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/overflow.h>
   7#include <linux/sizes.h>
   8#include <linux/bitfield.h>
   9#include <net/xfrm.h>
  10
  11#include "../nfp_app.h"
  12#include "../nfp_net.h"
  13#include "../nfp_net_dp.h"
  14#include "../crypto/crypto.h"
  15#include "../crypto/fw.h"
  16#include "nfdk.h"
  17
  18static int nfp_nfdk_tx_ring_should_wake(struct nfp_net_tx_ring *tx_ring)
  19{
  20	return !nfp_net_tx_full(tx_ring, NFDK_TX_DESC_STOP_CNT * 2);
  21}
  22
  23static int nfp_nfdk_tx_ring_should_stop(struct nfp_net_tx_ring *tx_ring)
  24{
  25	return nfp_net_tx_full(tx_ring, NFDK_TX_DESC_STOP_CNT);
  26}
  27
  28static void nfp_nfdk_tx_ring_stop(struct netdev_queue *nd_q,
  29				  struct nfp_net_tx_ring *tx_ring)
  30{
  31	netif_tx_stop_queue(nd_q);
  32
  33	/* We can race with the TX completion out of NAPI so recheck */
  34	smp_mb();
  35	if (unlikely(nfp_nfdk_tx_ring_should_wake(tx_ring)))
  36		netif_tx_start_queue(nd_q);
  37}
  38
  39static __le64
  40nfp_nfdk_tx_tso(struct nfp_net_r_vector *r_vec, struct nfp_nfdk_tx_buf *txbuf,
  41		struct sk_buff *skb)
  42{
  43	u32 segs, hdrlen, l3_offset, l4_offset, l4_hdrlen;
  44	struct nfp_nfdk_tx_desc txd;
  45	u16 mss;
  46
  47	if (!skb->encapsulation) {
  48		l3_offset = skb_network_offset(skb);
  49		l4_offset = skb_transport_offset(skb);
  50		l4_hdrlen = (skb_shinfo(skb)->gso_type & SKB_GSO_UDP_L4) ?
  51			    sizeof(struct udphdr) : tcp_hdrlen(skb);
  52	} else {
  53		l3_offset = skb_inner_network_offset(skb);
  54		l4_offset = skb_inner_transport_offset(skb);
  55		l4_hdrlen = (skb_shinfo(skb)->gso_type & SKB_GSO_UDP_L4) ?
  56			    sizeof(struct udphdr) : inner_tcp_hdrlen(skb);
  57	}
  58
  59	hdrlen = l4_offset + l4_hdrlen;
  60	segs = skb_shinfo(skb)->gso_segs;
  61	mss = skb_shinfo(skb)->gso_size & NFDK_DESC_TX_MSS_MASK;
  62
  63	txd.l3_offset = l3_offset;
  64	txd.l4_offset = l4_offset;
  65	txd.lso_meta_res = 0;
  66	txd.mss = cpu_to_le16(mss);
  67	txd.lso_hdrlen = hdrlen;
  68	txd.lso_totsegs = segs;
  69
  70	txbuf->pkt_cnt = segs;
  71	txbuf->real_len = skb->len + hdrlen * (txbuf->pkt_cnt - 1);
  72
  73	u64_stats_update_begin(&r_vec->tx_sync);
  74	r_vec->tx_lso++;
  75	u64_stats_update_end(&r_vec->tx_sync);
  76
  77	return txd.raw;
  78}
  79
  80static u8
  81nfp_nfdk_tx_csum(struct nfp_net_dp *dp, struct nfp_net_r_vector *r_vec,
  82		 unsigned int pkt_cnt, struct sk_buff *skb, u64 flags)
  83{
  84	struct ipv6hdr *ipv6h;
  85	struct iphdr *iph;
  86
  87	if (!(dp->ctrl & NFP_NET_CFG_CTRL_TXCSUM))
  88		return flags;
  89
  90	if (skb->ip_summed != CHECKSUM_PARTIAL)
  91		return flags;
  92
  93	flags |= NFDK_DESC_TX_L4_CSUM;
  94
  95	iph = skb->encapsulation ? inner_ip_hdr(skb) : ip_hdr(skb);
  96	ipv6h = skb->encapsulation ? inner_ipv6_hdr(skb) : ipv6_hdr(skb);
  97
  98	/* L3 checksum offloading flag is not required for ipv6 */
  99	if (iph->version == 4) {
 100		flags |= NFDK_DESC_TX_L3_CSUM;
 101	} else if (ipv6h->version != 6) {
 102		nn_dp_warn(dp, "partial checksum but ipv=%x!\n", iph->version);
 103		return flags;
 104	}
 105
 106	u64_stats_update_begin(&r_vec->tx_sync);
 107	if (!skb->encapsulation) {
 108		r_vec->hw_csum_tx += pkt_cnt;
 109	} else {
 110		flags |= NFDK_DESC_TX_ENCAP;
 111		r_vec->hw_csum_tx_inner += pkt_cnt;
 112	}
 113	u64_stats_update_end(&r_vec->tx_sync);
 114
 115	return flags;
 116}
 117
 118static int
 119nfp_nfdk_tx_maybe_close_block(struct nfp_net_tx_ring *tx_ring,
 120			      struct sk_buff *skb)
 121{
 122	unsigned int n_descs, wr_p, nop_slots;
 123	const skb_frag_t *frag, *fend;
 124	struct nfp_nfdk_tx_desc *txd;
 125	unsigned int nr_frags;
 126	unsigned int wr_idx;
 127	int err;
 128
 129recount_descs:
 130	n_descs = nfp_nfdk_headlen_to_segs(skb_headlen(skb));
 131	nr_frags = skb_shinfo(skb)->nr_frags;
 132	frag = skb_shinfo(skb)->frags;
 133	fend = frag + nr_frags;
 134	for (; frag < fend; frag++)
 135		n_descs += DIV_ROUND_UP(skb_frag_size(frag),
 136					NFDK_TX_MAX_DATA_PER_DESC);
 137
 138	if (unlikely(n_descs > NFDK_TX_DESC_GATHER_MAX)) {
 139		if (skb_is_nonlinear(skb)) {
 140			err = skb_linearize(skb);
 141			if (err)
 142				return err;
 143			goto recount_descs;
 144		}
 145		return -EINVAL;
 146	}
 147
 148	/* Under count by 1 (don't count meta) for the round down to work out */
 149	n_descs += !!skb_is_gso(skb);
 150
 151	if (round_down(tx_ring->wr_p, NFDK_TX_DESC_BLOCK_CNT) !=
 152	    round_down(tx_ring->wr_p + n_descs, NFDK_TX_DESC_BLOCK_CNT))
 153		goto close_block;
 154
 155	if ((u32)tx_ring->data_pending + skb->len > NFDK_TX_MAX_DATA_PER_BLOCK)
 156		goto close_block;
 157
 158	return 0;
 159
 160close_block:
 161	wr_p = tx_ring->wr_p;
 162	nop_slots = D_BLOCK_CPL(wr_p);
 163
 164	wr_idx = D_IDX(tx_ring, wr_p);
 165	tx_ring->ktxbufs[wr_idx].skb = NULL;
 166	txd = &tx_ring->ktxds[wr_idx];
 167
 168	memset(txd, 0, array_size(nop_slots, sizeof(struct nfp_nfdk_tx_desc)));
 169
 170	tx_ring->data_pending = 0;
 171	tx_ring->wr_p += nop_slots;
 172	tx_ring->wr_ptr_add += nop_slots;
 173
 174	return 0;
 175}
 176
 177static int
 178nfp_nfdk_prep_tx_meta(struct nfp_net_dp *dp, struct nfp_app *app,
 179		      struct sk_buff *skb, bool *ipsec)
 180{
 181	struct metadata_dst *md_dst = skb_metadata_dst(skb);
 182	struct nfp_ipsec_offload offload_info;
 183	unsigned char *data;
 184	bool vlan_insert;
 185	u32 meta_id = 0;
 186	int md_bytes;
 187
 188#ifdef CONFIG_NFP_NET_IPSEC
 189	if (xfrm_offload(skb))
 190		*ipsec = nfp_net_ipsec_tx_prep(dp, skb, &offload_info);
 191#endif
 192
 193	if (unlikely(md_dst && md_dst->type != METADATA_HW_PORT_MUX))
 194		md_dst = NULL;
 195
 196	vlan_insert = skb_vlan_tag_present(skb) && (dp->ctrl & NFP_NET_CFG_CTRL_TXVLAN_V2);
 197
 198	if (!(md_dst || vlan_insert || *ipsec))
 199		return 0;
 200
 201	md_bytes = sizeof(meta_id) +
 202		   (!!md_dst ? NFP_NET_META_PORTID_SIZE : 0) +
 203		   (vlan_insert ? NFP_NET_META_VLAN_SIZE : 0) +
 204		   (*ipsec ? NFP_NET_META_IPSEC_FIELD_SIZE : 0);
 205
 206	if (unlikely(skb_cow_head(skb, md_bytes)))
 207		return -ENOMEM;
 208
 209	data = skb_push(skb, md_bytes) + md_bytes;
 210	if (md_dst) {
 211		data -= NFP_NET_META_PORTID_SIZE;
 212		put_unaligned_be32(md_dst->u.port_info.port_id, data);
 213		meta_id = NFP_NET_META_PORTID;
 214	}
 215	if (vlan_insert) {
 216		data -= NFP_NET_META_VLAN_SIZE;
 217		/* data type of skb->vlan_proto is __be16
 218		 * so it fills metadata without calling put_unaligned_be16
 219		 */
 220		memcpy(data, &skb->vlan_proto, sizeof(skb->vlan_proto));
 221		put_unaligned_be16(skb_vlan_tag_get(skb), data + sizeof(skb->vlan_proto));
 222		meta_id <<= NFP_NET_META_FIELD_SIZE;
 223		meta_id |= NFP_NET_META_VLAN;
 224	}
 225
 226	if (*ipsec) {
 227		data -= NFP_NET_META_IPSEC_SIZE;
 228		put_unaligned_be32(offload_info.seq_hi, data);
 229		data -= NFP_NET_META_IPSEC_SIZE;
 230		put_unaligned_be32(offload_info.seq_low, data);
 231		data -= NFP_NET_META_IPSEC_SIZE;
 232		put_unaligned_be32(offload_info.handle - 1, data);
 233		meta_id <<= NFP_NET_META_IPSEC_FIELD_SIZE;
 234		meta_id |= NFP_NET_META_IPSEC << 8 | NFP_NET_META_IPSEC << 4 | NFP_NET_META_IPSEC;
 235	}
 236
 237	meta_id = FIELD_PREP(NFDK_META_LEN, md_bytes) |
 238		  FIELD_PREP(NFDK_META_FIELDS, meta_id);
 239
 240	data -= sizeof(meta_id);
 241	put_unaligned_be32(meta_id, data);
 242
 243	return NFDK_DESC_TX_CHAIN_META;
 244}
 245
 246/**
 247 * nfp_nfdk_tx() - Main transmit entry point
 248 * @skb:    SKB to transmit
 249 * @netdev: netdev structure
 250 *
 251 * Return: NETDEV_TX_OK on success.
 252 */
 253netdev_tx_t nfp_nfdk_tx(struct sk_buff *skb, struct net_device *netdev)
 254{
 255	struct nfp_net *nn = netdev_priv(netdev);
 256	struct nfp_nfdk_tx_buf *txbuf, *etxbuf;
 257	u32 cnt, tmp_dlen, dlen_type = 0;
 258	struct nfp_net_tx_ring *tx_ring;
 259	struct nfp_net_r_vector *r_vec;
 260	const skb_frag_t *frag, *fend;
 261	struct nfp_nfdk_tx_desc *txd;
 262	unsigned int real_len, qidx;
 263	unsigned int dma_len, type;
 264	struct netdev_queue *nd_q;
 265	struct nfp_net_dp *dp;
 266	int nr_frags, wr_idx;
 267	dma_addr_t dma_addr;
 268	bool ipsec = false;
 269	u64 metadata;
 270
 271	dp = &nn->dp;
 272	qidx = skb_get_queue_mapping(skb);
 273	tx_ring = &dp->tx_rings[qidx];
 274	r_vec = tx_ring->r_vec;
 275	nd_q = netdev_get_tx_queue(dp->netdev, qidx);
 276
 277	/* Don't bother counting frags, assume the worst */
 278	if (unlikely(nfp_net_tx_full(tx_ring, NFDK_TX_DESC_STOP_CNT))) {
 279		nn_dp_warn(dp, "TX ring %d busy. wrp=%u rdp=%u\n",
 280			   qidx, tx_ring->wr_p, tx_ring->rd_p);
 281		netif_tx_stop_queue(nd_q);
 282		nfp_net_tx_xmit_more_flush(tx_ring);
 283		u64_stats_update_begin(&r_vec->tx_sync);
 284		r_vec->tx_busy++;
 285		u64_stats_update_end(&r_vec->tx_sync);
 286		return NETDEV_TX_BUSY;
 287	}
 288
 289	metadata = nfp_nfdk_prep_tx_meta(dp, nn->app, skb, &ipsec);
 290	if (unlikely((int)metadata < 0))
 291		goto err_flush;
 292
 293	if (nfp_nfdk_tx_maybe_close_block(tx_ring, skb))
 294		goto err_flush;
 295
 296	/* nr_frags will change after skb_linearize so we get nr_frags after
 297	 * nfp_nfdk_tx_maybe_close_block function
 298	 */
 299	nr_frags = skb_shinfo(skb)->nr_frags;
 300	/* DMA map all */
 301	wr_idx = D_IDX(tx_ring, tx_ring->wr_p);
 302	txd = &tx_ring->ktxds[wr_idx];
 303	txbuf = &tx_ring->ktxbufs[wr_idx];
 304
 305	dma_len = skb_headlen(skb);
 306	if (skb_is_gso(skb))
 307		type = NFDK_DESC_TX_TYPE_TSO;
 308	else if (!nr_frags && dma_len <= NFDK_TX_MAX_DATA_PER_HEAD)
 309		type = NFDK_DESC_TX_TYPE_SIMPLE;
 310	else
 311		type = NFDK_DESC_TX_TYPE_GATHER;
 312
 313	dma_addr = dma_map_single(dp->dev, skb->data, dma_len, DMA_TO_DEVICE);
 314	if (dma_mapping_error(dp->dev, dma_addr))
 315		goto err_warn_dma;
 316
 317	txbuf->skb = skb;
 318	txbuf++;
 319
 320	txbuf->dma_addr = dma_addr;
 321	txbuf++;
 322
 323	/* FIELD_PREP() implicitly truncates to chunk */
 324	dma_len -= 1;
 325
 326	/* We will do our best to pass as much data as we can in descriptor
 327	 * and we need to make sure the first descriptor includes whole head
 328	 * since there is limitation in firmware side. Sometimes the value of
 329	 * dma_len bitwise and NFDK_DESC_TX_DMA_LEN_HEAD will less than
 330	 * headlen.
 331	 */
 332	dlen_type = FIELD_PREP(NFDK_DESC_TX_DMA_LEN_HEAD,
 333			       dma_len > NFDK_DESC_TX_DMA_LEN_HEAD ?
 334			       NFDK_DESC_TX_DMA_LEN_HEAD : dma_len) |
 335		    FIELD_PREP(NFDK_DESC_TX_TYPE_HEAD, type);
 336
 337	txd->dma_len_type = cpu_to_le16(dlen_type);
 338	nfp_desc_set_dma_addr_48b(txd, dma_addr);
 339
 340	/* starts at bit 0 */
 341	BUILD_BUG_ON(!(NFDK_DESC_TX_DMA_LEN_HEAD & 1));
 342
 343	/* Preserve the original dlen_type, this way below the EOP logic
 344	 * can use dlen_type.
 345	 */
 346	tmp_dlen = dlen_type & NFDK_DESC_TX_DMA_LEN_HEAD;
 347	dma_len -= tmp_dlen;
 348	dma_addr += tmp_dlen + 1;
 349	txd++;
 350
 351	/* The rest of the data (if any) will be in larger dma descritors
 352	 * and is handled with the fragment loop.
 353	 */
 354	frag = skb_shinfo(skb)->frags;
 355	fend = frag + nr_frags;
 356
 357	while (true) {
 358		while (dma_len > 0) {
 359			dma_len -= 1;
 360			dlen_type = FIELD_PREP(NFDK_DESC_TX_DMA_LEN, dma_len);
 361
 362			txd->dma_len_type = cpu_to_le16(dlen_type);
 363			nfp_desc_set_dma_addr_48b(txd, dma_addr);
 364
 365			dma_len -= dlen_type;
 366			dma_addr += dlen_type + 1;
 367			txd++;
 368		}
 369
 370		if (frag >= fend)
 371			break;
 372
 373		dma_len = skb_frag_size(frag);
 374		dma_addr = skb_frag_dma_map(dp->dev, frag, 0, dma_len,
 375					    DMA_TO_DEVICE);
 376		if (dma_mapping_error(dp->dev, dma_addr))
 377			goto err_unmap;
 378
 379		txbuf->dma_addr = dma_addr;
 380		txbuf++;
 381
 382		frag++;
 383	}
 384
 385	(txd - 1)->dma_len_type = cpu_to_le16(dlen_type | NFDK_DESC_TX_EOP);
 386
 387	if (ipsec)
 388		metadata = nfp_nfdk_ipsec_tx(metadata, skb);
 389
 390	if (!skb_is_gso(skb)) {
 391		real_len = skb->len;
 392		/* Metadata desc */
 393		if (!ipsec)
 394			metadata = nfp_nfdk_tx_csum(dp, r_vec, 1, skb, metadata);
 395		txd->raw = cpu_to_le64(metadata);
 396		txd++;
 397	} else {
 398		/* lso desc should be placed after metadata desc */
 399		(txd + 1)->raw = nfp_nfdk_tx_tso(r_vec, txbuf, skb);
 400		real_len = txbuf->real_len;
 401		/* Metadata desc */
 402		if (!ipsec)
 403			metadata = nfp_nfdk_tx_csum(dp, r_vec, txbuf->pkt_cnt, skb, metadata);
 404		txd->raw = cpu_to_le64(metadata);
 405		txd += 2;
 406		txbuf++;
 407	}
 408
 409	cnt = txd - tx_ring->ktxds - wr_idx;
 410	if (unlikely(round_down(wr_idx, NFDK_TX_DESC_BLOCK_CNT) !=
 411		     round_down(wr_idx + cnt - 1, NFDK_TX_DESC_BLOCK_CNT)))
 412		goto err_warn_overflow;
 413
 414	skb_tx_timestamp(skb);
 415
 416	tx_ring->wr_p += cnt;
 417	if (tx_ring->wr_p % NFDK_TX_DESC_BLOCK_CNT)
 418		tx_ring->data_pending += skb->len;
 419	else
 420		tx_ring->data_pending = 0;
 421
 422	if (nfp_nfdk_tx_ring_should_stop(tx_ring))
 423		nfp_nfdk_tx_ring_stop(nd_q, tx_ring);
 424
 425	tx_ring->wr_ptr_add += cnt;
 426	if (__netdev_tx_sent_queue(nd_q, real_len, netdev_xmit_more()))
 427		nfp_net_tx_xmit_more_flush(tx_ring);
 428
 429	return NETDEV_TX_OK;
 430
 431err_warn_overflow:
 432	WARN_ONCE(1, "unable to fit packet into a descriptor wr_idx:%d head:%d frags:%d cnt:%d",
 433		  wr_idx, skb_headlen(skb), nr_frags, cnt);
 434	if (skb_is_gso(skb))
 435		txbuf--;
 436err_unmap:
 437	/* txbuf pointed to the next-to-use */
 438	etxbuf = txbuf;
 439	/* first txbuf holds the skb */
 440	txbuf = &tx_ring->ktxbufs[wr_idx + 1];
 441	if (txbuf < etxbuf) {
 442		dma_unmap_single(dp->dev, txbuf->dma_addr,
 443				 skb_headlen(skb), DMA_TO_DEVICE);
 444		txbuf->raw = 0;
 445		txbuf++;
 446	}
 447	frag = skb_shinfo(skb)->frags;
 448	while (etxbuf < txbuf) {
 449		dma_unmap_page(dp->dev, txbuf->dma_addr,
 450			       skb_frag_size(frag), DMA_TO_DEVICE);
 451		txbuf->raw = 0;
 452		frag++;
 453		txbuf++;
 454	}
 455err_warn_dma:
 456	nn_dp_warn(dp, "Failed to map DMA TX buffer\n");
 457err_flush:
 458	nfp_net_tx_xmit_more_flush(tx_ring);
 459	u64_stats_update_begin(&r_vec->tx_sync);
 460	r_vec->tx_errors++;
 461	u64_stats_update_end(&r_vec->tx_sync);
 462	dev_kfree_skb_any(skb);
 463	return NETDEV_TX_OK;
 464}
 465
 466/**
 467 * nfp_nfdk_tx_complete() - Handled completed TX packets
 468 * @tx_ring:	TX ring structure
 469 * @budget:	NAPI budget (only used as bool to determine if in NAPI context)
 470 */
 471static void nfp_nfdk_tx_complete(struct nfp_net_tx_ring *tx_ring, int budget)
 472{
 473	struct nfp_net_r_vector *r_vec = tx_ring->r_vec;
 474	struct nfp_net_dp *dp = &r_vec->nfp_net->dp;
 475	u32 done_pkts = 0, done_bytes = 0;
 476	struct nfp_nfdk_tx_buf *ktxbufs;
 477	struct device *dev = dp->dev;
 478	struct netdev_queue *nd_q;
 479	u32 rd_p, qcp_rd_p;
 480	int todo;
 481
 482	rd_p = tx_ring->rd_p;
 483	if (tx_ring->wr_p == rd_p)
 484		return;
 485
 486	/* Work out how many descriptors have been transmitted */
 487	qcp_rd_p = nfp_net_read_tx_cmpl(tx_ring, dp);
 488
 489	if (qcp_rd_p == tx_ring->qcp_rd_p)
 490		return;
 491
 492	todo = D_IDX(tx_ring, qcp_rd_p - tx_ring->qcp_rd_p);
 493	ktxbufs = tx_ring->ktxbufs;
 494
 495	while (todo > 0) {
 496		const skb_frag_t *frag, *fend;
 497		unsigned int size, n_descs = 1;
 498		struct nfp_nfdk_tx_buf *txbuf;
 499		struct sk_buff *skb;
 500
 501		txbuf = &ktxbufs[D_IDX(tx_ring, rd_p)];
 502		skb = txbuf->skb;
 503		txbuf++;
 504
 505		/* Closed block */
 506		if (!skb) {
 507			n_descs = D_BLOCK_CPL(rd_p);
 508			goto next;
 509		}
 510
 511		/* Unmap head */
 512		size = skb_headlen(skb);
 513		n_descs += nfp_nfdk_headlen_to_segs(size);
 514		dma_unmap_single(dev, txbuf->dma_addr, size, DMA_TO_DEVICE);
 515		txbuf++;
 516
 517		/* Unmap frags */
 518		frag = skb_shinfo(skb)->frags;
 519		fend = frag + skb_shinfo(skb)->nr_frags;
 520		for (; frag < fend; frag++) {
 521			size = skb_frag_size(frag);
 522			n_descs += DIV_ROUND_UP(size,
 523						NFDK_TX_MAX_DATA_PER_DESC);
 524			dma_unmap_page(dev, txbuf->dma_addr,
 525				       skb_frag_size(frag), DMA_TO_DEVICE);
 526			txbuf++;
 527		}
 528
 529		if (!skb_is_gso(skb)) {
 530			done_bytes += skb->len;
 531			done_pkts++;
 532		} else {
 533			done_bytes += txbuf->real_len;
 534			done_pkts += txbuf->pkt_cnt;
 535			n_descs++;
 536		}
 537
 538		napi_consume_skb(skb, budget);
 539next:
 540		rd_p += n_descs;
 541		todo -= n_descs;
 542	}
 543
 544	tx_ring->rd_p = rd_p;
 545	tx_ring->qcp_rd_p = qcp_rd_p;
 546
 547	u64_stats_update_begin(&r_vec->tx_sync);
 548	r_vec->tx_bytes += done_bytes;
 549	r_vec->tx_pkts += done_pkts;
 550	u64_stats_update_end(&r_vec->tx_sync);
 551
 552	if (!dp->netdev)
 553		return;
 554
 555	nd_q = netdev_get_tx_queue(dp->netdev, tx_ring->idx);
 556	netdev_tx_completed_queue(nd_q, done_pkts, done_bytes);
 557	if (nfp_nfdk_tx_ring_should_wake(tx_ring)) {
 558		/* Make sure TX thread will see updated tx_ring->rd_p */
 559		smp_mb();
 560
 561		if (unlikely(netif_tx_queue_stopped(nd_q)))
 562			netif_tx_wake_queue(nd_q);
 563	}
 564
 565	WARN_ONCE(tx_ring->wr_p - tx_ring->rd_p > tx_ring->cnt,
 566		  "TX ring corruption rd_p=%u wr_p=%u cnt=%u\n",
 567		  tx_ring->rd_p, tx_ring->wr_p, tx_ring->cnt);
 568}
 569
 570/* Receive processing */
 571static void *
 572nfp_nfdk_napi_alloc_one(struct nfp_net_dp *dp, dma_addr_t *dma_addr)
 573{
 574	void *frag;
 575
 576	if (!dp->xdp_prog) {
 577		frag = napi_alloc_frag(dp->fl_bufsz);
 578		if (unlikely(!frag))
 579			return NULL;
 580	} else {
 581		struct page *page;
 582
 583		page = dev_alloc_page();
 584		if (unlikely(!page))
 585			return NULL;
 586		frag = page_address(page);
 587	}
 588
 589	*dma_addr = nfp_net_dma_map_rx(dp, frag);
 590	if (dma_mapping_error(dp->dev, *dma_addr)) {
 591		nfp_net_free_frag(frag, dp->xdp_prog);
 592		nn_dp_warn(dp, "Failed to map DMA RX buffer\n");
 593		return NULL;
 594	}
 595
 596	return frag;
 597}
 598
 599/**
 600 * nfp_nfdk_rx_give_one() - Put mapped skb on the software and hardware rings
 601 * @dp:		NFP Net data path struct
 602 * @rx_ring:	RX ring structure
 603 * @frag:	page fragment buffer
 604 * @dma_addr:	DMA address of skb mapping
 605 */
 606static void
 607nfp_nfdk_rx_give_one(const struct nfp_net_dp *dp,
 608		     struct nfp_net_rx_ring *rx_ring,
 609		     void *frag, dma_addr_t dma_addr)
 610{
 611	unsigned int wr_idx;
 612
 613	wr_idx = D_IDX(rx_ring, rx_ring->wr_p);
 614
 615	nfp_net_dma_sync_dev_rx(dp, dma_addr);
 616
 617	/* Stash SKB and DMA address away */
 618	rx_ring->rxbufs[wr_idx].frag = frag;
 619	rx_ring->rxbufs[wr_idx].dma_addr = dma_addr;
 620
 621	/* Fill freelist descriptor */
 622	rx_ring->rxds[wr_idx].fld.reserved = 0;
 623	rx_ring->rxds[wr_idx].fld.meta_len_dd = 0;
 624	nfp_desc_set_dma_addr_48b(&rx_ring->rxds[wr_idx].fld,
 625				  dma_addr + dp->rx_dma_off);
 626
 627	rx_ring->wr_p++;
 628	if (!(rx_ring->wr_p % NFP_NET_FL_BATCH)) {
 629		/* Update write pointer of the freelist queue. Make
 630		 * sure all writes are flushed before telling the hardware.
 631		 */
 632		wmb();
 633		nfp_qcp_wr_ptr_add(rx_ring->qcp_fl, NFP_NET_FL_BATCH);
 634	}
 635}
 636
 637/**
 638 * nfp_nfdk_rx_ring_fill_freelist() - Give buffers from the ring to FW
 639 * @dp:	     NFP Net data path struct
 640 * @rx_ring: RX ring to fill
 641 */
 642void nfp_nfdk_rx_ring_fill_freelist(struct nfp_net_dp *dp,
 643				    struct nfp_net_rx_ring *rx_ring)
 644{
 645	unsigned int i;
 646
 647	for (i = 0; i < rx_ring->cnt - 1; i++)
 648		nfp_nfdk_rx_give_one(dp, rx_ring, rx_ring->rxbufs[i].frag,
 649				     rx_ring->rxbufs[i].dma_addr);
 650}
 651
 652/**
 653 * nfp_nfdk_rx_csum_has_errors() - group check if rxd has any csum errors
 654 * @flags: RX descriptor flags field in CPU byte order
 655 */
 656static int nfp_nfdk_rx_csum_has_errors(u16 flags)
 657{
 658	u16 csum_all_checked, csum_all_ok;
 659
 660	csum_all_checked = flags & __PCIE_DESC_RX_CSUM_ALL;
 661	csum_all_ok = flags & __PCIE_DESC_RX_CSUM_ALL_OK;
 662
 663	return csum_all_checked != (csum_all_ok << PCIE_DESC_RX_CSUM_OK_SHIFT);
 664}
 665
 666/**
 667 * nfp_nfdk_rx_csum() - set SKB checksum field based on RX descriptor flags
 668 * @dp:  NFP Net data path struct
 669 * @r_vec: per-ring structure
 670 * @rxd: Pointer to RX descriptor
 671 * @meta: Parsed metadata prepend
 672 * @skb: Pointer to SKB
 673 */
 674static void
 675nfp_nfdk_rx_csum(struct nfp_net_dp *dp, struct nfp_net_r_vector *r_vec,
 676		 struct nfp_net_rx_desc *rxd, struct nfp_meta_parsed *meta,
 677		 struct sk_buff *skb)
 678{
 679	skb_checksum_none_assert(skb);
 680
 681	if (!(dp->netdev->features & NETIF_F_RXCSUM))
 682		return;
 683
 684	if (meta->csum_type) {
 685		skb->ip_summed = meta->csum_type;
 686		skb->csum = meta->csum;
 687		u64_stats_update_begin(&r_vec->rx_sync);
 688		r_vec->hw_csum_rx_complete++;
 689		u64_stats_update_end(&r_vec->rx_sync);
 690		return;
 691	}
 692
 693	if (nfp_nfdk_rx_csum_has_errors(le16_to_cpu(rxd->rxd.flags))) {
 694		u64_stats_update_begin(&r_vec->rx_sync);
 695		r_vec->hw_csum_rx_error++;
 696		u64_stats_update_end(&r_vec->rx_sync);
 697		return;
 698	}
 699
 700	/* Assume that the firmware will never report inner CSUM_OK unless outer
 701	 * L4 headers were successfully parsed. FW will always report zero UDP
 702	 * checksum as CSUM_OK.
 703	 */
 704	if (rxd->rxd.flags & PCIE_DESC_RX_TCP_CSUM_OK ||
 705	    rxd->rxd.flags & PCIE_DESC_RX_UDP_CSUM_OK) {
 706		__skb_incr_checksum_unnecessary(skb);
 707		u64_stats_update_begin(&r_vec->rx_sync);
 708		r_vec->hw_csum_rx_ok++;
 709		u64_stats_update_end(&r_vec->rx_sync);
 710	}
 711
 712	if (rxd->rxd.flags & PCIE_DESC_RX_I_TCP_CSUM_OK ||
 713	    rxd->rxd.flags & PCIE_DESC_RX_I_UDP_CSUM_OK) {
 714		__skb_incr_checksum_unnecessary(skb);
 715		u64_stats_update_begin(&r_vec->rx_sync);
 716		r_vec->hw_csum_rx_inner_ok++;
 717		u64_stats_update_end(&r_vec->rx_sync);
 718	}
 719}
 720
 721static void
 722nfp_nfdk_set_hash(struct net_device *netdev, struct nfp_meta_parsed *meta,
 723		  unsigned int type, __be32 *hash)
 724{
 725	if (!(netdev->features & NETIF_F_RXHASH))
 726		return;
 727
 728	switch (type) {
 729	case NFP_NET_RSS_IPV4:
 730	case NFP_NET_RSS_IPV6:
 731	case NFP_NET_RSS_IPV6_EX:
 732		meta->hash_type = PKT_HASH_TYPE_L3;
 733		break;
 734	default:
 735		meta->hash_type = PKT_HASH_TYPE_L4;
 736		break;
 737	}
 738
 739	meta->hash = get_unaligned_be32(hash);
 740}
 741
 742static bool
 743nfp_nfdk_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_nfdk_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 could 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_nfdk_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_nfdk_rx_give_one(dp, rx_ring, rxbuf->frag, rxbuf->dma_addr);
 831	if (skb)
 832		dev_kfree_skb_any(skb);
 833}
 834
 835static bool nfp_nfdk_xdp_complete(struct nfp_net_tx_ring *tx_ring)
 836{
 837	struct nfp_net_r_vector *r_vec = tx_ring->r_vec;
 838	struct nfp_net_dp *dp = &r_vec->nfp_net->dp;
 839	struct nfp_net_rx_ring *rx_ring;
 840	u32 qcp_rd_p, done = 0;
 841	bool done_all;
 842	int todo;
 843
 844	/* Work out how many descriptors have been transmitted */
 845	qcp_rd_p = nfp_net_read_tx_cmpl(tx_ring, dp);
 846	if (qcp_rd_p == tx_ring->qcp_rd_p)
 847		return true;
 848
 849	todo = D_IDX(tx_ring, qcp_rd_p - tx_ring->qcp_rd_p);
 850
 851	done_all = todo <= NFP_NET_XDP_MAX_COMPLETE;
 852	todo = min(todo, NFP_NET_XDP_MAX_COMPLETE);
 853
 854	rx_ring = r_vec->rx_ring;
 855	while (todo > 0) {
 856		int idx = D_IDX(tx_ring, tx_ring->rd_p + done);
 857		struct nfp_nfdk_tx_buf *txbuf;
 858		unsigned int step = 1;
 859
 860		txbuf = &tx_ring->ktxbufs[idx];
 861		if (!txbuf->raw)
 862			goto next;
 863
 864		if (NFDK_TX_BUF_INFO(txbuf->val) != NFDK_TX_BUF_INFO_SOP) {
 865			WARN_ONCE(1, "Unexpected TX buffer in XDP TX ring\n");
 866			goto next;
 867		}
 868
 869		/* Two successive txbufs are used to stash virtual and dma
 870		 * address respectively, recycle and clean them here.
 871		 */
 872		nfp_nfdk_rx_give_one(dp, rx_ring,
 873				     (void *)NFDK_TX_BUF_PTR(txbuf[0].val),
 874				     txbuf[1].dma_addr);
 875		txbuf[0].raw = 0;
 876		txbuf[1].raw = 0;
 877		step = 2;
 878
 879		u64_stats_update_begin(&r_vec->tx_sync);
 880		/* Note: tx_bytes not accumulated. */
 881		r_vec->tx_pkts++;
 882		u64_stats_update_end(&r_vec->tx_sync);
 883next:
 884		todo -= step;
 885		done += step;
 886	}
 887
 888	tx_ring->qcp_rd_p = D_IDX(tx_ring, tx_ring->qcp_rd_p + done);
 889	tx_ring->rd_p += done;
 890
 891	WARN_ONCE(tx_ring->wr_p - tx_ring->rd_p > tx_ring->cnt,
 892		  "XDP TX ring corruption rd_p=%u wr_p=%u cnt=%u\n",
 893		  tx_ring->rd_p, tx_ring->wr_p, tx_ring->cnt);
 894
 895	return done_all;
 896}
 897
 898static bool
 899nfp_nfdk_tx_xdp_buf(struct nfp_net_dp *dp, struct nfp_net_rx_ring *rx_ring,
 900		    struct nfp_net_tx_ring *tx_ring,
 901		    struct nfp_net_rx_buf *rxbuf, unsigned int dma_off,
 902		    unsigned int pkt_len, bool *completed)
 903{
 904	unsigned int dma_map_sz = dp->fl_bufsz - NFP_NET_RX_BUF_NON_DATA;
 905	unsigned int dma_len, type, cnt, dlen_type, tmp_dlen;
 906	struct nfp_nfdk_tx_buf *txbuf;
 907	struct nfp_nfdk_tx_desc *txd;
 908	unsigned int n_descs;
 909	dma_addr_t dma_addr;
 910	int wr_idx;
 911
 912	/* Reject if xdp_adjust_tail grow packet beyond DMA area */
 913	if (pkt_len + dma_off > dma_map_sz)
 914		return false;
 915
 916	/* Make sure there's still at least one block available after
 917	 * aligning to block boundary, so that the txds used below
 918	 * won't wrap around the tx_ring.
 919	 */
 920	if (unlikely(nfp_net_tx_full(tx_ring, NFDK_TX_DESC_STOP_CNT))) {
 921		if (!*completed) {
 922			nfp_nfdk_xdp_complete(tx_ring);
 923			*completed = true;
 924		}
 925
 926		if (unlikely(nfp_net_tx_full(tx_ring, NFDK_TX_DESC_STOP_CNT))) {
 927			nfp_nfdk_rx_drop(dp, rx_ring->r_vec, rx_ring, rxbuf,
 928					 NULL);
 929			return false;
 930		}
 931	}
 932
 933	/* Check if cross block boundary */
 934	n_descs = nfp_nfdk_headlen_to_segs(pkt_len);
 935	if ((round_down(tx_ring->wr_p, NFDK_TX_DESC_BLOCK_CNT) !=
 936	     round_down(tx_ring->wr_p + n_descs, NFDK_TX_DESC_BLOCK_CNT)) ||
 937	    ((u32)tx_ring->data_pending + pkt_len >
 938	     NFDK_TX_MAX_DATA_PER_BLOCK)) {
 939		unsigned int nop_slots = D_BLOCK_CPL(tx_ring->wr_p);
 940
 941		wr_idx = D_IDX(tx_ring, tx_ring->wr_p);
 942		txd = &tx_ring->ktxds[wr_idx];
 943		memset(txd, 0,
 944		       array_size(nop_slots, sizeof(struct nfp_nfdk_tx_desc)));
 945
 946		tx_ring->data_pending = 0;
 947		tx_ring->wr_p += nop_slots;
 948		tx_ring->wr_ptr_add += nop_slots;
 949	}
 950
 951	wr_idx = D_IDX(tx_ring, tx_ring->wr_p);
 952
 953	txbuf = &tx_ring->ktxbufs[wr_idx];
 954
 955	txbuf[0].val = (unsigned long)rxbuf->frag | NFDK_TX_BUF_INFO_SOP;
 956	txbuf[1].dma_addr = rxbuf->dma_addr;
 957	/* Note: pkt len not stored */
 958
 959	dma_sync_single_for_device(dp->dev, rxbuf->dma_addr + dma_off,
 960				   pkt_len, DMA_BIDIRECTIONAL);
 961
 962	/* Build TX descriptor */
 963	txd = &tx_ring->ktxds[wr_idx];
 964	dma_len = pkt_len;
 965	dma_addr = rxbuf->dma_addr + dma_off;
 966
 967	if (dma_len <= NFDK_TX_MAX_DATA_PER_HEAD)
 968		type = NFDK_DESC_TX_TYPE_SIMPLE;
 969	else
 970		type = NFDK_DESC_TX_TYPE_GATHER;
 971
 972	/* FIELD_PREP() implicitly truncates to chunk */
 973	dma_len -= 1;
 974	dlen_type = FIELD_PREP(NFDK_DESC_TX_DMA_LEN_HEAD,
 975			       dma_len > NFDK_DESC_TX_DMA_LEN_HEAD ?
 976			       NFDK_DESC_TX_DMA_LEN_HEAD : dma_len) |
 977		    FIELD_PREP(NFDK_DESC_TX_TYPE_HEAD, type);
 978
 979	txd->dma_len_type = cpu_to_le16(dlen_type);
 980	nfp_desc_set_dma_addr_48b(txd, dma_addr);
 981
 982	tmp_dlen = dlen_type & NFDK_DESC_TX_DMA_LEN_HEAD;
 983	dma_len -= tmp_dlen;
 984	dma_addr += tmp_dlen + 1;
 985	txd++;
 986
 987	while (dma_len > 0) {
 988		dma_len -= 1;
 989		dlen_type = FIELD_PREP(NFDK_DESC_TX_DMA_LEN, dma_len);
 990		txd->dma_len_type = cpu_to_le16(dlen_type);
 991		nfp_desc_set_dma_addr_48b(txd, dma_addr);
 992
 993		dlen_type &= NFDK_DESC_TX_DMA_LEN;
 994		dma_len -= dlen_type;
 995		dma_addr += dlen_type + 1;
 996		txd++;
 997	}
 998
 999	(txd - 1)->dma_len_type = cpu_to_le16(dlen_type | NFDK_DESC_TX_EOP);
1000
1001	/* Metadata desc */
1002	txd->raw = 0;
1003	txd++;
1004
1005	cnt = txd - tx_ring->ktxds - wr_idx;
1006	tx_ring->wr_p += cnt;
1007	if (tx_ring->wr_p % NFDK_TX_DESC_BLOCK_CNT)
1008		tx_ring->data_pending += pkt_len;
1009	else
1010		tx_ring->data_pending = 0;
1011
1012	tx_ring->wr_ptr_add += cnt;
1013	return true;
1014}
1015
1016/**
1017 * nfp_nfdk_rx() - receive up to @budget packets on @rx_ring
1018 * @rx_ring:   RX ring to receive from
1019 * @budget:    NAPI budget
1020 *
1021 * Note, this function is separated out from the napi poll function to
1022 * more cleanly separate packet receive code from other bookkeeping
1023 * functions performed in the napi poll function.
1024 *
1025 * Return: Number of packets received.
1026 */
1027static int nfp_nfdk_rx(struct nfp_net_rx_ring *rx_ring, int budget)
1028{
1029	struct nfp_net_r_vector *r_vec = rx_ring->r_vec;
1030	struct nfp_net_dp *dp = &r_vec->nfp_net->dp;
1031	struct nfp_net_tx_ring *tx_ring;
1032	struct bpf_prog *xdp_prog;
1033	bool xdp_tx_cmpl = false;
1034	unsigned int true_bufsz;
1035	struct sk_buff *skb;
1036	int pkts_polled = 0;
1037	struct xdp_buff xdp;
1038	int idx;
1039
1040	xdp_prog = READ_ONCE(dp->xdp_prog);
1041	true_bufsz = xdp_prog ? PAGE_SIZE : dp->fl_bufsz;
1042	xdp_init_buff(&xdp, PAGE_SIZE - NFP_NET_RX_BUF_HEADROOM,
1043		      &rx_ring->xdp_rxq);
1044	tx_ring = r_vec->xdp_ring;
1045
1046	while (pkts_polled < budget) {
1047		unsigned int meta_len, data_len, meta_off, pkt_len, pkt_off;
1048		struct nfp_net_rx_buf *rxbuf;
1049		struct nfp_net_rx_desc *rxd;
1050		struct nfp_meta_parsed meta;
1051		bool redir_egress = false;
1052		struct net_device *netdev;
1053		dma_addr_t new_dma_addr;
1054		u32 meta_len_xdp = 0;
1055		void *new_frag;
1056
1057		idx = D_IDX(rx_ring, rx_ring->rd_p);
1058
1059		rxd = &rx_ring->rxds[idx];
1060		if (!(rxd->rxd.meta_len_dd & PCIE_DESC_RX_DD))
1061			break;
1062
1063		/* Memory barrier to ensure that we won't do other reads
1064		 * before the DD bit.
1065		 */
1066		dma_rmb();
1067
1068		memset(&meta, 0, sizeof(meta));
1069
1070		rx_ring->rd_p++;
1071		pkts_polled++;
1072
1073		rxbuf =	&rx_ring->rxbufs[idx];
1074		/*         < meta_len >
1075		 *  <-- [rx_offset] -->
1076		 *  ---------------------------------------------------------
1077		 * | [XX] |  metadata  |             packet           | XXXX |
1078		 *  ---------------------------------------------------------
1079		 *         <---------------- data_len --------------->
1080		 *
1081		 * The rx_offset is fixed for all packets, the meta_len can vary
1082		 * on a packet by packet basis. If rx_offset is set to zero
1083		 * (_RX_OFFSET_DYNAMIC) metadata starts at the beginning of the
1084		 * buffer and is immediately followed by the packet (no [XX]).
1085		 */
1086		meta_len = rxd->rxd.meta_len_dd & PCIE_DESC_RX_META_LEN_MASK;
1087		data_len = le16_to_cpu(rxd->rxd.data_len);
1088		pkt_len = data_len - meta_len;
1089
1090		pkt_off = NFP_NET_RX_BUF_HEADROOM + dp->rx_dma_off;
1091		if (dp->rx_offset == NFP_NET_CFG_RX_OFFSET_DYNAMIC)
1092			pkt_off += meta_len;
1093		else
1094			pkt_off += dp->rx_offset;
1095		meta_off = pkt_off - meta_len;
1096
1097		/* Stats update */
1098		u64_stats_update_begin(&r_vec->rx_sync);
1099		r_vec->rx_pkts++;
1100		r_vec->rx_bytes += pkt_len;
1101		u64_stats_update_end(&r_vec->rx_sync);
1102
1103		if (unlikely(meta_len > NFP_NET_MAX_PREPEND ||
1104			     (dp->rx_offset && meta_len > dp->rx_offset))) {
1105			nn_dp_warn(dp, "oversized RX packet metadata %u\n",
1106				   meta_len);
1107			nfp_nfdk_rx_drop(dp, r_vec, rx_ring, rxbuf, NULL);
1108			continue;
1109		}
1110
1111		nfp_net_dma_sync_cpu_rx(dp, rxbuf->dma_addr + meta_off,
1112					data_len);
1113
1114		if (meta_len) {
1115			if (unlikely(nfp_nfdk_parse_meta(dp->netdev, &meta,
1116							 rxbuf->frag + meta_off,
1117							 rxbuf->frag + pkt_off,
1118							 pkt_len, meta_len))) {
1119				nn_dp_warn(dp, "invalid RX packet metadata\n");
1120				nfp_nfdk_rx_drop(dp, r_vec, rx_ring, rxbuf,
1121						 NULL);
1122				continue;
1123			}
1124		}
1125
1126		if (xdp_prog && !meta.portid) {
1127			void *orig_data = rxbuf->frag + pkt_off;
1128			unsigned int dma_off;
1129			int act;
1130
1131			xdp_prepare_buff(&xdp,
1132					 rxbuf->frag + NFP_NET_RX_BUF_HEADROOM,
1133					 pkt_off - NFP_NET_RX_BUF_HEADROOM,
1134					 pkt_len, true);
1135
1136			act = bpf_prog_run_xdp(xdp_prog, &xdp);
1137
1138			pkt_len = xdp.data_end - xdp.data;
1139			pkt_off += xdp.data - orig_data;
1140
1141			switch (act) {
1142			case XDP_PASS:
1143				meta_len_xdp = xdp.data - xdp.data_meta;
1144				break;
1145			case XDP_TX:
1146				dma_off = pkt_off - NFP_NET_RX_BUF_HEADROOM;
1147				if (unlikely(!nfp_nfdk_tx_xdp_buf(dp, rx_ring,
1148								  tx_ring,
1149								  rxbuf,
1150								  dma_off,
1151								  pkt_len,
1152								  &xdp_tx_cmpl)))
1153					trace_xdp_exception(dp->netdev,
1154							    xdp_prog, act);
1155				continue;
1156			default:
1157				bpf_warn_invalid_xdp_action(dp->netdev, xdp_prog, act);
1158				fallthrough;
1159			case XDP_ABORTED:
1160				trace_xdp_exception(dp->netdev, xdp_prog, act);
1161				fallthrough;
1162			case XDP_DROP:
1163				nfp_nfdk_rx_give_one(dp, rx_ring, rxbuf->frag,
1164						     rxbuf->dma_addr);
1165				continue;
1166			}
1167		}
1168
1169		if (likely(!meta.portid)) {
1170			netdev = dp->netdev;
1171		} else if (meta.portid == NFP_META_PORT_ID_CTRL) {
1172			struct nfp_net *nn = netdev_priv(dp->netdev);
1173
1174			nfp_app_ctrl_rx_raw(nn->app, rxbuf->frag + pkt_off,
1175					    pkt_len);
1176			nfp_nfdk_rx_give_one(dp, rx_ring, rxbuf->frag,
1177					     rxbuf->dma_addr);
1178			continue;
1179		} else {
1180			struct nfp_net *nn;
1181
1182			nn = netdev_priv(dp->netdev);
1183			netdev = nfp_app_dev_get(nn->app, meta.portid,
1184						 &redir_egress);
1185			if (unlikely(!netdev)) {
1186				nfp_nfdk_rx_drop(dp, r_vec, rx_ring, rxbuf,
1187						 NULL);
1188				continue;
1189			}
1190
1191			if (nfp_netdev_is_nfp_repr(netdev))
1192				nfp_repr_inc_rx_stats(netdev, pkt_len);
1193		}
1194
1195		skb = napi_build_skb(rxbuf->frag, true_bufsz);
1196		if (unlikely(!skb)) {
1197			nfp_nfdk_rx_drop(dp, r_vec, rx_ring, rxbuf, NULL);
1198			continue;
1199		}
1200		new_frag = nfp_nfdk_napi_alloc_one(dp, &new_dma_addr);
1201		if (unlikely(!new_frag)) {
1202			nfp_nfdk_rx_drop(dp, r_vec, rx_ring, rxbuf, skb);
1203			continue;
1204		}
1205
1206		nfp_net_dma_unmap_rx(dp, rxbuf->dma_addr);
1207
1208		nfp_nfdk_rx_give_one(dp, rx_ring, new_frag, new_dma_addr);
1209
1210		skb_reserve(skb, pkt_off);
1211		skb_put(skb, pkt_len);
1212
1213		skb->mark = meta.mark;
1214		skb_set_hash(skb, meta.hash, meta.hash_type);
1215
1216		skb_record_rx_queue(skb, rx_ring->idx);
1217		skb->protocol = eth_type_trans(skb, netdev);
1218
1219		nfp_nfdk_rx_csum(dp, r_vec, rxd, &meta, skb);
1220
1221		if (unlikely(!nfp_net_vlan_strip(skb, rxd, &meta))) {
1222			nfp_nfdk_rx_drop(dp, r_vec, rx_ring, NULL, skb);
1223			continue;
1224		}
1225
1226#ifdef CONFIG_NFP_NET_IPSEC
1227		if (meta.ipsec_saidx != 0 && unlikely(nfp_net_ipsec_rx(&meta, skb))) {
1228			nfp_nfdk_rx_drop(dp, r_vec, rx_ring, NULL, skb);
1229			continue;
1230		}
1231#endif
1232
1233		if (meta_len_xdp)
1234			skb_metadata_set(skb, meta_len_xdp);
1235
1236		if (likely(!redir_egress)) {
1237			napi_gro_receive(&rx_ring->r_vec->napi, skb);
1238		} else {
1239			skb->dev = netdev;
1240			skb_reset_network_header(skb);
1241			__skb_push(skb, ETH_HLEN);
1242			dev_queue_xmit(skb);
1243		}
1244	}
1245
1246	if (xdp_prog) {
1247		if (tx_ring->wr_ptr_add)
1248			nfp_net_tx_xmit_more_flush(tx_ring);
1249		else if (unlikely(tx_ring->wr_p != tx_ring->rd_p) &&
1250			 !xdp_tx_cmpl)
1251			if (!nfp_nfdk_xdp_complete(tx_ring))
1252				pkts_polled = budget;
1253	}
1254
1255	return pkts_polled;
1256}
1257
1258/**
1259 * nfp_nfdk_poll() - napi poll function
1260 * @napi:    NAPI structure
1261 * @budget:  NAPI budget
1262 *
1263 * Return: number of packets polled.
1264 */
1265int nfp_nfdk_poll(struct napi_struct *napi, int budget)
1266{
1267	struct nfp_net_r_vector *r_vec =
1268		container_of(napi, struct nfp_net_r_vector, napi);
1269	unsigned int pkts_polled = 0;
1270
1271	if (r_vec->tx_ring)
1272		nfp_nfdk_tx_complete(r_vec->tx_ring, budget);
1273	if (r_vec->rx_ring)
1274		pkts_polled = nfp_nfdk_rx(r_vec->rx_ring, budget);
1275
1276	if (pkts_polled < budget)
1277		if (napi_complete_done(napi, pkts_polled))
1278			nfp_net_irq_unmask(r_vec->nfp_net, r_vec->irq_entry);
1279
1280	if (r_vec->nfp_net->rx_coalesce_adapt_on && r_vec->rx_ring) {
1281		struct dim_sample dim_sample = {};
1282		u64 pkts, bytes;
1283
1284		pkts = r_vec->rx_pkts;
1285		bytes = r_vec->rx_bytes;
1286
1287		dim_update_sample(r_vec->event_ctr, pkts, bytes, &dim_sample);
1288		net_dim(&r_vec->rx_dim, &dim_sample);
1289	}
1290
1291	if (r_vec->nfp_net->tx_coalesce_adapt_on && r_vec->tx_ring) {
1292		struct dim_sample dim_sample = {};
1293		u64 pkts, bytes;
1294
1295		pkts = r_vec->tx_pkts;
1296		bytes = r_vec->tx_bytes;
1297
1298		dim_update_sample(r_vec->event_ctr, pkts, bytes, &dim_sample);
1299		net_dim(&r_vec->tx_dim, &dim_sample);
1300	}
1301
1302	return pkts_polled;
1303}
1304
1305/* Control device data path
1306 */
1307
1308bool
1309nfp_nfdk_ctrl_tx_one(struct nfp_net *nn, struct nfp_net_r_vector *r_vec,
1310		     struct sk_buff *skb, bool old)
1311{
1312	u32 cnt, tmp_dlen, dlen_type = 0;
1313	struct nfp_net_tx_ring *tx_ring;
1314	struct nfp_nfdk_tx_buf *txbuf;
1315	struct nfp_nfdk_tx_desc *txd;
1316	unsigned int dma_len, type;
1317	struct nfp_net_dp *dp;
1318	dma_addr_t dma_addr;
1319	u64 metadata = 0;
1320	int wr_idx;
1321
1322	dp = &r_vec->nfp_net->dp;
1323	tx_ring = r_vec->tx_ring;
1324
1325	if (WARN_ON_ONCE(skb_shinfo(skb)->nr_frags)) {
1326		nn_dp_warn(dp, "Driver's CTRL TX does not implement gather\n");
1327		goto err_free;
1328	}
1329
1330	/* Don't bother counting frags, assume the worst */
1331	if (unlikely(nfp_net_tx_full(tx_ring, NFDK_TX_DESC_STOP_CNT))) {
1332		u64_stats_update_begin(&r_vec->tx_sync);
1333		r_vec->tx_busy++;
1334		u64_stats_update_end(&r_vec->tx_sync);
1335		if (!old)
1336			__skb_queue_tail(&r_vec->queue, skb);
1337		else
1338			__skb_queue_head(&r_vec->queue, skb);
1339		return NETDEV_TX_BUSY;
1340	}
1341
1342	if (nfp_app_ctrl_has_meta(nn->app)) {
1343		if (unlikely(skb_headroom(skb) < 8)) {
1344			nn_dp_warn(dp, "CTRL TX on skb without headroom\n");
1345			goto err_free;
1346		}
1347		metadata = NFDK_DESC_TX_CHAIN_META;
1348		put_unaligned_be32(NFP_META_PORT_ID_CTRL, skb_push(skb, 4));
1349		put_unaligned_be32(FIELD_PREP(NFDK_META_LEN, 8) |
1350				   FIELD_PREP(NFDK_META_FIELDS,
1351					      NFP_NET_META_PORTID),
1352				   skb_push(skb, 4));
1353	}
1354
1355	if (nfp_nfdk_tx_maybe_close_block(tx_ring, skb))
1356		goto err_free;
1357
1358	/* DMA map all */
1359	wr_idx = D_IDX(tx_ring, tx_ring->wr_p);
1360	txd = &tx_ring->ktxds[wr_idx];
1361	txbuf = &tx_ring->ktxbufs[wr_idx];
1362
1363	dma_len = skb_headlen(skb);
1364	if (dma_len <= NFDK_TX_MAX_DATA_PER_HEAD)
1365		type = NFDK_DESC_TX_TYPE_SIMPLE;
1366	else
1367		type = NFDK_DESC_TX_TYPE_GATHER;
1368
1369	dma_addr = dma_map_single(dp->dev, skb->data, dma_len, DMA_TO_DEVICE);
1370	if (dma_mapping_error(dp->dev, dma_addr))
1371		goto err_warn_dma;
1372
1373	txbuf->skb = skb;
1374	txbuf++;
1375
1376	txbuf->dma_addr = dma_addr;
1377	txbuf++;
1378
1379	dma_len -= 1;
1380	dlen_type = FIELD_PREP(NFDK_DESC_TX_DMA_LEN_HEAD,
1381			       dma_len > NFDK_DESC_TX_DMA_LEN_HEAD ?
1382			       NFDK_DESC_TX_DMA_LEN_HEAD : dma_len) |
1383		    FIELD_PREP(NFDK_DESC_TX_TYPE_HEAD, type);
1384
1385	txd->dma_len_type = cpu_to_le16(dlen_type);
1386	nfp_desc_set_dma_addr_48b(txd, dma_addr);
1387
1388	tmp_dlen = dlen_type & NFDK_DESC_TX_DMA_LEN_HEAD;
1389	dma_len -= tmp_dlen;
1390	dma_addr += tmp_dlen + 1;
1391	txd++;
1392
1393	while (dma_len > 0) {
1394		dma_len -= 1;
1395		dlen_type = FIELD_PREP(NFDK_DESC_TX_DMA_LEN, dma_len);
1396		txd->dma_len_type = cpu_to_le16(dlen_type);
1397		nfp_desc_set_dma_addr_48b(txd, dma_addr);
1398
1399		dlen_type &= NFDK_DESC_TX_DMA_LEN;
1400		dma_len -= dlen_type;
1401		dma_addr += dlen_type + 1;
1402		txd++;
1403	}
1404
1405	(txd - 1)->dma_len_type = cpu_to_le16(dlen_type | NFDK_DESC_TX_EOP);
1406
1407	/* Metadata desc */
1408	txd->raw = cpu_to_le64(metadata);
1409	txd++;
1410
1411	cnt = txd - tx_ring->ktxds - wr_idx;
1412	if (unlikely(round_down(wr_idx, NFDK_TX_DESC_BLOCK_CNT) !=
1413		     round_down(wr_idx + cnt - 1, NFDK_TX_DESC_BLOCK_CNT)))
1414		goto err_warn_overflow;
1415
1416	tx_ring->wr_p += cnt;
1417	if (tx_ring->wr_p % NFDK_TX_DESC_BLOCK_CNT)
1418		tx_ring->data_pending += skb->len;
1419	else
1420		tx_ring->data_pending = 0;
1421
1422	tx_ring->wr_ptr_add += cnt;
1423	nfp_net_tx_xmit_more_flush(tx_ring);
1424
1425	return NETDEV_TX_OK;
1426
1427err_warn_overflow:
1428	WARN_ONCE(1, "unable to fit packet into a descriptor wr_idx:%d head:%d frags:%d cnt:%d",
1429		  wr_idx, skb_headlen(skb), 0, cnt);
1430	txbuf--;
1431	dma_unmap_single(dp->dev, txbuf->dma_addr,
1432			 skb_headlen(skb), DMA_TO_DEVICE);
1433	txbuf->raw = 0;
1434err_warn_dma:
1435	nn_dp_warn(dp, "Failed to map DMA TX buffer\n");
1436err_free:
1437	u64_stats_update_begin(&r_vec->tx_sync);
1438	r_vec->tx_errors++;
1439	u64_stats_update_end(&r_vec->tx_sync);
1440	dev_kfree_skb_any(skb);
1441	return NETDEV_TX_OK;
1442}
1443
1444static void __nfp_ctrl_tx_queued(struct nfp_net_r_vector *r_vec)
1445{
1446	struct sk_buff *skb;
1447
1448	while ((skb = __skb_dequeue(&r_vec->queue)))
1449		if (nfp_nfdk_ctrl_tx_one(r_vec->nfp_net, r_vec, skb, true))
1450			return;
1451}
1452
1453static bool
1454nfp_ctrl_meta_ok(struct nfp_net *nn, void *data, unsigned int meta_len)
1455{
1456	u32 meta_type, meta_tag;
1457
1458	if (!nfp_app_ctrl_has_meta(nn->app))
1459		return !meta_len;
1460
1461	if (meta_len != 8)
1462		return false;
1463
1464	meta_type = get_unaligned_be32(data);
1465	meta_tag = get_unaligned_be32(data + 4);
1466
1467	return (meta_type == NFP_NET_META_PORTID &&
1468		meta_tag == NFP_META_PORT_ID_CTRL);
1469}
1470
1471static bool
1472nfp_ctrl_rx_one(struct nfp_net *nn, struct nfp_net_dp *dp,
1473		struct nfp_net_r_vector *r_vec, struct nfp_net_rx_ring *rx_ring)
1474{
1475	unsigned int meta_len, data_len, meta_off, pkt_len, pkt_off;
1476	struct nfp_net_rx_buf *rxbuf;
1477	struct nfp_net_rx_desc *rxd;
1478	dma_addr_t new_dma_addr;
1479	struct sk_buff *skb;
1480	void *new_frag;
1481	int idx;
1482
1483	idx = D_IDX(rx_ring, rx_ring->rd_p);
1484
1485	rxd = &rx_ring->rxds[idx];
1486	if (!(rxd->rxd.meta_len_dd & PCIE_DESC_RX_DD))
1487		return false;
1488
1489	/* Memory barrier to ensure that we won't do other reads
1490	 * before the DD bit.
1491	 */
1492	dma_rmb();
1493
1494	rx_ring->rd_p++;
1495
1496	rxbuf =	&rx_ring->rxbufs[idx];
1497	meta_len = rxd->rxd.meta_len_dd & PCIE_DESC_RX_META_LEN_MASK;
1498	data_len = le16_to_cpu(rxd->rxd.data_len);
1499	pkt_len = data_len - meta_len;
1500
1501	pkt_off = NFP_NET_RX_BUF_HEADROOM + dp->rx_dma_off;
1502	if (dp->rx_offset == NFP_NET_CFG_RX_OFFSET_DYNAMIC)
1503		pkt_off += meta_len;
1504	else
1505		pkt_off += dp->rx_offset;
1506	meta_off = pkt_off - meta_len;
1507
1508	/* Stats update */
1509	u64_stats_update_begin(&r_vec->rx_sync);
1510	r_vec->rx_pkts++;
1511	r_vec->rx_bytes += pkt_len;
1512	u64_stats_update_end(&r_vec->rx_sync);
1513
1514	nfp_net_dma_sync_cpu_rx(dp, rxbuf->dma_addr + meta_off,	data_len);
1515
1516	if (unlikely(!nfp_ctrl_meta_ok(nn, rxbuf->frag + meta_off, meta_len))) {
1517		nn_dp_warn(dp, "incorrect metadata for ctrl packet (%d)\n",
1518			   meta_len);
1519		nfp_nfdk_rx_drop(dp, r_vec, rx_ring, rxbuf, NULL);
1520		return true;
1521	}
1522
1523	skb = build_skb(rxbuf->frag, dp->fl_bufsz);
1524	if (unlikely(!skb)) {
1525		nfp_nfdk_rx_drop(dp, r_vec, rx_ring, rxbuf, NULL);
1526		return true;
1527	}
1528	new_frag = nfp_nfdk_napi_alloc_one(dp, &new_dma_addr);
1529	if (unlikely(!new_frag)) {
1530		nfp_nfdk_rx_drop(dp, r_vec, rx_ring, rxbuf, skb);
1531		return true;
1532	}
1533
1534	nfp_net_dma_unmap_rx(dp, rxbuf->dma_addr);
1535
1536	nfp_nfdk_rx_give_one(dp, rx_ring, new_frag, new_dma_addr);
1537
1538	skb_reserve(skb, pkt_off);
1539	skb_put(skb, pkt_len);
1540
1541	nfp_app_ctrl_rx(nn->app, skb);
1542
1543	return true;
1544}
1545
1546static bool nfp_ctrl_rx(struct nfp_net_r_vector *r_vec)
1547{
1548	struct nfp_net_rx_ring *rx_ring = r_vec->rx_ring;
1549	struct nfp_net *nn = r_vec->nfp_net;
1550	struct nfp_net_dp *dp = &nn->dp;
1551	unsigned int budget = 512;
1552
1553	while (nfp_ctrl_rx_one(nn, dp, r_vec, rx_ring) && budget--)
1554		continue;
1555
1556	return budget;
1557}
1558
1559void nfp_nfdk_ctrl_poll(struct tasklet_struct *t)
1560{
1561	struct nfp_net_r_vector *r_vec = from_tasklet(r_vec, t, tasklet);
1562
1563	spin_lock(&r_vec->lock);
1564	nfp_nfdk_tx_complete(r_vec->tx_ring, 0);
1565	__nfp_ctrl_tx_queued(r_vec);
1566	spin_unlock(&r_vec->lock);
1567
1568	if (nfp_ctrl_rx(r_vec)) {
1569		nfp_net_irq_unmask(r_vec->nfp_net, r_vec->irq_entry);
1570	} else {
1571		tasklet_schedule(&r_vec->tasklet);
1572		nn_dp_warn(&r_vec->nfp_net->dp,
1573			   "control message budget exceeded!\n");
1574	}
1575}