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 <net/ip.h>
  20 #include <net/checksum.h>
  21 #include "net_driver.h"
  22 #include "efx.h"
  23 #include "nic.h"
  24 #include "selftest.h"
  25 #include "workarounds.h"
  26
  27 /* Number of RX descriptors pushed at once. */
  28 #define EFX_RX_BATCH  8
  29
  30 /* Maximum size of a buffer sharing a page */
  31 #define EFX_RX_HALF_PAGE ((PAGE_SIZE >> 1) - sizeof(struct efx_rx_page_state))
  32
  33 /* Size of buffer allocated for skb header area. */
  34 #define EFX_SKB_HEADERS  64u
  35
  36 /*
  37  * rx_alloc_method - RX buffer allocation method
  38  *
  39  * This driver supports two methods for allocating and using RX buffers:
  40  * each RX buffer may be backed by an skb or by an order-n page.
  41  *
  42  * When GRO is in use then the second method has a lower overhead,
  43  * since we don't have to allocate then free skbs on reassembled frames.
  44  *
  45  * Values:
  46  *   - RX_ALLOC_METHOD_AUTO = 0
  47  *   - RX_ALLOC_METHOD_SKB  = 1
  48  *   - RX_ALLOC_METHOD_PAGE = 2
  49  *
  50  * The heuristic for %RX_ALLOC_METHOD_AUTO is a simple hysteresis count
  51  * controlled by the parameters below.
  52  *
  53  *   - Since pushing and popping descriptors are separated by the rx_queue
  54  *     size, so the watermarks should be ~rxd_size.
  55  *   - The performance win by using page-based allocation for GRO is less
  56  *     than the performance hit of using page-based allocation of non-GRO,
  57  *     so the watermarks should reflect this.
  58  *
  59  * Per channel we maintain a single variable, updated by each channel:
  60  *
  61  *   rx_alloc_level += (gro_performed ? RX_ALLOC_FACTOR_GRO :
  62  *                      RX_ALLOC_FACTOR_SKB)
  63  * Per NAPI poll interval, we constrain rx_alloc_level to 0..MAX (which
  64  * limits the hysteresis), and update the allocation strategy:
  65  *
  66  *   rx_alloc_method = (rx_alloc_level > RX_ALLOC_LEVEL_GRO ?
  67  *                      RX_ALLOC_METHOD_PAGE : RX_ALLOC_METHOD_SKB)
  68  */
  69 static int rx_alloc_method = RX_ALLOC_METHOD_AUTO;
  70
  71 #define RX_ALLOC_LEVEL_GRO 0x2000
  72 #define RX_ALLOC_LEVEL_MAX 0x3000
  73 #define RX_ALLOC_FACTOR_GRO 1
  74 #define RX_ALLOC_FACTOR_SKB (-2)
  75
  76 /* This is the percentage fill level below which new RX descriptors
  77  * will be added to the RX descriptor ring.
  78  */
  79 static unsigned int rx_refill_threshold = 90;
  80
  81 /* This is the percentage fill level to which an RX queue will be refilled
  82  * when the "RX refill threshold" is reached.
  83  */
  84 static unsigned int rx_refill_limit = 95;
  85
  86 /*
  87  * RX maximum head room required.
  88  *
  89  * This must be at least 1 to prevent overflow and at least 2 to allow
  90  * pipelined receives.
  91  */
  92 #define EFX_RXD_HEAD_ROOM 2
  93
  94 /* Offset of ethernet header within page */
  95 static inline unsigned int efx_rx_buf_offset(struct efx_nic *efx,
  96                                              struct efx_rx_buffer *buf)
  97 {
  98         return buf->page_offset + efx->type->rx_buffer_hash_size;
  99 }
 100 static inline unsigned int efx_rx_buf_size(struct efx_nic *efx)
 101 {
 102         return PAGE_SIZE << efx->rx_buffer_order;
 103 }
 104
 105 static u8 *efx_rx_buf_eh(struct efx_nic *efx, struct efx_rx_buffer *buf)
 106 {
 107         if (buf->is_page)
 108                 return page_address(buf->u.page) + efx_rx_buf_offset(efx, buf);
 109         else
 110                 return ((u8 *)buf->u.skb->data +
 111                         efx->type->rx_buffer_hash_size);
 112 }
 113
 114 static inline u32 efx_rx_buf_hash(const u8 *eh)
 115 {
 116         /* The ethernet header is always directly after any hash. */
 117 #if defined(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS) || NET_IP_ALIGN % 4 == 0
 118         return __le32_to_cpup((const __le32 *)(eh - 4));
 119 #else
 120         const u8 *data = eh - 4;
 121         return ((u32)data[0]       |
 122                 (u32)data[1] << 8  |
 123                 (u32)data[2] << 16 |
 124                 (u32)data[3] << 24);
 125 #endif
 126 }
 127
 128 /**
 129  * efx_init_rx_buffers_skb - create EFX_RX_BATCH skb-based RX buffers
 130  *
 131  * @rx_queue:           Efx RX queue
 132  *
 133  * This allocates EFX_RX_BATCH skbs, maps them for DMA, and populates a
 134  * struct efx_rx_buffer for each one. Return a negative error code or 0
 135  * on success. May fail having only inserted fewer than EFX_RX_BATCH
 136  * buffers.
 137  */
 138 static int efx_init_rx_buffers_skb(struct efx_rx_queue *rx_queue)
 139 {
 140         struct efx_nic *efx = rx_queue->efx;
 141         struct net_device *net_dev = efx->net_dev;
 142         struct efx_rx_buffer *rx_buf;
 143         struct sk_buff *skb;
 144         int skb_len = efx->rx_buffer_len;
 145         unsigned index, count;
 146
 147         for (count = 0; count < EFX_RX_BATCH; ++count) {
 148                 index = rx_queue->added_count & rx_queue->ptr_mask;
 149                 rx_buf = efx_rx_buffer(rx_queue, index);
 150
 151                 rx_buf->u.skb = skb = netdev_alloc_skb(net_dev, skb_len);
 152                 if (unlikely(!skb))
 153                         return -ENOMEM;
 154
 155                 /* Adjust the SKB for padding */
 156                 skb_reserve(skb, NET_IP_ALIGN);
 157                 rx_buf->len = skb_len - NET_IP_ALIGN;
 158                 rx_buf->is_page = false;
 159
 160                 rx_buf->dma_addr = pci_map_single(efx->pci_dev,
 161                                                   skb->data, rx_buf->len,
 162                                                   PCI_DMA_FROMDEVICE);
 163                 if (unlikely(pci_dma_mapping_error(efx->pci_dev,
 164                                                    rx_buf->dma_addr))) {
 165                         dev_kfree_skb_any(skb);
 166                         rx_buf->u.skb = NULL;
 167                         return -EIO;
 168                 }
 169
 170                 ++rx_queue->added_count;
 171                 ++rx_queue->alloc_skb_count;
 172         }
 173
 174         return 0;
 175 }
 176
 177 /**
 178  * efx_init_rx_buffers_page - create EFX_RX_BATCH page-based RX buffers
 179  *
 180  * @rx_queue:           Efx RX queue
 181  *
 182  * This allocates memory for EFX_RX_BATCH receive buffers, maps them for DMA,
 183  * and populates struct efx_rx_buffers for each one. Return a negative error
 184  * code or 0 on success. If a single page can be split between two buffers,
 185  * then the page will either be inserted fully, or not at at all.
 186  */
 187 static int efx_init_rx_buffers_page(struct efx_rx_queue *rx_queue)
 188 {
 189         struct efx_nic *efx = rx_queue->efx;
 190         struct efx_rx_buffer *rx_buf;
 191         struct page *page;
 192         void *page_addr;
 193         unsigned int page_offset;
 194         struct efx_rx_page_state *state;
 195         dma_addr_t dma_addr;
 196         unsigned index, count;
 197
 198         /* We can split a page between two buffers */
 199         BUILD_BUG_ON(EFX_RX_BATCH & 1);
 200
 201         for (count = 0; count < EFX_RX_BATCH; ++count) {
 202                 page = alloc_pages(__GFP_COLD | __GFP_COMP | GFP_ATOMIC,
 203                                    efx->rx_buffer_order);
 204                 if (unlikely(page == NULL))
 205                         return -ENOMEM;
 206                 dma_addr = pci_map_page(efx->pci_dev, page, 0,
 207                                         efx_rx_buf_size(efx),
 208                                         PCI_DMA_FROMDEVICE);
 209                 if (unlikely(pci_dma_mapping_error(efx->pci_dev, dma_addr))) {
 210                         __free_pages(page, efx->rx_buffer_order);
 211                         return -EIO;
 212                 }
 213                 page_addr = page_address(page);
 214                 state = page_addr;
 215                 state->refcnt = 0;
 216                 state->dma_addr = dma_addr;
 217
 218                 page_addr += sizeof(struct efx_rx_page_state);
 219                 dma_addr += sizeof(struct efx_rx_page_state);
 220                 page_offset = sizeof(struct efx_rx_page_state);
 221
 222         split:
 223                 index = rx_queue->added_count & rx_queue->ptr_mask;
 224                 rx_buf = efx_rx_buffer(rx_queue, index);
 225                 rx_buf->dma_addr = dma_addr + EFX_PAGE_IP_ALIGN;
 226                 rx_buf->u.page = page;
 227                 rx_buf->page_offset = page_offset + EFX_PAGE_IP_ALIGN;
 228                 rx_buf->len = efx->rx_buffer_len - EFX_PAGE_IP_ALIGN;
 229                 rx_buf->is_page = true;
 230                 ++rx_queue->added_count;
 231                 ++rx_queue->alloc_page_count;
 232                 ++state->refcnt;
 233
 234                 if ((~count & 1) && (efx->rx_buffer_len <= EFX_RX_HALF_PAGE)) {
 235                         /* Use the second half of the page */
 236                         get_page(page);
 237                         dma_addr += (PAGE_SIZE >> 1);
 238                         page_addr += (PAGE_SIZE >> 1);
 239                         page_offset += (PAGE_SIZE >> 1);
 240                         ++count;
 241                         goto split;
 242                 }
 243         }
 244
 245         return 0;
 246 }
 247
 248 static void efx_unmap_rx_buffer(struct efx_nic *efx,
 249                                 struct efx_rx_buffer *rx_buf,
 250                                 unsigned int used_len)
 251 {
 252         if (rx_buf->is_page && rx_buf->u.page) {
 253                 struct efx_rx_page_state *state;
 254
 255                 state = page_address(rx_buf->u.page);
 256                 if (--state->refcnt == 0) {
 257                         pci_unmap_page(efx->pci_dev,
 258                                        state->dma_addr,
 259                                        efx_rx_buf_size(efx),
 260                                        PCI_DMA_FROMDEVICE);
 261                 } else if (used_len) {
 262                         dma_sync_single_for_cpu(&efx->pci_dev->dev,
 263                                                 rx_buf->dma_addr, used_len,
 264                                                 DMA_FROM_DEVICE);
 265                 }
 266         } else if (!rx_buf->is_page && rx_buf->u.skb) {
 267                 pci_unmap_single(efx->pci_dev, rx_buf->dma_addr,
 268                                  rx_buf->len, PCI_DMA_FROMDEVICE);
 269         }
 270 }
 271
 272 static void efx_free_rx_buffer(struct efx_nic *efx,
 273                                struct efx_rx_buffer *rx_buf)
 274 {
 275         if (rx_buf->is_page && rx_buf->u.page) {
 276                 __free_pages(rx_buf->u.page, efx->rx_buffer_order);
 277                 rx_buf->u.page = NULL;
 278         } else if (!rx_buf->is_page && rx_buf->u.skb) {
 279                 dev_kfree_skb_any(rx_buf->u.skb);
 280                 rx_buf->u.skb = NULL;
 281         }
 282 }
 283
 284 static void efx_fini_rx_buffer(struct efx_rx_queue *rx_queue,
 285                                struct efx_rx_buffer *rx_buf)
 286 {
 287         efx_unmap_rx_buffer(rx_queue->efx, rx_buf, 0);
 288         efx_free_rx_buffer(rx_queue->efx, rx_buf);
 289 }
 290
 291 /* Attempt to resurrect the other receive buffer that used to share this page,
 292  * which had previously been passed up to the kernel and freed. */
 293 static void efx_resurrect_rx_buffer(struct efx_rx_queue *rx_queue,
 294                                     struct efx_rx_buffer *rx_buf)
 295 {
 296         struct efx_rx_page_state *state = page_address(rx_buf->u.page);
 297         struct efx_rx_buffer *new_buf;
 298         unsigned fill_level, index;
 299
 300         /* +1 because efx_rx_packet() incremented removed_count. +1 because
 301          * we'd like to insert an additional descriptor whilst leaving
 302          * EFX_RXD_HEAD_ROOM for the non-recycle path */
 303         fill_level = (rx_queue->added_count - rx_queue->removed_count + 2);
 304         if (unlikely(fill_level > rx_queue->max_fill)) {
 305                 /* We could place "state" on a list, and drain the list in
 306                  * efx_fast_push_rx_descriptors(). For now, this will do. */
 307                 return;
 308         }
 309
 310         ++state->refcnt;
 311         get_page(rx_buf->u.page);
 312
 313         index = rx_queue->added_count & rx_queue->ptr_mask;
 314         new_buf = efx_rx_buffer(rx_queue, index);
 315         new_buf->dma_addr = rx_buf->dma_addr ^ (PAGE_SIZE >> 1);
 316         new_buf->u.page = rx_buf->u.page;
 317         new_buf->len = rx_buf->len;
 318         new_buf->is_page = true;
 319         ++rx_queue->added_count;
 320 }
 321
 322 /* Recycle the given rx buffer directly back into the rx_queue. There is
 323  * always room to add this buffer, because we've just popped a buffer. */
 324 static void efx_recycle_rx_buffer(struct efx_channel *channel,
 325                                   struct efx_rx_buffer *rx_buf)
 326 {
 327         struct efx_nic *efx = channel->efx;
 328         struct efx_rx_queue *rx_queue = efx_channel_get_rx_queue(channel);
 329         struct efx_rx_buffer *new_buf;
 330         unsigned index;
 331
 332         if (rx_buf->is_page && efx->rx_buffer_len <= EFX_RX_HALF_PAGE &&
 333             page_count(rx_buf->u.page) == 1)
 334                 efx_resurrect_rx_buffer(rx_queue, rx_buf);
 335
 336         index = rx_queue->added_count & rx_queue->ptr_mask;
 337         new_buf = efx_rx_buffer(rx_queue, index);
 338
 339         memcpy(new_buf, rx_buf, sizeof(*new_buf));
 340         rx_buf->u.page = NULL;
 341         ++rx_queue->added_count;
 342 }
 343
 344 /**
 345  * efx_fast_push_rx_descriptors - push new RX descriptors quickly
 346  * @rx_queue:           RX descriptor queue
 347  * This will aim to fill the RX descriptor queue up to
 348  * @rx_queue->@fast_fill_limit. If there is insufficient atomic
 349  * memory to do so, a slow fill will be scheduled.
 350  *
 351  * The caller must provide serialisation (none is used here). In practise,
 352  * this means this function must run from the NAPI handler, or be called
 353  * when NAPI is disabled.
 354  */
 355 void efx_fast_push_rx_descriptors(struct efx_rx_queue *rx_queue)
 356 {
 357         struct efx_channel *channel = efx_rx_queue_channel(rx_queue);
 358         unsigned fill_level;
 359         int space, rc = 0;
 360
 361         /* Calculate current fill level, and exit if we don't need to fill */
 362         fill_level = (rx_queue->added_count - rx_queue->removed_count);
 363         EFX_BUG_ON_PARANOID(fill_level > rx_queue->efx->rxq_entries);
 364         if (fill_level >= rx_queue->fast_fill_trigger)
 365                 goto out;
 366
 367         /* Record minimum fill level */
 368         if (unlikely(fill_level < rx_queue->min_fill)) {
 369                 if (fill_level)
 370                         rx_queue->min_fill = fill_level;
 371         }
 372
 373         space = rx_queue->fast_fill_limit - fill_level;
 374         if (space < EFX_RX_BATCH)
 375                 goto out;
 376
 377         netif_vdbg(rx_queue->efx, rx_status, rx_queue->efx->net_dev,
 378                    "RX queue %d fast-filling descriptor ring from"
 379                    " level %d to level %d using %s allocation\n",
 380                    efx_rx_queue_index(rx_queue), fill_level,
 381                    rx_queue->fast_fill_limit,
 382                    channel->rx_alloc_push_pages ? "page" : "skb");
 383
 384         do {
 385                 if (channel->rx_alloc_push_pages)
 386                         rc = efx_init_rx_buffers_page(rx_queue);
 387                 else
 388                         rc = efx_init_rx_buffers_skb(rx_queue);
 389                 if (unlikely(rc)) {
 390                         /* Ensure that we don't leave the rx queue empty */
 391                         if (rx_queue->added_count == rx_queue->removed_count)
 392                                 efx_schedule_slow_fill(rx_queue);
 393                         goto out;
 394                 }
 395         } while ((space -= EFX_RX_BATCH) >= EFX_RX_BATCH);
 396
 397         netif_vdbg(rx_queue->efx, rx_status, rx_queue->efx->net_dev,
 398                    "RX queue %d fast-filled descriptor ring "
 399                    "to level %d\n", efx_rx_queue_index(rx_queue),
 400                    rx_queue->added_count - rx_queue->removed_count);
 401
 402  out:
 403         if (rx_queue->notified_count != rx_queue->added_count)
 404                 efx_nic_notify_rx_desc(rx_queue);
 405 }
 406
 407 void efx_rx_slow_fill(unsigned long context)
 408 {
 409         struct efx_rx_queue *rx_queue = (struct efx_rx_queue *)context;
 410         struct efx_channel *channel = efx_rx_queue_channel(rx_queue);
 411
 412         /* Post an event to cause NAPI to run and refill the queue */
 413         efx_nic_generate_fill_event(channel);
 414         ++rx_queue->slow_fill_count;
 415 }
 416
 417 static void efx_rx_packet__check_len(struct efx_rx_queue *rx_queue,
 418                                      struct efx_rx_buffer *rx_buf,
 419                                      int len, bool *discard,
 420                                      bool *leak_packet)
 421 {
 422         struct efx_nic *efx = rx_queue->efx;
 423         unsigned max_len = rx_buf->len - efx->type->rx_buffer_padding;
 424
 425         if (likely(len <= max_len))
 426                 return;
 427
 428         /* The packet must be discarded, but this is only a fatal error
 429          * if the caller indicated it was
 430          */
 431         *discard = true;
 432
 433         if ((len > rx_buf->len) && EFX_WORKAROUND_8071(efx)) {
 434                 if (net_ratelimit())
 435                         netif_err(efx, rx_err, efx->net_dev,
 436                                   " RX queue %d seriously overlength "
 437                                   "RX event (0x%x > 0x%x+0x%x). Leaking\n",
 438                                   efx_rx_queue_index(rx_queue), len, max_len,
 439                                   efx->type->rx_buffer_padding);
 440                 /* If this buffer was skb-allocated, then the meta
 441                  * data at the end of the skb will be trashed. So
 442                  * we have no choice but to leak the fragment.
 443                  */
 444                 *leak_packet = !rx_buf->is_page;
 445                 efx_schedule_reset(efx, RESET_TYPE_RX_RECOVERY);
 446         } else {
 447                 if (net_ratelimit())
 448                         netif_err(efx, rx_err, efx->net_dev,
 449                                   " RX queue %d overlength RX event "
 450                                   "(0x%x > 0x%x)\n",
 451                                   efx_rx_queue_index(rx_queue), len, max_len);
 452         }
 453
 454         efx_rx_queue_channel(rx_queue)->n_rx_overlength++;
 455 }
 456
 457 /* Pass a received packet up through the generic GRO stack
 458  *
 459  * Handles driverlink veto, and passes the fragment up via
 460  * the appropriate GRO method
 461  */
 462 static void efx_rx_packet_gro(struct efx_channel *channel,
 463                               struct efx_rx_buffer *rx_buf,
 464                               const u8 *eh, bool checksummed)
 465 {
 466         struct napi_struct *napi = &channel->napi_str;
 467         gro_result_t gro_result;
 468
 469         /* Pass the skb/page into the GRO engine */
 470         if (rx_buf->is_page) {
 471                 struct efx_nic *efx = channel->efx;
 472                 struct page *page = rx_buf->u.page;
 473                 struct sk_buff *skb;
 474
 475                 rx_buf->u.page = NULL;
 476
 477                 skb = napi_get_frags(napi);
 478                 if (!skb) {
 479                         put_page(page);
 480                         return;
 481                 }
 482
 483                 if (efx->net_dev->features & NETIF_F_RXHASH)
 484                         skb->rxhash = efx_rx_buf_hash(eh);
 485
 486                 skb_frag_set_page(skb, 0, page);
 487                 skb_shinfo(skb)->frags[0].page_offset =
 488                         efx_rx_buf_offset(efx, rx_buf);
 489                 skb_frag_size_set(&skb_shinfo(skb)->frags[0], rx_buf->len);
 490                 skb_shinfo(skb)->nr_frags = 1;
 491
 492                 skb->len = rx_buf->len;
 493                 skb->data_len = rx_buf->len;
 494                 skb->truesize += rx_buf->len;
 495                 skb->ip_summed =
 496                         checksummed ? CHECKSUM_UNNECESSARY : CHECKSUM_NONE;
 497
 498                 skb_record_rx_queue(skb, channel->channel);
 499
 500                 gro_result = napi_gro_frags(napi);
 501         } else {
 502                 struct sk_buff *skb = rx_buf->u.skb;
 503
 504                 EFX_BUG_ON_PARANOID(!checksummed);
 505                 rx_buf->u.skb = NULL;
 506                 skb->ip_summed = CHECKSUM_UNNECESSARY;
 507
 508                 gro_result = napi_gro_receive(napi, skb);
 509         }
 510
 511         if (gro_result == GRO_NORMAL) {
 512                 channel->rx_alloc_level += RX_ALLOC_FACTOR_SKB;
 513         } else if (gro_result != GRO_DROP) {
 514                 channel->rx_alloc_level += RX_ALLOC_FACTOR_GRO;
 515                 channel->irq_mod_score += 2;
 516         }
 517 }
 518
 519 void efx_rx_packet(struct efx_rx_queue *rx_queue, unsigned int index,
 520                    unsigned int len, bool checksummed, bool discard)
 521 {
 522         struct efx_nic *efx = rx_queue->efx;
 523         struct efx_channel *channel = efx_rx_queue_channel(rx_queue);
 524         struct efx_rx_buffer *rx_buf;
 525         bool leak_packet = false;
 526
 527         rx_buf = efx_rx_buffer(rx_queue, index);
 528
 529         /* This allows the refill path to post another buffer.
 530          * EFX_RXD_HEAD_ROOM ensures that the slot we are using
 531          * isn't overwritten yet.
 532          */
 533         rx_queue->removed_count++;
 534
 535         /* Validate the length encoded in the event vs the descriptor pushed */
 536         efx_rx_packet__check_len(rx_queue, rx_buf, len,
 537                                  &discard, &leak_packet);
 538
 539         netif_vdbg(efx, rx_status, efx->net_dev,
 540                    "RX queue %d received id %x at %llx+%x %s%s\n",
 541                    efx_rx_queue_index(rx_queue), index,
 542                    (unsigned long long)rx_buf->dma_addr, len,
 543                    (checksummed ? " [SUMMED]" : ""),
 544                    (discard ? " [DISCARD]" : ""));
 545
 546         /* Discard packet, if instructed to do so */
 547         if (unlikely(discard)) {
 548                 if (unlikely(leak_packet))
 549                         channel->n_skbuff_leaks++;
 550                 else
 551                         efx_recycle_rx_buffer(channel, rx_buf);
 552
 553                 /* Don't hold off the previous receive */
 554                 rx_buf = NULL;
 555                 goto out;
 556         }
 557
 558         /* Release and/or sync DMA mapping - assumes all RX buffers
 559          * consumed in-order per RX queue
 560          */
 561         efx_unmap_rx_buffer(efx, rx_buf, len);
 562
 563         /* Prefetch nice and early so data will (hopefully) be in cache by
 564          * the time we look at it.
 565          */
 566         prefetch(efx_rx_buf_eh(efx, rx_buf));
 567
 568         /* Pipeline receives so that we give time for packet headers to be
 569          * prefetched into cache.
 570          */
 571         rx_buf->len = len - efx->type->rx_buffer_hash_size;
 572 out:
 573         if (channel->rx_pkt)
 574                 __efx_rx_packet(channel,
 575                                 channel->rx_pkt, channel->rx_pkt_csummed);
 576         channel->rx_pkt = rx_buf;
 577         channel->rx_pkt_csummed = checksummed;
 578 }
 579
 580 /* Handle a received packet.  Second half: Touches packet payload. */
 581 void __efx_rx_packet(struct efx_channel *channel,
 582                      struct efx_rx_buffer *rx_buf, bool checksummed)
 583 {
 584         struct efx_nic *efx = channel->efx;
 585         struct sk_buff *skb;
 586         u8 *eh = efx_rx_buf_eh(efx, rx_buf);
 587
 588         /* If we're in loopback test, then pass the packet directly to the
 589          * loopback layer, and free the rx_buf here
 590          */
 591         if (unlikely(efx->loopback_selftest)) {
 592                 efx_loopback_rx_packet(efx, eh, rx_buf->len);
 593                 efx_free_rx_buffer(efx, rx_buf);
 594                 return;
 595         }
 596
 597         if (!rx_buf->is_page) {
 598                 skb = rx_buf->u.skb;
 599
 600                 prefetch(skb_shinfo(skb));
 601
 602                 skb_reserve(skb, efx->type->rx_buffer_hash_size);
 603                 skb_put(skb, rx_buf->len);
 604
 605                 if (efx->net_dev->features & NETIF_F_RXHASH)
 606                         skb->rxhash = efx_rx_buf_hash(eh);
 607
 608                 /* Move past the ethernet header. rx_buf->data still points
 609                  * at the ethernet header */
 610                 skb->protocol = eth_type_trans(skb, efx->net_dev);
 611
 612                 skb_record_rx_queue(skb, channel->channel);
 613         }
 614
 615         if (unlikely(!(efx->net_dev->features & NETIF_F_RXCSUM)))
 616                 checksummed = false;
 617
 618         if (likely(checksummed || rx_buf->is_page)) {
 619                 efx_rx_packet_gro(channel, rx_buf, eh, checksummed);
 620                 return;
 621         }
 622
 623         /* We now own the SKB */
 624         skb = rx_buf->u.skb;
 625         rx_buf->u.skb = NULL;
 626
 627         /* Set the SKB flags */
 628         skb_checksum_none_assert(skb);
 629
 630         /* Pass the packet up */
 631         netif_receive_skb(skb);
 632
 633         /* Update allocation strategy method */
 634         channel->rx_alloc_level += RX_ALLOC_FACTOR_SKB;
 635 }
 636
 637 void efx_rx_strategy(struct efx_channel *channel)
 638 {
 639         enum efx_rx_alloc_method method = rx_alloc_method;
 640
 641         /* Only makes sense to use page based allocation if GRO is enabled */
 642         if (!(channel->efx->net_dev->features & NETIF_F_GRO)) {
 643                 method = RX_ALLOC_METHOD_SKB;
 644         } else if (method == RX_ALLOC_METHOD_AUTO) {
 645                 /* Constrain the rx_alloc_level */
 646                 if (channel->rx_alloc_level < 0)
 647                         channel->rx_alloc_level = 0;
 648                 else if (channel->rx_alloc_level > RX_ALLOC_LEVEL_MAX)
 649                         channel->rx_alloc_level = RX_ALLOC_LEVEL_MAX;
 650
 651                 /* Decide on the allocation method */
 652                 method = ((channel->rx_alloc_level > RX_ALLOC_LEVEL_GRO) ?
 653                           RX_ALLOC_METHOD_PAGE : RX_ALLOC_METHOD_SKB);
 654         }
 655
 656         /* Push the option */
 657         channel->rx_alloc_push_pages = (method == RX_ALLOC_METHOD_PAGE);
 658 }
 659
 660 int efx_probe_rx_queue(struct efx_rx_queue *rx_queue)
 661 {
 662         struct efx_nic *efx = rx_queue->efx;
 663         unsigned int entries;
 664         int rc;
 665
 666         /* Create the smallest power-of-two aligned ring */
 667         entries = max(roundup_pow_of_two(efx->rxq_entries), EFX_MIN_DMAQ_SIZE);
 668         EFX_BUG_ON_PARANOID(entries > EFX_MAX_DMAQ_SIZE);
 669         rx_queue->ptr_mask = entries - 1;
 670
 671         netif_dbg(efx, probe, efx->net_dev,
 672                   "creating RX queue %d size %#x mask %#x\n",
 673                   efx_rx_queue_index(rx_queue), efx->rxq_entries,
 674                   rx_queue->ptr_mask);
 675
 676         /* Allocate RX buffers */
 677         rx_queue->buffer = kzalloc(entries * sizeof(*rx_queue->buffer),
 678                                    GFP_KERNEL);
 679         if (!rx_queue->buffer)
 680                 return -ENOMEM;
 681
 682         rc = efx_nic_probe_rx(rx_queue);
 683         if (rc) {
 684                 kfree(rx_queue->buffer);
 685                 rx_queue->buffer = NULL;
 686         }
 687         return rc;
 688 }
 689
 690 void efx_init_rx_queue(struct efx_rx_queue *rx_queue)
 691 {
 692         struct efx_nic *efx = rx_queue->efx;
 693         unsigned int max_fill, trigger, limit;
 694
 695         netif_dbg(rx_queue->efx, drv, rx_queue->efx->net_dev,
 696                   "initialising RX queue %d\n", efx_rx_queue_index(rx_queue));
 697
 698         /* Initialise ptr fields */
 699         rx_queue->added_count = 0;
 700         rx_queue->notified_count = 0;
 701         rx_queue->removed_count = 0;
 702         rx_queue->min_fill = -1U;
 703
 704         /* Initialise limit fields */
 705         max_fill = efx->rxq_entries - EFX_RXD_HEAD_ROOM;
 706         trigger = max_fill * min(rx_refill_threshold, 100U) / 100U;
 707         limit = max_fill * min(rx_refill_limit, 100U) / 100U;
 708
 709         rx_queue->max_fill = max_fill;
 710         rx_queue->fast_fill_trigger = trigger;
 711         rx_queue->fast_fill_limit = limit;
 712
 713         /* Set up RX descriptor ring */
 714         efx_nic_init_rx(rx_queue);
 715 }
 716
 717 void efx_fini_rx_queue(struct efx_rx_queue *rx_queue)
 718 {
 719         int i;
 720         struct efx_rx_buffer *rx_buf;
 721
 722         netif_dbg(rx_queue->efx, drv, rx_queue->efx->net_dev,
 723                   "shutting down RX queue %d\n", efx_rx_queue_index(rx_queue));
 724
 725         del_timer_sync(&rx_queue->slow_fill);
 726         efx_nic_fini_rx(rx_queue);
 727
 728         /* Release RX buffers NB start at index 0 not current HW ptr */
 729         if (rx_queue->buffer) {
 730                 for (i = 0; i <= rx_queue->ptr_mask; i++) {
 731                         rx_buf = efx_rx_buffer(rx_queue, i);
 732                         efx_fini_rx_buffer(rx_queue, rx_buf);
 733                 }
 734         }
 735 }
 736
 737 void efx_remove_rx_queue(struct efx_rx_queue *rx_queue)
 738 {
 739         netif_dbg(rx_queue->efx, drv, rx_queue->efx->net_dev,
 740                   "destroying RX queue %d\n", efx_rx_queue_index(rx_queue));
 741
 742         efx_nic_remove_rx(rx_queue);
 743
 744         kfree(rx_queue->buffer);
 745         rx_queue->buffer = NULL;
 746 }
 747
 748
 749 module_param(rx_alloc_method, int, 0644);
 750 MODULE_PARM_DESC(rx_alloc_method, "Allocation method used for RX buffers");
 751
 752 module_param(rx_refill_threshold, uint, 0444);
 753 MODULE_PARM_DESC(rx_refill_threshold,
 754                  "RX descriptor ring fast/slow fill threshold (%)");
 755