net/ipv4/tcp_offload.c

0001 // SPDX-License-Identifier: GPL-2.0-or-later
0002 /*
0003  *  IPV4 GSO/GRO offload support
0004  *  Linux INET implementation
0005  *
0006  *  TCPv4 GSO/GRO support
0007  */
0008
0009 #include <linux/indirect_call_wrapper.h>
0010 #include <linux/skbuff.h>
0011 #include <net/gro.h>
0012 #include <net/tcp.h>
0013 #include <net/protocol.h>
0014
0015 static void tcp_gso_tstamp(struct sk_buff *skb, unsigned int ts_seq,
0016                unsigned int seq, unsigned int mss)
0017 {
0018     while (skb) {
0019         if (before(ts_seq, seq + mss)) {
0020             skb_shinfo(skb)->tx_flags |= SKBTX_SW_TSTAMP;
0021             skb_shinfo(skb)->tskey = ts_seq;
0022             return;
0023         }
0024
0025         skb = skb->next;
0026         seq += mss;
0027     }
0028 }
0029
0030 static struct sk_buff *tcp4_gso_segment(struct sk_buff *skb,
0031                     netdev_features_t features)
0032 {
0033     if (!(skb_shinfo(skb)->gso_type & SKB_GSO_TCPV4))
0034         return ERR_PTR(-EINVAL);
0035
0036     if (!pskb_may_pull(skb, sizeof(struct tcphdr)))
0037         return ERR_PTR(-EINVAL);
0038
0039     if (unlikely(skb->ip_summed != CHECKSUM_PARTIAL)) {
0040         const struct iphdr *iph = ip_hdr(skb);
0041         struct tcphdr *th = tcp_hdr(skb);
0042
0043         /* Set up checksum pseudo header, usually expect stack to
0044          * have done this already.
0045          */
0046
0047         th->check = 0;
0048         skb->ip_summed = CHECKSUM_PARTIAL;
0049         __tcp_v4_send_check(skb, iph->saddr, iph->daddr);
0050     }
0051
0052     return tcp_gso_segment(skb, features);
0053 }
0054
0055 struct sk_buff *tcp_gso_segment(struct sk_buff *skb,
0056                 netdev_features_t features)
0057 {
0058     struct sk_buff *segs = ERR_PTR(-EINVAL);
0059     unsigned int sum_truesize = 0;
0060     struct tcphdr *th;
0061     unsigned int thlen;
0062     unsigned int seq;
0063     __be32 delta;
0064     unsigned int oldlen;
0065     unsigned int mss;
0066     struct sk_buff *gso_skb = skb;
0067     __sum16 newcheck;
0068     bool ooo_okay, copy_destructor;
0069
0070     th = tcp_hdr(skb);
0071     thlen = th->doff * 4;
0072     if (thlen < sizeof(*th))
0073         goto out;
0074
0075     if (!pskb_may_pull(skb, thlen))
0076         goto out;
0077
0078     oldlen = (u16)~skb->len;
0079     __skb_pull(skb, thlen);
0080
0081     mss = skb_shinfo(skb)->gso_size;
0082     if (unlikely(skb->len <= mss))
0083         goto out;
0084
0085     if (skb_gso_ok(skb, features | NETIF_F_GSO_ROBUST)) {
0086         /* Packet is from an untrusted source, reset gso_segs. */
0087
0088         skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss);
0089
0090         segs = NULL;
0091         goto out;
0092     }
0093
0094     copy_destructor = gso_skb->destructor == tcp_wfree;
0095     ooo_okay = gso_skb->ooo_okay;
0096     /* All segments but the first should have ooo_okay cleared */
0097     skb->ooo_okay = 0;
0098
0099     segs = skb_segment(skb, features);
0100     if (IS_ERR(segs))
0101         goto out;
0102
0103     /* Only first segment might have ooo_okay set */
0104     segs->ooo_okay = ooo_okay;
0105
0106     /* GSO partial and frag_list segmentation only requires splitting
0107      * the frame into an MSS multiple and possibly a remainder, both
0108      * cases return a GSO skb. So update the mss now.
0109      */
0110     if (skb_is_gso(segs))
0111         mss *= skb_shinfo(segs)->gso_segs;
0112
0113     delta = htonl(oldlen + (thlen + mss));
0114
0115     skb = segs;
0116     th = tcp_hdr(skb);
0117     seq = ntohl(th->seq);
0118
0119     if (unlikely(skb_shinfo(gso_skb)->tx_flags & SKBTX_SW_TSTAMP))
0120         tcp_gso_tstamp(segs, skb_shinfo(gso_skb)->tskey, seq, mss);
0121
0122     newcheck = ~csum_fold((__force __wsum)((__force u32)th->check +
0123                            (__force u32)delta));
0124
0125     while (skb->next) {
0126         th->fin = th->psh = 0;
0127         th->check = newcheck;
0128
0129         if (skb->ip_summed == CHECKSUM_PARTIAL)
0130             gso_reset_checksum(skb, ~th->check);
0131         else
0132             th->check = gso_make_checksum(skb, ~th->check);
0133
0134         seq += mss;
0135         if (copy_destructor) {
0136             skb->destructor = gso_skb->destructor;
0137             skb->sk = gso_skb->sk;
0138             sum_truesize += skb->truesize;
0139         }
0140         skb = skb->next;
0141         th = tcp_hdr(skb);
0142
0143         th->seq = htonl(seq);
0144         th->cwr = 0;
0145     }
0146
0147     /* Following permits TCP Small Queues to work well with GSO :
0148      * The callback to TCP stack will be called at the time last frag
0149      * is freed at TX completion, and not right now when gso_skb
0150      * is freed by GSO engine
0151      */
0152     if (copy_destructor) {
0153         int delta;
0154
0155         swap(gso_skb->sk, skb->sk);
0156         swap(gso_skb->destructor, skb->destructor);
0157         sum_truesize += skb->truesize;
0158         delta = sum_truesize - gso_skb->truesize;
0159         /* In some pathological cases, delta can be negative.
0160          * We need to either use refcount_add() or refcount_sub_and_test()
0161          */
0162         if (likely(delta >= 0))
0163             refcount_add(delta, &skb->sk->sk_wmem_alloc);
0164         else
0165             WARN_ON_ONCE(refcount_sub_and_test(-delta, &skb->sk->sk_wmem_alloc));
0166     }
0167
0168     delta = htonl(oldlen + (skb_tail_pointer(skb) -
0169                 skb_transport_header(skb)) +
0170               skb->data_len);
0171     th->check = ~csum_fold((__force __wsum)((__force u32)th->check +
0172                 (__force u32)delta));
0173     if (skb->ip_summed == CHECKSUM_PARTIAL)
0174         gso_reset_checksum(skb, ~th->check);
0175     else
0176         th->check = gso_make_checksum(skb, ~th->check);
0177 out:
0178     return segs;
0179 }
0180
0181 struct sk_buff *tcp_gro_receive(struct list_head *head, struct sk_buff *skb)
0182 {
0183     struct sk_buff *pp = NULL;
0184     struct sk_buff *p;
0185     struct tcphdr *th;
0186     struct tcphdr *th2;
0187     unsigned int len;
0188     unsigned int thlen;
0189     __be32 flags;
0190     unsigned int mss = 1;
0191     unsigned int hlen;
0192     unsigned int off;
0193     int flush = 1;
0194     int i;
0195
0196     off = skb_gro_offset(skb);
0197     hlen = off + sizeof(*th);
0198     th = skb_gro_header_fast(skb, off);
0199     if (skb_gro_header_hard(skb, hlen)) {
0200         th = skb_gro_header_slow(skb, hlen, off);
0201         if (unlikely(!th))
0202             goto out;
0203     }
0204
0205     thlen = th->doff * 4;
0206     if (thlen < sizeof(*th))
0207         goto out;
0208
0209     hlen = off + thlen;
0210     if (skb_gro_header_hard(skb, hlen)) {
0211         th = skb_gro_header_slow(skb, hlen, off);
0212         if (unlikely(!th))
0213             goto out;
0214     }
0215
0216     skb_gro_pull(skb, thlen);
0217
0218     len = skb_gro_len(skb);
0219     flags = tcp_flag_word(th);
0220
0221     list_for_each_entry(p, head, list) {
0222         if (!NAPI_GRO_CB(p)->same_flow)
0223             continue;
0224
0225         th2 = tcp_hdr(p);
0226
0227         if (*(u32 *)&th->source ^ *(u32 *)&th2->source) {
0228             NAPI_GRO_CB(p)->same_flow = 0;
0229             continue;
0230         }
0231
0232         goto found;
0233     }
0234     p = NULL;
0235     goto out_check_final;
0236
0237 found:
0238     /* Include the IP ID check below from the inner most IP hdr */
0239     flush = NAPI_GRO_CB(p)->flush;
0240     flush |= (__force int)(flags & TCP_FLAG_CWR);
0241     flush |= (__force int)((flags ^ tcp_flag_word(th2)) &
0242           ~(TCP_FLAG_CWR | TCP_FLAG_FIN | TCP_FLAG_PSH));
0243     flush |= (__force int)(th->ack_seq ^ th2->ack_seq);
0244     for (i = sizeof(*th); i < thlen; i += 4)
0245         flush |= *(u32 *)((u8 *)th + i) ^
0246              *(u32 *)((u8 *)th2 + i);
0247
0248     /* When we receive our second frame we can made a decision on if we
0249      * continue this flow as an atomic flow with a fixed ID or if we use
0250      * an incrementing ID.
0251      */
0252     if (NAPI_GRO_CB(p)->flush_id != 1 ||
0253         NAPI_GRO_CB(p)->count != 1 ||
0254         !NAPI_GRO_CB(p)->is_atomic)
0255         flush |= NAPI_GRO_CB(p)->flush_id;
0256     else
0257         NAPI_GRO_CB(p)->is_atomic = false;
0258
0259     mss = skb_shinfo(p)->gso_size;
0260
0261     flush |= (len - 1) >= mss;
0262     flush |= (ntohl(th2->seq) + skb_gro_len(p)) ^ ntohl(th->seq);
0263 #ifdef CONFIG_TLS_DEVICE
0264     flush |= p->decrypted ^ skb->decrypted;
0265 #endif
0266
0267     if (flush || skb_gro_receive(p, skb)) {
0268         mss = 1;
0269         goto out_check_final;
0270     }
0271
0272     tcp_flag_word(th2) |= flags & (TCP_FLAG_FIN | TCP_FLAG_PSH);
0273
0274 out_check_final:
0275     flush = len < mss;
0276     flush |= (__force int)(flags & (TCP_FLAG_URG | TCP_FLAG_PSH |
0277                     TCP_FLAG_RST | TCP_FLAG_SYN |
0278                     TCP_FLAG_FIN));
0279
0280     if (p && (!NAPI_GRO_CB(skb)->same_flow || flush))
0281         pp = p;
0282
0283 out:
0284     NAPI_GRO_CB(skb)->flush |= (flush != 0);
0285
0286     return pp;
0287 }
0288
0289 int tcp_gro_complete(struct sk_buff *skb)
0290 {
0291     struct tcphdr *th = tcp_hdr(skb);
0292
0293     skb->csum_start = (unsigned char *)th - skb->head;
0294     skb->csum_offset = offsetof(struct tcphdr, check);
0295     skb->ip_summed = CHECKSUM_PARTIAL;
0296
0297     skb_shinfo(skb)->gso_segs = NAPI_GRO_CB(skb)->count;
0298
0299     if (th->cwr)
0300         skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
0301
0302     if (skb->encapsulation)
0303         skb->inner_transport_header = skb->transport_header;
0304
0305     return 0;
0306 }
0307 EXPORT_SYMBOL(tcp_gro_complete);
0308
0309 INDIRECT_CALLABLE_SCOPE
0310 struct sk_buff *tcp4_gro_receive(struct list_head *head, struct sk_buff *skb)
0311 {
0312     /* Don't bother verifying checksum if we're going to flush anyway. */
0313     if (!NAPI_GRO_CB(skb)->flush &&
0314         skb_gro_checksum_validate(skb, IPPROTO_TCP,
0315                       inet_gro_compute_pseudo)) {
0316         NAPI_GRO_CB(skb)->flush = 1;
0317         return NULL;
0318     }
0319
0320     return tcp_gro_receive(head, skb);
0321 }
0322
0323 INDIRECT_CALLABLE_SCOPE int tcp4_gro_complete(struct sk_buff *skb, int thoff)
0324 {
0325     const struct iphdr *iph = ip_hdr(skb);
0326     struct tcphdr *th = tcp_hdr(skb);
0327
0328     th->check = ~tcp_v4_check(skb->len - thoff, iph->saddr,
0329                   iph->daddr, 0);
0330     skb_shinfo(skb)->gso_type |= SKB_GSO_TCPV4;
0331
0332     if (NAPI_GRO_CB(skb)->is_atomic)
0333         skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_FIXEDID;
0334
0335     return tcp_gro_complete(skb);
0336 }
0337
0338 static const struct net_offload tcpv4_offload = {
0339     .callbacks = {
0340         .gso_segment    =   tcp4_gso_segment,
0341         .gro_receive    =   tcp4_gro_receive,
0342         .gro_complete   =   tcp4_gro_complete,
0343     },
0344 };
0345
0346 int __init tcpv4_offload_init(void)
0347 {
0348     return inet_add_offload(&tcpv4_offload, IPPROTO_TCP);
0349 }