drivers/net/ethernet/sfc/rx.c

   1 /****************************************************************************
   2  * Driver for Solarflare Solarstorm network controllers and boards
   3  * Copyright 2005-2006 Fen Systems Ltd.
   4  * Copyright 2005-2011 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 "nic.h"
  25 #include "selftest.h"
  26 #include "workarounds.h"
  27
  28 /* Preferred number of descriptors to fill at once */
  29 #define EFX_RX_PREFERRED_BATCH 8U
  30
  31 /* Number of RX buffers to recycle pages for.  When creating the RX page recycle
  32  * ring, this number is divided by the number of buffers per page to calculate
  33  * the number of pages to store in the RX page recycle ring.
  34  */
  35 #define EFX_RECYCLE_RING_SIZE_IOMMU 4096
  36 #define EFX_RECYCLE_RING_SIZE_NOIOMMU (2 * EFX_RX_PREFERRED_BATCH)
  37
  38 /* Size of buffer allocated for skb header area. */
  39 #define EFX_SKB_HEADERS  64u
  40
  41 /* This is the percentage fill level below which new RX descriptors
  42  * will be added to the RX descriptor ring.
  43  */
  44 static unsigned int rx_refill_threshold;
  45
  46 /* Each packet can consume up to ceil(max_frame_len / buffer_size) buffers */
  47 #define EFX_RX_MAX_FRAGS DIV_ROUND_UP(EFX_MAX_FRAME_LEN(EFX_MAX_MTU), \
  48                                       EFX_RX_USR_BUF_SIZE)
  49
  50 /*
  51  * RX maximum head room required.
  52  *
  53  * This must be at least 1 to prevent overflow, plus one packet-worth
  54  * to allow pipelined receives.
  55  */
  56 #define EFX_RXD_HEAD_ROOM (1 + EFX_RX_MAX_FRAGS)
  57
  58 static inline u8 *efx_rx_buf_va(struct efx_rx_buffer *buf)
  59 {
  60         return page_address(buf->page) + buf->page_offset;
  61 }
  62
  63 static inline u32 efx_rx_buf_hash(const u8 *eh)
  64 {
  65         /* The ethernet header is always directly after any hash. */
  66 #if defined(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS) || NET_IP_ALIGN % 4 == 0
  67         return __le32_to_cpup((const __le32 *)(eh - 4));
  68 #else
  69         const u8 *data = eh - 4;
  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_buffers(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 /**
 298  * efx_fast_push_rx_descriptors - push new RX descriptors quickly
 299  * @rx_queue:           RX descriptor queue
 300  *
 301  * This will aim to fill the RX descriptor queue up to
 302  * @rx_queue->@max_fill. If there is insufficient atomic
 303  * memory to do so, a slow fill will be scheduled.
 304  *
 305  * The caller must provide serialisation (none is used here). In practise,
 306  * this means this function must run from the NAPI handler, or be called
 307  * when NAPI is disabled.
 308  */
 309 void efx_fast_push_rx_descriptors(struct efx_rx_queue *rx_queue)
 310 {
 311         struct efx_nic *efx = rx_queue->efx;
 312         unsigned int fill_level, batch_size;
 313         int space, rc = 0;
 314
 315         /* Calculate current fill level, and exit if we don't need to fill */
 316         fill_level = (rx_queue->added_count - rx_queue->removed_count);
 317         EFX_BUG_ON_PARANOID(fill_level > rx_queue->efx->rxq_entries);
 318         if (fill_level >= rx_queue->fast_fill_trigger)
 319                 goto out;
 320
 321         /* Record minimum fill level */
 322         if (unlikely(fill_level < rx_queue->min_fill)) {
 323                 if (fill_level)
 324                         rx_queue->min_fill = fill_level;
 325         }
 326
 327         batch_size = efx->rx_pages_per_batch * efx->rx_bufs_per_page;
 328         space = rx_queue->max_fill - fill_level;
 329         EFX_BUG_ON_PARANOID(space < batch_size);
 330
 331         netif_vdbg(rx_queue->efx, rx_status, rx_queue->efx->net_dev,
 332                    "RX queue %d fast-filling descriptor ring from"
 333                    " level %d to level %d\n",
 334                    efx_rx_queue_index(rx_queue), fill_level,
 335                    rx_queue->max_fill);
 336
 337
 338         do {
 339                 rc = efx_init_rx_buffers(rx_queue);
 340                 if (unlikely(rc)) {
 341                         /* Ensure that we don't leave the rx queue empty */
 342                         if (rx_queue->added_count == rx_queue->removed_count)
 343                                 efx_schedule_slow_fill(rx_queue);
 344                         goto out;
 345                 }
 346         } while ((space -= batch_size) >= batch_size);
 347
 348         netif_vdbg(rx_queue->efx, rx_status, rx_queue->efx->net_dev,
 349                    "RX queue %d fast-filled descriptor ring "
 350                    "to level %d\n", efx_rx_queue_index(rx_queue),
 351                    rx_queue->added_count - rx_queue->removed_count);
 352
 353  out:
 354         if (rx_queue->notified_count != rx_queue->added_count)
 355                 efx_nic_notify_rx_desc(rx_queue);
 356 }
 357
 358 void efx_rx_slow_fill(unsigned long context)
 359 {
 360         struct efx_rx_queue *rx_queue = (struct efx_rx_queue *)context;
 361
 362         /* Post an event to cause NAPI to run and refill the queue */
 363         efx_nic_generate_fill_event(rx_queue);
 364         ++rx_queue->slow_fill_count;
 365 }
 366
 367 static void efx_rx_packet__check_len(struct efx_rx_queue *rx_queue,
 368                                      struct efx_rx_buffer *rx_buf,
 369                                      int len)
 370 {
 371         struct efx_nic *efx = rx_queue->efx;
 372         unsigned max_len = rx_buf->len - efx->type->rx_buffer_padding;
 373
 374         if (likely(len <= max_len))
 375                 return;
 376
 377         /* The packet must be discarded, but this is only a fatal error
 378          * if the caller indicated it was
 379          */
 380         rx_buf->flags |= EFX_RX_PKT_DISCARD;
 381
 382         if ((len > rx_buf->len) && EFX_WORKAROUND_8071(efx)) {
 383                 if (net_ratelimit())
 384                         netif_err(efx, rx_err, efx->net_dev,
 385                                   " RX queue %d seriously overlength "
 386                                   "RX event (0x%x > 0x%x+0x%x). Leaking\n",
 387                                   efx_rx_queue_index(rx_queue), len, max_len,
 388                                   efx->type->rx_buffer_padding);
 389                 efx_schedule_reset(efx, RESET_TYPE_RX_RECOVERY);
 390         } else {
 391                 if (net_ratelimit())
 392                         netif_err(efx, rx_err, efx->net_dev,
 393                                   " RX queue %d overlength RX event "
 394                                   "(0x%x > 0x%x)\n",
 395                                   efx_rx_queue_index(rx_queue), len, max_len);
 396         }
 397
 398         efx_rx_queue_channel(rx_queue)->n_rx_overlength++;
 399 }
 400
 401 /* Pass a received packet up through GRO.  GRO can handle pages
 402  * regardless of checksum state and skbs with a good checksum.
 403  */
 404 static void
 405 efx_rx_packet_gro(struct efx_channel *channel, struct efx_rx_buffer *rx_buf,
 406                   unsigned int n_frags, u8 *eh)
 407 {
 408         struct napi_struct *napi = &channel->napi_str;
 409         gro_result_t gro_result;
 410         struct efx_nic *efx = channel->efx;
 411         struct sk_buff *skb;
 412
 413         skb = napi_get_frags(napi);
 414         if (unlikely(!skb)) {
 415                 while (n_frags--) {
 416                         put_page(rx_buf->page);
 417                         rx_buf->page = NULL;
 418                         rx_buf = efx_rx_buf_next(&channel->rx_queue, rx_buf);
 419                 }
 420                 return;
 421         }
 422
 423         if (efx->net_dev->features & NETIF_F_RXHASH)
 424                 skb->rxhash = efx_rx_buf_hash(eh);
 425         skb->ip_summed = ((rx_buf->flags & EFX_RX_PKT_CSUMMED) ?
 426                           CHECKSUM_UNNECESSARY : CHECKSUM_NONE);
 427
 428         for (;;) {
 429                 skb_fill_page_desc(skb, skb_shinfo(skb)->nr_frags,
 430                                    rx_buf->page, rx_buf->page_offset,
 431                                    rx_buf->len);
 432                 rx_buf->page = NULL;
 433                 skb->len += rx_buf->len;
 434                 if (skb_shinfo(skb)->nr_frags == n_frags)
 435                         break;
 436
 437                 rx_buf = efx_rx_buf_next(&channel->rx_queue, rx_buf);
 438         }
 439
 440         skb->data_len = skb->len;
 441         skb->truesize += n_frags * efx->rx_buffer_truesize;
 442
 443         skb_record_rx_queue(skb, channel->rx_queue.core_index);
 444
 445         gro_result = napi_gro_frags(napi);
 446         if (gro_result != GRO_DROP)
 447                 channel->irq_mod_score += 2;
 448 }
 449
 450 /* Allocate and construct an SKB around page fragments */
 451 static struct sk_buff *efx_rx_mk_skb(struct efx_channel *channel,
 452                                      struct efx_rx_buffer *rx_buf,
 453                                      unsigned int n_frags,
 454                                      u8 *eh, int hdr_len)
 455 {
 456         struct efx_nic *efx = channel->efx;
 457         struct sk_buff *skb;
 458
 459         /* Allocate an SKB to store the headers */
 460         skb = netdev_alloc_skb(efx->net_dev, hdr_len + EFX_PAGE_SKB_ALIGN);
 461         if (unlikely(skb == NULL))
 462                 return NULL;
 463
 464         EFX_BUG_ON_PARANOID(rx_buf->len < hdr_len);
 465
 466         skb_reserve(skb, EFX_PAGE_SKB_ALIGN);
 467         memcpy(__skb_put(skb, hdr_len), eh, hdr_len);
 468
 469         /* Append the remaining page(s) onto the frag list */
 470         if (rx_buf->len > hdr_len) {
 471                 rx_buf->page_offset += hdr_len;
 472                 rx_buf->len -= hdr_len;
 473
 474                 for (;;) {
 475                         skb_fill_page_desc(skb, skb_shinfo(skb)->nr_frags,
 476                                            rx_buf->page, rx_buf->page_offset,
 477                                            rx_buf->len);
 478                         rx_buf->page = NULL;
 479                         skb->len += rx_buf->len;
 480                         skb->data_len += rx_buf->len;
 481                         if (skb_shinfo(skb)->nr_frags == n_frags)
 482                                 break;
 483
 484                         rx_buf = efx_rx_buf_next(&channel->rx_queue, rx_buf);
 485                 }
 486         } else {
 487                 __free_pages(rx_buf->page, efx->rx_buffer_order);
 488                 rx_buf->page = NULL;
 489                 n_frags = 0;
 490         }
 491
 492         skb->truesize += n_frags * efx->rx_buffer_truesize;
 493
 494         /* Move past the ethernet header */
 495         skb->protocol = eth_type_trans(skb, efx->net_dev);
 496
 497         return skb;
 498 }
 499
 500 void efx_rx_packet(struct efx_rx_queue *rx_queue, unsigned int index,
 501                    unsigned int n_frags, unsigned int len, u16 flags)
 502 {
 503         struct efx_nic *efx = rx_queue->efx;
 504         struct efx_channel *channel = efx_rx_queue_channel(rx_queue);
 505         struct efx_rx_buffer *rx_buf;
 506
 507         rx_buf = efx_rx_buffer(rx_queue, index);
 508         rx_buf->flags |= flags;
 509
 510         /* Validate the number of fragments and completed length */
 511         if (n_frags == 1) {
 512                 efx_rx_packet__check_len(rx_queue, rx_buf, len);
 513         } else if (unlikely(n_frags > EFX_RX_MAX_FRAGS) ||
 514                    unlikely(len <= (n_frags - 1) * EFX_RX_USR_BUF_SIZE) ||
 515                    unlikely(len > n_frags * EFX_RX_USR_BUF_SIZE) ||
 516                    unlikely(!efx->rx_scatter)) {
 517                 /* If this isn't an explicit discard request, either
 518                  * the hardware or the driver is broken.
 519                  */
 520                 WARN_ON(!(len == 0 && rx_buf->flags & EFX_RX_PKT_DISCARD));
 521                 rx_buf->flags |= EFX_RX_PKT_DISCARD;
 522         }
 523
 524         netif_vdbg(efx, rx_status, efx->net_dev,
 525                    "RX queue %d received ids %x-%x len %d %s%s\n",
 526                    efx_rx_queue_index(rx_queue), index,
 527                    (index + n_frags - 1) & rx_queue->ptr_mask, len,
 528                    (rx_buf->flags & EFX_RX_PKT_CSUMMED) ? " [SUMMED]" : "",
 529                    (rx_buf->flags & EFX_RX_PKT_DISCARD) ? " [DISCARD]" : "");
 530
 531         /* Discard packet, if instructed to do so.  Process the
 532          * previous receive first.
 533          */
 534         if (unlikely(rx_buf->flags & EFX_RX_PKT_DISCARD)) {
 535                 efx_rx_flush_packet(channel);
 536                 put_page(rx_buf->page);
 537                 efx_recycle_rx_buffers(channel, rx_buf, n_frags);
 538                 return;
 539         }
 540
 541         if (n_frags == 1)
 542                 rx_buf->len = len;
 543
 544         /* Release and/or sync the DMA mapping - assumes all RX buffers
 545          * consumed in-order per RX queue.
 546          */
 547         efx_sync_rx_buffer(efx, rx_buf, rx_buf->len);
 548
 549         /* Prefetch nice and early so data will (hopefully) be in cache by
 550          * the time we look at it.
 551          */
 552         prefetch(efx_rx_buf_va(rx_buf));
 553
 554         rx_buf->page_offset += efx->type->rx_buffer_hash_size;
 555         rx_buf->len -= efx->type->rx_buffer_hash_size;
 556
 557         if (n_frags > 1) {
 558                 /* Release/sync DMA mapping for additional fragments.
 559                  * Fix length for last fragment.
 560                  */
 561                 unsigned int tail_frags = n_frags - 1;
 562
 563                 for (;;) {
 564                         rx_buf = efx_rx_buf_next(rx_queue, rx_buf);
 565                         if (--tail_frags == 0)
 566                                 break;
 567                         efx_sync_rx_buffer(efx, rx_buf, EFX_RX_USR_BUF_SIZE);
 568                 }
 569                 rx_buf->len = len - (n_frags - 1) * EFX_RX_USR_BUF_SIZE;
 570                 efx_sync_rx_buffer(efx, rx_buf, rx_buf->len);
 571         }
 572
 573         /* All fragments have been DMA-synced, so recycle buffers and pages. */
 574         rx_buf = efx_rx_buffer(rx_queue, index);
 575         efx_recycle_rx_buffers(channel, rx_buf, n_frags);
 576
 577         /* Pipeline receives so that we give time for packet headers to be
 578          * prefetched into cache.
 579          */
 580         efx_rx_flush_packet(channel);
 581         channel->rx_pkt_n_frags = n_frags;
 582         channel->rx_pkt_index = index;
 583 }
 584
 585 static void efx_rx_deliver(struct efx_channel *channel, u8 *eh,
 586                            struct efx_rx_buffer *rx_buf,
 587                            unsigned int n_frags)
 588 {
 589         struct sk_buff *skb;
 590         u16 hdr_len = min_t(u16, rx_buf->len, EFX_SKB_HEADERS);
 591
 592         skb = efx_rx_mk_skb(channel, rx_buf, n_frags, eh, hdr_len);
 593         if (unlikely(skb == NULL)) {
 594                 efx_free_rx_buffer(rx_buf);
 595                 return;
 596         }
 597         skb_record_rx_queue(skb, channel->rx_queue.core_index);
 598
 599         /* Set the SKB flags */
 600         skb_checksum_none_assert(skb);
 601
 602         if (channel->type->receive_skb)
 603                 if (channel->type->receive_skb(channel, skb))
 604                         return;
 605
 606         /* Pass the packet up */
 607         netif_receive_skb(skb);
 608 }
 609
 610 /* Handle a received packet.  Second half: Touches packet payload. */
 611 void __efx_rx_packet(struct efx_channel *channel)
 612 {
 613         struct efx_nic *efx = channel->efx;
 614         struct efx_rx_buffer *rx_buf =
 615                 efx_rx_buffer(&channel->rx_queue, channel->rx_pkt_index);
 616         u8 *eh = efx_rx_buf_va(rx_buf);
 617
 618         /* If we're in loopback test, then pass the packet directly to the
 619          * loopback layer, and free the rx_buf here
 620          */
 621         if (unlikely(efx->loopback_selftest)) {
 622                 efx_loopback_rx_packet(efx, eh, rx_buf->len);
 623                 efx_free_rx_buffer(rx_buf);
 624                 goto out;
 625         }
 626
 627         if (unlikely(!(efx->net_dev->features & NETIF_F_RXCSUM)))
 628                 rx_buf->flags &= ~EFX_RX_PKT_CSUMMED;
 629
 630         if (!channel->type->receive_skb)
 631                 efx_rx_packet_gro(channel, rx_buf, channel->rx_pkt_n_frags, eh);
 632         else
 633                 efx_rx_deliver(channel, eh, rx_buf, channel->rx_pkt_n_frags);
 634 out:
 635         channel->rx_pkt_n_frags = 0;
 636 }
 637
 638 int efx_probe_rx_queue(struct efx_rx_queue *rx_queue)
 639 {
 640         struct efx_nic *efx = rx_queue->efx;
 641         unsigned int entries;
 642         int rc;
 643
 644         /* Create the smallest power-of-two aligned ring */
 645         entries = max(roundup_pow_of_two(efx->rxq_entries), EFX_MIN_DMAQ_SIZE);
 646         EFX_BUG_ON_PARANOID(entries > EFX_MAX_DMAQ_SIZE);
 647         rx_queue->ptr_mask = entries - 1;
 648
 649         netif_dbg(efx, probe, efx->net_dev,
 650                   "creating RX queue %d size %#x mask %#x\n",
 651                   efx_rx_queue_index(rx_queue), efx->rxq_entries,
 652                   rx_queue->ptr_mask);
 653
 654         /* Allocate RX buffers */
 655         rx_queue->buffer = kcalloc(entries, sizeof(*rx_queue->buffer),
 656                                    GFP_KERNEL);
 657         if (!rx_queue->buffer)
 658                 return -ENOMEM;
 659
 660         rc = efx_nic_probe_rx(rx_queue);
 661         if (rc) {
 662                 kfree(rx_queue->buffer);
 663                 rx_queue->buffer = NULL;
 664         }
 665
 666         return rc;
 667 }
 668
 669 static void efx_init_rx_recycle_ring(struct efx_nic *efx,
 670                                      struct efx_rx_queue *rx_queue)
 671 {
 672         unsigned int bufs_in_recycle_ring, page_ring_size;
 673
 674         /* Set the RX recycle ring size */
 675 #ifdef CONFIG_PPC64
 676         bufs_in_recycle_ring = EFX_RECYCLE_RING_SIZE_IOMMU;
 677 #else
 678         if (efx->pci_dev->dev.iommu_group)
 679                 bufs_in_recycle_ring = EFX_RECYCLE_RING_SIZE_IOMMU;
 680         else
 681                 bufs_in_recycle_ring = EFX_RECYCLE_RING_SIZE_NOIOMMU;
 682 #endif /* CONFIG_PPC64 */
 683
 684         page_ring_size = roundup_pow_of_two(bufs_in_recycle_ring /
 685                                             efx->rx_bufs_per_page);
 686         rx_queue->page_ring = kcalloc(page_ring_size,
 687                                       sizeof(*rx_queue->page_ring), GFP_KERNEL);
 688         rx_queue->page_ptr_mask = page_ring_size - 1;
 689 }
 690
 691 void efx_init_rx_queue(struct efx_rx_queue *rx_queue)
 692 {
 693         struct efx_nic *efx = rx_queue->efx;
 694         unsigned int max_fill, trigger, max_trigger;
 695
 696         netif_dbg(rx_queue->efx, drv, rx_queue->efx->net_dev,
 697                   "initialising RX queue %d\n", efx_rx_queue_index(rx_queue));
 698
 699         /* Initialise ptr fields */
 700         rx_queue->added_count = 0;
 701         rx_queue->notified_count = 0;
 702         rx_queue->removed_count = 0;
 703         rx_queue->min_fill = -1U;
 704         efx_init_rx_recycle_ring(efx, rx_queue);
 705
 706         rx_queue->page_remove = 0;
 707         rx_queue->page_add = rx_queue->page_ptr_mask + 1;
 708         rx_queue->page_recycle_count = 0;
 709         rx_queue->page_recycle_failed = 0;
 710         rx_queue->page_recycle_full = 0;
 711
 712         /* Initialise limit fields */
 713         max_fill = efx->rxq_entries - EFX_RXD_HEAD_ROOM;
 714         max_trigger =
 715                 max_fill - efx->rx_pages_per_batch * efx->rx_bufs_per_page;
 716         if (rx_refill_threshold != 0) {
 717                 trigger = max_fill * min(rx_refill_threshold, 100U) / 100U;
 718                 if (trigger > max_trigger)
 719                         trigger = max_trigger;
 720         } else {
 721                 trigger = max_trigger;
 722         }
 723
 724         rx_queue->max_fill = max_fill;
 725         rx_queue->fast_fill_trigger = trigger;
 726
 727         /* Set up RX descriptor ring */
 728         rx_queue->enabled = true;
 729         efx_nic_init_rx(rx_queue);
 730 }
 731
 732 void efx_fini_rx_queue(struct efx_rx_queue *rx_queue)
 733 {
 734         int i;
 735         struct efx_nic *efx = rx_queue->efx;
 736         struct efx_rx_buffer *rx_buf;
 737
 738         netif_dbg(rx_queue->efx, drv, rx_queue->efx->net_dev,
 739                   "shutting down RX queue %d\n", efx_rx_queue_index(rx_queue));
 740
 741         /* A flush failure might have left rx_queue->enabled */
 742         rx_queue->enabled = false;
 743
 744         del_timer_sync(&rx_queue->slow_fill);
 745         efx_nic_fini_rx(rx_queue);
 746
 747         /* Release RX buffers from the current read ptr to the write ptr */
 748         if (rx_queue->buffer) {
 749                 for (i = rx_queue->removed_count; i < rx_queue->added_count;
 750                      i++) {
 751                         unsigned index = i & rx_queue->ptr_mask;
 752                         rx_buf = efx_rx_buffer(rx_queue, index);
 753                         efx_fini_rx_buffer(rx_queue, rx_buf);
 754                 }
 755         }
 756
 757         /* Unmap and release the pages in the recycle ring. Remove the ring. */
 758         for (i = 0; i <= rx_queue->page_ptr_mask; i++) {
 759                 struct page *page = rx_queue->page_ring[i];
 760                 struct efx_rx_page_state *state;
 761
 762                 if (page == NULL)
 763                         continue;
 764
 765                 state = page_address(page);
 766                 dma_unmap_page(&efx->pci_dev->dev, state->dma_addr,
 767                                PAGE_SIZE << efx->rx_buffer_order,
 768                                DMA_FROM_DEVICE);
 769                 put_page(page);
 770         }
 771         kfree(rx_queue->page_ring);
 772         rx_queue->page_ring = NULL;
 773 }
 774
 775 void efx_remove_rx_queue(struct efx_rx_queue *rx_queue)
 776 {
 777         netif_dbg(rx_queue->efx, drv, rx_queue->efx->net_dev,
 778                   "destroying RX queue %d\n", efx_rx_queue_index(rx_queue));
 779
 780         efx_nic_remove_rx(rx_queue);
 781
 782         kfree(rx_queue->buffer);
 783         rx_queue->buffer = NULL;
 784 }
 785
 786
 787 module_param(rx_refill_threshold, uint, 0444);
 788 MODULE_PARM_DESC(rx_refill_threshold,
 789                  "RX descriptor ring refill threshold (%)");
 790