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/ip.h>
  14 #include <linux/tcp.h>
  15 #include <linux/udp.h>
  16 #include <net/ip.h>
  17 #include <net/checksum.h>
  18 #include "net_driver.h"
  19 #include "efx.h"
  20 #include "nic.h"
  21 #include "selftest.h"
  22 #include "workarounds.h"
  23
  24 /* Number of RX descriptors pushed at once. */
  25 #define EFX_RX_BATCH  8
  26
  27 /* Size of buffer allocated for skb header area. */
  28 #define EFX_SKB_HEADERS  64u
  29
  30 /*
  31  * rx_alloc_method - RX buffer allocation method
  32  *
  33  * This driver supports two methods for allocating and using RX buffers:
  34  * each RX buffer may be backed by an skb or by an order-n page.
  35  *
  36  * When LRO is in use then the second method has a lower overhead,
  37  * since we don't have to allocate then free skbs on reassembled frames.
  38  *
  39  * Values:
  40  *   - RX_ALLOC_METHOD_AUTO = 0
  41  *   - RX_ALLOC_METHOD_SKB  = 1
  42  *   - RX_ALLOC_METHOD_PAGE = 2
  43  *
  44  * The heuristic for %RX_ALLOC_METHOD_AUTO is a simple hysteresis count
  45  * controlled by the parameters below.
  46  *
  47  *   - Since pushing and popping descriptors are separated by the rx_queue
  48  *     size, so the watermarks should be ~rxd_size.
  49  *   - The performance win by using page-based allocation for LRO is less
  50  *     than the performance hit of using page-based allocation of non-LRO,
  51  *     so the watermarks should reflect this.
  52  *
  53  * Per channel we maintain a single variable, updated by each channel:
  54  *
  55  *   rx_alloc_level += (lro_performed ? RX_ALLOC_FACTOR_LRO :
  56  *                      RX_ALLOC_FACTOR_SKB)
  57  * Per NAPI poll interval, we constrain rx_alloc_level to 0..MAX (which
  58  * limits the hysteresis), and update the allocation strategy:
  59  *
  60  *   rx_alloc_method = (rx_alloc_level > RX_ALLOC_LEVEL_LRO ?
  61  *                      RX_ALLOC_METHOD_PAGE : RX_ALLOC_METHOD_SKB)
  62  */
  63 static int rx_alloc_method = RX_ALLOC_METHOD_AUTO;
  64
  65 #define RX_ALLOC_LEVEL_LRO 0x2000
  66 #define RX_ALLOC_LEVEL_MAX 0x3000
  67 #define RX_ALLOC_FACTOR_LRO 1
  68 #define RX_ALLOC_FACTOR_SKB (-2)
  69
  70 /* This is the percentage fill level below which new RX descriptors
  71  * will be added to the RX descriptor ring.
  72  */
  73 static unsigned int rx_refill_threshold = 90;
  74
  75 /* This is the percentage fill level to which an RX queue will be refilled
  76  * when the "RX refill threshold" is reached.
  77  */
  78 static unsigned int rx_refill_limit = 95;
  79
  80 /*
  81  * RX maximum head room required.
  82  *
  83  * This must be at least 1 to prevent overflow and at least 2 to allow
  84  * pipelined receives.
  85  */
  86 #define EFX_RXD_HEAD_ROOM 2
  87
  88 static inline unsigned int efx_rx_buf_offset(struct efx_rx_buffer *buf)
  89 {
  90         /* Offset is always within one page, so we don't need to consider
  91          * the page order.
  92          */
  93         return (__force unsigned long) buf->data & (PAGE_SIZE - 1);
  94 }
  95 static inline unsigned int efx_rx_buf_size(struct efx_nic *efx)
  96 {
  97         return PAGE_SIZE << efx->rx_buffer_order;
  98 }
  99
 100
 101 /**
 102  * efx_init_rx_buffer_skb - create new RX buffer using skb-based allocation
 103  *
 104  * @rx_queue:           Efx RX queue
 105  * @rx_buf:             RX buffer structure to populate
 106  *
 107  * This allocates memory for a new receive buffer, maps it for DMA,
 108  * and populates a struct efx_rx_buffer with the relevant
 109  * information.  Return a negative error code or 0 on success.
 110  */
 111 static int efx_init_rx_buffer_skb(struct efx_rx_queue *rx_queue,
 112                                   struct efx_rx_buffer *rx_buf)
 113 {
 114         struct efx_nic *efx = rx_queue->efx;
 115         struct net_device *net_dev = efx->net_dev;
 116         int skb_len = efx->rx_buffer_len;
 117
 118         rx_buf->skb = netdev_alloc_skb(net_dev, skb_len);
 119         if (unlikely(!rx_buf->skb))
 120                 return -ENOMEM;
 121
 122         /* Adjust the SKB for padding and checksum */
 123         skb_reserve(rx_buf->skb, NET_IP_ALIGN);
 124         rx_buf->len = skb_len - NET_IP_ALIGN;
 125         rx_buf->data = (char *)rx_buf->skb->data;
 126         rx_buf->skb->ip_summed = CHECKSUM_UNNECESSARY;
 127
 128         rx_buf->dma_addr = pci_map_single(efx->pci_dev,
 129                                           rx_buf->data, rx_buf->len,
 130                                           PCI_DMA_FROMDEVICE);
 131
 132         if (unlikely(pci_dma_mapping_error(efx->pci_dev, rx_buf->dma_addr))) {
 133                 dev_kfree_skb_any(rx_buf->skb);
 134                 rx_buf->skb = NULL;
 135                 return -EIO;
 136         }
 137
 138         return 0;
 139 }
 140
 141 /**
 142  * efx_init_rx_buffer_page - create new RX buffer using page-based allocation
 143  *
 144  * @rx_queue:           Efx RX queue
 145  * @rx_buf:             RX buffer structure to populate
 146  *
 147  * This allocates memory for a new receive buffer, maps it for DMA,
 148  * and populates a struct efx_rx_buffer with the relevant
 149  * information.  Return a negative error code or 0 on success.
 150  */
 151 static int efx_init_rx_buffer_page(struct efx_rx_queue *rx_queue,
 152                                    struct efx_rx_buffer *rx_buf)
 153 {
 154         struct efx_nic *efx = rx_queue->efx;
 155         int bytes, space, offset;
 156
 157         bytes = efx->rx_buffer_len - EFX_PAGE_IP_ALIGN;
 158
 159         /* If there is space left in the previously allocated page,
 160          * then use it. Otherwise allocate a new one */
 161         rx_buf->page = rx_queue->buf_page;
 162         if (rx_buf->page == NULL) {
 163                 dma_addr_t dma_addr;
 164
 165                 rx_buf->page = alloc_pages(__GFP_COLD | __GFP_COMP | GFP_ATOMIC,
 166                                            efx->rx_buffer_order);
 167                 if (unlikely(rx_buf->page == NULL))
 168                         return -ENOMEM;
 169
 170                 dma_addr = pci_map_page(efx->pci_dev, rx_buf->page,
 171                                         0, efx_rx_buf_size(efx),
 172                                         PCI_DMA_FROMDEVICE);
 173
 174                 if (unlikely(pci_dma_mapping_error(efx->pci_dev, dma_addr))) {
 175                         __free_pages(rx_buf->page, efx->rx_buffer_order);
 176                         rx_buf->page = NULL;
 177                         return -EIO;
 178                 }
 179
 180                 rx_queue->buf_page = rx_buf->page;
 181                 rx_queue->buf_dma_addr = dma_addr;
 182                 rx_queue->buf_data = (page_address(rx_buf->page) +
 183                                       EFX_PAGE_IP_ALIGN);
 184         }
 185
 186         rx_buf->len = bytes;
 187         rx_buf->data = rx_queue->buf_data;
 188         offset = efx_rx_buf_offset(rx_buf);
 189         rx_buf->dma_addr = rx_queue->buf_dma_addr + offset;
 190
 191         /* Try to pack multiple buffers per page */
 192         if (efx->rx_buffer_order == 0) {
 193                 /* The next buffer starts on the next 512 byte boundary */
 194                 rx_queue->buf_data += ((bytes + 0x1ff) & ~0x1ff);
 195                 offset += ((bytes + 0x1ff) & ~0x1ff);
 196
 197                 space = efx_rx_buf_size(efx) - offset;
 198                 if (space >= bytes) {
 199                         /* Refs dropped on kernel releasing each skb */
 200                         get_page(rx_queue->buf_page);
 201                         goto out;
 202                 }
 203         }
 204
 205         /* This is the final RX buffer for this page, so mark it for
 206          * unmapping */
 207         rx_queue->buf_page = NULL;
 208         rx_buf->unmap_addr = rx_queue->buf_dma_addr;
 209
 210  out:
 211         return 0;
 212 }
 213
 214 /* This allocates memory for a new receive buffer, maps it for DMA,
 215  * and populates a struct efx_rx_buffer with the relevant
 216  * information.
 217  */
 218 static int efx_init_rx_buffer(struct efx_rx_queue *rx_queue,
 219                               struct efx_rx_buffer *new_rx_buf)
 220 {
 221         int rc = 0;
 222
 223         if (rx_queue->channel->rx_alloc_push_pages) {
 224                 new_rx_buf->skb = NULL;
 225                 rc = efx_init_rx_buffer_page(rx_queue, new_rx_buf);
 226                 rx_queue->alloc_page_count++;
 227         } else {
 228                 new_rx_buf->page = NULL;
 229                 rc = efx_init_rx_buffer_skb(rx_queue, new_rx_buf);
 230                 rx_queue->alloc_skb_count++;
 231         }
 232
 233         if (unlikely(rc < 0))
 234                 EFX_LOG_RL(rx_queue->efx, "%s RXQ[%d] =%d\n", __func__,
 235                            rx_queue->queue, rc);
 236         return rc;
 237 }
 238
 239 static void efx_unmap_rx_buffer(struct efx_nic *efx,
 240                                 struct efx_rx_buffer *rx_buf)
 241 {
 242         if (rx_buf->page) {
 243                 EFX_BUG_ON_PARANOID(rx_buf->skb);
 244                 if (rx_buf->unmap_addr) {
 245                         pci_unmap_page(efx->pci_dev, rx_buf->unmap_addr,
 246                                        efx_rx_buf_size(efx),
 247                                        PCI_DMA_FROMDEVICE);
 248                         rx_buf->unmap_addr = 0;
 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 /**
 276  * efx_fast_push_rx_descriptors - push new RX descriptors quickly
 277  * @rx_queue:           RX descriptor queue
 278  * @retry:              Recheck the fill level
 279  * This will aim to fill the RX descriptor queue up to
 280  * @rx_queue->@fast_fill_limit. If there is insufficient atomic
 281  * memory to do so, the caller should retry.
 282  */
 283 static int __efx_fast_push_rx_descriptors(struct efx_rx_queue *rx_queue,
 284                                           int retry)
 285 {
 286         struct efx_rx_buffer *rx_buf;
 287         unsigned fill_level, index;
 288         int i, space, rc = 0;
 289
 290         /* Calculate current fill level.  Do this outside the lock,
 291          * because most of the time we'll end up not wanting to do the
 292          * fill anyway.
 293          */
 294         fill_level = (rx_queue->added_count - rx_queue->removed_count);
 295         EFX_BUG_ON_PARANOID(fill_level > EFX_RXQ_SIZE);
 296
 297         /* Don't fill if we don't need to */
 298         if (fill_level >= rx_queue->fast_fill_trigger)
 299                 return 0;
 300
 301         /* Record minimum fill level */
 302         if (unlikely(fill_level < rx_queue->min_fill)) {
 303                 if (fill_level)
 304                         rx_queue->min_fill = fill_level;
 305         }
 306
 307         /* Acquire RX add lock.  If this lock is contended, then a fast
 308          * fill must already be in progress (e.g. in the refill
 309          * tasklet), so we don't need to do anything
 310          */
 311         if (!spin_trylock_bh(&rx_queue->add_lock))
 312                 return -1;
 313
 314  retry:
 315         /* Recalculate current fill level now that we have the lock */
 316         fill_level = (rx_queue->added_count - rx_queue->removed_count);
 317         EFX_BUG_ON_PARANOID(fill_level > EFX_RXQ_SIZE);
 318         space = rx_queue->fast_fill_limit - fill_level;
 319         if (space < EFX_RX_BATCH)
 320                 goto out_unlock;
 321
 322         EFX_TRACE(rx_queue->efx, "RX queue %d fast-filling descriptor ring from"
 323                   " level %d to level %d using %s allocation\n",
 324                   rx_queue->queue, fill_level, rx_queue->fast_fill_limit,
 325                   rx_queue->channel->rx_alloc_push_pages ? "page" : "skb");
 326
 327         do {
 328                 for (i = 0; i < EFX_RX_BATCH; ++i) {
 329                         index = rx_queue->added_count & EFX_RXQ_MASK;
 330                         rx_buf = efx_rx_buffer(rx_queue, index);
 331                         rc = efx_init_rx_buffer(rx_queue, rx_buf);
 332                         if (unlikely(rc))
 333                                 goto out;
 334                         ++rx_queue->added_count;
 335                 }
 336         } while ((space -= EFX_RX_BATCH) >= EFX_RX_BATCH);
 337
 338         EFX_TRACE(rx_queue->efx, "RX queue %d fast-filled descriptor ring "
 339                   "to level %d\n", rx_queue->queue,
 340                   rx_queue->added_count - rx_queue->removed_count);
 341
 342  out:
 343         /* Send write pointer to card. */
 344         efx_nic_notify_rx_desc(rx_queue);
 345
 346         /* If the fast fill is running inside from the refill tasklet, then
 347          * for SMP systems it may be running on a different CPU to
 348          * RX event processing, which means that the fill level may now be
 349          * out of date. */
 350         if (unlikely(retry && (rc == 0)))
 351                 goto retry;
 352
 353  out_unlock:
 354         spin_unlock_bh(&rx_queue->add_lock);
 355
 356         return rc;
 357 }
 358
 359 /**
 360  * efx_fast_push_rx_descriptors - push new RX descriptors quickly
 361  * @rx_queue:           RX descriptor queue
 362  *
 363  * This will aim to fill the RX descriptor queue up to
 364  * @rx_queue->@fast_fill_limit.  If there is insufficient memory to do so,
 365  * it will schedule a work item to immediately continue the fast fill
 366  */
 367 void efx_fast_push_rx_descriptors(struct efx_rx_queue *rx_queue)
 368 {
 369         int rc;
 370
 371         rc = __efx_fast_push_rx_descriptors(rx_queue, 0);
 372         if (unlikely(rc)) {
 373                 /* Schedule the work item to run immediately. The hope is
 374                  * that work is immediately pending to free some memory
 375                  * (e.g. an RX event or TX completion)
 376                  */
 377                 efx_schedule_slow_fill(rx_queue, 0);
 378         }
 379 }
 380
 381 void efx_rx_work(struct work_struct *data)
 382 {
 383         struct efx_rx_queue *rx_queue;
 384         int rc;
 385
 386         rx_queue = container_of(data, struct efx_rx_queue, work.work);
 387
 388         if (unlikely(!rx_queue->channel->enabled))
 389                 return;
 390
 391         EFX_TRACE(rx_queue->efx, "RX queue %d worker thread executing on CPU "
 392                   "%d\n", rx_queue->queue, raw_smp_processor_id());
 393
 394         ++rx_queue->slow_fill_count;
 395         /* Push new RX descriptors, allowing at least 1 jiffy for
 396          * the kernel to free some more memory. */
 397         rc = __efx_fast_push_rx_descriptors(rx_queue, 1);
 398         if (rc)
 399                 efx_schedule_slow_fill(rx_queue, 1);
 400 }
 401
 402 static void efx_rx_packet__check_len(struct efx_rx_queue *rx_queue,
 403                                      struct efx_rx_buffer *rx_buf,
 404                                      int len, bool *discard,
 405                                      bool *leak_packet)
 406 {
 407         struct efx_nic *efx = rx_queue->efx;
 408         unsigned max_len = rx_buf->len - efx->type->rx_buffer_padding;
 409
 410         if (likely(len <= max_len))
 411                 return;
 412
 413         /* The packet must be discarded, but this is only a fatal error
 414          * if the caller indicated it was
 415          */
 416         *discard = true;
 417
 418         if ((len > rx_buf->len) && EFX_WORKAROUND_8071(efx)) {
 419                 EFX_ERR_RL(efx, " RX queue %d seriously overlength "
 420                            "RX event (0x%x > 0x%x+0x%x). Leaking\n",
 421                            rx_queue->queue, len, max_len,
 422                            efx->type->rx_buffer_padding);
 423                 /* If this buffer was skb-allocated, then the meta
 424                  * data at the end of the skb will be trashed. So
 425                  * we have no choice but to leak the fragment.
 426                  */
 427                 *leak_packet = (rx_buf->skb != NULL);
 428                 efx_schedule_reset(efx, RESET_TYPE_RX_RECOVERY);
 429         } else {
 430                 EFX_ERR_RL(efx, " RX queue %d overlength RX event "
 431                            "(0x%x > 0x%x)\n", rx_queue->queue, len, max_len);
 432         }
 433
 434         rx_queue->channel->n_rx_overlength++;
 435 }
 436
 437 /* Pass a received packet up through the generic LRO stack
 438  *
 439  * Handles driverlink veto, and passes the fragment up via
 440  * the appropriate LRO method
 441  */
 442 static void efx_rx_packet_lro(struct efx_channel *channel,
 443                               struct efx_rx_buffer *rx_buf,
 444                               bool checksummed)
 445 {
 446         struct napi_struct *napi = &channel->napi_str;
 447         gro_result_t gro_result;
 448
 449         /* Pass the skb/page into the LRO engine */
 450         if (rx_buf->page) {
 451                 struct page *page = rx_buf->page;
 452                 struct sk_buff *skb;
 453
 454                 EFX_BUG_ON_PARANOID(rx_buf->skb);
 455                 rx_buf->page = NULL;
 456
 457                 skb = napi_get_frags(napi);
 458                 if (!skb) {
 459                         put_page(page);
 460                         return;
 461                 }
 462
 463                 skb_shinfo(skb)->frags[0].page = page;
 464                 skb_shinfo(skb)->frags[0].page_offset =
 465                         efx_rx_buf_offset(rx_buf);
 466                 skb_shinfo(skb)->frags[0].size = rx_buf->len;
 467                 skb_shinfo(skb)->nr_frags = 1;
 468
 469                 skb->len = rx_buf->len;
 470                 skb->data_len = rx_buf->len;
 471                 skb->truesize += rx_buf->len;
 472                 skb->ip_summed =
 473                         checksummed ? CHECKSUM_UNNECESSARY : CHECKSUM_NONE;
 474
 475                 skb_record_rx_queue(skb, channel->channel);
 476
 477                 gro_result = napi_gro_frags(napi);
 478         } else {
 479                 struct sk_buff *skb = rx_buf->skb;
 480
 481                 EFX_BUG_ON_PARANOID(!skb);
 482                 EFX_BUG_ON_PARANOID(!checksummed);
 483                 rx_buf->skb = NULL;
 484
 485                 gro_result = napi_gro_receive(napi, skb);
 486         }
 487
 488         if (gro_result == GRO_NORMAL) {
 489                 channel->rx_alloc_level += RX_ALLOC_FACTOR_SKB;
 490         } else if (gro_result != GRO_DROP) {
 491                 channel->rx_alloc_level += RX_ALLOC_FACTOR_LRO;
 492                 channel->irq_mod_score += 2;
 493         }
 494 }
 495
 496 void efx_rx_packet(struct efx_rx_queue *rx_queue, unsigned int index,
 497                    unsigned int len, bool checksummed, bool discard)
 498 {
 499         struct efx_nic *efx = rx_queue->efx;
 500         struct efx_rx_buffer *rx_buf;
 501         bool leak_packet = false;
 502
 503         rx_buf = efx_rx_buffer(rx_queue, index);
 504         EFX_BUG_ON_PARANOID(!rx_buf->data);
 505         EFX_BUG_ON_PARANOID(rx_buf->skb && rx_buf->page);
 506         EFX_BUG_ON_PARANOID(!(rx_buf->skb || rx_buf->page));
 507
 508         /* This allows the refill path to post another buffer.
 509          * EFX_RXD_HEAD_ROOM ensures that the slot we are using
 510          * isn't overwritten yet.
 511          */
 512         rx_queue->removed_count++;
 513
 514         /* Validate the length encoded in the event vs the descriptor pushed */
 515         efx_rx_packet__check_len(rx_queue, rx_buf, len,
 516                                  &discard, &leak_packet);
 517
 518         EFX_TRACE(efx, "RX queue %d received id %x at %llx+%x %s%s\n",
 519                   rx_queue->queue, index,
 520                   (unsigned long long)rx_buf->dma_addr, len,
 521                   (checksummed ? " [SUMMED]" : ""),
 522                   (discard ? " [DISCARD]" : ""));
 523
 524         /* Discard packet, if instructed to do so */
 525         if (unlikely(discard)) {
 526                 if (unlikely(leak_packet))
 527                         rx_queue->channel->n_skbuff_leaks++;
 528                 else
 529                         /* We haven't called efx_unmap_rx_buffer yet,
 530                          * so fini the entire rx_buffer here */
 531                         efx_fini_rx_buffer(rx_queue, rx_buf);
 532                 return;
 533         }
 534
 535         /* Release card resources - assumes all RX buffers consumed in-order
 536          * per RX queue
 537          */
 538         efx_unmap_rx_buffer(efx, rx_buf);
 539
 540         /* Prefetch nice and early so data will (hopefully) be in cache by
 541          * the time we look at it.
 542          */
 543         prefetch(rx_buf->data);
 544
 545         /* Pipeline receives so that we give time for packet headers to be
 546          * prefetched into cache.
 547          */
 548         rx_buf->len = len;
 549         if (rx_queue->channel->rx_pkt)
 550                 __efx_rx_packet(rx_queue->channel,
 551                                 rx_queue->channel->rx_pkt,
 552                                 rx_queue->channel->rx_pkt_csummed);
 553         rx_queue->channel->rx_pkt = rx_buf;
 554         rx_queue->channel->rx_pkt_csummed = checksummed;
 555 }
 556
 557 /* Handle a received packet.  Second half: Touches packet payload. */
 558 void __efx_rx_packet(struct efx_channel *channel,
 559                      struct efx_rx_buffer *rx_buf, bool checksummed)
 560 {
 561         struct efx_nic *efx = channel->efx;
 562         struct sk_buff *skb;
 563
 564         /* If we're in loopback test, then pass the packet directly to the
 565          * loopback layer, and free the rx_buf here
 566          */
 567         if (unlikely(efx->loopback_selftest)) {
 568                 efx_loopback_rx_packet(efx, rx_buf->data, rx_buf->len);
 569                 efx_free_rx_buffer(efx, rx_buf);
 570                 return;
 571         }
 572
 573         if (rx_buf->skb) {
 574                 prefetch(skb_shinfo(rx_buf->skb));
 575
 576                 skb_put(rx_buf->skb, rx_buf->len);
 577
 578                 /* Move past the ethernet header. rx_buf->data still points
 579                  * at the ethernet header */
 580                 rx_buf->skb->protocol = eth_type_trans(rx_buf->skb,
 581                                                        efx->net_dev);
 582
 583                 skb_record_rx_queue(rx_buf->skb, channel->channel);
 584         }
 585
 586         if (likely(checksummed || rx_buf->page)) {
 587                 efx_rx_packet_lro(channel, rx_buf, checksummed);
 588                 return;
 589         }
 590
 591         /* We now own the SKB */
 592         skb = rx_buf->skb;
 593         rx_buf->skb = NULL;
 594         EFX_BUG_ON_PARANOID(!skb);
 595
 596         /* Set the SKB flags */
 597         skb->ip_summed = CHECKSUM_NONE;
 598
 599         /* Pass the packet up */
 600         netif_receive_skb(skb);
 601
 602         /* Update allocation strategy method */
 603         channel->rx_alloc_level += RX_ALLOC_FACTOR_SKB;
 604 }
 605
 606 void efx_rx_strategy(struct efx_channel *channel)
 607 {
 608         enum efx_rx_alloc_method method = rx_alloc_method;
 609
 610         /* Only makes sense to use page based allocation if LRO is enabled */
 611         if (!(channel->efx->net_dev->features & NETIF_F_GRO)) {
 612                 method = RX_ALLOC_METHOD_SKB;
 613         } else if (method == RX_ALLOC_METHOD_AUTO) {
 614                 /* Constrain the rx_alloc_level */
 615                 if (channel->rx_alloc_level < 0)
 616                         channel->rx_alloc_level = 0;
 617                 else if (channel->rx_alloc_level > RX_ALLOC_LEVEL_MAX)
 618                         channel->rx_alloc_level = RX_ALLOC_LEVEL_MAX;
 619
 620                 /* Decide on the allocation method */
 621                 method = ((channel->rx_alloc_level > RX_ALLOC_LEVEL_LRO) ?
 622                           RX_ALLOC_METHOD_PAGE : RX_ALLOC_METHOD_SKB);
 623         }
 624
 625         /* Push the option */
 626         channel->rx_alloc_push_pages = (method == RX_ALLOC_METHOD_PAGE);
 627 }
 628
 629 int efx_probe_rx_queue(struct efx_rx_queue *rx_queue)
 630 {
 631         struct efx_nic *efx = rx_queue->efx;
 632         unsigned int rxq_size;
 633         int rc;
 634
 635         EFX_LOG(efx, "creating RX queue %d\n", rx_queue->queue);
 636
 637         /* Allocate RX buffers */
 638         rxq_size = EFX_RXQ_SIZE * sizeof(*rx_queue->buffer);
 639         rx_queue->buffer = kzalloc(rxq_size, GFP_KERNEL);
 640         if (!rx_queue->buffer)
 641                 return -ENOMEM;
 642
 643         rc = efx_nic_probe_rx(rx_queue);
 644         if (rc) {
 645                 kfree(rx_queue->buffer);
 646                 rx_queue->buffer = NULL;
 647         }
 648         return rc;
 649 }
 650
 651 void efx_init_rx_queue(struct efx_rx_queue *rx_queue)
 652 {
 653         unsigned int max_fill, trigger, limit;
 654
 655         EFX_LOG(rx_queue->efx, "initialising RX queue %d\n", rx_queue->queue);
 656
 657         /* Initialise ptr fields */
 658         rx_queue->added_count = 0;
 659         rx_queue->notified_count = 0;
 660         rx_queue->removed_count = 0;
 661         rx_queue->min_fill = -1U;
 662         rx_queue->min_overfill = -1U;
 663
 664         /* Initialise limit fields */
 665         max_fill = EFX_RXQ_SIZE - EFX_RXD_HEAD_ROOM;
 666         trigger = max_fill * min(rx_refill_threshold, 100U) / 100U;
 667         limit = max_fill * min(rx_refill_limit, 100U) / 100U;
 668
 669         rx_queue->max_fill = max_fill;
 670         rx_queue->fast_fill_trigger = trigger;
 671         rx_queue->fast_fill_limit = limit;
 672
 673         /* Set up RX descriptor ring */
 674         efx_nic_init_rx(rx_queue);
 675 }
 676
 677 void efx_fini_rx_queue(struct efx_rx_queue *rx_queue)
 678 {
 679         int i;
 680         struct efx_rx_buffer *rx_buf;
 681
 682         EFX_LOG(rx_queue->efx, "shutting down RX queue %d\n", rx_queue->queue);
 683
 684         efx_nic_fini_rx(rx_queue);
 685
 686         /* Release RX buffers NB start at index 0 not current HW ptr */
 687         if (rx_queue->buffer) {
 688                 for (i = 0; i <= EFX_RXQ_MASK; i++) {
 689                         rx_buf = efx_rx_buffer(rx_queue, i);
 690                         efx_fini_rx_buffer(rx_queue, rx_buf);
 691                 }
 692         }
 693
 694         /* For a page that is part-way through splitting into RX buffers */
 695         if (rx_queue->buf_page != NULL) {
 696                 pci_unmap_page(rx_queue->efx->pci_dev, rx_queue->buf_dma_addr,
 697                                efx_rx_buf_size(rx_queue->efx),
 698                                PCI_DMA_FROMDEVICE);
 699                 __free_pages(rx_queue->buf_page,
 700                              rx_queue->efx->rx_buffer_order);
 701                 rx_queue->buf_page = NULL;
 702         }
 703 }
 704
 705 void efx_remove_rx_queue(struct efx_rx_queue *rx_queue)
 706 {
 707         EFX_LOG(rx_queue->efx, "destroying RX queue %d\n", rx_queue->queue);
 708
 709         efx_nic_remove_rx(rx_queue);
 710
 711         kfree(rx_queue->buffer);
 712         rx_queue->buffer = NULL;
 713 }
 714
 715
 716 module_param(rx_alloc_method, int, 0644);
 717 MODULE_PARM_DESC(rx_alloc_method, "Allocation method used for RX buffers");
 718
 719 module_param(rx_refill_threshold, uint, 0444);
 720 MODULE_PARM_DESC(rx_refill_threshold,
 721                  "RX descriptor ring fast/slow fill threshold (%)");
 722