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 ownership - if any - completely transferred to the aggregated packet */
 196         skb->destructor = NULL;
 197         skb->sk = NULL;
 198         delta_truesize = skb->truesize;
 199         if (offset > headlen) {
 200                 unsigned int eat = offset - headlen;
 201
 202                 skb_frag_off_add(&skbinfo->frags[0], eat);
 203                 skb_frag_size_sub(&skbinfo->frags[0], eat);
 204                 skb->data_len -= eat;
 205                 skb->len -= eat;
 206                 offset = headlen;
 207         }
 208
 209         __skb_pull(skb, offset);
 210
 211         if (NAPI_GRO_CB(p)->last == p)
 212                 skb_shinfo(p)->frag_list = skb;
 213         else
 214                 NAPI_GRO_CB(p)->last->next = skb;
 215         NAPI_GRO_CB(p)->last = skb;
 216         __skb_header_release(skb);
 217         lp = p;
 218
 219 done:
 220         NAPI_GRO_CB(p)->count += segs;
 221         p->data_len += len;
 222         p->truesize += delta_truesize;
 223         p->len += len;
 224         if (lp != p) {
 225                 lp->data_len += len;
 226                 lp->truesize += delta_truesize;
 227                 lp->len += len;
 228         }
 229         NAPI_GRO_CB(skb)->same_flow = 1;
 230         return 0;
 231 }
 232
 233
 234 static void napi_gro_complete(struct napi_struct *napi, struct sk_buff *skb)
 235 {
 236         struct list_head *head = &net_hotdata.offload_base;
 237         struct packet_offload *ptype;
 238         __be16 type = skb->protocol;
 239         int err = -ENOENT;
 240
 241         BUILD_BUG_ON(sizeof(struct napi_gro_cb) > sizeof(skb->cb));
 242
 243         if (NAPI_GRO_CB(skb)->count == 1) {
 244                 skb_shinfo(skb)->gso_size = 0;
 245                 goto out;
 246         }
 247
 248         rcu_read_lock();
 249         list_for_each_entry_rcu(ptype, head, list) {
 250                 if (ptype->type != type || !ptype->callbacks.gro_complete)
 251                         continue;
 252
 253                 err = INDIRECT_CALL_INET(ptype->callbacks.gro_complete,
 254                                          ipv6_gro_complete, inet_gro_complete,
 255                                          skb, 0);
 256                 break;
 257         }
 258         rcu_read_unlock();
 259
 260         if (err) {
 261                 WARN_ON(&ptype->list == head);
 262                 kfree_skb(skb);
 263                 return;
 264         }
 265
 266 out:
 267         gro_normal_one(napi, skb, NAPI_GRO_CB(skb)->count);
 268 }
 269
 270 static void __napi_gro_flush_chain(struct napi_struct *napi, u32 index,
 271                                    bool flush_old)
 272 {
 273         struct list_head *head = &napi->gro_hash[index].list;
 274         struct sk_buff *skb, *p;
 275
 276         list_for_each_entry_safe_reverse(skb, p, head, list) {
 277                 if (flush_old && NAPI_GRO_CB(skb)->age == jiffies)
 278                         return;
 279                 skb_list_del_init(skb);
 280                 napi_gro_complete(napi, skb);
 281                 napi->gro_hash[index].count--;
 282         }
 283
 284         if (!napi->gro_hash[index].count)
 285                 __clear_bit(index, &napi->gro_bitmask);
 286 }
 287
 288 /* napi->gro_hash[].list contains packets ordered by age.
 289  * youngest packets at the head of it.
 290  * Complete skbs in reverse order to reduce latencies.
 291  */
 292 void napi_gro_flush(struct napi_struct *napi, bool flush_old)
 293 {
 294         unsigned long bitmask = napi->gro_bitmask;
 295         unsigned int i, base = ~0U;
 296
 297         while ((i = ffs(bitmask)) != 0) {
 298                 bitmask >>= i;
 299                 base += i;
 300                 __napi_gro_flush_chain(napi, base, flush_old);
 301         }
 302 }
 303 EXPORT_SYMBOL(napi_gro_flush);
 304
 305 static unsigned long gro_list_prepare_tc_ext(const struct sk_buff *skb,
 306                                              const struct sk_buff *p,
 307                                              unsigned long diffs)
 308 {
 309 #if IS_ENABLED(CONFIG_NET_TC_SKB_EXT)
 310         struct tc_skb_ext *skb_ext;
 311         struct tc_skb_ext *p_ext;
 312
 313         skb_ext = skb_ext_find(skb, TC_SKB_EXT);
 314         p_ext = skb_ext_find(p, TC_SKB_EXT);
 315
 316         diffs |= (!!p_ext) ^ (!!skb_ext);
 317         if (!diffs && unlikely(skb_ext))
 318                 diffs |= p_ext->chain ^ skb_ext->chain;
 319 #endif
 320         return diffs;
 321 }
 322
 323 static void gro_list_prepare(const struct list_head *head,
 324                              const struct sk_buff *skb)
 325 {
 326         unsigned int maclen = skb->dev->hard_header_len;
 327         u32 hash = skb_get_hash_raw(skb);
 328         struct sk_buff *p;
 329
 330         list_for_each_entry(p, head, list) {
 331                 unsigned long diffs;
 332
 333                 NAPI_GRO_CB(p)->flush = 0;
 334
 335                 if (hash != skb_get_hash_raw(p)) {
 336                         NAPI_GRO_CB(p)->same_flow = 0;
 337                         continue;
 338                 }
 339
 340                 diffs = (unsigned long)p->dev ^ (unsigned long)skb->dev;
 341                 diffs |= p->vlan_all ^ skb->vlan_all;
 342                 diffs |= skb_metadata_differs(p, skb);
 343                 if (maclen == ETH_HLEN)
 344                         diffs |= compare_ether_header(skb_mac_header(p),
 345                                                       skb_mac_header(skb));
 346                 else if (!diffs)
 347                         diffs = memcmp(skb_mac_header(p),
 348                                        skb_mac_header(skb),
 349                                        maclen);
 350
 351                 /* in most common scenarions 'slow_gro' is 0
 352                  * otherwise we are already on some slower paths
 353                  * either skip all the infrequent tests altogether or
 354                  * avoid trying too hard to skip each of them individually
 355                  */
 356                 if (!diffs && unlikely(skb->slow_gro | p->slow_gro)) {
 357                         diffs |= p->sk != skb->sk;
 358                         diffs |= skb_metadata_dst_cmp(p, skb);
 359                         diffs |= skb_get_nfct(p) ^ skb_get_nfct(skb);
 360
 361                         diffs |= gro_list_prepare_tc_ext(skb, p, diffs);
 362                 }
 363
 364                 NAPI_GRO_CB(p)->same_flow = !diffs;
 365         }
 366 }
 367
 368 static inline void skb_gro_reset_offset(struct sk_buff *skb, u32 nhoff)
 369 {
 370         const struct skb_shared_info *pinfo;
 371         const skb_frag_t *frag0;
 372         unsigned int headlen;
 373
 374         NAPI_GRO_CB(skb)->data_offset = 0;
 375         headlen = skb_headlen(skb);
 376         NAPI_GRO_CB(skb)->frag0 = skb->data;
 377         NAPI_GRO_CB(skb)->frag0_len = headlen;
 378         if (headlen)
 379                 return;
 380
 381         pinfo = skb_shinfo(skb);
 382         frag0 = &pinfo->frags[0];
 383
 384         if (pinfo->nr_frags && !PageHighMem(skb_frag_page(frag0)) &&
 385             (!NET_IP_ALIGN || !((skb_frag_off(frag0) + nhoff) & 3))) {
 386                 NAPI_GRO_CB(skb)->frag0 = skb_frag_address(frag0);
 387                 NAPI_GRO_CB(skb)->frag0_len = min_t(unsigned int,
 388                                                     skb_frag_size(frag0),
 389                                                     skb->end - skb->tail);
 390         }
 391 }
 392
 393 static void gro_pull_from_frag0(struct sk_buff *skb, int grow)
 394 {
 395         struct skb_shared_info *pinfo = skb_shinfo(skb);
 396
 397         BUG_ON(skb->end - skb->tail < grow);
 398
 399         memcpy(skb_tail_pointer(skb), NAPI_GRO_CB(skb)->frag0, grow);
 400
 401         skb->data_len -= grow;
 402         skb->tail += grow;
 403
 404         skb_frag_off_add(&pinfo->frags[0], grow);
 405         skb_frag_size_sub(&pinfo->frags[0], grow);
 406
 407         if (unlikely(!skb_frag_size(&pinfo->frags[0]))) {
 408                 skb_frag_unref(skb, 0);
 409                 memmove(pinfo->frags, pinfo->frags + 1,
 410                         --pinfo->nr_frags * sizeof(pinfo->frags[0]));
 411         }
 412 }
 413
 414 static void gro_try_pull_from_frag0(struct sk_buff *skb)
 415 {
 416         int grow = skb_gro_offset(skb) - skb_headlen(skb);
 417
 418         if (grow > 0)
 419                 gro_pull_from_frag0(skb, grow);
 420 }
 421
 422 static void gro_flush_oldest(struct napi_struct *napi, struct list_head *head)
 423 {
 424         struct sk_buff *oldest;
 425
 426         oldest = list_last_entry(head, struct sk_buff, list);
 427
 428         /* We are called with head length >= MAX_GRO_SKBS, so this is
 429          * impossible.
 430          */
 431         if (WARN_ON_ONCE(!oldest))
 432                 return;
 433
 434         /* Do not adjust napi->gro_hash[].count, caller is adding a new
 435          * SKB to the chain.
 436          */
 437         skb_list_del_init(oldest);
 438         napi_gro_complete(napi, oldest);
 439 }
 440
 441 static enum gro_result dev_gro_receive(struct napi_struct *napi, struct sk_buff *skb)
 442 {
 443         u32 bucket = skb_get_hash_raw(skb) & (GRO_HASH_BUCKETS - 1);
 444         struct gro_list *gro_list = &napi->gro_hash[bucket];
 445         struct list_head *head = &net_hotdata.offload_base;
 446         struct packet_offload *ptype;
 447         __be16 type = skb->protocol;
 448         struct sk_buff *pp = NULL;
 449         enum gro_result ret;
 450         int same_flow;
 451
 452         if (netif_elide_gro(skb->dev))
 453                 goto normal;
 454
 455         gro_list_prepare(&gro_list->list, skb);
 456
 457         rcu_read_lock();
 458         list_for_each_entry_rcu(ptype, head, list) {
 459                 if (ptype->type == type && ptype->callbacks.gro_receive)
 460                         goto found_ptype;
 461         }
 462         rcu_read_unlock();
 463         goto normal;
 464
 465 found_ptype:
 466         skb_set_network_header(skb, skb_gro_offset(skb));
 467         skb_reset_mac_len(skb);
 468         BUILD_BUG_ON(sizeof_field(struct napi_gro_cb, zeroed) != sizeof(u32));
 469         BUILD_BUG_ON(!IS_ALIGNED(offsetof(struct napi_gro_cb, zeroed),
 470                                         sizeof(u32))); /* Avoid slow unaligned acc */
 471         *(u32 *)&NAPI_GRO_CB(skb)->zeroed = 0;
 472         NAPI_GRO_CB(skb)->flush = skb_has_frag_list(skb);
 473         NAPI_GRO_CB(skb)->is_atomic = 1;
 474         NAPI_GRO_CB(skb)->count = 1;
 475         if (unlikely(skb_is_gso(skb))) {
 476                 NAPI_GRO_CB(skb)->count = skb_shinfo(skb)->gso_segs;
 477                 /* Only support TCP and non DODGY users. */
 478                 if (!skb_is_gso_tcp(skb) ||
 479                     (skb_shinfo(skb)->gso_type & SKB_GSO_DODGY))
 480                         NAPI_GRO_CB(skb)->flush = 1;
 481         }
 482
 483         /* Setup for GRO checksum validation */
 484         switch (skb->ip_summed) {
 485         case CHECKSUM_COMPLETE:
 486                 NAPI_GRO_CB(skb)->csum = skb->csum;
 487                 NAPI_GRO_CB(skb)->csum_valid = 1;
 488                 break;
 489         case CHECKSUM_UNNECESSARY:
 490                 NAPI_GRO_CB(skb)->csum_cnt = skb->csum_level + 1;
 491                 break;
 492         }
 493
 494         pp = INDIRECT_CALL_INET(ptype->callbacks.gro_receive,
 495                                 ipv6_gro_receive, inet_gro_receive,
 496                                 &gro_list->list, skb);
 497
 498         rcu_read_unlock();
 499
 500         if (PTR_ERR(pp) == -EINPROGRESS) {
 501                 ret = GRO_CONSUMED;
 502                 goto ok;
 503         }
 504
 505         same_flow = NAPI_GRO_CB(skb)->same_flow;
 506         ret = NAPI_GRO_CB(skb)->free ? GRO_MERGED_FREE : GRO_MERGED;
 507
 508         if (pp) {
 509                 skb_list_del_init(pp);
 510                 napi_gro_complete(napi, pp);
 511                 gro_list->count--;
 512         }
 513
 514         if (same_flow)
 515                 goto ok;
 516
 517         if (NAPI_GRO_CB(skb)->flush)
 518                 goto normal;
 519
 520         if (unlikely(gro_list->count >= MAX_GRO_SKBS))
 521                 gro_flush_oldest(napi, &gro_list->list);
 522         else
 523                 gro_list->count++;
 524
 525         /* Must be called before setting NAPI_GRO_CB(skb)->{age|last} */
 526         gro_try_pull_from_frag0(skb);
 527         NAPI_GRO_CB(skb)->age = jiffies;
 528         NAPI_GRO_CB(skb)->last = skb;
 529         if (!skb_is_gso(skb))
 530                 skb_shinfo(skb)->gso_size = skb_gro_len(skb);
 531         list_add(&skb->list, &gro_list->list);
 532         ret = GRO_HELD;
 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         gro_try_pull_from_frag0(skb);
 546         goto ok;
 547 }
 548
 549 struct packet_offload *gro_find_receive_by_type(__be16 type)
 550 {
 551         struct list_head *offload_head = &net_hotdata.offload_base;
 552         struct packet_offload *ptype;
 553
 554         list_for_each_entry_rcu(ptype, offload_head, list) {
 555                 if (ptype->type != type || !ptype->callbacks.gro_receive)
 556                         continue;
 557                 return ptype;
 558         }
 559         return NULL;
 560 }
 561 EXPORT_SYMBOL(gro_find_receive_by_type);
 562
 563 struct packet_offload *gro_find_complete_by_type(__be16 type)
 564 {
 565         struct list_head *offload_head = &net_hotdata.offload_base;
 566         struct packet_offload *ptype;
 567
 568         list_for_each_entry_rcu(ptype, offload_head, list) {
 569                 if (ptype->type != type || !ptype->callbacks.gro_complete)
 570                         continue;
 571                 return ptype;
 572         }
 573         return NULL;
 574 }
 575 EXPORT_SYMBOL(gro_find_complete_by_type);
 576
 577 static gro_result_t napi_skb_finish(struct napi_struct *napi,
 578                                     struct sk_buff *skb,
 579                                     gro_result_t ret)
 580 {
 581         switch (ret) {
 582         case GRO_NORMAL:
 583                 gro_normal_one(napi, skb, 1);
 584                 break;
 585
 586         case GRO_MERGED_FREE:
 587                 if (NAPI_GRO_CB(skb)->free == NAPI_GRO_FREE_STOLEN_HEAD)
 588                         napi_skb_free_stolen_head(skb);
 589                 else if (skb->fclone != SKB_FCLONE_UNAVAILABLE)
 590                         __kfree_skb(skb);
 591                 else
 592                         __napi_kfree_skb(skb, SKB_CONSUMED);
 593                 break;
 594
 595         case GRO_HELD:
 596         case GRO_MERGED:
 597         case GRO_CONSUMED:
 598                 break;
 599         }
 600
 601         return ret;
 602 }
 603
 604 gro_result_t napi_gro_receive(struct napi_struct *napi, struct sk_buff *skb)
 605 {
 606         gro_result_t ret;
 607
 608         skb_mark_napi_id(skb, napi);
 609         trace_napi_gro_receive_entry(skb);
 610
 611         skb_gro_reset_offset(skb, 0);
 612
 613         ret = napi_skb_finish(napi, skb, dev_gro_receive(napi, skb));
 614         trace_napi_gro_receive_exit(ret);
 615
 616         return ret;
 617 }
 618 EXPORT_SYMBOL(napi_gro_receive);
 619
 620 static void napi_reuse_skb(struct napi_struct *napi, struct sk_buff *skb)
 621 {
 622         if (unlikely(skb->pfmemalloc)) {
 623                 consume_skb(skb);
 624                 return;
 625         }
 626         __skb_pull(skb, skb_headlen(skb));
 627         /* restore the reserve we had after netdev_alloc_skb_ip_align() */
 628         skb_reserve(skb, NET_SKB_PAD + NET_IP_ALIGN - skb_headroom(skb));
 629         __vlan_hwaccel_clear_tag(skb);
 630         skb->dev = napi->dev;
 631         skb->skb_iif = 0;
 632
 633         /* eth_type_trans() assumes pkt_type is PACKET_HOST */
 634         skb->pkt_type = PACKET_HOST;
 635
 636         skb->encapsulation = 0;
 637         skb_shinfo(skb)->gso_type = 0;
 638         skb_shinfo(skb)->gso_size = 0;
 639         if (unlikely(skb->slow_gro)) {
 640                 skb_orphan(skb);
 641                 skb_ext_reset(skb);
 642                 nf_reset_ct(skb);
 643                 skb->slow_gro = 0;
 644         }
 645
 646         napi->skb = skb;
 647 }
 648
 649 struct sk_buff *napi_get_frags(struct napi_struct *napi)
 650 {
 651         struct sk_buff *skb = napi->skb;
 652
 653         if (!skb) {
 654                 skb = napi_alloc_skb(napi, GRO_MAX_HEAD);
 655                 if (skb) {
 656                         napi->skb = skb;
 657                         skb_mark_napi_id(skb, napi);
 658                 }
 659         }
 660         return skb;
 661 }
 662 EXPORT_SYMBOL(napi_get_frags);
 663
 664 static gro_result_t napi_frags_finish(struct napi_struct *napi,
 665                                       struct sk_buff *skb,
 666                                       gro_result_t ret)
 667 {
 668         switch (ret) {
 669         case GRO_NORMAL:
 670         case GRO_HELD:
 671                 __skb_push(skb, ETH_HLEN);
 672                 skb->protocol = eth_type_trans(skb, skb->dev);
 673                 if (ret == GRO_NORMAL)
 674                         gro_normal_one(napi, skb, 1);
 675                 break;
 676
 677         case GRO_MERGED_FREE:
 678                 if (NAPI_GRO_CB(skb)->free == NAPI_GRO_FREE_STOLEN_HEAD)
 679                         napi_skb_free_stolen_head(skb);
 680                 else
 681                         napi_reuse_skb(napi, skb);
 682                 break;
 683
 684         case GRO_MERGED:
 685         case GRO_CONSUMED:
 686                 break;
 687         }
 688
 689         return ret;
 690 }
 691
 692 /* Upper GRO stack assumes network header starts at gro_offset=0
 693  * Drivers could call both napi_gro_frags() and napi_gro_receive()
 694  * We copy ethernet header into skb->data to have a common layout.
 695  */
 696 static struct sk_buff *napi_frags_skb(struct napi_struct *napi)
 697 {
 698         struct sk_buff *skb = napi->skb;
 699         const struct ethhdr *eth;
 700         unsigned int hlen = sizeof(*eth);
 701
 702         napi->skb = NULL;
 703
 704         skb_reset_mac_header(skb);
 705         skb_gro_reset_offset(skb, hlen);
 706
 707         if (unlikely(!skb_gro_may_pull(skb, hlen))) {
 708                 eth = skb_gro_header_slow(skb, hlen, 0);
 709                 if (unlikely(!eth)) {
 710                         net_warn_ratelimited("%s: dropping impossible skb from %s\n",
 711                                              __func__, napi->dev->name);
 712                         napi_reuse_skb(napi, skb);
 713                         return NULL;
 714                 }
 715         } else {
 716                 eth = (const struct ethhdr *)skb->data;
 717
 718                 if (NAPI_GRO_CB(skb)->frag0 != skb->data)
 719                         gro_pull_from_frag0(skb, hlen);
 720
 721                 NAPI_GRO_CB(skb)->frag0 += hlen;
 722                 NAPI_GRO_CB(skb)->frag0_len -= hlen;
 723         }
 724         __skb_pull(skb, hlen);
 725
 726         /*
 727          * This works because the only protocols we care about don't require
 728          * special handling.
 729          * We'll fix it up properly in napi_frags_finish()
 730          */
 731         skb->protocol = eth->h_proto;
 732
 733         return skb;
 734 }
 735
 736 gro_result_t napi_gro_frags(struct napi_struct *napi)
 737 {
 738         gro_result_t ret;
 739         struct sk_buff *skb = napi_frags_skb(napi);
 740
 741         trace_napi_gro_frags_entry(skb);
 742
 743         ret = napi_frags_finish(napi, skb, dev_gro_receive(napi, skb));
 744         trace_napi_gro_frags_exit(ret);
 745
 746         return ret;
 747 }
 748 EXPORT_SYMBOL(napi_gro_frags);
 749
 750 /* Compute the checksum from gro_offset and return the folded value
 751  * after adding in any pseudo checksum.
 752  */
 753 __sum16 __skb_gro_checksum_complete(struct sk_buff *skb)
 754 {
 755         __wsum wsum;
 756         __sum16 sum;
 757
 758         wsum = skb_checksum(skb, skb_gro_offset(skb), skb_gro_len(skb), 0);
 759
 760         /* NAPI_GRO_CB(skb)->csum holds pseudo checksum */
 761         sum = csum_fold(csum_add(NAPI_GRO_CB(skb)->csum, wsum));
 762         /* See comments in __skb_checksum_complete(). */
 763         if (likely(!sum)) {
 764                 if (unlikely(skb->ip_summed == CHECKSUM_COMPLETE) &&
 765                     !skb->csum_complete_sw)
 766                         netdev_rx_csum_fault(skb->dev, skb);
 767         }
 768
 769         NAPI_GRO_CB(skb)->csum = wsum;
 770         NAPI_GRO_CB(skb)->csum_valid = 1;
 771
 772         return sum;
 773 }
 774 EXPORT_SYMBOL(__skb_gro_checksum_complete);