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 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 int skb_gro_receive(struct sk_buff *p, struct sk_buff *skb)
  97 {
  98         struct skb_shared_info *pinfo, *skbinfo = skb_shinfo(skb);
  99         unsigned int offset = skb_gro_offset(skb);
 100         unsigned int headlen = skb_headlen(skb);
 101         unsigned int len = skb_gro_len(skb);
 102         unsigned int delta_truesize;
 103         unsigned int gro_max_size;
 104         unsigned int new_truesize;
 105         struct sk_buff *lp;
 106         int segs;
 107
 108         /* Do not splice page pool based packets w/ non-page pool
 109          * packets. This can result in reference count issues as page
 110          * pool pages will not decrement the reference count and will
 111          * instead be immediately returned to the pool or have frag
 112          * count decremented.
 113          */
 114         if (p->pp_recycle != skb->pp_recycle)
 115                 return -ETOOMANYREFS;
 116
 117         /* pairs with WRITE_ONCE() in netif_set_gro(_ipv4)_max_size() */
 118         gro_max_size = p->protocol == htons(ETH_P_IPV6) ?
 119                         READ_ONCE(p->dev->gro_max_size) :
 120                         READ_ONCE(p->dev->gro_ipv4_max_size);
 121
 122         if (unlikely(p->len + len >= gro_max_size || NAPI_GRO_CB(skb)->flush))
 123                 return -E2BIG;
 124
 125         if (unlikely(p->len + len >= GRO_LEGACY_MAX_SIZE)) {
 126                 if (NAPI_GRO_CB(skb)->proto != IPPROTO_TCP ||
 127                     (p->protocol == htons(ETH_P_IPV6) &&
 128                      skb_headroom(p) < sizeof(struct hop_jumbo_hdr)) ||
 129                     p->encapsulation)
 130                         return -E2BIG;
 131         }
 132
 133         segs = NAPI_GRO_CB(skb)->count;
 134         lp = NAPI_GRO_CB(p)->last;
 135         pinfo = skb_shinfo(lp);
 136
 137         if (headlen <= offset) {
 138                 skb_frag_t *frag;
 139                 skb_frag_t *frag2;
 140                 int i = skbinfo->nr_frags;
 141                 int nr_frags = pinfo->nr_frags + i;
 142
 143                 if (nr_frags > MAX_SKB_FRAGS)
 144                         goto merge;
 145
 146                 offset -= headlen;
 147                 pinfo->nr_frags = nr_frags;
 148                 skbinfo->nr_frags = 0;
 149
 150                 frag = pinfo->frags + nr_frags;
 151                 frag2 = skbinfo->frags + i;
 152                 do {
 153                         *--frag = *--frag2;
 154                 } while (--i);
 155
 156                 skb_frag_off_add(frag, offset);
 157                 skb_frag_size_sub(frag, offset);
 158
 159                 /* all fragments truesize : remove (head size + sk_buff) */
 160                 new_truesize = SKB_TRUESIZE(skb_end_offset(skb));
 161                 delta_truesize = skb->truesize - new_truesize;
 162
 163                 skb->truesize = new_truesize;
 164                 skb->len -= skb->data_len;
 165                 skb->data_len = 0;
 166
 167                 NAPI_GRO_CB(skb)->free = NAPI_GRO_FREE;
 168                 goto done;
 169         } else if (skb->head_frag) {
 170                 int nr_frags = pinfo->nr_frags;
 171                 skb_frag_t *frag = pinfo->frags + nr_frags;
 172                 struct page *page = virt_to_head_page(skb->head);
 173                 unsigned int first_size = headlen - offset;
 174                 unsigned int first_offset;
 175
 176                 if (nr_frags + 1 + skbinfo->nr_frags > MAX_SKB_FRAGS)
 177                         goto merge;
 178
 179                 first_offset = skb->data -
 180                                (unsigned char *)page_address(page) +
 181                                offset;
 182
 183                 pinfo->nr_frags = nr_frags + 1 + skbinfo->nr_frags;
 184
 185                 skb_frag_fill_page_desc(frag, page, first_offset, first_size);
 186
 187                 memcpy(frag + 1, skbinfo->frags, sizeof(*frag) * skbinfo->nr_frags);
 188                 /* We dont need to clear skbinfo->nr_frags here */
 189
 190                 new_truesize = SKB_DATA_ALIGN(sizeof(struct sk_buff));
 191                 delta_truesize = skb->truesize - new_truesize;
 192                 skb->truesize = new_truesize;
 193                 NAPI_GRO_CB(skb)->free = NAPI_GRO_FREE_STOLEN_HEAD;
 194                 goto done;
 195         }
 196
 197 merge:
 198         /* sk ownership - if any - completely transferred to the aggregated packet */
 199         skb->destructor = NULL;
 200         skb->sk = NULL;
 201         delta_truesize = skb->truesize;
 202         if (offset > headlen) {
 203                 unsigned int eat = offset - headlen;
 204
 205                 skb_frag_off_add(&skbinfo->frags[0], eat);
 206                 skb_frag_size_sub(&skbinfo->frags[0], eat);
 207                 skb->data_len -= eat;
 208                 skb->len -= eat;
 209                 offset = headlen;
 210         }
 211
 212         __skb_pull(skb, offset);
 213
 214         if (NAPI_GRO_CB(p)->last == p)
 215                 skb_shinfo(p)->frag_list = skb;
 216         else
 217                 NAPI_GRO_CB(p)->last->next = skb;
 218         NAPI_GRO_CB(p)->last = skb;
 219         __skb_header_release(skb);
 220         lp = p;
 221
 222 done:
 223         NAPI_GRO_CB(p)->count += segs;
 224         p->data_len += len;
 225         p->truesize += delta_truesize;
 226         p->len += len;
 227         if (lp != p) {
 228                 lp->data_len += len;
 229                 lp->truesize += delta_truesize;
 230                 lp->len += len;
 231         }
 232         NAPI_GRO_CB(skb)->same_flow = 1;
 233         return 0;
 234 }
 235
 236
 237 static void napi_gro_complete(struct napi_struct *napi, struct sk_buff *skb)
 238 {
 239         struct packet_offload *ptype;
 240         __be16 type = skb->protocol;
 241         struct list_head *head = &offload_base;
 242         int err = -ENOENT;
 243
 244         BUILD_BUG_ON(sizeof(struct napi_gro_cb) > sizeof(skb->cb));
 245
 246         if (NAPI_GRO_CB(skb)->count == 1) {
 247                 skb_shinfo(skb)->gso_size = 0;
 248                 goto out;
 249         }
 250
 251         rcu_read_lock();
 252         list_for_each_entry_rcu(ptype, head, list) {
 253                 if (ptype->type != type || !ptype->callbacks.gro_complete)
 254                         continue;
 255
 256                 err = INDIRECT_CALL_INET(ptype->callbacks.gro_complete,
 257                                          ipv6_gro_complete, inet_gro_complete,
 258                                          skb, 0);
 259                 break;
 260         }
 261         rcu_read_unlock();
 262
 263         if (err) {
 264                 WARN_ON(&ptype->list == head);
 265                 kfree_skb(skb);
 266                 return;
 267         }
 268
 269 out:
 270         gro_normal_one(napi, skb, NAPI_GRO_CB(skb)->count);
 271 }
 272
 273 static void __napi_gro_flush_chain(struct napi_struct *napi, u32 index,
 274                                    bool flush_old)
 275 {
 276         struct list_head *head = &napi->gro_hash[index].list;
 277         struct sk_buff *skb, *p;
 278
 279         list_for_each_entry_safe_reverse(skb, p, head, list) {
 280                 if (flush_old && NAPI_GRO_CB(skb)->age == jiffies)
 281                         return;
 282                 skb_list_del_init(skb);
 283                 napi_gro_complete(napi, skb);
 284                 napi->gro_hash[index].count--;
 285         }
 286
 287         if (!napi->gro_hash[index].count)
 288                 __clear_bit(index, &napi->gro_bitmask);
 289 }
 290
 291 /* napi->gro_hash[].list contains packets ordered by age.
 292  * youngest packets at the head of it.
 293  * Complete skbs in reverse order to reduce latencies.
 294  */
 295 void napi_gro_flush(struct napi_struct *napi, bool flush_old)
 296 {
 297         unsigned long bitmask = napi->gro_bitmask;
 298         unsigned int i, base = ~0U;
 299
 300         while ((i = ffs(bitmask)) != 0) {
 301                 bitmask >>= i;
 302                 base += i;
 303                 __napi_gro_flush_chain(napi, base, flush_old);
 304         }
 305 }
 306 EXPORT_SYMBOL(napi_gro_flush);
 307
 308 static unsigned long gro_list_prepare_tc_ext(const struct sk_buff *skb,
 309                                              const struct sk_buff *p,
 310                                              unsigned long diffs)
 311 {
 312 #if IS_ENABLED(CONFIG_NET_TC_SKB_EXT)
 313         struct tc_skb_ext *skb_ext;
 314         struct tc_skb_ext *p_ext;
 315
 316         skb_ext = skb_ext_find(skb, TC_SKB_EXT);
 317         p_ext = skb_ext_find(p, TC_SKB_EXT);
 318
 319         diffs |= (!!p_ext) ^ (!!skb_ext);
 320         if (!diffs && unlikely(skb_ext))
 321                 diffs |= p_ext->chain ^ skb_ext->chain;
 322 #endif
 323         return diffs;
 324 }
 325
 326 static void gro_list_prepare(const struct list_head *head,
 327                              const struct sk_buff *skb)
 328 {
 329         unsigned int maclen = skb->dev->hard_header_len;
 330         u32 hash = skb_get_hash_raw(skb);
 331         struct sk_buff *p;
 332
 333         list_for_each_entry(p, head, list) {
 334                 unsigned long diffs;
 335
 336                 NAPI_GRO_CB(p)->flush = 0;
 337
 338                 if (hash != skb_get_hash_raw(p)) {
 339                         NAPI_GRO_CB(p)->same_flow = 0;
 340                         continue;
 341                 }
 342
 343                 diffs = (unsigned long)p->dev ^ (unsigned long)skb->dev;
 344                 diffs |= p->vlan_all ^ skb->vlan_all;
 345                 diffs |= skb_metadata_differs(p, skb);
 346                 if (maclen == ETH_HLEN)
 347                         diffs |= compare_ether_header(skb_mac_header(p),
 348                                                       skb_mac_header(skb));
 349                 else if (!diffs)
 350                         diffs = memcmp(skb_mac_header(p),
 351                                        skb_mac_header(skb),
 352                                        maclen);
 353
 354                 /* in most common scenarions 'slow_gro' is 0
 355                  * otherwise we are already on some slower paths
 356                  * either skip all the infrequent tests altogether or
 357                  * avoid trying too hard to skip each of them individually
 358                  */
 359                 if (!diffs && unlikely(skb->slow_gro | p->slow_gro)) {
 360                         diffs |= p->sk != skb->sk;
 361                         diffs |= skb_metadata_dst_cmp(p, skb);
 362                         diffs |= skb_get_nfct(p) ^ skb_get_nfct(skb);
 363
 364                         diffs |= gro_list_prepare_tc_ext(skb, p, diffs);
 365                 }
 366
 367                 NAPI_GRO_CB(p)->same_flow = !diffs;
 368         }
 369 }
 370
 371 static inline void skb_gro_reset_offset(struct sk_buff *skb, u32 nhoff)
 372 {
 373         const struct skb_shared_info *pinfo = skb_shinfo(skb);
 374         const skb_frag_t *frag0 = &pinfo->frags[0];
 375
 376         NAPI_GRO_CB(skb)->data_offset = 0;
 377         NAPI_GRO_CB(skb)->frag0 = NULL;
 378         NAPI_GRO_CB(skb)->frag0_len = 0;
 379
 380         if (!skb_headlen(skb) && pinfo->nr_frags &&
 381             !PageHighMem(skb_frag_page(frag0)) &&
 382             (!NET_IP_ALIGN || !((skb_frag_off(frag0) + nhoff) & 3))) {
 383                 NAPI_GRO_CB(skb)->frag0 = skb_frag_address(frag0);
 384                 NAPI_GRO_CB(skb)->frag0_len = min_t(unsigned int,
 385                                                     skb_frag_size(frag0),
 386                                                     skb->end - skb->tail);
 387         }
 388 }
 389
 390 static void gro_pull_from_frag0(struct sk_buff *skb, int grow)
 391 {
 392         struct skb_shared_info *pinfo = skb_shinfo(skb);
 393
 394         BUG_ON(skb->end - skb->tail < grow);
 395
 396         memcpy(skb_tail_pointer(skb), NAPI_GRO_CB(skb)->frag0, grow);
 397
 398         skb->data_len -= grow;
 399         skb->tail += grow;
 400
 401         skb_frag_off_add(&pinfo->frags[0], grow);
 402         skb_frag_size_sub(&pinfo->frags[0], grow);
 403
 404         if (unlikely(!skb_frag_size(&pinfo->frags[0]))) {
 405                 skb_frag_unref(skb, 0);
 406                 memmove(pinfo->frags, pinfo->frags + 1,
 407                         --pinfo->nr_frags * sizeof(pinfo->frags[0]));
 408         }
 409 }
 410
 411 static void gro_try_pull_from_frag0(struct sk_buff *skb)
 412 {
 413         int grow = skb_gro_offset(skb) - skb_headlen(skb);
 414
 415         if (grow > 0)
 416                 gro_pull_from_frag0(skb, grow);
 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
 449         if (netif_elide_gro(skb->dev))
 450                 goto normal;
 451
 452         gro_list_prepare(&gro_list->list, skb);
 453
 454         rcu_read_lock();
 455         list_for_each_entry_rcu(ptype, head, list) {
 456                 if (ptype->type == type && ptype->callbacks.gro_receive)
 457                         goto found_ptype;
 458         }
 459         rcu_read_unlock();
 460         goto normal;
 461
 462 found_ptype:
 463         skb_set_network_header(skb, skb_gro_offset(skb));
 464         skb_reset_mac_len(skb);
 465         BUILD_BUG_ON(sizeof_field(struct napi_gro_cb, zeroed) != sizeof(u32));
 466         BUILD_BUG_ON(!IS_ALIGNED(offsetof(struct napi_gro_cb, zeroed),
 467                                         sizeof(u32))); /* Avoid slow unaligned acc */
 468         *(u32 *)&NAPI_GRO_CB(skb)->zeroed = 0;
 469         NAPI_GRO_CB(skb)->flush = skb_has_frag_list(skb);
 470         NAPI_GRO_CB(skb)->is_atomic = 1;
 471         NAPI_GRO_CB(skb)->count = 1;
 472         if (unlikely(skb_is_gso(skb))) {
 473                 NAPI_GRO_CB(skb)->count = skb_shinfo(skb)->gso_segs;
 474                 /* Only support TCP and non DODGY users. */
 475                 if (!skb_is_gso_tcp(skb) ||
 476                     (skb_shinfo(skb)->gso_type & SKB_GSO_DODGY))
 477                         NAPI_GRO_CB(skb)->flush = 1;
 478         }
 479
 480         /* Setup for GRO checksum validation */
 481         switch (skb->ip_summed) {
 482         case CHECKSUM_COMPLETE:
 483                 NAPI_GRO_CB(skb)->csum = skb->csum;
 484                 NAPI_GRO_CB(skb)->csum_valid = 1;
 485                 break;
 486         case CHECKSUM_UNNECESSARY:
 487                 NAPI_GRO_CB(skb)->csum_cnt = skb->csum_level + 1;
 488                 break;
 489         }
 490
 491         pp = INDIRECT_CALL_INET(ptype->callbacks.gro_receive,
 492                                 ipv6_gro_receive, inet_gro_receive,
 493                                 &gro_list->list, skb);
 494
 495         rcu_read_unlock();
 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         /* Must be called before setting NAPI_GRO_CB(skb)->{age|last} */
 523         gro_try_pull_from_frag0(skb);
 524         NAPI_GRO_CB(skb)->age = jiffies;
 525         NAPI_GRO_CB(skb)->last = skb;
 526         if (!skb_is_gso(skb))
 527                 skb_shinfo(skb)->gso_size = skb_gro_len(skb);
 528         list_add(&skb->list, &gro_list->list);
 529         ret = GRO_HELD;
 530 ok:
 531         if (gro_list->count) {
 532                 if (!test_bit(bucket, &napi->gro_bitmask))
 533                         __set_bit(bucket, &napi->gro_bitmask);
 534         } else if (test_bit(bucket, &napi->gro_bitmask)) {
 535                 __clear_bit(bucket, &napi->gro_bitmask);
 536         }
 537
 538         return ret;
 539
 540 normal:
 541         ret = GRO_NORMAL;
 542         gro_try_pull_from_frag0(skb);
 543         goto ok;
 544 }
 545
 546 struct packet_offload *gro_find_receive_by_type(__be16 type)
 547 {
 548         struct list_head *offload_head = &offload_base;
 549         struct packet_offload *ptype;
 550
 551         list_for_each_entry_rcu(ptype, offload_head, list) {
 552                 if (ptype->type != type || !ptype->callbacks.gro_receive)
 553                         continue;
 554                 return ptype;
 555         }
 556         return NULL;
 557 }
 558 EXPORT_SYMBOL(gro_find_receive_by_type);
 559
 560 struct packet_offload *gro_find_complete_by_type(__be16 type)
 561 {
 562         struct list_head *offload_head = &offload_base;
 563         struct packet_offload *ptype;
 564
 565         list_for_each_entry_rcu(ptype, offload_head, list) {
 566                 if (ptype->type != type || !ptype->callbacks.gro_complete)
 567                         continue;
 568                 return ptype;
 569         }
 570         return NULL;
 571 }
 572 EXPORT_SYMBOL(gro_find_complete_by_type);
 573
 574 static gro_result_t napi_skb_finish(struct napi_struct *napi,
 575                                     struct sk_buff *skb,
 576                                     gro_result_t ret)
 577 {
 578         switch (ret) {
 579         case GRO_NORMAL:
 580                 gro_normal_one(napi, skb, 1);
 581                 break;
 582
 583         case GRO_MERGED_FREE:
 584                 if (NAPI_GRO_CB(skb)->free == NAPI_GRO_FREE_STOLEN_HEAD)
 585                         napi_skb_free_stolen_head(skb);
 586                 else if (skb->fclone != SKB_FCLONE_UNAVAILABLE)
 587                         __kfree_skb(skb);
 588                 else
 589                         __napi_kfree_skb(skb, SKB_CONSUMED);
 590                 break;
 591
 592         case GRO_HELD:
 593         case GRO_MERGED:
 594         case GRO_CONSUMED:
 595                 break;
 596         }
 597
 598         return ret;
 599 }
 600
 601 gro_result_t napi_gro_receive(struct napi_struct *napi, struct sk_buff *skb)
 602 {
 603         gro_result_t ret;
 604
 605         skb_mark_napi_id(skb, napi);
 606         trace_napi_gro_receive_entry(skb);
 607
 608         skb_gro_reset_offset(skb, 0);
 609
 610         ret = napi_skb_finish(napi, skb, dev_gro_receive(napi, skb));
 611         trace_napi_gro_receive_exit(ret);
 612
 613         return ret;
 614 }
 615 EXPORT_SYMBOL(napi_gro_receive);
 616
 617 static void napi_reuse_skb(struct napi_struct *napi, struct sk_buff *skb)
 618 {
 619         if (unlikely(skb->pfmemalloc)) {
 620                 consume_skb(skb);
 621                 return;
 622         }
 623         __skb_pull(skb, skb_headlen(skb));
 624         /* restore the reserve we had after netdev_alloc_skb_ip_align() */
 625         skb_reserve(skb, NET_SKB_PAD + NET_IP_ALIGN - skb_headroom(skb));
 626         __vlan_hwaccel_clear_tag(skb);
 627         skb->dev = napi->dev;
 628         skb->skb_iif = 0;
 629
 630         /* eth_type_trans() assumes pkt_type is PACKET_HOST */
 631         skb->pkt_type = PACKET_HOST;
 632
 633         skb->encapsulation = 0;
 634         skb_shinfo(skb)->gso_type = 0;
 635         skb_shinfo(skb)->gso_size = 0;
 636         if (unlikely(skb->slow_gro)) {
 637                 skb_orphan(skb);
 638                 skb_ext_reset(skb);
 639                 nf_reset_ct(skb);
 640                 skb->slow_gro = 0;
 641         }
 642
 643         napi->skb = skb;
 644 }
 645
 646 struct sk_buff *napi_get_frags(struct napi_struct *napi)
 647 {
 648         struct sk_buff *skb = napi->skb;
 649
 650         if (!skb) {
 651                 skb = napi_alloc_skb(napi, GRO_MAX_HEAD);
 652                 if (skb) {
 653                         napi->skb = skb;
 654                         skb_mark_napi_id(skb, napi);
 655                 }
 656         }
 657         return skb;
 658 }
 659 EXPORT_SYMBOL(napi_get_frags);
 660
 661 static gro_result_t napi_frags_finish(struct napi_struct *napi,
 662                                       struct sk_buff *skb,
 663                                       gro_result_t ret)
 664 {
 665         switch (ret) {
 666         case GRO_NORMAL:
 667         case GRO_HELD:
 668                 __skb_push(skb, ETH_HLEN);
 669                 skb->protocol = eth_type_trans(skb, skb->dev);
 670                 if (ret == GRO_NORMAL)
 671                         gro_normal_one(napi, skb, 1);
 672                 break;
 673
 674         case GRO_MERGED_FREE:
 675                 if (NAPI_GRO_CB(skb)->free == NAPI_GRO_FREE_STOLEN_HEAD)
 676                         napi_skb_free_stolen_head(skb);
 677                 else
 678                         napi_reuse_skb(napi, skb);
 679                 break;
 680
 681         case GRO_MERGED:
 682         case GRO_CONSUMED:
 683                 break;
 684         }
 685
 686         return ret;
 687 }
 688
 689 /* Upper GRO stack assumes network header starts at gro_offset=0
 690  * Drivers could call both napi_gro_frags() and napi_gro_receive()
 691  * We copy ethernet header into skb->data to have a common layout.
 692  */
 693 static struct sk_buff *napi_frags_skb(struct napi_struct *napi)
 694 {
 695         struct sk_buff *skb = napi->skb;
 696         const struct ethhdr *eth;
 697         unsigned int hlen = sizeof(*eth);
 698
 699         napi->skb = NULL;
 700
 701         skb_reset_mac_header(skb);
 702         skb_gro_reset_offset(skb, hlen);
 703
 704         if (unlikely(skb_gro_header_hard(skb, hlen))) {
 705                 eth = skb_gro_header_slow(skb, hlen, 0);
 706                 if (unlikely(!eth)) {
 707                         net_warn_ratelimited("%s: dropping impossible skb from %s\n",
 708                                              __func__, napi->dev->name);
 709                         napi_reuse_skb(napi, skb);
 710                         return NULL;
 711                 }
 712         } else {
 713                 eth = (const struct ethhdr *)skb->data;
 714                 gro_pull_from_frag0(skb, hlen);
 715                 NAPI_GRO_CB(skb)->frag0 += hlen;
 716                 NAPI_GRO_CB(skb)->frag0_len -= hlen;
 717         }
 718         __skb_pull(skb, hlen);
 719
 720         /*
 721          * This works because the only protocols we care about don't require
 722          * special handling.
 723          * We'll fix it up properly in napi_frags_finish()
 724          */
 725         skb->protocol = eth->h_proto;
 726
 727         return skb;
 728 }
 729
 730 gro_result_t napi_gro_frags(struct napi_struct *napi)
 731 {
 732         gro_result_t ret;
 733         struct sk_buff *skb = napi_frags_skb(napi);
 734
 735         trace_napi_gro_frags_entry(skb);
 736
 737         ret = napi_frags_finish(napi, skb, dev_gro_receive(napi, skb));
 738         trace_napi_gro_frags_exit(ret);
 739
 740         return ret;
 741 }
 742 EXPORT_SYMBOL(napi_gro_frags);
 743
 744 /* Compute the checksum from gro_offset and return the folded value
 745  * after adding in any pseudo checksum.
 746  */
 747 __sum16 __skb_gro_checksum_complete(struct sk_buff *skb)
 748 {
 749         __wsum wsum;
 750         __sum16 sum;
 751
 752         wsum = skb_checksum(skb, skb_gro_offset(skb), skb_gro_len(skb), 0);
 753
 754         /* NAPI_GRO_CB(skb)->csum holds pseudo checksum */
 755         sum = csum_fold(csum_add(NAPI_GRO_CB(skb)->csum, wsum));
 756         /* See comments in __skb_checksum_complete(). */
 757         if (likely(!sum)) {
 758                 if (unlikely(skb->ip_summed == CHECKSUM_COMPLETE) &&
 759                     !skb->csum_complete_sw)
 760                         netdev_rx_csum_fault(skb->dev, skb);
 761         }
 762
 763         NAPI_GRO_CB(skb)->csum = wsum;
 764         NAPI_GRO_CB(skb)->csum_valid = 1;
 765
 766         return sum;
 767 }
 768 EXPORT_SYMBOL(__skb_gro_checksum_complete);