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 #include <linux/skbuff_ref.h> 7 8 #define MAX_GRO_SKBS 8 9 10 /* This should be increased if a protocol with a bigger head is added. */ 11 #define GRO_MAX_HEAD (MAX_HEADER + 128) 12 13 static DEFINE_SPINLOCK(offload_lock); 14 15 /** 16 * dev_add_offload - register offload handlers 17 * @po: protocol offload declaration 18 * 19 * Add protocol offload handlers to the networking stack. The passed 20 * &proto_offload is linked into kernel lists and may not be freed until 21 * it has been removed from the kernel lists. 22 * 23 * This call does not sleep therefore it can not 24 * guarantee all CPU's that are in middle of receiving packets 25 * will see the new offload handlers (until the next received packet). 26 */ 27 void dev_add_offload(struct packet_offload *po) 28 { 29 struct packet_offload *elem; 30 31 spin_lock(&offload_lock); 32 list_for_each_entry(elem, &net_hotdata.offload_base, list) { 33 if (po->priority < elem->priority) 34 break; 35 } 36 list_add_rcu(&po->list, elem->list.prev); 37 spin_unlock(&offload_lock); 38 } 39 EXPORT_SYMBOL(dev_add_offload); 40 41 /** 42 * __dev_remove_offload - remove offload handler 43 * @po: packet offload declaration 44 * 45 * Remove a protocol offload handler that was previously added to the 46 * kernel offload handlers by dev_add_offload(). The passed &offload_type 47 * is removed from the kernel lists and can be freed or reused once this 48 * function returns. 49 * 50 * The packet type might still be in use by receivers 51 * and must not be freed until after all the CPU's have gone 52 * through a quiescent state. 53 */ 54 static void __dev_remove_offload(struct packet_offload *po) 55 { 56 struct list_head *head = &net_hotdata.offload_base; 57 struct packet_offload *po1; 58 59 spin_lock(&offload_lock); 60 61 list_for_each_entry(po1, head, list) { 62 if (po == po1) { 63 list_del_rcu(&po->list); 64 goto out; 65 } 66 } 67 68 pr_warn("dev_remove_offload: %p not found\n", po); 69 out: 70 spin_unlock(&offload_lock); 71 } 72 73 /** 74 * dev_remove_offload - remove packet offload handler 75 * @po: packet offload declaration 76 * 77 * Remove a packet offload handler that was previously added to the kernel 78 * offload handlers by dev_add_offload(). The passed &offload_type is 79 * removed from the kernel lists and can be freed or reused once this 80 * function returns. 81 * 82 * This call sleeps to guarantee that no CPU is looking at the packet 83 * type after return. 84 */ 85 void dev_remove_offload(struct packet_offload *po) 86 { 87 __dev_remove_offload(po); 88 89 synchronize_net(); 90 } 91 EXPORT_SYMBOL(dev_remove_offload); 92 93 94 int skb_gro_receive(struct sk_buff *p, struct sk_buff *skb) 95 { 96 struct skb_shared_info *pinfo, *skbinfo = skb_shinfo(skb); 97 unsigned int offset = skb_gro_offset(skb); 98 unsigned int headlen = skb_headlen(skb); 99 unsigned int len = skb_gro_len(skb); 100 unsigned int delta_truesize; 101 unsigned int gro_max_size; 102 unsigned int new_truesize; 103 struct sk_buff *lp; 104 int segs; 105 106 /* Do not splice page pool based packets w/ non-page pool 107 * packets. This can result in reference count issues as page 108 * pool pages will not decrement the reference count and will 109 * instead be immediately returned to the pool or have frag 110 * count decremented. 111 */ 112 if (p->pp_recycle != skb->pp_recycle) 113 return -ETOOMANYREFS; 114 115 /* pairs with WRITE_ONCE() in netif_set_gro(_ipv4)_max_size() */ 116 gro_max_size = p->protocol == htons(ETH_P_IPV6) ? 117 READ_ONCE(p->dev->gro_max_size) : 118 READ_ONCE(p->dev->gro_ipv4_max_size); 119 120 if (unlikely(p->len + len >= gro_max_size || NAPI_GRO_CB(skb)->flush)) 121 return -E2BIG; 122 123 if (unlikely(p->len + len >= GRO_LEGACY_MAX_SIZE)) { 124 if (NAPI_GRO_CB(skb)->proto != IPPROTO_TCP || 125 (p->protocol == htons(ETH_P_IPV6) && 126 skb_headroom(p) < sizeof(struct hop_jumbo_hdr)) || 127 p->encapsulation) 128 return -E2BIG; 129 } 130 131 segs = NAPI_GRO_CB(skb)->count; 132 lp = NAPI_GRO_CB(p)->last; 133 pinfo = skb_shinfo(lp); 134 135 if (headlen <= offset) { 136 skb_frag_t *frag; 137 skb_frag_t *frag2; 138 int i = skbinfo->nr_frags; 139 int nr_frags = pinfo->nr_frags + i; 140 141 if (nr_frags > MAX_SKB_FRAGS) 142 goto merge; 143 144 offset -= headlen; 145 pinfo->nr_frags = nr_frags; 146 skbinfo->nr_frags = 0; 147 148 frag = pinfo->frags + nr_frags; 149 frag2 = skbinfo->frags + i; 150 do { 151 *--frag = *--frag2; 152 } while (--i); 153 154 skb_frag_off_add(frag, offset); 155 skb_frag_size_sub(frag, offset); 156 157 /* all fragments truesize : remove (head size + sk_buff) */ 158 new_truesize = SKB_TRUESIZE(skb_end_offset(skb)); 159 delta_truesize = skb->truesize - new_truesize; 160 161 skb->truesize = new_truesize; 162 skb->len -= skb->data_len; 163 skb->data_len = 0; 164 165 NAPI_GRO_CB(skb)->free = NAPI_GRO_FREE; 166 goto done; 167 } else if (skb->head_frag) { 168 int nr_frags = pinfo->nr_frags; 169 skb_frag_t *frag = pinfo->frags + nr_frags; 170 struct page *page = virt_to_head_page(skb->head); 171 unsigned int first_size = headlen - offset; 172 unsigned int first_offset; 173 174 if (nr_frags + 1 + skbinfo->nr_frags > MAX_SKB_FRAGS) 175 goto merge; 176 177 first_offset = skb->data - 178 (unsigned char *)page_address(page) + 179 offset; 180 181 pinfo->nr_frags = nr_frags + 1 + skbinfo->nr_frags; 182 183 skb_frag_fill_page_desc(frag, page, first_offset, first_size); 184 185 memcpy(frag + 1, skbinfo->frags, sizeof(*frag) * skbinfo->nr_frags); 186 /* We dont need to clear skbinfo->nr_frags here */ 187 188 new_truesize = SKB_DATA_ALIGN(sizeof(struct sk_buff)); 189 delta_truesize = skb->truesize - new_truesize; 190 skb->truesize = new_truesize; 191 NAPI_GRO_CB(skb)->free = NAPI_GRO_FREE_STOLEN_HEAD; 192 goto done; 193 } 194 195 merge: 196 /* sk ownership - if any - completely transferred to the aggregated packet */ 197 skb->destructor = NULL; 198 skb->sk = NULL; 199 delta_truesize = skb->truesize; 200 if (offset > headlen) { 201 unsigned int eat = offset - headlen; 202 203 skb_frag_off_add(&skbinfo->frags[0], eat); 204 skb_frag_size_sub(&skbinfo->frags[0], eat); 205 skb->data_len -= eat; 206 skb->len -= eat; 207 offset = headlen; 208 } 209 210 __skb_pull(skb, offset); 211 212 if (NAPI_GRO_CB(p)->last == p) 213 skb_shinfo(p)->frag_list = skb; 214 else 215 NAPI_GRO_CB(p)->last->next = skb; 216 NAPI_GRO_CB(p)->last = skb; 217 __skb_header_release(skb); 218 lp = p; 219 220 done: 221 NAPI_GRO_CB(p)->count += segs; 222 p->data_len += len; 223 p->truesize += delta_truesize; 224 p->len += len; 225 if (lp != p) { 226 lp->data_len += len; 227 lp->truesize += delta_truesize; 228 lp->len += len; 229 } 230 NAPI_GRO_CB(skb)->same_flow = 1; 231 return 0; 232 } 233 234 235 static void napi_gro_complete(struct napi_struct *napi, struct sk_buff *skb) 236 { 237 struct list_head *head = &net_hotdata.offload_base; 238 struct packet_offload *ptype; 239 __be16 type = skb->protocol; 240 int err = -ENOENT; 241 242 BUILD_BUG_ON(sizeof(struct napi_gro_cb) > sizeof(skb->cb)); 243 244 if (NAPI_GRO_CB(skb)->count == 1) { 245 skb_shinfo(skb)->gso_size = 0; 246 goto out; 247 } 248 249 rcu_read_lock(); 250 list_for_each_entry_rcu(ptype, head, list) { 251 if (ptype->type != type || !ptype->callbacks.gro_complete) 252 continue; 253 254 err = INDIRECT_CALL_INET(ptype->callbacks.gro_complete, 255 ipv6_gro_complete, inet_gro_complete, 256 skb, 0); 257 break; 258 } 259 rcu_read_unlock(); 260 261 if (err) { 262 WARN_ON(&ptype->list == head); 263 kfree_skb(skb); 264 return; 265 } 266 267 out: 268 gro_normal_one(napi, skb, NAPI_GRO_CB(skb)->count); 269 } 270 271 static void __napi_gro_flush_chain(struct napi_struct *napi, u32 index, 272 bool flush_old) 273 { 274 struct list_head *head = &napi->gro_hash[index].list; 275 struct sk_buff *skb, *p; 276 277 list_for_each_entry_safe_reverse(skb, p, head, list) { 278 if (flush_old && NAPI_GRO_CB(skb)->age == jiffies) 279 return; 280 skb_list_del_init(skb); 281 napi_gro_complete(napi, skb); 282 napi->gro_hash[index].count--; 283 } 284 285 if (!napi->gro_hash[index].count) 286 __clear_bit(index, &napi->gro_bitmask); 287 } 288 289 /* napi->gro_hash[].list contains packets ordered by age. 290 * youngest packets at the head of it. 291 * Complete skbs in reverse order to reduce latencies. 292 */ 293 void napi_gro_flush(struct napi_struct *napi, bool flush_old) 294 { 295 unsigned long bitmask = napi->gro_bitmask; 296 unsigned int i, base = ~0U; 297 298 while ((i = ffs(bitmask)) != 0) { 299 bitmask >>= i; 300 base += i; 301 __napi_gro_flush_chain(napi, base, flush_old); 302 } 303 } 304 EXPORT_SYMBOL(napi_gro_flush); 305 306 static unsigned long gro_list_prepare_tc_ext(const struct sk_buff *skb, 307 const struct sk_buff *p, 308 unsigned long diffs) 309 { 310 #if IS_ENABLED(CONFIG_NET_TC_SKB_EXT) 311 struct tc_skb_ext *skb_ext; 312 struct tc_skb_ext *p_ext; 313 314 skb_ext = skb_ext_find(skb, TC_SKB_EXT); 315 p_ext = skb_ext_find(p, TC_SKB_EXT); 316 317 diffs |= (!!p_ext) ^ (!!skb_ext); 318 if (!diffs && unlikely(skb_ext)) 319 diffs |= p_ext->chain ^ skb_ext->chain; 320 #endif 321 return diffs; 322 } 323 324 static void gro_list_prepare(const struct list_head *head, 325 const struct sk_buff *skb) 326 { 327 unsigned int maclen = skb->dev->hard_header_len; 328 u32 hash = skb_get_hash_raw(skb); 329 struct sk_buff *p; 330 331 list_for_each_entry(p, head, list) { 332 unsigned long diffs; 333 334 NAPI_GRO_CB(p)->flush = 0; 335 336 if (hash != skb_get_hash_raw(p)) { 337 NAPI_GRO_CB(p)->same_flow = 0; 338 continue; 339 } 340 341 diffs = (unsigned long)p->dev ^ (unsigned long)skb->dev; 342 diffs |= p->vlan_all ^ skb->vlan_all; 343 diffs |= skb_metadata_differs(p, skb); 344 if (maclen == ETH_HLEN) 345 diffs |= compare_ether_header(skb_mac_header(p), 346 skb_mac_header(skb)); 347 else if (!diffs) 348 diffs = memcmp(skb_mac_header(p), 349 skb_mac_header(skb), 350 maclen); 351 352 /* in most common scenarions 'slow_gro' is 0 353 * otherwise we are already on some slower paths 354 * either skip all the infrequent tests altogether or 355 * avoid trying too hard to skip each of them individually 356 */ 357 if (!diffs && unlikely(skb->slow_gro | p->slow_gro)) { 358 diffs |= p->sk != skb->sk; 359 diffs |= skb_metadata_dst_cmp(p, skb); 360 diffs |= skb_get_nfct(p) ^ skb_get_nfct(skb); 361 362 diffs |= gro_list_prepare_tc_ext(skb, p, diffs); 363 } 364 365 NAPI_GRO_CB(p)->same_flow = !diffs; 366 } 367 } 368 369 static inline void skb_gro_reset_offset(struct sk_buff *skb, u32 nhoff) 370 { 371 const struct skb_shared_info *pinfo; 372 const skb_frag_t *frag0; 373 unsigned int headlen; 374 375 NAPI_GRO_CB(skb)->network_offset = 0; 376 NAPI_GRO_CB(skb)->data_offset = 0; 377 headlen = skb_headlen(skb); 378 NAPI_GRO_CB(skb)->frag0 = skb->data; 379 NAPI_GRO_CB(skb)->frag0_len = headlen; 380 if (headlen) 381 return; 382 383 pinfo = skb_shinfo(skb); 384 frag0 = &pinfo->frags[0]; 385 386 if (pinfo->nr_frags && !PageHighMem(skb_frag_page(frag0)) && 387 (!NET_IP_ALIGN || !((skb_frag_off(frag0) + nhoff) & 3))) { 388 NAPI_GRO_CB(skb)->frag0 = skb_frag_address(frag0); 389 NAPI_GRO_CB(skb)->frag0_len = min_t(unsigned int, 390 skb_frag_size(frag0), 391 skb->end - skb->tail); 392 } 393 } 394 395 static void gro_pull_from_frag0(struct sk_buff *skb, int grow) 396 { 397 struct skb_shared_info *pinfo = skb_shinfo(skb); 398 399 BUG_ON(skb->end - skb->tail < grow); 400 401 memcpy(skb_tail_pointer(skb), NAPI_GRO_CB(skb)->frag0, grow); 402 403 skb->data_len -= grow; 404 skb->tail += grow; 405 406 skb_frag_off_add(&pinfo->frags[0], grow); 407 skb_frag_size_sub(&pinfo->frags[0], grow); 408 409 if (unlikely(!skb_frag_size(&pinfo->frags[0]))) { 410 skb_frag_unref(skb, 0); 411 memmove(pinfo->frags, pinfo->frags + 1, 412 --pinfo->nr_frags * sizeof(pinfo->frags[0])); 413 } 414 } 415 416 static void gro_try_pull_from_frag0(struct sk_buff *skb) 417 { 418 int grow = skb_gro_offset(skb) - skb_headlen(skb); 419 420 if (grow > 0) 421 gro_pull_from_frag0(skb, grow); 422 } 423 424 static void gro_flush_oldest(struct napi_struct *napi, struct list_head *head) 425 { 426 struct sk_buff *oldest; 427 428 oldest = list_last_entry(head, struct sk_buff, list); 429 430 /* We are called with head length >= MAX_GRO_SKBS, so this is 431 * impossible. 432 */ 433 if (WARN_ON_ONCE(!oldest)) 434 return; 435 436 /* Do not adjust napi->gro_hash[].count, caller is adding a new 437 * SKB to the chain. 438 */ 439 skb_list_del_init(oldest); 440 napi_gro_complete(napi, oldest); 441 } 442 443 static enum gro_result dev_gro_receive(struct napi_struct *napi, struct sk_buff *skb) 444 { 445 u32 bucket = skb_get_hash_raw(skb) & (GRO_HASH_BUCKETS - 1); 446 struct gro_list *gro_list = &napi->gro_hash[bucket]; 447 struct list_head *head = &net_hotdata.offload_base; 448 struct packet_offload *ptype; 449 __be16 type = skb->protocol; 450 struct sk_buff *pp = NULL; 451 enum gro_result ret; 452 int same_flow; 453 454 if (netif_elide_gro(skb->dev)) 455 goto normal; 456 457 gro_list_prepare(&gro_list->list, skb); 458 459 rcu_read_lock(); 460 list_for_each_entry_rcu(ptype, head, list) { 461 if (ptype->type == type && ptype->callbacks.gro_receive) 462 goto found_ptype; 463 } 464 rcu_read_unlock(); 465 goto normal; 466 467 found_ptype: 468 skb_set_network_header(skb, skb_gro_offset(skb)); 469 skb_reset_mac_len(skb); 470 BUILD_BUG_ON(sizeof_field(struct napi_gro_cb, zeroed) != sizeof(u32)); 471 BUILD_BUG_ON(!IS_ALIGNED(offsetof(struct napi_gro_cb, zeroed), 472 sizeof(u32))); /* Avoid slow unaligned acc */ 473 *(u32 *)&NAPI_GRO_CB(skb)->zeroed = 0; 474 NAPI_GRO_CB(skb)->flush = skb_has_frag_list(skb); 475 NAPI_GRO_CB(skb)->is_atomic = 1; 476 NAPI_GRO_CB(skb)->count = 1; 477 if (unlikely(skb_is_gso(skb))) { 478 NAPI_GRO_CB(skb)->count = skb_shinfo(skb)->gso_segs; 479 /* Only support TCP and non DODGY users. */ 480 if (!skb_is_gso_tcp(skb) || 481 (skb_shinfo(skb)->gso_type & SKB_GSO_DODGY)) 482 NAPI_GRO_CB(skb)->flush = 1; 483 } 484 485 /* Setup for GRO checksum validation */ 486 switch (skb->ip_summed) { 487 case CHECKSUM_COMPLETE: 488 NAPI_GRO_CB(skb)->csum = skb->csum; 489 NAPI_GRO_CB(skb)->csum_valid = 1; 490 break; 491 case CHECKSUM_UNNECESSARY: 492 NAPI_GRO_CB(skb)->csum_cnt = skb->csum_level + 1; 493 break; 494 } 495 496 pp = INDIRECT_CALL_INET(ptype->callbacks.gro_receive, 497 ipv6_gro_receive, inet_gro_receive, 498 &gro_list->list, skb); 499 500 rcu_read_unlock(); 501 502 if (PTR_ERR(pp) == -EINPROGRESS) { 503 ret = GRO_CONSUMED; 504 goto ok; 505 } 506 507 same_flow = NAPI_GRO_CB(skb)->same_flow; 508 ret = NAPI_GRO_CB(skb)->free ? GRO_MERGED_FREE : GRO_MERGED; 509 510 if (pp) { 511 skb_list_del_init(pp); 512 napi_gro_complete(napi, pp); 513 gro_list->count--; 514 } 515 516 if (same_flow) 517 goto ok; 518 519 if (NAPI_GRO_CB(skb)->flush) 520 goto normal; 521 522 if (unlikely(gro_list->count >= MAX_GRO_SKBS)) 523 gro_flush_oldest(napi, &gro_list->list); 524 else 525 gro_list->count++; 526 527 /* Must be called before setting NAPI_GRO_CB(skb)->{age|last} */ 528 gro_try_pull_from_frag0(skb); 529 NAPI_GRO_CB(skb)->age = jiffies; 530 NAPI_GRO_CB(skb)->last = skb; 531 if (!skb_is_gso(skb)) 532 skb_shinfo(skb)->gso_size = skb_gro_len(skb); 533 list_add(&skb->list, &gro_list->list); 534 ret = GRO_HELD; 535 ok: 536 if (gro_list->count) { 537 if (!test_bit(bucket, &napi->gro_bitmask)) 538 __set_bit(bucket, &napi->gro_bitmask); 539 } else if (test_bit(bucket, &napi->gro_bitmask)) { 540 __clear_bit(bucket, &napi->gro_bitmask); 541 } 542 543 return ret; 544 545 normal: 546 ret = GRO_NORMAL; 547 gro_try_pull_from_frag0(skb); 548 goto ok; 549 } 550 551 struct packet_offload *gro_find_receive_by_type(__be16 type) 552 { 553 struct list_head *offload_head = &net_hotdata.offload_base; 554 struct packet_offload *ptype; 555 556 list_for_each_entry_rcu(ptype, offload_head, list) { 557 if (ptype->type != type || !ptype->callbacks.gro_receive) 558 continue; 559 return ptype; 560 } 561 return NULL; 562 } 563 EXPORT_SYMBOL(gro_find_receive_by_type); 564 565 struct packet_offload *gro_find_complete_by_type(__be16 type) 566 { 567 struct list_head *offload_head = &net_hotdata.offload_base; 568 struct packet_offload *ptype; 569 570 list_for_each_entry_rcu(ptype, offload_head, list) { 571 if (ptype->type != type || !ptype->callbacks.gro_complete) 572 continue; 573 return ptype; 574 } 575 return NULL; 576 } 577 EXPORT_SYMBOL(gro_find_complete_by_type); 578 579 static gro_result_t napi_skb_finish(struct napi_struct *napi, 580 struct sk_buff *skb, 581 gro_result_t ret) 582 { 583 switch (ret) { 584 case GRO_NORMAL: 585 gro_normal_one(napi, skb, 1); 586 break; 587 588 case GRO_MERGED_FREE: 589 if (NAPI_GRO_CB(skb)->free == NAPI_GRO_FREE_STOLEN_HEAD) 590 napi_skb_free_stolen_head(skb); 591 else if (skb->fclone != SKB_FCLONE_UNAVAILABLE) 592 __kfree_skb(skb); 593 else 594 __napi_kfree_skb(skb, SKB_CONSUMED); 595 break; 596 597 case GRO_HELD: 598 case GRO_MERGED: 599 case GRO_CONSUMED: 600 break; 601 } 602 603 return ret; 604 } 605 606 gro_result_t napi_gro_receive(struct napi_struct *napi, struct sk_buff *skb) 607 { 608 gro_result_t ret; 609 610 skb_mark_napi_id(skb, napi); 611 trace_napi_gro_receive_entry(skb); 612 613 skb_gro_reset_offset(skb, 0); 614 615 ret = napi_skb_finish(napi, skb, dev_gro_receive(napi, skb)); 616 trace_napi_gro_receive_exit(ret); 617 618 return ret; 619 } 620 EXPORT_SYMBOL(napi_gro_receive); 621 622 static void napi_reuse_skb(struct napi_struct *napi, struct sk_buff *skb) 623 { 624 if (unlikely(skb->pfmemalloc)) { 625 consume_skb(skb); 626 return; 627 } 628 __skb_pull(skb, skb_headlen(skb)); 629 /* restore the reserve we had after netdev_alloc_skb_ip_align() */ 630 skb_reserve(skb, NET_SKB_PAD + NET_IP_ALIGN - skb_headroom(skb)); 631 __vlan_hwaccel_clear_tag(skb); 632 skb->dev = napi->dev; 633 skb->skb_iif = 0; 634 635 /* eth_type_trans() assumes pkt_type is PACKET_HOST */ 636 skb->pkt_type = PACKET_HOST; 637 638 skb->encapsulation = 0; 639 skb_shinfo(skb)->gso_type = 0; 640 skb_shinfo(skb)->gso_size = 0; 641 if (unlikely(skb->slow_gro)) { 642 skb_orphan(skb); 643 skb_ext_reset(skb); 644 nf_reset_ct(skb); 645 skb->slow_gro = 0; 646 } 647 648 napi->skb = skb; 649 } 650 651 struct sk_buff *napi_get_frags(struct napi_struct *napi) 652 { 653 struct sk_buff *skb = napi->skb; 654 655 if (!skb) { 656 skb = napi_alloc_skb(napi, GRO_MAX_HEAD); 657 if (skb) { 658 napi->skb = skb; 659 skb_mark_napi_id(skb, napi); 660 } 661 } 662 return skb; 663 } 664 EXPORT_SYMBOL(napi_get_frags); 665 666 static gro_result_t napi_frags_finish(struct napi_struct *napi, 667 struct sk_buff *skb, 668 gro_result_t ret) 669 { 670 switch (ret) { 671 case GRO_NORMAL: 672 case GRO_HELD: 673 __skb_push(skb, ETH_HLEN); 674 skb->protocol = eth_type_trans(skb, skb->dev); 675 if (ret == GRO_NORMAL) 676 gro_normal_one(napi, skb, 1); 677 break; 678 679 case GRO_MERGED_FREE: 680 if (NAPI_GRO_CB(skb)->free == NAPI_GRO_FREE_STOLEN_HEAD) 681 napi_skb_free_stolen_head(skb); 682 else 683 napi_reuse_skb(napi, skb); 684 break; 685 686 case GRO_MERGED: 687 case GRO_CONSUMED: 688 break; 689 } 690 691 return ret; 692 } 693 694 /* Upper GRO stack assumes network header starts at gro_offset=0 695 * Drivers could call both napi_gro_frags() and napi_gro_receive() 696 * We copy ethernet header into skb->data to have a common layout. 697 */ 698 static struct sk_buff *napi_frags_skb(struct napi_struct *napi) 699 { 700 struct sk_buff *skb = napi->skb; 701 const struct ethhdr *eth; 702 unsigned int hlen = sizeof(*eth); 703 704 napi->skb = NULL; 705 706 skb_reset_mac_header(skb); 707 skb_gro_reset_offset(skb, hlen); 708 709 if (unlikely(!skb_gro_may_pull(skb, hlen))) { 710 eth = skb_gro_header_slow(skb, hlen, 0); 711 if (unlikely(!eth)) { 712 net_warn_ratelimited("%s: dropping impossible skb from %s\n", 713 __func__, napi->dev->name); 714 napi_reuse_skb(napi, skb); 715 return NULL; 716 } 717 } else { 718 eth = (const struct ethhdr *)skb->data; 719 720 if (NAPI_GRO_CB(skb)->frag0 != skb->data) 721 gro_pull_from_frag0(skb, hlen); 722 723 NAPI_GRO_CB(skb)->frag0 += hlen; 724 NAPI_GRO_CB(skb)->frag0_len -= hlen; 725 } 726 __skb_pull(skb, hlen); 727 728 /* 729 * This works because the only protocols we care about don't require 730 * special handling. 731 * We'll fix it up properly in napi_frags_finish() 732 */ 733 skb->protocol = eth->h_proto; 734 735 return skb; 736 } 737 738 gro_result_t napi_gro_frags(struct napi_struct *napi) 739 { 740 gro_result_t ret; 741 struct sk_buff *skb = napi_frags_skb(napi); 742 743 trace_napi_gro_frags_entry(skb); 744 745 ret = napi_frags_finish(napi, skb, dev_gro_receive(napi, skb)); 746 trace_napi_gro_frags_exit(ret); 747 748 return ret; 749 } 750 EXPORT_SYMBOL(napi_gro_frags); 751 752 /* Compute the checksum from gro_offset and return the folded value 753 * after adding in any pseudo checksum. 754 */ 755 __sum16 __skb_gro_checksum_complete(struct sk_buff *skb) 756 { 757 __wsum wsum; 758 __sum16 sum; 759 760 wsum = skb_checksum(skb, skb_gro_offset(skb), skb_gro_len(skb), 0); 761 762 /* NAPI_GRO_CB(skb)->csum holds pseudo checksum */ 763 sum = csum_fold(csum_add(NAPI_GRO_CB(skb)->csum, wsum)); 764 /* See comments in __skb_checksum_complete(). */ 765 if (likely(!sum)) { 766 if (unlikely(skb->ip_summed == CHECKSUM_COMPLETE) && 767 !skb->csum_complete_sw) 768 netdev_rx_csum_fault(skb->dev, skb); 769 } 770 771 NAPI_GRO_CB(skb)->csum = wsum; 772 NAPI_GRO_CB(skb)->csum_valid = 1; 773 774 return sum; 775 } 776 EXPORT_SYMBOL(__skb_gro_checksum_complete); 777