net/ipv4/tcp_offload.c

   1 // SPDX-License-Identifier: GPL-2.0-or-later
   2 /*
   3  *      IPV4 GSO/GRO offload support
   4  *      Linux INET implementation
   5  *
   6  *      TCPv4 GSO/GRO support
   7  */
   8
   9 #include <linux/indirect_call_wrapper.h>
  10 #include <linux/skbuff.h>
  11 #include <net/gro.h>
  12 #include <net/gso.h>
  13 #include <net/tcp.h>
  14 #include <net/protocol.h>
  15
  16 static void tcp_gso_tstamp(struct sk_buff *skb, unsigned int ts_seq,
  17                            unsigned int seq, unsigned int mss)
  18 {
  19         while (skb) {
  20                 if (before(ts_seq, seq + mss)) {
  21                         skb_shinfo(skb)->tx_flags |= SKBTX_SW_TSTAMP;
  22                         skb_shinfo(skb)->tskey = ts_seq;
  23                         return;
  24                 }
  25
  26                 skb = skb->next;
  27                 seq += mss;
  28         }
  29 }
  30
  31 static struct sk_buff *tcp4_gso_segment(struct sk_buff *skb,
  32                                         netdev_features_t features)
  33 {
  34         if (!(skb_shinfo(skb)->gso_type & SKB_GSO_TCPV4))
  35                 return ERR_PTR(-EINVAL);
  36
  37         if (!pskb_may_pull(skb, sizeof(struct tcphdr)))
  38                 return ERR_PTR(-EINVAL);
  39
  40         if (unlikely(skb->ip_summed != CHECKSUM_PARTIAL)) {
  41                 const struct iphdr *iph = ip_hdr(skb);
  42                 struct tcphdr *th = tcp_hdr(skb);
  43
  44                 /* Set up checksum pseudo header, usually expect stack to
  45                  * have done this already.
  46                  */
  47
  48                 th->check = 0;
  49                 skb->ip_summed = CHECKSUM_PARTIAL;
  50                 __tcp_v4_send_check(skb, iph->saddr, iph->daddr);
  51         }
  52
  53         return tcp_gso_segment(skb, features);
  54 }
  55
  56 struct sk_buff *tcp_gso_segment(struct sk_buff *skb,
  57                                 netdev_features_t features)
  58 {
  59         struct sk_buff *segs = ERR_PTR(-EINVAL);
  60         unsigned int sum_truesize = 0;
  61         struct tcphdr *th;
  62         unsigned int thlen;
  63         unsigned int seq;
  64         unsigned int oldlen;
  65         unsigned int mss;
  66         struct sk_buff *gso_skb = skb;
  67         __sum16 newcheck;
  68         bool ooo_okay, copy_destructor;
  69         __wsum delta;
  70
  71         th = tcp_hdr(skb);
  72         thlen = th->doff * 4;
  73         if (thlen < sizeof(*th))
  74                 goto out;
  75
  76         if (!pskb_may_pull(skb, thlen))
  77                 goto out;
  78
  79         oldlen = ~skb->len;
  80         __skb_pull(skb, thlen);
  81
  82         mss = skb_shinfo(skb)->gso_size;
  83         if (unlikely(skb->len <= mss))
  84                 goto out;
  85
  86         if (skb_gso_ok(skb, features | NETIF_F_GSO_ROBUST)) {
  87                 /* Packet is from an untrusted source, reset gso_segs. */
  88
  89                 skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss);
  90
  91                 segs = NULL;
  92                 goto out;
  93         }
  94
  95         copy_destructor = gso_skb->destructor == tcp_wfree;
  96         ooo_okay = gso_skb->ooo_okay;
  97         /* All segments but the first should have ooo_okay cleared */
  98         skb->ooo_okay = 0;
  99
 100         segs = skb_segment(skb, features);
 101         if (IS_ERR(segs))
 102                 goto out;
 103
 104         /* Only first segment might have ooo_okay set */
 105         segs->ooo_okay = ooo_okay;
 106
 107         /* GSO partial and frag_list segmentation only requires splitting
 108          * the frame into an MSS multiple and possibly a remainder, both
 109          * cases return a GSO skb. So update the mss now.
 110          */
 111         if (skb_is_gso(segs))
 112                 mss *= skb_shinfo(segs)->gso_segs;
 113
 114         delta = (__force __wsum)htonl(oldlen + thlen + mss);
 115
 116         skb = segs;
 117         th = tcp_hdr(skb);
 118         seq = ntohl(th->seq);
 119
 120         if (unlikely(skb_shinfo(gso_skb)->tx_flags & SKBTX_SW_TSTAMP))
 121                 tcp_gso_tstamp(segs, skb_shinfo(gso_skb)->tskey, seq, mss);
 122
 123         newcheck = ~csum_fold(csum_add(csum_unfold(th->check), delta));
 124
 125         while (skb->next) {
 126                 th->fin = th->psh = 0;
 127                 th->check = newcheck;
 128
 129                 if (skb->ip_summed == CHECKSUM_PARTIAL)
 130                         gso_reset_checksum(skb, ~th->check);
 131                 else
 132                         th->check = gso_make_checksum(skb, ~th->check);
 133
 134                 seq += mss;
 135                 if (copy_destructor) {
 136                         skb->destructor = gso_skb->destructor;
 137                         skb->sk = gso_skb->sk;
 138                         sum_truesize += skb->truesize;
 139                 }
 140                 skb = skb->next;
 141                 th = tcp_hdr(skb);
 142
 143                 th->seq = htonl(seq);
 144                 th->cwr = 0;
 145         }
 146
 147         /* Following permits TCP Small Queues to work well with GSO :
 148          * The callback to TCP stack will be called at the time last frag
 149          * is freed at TX completion, and not right now when gso_skb
 150          * is freed by GSO engine
 151          */
 152         if (copy_destructor) {
 153                 int delta;
 154
 155                 swap(gso_skb->sk, skb->sk);
 156                 swap(gso_skb->destructor, skb->destructor);
 157                 sum_truesize += skb->truesize;
 158                 delta = sum_truesize - gso_skb->truesize;
 159                 /* In some pathological cases, delta can be negative.
 160                  * We need to either use refcount_add() or refcount_sub_and_test()
 161                  */
 162                 if (likely(delta >= 0))
 163                         refcount_add(delta, &skb->sk->sk_wmem_alloc);
 164                 else
 165                         WARN_ON_ONCE(refcount_sub_and_test(-delta, &skb->sk->sk_wmem_alloc));
 166         }
 167
 168         delta = (__force __wsum)htonl(oldlen +
 169                                       (skb_tail_pointer(skb) -
 170                                        skb_transport_header(skb)) +
 171                                       skb->data_len);
 172         th->check = ~csum_fold(csum_add(csum_unfold(th->check), delta));
 173         if (skb->ip_summed == CHECKSUM_PARTIAL)
 174                 gso_reset_checksum(skb, ~th->check);
 175         else
 176                 th->check = gso_make_checksum(skb, ~th->check);
 177 out:
 178         return segs;
 179 }
 180
 181 struct sk_buff *tcp_gro_receive(struct list_head *head, struct sk_buff *skb)
 182 {
 183         struct sk_buff *pp = NULL;
 184         struct sk_buff *p;
 185         struct tcphdr *th;
 186         struct tcphdr *th2;
 187         unsigned int len;
 188         unsigned int thlen;
 189         __be32 flags;
 190         unsigned int mss = 1;
 191         unsigned int hlen;
 192         unsigned int off;
 193         int flush = 1;
 194         int i;
 195
 196         off = skb_gro_offset(skb);
 197         hlen = off + sizeof(*th);
 198         th = skb_gro_header(skb, hlen, off);
 199         if (unlikely(!th))
 200                 goto out;
 201
 202         thlen = th->doff * 4;
 203         if (thlen < sizeof(*th))
 204                 goto out;
 205
 206         hlen = off + thlen;
 207         if (!skb_gro_may_pull(skb, hlen)) {
 208                 th = skb_gro_header_slow(skb, hlen, off);
 209                 if (unlikely(!th))
 210                         goto out;
 211         }
 212
 213         skb_gro_pull(skb, thlen);
 214
 215         len = skb_gro_len(skb);
 216         flags = tcp_flag_word(th);
 217
 218         list_for_each_entry(p, head, list) {
 219                 if (!NAPI_GRO_CB(p)->same_flow)
 220                         continue;
 221
 222                 th2 = tcp_hdr(p);
 223
 224                 if (*(u32 *)&th->source ^ *(u32 *)&th2->source) {
 225                         NAPI_GRO_CB(p)->same_flow = 0;
 226                         continue;
 227                 }
 228
 229                 goto found;
 230         }
 231         p = NULL;
 232         goto out_check_final;
 233
 234 found:
 235         /* Include the IP ID check below from the inner most IP hdr */
 236         flush = NAPI_GRO_CB(p)->flush;
 237         flush |= (__force int)(flags & TCP_FLAG_CWR);
 238         flush |= (__force int)((flags ^ tcp_flag_word(th2)) &
 239                   ~(TCP_FLAG_CWR | TCP_FLAG_FIN | TCP_FLAG_PSH));
 240         flush |= (__force int)(th->ack_seq ^ th2->ack_seq);
 241         for (i = sizeof(*th); i < thlen; i += 4)
 242                 flush |= *(u32 *)((u8 *)th + i) ^
 243                          *(u32 *)((u8 *)th2 + i);
 244
 245         /* When we receive our second frame we can made a decision on if we
 246          * continue this flow as an atomic flow with a fixed ID or if we use
 247          * an incrementing ID.
 248          */
 249         if (NAPI_GRO_CB(p)->flush_id != 1 ||
 250             NAPI_GRO_CB(p)->count != 1 ||
 251             !NAPI_GRO_CB(p)->is_atomic)
 252                 flush |= NAPI_GRO_CB(p)->flush_id;
 253         else
 254                 NAPI_GRO_CB(p)->is_atomic = false;
 255
 256         mss = skb_shinfo(p)->gso_size;
 257
 258         /* If skb is a GRO packet, make sure its gso_size matches prior packet mss.
 259          * If it is a single frame, do not aggregate it if its length
 260          * is bigger than our mss.
 261          */
 262         if (unlikely(skb_is_gso(skb)))
 263                 flush |= (mss != skb_shinfo(skb)->gso_size);
 264         else
 265                 flush |= (len - 1) >= mss;
 266
 267         flush |= (ntohl(th2->seq) + skb_gro_len(p)) ^ ntohl(th->seq);
 268 #ifdef CONFIG_TLS_DEVICE
 269         flush |= p->decrypted ^ skb->decrypted;
 270 #endif
 271
 272         if (flush || skb_gro_receive(p, skb)) {
 273                 mss = 1;
 274                 goto out_check_final;
 275         }
 276
 277         tcp_flag_word(th2) |= flags & (TCP_FLAG_FIN | TCP_FLAG_PSH);
 278
 279 out_check_final:
 280         /* Force a flush if last segment is smaller than mss. */
 281         if (unlikely(skb_is_gso(skb)))
 282                 flush = len != NAPI_GRO_CB(skb)->count * skb_shinfo(skb)->gso_size;
 283         else
 284                 flush = len < mss;
 285
 286         flush |= (__force int)(flags & (TCP_FLAG_URG | TCP_FLAG_PSH |
 287                                         TCP_FLAG_RST | TCP_FLAG_SYN |
 288                                         TCP_FLAG_FIN));
 289
 290         if (p && (!NAPI_GRO_CB(skb)->same_flow || flush))
 291                 pp = p;
 292
 293 out:
 294         NAPI_GRO_CB(skb)->flush |= (flush != 0);
 295
 296         return pp;
 297 }
 298
 299 void tcp_gro_complete(struct sk_buff *skb)
 300 {
 301         struct tcphdr *th = tcp_hdr(skb);
 302         struct skb_shared_info *shinfo;
 303
 304         if (skb->encapsulation)
 305                 skb->inner_transport_header = skb->transport_header;
 306
 307         skb->csum_start = (unsigned char *)th - skb->head;
 308         skb->csum_offset = offsetof(struct tcphdr, check);
 309         skb->ip_summed = CHECKSUM_PARTIAL;
 310
 311         shinfo = skb_shinfo(skb);
 312         shinfo->gso_segs = NAPI_GRO_CB(skb)->count;
 313
 314         if (th->cwr)
 315                 shinfo->gso_type |= SKB_GSO_TCP_ECN;
 316 }
 317 EXPORT_SYMBOL(tcp_gro_complete);
 318
 319 INDIRECT_CALLABLE_SCOPE
 320 struct sk_buff *tcp4_gro_receive(struct list_head *head, struct sk_buff *skb)
 321 {
 322         /* Don't bother verifying checksum if we're going to flush anyway. */
 323         if (!NAPI_GRO_CB(skb)->flush &&
 324             skb_gro_checksum_validate(skb, IPPROTO_TCP,
 325                                       inet_gro_compute_pseudo)) {
 326                 NAPI_GRO_CB(skb)->flush = 1;
 327                 return NULL;
 328         }
 329
 330         return tcp_gro_receive(head, skb);
 331 }
 332
 333 INDIRECT_CALLABLE_SCOPE int tcp4_gro_complete(struct sk_buff *skb, int thoff)
 334 {
 335         const struct iphdr *iph = ip_hdr(skb);
 336         struct tcphdr *th = tcp_hdr(skb);
 337
 338         th->check = ~tcp_v4_check(skb->len - thoff, iph->saddr,
 339                                   iph->daddr, 0);
 340
 341         skb_shinfo(skb)->gso_type |= SKB_GSO_TCPV4 |
 342                         (NAPI_GRO_CB(skb)->is_atomic * SKB_GSO_TCP_FIXEDID);
 343
 344         tcp_gro_complete(skb);
 345         return 0;
 346 }
 347
 348 int __init tcpv4_offload_init(void)
 349 {
 350         net_hotdata.tcpv4_offload = (struct net_offload) {
 351                 .callbacks = {
 352                         .gso_segment    =       tcp4_gso_segment,
 353                         .gro_receive    =       tcp4_gro_receive,
 354                         .gro_complete   =       tcp4_gro_complete,
 355                 },
 356         };
 357         return inet_add_offload(&net_hotdata.tcpv4_offload, IPPROTO_TCP);
 358 }