (PARSING_AND_ERR_FLAGS_L4CHKSMWASCALCULATED_MASK <<
1042		    PARSING_AND_ERR_FLAGS_L4CHKSMWASCALCULATED_SHIFT)) {
1043		csum_flag |= PARSING_AND_ERR_FLAGS_L4CHKSMERROR_MASK <<
1044			     PARSING_AND_ERR_FLAGS_L4CHKSMERROR_SHIFT;
1045		csum = QEDE_CSUM_UNNECESSARY;
1046	}
1047
1048	csum_flag |= PARSING_AND_ERR_FLAGS_IPHDRERROR_MASK <<
1049		     PARSING_AND_ERR_FLAGS_IPHDRERROR_SHIFT;
1050
1051	if (csum_flag & flag)
1052		return QEDE_CSUM_ERROR;
1053
1054	return csum;
1055}
1056
1057static u8 qede_check_csum(u16 flag)
1058{
1059	if (!qede_tunn_exist(flag))
1060		return qede_check_notunn_csum(flag);
1061	else
1062		return qede_check_tunn_csum(flag);
1063}
1064
1065static bool qede_pkt_is_ip_fragmented(struct eth_fast_path_rx_reg_cqe *cqe,
1066				      u16 flag)
1067{
1068	u8 tun_pars_flg = cqe->tunnel_pars_flags.flags;
1069
1070	if ((tun_pars_flg & (ETH_TUNNEL_PARSING_FLAGS_IPV4_FRAGMENT_MASK <<
1071			     ETH_TUNNEL_PARSING_FLAGS_IPV4_FRAGMENT_SHIFT)) ||
1072	    (flag & (PARSING_AND_ERR_FLAGS_IPV4FRAG_MASK <<
1073		     PARSING_AND_ERR_FLAGS_IPV4FRAG_SHIFT)))
1074		return true;
1075
1076	return false;
1077}
1078
1079/* Return true iff packet is to be passed to stack */
1080static bool qede_rx_xdp(struct qede_dev *edev,
1081			struct qede_fastpath *fp,
1082			struct qede_rx_queue *rxq,
1083			struct bpf_prog *prog,
1084			struct sw_rx_data *bd,
1085			struct eth_fast_path_rx_reg_cqe *cqe,
1086			u16 *data_offset, u16 *len)
1087{
1088	struct xdp_buff xdp;
1089	enum xdp_action act;
1090
1091	xdp_init_buff(&xdp, rxq->rx_buf_seg_size, &rxq->xdp_rxq);
1092	xdp_prepare_buff(&xdp, page_address(bd->data), *data_offset,
1093			 *len, false);
1094
1095	act = bpf_prog_run_xdp(prog, &xdp);
1096
1097	/* Recalculate, as XDP might have changed the headers */
1098	*data_offset = xdp.data - xdp.data_hard_start;
1099	*len = xdp.data_end - xdp.data;
1100
1101	if (act == XDP_PASS)
1102		return true;
1103
1104	/* Count number of packets not to be passed to stack */
1105	rxq->xdp_no_pass++;
1106
1107	switch (act) {
1108	case XDP_TX:
1109		/* We need the replacement buffer before transmit. */
1110		if (unlikely(qede_alloc_rx_buffer(rxq, true))) {
1111			qede_recycle_rx_bd_ring(rxq, 1);
1112
1113			trace_xdp_exception(edev->ndev, prog, act);
1114			break;
1115		}
1116
1117		/* Now if there's a transmission problem, we'd still have to
1118		 * throw current buffer, as replacement was already allocated.
1119		 */
1120		if (unlikely(qede_xdp_xmit(fp->xdp_tx, bd->mapping,
1121					   *data_offset, *len, bd->data,
1122					   NULL))) {
1123			dma_unmap_page(rxq->dev, bd->mapping, PAGE_SIZE,
1124				       rxq->data_direction);
1125			__free_page(bd->data);
1126
1127			trace_xdp_exception(edev->ndev, prog, act);
1128		} else {
1129			dma_sync_single_for_device(rxq->dev,
1130						   bd->mapping + *data_offset,
1131						   *len, rxq->data_direction);
1132			fp->xdp_xmit |= QEDE_XDP_TX;
1133		}
1134
1135		/* Regardless, we've consumed an Rx BD */
1136		qede_rx_bd_ring_consume(rxq);
1137		break;
1138	case XDP_REDIRECT:
1139		/* We need the replacement buffer before transmit. */
1140		if (unlikely(qede_alloc_rx_buffer(rxq, true))) {
1141			qede_recycle_rx_bd_ring(rxq, 1);
1142
1143			trace_xdp_exception(edev->ndev, prog, act);
1144			break;
1145		}
1146
1147		dma_unmap_page(rxq->dev, bd->mapping, PAGE_SIZE,
1148			       rxq->data_direction);
1149
1150		if (unlikely(xdp_do_redirect(edev->ndev, &xdp, prog)))
1151			DP_NOTICE(edev, "Failed to redirect the packet\n");
1152		else
1153			fp->xdp_xmit |= QEDE_XDP_REDIRECT;
1154
1155		qede_rx_bd_ring_consume(rxq);
1156		break;
1157	default:
1158		bpf_warn_invalid_xdp_action(edev->ndev, prog, act);
1159		fallthrough;
1160	case XDP_ABORTED:
1161		trace_xdp_exception(edev->ndev, prog, act);
1162		fallthrough;
1163	case XDP_DROP:
1164		qede_recycle_rx_bd_ring(rxq, cqe->bd_num);
1165	}
1166
1167	return false;
1168}
1169
1170static int qede_rx_build_jumbo(struct qede_dev *edev,
1171			       struct qede_rx_queue *rxq,
1172			       struct sk_buff *skb,
1173			       struct eth_fast_path_rx_reg_cqe *cqe,
1174			       u16 first_bd_len)
1175{
1176	u16 pkt_len = le16_to_cpu(cqe->pkt_len);
1177	struct sw_rx_data *bd;
1178	u16 bd_cons_idx;
1179	u8 num_frags;
1180
1181	pkt_len -= first_bd_len;
1182
1183	/* We've already used one BD for the SKB. Now take care of the rest */
1184	for (num_frags = cqe->bd_num - 1; num_frags > 0; num_frags--) {
1185		u16 cur_size = pkt_len > rxq->rx_buf_size ? rxq->rx_buf_size :
1186		    pkt_len;
1187
1188		if (unlikely(!cur_size)) {
1189			DP_ERR(edev,
1190			       "Still got %d BDs for mapping jumbo, but length became 0\n",
1191			       num_frags);
1192			goto out;
1193		}
1194
1195		/* We need a replacement buffer for each BD */
1196		if (unlikely(qede_alloc_rx_buffer(rxq, true)))
1197			goto out;
1198
1199		/* Now that we've allocated the replacement buffer,
1200		 * we can safely consume the next BD and map it to the SKB.
1201		 */
1202		bd_cons_idx = rxq->sw_rx_cons & NUM_RX_BDS_MAX;
1203		bd = &rxq->sw_rx_ring[bd_cons_idx];
1204		qede_rx_bd_ring_consume(rxq);
1205
1206		dma_unmap_page(rxq->dev, bd->mapping,
1207			       PAGE_SIZE, DMA_FROM_DEVICE);
1208
1209		skb_add_rx_frag(skb, skb_shinfo(skb)->nr_frags, bd->data,
1210				rxq->rx_headroom, cur_size, PAGE_SIZE);
1211
1212		pkt_len -= cur_size;
1213	}
1214
1215	if (unlikely(pkt_len))
1216		DP_ERR(edev,
1217		       "Mapped all BDs of jumbo, but still have %d bytes\n",
1218		       pkt_len);
1219
1220out:
1221	return num_frags;
1222}
1223
1224static int qede_rx_process_tpa_cqe(struct qede_dev *edev,
1225				   struct qede_fastpath *fp,
1226				   struct qede_rx_queue *rxq,
1227				   union eth_rx_cqe *cqe,
1228				   enum eth_rx_cqe_type type)
1229{
1230	switch (type) {
1231	case ETH_RX_CQE_TYPE_TPA_START:
1232		qede_tpa_start(edev, rxq, &cqe->fast_path_tpa_start);
1233		return 0;
1234	case ETH_RX_CQE_TYPE_TPA_CONT:
1235		qede_tpa_cont(edev, rxq, &cqe->fast_path_tpa_cont);
1236		return 0;
1237	case ETH_RX_CQE_TYPE_TPA_END:
1238		return qede_tpa_end(edev, fp, &cqe->fast_path_tpa_end);
1239	default:
1240		return 0;
1241	}
1242}
1243
1244static int qede_rx_process_cqe(struct qede_dev *edev,
1245			       struct qede_fastpath *fp,
1246			       struct qede_rx_queue *rxq)
1247{
1248	struct bpf_prog *xdp_prog = READ_ONCE(rxq->xdp_prog);
1249	struct eth_fast_path_rx_reg_cqe *fp_cqe;
1250	u16 len, pad, bd_cons_idx, parse_flag;
1251	enum eth_rx_cqe_type cqe_type;
1252	union eth_rx_cqe *cqe;
1253	struct sw_rx_data *bd;
1254	struct sk_buff *skb;
1255	__le16 flags;
1256	u8 csum_flag;
1257
1258	/* Get the CQE from the completion ring */
1259	cqe = (union eth_rx_cqe *)qed_chain_consume(&rxq->rx_comp_ring);
1260	cqe_type = cqe->fast_path_regular.type;
1261
1262	/* Process an unlikely slowpath event */
1263	if (unlikely(cqe_type == ETH_RX_CQE_TYPE_SLOW_PATH)) {
1264		struct eth_slow_path_rx_cqe *sp_cqe;
1265
1266		sp_cqe = (struct eth_slow_path_rx_cqe *)cqe;
1267		edev->ops->eth_cqe_completion(edev->cdev, fp->id, sp_cqe);
1268		return 0;
1269	}
1270
1271	/* Handle TPA cqes */
1272	if (cqe_type != ETH_RX_CQE_TYPE_REGULAR)
1273		return qede_rx_process_tpa_cqe(edev, fp, rxq, cqe, cqe_type);
1274
1275	/* Get the data from the SW ring; Consume it only after it's evident
1276	 * we wouldn't recycle it.
1277	 */
1278	bd_cons_idx = rxq->sw_rx_cons & NUM_RX_BDS_MAX;
1279	bd = &rxq->sw_rx_ring[bd_cons_idx];
1280
1281	fp_cqe = &cqe->fast_path_regular;
1282	len = le16_to_cpu(fp_cqe->len_on_first_bd);
1283	pad = fp_cqe->placement_offset + rxq->rx_headroom;
1284
1285	/* Run eBPF program if one is attached */
1286	if (xdp_prog)
1287		if (!qede_rx_xdp(edev, fp, rxq, xdp_prog, bd, fp_cqe,
1288				 &pad, &len))
1289			return 0;
1290
1291	/* If this is an error packet then drop it */
1292	flags = cqe->fast_path_regular.pars_flags.flags;
1293	parse_flag = le16_to_cpu(flags);
1294
1295	csum_flag = qede_check_csum(parse_flag);
1296	if (unlikely(csum_flag == QEDE_CSUM_ERROR)) {
1297		if (qede_pkt_is_ip_fragmented(fp_cqe, parse_flag))
1298			rxq->rx_ip_frags++;
1299		else
1300			rxq->rx_hw_errors++;
1301	}
1302
1303	/* Basic validation passed; Need to prepare an SKB. This would also
1304	 * guarantee to finally consume the first BD upon success.
1305	 */
1306	skb = qede_rx_build_skb(edev, rxq, bd, len, pad);
1307	if (!skb) {
1308		rxq->rx_alloc_errors++;
1309		qede_recycle_rx_bd_ring(rxq, fp_cqe->bd_num);
1310		return 0;
1311	}
1312
1313	/* In case of Jumbo packet, several PAGE_SIZEd buffers will be pointed
1314	 * by a single cqe.
1315	 */
1316	if (fp_cqe->bd_num > 1) {
1317		u16 unmapped_frags = qede_rx_build_jumbo(edev, rxq, skb,
1318							 fp_cqe, len);
1319
1320		if (unlikely(unmapped_frags > 0)) {
1321			qede_recycle_rx_bd_ring(rxq, unmapped_frags);
1322			dev_kfree_skb_any(skb);
1323			return 0;
1324		}
1325	}
1326
1327	/* The SKB contains all the data. Now prepare meta-magic */
1328	skb->protocol = eth_type_trans(skb, edev->ndev);
1329	qede_get_rxhash(skb, fp_cqe->bitfields, fp_cqe->rss_hash);
1330	qede_set_skb_csum(skb, csum_flag);
1331	skb_record_rx_queue(skb, rxq->rxq_id);
1332	qede_ptp_record_rx_ts(edev, cqe, skb);
1333
1334	/* SKB is prepared - pass it to stack */
1335	qede_skb_receive(edev, fp, rxq, skb, le16_to_cpu(fp_cqe->vlan_tag));
1336
1337	return 1;
1338}
1339
1340static int qede_rx_int(struct qede_fastpath *fp, int budget)
1341{
1342	struct qede_rx_queue *rxq = fp->rxq;
1343	struct qede_dev *edev = fp->edev;
1344	int work_done = 0, rcv_pkts = 0;
1345	u16 hw_comp_cons, sw_comp_cons;
1346
1347	hw_comp_cons = le16_to_cpu(*rxq->hw_cons_ptr);
1348	sw_comp_cons = qed_chain_get_cons_idx(&rxq->rx_comp_ring);
1349
1350	/* Memory barrier to prevent the CPU from doing speculative reads of CQE
1351	 * / BD in the while-loop before reading hw_comp_cons. If the CQE is
1352	 * read before it is written by FW, then FW writes CQE and SB, and then
1353	 * the CPU reads the hw_comp_cons, it will use an old CQE.
1354	 */
1355	rmb();
1356
1357	/* Loop to complete all indicated BDs */
1358	while ((sw_comp_cons != hw_comp_cons) && (work_done < budget)) {
1359		rcv_pkts += qede_rx_process_cqe(edev, fp, rxq);
1360		qed_chain_recycle_consumed(&rxq->rx_comp_ring);
1361		sw_comp_cons = qed_chain_get_cons_idx(&rxq->rx_comp_ring);
1362		work_done++;
1363	}
1364
1365	rxq->rcv_pkts += rcv_pkts;
1366
1367	/* Allocate replacement buffers */
1368	while (rxq->num_rx_buffers - rxq->filled_buffers)
1369		if (qede_alloc_rx_buffer(rxq, false))
1370			break;
1371
1372	/* Update producers */
1373	qede_update_rx_prod(edev, rxq);
1374
1375	return work_done;
1376}
1377
1378static bool qede_poll_is_more_work(struct qede_fastpath *fp)
1379{
1380	qed_sb_update_sb_idx(fp->sb_info);
1381
1382	/* *_has_*_work() reads the status block, thus we need to ensure that
1383	 * status block indices have been actually read (qed_sb_update_sb_idx)
1384	 * prior to this check (*_has_*_work) so that we won't write the
1385	 * "newer" value of the status block to HW (if there was a DMA right
1386	 * after qede_has_rx_work and if there is no rmb, the memory reading
1387	 * (qed_sb_update_sb_idx) may be postponed to right before *_ack_sb).
1388	 * In this case there will never be another interrupt until there is
1389	 * another update of the status block, while there is still unhandled
1390	 * work.
1391	 */
1392	rmb();
1393
1394	if (likely(fp->type & QEDE_FASTPATH_RX))
1395		if (qede_has_rx_work(fp->rxq))
1396			return true;
1397
1398	if (fp->type & QEDE_FASTPATH_XDP)
1399		if (qede_txq_has_work(fp->xdp_tx))
1400			return true;
1401
1402	if (likely(fp->type & QEDE_FASTPATH_TX)) {
1403		int cos;
1404
1405		for_each_cos_in_txq(fp->edev, cos) {
1406			if (qede_txq_has_work(&fp->txq[cos]))
1407				return true;
1408		}
1409	}
1410
1411	return false;
1412}
1413
1414/*********************
1415 * NDO & API related *
1416 *********************/
1417int qede_poll(struct napi_struct *napi, int budget)
1418{
1419	struct qede_fastpath *fp = container_of(napi, struct qede_fastpath,
1420						napi);
1421	struct qede_dev *edev = fp->edev;
1422	int rx_work_done = 0;
1423	u16 xdp_prod;
1424
1425	fp->xdp_xmit = 0;
1426
1427	if (likely(fp->type & QEDE_FASTPATH_TX)) {
1428		int cos;
1429
1430		for_each_cos_in_txq(fp->edev, cos) {
1431			if (qede_txq_has_work(&fp->txq[cos]))
1432				qede_tx_int(edev, &fp->txq[cos]);
1433		}
1434	}
1435
1436	if ((fp->type & QEDE_FASTPATH_XDP) && qede_txq_has_work(fp->xdp_tx))
1437		qede_xdp_tx_int(edev, fp->xdp_tx);
1438
1439	rx_work_done = (likely(fp->type & QEDE_FASTPATH_RX) &&
1440			qede_has_rx_work(fp->rxq)) ?
1441			qede_rx_int(fp, budget) : 0;
1442
1443	if (fp->xdp_xmit & QEDE_XDP_REDIRECT)
1444		xdp_do_flush();
1445
1446	/* Handle case where we are called by netpoll with a budget of 0 */
1447	if (rx_work_done < budget || !budget) {
1448		if (!qede_poll_is_more_work(fp)) {
1449			napi_complete_done(napi, rx_work_done);
1450
1451			/* Update and reenable interrupts */
1452			qed_sb_ack(fp->sb_info, IGU_INT_ENABLE, 1);
1453		} else {
1454			rx_work_done = budget;
1455		}
1456	}
1457
1458	if (fp->xdp_xmit & QEDE_XDP_TX) {
1459		xdp_prod = qed_chain_get_prod_idx(&fp->xdp_tx->tx_pbl);
1460
1461		fp->xdp_tx->tx_db.data.bd_prod = cpu_to_le16(xdp_prod);
1462		qede_update_tx_producer(fp->xdp_tx);
1463	}
1464
1465	return rx_work_done;
1466}
1467
1468irqreturn_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 */
1479netdev_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
1746u16 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
1761netdev_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}