dp.c

   1// SPDX-License-Identifier: (GPL-2.0-only OR BSD-2-Clause)
   2/* Copyright (C) 2015-2019 Netronome Systems, Inc. */
   3
   4#include <linux/bpf_trace.h>
   5#include <linux/netdevice.h>
   6#include <linux/bitfield.h>
   7#include <net/xfrm.h>
   8
   9#include "../nfp_app.h"
  10#include "../nfp_net.h"
  11#include "../nfp_net_dp.h"
  12#include "../nfp_net_xsk.h"
  13#include "../crypto/crypto.h"
  14#include "../crypto/fw.h"
  15#include "nfd3.h"
  16
  17/* Transmit processing
  18 *
  19 * One queue controller peripheral queue is used for transmit.  The
  20 * driver en-queues packets for transmit by advancing the write
  21 * pointer.  The device indicates that packets have transmitted by
  22 * advancing the read pointer.  The driver maintains a local copy of
  23 * the read and write pointer in @struct nfp_net_tx_ring.  The driver
  24 * keeps @wr_p in sync with the queue controller write pointer and can
  25 * determine how many packets have been transmitted by comparing its
  26 * copy of the read pointer @rd_p with the read pointer maintained by
  27 * the queue controller peripheral.
  28 */
  29
  30/* Wrappers for deciding when to stop and restart TX queues */
  31static int nfp_nfd3_tx_ring_should_wake(struct nfp_net_tx_ring *tx_ring)
  32{
  33	return !nfp_net_tx_full(tx_ring, MAX_SKB_FRAGS * 4);
  34}
  35
  36static int nfp_nfd3_tx_ring_should_stop(struct nfp_net_tx_ring *tx_ring)
  37{
  38	return nfp_net_tx_full(tx_ring, MAX_SKB_FRAGS + 1);
  39}
  40
  41/**
  42 * nfp_nfd3_tx_ring_stop() - stop tx ring
  43 * @nd_q:    netdev queue
  44 * @tx_ring: driver tx queue structure
  45 *
  46 * Safely stop TX ring.  Remember that while we are running .start_xmit()
  47 * someone else may be cleaning the TX ring completions so we need to be
  48 * extra careful here.
  49 */
  50static void
  51nfp_nfd3_tx_ring_stop(struct netdev_queue *nd_q,
  52		      struct nfp_net_tx_ring *tx_ring)
  53{
  54	netif_tx_stop_queue(nd_q);
  55
  56	/* We can race with the TX completion out of NAPI so recheck */
  57	smp_mb();
  58	if (unlikely(nfp_nfd3_tx_ring_should_wake(tx_ring)))
  59		netif_tx_start_queue(nd_q);
  60}
  61
  62/**
  63 * nfp_nfd3_tx_tso() - Set up Tx descriptor for LSO
  64 * @r_vec: per-ring structure
  65 * @txbuf: Pointer to driver soft TX descriptor
  66 * @txd: Pointer to HW TX descriptor
  67 * @skb: Pointer to SKB
  68 * @md_bytes: Prepend length
  69 *
  70 * Set up Tx descriptor for LSO, do nothing for non-LSO skbs.
  71 * Return error on packet header greater than maximum supported LSO header size.
  72 */
  73static void
  74nfp_nfd3_tx_tso(struct nfp_net_r_vector *r_vec, struct nfp_nfd3_tx_buf *txbuf,
  75		struct nfp_nfd3_tx_desc *txd, struct sk_buff *skb, u32 md_bytes)
  76{
  77	u32 l3_offset, l4_offset, hdrlen, l4_hdrlen;
  78	u16 mss;
  79
  80	if (!skb_is_gso(skb))
  81		return;
  82
  83	if (!skb->encapsulation) {
  84		l3_offset = skb_network_offset(skb);
  85		l4_offset = skb_transport_offset(skb);
  86		l4_hdrlen = (skb_shinfo(skb)->gso_type & SKB_GSO_UDP_L4) ?
  87			    sizeof(struct udphdr) : tcp_hdrlen(skb);
  88	} else {
  89		l3_offset = skb_inner_network_offset(skb);
  90		l4_offset = skb_inner_transport_offset(skb);
  91		l4_hdrlen = (skb_shinfo(skb)->gso_type & SKB_GSO_UDP_L4) ?
  92			    sizeof(struct udphdr) : inner_tcp_hdrlen(skb);
  93	}
  94
  95	hdrlen = l4_offset + l4_hdrlen;
  96	txbuf->pkt_cnt = skb_shinfo(skb)->gso_segs;
  97	txbuf->real_len += hdrlen * (txbuf->pkt_cnt - 1);
  98
  99	mss = skb_shinfo(skb)->gso_size & NFD3_DESC_TX_MSS_MASK;
 100	txd->l3_offset = l3_offset - md_bytes;
 101	txd->l4_offset = l4_offset - md_bytes;
 102	txd->lso_hdrlen = hdrlen - md_bytes;
 103	txd->mss = cpu_to_le16(mss);
 104	txd->flags |= NFD3_DESC_TX_LSO;
 105
 106	u64_stats_update_begin(&r_vec->tx_sync);
 107	r_vec->tx_lso++;
 108	u64_stats_update_end(&r_vec->tx_sync);
 109}
 110
 111/**
 112 * nfp_nfd3_tx_csum() - Set TX CSUM offload flags in TX descriptor
 113 * @dp:  NFP Net data path struct
 114 * @r_vec: per-ring structure
 115 * @txbuf: Pointer to driver soft TX descriptor
 116 * @txd: Pointer to TX descriptor
 117 * @skb: Pointer to SKB
 118 *
 119 * This function sets the TX checksum flags in the TX descriptor based
 120 * on the configuration and the protocol of the packet to be transmitted.
 121 */
 122static void
 123nfp_nfd3_tx_csum(struct nfp_net_dp *dp, struct nfp_net_r_vector *r_vec,
 124		 struct nfp_nfd3_tx_buf *txbuf, struct nfp_nfd3_tx_desc *txd,
 125		 struct sk_buff *skb)
 126{
 127	struct ipv6hdr *ipv6h;
 128	struct iphdr *iph;
 129	u8 l4_hdr;
 130
 131	if (!(dp->ctrl & NFP_NET_CFG_CTRL_TXCSUM))
 132		return;
 133
 134	if (skb->ip_summed != CHECKSUM_PARTIAL)
 135		return;
 136
 137	txd->flags |= NFD3_DESC_TX_CSUM;
 138	if (skb->encapsulation)
 139		txd->flags |= NFD3_DESC_TX_ENCAP;
 140
 141	iph = skb->encapsulation ? inner_ip_hdr(skb) : ip_hdr(skb);
 142	ipv6h = skb->encapsulation ? inner_ipv6_hdr(skb) : ipv6_hdr(skb);
 143
 144	if (iph->version == 4) {
 145		txd->flags |= NFD3_DESC_TX_IP4_CSUM;
 146		l4_hdr = iph->protocol;
 147	} else if (ipv6h->version == 6) {
 148		l4_hdr = ipv6h->nexthdr;
 149	} else {
 150		nn_dp_warn(dp, "partial checksum but ipv=%x!\n", iph->version);
 151		return;
 152	}
 153
 154	switch (l4_hdr) {
 155	case IPPROTO_TCP:
 156		txd->flags |= NFD3_DESC_TX_TCP_CSUM;
 157		break;
 158	case IPPROTO_UDP:
 159		txd->flags |= NFD3_DESC_TX_UDP_CSUM;
 160		break;
 161	default:
 162		nn_dp_warn(dp, "partial checksum but l4 proto=%x!\n", l4_hdr);
 163		return;
 164	}
 165
 166	u64_stats_update_begin(&r_vec->tx_sync);
 167	if (skb->encapsulation)
 168		r_vec->hw_csum_tx_inner += txbuf->pkt_cnt;
 169	else
 170		r_vec->hw_csum_tx += txbuf->pkt_cnt;
 171	u64_stats_update_end(&r_vec->tx_sync);
 172}
 173
 174static int nfp_nfd3_prep_tx_meta(struct nfp_net_dp *dp, struct sk_buff *skb,
 175				 u64 tls_handle, bool *ipsec)
 176{
 177	struct metadata_dst *md_dst = skb_metadata_dst(skb);
 178	struct nfp_ipsec_offload offload_info;
 179	unsigned char *data;
 180	bool vlan_insert;
 181	u32 meta_id = 0;
 182	int md_bytes;
 183
 184#ifdef CONFIG_NFP_NET_IPSEC
 185	if (xfrm_offload(skb))
 186		*ipsec = nfp_net_ipsec_tx_prep(dp, skb, &offload_info);
 187#endif
 188
 189	if (unlikely(md_dst && md_dst->type != METADATA_HW_PORT_MUX))
 190		md_dst = NULL;
 191
 192	vlan_insert = skb_vlan_tag_present(skb) && (dp->ctrl & NFP_NET_CFG_CTRL_TXVLAN_V2);
 193
 194	if (!(md_dst || tls_handle || vlan_insert || *ipsec))
 195		return 0;
 196
 197	md_bytes = sizeof(meta_id) +
 198		   (!!md_dst ? NFP_NET_META_PORTID_SIZE : 0) +
 199		   (!!tls_handle ? NFP_NET_META_CONN_HANDLE_SIZE : 0) +
 200		   (vlan_insert ? NFP_NET_META_VLAN_SIZE : 0) +
 201		   (*ipsec ? NFP_NET_META_IPSEC_FIELD_SIZE : 0);
 202
 203	if (unlikely(skb_cow_head(skb, md_bytes)))
 204		return -ENOMEM;
 205
 206	data = skb_push(skb, md_bytes) + md_bytes;
 207	if (md_dst) {
 208		data -= NFP_NET_META_PORTID_SIZE;
 209		put_unaligned_be32(md_dst->u.port_info.port_id, data);
 210		meta_id = NFP_NET_META_PORTID;
 211	}
 212	if (tls_handle) {
 213		/* conn handle is opaque, we just use u64 to be able to quickly
 214		 * compare it to zero
 215		 */
 216		data -= NFP_NET_META_CONN_HANDLE_SIZE;
 217		memcpy(data, &tls_handle, sizeof(tls_handle));
 218		meta_id <<= NFP_NET_META_FIELD_SIZE;
 219		meta_id |= NFP_NET_META_CONN_HANDLE;
 220	}
 221	if (vlan_insert) {
 222		data -= NFP_NET_META_VLAN_SIZE;
 223		/* data type of skb->vlan_proto is __be16
 224		 * so it fills metadata without calling put_unaligned_be16
 225		 */
 226		memcpy(data, &skb->vlan_proto, sizeof(skb->vlan_proto));
 227		put_unaligned_be16(skb_vlan_tag_get(skb), data + sizeof(skb->vlan_proto));
 228		meta_id <<= NFP_NET_META_FIELD_SIZE;
 229		meta_id |= NFP_NET_META_VLAN;
 230	}
 231	if (*ipsec) {
 232		data -= NFP_NET_META_IPSEC_SIZE;
 233		put_unaligned_be32(offload_info.seq_hi, data);
 234		data -= NFP_NET_META_IPSEC_SIZE;
 235		put_unaligned_be32(offload_info.seq_low, data);
 236		data -= NFP_NET_META_IPSEC_SIZE;
 237		put_unaligned_be32(offload_info.handle - 1, data);
 238		meta_id <<= NFP_NET_META_IPSEC_FIELD_SIZE;
 239		meta_id |= NFP_NET_META_IPSEC << 8 | NFP_NET_META_IPSEC << 4 | NFP_NET_META_IPSEC;
 240	}
 241
 242	data -= sizeof(meta_id);
 243	put_unaligned_be32(meta_id, data);
 244
 245	return md_bytes;
 246}
 247
 248/**
 249 * nfp_nfd3_tx() - Main transmit entry point
 250 * @skb:    SKB to transmit
 251 * @netdev: netdev structure
 252 *
 253 * Return: NETDEV_TX_OK on success.
 254 */
 255netdev_tx_t nfp_nfd3_tx(struct sk_buff *skb, struct net_device *netdev)
 256{
 257	struct nfp_net *nn = netdev_priv(netdev);
 258	int f, nr_frags, wr_idx, md_bytes;
 259	struct nfp_net_tx_ring *tx_ring;
 260	struct nfp_net_r_vector *r_vec;
 261	struct nfp_nfd3_tx_buf *txbuf;
 262	struct nfp_nfd3_tx_desc *txd;
 263	struct netdev_queue *nd_q;
 264	const skb_frag_t *frag;
 265	struct nfp_net_dp *dp;
 266	dma_addr_t dma_addr;
 267	unsigned int fsize;
 268	u64 tls_handle = 0;
 269	bool ipsec = false;
 270	u16 qidx;
 271
 272	dp = &nn->dp;
 273	qidx = skb_get_queue_mapping(skb);
 274	tx_ring = &dp->tx_rings[qidx];
 275	r_vec = tx_ring->r_vec;
 276
 277	nr_frags = skb_shinfo(skb)->nr_frags;
 278
 279	if (unlikely(nfp_net_tx_full(tx_ring, nr_frags + 1))) {
 280		nn_dp_warn(dp, "TX ring %d busy. wrp=%u rdp=%u\n",
 281			   qidx, tx_ring->wr_p, tx_ring->rd_p);
 282		nd_q = netdev_get_tx_queue(dp->netdev, qidx);
 283		netif_tx_stop_queue(nd_q);
 284		nfp_net_tx_xmit_more_flush(tx_ring);
 285		u64_stats_update_begin(&r_vec->tx_sync);
 286		r_vec->tx_busy++;
 287		u64_stats_update_end(&r_vec->tx_sync);
 288		return NETDEV_TX_BUSY;
 289	}
 290
 291	skb = nfp_net_tls_tx(dp, r_vec, skb, &tls_handle, &nr_frags);
 292	if (unlikely(!skb)) {
 293		nfp_net_tx_xmit_more_flush(tx_ring);
 294		return NETDEV_TX_OK;
 295	}
 296
 297	md_bytes = nfp_nfd3_prep_tx_meta(dp, skb, tls_handle, &ipsec);
 298	if (unlikely(md_bytes < 0))
 299		goto err_flush;
 300
 301	/* Start with the head skbuf */
 302	dma_addr = dma_map_single(dp->dev, skb->data, skb_headlen(skb),
 303				  DMA_TO_DEVICE);
 304	if (dma_mapping_error(dp->dev, dma_addr))
 305		goto err_dma_err;
 306
 307	wr_idx = D_IDX(tx_ring, tx_ring->wr_p);
 308
 309	/* Stash the soft descriptor of the head then initialize it */
 310	txbuf = &tx_ring->txbufs[wr_idx];
 311	txbuf->skb = skb;
 312	txbuf->dma_addr = dma_addr;
 313	txbuf->fidx = -1;
 314	txbuf->pkt_cnt = 1;
 315	txbuf->real_len = skb->len;
 316
 317	/* Build TX descriptor */
 318	txd = &tx_ring->txds[wr_idx];
 319	txd->offset_eop = (nr_frags ? 0 : NFD3_DESC_TX_EOP) | md_bytes;
 320	txd->dma_len = cpu_to_le16(skb_headlen(skb));
 321	nfp_desc_set_dma_addr_40b(txd, dma_addr);
 322	txd->data_len = cpu_to_le16(skb->len);
 323
 324	txd->flags = 0;
 325	txd->mss = 0;
 326	txd->lso_hdrlen = 0;
 327
 328	/* Do not reorder - tso may adjust pkt cnt, vlan may override fields */
 329	nfp_nfd3_tx_tso(r_vec, txbuf, txd, skb, md_bytes);
 330	if (ipsec)
 331		nfp_nfd3_ipsec_tx(txd, skb);
 332	else
 333		nfp_nfd3_tx_csum(dp, r_vec, txbuf, txd, skb);
 334	if (skb_vlan_tag_present(skb) && dp->ctrl & NFP_NET_CFG_CTRL_TXVLAN) {
 335		txd->flags |= NFD3_DESC_TX_VLAN;
 336		txd->vlan = cpu_to_le16(skb_vlan_tag_get(skb));
 337	}
 338
 339	/* Gather DMA */
 340	if (nr_frags > 0) {
 341		__le64 second_half;
 342
 343		/* all descs must match except for in addr, length and eop */
 344		second_half = txd->vals8[1];
 345
 346		for (f = 0; f < nr_frags; f++) {
 347			frag = &skb_shinfo(skb)->frags[f];
 348			fsize = skb_frag_size(frag);
 349
 350			dma_addr = skb_frag_dma_map(dp->dev, frag, 0,
 351						    fsize, DMA_TO_DEVICE);
 352			if (dma_mapping_error(dp->dev, dma_addr))
 353				goto err_unmap;
 354
 355			wr_idx = D_IDX(tx_ring, wr_idx + 1);
 356			tx_ring->txbufs[wr_idx].skb = skb;
 357			tx_ring->txbufs[wr_idx].dma_addr = dma_addr;
 358			tx_ring->txbufs[wr_idx].fidx = f;
 359
 360			txd = &tx_ring->txds[wr_idx];
 361			txd->dma_len = cpu_to_le16(fsize);
 362			nfp_desc_set_dma_addr_40b(txd, dma_addr);
 363			txd->offset_eop = md_bytes |
 364				((f == nr_frags - 1) ? NFD3_DESC_TX_EOP : 0);
 365			txd->vals8[1] = second_half;
 366		}
 367
 368		u64_stats_update_begin(&r_vec->tx_sync);
 369		r_vec->tx_gather++;
 370		u64_stats_update_end(&r_vec->tx_sync);
 371	}
 372
 373	skb_tx_timestamp(skb);
 374
 375	nd_q = netdev_get_tx_queue(dp->netdev, tx_ring->idx);
 376
 377	tx_ring->wr_p += nr_frags + 1;
 378	if (nfp_nfd3_tx_ring_should_stop(tx_ring))
 379		nfp_nfd3_tx_ring_stop(nd_q, tx_ring);
 380
 381	tx_ring->wr_ptr_add += nr_frags + 1;
 382	if (__netdev_tx_sent_queue(nd_q, txbuf->real_len, netdev_xmit_more()))
 383		nfp_net_tx_xmit_more_flush(tx_ring);
 384
 385	return NETDEV_TX_OK;
 386
 387err_unmap:
 388	while (--f >= 0) {
 389		frag = &skb_shinfo(skb)->frags[f];
 390		dma_unmap_page(dp->dev, tx_ring->txbufs[wr_idx].dma_addr,
 391			       skb_frag_size(frag), DMA_TO_DEVICE);
 392		tx_ring->txbufs[wr_idx].skb = NULL;
 393		tx_ring->txbufs[wr_idx].dma_addr = 0;
 394		tx_ring->txbufs[wr_idx].fidx = -2;
 395		wr_idx = wr_idx - 1;
 396		if (wr_idx < 0)
 397			wr_idx += tx_ring->cnt;
 398	}
 399	dma_unmap_single(dp->dev, tx_ring->txbufs[wr_idx].dma_addr,
 400			 skb_headlen(skb), DMA_TO_DEVICE);
 401	tx_ring->txbufs[wr_idx].skb = NULL;
 402	tx_ring->txbufs[wr_idx].dma_addr = 0;
 403	tx_ring->txbufs[wr_idx].fidx = -2;
 404err_dma_err:
 405	nn_dp_warn(dp, "Failed to map DMA TX buffer\n");
 406err_flush:
 407	nfp_net_tx_xmit_more_flush(tx_ring);
 408	u64_stats_update_begin(&r_vec->tx_sync);
 409	r_vec->tx_errors++;
 410	u64_stats_update_end(&r_vec->tx_sync);
 411	nfp_net_tls_tx_undo(skb, tls_handle);
 412	dev_kfree_skb_any(skb);
 413	return NETDEV_TX_OK;
 414}
 415
 416/**
 417 * nfp_nfd3_tx_complete() - Handled completed TX packets
 418 * @tx_ring:	TX ring structure
 419 * @budget:	NAPI budget (only used as bool to determine if in NAPI context)
 420 */
 421void nfp_nfd3_tx_complete(struct nfp_net_tx_ring *tx_ring, int budget)
 422{
 423	struct nfp_net_r_vector *r_vec = tx_ring->r_vec;
 424	struct nfp_net_dp *dp = &r_vec->nfp_net->dp;
 425	u32 done_pkts = 0, done_bytes = 0;
 426	struct netdev_queue *nd_q;
 427	u32 qcp_rd_p;
 428	int todo;
 429
 430	if (tx_ring->wr_p == tx_ring->rd_p)
 431		return;
 432
 433	/* Work out how many descriptors have been transmitted */
 434	qcp_rd_p = nfp_net_read_tx_cmpl(tx_ring, dp);
 435
 436	if (qcp_rd_p == tx_ring->qcp_rd_p)
 437		return;
 438
 439	todo = D_IDX(tx_ring, qcp_rd_p - tx_ring->qcp_rd_p);
 440
 441	while (todo--) {
 442		const skb_frag_t *frag;
 443		struct nfp_nfd3_tx_buf *tx_buf;
 444		struct sk_buff *skb;
 445		int fidx, nr_frags;
 446		int idx;
 447
 448		idx = D_IDX(tx_ring, tx_ring->rd_p++);
 449		tx_buf = &tx_ring->txbufs[idx];
 450
 451		skb = tx_buf->skb;
 452		if (!skb)
 453			continue;
 454
 455		nr_frags = skb_shinfo(skb)->nr_frags;
 456		fidx = tx_buf->fidx;
 457
 458		if (fidx == -1) {
 459			/* unmap head */
 460			dma_unmap_single(dp->dev, tx_buf->dma_addr,
 461					 skb_headlen(skb), DMA_TO_DEVICE);
 462
 463			done_pkts += tx_buf->pkt_cnt;
 464			done_bytes += tx_buf->real_len;
 465		} else {
 466			/* unmap fragment */
 467			frag = &skb_shinfo(skb)->frags[fidx];
 468			dma_unmap_page(dp->dev, tx_buf->dma_addr,
 469				       skb_frag_size(frag), DMA_TO_DEVICE);
 470		}
 471
 472		/* check for last gather fragment */
 473		if (fidx == nr_frags - 1)
 474			napi_consume_skb(skb, budget);
 475
 476		tx_buf->dma_addr = 0;
 477		tx_buf->skb = NULL;
 478		tx_buf->fidx = -2;
 479	}
 480
 481	tx_ring->qcp_rd_p = qcp_rd_p;
 482
 483	u64_stats_update_begin(&r_vec->tx_sync);
 484	r_vec->tx_bytes += done_bytes;
 485	r_vec->tx_pkts += done_pkts;
 486	u64_stats_update_end(&r_vec->tx_sync);
 487
 488	if (!dp->netdev)
 489		return;
 490
 491	nd_q = netdev_get_tx_queue(dp->netdev, tx_ring->idx);
 492	netdev_tx_completed_queue(nd_q, done_pkts, done_bytes);
 493	if (nfp_nfd3_tx_ring_should_wake(tx_ring)) {
 494		/* Make sure TX thread will see updated tx_ring->rd_p */
 495		smp_mb();
 496
 497		if (unlikely(netif_tx_queue_stopped(nd_q)))
 498			netif_tx_wake_queue(nd_q);
 499	}
 500
 501	WARN_ONCE(tx_ring->wr_p - tx_ring->rd_p > tx_ring->cnt,
 502		  "TX ring corruption rd_p=%u wr_p=%u cnt=%u\n",
 503		  tx_ring->rd_p, tx_ring->wr_p, tx_ring->cnt);
 504}
 505
 506static bool nfp_nfd3_xdp_complete(struct nfp_net_tx_ring *tx_ring)
 507{
 508	struct nfp_net_r_vector *r_vec = tx_ring->r_vec;
 509	struct nfp_net_dp *dp = &r_vec->nfp_net->dp;
 510	u32 done_pkts = 0, done_bytes = 0;
 511	bool done_all;
 512	int idx, todo;
 513	u32 qcp_rd_p;
 514
 515	/* Work out how many descriptors have been transmitted */
 516	qcp_rd_p = nfp_net_read_tx_cmpl(tx_ring, dp);
 517
 518	if (qcp_rd_p == tx_ring->qcp_rd_p)
 519		return true;
 520
 521	todo = D_IDX(tx_ring, qcp_rd_p - tx_ring->qcp_rd_p);
 522
 523	done_all = todo <= NFP_NET_XDP_MAX_COMPLETE;
 524	todo = min(todo, NFP_NET_XDP_MAX_COMPLETE);
 525
 526	tx_ring->qcp_rd_p = D_IDX(tx_ring, tx_ring->qcp_rd_p + todo);
 527
 528	done_pkts = todo;
 529	while (todo--) {
 530		idx = D_IDX(tx_ring, tx_ring->rd_p);
 531		tx_ring->rd_p++;
 532
 533		done_bytes += tx_ring->txbufs[idx].real_len;
 534	}
 535
 536	u64_stats_update_begin(&r_vec->tx_sync);
 537	r_vec->tx_bytes += done_bytes;
 538	r_vec->tx_pkts += done_pkts;
 539	u64_stats_update_end(&r_vec->tx_sync);
 540
 541	WARN_ONCE(tx_ring->wr_p - tx_ring->rd_p > tx_ring->cnt,
 542		  "XDP TX ring corruption rd_p=%u wr_p=%u cnt=%u\n",
 543		  tx_ring->rd_p, tx_ring->wr_p, tx_ring->cnt);
 544
 545	return done_all;
 546}
 547
 548/* Receive processing
 549 */
 550
 551static void *
 552nfp_nfd3_napi_alloc_one(struct nfp_net_dp *dp, dma_addr_t *dma_addr)
 553{
 554	void *frag;
 555
 556	if (!dp->xdp_prog) {
 557		frag = napi_alloc_frag(dp->fl_bufsz);
 558		if (unlikely(!frag))
 559			return NULL;
 560	} else {
 561		struct page *page;
 562
 563		page = dev_alloc_page();
 564		if (unlikely(!page))
 565			return NULL;
 566		frag = page_address(page);
 567	}
 568
 569	*dma_addr = nfp_net_dma_map_rx(dp, frag);
 570	if (dma_mapping_error(dp->dev, *dma_addr)) {
 571		nfp_net_free_frag(frag, dp->xdp_prog);
 572		nn_dp_warn(dp, "Failed to map DMA RX buffer\n");
 573		return NULL;
 574	}
 575
 576	return frag;
 577}
 578
 579/**
 580 * nfp_nfd3_rx_give_one() - Put mapped skb on the software and hardware rings
 581 * @dp:		NFP Net data path struct
 582 * @rx_ring:	RX ring structure
 583 * @frag:	page fragment buffer
 584 * @dma_addr:	DMA address of skb mapping
 585 */
 586static void
 587nfp_nfd3_rx_give_one(const struct nfp_net_dp *dp,
 588		     struct nfp_net_rx_ring *rx_ring,
 589		     void *frag, dma_addr_t dma_addr)
 590{
 591	unsigned int wr_idx;
 592
 593	wr_idx = D_IDX(rx_ring, rx_ring->wr_p);
 594
 595	nfp_net_dma_sync_dev_rx(dp, dma_addr);
 596
 597	/* Stash SKB and DMA address away */
 598	rx_ring->rxbufs[wr_idx].frag = frag;
 599	rx_ring->rxbufs[wr_idx].dma_addr = dma_addr;
 600
 601	/* Fill freelist descriptor */
 602	rx_ring->rxds[wr_idx].fld.reserved = 0;
 603	rx_ring->rxds[wr_idx].fld.meta_len_dd = 0;
 604	/* DMA address is expanded to 48-bit width in freelist for NFP3800,
 605	 * so the *_48b macro is used accordingly, it's also OK to fill
 606	 * a 40-bit address since the top 8 bits are get set to 0.
 607	 */
 608	nfp_desc_set_dma_addr_48b(&rx_ring->rxds[wr_idx].fld,
 609				  dma_addr + dp->rx_dma_off);
 610
 611	rx_ring->wr_p++;
 612	if (!(rx_ring->wr_p % NFP_NET_FL_BATCH)) {
 613		/* Update write pointer of the freelist queue. Make
 614		 * sure all writes are flushed before telling the hardware.
 615		 */
 616		wmb();
 617		nfp_qcp_wr_ptr_add(rx_ring->qcp_fl, NFP_NET_FL_BATCH);
 618	}
 619}
 620
 621/**
 622 * nfp_nfd3_rx_ring_fill_freelist() - Give buffers from the ring to FW
 623 * @dp:	     NFP Net data path struct
 624 * @rx_ring: RX ring to fill
 625 */
 626void nfp_nfd3_rx_ring_fill_freelist(struct nfp_net_dp *dp,
 627				    struct nfp_net_rx_ring *rx_ring)
 628{
 629	unsigned int i;
 630
 631	if (nfp_net_has_xsk_pool_slow(dp, rx_ring->idx))
 632		return nfp_net_xsk_rx_ring_fill_freelist(rx_ring);
 633
 634	for (i = 0; i < rx_ring->cnt - 1; i++)
 635		nfp_nfd3_rx_give_one(dp, rx_ring, rx_ring->rxbufs[i].frag,
 636				     rx_ring->rxbufs[i].dma_addr);
 637}
 638
 639/**
 640 * nfp_nfd3_rx_csum_has_errors() - group check if rxd has any csum errors
 641 * @flags: RX descriptor flags field in CPU byte order
 642 */
 643static int nfp_nfd3_rx_csum_has_errors(u16 flags)
 644{
 645	u16 csum_all_checked, csum_all_ok;
 646
 647	csum_all_checked = flags & __PCIE_DESC_RX_CSUM_ALL;
 648	csum_all_ok = flags & __PCIE_DESC_RX_CSUM_ALL_OK;
 649
 650	return csum_all_checked != (csum_all_ok << PCIE_DESC_RX_CSUM_OK_SHIFT);
 651}
 652
 653/**
 654 * nfp_nfd3_rx_csum() - set SKB checksum field based on RX descriptor flags
 655 * @dp:  NFP Net data path struct
 656 * @r_vec: per-ring structure
 657 * @rxd: Pointer to RX descriptor
 658 * @meta: Parsed metadata prepend
 659 * @skb: Pointer to SKB
 660 */
 661void
 662nfp_nfd3_rx_csum(const struct nfp_net_dp *dp, struct nfp_net_r_vector *r_vec,
 663		 const struct nfp_net_rx_desc *rxd,
 664		 const struct nfp_meta_parsed *meta, struct sk_buff *skb)
 665{
 666	skb_checksum_none_assert(skb);
 667
 668	if (!(dp->netdev->features & NETIF_F_RXCSUM))
 669		return;
 670
 671	if (meta->csum_type) {
 672		skb->ip_summed = meta->csum_type;
 673		skb->csum = meta->csum;
 674		u64_stats_update_begin(&r_vec->rx_sync);
 675		r_vec->hw_csum_rx_complete++;
 676		u64_stats_update_end(&r_vec->rx_sync);
 677		return;
 678	}
 679
 680	if (nfp_nfd3_rx_csum_has_errors(le16_to_cpu(rxd->rxd.flags))) {
 681		u64_stats_update_begin(&r_vec->rx_sync);
 682		r_vec->hw_csum_rx_error++;
 683		u64_stats_update_end(&r_vec->rx_sync);
 684		return;
 685	}
 686
 687	/* Assume that the firmware will never report inner CSUM_OK unless outer
 688	 * L4 headers were successfully parsed. FW will always report zero UDP
 689	 * checksum as CSUM_OK.
 690	 */
 691	if (rxd->rxd.flags & PCIE_DESC_RX_TCP_CSUM_OK ||
 692	    rxd->rxd.flags & PCIE_DESC_RX_UDP_CSUM_OK) {
 693		__skb_incr_checksum_unnecessary(skb);
 694		u64_stats_update_begin(&r_vec->rx_sync);
 695		r_vec->hw_csum_rx_ok++;
 696		u64_stats_update_end(&r_vec->rx_sync);
 697	}
 698
 699	if (rxd->rxd.flags & PCIE_DESC_RX_I_TCP_CSUM_OK ||
 700	    rxd->rxd.flags & PCIE_DESC_RX_I_UDP_CSUM_OK) {
 701		__skb_incr_checksum_unnecessary(skb);
 702		u64_stats_update_begin(&r_vec->rx_sync);
 703		r_vec->hw_csum_rx_inner_ok++;
 704		u64_stats_update_end(&r_vec->rx_sync);
 705	}
 706}
 707
 708static void
 709nfp_nfd3_set_hash(struct net_device *netdev, struct nfp_meta_parsed *meta,
 710		  unsigned int type, __be32 *hash)
 711{
 712	if (!(netdev->features & NETIF_F_RXHASH))
 713		return;
 714
 715	switch (type) {
 716	case NFP_NET_RSS_IPV4:
 717	case NFP_NET_RSS_IPV6:
 718	case NFP_NET_RSS_IPV6_EX:
 719		meta->hash_type = PKT_HASH_TYPE_L3;
 720		break;
 721	default:
 722		meta->hash_type = PKT_HASH_TYPE_L4;
 723		break;
 724	}
 725
 726	meta->hash = get_unaligned_be32(hash);
 727}
 728
 729static void
 730nfp_nfd3_set_hash_desc(struct net_device *netdev, struct nfp_meta_parsed *meta,
 731		       void *data, struct nfp_net_rx_desc *rxd)
 732{
 733	struct nfp_net_rx_hash *rx_hash = data;
 734
 735	if (!(rxd->rxd.flags & PCIE_DESC_RX_RSS))
 736		return;
 737
 738	nfp_nfd3_set_hash(netdev, meta, get_unaligned_be32(&rx_hash->hash_type),
 739			  &rx_hash->hash);
 740}
 741
 742bool
 743nfp_nfd3_parse_meta(struct net_device *netdev, struct nfp_meta_parsed *meta,
 744		    void *data, void *pkt, unsigned int pkt_len, int meta_len)
 745{
 746	u32 meta_info, vlan_info;
 747
 748	meta_info = get_unaligned_be32(data);
 749	data += 4;
 750
 751	while (meta_info) {
 752		switch (meta_info & NFP_NET_META_FIELD_MASK) {
 753		case NFP_NET_META_HASH:
 754			meta_info >>= NFP_NET_META_FIELD_SIZE;
 755			nfp_nfd3_set_hash(netdev, meta,
 756					  meta_info & NFP_NET_META_FIELD_MASK,
 757					  (__be32 *)data);
 758			data += 4;
 759			break;
 760		case NFP_NET_META_MARK:
 761			meta->mark = get_unaligned_be32(data);
 762			data += 4;
 763			break;
 764		case NFP_NET_META_VLAN:
 765			vlan_info = get_unaligned_be32(data);
 766			if (FIELD_GET(NFP_NET_META_VLAN_STRIP, vlan_info)) {
 767				meta->vlan.stripped = true;
 768				meta->vlan.tpid = FIELD_GET(NFP_NET_META_VLAN_TPID_MASK,
 769							    vlan_info);
 770				meta->vlan.tci = FIELD_GET(NFP_NET_META_VLAN_TCI_MASK,
 771							   vlan_info);
 772			}
 773			data += 4;
 774			break;
 775		case NFP_NET_META_PORTID:
 776			meta->portid = get_unaligned_be32(data);
 777			data += 4;
 778			break;
 779		case NFP_NET_META_CSUM:
 780			meta->csum_type = CHECKSUM_COMPLETE;
 781			meta->csum =
 782				(__force __wsum)get_unaligned((u32 *)data);
 783			data += 4;
 784			break;
 785		case NFP_NET_META_RESYNC_INFO:
 786			if (nfp_net_tls_rx_resync_req(netdev, data, pkt,
 787						      pkt_len))
 788				return false;
 789			data += sizeof(struct nfp_net_tls_resync_req);
 790			break;
 791#ifdef CONFIG_NFP_NET_IPSEC
 792		case NFP_NET_META_IPSEC:
 793			/* Note: IPsec packet will have zero saidx, so need add 1
 794			 * to indicate packet is IPsec packet within driver.
 795			 */
 796			meta->ipsec_saidx = get_unaligned_be32(data) + 1;
 797			data += 4;
 798			break;
 799#endif
 800		default:
 801			return true;
 802		}
 803
 804		meta_info >>= NFP_NET_META_FIELD_SIZE;
 805	}
 806
 807	return data != pkt;
 808}
 809
 810static void
 811nfp_nfd3_rx_drop(const struct nfp_net_dp *dp, struct nfp_net_r_vector *r_vec,
 812		 struct nfp_net_rx_ring *rx_ring, struct nfp_net_rx_buf *rxbuf,
 813		 struct sk_buff *skb)
 814{
 815	u64_stats_update_begin(&r_vec->rx_sync);
 816	r_vec->rx_drops++;
 817	/* If we have both skb and rxbuf the replacement buffer allocation
 818	 * must have failed, count this as an alloc failure.
 819	 */
 820	if (skb && rxbuf)
 821		r_vec->rx_replace_buf_alloc_fail++;
 822	u64_stats_update_end(&r_vec->rx_sync);
 823
 824	/* skb is build based on the frag, free_skb() would free the frag
 825	 * so to be able to reuse it we need an extra ref.
 826	 */
 827	if (skb && rxbuf && skb->head == rxbuf->frag)
 828		page_ref_inc(virt_to_head_page(rxbuf->frag));
 829	if (rxbuf)
 830		nfp_nfd3_rx_give_one(dp, rx_ring, rxbuf->frag, rxbuf->dma_addr);
 831	if (skb)
 832		dev_kfree_skb_any(skb);
 833}
 834
 835static bool
 836nfp_nfd3_tx_xdp_buf(struct nfp_net_dp *dp, struct nfp_net_rx_ring *rx_ring,
 837		    struct nfp_net_tx_ring *tx_ring,
 838		    struct nfp_net_rx_buf *rxbuf, unsigned int dma_off,
 839		    unsigned int pkt_len, bool *completed)
 840{
 841	unsigned int dma_map_sz = dp->fl_bufsz - NFP_NET_RX_BUF_NON_DATA;
 842	struct nfp_nfd3_tx_buf *txbuf;
 843	struct nfp_nfd3_tx_desc *txd;
 844	int wr_idx;
 845
 846	/* Reject if xdp_adjust_tail grow packet beyond DMA area */
 847	if (pkt_len + dma_off > dma_map_sz)
 848		return false;
 849
 850	if (unlikely(nfp_net_tx_full(tx_ring, 1))) {
 851		if (!*completed) {
 852			nfp_nfd3_xdp_complete(tx_ring);
 853			*completed = true;
 854		}
 855
 856		if (unlikely(nfp_net_tx_full(tx_ring, 1))) {
 857			nfp_nfd3_rx_drop(dp, rx_ring->r_vec, rx_ring, rxbuf,
 858					 NULL);
 859			return false;
 860		}
 861	}
 862
 863	wr_idx = D_IDX(tx_ring, tx_ring->wr_p);
 864
 865	/* Stash the soft descriptor of the head then initialize it */
 866	txbuf = &tx_ring->txbufs[wr_idx];
 867
 868	nfp_nfd3_rx_give_one(dp, rx_ring, txbuf->frag, txbuf->dma_addr);
 869
 870	txbuf->frag = rxbuf->frag;
 871	txbuf->dma_addr = rxbuf->dma_addr;
 872	txbuf->fidx = -1;
 873	txbuf->pkt_cnt = 1;
 874	txbuf->real_len = pkt_len;
 875
 876	dma_sync_single_for_device(dp->dev, rxbuf->dma_addr + dma_off,
 877				   pkt_len, DMA_BIDIRECTIONAL);
 878
 879	/* Build TX descriptor */
 880	txd = &tx_ring->txds[wr_idx];
 881	txd->offset_eop = NFD3_DESC_TX_EOP;
 882	txd->dma_len = cpu_to_le16(pkt_len);
 883	nfp_desc_set_dma_addr_40b(txd, rxbuf->dma_addr + dma_off);
 884	txd->data_len = cpu_to_le16(pkt_len);
 885
 886	txd->flags = 0;
 887	txd->mss = 0;
 888	txd->lso_hdrlen = 0;
 889
 890	tx_ring->wr_p++;
 891	tx_ring->wr_ptr_add++;
 892	return true;
 893}
 894
 895/**
 896 * nfp_nfd3_rx() - receive up to @budget packets on @rx_ring
 897 * @rx_ring:   RX ring to receive from
 898 * @budget:    NAPI budget
 899 *
 900 * Note, this function is separated out from the napi poll function to
 901 * more cleanly separate packet receive code from other bookkeeping
 902 * functions performed in the napi poll function.
 903 *
 904 * Return: Number of packets received.
 905 */
 906static int nfp_nfd3_rx(struct nfp_net_rx_ring *rx_ring, int budget)
 907{
 908	struct nfp_net_r_vector *r_vec = rx_ring->r_vec;
 909	struct nfp_net_dp *dp = &r_vec->nfp_net->dp;
 910	struct nfp_net_tx_ring *tx_ring;
 911	struct bpf_prog *xdp_prog;
 912	int idx, pkts_polled = 0;
 913	bool xdp_tx_cmpl = false;
 914	unsigned int true_bufsz;
 915	struct sk_buff *skb;
 916	struct xdp_buff xdp;
 917
 918	xdp_prog = READ_ONCE(dp->xdp_prog);
 919	true_bufsz = xdp_prog ? PAGE_SIZE : dp->fl_bufsz;
 920	xdp_init_buff(&xdp, PAGE_SIZE - NFP_NET_RX_BUF_HEADROOM,
 921		      &rx_ring->xdp_rxq);
 922	tx_ring = r_vec->xdp_ring;
 923
 924	while (pkts_polled < budget) {
 925		unsigned int meta_len, data_len, meta_off, pkt_len, pkt_off;
 926		struct nfp_net_rx_buf *rxbuf;
 927		struct nfp_net_rx_desc *rxd;
 928		struct nfp_meta_parsed meta;
 929		bool redir_egress = false;
 930		struct net_device *netdev;
 931		dma_addr_t new_dma_addr;
 932		u32 meta_len_xdp = 0;
 933		void *new_frag;
 934
 935		idx = D_IDX(rx_ring, rx_ring->rd_p);
 936
 937		rxd = &rx_ring->rxds[idx];
 938		if (!(rxd->rxd.meta_len_dd & PCIE_DESC_RX_DD))
 939			break;
 940
 941		/* Memory barrier to ensure that we won't do other reads
 942		 * before the DD bit.
 943		 */
 944		dma_rmb();
 945
 946		memset(&meta, 0, sizeof(meta));
 947
 948		rx_ring->rd_p++;
 949		pkts_polled++;
 950
 951		rxbuf =	&rx_ring->rxbufs[idx];
 952		/*         < meta_len >
 953		 *  <-- [rx_offset] -->
 954		 *  ---------------------------------------------------------
 955		 * | [XX] |  metadata  |             packet           | XXXX |
 956		 *  ---------------------------------------------------------
 957		 *         <---------------- data_len --------------->
 958		 *
 959		 * The rx_offset is fixed for all packets, the meta_len can vary
 960		 * on a packet by packet basis. If rx_offset is set to zero
 961		 * (_RX_OFFSET_DYNAMIC) metadata starts at the beginning of the
 962		 * buffer and is immediately followed by the packet (no [XX]).
 963		 */
 964		meta_len = rxd->rxd.meta_len_dd & PCIE_DESC_RX_META_LEN_MASK;
 965		data_len = le16_to_cpu(rxd->rxd.data_len);
 966		pkt_len = data_len - meta_len;
 967
 968		pkt_off = NFP_NET_RX_BUF_HEADROOM + dp->rx_dma_off;
 969		if (dp->rx_offset == NFP_NET_CFG_RX_OFFSET_DYNAMIC)
 970			pkt_off += meta_len;
 971		else
 972			pkt_off += dp->rx_offset;
 973		meta_off = pkt_off - meta_len;
 974
 975		/* Stats update */
 976		u64_stats_update_begin(&r_vec->rx_sync);
 977		r_vec->rx_pkts++;
 978		r_vec->rx_bytes += pkt_len;
 979		u64_stats_update_end(&r_vec->rx_sync);
 980
 981		if (unlikely(meta_len > NFP_NET_MAX_PREPEND ||
 982			     (dp->rx_offset && meta_len > dp->rx_offset))) {
 983			nn_dp_warn(dp, "oversized RX packet metadata %u\n",
 984				   meta_len);
 985			nfp_nfd3_rx_drop(dp, r_vec, rx_ring, rxbuf, NULL);
 986			continue;
 987		}
 988
 989		nfp_net_dma_sync_cpu_rx(dp, rxbuf->dma_addr + meta_off,
 990					data_len);
 991
 992		if (!dp->chained_metadata_format) {
 993			nfp_nfd3_set_hash_desc(dp->netdev, &meta,
 994					       rxbuf->frag + meta_off, rxd);
 995		} else if (meta_len) {
 996			if (unlikely(nfp_nfd3_parse_meta(dp->netdev, &meta,
 997							 rxbuf->frag + meta_off,
 998							 rxbuf->frag + pkt_off,
 999							 pkt_len, meta_len))) {
1000				nn_dp_warn(dp, "invalid RX packet metadata\n");
1001				nfp_nfd3_rx_drop(dp, r_vec, rx_ring, rxbuf,
1002						 NULL);
1003				continue;
1004			}
1005		}
1006
1007		if (xdp_prog && !meta.portid) {
1008			void *orig_data = rxbuf->frag + pkt_off;
1009			unsigned int dma_off;
1010			int act;
1011
1012			xdp_prepare_buff(&xdp,
1013					 rxbuf->frag + NFP_NET_RX_BUF_HEADROOM,
1014					 pkt_off - NFP_NET_RX_BUF_HEADROOM,
1015					 pkt_len, true);
1016
1017			act = bpf_prog_run_xdp(xdp_prog, &xdp);
1018
1019			pkt_len = xdp.data_end - xdp.data;
1020			pkt_off += xdp.data - orig_data;
1021
1022			switch (act) {
1023			case XDP_PASS:
1024				meta_len_xdp = xdp.data - xdp.data_meta;
1025				break;
1026			case XDP_TX:
1027				dma_off = pkt_off - NFP_NET_RX_BUF_HEADROOM;
1028				if (unlikely(!nfp_nfd3_tx_xdp_buf(dp, rx_ring,
1029								  tx_ring,
1030								  rxbuf,
1031								  dma_off,
1032								  pkt_len,
1033								  &xdp_tx_cmpl)))
1034					trace_xdp_exception(dp->netdev,
1035							    xdp_prog, act);
1036				continue;
1037			default:
1038				bpf_warn_invalid_xdp_action(dp->netdev, xdp_prog, act);
1039				fallthrough;
1040			case XDP_ABORTED:
1041				trace_xdp_exception(dp->netdev, xdp_prog, act);
1042				fallthrough;
1043			case XDP_DROP:
1044				nfp_nfd3_rx_give_one(dp, rx_ring, rxbuf->frag,
1045						     rxbuf->dma_addr);
1046				continue;
1047			}
1048		}
1049
1050		if (likely(!meta.portid)) {
1051			netdev = dp->netdev;
1052		} else if (meta.portid == NFP_META_PORT_ID_CTRL) {
1053			struct nfp_net *nn = netdev_priv(dp->netdev);
1054
1055			nfp_app_ctrl_rx_raw(nn->app, rxbuf->frag + pkt_off,
1056					    pkt_len);
1057			nfp_nfd3_rx_give_one(dp, rx_ring, rxbuf->frag,
1058					     rxbuf->dma_addr);
1059			continue;
1060		} else {
1061			struct nfp_net *nn;
1062
1063			nn = netdev_priv(dp->netdev);
1064			netdev = nfp_app_dev_get(nn->app, meta.portid,
1065						 &redir_egress);
1066			if (unlikely(!netdev)) {
1067				nfp_nfd3_rx_drop(dp, r_vec, rx_ring, rxbuf,
1068						 NULL);
1069				continue;
1070			}
1071
1072			if (nfp_netdev_is_nfp_repr(netdev))
1073				nfp_repr_inc_rx_stats(netdev, pkt_len);
1074		}
1075
1076		skb = napi_build_skb(rxbuf->frag, true_bufsz);
1077		if (unlikely(!skb)) {
1078			nfp_nfd3_rx_drop(dp, r_vec, rx_ring, rxbuf, NULL);
1079			continue;
1080		}
1081		new_frag = nfp_nfd3_napi_alloc_one(dp, &new_dma_addr);
1082		if (unlikely(!new_frag)) {
1083			nfp_nfd3_rx_drop(dp, r_vec, rx_ring, rxbuf, skb);
1084			continue;
1085		}
1086
1087		nfp_net_dma_unmap_rx(dp, rxbuf->dma_addr);
1088
1089		nfp_nfd3_rx_give_one(dp, rx_ring, new_frag, new_dma_addr);
1090
1091		skb_reserve(skb, pkt_off);
1092		skb_put(skb, pkt_len);
1093
1094		skb->mark = meta.mark;
1095		skb_set_hash(skb, meta.hash, meta.hash_type);
1096
1097		skb_record_rx_queue(skb, rx_ring->idx);
1098		skb->protocol = eth_type_trans(skb, netdev);
1099
1100		nfp_nfd3_rx_csum(dp, r_vec, rxd, &meta, skb);
1101
1102#ifdef CONFIG_TLS_DEVICE
1103		if (rxd->rxd.flags & PCIE_DESC_RX_DECRYPTED) {
1104			skb->decrypted = true;
1105			u64_stats_update_begin(&r_vec->rx_sync);
1106			r_vec->hw_tls_rx++;
1107			u64_stats_update_end(&r_vec->rx_sync);
1108		}
1109#endif
1110
1111		if (unlikely(!nfp_net_vlan_strip(skb, rxd, &meta))) {
1112			nfp_nfd3_rx_drop(dp, r_vec, rx_ring, NULL, skb);
1113			continue;
1114		}
1115
1116#ifdef CONFIG_NFP_NET_IPSEC
1117		if (meta.ipsec_saidx != 0 && unlikely(nfp_net_ipsec_rx(&meta, skb))) {
1118			nfp_nfd3_rx_drop(dp, r_vec, rx_ring, NULL, skb);
1119			continue;
1120		}
1121#endif
1122
1123		if (meta_len_xdp)
1124			skb_metadata_set(skb, meta_len_xdp);
1125
1126		if (likely(!redir_egress)) {
1127			napi_gro_receive(&rx_ring->r_vec->napi, skb);
1128		} else {
1129			skb->dev = netdev;
1130			skb_reset_network_header(skb);
1131			__skb_push(skb, ETH_HLEN);
1132			dev_queue_xmit(skb);
1133		}
1134	}
1135
1136	if (xdp_prog) {
1137		if (tx_ring->wr_ptr_add)
1138			nfp_net_tx_xmit_more_flush(tx_ring);
1139		else if (unlikely(tx_ring->wr_p != tx_ring->rd_p) &&
1140			 !xdp_tx_cmpl)
1141			if (!nfp_nfd3_xdp_complete(tx_ring))
1142				pkts_polled = budget;
1143	}
1144
1145	return pkts_polled;
1146}
1147
1148/**
1149 * nfp_nfd3_poll() - napi poll function
1150 * @napi:    NAPI structure
1151 * @budget:  NAPI budget
1152 *
1153 * Return: number of packets polled.
1154 */
1155int nfp_nfd3_poll(struct napi_struct *napi, int budget)
1156{
1157	struct nfp_net_r_vector *r_vec =
1158		container_of(napi, struct nfp_net_r_vector, napi);
1159	unsigned int pkts_polled = 0;
1160
1161	if (r_vec->tx_ring)
1162		nfp_nfd3_tx_complete(r_vec->tx_ring, budget);
1163	if (r_vec->rx_ring)
1164		pkts_polled = nfp_nfd3_rx(r_vec->rx_ring, budget);
1165
1166	if (pkts_polled < budget)
1167		if (napi_complete_done(napi, pkts_polled))
1168			nfp_net_irq_unmask(r_vec->nfp_net, r_vec->irq_entry);
1169
1170	if (r_vec->nfp_net->rx_coalesce_adapt_on && r_vec->rx_ring) {
1171		struct dim_sample dim_sample = {};
1172		u64 pkts, bytes;
1173
1174		pkts = r_vec->rx_pkts;
1175		bytes = r_vec->rx_bytes;
1176
1177		dim_update_sample(r_vec->event_ctr, pkts, bytes, &dim_sample);
1178		net_dim(&r_vec->rx_dim, &dim_sample);
1179	}
1180
1181	if (r_vec->nfp_net->tx_coalesce_adapt_on && r_vec->tx_ring) {
1182		struct dim_sample dim_sample = {};
1183		u64 pkts, bytes;
1184
1185		pkts = r_vec->tx_pkts;
1186		bytes = r_vec->tx_bytes;
1187
1188		dim_update_sample(r_vec->event_ctr, pkts, bytes, &dim_sample);
1189		net_dim(&r_vec->tx_dim, &dim_sample);
1190	}
1191
1192	return pkts_polled;
1193}
1194
1195/* Control device data path
1196 */
1197
1198bool
1199nfp_nfd3_ctrl_tx_one(struct nfp_net *nn, struct nfp_net_r_vector *r_vec,
1200		     struct sk_buff *skb, bool old)
1201{
1202	unsigned int real_len = skb->len, meta_len = 0;
1203	struct nfp_net_tx_ring *tx_ring;
1204	struct nfp_nfd3_tx_buf *txbuf;
1205	struct nfp_nfd3_tx_desc *txd;
1206	struct nfp_net_dp *dp;
1207	dma_addr_t dma_addr;
1208	int wr_idx;
1209
1210	dp = &r_vec->nfp_net->dp;
1211	tx_ring = r_vec->tx_ring;
1212
1213	if (WARN_ON_ONCE(skb_shinfo(skb)->nr_frags)) {
1214		nn_dp_warn(dp, "Driver's CTRL TX does not implement gather\n");
1215		goto err_free;
1216	}
1217
1218	if (unlikely(nfp_net_tx_full(tx_ring, 1))) {
1219		u64_stats_update_begin(&r_vec->tx_sync);
1220		r_vec->tx_busy++;
1221		u64_stats_update_end(&r_vec->tx_sync);
1222		if (!old)
1223			__skb_queue_tail(&r_vec->queue, skb);
1224		else
1225			__skb_queue_head(&r_vec->queue, skb);
1226		return true;
1227	}
1228
1229	if (nfp_app_ctrl_has_meta(nn->app)) {
1230		if (unlikely(skb_headroom(skb) < 8)) {
1231			nn_dp_warn(dp, "CTRL TX on skb without headroom\n");
1232			goto err_free;
1233		}
1234		meta_len = 8;
1235		put_unaligned_be32(NFP_META_PORT_ID_CTRL, skb_push(skb, 4));
1236		put_unaligned_be32(NFP_NET_META_PORTID, skb_push(skb, 4));
1237	}
1238
1239	/* Start with the head skbuf */
1240	dma_addr = dma_map_single(dp->dev, skb->data, skb_headlen(skb),
1241				  DMA_TO_DEVICE);
1242	if (dma_mapping_error(dp->dev, dma_addr))
1243		goto err_dma_warn;
1244
1245	wr_idx = D_IDX(tx_ring, tx_ring->wr_p);
1246
1247	/* Stash the soft descriptor of the head then initialize it */
1248	txbuf = &tx_ring->txbufs[wr_idx];
1249	txbuf->skb = skb;
1250	txbuf->dma_addr = dma_addr;
1251	txbuf->fidx = -1;
1252	txbuf->pkt_cnt = 1;
1253	txbuf->real_len = real_len;
1254
1255	/* Build TX descriptor */
1256	txd = &tx_ring->txds[wr_idx];
1257	txd->offset_eop = meta_len | NFD3_DESC_TX_EOP;
1258	txd->dma_len = cpu_to_le16(skb_headlen(skb));
1259	nfp_desc_set_dma_addr_40b(txd, dma_addr);
1260	txd->data_len = cpu_to_le16(skb->len);
1261
1262	txd->flags = 0;
1263	txd->mss = 0;
1264	txd->lso_hdrlen = 0;
1265
1266	tx_ring->wr_p++;
1267	tx_ring->wr_ptr_add++;
1268	nfp_net_tx_xmit_more_flush(tx_ring);
1269
1270	return false;
1271
1272err_dma_warn:
1273	nn_dp_warn(dp, "Failed to DMA map TX CTRL buffer\n");
1274err_free:
1275	u64_stats_update_begin(&r_vec->tx_sync);
1276	r_vec->tx_errors++;
1277	u64_stats_update_end(&r_vec->tx_sync);
1278	dev_kfree_skb_any(skb);
1279	return false;
1280}
1281
1282static void __nfp_ctrl_tx_queued(struct nfp_net_r_vector *r_vec)
1283{
1284	struct sk_buff *skb;
1285
1286	while ((skb = __skb_dequeue(&r_vec->queue)))
1287		if (nfp_nfd3_ctrl_tx_one(r_vec->nfp_net, r_vec, skb, true))
1288			return;
1289}
1290
1291static bool
1292nfp_ctrl_meta_ok(struct nfp_net *nn, void *data, unsigned int meta_len)
1293{
1294	u32 meta_type, meta_tag;
1295
1296	if (!nfp_app_ctrl_has_meta(nn->app))
1297		return !meta_len;
1298
1299	if (meta_len != 8)
1300		return false;
1301
1302	meta_type = get_unaligned_be32(data);
1303	meta_tag = get_unaligned_be32(data + 4);
1304
1305	return (meta_type == NFP_NET_META_PORTID &&
1306		meta_tag == NFP_META_PORT_ID_CTRL);
1307}
1308
1309static bool
1310nfp_ctrl_rx_one(struct nfp_net *nn, struct nfp_net_dp *dp,
1311		struct nfp_net_r_vector *r_vec, struct nfp_net_rx_ring *rx_ring)
1312{
1313	unsigned int meta_len, data_len, meta_off, pkt_len, pkt_off;
1314	struct nfp_net_rx_buf *rxbuf;
1315	struct nfp_net_rx_desc *rxd;
1316	dma_addr_t new_dma_addr;
1317	struct sk_buff *skb;
1318	void *new_frag;
1319	int idx;
1320
1321	idx = D_IDX(rx_ring, rx_ring->rd_p);
1322
1323	rxd = &rx_ring->rxds[idx];
1324	if (!(rxd->rxd.meta_len_dd & PCIE_DESC_RX_DD))
1325		return false;
1326
1327	/* Memory barrier to ensure that we won't do other reads
1328	 * before the DD bit.
1329	 */
1330	dma_rmb();
1331
1332	rx_ring->rd_p++;
1333
1334	rxbuf =	&rx_ring->rxbufs[idx];
1335	meta_len = rxd->rxd.meta_len_dd & PCIE_DESC_RX_META_LEN_MASK;
1336	data_len = le16_to_cpu(rxd->rxd.data_len);
1337	pkt_len = data_len - meta_len;
1338
1339	pkt_off = NFP_NET_RX_BUF_HEADROOM + dp->rx_dma_off;
1340	if (dp->rx_offset == NFP_NET_CFG_RX_OFFSET_DYNAMIC)
1341		pkt_off += meta_len;
1342	else
1343		pkt_off += dp->rx_offset;
1344	meta_off = pkt_off - meta_len;
1345
1346	/* Stats update */
1347	u64_stats_update_begin(&r_vec->rx_sync);
1348	r_vec->rx_pkts++;
1349	r_vec->rx_bytes += pkt_len;
1350	u64_stats_update_end(&r_vec->rx_sync);
1351
1352	nfp_net_dma_sync_cpu_rx(dp, rxbuf->dma_addr + meta_off,	data_len);
1353
1354	if (unlikely(!nfp_ctrl_meta_ok(nn, rxbuf->frag + meta_off, meta_len))) {
1355		nn_dp_warn(dp, "incorrect metadata for ctrl packet (%d)\n",
1356			   meta_len);
1357		nfp_nfd3_rx_drop(dp, r_vec, rx_ring, rxbuf, NULL);
1358		return true;
1359	}
1360
1361	skb = build_skb(rxbuf->frag, dp->fl_bufsz);
1362	if (unlikely(!skb)) {
1363		nfp_nfd3_rx_drop(dp, r_vec, rx_ring, rxbuf, NULL);
1364		return true;
1365	}
1366	new_frag = nfp_nfd3_napi_alloc_one(dp, &new_dma_addr);
1367	if (unlikely(!new_frag)) {
1368		nfp_nfd3_rx_drop(dp, r_vec, rx_ring, rxbuf, skb);
1369		return true;
1370	}
1371
1372	nfp_net_dma_unmap_rx(dp, rxbuf->dma_addr);
1373
1374	nfp_nfd3_rx_give_one(dp, rx_ring, new_frag, new_dma_addr);
1375
1376	skb_reserve(skb, pkt_off);
1377	skb_put(skb, pkt_len);
1378
1379	nfp_app_ctrl_rx(nn->app, skb);
1380
1381	return true;
1382}
1383
1384static bool nfp_ctrl_rx(struct nfp_net_r_vector *r_vec)
1385{
1386	struct nfp_net_rx_ring *rx_ring = r_vec->rx_ring;
1387	struct nfp_net *nn = r_vec->nfp_net;
1388	struct nfp_net_dp *dp = &nn->dp;
1389	unsigned int budget = 512;
1390
1391	while (nfp_ctrl_rx_one(nn, dp, r_vec, rx_ring) && budget--)
1392		continue;
1393
1394	return budget;
1395}
1396
1397void nfp_nfd3_ctrl_poll(struct tasklet_struct *t)
1398{
1399	struct nfp_net_r_vector *r_vec = from_tasklet(r_vec, t, tasklet);
1400
1401	spin_lock(&r_vec->lock);
1402	nfp_nfd3_tx_complete(r_vec->tx_ring, 0);
1403	__nfp_ctrl_tx_queued(r_vec);
1404	spin_unlock(&r_vec->lock);
1405
1406	if (nfp_ctrl_rx(r_vec)) {
1407		nfp_net_irq_unmask(r_vec->nfp_net, r_vec->irq_entry);
1408	} else {
1409		tasklet_schedule(&r_vec->tasklet);
1410		nn_dp_warn(&r_vec->nfp_net->dp,
1411			   "control message budget exceeded!\n");
1412	}
1413}