Linux Audio

Check our new training course

Loading...
v6.13.7
   1// SPDX-License-Identifier: GPL-2.0
   2/* Copyright(c) 2013 - 2019 Intel Corporation. */
   3
   4#include <linux/types.h>
   5#include <linux/module.h>
   6#include <net/ipv6.h>
   7#include <net/ip.h>
   8#include <net/tcp.h>
   9#include <linux/if_macvlan.h>
  10#include <linux/prefetch.h>
  11
  12#include "fm10k.h"
  13
  14#define DRV_SUMMARY	"Intel(R) Ethernet Switch Host Interface Driver"
  15char fm10k_driver_name[] = "fm10k";
  16static const char fm10k_driver_string[] = DRV_SUMMARY;
  17static const char fm10k_copyright[] =
  18	"Copyright(c) 2013 - 2019 Intel Corporation.";
  19
 
  20MODULE_DESCRIPTION(DRV_SUMMARY);
  21MODULE_LICENSE("GPL v2");
  22
  23/* single workqueue for entire fm10k driver */
  24struct workqueue_struct *fm10k_workqueue;
  25
  26/**
  27 * fm10k_init_module - Driver Registration Routine
  28 *
  29 * fm10k_init_module is the first routine called when the driver is
  30 * loaded.  All it does is register with the PCI subsystem.
  31 **/
  32static int __init fm10k_init_module(void)
  33{
  34	int ret;
  35
  36	pr_info("%s\n", fm10k_driver_string);
  37	pr_info("%s\n", fm10k_copyright);
  38
  39	/* create driver workqueue */
  40	fm10k_workqueue = alloc_workqueue("%s", WQ_MEM_RECLAIM, 0,
  41					  fm10k_driver_name);
  42	if (!fm10k_workqueue)
  43		return -ENOMEM;
  44
  45	fm10k_dbg_init();
  46
  47	ret = fm10k_register_pci_driver();
  48	if (ret) {
  49		fm10k_dbg_exit();
  50		destroy_workqueue(fm10k_workqueue);
  51	}
  52
  53	return ret;
  54}
  55module_init(fm10k_init_module);
  56
  57/**
  58 * fm10k_exit_module - Driver Exit Cleanup Routine
  59 *
  60 * fm10k_exit_module is called just before the driver is removed
  61 * from memory.
  62 **/
  63static void __exit fm10k_exit_module(void)
  64{
  65	fm10k_unregister_pci_driver();
  66
  67	fm10k_dbg_exit();
  68
  69	/* destroy driver workqueue */
  70	destroy_workqueue(fm10k_workqueue);
  71}
  72module_exit(fm10k_exit_module);
  73
  74static bool fm10k_alloc_mapped_page(struct fm10k_ring *rx_ring,
  75				    struct fm10k_rx_buffer *bi)
  76{
  77	struct page *page = bi->page;
  78	dma_addr_t dma;
  79
  80	/* Only page will be NULL if buffer was consumed */
  81	if (likely(page))
  82		return true;
  83
  84	/* alloc new page for storage */
  85	page = dev_alloc_page();
  86	if (unlikely(!page)) {
  87		rx_ring->rx_stats.alloc_failed++;
  88		return false;
  89	}
  90
  91	/* map page for use */
  92	dma = dma_map_page(rx_ring->dev, page, 0, PAGE_SIZE, DMA_FROM_DEVICE);
  93
  94	/* if mapping failed free memory back to system since
  95	 * there isn't much point in holding memory we can't use
  96	 */
  97	if (dma_mapping_error(rx_ring->dev, dma)) {
  98		__free_page(page);
  99
 100		rx_ring->rx_stats.alloc_failed++;
 101		return false;
 102	}
 103
 104	bi->dma = dma;
 105	bi->page = page;
 106	bi->page_offset = 0;
 107
 108	return true;
 109}
 110
 111/**
 112 * fm10k_alloc_rx_buffers - Replace used receive buffers
 113 * @rx_ring: ring to place buffers on
 114 * @cleaned_count: number of buffers to replace
 115 **/
 116void fm10k_alloc_rx_buffers(struct fm10k_ring *rx_ring, u16 cleaned_count)
 117{
 118	union fm10k_rx_desc *rx_desc;
 119	struct fm10k_rx_buffer *bi;
 120	u16 i = rx_ring->next_to_use;
 121
 122	/* nothing to do */
 123	if (!cleaned_count)
 124		return;
 125
 126	rx_desc = FM10K_RX_DESC(rx_ring, i);
 127	bi = &rx_ring->rx_buffer[i];
 128	i -= rx_ring->count;
 129
 130	do {
 131		if (!fm10k_alloc_mapped_page(rx_ring, bi))
 132			break;
 133
 134		/* Refresh the desc even if buffer_addrs didn't change
 135		 * because each write-back erases this info.
 136		 */
 137		rx_desc->q.pkt_addr = cpu_to_le64(bi->dma + bi->page_offset);
 138
 139		rx_desc++;
 140		bi++;
 141		i++;
 142		if (unlikely(!i)) {
 143			rx_desc = FM10K_RX_DESC(rx_ring, 0);
 144			bi = rx_ring->rx_buffer;
 145			i -= rx_ring->count;
 146		}
 147
 148		/* clear the status bits for the next_to_use descriptor */
 149		rx_desc->d.staterr = 0;
 150
 151		cleaned_count--;
 152	} while (cleaned_count);
 153
 154	i += rx_ring->count;
 155
 156	if (rx_ring->next_to_use != i) {
 157		/* record the next descriptor to use */
 158		rx_ring->next_to_use = i;
 159
 160		/* update next to alloc since we have filled the ring */
 161		rx_ring->next_to_alloc = i;
 162
 163		/* Force memory writes to complete before letting h/w
 164		 * know there are new descriptors to fetch.  (Only
 165		 * applicable for weak-ordered memory model archs,
 166		 * such as IA-64).
 167		 */
 168		wmb();
 169
 170		/* notify hardware of new descriptors */
 171		writel(i, rx_ring->tail);
 172	}
 173}
 174
 175/**
 176 * fm10k_reuse_rx_page - page flip buffer and store it back on the ring
 177 * @rx_ring: rx descriptor ring to store buffers on
 178 * @old_buff: donor buffer to have page reused
 179 *
 180 * Synchronizes page for reuse by the interface
 181 **/
 182static void fm10k_reuse_rx_page(struct fm10k_ring *rx_ring,
 183				struct fm10k_rx_buffer *old_buff)
 184{
 185	struct fm10k_rx_buffer *new_buff;
 186	u16 nta = rx_ring->next_to_alloc;
 187
 188	new_buff = &rx_ring->rx_buffer[nta];
 189
 190	/* update, and store next to alloc */
 191	nta++;
 192	rx_ring->next_to_alloc = (nta < rx_ring->count) ? nta : 0;
 193
 194	/* transfer page from old buffer to new buffer */
 195	*new_buff = *old_buff;
 196
 197	/* sync the buffer for use by the device */
 198	dma_sync_single_range_for_device(rx_ring->dev, old_buff->dma,
 199					 old_buff->page_offset,
 200					 FM10K_RX_BUFSZ,
 201					 DMA_FROM_DEVICE);
 202}
 203
 
 
 
 
 
 204static bool fm10k_can_reuse_rx_page(struct fm10k_rx_buffer *rx_buffer,
 205				    struct page *page,
 206				    unsigned int __maybe_unused truesize)
 207{
 208	/* avoid re-using remote and pfmemalloc pages */
 209	if (!dev_page_is_reusable(page))
 210		return false;
 211
 212#if (PAGE_SIZE < 8192)
 213	/* if we are only owner of page we can reuse it */
 214	if (unlikely(page_count(page) != 1))
 215		return false;
 216
 217	/* flip page offset to other buffer */
 218	rx_buffer->page_offset ^= FM10K_RX_BUFSZ;
 219#else
 220	/* move offset up to the next cache line */
 221	rx_buffer->page_offset += truesize;
 222
 223	if (rx_buffer->page_offset > (PAGE_SIZE - FM10K_RX_BUFSZ))
 224		return false;
 225#endif
 226
 227	/* Even if we own the page, we are not allowed to use atomic_set()
 228	 * This would break get_page_unless_zero() users.
 229	 */
 230	page_ref_inc(page);
 231
 232	return true;
 233}
 234
 235/**
 236 * fm10k_add_rx_frag - Add contents of Rx buffer to sk_buff
 237 * @rx_buffer: buffer containing page to add
 238 * @size: packet size from rx_desc
 239 * @rx_desc: descriptor containing length of buffer written by hardware
 240 * @skb: sk_buff to place the data into
 241 *
 242 * This function will add the data contained in rx_buffer->page to the skb.
 243 * This is done either through a direct copy if the data in the buffer is
 244 * less than the skb header size, otherwise it will just attach the page as
 245 * a frag to the skb.
 246 *
 247 * The function will then update the page offset if necessary and return
 248 * true if the buffer can be reused by the interface.
 249 **/
 250static bool fm10k_add_rx_frag(struct fm10k_rx_buffer *rx_buffer,
 251			      unsigned int size,
 252			      union fm10k_rx_desc *rx_desc,
 253			      struct sk_buff *skb)
 254{
 255	struct page *page = rx_buffer->page;
 256	unsigned char *va = page_address(page) + rx_buffer->page_offset;
 257#if (PAGE_SIZE < 8192)
 258	unsigned int truesize = FM10K_RX_BUFSZ;
 259#else
 260	unsigned int truesize = ALIGN(size, 512);
 261#endif
 262	unsigned int pull_len;
 263
 264	if (unlikely(skb_is_nonlinear(skb)))
 265		goto add_tail_frag;
 266
 267	if (likely(size <= FM10K_RX_HDR_LEN)) {
 268		memcpy(__skb_put(skb, size), va, ALIGN(size, sizeof(long)));
 269
 270		/* page is reusable, we can reuse buffer as-is */
 271		if (dev_page_is_reusable(page))
 272			return true;
 273
 274		/* this page cannot be reused so discard it */
 275		__free_page(page);
 276		return false;
 277	}
 278
 279	/* we need the header to contain the greater of either ETH_HLEN or
 280	 * 60 bytes if the skb->len is less than 60 for skb_pad.
 281	 */
 282	pull_len = eth_get_headlen(skb->dev, va, FM10K_RX_HDR_LEN);
 283
 284	/* align pull length to size of long to optimize memcpy performance */
 285	memcpy(__skb_put(skb, pull_len), va, ALIGN(pull_len, sizeof(long)));
 286
 287	/* update all of the pointers */
 288	va += pull_len;
 289	size -= pull_len;
 290
 291add_tail_frag:
 292	skb_add_rx_frag(skb, skb_shinfo(skb)->nr_frags, page,
 293			(unsigned long)va & ~PAGE_MASK, size, truesize);
 294
 295	return fm10k_can_reuse_rx_page(rx_buffer, page, truesize);
 296}
 297
 298static struct sk_buff *fm10k_fetch_rx_buffer(struct fm10k_ring *rx_ring,
 299					     union fm10k_rx_desc *rx_desc,
 300					     struct sk_buff *skb)
 301{
 302	unsigned int size = le16_to_cpu(rx_desc->w.length);
 303	struct fm10k_rx_buffer *rx_buffer;
 304	struct page *page;
 305
 306	rx_buffer = &rx_ring->rx_buffer[rx_ring->next_to_clean];
 307	page = rx_buffer->page;
 308	prefetchw(page);
 309
 310	if (likely(!skb)) {
 311		void *page_addr = page_address(page) +
 312				  rx_buffer->page_offset;
 313
 314		/* prefetch first cache line of first page */
 315		net_prefetch(page_addr);
 
 
 
 316
 317		/* allocate a skb to store the frags */
 318		skb = napi_alloc_skb(&rx_ring->q_vector->napi,
 319				     FM10K_RX_HDR_LEN);
 320		if (unlikely(!skb)) {
 321			rx_ring->rx_stats.alloc_failed++;
 322			return NULL;
 323		}
 324
 325		/* we will be copying header into skb->data in
 326		 * pskb_may_pull so it is in our interest to prefetch
 327		 * it now to avoid a possible cache miss
 328		 */
 329		prefetchw(skb->data);
 330	}
 331
 332	/* we are reusing so sync this buffer for CPU use */
 333	dma_sync_single_range_for_cpu(rx_ring->dev,
 334				      rx_buffer->dma,
 335				      rx_buffer->page_offset,
 336				      size,
 337				      DMA_FROM_DEVICE);
 338
 339	/* pull page into skb */
 340	if (fm10k_add_rx_frag(rx_buffer, size, rx_desc, skb)) {
 341		/* hand second half of page back to the ring */
 342		fm10k_reuse_rx_page(rx_ring, rx_buffer);
 343	} else {
 344		/* we are not reusing the buffer so unmap it */
 345		dma_unmap_page(rx_ring->dev, rx_buffer->dma,
 346			       PAGE_SIZE, DMA_FROM_DEVICE);
 347	}
 348
 349	/* clear contents of rx_buffer */
 350	rx_buffer->page = NULL;
 351
 352	return skb;
 353}
 354
 355static inline void fm10k_rx_checksum(struct fm10k_ring *ring,
 356				     union fm10k_rx_desc *rx_desc,
 357				     struct sk_buff *skb)
 358{
 359	skb_checksum_none_assert(skb);
 360
 361	/* Rx checksum disabled via ethtool */
 362	if (!(ring->netdev->features & NETIF_F_RXCSUM))
 363		return;
 364
 365	/* TCP/UDP checksum error bit is set */
 366	if (fm10k_test_staterr(rx_desc,
 367			       FM10K_RXD_STATUS_L4E |
 368			       FM10K_RXD_STATUS_L4E2 |
 369			       FM10K_RXD_STATUS_IPE |
 370			       FM10K_RXD_STATUS_IPE2)) {
 371		ring->rx_stats.csum_err++;
 372		return;
 373	}
 374
 375	/* It must be a TCP or UDP packet with a valid checksum */
 376	if (fm10k_test_staterr(rx_desc, FM10K_RXD_STATUS_L4CS2))
 377		skb->encapsulation = true;
 378	else if (!fm10k_test_staterr(rx_desc, FM10K_RXD_STATUS_L4CS))
 379		return;
 380
 381	skb->ip_summed = CHECKSUM_UNNECESSARY;
 382
 383	ring->rx_stats.csum_good++;
 384}
 385
 386#define FM10K_RSS_L4_TYPES_MASK \
 387	(BIT(FM10K_RSSTYPE_IPV4_TCP) | \
 388	 BIT(FM10K_RSSTYPE_IPV4_UDP) | \
 389	 BIT(FM10K_RSSTYPE_IPV6_TCP) | \
 390	 BIT(FM10K_RSSTYPE_IPV6_UDP))
 391
 392static inline void fm10k_rx_hash(struct fm10k_ring *ring,
 393				 union fm10k_rx_desc *rx_desc,
 394				 struct sk_buff *skb)
 395{
 396	u16 rss_type;
 397
 398	if (!(ring->netdev->features & NETIF_F_RXHASH))
 399		return;
 400
 401	rss_type = le16_to_cpu(rx_desc->w.pkt_info) & FM10K_RXD_RSSTYPE_MASK;
 402	if (!rss_type)
 403		return;
 404
 405	skb_set_hash(skb, le32_to_cpu(rx_desc->d.rss),
 406		     (BIT(rss_type) & FM10K_RSS_L4_TYPES_MASK) ?
 407		     PKT_HASH_TYPE_L4 : PKT_HASH_TYPE_L3);
 408}
 409
 410static void fm10k_type_trans(struct fm10k_ring *rx_ring,
 411			     union fm10k_rx_desc __maybe_unused *rx_desc,
 412			     struct sk_buff *skb)
 413{
 414	struct net_device *dev = rx_ring->netdev;
 415	struct fm10k_l2_accel *l2_accel = rcu_dereference_bh(rx_ring->l2_accel);
 416
 417	/* check to see if DGLORT belongs to a MACVLAN */
 418	if (l2_accel) {
 419		u16 idx = le16_to_cpu(FM10K_CB(skb)->fi.w.dglort) - 1;
 420
 421		idx -= l2_accel->dglort;
 422		if (idx < l2_accel->size && l2_accel->macvlan[idx])
 423			dev = l2_accel->macvlan[idx];
 424		else
 425			l2_accel = NULL;
 426	}
 427
 428	/* Record Rx queue, or update macvlan statistics */
 429	if (!l2_accel)
 430		skb_record_rx_queue(skb, rx_ring->queue_index);
 431	else
 432		macvlan_count_rx(netdev_priv(dev), skb->len + ETH_HLEN, true,
 433				 false);
 434
 435	skb->protocol = eth_type_trans(skb, dev);
 436}
 437
 438/**
 439 * fm10k_process_skb_fields - Populate skb header fields from Rx descriptor
 440 * @rx_ring: rx descriptor ring packet is being transacted on
 441 * @rx_desc: pointer to the EOP Rx descriptor
 442 * @skb: pointer to current skb being populated
 443 *
 444 * This function checks the ring, descriptor, and packet information in
 445 * order to populate the hash, checksum, VLAN, timestamp, protocol, and
 446 * other fields within the skb.
 447 **/
 448static unsigned int fm10k_process_skb_fields(struct fm10k_ring *rx_ring,
 449					     union fm10k_rx_desc *rx_desc,
 450					     struct sk_buff *skb)
 451{
 452	unsigned int len = skb->len;
 453
 454	fm10k_rx_hash(rx_ring, rx_desc, skb);
 455
 456	fm10k_rx_checksum(rx_ring, rx_desc, skb);
 457
 458	FM10K_CB(skb)->tstamp = rx_desc->q.timestamp;
 459
 460	FM10K_CB(skb)->fi.w.vlan = rx_desc->w.vlan;
 461
 462	FM10K_CB(skb)->fi.d.glort = rx_desc->d.glort;
 463
 464	if (rx_desc->w.vlan) {
 465		u16 vid = le16_to_cpu(rx_desc->w.vlan);
 466
 467		if ((vid & VLAN_VID_MASK) != rx_ring->vid)
 468			__vlan_hwaccel_put_tag(skb, htons(ETH_P_8021Q), vid);
 469		else if (vid & VLAN_PRIO_MASK)
 470			__vlan_hwaccel_put_tag(skb, htons(ETH_P_8021Q),
 471					       vid & VLAN_PRIO_MASK);
 472	}
 473
 474	fm10k_type_trans(rx_ring, rx_desc, skb);
 475
 476	return len;
 477}
 478
 479/**
 480 * fm10k_is_non_eop - process handling of non-EOP buffers
 481 * @rx_ring: Rx ring being processed
 482 * @rx_desc: Rx descriptor for current buffer
 483 *
 484 * This function updates next to clean.  If the buffer is an EOP buffer
 485 * this function exits returning false, otherwise it will place the
 486 * sk_buff in the next buffer to be chained and return true indicating
 487 * that this is in fact a non-EOP buffer.
 488 **/
 489static bool fm10k_is_non_eop(struct fm10k_ring *rx_ring,
 490			     union fm10k_rx_desc *rx_desc)
 491{
 492	u32 ntc = rx_ring->next_to_clean + 1;
 493
 494	/* fetch, update, and store next to clean */
 495	ntc = (ntc < rx_ring->count) ? ntc : 0;
 496	rx_ring->next_to_clean = ntc;
 497
 498	prefetch(FM10K_RX_DESC(rx_ring, ntc));
 499
 500	if (likely(fm10k_test_staterr(rx_desc, FM10K_RXD_STATUS_EOP)))
 501		return false;
 502
 503	return true;
 504}
 505
 506/**
 507 * fm10k_cleanup_headers - Correct corrupted or empty headers
 508 * @rx_ring: rx descriptor ring packet is being transacted on
 509 * @rx_desc: pointer to the EOP Rx descriptor
 510 * @skb: pointer to current skb being fixed
 511 *
 512 * Address the case where we are pulling data in on pages only
 513 * and as such no data is present in the skb header.
 514 *
 515 * In addition if skb is not at least 60 bytes we need to pad it so that
 516 * it is large enough to qualify as a valid Ethernet frame.
 517 *
 518 * Returns true if an error was encountered and skb was freed.
 519 **/
 520static bool fm10k_cleanup_headers(struct fm10k_ring *rx_ring,
 521				  union fm10k_rx_desc *rx_desc,
 522				  struct sk_buff *skb)
 523{
 524	if (unlikely((fm10k_test_staterr(rx_desc,
 525					 FM10K_RXD_STATUS_RXE)))) {
 526#define FM10K_TEST_RXD_BIT(rxd, bit) \
 527	((rxd)->w.csum_err & cpu_to_le16(bit))
 528		if (FM10K_TEST_RXD_BIT(rx_desc, FM10K_RXD_ERR_SWITCH_ERROR))
 529			rx_ring->rx_stats.switch_errors++;
 530		if (FM10K_TEST_RXD_BIT(rx_desc, FM10K_RXD_ERR_NO_DESCRIPTOR))
 531			rx_ring->rx_stats.drops++;
 532		if (FM10K_TEST_RXD_BIT(rx_desc, FM10K_RXD_ERR_PP_ERROR))
 533			rx_ring->rx_stats.pp_errors++;
 534		if (FM10K_TEST_RXD_BIT(rx_desc, FM10K_RXD_ERR_SWITCH_READY))
 535			rx_ring->rx_stats.link_errors++;
 536		if (FM10K_TEST_RXD_BIT(rx_desc, FM10K_RXD_ERR_TOO_BIG))
 537			rx_ring->rx_stats.length_errors++;
 538		dev_kfree_skb_any(skb);
 539		rx_ring->rx_stats.errors++;
 540		return true;
 541	}
 542
 543	/* if eth_skb_pad returns an error the skb was freed */
 544	if (eth_skb_pad(skb))
 545		return true;
 546
 547	return false;
 548}
 549
 550/**
 551 * fm10k_receive_skb - helper function to handle rx indications
 552 * @q_vector: structure containing interrupt and ring information
 553 * @skb: packet to send up
 554 **/
 555static void fm10k_receive_skb(struct fm10k_q_vector *q_vector,
 556			      struct sk_buff *skb)
 557{
 558	napi_gro_receive(&q_vector->napi, skb);
 559}
 560
 561static int fm10k_clean_rx_irq(struct fm10k_q_vector *q_vector,
 562			      struct fm10k_ring *rx_ring,
 563			      int budget)
 564{
 565	struct sk_buff *skb = rx_ring->skb;
 566	unsigned int total_bytes = 0, total_packets = 0;
 567	u16 cleaned_count = fm10k_desc_unused(rx_ring);
 568
 569	while (likely(total_packets < budget)) {
 570		union fm10k_rx_desc *rx_desc;
 571
 572		/* return some buffers to hardware, one at a time is too slow */
 573		if (cleaned_count >= FM10K_RX_BUFFER_WRITE) {
 574			fm10k_alloc_rx_buffers(rx_ring, cleaned_count);
 575			cleaned_count = 0;
 576		}
 577
 578		rx_desc = FM10K_RX_DESC(rx_ring, rx_ring->next_to_clean);
 579
 580		if (!rx_desc->d.staterr)
 581			break;
 582
 583		/* This memory barrier is needed to keep us from reading
 584		 * any other fields out of the rx_desc until we know the
 585		 * descriptor has been written back
 586		 */
 587		dma_rmb();
 588
 589		/* retrieve a buffer from the ring */
 590		skb = fm10k_fetch_rx_buffer(rx_ring, rx_desc, skb);
 591
 592		/* exit if we failed to retrieve a buffer */
 593		if (!skb)
 594			break;
 595
 596		cleaned_count++;
 597
 598		/* fetch next buffer in frame if non-eop */
 599		if (fm10k_is_non_eop(rx_ring, rx_desc))
 600			continue;
 601
 602		/* verify the packet layout is correct */
 603		if (fm10k_cleanup_headers(rx_ring, rx_desc, skb)) {
 604			skb = NULL;
 605			continue;
 606		}
 607
 608		/* populate checksum, timestamp, VLAN, and protocol */
 609		total_bytes += fm10k_process_skb_fields(rx_ring, rx_desc, skb);
 610
 611		fm10k_receive_skb(q_vector, skb);
 612
 613		/* reset skb pointer */
 614		skb = NULL;
 615
 616		/* update budget accounting */
 617		total_packets++;
 618	}
 619
 620	/* place incomplete frames back on ring for completion */
 621	rx_ring->skb = skb;
 622
 623	u64_stats_update_begin(&rx_ring->syncp);
 624	rx_ring->stats.packets += total_packets;
 625	rx_ring->stats.bytes += total_bytes;
 626	u64_stats_update_end(&rx_ring->syncp);
 627	q_vector->rx.total_packets += total_packets;
 628	q_vector->rx.total_bytes += total_bytes;
 629
 630	return total_packets;
 631}
 632
 633#define VXLAN_HLEN (sizeof(struct udphdr) + 8)
 634static struct ethhdr *fm10k_port_is_vxlan(struct sk_buff *skb)
 635{
 636	struct fm10k_intfc *interface = netdev_priv(skb->dev);
 637
 638	if (interface->vxlan_port != udp_hdr(skb)->dest)
 639		return NULL;
 640
 641	/* return offset of udp_hdr plus 8 bytes for VXLAN header */
 642	return (struct ethhdr *)(skb_transport_header(skb) + VXLAN_HLEN);
 643}
 644
 645#define FM10K_NVGRE_RESERVED0_FLAGS htons(0x9FFF)
 646#define NVGRE_TNI htons(0x2000)
 647struct fm10k_nvgre_hdr {
 648	__be16 flags;
 649	__be16 proto;
 650	__be32 tni;
 651};
 652
 653static struct ethhdr *fm10k_gre_is_nvgre(struct sk_buff *skb)
 654{
 655	struct fm10k_nvgre_hdr *nvgre_hdr;
 656	int hlen = ip_hdrlen(skb);
 657
 658	/* currently only IPv4 is supported due to hlen above */
 659	if (vlan_get_protocol(skb) != htons(ETH_P_IP))
 660		return NULL;
 661
 662	/* our transport header should be NVGRE */
 663	nvgre_hdr = (struct fm10k_nvgre_hdr *)(skb_network_header(skb) + hlen);
 664
 665	/* verify all reserved flags are 0 */
 666	if (nvgre_hdr->flags & FM10K_NVGRE_RESERVED0_FLAGS)
 667		return NULL;
 668
 669	/* report start of ethernet header */
 670	if (nvgre_hdr->flags & NVGRE_TNI)
 671		return (struct ethhdr *)(nvgre_hdr + 1);
 672
 673	return (struct ethhdr *)(&nvgre_hdr->tni);
 674}
 675
 676__be16 fm10k_tx_encap_offload(struct sk_buff *skb)
 677{
 678	u8 l4_hdr = 0, inner_l4_hdr = 0, inner_l4_hlen;
 679	struct ethhdr *eth_hdr;
 680
 681	if (skb->inner_protocol_type != ENCAP_TYPE_ETHER ||
 682	    skb->inner_protocol != htons(ETH_P_TEB))
 683		return 0;
 684
 685	switch (vlan_get_protocol(skb)) {
 686	case htons(ETH_P_IP):
 687		l4_hdr = ip_hdr(skb)->protocol;
 688		break;
 689	case htons(ETH_P_IPV6):
 690		l4_hdr = ipv6_hdr(skb)->nexthdr;
 691		break;
 692	default:
 693		return 0;
 694	}
 695
 696	switch (l4_hdr) {
 697	case IPPROTO_UDP:
 698		eth_hdr = fm10k_port_is_vxlan(skb);
 699		break;
 700	case IPPROTO_GRE:
 701		eth_hdr = fm10k_gre_is_nvgre(skb);
 702		break;
 703	default:
 704		return 0;
 705	}
 706
 707	if (!eth_hdr)
 708		return 0;
 709
 710	switch (eth_hdr->h_proto) {
 711	case htons(ETH_P_IP):
 712		inner_l4_hdr = inner_ip_hdr(skb)->protocol;
 713		break;
 714	case htons(ETH_P_IPV6):
 715		inner_l4_hdr = inner_ipv6_hdr(skb)->nexthdr;
 716		break;
 717	default:
 718		return 0;
 719	}
 720
 721	switch (inner_l4_hdr) {
 722	case IPPROTO_TCP:
 723		inner_l4_hlen = inner_tcp_hdrlen(skb);
 724		break;
 725	case IPPROTO_UDP:
 726		inner_l4_hlen = 8;
 727		break;
 728	default:
 729		return 0;
 730	}
 731
 732	/* The hardware allows tunnel offloads only if the combined inner and
 733	 * outer header is 184 bytes or less
 734	 */
 735	if (skb_inner_transport_header(skb) + inner_l4_hlen -
 736	    skb_mac_header(skb) > FM10K_TUNNEL_HEADER_LENGTH)
 737		return 0;
 738
 739	return eth_hdr->h_proto;
 740}
 741
 742static int fm10k_tso(struct fm10k_ring *tx_ring,
 743		     struct fm10k_tx_buffer *first)
 744{
 745	struct sk_buff *skb = first->skb;
 746	struct fm10k_tx_desc *tx_desc;
 747	unsigned char *th;
 748	u8 hdrlen;
 749
 750	if (skb->ip_summed != CHECKSUM_PARTIAL)
 751		return 0;
 752
 753	if (!skb_is_gso(skb))
 754		return 0;
 755
 756	/* compute header lengths */
 757	if (skb->encapsulation) {
 758		if (!fm10k_tx_encap_offload(skb))
 759			goto err_vxlan;
 760		th = skb_inner_transport_header(skb);
 761	} else {
 762		th = skb_transport_header(skb);
 763	}
 764
 765	/* compute offset from SOF to transport header and add header len */
 766	hdrlen = (th - skb->data) + (((struct tcphdr *)th)->doff << 2);
 767
 768	first->tx_flags |= FM10K_TX_FLAGS_CSUM;
 769
 770	/* update gso size and bytecount with header size */
 771	first->gso_segs = skb_shinfo(skb)->gso_segs;
 772	first->bytecount += (first->gso_segs - 1) * hdrlen;
 773
 774	/* populate Tx descriptor header size and mss */
 775	tx_desc = FM10K_TX_DESC(tx_ring, tx_ring->next_to_use);
 776	tx_desc->hdrlen = hdrlen;
 777	tx_desc->mss = cpu_to_le16(skb_shinfo(skb)->gso_size);
 778
 779	return 1;
 780
 781err_vxlan:
 782	tx_ring->netdev->features &= ~NETIF_F_GSO_UDP_TUNNEL;
 783	if (net_ratelimit())
 784		netdev_err(tx_ring->netdev,
 785			   "TSO requested for unsupported tunnel, disabling offload\n");
 786	return -1;
 787}
 788
 789static void fm10k_tx_csum(struct fm10k_ring *tx_ring,
 790			  struct fm10k_tx_buffer *first)
 791{
 792	struct sk_buff *skb = first->skb;
 793	struct fm10k_tx_desc *tx_desc;
 794	union {
 795		struct iphdr *ipv4;
 796		struct ipv6hdr *ipv6;
 797		u8 *raw;
 798	} network_hdr;
 799	u8 *transport_hdr;
 800	__be16 frag_off;
 801	__be16 protocol;
 802	u8 l4_hdr = 0;
 803
 804	if (skb->ip_summed != CHECKSUM_PARTIAL)
 805		goto no_csum;
 806
 807	if (skb->encapsulation) {
 808		protocol = fm10k_tx_encap_offload(skb);
 809		if (!protocol) {
 810			if (skb_checksum_help(skb)) {
 811				dev_warn(tx_ring->dev,
 812					 "failed to offload encap csum!\n");
 813				tx_ring->tx_stats.csum_err++;
 814			}
 815			goto no_csum;
 816		}
 817		network_hdr.raw = skb_inner_network_header(skb);
 818		transport_hdr = skb_inner_transport_header(skb);
 819	} else {
 820		protocol = vlan_get_protocol(skb);
 821		network_hdr.raw = skb_network_header(skb);
 822		transport_hdr = skb_transport_header(skb);
 823	}
 824
 825	switch (protocol) {
 826	case htons(ETH_P_IP):
 827		l4_hdr = network_hdr.ipv4->protocol;
 828		break;
 829	case htons(ETH_P_IPV6):
 830		l4_hdr = network_hdr.ipv6->nexthdr;
 831		if (likely((transport_hdr - network_hdr.raw) ==
 832			   sizeof(struct ipv6hdr)))
 833			break;
 834		ipv6_skip_exthdr(skb, network_hdr.raw - skb->data +
 835				      sizeof(struct ipv6hdr),
 836				 &l4_hdr, &frag_off);
 837		if (unlikely(frag_off))
 838			l4_hdr = NEXTHDR_FRAGMENT;
 839		break;
 840	default:
 841		break;
 842	}
 843
 844	switch (l4_hdr) {
 845	case IPPROTO_TCP:
 846	case IPPROTO_UDP:
 847		break;
 848	case IPPROTO_GRE:
 849		if (skb->encapsulation)
 850			break;
 851		fallthrough;
 852	default:
 853		if (unlikely(net_ratelimit())) {
 854			dev_warn(tx_ring->dev,
 855				 "partial checksum, version=%d l4 proto=%x\n",
 856				 protocol, l4_hdr);
 857		}
 858		skb_checksum_help(skb);
 859		tx_ring->tx_stats.csum_err++;
 860		goto no_csum;
 861	}
 862
 863	/* update TX checksum flag */
 864	first->tx_flags |= FM10K_TX_FLAGS_CSUM;
 865	tx_ring->tx_stats.csum_good++;
 866
 867no_csum:
 868	/* populate Tx descriptor header size and mss */
 869	tx_desc = FM10K_TX_DESC(tx_ring, tx_ring->next_to_use);
 870	tx_desc->hdrlen = 0;
 871	tx_desc->mss = 0;
 872}
 873
 874#define FM10K_SET_FLAG(_input, _flag, _result) \
 875	((_flag <= _result) ? \
 876	 ((u32)(_input & _flag) * (_result / _flag)) : \
 877	 ((u32)(_input & _flag) / (_flag / _result)))
 878
 879static u8 fm10k_tx_desc_flags(struct sk_buff *skb, u32 tx_flags)
 880{
 881	/* set type for advanced descriptor with frame checksum insertion */
 882	u32 desc_flags = 0;
 883
 884	/* set checksum offload bits */
 885	desc_flags |= FM10K_SET_FLAG(tx_flags, FM10K_TX_FLAGS_CSUM,
 886				     FM10K_TXD_FLAG_CSUM);
 887
 888	return desc_flags;
 889}
 890
 891static bool fm10k_tx_desc_push(struct fm10k_ring *tx_ring,
 892			       struct fm10k_tx_desc *tx_desc, u16 i,
 893			       dma_addr_t dma, unsigned int size, u8 desc_flags)
 894{
 895	/* set RS and INT for last frame in a cache line */
 896	if ((++i & (FM10K_TXD_WB_FIFO_SIZE - 1)) == 0)
 897		desc_flags |= FM10K_TXD_FLAG_RS | FM10K_TXD_FLAG_INT;
 898
 899	/* record values to descriptor */
 900	tx_desc->buffer_addr = cpu_to_le64(dma);
 901	tx_desc->flags = desc_flags;
 902	tx_desc->buflen = cpu_to_le16(size);
 903
 904	/* return true if we just wrapped the ring */
 905	return i == tx_ring->count;
 906}
 907
 908static int __fm10k_maybe_stop_tx(struct fm10k_ring *tx_ring, u16 size)
 909{
 910	netif_stop_subqueue(tx_ring->netdev, tx_ring->queue_index);
 911
 912	/* Memory barrier before checking head and tail */
 913	smp_mb();
 914
 915	/* Check again in a case another CPU has just made room available */
 916	if (likely(fm10k_desc_unused(tx_ring) < size))
 917		return -EBUSY;
 918
 919	/* A reprieve! - use start_queue because it doesn't call schedule */
 920	netif_start_subqueue(tx_ring->netdev, tx_ring->queue_index);
 921	++tx_ring->tx_stats.restart_queue;
 922	return 0;
 923}
 924
 925static inline int fm10k_maybe_stop_tx(struct fm10k_ring *tx_ring, u16 size)
 926{
 927	if (likely(fm10k_desc_unused(tx_ring) >= size))
 928		return 0;
 929	return __fm10k_maybe_stop_tx(tx_ring, size);
 930}
 931
 932static void fm10k_tx_map(struct fm10k_ring *tx_ring,
 933			 struct fm10k_tx_buffer *first)
 934{
 935	struct sk_buff *skb = first->skb;
 936	struct fm10k_tx_buffer *tx_buffer;
 937	struct fm10k_tx_desc *tx_desc;
 938	skb_frag_t *frag;
 939	unsigned char *data;
 940	dma_addr_t dma;
 941	unsigned int data_len, size;
 942	u32 tx_flags = first->tx_flags;
 943	u16 i = tx_ring->next_to_use;
 944	u8 flags = fm10k_tx_desc_flags(skb, tx_flags);
 945
 946	tx_desc = FM10K_TX_DESC(tx_ring, i);
 947
 948	/* add HW VLAN tag */
 949	if (skb_vlan_tag_present(skb))
 950		tx_desc->vlan = cpu_to_le16(skb_vlan_tag_get(skb));
 951	else
 952		tx_desc->vlan = 0;
 953
 954	size = skb_headlen(skb);
 955	data = skb->data;
 956
 957	dma = dma_map_single(tx_ring->dev, data, size, DMA_TO_DEVICE);
 958
 959	data_len = skb->data_len;
 960	tx_buffer = first;
 961
 962	for (frag = &skb_shinfo(skb)->frags[0];; frag++) {
 963		if (dma_mapping_error(tx_ring->dev, dma))
 964			goto dma_error;
 965
 966		/* record length, and DMA address */
 967		dma_unmap_len_set(tx_buffer, len, size);
 968		dma_unmap_addr_set(tx_buffer, dma, dma);
 969
 970		while (unlikely(size > FM10K_MAX_DATA_PER_TXD)) {
 971			if (fm10k_tx_desc_push(tx_ring, tx_desc++, i++, dma,
 972					       FM10K_MAX_DATA_PER_TXD, flags)) {
 973				tx_desc = FM10K_TX_DESC(tx_ring, 0);
 974				i = 0;
 975			}
 976
 977			dma += FM10K_MAX_DATA_PER_TXD;
 978			size -= FM10K_MAX_DATA_PER_TXD;
 979		}
 980
 981		if (likely(!data_len))
 982			break;
 983
 984		if (fm10k_tx_desc_push(tx_ring, tx_desc++, i++,
 985				       dma, size, flags)) {
 986			tx_desc = FM10K_TX_DESC(tx_ring, 0);
 987			i = 0;
 988		}
 989
 990		size = skb_frag_size(frag);
 991		data_len -= size;
 992
 993		dma = skb_frag_dma_map(tx_ring->dev, frag, 0, size,
 994				       DMA_TO_DEVICE);
 995
 996		tx_buffer = &tx_ring->tx_buffer[i];
 997	}
 998
 999	/* write last descriptor with LAST bit set */
1000	flags |= FM10K_TXD_FLAG_LAST;
1001
1002	if (fm10k_tx_desc_push(tx_ring, tx_desc, i++, dma, size, flags))
1003		i = 0;
1004
1005	/* record bytecount for BQL */
1006	netdev_tx_sent_queue(txring_txq(tx_ring), first->bytecount);
1007
1008	/* record SW timestamp if HW timestamp is not available */
1009	skb_tx_timestamp(first->skb);
1010
1011	/* Force memory writes to complete before letting h/w know there
1012	 * are new descriptors to fetch.  (Only applicable for weak-ordered
1013	 * memory model archs, such as IA-64).
1014	 *
1015	 * We also need this memory barrier to make certain all of the
1016	 * status bits have been updated before next_to_watch is written.
1017	 */
1018	wmb();
1019
1020	/* set next_to_watch value indicating a packet is present */
1021	first->next_to_watch = tx_desc;
1022
1023	tx_ring->next_to_use = i;
1024
1025	/* Make sure there is space in the ring for the next send. */
1026	fm10k_maybe_stop_tx(tx_ring, DESC_NEEDED);
1027
1028	/* notify HW of packet */
1029	if (netif_xmit_stopped(txring_txq(tx_ring)) || !netdev_xmit_more()) {
1030		writel(i, tx_ring->tail);
1031	}
1032
1033	return;
1034dma_error:
1035	dev_err(tx_ring->dev, "TX DMA map failed\n");
1036
1037	/* clear dma mappings for failed tx_buffer map */
1038	for (;;) {
1039		tx_buffer = &tx_ring->tx_buffer[i];
1040		fm10k_unmap_and_free_tx_resource(tx_ring, tx_buffer);
1041		if (tx_buffer == first)
1042			break;
1043		if (i == 0)
1044			i = tx_ring->count;
1045		i--;
1046	}
1047
1048	tx_ring->next_to_use = i;
1049}
1050
1051netdev_tx_t fm10k_xmit_frame_ring(struct sk_buff *skb,
1052				  struct fm10k_ring *tx_ring)
1053{
1054	u16 count = TXD_USE_COUNT(skb_headlen(skb));
1055	struct fm10k_tx_buffer *first;
1056	unsigned short f;
1057	u32 tx_flags = 0;
1058	int tso;
1059
1060	/* need: 1 descriptor per page * PAGE_SIZE/FM10K_MAX_DATA_PER_TXD,
1061	 *       + 1 desc for skb_headlen/FM10K_MAX_DATA_PER_TXD,
1062	 *       + 2 desc gap to keep tail from touching head
1063	 * otherwise try next time
1064	 */
1065	for (f = 0; f < skb_shinfo(skb)->nr_frags; f++) {
1066		skb_frag_t *frag = &skb_shinfo(skb)->frags[f];
1067
1068		count += TXD_USE_COUNT(skb_frag_size(frag));
1069	}
1070
1071	if (fm10k_maybe_stop_tx(tx_ring, count + 3)) {
1072		tx_ring->tx_stats.tx_busy++;
1073		return NETDEV_TX_BUSY;
1074	}
1075
1076	/* record the location of the first descriptor for this packet */
1077	first = &tx_ring->tx_buffer[tx_ring->next_to_use];
1078	first->skb = skb;
1079	first->bytecount = max_t(unsigned int, skb->len, ETH_ZLEN);
1080	first->gso_segs = 1;
1081
1082	/* record initial flags and protocol */
1083	first->tx_flags = tx_flags;
1084
1085	tso = fm10k_tso(tx_ring, first);
1086	if (tso < 0)
1087		goto out_drop;
1088	else if (!tso)
1089		fm10k_tx_csum(tx_ring, first);
1090
1091	fm10k_tx_map(tx_ring, first);
1092
1093	return NETDEV_TX_OK;
1094
1095out_drop:
1096	dev_kfree_skb_any(first->skb);
1097	first->skb = NULL;
1098
1099	return NETDEV_TX_OK;
1100}
1101
1102static u64 fm10k_get_tx_completed(struct fm10k_ring *ring)
1103{
1104	return ring->stats.packets;
1105}
1106
1107/**
1108 * fm10k_get_tx_pending - how many Tx descriptors not processed
1109 * @ring: the ring structure
1110 * @in_sw: is tx_pending being checked in SW or in HW?
1111 */
1112u64 fm10k_get_tx_pending(struct fm10k_ring *ring, bool in_sw)
1113{
1114	struct fm10k_intfc *interface = ring->q_vector->interface;
1115	struct fm10k_hw *hw = &interface->hw;
1116	u32 head, tail;
1117
1118	if (likely(in_sw)) {
1119		head = ring->next_to_clean;
1120		tail = ring->next_to_use;
1121	} else {
1122		head = fm10k_read_reg(hw, FM10K_TDH(ring->reg_idx));
1123		tail = fm10k_read_reg(hw, FM10K_TDT(ring->reg_idx));
1124	}
1125
1126	return ((head <= tail) ? tail : tail + ring->count) - head;
1127}
1128
1129bool fm10k_check_tx_hang(struct fm10k_ring *tx_ring)
1130{
1131	u32 tx_done = fm10k_get_tx_completed(tx_ring);
1132	u32 tx_done_old = tx_ring->tx_stats.tx_done_old;
1133	u32 tx_pending = fm10k_get_tx_pending(tx_ring, true);
1134
1135	clear_check_for_tx_hang(tx_ring);
1136
1137	/* Check for a hung queue, but be thorough. This verifies
1138	 * that a transmit has been completed since the previous
1139	 * check AND there is at least one packet pending. By
1140	 * requiring this to fail twice we avoid races with
1141	 * clearing the ARMED bit and conditions where we
1142	 * run the check_tx_hang logic with a transmit completion
1143	 * pending but without time to complete it yet.
1144	 */
1145	if (!tx_pending || (tx_done_old != tx_done)) {
1146		/* update completed stats and continue */
1147		tx_ring->tx_stats.tx_done_old = tx_done;
1148		/* reset the countdown */
1149		clear_bit(__FM10K_HANG_CHECK_ARMED, tx_ring->state);
1150
1151		return false;
1152	}
1153
1154	/* make sure it is true for two checks in a row */
1155	return test_and_set_bit(__FM10K_HANG_CHECK_ARMED, tx_ring->state);
1156}
1157
1158/**
1159 * fm10k_tx_timeout_reset - initiate reset due to Tx timeout
1160 * @interface: driver private struct
1161 **/
1162void fm10k_tx_timeout_reset(struct fm10k_intfc *interface)
1163{
1164	/* Do the reset outside of interrupt context */
1165	if (!test_bit(__FM10K_DOWN, interface->state)) {
1166		interface->tx_timeout_count++;
1167		set_bit(FM10K_FLAG_RESET_REQUESTED, interface->flags);
1168		fm10k_service_event_schedule(interface);
1169	}
1170}
1171
1172/**
1173 * fm10k_clean_tx_irq - Reclaim resources after transmit completes
1174 * @q_vector: structure containing interrupt and ring information
1175 * @tx_ring: tx ring to clean
1176 * @napi_budget: Used to determine if we are in netpoll
1177 **/
1178static bool fm10k_clean_tx_irq(struct fm10k_q_vector *q_vector,
1179			       struct fm10k_ring *tx_ring, int napi_budget)
1180{
1181	struct fm10k_intfc *interface = q_vector->interface;
1182	struct fm10k_tx_buffer *tx_buffer;
1183	struct fm10k_tx_desc *tx_desc;
1184	unsigned int total_bytes = 0, total_packets = 0;
1185	unsigned int budget = q_vector->tx.work_limit;
1186	unsigned int i = tx_ring->next_to_clean;
1187
1188	if (test_bit(__FM10K_DOWN, interface->state))
1189		return true;
1190
1191	tx_buffer = &tx_ring->tx_buffer[i];
1192	tx_desc = FM10K_TX_DESC(tx_ring, i);
1193	i -= tx_ring->count;
1194
1195	do {
1196		struct fm10k_tx_desc *eop_desc = tx_buffer->next_to_watch;
1197
1198		/* if next_to_watch is not set then there is no work pending */
1199		if (!eop_desc)
1200			break;
1201
1202		/* prevent any other reads prior to eop_desc */
1203		smp_rmb();
1204
1205		/* if DD is not set pending work has not been completed */
1206		if (!(eop_desc->flags & FM10K_TXD_FLAG_DONE))
1207			break;
1208
1209		/* clear next_to_watch to prevent false hangs */
1210		tx_buffer->next_to_watch = NULL;
1211
1212		/* update the statistics for this packet */
1213		total_bytes += tx_buffer->bytecount;
1214		total_packets += tx_buffer->gso_segs;
1215
1216		/* free the skb */
1217		napi_consume_skb(tx_buffer->skb, napi_budget);
1218
1219		/* unmap skb header data */
1220		dma_unmap_single(tx_ring->dev,
1221				 dma_unmap_addr(tx_buffer, dma),
1222				 dma_unmap_len(tx_buffer, len),
1223				 DMA_TO_DEVICE);
1224
1225		/* clear tx_buffer data */
1226		tx_buffer->skb = NULL;
1227		dma_unmap_len_set(tx_buffer, len, 0);
1228
1229		/* unmap remaining buffers */
1230		while (tx_desc != eop_desc) {
1231			tx_buffer++;
1232			tx_desc++;
1233			i++;
1234			if (unlikely(!i)) {
1235				i -= tx_ring->count;
1236				tx_buffer = tx_ring->tx_buffer;
1237				tx_desc = FM10K_TX_DESC(tx_ring, 0);
1238			}
1239
1240			/* unmap any remaining paged data */
1241			if (dma_unmap_len(tx_buffer, len)) {
1242				dma_unmap_page(tx_ring->dev,
1243					       dma_unmap_addr(tx_buffer, dma),
1244					       dma_unmap_len(tx_buffer, len),
1245					       DMA_TO_DEVICE);
1246				dma_unmap_len_set(tx_buffer, len, 0);
1247			}
1248		}
1249
1250		/* move us one more past the eop_desc for start of next pkt */
1251		tx_buffer++;
1252		tx_desc++;
1253		i++;
1254		if (unlikely(!i)) {
1255			i -= tx_ring->count;
1256			tx_buffer = tx_ring->tx_buffer;
1257			tx_desc = FM10K_TX_DESC(tx_ring, 0);
1258		}
1259
1260		/* issue prefetch for next Tx descriptor */
1261		prefetch(tx_desc);
1262
1263		/* update budget accounting */
1264		budget--;
1265	} while (likely(budget));
1266
1267	i += tx_ring->count;
1268	tx_ring->next_to_clean = i;
1269	u64_stats_update_begin(&tx_ring->syncp);
1270	tx_ring->stats.bytes += total_bytes;
1271	tx_ring->stats.packets += total_packets;
1272	u64_stats_update_end(&tx_ring->syncp);
1273	q_vector->tx.total_bytes += total_bytes;
1274	q_vector->tx.total_packets += total_packets;
1275
1276	if (check_for_tx_hang(tx_ring) && fm10k_check_tx_hang(tx_ring)) {
1277		/* schedule immediate reset if we believe we hung */
1278		struct fm10k_hw *hw = &interface->hw;
1279
1280		netif_err(interface, drv, tx_ring->netdev,
1281			  "Detected Tx Unit Hang\n"
1282			  "  Tx Queue             <%d>\n"
1283			  "  TDH, TDT             <%x>, <%x>\n"
1284			  "  next_to_use          <%x>\n"
1285			  "  next_to_clean        <%x>\n",
1286			  tx_ring->queue_index,
1287			  fm10k_read_reg(hw, FM10K_TDH(tx_ring->reg_idx)),
1288			  fm10k_read_reg(hw, FM10K_TDT(tx_ring->reg_idx)),
1289			  tx_ring->next_to_use, i);
1290
1291		netif_stop_subqueue(tx_ring->netdev,
1292				    tx_ring->queue_index);
1293
1294		netif_info(interface, probe, tx_ring->netdev,
1295			   "tx hang %d detected on queue %d, resetting interface\n",
1296			   interface->tx_timeout_count + 1,
1297			   tx_ring->queue_index);
1298
1299		fm10k_tx_timeout_reset(interface);
1300
1301		/* the netdev is about to reset, no point in enabling stuff */
1302		return true;
1303	}
1304
1305	/* notify netdev of completed buffers */
1306	netdev_tx_completed_queue(txring_txq(tx_ring),
1307				  total_packets, total_bytes);
1308
1309#define TX_WAKE_THRESHOLD min_t(u16, FM10K_MIN_TXD - 1, DESC_NEEDED * 2)
1310	if (unlikely(total_packets && netif_carrier_ok(tx_ring->netdev) &&
1311		     (fm10k_desc_unused(tx_ring) >= TX_WAKE_THRESHOLD))) {
1312		/* Make sure that anybody stopping the queue after this
1313		 * sees the new next_to_clean.
1314		 */
1315		smp_mb();
1316		if (__netif_subqueue_stopped(tx_ring->netdev,
1317					     tx_ring->queue_index) &&
1318		    !test_bit(__FM10K_DOWN, interface->state)) {
1319			netif_wake_subqueue(tx_ring->netdev,
1320					    tx_ring->queue_index);
1321			++tx_ring->tx_stats.restart_queue;
1322		}
1323	}
1324
1325	return !!budget;
1326}
1327
1328/**
1329 * fm10k_update_itr - update the dynamic ITR value based on packet size
1330 *
1331 *      Stores a new ITR value based on strictly on packet size.  The
1332 *      divisors and thresholds used by this function were determined based
1333 *      on theoretical maximum wire speed and testing data, in order to
1334 *      minimize response time while increasing bulk throughput.
1335 *
1336 * @ring_container: Container for rings to have ITR updated
1337 **/
1338static void fm10k_update_itr(struct fm10k_ring_container *ring_container)
1339{
1340	unsigned int avg_wire_size, packets, itr_round;
1341
1342	/* Only update ITR if we are using adaptive setting */
1343	if (!ITR_IS_ADAPTIVE(ring_container->itr))
1344		goto clear_counts;
1345
1346	packets = ring_container->total_packets;
1347	if (!packets)
1348		goto clear_counts;
1349
1350	avg_wire_size = ring_container->total_bytes / packets;
1351
1352	/* The following is a crude approximation of:
1353	 *  wmem_default / (size + overhead) = desired_pkts_per_int
1354	 *  rate / bits_per_byte / (size + ethernet overhead) = pkt_rate
1355	 *  (desired_pkt_rate / pkt_rate) * usecs_per_sec = ITR value
1356	 *
1357	 * Assuming wmem_default is 212992 and overhead is 640 bytes per
1358	 * packet, (256 skb, 64 headroom, 320 shared info), we can reduce the
1359	 * formula down to
1360	 *
1361	 *  (34 * (size + 24)) / (size + 640) = ITR
1362	 *
1363	 * We first do some math on the packet size and then finally bitshift
1364	 * by 8 after rounding up. We also have to account for PCIe link speed
1365	 * difference as ITR scales based on this.
1366	 */
1367	if (avg_wire_size <= 360) {
1368		/* Start at 250K ints/sec and gradually drop to 77K ints/sec */
1369		avg_wire_size *= 8;
1370		avg_wire_size += 376;
1371	} else if (avg_wire_size <= 1152) {
1372		/* 77K ints/sec to 45K ints/sec */
1373		avg_wire_size *= 3;
1374		avg_wire_size += 2176;
1375	} else if (avg_wire_size <= 1920) {
1376		/* 45K ints/sec to 38K ints/sec */
1377		avg_wire_size += 4480;
1378	} else {
1379		/* plateau at a limit of 38K ints/sec */
1380		avg_wire_size = 6656;
1381	}
1382
1383	/* Perform final bitshift for division after rounding up to ensure
1384	 * that the calculation will never get below a 1. The bit shift
1385	 * accounts for changes in the ITR due to PCIe link speed.
1386	 */
1387	itr_round = READ_ONCE(ring_container->itr_scale) + 8;
1388	avg_wire_size += BIT(itr_round) - 1;
1389	avg_wire_size >>= itr_round;
1390
1391	/* write back value and retain adaptive flag */
1392	ring_container->itr = avg_wire_size | FM10K_ITR_ADAPTIVE;
1393
1394clear_counts:
1395	ring_container->total_bytes = 0;
1396	ring_container->total_packets = 0;
1397}
1398
1399static void fm10k_qv_enable(struct fm10k_q_vector *q_vector)
1400{
1401	/* Enable auto-mask and clear the current mask */
1402	u32 itr = FM10K_ITR_ENABLE;
1403
1404	/* Update Tx ITR */
1405	fm10k_update_itr(&q_vector->tx);
1406
1407	/* Update Rx ITR */
1408	fm10k_update_itr(&q_vector->rx);
1409
1410	/* Store Tx itr in timer slot 0 */
1411	itr |= (q_vector->tx.itr & FM10K_ITR_MAX);
1412
1413	/* Shift Rx itr to timer slot 1 */
1414	itr |= (q_vector->rx.itr & FM10K_ITR_MAX) << FM10K_ITR_INTERVAL1_SHIFT;
1415
1416	/* Write the final value to the ITR register */
1417	writel(itr, q_vector->itr);
1418}
1419
1420static int fm10k_poll(struct napi_struct *napi, int budget)
1421{
1422	struct fm10k_q_vector *q_vector =
1423			       container_of(napi, struct fm10k_q_vector, napi);
1424	struct fm10k_ring *ring;
1425	int per_ring_budget, work_done = 0;
1426	bool clean_complete = true;
1427
1428	fm10k_for_each_ring(ring, q_vector->tx) {
1429		if (!fm10k_clean_tx_irq(q_vector, ring, budget))
1430			clean_complete = false;
1431	}
1432
1433	/* Handle case where we are called by netpoll with a budget of 0 */
1434	if (budget <= 0)
1435		return budget;
1436
1437	/* attempt to distribute budget to each queue fairly, but don't
1438	 * allow the budget to go below 1 because we'll exit polling
1439	 */
1440	if (q_vector->rx.count > 1)
1441		per_ring_budget = max(budget / q_vector->rx.count, 1);
1442	else
1443		per_ring_budget = budget;
1444
1445	fm10k_for_each_ring(ring, q_vector->rx) {
1446		int work = fm10k_clean_rx_irq(q_vector, ring, per_ring_budget);
1447
1448		work_done += work;
1449		if (work >= per_ring_budget)
1450			clean_complete = false;
1451	}
1452
1453	/* If all work not completed, return budget and keep polling */
1454	if (!clean_complete)
1455		return budget;
1456
1457	/* Exit the polling mode, but don't re-enable interrupts if stack might
1458	 * poll us due to busy-polling
1459	 */
1460	if (likely(napi_complete_done(napi, work_done)))
1461		fm10k_qv_enable(q_vector);
1462
1463	return min(work_done, budget - 1);
1464}
1465
1466/**
1467 * fm10k_set_qos_queues: Allocate queues for a QOS-enabled device
1468 * @interface: board private structure to initialize
1469 *
1470 * When QoS (Quality of Service) is enabled, allocate queues for
1471 * each traffic class.  If multiqueue isn't available,then abort QoS
1472 * initialization.
1473 *
1474 * This function handles all combinations of Qos and RSS.
1475 *
1476 **/
1477static bool fm10k_set_qos_queues(struct fm10k_intfc *interface)
1478{
1479	struct net_device *dev = interface->netdev;
1480	struct fm10k_ring_feature *f;
1481	int rss_i, i;
1482	int pcs;
1483
1484	/* Map queue offset and counts onto allocated tx queues */
1485	pcs = netdev_get_num_tc(dev);
1486
1487	if (pcs <= 1)
1488		return false;
1489
1490	/* set QoS mask and indices */
1491	f = &interface->ring_feature[RING_F_QOS];
1492	f->indices = pcs;
1493	f->mask = BIT(fls(pcs - 1)) - 1;
1494
1495	/* determine the upper limit for our current DCB mode */
1496	rss_i = interface->hw.mac.max_queues / pcs;
1497	rss_i = BIT(fls(rss_i) - 1);
1498
1499	/* set RSS mask and indices */
1500	f = &interface->ring_feature[RING_F_RSS];
1501	rss_i = min_t(u16, rss_i, f->limit);
1502	f->indices = rss_i;
1503	f->mask = BIT(fls(rss_i - 1)) - 1;
1504
1505	/* configure pause class to queue mapping */
1506	for (i = 0; i < pcs; i++)
1507		netdev_set_tc_queue(dev, i, rss_i, rss_i * i);
1508
1509	interface->num_rx_queues = rss_i * pcs;
1510	interface->num_tx_queues = rss_i * pcs;
1511
1512	return true;
1513}
1514
1515/**
1516 * fm10k_set_rss_queues: Allocate queues for RSS
1517 * @interface: board private structure to initialize
1518 *
1519 * This is our "base" multiqueue mode.  RSS (Receive Side Scaling) will try
1520 * to allocate one Rx queue per CPU, and if available, one Tx queue per CPU.
1521 *
1522 **/
1523static bool fm10k_set_rss_queues(struct fm10k_intfc *interface)
1524{
1525	struct fm10k_ring_feature *f;
1526	u16 rss_i;
1527
1528	f = &interface->ring_feature[RING_F_RSS];
1529	rss_i = min_t(u16, interface->hw.mac.max_queues, f->limit);
1530
1531	/* record indices and power of 2 mask for RSS */
1532	f->indices = rss_i;
1533	f->mask = BIT(fls(rss_i - 1)) - 1;
1534
1535	interface->num_rx_queues = rss_i;
1536	interface->num_tx_queues = rss_i;
1537
1538	return true;
1539}
1540
1541/**
1542 * fm10k_set_num_queues: Allocate queues for device, feature dependent
1543 * @interface: board private structure to initialize
1544 *
1545 * This is the top level queue allocation routine.  The order here is very
1546 * important, starting with the "most" number of features turned on at once,
1547 * and ending with the smallest set of features.  This way large combinations
1548 * can be allocated if they're turned on, and smaller combinations are the
1549 * fall through conditions.
1550 *
1551 **/
1552static void fm10k_set_num_queues(struct fm10k_intfc *interface)
1553{
1554	/* Attempt to setup QoS and RSS first */
1555	if (fm10k_set_qos_queues(interface))
1556		return;
1557
1558	/* If we don't have QoS, just fallback to only RSS. */
1559	fm10k_set_rss_queues(interface);
1560}
1561
1562/**
1563 * fm10k_reset_num_queues - Reset the number of queues to zero
1564 * @interface: board private structure
1565 *
1566 * This function should be called whenever we need to reset the number of
1567 * queues after an error condition.
1568 */
1569static void fm10k_reset_num_queues(struct fm10k_intfc *interface)
1570{
1571	interface->num_tx_queues = 0;
1572	interface->num_rx_queues = 0;
1573	interface->num_q_vectors = 0;
1574}
1575
1576/**
1577 * fm10k_alloc_q_vector - Allocate memory for a single interrupt vector
1578 * @interface: board private structure to initialize
1579 * @v_count: q_vectors allocated on interface, used for ring interleaving
1580 * @v_idx: index of vector in interface struct
1581 * @txr_count: total number of Tx rings to allocate
1582 * @txr_idx: index of first Tx ring to allocate
1583 * @rxr_count: total number of Rx rings to allocate
1584 * @rxr_idx: index of first Rx ring to allocate
1585 *
1586 * We allocate one q_vector.  If allocation fails we return -ENOMEM.
1587 **/
1588static int fm10k_alloc_q_vector(struct fm10k_intfc *interface,
1589				unsigned int v_count, unsigned int v_idx,
1590				unsigned int txr_count, unsigned int txr_idx,
1591				unsigned int rxr_count, unsigned int rxr_idx)
1592{
1593	struct fm10k_q_vector *q_vector;
1594	struct fm10k_ring *ring;
1595	int ring_count;
1596
1597	ring_count = txr_count + rxr_count;
1598
1599	/* allocate q_vector and rings */
1600	q_vector = kzalloc(struct_size(q_vector, ring, ring_count), GFP_KERNEL);
1601	if (!q_vector)
1602		return -ENOMEM;
1603
1604	/* initialize NAPI */
1605	netif_napi_add(interface->netdev, &q_vector->napi, fm10k_poll);
 
1606
1607	/* tie q_vector and interface together */
1608	interface->q_vector[v_idx] = q_vector;
1609	q_vector->interface = interface;
1610	q_vector->v_idx = v_idx;
1611
1612	/* initialize pointer to rings */
1613	ring = q_vector->ring;
1614
1615	/* save Tx ring container info */
1616	q_vector->tx.ring = ring;
1617	q_vector->tx.work_limit = FM10K_DEFAULT_TX_WORK;
1618	q_vector->tx.itr = interface->tx_itr;
1619	q_vector->tx.itr_scale = interface->hw.mac.itr_scale;
1620	q_vector->tx.count = txr_count;
1621
1622	while (txr_count) {
1623		/* assign generic ring traits */
1624		ring->dev = &interface->pdev->dev;
1625		ring->netdev = interface->netdev;
1626
1627		/* configure backlink on ring */
1628		ring->q_vector = q_vector;
1629
1630		/* apply Tx specific ring traits */
1631		ring->count = interface->tx_ring_count;
1632		ring->queue_index = txr_idx;
1633
1634		/* assign ring to interface */
1635		interface->tx_ring[txr_idx] = ring;
1636
1637		/* update count and index */
1638		txr_count--;
1639		txr_idx += v_count;
1640
1641		/* push pointer to next ring */
1642		ring++;
1643	}
1644
1645	/* save Rx ring container info */
1646	q_vector->rx.ring = ring;
1647	q_vector->rx.itr = interface->rx_itr;
1648	q_vector->rx.itr_scale = interface->hw.mac.itr_scale;
1649	q_vector->rx.count = rxr_count;
1650
1651	while (rxr_count) {
1652		/* assign generic ring traits */
1653		ring->dev = &interface->pdev->dev;
1654		ring->netdev = interface->netdev;
1655		rcu_assign_pointer(ring->l2_accel, interface->l2_accel);
1656
1657		/* configure backlink on ring */
1658		ring->q_vector = q_vector;
1659
1660		/* apply Rx specific ring traits */
1661		ring->count = interface->rx_ring_count;
1662		ring->queue_index = rxr_idx;
1663
1664		/* assign ring to interface */
1665		interface->rx_ring[rxr_idx] = ring;
1666
1667		/* update count and index */
1668		rxr_count--;
1669		rxr_idx += v_count;
1670
1671		/* push pointer to next ring */
1672		ring++;
1673	}
1674
1675	fm10k_dbg_q_vector_init(q_vector);
1676
1677	return 0;
1678}
1679
1680/**
1681 * fm10k_free_q_vector - Free memory allocated for specific interrupt vector
1682 * @interface: board private structure to initialize
1683 * @v_idx: Index of vector to be freed
1684 *
1685 * This function frees the memory allocated to the q_vector.  In addition if
1686 * NAPI is enabled it will delete any references to the NAPI struct prior
1687 * to freeing the q_vector.
1688 **/
1689static void fm10k_free_q_vector(struct fm10k_intfc *interface, int v_idx)
1690{
1691	struct fm10k_q_vector *q_vector = interface->q_vector[v_idx];
1692	struct fm10k_ring *ring;
1693
1694	fm10k_dbg_q_vector_exit(q_vector);
1695
1696	fm10k_for_each_ring(ring, q_vector->tx)
1697		interface->tx_ring[ring->queue_index] = NULL;
1698
1699	fm10k_for_each_ring(ring, q_vector->rx)
1700		interface->rx_ring[ring->queue_index] = NULL;
1701
1702	interface->q_vector[v_idx] = NULL;
1703	netif_napi_del(&q_vector->napi);
1704	kfree_rcu(q_vector, rcu);
1705}
1706
1707/**
1708 * fm10k_alloc_q_vectors - Allocate memory for interrupt vectors
1709 * @interface: board private structure to initialize
1710 *
1711 * We allocate one q_vector per queue interrupt.  If allocation fails we
1712 * return -ENOMEM.
1713 **/
1714static int fm10k_alloc_q_vectors(struct fm10k_intfc *interface)
1715{
1716	unsigned int q_vectors = interface->num_q_vectors;
1717	unsigned int rxr_remaining = interface->num_rx_queues;
1718	unsigned int txr_remaining = interface->num_tx_queues;
1719	unsigned int rxr_idx = 0, txr_idx = 0, v_idx = 0;
1720	int err;
1721
1722	if (q_vectors >= (rxr_remaining + txr_remaining)) {
1723		for (; rxr_remaining; v_idx++) {
1724			err = fm10k_alloc_q_vector(interface, q_vectors, v_idx,
1725						   0, 0, 1, rxr_idx);
1726			if (err)
1727				goto err_out;
1728
1729			/* update counts and index */
1730			rxr_remaining--;
1731			rxr_idx++;
1732		}
1733	}
1734
1735	for (; v_idx < q_vectors; v_idx++) {
1736		int rqpv = DIV_ROUND_UP(rxr_remaining, q_vectors - v_idx);
1737		int tqpv = DIV_ROUND_UP(txr_remaining, q_vectors - v_idx);
1738
1739		err = fm10k_alloc_q_vector(interface, q_vectors, v_idx,
1740					   tqpv, txr_idx,
1741					   rqpv, rxr_idx);
1742
1743		if (err)
1744			goto err_out;
1745
1746		/* update counts and index */
1747		rxr_remaining -= rqpv;
1748		txr_remaining -= tqpv;
1749		rxr_idx++;
1750		txr_idx++;
1751	}
1752
1753	return 0;
1754
1755err_out:
1756	fm10k_reset_num_queues(interface);
1757
1758	while (v_idx--)
1759		fm10k_free_q_vector(interface, v_idx);
1760
1761	return -ENOMEM;
1762}
1763
1764/**
1765 * fm10k_free_q_vectors - Free memory allocated for interrupt vectors
1766 * @interface: board private structure to initialize
1767 *
1768 * This function frees the memory allocated to the q_vectors.  In addition if
1769 * NAPI is enabled it will delete any references to the NAPI struct prior
1770 * to freeing the q_vector.
1771 **/
1772static void fm10k_free_q_vectors(struct fm10k_intfc *interface)
1773{
1774	int v_idx = interface->num_q_vectors;
1775
1776	fm10k_reset_num_queues(interface);
1777
1778	while (v_idx--)
1779		fm10k_free_q_vector(interface, v_idx);
1780}
1781
1782/**
1783 * fm10k_reset_msix_capability - reset MSI-X capability
1784 * @interface: board private structure to initialize
1785 *
1786 * Reset the MSI-X capability back to its starting state
1787 **/
1788static void fm10k_reset_msix_capability(struct fm10k_intfc *interface)
1789{
1790	pci_disable_msix(interface->pdev);
1791	kfree(interface->msix_entries);
1792	interface->msix_entries = NULL;
1793}
1794
1795/**
1796 * fm10k_init_msix_capability - configure MSI-X capability
1797 * @interface: board private structure to initialize
1798 *
1799 * Attempt to configure the interrupts using the best available
1800 * capabilities of the hardware and the kernel.
1801 **/
1802static int fm10k_init_msix_capability(struct fm10k_intfc *interface)
1803{
1804	struct fm10k_hw *hw = &interface->hw;
1805	int v_budget, vector;
1806
1807	/* It's easy to be greedy for MSI-X vectors, but it really
1808	 * doesn't do us much good if we have a lot more vectors
1809	 * than CPU's.  So let's be conservative and only ask for
1810	 * (roughly) the same number of vectors as there are CPU's.
1811	 * the default is to use pairs of vectors
1812	 */
1813	v_budget = max(interface->num_rx_queues, interface->num_tx_queues);
1814	v_budget = min_t(u16, v_budget, num_online_cpus());
1815
1816	/* account for vectors not related to queues */
1817	v_budget += NON_Q_VECTORS;
1818
1819	/* At the same time, hardware can only support a maximum of
1820	 * hw.mac->max_msix_vectors vectors.  With features
1821	 * such as RSS and VMDq, we can easily surpass the number of Rx and Tx
1822	 * descriptor queues supported by our device.  Thus, we cap it off in
1823	 * those rare cases where the cpu count also exceeds our vector limit.
1824	 */
1825	v_budget = min_t(int, v_budget, hw->mac.max_msix_vectors);
1826
1827	/* A failure in MSI-X entry allocation is fatal. */
1828	interface->msix_entries = kcalloc(v_budget, sizeof(struct msix_entry),
1829					  GFP_KERNEL);
1830	if (!interface->msix_entries)
1831		return -ENOMEM;
1832
1833	/* populate entry values */
1834	for (vector = 0; vector < v_budget; vector++)
1835		interface->msix_entries[vector].entry = vector;
1836
1837	/* Attempt to enable MSI-X with requested value */
1838	v_budget = pci_enable_msix_range(interface->pdev,
1839					 interface->msix_entries,
1840					 MIN_MSIX_COUNT(hw),
1841					 v_budget);
1842	if (v_budget < 0) {
1843		kfree(interface->msix_entries);
1844		interface->msix_entries = NULL;
1845		return v_budget;
1846	}
1847
1848	/* record the number of queues available for q_vectors */
1849	interface->num_q_vectors = v_budget - NON_Q_VECTORS;
1850
1851	return 0;
1852}
1853
1854/**
1855 * fm10k_cache_ring_qos - Descriptor ring to register mapping for QoS
1856 * @interface: Interface structure continaining rings and devices
1857 *
1858 * Cache the descriptor ring offsets for Qos
1859 **/
1860static bool fm10k_cache_ring_qos(struct fm10k_intfc *interface)
1861{
1862	struct net_device *dev = interface->netdev;
1863	int pc, offset, rss_i, i;
1864	u16 pc_stride = interface->ring_feature[RING_F_QOS].mask + 1;
1865	u8 num_pcs = netdev_get_num_tc(dev);
1866
1867	if (num_pcs <= 1)
1868		return false;
1869
1870	rss_i = interface->ring_feature[RING_F_RSS].indices;
1871
1872	for (pc = 0, offset = 0; pc < num_pcs; pc++, offset += rss_i) {
1873		int q_idx = pc;
1874
1875		for (i = 0; i < rss_i; i++) {
1876			interface->tx_ring[offset + i]->reg_idx = q_idx;
1877			interface->tx_ring[offset + i]->qos_pc = pc;
1878			interface->rx_ring[offset + i]->reg_idx = q_idx;
1879			interface->rx_ring[offset + i]->qos_pc = pc;
1880			q_idx += pc_stride;
1881		}
1882	}
1883
1884	return true;
1885}
1886
1887/**
1888 * fm10k_cache_ring_rss - Descriptor ring to register mapping for RSS
1889 * @interface: Interface structure continaining rings and devices
1890 *
1891 * Cache the descriptor ring offsets for RSS
1892 **/
1893static void fm10k_cache_ring_rss(struct fm10k_intfc *interface)
1894{
1895	int i;
1896
1897	for (i = 0; i < interface->num_rx_queues; i++)
1898		interface->rx_ring[i]->reg_idx = i;
1899
1900	for (i = 0; i < interface->num_tx_queues; i++)
1901		interface->tx_ring[i]->reg_idx = i;
1902}
1903
1904/**
1905 * fm10k_assign_rings - Map rings to network devices
1906 * @interface: Interface structure containing rings and devices
1907 *
1908 * This function is meant to go though and configure both the network
1909 * devices so that they contain rings, and configure the rings so that
1910 * they function with their network devices.
1911 **/
1912static void fm10k_assign_rings(struct fm10k_intfc *interface)
1913{
1914	if (fm10k_cache_ring_qos(interface))
1915		return;
1916
1917	fm10k_cache_ring_rss(interface);
1918}
1919
1920static void fm10k_init_reta(struct fm10k_intfc *interface)
1921{
1922	u16 i, rss_i = interface->ring_feature[RING_F_RSS].indices;
1923	u32 reta;
1924
1925	/* If the Rx flow indirection table has been configured manually, we
1926	 * need to maintain it when possible.
1927	 */
1928	if (netif_is_rxfh_configured(interface->netdev)) {
1929		for (i = FM10K_RETA_SIZE; i--;) {
1930			reta = interface->reta[i];
1931			if ((((reta << 24) >> 24) < rss_i) &&
1932			    (((reta << 16) >> 24) < rss_i) &&
1933			    (((reta <<  8) >> 24) < rss_i) &&
1934			    (((reta)       >> 24) < rss_i))
1935				continue;
1936
1937			/* this should never happen */
1938			dev_err(&interface->pdev->dev,
1939				"RSS indirection table assigned flows out of queue bounds. Reconfiguring.\n");
1940			goto repopulate_reta;
1941		}
1942
1943		/* do nothing if all of the elements are in bounds */
1944		return;
1945	}
1946
1947repopulate_reta:
1948	fm10k_write_reta(interface, NULL);
1949}
1950
1951/**
1952 * fm10k_init_queueing_scheme - Determine proper queueing scheme
1953 * @interface: board private structure to initialize
1954 *
1955 * We determine which queueing scheme to use based on...
1956 * - Hardware queue count (num_*_queues)
1957 *   - defined by miscellaneous hardware support/features (RSS, etc.)
1958 **/
1959int fm10k_init_queueing_scheme(struct fm10k_intfc *interface)
1960{
1961	int err;
1962
1963	/* Number of supported queues */
1964	fm10k_set_num_queues(interface);
1965
1966	/* Configure MSI-X capability */
1967	err = fm10k_init_msix_capability(interface);
1968	if (err) {
1969		dev_err(&interface->pdev->dev,
1970			"Unable to initialize MSI-X capability\n");
1971		goto err_init_msix;
1972	}
1973
1974	/* Allocate memory for queues */
1975	err = fm10k_alloc_q_vectors(interface);
1976	if (err) {
1977		dev_err(&interface->pdev->dev,
1978			"Unable to allocate queue vectors\n");
1979		goto err_alloc_q_vectors;
1980	}
1981
1982	/* Map rings to devices, and map devices to physical queues */
1983	fm10k_assign_rings(interface);
1984
1985	/* Initialize RSS redirection table */
1986	fm10k_init_reta(interface);
1987
1988	return 0;
1989
1990err_alloc_q_vectors:
1991	fm10k_reset_msix_capability(interface);
1992err_init_msix:
1993	fm10k_reset_num_queues(interface);
1994	return err;
1995}
1996
1997/**
1998 * fm10k_clear_queueing_scheme - Clear the current queueing scheme settings
1999 * @interface: board private structure to clear queueing scheme on
2000 *
2001 * We go through and clear queueing specific resources and reset the structure
2002 * to pre-load conditions
2003 **/
2004void fm10k_clear_queueing_scheme(struct fm10k_intfc *interface)
2005{
2006	fm10k_free_q_vectors(interface);
2007	fm10k_reset_msix_capability(interface);
2008}
v5.9
   1// SPDX-License-Identifier: GPL-2.0
   2/* Copyright(c) 2013 - 2019 Intel Corporation. */
   3
   4#include <linux/types.h>
   5#include <linux/module.h>
   6#include <net/ipv6.h>
   7#include <net/ip.h>
   8#include <net/tcp.h>
   9#include <linux/if_macvlan.h>
  10#include <linux/prefetch.h>
  11
  12#include "fm10k.h"
  13
  14#define DRV_SUMMARY	"Intel(R) Ethernet Switch Host Interface Driver"
  15char fm10k_driver_name[] = "fm10k";
  16static const char fm10k_driver_string[] = DRV_SUMMARY;
  17static const char fm10k_copyright[] =
  18	"Copyright(c) 2013 - 2019 Intel Corporation.";
  19
  20MODULE_AUTHOR("Intel Corporation, <linux.nics@intel.com>");
  21MODULE_DESCRIPTION(DRV_SUMMARY);
  22MODULE_LICENSE("GPL v2");
  23
  24/* single workqueue for entire fm10k driver */
  25struct workqueue_struct *fm10k_workqueue;
  26
  27/**
  28 * fm10k_init_module - Driver Registration Routine
  29 *
  30 * fm10k_init_module is the first routine called when the driver is
  31 * loaded.  All it does is register with the PCI subsystem.
  32 **/
  33static int __init fm10k_init_module(void)
  34{
 
 
  35	pr_info("%s\n", fm10k_driver_string);
  36	pr_info("%s\n", fm10k_copyright);
  37
  38	/* create driver workqueue */
  39	fm10k_workqueue = alloc_workqueue("%s", WQ_MEM_RECLAIM, 0,
  40					  fm10k_driver_name);
  41	if (!fm10k_workqueue)
  42		return -ENOMEM;
  43
  44	fm10k_dbg_init();
  45
  46	return fm10k_register_pci_driver();
 
 
 
 
 
 
  47}
  48module_init(fm10k_init_module);
  49
  50/**
  51 * fm10k_exit_module - Driver Exit Cleanup Routine
  52 *
  53 * fm10k_exit_module is called just before the driver is removed
  54 * from memory.
  55 **/
  56static void __exit fm10k_exit_module(void)
  57{
  58	fm10k_unregister_pci_driver();
  59
  60	fm10k_dbg_exit();
  61
  62	/* destroy driver workqueue */
  63	destroy_workqueue(fm10k_workqueue);
  64}
  65module_exit(fm10k_exit_module);
  66
  67static bool fm10k_alloc_mapped_page(struct fm10k_ring *rx_ring,
  68				    struct fm10k_rx_buffer *bi)
  69{
  70	struct page *page = bi->page;
  71	dma_addr_t dma;
  72
  73	/* Only page will be NULL if buffer was consumed */
  74	if (likely(page))
  75		return true;
  76
  77	/* alloc new page for storage */
  78	page = dev_alloc_page();
  79	if (unlikely(!page)) {
  80		rx_ring->rx_stats.alloc_failed++;
  81		return false;
  82	}
  83
  84	/* map page for use */
  85	dma = dma_map_page(rx_ring->dev, page, 0, PAGE_SIZE, DMA_FROM_DEVICE);
  86
  87	/* if mapping failed free memory back to system since
  88	 * there isn't much point in holding memory we can't use
  89	 */
  90	if (dma_mapping_error(rx_ring->dev, dma)) {
  91		__free_page(page);
  92
  93		rx_ring->rx_stats.alloc_failed++;
  94		return false;
  95	}
  96
  97	bi->dma = dma;
  98	bi->page = page;
  99	bi->page_offset = 0;
 100
 101	return true;
 102}
 103
 104/**
 105 * fm10k_alloc_rx_buffers - Replace used receive buffers
 106 * @rx_ring: ring to place buffers on
 107 * @cleaned_count: number of buffers to replace
 108 **/
 109void fm10k_alloc_rx_buffers(struct fm10k_ring *rx_ring, u16 cleaned_count)
 110{
 111	union fm10k_rx_desc *rx_desc;
 112	struct fm10k_rx_buffer *bi;
 113	u16 i = rx_ring->next_to_use;
 114
 115	/* nothing to do */
 116	if (!cleaned_count)
 117		return;
 118
 119	rx_desc = FM10K_RX_DESC(rx_ring, i);
 120	bi = &rx_ring->rx_buffer[i];
 121	i -= rx_ring->count;
 122
 123	do {
 124		if (!fm10k_alloc_mapped_page(rx_ring, bi))
 125			break;
 126
 127		/* Refresh the desc even if buffer_addrs didn't change
 128		 * because each write-back erases this info.
 129		 */
 130		rx_desc->q.pkt_addr = cpu_to_le64(bi->dma + bi->page_offset);
 131
 132		rx_desc++;
 133		bi++;
 134		i++;
 135		if (unlikely(!i)) {
 136			rx_desc = FM10K_RX_DESC(rx_ring, 0);
 137			bi = rx_ring->rx_buffer;
 138			i -= rx_ring->count;
 139		}
 140
 141		/* clear the status bits for the next_to_use descriptor */
 142		rx_desc->d.staterr = 0;
 143
 144		cleaned_count--;
 145	} while (cleaned_count);
 146
 147	i += rx_ring->count;
 148
 149	if (rx_ring->next_to_use != i) {
 150		/* record the next descriptor to use */
 151		rx_ring->next_to_use = i;
 152
 153		/* update next to alloc since we have filled the ring */
 154		rx_ring->next_to_alloc = i;
 155
 156		/* Force memory writes to complete before letting h/w
 157		 * know there are new descriptors to fetch.  (Only
 158		 * applicable for weak-ordered memory model archs,
 159		 * such as IA-64).
 160		 */
 161		wmb();
 162
 163		/* notify hardware of new descriptors */
 164		writel(i, rx_ring->tail);
 165	}
 166}
 167
 168/**
 169 * fm10k_reuse_rx_page - page flip buffer and store it back on the ring
 170 * @rx_ring: rx descriptor ring to store buffers on
 171 * @old_buff: donor buffer to have page reused
 172 *
 173 * Synchronizes page for reuse by the interface
 174 **/
 175static void fm10k_reuse_rx_page(struct fm10k_ring *rx_ring,
 176				struct fm10k_rx_buffer *old_buff)
 177{
 178	struct fm10k_rx_buffer *new_buff;
 179	u16 nta = rx_ring->next_to_alloc;
 180
 181	new_buff = &rx_ring->rx_buffer[nta];
 182
 183	/* update, and store next to alloc */
 184	nta++;
 185	rx_ring->next_to_alloc = (nta < rx_ring->count) ? nta : 0;
 186
 187	/* transfer page from old buffer to new buffer */
 188	*new_buff = *old_buff;
 189
 190	/* sync the buffer for use by the device */
 191	dma_sync_single_range_for_device(rx_ring->dev, old_buff->dma,
 192					 old_buff->page_offset,
 193					 FM10K_RX_BUFSZ,
 194					 DMA_FROM_DEVICE);
 195}
 196
 197static inline bool fm10k_page_is_reserved(struct page *page)
 198{
 199	return (page_to_nid(page) != numa_mem_id()) || page_is_pfmemalloc(page);
 200}
 201
 202static bool fm10k_can_reuse_rx_page(struct fm10k_rx_buffer *rx_buffer,
 203				    struct page *page,
 204				    unsigned int __maybe_unused truesize)
 205{
 206	/* avoid re-using remote pages */
 207	if (unlikely(fm10k_page_is_reserved(page)))
 208		return false;
 209
 210#if (PAGE_SIZE < 8192)
 211	/* if we are only owner of page we can reuse it */
 212	if (unlikely(page_count(page) != 1))
 213		return false;
 214
 215	/* flip page offset to other buffer */
 216	rx_buffer->page_offset ^= FM10K_RX_BUFSZ;
 217#else
 218	/* move offset up to the next cache line */
 219	rx_buffer->page_offset += truesize;
 220
 221	if (rx_buffer->page_offset > (PAGE_SIZE - FM10K_RX_BUFSZ))
 222		return false;
 223#endif
 224
 225	/* Even if we own the page, we are not allowed to use atomic_set()
 226	 * This would break get_page_unless_zero() users.
 227	 */
 228	page_ref_inc(page);
 229
 230	return true;
 231}
 232
 233/**
 234 * fm10k_add_rx_frag - Add contents of Rx buffer to sk_buff
 235 * @rx_buffer: buffer containing page to add
 236 * @size: packet size from rx_desc
 237 * @rx_desc: descriptor containing length of buffer written by hardware
 238 * @skb: sk_buff to place the data into
 239 *
 240 * This function will add the data contained in rx_buffer->page to the skb.
 241 * This is done either through a direct copy if the data in the buffer is
 242 * less than the skb header size, otherwise it will just attach the page as
 243 * a frag to the skb.
 244 *
 245 * The function will then update the page offset if necessary and return
 246 * true if the buffer can be reused by the interface.
 247 **/
 248static bool fm10k_add_rx_frag(struct fm10k_rx_buffer *rx_buffer,
 249			      unsigned int size,
 250			      union fm10k_rx_desc *rx_desc,
 251			      struct sk_buff *skb)
 252{
 253	struct page *page = rx_buffer->page;
 254	unsigned char *va = page_address(page) + rx_buffer->page_offset;
 255#if (PAGE_SIZE < 8192)
 256	unsigned int truesize = FM10K_RX_BUFSZ;
 257#else
 258	unsigned int truesize = ALIGN(size, 512);
 259#endif
 260	unsigned int pull_len;
 261
 262	if (unlikely(skb_is_nonlinear(skb)))
 263		goto add_tail_frag;
 264
 265	if (likely(size <= FM10K_RX_HDR_LEN)) {
 266		memcpy(__skb_put(skb, size), va, ALIGN(size, sizeof(long)));
 267
 268		/* page is not reserved, we can reuse buffer as-is */
 269		if (likely(!fm10k_page_is_reserved(page)))
 270			return true;
 271
 272		/* this page cannot be reused so discard it */
 273		__free_page(page);
 274		return false;
 275	}
 276
 277	/* we need the header to contain the greater of either ETH_HLEN or
 278	 * 60 bytes if the skb->len is less than 60 for skb_pad.
 279	 */
 280	pull_len = eth_get_headlen(skb->dev, va, FM10K_RX_HDR_LEN);
 281
 282	/* align pull length to size of long to optimize memcpy performance */
 283	memcpy(__skb_put(skb, pull_len), va, ALIGN(pull_len, sizeof(long)));
 284
 285	/* update all of the pointers */
 286	va += pull_len;
 287	size -= pull_len;
 288
 289add_tail_frag:
 290	skb_add_rx_frag(skb, skb_shinfo(skb)->nr_frags, page,
 291			(unsigned long)va & ~PAGE_MASK, size, truesize);
 292
 293	return fm10k_can_reuse_rx_page(rx_buffer, page, truesize);
 294}
 295
 296static struct sk_buff *fm10k_fetch_rx_buffer(struct fm10k_ring *rx_ring,
 297					     union fm10k_rx_desc *rx_desc,
 298					     struct sk_buff *skb)
 299{
 300	unsigned int size = le16_to_cpu(rx_desc->w.length);
 301	struct fm10k_rx_buffer *rx_buffer;
 302	struct page *page;
 303
 304	rx_buffer = &rx_ring->rx_buffer[rx_ring->next_to_clean];
 305	page = rx_buffer->page;
 306	prefetchw(page);
 307
 308	if (likely(!skb)) {
 309		void *page_addr = page_address(page) +
 310				  rx_buffer->page_offset;
 311
 312		/* prefetch first cache line of first page */
 313		prefetch(page_addr);
 314#if L1_CACHE_BYTES < 128
 315		prefetch((void *)((u8 *)page_addr + L1_CACHE_BYTES));
 316#endif
 317
 318		/* allocate a skb to store the frags */
 319		skb = napi_alloc_skb(&rx_ring->q_vector->napi,
 320				     FM10K_RX_HDR_LEN);
 321		if (unlikely(!skb)) {
 322			rx_ring->rx_stats.alloc_failed++;
 323			return NULL;
 324		}
 325
 326		/* we will be copying header into skb->data in
 327		 * pskb_may_pull so it is in our interest to prefetch
 328		 * it now to avoid a possible cache miss
 329		 */
 330		prefetchw(skb->data);
 331	}
 332
 333	/* we are reusing so sync this buffer for CPU use */
 334	dma_sync_single_range_for_cpu(rx_ring->dev,
 335				      rx_buffer->dma,
 336				      rx_buffer->page_offset,
 337				      size,
 338				      DMA_FROM_DEVICE);
 339
 340	/* pull page into skb */
 341	if (fm10k_add_rx_frag(rx_buffer, size, rx_desc, skb)) {
 342		/* hand second half of page back to the ring */
 343		fm10k_reuse_rx_page(rx_ring, rx_buffer);
 344	} else {
 345		/* we are not reusing the buffer so unmap it */
 346		dma_unmap_page(rx_ring->dev, rx_buffer->dma,
 347			       PAGE_SIZE, DMA_FROM_DEVICE);
 348	}
 349
 350	/* clear contents of rx_buffer */
 351	rx_buffer->page = NULL;
 352
 353	return skb;
 354}
 355
 356static inline void fm10k_rx_checksum(struct fm10k_ring *ring,
 357				     union fm10k_rx_desc *rx_desc,
 358				     struct sk_buff *skb)
 359{
 360	skb_checksum_none_assert(skb);
 361
 362	/* Rx checksum disabled via ethtool */
 363	if (!(ring->netdev->features & NETIF_F_RXCSUM))
 364		return;
 365
 366	/* TCP/UDP checksum error bit is set */
 367	if (fm10k_test_staterr(rx_desc,
 368			       FM10K_RXD_STATUS_L4E |
 369			       FM10K_RXD_STATUS_L4E2 |
 370			       FM10K_RXD_STATUS_IPE |
 371			       FM10K_RXD_STATUS_IPE2)) {
 372		ring->rx_stats.csum_err++;
 373		return;
 374	}
 375
 376	/* It must be a TCP or UDP packet with a valid checksum */
 377	if (fm10k_test_staterr(rx_desc, FM10K_RXD_STATUS_L4CS2))
 378		skb->encapsulation = true;
 379	else if (!fm10k_test_staterr(rx_desc, FM10K_RXD_STATUS_L4CS))
 380		return;
 381
 382	skb->ip_summed = CHECKSUM_UNNECESSARY;
 383
 384	ring->rx_stats.csum_good++;
 385}
 386
 387#define FM10K_RSS_L4_TYPES_MASK \
 388	(BIT(FM10K_RSSTYPE_IPV4_TCP) | \
 389	 BIT(FM10K_RSSTYPE_IPV4_UDP) | \
 390	 BIT(FM10K_RSSTYPE_IPV6_TCP) | \
 391	 BIT(FM10K_RSSTYPE_IPV6_UDP))
 392
 393static inline void fm10k_rx_hash(struct fm10k_ring *ring,
 394				 union fm10k_rx_desc *rx_desc,
 395				 struct sk_buff *skb)
 396{
 397	u16 rss_type;
 398
 399	if (!(ring->netdev->features & NETIF_F_RXHASH))
 400		return;
 401
 402	rss_type = le16_to_cpu(rx_desc->w.pkt_info) & FM10K_RXD_RSSTYPE_MASK;
 403	if (!rss_type)
 404		return;
 405
 406	skb_set_hash(skb, le32_to_cpu(rx_desc->d.rss),
 407		     (BIT(rss_type) & FM10K_RSS_L4_TYPES_MASK) ?
 408		     PKT_HASH_TYPE_L4 : PKT_HASH_TYPE_L3);
 409}
 410
 411static void fm10k_type_trans(struct fm10k_ring *rx_ring,
 412			     union fm10k_rx_desc __maybe_unused *rx_desc,
 413			     struct sk_buff *skb)
 414{
 415	struct net_device *dev = rx_ring->netdev;
 416	struct fm10k_l2_accel *l2_accel = rcu_dereference_bh(rx_ring->l2_accel);
 417
 418	/* check to see if DGLORT belongs to a MACVLAN */
 419	if (l2_accel) {
 420		u16 idx = le16_to_cpu(FM10K_CB(skb)->fi.w.dglort) - 1;
 421
 422		idx -= l2_accel->dglort;
 423		if (idx < l2_accel->size && l2_accel->macvlan[idx])
 424			dev = l2_accel->macvlan[idx];
 425		else
 426			l2_accel = NULL;
 427	}
 428
 429	/* Record Rx queue, or update macvlan statistics */
 430	if (!l2_accel)
 431		skb_record_rx_queue(skb, rx_ring->queue_index);
 432	else
 433		macvlan_count_rx(netdev_priv(dev), skb->len + ETH_HLEN, true,
 434				 false);
 435
 436	skb->protocol = eth_type_trans(skb, dev);
 437}
 438
 439/**
 440 * fm10k_process_skb_fields - Populate skb header fields from Rx descriptor
 441 * @rx_ring: rx descriptor ring packet is being transacted on
 442 * @rx_desc: pointer to the EOP Rx descriptor
 443 * @skb: pointer to current skb being populated
 444 *
 445 * This function checks the ring, descriptor, and packet information in
 446 * order to populate the hash, checksum, VLAN, timestamp, protocol, and
 447 * other fields within the skb.
 448 **/
 449static unsigned int fm10k_process_skb_fields(struct fm10k_ring *rx_ring,
 450					     union fm10k_rx_desc *rx_desc,
 451					     struct sk_buff *skb)
 452{
 453	unsigned int len = skb->len;
 454
 455	fm10k_rx_hash(rx_ring, rx_desc, skb);
 456
 457	fm10k_rx_checksum(rx_ring, rx_desc, skb);
 458
 459	FM10K_CB(skb)->tstamp = rx_desc->q.timestamp;
 460
 461	FM10K_CB(skb)->fi.w.vlan = rx_desc->w.vlan;
 462
 463	FM10K_CB(skb)->fi.d.glort = rx_desc->d.glort;
 464
 465	if (rx_desc->w.vlan) {
 466		u16 vid = le16_to_cpu(rx_desc->w.vlan);
 467
 468		if ((vid & VLAN_VID_MASK) != rx_ring->vid)
 469			__vlan_hwaccel_put_tag(skb, htons(ETH_P_8021Q), vid);
 470		else if (vid & VLAN_PRIO_MASK)
 471			__vlan_hwaccel_put_tag(skb, htons(ETH_P_8021Q),
 472					       vid & VLAN_PRIO_MASK);
 473	}
 474
 475	fm10k_type_trans(rx_ring, rx_desc, skb);
 476
 477	return len;
 478}
 479
 480/**
 481 * fm10k_is_non_eop - process handling of non-EOP buffers
 482 * @rx_ring: Rx ring being processed
 483 * @rx_desc: Rx descriptor for current buffer
 484 *
 485 * This function updates next to clean.  If the buffer is an EOP buffer
 486 * this function exits returning false, otherwise it will place the
 487 * sk_buff in the next buffer to be chained and return true indicating
 488 * that this is in fact a non-EOP buffer.
 489 **/
 490static bool fm10k_is_non_eop(struct fm10k_ring *rx_ring,
 491			     union fm10k_rx_desc *rx_desc)
 492{
 493	u32 ntc = rx_ring->next_to_clean + 1;
 494
 495	/* fetch, update, and store next to clean */
 496	ntc = (ntc < rx_ring->count) ? ntc : 0;
 497	rx_ring->next_to_clean = ntc;
 498
 499	prefetch(FM10K_RX_DESC(rx_ring, ntc));
 500
 501	if (likely(fm10k_test_staterr(rx_desc, FM10K_RXD_STATUS_EOP)))
 502		return false;
 503
 504	return true;
 505}
 506
 507/**
 508 * fm10k_cleanup_headers - Correct corrupted or empty headers
 509 * @rx_ring: rx descriptor ring packet is being transacted on
 510 * @rx_desc: pointer to the EOP Rx descriptor
 511 * @skb: pointer to current skb being fixed
 512 *
 513 * Address the case where we are pulling data in on pages only
 514 * and as such no data is present in the skb header.
 515 *
 516 * In addition if skb is not at least 60 bytes we need to pad it so that
 517 * it is large enough to qualify as a valid Ethernet frame.
 518 *
 519 * Returns true if an error was encountered and skb was freed.
 520 **/
 521static bool fm10k_cleanup_headers(struct fm10k_ring *rx_ring,
 522				  union fm10k_rx_desc *rx_desc,
 523				  struct sk_buff *skb)
 524{
 525	if (unlikely((fm10k_test_staterr(rx_desc,
 526					 FM10K_RXD_STATUS_RXE)))) {
 527#define FM10K_TEST_RXD_BIT(rxd, bit) \
 528	((rxd)->w.csum_err & cpu_to_le16(bit))
 529		if (FM10K_TEST_RXD_BIT(rx_desc, FM10K_RXD_ERR_SWITCH_ERROR))
 530			rx_ring->rx_stats.switch_errors++;
 531		if (FM10K_TEST_RXD_BIT(rx_desc, FM10K_RXD_ERR_NO_DESCRIPTOR))
 532			rx_ring->rx_stats.drops++;
 533		if (FM10K_TEST_RXD_BIT(rx_desc, FM10K_RXD_ERR_PP_ERROR))
 534			rx_ring->rx_stats.pp_errors++;
 535		if (FM10K_TEST_RXD_BIT(rx_desc, FM10K_RXD_ERR_SWITCH_READY))
 536			rx_ring->rx_stats.link_errors++;
 537		if (FM10K_TEST_RXD_BIT(rx_desc, FM10K_RXD_ERR_TOO_BIG))
 538			rx_ring->rx_stats.length_errors++;
 539		dev_kfree_skb_any(skb);
 540		rx_ring->rx_stats.errors++;
 541		return true;
 542	}
 543
 544	/* if eth_skb_pad returns an error the skb was freed */
 545	if (eth_skb_pad(skb))
 546		return true;
 547
 548	return false;
 549}
 550
 551/**
 552 * fm10k_receive_skb - helper function to handle rx indications
 553 * @q_vector: structure containing interrupt and ring information
 554 * @skb: packet to send up
 555 **/
 556static void fm10k_receive_skb(struct fm10k_q_vector *q_vector,
 557			      struct sk_buff *skb)
 558{
 559	napi_gro_receive(&q_vector->napi, skb);
 560}
 561
 562static int fm10k_clean_rx_irq(struct fm10k_q_vector *q_vector,
 563			      struct fm10k_ring *rx_ring,
 564			      int budget)
 565{
 566	struct sk_buff *skb = rx_ring->skb;
 567	unsigned int total_bytes = 0, total_packets = 0;
 568	u16 cleaned_count = fm10k_desc_unused(rx_ring);
 569
 570	while (likely(total_packets < budget)) {
 571		union fm10k_rx_desc *rx_desc;
 572
 573		/* return some buffers to hardware, one at a time is too slow */
 574		if (cleaned_count >= FM10K_RX_BUFFER_WRITE) {
 575			fm10k_alloc_rx_buffers(rx_ring, cleaned_count);
 576			cleaned_count = 0;
 577		}
 578
 579		rx_desc = FM10K_RX_DESC(rx_ring, rx_ring->next_to_clean);
 580
 581		if (!rx_desc->d.staterr)
 582			break;
 583
 584		/* This memory barrier is needed to keep us from reading
 585		 * any other fields out of the rx_desc until we know the
 586		 * descriptor has been written back
 587		 */
 588		dma_rmb();
 589
 590		/* retrieve a buffer from the ring */
 591		skb = fm10k_fetch_rx_buffer(rx_ring, rx_desc, skb);
 592
 593		/* exit if we failed to retrieve a buffer */
 594		if (!skb)
 595			break;
 596
 597		cleaned_count++;
 598
 599		/* fetch next buffer in frame if non-eop */
 600		if (fm10k_is_non_eop(rx_ring, rx_desc))
 601			continue;
 602
 603		/* verify the packet layout is correct */
 604		if (fm10k_cleanup_headers(rx_ring, rx_desc, skb)) {
 605			skb = NULL;
 606			continue;
 607		}
 608
 609		/* populate checksum, timestamp, VLAN, and protocol */
 610		total_bytes += fm10k_process_skb_fields(rx_ring, rx_desc, skb);
 611
 612		fm10k_receive_skb(q_vector, skb);
 613
 614		/* reset skb pointer */
 615		skb = NULL;
 616
 617		/* update budget accounting */
 618		total_packets++;
 619	}
 620
 621	/* place incomplete frames back on ring for completion */
 622	rx_ring->skb = skb;
 623
 624	u64_stats_update_begin(&rx_ring->syncp);
 625	rx_ring->stats.packets += total_packets;
 626	rx_ring->stats.bytes += total_bytes;
 627	u64_stats_update_end(&rx_ring->syncp);
 628	q_vector->rx.total_packets += total_packets;
 629	q_vector->rx.total_bytes += total_bytes;
 630
 631	return total_packets;
 632}
 633
 634#define VXLAN_HLEN (sizeof(struct udphdr) + 8)
 635static struct ethhdr *fm10k_port_is_vxlan(struct sk_buff *skb)
 636{
 637	struct fm10k_intfc *interface = netdev_priv(skb->dev);
 638
 639	if (interface->vxlan_port != udp_hdr(skb)->dest)
 640		return NULL;
 641
 642	/* return offset of udp_hdr plus 8 bytes for VXLAN header */
 643	return (struct ethhdr *)(skb_transport_header(skb) + VXLAN_HLEN);
 644}
 645
 646#define FM10K_NVGRE_RESERVED0_FLAGS htons(0x9FFF)
 647#define NVGRE_TNI htons(0x2000)
 648struct fm10k_nvgre_hdr {
 649	__be16 flags;
 650	__be16 proto;
 651	__be32 tni;
 652};
 653
 654static struct ethhdr *fm10k_gre_is_nvgre(struct sk_buff *skb)
 655{
 656	struct fm10k_nvgre_hdr *nvgre_hdr;
 657	int hlen = ip_hdrlen(skb);
 658
 659	/* currently only IPv4 is supported due to hlen above */
 660	if (vlan_get_protocol(skb) != htons(ETH_P_IP))
 661		return NULL;
 662
 663	/* our transport header should be NVGRE */
 664	nvgre_hdr = (struct fm10k_nvgre_hdr *)(skb_network_header(skb) + hlen);
 665
 666	/* verify all reserved flags are 0 */
 667	if (nvgre_hdr->flags & FM10K_NVGRE_RESERVED0_FLAGS)
 668		return NULL;
 669
 670	/* report start of ethernet header */
 671	if (nvgre_hdr->flags & NVGRE_TNI)
 672		return (struct ethhdr *)(nvgre_hdr + 1);
 673
 674	return (struct ethhdr *)(&nvgre_hdr->tni);
 675}
 676
 677__be16 fm10k_tx_encap_offload(struct sk_buff *skb)
 678{
 679	u8 l4_hdr = 0, inner_l4_hdr = 0, inner_l4_hlen;
 680	struct ethhdr *eth_hdr;
 681
 682	if (skb->inner_protocol_type != ENCAP_TYPE_ETHER ||
 683	    skb->inner_protocol != htons(ETH_P_TEB))
 684		return 0;
 685
 686	switch (vlan_get_protocol(skb)) {
 687	case htons(ETH_P_IP):
 688		l4_hdr = ip_hdr(skb)->protocol;
 689		break;
 690	case htons(ETH_P_IPV6):
 691		l4_hdr = ipv6_hdr(skb)->nexthdr;
 692		break;
 693	default:
 694		return 0;
 695	}
 696
 697	switch (l4_hdr) {
 698	case IPPROTO_UDP:
 699		eth_hdr = fm10k_port_is_vxlan(skb);
 700		break;
 701	case IPPROTO_GRE:
 702		eth_hdr = fm10k_gre_is_nvgre(skb);
 703		break;
 704	default:
 705		return 0;
 706	}
 707
 708	if (!eth_hdr)
 709		return 0;
 710
 711	switch (eth_hdr->h_proto) {
 712	case htons(ETH_P_IP):
 713		inner_l4_hdr = inner_ip_hdr(skb)->protocol;
 714		break;
 715	case htons(ETH_P_IPV6):
 716		inner_l4_hdr = inner_ipv6_hdr(skb)->nexthdr;
 717		break;
 718	default:
 719		return 0;
 720	}
 721
 722	switch (inner_l4_hdr) {
 723	case IPPROTO_TCP:
 724		inner_l4_hlen = inner_tcp_hdrlen(skb);
 725		break;
 726	case IPPROTO_UDP:
 727		inner_l4_hlen = 8;
 728		break;
 729	default:
 730		return 0;
 731	}
 732
 733	/* The hardware allows tunnel offloads only if the combined inner and
 734	 * outer header is 184 bytes or less
 735	 */
 736	if (skb_inner_transport_header(skb) + inner_l4_hlen -
 737	    skb_mac_header(skb) > FM10K_TUNNEL_HEADER_LENGTH)
 738		return 0;
 739
 740	return eth_hdr->h_proto;
 741}
 742
 743static int fm10k_tso(struct fm10k_ring *tx_ring,
 744		     struct fm10k_tx_buffer *first)
 745{
 746	struct sk_buff *skb = first->skb;
 747	struct fm10k_tx_desc *tx_desc;
 748	unsigned char *th;
 749	u8 hdrlen;
 750
 751	if (skb->ip_summed != CHECKSUM_PARTIAL)
 752		return 0;
 753
 754	if (!skb_is_gso(skb))
 755		return 0;
 756
 757	/* compute header lengths */
 758	if (skb->encapsulation) {
 759		if (!fm10k_tx_encap_offload(skb))
 760			goto err_vxlan;
 761		th = skb_inner_transport_header(skb);
 762	} else {
 763		th = skb_transport_header(skb);
 764	}
 765
 766	/* compute offset from SOF to transport header and add header len */
 767	hdrlen = (th - skb->data) + (((struct tcphdr *)th)->doff << 2);
 768
 769	first->tx_flags |= FM10K_TX_FLAGS_CSUM;
 770
 771	/* update gso size and bytecount with header size */
 772	first->gso_segs = skb_shinfo(skb)->gso_segs;
 773	first->bytecount += (first->gso_segs - 1) * hdrlen;
 774
 775	/* populate Tx descriptor header size and mss */
 776	tx_desc = FM10K_TX_DESC(tx_ring, tx_ring->next_to_use);
 777	tx_desc->hdrlen = hdrlen;
 778	tx_desc->mss = cpu_to_le16(skb_shinfo(skb)->gso_size);
 779
 780	return 1;
 781
 782err_vxlan:
 783	tx_ring->netdev->features &= ~NETIF_F_GSO_UDP_TUNNEL;
 784	if (net_ratelimit())
 785		netdev_err(tx_ring->netdev,
 786			   "TSO requested for unsupported tunnel, disabling offload\n");
 787	return -1;
 788}
 789
 790static void fm10k_tx_csum(struct fm10k_ring *tx_ring,
 791			  struct fm10k_tx_buffer *first)
 792{
 793	struct sk_buff *skb = first->skb;
 794	struct fm10k_tx_desc *tx_desc;
 795	union {
 796		struct iphdr *ipv4;
 797		struct ipv6hdr *ipv6;
 798		u8 *raw;
 799	} network_hdr;
 800	u8 *transport_hdr;
 801	__be16 frag_off;
 802	__be16 protocol;
 803	u8 l4_hdr = 0;
 804
 805	if (skb->ip_summed != CHECKSUM_PARTIAL)
 806		goto no_csum;
 807
 808	if (skb->encapsulation) {
 809		protocol = fm10k_tx_encap_offload(skb);
 810		if (!protocol) {
 811			if (skb_checksum_help(skb)) {
 812				dev_warn(tx_ring->dev,
 813					 "failed to offload encap csum!\n");
 814				tx_ring->tx_stats.csum_err++;
 815			}
 816			goto no_csum;
 817		}
 818		network_hdr.raw = skb_inner_network_header(skb);
 819		transport_hdr = skb_inner_transport_header(skb);
 820	} else {
 821		protocol = vlan_get_protocol(skb);
 822		network_hdr.raw = skb_network_header(skb);
 823		transport_hdr = skb_transport_header(skb);
 824	}
 825
 826	switch (protocol) {
 827	case htons(ETH_P_IP):
 828		l4_hdr = network_hdr.ipv4->protocol;
 829		break;
 830	case htons(ETH_P_IPV6):
 831		l4_hdr = network_hdr.ipv6->nexthdr;
 832		if (likely((transport_hdr - network_hdr.raw) ==
 833			   sizeof(struct ipv6hdr)))
 834			break;
 835		ipv6_skip_exthdr(skb, network_hdr.raw - skb->data +
 836				      sizeof(struct ipv6hdr),
 837				 &l4_hdr, &frag_off);
 838		if (unlikely(frag_off))
 839			l4_hdr = NEXTHDR_FRAGMENT;
 840		break;
 841	default:
 842		break;
 843	}
 844
 845	switch (l4_hdr) {
 846	case IPPROTO_TCP:
 847	case IPPROTO_UDP:
 848		break;
 849	case IPPROTO_GRE:
 850		if (skb->encapsulation)
 851			break;
 852		fallthrough;
 853	default:
 854		if (unlikely(net_ratelimit())) {
 855			dev_warn(tx_ring->dev,
 856				 "partial checksum, version=%d l4 proto=%x\n",
 857				 protocol, l4_hdr);
 858		}
 859		skb_checksum_help(skb);
 860		tx_ring->tx_stats.csum_err++;
 861		goto no_csum;
 862	}
 863
 864	/* update TX checksum flag */
 865	first->tx_flags |= FM10K_TX_FLAGS_CSUM;
 866	tx_ring->tx_stats.csum_good++;
 867
 868no_csum:
 869	/* populate Tx descriptor header size and mss */
 870	tx_desc = FM10K_TX_DESC(tx_ring, tx_ring->next_to_use);
 871	tx_desc->hdrlen = 0;
 872	tx_desc->mss = 0;
 873}
 874
 875#define FM10K_SET_FLAG(_input, _flag, _result) \
 876	((_flag <= _result) ? \
 877	 ((u32)(_input & _flag) * (_result / _flag)) : \
 878	 ((u32)(_input & _flag) / (_flag / _result)))
 879
 880static u8 fm10k_tx_desc_flags(struct sk_buff *skb, u32 tx_flags)
 881{
 882	/* set type for advanced descriptor with frame checksum insertion */
 883	u32 desc_flags = 0;
 884
 885	/* set checksum offload bits */
 886	desc_flags |= FM10K_SET_FLAG(tx_flags, FM10K_TX_FLAGS_CSUM,
 887				     FM10K_TXD_FLAG_CSUM);
 888
 889	return desc_flags;
 890}
 891
 892static bool fm10k_tx_desc_push(struct fm10k_ring *tx_ring,
 893			       struct fm10k_tx_desc *tx_desc, u16 i,
 894			       dma_addr_t dma, unsigned int size, u8 desc_flags)
 895{
 896	/* set RS and INT for last frame in a cache line */
 897	if ((++i & (FM10K_TXD_WB_FIFO_SIZE - 1)) == 0)
 898		desc_flags |= FM10K_TXD_FLAG_RS | FM10K_TXD_FLAG_INT;
 899
 900	/* record values to descriptor */
 901	tx_desc->buffer_addr = cpu_to_le64(dma);
 902	tx_desc->flags = desc_flags;
 903	tx_desc->buflen = cpu_to_le16(size);
 904
 905	/* return true if we just wrapped the ring */
 906	return i == tx_ring->count;
 907}
 908
 909static int __fm10k_maybe_stop_tx(struct fm10k_ring *tx_ring, u16 size)
 910{
 911	netif_stop_subqueue(tx_ring->netdev, tx_ring->queue_index);
 912
 913	/* Memory barrier before checking head and tail */
 914	smp_mb();
 915
 916	/* Check again in a case another CPU has just made room available */
 917	if (likely(fm10k_desc_unused(tx_ring) < size))
 918		return -EBUSY;
 919
 920	/* A reprieve! - use start_queue because it doesn't call schedule */
 921	netif_start_subqueue(tx_ring->netdev, tx_ring->queue_index);
 922	++tx_ring->tx_stats.restart_queue;
 923	return 0;
 924}
 925
 926static inline int fm10k_maybe_stop_tx(struct fm10k_ring *tx_ring, u16 size)
 927{
 928	if (likely(fm10k_desc_unused(tx_ring) >= size))
 929		return 0;
 930	return __fm10k_maybe_stop_tx(tx_ring, size);
 931}
 932
 933static void fm10k_tx_map(struct fm10k_ring *tx_ring,
 934			 struct fm10k_tx_buffer *first)
 935{
 936	struct sk_buff *skb = first->skb;
 937	struct fm10k_tx_buffer *tx_buffer;
 938	struct fm10k_tx_desc *tx_desc;
 939	skb_frag_t *frag;
 940	unsigned char *data;
 941	dma_addr_t dma;
 942	unsigned int data_len, size;
 943	u32 tx_flags = first->tx_flags;
 944	u16 i = tx_ring->next_to_use;
 945	u8 flags = fm10k_tx_desc_flags(skb, tx_flags);
 946
 947	tx_desc = FM10K_TX_DESC(tx_ring, i);
 948
 949	/* add HW VLAN tag */
 950	if (skb_vlan_tag_present(skb))
 951		tx_desc->vlan = cpu_to_le16(skb_vlan_tag_get(skb));
 952	else
 953		tx_desc->vlan = 0;
 954
 955	size = skb_headlen(skb);
 956	data = skb->data;
 957
 958	dma = dma_map_single(tx_ring->dev, data, size, DMA_TO_DEVICE);
 959
 960	data_len = skb->data_len;
 961	tx_buffer = first;
 962
 963	for (frag = &skb_shinfo(skb)->frags[0];; frag++) {
 964		if (dma_mapping_error(tx_ring->dev, dma))
 965			goto dma_error;
 966
 967		/* record length, and DMA address */
 968		dma_unmap_len_set(tx_buffer, len, size);
 969		dma_unmap_addr_set(tx_buffer, dma, dma);
 970
 971		while (unlikely(size > FM10K_MAX_DATA_PER_TXD)) {
 972			if (fm10k_tx_desc_push(tx_ring, tx_desc++, i++, dma,
 973					       FM10K_MAX_DATA_PER_TXD, flags)) {
 974				tx_desc = FM10K_TX_DESC(tx_ring, 0);
 975				i = 0;
 976			}
 977
 978			dma += FM10K_MAX_DATA_PER_TXD;
 979			size -= FM10K_MAX_DATA_PER_TXD;
 980		}
 981
 982		if (likely(!data_len))
 983			break;
 984
 985		if (fm10k_tx_desc_push(tx_ring, tx_desc++, i++,
 986				       dma, size, flags)) {
 987			tx_desc = FM10K_TX_DESC(tx_ring, 0);
 988			i = 0;
 989		}
 990
 991		size = skb_frag_size(frag);
 992		data_len -= size;
 993
 994		dma = skb_frag_dma_map(tx_ring->dev, frag, 0, size,
 995				       DMA_TO_DEVICE);
 996
 997		tx_buffer = &tx_ring->tx_buffer[i];
 998	}
 999
1000	/* write last descriptor with LAST bit set */
1001	flags |= FM10K_TXD_FLAG_LAST;
1002
1003	if (fm10k_tx_desc_push(tx_ring, tx_desc, i++, dma, size, flags))
1004		i = 0;
1005
1006	/* record bytecount for BQL */
1007	netdev_tx_sent_queue(txring_txq(tx_ring), first->bytecount);
1008
1009	/* record SW timestamp if HW timestamp is not available */
1010	skb_tx_timestamp(first->skb);
1011
1012	/* Force memory writes to complete before letting h/w know there
1013	 * are new descriptors to fetch.  (Only applicable for weak-ordered
1014	 * memory model archs, such as IA-64).
1015	 *
1016	 * We also need this memory barrier to make certain all of the
1017	 * status bits have been updated before next_to_watch is written.
1018	 */
1019	wmb();
1020
1021	/* set next_to_watch value indicating a packet is present */
1022	first->next_to_watch = tx_desc;
1023
1024	tx_ring->next_to_use = i;
1025
1026	/* Make sure there is space in the ring for the next send. */
1027	fm10k_maybe_stop_tx(tx_ring, DESC_NEEDED);
1028
1029	/* notify HW of packet */
1030	if (netif_xmit_stopped(txring_txq(tx_ring)) || !netdev_xmit_more()) {
1031		writel(i, tx_ring->tail);
1032	}
1033
1034	return;
1035dma_error:
1036	dev_err(tx_ring->dev, "TX DMA map failed\n");
1037
1038	/* clear dma mappings for failed tx_buffer map */
1039	for (;;) {
1040		tx_buffer = &tx_ring->tx_buffer[i];
1041		fm10k_unmap_and_free_tx_resource(tx_ring, tx_buffer);
1042		if (tx_buffer == first)
1043			break;
1044		if (i == 0)
1045			i = tx_ring->count;
1046		i--;
1047	}
1048
1049	tx_ring->next_to_use = i;
1050}
1051
1052netdev_tx_t fm10k_xmit_frame_ring(struct sk_buff *skb,
1053				  struct fm10k_ring *tx_ring)
1054{
1055	u16 count = TXD_USE_COUNT(skb_headlen(skb));
1056	struct fm10k_tx_buffer *first;
1057	unsigned short f;
1058	u32 tx_flags = 0;
1059	int tso;
1060
1061	/* need: 1 descriptor per page * PAGE_SIZE/FM10K_MAX_DATA_PER_TXD,
1062	 *       + 1 desc for skb_headlen/FM10K_MAX_DATA_PER_TXD,
1063	 *       + 2 desc gap to keep tail from touching head
1064	 * otherwise try next time
1065	 */
1066	for (f = 0; f < skb_shinfo(skb)->nr_frags; f++) {
1067		skb_frag_t *frag = &skb_shinfo(skb)->frags[f];
1068
1069		count += TXD_USE_COUNT(skb_frag_size(frag));
1070	}
1071
1072	if (fm10k_maybe_stop_tx(tx_ring, count + 3)) {
1073		tx_ring->tx_stats.tx_busy++;
1074		return NETDEV_TX_BUSY;
1075	}
1076
1077	/* record the location of the first descriptor for this packet */
1078	first = &tx_ring->tx_buffer[tx_ring->next_to_use];
1079	first->skb = skb;
1080	first->bytecount = max_t(unsigned int, skb->len, ETH_ZLEN);
1081	first->gso_segs = 1;
1082
1083	/* record initial flags and protocol */
1084	first->tx_flags = tx_flags;
1085
1086	tso = fm10k_tso(tx_ring, first);
1087	if (tso < 0)
1088		goto out_drop;
1089	else if (!tso)
1090		fm10k_tx_csum(tx_ring, first);
1091
1092	fm10k_tx_map(tx_ring, first);
1093
1094	return NETDEV_TX_OK;
1095
1096out_drop:
1097	dev_kfree_skb_any(first->skb);
1098	first->skb = NULL;
1099
1100	return NETDEV_TX_OK;
1101}
1102
1103static u64 fm10k_get_tx_completed(struct fm10k_ring *ring)
1104{
1105	return ring->stats.packets;
1106}
1107
1108/**
1109 * fm10k_get_tx_pending - how many Tx descriptors not processed
1110 * @ring: the ring structure
1111 * @in_sw: is tx_pending being checked in SW or in HW?
1112 */
1113u64 fm10k_get_tx_pending(struct fm10k_ring *ring, bool in_sw)
1114{
1115	struct fm10k_intfc *interface = ring->q_vector->interface;
1116	struct fm10k_hw *hw = &interface->hw;
1117	u32 head, tail;
1118
1119	if (likely(in_sw)) {
1120		head = ring->next_to_clean;
1121		tail = ring->next_to_use;
1122	} else {
1123		head = fm10k_read_reg(hw, FM10K_TDH(ring->reg_idx));
1124		tail = fm10k_read_reg(hw, FM10K_TDT(ring->reg_idx));
1125	}
1126
1127	return ((head <= tail) ? tail : tail + ring->count) - head;
1128}
1129
1130bool fm10k_check_tx_hang(struct fm10k_ring *tx_ring)
1131{
1132	u32 tx_done = fm10k_get_tx_completed(tx_ring);
1133	u32 tx_done_old = tx_ring->tx_stats.tx_done_old;
1134	u32 tx_pending = fm10k_get_tx_pending(tx_ring, true);
1135
1136	clear_check_for_tx_hang(tx_ring);
1137
1138	/* Check for a hung queue, but be thorough. This verifies
1139	 * that a transmit has been completed since the previous
1140	 * check AND there is at least one packet pending. By
1141	 * requiring this to fail twice we avoid races with
1142	 * clearing the ARMED bit and conditions where we
1143	 * run the check_tx_hang logic with a transmit completion
1144	 * pending but without time to complete it yet.
1145	 */
1146	if (!tx_pending || (tx_done_old != tx_done)) {
1147		/* update completed stats and continue */
1148		tx_ring->tx_stats.tx_done_old = tx_done;
1149		/* reset the countdown */
1150		clear_bit(__FM10K_HANG_CHECK_ARMED, tx_ring->state);
1151
1152		return false;
1153	}
1154
1155	/* make sure it is true for two checks in a row */
1156	return test_and_set_bit(__FM10K_HANG_CHECK_ARMED, tx_ring->state);
1157}
1158
1159/**
1160 * fm10k_tx_timeout_reset - initiate reset due to Tx timeout
1161 * @interface: driver private struct
1162 **/
1163void fm10k_tx_timeout_reset(struct fm10k_intfc *interface)
1164{
1165	/* Do the reset outside of interrupt context */
1166	if (!test_bit(__FM10K_DOWN, interface->state)) {
1167		interface->tx_timeout_count++;
1168		set_bit(FM10K_FLAG_RESET_REQUESTED, interface->flags);
1169		fm10k_service_event_schedule(interface);
1170	}
1171}
1172
1173/**
1174 * fm10k_clean_tx_irq - Reclaim resources after transmit completes
1175 * @q_vector: structure containing interrupt and ring information
1176 * @tx_ring: tx ring to clean
1177 * @napi_budget: Used to determine if we are in netpoll
1178 **/
1179static bool fm10k_clean_tx_irq(struct fm10k_q_vector *q_vector,
1180			       struct fm10k_ring *tx_ring, int napi_budget)
1181{
1182	struct fm10k_intfc *interface = q_vector->interface;
1183	struct fm10k_tx_buffer *tx_buffer;
1184	struct fm10k_tx_desc *tx_desc;
1185	unsigned int total_bytes = 0, total_packets = 0;
1186	unsigned int budget = q_vector->tx.work_limit;
1187	unsigned int i = tx_ring->next_to_clean;
1188
1189	if (test_bit(__FM10K_DOWN, interface->state))
1190		return true;
1191
1192	tx_buffer = &tx_ring->tx_buffer[i];
1193	tx_desc = FM10K_TX_DESC(tx_ring, i);
1194	i -= tx_ring->count;
1195
1196	do {
1197		struct fm10k_tx_desc *eop_desc = tx_buffer->next_to_watch;
1198
1199		/* if next_to_watch is not set then there is no work pending */
1200		if (!eop_desc)
1201			break;
1202
1203		/* prevent any other reads prior to eop_desc */
1204		smp_rmb();
1205
1206		/* if DD is not set pending work has not been completed */
1207		if (!(eop_desc->flags & FM10K_TXD_FLAG_DONE))
1208			break;
1209
1210		/* clear next_to_watch to prevent false hangs */
1211		tx_buffer->next_to_watch = NULL;
1212
1213		/* update the statistics for this packet */
1214		total_bytes += tx_buffer->bytecount;
1215		total_packets += tx_buffer->gso_segs;
1216
1217		/* free the skb */
1218		napi_consume_skb(tx_buffer->skb, napi_budget);
1219
1220		/* unmap skb header data */
1221		dma_unmap_single(tx_ring->dev,
1222				 dma_unmap_addr(tx_buffer, dma),
1223				 dma_unmap_len(tx_buffer, len),
1224				 DMA_TO_DEVICE);
1225
1226		/* clear tx_buffer data */
1227		tx_buffer->skb = NULL;
1228		dma_unmap_len_set(tx_buffer, len, 0);
1229
1230		/* unmap remaining buffers */
1231		while (tx_desc != eop_desc) {
1232			tx_buffer++;
1233			tx_desc++;
1234			i++;
1235			if (unlikely(!i)) {
1236				i -= tx_ring->count;
1237				tx_buffer = tx_ring->tx_buffer;
1238				tx_desc = FM10K_TX_DESC(tx_ring, 0);
1239			}
1240
1241			/* unmap any remaining paged data */
1242			if (dma_unmap_len(tx_buffer, len)) {
1243				dma_unmap_page(tx_ring->dev,
1244					       dma_unmap_addr(tx_buffer, dma),
1245					       dma_unmap_len(tx_buffer, len),
1246					       DMA_TO_DEVICE);
1247				dma_unmap_len_set(tx_buffer, len, 0);
1248			}
1249		}
1250
1251		/* move us one more past the eop_desc for start of next pkt */
1252		tx_buffer++;
1253		tx_desc++;
1254		i++;
1255		if (unlikely(!i)) {
1256			i -= tx_ring->count;
1257			tx_buffer = tx_ring->tx_buffer;
1258			tx_desc = FM10K_TX_DESC(tx_ring, 0);
1259		}
1260
1261		/* issue prefetch for next Tx descriptor */
1262		prefetch(tx_desc);
1263
1264		/* update budget accounting */
1265		budget--;
1266	} while (likely(budget));
1267
1268	i += tx_ring->count;
1269	tx_ring->next_to_clean = i;
1270	u64_stats_update_begin(&tx_ring->syncp);
1271	tx_ring->stats.bytes += total_bytes;
1272	tx_ring->stats.packets += total_packets;
1273	u64_stats_update_end(&tx_ring->syncp);
1274	q_vector->tx.total_bytes += total_bytes;
1275	q_vector->tx.total_packets += total_packets;
1276
1277	if (check_for_tx_hang(tx_ring) && fm10k_check_tx_hang(tx_ring)) {
1278		/* schedule immediate reset if we believe we hung */
1279		struct fm10k_hw *hw = &interface->hw;
1280
1281		netif_err(interface, drv, tx_ring->netdev,
1282			  "Detected Tx Unit Hang\n"
1283			  "  Tx Queue             <%d>\n"
1284			  "  TDH, TDT             <%x>, <%x>\n"
1285			  "  next_to_use          <%x>\n"
1286			  "  next_to_clean        <%x>\n",
1287			  tx_ring->queue_index,
1288			  fm10k_read_reg(hw, FM10K_TDH(tx_ring->reg_idx)),
1289			  fm10k_read_reg(hw, FM10K_TDT(tx_ring->reg_idx)),
1290			  tx_ring->next_to_use, i);
1291
1292		netif_stop_subqueue(tx_ring->netdev,
1293				    tx_ring->queue_index);
1294
1295		netif_info(interface, probe, tx_ring->netdev,
1296			   "tx hang %d detected on queue %d, resetting interface\n",
1297			   interface->tx_timeout_count + 1,
1298			   tx_ring->queue_index);
1299
1300		fm10k_tx_timeout_reset(interface);
1301
1302		/* the netdev is about to reset, no point in enabling stuff */
1303		return true;
1304	}
1305
1306	/* notify netdev of completed buffers */
1307	netdev_tx_completed_queue(txring_txq(tx_ring),
1308				  total_packets, total_bytes);
1309
1310#define TX_WAKE_THRESHOLD min_t(u16, FM10K_MIN_TXD - 1, DESC_NEEDED * 2)
1311	if (unlikely(total_packets && netif_carrier_ok(tx_ring->netdev) &&
1312		     (fm10k_desc_unused(tx_ring) >= TX_WAKE_THRESHOLD))) {
1313		/* Make sure that anybody stopping the queue after this
1314		 * sees the new next_to_clean.
1315		 */
1316		smp_mb();
1317		if (__netif_subqueue_stopped(tx_ring->netdev,
1318					     tx_ring->queue_index) &&
1319		    !test_bit(__FM10K_DOWN, interface->state)) {
1320			netif_wake_subqueue(tx_ring->netdev,
1321					    tx_ring->queue_index);
1322			++tx_ring->tx_stats.restart_queue;
1323		}
1324	}
1325
1326	return !!budget;
1327}
1328
1329/**
1330 * fm10k_update_itr - update the dynamic ITR value based on packet size
1331 *
1332 *      Stores a new ITR value based on strictly on packet size.  The
1333 *      divisors and thresholds used by this function were determined based
1334 *      on theoretical maximum wire speed and testing data, in order to
1335 *      minimize response time while increasing bulk throughput.
1336 *
1337 * @ring_container: Container for rings to have ITR updated
1338 **/
1339static void fm10k_update_itr(struct fm10k_ring_container *ring_container)
1340{
1341	unsigned int avg_wire_size, packets, itr_round;
1342
1343	/* Only update ITR if we are using adaptive setting */
1344	if (!ITR_IS_ADAPTIVE(ring_container->itr))
1345		goto clear_counts;
1346
1347	packets = ring_container->total_packets;
1348	if (!packets)
1349		goto clear_counts;
1350
1351	avg_wire_size = ring_container->total_bytes / packets;
1352
1353	/* The following is a crude approximation of:
1354	 *  wmem_default / (size + overhead) = desired_pkts_per_int
1355	 *  rate / bits_per_byte / (size + ethernet overhead) = pkt_rate
1356	 *  (desired_pkt_rate / pkt_rate) * usecs_per_sec = ITR value
1357	 *
1358	 * Assuming wmem_default is 212992 and overhead is 640 bytes per
1359	 * packet, (256 skb, 64 headroom, 320 shared info), we can reduce the
1360	 * formula down to
1361	 *
1362	 *  (34 * (size + 24)) / (size + 640) = ITR
1363	 *
1364	 * We first do some math on the packet size and then finally bitshift
1365	 * by 8 after rounding up. We also have to account for PCIe link speed
1366	 * difference as ITR scales based on this.
1367	 */
1368	if (avg_wire_size <= 360) {
1369		/* Start at 250K ints/sec and gradually drop to 77K ints/sec */
1370		avg_wire_size *= 8;
1371		avg_wire_size += 376;
1372	} else if (avg_wire_size <= 1152) {
1373		/* 77K ints/sec to 45K ints/sec */
1374		avg_wire_size *= 3;
1375		avg_wire_size += 2176;
1376	} else if (avg_wire_size <= 1920) {
1377		/* 45K ints/sec to 38K ints/sec */
1378		avg_wire_size += 4480;
1379	} else {
1380		/* plateau at a limit of 38K ints/sec */
1381		avg_wire_size = 6656;
1382	}
1383
1384	/* Perform final bitshift for division after rounding up to ensure
1385	 * that the calculation will never get below a 1. The bit shift
1386	 * accounts for changes in the ITR due to PCIe link speed.
1387	 */
1388	itr_round = READ_ONCE(ring_container->itr_scale) + 8;
1389	avg_wire_size += BIT(itr_round) - 1;
1390	avg_wire_size >>= itr_round;
1391
1392	/* write back value and retain adaptive flag */
1393	ring_container->itr = avg_wire_size | FM10K_ITR_ADAPTIVE;
1394
1395clear_counts:
1396	ring_container->total_bytes = 0;
1397	ring_container->total_packets = 0;
1398}
1399
1400static void fm10k_qv_enable(struct fm10k_q_vector *q_vector)
1401{
1402	/* Enable auto-mask and clear the current mask */
1403	u32 itr = FM10K_ITR_ENABLE;
1404
1405	/* Update Tx ITR */
1406	fm10k_update_itr(&q_vector->tx);
1407
1408	/* Update Rx ITR */
1409	fm10k_update_itr(&q_vector->rx);
1410
1411	/* Store Tx itr in timer slot 0 */
1412	itr |= (q_vector->tx.itr & FM10K_ITR_MAX);
1413
1414	/* Shift Rx itr to timer slot 1 */
1415	itr |= (q_vector->rx.itr & FM10K_ITR_MAX) << FM10K_ITR_INTERVAL1_SHIFT;
1416
1417	/* Write the final value to the ITR register */
1418	writel(itr, q_vector->itr);
1419}
1420
1421static int fm10k_poll(struct napi_struct *napi, int budget)
1422{
1423	struct fm10k_q_vector *q_vector =
1424			       container_of(napi, struct fm10k_q_vector, napi);
1425	struct fm10k_ring *ring;
1426	int per_ring_budget, work_done = 0;
1427	bool clean_complete = true;
1428
1429	fm10k_for_each_ring(ring, q_vector->tx) {
1430		if (!fm10k_clean_tx_irq(q_vector, ring, budget))
1431			clean_complete = false;
1432	}
1433
1434	/* Handle case where we are called by netpoll with a budget of 0 */
1435	if (budget <= 0)
1436		return budget;
1437
1438	/* attempt to distribute budget to each queue fairly, but don't
1439	 * allow the budget to go below 1 because we'll exit polling
1440	 */
1441	if (q_vector->rx.count > 1)
1442		per_ring_budget = max(budget / q_vector->rx.count, 1);
1443	else
1444		per_ring_budget = budget;
1445
1446	fm10k_for_each_ring(ring, q_vector->rx) {
1447		int work = fm10k_clean_rx_irq(q_vector, ring, per_ring_budget);
1448
1449		work_done += work;
1450		if (work >= per_ring_budget)
1451			clean_complete = false;
1452	}
1453
1454	/* If all work not completed, return budget and keep polling */
1455	if (!clean_complete)
1456		return budget;
1457
1458	/* Exit the polling mode, but don't re-enable interrupts if stack might
1459	 * poll us due to busy-polling
1460	 */
1461	if (likely(napi_complete_done(napi, work_done)))
1462		fm10k_qv_enable(q_vector);
1463
1464	return min(work_done, budget - 1);
1465}
1466
1467/**
1468 * fm10k_set_qos_queues: Allocate queues for a QOS-enabled device
1469 * @interface: board private structure to initialize
1470 *
1471 * When QoS (Quality of Service) is enabled, allocate queues for
1472 * each traffic class.  If multiqueue isn't available,then abort QoS
1473 * initialization.
1474 *
1475 * This function handles all combinations of Qos and RSS.
1476 *
1477 **/
1478static bool fm10k_set_qos_queues(struct fm10k_intfc *interface)
1479{
1480	struct net_device *dev = interface->netdev;
1481	struct fm10k_ring_feature *f;
1482	int rss_i, i;
1483	int pcs;
1484
1485	/* Map queue offset and counts onto allocated tx queues */
1486	pcs = netdev_get_num_tc(dev);
1487
1488	if (pcs <= 1)
1489		return false;
1490
1491	/* set QoS mask and indices */
1492	f = &interface->ring_feature[RING_F_QOS];
1493	f->indices = pcs;
1494	f->mask = BIT(fls(pcs - 1)) - 1;
1495
1496	/* determine the upper limit for our current DCB mode */
1497	rss_i = interface->hw.mac.max_queues / pcs;
1498	rss_i = BIT(fls(rss_i) - 1);
1499
1500	/* set RSS mask and indices */
1501	f = &interface->ring_feature[RING_F_RSS];
1502	rss_i = min_t(u16, rss_i, f->limit);
1503	f->indices = rss_i;
1504	f->mask = BIT(fls(rss_i - 1)) - 1;
1505
1506	/* configure pause class to queue mapping */
1507	for (i = 0; i < pcs; i++)
1508		netdev_set_tc_queue(dev, i, rss_i, rss_i * i);
1509
1510	interface->num_rx_queues = rss_i * pcs;
1511	interface->num_tx_queues = rss_i * pcs;
1512
1513	return true;
1514}
1515
1516/**
1517 * fm10k_set_rss_queues: Allocate queues for RSS
1518 * @interface: board private structure to initialize
1519 *
1520 * This is our "base" multiqueue mode.  RSS (Receive Side Scaling) will try
1521 * to allocate one Rx queue per CPU, and if available, one Tx queue per CPU.
1522 *
1523 **/
1524static bool fm10k_set_rss_queues(struct fm10k_intfc *interface)
1525{
1526	struct fm10k_ring_feature *f;
1527	u16 rss_i;
1528
1529	f = &interface->ring_feature[RING_F_RSS];
1530	rss_i = min_t(u16, interface->hw.mac.max_queues, f->limit);
1531
1532	/* record indices and power of 2 mask for RSS */
1533	f->indices = rss_i;
1534	f->mask = BIT(fls(rss_i - 1)) - 1;
1535
1536	interface->num_rx_queues = rss_i;
1537	interface->num_tx_queues = rss_i;
1538
1539	return true;
1540}
1541
1542/**
1543 * fm10k_set_num_queues: Allocate queues for device, feature dependent
1544 * @interface: board private structure to initialize
1545 *
1546 * This is the top level queue allocation routine.  The order here is very
1547 * important, starting with the "most" number of features turned on at once,
1548 * and ending with the smallest set of features.  This way large combinations
1549 * can be allocated if they're turned on, and smaller combinations are the
1550 * fall through conditions.
1551 *
1552 **/
1553static void fm10k_set_num_queues(struct fm10k_intfc *interface)
1554{
1555	/* Attempt to setup QoS and RSS first */
1556	if (fm10k_set_qos_queues(interface))
1557		return;
1558
1559	/* If we don't have QoS, just fallback to only RSS. */
1560	fm10k_set_rss_queues(interface);
1561}
1562
1563/**
1564 * fm10k_reset_num_queues - Reset the number of queues to zero
1565 * @interface: board private structure
1566 *
1567 * This function should be called whenever we need to reset the number of
1568 * queues after an error condition.
1569 */
1570static void fm10k_reset_num_queues(struct fm10k_intfc *interface)
1571{
1572	interface->num_tx_queues = 0;
1573	interface->num_rx_queues = 0;
1574	interface->num_q_vectors = 0;
1575}
1576
1577/**
1578 * fm10k_alloc_q_vector - Allocate memory for a single interrupt vector
1579 * @interface: board private structure to initialize
1580 * @v_count: q_vectors allocated on interface, used for ring interleaving
1581 * @v_idx: index of vector in interface struct
1582 * @txr_count: total number of Tx rings to allocate
1583 * @txr_idx: index of first Tx ring to allocate
1584 * @rxr_count: total number of Rx rings to allocate
1585 * @rxr_idx: index of first Rx ring to allocate
1586 *
1587 * We allocate one q_vector.  If allocation fails we return -ENOMEM.
1588 **/
1589static int fm10k_alloc_q_vector(struct fm10k_intfc *interface,
1590				unsigned int v_count, unsigned int v_idx,
1591				unsigned int txr_count, unsigned int txr_idx,
1592				unsigned int rxr_count, unsigned int rxr_idx)
1593{
1594	struct fm10k_q_vector *q_vector;
1595	struct fm10k_ring *ring;
1596	int ring_count;
1597
1598	ring_count = txr_count + rxr_count;
1599
1600	/* allocate q_vector and rings */
1601	q_vector = kzalloc(struct_size(q_vector, ring, ring_count), GFP_KERNEL);
1602	if (!q_vector)
1603		return -ENOMEM;
1604
1605	/* initialize NAPI */
1606	netif_napi_add(interface->netdev, &q_vector->napi,
1607		       fm10k_poll, NAPI_POLL_WEIGHT);
1608
1609	/* tie q_vector and interface together */
1610	interface->q_vector[v_idx] = q_vector;
1611	q_vector->interface = interface;
1612	q_vector->v_idx = v_idx;
1613
1614	/* initialize pointer to rings */
1615	ring = q_vector->ring;
1616
1617	/* save Tx ring container info */
1618	q_vector->tx.ring = ring;
1619	q_vector->tx.work_limit = FM10K_DEFAULT_TX_WORK;
1620	q_vector->tx.itr = interface->tx_itr;
1621	q_vector->tx.itr_scale = interface->hw.mac.itr_scale;
1622	q_vector->tx.count = txr_count;
1623
1624	while (txr_count) {
1625		/* assign generic ring traits */
1626		ring->dev = &interface->pdev->dev;
1627		ring->netdev = interface->netdev;
1628
1629		/* configure backlink on ring */
1630		ring->q_vector = q_vector;
1631
1632		/* apply Tx specific ring traits */
1633		ring->count = interface->tx_ring_count;
1634		ring->queue_index = txr_idx;
1635
1636		/* assign ring to interface */
1637		interface->tx_ring[txr_idx] = ring;
1638
1639		/* update count and index */
1640		txr_count--;
1641		txr_idx += v_count;
1642
1643		/* push pointer to next ring */
1644		ring++;
1645	}
1646
1647	/* save Rx ring container info */
1648	q_vector->rx.ring = ring;
1649	q_vector->rx.itr = interface->rx_itr;
1650	q_vector->rx.itr_scale = interface->hw.mac.itr_scale;
1651	q_vector->rx.count = rxr_count;
1652
1653	while (rxr_count) {
1654		/* assign generic ring traits */
1655		ring->dev = &interface->pdev->dev;
1656		ring->netdev = interface->netdev;
1657		rcu_assign_pointer(ring->l2_accel, interface->l2_accel);
1658
1659		/* configure backlink on ring */
1660		ring->q_vector = q_vector;
1661
1662		/* apply Rx specific ring traits */
1663		ring->count = interface->rx_ring_count;
1664		ring->queue_index = rxr_idx;
1665
1666		/* assign ring to interface */
1667		interface->rx_ring[rxr_idx] = ring;
1668
1669		/* update count and index */
1670		rxr_count--;
1671		rxr_idx += v_count;
1672
1673		/* push pointer to next ring */
1674		ring++;
1675	}
1676
1677	fm10k_dbg_q_vector_init(q_vector);
1678
1679	return 0;
1680}
1681
1682/**
1683 * fm10k_free_q_vector - Free memory allocated for specific interrupt vector
1684 * @interface: board private structure to initialize
1685 * @v_idx: Index of vector to be freed
1686 *
1687 * This function frees the memory allocated to the q_vector.  In addition if
1688 * NAPI is enabled it will delete any references to the NAPI struct prior
1689 * to freeing the q_vector.
1690 **/
1691static void fm10k_free_q_vector(struct fm10k_intfc *interface, int v_idx)
1692{
1693	struct fm10k_q_vector *q_vector = interface->q_vector[v_idx];
1694	struct fm10k_ring *ring;
1695
1696	fm10k_dbg_q_vector_exit(q_vector);
1697
1698	fm10k_for_each_ring(ring, q_vector->tx)
1699		interface->tx_ring[ring->queue_index] = NULL;
1700
1701	fm10k_for_each_ring(ring, q_vector->rx)
1702		interface->rx_ring[ring->queue_index] = NULL;
1703
1704	interface->q_vector[v_idx] = NULL;
1705	netif_napi_del(&q_vector->napi);
1706	kfree_rcu(q_vector, rcu);
1707}
1708
1709/**
1710 * fm10k_alloc_q_vectors - Allocate memory for interrupt vectors
1711 * @interface: board private structure to initialize
1712 *
1713 * We allocate one q_vector per queue interrupt.  If allocation fails we
1714 * return -ENOMEM.
1715 **/
1716static int fm10k_alloc_q_vectors(struct fm10k_intfc *interface)
1717{
1718	unsigned int q_vectors = interface->num_q_vectors;
1719	unsigned int rxr_remaining = interface->num_rx_queues;
1720	unsigned int txr_remaining = interface->num_tx_queues;
1721	unsigned int rxr_idx = 0, txr_idx = 0, v_idx = 0;
1722	int err;
1723
1724	if (q_vectors >= (rxr_remaining + txr_remaining)) {
1725		for (; rxr_remaining; v_idx++) {
1726			err = fm10k_alloc_q_vector(interface, q_vectors, v_idx,
1727						   0, 0, 1, rxr_idx);
1728			if (err)
1729				goto err_out;
1730
1731			/* update counts and index */
1732			rxr_remaining--;
1733			rxr_idx++;
1734		}
1735	}
1736
1737	for (; v_idx < q_vectors; v_idx++) {
1738		int rqpv = DIV_ROUND_UP(rxr_remaining, q_vectors - v_idx);
1739		int tqpv = DIV_ROUND_UP(txr_remaining, q_vectors - v_idx);
1740
1741		err = fm10k_alloc_q_vector(interface, q_vectors, v_idx,
1742					   tqpv, txr_idx,
1743					   rqpv, rxr_idx);
1744
1745		if (err)
1746			goto err_out;
1747
1748		/* update counts and index */
1749		rxr_remaining -= rqpv;
1750		txr_remaining -= tqpv;
1751		rxr_idx++;
1752		txr_idx++;
1753	}
1754
1755	return 0;
1756
1757err_out:
1758	fm10k_reset_num_queues(interface);
1759
1760	while (v_idx--)
1761		fm10k_free_q_vector(interface, v_idx);
1762
1763	return -ENOMEM;
1764}
1765
1766/**
1767 * fm10k_free_q_vectors - Free memory allocated for interrupt vectors
1768 * @interface: board private structure to initialize
1769 *
1770 * This function frees the memory allocated to the q_vectors.  In addition if
1771 * NAPI is enabled it will delete any references to the NAPI struct prior
1772 * to freeing the q_vector.
1773 **/
1774static void fm10k_free_q_vectors(struct fm10k_intfc *interface)
1775{
1776	int v_idx = interface->num_q_vectors;
1777
1778	fm10k_reset_num_queues(interface);
1779
1780	while (v_idx--)
1781		fm10k_free_q_vector(interface, v_idx);
1782}
1783
1784/**
1785 * f10k_reset_msix_capability - reset MSI-X capability
1786 * @interface: board private structure to initialize
1787 *
1788 * Reset the MSI-X capability back to its starting state
1789 **/
1790static void fm10k_reset_msix_capability(struct fm10k_intfc *interface)
1791{
1792	pci_disable_msix(interface->pdev);
1793	kfree(interface->msix_entries);
1794	interface->msix_entries = NULL;
1795}
1796
1797/**
1798 * f10k_init_msix_capability - configure MSI-X capability
1799 * @interface: board private structure to initialize
1800 *
1801 * Attempt to configure the interrupts using the best available
1802 * capabilities of the hardware and the kernel.
1803 **/
1804static int fm10k_init_msix_capability(struct fm10k_intfc *interface)
1805{
1806	struct fm10k_hw *hw = &interface->hw;
1807	int v_budget, vector;
1808
1809	/* It's easy to be greedy for MSI-X vectors, but it really
1810	 * doesn't do us much good if we have a lot more vectors
1811	 * than CPU's.  So let's be conservative and only ask for
1812	 * (roughly) the same number of vectors as there are CPU's.
1813	 * the default is to use pairs of vectors
1814	 */
1815	v_budget = max(interface->num_rx_queues, interface->num_tx_queues);
1816	v_budget = min_t(u16, v_budget, num_online_cpus());
1817
1818	/* account for vectors not related to queues */
1819	v_budget += NON_Q_VECTORS;
1820
1821	/* At the same time, hardware can only support a maximum of
1822	 * hw.mac->max_msix_vectors vectors.  With features
1823	 * such as RSS and VMDq, we can easily surpass the number of Rx and Tx
1824	 * descriptor queues supported by our device.  Thus, we cap it off in
1825	 * those rare cases where the cpu count also exceeds our vector limit.
1826	 */
1827	v_budget = min_t(int, v_budget, hw->mac.max_msix_vectors);
1828
1829	/* A failure in MSI-X entry allocation is fatal. */
1830	interface->msix_entries = kcalloc(v_budget, sizeof(struct msix_entry),
1831					  GFP_KERNEL);
1832	if (!interface->msix_entries)
1833		return -ENOMEM;
1834
1835	/* populate entry values */
1836	for (vector = 0; vector < v_budget; vector++)
1837		interface->msix_entries[vector].entry = vector;
1838
1839	/* Attempt to enable MSI-X with requested value */
1840	v_budget = pci_enable_msix_range(interface->pdev,
1841					 interface->msix_entries,
1842					 MIN_MSIX_COUNT(hw),
1843					 v_budget);
1844	if (v_budget < 0) {
1845		kfree(interface->msix_entries);
1846		interface->msix_entries = NULL;
1847		return v_budget;
1848	}
1849
1850	/* record the number of queues available for q_vectors */
1851	interface->num_q_vectors = v_budget - NON_Q_VECTORS;
1852
1853	return 0;
1854}
1855
1856/**
1857 * fm10k_cache_ring_qos - Descriptor ring to register mapping for QoS
1858 * @interface: Interface structure continaining rings and devices
1859 *
1860 * Cache the descriptor ring offsets for Qos
1861 **/
1862static bool fm10k_cache_ring_qos(struct fm10k_intfc *interface)
1863{
1864	struct net_device *dev = interface->netdev;
1865	int pc, offset, rss_i, i;
1866	u16 pc_stride = interface->ring_feature[RING_F_QOS].mask + 1;
1867	u8 num_pcs = netdev_get_num_tc(dev);
1868
1869	if (num_pcs <= 1)
1870		return false;
1871
1872	rss_i = interface->ring_feature[RING_F_RSS].indices;
1873
1874	for (pc = 0, offset = 0; pc < num_pcs; pc++, offset += rss_i) {
1875		int q_idx = pc;
1876
1877		for (i = 0; i < rss_i; i++) {
1878			interface->tx_ring[offset + i]->reg_idx = q_idx;
1879			interface->tx_ring[offset + i]->qos_pc = pc;
1880			interface->rx_ring[offset + i]->reg_idx = q_idx;
1881			interface->rx_ring[offset + i]->qos_pc = pc;
1882			q_idx += pc_stride;
1883		}
1884	}
1885
1886	return true;
1887}
1888
1889/**
1890 * fm10k_cache_ring_rss - Descriptor ring to register mapping for RSS
1891 * @interface: Interface structure continaining rings and devices
1892 *
1893 * Cache the descriptor ring offsets for RSS
1894 **/
1895static void fm10k_cache_ring_rss(struct fm10k_intfc *interface)
1896{
1897	int i;
1898
1899	for (i = 0; i < interface->num_rx_queues; i++)
1900		interface->rx_ring[i]->reg_idx = i;
1901
1902	for (i = 0; i < interface->num_tx_queues; i++)
1903		interface->tx_ring[i]->reg_idx = i;
1904}
1905
1906/**
1907 * fm10k_assign_rings - Map rings to network devices
1908 * @interface: Interface structure containing rings and devices
1909 *
1910 * This function is meant to go though and configure both the network
1911 * devices so that they contain rings, and configure the rings so that
1912 * they function with their network devices.
1913 **/
1914static void fm10k_assign_rings(struct fm10k_intfc *interface)
1915{
1916	if (fm10k_cache_ring_qos(interface))
1917		return;
1918
1919	fm10k_cache_ring_rss(interface);
1920}
1921
1922static void fm10k_init_reta(struct fm10k_intfc *interface)
1923{
1924	u16 i, rss_i = interface->ring_feature[RING_F_RSS].indices;
1925	u32 reta;
1926
1927	/* If the Rx flow indirection table has been configured manually, we
1928	 * need to maintain it when possible.
1929	 */
1930	if (netif_is_rxfh_configured(interface->netdev)) {
1931		for (i = FM10K_RETA_SIZE; i--;) {
1932			reta = interface->reta[i];
1933			if ((((reta << 24) >> 24) < rss_i) &&
1934			    (((reta << 16) >> 24) < rss_i) &&
1935			    (((reta <<  8) >> 24) < rss_i) &&
1936			    (((reta)       >> 24) < rss_i))
1937				continue;
1938
1939			/* this should never happen */
1940			dev_err(&interface->pdev->dev,
1941				"RSS indirection table assigned flows out of queue bounds. Reconfiguring.\n");
1942			goto repopulate_reta;
1943		}
1944
1945		/* do nothing if all of the elements are in bounds */
1946		return;
1947	}
1948
1949repopulate_reta:
1950	fm10k_write_reta(interface, NULL);
1951}
1952
1953/**
1954 * fm10k_init_queueing_scheme - Determine proper queueing scheme
1955 * @interface: board private structure to initialize
1956 *
1957 * We determine which queueing scheme to use based on...
1958 * - Hardware queue count (num_*_queues)
1959 *   - defined by miscellaneous hardware support/features (RSS, etc.)
1960 **/
1961int fm10k_init_queueing_scheme(struct fm10k_intfc *interface)
1962{
1963	int err;
1964
1965	/* Number of supported queues */
1966	fm10k_set_num_queues(interface);
1967
1968	/* Configure MSI-X capability */
1969	err = fm10k_init_msix_capability(interface);
1970	if (err) {
1971		dev_err(&interface->pdev->dev,
1972			"Unable to initialize MSI-X capability\n");
1973		goto err_init_msix;
1974	}
1975
1976	/* Allocate memory for queues */
1977	err = fm10k_alloc_q_vectors(interface);
1978	if (err) {
1979		dev_err(&interface->pdev->dev,
1980			"Unable to allocate queue vectors\n");
1981		goto err_alloc_q_vectors;
1982	}
1983
1984	/* Map rings to devices, and map devices to physical queues */
1985	fm10k_assign_rings(interface);
1986
1987	/* Initialize RSS redirection table */
1988	fm10k_init_reta(interface);
1989
1990	return 0;
1991
1992err_alloc_q_vectors:
1993	fm10k_reset_msix_capability(interface);
1994err_init_msix:
1995	fm10k_reset_num_queues(interface);
1996	return err;
1997}
1998
1999/**
2000 * fm10k_clear_queueing_scheme - Clear the current queueing scheme settings
2001 * @interface: board private structure to clear queueing scheme on
2002 *
2003 * We go through and clear queueing specific resources and reset the structure
2004 * to pre-load conditions
2005 **/
2006void fm10k_clear_queueing_scheme(struct fm10k_intfc *interface)
2007{
2008	fm10k_free_q_vectors(interface);
2009	fm10k_reset_msix_capability(interface);
2010}