inet_sock.h

  1/* SPDX-License-Identifier: GPL-2.0-or-later */
  2/*
  3 * INET		An implementation of the TCP/IP protocol suite for the LINUX
  4 *		operating system.  INET is implemented using the  BSD Socket
  5 *		interface as the means of communication with the user level.
  6 *
  7 *		Definitions for inet_sock
  8 *
  9 * Authors:	Many, reorganised here by
 10 * 		Arnaldo Carvalho de Melo <acme@mandriva.com>
 11 */
 12#ifndef _INET_SOCK_H
 13#define _INET_SOCK_H
 14
 15#include <linux/bitops.h>
 16#include <linux/string.h>
 17#include <linux/types.h>
 18#include <linux/jhash.h>
 19#include <linux/netdevice.h>
 20
 21#include <net/flow.h>
 22#include <net/inet_dscp.h>
 23#include <net/sock.h>
 24#include <net/request_sock.h>
 25#include <net/netns/hash.h>
 26#include <net/tcp_states.h>
 27#include <net/l3mdev.h>
 28
 29/** struct ip_options - IP Options
 30 *
 31 * @faddr - Saved first hop address
 32 * @nexthop - Saved nexthop address in LSRR and SSRR
 33 * @is_strictroute - Strict source route
 34 * @srr_is_hit - Packet destination addr was our one
 35 * @is_changed - IP checksum more not valid
 36 * @rr_needaddr - Need to record addr of outgoing dev
 37 * @ts_needtime - Need to record timestamp
 38 * @ts_needaddr - Need to record addr of outgoing dev
 39 */
 40struct ip_options {
 41	__be32		faddr;
 42	__be32		nexthop;
 43	unsigned char	optlen;
 44	unsigned char	srr;
 45	unsigned char	rr;
 46	unsigned char	ts;
 47	unsigned char	is_strictroute:1,
 48			srr_is_hit:1,
 49			is_changed:1,
 50			rr_needaddr:1,
 51			ts_needtime:1,
 52			ts_needaddr:1;
 53	unsigned char	router_alert;
 54	unsigned char	cipso;
 55	unsigned char	__pad2;
 56	unsigned char	__data[];
 57};
 58
 59struct ip_options_rcu {
 60	struct rcu_head rcu;
 61	struct ip_options opt;
 62};
 63
 64struct ip_options_data {
 65	struct ip_options_rcu	opt;
 66	char			data[40];
 67};
 68
 69struct inet_request_sock {
 70	struct request_sock	req;
 71#define ir_loc_addr		req.__req_common.skc_rcv_saddr
 72#define ir_rmt_addr		req.__req_common.skc_daddr
 73#define ir_num			req.__req_common.skc_num
 74#define ir_rmt_port		req.__req_common.skc_dport
 75#define ir_v6_rmt_addr		req.__req_common.skc_v6_daddr
 76#define ir_v6_loc_addr		req.__req_common.skc_v6_rcv_saddr
 77#define ir_iif			req.__req_common.skc_bound_dev_if
 78#define ir_cookie		req.__req_common.skc_cookie
 79#define ireq_net		req.__req_common.skc_net
 80#define ireq_state		req.__req_common.skc_state
 81#define ireq_family		req.__req_common.skc_family
 82
 83	u16			snd_wscale : 4,
 84				rcv_wscale : 4,
 85				tstamp_ok  : 1,
 86				sack_ok	   : 1,
 87				wscale_ok  : 1,
 88				ecn_ok	   : 1,
 89				acked	   : 1,
 90				no_srccheck: 1,
 91				smc_ok	   : 1;
 92	u32                     ir_mark;
 93	union {
 94		struct ip_options_rcu __rcu	*ireq_opt;
 95#if IS_ENABLED(CONFIG_IPV6)
 96		struct {
 97			struct ipv6_txoptions	*ipv6_opt;
 98			struct sk_buff		*pktopts;
 99		};
100#endif
101	};
102};
103
104static inline struct inet_request_sock *inet_rsk(const struct request_sock *sk)
105{
106	return (struct inet_request_sock *)sk;
107}
108
109static inline u32 inet_request_mark(const struct sock *sk, struct sk_buff *skb)
110{
111	u32 mark = READ_ONCE(sk->sk_mark);
112
113	if (!mark && READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_fwmark_accept))
114		return skb->mark;
115
116	return mark;
117}
118
119static inline int inet_request_bound_dev_if(const struct sock *sk,
120					    struct sk_buff *skb)
121{
122	int bound_dev_if = READ_ONCE(sk->sk_bound_dev_if);
123#ifdef CONFIG_NET_L3_MASTER_DEV
124	struct net *net = sock_net(sk);
125
126	if (!bound_dev_if && READ_ONCE(net->ipv4.sysctl_tcp_l3mdev_accept))
127		return l3mdev_master_ifindex_by_index(net, skb->skb_iif);
128#endif
129
130	return bound_dev_if;
131}
132
133static inline int inet_sk_bound_l3mdev(const struct sock *sk)
134{
135#ifdef CONFIG_NET_L3_MASTER_DEV
136	struct net *net = sock_net(sk);
137
138	if (!READ_ONCE(net->ipv4.sysctl_tcp_l3mdev_accept))
139		return l3mdev_master_ifindex_by_index(net,
140						      sk->sk_bound_dev_if);
141#endif
142
143	return 0;
144}
145
146static inline bool inet_bound_dev_eq(bool l3mdev_accept, int bound_dev_if,
147				     int dif, int sdif)
148{
149	if (!bound_dev_if)
150		return !sdif || l3mdev_accept;
151	return bound_dev_if == dif || bound_dev_if == sdif;
152}
153
154static inline bool inet_sk_bound_dev_eq(const struct net *net,
155					int bound_dev_if,
156					int dif, int sdif)
157{
158#if IS_ENABLED(CONFIG_NET_L3_MASTER_DEV)
159	return inet_bound_dev_eq(!!READ_ONCE(net->ipv4.sysctl_tcp_l3mdev_accept),
160				 bound_dev_if, dif, sdif);
161#else
162	return inet_bound_dev_eq(true, bound_dev_if, dif, sdif);
163#endif
164}
165
166struct inet_cork {
167	unsigned int		flags;
168	__be32			addr;
169	struct ip_options	*opt;
170	unsigned int		fragsize;
171	int			length; /* Total length of all frames */
172	struct dst_entry	*dst;
173	u8			tx_flags;
174	__u8			ttl;
175	__s16			tos;
176	u32			priority;
177	__u16			gso_size;
178	u32			ts_opt_id;
179	u64			transmit_time;
180	u32			mark;
181};
182
183struct inet_cork_full {
184	struct inet_cork	base;
185	struct flowi		fl;
186};
187
188struct ip_mc_socklist;
189struct ipv6_pinfo;
190struct rtable;
191
192/** struct inet_sock - representation of INET sockets
193 *
194 * @sk - ancestor class
195 * @pinet6 - pointer to IPv6 control block
196 * @inet_daddr - Foreign IPv4 addr
197 * @inet_rcv_saddr - Bound local IPv4 addr
198 * @inet_dport - Destination port
199 * @inet_num - Local port
200 * @inet_flags - various atomic flags
201 * @inet_saddr - Sending source
202 * @uc_ttl - Unicast TTL
203 * @inet_sport - Source port
204 * @inet_id - ID counter for DF pkts
205 * @tos - TOS
206 * @mc_ttl - Multicasting TTL
207 * @uc_index - Unicast outgoing device index
208 * @mc_index - Multicast device index
209 * @mc_list - Group array
210 * @cork - info to build ip hdr on each ip frag while socket is corked
211 */
212struct inet_sock {
213	/* sk and pinet6 has to be the first two members of inet_sock */
214	struct sock		sk;
215#if IS_ENABLED(CONFIG_IPV6)
216	struct ipv6_pinfo	*pinet6;
217	struct ipv6_fl_socklist __rcu *ipv6_fl_list;
218#endif
219	/* Socket demultiplex comparisons on incoming packets. */
220#define inet_daddr		sk.__sk_common.skc_daddr
221#define inet_rcv_saddr		sk.__sk_common.skc_rcv_saddr
222#define inet_dport		sk.__sk_common.skc_dport
223#define inet_num		sk.__sk_common.skc_num
224
225	unsigned long		inet_flags;
226	__be32			inet_saddr;
227	__s16			uc_ttl;
228	__be16			inet_sport;
229	struct ip_options_rcu __rcu	*inet_opt;
230	atomic_t		inet_id;
231
232	__u8			tos;
233	__u8			min_ttl;
234	__u8			mc_ttl;
235	__u8			pmtudisc;
236	__u8			rcv_tos;
237	__u8			convert_csum;
238	int			uc_index;
239	int			mc_index;
240	__be32			mc_addr;
241	u32			local_port_range;	/* high << 16 | low */
242
243	struct ip_mc_socklist __rcu	*mc_list;
244	struct inet_cork_full	cork;
245};
246
247#define IPCORK_OPT		1	/* ip-options has been held in ipcork.opt */
248#define IPCORK_TS_OPT_ID	2	/* ts_opt_id field is valid, overriding sk_tskey */
249
250enum {
251	INET_FLAGS_PKTINFO	= 0,
252	INET_FLAGS_TTL		= 1,
253	INET_FLAGS_TOS		= 2,
254	INET_FLAGS_RECVOPTS	= 3,
255	INET_FLAGS_RETOPTS	= 4,
256	INET_FLAGS_PASSSEC	= 5,
257	INET_FLAGS_ORIGDSTADDR	= 6,
258	INET_FLAGS_CHECKSUM	= 7,
259	INET_FLAGS_RECVFRAGSIZE	= 8,
260
261	INET_FLAGS_RECVERR	= 9,
262	INET_FLAGS_RECVERR_RFC4884 = 10,
263	INET_FLAGS_FREEBIND	= 11,
264	INET_FLAGS_HDRINCL	= 12,
265	INET_FLAGS_MC_LOOP	= 13,
266	INET_FLAGS_MC_ALL	= 14,
267	INET_FLAGS_TRANSPARENT	= 15,
268	INET_FLAGS_IS_ICSK	= 16,
269	INET_FLAGS_NODEFRAG	= 17,
270	INET_FLAGS_BIND_ADDRESS_NO_PORT = 18,
271	INET_FLAGS_DEFER_CONNECT = 19,
272	INET_FLAGS_MC6_LOOP	= 20,
273	INET_FLAGS_RECVERR6_RFC4884 = 21,
274	INET_FLAGS_MC6_ALL	= 22,
275	INET_FLAGS_AUTOFLOWLABEL_SET = 23,
276	INET_FLAGS_AUTOFLOWLABEL = 24,
277	INET_FLAGS_DONTFRAG	= 25,
278	INET_FLAGS_RECVERR6	= 26,
279	INET_FLAGS_REPFLOW	= 27,
280	INET_FLAGS_RTALERT_ISOLATE = 28,
281	INET_FLAGS_SNDFLOW	= 29,
282	INET_FLAGS_RTALERT	= 30,
283};
284
285/* cmsg flags for inet */
286#define IP_CMSG_PKTINFO		BIT(INET_FLAGS_PKTINFO)
287#define IP_CMSG_TTL		BIT(INET_FLAGS_TTL)
288#define IP_CMSG_TOS		BIT(INET_FLAGS_TOS)
289#define IP_CMSG_RECVOPTS	BIT(INET_FLAGS_RECVOPTS)
290#define IP_CMSG_RETOPTS		BIT(INET_FLAGS_RETOPTS)
291#define IP_CMSG_PASSSEC		BIT(INET_FLAGS_PASSSEC)
292#define IP_CMSG_ORIGDSTADDR	BIT(INET_FLAGS_ORIGDSTADDR)
293#define IP_CMSG_CHECKSUM	BIT(INET_FLAGS_CHECKSUM)
294#define IP_CMSG_RECVFRAGSIZE	BIT(INET_FLAGS_RECVFRAGSIZE)
295
296#define IP_CMSG_ALL	(IP_CMSG_PKTINFO | IP_CMSG_TTL |		\
297			 IP_CMSG_TOS | IP_CMSG_RECVOPTS |		\
298			 IP_CMSG_RETOPTS | IP_CMSG_PASSSEC |		\
299			 IP_CMSG_ORIGDSTADDR | IP_CMSG_CHECKSUM |	\
300			 IP_CMSG_RECVFRAGSIZE)
301
302static inline unsigned long inet_cmsg_flags(const struct inet_sock *inet)
303{
304	return READ_ONCE(inet->inet_flags) & IP_CMSG_ALL;
305}
306
307static inline dscp_t inet_sk_dscp(const struct inet_sock *inet)
308{
309	return inet_dsfield_to_dscp(READ_ONCE(inet->tos));
310}
311
312#define inet_test_bit(nr, sk)			\
313	test_bit(INET_FLAGS_##nr, &inet_sk(sk)->inet_flags)
314#define inet_set_bit(nr, sk)			\
315	set_bit(INET_FLAGS_##nr, &inet_sk(sk)->inet_flags)
316#define inet_clear_bit(nr, sk)			\
317	clear_bit(INET_FLAGS_##nr, &inet_sk(sk)->inet_flags)
318#define inet_assign_bit(nr, sk, val)		\
319	assign_bit(INET_FLAGS_##nr, &inet_sk(sk)->inet_flags, val)
320
321/**
322 * sk_to_full_sk - Access to a full socket
323 * @sk: pointer to a socket
324 *
325 * SYNACK messages might be attached to request sockets.
326 * Some places want to reach the listener in this case.
327 */
328static inline struct sock *sk_to_full_sk(struct sock *sk)
329{
330#ifdef CONFIG_INET
331	if (sk && READ_ONCE(sk->sk_state) == TCP_NEW_SYN_RECV)
332		sk = inet_reqsk(sk)->rsk_listener;
333	if (sk && READ_ONCE(sk->sk_state) == TCP_TIME_WAIT)
334		sk = NULL;
335#endif
336	return sk;
337}
338
339/* sk_to_full_sk() variant with a const argument */
340static inline const struct sock *sk_const_to_full_sk(const struct sock *sk)
341{
342#ifdef CONFIG_INET
343	if (sk && READ_ONCE(sk->sk_state) == TCP_NEW_SYN_RECV)
344		sk = ((const struct request_sock *)sk)->rsk_listener;
345	if (sk && READ_ONCE(sk->sk_state) == TCP_TIME_WAIT)
346		sk = NULL;
347#endif
348	return sk;
349}
350
351static inline struct sock *skb_to_full_sk(const struct sk_buff *skb)
352{
353	return sk_to_full_sk(skb->sk);
354}
355
356#define inet_sk(ptr) container_of_const(ptr, struct inet_sock, sk)
357
358int inet_sk_rebuild_header(struct sock *sk);
359
360/**
361 * inet_sk_state_load - read sk->sk_state for lockless contexts
362 * @sk: socket pointer
363 *
364 * Paired with inet_sk_state_store(). Used in places we don't hold socket lock:
365 * tcp_diag_get_info(), tcp_get_info(), tcp_poll(), get_tcp4_sock() ...
366 */
367static inline int inet_sk_state_load(const struct sock *sk)
368{
369	/* state change might impact lockless readers. */
370	return smp_load_acquire(&sk->sk_state);
371}
372
373/**
374 * inet_sk_state_store - update sk->sk_state
375 * @sk: socket pointer
376 * @newstate: new state
377 *
378 * Paired with inet_sk_state_load(). Should be used in contexts where
379 * state change might impact lockless readers.
380 */
381void inet_sk_state_store(struct sock *sk, int newstate);
382
383void inet_sk_set_state(struct sock *sk, int state);
384
385static inline unsigned int __inet_ehashfn(const __be32 laddr,
386					  const __u16 lport,
387					  const __be32 faddr,
388					  const __be16 fport,
389					  u32 initval)
390{
391	return jhash_3words((__force __u32) laddr,
392			    (__force __u32) faddr,
393			    ((__u32) lport) << 16 | (__force __u32)fport,
394			    initval);
395}
396
397struct request_sock *inet_reqsk_alloc(const struct request_sock_ops *ops,
398				      struct sock *sk_listener,
399				      bool attach_listener);
400
401static inline __u8 inet_sk_flowi_flags(const struct sock *sk)
402{
403	__u8 flags = 0;
404
405	if (inet_test_bit(TRANSPARENT, sk) || inet_test_bit(HDRINCL, sk))
406		flags |= FLOWI_FLAG_ANYSRC;
407	return flags;
408}
409
410static inline void inet_inc_convert_csum(struct sock *sk)
411{
412	inet_sk(sk)->convert_csum++;
413}
414
415static inline void inet_dec_convert_csum(struct sock *sk)
416{
417	if (inet_sk(sk)->convert_csum > 0)
418		inet_sk(sk)->convert_csum--;
419}
420
421static inline bool inet_get_convert_csum(struct sock *sk)
422{
423	return !!inet_sk(sk)->convert_csum;
424}
425
426
427static inline bool inet_can_nonlocal_bind(struct net *net,
428					  struct inet_sock *inet)
429{
430	return READ_ONCE(net->ipv4.sysctl_ip_nonlocal_bind) ||
431		test_bit(INET_FLAGS_FREEBIND, &inet->inet_flags) ||
432		test_bit(INET_FLAGS_TRANSPARENT, &inet->inet_flags);
433}
434
435static inline bool inet_addr_valid_or_nonlocal(struct net *net,
436					       struct inet_sock *inet,
437					       __be32 addr,
438					       int addr_type)
439{
440	return inet_can_nonlocal_bind(net, inet) ||
441		addr == htonl(INADDR_ANY) ||
442		addr_type == RTN_LOCAL ||
443		addr_type == RTN_MULTICAST ||
444		addr_type == RTN_BROADCAST;
445}
446
447#endif	/* _INET_SOCK_H */