2 * IPV4 GSO/GRO offload support
3 * Linux INET implementation
5 * This program is free software; you can redistribute it and/or
6 * modify it under the terms of the GNU General Public License
7 * as published by the Free Software Foundation; either version
8 * 2 of the License, or (at your option) any later version.
13 #include <linux/skbuff.h>
15 #include <net/protocol.h>
17 static struct sk_buff
*__skb_udp_tunnel_segment(struct sk_buff
*skb
,
18 netdev_features_t features
,
19 struct sk_buff
*(*gso_inner_segment
)(struct sk_buff
*skb
,
20 netdev_features_t features
),
21 __be16 new_protocol
, bool is_ipv6
)
23 int tnl_hlen
= skb_inner_mac_header(skb
) - skb_transport_header(skb
);
24 bool remcsum
, need_csum
, offload_csum
, gso_partial
;
25 struct sk_buff
*segs
= ERR_PTR(-EINVAL
);
26 struct udphdr
*uh
= udp_hdr(skb
);
27 u16 mac_offset
= skb
->mac_header
;
28 __be16 protocol
= skb
->protocol
;
29 u16 mac_len
= skb
->mac_len
;
30 int udp_offset
, outer_hlen
;
34 if (unlikely(!pskb_may_pull(skb
, tnl_hlen
)))
37 /* Adjust partial header checksum to negate old length.
38 * We cannot rely on the value contained in uh->len as it is
39 * possible that the actual value exceeds the boundaries of the
40 * 16 bit length field due to the header being added outside of an
41 * IP or IPv6 frame that was already limited to 64K - 1.
43 if (skb_shinfo(skb
)->gso_type
& SKB_GSO_PARTIAL
)
44 partial
= (__force __wsum
)uh
->len
;
46 partial
= (__force __wsum
)htonl(skb
->len
);
47 partial
= csum_sub(csum_unfold(uh
->check
), partial
);
49 /* setup inner skb. */
50 skb
->encapsulation
= 0;
51 SKB_GSO_CB(skb
)->encap_level
= 0;
52 __skb_pull(skb
, tnl_hlen
);
53 skb_reset_mac_header(skb
);
54 skb_set_network_header(skb
, skb_inner_network_offset(skb
));
55 skb
->mac_len
= skb_inner_network_offset(skb
);
56 skb
->protocol
= new_protocol
;
58 need_csum
= !!(skb_shinfo(skb
)->gso_type
& SKB_GSO_UDP_TUNNEL_CSUM
);
59 skb
->encap_hdr_csum
= need_csum
;
61 remcsum
= !!(skb_shinfo(skb
)->gso_type
& SKB_GSO_TUNNEL_REMCSUM
);
62 skb
->remcsum_offload
= remcsum
;
64 need_ipsec
= skb_dst(skb
) && dst_xfrm(skb_dst(skb
));
65 /* Try to offload checksum if possible */
66 offload_csum
= !!(need_csum
&&
69 (is_ipv6
? (NETIF_F_HW_CSUM
| NETIF_F_IPV6_CSUM
) :
70 (NETIF_F_HW_CSUM
| NETIF_F_IP_CSUM
))));
72 features
&= skb
->dev
->hw_enc_features
;
74 /* The only checksum offload we care about from here on out is the
75 * outer one so strip the existing checksum feature flags and
76 * instead set the flag based on our outer checksum offload value.
79 features
&= ~NETIF_F_CSUM_MASK
;
80 if (!need_csum
|| offload_csum
)
81 features
|= NETIF_F_HW_CSUM
;
84 /* segment inner packet. */
85 segs
= gso_inner_segment(skb
, features
);
86 if (IS_ERR_OR_NULL(segs
)) {
87 skb_gso_error_unwind(skb
, protocol
, tnl_hlen
, mac_offset
,
92 gso_partial
= !!(skb_shinfo(segs
)->gso_type
& SKB_GSO_PARTIAL
);
94 outer_hlen
= skb_tnl_header_len(skb
);
95 udp_offset
= outer_hlen
- tnl_hlen
;
101 skb
->ip_summed
= CHECKSUM_NONE
;
103 /* Set up inner headers if we are offloading inner checksum */
104 if (skb
->ip_summed
== CHECKSUM_PARTIAL
) {
105 skb_reset_inner_headers(skb
);
106 skb
->encapsulation
= 1;
109 skb
->mac_len
= mac_len
;
110 skb
->protocol
= protocol
;
112 __skb_push(skb
, outer_hlen
);
113 skb_reset_mac_header(skb
);
114 skb_set_network_header(skb
, mac_len
);
115 skb_set_transport_header(skb
, udp_offset
);
116 len
= skb
->len
- udp_offset
;
119 /* If we are only performing partial GSO the inner header
120 * will be using a length value equal to only one MSS sized
121 * segment instead of the entire frame.
123 if (gso_partial
&& skb_is_gso(skb
)) {
124 uh
->len
= htons(skb_shinfo(skb
)->gso_size
+
125 SKB_GSO_CB(skb
)->data_offset
+
126 skb
->head
- (unsigned char *)uh
);
128 uh
->len
= htons(len
);
134 uh
->check
= ~csum_fold(csum_add(partial
,
135 (__force __wsum
)htonl(len
)));
137 if (skb
->encapsulation
|| !offload_csum
) {
138 uh
->check
= gso_make_checksum(skb
, ~uh
->check
);
140 uh
->check
= CSUM_MANGLED_0
;
142 skb
->ip_summed
= CHECKSUM_PARTIAL
;
143 skb
->csum_start
= skb_transport_header(skb
) - skb
->head
;
144 skb
->csum_offset
= offsetof(struct udphdr
, check
);
146 } while ((skb
= skb
->next
));
151 struct sk_buff
*skb_udp_tunnel_segment(struct sk_buff
*skb
,
152 netdev_features_t features
,
155 __be16 protocol
= skb
->protocol
;
156 const struct net_offload
**offloads
;
157 const struct net_offload
*ops
;
158 struct sk_buff
*segs
= ERR_PTR(-EINVAL
);
159 struct sk_buff
*(*gso_inner_segment
)(struct sk_buff
*skb
,
160 netdev_features_t features
);
164 switch (skb
->inner_protocol_type
) {
165 case ENCAP_TYPE_ETHER
:
166 protocol
= skb
->inner_protocol
;
167 gso_inner_segment
= skb_mac_gso_segment
;
169 case ENCAP_TYPE_IPPROTO
:
170 offloads
= is_ipv6
? inet6_offloads
: inet_offloads
;
171 ops
= rcu_dereference(offloads
[skb
->inner_ipproto
]);
172 if (!ops
|| !ops
->callbacks
.gso_segment
)
174 gso_inner_segment
= ops
->callbacks
.gso_segment
;
180 segs
= __skb_udp_tunnel_segment(skb
, features
, gso_inner_segment
,
188 EXPORT_SYMBOL(skb_udp_tunnel_segment
);
190 struct sk_buff
*__udp_gso_segment(struct sk_buff
*gso_skb
,
191 netdev_features_t features
)
193 struct sock
*sk
= gso_skb
->sk
;
194 unsigned int sum_truesize
= 0;
195 struct sk_buff
*segs
, *seg
;
202 mss
= skb_shinfo(gso_skb
)->gso_size
;
203 if (gso_skb
->len
<= sizeof(*uh
) + mss
)
204 return ERR_PTR(-EINVAL
);
206 skb_pull(gso_skb
, sizeof(*uh
));
208 /* clear destructor to avoid skb_segment assigning it to tail */
209 copy_dtor
= gso_skb
->destructor
== sock_wfree
;
211 gso_skb
->destructor
= NULL
;
213 segs
= skb_segment(gso_skb
, features
);
214 if (unlikely(IS_ERR_OR_NULL(segs
))) {
216 gso_skb
->destructor
= sock_wfree
;
220 /* GSO partial and frag_list segmentation only requires splitting
221 * the frame into an MSS multiple and possibly a remainder, both
222 * cases return a GSO skb. So update the mss now.
224 if (skb_is_gso(segs
))
225 mss
*= skb_shinfo(segs
)->gso_segs
;
230 /* compute checksum adjustment based on old length versus new */
231 newlen
= htons(sizeof(*uh
) + mss
);
232 check
= csum16_add(csum16_sub(uh
->check
, uh
->len
), newlen
);
236 seg
->destructor
= sock_wfree
;
238 sum_truesize
+= seg
->truesize
;
247 if (seg
->ip_summed
== CHECKSUM_PARTIAL
)
248 gso_reset_checksum(seg
, ~check
);
250 uh
->check
= gso_make_checksum(seg
, ~check
) ? :
257 /* last packet can be partial gso_size, account for that in checksum */
258 newlen
= htons(skb_tail_pointer(seg
) - skb_transport_header(seg
) +
260 check
= csum16_add(csum16_sub(uh
->check
, uh
->len
), newlen
);
265 if (seg
->ip_summed
== CHECKSUM_PARTIAL
)
266 gso_reset_checksum(seg
, ~check
);
268 uh
->check
= gso_make_checksum(seg
, ~check
) ? : CSUM_MANGLED_0
;
270 /* update refcount for the packet */
272 int delta
= sum_truesize
- gso_skb
->truesize
;
274 /* In some pathological cases, delta can be negative.
275 * We need to either use refcount_add() or refcount_sub_and_test()
277 if (likely(delta
>= 0))
278 refcount_add(delta
, &sk
->sk_wmem_alloc
);
280 WARN_ON_ONCE(refcount_sub_and_test(-delta
, &sk
->sk_wmem_alloc
));
284 EXPORT_SYMBOL_GPL(__udp_gso_segment
);
286 static struct sk_buff
*udp4_ufo_fragment(struct sk_buff
*skb
,
287 netdev_features_t features
)
289 struct sk_buff
*segs
= ERR_PTR(-EINVAL
);
295 if (skb
->encapsulation
&&
296 (skb_shinfo(skb
)->gso_type
&
297 (SKB_GSO_UDP_TUNNEL
|SKB_GSO_UDP_TUNNEL_CSUM
))) {
298 segs
= skb_udp_tunnel_segment(skb
, features
, false);
302 if (!(skb_shinfo(skb
)->gso_type
& (SKB_GSO_UDP
| SKB_GSO_UDP_L4
)))
305 if (!pskb_may_pull(skb
, sizeof(struct udphdr
)))
308 if (skb_shinfo(skb
)->gso_type
& SKB_GSO_UDP_L4
)
309 return __udp_gso_segment(skb
, features
);
311 mss
= skb_shinfo(skb
)->gso_size
;
312 if (unlikely(skb
->len
<= mss
))
315 /* Do software UFO. Complete and fill in the UDP checksum as
316 * HW cannot do checksum of UDP packets sent as multiple
324 csum
= skb_checksum(skb
, 0, skb
->len
, 0);
325 uh
->check
= udp_v4_check(skb
->len
, iph
->saddr
, iph
->daddr
, csum
);
327 uh
->check
= CSUM_MANGLED_0
;
329 skb
->ip_summed
= CHECKSUM_UNNECESSARY
;
331 /* If there is no outer header we can fake a checksum offload
332 * due to the fact that we have already done the checksum in
333 * software prior to segmenting the frame.
335 if (!skb
->encap_hdr_csum
)
336 features
|= NETIF_F_HW_CSUM
;
338 /* Fragment the skb. IP headers of the fragments are updated in
341 segs
= skb_segment(skb
, features
);
346 struct sk_buff
*udp_gro_receive(struct list_head
*head
, struct sk_buff
*skb
,
347 struct udphdr
*uh
, udp_lookup_t lookup
)
349 struct sk_buff
*pp
= NULL
;
352 unsigned int off
= skb_gro_offset(skb
);
356 if (NAPI_GRO_CB(skb
)->encap_mark
||
357 (skb
->ip_summed
!= CHECKSUM_PARTIAL
&&
358 NAPI_GRO_CB(skb
)->csum_cnt
== 0 &&
359 !NAPI_GRO_CB(skb
)->csum_valid
))
362 /* mark that this skb passed once through the tunnel gro layer */
363 NAPI_GRO_CB(skb
)->encap_mark
= 1;
366 sk
= (*lookup
)(skb
, uh
->source
, uh
->dest
);
368 if (sk
&& udp_sk(sk
)->gro_receive
)
375 list_for_each_entry(p
, head
, list
) {
376 if (!NAPI_GRO_CB(p
)->same_flow
)
379 uh2
= (struct udphdr
*)(p
->data
+ off
);
381 /* Match ports and either checksums are either both zero
384 if ((*(u32
*)&uh
->source
!= *(u32
*)&uh2
->source
) ||
385 (!uh
->check
^ !uh2
->check
)) {
386 NAPI_GRO_CB(p
)->same_flow
= 0;
391 skb_gro_pull(skb
, sizeof(struct udphdr
)); /* pull encapsulating udp header */
392 skb_gro_postpull_rcsum(skb
, uh
, sizeof(struct udphdr
));
393 pp
= call_gro_receive_sk(udp_sk(sk
)->gro_receive
, sk
, head
, skb
);
398 skb_gro_flush_final(skb
, pp
, flush
);
401 EXPORT_SYMBOL(udp_gro_receive
);
403 static struct sk_buff
*udp4_gro_receive(struct list_head
*head
,
406 struct udphdr
*uh
= udp_gro_udphdr(skb
);
411 /* Don't bother verifying checksum if we're going to flush anyway. */
412 if (NAPI_GRO_CB(skb
)->flush
)
415 if (skb_gro_checksum_validate_zero_check(skb
, IPPROTO_UDP
, uh
->check
,
416 inet_gro_compute_pseudo
))
419 skb_gro_checksum_try_convert(skb
, IPPROTO_UDP
, uh
->check
,
420 inet_gro_compute_pseudo
);
422 NAPI_GRO_CB(skb
)->is_ipv6
= 0;
423 return udp_gro_receive(head
, skb
, uh
, udp4_lib_lookup_skb
);
426 NAPI_GRO_CB(skb
)->flush
= 1;
430 int udp_gro_complete(struct sk_buff
*skb
, int nhoff
,
433 __be16 newlen
= htons(skb
->len
- nhoff
);
434 struct udphdr
*uh
= (struct udphdr
*)(skb
->data
+ nhoff
);
440 /* Set encapsulation before calling into inner gro_complete() functions
441 * to make them set up the inner offsets.
443 skb
->encapsulation
= 1;
446 sk
= (*lookup
)(skb
, uh
->source
, uh
->dest
);
447 if (sk
&& udp_sk(sk
)->gro_complete
)
448 err
= udp_sk(sk
)->gro_complete(sk
, skb
,
449 nhoff
+ sizeof(struct udphdr
));
452 if (skb
->remcsum_offload
)
453 skb_shinfo(skb
)->gso_type
|= SKB_GSO_TUNNEL_REMCSUM
;
457 EXPORT_SYMBOL(udp_gro_complete
);
459 static int udp4_gro_complete(struct sk_buff
*skb
, int nhoff
)
461 const struct iphdr
*iph
= ip_hdr(skb
);
462 struct udphdr
*uh
= (struct udphdr
*)(skb
->data
+ nhoff
);
465 skb_shinfo(skb
)->gso_type
|= SKB_GSO_UDP_TUNNEL_CSUM
;
466 uh
->check
= ~udp_v4_check(skb
->len
- nhoff
, iph
->saddr
,
469 skb_shinfo(skb
)->gso_type
|= SKB_GSO_UDP_TUNNEL
;
472 return udp_gro_complete(skb
, nhoff
, udp4_lib_lookup_skb
);
475 static const struct net_offload udpv4_offload
= {
477 .gso_segment
= udp4_ufo_fragment
,
478 .gro_receive
= udp4_gro_receive
,
479 .gro_complete
= udp4_gro_complete
,
483 int __init
udpv4_offload_init(void)
485 return inet_add_offload(&udpv4_offload
, IPPROTO_UDP
);