drivers/net/ethernet/sfc/rx.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /****************************************************************************
   3  * Driver for Solarflare network controllers and boards
   4  * Copyright 2005-2006 Fen Systems Ltd.
   5  * Copyright 2005-2013 Solarflare Communications Inc.
   6  */
   7
   8 #include <linux/socket.h>
   9 #include <linux/in.h>
  10 #include <linux/slab.h>
  11 #include <linux/ip.h>
  12 #include <linux/ipv6.h>
  13 #include <linux/tcp.h>
  14 #include <linux/udp.h>
  15 #include <linux/prefetch.h>
  16 #include <linux/moduleparam.h>
  17 #include <linux/iommu.h>
  18 #include <net/ip.h>
  19 #include <net/checksum.h>
  20 #include "net_driver.h"
  21 #include "efx.h"
  22 #include "filter.h"
  23 #include "nic.h"
  24 #include "selftest.h"
  25 #include "workarounds.h"
  26
  27 /* Preferred number of descriptors to fill at once */
  28 #define EFX_RX_PREFERRED_BATCH 8U
  29
  30 /* Number of RX buffers to recycle pages for.  When creating the RX page recycle
  31  * ring, this number is divided by the number of buffers per page to calculate
  32  * the number of pages to store in the RX page recycle ring.
  33  */
  34 #define EFX_RECYCLE_RING_SIZE_IOMMU 4096
  35 #define EFX_RECYCLE_RING_SIZE_NOIOMMU (2 * EFX_RX_PREFERRED_BATCH)
  36
  37 /* Size of buffer allocated for skb header area. */
  38 #define EFX_SKB_HEADERS  128u
  39
  40 /* This is the percentage fill level below which new RX descriptors
  41  * will be added to the RX descriptor ring.
  42  */
  43 static unsigned int rx_refill_threshold;
  44
  45 /* Each packet can consume up to ceil(max_frame_len / buffer_size) buffers */
  46 #define EFX_RX_MAX_FRAGS DIV_ROUND_UP(EFX_MAX_FRAME_LEN(EFX_MAX_MTU), \
  47                                       EFX_RX_USR_BUF_SIZE)
  48
  49 /*
  50  * RX maximum head room required.
  51  *
  52  * This must be at least 1 to prevent overflow, plus one packet-worth
  53  * to allow pipelined receives.
  54  */
  55 #define EFX_RXD_HEAD_ROOM (1 + EFX_RX_MAX_FRAGS)
  56
  57 static inline u8 *efx_rx_buf_va(struct efx_rx_buffer *buf)
  58 {
  59         return page_address(buf->page) + buf->page_offset;
  60 }
  61
  62 static inline u32 efx_rx_buf_hash(struct efx_nic *efx, const u8 *eh)
  63 {
  64 #if defined(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS)
  65         return __le32_to_cpup((const __le32 *)(eh + efx->rx_packet_hash_offset));
  66 #else
  67         const u8 *data = eh + efx->rx_packet_hash_offset;
  68         return (u32)data[0]       |
  69                (u32)data[1] << 8  |
  70                (u32)data[2] << 16 |
  71                (u32)data[3] << 24;
  72 #endif
  73 }
  74
  75 static inline struct efx_rx_buffer *
  76 efx_rx_buf_next(struct efx_rx_queue *rx_queue, struct efx_rx_buffer *rx_buf)
  77 {
  78         if (unlikely(rx_buf == efx_rx_buffer(rx_queue, rx_queue->ptr_mask)))
  79                 return efx_rx_buffer(rx_queue, 0);
  80         else
  81                 return rx_buf + 1;
  82 }
  83
  84 static inline void efx_sync_rx_buffer(struct efx_nic *efx,
  85                                       struct efx_rx_buffer *rx_buf,
  86                                       unsigned int len)
  87 {
  88         dma_sync_single_for_cpu(&efx->pci_dev->dev, rx_buf->dma_addr, len,
  89                                 DMA_FROM_DEVICE);
  90 }
  91
  92 void efx_rx_config_page_split(struct efx_nic *efx)
  93 {
  94         efx->rx_page_buf_step = ALIGN(efx->rx_dma_len + efx->rx_ip_align,
  95                                       EFX_RX_BUF_ALIGNMENT);
  96         efx->rx_bufs_per_page = efx->rx_buffer_order ? 1 :
  97                 ((PAGE_SIZE - sizeof(struct efx_rx_page_state)) /
  98                  efx->rx_page_buf_step);
  99         efx->rx_buffer_truesize = (PAGE_SIZE << efx->rx_buffer_order) /
 100                 efx->rx_bufs_per_page;
 101         efx->rx_pages_per_batch = DIV_ROUND_UP(EFX_RX_PREFERRED_BATCH,
 102                                                efx->rx_bufs_per_page);
 103 }
 104
 105 /* Check the RX page recycle ring for a page that can be reused. */
 106 static struct page *efx_reuse_page(struct efx_rx_queue *rx_queue)
 107 {
 108         struct efx_nic *efx = rx_queue->efx;
 109         struct page *page;
 110         struct efx_rx_page_state *state;
 111         unsigned index;
 112
 113         index = rx_queue->page_remove & rx_queue->page_ptr_mask;
 114         page = rx_queue->page_ring[index];
 115         if (page == NULL)
 116                 return NULL;
 117
 118         rx_queue->page_ring[index] = NULL;
 119         /* page_remove cannot exceed page_add. */
 120         if (rx_queue->page_remove != rx_queue->page_add)
 121                 ++rx_queue->page_remove;
 122
 123         /* If page_count is 1 then we hold the only reference to this page. */
 124         if (page_count(page) == 1) {
 125                 ++rx_queue->page_recycle_count;
 126                 return page;
 127         } else {
 128                 state = page_address(page);
 129                 dma_unmap_page(&efx->pci_dev->dev, state->dma_addr,
 130                                PAGE_SIZE << efx->rx_buffer_order,
 131                                DMA_FROM_DEVICE);
 132                 put_page(page);
 133                 ++rx_queue->page_recycle_failed;
 134         }
 135
 136         return NULL;
 137 }
 138
 139 /**
 140  * efx_init_rx_buffers - create EFX_RX_BATCH page-based RX buffers
 141  *
 142  * @rx_queue:           Efx RX queue
 143  *
 144  * This allocates a batch of pages, maps them for DMA, and populates
 145  * struct efx_rx_buffers for each one. Return a negative error code or
 146  * 0 on success. If a single page can be used for multiple buffers,
 147  * then the page will either be inserted fully, or not at all.
 148  */
 149 static int efx_init_rx_buffers(struct efx_rx_queue *rx_queue, bool atomic)
 150 {
 151         struct efx_nic *efx = rx_queue->efx;
 152         struct efx_rx_buffer *rx_buf;
 153         struct page *page;
 154         unsigned int page_offset;
 155         struct efx_rx_page_state *state;
 156         dma_addr_t dma_addr;
 157         unsigned index, count;
 158
 159         count = 0;
 160         do {
 161                 page = efx_reuse_page(rx_queue);
 162                 if (page == NULL) {
 163                         page = alloc_pages(__GFP_COMP |
 164                                            (atomic ? GFP_ATOMIC : GFP_KERNEL),
 165                                            efx->rx_buffer_order);
 166                         if (unlikely(page == NULL))
 167                                 return -ENOMEM;
 168                         dma_addr =
 169                                 dma_map_page(&efx->pci_dev->dev, page, 0,
 170                                              PAGE_SIZE << efx->rx_buffer_order,
 171                                              DMA_FROM_DEVICE);
 172                         if (unlikely(dma_mapping_error(&efx->pci_dev->dev,
 173                                                        dma_addr))) {
 174                                 __free_pages(page, efx->rx_buffer_order);
 175                                 return -EIO;
 176                         }
 177                         state = page_address(page);
 178                         state->dma_addr = dma_addr;
 179                 } else {
 180                         state = page_address(page);
 181                         dma_addr = state->dma_addr;
 182                 }
 183
 184                 dma_addr += sizeof(struct efx_rx_page_state);
 185                 page_offset = sizeof(struct efx_rx_page_state);
 186
 187                 do {
 188                         index = rx_queue->added_count & rx_queue->ptr_mask;
 189                         rx_buf = efx_rx_buffer(rx_queue, index);
 190                         rx_buf->dma_addr = dma_addr + efx->rx_ip_align;
 191                         rx_buf->page = page;
 192                         rx_buf->page_offset = page_offset + efx->rx_ip_align;
 193                         rx_buf->len = efx->rx_dma_len;
 194                         rx_buf->flags = 0;
 195                         ++rx_queue->added_count;
 196                         get_page(page);
 197                         dma_addr += efx->rx_page_buf_step;
 198                         page_offset += efx->rx_page_buf_step;
 199                 } while (page_offset + efx->rx_page_buf_step <= PAGE_SIZE);
 200
 201                 rx_buf->flags = EFX_RX_BUF_LAST_IN_PAGE;
 202         } while (++count < efx->rx_pages_per_batch);
 203
 204         return 0;
 205 }
 206
 207 /* Unmap a DMA-mapped page.  This function is only called for the final RX
 208  * buffer in a page.
 209  */
 210 static void efx_unmap_rx_buffer(struct efx_nic *efx,
 211                                 struct efx_rx_buffer *rx_buf)
 212 {
 213         struct page *page = rx_buf->page;
 214
 215         if (page) {
 216                 struct efx_rx_page_state *state = page_address(page);
 217                 dma_unmap_page(&efx->pci_dev->dev,
 218                                state->dma_addr,
 219                                PAGE_SIZE << efx->rx_buffer_order,
 220                                DMA_FROM_DEVICE);
 221         }
 222 }
 223
 224 static void efx_free_rx_buffers(struct efx_rx_queue *rx_queue,
 225                                 struct efx_rx_buffer *rx_buf,
 226                                 unsigned int num_bufs)
 227 {
 228         do {
 229                 if (rx_buf->page) {
 230                         put_page(rx_buf->page);
 231                         rx_buf->page = NULL;
 232                 }
 233                 rx_buf = efx_rx_buf_next(rx_queue, rx_buf);
 234         } while (--num_bufs);
 235 }
 236
 237 /* Attempt to recycle the page if there is an RX recycle ring; the page can
 238  * only be added if this is the final RX buffer, to prevent pages being used in
 239  * the descriptor ring and appearing in the recycle ring simultaneously.
 240  */
 241 static void efx_recycle_rx_page(struct efx_channel *channel,
 242                                 struct efx_rx_buffer *rx_buf)
 243 {
 244         struct page *page = rx_buf->page;
 245         struct efx_rx_queue *rx_queue = efx_channel_get_rx_queue(channel);
 246         struct efx_nic *efx = rx_queue->efx;
 247         unsigned index;
 248
 249         /* Only recycle the page after processing the final buffer. */
 250         if (!(rx_buf->flags & EFX_RX_BUF_LAST_IN_PAGE))
 251                 return;
 252
 253         index = rx_queue->page_add & rx_queue->page_ptr_mask;
 254         if (rx_queue->page_ring[index] == NULL) {
 255                 unsigned read_index = rx_queue->page_remove &
 256                         rx_queue->page_ptr_mask;
 257
 258                 /* The next slot in the recycle ring is available, but
 259                  * increment page_remove if the read pointer currently
 260                  * points here.
 261                  */
 262                 if (read_index == index)
 263                         ++rx_queue->page_remove;
 264                 rx_queue->page_ring[index] = page;
 265                 ++rx_queue->page_add;
 266                 return;
 267         }
 268         ++rx_queue->page_recycle_full;
 269         efx_unmap_rx_buffer(efx, rx_buf);
 270         put_page(rx_buf->page);
 271 }
 272
 273 static void efx_fini_rx_buffer(struct efx_rx_queue *rx_queue,
 274                                struct efx_rx_buffer *rx_buf)
 275 {
 276         /* Release the page reference we hold for the buffer. */
 277         if (rx_buf->page)
 278                 put_page(rx_buf->page);
 279
 280         /* If this is the last buffer in a page, unmap and free it. */
 281         if (rx_buf->flags & EFX_RX_BUF_LAST_IN_PAGE) {
 282                 efx_unmap_rx_buffer(rx_queue->efx, rx_buf);
 283                 efx_free_rx_buffers(rx_queue, rx_buf, 1);
 284         }
 285         rx_buf->page = NULL;
 286 }
 287
 288 /* Recycle the pages that are used by buffers that have just been received. */
 289 static void efx_recycle_rx_pages(struct efx_channel *channel,
 290                                  struct efx_rx_buffer *rx_buf,
 291                                  unsigned int n_frags)
 292 {
 293         struct efx_rx_queue *rx_queue = efx_channel_get_rx_queue(channel);
 294
 295         do {
 296                 efx_recycle_rx_page(channel, rx_buf);
 297                 rx_buf = efx_rx_buf_next(rx_queue, rx_buf);
 298         } while (--n_frags);
 299 }
 300
 301 static void efx_discard_rx_packet(struct efx_channel *channel,
 302                                   struct efx_rx_buffer *rx_buf,
 303                                   unsigned int n_frags)
 304 {
 305         struct efx_rx_queue *rx_queue = efx_channel_get_rx_queue(channel);
 306
 307         efx_recycle_rx_pages(channel, rx_buf, n_frags);
 308
 309         efx_free_rx_buffers(rx_queue, rx_buf, n_frags);
 310 }
 311
 312 /**
 313  * efx_fast_push_rx_descriptors - push new RX descriptors quickly
 314  * @rx_queue:           RX descriptor queue
 315  *
 316  * This will aim to fill the RX descriptor queue up to
 317  * @rx_queue->@max_fill. If there is insufficient atomic
 318  * memory to do so, a slow fill will be scheduled.
 319  *
 320  * The caller must provide serialisation (none is used here). In practise,
 321  * this means this function must run from the NAPI handler, or be called
 322  * when NAPI is disabled.
 323  */
 324 void efx_fast_push_rx_descriptors(struct efx_rx_queue *rx_queue, bool atomic)
 325 {
 326         struct efx_nic *efx = rx_queue->efx;
 327         unsigned int fill_level, batch_size;
 328         int space, rc = 0;
 329
 330         if (!rx_queue->refill_enabled)
 331                 return;
 332
 333         /* Calculate current fill level, and exit if we don't need to fill */
 334         fill_level = (rx_queue->added_count - rx_queue->removed_count);
 335         EFX_WARN_ON_ONCE_PARANOID(fill_level > rx_queue->efx->rxq_entries);
 336         if (fill_level >= rx_queue->fast_fill_trigger)
 337                 goto out;
 338
 339         /* Record minimum fill level */
 340         if (unlikely(fill_level < rx_queue->min_fill)) {
 341                 if (fill_level)
 342                         rx_queue->min_fill = fill_level;
 343         }
 344
 345         batch_size = efx->rx_pages_per_batch * efx->rx_bufs_per_page;
 346         space = rx_queue->max_fill - fill_level;
 347         EFX_WARN_ON_ONCE_PARANOID(space < batch_size);
 348
 349         netif_vdbg(rx_queue->efx, rx_status, rx_queue->efx->net_dev,
 350                    "RX queue %d fast-filling descriptor ring from"
 351                    " level %d to level %d\n",
 352                    efx_rx_queue_index(rx_queue), fill_level,
 353                    rx_queue->max_fill);
 354
 355
 356         do {
 357                 rc = efx_init_rx_buffers(rx_queue, atomic);
 358                 if (unlikely(rc)) {
 359                         /* Ensure that we don't leave the rx queue empty */
 360                         efx_schedule_slow_fill(rx_queue);
 361                         goto out;
 362                 }
 363         } while ((space -= batch_size) >= batch_size);
 364
 365         netif_vdbg(rx_queue->efx, rx_status, rx_queue->efx->net_dev,
 366                    "RX queue %d fast-filled descriptor ring "
 367                    "to level %d\n", efx_rx_queue_index(rx_queue),
 368                    rx_queue->added_count - rx_queue->removed_count);
 369
 370  out:
 371         if (rx_queue->notified_count != rx_queue->added_count)
 372                 efx_nic_notify_rx_desc(rx_queue);
 373 }
 374
 375 void efx_rx_slow_fill(struct timer_list *t)
 376 {
 377         struct efx_rx_queue *rx_queue = from_timer(rx_queue, t, slow_fill);
 378
 379         /* Post an event to cause NAPI to run and refill the queue */
 380         efx_nic_generate_fill_event(rx_queue);
 381         ++rx_queue->slow_fill_count;
 382 }
 383
 384 static void efx_rx_packet__check_len(struct efx_rx_queue *rx_queue,
 385                                      struct efx_rx_buffer *rx_buf,
 386                                      int len)
 387 {
 388         struct efx_nic *efx = rx_queue->efx;
 389         unsigned max_len = rx_buf->len - efx->type->rx_buffer_padding;
 390
 391         if (likely(len <= max_len))
 392                 return;
 393
 394         /* The packet must be discarded, but this is only a fatal error
 395          * if the caller indicated it was
 396          */
 397         rx_buf->flags |= EFX_RX_PKT_DISCARD;
 398
 399         if (net_ratelimit())
 400                 netif_err(efx, rx_err, efx->net_dev,
 401                           "RX queue %d overlength RX event (%#x > %#x)\n",
 402                           efx_rx_queue_index(rx_queue), len, max_len);
 403
 404         efx_rx_queue_channel(rx_queue)->n_rx_overlength++;
 405 }
 406
 407 /* Pass a received packet up through GRO.  GRO can handle pages
 408  * regardless of checksum state and skbs with a good checksum.
 409  */
 410 static void
 411 efx_rx_packet_gro(struct efx_channel *channel, struct efx_rx_buffer *rx_buf,
 412                   unsigned int n_frags, u8 *eh)
 413 {
 414         struct napi_struct *napi = &channel->napi_str;
 415         gro_result_t gro_result;
 416         struct efx_nic *efx = channel->efx;
 417         struct sk_buff *skb;
 418
 419         skb = napi_get_frags(napi);
 420         if (unlikely(!skb)) {
 421                 struct efx_rx_queue *rx_queue;
 422
 423                 rx_queue = efx_channel_get_rx_queue(channel);
 424                 efx_free_rx_buffers(rx_queue, rx_buf, n_frags);
 425                 return;
 426         }
 427
 428         if (efx->net_dev->features & NETIF_F_RXHASH)
 429                 skb_set_hash(skb, efx_rx_buf_hash(efx, eh),
 430                              PKT_HASH_TYPE_L3);
 431         skb->ip_summed = ((rx_buf->flags & EFX_RX_PKT_CSUMMED) ?
 432                           CHECKSUM_UNNECESSARY : CHECKSUM_NONE);
 433         skb->csum_level = !!(rx_buf->flags & EFX_RX_PKT_CSUM_LEVEL);
 434
 435         for (;;) {
 436                 skb_fill_page_desc(skb, skb_shinfo(skb)->nr_frags,
 437                                    rx_buf->page, rx_buf->page_offset,
 438                                    rx_buf->len);
 439                 rx_buf->page = NULL;
 440                 skb->len += rx_buf->len;
 441                 if (skb_shinfo(skb)->nr_frags == n_frags)
 442                         break;
 443
 444                 rx_buf = efx_rx_buf_next(&channel->rx_queue, rx_buf);
 445         }
 446
 447         skb->data_len = skb->len;
 448         skb->truesize += n_frags * efx->rx_buffer_truesize;
 449
 450         skb_record_rx_queue(skb, channel->rx_queue.core_index);
 451
 452         gro_result = napi_gro_frags(napi);
 453         if (gro_result != GRO_DROP)
 454                 channel->irq_mod_score += 2;
 455 }
 456
 457 /* Allocate and construct an SKB around page fragments */
 458 static struct sk_buff *efx_rx_mk_skb(struct efx_channel *channel,
 459                                      struct efx_rx_buffer *rx_buf,
 460                                      unsigned int n_frags,
 461                                      u8 *eh, int hdr_len)
 462 {
 463         struct efx_nic *efx = channel->efx;
 464         struct sk_buff *skb;
 465
 466         /* Allocate an SKB to store the headers */
 467         skb = netdev_alloc_skb(efx->net_dev,
 468                                efx->rx_ip_align + efx->rx_prefix_size +
 469                                hdr_len);
 470         if (unlikely(skb == NULL)) {
 471                 atomic_inc(&efx->n_rx_noskb_drops);
 472                 return NULL;
 473         }
 474
 475         EFX_WARN_ON_ONCE_PARANOID(rx_buf->len < hdr_len);
 476
 477         memcpy(skb->data + efx->rx_ip_align, eh - efx->rx_prefix_size,
 478                efx->rx_prefix_size + hdr_len);
 479         skb_reserve(skb, efx->rx_ip_align + efx->rx_prefix_size);
 480         __skb_put(skb, hdr_len);
 481
 482         /* Append the remaining page(s) onto the frag list */
 483         if (rx_buf->len > hdr_len) {
 484                 rx_buf->page_offset += hdr_len;
 485                 rx_buf->len -= hdr_len;
 486
 487                 for (;;) {
 488                         skb_fill_page_desc(skb, skb_shinfo(skb)->nr_frags,
 489                                            rx_buf->page, rx_buf->page_offset,
 490                                            rx_buf->len);
 491                         rx_buf->page = NULL;
 492                         skb->len += rx_buf->len;
 493                         skb->data_len += rx_buf->len;
 494                         if (skb_shinfo(skb)->nr_frags == n_frags)
 495                                 break;
 496
 497                         rx_buf = efx_rx_buf_next(&channel->rx_queue, rx_buf);
 498                 }
 499         } else {
 500                 __free_pages(rx_buf->page, efx->rx_buffer_order);
 501                 rx_buf->page = NULL;
 502                 n_frags = 0;
 503         }
 504
 505         skb->truesize += n_frags * efx->rx_buffer_truesize;
 506
 507         /* Move past the ethernet header */
 508         skb->protocol = eth_type_trans(skb, efx->net_dev);
 509
 510         skb_mark_napi_id(skb, &channel->napi_str);
 511
 512         return skb;
 513 }
 514
 515 void efx_rx_packet(struct efx_rx_queue *rx_queue, unsigned int index,
 516                    unsigned int n_frags, unsigned int len, u16 flags)
 517 {
 518         struct efx_nic *efx = rx_queue->efx;
 519         struct efx_channel *channel = efx_rx_queue_channel(rx_queue);
 520         struct efx_rx_buffer *rx_buf;
 521
 522         rx_queue->rx_packets++;
 523
 524         rx_buf = efx_rx_buffer(rx_queue, index);
 525         rx_buf->flags |= flags;
 526
 527         /* Validate the number of fragments and completed length */
 528         if (n_frags == 1) {
 529                 if (!(flags & EFX_RX_PKT_PREFIX_LEN))
 530                         efx_rx_packet__check_len(rx_queue, rx_buf, len);
 531         } else if (unlikely(n_frags > EFX_RX_MAX_FRAGS) ||
 532                    unlikely(len <= (n_frags - 1) * efx->rx_dma_len) ||
 533                    unlikely(len > n_frags * efx->rx_dma_len) ||
 534                    unlikely(!efx->rx_scatter)) {
 535                 /* If this isn't an explicit discard request, either
 536                  * the hardware or the driver is broken.
 537                  */
 538                 WARN_ON(!(len == 0 && rx_buf->flags & EFX_RX_PKT_DISCARD));
 539                 rx_buf->flags |= EFX_RX_PKT_DISCARD;
 540         }
 541
 542         netif_vdbg(efx, rx_status, efx->net_dev,
 543                    "RX queue %d received ids %x-%x len %d %s%s\n",
 544                    efx_rx_queue_index(rx_queue), index,
 545                    (index + n_frags - 1) & rx_queue->ptr_mask, len,
 546                    (rx_buf->flags & EFX_RX_PKT_CSUMMED) ? " [SUMMED]" : "",
 547                    (rx_buf->flags & EFX_RX_PKT_DISCARD) ? " [DISCARD]" : "");
 548
 549         /* Discard packet, if instructed to do so.  Process the
 550          * previous receive first.
 551          */
 552         if (unlikely(rx_buf->flags & EFX_RX_PKT_DISCARD)) {
 553                 efx_rx_flush_packet(channel);
 554                 efx_discard_rx_packet(channel, rx_buf, n_frags);
 555                 return;
 556         }
 557
 558         if (n_frags == 1 && !(flags & EFX_RX_PKT_PREFIX_LEN))
 559                 rx_buf->len = len;
 560
 561         /* Release and/or sync the DMA mapping - assumes all RX buffers
 562          * consumed in-order per RX queue.
 563          */
 564         efx_sync_rx_buffer(efx, rx_buf, rx_buf->len);
 565
 566         /* Prefetch nice and early so data will (hopefully) be in cache by
 567          * the time we look at it.
 568          */
 569         prefetch(efx_rx_buf_va(rx_buf));
 570
 571         rx_buf->page_offset += efx->rx_prefix_size;
 572         rx_buf->len -= efx->rx_prefix_size;
 573
 574         if (n_frags > 1) {
 575                 /* Release/sync DMA mapping for additional fragments.
 576                  * Fix length for last fragment.
 577                  */
 578                 unsigned int tail_frags = n_frags - 1;
 579
 580                 for (;;) {
 581                         rx_buf = efx_rx_buf_next(rx_queue, rx_buf);
 582                         if (--tail_frags == 0)
 583                                 break;
 584                         efx_sync_rx_buffer(efx, rx_buf, efx->rx_dma_len);
 585                 }
 586                 rx_buf->len = len - (n_frags - 1) * efx->rx_dma_len;
 587                 efx_sync_rx_buffer(efx, rx_buf, rx_buf->len);
 588         }
 589
 590         /* All fragments have been DMA-synced, so recycle pages. */
 591         rx_buf = efx_rx_buffer(rx_queue, index);
 592         efx_recycle_rx_pages(channel, rx_buf, n_frags);
 593
 594         /* Pipeline receives so that we give time for packet headers to be
 595          * prefetched into cache.
 596          */
 597         efx_rx_flush_packet(channel);
 598         channel->rx_pkt_n_frags = n_frags;
 599         channel->rx_pkt_index = index;
 600 }
 601
 602 static void efx_rx_deliver(struct efx_channel *channel, u8 *eh,
 603                            struct efx_rx_buffer *rx_buf,
 604                            unsigned int n_frags)
 605 {
 606         struct sk_buff *skb;
 607         u16 hdr_len = min_t(u16, rx_buf->len, EFX_SKB_HEADERS);
 608
 609         skb = efx_rx_mk_skb(channel, rx_buf, n_frags, eh, hdr_len);
 610         if (unlikely(skb == NULL)) {
 611                 struct efx_rx_queue *rx_queue;
 612
 613                 rx_queue = efx_channel_get_rx_queue(channel);
 614                 efx_free_rx_buffers(rx_queue, rx_buf, n_frags);
 615                 return;
 616         }
 617         skb_record_rx_queue(skb, channel->rx_queue.core_index);
 618
 619         /* Set the SKB flags */
 620         skb_checksum_none_assert(skb);
 621         if (likely(rx_buf->flags & EFX_RX_PKT_CSUMMED)) {
 622                 skb->ip_summed = CHECKSUM_UNNECESSARY;
 623                 skb->csum_level = !!(rx_buf->flags & EFX_RX_PKT_CSUM_LEVEL);
 624         }
 625
 626         efx_rx_skb_attach_timestamp(channel, skb);
 627
 628         if (channel->type->receive_skb)
 629                 if (channel->type->receive_skb(channel, skb))
 630                         return;
 631
 632         /* Pass the packet up */
 633         if (channel->rx_list != NULL)
 634                 /* Add to list, will pass up later */
 635                 list_add_tail(&skb->list, channel->rx_list);
 636         else
 637                 /* No list, so pass it up now */
 638                 netif_receive_skb(skb);
 639 }
 640
 641 /* Handle a received packet.  Second half: Touches packet payload. */
 642 void __efx_rx_packet(struct efx_channel *channel)
 643 {
 644         struct efx_nic *efx = channel->efx;
 645         struct efx_rx_buffer *rx_buf =
 646                 efx_rx_buffer(&channel->rx_queue, channel->rx_pkt_index);
 647         u8 *eh = efx_rx_buf_va(rx_buf);
 648
 649         /* Read length from the prefix if necessary.  This already
 650          * excludes the length of the prefix itself.
 651          */
 652         if (rx_buf->flags & EFX_RX_PKT_PREFIX_LEN)
 653                 rx_buf->len = le16_to_cpup((__le16 *)
 654                                            (eh + efx->rx_packet_len_offset));
 655
 656         /* If we're in loopback test, then pass the packet directly to the
 657          * loopback layer, and free the rx_buf here
 658          */
 659         if (unlikely(efx->loopback_selftest)) {
 660                 struct efx_rx_queue *rx_queue;
 661
 662                 efx_loopback_rx_packet(efx, eh, rx_buf->len);
 663                 rx_queue = efx_channel_get_rx_queue(channel);
 664                 efx_free_rx_buffers(rx_queue, rx_buf,
 665                                     channel->rx_pkt_n_frags);
 666                 goto out;
 667         }
 668
 669         if (unlikely(!(efx->net_dev->features & NETIF_F_RXCSUM)))
 670                 rx_buf->flags &= ~EFX_RX_PKT_CSUMMED;
 671
 672         if ((rx_buf->flags & EFX_RX_PKT_TCP) && !channel->type->receive_skb)
 673                 efx_rx_packet_gro(channel, rx_buf, channel->rx_pkt_n_frags, eh);
 674         else
 675                 efx_rx_deliver(channel, eh, rx_buf, channel->rx_pkt_n_frags);
 676 out:
 677         channel->rx_pkt_n_frags = 0;
 678 }
 679
 680 int efx_probe_rx_queue(struct efx_rx_queue *rx_queue)
 681 {
 682         struct efx_nic *efx = rx_queue->efx;
 683         unsigned int entries;
 684         int rc;
 685
 686         /* Create the smallest power-of-two aligned ring */
 687         entries = max(roundup_pow_of_two(efx->rxq_entries), EFX_MIN_DMAQ_SIZE);
 688         EFX_WARN_ON_PARANOID(entries > EFX_MAX_DMAQ_SIZE);
 689         rx_queue->ptr_mask = entries - 1;
 690
 691         netif_dbg(efx, probe, efx->net_dev,
 692                   "creating RX queue %d size %#x mask %#x\n",
 693                   efx_rx_queue_index(rx_queue), efx->rxq_entries,
 694                   rx_queue->ptr_mask);
 695
 696         /* Allocate RX buffers */
 697         rx_queue->buffer = kcalloc(entries, sizeof(*rx_queue->buffer),
 698                                    GFP_KERNEL);
 699         if (!rx_queue->buffer)
 700                 return -ENOMEM;
 701
 702         rc = efx_nic_probe_rx(rx_queue);
 703         if (rc) {
 704                 kfree(rx_queue->buffer);
 705                 rx_queue->buffer = NULL;
 706         }
 707
 708         return rc;
 709 }
 710
 711 static void efx_init_rx_recycle_ring(struct efx_nic *efx,
 712                                      struct efx_rx_queue *rx_queue)
 713 {
 714         unsigned int bufs_in_recycle_ring, page_ring_size;
 715
 716         /* Set the RX recycle ring size */
 717 #ifdef CONFIG_PPC64
 718         bufs_in_recycle_ring = EFX_RECYCLE_RING_SIZE_IOMMU;
 719 #else
 720         if (iommu_present(&pci_bus_type))
 721                 bufs_in_recycle_ring = EFX_RECYCLE_RING_SIZE_IOMMU;
 722         else
 723                 bufs_in_recycle_ring = EFX_RECYCLE_RING_SIZE_NOIOMMU;
 724 #endif /* CONFIG_PPC64 */
 725
 726         page_ring_size = roundup_pow_of_two(bufs_in_recycle_ring /
 727                                             efx->rx_bufs_per_page);
 728         rx_queue->page_ring = kcalloc(page_ring_size,
 729                                       sizeof(*rx_queue->page_ring), GFP_KERNEL);
 730         rx_queue->page_ptr_mask = page_ring_size - 1;
 731 }
 732
 733 void efx_init_rx_queue(struct efx_rx_queue *rx_queue)
 734 {
 735         struct efx_nic *efx = rx_queue->efx;
 736         unsigned int max_fill, trigger, max_trigger;
 737
 738         netif_dbg(rx_queue->efx, drv, rx_queue->efx->net_dev,
 739                   "initialising RX queue %d\n", efx_rx_queue_index(rx_queue));
 740
 741         /* Initialise ptr fields */
 742         rx_queue->added_count = 0;
 743         rx_queue->notified_count = 0;
 744         rx_queue->removed_count = 0;
 745         rx_queue->min_fill = -1U;
 746         efx_init_rx_recycle_ring(efx, rx_queue);
 747
 748         rx_queue->page_remove = 0;
 749         rx_queue->page_add = rx_queue->page_ptr_mask + 1;
 750         rx_queue->page_recycle_count = 0;
 751         rx_queue->page_recycle_failed = 0;
 752         rx_queue->page_recycle_full = 0;
 753
 754         /* Initialise limit fields */
 755         max_fill = efx->rxq_entries - EFX_RXD_HEAD_ROOM;
 756         max_trigger =
 757                 max_fill - efx->rx_pages_per_batch * efx->rx_bufs_per_page;
 758         if (rx_refill_threshold != 0) {
 759                 trigger = max_fill * min(rx_refill_threshold, 100U) / 100U;
 760                 if (trigger > max_trigger)
 761                         trigger = max_trigger;
 762         } else {
 763                 trigger = max_trigger;
 764         }
 765
 766         rx_queue->max_fill = max_fill;
 767         rx_queue->fast_fill_trigger = trigger;
 768         rx_queue->refill_enabled = true;
 769
 770         /* Set up RX descriptor ring */
 771         efx_nic_init_rx(rx_queue);
 772 }
 773
 774 void efx_fini_rx_queue(struct efx_rx_queue *rx_queue)
 775 {
 776         int i;
 777         struct efx_nic *efx = rx_queue->efx;
 778         struct efx_rx_buffer *rx_buf;
 779
 780         netif_dbg(rx_queue->efx, drv, rx_queue->efx->net_dev,
 781                   "shutting down RX queue %d\n", efx_rx_queue_index(rx_queue));
 782
 783         del_timer_sync(&rx_queue->slow_fill);
 784
 785         /* Release RX buffers from the current read ptr to the write ptr */
 786         if (rx_queue->buffer) {
 787                 for (i = rx_queue->removed_count; i < rx_queue->added_count;
 788                      i++) {
 789                         unsigned index = i & rx_queue->ptr_mask;
 790                         rx_buf = efx_rx_buffer(rx_queue, index);
 791                         efx_fini_rx_buffer(rx_queue, rx_buf);
 792                 }
 793         }
 794
 795         /* Unmap and release the pages in the recycle ring. Remove the ring. */
 796         for (i = 0; i <= rx_queue->page_ptr_mask; i++) {
 797                 struct page *page = rx_queue->page_ring[i];
 798                 struct efx_rx_page_state *state;
 799
 800                 if (page == NULL)
 801                         continue;
 802
 803                 state = page_address(page);
 804                 dma_unmap_page(&efx->pci_dev->dev, state->dma_addr,
 805                                PAGE_SIZE << efx->rx_buffer_order,
 806                                DMA_FROM_DEVICE);
 807                 put_page(page);
 808         }
 809         kfree(rx_queue->page_ring);
 810         rx_queue->page_ring = NULL;
 811 }
 812
 813 void efx_remove_rx_queue(struct efx_rx_queue *rx_queue)
 814 {
 815         netif_dbg(rx_queue->efx, drv, rx_queue->efx->net_dev,
 816                   "destroying RX queue %d\n", efx_rx_queue_index(rx_queue));
 817
 818         efx_nic_remove_rx(rx_queue);
 819
 820         kfree(rx_queue->buffer);
 821         rx_queue->buffer = NULL;
 822 }
 823
 824
 825 module_param(rx_refill_threshold, uint, 0444);
 826 MODULE_PARM_DESC(rx_refill_threshold,
 827                  "RX descriptor ring refill threshold (%)");
 828
 829 #ifdef CONFIG_RFS_ACCEL
 830
 831 static void efx_filter_rfs_work(struct work_struct *data)
 832 {
 833         struct efx_async_filter_insertion *req = container_of(data, struct efx_async_filter_insertion,
 834                                                               work);
 835         struct efx_nic *efx = netdev_priv(req->net_dev);
 836         struct efx_channel *channel = efx_get_channel(efx, req->rxq_index);
 837         int slot_idx = req - efx->rps_slot;
 838         struct efx_arfs_rule *rule;
 839         u16 arfs_id = 0;
 840         int rc;
 841
 842         rc = efx->type->filter_insert(efx, &req->spec, true);
 843         if (rc >= 0)
 844                 rc %= efx->type->max_rx_ip_filters;
 845         if (efx->rps_hash_table) {
 846                 spin_lock_bh(&efx->rps_hash_lock);
 847                 rule = efx_rps_hash_find(efx, &req->spec);
 848                 /* The rule might have already gone, if someone else's request
 849                  * for the same spec was already worked and then expired before
 850                  * we got around to our work.  In that case we have nothing
 851                  * tying us to an arfs_id, meaning that as soon as the filter
 852                  * is considered for expiry it will be removed.
 853                  */
 854                 if (rule) {
 855                         if (rc < 0)
 856                                 rule->filter_id = EFX_ARFS_FILTER_ID_ERROR;
 857                         else
 858                                 rule->filter_id = rc;
 859                         arfs_id = rule->arfs_id;
 860                 }
 861                 spin_unlock_bh(&efx->rps_hash_lock);
 862         }
 863         if (rc >= 0) {
 864                 /* Remember this so we can check whether to expire the filter
 865                  * later.
 866                  */
 867                 mutex_lock(&efx->rps_mutex);
 868                 channel->rps_flow_id[rc] = req->flow_id;
 869                 ++channel->rfs_filters_added;
 870                 mutex_unlock(&efx->rps_mutex);
 871
 872                 if (req->spec.ether_type == htons(ETH_P_IP))
 873                         netif_info(efx, rx_status, efx->net_dev,
 874                                    "steering %s %pI4:%u:%pI4:%u to queue %u [flow %u filter %d id %u]\n",
 875                                    (req->spec.ip_proto == IPPROTO_TCP) ? "TCP" : "UDP",
 876                                    req->spec.rem_host, ntohs(req->spec.rem_port),
 877                                    req->spec.loc_host, ntohs(req->spec.loc_port),
 878                                    req->rxq_index, req->flow_id, rc, arfs_id);
 879                 else
 880                         netif_info(efx, rx_status, efx->net_dev,
 881                                    "steering %s [%pI6]:%u:[%pI6]:%u to queue %u [flow %u filter %d id %u]\n",
 882                                    (req->spec.ip_proto == IPPROTO_TCP) ? "TCP" : "UDP",
 883                                    req->spec.rem_host, ntohs(req->spec.rem_port),
 884                                    req->spec.loc_host, ntohs(req->spec.loc_port),
 885                                    req->rxq_index, req->flow_id, rc, arfs_id);
 886         }
 887
 888         /* Release references */
 889         clear_bit(slot_idx, &efx->rps_slot_map);
 890         dev_put(req->net_dev);
 891 }
 892
 893 int efx_filter_rfs(struct net_device *net_dev, const struct sk_buff *skb,
 894                    u16 rxq_index, u32 flow_id)
 895 {
 896         struct efx_nic *efx = netdev_priv(net_dev);
 897         struct efx_async_filter_insertion *req;
 898         struct efx_arfs_rule *rule;
 899         struct flow_keys fk;
 900         int slot_idx;
 901         bool new;
 902         int rc;
 903
 904         /* find a free slot */
 905         for (slot_idx = 0; slot_idx < EFX_RPS_MAX_IN_FLIGHT; slot_idx++)
 906                 if (!test_and_set_bit(slot_idx, &efx->rps_slot_map))
 907                         break;
 908         if (slot_idx >= EFX_RPS_MAX_IN_FLIGHT)
 909                 return -EBUSY;
 910
 911         if (flow_id == RPS_FLOW_ID_INVALID) {
 912                 rc = -EINVAL;
 913                 goto out_clear;
 914         }
 915
 916         if (!skb_flow_dissect_flow_keys(skb, &fk, 0)) {
 917                 rc = -EPROTONOSUPPORT;
 918                 goto out_clear;
 919         }
 920
 921         if (fk.basic.n_proto != htons(ETH_P_IP) && fk.basic.n_proto != htons(ETH_P_IPV6)) {
 922                 rc = -EPROTONOSUPPORT;
 923                 goto out_clear;
 924         }
 925         if (fk.control.flags & FLOW_DIS_IS_FRAGMENT) {
 926                 rc = -EPROTONOSUPPORT;
 927                 goto out_clear;
 928         }
 929
 930         req = efx->rps_slot + slot_idx;
 931         efx_filter_init_rx(&req->spec, EFX_FILTER_PRI_HINT,
 932                            efx->rx_scatter ? EFX_FILTER_FLAG_RX_SCATTER : 0,
 933                            rxq_index);
 934         req->spec.match_flags =
 935                 EFX_FILTER_MATCH_ETHER_TYPE | EFX_FILTER_MATCH_IP_PROTO |
 936                 EFX_FILTER_MATCH_LOC_HOST | EFX_FILTER_MATCH_LOC_PORT |
 937                 EFX_FILTER_MATCH_REM_HOST | EFX_FILTER_MATCH_REM_PORT;
 938         req->spec.ether_type = fk.basic.n_proto;
 939         req->spec.ip_proto = fk.basic.ip_proto;
 940
 941         if (fk.basic.n_proto == htons(ETH_P_IP)) {
 942                 req->spec.rem_host[0] = fk.addrs.v4addrs.src;
 943                 req->spec.loc_host[0] = fk.addrs.v4addrs.dst;
 944         } else {
 945                 memcpy(req->spec.rem_host, &fk.addrs.v6addrs.src,
 946                        sizeof(struct in6_addr));
 947                 memcpy(req->spec.loc_host, &fk.addrs.v6addrs.dst,
 948                        sizeof(struct in6_addr));
 949         }
 950
 951         req->spec.rem_port = fk.ports.src;
 952         req->spec.loc_port = fk.ports.dst;
 953
 954         if (efx->rps_hash_table) {
 955                 /* Add it to ARFS hash table */
 956                 spin_lock(&efx->rps_hash_lock);
 957                 rule = efx_rps_hash_add(efx, &req->spec, &new);
 958                 if (!rule) {
 959                         rc = -ENOMEM;
 960                         goto out_unlock;
 961                 }
 962                 if (new)
 963                         rule->arfs_id = efx->rps_next_id++ % RPS_NO_FILTER;
 964                 rc = rule->arfs_id;
 965                 /* Skip if existing or pending filter already does the right thing */
 966                 if (!new && rule->rxq_index == rxq_index &&
 967                     rule->filter_id >= EFX_ARFS_FILTER_ID_PENDING)
 968                         goto out_unlock;
 969                 rule->rxq_index = rxq_index;
 970                 rule->filter_id = EFX_ARFS_FILTER_ID_PENDING;
 971                 spin_unlock(&efx->rps_hash_lock);
 972         } else {
 973                 /* Without an ARFS hash table, we just use arfs_id 0 for all
 974                  * filters.  This means if multiple flows hash to the same
 975                  * flow_id, all but the most recently touched will be eligible
 976                  * for expiry.
 977                  */
 978                 rc = 0;
 979         }
 980
 981         /* Queue the request */
 982         dev_hold(req->net_dev = net_dev);
 983         INIT_WORK(&req->work, efx_filter_rfs_work);
 984         req->rxq_index = rxq_index;
 985         req->flow_id = flow_id;
 986         schedule_work(&req->work);
 987         return rc;
 988 out_unlock:
 989         spin_unlock(&efx->rps_hash_lock);
 990 out_clear:
 991         clear_bit(slot_idx, &efx->rps_slot_map);
 992         return rc;
 993 }
 994
 995 bool __efx_filter_rfs_expire(struct efx_nic *efx, unsigned int quota)
 996 {
 997         bool (*expire_one)(struct efx_nic *efx, u32 flow_id, unsigned int index);
 998         unsigned int channel_idx, index, size;
 999         u32 flow_id;
1000
1001         if (!mutex_trylock(&efx->rps_mutex))
1002                 return false;
1003         expire_one = efx->type->filter_rfs_expire_one;
1004         channel_idx = efx->rps_expire_channel;
1005         index = efx->rps_expire_index;
1006         size = efx->type->max_rx_ip_filters;
1007         while (quota--) {
1008                 struct efx_channel *channel = efx_get_channel(efx, channel_idx);
1009                 flow_id = channel->rps_flow_id[index];
1010
1011                 if (flow_id != RPS_FLOW_ID_INVALID &&
1012                     expire_one(efx, flow_id, index)) {
1013                         netif_info(efx, rx_status, efx->net_dev,
1014                                    "expired filter %d [queue %u flow %u]\n",
1015                                    index, channel_idx, flow_id);
1016                         channel->rps_flow_id[index] = RPS_FLOW_ID_INVALID;
1017                 }
1018                 if (++index == size) {
1019                         if (++channel_idx == efx->n_channels)
1020                                 channel_idx = 0;
1021                         index = 0;
1022                 }
1023         }
1024         efx->rps_expire_channel = channel_idx;
1025         efx->rps_expire_index = index;
1026
1027         mutex_unlock(&efx->rps_mutex);
1028         return true;
1029 }
1030
1031 #endif /* CONFIG_RFS_ACCEL */
1032
1033 /**
1034  * efx_filter_is_mc_recipient - test whether spec is a multicast recipient
1035  * @spec: Specification to test
1036  *
1037  * Return: %true if the specification is a non-drop RX filter that
1038  * matches a local MAC address I/G bit value of 1 or matches a local
1039  * IPv4 or IPv6 address value in the respective multicast address
1040  * range.  Otherwise %false.
1041  */
1042 bool efx_filter_is_mc_recipient(const struct efx_filter_spec *spec)
1043 {
1044         if (!(spec->flags & EFX_FILTER_FLAG_RX) ||
1045             spec->dmaq_id == EFX_FILTER_RX_DMAQ_ID_DROP)
1046                 return false;
1047
1048         if (spec->match_flags &
1049             (EFX_FILTER_MATCH_LOC_MAC | EFX_FILTER_MATCH_LOC_MAC_IG) &&
1050             is_multicast_ether_addr(spec->loc_mac))
1051                 return true;
1052
1053         if ((spec->match_flags &
1054              (EFX_FILTER_MATCH_ETHER_TYPE | EFX_FILTER_MATCH_LOC_HOST)) ==
1055             (EFX_FILTER_MATCH_ETHER_TYPE | EFX_FILTER_MATCH_LOC_HOST)) {
1056                 if (spec->ether_type == htons(ETH_P_IP) &&
1057                     ipv4_is_multicast(spec->loc_host[0]))
1058                         return true;
1059                 if (spec->ether_type == htons(ETH_P_IPV6) &&
1060                     ((const u8 *)spec->loc_host)[0] == 0xff)
1061                         return true;
1062         }
1063
1064         return false;
1065 }