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/indirect_call_wrapper.h>
  14 #include <linux/skbuff.h>
  15 #include <net/tcp.h>
  16 #include <net/protocol.h>
  17
  18 static void tcp_gso_tstamp(struct sk_buff *skb, unsigned int ts_seq,
  19                            unsigned int seq, unsigned int mss)
  20 {
  21         while (skb) {
  22                 if (before(ts_seq, seq + mss)) {
  23                         skb_shinfo(skb)->tx_flags |= SKBTX_SW_TSTAMP;
  24                         skb_shinfo(skb)->tskey = ts_seq;
  25                         return;
  26                 }
  27
  28                 skb = skb->next;
  29                 seq += mss;
  30         }
  31 }
  32
  33 static struct sk_buff *tcp4_gso_segment(struct sk_buff *skb,
  34                                         netdev_features_t features)
  35 {
  36         if (!(skb_shinfo(skb)->gso_type & SKB_GSO_TCPV4))
  37                 return ERR_PTR(-EINVAL);
  38
  39         if (!pskb_may_pull(skb, sizeof(struct tcphdr)))
  40                 return ERR_PTR(-EINVAL);
  41
  42         if (unlikely(skb->ip_summed != CHECKSUM_PARTIAL)) {
  43                 const struct iphdr *iph = ip_hdr(skb);
  44                 struct tcphdr *th = tcp_hdr(skb);
  45
  46                 /* Set up checksum pseudo header, usually expect stack to
  47                  * have done this already.
  48                  */
  49
  50                 th->check = 0;
  51                 skb->ip_summed = CHECKSUM_PARTIAL;
  52                 __tcp_v4_send_check(skb, iph->saddr, iph->daddr);
  53         }
  54
  55         return tcp_gso_segment(skb, features);
  56 }
  57
  58 struct sk_buff *tcp_gso_segment(struct sk_buff *skb,
  59                                 netdev_features_t features)
  60 {
  61         struct sk_buff *segs = ERR_PTR(-EINVAL);
  62         unsigned int sum_truesize = 0;
  63         struct tcphdr *th;
  64         unsigned int thlen;
  65         unsigned int seq;
  66         __be32 delta;
  67         unsigned int oldlen;
  68         unsigned int mss;
  69         struct sk_buff *gso_skb = skb;
  70         __sum16 newcheck;
  71         bool ooo_okay, copy_destructor;
  72
  73         th = tcp_hdr(skb);
  74         thlen = th->doff * 4;
  75         if (thlen < sizeof(*th))
  76                 goto out;
  77
  78         if (!pskb_may_pull(skb, thlen))
  79                 goto out;
  80
  81         oldlen = (u16)~skb->len;
  82         __skb_pull(skb, thlen);
  83
  84         mss = skb_shinfo(skb)->gso_size;
  85         if (unlikely(skb->len <= mss))
  86                 goto out;
  87
  88         if (skb_gso_ok(skb, features | NETIF_F_GSO_ROBUST)) {
  89                 /* Packet is from an untrusted source, reset gso_segs. */
  90
  91                 skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss);
  92
  93                 segs = NULL;
  94                 goto out;
  95         }
  96
  97         copy_destructor = gso_skb->destructor == tcp_wfree;
  98         ooo_okay = gso_skb->ooo_okay;
  99         /* All segments but the first should have ooo_okay cleared */
 100         skb->ooo_okay = 0;
 101
 102         segs = skb_segment(skb, features);
 103         if (IS_ERR(segs))
 104                 goto out;
 105
 106         /* Only first segment might have ooo_okay set */
 107         segs->ooo_okay = ooo_okay;
 108
 109         /* GSO partial and frag_list segmentation only requires splitting
 110          * the frame into an MSS multiple and possibly a remainder, both
 111          * cases return a GSO skb. So update the mss now.
 112          */
 113         if (skb_is_gso(segs))
 114                 mss *= skb_shinfo(segs)->gso_segs;
 115
 116         delta = htonl(oldlen + (thlen + mss));
 117
 118         skb = segs;
 119         th = tcp_hdr(skb);
 120         seq = ntohl(th->seq);
 121
 122         if (unlikely(skb_shinfo(gso_skb)->tx_flags & SKBTX_SW_TSTAMP))
 123                 tcp_gso_tstamp(segs, skb_shinfo(gso_skb)->tskey, seq, mss);
 124
 125         newcheck = ~csum_fold((__force __wsum)((__force u32)th->check +
 126                                                (__force u32)delta));
 127
 128         while (skb->next) {
 129                 th->fin = th->psh = 0;
 130                 th->check = newcheck;
 131
 132                 if (skb->ip_summed == CHECKSUM_PARTIAL)
 133                         gso_reset_checksum(skb, ~th->check);
 134                 else
 135                         th->check = gso_make_checksum(skb, ~th->check);
 136
 137                 seq += mss;
 138                 if (copy_destructor) {
 139                         skb->destructor = gso_skb->destructor;
 140                         skb->sk = gso_skb->sk;
 141                         sum_truesize += skb->truesize;
 142                 }
 143                 skb = skb->next;
 144                 th = tcp_hdr(skb);
 145
 146                 th->seq = htonl(seq);
 147                 th->cwr = 0;
 148         }
 149
 150         /* Following permits TCP Small Queues to work well with GSO :
 151          * The callback to TCP stack will be called at the time last frag
 152          * is freed at TX completion, and not right now when gso_skb
 153          * is freed by GSO engine
 154          */
 155         if (copy_destructor) {
 156                 int delta;
 157
 158                 swap(gso_skb->sk, skb->sk);
 159                 swap(gso_skb->destructor, skb->destructor);
 160                 sum_truesize += skb->truesize;
 161                 delta = sum_truesize - gso_skb->truesize;
 162                 /* In some pathological cases, delta can be negative.
 163                  * We need to either use refcount_add() or refcount_sub_and_test()
 164                  */
 165                 if (likely(delta >= 0))
 166                         refcount_add(delta, &skb->sk->sk_wmem_alloc);
 167                 else
 168                         WARN_ON_ONCE(refcount_sub_and_test(-delta, &skb->sk->sk_wmem_alloc));
 169         }
 170
 171         delta = htonl(oldlen + (skb_tail_pointer(skb) -
 172                                 skb_transport_header(skb)) +
 173                       skb->data_len);
 174         th->check = ~csum_fold((__force __wsum)((__force u32)th->check +
 175                                 (__force u32)delta));
 176         if (skb->ip_summed == CHECKSUM_PARTIAL)
 177                 gso_reset_checksum(skb, ~th->check);
 178         else
 179                 th->check = gso_make_checksum(skb, ~th->check);
 180 out:
 181         return segs;
 182 }
 183
 184 struct sk_buff *tcp_gro_receive(struct list_head *head, struct sk_buff *skb)
 185 {
 186         struct sk_buff *pp = NULL;
 187         struct sk_buff *p;
 188         struct tcphdr *th;
 189         struct tcphdr *th2;
 190         unsigned int len;
 191         unsigned int thlen;
 192         __be32 flags;
 193         unsigned int mss = 1;
 194         unsigned int hlen;
 195         unsigned int off;
 196         int flush = 1;
 197         int i;
 198
 199         off = skb_gro_offset(skb);
 200         hlen = off + sizeof(*th);
 201         th = skb_gro_header_fast(skb, off);
 202         if (skb_gro_header_hard(skb, hlen)) {
 203                 th = skb_gro_header_slow(skb, hlen, off);
 204                 if (unlikely(!th))
 205                         goto out;
 206         }
 207
 208         thlen = th->doff * 4;
 209         if (thlen < sizeof(*th))
 210                 goto out;
 211
 212         hlen = off + thlen;
 213         if (skb_gro_header_hard(skb, hlen)) {
 214                 th = skb_gro_header_slow(skb, hlen, off);
 215                 if (unlikely(!th))
 216                         goto out;
 217         }
 218
 219         skb_gro_pull(skb, thlen);
 220
 221         len = skb_gro_len(skb);
 222         flags = tcp_flag_word(th);
 223
 224         list_for_each_entry(p, head, list) {
 225                 if (!NAPI_GRO_CB(p)->same_flow)
 226                         continue;
 227
 228                 th2 = tcp_hdr(p);
 229
 230                 if (*(u32 *)&th->source ^ *(u32 *)&th2->source) {
 231                         NAPI_GRO_CB(p)->same_flow = 0;
 232                         continue;
 233                 }
 234
 235                 goto found;
 236         }
 237         p = NULL;
 238         goto out_check_final;
 239
 240 found:
 241         /* Include the IP ID check below from the inner most IP hdr */
 242         flush = NAPI_GRO_CB(p)->flush;
 243         flush |= (__force int)(flags & TCP_FLAG_CWR);
 244         flush |= (__force int)((flags ^ tcp_flag_word(th2)) &
 245                   ~(TCP_FLAG_CWR | TCP_FLAG_FIN | TCP_FLAG_PSH));
 246         flush |= (__force int)(th->ack_seq ^ th2->ack_seq);
 247         for (i = sizeof(*th); i < thlen; i += 4)
 248                 flush |= *(u32 *)((u8 *)th + i) ^
 249                          *(u32 *)((u8 *)th2 + i);
 250
 251         /* When we receive our second frame we can made a decision on if we
 252          * continue this flow as an atomic flow with a fixed ID or if we use
 253          * an incrementing ID.
 254          */
 255         if (NAPI_GRO_CB(p)->flush_id != 1 ||
 256             NAPI_GRO_CB(p)->count != 1 ||
 257             !NAPI_GRO_CB(p)->is_atomic)
 258                 flush |= NAPI_GRO_CB(p)->flush_id;
 259         else
 260                 NAPI_GRO_CB(p)->is_atomic = false;
 261
 262         mss = skb_shinfo(p)->gso_size;
 263
 264         flush |= (len - 1) >= mss;
 265         flush |= (ntohl(th2->seq) + skb_gro_len(p)) ^ ntohl(th->seq);
 266 #ifdef CONFIG_TLS_DEVICE
 267         flush |= p->decrypted ^ skb->decrypted;
 268 #endif
 269
 270         if (flush || skb_gro_receive(p, skb)) {
 271                 mss = 1;
 272                 goto out_check_final;
 273         }
 274
 275         tcp_flag_word(th2) |= flags & (TCP_FLAG_FIN | TCP_FLAG_PSH);
 276
 277 out_check_final:
 278         flush = len < mss;
 279         flush |= (__force int)(flags & (TCP_FLAG_URG | TCP_FLAG_PSH |
 280                                         TCP_FLAG_RST | TCP_FLAG_SYN |
 281                                         TCP_FLAG_FIN));
 282
 283         if (p && (!NAPI_GRO_CB(skb)->same_flow || flush))
 284                 pp = p;
 285
 286 out:
 287         NAPI_GRO_CB(skb)->flush |= (flush != 0);
 288
 289         return pp;
 290 }
 291
 292 int tcp_gro_complete(struct sk_buff *skb)
 293 {
 294         struct tcphdr *th = tcp_hdr(skb);
 295
 296         skb->csum_start = (unsigned char *)th - skb->head;
 297         skb->csum_offset = offsetof(struct tcphdr, check);
 298         skb->ip_summed = CHECKSUM_PARTIAL;
 299
 300         skb_shinfo(skb)->gso_segs = NAPI_GRO_CB(skb)->count;
 301
 302         if (th->cwr)
 303                 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
 304
 305         return 0;
 306 }
 307 EXPORT_SYMBOL(tcp_gro_complete);
 308
 309 INDIRECT_CALLABLE_SCOPE
 310 struct sk_buff *tcp4_gro_receive(struct list_head *head, struct sk_buff *skb)
 311 {
 312         /* Don't bother verifying checksum if we're going to flush anyway. */
 313         if (!NAPI_GRO_CB(skb)->flush &&
 314             skb_gro_checksum_validate(skb, IPPROTO_TCP,
 315                                       inet_gro_compute_pseudo)) {
 316                 NAPI_GRO_CB(skb)->flush = 1;
 317                 return NULL;
 318         }
 319
 320         return tcp_gro_receive(head, skb);
 321 }
 322
 323 INDIRECT_CALLABLE_SCOPE int tcp4_gro_complete(struct sk_buff *skb, int thoff)
 324 {
 325         const struct iphdr *iph = ip_hdr(skb);
 326         struct tcphdr *th = tcp_hdr(skb);
 327
 328         th->check = ~tcp_v4_check(skb->len - thoff, iph->saddr,
 329                                   iph->daddr, 0);
 330         skb_shinfo(skb)->gso_type |= SKB_GSO_TCPV4;
 331
 332         if (NAPI_GRO_CB(skb)->is_atomic)
 333                 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_FIXEDID;
 334
 335         return tcp_gro_complete(skb);
 336 }
 337
 338 static const struct net_offload tcpv4_offload = {
 339         .callbacks = {
 340                 .gso_segment    =       tcp4_gso_segment,
 341                 .gro_receive    =       tcp4_gro_receive,
 342                 .gro_complete   =       tcp4_gro_complete,
 343         },
 344 };
 345
 346 int __init tcpv4_offload_init(void)
 347 {
 348         return inet_add_offload(&tcpv4_offload, IPPROTO_TCP);
 349 }