xref: /openbmc/linux/net/core/gro.c (revision a93fbb00)
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 
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))
168 		return -E2BIG;
169 
170 	lp = NAPI_GRO_CB(p)->last;
171 	pinfo = skb_shinfo(lp);
172 
173 	if (headlen <= offset) {
174 		skb_frag_t *frag;
175 		skb_frag_t *frag2;
176 		int i = skbinfo->nr_frags;
177 		int nr_frags = pinfo->nr_frags + i;
178 
179 		if (nr_frags > MAX_SKB_FRAGS)
180 			goto merge;
181 
182 		offset -= headlen;
183 		pinfo->nr_frags = nr_frags;
184 		skbinfo->nr_frags = 0;
185 
186 		frag = pinfo->frags + nr_frags;
187 		frag2 = skbinfo->frags + i;
188 		do {
189 			*--frag = *--frag2;
190 		} while (--i);
191 
192 		skb_frag_off_add(frag, offset);
193 		skb_frag_size_sub(frag, offset);
194 
195 		/* all fragments truesize : remove (head size + sk_buff) */
196 		new_truesize = SKB_TRUESIZE(skb_end_offset(skb));
197 		delta_truesize = skb->truesize - new_truesize;
198 
199 		skb->truesize = new_truesize;
200 		skb->len -= skb->data_len;
201 		skb->data_len = 0;
202 
203 		NAPI_GRO_CB(skb)->free = NAPI_GRO_FREE;
204 		goto done;
205 	} else if (skb->head_frag) {
206 		int nr_frags = pinfo->nr_frags;
207 		skb_frag_t *frag = pinfo->frags + nr_frags;
208 		struct page *page = virt_to_head_page(skb->head);
209 		unsigned int first_size = headlen - offset;
210 		unsigned int first_offset;
211 
212 		if (nr_frags + 1 + skbinfo->nr_frags > MAX_SKB_FRAGS)
213 			goto merge;
214 
215 		first_offset = skb->data -
216 			       (unsigned char *)page_address(page) +
217 			       offset;
218 
219 		pinfo->nr_frags = nr_frags + 1 + skbinfo->nr_frags;
220 
221 		__skb_frag_set_page(frag, page);
222 		skb_frag_off_set(frag, first_offset);
223 		skb_frag_size_set(frag, first_size);
224 
225 		memcpy(frag + 1, skbinfo->frags, sizeof(*frag) * skbinfo->nr_frags);
226 		/* We dont need to clear skbinfo->nr_frags here */
227 
228 		new_truesize = SKB_DATA_ALIGN(sizeof(struct sk_buff));
229 		delta_truesize = skb->truesize - new_truesize;
230 		skb->truesize = new_truesize;
231 		NAPI_GRO_CB(skb)->free = NAPI_GRO_FREE_STOLEN_HEAD;
232 		goto done;
233 	}
234 
235 merge:
236 	/* sk owenrship - if any - completely transferred to the aggregated packet */
237 	skb->destructor = NULL;
238 	delta_truesize = skb->truesize;
239 	if (offset > headlen) {
240 		unsigned int eat = offset - headlen;
241 
242 		skb_frag_off_add(&skbinfo->frags[0], eat);
243 		skb_frag_size_sub(&skbinfo->frags[0], eat);
244 		skb->data_len -= eat;
245 		skb->len -= eat;
246 		offset = headlen;
247 	}
248 
249 	__skb_pull(skb, offset);
250 
251 	if (NAPI_GRO_CB(p)->last == p)
252 		skb_shinfo(p)->frag_list = skb;
253 	else
254 		NAPI_GRO_CB(p)->last->next = skb;
255 	NAPI_GRO_CB(p)->last = skb;
256 	__skb_header_release(skb);
257 	lp = p;
258 
259 done:
260 	NAPI_GRO_CB(p)->count++;
261 	p->data_len += len;
262 	p->truesize += delta_truesize;
263 	p->len += len;
264 	if (lp != p) {
265 		lp->data_len += len;
266 		lp->truesize += delta_truesize;
267 		lp->len += len;
268 	}
269 	NAPI_GRO_CB(skb)->same_flow = 1;
270 	return 0;
271 }
272 
273 
274 static void napi_gro_complete(struct napi_struct *napi, struct sk_buff *skb)
275 {
276 	struct packet_offload *ptype;
277 	__be16 type = skb->protocol;
278 	struct list_head *head = &offload_base;
279 	int err = -ENOENT;
280 
281 	BUILD_BUG_ON(sizeof(struct napi_gro_cb) > sizeof(skb->cb));
282 
283 	if (NAPI_GRO_CB(skb)->count == 1) {
284 		skb_shinfo(skb)->gso_size = 0;
285 		goto out;
286 	}
287 
288 	rcu_read_lock();
289 	list_for_each_entry_rcu(ptype, head, list) {
290 		if (ptype->type != type || !ptype->callbacks.gro_complete)
291 			continue;
292 
293 		err = INDIRECT_CALL_INET(ptype->callbacks.gro_complete,
294 					 ipv6_gro_complete, inet_gro_complete,
295 					 skb, 0);
296 		break;
297 	}
298 	rcu_read_unlock();
299 
300 	if (err) {
301 		WARN_ON(&ptype->list == head);
302 		kfree_skb(skb);
303 		return;
304 	}
305 
306 out:
307 	gro_normal_one(napi, skb, NAPI_GRO_CB(skb)->count);
308 }
309 
310 static void __napi_gro_flush_chain(struct napi_struct *napi, u32 index,
311 				   bool flush_old)
312 {
313 	struct list_head *head = &napi->gro_hash[index].list;
314 	struct sk_buff *skb, *p;
315 
316 	list_for_each_entry_safe_reverse(skb, p, head, list) {
317 		if (flush_old && NAPI_GRO_CB(skb)->age == jiffies)
318 			return;
319 		skb_list_del_init(skb);
320 		napi_gro_complete(napi, skb);
321 		napi->gro_hash[index].count--;
322 	}
323 
324 	if (!napi->gro_hash[index].count)
325 		__clear_bit(index, &napi->gro_bitmask);
326 }
327 
328 /* napi->gro_hash[].list contains packets ordered by age.
329  * youngest packets at the head of it.
330  * Complete skbs in reverse order to reduce latencies.
331  */
332 void napi_gro_flush(struct napi_struct *napi, bool flush_old)
333 {
334 	unsigned long bitmask = napi->gro_bitmask;
335 	unsigned int i, base = ~0U;
336 
337 	while ((i = ffs(bitmask)) != 0) {
338 		bitmask >>= i;
339 		base += i;
340 		__napi_gro_flush_chain(napi, base, flush_old);
341 	}
342 }
343 EXPORT_SYMBOL(napi_gro_flush);
344 
345 static void gro_list_prepare(const struct list_head *head,
346 			     const struct sk_buff *skb)
347 {
348 	unsigned int maclen = skb->dev->hard_header_len;
349 	u32 hash = skb_get_hash_raw(skb);
350 	struct sk_buff *p;
351 
352 	list_for_each_entry(p, head, list) {
353 		unsigned long diffs;
354 
355 		NAPI_GRO_CB(p)->flush = 0;
356 
357 		if (hash != skb_get_hash_raw(p)) {
358 			NAPI_GRO_CB(p)->same_flow = 0;
359 			continue;
360 		}
361 
362 		diffs = (unsigned long)p->dev ^ (unsigned long)skb->dev;
363 		diffs |= skb_vlan_tag_present(p) ^ skb_vlan_tag_present(skb);
364 		if (skb_vlan_tag_present(p))
365 			diffs |= skb_vlan_tag_get(p) ^ skb_vlan_tag_get(skb);
366 		diffs |= skb_metadata_differs(p, skb);
367 		if (maclen == ETH_HLEN)
368 			diffs |= compare_ether_header(skb_mac_header(p),
369 						      skb_mac_header(skb));
370 		else if (!diffs)
371 			diffs = memcmp(skb_mac_header(p),
372 				       skb_mac_header(skb),
373 				       maclen);
374 
375 		/* in most common scenarions 'slow_gro' is 0
376 		 * otherwise we are already on some slower paths
377 		 * either skip all the infrequent tests altogether or
378 		 * avoid trying too hard to skip each of them individually
379 		 */
380 		if (!diffs && unlikely(skb->slow_gro | p->slow_gro)) {
381 #if IS_ENABLED(CONFIG_SKB_EXTENSIONS) && IS_ENABLED(CONFIG_NET_TC_SKB_EXT)
382 			struct tc_skb_ext *skb_ext;
383 			struct tc_skb_ext *p_ext;
384 #endif
385 
386 			diffs |= p->sk != skb->sk;
387 			diffs |= skb_metadata_dst_cmp(p, skb);
388 			diffs |= skb_get_nfct(p) ^ skb_get_nfct(skb);
389 
390 #if IS_ENABLED(CONFIG_SKB_EXTENSIONS) && IS_ENABLED(CONFIG_NET_TC_SKB_EXT)
391 			skb_ext = skb_ext_find(skb, TC_SKB_EXT);
392 			p_ext = skb_ext_find(p, TC_SKB_EXT);
393 
394 			diffs |= (!!p_ext) ^ (!!skb_ext);
395 			if (!diffs && unlikely(skb_ext))
396 				diffs |= p_ext->chain ^ skb_ext->chain;
397 #endif
398 		}
399 
400 		NAPI_GRO_CB(p)->same_flow = !diffs;
401 	}
402 }
403 
404 static inline void skb_gro_reset_offset(struct sk_buff *skb, u32 nhoff)
405 {
406 	const struct skb_shared_info *pinfo = skb_shinfo(skb);
407 	const skb_frag_t *frag0 = &pinfo->frags[0];
408 
409 	NAPI_GRO_CB(skb)->data_offset = 0;
410 	NAPI_GRO_CB(skb)->frag0 = NULL;
411 	NAPI_GRO_CB(skb)->frag0_len = 0;
412 
413 	if (!skb_headlen(skb) && pinfo->nr_frags &&
414 	    !PageHighMem(skb_frag_page(frag0)) &&
415 	    (!NET_IP_ALIGN || !((skb_frag_off(frag0) + nhoff) & 3))) {
416 		NAPI_GRO_CB(skb)->frag0 = skb_frag_address(frag0);
417 		NAPI_GRO_CB(skb)->frag0_len = min_t(unsigned int,
418 						    skb_frag_size(frag0),
419 						    skb->end - skb->tail);
420 	}
421 }
422 
423 static void gro_pull_from_frag0(struct sk_buff *skb, int grow)
424 {
425 	struct skb_shared_info *pinfo = skb_shinfo(skb);
426 
427 	BUG_ON(skb->end - skb->tail < grow);
428 
429 	memcpy(skb_tail_pointer(skb), NAPI_GRO_CB(skb)->frag0, grow);
430 
431 	skb->data_len -= grow;
432 	skb->tail += grow;
433 
434 	skb_frag_off_add(&pinfo->frags[0], grow);
435 	skb_frag_size_sub(&pinfo->frags[0], grow);
436 
437 	if (unlikely(!skb_frag_size(&pinfo->frags[0]))) {
438 		skb_frag_unref(skb, 0);
439 		memmove(pinfo->frags, pinfo->frags + 1,
440 			--pinfo->nr_frags * sizeof(pinfo->frags[0]));
441 	}
442 }
443 
444 static void gro_flush_oldest(struct napi_struct *napi, struct list_head *head)
445 {
446 	struct sk_buff *oldest;
447 
448 	oldest = list_last_entry(head, struct sk_buff, list);
449 
450 	/* We are called with head length >= MAX_GRO_SKBS, so this is
451 	 * impossible.
452 	 */
453 	if (WARN_ON_ONCE(!oldest))
454 		return;
455 
456 	/* Do not adjust napi->gro_hash[].count, caller is adding a new
457 	 * SKB to the chain.
458 	 */
459 	skb_list_del_init(oldest);
460 	napi_gro_complete(napi, oldest);
461 }
462 
463 static enum gro_result dev_gro_receive(struct napi_struct *napi, struct sk_buff *skb)
464 {
465 	u32 bucket = skb_get_hash_raw(skb) & (GRO_HASH_BUCKETS - 1);
466 	struct gro_list *gro_list = &napi->gro_hash[bucket];
467 	struct list_head *head = &offload_base;
468 	struct packet_offload *ptype;
469 	__be16 type = skb->protocol;
470 	struct sk_buff *pp = NULL;
471 	enum gro_result ret;
472 	int same_flow;
473 	int grow;
474 
475 	if (netif_elide_gro(skb->dev))
476 		goto normal;
477 
478 	gro_list_prepare(&gro_list->list, skb);
479 
480 	rcu_read_lock();
481 	list_for_each_entry_rcu(ptype, head, list) {
482 		if (ptype->type != type || !ptype->callbacks.gro_receive)
483 			continue;
484 
485 		skb_set_network_header(skb, skb_gro_offset(skb));
486 		skb_reset_mac_len(skb);
487 		BUILD_BUG_ON(sizeof_field(struct napi_gro_cb, zeroed) != sizeof(u32));
488 		BUILD_BUG_ON(!IS_ALIGNED(offsetof(struct napi_gro_cb, zeroed),
489 					 sizeof(u32))); /* Avoid slow unaligned acc */
490 		*(u32 *)&NAPI_GRO_CB(skb)->zeroed = 0;
491 		NAPI_GRO_CB(skb)->flush = skb_is_gso(skb) || skb_has_frag_list(skb);
492 		NAPI_GRO_CB(skb)->is_atomic = 1;
493 
494 		/* Setup for GRO checksum validation */
495 		switch (skb->ip_summed) {
496 		case CHECKSUM_COMPLETE:
497 			NAPI_GRO_CB(skb)->csum = skb->csum;
498 			NAPI_GRO_CB(skb)->csum_valid = 1;
499 			break;
500 		case CHECKSUM_UNNECESSARY:
501 			NAPI_GRO_CB(skb)->csum_cnt = skb->csum_level + 1;
502 			break;
503 		}
504 
505 		pp = INDIRECT_CALL_INET(ptype->callbacks.gro_receive,
506 					ipv6_gro_receive, inet_gro_receive,
507 					&gro_list->list, skb);
508 		break;
509 	}
510 	rcu_read_unlock();
511 
512 	if (&ptype->list == head)
513 		goto normal;
514 
515 	if (PTR_ERR(pp) == -EINPROGRESS) {
516 		ret = GRO_CONSUMED;
517 		goto ok;
518 	}
519 
520 	same_flow = NAPI_GRO_CB(skb)->same_flow;
521 	ret = NAPI_GRO_CB(skb)->free ? GRO_MERGED_FREE : GRO_MERGED;
522 
523 	if (pp) {
524 		skb_list_del_init(pp);
525 		napi_gro_complete(napi, pp);
526 		gro_list->count--;
527 	}
528 
529 	if (same_flow)
530 		goto ok;
531 
532 	if (NAPI_GRO_CB(skb)->flush)
533 		goto normal;
534 
535 	if (unlikely(gro_list->count >= MAX_GRO_SKBS))
536 		gro_flush_oldest(napi, &gro_list->list);
537 	else
538 		gro_list->count++;
539 
540 	NAPI_GRO_CB(skb)->count = 1;
541 	NAPI_GRO_CB(skb)->age = jiffies;
542 	NAPI_GRO_CB(skb)->last = skb;
543 	skb_shinfo(skb)->gso_size = skb_gro_len(skb);
544 	list_add(&skb->list, &gro_list->list);
545 	ret = GRO_HELD;
546 
547 pull:
548 	grow = skb_gro_offset(skb) - skb_headlen(skb);
549 	if (grow > 0)
550 		gro_pull_from_frag0(skb, grow);
551 ok:
552 	if (gro_list->count) {
553 		if (!test_bit(bucket, &napi->gro_bitmask))
554 			__set_bit(bucket, &napi->gro_bitmask);
555 	} else if (test_bit(bucket, &napi->gro_bitmask)) {
556 		__clear_bit(bucket, &napi->gro_bitmask);
557 	}
558 
559 	return ret;
560 
561 normal:
562 	ret = GRO_NORMAL;
563 	goto pull;
564 }
565 
566 struct packet_offload *gro_find_receive_by_type(__be16 type)
567 {
568 	struct list_head *offload_head = &offload_base;
569 	struct packet_offload *ptype;
570 
571 	list_for_each_entry_rcu(ptype, offload_head, list) {
572 		if (ptype->type != type || !ptype->callbacks.gro_receive)
573 			continue;
574 		return ptype;
575 	}
576 	return NULL;
577 }
578 EXPORT_SYMBOL(gro_find_receive_by_type);
579 
580 struct packet_offload *gro_find_complete_by_type(__be16 type)
581 {
582 	struct list_head *offload_head = &offload_base;
583 	struct packet_offload *ptype;
584 
585 	list_for_each_entry_rcu(ptype, offload_head, list) {
586 		if (ptype->type != type || !ptype->callbacks.gro_complete)
587 			continue;
588 		return ptype;
589 	}
590 	return NULL;
591 }
592 EXPORT_SYMBOL(gro_find_complete_by_type);
593 
594 static gro_result_t napi_skb_finish(struct napi_struct *napi,
595 				    struct sk_buff *skb,
596 				    gro_result_t ret)
597 {
598 	switch (ret) {
599 	case GRO_NORMAL:
600 		gro_normal_one(napi, skb, 1);
601 		break;
602 
603 	case GRO_MERGED_FREE:
604 		if (NAPI_GRO_CB(skb)->free == NAPI_GRO_FREE_STOLEN_HEAD)
605 			napi_skb_free_stolen_head(skb);
606 		else if (skb->fclone != SKB_FCLONE_UNAVAILABLE)
607 			__kfree_skb(skb);
608 		else
609 			__kfree_skb_defer(skb);
610 		break;
611 
612 	case GRO_HELD:
613 	case GRO_MERGED:
614 	case GRO_CONSUMED:
615 		break;
616 	}
617 
618 	return ret;
619 }
620 
621 gro_result_t napi_gro_receive(struct napi_struct *napi, struct sk_buff *skb)
622 {
623 	gro_result_t ret;
624 
625 	skb_mark_napi_id(skb, napi);
626 	trace_napi_gro_receive_entry(skb);
627 
628 	skb_gro_reset_offset(skb, 0);
629 
630 	ret = napi_skb_finish(napi, skb, dev_gro_receive(napi, skb));
631 	trace_napi_gro_receive_exit(ret);
632 
633 	return ret;
634 }
635 EXPORT_SYMBOL(napi_gro_receive);
636 
637 static void napi_reuse_skb(struct napi_struct *napi, struct sk_buff *skb)
638 {
639 	if (unlikely(skb->pfmemalloc)) {
640 		consume_skb(skb);
641 		return;
642 	}
643 	__skb_pull(skb, skb_headlen(skb));
644 	/* restore the reserve we had after netdev_alloc_skb_ip_align() */
645 	skb_reserve(skb, NET_SKB_PAD + NET_IP_ALIGN - skb_headroom(skb));
646 	__vlan_hwaccel_clear_tag(skb);
647 	skb->dev = napi->dev;
648 	skb->skb_iif = 0;
649 
650 	/* eth_type_trans() assumes pkt_type is PACKET_HOST */
651 	skb->pkt_type = PACKET_HOST;
652 
653 	skb->encapsulation = 0;
654 	skb_shinfo(skb)->gso_type = 0;
655 	if (unlikely(skb->slow_gro)) {
656 		skb_orphan(skb);
657 		skb_ext_reset(skb);
658 		nf_reset_ct(skb);
659 		skb->slow_gro = 0;
660 	}
661 
662 	napi->skb = skb;
663 }
664 
665 struct sk_buff *napi_get_frags(struct napi_struct *napi)
666 {
667 	struct sk_buff *skb = napi->skb;
668 
669 	if (!skb) {
670 		skb = napi_alloc_skb(napi, GRO_MAX_HEAD);
671 		if (skb) {
672 			napi->skb = skb;
673 			skb_mark_napi_id(skb, napi);
674 		}
675 	}
676 	return skb;
677 }
678 EXPORT_SYMBOL(napi_get_frags);
679 
680 static gro_result_t napi_frags_finish(struct napi_struct *napi,
681 				      struct sk_buff *skb,
682 				      gro_result_t ret)
683 {
684 	switch (ret) {
685 	case GRO_NORMAL:
686 	case GRO_HELD:
687 		__skb_push(skb, ETH_HLEN);
688 		skb->protocol = eth_type_trans(skb, skb->dev);
689 		if (ret == GRO_NORMAL)
690 			gro_normal_one(napi, skb, 1);
691 		break;
692 
693 	case GRO_MERGED_FREE:
694 		if (NAPI_GRO_CB(skb)->free == NAPI_GRO_FREE_STOLEN_HEAD)
695 			napi_skb_free_stolen_head(skb);
696 		else
697 			napi_reuse_skb(napi, skb);
698 		break;
699 
700 	case GRO_MERGED:
701 	case GRO_CONSUMED:
702 		break;
703 	}
704 
705 	return ret;
706 }
707 
708 /* Upper GRO stack assumes network header starts at gro_offset=0
709  * Drivers could call both napi_gro_frags() and napi_gro_receive()
710  * We copy ethernet header into skb->data to have a common layout.
711  */
712 static struct sk_buff *napi_frags_skb(struct napi_struct *napi)
713 {
714 	struct sk_buff *skb = napi->skb;
715 	const struct ethhdr *eth;
716 	unsigned int hlen = sizeof(*eth);
717 
718 	napi->skb = NULL;
719 
720 	skb_reset_mac_header(skb);
721 	skb_gro_reset_offset(skb, hlen);
722 
723 	if (unlikely(skb_gro_header_hard(skb, hlen))) {
724 		eth = skb_gro_header_slow(skb, hlen, 0);
725 		if (unlikely(!eth)) {
726 			net_warn_ratelimited("%s: dropping impossible skb from %s\n",
727 					     __func__, napi->dev->name);
728 			napi_reuse_skb(napi, skb);
729 			return NULL;
730 		}
731 	} else {
732 		eth = (const struct ethhdr *)skb->data;
733 		gro_pull_from_frag0(skb, hlen);
734 		NAPI_GRO_CB(skb)->frag0 += hlen;
735 		NAPI_GRO_CB(skb)->frag0_len -= hlen;
736 	}
737 	__skb_pull(skb, hlen);
738 
739 	/*
740 	 * This works because the only protocols we care about don't require
741 	 * special handling.
742 	 * We'll fix it up properly in napi_frags_finish()
743 	 */
744 	skb->protocol = eth->h_proto;
745 
746 	return skb;
747 }
748 
749 gro_result_t napi_gro_frags(struct napi_struct *napi)
750 {
751 	gro_result_t ret;
752 	struct sk_buff *skb = napi_frags_skb(napi);
753 
754 	trace_napi_gro_frags_entry(skb);
755 
756 	ret = napi_frags_finish(napi, skb, dev_gro_receive(napi, skb));
757 	trace_napi_gro_frags_exit(ret);
758 
759 	return ret;
760 }
761 EXPORT_SYMBOL(napi_gro_frags);
762 
763 /* Compute the checksum from gro_offset and return the folded value
764  * after adding in any pseudo checksum.
765  */
766 __sum16 __skb_gro_checksum_complete(struct sk_buff *skb)
767 {
768 	__wsum wsum;
769 	__sum16 sum;
770 
771 	wsum = skb_checksum(skb, skb_gro_offset(skb), skb_gro_len(skb), 0);
772 
773 	/* NAPI_GRO_CB(skb)->csum holds pseudo checksum */
774 	sum = csum_fold(csum_add(NAPI_GRO_CB(skb)->csum, wsum));
775 	/* See comments in __skb_checksum_complete(). */
776 	if (likely(!sum)) {
777 		if (unlikely(skb->ip_summed == CHECKSUM_COMPLETE) &&
778 		    !skb->csum_complete_sw)
779 			netdev_rx_csum_fault(skb->dev, skb);
780 	}
781 
782 	NAPI_GRO_CB(skb)->csum = wsum;
783 	NAPI_GRO_CB(skb)->csum_valid = 1;
784 
785 	return sum;
786 }
787 EXPORT_SYMBOL(__skb_gro_checksum_complete);
788