whiterose

linux unikernel
Log | Files | Refs | README | LICENSE | git clone https://git.ne02ptzero.me/git/whiterose

ip_output.c (41119B)


      1 /*
      2  * INET		An implementation of the TCP/IP protocol suite for the LINUX
      3  *		operating system.  INET is implemented using the  BSD Socket
      4  *		interface as the means of communication with the user level.
      5  *
      6  *		The Internet Protocol (IP) output module.
      7  *
      8  * Authors:	Ross Biro
      9  *		Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
     10  *		Donald Becker, <becker@super.org>
     11  *		Alan Cox, <Alan.Cox@linux.org>
     12  *		Richard Underwood
     13  *		Stefan Becker, <stefanb@yello.ping.de>
     14  *		Jorge Cwik, <jorge@laser.satlink.net>
     15  *		Arnt Gulbrandsen, <agulbra@nvg.unit.no>
     16  *		Hirokazu Takahashi, <taka@valinux.co.jp>
     17  *
     18  *	See ip_input.c for original log
     19  *
     20  *	Fixes:
     21  *		Alan Cox	:	Missing nonblock feature in ip_build_xmit.
     22  *		Mike Kilburn	:	htons() missing in ip_build_xmit.
     23  *		Bradford Johnson:	Fix faulty handling of some frames when
     24  *					no route is found.
     25  *		Alexander Demenshin:	Missing sk/skb free in ip_queue_xmit
     26  *					(in case if packet not accepted by
     27  *					output firewall rules)
     28  *		Mike McLagan	:	Routing by source
     29  *		Alexey Kuznetsov:	use new route cache
     30  *		Andi Kleen:		Fix broken PMTU recovery and remove
     31  *					some redundant tests.
     32  *	Vitaly E. Lavrov	:	Transparent proxy revived after year coma.
     33  *		Andi Kleen	: 	Replace ip_reply with ip_send_reply.
     34  *		Andi Kleen	:	Split fast and slow ip_build_xmit path
     35  *					for decreased register pressure on x86
     36  *					and more readibility.
     37  *		Marc Boucher	:	When call_out_firewall returns FW_QUEUE,
     38  *					silently drop skb instead of failing with -EPERM.
     39  *		Detlev Wengorz	:	Copy protocol for fragments.
     40  *		Hirokazu Takahashi:	HW checksumming for outgoing UDP
     41  *					datagrams.
     42  *		Hirokazu Takahashi:	sendfile() on UDP works now.
     43  */
     44 
     45 #include <linux/uaccess.h>
     46 #include <linux/module.h>
     47 #include <linux/types.h>
     48 #include <linux/kernel.h>
     49 #include <linux/mm.h>
     50 #include <linux/string.h>
     51 #include <linux/errno.h>
     52 #include <linux/highmem.h>
     53 #include <linux/slab.h>
     54 
     55 #include <linux/socket.h>
     56 #include <linux/sockios.h>
     57 #include <linux/in.h>
     58 #include <linux/inet.h>
     59 #include <linux/netdevice.h>
     60 #include <linux/etherdevice.h>
     61 #include <linux/proc_fs.h>
     62 #include <linux/stat.h>
     63 #include <linux/init.h>
     64 
     65 #include <net/snmp.h>
     66 #include <net/ip.h>
     67 #include <net/protocol.h>
     68 #include <net/route.h>
     69 #include <net/xfrm.h>
     70 #include <linux/skbuff.h>
     71 #include <net/sock.h>
     72 #include <net/arp.h>
     73 #include <net/icmp.h>
     74 #include <net/checksum.h>
     75 #include <net/inetpeer.h>
     76 #include <net/lwtunnel.h>
     77 #include <linux/bpf-cgroup.h>
     78 #include <linux/igmp.h>
     79 #include <linux/netfilter_ipv4.h>
     80 #include <linux/netfilter_bridge.h>
     81 #include <linux/netlink.h>
     82 #include <linux/tcp.h>
     83 
     84 static int
     85 ip_fragment(struct net *net, struct sock *sk, struct sk_buff *skb,
     86 	    unsigned int mtu,
     87 	    int (*output)(struct net *, struct sock *, struct sk_buff *));
     88 
     89 /* Generate a checksum for an outgoing IP datagram. */
     90 void ip_send_check(struct iphdr *iph)
     91 {
     92 	iph->check = 0;
     93 	iph->check = ip_fast_csum((unsigned char *)iph, iph->ihl);
     94 }
     95 EXPORT_SYMBOL(ip_send_check);
     96 
     97 int __ip_local_out(struct net *net, struct sock *sk, struct sk_buff *skb)
     98 {
     99 	struct iphdr *iph = ip_hdr(skb);
    100 
    101 	iph->tot_len = htons(skb->len);
    102 	ip_send_check(iph);
    103 
    104 	/* if egress device is enslaved to an L3 master device pass the
    105 	 * skb to its handler for processing
    106 	 */
    107 	skb = l3mdev_ip_out(sk, skb);
    108 	if (unlikely(!skb))
    109 		return 0;
    110 
    111 	skb->protocol = htons(ETH_P_IP);
    112 
    113 	return nf_hook(NFPROTO_IPV4, NF_INET_LOCAL_OUT,
    114 		       net, sk, skb, NULL, skb_dst(skb)->dev,
    115 		       dst_output);
    116 }
    117 
    118 int ip_local_out(struct net *net, struct sock *sk, struct sk_buff *skb)
    119 {
    120 	int err;
    121 
    122 	err = __ip_local_out(net, sk, skb);
    123 	if (likely(err == 1))
    124 		err = dst_output(net, sk, skb);
    125 
    126 	return err;
    127 }
    128 EXPORT_SYMBOL_GPL(ip_local_out);
    129 
    130 static inline int ip_select_ttl(struct inet_sock *inet, struct dst_entry *dst)
    131 {
    132 	int ttl = inet->uc_ttl;
    133 
    134 	if (ttl < 0)
    135 		ttl = ip4_dst_hoplimit(dst);
    136 	return ttl;
    137 }
    138 
    139 /*
    140  *		Add an ip header to a skbuff and send it out.
    141  *
    142  */
    143 int ip_build_and_send_pkt(struct sk_buff *skb, const struct sock *sk,
    144 			  __be32 saddr, __be32 daddr, struct ip_options_rcu *opt)
    145 {
    146 	struct inet_sock *inet = inet_sk(sk);
    147 	struct rtable *rt = skb_rtable(skb);
    148 	struct net *net = sock_net(sk);
    149 	struct iphdr *iph;
    150 
    151 	/* Build the IP header. */
    152 	skb_push(skb, sizeof(struct iphdr) + (opt ? opt->opt.optlen : 0));
    153 	skb_reset_network_header(skb);
    154 	iph = ip_hdr(skb);
    155 	iph->version  = 4;
    156 	iph->ihl      = 5;
    157 	iph->tos      = inet->tos;
    158 	iph->ttl      = ip_select_ttl(inet, &rt->dst);
    159 	iph->daddr    = (opt && opt->opt.srr ? opt->opt.faddr : daddr);
    160 	iph->saddr    = saddr;
    161 	iph->protocol = sk->sk_protocol;
    162 	if (ip_dont_fragment(sk, &rt->dst)) {
    163 		iph->frag_off = htons(IP_DF);
    164 		iph->id = 0;
    165 	} else {
    166 		iph->frag_off = 0;
    167 		__ip_select_ident(net, iph, 1);
    168 	}
    169 
    170 	if (opt && opt->opt.optlen) {
    171 		iph->ihl += opt->opt.optlen>>2;
    172 		ip_options_build(skb, &opt->opt, daddr, rt, 0);
    173 	}
    174 
    175 	skb->priority = sk->sk_priority;
    176 	if (!skb->mark)
    177 		skb->mark = sk->sk_mark;
    178 
    179 	/* Send it out. */
    180 	return ip_local_out(net, skb->sk, skb);
    181 }
    182 EXPORT_SYMBOL_GPL(ip_build_and_send_pkt);
    183 
    184 static int ip_finish_output2(struct net *net, struct sock *sk, struct sk_buff *skb)
    185 {
    186 	struct dst_entry *dst = skb_dst(skb);
    187 	struct rtable *rt = (struct rtable *)dst;
    188 	struct net_device *dev = dst->dev;
    189 	unsigned int hh_len = LL_RESERVED_SPACE(dev);
    190 	struct neighbour *neigh;
    191 	u32 nexthop;
    192 
    193 	if (rt->rt_type == RTN_MULTICAST) {
    194 		IP_UPD_PO_STATS(net, IPSTATS_MIB_OUTMCAST, skb->len);
    195 	} else if (rt->rt_type == RTN_BROADCAST)
    196 		IP_UPD_PO_STATS(net, IPSTATS_MIB_OUTBCAST, skb->len);
    197 
    198 	/* Be paranoid, rather than too clever. */
    199 	if (unlikely(skb_headroom(skb) < hh_len && dev->header_ops)) {
    200 		struct sk_buff *skb2;
    201 
    202 		skb2 = skb_realloc_headroom(skb, LL_RESERVED_SPACE(dev));
    203 		if (!skb2) {
    204 			kfree_skb(skb);
    205 			return -ENOMEM;
    206 		}
    207 		if (skb->sk)
    208 			skb_set_owner_w(skb2, skb->sk);
    209 		consume_skb(skb);
    210 		skb = skb2;
    211 	}
    212 
    213 	if (lwtunnel_xmit_redirect(dst->lwtstate)) {
    214 		int res = lwtunnel_xmit(skb);
    215 
    216 		if (res < 0 || res == LWTUNNEL_XMIT_DONE)
    217 			return res;
    218 	}
    219 
    220 	rcu_read_lock_bh();
    221 	nexthop = (__force u32) rt_nexthop(rt, ip_hdr(skb)->daddr);
    222 	neigh = __ipv4_neigh_lookup_noref(dev, nexthop);
    223 	if (unlikely(!neigh))
    224 		neigh = __neigh_create(&arp_tbl, &nexthop, dev, false);
    225 	if (!IS_ERR(neigh)) {
    226 		int res;
    227 
    228 		sock_confirm_neigh(skb, neigh);
    229 		res = neigh_output(neigh, skb);
    230 
    231 		rcu_read_unlock_bh();
    232 		return res;
    233 	}
    234 	rcu_read_unlock_bh();
    235 
    236 	net_dbg_ratelimited("%s: No header cache and no neighbour!\n",
    237 			    __func__);
    238 	kfree_skb(skb);
    239 	return -EINVAL;
    240 }
    241 
    242 static int ip_finish_output_gso(struct net *net, struct sock *sk,
    243 				struct sk_buff *skb, unsigned int mtu)
    244 {
    245 	netdev_features_t features;
    246 	struct sk_buff *segs;
    247 	int ret = 0;
    248 
    249 	/* common case: seglen is <= mtu
    250 	 */
    251 	if (skb_gso_validate_network_len(skb, mtu))
    252 		return ip_finish_output2(net, sk, skb);
    253 
    254 	/* Slowpath -  GSO segment length exceeds the egress MTU.
    255 	 *
    256 	 * This can happen in several cases:
    257 	 *  - Forwarding of a TCP GRO skb, when DF flag is not set.
    258 	 *  - Forwarding of an skb that arrived on a virtualization interface
    259 	 *    (virtio-net/vhost/tap) with TSO/GSO size set by other network
    260 	 *    stack.
    261 	 *  - Local GSO skb transmitted on an NETIF_F_TSO tunnel stacked over an
    262 	 *    interface with a smaller MTU.
    263 	 *  - Arriving GRO skb (or GSO skb in a virtualized environment) that is
    264 	 *    bridged to a NETIF_F_TSO tunnel stacked over an interface with an
    265 	 *    insufficent MTU.
    266 	 */
    267 	features = netif_skb_features(skb);
    268 	BUILD_BUG_ON(sizeof(*IPCB(skb)) > SKB_SGO_CB_OFFSET);
    269 	segs = skb_gso_segment(skb, features & ~NETIF_F_GSO_MASK);
    270 	if (IS_ERR_OR_NULL(segs)) {
    271 		kfree_skb(skb);
    272 		return -ENOMEM;
    273 	}
    274 
    275 	consume_skb(skb);
    276 
    277 	do {
    278 		struct sk_buff *nskb = segs->next;
    279 		int err;
    280 
    281 		skb_mark_not_on_list(segs);
    282 		err = ip_fragment(net, sk, segs, mtu, ip_finish_output2);
    283 
    284 		if (err && ret == 0)
    285 			ret = err;
    286 		segs = nskb;
    287 	} while (segs);
    288 
    289 	return ret;
    290 }
    291 
    292 static int ip_finish_output(struct net *net, struct sock *sk, struct sk_buff *skb)
    293 {
    294 	unsigned int mtu;
    295 	int ret;
    296 
    297 	ret = BPF_CGROUP_RUN_PROG_INET_EGRESS(sk, skb);
    298 	if (ret) {
    299 		kfree_skb(skb);
    300 		return ret;
    301 	}
    302 
    303 #if defined(CONFIG_NETFILTER) && defined(CONFIG_XFRM)
    304 	/* Policy lookup after SNAT yielded a new policy */
    305 	if (skb_dst(skb)->xfrm) {
    306 		IPCB(skb)->flags |= IPSKB_REROUTED;
    307 		return dst_output(net, sk, skb);
    308 	}
    309 #endif
    310 	mtu = ip_skb_dst_mtu(sk, skb);
    311 	if (skb_is_gso(skb))
    312 		return ip_finish_output_gso(net, sk, skb, mtu);
    313 
    314 	if (skb->len > mtu || (IPCB(skb)->flags & IPSKB_FRAG_PMTU))
    315 		return ip_fragment(net, sk, skb, mtu, ip_finish_output2);
    316 
    317 	return ip_finish_output2(net, sk, skb);
    318 }
    319 
    320 static int ip_mc_finish_output(struct net *net, struct sock *sk,
    321 			       struct sk_buff *skb)
    322 {
    323 	int ret;
    324 
    325 	ret = BPF_CGROUP_RUN_PROG_INET_EGRESS(sk, skb);
    326 	if (ret) {
    327 		kfree_skb(skb);
    328 		return ret;
    329 	}
    330 
    331 	return dev_loopback_xmit(net, sk, skb);
    332 }
    333 
    334 int ip_mc_output(struct net *net, struct sock *sk, struct sk_buff *skb)
    335 {
    336 	struct rtable *rt = skb_rtable(skb);
    337 	struct net_device *dev = rt->dst.dev;
    338 
    339 	/*
    340 	 *	If the indicated interface is up and running, send the packet.
    341 	 */
    342 	IP_UPD_PO_STATS(net, IPSTATS_MIB_OUT, skb->len);
    343 
    344 	skb->dev = dev;
    345 	skb->protocol = htons(ETH_P_IP);
    346 
    347 	/*
    348 	 *	Multicasts are looped back for other local users
    349 	 */
    350 
    351 	if (rt->rt_flags&RTCF_MULTICAST) {
    352 		if (sk_mc_loop(sk)
    353 #ifdef CONFIG_IP_MROUTE
    354 		/* Small optimization: do not loopback not local frames,
    355 		   which returned after forwarding; they will be  dropped
    356 		   by ip_mr_input in any case.
    357 		   Note, that local frames are looped back to be delivered
    358 		   to local recipients.
    359 
    360 		   This check is duplicated in ip_mr_input at the moment.
    361 		 */
    362 		    &&
    363 		    ((rt->rt_flags & RTCF_LOCAL) ||
    364 		     !(IPCB(skb)->flags & IPSKB_FORWARDED))
    365 #endif
    366 		   ) {
    367 			struct sk_buff *newskb = skb_clone(skb, GFP_ATOMIC);
    368 			if (newskb)
    369 				NF_HOOK(NFPROTO_IPV4, NF_INET_POST_ROUTING,
    370 					net, sk, newskb, NULL, newskb->dev,
    371 					ip_mc_finish_output);
    372 		}
    373 
    374 		/* Multicasts with ttl 0 must not go beyond the host */
    375 
    376 		if (ip_hdr(skb)->ttl == 0) {
    377 			kfree_skb(skb);
    378 			return 0;
    379 		}
    380 	}
    381 
    382 	if (rt->rt_flags&RTCF_BROADCAST) {
    383 		struct sk_buff *newskb = skb_clone(skb, GFP_ATOMIC);
    384 		if (newskb)
    385 			NF_HOOK(NFPROTO_IPV4, NF_INET_POST_ROUTING,
    386 				net, sk, newskb, NULL, newskb->dev,
    387 				ip_mc_finish_output);
    388 	}
    389 
    390 	return NF_HOOK_COND(NFPROTO_IPV4, NF_INET_POST_ROUTING,
    391 			    net, sk, skb, NULL, skb->dev,
    392 			    ip_finish_output,
    393 			    !(IPCB(skb)->flags & IPSKB_REROUTED));
    394 }
    395 
    396 int ip_output(struct net *net, struct sock *sk, struct sk_buff *skb)
    397 {
    398 	struct net_device *dev = skb_dst(skb)->dev;
    399 
    400 	IP_UPD_PO_STATS(net, IPSTATS_MIB_OUT, skb->len);
    401 
    402 	skb->dev = dev;
    403 	skb->protocol = htons(ETH_P_IP);
    404 
    405 	return NF_HOOK_COND(NFPROTO_IPV4, NF_INET_POST_ROUTING,
    406 			    net, sk, skb, NULL, dev,
    407 			    ip_finish_output,
    408 			    !(IPCB(skb)->flags & IPSKB_REROUTED));
    409 }
    410 
    411 /*
    412  * copy saddr and daddr, possibly using 64bit load/stores
    413  * Equivalent to :
    414  *   iph->saddr = fl4->saddr;
    415  *   iph->daddr = fl4->daddr;
    416  */
    417 static void ip_copy_addrs(struct iphdr *iph, const struct flowi4 *fl4)
    418 {
    419 	BUILD_BUG_ON(offsetof(typeof(*fl4), daddr) !=
    420 		     offsetof(typeof(*fl4), saddr) + sizeof(fl4->saddr));
    421 	memcpy(&iph->saddr, &fl4->saddr,
    422 	       sizeof(fl4->saddr) + sizeof(fl4->daddr));
    423 }
    424 
    425 /* Note: skb->sk can be different from sk, in case of tunnels */
    426 int __ip_queue_xmit(struct sock *sk, struct sk_buff *skb, struct flowi *fl,
    427 		    __u8 tos)
    428 {
    429 	struct inet_sock *inet = inet_sk(sk);
    430 	struct net *net = sock_net(sk);
    431 	struct ip_options_rcu *inet_opt;
    432 	struct flowi4 *fl4;
    433 	struct rtable *rt;
    434 	struct iphdr *iph;
    435 	int res;
    436 
    437 	/* Skip all of this if the packet is already routed,
    438 	 * f.e. by something like SCTP.
    439 	 */
    440 	rcu_read_lock();
    441 	inet_opt = rcu_dereference(inet->inet_opt);
    442 	fl4 = &fl->u.ip4;
    443 	rt = skb_rtable(skb);
    444 	if (rt)
    445 		goto packet_routed;
    446 
    447 	/* Make sure we can route this packet. */
    448 	rt = (struct rtable *)__sk_dst_check(sk, 0);
    449 	if (!rt) {
    450 		__be32 daddr;
    451 
    452 		/* Use correct destination address if we have options. */
    453 		daddr = inet->inet_daddr;
    454 		if (inet_opt && inet_opt->opt.srr)
    455 			daddr = inet_opt->opt.faddr;
    456 
    457 		/* If this fails, retransmit mechanism of transport layer will
    458 		 * keep trying until route appears or the connection times
    459 		 * itself out.
    460 		 */
    461 		rt = ip_route_output_ports(net, fl4, sk,
    462 					   daddr, inet->inet_saddr,
    463 					   inet->inet_dport,
    464 					   inet->inet_sport,
    465 					   sk->sk_protocol,
    466 					   RT_CONN_FLAGS_TOS(sk, tos),
    467 					   sk->sk_bound_dev_if);
    468 		if (IS_ERR(rt))
    469 			goto no_route;
    470 		sk_setup_caps(sk, &rt->dst);
    471 	}
    472 	skb_dst_set_noref(skb, &rt->dst);
    473 
    474 packet_routed:
    475 	if (inet_opt && inet_opt->opt.is_strictroute && rt->rt_uses_gateway)
    476 		goto no_route;
    477 
    478 	/* OK, we know where to send it, allocate and build IP header. */
    479 	skb_push(skb, sizeof(struct iphdr) + (inet_opt ? inet_opt->opt.optlen : 0));
    480 	skb_reset_network_header(skb);
    481 	iph = ip_hdr(skb);
    482 	*((__be16 *)iph) = htons((4 << 12) | (5 << 8) | (tos & 0xff));
    483 	if (ip_dont_fragment(sk, &rt->dst) && !skb->ignore_df)
    484 		iph->frag_off = htons(IP_DF);
    485 	else
    486 		iph->frag_off = 0;
    487 	iph->ttl      = ip_select_ttl(inet, &rt->dst);
    488 	iph->protocol = sk->sk_protocol;
    489 	ip_copy_addrs(iph, fl4);
    490 
    491 	/* Transport layer set skb->h.foo itself. */
    492 
    493 	if (inet_opt && inet_opt->opt.optlen) {
    494 		iph->ihl += inet_opt->opt.optlen >> 2;
    495 		ip_options_build(skb, &inet_opt->opt, inet->inet_daddr, rt, 0);
    496 	}
    497 
    498 	ip_select_ident_segs(net, skb, sk,
    499 			     skb_shinfo(skb)->gso_segs ?: 1);
    500 
    501 	/* TODO : should we use skb->sk here instead of sk ? */
    502 	skb->priority = sk->sk_priority;
    503 	skb->mark = sk->sk_mark;
    504 
    505 	res = ip_local_out(net, sk, skb);
    506 	rcu_read_unlock();
    507 	return res;
    508 
    509 no_route:
    510 	rcu_read_unlock();
    511 	IP_INC_STATS(net, IPSTATS_MIB_OUTNOROUTES);
    512 	kfree_skb(skb);
    513 	return -EHOSTUNREACH;
    514 }
    515 EXPORT_SYMBOL(__ip_queue_xmit);
    516 
    517 static void ip_copy_metadata(struct sk_buff *to, struct sk_buff *from)
    518 {
    519 	to->pkt_type = from->pkt_type;
    520 	to->priority = from->priority;
    521 	to->protocol = from->protocol;
    522 	skb_dst_drop(to);
    523 	skb_dst_copy(to, from);
    524 	to->dev = from->dev;
    525 	to->mark = from->mark;
    526 
    527 	skb_copy_hash(to, from);
    528 
    529 	/* Copy the flags to each fragment. */
    530 	IPCB(to)->flags = IPCB(from)->flags;
    531 
    532 #ifdef CONFIG_NET_SCHED
    533 	to->tc_index = from->tc_index;
    534 #endif
    535 	nf_copy(to, from);
    536 	skb_ext_copy(to, from);
    537 #if IS_ENABLED(CONFIG_IP_VS)
    538 	to->ipvs_property = from->ipvs_property;
    539 #endif
    540 	skb_copy_secmark(to, from);
    541 }
    542 
    543 static int ip_fragment(struct net *net, struct sock *sk, struct sk_buff *skb,
    544 		       unsigned int mtu,
    545 		       int (*output)(struct net *, struct sock *, struct sk_buff *))
    546 {
    547 	struct iphdr *iph = ip_hdr(skb);
    548 
    549 	if ((iph->frag_off & htons(IP_DF)) == 0)
    550 		return ip_do_fragment(net, sk, skb, output);
    551 
    552 	if (unlikely(!skb->ignore_df ||
    553 		     (IPCB(skb)->frag_max_size &&
    554 		      IPCB(skb)->frag_max_size > mtu))) {
    555 		IP_INC_STATS(net, IPSTATS_MIB_FRAGFAILS);
    556 		icmp_send(skb, ICMP_DEST_UNREACH, ICMP_FRAG_NEEDED,
    557 			  htonl(mtu));
    558 		kfree_skb(skb);
    559 		return -EMSGSIZE;
    560 	}
    561 
    562 	return ip_do_fragment(net, sk, skb, output);
    563 }
    564 
    565 /*
    566  *	This IP datagram is too large to be sent in one piece.  Break it up into
    567  *	smaller pieces (each of size equal to IP header plus
    568  *	a block of the data of the original IP data part) that will yet fit in a
    569  *	single device frame, and queue such a frame for sending.
    570  */
    571 
    572 int ip_do_fragment(struct net *net, struct sock *sk, struct sk_buff *skb,
    573 		   int (*output)(struct net *, struct sock *, struct sk_buff *))
    574 {
    575 	struct iphdr *iph;
    576 	int ptr;
    577 	struct sk_buff *skb2;
    578 	unsigned int mtu, hlen, left, len, ll_rs;
    579 	int offset;
    580 	__be16 not_last_frag;
    581 	struct rtable *rt = skb_rtable(skb);
    582 	int err = 0;
    583 
    584 	/* for offloaded checksums cleanup checksum before fragmentation */
    585 	if (skb->ip_summed == CHECKSUM_PARTIAL &&
    586 	    (err = skb_checksum_help(skb)))
    587 		goto fail;
    588 
    589 	/*
    590 	 *	Point into the IP datagram header.
    591 	 */
    592 
    593 	iph = ip_hdr(skb);
    594 
    595 	mtu = ip_skb_dst_mtu(sk, skb);
    596 	if (IPCB(skb)->frag_max_size && IPCB(skb)->frag_max_size < mtu)
    597 		mtu = IPCB(skb)->frag_max_size;
    598 
    599 	/*
    600 	 *	Setup starting values.
    601 	 */
    602 
    603 	hlen = iph->ihl * 4;
    604 	mtu = mtu - hlen;	/* Size of data space */
    605 	IPCB(skb)->flags |= IPSKB_FRAG_COMPLETE;
    606 	ll_rs = LL_RESERVED_SPACE(rt->dst.dev);
    607 
    608 	/* When frag_list is given, use it. First, check its validity:
    609 	 * some transformers could create wrong frag_list or break existing
    610 	 * one, it is not prohibited. In this case fall back to copying.
    611 	 *
    612 	 * LATER: this step can be merged to real generation of fragments,
    613 	 * we can switch to copy when see the first bad fragment.
    614 	 */
    615 	if (skb_has_frag_list(skb)) {
    616 		struct sk_buff *frag, *frag2;
    617 		unsigned int first_len = skb_pagelen(skb);
    618 
    619 		if (first_len - hlen > mtu ||
    620 		    ((first_len - hlen) & 7) ||
    621 		    ip_is_fragment(iph) ||
    622 		    skb_cloned(skb) ||
    623 		    skb_headroom(skb) < ll_rs)
    624 			goto slow_path;
    625 
    626 		skb_walk_frags(skb, frag) {
    627 			/* Correct geometry. */
    628 			if (frag->len > mtu ||
    629 			    ((frag->len & 7) && frag->next) ||
    630 			    skb_headroom(frag) < hlen + ll_rs)
    631 				goto slow_path_clean;
    632 
    633 			/* Partially cloned skb? */
    634 			if (skb_shared(frag))
    635 				goto slow_path_clean;
    636 
    637 			BUG_ON(frag->sk);
    638 			if (skb->sk) {
    639 				frag->sk = skb->sk;
    640 				frag->destructor = sock_wfree;
    641 			}
    642 			skb->truesize -= frag->truesize;
    643 		}
    644 
    645 		/* Everything is OK. Generate! */
    646 
    647 		err = 0;
    648 		offset = 0;
    649 		frag = skb_shinfo(skb)->frag_list;
    650 		skb_frag_list_init(skb);
    651 		skb->data_len = first_len - skb_headlen(skb);
    652 		skb->len = first_len;
    653 		iph->tot_len = htons(first_len);
    654 		iph->frag_off = htons(IP_MF);
    655 		ip_send_check(iph);
    656 
    657 		for (;;) {
    658 			/* Prepare header of the next frame,
    659 			 * before previous one went down. */
    660 			if (frag) {
    661 				frag->ip_summed = CHECKSUM_NONE;
    662 				skb_reset_transport_header(frag);
    663 				__skb_push(frag, hlen);
    664 				skb_reset_network_header(frag);
    665 				memcpy(skb_network_header(frag), iph, hlen);
    666 				iph = ip_hdr(frag);
    667 				iph->tot_len = htons(frag->len);
    668 				ip_copy_metadata(frag, skb);
    669 				if (offset == 0)
    670 					ip_options_fragment(frag);
    671 				offset += skb->len - hlen;
    672 				iph->frag_off = htons(offset>>3);
    673 				if (frag->next)
    674 					iph->frag_off |= htons(IP_MF);
    675 				/* Ready, complete checksum */
    676 				ip_send_check(iph);
    677 			}
    678 
    679 			err = output(net, sk, skb);
    680 
    681 			if (!err)
    682 				IP_INC_STATS(net, IPSTATS_MIB_FRAGCREATES);
    683 			if (err || !frag)
    684 				break;
    685 
    686 			skb = frag;
    687 			frag = skb->next;
    688 			skb_mark_not_on_list(skb);
    689 		}
    690 
    691 		if (err == 0) {
    692 			IP_INC_STATS(net, IPSTATS_MIB_FRAGOKS);
    693 			return 0;
    694 		}
    695 
    696 		while (frag) {
    697 			skb = frag->next;
    698 			kfree_skb(frag);
    699 			frag = skb;
    700 		}
    701 		IP_INC_STATS(net, IPSTATS_MIB_FRAGFAILS);
    702 		return err;
    703 
    704 slow_path_clean:
    705 		skb_walk_frags(skb, frag2) {
    706 			if (frag2 == frag)
    707 				break;
    708 			frag2->sk = NULL;
    709 			frag2->destructor = NULL;
    710 			skb->truesize += frag2->truesize;
    711 		}
    712 	}
    713 
    714 slow_path:
    715 	iph = ip_hdr(skb);
    716 
    717 	left = skb->len - hlen;		/* Space per frame */
    718 	ptr = hlen;		/* Where to start from */
    719 
    720 	/*
    721 	 *	Fragment the datagram.
    722 	 */
    723 
    724 	offset = (ntohs(iph->frag_off) & IP_OFFSET) << 3;
    725 	not_last_frag = iph->frag_off & htons(IP_MF);
    726 
    727 	/*
    728 	 *	Keep copying data until we run out.
    729 	 */
    730 
    731 	while (left > 0) {
    732 		len = left;
    733 		/* IF: it doesn't fit, use 'mtu' - the data space left */
    734 		if (len > mtu)
    735 			len = mtu;
    736 		/* IF: we are not sending up to and including the packet end
    737 		   then align the next start on an eight byte boundary */
    738 		if (len < left)	{
    739 			len &= ~7;
    740 		}
    741 
    742 		/* Allocate buffer */
    743 		skb2 = alloc_skb(len + hlen + ll_rs, GFP_ATOMIC);
    744 		if (!skb2) {
    745 			err = -ENOMEM;
    746 			goto fail;
    747 		}
    748 
    749 		/*
    750 		 *	Set up data on packet
    751 		 */
    752 
    753 		ip_copy_metadata(skb2, skb);
    754 		skb_reserve(skb2, ll_rs);
    755 		skb_put(skb2, len + hlen);
    756 		skb_reset_network_header(skb2);
    757 		skb2->transport_header = skb2->network_header + hlen;
    758 
    759 		/*
    760 		 *	Charge the memory for the fragment to any owner
    761 		 *	it might possess
    762 		 */
    763 
    764 		if (skb->sk)
    765 			skb_set_owner_w(skb2, skb->sk);
    766 
    767 		/*
    768 		 *	Copy the packet header into the new buffer.
    769 		 */
    770 
    771 		skb_copy_from_linear_data(skb, skb_network_header(skb2), hlen);
    772 
    773 		/*
    774 		 *	Copy a block of the IP datagram.
    775 		 */
    776 		if (skb_copy_bits(skb, ptr, skb_transport_header(skb2), len))
    777 			BUG();
    778 		left -= len;
    779 
    780 		/*
    781 		 *	Fill in the new header fields.
    782 		 */
    783 		iph = ip_hdr(skb2);
    784 		iph->frag_off = htons((offset >> 3));
    785 
    786 		if (IPCB(skb)->flags & IPSKB_FRAG_PMTU)
    787 			iph->frag_off |= htons(IP_DF);
    788 
    789 		/* ANK: dirty, but effective trick. Upgrade options only if
    790 		 * the segment to be fragmented was THE FIRST (otherwise,
    791 		 * options are already fixed) and make it ONCE
    792 		 * on the initial skb, so that all the following fragments
    793 		 * will inherit fixed options.
    794 		 */
    795 		if (offset == 0)
    796 			ip_options_fragment(skb);
    797 
    798 		/*
    799 		 *	Added AC : If we are fragmenting a fragment that's not the
    800 		 *		   last fragment then keep MF on each bit
    801 		 */
    802 		if (left > 0 || not_last_frag)
    803 			iph->frag_off |= htons(IP_MF);
    804 		ptr += len;
    805 		offset += len;
    806 
    807 		/*
    808 		 *	Put this fragment into the sending queue.
    809 		 */
    810 		iph->tot_len = htons(len + hlen);
    811 
    812 		ip_send_check(iph);
    813 
    814 		err = output(net, sk, skb2);
    815 		if (err)
    816 			goto fail;
    817 
    818 		IP_INC_STATS(net, IPSTATS_MIB_FRAGCREATES);
    819 	}
    820 	consume_skb(skb);
    821 	IP_INC_STATS(net, IPSTATS_MIB_FRAGOKS);
    822 	return err;
    823 
    824 fail:
    825 	kfree_skb(skb);
    826 	IP_INC_STATS(net, IPSTATS_MIB_FRAGFAILS);
    827 	return err;
    828 }
    829 EXPORT_SYMBOL(ip_do_fragment);
    830 
    831 int
    832 ip_generic_getfrag(void *from, char *to, int offset, int len, int odd, struct sk_buff *skb)
    833 {
    834 	struct msghdr *msg = from;
    835 
    836 	if (skb->ip_summed == CHECKSUM_PARTIAL) {
    837 		if (!copy_from_iter_full(to, len, &msg->msg_iter))
    838 			return -EFAULT;
    839 	} else {
    840 		__wsum csum = 0;
    841 		if (!csum_and_copy_from_iter_full(to, len, &csum, &msg->msg_iter))
    842 			return -EFAULT;
    843 		skb->csum = csum_block_add(skb->csum, csum, odd);
    844 	}
    845 	return 0;
    846 }
    847 EXPORT_SYMBOL(ip_generic_getfrag);
    848 
    849 static inline __wsum
    850 csum_page(struct page *page, int offset, int copy)
    851 {
    852 	char *kaddr;
    853 	__wsum csum;
    854 	kaddr = kmap(page);
    855 	csum = csum_partial(kaddr + offset, copy, 0);
    856 	kunmap(page);
    857 	return csum;
    858 }
    859 
    860 static int __ip_append_data(struct sock *sk,
    861 			    struct flowi4 *fl4,
    862 			    struct sk_buff_head *queue,
    863 			    struct inet_cork *cork,
    864 			    struct page_frag *pfrag,
    865 			    int getfrag(void *from, char *to, int offset,
    866 					int len, int odd, struct sk_buff *skb),
    867 			    void *from, int length, int transhdrlen,
    868 			    unsigned int flags)
    869 {
    870 	struct inet_sock *inet = inet_sk(sk);
    871 	struct ubuf_info *uarg = NULL;
    872 	struct sk_buff *skb;
    873 
    874 	struct ip_options *opt = cork->opt;
    875 	int hh_len;
    876 	int exthdrlen;
    877 	int mtu;
    878 	int copy;
    879 	int err;
    880 	int offset = 0;
    881 	unsigned int maxfraglen, fragheaderlen, maxnonfragsize;
    882 	int csummode = CHECKSUM_NONE;
    883 	struct rtable *rt = (struct rtable *)cork->dst;
    884 	unsigned int wmem_alloc_delta = 0;
    885 	bool paged, extra_uref;
    886 	u32 tskey = 0;
    887 
    888 	skb = skb_peek_tail(queue);
    889 
    890 	exthdrlen = !skb ? rt->dst.header_len : 0;
    891 	mtu = cork->gso_size ? IP_MAX_MTU : cork->fragsize;
    892 	paged = !!cork->gso_size;
    893 
    894 	if (cork->tx_flags & SKBTX_ANY_SW_TSTAMP &&
    895 	    sk->sk_tsflags & SOF_TIMESTAMPING_OPT_ID)
    896 		tskey = sk->sk_tskey++;
    897 
    898 	hh_len = LL_RESERVED_SPACE(rt->dst.dev);
    899 
    900 	fragheaderlen = sizeof(struct iphdr) + (opt ? opt->optlen : 0);
    901 	maxfraglen = ((mtu - fragheaderlen) & ~7) + fragheaderlen;
    902 	maxnonfragsize = ip_sk_ignore_df(sk) ? 0xFFFF : mtu;
    903 
    904 	if (cork->length + length > maxnonfragsize - fragheaderlen) {
    905 		ip_local_error(sk, EMSGSIZE, fl4->daddr, inet->inet_dport,
    906 			       mtu - (opt ? opt->optlen : 0));
    907 		return -EMSGSIZE;
    908 	}
    909 
    910 	/*
    911 	 * transhdrlen > 0 means that this is the first fragment and we wish
    912 	 * it won't be fragmented in the future.
    913 	 */
    914 	if (transhdrlen &&
    915 	    length + fragheaderlen <= mtu &&
    916 	    rt->dst.dev->features & (NETIF_F_HW_CSUM | NETIF_F_IP_CSUM) &&
    917 	    (!(flags & MSG_MORE) || cork->gso_size) &&
    918 	    (!exthdrlen || (rt->dst.dev->features & NETIF_F_HW_ESP_TX_CSUM)))
    919 		csummode = CHECKSUM_PARTIAL;
    920 
    921 	if (flags & MSG_ZEROCOPY && length && sock_flag(sk, SOCK_ZEROCOPY)) {
    922 		uarg = sock_zerocopy_realloc(sk, length, skb_zcopy(skb));
    923 		if (!uarg)
    924 			return -ENOBUFS;
    925 		extra_uref = true;
    926 		if (rt->dst.dev->features & NETIF_F_SG &&
    927 		    csummode == CHECKSUM_PARTIAL) {
    928 			paged = true;
    929 		} else {
    930 			uarg->zerocopy = 0;
    931 			skb_zcopy_set(skb, uarg, &extra_uref);
    932 		}
    933 	}
    934 
    935 	cork->length += length;
    936 
    937 	/* So, what's going on in the loop below?
    938 	 *
    939 	 * We use calculated fragment length to generate chained skb,
    940 	 * each of segments is IP fragment ready for sending to network after
    941 	 * adding appropriate IP header.
    942 	 */
    943 
    944 	if (!skb)
    945 		goto alloc_new_skb;
    946 
    947 	while (length > 0) {
    948 		/* Check if the remaining data fits into current packet. */
    949 		copy = mtu - skb->len;
    950 		if (copy < length)
    951 			copy = maxfraglen - skb->len;
    952 		if (copy <= 0) {
    953 			char *data;
    954 			unsigned int datalen;
    955 			unsigned int fraglen;
    956 			unsigned int fraggap;
    957 			unsigned int alloclen;
    958 			unsigned int pagedlen;
    959 			struct sk_buff *skb_prev;
    960 alloc_new_skb:
    961 			skb_prev = skb;
    962 			if (skb_prev)
    963 				fraggap = skb_prev->len - maxfraglen;
    964 			else
    965 				fraggap = 0;
    966 
    967 			/*
    968 			 * If remaining data exceeds the mtu,
    969 			 * we know we need more fragment(s).
    970 			 */
    971 			datalen = length + fraggap;
    972 			if (datalen > mtu - fragheaderlen)
    973 				datalen = maxfraglen - fragheaderlen;
    974 			fraglen = datalen + fragheaderlen;
    975 			pagedlen = 0;
    976 
    977 			if ((flags & MSG_MORE) &&
    978 			    !(rt->dst.dev->features&NETIF_F_SG))
    979 				alloclen = mtu;
    980 			else if (!paged)
    981 				alloclen = fraglen;
    982 			else {
    983 				alloclen = min_t(int, fraglen, MAX_HEADER);
    984 				pagedlen = fraglen - alloclen;
    985 			}
    986 
    987 			alloclen += exthdrlen;
    988 
    989 			/* The last fragment gets additional space at tail.
    990 			 * Note, with MSG_MORE we overallocate on fragments,
    991 			 * because we have no idea what fragment will be
    992 			 * the last.
    993 			 */
    994 			if (datalen == length + fraggap)
    995 				alloclen += rt->dst.trailer_len;
    996 
    997 			if (transhdrlen) {
    998 				skb = sock_alloc_send_skb(sk,
    999 						alloclen + hh_len + 15,
   1000 						(flags & MSG_DONTWAIT), &err);
   1001 			} else {
   1002 				skb = NULL;
   1003 				if (refcount_read(&sk->sk_wmem_alloc) + wmem_alloc_delta <=
   1004 				    2 * sk->sk_sndbuf)
   1005 					skb = alloc_skb(alloclen + hh_len + 15,
   1006 							sk->sk_allocation);
   1007 				if (unlikely(!skb))
   1008 					err = -ENOBUFS;
   1009 			}
   1010 			if (!skb)
   1011 				goto error;
   1012 
   1013 			/*
   1014 			 *	Fill in the control structures
   1015 			 */
   1016 			skb->ip_summed = csummode;
   1017 			skb->csum = 0;
   1018 			skb_reserve(skb, hh_len);
   1019 
   1020 			/*
   1021 			 *	Find where to start putting bytes.
   1022 			 */
   1023 			data = skb_put(skb, fraglen + exthdrlen - pagedlen);
   1024 			skb_set_network_header(skb, exthdrlen);
   1025 			skb->transport_header = (skb->network_header +
   1026 						 fragheaderlen);
   1027 			data += fragheaderlen + exthdrlen;
   1028 
   1029 			if (fraggap) {
   1030 				skb->csum = skb_copy_and_csum_bits(
   1031 					skb_prev, maxfraglen,
   1032 					data + transhdrlen, fraggap, 0);
   1033 				skb_prev->csum = csum_sub(skb_prev->csum,
   1034 							  skb->csum);
   1035 				data += fraggap;
   1036 				pskb_trim_unique(skb_prev, maxfraglen);
   1037 			}
   1038 
   1039 			copy = datalen - transhdrlen - fraggap - pagedlen;
   1040 			if (copy > 0 && getfrag(from, data + transhdrlen, offset, copy, fraggap, skb) < 0) {
   1041 				err = -EFAULT;
   1042 				kfree_skb(skb);
   1043 				goto error;
   1044 			}
   1045 
   1046 			offset += copy;
   1047 			length -= copy + transhdrlen;
   1048 			transhdrlen = 0;
   1049 			exthdrlen = 0;
   1050 			csummode = CHECKSUM_NONE;
   1051 
   1052 			/* only the initial fragment is time stamped */
   1053 			skb_shinfo(skb)->tx_flags = cork->tx_flags;
   1054 			cork->tx_flags = 0;
   1055 			skb_shinfo(skb)->tskey = tskey;
   1056 			tskey = 0;
   1057 			skb_zcopy_set(skb, uarg, &extra_uref);
   1058 
   1059 			if ((flags & MSG_CONFIRM) && !skb_prev)
   1060 				skb_set_dst_pending_confirm(skb, 1);
   1061 
   1062 			/*
   1063 			 * Put the packet on the pending queue.
   1064 			 */
   1065 			if (!skb->destructor) {
   1066 				skb->destructor = sock_wfree;
   1067 				skb->sk = sk;
   1068 				wmem_alloc_delta += skb->truesize;
   1069 			}
   1070 			__skb_queue_tail(queue, skb);
   1071 			continue;
   1072 		}
   1073 
   1074 		if (copy > length)
   1075 			copy = length;
   1076 
   1077 		if (!(rt->dst.dev->features&NETIF_F_SG) &&
   1078 		    skb_tailroom(skb) >= copy) {
   1079 			unsigned int off;
   1080 
   1081 			off = skb->len;
   1082 			if (getfrag(from, skb_put(skb, copy),
   1083 					offset, copy, off, skb) < 0) {
   1084 				__skb_trim(skb, off);
   1085 				err = -EFAULT;
   1086 				goto error;
   1087 			}
   1088 		} else if (!uarg || !uarg->zerocopy) {
   1089 			int i = skb_shinfo(skb)->nr_frags;
   1090 
   1091 			err = -ENOMEM;
   1092 			if (!sk_page_frag_refill(sk, pfrag))
   1093 				goto error;
   1094 
   1095 			if (!skb_can_coalesce(skb, i, pfrag->page,
   1096 					      pfrag->offset)) {
   1097 				err = -EMSGSIZE;
   1098 				if (i == MAX_SKB_FRAGS)
   1099 					goto error;
   1100 
   1101 				__skb_fill_page_desc(skb, i, pfrag->page,
   1102 						     pfrag->offset, 0);
   1103 				skb_shinfo(skb)->nr_frags = ++i;
   1104 				get_page(pfrag->page);
   1105 			}
   1106 			copy = min_t(int, copy, pfrag->size - pfrag->offset);
   1107 			if (getfrag(from,
   1108 				    page_address(pfrag->page) + pfrag->offset,
   1109 				    offset, copy, skb->len, skb) < 0)
   1110 				goto error_efault;
   1111 
   1112 			pfrag->offset += copy;
   1113 			skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
   1114 			skb->len += copy;
   1115 			skb->data_len += copy;
   1116 			skb->truesize += copy;
   1117 			wmem_alloc_delta += copy;
   1118 		} else {
   1119 			err = skb_zerocopy_iter_dgram(skb, from, copy);
   1120 			if (err < 0)
   1121 				goto error;
   1122 		}
   1123 		offset += copy;
   1124 		length -= copy;
   1125 	}
   1126 
   1127 	if (wmem_alloc_delta)
   1128 		refcount_add(wmem_alloc_delta, &sk->sk_wmem_alloc);
   1129 	return 0;
   1130 
   1131 error_efault:
   1132 	err = -EFAULT;
   1133 error:
   1134 	if (uarg)
   1135 		sock_zerocopy_put_abort(uarg, extra_uref);
   1136 	cork->length -= length;
   1137 	IP_INC_STATS(sock_net(sk), IPSTATS_MIB_OUTDISCARDS);
   1138 	refcount_add(wmem_alloc_delta, &sk->sk_wmem_alloc);
   1139 	return err;
   1140 }
   1141 
   1142 static int ip_setup_cork(struct sock *sk, struct inet_cork *cork,
   1143 			 struct ipcm_cookie *ipc, struct rtable **rtp)
   1144 {
   1145 	struct ip_options_rcu *opt;
   1146 	struct rtable *rt;
   1147 
   1148 	rt = *rtp;
   1149 	if (unlikely(!rt))
   1150 		return -EFAULT;
   1151 
   1152 	/*
   1153 	 * setup for corking.
   1154 	 */
   1155 	opt = ipc->opt;
   1156 	if (opt) {
   1157 		if (!cork->opt) {
   1158 			cork->opt = kmalloc(sizeof(struct ip_options) + 40,
   1159 					    sk->sk_allocation);
   1160 			if (unlikely(!cork->opt))
   1161 				return -ENOBUFS;
   1162 		}
   1163 		memcpy(cork->opt, &opt->opt, sizeof(struct ip_options) + opt->opt.optlen);
   1164 		cork->flags |= IPCORK_OPT;
   1165 		cork->addr = ipc->addr;
   1166 	}
   1167 
   1168 	/*
   1169 	 * We steal reference to this route, caller should not release it
   1170 	 */
   1171 	*rtp = NULL;
   1172 	cork->fragsize = ip_sk_use_pmtu(sk) ?
   1173 			 dst_mtu(&rt->dst) : rt->dst.dev->mtu;
   1174 
   1175 	cork->gso_size = ipc->gso_size;
   1176 	cork->dst = &rt->dst;
   1177 	cork->length = 0;
   1178 	cork->ttl = ipc->ttl;
   1179 	cork->tos = ipc->tos;
   1180 	cork->priority = ipc->priority;
   1181 	cork->transmit_time = ipc->sockc.transmit_time;
   1182 	cork->tx_flags = 0;
   1183 	sock_tx_timestamp(sk, ipc->sockc.tsflags, &cork->tx_flags);
   1184 
   1185 	return 0;
   1186 }
   1187 
   1188 /*
   1189  *	ip_append_data() and ip_append_page() can make one large IP datagram
   1190  *	from many pieces of data. Each pieces will be holded on the socket
   1191  *	until ip_push_pending_frames() is called. Each piece can be a page
   1192  *	or non-page data.
   1193  *
   1194  *	Not only UDP, other transport protocols - e.g. raw sockets - can use
   1195  *	this interface potentially.
   1196  *
   1197  *	LATER: length must be adjusted by pad at tail, when it is required.
   1198  */
   1199 int ip_append_data(struct sock *sk, struct flowi4 *fl4,
   1200 		   int getfrag(void *from, char *to, int offset, int len,
   1201 			       int odd, struct sk_buff *skb),
   1202 		   void *from, int length, int transhdrlen,
   1203 		   struct ipcm_cookie *ipc, struct rtable **rtp,
   1204 		   unsigned int flags)
   1205 {
   1206 	struct inet_sock *inet = inet_sk(sk);
   1207 	int err;
   1208 
   1209 	if (flags&MSG_PROBE)
   1210 		return 0;
   1211 
   1212 	if (skb_queue_empty(&sk->sk_write_queue)) {
   1213 		err = ip_setup_cork(sk, &inet->cork.base, ipc, rtp);
   1214 		if (err)
   1215 			return err;
   1216 	} else {
   1217 		transhdrlen = 0;
   1218 	}
   1219 
   1220 	return __ip_append_data(sk, fl4, &sk->sk_write_queue, &inet->cork.base,
   1221 				sk_page_frag(sk), getfrag,
   1222 				from, length, transhdrlen, flags);
   1223 }
   1224 
   1225 ssize_t	ip_append_page(struct sock *sk, struct flowi4 *fl4, struct page *page,
   1226 		       int offset, size_t size, int flags)
   1227 {
   1228 	struct inet_sock *inet = inet_sk(sk);
   1229 	struct sk_buff *skb;
   1230 	struct rtable *rt;
   1231 	struct ip_options *opt = NULL;
   1232 	struct inet_cork *cork;
   1233 	int hh_len;
   1234 	int mtu;
   1235 	int len;
   1236 	int err;
   1237 	unsigned int maxfraglen, fragheaderlen, fraggap, maxnonfragsize;
   1238 
   1239 	if (inet->hdrincl)
   1240 		return -EPERM;
   1241 
   1242 	if (flags&MSG_PROBE)
   1243 		return 0;
   1244 
   1245 	if (skb_queue_empty(&sk->sk_write_queue))
   1246 		return -EINVAL;
   1247 
   1248 	cork = &inet->cork.base;
   1249 	rt = (struct rtable *)cork->dst;
   1250 	if (cork->flags & IPCORK_OPT)
   1251 		opt = cork->opt;
   1252 
   1253 	if (!(rt->dst.dev->features&NETIF_F_SG))
   1254 		return -EOPNOTSUPP;
   1255 
   1256 	hh_len = LL_RESERVED_SPACE(rt->dst.dev);
   1257 	mtu = cork->gso_size ? IP_MAX_MTU : cork->fragsize;
   1258 
   1259 	fragheaderlen = sizeof(struct iphdr) + (opt ? opt->optlen : 0);
   1260 	maxfraglen = ((mtu - fragheaderlen) & ~7) + fragheaderlen;
   1261 	maxnonfragsize = ip_sk_ignore_df(sk) ? 0xFFFF : mtu;
   1262 
   1263 	if (cork->length + size > maxnonfragsize - fragheaderlen) {
   1264 		ip_local_error(sk, EMSGSIZE, fl4->daddr, inet->inet_dport,
   1265 			       mtu - (opt ? opt->optlen : 0));
   1266 		return -EMSGSIZE;
   1267 	}
   1268 
   1269 	skb = skb_peek_tail(&sk->sk_write_queue);
   1270 	if (!skb)
   1271 		return -EINVAL;
   1272 
   1273 	cork->length += size;
   1274 
   1275 	while (size > 0) {
   1276 		/* Check if the remaining data fits into current packet. */
   1277 		len = mtu - skb->len;
   1278 		if (len < size)
   1279 			len = maxfraglen - skb->len;
   1280 
   1281 		if (len <= 0) {
   1282 			struct sk_buff *skb_prev;
   1283 			int alloclen;
   1284 
   1285 			skb_prev = skb;
   1286 			fraggap = skb_prev->len - maxfraglen;
   1287 
   1288 			alloclen = fragheaderlen + hh_len + fraggap + 15;
   1289 			skb = sock_wmalloc(sk, alloclen, 1, sk->sk_allocation);
   1290 			if (unlikely(!skb)) {
   1291 				err = -ENOBUFS;
   1292 				goto error;
   1293 			}
   1294 
   1295 			/*
   1296 			 *	Fill in the control structures
   1297 			 */
   1298 			skb->ip_summed = CHECKSUM_NONE;
   1299 			skb->csum = 0;
   1300 			skb_reserve(skb, hh_len);
   1301 
   1302 			/*
   1303 			 *	Find where to start putting bytes.
   1304 			 */
   1305 			skb_put(skb, fragheaderlen + fraggap);
   1306 			skb_reset_network_header(skb);
   1307 			skb->transport_header = (skb->network_header +
   1308 						 fragheaderlen);
   1309 			if (fraggap) {
   1310 				skb->csum = skb_copy_and_csum_bits(skb_prev,
   1311 								   maxfraglen,
   1312 						    skb_transport_header(skb),
   1313 								   fraggap, 0);
   1314 				skb_prev->csum = csum_sub(skb_prev->csum,
   1315 							  skb->csum);
   1316 				pskb_trim_unique(skb_prev, maxfraglen);
   1317 			}
   1318 
   1319 			/*
   1320 			 * Put the packet on the pending queue.
   1321 			 */
   1322 			__skb_queue_tail(&sk->sk_write_queue, skb);
   1323 			continue;
   1324 		}
   1325 
   1326 		if (len > size)
   1327 			len = size;
   1328 
   1329 		if (skb_append_pagefrags(skb, page, offset, len)) {
   1330 			err = -EMSGSIZE;
   1331 			goto error;
   1332 		}
   1333 
   1334 		if (skb->ip_summed == CHECKSUM_NONE) {
   1335 			__wsum csum;
   1336 			csum = csum_page(page, offset, len);
   1337 			skb->csum = csum_block_add(skb->csum, csum, skb->len);
   1338 		}
   1339 
   1340 		skb->len += len;
   1341 		skb->data_len += len;
   1342 		skb->truesize += len;
   1343 		refcount_add(len, &sk->sk_wmem_alloc);
   1344 		offset += len;
   1345 		size -= len;
   1346 	}
   1347 	return 0;
   1348 
   1349 error:
   1350 	cork->length -= size;
   1351 	IP_INC_STATS(sock_net(sk), IPSTATS_MIB_OUTDISCARDS);
   1352 	return err;
   1353 }
   1354 
   1355 static void ip_cork_release(struct inet_cork *cork)
   1356 {
   1357 	cork->flags &= ~IPCORK_OPT;
   1358 	kfree(cork->opt);
   1359 	cork->opt = NULL;
   1360 	dst_release(cork->dst);
   1361 	cork->dst = NULL;
   1362 }
   1363 
   1364 /*
   1365  *	Combined all pending IP fragments on the socket as one IP datagram
   1366  *	and push them out.
   1367  */
   1368 struct sk_buff *__ip_make_skb(struct sock *sk,
   1369 			      struct flowi4 *fl4,
   1370 			      struct sk_buff_head *queue,
   1371 			      struct inet_cork *cork)
   1372 {
   1373 	struct sk_buff *skb, *tmp_skb;
   1374 	struct sk_buff **tail_skb;
   1375 	struct inet_sock *inet = inet_sk(sk);
   1376 	struct net *net = sock_net(sk);
   1377 	struct ip_options *opt = NULL;
   1378 	struct rtable *rt = (struct rtable *)cork->dst;
   1379 	struct iphdr *iph;
   1380 	__be16 df = 0;
   1381 	__u8 ttl;
   1382 
   1383 	skb = __skb_dequeue(queue);
   1384 	if (!skb)
   1385 		goto out;
   1386 	tail_skb = &(skb_shinfo(skb)->frag_list);
   1387 
   1388 	/* move skb->data to ip header from ext header */
   1389 	if (skb->data < skb_network_header(skb))
   1390 		__skb_pull(skb, skb_network_offset(skb));
   1391 	while ((tmp_skb = __skb_dequeue(queue)) != NULL) {
   1392 		__skb_pull(tmp_skb, skb_network_header_len(skb));
   1393 		*tail_skb = tmp_skb;
   1394 		tail_skb = &(tmp_skb->next);
   1395 		skb->len += tmp_skb->len;
   1396 		skb->data_len += tmp_skb->len;
   1397 		skb->truesize += tmp_skb->truesize;
   1398 		tmp_skb->destructor = NULL;
   1399 		tmp_skb->sk = NULL;
   1400 	}
   1401 
   1402 	/* Unless user demanded real pmtu discovery (IP_PMTUDISC_DO), we allow
   1403 	 * to fragment the frame generated here. No matter, what transforms
   1404 	 * how transforms change size of the packet, it will come out.
   1405 	 */
   1406 	skb->ignore_df = ip_sk_ignore_df(sk);
   1407 
   1408 	/* DF bit is set when we want to see DF on outgoing frames.
   1409 	 * If ignore_df is set too, we still allow to fragment this frame
   1410 	 * locally. */
   1411 	if (inet->pmtudisc == IP_PMTUDISC_DO ||
   1412 	    inet->pmtudisc == IP_PMTUDISC_PROBE ||
   1413 	    (skb->len <= dst_mtu(&rt->dst) &&
   1414 	     ip_dont_fragment(sk, &rt->dst)))
   1415 		df = htons(IP_DF);
   1416 
   1417 	if (cork->flags & IPCORK_OPT)
   1418 		opt = cork->opt;
   1419 
   1420 	if (cork->ttl != 0)
   1421 		ttl = cork->ttl;
   1422 	else if (rt->rt_type == RTN_MULTICAST)
   1423 		ttl = inet->mc_ttl;
   1424 	else
   1425 		ttl = ip_select_ttl(inet, &rt->dst);
   1426 
   1427 	iph = ip_hdr(skb);
   1428 	iph->version = 4;
   1429 	iph->ihl = 5;
   1430 	iph->tos = (cork->tos != -1) ? cork->tos : inet->tos;
   1431 	iph->frag_off = df;
   1432 	iph->ttl = ttl;
   1433 	iph->protocol = sk->sk_protocol;
   1434 	ip_copy_addrs(iph, fl4);
   1435 	ip_select_ident(net, skb, sk);
   1436 
   1437 	if (opt) {
   1438 		iph->ihl += opt->optlen>>2;
   1439 		ip_options_build(skb, opt, cork->addr, rt, 0);
   1440 	}
   1441 
   1442 	skb->priority = (cork->tos != -1) ? cork->priority: sk->sk_priority;
   1443 	skb->mark = sk->sk_mark;
   1444 	skb->tstamp = cork->transmit_time;
   1445 	/*
   1446 	 * Steal rt from cork.dst to avoid a pair of atomic_inc/atomic_dec
   1447 	 * on dst refcount
   1448 	 */
   1449 	cork->dst = NULL;
   1450 	skb_dst_set(skb, &rt->dst);
   1451 
   1452 	if (iph->protocol == IPPROTO_ICMP)
   1453 		icmp_out_count(net, ((struct icmphdr *)
   1454 			skb_transport_header(skb))->type);
   1455 
   1456 	ip_cork_release(cork);
   1457 out:
   1458 	return skb;
   1459 }
   1460 
   1461 int ip_send_skb(struct net *net, struct sk_buff *skb)
   1462 {
   1463 	int err;
   1464 
   1465 	err = ip_local_out(net, skb->sk, skb);
   1466 	if (err) {
   1467 		if (err > 0)
   1468 			err = net_xmit_errno(err);
   1469 		if (err)
   1470 			IP_INC_STATS(net, IPSTATS_MIB_OUTDISCARDS);
   1471 	}
   1472 
   1473 	return err;
   1474 }
   1475 
   1476 int ip_push_pending_frames(struct sock *sk, struct flowi4 *fl4)
   1477 {
   1478 	struct sk_buff *skb;
   1479 
   1480 	skb = ip_finish_skb(sk, fl4);
   1481 	if (!skb)
   1482 		return 0;
   1483 
   1484 	/* Netfilter gets whole the not fragmented skb. */
   1485 	return ip_send_skb(sock_net(sk), skb);
   1486 }
   1487 
   1488 /*
   1489  *	Throw away all pending data on the socket.
   1490  */
   1491 static void __ip_flush_pending_frames(struct sock *sk,
   1492 				      struct sk_buff_head *queue,
   1493 				      struct inet_cork *cork)
   1494 {
   1495 	struct sk_buff *skb;
   1496 
   1497 	while ((skb = __skb_dequeue_tail(queue)) != NULL)
   1498 		kfree_skb(skb);
   1499 
   1500 	ip_cork_release(cork);
   1501 }
   1502 
   1503 void ip_flush_pending_frames(struct sock *sk)
   1504 {
   1505 	__ip_flush_pending_frames(sk, &sk->sk_write_queue, &inet_sk(sk)->cork.base);
   1506 }
   1507 
   1508 struct sk_buff *ip_make_skb(struct sock *sk,
   1509 			    struct flowi4 *fl4,
   1510 			    int getfrag(void *from, char *to, int offset,
   1511 					int len, int odd, struct sk_buff *skb),
   1512 			    void *from, int length, int transhdrlen,
   1513 			    struct ipcm_cookie *ipc, struct rtable **rtp,
   1514 			    struct inet_cork *cork, unsigned int flags)
   1515 {
   1516 	struct sk_buff_head queue;
   1517 	int err;
   1518 
   1519 	if (flags & MSG_PROBE)
   1520 		return NULL;
   1521 
   1522 	__skb_queue_head_init(&queue);
   1523 
   1524 	cork->flags = 0;
   1525 	cork->addr = 0;
   1526 	cork->opt = NULL;
   1527 	err = ip_setup_cork(sk, cork, ipc, rtp);
   1528 	if (err)
   1529 		return ERR_PTR(err);
   1530 
   1531 	err = __ip_append_data(sk, fl4, &queue, cork,
   1532 			       &current->task_frag, getfrag,
   1533 			       from, length, transhdrlen, flags);
   1534 	if (err) {
   1535 		__ip_flush_pending_frames(sk, &queue, cork);
   1536 		return ERR_PTR(err);
   1537 	}
   1538 
   1539 	return __ip_make_skb(sk, fl4, &queue, cork);
   1540 }
   1541 
   1542 /*
   1543  *	Fetch data from kernel space and fill in checksum if needed.
   1544  */
   1545 static int ip_reply_glue_bits(void *dptr, char *to, int offset,
   1546 			      int len, int odd, struct sk_buff *skb)
   1547 {
   1548 	__wsum csum;
   1549 
   1550 	csum = csum_partial_copy_nocheck(dptr+offset, to, len, 0);
   1551 	skb->csum = csum_block_add(skb->csum, csum, odd);
   1552 	return 0;
   1553 }
   1554 
   1555 /*
   1556  *	Generic function to send a packet as reply to another packet.
   1557  *	Used to send some TCP resets/acks so far.
   1558  */
   1559 void ip_send_unicast_reply(struct sock *sk, struct sk_buff *skb,
   1560 			   const struct ip_options *sopt,
   1561 			   __be32 daddr, __be32 saddr,
   1562 			   const struct ip_reply_arg *arg,
   1563 			   unsigned int len)
   1564 {
   1565 	struct ip_options_data replyopts;
   1566 	struct ipcm_cookie ipc;
   1567 	struct flowi4 fl4;
   1568 	struct rtable *rt = skb_rtable(skb);
   1569 	struct net *net = sock_net(sk);
   1570 	struct sk_buff *nskb;
   1571 	int err;
   1572 	int oif;
   1573 
   1574 	if (__ip_options_echo(net, &replyopts.opt.opt, skb, sopt))
   1575 		return;
   1576 
   1577 	ipcm_init(&ipc);
   1578 	ipc.addr = daddr;
   1579 
   1580 	if (replyopts.opt.opt.optlen) {
   1581 		ipc.opt = &replyopts.opt;
   1582 
   1583 		if (replyopts.opt.opt.srr)
   1584 			daddr = replyopts.opt.opt.faddr;
   1585 	}
   1586 
   1587 	oif = arg->bound_dev_if;
   1588 	if (!oif && netif_index_is_l3_master(net, skb->skb_iif))
   1589 		oif = skb->skb_iif;
   1590 
   1591 	flowi4_init_output(&fl4, oif,
   1592 			   IP4_REPLY_MARK(net, skb->mark) ?: sk->sk_mark,
   1593 			   RT_TOS(arg->tos),
   1594 			   RT_SCOPE_UNIVERSE, ip_hdr(skb)->protocol,
   1595 			   ip_reply_arg_flowi_flags(arg),
   1596 			   daddr, saddr,
   1597 			   tcp_hdr(skb)->source, tcp_hdr(skb)->dest,
   1598 			   arg->uid);
   1599 	security_skb_classify_flow(skb, flowi4_to_flowi(&fl4));
   1600 	rt = ip_route_output_key(net, &fl4);
   1601 	if (IS_ERR(rt))
   1602 		return;
   1603 
   1604 	inet_sk(sk)->tos = arg->tos;
   1605 
   1606 	sk->sk_priority = skb->priority;
   1607 	sk->sk_protocol = ip_hdr(skb)->protocol;
   1608 	sk->sk_bound_dev_if = arg->bound_dev_if;
   1609 	sk->sk_sndbuf = sysctl_wmem_default;
   1610 	sk->sk_mark = fl4.flowi4_mark;
   1611 	err = ip_append_data(sk, &fl4, ip_reply_glue_bits, arg->iov->iov_base,
   1612 			     len, 0, &ipc, &rt, MSG_DONTWAIT);
   1613 	if (unlikely(err)) {
   1614 		ip_flush_pending_frames(sk);
   1615 		goto out;
   1616 	}
   1617 
   1618 	nskb = skb_peek(&sk->sk_write_queue);
   1619 	if (nskb) {
   1620 		if (arg->csumoffset >= 0)
   1621 			*((__sum16 *)skb_transport_header(nskb) +
   1622 			  arg->csumoffset) = csum_fold(csum_add(nskb->csum,
   1623 								arg->csum));
   1624 		nskb->ip_summed = CHECKSUM_NONE;
   1625 		ip_push_pending_frames(sk, &fl4);
   1626 	}
   1627 out:
   1628 	ip_rt_put(rt);
   1629 }
   1630 
   1631 void __init ip_init(void)
   1632 {
   1633 	ip_rt_init();
   1634 	inet_initpeers();
   1635 
   1636 #if defined(CONFIG_IP_MULTICAST)
   1637 	igmp_mc_init();
   1638 #endif
   1639 }