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         /* Offset is always within one page, so we don't need to consider
  99          * the page order.
 100          */
 101         return ((unsigned int) buf->dma_addr & (PAGE_SIZE - 1)) +
 102                 efx->type->rx_buffer_hash_size;
 103 }
 104 static inline unsigned int efx_rx_buf_size(struct efx_nic *efx)
 105 {
 106         return PAGE_SIZE << efx->rx_buffer_order;
 107 }
 108
 109 static u8 *efx_rx_buf_eh(struct efx_nic *efx, struct efx_rx_buffer *buf)
 110 {
 111         if (buf->flags & EFX_RX_BUF_PAGE)
 112                 return page_address(buf->u.page) + efx_rx_buf_offset(efx, buf);
 113         else
 114                 return (u8 *)buf->u.skb->data + efx->type->rx_buffer_hash_size;
 115 }
 116
 117 static inline u32 efx_rx_buf_hash(const u8 *eh)
 118 {
 119         /* The ethernet header is always directly after any hash. */
 120 #if defined(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS) || NET_IP_ALIGN % 4 == 0
 121         return __le32_to_cpup((const __le32 *)(eh - 4));
 122 #else
 123         const u8 *data = eh - 4;
 124         return (u32)data[0]       |
 125                (u32)data[1] << 8  |
 126                (u32)data[2] << 16 |
 127                (u32)data[3] << 24;
 128 #endif
 129 }
 130
 131 /**
 132  * efx_init_rx_buffers_skb - create EFX_RX_BATCH skb-based RX buffers
 133  *
 134  * @rx_queue:           Efx RX queue
 135  *
 136  * This allocates EFX_RX_BATCH skbs, maps them for DMA, and populates a
 137  * struct efx_rx_buffer for each one. Return a negative error code or 0
 138  * on success. May fail having only inserted fewer than EFX_RX_BATCH
 139  * buffers.
 140  */
 141 static int efx_init_rx_buffers_skb(struct efx_rx_queue *rx_queue)
 142 {
 143         struct efx_nic *efx = rx_queue->efx;
 144         struct net_device *net_dev = efx->net_dev;
 145         struct efx_rx_buffer *rx_buf;
 146         struct sk_buff *skb;
 147         int skb_len = efx->rx_buffer_len;
 148         unsigned index, count;
 149
 150         for (count = 0; count < EFX_RX_BATCH; ++count) {
 151                 index = rx_queue->added_count & rx_queue->ptr_mask;
 152                 rx_buf = efx_rx_buffer(rx_queue, index);
 153
 154                 rx_buf->u.skb = skb = netdev_alloc_skb(net_dev, skb_len);
 155                 if (unlikely(!skb))
 156                         return -ENOMEM;
 157
 158                 /* Adjust the SKB for padding */
 159                 skb_reserve(skb, NET_IP_ALIGN);
 160                 rx_buf->len = skb_len - NET_IP_ALIGN;
 161                 rx_buf->flags = 0;
 162
 163                 rx_buf->dma_addr = pci_map_single(efx->pci_dev,
 164                                                   skb->data, rx_buf->len,
 165                                                   PCI_DMA_FROMDEVICE);
 166                 if (unlikely(pci_dma_mapping_error(efx->pci_dev,
 167                                                    rx_buf->dma_addr))) {
 168                         dev_kfree_skb_any(skb);
 169                         rx_buf->u.skb = NULL;
 170                         return -EIO;
 171                 }
 172
 173                 ++rx_queue->added_count;
 174                 ++rx_queue->alloc_skb_count;
 175         }
 176
 177         return 0;
 178 }
 179
 180 /**
 181  * efx_init_rx_buffers_page - create EFX_RX_BATCH page-based RX buffers
 182  *
 183  * @rx_queue:           Efx RX queue
 184  *
 185  * This allocates memory for EFX_RX_BATCH receive buffers, maps them for DMA,
 186  * and populates struct efx_rx_buffers for each one. Return a negative error
 187  * code or 0 on success. If a single page can be split between two buffers,
 188  * then the page will either be inserted fully, or not at at all.
 189  */
 190 static int efx_init_rx_buffers_page(struct efx_rx_queue *rx_queue)
 191 {
 192         struct efx_nic *efx = rx_queue->efx;
 193         struct efx_rx_buffer *rx_buf;
 194         struct page *page;
 195         void *page_addr;
 196         struct efx_rx_page_state *state;
 197         dma_addr_t dma_addr;
 198         unsigned index, count;
 199
 200         /* We can split a page between two buffers */
 201         BUILD_BUG_ON(EFX_RX_BATCH & 1);
 202
 203         for (count = 0; count < EFX_RX_BATCH; ++count) {
 204                 page = alloc_pages(__GFP_COLD | __GFP_COMP | GFP_ATOMIC,
 205                                    efx->rx_buffer_order);
 206                 if (unlikely(page == NULL))
 207                         return -ENOMEM;
 208                 dma_addr = pci_map_page(efx->pci_dev, page, 0,
 209                                         efx_rx_buf_size(efx),
 210                                         PCI_DMA_FROMDEVICE);
 211                 if (unlikely(pci_dma_mapping_error(efx->pci_dev, dma_addr))) {
 212                         __free_pages(page, efx->rx_buffer_order);
 213                         return -EIO;
 214                 }
 215                 page_addr = page_address(page);
 216                 state = page_addr;
 217                 state->refcnt = 0;
 218                 state->dma_addr = dma_addr;
 219
 220                 page_addr += sizeof(struct efx_rx_page_state);
 221                 dma_addr += sizeof(struct efx_rx_page_state);
 222
 223         split:
 224                 index = rx_queue->added_count & rx_queue->ptr_mask;
 225                 rx_buf = efx_rx_buffer(rx_queue, index);
 226                 rx_buf->dma_addr = dma_addr + EFX_PAGE_IP_ALIGN;
 227                 rx_buf->u.page = page;
 228                 rx_buf->len = efx->rx_buffer_len - EFX_PAGE_IP_ALIGN;
 229                 rx_buf->flags = EFX_RX_BUF_PAGE;
 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                         ++count;
 240                         goto split;
 241                 }
 242         }
 243
 244         return 0;
 245 }
 246
 247 static void efx_unmap_rx_buffer(struct efx_nic *efx,
 248                                 struct efx_rx_buffer *rx_buf)
 249 {
 250         if ((rx_buf->flags & EFX_RX_BUF_PAGE) && rx_buf->u.page) {
 251                 struct efx_rx_page_state *state;
 252
 253                 state = page_address(rx_buf->u.page);
 254                 if (--state->refcnt == 0) {
 255                         pci_unmap_page(efx->pci_dev,
 256                                        state->dma_addr,
 257                                        efx_rx_buf_size(efx),
 258                                        PCI_DMA_FROMDEVICE);
 259                 }
 260         } else if (!(rx_buf->flags & EFX_RX_BUF_PAGE) && rx_buf->u.skb) {
 261                 pci_unmap_single(efx->pci_dev, rx_buf->dma_addr,
 262                                  rx_buf->len, PCI_DMA_FROMDEVICE);
 263         }
 264 }
 265
 266 static void efx_free_rx_buffer(struct efx_nic *efx,
 267                                struct efx_rx_buffer *rx_buf)
 268 {
 269         if ((rx_buf->flags & EFX_RX_BUF_PAGE) && rx_buf->u.page) {
 270                 __free_pages(rx_buf->u.page, efx->rx_buffer_order);
 271                 rx_buf->u.page = NULL;
 272         } else if (!(rx_buf->flags & EFX_RX_BUF_PAGE) && rx_buf->u.skb) {
 273                 dev_kfree_skb_any(rx_buf->u.skb);
 274                 rx_buf->u.skb = NULL;
 275         }
 276 }
 277
 278 static void efx_fini_rx_buffer(struct efx_rx_queue *rx_queue,
 279                                struct efx_rx_buffer *rx_buf)
 280 {
 281         efx_unmap_rx_buffer(rx_queue->efx, rx_buf);
 282         efx_free_rx_buffer(rx_queue->efx, rx_buf);
 283 }
 284
 285 /* Attempt to resurrect the other receive buffer that used to share this page,
 286  * which had previously been passed up to the kernel and freed. */
 287 static void efx_resurrect_rx_buffer(struct efx_rx_queue *rx_queue,
 288                                     struct efx_rx_buffer *rx_buf)
 289 {
 290         struct efx_rx_page_state *state = page_address(rx_buf->u.page);
 291         struct efx_rx_buffer *new_buf;
 292         unsigned fill_level, index;
 293
 294         /* +1 because efx_rx_packet() incremented removed_count. +1 because
 295          * we'd like to insert an additional descriptor whilst leaving
 296          * EFX_RXD_HEAD_ROOM for the non-recycle path */
 297         fill_level = (rx_queue->added_count - rx_queue->removed_count + 2);
 298         if (unlikely(fill_level > rx_queue->max_fill)) {
 299                 /* We could place "state" on a list, and drain the list in
 300                  * efx_fast_push_rx_descriptors(). For now, this will do. */
 301                 return;
 302         }
 303
 304         ++state->refcnt;
 305         get_page(rx_buf->u.page);
 306
 307         index = rx_queue->added_count & rx_queue->ptr_mask;
 308         new_buf = efx_rx_buffer(rx_queue, index);
 309         new_buf->dma_addr = rx_buf->dma_addr ^ (PAGE_SIZE >> 1);
 310         new_buf->u.page = rx_buf->u.page;
 311         new_buf->len = rx_buf->len;
 312         new_buf->flags = EFX_RX_BUF_PAGE;
 313         ++rx_queue->added_count;
 314 }
 315
 316 /* Recycle the given rx buffer directly back into the rx_queue. There is
 317  * always room to add this buffer, because we've just popped a buffer. */
 318 static void efx_recycle_rx_buffer(struct efx_channel *channel,
 319                                   struct efx_rx_buffer *rx_buf)
 320 {
 321         struct efx_nic *efx = channel->efx;
 322         struct efx_rx_queue *rx_queue = efx_channel_get_rx_queue(channel);
 323         struct efx_rx_buffer *new_buf;
 324         unsigned index;
 325
 326         rx_buf->flags &= EFX_RX_BUF_PAGE;
 327
 328         if ((rx_buf->flags & EFX_RX_BUF_PAGE) &&
 329             efx->rx_buffer_len <= EFX_RX_HALF_PAGE &&
 330             page_count(rx_buf->u.page) == 1)
 331                 efx_resurrect_rx_buffer(rx_queue, rx_buf);
 332
 333         index = rx_queue->added_count & rx_queue->ptr_mask;
 334         new_buf = efx_rx_buffer(rx_queue, index);
 335
 336         memcpy(new_buf, rx_buf, sizeof(*new_buf));
 337         rx_buf->u.page = NULL;
 338         ++rx_queue->added_count;
 339 }
 340
 341 /**
 342  * efx_fast_push_rx_descriptors - push new RX descriptors quickly
 343  * @rx_queue:           RX descriptor queue
 344  * This will aim to fill the RX descriptor queue up to
 345  * @rx_queue->@fast_fill_limit. If there is insufficient atomic
 346  * memory to do so, a slow fill will be scheduled.
 347  *
 348  * The caller must provide serialisation (none is used here). In practise,
 349  * this means this function must run from the NAPI handler, or be called
 350  * when NAPI is disabled.
 351  */
 352 void efx_fast_push_rx_descriptors(struct efx_rx_queue *rx_queue)
 353 {
 354         struct efx_channel *channel = efx_rx_queue_channel(rx_queue);
 355         unsigned fill_level;
 356         int space, rc = 0;
 357
 358         /* Calculate current fill level, and exit if we don't need to fill */
 359         fill_level = (rx_queue->added_count - rx_queue->removed_count);
 360         EFX_BUG_ON_PARANOID(fill_level > rx_queue->efx->rxq_entries);
 361         if (fill_level >= rx_queue->fast_fill_trigger)
 362                 goto out;
 363
 364         /* Record minimum fill level */
 365         if (unlikely(fill_level < rx_queue->min_fill)) {
 366                 if (fill_level)
 367                         rx_queue->min_fill = fill_level;
 368         }
 369
 370         space = rx_queue->fast_fill_limit - fill_level;
 371         if (space < EFX_RX_BATCH)
 372                 goto out;
 373
 374         netif_vdbg(rx_queue->efx, rx_status, rx_queue->efx->net_dev,
 375                    "RX queue %d fast-filling descriptor ring from"
 376                    " level %d to level %d using %s allocation\n",
 377                    efx_rx_queue_index(rx_queue), fill_level,
 378                    rx_queue->fast_fill_limit,
 379                    channel->rx_alloc_push_pages ? "page" : "skb");
 380
 381         do {
 382                 if (channel->rx_alloc_push_pages)
 383                         rc = efx_init_rx_buffers_page(rx_queue);
 384                 else
 385                         rc = efx_init_rx_buffers_skb(rx_queue);
 386                 if (unlikely(rc)) {
 387                         /* Ensure that we don't leave the rx queue empty */
 388                         if (rx_queue->added_count == rx_queue->removed_count)
 389                                 efx_schedule_slow_fill(rx_queue);
 390                         goto out;
 391                 }
 392         } while ((space -= EFX_RX_BATCH) >= EFX_RX_BATCH);
 393
 394         netif_vdbg(rx_queue->efx, rx_status, rx_queue->efx->net_dev,
 395                    "RX queue %d fast-filled descriptor ring "
 396                    "to level %d\n", efx_rx_queue_index(rx_queue),
 397                    rx_queue->added_count - rx_queue->removed_count);
 398
 399  out:
 400         if (rx_queue->notified_count != rx_queue->added_count)
 401                 efx_nic_notify_rx_desc(rx_queue);
 402 }
 403
 404 void efx_rx_slow_fill(unsigned long context)
 405 {
 406         struct efx_rx_queue *rx_queue = (struct efx_rx_queue *)context;
 407
 408         /* Post an event to cause NAPI to run and refill the queue */
 409         efx_nic_generate_fill_event(rx_queue);
 410         ++rx_queue->slow_fill_count;
 411 }
 412
 413 static void efx_rx_packet__check_len(struct efx_rx_queue *rx_queue,
 414                                      struct efx_rx_buffer *rx_buf,
 415                                      int len, bool *leak_packet)
 416 {
 417         struct efx_nic *efx = rx_queue->efx;
 418         unsigned max_len = rx_buf->len - efx->type->rx_buffer_padding;
 419
 420         if (likely(len <= max_len))
 421                 return;
 422
 423         /* The packet must be discarded, but this is only a fatal error
 424          * if the caller indicated it was
 425          */
 426         rx_buf->flags |= EFX_RX_PKT_DISCARD;
 427
 428         if ((len > rx_buf->len) && EFX_WORKAROUND_8071(efx)) {
 429                 if (net_ratelimit())
 430                         netif_err(efx, rx_err, efx->net_dev,
 431                                   " RX queue %d seriously overlength "
 432                                   "RX event (0x%x > 0x%x+0x%x). Leaking\n",
 433                                   efx_rx_queue_index(rx_queue), len, max_len,
 434                                   efx->type->rx_buffer_padding);
 435                 /* If this buffer was skb-allocated, then the meta
 436                  * data at the end of the skb will be trashed. So
 437                  * we have no choice but to leak the fragment.
 438                  */
 439                 *leak_packet = !(rx_buf->flags & EFX_RX_BUF_PAGE);
 440                 efx_schedule_reset(efx, RESET_TYPE_RX_RECOVERY);
 441         } else {
 442                 if (net_ratelimit())
 443                         netif_err(efx, rx_err, efx->net_dev,
 444                                   " RX queue %d overlength RX event "
 445                                   "(0x%x > 0x%x)\n",
 446                                   efx_rx_queue_index(rx_queue), len, max_len);
 447         }
 448
 449         efx_rx_queue_channel(rx_queue)->n_rx_overlength++;
 450 }
 451
 452 /* Pass a received packet up through the generic GRO stack
 453  *
 454  * Handles driverlink veto, and passes the fragment up via
 455  * the appropriate GRO method
 456  */
 457 static void efx_rx_packet_gro(struct efx_channel *channel,
 458                               struct efx_rx_buffer *rx_buf,
 459                               const u8 *eh)
 460 {
 461         struct napi_struct *napi = &channel->napi_str;
 462         gro_result_t gro_result;
 463
 464         /* Pass the skb/page into the GRO engine */
 465         if (rx_buf->flags & EFX_RX_BUF_PAGE) {
 466                 struct efx_nic *efx = channel->efx;
 467                 struct page *page = rx_buf->u.page;
 468                 struct sk_buff *skb;
 469
 470                 rx_buf->u.page = NULL;
 471
 472                 skb = napi_get_frags(napi);
 473                 if (!skb) {
 474                         put_page(page);
 475                         return;
 476                 }
 477
 478                 if (efx->net_dev->features & NETIF_F_RXHASH)
 479                         skb->rxhash = efx_rx_buf_hash(eh);
 480
 481                 skb_fill_page_desc(skb, 0, page,
 482                                    efx_rx_buf_offset(efx, rx_buf), rx_buf->len);
 483
 484                 skb->len = rx_buf->len;
 485                 skb->data_len = rx_buf->len;
 486                 skb->truesize += rx_buf->len;
 487                 skb->ip_summed = ((rx_buf->flags & EFX_RX_PKT_CSUMMED) ?
 488                                   CHECKSUM_UNNECESSARY : CHECKSUM_NONE);
 489
 490                 skb_record_rx_queue(skb, channel->channel);
 491
 492                 gro_result = napi_gro_frags(napi);
 493         } else {
 494                 struct sk_buff *skb = rx_buf->u.skb;
 495
 496                 EFX_BUG_ON_PARANOID(!(rx_buf->flags & EFX_RX_PKT_CSUMMED));
 497                 rx_buf->u.skb = NULL;
 498                 skb->ip_summed = CHECKSUM_UNNECESSARY;
 499
 500                 gro_result = napi_gro_receive(napi, skb);
 501         }
 502
 503         if (gro_result == GRO_NORMAL) {
 504                 channel->rx_alloc_level += RX_ALLOC_FACTOR_SKB;
 505         } else if (gro_result != GRO_DROP) {
 506                 channel->rx_alloc_level += RX_ALLOC_FACTOR_GRO;
 507                 channel->irq_mod_score += 2;
 508         }
 509 }
 510
 511 void efx_rx_packet(struct efx_rx_queue *rx_queue, unsigned int index,
 512                    unsigned int len, u16 flags)
 513 {
 514         struct efx_nic *efx = rx_queue->efx;
 515         struct efx_channel *channel = efx_rx_queue_channel(rx_queue);
 516         struct efx_rx_buffer *rx_buf;
 517         bool leak_packet = false;
 518
 519         rx_buf = efx_rx_buffer(rx_queue, index);
 520         rx_buf->flags |= flags;
 521
 522         /* This allows the refill path to post another buffer.
 523          * EFX_RXD_HEAD_ROOM ensures that the slot we are using
 524          * isn't overwritten yet.
 525          */
 526         rx_queue->removed_count++;
 527
 528         /* Validate the length encoded in the event vs the descriptor pushed */
 529         efx_rx_packet__check_len(rx_queue, rx_buf, len, &leak_packet);
 530
 531         netif_vdbg(efx, rx_status, efx->net_dev,
 532                    "RX queue %d received id %x at %llx+%x %s%s\n",
 533                    efx_rx_queue_index(rx_queue), index,
 534                    (unsigned long long)rx_buf->dma_addr, len,
 535                    (rx_buf->flags & EFX_RX_PKT_CSUMMED) ? " [SUMMED]" : "",
 536                    (rx_buf->flags & EFX_RX_PKT_DISCARD) ? " [DISCARD]" : "");
 537
 538         /* Discard packet, if instructed to do so */
 539         if (unlikely(rx_buf->flags & EFX_RX_PKT_DISCARD)) {
 540                 if (unlikely(leak_packet))
 541                         channel->n_skbuff_leaks++;
 542                 else
 543                         efx_recycle_rx_buffer(channel, rx_buf);
 544
 545                 /* Don't hold off the previous receive */
 546                 rx_buf = NULL;
 547                 goto out;
 548         }
 549
 550         /* Release card resources - assumes all RX buffers consumed in-order
 551          * per RX queue
 552          */
 553         efx_unmap_rx_buffer(efx, rx_buf);
 554
 555         /* Prefetch nice and early so data will (hopefully) be in cache by
 556          * the time we look at it.
 557          */
 558         prefetch(efx_rx_buf_eh(efx, rx_buf));
 559
 560         /* Pipeline receives so that we give time for packet headers to be
 561          * prefetched into cache.
 562          */
 563         rx_buf->len = len - efx->type->rx_buffer_hash_size;
 564 out:
 565         if (channel->rx_pkt)
 566                 __efx_rx_packet(channel, channel->rx_pkt);
 567         channel->rx_pkt = rx_buf;
 568 }
 569
 570 static void efx_rx_deliver(struct efx_channel *channel,
 571                            struct efx_rx_buffer *rx_buf)
 572 {
 573         struct sk_buff *skb;
 574
 575         /* We now own the SKB */
 576         skb = rx_buf->u.skb;
 577         rx_buf->u.skb = NULL;
 578
 579         /* Set the SKB flags */
 580         skb_checksum_none_assert(skb);
 581
 582         /* Pass the packet up */
 583         netif_receive_skb(skb);
 584
 585         /* Update allocation strategy method */
 586         channel->rx_alloc_level += RX_ALLOC_FACTOR_SKB;
 587 }
 588
 589 /* Handle a received packet.  Second half: Touches packet payload. */
 590 void __efx_rx_packet(struct efx_channel *channel, struct efx_rx_buffer *rx_buf)
 591 {
 592         struct efx_nic *efx = channel->efx;
 593         u8 *eh = efx_rx_buf_eh(efx, rx_buf);
 594
 595         /* If we're in loopback test, then pass the packet directly to the
 596          * loopback layer, and free the rx_buf here
 597          */
 598         if (unlikely(efx->loopback_selftest)) {
 599                 efx_loopback_rx_packet(efx, eh, rx_buf->len);
 600                 efx_free_rx_buffer(efx, rx_buf);
 601                 return;
 602         }
 603
 604         if (!(rx_buf->flags & EFX_RX_BUF_PAGE)) {
 605                 struct sk_buff *skb = rx_buf->u.skb;
 606
 607                 prefetch(skb_shinfo(skb));
 608
 609                 skb_reserve(skb, efx->type->rx_buffer_hash_size);
 610                 skb_put(skb, rx_buf->len);
 611
 612                 if (efx->net_dev->features & NETIF_F_RXHASH)
 613                         skb->rxhash = efx_rx_buf_hash(eh);
 614
 615                 /* Move past the ethernet header. rx_buf->data still points
 616                  * at the ethernet header */
 617                 skb->protocol = eth_type_trans(skb, efx->net_dev);
 618
 619                 skb_record_rx_queue(skb, channel->channel);
 620         }
 621
 622         if (unlikely(!(efx->net_dev->features & NETIF_F_RXCSUM)))
 623                 rx_buf->flags &= ~EFX_RX_PKT_CSUMMED;
 624
 625         if (likely(rx_buf->flags & (EFX_RX_BUF_PAGE | EFX_RX_PKT_CSUMMED)))
 626                 efx_rx_packet_gro(channel, rx_buf, eh);
 627         else
 628                 efx_rx_deliver(channel, rx_buf);
 629 }
 630
 631 void efx_rx_strategy(struct efx_channel *channel)
 632 {
 633         enum efx_rx_alloc_method method = rx_alloc_method;
 634
 635         /* Only makes sense to use page based allocation if GRO is enabled */
 636         if (!(channel->efx->net_dev->features & NETIF_F_GRO)) {
 637                 method = RX_ALLOC_METHOD_SKB;
 638         } else if (method == RX_ALLOC_METHOD_AUTO) {
 639                 /* Constrain the rx_alloc_level */
 640                 if (channel->rx_alloc_level < 0)
 641                         channel->rx_alloc_level = 0;
 642                 else if (channel->rx_alloc_level > RX_ALLOC_LEVEL_MAX)
 643                         channel->rx_alloc_level = RX_ALLOC_LEVEL_MAX;
 644
 645                 /* Decide on the allocation method */
 646                 method = ((channel->rx_alloc_level > RX_ALLOC_LEVEL_GRO) ?
 647                           RX_ALLOC_METHOD_PAGE : RX_ALLOC_METHOD_SKB);
 648         }
 649
 650         /* Push the option */
 651         channel->rx_alloc_push_pages = (method == RX_ALLOC_METHOD_PAGE);
 652 }
 653
 654 int efx_probe_rx_queue(struct efx_rx_queue *rx_queue)
 655 {
 656         struct efx_nic *efx = rx_queue->efx;
 657         unsigned int entries;
 658         int rc;
 659
 660         /* Create the smallest power-of-two aligned ring */
 661         entries = max(roundup_pow_of_two(efx->rxq_entries), EFX_MIN_DMAQ_SIZE);
 662         EFX_BUG_ON_PARANOID(entries > EFX_MAX_DMAQ_SIZE);
 663         rx_queue->ptr_mask = entries - 1;
 664
 665         netif_dbg(efx, probe, efx->net_dev,
 666                   "creating RX queue %d size %#x mask %#x\n",
 667                   efx_rx_queue_index(rx_queue), efx->rxq_entries,
 668                   rx_queue->ptr_mask);
 669
 670         /* Allocate RX buffers */
 671         rx_queue->buffer = kcalloc(entries, sizeof(*rx_queue->buffer),
 672                                    GFP_KERNEL);
 673         if (!rx_queue->buffer)
 674                 return -ENOMEM;
 675
 676         rc = efx_nic_probe_rx(rx_queue);
 677         if (rc) {
 678                 kfree(rx_queue->buffer);
 679                 rx_queue->buffer = NULL;
 680         }
 681         return rc;
 682 }
 683
 684 void efx_init_rx_queue(struct efx_rx_queue *rx_queue)
 685 {
 686         struct efx_nic *efx = rx_queue->efx;
 687         unsigned int max_fill, trigger, limit;
 688
 689         netif_dbg(rx_queue->efx, drv, rx_queue->efx->net_dev,
 690                   "initialising RX queue %d\n", efx_rx_queue_index(rx_queue));
 691
 692         /* Initialise ptr fields */
 693         rx_queue->added_count = 0;
 694         rx_queue->notified_count = 0;
 695         rx_queue->removed_count = 0;
 696         rx_queue->min_fill = -1U;
 697
 698         /* Initialise limit fields */
 699         max_fill = efx->rxq_entries - EFX_RXD_HEAD_ROOM;
 700         trigger = max_fill * min(rx_refill_threshold, 100U) / 100U;
 701         limit = max_fill * min(rx_refill_limit, 100U) / 100U;
 702
 703         rx_queue->max_fill = max_fill;
 704         rx_queue->fast_fill_trigger = trigger;
 705         rx_queue->fast_fill_limit = limit;
 706
 707         /* Set up RX descriptor ring */
 708         rx_queue->enabled = true;
 709         efx_nic_init_rx(rx_queue);
 710 }
 711
 712 void efx_fini_rx_queue(struct efx_rx_queue *rx_queue)
 713 {
 714         int i;
 715         struct efx_rx_buffer *rx_buf;
 716
 717         netif_dbg(rx_queue->efx, drv, rx_queue->efx->net_dev,
 718                   "shutting down RX queue %d\n", efx_rx_queue_index(rx_queue));
 719
 720         /* A flush failure might have left rx_queue->enabled */
 721         rx_queue->enabled = false;
 722
 723         del_timer_sync(&rx_queue->slow_fill);
 724         efx_nic_fini_rx(rx_queue);
 725
 726         /* Release RX buffers NB start at index 0 not current HW ptr */
 727         if (rx_queue->buffer) {
 728                 for (i = 0; i <= rx_queue->ptr_mask; i++) {
 729                         rx_buf = efx_rx_buffer(rx_queue, i);
 730                         efx_fini_rx_buffer(rx_queue, rx_buf);
 731                 }
 732         }
 733 }
 734
 735 void efx_remove_rx_queue(struct efx_rx_queue *rx_queue)
 736 {
 737         netif_dbg(rx_queue->efx, drv, rx_queue->efx->net_dev,
 738                   "destroying RX queue %d\n", efx_rx_queue_index(rx_queue));
 739
 740         efx_nic_remove_rx(rx_queue);
 741
 742         kfree(rx_queue->buffer);
 743         rx_queue->buffer = NULL;
 744 }
 745
 746
 747 module_param(rx_alloc_method, int, 0644);
 748 MODULE_PARM_DESC(rx_alloc_method, "Allocation method used for RX buffers");
 749
 750 module_param(rx_refill_threshold, uint, 0444);
 751 MODULE_PARM_DESC(rx_refill_threshold,
 752                  "RX descriptor ring fast/slow fill threshold (%)");
 753