drivers/net/ethernet/sfc/falcon/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 EF4_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 EF4_RECYCLE_RING_SIZE_IOMMU 4096
  35 #define EF4_RECYCLE_RING_SIZE_NOIOMMU (2 * EF4_RX_PREFERRED_BATCH)
  36
  37 /* Size of buffer allocated for skb header area. */
  38 #define EF4_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 EF4_RX_MAX_FRAGS DIV_ROUND_UP(EF4_MAX_FRAME_LEN(EF4_MAX_MTU), \
  47                                       EF4_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 EF4_RXD_HEAD_ROOM (1 + EF4_RX_MAX_FRAGS)
  56
  57 static inline u8 *ef4_rx_buf_va(struct ef4_rx_buffer *buf)
  58 {
  59         return page_address(buf->page) + buf->page_offset;
  60 }
  61
  62 static inline u32 ef4_rx_buf_hash(struct ef4_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 ef4_rx_buffer *
  76 ef4_rx_buf_next(struct ef4_rx_queue *rx_queue, struct ef4_rx_buffer *rx_buf)
  77 {
  78         if (unlikely(rx_buf == ef4_rx_buffer(rx_queue, rx_queue->ptr_mask)))
  79                 return ef4_rx_buffer(rx_queue, 0);
  80         else
  81                 return rx_buf + 1;
  82 }
  83
  84 static inline void ef4_sync_rx_buffer(struct ef4_nic *efx,
  85                                       struct ef4_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 ef4_rx_config_page_split(struct ef4_nic *efx)
  93 {
  94         efx->rx_page_buf_step = ALIGN(efx->rx_dma_len + efx->rx_ip_align,
  95                                       EF4_RX_BUF_ALIGNMENT);
  96         efx->rx_bufs_per_page = efx->rx_buffer_order ? 1 :
  97                 ((PAGE_SIZE - sizeof(struct ef4_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(EF4_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 *ef4_reuse_page(struct ef4_rx_queue *rx_queue)
 107 {
 108         struct ef4_nic *efx = rx_queue->efx;
 109         struct page *page;
 110         struct ef4_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  * ef4_init_rx_buffers - create EF4_RX_BATCH page-based RX buffers
 141  *
 142  * @rx_queue:           Efx RX queue
 143  * @atomic:             control memory allocation flags
 144  *
 145  * This allocates a batch of pages, maps them for DMA, and populates
 146  * struct ef4_rx_buffers for each one. Return a negative error code or
 147  * 0 on success. If a single page can be used for multiple buffers,
 148  * then the page will either be inserted fully, or not at all.
 149  */
 150 static int ef4_init_rx_buffers(struct ef4_rx_queue *rx_queue, bool atomic)
 151 {
 152         struct ef4_nic *efx = rx_queue->efx;
 153         struct ef4_rx_buffer *rx_buf;
 154         struct page *page;
 155         unsigned int page_offset;
 156         struct ef4_rx_page_state *state;
 157         dma_addr_t dma_addr;
 158         unsigned index, count;
 159
 160         count = 0;
 161         do {
 162                 page = ef4_reuse_page(rx_queue);
 163                 if (page == NULL) {
 164                         page = alloc_pages(__GFP_COMP |
 165                                            (atomic ? GFP_ATOMIC : GFP_KERNEL),
 166                                            efx->rx_buffer_order);
 167                         if (unlikely(page == NULL))
 168                                 return -ENOMEM;
 169                         dma_addr =
 170                                 dma_map_page(&efx->pci_dev->dev, page, 0,
 171                                              PAGE_SIZE << efx->rx_buffer_order,
 172                                              DMA_FROM_DEVICE);
 173                         if (unlikely(dma_mapping_error(&efx->pci_dev->dev,
 174                                                        dma_addr))) {
 175                                 __free_pages(page, efx->rx_buffer_order);
 176                                 return -EIO;
 177                         }
 178                         state = page_address(page);
 179                         state->dma_addr = dma_addr;
 180                 } else {
 181                         state = page_address(page);
 182                         dma_addr = state->dma_addr;
 183                 }
 184
 185                 dma_addr += sizeof(struct ef4_rx_page_state);
 186                 page_offset = sizeof(struct ef4_rx_page_state);
 187
 188                 do {
 189                         index = rx_queue->added_count & rx_queue->ptr_mask;
 190                         rx_buf = ef4_rx_buffer(rx_queue, index);
 191                         rx_buf->dma_addr = dma_addr + efx->rx_ip_align;
 192                         rx_buf->page = page;
 193                         rx_buf->page_offset = page_offset + efx->rx_ip_align;
 194                         rx_buf->len = efx->rx_dma_len;
 195                         rx_buf->flags = 0;
 196                         ++rx_queue->added_count;
 197                         get_page(page);
 198                         dma_addr += efx->rx_page_buf_step;
 199                         page_offset += efx->rx_page_buf_step;
 200                 } while (page_offset + efx->rx_page_buf_step <= PAGE_SIZE);
 201
 202                 rx_buf->flags = EF4_RX_BUF_LAST_IN_PAGE;
 203         } while (++count < efx->rx_pages_per_batch);
 204
 205         return 0;
 206 }
 207
 208 /* Unmap a DMA-mapped page.  This function is only called for the final RX
 209  * buffer in a page.
 210  */
 211 static void ef4_unmap_rx_buffer(struct ef4_nic *efx,
 212                                 struct ef4_rx_buffer *rx_buf)
 213 {
 214         struct page *page = rx_buf->page;
 215
 216         if (page) {
 217                 struct ef4_rx_page_state *state = page_address(page);
 218                 dma_unmap_page(&efx->pci_dev->dev,
 219                                state->dma_addr,
 220                                PAGE_SIZE << efx->rx_buffer_order,
 221                                DMA_FROM_DEVICE);
 222         }
 223 }
 224
 225 static void ef4_free_rx_buffers(struct ef4_rx_queue *rx_queue,
 226                                 struct ef4_rx_buffer *rx_buf,
 227                                 unsigned int num_bufs)
 228 {
 229         do {
 230                 if (rx_buf->page) {
 231                         put_page(rx_buf->page);
 232                         rx_buf->page = NULL;
 233                 }
 234                 rx_buf = ef4_rx_buf_next(rx_queue, rx_buf);
 235         } while (--num_bufs);
 236 }
 237
 238 /* Attempt to recycle the page if there is an RX recycle ring; the page can
 239  * only be added if this is the final RX buffer, to prevent pages being used in
 240  * the descriptor ring and appearing in the recycle ring simultaneously.
 241  */
 242 static void ef4_recycle_rx_page(struct ef4_channel *channel,
 243                                 struct ef4_rx_buffer *rx_buf)
 244 {
 245         struct page *page = rx_buf->page;
 246         struct ef4_rx_queue *rx_queue = ef4_channel_get_rx_queue(channel);
 247         struct ef4_nic *efx = rx_queue->efx;
 248         unsigned index;
 249
 250         /* Only recycle the page after processing the final buffer. */
 251         if (!(rx_buf->flags & EF4_RX_BUF_LAST_IN_PAGE))
 252                 return;
 253
 254         index = rx_queue->page_add & rx_queue->page_ptr_mask;
 255         if (rx_queue->page_ring[index] == NULL) {
 256                 unsigned read_index = rx_queue->page_remove &
 257                         rx_queue->page_ptr_mask;
 258
 259                 /* The next slot in the recycle ring is available, but
 260                  * increment page_remove if the read pointer currently
 261                  * points here.
 262                  */
 263                 if (read_index == index)
 264                         ++rx_queue->page_remove;
 265                 rx_queue->page_ring[index] = page;
 266                 ++rx_queue->page_add;
 267                 return;
 268         }
 269         ++rx_queue->page_recycle_full;
 270         ef4_unmap_rx_buffer(efx, rx_buf);
 271         put_page(rx_buf->page);
 272 }
 273
 274 static void ef4_fini_rx_buffer(struct ef4_rx_queue *rx_queue,
 275                                struct ef4_rx_buffer *rx_buf)
 276 {
 277         /* Release the page reference we hold for the buffer. */
 278         if (rx_buf->page)
 279                 put_page(rx_buf->page);
 280
 281         /* If this is the last buffer in a page, unmap and free it. */
 282         if (rx_buf->flags & EF4_RX_BUF_LAST_IN_PAGE) {
 283                 ef4_unmap_rx_buffer(rx_queue->efx, rx_buf);
 284                 ef4_free_rx_buffers(rx_queue, rx_buf, 1);
 285         }
 286         rx_buf->page = NULL;
 287 }
 288
 289 /* Recycle the pages that are used by buffers that have just been received. */
 290 static void ef4_recycle_rx_pages(struct ef4_channel *channel,
 291                                  struct ef4_rx_buffer *rx_buf,
 292                                  unsigned int n_frags)
 293 {
 294         struct ef4_rx_queue *rx_queue = ef4_channel_get_rx_queue(channel);
 295
 296         do {
 297                 ef4_recycle_rx_page(channel, rx_buf);
 298                 rx_buf = ef4_rx_buf_next(rx_queue, rx_buf);
 299         } while (--n_frags);
 300 }
 301
 302 static void ef4_discard_rx_packet(struct ef4_channel *channel,
 303                                   struct ef4_rx_buffer *rx_buf,
 304                                   unsigned int n_frags)
 305 {
 306         struct ef4_rx_queue *rx_queue = ef4_channel_get_rx_queue(channel);
 307
 308         ef4_recycle_rx_pages(channel, rx_buf, n_frags);
 309
 310         ef4_free_rx_buffers(rx_queue, rx_buf, n_frags);
 311 }
 312
 313 /**
 314  * ef4_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  * @atomic: control memory allocation flags
 321  *
 322  * The caller must provide serialisation (none is used here). In practise,
 323  * this means this function must run from the NAPI handler, or be called
 324  * when NAPI is disabled.
 325  */
 326 void ef4_fast_push_rx_descriptors(struct ef4_rx_queue *rx_queue, bool atomic)
 327 {
 328         struct ef4_nic *efx = rx_queue->efx;
 329         unsigned int fill_level, batch_size;
 330         int space, rc = 0;
 331
 332         if (!rx_queue->refill_enabled)
 333                 return;
 334
 335         /* Calculate current fill level, and exit if we don't need to fill */
 336         fill_level = (rx_queue->added_count - rx_queue->removed_count);
 337         EF4_BUG_ON_PARANOID(fill_level > rx_queue->efx->rxq_entries);
 338         if (fill_level >= rx_queue->fast_fill_trigger)
 339                 goto out;
 340
 341         /* Record minimum fill level */
 342         if (unlikely(fill_level < rx_queue->min_fill)) {
 343                 if (fill_level)
 344                         rx_queue->min_fill = fill_level;
 345         }
 346
 347         batch_size = efx->rx_pages_per_batch * efx->rx_bufs_per_page;
 348         space = rx_queue->max_fill - fill_level;
 349         EF4_BUG_ON_PARANOID(space < batch_size);
 350
 351         netif_vdbg(rx_queue->efx, rx_status, rx_queue->efx->net_dev,
 352                    "RX queue %d fast-filling descriptor ring from"
 353                    " level %d to level %d\n",
 354                    ef4_rx_queue_index(rx_queue), fill_level,
 355                    rx_queue->max_fill);
 356
 357
 358         do {
 359                 rc = ef4_init_rx_buffers(rx_queue, atomic);
 360                 if (unlikely(rc)) {
 361                         /* Ensure that we don't leave the rx queue empty */
 362                         if (rx_queue->added_count == rx_queue->removed_count)
 363                                 ef4_schedule_slow_fill(rx_queue);
 364                         goto out;
 365                 }
 366         } while ((space -= batch_size) >= batch_size);
 367
 368         netif_vdbg(rx_queue->efx, rx_status, rx_queue->efx->net_dev,
 369                    "RX queue %d fast-filled descriptor ring "
 370                    "to level %d\n", ef4_rx_queue_index(rx_queue),
 371                    rx_queue->added_count - rx_queue->removed_count);
 372
 373  out:
 374         if (rx_queue->notified_count != rx_queue->added_count)
 375                 ef4_nic_notify_rx_desc(rx_queue);
 376 }
 377
 378 void ef4_rx_slow_fill(struct timer_list *t)
 379 {
 380         struct ef4_rx_queue *rx_queue = from_timer(rx_queue, t, slow_fill);
 381
 382         /* Post an event to cause NAPI to run and refill the queue */
 383         ef4_nic_generate_fill_event(rx_queue);
 384         ++rx_queue->slow_fill_count;
 385 }
 386
 387 static void ef4_rx_packet__check_len(struct ef4_rx_queue *rx_queue,
 388                                      struct ef4_rx_buffer *rx_buf,
 389                                      int len)
 390 {
 391         struct ef4_nic *efx = rx_queue->efx;
 392         unsigned max_len = rx_buf->len - efx->type->rx_buffer_padding;
 393
 394         if (likely(len <= max_len))
 395                 return;
 396
 397         /* The packet must be discarded, but this is only a fatal error
 398          * if the caller indicated it was
 399          */
 400         rx_buf->flags |= EF4_RX_PKT_DISCARD;
 401
 402         if ((len > rx_buf->len) && EF4_WORKAROUND_8071(efx)) {
 403                 if (net_ratelimit())
 404                         netif_err(efx, rx_err, efx->net_dev,
 405                                   " RX queue %d seriously overlength "
 406                                   "RX event (0x%x > 0x%x+0x%x). Leaking\n",
 407                                   ef4_rx_queue_index(rx_queue), len, max_len,
 408                                   efx->type->rx_buffer_padding);
 409                 ef4_schedule_reset(efx, RESET_TYPE_RX_RECOVERY);
 410         } else {
 411                 if (net_ratelimit())
 412                         netif_err(efx, rx_err, efx->net_dev,
 413                                   " RX queue %d overlength RX event "
 414                                   "(0x%x > 0x%x)\n",
 415                                   ef4_rx_queue_index(rx_queue), len, max_len);
 416         }
 417
 418         ef4_rx_queue_channel(rx_queue)->n_rx_overlength++;
 419 }
 420
 421 /* Pass a received packet up through GRO.  GRO can handle pages
 422  * regardless of checksum state and skbs with a good checksum.
 423  */
 424 static void
 425 ef4_rx_packet_gro(struct ef4_channel *channel, struct ef4_rx_buffer *rx_buf,
 426                   unsigned int n_frags, u8 *eh)
 427 {
 428         struct napi_struct *napi = &channel->napi_str;
 429         struct ef4_nic *efx = channel->efx;
 430         struct sk_buff *skb;
 431
 432         skb = napi_get_frags(napi);
 433         if (unlikely(!skb)) {
 434                 struct ef4_rx_queue *rx_queue;
 435
 436                 rx_queue = ef4_channel_get_rx_queue(channel);
 437                 ef4_free_rx_buffers(rx_queue, rx_buf, n_frags);
 438                 return;
 439         }
 440
 441         if (efx->net_dev->features & NETIF_F_RXHASH)
 442                 skb_set_hash(skb, ef4_rx_buf_hash(efx, eh),
 443                              PKT_HASH_TYPE_L3);
 444         skb->ip_summed = ((rx_buf->flags & EF4_RX_PKT_CSUMMED) ?
 445                           CHECKSUM_UNNECESSARY : CHECKSUM_NONE);
 446
 447         for (;;) {
 448                 skb_fill_page_desc(skb, skb_shinfo(skb)->nr_frags,
 449                                    rx_buf->page, rx_buf->page_offset,
 450                                    rx_buf->len);
 451                 rx_buf->page = NULL;
 452                 skb->len += rx_buf->len;
 453                 if (skb_shinfo(skb)->nr_frags == n_frags)
 454                         break;
 455
 456                 rx_buf = ef4_rx_buf_next(&channel->rx_queue, rx_buf);
 457         }
 458
 459         skb->data_len = skb->len;
 460         skb->truesize += n_frags * efx->rx_buffer_truesize;
 461
 462         skb_record_rx_queue(skb, channel->rx_queue.core_index);
 463
 464         napi_gro_frags(napi);
 465 }
 466
 467 /* Allocate and construct an SKB around page fragments */
 468 static struct sk_buff *ef4_rx_mk_skb(struct ef4_channel *channel,
 469                                      struct ef4_rx_buffer *rx_buf,
 470                                      unsigned int n_frags,
 471                                      u8 *eh, int hdr_len)
 472 {
 473         struct ef4_nic *efx = channel->efx;
 474         struct sk_buff *skb;
 475
 476         /* Allocate an SKB to store the headers */
 477         skb = netdev_alloc_skb(efx->net_dev,
 478                                efx->rx_ip_align + efx->rx_prefix_size +
 479                                hdr_len);
 480         if (unlikely(skb == NULL)) {
 481                 atomic_inc(&efx->n_rx_noskb_drops);
 482                 return NULL;
 483         }
 484
 485         EF4_BUG_ON_PARANOID(rx_buf->len < hdr_len);
 486
 487         memcpy(skb->data + efx->rx_ip_align, eh - efx->rx_prefix_size,
 488                efx->rx_prefix_size + hdr_len);
 489         skb_reserve(skb, efx->rx_ip_align + efx->rx_prefix_size);
 490         __skb_put(skb, hdr_len);
 491
 492         /* Append the remaining page(s) onto the frag list */
 493         if (rx_buf->len > hdr_len) {
 494                 rx_buf->page_offset += hdr_len;
 495                 rx_buf->len -= hdr_len;
 496
 497                 for (;;) {
 498                         skb_fill_page_desc(skb, skb_shinfo(skb)->nr_frags,
 499                                            rx_buf->page, rx_buf->page_offset,
 500                                            rx_buf->len);
 501                         rx_buf->page = NULL;
 502                         skb->len += rx_buf->len;
 503                         skb->data_len += rx_buf->len;
 504                         if (skb_shinfo(skb)->nr_frags == n_frags)
 505                                 break;
 506
 507                         rx_buf = ef4_rx_buf_next(&channel->rx_queue, rx_buf);
 508                 }
 509         } else {
 510                 __free_pages(rx_buf->page, efx->rx_buffer_order);
 511                 rx_buf->page = NULL;
 512                 n_frags = 0;
 513         }
 514
 515         skb->truesize += n_frags * efx->rx_buffer_truesize;
 516
 517         /* Move past the ethernet header */
 518         skb->protocol = eth_type_trans(skb, efx->net_dev);
 519
 520         skb_mark_napi_id(skb, &channel->napi_str);
 521
 522         return skb;
 523 }
 524
 525 void ef4_rx_packet(struct ef4_rx_queue *rx_queue, unsigned int index,
 526                    unsigned int n_frags, unsigned int len, u16 flags)
 527 {
 528         struct ef4_nic *efx = rx_queue->efx;
 529         struct ef4_channel *channel = ef4_rx_queue_channel(rx_queue);
 530         struct ef4_rx_buffer *rx_buf;
 531
 532         rx_queue->rx_packets++;
 533
 534         rx_buf = ef4_rx_buffer(rx_queue, index);
 535         rx_buf->flags |= flags;
 536
 537         /* Validate the number of fragments and completed length */
 538         if (n_frags == 1) {
 539                 if (!(flags & EF4_RX_PKT_PREFIX_LEN))
 540                         ef4_rx_packet__check_len(rx_queue, rx_buf, len);
 541         } else if (unlikely(n_frags > EF4_RX_MAX_FRAGS) ||
 542                    unlikely(len <= (n_frags - 1) * efx->rx_dma_len) ||
 543                    unlikely(len > n_frags * efx->rx_dma_len) ||
 544                    unlikely(!efx->rx_scatter)) {
 545                 /* If this isn't an explicit discard request, either
 546                  * the hardware or the driver is broken.
 547                  */
 548                 WARN_ON(!(len == 0 && rx_buf->flags & EF4_RX_PKT_DISCARD));
 549                 rx_buf->flags |= EF4_RX_PKT_DISCARD;
 550         }
 551
 552         netif_vdbg(efx, rx_status, efx->net_dev,
 553                    "RX queue %d received ids %x-%x len %d %s%s\n",
 554                    ef4_rx_queue_index(rx_queue), index,
 555                    (index + n_frags - 1) & rx_queue->ptr_mask, len,
 556                    (rx_buf->flags & EF4_RX_PKT_CSUMMED) ? " [SUMMED]" : "",
 557                    (rx_buf->flags & EF4_RX_PKT_DISCARD) ? " [DISCARD]" : "");
 558
 559         /* Discard packet, if instructed to do so.  Process the
 560          * previous receive first.
 561          */
 562         if (unlikely(rx_buf->flags & EF4_RX_PKT_DISCARD)) {
 563                 ef4_rx_flush_packet(channel);
 564                 ef4_discard_rx_packet(channel, rx_buf, n_frags);
 565                 return;
 566         }
 567
 568         if (n_frags == 1 && !(flags & EF4_RX_PKT_PREFIX_LEN))
 569                 rx_buf->len = len;
 570
 571         /* Release and/or sync the DMA mapping - assumes all RX buffers
 572          * consumed in-order per RX queue.
 573          */
 574         ef4_sync_rx_buffer(efx, rx_buf, rx_buf->len);
 575
 576         /* Prefetch nice and early so data will (hopefully) be in cache by
 577          * the time we look at it.
 578          */
 579         prefetch(ef4_rx_buf_va(rx_buf));
 580
 581         rx_buf->page_offset += efx->rx_prefix_size;
 582         rx_buf->len -= efx->rx_prefix_size;
 583
 584         if (n_frags > 1) {
 585                 /* Release/sync DMA mapping for additional fragments.
 586                  * Fix length for last fragment.
 587                  */
 588                 unsigned int tail_frags = n_frags - 1;
 589
 590                 for (;;) {
 591                         rx_buf = ef4_rx_buf_next(rx_queue, rx_buf);
 592                         if (--tail_frags == 0)
 593                                 break;
 594                         ef4_sync_rx_buffer(efx, rx_buf, efx->rx_dma_len);
 595                 }
 596                 rx_buf->len = len - (n_frags - 1) * efx->rx_dma_len;
 597                 ef4_sync_rx_buffer(efx, rx_buf, rx_buf->len);
 598         }
 599
 600         /* All fragments have been DMA-synced, so recycle pages. */
 601         rx_buf = ef4_rx_buffer(rx_queue, index);
 602         ef4_recycle_rx_pages(channel, rx_buf, n_frags);
 603
 604         /* Pipeline receives so that we give time for packet headers to be
 605          * prefetched into cache.
 606          */
 607         ef4_rx_flush_packet(channel);
 608         channel->rx_pkt_n_frags = n_frags;
 609         channel->rx_pkt_index = index;
 610 }
 611
 612 static void ef4_rx_deliver(struct ef4_channel *channel, u8 *eh,
 613                            struct ef4_rx_buffer *rx_buf,
 614                            unsigned int n_frags)
 615 {
 616         struct sk_buff *skb;
 617         u16 hdr_len = min_t(u16, rx_buf->len, EF4_SKB_HEADERS);
 618
 619         skb = ef4_rx_mk_skb(channel, rx_buf, n_frags, eh, hdr_len);
 620         if (unlikely(skb == NULL)) {
 621                 struct ef4_rx_queue *rx_queue;
 622
 623                 rx_queue = ef4_channel_get_rx_queue(channel);
 624                 ef4_free_rx_buffers(rx_queue, rx_buf, n_frags);
 625                 return;
 626         }
 627         skb_record_rx_queue(skb, channel->rx_queue.core_index);
 628
 629         /* Set the SKB flags */
 630         skb_checksum_none_assert(skb);
 631         if (likely(rx_buf->flags & EF4_RX_PKT_CSUMMED))
 632                 skb->ip_summed = CHECKSUM_UNNECESSARY;
 633
 634         if (channel->type->receive_skb)
 635                 if (channel->type->receive_skb(channel, skb))
 636                         return;
 637
 638         /* Pass the packet up */
 639         netif_receive_skb(skb);
 640 }
 641
 642 /* Handle a received packet.  Second half: Touches packet payload. */
 643 void __ef4_rx_packet(struct ef4_channel *channel)
 644 {
 645         struct ef4_nic *efx = channel->efx;
 646         struct ef4_rx_buffer *rx_buf =
 647                 ef4_rx_buffer(&channel->rx_queue, channel->rx_pkt_index);
 648         u8 *eh = ef4_rx_buf_va(rx_buf);
 649
 650         /* Read length from the prefix if necessary.  This already
 651          * excludes the length of the prefix itself.
 652          */
 653         if (rx_buf->flags & EF4_RX_PKT_PREFIX_LEN)
 654                 rx_buf->len = le16_to_cpup((__le16 *)
 655                                            (eh + efx->rx_packet_len_offset));
 656
 657         /* If we're in loopback test, then pass the packet directly to the
 658          * loopback layer, and free the rx_buf here
 659          */
 660         if (unlikely(efx->loopback_selftest)) {
 661                 struct ef4_rx_queue *rx_queue;
 662
 663                 ef4_loopback_rx_packet(efx, eh, rx_buf->len);
 664                 rx_queue = ef4_channel_get_rx_queue(channel);
 665                 ef4_free_rx_buffers(rx_queue, rx_buf,
 666                                     channel->rx_pkt_n_frags);
 667                 goto out;
 668         }
 669
 670         if (unlikely(!(efx->net_dev->features & NETIF_F_RXCSUM)))
 671                 rx_buf->flags &= ~EF4_RX_PKT_CSUMMED;
 672
 673         if ((rx_buf->flags & EF4_RX_PKT_TCP) && !channel->type->receive_skb)
 674                 ef4_rx_packet_gro(channel, rx_buf, channel->rx_pkt_n_frags, eh);
 675         else
 676                 ef4_rx_deliver(channel, eh, rx_buf, channel->rx_pkt_n_frags);
 677 out:
 678         channel->rx_pkt_n_frags = 0;
 679 }
 680
 681 int ef4_probe_rx_queue(struct ef4_rx_queue *rx_queue)
 682 {
 683         struct ef4_nic *efx = rx_queue->efx;
 684         unsigned int entries;
 685         int rc;
 686
 687         /* Create the smallest power-of-two aligned ring */
 688         entries = max(roundup_pow_of_two(efx->rxq_entries), EF4_MIN_DMAQ_SIZE);
 689         EF4_BUG_ON_PARANOID(entries > EF4_MAX_DMAQ_SIZE);
 690         rx_queue->ptr_mask = entries - 1;
 691
 692         netif_dbg(efx, probe, efx->net_dev,
 693                   "creating RX queue %d size %#x mask %#x\n",
 694                   ef4_rx_queue_index(rx_queue), efx->rxq_entries,
 695                   rx_queue->ptr_mask);
 696
 697         /* Allocate RX buffers */
 698         rx_queue->buffer = kcalloc(entries, sizeof(*rx_queue->buffer),
 699                                    GFP_KERNEL);
 700         if (!rx_queue->buffer)
 701                 return -ENOMEM;
 702
 703         rc = ef4_nic_probe_rx(rx_queue);
 704         if (rc) {
 705                 kfree(rx_queue->buffer);
 706                 rx_queue->buffer = NULL;
 707         }
 708
 709         return rc;
 710 }
 711
 712 static void ef4_init_rx_recycle_ring(struct ef4_nic *efx,
 713                                      struct ef4_rx_queue *rx_queue)
 714 {
 715         unsigned int bufs_in_recycle_ring, page_ring_size;
 716
 717         /* Set the RX recycle ring size */
 718 #ifdef CONFIG_PPC64
 719         bufs_in_recycle_ring = EF4_RECYCLE_RING_SIZE_IOMMU;
 720 #else
 721         if (iommu_present(&pci_bus_type))
 722                 bufs_in_recycle_ring = EF4_RECYCLE_RING_SIZE_IOMMU;
 723         else
 724                 bufs_in_recycle_ring = EF4_RECYCLE_RING_SIZE_NOIOMMU;
 725 #endif /* CONFIG_PPC64 */
 726
 727         page_ring_size = roundup_pow_of_two(bufs_in_recycle_ring /
 728                                             efx->rx_bufs_per_page);
 729         rx_queue->page_ring = kcalloc(page_ring_size,
 730                                       sizeof(*rx_queue->page_ring), GFP_KERNEL);
 731         rx_queue->page_ptr_mask = page_ring_size - 1;
 732 }
 733
 734 void ef4_init_rx_queue(struct ef4_rx_queue *rx_queue)
 735 {
 736         struct ef4_nic *efx = rx_queue->efx;
 737         unsigned int max_fill, trigger, max_trigger;
 738
 739         netif_dbg(rx_queue->efx, drv, rx_queue->efx->net_dev,
 740                   "initialising RX queue %d\n", ef4_rx_queue_index(rx_queue));
 741
 742         /* Initialise ptr fields */
 743         rx_queue->added_count = 0;
 744         rx_queue->notified_count = 0;
 745         rx_queue->removed_count = 0;
 746         rx_queue->min_fill = -1U;
 747         ef4_init_rx_recycle_ring(efx, rx_queue);
 748
 749         rx_queue->page_remove = 0;
 750         rx_queue->page_add = rx_queue->page_ptr_mask + 1;
 751         rx_queue->page_recycle_count = 0;
 752         rx_queue->page_recycle_failed = 0;
 753         rx_queue->page_recycle_full = 0;
 754
 755         /* Initialise limit fields */
 756         max_fill = efx->rxq_entries - EF4_RXD_HEAD_ROOM;
 757         max_trigger =
 758                 max_fill - efx->rx_pages_per_batch * efx->rx_bufs_per_page;
 759         if (rx_refill_threshold != 0) {
 760                 trigger = max_fill * min(rx_refill_threshold, 100U) / 100U;
 761                 if (trigger > max_trigger)
 762                         trigger = max_trigger;
 763         } else {
 764                 trigger = max_trigger;
 765         }
 766
 767         rx_queue->max_fill = max_fill;
 768         rx_queue->fast_fill_trigger = trigger;
 769         rx_queue->refill_enabled = true;
 770
 771         /* Set up RX descriptor ring */
 772         ef4_nic_init_rx(rx_queue);
 773 }
 774
 775 void ef4_fini_rx_queue(struct ef4_rx_queue *rx_queue)
 776 {
 777         int i;
 778         struct ef4_nic *efx = rx_queue->efx;
 779         struct ef4_rx_buffer *rx_buf;
 780
 781         netif_dbg(rx_queue->efx, drv, rx_queue->efx->net_dev,
 782                   "shutting down RX queue %d\n", ef4_rx_queue_index(rx_queue));
 783
 784         del_timer_sync(&rx_queue->slow_fill);
 785
 786         /* Release RX buffers from the current read ptr to the write ptr */
 787         if (rx_queue->buffer) {
 788                 for (i = rx_queue->removed_count; i < rx_queue->added_count;
 789                      i++) {
 790                         unsigned index = i & rx_queue->ptr_mask;
 791                         rx_buf = ef4_rx_buffer(rx_queue, index);
 792                         ef4_fini_rx_buffer(rx_queue, rx_buf);
 793                 }
 794         }
 795
 796         /* Unmap and release the pages in the recycle ring. Remove the ring. */
 797         for (i = 0; i <= rx_queue->page_ptr_mask; i++) {
 798                 struct page *page = rx_queue->page_ring[i];
 799                 struct ef4_rx_page_state *state;
 800
 801                 if (page == NULL)
 802                         continue;
 803
 804                 state = page_address(page);
 805                 dma_unmap_page(&efx->pci_dev->dev, state->dma_addr,
 806                                PAGE_SIZE << efx->rx_buffer_order,
 807                                DMA_FROM_DEVICE);
 808                 put_page(page);
 809         }
 810         kfree(rx_queue->page_ring);
 811         rx_queue->page_ring = NULL;
 812 }
 813
 814 void ef4_remove_rx_queue(struct ef4_rx_queue *rx_queue)
 815 {
 816         netif_dbg(rx_queue->efx, drv, rx_queue->efx->net_dev,
 817                   "destroying RX queue %d\n", ef4_rx_queue_index(rx_queue));
 818
 819         ef4_nic_remove_rx(rx_queue);
 820
 821         kfree(rx_queue->buffer);
 822         rx_queue->buffer = NULL;
 823 }
 824
 825
 826 module_param(rx_refill_threshold, uint, 0444);
 827 MODULE_PARM_DESC(rx_refill_threshold,
 828                  "RX descriptor ring refill threshold (%)");
 829
 830 #ifdef CONFIG_RFS_ACCEL
 831
 832 int ef4_filter_rfs(struct net_device *net_dev, const struct sk_buff *skb,
 833                    u16 rxq_index, u32 flow_id)
 834 {
 835         struct ef4_nic *efx = netdev_priv(net_dev);
 836         struct ef4_channel *channel;
 837         struct ef4_filter_spec spec;
 838         struct flow_keys fk;
 839         int rc;
 840
 841         if (flow_id == RPS_FLOW_ID_INVALID)
 842                 return -EINVAL;
 843
 844         if (!skb_flow_dissect_flow_keys(skb, &fk, 0))
 845                 return -EPROTONOSUPPORT;
 846
 847         if (fk.basic.n_proto != htons(ETH_P_IP) && fk.basic.n_proto != htons(ETH_P_IPV6))
 848                 return -EPROTONOSUPPORT;
 849         if (fk.control.flags & FLOW_DIS_IS_FRAGMENT)
 850                 return -EPROTONOSUPPORT;
 851
 852         ef4_filter_init_rx(&spec, EF4_FILTER_PRI_HINT,
 853                            efx->rx_scatter ? EF4_FILTER_FLAG_RX_SCATTER : 0,
 854                            rxq_index);
 855         spec.match_flags =
 856                 EF4_FILTER_MATCH_ETHER_TYPE | EF4_FILTER_MATCH_IP_PROTO |
 857                 EF4_FILTER_MATCH_LOC_HOST | EF4_FILTER_MATCH_LOC_PORT |
 858                 EF4_FILTER_MATCH_REM_HOST | EF4_FILTER_MATCH_REM_PORT;
 859         spec.ether_type = fk.basic.n_proto;
 860         spec.ip_proto = fk.basic.ip_proto;
 861
 862         if (fk.basic.n_proto == htons(ETH_P_IP)) {
 863                 spec.rem_host[0] = fk.addrs.v4addrs.src;
 864                 spec.loc_host[0] = fk.addrs.v4addrs.dst;
 865         } else {
 866                 memcpy(spec.rem_host, &fk.addrs.v6addrs.src, sizeof(struct in6_addr));
 867                 memcpy(spec.loc_host, &fk.addrs.v6addrs.dst, sizeof(struct in6_addr));
 868         }
 869
 870         spec.rem_port = fk.ports.src;
 871         spec.loc_port = fk.ports.dst;
 872
 873         rc = efx->type->filter_rfs_insert(efx, &spec);
 874         if (rc < 0)
 875                 return rc;
 876
 877         /* Remember this so we can check whether to expire the filter later */
 878         channel = ef4_get_channel(efx, rxq_index);
 879         channel->rps_flow_id[rc] = flow_id;
 880         ++channel->rfs_filters_added;
 881
 882         if (spec.ether_type == htons(ETH_P_IP))
 883                 netif_info(efx, rx_status, efx->net_dev,
 884                            "steering %s %pI4:%u:%pI4:%u to queue %u [flow %u filter %d]\n",
 885                            (spec.ip_proto == IPPROTO_TCP) ? "TCP" : "UDP",
 886                            spec.rem_host, ntohs(spec.rem_port), spec.loc_host,
 887                            ntohs(spec.loc_port), rxq_index, flow_id, rc);
 888         else
 889                 netif_info(efx, rx_status, efx->net_dev,
 890                            "steering %s [%pI6]:%u:[%pI6]:%u to queue %u [flow %u filter %d]\n",
 891                            (spec.ip_proto == IPPROTO_TCP) ? "TCP" : "UDP",
 892                            spec.rem_host, ntohs(spec.rem_port), spec.loc_host,
 893                            ntohs(spec.loc_port), rxq_index, flow_id, rc);
 894
 895         return rc;
 896 }
 897
 898 bool __ef4_filter_rfs_expire(struct ef4_nic *efx, unsigned int quota)
 899 {
 900         bool (*expire_one)(struct ef4_nic *efx, u32 flow_id, unsigned int index);
 901         unsigned int channel_idx, index, size;
 902         u32 flow_id;
 903
 904         if (!spin_trylock_bh(&efx->filter_lock))
 905                 return false;
 906
 907         expire_one = efx->type->filter_rfs_expire_one;
 908         channel_idx = efx->rps_expire_channel;
 909         index = efx->rps_expire_index;
 910         size = efx->type->max_rx_ip_filters;
 911         while (quota--) {
 912                 struct ef4_channel *channel = ef4_get_channel(efx, channel_idx);
 913                 flow_id = channel->rps_flow_id[index];
 914
 915                 if (flow_id != RPS_FLOW_ID_INVALID &&
 916                     expire_one(efx, flow_id, index)) {
 917                         netif_info(efx, rx_status, efx->net_dev,
 918                                    "expired filter %d [queue %u flow %u]\n",
 919                                    index, channel_idx, flow_id);
 920                         channel->rps_flow_id[index] = RPS_FLOW_ID_INVALID;
 921                 }
 922                 if (++index == size) {
 923                         if (++channel_idx == efx->n_channels)
 924                                 channel_idx = 0;
 925                         index = 0;
 926                 }
 927         }
 928         efx->rps_expire_channel = channel_idx;
 929         efx->rps_expire_index = index;
 930
 931         spin_unlock_bh(&efx->filter_lock);
 932         return true;
 933 }
 934
 935 #endif /* CONFIG_RFS_ACCEL */
 936
 937 /**
 938  * ef4_filter_is_mc_recipient - test whether spec is a multicast recipient
 939  * @spec: Specification to test
 940  *
 941  * Return: %true if the specification is a non-drop RX filter that
 942  * matches a local MAC address I/G bit value of 1 or matches a local
 943  * IPv4 or IPv6 address value in the respective multicast address
 944  * range.  Otherwise %false.
 945  */
 946 bool ef4_filter_is_mc_recipient(const struct ef4_filter_spec *spec)
 947 {
 948         if (!(spec->flags & EF4_FILTER_FLAG_RX) ||
 949             spec->dmaq_id == EF4_FILTER_RX_DMAQ_ID_DROP)
 950                 return false;
 951
 952         if (spec->match_flags &
 953             (EF4_FILTER_MATCH_LOC_MAC | EF4_FILTER_MATCH_LOC_MAC_IG) &&
 954             is_multicast_ether_addr(spec->loc_mac))
 955                 return true;
 956
 957         if ((spec->match_flags &
 958              (EF4_FILTER_MATCH_ETHER_TYPE | EF4_FILTER_MATCH_LOC_HOST)) ==
 959             (EF4_FILTER_MATCH_ETHER_TYPE | EF4_FILTER_MATCH_LOC_HOST)) {
 960                 if (spec->ether_type == htons(ETH_P_IP) &&
 961                     ipv4_is_multicast(spec->loc_host[0]))
 962                         return true;
 963                 if (spec->ether_type == htons(ETH_P_IPV6) &&
 964                     ((const u8 *)spec->loc_host)[0] == 0xff)
 965                         return true;
 966         }
 967
 968         return false;
 969 }