Eric Dumazet | e456a18 | 2021-11-15 09:05:53 -0800 | [diff] [blame] | 1 | // SPDX-License-Identifier: GPL-2.0-or-later |
| 2 | #include <net/gro.h> |
Eric Dumazet | 587652b | 2021-11-15 09:05:54 -0800 | [diff] [blame] | 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 | static struct list_head offload_base __read_mostly = LIST_HEAD_INIT(offload_base); |
| 14 | /* Maximum number of GRO_NORMAL skbs to batch up for list-RX */ |
| 15 | int gro_normal_batch __read_mostly = 8; |
| 16 | |
| 17 | /** |
| 18 | * dev_add_offload - register offload handlers |
| 19 | * @po: protocol offload declaration |
| 20 | * |
| 21 | * Add protocol offload handlers to the networking stack. The passed |
| 22 | * &proto_offload is linked into kernel lists and may not be freed until |
| 23 | * it has been removed from the kernel lists. |
| 24 | * |
| 25 | * This call does not sleep therefore it can not |
| 26 | * guarantee all CPU's that are in middle of receiving packets |
| 27 | * will see the new offload handlers (until the next received packet). |
| 28 | */ |
| 29 | void dev_add_offload(struct packet_offload *po) |
| 30 | { |
| 31 | struct packet_offload *elem; |
| 32 | |
| 33 | spin_lock(&offload_lock); |
| 34 | list_for_each_entry(elem, &offload_base, list) { |
| 35 | if (po->priority < elem->priority) |
| 36 | break; |
| 37 | } |
| 38 | list_add_rcu(&po->list, elem->list.prev); |
| 39 | spin_unlock(&offload_lock); |
| 40 | } |
| 41 | EXPORT_SYMBOL(dev_add_offload); |
| 42 | |
| 43 | /** |
| 44 | * __dev_remove_offload - remove offload handler |
| 45 | * @po: packet offload declaration |
| 46 | * |
| 47 | * Remove a protocol offload handler that was previously added to the |
| 48 | * kernel offload handlers by dev_add_offload(). The passed &offload_type |
| 49 | * is removed from the kernel lists and can be freed or reused once this |
| 50 | * function returns. |
| 51 | * |
| 52 | * The packet type might still be in use by receivers |
| 53 | * and must not be freed until after all the CPU's have gone |
| 54 | * through a quiescent state. |
| 55 | */ |
| 56 | static void __dev_remove_offload(struct packet_offload *po) |
| 57 | { |
| 58 | struct list_head *head = &offload_base; |
| 59 | struct packet_offload *po1; |
| 60 | |
| 61 | spin_lock(&offload_lock); |
| 62 | |
| 63 | list_for_each_entry(po1, head, list) { |
| 64 | if (po == po1) { |
| 65 | list_del_rcu(&po->list); |
| 66 | goto out; |
| 67 | } |
| 68 | } |
| 69 | |
| 70 | pr_warn("dev_remove_offload: %p not found\n", po); |
| 71 | out: |
| 72 | spin_unlock(&offload_lock); |
| 73 | } |
| 74 | |
| 75 | /** |
| 76 | * dev_remove_offload - remove packet offload handler |
| 77 | * @po: packet offload declaration |
| 78 | * |
| 79 | * Remove a packet offload handler that was previously added to the kernel |
| 80 | * offload handlers by dev_add_offload(). The passed &offload_type is |
| 81 | * removed from the kernel lists and can be freed or reused once this |
| 82 | * function returns. |
| 83 | * |
| 84 | * This call sleeps to guarantee that no CPU is looking at the packet |
| 85 | * type after return. |
| 86 | */ |
| 87 | void dev_remove_offload(struct packet_offload *po) |
| 88 | { |
| 89 | __dev_remove_offload(po); |
| 90 | |
| 91 | synchronize_net(); |
| 92 | } |
| 93 | EXPORT_SYMBOL(dev_remove_offload); |
| 94 | |
| 95 | /** |
Steffen Klassert | 23c7f8d | 2022-03-07 13:11:41 +0100 | [diff] [blame] | 96 | * skb_eth_gso_segment - segmentation handler for ethernet protocols. |
| 97 | * @skb: buffer to segment |
| 98 | * @features: features for the output path (see dev->features) |
| 99 | * @type: Ethernet Protocol ID |
| 100 | */ |
| 101 | struct sk_buff *skb_eth_gso_segment(struct sk_buff *skb, |
| 102 | netdev_features_t features, __be16 type) |
| 103 | { |
| 104 | struct sk_buff *segs = ERR_PTR(-EPROTONOSUPPORT); |
| 105 | struct packet_offload *ptype; |
| 106 | |
| 107 | rcu_read_lock(); |
| 108 | list_for_each_entry_rcu(ptype, &offload_base, list) { |
| 109 | if (ptype->type == type && ptype->callbacks.gso_segment) { |
| 110 | segs = ptype->callbacks.gso_segment(skb, features); |
| 111 | break; |
| 112 | } |
| 113 | } |
| 114 | rcu_read_unlock(); |
| 115 | |
| 116 | return segs; |
| 117 | } |
| 118 | EXPORT_SYMBOL(skb_eth_gso_segment); |
| 119 | |
| 120 | /** |
Eric Dumazet | 587652b | 2021-11-15 09:05:54 -0800 | [diff] [blame] | 121 | * skb_mac_gso_segment - mac layer segmentation handler. |
| 122 | * @skb: buffer to segment |
| 123 | * @features: features for the output path (see dev->features) |
| 124 | */ |
| 125 | struct sk_buff *skb_mac_gso_segment(struct sk_buff *skb, |
| 126 | netdev_features_t features) |
| 127 | { |
| 128 | struct sk_buff *segs = ERR_PTR(-EPROTONOSUPPORT); |
| 129 | struct packet_offload *ptype; |
| 130 | int vlan_depth = skb->mac_len; |
| 131 | __be16 type = skb_network_protocol(skb, &vlan_depth); |
| 132 | |
| 133 | if (unlikely(!type)) |
| 134 | return ERR_PTR(-EINVAL); |
| 135 | |
| 136 | __skb_pull(skb, vlan_depth); |
| 137 | |
| 138 | rcu_read_lock(); |
| 139 | list_for_each_entry_rcu(ptype, &offload_base, list) { |
| 140 | if (ptype->type == type && ptype->callbacks.gso_segment) { |
| 141 | segs = ptype->callbacks.gso_segment(skb, features); |
| 142 | break; |
| 143 | } |
| 144 | } |
| 145 | rcu_read_unlock(); |
| 146 | |
| 147 | __skb_push(skb, skb->data - skb_mac_header(skb)); |
| 148 | |
| 149 | return segs; |
| 150 | } |
| 151 | EXPORT_SYMBOL(skb_mac_gso_segment); |
Eric Dumazet | e456a18 | 2021-11-15 09:05:53 -0800 | [diff] [blame] | 152 | |
| 153 | int skb_gro_receive(struct sk_buff *p, struct sk_buff *skb) |
| 154 | { |
| 155 | struct skb_shared_info *pinfo, *skbinfo = skb_shinfo(skb); |
| 156 | unsigned int offset = skb_gro_offset(skb); |
| 157 | unsigned int headlen = skb_headlen(skb); |
| 158 | unsigned int len = skb_gro_len(skb); |
| 159 | unsigned int delta_truesize; |
Coco Li | eac1b93 | 2022-01-05 02:48:38 -0800 | [diff] [blame] | 160 | unsigned int gro_max_size; |
Eric Dumazet | e456a18 | 2021-11-15 09:05:53 -0800 | [diff] [blame] | 161 | unsigned int new_truesize; |
| 162 | struct sk_buff *lp; |
| 163 | |
Coco Li | eac1b93 | 2022-01-05 02:48:38 -0800 | [diff] [blame] | 164 | /* pairs with WRITE_ONCE() in netif_set_gro_max_size() */ |
| 165 | gro_max_size = READ_ONCE(p->dev->gro_max_size); |
| 166 | |
| 167 | if (unlikely(p->len + len >= gro_max_size || NAPI_GRO_CB(skb)->flush)) |
Eric Dumazet | e456a18 | 2021-11-15 09:05:53 -0800 | [diff] [blame] | 168 | return -E2BIG; |
| 169 | |
Alexander Duyck | 0fe79f2 | 2022-05-13 11:34:03 -0700 | [diff] [blame] | 170 | if (unlikely(p->len + len >= GRO_LEGACY_MAX_SIZE)) { |
| 171 | if (p->protocol != htons(ETH_P_IPV6) || |
| 172 | skb_headroom(p) < sizeof(struct hop_jumbo_hdr) || |
| 173 | ipv6_hdr(p)->nexthdr != IPPROTO_TCP || |
| 174 | p->encapsulation) |
| 175 | return -E2BIG; |
| 176 | } |
| 177 | |
Eric Dumazet | e456a18 | 2021-11-15 09:05:53 -0800 | [diff] [blame] | 178 | lp = NAPI_GRO_CB(p)->last; |
| 179 | pinfo = skb_shinfo(lp); |
| 180 | |
| 181 | if (headlen <= offset) { |
| 182 | skb_frag_t *frag; |
| 183 | skb_frag_t *frag2; |
| 184 | int i = skbinfo->nr_frags; |
| 185 | int nr_frags = pinfo->nr_frags + i; |
| 186 | |
| 187 | if (nr_frags > MAX_SKB_FRAGS) |
| 188 | goto merge; |
| 189 | |
| 190 | offset -= headlen; |
| 191 | pinfo->nr_frags = nr_frags; |
| 192 | skbinfo->nr_frags = 0; |
| 193 | |
| 194 | frag = pinfo->frags + nr_frags; |
| 195 | frag2 = skbinfo->frags + i; |
| 196 | do { |
| 197 | *--frag = *--frag2; |
| 198 | } while (--i); |
| 199 | |
| 200 | skb_frag_off_add(frag, offset); |
| 201 | skb_frag_size_sub(frag, offset); |
| 202 | |
| 203 | /* all fragments truesize : remove (head size + sk_buff) */ |
| 204 | new_truesize = SKB_TRUESIZE(skb_end_offset(skb)); |
| 205 | delta_truesize = skb->truesize - new_truesize; |
| 206 | |
| 207 | skb->truesize = new_truesize; |
| 208 | skb->len -= skb->data_len; |
| 209 | skb->data_len = 0; |
| 210 | |
| 211 | NAPI_GRO_CB(skb)->free = NAPI_GRO_FREE; |
| 212 | goto done; |
| 213 | } else if (skb->head_frag) { |
| 214 | int nr_frags = pinfo->nr_frags; |
| 215 | skb_frag_t *frag = pinfo->frags + nr_frags; |
| 216 | struct page *page = virt_to_head_page(skb->head); |
| 217 | unsigned int first_size = headlen - offset; |
| 218 | unsigned int first_offset; |
| 219 | |
| 220 | if (nr_frags + 1 + skbinfo->nr_frags > MAX_SKB_FRAGS) |
| 221 | goto merge; |
| 222 | |
| 223 | first_offset = skb->data - |
| 224 | (unsigned char *)page_address(page) + |
| 225 | offset; |
| 226 | |
| 227 | pinfo->nr_frags = nr_frags + 1 + skbinfo->nr_frags; |
| 228 | |
| 229 | __skb_frag_set_page(frag, page); |
| 230 | skb_frag_off_set(frag, first_offset); |
| 231 | skb_frag_size_set(frag, first_size); |
| 232 | |
| 233 | memcpy(frag + 1, skbinfo->frags, sizeof(*frag) * skbinfo->nr_frags); |
| 234 | /* We dont need to clear skbinfo->nr_frags here */ |
| 235 | |
| 236 | new_truesize = SKB_DATA_ALIGN(sizeof(struct sk_buff)); |
| 237 | delta_truesize = skb->truesize - new_truesize; |
| 238 | skb->truesize = new_truesize; |
| 239 | NAPI_GRO_CB(skb)->free = NAPI_GRO_FREE_STOLEN_HEAD; |
| 240 | goto done; |
| 241 | } |
| 242 | |
| 243 | merge: |
| 244 | /* sk owenrship - if any - completely transferred to the aggregated packet */ |
| 245 | skb->destructor = NULL; |
| 246 | delta_truesize = skb->truesize; |
| 247 | if (offset > headlen) { |
| 248 | unsigned int eat = offset - headlen; |
| 249 | |
| 250 | skb_frag_off_add(&skbinfo->frags[0], eat); |
| 251 | skb_frag_size_sub(&skbinfo->frags[0], eat); |
| 252 | skb->data_len -= eat; |
| 253 | skb->len -= eat; |
| 254 | offset = headlen; |
| 255 | } |
| 256 | |
| 257 | __skb_pull(skb, offset); |
| 258 | |
| 259 | if (NAPI_GRO_CB(p)->last == p) |
| 260 | skb_shinfo(p)->frag_list = skb; |
| 261 | else |
| 262 | NAPI_GRO_CB(p)->last->next = skb; |
| 263 | NAPI_GRO_CB(p)->last = skb; |
| 264 | __skb_header_release(skb); |
| 265 | lp = p; |
| 266 | |
| 267 | done: |
| 268 | NAPI_GRO_CB(p)->count++; |
| 269 | p->data_len += len; |
| 270 | p->truesize += delta_truesize; |
| 271 | p->len += len; |
| 272 | if (lp != p) { |
| 273 | lp->data_len += len; |
| 274 | lp->truesize += delta_truesize; |
| 275 | lp->len += len; |
| 276 | } |
| 277 | NAPI_GRO_CB(skb)->same_flow = 1; |
| 278 | return 0; |
| 279 | } |
Eric Dumazet | 587652b | 2021-11-15 09:05:54 -0800 | [diff] [blame] | 280 | |
| 281 | |
| 282 | static void napi_gro_complete(struct napi_struct *napi, struct sk_buff *skb) |
| 283 | { |
| 284 | struct packet_offload *ptype; |
| 285 | __be16 type = skb->protocol; |
| 286 | struct list_head *head = &offload_base; |
| 287 | int err = -ENOENT; |
| 288 | |
| 289 | BUILD_BUG_ON(sizeof(struct napi_gro_cb) > sizeof(skb->cb)); |
| 290 | |
| 291 | if (NAPI_GRO_CB(skb)->count == 1) { |
| 292 | skb_shinfo(skb)->gso_size = 0; |
| 293 | goto out; |
| 294 | } |
| 295 | |
| 296 | rcu_read_lock(); |
| 297 | list_for_each_entry_rcu(ptype, head, list) { |
| 298 | if (ptype->type != type || !ptype->callbacks.gro_complete) |
| 299 | continue; |
| 300 | |
| 301 | err = INDIRECT_CALL_INET(ptype->callbacks.gro_complete, |
| 302 | ipv6_gro_complete, inet_gro_complete, |
| 303 | skb, 0); |
| 304 | break; |
| 305 | } |
| 306 | rcu_read_unlock(); |
| 307 | |
| 308 | if (err) { |
| 309 | WARN_ON(&ptype->list == head); |
| 310 | kfree_skb(skb); |
| 311 | return; |
| 312 | } |
| 313 | |
| 314 | out: |
| 315 | gro_normal_one(napi, skb, NAPI_GRO_CB(skb)->count); |
| 316 | } |
| 317 | |
| 318 | static void __napi_gro_flush_chain(struct napi_struct *napi, u32 index, |
| 319 | bool flush_old) |
| 320 | { |
| 321 | struct list_head *head = &napi->gro_hash[index].list; |
| 322 | struct sk_buff *skb, *p; |
| 323 | |
| 324 | list_for_each_entry_safe_reverse(skb, p, head, list) { |
| 325 | if (flush_old && NAPI_GRO_CB(skb)->age == jiffies) |
| 326 | return; |
| 327 | skb_list_del_init(skb); |
| 328 | napi_gro_complete(napi, skb); |
| 329 | napi->gro_hash[index].count--; |
| 330 | } |
| 331 | |
| 332 | if (!napi->gro_hash[index].count) |
| 333 | __clear_bit(index, &napi->gro_bitmask); |
| 334 | } |
| 335 | |
| 336 | /* napi->gro_hash[].list contains packets ordered by age. |
| 337 | * youngest packets at the head of it. |
| 338 | * Complete skbs in reverse order to reduce latencies. |
| 339 | */ |
| 340 | void napi_gro_flush(struct napi_struct *napi, bool flush_old) |
| 341 | { |
| 342 | unsigned long bitmask = napi->gro_bitmask; |
| 343 | unsigned int i, base = ~0U; |
| 344 | |
| 345 | while ((i = ffs(bitmask)) != 0) { |
| 346 | bitmask >>= i; |
| 347 | base += i; |
| 348 | __napi_gro_flush_chain(napi, base, flush_old); |
| 349 | } |
| 350 | } |
| 351 | EXPORT_SYMBOL(napi_gro_flush); |
| 352 | |
| 353 | static void gro_list_prepare(const struct list_head *head, |
| 354 | const struct sk_buff *skb) |
| 355 | { |
| 356 | unsigned int maclen = skb->dev->hard_header_len; |
| 357 | u32 hash = skb_get_hash_raw(skb); |
| 358 | struct sk_buff *p; |
| 359 | |
| 360 | list_for_each_entry(p, head, list) { |
| 361 | unsigned long diffs; |
| 362 | |
| 363 | NAPI_GRO_CB(p)->flush = 0; |
| 364 | |
| 365 | if (hash != skb_get_hash_raw(p)) { |
| 366 | NAPI_GRO_CB(p)->same_flow = 0; |
| 367 | continue; |
| 368 | } |
| 369 | |
| 370 | diffs = (unsigned long)p->dev ^ (unsigned long)skb->dev; |
| 371 | diffs |= skb_vlan_tag_present(p) ^ skb_vlan_tag_present(skb); |
| 372 | if (skb_vlan_tag_present(p)) |
| 373 | diffs |= skb_vlan_tag_get(p) ^ skb_vlan_tag_get(skb); |
| 374 | diffs |= skb_metadata_differs(p, skb); |
| 375 | if (maclen == ETH_HLEN) |
| 376 | diffs |= compare_ether_header(skb_mac_header(p), |
| 377 | skb_mac_header(skb)); |
| 378 | else if (!diffs) |
| 379 | diffs = memcmp(skb_mac_header(p), |
| 380 | skb_mac_header(skb), |
| 381 | maclen); |
| 382 | |
| 383 | /* in most common scenarions 'slow_gro' is 0 |
| 384 | * otherwise we are already on some slower paths |
| 385 | * either skip all the infrequent tests altogether or |
| 386 | * avoid trying too hard to skip each of them individually |
| 387 | */ |
| 388 | if (!diffs && unlikely(skb->slow_gro | p->slow_gro)) { |
| 389 | #if IS_ENABLED(CONFIG_SKB_EXTENSIONS) && IS_ENABLED(CONFIG_NET_TC_SKB_EXT) |
| 390 | struct tc_skb_ext *skb_ext; |
| 391 | struct tc_skb_ext *p_ext; |
| 392 | #endif |
| 393 | |
| 394 | diffs |= p->sk != skb->sk; |
| 395 | diffs |= skb_metadata_dst_cmp(p, skb); |
| 396 | diffs |= skb_get_nfct(p) ^ skb_get_nfct(skb); |
| 397 | |
| 398 | #if IS_ENABLED(CONFIG_SKB_EXTENSIONS) && IS_ENABLED(CONFIG_NET_TC_SKB_EXT) |
| 399 | skb_ext = skb_ext_find(skb, TC_SKB_EXT); |
| 400 | p_ext = skb_ext_find(p, TC_SKB_EXT); |
| 401 | |
| 402 | diffs |= (!!p_ext) ^ (!!skb_ext); |
| 403 | if (!diffs && unlikely(skb_ext)) |
| 404 | diffs |= p_ext->chain ^ skb_ext->chain; |
| 405 | #endif |
| 406 | } |
| 407 | |
| 408 | NAPI_GRO_CB(p)->same_flow = !diffs; |
| 409 | } |
| 410 | } |
| 411 | |
| 412 | static inline void skb_gro_reset_offset(struct sk_buff *skb, u32 nhoff) |
| 413 | { |
| 414 | const struct skb_shared_info *pinfo = skb_shinfo(skb); |
| 415 | const skb_frag_t *frag0 = &pinfo->frags[0]; |
| 416 | |
| 417 | NAPI_GRO_CB(skb)->data_offset = 0; |
| 418 | NAPI_GRO_CB(skb)->frag0 = NULL; |
| 419 | NAPI_GRO_CB(skb)->frag0_len = 0; |
| 420 | |
| 421 | if (!skb_headlen(skb) && pinfo->nr_frags && |
| 422 | !PageHighMem(skb_frag_page(frag0)) && |
| 423 | (!NET_IP_ALIGN || !((skb_frag_off(frag0) + nhoff) & 3))) { |
| 424 | NAPI_GRO_CB(skb)->frag0 = skb_frag_address(frag0); |
| 425 | NAPI_GRO_CB(skb)->frag0_len = min_t(unsigned int, |
| 426 | skb_frag_size(frag0), |
| 427 | skb->end - skb->tail); |
| 428 | } |
| 429 | } |
| 430 | |
| 431 | static void gro_pull_from_frag0(struct sk_buff *skb, int grow) |
| 432 | { |
| 433 | struct skb_shared_info *pinfo = skb_shinfo(skb); |
| 434 | |
| 435 | BUG_ON(skb->end - skb->tail < grow); |
| 436 | |
| 437 | memcpy(skb_tail_pointer(skb), NAPI_GRO_CB(skb)->frag0, grow); |
| 438 | |
| 439 | skb->data_len -= grow; |
| 440 | skb->tail += grow; |
| 441 | |
| 442 | skb_frag_off_add(&pinfo->frags[0], grow); |
| 443 | skb_frag_size_sub(&pinfo->frags[0], grow); |
| 444 | |
| 445 | if (unlikely(!skb_frag_size(&pinfo->frags[0]))) { |
| 446 | skb_frag_unref(skb, 0); |
| 447 | memmove(pinfo->frags, pinfo->frags + 1, |
| 448 | --pinfo->nr_frags * sizeof(pinfo->frags[0])); |
| 449 | } |
| 450 | } |
| 451 | |
| 452 | static void gro_flush_oldest(struct napi_struct *napi, struct list_head *head) |
| 453 | { |
| 454 | struct sk_buff *oldest; |
| 455 | |
| 456 | oldest = list_last_entry(head, struct sk_buff, list); |
| 457 | |
| 458 | /* We are called with head length >= MAX_GRO_SKBS, so this is |
| 459 | * impossible. |
| 460 | */ |
| 461 | if (WARN_ON_ONCE(!oldest)) |
| 462 | return; |
| 463 | |
| 464 | /* Do not adjust napi->gro_hash[].count, caller is adding a new |
| 465 | * SKB to the chain. |
| 466 | */ |
| 467 | skb_list_del_init(oldest); |
| 468 | napi_gro_complete(napi, oldest); |
| 469 | } |
| 470 | |
| 471 | static enum gro_result dev_gro_receive(struct napi_struct *napi, struct sk_buff *skb) |
| 472 | { |
| 473 | u32 bucket = skb_get_hash_raw(skb) & (GRO_HASH_BUCKETS - 1); |
| 474 | struct gro_list *gro_list = &napi->gro_hash[bucket]; |
| 475 | struct list_head *head = &offload_base; |
| 476 | struct packet_offload *ptype; |
| 477 | __be16 type = skb->protocol; |
| 478 | struct sk_buff *pp = NULL; |
| 479 | enum gro_result ret; |
| 480 | int same_flow; |
| 481 | int grow; |
| 482 | |
| 483 | if (netif_elide_gro(skb->dev)) |
| 484 | goto normal; |
| 485 | |
| 486 | gro_list_prepare(&gro_list->list, skb); |
| 487 | |
| 488 | rcu_read_lock(); |
| 489 | list_for_each_entry_rcu(ptype, head, list) { |
| 490 | if (ptype->type != type || !ptype->callbacks.gro_receive) |
| 491 | continue; |
| 492 | |
| 493 | skb_set_network_header(skb, skb_gro_offset(skb)); |
| 494 | skb_reset_mac_len(skb); |
Paolo Abeni | de5a1f3 | 2022-02-04 12:28:37 +0100 | [diff] [blame] | 495 | BUILD_BUG_ON(sizeof_field(struct napi_gro_cb, zeroed) != sizeof(u32)); |
| 496 | BUILD_BUG_ON(!IS_ALIGNED(offsetof(struct napi_gro_cb, zeroed), |
| 497 | sizeof(u32))); /* Avoid slow unaligned acc */ |
| 498 | *(u32 *)&NAPI_GRO_CB(skb)->zeroed = 0; |
Eric Dumazet | 587652b | 2021-11-15 09:05:54 -0800 | [diff] [blame] | 499 | NAPI_GRO_CB(skb)->flush = skb_is_gso(skb) || skb_has_frag_list(skb); |
Eric Dumazet | 587652b | 2021-11-15 09:05:54 -0800 | [diff] [blame] | 500 | NAPI_GRO_CB(skb)->is_atomic = 1; |
Eric Dumazet | 587652b | 2021-11-15 09:05:54 -0800 | [diff] [blame] | 501 | |
| 502 | /* Setup for GRO checksum validation */ |
| 503 | switch (skb->ip_summed) { |
| 504 | case CHECKSUM_COMPLETE: |
| 505 | NAPI_GRO_CB(skb)->csum = skb->csum; |
| 506 | NAPI_GRO_CB(skb)->csum_valid = 1; |
Eric Dumazet | 587652b | 2021-11-15 09:05:54 -0800 | [diff] [blame] | 507 | break; |
| 508 | case CHECKSUM_UNNECESSARY: |
| 509 | NAPI_GRO_CB(skb)->csum_cnt = skb->csum_level + 1; |
Eric Dumazet | 587652b | 2021-11-15 09:05:54 -0800 | [diff] [blame] | 510 | break; |
Eric Dumazet | 587652b | 2021-11-15 09:05:54 -0800 | [diff] [blame] | 511 | } |
| 512 | |
| 513 | pp = INDIRECT_CALL_INET(ptype->callbacks.gro_receive, |
| 514 | ipv6_gro_receive, inet_gro_receive, |
| 515 | &gro_list->list, skb); |
| 516 | break; |
| 517 | } |
| 518 | rcu_read_unlock(); |
| 519 | |
| 520 | if (&ptype->list == head) |
| 521 | goto normal; |
| 522 | |
| 523 | if (PTR_ERR(pp) == -EINPROGRESS) { |
| 524 | ret = GRO_CONSUMED; |
| 525 | goto ok; |
| 526 | } |
| 527 | |
| 528 | same_flow = NAPI_GRO_CB(skb)->same_flow; |
| 529 | ret = NAPI_GRO_CB(skb)->free ? GRO_MERGED_FREE : GRO_MERGED; |
| 530 | |
| 531 | if (pp) { |
| 532 | skb_list_del_init(pp); |
| 533 | napi_gro_complete(napi, pp); |
| 534 | gro_list->count--; |
| 535 | } |
| 536 | |
| 537 | if (same_flow) |
| 538 | goto ok; |
| 539 | |
| 540 | if (NAPI_GRO_CB(skb)->flush) |
| 541 | goto normal; |
| 542 | |
| 543 | if (unlikely(gro_list->count >= MAX_GRO_SKBS)) |
| 544 | gro_flush_oldest(napi, &gro_list->list); |
| 545 | else |
| 546 | gro_list->count++; |
| 547 | |
| 548 | NAPI_GRO_CB(skb)->count = 1; |
| 549 | NAPI_GRO_CB(skb)->age = jiffies; |
| 550 | NAPI_GRO_CB(skb)->last = skb; |
| 551 | skb_shinfo(skb)->gso_size = skb_gro_len(skb); |
| 552 | list_add(&skb->list, &gro_list->list); |
| 553 | ret = GRO_HELD; |
| 554 | |
| 555 | pull: |
| 556 | grow = skb_gro_offset(skb) - skb_headlen(skb); |
| 557 | if (grow > 0) |
| 558 | gro_pull_from_frag0(skb, grow); |
| 559 | ok: |
| 560 | if (gro_list->count) { |
| 561 | if (!test_bit(bucket, &napi->gro_bitmask)) |
| 562 | __set_bit(bucket, &napi->gro_bitmask); |
| 563 | } else if (test_bit(bucket, &napi->gro_bitmask)) { |
| 564 | __clear_bit(bucket, &napi->gro_bitmask); |
| 565 | } |
| 566 | |
| 567 | return ret; |
| 568 | |
| 569 | normal: |
| 570 | ret = GRO_NORMAL; |
| 571 | goto pull; |
| 572 | } |
| 573 | |
| 574 | struct packet_offload *gro_find_receive_by_type(__be16 type) |
| 575 | { |
| 576 | struct list_head *offload_head = &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 = &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 napi_skb_finish(struct napi_struct *napi, |
| 603 | struct sk_buff *skb, |
| 604 | gro_result_t ret) |
| 605 | { |
| 606 | switch (ret) { |
| 607 | case GRO_NORMAL: |
| 608 | gro_normal_one(napi, skb, 1); |
| 609 | break; |
| 610 | |
| 611 | case GRO_MERGED_FREE: |
| 612 | if (NAPI_GRO_CB(skb)->free == NAPI_GRO_FREE_STOLEN_HEAD) |
| 613 | napi_skb_free_stolen_head(skb); |
| 614 | else if (skb->fclone != SKB_FCLONE_UNAVAILABLE) |
| 615 | __kfree_skb(skb); |
| 616 | else |
| 617 | __kfree_skb_defer(skb); |
| 618 | break; |
| 619 | |
| 620 | case GRO_HELD: |
| 621 | case GRO_MERGED: |
| 622 | case GRO_CONSUMED: |
| 623 | break; |
| 624 | } |
| 625 | |
| 626 | return ret; |
| 627 | } |
| 628 | |
| 629 | gro_result_t napi_gro_receive(struct napi_struct *napi, struct sk_buff *skb) |
| 630 | { |
| 631 | gro_result_t ret; |
| 632 | |
| 633 | skb_mark_napi_id(skb, napi); |
| 634 | trace_napi_gro_receive_entry(skb); |
| 635 | |
| 636 | skb_gro_reset_offset(skb, 0); |
| 637 | |
| 638 | ret = napi_skb_finish(napi, skb, dev_gro_receive(napi, skb)); |
| 639 | trace_napi_gro_receive_exit(ret); |
| 640 | |
| 641 | return ret; |
| 642 | } |
| 643 | EXPORT_SYMBOL(napi_gro_receive); |
| 644 | |
| 645 | static void napi_reuse_skb(struct napi_struct *napi, struct sk_buff *skb) |
| 646 | { |
| 647 | if (unlikely(skb->pfmemalloc)) { |
| 648 | consume_skb(skb); |
| 649 | return; |
| 650 | } |
| 651 | __skb_pull(skb, skb_headlen(skb)); |
| 652 | /* restore the reserve we had after netdev_alloc_skb_ip_align() */ |
| 653 | skb_reserve(skb, NET_SKB_PAD + NET_IP_ALIGN - skb_headroom(skb)); |
| 654 | __vlan_hwaccel_clear_tag(skb); |
| 655 | skb->dev = napi->dev; |
| 656 | skb->skb_iif = 0; |
| 657 | |
| 658 | /* eth_type_trans() assumes pkt_type is PACKET_HOST */ |
| 659 | skb->pkt_type = PACKET_HOST; |
| 660 | |
| 661 | skb->encapsulation = 0; |
| 662 | skb_shinfo(skb)->gso_type = 0; |
Eric Dumazet | 587652b | 2021-11-15 09:05:54 -0800 | [diff] [blame] | 663 | if (unlikely(skb->slow_gro)) { |
| 664 | skb_orphan(skb); |
| 665 | skb_ext_reset(skb); |
| 666 | nf_reset_ct(skb); |
| 667 | skb->slow_gro = 0; |
| 668 | } |
| 669 | |
| 670 | napi->skb = skb; |
| 671 | } |
| 672 | |
| 673 | struct sk_buff *napi_get_frags(struct napi_struct *napi) |
| 674 | { |
| 675 | struct sk_buff *skb = napi->skb; |
| 676 | |
| 677 | if (!skb) { |
| 678 | skb = napi_alloc_skb(napi, GRO_MAX_HEAD); |
| 679 | if (skb) { |
| 680 | napi->skb = skb; |
| 681 | skb_mark_napi_id(skb, napi); |
| 682 | } |
| 683 | } |
| 684 | return skb; |
| 685 | } |
| 686 | EXPORT_SYMBOL(napi_get_frags); |
| 687 | |
| 688 | static gro_result_t napi_frags_finish(struct napi_struct *napi, |
| 689 | struct sk_buff *skb, |
| 690 | gro_result_t ret) |
| 691 | { |
| 692 | switch (ret) { |
| 693 | case GRO_NORMAL: |
| 694 | case GRO_HELD: |
| 695 | __skb_push(skb, ETH_HLEN); |
| 696 | skb->protocol = eth_type_trans(skb, skb->dev); |
| 697 | if (ret == GRO_NORMAL) |
| 698 | gro_normal_one(napi, skb, 1); |
| 699 | break; |
| 700 | |
| 701 | case GRO_MERGED_FREE: |
| 702 | if (NAPI_GRO_CB(skb)->free == NAPI_GRO_FREE_STOLEN_HEAD) |
| 703 | napi_skb_free_stolen_head(skb); |
| 704 | else |
| 705 | napi_reuse_skb(napi, skb); |
| 706 | break; |
| 707 | |
| 708 | case GRO_MERGED: |
| 709 | case GRO_CONSUMED: |
| 710 | break; |
| 711 | } |
| 712 | |
| 713 | return ret; |
| 714 | } |
| 715 | |
| 716 | /* Upper GRO stack assumes network header starts at gro_offset=0 |
| 717 | * Drivers could call both napi_gro_frags() and napi_gro_receive() |
| 718 | * We copy ethernet header into skb->data to have a common layout. |
| 719 | */ |
| 720 | static struct sk_buff *napi_frags_skb(struct napi_struct *napi) |
| 721 | { |
| 722 | struct sk_buff *skb = napi->skb; |
| 723 | const struct ethhdr *eth; |
| 724 | unsigned int hlen = sizeof(*eth); |
| 725 | |
| 726 | napi->skb = NULL; |
| 727 | |
| 728 | skb_reset_mac_header(skb); |
| 729 | skb_gro_reset_offset(skb, hlen); |
| 730 | |
| 731 | if (unlikely(skb_gro_header_hard(skb, hlen))) { |
| 732 | eth = skb_gro_header_slow(skb, hlen, 0); |
| 733 | if (unlikely(!eth)) { |
| 734 | net_warn_ratelimited("%s: dropping impossible skb from %s\n", |
| 735 | __func__, napi->dev->name); |
| 736 | napi_reuse_skb(napi, skb); |
| 737 | return NULL; |
| 738 | } |
| 739 | } else { |
| 740 | eth = (const struct ethhdr *)skb->data; |
| 741 | gro_pull_from_frag0(skb, hlen); |
| 742 | NAPI_GRO_CB(skb)->frag0 += hlen; |
| 743 | NAPI_GRO_CB(skb)->frag0_len -= hlen; |
| 744 | } |
| 745 | __skb_pull(skb, hlen); |
| 746 | |
| 747 | /* |
| 748 | * This works because the only protocols we care about don't require |
| 749 | * special handling. |
| 750 | * We'll fix it up properly in napi_frags_finish() |
| 751 | */ |
| 752 | skb->protocol = eth->h_proto; |
| 753 | |
| 754 | return skb; |
| 755 | } |
| 756 | |
| 757 | gro_result_t napi_gro_frags(struct napi_struct *napi) |
| 758 | { |
| 759 | gro_result_t ret; |
| 760 | struct sk_buff *skb = napi_frags_skb(napi); |
| 761 | |
| 762 | trace_napi_gro_frags_entry(skb); |
| 763 | |
| 764 | ret = napi_frags_finish(napi, skb, dev_gro_receive(napi, skb)); |
| 765 | trace_napi_gro_frags_exit(ret); |
| 766 | |
| 767 | return ret; |
| 768 | } |
| 769 | EXPORT_SYMBOL(napi_gro_frags); |
| 770 | |
| 771 | /* Compute the checksum from gro_offset and return the folded value |
| 772 | * after adding in any pseudo checksum. |
| 773 | */ |
| 774 | __sum16 __skb_gro_checksum_complete(struct sk_buff *skb) |
| 775 | { |
| 776 | __wsum wsum; |
| 777 | __sum16 sum; |
| 778 | |
| 779 | wsum = skb_checksum(skb, skb_gro_offset(skb), skb_gro_len(skb), 0); |
| 780 | |
| 781 | /* NAPI_GRO_CB(skb)->csum holds pseudo checksum */ |
| 782 | sum = csum_fold(csum_add(NAPI_GRO_CB(skb)->csum, wsum)); |
| 783 | /* See comments in __skb_checksum_complete(). */ |
| 784 | if (likely(!sum)) { |
| 785 | if (unlikely(skb->ip_summed == CHECKSUM_COMPLETE) && |
| 786 | !skb->csum_complete_sw) |
| 787 | netdev_rx_csum_fault(skb->dev, skb); |
| 788 | } |
| 789 | |
| 790 | NAPI_GRO_CB(skb)->csum = wsum; |
| 791 | NAPI_GRO_CB(skb)->csum_valid = 1; |
| 792 | |
| 793 | return sum; |
| 794 | } |
| 795 | EXPORT_SYMBOL(__skb_gro_checksum_complete); |