drivers/net/sfc/rx.c

   1 /****************************************************************************
   2  * Driver for Solarflare Solarstorm network controllers and boards
   3  * Copyright 2005-2006 Fen Systems Ltd.
   4  * Copyright 2005-2009 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 <net/ip.h>
  18 #include <net/checksum.h>
  19 #include "net_driver.h"
  20 #include "efx.h"
  21 #include "nic.h"
  22 #include "selftest.h"
  23 #include "workarounds.h"
  24
  25 /* Number of RX descriptors pushed at once. */
  26 #define EFX_RX_BATCH  8
  27
  28 /* Maximum size of a buffer sharing a page */
  29 #define EFX_RX_HALF_PAGE ((PAGE_SIZE >> 1) - sizeof(struct efx_rx_page_state))
  30
  31 /* Size of buffer allocated for skb header area. */
  32 #define EFX_SKB_HEADERS  64u
  33
  34 /*
  35  * rx_alloc_method - RX buffer allocation method
  36  *
  37  * This driver supports two methods for allocating and using RX buffers:
  38  * each RX buffer may be backed by an skb or by an order-n page.
  39  *
  40  * When LRO is in use then the second method has a lower overhead,
  41  * since we don't have to allocate then free skbs on reassembled frames.
  42  *
  43  * Values:
  44  *   - RX_ALLOC_METHOD_AUTO = 0
  45  *   - RX_ALLOC_METHOD_SKB  = 1
  46  *   - RX_ALLOC_METHOD_PAGE = 2
  47  *
  48  * The heuristic for %RX_ALLOC_METHOD_AUTO is a simple hysteresis count
  49  * controlled by the parameters below.
  50  *
  51  *   - Since pushing and popping descriptors are separated by the rx_queue
  52  *     size, so the watermarks should be ~rxd_size.
  53  *   - The performance win by using page-based allocation for LRO is less
  54  *     than the performance hit of using page-based allocation of non-LRO,
  55  *     so the watermarks should reflect this.
  56  *
  57  * Per channel we maintain a single variable, updated by each channel:
  58  *
  59  *   rx_alloc_level += (lro_performed ? RX_ALLOC_FACTOR_LRO :
  60  *                      RX_ALLOC_FACTOR_SKB)
  61  * Per NAPI poll interval, we constrain rx_alloc_level to 0..MAX (which
  62  * limits the hysteresis), and update the allocation strategy:
  63  *
  64  *   rx_alloc_method = (rx_alloc_level > RX_ALLOC_LEVEL_LRO ?
  65  *                      RX_ALLOC_METHOD_PAGE : RX_ALLOC_METHOD_SKB)
  66  */
  67 static int rx_alloc_method = RX_ALLOC_METHOD_AUTO;
  68
  69 #define RX_ALLOC_LEVEL_LRO 0x2000
  70 #define RX_ALLOC_LEVEL_MAX 0x3000
  71 #define RX_ALLOC_FACTOR_LRO 1
  72 #define RX_ALLOC_FACTOR_SKB (-2)
  73
  74 /* This is the percentage fill level below which new RX descriptors
  75  * will be added to the RX descriptor ring.
  76  */
  77 static unsigned int rx_refill_threshold = 90;
  78
  79 /* This is the percentage fill level to which an RX queue will be refilled
  80  * when the "RX refill threshold" is reached.
  81  */
  82 static unsigned int rx_refill_limit = 95;
  83
  84 /*
  85  * RX maximum head room required.
  86  *
  87  * This must be at least 1 to prevent overflow and at least 2 to allow
  88  * pipelined receives.
  89  */
  90 #define EFX_RXD_HEAD_ROOM 2
  91
  92 static inline unsigned int efx_rx_buf_offset(struct efx_rx_buffer *buf)
  93 {
  94         /* Offset is always within one page, so we don't need to consider
  95          * the page order.
  96          */
  97         return (__force unsigned long) buf->data & (PAGE_SIZE - 1);
  98 }
  99 static inline unsigned int efx_rx_buf_size(struct efx_nic *efx)
 100 {
 101         return PAGE_SIZE << efx->rx_buffer_order;
 102 }
 103
 104 static inline u32 efx_rx_buf_hash(struct efx_rx_buffer *buf)
 105 {
 106 #if defined(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS) || NET_IP_ALIGN % 4 == 0
 107         return __le32_to_cpup((const __le32 *)(buf->data - 4));
 108 #else
 109         const u8 *data = (const u8 *)(buf->data - 4);
 110         return ((u32)data[0]       |
 111                 (u32)data[1] << 8  |
 112                 (u32)data[2] << 16 |
 113                 (u32)data[3] << 24);
 114 #endif
 115 }
 116
 117 /**
 118  * efx_init_rx_buffers_skb - create EFX_RX_BATCH skb-based RX buffers
 119  *
 120  * @rx_queue:           Efx RX queue
 121  *
 122  * This allocates EFX_RX_BATCH skbs, maps them for DMA, and populates a
 123  * struct efx_rx_buffer for each one. Return a negative error code or 0
 124  * on success. May fail having only inserted fewer than EFX_RX_BATCH
 125  * buffers.
 126  */
 127 static int efx_init_rx_buffers_skb(struct efx_rx_queue *rx_queue)
 128 {
 129         struct efx_nic *efx = rx_queue->efx;
 130         struct net_device *net_dev = efx->net_dev;
 131         struct efx_rx_buffer *rx_buf;
 132         int skb_len = efx->rx_buffer_len;
 133         unsigned index, count;
 134
 135         for (count = 0; count < EFX_RX_BATCH; ++count) {
 136                 index = rx_queue->added_count & EFX_RXQ_MASK;
 137                 rx_buf = efx_rx_buffer(rx_queue, index);
 138
 139                 rx_buf->skb = netdev_alloc_skb(net_dev, skb_len);
 140                 if (unlikely(!rx_buf->skb))
 141                         return -ENOMEM;
 142                 rx_buf->page = NULL;
 143
 144                 /* Adjust the SKB for padding and checksum */
 145                 skb_reserve(rx_buf->skb, NET_IP_ALIGN);
 146                 rx_buf->len = skb_len - NET_IP_ALIGN;
 147                 rx_buf->data = (char *)rx_buf->skb->data;
 148                 rx_buf->skb->ip_summed = CHECKSUM_UNNECESSARY;
 149
 150                 rx_buf->dma_addr = pci_map_single(efx->pci_dev,
 151                                                   rx_buf->data, rx_buf->len,
 152                                                   PCI_DMA_FROMDEVICE);
 153                 if (unlikely(pci_dma_mapping_error(efx->pci_dev,
 154                                                    rx_buf->dma_addr))) {
 155                         dev_kfree_skb_any(rx_buf->skb);
 156                         rx_buf->skb = NULL;
 157                         return -EIO;
 158                 }
 159
 160                 ++rx_queue->added_count;
 161                 ++rx_queue->alloc_skb_count;
 162         }
 163
 164         return 0;
 165 }
 166
 167 /**
 168  * efx_init_rx_buffers_page - create EFX_RX_BATCH page-based RX buffers
 169  *
 170  * @rx_queue:           Efx RX queue
 171  *
 172  * This allocates memory for EFX_RX_BATCH receive buffers, maps them for DMA,
 173  * and populates struct efx_rx_buffers for each one. Return a negative error
 174  * code or 0 on success. If a single page can be split between two buffers,
 175  * then the page will either be inserted fully, or not at at all.
 176  */
 177 static int efx_init_rx_buffers_page(struct efx_rx_queue *rx_queue)
 178 {
 179         struct efx_nic *efx = rx_queue->efx;
 180         struct efx_rx_buffer *rx_buf;
 181         struct page *page;
 182         void *page_addr;
 183         struct efx_rx_page_state *state;
 184         dma_addr_t dma_addr;
 185         unsigned index, count;
 186
 187         /* We can split a page between two buffers */
 188         BUILD_BUG_ON(EFX_RX_BATCH & 1);
 189
 190         for (count = 0; count < EFX_RX_BATCH; ++count) {
 191                 page = alloc_pages(__GFP_COLD | __GFP_COMP | GFP_ATOMIC,
 192                                    efx->rx_buffer_order);
 193                 if (unlikely(page == NULL))
 194                         return -ENOMEM;
 195                 dma_addr = pci_map_page(efx->pci_dev, page, 0,
 196                                         efx_rx_buf_size(efx),
 197                                         PCI_DMA_FROMDEVICE);
 198                 if (unlikely(pci_dma_mapping_error(efx->pci_dev, dma_addr))) {
 199                         __free_pages(page, efx->rx_buffer_order);
 200                         return -EIO;
 201                 }
 202                 page_addr = page_address(page);
 203                 state = page_addr;
 204                 state->refcnt = 0;
 205                 state->dma_addr = dma_addr;
 206
 207                 page_addr += sizeof(struct efx_rx_page_state);
 208                 dma_addr += sizeof(struct efx_rx_page_state);
 209
 210         split:
 211                 index = rx_queue->added_count & EFX_RXQ_MASK;
 212                 rx_buf = efx_rx_buffer(rx_queue, index);
 213                 rx_buf->dma_addr = dma_addr + EFX_PAGE_IP_ALIGN;
 214                 rx_buf->skb = NULL;
 215                 rx_buf->page = page;
 216                 rx_buf->data = page_addr + EFX_PAGE_IP_ALIGN;
 217                 rx_buf->len = efx->rx_buffer_len - EFX_PAGE_IP_ALIGN;
 218                 ++rx_queue->added_count;
 219                 ++rx_queue->alloc_page_count;
 220                 ++state->refcnt;
 221
 222                 if ((~count & 1) && (efx->rx_buffer_len <= EFX_RX_HALF_PAGE)) {
 223                         /* Use the second half of the page */
 224                         get_page(page);
 225                         dma_addr += (PAGE_SIZE >> 1);
 226                         page_addr += (PAGE_SIZE >> 1);
 227                         ++count;
 228                         goto split;
 229                 }
 230         }
 231
 232         return 0;
 233 }
 234
 235 static void efx_unmap_rx_buffer(struct efx_nic *efx,
 236                                 struct efx_rx_buffer *rx_buf)
 237 {
 238         if (rx_buf->page) {
 239                 struct efx_rx_page_state *state;
 240
 241                 EFX_BUG_ON_PARANOID(rx_buf->skb);
 242
 243                 state = page_address(rx_buf->page);
 244                 if (--state->refcnt == 0) {
 245                         pci_unmap_page(efx->pci_dev,
 246                                        state->dma_addr,
 247                                        efx_rx_buf_size(efx),
 248                                        PCI_DMA_FROMDEVICE);
 249                 }
 250         } else if (likely(rx_buf->skb)) {
 251                 pci_unmap_single(efx->pci_dev, rx_buf->dma_addr,
 252                                  rx_buf->len, PCI_DMA_FROMDEVICE);
 253         }
 254 }
 255
 256 static void efx_free_rx_buffer(struct efx_nic *efx,
 257                                struct efx_rx_buffer *rx_buf)
 258 {
 259         if (rx_buf->page) {
 260                 __free_pages(rx_buf->page, efx->rx_buffer_order);
 261                 rx_buf->page = NULL;
 262         } else if (likely(rx_buf->skb)) {
 263                 dev_kfree_skb_any(rx_buf->skb);
 264                 rx_buf->skb = NULL;
 265         }
 266 }
 267
 268 static void efx_fini_rx_buffer(struct efx_rx_queue *rx_queue,
 269                                struct efx_rx_buffer *rx_buf)
 270 {
 271         efx_unmap_rx_buffer(rx_queue->efx, rx_buf);
 272         efx_free_rx_buffer(rx_queue->efx, rx_buf);
 273 }
 274
 275 /* Attempt to resurrect the other receive buffer that used to share this page,
 276  * which had previously been passed up to the kernel and freed. */
 277 static void efx_resurrect_rx_buffer(struct efx_rx_queue *rx_queue,
 278                                     struct efx_rx_buffer *rx_buf)
 279 {
 280         struct efx_rx_page_state *state = page_address(rx_buf->page);
 281         struct efx_rx_buffer *new_buf;
 282         unsigned fill_level, index;
 283
 284         /* +1 because efx_rx_packet() incremented removed_count. +1 because
 285          * we'd like to insert an additional descriptor whilst leaving
 286          * EFX_RXD_HEAD_ROOM for the non-recycle path */
 287         fill_level = (rx_queue->added_count - rx_queue->removed_count + 2);
 288         if (unlikely(fill_level >= EFX_RXQ_SIZE - EFX_RXD_HEAD_ROOM)) {
 289                 /* We could place "state" on a list, and drain the list in
 290                  * efx_fast_push_rx_descriptors(). For now, this will do. */
 291                 return;
 292         }
 293
 294         ++state->refcnt;
 295         get_page(rx_buf->page);
 296
 297         index = rx_queue->added_count & EFX_RXQ_MASK;
 298         new_buf = efx_rx_buffer(rx_queue, index);
 299         new_buf->dma_addr = rx_buf->dma_addr ^ (PAGE_SIZE >> 1);
 300         new_buf->skb = NULL;
 301         new_buf->page = rx_buf->page;
 302         new_buf->data = (void *)
 303                 ((__force unsigned long)rx_buf->data ^ (PAGE_SIZE >> 1));
 304         new_buf->len = rx_buf->len;
 305         ++rx_queue->added_count;
 306 }
 307
 308 /* Recycle the given rx buffer directly back into the rx_queue. There is
 309  * always room to add this buffer, because we've just popped a buffer. */
 310 static void efx_recycle_rx_buffer(struct efx_channel *channel,
 311                                   struct efx_rx_buffer *rx_buf)
 312 {
 313         struct efx_nic *efx = channel->efx;
 314         struct efx_rx_queue *rx_queue = &efx->rx_queue[channel->channel];
 315         struct efx_rx_buffer *new_buf;
 316         unsigned index;
 317
 318         if (rx_buf->page != NULL && efx->rx_buffer_len <= EFX_RX_HALF_PAGE &&
 319             page_count(rx_buf->page) == 1)
 320                 efx_resurrect_rx_buffer(rx_queue, rx_buf);
 321
 322         index = rx_queue->added_count & EFX_RXQ_MASK;
 323         new_buf = efx_rx_buffer(rx_queue, index);
 324
 325         memcpy(new_buf, rx_buf, sizeof(*new_buf));
 326         rx_buf->page = NULL;
 327         rx_buf->skb = NULL;
 328         ++rx_queue->added_count;
 329 }
 330
 331 /**
 332  * efx_fast_push_rx_descriptors - push new RX descriptors quickly
 333  * @rx_queue:           RX descriptor queue
 334  * This will aim to fill the RX descriptor queue up to
 335  * @rx_queue->@fast_fill_limit. If there is insufficient atomic
 336  * memory to do so, a slow fill will be scheduled.
 337  *
 338  * The caller must provide serialisation (none is used here). In practise,
 339  * this means this function must run from the NAPI handler, or be called
 340  * when NAPI is disabled.
 341  */
 342 void efx_fast_push_rx_descriptors(struct efx_rx_queue *rx_queue)
 343 {
 344         struct efx_channel *channel = rx_queue->channel;
 345         unsigned fill_level;
 346         int space, rc = 0;
 347
 348         /* Calculate current fill level, and exit if we don't need to fill */
 349         fill_level = (rx_queue->added_count - rx_queue->removed_count);
 350         EFX_BUG_ON_PARANOID(fill_level > EFX_RXQ_SIZE);
 351         if (fill_level >= rx_queue->fast_fill_trigger)
 352                 goto out;
 353
 354         /* Record minimum fill level */
 355         if (unlikely(fill_level < rx_queue->min_fill)) {
 356                 if (fill_level)
 357                         rx_queue->min_fill = fill_level;
 358         }
 359
 360         space = rx_queue->fast_fill_limit - fill_level;
 361         if (space < EFX_RX_BATCH)
 362                 goto out;
 363
 364         netif_vdbg(rx_queue->efx, rx_status, rx_queue->efx->net_dev,
 365                    "RX queue %d fast-filling descriptor ring from"
 366                    " level %d to level %d using %s allocation\n",
 367                    rx_queue->queue, fill_level, rx_queue->fast_fill_limit,
 368                    channel->rx_alloc_push_pages ? "page" : "skb");
 369
 370         do {
 371                 if (channel->rx_alloc_push_pages)
 372                         rc = efx_init_rx_buffers_page(rx_queue);
 373                 else
 374                         rc = efx_init_rx_buffers_skb(rx_queue);
 375                 if (unlikely(rc)) {
 376                         /* Ensure that we don't leave the rx queue empty */
 377                         if (rx_queue->added_count == rx_queue->removed_count)
 378                                 efx_schedule_slow_fill(rx_queue);
 379                         goto out;
 380                 }
 381         } while ((space -= EFX_RX_BATCH) >= EFX_RX_BATCH);
 382
 383         netif_vdbg(rx_queue->efx, rx_status, rx_queue->efx->net_dev,
 384                    "RX queue %d fast-filled descriptor ring "
 385                    "to level %d\n", rx_queue->queue,
 386                    rx_queue->added_count - rx_queue->removed_count);
 387
 388  out:
 389         if (rx_queue->notified_count != rx_queue->added_count)
 390                 efx_nic_notify_rx_desc(rx_queue);
 391 }
 392
 393 void efx_rx_slow_fill(unsigned long context)
 394 {
 395         struct efx_rx_queue *rx_queue = (struct efx_rx_queue *)context;
 396         struct efx_channel *channel = rx_queue->channel;
 397
 398         /* Post an event to cause NAPI to run and refill the queue */
 399         efx_nic_generate_fill_event(channel);
 400         ++rx_queue->slow_fill_count;
 401 }
 402
 403 static void efx_rx_packet__check_len(struct efx_rx_queue *rx_queue,
 404                                      struct efx_rx_buffer *rx_buf,
 405                                      int len, bool *discard,
 406                                      bool *leak_packet)
 407 {
 408         struct efx_nic *efx = rx_queue->efx;
 409         unsigned max_len = rx_buf->len - efx->type->rx_buffer_padding;
 410
 411         if (likely(len <= max_len))
 412                 return;
 413
 414         /* The packet must be discarded, but this is only a fatal error
 415          * if the caller indicated it was
 416          */
 417         *discard = true;
 418
 419         if ((len > rx_buf->len) && EFX_WORKAROUND_8071(efx)) {
 420                 if (net_ratelimit())
 421                         netif_err(efx, rx_err, efx->net_dev,
 422                                   " RX queue %d seriously overlength "
 423                                   "RX event (0x%x > 0x%x+0x%x). Leaking\n",
 424                                   rx_queue->queue, len, max_len,
 425                                   efx->type->rx_buffer_padding);
 426                 /* If this buffer was skb-allocated, then the meta
 427                  * data at the end of the skb will be trashed. So
 428                  * we have no choice but to leak the fragment.
 429                  */
 430                 *leak_packet = (rx_buf->skb != NULL);
 431                 efx_schedule_reset(efx, RESET_TYPE_RX_RECOVERY);
 432         } else {
 433                 if (net_ratelimit())
 434                         netif_err(efx, rx_err, efx->net_dev,
 435                                   " RX queue %d overlength RX event "
 436                                   "(0x%x > 0x%x)\n",
 437                                   rx_queue->queue, len, max_len);
 438         }
 439
 440         rx_queue->channel->n_rx_overlength++;
 441 }
 442
 443 /* Pass a received packet up through the generic LRO stack
 444  *
 445  * Handles driverlink veto, and passes the fragment up via
 446  * the appropriate LRO method
 447  */
 448 static void efx_rx_packet_lro(struct efx_channel *channel,
 449                               struct efx_rx_buffer *rx_buf,
 450                               bool checksummed)
 451 {
 452         struct napi_struct *napi = &channel->napi_str;
 453         gro_result_t gro_result;
 454
 455         /* Pass the skb/page into the LRO engine */
 456         if (rx_buf->page) {
 457                 struct efx_nic *efx = channel->efx;
 458                 struct page *page = rx_buf->page;
 459                 struct sk_buff *skb;
 460
 461                 EFX_BUG_ON_PARANOID(rx_buf->skb);
 462                 rx_buf->page = NULL;
 463
 464                 skb = napi_get_frags(napi);
 465                 if (!skb) {
 466                         put_page(page);
 467                         return;
 468                 }
 469
 470                 if (efx->net_dev->features & NETIF_F_RXHASH)
 471                         skb->rxhash = efx_rx_buf_hash(rx_buf);
 472
 473                 skb_shinfo(skb)->frags[0].page = page;
 474                 skb_shinfo(skb)->frags[0].page_offset =
 475                         efx_rx_buf_offset(rx_buf);
 476                 skb_shinfo(skb)->frags[0].size = rx_buf->len;
 477                 skb_shinfo(skb)->nr_frags = 1;
 478
 479                 skb->len = rx_buf->len;
 480                 skb->data_len = rx_buf->len;
 481                 skb->truesize += rx_buf->len;
 482                 skb->ip_summed =
 483                         checksummed ? CHECKSUM_UNNECESSARY : CHECKSUM_NONE;
 484
 485                 skb_record_rx_queue(skb, channel->channel);
 486
 487                 gro_result = napi_gro_frags(napi);
 488         } else {
 489                 struct sk_buff *skb = rx_buf->skb;
 490
 491                 EFX_BUG_ON_PARANOID(!skb);
 492                 EFX_BUG_ON_PARANOID(!checksummed);
 493                 rx_buf->skb = NULL;
 494
 495                 gro_result = napi_gro_receive(napi, skb);
 496         }
 497
 498         if (gro_result == GRO_NORMAL) {
 499                 channel->rx_alloc_level += RX_ALLOC_FACTOR_SKB;
 500         } else if (gro_result != GRO_DROP) {
 501                 channel->rx_alloc_level += RX_ALLOC_FACTOR_LRO;
 502                 channel->irq_mod_score += 2;
 503         }
 504 }
 505
 506 void efx_rx_packet(struct efx_rx_queue *rx_queue, unsigned int index,
 507                    unsigned int len, bool checksummed, bool discard)
 508 {
 509         struct efx_nic *efx = rx_queue->efx;
 510         struct efx_channel *channel = rx_queue->channel;
 511         struct efx_rx_buffer *rx_buf;
 512         bool leak_packet = false;
 513
 514         rx_buf = efx_rx_buffer(rx_queue, index);
 515         EFX_BUG_ON_PARANOID(!rx_buf->data);
 516         EFX_BUG_ON_PARANOID(rx_buf->skb && rx_buf->page);
 517         EFX_BUG_ON_PARANOID(!(rx_buf->skb || rx_buf->page));
 518
 519         /* This allows the refill path to post another buffer.
 520          * EFX_RXD_HEAD_ROOM ensures that the slot we are using
 521          * isn't overwritten yet.
 522          */
 523         rx_queue->removed_count++;
 524
 525         /* Validate the length encoded in the event vs the descriptor pushed */
 526         efx_rx_packet__check_len(rx_queue, rx_buf, len,
 527                                  &discard, &leak_packet);
 528
 529         netif_vdbg(efx, rx_status, efx->net_dev,
 530                    "RX queue %d received id %x at %llx+%x %s%s\n",
 531                    rx_queue->queue, index,
 532                    (unsigned long long)rx_buf->dma_addr, len,
 533                    (checksummed ? " [SUMMED]" : ""),
 534                    (discard ? " [DISCARD]" : ""));
 535
 536         /* Discard packet, if instructed to do so */
 537         if (unlikely(discard)) {
 538                 if (unlikely(leak_packet))
 539                         channel->n_skbuff_leaks++;
 540                 else
 541                         efx_recycle_rx_buffer(channel, rx_buf);
 542
 543                 /* Don't hold off the previous receive */
 544                 rx_buf = NULL;
 545                 goto out;
 546         }
 547
 548         /* Release card resources - assumes all RX buffers consumed in-order
 549          * per RX queue
 550          */
 551         efx_unmap_rx_buffer(efx, rx_buf);
 552
 553         /* Prefetch nice and early so data will (hopefully) be in cache by
 554          * the time we look at it.
 555          */
 556         prefetch(rx_buf->data);
 557
 558         /* Pipeline receives so that we give time for packet headers to be
 559          * prefetched into cache.
 560          */
 561         rx_buf->len = len;
 562 out:
 563         if (rx_queue->channel->rx_pkt)
 564                 __efx_rx_packet(rx_queue->channel,
 565                                 rx_queue->channel->rx_pkt,
 566                                 rx_queue->channel->rx_pkt_csummed);
 567         rx_queue->channel->rx_pkt = rx_buf;
 568         rx_queue->channel->rx_pkt_csummed = checksummed;
 569 }
 570
 571 /* Handle a received packet.  Second half: Touches packet payload. */
 572 void __efx_rx_packet(struct efx_channel *channel,
 573                      struct efx_rx_buffer *rx_buf, bool checksummed)
 574 {
 575         struct efx_nic *efx = channel->efx;
 576         struct sk_buff *skb;
 577
 578         rx_buf->data += efx->type->rx_buffer_hash_size;
 579         rx_buf->len -= efx->type->rx_buffer_hash_size;
 580
 581         /* If we're in loopback test, then pass the packet directly to the
 582          * loopback layer, and free the rx_buf here
 583          */
 584         if (unlikely(efx->loopback_selftest)) {
 585                 efx_loopback_rx_packet(efx, rx_buf->data, rx_buf->len);
 586                 efx_free_rx_buffer(efx, rx_buf);
 587                 return;
 588         }
 589
 590         if (rx_buf->skb) {
 591                 prefetch(skb_shinfo(rx_buf->skb));
 592
 593                 skb_reserve(rx_buf->skb, efx->type->rx_buffer_hash_size);
 594                 skb_put(rx_buf->skb, rx_buf->len);
 595
 596                 if (efx->net_dev->features & NETIF_F_RXHASH)
 597                         rx_buf->skb->rxhash = efx_rx_buf_hash(rx_buf);
 598
 599                 /* Move past the ethernet header. rx_buf->data still points
 600                  * at the ethernet header */
 601                 rx_buf->skb->protocol = eth_type_trans(rx_buf->skb,
 602                                                        efx->net_dev);
 603
 604                 skb_record_rx_queue(rx_buf->skb, channel->channel);
 605         }
 606
 607         if (likely(checksummed || rx_buf->page)) {
 608                 efx_rx_packet_lro(channel, rx_buf, checksummed);
 609                 return;
 610         }
 611
 612         /* We now own the SKB */
 613         skb = rx_buf->skb;
 614         rx_buf->skb = NULL;
 615         EFX_BUG_ON_PARANOID(!skb);
 616
 617         /* Set the SKB flags */
 618         skb->ip_summed = CHECKSUM_NONE;
 619
 620         /* Pass the packet up */
 621         netif_receive_skb(skb);
 622
 623         /* Update allocation strategy method */
 624         channel->rx_alloc_level += RX_ALLOC_FACTOR_SKB;
 625 }
 626
 627 void efx_rx_strategy(struct efx_channel *channel)
 628 {
 629         enum efx_rx_alloc_method method = rx_alloc_method;
 630
 631         /* Only makes sense to use page based allocation if LRO is enabled */
 632         if (!(channel->efx->net_dev->features & NETIF_F_GRO)) {
 633                 method = RX_ALLOC_METHOD_SKB;
 634         } else if (method == RX_ALLOC_METHOD_AUTO) {
 635                 /* Constrain the rx_alloc_level */
 636                 if (channel->rx_alloc_level < 0)
 637                         channel->rx_alloc_level = 0;
 638                 else if (channel->rx_alloc_level > RX_ALLOC_LEVEL_MAX)
 639                         channel->rx_alloc_level = RX_ALLOC_LEVEL_MAX;
 640
 641                 /* Decide on the allocation method */
 642                 method = ((channel->rx_alloc_level > RX_ALLOC_LEVEL_LRO) ?
 643                           RX_ALLOC_METHOD_PAGE : RX_ALLOC_METHOD_SKB);
 644         }
 645
 646         /* Push the option */
 647         channel->rx_alloc_push_pages = (method == RX_ALLOC_METHOD_PAGE);
 648 }
 649
 650 int efx_probe_rx_queue(struct efx_rx_queue *rx_queue)
 651 {
 652         struct efx_nic *efx = rx_queue->efx;
 653         unsigned int rxq_size;
 654         int rc;
 655
 656         netif_dbg(efx, probe, efx->net_dev,
 657                   "creating RX queue %d\n", rx_queue->queue);
 658
 659         /* Allocate RX buffers */
 660         rxq_size = EFX_RXQ_SIZE * sizeof(*rx_queue->buffer);
 661         rx_queue->buffer = kzalloc(rxq_size, GFP_KERNEL);
 662         if (!rx_queue->buffer)
 663                 return -ENOMEM;
 664
 665         rc = efx_nic_probe_rx(rx_queue);
 666         if (rc) {
 667                 kfree(rx_queue->buffer);
 668                 rx_queue->buffer = NULL;
 669         }
 670         return rc;
 671 }
 672
 673 void efx_init_rx_queue(struct efx_rx_queue *rx_queue)
 674 {
 675         unsigned int max_fill, trigger, limit;
 676
 677         netif_dbg(rx_queue->efx, drv, rx_queue->efx->net_dev,
 678                   "initialising RX queue %d\n", rx_queue->queue);
 679
 680         /* Initialise ptr fields */
 681         rx_queue->added_count = 0;
 682         rx_queue->notified_count = 0;
 683         rx_queue->removed_count = 0;
 684         rx_queue->min_fill = -1U;
 685         rx_queue->min_overfill = -1U;
 686
 687         /* Initialise limit fields */
 688         max_fill = EFX_RXQ_SIZE - EFX_RXD_HEAD_ROOM;
 689         trigger = max_fill * min(rx_refill_threshold, 100U) / 100U;
 690         limit = max_fill * min(rx_refill_limit, 100U) / 100U;
 691
 692         rx_queue->max_fill = max_fill;
 693         rx_queue->fast_fill_trigger = trigger;
 694         rx_queue->fast_fill_limit = limit;
 695
 696         /* Set up RX descriptor ring */
 697         efx_nic_init_rx(rx_queue);
 698 }
 699
 700 void efx_fini_rx_queue(struct efx_rx_queue *rx_queue)
 701 {
 702         int i;
 703         struct efx_rx_buffer *rx_buf;
 704
 705         netif_dbg(rx_queue->efx, drv, rx_queue->efx->net_dev,
 706                   "shutting down RX queue %d\n", rx_queue->queue);
 707
 708         del_timer_sync(&rx_queue->slow_fill);
 709         efx_nic_fini_rx(rx_queue);
 710
 711         /* Release RX buffers NB start at index 0 not current HW ptr */
 712         if (rx_queue->buffer) {
 713                 for (i = 0; i <= EFX_RXQ_MASK; i++) {
 714                         rx_buf = efx_rx_buffer(rx_queue, i);
 715                         efx_fini_rx_buffer(rx_queue, rx_buf);
 716                 }
 717         }
 718 }
 719
 720 void efx_remove_rx_queue(struct efx_rx_queue *rx_queue)
 721 {
 722         netif_dbg(rx_queue->efx, drv, rx_queue->efx->net_dev,
 723                   "destroying RX queue %d\n", rx_queue->queue);
 724
 725         efx_nic_remove_rx(rx_queue);
 726
 727         kfree(rx_queue->buffer);
 728         rx_queue->buffer = NULL;
 729 }
 730
 731
 732 module_param(rx_alloc_method, int, 0644);
 733 MODULE_PARM_DESC(rx_alloc_method, "Allocation method used for RX buffers");
 734
 735 module_param(rx_refill_threshold, uint, 0444);
 736 MODULE_PARM_DESC(rx_refill_threshold,
 737                  "RX descriptor ring fast/slow fill threshold (%)");
 738