drivers/net/ethernet/sfc/rx.c

   1 /****************************************************************************
   2  * Driver for Solarflare network controllers and boards
   3  * Copyright 2005-2006 Fen Systems Ltd.
   4  * Copyright 2005-2013 Solarflare Communications Inc.
   5  *
   6  * This program is free software; you can redistribute it and/or modify it
   7  * under the terms of the GNU General Public License version 2 as published
   8  * by the Free Software Foundation, incorporated herein by reference.
   9  */
  10
  11 #include <linux/socket.h>
  12 #include <linux/in.h>
  13 #include <linux/slab.h>
  14 #include <linux/ip.h>
  15 #include <linux/tcp.h>
  16 #include <linux/udp.h>
  17 #include <linux/prefetch.h>
  18 #include <linux/moduleparam.h>
  19 #include <linux/iommu.h>
  20 #include <net/ip.h>
  21 #include <net/checksum.h>
  22 #include "net_driver.h"
  23 #include "efx.h"
  24 #include "filter.h"
  25 #include "nic.h"
  26 #include "selftest.h"
  27 #include "workarounds.h"
  28
  29 /* Preferred number of descriptors to fill at once */
  30 #define EFX_RX_PREFERRED_BATCH 8U
  31
  32 /* Number of RX buffers to recycle pages for.  When creating the RX page recycle
  33  * ring, this number is divided by the number of buffers per page to calculate
  34  * the number of pages to store in the RX page recycle ring.
  35  */
  36 #define EFX_RECYCLE_RING_SIZE_IOMMU 4096
  37 #define EFX_RECYCLE_RING_SIZE_NOIOMMU (2 * EFX_RX_PREFERRED_BATCH)
  38
  39 /* Size of buffer allocated for skb header area. */
  40 #define EFX_SKB_HEADERS  128u
  41
  42 /* This is the percentage fill level below which new RX descriptors
  43  * will be added to the RX descriptor ring.
  44  */
  45 static unsigned int rx_refill_threshold;
  46
  47 /* Each packet can consume up to ceil(max_frame_len / buffer_size) buffers */
  48 #define EFX_RX_MAX_FRAGS DIV_ROUND_UP(EFX_MAX_FRAME_LEN(EFX_MAX_MTU), \
  49                                       EFX_RX_USR_BUF_SIZE)
  50
  51 /*
  52  * RX maximum head room required.
  53  *
  54  * This must be at least 1 to prevent overflow, plus one packet-worth
  55  * to allow pipelined receives.
  56  */
  57 #define EFX_RXD_HEAD_ROOM (1 + EFX_RX_MAX_FRAGS)
  58
  59 static inline u8 *efx_rx_buf_va(struct efx_rx_buffer *buf)
  60 {
  61         return page_address(buf->page) + buf->page_offset;
  62 }
  63
  64 static inline u32 efx_rx_buf_hash(struct efx_nic *efx, const u8 *eh)
  65 {
  66 #if defined(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS)
  67         return __le32_to_cpup((const __le32 *)(eh + efx->rx_packet_hash_offset));
  68 #else
  69         const u8 *data = eh + efx->rx_packet_hash_offset;
  70         return (u32)data[0]       |
  71                (u32)data[1] << 8  |
  72                (u32)data[2] << 16 |
  73                (u32)data[3] << 24;
  74 #endif
  75 }
  76
  77 static inline struct efx_rx_buffer *
  78 efx_rx_buf_next(struct efx_rx_queue *rx_queue, struct efx_rx_buffer *rx_buf)
  79 {
  80         if (unlikely(rx_buf == efx_rx_buffer(rx_queue, rx_queue->ptr_mask)))
  81                 return efx_rx_buffer(rx_queue, 0);
  82         else
  83                 return rx_buf + 1;
  84 }
  85
  86 static inline void efx_sync_rx_buffer(struct efx_nic *efx,
  87                                       struct efx_rx_buffer *rx_buf,
  88                                       unsigned int len)
  89 {
  90         dma_sync_single_for_cpu(&efx->pci_dev->dev, rx_buf->dma_addr, len,
  91                                 DMA_FROM_DEVICE);
  92 }
  93
  94 void efx_rx_config_page_split(struct efx_nic *efx)
  95 {
  96         efx->rx_page_buf_step = ALIGN(efx->rx_dma_len + NET_IP_ALIGN,
  97                                       EFX_RX_BUF_ALIGNMENT);
  98         efx->rx_bufs_per_page = efx->rx_buffer_order ? 1 :
  99                 ((PAGE_SIZE - sizeof(struct efx_rx_page_state)) /
 100                  efx->rx_page_buf_step);
 101         efx->rx_buffer_truesize = (PAGE_SIZE << efx->rx_buffer_order) /
 102                 efx->rx_bufs_per_page;
 103         efx->rx_pages_per_batch = DIV_ROUND_UP(EFX_RX_PREFERRED_BATCH,
 104                                                efx->rx_bufs_per_page);
 105 }
 106
 107 /* Check the RX page recycle ring for a page that can be reused. */
 108 static struct page *efx_reuse_page(struct efx_rx_queue *rx_queue)
 109 {
 110         struct efx_nic *efx = rx_queue->efx;
 111         struct page *page;
 112         struct efx_rx_page_state *state;
 113         unsigned index;
 114
 115         index = rx_queue->page_remove & rx_queue->page_ptr_mask;
 116         page = rx_queue->page_ring[index];
 117         if (page == NULL)
 118                 return NULL;
 119
 120         rx_queue->page_ring[index] = NULL;
 121         /* page_remove cannot exceed page_add. */
 122         if (rx_queue->page_remove != rx_queue->page_add)
 123                 ++rx_queue->page_remove;
 124
 125         /* If page_count is 1 then we hold the only reference to this page. */
 126         if (page_count(page) == 1) {
 127                 ++rx_queue->page_recycle_count;
 128                 return page;
 129         } else {
 130                 state = page_address(page);
 131                 dma_unmap_page(&efx->pci_dev->dev, state->dma_addr,
 132                                PAGE_SIZE << efx->rx_buffer_order,
 133                                DMA_FROM_DEVICE);
 134                 put_page(page);
 135                 ++rx_queue->page_recycle_failed;
 136         }
 137
 138         return NULL;
 139 }
 140
 141 /**
 142  * efx_init_rx_buffers - create EFX_RX_BATCH page-based RX buffers
 143  *
 144  * @rx_queue:           Efx RX queue
 145  *
 146  * This allocates a batch of pages, maps them for DMA, and populates
 147  * struct efx_rx_buffers for each one. Return a negative error code or
 148  * 0 on success. If a single page can be used for multiple buffers,
 149  * then the page will either be inserted fully, or not at all.
 150  */
 151 static int efx_init_rx_buffers(struct efx_rx_queue *rx_queue)
 152 {
 153         struct efx_nic *efx = rx_queue->efx;
 154         struct efx_rx_buffer *rx_buf;
 155         struct page *page;
 156         unsigned int page_offset;
 157         struct efx_rx_page_state *state;
 158         dma_addr_t dma_addr;
 159         unsigned index, count;
 160
 161         count = 0;
 162         do {
 163                 page = efx_reuse_page(rx_queue);
 164                 if (page == NULL) {
 165                         page = alloc_pages(__GFP_COLD | __GFP_COMP | GFP_ATOMIC,
 166                                            efx->rx_buffer_order);
 167                         if (unlikely(page == NULL))
 168                                 return -ENOMEM;
 169                         dma_addr =
 170                                 dma_map_page(&efx->pci_dev->dev, page, 0,
 171                                              PAGE_SIZE << efx->rx_buffer_order,
 172                                              DMA_FROM_DEVICE);
 173                         if (unlikely(dma_mapping_error(&efx->pci_dev->dev,
 174                                                        dma_addr))) {
 175                                 __free_pages(page, efx->rx_buffer_order);
 176                                 return -EIO;
 177                         }
 178                         state = page_address(page);
 179                         state->dma_addr = dma_addr;
 180                 } else {
 181                         state = page_address(page);
 182                         dma_addr = state->dma_addr;
 183                 }
 184
 185                 dma_addr += sizeof(struct efx_rx_page_state);
 186                 page_offset = sizeof(struct efx_rx_page_state);
 187
 188                 do {
 189                         index = rx_queue->added_count & rx_queue->ptr_mask;
 190                         rx_buf = efx_rx_buffer(rx_queue, index);
 191                         rx_buf->dma_addr = dma_addr + NET_IP_ALIGN;
 192                         rx_buf->page = page;
 193                         rx_buf->page_offset = page_offset + NET_IP_ALIGN;
 194                         rx_buf->len = efx->rx_dma_len;
 195                         rx_buf->flags = 0;
 196                         ++rx_queue->added_count;
 197                         get_page(page);
 198                         dma_addr += efx->rx_page_buf_step;
 199                         page_offset += efx->rx_page_buf_step;
 200                 } while (page_offset + efx->rx_page_buf_step <= PAGE_SIZE);
 201
 202                 rx_buf->flags = EFX_RX_BUF_LAST_IN_PAGE;
 203         } while (++count < efx->rx_pages_per_batch);
 204
 205         return 0;
 206 }
 207
 208 /* Unmap a DMA-mapped page.  This function is only called for the final RX
 209  * buffer in a page.
 210  */
 211 static void efx_unmap_rx_buffer(struct efx_nic *efx,
 212                                 struct efx_rx_buffer *rx_buf)
 213 {
 214         struct page *page = rx_buf->page;
 215
 216         if (page) {
 217                 struct efx_rx_page_state *state = page_address(page);
 218                 dma_unmap_page(&efx->pci_dev->dev,
 219                                state->dma_addr,
 220                                PAGE_SIZE << efx->rx_buffer_order,
 221                                DMA_FROM_DEVICE);
 222         }
 223 }
 224
 225 static void efx_free_rx_buffer(struct efx_rx_buffer *rx_buf)
 226 {
 227         if (rx_buf->page) {
 228                 put_page(rx_buf->page);
 229                 rx_buf->page = NULL;
 230         }
 231 }
 232
 233 /* Attempt to recycle the page if there is an RX recycle ring; the page can
 234  * only be added if this is the final RX buffer, to prevent pages being used in
 235  * the descriptor ring and appearing in the recycle ring simultaneously.
 236  */
 237 static void efx_recycle_rx_page(struct efx_channel *channel,
 238                                 struct efx_rx_buffer *rx_buf)
 239 {
 240         struct page *page = rx_buf->page;
 241         struct efx_rx_queue *rx_queue = efx_channel_get_rx_queue(channel);
 242         struct efx_nic *efx = rx_queue->efx;
 243         unsigned index;
 244
 245         /* Only recycle the page after processing the final buffer. */
 246         if (!(rx_buf->flags & EFX_RX_BUF_LAST_IN_PAGE))
 247                 return;
 248
 249         index = rx_queue->page_add & rx_queue->page_ptr_mask;
 250         if (rx_queue->page_ring[index] == NULL) {
 251                 unsigned read_index = rx_queue->page_remove &
 252                         rx_queue->page_ptr_mask;
 253
 254                 /* The next slot in the recycle ring is available, but
 255                  * increment page_remove if the read pointer currently
 256                  * points here.
 257                  */
 258                 if (read_index == index)
 259                         ++rx_queue->page_remove;
 260                 rx_queue->page_ring[index] = page;
 261                 ++rx_queue->page_add;
 262                 return;
 263         }
 264         ++rx_queue->page_recycle_full;
 265         efx_unmap_rx_buffer(efx, rx_buf);
 266         put_page(rx_buf->page);
 267 }
 268
 269 static void efx_fini_rx_buffer(struct efx_rx_queue *rx_queue,
 270                                struct efx_rx_buffer *rx_buf)
 271 {
 272         /* Release the page reference we hold for the buffer. */
 273         if (rx_buf->page)
 274                 put_page(rx_buf->page);
 275
 276         /* If this is the last buffer in a page, unmap and free it. */
 277         if (rx_buf->flags & EFX_RX_BUF_LAST_IN_PAGE) {
 278                 efx_unmap_rx_buffer(rx_queue->efx, rx_buf);
 279                 efx_free_rx_buffer(rx_buf);
 280         }
 281         rx_buf->page = NULL;
 282 }
 283
 284 /* Recycle the pages that are used by buffers that have just been received. */
 285 static void efx_recycle_rx_pages(struct efx_channel *channel,
 286                                  struct efx_rx_buffer *rx_buf,
 287                                  unsigned int n_frags)
 288 {
 289         struct efx_rx_queue *rx_queue = efx_channel_get_rx_queue(channel);
 290
 291         do {
 292                 efx_recycle_rx_page(channel, rx_buf);
 293                 rx_buf = efx_rx_buf_next(rx_queue, rx_buf);
 294         } while (--n_frags);
 295 }
 296
 297 static void efx_discard_rx_packet(struct efx_channel *channel,
 298                                   struct efx_rx_buffer *rx_buf,
 299                                   unsigned int n_frags)
 300 {
 301         struct efx_rx_queue *rx_queue = efx_channel_get_rx_queue(channel);
 302
 303         efx_recycle_rx_pages(channel, rx_buf, n_frags);
 304
 305         do {
 306                 efx_free_rx_buffer(rx_buf);
 307                 rx_buf = efx_rx_buf_next(rx_queue, rx_buf);
 308         } while (--n_frags);
 309 }
 310
 311 /**
 312  * efx_fast_push_rx_descriptors - push new RX descriptors quickly
 313  * @rx_queue:           RX descriptor queue
 314  *
 315  * This will aim to fill the RX descriptor queue up to
 316  * @rx_queue->@max_fill. If there is insufficient atomic
 317  * memory to do so, a slow fill will be scheduled.
 318  *
 319  * The caller must provide serialisation (none is used here). In practise,
 320  * this means this function must run from the NAPI handler, or be called
 321  * when NAPI is disabled.
 322  */
 323 void efx_fast_push_rx_descriptors(struct efx_rx_queue *rx_queue)
 324 {
 325         struct efx_nic *efx = rx_queue->efx;
 326         unsigned int fill_level, batch_size;
 327         int space, rc = 0;
 328
 329         if (!rx_queue->refill_enabled)
 330                 return;
 331
 332         /* Calculate current fill level, and exit if we don't need to fill */
 333         fill_level = (rx_queue->added_count - rx_queue->removed_count);
 334         EFX_BUG_ON_PARANOID(fill_level > rx_queue->efx->rxq_entries);
 335         if (fill_level >= rx_queue->fast_fill_trigger)
 336                 goto out;
 337
 338         /* Record minimum fill level */
 339         if (unlikely(fill_level < rx_queue->min_fill)) {
 340                 if (fill_level)
 341                         rx_queue->min_fill = fill_level;
 342         }
 343
 344         batch_size = efx->rx_pages_per_batch * efx->rx_bufs_per_page;
 345         space = rx_queue->max_fill - fill_level;
 346         EFX_BUG_ON_PARANOID(space < batch_size);
 347
 348         netif_vdbg(rx_queue->efx, rx_status, rx_queue->efx->net_dev,
 349                    "RX queue %d fast-filling descriptor ring from"
 350                    " level %d to level %d\n",
 351                    efx_rx_queue_index(rx_queue), fill_level,
 352                    rx_queue->max_fill);
 353
 354
 355         do {
 356                 rc = efx_init_rx_buffers(rx_queue);
 357                 if (unlikely(rc)) {
 358                         /* Ensure that we don't leave the rx queue empty */
 359                         if (rx_queue->added_count == rx_queue->removed_count)
 360                                 efx_schedule_slow_fill(rx_queue);
 361                         goto out;
 362                 }
 363         } while ((space -= batch_size) >= batch_size);
 364
 365         netif_vdbg(rx_queue->efx, rx_status, rx_queue->efx->net_dev,
 366                    "RX queue %d fast-filled descriptor ring "
 367                    "to level %d\n", efx_rx_queue_index(rx_queue),
 368                    rx_queue->added_count - rx_queue->removed_count);
 369
 370  out:
 371         if (rx_queue->notified_count != rx_queue->added_count)
 372                 efx_nic_notify_rx_desc(rx_queue);
 373 }
 374
 375 void efx_rx_slow_fill(unsigned long context)
 376 {
 377         struct efx_rx_queue *rx_queue = (struct efx_rx_queue *)context;
 378
 379         /* Post an event to cause NAPI to run and refill the queue */
 380         efx_nic_generate_fill_event(rx_queue);
 381         ++rx_queue->slow_fill_count;
 382 }
 383
 384 static void efx_rx_packet__check_len(struct efx_rx_queue *rx_queue,
 385                                      struct efx_rx_buffer *rx_buf,
 386                                      int len)
 387 {
 388         struct efx_nic *efx = rx_queue->efx;
 389         unsigned max_len = rx_buf->len - efx->type->rx_buffer_padding;
 390
 391         if (likely(len <= max_len))
 392                 return;
 393
 394         /* The packet must be discarded, but this is only a fatal error
 395          * if the caller indicated it was
 396          */
 397         rx_buf->flags |= EFX_RX_PKT_DISCARD;
 398
 399         if ((len > rx_buf->len) && EFX_WORKAROUND_8071(efx)) {
 400                 if (net_ratelimit())
 401                         netif_err(efx, rx_err, efx->net_dev,
 402                                   " RX queue %d seriously overlength "
 403                                   "RX event (0x%x > 0x%x+0x%x). Leaking\n",
 404                                   efx_rx_queue_index(rx_queue), len, max_len,
 405                                   efx->type->rx_buffer_padding);
 406                 efx_schedule_reset(efx, RESET_TYPE_RX_RECOVERY);
 407         } else {
 408                 if (net_ratelimit())
 409                         netif_err(efx, rx_err, efx->net_dev,
 410                                   " RX queue %d overlength RX event "
 411                                   "(0x%x > 0x%x)\n",
 412                                   efx_rx_queue_index(rx_queue), len, max_len);
 413         }
 414
 415         efx_rx_queue_channel(rx_queue)->n_rx_overlength++;
 416 }
 417
 418 /* Pass a received packet up through GRO.  GRO can handle pages
 419  * regardless of checksum state and skbs with a good checksum.
 420  */
 421 static void
 422 efx_rx_packet_gro(struct efx_channel *channel, struct efx_rx_buffer *rx_buf,
 423                   unsigned int n_frags, u8 *eh)
 424 {
 425         struct napi_struct *napi = &channel->napi_str;
 426         gro_result_t gro_result;
 427         struct efx_nic *efx = channel->efx;
 428         struct sk_buff *skb;
 429
 430         skb = napi_get_frags(napi);
 431         if (unlikely(!skb)) {
 432                 while (n_frags--) {
 433                         put_page(rx_buf->page);
 434                         rx_buf->page = NULL;
 435                         rx_buf = efx_rx_buf_next(&channel->rx_queue, rx_buf);
 436                 }
 437                 return;
 438         }
 439
 440         if (efx->net_dev->features & NETIF_F_RXHASH)
 441                 skb->rxhash = efx_rx_buf_hash(efx, eh);
 442         skb->ip_summed = ((rx_buf->flags & EFX_RX_PKT_CSUMMED) ?
 443                           CHECKSUM_UNNECESSARY : CHECKSUM_NONE);
 444
 445         for (;;) {
 446                 skb_fill_page_desc(skb, skb_shinfo(skb)->nr_frags,
 447                                    rx_buf->page, rx_buf->page_offset,
 448                                    rx_buf->len);
 449                 rx_buf->page = NULL;
 450                 skb->len += rx_buf->len;
 451                 if (skb_shinfo(skb)->nr_frags == n_frags)
 452                         break;
 453
 454                 rx_buf = efx_rx_buf_next(&channel->rx_queue, rx_buf);
 455         }
 456
 457         skb->data_len = skb->len;
 458         skb->truesize += n_frags * efx->rx_buffer_truesize;
 459
 460         skb_record_rx_queue(skb, channel->rx_queue.core_index);
 461
 462         gro_result = napi_gro_frags(napi);
 463         if (gro_result != GRO_DROP)
 464                 channel->irq_mod_score += 2;
 465 }
 466
 467 /* Allocate and construct an SKB around page fragments */
 468 static struct sk_buff *efx_rx_mk_skb(struct efx_channel *channel,
 469                                      struct efx_rx_buffer *rx_buf,
 470                                      unsigned int n_frags,
 471                                      u8 *eh, int hdr_len)
 472 {
 473         struct efx_nic *efx = channel->efx;
 474         struct sk_buff *skb;
 475
 476         /* Allocate an SKB to store the headers */
 477         skb = netdev_alloc_skb(efx->net_dev, hdr_len + EFX_PAGE_SKB_ALIGN);
 478         if (unlikely(skb == NULL))
 479                 return NULL;
 480
 481         EFX_BUG_ON_PARANOID(rx_buf->len < hdr_len);
 482
 483         skb_reserve(skb, EFX_PAGE_SKB_ALIGN);
 484         memcpy(__skb_put(skb, hdr_len), eh, hdr_len);
 485
 486         /* Append the remaining page(s) onto the frag list */
 487         if (rx_buf->len > hdr_len) {
 488                 rx_buf->page_offset += hdr_len;
 489                 rx_buf->len -= hdr_len;
 490
 491                 for (;;) {
 492                         skb_fill_page_desc(skb, skb_shinfo(skb)->nr_frags,
 493                                            rx_buf->page, rx_buf->page_offset,
 494                                            rx_buf->len);
 495                         rx_buf->page = NULL;
 496                         skb->len += rx_buf->len;
 497                         skb->data_len += rx_buf->len;
 498                         if (skb_shinfo(skb)->nr_frags == n_frags)
 499                                 break;
 500
 501                         rx_buf = efx_rx_buf_next(&channel->rx_queue, rx_buf);
 502                 }
 503         } else {
 504                 __free_pages(rx_buf->page, efx->rx_buffer_order);
 505                 rx_buf->page = NULL;
 506                 n_frags = 0;
 507         }
 508
 509         skb->truesize += n_frags * efx->rx_buffer_truesize;
 510
 511         /* Move past the ethernet header */
 512         skb->protocol = eth_type_trans(skb, efx->net_dev);
 513
 514         return skb;
 515 }
 516
 517 void efx_rx_packet(struct efx_rx_queue *rx_queue, unsigned int index,
 518                    unsigned int n_frags, unsigned int len, u16 flags)
 519 {
 520         struct efx_nic *efx = rx_queue->efx;
 521         struct efx_channel *channel = efx_rx_queue_channel(rx_queue);
 522         struct efx_rx_buffer *rx_buf;
 523
 524         rx_buf = efx_rx_buffer(rx_queue, index);
 525         rx_buf->flags |= flags;
 526
 527         /* Validate the number of fragments and completed length */
 528         if (n_frags == 1) {
 529                 if (!(flags & EFX_RX_PKT_PREFIX_LEN))
 530                         efx_rx_packet__check_len(rx_queue, rx_buf, len);
 531         } else if (unlikely(n_frags > EFX_RX_MAX_FRAGS) ||
 532                    unlikely(len <= (n_frags - 1) * efx->rx_dma_len) ||
 533                    unlikely(len > n_frags * efx->rx_dma_len) ||
 534                    unlikely(!efx->rx_scatter)) {
 535                 /* If this isn't an explicit discard request, either
 536                  * the hardware or the driver is broken.
 537                  */
 538                 WARN_ON(!(len == 0 && rx_buf->flags & EFX_RX_PKT_DISCARD));
 539                 rx_buf->flags |= EFX_RX_PKT_DISCARD;
 540         }
 541
 542         netif_vdbg(efx, rx_status, efx->net_dev,
 543                    "RX queue %d received ids %x-%x len %d %s%s\n",
 544                    efx_rx_queue_index(rx_queue), index,
 545                    (index + n_frags - 1) & rx_queue->ptr_mask, len,
 546                    (rx_buf->flags & EFX_RX_PKT_CSUMMED) ? " [SUMMED]" : "",
 547                    (rx_buf->flags & EFX_RX_PKT_DISCARD) ? " [DISCARD]" : "");
 548
 549         /* Discard packet, if instructed to do so.  Process the
 550          * previous receive first.
 551          */
 552         if (unlikely(rx_buf->flags & EFX_RX_PKT_DISCARD)) {
 553                 efx_rx_flush_packet(channel);
 554                 efx_discard_rx_packet(channel, rx_buf, n_frags);
 555                 return;
 556         }
 557
 558         if (n_frags == 1 && !(flags & EFX_RX_PKT_PREFIX_LEN))
 559                 rx_buf->len = len;
 560
 561         /* Release and/or sync the DMA mapping - assumes all RX buffers
 562          * consumed in-order per RX queue.
 563          */
 564         efx_sync_rx_buffer(efx, rx_buf, rx_buf->len);
 565
 566         /* Prefetch nice and early so data will (hopefully) be in cache by
 567          * the time we look at it.
 568          */
 569         prefetch(efx_rx_buf_va(rx_buf));
 570
 571         rx_buf->page_offset += efx->rx_prefix_size;
 572         rx_buf->len -= efx->rx_prefix_size;
 573
 574         if (n_frags > 1) {
 575                 /* Release/sync DMA mapping for additional fragments.
 576                  * Fix length for last fragment.
 577                  */
 578                 unsigned int tail_frags = n_frags - 1;
 579
 580                 for (;;) {
 581                         rx_buf = efx_rx_buf_next(rx_queue, rx_buf);
 582                         if (--tail_frags == 0)
 583                                 break;
 584                         efx_sync_rx_buffer(efx, rx_buf, efx->rx_dma_len);
 585                 }
 586                 rx_buf->len = len - (n_frags - 1) * efx->rx_dma_len;
 587                 efx_sync_rx_buffer(efx, rx_buf, rx_buf->len);
 588         }
 589
 590         /* All fragments have been DMA-synced, so recycle pages. */
 591         rx_buf = efx_rx_buffer(rx_queue, index);
 592         efx_recycle_rx_pages(channel, rx_buf, n_frags);
 593
 594         /* Pipeline receives so that we give time for packet headers to be
 595          * prefetched into cache.
 596          */
 597         efx_rx_flush_packet(channel);
 598         channel->rx_pkt_n_frags = n_frags;
 599         channel->rx_pkt_index = index;
 600 }
 601
 602 static void efx_rx_deliver(struct efx_channel *channel, u8 *eh,
 603                            struct efx_rx_buffer *rx_buf,
 604                            unsigned int n_frags)
 605 {
 606         struct sk_buff *skb;
 607         u16 hdr_len = min_t(u16, rx_buf->len, EFX_SKB_HEADERS);
 608
 609         skb = efx_rx_mk_skb(channel, rx_buf, n_frags, eh, hdr_len);
 610         if (unlikely(skb == NULL)) {
 611                 efx_free_rx_buffer(rx_buf);
 612                 return;
 613         }
 614         skb_record_rx_queue(skb, channel->rx_queue.core_index);
 615
 616         /* Set the SKB flags */
 617         skb_checksum_none_assert(skb);
 618         if (likely(rx_buf->flags & EFX_RX_PKT_CSUMMED))
 619                 skb->ip_summed = CHECKSUM_UNNECESSARY;
 620
 621         if (channel->type->receive_skb)
 622                 if (channel->type->receive_skb(channel, skb))
 623                         return;
 624
 625         /* Pass the packet up */
 626         netif_receive_skb(skb);
 627 }
 628
 629 /* Handle a received packet.  Second half: Touches packet payload. */
 630 void __efx_rx_packet(struct efx_channel *channel)
 631 {
 632         struct efx_nic *efx = channel->efx;
 633         struct efx_rx_buffer *rx_buf =
 634                 efx_rx_buffer(&channel->rx_queue, channel->rx_pkt_index);
 635         u8 *eh = efx_rx_buf_va(rx_buf);
 636
 637         /* Read length from the prefix if necessary.  This already
 638          * excludes the length of the prefix itself.
 639          */
 640         if (rx_buf->flags & EFX_RX_PKT_PREFIX_LEN)
 641                 rx_buf->len = le16_to_cpup((__le16 *)
 642                                            (eh + efx->rx_packet_len_offset));
 643
 644         /* If we're in loopback test, then pass the packet directly to the
 645          * loopback layer, and free the rx_buf here
 646          */
 647         if (unlikely(efx->loopback_selftest)) {
 648                 efx_loopback_rx_packet(efx, eh, rx_buf->len);
 649                 efx_free_rx_buffer(rx_buf);
 650                 goto out;
 651         }
 652
 653         if (unlikely(!(efx->net_dev->features & NETIF_F_RXCSUM)))
 654                 rx_buf->flags &= ~EFX_RX_PKT_CSUMMED;
 655
 656         if ((rx_buf->flags & EFX_RX_PKT_TCP) && !channel->type->receive_skb)
 657                 efx_rx_packet_gro(channel, rx_buf, channel->rx_pkt_n_frags, eh);
 658         else
 659                 efx_rx_deliver(channel, eh, rx_buf, channel->rx_pkt_n_frags);
 660 out:
 661         channel->rx_pkt_n_frags = 0;
 662 }
 663
 664 int efx_probe_rx_queue(struct efx_rx_queue *rx_queue)
 665 {
 666         struct efx_nic *efx = rx_queue->efx;
 667         unsigned int entries;
 668         int rc;
 669
 670         /* Create the smallest power-of-two aligned ring */
 671         entries = max(roundup_pow_of_two(efx->rxq_entries), EFX_MIN_DMAQ_SIZE);
 672         EFX_BUG_ON_PARANOID(entries > EFX_MAX_DMAQ_SIZE);
 673         rx_queue->ptr_mask = entries - 1;
 674
 675         netif_dbg(efx, probe, efx->net_dev,
 676                   "creating RX queue %d size %#x mask %#x\n",
 677                   efx_rx_queue_index(rx_queue), efx->rxq_entries,
 678                   rx_queue->ptr_mask);
 679
 680         /* Allocate RX buffers */
 681         rx_queue->buffer = kcalloc(entries, sizeof(*rx_queue->buffer),
 682                                    GFP_KERNEL);
 683         if (!rx_queue->buffer)
 684                 return -ENOMEM;
 685
 686         rc = efx_nic_probe_rx(rx_queue);
 687         if (rc) {
 688                 kfree(rx_queue->buffer);
 689                 rx_queue->buffer = NULL;
 690         }
 691
 692         return rc;
 693 }
 694
 695 static void efx_init_rx_recycle_ring(struct efx_nic *efx,
 696                                      struct efx_rx_queue *rx_queue)
 697 {
 698         unsigned int bufs_in_recycle_ring, page_ring_size;
 699
 700         /* Set the RX recycle ring size */
 701 #ifdef CONFIG_PPC64
 702         bufs_in_recycle_ring = EFX_RECYCLE_RING_SIZE_IOMMU;
 703 #else
 704         if (iommu_present(&pci_bus_type))
 705                 bufs_in_recycle_ring = EFX_RECYCLE_RING_SIZE_IOMMU;
 706         else
 707                 bufs_in_recycle_ring = EFX_RECYCLE_RING_SIZE_NOIOMMU;
 708 #endif /* CONFIG_PPC64 */
 709
 710         page_ring_size = roundup_pow_of_two(bufs_in_recycle_ring /
 711                                             efx->rx_bufs_per_page);
 712         rx_queue->page_ring = kcalloc(page_ring_size,
 713                                       sizeof(*rx_queue->page_ring), GFP_KERNEL);
 714         rx_queue->page_ptr_mask = page_ring_size - 1;
 715 }
 716
 717 void efx_init_rx_queue(struct efx_rx_queue *rx_queue)
 718 {
 719         struct efx_nic *efx = rx_queue->efx;
 720         unsigned int max_fill, trigger, max_trigger;
 721
 722         netif_dbg(rx_queue->efx, drv, rx_queue->efx->net_dev,
 723                   "initialising RX queue %d\n", efx_rx_queue_index(rx_queue));
 724
 725         /* Initialise ptr fields */
 726         rx_queue->added_count = 0;
 727         rx_queue->notified_count = 0;
 728         rx_queue->removed_count = 0;
 729         rx_queue->min_fill = -1U;
 730         efx_init_rx_recycle_ring(efx, rx_queue);
 731
 732         rx_queue->page_remove = 0;
 733         rx_queue->page_add = rx_queue->page_ptr_mask + 1;
 734         rx_queue->page_recycle_count = 0;
 735         rx_queue->page_recycle_failed = 0;
 736         rx_queue->page_recycle_full = 0;
 737
 738         /* Initialise limit fields */
 739         max_fill = efx->rxq_entries - EFX_RXD_HEAD_ROOM;
 740         max_trigger =
 741                 max_fill - efx->rx_pages_per_batch * efx->rx_bufs_per_page;
 742         if (rx_refill_threshold != 0) {
 743                 trigger = max_fill * min(rx_refill_threshold, 100U) / 100U;
 744                 if (trigger > max_trigger)
 745                         trigger = max_trigger;
 746         } else {
 747                 trigger = max_trigger;
 748         }
 749
 750         rx_queue->max_fill = max_fill;
 751         rx_queue->fast_fill_trigger = trigger;
 752         rx_queue->refill_enabled = true;
 753
 754         /* Set up RX descriptor ring */
 755         efx_nic_init_rx(rx_queue);
 756 }
 757
 758 void efx_fini_rx_queue(struct efx_rx_queue *rx_queue)
 759 {
 760         int i;
 761         struct efx_nic *efx = rx_queue->efx;
 762         struct efx_rx_buffer *rx_buf;
 763
 764         netif_dbg(rx_queue->efx, drv, rx_queue->efx->net_dev,
 765                   "shutting down RX queue %d\n", efx_rx_queue_index(rx_queue));
 766
 767         del_timer_sync(&rx_queue->slow_fill);
 768
 769         /* Release RX buffers from the current read ptr to the write ptr */
 770         if (rx_queue->buffer) {
 771                 for (i = rx_queue->removed_count; i < rx_queue->added_count;
 772                      i++) {
 773                         unsigned index = i & rx_queue->ptr_mask;
 774                         rx_buf = efx_rx_buffer(rx_queue, index);
 775                         efx_fini_rx_buffer(rx_queue, rx_buf);
 776                 }
 777         }
 778
 779         /* Unmap and release the pages in the recycle ring. Remove the ring. */
 780         for (i = 0; i <= rx_queue->page_ptr_mask; i++) {
 781                 struct page *page = rx_queue->page_ring[i];
 782                 struct efx_rx_page_state *state;
 783
 784                 if (page == NULL)
 785                         continue;
 786
 787                 state = page_address(page);
 788                 dma_unmap_page(&efx->pci_dev->dev, state->dma_addr,
 789                                PAGE_SIZE << efx->rx_buffer_order,
 790                                DMA_FROM_DEVICE);
 791                 put_page(page);
 792         }
 793         kfree(rx_queue->page_ring);
 794         rx_queue->page_ring = NULL;
 795 }
 796
 797 void efx_remove_rx_queue(struct efx_rx_queue *rx_queue)
 798 {
 799         netif_dbg(rx_queue->efx, drv, rx_queue->efx->net_dev,
 800                   "destroying RX queue %d\n", efx_rx_queue_index(rx_queue));
 801
 802         efx_nic_remove_rx(rx_queue);
 803
 804         kfree(rx_queue->buffer);
 805         rx_queue->buffer = NULL;
 806 }
 807
 808
 809 module_param(rx_refill_threshold, uint, 0444);
 810 MODULE_PARM_DESC(rx_refill_threshold,
 811                  "RX descriptor ring refill threshold (%)");
 812
 813 #ifdef CONFIG_RFS_ACCEL
 814
 815 int efx_filter_rfs(struct net_device *net_dev, const struct sk_buff *skb,
 816                    u16 rxq_index, u32 flow_id)
 817 {
 818         struct efx_nic *efx = netdev_priv(net_dev);
 819         struct efx_channel *channel;
 820         struct efx_filter_spec spec;
 821         const struct iphdr *ip;
 822         const __be16 *ports;
 823         int nhoff;
 824         int rc;
 825
 826         nhoff = skb_network_offset(skb);
 827
 828         if (skb->protocol == htons(ETH_P_8021Q)) {
 829                 EFX_BUG_ON_PARANOID(skb_headlen(skb) <
 830                                     nhoff + sizeof(struct vlan_hdr));
 831                 if (((const struct vlan_hdr *)skb->data + nhoff)->
 832                     h_vlan_encapsulated_proto != htons(ETH_P_IP))
 833                         return -EPROTONOSUPPORT;
 834
 835                 /* This is IP over 802.1q VLAN.  We can't filter on the
 836                  * IP 5-tuple and the vlan together, so just strip the
 837                  * vlan header and filter on the IP part.
 838                  */
 839                 nhoff += sizeof(struct vlan_hdr);
 840         } else if (skb->protocol != htons(ETH_P_IP)) {
 841                 return -EPROTONOSUPPORT;
 842         }
 843
 844         /* RFS must validate the IP header length before calling us */
 845         EFX_BUG_ON_PARANOID(skb_headlen(skb) < nhoff + sizeof(*ip));
 846         ip = (const struct iphdr *)(skb->data + nhoff);
 847         if (ip_is_fragment(ip))
 848                 return -EPROTONOSUPPORT;
 849         EFX_BUG_ON_PARANOID(skb_headlen(skb) < nhoff + 4 * ip->ihl + 4);
 850         ports = (const __be16 *)(skb->data + nhoff + 4 * ip->ihl);
 851
 852         efx_filter_init_rx(&spec, EFX_FILTER_PRI_HINT,
 853                            efx->rx_scatter ? EFX_FILTER_FLAG_RX_SCATTER : 0,
 854                            rxq_index);
 855         rc = efx_filter_set_ipv4_full(&spec, ip->protocol,
 856                                       ip->daddr, ports[1], ip->saddr, ports[0]);
 857         if (rc)
 858                 return rc;
 859
 860         rc = efx->type->filter_rfs_insert(efx, &spec);
 861         if (rc < 0)
 862                 return rc;
 863
 864         /* Remember this so we can check whether to expire the filter later */
 865         efx->rps_flow_id[rc] = flow_id;
 866         channel = efx_get_channel(efx, skb_get_rx_queue(skb));
 867         ++channel->rfs_filters_added;
 868
 869         netif_info(efx, rx_status, efx->net_dev,
 870                    "steering %s %pI4:%u:%pI4:%u to queue %u [flow %u filter %d]\n",
 871                    (ip->protocol == IPPROTO_TCP) ? "TCP" : "UDP",
 872                    &ip->saddr, ntohs(ports[0]), &ip->daddr, ntohs(ports[1]),
 873                    rxq_index, flow_id, rc);
 874
 875         return rc;
 876 }
 877
 878 bool __efx_filter_rfs_expire(struct efx_nic *efx, unsigned int quota)
 879 {
 880         bool (*expire_one)(struct efx_nic *efx, u32 flow_id, unsigned int index);
 881         unsigned int index, size;
 882         u32 flow_id;
 883
 884         if (!spin_trylock_bh(&efx->filter_lock))
 885                 return false;
 886
 887         expire_one = efx->type->filter_rfs_expire_one;
 888         index = efx->rps_expire_index;
 889         size = efx->type->max_rx_ip_filters;
 890         while (quota--) {
 891                 flow_id = efx->rps_flow_id[index];
 892                 if (expire_one(efx, flow_id, index))
 893                         netif_info(efx, rx_status, efx->net_dev,
 894                                    "expired filter %d [flow %u]\n",
 895                                    index, flow_id);
 896                 if (++index == size)
 897                         index = 0;
 898         }
 899         efx->rps_expire_index = index;
 900
 901         spin_unlock_bh(&efx->filter_lock);
 902         return true;
 903 }
 904
 905 #endif /* CONFIG_RFS_ACCEL */
 906
 907 /**
 908  * efx_filter_is_mc_recipient - test whether spec is a multicast recipient
 909  * @spec: Specification to test
 910  *
 911  * Return: %true if the specification is a non-drop RX filter that
 912  * matches a local MAC address I/G bit value of 1 or matches a local
 913  * IPv4 or IPv6 address value in the respective multicast address
 914  * range.  Otherwise %false.
 915  */
 916 bool efx_filter_is_mc_recipient(const struct efx_filter_spec *spec)
 917 {
 918         if (!(spec->flags & EFX_FILTER_FLAG_RX) ||
 919             spec->dmaq_id == EFX_FILTER_RX_DMAQ_ID_DROP)
 920                 return false;
 921
 922         if (spec->match_flags &
 923             (EFX_FILTER_MATCH_LOC_MAC | EFX_FILTER_MATCH_LOC_MAC_IG) &&
 924             is_multicast_ether_addr(spec->loc_mac))
 925                 return true;
 926
 927         if ((spec->match_flags &
 928              (EFX_FILTER_MATCH_ETHER_TYPE | EFX_FILTER_MATCH_LOC_HOST)) ==
 929             (EFX_FILTER_MATCH_ETHER_TYPE | EFX_FILTER_MATCH_LOC_HOST)) {
 930                 if (spec->ether_type == htons(ETH_P_IP) &&
 931                     ipv4_is_multicast(spec->loc_host[0]))
 932                         return true;
 933                 if (spec->ether_type == htons(ETH_P_IPV6) &&
 934                     ((const u8 *)spec->loc_host)[0] == 0xff)
 935                         return true;
 936         }
 937
 938         return false;
 939 }