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_buffer(struct efx_rx_buffer *rx_buf)
 228 {
 229         if (rx_buf->page) {
 230                 put_page(rx_buf->page);
 231                 rx_buf->page = NULL;
 232         }
 233 }
 234
 235 /* Attempt to recycle the page if there is an RX recycle ring; the page can
 236  * only be added if this is the final RX buffer, to prevent pages being used in
 237  * the descriptor ring and appearing in the recycle ring simultaneously.
 238  */
 239 static void efx_recycle_rx_page(struct efx_channel *channel,
 240                                 struct efx_rx_buffer *rx_buf)
 241 {
 242         struct page *page = rx_buf->page;
 243         struct efx_rx_queue *rx_queue = efx_channel_get_rx_queue(channel);
 244         struct efx_nic *efx = rx_queue->efx;
 245         unsigned index;
 246
 247         /* Only recycle the page after processing the final buffer. */
 248         if (!(rx_buf->flags & EFX_RX_BUF_LAST_IN_PAGE))
 249                 return;
 250
 251         index = rx_queue->page_add & rx_queue->page_ptr_mask;
 252         if (rx_queue->page_ring[index] == NULL) {
 253                 unsigned read_index = rx_queue->page_remove &
 254                         rx_queue->page_ptr_mask;
 255
 256                 /* The next slot in the recycle ring is available, but
 257                  * increment page_remove if the read pointer currently
 258                  * points here.
 259                  */
 260                 if (read_index == index)
 261                         ++rx_queue->page_remove;
 262                 rx_queue->page_ring[index] = page;
 263                 ++rx_queue->page_add;
 264                 return;
 265         }
 266         ++rx_queue->page_recycle_full;
 267         efx_unmap_rx_buffer(efx, rx_buf);
 268         put_page(rx_buf->page);
 269 }
 270
 271 static void efx_fini_rx_buffer(struct efx_rx_queue *rx_queue,
 272                                struct efx_rx_buffer *rx_buf)
 273 {
 274         /* Release the page reference we hold for the buffer. */
 275         if (rx_buf->page)
 276                 put_page(rx_buf->page);
 277
 278         /* If this is the last buffer in a page, unmap and free it. */
 279         if (rx_buf->flags & EFX_RX_BUF_LAST_IN_PAGE) {
 280                 efx_unmap_rx_buffer(rx_queue->efx, rx_buf);
 281                 efx_free_rx_buffer(rx_buf);
 282         }
 283         rx_buf->page = NULL;
 284 }
 285
 286 /* Recycle the pages that are used by buffers that have just been received. */
 287 static void efx_recycle_rx_pages(struct efx_channel *channel,
 288                                  struct efx_rx_buffer *rx_buf,
 289                                  unsigned int n_frags)
 290 {
 291         struct efx_rx_queue *rx_queue = efx_channel_get_rx_queue(channel);
 292
 293         do {
 294                 efx_recycle_rx_page(channel, rx_buf);
 295                 rx_buf = efx_rx_buf_next(rx_queue, rx_buf);
 296         } while (--n_frags);
 297 }
 298
 299 static void efx_discard_rx_packet(struct efx_channel *channel,
 300                                   struct efx_rx_buffer *rx_buf,
 301                                   unsigned int n_frags)
 302 {
 303         struct efx_rx_queue *rx_queue = efx_channel_get_rx_queue(channel);
 304
 305         efx_recycle_rx_pages(channel, rx_buf, n_frags);
 306
 307         do {
 308                 efx_free_rx_buffer(rx_buf);
 309                 rx_buf = efx_rx_buf_next(rx_queue, rx_buf);
 310         } while (--n_frags);
 311 }
 312
 313 /**
 314  * efx_fast_push_rx_descriptors - push new RX descriptors quickly
 315  * @rx_queue:           RX descriptor queue
 316  *
 317  * This will aim to fill the RX descriptor queue up to
 318  * @rx_queue->@max_fill. If there is insufficient atomic
 319  * memory to do so, a slow fill will be scheduled.
 320  *
 321  * The caller must provide serialisation (none is used here). In practise,
 322  * this means this function must run from the NAPI handler, or be called
 323  * when NAPI is disabled.
 324  */
 325 void efx_fast_push_rx_descriptors(struct efx_rx_queue *rx_queue, bool atomic)
 326 {
 327         struct efx_nic *efx = rx_queue->efx;
 328         unsigned int fill_level, batch_size;
 329         int space, rc = 0;
 330
 331         if (!rx_queue->refill_enabled)
 332                 return;
 333
 334         /* Calculate current fill level, and exit if we don't need to fill */
 335         fill_level = (rx_queue->added_count - rx_queue->removed_count);
 336         EFX_BUG_ON_PARANOID(fill_level > rx_queue->efx->rxq_entries);
 337         if (fill_level >= rx_queue->fast_fill_trigger)
 338                 goto out;
 339
 340         /* Record minimum fill level */
 341         if (unlikely(fill_level < rx_queue->min_fill)) {
 342                 if (fill_level)
 343                         rx_queue->min_fill = fill_level;
 344         }
 345
 346         batch_size = efx->rx_pages_per_batch * efx->rx_bufs_per_page;
 347         space = rx_queue->max_fill - fill_level;
 348         EFX_BUG_ON_PARANOID(space < batch_size);
 349
 350         netif_vdbg(rx_queue->efx, rx_status, rx_queue->efx->net_dev,
 351                    "RX queue %d fast-filling descriptor ring from"
 352                    " level %d to level %d\n",
 353                    efx_rx_queue_index(rx_queue), fill_level,
 354                    rx_queue->max_fill);
 355
 356
 357         do {
 358                 rc = efx_init_rx_buffers(rx_queue, atomic);
 359                 if (unlikely(rc)) {
 360                         /* Ensure that we don't leave the rx queue empty */
 361                         if (rx_queue->added_count == rx_queue->removed_count)
 362                                 efx_schedule_slow_fill(rx_queue);
 363                         goto out;
 364                 }
 365         } while ((space -= batch_size) >= batch_size);
 366
 367         netif_vdbg(rx_queue->efx, rx_status, rx_queue->efx->net_dev,
 368                    "RX queue %d fast-filled descriptor ring "
 369                    "to level %d\n", efx_rx_queue_index(rx_queue),
 370                    rx_queue->added_count - rx_queue->removed_count);
 371
 372  out:
 373         if (rx_queue->notified_count != rx_queue->added_count)
 374                 efx_nic_notify_rx_desc(rx_queue);
 375 }
 376
 377 void efx_rx_slow_fill(unsigned long context)
 378 {
 379         struct efx_rx_queue *rx_queue = (struct efx_rx_queue *)context;
 380
 381         /* Post an event to cause NAPI to run and refill the queue */
 382         efx_nic_generate_fill_event(rx_queue);
 383         ++rx_queue->slow_fill_count;
 384 }
 385
 386 static void efx_rx_packet__check_len(struct efx_rx_queue *rx_queue,
 387                                      struct efx_rx_buffer *rx_buf,
 388                                      int len)
 389 {
 390         struct efx_nic *efx = rx_queue->efx;
 391         unsigned max_len = rx_buf->len - efx->type->rx_buffer_padding;
 392
 393         if (likely(len <= max_len))
 394                 return;
 395
 396         /* The packet must be discarded, but this is only a fatal error
 397          * if the caller indicated it was
 398          */
 399         rx_buf->flags |= EFX_RX_PKT_DISCARD;
 400
 401         if ((len > rx_buf->len) && EFX_WORKAROUND_8071(efx)) {
 402                 if (net_ratelimit())
 403                         netif_err(efx, rx_err, efx->net_dev,
 404                                   " RX queue %d seriously overlength "
 405                                   "RX event (0x%x > 0x%x+0x%x). Leaking\n",
 406                                   efx_rx_queue_index(rx_queue), len, max_len,
 407                                   efx->type->rx_buffer_padding);
 408                 efx_schedule_reset(efx, RESET_TYPE_RX_RECOVERY);
 409         } else {
 410                 if (net_ratelimit())
 411                         netif_err(efx, rx_err, efx->net_dev,
 412                                   " RX queue %d overlength RX event "
 413                                   "(0x%x > 0x%x)\n",
 414                                   efx_rx_queue_index(rx_queue), len, max_len);
 415         }
 416
 417         efx_rx_queue_channel(rx_queue)->n_rx_overlength++;
 418 }
 419
 420 /* Pass a received packet up through GRO.  GRO can handle pages
 421  * regardless of checksum state and skbs with a good checksum.
 422  */
 423 static void
 424 efx_rx_packet_gro(struct efx_channel *channel, struct efx_rx_buffer *rx_buf,
 425                   unsigned int n_frags, u8 *eh)
 426 {
 427         struct napi_struct *napi = &channel->napi_str;
 428         gro_result_t gro_result;
 429         struct efx_nic *efx = channel->efx;
 430         struct sk_buff *skb;
 431
 432         skb = napi_get_frags(napi);
 433         if (unlikely(!skb)) {
 434                 while (n_frags--) {
 435                         put_page(rx_buf->page);
 436                         rx_buf->page = NULL;
 437                         rx_buf = efx_rx_buf_next(&channel->rx_queue, rx_buf);
 438                 }
 439                 return;
 440         }
 441
 442         if (efx->net_dev->features & NETIF_F_RXHASH)
 443                 skb_set_hash(skb, efx_rx_buf_hash(efx, eh),
 444                              PKT_HASH_TYPE_L3);
 445         skb->ip_summed = ((rx_buf->flags & EFX_RX_PKT_CSUMMED) ?
 446                           CHECKSUM_UNNECESSARY : CHECKSUM_NONE);
 447
 448         for (;;) {
 449                 skb_fill_page_desc(skb, skb_shinfo(skb)->nr_frags,
 450                                    rx_buf->page, rx_buf->page_offset,
 451                                    rx_buf->len);
 452                 rx_buf->page = NULL;
 453                 skb->len += rx_buf->len;
 454                 if (skb_shinfo(skb)->nr_frags == n_frags)
 455                         break;
 456
 457                 rx_buf = efx_rx_buf_next(&channel->rx_queue, rx_buf);
 458         }
 459
 460         skb->data_len = skb->len;
 461         skb->truesize += n_frags * efx->rx_buffer_truesize;
 462
 463         skb_record_rx_queue(skb, channel->rx_queue.core_index);
 464
 465         gro_result = napi_gro_frags(napi);
 466         if (gro_result != GRO_DROP)
 467                 channel->irq_mod_score += 2;
 468 }
 469
 470 /* Allocate and construct an SKB around page fragments */
 471 static struct sk_buff *efx_rx_mk_skb(struct efx_channel *channel,
 472                                      struct efx_rx_buffer *rx_buf,
 473                                      unsigned int n_frags,
 474                                      u8 *eh, int hdr_len)
 475 {
 476         struct efx_nic *efx = channel->efx;
 477         struct sk_buff *skb;
 478
 479         /* Allocate an SKB to store the headers */
 480         skb = netdev_alloc_skb(efx->net_dev,
 481                                efx->rx_ip_align + efx->rx_prefix_size +
 482                                hdr_len);
 483         if (unlikely(skb == NULL))
 484                 return NULL;
 485
 486         EFX_BUG_ON_PARANOID(rx_buf->len < hdr_len);
 487
 488         memcpy(skb->data + efx->rx_ip_align, eh - efx->rx_prefix_size,
 489                efx->rx_prefix_size + hdr_len);
 490         skb_reserve(skb, efx->rx_ip_align + efx->rx_prefix_size);
 491         __skb_put(skb, hdr_len);
 492
 493         /* Append the remaining page(s) onto the frag list */
 494         if (rx_buf->len > hdr_len) {
 495                 rx_buf->page_offset += hdr_len;
 496                 rx_buf->len -= hdr_len;
 497
 498                 for (;;) {
 499                         skb_fill_page_desc(skb, skb_shinfo(skb)->nr_frags,
 500                                            rx_buf->page, rx_buf->page_offset,
 501                                            rx_buf->len);
 502                         rx_buf->page = NULL;
 503                         skb->len += rx_buf->len;
 504                         skb->data_len += rx_buf->len;
 505                         if (skb_shinfo(skb)->nr_frags == n_frags)
 506                                 break;
 507
 508                         rx_buf = efx_rx_buf_next(&channel->rx_queue, rx_buf);
 509                 }
 510         } else {
 511                 __free_pages(rx_buf->page, efx->rx_buffer_order);
 512                 rx_buf->page = NULL;
 513                 n_frags = 0;
 514         }
 515
 516         skb->truesize += n_frags * efx->rx_buffer_truesize;
 517
 518         /* Move past the ethernet header */
 519         skb->protocol = eth_type_trans(skb, efx->net_dev);
 520
 521         return skb;
 522 }
 523
 524 void efx_rx_packet(struct efx_rx_queue *rx_queue, unsigned int index,
 525                    unsigned int n_frags, unsigned int len, u16 flags)
 526 {
 527         struct efx_nic *efx = rx_queue->efx;
 528         struct efx_channel *channel = efx_rx_queue_channel(rx_queue);
 529         struct efx_rx_buffer *rx_buf;
 530
 531         rx_buf = efx_rx_buffer(rx_queue, index);
 532         rx_buf->flags |= flags;
 533
 534         /* Validate the number of fragments and completed length */
 535         if (n_frags == 1) {
 536                 if (!(flags & EFX_RX_PKT_PREFIX_LEN))
 537                         efx_rx_packet__check_len(rx_queue, rx_buf, len);
 538         } else if (unlikely(n_frags > EFX_RX_MAX_FRAGS) ||
 539                    unlikely(len <= (n_frags - 1) * efx->rx_dma_len) ||
 540                    unlikely(len > n_frags * efx->rx_dma_len) ||
 541                    unlikely(!efx->rx_scatter)) {
 542                 /* If this isn't an explicit discard request, either
 543                  * the hardware or the driver is broken.
 544                  */
 545                 WARN_ON(!(len == 0 && rx_buf->flags & EFX_RX_PKT_DISCARD));
 546                 rx_buf->flags |= EFX_RX_PKT_DISCARD;
 547         }
 548
 549         netif_vdbg(efx, rx_status, efx->net_dev,
 550                    "RX queue %d received ids %x-%x len %d %s%s\n",
 551                    efx_rx_queue_index(rx_queue), index,
 552                    (index + n_frags - 1) & rx_queue->ptr_mask, len,
 553                    (rx_buf->flags & EFX_RX_PKT_CSUMMED) ? " [SUMMED]" : "",
 554                    (rx_buf->flags & EFX_RX_PKT_DISCARD) ? " [DISCARD]" : "");
 555
 556         /* Discard packet, if instructed to do so.  Process the
 557          * previous receive first.
 558          */
 559         if (unlikely(rx_buf->flags & EFX_RX_PKT_DISCARD)) {
 560                 efx_rx_flush_packet(channel);
 561                 efx_discard_rx_packet(channel, rx_buf, n_frags);
 562                 return;
 563         }
 564
 565         if (n_frags == 1 && !(flags & EFX_RX_PKT_PREFIX_LEN))
 566                 rx_buf->len = len;
 567
 568         /* Release and/or sync the DMA mapping - assumes all RX buffers
 569          * consumed in-order per RX queue.
 570          */
 571         efx_sync_rx_buffer(efx, rx_buf, rx_buf->len);
 572
 573         /* Prefetch nice and early so data will (hopefully) be in cache by
 574          * the time we look at it.
 575          */
 576         prefetch(efx_rx_buf_va(rx_buf));
 577
 578         rx_buf->page_offset += efx->rx_prefix_size;
 579         rx_buf->len -= efx->rx_prefix_size;
 580
 581         if (n_frags > 1) {
 582                 /* Release/sync DMA mapping for additional fragments.
 583                  * Fix length for last fragment.
 584                  */
 585                 unsigned int tail_frags = n_frags - 1;
 586
 587                 for (;;) {
 588                         rx_buf = efx_rx_buf_next(rx_queue, rx_buf);
 589                         if (--tail_frags == 0)
 590                                 break;
 591                         efx_sync_rx_buffer(efx, rx_buf, efx->rx_dma_len);
 592                 }
 593                 rx_buf->len = len - (n_frags - 1) * efx->rx_dma_len;
 594                 efx_sync_rx_buffer(efx, rx_buf, rx_buf->len);
 595         }
 596
 597         /* All fragments have been DMA-synced, so recycle pages. */
 598         rx_buf = efx_rx_buffer(rx_queue, index);
 599         efx_recycle_rx_pages(channel, rx_buf, n_frags);
 600
 601         /* Pipeline receives so that we give time for packet headers to be
 602          * prefetched into cache.
 603          */
 604         efx_rx_flush_packet(channel);
 605         channel->rx_pkt_n_frags = n_frags;
 606         channel->rx_pkt_index = index;
 607 }
 608
 609 static void efx_rx_deliver(struct efx_channel *channel, u8 *eh,
 610                            struct efx_rx_buffer *rx_buf,
 611                            unsigned int n_frags)
 612 {
 613         struct sk_buff *skb;
 614         u16 hdr_len = min_t(u16, rx_buf->len, EFX_SKB_HEADERS);
 615
 616         skb = efx_rx_mk_skb(channel, rx_buf, n_frags, eh, hdr_len);
 617         if (unlikely(skb == NULL)) {
 618                 efx_free_rx_buffer(rx_buf);
 619                 return;
 620         }
 621         skb_record_rx_queue(skb, channel->rx_queue.core_index);
 622
 623         /* Set the SKB flags */
 624         skb_checksum_none_assert(skb);
 625         if (likely(rx_buf->flags & EFX_RX_PKT_CSUMMED))
 626                 skb->ip_summed = CHECKSUM_UNNECESSARY;
 627
 628         efx_rx_skb_attach_timestamp(channel, skb);
 629
 630         if (channel->type->receive_skb)
 631                 if (channel->type->receive_skb(channel, skb))
 632                         return;
 633
 634         /* Pass the packet up */
 635         netif_receive_skb(skb);
 636 }
 637
 638 /* Handle a received packet.  Second half: Touches packet payload. */
 639 void __efx_rx_packet(struct efx_channel *channel)
 640 {
 641         struct efx_nic *efx = channel->efx;
 642         struct efx_rx_buffer *rx_buf =
 643                 efx_rx_buffer(&channel->rx_queue, channel->rx_pkt_index);
 644         u8 *eh = efx_rx_buf_va(rx_buf);
 645
 646         /* Read length from the prefix if necessary.  This already
 647          * excludes the length of the prefix itself.
 648          */
 649         if (rx_buf->flags & EFX_RX_PKT_PREFIX_LEN)
 650                 rx_buf->len = le16_to_cpup((__le16 *)
 651                                            (eh + efx->rx_packet_len_offset));
 652
 653         /* If we're in loopback test, then pass the packet directly to the
 654          * loopback layer, and free the rx_buf here
 655          */
 656         if (unlikely(efx->loopback_selftest)) {
 657                 efx_loopback_rx_packet(efx, eh, rx_buf->len);
 658                 efx_free_rx_buffer(rx_buf);
 659                 goto out;
 660         }
 661
 662         if (unlikely(!(efx->net_dev->features & NETIF_F_RXCSUM)))
 663                 rx_buf->flags &= ~EFX_RX_PKT_CSUMMED;
 664
 665         if ((rx_buf->flags & EFX_RX_PKT_TCP) && !channel->type->receive_skb)
 666                 efx_rx_packet_gro(channel, rx_buf, channel->rx_pkt_n_frags, eh);
 667         else
 668                 efx_rx_deliver(channel, eh, rx_buf, channel->rx_pkt_n_frags);
 669 out:
 670         channel->rx_pkt_n_frags = 0;
 671 }
 672
 673 int efx_probe_rx_queue(struct efx_rx_queue *rx_queue)
 674 {
 675         struct efx_nic *efx = rx_queue->efx;
 676         unsigned int entries;
 677         int rc;
 678
 679         /* Create the smallest power-of-two aligned ring */
 680         entries = max(roundup_pow_of_two(efx->rxq_entries), EFX_MIN_DMAQ_SIZE);
 681         EFX_BUG_ON_PARANOID(entries > EFX_MAX_DMAQ_SIZE);
 682         rx_queue->ptr_mask = entries - 1;
 683
 684         netif_dbg(efx, probe, efx->net_dev,
 685                   "creating RX queue %d size %#x mask %#x\n",
 686                   efx_rx_queue_index(rx_queue), efx->rxq_entries,
 687                   rx_queue->ptr_mask);
 688
 689         /* Allocate RX buffers */
 690         rx_queue->buffer = kcalloc(entries, sizeof(*rx_queue->buffer),
 691                                    GFP_KERNEL);
 692         if (!rx_queue->buffer)
 693                 return -ENOMEM;
 694
 695         rc = efx_nic_probe_rx(rx_queue);
 696         if (rc) {
 697                 kfree(rx_queue->buffer);
 698                 rx_queue->buffer = NULL;
 699         }
 700
 701         return rc;
 702 }
 703
 704 static void efx_init_rx_recycle_ring(struct efx_nic *efx,
 705                                      struct efx_rx_queue *rx_queue)
 706 {
 707         unsigned int bufs_in_recycle_ring, page_ring_size;
 708
 709         /* Set the RX recycle ring size */
 710 #ifdef CONFIG_PPC64
 711         bufs_in_recycle_ring = EFX_RECYCLE_RING_SIZE_IOMMU;
 712 #else
 713         if (iommu_present(&pci_bus_type))
 714                 bufs_in_recycle_ring = EFX_RECYCLE_RING_SIZE_IOMMU;
 715         else
 716                 bufs_in_recycle_ring = EFX_RECYCLE_RING_SIZE_NOIOMMU;
 717 #endif /* CONFIG_PPC64 */
 718
 719         page_ring_size = roundup_pow_of_two(bufs_in_recycle_ring /
 720                                             efx->rx_bufs_per_page);
 721         rx_queue->page_ring = kcalloc(page_ring_size,
 722                                       sizeof(*rx_queue->page_ring), GFP_KERNEL);
 723         rx_queue->page_ptr_mask = page_ring_size - 1;
 724 }
 725
 726 void efx_init_rx_queue(struct efx_rx_queue *rx_queue)
 727 {
 728         struct efx_nic *efx = rx_queue->efx;
 729         unsigned int max_fill, trigger, max_trigger;
 730
 731         netif_dbg(rx_queue->efx, drv, rx_queue->efx->net_dev,
 732                   "initialising RX queue %d\n", efx_rx_queue_index(rx_queue));
 733
 734         /* Initialise ptr fields */
 735         rx_queue->added_count = 0;
 736         rx_queue->notified_count = 0;
 737         rx_queue->removed_count = 0;
 738         rx_queue->min_fill = -1U;
 739         efx_init_rx_recycle_ring(efx, rx_queue);
 740
 741         rx_queue->page_remove = 0;
 742         rx_queue->page_add = rx_queue->page_ptr_mask + 1;
 743         rx_queue->page_recycle_count = 0;
 744         rx_queue->page_recycle_failed = 0;
 745         rx_queue->page_recycle_full = 0;
 746
 747         /* Initialise limit fields */
 748         max_fill = efx->rxq_entries - EFX_RXD_HEAD_ROOM;
 749         max_trigger =
 750                 max_fill - efx->rx_pages_per_batch * efx->rx_bufs_per_page;
 751         if (rx_refill_threshold != 0) {
 752                 trigger = max_fill * min(rx_refill_threshold, 100U) / 100U;
 753                 if (trigger > max_trigger)
 754                         trigger = max_trigger;
 755         } else {
 756                 trigger = max_trigger;
 757         }
 758
 759         rx_queue->max_fill = max_fill;
 760         rx_queue->fast_fill_trigger = trigger;
 761         rx_queue->refill_enabled = true;
 762
 763         /* Set up RX descriptor ring */
 764         efx_nic_init_rx(rx_queue);
 765 }
 766
 767 void efx_fini_rx_queue(struct efx_rx_queue *rx_queue)
 768 {
 769         int i;
 770         struct efx_nic *efx = rx_queue->efx;
 771         struct efx_rx_buffer *rx_buf;
 772
 773         netif_dbg(rx_queue->efx, drv, rx_queue->efx->net_dev,
 774                   "shutting down RX queue %d\n", efx_rx_queue_index(rx_queue));
 775
 776         del_timer_sync(&rx_queue->slow_fill);
 777
 778         /* Release RX buffers from the current read ptr to the write ptr */
 779         if (rx_queue->buffer) {
 780                 for (i = rx_queue->removed_count; i < rx_queue->added_count;
 781                      i++) {
 782                         unsigned index = i & rx_queue->ptr_mask;
 783                         rx_buf = efx_rx_buffer(rx_queue, index);
 784                         efx_fini_rx_buffer(rx_queue, rx_buf);
 785                 }
 786         }
 787
 788         /* Unmap and release the pages in the recycle ring. Remove the ring. */
 789         for (i = 0; i <= rx_queue->page_ptr_mask; i++) {
 790                 struct page *page = rx_queue->page_ring[i];
 791                 struct efx_rx_page_state *state;
 792
 793                 if (page == NULL)
 794                         continue;
 795
 796                 state = page_address(page);
 797                 dma_unmap_page(&efx->pci_dev->dev, state->dma_addr,
 798                                PAGE_SIZE << efx->rx_buffer_order,
 799                                DMA_FROM_DEVICE);
 800                 put_page(page);
 801         }
 802         kfree(rx_queue->page_ring);
 803         rx_queue->page_ring = NULL;
 804 }
 805
 806 void efx_remove_rx_queue(struct efx_rx_queue *rx_queue)
 807 {
 808         netif_dbg(rx_queue->efx, drv, rx_queue->efx->net_dev,
 809                   "destroying RX queue %d\n", efx_rx_queue_index(rx_queue));
 810
 811         efx_nic_remove_rx(rx_queue);
 812
 813         kfree(rx_queue->buffer);
 814         rx_queue->buffer = NULL;
 815 }
 816
 817
 818 module_param(rx_refill_threshold, uint, 0444);
 819 MODULE_PARM_DESC(rx_refill_threshold,
 820                  "RX descriptor ring refill threshold (%)");
 821
 822 #ifdef CONFIG_RFS_ACCEL
 823
 824 int efx_filter_rfs(struct net_device *net_dev, const struct sk_buff *skb,
 825                    u16 rxq_index, u32 flow_id)
 826 {
 827         struct efx_nic *efx = netdev_priv(net_dev);
 828         struct efx_channel *channel;
 829         struct efx_filter_spec spec;
 830         const __be16 *ports;
 831         __be16 ether_type;
 832         int nhoff;
 833         int rc;
 834
 835         /* The core RPS/RFS code has already parsed and validated
 836          * VLAN, IP and transport headers.  We assume they are in the
 837          * header area.
 838          */
 839
 840         if (skb->protocol == htons(ETH_P_8021Q)) {
 841                 const struct vlan_hdr *vh =
 842                         (const struct vlan_hdr *)skb->data;
 843
 844                 /* We can't filter on the IP 5-tuple and the vlan
 845                  * together, so just strip the vlan header and filter
 846                  * on the IP part.
 847                  */
 848                 EFX_BUG_ON_PARANOID(skb_headlen(skb) < sizeof(*vh));
 849                 ether_type = vh->h_vlan_encapsulated_proto;
 850                 nhoff = sizeof(struct vlan_hdr);
 851         } else {
 852                 ether_type = skb->protocol;
 853                 nhoff = 0;
 854         }
 855
 856         if (ether_type != htons(ETH_P_IP) && ether_type != htons(ETH_P_IPV6))
 857                 return -EPROTONOSUPPORT;
 858
 859         efx_filter_init_rx(&spec, EFX_FILTER_PRI_HINT,
 860                            efx->rx_scatter ? EFX_FILTER_FLAG_RX_SCATTER : 0,
 861                            rxq_index);
 862         spec.match_flags =
 863                 EFX_FILTER_MATCH_ETHER_TYPE | EFX_FILTER_MATCH_IP_PROTO |
 864                 EFX_FILTER_MATCH_LOC_HOST | EFX_FILTER_MATCH_LOC_PORT |
 865                 EFX_FILTER_MATCH_REM_HOST | EFX_FILTER_MATCH_REM_PORT;
 866         spec.ether_type = ether_type;
 867
 868         if (ether_type == htons(ETH_P_IP)) {
 869                 const struct iphdr *ip =
 870                         (const struct iphdr *)(skb->data + nhoff);
 871
 872                 EFX_BUG_ON_PARANOID(skb_headlen(skb) < nhoff + sizeof(*ip));
 873                 if (ip_is_fragment(ip))
 874                         return -EPROTONOSUPPORT;
 875                 spec.ip_proto = ip->protocol;
 876                 spec.rem_host[0] = ip->saddr;
 877                 spec.loc_host[0] = ip->daddr;
 878                 EFX_BUG_ON_PARANOID(skb_headlen(skb) < nhoff + 4 * ip->ihl + 4);
 879                 ports = (const __be16 *)(skb->data + nhoff + 4 * ip->ihl);
 880         } else {
 881                 const struct ipv6hdr *ip6 =
 882                         (const struct ipv6hdr *)(skb->data + nhoff);
 883
 884                 EFX_BUG_ON_PARANOID(skb_headlen(skb) <
 885                                     nhoff + sizeof(*ip6) + 4);
 886                 spec.ip_proto = ip6->nexthdr;
 887                 memcpy(spec.rem_host, &ip6->saddr, sizeof(ip6->saddr));
 888                 memcpy(spec.loc_host, &ip6->daddr, sizeof(ip6->daddr));
 889                 ports = (const __be16 *)(ip6 + 1);
 890         }
 891
 892         spec.rem_port = ports[0];
 893         spec.loc_port = ports[1];
 894
 895         rc = efx->type->filter_rfs_insert(efx, &spec);
 896         if (rc < 0)
 897                 return rc;
 898
 899         /* Remember this so we can check whether to expire the filter later */
 900         efx->rps_flow_id[rc] = flow_id;
 901         channel = efx_get_channel(efx, skb_get_rx_queue(skb));
 902         ++channel->rfs_filters_added;
 903
 904         if (ether_type == htons(ETH_P_IP))
 905                 netif_info(efx, rx_status, efx->net_dev,
 906                            "steering %s %pI4:%u:%pI4:%u to queue %u [flow %u filter %d]\n",
 907                            (spec.ip_proto == IPPROTO_TCP) ? "TCP" : "UDP",
 908                            spec.rem_host, ntohs(ports[0]), spec.loc_host,
 909                            ntohs(ports[1]), rxq_index, flow_id, rc);
 910         else
 911                 netif_info(efx, rx_status, efx->net_dev,
 912                            "steering %s [%pI6]:%u:[%pI6]:%u to queue %u [flow %u filter %d]\n",
 913                            (spec.ip_proto == IPPROTO_TCP) ? "TCP" : "UDP",
 914                            spec.rem_host, ntohs(ports[0]), spec.loc_host,
 915                            ntohs(ports[1]), rxq_index, flow_id, rc);
 916
 917         return rc;
 918 }
 919
 920 bool __efx_filter_rfs_expire(struct efx_nic *efx, unsigned int quota)
 921 {
 922         bool (*expire_one)(struct efx_nic *efx, u32 flow_id, unsigned int index);
 923         unsigned int index, size;
 924         u32 flow_id;
 925
 926         if (!spin_trylock_bh(&efx->filter_lock))
 927                 return false;
 928
 929         expire_one = efx->type->filter_rfs_expire_one;
 930         index = efx->rps_expire_index;
 931         size = efx->type->max_rx_ip_filters;
 932         while (quota--) {
 933                 flow_id = efx->rps_flow_id[index];
 934                 if (expire_one(efx, flow_id, index))
 935                         netif_info(efx, rx_status, efx->net_dev,
 936                                    "expired filter %d [flow %u]\n",
 937                                    index, flow_id);
 938                 if (++index == size)
 939                         index = 0;
 940         }
 941         efx->rps_expire_index = index;
 942
 943         spin_unlock_bh(&efx->filter_lock);
 944         return true;
 945 }
 946
 947 #endif /* CONFIG_RFS_ACCEL */
 948
 949 /**
 950  * efx_filter_is_mc_recipient - test whether spec is a multicast recipient
 951  * @spec: Specification to test
 952  *
 953  * Return: %true if the specification is a non-drop RX filter that
 954  * matches a local MAC address I/G bit value of 1 or matches a local
 955  * IPv4 or IPv6 address value in the respective multicast address
 956  * range.  Otherwise %false.
 957  */
 958 bool efx_filter_is_mc_recipient(const struct efx_filter_spec *spec)
 959 {
 960         if (!(spec->flags & EFX_FILTER_FLAG_RX) ||
 961             spec->dmaq_id == EFX_FILTER_RX_DMAQ_ID_DROP)
 962                 return false;
 963
 964         if (spec->match_flags &
 965             (EFX_FILTER_MATCH_LOC_MAC | EFX_FILTER_MATCH_LOC_MAC_IG) &&
 966             is_multicast_ether_addr(spec->loc_mac))
 967                 return true;
 968
 969         if ((spec->match_flags &
 970              (EFX_FILTER_MATCH_ETHER_TYPE | EFX_FILTER_MATCH_LOC_HOST)) ==
 971             (EFX_FILTER_MATCH_ETHER_TYPE | EFX_FILTER_MATCH_LOC_HOST)) {
 972                 if (spec->ether_type == htons(ETH_P_IP) &&
 973                     ipv4_is_multicast(spec->loc_host[0]))
 974                         return true;
 975                 if (spec->ether_type == htons(ETH_P_IPV6) &&
 976                     ((const u8 *)spec->loc_host)[0] == 0xff)
 977                         return true;
 978         }
 979
 980         return false;
 981 }