ice_txrx.c - drivers/net/ethernet/intel/ice/ice_txrx.c - Linux diff v6.13.7

   1// SPDX-License-Identifier: GPL-2.0
   2/* Copyright (c) 2018, Intel Corporation. */
   3
   4/* The driver transmit and receive code */
   5
   6#include <linux/mm.h>
   7#include <linux/netdevice.h>
   8#include <linux/prefetch.h>
   9#include <linux/bpf_trace.h>
  10#include <net/dsfield.h>
  11#include <net/mpls.h>
  12#include <net/xdp.h>
  13#include "ice_txrx_lib.h"
  14#include "ice_lib.h"
  15#include "ice.h"
  16#include "ice_trace.h"
  17#include "ice_dcb_lib.h"
  18#include "ice_xsk.h"
  19#include "ice_eswitch.h"
  20
  21#define ICE_RX_HDR_SIZE		256
  22
  23#define FDIR_DESC_RXDID 0x40
  24#define ICE_FDIR_CLEAN_DELAY 10
  25
  26/**
  27 * ice_prgm_fdir_fltr - Program a Flow Director filter
  28 * @vsi: VSI to send dummy packet
  29 * @fdir_desc: flow director descriptor
  30 * @raw_packet: allocated buffer for flow director
  31 */
  32int
  33ice_prgm_fdir_fltr(struct ice_vsi *vsi, struct ice_fltr_desc *fdir_desc,
  34		   u8 *raw_packet)
  35{
  36	struct ice_tx_buf *tx_buf, *first;
  37	struct ice_fltr_desc *f_desc;
  38	struct ice_tx_desc *tx_desc;
  39	struct ice_tx_ring *tx_ring;
  40	struct device *dev;
  41	dma_addr_t dma;
  42	u32 td_cmd;
  43	u16 i;
  44
  45	/* VSI and Tx ring */
  46	if (!vsi)
  47		return -ENOENT;
  48	tx_ring = vsi->tx_rings[0];
  49	if (!tx_ring || !tx_ring->desc)
  50		return -ENOENT;
  51	dev = tx_ring->dev;
  52
  53	/* we are using two descriptors to add/del a filter and we can wait */
  54	for (i = ICE_FDIR_CLEAN_DELAY; ICE_DESC_UNUSED(tx_ring) < 2; i--) {
  55		if (!i)
  56			return -EAGAIN;
  57		msleep_interruptible(1);
  58	}
  59
  60	dma = dma_map_single(dev, raw_packet, ICE_FDIR_MAX_RAW_PKT_SIZE,
  61			     DMA_TO_DEVICE);
  62
  63	if (dma_mapping_error(dev, dma))
  64		return -EINVAL;
  65
  66	/* grab the next descriptor */
  67	i = tx_ring->next_to_use;
  68	first = &tx_ring->tx_buf[i];
  69	f_desc = ICE_TX_FDIRDESC(tx_ring, i);
  70	memcpy(f_desc, fdir_desc, sizeof(*f_desc));
  71
  72	i++;
  73	i = (i < tx_ring->count) ? i : 0;
  74	tx_desc = ICE_TX_DESC(tx_ring, i);
  75	tx_buf = &tx_ring->tx_buf[i];
  76
  77	i++;
  78	tx_ring->next_to_use = (i < tx_ring->count) ? i : 0;
  79
  80	memset(tx_buf, 0, sizeof(*tx_buf));
  81	dma_unmap_len_set(tx_buf, len, ICE_FDIR_MAX_RAW_PKT_SIZE);
  82	dma_unmap_addr_set(tx_buf, dma, dma);
  83
  84	tx_desc->buf_addr = cpu_to_le64(dma);
  85	td_cmd = ICE_TXD_LAST_DESC_CMD | ICE_TX_DESC_CMD_DUMMY |
  86		 ICE_TX_DESC_CMD_RE;
  87
  88	tx_buf->type = ICE_TX_BUF_DUMMY;
  89	tx_buf->raw_buf = raw_packet;
  90
  91	tx_desc->cmd_type_offset_bsz =
  92		ice_build_ctob(td_cmd, 0, ICE_FDIR_MAX_RAW_PKT_SIZE, 0);
  93
  94	/* Force memory write to complete before letting h/w know
  95	 * there are new descriptors to fetch.
  96	 */
  97	wmb();
  98
  99	/* mark the data descriptor to be watched */
 100	first->next_to_watch = tx_desc;
 101
 102	writel(tx_ring->next_to_use, tx_ring->tail);
 103
 104	return 0;
 105}
 106
 107/**
 108 * ice_unmap_and_free_tx_buf - Release a Tx buffer
 109 * @ring: the ring that owns the buffer
 110 * @tx_buf: the buffer to free
 111 */
 112static void
 113ice_unmap_and_free_tx_buf(struct ice_tx_ring *ring, struct ice_tx_buf *tx_buf)
 114{
 115	if (dma_unmap_len(tx_buf, len))
 
 
 
 
 
 
 
 
 
 
 
 
 116		dma_unmap_page(ring->dev,
 117			       dma_unmap_addr(tx_buf, dma),
 118			       dma_unmap_len(tx_buf, len),
 119			       DMA_TO_DEVICE);
 120
 121	switch (tx_buf->type) {
 122	case ICE_TX_BUF_DUMMY:
 123		devm_kfree(ring->dev, tx_buf->raw_buf);
 124		break;
 125	case ICE_TX_BUF_SKB:
 126		dev_kfree_skb_any(tx_buf->skb);
 127		break;
 128	case ICE_TX_BUF_XDP_TX:
 129		page_frag_free(tx_buf->raw_buf);
 130		break;
 131	case ICE_TX_BUF_XDP_XMIT:
 132		xdp_return_frame(tx_buf->xdpf);
 133		break;
 134	}
 135
 136	tx_buf->next_to_watch = NULL;
 137	tx_buf->type = ICE_TX_BUF_EMPTY;
 138	dma_unmap_len_set(tx_buf, len, 0);
 139	/* tx_buf must be completely set up in the transmit path */
 140}
 141
 142static struct netdev_queue *txring_txq(const struct ice_tx_ring *ring)
 143{
 144	return netdev_get_tx_queue(ring->netdev, ring->q_index);
 145}
 146
 147/**
 148 * ice_clean_tx_ring - Free any empty Tx buffers
 149 * @tx_ring: ring to be cleaned
 150 */
 151void ice_clean_tx_ring(struct ice_tx_ring *tx_ring)
 152{
 153	u32 size;
 154	u16 i;
 155
 156	if (ice_ring_is_xdp(tx_ring) && tx_ring->xsk_pool) {
 157		ice_xsk_clean_xdp_ring(tx_ring);
 158		goto tx_skip_free;
 159	}
 160
 161	/* ring already cleared, nothing to do */
 162	if (!tx_ring->tx_buf)
 163		return;
 164
 165	/* Free all the Tx ring sk_buffs */
 166	for (i = 0; i < tx_ring->count; i++)
 167		ice_unmap_and_free_tx_buf(tx_ring, &tx_ring->tx_buf[i]);
 168
 169tx_skip_free:
 170	memset(tx_ring->tx_buf, 0, sizeof(*tx_ring->tx_buf) * tx_ring->count);
 171
 172	size = ALIGN(tx_ring->count * sizeof(struct ice_tx_desc),
 173		     PAGE_SIZE);
 174	/* Zero out the descriptor ring */
 175	memset(tx_ring->desc, 0, size);
 176
 177	tx_ring->next_to_use = 0;
 178	tx_ring->next_to_clean = 0;
 
 
 179
 180	if (!tx_ring->netdev)
 181		return;
 182
 183	/* cleanup Tx queue statistics */
 184	netdev_tx_reset_queue(txring_txq(tx_ring));
 185}
 186
 187/**
 188 * ice_free_tx_ring - Free Tx resources per queue
 189 * @tx_ring: Tx descriptor ring for a specific queue
 190 *
 191 * Free all transmit software resources
 192 */
 193void ice_free_tx_ring(struct ice_tx_ring *tx_ring)
 194{
 195	u32 size;
 196
 197	ice_clean_tx_ring(tx_ring);
 198	devm_kfree(tx_ring->dev, tx_ring->tx_buf);
 199	tx_ring->tx_buf = NULL;
 200
 201	if (tx_ring->desc) {
 202		size = ALIGN(tx_ring->count * sizeof(struct ice_tx_desc),
 203			     PAGE_SIZE);
 204		dmam_free_coherent(tx_ring->dev, size,
 205				   tx_ring->desc, tx_ring->dma);
 206		tx_ring->desc = NULL;
 207	}
 208}
 209
 210/**
 211 * ice_clean_tx_irq - Reclaim resources after transmit completes
 212 * @tx_ring: Tx ring to clean
 213 * @napi_budget: Used to determine if we are in netpoll
 214 *
 215 * Returns true if there's any budget left (e.g. the clean is finished)
 216 */
 217static bool ice_clean_tx_irq(struct ice_tx_ring *tx_ring, int napi_budget)
 218{
 219	unsigned int total_bytes = 0, total_pkts = 0;
 220	unsigned int budget = ICE_DFLT_IRQ_WORK;
 221	struct ice_vsi *vsi = tx_ring->vsi;
 222	s16 i = tx_ring->next_to_clean;
 223	struct ice_tx_desc *tx_desc;
 224	struct ice_tx_buf *tx_buf;
 225
 226	/* get the bql data ready */
 227	netdev_txq_bql_complete_prefetchw(txring_txq(tx_ring));
 228
 229	tx_buf = &tx_ring->tx_buf[i];
 230	tx_desc = ICE_TX_DESC(tx_ring, i);
 231	i -= tx_ring->count;
 232
 233	prefetch(&vsi->state);
 234
 235	do {
 236		struct ice_tx_desc *eop_desc = tx_buf->next_to_watch;
 237
 238		/* if next_to_watch is not set then there is no work pending */
 239		if (!eop_desc)
 240			break;
 241
 242		/* follow the guidelines of other drivers */
 243		prefetchw(&tx_buf->skb->users);
 244
 245		smp_rmb();	/* prevent any other reads prior to eop_desc */
 246
 247		ice_trace(clean_tx_irq, tx_ring, tx_desc, tx_buf);
 248		/* if the descriptor isn't done, no work yet to do */
 249		if (!(eop_desc->cmd_type_offset_bsz &
 250		      cpu_to_le64(ICE_TX_DESC_DTYPE_DESC_DONE)))
 251			break;
 252
 253		/* clear next_to_watch to prevent false hangs */
 254		tx_buf->next_to_watch = NULL;
 255
 256		/* update the statistics for this packet */
 257		total_bytes += tx_buf->bytecount;
 258		total_pkts += tx_buf->gso_segs;
 259
 260		/* free the skb */
 261		napi_consume_skb(tx_buf->skb, napi_budget);
 262
 263		/* unmap skb header data */
 264		dma_unmap_single(tx_ring->dev,
 265				 dma_unmap_addr(tx_buf, dma),
 266				 dma_unmap_len(tx_buf, len),
 267				 DMA_TO_DEVICE);
 268
 269		/* clear tx_buf data */
 270		tx_buf->type = ICE_TX_BUF_EMPTY;
 271		dma_unmap_len_set(tx_buf, len, 0);
 272
 273		/* unmap remaining buffers */
 274		while (tx_desc != eop_desc) {
 275			ice_trace(clean_tx_irq_unmap, tx_ring, tx_desc, tx_buf);
 276			tx_buf++;
 277			tx_desc++;
 278			i++;
 279			if (unlikely(!i)) {
 280				i -= tx_ring->count;
 281				tx_buf = tx_ring->tx_buf;
 282				tx_desc = ICE_TX_DESC(tx_ring, 0);
 283			}
 284
 285			/* unmap any remaining paged data */
 286			if (dma_unmap_len(tx_buf, len)) {
 287				dma_unmap_page(tx_ring->dev,
 288					       dma_unmap_addr(tx_buf, dma),
 289					       dma_unmap_len(tx_buf, len),
 290					       DMA_TO_DEVICE);
 291				dma_unmap_len_set(tx_buf, len, 0);
 292			}
 293		}
 294		ice_trace(clean_tx_irq_unmap_eop, tx_ring, tx_desc, tx_buf);
 295
 296		/* move us one more past the eop_desc for start of next pkt */
 297		tx_buf++;
 298		tx_desc++;
 299		i++;
 300		if (unlikely(!i)) {
 301			i -= tx_ring->count;
 302			tx_buf = tx_ring->tx_buf;
 303			tx_desc = ICE_TX_DESC(tx_ring, 0);
 304		}
 305
 306		prefetch(tx_desc);
 307
 308		/* update budget accounting */
 309		budget--;
 310	} while (likely(budget));
 311
 312	i += tx_ring->count;
 313	tx_ring->next_to_clean = i;
 314
 315	ice_update_tx_ring_stats(tx_ring, total_pkts, total_bytes);
 316	netdev_tx_completed_queue(txring_txq(tx_ring), total_pkts, total_bytes);
 317
 318#define TX_WAKE_THRESHOLD ((s16)(DESC_NEEDED * 2))
 319	if (unlikely(total_pkts && netif_carrier_ok(tx_ring->netdev) &&
 320		     (ICE_DESC_UNUSED(tx_ring) >= TX_WAKE_THRESHOLD))) {
 321		/* Make sure that anybody stopping the queue after this
 322		 * sees the new next_to_clean.
 323		 */
 324		smp_mb();
 325		if (netif_tx_queue_stopped(txring_txq(tx_ring)) &&
 326		    !test_bit(ICE_VSI_DOWN, vsi->state)) {
 327			netif_tx_wake_queue(txring_txq(tx_ring));
 328			++tx_ring->ring_stats->tx_stats.restart_q;
 329		}
 330	}
 331
 332	return !!budget;
 333}
 334
 335/**
 336 * ice_setup_tx_ring - Allocate the Tx descriptors
 337 * @tx_ring: the Tx ring to set up
 338 *
 339 * Return 0 on success, negative on error
 340 */
 341int ice_setup_tx_ring(struct ice_tx_ring *tx_ring)
 342{
 343	struct device *dev = tx_ring->dev;
 344	u32 size;
 345
 346	if (!dev)
 347		return -ENOMEM;
 348
 349	/* warn if we are about to overwrite the pointer */
 350	WARN_ON(tx_ring->tx_buf);
 351	tx_ring->tx_buf =
 352		devm_kcalloc(dev, sizeof(*tx_ring->tx_buf), tx_ring->count,
 353			     GFP_KERNEL);
 354	if (!tx_ring->tx_buf)
 355		return -ENOMEM;
 356
 357	/* round up to nearest page */
 358	size = ALIGN(tx_ring->count * sizeof(struct ice_tx_desc),
 359		     PAGE_SIZE);
 360	tx_ring->desc = dmam_alloc_coherent(dev, size, &tx_ring->dma,
 361					    GFP_KERNEL);
 362	if (!tx_ring->desc) {
 363		dev_err(dev, "Unable to allocate memory for the Tx descriptor ring, size=%d\n",
 364			size);
 365		goto err;
 366	}
 367
 368	tx_ring->next_to_use = 0;
 369	tx_ring->next_to_clean = 0;
 370	tx_ring->ring_stats->tx_stats.prev_pkt = -1;
 371	return 0;
 372
 373err:
 374	devm_kfree(dev, tx_ring->tx_buf);
 375	tx_ring->tx_buf = NULL;
 376	return -ENOMEM;
 377}
 378
 379/**
 380 * ice_clean_rx_ring - Free Rx buffers
 381 * @rx_ring: ring to be cleaned
 382 */
 383void ice_clean_rx_ring(struct ice_rx_ring *rx_ring)
 384{
 385	struct xdp_buff *xdp = &rx_ring->xdp;
 386	struct device *dev = rx_ring->dev;
 387	u32 size;
 388	u16 i;
 389
 390	/* ring already cleared, nothing to do */
 391	if (!rx_ring->rx_buf)
 392		return;
 393
 
 
 
 
 
 394	if (rx_ring->xsk_pool) {
 395		ice_xsk_clean_rx_ring(rx_ring);
 396		goto rx_skip_free;
 397	}
 398
 399	if (xdp->data) {
 400		xdp_return_buff(xdp);
 401		xdp->data = NULL;
 402	}
 403
 404	/* Free all the Rx ring sk_buffs */
 405	for (i = 0; i < rx_ring->count; i++) {
 406		struct ice_rx_buf *rx_buf = &rx_ring->rx_buf[i];
 407
 408		if (!rx_buf->page)
 409			continue;
 410
 411		/* Invalidate cache lines that may have been written to by
 412		 * device so that we avoid corrupting memory.
 413		 */
 414		dma_sync_single_range_for_cpu(dev, rx_buf->dma,
 415					      rx_buf->page_offset,
 416					      rx_ring->rx_buf_len,
 417					      DMA_FROM_DEVICE);
 418
 419		/* free resources associated with mapping */
 420		dma_unmap_page_attrs(dev, rx_buf->dma, ice_rx_pg_size(rx_ring),
 421				     DMA_FROM_DEVICE, ICE_RX_DMA_ATTR);
 422		__page_frag_cache_drain(rx_buf->page, rx_buf->pagecnt_bias);
 423
 424		rx_buf->page = NULL;
 425		rx_buf->page_offset = 0;
 426	}
 427
 428rx_skip_free:
 429	if (rx_ring->xsk_pool)
 430		memset(rx_ring->xdp_buf, 0, array_size(rx_ring->count, sizeof(*rx_ring->xdp_buf)));
 431	else
 432		memset(rx_ring->rx_buf, 0, array_size(rx_ring->count, sizeof(*rx_ring->rx_buf)));
 433
 434	/* Zero out the descriptor ring */
 435	size = ALIGN(rx_ring->count * sizeof(union ice_32byte_rx_desc),
 436		     PAGE_SIZE);
 437	memset(rx_ring->desc, 0, size);
 438
 439	rx_ring->next_to_alloc = 0;
 440	rx_ring->next_to_clean = 0;
 441	rx_ring->first_desc = 0;
 442	rx_ring->next_to_use = 0;
 443}
 444
 445/**
 446 * ice_free_rx_ring - Free Rx resources
 447 * @rx_ring: ring to clean the resources from
 448 *
 449 * Free all receive software resources
 450 */
 451void ice_free_rx_ring(struct ice_rx_ring *rx_ring)
 452{
 453	u32 size;
 454
 455	ice_clean_rx_ring(rx_ring);
 456	if (rx_ring->vsi->type == ICE_VSI_PF)
 457		if (xdp_rxq_info_is_reg(&rx_ring->xdp_rxq))
 458			xdp_rxq_info_unreg(&rx_ring->xdp_rxq);
 459	WRITE_ONCE(rx_ring->xdp_prog, NULL);
 460	if (rx_ring->xsk_pool) {
 461		kfree(rx_ring->xdp_buf);
 462		rx_ring->xdp_buf = NULL;
 463	} else {
 464		kfree(rx_ring->rx_buf);
 465		rx_ring->rx_buf = NULL;
 466	}
 467
 468	if (rx_ring->desc) {
 469		size = ALIGN(rx_ring->count * sizeof(union ice_32byte_rx_desc),
 470			     PAGE_SIZE);
 471		dmam_free_coherent(rx_ring->dev, size,
 472				   rx_ring->desc, rx_ring->dma);
 473		rx_ring->desc = NULL;
 474	}
 475}
 476
 477/**
 478 * ice_setup_rx_ring - Allocate the Rx descriptors
 479 * @rx_ring: the Rx ring to set up
 480 *
 481 * Return 0 on success, negative on error
 482 */
 483int ice_setup_rx_ring(struct ice_rx_ring *rx_ring)
 484{
 485	struct device *dev = rx_ring->dev;
 486	u32 size;
 487
 488	if (!dev)
 489		return -ENOMEM;
 490
 491	/* warn if we are about to overwrite the pointer */
 492	WARN_ON(rx_ring->rx_buf);
 493	rx_ring->rx_buf =
 494		kcalloc(rx_ring->count, sizeof(*rx_ring->rx_buf), GFP_KERNEL);
 495	if (!rx_ring->rx_buf)
 496		return -ENOMEM;
 497
 498	/* round up to nearest page */
 499	size = ALIGN(rx_ring->count * sizeof(union ice_32byte_rx_desc),
 500		     PAGE_SIZE);
 501	rx_ring->desc = dmam_alloc_coherent(dev, size, &rx_ring->dma,
 502					    GFP_KERNEL);
 503	if (!rx_ring->desc) {
 504		dev_err(dev, "Unable to allocate memory for the Rx descriptor ring, size=%d\n",
 505			size);
 506		goto err;
 507	}
 508
 509	rx_ring->next_to_use = 0;
 510	rx_ring->next_to_clean = 0;
 511	rx_ring->first_desc = 0;
 512
 513	if (ice_is_xdp_ena_vsi(rx_ring->vsi))
 514		WRITE_ONCE(rx_ring->xdp_prog, rx_ring->vsi->xdp_prog);
 515
 
 
 
 
 
 516	return 0;
 517
 518err:
 519	kfree(rx_ring->rx_buf);
 520	rx_ring->rx_buf = NULL;
 521	return -ENOMEM;
 522}
 523
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 524/**
 525 * ice_run_xdp - Executes an XDP program on initialized xdp_buff
 526 * @rx_ring: Rx ring
 527 * @xdp: xdp_buff used as input to the XDP program
 528 * @xdp_prog: XDP program to run
 529 * @xdp_ring: ring to be used for XDP_TX action
 530 * @eop_desc: Last descriptor in packet to read metadata from
 531 *
 532 * Returns any of ICE_XDP_{PASS, CONSUMED, TX, REDIR}
 533 */
 534static u32
 535ice_run_xdp(struct ice_rx_ring *rx_ring, struct xdp_buff *xdp,
 536	    struct bpf_prog *xdp_prog, struct ice_tx_ring *xdp_ring,
 537	    union ice_32b_rx_flex_desc *eop_desc)
 538{
 539	unsigned int ret = ICE_XDP_PASS;
 540	u32 act;
 541
 542	if (!xdp_prog)
 543		goto exit;
 544
 545	ice_xdp_meta_set_desc(xdp, eop_desc);
 546
 547	act = bpf_prog_run_xdp(xdp_prog, xdp);
 548	switch (act) {
 549	case XDP_PASS:
 550		break;
 551	case XDP_TX:
 552		if (static_branch_unlikely(&ice_xdp_locking_key))
 553			spin_lock(&xdp_ring->tx_lock);
 554		ret = __ice_xmit_xdp_ring(xdp, xdp_ring, false);
 555		if (static_branch_unlikely(&ice_xdp_locking_key))
 556			spin_unlock(&xdp_ring->tx_lock);
 557		if (ret == ICE_XDP_CONSUMED)
 558			goto out_failure;
 559		break;
 560	case XDP_REDIRECT:
 561		if (xdp_do_redirect(rx_ring->netdev, xdp, xdp_prog))
 
 562			goto out_failure;
 563		ret = ICE_XDP_REDIR;
 564		break;
 565	default:
 566		bpf_warn_invalid_xdp_action(rx_ring->netdev, xdp_prog, act);
 567		fallthrough;
 568	case XDP_ABORTED:
 569out_failure:
 570		trace_xdp_exception(rx_ring->netdev, xdp_prog, act);
 571		fallthrough;
 572	case XDP_DROP:
 573		ret = ICE_XDP_CONSUMED;
 574	}
 575exit:
 576	return ret;
 577}
 578
 579/**
 580 * ice_xmit_xdp_ring - submit frame to XDP ring for transmission
 581 * @xdpf: XDP frame that will be converted to XDP buff
 582 * @xdp_ring: XDP ring for transmission
 583 */
 584static int ice_xmit_xdp_ring(const struct xdp_frame *xdpf,
 585			     struct ice_tx_ring *xdp_ring)
 586{
 587	struct xdp_buff xdp;
 588
 589	xdp.data_hard_start = (void *)xdpf;
 590	xdp.data = xdpf->data;
 591	xdp.data_end = xdp.data + xdpf->len;
 592	xdp.frame_sz = xdpf->frame_sz;
 593	xdp.flags = xdpf->flags;
 594
 595	return __ice_xmit_xdp_ring(&xdp, xdp_ring, true);
 596}
 597
 598/**
 599 * ice_xdp_xmit - submit packets to XDP ring for transmission
 600 * @dev: netdev
 601 * @n: number of XDP frames to be transmitted
 602 * @frames: XDP frames to be transmitted
 603 * @flags: transmit flags
 604 *
 605 * Returns number of frames successfully sent. Failed frames
 606 * will be free'ed by XDP core.
 607 * For error cases, a negative errno code is returned and no-frames
 608 * are transmitted (caller must handle freeing frames).
 609 */
 610int
 611ice_xdp_xmit(struct net_device *dev, int n, struct xdp_frame **frames,
 612	     u32 flags)
 613{
 614	struct ice_netdev_priv *np = netdev_priv(dev);
 615	unsigned int queue_index = smp_processor_id();
 616	struct ice_vsi *vsi = np->vsi;
 617	struct ice_tx_ring *xdp_ring;
 618	struct ice_tx_buf *tx_buf;
 619	int nxmit = 0, i;
 620
 621	if (test_bit(ICE_VSI_DOWN, vsi->state))
 622		return -ENETDOWN;
 623
 624	if (!ice_is_xdp_ena_vsi(vsi))
 625		return -ENXIO;
 626
 627	if (unlikely(flags & ~XDP_XMIT_FLAGS_MASK))
 628		return -EINVAL;
 629
 630	if (static_branch_unlikely(&ice_xdp_locking_key)) {
 631		queue_index %= vsi->num_xdp_txq;
 632		xdp_ring = vsi->xdp_rings[queue_index];
 633		spin_lock(&xdp_ring->tx_lock);
 634	} else {
 635		/* Generally, should not happen */
 636		if (unlikely(queue_index >= vsi->num_xdp_txq))
 637			return -ENXIO;
 638		xdp_ring = vsi->xdp_rings[queue_index];
 639	}
 640
 641	tx_buf = &xdp_ring->tx_buf[xdp_ring->next_to_use];
 642	for (i = 0; i < n; i++) {
 643		const struct xdp_frame *xdpf = frames[i];
 644		int err;
 645
 646		err = ice_xmit_xdp_ring(xdpf, xdp_ring);
 647		if (err != ICE_XDP_TX)
 648			break;
 649		nxmit++;
 650	}
 651
 652	tx_buf->rs_idx = ice_set_rs_bit(xdp_ring);
 653	if (unlikely(flags & XDP_XMIT_FLUSH))
 654		ice_xdp_ring_update_tail(xdp_ring);
 655
 656	if (static_branch_unlikely(&ice_xdp_locking_key))
 657		spin_unlock(&xdp_ring->tx_lock);
 658
 659	return nxmit;
 660}
 661
 662/**
 663 * ice_alloc_mapped_page - recycle or make a new page
 664 * @rx_ring: ring to use
 665 * @bi: rx_buf struct to modify
 666 *
 667 * Returns true if the page was successfully allocated or
 668 * reused.
 669 */
 670static bool
 671ice_alloc_mapped_page(struct ice_rx_ring *rx_ring, struct ice_rx_buf *bi)
 672{
 673	struct page *page = bi->page;
 674	dma_addr_t dma;
 675
 676	/* since we are recycling buffers we should seldom need to alloc */
 677	if (likely(page))
 678		return true;
 679
 680	/* alloc new page for storage */
 681	page = dev_alloc_pages(ice_rx_pg_order(rx_ring));
 682	if (unlikely(!page)) {
 683		rx_ring->ring_stats->rx_stats.alloc_page_failed++;
 684		return false;
 685	}
 686
 687	/* map page for use */
 688	dma = dma_map_page_attrs(rx_ring->dev, page, 0, ice_rx_pg_size(rx_ring),
 689				 DMA_FROM_DEVICE, ICE_RX_DMA_ATTR);
 690
 691	/* if mapping failed free memory back to system since
 692	 * there isn't much point in holding memory we can't use
 693	 */
 694	if (dma_mapping_error(rx_ring->dev, dma)) {
 695		__free_pages(page, ice_rx_pg_order(rx_ring));
 696		rx_ring->ring_stats->rx_stats.alloc_page_failed++;
 697		return false;
 698	}
 699
 700	bi->dma = dma;
 701	bi->page = page;
 702	bi->page_offset = rx_ring->rx_offset;
 703	page_ref_add(page, USHRT_MAX - 1);
 704	bi->pagecnt_bias = USHRT_MAX;
 705
 706	return true;
 707}
 708
 709/**
 710 * ice_alloc_rx_bufs - Replace used receive buffers
 711 * @rx_ring: ring to place buffers on
 712 * @cleaned_count: number of buffers to replace
 713 *
 714 * Returns false if all allocations were successful, true if any fail. Returning
 715 * true signals to the caller that we didn't replace cleaned_count buffers and
 716 * there is more work to do.
 717 *
 718 * First, try to clean "cleaned_count" Rx buffers. Then refill the cleaned Rx
 719 * buffers. Then bump tail at most one time. Grouping like this lets us avoid
 720 * multiple tail writes per call.
 721 */
 722bool ice_alloc_rx_bufs(struct ice_rx_ring *rx_ring, unsigned int cleaned_count)
 723{
 724	union ice_32b_rx_flex_desc *rx_desc;
 725	u16 ntu = rx_ring->next_to_use;
 726	struct ice_rx_buf *bi;
 727
 728	/* do nothing if no valid netdev defined */
 729	if ((!rx_ring->netdev && rx_ring->vsi->type != ICE_VSI_CTRL) ||
 730	    !cleaned_count)
 731		return false;
 732
 733	/* get the Rx descriptor and buffer based on next_to_use */
 734	rx_desc = ICE_RX_DESC(rx_ring, ntu);
 735	bi = &rx_ring->rx_buf[ntu];
 736
 737	do {
 738		/* if we fail here, we have work remaining */
 739		if (!ice_alloc_mapped_page(rx_ring, bi))
 740			break;
 741
 742		/* sync the buffer for use by the device */
 743		dma_sync_single_range_for_device(rx_ring->dev, bi->dma,
 744						 bi->page_offset,
 745						 rx_ring->rx_buf_len,
 746						 DMA_FROM_DEVICE);
 747
 748		/* Refresh the desc even if buffer_addrs didn't change
 749		 * because each write-back erases this info.
 750		 */
 751		rx_desc->read.pkt_addr = cpu_to_le64(bi->dma + bi->page_offset);
 752
 753		rx_desc++;
 754		bi++;
 755		ntu++;
 756		if (unlikely(ntu == rx_ring->count)) {
 757			rx_desc = ICE_RX_DESC(rx_ring, 0);
 758			bi = rx_ring->rx_buf;
 759			ntu = 0;
 760		}
 761
 762		/* clear the status bits for the next_to_use descriptor */
 763		rx_desc->wb.status_error0 = 0;
 764
 765		cleaned_count--;
 766	} while (cleaned_count);
 767
 768	if (rx_ring->next_to_use != ntu)
 769		ice_release_rx_desc(rx_ring, ntu);
 770
 771	return !!cleaned_count;
 772}
 773
 774/**
 775 * ice_rx_buf_adjust_pg_offset - Prepare Rx buffer for reuse
 776 * @rx_buf: Rx buffer to adjust
 777 * @size: Size of adjustment
 778 *
 779 * Update the offset within page so that Rx buf will be ready to be reused.
 780 * For systems with PAGE_SIZE < 8192 this function will flip the page offset
 781 * so the second half of page assigned to Rx buffer will be used, otherwise
 782 * the offset is moved by "size" bytes
 783 */
 784static void
 785ice_rx_buf_adjust_pg_offset(struct ice_rx_buf *rx_buf, unsigned int size)
 786{
 787#if (PAGE_SIZE < 8192)
 788	/* flip page offset to other buffer */
 789	rx_buf->page_offset ^= size;
 790#else
 791	/* move offset up to the next cache line */
 792	rx_buf->page_offset += size;
 793#endif
 794}
 795
 796/**
 797 * ice_can_reuse_rx_page - Determine if page can be reused for another Rx
 798 * @rx_buf: buffer containing the page
 
 799 *
 800 * If page is reusable, we have a green light for calling ice_reuse_rx_page,
 801 * which will assign the current buffer to the buffer that next_to_alloc is
 802 * pointing to; otherwise, the DMA mapping needs to be destroyed and
 803 * page freed
 804 */
 805static bool
 806ice_can_reuse_rx_page(struct ice_rx_buf *rx_buf)
 807{
 808	unsigned int pagecnt_bias = rx_buf->pagecnt_bias;
 809	struct page *page = rx_buf->page;
 810
 811	/* avoid re-using remote and pfmemalloc pages */
 812	if (!dev_page_is_reusable(page))
 813		return false;
 814
 
 815	/* if we are only owner of page we can reuse it */
 816	if (unlikely(rx_buf->pgcnt - pagecnt_bias > 1))
 817		return false;
 818#if (PAGE_SIZE >= 8192)
 819#define ICE_LAST_OFFSET \
 820	(SKB_WITH_OVERHEAD(PAGE_SIZE) - ICE_RXBUF_3072)
 821	if (rx_buf->page_offset > ICE_LAST_OFFSET)
 822		return false;
 823#endif /* PAGE_SIZE >= 8192) */
 824
 825	/* If we have drained the page fragment pool we need to update
 826	 * the pagecnt_bias and page count so that we fully restock the
 827	 * number of references the driver holds.
 828	 */
 829	if (unlikely(pagecnt_bias == 1)) {
 830		page_ref_add(page, USHRT_MAX - 1);
 831		rx_buf->pagecnt_bias = USHRT_MAX;
 832	}
 833
 834	return true;
 835}
 836
 837/**
 838 * ice_add_xdp_frag - Add contents of Rx buffer to xdp buf as a frag
 839 * @rx_ring: Rx descriptor ring to transact packets on
 840 * @xdp: xdp buff to place the data into
 841 * @rx_buf: buffer containing page to add
 
 842 * @size: packet length from rx_desc
 843 *
 844 * This function will add the data contained in rx_buf->page to the xdp buf.
 845 * It will just attach the page as a frag.
 
 846 */
 847static int
 848ice_add_xdp_frag(struct ice_rx_ring *rx_ring, struct xdp_buff *xdp,
 849		 struct ice_rx_buf *rx_buf, const unsigned int size)
 850{
 851	struct skb_shared_info *sinfo = xdp_get_shared_info_from_buff(xdp);
 
 
 
 
 852
 853	if (!size)
 854		return 0;
 855
 856	if (!xdp_buff_has_frags(xdp)) {
 857		sinfo->nr_frags = 0;
 858		sinfo->xdp_frags_size = 0;
 859		xdp_buff_set_frags_flag(xdp);
 860	}
 861
 862	if (unlikely(sinfo->nr_frags == MAX_SKB_FRAGS))
 863		return -ENOMEM;
 864
 865	__skb_fill_page_desc_noacc(sinfo, sinfo->nr_frags++, rx_buf->page,
 866				   rx_buf->page_offset, size);
 867	sinfo->xdp_frags_size += size;
 868	/* remember frag count before XDP prog execution; bpf_xdp_adjust_tail()
 869	 * can pop off frags but driver has to handle it on its own
 870	 */
 871	rx_ring->nr_frags = sinfo->nr_frags;
 872
 873	if (page_is_pfmemalloc(rx_buf->page))
 874		xdp_buff_set_frag_pfmemalloc(xdp);
 875
 876	return 0;
 
 877}
 878
 879/**
 880 * ice_reuse_rx_page - page flip buffer and store it back on the ring
 881 * @rx_ring: Rx descriptor ring to store buffers on
 882 * @old_buf: donor buffer to have page reused
 883 *
 884 * Synchronizes page for reuse by the adapter
 885 */
 886static void
 887ice_reuse_rx_page(struct ice_rx_ring *rx_ring, struct ice_rx_buf *old_buf)
 888{
 889	u16 nta = rx_ring->next_to_alloc;
 890	struct ice_rx_buf *new_buf;
 891
 892	new_buf = &rx_ring->rx_buf[nta];
 893
 894	/* update, and store next to alloc */
 895	nta++;
 896	rx_ring->next_to_alloc = (nta < rx_ring->count) ? nta : 0;
 897
 898	/* Transfer page from old buffer to new buffer.
 899	 * Move each member individually to avoid possible store
 900	 * forwarding stalls and unnecessary copy of skb.
 901	 */
 902	new_buf->dma = old_buf->dma;
 903	new_buf->page = old_buf->page;
 904	new_buf->page_offset = old_buf->page_offset;
 905	new_buf->pagecnt_bias = old_buf->pagecnt_bias;
 906}
 907
 908/**
 909 * ice_get_rx_buf - Fetch Rx buffer and synchronize data for use
 910 * @rx_ring: Rx descriptor ring to transact packets on
 911 * @size: size of buffer to add to skb
 912 * @ntc: index of next to clean element
 913 *
 914 * This function will pull an Rx buffer from the ring and synchronize it
 915 * for use by the CPU.
 916 */
 917static struct ice_rx_buf *
 918ice_get_rx_buf(struct ice_rx_ring *rx_ring, const unsigned int size,
 919	       const unsigned int ntc)
 920{
 921	struct ice_rx_buf *rx_buf;
 922
 923	rx_buf = &rx_ring->rx_buf[ntc];
 
 
 
 
 
 
 924	prefetchw(rx_buf->page);
 925
 926	if (!size)
 927		return rx_buf;
 928	/* we are reusing so sync this buffer for CPU use */
 929	dma_sync_single_range_for_cpu(rx_ring->dev, rx_buf->dma,
 930				      rx_buf->page_offset, size,
 931				      DMA_FROM_DEVICE);
 932
 933	/* We have pulled a buffer for use, so decrement pagecnt_bias */
 934	rx_buf->pagecnt_bias--;
 935
 936	return rx_buf;
 937}
 938
 939/**
 940 * ice_get_pgcnts - grab page_count() for gathered fragments
 941 * @rx_ring: Rx descriptor ring to store the page counts on
 942 *
 943 * This function is intended to be called right before running XDP
 944 * program so that the page recycling mechanism will be able to take
 945 * a correct decision regarding underlying pages; this is done in such
 946 * way as XDP program can change the refcount of page
 947 */
 948static void ice_get_pgcnts(struct ice_rx_ring *rx_ring)
 949{
 950	u32 nr_frags = rx_ring->nr_frags + 1;
 951	u32 idx = rx_ring->first_desc;
 952	struct ice_rx_buf *rx_buf;
 953	u32 cnt = rx_ring->count;
 954
 955	for (int i = 0; i < nr_frags; i++) {
 956		rx_buf = &rx_ring->rx_buf[idx];
 957		rx_buf->pgcnt = page_count(rx_buf->page);
 958
 959		if (++idx == cnt)
 960			idx = 0;
 961	}
 962}
 963
 964/**
 965 * ice_build_skb - Build skb around an existing buffer
 966 * @rx_ring: Rx descriptor ring to transact packets on
 
 967 * @xdp: xdp_buff pointing to the data
 968 *
 969 * This function builds an skb around an existing XDP buffer, taking care
 970 * to set up the skb correctly and avoid any memcpy overhead. Driver has
 971 * already combined frags (if any) to skb_shared_info.
 972 */
 973static struct sk_buff *
 974ice_build_skb(struct ice_rx_ring *rx_ring, struct xdp_buff *xdp)
 
 975{
 976	u8 metasize = xdp->data - xdp->data_meta;
 977	struct skb_shared_info *sinfo = NULL;
 978	unsigned int nr_frags;
 
 
 
 
 
 979	struct sk_buff *skb;
 980
 981	if (unlikely(xdp_buff_has_frags(xdp))) {
 982		sinfo = xdp_get_shared_info_from_buff(xdp);
 983		nr_frags = sinfo->nr_frags;
 984	}
 985
 986	/* Prefetch first cache line of first page. If xdp->data_meta
 987	 * is unused, this points exactly as xdp->data, otherwise we
 988	 * likely have a consumer accessing first few bytes of meta
 989	 * data, and then actual data.
 990	 */
 991	net_prefetch(xdp->data_meta);
 992	/* build an skb around the page buffer */
 993	skb = napi_build_skb(xdp->data_hard_start, xdp->frame_sz);
 994	if (unlikely(!skb))
 995		return NULL;
 996
 997	/* must to record Rx queue, otherwise OS features such as
 998	 * symmetric queue won't work
 999	 */
1000	skb_record_rx_queue(skb, rx_ring->q_index);
1001
1002	/* update pointers within the skb to store the data */
1003	skb_reserve(skb, xdp->data - xdp->data_hard_start);
1004	__skb_put(skb, xdp->data_end - xdp->data);
1005	if (metasize)
1006		skb_metadata_set(skb, metasize);
1007
1008	if (unlikely(xdp_buff_has_frags(xdp)))
1009		xdp_update_skb_shared_info(skb, nr_frags,
1010					   sinfo->xdp_frags_size,
1011					   nr_frags * xdp->frame_sz,
1012					   xdp_buff_is_frag_pfmemalloc(xdp));
1013
1014	return skb;
1015}
1016
1017/**
1018 * ice_construct_skb - Allocate skb and populate it
1019 * @rx_ring: Rx descriptor ring to transact packets on
 
1020 * @xdp: xdp_buff pointing to the data
1021 *
1022 * This function allocates an skb. It then populates it with the page
1023 * data from the current receive descriptor, taking care to set up the
1024 * skb correctly.
1025 */
1026static struct sk_buff *
1027ice_construct_skb(struct ice_rx_ring *rx_ring, struct xdp_buff *xdp)
 
1028{
 
1029	unsigned int size = xdp->data_end - xdp->data;
1030	struct skb_shared_info *sinfo = NULL;
1031	struct ice_rx_buf *rx_buf;
1032	unsigned int nr_frags = 0;
1033	unsigned int headlen;
1034	struct sk_buff *skb;
1035
1036	/* prefetch first cache line of first page */
1037	net_prefetch(xdp->data);
1038
1039	if (unlikely(xdp_buff_has_frags(xdp))) {
1040		sinfo = xdp_get_shared_info_from_buff(xdp);
1041		nr_frags = sinfo->nr_frags;
1042	}
1043
1044	/* allocate a skb to store the frags */
1045	skb = napi_alloc_skb(&rx_ring->q_vector->napi, ICE_RX_HDR_SIZE);
 
 
1046	if (unlikely(!skb))
1047		return NULL;
1048
1049	rx_buf = &rx_ring->rx_buf[rx_ring->first_desc];
1050	skb_record_rx_queue(skb, rx_ring->q_index);
1051	/* Determine available headroom for copy */
1052	headlen = size;
1053	if (headlen > ICE_RX_HDR_SIZE)
1054		headlen = eth_get_headlen(skb->dev, xdp->data, ICE_RX_HDR_SIZE);
1055
1056	/* align pull length to size of long to optimize memcpy performance */
1057	memcpy(__skb_put(skb, headlen), xdp->data, ALIGN(headlen,
1058							 sizeof(long)));
 
 
 
 
 
1059
1060	/* if we exhaust the linear part then add what is left as a frag */
1061	size -= headlen;
1062	if (size) {
1063		/* besides adding here a partial frag, we are going to add
1064		 * frags from xdp_buff, make sure there is enough space for
1065		 * them
1066		 */
1067		if (unlikely(nr_frags >= MAX_SKB_FRAGS - 1)) {
1068			dev_kfree_skb(skb);
1069			return NULL;
1070		}
1071		skb_add_rx_frag(skb, 0, rx_buf->page,
1072				rx_buf->page_offset + headlen, size,
1073				xdp->frame_sz);
 
1074	} else {
1075		/* buffer is unused, restore biased page count in Rx buffer;
1076		 * data was copied onto skb's linear part so there's no
1077		 * need for adjusting page offset and we can reuse this buffer
1078		 * as-is
1079		 */
1080		rx_buf->pagecnt_bias++;
1081	}
1082
1083	if (unlikely(xdp_buff_has_frags(xdp))) {
1084		struct skb_shared_info *skinfo = skb_shinfo(skb);
1085
1086		memcpy(&skinfo->frags[skinfo->nr_frags], &sinfo->frags[0],
1087		       sizeof(skb_frag_t) * nr_frags);
1088
1089		xdp_update_skb_shared_info(skb, skinfo->nr_frags + nr_frags,
1090					   sinfo->xdp_frags_size,
1091					   nr_frags * xdp->frame_sz,
1092					   xdp_buff_is_frag_pfmemalloc(xdp));
1093	}
1094
1095	return skb;
1096}
1097
1098/**
1099 * ice_put_rx_buf - Clean up used buffer and either recycle or free
1100 * @rx_ring: Rx descriptor ring to transact packets on
1101 * @rx_buf: Rx buffer to pull data from
 
1102 *
1103 * This function will clean up the contents of the rx_buf. It will either
1104 * recycle the buffer or unmap it and free the associated resources.
 
1105 */
1106static void
1107ice_put_rx_buf(struct ice_rx_ring *rx_ring, struct ice_rx_buf *rx_buf)
 
1108{
 
 
 
 
 
 
1109	if (!rx_buf)
1110		return;
1111
1112	if (ice_can_reuse_rx_page(rx_buf)) {
1113		/* hand second half of page back to the ring */
1114		ice_reuse_rx_page(rx_ring, rx_buf);
1115	} else {
1116		/* we are not reusing the buffer so unmap it */
1117		dma_unmap_page_attrs(rx_ring->dev, rx_buf->dma,
1118				     ice_rx_pg_size(rx_ring), DMA_FROM_DEVICE,
1119				     ICE_RX_DMA_ATTR);
1120		__page_frag_cache_drain(rx_buf->page, rx_buf->pagecnt_bias);
1121	}
1122
1123	/* clear contents of buffer_info */
1124	rx_buf->page = NULL;
1125}
1126
1127/**
1128 * ice_put_rx_mbuf - ice_put_rx_buf() caller, for all frame frags
1129 * @rx_ring: Rx ring with all the auxiliary data
1130 * @xdp: XDP buffer carrying linear + frags part
1131 * @xdp_xmit: XDP_TX/XDP_REDIRECT verdict storage
1132 * @ntc: a current next_to_clean value to be stored at rx_ring
1133 * @verdict: return code from XDP program execution
1134 *
1135 * Walk through gathered fragments and satisfy internal page
1136 * recycle mechanism; we take here an action related to verdict
1137 * returned by XDP program;
1138 */
1139static void ice_put_rx_mbuf(struct ice_rx_ring *rx_ring, struct xdp_buff *xdp,
1140			    u32 *xdp_xmit, u32 ntc, u32 verdict)
1141{
1142	u32 nr_frags = rx_ring->nr_frags + 1;
1143	u32 idx = rx_ring->first_desc;
1144	u32 cnt = rx_ring->count;
1145	u32 post_xdp_frags = 1;
1146	struct ice_rx_buf *buf;
1147	int i;
1148
1149	if (unlikely(xdp_buff_has_frags(xdp)))
1150		post_xdp_frags += xdp_get_shared_info_from_buff(xdp)->nr_frags;
1151
1152	for (i = 0; i < post_xdp_frags; i++) {
1153		buf = &rx_ring->rx_buf[idx];
1154
1155		if (verdict & (ICE_XDP_TX | ICE_XDP_REDIR)) {
1156			ice_rx_buf_adjust_pg_offset(buf, xdp->frame_sz);
1157			*xdp_xmit |= verdict;
1158		} else if (verdict & ICE_XDP_CONSUMED) {
1159			buf->pagecnt_bias++;
1160		} else if (verdict == ICE_XDP_PASS) {
1161			ice_rx_buf_adjust_pg_offset(buf, xdp->frame_sz);
1162		}
1163
1164		ice_put_rx_buf(rx_ring, buf);
1165
1166		if (++idx == cnt)
1167			idx = 0;
1168	}
1169	/* handle buffers that represented frags released by XDP prog;
1170	 * for these we keep pagecnt_bias as-is; refcount from struct page
1171	 * has been decremented within XDP prog and we do not have to increase
1172	 * the biased refcnt
1173	 */
1174	for (; i < nr_frags; i++) {
1175		buf = &rx_ring->rx_buf[idx];
1176		ice_put_rx_buf(rx_ring, buf);
1177		if (++idx == cnt)
1178			idx = 0;
1179	}
1180
1181	xdp->data = NULL;
1182	rx_ring->first_desc = ntc;
1183	rx_ring->nr_frags = 0;
1184}
1185
1186/**
1187 * ice_clean_rx_irq - Clean completed descriptors from Rx ring - bounce buf
1188 * @rx_ring: Rx descriptor ring to transact packets on
1189 * @budget: Total limit on number of packets to process
1190 *
1191 * This function provides a "bounce buffer" approach to Rx interrupt
1192 * processing. The advantage to this is that on systems that have
1193 * expensive overhead for IOMMU access this provides a means of avoiding
1194 * it by maintaining the mapping of the page to the system.
1195 *
1196 * Returns amount of work completed
1197 */
1198int ice_clean_rx_irq(struct ice_rx_ring *rx_ring, int budget)
1199{
1200	unsigned int total_rx_bytes = 0, total_rx_pkts = 0;
 
1201	unsigned int offset = rx_ring->rx_offset;
1202	struct xdp_buff *xdp = &rx_ring->xdp;
1203	struct ice_tx_ring *xdp_ring = NULL;
 
 
1204	struct bpf_prog *xdp_prog = NULL;
1205	u32 ntc = rx_ring->next_to_clean;
1206	u32 cached_ntu, xdp_verdict;
1207	u32 cnt = rx_ring->count;
1208	u32 xdp_xmit = 0;
1209	bool failure;
1210
 
 
 
 
 
 
1211	xdp_prog = READ_ONCE(rx_ring->xdp_prog);
1212	if (xdp_prog) {
1213		xdp_ring = rx_ring->xdp_ring;
1214		cached_ntu = xdp_ring->next_to_use;
1215	}
1216
1217	/* start the loop to process Rx packets bounded by 'budget' */
1218	while (likely(total_rx_pkts < (unsigned int)budget)) {
1219		union ice_32b_rx_flex_desc *rx_desc;
1220		struct ice_rx_buf *rx_buf;
1221		struct sk_buff *skb;
1222		unsigned int size;
1223		u16 stat_err_bits;
1224		u16 vlan_tci;
 
 
1225
1226		/* get the Rx desc from Rx ring based on 'next_to_clean' */
1227		rx_desc = ICE_RX_DESC(rx_ring, ntc);
1228
1229		/* status_error_len will always be zero for unused descriptors
1230		 * because it's cleared in cleanup, and overlaps with hdr_addr
1231		 * which is always zero because packet split isn't used, if the
1232		 * hardware wrote DD then it will be non-zero
1233		 */
1234		stat_err_bits = BIT(ICE_RX_FLEX_DESC_STATUS0_DD_S);
1235		if (!ice_test_staterr(rx_desc->wb.status_error0, stat_err_bits))
1236			break;
1237
1238		/* This memory barrier is needed to keep us from reading
1239		 * any other fields out of the rx_desc until we know the
1240		 * DD bit is set.
1241		 */
1242		dma_rmb();
1243
1244		ice_trace(clean_rx_irq, rx_ring, rx_desc);
1245		if (rx_desc->wb.rxdid == FDIR_DESC_RXDID || !rx_ring->netdev) {
1246			struct ice_vsi *ctrl_vsi = rx_ring->vsi;
1247
1248			if (rx_desc->wb.rxdid == FDIR_DESC_RXDID &&
1249			    ctrl_vsi->vf)
1250				ice_vc_fdir_irq_handler(ctrl_vsi, rx_desc);
1251			if (++ntc == cnt)
1252				ntc = 0;
1253			rx_ring->first_desc = ntc;
1254			continue;
1255		}
1256
1257		size = le16_to_cpu(rx_desc->wb.pkt_len) &
1258			ICE_RX_FLX_DESC_PKT_LEN_M;
1259
1260		/* retrieve a buffer from the ring */
1261		rx_buf = ice_get_rx_buf(rx_ring, size, ntc);
1262
1263		if (!xdp->data) {
1264			void *hard_start;
1265
1266			hard_start = page_address(rx_buf->page) + rx_buf->page_offset -
1267				     offset;
1268			xdp_prepare_buff(xdp, hard_start, offset, size, !!offset);
1269			xdp_buff_clear_frags_flag(xdp);
1270		} else if (ice_add_xdp_frag(rx_ring, xdp, rx_buf, size)) {
1271			ice_put_rx_mbuf(rx_ring, xdp, NULL, ntc, ICE_XDP_CONSUMED);
1272			break;
1273		}
1274		if (++ntc == cnt)
1275			ntc = 0;
1276
1277		/* skip if it is NOP desc */
1278		if (ice_is_non_eop(rx_ring, rx_desc))
1279			continue;
 
 
 
 
1280
1281		ice_get_pgcnts(rx_ring);
1282		xdp_verdict = ice_run_xdp(rx_ring, xdp, xdp_prog, xdp_ring, rx_desc);
1283		if (xdp_verdict == ICE_XDP_PASS)
1284			goto construct_skb;
1285		total_rx_bytes += xdp_get_buff_len(xdp);
1286		total_rx_pkts++;
1287
1288		ice_put_rx_mbuf(rx_ring, xdp, &xdp_xmit, ntc, xdp_verdict);
 
 
 
 
 
 
 
 
 
 
1289
 
 
1290		continue;
1291construct_skb:
1292		if (likely(ice_ring_uses_build_skb(rx_ring)))
1293			skb = ice_build_skb(rx_ring, xdp);
1294		else
1295			skb = ice_construct_skb(rx_ring, xdp);
 
 
 
 
1296		/* exit if we failed to retrieve a buffer */
1297		if (!skb) {
1298			rx_ring->ring_stats->rx_stats.alloc_page_failed++;
1299			xdp_verdict = ICE_XDP_CONSUMED;
 
 
1300		}
1301		ice_put_rx_mbuf(rx_ring, xdp, &xdp_xmit, ntc, xdp_verdict);
1302
1303		if (!skb)
1304			break;
 
 
 
 
1305
1306		stat_err_bits = BIT(ICE_RX_FLEX_DESC_STATUS0_RXE_S);
1307		if (unlikely(ice_test_staterr(rx_desc->wb.status_error0,
1308					      stat_err_bits))) {
1309			dev_kfree_skb_any(skb);
1310			continue;
1311		}
1312
1313		vlan_tci = ice_get_vlan_tci(rx_desc);
1314
1315		/* pad the skb if needed, to make a valid ethernet frame */
1316		if (eth_skb_pad(skb))
 
1317			continue;
 
1318
1319		/* probably a little skewed due to removing CRC */
1320		total_rx_bytes += skb->len;
1321
1322		/* populate checksum, VLAN, and protocol */
1323		ice_process_skb_fields(rx_ring, rx_desc, skb);
 
 
 
1324
1325		ice_trace(clean_rx_irq_indicate, rx_ring, rx_desc, skb);
1326		/* send completed skb up the stack */
1327		ice_receive_skb(rx_ring, skb, vlan_tci);
 
1328
1329		/* update budget accounting */
1330		total_rx_pkts++;
1331	}
1332
1333	rx_ring->next_to_clean = ntc;
1334	/* return up to cleaned_count buffers to hardware */
1335	failure = ice_alloc_rx_bufs(rx_ring, ICE_RX_DESC_UNUSED(rx_ring));
1336
1337	if (xdp_xmit)
1338		ice_finalize_xdp_rx(xdp_ring, xdp_xmit, cached_ntu);
 
1339
1340	if (rx_ring->ring_stats)
1341		ice_update_rx_ring_stats(rx_ring, total_rx_pkts,
1342					 total_rx_bytes);
1343
1344	/* guarantee a trip back through this routine if there was a failure */
1345	return failure ? budget : (int)total_rx_pkts;
1346}
1347
1348static void __ice_update_sample(struct ice_q_vector *q_vector,
1349				struct ice_ring_container *rc,
1350				struct dim_sample *sample,
1351				bool is_tx)
1352{
1353	u64 packets = 0, bytes = 0;
1354
1355	if (is_tx) {
1356		struct ice_tx_ring *tx_ring;
1357
1358		ice_for_each_tx_ring(tx_ring, *rc) {
1359			struct ice_ring_stats *ring_stats;
1360
1361			ring_stats = tx_ring->ring_stats;
1362			if (!ring_stats)
1363				continue;
1364			packets += ring_stats->stats.pkts;
1365			bytes += ring_stats->stats.bytes;
1366		}
1367	} else {
1368		struct ice_rx_ring *rx_ring;
1369
1370		ice_for_each_rx_ring(rx_ring, *rc) {
1371			struct ice_ring_stats *ring_stats;
1372
1373			ring_stats = rx_ring->ring_stats;
1374			if (!ring_stats)
1375				continue;
1376			packets += ring_stats->stats.pkts;
1377			bytes += ring_stats->stats.bytes;
1378		}
1379	}
1380
1381	dim_update_sample(q_vector->total_events, packets, bytes, sample);
1382	sample->comp_ctr = 0;
1383
1384	/* if dim settings get stale, like when not updated for 1
1385	 * second or longer, force it to start again. This addresses the
1386	 * frequent case of an idle queue being switched to by the
1387	 * scheduler. The 1,000 here means 1,000 milliseconds.
1388	 */
1389	if (ktime_ms_delta(sample->time, rc->dim.start_sample.time) >= 1000)
1390		rc->dim.state = DIM_START_MEASURE;
1391}
1392
1393/**
1394 * ice_net_dim - Update net DIM algorithm
1395 * @q_vector: the vector associated with the interrupt
1396 *
1397 * Create a DIM sample and notify net_dim() so that it can possibly decide
1398 * a new ITR value based on incoming packets, bytes, and interrupts.
1399 *
1400 * This function is a no-op if the ring is not configured to dynamic ITR.
1401 */
1402static void ice_net_dim(struct ice_q_vector *q_vector)
1403{
1404	struct ice_ring_container *tx = &q_vector->tx;
1405	struct ice_ring_container *rx = &q_vector->rx;
1406
1407	if (ITR_IS_DYNAMIC(tx)) {
1408		struct dim_sample dim_sample;
1409
1410		__ice_update_sample(q_vector, tx, &dim_sample, true);
1411		net_dim(&tx->dim, &dim_sample);
1412	}
1413
1414	if (ITR_IS_DYNAMIC(rx)) {
1415		struct dim_sample dim_sample;
1416
1417		__ice_update_sample(q_vector, rx, &dim_sample, false);
1418		net_dim(&rx->dim, &dim_sample);
1419	}
1420}
1421
1422/**
1423 * ice_buildreg_itr - build value for writing to the GLINT_DYN_CTL register
1424 * @itr_idx: interrupt throttling index
1425 * @itr: interrupt throttling value in usecs
1426 */
1427static u32 ice_buildreg_itr(u16 itr_idx, u16 itr)
1428{
1429	/* The ITR value is reported in microseconds, and the register value is
1430	 * recorded in 2 microsecond units. For this reason we only need to
1431	 * shift by the GLINT_DYN_CTL_INTERVAL_S - ICE_ITR_GRAN_S to apply this
1432	 * granularity as a shift instead of division. The mask makes sure the
1433	 * ITR value is never odd so we don't accidentally write into the field
1434	 * prior to the ITR field.
1435	 */
1436	itr &= ICE_ITR_MASK;
1437
1438	return GLINT_DYN_CTL_INTENA_M | GLINT_DYN_CTL_CLEARPBA_M |
1439		(itr_idx << GLINT_DYN_CTL_ITR_INDX_S) |
1440		(itr << (GLINT_DYN_CTL_INTERVAL_S - ICE_ITR_GRAN_S));
1441}
1442
1443/**
1444 * ice_enable_interrupt - re-enable MSI-X interrupt
1445 * @q_vector: the vector associated with the interrupt to enable
1446 *
1447 * If the VSI is down, the interrupt will not be re-enabled. Also,
1448 * when enabling the interrupt always reset the wb_on_itr to false
1449 * and trigger a software interrupt to clean out internal state.
1450 */
1451static void ice_enable_interrupt(struct ice_q_vector *q_vector)
1452{
1453	struct ice_vsi *vsi = q_vector->vsi;
1454	bool wb_en = q_vector->wb_on_itr;
1455	u32 itr_val;
1456
1457	if (test_bit(ICE_DOWN, vsi->state))
1458		return;
1459
1460	/* trigger an ITR delayed software interrupt when exiting busy poll, to
1461	 * make sure to catch any pending cleanups that might have been missed
1462	 * due to interrupt state transition. If busy poll or poll isn't
1463	 * enabled, then don't update ITR, and just enable the interrupt.
1464	 */
1465	if (!wb_en) {
1466		itr_val = ice_buildreg_itr(ICE_ITR_NONE, 0);
1467	} else {
1468		q_vector->wb_on_itr = false;
1469
1470		/* do two things here with a single write. Set up the third ITR
1471		 * index to be used for software interrupt moderation, and then
1472		 * trigger a software interrupt with a rate limit of 20K on
1473		 * software interrupts, this will help avoid high interrupt
1474		 * loads due to frequently polling and exiting polling.
1475		 */
1476		itr_val = ice_buildreg_itr(ICE_IDX_ITR2, ICE_ITR_20K);
1477		itr_val |= GLINT_DYN_CTL_SWINT_TRIG_M |
1478			   ICE_IDX_ITR2 << GLINT_DYN_CTL_SW_ITR_INDX_S |
1479			   GLINT_DYN_CTL_SW_ITR_INDX_ENA_M;
1480	}
1481	wr32(&vsi->back->hw, GLINT_DYN_CTL(q_vector->reg_idx), itr_val);
1482}
1483
1484/**
1485 * ice_set_wb_on_itr - set WB_ON_ITR for this q_vector
1486 * @q_vector: q_vector to set WB_ON_ITR on
1487 *
1488 * We need to tell hardware to write-back completed descriptors even when
1489 * interrupts are disabled. Descriptors will be written back on cache line
1490 * boundaries without WB_ON_ITR enabled, but if we don't enable WB_ON_ITR
1491 * descriptors may not be written back if they don't fill a cache line until
1492 * the next interrupt.
1493 *
1494 * This sets the write-back frequency to whatever was set previously for the
1495 * ITR indices. Also, set the INTENA_MSK bit to make sure hardware knows we
1496 * aren't meddling with the INTENA_M bit.
1497 */
1498static void ice_set_wb_on_itr(struct ice_q_vector *q_vector)
1499{
1500	struct ice_vsi *vsi = q_vector->vsi;
1501
1502	/* already in wb_on_itr mode no need to change it */
1503	if (q_vector->wb_on_itr)
1504		return;
1505
1506	/* use previously set ITR values for all of the ITR indices by
1507	 * specifying ICE_ITR_NONE, which will vary in adaptive (AIM) mode and
1508	 * be static in non-adaptive mode (user configured)
1509	 */
1510	wr32(&vsi->back->hw, GLINT_DYN_CTL(q_vector->reg_idx),
1511	     FIELD_PREP(GLINT_DYN_CTL_ITR_INDX_M, ICE_ITR_NONE) |
1512	     FIELD_PREP(GLINT_DYN_CTL_INTENA_MSK_M, 1) |
1513	     FIELD_PREP(GLINT_DYN_CTL_WB_ON_ITR_M, 1));
1514
1515	q_vector->wb_on_itr = true;
1516}
1517
1518/**
1519 * ice_napi_poll - NAPI polling Rx/Tx cleanup routine
1520 * @napi: napi struct with our devices info in it
1521 * @budget: amount of work driver is allowed to do this pass, in packets
1522 *
1523 * This function will clean all queues associated with a q_vector.
1524 *
1525 * Returns the amount of work done
1526 */
1527int ice_napi_poll(struct napi_struct *napi, int budget)
1528{
1529	struct ice_q_vector *q_vector =
1530				container_of(napi, struct ice_q_vector, napi);
1531	struct ice_tx_ring *tx_ring;
1532	struct ice_rx_ring *rx_ring;
1533	bool clean_complete = true;
1534	int budget_per_ring;
1535	int work_done = 0;
1536
1537	/* Since the actual Tx work is minimal, we can give the Tx a larger
1538	 * budget and be more aggressive about cleaning up the Tx descriptors.
1539	 */
1540	ice_for_each_tx_ring(tx_ring, q_vector->tx) {
1541		struct xsk_buff_pool *xsk_pool = READ_ONCE(tx_ring->xsk_pool);
1542		bool wd;
1543
1544		if (xsk_pool)
1545			wd = ice_xmit_zc(tx_ring, xsk_pool);
1546		else if (ice_ring_is_xdp(tx_ring))
1547			wd = true;
1548		else
1549			wd = ice_clean_tx_irq(tx_ring, budget);
1550
1551		if (!wd)
1552			clean_complete = false;
1553	}
1554
1555	/* Handle case where we are called by netpoll with a budget of 0 */
1556	if (unlikely(budget <= 0))
1557		return budget;
1558
1559	/* normally we have 1 Rx ring per q_vector */
1560	if (unlikely(q_vector->num_ring_rx > 1))
1561		/* We attempt to distribute budget to each Rx queue fairly, but
1562		 * don't allow the budget to go below 1 because that would exit
1563		 * polling early.
1564		 */
1565		budget_per_ring = max_t(int, budget / q_vector->num_ring_rx, 1);
1566	else
1567		/* Max of 1 Rx ring in this q_vector so give it the budget */
1568		budget_per_ring = budget;
1569
1570	ice_for_each_rx_ring(rx_ring, q_vector->rx) {
1571		struct xsk_buff_pool *xsk_pool = READ_ONCE(rx_ring->xsk_pool);
1572		int cleaned;
1573
1574		/* A dedicated path for zero-copy allows making a single
1575		 * comparison in the irq context instead of many inside the
1576		 * ice_clean_rx_irq function and makes the codebase cleaner.
1577		 */
1578		cleaned = rx_ring->xsk_pool ?
1579			  ice_clean_rx_irq_zc(rx_ring, xsk_pool, budget_per_ring) :
1580			  ice_clean_rx_irq(rx_ring, budget_per_ring);
1581		work_done += cleaned;
1582		/* if we clean as many as budgeted, we must not be done */
1583		if (cleaned >= budget_per_ring)
1584			clean_complete = false;
1585	}
1586
1587	/* If work not completed, return budget and polling will return */
1588	if (!clean_complete) {
1589		/* Set the writeback on ITR so partial completions of
1590		 * cache-lines will still continue even if we're polling.
1591		 */
1592		ice_set_wb_on_itr(q_vector);
1593		return budget;
1594	}
1595
1596	/* Exit the polling mode, but don't re-enable interrupts if stack might
1597	 * poll us due to busy-polling
1598	 */
1599	if (napi_complete_done(napi, work_done)) {
1600		ice_net_dim(q_vector);
1601		ice_enable_interrupt(q_vector);
1602	} else {
1603		ice_set_wb_on_itr(q_vector);
1604	}
1605
1606	return min_t(int, work_done, budget - 1);
1607}
1608
1609/**
1610 * __ice_maybe_stop_tx - 2nd level check for Tx stop conditions
1611 * @tx_ring: the ring to be checked
1612 * @size: the size buffer we want to assure is available
1613 *
1614 * Returns -EBUSY if a stop is needed, else 0
1615 */
1616static int __ice_maybe_stop_tx(struct ice_tx_ring *tx_ring, unsigned int size)
1617{
1618	netif_tx_stop_queue(txring_txq(tx_ring));
1619	/* Memory barrier before checking head and tail */
1620	smp_mb();
1621
1622	/* Check again in a case another CPU has just made room available. */
1623	if (likely(ICE_DESC_UNUSED(tx_ring) < size))
1624		return -EBUSY;
1625
1626	/* A reprieve! - use start_queue because it doesn't call schedule */
1627	netif_tx_start_queue(txring_txq(tx_ring));
1628	++tx_ring->ring_stats->tx_stats.restart_q;
1629	return 0;
1630}
1631
1632/**
1633 * ice_maybe_stop_tx - 1st level check for Tx stop conditions
1634 * @tx_ring: the ring to be checked
1635 * @size:    the size buffer we want to assure is available
1636 *
1637 * Returns 0 if stop is not needed
1638 */
1639static int ice_maybe_stop_tx(struct ice_tx_ring *tx_ring, unsigned int size)
1640{
1641	if (likely(ICE_DESC_UNUSED(tx_ring) >= size))
1642		return 0;
1643
1644	return __ice_maybe_stop_tx(tx_ring, size);
1645}
1646
1647/**
1648 * ice_tx_map - Build the Tx descriptor
1649 * @tx_ring: ring to send buffer on
1650 * @first: first buffer info buffer to use
1651 * @off: pointer to struct that holds offload parameters
1652 *
1653 * This function loops over the skb data pointed to by *first
1654 * and gets a physical address for each memory location and programs
1655 * it and the length into the transmit descriptor.
1656 */
1657static void
1658ice_tx_map(struct ice_tx_ring *tx_ring, struct ice_tx_buf *first,
1659	   struct ice_tx_offload_params *off)
1660{
1661	u64 td_offset, td_tag, td_cmd;
1662	u16 i = tx_ring->next_to_use;
1663	unsigned int data_len, size;
1664	struct ice_tx_desc *tx_desc;
1665	struct ice_tx_buf *tx_buf;
1666	struct sk_buff *skb;
1667	skb_frag_t *frag;
1668	dma_addr_t dma;
1669	bool kick;
1670
1671	td_tag = off->td_l2tag1;
1672	td_cmd = off->td_cmd;
1673	td_offset = off->td_offset;
1674	skb = first->skb;
1675
1676	data_len = skb->data_len;
1677	size = skb_headlen(skb);
1678
1679	tx_desc = ICE_TX_DESC(tx_ring, i);
1680
1681	if (first->tx_flags & ICE_TX_FLAGS_HW_VLAN) {
1682		td_cmd |= (u64)ICE_TX_DESC_CMD_IL2TAG1;
1683		td_tag = first->vid;
 
1684	}
1685
1686	dma = dma_map_single(tx_ring->dev, skb->data, size, DMA_TO_DEVICE);
1687
1688	tx_buf = first;
1689
1690	for (frag = &skb_shinfo(skb)->frags[0];; frag++) {
1691		unsigned int max_data = ICE_MAX_DATA_PER_TXD_ALIGNED;
1692
1693		if (dma_mapping_error(tx_ring->dev, dma))
1694			goto dma_error;
1695
1696		/* record length, and DMA address */
1697		dma_unmap_len_set(tx_buf, len, size);
1698		dma_unmap_addr_set(tx_buf, dma, dma);
1699
1700		/* align size to end of page */
1701		max_data += -dma & (ICE_MAX_READ_REQ_SIZE - 1);
1702		tx_desc->buf_addr = cpu_to_le64(dma);
1703
1704		/* account for data chunks larger than the hardware
1705		 * can handle
1706		 */
1707		while (unlikely(size > ICE_MAX_DATA_PER_TXD)) {
1708			tx_desc->cmd_type_offset_bsz =
1709				ice_build_ctob(td_cmd, td_offset, max_data,
1710					       td_tag);
1711
1712			tx_desc++;
1713			i++;
1714
1715			if (i == tx_ring->count) {
1716				tx_desc = ICE_TX_DESC(tx_ring, 0);
1717				i = 0;
1718			}
1719
1720			dma += max_data;
1721			size -= max_data;
1722
1723			max_data = ICE_MAX_DATA_PER_TXD_ALIGNED;
1724			tx_desc->buf_addr = cpu_to_le64(dma);
1725		}
1726
1727		if (likely(!data_len))
1728			break;
1729
1730		tx_desc->cmd_type_offset_bsz = ice_build_ctob(td_cmd, td_offset,
1731							      size, td_tag);
1732
1733		tx_desc++;
1734		i++;
1735
1736		if (i == tx_ring->count) {
1737			tx_desc = ICE_TX_DESC(tx_ring, 0);
1738			i = 0;
1739		}
1740
1741		size = skb_frag_size(frag);
1742		data_len -= size;
1743
1744		dma = skb_frag_dma_map(tx_ring->dev, frag, 0, size,
1745				       DMA_TO_DEVICE);
1746
1747		tx_buf = &tx_ring->tx_buf[i];
1748		tx_buf->type = ICE_TX_BUF_FRAG;
1749	}
1750
1751	/* record SW timestamp if HW timestamp is not available */
1752	skb_tx_timestamp(first->skb);
1753
1754	i++;
1755	if (i == tx_ring->count)
1756		i = 0;
1757
1758	/* write last descriptor with RS and EOP bits */
1759	td_cmd |= (u64)ICE_TXD_LAST_DESC_CMD;
1760	tx_desc->cmd_type_offset_bsz =
1761			ice_build_ctob(td_cmd, td_offset, size, td_tag);
1762
1763	/* Force memory writes to complete before letting h/w know there
1764	 * are new descriptors to fetch.
1765	 *
1766	 * We also use this memory barrier to make certain all of the
1767	 * status bits have been updated before next_to_watch is written.
1768	 */
1769	wmb();
1770
1771	/* set next_to_watch value indicating a packet is present */
1772	first->next_to_watch = tx_desc;
1773
1774	tx_ring->next_to_use = i;
1775
1776	ice_maybe_stop_tx(tx_ring, DESC_NEEDED);
1777
1778	/* notify HW of packet */
1779	kick = __netdev_tx_sent_queue(txring_txq(tx_ring), first->bytecount,
1780				      netdev_xmit_more());
1781	if (kick)
1782		/* notify HW of packet */
1783		writel(i, tx_ring->tail);
1784
1785	return;
1786
1787dma_error:
1788	/* clear DMA mappings for failed tx_buf map */
1789	for (;;) {
1790		tx_buf = &tx_ring->tx_buf[i];
1791		ice_unmap_and_free_tx_buf(tx_ring, tx_buf);
1792		if (tx_buf == first)
1793			break;
1794		if (i == 0)
1795			i = tx_ring->count;
1796		i--;
1797	}
1798
1799	tx_ring->next_to_use = i;
1800}
1801
1802/**
1803 * ice_tx_csum - Enable Tx checksum offloads
1804 * @first: pointer to the first descriptor
1805 * @off: pointer to struct that holds offload parameters
1806 *
1807 * Returns 0 or error (negative) if checksum offload can't happen, 1 otherwise.
1808 */
1809static
1810int ice_tx_csum(struct ice_tx_buf *first, struct ice_tx_offload_params *off)
1811{
1812	u32 l4_len = 0, l3_len = 0, l2_len = 0;
1813	struct sk_buff *skb = first->skb;
1814	union {
1815		struct iphdr *v4;
1816		struct ipv6hdr *v6;
1817		unsigned char *hdr;
1818	} ip;
1819	union {
1820		struct tcphdr *tcp;
1821		unsigned char *hdr;
1822	} l4;
1823	__be16 frag_off, protocol;
1824	unsigned char *exthdr;
1825	u32 offset, cmd = 0;
1826	u8 l4_proto = 0;
1827
1828	if (skb->ip_summed != CHECKSUM_PARTIAL)
1829		return 0;
1830
1831	protocol = vlan_get_protocol(skb);
1832
1833	if (eth_p_mpls(protocol)) {
1834		ip.hdr = skb_inner_network_header(skb);
1835		l4.hdr = skb_checksum_start(skb);
1836	} else {
1837		ip.hdr = skb_network_header(skb);
1838		l4.hdr = skb_transport_header(skb);
1839	}
1840
1841	/* compute outer L2 header size */
1842	l2_len = ip.hdr - skb->data;
1843	offset = (l2_len / 2) << ICE_TX_DESC_LEN_MACLEN_S;
1844
1845	/* set the tx_flags to indicate the IP protocol type. this is
1846	 * required so that checksum header computation below is accurate.
1847	 */
1848	if (ip.v4->version == 4)
1849		first->tx_flags |= ICE_TX_FLAGS_IPV4;
1850	else if (ip.v6->version == 6)
1851		first->tx_flags |= ICE_TX_FLAGS_IPV6;
1852
1853	if (skb->encapsulation) {
1854		bool gso_ena = false;
1855		u32 tunnel = 0;
1856
1857		/* define outer network header type */
1858		if (first->tx_flags & ICE_TX_FLAGS_IPV4) {
1859			tunnel |= (first->tx_flags & ICE_TX_FLAGS_TSO) ?
1860				  ICE_TX_CTX_EIPT_IPV4 :
1861				  ICE_TX_CTX_EIPT_IPV4_NO_CSUM;
1862			l4_proto = ip.v4->protocol;
1863		} else if (first->tx_flags & ICE_TX_FLAGS_IPV6) {
1864			int ret;
1865
1866			tunnel |= ICE_TX_CTX_EIPT_IPV6;
1867			exthdr = ip.hdr + sizeof(*ip.v6);
1868			l4_proto = ip.v6->nexthdr;
1869			ret = ipv6_skip_exthdr(skb, exthdr - skb->data,
1870					       &l4_proto, &frag_off);
1871			if (ret < 0)
1872				return -1;
1873		}
1874
1875		/* define outer transport */
1876		switch (l4_proto) {
1877		case IPPROTO_UDP:
1878			tunnel |= ICE_TXD_CTX_UDP_TUNNELING;
1879			first->tx_flags |= ICE_TX_FLAGS_TUNNEL;
1880			break;
1881		case IPPROTO_GRE:
1882			tunnel |= ICE_TXD_CTX_GRE_TUNNELING;
1883			first->tx_flags |= ICE_TX_FLAGS_TUNNEL;
1884			break;
1885		case IPPROTO_IPIP:
1886		case IPPROTO_IPV6:
1887			first->tx_flags |= ICE_TX_FLAGS_TUNNEL;
1888			l4.hdr = skb_inner_network_header(skb);
1889			break;
1890		default:
1891			if (first->tx_flags & ICE_TX_FLAGS_TSO)
1892				return -1;
1893
1894			skb_checksum_help(skb);
1895			return 0;
1896		}
1897
1898		/* compute outer L3 header size */
1899		tunnel |= ((l4.hdr - ip.hdr) / 4) <<
1900			  ICE_TXD_CTX_QW0_EIPLEN_S;
1901
1902		/* switch IP header pointer from outer to inner header */
1903		ip.hdr = skb_inner_network_header(skb);
1904
1905		/* compute tunnel header size */
1906		tunnel |= ((ip.hdr - l4.hdr) / 2) <<
1907			   ICE_TXD_CTX_QW0_NATLEN_S;
1908
1909		gso_ena = skb_shinfo(skb)->gso_type & SKB_GSO_PARTIAL;
1910		/* indicate if we need to offload outer UDP header */
1911		if ((first->tx_flags & ICE_TX_FLAGS_TSO) && !gso_ena &&
1912		    (skb_shinfo(skb)->gso_type & SKB_GSO_UDP_TUNNEL_CSUM))
1913			tunnel |= ICE_TXD_CTX_QW0_L4T_CS_M;
1914
1915		/* record tunnel offload values */
1916		off->cd_tunnel_params |= tunnel;
1917
1918		/* set DTYP=1 to indicate that it's an Tx context descriptor
1919		 * in IPsec tunnel mode with Tx offloads in Quad word 1
1920		 */
1921		off->cd_qw1 |= (u64)ICE_TX_DESC_DTYPE_CTX;
1922
1923		/* switch L4 header pointer from outer to inner */
1924		l4.hdr = skb_inner_transport_header(skb);
1925		l4_proto = 0;
1926
1927		/* reset type as we transition from outer to inner headers */
1928		first->tx_flags &= ~(ICE_TX_FLAGS_IPV4 | ICE_TX_FLAGS_IPV6);
1929		if (ip.v4->version == 4)
1930			first->tx_flags |= ICE_TX_FLAGS_IPV4;
1931		if (ip.v6->version == 6)
1932			first->tx_flags |= ICE_TX_FLAGS_IPV6;
1933	}
1934
1935	/* Enable IP checksum offloads */
1936	if (first->tx_flags & ICE_TX_FLAGS_IPV4) {
1937		l4_proto = ip.v4->protocol;
1938		/* the stack computes the IP header already, the only time we
1939		 * need the hardware to recompute it is in the case of TSO.
1940		 */
1941		if (first->tx_flags & ICE_TX_FLAGS_TSO)
1942			cmd |= ICE_TX_DESC_CMD_IIPT_IPV4_CSUM;
1943		else
1944			cmd |= ICE_TX_DESC_CMD_IIPT_IPV4;
1945
1946	} else if (first->tx_flags & ICE_TX_FLAGS_IPV6) {
1947		cmd |= ICE_TX_DESC_CMD_IIPT_IPV6;
1948		exthdr = ip.hdr + sizeof(*ip.v6);
1949		l4_proto = ip.v6->nexthdr;
1950		if (l4.hdr != exthdr)
1951			ipv6_skip_exthdr(skb, exthdr - skb->data, &l4_proto,
1952					 &frag_off);
1953	} else {
1954		return -1;
1955	}
1956
1957	/* compute inner L3 header size */
1958	l3_len = l4.hdr - ip.hdr;
1959	offset |= (l3_len / 4) << ICE_TX_DESC_LEN_IPLEN_S;
1960
1961	/* Enable L4 checksum offloads */
1962	switch (l4_proto) {
1963	case IPPROTO_TCP:
1964		/* enable checksum offloads */
1965		cmd |= ICE_TX_DESC_CMD_L4T_EOFT_TCP;
1966		l4_len = l4.tcp->doff;
1967		offset |= l4_len << ICE_TX_DESC_LEN_L4_LEN_S;
1968		break;
1969	case IPPROTO_UDP:
1970		/* enable UDP checksum offload */
1971		cmd |= ICE_TX_DESC_CMD_L4T_EOFT_UDP;
1972		l4_len = (sizeof(struct udphdr) >> 2);
1973		offset |= l4_len << ICE_TX_DESC_LEN_L4_LEN_S;
1974		break;
1975	case IPPROTO_SCTP:
1976		/* enable SCTP checksum offload */
1977		cmd |= ICE_TX_DESC_CMD_L4T_EOFT_SCTP;
1978		l4_len = sizeof(struct sctphdr) >> 2;
1979		offset |= l4_len << ICE_TX_DESC_LEN_L4_LEN_S;
1980		break;
1981
1982	default:
1983		if (first->tx_flags & ICE_TX_FLAGS_TSO)
1984			return -1;
1985		skb_checksum_help(skb);
1986		return 0;
1987	}
1988
1989	off->td_cmd |= cmd;
1990	off->td_offset |= offset;
1991	return 1;
1992}
1993
1994/**
1995 * ice_tx_prepare_vlan_flags - prepare generic Tx VLAN tagging flags for HW
1996 * @tx_ring: ring to send buffer on
1997 * @first: pointer to struct ice_tx_buf
1998 *
1999 * Checks the skb and set up correspondingly several generic transmit flags
2000 * related to VLAN tagging for the HW, such as VLAN, DCB, etc.
2001 */
2002static void
2003ice_tx_prepare_vlan_flags(struct ice_tx_ring *tx_ring, struct ice_tx_buf *first)
2004{
2005	struct sk_buff *skb = first->skb;
2006
2007	/* nothing left to do, software offloaded VLAN */
2008	if (!skb_vlan_tag_present(skb) && eth_type_vlan(skb->protocol))
2009		return;
2010
2011	/* the VLAN ethertype/tpid is determined by VSI configuration and netdev
2012	 * feature flags, which the driver only allows either 802.1Q or 802.1ad
2013	 * VLAN offloads exclusively so we only care about the VLAN ID here
2014	 */
2015	if (skb_vlan_tag_present(skb)) {
2016		first->vid = skb_vlan_tag_get(skb);
2017		if (tx_ring->flags & ICE_TX_FLAGS_RING_VLAN_L2TAG2)
2018			first->tx_flags |= ICE_TX_FLAGS_HW_OUTER_SINGLE_VLAN;
2019		else
2020			first->tx_flags |= ICE_TX_FLAGS_HW_VLAN;
2021	}
2022
2023	ice_tx_prepare_vlan_flags_dcb(tx_ring, first);
2024}
2025
2026/**
2027 * ice_tso - computes mss and TSO length to prepare for TSO
2028 * @first: pointer to struct ice_tx_buf
2029 * @off: pointer to struct that holds offload parameters
2030 *
2031 * Returns 0 or error (negative) if TSO can't happen, 1 otherwise.
2032 */
2033static
2034int ice_tso(struct ice_tx_buf *first, struct ice_tx_offload_params *off)
2035{
2036	struct sk_buff *skb = first->skb;
2037	union {
2038		struct iphdr *v4;
2039		struct ipv6hdr *v6;
2040		unsigned char *hdr;
2041	} ip;
2042	union {
2043		struct tcphdr *tcp;
2044		struct udphdr *udp;
2045		unsigned char *hdr;
2046	} l4;
2047	u64 cd_mss, cd_tso_len;
2048	__be16 protocol;
2049	u32 paylen;
2050	u8 l4_start;
2051	int err;
2052
2053	if (skb->ip_summed != CHECKSUM_PARTIAL)
2054		return 0;
2055
2056	if (!skb_is_gso(skb))
2057		return 0;
2058
2059	err = skb_cow_head(skb, 0);
2060	if (err < 0)
2061		return err;
2062
 
2063	protocol = vlan_get_protocol(skb);
2064
2065	if (eth_p_mpls(protocol))
2066		ip.hdr = skb_inner_network_header(skb);
2067	else
2068		ip.hdr = skb_network_header(skb);
2069	l4.hdr = skb_checksum_start(skb);
2070
2071	/* initialize outer IP header fields */
2072	if (ip.v4->version == 4) {
2073		ip.v4->tot_len = 0;
2074		ip.v4->check = 0;
2075	} else {
2076		ip.v6->payload_len = 0;
2077	}
2078
2079	if (skb_shinfo(skb)->gso_type & (SKB_GSO_GRE |
2080					 SKB_GSO_GRE_CSUM |
2081					 SKB_GSO_IPXIP4 |
2082					 SKB_GSO_IPXIP6 |
2083					 SKB_GSO_UDP_TUNNEL |
2084					 SKB_GSO_UDP_TUNNEL_CSUM)) {
2085		if (!(skb_shinfo(skb)->gso_type & SKB_GSO_PARTIAL) &&
2086		    (skb_shinfo(skb)->gso_type & SKB_GSO_UDP_TUNNEL_CSUM)) {
2087			l4.udp->len = 0;
2088
2089			/* determine offset of outer transport header */
2090			l4_start = (u8)(l4.hdr - skb->data);
2091
2092			/* remove payload length from outer checksum */
2093			paylen = skb->len - l4_start;
2094			csum_replace_by_diff(&l4.udp->check,
2095					     (__force __wsum)htonl(paylen));
2096		}
2097
2098		/* reset pointers to inner headers */
 
 
2099		ip.hdr = skb_inner_network_header(skb);
2100		l4.hdr = skb_inner_transport_header(skb);
2101
2102		/* initialize inner IP header fields */
2103		if (ip.v4->version == 4) {
2104			ip.v4->tot_len = 0;
2105			ip.v4->check = 0;
2106		} else {
2107			ip.v6->payload_len = 0;
2108		}
2109	}
2110
2111	/* determine offset of transport header */
2112	l4_start = (u8)(l4.hdr - skb->data);
2113
2114	/* remove payload length from checksum */
2115	paylen = skb->len - l4_start;
2116
2117	if (skb_shinfo(skb)->gso_type & SKB_GSO_UDP_L4) {
2118		csum_replace_by_diff(&l4.udp->check,
2119				     (__force __wsum)htonl(paylen));
2120		/* compute length of UDP segmentation header */
2121		off->header_len = (u8)sizeof(l4.udp) + l4_start;
2122	} else {
2123		csum_replace_by_diff(&l4.tcp->check,
2124				     (__force __wsum)htonl(paylen));
2125		/* compute length of TCP segmentation header */
2126		off->header_len = (u8)((l4.tcp->doff * 4) + l4_start);
2127	}
2128
2129	/* update gso_segs and bytecount */
2130	first->gso_segs = skb_shinfo(skb)->gso_segs;
2131	first->bytecount += (first->gso_segs - 1) * off->header_len;
2132
2133	cd_tso_len = skb->len - off->header_len;
2134	cd_mss = skb_shinfo(skb)->gso_size;
2135
2136	/* record cdesc_qw1 with TSO parameters */
2137	off->cd_qw1 |= (u64)(ICE_TX_DESC_DTYPE_CTX |
2138			     (ICE_TX_CTX_DESC_TSO << ICE_TXD_CTX_QW1_CMD_S) |
2139			     (cd_tso_len << ICE_TXD_CTX_QW1_TSO_LEN_S) |
2140			     (cd_mss << ICE_TXD_CTX_QW1_MSS_S));
2141	first->tx_flags |= ICE_TX_FLAGS_TSO;
2142	return 1;
2143}
2144
2145/**
2146 * ice_txd_use_count  - estimate the number of descriptors needed for Tx
2147 * @size: transmit request size in bytes
2148 *
2149 * Due to hardware alignment restrictions (4K alignment), we need to
2150 * assume that we can have no more than 12K of data per descriptor, even
2151 * though each descriptor can take up to 16K - 1 bytes of aligned memory.
2152 * Thus, we need to divide by 12K. But division is slow! Instead,
2153 * we decompose the operation into shifts and one relatively cheap
2154 * multiply operation.
2155 *
2156 * To divide by 12K, we first divide by 4K, then divide by 3:
2157 *     To divide by 4K, shift right by 12 bits
2158 *     To divide by 3, multiply by 85, then divide by 256
2159 *     (Divide by 256 is done by shifting right by 8 bits)
2160 * Finally, we add one to round up. Because 256 isn't an exact multiple of
2161 * 3, we'll underestimate near each multiple of 12K. This is actually more
2162 * accurate as we have 4K - 1 of wiggle room that we can fit into the last
2163 * segment. For our purposes this is accurate out to 1M which is orders of
2164 * magnitude greater than our largest possible GSO size.
2165 *
2166 * This would then be implemented as:
2167 *     return (((size >> 12) * 85) >> 8) + ICE_DESCS_FOR_SKB_DATA_PTR;
2168 *
2169 * Since multiplication and division are commutative, we can reorder
2170 * operations into:
2171 *     return ((size * 85) >> 20) + ICE_DESCS_FOR_SKB_DATA_PTR;
2172 */
2173static unsigned int ice_txd_use_count(unsigned int size)
2174{
2175	return ((size * 85) >> 20) + ICE_DESCS_FOR_SKB_DATA_PTR;
2176}
2177
2178/**
2179 * ice_xmit_desc_count - calculate number of Tx descriptors needed
2180 * @skb: send buffer
2181 *
2182 * Returns number of data descriptors needed for this skb.
2183 */
2184static unsigned int ice_xmit_desc_count(struct sk_buff *skb)
2185{
2186	const skb_frag_t *frag = &skb_shinfo(skb)->frags[0];
2187	unsigned int nr_frags = skb_shinfo(skb)->nr_frags;
2188	unsigned int count = 0, size = skb_headlen(skb);
2189
2190	for (;;) {
2191		count += ice_txd_use_count(size);
2192
2193		if (!nr_frags--)
2194			break;
2195
2196		size = skb_frag_size(frag++);
2197	}
2198
2199	return count;
2200}
2201
2202/**
2203 * __ice_chk_linearize - Check if there are more than 8 buffers per packet
2204 * @skb: send buffer
2205 *
2206 * Note: This HW can't DMA more than 8 buffers to build a packet on the wire
2207 * and so we need to figure out the cases where we need to linearize the skb.
2208 *
2209 * For TSO we need to count the TSO header and segment payload separately.
2210 * As such we need to check cases where we have 7 fragments or more as we
2211 * can potentially require 9 DMA transactions, 1 for the TSO header, 1 for
2212 * the segment payload in the first descriptor, and another 7 for the
2213 * fragments.
2214 */
2215static bool __ice_chk_linearize(struct sk_buff *skb)
2216{
2217	const skb_frag_t *frag, *stale;
2218	int nr_frags, sum;
2219
2220	/* no need to check if number of frags is less than 7 */
2221	nr_frags = skb_shinfo(skb)->nr_frags;
2222	if (nr_frags < (ICE_MAX_BUF_TXD - 1))
2223		return false;
2224
2225	/* We need to walk through the list and validate that each group
2226	 * of 6 fragments totals at least gso_size.
2227	 */
2228	nr_frags -= ICE_MAX_BUF_TXD - 2;
2229	frag = &skb_shinfo(skb)->frags[0];
2230
2231	/* Initialize size to the negative value of gso_size minus 1. We
2232	 * use this as the worst case scenario in which the frag ahead
2233	 * of us only provides one byte which is why we are limited to 6
2234	 * descriptors for a single transmit as the header and previous
2235	 * fragment are already consuming 2 descriptors.
2236	 */
2237	sum = 1 - skb_shinfo(skb)->gso_size;
2238
2239	/* Add size of frags 0 through 4 to create our initial sum */
2240	sum += skb_frag_size(frag++);
2241	sum += skb_frag_size(frag++);
2242	sum += skb_frag_size(frag++);
2243	sum += skb_frag_size(frag++);
2244	sum += skb_frag_size(frag++);
2245
2246	/* Walk through fragments adding latest fragment, testing it, and
2247	 * then removing stale fragments from the sum.
2248	 */
2249	for (stale = &skb_shinfo(skb)->frags[0];; stale++) {
2250		int stale_size = skb_frag_size(stale);
2251
2252		sum += skb_frag_size(frag++);
2253
2254		/* The stale fragment may present us with a smaller
2255		 * descriptor than the actual fragment size. To account
2256		 * for that we need to remove all the data on the front and
2257		 * figure out what the remainder would be in the last
2258		 * descriptor associated with the fragment.
2259		 */
2260		if (stale_size > ICE_MAX_DATA_PER_TXD) {
2261			int align_pad = -(skb_frag_off(stale)) &
2262					(ICE_MAX_READ_REQ_SIZE - 1);
2263
2264			sum -= align_pad;
2265			stale_size -= align_pad;
2266
2267			do {
2268				sum -= ICE_MAX_DATA_PER_TXD_ALIGNED;
2269				stale_size -= ICE_MAX_DATA_PER_TXD_ALIGNED;
2270			} while (stale_size > ICE_MAX_DATA_PER_TXD);
2271		}
2272
2273		/* if sum is negative we failed to make sufficient progress */
2274		if (sum < 0)
2275			return true;
2276
2277		if (!nr_frags--)
2278			break;
2279
2280		sum -= stale_size;
2281	}
2282
2283	return false;
2284}
2285
2286/**
2287 * ice_chk_linearize - Check if there are more than 8 fragments per packet
2288 * @skb:      send buffer
2289 * @count:    number of buffers used
2290 *
2291 * Note: Our HW can't scatter-gather more than 8 fragments to build
2292 * a packet on the wire and so we need to figure out the cases where we
2293 * need to linearize the skb.
2294 */
2295static bool ice_chk_linearize(struct sk_buff *skb, unsigned int count)
2296{
2297	/* Both TSO and single send will work if count is less than 8 */
2298	if (likely(count < ICE_MAX_BUF_TXD))
2299		return false;
2300
2301	if (skb_is_gso(skb))
2302		return __ice_chk_linearize(skb);
2303
2304	/* we can support up to 8 data buffers for a single send */
2305	return count != ICE_MAX_BUF_TXD;
2306}
2307
2308/**
2309 * ice_tstamp - set up context descriptor for hardware timestamp
2310 * @tx_ring: pointer to the Tx ring to send buffer on
2311 * @skb: pointer to the SKB we're sending
2312 * @first: Tx buffer
2313 * @off: Tx offload parameters
2314 */
2315static void
2316ice_tstamp(struct ice_tx_ring *tx_ring, struct sk_buff *skb,
2317	   struct ice_tx_buf *first, struct ice_tx_offload_params *off)
2318{
2319	s8 idx;
2320
2321	/* only timestamp the outbound packet if the user has requested it */
2322	if (likely(!(skb_shinfo(skb)->tx_flags & SKBTX_HW_TSTAMP)))
2323		return;
2324
 
 
 
2325	/* Tx timestamps cannot be sampled when doing TSO */
2326	if (first->tx_flags & ICE_TX_FLAGS_TSO)
2327		return;
2328
2329	/* Grab an open timestamp slot */
2330	idx = ice_ptp_request_ts(tx_ring->tx_tstamps, skb);
2331	if (idx < 0) {
2332		tx_ring->vsi->back->ptp.tx_hwtstamp_skipped++;
2333		return;
2334	}
2335
2336	off->cd_qw1 |= (u64)(ICE_TX_DESC_DTYPE_CTX |
2337			     (ICE_TX_CTX_DESC_TSYN << ICE_TXD_CTX_QW1_CMD_S) |
2338			     ((u64)idx << ICE_TXD_CTX_QW1_TSO_LEN_S));
2339	first->tx_flags |= ICE_TX_FLAGS_TSYN;
2340}
2341
2342/**
2343 * ice_xmit_frame_ring - Sends buffer on Tx ring
2344 * @skb: send buffer
2345 * @tx_ring: ring to send buffer on
2346 *
2347 * Returns NETDEV_TX_OK if sent, else an error code
2348 */
2349static netdev_tx_t
2350ice_xmit_frame_ring(struct sk_buff *skb, struct ice_tx_ring *tx_ring)
2351{
2352	struct ice_tx_offload_params offload = { 0 };
2353	struct ice_vsi *vsi = tx_ring->vsi;
2354	struct ice_tx_buf *first;
2355	struct ethhdr *eth;
2356	unsigned int count;
2357	int tso, csum;
2358
2359	ice_trace(xmit_frame_ring, tx_ring, skb);
2360
2361	if (unlikely(ipv6_hopopt_jumbo_remove(skb)))
2362		goto out_drop;
2363
2364	count = ice_xmit_desc_count(skb);
2365	if (ice_chk_linearize(skb, count)) {
2366		if (__skb_linearize(skb))
2367			goto out_drop;
2368		count = ice_txd_use_count(skb->len);
2369		tx_ring->ring_stats->tx_stats.tx_linearize++;
2370	}
2371
2372	/* need: 1 descriptor per page * PAGE_SIZE/ICE_MAX_DATA_PER_TXD,
2373	 *       + 1 desc for skb_head_len/ICE_MAX_DATA_PER_TXD,
2374	 *       + 4 desc gap to avoid the cache line where head is,
2375	 *       + 1 desc for context descriptor,
2376	 * otherwise try next time
2377	 */
2378	if (ice_maybe_stop_tx(tx_ring, count + ICE_DESCS_PER_CACHE_LINE +
2379			      ICE_DESCS_FOR_CTX_DESC)) {
2380		tx_ring->ring_stats->tx_stats.tx_busy++;
2381		return NETDEV_TX_BUSY;
2382	}
2383
2384	/* prefetch for bql data which is infrequently used */
2385	netdev_txq_bql_enqueue_prefetchw(txring_txq(tx_ring));
2386
2387	offload.tx_ring = tx_ring;
2388
2389	/* record the location of the first descriptor for this packet */
2390	first = &tx_ring->tx_buf[tx_ring->next_to_use];
2391	first->skb = skb;
2392	first->type = ICE_TX_BUF_SKB;
2393	first->bytecount = max_t(unsigned int, skb->len, ETH_ZLEN);
2394	first->gso_segs = 1;
2395	first->tx_flags = 0;
2396
2397	/* prepare the VLAN tagging flags for Tx */
2398	ice_tx_prepare_vlan_flags(tx_ring, first);
2399	if (first->tx_flags & ICE_TX_FLAGS_HW_OUTER_SINGLE_VLAN) {
2400		offload.cd_qw1 |= (u64)(ICE_TX_DESC_DTYPE_CTX |
2401					(ICE_TX_CTX_DESC_IL2TAG2 <<
2402					ICE_TXD_CTX_QW1_CMD_S));
2403		offload.cd_l2tag2 = first->vid;
 
2404	}
2405
2406	/* set up TSO offload */
2407	tso = ice_tso(first, &offload);
2408	if (tso < 0)
2409		goto out_drop;
2410
2411	/* always set up Tx checksum offload */
2412	csum = ice_tx_csum(first, &offload);
2413	if (csum < 0)
2414		goto out_drop;
2415
2416	/* allow CONTROL frames egress from main VSI if FW LLDP disabled */
2417	eth = (struct ethhdr *)skb_mac_header(skb);
2418	if (unlikely((skb->priority == TC_PRIO_CONTROL ||
2419		      eth->h_proto == htons(ETH_P_LLDP)) &&
2420		     vsi->type == ICE_VSI_PF &&
2421		     vsi->port_info->qos_cfg.is_sw_lldp))
2422		offload.cd_qw1 |= (u64)(ICE_TX_DESC_DTYPE_CTX |
2423					ICE_TX_CTX_DESC_SWTCH_UPLINK <<
2424					ICE_TXD_CTX_QW1_CMD_S);
2425
2426	ice_tstamp(tx_ring, skb, first, &offload);
2427	if (ice_is_switchdev_running(vsi->back) && vsi->type != ICE_VSI_SF)
2428		ice_eswitch_set_target_vsi(skb, &offload);
2429
2430	if (offload.cd_qw1 & ICE_TX_DESC_DTYPE_CTX) {
2431		struct ice_tx_ctx_desc *cdesc;
2432		u16 i = tx_ring->next_to_use;
2433
2434		/* grab the next descriptor */
2435		cdesc = ICE_TX_CTX_DESC(tx_ring, i);
2436		i++;
2437		tx_ring->next_to_use = (i < tx_ring->count) ? i : 0;
2438
2439		/* setup context descriptor */
2440		cdesc->tunneling_params = cpu_to_le32(offload.cd_tunnel_params);
2441		cdesc->l2tag2 = cpu_to_le16(offload.cd_l2tag2);
2442		cdesc->rsvd = cpu_to_le16(0);
2443		cdesc->qw1 = cpu_to_le64(offload.cd_qw1);
2444	}
2445
2446	ice_tx_map(tx_ring, first, &offload);
2447	return NETDEV_TX_OK;
2448
2449out_drop:
2450	ice_trace(xmit_frame_ring_drop, tx_ring, skb);
2451	dev_kfree_skb_any(skb);
2452	return NETDEV_TX_OK;
2453}
2454
2455/**
2456 * ice_start_xmit - Selects the correct VSI and Tx queue to send buffer
2457 * @skb: send buffer
2458 * @netdev: network interface device structure
2459 *
2460 * Returns NETDEV_TX_OK if sent, else an error code
2461 */
2462netdev_tx_t ice_start_xmit(struct sk_buff *skb, struct net_device *netdev)
2463{
2464	struct ice_netdev_priv *np = netdev_priv(netdev);
2465	struct ice_vsi *vsi = np->vsi;
2466	struct ice_tx_ring *tx_ring;
2467
2468	tx_ring = vsi->tx_rings[skb->queue_mapping];
2469
2470	/* hardware can't handle really short frames, hardware padding works
2471	 * beyond this point
2472	 */
2473	if (skb_put_padto(skb, ICE_MIN_TX_LEN))
2474		return NETDEV_TX_OK;
2475
2476	return ice_xmit_frame_ring(skb, tx_ring);
2477}
2478
2479/**
2480 * ice_get_dscp_up - return the UP/TC value for a SKB
2481 * @dcbcfg: DCB config that contains DSCP to UP/TC mapping
2482 * @skb: SKB to query for info to determine UP/TC
2483 *
2484 * This function is to only be called when the PF is in L3 DSCP PFC mode
2485 */
2486static u8 ice_get_dscp_up(struct ice_dcbx_cfg *dcbcfg, struct sk_buff *skb)
2487{
2488	u8 dscp = 0;
2489
2490	if (skb->protocol == htons(ETH_P_IP))
2491		dscp = ipv4_get_dsfield(ip_hdr(skb)) >> 2;
2492	else if (skb->protocol == htons(ETH_P_IPV6))
2493		dscp = ipv6_get_dsfield(ipv6_hdr(skb)) >> 2;
2494
2495	return dcbcfg->dscp_map[dscp];
2496}
2497
2498u16
2499ice_select_queue(struct net_device *netdev, struct sk_buff *skb,
2500		 struct net_device *sb_dev)
2501{
2502	struct ice_pf *pf = ice_netdev_to_pf(netdev);
2503	struct ice_dcbx_cfg *dcbcfg;
2504
2505	dcbcfg = &pf->hw.port_info->qos_cfg.local_dcbx_cfg;
2506	if (dcbcfg->pfc_mode == ICE_QOS_MODE_DSCP)
2507		skb->priority = ice_get_dscp_up(dcbcfg, skb);
2508
2509	return netdev_pick_tx(netdev, skb, sb_dev);
2510}
2511
2512/**
2513 * ice_clean_ctrl_tx_irq - interrupt handler for flow director Tx queue
2514 * @tx_ring: tx_ring to clean
2515 */
2516void ice_clean_ctrl_tx_irq(struct ice_tx_ring *tx_ring)
2517{
2518	struct ice_vsi *vsi = tx_ring->vsi;
2519	s16 i = tx_ring->next_to_clean;
2520	int budget = ICE_DFLT_IRQ_WORK;
2521	struct ice_tx_desc *tx_desc;
2522	struct ice_tx_buf *tx_buf;
2523
2524	tx_buf = &tx_ring->tx_buf[i];
2525	tx_desc = ICE_TX_DESC(tx_ring, i);
2526	i -= tx_ring->count;
2527
2528	do {
2529		struct ice_tx_desc *eop_desc = tx_buf->next_to_watch;
2530
2531		/* if next_to_watch is not set then there is no pending work */
2532		if (!eop_desc)
2533			break;
2534
2535		/* prevent any other reads prior to eop_desc */
2536		smp_rmb();
2537
2538		/* if the descriptor isn't done, no work to do */
2539		if (!(eop_desc->cmd_type_offset_bsz &
2540		      cpu_to_le64(ICE_TX_DESC_DTYPE_DESC_DONE)))
2541			break;
2542
2543		/* clear next_to_watch to prevent false hangs */
2544		tx_buf->next_to_watch = NULL;
2545		tx_desc->buf_addr = 0;
2546		tx_desc->cmd_type_offset_bsz = 0;
2547
2548		/* move past filter desc */
2549		tx_buf++;
2550		tx_desc++;
2551		i++;
2552		if (unlikely(!i)) {
2553			i -= tx_ring->count;
2554			tx_buf = tx_ring->tx_buf;
2555			tx_desc = ICE_TX_DESC(tx_ring, 0);
2556		}
2557
2558		/* unmap the data header */
2559		if (dma_unmap_len(tx_buf, len))
2560			dma_unmap_single(tx_ring->dev,
2561					 dma_unmap_addr(tx_buf, dma),
2562					 dma_unmap_len(tx_buf, len),
2563					 DMA_TO_DEVICE);
2564		if (tx_buf->type == ICE_TX_BUF_DUMMY)
2565			devm_kfree(tx_ring->dev, tx_buf->raw_buf);
2566
2567		/* clear next_to_watch to prevent false hangs */
2568		tx_buf->type = ICE_TX_BUF_EMPTY;
2569		tx_buf->tx_flags = 0;
2570		tx_buf->next_to_watch = NULL;
2571		dma_unmap_len_set(tx_buf, len, 0);
2572		tx_desc->buf_addr = 0;
2573		tx_desc->cmd_type_offset_bsz = 0;
2574
2575		/* move past eop_desc for start of next FD desc */
2576		tx_buf++;
2577		tx_desc++;
2578		i++;
2579		if (unlikely(!i)) {
2580			i -= tx_ring->count;
2581			tx_buf = tx_ring->tx_buf;
2582			tx_desc = ICE_TX_DESC(tx_ring, 0);
2583		}
2584
2585		budget--;
2586	} while (likely(budget));
2587
2588	i += tx_ring->count;
2589	tx_ring->next_to_clean = i;
2590
2591	/* re-enable interrupt if needed */
2592	ice_irq_dynamic_ena(&vsi->back->hw, vsi, vsi->q_vectors[0]);
2593}

   1// SPDX-License-Identifier: GPL-2.0
   2/* Copyright (c) 2018, Intel Corporation. */
   3
   4/* The driver transmit and receive code */
   5
   6#include <linux/mm.h>
   7#include <linux/netdevice.h>
   8#include <linux/prefetch.h>
   9#include <linux/bpf_trace.h>
  10#include <net/dsfield.h>
  11#include <net/mpls.h>
  12#include <net/xdp.h>
  13#include "ice_txrx_lib.h"
  14#include "ice_lib.h"
  15#include "ice.h"
  16#include "ice_trace.h"
  17#include "ice_dcb_lib.h"
  18#include "ice_xsk.h"
  19#include "ice_eswitch.h"
  20
  21#define ICE_RX_HDR_SIZE		256
  22
  23#define FDIR_DESC_RXDID 0x40
  24#define ICE_FDIR_CLEAN_DELAY 10
  25
  26/**
  27 * ice_prgm_fdir_fltr - Program a Flow Director filter
  28 * @vsi: VSI to send dummy packet
  29 * @fdir_desc: flow director descriptor
  30 * @raw_packet: allocated buffer for flow director
  31 */
  32int
  33ice_prgm_fdir_fltr(struct ice_vsi *vsi, struct ice_fltr_desc *fdir_desc,
  34		   u8 *raw_packet)
  35{
  36	struct ice_tx_buf *tx_buf, *first;
  37	struct ice_fltr_desc *f_desc;
  38	struct ice_tx_desc *tx_desc;
  39	struct ice_tx_ring *tx_ring;
  40	struct device *dev;
  41	dma_addr_t dma;
  42	u32 td_cmd;
  43	u16 i;
  44
  45	/* VSI and Tx ring */
  46	if (!vsi)
  47		return -ENOENT;
  48	tx_ring = vsi->tx_rings[0];
  49	if (!tx_ring || !tx_ring->desc)
  50		return -ENOENT;
  51	dev = tx_ring->dev;
  52
  53	/* we are using two descriptors to add/del a filter and we can wait */
  54	for (i = ICE_FDIR_CLEAN_DELAY; ICE_DESC_UNUSED(tx_ring) < 2; i--) {
  55		if (!i)
  56			return -EAGAIN;
  57		msleep_interruptible(1);
  58	}
  59
  60	dma = dma_map_single(dev, raw_packet, ICE_FDIR_MAX_RAW_PKT_SIZE,
  61			     DMA_TO_DEVICE);
  62
  63	if (dma_mapping_error(dev, dma))
  64		return -EINVAL;
  65
  66	/* grab the next descriptor */
  67	i = tx_ring->next_to_use;
  68	first = &tx_ring->tx_buf[i];
  69	f_desc = ICE_TX_FDIRDESC(tx_ring, i);
  70	memcpy(f_desc, fdir_desc, sizeof(*f_desc));
  71
  72	i++;
  73	i = (i < tx_ring->count) ? i : 0;
  74	tx_desc = ICE_TX_DESC(tx_ring, i);
  75	tx_buf = &tx_ring->tx_buf[i];
  76
  77	i++;
  78	tx_ring->next_to_use = (i < tx_ring->count) ? i : 0;
  79
  80	memset(tx_buf, 0, sizeof(*tx_buf));
  81	dma_unmap_len_set(tx_buf, len, ICE_FDIR_MAX_RAW_PKT_SIZE);
  82	dma_unmap_addr_set(tx_buf, dma, dma);
  83
  84	tx_desc->buf_addr = cpu_to_le64(dma);
  85	td_cmd = ICE_TXD_LAST_DESC_CMD | ICE_TX_DESC_CMD_DUMMY |
  86		 ICE_TX_DESC_CMD_RE;
  87
  88	tx_buf->tx_flags = ICE_TX_FLAGS_DUMMY_PKT;
  89	tx_buf->raw_buf = raw_packet;
  90
  91	tx_desc->cmd_type_offset_bsz =
  92		ice_build_ctob(td_cmd, 0, ICE_FDIR_MAX_RAW_PKT_SIZE, 0);
  93
  94	/* Force memory write to complete before letting h/w know
  95	 * there are new descriptors to fetch.
  96	 */
  97	wmb();
  98
  99	/* mark the data descriptor to be watched */
 100	first->next_to_watch = tx_desc;
 101
 102	writel(tx_ring->next_to_use, tx_ring->tail);
 103
 104	return 0;
 105}
 106
 107/**
 108 * ice_unmap_and_free_tx_buf - Release a Tx buffer
 109 * @ring: the ring that owns the buffer
 110 * @tx_buf: the buffer to free
 111 */
 112static void
 113ice_unmap_and_free_tx_buf(struct ice_tx_ring *ring, struct ice_tx_buf *tx_buf)
 114{
 115	if (tx_buf->skb) {
 116		if (tx_buf->tx_flags & ICE_TX_FLAGS_DUMMY_PKT)
 117			devm_kfree(ring->dev, tx_buf->raw_buf);
 118		else if (ice_ring_is_xdp(ring))
 119			page_frag_free(tx_buf->raw_buf);
 120		else
 121			dev_kfree_skb_any(tx_buf->skb);
 122		if (dma_unmap_len(tx_buf, len))
 123			dma_unmap_single(ring->dev,
 124					 dma_unmap_addr(tx_buf, dma),
 125					 dma_unmap_len(tx_buf, len),
 126					 DMA_TO_DEVICE);
 127	} else if (dma_unmap_len(tx_buf, len)) {
 128		dma_unmap_page(ring->dev,
 129			       dma_unmap_addr(tx_buf, dma),
 130			       dma_unmap_len(tx_buf, len),
 131			       DMA_TO_DEVICE);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 132	}
 133
 134	tx_buf->next_to_watch = NULL;
 135	tx_buf->skb = NULL;
 136	dma_unmap_len_set(tx_buf, len, 0);
 137	/* tx_buf must be completely set up in the transmit path */
 138}
 139
 140static struct netdev_queue *txring_txq(const struct ice_tx_ring *ring)
 141{
 142	return netdev_get_tx_queue(ring->netdev, ring->q_index);
 143}
 144
 145/**
 146 * ice_clean_tx_ring - Free any empty Tx buffers
 147 * @tx_ring: ring to be cleaned
 148 */
 149void ice_clean_tx_ring(struct ice_tx_ring *tx_ring)
 150{
 151	u32 size;
 152	u16 i;
 153
 154	if (ice_ring_is_xdp(tx_ring) && tx_ring->xsk_pool) {
 155		ice_xsk_clean_xdp_ring(tx_ring);
 156		goto tx_skip_free;
 157	}
 158
 159	/* ring already cleared, nothing to do */
 160	if (!tx_ring->tx_buf)
 161		return;
 162
 163	/* Free all the Tx ring sk_buffs */
 164	for (i = 0; i < tx_ring->count; i++)
 165		ice_unmap_and_free_tx_buf(tx_ring, &tx_ring->tx_buf[i]);
 166
 167tx_skip_free:
 168	memset(tx_ring->tx_buf, 0, sizeof(*tx_ring->tx_buf) * tx_ring->count);
 169
 170	size = ALIGN(tx_ring->count * sizeof(struct ice_tx_desc),
 171		     PAGE_SIZE);
 172	/* Zero out the descriptor ring */
 173	memset(tx_ring->desc, 0, size);
 174
 175	tx_ring->next_to_use = 0;
 176	tx_ring->next_to_clean = 0;
 177	tx_ring->next_dd = ICE_RING_QUARTER(tx_ring) - 1;
 178	tx_ring->next_rs = ICE_RING_QUARTER(tx_ring) - 1;
 179
 180	if (!tx_ring->netdev)
 181		return;
 182
 183	/* cleanup Tx queue statistics */
 184	netdev_tx_reset_queue(txring_txq(tx_ring));
 185}
 186
 187/**
 188 * ice_free_tx_ring - Free Tx resources per queue
 189 * @tx_ring: Tx descriptor ring for a specific queue
 190 *
 191 * Free all transmit software resources
 192 */
 193void ice_free_tx_ring(struct ice_tx_ring *tx_ring)
 194{
 195	u32 size;
 196
 197	ice_clean_tx_ring(tx_ring);
 198	devm_kfree(tx_ring->dev, tx_ring->tx_buf);
 199	tx_ring->tx_buf = NULL;
 200
 201	if (tx_ring->desc) {
 202		size = ALIGN(tx_ring->count * sizeof(struct ice_tx_desc),
 203			     PAGE_SIZE);
 204		dmam_free_coherent(tx_ring->dev, size,
 205				   tx_ring->desc, tx_ring->dma);
 206		tx_ring->desc = NULL;
 207	}
 208}
 209
 210/**
 211 * ice_clean_tx_irq - Reclaim resources after transmit completes
 212 * @tx_ring: Tx ring to clean
 213 * @napi_budget: Used to determine if we are in netpoll
 214 *
 215 * Returns true if there's any budget left (e.g. the clean is finished)
 216 */
 217static bool ice_clean_tx_irq(struct ice_tx_ring *tx_ring, int napi_budget)
 218{
 219	unsigned int total_bytes = 0, total_pkts = 0;
 220	unsigned int budget = ICE_DFLT_IRQ_WORK;
 221	struct ice_vsi *vsi = tx_ring->vsi;
 222	s16 i = tx_ring->next_to_clean;
 223	struct ice_tx_desc *tx_desc;
 224	struct ice_tx_buf *tx_buf;
 225
 226	/* get the bql data ready */
 227	netdev_txq_bql_complete_prefetchw(txring_txq(tx_ring));
 228
 229	tx_buf = &tx_ring->tx_buf[i];
 230	tx_desc = ICE_TX_DESC(tx_ring, i);
 231	i -= tx_ring->count;
 232
 233	prefetch(&vsi->state);
 234
 235	do {
 236		struct ice_tx_desc *eop_desc = tx_buf->next_to_watch;
 237
 238		/* if next_to_watch is not set then there is no work pending */
 239		if (!eop_desc)
 240			break;
 241
 242		/* follow the guidelines of other drivers */
 243		prefetchw(&tx_buf->skb->users);
 244
 245		smp_rmb();	/* prevent any other reads prior to eop_desc */
 246
 247		ice_trace(clean_tx_irq, tx_ring, tx_desc, tx_buf);
 248		/* if the descriptor isn't done, no work yet to do */
 249		if (!(eop_desc->cmd_type_offset_bsz &
 250		      cpu_to_le64(ICE_TX_DESC_DTYPE_DESC_DONE)))
 251			break;
 252
 253		/* clear next_to_watch to prevent false hangs */
 254		tx_buf->next_to_watch = NULL;
 255
 256		/* update the statistics for this packet */
 257		total_bytes += tx_buf->bytecount;
 258		total_pkts += tx_buf->gso_segs;
 259
 260		/* free the skb */
 261		napi_consume_skb(tx_buf->skb, napi_budget);
 262
 263		/* unmap skb header data */
 264		dma_unmap_single(tx_ring->dev,
 265				 dma_unmap_addr(tx_buf, dma),
 266				 dma_unmap_len(tx_buf, len),
 267				 DMA_TO_DEVICE);
 268
 269		/* clear tx_buf data */
 270		tx_buf->skb = NULL;
 271		dma_unmap_len_set(tx_buf, len, 0);
 272
 273		/* unmap remaining buffers */
 274		while (tx_desc != eop_desc) {
 275			ice_trace(clean_tx_irq_unmap, tx_ring, tx_desc, tx_buf);
 276			tx_buf++;
 277			tx_desc++;
 278			i++;
 279			if (unlikely(!i)) {
 280				i -= tx_ring->count;
 281				tx_buf = tx_ring->tx_buf;
 282				tx_desc = ICE_TX_DESC(tx_ring, 0);
 283			}
 284
 285			/* unmap any remaining paged data */
 286			if (dma_unmap_len(tx_buf, len)) {
 287				dma_unmap_page(tx_ring->dev,
 288					       dma_unmap_addr(tx_buf, dma),
 289					       dma_unmap_len(tx_buf, len),
 290					       DMA_TO_DEVICE);
 291				dma_unmap_len_set(tx_buf, len, 0);
 292			}
 293		}
 294		ice_trace(clean_tx_irq_unmap_eop, tx_ring, tx_desc, tx_buf);
 295
 296		/* move us one more past the eop_desc for start of next pkt */
 297		tx_buf++;
 298		tx_desc++;
 299		i++;
 300		if (unlikely(!i)) {
 301			i -= tx_ring->count;
 302			tx_buf = tx_ring->tx_buf;
 303			tx_desc = ICE_TX_DESC(tx_ring, 0);
 304		}
 305
 306		prefetch(tx_desc);
 307
 308		/* update budget accounting */
 309		budget--;
 310	} while (likely(budget));
 311
 312	i += tx_ring->count;
 313	tx_ring->next_to_clean = i;
 314
 315	ice_update_tx_ring_stats(tx_ring, total_pkts, total_bytes);
 316	netdev_tx_completed_queue(txring_txq(tx_ring), total_pkts, total_bytes);
 317
 318#define TX_WAKE_THRESHOLD ((s16)(DESC_NEEDED * 2))
 319	if (unlikely(total_pkts && netif_carrier_ok(tx_ring->netdev) &&
 320		     (ICE_DESC_UNUSED(tx_ring) >= TX_WAKE_THRESHOLD))) {
 321		/* Make sure that anybody stopping the queue after this
 322		 * sees the new next_to_clean.
 323		 */
 324		smp_mb();
 325		if (netif_tx_queue_stopped(txring_txq(tx_ring)) &&
 326		    !test_bit(ICE_VSI_DOWN, vsi->state)) {
 327			netif_tx_wake_queue(txring_txq(tx_ring));
 328			++tx_ring->ring_stats->tx_stats.restart_q;
 329		}
 330	}
 331
 332	return !!budget;
 333}
 334
 335/**
 336 * ice_setup_tx_ring - Allocate the Tx descriptors
 337 * @tx_ring: the Tx ring to set up
 338 *
 339 * Return 0 on success, negative on error
 340 */
 341int ice_setup_tx_ring(struct ice_tx_ring *tx_ring)
 342{
 343	struct device *dev = tx_ring->dev;
 344	u32 size;
 345
 346	if (!dev)
 347		return -ENOMEM;
 348
 349	/* warn if we are about to overwrite the pointer */
 350	WARN_ON(tx_ring->tx_buf);
 351	tx_ring->tx_buf =
 352		devm_kcalloc(dev, sizeof(*tx_ring->tx_buf), tx_ring->count,
 353			     GFP_KERNEL);
 354	if (!tx_ring->tx_buf)
 355		return -ENOMEM;
 356
 357	/* round up to nearest page */
 358	size = ALIGN(tx_ring->count * sizeof(struct ice_tx_desc),
 359		     PAGE_SIZE);
 360	tx_ring->desc = dmam_alloc_coherent(dev, size, &tx_ring->dma,
 361					    GFP_KERNEL);
 362	if (!tx_ring->desc) {
 363		dev_err(dev, "Unable to allocate memory for the Tx descriptor ring, size=%d\n",
 364			size);
 365		goto err;
 366	}
 367
 368	tx_ring->next_to_use = 0;
 369	tx_ring->next_to_clean = 0;
 370	tx_ring->ring_stats->tx_stats.prev_pkt = -1;
 371	return 0;
 372
 373err:
 374	devm_kfree(dev, tx_ring->tx_buf);
 375	tx_ring->tx_buf = NULL;
 376	return -ENOMEM;
 377}
 378
 379/**
 380 * ice_clean_rx_ring - Free Rx buffers
 381 * @rx_ring: ring to be cleaned
 382 */
 383void ice_clean_rx_ring(struct ice_rx_ring *rx_ring)
 384{
 
 385	struct device *dev = rx_ring->dev;
 386	u32 size;
 387	u16 i;
 388
 389	/* ring already cleared, nothing to do */
 390	if (!rx_ring->rx_buf)
 391		return;
 392
 393	if (rx_ring->skb) {
 394		dev_kfree_skb(rx_ring->skb);
 395		rx_ring->skb = NULL;
 396	}
 397
 398	if (rx_ring->xsk_pool) {
 399		ice_xsk_clean_rx_ring(rx_ring);
 400		goto rx_skip_free;
 401	}
 402
 
 
 
 
 
 403	/* Free all the Rx ring sk_buffs */
 404	for (i = 0; i < rx_ring->count; i++) {
 405		struct ice_rx_buf *rx_buf = &rx_ring->rx_buf[i];
 406
 407		if (!rx_buf->page)
 408			continue;
 409
 410		/* Invalidate cache lines that may have been written to by
 411		 * device so that we avoid corrupting memory.
 412		 */
 413		dma_sync_single_range_for_cpu(dev, rx_buf->dma,
 414					      rx_buf->page_offset,
 415					      rx_ring->rx_buf_len,
 416					      DMA_FROM_DEVICE);
 417
 418		/* free resources associated with mapping */
 419		dma_unmap_page_attrs(dev, rx_buf->dma, ice_rx_pg_size(rx_ring),
 420				     DMA_FROM_DEVICE, ICE_RX_DMA_ATTR);
 421		__page_frag_cache_drain(rx_buf->page, rx_buf->pagecnt_bias);
 422
 423		rx_buf->page = NULL;
 424		rx_buf->page_offset = 0;
 425	}
 426
 427rx_skip_free:
 428	if (rx_ring->xsk_pool)
 429		memset(rx_ring->xdp_buf, 0, array_size(rx_ring->count, sizeof(*rx_ring->xdp_buf)));
 430	else
 431		memset(rx_ring->rx_buf, 0, array_size(rx_ring->count, sizeof(*rx_ring->rx_buf)));
 432
 433	/* Zero out the descriptor ring */
 434	size = ALIGN(rx_ring->count * sizeof(union ice_32byte_rx_desc),
 435		     PAGE_SIZE);
 436	memset(rx_ring->desc, 0, size);
 437
 438	rx_ring->next_to_alloc = 0;
 439	rx_ring->next_to_clean = 0;
 
 440	rx_ring->next_to_use = 0;
 441}
 442
 443/**
 444 * ice_free_rx_ring - Free Rx resources
 445 * @rx_ring: ring to clean the resources from
 446 *
 447 * Free all receive software resources
 448 */
 449void ice_free_rx_ring(struct ice_rx_ring *rx_ring)
 450{
 451	u32 size;
 452
 453	ice_clean_rx_ring(rx_ring);
 454	if (rx_ring->vsi->type == ICE_VSI_PF)
 455		if (xdp_rxq_info_is_reg(&rx_ring->xdp_rxq))
 456			xdp_rxq_info_unreg(&rx_ring->xdp_rxq);
 457	rx_ring->xdp_prog = NULL;
 458	if (rx_ring->xsk_pool) {
 459		kfree(rx_ring->xdp_buf);
 460		rx_ring->xdp_buf = NULL;
 461	} else {
 462		kfree(rx_ring->rx_buf);
 463		rx_ring->rx_buf = NULL;
 464	}
 465
 466	if (rx_ring->desc) {
 467		size = ALIGN(rx_ring->count * sizeof(union ice_32byte_rx_desc),
 468			     PAGE_SIZE);
 469		dmam_free_coherent(rx_ring->dev, size,
 470				   rx_ring->desc, rx_ring->dma);
 471		rx_ring->desc = NULL;
 472	}
 473}
 474
 475/**
 476 * ice_setup_rx_ring - Allocate the Rx descriptors
 477 * @rx_ring: the Rx ring to set up
 478 *
 479 * Return 0 on success, negative on error
 480 */
 481int ice_setup_rx_ring(struct ice_rx_ring *rx_ring)
 482{
 483	struct device *dev = rx_ring->dev;
 484	u32 size;
 485
 486	if (!dev)
 487		return -ENOMEM;
 488
 489	/* warn if we are about to overwrite the pointer */
 490	WARN_ON(rx_ring->rx_buf);
 491	rx_ring->rx_buf =
 492		kcalloc(rx_ring->count, sizeof(*rx_ring->rx_buf), GFP_KERNEL);
 493	if (!rx_ring->rx_buf)
 494		return -ENOMEM;
 495
 496	/* round up to nearest page */
 497	size = ALIGN(rx_ring->count * sizeof(union ice_32byte_rx_desc),
 498		     PAGE_SIZE);
 499	rx_ring->desc = dmam_alloc_coherent(dev, size, &rx_ring->dma,
 500					    GFP_KERNEL);
 501	if (!rx_ring->desc) {
 502		dev_err(dev, "Unable to allocate memory for the Rx descriptor ring, size=%d\n",
 503			size);
 504		goto err;
 505	}
 506
 507	rx_ring->next_to_use = 0;
 508	rx_ring->next_to_clean = 0;
 
 509
 510	if (ice_is_xdp_ena_vsi(rx_ring->vsi))
 511		WRITE_ONCE(rx_ring->xdp_prog, rx_ring->vsi->xdp_prog);
 512
 513	if (rx_ring->vsi->type == ICE_VSI_PF &&
 514	    !xdp_rxq_info_is_reg(&rx_ring->xdp_rxq))
 515		if (xdp_rxq_info_reg(&rx_ring->xdp_rxq, rx_ring->netdev,
 516				     rx_ring->q_index, rx_ring->q_vector->napi.napi_id))
 517			goto err;
 518	return 0;
 519
 520err:
 521	kfree(rx_ring->rx_buf);
 522	rx_ring->rx_buf = NULL;
 523	return -ENOMEM;
 524}
 525
 526static unsigned int
 527ice_rx_frame_truesize(struct ice_rx_ring *rx_ring, unsigned int __maybe_unused size)
 528{
 529	unsigned int truesize;
 530
 531#if (PAGE_SIZE < 8192)
 532	truesize = ice_rx_pg_size(rx_ring) / 2; /* Must be power-of-2 */
 533#else
 534	truesize = rx_ring->rx_offset ?
 535		SKB_DATA_ALIGN(rx_ring->rx_offset + size) +
 536		SKB_DATA_ALIGN(sizeof(struct skb_shared_info)) :
 537		SKB_DATA_ALIGN(size);
 538#endif
 539	return truesize;
 540}
 541
 542/**
 543 * ice_run_xdp - Executes an XDP program on initialized xdp_buff
 544 * @rx_ring: Rx ring
 545 * @xdp: xdp_buff used as input to the XDP program
 546 * @xdp_prog: XDP program to run
 547 * @xdp_ring: ring to be used for XDP_TX action
 
 548 *
 549 * Returns any of ICE_XDP_{PASS, CONSUMED, TX, REDIR}
 550 */
 551static int
 552ice_run_xdp(struct ice_rx_ring *rx_ring, struct xdp_buff *xdp,
 553	    struct bpf_prog *xdp_prog, struct ice_tx_ring *xdp_ring)
 
 554{
 555	int err;
 556	u32 act;
 557
 
 
 
 
 
 558	act = bpf_prog_run_xdp(xdp_prog, xdp);
 559	switch (act) {
 560	case XDP_PASS:
 561		return ICE_XDP_PASS;
 562	case XDP_TX:
 563		if (static_branch_unlikely(&ice_xdp_locking_key))
 564			spin_lock(&xdp_ring->tx_lock);
 565		err = ice_xmit_xdp_ring(xdp->data, xdp->data_end - xdp->data, xdp_ring);
 566		if (static_branch_unlikely(&ice_xdp_locking_key))
 567			spin_unlock(&xdp_ring->tx_lock);
 568		if (err == ICE_XDP_CONSUMED)
 569			goto out_failure;
 570		return err;
 571	case XDP_REDIRECT:
 572		err = xdp_do_redirect(rx_ring->netdev, xdp, xdp_prog);
 573		if (err)
 574			goto out_failure;
 575		return ICE_XDP_REDIR;
 
 576	default:
 577		bpf_warn_invalid_xdp_action(rx_ring->netdev, xdp_prog, act);
 578		fallthrough;
 579	case XDP_ABORTED:
 580out_failure:
 581		trace_xdp_exception(rx_ring->netdev, xdp_prog, act);
 582		fallthrough;
 583	case XDP_DROP:
 584		return ICE_XDP_CONSUMED;
 585	}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 586}
 587
 588/**
 589 * ice_xdp_xmit - submit packets to XDP ring for transmission
 590 * @dev: netdev
 591 * @n: number of XDP frames to be transmitted
 592 * @frames: XDP frames to be transmitted
 593 * @flags: transmit flags
 594 *
 595 * Returns number of frames successfully sent. Failed frames
 596 * will be free'ed by XDP core.
 597 * For error cases, a negative errno code is returned and no-frames
 598 * are transmitted (caller must handle freeing frames).
 599 */
 600int
 601ice_xdp_xmit(struct net_device *dev, int n, struct xdp_frame **frames,
 602	     u32 flags)
 603{
 604	struct ice_netdev_priv *np = netdev_priv(dev);
 605	unsigned int queue_index = smp_processor_id();
 606	struct ice_vsi *vsi = np->vsi;
 607	struct ice_tx_ring *xdp_ring;
 
 608	int nxmit = 0, i;
 609
 610	if (test_bit(ICE_VSI_DOWN, vsi->state))
 611		return -ENETDOWN;
 612
 613	if (!ice_is_xdp_ena_vsi(vsi))
 614		return -ENXIO;
 615
 616	if (unlikely(flags & ~XDP_XMIT_FLAGS_MASK))
 617		return -EINVAL;
 618
 619	if (static_branch_unlikely(&ice_xdp_locking_key)) {
 620		queue_index %= vsi->num_xdp_txq;
 621		xdp_ring = vsi->xdp_rings[queue_index];
 622		spin_lock(&xdp_ring->tx_lock);
 623	} else {
 624		/* Generally, should not happen */
 625		if (unlikely(queue_index >= vsi->num_xdp_txq))
 626			return -ENXIO;
 627		xdp_ring = vsi->xdp_rings[queue_index];
 628	}
 629
 
 630	for (i = 0; i < n; i++) {
 631		struct xdp_frame *xdpf = frames[i];
 632		int err;
 633
 634		err = ice_xmit_xdp_ring(xdpf->data, xdpf->len, xdp_ring);
 635		if (err != ICE_XDP_TX)
 636			break;
 637		nxmit++;
 638	}
 639
 
 640	if (unlikely(flags & XDP_XMIT_FLUSH))
 641		ice_xdp_ring_update_tail(xdp_ring);
 642
 643	if (static_branch_unlikely(&ice_xdp_locking_key))
 644		spin_unlock(&xdp_ring->tx_lock);
 645
 646	return nxmit;
 647}
 648
 649/**
 650 * ice_alloc_mapped_page - recycle or make a new page
 651 * @rx_ring: ring to use
 652 * @bi: rx_buf struct to modify
 653 *
 654 * Returns true if the page was successfully allocated or
 655 * reused.
 656 */
 657static bool
 658ice_alloc_mapped_page(struct ice_rx_ring *rx_ring, struct ice_rx_buf *bi)
 659{
 660	struct page *page = bi->page;
 661	dma_addr_t dma;
 662
 663	/* since we are recycling buffers we should seldom need to alloc */
 664	if (likely(page))
 665		return true;
 666
 667	/* alloc new page for storage */
 668	page = dev_alloc_pages(ice_rx_pg_order(rx_ring));
 669	if (unlikely(!page)) {
 670		rx_ring->ring_stats->rx_stats.alloc_page_failed++;
 671		return false;
 672	}
 673
 674	/* map page for use */
 675	dma = dma_map_page_attrs(rx_ring->dev, page, 0, ice_rx_pg_size(rx_ring),
 676				 DMA_FROM_DEVICE, ICE_RX_DMA_ATTR);
 677
 678	/* if mapping failed free memory back to system since
 679	 * there isn't much point in holding memory we can't use
 680	 */
 681	if (dma_mapping_error(rx_ring->dev, dma)) {
 682		__free_pages(page, ice_rx_pg_order(rx_ring));
 683		rx_ring->ring_stats->rx_stats.alloc_page_failed++;
 684		return false;
 685	}
 686
 687	bi->dma = dma;
 688	bi->page = page;
 689	bi->page_offset = rx_ring->rx_offset;
 690	page_ref_add(page, USHRT_MAX - 1);
 691	bi->pagecnt_bias = USHRT_MAX;
 692
 693	return true;
 694}
 695
 696/**
 697 * ice_alloc_rx_bufs - Replace used receive buffers
 698 * @rx_ring: ring to place buffers on
 699 * @cleaned_count: number of buffers to replace
 700 *
 701 * Returns false if all allocations were successful, true if any fail. Returning
 702 * true signals to the caller that we didn't replace cleaned_count buffers and
 703 * there is more work to do.
 704 *
 705 * First, try to clean "cleaned_count" Rx buffers. Then refill the cleaned Rx
 706 * buffers. Then bump tail at most one time. Grouping like this lets us avoid
 707 * multiple tail writes per call.
 708 */
 709bool ice_alloc_rx_bufs(struct ice_rx_ring *rx_ring, u16 cleaned_count)
 710{
 711	union ice_32b_rx_flex_desc *rx_desc;
 712	u16 ntu = rx_ring->next_to_use;
 713	struct ice_rx_buf *bi;
 714
 715	/* do nothing if no valid netdev defined */
 716	if ((!rx_ring->netdev && rx_ring->vsi->type != ICE_VSI_CTRL) ||
 717	    !cleaned_count)
 718		return false;
 719
 720	/* get the Rx descriptor and buffer based on next_to_use */
 721	rx_desc = ICE_RX_DESC(rx_ring, ntu);
 722	bi = &rx_ring->rx_buf[ntu];
 723
 724	do {
 725		/* if we fail here, we have work remaining */
 726		if (!ice_alloc_mapped_page(rx_ring, bi))
 727			break;
 728
 729		/* sync the buffer for use by the device */
 730		dma_sync_single_range_for_device(rx_ring->dev, bi->dma,
 731						 bi->page_offset,
 732						 rx_ring->rx_buf_len,
 733						 DMA_FROM_DEVICE);
 734
 735		/* Refresh the desc even if buffer_addrs didn't change
 736		 * because each write-back erases this info.
 737		 */
 738		rx_desc->read.pkt_addr = cpu_to_le64(bi->dma + bi->page_offset);
 739
 740		rx_desc++;
 741		bi++;
 742		ntu++;
 743		if (unlikely(ntu == rx_ring->count)) {
 744			rx_desc = ICE_RX_DESC(rx_ring, 0);
 745			bi = rx_ring->rx_buf;
 746			ntu = 0;
 747		}
 748
 749		/* clear the status bits for the next_to_use descriptor */
 750		rx_desc->wb.status_error0 = 0;
 751
 752		cleaned_count--;
 753	} while (cleaned_count);
 754
 755	if (rx_ring->next_to_use != ntu)
 756		ice_release_rx_desc(rx_ring, ntu);
 757
 758	return !!cleaned_count;
 759}
 760
 761/**
 762 * ice_rx_buf_adjust_pg_offset - Prepare Rx buffer for reuse
 763 * @rx_buf: Rx buffer to adjust
 764 * @size: Size of adjustment
 765 *
 766 * Update the offset within page so that Rx buf will be ready to be reused.
 767 * For systems with PAGE_SIZE < 8192 this function will flip the page offset
 768 * so the second half of page assigned to Rx buffer will be used, otherwise
 769 * the offset is moved by "size" bytes
 770 */
 771static void
 772ice_rx_buf_adjust_pg_offset(struct ice_rx_buf *rx_buf, unsigned int size)
 773{
 774#if (PAGE_SIZE < 8192)
 775	/* flip page offset to other buffer */
 776	rx_buf->page_offset ^= size;
 777#else
 778	/* move offset up to the next cache line */
 779	rx_buf->page_offset += size;
 780#endif
 781}
 782
 783/**
 784 * ice_can_reuse_rx_page - Determine if page can be reused for another Rx
 785 * @rx_buf: buffer containing the page
 786 * @rx_buf_pgcnt: rx_buf page refcount pre xdp_do_redirect() call
 787 *
 788 * If page is reusable, we have a green light for calling ice_reuse_rx_page,
 789 * which will assign the current buffer to the buffer that next_to_alloc is
 790 * pointing to; otherwise, the DMA mapping needs to be destroyed and
 791 * page freed
 792 */
 793static bool
 794ice_can_reuse_rx_page(struct ice_rx_buf *rx_buf, int rx_buf_pgcnt)
 795{
 796	unsigned int pagecnt_bias = rx_buf->pagecnt_bias;
 797	struct page *page = rx_buf->page;
 798
 799	/* avoid re-using remote and pfmemalloc pages */
 800	if (!dev_page_is_reusable(page))
 801		return false;
 802
 803#if (PAGE_SIZE < 8192)
 804	/* if we are only owner of page we can reuse it */
 805	if (unlikely((rx_buf_pgcnt - pagecnt_bias) > 1))
 806		return false;
 807#else
 808#define ICE_LAST_OFFSET \
 809	(SKB_WITH_OVERHEAD(PAGE_SIZE) - ICE_RXBUF_2048)
 810	if (rx_buf->page_offset > ICE_LAST_OFFSET)
 811		return false;
 812#endif /* PAGE_SIZE < 8192) */
 813
 814	/* If we have drained the page fragment pool we need to update
 815	 * the pagecnt_bias and page count so that we fully restock the
 816	 * number of references the driver holds.
 817	 */
 818	if (unlikely(pagecnt_bias == 1)) {
 819		page_ref_add(page, USHRT_MAX - 1);
 820		rx_buf->pagecnt_bias = USHRT_MAX;
 821	}
 822
 823	return true;
 824}
 825
 826/**
 827 * ice_add_rx_frag - Add contents of Rx buffer to sk_buff as a frag
 828 * @rx_ring: Rx descriptor ring to transact packets on
 
 829 * @rx_buf: buffer containing page to add
 830 * @skb: sk_buff to place the data into
 831 * @size: packet length from rx_desc
 832 *
 833 * This function will add the data contained in rx_buf->page to the skb.
 834 * It will just attach the page as a frag to the skb.
 835 * The function will then update the page offset.
 836 */
 837static void
 838ice_add_rx_frag(struct ice_rx_ring *rx_ring, struct ice_rx_buf *rx_buf,
 839		struct sk_buff *skb, unsigned int size)
 840{
 841#if (PAGE_SIZE >= 8192)
 842	unsigned int truesize = SKB_DATA_ALIGN(size + rx_ring->rx_offset);
 843#else
 844	unsigned int truesize = ice_rx_pg_size(rx_ring) / 2;
 845#endif
 846
 847	if (!size)
 848		return;
 849	skb_add_rx_frag(skb, skb_shinfo(skb)->nr_frags, rx_buf->page,
 850			rx_buf->page_offset, size, truesize);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 851
 852	/* page is being used so we must update the page offset */
 853	ice_rx_buf_adjust_pg_offset(rx_buf, truesize);
 854}
 855
 856/**
 857 * ice_reuse_rx_page - page flip buffer and store it back on the ring
 858 * @rx_ring: Rx descriptor ring to store buffers on
 859 * @old_buf: donor buffer to have page reused
 860 *
 861 * Synchronizes page for reuse by the adapter
 862 */
 863static void
 864ice_reuse_rx_page(struct ice_rx_ring *rx_ring, struct ice_rx_buf *old_buf)
 865{
 866	u16 nta = rx_ring->next_to_alloc;
 867	struct ice_rx_buf *new_buf;
 868
 869	new_buf = &rx_ring->rx_buf[nta];
 870
 871	/* update, and store next to alloc */
 872	nta++;
 873	rx_ring->next_to_alloc = (nta < rx_ring->count) ? nta : 0;
 874
 875	/* Transfer page from old buffer to new buffer.
 876	 * Move each member individually to avoid possible store
 877	 * forwarding stalls and unnecessary copy of skb.
 878	 */
 879	new_buf->dma = old_buf->dma;
 880	new_buf->page = old_buf->page;
 881	new_buf->page_offset = old_buf->page_offset;
 882	new_buf->pagecnt_bias = old_buf->pagecnt_bias;
 883}
 884
 885/**
 886 * ice_get_rx_buf - Fetch Rx buffer and synchronize data for use
 887 * @rx_ring: Rx descriptor ring to transact packets on
 888 * @size: size of buffer to add to skb
 889 * @rx_buf_pgcnt: rx_buf page refcount
 890 *
 891 * This function will pull an Rx buffer from the ring and synchronize it
 892 * for use by the CPU.
 893 */
 894static struct ice_rx_buf *
 895ice_get_rx_buf(struct ice_rx_ring *rx_ring, const unsigned int size,
 896	       int *rx_buf_pgcnt)
 897{
 898	struct ice_rx_buf *rx_buf;
 899
 900	rx_buf = &rx_ring->rx_buf[rx_ring->next_to_clean];
 901	*rx_buf_pgcnt =
 902#if (PAGE_SIZE < 8192)
 903		page_count(rx_buf->page);
 904#else
 905		0;
 906#endif
 907	prefetchw(rx_buf->page);
 908
 909	if (!size)
 910		return rx_buf;
 911	/* we are reusing so sync this buffer for CPU use */
 912	dma_sync_single_range_for_cpu(rx_ring->dev, rx_buf->dma,
 913				      rx_buf->page_offset, size,
 914				      DMA_FROM_DEVICE);
 915
 916	/* We have pulled a buffer for use, so decrement pagecnt_bias */
 917	rx_buf->pagecnt_bias--;
 918
 919	return rx_buf;
 920}
 921
 922/**
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 923 * ice_build_skb - Build skb around an existing buffer
 924 * @rx_ring: Rx descriptor ring to transact packets on
 925 * @rx_buf: Rx buffer to pull data from
 926 * @xdp: xdp_buff pointing to the data
 927 *
 928 * This function builds an skb around an existing Rx buffer, taking care
 929 * to set up the skb correctly and avoid any memcpy overhead.
 
 930 */
 931static struct sk_buff *
 932ice_build_skb(struct ice_rx_ring *rx_ring, struct ice_rx_buf *rx_buf,
 933	      struct xdp_buff *xdp)
 934{
 935	u8 metasize = xdp->data - xdp->data_meta;
 936#if (PAGE_SIZE < 8192)
 937	unsigned int truesize = ice_rx_pg_size(rx_ring) / 2;
 938#else
 939	unsigned int truesize = SKB_DATA_ALIGN(sizeof(struct skb_shared_info)) +
 940				SKB_DATA_ALIGN(xdp->data_end -
 941					       xdp->data_hard_start);
 942#endif
 943	struct sk_buff *skb;
 944
 
 
 
 
 
 945	/* Prefetch first cache line of first page. If xdp->data_meta
 946	 * is unused, this points exactly as xdp->data, otherwise we
 947	 * likely have a consumer accessing first few bytes of meta
 948	 * data, and then actual data.
 949	 */
 950	net_prefetch(xdp->data_meta);
 951	/* build an skb around the page buffer */
 952	skb = napi_build_skb(xdp->data_hard_start, truesize);
 953	if (unlikely(!skb))
 954		return NULL;
 955
 956	/* must to record Rx queue, otherwise OS features such as
 957	 * symmetric queue won't work
 958	 */
 959	skb_record_rx_queue(skb, rx_ring->q_index);
 960
 961	/* update pointers within the skb to store the data */
 962	skb_reserve(skb, xdp->data - xdp->data_hard_start);
 963	__skb_put(skb, xdp->data_end - xdp->data);
 964	if (metasize)
 965		skb_metadata_set(skb, metasize);
 966
 967	/* buffer is used by skb, update page_offset */
 968	ice_rx_buf_adjust_pg_offset(rx_buf, truesize);
 
 
 
 969
 970	return skb;
 971}
 972
 973/**
 974 * ice_construct_skb - Allocate skb and populate it
 975 * @rx_ring: Rx descriptor ring to transact packets on
 976 * @rx_buf: Rx buffer to pull data from
 977 * @xdp: xdp_buff pointing to the data
 978 *
 979 * This function allocates an skb. It then populates it with the page
 980 * data from the current receive descriptor, taking care to set up the
 981 * skb correctly.
 982 */
 983static struct sk_buff *
 984ice_construct_skb(struct ice_rx_ring *rx_ring, struct ice_rx_buf *rx_buf,
 985		  struct xdp_buff *xdp)
 986{
 987	unsigned int metasize = xdp->data - xdp->data_meta;
 988	unsigned int size = xdp->data_end - xdp->data;
 
 
 
 989	unsigned int headlen;
 990	struct sk_buff *skb;
 991
 992	/* prefetch first cache line of first page */
 993	net_prefetch(xdp->data_meta);
 
 
 
 
 
 994
 995	/* allocate a skb to store the frags */
 996	skb = __napi_alloc_skb(&rx_ring->q_vector->napi,
 997			       ICE_RX_HDR_SIZE + metasize,
 998			       GFP_ATOMIC | __GFP_NOWARN);
 999	if (unlikely(!skb))
1000		return NULL;
1001
 
1002	skb_record_rx_queue(skb, rx_ring->q_index);
1003	/* Determine available headroom for copy */
1004	headlen = size;
1005	if (headlen > ICE_RX_HDR_SIZE)
1006		headlen = eth_get_headlen(skb->dev, xdp->data, ICE_RX_HDR_SIZE);
1007
1008	/* align pull length to size of long to optimize memcpy performance */
1009	memcpy(__skb_put(skb, headlen + metasize), xdp->data_meta,
1010	       ALIGN(headlen + metasize, sizeof(long)));
1011
1012	if (metasize) {
1013		skb_metadata_set(skb, metasize);
1014		__skb_pull(skb, metasize);
1015	}
1016
1017	/* if we exhaust the linear part then add what is left as a frag */
1018	size -= headlen;
1019	if (size) {
1020#if (PAGE_SIZE >= 8192)
1021		unsigned int truesize = SKB_DATA_ALIGN(size);
1022#else
1023		unsigned int truesize = ice_rx_pg_size(rx_ring) / 2;
1024#endif
 
 
 
1025		skb_add_rx_frag(skb, 0, rx_buf->page,
1026				rx_buf->page_offset + headlen, size, truesize);
1027		/* buffer is used by skb, update page_offset */
1028		ice_rx_buf_adjust_pg_offset(rx_buf, truesize);
1029	} else {
1030		/* buffer is unused, reset bias back to rx_buf; data was copied
1031		 * onto skb's linear part so there's no need for adjusting
1032		 * page offset and we can reuse this buffer as-is
 
1033		 */
1034		rx_buf->pagecnt_bias++;
1035	}
1036
 
 
 
 
 
 
 
 
 
 
 
 
1037	return skb;
1038}
1039
1040/**
1041 * ice_put_rx_buf - Clean up used buffer and either recycle or free
1042 * @rx_ring: Rx descriptor ring to transact packets on
1043 * @rx_buf: Rx buffer to pull data from
1044 * @rx_buf_pgcnt: Rx buffer page count pre xdp_do_redirect()
1045 *
1046 * This function will update next_to_clean and then clean up the contents
1047 * of the rx_buf. It will either recycle the buffer or unmap it and free
1048 * the associated resources.
1049 */
1050static void
1051ice_put_rx_buf(struct ice_rx_ring *rx_ring, struct ice_rx_buf *rx_buf,
1052	       int rx_buf_pgcnt)
1053{
1054	u16 ntc = rx_ring->next_to_clean + 1;
1055
1056	/* fetch, update, and store next to clean */
1057	ntc = (ntc < rx_ring->count) ? ntc : 0;
1058	rx_ring->next_to_clean = ntc;
1059
1060	if (!rx_buf)
1061		return;
1062
1063	if (ice_can_reuse_rx_page(rx_buf, rx_buf_pgcnt)) {
1064		/* hand second half of page back to the ring */
1065		ice_reuse_rx_page(rx_ring, rx_buf);
1066	} else {
1067		/* we are not reusing the buffer so unmap it */
1068		dma_unmap_page_attrs(rx_ring->dev, rx_buf->dma,
1069				     ice_rx_pg_size(rx_ring), DMA_FROM_DEVICE,
1070				     ICE_RX_DMA_ATTR);
1071		__page_frag_cache_drain(rx_buf->page, rx_buf->pagecnt_bias);
1072	}
1073
1074	/* clear contents of buffer_info */
1075	rx_buf->page = NULL;
1076}
1077
1078/**
1079 * ice_is_non_eop - process handling of non-EOP buffers
1080 * @rx_ring: Rx ring being processed
1081 * @rx_desc: Rx descriptor for current buffer
1082 *
1083 * If the buffer is an EOP buffer, this function exits returning false,
1084 * otherwise return true indicating that this is in fact a non-EOP buffer.
1085 */
1086static bool
1087ice_is_non_eop(struct ice_rx_ring *rx_ring, union ice_32b_rx_flex_desc *rx_desc)
1088{
1089	/* if we are the last buffer then there is nothing else to do */
1090#define ICE_RXD_EOF BIT(ICE_RX_FLEX_DESC_STATUS0_EOF_S)
1091	if (likely(ice_test_staterr(rx_desc->wb.status_error0, ICE_RXD_EOF)))
1092		return false;
1093
1094	rx_ring->ring_stats->rx_stats.non_eop_descs++;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1095
1096	return true;
 
 
1097}
1098
1099/**
1100 * ice_clean_rx_irq - Clean completed descriptors from Rx ring - bounce buf
1101 * @rx_ring: Rx descriptor ring to transact packets on
1102 * @budget: Total limit on number of packets to process
1103 *
1104 * This function provides a "bounce buffer" approach to Rx interrupt
1105 * processing. The advantage to this is that on systems that have
1106 * expensive overhead for IOMMU access this provides a means of avoiding
1107 * it by maintaining the mapping of the page to the system.
1108 *
1109 * Returns amount of work completed
1110 */
1111int ice_clean_rx_irq(struct ice_rx_ring *rx_ring, int budget)
1112{
1113	unsigned int total_rx_bytes = 0, total_rx_pkts = 0, frame_sz = 0;
1114	u16 cleaned_count = ICE_DESC_UNUSED(rx_ring);
1115	unsigned int offset = rx_ring->rx_offset;
 
1116	struct ice_tx_ring *xdp_ring = NULL;
1117	unsigned int xdp_res, xdp_xmit = 0;
1118	struct sk_buff *skb = rx_ring->skb;
1119	struct bpf_prog *xdp_prog = NULL;
1120	struct xdp_buff xdp;
 
 
 
1121	bool failure;
1122
1123	/* Frame size depend on rx_ring setup when PAGE_SIZE=4K */
1124#if (PAGE_SIZE < 8192)
1125	frame_sz = ice_rx_frame_truesize(rx_ring, 0);
1126#endif
1127	xdp_init_buff(&xdp, frame_sz, &rx_ring->xdp_rxq);
1128
1129	xdp_prog = READ_ONCE(rx_ring->xdp_prog);
1130	if (xdp_prog)
1131		xdp_ring = rx_ring->xdp_ring;
 
 
1132
1133	/* start the loop to process Rx packets bounded by 'budget' */
1134	while (likely(total_rx_pkts < (unsigned int)budget)) {
1135		union ice_32b_rx_flex_desc *rx_desc;
1136		struct ice_rx_buf *rx_buf;
1137		unsigned char *hard_start;
1138		unsigned int size;
1139		u16 stat_err_bits;
1140		int rx_buf_pgcnt;
1141		u16 vlan_tag = 0;
1142		u16 rx_ptype;
1143
1144		/* get the Rx desc from Rx ring based on 'next_to_clean' */
1145		rx_desc = ICE_RX_DESC(rx_ring, rx_ring->next_to_clean);
1146
1147		/* status_error_len will always be zero for unused descriptors
1148		 * because it's cleared in cleanup, and overlaps with hdr_addr
1149		 * which is always zero because packet split isn't used, if the
1150		 * hardware wrote DD then it will be non-zero
1151		 */
1152		stat_err_bits = BIT(ICE_RX_FLEX_DESC_STATUS0_DD_S);
1153		if (!ice_test_staterr(rx_desc->wb.status_error0, stat_err_bits))
1154			break;
1155
1156		/* This memory barrier is needed to keep us from reading
1157		 * any other fields out of the rx_desc until we know the
1158		 * DD bit is set.
1159		 */
1160		dma_rmb();
1161
1162		ice_trace(clean_rx_irq, rx_ring, rx_desc);
1163		if (rx_desc->wb.rxdid == FDIR_DESC_RXDID || !rx_ring->netdev) {
1164			struct ice_vsi *ctrl_vsi = rx_ring->vsi;
1165
1166			if (rx_desc->wb.rxdid == FDIR_DESC_RXDID &&
1167			    ctrl_vsi->vf)
1168				ice_vc_fdir_irq_handler(ctrl_vsi, rx_desc);
1169			ice_put_rx_buf(rx_ring, NULL, 0);
1170			cleaned_count++;
 
1171			continue;
1172		}
1173
1174		size = le16_to_cpu(rx_desc->wb.pkt_len) &
1175			ICE_RX_FLX_DESC_PKT_LEN_M;
1176
1177		/* retrieve a buffer from the ring */
1178		rx_buf = ice_get_rx_buf(rx_ring, size, &rx_buf_pgcnt);
 
 
 
1179
1180		if (!size) {
1181			xdp.data = NULL;
1182			xdp.data_end = NULL;
1183			xdp.data_hard_start = NULL;
1184			xdp.data_meta = NULL;
1185			goto construct_skb;
 
1186		}
 
 
1187
1188		hard_start = page_address(rx_buf->page) + rx_buf->page_offset -
1189			     offset;
1190		xdp_prepare_buff(&xdp, hard_start, offset, size, true);
1191#if (PAGE_SIZE > 4096)
1192		/* At larger PAGE_SIZE, frame_sz depend on len size */
1193		xdp.frame_sz = ice_rx_frame_truesize(rx_ring, size);
1194#endif
1195
1196		if (!xdp_prog)
 
 
1197			goto construct_skb;
 
 
1198
1199		xdp_res = ice_run_xdp(rx_ring, &xdp, xdp_prog, xdp_ring);
1200		if (!xdp_res)
1201			goto construct_skb;
1202		if (xdp_res & (ICE_XDP_TX | ICE_XDP_REDIR)) {
1203			xdp_xmit |= xdp_res;
1204			ice_rx_buf_adjust_pg_offset(rx_buf, xdp.frame_sz);
1205		} else {
1206			rx_buf->pagecnt_bias++;
1207		}
1208		total_rx_bytes += size;
1209		total_rx_pkts++;
1210
1211		cleaned_count++;
1212		ice_put_rx_buf(rx_ring, rx_buf, rx_buf_pgcnt);
1213		continue;
1214construct_skb:
1215		if (skb) {
1216			ice_add_rx_frag(rx_ring, rx_buf, skb, size);
1217		} else if (likely(xdp.data)) {
1218			if (ice_ring_uses_build_skb(rx_ring))
1219				skb = ice_build_skb(rx_ring, rx_buf, &xdp);
1220			else
1221				skb = ice_construct_skb(rx_ring, rx_buf, &xdp);
1222		}
1223		/* exit if we failed to retrieve a buffer */
1224		if (!skb) {
1225			rx_ring->ring_stats->rx_stats.alloc_buf_failed++;
1226			if (rx_buf)
1227				rx_buf->pagecnt_bias++;
1228			break;
1229		}
 
1230
1231		ice_put_rx_buf(rx_ring, rx_buf, rx_buf_pgcnt);
1232		cleaned_count++;
1233
1234		/* skip if it is NOP desc */
1235		if (ice_is_non_eop(rx_ring, rx_desc))
1236			continue;
1237
1238		stat_err_bits = BIT(ICE_RX_FLEX_DESC_STATUS0_RXE_S);
1239		if (unlikely(ice_test_staterr(rx_desc->wb.status_error0,
1240					      stat_err_bits))) {
1241			dev_kfree_skb_any(skb);
1242			continue;
1243		}
1244
1245		vlan_tag = ice_get_vlan_tag_from_rx_desc(rx_desc);
1246
1247		/* pad the skb if needed, to make a valid ethernet frame */
1248		if (eth_skb_pad(skb)) {
1249			skb = NULL;
1250			continue;
1251		}
1252
1253		/* probably a little skewed due to removing CRC */
1254		total_rx_bytes += skb->len;
1255
1256		/* populate checksum, VLAN, and protocol */
1257		rx_ptype = le16_to_cpu(rx_desc->wb.ptype_flex_flags0) &
1258			ICE_RX_FLEX_DESC_PTYPE_M;
1259
1260		ice_process_skb_fields(rx_ring, rx_desc, skb, rx_ptype);
1261
1262		ice_trace(clean_rx_irq_indicate, rx_ring, rx_desc, skb);
1263		/* send completed skb up the stack */
1264		ice_receive_skb(rx_ring, skb, vlan_tag);
1265		skb = NULL;
1266
1267		/* update budget accounting */
1268		total_rx_pkts++;
1269	}
1270
 
1271	/* return up to cleaned_count buffers to hardware */
1272	failure = ice_alloc_rx_bufs(rx_ring, cleaned_count);
1273
1274	if (xdp_prog)
1275		ice_finalize_xdp_rx(xdp_ring, xdp_xmit);
1276	rx_ring->skb = skb;
1277
1278	if (rx_ring->ring_stats)
1279		ice_update_rx_ring_stats(rx_ring, total_rx_pkts,
1280					 total_rx_bytes);
1281
1282	/* guarantee a trip back through this routine if there was a failure */
1283	return failure ? budget : (int)total_rx_pkts;
1284}
1285
1286static void __ice_update_sample(struct ice_q_vector *q_vector,
1287				struct ice_ring_container *rc,
1288				struct dim_sample *sample,
1289				bool is_tx)
1290{
1291	u64 packets = 0, bytes = 0;
1292
1293	if (is_tx) {
1294		struct ice_tx_ring *tx_ring;
1295
1296		ice_for_each_tx_ring(tx_ring, *rc) {
1297			struct ice_ring_stats *ring_stats;
1298
1299			ring_stats = tx_ring->ring_stats;
1300			if (!ring_stats)
1301				continue;
1302			packets += ring_stats->stats.pkts;
1303			bytes += ring_stats->stats.bytes;
1304		}
1305	} else {
1306		struct ice_rx_ring *rx_ring;
1307
1308		ice_for_each_rx_ring(rx_ring, *rc) {
1309			struct ice_ring_stats *ring_stats;
1310
1311			ring_stats = rx_ring->ring_stats;
1312			if (!ring_stats)
1313				continue;
1314			packets += ring_stats->stats.pkts;
1315			bytes += ring_stats->stats.bytes;
1316		}
1317	}
1318
1319	dim_update_sample(q_vector->total_events, packets, bytes, sample);
1320	sample->comp_ctr = 0;
1321
1322	/* if dim settings get stale, like when not updated for 1
1323	 * second or longer, force it to start again. This addresses the
1324	 * frequent case of an idle queue being switched to by the
1325	 * scheduler. The 1,000 here means 1,000 milliseconds.
1326	 */
1327	if (ktime_ms_delta(sample->time, rc->dim.start_sample.time) >= 1000)
1328		rc->dim.state = DIM_START_MEASURE;
1329}
1330
1331/**
1332 * ice_net_dim - Update net DIM algorithm
1333 * @q_vector: the vector associated with the interrupt
1334 *
1335 * Create a DIM sample and notify net_dim() so that it can possibly decide
1336 * a new ITR value based on incoming packets, bytes, and interrupts.
1337 *
1338 * This function is a no-op if the ring is not configured to dynamic ITR.
1339 */
1340static void ice_net_dim(struct ice_q_vector *q_vector)
1341{
1342	struct ice_ring_container *tx = &q_vector->tx;
1343	struct ice_ring_container *rx = &q_vector->rx;
1344
1345	if (ITR_IS_DYNAMIC(tx)) {
1346		struct dim_sample dim_sample;
1347
1348		__ice_update_sample(q_vector, tx, &dim_sample, true);
1349		net_dim(&tx->dim, dim_sample);
1350	}
1351
1352	if (ITR_IS_DYNAMIC(rx)) {
1353		struct dim_sample dim_sample;
1354
1355		__ice_update_sample(q_vector, rx, &dim_sample, false);
1356		net_dim(&rx->dim, dim_sample);
1357	}
1358}
1359
1360/**
1361 * ice_buildreg_itr - build value for writing to the GLINT_DYN_CTL register
1362 * @itr_idx: interrupt throttling index
1363 * @itr: interrupt throttling value in usecs
1364 */
1365static u32 ice_buildreg_itr(u16 itr_idx, u16 itr)
1366{
1367	/* The ITR value is reported in microseconds, and the register value is
1368	 * recorded in 2 microsecond units. For this reason we only need to
1369	 * shift by the GLINT_DYN_CTL_INTERVAL_S - ICE_ITR_GRAN_S to apply this
1370	 * granularity as a shift instead of division. The mask makes sure the
1371	 * ITR value is never odd so we don't accidentally write into the field
1372	 * prior to the ITR field.
1373	 */
1374	itr &= ICE_ITR_MASK;
1375
1376	return GLINT_DYN_CTL_INTENA_M | GLINT_DYN_CTL_CLEARPBA_M |
1377		(itr_idx << GLINT_DYN_CTL_ITR_INDX_S) |
1378		(itr << (GLINT_DYN_CTL_INTERVAL_S - ICE_ITR_GRAN_S));
1379}
1380
1381/**
1382 * ice_enable_interrupt - re-enable MSI-X interrupt
1383 * @q_vector: the vector associated with the interrupt to enable
1384 *
1385 * If the VSI is down, the interrupt will not be re-enabled. Also,
1386 * when enabling the interrupt always reset the wb_on_itr to false
1387 * and trigger a software interrupt to clean out internal state.
1388 */
1389static void ice_enable_interrupt(struct ice_q_vector *q_vector)
1390{
1391	struct ice_vsi *vsi = q_vector->vsi;
1392	bool wb_en = q_vector->wb_on_itr;
1393	u32 itr_val;
1394
1395	if (test_bit(ICE_DOWN, vsi->state))
1396		return;
1397
1398	/* trigger an ITR delayed software interrupt when exiting busy poll, to
1399	 * make sure to catch any pending cleanups that might have been missed
1400	 * due to interrupt state transition. If busy poll or poll isn't
1401	 * enabled, then don't update ITR, and just enable the interrupt.
1402	 */
1403	if (!wb_en) {
1404		itr_val = ice_buildreg_itr(ICE_ITR_NONE, 0);
1405	} else {
1406		q_vector->wb_on_itr = false;
1407
1408		/* do two things here with a single write. Set up the third ITR
1409		 * index to be used for software interrupt moderation, and then
1410		 * trigger a software interrupt with a rate limit of 20K on
1411		 * software interrupts, this will help avoid high interrupt
1412		 * loads due to frequently polling and exiting polling.
1413		 */
1414		itr_val = ice_buildreg_itr(ICE_IDX_ITR2, ICE_ITR_20K);
1415		itr_val |= GLINT_DYN_CTL_SWINT_TRIG_M |
1416			   ICE_IDX_ITR2 << GLINT_DYN_CTL_SW_ITR_INDX_S |
1417			   GLINT_DYN_CTL_SW_ITR_INDX_ENA_M;
1418	}
1419	wr32(&vsi->back->hw, GLINT_DYN_CTL(q_vector->reg_idx), itr_val);
1420}
1421
1422/**
1423 * ice_set_wb_on_itr - set WB_ON_ITR for this q_vector
1424 * @q_vector: q_vector to set WB_ON_ITR on
1425 *
1426 * We need to tell hardware to write-back completed descriptors even when
1427 * interrupts are disabled. Descriptors will be written back on cache line
1428 * boundaries without WB_ON_ITR enabled, but if we don't enable WB_ON_ITR
1429 * descriptors may not be written back if they don't fill a cache line until
1430 * the next interrupt.
1431 *
1432 * This sets the write-back frequency to whatever was set previously for the
1433 * ITR indices. Also, set the INTENA_MSK bit to make sure hardware knows we
1434 * aren't meddling with the INTENA_M bit.
1435 */
1436static void ice_set_wb_on_itr(struct ice_q_vector *q_vector)
1437{
1438	struct ice_vsi *vsi = q_vector->vsi;
1439
1440	/* already in wb_on_itr mode no need to change it */
1441	if (q_vector->wb_on_itr)
1442		return;
1443
1444	/* use previously set ITR values for all of the ITR indices by
1445	 * specifying ICE_ITR_NONE, which will vary in adaptive (AIM) mode and
1446	 * be static in non-adaptive mode (user configured)
1447	 */
1448	wr32(&vsi->back->hw, GLINT_DYN_CTL(q_vector->reg_idx),
1449	     ((ICE_ITR_NONE << GLINT_DYN_CTL_ITR_INDX_S) &
1450	      GLINT_DYN_CTL_ITR_INDX_M) | GLINT_DYN_CTL_INTENA_MSK_M |
1451	     GLINT_DYN_CTL_WB_ON_ITR_M);
1452
1453	q_vector->wb_on_itr = true;
1454}
1455
1456/**
1457 * ice_napi_poll - NAPI polling Rx/Tx cleanup routine
1458 * @napi: napi struct with our devices info in it
1459 * @budget: amount of work driver is allowed to do this pass, in packets
1460 *
1461 * This function will clean all queues associated with a q_vector.
1462 *
1463 * Returns the amount of work done
1464 */
1465int ice_napi_poll(struct napi_struct *napi, int budget)
1466{
1467	struct ice_q_vector *q_vector =
1468				container_of(napi, struct ice_q_vector, napi);
1469	struct ice_tx_ring *tx_ring;
1470	struct ice_rx_ring *rx_ring;
1471	bool clean_complete = true;
1472	int budget_per_ring;
1473	int work_done = 0;
1474
1475	/* Since the actual Tx work is minimal, we can give the Tx a larger
1476	 * budget and be more aggressive about cleaning up the Tx descriptors.
1477	 */
1478	ice_for_each_tx_ring(tx_ring, q_vector->tx) {
 
1479		bool wd;
1480
1481		if (tx_ring->xsk_pool)
1482			wd = ice_xmit_zc(tx_ring);
1483		else if (ice_ring_is_xdp(tx_ring))
1484			wd = true;
1485		else
1486			wd = ice_clean_tx_irq(tx_ring, budget);
1487
1488		if (!wd)
1489			clean_complete = false;
1490	}
1491
1492	/* Handle case where we are called by netpoll with a budget of 0 */
1493	if (unlikely(budget <= 0))
1494		return budget;
1495
1496	/* normally we have 1 Rx ring per q_vector */
1497	if (unlikely(q_vector->num_ring_rx > 1))
1498		/* We attempt to distribute budget to each Rx queue fairly, but
1499		 * don't allow the budget to go below 1 because that would exit
1500		 * polling early.
1501		 */
1502		budget_per_ring = max_t(int, budget / q_vector->num_ring_rx, 1);
1503	else
1504		/* Max of 1 Rx ring in this q_vector so give it the budget */
1505		budget_per_ring = budget;
1506
1507	ice_for_each_rx_ring(rx_ring, q_vector->rx) {
 
1508		int cleaned;
1509
1510		/* A dedicated path for zero-copy allows making a single
1511		 * comparison in the irq context instead of many inside the
1512		 * ice_clean_rx_irq function and makes the codebase cleaner.
1513		 */
1514		cleaned = rx_ring->xsk_pool ?
1515			  ice_clean_rx_irq_zc(rx_ring, budget_per_ring) :
1516			  ice_clean_rx_irq(rx_ring, budget_per_ring);
1517		work_done += cleaned;
1518		/* if we clean as many as budgeted, we must not be done */
1519		if (cleaned >= budget_per_ring)
1520			clean_complete = false;
1521	}
1522
1523	/* If work not completed, return budget and polling will return */
1524	if (!clean_complete) {
1525		/* Set the writeback on ITR so partial completions of
1526		 * cache-lines will still continue even if we're polling.
1527		 */
1528		ice_set_wb_on_itr(q_vector);
1529		return budget;
1530	}
1531
1532	/* Exit the polling mode, but don't re-enable interrupts if stack might
1533	 * poll us due to busy-polling
1534	 */
1535	if (napi_complete_done(napi, work_done)) {
1536		ice_net_dim(q_vector);
1537		ice_enable_interrupt(q_vector);
1538	} else {
1539		ice_set_wb_on_itr(q_vector);
1540	}
1541
1542	return min_t(int, work_done, budget - 1);
1543}
1544
1545/**
1546 * __ice_maybe_stop_tx - 2nd level check for Tx stop conditions
1547 * @tx_ring: the ring to be checked
1548 * @size: the size buffer we want to assure is available
1549 *
1550 * Returns -EBUSY if a stop is needed, else 0
1551 */
1552static int __ice_maybe_stop_tx(struct ice_tx_ring *tx_ring, unsigned int size)
1553{
1554	netif_tx_stop_queue(txring_txq(tx_ring));
1555	/* Memory barrier before checking head and tail */
1556	smp_mb();
1557
1558	/* Check again in a case another CPU has just made room available. */
1559	if (likely(ICE_DESC_UNUSED(tx_ring) < size))
1560		return -EBUSY;
1561
1562	/* A reprieve! - use start_queue because it doesn't call schedule */
1563	netif_tx_start_queue(txring_txq(tx_ring));
1564	++tx_ring->ring_stats->tx_stats.restart_q;
1565	return 0;
1566}
1567
1568/**
1569 * ice_maybe_stop_tx - 1st level check for Tx stop conditions
1570 * @tx_ring: the ring to be checked
1571 * @size:    the size buffer we want to assure is available
1572 *
1573 * Returns 0 if stop is not needed
1574 */
1575static int ice_maybe_stop_tx(struct ice_tx_ring *tx_ring, unsigned int size)
1576{
1577	if (likely(ICE_DESC_UNUSED(tx_ring) >= size))
1578		return 0;
1579
1580	return __ice_maybe_stop_tx(tx_ring, size);
1581}
1582
1583/**
1584 * ice_tx_map - Build the Tx descriptor
1585 * @tx_ring: ring to send buffer on
1586 * @first: first buffer info buffer to use
1587 * @off: pointer to struct that holds offload parameters
1588 *
1589 * This function loops over the skb data pointed to by *first
1590 * and gets a physical address for each memory location and programs
1591 * it and the length into the transmit descriptor.
1592 */
1593static void
1594ice_tx_map(struct ice_tx_ring *tx_ring, struct ice_tx_buf *first,
1595	   struct ice_tx_offload_params *off)
1596{
1597	u64 td_offset, td_tag, td_cmd;
1598	u16 i = tx_ring->next_to_use;
1599	unsigned int data_len, size;
1600	struct ice_tx_desc *tx_desc;
1601	struct ice_tx_buf *tx_buf;
1602	struct sk_buff *skb;
1603	skb_frag_t *frag;
1604	dma_addr_t dma;
1605	bool kick;
1606
1607	td_tag = off->td_l2tag1;
1608	td_cmd = off->td_cmd;
1609	td_offset = off->td_offset;
1610	skb = first->skb;
1611
1612	data_len = skb->data_len;
1613	size = skb_headlen(skb);
1614
1615	tx_desc = ICE_TX_DESC(tx_ring, i);
1616
1617	if (first->tx_flags & ICE_TX_FLAGS_HW_VLAN) {
1618		td_cmd |= (u64)ICE_TX_DESC_CMD_IL2TAG1;
1619		td_tag = (first->tx_flags & ICE_TX_FLAGS_VLAN_M) >>
1620			  ICE_TX_FLAGS_VLAN_S;
1621	}
1622
1623	dma = dma_map_single(tx_ring->dev, skb->data, size, DMA_TO_DEVICE);
1624
1625	tx_buf = first;
1626
1627	for (frag = &skb_shinfo(skb)->frags[0];; frag++) {
1628		unsigned int max_data = ICE_MAX_DATA_PER_TXD_ALIGNED;
1629
1630		if (dma_mapping_error(tx_ring->dev, dma))
1631			goto dma_error;
1632
1633		/* record length, and DMA address */
1634		dma_unmap_len_set(tx_buf, len, size);
1635		dma_unmap_addr_set(tx_buf, dma, dma);
1636
1637		/* align size to end of page */
1638		max_data += -dma & (ICE_MAX_READ_REQ_SIZE - 1);
1639		tx_desc->buf_addr = cpu_to_le64(dma);
1640
1641		/* account for data chunks larger than the hardware
1642		 * can handle
1643		 */
1644		while (unlikely(size > ICE_MAX_DATA_PER_TXD)) {
1645			tx_desc->cmd_type_offset_bsz =
1646				ice_build_ctob(td_cmd, td_offset, max_data,
1647					       td_tag);
1648
1649			tx_desc++;
1650			i++;
1651
1652			if (i == tx_ring->count) {
1653				tx_desc = ICE_TX_DESC(tx_ring, 0);
1654				i = 0;
1655			}
1656
1657			dma += max_data;
1658			size -= max_data;
1659
1660			max_data = ICE_MAX_DATA_PER_TXD_ALIGNED;
1661			tx_desc->buf_addr = cpu_to_le64(dma);
1662		}
1663
1664		if (likely(!data_len))
1665			break;
1666
1667		tx_desc->cmd_type_offset_bsz = ice_build_ctob(td_cmd, td_offset,
1668							      size, td_tag);
1669
1670		tx_desc++;
1671		i++;
1672
1673		if (i == tx_ring->count) {
1674			tx_desc = ICE_TX_DESC(tx_ring, 0);
1675			i = 0;
1676		}
1677
1678		size = skb_frag_size(frag);
1679		data_len -= size;
1680
1681		dma = skb_frag_dma_map(tx_ring->dev, frag, 0, size,
1682				       DMA_TO_DEVICE);
1683
1684		tx_buf = &tx_ring->tx_buf[i];
 
1685	}
1686
1687	/* record SW timestamp if HW timestamp is not available */
1688	skb_tx_timestamp(first->skb);
1689
1690	i++;
1691	if (i == tx_ring->count)
1692		i = 0;
1693
1694	/* write last descriptor with RS and EOP bits */
1695	td_cmd |= (u64)ICE_TXD_LAST_DESC_CMD;
1696	tx_desc->cmd_type_offset_bsz =
1697			ice_build_ctob(td_cmd, td_offset, size, td_tag);
1698
1699	/* Force memory writes to complete before letting h/w know there
1700	 * are new descriptors to fetch.
1701	 *
1702	 * We also use this memory barrier to make certain all of the
1703	 * status bits have been updated before next_to_watch is written.
1704	 */
1705	wmb();
1706
1707	/* set next_to_watch value indicating a packet is present */
1708	first->next_to_watch = tx_desc;
1709
1710	tx_ring->next_to_use = i;
1711
1712	ice_maybe_stop_tx(tx_ring, DESC_NEEDED);
1713
1714	/* notify HW of packet */
1715	kick = __netdev_tx_sent_queue(txring_txq(tx_ring), first->bytecount,
1716				      netdev_xmit_more());
1717	if (kick)
1718		/* notify HW of packet */
1719		writel(i, tx_ring->tail);
1720
1721	return;
1722
1723dma_error:
1724	/* clear DMA mappings for failed tx_buf map */
1725	for (;;) {
1726		tx_buf = &tx_ring->tx_buf[i];
1727		ice_unmap_and_free_tx_buf(tx_ring, tx_buf);
1728		if (tx_buf == first)
1729			break;
1730		if (i == 0)
1731			i = tx_ring->count;
1732		i--;
1733	}
1734
1735	tx_ring->next_to_use = i;
1736}
1737
1738/**
1739 * ice_tx_csum - Enable Tx checksum offloads
1740 * @first: pointer to the first descriptor
1741 * @off: pointer to struct that holds offload parameters
1742 *
1743 * Returns 0 or error (negative) if checksum offload can't happen, 1 otherwise.
1744 */
1745static
1746int ice_tx_csum(struct ice_tx_buf *first, struct ice_tx_offload_params *off)
1747{
1748	u32 l4_len = 0, l3_len = 0, l2_len = 0;
1749	struct sk_buff *skb = first->skb;
1750	union {
1751		struct iphdr *v4;
1752		struct ipv6hdr *v6;
1753		unsigned char *hdr;
1754	} ip;
1755	union {
1756		struct tcphdr *tcp;
1757		unsigned char *hdr;
1758	} l4;
1759	__be16 frag_off, protocol;
1760	unsigned char *exthdr;
1761	u32 offset, cmd = 0;
1762	u8 l4_proto = 0;
1763
1764	if (skb->ip_summed != CHECKSUM_PARTIAL)
1765		return 0;
1766
1767	protocol = vlan_get_protocol(skb);
1768
1769	if (eth_p_mpls(protocol)) {
1770		ip.hdr = skb_inner_network_header(skb);
1771		l4.hdr = skb_checksum_start(skb);
1772	} else {
1773		ip.hdr = skb_network_header(skb);
1774		l4.hdr = skb_transport_header(skb);
1775	}
1776
1777	/* compute outer L2 header size */
1778	l2_len = ip.hdr - skb->data;
1779	offset = (l2_len / 2) << ICE_TX_DESC_LEN_MACLEN_S;
1780
1781	/* set the tx_flags to indicate the IP protocol type. this is
1782	 * required so that checksum header computation below is accurate.
1783	 */
1784	if (ip.v4->version == 4)
1785		first->tx_flags |= ICE_TX_FLAGS_IPV4;
1786	else if (ip.v6->version == 6)
1787		first->tx_flags |= ICE_TX_FLAGS_IPV6;
1788
1789	if (skb->encapsulation) {
1790		bool gso_ena = false;
1791		u32 tunnel = 0;
1792
1793		/* define outer network header type */
1794		if (first->tx_flags & ICE_TX_FLAGS_IPV4) {
1795			tunnel |= (first->tx_flags & ICE_TX_FLAGS_TSO) ?
1796				  ICE_TX_CTX_EIPT_IPV4 :
1797				  ICE_TX_CTX_EIPT_IPV4_NO_CSUM;
1798			l4_proto = ip.v4->protocol;
1799		} else if (first->tx_flags & ICE_TX_FLAGS_IPV6) {
1800			int ret;
1801
1802			tunnel |= ICE_TX_CTX_EIPT_IPV6;
1803			exthdr = ip.hdr + sizeof(*ip.v6);
1804			l4_proto = ip.v6->nexthdr;
1805			ret = ipv6_skip_exthdr(skb, exthdr - skb->data,
1806					       &l4_proto, &frag_off);
1807			if (ret < 0)
1808				return -1;
1809		}
1810
1811		/* define outer transport */
1812		switch (l4_proto) {
1813		case IPPROTO_UDP:
1814			tunnel |= ICE_TXD_CTX_UDP_TUNNELING;
1815			first->tx_flags |= ICE_TX_FLAGS_TUNNEL;
1816			break;
1817		case IPPROTO_GRE:
1818			tunnel |= ICE_TXD_CTX_GRE_TUNNELING;
1819			first->tx_flags |= ICE_TX_FLAGS_TUNNEL;
1820			break;
1821		case IPPROTO_IPIP:
1822		case IPPROTO_IPV6:
1823			first->tx_flags |= ICE_TX_FLAGS_TUNNEL;
1824			l4.hdr = skb_inner_network_header(skb);
1825			break;
1826		default:
1827			if (first->tx_flags & ICE_TX_FLAGS_TSO)
1828				return -1;
1829
1830			skb_checksum_help(skb);
1831			return 0;
1832		}
1833
1834		/* compute outer L3 header size */
1835		tunnel |= ((l4.hdr - ip.hdr) / 4) <<
1836			  ICE_TXD_CTX_QW0_EIPLEN_S;
1837
1838		/* switch IP header pointer from outer to inner header */
1839		ip.hdr = skb_inner_network_header(skb);
1840
1841		/* compute tunnel header size */
1842		tunnel |= ((ip.hdr - l4.hdr) / 2) <<
1843			   ICE_TXD_CTX_QW0_NATLEN_S;
1844
1845		gso_ena = skb_shinfo(skb)->gso_type & SKB_GSO_PARTIAL;
1846		/* indicate if we need to offload outer UDP header */
1847		if ((first->tx_flags & ICE_TX_FLAGS_TSO) && !gso_ena &&
1848		    (skb_shinfo(skb)->gso_type & SKB_GSO_UDP_TUNNEL_CSUM))
1849			tunnel |= ICE_TXD_CTX_QW0_L4T_CS_M;
1850
1851		/* record tunnel offload values */
1852		off->cd_tunnel_params |= tunnel;
1853
1854		/* set DTYP=1 to indicate that it's an Tx context descriptor
1855		 * in IPsec tunnel mode with Tx offloads in Quad word 1
1856		 */
1857		off->cd_qw1 |= (u64)ICE_TX_DESC_DTYPE_CTX;
1858
1859		/* switch L4 header pointer from outer to inner */
1860		l4.hdr = skb_inner_transport_header(skb);
1861		l4_proto = 0;
1862
1863		/* reset type as we transition from outer to inner headers */
1864		first->tx_flags &= ~(ICE_TX_FLAGS_IPV4 | ICE_TX_FLAGS_IPV6);
1865		if (ip.v4->version == 4)
1866			first->tx_flags |= ICE_TX_FLAGS_IPV4;
1867		if (ip.v6->version == 6)
1868			first->tx_flags |= ICE_TX_FLAGS_IPV6;
1869	}
1870
1871	/* Enable IP checksum offloads */
1872	if (first->tx_flags & ICE_TX_FLAGS_IPV4) {
1873		l4_proto = ip.v4->protocol;
1874		/* the stack computes the IP header already, the only time we
1875		 * need the hardware to recompute it is in the case of TSO.
1876		 */
1877		if (first->tx_flags & ICE_TX_FLAGS_TSO)
1878			cmd |= ICE_TX_DESC_CMD_IIPT_IPV4_CSUM;
1879		else
1880			cmd |= ICE_TX_DESC_CMD_IIPT_IPV4;
1881
1882	} else if (first->tx_flags & ICE_TX_FLAGS_IPV6) {
1883		cmd |= ICE_TX_DESC_CMD_IIPT_IPV6;
1884		exthdr = ip.hdr + sizeof(*ip.v6);
1885		l4_proto = ip.v6->nexthdr;
1886		if (l4.hdr != exthdr)
1887			ipv6_skip_exthdr(skb, exthdr - skb->data, &l4_proto,
1888					 &frag_off);
1889	} else {
1890		return -1;
1891	}
1892
1893	/* compute inner L3 header size */
1894	l3_len = l4.hdr - ip.hdr;
1895	offset |= (l3_len / 4) << ICE_TX_DESC_LEN_IPLEN_S;
1896
1897	/* Enable L4 checksum offloads */
1898	switch (l4_proto) {
1899	case IPPROTO_TCP:
1900		/* enable checksum offloads */
1901		cmd |= ICE_TX_DESC_CMD_L4T_EOFT_TCP;
1902		l4_len = l4.tcp->doff;
1903		offset |= l4_len << ICE_TX_DESC_LEN_L4_LEN_S;
1904		break;
1905	case IPPROTO_UDP:
1906		/* enable UDP checksum offload */
1907		cmd |= ICE_TX_DESC_CMD_L4T_EOFT_UDP;
1908		l4_len = (sizeof(struct udphdr) >> 2);
1909		offset |= l4_len << ICE_TX_DESC_LEN_L4_LEN_S;
1910		break;
1911	case IPPROTO_SCTP:
1912		/* enable SCTP checksum offload */
1913		cmd |= ICE_TX_DESC_CMD_L4T_EOFT_SCTP;
1914		l4_len = sizeof(struct sctphdr) >> 2;
1915		offset |= l4_len << ICE_TX_DESC_LEN_L4_LEN_S;
1916		break;
1917
1918	default:
1919		if (first->tx_flags & ICE_TX_FLAGS_TSO)
1920			return -1;
1921		skb_checksum_help(skb);
1922		return 0;
1923	}
1924
1925	off->td_cmd |= cmd;
1926	off->td_offset |= offset;
1927	return 1;
1928}
1929
1930/**
1931 * ice_tx_prepare_vlan_flags - prepare generic Tx VLAN tagging flags for HW
1932 * @tx_ring: ring to send buffer on
1933 * @first: pointer to struct ice_tx_buf
1934 *
1935 * Checks the skb and set up correspondingly several generic transmit flags
1936 * related to VLAN tagging for the HW, such as VLAN, DCB, etc.
1937 */
1938static void
1939ice_tx_prepare_vlan_flags(struct ice_tx_ring *tx_ring, struct ice_tx_buf *first)
1940{
1941	struct sk_buff *skb = first->skb;
1942
1943	/* nothing left to do, software offloaded VLAN */
1944	if (!skb_vlan_tag_present(skb) && eth_type_vlan(skb->protocol))
1945		return;
1946
1947	/* the VLAN ethertype/tpid is determined by VSI configuration and netdev
1948	 * feature flags, which the driver only allows either 802.1Q or 802.1ad
1949	 * VLAN offloads exclusively so we only care about the VLAN ID here
1950	 */
1951	if (skb_vlan_tag_present(skb)) {
1952		first->tx_flags |= skb_vlan_tag_get(skb) << ICE_TX_FLAGS_VLAN_S;
1953		if (tx_ring->flags & ICE_TX_FLAGS_RING_VLAN_L2TAG2)
1954			first->tx_flags |= ICE_TX_FLAGS_HW_OUTER_SINGLE_VLAN;
1955		else
1956			first->tx_flags |= ICE_TX_FLAGS_HW_VLAN;
1957	}
1958
1959	ice_tx_prepare_vlan_flags_dcb(tx_ring, first);
1960}
1961
1962/**
1963 * ice_tso - computes mss and TSO length to prepare for TSO
1964 * @first: pointer to struct ice_tx_buf
1965 * @off: pointer to struct that holds offload parameters
1966 *
1967 * Returns 0 or error (negative) if TSO can't happen, 1 otherwise.
1968 */
1969static
1970int ice_tso(struct ice_tx_buf *first, struct ice_tx_offload_params *off)
1971{
1972	struct sk_buff *skb = first->skb;
1973	union {
1974		struct iphdr *v4;
1975		struct ipv6hdr *v6;
1976		unsigned char *hdr;
1977	} ip;
1978	union {
1979		struct tcphdr *tcp;
1980		struct udphdr *udp;
1981		unsigned char *hdr;
1982	} l4;
1983	u64 cd_mss, cd_tso_len;
1984	__be16 protocol;
1985	u32 paylen;
1986	u8 l4_start;
1987	int err;
1988
1989	if (skb->ip_summed != CHECKSUM_PARTIAL)
1990		return 0;
1991
1992	if (!skb_is_gso(skb))
1993		return 0;
1994
1995	err = skb_cow_head(skb, 0);
1996	if (err < 0)
1997		return err;
1998
1999	/* cppcheck-suppress unreadVariable */
2000	protocol = vlan_get_protocol(skb);
2001
2002	if (eth_p_mpls(protocol))
2003		ip.hdr = skb_inner_network_header(skb);
2004	else
2005		ip.hdr = skb_network_header(skb);
2006	l4.hdr = skb_checksum_start(skb);
2007
2008	/* initialize outer IP header fields */
2009	if (ip.v4->version == 4) {
2010		ip.v4->tot_len = 0;
2011		ip.v4->check = 0;
2012	} else {
2013		ip.v6->payload_len = 0;
2014	}
2015
2016	if (skb_shinfo(skb)->gso_type & (SKB_GSO_GRE |
2017					 SKB_GSO_GRE_CSUM |
2018					 SKB_GSO_IPXIP4 |
2019					 SKB_GSO_IPXIP6 |
2020					 SKB_GSO_UDP_TUNNEL |
2021					 SKB_GSO_UDP_TUNNEL_CSUM)) {
2022		if (!(skb_shinfo(skb)->gso_type & SKB_GSO_PARTIAL) &&
2023		    (skb_shinfo(skb)->gso_type & SKB_GSO_UDP_TUNNEL_CSUM)) {
2024			l4.udp->len = 0;
2025
2026			/* determine offset of outer transport header */
2027			l4_start = (u8)(l4.hdr - skb->data);
2028
2029			/* remove payload length from outer checksum */
2030			paylen = skb->len - l4_start;
2031			csum_replace_by_diff(&l4.udp->check,
2032					     (__force __wsum)htonl(paylen));
2033		}
2034
2035		/* reset pointers to inner headers */
2036
2037		/* cppcheck-suppress unreadVariable */
2038		ip.hdr = skb_inner_network_header(skb);
2039		l4.hdr = skb_inner_transport_header(skb);
2040
2041		/* initialize inner IP header fields */
2042		if (ip.v4->version == 4) {
2043			ip.v4->tot_len = 0;
2044			ip.v4->check = 0;
2045		} else {
2046			ip.v6->payload_len = 0;
2047		}
2048	}
2049
2050	/* determine offset of transport header */
2051	l4_start = (u8)(l4.hdr - skb->data);
2052
2053	/* remove payload length from checksum */
2054	paylen = skb->len - l4_start;
2055
2056	if (skb_shinfo(skb)->gso_type & SKB_GSO_UDP_L4) {
2057		csum_replace_by_diff(&l4.udp->check,
2058				     (__force __wsum)htonl(paylen));
2059		/* compute length of UDP segmentation header */
2060		off->header_len = (u8)sizeof(l4.udp) + l4_start;
2061	} else {
2062		csum_replace_by_diff(&l4.tcp->check,
2063				     (__force __wsum)htonl(paylen));
2064		/* compute length of TCP segmentation header */
2065		off->header_len = (u8)((l4.tcp->doff * 4) + l4_start);
2066	}
2067
2068	/* update gso_segs and bytecount */
2069	first->gso_segs = skb_shinfo(skb)->gso_segs;
2070	first->bytecount += (first->gso_segs - 1) * off->header_len;
2071
2072	cd_tso_len = skb->len - off->header_len;
2073	cd_mss = skb_shinfo(skb)->gso_size;
2074
2075	/* record cdesc_qw1 with TSO parameters */
2076	off->cd_qw1 |= (u64)(ICE_TX_DESC_DTYPE_CTX |
2077			     (ICE_TX_CTX_DESC_TSO << ICE_TXD_CTX_QW1_CMD_S) |
2078			     (cd_tso_len << ICE_TXD_CTX_QW1_TSO_LEN_S) |
2079			     (cd_mss << ICE_TXD_CTX_QW1_MSS_S));
2080	first->tx_flags |= ICE_TX_FLAGS_TSO;
2081	return 1;
2082}
2083
2084/**
2085 * ice_txd_use_count  - estimate the number of descriptors needed for Tx
2086 * @size: transmit request size in bytes
2087 *
2088 * Due to hardware alignment restrictions (4K alignment), we need to
2089 * assume that we can have no more than 12K of data per descriptor, even
2090 * though each descriptor can take up to 16K - 1 bytes of aligned memory.
2091 * Thus, we need to divide by 12K. But division is slow! Instead,
2092 * we decompose the operation into shifts and one relatively cheap
2093 * multiply operation.
2094 *
2095 * To divide by 12K, we first divide by 4K, then divide by 3:
2096 *     To divide by 4K, shift right by 12 bits
2097 *     To divide by 3, multiply by 85, then divide by 256
2098 *     (Divide by 256 is done by shifting right by 8 bits)
2099 * Finally, we add one to round up. Because 256 isn't an exact multiple of
2100 * 3, we'll underestimate near each multiple of 12K. This is actually more
2101 * accurate as we have 4K - 1 of wiggle room that we can fit into the last
2102 * segment. For our purposes this is accurate out to 1M which is orders of
2103 * magnitude greater than our largest possible GSO size.
2104 *
2105 * This would then be implemented as:
2106 *     return (((size >> 12) * 85) >> 8) + ICE_DESCS_FOR_SKB_DATA_PTR;
2107 *
2108 * Since multiplication and division are commutative, we can reorder
2109 * operations into:
2110 *     return ((size * 85) >> 20) + ICE_DESCS_FOR_SKB_DATA_PTR;
2111 */
2112static unsigned int ice_txd_use_count(unsigned int size)
2113{
2114	return ((size * 85) >> 20) + ICE_DESCS_FOR_SKB_DATA_PTR;
2115}
2116
2117/**
2118 * ice_xmit_desc_count - calculate number of Tx descriptors needed
2119 * @skb: send buffer
2120 *
2121 * Returns number of data descriptors needed for this skb.
2122 */
2123static unsigned int ice_xmit_desc_count(struct sk_buff *skb)
2124{
2125	const skb_frag_t *frag = &skb_shinfo(skb)->frags[0];
2126	unsigned int nr_frags = skb_shinfo(skb)->nr_frags;
2127	unsigned int count = 0, size = skb_headlen(skb);
2128
2129	for (;;) {
2130		count += ice_txd_use_count(size);
2131
2132		if (!nr_frags--)
2133			break;
2134
2135		size = skb_frag_size(frag++);
2136	}
2137
2138	return count;
2139}
2140
2141/**
2142 * __ice_chk_linearize - Check if there are more than 8 buffers per packet
2143 * @skb: send buffer
2144 *
2145 * Note: This HW can't DMA more than 8 buffers to build a packet on the wire
2146 * and so we need to figure out the cases where we need to linearize the skb.
2147 *
2148 * For TSO we need to count the TSO header and segment payload separately.
2149 * As such we need to check cases where we have 7 fragments or more as we
2150 * can potentially require 9 DMA transactions, 1 for the TSO header, 1 for
2151 * the segment payload in the first descriptor, and another 7 for the
2152 * fragments.
2153 */
2154static bool __ice_chk_linearize(struct sk_buff *skb)
2155{
2156	const skb_frag_t *frag, *stale;
2157	int nr_frags, sum;
2158
2159	/* no need to check if number of frags is less than 7 */
2160	nr_frags = skb_shinfo(skb)->nr_frags;
2161	if (nr_frags < (ICE_MAX_BUF_TXD - 1))
2162		return false;
2163
2164	/* We need to walk through the list and validate that each group
2165	 * of 6 fragments totals at least gso_size.
2166	 */
2167	nr_frags -= ICE_MAX_BUF_TXD - 2;
2168	frag = &skb_shinfo(skb)->frags[0];
2169
2170	/* Initialize size to the negative value of gso_size minus 1. We
2171	 * use this as the worst case scenario in which the frag ahead
2172	 * of us only provides one byte which is why we are limited to 6
2173	 * descriptors for a single transmit as the header and previous
2174	 * fragment are already consuming 2 descriptors.
2175	 */
2176	sum = 1 - skb_shinfo(skb)->gso_size;
2177
2178	/* Add size of frags 0 through 4 to create our initial sum */
2179	sum += skb_frag_size(frag++);
2180	sum += skb_frag_size(frag++);
2181	sum += skb_frag_size(frag++);
2182	sum += skb_frag_size(frag++);
2183	sum += skb_frag_size(frag++);
2184
2185	/* Walk through fragments adding latest fragment, testing it, and
2186	 * then removing stale fragments from the sum.
2187	 */
2188	for (stale = &skb_shinfo(skb)->frags[0];; stale++) {
2189		int stale_size = skb_frag_size(stale);
2190
2191		sum += skb_frag_size(frag++);
2192
2193		/* The stale fragment may present us with a smaller
2194		 * descriptor than the actual fragment size. To account
2195		 * for that we need to remove all the data on the front and
2196		 * figure out what the remainder would be in the last
2197		 * descriptor associated with the fragment.
2198		 */
2199		if (stale_size > ICE_MAX_DATA_PER_TXD) {
2200			int align_pad = -(skb_frag_off(stale)) &
2201					(ICE_MAX_READ_REQ_SIZE - 1);
2202
2203			sum -= align_pad;
2204			stale_size -= align_pad;
2205
2206			do {
2207				sum -= ICE_MAX_DATA_PER_TXD_ALIGNED;
2208				stale_size -= ICE_MAX_DATA_PER_TXD_ALIGNED;
2209			} while (stale_size > ICE_MAX_DATA_PER_TXD);
2210		}
2211
2212		/* if sum is negative we failed to make sufficient progress */
2213		if (sum < 0)
2214			return true;
2215
2216		if (!nr_frags--)
2217			break;
2218
2219		sum -= stale_size;
2220	}
2221
2222	return false;
2223}
2224
2225/**
2226 * ice_chk_linearize - Check if there are more than 8 fragments per packet
2227 * @skb:      send buffer
2228 * @count:    number of buffers used
2229 *
2230 * Note: Our HW can't scatter-gather more than 8 fragments to build
2231 * a packet on the wire and so we need to figure out the cases where we
2232 * need to linearize the skb.
2233 */
2234static bool ice_chk_linearize(struct sk_buff *skb, unsigned int count)
2235{
2236	/* Both TSO and single send will work if count is less than 8 */
2237	if (likely(count < ICE_MAX_BUF_TXD))
2238		return false;
2239
2240	if (skb_is_gso(skb))
2241		return __ice_chk_linearize(skb);
2242
2243	/* we can support up to 8 data buffers for a single send */
2244	return count != ICE_MAX_BUF_TXD;
2245}
2246
2247/**
2248 * ice_tstamp - set up context descriptor for hardware timestamp
2249 * @tx_ring: pointer to the Tx ring to send buffer on
2250 * @skb: pointer to the SKB we're sending
2251 * @first: Tx buffer
2252 * @off: Tx offload parameters
2253 */
2254static void
2255ice_tstamp(struct ice_tx_ring *tx_ring, struct sk_buff *skb,
2256	   struct ice_tx_buf *first, struct ice_tx_offload_params *off)
2257{
2258	s8 idx;
2259
2260	/* only timestamp the outbound packet if the user has requested it */
2261	if (likely(!(skb_shinfo(skb)->tx_flags & SKBTX_HW_TSTAMP)))
2262		return;
2263
2264	if (!tx_ring->ptp_tx)
2265		return;
2266
2267	/* Tx timestamps cannot be sampled when doing TSO */
2268	if (first->tx_flags & ICE_TX_FLAGS_TSO)
2269		return;
2270
2271	/* Grab an open timestamp slot */
2272	idx = ice_ptp_request_ts(tx_ring->tx_tstamps, skb);
2273	if (idx < 0) {
2274		tx_ring->vsi->back->ptp.tx_hwtstamp_skipped++;
2275		return;
2276	}
2277
2278	off->cd_qw1 |= (u64)(ICE_TX_DESC_DTYPE_CTX |
2279			     (ICE_TX_CTX_DESC_TSYN << ICE_TXD_CTX_QW1_CMD_S) |
2280			     ((u64)idx << ICE_TXD_CTX_QW1_TSO_LEN_S));
2281	first->tx_flags |= ICE_TX_FLAGS_TSYN;
2282}
2283
2284/**
2285 * ice_xmit_frame_ring - Sends buffer on Tx ring
2286 * @skb: send buffer
2287 * @tx_ring: ring to send buffer on
2288 *
2289 * Returns NETDEV_TX_OK if sent, else an error code
2290 */
2291static netdev_tx_t
2292ice_xmit_frame_ring(struct sk_buff *skb, struct ice_tx_ring *tx_ring)
2293{
2294	struct ice_tx_offload_params offload = { 0 };
2295	struct ice_vsi *vsi = tx_ring->vsi;
2296	struct ice_tx_buf *first;
2297	struct ethhdr *eth;
2298	unsigned int count;
2299	int tso, csum;
2300
2301	ice_trace(xmit_frame_ring, tx_ring, skb);
2302
 
 
 
2303	count = ice_xmit_desc_count(skb);
2304	if (ice_chk_linearize(skb, count)) {
2305		if (__skb_linearize(skb))
2306			goto out_drop;
2307		count = ice_txd_use_count(skb->len);
2308		tx_ring->ring_stats->tx_stats.tx_linearize++;
2309	}
2310
2311	/* need: 1 descriptor per page * PAGE_SIZE/ICE_MAX_DATA_PER_TXD,
2312	 *       + 1 desc for skb_head_len/ICE_MAX_DATA_PER_TXD,
2313	 *       + 4 desc gap to avoid the cache line where head is,
2314	 *       + 1 desc for context descriptor,
2315	 * otherwise try next time
2316	 */
2317	if (ice_maybe_stop_tx(tx_ring, count + ICE_DESCS_PER_CACHE_LINE +
2318			      ICE_DESCS_FOR_CTX_DESC)) {
2319		tx_ring->ring_stats->tx_stats.tx_busy++;
2320		return NETDEV_TX_BUSY;
2321	}
2322
2323	/* prefetch for bql data which is infrequently used */
2324	netdev_txq_bql_enqueue_prefetchw(txring_txq(tx_ring));
2325
2326	offload.tx_ring = tx_ring;
2327
2328	/* record the location of the first descriptor for this packet */
2329	first = &tx_ring->tx_buf[tx_ring->next_to_use];
2330	first->skb = skb;
 
2331	first->bytecount = max_t(unsigned int, skb->len, ETH_ZLEN);
2332	first->gso_segs = 1;
2333	first->tx_flags = 0;
2334
2335	/* prepare the VLAN tagging flags for Tx */
2336	ice_tx_prepare_vlan_flags(tx_ring, first);
2337	if (first->tx_flags & ICE_TX_FLAGS_HW_OUTER_SINGLE_VLAN) {
2338		offload.cd_qw1 |= (u64)(ICE_TX_DESC_DTYPE_CTX |
2339					(ICE_TX_CTX_DESC_IL2TAG2 <<
2340					ICE_TXD_CTX_QW1_CMD_S));
2341		offload.cd_l2tag2 = (first->tx_flags & ICE_TX_FLAGS_VLAN_M) >>
2342			ICE_TX_FLAGS_VLAN_S;
2343	}
2344
2345	/* set up TSO offload */
2346	tso = ice_tso(first, &offload);
2347	if (tso < 0)
2348		goto out_drop;
2349
2350	/* always set up Tx checksum offload */
2351	csum = ice_tx_csum(first, &offload);
2352	if (csum < 0)
2353		goto out_drop;
2354
2355	/* allow CONTROL frames egress from main VSI if FW LLDP disabled */
2356	eth = (struct ethhdr *)skb_mac_header(skb);
2357	if (unlikely((skb->priority == TC_PRIO_CONTROL ||
2358		      eth->h_proto == htons(ETH_P_LLDP)) &&
2359		     vsi->type == ICE_VSI_PF &&
2360		     vsi->port_info->qos_cfg.is_sw_lldp))
2361		offload.cd_qw1 |= (u64)(ICE_TX_DESC_DTYPE_CTX |
2362					ICE_TX_CTX_DESC_SWTCH_UPLINK <<
2363					ICE_TXD_CTX_QW1_CMD_S);
2364
2365	ice_tstamp(tx_ring, skb, first, &offload);
2366	if (ice_is_switchdev_running(vsi->back))
2367		ice_eswitch_set_target_vsi(skb, &offload);
2368
2369	if (offload.cd_qw1 & ICE_TX_DESC_DTYPE_CTX) {
2370		struct ice_tx_ctx_desc *cdesc;
2371		u16 i = tx_ring->next_to_use;
2372
2373		/* grab the next descriptor */
2374		cdesc = ICE_TX_CTX_DESC(tx_ring, i);
2375		i++;
2376		tx_ring->next_to_use = (i < tx_ring->count) ? i : 0;
2377
2378		/* setup context descriptor */
2379		cdesc->tunneling_params = cpu_to_le32(offload.cd_tunnel_params);
2380		cdesc->l2tag2 = cpu_to_le16(offload.cd_l2tag2);
2381		cdesc->rsvd = cpu_to_le16(0);
2382		cdesc->qw1 = cpu_to_le64(offload.cd_qw1);
2383	}
2384
2385	ice_tx_map(tx_ring, first, &offload);
2386	return NETDEV_TX_OK;
2387
2388out_drop:
2389	ice_trace(xmit_frame_ring_drop, tx_ring, skb);
2390	dev_kfree_skb_any(skb);
2391	return NETDEV_TX_OK;
2392}
2393
2394/**
2395 * ice_start_xmit - Selects the correct VSI and Tx queue to send buffer
2396 * @skb: send buffer
2397 * @netdev: network interface device structure
2398 *
2399 * Returns NETDEV_TX_OK if sent, else an error code
2400 */
2401netdev_tx_t ice_start_xmit(struct sk_buff *skb, struct net_device *netdev)
2402{
2403	struct ice_netdev_priv *np = netdev_priv(netdev);
2404	struct ice_vsi *vsi = np->vsi;
2405	struct ice_tx_ring *tx_ring;
2406
2407	tx_ring = vsi->tx_rings[skb->queue_mapping];
2408
2409	/* hardware can't handle really short frames, hardware padding works
2410	 * beyond this point
2411	 */
2412	if (skb_put_padto(skb, ICE_MIN_TX_LEN))
2413		return NETDEV_TX_OK;
2414
2415	return ice_xmit_frame_ring(skb, tx_ring);
2416}
2417
2418/**
2419 * ice_get_dscp_up - return the UP/TC value for a SKB
2420 * @dcbcfg: DCB config that contains DSCP to UP/TC mapping
2421 * @skb: SKB to query for info to determine UP/TC
2422 *
2423 * This function is to only be called when the PF is in L3 DSCP PFC mode
2424 */
2425static u8 ice_get_dscp_up(struct ice_dcbx_cfg *dcbcfg, struct sk_buff *skb)
2426{
2427	u8 dscp = 0;
2428
2429	if (skb->protocol == htons(ETH_P_IP))
2430		dscp = ipv4_get_dsfield(ip_hdr(skb)) >> 2;
2431	else if (skb->protocol == htons(ETH_P_IPV6))
2432		dscp = ipv6_get_dsfield(ipv6_hdr(skb)) >> 2;
2433
2434	return dcbcfg->dscp_map[dscp];
2435}
2436
2437u16
2438ice_select_queue(struct net_device *netdev, struct sk_buff *skb,
2439		 struct net_device *sb_dev)
2440{
2441	struct ice_pf *pf = ice_netdev_to_pf(netdev);
2442	struct ice_dcbx_cfg *dcbcfg;
2443
2444	dcbcfg = &pf->hw.port_info->qos_cfg.local_dcbx_cfg;
2445	if (dcbcfg->pfc_mode == ICE_QOS_MODE_DSCP)
2446		skb->priority = ice_get_dscp_up(dcbcfg, skb);
2447
2448	return netdev_pick_tx(netdev, skb, sb_dev);
2449}
2450
2451/**
2452 * ice_clean_ctrl_tx_irq - interrupt handler for flow director Tx queue
2453 * @tx_ring: tx_ring to clean
2454 */
2455void ice_clean_ctrl_tx_irq(struct ice_tx_ring *tx_ring)
2456{
2457	struct ice_vsi *vsi = tx_ring->vsi;
2458	s16 i = tx_ring->next_to_clean;
2459	int budget = ICE_DFLT_IRQ_WORK;
2460	struct ice_tx_desc *tx_desc;
2461	struct ice_tx_buf *tx_buf;
2462
2463	tx_buf = &tx_ring->tx_buf[i];
2464	tx_desc = ICE_TX_DESC(tx_ring, i);
2465	i -= tx_ring->count;
2466
2467	do {
2468		struct ice_tx_desc *eop_desc = tx_buf->next_to_watch;
2469
2470		/* if next_to_watch is not set then there is no pending work */
2471		if (!eop_desc)
2472			break;
2473
2474		/* prevent any other reads prior to eop_desc */
2475		smp_rmb();
2476
2477		/* if the descriptor isn't done, no work to do */
2478		if (!(eop_desc->cmd_type_offset_bsz &
2479		      cpu_to_le64(ICE_TX_DESC_DTYPE_DESC_DONE)))
2480			break;
2481
2482		/* clear next_to_watch to prevent false hangs */
2483		tx_buf->next_to_watch = NULL;
2484		tx_desc->buf_addr = 0;
2485		tx_desc->cmd_type_offset_bsz = 0;
2486
2487		/* move past filter desc */
2488		tx_buf++;
2489		tx_desc++;
2490		i++;
2491		if (unlikely(!i)) {
2492			i -= tx_ring->count;
2493			tx_buf = tx_ring->tx_buf;
2494			tx_desc = ICE_TX_DESC(tx_ring, 0);
2495		}
2496
2497		/* unmap the data header */
2498		if (dma_unmap_len(tx_buf, len))
2499			dma_unmap_single(tx_ring->dev,
2500					 dma_unmap_addr(tx_buf, dma),
2501					 dma_unmap_len(tx_buf, len),
2502					 DMA_TO_DEVICE);
2503		if (tx_buf->tx_flags & ICE_TX_FLAGS_DUMMY_PKT)
2504			devm_kfree(tx_ring->dev, tx_buf->raw_buf);
2505
2506		/* clear next_to_watch to prevent false hangs */
2507		tx_buf->raw_buf = NULL;
2508		tx_buf->tx_flags = 0;
2509		tx_buf->next_to_watch = NULL;
2510		dma_unmap_len_set(tx_buf, len, 0);
2511		tx_desc->buf_addr = 0;
2512		tx_desc->cmd_type_offset_bsz = 0;
2513
2514		/* move past eop_desc for start of next FD desc */
2515		tx_buf++;
2516		tx_desc++;
2517		i++;
2518		if (unlikely(!i)) {
2519			i -= tx_ring->count;
2520			tx_buf = tx_ring->tx_buf;
2521			tx_desc = ICE_TX_DESC(tx_ring, 0);
2522		}
2523
2524		budget--;
2525	} while (likely(budget));
2526
2527	i += tx_ring->count;
2528	tx_ring->next_to_clean = i;
2529
2530	/* re-enable interrupt if needed */
2531	ice_irq_dynamic_ena(&vsi->back->hw, vsi, vsi->q_vectors[0]);
2532}