Linux Audio

Check our new training course

Real-Time Linux with PREEMPT_RT training

Feb 18-20, 2025
Register
Loading...
v6.13.7
  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);
v6.8
  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);
 13struct 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
 96int skb_gro_receive(struct sk_buff *p, struct sk_buff *skb)
 97{
 98	struct skb_shared_info *pinfo, *skbinfo = skb_shinfo(skb);
 99	unsigned int offset = skb_gro_offset(skb);
100	unsigned int headlen = skb_headlen(skb);
101	unsigned int len = skb_gro_len(skb);
102	unsigned int delta_truesize;
103	unsigned int gro_max_size;
104	unsigned int new_truesize;
105	struct sk_buff *lp;
106	int segs;
107
108	/* Do not splice page pool based packets w/ non-page pool
109	 * packets. This can result in reference count issues as page
110	 * pool pages will not decrement the reference count and will
111	 * instead be immediately returned to the pool or have frag
112	 * count decremented.
113	 */
114	if (p->pp_recycle != skb->pp_recycle)
115		return -ETOOMANYREFS;
116
117	/* pairs with WRITE_ONCE() in netif_set_gro(_ipv4)_max_size() */
118	gro_max_size = p->protocol == htons(ETH_P_IPV6) ?
119			READ_ONCE(p->dev->gro_max_size) :
120			READ_ONCE(p->dev->gro_ipv4_max_size);
121
122	if (unlikely(p->len + len >= gro_max_size || NAPI_GRO_CB(skb)->flush))
123		return -E2BIG;
124
125	if (unlikely(p->len + len >= GRO_LEGACY_MAX_SIZE)) {
126		if (NAPI_GRO_CB(skb)->proto != IPPROTO_TCP ||
127		    (p->protocol == htons(ETH_P_IPV6) &&
128		     skb_headroom(p) < sizeof(struct hop_jumbo_hdr)) ||
129		    p->encapsulation)
130			return -E2BIG;
131	}
132
133	segs = NAPI_GRO_CB(skb)->count;
134	lp = NAPI_GRO_CB(p)->last;
135	pinfo = skb_shinfo(lp);
136
137	if (headlen <= offset) {
138		skb_frag_t *frag;
139		skb_frag_t *frag2;
140		int i = skbinfo->nr_frags;
141		int nr_frags = pinfo->nr_frags + i;
142
143		if (nr_frags > MAX_SKB_FRAGS)
144			goto merge;
145
146		offset -= headlen;
147		pinfo->nr_frags = nr_frags;
148		skbinfo->nr_frags = 0;
149
150		frag = pinfo->frags + nr_frags;
151		frag2 = skbinfo->frags + i;
152		do {
153			*--frag = *--frag2;
154		} while (--i);
155
156		skb_frag_off_add(frag, offset);
157		skb_frag_size_sub(frag, offset);
158
159		/* all fragments truesize : remove (head size + sk_buff) */
160		new_truesize = SKB_TRUESIZE(skb_end_offset(skb));
161		delta_truesize = skb->truesize - new_truesize;
162
163		skb->truesize = new_truesize;
164		skb->len -= skb->data_len;
165		skb->data_len = 0;
166
167		NAPI_GRO_CB(skb)->free = NAPI_GRO_FREE;
168		goto done;
169	} else if (skb->head_frag) {
170		int nr_frags = pinfo->nr_frags;
171		skb_frag_t *frag = pinfo->frags + nr_frags;
172		struct page *page = virt_to_head_page(skb->head);
173		unsigned int first_size = headlen - offset;
174		unsigned int first_offset;
175
176		if (nr_frags + 1 + skbinfo->nr_frags > MAX_SKB_FRAGS)
177			goto merge;
178
179		first_offset = skb->data -
180			       (unsigned char *)page_address(page) +
181			       offset;
182
183		pinfo->nr_frags = nr_frags + 1 + skbinfo->nr_frags;
184
185		skb_frag_fill_page_desc(frag, page, first_offset, first_size);
186
187		memcpy(frag + 1, skbinfo->frags, sizeof(*frag) * skbinfo->nr_frags);
188		/* We dont need to clear skbinfo->nr_frags here */
189
190		new_truesize = SKB_DATA_ALIGN(sizeof(struct sk_buff));
191		delta_truesize = skb->truesize - new_truesize;
192		skb->truesize = new_truesize;
193		NAPI_GRO_CB(skb)->free = NAPI_GRO_FREE_STOLEN_HEAD;
194		goto done;
195	}
196
197merge:
198	/* sk owenrship - if any - completely transferred to the aggregated packet */
199	skb->destructor = NULL;
 
200	delta_truesize = skb->truesize;
201	if (offset > headlen) {
202		unsigned int eat = offset - headlen;
203
204		skb_frag_off_add(&skbinfo->frags[0], eat);
205		skb_frag_size_sub(&skbinfo->frags[0], eat);
206		skb->data_len -= eat;
207		skb->len -= eat;
208		offset = headlen;
209	}
210
211	__skb_pull(skb, offset);
212
213	if (NAPI_GRO_CB(p)->last == p)
214		skb_shinfo(p)->frag_list = skb;
215	else
216		NAPI_GRO_CB(p)->last->next = skb;
217	NAPI_GRO_CB(p)->last = skb;
218	__skb_header_release(skb);
219	lp = p;
220
221done:
222	NAPI_GRO_CB(p)->count += segs;
223	p->data_len += len;
224	p->truesize += delta_truesize;
225	p->len += len;
226	if (lp != p) {
227		lp->data_len += len;
228		lp->truesize += delta_truesize;
229		lp->len += len;
230	}
231	NAPI_GRO_CB(skb)->same_flow = 1;
232	return 0;
233}
234
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
235
236static void napi_gro_complete(struct napi_struct *napi, struct sk_buff *skb)
237{
 
238	struct packet_offload *ptype;
239	__be16 type = skb->protocol;
240	struct list_head *head = &offload_base;
241	int err = -ENOENT;
242
243	BUILD_BUG_ON(sizeof(struct napi_gro_cb) > sizeof(skb->cb));
244
245	if (NAPI_GRO_CB(skb)->count == 1) {
246		skb_shinfo(skb)->gso_size = 0;
247		goto out;
248	}
249
250	rcu_read_lock();
251	list_for_each_entry_rcu(ptype, head, list) {
252		if (ptype->type != type || !ptype->callbacks.gro_complete)
253			continue;
254
255		err = INDIRECT_CALL_INET(ptype->callbacks.gro_complete,
256					 ipv6_gro_complete, inet_gro_complete,
257					 skb, 0);
258		break;
259	}
260	rcu_read_unlock();
261
262	if (err) {
263		WARN_ON(&ptype->list == head);
264		kfree_skb(skb);
265		return;
266	}
267
268out:
269	gro_normal_one(napi, skb, NAPI_GRO_CB(skb)->count);
270}
271
272static void __napi_gro_flush_chain(struct napi_struct *napi, u32 index,
273				   bool flush_old)
274{
275	struct list_head *head = &napi->gro_hash[index].list;
276	struct sk_buff *skb, *p;
277
278	list_for_each_entry_safe_reverse(skb, p, head, list) {
279		if (flush_old && NAPI_GRO_CB(skb)->age == jiffies)
280			return;
281		skb_list_del_init(skb);
282		napi_gro_complete(napi, skb);
283		napi->gro_hash[index].count--;
284	}
285
286	if (!napi->gro_hash[index].count)
287		__clear_bit(index, &napi->gro_bitmask);
288}
289
290/* napi->gro_hash[].list contains packets ordered by age.
291 * youngest packets at the head of it.
292 * Complete skbs in reverse order to reduce latencies.
293 */
294void napi_gro_flush(struct napi_struct *napi, bool flush_old)
295{
296	unsigned long bitmask = napi->gro_bitmask;
297	unsigned int i, base = ~0U;
298
299	while ((i = ffs(bitmask)) != 0) {
300		bitmask >>= i;
301		base += i;
302		__napi_gro_flush_chain(napi, base, flush_old);
303	}
304}
305EXPORT_SYMBOL(napi_gro_flush);
306
307static unsigned long gro_list_prepare_tc_ext(const struct sk_buff *skb,
308					     const struct sk_buff *p,
309					     unsigned long diffs)
310{
311#if IS_ENABLED(CONFIG_NET_TC_SKB_EXT)
312	struct tc_skb_ext *skb_ext;
313	struct tc_skb_ext *p_ext;
314
315	skb_ext = skb_ext_find(skb, TC_SKB_EXT);
316	p_ext = skb_ext_find(p, TC_SKB_EXT);
317
318	diffs |= (!!p_ext) ^ (!!skb_ext);
319	if (!diffs && unlikely(skb_ext))
320		diffs |= p_ext->chain ^ skb_ext->chain;
321#endif
322	return diffs;
323}
324
325static void gro_list_prepare(const struct list_head *head,
326			     const struct sk_buff *skb)
327{
328	unsigned int maclen = skb->dev->hard_header_len;
329	u32 hash = skb_get_hash_raw(skb);
330	struct sk_buff *p;
331
332	list_for_each_entry(p, head, list) {
333		unsigned long diffs;
334
335		NAPI_GRO_CB(p)->flush = 0;
336
337		if (hash != skb_get_hash_raw(p)) {
338			NAPI_GRO_CB(p)->same_flow = 0;
339			continue;
340		}
341
342		diffs = (unsigned long)p->dev ^ (unsigned long)skb->dev;
343		diffs |= p->vlan_all ^ skb->vlan_all;
344		diffs |= skb_metadata_differs(p, skb);
345		if (maclen == ETH_HLEN)
346			diffs |= compare_ether_header(skb_mac_header(p),
347						      skb_mac_header(skb));
348		else if (!diffs)
349			diffs = memcmp(skb_mac_header(p),
350				       skb_mac_header(skb),
351				       maclen);
352
353		/* in most common scenarions 'slow_gro' is 0
354		 * otherwise we are already on some slower paths
355		 * either skip all the infrequent tests altogether or
356		 * avoid trying too hard to skip each of them individually
357		 */
358		if (!diffs && unlikely(skb->slow_gro | p->slow_gro)) {
359			diffs |= p->sk != skb->sk;
360			diffs |= skb_metadata_dst_cmp(p, skb);
361			diffs |= skb_get_nfct(p) ^ skb_get_nfct(skb);
362
363			diffs |= gro_list_prepare_tc_ext(skb, p, diffs);
364		}
365
366		NAPI_GRO_CB(p)->same_flow = !diffs;
367	}
368}
369
370static inline void skb_gro_reset_offset(struct sk_buff *skb, u32 nhoff)
371{
372	const struct skb_shared_info *pinfo = skb_shinfo(skb);
373	const skb_frag_t *frag0 = &pinfo->frags[0];
 
374
 
375	NAPI_GRO_CB(skb)->data_offset = 0;
376	NAPI_GRO_CB(skb)->frag0 = NULL;
377	NAPI_GRO_CB(skb)->frag0_len = 0;
 
 
 
378
379	if (!skb_headlen(skb) && pinfo->nr_frags &&
 
 
 
380	    !PageHighMem(skb_frag_page(frag0)) &&
381	    (!NET_IP_ALIGN || !((skb_frag_off(frag0) + nhoff) & 3))) {
382		NAPI_GRO_CB(skb)->frag0 = skb_frag_address(frag0);
383		NAPI_GRO_CB(skb)->frag0_len = min_t(unsigned int,
384						    skb_frag_size(frag0),
385						    skb->end - skb->tail);
386	}
387}
388
389static void gro_pull_from_frag0(struct sk_buff *skb, int grow)
390{
391	struct skb_shared_info *pinfo = skb_shinfo(skb);
392
393	BUG_ON(skb->end - skb->tail < grow);
394
395	memcpy(skb_tail_pointer(skb), NAPI_GRO_CB(skb)->frag0, grow);
396
397	skb->data_len -= grow;
398	skb->tail += grow;
399
400	skb_frag_off_add(&pinfo->frags[0], grow);
401	skb_frag_size_sub(&pinfo->frags[0], grow);
402
403	if (unlikely(!skb_frag_size(&pinfo->frags[0]))) {
404		skb_frag_unref(skb, 0);
405		memmove(pinfo->frags, pinfo->frags + 1,
406			--pinfo->nr_frags * sizeof(pinfo->frags[0]));
407	}
408}
409
410static void gro_try_pull_from_frag0(struct sk_buff *skb)
411{
412	int grow = skb_gro_offset(skb) - skb_headlen(skb);
413
414	if (grow > 0)
415		gro_pull_from_frag0(skb, grow);
416}
417
418static void gro_flush_oldest(struct napi_struct *napi, struct list_head *head)
419{
420	struct sk_buff *oldest;
421
422	oldest = list_last_entry(head, struct sk_buff, list);
423
424	/* We are called with head length >= MAX_GRO_SKBS, so this is
425	 * impossible.
426	 */
427	if (WARN_ON_ONCE(!oldest))
428		return;
429
430	/* Do not adjust napi->gro_hash[].count, caller is adding a new
431	 * SKB to the chain.
432	 */
433	skb_list_del_init(oldest);
434	napi_gro_complete(napi, oldest);
435}
436
437static enum gro_result dev_gro_receive(struct napi_struct *napi, struct sk_buff *skb)
438{
439	u32 bucket = skb_get_hash_raw(skb) & (GRO_HASH_BUCKETS - 1);
440	struct gro_list *gro_list = &napi->gro_hash[bucket];
441	struct list_head *head = &offload_base;
442	struct packet_offload *ptype;
443	__be16 type = skb->protocol;
444	struct sk_buff *pp = NULL;
445	enum gro_result ret;
446	int same_flow;
447
448	if (netif_elide_gro(skb->dev))
449		goto normal;
450
451	gro_list_prepare(&gro_list->list, skb);
452
453	rcu_read_lock();
454	list_for_each_entry_rcu(ptype, head, list) {
455		if (ptype->type == type && ptype->callbacks.gro_receive)
456			goto found_ptype;
457	}
458	rcu_read_unlock();
459	goto normal;
460
461found_ptype:
462	skb_set_network_header(skb, skb_gro_offset(skb));
463	skb_reset_mac_len(skb);
464	BUILD_BUG_ON(sizeof_field(struct napi_gro_cb, zeroed) != sizeof(u32));
465	BUILD_BUG_ON(!IS_ALIGNED(offsetof(struct napi_gro_cb, zeroed),
466					sizeof(u32))); /* Avoid slow unaligned acc */
467	*(u32 *)&NAPI_GRO_CB(skb)->zeroed = 0;
468	NAPI_GRO_CB(skb)->flush = skb_has_frag_list(skb);
469	NAPI_GRO_CB(skb)->is_atomic = 1;
470	NAPI_GRO_CB(skb)->count = 1;
471	if (unlikely(skb_is_gso(skb))) {
472		NAPI_GRO_CB(skb)->count = skb_shinfo(skb)->gso_segs;
473		/* Only support TCP and non DODGY users. */
474		if (!skb_is_gso_tcp(skb) ||
475		    (skb_shinfo(skb)->gso_type & SKB_GSO_DODGY))
476			NAPI_GRO_CB(skb)->flush = 1;
477	}
478
479	/* Setup for GRO checksum validation */
480	switch (skb->ip_summed) {
481	case CHECKSUM_COMPLETE:
482		NAPI_GRO_CB(skb)->csum = skb->csum;
483		NAPI_GRO_CB(skb)->csum_valid = 1;
484		break;
485	case CHECKSUM_UNNECESSARY:
486		NAPI_GRO_CB(skb)->csum_cnt = skb->csum_level + 1;
487		break;
488	}
489
490	pp = INDIRECT_CALL_INET(ptype->callbacks.gro_receive,
491				ipv6_gro_receive, inet_gro_receive,
492				&gro_list->list, skb);
493
494	rcu_read_unlock();
495
496	if (PTR_ERR(pp) == -EINPROGRESS) {
497		ret = GRO_CONSUMED;
498		goto ok;
499	}
500
501	same_flow = NAPI_GRO_CB(skb)->same_flow;
502	ret = NAPI_GRO_CB(skb)->free ? GRO_MERGED_FREE : GRO_MERGED;
503
504	if (pp) {
505		skb_list_del_init(pp);
506		napi_gro_complete(napi, pp);
507		gro_list->count--;
508	}
509
510	if (same_flow)
511		goto ok;
512
513	if (NAPI_GRO_CB(skb)->flush)
514		goto normal;
515
516	if (unlikely(gro_list->count >= MAX_GRO_SKBS))
517		gro_flush_oldest(napi, &gro_list->list);
518	else
519		gro_list->count++;
520
521	/* Must be called before setting NAPI_GRO_CB(skb)->{age|last} */
522	gro_try_pull_from_frag0(skb);
523	NAPI_GRO_CB(skb)->age = jiffies;
524	NAPI_GRO_CB(skb)->last = skb;
525	if (!skb_is_gso(skb))
526		skb_shinfo(skb)->gso_size = skb_gro_len(skb);
527	list_add(&skb->list, &gro_list->list);
528	ret = GRO_HELD;
529ok:
530	if (gro_list->count) {
531		if (!test_bit(bucket, &napi->gro_bitmask))
532			__set_bit(bucket, &napi->gro_bitmask);
533	} else if (test_bit(bucket, &napi->gro_bitmask)) {
534		__clear_bit(bucket, &napi->gro_bitmask);
535	}
536
537	return ret;
538
539normal:
540	ret = GRO_NORMAL;
541	gro_try_pull_from_frag0(skb);
542	goto ok;
543}
544
545struct packet_offload *gro_find_receive_by_type(__be16 type)
546{
547	struct list_head *offload_head = &offload_base;
548	struct packet_offload *ptype;
549
550	list_for_each_entry_rcu(ptype, offload_head, list) {
551		if (ptype->type != type || !ptype->callbacks.gro_receive)
552			continue;
553		return ptype;
554	}
555	return NULL;
556}
557EXPORT_SYMBOL(gro_find_receive_by_type);
558
559struct packet_offload *gro_find_complete_by_type(__be16 type)
560{
561	struct list_head *offload_head = &offload_base;
562	struct packet_offload *ptype;
563
564	list_for_each_entry_rcu(ptype, offload_head, list) {
565		if (ptype->type != type || !ptype->callbacks.gro_complete)
566			continue;
567		return ptype;
568	}
569	return NULL;
570}
571EXPORT_SYMBOL(gro_find_complete_by_type);
572
573static gro_result_t napi_skb_finish(struct napi_struct *napi,
574				    struct sk_buff *skb,
575				    gro_result_t ret)
576{
577	switch (ret) {
578	case GRO_NORMAL:
579		gro_normal_one(napi, skb, 1);
580		break;
581
582	case GRO_MERGED_FREE:
583		if (NAPI_GRO_CB(skb)->free == NAPI_GRO_FREE_STOLEN_HEAD)
584			napi_skb_free_stolen_head(skb);
585		else if (skb->fclone != SKB_FCLONE_UNAVAILABLE)
586			__kfree_skb(skb);
587		else
588			__napi_kfree_skb(skb, SKB_CONSUMED);
589		break;
590
591	case GRO_HELD:
592	case GRO_MERGED:
593	case GRO_CONSUMED:
594		break;
595	}
596
597	return ret;
598}
599
600gro_result_t napi_gro_receive(struct napi_struct *napi, struct sk_buff *skb)
601{
602	gro_result_t ret;
603
604	skb_mark_napi_id(skb, napi);
605	trace_napi_gro_receive_entry(skb);
606
607	skb_gro_reset_offset(skb, 0);
608
609	ret = napi_skb_finish(napi, skb, dev_gro_receive(napi, skb));
610	trace_napi_gro_receive_exit(ret);
611
612	return ret;
613}
614EXPORT_SYMBOL(napi_gro_receive);
615
616static void napi_reuse_skb(struct napi_struct *napi, struct sk_buff *skb)
617{
618	if (unlikely(skb->pfmemalloc)) {
619		consume_skb(skb);
620		return;
621	}
622	__skb_pull(skb, skb_headlen(skb));
623	/* restore the reserve we had after netdev_alloc_skb_ip_align() */
624	skb_reserve(skb, NET_SKB_PAD + NET_IP_ALIGN - skb_headroom(skb));
625	__vlan_hwaccel_clear_tag(skb);
626	skb->dev = napi->dev;
627	skb->skb_iif = 0;
628
629	/* eth_type_trans() assumes pkt_type is PACKET_HOST */
630	skb->pkt_type = PACKET_HOST;
631
632	skb->encapsulation = 0;
 
633	skb_shinfo(skb)->gso_type = 0;
634	skb_shinfo(skb)->gso_size = 0;
635	if (unlikely(skb->slow_gro)) {
636		skb_orphan(skb);
637		skb_ext_reset(skb);
638		nf_reset_ct(skb);
639		skb->slow_gro = 0;
640	}
641
642	napi->skb = skb;
643}
644
645struct sk_buff *napi_get_frags(struct napi_struct *napi)
646{
647	struct sk_buff *skb = napi->skb;
648
649	if (!skb) {
650		skb = napi_alloc_skb(napi, GRO_MAX_HEAD);
651		if (skb) {
652			napi->skb = skb;
653			skb_mark_napi_id(skb, napi);
654		}
655	}
656	return skb;
657}
658EXPORT_SYMBOL(napi_get_frags);
659
660static gro_result_t napi_frags_finish(struct napi_struct *napi,
661				      struct sk_buff *skb,
662				      gro_result_t ret)
663{
664	switch (ret) {
665	case GRO_NORMAL:
666	case GRO_HELD:
667		__skb_push(skb, ETH_HLEN);
668		skb->protocol = eth_type_trans(skb, skb->dev);
669		if (ret == GRO_NORMAL)
670			gro_normal_one(napi, skb, 1);
671		break;
672
673	case GRO_MERGED_FREE:
674		if (NAPI_GRO_CB(skb)->free == NAPI_GRO_FREE_STOLEN_HEAD)
675			napi_skb_free_stolen_head(skb);
676		else
677			napi_reuse_skb(napi, skb);
678		break;
679
680	case GRO_MERGED:
681	case GRO_CONSUMED:
682		break;
683	}
684
685	return ret;
686}
687
688/* Upper GRO stack assumes network header starts at gro_offset=0
689 * Drivers could call both napi_gro_frags() and napi_gro_receive()
690 * We copy ethernet header into skb->data to have a common layout.
691 */
692static struct sk_buff *napi_frags_skb(struct napi_struct *napi)
693{
694	struct sk_buff *skb = napi->skb;
695	const struct ethhdr *eth;
696	unsigned int hlen = sizeof(*eth);
697
698	napi->skb = NULL;
699
700	skb_reset_mac_header(skb);
701	skb_gro_reset_offset(skb, hlen);
702
703	if (unlikely(skb_gro_header_hard(skb, hlen))) {
704		eth = skb_gro_header_slow(skb, hlen, 0);
705		if (unlikely(!eth)) {
706			net_warn_ratelimited("%s: dropping impossible skb from %s\n",
707					     __func__, napi->dev->name);
708			napi_reuse_skb(napi, skb);
709			return NULL;
710		}
711	} else {
712		eth = (const struct ethhdr *)skb->data;
713		gro_pull_from_frag0(skb, hlen);
 
 
 
714		NAPI_GRO_CB(skb)->frag0 += hlen;
715		NAPI_GRO_CB(skb)->frag0_len -= hlen;
716	}
717	__skb_pull(skb, hlen);
718
719	/*
720	 * This works because the only protocols we care about don't require
721	 * special handling.
722	 * We'll fix it up properly in napi_frags_finish()
723	 */
724	skb->protocol = eth->h_proto;
725
726	return skb;
727}
728
729gro_result_t napi_gro_frags(struct napi_struct *napi)
730{
731	gro_result_t ret;
732	struct sk_buff *skb = napi_frags_skb(napi);
733
734	trace_napi_gro_frags_entry(skb);
735
736	ret = napi_frags_finish(napi, skb, dev_gro_receive(napi, skb));
737	trace_napi_gro_frags_exit(ret);
738
739	return ret;
740}
741EXPORT_SYMBOL(napi_gro_frags);
742
743/* Compute the checksum from gro_offset and return the folded value
744 * after adding in any pseudo checksum.
745 */
746__sum16 __skb_gro_checksum_complete(struct sk_buff *skb)
747{
748	__wsum wsum;
749	__sum16 sum;
750
751	wsum = skb_checksum(skb, skb_gro_offset(skb), skb_gro_len(skb), 0);
752
753	/* NAPI_GRO_CB(skb)->csum holds pseudo checksum */
754	sum = csum_fold(csum_add(NAPI_GRO_CB(skb)->csum, wsum));
755	/* See comments in __skb_checksum_complete(). */
756	if (likely(!sum)) {
757		if (unlikely(skb->ip_summed == CHECKSUM_COMPLETE) &&
758		    !skb->csum_complete_sw)
759			netdev_rx_csum_fault(skb->dev, skb);
760	}
761
762	NAPI_GRO_CB(skb)->csum = wsum;
763	NAPI_GRO_CB(skb)->csum_valid = 1;
764
765	return sum;
766}
767EXPORT_SYMBOL(__skb_gro_checksum_complete);