Loading...
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
10static DEFINE_SPINLOCK(offload_lock);
11
12/**
13 * dev_add_offload - register offload handlers
14 * @po: protocol offload declaration
15 *
16 * Add protocol offload handlers to the networking stack. The passed
17 * &proto_offload is linked into kernel lists and may not be freed until
18 * it has been removed from the kernel lists.
19 *
20 * This call does not sleep therefore it can not
21 * guarantee all CPU's that are in middle of receiving packets
22 * will see the new offload handlers (until the next received packet).
23 */
24void dev_add_offload(struct packet_offload *po)
25{
26 struct packet_offload *elem;
27
28 spin_lock(&offload_lock);
29 list_for_each_entry(elem, &net_hotdata.offload_base, list) {
30 if (po->priority < elem->priority)
31 break;
32 }
33 list_add_rcu(&po->list, elem->list.prev);
34 spin_unlock(&offload_lock);
35}
36EXPORT_SYMBOL(dev_add_offload);
37
38/**
39 * __dev_remove_offload - remove offload handler
40 * @po: packet offload declaration
41 *
42 * Remove a protocol offload handler that was previously added to the
43 * kernel offload handlers by dev_add_offload(). The passed &offload_type
44 * is removed from the kernel lists and can be freed or reused once this
45 * function returns.
46 *
47 * The packet type might still be in use by receivers
48 * and must not be freed until after all the CPU's have gone
49 * through a quiescent state.
50 */
51static void __dev_remove_offload(struct packet_offload *po)
52{
53 struct list_head *head = &net_hotdata.offload_base;
54 struct packet_offload *po1;
55
56 spin_lock(&offload_lock);
57
58 list_for_each_entry(po1, head, list) {
59 if (po == po1) {
60 list_del_rcu(&po->list);
61 goto out;
62 }
63 }
64
65 pr_warn("dev_remove_offload: %p not found\n", po);
66out:
67 spin_unlock(&offload_lock);
68}
69
70/**
71 * dev_remove_offload - remove packet offload handler
72 * @po: packet offload declaration
73 *
74 * Remove a packet offload handler that was previously added to the kernel
75 * offload handlers by dev_add_offload(). The passed &offload_type is
76 * removed from the kernel lists and can be freed or reused once this
77 * function returns.
78 *
79 * This call sleeps to guarantee that no CPU is looking at the packet
80 * type after return.
81 */
82void dev_remove_offload(struct packet_offload *po)
83{
84 __dev_remove_offload(po);
85
86 synchronize_net();
87}
88EXPORT_SYMBOL(dev_remove_offload);
89
90
91int skb_gro_receive(struct sk_buff *p, struct sk_buff *skb)
92{
93 struct skb_shared_info *pinfo, *skbinfo = skb_shinfo(skb);
94 unsigned int offset = skb_gro_offset(skb);
95 unsigned int headlen = skb_headlen(skb);
96 unsigned int len = skb_gro_len(skb);
97 unsigned int delta_truesize;
98 unsigned int new_truesize;
99 struct sk_buff *lp;
100 int segs;
101
102 /* Do not splice page pool based packets w/ non-page pool
103 * packets. This can result in reference count issues as page
104 * pool pages will not decrement the reference count and will
105 * instead be immediately returned to the pool or have frag
106 * count decremented.
107 */
108 if (p->pp_recycle != skb->pp_recycle)
109 return -ETOOMANYREFS;
110
111 if (unlikely(p->len + len >= netif_get_gro_max_size(p->dev, p) ||
112 NAPI_GRO_CB(skb)->flush))
113 return -E2BIG;
114
115 if (unlikely(p->len + len >= GRO_LEGACY_MAX_SIZE)) {
116 if (NAPI_GRO_CB(skb)->proto != IPPROTO_TCP ||
117 (p->protocol == htons(ETH_P_IPV6) &&
118 skb_headroom(p) < sizeof(struct hop_jumbo_hdr)) ||
119 p->encapsulation)
120 return -E2BIG;
121 }
122
123 segs = NAPI_GRO_CB(skb)->count;
124 lp = NAPI_GRO_CB(p)->last;
125 pinfo = skb_shinfo(lp);
126
127 if (headlen <= offset) {
128 skb_frag_t *frag;
129 skb_frag_t *frag2;
130 int i = skbinfo->nr_frags;
131 int nr_frags = pinfo->nr_frags + i;
132
133 if (nr_frags > MAX_SKB_FRAGS)
134 goto merge;
135
136 offset -= headlen;
137 pinfo->nr_frags = nr_frags;
138 skbinfo->nr_frags = 0;
139
140 frag = pinfo->frags + nr_frags;
141 frag2 = skbinfo->frags + i;
142 do {
143 *--frag = *--frag2;
144 } while (--i);
145
146 skb_frag_off_add(frag, offset);
147 skb_frag_size_sub(frag, offset);
148
149 /* all fragments truesize : remove (head size + sk_buff) */
150 new_truesize = SKB_TRUESIZE(skb_end_offset(skb));
151 delta_truesize = skb->truesize - new_truesize;
152
153 skb->truesize = new_truesize;
154 skb->len -= skb->data_len;
155 skb->data_len = 0;
156
157 NAPI_GRO_CB(skb)->free = NAPI_GRO_FREE;
158 goto done;
159 } else if (skb->head_frag) {
160 int nr_frags = pinfo->nr_frags;
161 skb_frag_t *frag = pinfo->frags + nr_frags;
162 struct page *page = virt_to_head_page(skb->head);
163 unsigned int first_size = headlen - offset;
164 unsigned int first_offset;
165
166 if (nr_frags + 1 + skbinfo->nr_frags > MAX_SKB_FRAGS)
167 goto merge;
168
169 first_offset = skb->data -
170 (unsigned char *)page_address(page) +
171 offset;
172
173 pinfo->nr_frags = nr_frags + 1 + skbinfo->nr_frags;
174
175 skb_frag_fill_page_desc(frag, page, first_offset, first_size);
176
177 memcpy(frag + 1, skbinfo->frags, sizeof(*frag) * skbinfo->nr_frags);
178 /* We dont need to clear skbinfo->nr_frags here */
179
180 new_truesize = SKB_DATA_ALIGN(sizeof(struct sk_buff));
181 delta_truesize = skb->truesize - new_truesize;
182 skb->truesize = new_truesize;
183 NAPI_GRO_CB(skb)->free = NAPI_GRO_FREE_STOLEN_HEAD;
184 goto done;
185 }
186
187merge:
188 /* sk ownership - if any - completely transferred to the aggregated packet */
189 skb->destructor = NULL;
190 skb->sk = NULL;
191 delta_truesize = skb->truesize;
192 if (offset > headlen) {
193 unsigned int eat = offset - headlen;
194
195 skb_frag_off_add(&skbinfo->frags[0], eat);
196 skb_frag_size_sub(&skbinfo->frags[0], eat);
197 skb->data_len -= eat;
198 skb->len -= eat;
199 offset = headlen;
200 }
201
202 __skb_pull(skb, offset);
203
204 if (NAPI_GRO_CB(p)->last == p)
205 skb_shinfo(p)->frag_list = skb;
206 else
207 NAPI_GRO_CB(p)->last->next = skb;
208 NAPI_GRO_CB(p)->last = skb;
209 __skb_header_release(skb);
210 lp = p;
211
212done:
213 NAPI_GRO_CB(p)->count += segs;
214 p->data_len += len;
215 p->truesize += delta_truesize;
216 p->len += len;
217 if (lp != p) {
218 lp->data_len += len;
219 lp->truesize += delta_truesize;
220 lp->len += len;
221 }
222 NAPI_GRO_CB(skb)->same_flow = 1;
223 return 0;
224}
225
226int skb_gro_receive_list(struct sk_buff *p, struct sk_buff *skb)
227{
228 if (unlikely(p->len + skb->len >= 65536))
229 return -E2BIG;
230
231 if (NAPI_GRO_CB(p)->last == p)
232 skb_shinfo(p)->frag_list = skb;
233 else
234 NAPI_GRO_CB(p)->last->next = skb;
235
236 skb_pull(skb, skb_gro_offset(skb));
237
238 NAPI_GRO_CB(p)->last = skb;
239 NAPI_GRO_CB(p)->count++;
240 p->data_len += skb->len;
241
242 /* sk ownership - if any - completely transferred to the aggregated packet */
243 skb->destructor = NULL;
244 skb->sk = NULL;
245 p->truesize += skb->truesize;
246 p->len += skb->len;
247
248 NAPI_GRO_CB(skb)->same_flow = 1;
249
250 return 0;
251}
252
253
254static void napi_gro_complete(struct napi_struct *napi, struct sk_buff *skb)
255{
256 struct list_head *head = &net_hotdata.offload_base;
257 struct packet_offload *ptype;
258 __be16 type = skb->protocol;
259 int err = -ENOENT;
260
261 BUILD_BUG_ON(sizeof(struct napi_gro_cb) > sizeof(skb->cb));
262
263 if (NAPI_GRO_CB(skb)->count == 1) {
264 skb_shinfo(skb)->gso_size = 0;
265 goto out;
266 }
267
268 rcu_read_lock();
269 list_for_each_entry_rcu(ptype, head, list) {
270 if (ptype->type != type || !ptype->callbacks.gro_complete)
271 continue;
272
273 err = INDIRECT_CALL_INET(ptype->callbacks.gro_complete,
274 ipv6_gro_complete, inet_gro_complete,
275 skb, 0);
276 break;
277 }
278 rcu_read_unlock();
279
280 if (err) {
281 WARN_ON(&ptype->list == head);
282 kfree_skb(skb);
283 return;
284 }
285
286out:
287 gro_normal_one(napi, skb, NAPI_GRO_CB(skb)->count);
288}
289
290static void __napi_gro_flush_chain(struct napi_struct *napi, u32 index,
291 bool flush_old)
292{
293 struct list_head *head = &napi->gro_hash[index].list;
294 struct sk_buff *skb, *p;
295
296 list_for_each_entry_safe_reverse(skb, p, head, list) {
297 if (flush_old && NAPI_GRO_CB(skb)->age == jiffies)
298 return;
299 skb_list_del_init(skb);
300 napi_gro_complete(napi, skb);
301 napi->gro_hash[index].count--;
302 }
303
304 if (!napi->gro_hash[index].count)
305 __clear_bit(index, &napi->gro_bitmask);
306}
307
308/* napi->gro_hash[].list contains packets ordered by age.
309 * youngest packets at the head of it.
310 * Complete skbs in reverse order to reduce latencies.
311 */
312void napi_gro_flush(struct napi_struct *napi, bool flush_old)
313{
314 unsigned long bitmask = napi->gro_bitmask;
315 unsigned int i, base = ~0U;
316
317 while ((i = ffs(bitmask)) != 0) {
318 bitmask >>= i;
319 base += i;
320 __napi_gro_flush_chain(napi, base, flush_old);
321 }
322}
323EXPORT_SYMBOL(napi_gro_flush);
324
325static unsigned long gro_list_prepare_tc_ext(const struct sk_buff *skb,
326 const struct sk_buff *p,
327 unsigned long diffs)
328{
329#if IS_ENABLED(CONFIG_NET_TC_SKB_EXT)
330 struct tc_skb_ext *skb_ext;
331 struct tc_skb_ext *p_ext;
332
333 skb_ext = skb_ext_find(skb, TC_SKB_EXT);
334 p_ext = skb_ext_find(p, TC_SKB_EXT);
335
336 diffs |= (!!p_ext) ^ (!!skb_ext);
337 if (!diffs && unlikely(skb_ext))
338 diffs |= p_ext->chain ^ skb_ext->chain;
339#endif
340 return diffs;
341}
342
343static void gro_list_prepare(const struct list_head *head,
344 const struct sk_buff *skb)
345{
346 unsigned int maclen = skb->dev->hard_header_len;
347 u32 hash = skb_get_hash_raw(skb);
348 struct sk_buff *p;
349
350 list_for_each_entry(p, head, list) {
351 unsigned long diffs;
352
353 if (hash != skb_get_hash_raw(p)) {
354 NAPI_GRO_CB(p)->same_flow = 0;
355 continue;
356 }
357
358 diffs = (unsigned long)p->dev ^ (unsigned long)skb->dev;
359 diffs |= p->vlan_all ^ skb->vlan_all;
360 diffs |= skb_metadata_differs(p, skb);
361 if (maclen == ETH_HLEN)
362 diffs |= compare_ether_header(skb_mac_header(p),
363 skb_mac_header(skb));
364 else if (!diffs)
365 diffs = memcmp(skb_mac_header(p),
366 skb_mac_header(skb),
367 maclen);
368
369 /* in most common scenarios 'slow_gro' is 0
370 * otherwise we are already on some slower paths
371 * either skip all the infrequent tests altogether or
372 * avoid trying too hard to skip each of them individually
373 */
374 if (!diffs && unlikely(skb->slow_gro | p->slow_gro)) {
375 diffs |= p->sk != skb->sk;
376 diffs |= skb_metadata_dst_cmp(p, skb);
377 diffs |= skb_get_nfct(p) ^ skb_get_nfct(skb);
378
379 diffs |= gro_list_prepare_tc_ext(skb, p, diffs);
380 }
381
382 NAPI_GRO_CB(p)->same_flow = !diffs;
383 }
384}
385
386static inline void skb_gro_reset_offset(struct sk_buff *skb, u32 nhoff)
387{
388 const struct skb_shared_info *pinfo;
389 const skb_frag_t *frag0;
390 unsigned int headlen;
391
392 NAPI_GRO_CB(skb)->network_offset = 0;
393 NAPI_GRO_CB(skb)->data_offset = 0;
394 headlen = skb_headlen(skb);
395 NAPI_GRO_CB(skb)->frag0 = skb->data;
396 NAPI_GRO_CB(skb)->frag0_len = headlen;
397 if (headlen)
398 return;
399
400 pinfo = skb_shinfo(skb);
401 frag0 = &pinfo->frags[0];
402
403 if (pinfo->nr_frags && skb_frag_page(frag0) &&
404 !PageHighMem(skb_frag_page(frag0)) &&
405 (!NET_IP_ALIGN || !((skb_frag_off(frag0) + nhoff) & 3))) {
406 NAPI_GRO_CB(skb)->frag0 = skb_frag_address(frag0);
407 NAPI_GRO_CB(skb)->frag0_len = min_t(unsigned int,
408 skb_frag_size(frag0),
409 skb->end - skb->tail);
410 }
411}
412
413static void gro_pull_from_frag0(struct sk_buff *skb, int grow)
414{
415 struct skb_shared_info *pinfo = skb_shinfo(skb);
416
417 BUG_ON(skb->end - skb->tail < grow);
418
419 memcpy(skb_tail_pointer(skb), NAPI_GRO_CB(skb)->frag0, grow);
420
421 skb->data_len -= grow;
422 skb->tail += grow;
423
424 skb_frag_off_add(&pinfo->frags[0], grow);
425 skb_frag_size_sub(&pinfo->frags[0], grow);
426
427 if (unlikely(!skb_frag_size(&pinfo->frags[0]))) {
428 skb_frag_unref(skb, 0);
429 memmove(pinfo->frags, pinfo->frags + 1,
430 --pinfo->nr_frags * sizeof(pinfo->frags[0]));
431 }
432}
433
434static void gro_try_pull_from_frag0(struct sk_buff *skb)
435{
436 int grow = skb_gro_offset(skb) - skb_headlen(skb);
437
438 if (grow > 0)
439 gro_pull_from_frag0(skb, grow);
440}
441
442static void gro_flush_oldest(struct napi_struct *napi, struct list_head *head)
443{
444 struct sk_buff *oldest;
445
446 oldest = list_last_entry(head, struct sk_buff, list);
447
448 /* We are called with head length >= MAX_GRO_SKBS, so this is
449 * impossible.
450 */
451 if (WARN_ON_ONCE(!oldest))
452 return;
453
454 /* Do not adjust napi->gro_hash[].count, caller is adding a new
455 * SKB to the chain.
456 */
457 skb_list_del_init(oldest);
458 napi_gro_complete(napi, oldest);
459}
460
461static enum gro_result dev_gro_receive(struct napi_struct *napi, struct sk_buff *skb)
462{
463 u32 bucket = skb_get_hash_raw(skb) & (GRO_HASH_BUCKETS - 1);
464 struct gro_list *gro_list = &napi->gro_hash[bucket];
465 struct list_head *head = &net_hotdata.offload_base;
466 struct packet_offload *ptype;
467 __be16 type = skb->protocol;
468 struct sk_buff *pp = NULL;
469 enum gro_result ret;
470 int same_flow;
471
472 if (netif_elide_gro(skb->dev))
473 goto normal;
474
475 gro_list_prepare(&gro_list->list, skb);
476
477 rcu_read_lock();
478 list_for_each_entry_rcu(ptype, head, list) {
479 if (ptype->type == type && ptype->callbacks.gro_receive)
480 goto found_ptype;
481 }
482 rcu_read_unlock();
483 goto normal;
484
485found_ptype:
486 skb_set_network_header(skb, skb_gro_offset(skb));
487 skb_reset_mac_len(skb);
488 BUILD_BUG_ON(sizeof_field(struct napi_gro_cb, zeroed) != sizeof(u32));
489 BUILD_BUG_ON(!IS_ALIGNED(offsetof(struct napi_gro_cb, zeroed),
490 sizeof(u32))); /* Avoid slow unaligned acc */
491 *(u32 *)&NAPI_GRO_CB(skb)->zeroed = 0;
492 NAPI_GRO_CB(skb)->flush = skb_has_frag_list(skb);
493 NAPI_GRO_CB(skb)->count = 1;
494 if (unlikely(skb_is_gso(skb))) {
495 NAPI_GRO_CB(skb)->count = skb_shinfo(skb)->gso_segs;
496 /* Only support TCP and non DODGY users. */
497 if (!skb_is_gso_tcp(skb) ||
498 (skb_shinfo(skb)->gso_type & SKB_GSO_DODGY))
499 NAPI_GRO_CB(skb)->flush = 1;
500 }
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;
507 break;
508 case CHECKSUM_UNNECESSARY:
509 NAPI_GRO_CB(skb)->csum_cnt = skb->csum_level + 1;
510 break;
511 }
512
513 pp = INDIRECT_CALL_INET(ptype->callbacks.gro_receive,
514 ipv6_gro_receive, inet_gro_receive,
515 &gro_list->list, skb);
516
517 rcu_read_unlock();
518
519 if (PTR_ERR(pp) == -EINPROGRESS) {
520 ret = GRO_CONSUMED;
521 goto ok;
522 }
523
524 same_flow = NAPI_GRO_CB(skb)->same_flow;
525 ret = NAPI_GRO_CB(skb)->free ? GRO_MERGED_FREE : GRO_MERGED;
526
527 if (pp) {
528 skb_list_del_init(pp);
529 napi_gro_complete(napi, pp);
530 gro_list->count--;
531 }
532
533 if (same_flow)
534 goto ok;
535
536 if (NAPI_GRO_CB(skb)->flush)
537 goto normal;
538
539 if (unlikely(gro_list->count >= MAX_GRO_SKBS))
540 gro_flush_oldest(napi, &gro_list->list);
541 else
542 gro_list->count++;
543
544 /* Must be called before setting NAPI_GRO_CB(skb)->{age|last} */
545 gro_try_pull_from_frag0(skb);
546 NAPI_GRO_CB(skb)->age = jiffies;
547 NAPI_GRO_CB(skb)->last = skb;
548 if (!skb_is_gso(skb))
549 skb_shinfo(skb)->gso_size = skb_gro_len(skb);
550 list_add(&skb->list, &gro_list->list);
551 ret = GRO_HELD;
552ok:
553 if (gro_list->count) {
554 if (!test_bit(bucket, &napi->gro_bitmask))
555 __set_bit(bucket, &napi->gro_bitmask);
556 } else if (test_bit(bucket, &napi->gro_bitmask)) {
557 __clear_bit(bucket, &napi->gro_bitmask);
558 }
559
560 return ret;
561
562normal:
563 ret = GRO_NORMAL;
564 gro_try_pull_from_frag0(skb);
565 goto ok;
566}
567
568struct packet_offload *gro_find_receive_by_type(__be16 type)
569{
570 struct list_head *offload_head = &net_hotdata.offload_base;
571 struct packet_offload *ptype;
572
573 list_for_each_entry_rcu(ptype, offload_head, list) {
574 if (ptype->type != type || !ptype->callbacks.gro_receive)
575 continue;
576 return ptype;
577 }
578 return NULL;
579}
580EXPORT_SYMBOL(gro_find_receive_by_type);
581
582struct packet_offload *gro_find_complete_by_type(__be16 type)
583{
584 struct list_head *offload_head = &net_hotdata.offload_base;
585 struct packet_offload *ptype;
586
587 list_for_each_entry_rcu(ptype, offload_head, list) {
588 if (ptype->type != type || !ptype->callbacks.gro_complete)
589 continue;
590 return ptype;
591 }
592 return NULL;
593}
594EXPORT_SYMBOL(gro_find_complete_by_type);
595
596static gro_result_t napi_skb_finish(struct napi_struct *napi,
597 struct sk_buff *skb,
598 gro_result_t ret)
599{
600 switch (ret) {
601 case GRO_NORMAL:
602 gro_normal_one(napi, skb, 1);
603 break;
604
605 case GRO_MERGED_FREE:
606 if (NAPI_GRO_CB(skb)->free == NAPI_GRO_FREE_STOLEN_HEAD)
607 napi_skb_free_stolen_head(skb);
608 else if (skb->fclone != SKB_FCLONE_UNAVAILABLE)
609 __kfree_skb(skb);
610 else
611 __napi_kfree_skb(skb, SKB_CONSUMED);
612 break;
613
614 case GRO_HELD:
615 case GRO_MERGED:
616 case GRO_CONSUMED:
617 break;
618 }
619
620 return ret;
621}
622
623gro_result_t napi_gro_receive(struct napi_struct *napi, struct sk_buff *skb)
624{
625 gro_result_t ret;
626
627 skb_mark_napi_id(skb, napi);
628 trace_napi_gro_receive_entry(skb);
629
630 skb_gro_reset_offset(skb, 0);
631
632 ret = napi_skb_finish(napi, skb, dev_gro_receive(napi, skb));
633 trace_napi_gro_receive_exit(ret);
634
635 return ret;
636}
637EXPORT_SYMBOL(napi_gro_receive);
638
639static void napi_reuse_skb(struct napi_struct *napi, struct sk_buff *skb)
640{
641 if (unlikely(skb->pfmemalloc)) {
642 consume_skb(skb);
643 return;
644 }
645 __skb_pull(skb, skb_headlen(skb));
646 /* restore the reserve we had after netdev_alloc_skb_ip_align() */
647 skb_reserve(skb, NET_SKB_PAD + NET_IP_ALIGN - skb_headroom(skb));
648 __vlan_hwaccel_clear_tag(skb);
649 skb->dev = napi->dev;
650 skb->skb_iif = 0;
651
652 /* eth_type_trans() assumes pkt_type is PACKET_HOST */
653 skb->pkt_type = PACKET_HOST;
654
655 skb->encapsulation = 0;
656 skb->ip_summed = CHECKSUM_NONE;
657 skb_shinfo(skb)->gso_type = 0;
658 skb_shinfo(skb)->gso_size = 0;
659 if (unlikely(skb->slow_gro)) {
660 skb_orphan(skb);
661 skb_ext_reset(skb);
662 nf_reset_ct(skb);
663 skb->slow_gro = 0;
664 }
665
666 napi->skb = skb;
667}
668
669struct sk_buff *napi_get_frags(struct napi_struct *napi)
670{
671 struct sk_buff *skb = napi->skb;
672
673 if (!skb) {
674 skb = napi_alloc_skb(napi, GRO_MAX_HEAD);
675 if (skb) {
676 napi->skb = skb;
677 skb_mark_napi_id(skb, napi);
678 }
679 }
680 return skb;
681}
682EXPORT_SYMBOL(napi_get_frags);
683
684static gro_result_t napi_frags_finish(struct napi_struct *napi,
685 struct sk_buff *skb,
686 gro_result_t ret)
687{
688 switch (ret) {
689 case GRO_NORMAL:
690 case GRO_HELD:
691 __skb_push(skb, ETH_HLEN);
692 skb->protocol = eth_type_trans(skb, skb->dev);
693 if (ret == GRO_NORMAL)
694 gro_normal_one(napi, skb, 1);
695 break;
696
697 case GRO_MERGED_FREE:
698 if (NAPI_GRO_CB(skb)->free == NAPI_GRO_FREE_STOLEN_HEAD)
699 napi_skb_free_stolen_head(skb);
700 else
701 napi_reuse_skb(napi, skb);
702 break;
703
704 case GRO_MERGED:
705 case GRO_CONSUMED:
706 break;
707 }
708
709 return ret;
710}
711
712/* Upper GRO stack assumes network header starts at gro_offset=0
713 * Drivers could call both napi_gro_frags() and napi_gro_receive()
714 * We copy ethernet header into skb->data to have a common layout.
715 */
716static struct sk_buff *napi_frags_skb(struct napi_struct *napi)
717{
718 struct sk_buff *skb = napi->skb;
719 const struct ethhdr *eth;
720 unsigned int hlen = sizeof(*eth);
721
722 napi->skb = NULL;
723
724 skb_reset_mac_header(skb);
725 skb_gro_reset_offset(skb, hlen);
726
727 if (unlikely(!skb_gro_may_pull(skb, hlen))) {
728 eth = skb_gro_header_slow(skb, hlen, 0);
729 if (unlikely(!eth)) {
730 net_warn_ratelimited("%s: dropping impossible skb from %s\n",
731 __func__, napi->dev->name);
732 napi_reuse_skb(napi, skb);
733 return NULL;
734 }
735 } else {
736 eth = (const struct ethhdr *)skb->data;
737
738 if (NAPI_GRO_CB(skb)->frag0 != skb->data)
739 gro_pull_from_frag0(skb, hlen);
740
741 NAPI_GRO_CB(skb)->frag0 += hlen;
742 NAPI_GRO_CB(skb)->frag0_len -= hlen;
743 }
744 __skb_pull(skb, hlen);
745
746 /*
747 * This works because the only protocols we care about don't require
748 * special handling.
749 * We'll fix it up properly in napi_frags_finish()
750 */
751 skb->protocol = eth->h_proto;
752
753 return skb;
754}
755
756gro_result_t napi_gro_frags(struct napi_struct *napi)
757{
758 gro_result_t ret;
759 struct sk_buff *skb = napi_frags_skb(napi);
760
761 trace_napi_gro_frags_entry(skb);
762
763 ret = napi_frags_finish(napi, skb, dev_gro_receive(napi, skb));
764 trace_napi_gro_frags_exit(ret);
765
766 return ret;
767}
768EXPORT_SYMBOL(napi_gro_frags);
769
770/* Compute the checksum from gro_offset and return the folded value
771 * after adding in any pseudo checksum.
772 */
773__sum16 __skb_gro_checksum_complete(struct sk_buff *skb)
774{
775 __wsum wsum;
776 __sum16 sum;
777
778 wsum = skb_checksum(skb, skb_gro_offset(skb), skb_gro_len(skb), 0);
779
780 /* NAPI_GRO_CB(skb)->csum holds pseudo checksum */
781 sum = csum_fold(csum_add(NAPI_GRO_CB(skb)->csum, wsum));
782 /* See comments in __skb_checksum_complete(). */
783 if (likely(!sum)) {
784 if (unlikely(skb->ip_summed == CHECKSUM_COMPLETE) &&
785 !skb->csum_complete_sw)
786 netdev_rx_csum_fault(skb->dev, skb);
787 }
788
789 NAPI_GRO_CB(skb)->csum = wsum;
790 NAPI_GRO_CB(skb)->csum_valid = 1;
791
792 return sum;
793}
794EXPORT_SYMBOL(__skb_gro_checksum_complete);
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
12static DEFINE_SPINLOCK(offload_lock);
13static 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 */
15int 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 */
29void 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}
41EXPORT_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 */
56static 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);
71out:
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 */
87void dev_remove_offload(struct packet_offload *po)
88{
89 __dev_remove_offload(po);
90
91 synchronize_net();
92}
93EXPORT_SYMBOL(dev_remove_offload);
94
95/**
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 */
101struct 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}
118EXPORT_SYMBOL(skb_eth_gso_segment);
119
120/**
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 */
125struct 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}
151EXPORT_SYMBOL(skb_mac_gso_segment);
152
153int 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;
160 unsigned int gro_max_size;
161 unsigned int new_truesize;
162 struct sk_buff *lp;
163 int segs;
164
165 /* Do not splice page pool based packets w/ non-page pool
166 * packets. This can result in reference count issues as page
167 * pool pages will not decrement the reference count and will
168 * instead be immediately returned to the pool or have frag
169 * count decremented.
170 */
171 if (p->pp_recycle != skb->pp_recycle)
172 return -ETOOMANYREFS;
173
174 /* pairs with WRITE_ONCE() in netif_set_gro_max_size() */
175 gro_max_size = READ_ONCE(p->dev->gro_max_size);
176
177 if (unlikely(p->len + len >= gro_max_size || NAPI_GRO_CB(skb)->flush))
178 return -E2BIG;
179
180 if (unlikely(p->len + len >= GRO_LEGACY_MAX_SIZE)) {
181 if (p->protocol != htons(ETH_P_IPV6) ||
182 skb_headroom(p) < sizeof(struct hop_jumbo_hdr) ||
183 ipv6_hdr(p)->nexthdr != IPPROTO_TCP ||
184 p->encapsulation)
185 return -E2BIG;
186 }
187
188 segs = NAPI_GRO_CB(skb)->count;
189 lp = NAPI_GRO_CB(p)->last;
190 pinfo = skb_shinfo(lp);
191
192 if (headlen <= offset) {
193 skb_frag_t *frag;
194 skb_frag_t *frag2;
195 int i = skbinfo->nr_frags;
196 int nr_frags = pinfo->nr_frags + i;
197
198 if (nr_frags > MAX_SKB_FRAGS)
199 goto merge;
200
201 offset -= headlen;
202 pinfo->nr_frags = nr_frags;
203 skbinfo->nr_frags = 0;
204
205 frag = pinfo->frags + nr_frags;
206 frag2 = skbinfo->frags + i;
207 do {
208 *--frag = *--frag2;
209 } while (--i);
210
211 skb_frag_off_add(frag, offset);
212 skb_frag_size_sub(frag, offset);
213
214 /* all fragments truesize : remove (head size + sk_buff) */
215 new_truesize = SKB_TRUESIZE(skb_end_offset(skb));
216 delta_truesize = skb->truesize - new_truesize;
217
218 skb->truesize = new_truesize;
219 skb->len -= skb->data_len;
220 skb->data_len = 0;
221
222 NAPI_GRO_CB(skb)->free = NAPI_GRO_FREE;
223 goto done;
224 } else if (skb->head_frag) {
225 int nr_frags = pinfo->nr_frags;
226 skb_frag_t *frag = pinfo->frags + nr_frags;
227 struct page *page = virt_to_head_page(skb->head);
228 unsigned int first_size = headlen - offset;
229 unsigned int first_offset;
230
231 if (nr_frags + 1 + skbinfo->nr_frags > MAX_SKB_FRAGS)
232 goto merge;
233
234 first_offset = skb->data -
235 (unsigned char *)page_address(page) +
236 offset;
237
238 pinfo->nr_frags = nr_frags + 1 + skbinfo->nr_frags;
239
240 __skb_frag_set_page(frag, page);
241 skb_frag_off_set(frag, first_offset);
242 skb_frag_size_set(frag, first_size);
243
244 memcpy(frag + 1, skbinfo->frags, sizeof(*frag) * skbinfo->nr_frags);
245 /* We dont need to clear skbinfo->nr_frags here */
246
247 new_truesize = SKB_DATA_ALIGN(sizeof(struct sk_buff));
248 delta_truesize = skb->truesize - new_truesize;
249 skb->truesize = new_truesize;
250 NAPI_GRO_CB(skb)->free = NAPI_GRO_FREE_STOLEN_HEAD;
251 goto done;
252 }
253
254merge:
255 /* sk owenrship - if any - completely transferred to the aggregated packet */
256 skb->destructor = NULL;
257 delta_truesize = skb->truesize;
258 if (offset > headlen) {
259 unsigned int eat = offset - headlen;
260
261 skb_frag_off_add(&skbinfo->frags[0], eat);
262 skb_frag_size_sub(&skbinfo->frags[0], eat);
263 skb->data_len -= eat;
264 skb->len -= eat;
265 offset = headlen;
266 }
267
268 __skb_pull(skb, offset);
269
270 if (NAPI_GRO_CB(p)->last == p)
271 skb_shinfo(p)->frag_list = skb;
272 else
273 NAPI_GRO_CB(p)->last->next = skb;
274 NAPI_GRO_CB(p)->last = skb;
275 __skb_header_release(skb);
276 lp = p;
277
278done:
279 NAPI_GRO_CB(p)->count += segs;
280 p->data_len += len;
281 p->truesize += delta_truesize;
282 p->len += len;
283 if (lp != p) {
284 lp->data_len += len;
285 lp->truesize += delta_truesize;
286 lp->len += len;
287 }
288 NAPI_GRO_CB(skb)->same_flow = 1;
289 return 0;
290}
291
292
293static void napi_gro_complete(struct napi_struct *napi, struct sk_buff *skb)
294{
295 struct packet_offload *ptype;
296 __be16 type = skb->protocol;
297 struct list_head *head = &offload_base;
298 int err = -ENOENT;
299
300 BUILD_BUG_ON(sizeof(struct napi_gro_cb) > sizeof(skb->cb));
301
302 if (NAPI_GRO_CB(skb)->count == 1) {
303 skb_shinfo(skb)->gso_size = 0;
304 goto out;
305 }
306
307 rcu_read_lock();
308 list_for_each_entry_rcu(ptype, head, list) {
309 if (ptype->type != type || !ptype->callbacks.gro_complete)
310 continue;
311
312 err = INDIRECT_CALL_INET(ptype->callbacks.gro_complete,
313 ipv6_gro_complete, inet_gro_complete,
314 skb, 0);
315 break;
316 }
317 rcu_read_unlock();
318
319 if (err) {
320 WARN_ON(&ptype->list == head);
321 kfree_skb(skb);
322 return;
323 }
324
325out:
326 gro_normal_one(napi, skb, NAPI_GRO_CB(skb)->count);
327}
328
329static void __napi_gro_flush_chain(struct napi_struct *napi, u32 index,
330 bool flush_old)
331{
332 struct list_head *head = &napi->gro_hash[index].list;
333 struct sk_buff *skb, *p;
334
335 list_for_each_entry_safe_reverse(skb, p, head, list) {
336 if (flush_old && NAPI_GRO_CB(skb)->age == jiffies)
337 return;
338 skb_list_del_init(skb);
339 napi_gro_complete(napi, skb);
340 napi->gro_hash[index].count--;
341 }
342
343 if (!napi->gro_hash[index].count)
344 __clear_bit(index, &napi->gro_bitmask);
345}
346
347/* napi->gro_hash[].list contains packets ordered by age.
348 * youngest packets at the head of it.
349 * Complete skbs in reverse order to reduce latencies.
350 */
351void napi_gro_flush(struct napi_struct *napi, bool flush_old)
352{
353 unsigned long bitmask = napi->gro_bitmask;
354 unsigned int i, base = ~0U;
355
356 while ((i = ffs(bitmask)) != 0) {
357 bitmask >>= i;
358 base += i;
359 __napi_gro_flush_chain(napi, base, flush_old);
360 }
361}
362EXPORT_SYMBOL(napi_gro_flush);
363
364static void gro_list_prepare(const struct list_head *head,
365 const struct sk_buff *skb)
366{
367 unsigned int maclen = skb->dev->hard_header_len;
368 u32 hash = skb_get_hash_raw(skb);
369 struct sk_buff *p;
370
371 list_for_each_entry(p, head, list) {
372 unsigned long diffs;
373
374 NAPI_GRO_CB(p)->flush = 0;
375
376 if (hash != skb_get_hash_raw(p)) {
377 NAPI_GRO_CB(p)->same_flow = 0;
378 continue;
379 }
380
381 diffs = (unsigned long)p->dev ^ (unsigned long)skb->dev;
382 diffs |= p->vlan_all ^ skb->vlan_all;
383 diffs |= skb_metadata_differs(p, skb);
384 if (maclen == ETH_HLEN)
385 diffs |= compare_ether_header(skb_mac_header(p),
386 skb_mac_header(skb));
387 else if (!diffs)
388 diffs = memcmp(skb_mac_header(p),
389 skb_mac_header(skb),
390 maclen);
391
392 /* in most common scenarions 'slow_gro' is 0
393 * otherwise we are already on some slower paths
394 * either skip all the infrequent tests altogether or
395 * avoid trying too hard to skip each of them individually
396 */
397 if (!diffs && unlikely(skb->slow_gro | p->slow_gro)) {
398#if IS_ENABLED(CONFIG_SKB_EXTENSIONS) && IS_ENABLED(CONFIG_NET_TC_SKB_EXT)
399 struct tc_skb_ext *skb_ext;
400 struct tc_skb_ext *p_ext;
401#endif
402
403 diffs |= p->sk != skb->sk;
404 diffs |= skb_metadata_dst_cmp(p, skb);
405 diffs |= skb_get_nfct(p) ^ skb_get_nfct(skb);
406
407#if IS_ENABLED(CONFIG_SKB_EXTENSIONS) && IS_ENABLED(CONFIG_NET_TC_SKB_EXT)
408 skb_ext = skb_ext_find(skb, TC_SKB_EXT);
409 p_ext = skb_ext_find(p, TC_SKB_EXT);
410
411 diffs |= (!!p_ext) ^ (!!skb_ext);
412 if (!diffs && unlikely(skb_ext))
413 diffs |= p_ext->chain ^ skb_ext->chain;
414#endif
415 }
416
417 NAPI_GRO_CB(p)->same_flow = !diffs;
418 }
419}
420
421static inline void skb_gro_reset_offset(struct sk_buff *skb, u32 nhoff)
422{
423 const struct skb_shared_info *pinfo = skb_shinfo(skb);
424 const skb_frag_t *frag0 = &pinfo->frags[0];
425
426 NAPI_GRO_CB(skb)->data_offset = 0;
427 NAPI_GRO_CB(skb)->frag0 = NULL;
428 NAPI_GRO_CB(skb)->frag0_len = 0;
429
430 if (!skb_headlen(skb) && pinfo->nr_frags &&
431 !PageHighMem(skb_frag_page(frag0)) &&
432 (!NET_IP_ALIGN || !((skb_frag_off(frag0) + nhoff) & 3))) {
433 NAPI_GRO_CB(skb)->frag0 = skb_frag_address(frag0);
434 NAPI_GRO_CB(skb)->frag0_len = min_t(unsigned int,
435 skb_frag_size(frag0),
436 skb->end - skb->tail);
437 }
438}
439
440static void gro_pull_from_frag0(struct sk_buff *skb, int grow)
441{
442 struct skb_shared_info *pinfo = skb_shinfo(skb);
443
444 BUG_ON(skb->end - skb->tail < grow);
445
446 memcpy(skb_tail_pointer(skb), NAPI_GRO_CB(skb)->frag0, grow);
447
448 skb->data_len -= grow;
449 skb->tail += grow;
450
451 skb_frag_off_add(&pinfo->frags[0], grow);
452 skb_frag_size_sub(&pinfo->frags[0], grow);
453
454 if (unlikely(!skb_frag_size(&pinfo->frags[0]))) {
455 skb_frag_unref(skb, 0);
456 memmove(pinfo->frags, pinfo->frags + 1,
457 --pinfo->nr_frags * sizeof(pinfo->frags[0]));
458 }
459}
460
461static void gro_flush_oldest(struct napi_struct *napi, struct list_head *head)
462{
463 struct sk_buff *oldest;
464
465 oldest = list_last_entry(head, struct sk_buff, list);
466
467 /* We are called with head length >= MAX_GRO_SKBS, so this is
468 * impossible.
469 */
470 if (WARN_ON_ONCE(!oldest))
471 return;
472
473 /* Do not adjust napi->gro_hash[].count, caller is adding a new
474 * SKB to the chain.
475 */
476 skb_list_del_init(oldest);
477 napi_gro_complete(napi, oldest);
478}
479
480static enum gro_result dev_gro_receive(struct napi_struct *napi, struct sk_buff *skb)
481{
482 u32 bucket = skb_get_hash_raw(skb) & (GRO_HASH_BUCKETS - 1);
483 struct gro_list *gro_list = &napi->gro_hash[bucket];
484 struct list_head *head = &offload_base;
485 struct packet_offload *ptype;
486 __be16 type = skb->protocol;
487 struct sk_buff *pp = NULL;
488 enum gro_result ret;
489 int same_flow;
490 int grow;
491
492 if (netif_elide_gro(skb->dev))
493 goto normal;
494
495 gro_list_prepare(&gro_list->list, skb);
496
497 rcu_read_lock();
498 list_for_each_entry_rcu(ptype, head, list) {
499 if (ptype->type == type && ptype->callbacks.gro_receive)
500 goto found_ptype;
501 }
502 rcu_read_unlock();
503 goto normal;
504
505found_ptype:
506 skb_set_network_header(skb, skb_gro_offset(skb));
507 skb_reset_mac_len(skb);
508 BUILD_BUG_ON(sizeof_field(struct napi_gro_cb, zeroed) != sizeof(u32));
509 BUILD_BUG_ON(!IS_ALIGNED(offsetof(struct napi_gro_cb, zeroed),
510 sizeof(u32))); /* Avoid slow unaligned acc */
511 *(u32 *)&NAPI_GRO_CB(skb)->zeroed = 0;
512 NAPI_GRO_CB(skb)->flush = skb_has_frag_list(skb);
513 NAPI_GRO_CB(skb)->is_atomic = 1;
514 NAPI_GRO_CB(skb)->count = 1;
515 if (unlikely(skb_is_gso(skb))) {
516 NAPI_GRO_CB(skb)->count = skb_shinfo(skb)->gso_segs;
517 /* Only support TCP and non DODGY users. */
518 if (!skb_is_gso_tcp(skb) ||
519 (skb_shinfo(skb)->gso_type & SKB_GSO_DODGY))
520 NAPI_GRO_CB(skb)->flush = 1;
521 }
522
523 /* Setup for GRO checksum validation */
524 switch (skb->ip_summed) {
525 case CHECKSUM_COMPLETE:
526 NAPI_GRO_CB(skb)->csum = skb->csum;
527 NAPI_GRO_CB(skb)->csum_valid = 1;
528 break;
529 case CHECKSUM_UNNECESSARY:
530 NAPI_GRO_CB(skb)->csum_cnt = skb->csum_level + 1;
531 break;
532 }
533
534 pp = INDIRECT_CALL_INET(ptype->callbacks.gro_receive,
535 ipv6_gro_receive, inet_gro_receive,
536 &gro_list->list, skb);
537
538 rcu_read_unlock();
539
540 if (PTR_ERR(pp) == -EINPROGRESS) {
541 ret = GRO_CONSUMED;
542 goto ok;
543 }
544
545 same_flow = NAPI_GRO_CB(skb)->same_flow;
546 ret = NAPI_GRO_CB(skb)->free ? GRO_MERGED_FREE : GRO_MERGED;
547
548 if (pp) {
549 skb_list_del_init(pp);
550 napi_gro_complete(napi, pp);
551 gro_list->count--;
552 }
553
554 if (same_flow)
555 goto ok;
556
557 if (NAPI_GRO_CB(skb)->flush)
558 goto normal;
559
560 if (unlikely(gro_list->count >= MAX_GRO_SKBS))
561 gro_flush_oldest(napi, &gro_list->list);
562 else
563 gro_list->count++;
564
565 NAPI_GRO_CB(skb)->age = jiffies;
566 NAPI_GRO_CB(skb)->last = skb;
567 if (!skb_is_gso(skb))
568 skb_shinfo(skb)->gso_size = skb_gro_len(skb);
569 list_add(&skb->list, &gro_list->list);
570 ret = GRO_HELD;
571
572pull:
573 grow = skb_gro_offset(skb) - skb_headlen(skb);
574 if (grow > 0)
575 gro_pull_from_frag0(skb, grow);
576ok:
577 if (gro_list->count) {
578 if (!test_bit(bucket, &napi->gro_bitmask))
579 __set_bit(bucket, &napi->gro_bitmask);
580 } else if (test_bit(bucket, &napi->gro_bitmask)) {
581 __clear_bit(bucket, &napi->gro_bitmask);
582 }
583
584 return ret;
585
586normal:
587 ret = GRO_NORMAL;
588 goto pull;
589}
590
591struct packet_offload *gro_find_receive_by_type(__be16 type)
592{
593 struct list_head *offload_head = &offload_base;
594 struct packet_offload *ptype;
595
596 list_for_each_entry_rcu(ptype, offload_head, list) {
597 if (ptype->type != type || !ptype->callbacks.gro_receive)
598 continue;
599 return ptype;
600 }
601 return NULL;
602}
603EXPORT_SYMBOL(gro_find_receive_by_type);
604
605struct packet_offload *gro_find_complete_by_type(__be16 type)
606{
607 struct list_head *offload_head = &offload_base;
608 struct packet_offload *ptype;
609
610 list_for_each_entry_rcu(ptype, offload_head, list) {
611 if (ptype->type != type || !ptype->callbacks.gro_complete)
612 continue;
613 return ptype;
614 }
615 return NULL;
616}
617EXPORT_SYMBOL(gro_find_complete_by_type);
618
619static gro_result_t napi_skb_finish(struct napi_struct *napi,
620 struct sk_buff *skb,
621 gro_result_t ret)
622{
623 switch (ret) {
624 case GRO_NORMAL:
625 gro_normal_one(napi, skb, 1);
626 break;
627
628 case GRO_MERGED_FREE:
629 if (NAPI_GRO_CB(skb)->free == NAPI_GRO_FREE_STOLEN_HEAD)
630 napi_skb_free_stolen_head(skb);
631 else if (skb->fclone != SKB_FCLONE_UNAVAILABLE)
632 __kfree_skb(skb);
633 else
634 __kfree_skb_defer(skb);
635 break;
636
637 case GRO_HELD:
638 case GRO_MERGED:
639 case GRO_CONSUMED:
640 break;
641 }
642
643 return ret;
644}
645
646gro_result_t napi_gro_receive(struct napi_struct *napi, struct sk_buff *skb)
647{
648 gro_result_t ret;
649
650 skb_mark_napi_id(skb, napi);
651 trace_napi_gro_receive_entry(skb);
652
653 skb_gro_reset_offset(skb, 0);
654
655 ret = napi_skb_finish(napi, skb, dev_gro_receive(napi, skb));
656 trace_napi_gro_receive_exit(ret);
657
658 return ret;
659}
660EXPORT_SYMBOL(napi_gro_receive);
661
662static void napi_reuse_skb(struct napi_struct *napi, struct sk_buff *skb)
663{
664 if (unlikely(skb->pfmemalloc)) {
665 consume_skb(skb);
666 return;
667 }
668 __skb_pull(skb, skb_headlen(skb));
669 /* restore the reserve we had after netdev_alloc_skb_ip_align() */
670 skb_reserve(skb, NET_SKB_PAD + NET_IP_ALIGN - skb_headroom(skb));
671 __vlan_hwaccel_clear_tag(skb);
672 skb->dev = napi->dev;
673 skb->skb_iif = 0;
674
675 /* eth_type_trans() assumes pkt_type is PACKET_HOST */
676 skb->pkt_type = PACKET_HOST;
677
678 skb->encapsulation = 0;
679 skb_shinfo(skb)->gso_type = 0;
680 skb_shinfo(skb)->gso_size = 0;
681 if (unlikely(skb->slow_gro)) {
682 skb_orphan(skb);
683 skb_ext_reset(skb);
684 nf_reset_ct(skb);
685 skb->slow_gro = 0;
686 }
687
688 napi->skb = skb;
689}
690
691struct sk_buff *napi_get_frags(struct napi_struct *napi)
692{
693 struct sk_buff *skb = napi->skb;
694
695 if (!skb) {
696 skb = napi_alloc_skb(napi, GRO_MAX_HEAD);
697 if (skb) {
698 napi->skb = skb;
699 skb_mark_napi_id(skb, napi);
700 }
701 }
702 return skb;
703}
704EXPORT_SYMBOL(napi_get_frags);
705
706static gro_result_t napi_frags_finish(struct napi_struct *napi,
707 struct sk_buff *skb,
708 gro_result_t ret)
709{
710 switch (ret) {
711 case GRO_NORMAL:
712 case GRO_HELD:
713 __skb_push(skb, ETH_HLEN);
714 skb->protocol = eth_type_trans(skb, skb->dev);
715 if (ret == GRO_NORMAL)
716 gro_normal_one(napi, skb, 1);
717 break;
718
719 case GRO_MERGED_FREE:
720 if (NAPI_GRO_CB(skb)->free == NAPI_GRO_FREE_STOLEN_HEAD)
721 napi_skb_free_stolen_head(skb);
722 else
723 napi_reuse_skb(napi, skb);
724 break;
725
726 case GRO_MERGED:
727 case GRO_CONSUMED:
728 break;
729 }
730
731 return ret;
732}
733
734/* Upper GRO stack assumes network header starts at gro_offset=0
735 * Drivers could call both napi_gro_frags() and napi_gro_receive()
736 * We copy ethernet header into skb->data to have a common layout.
737 */
738static struct sk_buff *napi_frags_skb(struct napi_struct *napi)
739{
740 struct sk_buff *skb = napi->skb;
741 const struct ethhdr *eth;
742 unsigned int hlen = sizeof(*eth);
743
744 napi->skb = NULL;
745
746 skb_reset_mac_header(skb);
747 skb_gro_reset_offset(skb, hlen);
748
749 if (unlikely(skb_gro_header_hard(skb, hlen))) {
750 eth = skb_gro_header_slow(skb, hlen, 0);
751 if (unlikely(!eth)) {
752 net_warn_ratelimited("%s: dropping impossible skb from %s\n",
753 __func__, napi->dev->name);
754 napi_reuse_skb(napi, skb);
755 return NULL;
756 }
757 } else {
758 eth = (const struct ethhdr *)skb->data;
759 gro_pull_from_frag0(skb, hlen);
760 NAPI_GRO_CB(skb)->frag0 += hlen;
761 NAPI_GRO_CB(skb)->frag0_len -= hlen;
762 }
763 __skb_pull(skb, hlen);
764
765 /*
766 * This works because the only protocols we care about don't require
767 * special handling.
768 * We'll fix it up properly in napi_frags_finish()
769 */
770 skb->protocol = eth->h_proto;
771
772 return skb;
773}
774
775gro_result_t napi_gro_frags(struct napi_struct *napi)
776{
777 gro_result_t ret;
778 struct sk_buff *skb = napi_frags_skb(napi);
779
780 trace_napi_gro_frags_entry(skb);
781
782 ret = napi_frags_finish(napi, skb, dev_gro_receive(napi, skb));
783 trace_napi_gro_frags_exit(ret);
784
785 return ret;
786}
787EXPORT_SYMBOL(napi_gro_frags);
788
789/* Compute the checksum from gro_offset and return the folded value
790 * after adding in any pseudo checksum.
791 */
792__sum16 __skb_gro_checksum_complete(struct sk_buff *skb)
793{
794 __wsum wsum;
795 __sum16 sum;
796
797 wsum = skb_checksum(skb, skb_gro_offset(skb), skb_gro_len(skb), 0);
798
799 /* NAPI_GRO_CB(skb)->csum holds pseudo checksum */
800 sum = csum_fold(csum_add(NAPI_GRO_CB(skb)->csum, wsum));
801 /* See comments in __skb_checksum_complete(). */
802 if (likely(!sum)) {
803 if (unlikely(skb->ip_summed == CHECKSUM_COMPLETE) &&
804 !skb->csum_complete_sw)
805 netdev_rx_csum_fault(skb->dev, skb);
806 }
807
808 NAPI_GRO_CB(skb)->csum = wsum;
809 NAPI_GRO_CB(skb)->csum_valid = 1;
810
811 return sum;
812}
813EXPORT_SYMBOL(__skb_gro_checksum_complete);