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.
10 * TCPv4 GSO/GRO support
13 #include <linux/skbuff.h>
15 #include <net/protocol.h>
17 static void tcp_gso_tstamp(struct sk_buff *skb, unsigned int ts_seq,
18 unsigned int seq, unsigned int mss)
21 if (before(ts_seq, seq + mss)) {
22 skb_shinfo(skb)->tx_flags |= SKBTX_SW_TSTAMP;
23 skb_shinfo(skb)->tskey = ts_seq;
32 static struct sk_buff *tcp4_gso_segment(struct sk_buff *skb,
33 netdev_features_t features)
35 if (!(skb_shinfo(skb)->gso_type & SKB_GSO_TCPV4))
36 return ERR_PTR(-EINVAL);
38 if (!pskb_may_pull(skb, sizeof(struct tcphdr)))
39 return ERR_PTR(-EINVAL);
41 if (unlikely(skb->ip_summed != CHECKSUM_PARTIAL)) {
42 const struct iphdr *iph = ip_hdr(skb);
43 struct tcphdr *th = tcp_hdr(skb);
45 /* Set up checksum pseudo header, usually expect stack to
46 * have done this already.
50 skb->ip_summed = CHECKSUM_PARTIAL;
51 __tcp_v4_send_check(skb, iph->saddr, iph->daddr);
54 return tcp_gso_segment(skb, features);
57 struct sk_buff *tcp_gso_segment(struct sk_buff *skb,
58 netdev_features_t features)
60 struct sk_buff *segs = ERR_PTR(-EINVAL);
61 unsigned int sum_truesize = 0;
68 struct sk_buff *gso_skb = skb;
70 bool ooo_okay, copy_destructor;
74 if (thlen < sizeof(*th))
77 if (!pskb_may_pull(skb, thlen))
80 oldlen = (u16)~skb->len;
81 __skb_pull(skb, thlen);
83 mss = skb_shinfo(skb)->gso_size;
84 if (unlikely(skb->len <= mss))
87 if (skb_gso_ok(skb, features | NETIF_F_GSO_ROBUST)) {
88 /* Packet is from an untrusted source, reset gso_segs. */
90 skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss);
96 copy_destructor = gso_skb->destructor == tcp_wfree;
97 ooo_okay = gso_skb->ooo_okay;
98 /* All segments but the first should have ooo_okay cleared */
101 segs = skb_segment(skb, features);
105 /* Only first segment might have ooo_okay set */
106 segs->ooo_okay = ooo_okay;
108 /* GSO partial and frag_list segmentation only requires splitting
109 * the frame into an MSS multiple and possibly a remainder, both
110 * cases return a GSO skb. So update the mss now.
112 if (skb_is_gso(segs))
113 mss *= skb_shinfo(segs)->gso_segs;
115 delta = htonl(oldlen + (thlen + mss));
119 seq = ntohl(th->seq);
121 if (unlikely(skb_shinfo(gso_skb)->tx_flags & SKBTX_SW_TSTAMP))
122 tcp_gso_tstamp(segs, skb_shinfo(gso_skb)->tskey, seq, mss);
124 newcheck = ~csum_fold((__force __wsum)((__force u32)th->check +
125 (__force u32)delta));
128 th->fin = th->psh = 0;
129 th->check = newcheck;
131 if (skb->ip_summed == CHECKSUM_PARTIAL)
132 gso_reset_checksum(skb, ~th->check);
134 th->check = gso_make_checksum(skb, ~th->check);
137 if (copy_destructor) {
138 skb->destructor = gso_skb->destructor;
139 skb->sk = gso_skb->sk;
140 sum_truesize += skb->truesize;
145 th->seq = htonl(seq);
149 /* Following permits TCP Small Queues to work well with GSO :
150 * The callback to TCP stack will be called at the time last frag
151 * is freed at TX completion, and not right now when gso_skb
152 * is freed by GSO engine
154 if (copy_destructor) {
155 swap(gso_skb->sk, skb->sk);
156 swap(gso_skb->destructor, skb->destructor);
157 sum_truesize += skb->truesize;
158 atomic_add(sum_truesize - gso_skb->truesize,
159 &skb->sk->sk_wmem_alloc);
162 delta = htonl(oldlen + (skb_tail_pointer(skb) -
163 skb_transport_header(skb)) +
165 th->check = ~csum_fold((__force __wsum)((__force u32)th->check +
166 (__force u32)delta));
167 if (skb->ip_summed == CHECKSUM_PARTIAL)
168 gso_reset_checksum(skb, ~th->check);
170 th->check = gso_make_checksum(skb, ~th->check);
175 struct sk_buff **tcp_gro_receive(struct sk_buff **head, struct sk_buff *skb)
177 struct sk_buff **pp = NULL;
184 unsigned int mss = 1;
190 off = skb_gro_offset(skb);
191 hlen = off + sizeof(*th);
192 th = skb_gro_header_fast(skb, off);
193 if (skb_gro_header_hard(skb, hlen)) {
194 th = skb_gro_header_slow(skb, hlen, off);
199 thlen = th->doff * 4;
200 if (thlen < sizeof(*th))
204 if (skb_gro_header_hard(skb, hlen)) {
205 th = skb_gro_header_slow(skb, hlen, off);
210 skb_gro_pull(skb, thlen);
212 len = skb_gro_len(skb);
213 flags = tcp_flag_word(th);
215 for (; (p = *head); head = &p->next) {
216 if (!NAPI_GRO_CB(p)->same_flow)
221 if (*(u32 *)&th->source ^ *(u32 *)&th2->source) {
222 NAPI_GRO_CB(p)->same_flow = 0;
229 goto out_check_final;
232 /* Include the IP ID check below from the inner most IP hdr */
233 flush = NAPI_GRO_CB(p)->flush;
234 flush |= (__force int)(flags & TCP_FLAG_CWR);
235 flush |= (__force int)((flags ^ tcp_flag_word(th2)) &
236 ~(TCP_FLAG_CWR | TCP_FLAG_FIN | TCP_FLAG_PSH));
237 flush |= (__force int)(th->ack_seq ^ th2->ack_seq);
238 for (i = sizeof(*th); i < thlen; i += 4)
239 flush |= *(u32 *)((u8 *)th + i) ^
240 *(u32 *)((u8 *)th2 + i);
242 /* When we receive our second frame we can made a decision on if we
243 * continue this flow as an atomic flow with a fixed ID or if we use
244 * an incrementing ID.
246 if (NAPI_GRO_CB(p)->flush_id != 1 ||
247 NAPI_GRO_CB(p)->count != 1 ||
248 !NAPI_GRO_CB(p)->is_atomic)
249 flush |= NAPI_GRO_CB(p)->flush_id;
251 NAPI_GRO_CB(p)->is_atomic = false;
253 mss = skb_shinfo(p)->gso_size;
255 flush |= (len - 1) >= mss;
256 flush |= (ntohl(th2->seq) + skb_gro_len(p)) ^ ntohl(th->seq);
258 if (flush || skb_gro_receive(head, skb)) {
260 goto out_check_final;
265 tcp_flag_word(th2) |= flags & (TCP_FLAG_FIN | TCP_FLAG_PSH);
269 flush |= (__force int)(flags & (TCP_FLAG_URG | TCP_FLAG_PSH |
270 TCP_FLAG_RST | TCP_FLAG_SYN |
273 if (p && (!NAPI_GRO_CB(skb)->same_flow || flush))
277 NAPI_GRO_CB(skb)->flush |= (flush != 0);
282 int tcp_gro_complete(struct sk_buff *skb)
284 struct tcphdr *th = tcp_hdr(skb);
286 skb->csum_start = (unsigned char *)th - skb->head;
287 skb->csum_offset = offsetof(struct tcphdr, check);
288 skb->ip_summed = CHECKSUM_PARTIAL;
290 skb_shinfo(skb)->gso_segs = NAPI_GRO_CB(skb)->count;
293 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
297 EXPORT_SYMBOL(tcp_gro_complete);
299 static struct sk_buff **tcp4_gro_receive(struct sk_buff **head, struct sk_buff *skb)
301 /* Don't bother verifying checksum if we're going to flush anyway. */
302 if (!NAPI_GRO_CB(skb)->flush &&
303 skb_gro_checksum_validate(skb, IPPROTO_TCP,
304 inet_gro_compute_pseudo)) {
305 NAPI_GRO_CB(skb)->flush = 1;
309 return tcp_gro_receive(head, skb);
312 static int tcp4_gro_complete(struct sk_buff *skb, int thoff)
314 const struct iphdr *iph = ip_hdr(skb);
315 struct tcphdr *th = tcp_hdr(skb);
317 th->check = ~tcp_v4_check(skb->len - thoff, iph->saddr,
319 skb_shinfo(skb)->gso_type |= SKB_GSO_TCPV4;
321 if (NAPI_GRO_CB(skb)->is_atomic)
322 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_FIXEDID;
324 return tcp_gro_complete(skb);
327 static const struct net_offload tcpv4_offload = {
329 .gso_segment = tcp4_gso_segment,
330 .gro_receive = tcp4_gro_receive,
331 .gro_complete = tcp4_gro_complete,
335 int __init tcpv4_offload_init(void)
337 return inet_add_offload(&tcpv4_offload, IPPROTO_TCP);