net/core/gro.c

   1 // SPDX-License-Identifier: GPL-2.0-or-later
   2 #include <net/gro.h>
   3 #include <net/dst_metadata.h>
   4 #include <net/busy_poll.h>
   5 #include <trace/events/net.h>
   6
   7 #define MAX_GRO_SKBS 8
   8
   9 /* This should be increased if a protocol with a bigger head is added. */
  10 #define GRO_MAX_HEAD (MAX_HEADER + 128)
  11
  12 static DEFINE_SPINLOCK(offload_lock);
  13 static struct list_head offload_base __read_mostly = LIST_HEAD_INIT(offload_base);
  14 /* Maximum number of GRO_NORMAL skbs to batch up for list-RX */
  15 int gro_normal_batch __read_mostly = 8;
  16
  17 /**
  18  *      dev_add_offload - register offload handlers
  19  *      @po: protocol offload declaration
  20  *
  21  *      Add protocol offload handlers to the networking stack. The passed
  22  *      &proto_offload is linked into kernel lists and may not be freed until
  23  *      it has been removed from the kernel lists.
  24  *
  25  *      This call does not sleep therefore it can not
  26  *      guarantee all CPU's that are in middle of receiving packets
  27  *      will see the new offload handlers (until the next received packet).
  28  */
  29 void dev_add_offload(struct packet_offload *po)
  30 {
  31         struct packet_offload *elem;
  32
  33         spin_lock(&offload_lock);
  34         list_for_each_entry(elem, &offload_base, list) {
  35                 if (po->priority < elem->priority)
  36                         break;
  37         }
  38         list_add_rcu(&po->list, elem->list.prev);
  39         spin_unlock(&offload_lock);
  40 }
  41 EXPORT_SYMBOL(dev_add_offload);
  42
  43 /**
  44  *      __dev_remove_offload     - remove offload handler
  45  *      @po: packet offload declaration
  46  *
  47  *      Remove a protocol offload handler that was previously added to the
  48  *      kernel offload handlers by dev_add_offload(). The passed &offload_type
  49  *      is removed from the kernel lists and can be freed or reused once this
  50  *      function returns.
  51  *
  52  *      The packet type might still be in use by receivers
  53  *      and must not be freed until after all the CPU's have gone
  54  *      through a quiescent state.
  55  */
  56 static void __dev_remove_offload(struct packet_offload *po)
  57 {
  58         struct list_head *head = &offload_base;
  59         struct packet_offload *po1;
  60
  61         spin_lock(&offload_lock);
  62
  63         list_for_each_entry(po1, head, list) {
  64                 if (po == po1) {
  65                         list_del_rcu(&po->list);
  66                         goto out;
  67                 }
  68         }
  69
  70         pr_warn("dev_remove_offload: %p not found\n", po);
  71 out:
  72         spin_unlock(&offload_lock);
  73 }
  74
  75 /**
  76  *      dev_remove_offload       - remove packet offload handler
  77  *      @po: packet offload declaration
  78  *
  79  *      Remove a packet offload handler that was previously added to the kernel
  80  *      offload handlers by dev_add_offload(). The passed &offload_type is
  81  *      removed from the kernel lists and can be freed or reused once this
  82  *      function returns.
  83  *
  84  *      This call sleeps to guarantee that no CPU is looking at the packet
  85  *      type after return.
  86  */
  87 void dev_remove_offload(struct packet_offload *po)
  88 {
  89         __dev_remove_offload(po);
  90
  91         synchronize_net();
  92 }
  93 EXPORT_SYMBOL(dev_remove_offload);
  94
  95 /**
  96  *      skb_mac_gso_segment - mac layer segmentation handler.
  97  *      @skb: buffer to segment
  98  *      @features: features for the output path (see dev->features)
  99  */
 100 struct sk_buff *skb_mac_gso_segment(struct sk_buff *skb,
 101                                     netdev_features_t features)
 102 {
 103         struct sk_buff *segs = ERR_PTR(-EPROTONOSUPPORT);
 104         struct packet_offload *ptype;
 105         int vlan_depth = skb->mac_len;
 106         __be16 type = skb_network_protocol(skb, &vlan_depth);
 107
 108         if (unlikely(!type))
 109                 return ERR_PTR(-EINVAL);
 110
 111         __skb_pull(skb, vlan_depth);
 112
 113         rcu_read_lock();
 114         list_for_each_entry_rcu(ptype, &offload_base, list) {
 115                 if (ptype->type == type && ptype->callbacks.gso_segment) {
 116                         segs = ptype->callbacks.gso_segment(skb, features);
 117                         break;
 118                 }
 119         }
 120         rcu_read_unlock();
 121
 122         __skb_push(skb, skb->data - skb_mac_header(skb));
 123
 124         return segs;
 125 }
 126 EXPORT_SYMBOL(skb_mac_gso_segment);
 127
 128 int skb_gro_receive(struct sk_buff *p, struct sk_buff *skb)
 129 {
 130         struct skb_shared_info *pinfo, *skbinfo = skb_shinfo(skb);
 131         unsigned int offset = skb_gro_offset(skb);
 132         unsigned int headlen = skb_headlen(skb);
 133         unsigned int len = skb_gro_len(skb);
 134         unsigned int delta_truesize;
 135         unsigned int gro_max_size;
 136         unsigned int new_truesize;
 137         struct sk_buff *lp;
 138
 139         /* pairs with WRITE_ONCE() in netif_set_gro_max_size() */
 140         gro_max_size = READ_ONCE(p->dev->gro_max_size);
 141
 142         if (unlikely(p->len + len >= gro_max_size || NAPI_GRO_CB(skb)->flush))
 143                 return -E2BIG;
 144
 145         lp = NAPI_GRO_CB(p)->last;
 146         pinfo = skb_shinfo(lp);
 147
 148         if (headlen <= offset) {
 149                 skb_frag_t *frag;
 150                 skb_frag_t *frag2;
 151                 int i = skbinfo->nr_frags;
 152                 int nr_frags = pinfo->nr_frags + i;
 153
 154                 if (nr_frags > MAX_SKB_FRAGS)
 155                         goto merge;
 156
 157                 offset -= headlen;
 158                 pinfo->nr_frags = nr_frags;
 159                 skbinfo->nr_frags = 0;
 160
 161                 frag = pinfo->frags + nr_frags;
 162                 frag2 = skbinfo->frags + i;
 163                 do {
 164                         *--frag = *--frag2;
 165                 } while (--i);
 166
 167                 skb_frag_off_add(frag, offset);
 168                 skb_frag_size_sub(frag, offset);
 169
 170                 /* all fragments truesize : remove (head size + sk_buff) */
 171                 new_truesize = SKB_TRUESIZE(skb_end_offset(skb));
 172                 delta_truesize = skb->truesize - new_truesize;
 173
 174                 skb->truesize = new_truesize;
 175                 skb->len -= skb->data_len;
 176                 skb->data_len = 0;
 177
 178                 NAPI_GRO_CB(skb)->free = NAPI_GRO_FREE;
 179                 goto done;
 180         } else if (skb->head_frag) {
 181                 int nr_frags = pinfo->nr_frags;
 182                 skb_frag_t *frag = pinfo->frags + nr_frags;
 183                 struct page *page = virt_to_head_page(skb->head);
 184                 unsigned int first_size = headlen - offset;
 185                 unsigned int first_offset;
 186
 187                 if (nr_frags + 1 + skbinfo->nr_frags > MAX_SKB_FRAGS)
 188                         goto merge;
 189
 190                 first_offset = skb->data -
 191                                (unsigned char *)page_address(page) +
 192                                offset;
 193
 194                 pinfo->nr_frags = nr_frags + 1 + skbinfo->nr_frags;
 195
 196                 __skb_frag_set_page(frag, page);
 197                 skb_frag_off_set(frag, first_offset);
 198                 skb_frag_size_set(frag, first_size);
 199
 200                 memcpy(frag + 1, skbinfo->frags, sizeof(*frag) * skbinfo->nr_frags);
 201                 /* We dont need to clear skbinfo->nr_frags here */
 202
 203                 new_truesize = SKB_DATA_ALIGN(sizeof(struct sk_buff));
 204                 delta_truesize = skb->truesize - new_truesize;
 205                 skb->truesize = new_truesize;
 206                 NAPI_GRO_CB(skb)->free = NAPI_GRO_FREE_STOLEN_HEAD;
 207                 goto done;
 208         }
 209
 210 merge:
 211         /* sk owenrship - if any - completely transferred to the aggregated packet */
 212         skb->destructor = NULL;
 213         delta_truesize = skb->truesize;
 214         if (offset > headlen) {
 215                 unsigned int eat = offset - headlen;
 216
 217                 skb_frag_off_add(&skbinfo->frags[0], eat);
 218                 skb_frag_size_sub(&skbinfo->frags[0], eat);
 219                 skb->data_len -= eat;
 220                 skb->len -= eat;
 221                 offset = headlen;
 222         }
 223
 224         __skb_pull(skb, offset);
 225
 226         if (NAPI_GRO_CB(p)->last == p)
 227                 skb_shinfo(p)->frag_list = skb;
 228         else
 229                 NAPI_GRO_CB(p)->last->next = skb;
 230         NAPI_GRO_CB(p)->last = skb;
 231         __skb_header_release(skb);
 232         lp = p;
 233
 234 done:
 235         NAPI_GRO_CB(p)->count++;
 236         p->data_len += len;
 237         p->truesize += delta_truesize;
 238         p->len += len;
 239         if (lp != p) {
 240                 lp->data_len += len;
 241                 lp->truesize += delta_truesize;
 242                 lp->len += len;
 243         }
 244         NAPI_GRO_CB(skb)->same_flow = 1;
 245         return 0;
 246 }
 247
 248
 249 static void napi_gro_complete(struct napi_struct *napi, struct sk_buff *skb)
 250 {
 251         struct packet_offload *ptype;
 252         __be16 type = skb->protocol;
 253         struct list_head *head = &offload_base;
 254         int err = -ENOENT;
 255
 256         BUILD_BUG_ON(sizeof(struct napi_gro_cb) > sizeof(skb->cb));
 257
 258         if (NAPI_GRO_CB(skb)->count == 1) {
 259                 skb_shinfo(skb)->gso_size = 0;
 260                 goto out;
 261         }
 262
 263         rcu_read_lock();
 264         list_for_each_entry_rcu(ptype, head, list) {
 265                 if (ptype->type != type || !ptype->callbacks.gro_complete)
 266                         continue;
 267
 268                 err = INDIRECT_CALL_INET(ptype->callbacks.gro_complete,
 269                                          ipv6_gro_complete, inet_gro_complete,
 270                                          skb, 0);
 271                 break;
 272         }
 273         rcu_read_unlock();
 274
 275         if (err) {
 276                 WARN_ON(&ptype->list == head);
 277                 kfree_skb(skb);
 278                 return;
 279         }
 280
 281 out:
 282         gro_normal_one(napi, skb, NAPI_GRO_CB(skb)->count);
 283 }
 284
 285 static void __napi_gro_flush_chain(struct napi_struct *napi, u32 index,
 286                                    bool flush_old)
 287 {
 288         struct list_head *head = &napi->gro_hash[index].list;
 289         struct sk_buff *skb, *p;
 290
 291         list_for_each_entry_safe_reverse(skb, p, head, list) {
 292                 if (flush_old && NAPI_GRO_CB(skb)->age == jiffies)
 293                         return;
 294                 skb_list_del_init(skb);
 295                 napi_gro_complete(napi, skb);
 296                 napi->gro_hash[index].count--;
 297         }
 298
 299         if (!napi->gro_hash[index].count)
 300                 __clear_bit(index, &napi->gro_bitmask);
 301 }
 302
 303 /* napi->gro_hash[].list contains packets ordered by age.
 304  * youngest packets at the head of it.
 305  * Complete skbs in reverse order to reduce latencies.
 306  */
 307 void napi_gro_flush(struct napi_struct *napi, bool flush_old)
 308 {
 309         unsigned long bitmask = napi->gro_bitmask;
 310         unsigned int i, base = ~0U;
 311
 312         while ((i = ffs(bitmask)) != 0) {
 313                 bitmask >>= i;
 314                 base += i;
 315                 __napi_gro_flush_chain(napi, base, flush_old);
 316         }
 317 }
 318 EXPORT_SYMBOL(napi_gro_flush);
 319
 320 static void gro_list_prepare(const struct list_head *head,
 321                              const struct sk_buff *skb)
 322 {
 323         unsigned int maclen = skb->dev->hard_header_len;
 324         u32 hash = skb_get_hash_raw(skb);
 325         struct sk_buff *p;
 326
 327         list_for_each_entry(p, head, list) {
 328                 unsigned long diffs;
 329
 330                 NAPI_GRO_CB(p)->flush = 0;
 331
 332                 if (hash != skb_get_hash_raw(p)) {
 333                         NAPI_GRO_CB(p)->same_flow = 0;
 334                         continue;
 335                 }
 336
 337                 diffs = (unsigned long)p->dev ^ (unsigned long)skb->dev;
 338                 diffs |= skb_vlan_tag_present(p) ^ skb_vlan_tag_present(skb);
 339                 if (skb_vlan_tag_present(p))
 340                         diffs |= skb_vlan_tag_get(p) ^ skb_vlan_tag_get(skb);
 341                 diffs |= skb_metadata_differs(p, skb);
 342                 if (maclen == ETH_HLEN)
 343                         diffs |= compare_ether_header(skb_mac_header(p),
 344                                                       skb_mac_header(skb));
 345                 else if (!diffs)
 346                         diffs = memcmp(skb_mac_header(p),
 347                                        skb_mac_header(skb),
 348                                        maclen);
 349
 350                 /* in most common scenarions 'slow_gro' is 0
 351                  * otherwise we are already on some slower paths
 352                  * either skip all the infrequent tests altogether or
 353                  * avoid trying too hard to skip each of them individually
 354                  */
 355                 if (!diffs && unlikely(skb->slow_gro | p->slow_gro)) {
 356 #if IS_ENABLED(CONFIG_SKB_EXTENSIONS) && IS_ENABLED(CONFIG_NET_TC_SKB_EXT)
 357                         struct tc_skb_ext *skb_ext;
 358                         struct tc_skb_ext *p_ext;
 359 #endif
 360
 361                         diffs |= p->sk != skb->sk;
 362                         diffs |= skb_metadata_dst_cmp(p, skb);
 363                         diffs |= skb_get_nfct(p) ^ skb_get_nfct(skb);
 364
 365 #if IS_ENABLED(CONFIG_SKB_EXTENSIONS) && IS_ENABLED(CONFIG_NET_TC_SKB_EXT)
 366                         skb_ext = skb_ext_find(skb, TC_SKB_EXT);
 367                         p_ext = skb_ext_find(p, TC_SKB_EXT);
 368
 369                         diffs |= (!!p_ext) ^ (!!skb_ext);
 370                         if (!diffs && unlikely(skb_ext))
 371                                 diffs |= p_ext->chain ^ skb_ext->chain;
 372 #endif
 373                 }
 374
 375                 NAPI_GRO_CB(p)->same_flow = !diffs;
 376         }
 377 }
 378
 379 static inline void skb_gro_reset_offset(struct sk_buff *skb, u32 nhoff)
 380 {
 381         const struct skb_shared_info *pinfo = skb_shinfo(skb);
 382         const skb_frag_t *frag0 = &pinfo->frags[0];
 383
 384         NAPI_GRO_CB(skb)->data_offset = 0;
 385         NAPI_GRO_CB(skb)->frag0 = NULL;
 386         NAPI_GRO_CB(skb)->frag0_len = 0;
 387
 388         if (!skb_headlen(skb) && pinfo->nr_frags &&
 389             !PageHighMem(skb_frag_page(frag0)) &&
 390             (!NET_IP_ALIGN || !((skb_frag_off(frag0) + nhoff) & 3))) {
 391                 NAPI_GRO_CB(skb)->frag0 = skb_frag_address(frag0);
 392                 NAPI_GRO_CB(skb)->frag0_len = min_t(unsigned int,
 393                                                     skb_frag_size(frag0),
 394                                                     skb->end - skb->tail);
 395         }
 396 }
 397
 398 static void gro_pull_from_frag0(struct sk_buff *skb, int grow)
 399 {
 400         struct skb_shared_info *pinfo = skb_shinfo(skb);
 401
 402         BUG_ON(skb->end - skb->tail < grow);
 403
 404         memcpy(skb_tail_pointer(skb), NAPI_GRO_CB(skb)->frag0, grow);
 405
 406         skb->data_len -= grow;
 407         skb->tail += grow;
 408
 409         skb_frag_off_add(&pinfo->frags[0], grow);
 410         skb_frag_size_sub(&pinfo->frags[0], grow);
 411
 412         if (unlikely(!skb_frag_size(&pinfo->frags[0]))) {
 413                 skb_frag_unref(skb, 0);
 414                 memmove(pinfo->frags, pinfo->frags + 1,
 415                         --pinfo->nr_frags * sizeof(pinfo->frags[0]));
 416         }
 417 }
 418
 419 static void gro_flush_oldest(struct napi_struct *napi, struct list_head *head)
 420 {
 421         struct sk_buff *oldest;
 422
 423         oldest = list_last_entry(head, struct sk_buff, list);
 424
 425         /* We are called with head length >= MAX_GRO_SKBS, so this is
 426          * impossible.
 427          */
 428         if (WARN_ON_ONCE(!oldest))
 429                 return;
 430
 431         /* Do not adjust napi->gro_hash[].count, caller is adding a new
 432          * SKB to the chain.
 433          */
 434         skb_list_del_init(oldest);
 435         napi_gro_complete(napi, oldest);
 436 }
 437
 438 static enum gro_result dev_gro_receive(struct napi_struct *napi, struct sk_buff *skb)
 439 {
 440         u32 bucket = skb_get_hash_raw(skb) & (GRO_HASH_BUCKETS - 1);
 441         struct gro_list *gro_list = &napi->gro_hash[bucket];
 442         struct list_head *head = &offload_base;
 443         struct packet_offload *ptype;
 444         __be16 type = skb->protocol;
 445         struct sk_buff *pp = NULL;
 446         enum gro_result ret;
 447         int same_flow;
 448         int grow;
 449
 450         if (netif_elide_gro(skb->dev))
 451                 goto normal;
 452
 453         gro_list_prepare(&gro_list->list, skb);
 454
 455         rcu_read_lock();
 456         list_for_each_entry_rcu(ptype, head, list) {
 457                 if (ptype->type != type || !ptype->callbacks.gro_receive)
 458                         continue;
 459
 460                 skb_set_network_header(skb, skb_gro_offset(skb));
 461                 skb_reset_mac_len(skb);
 462                 NAPI_GRO_CB(skb)->same_flow = 0;
 463                 NAPI_GRO_CB(skb)->flush = skb_is_gso(skb) || skb_has_frag_list(skb);
 464                 NAPI_GRO_CB(skb)->free = 0;
 465                 NAPI_GRO_CB(skb)->encap_mark = 0;
 466                 NAPI_GRO_CB(skb)->recursion_counter = 0;
 467                 NAPI_GRO_CB(skb)->is_fou = 0;
 468                 NAPI_GRO_CB(skb)->is_atomic = 1;
 469                 NAPI_GRO_CB(skb)->gro_remcsum_start = 0;
 470
 471                 /* Setup for GRO checksum validation */
 472                 switch (skb->ip_summed) {
 473                 case CHECKSUM_COMPLETE:
 474                         NAPI_GRO_CB(skb)->csum = skb->csum;
 475                         NAPI_GRO_CB(skb)->csum_valid = 1;
 476                         NAPI_GRO_CB(skb)->csum_cnt = 0;
 477                         break;
 478                 case CHECKSUM_UNNECESSARY:
 479                         NAPI_GRO_CB(skb)->csum_cnt = skb->csum_level + 1;
 480                         NAPI_GRO_CB(skb)->csum_valid = 0;
 481                         break;
 482                 default:
 483                         NAPI_GRO_CB(skb)->csum_cnt = 0;
 484                         NAPI_GRO_CB(skb)->csum_valid = 0;
 485                 }
 486
 487                 pp = INDIRECT_CALL_INET(ptype->callbacks.gro_receive,
 488                                         ipv6_gro_receive, inet_gro_receive,
 489                                         &gro_list->list, skb);
 490                 break;
 491         }
 492         rcu_read_unlock();
 493
 494         if (&ptype->list == head)
 495                 goto normal;
 496
 497         if (PTR_ERR(pp) == -EINPROGRESS) {
 498                 ret = GRO_CONSUMED;
 499                 goto ok;
 500         }
 501
 502         same_flow = NAPI_GRO_CB(skb)->same_flow;
 503         ret = NAPI_GRO_CB(skb)->free ? GRO_MERGED_FREE : GRO_MERGED;
 504
 505         if (pp) {
 506                 skb_list_del_init(pp);
 507                 napi_gro_complete(napi, pp);
 508                 gro_list->count--;
 509         }
 510
 511         if (same_flow)
 512                 goto ok;
 513
 514         if (NAPI_GRO_CB(skb)->flush)
 515                 goto normal;
 516
 517         if (unlikely(gro_list->count >= MAX_GRO_SKBS))
 518                 gro_flush_oldest(napi, &gro_list->list);
 519         else
 520                 gro_list->count++;
 521
 522         NAPI_GRO_CB(skb)->count = 1;
 523         NAPI_GRO_CB(skb)->age = jiffies;
 524         NAPI_GRO_CB(skb)->last = skb;
 525         skb_shinfo(skb)->gso_size = skb_gro_len(skb);
 526         list_add(&skb->list, &gro_list->list);
 527         ret = GRO_HELD;
 528
 529 pull:
 530         grow = skb_gro_offset(skb) - skb_headlen(skb);
 531         if (grow > 0)
 532                 gro_pull_from_frag0(skb, grow);
 533 ok:
 534         if (gro_list->count) {
 535                 if (!test_bit(bucket, &napi->gro_bitmask))
 536                         __set_bit(bucket, &napi->gro_bitmask);
 537         } else if (test_bit(bucket, &napi->gro_bitmask)) {
 538                 __clear_bit(bucket, &napi->gro_bitmask);
 539         }
 540
 541         return ret;
 542
 543 normal:
 544         ret = GRO_NORMAL;
 545         goto pull;
 546 }
 547
 548 struct packet_offload *gro_find_receive_by_type(__be16 type)
 549 {
 550         struct list_head *offload_head = &offload_base;
 551         struct packet_offload *ptype;
 552
 553         list_for_each_entry_rcu(ptype, offload_head, list) {
 554                 if (ptype->type != type || !ptype->callbacks.gro_receive)
 555                         continue;
 556                 return ptype;
 557         }
 558         return NULL;
 559 }
 560 EXPORT_SYMBOL(gro_find_receive_by_type);
 561
 562 struct packet_offload *gro_find_complete_by_type(__be16 type)
 563 {
 564         struct list_head *offload_head = &offload_base;
 565         struct packet_offload *ptype;
 566
 567         list_for_each_entry_rcu(ptype, offload_head, list) {
 568                 if (ptype->type != type || !ptype->callbacks.gro_complete)
 569                         continue;
 570                 return ptype;
 571         }
 572         return NULL;
 573 }
 574 EXPORT_SYMBOL(gro_find_complete_by_type);
 575
 576 static gro_result_t napi_skb_finish(struct napi_struct *napi,
 577                                     struct sk_buff *skb,
 578                                     gro_result_t ret)
 579 {
 580         switch (ret) {
 581         case GRO_NORMAL:
 582                 gro_normal_one(napi, skb, 1);
 583                 break;
 584
 585         case GRO_MERGED_FREE:
 586                 if (NAPI_GRO_CB(skb)->free == NAPI_GRO_FREE_STOLEN_HEAD)
 587                         napi_skb_free_stolen_head(skb);
 588                 else if (skb->fclone != SKB_FCLONE_UNAVAILABLE)
 589                         __kfree_skb(skb);
 590                 else
 591                         __kfree_skb_defer(skb);
 592                 break;
 593
 594         case GRO_HELD:
 595         case GRO_MERGED:
 596         case GRO_CONSUMED:
 597                 break;
 598         }
 599
 600         return ret;
 601 }
 602
 603 gro_result_t napi_gro_receive(struct napi_struct *napi, struct sk_buff *skb)
 604 {
 605         gro_result_t ret;
 606
 607         skb_mark_napi_id(skb, napi);
 608         trace_napi_gro_receive_entry(skb);
 609
 610         skb_gro_reset_offset(skb, 0);
 611
 612         ret = napi_skb_finish(napi, skb, dev_gro_receive(napi, skb));
 613         trace_napi_gro_receive_exit(ret);
 614
 615         return ret;
 616 }
 617 EXPORT_SYMBOL(napi_gro_receive);
 618
 619 static void napi_reuse_skb(struct napi_struct *napi, struct sk_buff *skb)
 620 {
 621         if (unlikely(skb->pfmemalloc)) {
 622                 consume_skb(skb);
 623                 return;
 624         }
 625         __skb_pull(skb, skb_headlen(skb));
 626         /* restore the reserve we had after netdev_alloc_skb_ip_align() */
 627         skb_reserve(skb, NET_SKB_PAD + NET_IP_ALIGN - skb_headroom(skb));
 628         __vlan_hwaccel_clear_tag(skb);
 629         skb->dev = napi->dev;
 630         skb->skb_iif = 0;
 631
 632         /* eth_type_trans() assumes pkt_type is PACKET_HOST */
 633         skb->pkt_type = PACKET_HOST;
 634
 635         skb->encapsulation = 0;
 636         skb_shinfo(skb)->gso_type = 0;
 637         skb->truesize = SKB_TRUESIZE(skb_end_offset(skb));
 638         if (unlikely(skb->slow_gro)) {
 639                 skb_orphan(skb);
 640                 skb_ext_reset(skb);
 641                 nf_reset_ct(skb);
 642                 skb->slow_gro = 0;
 643         }
 644
 645         napi->skb = skb;
 646 }
 647
 648 struct sk_buff *napi_get_frags(struct napi_struct *napi)
 649 {
 650         struct sk_buff *skb = napi->skb;
 651
 652         if (!skb) {
 653                 skb = napi_alloc_skb(napi, GRO_MAX_HEAD);
 654                 if (skb) {
 655                         napi->skb = skb;
 656                         skb_mark_napi_id(skb, napi);
 657                 }
 658         }
 659         return skb;
 660 }
 661 EXPORT_SYMBOL(napi_get_frags);
 662
 663 static gro_result_t napi_frags_finish(struct napi_struct *napi,
 664                                       struct sk_buff *skb,
 665                                       gro_result_t ret)
 666 {
 667         switch (ret) {
 668         case GRO_NORMAL:
 669         case GRO_HELD:
 670                 __skb_push(skb, ETH_HLEN);
 671                 skb->protocol = eth_type_trans(skb, skb->dev);
 672                 if (ret == GRO_NORMAL)
 673                         gro_normal_one(napi, skb, 1);
 674                 break;
 675
 676         case GRO_MERGED_FREE:
 677                 if (NAPI_GRO_CB(skb)->free == NAPI_GRO_FREE_STOLEN_HEAD)
 678                         napi_skb_free_stolen_head(skb);
 679                 else
 680                         napi_reuse_skb(napi, skb);
 681                 break;
 682
 683         case GRO_MERGED:
 684         case GRO_CONSUMED:
 685                 break;
 686         }
 687
 688         return ret;
 689 }
 690
 691 /* Upper GRO stack assumes network header starts at gro_offset=0
 692  * Drivers could call both napi_gro_frags() and napi_gro_receive()
 693  * We copy ethernet header into skb->data to have a common layout.
 694  */
 695 static struct sk_buff *napi_frags_skb(struct napi_struct *napi)
 696 {
 697         struct sk_buff *skb = napi->skb;
 698         const struct ethhdr *eth;
 699         unsigned int hlen = sizeof(*eth);
 700
 701         napi->skb = NULL;
 702
 703         skb_reset_mac_header(skb);
 704         skb_gro_reset_offset(skb, hlen);
 705
 706         if (unlikely(skb_gro_header_hard(skb, hlen))) {
 707                 eth = skb_gro_header_slow(skb, hlen, 0);
 708                 if (unlikely(!eth)) {
 709                         net_warn_ratelimited("%s: dropping impossible skb from %s\n",
 710                                              __func__, napi->dev->name);
 711                         napi_reuse_skb(napi, skb);
 712                         return NULL;
 713                 }
 714         } else {
 715                 eth = (const struct ethhdr *)skb->data;
 716                 gro_pull_from_frag0(skb, hlen);
 717                 NAPI_GRO_CB(skb)->frag0 += hlen;
 718                 NAPI_GRO_CB(skb)->frag0_len -= hlen;
 719         }
 720         __skb_pull(skb, hlen);
 721
 722         /*
 723          * This works because the only protocols we care about don't require
 724          * special handling.
 725          * We'll fix it up properly in napi_frags_finish()
 726          */
 727         skb->protocol = eth->h_proto;
 728
 729         return skb;
 730 }
 731
 732 gro_result_t napi_gro_frags(struct napi_struct *napi)
 733 {
 734         gro_result_t ret;
 735         struct sk_buff *skb = napi_frags_skb(napi);
 736
 737         trace_napi_gro_frags_entry(skb);
 738
 739         ret = napi_frags_finish(napi, skb, dev_gro_receive(napi, skb));
 740         trace_napi_gro_frags_exit(ret);
 741
 742         return ret;
 743 }
 744 EXPORT_SYMBOL(napi_gro_frags);
 745
 746 /* Compute the checksum from gro_offset and return the folded value
 747  * after adding in any pseudo checksum.
 748  */
 749 __sum16 __skb_gro_checksum_complete(struct sk_buff *skb)
 750 {
 751         __wsum wsum;
 752         __sum16 sum;
 753
 754         wsum = skb_checksum(skb, skb_gro_offset(skb), skb_gro_len(skb), 0);
 755
 756         /* NAPI_GRO_CB(skb)->csum holds pseudo checksum */
 757         sum = csum_fold(csum_add(NAPI_GRO_CB(skb)->csum, wsum));
 758         /* See comments in __skb_checksum_complete(). */
 759         if (likely(!sum)) {
 760                 if (unlikely(skb->ip_summed == CHECKSUM_COMPLETE) &&
 761                     !skb->csum_complete_sw)
 762                         netdev_rx_csum_fault(skb->dev, skb);
 763         }
 764
 765         NAPI_GRO_CB(skb)->csum = wsum;
 766         NAPI_GRO_CB(skb)->csum_valid = 1;
 767
 768         return sum;
 769 }
 770 EXPORT_SYMBOL(__skb_gro_checksum_complete);