xref: /linux/net/core/gro.c (revision a33682e4e78e249155abbe5e8ee880d5760b5e28)
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 	/* 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(_ipv4)_max_size() */
175 	gro_max_size = p->protocol == htons(ETH_P_IPV6) ?
176 			READ_ONCE(p->dev->gro_max_size) :
177 			READ_ONCE(p->dev->gro_ipv4_max_size);
178 
179 	if (unlikely(p->len + len >= gro_max_size || NAPI_GRO_CB(skb)->flush))
180 		return -E2BIG;
181 
182 	if (unlikely(p->len + len >= GRO_LEGACY_MAX_SIZE)) {
183 		if (NAPI_GRO_CB(skb)->proto != IPPROTO_TCP ||
184 		    (p->protocol == htons(ETH_P_IPV6) &&
185 		     skb_headroom(p) < sizeof(struct hop_jumbo_hdr)) ||
186 		    p->encapsulation)
187 			return -E2BIG;
188 	}
189 
190 	segs = NAPI_GRO_CB(skb)->count;
191 	lp = NAPI_GRO_CB(p)->last;
192 	pinfo = skb_shinfo(lp);
193 
194 	if (headlen <= offset) {
195 		skb_frag_t *frag;
196 		skb_frag_t *frag2;
197 		int i = skbinfo->nr_frags;
198 		int nr_frags = pinfo->nr_frags + i;
199 
200 		if (nr_frags > MAX_SKB_FRAGS)
201 			goto merge;
202 
203 		offset -= headlen;
204 		pinfo->nr_frags = nr_frags;
205 		skbinfo->nr_frags = 0;
206 
207 		frag = pinfo->frags + nr_frags;
208 		frag2 = skbinfo->frags + i;
209 		do {
210 			*--frag = *--frag2;
211 		} while (--i);
212 
213 		skb_frag_off_add(frag, offset);
214 		skb_frag_size_sub(frag, offset);
215 
216 		/* all fragments truesize : remove (head size + sk_buff) */
217 		new_truesize = SKB_TRUESIZE(skb_end_offset(skb));
218 		delta_truesize = skb->truesize - new_truesize;
219 
220 		skb->truesize = new_truesize;
221 		skb->len -= skb->data_len;
222 		skb->data_len = 0;
223 
224 		NAPI_GRO_CB(skb)->free = NAPI_GRO_FREE;
225 		goto done;
226 	} else if (skb->head_frag) {
227 		int nr_frags = pinfo->nr_frags;
228 		skb_frag_t *frag = pinfo->frags + nr_frags;
229 		struct page *page = virt_to_head_page(skb->head);
230 		unsigned int first_size = headlen - offset;
231 		unsigned int first_offset;
232 
233 		if (nr_frags + 1 + skbinfo->nr_frags > MAX_SKB_FRAGS)
234 			goto merge;
235 
236 		first_offset = skb->data -
237 			       (unsigned char *)page_address(page) +
238 			       offset;
239 
240 		pinfo->nr_frags = nr_frags + 1 + skbinfo->nr_frags;
241 
242 		skb_frag_fill_page_desc(frag, page, first_offset, 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 
254 merge:
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 
278 done:
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 
293 static 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 
325 out:
326 	gro_normal_one(napi, skb, NAPI_GRO_CB(skb)->count);
327 }
328 
329 static 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  */
351 void 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 }
362 EXPORT_SYMBOL(napi_gro_flush);
363 
364 static 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 
421 static 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 
440 static 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 
461 static void gro_try_pull_from_frag0(struct sk_buff *skb)
462 {
463 	int grow = skb_gro_offset(skb) - skb_headlen(skb);
464 
465 	if (grow > 0)
466 		gro_pull_from_frag0(skb, grow);
467 }
468 
469 static void gro_flush_oldest(struct napi_struct *napi, struct list_head *head)
470 {
471 	struct sk_buff *oldest;
472 
473 	oldest = list_last_entry(head, struct sk_buff, list);
474 
475 	/* We are called with head length >= MAX_GRO_SKBS, so this is
476 	 * impossible.
477 	 */
478 	if (WARN_ON_ONCE(!oldest))
479 		return;
480 
481 	/* Do not adjust napi->gro_hash[].count, caller is adding a new
482 	 * SKB to the chain.
483 	 */
484 	skb_list_del_init(oldest);
485 	napi_gro_complete(napi, oldest);
486 }
487 
488 static enum gro_result dev_gro_receive(struct napi_struct *napi, struct sk_buff *skb)
489 {
490 	u32 bucket = skb_get_hash_raw(skb) & (GRO_HASH_BUCKETS - 1);
491 	struct gro_list *gro_list = &napi->gro_hash[bucket];
492 	struct list_head *head = &offload_base;
493 	struct packet_offload *ptype;
494 	__be16 type = skb->protocol;
495 	struct sk_buff *pp = NULL;
496 	enum gro_result ret;
497 	int same_flow;
498 
499 	if (netif_elide_gro(skb->dev))
500 		goto normal;
501 
502 	gro_list_prepare(&gro_list->list, skb);
503 
504 	rcu_read_lock();
505 	list_for_each_entry_rcu(ptype, head, list) {
506 		if (ptype->type == type && ptype->callbacks.gro_receive)
507 			goto found_ptype;
508 	}
509 	rcu_read_unlock();
510 	goto normal;
511 
512 found_ptype:
513 	skb_set_network_header(skb, skb_gro_offset(skb));
514 	skb_reset_mac_len(skb);
515 	BUILD_BUG_ON(sizeof_field(struct napi_gro_cb, zeroed) != sizeof(u32));
516 	BUILD_BUG_ON(!IS_ALIGNED(offsetof(struct napi_gro_cb, zeroed),
517 					sizeof(u32))); /* Avoid slow unaligned acc */
518 	*(u32 *)&NAPI_GRO_CB(skb)->zeroed = 0;
519 	NAPI_GRO_CB(skb)->flush = skb_has_frag_list(skb);
520 	NAPI_GRO_CB(skb)->is_atomic = 1;
521 	NAPI_GRO_CB(skb)->count = 1;
522 	if (unlikely(skb_is_gso(skb))) {
523 		NAPI_GRO_CB(skb)->count = skb_shinfo(skb)->gso_segs;
524 		/* Only support TCP and non DODGY users. */
525 		if (!skb_is_gso_tcp(skb) ||
526 		    (skb_shinfo(skb)->gso_type & SKB_GSO_DODGY))
527 			NAPI_GRO_CB(skb)->flush = 1;
528 	}
529 
530 	/* Setup for GRO checksum validation */
531 	switch (skb->ip_summed) {
532 	case CHECKSUM_COMPLETE:
533 		NAPI_GRO_CB(skb)->csum = skb->csum;
534 		NAPI_GRO_CB(skb)->csum_valid = 1;
535 		break;
536 	case CHECKSUM_UNNECESSARY:
537 		NAPI_GRO_CB(skb)->csum_cnt = skb->csum_level + 1;
538 		break;
539 	}
540 
541 	pp = INDIRECT_CALL_INET(ptype->callbacks.gro_receive,
542 				ipv6_gro_receive, inet_gro_receive,
543 				&gro_list->list, skb);
544 
545 	rcu_read_unlock();
546 
547 	if (PTR_ERR(pp) == -EINPROGRESS) {
548 		ret = GRO_CONSUMED;
549 		goto ok;
550 	}
551 
552 	same_flow = NAPI_GRO_CB(skb)->same_flow;
553 	ret = NAPI_GRO_CB(skb)->free ? GRO_MERGED_FREE : GRO_MERGED;
554 
555 	if (pp) {
556 		skb_list_del_init(pp);
557 		napi_gro_complete(napi, pp);
558 		gro_list->count--;
559 	}
560 
561 	if (same_flow)
562 		goto ok;
563 
564 	if (NAPI_GRO_CB(skb)->flush)
565 		goto normal;
566 
567 	if (unlikely(gro_list->count >= MAX_GRO_SKBS))
568 		gro_flush_oldest(napi, &gro_list->list);
569 	else
570 		gro_list->count++;
571 
572 	/* Must be called before setting NAPI_GRO_CB(skb)->{age|last} */
573 	gro_try_pull_from_frag0(skb);
574 	NAPI_GRO_CB(skb)->age = jiffies;
575 	NAPI_GRO_CB(skb)->last = skb;
576 	if (!skb_is_gso(skb))
577 		skb_shinfo(skb)->gso_size = skb_gro_len(skb);
578 	list_add(&skb->list, &gro_list->list);
579 	ret = GRO_HELD;
580 ok:
581 	if (gro_list->count) {
582 		if (!test_bit(bucket, &napi->gro_bitmask))
583 			__set_bit(bucket, &napi->gro_bitmask);
584 	} else if (test_bit(bucket, &napi->gro_bitmask)) {
585 		__clear_bit(bucket, &napi->gro_bitmask);
586 	}
587 
588 	return ret;
589 
590 normal:
591 	ret = GRO_NORMAL;
592 	gro_try_pull_from_frag0(skb);
593 	goto ok;
594 }
595 
596 struct packet_offload *gro_find_receive_by_type(__be16 type)
597 {
598 	struct list_head *offload_head = &offload_base;
599 	struct packet_offload *ptype;
600 
601 	list_for_each_entry_rcu(ptype, offload_head, list) {
602 		if (ptype->type != type || !ptype->callbacks.gro_receive)
603 			continue;
604 		return ptype;
605 	}
606 	return NULL;
607 }
608 EXPORT_SYMBOL(gro_find_receive_by_type);
609 
610 struct packet_offload *gro_find_complete_by_type(__be16 type)
611 {
612 	struct list_head *offload_head = &offload_base;
613 	struct packet_offload *ptype;
614 
615 	list_for_each_entry_rcu(ptype, offload_head, list) {
616 		if (ptype->type != type || !ptype->callbacks.gro_complete)
617 			continue;
618 		return ptype;
619 	}
620 	return NULL;
621 }
622 EXPORT_SYMBOL(gro_find_complete_by_type);
623 
624 static gro_result_t napi_skb_finish(struct napi_struct *napi,
625 				    struct sk_buff *skb,
626 				    gro_result_t ret)
627 {
628 	switch (ret) {
629 	case GRO_NORMAL:
630 		gro_normal_one(napi, skb, 1);
631 		break;
632 
633 	case GRO_MERGED_FREE:
634 		if (NAPI_GRO_CB(skb)->free == NAPI_GRO_FREE_STOLEN_HEAD)
635 			napi_skb_free_stolen_head(skb);
636 		else if (skb->fclone != SKB_FCLONE_UNAVAILABLE)
637 			__kfree_skb(skb);
638 		else
639 			__napi_kfree_skb(skb, SKB_CONSUMED);
640 		break;
641 
642 	case GRO_HELD:
643 	case GRO_MERGED:
644 	case GRO_CONSUMED:
645 		break;
646 	}
647 
648 	return ret;
649 }
650 
651 gro_result_t napi_gro_receive(struct napi_struct *napi, struct sk_buff *skb)
652 {
653 	gro_result_t ret;
654 
655 	skb_mark_napi_id(skb, napi);
656 	trace_napi_gro_receive_entry(skb);
657 
658 	skb_gro_reset_offset(skb, 0);
659 
660 	ret = napi_skb_finish(napi, skb, dev_gro_receive(napi, skb));
661 	trace_napi_gro_receive_exit(ret);
662 
663 	return ret;
664 }
665 EXPORT_SYMBOL(napi_gro_receive);
666 
667 static void napi_reuse_skb(struct napi_struct *napi, struct sk_buff *skb)
668 {
669 	if (unlikely(skb->pfmemalloc)) {
670 		consume_skb(skb);
671 		return;
672 	}
673 	__skb_pull(skb, skb_headlen(skb));
674 	/* restore the reserve we had after netdev_alloc_skb_ip_align() */
675 	skb_reserve(skb, NET_SKB_PAD + NET_IP_ALIGN - skb_headroom(skb));
676 	__vlan_hwaccel_clear_tag(skb);
677 	skb->dev = napi->dev;
678 	skb->skb_iif = 0;
679 
680 	/* eth_type_trans() assumes pkt_type is PACKET_HOST */
681 	skb->pkt_type = PACKET_HOST;
682 
683 	skb->encapsulation = 0;
684 	skb_shinfo(skb)->gso_type = 0;
685 	skb_shinfo(skb)->gso_size = 0;
686 	if (unlikely(skb->slow_gro)) {
687 		skb_orphan(skb);
688 		skb_ext_reset(skb);
689 		nf_reset_ct(skb);
690 		skb->slow_gro = 0;
691 	}
692 
693 	napi->skb = skb;
694 }
695 
696 struct sk_buff *napi_get_frags(struct napi_struct *napi)
697 {
698 	struct sk_buff *skb = napi->skb;
699 
700 	if (!skb) {
701 		skb = napi_alloc_skb(napi, GRO_MAX_HEAD);
702 		if (skb) {
703 			napi->skb = skb;
704 			skb_mark_napi_id(skb, napi);
705 		}
706 	}
707 	return skb;
708 }
709 EXPORT_SYMBOL(napi_get_frags);
710 
711 static gro_result_t napi_frags_finish(struct napi_struct *napi,
712 				      struct sk_buff *skb,
713 				      gro_result_t ret)
714 {
715 	switch (ret) {
716 	case GRO_NORMAL:
717 	case GRO_HELD:
718 		__skb_push(skb, ETH_HLEN);
719 		skb->protocol = eth_type_trans(skb, skb->dev);
720 		if (ret == GRO_NORMAL)
721 			gro_normal_one(napi, skb, 1);
722 		break;
723 
724 	case GRO_MERGED_FREE:
725 		if (NAPI_GRO_CB(skb)->free == NAPI_GRO_FREE_STOLEN_HEAD)
726 			napi_skb_free_stolen_head(skb);
727 		else
728 			napi_reuse_skb(napi, skb);
729 		break;
730 
731 	case GRO_MERGED:
732 	case GRO_CONSUMED:
733 		break;
734 	}
735 
736 	return ret;
737 }
738 
739 /* Upper GRO stack assumes network header starts at gro_offset=0
740  * Drivers could call both napi_gro_frags() and napi_gro_receive()
741  * We copy ethernet header into skb->data to have a common layout.
742  */
743 static struct sk_buff *napi_frags_skb(struct napi_struct *napi)
744 {
745 	struct sk_buff *skb = napi->skb;
746 	const struct ethhdr *eth;
747 	unsigned int hlen = sizeof(*eth);
748 
749 	napi->skb = NULL;
750 
751 	skb_reset_mac_header(skb);
752 	skb_gro_reset_offset(skb, hlen);
753 
754 	if (unlikely(skb_gro_header_hard(skb, hlen))) {
755 		eth = skb_gro_header_slow(skb, hlen, 0);
756 		if (unlikely(!eth)) {
757 			net_warn_ratelimited("%s: dropping impossible skb from %s\n",
758 					     __func__, napi->dev->name);
759 			napi_reuse_skb(napi, skb);
760 			return NULL;
761 		}
762 	} else {
763 		eth = (const struct ethhdr *)skb->data;
764 		gro_pull_from_frag0(skb, hlen);
765 		NAPI_GRO_CB(skb)->frag0 += hlen;
766 		NAPI_GRO_CB(skb)->frag0_len -= hlen;
767 	}
768 	__skb_pull(skb, hlen);
769 
770 	/*
771 	 * This works because the only protocols we care about don't require
772 	 * special handling.
773 	 * We'll fix it up properly in napi_frags_finish()
774 	 */
775 	skb->protocol = eth->h_proto;
776 
777 	return skb;
778 }
779 
780 gro_result_t napi_gro_frags(struct napi_struct *napi)
781 {
782 	gro_result_t ret;
783 	struct sk_buff *skb = napi_frags_skb(napi);
784 
785 	trace_napi_gro_frags_entry(skb);
786 
787 	ret = napi_frags_finish(napi, skb, dev_gro_receive(napi, skb));
788 	trace_napi_gro_frags_exit(ret);
789 
790 	return ret;
791 }
792 EXPORT_SYMBOL(napi_gro_frags);
793 
794 /* Compute the checksum from gro_offset and return the folded value
795  * after adding in any pseudo checksum.
796  */
797 __sum16 __skb_gro_checksum_complete(struct sk_buff *skb)
798 {
799 	__wsum wsum;
800 	__sum16 sum;
801 
802 	wsum = skb_checksum(skb, skb_gro_offset(skb), skb_gro_len(skb), 0);
803 
804 	/* NAPI_GRO_CB(skb)->csum holds pseudo checksum */
805 	sum = csum_fold(csum_add(NAPI_GRO_CB(skb)->csum, wsum));
806 	/* See comments in __skb_checksum_complete(). */
807 	if (likely(!sum)) {
808 		if (unlikely(skb->ip_summed == CHECKSUM_COMPLETE) &&
809 		    !skb->csum_complete_sw)
810 			netdev_rx_csum_fault(skb->dev, skb);
811 	}
812 
813 	NAPI_GRO_CB(skb)->csum = wsum;
814 	NAPI_GRO_CB(skb)->csum_valid = 1;
815 
816 	return sum;
817 }
818 EXPORT_SYMBOL(__skb_gro_checksum_complete);
819