xref: /openbmc/linux/net/core/gro.c (revision c699ce1a)
1 // SPDX-License-Identifier: GPL-2.0-or-later
2 #include <net/gro.h>
3 #include <net/dst_metadata.h>
4 #include <net/busy_poll.h>
5 #include <trace/events/net.h>
6 
7 #define MAX_GRO_SKBS 8
8 
9 /* This should be increased if a protocol with a bigger head is added. */
10 #define GRO_MAX_HEAD (MAX_HEADER + 128)
11 
12 static DEFINE_SPINLOCK(offload_lock);
13 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 /**
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 /**
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);
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;
160 	unsigned int gro_max_size;
161 	unsigned int new_truesize;
162 	struct sk_buff *lp;
163 	int segs;
164 
165 	/* pairs with WRITE_ONCE() in netif_set_gro_max_size() */
166 	gro_max_size = READ_ONCE(p->dev->gro_max_size);
167 
168 	if (unlikely(p->len + len >= gro_max_size || NAPI_GRO_CB(skb)->flush))
169 		return -E2BIG;
170 
171 	if (unlikely(p->len + len >= GRO_LEGACY_MAX_SIZE)) {
172 		if (p->protocol != htons(ETH_P_IPV6) ||
173 		    skb_headroom(p) < sizeof(struct hop_jumbo_hdr) ||
174 		    ipv6_hdr(p)->nexthdr != IPPROTO_TCP ||
175 		    p->encapsulation)
176 			return -E2BIG;
177 	}
178 
179 	segs = NAPI_GRO_CB(skb)->count;
180 	lp = NAPI_GRO_CB(p)->last;
181 	pinfo = skb_shinfo(lp);
182 
183 	if (headlen <= offset) {
184 		skb_frag_t *frag;
185 		skb_frag_t *frag2;
186 		int i = skbinfo->nr_frags;
187 		int nr_frags = pinfo->nr_frags + i;
188 
189 		if (nr_frags > MAX_SKB_FRAGS)
190 			goto merge;
191 
192 		offset -= headlen;
193 		pinfo->nr_frags = nr_frags;
194 		skbinfo->nr_frags = 0;
195 
196 		frag = pinfo->frags + nr_frags;
197 		frag2 = skbinfo->frags + i;
198 		do {
199 			*--frag = *--frag2;
200 		} while (--i);
201 
202 		skb_frag_off_add(frag, offset);
203 		skb_frag_size_sub(frag, offset);
204 
205 		/* all fragments truesize : remove (head size + sk_buff) */
206 		new_truesize = SKB_TRUESIZE(skb_end_offset(skb));
207 		delta_truesize = skb->truesize - new_truesize;
208 
209 		skb->truesize = new_truesize;
210 		skb->len -= skb->data_len;
211 		skb->data_len = 0;
212 
213 		NAPI_GRO_CB(skb)->free = NAPI_GRO_FREE;
214 		goto done;
215 	} else if (skb->head_frag) {
216 		int nr_frags = pinfo->nr_frags;
217 		skb_frag_t *frag = pinfo->frags + nr_frags;
218 		struct page *page = virt_to_head_page(skb->head);
219 		unsigned int first_size = headlen - offset;
220 		unsigned int first_offset;
221 
222 		if (nr_frags + 1 + skbinfo->nr_frags > MAX_SKB_FRAGS)
223 			goto merge;
224 
225 		first_offset = skb->data -
226 			       (unsigned char *)page_address(page) +
227 			       offset;
228 
229 		pinfo->nr_frags = nr_frags + 1 + skbinfo->nr_frags;
230 
231 		__skb_frag_set_page(frag, page);
232 		skb_frag_off_set(frag, first_offset);
233 		skb_frag_size_set(frag, first_size);
234 
235 		memcpy(frag + 1, skbinfo->frags, sizeof(*frag) * skbinfo->nr_frags);
236 		/* We dont need to clear skbinfo->nr_frags here */
237 
238 		new_truesize = SKB_DATA_ALIGN(sizeof(struct sk_buff));
239 		delta_truesize = skb->truesize - new_truesize;
240 		skb->truesize = new_truesize;
241 		NAPI_GRO_CB(skb)->free = NAPI_GRO_FREE_STOLEN_HEAD;
242 		goto done;
243 	}
244 
245 merge:
246 	/* sk owenrship - if any - completely transferred to the aggregated packet */
247 	skb->destructor = NULL;
248 	delta_truesize = skb->truesize;
249 	if (offset > headlen) {
250 		unsigned int eat = offset - headlen;
251 
252 		skb_frag_off_add(&skbinfo->frags[0], eat);
253 		skb_frag_size_sub(&skbinfo->frags[0], eat);
254 		skb->data_len -= eat;
255 		skb->len -= eat;
256 		offset = headlen;
257 	}
258 
259 	__skb_pull(skb, offset);
260 
261 	if (NAPI_GRO_CB(p)->last == p)
262 		skb_shinfo(p)->frag_list = skb;
263 	else
264 		NAPI_GRO_CB(p)->last->next = skb;
265 	NAPI_GRO_CB(p)->last = skb;
266 	__skb_header_release(skb);
267 	lp = p;
268 
269 done:
270 	NAPI_GRO_CB(p)->count += segs;
271 	p->data_len += len;
272 	p->truesize += delta_truesize;
273 	p->len += len;
274 	if (lp != p) {
275 		lp->data_len += len;
276 		lp->truesize += delta_truesize;
277 		lp->len += len;
278 	}
279 	NAPI_GRO_CB(skb)->same_flow = 1;
280 	return 0;
281 }
282 
283 
284 static void napi_gro_complete(struct napi_struct *napi, struct sk_buff *skb)
285 {
286 	struct packet_offload *ptype;
287 	__be16 type = skb->protocol;
288 	struct list_head *head = &offload_base;
289 	int err = -ENOENT;
290 
291 	BUILD_BUG_ON(sizeof(struct napi_gro_cb) > sizeof(skb->cb));
292 
293 	if (NAPI_GRO_CB(skb)->count == 1) {
294 		skb_shinfo(skb)->gso_size = 0;
295 		goto out;
296 	}
297 
298 	rcu_read_lock();
299 	list_for_each_entry_rcu(ptype, head, list) {
300 		if (ptype->type != type || !ptype->callbacks.gro_complete)
301 			continue;
302 
303 		err = INDIRECT_CALL_INET(ptype->callbacks.gro_complete,
304 					 ipv6_gro_complete, inet_gro_complete,
305 					 skb, 0);
306 		break;
307 	}
308 	rcu_read_unlock();
309 
310 	if (err) {
311 		WARN_ON(&ptype->list == head);
312 		kfree_skb(skb);
313 		return;
314 	}
315 
316 out:
317 	gro_normal_one(napi, skb, NAPI_GRO_CB(skb)->count);
318 }
319 
320 static void __napi_gro_flush_chain(struct napi_struct *napi, u32 index,
321 				   bool flush_old)
322 {
323 	struct list_head *head = &napi->gro_hash[index].list;
324 	struct sk_buff *skb, *p;
325 
326 	list_for_each_entry_safe_reverse(skb, p, head, list) {
327 		if (flush_old && NAPI_GRO_CB(skb)->age == jiffies)
328 			return;
329 		skb_list_del_init(skb);
330 		napi_gro_complete(napi, skb);
331 		napi->gro_hash[index].count--;
332 	}
333 
334 	if (!napi->gro_hash[index].count)
335 		__clear_bit(index, &napi->gro_bitmask);
336 }
337 
338 /* napi->gro_hash[].list contains packets ordered by age.
339  * youngest packets at the head of it.
340  * Complete skbs in reverse order to reduce latencies.
341  */
342 void napi_gro_flush(struct napi_struct *napi, bool flush_old)
343 {
344 	unsigned long bitmask = napi->gro_bitmask;
345 	unsigned int i, base = ~0U;
346 
347 	while ((i = ffs(bitmask)) != 0) {
348 		bitmask >>= i;
349 		base += i;
350 		__napi_gro_flush_chain(napi, base, flush_old);
351 	}
352 }
353 EXPORT_SYMBOL(napi_gro_flush);
354 
355 static void gro_list_prepare(const struct list_head *head,
356 			     const struct sk_buff *skb)
357 {
358 	unsigned int maclen = skb->dev->hard_header_len;
359 	u32 hash = skb_get_hash_raw(skb);
360 	struct sk_buff *p;
361 
362 	list_for_each_entry(p, head, list) {
363 		unsigned long diffs;
364 
365 		NAPI_GRO_CB(p)->flush = 0;
366 
367 		if (hash != skb_get_hash_raw(p)) {
368 			NAPI_GRO_CB(p)->same_flow = 0;
369 			continue;
370 		}
371 
372 		diffs = (unsigned long)p->dev ^ (unsigned long)skb->dev;
373 		diffs |= p->vlan_all ^ skb->vlan_all;
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 			goto found_ptype;
492 	}
493 	rcu_read_unlock();
494 	goto normal;
495 
496 found_ptype:
497 	skb_set_network_header(skb, skb_gro_offset(skb));
498 	skb_reset_mac_len(skb);
499 	BUILD_BUG_ON(sizeof_field(struct napi_gro_cb, zeroed) != sizeof(u32));
500 	BUILD_BUG_ON(!IS_ALIGNED(offsetof(struct napi_gro_cb, zeroed),
501 					sizeof(u32))); /* Avoid slow unaligned acc */
502 	*(u32 *)&NAPI_GRO_CB(skb)->zeroed = 0;
503 	NAPI_GRO_CB(skb)->flush = skb_has_frag_list(skb);
504 	NAPI_GRO_CB(skb)->is_atomic = 1;
505 	NAPI_GRO_CB(skb)->count = 1;
506 	if (unlikely(skb_is_gso(skb))) {
507 		NAPI_GRO_CB(skb)->count = skb_shinfo(skb)->gso_segs;
508 		/* Only support TCP at the moment. */
509 		if (!skb_is_gso_tcp(skb))
510 			NAPI_GRO_CB(skb)->flush = 1;
511 	}
512 
513 	/* Setup for GRO checksum validation */
514 	switch (skb->ip_summed) {
515 	case CHECKSUM_COMPLETE:
516 		NAPI_GRO_CB(skb)->csum = skb->csum;
517 		NAPI_GRO_CB(skb)->csum_valid = 1;
518 		break;
519 	case CHECKSUM_UNNECESSARY:
520 		NAPI_GRO_CB(skb)->csum_cnt = skb->csum_level + 1;
521 		break;
522 	}
523 
524 	pp = INDIRECT_CALL_INET(ptype->callbacks.gro_receive,
525 				ipv6_gro_receive, inet_gro_receive,
526 				&gro_list->list, skb);
527 
528 	rcu_read_unlock();
529 
530 	if (PTR_ERR(pp) == -EINPROGRESS) {
531 		ret = GRO_CONSUMED;
532 		goto ok;
533 	}
534 
535 	same_flow = NAPI_GRO_CB(skb)->same_flow;
536 	ret = NAPI_GRO_CB(skb)->free ? GRO_MERGED_FREE : GRO_MERGED;
537 
538 	if (pp) {
539 		skb_list_del_init(pp);
540 		napi_gro_complete(napi, pp);
541 		gro_list->count--;
542 	}
543 
544 	if (same_flow)
545 		goto ok;
546 
547 	if (NAPI_GRO_CB(skb)->flush)
548 		goto normal;
549 
550 	if (unlikely(gro_list->count >= MAX_GRO_SKBS))
551 		gro_flush_oldest(napi, &gro_list->list);
552 	else
553 		gro_list->count++;
554 
555 	NAPI_GRO_CB(skb)->age = jiffies;
556 	NAPI_GRO_CB(skb)->last = skb;
557 	if (!skb_is_gso(skb))
558 		skb_shinfo(skb)->gso_size = skb_gro_len(skb);
559 	list_add(&skb->list, &gro_list->list);
560 	ret = GRO_HELD;
561 
562 pull:
563 	grow = skb_gro_offset(skb) - skb_headlen(skb);
564 	if (grow > 0)
565 		gro_pull_from_frag0(skb, grow);
566 ok:
567 	if (gro_list->count) {
568 		if (!test_bit(bucket, &napi->gro_bitmask))
569 			__set_bit(bucket, &napi->gro_bitmask);
570 	} else if (test_bit(bucket, &napi->gro_bitmask)) {
571 		__clear_bit(bucket, &napi->gro_bitmask);
572 	}
573 
574 	return ret;
575 
576 normal:
577 	ret = GRO_NORMAL;
578 	goto pull;
579 }
580 
581 struct packet_offload *gro_find_receive_by_type(__be16 type)
582 {
583 	struct list_head *offload_head = &offload_base;
584 	struct packet_offload *ptype;
585 
586 	list_for_each_entry_rcu(ptype, offload_head, list) {
587 		if (ptype->type != type || !ptype->callbacks.gro_receive)
588 			continue;
589 		return ptype;
590 	}
591 	return NULL;
592 }
593 EXPORT_SYMBOL(gro_find_receive_by_type);
594 
595 struct packet_offload *gro_find_complete_by_type(__be16 type)
596 {
597 	struct list_head *offload_head = &offload_base;
598 	struct packet_offload *ptype;
599 
600 	list_for_each_entry_rcu(ptype, offload_head, list) {
601 		if (ptype->type != type || !ptype->callbacks.gro_complete)
602 			continue;
603 		return ptype;
604 	}
605 	return NULL;
606 }
607 EXPORT_SYMBOL(gro_find_complete_by_type);
608 
609 static gro_result_t napi_skb_finish(struct napi_struct *napi,
610 				    struct sk_buff *skb,
611 				    gro_result_t ret)
612 {
613 	switch (ret) {
614 	case GRO_NORMAL:
615 		gro_normal_one(napi, skb, 1);
616 		break;
617 
618 	case GRO_MERGED_FREE:
619 		if (NAPI_GRO_CB(skb)->free == NAPI_GRO_FREE_STOLEN_HEAD)
620 			napi_skb_free_stolen_head(skb);
621 		else if (skb->fclone != SKB_FCLONE_UNAVAILABLE)
622 			__kfree_skb(skb);
623 		else
624 			__kfree_skb_defer(skb);
625 		break;
626 
627 	case GRO_HELD:
628 	case GRO_MERGED:
629 	case GRO_CONSUMED:
630 		break;
631 	}
632 
633 	return ret;
634 }
635 
636 gro_result_t napi_gro_receive(struct napi_struct *napi, struct sk_buff *skb)
637 {
638 	gro_result_t ret;
639 
640 	skb_mark_napi_id(skb, napi);
641 	trace_napi_gro_receive_entry(skb);
642 
643 	skb_gro_reset_offset(skb, 0);
644 
645 	ret = napi_skb_finish(napi, skb, dev_gro_receive(napi, skb));
646 	trace_napi_gro_receive_exit(ret);
647 
648 	return ret;
649 }
650 EXPORT_SYMBOL(napi_gro_receive);
651 
652 static void napi_reuse_skb(struct napi_struct *napi, struct sk_buff *skb)
653 {
654 	if (unlikely(skb->pfmemalloc)) {
655 		consume_skb(skb);
656 		return;
657 	}
658 	__skb_pull(skb, skb_headlen(skb));
659 	/* restore the reserve we had after netdev_alloc_skb_ip_align() */
660 	skb_reserve(skb, NET_SKB_PAD + NET_IP_ALIGN - skb_headroom(skb));
661 	__vlan_hwaccel_clear_tag(skb);
662 	skb->dev = napi->dev;
663 	skb->skb_iif = 0;
664 
665 	/* eth_type_trans() assumes pkt_type is PACKET_HOST */
666 	skb->pkt_type = PACKET_HOST;
667 
668 	skb->encapsulation = 0;
669 	skb_shinfo(skb)->gso_type = 0;
670 	skb_shinfo(skb)->gso_size = 0;
671 	if (unlikely(skb->slow_gro)) {
672 		skb_orphan(skb);
673 		skb_ext_reset(skb);
674 		nf_reset_ct(skb);
675 		skb->slow_gro = 0;
676 	}
677 
678 	napi->skb = skb;
679 }
680 
681 struct sk_buff *napi_get_frags(struct napi_struct *napi)
682 {
683 	struct sk_buff *skb = napi->skb;
684 
685 	if (!skb) {
686 		skb = napi_alloc_skb(napi, GRO_MAX_HEAD);
687 		if (skb) {
688 			napi->skb = skb;
689 			skb_mark_napi_id(skb, napi);
690 		}
691 	}
692 	return skb;
693 }
694 EXPORT_SYMBOL(napi_get_frags);
695 
696 static gro_result_t napi_frags_finish(struct napi_struct *napi,
697 				      struct sk_buff *skb,
698 				      gro_result_t ret)
699 {
700 	switch (ret) {
701 	case GRO_NORMAL:
702 	case GRO_HELD:
703 		__skb_push(skb, ETH_HLEN);
704 		skb->protocol = eth_type_trans(skb, skb->dev);
705 		if (ret == GRO_NORMAL)
706 			gro_normal_one(napi, skb, 1);
707 		break;
708 
709 	case GRO_MERGED_FREE:
710 		if (NAPI_GRO_CB(skb)->free == NAPI_GRO_FREE_STOLEN_HEAD)
711 			napi_skb_free_stolen_head(skb);
712 		else
713 			napi_reuse_skb(napi, skb);
714 		break;
715 
716 	case GRO_MERGED:
717 	case GRO_CONSUMED:
718 		break;
719 	}
720 
721 	return ret;
722 }
723 
724 /* Upper GRO stack assumes network header starts at gro_offset=0
725  * Drivers could call both napi_gro_frags() and napi_gro_receive()
726  * We copy ethernet header into skb->data to have a common layout.
727  */
728 static struct sk_buff *napi_frags_skb(struct napi_struct *napi)
729 {
730 	struct sk_buff *skb = napi->skb;
731 	const struct ethhdr *eth;
732 	unsigned int hlen = sizeof(*eth);
733 
734 	napi->skb = NULL;
735 
736 	skb_reset_mac_header(skb);
737 	skb_gro_reset_offset(skb, hlen);
738 
739 	if (unlikely(skb_gro_header_hard(skb, hlen))) {
740 		eth = skb_gro_header_slow(skb, hlen, 0);
741 		if (unlikely(!eth)) {
742 			net_warn_ratelimited("%s: dropping impossible skb from %s\n",
743 					     __func__, napi->dev->name);
744 			napi_reuse_skb(napi, skb);
745 			return NULL;
746 		}
747 	} else {
748 		eth = (const struct ethhdr *)skb->data;
749 		gro_pull_from_frag0(skb, hlen);
750 		NAPI_GRO_CB(skb)->frag0 += hlen;
751 		NAPI_GRO_CB(skb)->frag0_len -= hlen;
752 	}
753 	__skb_pull(skb, hlen);
754 
755 	/*
756 	 * This works because the only protocols we care about don't require
757 	 * special handling.
758 	 * We'll fix it up properly in napi_frags_finish()
759 	 */
760 	skb->protocol = eth->h_proto;
761 
762 	return skb;
763 }
764 
765 gro_result_t napi_gro_frags(struct napi_struct *napi)
766 {
767 	gro_result_t ret;
768 	struct sk_buff *skb = napi_frags_skb(napi);
769 
770 	trace_napi_gro_frags_entry(skb);
771 
772 	ret = napi_frags_finish(napi, skb, dev_gro_receive(napi, skb));
773 	trace_napi_gro_frags_exit(ret);
774 
775 	return ret;
776 }
777 EXPORT_SYMBOL(napi_gro_frags);
778 
779 /* Compute the checksum from gro_offset and return the folded value
780  * after adding in any pseudo checksum.
781  */
782 __sum16 __skb_gro_checksum_complete(struct sk_buff *skb)
783 {
784 	__wsum wsum;
785 	__sum16 sum;
786 
787 	wsum = skb_checksum(skb, skb_gro_offset(skb), skb_gro_len(skb), 0);
788 
789 	/* NAPI_GRO_CB(skb)->csum holds pseudo checksum */
790 	sum = csum_fold(csum_add(NAPI_GRO_CB(skb)->csum, wsum));
791 	/* See comments in __skb_checksum_complete(). */
792 	if (likely(!sum)) {
793 		if (unlikely(skb->ip_summed == CHECKSUM_COMPLETE) &&
794 		    !skb->csum_complete_sw)
795 			netdev_rx_csum_fault(skb->dev, skb);
796 	}
797 
798 	NAPI_GRO_CB(skb)->csum = wsum;
799 	NAPI_GRO_CB(skb)->csum_valid = 1;
800 
801 	return sum;
802 }
803 EXPORT_SYMBOL(__skb_gro_checksum_complete);
804