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.
6 * Definitions for the TCP module.
8 * Version: @(#)tcp.h 1.0.5 05/23/93
10 * Authors: Ross Biro, <bir7@leland.Stanford.Edu>
11 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
13 * This program is free software; you can redistribute it and/or
14 * modify it under the terms of the GNU General Public License
15 * as published by the Free Software Foundation; either version
16 * 2 of the License, or (at your option) any later version.
22 #define FASTRETRANS_DEBUG 1
24 /* Cancel timers, when they are not required. */
25 #undef TCP_CLEAR_TIMERS
27 #include <linux/config.h>
28 #include <linux/list.h>
29 #include <linux/tcp.h>
30 #include <linux/slab.h>
31 #include <linux/cache.h>
32 #include <linux/percpu.h>
33 #include <net/checksum.h>
37 #if defined(CONFIG_IPV6) || defined (CONFIG_IPV6_MODULE)
38 #include <linux/ipv6.h>
40 #include <linux/seq_file.h>
42 /* This is for all connections with a full identity, no wildcards.
43 * New scheme, half the table is for TIME_WAIT, the other half is
44 * for the rest. I'll experiment with dynamic table growth later.
46 struct tcp_ehash_bucket {
48 struct hlist_head chain;
49 } __attribute__((__aligned__(8)));
51 /* This is for listening sockets, thus all sockets which possess wildcards. */
52 #define TCP_LHTABLE_SIZE 32 /* Yes, really, this is all you need. */
54 /* There are a few simple rules, which allow for local port reuse by
55 * an application. In essence:
57 * 1) Sockets bound to different interfaces may share a local port.
58 * Failing that, goto test 2.
59 * 2) If all sockets have sk->sk_reuse set, and none of them are in
60 * TCP_LISTEN state, the port may be shared.
61 * Failing that, goto test 3.
62 * 3) If all sockets are bound to a specific inet_sk(sk)->rcv_saddr local
63 * address, and none of them are the same, the port may be
65 * Failing this, the port cannot be shared.
67 * The interesting point, is test #2. This is what an FTP server does
68 * all day. To optimize this case we use a specific flag bit defined
69 * below. As we add sockets to a bind bucket list, we perform a
70 * check of: (newsk->sk_reuse && (newsk->sk_state != TCP_LISTEN))
71 * As long as all sockets added to a bind bucket pass this test,
72 * the flag bit will be set.
73 * The resulting situation is that tcp_v[46]_verify_bind() can just check
74 * for this flag bit, if it is set and the socket trying to bind has
75 * sk->sk_reuse set, we don't even have to walk the owners list at all,
76 * we return that it is ok to bind this socket to the requested local port.
78 * Sounds like a lot of work, but it is worth it. In a more naive
79 * implementation (ie. current FreeBSD etc.) the entire list of ports
80 * must be walked for each data port opened by an ftp server. Needless
81 * to say, this does not scale at all. With a couple thousand FTP
82 * users logged onto your box, isn't it nice to know that new data
83 * ports are created in O(1) time? I thought so. ;-) -DaveM
85 struct tcp_bind_bucket {
87 signed short fastreuse;
88 struct hlist_node node;
89 struct hlist_head owners;
92 #define tb_for_each(tb, node, head) hlist_for_each_entry(tb, node, head, node)
94 struct tcp_bind_hashbucket {
96 struct hlist_head chain;
99 static inline struct tcp_bind_bucket *__tb_head(struct tcp_bind_hashbucket *head)
101 return hlist_entry(head->chain.first, struct tcp_bind_bucket, node);
104 static inline struct tcp_bind_bucket *tb_head(struct tcp_bind_hashbucket *head)
106 return hlist_empty(&head->chain) ? NULL : __tb_head(head);
109 extern struct tcp_hashinfo {
110 /* This is for sockets with full identity only. Sockets here will
111 * always be without wildcards and will have the following invariant:
113 * TCP_ESTABLISHED <= sk->sk_state < TCP_CLOSE
115 * First half of the table is for sockets not in TIME_WAIT, second half
116 * is for TIME_WAIT sockets only.
118 struct tcp_ehash_bucket *__tcp_ehash;
120 /* Ok, let's try this, I give up, we do need a local binding
121 * TCP hash as well as the others for fast bind/connect.
123 struct tcp_bind_hashbucket *__tcp_bhash;
125 int __tcp_bhash_size;
126 int __tcp_ehash_size;
128 /* All sockets in TCP_LISTEN state will be in here. This is the only
129 * table where wildcard'd TCP sockets can exist. Hash function here
130 * is just local port number.
132 struct hlist_head __tcp_listening_hash[TCP_LHTABLE_SIZE];
134 /* All the above members are written once at bootup and
135 * never written again _or_ are predominantly read-access.
137 * Now align to a new cache line as all the following members
140 rwlock_t __tcp_lhash_lock ____cacheline_aligned;
141 atomic_t __tcp_lhash_users;
142 wait_queue_head_t __tcp_lhash_wait;
143 spinlock_t __tcp_portalloc_lock;
146 #define tcp_ehash (tcp_hashinfo.__tcp_ehash)
147 #define tcp_bhash (tcp_hashinfo.__tcp_bhash)
148 #define tcp_ehash_size (tcp_hashinfo.__tcp_ehash_size)
149 #define tcp_bhash_size (tcp_hashinfo.__tcp_bhash_size)
150 #define tcp_listening_hash (tcp_hashinfo.__tcp_listening_hash)
151 #define tcp_lhash_lock (tcp_hashinfo.__tcp_lhash_lock)
152 #define tcp_lhash_users (tcp_hashinfo.__tcp_lhash_users)
153 #define tcp_lhash_wait (tcp_hashinfo.__tcp_lhash_wait)
154 #define tcp_portalloc_lock (tcp_hashinfo.__tcp_portalloc_lock)
156 /* SLAB cache for TCP socks */
157 extern kmem_cache_t *tcp_sk_cachep;
159 extern kmem_cache_t *tcp_bucket_cachep;
160 extern struct tcp_bind_bucket *tcp_bucket_create(struct tcp_bind_hashbucket *head,
161 unsigned short snum);
162 extern void tcp_bucket_destroy(struct tcp_bind_bucket *tb);
163 extern void tcp_bucket_unlock(struct sock *sk);
164 extern int tcp_port_rover;
165 extern struct sock *tcp_v4_lookup_listener(u32 addr, unsigned short hnum, int dif);
167 /* These are AF independent. */
168 static __inline__ int tcp_bhashfn(__u16 lport)
170 return (lport & (tcp_bhash_size - 1));
173 extern void tcp_bind_hash(struct sock *sk, struct tcp_bind_bucket *tb,
174 unsigned short snum);
176 #if (BITS_PER_LONG == 64)
177 #define TCP_ADDRCMP_ALIGN_BYTES 8
179 #define TCP_ADDRCMP_ALIGN_BYTES 4
182 /* This is a TIME_WAIT bucket. It works around the memory consumption
183 * problems of sockets in such a state on heavily loaded servers, but
184 * without violating the protocol specification.
186 struct tcp_tw_bucket {
188 * Now struct sock also uses sock_common, so please just
189 * don't add nothing before this first member (__tw_common) --acme
191 struct sock_common __tw_common;
192 #define tw_family __tw_common.skc_family
193 #define tw_state __tw_common.skc_state
194 #define tw_reuse __tw_common.skc_reuse
195 #define tw_bound_dev_if __tw_common.skc_bound_dev_if
196 #define tw_node __tw_common.skc_node
197 #define tw_bind_node __tw_common.skc_bind_node
198 #define tw_refcnt __tw_common.skc_refcnt
199 volatile unsigned char tw_substate;
200 unsigned char tw_rcv_wscale;
202 /* Socket demultiplex comparisons on incoming packets. */
203 /* these five are in inet_opt */
205 __attribute__((aligned(TCP_ADDRCMP_ALIGN_BYTES)));
209 /* And these are ours. */
216 long tw_ts_recent_stamp;
217 unsigned long tw_ttd;
218 struct tcp_bind_bucket *tw_tb;
219 struct hlist_node tw_death_node;
220 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
221 struct in6_addr tw_v6_daddr;
222 struct in6_addr tw_v6_rcv_saddr;
227 static __inline__ void tw_add_node(struct tcp_tw_bucket *tw,
228 struct hlist_head *list)
230 hlist_add_head(&tw->tw_node, list);
233 static __inline__ void tw_add_bind_node(struct tcp_tw_bucket *tw,
234 struct hlist_head *list)
236 hlist_add_head(&tw->tw_bind_node, list);
239 static inline int tw_dead_hashed(struct tcp_tw_bucket *tw)
241 return tw->tw_death_node.pprev != NULL;
244 static __inline__ void tw_dead_node_init(struct tcp_tw_bucket *tw)
246 tw->tw_death_node.pprev = NULL;
249 static __inline__ void __tw_del_dead_node(struct tcp_tw_bucket *tw)
251 __hlist_del(&tw->tw_death_node);
252 tw_dead_node_init(tw);
255 static __inline__ int tw_del_dead_node(struct tcp_tw_bucket *tw)
257 if (tw_dead_hashed(tw)) {
258 __tw_del_dead_node(tw);
264 #define tw_for_each(tw, node, head) \
265 hlist_for_each_entry(tw, node, head, tw_node)
267 #define tw_for_each_inmate(tw, node, jail) \
268 hlist_for_each_entry(tw, node, jail, tw_death_node)
270 #define tw_for_each_inmate_safe(tw, node, safe, jail) \
271 hlist_for_each_entry_safe(tw, node, safe, jail, tw_death_node)
273 #define tcptw_sk(__sk) ((struct tcp_tw_bucket *)(__sk))
275 static inline u32 tcp_v4_rcv_saddr(const struct sock *sk)
277 return likely(sk->sk_state != TCP_TIME_WAIT) ?
278 inet_sk(sk)->rcv_saddr : tcptw_sk(sk)->tw_rcv_saddr;
281 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
282 static inline struct in6_addr *__tcp_v6_rcv_saddr(const struct sock *sk)
284 return likely(sk->sk_state != TCP_TIME_WAIT) ?
285 &inet6_sk(sk)->rcv_saddr : &tcptw_sk(sk)->tw_v6_rcv_saddr;
288 static inline struct in6_addr *tcp_v6_rcv_saddr(const struct sock *sk)
290 return sk->sk_family == AF_INET6 ? __tcp_v6_rcv_saddr(sk) : NULL;
293 #define tcptw_sk_ipv6only(__sk) (tcptw_sk(__sk)->tw_v6_ipv6only)
295 static inline int tcp_v6_ipv6only(const struct sock *sk)
297 return likely(sk->sk_state != TCP_TIME_WAIT) ?
298 ipv6_only_sock(sk) : tcptw_sk_ipv6only(sk);
301 # define __tcp_v6_rcv_saddr(__sk) NULL
302 # define tcp_v6_rcv_saddr(__sk) NULL
303 # define tcptw_sk_ipv6only(__sk) 0
304 # define tcp_v6_ipv6only(__sk) 0
307 extern kmem_cache_t *tcp_timewait_cachep;
309 static inline void tcp_tw_put(struct tcp_tw_bucket *tw)
311 if (atomic_dec_and_test(&tw->tw_refcnt)) {
312 #ifdef INET_REFCNT_DEBUG
313 printk(KERN_DEBUG "tw_bucket %p released\n", tw);
315 kmem_cache_free(tcp_timewait_cachep, tw);
319 extern atomic_t tcp_orphan_count;
320 extern int tcp_tw_count;
321 extern void tcp_time_wait(struct sock *sk, int state, int timeo);
322 extern void tcp_tw_schedule(struct tcp_tw_bucket *tw, int timeo);
323 extern void tcp_tw_deschedule(struct tcp_tw_bucket *tw);
326 /* Socket demux engine toys. */
328 #define TCP_COMBINED_PORTS(__sport, __dport) \
329 (((__u32)(__sport)<<16) | (__u32)(__dport))
330 #else /* __LITTLE_ENDIAN */
331 #define TCP_COMBINED_PORTS(__sport, __dport) \
332 (((__u32)(__dport)<<16) | (__u32)(__sport))
335 #if (BITS_PER_LONG == 64)
337 #define TCP_V4_ADDR_COOKIE(__name, __saddr, __daddr) \
338 __u64 __name = (((__u64)(__saddr))<<32)|((__u64)(__daddr));
339 #else /* __LITTLE_ENDIAN */
340 #define TCP_V4_ADDR_COOKIE(__name, __saddr, __daddr) \
341 __u64 __name = (((__u64)(__daddr))<<32)|((__u64)(__saddr));
342 #endif /* __BIG_ENDIAN */
343 #define TCP_IPV4_MATCH(__sk, __cookie, __saddr, __daddr, __ports, __dif)\
344 (((*((__u64 *)&(inet_sk(__sk)->daddr)))== (__cookie)) && \
345 ((*((__u32 *)&(inet_sk(__sk)->dport)))== (__ports)) && \
346 (!((__sk)->sk_bound_dev_if) || ((__sk)->sk_bound_dev_if == (__dif))))
347 #define TCP_IPV4_TW_MATCH(__sk, __cookie, __saddr, __daddr, __ports, __dif)\
348 (((*((__u64 *)&(tcptw_sk(__sk)->tw_daddr))) == (__cookie)) && \
349 ((*((__u32 *)&(tcptw_sk(__sk)->tw_dport))) == (__ports)) && \
350 (!((__sk)->sk_bound_dev_if) || ((__sk)->sk_bound_dev_if == (__dif))))
351 #else /* 32-bit arch */
352 #define TCP_V4_ADDR_COOKIE(__name, __saddr, __daddr)
353 #define TCP_IPV4_MATCH(__sk, __cookie, __saddr, __daddr, __ports, __dif)\
354 ((inet_sk(__sk)->daddr == (__saddr)) && \
355 (inet_sk(__sk)->rcv_saddr == (__daddr)) && \
356 ((*((__u32 *)&(inet_sk(__sk)->dport)))== (__ports)) && \
357 (!((__sk)->sk_bound_dev_if) || ((__sk)->sk_bound_dev_if == (__dif))))
358 #define TCP_IPV4_TW_MATCH(__sk, __cookie, __saddr, __daddr, __ports, __dif)\
359 ((tcptw_sk(__sk)->tw_daddr == (__saddr)) && \
360 (tcptw_sk(__sk)->tw_rcv_saddr == (__daddr)) && \
361 ((*((__u32 *)&(tcptw_sk(__sk)->tw_dport))) == (__ports)) && \
362 (!((__sk)->sk_bound_dev_if) || ((__sk)->sk_bound_dev_if == (__dif))))
363 #endif /* 64-bit arch */
365 #define TCP_IPV6_MATCH(__sk, __saddr, __daddr, __ports, __dif) \
366 (((*((__u32 *)&(inet_sk(__sk)->dport)))== (__ports)) && \
367 ((__sk)->sk_family == AF_INET6) && \
368 !ipv6_addr_cmp(&inet6_sk(__sk)->daddr, (__saddr)) && \
369 !ipv6_addr_cmp(&inet6_sk(__sk)->rcv_saddr, (__daddr)) && \
370 (!((__sk)->sk_bound_dev_if) || ((__sk)->sk_bound_dev_if == (__dif))))
372 /* These can have wildcards, don't try too hard. */
373 static __inline__ int tcp_lhashfn(unsigned short num)
375 return num & (TCP_LHTABLE_SIZE - 1);
378 static __inline__ int tcp_sk_listen_hashfn(struct sock *sk)
380 return tcp_lhashfn(inet_sk(sk)->num);
383 #define MAX_TCP_HEADER (128 + MAX_HEADER)
386 * Never offer a window over 32767 without using window scaling. Some
387 * poor stacks do signed 16bit maths!
389 #define MAX_TCP_WINDOW 32767U
391 /* Minimal accepted MSS. It is (60+60+8) - (20+20). */
392 #define TCP_MIN_MSS 88U
394 /* Minimal RCV_MSS. */
395 #define TCP_MIN_RCVMSS 536U
397 /* After receiving this amount of duplicate ACKs fast retransmit starts. */
398 #define TCP_FASTRETRANS_THRESH 3
400 /* Maximal reordering. */
401 #define TCP_MAX_REORDERING 127
403 /* Maximal number of ACKs sent quickly to accelerate slow-start. */
404 #define TCP_MAX_QUICKACKS 16U
406 /* urg_data states */
407 #define TCP_URG_VALID 0x0100
408 #define TCP_URG_NOTYET 0x0200
409 #define TCP_URG_READ 0x0400
411 #define TCP_RETR1 3 /*
412 * This is how many retries it does before it
413 * tries to figure out if the gateway is
414 * down. Minimal RFC value is 3; it corresponds
415 * to ~3sec-8min depending on RTO.
418 #define TCP_RETR2 15 /*
419 * This should take at least
420 * 90 minutes to time out.
421 * RFC1122 says that the limit is 100 sec.
422 * 15 is ~13-30min depending on RTO.
425 #define TCP_SYN_RETRIES 5 /* number of times to retry active opening a
426 * connection: ~180sec is RFC minumum */
428 #define TCP_SYNACK_RETRIES 5 /* number of times to retry passive opening a
429 * connection: ~180sec is RFC minumum */
432 #define TCP_ORPHAN_RETRIES 7 /* number of times to retry on an orphaned
433 * socket. 7 is ~50sec-16min.
437 #define TCP_TIMEWAIT_LEN (60*HZ) /* how long to wait to destroy TIME-WAIT
438 * state, about 60 seconds */
439 #define TCP_FIN_TIMEOUT TCP_TIMEWAIT_LEN
440 /* BSD style FIN_WAIT2 deadlock breaker.
441 * It used to be 3min, new value is 60sec,
442 * to combine FIN-WAIT-2 timeout with
446 #define TCP_DELACK_MAX ((unsigned)(HZ/5)) /* maximal time to delay before sending an ACK */
448 #define TCP_DELACK_MIN ((unsigned)(HZ/25)) /* minimal time to delay before sending an ACK */
449 #define TCP_ATO_MIN ((unsigned)(HZ/25))
451 #define TCP_DELACK_MIN 4U
452 #define TCP_ATO_MIN 4U
454 #define TCP_RTO_MAX ((unsigned)(120*HZ))
455 #define TCP_RTO_MIN ((unsigned)(HZ/5))
456 #define TCP_TIMEOUT_INIT ((unsigned)(3*HZ)) /* RFC 1122 initial RTO value */
458 #define TCP_RESOURCE_PROBE_INTERVAL ((unsigned)(HZ/2U)) /* Maximal interval between probes
459 * for local resources.
462 #define TCP_KEEPALIVE_TIME (120*60*HZ) /* two hours */
463 #define TCP_KEEPALIVE_PROBES 9 /* Max of 9 keepalive probes */
464 #define TCP_KEEPALIVE_INTVL (75*HZ)
466 #define MAX_TCP_KEEPIDLE 32767
467 #define MAX_TCP_KEEPINTVL 32767
468 #define MAX_TCP_KEEPCNT 127
469 #define MAX_TCP_SYNCNT 127
471 #define TCP_SYNQ_INTERVAL (HZ/5) /* Period of SYNACK timer */
472 #define TCP_SYNQ_HSIZE 512 /* Size of SYNACK hash table */
474 #define TCP_PAWS_24DAYS (60 * 60 * 24 * 24)
475 #define TCP_PAWS_MSL 60 /* Per-host timestamps are invalidated
476 * after this time. It should be equal
477 * (or greater than) TCP_TIMEWAIT_LEN
478 * to provide reliability equal to one
479 * provided by timewait state.
481 #define TCP_PAWS_WINDOW 1 /* Replay window for per-host
482 * timestamps. It must be less than
483 * minimal timewait lifetime.
486 #define TCP_TW_RECYCLE_SLOTS_LOG 5
487 #define TCP_TW_RECYCLE_SLOTS (1<<TCP_TW_RECYCLE_SLOTS_LOG)
489 /* If time > 4sec, it is "slow" path, no recycling is required,
490 so that we select tick to get range about 4 seconds.
493 #if HZ <= 16 || HZ > 4096
494 # error Unsupported: HZ <= 16 or HZ > 4096
496 # define TCP_TW_RECYCLE_TICK (5+2-TCP_TW_RECYCLE_SLOTS_LOG)
498 # define TCP_TW_RECYCLE_TICK (6+2-TCP_TW_RECYCLE_SLOTS_LOG)
500 # define TCP_TW_RECYCLE_TICK (7+2-TCP_TW_RECYCLE_SLOTS_LOG)
502 # define TCP_TW_RECYCLE_TICK (8+2-TCP_TW_RECYCLE_SLOTS_LOG)
504 # define TCP_TW_RECYCLE_TICK (9+2-TCP_TW_RECYCLE_SLOTS_LOG)
506 # define TCP_TW_RECYCLE_TICK (10+2-TCP_TW_RECYCLE_SLOTS_LOG)
508 # define TCP_TW_RECYCLE_TICK (11+2-TCP_TW_RECYCLE_SLOTS_LOG)
510 # define TCP_TW_RECYCLE_TICK (12+2-TCP_TW_RECYCLE_SLOTS_LOG)
513 #define BICTCP_1_OVER_BETA 8 /*
515 * multiplicative decrease factor
517 #define BICTCP_MAX_INCREMENT 32 /*
518 * Limit on the amount of
519 * increment allowed during
522 #define BICTCP_FUNC_OF_MIN_INCR 11 /*
523 * log(B/Smin)/log(B/(B-1))+1,
527 #define BICTCP_B 4 /*
529 * go to point (max+min)/N
536 #define TCPOPT_NOP 1 /* Padding */
537 #define TCPOPT_EOL 0 /* End of options */
538 #define TCPOPT_MSS 2 /* Segment size negotiating */
539 #define TCPOPT_WINDOW 3 /* Window scaling */
540 #define TCPOPT_SACK_PERM 4 /* SACK Permitted */
541 #define TCPOPT_SACK 5 /* SACK Block */
542 #define TCPOPT_TIMESTAMP 8 /* Better RTT estimations/PAWS */
548 #define TCPOLEN_MSS 4
549 #define TCPOLEN_WINDOW 3
550 #define TCPOLEN_SACK_PERM 2
551 #define TCPOLEN_TIMESTAMP 10
553 /* But this is what stacks really send out. */
554 #define TCPOLEN_TSTAMP_ALIGNED 12
555 #define TCPOLEN_WSCALE_ALIGNED 4
556 #define TCPOLEN_SACKPERM_ALIGNED 4
557 #define TCPOLEN_SACK_BASE 2
558 #define TCPOLEN_SACK_BASE_ALIGNED 4
559 #define TCPOLEN_SACK_PERBLOCK 8
561 #define TCP_TIME_RETRANS 1 /* Retransmit timer */
562 #define TCP_TIME_DACK 2 /* Delayed ack timer */
563 #define TCP_TIME_PROBE0 3 /* Zero window probe timer */
564 #define TCP_TIME_KEEPOPEN 4 /* Keepalive timer */
566 /* Flags in tp->nonagle */
567 #define TCP_NAGLE_OFF 1 /* Nagle's algo is disabled */
568 #define TCP_NAGLE_CORK 2 /* Socket is corked */
569 #define TCP_NAGLE_PUSH 4 /* Cork is overriden for already queued data */
571 /* sysctl variables for tcp */
572 extern int sysctl_max_syn_backlog;
573 extern int sysctl_tcp_timestamps;
574 extern int sysctl_tcp_window_scaling;
575 extern int sysctl_tcp_sack;
576 extern int sysctl_tcp_fin_timeout;
577 extern int sysctl_tcp_tw_recycle;
578 extern int sysctl_tcp_keepalive_time;
579 extern int sysctl_tcp_keepalive_probes;
580 extern int sysctl_tcp_keepalive_intvl;
581 extern int sysctl_tcp_syn_retries;
582 extern int sysctl_tcp_synack_retries;
583 extern int sysctl_tcp_retries1;
584 extern int sysctl_tcp_retries2;
585 extern int sysctl_tcp_orphan_retries;
586 extern int sysctl_tcp_syncookies;
587 extern int sysctl_tcp_retrans_collapse;
588 extern int sysctl_tcp_stdurg;
589 extern int sysctl_tcp_rfc1337;
590 extern int sysctl_tcp_abort_on_overflow;
591 extern int sysctl_tcp_max_orphans;
592 extern int sysctl_tcp_max_tw_buckets;
593 extern int sysctl_tcp_fack;
594 extern int sysctl_tcp_reordering;
595 extern int sysctl_tcp_ecn;
596 extern int sysctl_tcp_dsack;
597 extern int sysctl_tcp_mem[3];
598 extern int sysctl_tcp_wmem[3];
599 extern int sysctl_tcp_rmem[3];
600 extern int sysctl_tcp_app_win;
601 extern int sysctl_tcp_adv_win_scale;
602 extern int sysctl_tcp_tw_reuse;
603 extern int sysctl_tcp_frto;
604 extern int sysctl_tcp_low_latency;
605 extern int sysctl_tcp_westwood;
606 extern int sysctl_tcp_vegas_cong_avoid;
607 extern int sysctl_tcp_vegas_alpha;
608 extern int sysctl_tcp_vegas_beta;
609 extern int sysctl_tcp_vegas_gamma;
610 extern int sysctl_tcp_nometrics_save;
611 extern int sysctl_tcp_bic;
612 extern int sysctl_tcp_bic_fast_convergence;
613 extern int sysctl_tcp_bic_low_window;
614 extern int sysctl_tcp_default_win_scale;
615 extern int sysctl_tcp_moderate_rcvbuf;
617 extern atomic_t tcp_memory_allocated;
618 extern atomic_t tcp_sockets_allocated;
619 extern int tcp_memory_pressure;
623 struct or_calltable {
625 int (*rtx_syn_ack) (struct sock *sk, struct open_request *req, struct dst_entry*);
626 void (*send_ack) (struct sk_buff *skb, struct open_request *req);
627 void (*destructor) (struct open_request *req);
628 void (*send_reset) (struct sk_buff *skb);
631 struct tcp_v4_open_req {
634 struct ip_options *opt;
637 #if defined(CONFIG_IPV6) || defined (CONFIG_IPV6_MODULE)
638 struct tcp_v6_open_req {
639 struct in6_addr loc_addr;
640 struct in6_addr rmt_addr;
641 struct sk_buff *pktopts;
646 /* this structure is too big */
647 struct open_request {
648 struct open_request *dl_next; /* Must be first member! */
655 __u16 snd_wscale : 4,
662 /* The following two fields can be easily recomputed I think -AK */
663 __u32 window_clamp; /* window clamp at creation time */
664 __u32 rcv_wnd; /* rcv_wnd offered first time */
666 unsigned long expires;
667 struct or_calltable *class;
670 struct tcp_v4_open_req v4_req;
671 #if defined(CONFIG_IPV6) || defined (CONFIG_IPV6_MODULE)
672 struct tcp_v6_open_req v6_req;
677 /* SLAB cache for open requests. */
678 extern kmem_cache_t *tcp_openreq_cachep;
680 #define tcp_openreq_alloc() kmem_cache_alloc(tcp_openreq_cachep, SLAB_ATOMIC)
681 #define tcp_openreq_fastfree(req) kmem_cache_free(tcp_openreq_cachep, req)
683 static inline void tcp_openreq_free(struct open_request *req)
685 req->class->destructor(req);
686 tcp_openreq_fastfree(req);
689 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
690 #define TCP_INET_FAMILY(fam) ((fam) == AF_INET)
692 #define TCP_INET_FAMILY(fam) 1
696 * Pointers to address related TCP functions
697 * (i.e. things that depend on the address family)
701 int (*queue_xmit) (struct sk_buff *skb,
704 void (*send_check) (struct sock *sk,
707 struct sk_buff *skb);
709 int (*rebuild_header) (struct sock *sk);
711 int (*conn_request) (struct sock *sk,
712 struct sk_buff *skb);
714 struct sock * (*syn_recv_sock) (struct sock *sk,
716 struct open_request *req,
717 struct dst_entry *dst);
719 int (*remember_stamp) (struct sock *sk);
721 __u16 net_header_len;
723 int (*setsockopt) (struct sock *sk,
729 int (*getsockopt) (struct sock *sk,
736 void (*addr2sockaddr) (struct sock *sk,
743 * The next routines deal with comparing 32 bit unsigned ints
744 * and worry about wraparound (automatic with unsigned arithmetic).
747 static inline int before(__u32 seq1, __u32 seq2)
749 return (__s32)(seq1-seq2) < 0;
752 static inline int after(__u32 seq1, __u32 seq2)
754 return (__s32)(seq2-seq1) < 0;
758 /* is s2<=s1<=s3 ? */
759 static inline int between(__u32 seq1, __u32 seq2, __u32 seq3)
761 return seq3 - seq2 >= seq1 - seq2;
765 extern struct proto tcp_prot;
767 DECLARE_SNMP_STAT(struct tcp_mib, tcp_statistics);
768 #define TCP_INC_STATS(field) SNMP_INC_STATS(tcp_statistics, field)
769 #define TCP_INC_STATS_BH(field) SNMP_INC_STATS_BH(tcp_statistics, field)
770 #define TCP_INC_STATS_USER(field) SNMP_INC_STATS_USER(tcp_statistics, field)
771 #define TCP_DEC_STATS(field) SNMP_DEC_STATS(tcp_statistics, field)
772 #define TCP_ADD_STATS_BH(field, val) SNMP_ADD_STATS_BH(tcp_statistics, field, val)
773 #define TCP_ADD_STATS_USER(field, val) SNMP_ADD_STATS_USER(tcp_statistics, field, val)
775 extern void tcp_put_port(struct sock *sk);
776 extern void tcp_inherit_port(struct sock *sk, struct sock *child);
778 extern void tcp_v4_err(struct sk_buff *skb, u32);
780 extern void tcp_shutdown (struct sock *sk, int how);
782 extern int tcp_v4_rcv(struct sk_buff *skb);
784 extern int tcp_v4_remember_stamp(struct sock *sk);
786 extern int tcp_v4_tw_remember_stamp(struct tcp_tw_bucket *tw);
788 extern int tcp_sendmsg(struct kiocb *iocb, struct sock *sk,
789 struct msghdr *msg, size_t size);
790 extern ssize_t tcp_sendpage(struct socket *sock, struct page *page, int offset, size_t size, int flags);
792 extern int tcp_ioctl(struct sock *sk,
796 extern int tcp_rcv_state_process(struct sock *sk,
801 extern int tcp_rcv_established(struct sock *sk,
806 extern void tcp_rcv_space_adjust(struct sock *sk);
815 static inline void tcp_schedule_ack(struct tcp_opt *tp)
817 tp->ack.pending |= TCP_ACK_SCHED;
820 static inline int tcp_ack_scheduled(struct tcp_opt *tp)
822 return tp->ack.pending&TCP_ACK_SCHED;
825 static __inline__ void tcp_dec_quickack_mode(struct tcp_opt *tp)
827 if (tp->ack.quick && --tp->ack.quick == 0) {
828 /* Leaving quickack mode we deflate ATO. */
829 tp->ack.ato = TCP_ATO_MIN;
833 extern void tcp_enter_quickack_mode(struct tcp_opt *tp);
835 static __inline__ void tcp_delack_init(struct tcp_opt *tp)
837 memset(&tp->ack, 0, sizeof(tp->ack));
840 static inline void tcp_clear_options(struct tcp_opt *tp)
842 tp->tstamp_ok = tp->sack_ok = tp->wscale_ok = tp->snd_wscale = 0;
854 extern enum tcp_tw_status tcp_timewait_state_process(struct tcp_tw_bucket *tw,
859 extern struct sock * tcp_check_req(struct sock *sk,struct sk_buff *skb,
860 struct open_request *req,
861 struct open_request **prev);
862 extern int tcp_child_process(struct sock *parent,
864 struct sk_buff *skb);
865 extern void tcp_enter_frto(struct sock *sk);
866 extern void tcp_enter_loss(struct sock *sk, int how);
867 extern void tcp_clear_retrans(struct tcp_opt *tp);
868 extern void tcp_update_metrics(struct sock *sk);
870 extern void tcp_close(struct sock *sk,
872 extern struct sock * tcp_accept(struct sock *sk, int flags, int *err);
873 extern unsigned int tcp_poll(struct file * file, struct socket *sock, struct poll_table_struct *wait);
875 extern int tcp_getsockopt(struct sock *sk, int level,
879 extern int tcp_setsockopt(struct sock *sk, int level,
880 int optname, char __user *optval,
882 extern void tcp_set_keepalive(struct sock *sk, int val);
883 extern int tcp_recvmsg(struct kiocb *iocb, struct sock *sk,
885 size_t len, int nonblock,
886 int flags, int *addr_len);
888 extern int tcp_listen_start(struct sock *sk);
890 extern void tcp_parse_options(struct sk_buff *skb,
895 * TCP v4 functions exported for the inet6 API
898 extern int tcp_v4_rebuild_header(struct sock *sk);
900 extern int tcp_v4_build_header(struct sock *sk,
901 struct sk_buff *skb);
903 extern void tcp_v4_send_check(struct sock *sk,
904 struct tcphdr *th, int len,
905 struct sk_buff *skb);
907 extern int tcp_v4_conn_request(struct sock *sk,
908 struct sk_buff *skb);
910 extern struct sock * tcp_create_openreq_child(struct sock *sk,
911 struct open_request *req,
912 struct sk_buff *skb);
914 extern struct sock * tcp_v4_syn_recv_sock(struct sock *sk,
916 struct open_request *req,
917 struct dst_entry *dst);
919 extern int tcp_v4_do_rcv(struct sock *sk,
920 struct sk_buff *skb);
922 extern int tcp_v4_connect(struct sock *sk,
923 struct sockaddr *uaddr,
926 extern int tcp_connect(struct sock *sk);
928 extern struct sk_buff * tcp_make_synack(struct sock *sk,
929 struct dst_entry *dst,
930 struct open_request *req);
932 extern int tcp_disconnect(struct sock *sk, int flags);
934 extern void tcp_unhash(struct sock *sk);
936 extern int tcp_v4_hash_connecting(struct sock *sk);
939 /* From syncookies.c */
940 extern struct sock *cookie_v4_check(struct sock *sk, struct sk_buff *skb,
941 struct ip_options *opt);
942 extern __u32 cookie_v4_init_sequence(struct sock *sk, struct sk_buff *skb,
947 extern int tcp_write_xmit(struct sock *, int nonagle);
948 extern int tcp_retransmit_skb(struct sock *, struct sk_buff *);
949 extern void tcp_xmit_retransmit_queue(struct sock *);
950 extern void tcp_simple_retransmit(struct sock *);
952 extern void tcp_send_probe0(struct sock *);
953 extern void tcp_send_partial(struct sock *);
954 extern int tcp_write_wakeup(struct sock *);
955 extern void tcp_send_fin(struct sock *sk);
956 extern void tcp_send_active_reset(struct sock *sk, int priority);
957 extern int tcp_send_synack(struct sock *);
958 extern int tcp_transmit_skb(struct sock *, struct sk_buff *);
959 extern void tcp_push_one(struct sock *, unsigned mss_now);
960 extern void tcp_send_ack(struct sock *sk);
961 extern void tcp_send_delayed_ack(struct sock *sk);
964 extern void tcp_init_xmit_timers(struct sock *);
965 extern void tcp_clear_xmit_timers(struct sock *);
967 extern void tcp_delete_keepalive_timer (struct sock *);
968 extern void tcp_reset_keepalive_timer (struct sock *, unsigned long);
969 extern int tcp_sync_mss(struct sock *sk, u32 pmtu);
971 extern const char timer_bug_msg[];
974 extern void tcp_get_info(struct sock *, struct tcp_info *);
976 /* Read 'sendfile()'-style from a TCP socket */
977 typedef int (*sk_read_actor_t)(read_descriptor_t *, struct sk_buff *,
978 unsigned int, size_t);
979 extern int tcp_read_sock(struct sock *sk, read_descriptor_t *desc,
980 sk_read_actor_t recv_actor);
982 static inline void tcp_clear_xmit_timer(struct sock *sk, int what)
984 struct tcp_opt *tp = tcp_sk(sk);
987 case TCP_TIME_RETRANS:
988 case TCP_TIME_PROBE0:
991 #ifdef TCP_CLEAR_TIMERS
992 sk_stop_timer(sk, &tp->retransmit_timer);
999 #ifdef TCP_CLEAR_TIMERS
1000 sk_stop_timer(sk, &tp->delack_timer);
1004 printk(timer_bug_msg);
1011 * Reset the retransmission timer
1013 static inline void tcp_reset_xmit_timer(struct sock *sk, int what, unsigned long when)
1015 struct tcp_opt *tp = tcp_sk(sk);
1017 if (when > TCP_RTO_MAX) {
1019 printk(KERN_DEBUG "reset_xmit_timer sk=%p %d when=0x%lx, caller=%p\n", sk, what, when, current_text_addr());
1025 case TCP_TIME_RETRANS:
1026 case TCP_TIME_PROBE0:
1028 tp->timeout = jiffies+when;
1029 sk_reset_timer(sk, &tp->retransmit_timer, tp->timeout);
1033 tp->ack.pending |= TCP_ACK_TIMER;
1034 tp->ack.timeout = jiffies+when;
1035 sk_reset_timer(sk, &tp->delack_timer, tp->ack.timeout);
1039 printk(timer_bug_msg);
1043 /* Compute the current effective MSS, taking SACKs and IP options,
1044 * and even PMTU discovery events into account.
1046 * LARGESEND note: !urg_mode is overkill, only frames up to snd_up
1047 * cannot be large. However, taking into account rare use of URG, this
1048 * is not a big flaw.
1051 static __inline__ unsigned int tcp_current_mss(struct sock *sk, int large)
1053 struct tcp_opt *tp = tcp_sk(sk);
1054 struct dst_entry *dst = __sk_dst_get(sk);
1055 int mss_now = large && (sk->sk_route_caps & NETIF_F_TSO) &&
1057 tp->mss_cache : tp->mss_cache_std;
1060 u32 mtu = dst_pmtu(dst);
1061 if (mtu != tp->pmtu_cookie ||
1062 tp->ext2_header_len != dst->header_len)
1063 mss_now = tcp_sync_mss(sk, mtu);
1066 mss_now -= (TCPOLEN_SACK_BASE_ALIGNED +
1067 (tp->eff_sacks * TCPOLEN_SACK_PERBLOCK));
1071 /* Initialize RCV_MSS value.
1072 * RCV_MSS is an our guess about MSS used by the peer.
1073 * We haven't any direct information about the MSS.
1074 * It's better to underestimate the RCV_MSS rather than overestimate.
1075 * Overestimations make us ACKing less frequently than needed.
1076 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
1079 static inline void tcp_initialize_rcv_mss(struct sock *sk)
1081 struct tcp_opt *tp = tcp_sk(sk);
1082 unsigned int hint = min(tp->advmss, tp->mss_cache_std);
1084 hint = min(hint, tp->rcv_wnd/2);
1085 hint = min(hint, TCP_MIN_RCVMSS);
1086 hint = max(hint, TCP_MIN_MSS);
1088 tp->ack.rcv_mss = hint;
1091 static __inline__ void __tcp_fast_path_on(struct tcp_opt *tp, u32 snd_wnd)
1093 tp->pred_flags = htonl((tp->tcp_header_len << 26) |
1094 ntohl(TCP_FLAG_ACK) |
1098 static __inline__ void tcp_fast_path_on(struct tcp_opt *tp)
1100 __tcp_fast_path_on(tp, tp->snd_wnd>>tp->snd_wscale);
1103 static inline void tcp_fast_path_check(struct sock *sk, struct tcp_opt *tp)
1105 if (skb_queue_len(&tp->out_of_order_queue) == 0 &&
1107 atomic_read(&sk->sk_rmem_alloc) < sk->sk_rcvbuf &&
1109 tcp_fast_path_on(tp);
1112 /* Compute the actual receive window we are currently advertising.
1113 * Rcv_nxt can be after the window if our peer push more data
1114 * than the offered window.
1116 static __inline__ u32 tcp_receive_window(struct tcp_opt *tp)
1118 s32 win = tp->rcv_wup + tp->rcv_wnd - tp->rcv_nxt;
1125 /* Choose a new window, without checks for shrinking, and without
1126 * scaling applied to the result. The caller does these things
1127 * if necessary. This is a "raw" window selection.
1129 extern u32 __tcp_select_window(struct sock *sk);
1131 /* TCP timestamps are only 32-bits, this causes a slight
1132 * complication on 64-bit systems since we store a snapshot
1133 * of jiffies in the buffer control blocks below. We decidely
1134 * only use of the low 32-bits of jiffies and hide the ugly
1135 * casts with the following macro.
1137 #define tcp_time_stamp ((__u32)(jiffies))
1139 /* This is what the send packet queueing engine uses to pass
1140 * TCP per-packet control information to the transmission
1141 * code. We also store the host-order sequence numbers in
1142 * here too. This is 36 bytes on 32-bit architectures,
1143 * 40 bytes on 64-bit machines, if this grows please adjust
1144 * skbuff.h:skbuff->cb[xxx] size appropriately.
1148 struct inet_skb_parm h4;
1149 #if defined(CONFIG_IPV6) || defined (CONFIG_IPV6_MODULE)
1150 struct inet6_skb_parm h6;
1152 } header; /* For incoming frames */
1153 __u32 seq; /* Starting sequence number */
1154 __u32 end_seq; /* SEQ + FIN + SYN + datalen */
1155 __u32 when; /* used to compute rtt's */
1156 __u8 flags; /* TCP header flags. */
1158 /* NOTE: These must match up to the flags byte in a
1161 #define TCPCB_FLAG_FIN 0x01
1162 #define TCPCB_FLAG_SYN 0x02
1163 #define TCPCB_FLAG_RST 0x04
1164 #define TCPCB_FLAG_PSH 0x08
1165 #define TCPCB_FLAG_ACK 0x10
1166 #define TCPCB_FLAG_URG 0x20
1167 #define TCPCB_FLAG_ECE 0x40
1168 #define TCPCB_FLAG_CWR 0x80
1170 __u8 sacked; /* State flags for SACK/FACK. */
1171 #define TCPCB_SACKED_ACKED 0x01 /* SKB ACK'd by a SACK block */
1172 #define TCPCB_SACKED_RETRANS 0x02 /* SKB retransmitted */
1173 #define TCPCB_LOST 0x04 /* SKB is lost */
1174 #define TCPCB_TAGBITS 0x07 /* All tag bits */
1176 #define TCPCB_EVER_RETRANS 0x80 /* Ever retransmitted frame */
1177 #define TCPCB_RETRANS (TCPCB_SACKED_RETRANS|TCPCB_EVER_RETRANS)
1179 #define TCPCB_URG 0x20 /* Urgent pointer advenced here */
1181 #define TCPCB_AT_TAIL (TCPCB_URG)
1183 __u16 urg_ptr; /* Valid w/URG flags is set. */
1184 __u32 ack_seq; /* Sequence number ACK'd */
1187 #define TCP_SKB_CB(__skb) ((struct tcp_skb_cb *)&((__skb)->cb[0]))
1189 #include <net/tcp_ecn.h>
1191 /* This determines how many packets are "in the network" to the best
1192 * of our knowledge. In many cases it is conservative, but where
1193 * detailed information is available from the receiver (via SACK
1194 * blocks etc.) we can make more aggressive calculations.
1196 * Use this for decisions involving congestion control, use just
1197 * tp->packets_out to determine if the send queue is empty or not.
1199 * Read this equation as:
1201 * "Packets sent once on transmission queue" MINUS
1202 * "Packets left network, but not honestly ACKed yet" PLUS
1203 * "Packets fast retransmitted"
1205 static __inline__ unsigned int tcp_packets_in_flight(struct tcp_opt *tp)
1207 return tp->packets_out - tp->left_out + tp->retrans_out;
1210 /* Recalculate snd_ssthresh, we want to set it to:
1213 * one half the current congestion window, but no
1214 * less than two segments
1217 * behave like Reno until low_window is reached,
1218 * then increase congestion window slowly
1220 static inline __u32 tcp_recalc_ssthresh(struct tcp_opt *tp)
1222 if (sysctl_tcp_bic) {
1223 if (sysctl_tcp_bic_fast_convergence &&
1224 tp->snd_cwnd < tp->bictcp.last_max_cwnd)
1225 tp->bictcp.last_max_cwnd
1226 = (tp->snd_cwnd * (2*BICTCP_1_OVER_BETA-1))
1227 / (BICTCP_1_OVER_BETA/2);
1229 tp->bictcp.last_max_cwnd = tp->snd_cwnd;
1231 if (tp->snd_cwnd > sysctl_tcp_bic_low_window)
1232 return max(tp->snd_cwnd - (tp->snd_cwnd/BICTCP_1_OVER_BETA),
1236 return max(tp->snd_cwnd >> 1U, 2U);
1239 /* Stop taking Vegas samples for now. */
1240 #define tcp_vegas_disable(__tp) ((__tp)->vegas.doing_vegas_now = 0)
1242 /* Is this TCP connection using Vegas (regardless of whether it is taking
1243 * Vegas measurements at the current time)?
1245 #define tcp_is_vegas(__tp) ((__tp)->vegas.do_vegas)
1247 static inline void tcp_vegas_enable(struct tcp_opt *tp)
1249 /* There are several situations when we must "re-start" Vegas:
1251 * o when a connection is established
1253 * o after fast recovery
1254 * o when we send a packet and there is no outstanding
1255 * unacknowledged data (restarting an idle connection)
1257 * In these circumstances we cannot do a Vegas calculation at the
1258 * end of the first RTT, because any calculation we do is using
1259 * stale info -- both the saved cwnd and congestion feedback are
1262 * Instead we must wait until the completion of an RTT during
1263 * which we actually receive ACKs.
1266 /* Begin taking Vegas samples next time we send something. */
1267 tp->vegas.doing_vegas_now = 1;
1269 /* Set the beginning of the next send window. */
1270 tp->vegas.beg_snd_nxt = tp->snd_nxt;
1272 tp->vegas.cntRTT = 0;
1273 tp->vegas.minRTT = 0x7fffffff;
1276 /* Should we be taking Vegas samples right now? */
1277 #define tcp_vegas_enabled(__tp) ((__tp)->vegas.doing_vegas_now)
1279 extern void tcp_vegas_init(struct tcp_opt *tp);
1281 static inline void tcp_set_ca_state(struct tcp_opt *tp, u8 ca_state)
1283 if (tcp_is_vegas(tp)) {
1284 if (ca_state == TCP_CA_Open)
1285 tcp_vegas_enable(tp);
1287 tcp_vegas_disable(tp);
1289 tp->ca_state = ca_state;
1292 /* If cwnd > ssthresh, we may raise ssthresh to be half-way to cwnd.
1293 * The exception is rate halving phase, when cwnd is decreasing towards
1296 static inline __u32 tcp_current_ssthresh(struct tcp_opt *tp)
1298 if ((1<<tp->ca_state)&(TCPF_CA_CWR|TCPF_CA_Recovery))
1299 return tp->snd_ssthresh;
1301 return max(tp->snd_ssthresh,
1302 ((tp->snd_cwnd >> 1) +
1303 (tp->snd_cwnd >> 2)));
1306 static inline void tcp_sync_left_out(struct tcp_opt *tp)
1308 if (tp->sack_ok && tp->sacked_out >= tp->packets_out - tp->lost_out)
1309 tp->sacked_out = tp->packets_out - tp->lost_out;
1310 tp->left_out = tp->sacked_out + tp->lost_out;
1313 extern void tcp_cwnd_application_limited(struct sock *sk);
1315 /* Congestion window validation. (RFC2861) */
1317 static inline void tcp_cwnd_validate(struct sock *sk, struct tcp_opt *tp)
1319 if (tp->packets_out >= tp->snd_cwnd) {
1320 /* Network is feed fully. */
1321 tp->snd_cwnd_used = 0;
1322 tp->snd_cwnd_stamp = tcp_time_stamp;
1324 /* Network starves. */
1325 if (tp->packets_out > tp->snd_cwnd_used)
1326 tp->snd_cwnd_used = tp->packets_out;
1328 if ((s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= tp->rto)
1329 tcp_cwnd_application_limited(sk);
1333 /* Set slow start threshould and cwnd not falling to slow start */
1334 static inline void __tcp_enter_cwr(struct tcp_opt *tp)
1336 tp->undo_marker = 0;
1337 tp->snd_ssthresh = tcp_recalc_ssthresh(tp);
1338 tp->snd_cwnd = min(tp->snd_cwnd,
1339 tcp_packets_in_flight(tp) + 1U);
1340 tp->snd_cwnd_cnt = 0;
1341 tp->high_seq = tp->snd_nxt;
1342 tp->snd_cwnd_stamp = tcp_time_stamp;
1343 TCP_ECN_queue_cwr(tp);
1346 static inline void tcp_enter_cwr(struct tcp_opt *tp)
1348 tp->prior_ssthresh = 0;
1349 if (tp->ca_state < TCP_CA_CWR) {
1350 __tcp_enter_cwr(tp);
1351 tcp_set_ca_state(tp, TCP_CA_CWR);
1355 extern __u32 tcp_init_cwnd(struct tcp_opt *tp, struct dst_entry *dst);
1357 /* Slow start with delack produces 3 packets of burst, so that
1358 * it is safe "de facto".
1360 static __inline__ __u32 tcp_max_burst(struct tcp_opt *tp)
1365 static __inline__ int tcp_minshall_check(struct tcp_opt *tp)
1367 return after(tp->snd_sml,tp->snd_una) &&
1368 !after(tp->snd_sml, tp->snd_nxt);
1371 static __inline__ void tcp_minshall_update(struct tcp_opt *tp, int mss, struct sk_buff *skb)
1374 tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1377 /* Return 0, if packet can be sent now without violation Nagle's rules:
1378 1. It is full sized.
1379 2. Or it contains FIN.
1380 3. Or TCP_NODELAY was set.
1381 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1382 With Minshall's modification: all sent small packets are ACKed.
1385 static __inline__ int
1386 tcp_nagle_check(struct tcp_opt *tp, struct sk_buff *skb, unsigned mss_now, int nonagle)
1388 return (skb->len < mss_now &&
1389 !(TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN) &&
1390 ((nonagle&TCP_NAGLE_CORK) ||
1393 tcp_minshall_check(tp))));
1396 /* This checks if the data bearing packet SKB (usually sk->sk_send_head)
1397 * should be put on the wire right now.
1399 static __inline__ int tcp_snd_test(struct tcp_opt *tp, struct sk_buff *skb,
1400 unsigned cur_mss, int nonagle)
1402 /* RFC 1122 - section 4.2.3.4
1406 * a) The right edge of this frame exceeds the window
1407 * b) There are packets in flight and we have a small segment
1408 * [SWS avoidance and Nagle algorithm]
1409 * (part of SWS is done on packetization)
1410 * Minshall version sounds: there are no _small_
1411 * segments in flight. (tcp_nagle_check)
1412 * c) We have too many packets 'in flight'
1414 * Don't use the nagle rule for urgent data (or
1415 * for the final FIN -DaveM).
1417 * Also, Nagle rule does not apply to frames, which
1418 * sit in the middle of queue (they have no chances
1419 * to get new data) and if room at tail of skb is
1420 * not enough to save something seriously (<32 for now).
1423 /* Don't be strict about the congestion window for the
1424 * final FIN frame. -DaveM
1426 return (((nonagle&TCP_NAGLE_PUSH) || tp->urg_mode
1427 || !tcp_nagle_check(tp, skb, cur_mss, nonagle)) &&
1428 ((tcp_packets_in_flight(tp) < tp->snd_cwnd) ||
1429 (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN)) &&
1430 !after(TCP_SKB_CB(skb)->end_seq, tp->snd_una + tp->snd_wnd));
1433 static __inline__ void tcp_check_probe_timer(struct sock *sk, struct tcp_opt *tp)
1435 if (!tp->packets_out && !tp->pending)
1436 tcp_reset_xmit_timer(sk, TCP_TIME_PROBE0, tp->rto);
1439 static __inline__ int tcp_skb_is_last(struct sock *sk, struct sk_buff *skb)
1441 return skb->next == (struct sk_buff *)&sk->sk_write_queue;
1444 /* Push out any pending frames which were held back due to
1445 * TCP_CORK or attempt at coalescing tiny packets.
1446 * The socket must be locked by the caller.
1448 static __inline__ void __tcp_push_pending_frames(struct sock *sk,
1453 struct sk_buff *skb = sk->sk_send_head;
1456 if (!tcp_skb_is_last(sk, skb))
1457 nonagle = TCP_NAGLE_PUSH;
1458 if (!tcp_snd_test(tp, skb, cur_mss, nonagle) ||
1459 tcp_write_xmit(sk, nonagle))
1460 tcp_check_probe_timer(sk, tp);
1462 tcp_cwnd_validate(sk, tp);
1465 static __inline__ void tcp_push_pending_frames(struct sock *sk,
1468 __tcp_push_pending_frames(sk, tp, tcp_current_mss(sk, 1), tp->nonagle);
1471 static __inline__ int tcp_may_send_now(struct sock *sk, struct tcp_opt *tp)
1473 struct sk_buff *skb = sk->sk_send_head;
1476 tcp_snd_test(tp, skb, tcp_current_mss(sk, 1),
1477 tcp_skb_is_last(sk, skb) ? TCP_NAGLE_PUSH : tp->nonagle));
1480 static __inline__ void tcp_init_wl(struct tcp_opt *tp, u32 ack, u32 seq)
1485 static __inline__ void tcp_update_wl(struct tcp_opt *tp, u32 ack, u32 seq)
1490 extern void tcp_destroy_sock(struct sock *sk);
1494 * Calculate(/check) TCP checksum
1496 static __inline__ u16 tcp_v4_check(struct tcphdr *th, int len,
1497 unsigned long saddr, unsigned long daddr,
1500 return csum_tcpudp_magic(saddr,daddr,len,IPPROTO_TCP,base);
1503 static __inline__ int __tcp_checksum_complete(struct sk_buff *skb)
1505 return (unsigned short)csum_fold(skb_checksum(skb, 0, skb->len, skb->csum));
1508 static __inline__ int tcp_checksum_complete(struct sk_buff *skb)
1510 return skb->ip_summed != CHECKSUM_UNNECESSARY &&
1511 __tcp_checksum_complete(skb);
1514 /* Prequeue for VJ style copy to user, combined with checksumming. */
1516 static __inline__ void tcp_prequeue_init(struct tcp_opt *tp)
1518 tp->ucopy.task = NULL;
1520 tp->ucopy.memory = 0;
1521 skb_queue_head_init(&tp->ucopy.prequeue);
1524 /* Packet is added to VJ-style prequeue for processing in process
1525 * context, if a reader task is waiting. Apparently, this exciting
1526 * idea (VJ's mail "Re: query about TCP header on tcp-ip" of 07 Sep 93)
1527 * failed somewhere. Latency? Burstiness? Well, at least now we will
1528 * see, why it failed. 8)8) --ANK
1530 * NOTE: is this not too big to inline?
1532 static __inline__ int tcp_prequeue(struct sock *sk, struct sk_buff *skb)
1534 struct tcp_opt *tp = tcp_sk(sk);
1536 if (!sysctl_tcp_low_latency && tp->ucopy.task) {
1537 __skb_queue_tail(&tp->ucopy.prequeue, skb);
1538 tp->ucopy.memory += skb->truesize;
1539 if (tp->ucopy.memory > sk->sk_rcvbuf) {
1540 struct sk_buff *skb1;
1542 BUG_ON(sock_owned_by_user(sk));
1544 while ((skb1 = __skb_dequeue(&tp->ucopy.prequeue)) != NULL) {
1545 sk->sk_backlog_rcv(sk, skb1);
1546 NET_INC_STATS_BH(LINUX_MIB_TCPPREQUEUEDROPPED);
1549 tp->ucopy.memory = 0;
1550 } else if (skb_queue_len(&tp->ucopy.prequeue) == 1) {
1551 wake_up_interruptible(sk->sk_sleep);
1552 if (!tcp_ack_scheduled(tp))
1553 tcp_reset_xmit_timer(sk, TCP_TIME_DACK, (3*TCP_RTO_MIN)/4);
1564 static char *statename[]={
1565 "Unused","Established","Syn Sent","Syn Recv",
1566 "Fin Wait 1","Fin Wait 2","Time Wait", "Close",
1567 "Close Wait","Last ACK","Listen","Closing"
1571 static __inline__ void tcp_set_state(struct sock *sk, int state)
1573 int oldstate = sk->sk_state;
1576 case TCP_ESTABLISHED:
1577 if (oldstate != TCP_ESTABLISHED)
1578 TCP_INC_STATS(TCP_MIB_CURRESTAB);
1582 if (oldstate == TCP_CLOSE_WAIT || oldstate == TCP_ESTABLISHED)
1583 TCP_INC_STATS(TCP_MIB_ESTABRESETS);
1585 sk->sk_prot->unhash(sk);
1586 if (tcp_sk(sk)->bind_hash &&
1587 !(sk->sk_userlocks & SOCK_BINDPORT_LOCK))
1591 if (oldstate==TCP_ESTABLISHED)
1592 TCP_DEC_STATS(TCP_MIB_CURRESTAB);
1595 /* Change state AFTER socket is unhashed to avoid closed
1596 * socket sitting in hash tables.
1598 sk->sk_state = state;
1601 SOCK_DEBUG(sk, "TCP sk=%p, State %s -> %s\n",sk, statename[oldstate],statename[state]);
1605 static __inline__ void tcp_done(struct sock *sk)
1607 tcp_set_state(sk, TCP_CLOSE);
1608 tcp_clear_xmit_timers(sk);
1610 sk->sk_shutdown = SHUTDOWN_MASK;
1612 if (!sock_flag(sk, SOCK_DEAD))
1613 sk->sk_state_change(sk);
1615 tcp_destroy_sock(sk);
1618 static __inline__ void tcp_sack_reset(struct tcp_opt *tp)
1625 static __inline__ void tcp_build_and_update_options(__u32 *ptr, struct tcp_opt *tp, __u32 tstamp)
1627 if (tp->tstamp_ok) {
1628 *ptr++ = __constant_htonl((TCPOPT_NOP << 24) |
1629 (TCPOPT_NOP << 16) |
1630 (TCPOPT_TIMESTAMP << 8) |
1632 *ptr++ = htonl(tstamp);
1633 *ptr++ = htonl(tp->ts_recent);
1635 if (tp->eff_sacks) {
1636 struct tcp_sack_block *sp = tp->dsack ? tp->duplicate_sack : tp->selective_acks;
1639 *ptr++ = __constant_htonl((TCPOPT_NOP << 24) |
1640 (TCPOPT_NOP << 16) |
1641 (TCPOPT_SACK << 8) |
1642 (TCPOLEN_SACK_BASE +
1643 (tp->eff_sacks * TCPOLEN_SACK_PERBLOCK)));
1644 for(this_sack = 0; this_sack < tp->eff_sacks; this_sack++) {
1645 *ptr++ = htonl(sp[this_sack].start_seq);
1646 *ptr++ = htonl(sp[this_sack].end_seq);
1655 /* Construct a tcp options header for a SYN or SYN_ACK packet.
1656 * If this is every changed make sure to change the definition of
1657 * MAX_SYN_SIZE to match the new maximum number of options that you
1660 static inline void tcp_syn_build_options(__u32 *ptr, int mss, int ts, int sack,
1661 int offer_wscale, int wscale, __u32 tstamp, __u32 ts_recent)
1663 /* We always get an MSS option.
1664 * The option bytes which will be seen in normal data
1665 * packets should timestamps be used, must be in the MSS
1666 * advertised. But we subtract them from tp->mss_cache so
1667 * that calculations in tcp_sendmsg are simpler etc.
1668 * So account for this fact here if necessary. If we
1669 * don't do this correctly, as a receiver we won't
1670 * recognize data packets as being full sized when we
1671 * should, and thus we won't abide by the delayed ACK
1673 * SACKs don't matter, we never delay an ACK when we
1674 * have any of those going out.
1676 *ptr++ = htonl((TCPOPT_MSS << 24) | (TCPOLEN_MSS << 16) | mss);
1679 *ptr++ = __constant_htonl((TCPOPT_SACK_PERM << 24) | (TCPOLEN_SACK_PERM << 16) |
1680 (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP);
1682 *ptr++ = __constant_htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) |
1683 (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP);
1684 *ptr++ = htonl(tstamp); /* TSVAL */
1685 *ptr++ = htonl(ts_recent); /* TSECR */
1687 *ptr++ = __constant_htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) |
1688 (TCPOPT_SACK_PERM << 8) | TCPOLEN_SACK_PERM);
1690 *ptr++ = htonl((TCPOPT_NOP << 24) | (TCPOPT_WINDOW << 16) | (TCPOLEN_WINDOW << 8) | (wscale));
1693 /* Determine a window scaling and initial window to offer.
1694 * Based on the assumption that the given amount of space
1695 * will be offered. Store the results in the tp structure.
1696 * NOTE: for smooth operation initial space offering should
1697 * be a multiple of mss if possible. We assume here that mss >= 1.
1698 * This MUST be enforced by all callers.
1700 static inline void tcp_select_initial_window(int __space, __u32 mss,
1702 __u32 *window_clamp,
1706 unsigned int space = (__space < 0 ? 0 : __space);
1708 /* If no clamp set the clamp to the max possible scaled window */
1709 if (*window_clamp == 0)
1710 (*window_clamp) = (65535 << 14);
1711 space = min(*window_clamp, space);
1713 /* Quantize space offering to a multiple of mss if possible. */
1715 space = (space / mss) * mss;
1717 /* NOTE: offering an initial window larger than 32767
1718 * will break some buggy TCP stacks. We try to be nice.
1719 * If we are not window scaling, then this truncates
1720 * our initial window offering to 32k. There should also
1721 * be a sysctl option to stop being nice.
1723 (*rcv_wnd) = min(space, MAX_TCP_WINDOW);
1726 /* See RFC1323 for an explanation of the limit to 14 */
1727 while (space > 65535 && (*rcv_wscale) < 14) {
1731 if (*rcv_wscale && sysctl_tcp_app_win && space>=mss &&
1732 space - max((space>>sysctl_tcp_app_win), mss>>*rcv_wscale) < 65536/2)
1735 *rcv_wscale = max((__u8)sysctl_tcp_default_win_scale,
1739 /* Set initial window to value enough for senders,
1740 * following RFC1414. Senders, not following this RFC,
1741 * will be satisfied with 2.
1743 if (mss > (1<<*rcv_wscale)) {
1747 else if (mss > 1460)
1749 if (*rcv_wnd > init_cwnd*mss)
1750 *rcv_wnd = init_cwnd*mss;
1752 /* Set the clamp no higher than max representable value */
1753 (*window_clamp) = min(65535U << (*rcv_wscale), *window_clamp);
1756 static inline int tcp_win_from_space(int space)
1758 return sysctl_tcp_adv_win_scale<=0 ?
1759 (space>>(-sysctl_tcp_adv_win_scale)) :
1760 space - (space>>sysctl_tcp_adv_win_scale);
1763 /* Note: caller must be prepared to deal with negative returns */
1764 static inline int tcp_space(struct sock *sk)
1766 return tcp_win_from_space(sk->sk_rcvbuf -
1767 atomic_read(&sk->sk_rmem_alloc));
1770 static inline int tcp_full_space( struct sock *sk)
1772 return tcp_win_from_space(sk->sk_rcvbuf);
1775 static inline void tcp_acceptq_queue(struct sock *sk, struct open_request *req,
1778 struct tcp_opt *tp = tcp_sk(sk);
1781 sk_acceptq_added(sk);
1783 if (!tp->accept_queue_tail) {
1784 tp->accept_queue = req;
1786 tp->accept_queue_tail->dl_next = req;
1788 tp->accept_queue_tail = req;
1789 req->dl_next = NULL;
1792 struct tcp_listen_opt
1794 u8 max_qlen_log; /* log_2 of maximal queued SYNs */
1799 struct open_request *syn_table[TCP_SYNQ_HSIZE];
1803 tcp_synq_removed(struct sock *sk, struct open_request *req)
1805 struct tcp_listen_opt *lopt = tcp_sk(sk)->listen_opt;
1807 if (--lopt->qlen == 0)
1808 tcp_delete_keepalive_timer(sk);
1809 if (req->retrans == 0)
1813 static inline void tcp_synq_added(struct sock *sk)
1815 struct tcp_listen_opt *lopt = tcp_sk(sk)->listen_opt;
1817 if (lopt->qlen++ == 0)
1818 tcp_reset_keepalive_timer(sk, TCP_TIMEOUT_INIT);
1822 static inline int tcp_synq_len(struct sock *sk)
1824 return tcp_sk(sk)->listen_opt->qlen;
1827 static inline int tcp_synq_young(struct sock *sk)
1829 return tcp_sk(sk)->listen_opt->qlen_young;
1832 static inline int tcp_synq_is_full(struct sock *sk)
1834 return tcp_synq_len(sk) >> tcp_sk(sk)->listen_opt->max_qlen_log;
1837 static inline void tcp_synq_unlink(struct tcp_opt *tp, struct open_request *req,
1838 struct open_request **prev)
1840 write_lock(&tp->syn_wait_lock);
1841 *prev = req->dl_next;
1842 write_unlock(&tp->syn_wait_lock);
1845 static inline void tcp_synq_drop(struct sock *sk, struct open_request *req,
1846 struct open_request **prev)
1848 tcp_synq_unlink(tcp_sk(sk), req, prev);
1849 tcp_synq_removed(sk, req);
1850 tcp_openreq_free(req);
1853 static __inline__ void tcp_openreq_init(struct open_request *req,
1855 struct sk_buff *skb)
1857 req->rcv_wnd = 0; /* So that tcp_send_synack() knows! */
1858 req->rcv_isn = TCP_SKB_CB(skb)->seq;
1859 req->mss = tp->mss_clamp;
1860 req->ts_recent = tp->saw_tstamp ? tp->rcv_tsval : 0;
1861 req->tstamp_ok = tp->tstamp_ok;
1862 req->sack_ok = tp->sack_ok;
1863 req->snd_wscale = tp->snd_wscale;
1864 req->wscale_ok = tp->wscale_ok;
1867 req->rmt_port = skb->h.th->source;
1870 extern void tcp_enter_memory_pressure(void);
1872 extern void tcp_listen_wlock(void);
1874 /* - We may sleep inside this lock.
1875 * - If sleeping is not required (or called from BH),
1876 * use plain read_(un)lock(&tcp_lhash_lock).
1879 static inline void tcp_listen_lock(void)
1881 /* read_lock synchronizes to candidates to writers */
1882 read_lock(&tcp_lhash_lock);
1883 atomic_inc(&tcp_lhash_users);
1884 read_unlock(&tcp_lhash_lock);
1887 static inline void tcp_listen_unlock(void)
1889 if (atomic_dec_and_test(&tcp_lhash_users))
1890 wake_up(&tcp_lhash_wait);
1893 static inline int keepalive_intvl_when(struct tcp_opt *tp)
1895 return tp->keepalive_intvl ? : sysctl_tcp_keepalive_intvl;
1898 static inline int keepalive_time_when(struct tcp_opt *tp)
1900 return tp->keepalive_time ? : sysctl_tcp_keepalive_time;
1903 static inline int tcp_fin_time(struct tcp_opt *tp)
1905 int fin_timeout = tp->linger2 ? : sysctl_tcp_fin_timeout;
1907 if (fin_timeout < (tp->rto<<2) - (tp->rto>>1))
1908 fin_timeout = (tp->rto<<2) - (tp->rto>>1);
1913 static inline int tcp_paws_check(struct tcp_opt *tp, int rst)
1915 if ((s32)(tp->rcv_tsval - tp->ts_recent) >= 0)
1917 if (xtime.tv_sec >= tp->ts_recent_stamp + TCP_PAWS_24DAYS)
1920 /* RST segments are not recommended to carry timestamp,
1921 and, if they do, it is recommended to ignore PAWS because
1922 "their cleanup function should take precedence over timestamps."
1923 Certainly, it is mistake. It is necessary to understand the reasons
1924 of this constraint to relax it: if peer reboots, clock may go
1925 out-of-sync and half-open connections will not be reset.
1926 Actually, the problem would be not existing if all
1927 the implementations followed draft about maintaining clock
1928 via reboots. Linux-2.2 DOES NOT!
1930 However, we can relax time bounds for RST segments to MSL.
1932 if (rst && xtime.tv_sec >= tp->ts_recent_stamp + TCP_PAWS_MSL)
1937 static inline void tcp_v4_setup_caps(struct sock *sk, struct dst_entry *dst)
1939 sk->sk_route_caps = dst->dev->features;
1940 if (sk->sk_route_caps & NETIF_F_TSO) {
1941 if (sk->sk_no_largesend || dst->header_len)
1942 sk->sk_route_caps &= ~NETIF_F_TSO;
1946 #define TCP_CHECK_TIMER(sk) do { } while (0)
1948 static inline int tcp_use_frto(const struct sock *sk)
1950 const struct tcp_opt *tp = tcp_sk(sk);
1952 /* F-RTO must be activated in sysctl and there must be some
1953 * unsent new data, and the advertised window should allow
1956 return (sysctl_tcp_frto && sk->sk_send_head &&
1957 !after(TCP_SKB_CB(sk->sk_send_head)->end_seq,
1958 tp->snd_una + tp->snd_wnd));
1961 static inline void tcp_mib_init(void)
1964 TCP_ADD_STATS_USER(TCP_MIB_RTOALGORITHM, 1);
1965 TCP_ADD_STATS_USER(TCP_MIB_RTOMIN, TCP_RTO_MIN*1000/HZ);
1966 TCP_ADD_STATS_USER(TCP_MIB_RTOMAX, TCP_RTO_MAX*1000/HZ);
1967 TCP_ADD_STATS_USER(TCP_MIB_MAXCONN, -1);
1971 enum tcp_seq_states {
1972 TCP_SEQ_STATE_LISTENING,
1973 TCP_SEQ_STATE_OPENREQ,
1974 TCP_SEQ_STATE_ESTABLISHED,
1975 TCP_SEQ_STATE_TIME_WAIT,
1978 struct tcp_seq_afinfo {
1979 struct module *owner;
1982 int (*seq_show) (struct seq_file *m, void *v);
1983 struct file_operations *seq_fops;
1986 struct tcp_iter_state {
1988 enum tcp_seq_states state;
1989 struct sock *syn_wait_sk;
1990 int bucket, sbucket, num, uid;
1991 struct seq_operations seq_ops;
1994 extern int tcp_proc_register(struct tcp_seq_afinfo *afinfo);
1995 extern void tcp_proc_unregister(struct tcp_seq_afinfo *afinfo);
1997 /* TCP Westwood functions and constants */
1999 #define TCP_WESTWOOD_INIT_RTT (20*HZ) /* maybe too conservative?! */
2000 #define TCP_WESTWOOD_RTT_MIN (HZ/20) /* 50ms */
2002 static inline void tcp_westwood_update_rtt(struct tcp_opt *tp, __u32 rtt_seq)
2004 if (sysctl_tcp_westwood)
2005 tp->westwood.rtt = rtt_seq;
2008 void __tcp_westwood_fast_bw(struct sock *, struct sk_buff *);
2009 void __tcp_westwood_slow_bw(struct sock *, struct sk_buff *);
2011 static inline void tcp_westwood_fast_bw(struct sock *sk, struct sk_buff *skb)
2013 if (sysctl_tcp_westwood)
2014 __tcp_westwood_fast_bw(sk, skb);
2017 static inline void tcp_westwood_slow_bw(struct sock *sk, struct sk_buff *skb)
2019 if (sysctl_tcp_westwood)
2020 __tcp_westwood_slow_bw(sk, skb);
2023 static inline __u32 __tcp_westwood_bw_rttmin(const struct tcp_opt *tp)
2025 return max((tp->westwood.bw_est) * (tp->westwood.rtt_min) /
2026 (__u32) (tp->mss_cache),
2030 static inline __u32 tcp_westwood_bw_rttmin(const struct tcp_opt *tp)
2032 return sysctl_tcp_westwood ? __tcp_westwood_bw_rttmin(tp) : 0;
2035 static inline int tcp_westwood_ssthresh(struct tcp_opt *tp)
2039 if (sysctl_tcp_westwood) {
2040 ssthresh = __tcp_westwood_bw_rttmin(tp);
2042 tp->snd_ssthresh = ssthresh;
2045 return (ssthresh != 0);
2048 static inline int tcp_westwood_cwnd(struct tcp_opt *tp)
2052 if (sysctl_tcp_westwood) {
2053 cwnd = __tcp_westwood_bw_rttmin(tp);
2055 tp->snd_cwnd = cwnd;