linux/net/ipv4/tcp_offload.c
<<
>>
Prefs
   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/tcp.h>
  12#include <net/protocol.h>
  13
  14static void tcp_gso_tstamp(struct sk_buff *skb, unsigned int ts_seq,
  15                           unsigned int seq, unsigned int mss)
  16{
  17        while (skb) {
  18                if (before(ts_seq, seq + mss)) {
  19                        skb_shinfo(skb)->tx_flags |= SKBTX_SW_TSTAMP;
  20                        skb_shinfo(skb)->tskey = ts_seq;
  21                        return;
  22                }
  23
  24                skb = skb->next;
  25                seq += mss;
  26        }
  27}
  28
  29static struct sk_buff *tcp4_gso_segment(struct sk_buff *skb,
  30                                        netdev_features_t features)
  31{
  32        if (!(skb_shinfo(skb)->gso_type & SKB_GSO_TCPV4))
  33                return ERR_PTR(-EINVAL);
  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
  54struct 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);
 176out:
 177        return segs;
 178}
 179
 180struct sk_buff *tcp_gro_receive(struct list_head *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        list_for_each_entry(p, head, list) {
 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        p = NULL;
 234        goto out_check_final;
 235
 236found:
 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#ifdef CONFIG_TLS_DEVICE
 263        flush |= p->decrypted ^ skb->decrypted;
 264#endif
 265
 266        if (flush || skb_gro_receive(p, skb)) {
 267                mss = 1;
 268                goto out_check_final;
 269        }
 270
 271        tcp_flag_word(th2) |= flags & (TCP_FLAG_FIN | TCP_FLAG_PSH);
 272
 273out_check_final:
 274        flush = len < mss;
 275        flush |= (__force int)(flags & (TCP_FLAG_URG | TCP_FLAG_PSH |
 276                                        TCP_FLAG_RST | TCP_FLAG_SYN |
 277                                        TCP_FLAG_FIN));
 278
 279        if (p && (!NAPI_GRO_CB(skb)->same_flow || flush))
 280                pp = p;
 281
 282out:
 283        NAPI_GRO_CB(skb)->flush |= (flush != 0);
 284
 285        return pp;
 286}
 287
 288int tcp_gro_complete(struct sk_buff *skb)
 289{
 290        struct tcphdr *th = tcp_hdr(skb);
 291
 292        skb->csum_start = (unsigned char *)th - skb->head;
 293        skb->csum_offset = offsetof(struct tcphdr, check);
 294        skb->ip_summed = CHECKSUM_PARTIAL;
 295
 296        skb_shinfo(skb)->gso_segs = NAPI_GRO_CB(skb)->count;
 297
 298        if (th->cwr)
 299                skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
 300
 301        if (skb->encapsulation)
 302                skb->inner_transport_header = skb->transport_header;
 303
 304        return 0;
 305}
 306EXPORT_SYMBOL(tcp_gro_complete);
 307
 308INDIRECT_CALLABLE_SCOPE
 309struct sk_buff *tcp4_gro_receive(struct list_head *head, struct sk_buff *skb)
 310{
 311        /* Don't bother verifying checksum if we're going to flush anyway. */
 312        if (!NAPI_GRO_CB(skb)->flush &&
 313            skb_gro_checksum_validate(skb, IPPROTO_TCP,
 314                                      inet_gro_compute_pseudo)) {
 315                NAPI_GRO_CB(skb)->flush = 1;
 316                return NULL;
 317        }
 318
 319        return tcp_gro_receive(head, skb);
 320}
 321
 322INDIRECT_CALLABLE_SCOPE int tcp4_gro_complete(struct sk_buff *skb, int thoff)
 323{
 324        const struct iphdr *iph = ip_hdr(skb);
 325        struct tcphdr *th = tcp_hdr(skb);
 326
 327        th->check = ~tcp_v4_check(skb->len - thoff, iph->saddr,
 328                                  iph->daddr, 0);
 329        skb_shinfo(skb)->gso_type |= SKB_GSO_TCPV4;
 330
 331        if (NAPI_GRO_CB(skb)->is_atomic)
 332                skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_FIXEDID;
 333
 334        return tcp_gro_complete(skb);
 335}
 336
 337static const struct net_offload tcpv4_offload = {
 338        .callbacks = {
 339                .gso_segment    =       tcp4_gso_segment,
 340                .gro_receive    =       tcp4_gro_receive,
 341                .gro_complete   =       tcp4_gro_complete,
 342        },
 343};
 344
 345int __init tcpv4_offload_init(void)
 346{
 347        return inet_add_offload(&tcpv4_offload, IPPROTO_TCP);
 348}
 349