net/ipv4/tcp_offload.c

   1 /*
   2  *      IPV4 GSO/GRO offload support
   3  *      Linux INET implementation
   4  *
   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.
   9  *
  10  *      TCPv4 GSO/GRO support
  11  */
  12
  13 #include <linux/skbuff.h>
  14 #include <net/tcp.h>
  15 #include <net/protocol.h>
  16
  17 static void tcp_gso_tstamp(struct sk_buff *skb, unsigned int ts_seq,
  18                            unsigned int seq, unsigned int mss)
  19 {
  20         while (skb) {
  21                 if (before(ts_seq, seq + mss)) {
  22                         skb_shinfo(skb)->tx_flags |= SKBTX_SW_TSTAMP;
  23                         skb_shinfo(skb)->tskey = ts_seq;
  24                         return;
  25                 }
  26
  27                 skb = skb->next;
  28                 seq += mss;
  29         }
  30 }
  31
  32 static struct sk_buff *tcp4_gso_segment(struct sk_buff *skb,
  33                                         netdev_features_t features)
  34 {
  35         if (!pskb_may_pull(skb, sizeof(struct tcphdr)))
  36                 return ERR_PTR(-EINVAL);
  37
  38         if (unlikely(skb->ip_summed != CHECKSUM_PARTIAL)) {
  39                 const struct iphdr *iph = ip_hdr(skb);
  40                 struct tcphdr *th = tcp_hdr(skb);
  41
  42                 /* Set up checksum pseudo header, usually expect stack to
  43                  * have done this already.
  44                  */
  45
  46                 th->check = 0;
  47                 skb->ip_summed = CHECKSUM_PARTIAL;
  48                 __tcp_v4_send_check(skb, iph->saddr, iph->daddr);
  49         }
  50
  51         return tcp_gso_segment(skb, features);
  52 }
  53
  54 struct sk_buff *tcp_gso_segment(struct sk_buff *skb,
  55                                 netdev_features_t features)
  56 {
  57         struct sk_buff *segs = ERR_PTR(-EINVAL);
  58         unsigned int sum_truesize = 0;
  59         struct tcphdr *th;
  60         unsigned int thlen;
  61         unsigned int seq;
  62         __be32 delta;
  63         unsigned int oldlen;
  64         unsigned int mss;
  65         struct sk_buff *gso_skb = skb;
  66         __sum16 newcheck;
  67         bool ooo_okay, copy_destructor;
  68
  69         th = tcp_hdr(skb);
  70         thlen = th->doff * 4;
  71         if (thlen < sizeof(*th))
  72                 goto out;
  73
  74         if (!pskb_may_pull(skb, thlen))
  75                 goto out;
  76
  77         oldlen = (u16)~skb->len;
  78         __skb_pull(skb, thlen);
  79
  80         mss = skb_shinfo(skb)->gso_size;
  81         if (unlikely(skb->len <= mss))
  82                 goto out;
  83
  84         if (skb_gso_ok(skb, features | NETIF_F_GSO_ROBUST)) {
  85                 /* Packet is from an untrusted source, reset gso_segs. */
  86
  87                 skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss);
  88
  89                 segs = NULL;
  90                 goto out;
  91         }
  92
  93         copy_destructor = gso_skb->destructor == tcp_wfree;
  94         ooo_okay = gso_skb->ooo_okay;
  95         /* All segments but the first should have ooo_okay cleared */
  96         skb->ooo_okay = 0;
  97
  98         segs = skb_segment(skb, features);
  99         if (IS_ERR(segs))
 100                 goto out;
 101
 102         /* Only first segment might have ooo_okay set */
 103         segs->ooo_okay = ooo_okay;
 104
 105         /* GSO partial and frag_list segmentation only requires splitting
 106          * the frame into an MSS multiple and possibly a remainder, both
 107          * cases return a GSO skb. So update the mss now.
 108          */
 109         if (skb_is_gso(segs))
 110                 mss *= skb_shinfo(segs)->gso_segs;
 111
 112         delta = htonl(oldlen + (thlen + mss));
 113
 114         skb = segs;
 115         th = tcp_hdr(skb);
 116         seq = ntohl(th->seq);
 117
 118         if (unlikely(skb_shinfo(gso_skb)->tx_flags & SKBTX_SW_TSTAMP))
 119                 tcp_gso_tstamp(segs, skb_shinfo(gso_skb)->tskey, seq, mss);
 120
 121         newcheck = ~csum_fold((__force __wsum)((__force u32)th->check +
 122                                                (__force u32)delta));
 123
 124         while (skb->next) {
 125                 th->fin = th->psh = 0;
 126                 th->check = newcheck;
 127
 128                 if (skb->ip_summed == CHECKSUM_PARTIAL)
 129                         gso_reset_checksum(skb, ~th->check);
 130                 else
 131                         th->check = gso_make_checksum(skb, ~th->check);
 132
 133                 seq += mss;
 134                 if (copy_destructor) {
 135                         skb->destructor = gso_skb->destructor;
 136                         skb->sk = gso_skb->sk;
 137                         sum_truesize += skb->truesize;
 138                 }
 139                 skb = skb->next;
 140                 th = tcp_hdr(skb);
 141
 142                 th->seq = htonl(seq);
 143                 th->cwr = 0;
 144         }
 145
 146         /* Following permits TCP Small Queues to work well with GSO :
 147          * The callback to TCP stack will be called at the time last frag
 148          * is freed at TX completion, and not right now when gso_skb
 149          * is freed by GSO engine
 150          */
 151         if (copy_destructor) {
 152                 int delta;
 153
 154                 swap(gso_skb->sk, skb->sk);
 155                 swap(gso_skb->destructor, skb->destructor);
 156                 sum_truesize += skb->truesize;
 157                 delta = sum_truesize - gso_skb->truesize;
 158                 /* In some pathological cases, delta can be negative.
 159                  * We need to either use refcount_add() or refcount_sub_and_test()
 160                  */
 161                 if (likely(delta >= 0))
 162                         refcount_add(delta, &skb->sk->sk_wmem_alloc);
 163                 else
 164                         WARN_ON_ONCE(refcount_sub_and_test(-delta, &skb->sk->sk_wmem_alloc));
 165         }
 166
 167         delta = htonl(oldlen + (skb_tail_pointer(skb) -
 168                                 skb_transport_header(skb)) +
 169                       skb->data_len);
 170         th->check = ~csum_fold((__force __wsum)((__force u32)th->check +
 171                                 (__force u32)delta));
 172         if (skb->ip_summed == CHECKSUM_PARTIAL)
 173                 gso_reset_checksum(skb, ~th->check);
 174         else
 175                 th->check = gso_make_checksum(skb, ~th->check);
 176 out:
 177         return segs;
 178 }
 179
 180 struct sk_buff **tcp_gro_receive(struct sk_buff **head, struct sk_buff *skb)
 181 {
 182         struct sk_buff **pp = NULL;
 183         struct sk_buff *p;
 184         struct tcphdr *th;
 185         struct tcphdr *th2;
 186         unsigned int len;
 187         unsigned int thlen;
 188         __be32 flags;
 189         unsigned int mss = 1;
 190         unsigned int hlen;
 191         unsigned int off;
 192         int flush = 1;
 193         int i;
 194
 195         off = skb_gro_offset(skb);
 196         hlen = off + sizeof(*th);
 197         th = skb_gro_header_fast(skb, off);
 198         if (skb_gro_header_hard(skb, hlen)) {
 199                 th = skb_gro_header_slow(skb, hlen, off);
 200                 if (unlikely(!th))
 201                         goto out;
 202         }
 203
 204         thlen = th->doff * 4;
 205         if (thlen < sizeof(*th))
 206                 goto out;
 207
 208         hlen = off + thlen;
 209         if (skb_gro_header_hard(skb, hlen)) {
 210                 th = skb_gro_header_slow(skb, hlen, off);
 211                 if (unlikely(!th))
 212                         goto out;
 213         }
 214
 215         skb_gro_pull(skb, thlen);
 216
 217         len = skb_gro_len(skb);
 218         flags = tcp_flag_word(th);
 219
 220         for (; (p = *head); head = &p->next) {
 221                 if (!NAPI_GRO_CB(p)->same_flow)
 222                         continue;
 223
 224                 th2 = tcp_hdr(p);
 225
 226                 if (*(u32 *)&th->source ^ *(u32 *)&th2->source) {
 227                         NAPI_GRO_CB(p)->same_flow = 0;
 228                         continue;
 229                 }
 230
 231                 goto found;
 232         }
 233
 234         goto out_check_final;
 235
 236 found:
 237         /* Include the IP ID check below from the inner most IP hdr */
 238         flush = NAPI_GRO_CB(p)->flush;
 239         flush |= (__force int)(flags & TCP_FLAG_CWR);
 240         flush |= (__force int)((flags ^ tcp_flag_word(th2)) &
 241                   ~(TCP_FLAG_CWR | TCP_FLAG_FIN | TCP_FLAG_PSH));
 242         flush |= (__force int)(th->ack_seq ^ th2->ack_seq);
 243         for (i = sizeof(*th); i < thlen; i += 4)
 244                 flush |= *(u32 *)((u8 *)th + i) ^
 245                          *(u32 *)((u8 *)th2 + i);
 246
 247         /* When we receive our second frame we can made a decision on if we
 248          * continue this flow as an atomic flow with a fixed ID or if we use
 249          * an incrementing ID.
 250          */
 251         if (NAPI_GRO_CB(p)->flush_id != 1 ||
 252             NAPI_GRO_CB(p)->count != 1 ||
 253             !NAPI_GRO_CB(p)->is_atomic)
 254                 flush |= NAPI_GRO_CB(p)->flush_id;
 255         else
 256                 NAPI_GRO_CB(p)->is_atomic = false;
 257
 258         mss = skb_shinfo(p)->gso_size;
 259
 260         flush |= (len - 1) >= mss;
 261         flush |= (ntohl(th2->seq) + skb_gro_len(p)) ^ ntohl(th->seq);
 262
 263         if (flush || skb_gro_receive(head, skb)) {
 264                 mss = 1;
 265                 goto out_check_final;
 266         }
 267
 268         p = *head;
 269         th2 = tcp_hdr(p);
 270         tcp_flag_word(th2) |= flags & (TCP_FLAG_FIN | TCP_FLAG_PSH);
 271
 272 out_check_final:
 273         flush = len < mss;
 274         flush |= (__force int)(flags & (TCP_FLAG_URG | TCP_FLAG_PSH |
 275                                         TCP_FLAG_RST | TCP_FLAG_SYN |
 276                                         TCP_FLAG_FIN));
 277
 278         if (p && (!NAPI_GRO_CB(skb)->same_flow || flush))
 279                 pp = head;
 280
 281 out:
 282         NAPI_GRO_CB(skb)->flush |= (flush != 0);
 283
 284         return pp;
 285 }
 286
 287 int tcp_gro_complete(struct sk_buff *skb)
 288 {
 289         struct tcphdr *th = tcp_hdr(skb);
 290
 291         skb->csum_start = (unsigned char *)th - skb->head;
 292         skb->csum_offset = offsetof(struct tcphdr, check);
 293         skb->ip_summed = CHECKSUM_PARTIAL;
 294
 295         skb_shinfo(skb)->gso_segs = NAPI_GRO_CB(skb)->count;
 296
 297         if (th->cwr)
 298                 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
 299
 300         return 0;
 301 }
 302 EXPORT_SYMBOL(tcp_gro_complete);
 303
 304 static struct sk_buff **tcp4_gro_receive(struct sk_buff **head, struct sk_buff *skb)
 305 {
 306         /* Don't bother verifying checksum if we're going to flush anyway. */
 307         if (!NAPI_GRO_CB(skb)->flush &&
 308             skb_gro_checksum_validate(skb, IPPROTO_TCP,
 309                                       inet_gro_compute_pseudo)) {
 310                 NAPI_GRO_CB(skb)->flush = 1;
 311                 return NULL;
 312         }
 313
 314         return tcp_gro_receive(head, skb);
 315 }
 316
 317 static int tcp4_gro_complete(struct sk_buff *skb, int thoff)
 318 {
 319         const struct iphdr *iph = ip_hdr(skb);
 320         struct tcphdr *th = tcp_hdr(skb);
 321
 322         th->check = ~tcp_v4_check(skb->len - thoff, iph->saddr,
 323                                   iph->daddr, 0);
 324         skb_shinfo(skb)->gso_type |= SKB_GSO_TCPV4;
 325
 326         if (NAPI_GRO_CB(skb)->is_atomic)
 327                 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_FIXEDID;
 328
 329         return tcp_gro_complete(skb);
 330 }
 331
 332 static const struct net_offload tcpv4_offload = {
 333         .callbacks = {
 334                 .gso_segment    =       tcp4_gso_segment,
 335                 .gro_receive    =       tcp4_gro_receive,
 336                 .gro_complete   =       tcp4_gro_complete,
 337         },
 338 };
 339
 340 int __init tcpv4_offload_init(void)
 341 {
 342         return inet_add_offload(&tcpv4_offload, IPPROTO_TCP);
 343 }