1 // SPDX-License-Identifier: GPL-2.0-or-later 2 #include <net/psp.h> 3 #include <net/gro.h> 4 #include <net/dst_metadata.h> 5 #include <net/busy_poll.h> 6 #include <trace/events/net.h> 7 #include <linux/skbuff_ref.h> 8 9 #define MAX_GRO_SKBS 8 10 11 static DEFINE_SPINLOCK(offload_lock); 12 13 /** 14 * dev_add_offload - register offload handlers 15 * @po: protocol offload declaration 16 * 17 * Add protocol offload handlers to the networking stack. The passed 18 * &proto_offload is linked into kernel lists and may not be freed until 19 * it has been removed from the kernel lists. 20 * 21 * This call does not sleep therefore it can not 22 * guarantee all CPU's that are in middle of receiving packets 23 * will see the new offload handlers (until the next received packet). 24 */ 25 void dev_add_offload(struct packet_offload *po) 26 { 27 struct packet_offload *elem; 28 29 spin_lock(&offload_lock); 30 list_for_each_entry(elem, &net_hotdata.offload_base, list) { 31 if (po->priority < elem->priority) 32 break; 33 } 34 list_add_rcu(&po->list, elem->list.prev); 35 spin_unlock(&offload_lock); 36 } 37 EXPORT_SYMBOL(dev_add_offload); 38 39 /** 40 * __dev_remove_offload - remove offload handler 41 * @po: packet offload declaration 42 * 43 * Remove a protocol offload handler that was previously added to the 44 * kernel offload handlers by dev_add_offload(). The passed &offload_type 45 * is removed from the kernel lists and can be freed or reused once this 46 * function returns. 47 * 48 * The packet type might still be in use by receivers 49 * and must not be freed until after all the CPU's have gone 50 * through a quiescent state. 51 */ 52 static void __dev_remove_offload(struct packet_offload *po) 53 { 54 struct list_head *head = &net_hotdata.offload_base; 55 struct packet_offload *po1; 56 57 spin_lock(&offload_lock); 58 59 list_for_each_entry(po1, head, list) { 60 if (po == po1) { 61 list_del_rcu(&po->list); 62 goto out; 63 } 64 } 65 66 pr_warn("dev_remove_offload: %p not found\n", po); 67 out: 68 spin_unlock(&offload_lock); 69 } 70 71 /** 72 * dev_remove_offload - remove packet offload handler 73 * @po: packet offload declaration 74 * 75 * Remove a packet offload handler that was previously added to the kernel 76 * offload handlers by dev_add_offload(). The passed &offload_type is 77 * removed from the kernel lists and can be freed or reused once this 78 * function returns. 79 * 80 * This call sleeps to guarantee that no CPU is looking at the packet 81 * type after return. 82 */ 83 void dev_remove_offload(struct packet_offload *po) 84 { 85 __dev_remove_offload(po); 86 87 synchronize_net(); 88 } 89 EXPORT_SYMBOL(dev_remove_offload); 90 91 92 int skb_gro_receive(struct sk_buff *p, struct sk_buff *skb) 93 { 94 struct skb_shared_info *pinfo, *skbinfo = skb_shinfo(skb); 95 unsigned int offset = skb_gro_offset(skb); 96 unsigned int headlen = skb_headlen(skb); 97 unsigned int len = skb_gro_len(skb); 98 unsigned int delta_truesize; 99 unsigned int new_truesize; 100 struct sk_buff *lp; 101 int segs; 102 103 /* Do not splice page pool based packets w/ non-page pool 104 * packets. This can result in reference count issues as page 105 * pool pages will not decrement the reference count and will 106 * instead be immediately returned to the pool or have frag 107 * count decremented. 108 */ 109 if (p->pp_recycle != skb->pp_recycle) 110 return -ETOOMANYREFS; 111 112 if (unlikely(p->len + len >= netif_get_gro_max_size(p->dev, p) || 113 NAPI_GRO_CB(skb)->flush)) 114 return -E2BIG; 115 116 if (unlikely(p->len + len >= GRO_LEGACY_MAX_SIZE)) { 117 if (NAPI_GRO_CB(skb)->proto != IPPROTO_TCP || 118 p->encapsulation) 119 return -E2BIG; 120 } 121 122 segs = NAPI_GRO_CB(skb)->count; 123 lp = NAPI_GRO_CB(p)->last; 124 pinfo = skb_shinfo(lp); 125 126 if (headlen <= offset) { 127 skb_frag_t *frag; 128 skb_frag_t *frag2; 129 int i = skbinfo->nr_frags; 130 int nr_frags = pinfo->nr_frags + i; 131 132 if (nr_frags > MAX_SKB_FRAGS) 133 goto merge; 134 135 offset -= headlen; 136 pinfo->nr_frags = nr_frags; 137 skbinfo->nr_frags = 0; 138 139 frag = pinfo->frags + nr_frags; 140 frag2 = skbinfo->frags + i; 141 do { 142 *--frag = *--frag2; 143 } while (--i); 144 145 skb_frag_off_add(frag, offset); 146 skb_frag_size_sub(frag, offset); 147 148 /* all fragments truesize : remove (head size + sk_buff) */ 149 new_truesize = SKB_TRUESIZE(skb_end_offset(skb)); 150 delta_truesize = skb->truesize - new_truesize; 151 152 skb->truesize = new_truesize; 153 skb->len -= skb->data_len; 154 skb->data_len = 0; 155 156 NAPI_GRO_CB(skb)->free = NAPI_GRO_FREE; 157 goto done; 158 } else if (skb->head_frag) { 159 int nr_frags = pinfo->nr_frags; 160 skb_frag_t *frag = pinfo->frags + nr_frags; 161 struct page *page = virt_to_head_page(skb->head); 162 unsigned int first_size = headlen - offset; 163 unsigned int first_offset; 164 165 if (nr_frags + 1 + skbinfo->nr_frags > MAX_SKB_FRAGS) 166 goto merge; 167 168 first_offset = skb->data - 169 (unsigned char *)page_address(page) + 170 offset; 171 172 pinfo->nr_frags = nr_frags + 1 + skbinfo->nr_frags; 173 174 skb_frag_fill_page_desc(frag, page, first_offset, first_size); 175 176 memcpy(frag + 1, skbinfo->frags, sizeof(*frag) * skbinfo->nr_frags); 177 /* We dont need to clear skbinfo->nr_frags here */ 178 179 new_truesize = SKB_DATA_ALIGN(sizeof(struct sk_buff)); 180 delta_truesize = skb->truesize - new_truesize; 181 skb->truesize = new_truesize; 182 NAPI_GRO_CB(skb)->free = NAPI_GRO_FREE_STOLEN_HEAD; 183 goto done; 184 } 185 186 merge: 187 /* sk ownership - if any - completely transferred to the aggregated packet */ 188 skb->destructor = NULL; 189 skb->sk = NULL; 190 delta_truesize = skb->truesize; 191 if (offset > headlen) { 192 unsigned int eat = offset - headlen; 193 194 skb_frag_off_add(&skbinfo->frags[0], eat); 195 skb_frag_size_sub(&skbinfo->frags[0], eat); 196 skb->data_len -= eat; 197 skb->len -= eat; 198 offset = headlen; 199 } 200 201 __skb_pull(skb, offset); 202 203 if (NAPI_GRO_CB(p)->last == p) 204 skb_shinfo(p)->frag_list = skb; 205 else 206 NAPI_GRO_CB(p)->last->next = skb; 207 NAPI_GRO_CB(p)->last = skb; 208 __skb_header_release(skb); 209 lp = p; 210 211 done: 212 NAPI_GRO_CB(p)->count += segs; 213 p->data_len += len; 214 p->truesize += delta_truesize; 215 p->len += len; 216 skb_shinfo(p)->flags |= skbinfo->flags & SKBFL_SHARED_FRAG; 217 if (lp != p) { 218 lp->data_len += len; 219 lp->truesize += delta_truesize; 220 lp->len += len; 221 skb_shinfo(lp)->flags |= skbinfo->flags & SKBFL_SHARED_FRAG; 222 } 223 NAPI_GRO_CB(skb)->same_flow = 1; 224 return 0; 225 } 226 227 int skb_gro_receive_list(struct sk_buff *p, struct sk_buff *skb) 228 { 229 if (unlikely(p->len + skb->len >= 65536)) 230 return -E2BIG; 231 232 if (NAPI_GRO_CB(p)->last == p) 233 skb_shinfo(p)->frag_list = skb; 234 else 235 NAPI_GRO_CB(p)->last->next = skb; 236 237 skb_pull(skb, skb_gro_offset(skb)); 238 239 NAPI_GRO_CB(p)->last = skb; 240 NAPI_GRO_CB(p)->count++; 241 p->data_len += skb->len; 242 243 /* sk ownership - if any - completely transferred to the aggregated packet */ 244 skb->destructor = NULL; 245 skb->sk = NULL; 246 p->truesize += skb->truesize; 247 p->len += skb->len; 248 249 skb_shinfo(p)->flags |= skb_shinfo(skb)->flags & SKBFL_SHARED_FRAG; 250 251 NAPI_GRO_CB(skb)->same_flow = 1; 252 253 return 0; 254 } 255 256 static void gro_complete(struct gro_node *gro, struct sk_buff *skb) 257 { 258 struct list_head *head = &net_hotdata.offload_base; 259 struct packet_offload *ptype; 260 __be16 type = skb->protocol; 261 int err = -ENOENT; 262 263 BUILD_BUG_ON(sizeof(struct napi_gro_cb) > sizeof(skb->cb)); 264 265 if (NAPI_GRO_CB(skb)->count == 1) { 266 skb_shinfo(skb)->gso_size = 0; 267 goto out; 268 } 269 270 /* NICs can feed encapsulated packets into GRO */ 271 skb->encapsulation = 0; 272 rcu_read_lock(); 273 list_for_each_entry_rcu(ptype, head, list) { 274 if (ptype->type != type || !ptype->callbacks.gro_complete) 275 continue; 276 277 err = INDIRECT_CALL_INET(ptype->callbacks.gro_complete, 278 ipv6_gro_complete, inet_gro_complete, 279 skb, 0); 280 break; 281 } 282 rcu_read_unlock(); 283 284 if (err) { 285 WARN_ON(&ptype->list == head); 286 kfree_skb(skb); 287 return; 288 } 289 290 out: 291 gro_normal_one(gro, skb, NAPI_GRO_CB(skb)->count); 292 } 293 294 static void __gro_flush_chain(struct gro_node *gro, u32 index, bool flush_old) 295 { 296 struct list_head *head = &gro->hash[index].list; 297 struct sk_buff *skb, *p; 298 299 list_for_each_entry_safe_reverse(skb, p, head, list) { 300 if (flush_old && NAPI_GRO_CB(skb)->age == jiffies) 301 return; 302 skb_list_del_init(skb); 303 gro_complete(gro, skb); 304 gro->hash[index].count--; 305 } 306 307 if (!gro->hash[index].count) 308 __clear_bit(index, &gro->bitmask); 309 } 310 311 /* 312 * gro->hash[].list contains packets ordered by age. 313 * youngest packets at the head of it. 314 * Complete skbs in reverse order to reduce latencies. 315 */ 316 void __gro_flush(struct gro_node *gro, bool flush_old) 317 { 318 unsigned long bitmask = gro->bitmask; 319 unsigned int i, base = ~0U; 320 321 while ((i = ffs(bitmask)) != 0) { 322 bitmask >>= i; 323 base += i; 324 __gro_flush_chain(gro, base, flush_old); 325 } 326 } 327 EXPORT_SYMBOL(__gro_flush); 328 329 static unsigned long gro_list_prepare_tc_ext(const struct sk_buff *skb, 330 const struct sk_buff *p, 331 unsigned long diffs) 332 { 333 #if IS_ENABLED(CONFIG_NET_TC_SKB_EXT) 334 struct tc_skb_ext *skb_ext; 335 struct tc_skb_ext *p_ext; 336 337 skb_ext = skb_ext_find(skb, TC_SKB_EXT); 338 p_ext = skb_ext_find(p, TC_SKB_EXT); 339 340 diffs |= (!!p_ext) ^ (!!skb_ext); 341 if (!diffs && unlikely(skb_ext)) 342 diffs |= p_ext->chain ^ skb_ext->chain; 343 #endif 344 return diffs; 345 } 346 347 static void gro_list_prepare(const struct list_head *head, 348 const struct sk_buff *skb) 349 { 350 unsigned int maclen = skb->dev->hard_header_len; 351 u32 hash = skb_get_hash_raw(skb); 352 struct sk_buff *p; 353 354 list_for_each_entry(p, head, list) { 355 unsigned long diffs; 356 357 if (hash != skb_get_hash_raw(p)) { 358 NAPI_GRO_CB(p)->same_flow = 0; 359 continue; 360 } 361 362 diffs = (unsigned long)p->dev ^ (unsigned long)skb->dev; 363 diffs |= p->vlan_all ^ skb->vlan_all; 364 diffs |= skb_metadata_differs(p, skb); 365 if (maclen == ETH_HLEN) 366 diffs |= compare_ether_header(skb_mac_header(p), 367 skb_mac_header(skb)); 368 else if (!diffs) 369 diffs = memcmp(skb_mac_header(p), 370 skb_mac_header(skb), 371 maclen); 372 373 /* in most common scenarios 'slow_gro' is 0 374 * otherwise we are already on some slower paths 375 * either skip all the infrequent tests altogether or 376 * avoid trying too hard to skip each of them individually 377 */ 378 if (!diffs && unlikely(skb->slow_gro | p->slow_gro)) { 379 diffs |= p->sk != skb->sk; 380 diffs |= skb_metadata_dst_cmp(p, skb); 381 diffs |= skb_get_nfct(p) ^ skb_get_nfct(skb); 382 383 diffs |= gro_list_prepare_tc_ext(skb, p, diffs); 384 diffs |= __psp_skb_coalesce_diff(skb, p, diffs); 385 } 386 387 NAPI_GRO_CB(p)->same_flow = !diffs; 388 } 389 } 390 391 static inline void skb_gro_reset_offset(struct sk_buff *skb, u32 nhoff) 392 { 393 const struct skb_shared_info *pinfo; 394 const skb_frag_t *frag0; 395 unsigned int headlen; 396 397 NAPI_GRO_CB(skb)->network_offset = 0; 398 NAPI_GRO_CB(skb)->data_offset = 0; 399 headlen = skb_headlen(skb); 400 NAPI_GRO_CB(skb)->frag0 = skb->data; 401 NAPI_GRO_CB(skb)->frag0_len = headlen; 402 if (headlen) 403 return; 404 405 pinfo = skb_shinfo(skb); 406 frag0 = &pinfo->frags[0]; 407 408 if (pinfo->nr_frags && skb_frag_page(frag0) && 409 !PageHighMem(skb_frag_page(frag0)) && 410 (!NET_IP_ALIGN || !((skb_frag_off(frag0) + nhoff) & 3))) { 411 NAPI_GRO_CB(skb)->frag0 = skb_frag_address(frag0); 412 NAPI_GRO_CB(skb)->frag0_len = min_t(unsigned int, 413 skb_frag_size(frag0), 414 skb->end - skb->tail); 415 } 416 } 417 418 static void gro_pull_from_frag0(struct sk_buff *skb, int grow) 419 { 420 struct skb_shared_info *pinfo = skb_shinfo(skb); 421 422 DEBUG_NET_WARN_ON_ONCE(skb->end - skb->tail < grow); 423 424 memcpy(skb_tail_pointer(skb), NAPI_GRO_CB(skb)->frag0, grow); 425 426 skb->data_len -= grow; 427 skb->tail += grow; 428 429 skb_frag_off_add(&pinfo->frags[0], grow); 430 skb_frag_size_sub(&pinfo->frags[0], grow); 431 432 if (unlikely(!skb_frag_size(&pinfo->frags[0]))) { 433 skb_frag_unref(skb, 0); 434 memmove(pinfo->frags, pinfo->frags + 1, 435 --pinfo->nr_frags * sizeof(pinfo->frags[0])); 436 } 437 } 438 439 static void gro_try_pull_from_frag0(struct sk_buff *skb) 440 { 441 int grow = skb_gro_offset(skb) - skb_headlen(skb); 442 443 if (grow > 0) 444 gro_pull_from_frag0(skb, grow); 445 } 446 447 static void gro_flush_oldest(struct gro_node *gro, struct list_head *head) 448 { 449 struct sk_buff *oldest; 450 451 oldest = list_last_entry(head, struct sk_buff, list); 452 453 /* We are called with head length >= MAX_GRO_SKBS, so this is 454 * impossible. 455 */ 456 if (WARN_ON_ONCE(!oldest)) 457 return; 458 459 /* Do not adjust napi->gro_hash[].count, caller is adding a new 460 * SKB to the chain. 461 */ 462 skb_list_del_init(oldest); 463 gro_complete(gro, oldest); 464 } 465 466 static enum gro_result dev_gro_receive(struct gro_node *gro, 467 struct sk_buff *skb) 468 { 469 u32 bucket = skb_get_hash_raw(skb) & (GRO_HASH_BUCKETS - 1); 470 struct list_head *head = &net_hotdata.offload_base; 471 struct gro_list *gro_list = &gro->hash[bucket]; 472 struct packet_offload *ptype; 473 __be16 type = skb->protocol; 474 struct sk_buff *pp = NULL; 475 enum gro_result ret; 476 int same_flow; 477 478 if (netif_elide_gro(skb->dev)) 479 goto normal; 480 481 gro_list_prepare(&gro_list->list, skb); 482 483 rcu_read_lock(); 484 list_for_each_entry_rcu(ptype, head, list) { 485 if (ptype->type == type && ptype->callbacks.gro_receive) 486 goto found_ptype; 487 } 488 rcu_read_unlock(); 489 goto normal; 490 491 found_ptype: 492 skb_set_network_header(skb, skb_gro_offset(skb)); 493 skb_reset_mac_len(skb); 494 BUILD_BUG_ON(sizeof_field(struct napi_gro_cb, zeroed) != sizeof(u32)); 495 BUILD_BUG_ON(!IS_ALIGNED(offsetof(struct napi_gro_cb, zeroed), 496 sizeof(u32))); /* Avoid slow unaligned acc */ 497 *(u32 *)&NAPI_GRO_CB(skb)->zeroed = 0; 498 NAPI_GRO_CB(skb)->flush = skb_has_frag_list(skb); 499 NAPI_GRO_CB(skb)->count = 1; 500 if (unlikely(skb_is_gso(skb))) { 501 NAPI_GRO_CB(skb)->count = skb_shinfo(skb)->gso_segs; 502 /* Only support TCP and non DODGY users. */ 503 if (!skb_is_gso_tcp(skb) || 504 (skb_shinfo(skb)->gso_type & SKB_GSO_DODGY)) 505 NAPI_GRO_CB(skb)->flush = 1; 506 } 507 508 /* Setup for GRO checksum validation */ 509 switch (skb->ip_summed) { 510 case CHECKSUM_COMPLETE: 511 NAPI_GRO_CB(skb)->csum = skb->csum; 512 NAPI_GRO_CB(skb)->csum_valid = 1; 513 break; 514 case CHECKSUM_UNNECESSARY: 515 NAPI_GRO_CB(skb)->csum_cnt = skb->csum_level + 1; 516 break; 517 } 518 519 pp = INDIRECT_CALL_INET(ptype->callbacks.gro_receive, 520 ipv6_gro_receive, inet_gro_receive, 521 &gro_list->list, skb); 522 523 rcu_read_unlock(); 524 525 if (PTR_ERR(pp) == -EINPROGRESS) { 526 ret = GRO_CONSUMED; 527 goto ok; 528 } 529 530 same_flow = NAPI_GRO_CB(skb)->same_flow; 531 ret = NAPI_GRO_CB(skb)->free ? GRO_MERGED_FREE : GRO_MERGED; 532 533 if (pp) { 534 skb_list_del_init(pp); 535 gro_complete(gro, pp); 536 gro_list->count--; 537 } 538 539 if (same_flow) 540 goto ok; 541 542 if (NAPI_GRO_CB(skb)->flush) 543 goto normal; 544 545 if (unlikely(gro_list->count >= MAX_GRO_SKBS)) 546 gro_flush_oldest(gro, &gro_list->list); 547 else 548 gro_list->count++; 549 550 /* Must be called before setting NAPI_GRO_CB(skb)->{age|last} */ 551 gro_try_pull_from_frag0(skb); 552 NAPI_GRO_CB(skb)->age = jiffies; 553 NAPI_GRO_CB(skb)->last = skb; 554 if (!skb_is_gso(skb)) 555 skb_shinfo(skb)->gso_size = skb_gro_len(skb); 556 list_add(&skb->list, &gro_list->list); 557 ret = GRO_HELD; 558 ok: 559 if (gro_list->count) { 560 if (!test_bit(bucket, &gro->bitmask)) 561 __set_bit(bucket, &gro->bitmask); 562 } else if (test_bit(bucket, &gro->bitmask)) { 563 __clear_bit(bucket, &gro->bitmask); 564 } 565 566 return ret; 567 568 normal: 569 ret = GRO_NORMAL; 570 gro_try_pull_from_frag0(skb); 571 goto ok; 572 } 573 574 struct packet_offload *gro_find_receive_by_type(__be16 type) 575 { 576 struct list_head *offload_head = &net_hotdata.offload_base; 577 struct packet_offload *ptype; 578 579 list_for_each_entry_rcu(ptype, offload_head, list) { 580 if (ptype->type != type || !ptype->callbacks.gro_receive) 581 continue; 582 return ptype; 583 } 584 return NULL; 585 } 586 EXPORT_SYMBOL(gro_find_receive_by_type); 587 588 struct packet_offload *gro_find_complete_by_type(__be16 type) 589 { 590 struct list_head *offload_head = &net_hotdata.offload_base; 591 struct packet_offload *ptype; 592 593 list_for_each_entry_rcu(ptype, offload_head, list) { 594 if (ptype->type != type || !ptype->callbacks.gro_complete) 595 continue; 596 return ptype; 597 } 598 return NULL; 599 } 600 EXPORT_SYMBOL(gro_find_complete_by_type); 601 602 static gro_result_t gro_skb_finish(struct gro_node *gro, struct sk_buff *skb, 603 gro_result_t ret) 604 { 605 switch (ret) { 606 case GRO_NORMAL: 607 gro_normal_one(gro, skb, 1); 608 break; 609 610 case GRO_MERGED_FREE: 611 if (NAPI_GRO_CB(skb)->free == NAPI_GRO_FREE_STOLEN_HEAD) 612 napi_skb_free_stolen_head(skb); 613 else if (skb->fclone != SKB_FCLONE_UNAVAILABLE) 614 __kfree_skb(skb); 615 else 616 __napi_kfree_skb(skb, SKB_CONSUMED); 617 break; 618 619 case GRO_HELD: 620 case GRO_MERGED: 621 case GRO_CONSUMED: 622 break; 623 } 624 625 return ret; 626 } 627 628 gro_result_t gro_receive_skb(struct gro_node *gro, struct sk_buff *skb) 629 { 630 gro_result_t ret; 631 632 __skb_mark_napi_id(skb, gro); 633 trace_napi_gro_receive_entry(skb); 634 635 skb_gro_reset_offset(skb, 0); 636 637 ret = gro_skb_finish(gro, skb, dev_gro_receive(gro, skb)); 638 trace_napi_gro_receive_exit(ret); 639 640 return ret; 641 } 642 EXPORT_SYMBOL(gro_receive_skb); 643 644 static void napi_reuse_skb(struct napi_struct *napi, struct sk_buff *skb) 645 { 646 struct skb_shared_info *shinfo; 647 648 if (unlikely(skb->pfmemalloc)) { 649 consume_skb(skb); 650 return; 651 } 652 __skb_pull(skb, skb_headlen(skb)); 653 /* restore the reserve we had after netdev_alloc_skb_ip_align() */ 654 skb_reserve(skb, NET_SKB_PAD + NET_IP_ALIGN - skb_headroom(skb)); 655 __vlan_hwaccel_clear_tag(skb); 656 skb->dev = napi->dev; 657 skb->skb_iif = 0; 658 659 /* eth_type_trans() assumes pkt_type is PACKET_HOST */ 660 skb->pkt_type = PACKET_HOST; 661 662 skb->encapsulation = 0; 663 skb->ip_summed = CHECKSUM_NONE; 664 665 shinfo = skb_shinfo(skb); 666 shinfo->gso_type = 0; 667 shinfo->gso_size = 0; 668 shinfo->hwtstamps.hwtstamp = 0; 669 670 if (unlikely(skb->slow_gro)) { 671 skb_orphan(skb); 672 skb_ext_reset(skb); 673 nf_reset_ct(skb); 674 skb->slow_gro = 0; 675 } 676 677 napi->skb = skb; 678 } 679 680 struct sk_buff *napi_get_frags(struct napi_struct *napi) 681 { 682 struct sk_buff *skb = napi->skb; 683 684 if (!skb) { 685 skb = napi_alloc_skb(napi, GRO_MAX_HEAD); 686 if (skb) { 687 napi->skb = skb; 688 skb_mark_napi_id(skb, napi); 689 } 690 } 691 return skb; 692 } 693 EXPORT_SYMBOL(napi_get_frags); 694 695 static gro_result_t napi_frags_finish(struct napi_struct *napi, 696 struct sk_buff *skb, 697 gro_result_t ret) 698 { 699 switch (ret) { 700 case GRO_NORMAL: 701 case GRO_HELD: 702 __skb_push(skb, ETH_HLEN); 703 skb->protocol = eth_type_trans(skb, skb->dev); 704 if (ret == GRO_NORMAL) 705 gro_normal_one(&napi->gro, skb, 1); 706 break; 707 708 case GRO_MERGED_FREE: 709 if (NAPI_GRO_CB(skb)->free == NAPI_GRO_FREE_STOLEN_HEAD) 710 napi_skb_free_stolen_head(skb); 711 else 712 napi_reuse_skb(napi, skb); 713 break; 714 715 case GRO_MERGED: 716 case GRO_CONSUMED: 717 break; 718 } 719 720 return ret; 721 } 722 723 /* Upper GRO stack assumes network header starts at gro_offset=0 724 * Drivers could call both napi_gro_frags() and napi_gro_receive() 725 * We copy ethernet header into skb->data to have a common layout. 726 */ 727 static struct sk_buff *napi_frags_skb(struct napi_struct *napi) 728 { 729 struct sk_buff *skb = napi->skb; 730 const struct ethhdr *eth; 731 unsigned int hlen = sizeof(*eth); 732 733 napi->skb = NULL; 734 735 skb_reset_mac_header(skb); 736 skb_gro_reset_offset(skb, hlen); 737 738 if (unlikely(!skb_gro_may_pull(skb, hlen))) { 739 eth = skb_gro_header_slow(skb, hlen, 0); 740 if (unlikely(!eth)) { 741 net_warn_ratelimited("%s: dropping impossible skb from %s\n", 742 __func__, napi->dev->name); 743 napi_reuse_skb(napi, skb); 744 return NULL; 745 } 746 } else { 747 eth = (const struct ethhdr *)skb->data; 748 749 if (NAPI_GRO_CB(skb)->frag0 != skb->data) 750 gro_pull_from_frag0(skb, hlen); 751 752 NAPI_GRO_CB(skb)->frag0 += hlen; 753 NAPI_GRO_CB(skb)->frag0_len -= hlen; 754 } 755 __skb_pull(skb, hlen); 756 757 /* 758 * This works because the only protocols we care about don't require 759 * special handling. 760 * We'll fix it up properly in napi_frags_finish() 761 */ 762 skb->protocol = eth->h_proto; 763 764 return skb; 765 } 766 767 gro_result_t napi_gro_frags(struct napi_struct *napi) 768 { 769 gro_result_t ret; 770 struct sk_buff *skb = napi_frags_skb(napi); 771 772 trace_napi_gro_frags_entry(skb); 773 774 ret = napi_frags_finish(napi, skb, dev_gro_receive(&napi->gro, skb)); 775 trace_napi_gro_frags_exit(ret); 776 777 return ret; 778 } 779 EXPORT_SYMBOL(napi_gro_frags); 780 781 /* Compute the checksum from gro_offset and return the folded value 782 * after adding in any pseudo checksum. 783 */ 784 __sum16 __skb_gro_checksum_complete(struct sk_buff *skb) 785 { 786 __wsum wsum; 787 __sum16 sum; 788 789 wsum = skb_checksum(skb, skb_gro_offset(skb), skb_gro_len(skb), 0); 790 791 /* NAPI_GRO_CB(skb)->csum holds pseudo checksum */ 792 sum = csum_fold(csum_add(NAPI_GRO_CB(skb)->csum, wsum)); 793 /* See comments in __skb_checksum_complete(). */ 794 if (likely(!sum)) { 795 if (unlikely(skb->ip_summed == CHECKSUM_COMPLETE) && 796 !skb->csum_complete_sw) 797 netdev_rx_csum_fault(skb->dev, skb); 798 } 799 800 NAPI_GRO_CB(skb)->csum = wsum; 801 NAPI_GRO_CB(skb)->csum_valid = 1; 802 803 return sum; 804 } 805 EXPORT_SYMBOL(__skb_gro_checksum_complete); 806 807 void gro_init(struct gro_node *gro) 808 { 809 for (u32 i = 0; i < GRO_HASH_BUCKETS; i++) { 810 INIT_LIST_HEAD(&gro->hash[i].list); 811 gro->hash[i].count = 0; 812 } 813 814 gro->bitmask = 0; 815 gro->cached_napi_id = 0; 816 817 INIT_LIST_HEAD(&gro->rx_list); 818 gro->rx_count = 0; 819 } 820 821 void gro_cleanup(struct gro_node *gro) 822 { 823 struct sk_buff *skb, *n; 824 825 for (u32 i = 0; i < GRO_HASH_BUCKETS; i++) { 826 list_for_each_entry_safe(skb, n, &gro->hash[i].list, list) 827 kfree_skb(skb); 828 829 gro->hash[i].count = 0; 830 } 831 832 gro->bitmask = 0; 833 gro->cached_napi_id = 0; 834 835 list_for_each_entry_safe(skb, n, &gro->rx_list, list) 836 kfree_skb(skb); 837 838 gro->rx_count = 0; 839 } 840