kernel/drivers/net/ethernet/sfc/rx.c

   1 /****************************************************************************
   2  * Driver for Solarflare network controllers and boards
   3  * Copyright 2005-2006 Fen Systems Ltd.
   4  * Copyright 2005-2013 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/ipv6.h>
  16 #include <linux/tcp.h>
  17 #include <linux/udp.h>
  18 #include <linux/prefetch.h>
  19 #include <linux/moduleparam.h>
  20 #include <linux/iommu.h>
  21 #include <net/ip.h>
  22 #include <net/checksum.h>
  23 #include "net_driver.h"
  24 #include "efx.h"
  25 #include "filter.h"
  26 #include "nic.h"
  27 #include "selftest.h"
  28 #include "workarounds.h"
  29
  30 /* Preferred number of descriptors to fill at once */
  31 #define EFX_RX_PREFERRED_BATCH 8U
  32
  33 /* Number of RX buffers to recycle pages for.  When creating the RX page recycle
  34  * ring, this number is divided by the number of buffers per page to calculate
  35  * the number of pages to store in the RX page recycle ring.
  36  */
  37 #define EFX_RECYCLE_RING_SIZE_IOMMU 4096
  38 #define EFX_RECYCLE_RING_SIZE_NOIOMMU (2 * EFX_RX_PREFERRED_BATCH)
  39
  40 /* Size of buffer allocated for skb header area. */
  41 #define EFX_SKB_HEADERS  128u
  42
  43 /* This is the percentage fill level below which new RX descriptors
  44  * will be added to the RX descriptor ring.
  45  */
  46 static unsigned int rx_refill_threshold;
  47
  48 /* Each packet can consume up to ceil(max_frame_len / buffer_size) buffers */
  49 #define EFX_RX_MAX_FRAGS DIV_ROUND_UP(EFX_MAX_FRAME_LEN(EFX_MAX_MTU), \
  50                                       EFX_RX_USR_BUF_SIZE)
  51
  52 /*
  53  * RX maximum head room required.
  54  *
  55  * This must be at least 1 to prevent overflow, plus one packet-worth
  56  * to allow pipelined receives.
  57  */
  58 #define EFX_RXD_HEAD_ROOM (1 + EFX_RX_MAX_FRAGS)
  59
  60 static inline u8 *efx_rx_buf_va(struct efx_rx_buffer *buf)
  61 {
  62         return page_address(buf->page) + buf->page_offset;
  63 }
  64
  65 static inline u32 efx_rx_buf_hash(struct efx_nic *efx, const u8 *eh)
  66 {
  67 #if defined(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS)
  68         return __le32_to_cpup((const __le32 *)(eh + efx->rx_packet_hash_offset));
  69 #else
  70         const u8 *data = eh + efx->rx_packet_hash_offset;
  71         return (u32)data[0]       |
  72                (u32)data[1] << 8  |
  73                (u32)data[2] << 16 |
  74                (u32)data[3] << 24;
  75 #endif
  76 }
  77
  78 static inline struct efx_rx_buffer *
  79 efx_rx_buf_next(struct efx_rx_queue *rx_queue, struct efx_rx_buffer *rx_buf)
  80 {
  81         if (unlikely(rx_buf == efx_rx_buffer(rx_queue, rx_queue->ptr_mask)))
  82                 return efx_rx_buffer(rx_queue, 0);
  83         else
  84                 return rx_buf + 1;
  85 }
  86
  87 static inline void efx_sync_rx_buffer(struct efx_nic *efx,
  88                                       struct efx_rx_buffer *rx_buf,
  89                                       unsigned int len)
  90 {
  91         dma_sync_single_for_cpu(&efx->pci_dev->dev, rx_buf->dma_addr, len,
  92                                 DMA_FROM_DEVICE);
  93 }
  94
  95 void efx_rx_config_page_split(struct efx_nic *efx)
  96 {
  97         efx->rx_page_buf_step = ALIGN(efx->rx_dma_len + efx->rx_ip_align,
  98                                       EFX_RX_BUF_ALIGNMENT);
  99         efx->rx_bufs_per_page = efx->rx_buffer_order ? 1 :
 100                 ((PAGE_SIZE - sizeof(struct efx_rx_page_state)) /
 101                  efx->rx_page_buf_step);
 102         efx->rx_buffer_truesize = (PAGE_SIZE << efx->rx_buffer_order) /
 103                 efx->rx_bufs_per_page;
 104         efx->rx_pages_per_batch = DIV_ROUND_UP(EFX_RX_PREFERRED_BATCH,
 105                                                efx->rx_bufs_per_page);
 106 }
 107
 108 /* Check the RX page recycle ring for a page that can be reused. */
 109 static struct page *efx_reuse_page(struct efx_rx_queue *rx_queue)
 110 {
 111         struct efx_nic *efx = rx_queue->efx;
 112         struct page *page;
 113         struct efx_rx_page_state *state;
 114         unsigned index;
 115
 116         index = rx_queue->page_remove & rx_queue->page_ptr_mask;
 117         page = rx_queue->page_ring[index];
 118         if (page == NULL)
 119                 return NULL;
 120
 121         rx_queue->page_ring[index] = NULL;
 122         /* page_remove cannot exceed page_add. */
 123         if (rx_queue->page_remove != rx_queue->page_add)
 124                 ++rx_queue->page_remove;
 125
 126         /* If page_count is 1 then we hold the only reference to this page. */
 127         if (page_count(page) == 1) {
 128                 ++rx_queue->page_recycle_count;
 129                 return page;
 130         } else {
 131                 state = page_address(page);
 132                 dma_unmap_page(&efx->pci_dev->dev, state->dma_addr,
 133                                PAGE_SIZE << efx->rx_buffer_order,
 134                                DMA_FROM_DEVICE);
 135                 put_page(page);
 136                 ++rx_queue->page_recycle_failed;
 137         }
 138
 139         return NULL;
 140 }
 141
 142 /**
 143  * efx_init_rx_buffers - create EFX_RX_BATCH page-based RX buffers
 144  *
 145  * @rx_queue:           Efx RX queue
 146  *
 147  * This allocates a batch of pages, maps them for DMA, and populates
 148  * struct efx_rx_buffers for each one. Return a negative error code or
 149  * 0 on success. If a single page can be used for multiple buffers,
 150  * then the page will either be inserted fully, or not at all.
 151  */
 152 static int efx_init_rx_buffers(struct efx_rx_queue *rx_queue, bool atomic)
 153 {
 154         struct efx_nic *efx = rx_queue->efx;
 155         struct efx_rx_buffer *rx_buf;
 156         struct page *page;
 157         unsigned int page_offset;
 158         struct efx_rx_page_state *state;
 159         dma_addr_t dma_addr;
 160         unsigned index, count;
 161
 162         count = 0;
 163         do {
 164                 page = efx_reuse_page(rx_queue);
 165                 if (page == NULL) {
 166                         page = alloc_pages(__GFP_COLD | __GFP_COMP |
 167                                            (atomic ? GFP_ATOMIC : GFP_KERNEL),
 168                                            efx->rx_buffer_order);
 169                         if (unlikely(page == NULL))
 170                                 return -ENOMEM;
 171                         dma_addr =
 172                                 dma_map_page(&efx->pci_dev->dev, page, 0,
 173                                              PAGE_SIZE << efx->rx_buffer_order,
 174                                              DMA_FROM_DEVICE);
 175                         if (unlikely(dma_mapping_error(&efx->pci_dev->dev,
 176                                                        dma_addr))) {
 177                                 __free_pages(page, efx->rx_buffer_order);
 178                                 return -EIO;
 179                         }
 180                         state = page_address(page);
 181                         state->dma_addr = dma_addr;
 182                 } else {
 183                         state = page_address(page);
 184                         dma_addr = state->dma_addr;
 185                 }
 186
 187                 dma_addr += sizeof(struct efx_rx_page_state);
 188                 page_offset = sizeof(struct efx_rx_page_state);
 189
 190                 do {
 191                         index = rx_queue->added_count & rx_queue->ptr_mask;
 192                         rx_buf = efx_rx_buffer(rx_queue, index);
 193                         rx_buf->dma_addr = dma_addr + efx->rx_ip_align;
 194                         rx_buf->page = page;
 195                         rx_buf->page_offset = page_offset + efx->rx_ip_align;
 196                         rx_buf->len = efx->rx_dma_len;
 197                         rx_buf->flags = 0;
 198                         ++rx_queue->added_count;
 199                         get_page(page);
 200                         dma_addr += efx->rx_page_buf_step;
 201                         page_offset += efx->rx_page_buf_step;
 202                 } while (page_offset + efx->rx_page_buf_step <= PAGE_SIZE);
 203
 204                 rx_buf->flags = EFX_RX_BUF_LAST_IN_PAGE;
 205         } while (++count < efx->rx_pages_per_batch);
 206
 207         return 0;
 208 }
 209
 210 /* Unmap a DMA-mapped page.  This function is only called for the final RX
 211  * buffer in a page.
 212  */
 213 static void efx_unmap_rx_buffer(struct efx_nic *efx,
 214                                 struct efx_rx_buffer *rx_buf)
 215 {
 216         struct page *page = rx_buf->page;
 217
 218         if (page) {
 219                 struct efx_rx_page_state *state = page_address(page);
 220                 dma_unmap_page(&efx->pci_dev->dev,
 221                                state->dma_addr,
 222                                PAGE_SIZE << efx->rx_buffer_order,
 223                                DMA_FROM_DEVICE);
 224         }
 225 }
 226
 227 static void efx_free_rx_buffers(struct efx_rx_queue *rx_queue,
 228                                 struct efx_rx_buffer *rx_buf,
 229                                 unsigned int num_bufs)
 230 {
 231         do {
 232                 if (rx_buf->page) {
 233                         put_page(rx_buf->page);
 234                         rx_buf->page = NULL;
 235                 }
 236                 rx_buf = efx_rx_buf_next(rx_queue, rx_buf);
 237         } while (--num_bufs);
 238 }
 239
 240 /* Attempt to recycle the page if there is an RX recycle ring; the page can
 241  * only be added if this is the final RX buffer, to prevent pages being used in
 242  * the descriptor ring and appearing in the recycle ring simultaneously.
 243  */
 244 static void efx_recycle_rx_page(struct efx_channel *channel,
 245                                 struct efx_rx_buffer *rx_buf)
 246 {
 247         struct page *page = rx_buf->page;
 248         struct efx_rx_queue *rx_queue = efx_channel_get_rx_queue(channel);
 249         struct efx_nic *efx = rx_queue->efx;
 250         unsigned index;
 251
 252         /* Only recycle the page after processing the final buffer. */
 253         if (!(rx_buf->flags & EFX_RX_BUF_LAST_IN_PAGE))
 254                 return;
 255
 256         index = rx_queue->page_add & rx_queue->page_ptr_mask;
 257         if (rx_queue->page_ring[index] == NULL) {
 258                 unsigned read_index = rx_queue->page_remove &
 259                         rx_queue->page_ptr_mask;
 260
 261                 /* The next slot in the recycle ring is available, but
 262                  * increment page_remove if the read pointer currently
 263                  * points here.
 264                  */
 265                 if (read_index == index)
 266                         ++rx_queue->page_remove;
 267                 rx_queue->page_ring[index] = page;
 268                 ++rx_queue->page_add;
 269                 return;
 270         }
 271         ++rx_queue->page_recycle_full;
 272         efx_unmap_rx_buffer(efx, rx_buf);
 273         put_page(rx_buf->page);
 274 }
 275
 276 static void efx_fini_rx_buffer(struct efx_rx_queue *rx_queue,
 277                                struct efx_rx_buffer *rx_buf)
 278 {
 279         /* Release the page reference we hold for the buffer. */
 280         if (rx_buf->page)
 281                 put_page(rx_buf->page);
 282
 283         /* If this is the last buffer in a page, unmap and free it. */
 284         if (rx_buf->flags & EFX_RX_BUF_LAST_IN_PAGE) {
 285                 efx_unmap_rx_buffer(rx_queue->efx, rx_buf);
 286                 efx_free_rx_buffers(rx_queue, rx_buf, 1);
 287         }
 288         rx_buf->page = NULL;
 289 }
 290
 291 /* Recycle the pages that are used by buffers that have just been received. */
 292 static void efx_recycle_rx_pages(struct efx_channel *channel,
 293                                  struct efx_rx_buffer *rx_buf,
 294                                  unsigned int n_frags)
 295 {
 296         struct efx_rx_queue *rx_queue = efx_channel_get_rx_queue(channel);
 297
 298         do {
 299                 efx_recycle_rx_page(channel, rx_buf);
 300                 rx_buf = efx_rx_buf_next(rx_queue, rx_buf);
 301         } while (--n_frags);
 302 }
 303
 304 static void efx_discard_rx_packet(struct efx_channel *channel,
 305                                   struct efx_rx_buffer *rx_buf,
 306                                   unsigned int n_frags)
 307 {
 308         struct efx_rx_queue *rx_queue = efx_channel_get_rx_queue(channel);
 309
 310         efx_recycle_rx_pages(channel, rx_buf, n_frags);
 311
 312         efx_free_rx_buffers(rx_queue, rx_buf, n_frags);
 313 }
 314
 315 /**
 316  * efx_fast_push_rx_descriptors - push new RX descriptors quickly
 317  * @rx_queue:           RX descriptor queue
 318  *
 319  * This will aim to fill the RX descriptor queue up to
 320  * @rx_queue->@max_fill. If there is insufficient atomic
 321  * memory to do so, a slow fill will be scheduled.
 322  *
 323  * The caller must provide serialisation (none is used here). In practise,
 324  * this means this function must run from the NAPI handler, or be called
 325  * when NAPI is disabled.
 326  */
 327 void efx_fast_push_rx_descriptors(struct efx_rx_queue *rx_queue, bool atomic)
 328 {
 329         struct efx_nic *efx = rx_queue->efx;
 330         unsigned int fill_level, batch_size;
 331         int space, rc = 0;
 332
 333         if (!rx_queue->refill_enabled)
 334                 return;
 335
 336         /* Calculate current fill level, and exit if we don't need to fill */
 337         fill_level = (rx_queue->added_count - rx_queue->removed_count);
 338         EFX_BUG_ON_PARANOID(fill_level > rx_queue->efx->rxq_entries);
 339         if (fill_level >= rx_queue->fast_fill_trigger)
 340                 goto out;
 341
 342         /* Record minimum fill level */
 343         if (unlikely(fill_level < rx_queue->min_fill)) {
 344                 if (fill_level)
 345                         rx_queue->min_fill = fill_level;
 346         }
 347
 348         batch_size = efx->rx_pages_per_batch * efx->rx_bufs_per_page;
 349         space = rx_queue->max_fill - fill_level;
 350         EFX_BUG_ON_PARANOID(space < batch_size);
 351
 352         netif_vdbg(rx_queue->efx, rx_status, rx_queue->efx->net_dev,
 353                    "RX queue %d fast-filling descriptor ring from"
 354                    " level %d to level %d\n",
 355                    efx_rx_queue_index(rx_queue), fill_level,
 356                    rx_queue->max_fill);
 357
 358
 359         do {
 360                 rc = efx_init_rx_buffers(rx_queue, atomic);
 361                 if (unlikely(rc)) {
 362                         /* Ensure that we don't leave the rx queue empty */
 363                         if (rx_queue->added_count == rx_queue->removed_count)
 364                                 efx_schedule_slow_fill(rx_queue);
 365                         goto out;
 366                 }
 367         } while ((space -= batch_size) >= batch_size);
 368
 369         netif_vdbg(rx_queue->efx, rx_status, rx_queue->efx->net_dev,
 370                    "RX queue %d fast-filled descriptor ring "
 371                    "to level %d\n", efx_rx_queue_index(rx_queue),
 372                    rx_queue->added_count - rx_queue->removed_count);
 373
 374  out:
 375         if (rx_queue->notified_count != rx_queue->added_count)
 376                 efx_nic_notify_rx_desc(rx_queue);
 377 }
 378
 379 void efx_rx_slow_fill(unsigned long context)
 380 {
 381         struct efx_rx_queue *rx_queue = (struct efx_rx_queue *)context;
 382
 383         /* Post an event to cause NAPI to run and refill the queue */
 384         efx_nic_generate_fill_event(rx_queue);
 385         ++rx_queue->slow_fill_count;
 386 }
 387
 388 static void efx_rx_packet__check_len(struct efx_rx_queue *rx_queue,
 389                                      struct efx_rx_buffer *rx_buf,
 390                                      int len)
 391 {
 392         struct efx_nic *efx = rx_queue->efx;
 393         unsigned max_len = rx_buf->len - efx->type->rx_buffer_padding;
 394
 395         if (likely(len <= max_len))
 396                 return;
 397
 398         /* The packet must be discarded, but this is only a fatal error
 399          * if the caller indicated it was
 400          */
 401         rx_buf->flags |= EFX_RX_PKT_DISCARD;
 402
 403         if ((len > rx_buf->len) && EFX_WORKAROUND_8071(efx)) {
 404                 if (net_ratelimit())
 405                         netif_err(efx, rx_err, efx->net_dev,
 406                                   " RX queue %d seriously overlength "
 407                                   "RX event (0x%x > 0x%x+0x%x). Leaking\n",
 408                                   efx_rx_queue_index(rx_queue), len, max_len,
 409                                   efx->type->rx_buffer_padding);
 410                 efx_schedule_reset(efx, RESET_TYPE_RX_RECOVERY);
 411         } else {
 412                 if (net_ratelimit())
 413                         netif_err(efx, rx_err, efx->net_dev,
 414                                   " RX queue %d overlength RX event "
 415                                   "(0x%x > 0x%x)\n",
 416                                   efx_rx_queue_index(rx_queue), len, max_len);
 417         }
 418
 419         efx_rx_queue_channel(rx_queue)->n_rx_overlength++;
 420 }
 421
 422 /* Pass a received packet up through GRO.  GRO can handle pages
 423  * regardless of checksum state and skbs with a good checksum.
 424  */
 425 static void
 426 efx_rx_packet_gro(struct efx_channel *channel, struct efx_rx_buffer *rx_buf,
 427                   unsigned int n_frags, u8 *eh)
 428 {
 429         struct napi_struct *napi = &channel->napi_str;
 430         gro_result_t gro_result;
 431         struct efx_nic *efx = channel->efx;
 432         struct sk_buff *skb;
 433
 434         skb = napi_get_frags(napi);
 435         if (unlikely(!skb)) {
 436                 struct efx_rx_queue *rx_queue;
 437
 438                 rx_queue = efx_channel_get_rx_queue(channel);
 439                 efx_free_rx_buffers(rx_queue, rx_buf, n_frags);
 440                 return;
 441         }
 442
 443         if (efx->net_dev->features & NETIF_F_RXHASH)
 444                 skb_set_hash(skb, efx_rx_buf_hash(efx, eh),
 445                              PKT_HASH_TYPE_L3);
 446         skb->ip_summed = ((rx_buf->flags & EFX_RX_PKT_CSUMMED) ?
 447                           CHECKSUM_UNNECESSARY : CHECKSUM_NONE);
 448
 449         for (;;) {
 450                 skb_fill_page_desc(skb, skb_shinfo(skb)->nr_frags,
 451                                    rx_buf->page, rx_buf->page_offset,
 452                                    rx_buf->len);
 453                 rx_buf->page = NULL;
 454                 skb->len += rx_buf->len;
 455                 if (skb_shinfo(skb)->nr_frags == n_frags)
 456                         break;
 457
 458                 rx_buf = efx_rx_buf_next(&channel->rx_queue, rx_buf);
 459         }
 460
 461         skb->data_len = skb->len;
 462         skb->truesize += n_frags * efx->rx_buffer_truesize;
 463
 464         skb_record_rx_queue(skb, channel->rx_queue.core_index);
 465
 466         skb_mark_napi_id(skb, &channel->napi_str);
 467         gro_result = napi_gro_frags(napi);
 468         if (gro_result != GRO_DROP)
 469                 channel->irq_mod_score += 2;
 470 }
 471
 472 /* Allocate and construct an SKB around page fragments */
 473 static struct sk_buff *efx_rx_mk_skb(struct efx_channel *channel,
 474                                      struct efx_rx_buffer *rx_buf,
 475                                      unsigned int n_frags,
 476                                      u8 *eh, int hdr_len)
 477 {
 478         struct efx_nic *efx = channel->efx;
 479         struct sk_buff *skb;
 480
 481         /* Allocate an SKB to store the headers */
 482         skb = netdev_alloc_skb(efx->net_dev,
 483                                efx->rx_ip_align + efx->rx_prefix_size +
 484                                hdr_len);
 485         if (unlikely(skb == NULL)) {
 486                 atomic_inc(&efx->n_rx_noskb_drops);
 487                 return NULL;
 488         }
 489
 490         EFX_BUG_ON_PARANOID(rx_buf->len < hdr_len);
 491
 492         memcpy(skb->data + efx->rx_ip_align, eh - efx->rx_prefix_size,
 493                efx->rx_prefix_size + hdr_len);
 494         skb_reserve(skb, efx->rx_ip_align + efx->rx_prefix_size);
 495         __skb_put(skb, hdr_len);
 496
 497         /* Append the remaining page(s) onto the frag list */
 498         if (rx_buf->len > hdr_len) {
 499                 rx_buf->page_offset += hdr_len;
 500                 rx_buf->len -= hdr_len;
 501
 502                 for (;;) {
 503                         skb_fill_page_desc(skb, skb_shinfo(skb)->nr_frags,
 504                                            rx_buf->page, rx_buf->page_offset,
 505                                            rx_buf->len);
 506                         rx_buf->page = NULL;
 507                         skb->len += rx_buf->len;
 508                         skb->data_len += rx_buf->len;
 509                         if (skb_shinfo(skb)->nr_frags == n_frags)
 510                                 break;
 511
 512                         rx_buf = efx_rx_buf_next(&channel->rx_queue, rx_buf);
 513                 }
 514         } else {
 515                 __free_pages(rx_buf->page, efx->rx_buffer_order);
 516                 rx_buf->page = NULL;
 517                 n_frags = 0;
 518         }
 519
 520         skb->truesize += n_frags * efx->rx_buffer_truesize;
 521
 522         /* Move past the ethernet header */
 523         skb->protocol = eth_type_trans(skb, efx->net_dev);
 524
 525         skb_mark_napi_id(skb, &channel->napi_str);
 526
 527         return skb;
 528 }
 529
 530 void efx_rx_packet(struct efx_rx_queue *rx_queue, unsigned int index,
 531                    unsigned int n_frags, unsigned int len, u16 flags)
 532 {
 533         struct efx_nic *efx = rx_queue->efx;
 534         struct efx_channel *channel = efx_rx_queue_channel(rx_queue);
 535         struct efx_rx_buffer *rx_buf;
 536
 537         rx_queue->rx_packets++;
 538
 539         rx_buf = efx_rx_buffer(rx_queue, index);
 540         rx_buf->flags |= flags;
 541
 542         /* Validate the number of fragments and completed length */
 543         if (n_frags == 1) {
 544                 if (!(flags & EFX_RX_PKT_PREFIX_LEN))
 545                         efx_rx_packet__check_len(rx_queue, rx_buf, len);
 546         } else if (unlikely(n_frags > EFX_RX_MAX_FRAGS) ||
 547                    unlikely(len <= (n_frags - 1) * efx->rx_dma_len) ||
 548                    unlikely(len > n_frags * efx->rx_dma_len) ||
 549                    unlikely(!efx->rx_scatter)) {
 550                 /* If this isn't an explicit discard request, either
 551                  * the hardware or the driver is broken.
 552                  */
 553                 WARN_ON(!(len == 0 && rx_buf->flags & EFX_RX_PKT_DISCARD));
 554                 rx_buf->flags |= EFX_RX_PKT_DISCARD;
 555         }
 556
 557         netif_vdbg(efx, rx_status, efx->net_dev,
 558                    "RX queue %d received ids %x-%x len %d %s%s\n",
 559                    efx_rx_queue_index(rx_queue), index,
 560                    (index + n_frags - 1) & rx_queue->ptr_mask, len,
 561                    (rx_buf->flags & EFX_RX_PKT_CSUMMED) ? " [SUMMED]" : "",
 562                    (rx_buf->flags & EFX_RX_PKT_DISCARD) ? " [DISCARD]" : "");
 563
 564         /* Discard packet, if instructed to do so.  Process the
 565          * previous receive first.
 566          */
 567         if (unlikely(rx_buf->flags & EFX_RX_PKT_DISCARD)) {
 568                 efx_rx_flush_packet(channel);
 569                 efx_discard_rx_packet(channel, rx_buf, n_frags);
 570                 return;
 571         }
 572
 573         if (n_frags == 1 && !(flags & EFX_RX_PKT_PREFIX_LEN))
 574                 rx_buf->len = len;
 575
 576         /* Release and/or sync the DMA mapping - assumes all RX buffers
 577          * consumed in-order per RX queue.
 578          */
 579         efx_sync_rx_buffer(efx, rx_buf, rx_buf->len);
 580
 581         /* Prefetch nice and early so data will (hopefully) be in cache by
 582          * the time we look at it.
 583          */
 584         prefetch(efx_rx_buf_va(rx_buf));
 585
 586         rx_buf->page_offset += efx->rx_prefix_size;
 587         rx_buf->len -= efx->rx_prefix_size;
 588
 589         if (n_frags > 1) {
 590                 /* Release/sync DMA mapping for additional fragments.
 591                  * Fix length for last fragment.
 592                  */
 593                 unsigned int tail_frags = n_frags - 1;
 594
 595                 for (;;) {
 596                         rx_buf = efx_rx_buf_next(rx_queue, rx_buf);
 597                         if (--tail_frags == 0)
 598                                 break;
 599                         efx_sync_rx_buffer(efx, rx_buf, efx->rx_dma_len);
 600                 }
 601                 rx_buf->len = len - (n_frags - 1) * efx->rx_dma_len;
 602                 efx_sync_rx_buffer(efx, rx_buf, rx_buf->len);
 603         }
 604
 605         /* All fragments have been DMA-synced, so recycle pages. */
 606         rx_buf = efx_rx_buffer(rx_queue, index);
 607         efx_recycle_rx_pages(channel, rx_buf, n_frags);
 608
 609         /* Pipeline receives so that we give time for packet headers to be
 610          * prefetched into cache.
 611          */
 612         efx_rx_flush_packet(channel);
 613         channel->rx_pkt_n_frags = n_frags;
 614         channel->rx_pkt_index = index;
 615 }
 616
 617 static void efx_rx_deliver(struct efx_channel *channel, u8 *eh,
 618                            struct efx_rx_buffer *rx_buf,
 619                            unsigned int n_frags)
 620 {
 621         struct sk_buff *skb;
 622         u16 hdr_len = min_t(u16, rx_buf->len, EFX_SKB_HEADERS);
 623
 624         skb = efx_rx_mk_skb(channel, rx_buf, n_frags, eh, hdr_len);
 625         if (unlikely(skb == NULL)) {
 626                 struct efx_rx_queue *rx_queue;
 627
 628                 rx_queue = efx_channel_get_rx_queue(channel);
 629                 efx_free_rx_buffers(rx_queue, rx_buf, n_frags);
 630                 return;
 631         }
 632         skb_record_rx_queue(skb, channel->rx_queue.core_index);
 633
 634         /* Set the SKB flags */
 635         skb_checksum_none_assert(skb);
 636         if (likely(rx_buf->flags & EFX_RX_PKT_CSUMMED))
 637                 skb->ip_summed = CHECKSUM_UNNECESSARY;
 638
 639         efx_rx_skb_attach_timestamp(channel, skb);
 640
 641         if (channel->type->receive_skb)
 642                 if (channel->type->receive_skb(channel, skb))
 643                         return;
 644
 645         /* Pass the packet up */
 646         netif_receive_skb(skb);
 647 }
 648
 649 /* Handle a received packet.  Second half: Touches packet payload. */
 650 void __efx_rx_packet(struct efx_channel *channel)
 651 {
 652         struct efx_nic *efx = channel->efx;
 653         struct efx_rx_buffer *rx_buf =
 654                 efx_rx_buffer(&channel->rx_queue, channel->rx_pkt_index);
 655         u8 *eh = efx_rx_buf_va(rx_buf);
 656
 657         /* Read length from the prefix if necessary.  This already
 658          * excludes the length of the prefix itself.
 659          */
 660         if (rx_buf->flags & EFX_RX_PKT_PREFIX_LEN)
 661                 rx_buf->len = le16_to_cpup((__le16 *)
 662                                            (eh + efx->rx_packet_len_offset));
 663
 664         /* If we're in loopback test, then pass the packet directly to the
 665          * loopback layer, and free the rx_buf here
 666          */
 667         if (unlikely(efx->loopback_selftest)) {
 668                 struct efx_rx_queue *rx_queue;
 669
 670                 efx_loopback_rx_packet(efx, eh, rx_buf->len);
 671                 rx_queue = efx_channel_get_rx_queue(channel);
 672                 efx_free_rx_buffers(rx_queue, rx_buf,
 673                                     channel->rx_pkt_n_frags);
 674                 goto out;
 675         }
 676
 677         if (unlikely(!(efx->net_dev->features & NETIF_F_RXCSUM)))
 678                 rx_buf->flags &= ~EFX_RX_PKT_CSUMMED;
 679
 680         if ((rx_buf->flags & EFX_RX_PKT_TCP) && !channel->type->receive_skb &&
 681             !efx_channel_busy_polling(channel))
 682                 efx_rx_packet_gro(channel, rx_buf, channel->rx_pkt_n_frags, eh);
 683         else
 684                 efx_rx_deliver(channel, eh, rx_buf, channel->rx_pkt_n_frags);
 685 out:
 686         channel->rx_pkt_n_frags = 0;
 687 }
 688
 689 int efx_probe_rx_queue(struct efx_rx_queue *rx_queue)
 690 {
 691         struct efx_nic *efx = rx_queue->efx;
 692         unsigned int entries;
 693         int rc;
 694
 695         /* Create the smallest power-of-two aligned ring */
 696         entries = max(roundup_pow_of_two(efx->rxq_entries), EFX_MIN_DMAQ_SIZE);
 697         EFX_BUG_ON_PARANOID(entries > EFX_MAX_DMAQ_SIZE);
 698         rx_queue->ptr_mask = entries - 1;
 699
 700         netif_dbg(efx, probe, efx->net_dev,
 701                   "creating RX queue %d size %#x mask %#x\n",
 702                   efx_rx_queue_index(rx_queue), efx->rxq_entries,
 703                   rx_queue->ptr_mask);
 704
 705         /* Allocate RX buffers */
 706         rx_queue->buffer = kcalloc(entries, sizeof(*rx_queue->buffer),
 707                                    GFP_KERNEL);
 708         if (!rx_queue->buffer)
 709                 return -ENOMEM;
 710
 711         rc = efx_nic_probe_rx(rx_queue);
 712         if (rc) {
 713                 kfree(rx_queue->buffer);
 714                 rx_queue->buffer = NULL;
 715         }
 716
 717         return rc;
 718 }
 719
 720 static void efx_init_rx_recycle_ring(struct efx_nic *efx,
 721                                      struct efx_rx_queue *rx_queue)
 722 {
 723         unsigned int bufs_in_recycle_ring, page_ring_size;
 724
 725         /* Set the RX recycle ring size */
 726 #ifdef CONFIG_PPC64
 727         bufs_in_recycle_ring = EFX_RECYCLE_RING_SIZE_IOMMU;
 728 #else
 729         if (iommu_present(&pci_bus_type))
 730                 bufs_in_recycle_ring = EFX_RECYCLE_RING_SIZE_IOMMU;
 731         else
 732                 bufs_in_recycle_ring = EFX_RECYCLE_RING_SIZE_NOIOMMU;
 733 #endif /* CONFIG_PPC64 */
 734
 735         page_ring_size = roundup_pow_of_two(bufs_in_recycle_ring /
 736                                             efx->rx_bufs_per_page);
 737         rx_queue->page_ring = kcalloc(page_ring_size,
 738                                       sizeof(*rx_queue->page_ring), GFP_KERNEL);
 739         rx_queue->page_ptr_mask = page_ring_size - 1;
 740 }
 741
 742 void efx_init_rx_queue(struct efx_rx_queue *rx_queue)
 743 {
 744         struct efx_nic *efx = rx_queue->efx;
 745         unsigned int max_fill, trigger, max_trigger;
 746
 747         netif_dbg(rx_queue->efx, drv, rx_queue->efx->net_dev,
 748                   "initialising RX queue %d\n", efx_rx_queue_index(rx_queue));
 749
 750         /* Initialise ptr fields */
 751         rx_queue->added_count = 0;
 752         rx_queue->notified_count = 0;
 753         rx_queue->removed_count = 0;
 754         rx_queue->min_fill = -1U;
 755         efx_init_rx_recycle_ring(efx, rx_queue);
 756
 757         rx_queue->page_remove = 0;
 758         rx_queue->page_add = rx_queue->page_ptr_mask + 1;
 759         rx_queue->page_recycle_count = 0;
 760         rx_queue->page_recycle_failed = 0;
 761         rx_queue->page_recycle_full = 0;
 762
 763         /* Initialise limit fields */
 764         max_fill = efx->rxq_entries - EFX_RXD_HEAD_ROOM;
 765         max_trigger =
 766                 max_fill - efx->rx_pages_per_batch * efx->rx_bufs_per_page;
 767         if (rx_refill_threshold != 0) {
 768                 trigger = max_fill * min(rx_refill_threshold, 100U) / 100U;
 769                 if (trigger > max_trigger)
 770                         trigger = max_trigger;
 771         } else {
 772                 trigger = max_trigger;
 773         }
 774
 775         rx_queue->max_fill = max_fill;
 776         rx_queue->fast_fill_trigger = trigger;
 777         rx_queue->refill_enabled = true;
 778
 779         /* Set up RX descriptor ring */
 780         efx_nic_init_rx(rx_queue);
 781 }
 782
 783 void efx_fini_rx_queue(struct efx_rx_queue *rx_queue)
 784 {
 785         int i;
 786         struct efx_nic *efx = rx_queue->efx;
 787         struct efx_rx_buffer *rx_buf;
 788
 789         netif_dbg(rx_queue->efx, drv, rx_queue->efx->net_dev,
 790                   "shutting down RX queue %d\n", efx_rx_queue_index(rx_queue));
 791
 792         del_timer_sync(&rx_queue->slow_fill);
 793
 794         /* Release RX buffers from the current read ptr to the write ptr */
 795         if (rx_queue->buffer) {
 796                 for (i = rx_queue->removed_count; i < rx_queue->added_count;
 797                      i++) {
 798                         unsigned index = i & rx_queue->ptr_mask;
 799                         rx_buf = efx_rx_buffer(rx_queue, index);
 800                         efx_fini_rx_buffer(rx_queue, rx_buf);
 801                 }
 802         }
 803
 804         /* Unmap and release the pages in the recycle ring. Remove the ring. */
 805         for (i = 0; i <= rx_queue->page_ptr_mask; i++) {
 806                 struct page *page = rx_queue->page_ring[i];
 807                 struct efx_rx_page_state *state;
 808
 809                 if (page == NULL)
 810                         continue;
 811
 812                 state = page_address(page);
 813                 dma_unmap_page(&efx->pci_dev->dev, state->dma_addr,
 814                                PAGE_SIZE << efx->rx_buffer_order,
 815                                DMA_FROM_DEVICE);
 816                 put_page(page);
 817         }
 818         kfree(rx_queue->page_ring);
 819         rx_queue->page_ring = NULL;
 820 }
 821
 822 void efx_remove_rx_queue(struct efx_rx_queue *rx_queue)
 823 {
 824         netif_dbg(rx_queue->efx, drv, rx_queue->efx->net_dev,
 825                   "destroying RX queue %d\n", efx_rx_queue_index(rx_queue));
 826
 827         efx_nic_remove_rx(rx_queue);
 828
 829         kfree(rx_queue->buffer);
 830         rx_queue->buffer = NULL;
 831 }
 832
 833
 834 module_param(rx_refill_threshold, uint, 0444);
 835 MODULE_PARM_DESC(rx_refill_threshold,
 836                  "RX descriptor ring refill threshold (%)");
 837
 838 #ifdef CONFIG_RFS_ACCEL
 839
 840 int efx_filter_rfs(struct net_device *net_dev, const struct sk_buff *skb,
 841                    u16 rxq_index, u32 flow_id)
 842 {
 843         struct efx_nic *efx = netdev_priv(net_dev);
 844         struct efx_channel *channel;
 845         struct efx_filter_spec spec;
 846         const __be16 *ports;
 847         __be16 ether_type;
 848         int nhoff;
 849         int rc;
 850
 851         /* The core RPS/RFS code has already parsed and validated
 852          * VLAN, IP and transport headers.  We assume they are in the
 853          * header area.
 854          */
 855
 856         if (skb->protocol == htons(ETH_P_8021Q)) {
 857                 const struct vlan_hdr *vh =
 858                         (const struct vlan_hdr *)skb->data;
 859
 860                 /* We can't filter on the IP 5-tuple and the vlan
 861                  * together, so just strip the vlan header and filter
 862                  * on the IP part.
 863                  */
 864                 EFX_BUG_ON_PARANOID(skb_headlen(skb) < sizeof(*vh));
 865                 ether_type = vh->h_vlan_encapsulated_proto;
 866                 nhoff = sizeof(struct vlan_hdr);
 867         } else {
 868                 ether_type = skb->protocol;
 869                 nhoff = 0;
 870         }
 871
 872         if (ether_type != htons(ETH_P_IP) && ether_type != htons(ETH_P_IPV6))
 873                 return -EPROTONOSUPPORT;
 874
 875         efx_filter_init_rx(&spec, EFX_FILTER_PRI_HINT,
 876                            efx->rx_scatter ? EFX_FILTER_FLAG_RX_SCATTER : 0,
 877                            rxq_index);
 878         spec.match_flags =
 879                 EFX_FILTER_MATCH_ETHER_TYPE | EFX_FILTER_MATCH_IP_PROTO |
 880                 EFX_FILTER_MATCH_LOC_HOST | EFX_FILTER_MATCH_LOC_PORT |
 881                 EFX_FILTER_MATCH_REM_HOST | EFX_FILTER_MATCH_REM_PORT;
 882         spec.ether_type = ether_type;
 883
 884         if (ether_type == htons(ETH_P_IP)) {
 885                 const struct iphdr *ip =
 886                         (const struct iphdr *)(skb->data + nhoff);
 887
 888                 EFX_BUG_ON_PARANOID(skb_headlen(skb) < nhoff + sizeof(*ip));
 889                 if (ip_is_fragment(ip))
 890                         return -EPROTONOSUPPORT;
 891                 spec.ip_proto = ip->protocol;
 892                 spec.rem_host[0] = ip->saddr;
 893                 spec.loc_host[0] = ip->daddr;
 894                 EFX_BUG_ON_PARANOID(skb_headlen(skb) < nhoff + 4 * ip->ihl + 4);
 895                 ports = (const __be16 *)(skb->data + nhoff + 4 * ip->ihl);
 896         } else {
 897                 const struct ipv6hdr *ip6 =
 898                         (const struct ipv6hdr *)(skb->data + nhoff);
 899
 900                 EFX_BUG_ON_PARANOID(skb_headlen(skb) <
 901                                     nhoff + sizeof(*ip6) + 4);
 902                 spec.ip_proto = ip6->nexthdr;
 903                 memcpy(spec.rem_host, &ip6->saddr, sizeof(ip6->saddr));
 904                 memcpy(spec.loc_host, &ip6->daddr, sizeof(ip6->daddr));
 905                 ports = (const __be16 *)(ip6 + 1);
 906         }
 907
 908         spec.rem_port = ports[0];
 909         spec.loc_port = ports[1];
 910
 911         rc = efx->type->filter_rfs_insert(efx, &spec);
 912         if (rc < 0)
 913                 return rc;
 914
 915         /* Remember this so we can check whether to expire the filter later */
 916         efx->rps_flow_id[rc] = flow_id;
 917         channel = efx_get_channel(efx, skb_get_rx_queue(skb));
 918         ++channel->rfs_filters_added;
 919
 920         if (ether_type == htons(ETH_P_IP))
 921                 netif_info(efx, rx_status, efx->net_dev,
 922                            "steering %s %pI4:%u:%pI4:%u to queue %u [flow %u filter %d]\n",
 923                            (spec.ip_proto == IPPROTO_TCP) ? "TCP" : "UDP",
 924                            spec.rem_host, ntohs(ports[0]), spec.loc_host,
 925                            ntohs(ports[1]), rxq_index, flow_id, rc);
 926         else
 927                 netif_info(efx, rx_status, efx->net_dev,
 928                            "steering %s [%pI6]:%u:[%pI6]:%u to queue %u [flow %u filter %d]\n",
 929                            (spec.ip_proto == IPPROTO_TCP) ? "TCP" : "UDP",
 930                            spec.rem_host, ntohs(ports[0]), spec.loc_host,
 931                            ntohs(ports[1]), rxq_index, flow_id, rc);
 932
 933         return rc;
 934 }
 935
 936 bool __efx_filter_rfs_expire(struct efx_nic *efx, unsigned int quota)
 937 {
 938         bool (*expire_one)(struct efx_nic *efx, u32 flow_id, unsigned int index);
 939         unsigned int index, size;
 940         u32 flow_id;
 941
 942         if (!spin_trylock_bh(&efx->filter_lock))
 943                 return false;
 944
 945         expire_one = efx->type->filter_rfs_expire_one;
 946         index = efx->rps_expire_index;
 947         size = efx->type->max_rx_ip_filters;
 948         while (quota--) {
 949                 flow_id = efx->rps_flow_id[index];
 950                 if (expire_one(efx, flow_id, index))
 951                         netif_info(efx, rx_status, efx->net_dev,
 952                                    "expired filter %d [flow %u]\n",
 953                                    index, flow_id);
 954                 if (++index == size)
 955                         index = 0;
 956         }
 957         efx->rps_expire_index = index;
 958
 959         spin_unlock_bh(&efx->filter_lock);
 960         return true;
 961 }
 962
 963 #endif /* CONFIG_RFS_ACCEL */
 964
 965 /**
 966  * efx_filter_is_mc_recipient - test whether spec is a multicast recipient
 967  * @spec: Specification to test
 968  *
 969  * Return: %true if the specification is a non-drop RX filter that
 970  * matches a local MAC address I/G bit value of 1 or matches a local
 971  * IPv4 or IPv6 address value in the respective multicast address
 972  * range.  Otherwise %false.
 973  */
 974 bool efx_filter_is_mc_recipient(const struct efx_filter_spec *spec)
 975 {
 976         if (!(spec->flags & EFX_FILTER_FLAG_RX) ||
 977             spec->dmaq_id == EFX_FILTER_RX_DMAQ_ID_DROP)
 978                 return false;
 979
 980         if (spec->match_flags &
 981             (EFX_FILTER_MATCH_LOC_MAC | EFX_FILTER_MATCH_LOC_MAC_IG) &&
 982             is_multicast_ether_addr(spec->loc_mac))
 983                 return true;
 984
 985         if ((spec->match_flags &
 986              (EFX_FILTER_MATCH_ETHER_TYPE | EFX_FILTER_MATCH_LOC_HOST)) ==
 987             (EFX_FILTER_MATCH_ETHER_TYPE | EFX_FILTER_MATCH_LOC_HOST)) {
 988                 if (spec->ether_type == htons(ETH_P_IP) &&
 989                     ipv4_is_multicast(spec->loc_host[0]))
 990                         return true;
 991                 if (spec->ether_type == htons(ETH_P_IPV6) &&
 992                     ((const u8 *)spec->loc_host)[0] == 0xff)
 993                         return true;
 994         }
 995
 996         return false;
 997 }