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
  14 /**
  15  *      dev_add_offload - register offload handlers
  16  *      @po: protocol offload declaration
  17  *
  18  *      Add protocol offload handlers to the networking stack. The passed
  19  *      &proto_offload is linked into kernel lists and may not be freed until
  20  *      it has been removed from the kernel lists.
  21  *
  22  *      This call does not sleep therefore it can not
  23  *      guarantee all CPU's that are in middle of receiving packets
  24  *      will see the new offload handlers (until the next received packet).
  25  */
  26 void dev_add_offload(struct packet_offload *po)
  27 {
  28         struct packet_offload *elem;
  29
  30         spin_lock(&offload_lock);
  31         list_for_each_entry(elem, &net_hotdata.offload_base, list) {
  32                 if (po->priority < elem->priority)
  33                         break;
  34         }
  35         list_add_rcu(&po->list, elem->list.prev);
  36         spin_unlock(&offload_lock);
  37 }
  38 EXPORT_SYMBOL(dev_add_offload);
  39
  40 /**
  41  *      __dev_remove_offload     - remove offload handler
  42  *      @po: packet offload declaration
  43  *
  44  *      Remove a protocol offload handler that was previously added to the
  45  *      kernel offload handlers by dev_add_offload(). The passed &offload_type
  46  *      is removed from the kernel lists and can be freed or reused once this
  47  *      function returns.
  48  *
  49  *      The packet type might still be in use by receivers
  50  *      and must not be freed until after all the CPU's have gone
  51  *      through a quiescent state.
  52  */
  53 static void __dev_remove_offload(struct packet_offload *po)
  54 {
  55         struct list_head *head = &net_hotdata.offload_base;
  56         struct packet_offload *po1;
  57
  58         spin_lock(&offload_lock);
  59
  60         list_for_each_entry(po1, head, list) {
  61                 if (po == po1) {
  62                         list_del_rcu(&po->list);
  63                         goto out;
  64                 }
  65         }
  66
  67         pr_warn("dev_remove_offload: %p not found\n", po);
  68 out:
  69         spin_unlock(&offload_lock);
  70 }
  71
  72 /**
  73  *      dev_remove_offload       - remove packet offload handler
  74  *      @po: packet offload declaration
  75  *
  76  *      Remove a packet offload handler that was previously added to the kernel
  77  *      offload handlers by dev_add_offload(). The passed &offload_type is
  78  *      removed from the kernel lists and can be freed or reused once this
  79  *      function returns.
  80  *
  81  *      This call sleeps to guarantee that no CPU is looking at the packet
  82  *      type after return.
  83  */
  84 void dev_remove_offload(struct packet_offload *po)
  85 {
  86         __dev_remove_offload(po);
  87
  88         synchronize_net();
  89 }
  90 EXPORT_SYMBOL(dev_remove_offload);
  91
  92
  93 int skb_gro_receive(struct sk_buff *p, struct sk_buff *skb)
  94 {
  95         struct skb_shared_info *pinfo, *skbinfo = skb_shinfo(skb);
  96         unsigned int offset = skb_gro_offset(skb);
  97         unsigned int headlen = skb_headlen(skb);
  98         unsigned int len = skb_gro_len(skb);
  99         unsigned int delta_truesize;
 100         unsigned int gro_max_size;
 101         unsigned int new_truesize;
 102         struct sk_buff *lp;
 103         int segs;
 104
 105         /* Do not splice page pool based packets w/ non-page pool
 106          * packets. This can result in reference count issues as page
 107          * pool pages will not decrement the reference count and will
 108          * instead be immediately returned to the pool or have frag
 109          * count decremented.
 110          */
 111         if (p->pp_recycle != skb->pp_recycle)
 112                 return -ETOOMANYREFS;
 113
 114         /* pairs with WRITE_ONCE() in netif_set_gro(_ipv4)_max_size() */
 115         gro_max_size = p->protocol == htons(ETH_P_IPV6) ?
 116                         READ_ONCE(p->dev->gro_max_size) :
 117                         READ_ONCE(p->dev->gro_ipv4_max_size);
 118
 119         if (unlikely(p->len + len >= gro_max_size || NAPI_GRO_CB(skb)->flush))
 120                 return -E2BIG;
 121
 122         if (unlikely(p->len + len >= GRO_LEGACY_MAX_SIZE)) {
 123                 if (NAPI_GRO_CB(skb)->proto != IPPROTO_TCP ||
 124                     (p->protocol == htons(ETH_P_IPV6) &&
 125                      skb_headroom(p) < sizeof(struct hop_jumbo_hdr)) ||
 126                     p->encapsulation)
 127                         return -E2BIG;
 128         }
 129
 130         segs = NAPI_GRO_CB(skb)->count;
 131         lp = NAPI_GRO_CB(p)->last;
 132         pinfo = skb_shinfo(lp);
 133
 134         if (headlen <= offset) {
 135                 skb_frag_t *frag;
 136                 skb_frag_t *frag2;
 137                 int i = skbinfo->nr_frags;
 138                 int nr_frags = pinfo->nr_frags + i;
 139
 140                 if (nr_frags > MAX_SKB_FRAGS)
 141                         goto merge;
 142
 143                 offset -= headlen;
 144                 pinfo->nr_frags = nr_frags;
 145                 skbinfo->nr_frags = 0;
 146
 147                 frag = pinfo->frags + nr_frags;
 148                 frag2 = skbinfo->frags + i;
 149                 do {
 150                         *--frag = *--frag2;
 151                 } while (--i);
 152
 153                 skb_frag_off_add(frag, offset);
 154                 skb_frag_size_sub(frag, offset);
 155
 156                 /* all fragments truesize : remove (head size + sk_buff) */
 157                 new_truesize = SKB_TRUESIZE(skb_end_offset(skb));
 158                 delta_truesize = skb->truesize - new_truesize;
 159
 160                 skb->truesize = new_truesize;
 161                 skb->len -= skb->data_len;
 162                 skb->data_len = 0;
 163
 164                 NAPI_GRO_CB(skb)->free = NAPI_GRO_FREE;
 165                 goto done;
 166         } else if (skb->head_frag) {
 167                 int nr_frags = pinfo->nr_frags;
 168                 skb_frag_t *frag = pinfo->frags + nr_frags;
 169                 struct page *page = virt_to_head_page(skb->head);
 170                 unsigned int first_size = headlen - offset;
 171                 unsigned int first_offset;
 172
 173                 if (nr_frags + 1 + skbinfo->nr_frags > MAX_SKB_FRAGS)
 174                         goto merge;
 175
 176                 first_offset = skb->data -
 177                                (unsigned char *)page_address(page) +
 178                                offset;
 179
 180                 pinfo->nr_frags = nr_frags + 1 + skbinfo->nr_frags;
 181
 182                 skb_frag_fill_page_desc(frag, page, first_offset, first_size);
 183
 184                 memcpy(frag + 1, skbinfo->frags, sizeof(*frag) * skbinfo->nr_frags);
 185                 /* We dont need to clear skbinfo->nr_frags here */
 186
 187                 new_truesize = SKB_DATA_ALIGN(sizeof(struct sk_buff));
 188                 delta_truesize = skb->truesize - new_truesize;
 189                 skb->truesize = new_truesize;
 190                 NAPI_GRO_CB(skb)->free = NAPI_GRO_FREE_STOLEN_HEAD;
 191                 goto done;
 192         }
 193
 194 merge:
 195         /* sk owenrship - if any - completely transferred to the aggregated packet */
 196         skb->destructor = NULL;
 197         delta_truesize = skb->truesize;
 198         if (offset > headlen) {
 199                 unsigned int eat = offset - headlen;
 200
 201                 skb_frag_off_add(&skbinfo->frags[0], eat);
 202                 skb_frag_size_sub(&skbinfo->frags[0], eat);
 203                 skb->data_len -= eat;
 204                 skb->len -= eat;
 205                 offset = headlen;
 206         }
 207
 208         __skb_pull(skb, offset);
 209
 210         if (NAPI_GRO_CB(p)->last == p)
 211                 skb_shinfo(p)->frag_list = skb;
 212         else
 213                 NAPI_GRO_CB(p)->last->next = skb;
 214         NAPI_GRO_CB(p)->last = skb;
 215         __skb_header_release(skb);
 216         lp = p;
 217
 218 done:
 219         NAPI_GRO_CB(p)->count += segs;
 220         p->data_len += len;
 221         p->truesize += delta_truesize;
 222         p->len += len;
 223         if (lp != p) {
 224                 lp->data_len += len;
 225                 lp->truesize += delta_truesize;
 226                 lp->len += len;
 227         }
 228         NAPI_GRO_CB(skb)->same_flow = 1;
 229         return 0;
 230 }
 231
 232
 233 static void napi_gro_complete(struct napi_struct *napi, struct sk_buff *skb)
 234 {
 235         struct list_head *head = &net_hotdata.offload_base;
 236         struct packet_offload *ptype;
 237         __be16 type = skb->protocol;
 238         int err = -ENOENT;
 239
 240         BUILD_BUG_ON(sizeof(struct napi_gro_cb) > sizeof(skb->cb));
 241
 242         if (NAPI_GRO_CB(skb)->count == 1) {
 243                 skb_shinfo(skb)->gso_size = 0;
 244                 goto out;
 245         }
 246
 247         rcu_read_lock();
 248         list_for_each_entry_rcu(ptype, head, list) {
 249                 if (ptype->type != type || !ptype->callbacks.gro_complete)
 250                         continue;
 251
 252                 err = INDIRECT_CALL_INET(ptype->callbacks.gro_complete,
 253                                          ipv6_gro_complete, inet_gro_complete,
 254                                          skb, 0);
 255                 break;
 256         }
 257         rcu_read_unlock();
 258
 259         if (err) {
 260                 WARN_ON(&ptype->list == head);
 261                 kfree_skb(skb);
 262                 return;
 263         }
 264
 265 out:
 266         gro_normal_one(napi, skb, NAPI_GRO_CB(skb)->count);
 267 }
 268
 269 static void __napi_gro_flush_chain(struct napi_struct *napi, u32 index,
 270                                    bool flush_old)
 271 {
 272         struct list_head *head = &napi->gro_hash[index].list;
 273         struct sk_buff *skb, *p;
 274
 275         list_for_each_entry_safe_reverse(skb, p, head, list) {
 276                 if (flush_old && NAPI_GRO_CB(skb)->age == jiffies)
 277                         return;
 278                 skb_list_del_init(skb);
 279                 napi_gro_complete(napi, skb);
 280                 napi->gro_hash[index].count--;
 281         }
 282
 283         if (!napi->gro_hash[index].count)
 284                 __clear_bit(index, &napi->gro_bitmask);
 285 }
 286
 287 /* napi->gro_hash[].list contains packets ordered by age.
 288  * youngest packets at the head of it.
 289  * Complete skbs in reverse order to reduce latencies.
 290  */
 291 void napi_gro_flush(struct napi_struct *napi, bool flush_old)
 292 {
 293         unsigned long bitmask = napi->gro_bitmask;
 294         unsigned int i, base = ~0U;
 295
 296         while ((i = ffs(bitmask)) != 0) {
 297                 bitmask >>= i;
 298                 base += i;
 299                 __napi_gro_flush_chain(napi, base, flush_old);
 300         }
 301 }
 302 EXPORT_SYMBOL(napi_gro_flush);
 303
 304 static unsigned long gro_list_prepare_tc_ext(const struct sk_buff *skb,
 305                                              const struct sk_buff *p,
 306                                              unsigned long diffs)
 307 {
 308 #if IS_ENABLED(CONFIG_NET_TC_SKB_EXT)
 309         struct tc_skb_ext *skb_ext;
 310         struct tc_skb_ext *p_ext;
 311
 312         skb_ext = skb_ext_find(skb, TC_SKB_EXT);
 313         p_ext = skb_ext_find(p, TC_SKB_EXT);
 314
 315         diffs |= (!!p_ext) ^ (!!skb_ext);
 316         if (!diffs && unlikely(skb_ext))
 317                 diffs |= p_ext->chain ^ skb_ext->chain;
 318 #endif
 319         return diffs;
 320 }
 321
 322 static void gro_list_prepare(const struct list_head *head,
 323                              const struct sk_buff *skb)
 324 {
 325         unsigned int maclen = skb->dev->hard_header_len;
 326         u32 hash = skb_get_hash_raw(skb);
 327         struct sk_buff *p;
 328
 329         list_for_each_entry(p, head, list) {
 330                 unsigned long diffs;
 331
 332                 NAPI_GRO_CB(p)->flush = 0;
 333
 334                 if (hash != skb_get_hash_raw(p)) {
 335                         NAPI_GRO_CB(p)->same_flow = 0;
 336                         continue;
 337                 }
 338
 339                 diffs = (unsigned long)p->dev ^ (unsigned long)skb->dev;
 340                 diffs |= p->vlan_all ^ skb->vlan_all;
 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                         diffs |= p->sk != skb->sk;
 357                         diffs |= skb_metadata_dst_cmp(p, skb);
 358                         diffs |= skb_get_nfct(p) ^ skb_get_nfct(skb);
 359
 360                         diffs |= gro_list_prepare_tc_ext(skb, p, diffs);
 361                 }
 362
 363                 NAPI_GRO_CB(p)->same_flow = !diffs;
 364         }
 365 }
 366
 367 static inline void skb_gro_reset_offset(struct sk_buff *skb, u32 nhoff)
 368 {
 369         const struct skb_shared_info *pinfo;
 370         const skb_frag_t *frag0;
 371         unsigned int headlen;
 372
 373         NAPI_GRO_CB(skb)->data_offset = 0;
 374         headlen = skb_headlen(skb);
 375         NAPI_GRO_CB(skb)->frag0 = skb->data;
 376         NAPI_GRO_CB(skb)->frag0_len = headlen;
 377         if (headlen)
 378                 return;
 379
 380         pinfo = skb_shinfo(skb);
 381         frag0 = &pinfo->frags[0];
 382
 383         if (pinfo->nr_frags && !PageHighMem(skb_frag_page(frag0)) &&
 384             (!NET_IP_ALIGN || !((skb_frag_off(frag0) + nhoff) & 3))) {
 385                 NAPI_GRO_CB(skb)->frag0 = skb_frag_address(frag0);
 386                 NAPI_GRO_CB(skb)->frag0_len = min_t(unsigned int,
 387                                                     skb_frag_size(frag0),
 388                                                     skb->end - skb->tail);
 389         }
 390 }
 391
 392 static void gro_pull_from_frag0(struct sk_buff *skb, int grow)
 393 {
 394         struct skb_shared_info *pinfo = skb_shinfo(skb);
 395
 396         BUG_ON(skb->end - skb->tail < grow);
 397
 398         memcpy(skb_tail_pointer(skb), NAPI_GRO_CB(skb)->frag0, grow);
 399
 400         skb->data_len -= grow;
 401         skb->tail += grow;
 402
 403         skb_frag_off_add(&pinfo->frags[0], grow);
 404         skb_frag_size_sub(&pinfo->frags[0], grow);
 405
 406         if (unlikely(!skb_frag_size(&pinfo->frags[0]))) {
 407                 skb_frag_unref(skb, 0);
 408                 memmove(pinfo->frags, pinfo->frags + 1,
 409                         --pinfo->nr_frags * sizeof(pinfo->frags[0]));
 410         }
 411 }
 412
 413 static void gro_try_pull_from_frag0(struct sk_buff *skb)
 414 {
 415         int grow = skb_gro_offset(skb) - skb_headlen(skb);
 416
 417         if (grow > 0)
 418                 gro_pull_from_frag0(skb, grow);
 419 }
 420
 421 static void gro_flush_oldest(struct napi_struct *napi, struct list_head *head)
 422 {
 423         struct sk_buff *oldest;
 424
 425         oldest = list_last_entry(head, struct sk_buff, list);
 426
 427         /* We are called with head length >= MAX_GRO_SKBS, so this is
 428          * impossible.
 429          */
 430         if (WARN_ON_ONCE(!oldest))
 431                 return;
 432
 433         /* Do not adjust napi->gro_hash[].count, caller is adding a new
 434          * SKB to the chain.
 435          */
 436         skb_list_del_init(oldest);
 437         napi_gro_complete(napi, oldest);
 438 }
 439
 440 static enum gro_result dev_gro_receive(struct napi_struct *napi, struct sk_buff *skb)
 441 {
 442         u32 bucket = skb_get_hash_raw(skb) & (GRO_HASH_BUCKETS - 1);
 443         struct gro_list *gro_list = &napi->gro_hash[bucket];
 444         struct list_head *head = &net_hotdata.offload_base;
 445         struct packet_offload *ptype;
 446         __be16 type = skb->protocol;
 447         struct sk_buff *pp = NULL;
 448         enum gro_result ret;
 449         int same_flow;
 450
 451         if (netif_elide_gro(skb->dev))
 452                 goto normal;
 453
 454         gro_list_prepare(&gro_list->list, skb);
 455
 456         rcu_read_lock();
 457         list_for_each_entry_rcu(ptype, head, list) {
 458                 if (ptype->type == type && ptype->callbacks.gro_receive)
 459                         goto found_ptype;
 460         }
 461         rcu_read_unlock();
 462         goto normal;
 463
 464 found_ptype:
 465         skb_set_network_header(skb, skb_gro_offset(skb));
 466         skb_reset_mac_len(skb);
 467         BUILD_BUG_ON(sizeof_field(struct napi_gro_cb, zeroed) != sizeof(u32));
 468         BUILD_BUG_ON(!IS_ALIGNED(offsetof(struct napi_gro_cb, zeroed),
 469                                         sizeof(u32))); /* Avoid slow unaligned acc */
 470         *(u32 *)&NAPI_GRO_CB(skb)->zeroed = 0;
 471         NAPI_GRO_CB(skb)->flush = skb_has_frag_list(skb);
 472         NAPI_GRO_CB(skb)->is_atomic = 1;
 473         NAPI_GRO_CB(skb)->count = 1;
 474         if (unlikely(skb_is_gso(skb))) {
 475                 NAPI_GRO_CB(skb)->count = skb_shinfo(skb)->gso_segs;
 476                 /* Only support TCP and non DODGY users. */
 477                 if (!skb_is_gso_tcp(skb) ||
 478                     (skb_shinfo(skb)->gso_type & SKB_GSO_DODGY))
 479                         NAPI_GRO_CB(skb)->flush = 1;
 480         }
 481
 482         /* Setup for GRO checksum validation */
 483         switch (skb->ip_summed) {
 484         case CHECKSUM_COMPLETE:
 485                 NAPI_GRO_CB(skb)->csum = skb->csum;
 486                 NAPI_GRO_CB(skb)->csum_valid = 1;
 487                 break;
 488         case CHECKSUM_UNNECESSARY:
 489                 NAPI_GRO_CB(skb)->csum_cnt = skb->csum_level + 1;
 490                 break;
 491         }
 492
 493         pp = INDIRECT_CALL_INET(ptype->callbacks.gro_receive,
 494                                 ipv6_gro_receive, inet_gro_receive,
 495                                 &gro_list->list, skb);
 496
 497         rcu_read_unlock();
 498
 499         if (PTR_ERR(pp) == -EINPROGRESS) {
 500                 ret = GRO_CONSUMED;
 501                 goto ok;
 502         }
 503
 504         same_flow = NAPI_GRO_CB(skb)->same_flow;
 505         ret = NAPI_GRO_CB(skb)->free ? GRO_MERGED_FREE : GRO_MERGED;
 506
 507         if (pp) {
 508                 skb_list_del_init(pp);
 509                 napi_gro_complete(napi, pp);
 510                 gro_list->count--;
 511         }
 512
 513         if (same_flow)
 514                 goto ok;
 515
 516         if (NAPI_GRO_CB(skb)->flush)
 517                 goto normal;
 518
 519         if (unlikely(gro_list->count >= MAX_GRO_SKBS))
 520                 gro_flush_oldest(napi, &gro_list->list);
 521         else
 522                 gro_list->count++;
 523
 524         /* Must be called before setting NAPI_GRO_CB(skb)->{age|last} */
 525         gro_try_pull_from_frag0(skb);
 526         NAPI_GRO_CB(skb)->age = jiffies;
 527         NAPI_GRO_CB(skb)->last = skb;
 528         if (!skb_is_gso(skb))
 529                 skb_shinfo(skb)->gso_size = skb_gro_len(skb);
 530         list_add(&skb->list, &gro_list->list);
 531         ret = GRO_HELD;
 532 ok:
 533         if (gro_list->count) {
 534                 if (!test_bit(bucket, &napi->gro_bitmask))
 535                         __set_bit(bucket, &napi->gro_bitmask);
 536         } else if (test_bit(bucket, &napi->gro_bitmask)) {
 537                 __clear_bit(bucket, &napi->gro_bitmask);
 538         }
 539
 540         return ret;
 541
 542 normal:
 543         ret = GRO_NORMAL;
 544         gro_try_pull_from_frag0(skb);
 545         goto ok;
 546 }
 547
 548 struct packet_offload *gro_find_receive_by_type(__be16 type)
 549 {
 550         struct list_head *offload_head = &net_hotdata.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 = &net_hotdata.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                         __napi_kfree_skb(skb, SKB_CONSUMED);
 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_shinfo(skb)->gso_size = 0;
 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_may_pull(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
 717                 if (NAPI_GRO_CB(skb)->frag0 != skb->data)
 718                         gro_pull_from_frag0(skb, hlen);
 719
 720                 NAPI_GRO_CB(skb)->frag0 += hlen;
 721                 NAPI_GRO_CB(skb)->frag0_len -= hlen;
 722         }
 723         __skb_pull(skb, hlen);
 724
 725         /*
 726          * This works because the only protocols we care about don't require
 727          * special handling.
 728          * We'll fix it up properly in napi_frags_finish()
 729          */
 730         skb->protocol = eth->h_proto;
 731
 732         return skb;
 733 }
 734
 735 gro_result_t napi_gro_frags(struct napi_struct *napi)
 736 {
 737         gro_result_t ret;
 738         struct sk_buff *skb = napi_frags_skb(napi);
 739
 740         trace_napi_gro_frags_entry(skb);
 741
 742         ret = napi_frags_finish(napi, skb, dev_gro_receive(napi, skb));
 743         trace_napi_gro_frags_exit(ret);
 744
 745         return ret;
 746 }
 747 EXPORT_SYMBOL(napi_gro_frags);
 748
 749 /* Compute the checksum from gro_offset and return the folded value
 750  * after adding in any pseudo checksum.
 751  */
 752 __sum16 __skb_gro_checksum_complete(struct sk_buff *skb)
 753 {
 754         __wsum wsum;
 755         __sum16 sum;
 756
 757         wsum = skb_checksum(skb, skb_gro_offset(skb), skb_gro_len(skb), 0);
 758
 759         /* NAPI_GRO_CB(skb)->csum holds pseudo checksum */
 760         sum = csum_fold(csum_add(NAPI_GRO_CB(skb)->csum, wsum));
 761         /* See comments in __skb_checksum_complete(). */
 762         if (likely(!sum)) {
 763                 if (unlikely(skb->ip_summed == CHECKSUM_COMPLETE) &&
 764                     !skb->csum_complete_sw)
 765                         netdev_rx_csum_fault(skb->dev, skb);
 766         }
 767
 768         NAPI_GRO_CB(skb)->csum = wsum;
 769         NAPI_GRO_CB(skb)->csum_valid = 1;
 770
 771         return sum;
 772 }
 773 EXPORT_SYMBOL(__skb_gro_checksum_complete);