drivers/net/sfc/rx.c

   1 /****************************************************************************
   2  * Driver for Solarflare Solarstorm network controllers and boards
   3  * Copyright 2005-2006 Fen Systems Ltd.
   4  * Copyright 2005-2008 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 "rx.h"
  20 #include "efx.h"
  21 #include "falcon.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 /* Size of buffer allocated for skb header area. */
  29 #define EFX_SKB_HEADERS  64u
  30
  31 /*
  32  * rx_alloc_method - RX buffer allocation method
  33  *
  34  * This driver supports two methods for allocating and using RX buffers:
  35  * each RX buffer may be backed by an skb or by an order-n page.
  36  *
  37  * When LRO is in use then the second method has a lower overhead,
  38  * since we don't have to allocate then free skbs on reassembled frames.
  39  *
  40  * Values:
  41  *   - RX_ALLOC_METHOD_AUTO = 0
  42  *   - RX_ALLOC_METHOD_SKB  = 1
  43  *   - RX_ALLOC_METHOD_PAGE = 2
  44  *
  45  * The heuristic for %RX_ALLOC_METHOD_AUTO is a simple hysteresis count
  46  * controlled by the parameters below.
  47  *
  48  *   - Since pushing and popping descriptors are separated by the rx_queue
  49  *     size, so the watermarks should be ~rxd_size.
  50  *   - The performance win by using page-based allocation for LRO is less
  51  *     than the performance hit of using page-based allocation of non-LRO,
  52  *     so the watermarks should reflect this.
  53  *
  54  * Per channel we maintain a single variable, updated by each channel:
  55  *
  56  *   rx_alloc_level += (lro_performed ? RX_ALLOC_FACTOR_LRO :
  57  *                      RX_ALLOC_FACTOR_SKB)
  58  * Per NAPI poll interval, we constrain rx_alloc_level to 0..MAX (which
  59  * limits the hysteresis), and update the allocation strategy:
  60  *
  61  *   rx_alloc_method = (rx_alloc_level > RX_ALLOC_LEVEL_LRO ?
  62  *                      RX_ALLOC_METHOD_PAGE : RX_ALLOC_METHOD_SKB)
  63  */
  64 static int rx_alloc_method = RX_ALLOC_METHOD_AUTO;
  65
  66 #define RX_ALLOC_LEVEL_LRO 0x2000
  67 #define RX_ALLOC_LEVEL_MAX 0x3000
  68 #define RX_ALLOC_FACTOR_LRO 1
  69 #define RX_ALLOC_FACTOR_SKB (-2)
  70
  71 /* This is the percentage fill level below which new RX descriptors
  72  * will be added to the RX descriptor ring.
  73  */
  74 static unsigned int rx_refill_threshold = 90;
  75
  76 /* This is the percentage fill level to which an RX queue will be refilled
  77  * when the "RX refill threshold" is reached.
  78  */
  79 static unsigned int rx_refill_limit = 95;
  80
  81 /*
  82  * RX maximum head room required.
  83  *
  84  * This must be at least 1 to prevent overflow and at least 2 to allow
  85  * pipelined receives.
  86  */
  87 #define EFX_RXD_HEAD_ROOM 2
  88
  89 static inline unsigned int efx_rx_buf_offset(struct efx_rx_buffer *buf)
  90 {
  91         /* Offset is always within one page, so we don't need to consider
  92          * the page order.
  93          */
  94         return (__force unsigned long) buf->data & (PAGE_SIZE - 1);
  95 }
  96 static inline unsigned int efx_rx_buf_size(struct efx_nic *efx)
  97 {
  98         return PAGE_SIZE << efx->rx_buffer_order;
  99 }
 100
 101
 102 /**
 103  * efx_init_rx_buffer_skb - create new RX buffer using skb-based allocation
 104  *
 105  * @rx_queue:           Efx RX queue
 106  * @rx_buf:             RX buffer structure to populate
 107  *
 108  * This allocates memory for a new receive buffer, maps it for DMA,
 109  * and populates a struct efx_rx_buffer with the relevant
 110  * information.  Return a negative error code or 0 on success.
 111  */
 112 static int efx_init_rx_buffer_skb(struct efx_rx_queue *rx_queue,
 113                                   struct efx_rx_buffer *rx_buf)
 114 {
 115         struct efx_nic *efx = rx_queue->efx;
 116         struct net_device *net_dev = efx->net_dev;
 117         int skb_len = efx->rx_buffer_len;
 118
 119         rx_buf->skb = netdev_alloc_skb(net_dev, skb_len);
 120         if (unlikely(!rx_buf->skb))
 121                 return -ENOMEM;
 122
 123         /* Adjust the SKB for padding and checksum */
 124         skb_reserve(rx_buf->skb, NET_IP_ALIGN);
 125         rx_buf->len = skb_len - NET_IP_ALIGN;
 126         rx_buf->data = (char *)rx_buf->skb->data;
 127         rx_buf->skb->ip_summed = CHECKSUM_UNNECESSARY;
 128
 129         rx_buf->dma_addr = pci_map_single(efx->pci_dev,
 130                                           rx_buf->data, rx_buf->len,
 131                                           PCI_DMA_FROMDEVICE);
 132
 133         if (unlikely(pci_dma_mapping_error(efx->pci_dev, rx_buf->dma_addr))) {
 134                 dev_kfree_skb_any(rx_buf->skb);
 135                 rx_buf->skb = NULL;
 136                 return -EIO;
 137         }
 138
 139         return 0;
 140 }
 141
 142 /**
 143  * efx_init_rx_buffer_page - create new RX buffer using page-based allocation
 144  *
 145  * @rx_queue:           Efx RX queue
 146  * @rx_buf:             RX buffer structure to populate
 147  *
 148  * This allocates memory for a new receive buffer, maps it for DMA,
 149  * and populates a struct efx_rx_buffer with the relevant
 150  * information.  Return a negative error code or 0 on success.
 151  */
 152 static int efx_init_rx_buffer_page(struct efx_rx_queue *rx_queue,
 153                                    struct efx_rx_buffer *rx_buf)
 154 {
 155         struct efx_nic *efx = rx_queue->efx;
 156         int bytes, space, offset;
 157
 158         bytes = efx->rx_buffer_len - EFX_PAGE_IP_ALIGN;
 159
 160         /* If there is space left in the previously allocated page,
 161          * then use it. Otherwise allocate a new one */
 162         rx_buf->page = rx_queue->buf_page;
 163         if (rx_buf->page == NULL) {
 164                 dma_addr_t dma_addr;
 165
 166                 rx_buf->page = alloc_pages(__GFP_COLD | __GFP_COMP | GFP_ATOMIC,
 167                                            efx->rx_buffer_order);
 168                 if (unlikely(rx_buf->page == NULL))
 169                         return -ENOMEM;
 170
 171                 dma_addr = pci_map_page(efx->pci_dev, rx_buf->page,
 172                                         0, efx_rx_buf_size(efx),
 173                                         PCI_DMA_FROMDEVICE);
 174
 175                 if (unlikely(pci_dma_mapping_error(efx->pci_dev, dma_addr))) {
 176                         __free_pages(rx_buf->page, efx->rx_buffer_order);
 177                         rx_buf->page = NULL;
 178                         return -EIO;
 179                 }
 180
 181                 rx_queue->buf_page = rx_buf->page;
 182                 rx_queue->buf_dma_addr = dma_addr;
 183                 rx_queue->buf_data = (page_address(rx_buf->page) +
 184                                       EFX_PAGE_IP_ALIGN);
 185         }
 186
 187         rx_buf->len = bytes;
 188         rx_buf->data = rx_queue->buf_data;
 189         offset = efx_rx_buf_offset(rx_buf);
 190         rx_buf->dma_addr = rx_queue->buf_dma_addr + offset;
 191
 192         /* Try to pack multiple buffers per page */
 193         if (efx->rx_buffer_order == 0) {
 194                 /* The next buffer starts on the next 512 byte boundary */
 195                 rx_queue->buf_data += ((bytes + 0x1ff) & ~0x1ff);
 196                 offset += ((bytes + 0x1ff) & ~0x1ff);
 197
 198                 space = efx_rx_buf_size(efx) - offset;
 199                 if (space >= bytes) {
 200                         /* Refs dropped on kernel releasing each skb */
 201                         get_page(rx_queue->buf_page);
 202                         goto out;
 203                 }
 204         }
 205
 206         /* This is the final RX buffer for this page, so mark it for
 207          * unmapping */
 208         rx_queue->buf_page = NULL;
 209         rx_buf->unmap_addr = rx_queue->buf_dma_addr;
 210
 211  out:
 212         return 0;
 213 }
 214
 215 /* This allocates memory for a new receive buffer, maps it for DMA,
 216  * and populates a struct efx_rx_buffer with the relevant
 217  * information.
 218  */
 219 static int efx_init_rx_buffer(struct efx_rx_queue *rx_queue,
 220                               struct efx_rx_buffer *new_rx_buf)
 221 {
 222         int rc = 0;
 223
 224         if (rx_queue->channel->rx_alloc_push_pages) {
 225                 new_rx_buf->skb = NULL;
 226                 rc = efx_init_rx_buffer_page(rx_queue, new_rx_buf);
 227                 rx_queue->alloc_page_count++;
 228         } else {
 229                 new_rx_buf->page = NULL;
 230                 rc = efx_init_rx_buffer_skb(rx_queue, new_rx_buf);
 231                 rx_queue->alloc_skb_count++;
 232         }
 233
 234         if (unlikely(rc < 0))
 235                 EFX_LOG_RL(rx_queue->efx, "%s RXQ[%d] =%d\n", __func__,
 236                            rx_queue->queue, rc);
 237         return rc;
 238 }
 239
 240 static void efx_unmap_rx_buffer(struct efx_nic *efx,
 241                                 struct efx_rx_buffer *rx_buf)
 242 {
 243         if (rx_buf->page) {
 244                 EFX_BUG_ON_PARANOID(rx_buf->skb);
 245                 if (rx_buf->unmap_addr) {
 246                         pci_unmap_page(efx->pci_dev, rx_buf->unmap_addr,
 247                                        efx_rx_buf_size(efx),
 248                                        PCI_DMA_FROMDEVICE);
 249                         rx_buf->unmap_addr = 0;
 250                 }
 251         } else if (likely(rx_buf->skb)) {
 252                 pci_unmap_single(efx->pci_dev, rx_buf->dma_addr,
 253                                  rx_buf->len, PCI_DMA_FROMDEVICE);
 254         }
 255 }
 256
 257 static void efx_free_rx_buffer(struct efx_nic *efx,
 258                                struct efx_rx_buffer *rx_buf)
 259 {
 260         if (rx_buf->page) {
 261                 __free_pages(rx_buf->page, efx->rx_buffer_order);
 262                 rx_buf->page = NULL;
 263         } else if (likely(rx_buf->skb)) {
 264                 dev_kfree_skb_any(rx_buf->skb);
 265                 rx_buf->skb = NULL;
 266         }
 267 }
 268
 269 static void efx_fini_rx_buffer(struct efx_rx_queue *rx_queue,
 270                                struct efx_rx_buffer *rx_buf)
 271 {
 272         efx_unmap_rx_buffer(rx_queue->efx, rx_buf);
 273         efx_free_rx_buffer(rx_queue->efx, rx_buf);
 274 }
 275
 276 /**
 277  * efx_fast_push_rx_descriptors - push new RX descriptors quickly
 278  * @rx_queue:           RX descriptor queue
 279  * @retry:              Recheck the fill level
 280  * This will aim to fill the RX descriptor queue up to
 281  * @rx_queue->@fast_fill_limit. If there is insufficient atomic
 282  * memory to do so, the caller should retry.
 283  */
 284 static int __efx_fast_push_rx_descriptors(struct efx_rx_queue *rx_queue,
 285                                           int retry)
 286 {
 287         struct efx_rx_buffer *rx_buf;
 288         unsigned fill_level, index;
 289         int i, space, rc = 0;
 290
 291         /* Calculate current fill level.  Do this outside the lock,
 292          * because most of the time we'll end up not wanting to do the
 293          * fill anyway.
 294          */
 295         fill_level = (rx_queue->added_count - rx_queue->removed_count);
 296         EFX_BUG_ON_PARANOID(fill_level > EFX_RXQ_SIZE);
 297
 298         /* Don't fill if we don't need to */
 299         if (fill_level >= rx_queue->fast_fill_trigger)
 300                 return 0;
 301
 302         /* Record minimum fill level */
 303         if (unlikely(fill_level < rx_queue->min_fill)) {
 304                 if (fill_level)
 305                         rx_queue->min_fill = fill_level;
 306         }
 307
 308         /* Acquire RX add lock.  If this lock is contended, then a fast
 309          * fill must already be in progress (e.g. in the refill
 310          * tasklet), so we don't need to do anything
 311          */
 312         if (!spin_trylock_bh(&rx_queue->add_lock))
 313                 return -1;
 314
 315  retry:
 316         /* Recalculate current fill level now that we have the lock */
 317         fill_level = (rx_queue->added_count - rx_queue->removed_count);
 318         EFX_BUG_ON_PARANOID(fill_level > EFX_RXQ_SIZE);
 319         space = rx_queue->fast_fill_limit - fill_level;
 320         if (space < EFX_RX_BATCH)
 321                 goto out_unlock;
 322
 323         EFX_TRACE(rx_queue->efx, "RX queue %d fast-filling descriptor ring from"
 324                   " level %d to level %d using %s allocation\n",
 325                   rx_queue->queue, fill_level, rx_queue->fast_fill_limit,
 326                   rx_queue->channel->rx_alloc_push_pages ? "page" : "skb");
 327
 328         do {
 329                 for (i = 0; i < EFX_RX_BATCH; ++i) {
 330                         index = rx_queue->added_count & EFX_RXQ_MASK;
 331                         rx_buf = efx_rx_buffer(rx_queue, index);
 332                         rc = efx_init_rx_buffer(rx_queue, rx_buf);
 333                         if (unlikely(rc))
 334                                 goto out;
 335                         ++rx_queue->added_count;
 336                 }
 337         } while ((space -= EFX_RX_BATCH) >= EFX_RX_BATCH);
 338
 339         EFX_TRACE(rx_queue->efx, "RX queue %d fast-filled descriptor ring "
 340                   "to level %d\n", rx_queue->queue,
 341                   rx_queue->added_count - rx_queue->removed_count);
 342
 343  out:
 344         /* Send write pointer to card. */
 345         falcon_notify_rx_desc(rx_queue);
 346
 347         /* If the fast fill is running inside from the refill tasklet, then
 348          * for SMP systems it may be running on a different CPU to
 349          * RX event processing, which means that the fill level may now be
 350          * out of date. */
 351         if (unlikely(retry && (rc == 0)))
 352                 goto retry;
 353
 354  out_unlock:
 355         spin_unlock_bh(&rx_queue->add_lock);
 356
 357         return rc;
 358 }
 359
 360 /**
 361  * efx_fast_push_rx_descriptors - push new RX descriptors quickly
 362  * @rx_queue:           RX descriptor queue
 363  *
 364  * This will aim to fill the RX descriptor queue up to
 365  * @rx_queue->@fast_fill_limit.  If there is insufficient memory to do so,
 366  * it will schedule a work item to immediately continue the fast fill
 367  */
 368 void efx_fast_push_rx_descriptors(struct efx_rx_queue *rx_queue)
 369 {
 370         int rc;
 371
 372         rc = __efx_fast_push_rx_descriptors(rx_queue, 0);
 373         if (unlikely(rc)) {
 374                 /* Schedule the work item to run immediately. The hope is
 375                  * that work is immediately pending to free some memory
 376                  * (e.g. an RX event or TX completion)
 377                  */
 378                 efx_schedule_slow_fill(rx_queue, 0);
 379         }
 380 }
 381
 382 void efx_rx_work(struct work_struct *data)
 383 {
 384         struct efx_rx_queue *rx_queue;
 385         int rc;
 386
 387         rx_queue = container_of(data, struct efx_rx_queue, work.work);
 388
 389         if (unlikely(!rx_queue->channel->enabled))
 390                 return;
 391
 392         EFX_TRACE(rx_queue->efx, "RX queue %d worker thread executing on CPU "
 393                   "%d\n", rx_queue->queue, raw_smp_processor_id());
 394
 395         ++rx_queue->slow_fill_count;
 396         /* Push new RX descriptors, allowing at least 1 jiffy for
 397          * the kernel to free some more memory. */
 398         rc = __efx_fast_push_rx_descriptors(rx_queue, 1);
 399         if (rc)
 400                 efx_schedule_slow_fill(rx_queue, 1);
 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                 EFX_ERR_RL(efx, " RX queue %d seriously overlength "
 421                            "RX event (0x%x > 0x%x+0x%x). Leaking\n",
 422                            rx_queue->queue, len, max_len,
 423                            efx->type->rx_buffer_padding);
 424                 /* If this buffer was skb-allocated, then the meta
 425                  * data at the end of the skb will be trashed. So
 426                  * we have no choice but to leak the fragment.
 427                  */
 428                 *leak_packet = (rx_buf->skb != NULL);
 429                 efx_schedule_reset(efx, RESET_TYPE_RX_RECOVERY);
 430         } else {
 431                 EFX_ERR_RL(efx, " RX queue %d overlength RX event "
 432                            "(0x%x > 0x%x)\n", rx_queue->queue, len, max_len);
 433         }
 434
 435         rx_queue->channel->n_rx_overlength++;
 436 }
 437
 438 /* Pass a received packet up through the generic LRO stack
 439  *
 440  * Handles driverlink veto, and passes the fragment up via
 441  * the appropriate LRO method
 442  */
 443 static void efx_rx_packet_lro(struct efx_channel *channel,
 444                               struct efx_rx_buffer *rx_buf,
 445                               bool checksummed)
 446 {
 447         struct napi_struct *napi = &channel->napi_str;
 448         gro_result_t gro_result;
 449
 450         /* Pass the skb/page into the LRO engine */
 451         if (rx_buf->page) {
 452                 struct sk_buff *skb = napi_get_frags(napi);
 453
 454                 if (!skb) {
 455                         put_page(rx_buf->page);
 456                         gro_result = GRO_DROP;
 457                         goto out;
 458                 }
 459
 460                 skb_shinfo(skb)->frags[0].page = rx_buf->page;
 461                 skb_shinfo(skb)->frags[0].page_offset =
 462                         efx_rx_buf_offset(rx_buf);
 463                 skb_shinfo(skb)->frags[0].size = rx_buf->len;
 464                 skb_shinfo(skb)->nr_frags = 1;
 465
 466                 skb->len = rx_buf->len;
 467                 skb->data_len = rx_buf->len;
 468                 skb->truesize += rx_buf->len;
 469                 skb->ip_summed =
 470                         checksummed ? CHECKSUM_UNNECESSARY : CHECKSUM_NONE;
 471
 472                 gro_result = napi_gro_frags(napi);
 473
 474 out:
 475                 EFX_BUG_ON_PARANOID(rx_buf->skb);
 476                 rx_buf->page = NULL;
 477         } else {
 478                 EFX_BUG_ON_PARANOID(!rx_buf->skb);
 479                 EFX_BUG_ON_PARANOID(!checksummed);
 480
 481                 gro_result = napi_gro_receive(napi, rx_buf->skb);
 482                 rx_buf->skb = NULL;
 483         }
 484
 485         if (gro_result == GRO_NORMAL) {
 486                 channel->rx_alloc_level += RX_ALLOC_FACTOR_SKB;
 487         } else if (gro_result != GRO_DROP) {
 488                 channel->rx_alloc_level += RX_ALLOC_FACTOR_LRO;
 489                 channel->irq_mod_score += 2;
 490         }
 491 }
 492
 493 void efx_rx_packet(struct efx_rx_queue *rx_queue, unsigned int index,
 494                    unsigned int len, bool checksummed, bool discard)
 495 {
 496         struct efx_nic *efx = rx_queue->efx;
 497         struct efx_rx_buffer *rx_buf;
 498         bool leak_packet = false;
 499
 500         rx_buf = efx_rx_buffer(rx_queue, index);
 501         EFX_BUG_ON_PARANOID(!rx_buf->data);
 502         EFX_BUG_ON_PARANOID(rx_buf->skb && rx_buf->page);
 503         EFX_BUG_ON_PARANOID(!(rx_buf->skb || rx_buf->page));
 504
 505         /* This allows the refill path to post another buffer.
 506          * EFX_RXD_HEAD_ROOM ensures that the slot we are using
 507          * isn't overwritten yet.
 508          */
 509         rx_queue->removed_count++;
 510
 511         /* Validate the length encoded in the event vs the descriptor pushed */
 512         efx_rx_packet__check_len(rx_queue, rx_buf, len,
 513                                  &discard, &leak_packet);
 514
 515         EFX_TRACE(efx, "RX queue %d received id %x at %llx+%x %s%s\n",
 516                   rx_queue->queue, index,
 517                   (unsigned long long)rx_buf->dma_addr, len,
 518                   (checksummed ? " [SUMMED]" : ""),
 519                   (discard ? " [DISCARD]" : ""));
 520
 521         /* Discard packet, if instructed to do so */
 522         if (unlikely(discard)) {
 523                 if (unlikely(leak_packet))
 524                         rx_queue->channel->n_skbuff_leaks++;
 525                 else
 526                         /* We haven't called efx_unmap_rx_buffer yet,
 527                          * so fini the entire rx_buffer here */
 528                         efx_fini_rx_buffer(rx_queue, rx_buf);
 529                 return;
 530         }
 531
 532         /* Release card resources - assumes all RX buffers consumed in-order
 533          * per RX queue
 534          */
 535         efx_unmap_rx_buffer(efx, rx_buf);
 536
 537         /* Prefetch nice and early so data will (hopefully) be in cache by
 538          * the time we look at it.
 539          */
 540         prefetch(rx_buf->data);
 541
 542         /* Pipeline receives so that we give time for packet headers to be
 543          * prefetched into cache.
 544          */
 545         rx_buf->len = len;
 546         if (rx_queue->channel->rx_pkt)
 547                 __efx_rx_packet(rx_queue->channel,
 548                                 rx_queue->channel->rx_pkt,
 549                                 rx_queue->channel->rx_pkt_csummed);
 550         rx_queue->channel->rx_pkt = rx_buf;
 551         rx_queue->channel->rx_pkt_csummed = checksummed;
 552 }
 553
 554 /* Handle a received packet.  Second half: Touches packet payload. */
 555 void __efx_rx_packet(struct efx_channel *channel,
 556                      struct efx_rx_buffer *rx_buf, bool checksummed)
 557 {
 558         struct efx_nic *efx = channel->efx;
 559         struct sk_buff *skb;
 560
 561         /* If we're in loopback test, then pass the packet directly to the
 562          * loopback layer, and free the rx_buf here
 563          */
 564         if (unlikely(efx->loopback_selftest)) {
 565                 efx_loopback_rx_packet(efx, rx_buf->data, rx_buf->len);
 566                 efx_free_rx_buffer(efx, rx_buf);
 567                 return;
 568         }
 569
 570         if (rx_buf->skb) {
 571                 prefetch(skb_shinfo(rx_buf->skb));
 572
 573                 skb_put(rx_buf->skb, rx_buf->len);
 574
 575                 /* Move past the ethernet header. rx_buf->data still points
 576                  * at the ethernet header */
 577                 rx_buf->skb->protocol = eth_type_trans(rx_buf->skb,
 578                                                        efx->net_dev);
 579         }
 580
 581         if (likely(checksummed || rx_buf->page)) {
 582                 efx_rx_packet_lro(channel, rx_buf, checksummed);
 583                 return;
 584         }
 585
 586         /* We now own the SKB */
 587         skb = rx_buf->skb;
 588         rx_buf->skb = NULL;
 589         EFX_BUG_ON_PARANOID(!skb);
 590
 591         /* Set the SKB flags */
 592         skb->ip_summed = CHECKSUM_NONE;
 593
 594         skb_record_rx_queue(skb, channel->channel);
 595
 596         /* Pass the packet up */
 597         netif_receive_skb(skb);
 598
 599         /* Update allocation strategy method */
 600         channel->rx_alloc_level += RX_ALLOC_FACTOR_SKB;
 601 }
 602
 603 void efx_rx_strategy(struct efx_channel *channel)
 604 {
 605         enum efx_rx_alloc_method method = rx_alloc_method;
 606
 607         /* Only makes sense to use page based allocation if LRO is enabled */
 608         if (!(channel->efx->net_dev->features & NETIF_F_GRO)) {
 609                 method = RX_ALLOC_METHOD_SKB;
 610         } else if (method == RX_ALLOC_METHOD_AUTO) {
 611                 /* Constrain the rx_alloc_level */
 612                 if (channel->rx_alloc_level < 0)
 613                         channel->rx_alloc_level = 0;
 614                 else if (channel->rx_alloc_level > RX_ALLOC_LEVEL_MAX)
 615                         channel->rx_alloc_level = RX_ALLOC_LEVEL_MAX;
 616
 617                 /* Decide on the allocation method */
 618                 method = ((channel->rx_alloc_level > RX_ALLOC_LEVEL_LRO) ?
 619                           RX_ALLOC_METHOD_PAGE : RX_ALLOC_METHOD_SKB);
 620         }
 621
 622         /* Push the option */
 623         channel->rx_alloc_push_pages = (method == RX_ALLOC_METHOD_PAGE);
 624 }
 625
 626 int efx_probe_rx_queue(struct efx_rx_queue *rx_queue)
 627 {
 628         struct efx_nic *efx = rx_queue->efx;
 629         unsigned int rxq_size;
 630         int rc;
 631
 632         EFX_LOG(efx, "creating RX queue %d\n", rx_queue->queue);
 633
 634         /* Allocate RX buffers */
 635         rxq_size = EFX_RXQ_SIZE * sizeof(*rx_queue->buffer);
 636         rx_queue->buffer = kzalloc(rxq_size, GFP_KERNEL);
 637         if (!rx_queue->buffer)
 638                 return -ENOMEM;
 639
 640         rc = falcon_probe_rx(rx_queue);
 641         if (rc) {
 642                 kfree(rx_queue->buffer);
 643                 rx_queue->buffer = NULL;
 644         }
 645         return rc;
 646 }
 647
 648 void efx_init_rx_queue(struct efx_rx_queue *rx_queue)
 649 {
 650         unsigned int max_fill, trigger, limit;
 651
 652         EFX_LOG(rx_queue->efx, "initialising RX queue %d\n", rx_queue->queue);
 653
 654         /* Initialise ptr fields */
 655         rx_queue->added_count = 0;
 656         rx_queue->notified_count = 0;
 657         rx_queue->removed_count = 0;
 658         rx_queue->min_fill = -1U;
 659         rx_queue->min_overfill = -1U;
 660
 661         /* Initialise limit fields */
 662         max_fill = EFX_RXQ_SIZE - EFX_RXD_HEAD_ROOM;
 663         trigger = max_fill * min(rx_refill_threshold, 100U) / 100U;
 664         limit = max_fill * min(rx_refill_limit, 100U) / 100U;
 665
 666         rx_queue->max_fill = max_fill;
 667         rx_queue->fast_fill_trigger = trigger;
 668         rx_queue->fast_fill_limit = limit;
 669
 670         /* Set up RX descriptor ring */
 671         falcon_init_rx(rx_queue);
 672 }
 673
 674 void efx_fini_rx_queue(struct efx_rx_queue *rx_queue)
 675 {
 676         int i;
 677         struct efx_rx_buffer *rx_buf;
 678
 679         EFX_LOG(rx_queue->efx, "shutting down RX queue %d\n", rx_queue->queue);
 680
 681         falcon_fini_rx(rx_queue);
 682
 683         /* Release RX buffers NB start at index 0 not current HW ptr */
 684         if (rx_queue->buffer) {
 685                 for (i = 0; i <= EFX_RXQ_MASK; i++) {
 686                         rx_buf = efx_rx_buffer(rx_queue, i);
 687                         efx_fini_rx_buffer(rx_queue, rx_buf);
 688                 }
 689         }
 690
 691         /* For a page that is part-way through splitting into RX buffers */
 692         if (rx_queue->buf_page != NULL) {
 693                 pci_unmap_page(rx_queue->efx->pci_dev, rx_queue->buf_dma_addr,
 694                                efx_rx_buf_size(rx_queue->efx),
 695                                PCI_DMA_FROMDEVICE);
 696                 __free_pages(rx_queue->buf_page,
 697                              rx_queue->efx->rx_buffer_order);
 698                 rx_queue->buf_page = NULL;
 699         }
 700 }
 701
 702 void efx_remove_rx_queue(struct efx_rx_queue *rx_queue)
 703 {
 704         EFX_LOG(rx_queue->efx, "destroying RX queue %d\n", rx_queue->queue);
 705
 706         falcon_remove_rx(rx_queue);
 707
 708         kfree(rx_queue->buffer);
 709         rx_queue->buffer = NULL;
 710 }
 711
 712
 713 module_param(rx_alloc_method, int, 0644);
 714 MODULE_PARM_DESC(rx_alloc_method, "Allocation method used for RX buffers");
 715
 716 module_param(rx_refill_threshold, uint, 0444);
 717 MODULE_PARM_DESC(rx_refill_threshold,
 718                  "RX descriptor ring fast/slow fill threshold (%)");
 719