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 * Implementation of the Transmission Control Protocol(TCP).
8 * Version: $Id: tcp.c,v 1.216 2002/02/01 22:01:04 davem Exp $
11 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
12 * Mark Evans, <evansmp@uhura.aston.ac.uk>
13 * Corey Minyard <wf-rch!minyard@relay.EU.net>
14 * Florian La Roche, <flla@stud.uni-sb.de>
15 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
16 * Linus Torvalds, <torvalds@cs.helsinki.fi>
17 * Alan Cox, <gw4pts@gw4pts.ampr.org>
18 * Matthew Dillon, <dillon@apollo.west.oic.com>
19 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
20 * Jorge Cwik, <jorge@laser.satlink.net>
23 * Alan Cox : Numerous verify_area() calls
24 * Alan Cox : Set the ACK bit on a reset
25 * Alan Cox : Stopped it crashing if it closed while
26 * sk->inuse=1 and was trying to connect
28 * Alan Cox : All icmp error handling was broken
29 * pointers passed where wrong and the
30 * socket was looked up backwards. Nobody
31 * tested any icmp error code obviously.
32 * Alan Cox : tcp_err() now handled properly. It
33 * wakes people on errors. poll
34 * behaves and the icmp error race
35 * has gone by moving it into sock.c
36 * Alan Cox : tcp_send_reset() fixed to work for
37 * everything not just packets for
39 * Alan Cox : tcp option processing.
40 * Alan Cox : Reset tweaked (still not 100%) [Had
42 * Herp Rosmanith : More reset fixes
43 * Alan Cox : No longer acks invalid rst frames.
44 * Acking any kind of RST is right out.
45 * Alan Cox : Sets an ignore me flag on an rst
46 * receive otherwise odd bits of prattle
48 * Alan Cox : Fixed another acking RST frame bug.
49 * Should stop LAN workplace lockups.
50 * Alan Cox : Some tidyups using the new skb list
52 * Alan Cox : sk->keepopen now seems to work
53 * Alan Cox : Pulls options out correctly on accepts
54 * Alan Cox : Fixed assorted sk->rqueue->next errors
55 * Alan Cox : PSH doesn't end a TCP read. Switched a
57 * Alan Cox : Tidied tcp_data to avoid a potential
59 * Alan Cox : Added some better commenting, as the
60 * tcp is hard to follow
61 * Alan Cox : Removed incorrect check for 20 * psh
62 * Michael O'Reilly : ack < copied bug fix.
63 * Johannes Stille : Misc tcp fixes (not all in yet).
64 * Alan Cox : FIN with no memory -> CRASH
65 * Alan Cox : Added socket option proto entries.
66 * Also added awareness of them to accept.
67 * Alan Cox : Added TCP options (SOL_TCP)
68 * Alan Cox : Switched wakeup calls to callbacks,
69 * so the kernel can layer network
71 * Alan Cox : Use ip_tos/ip_ttl settings.
72 * Alan Cox : Handle FIN (more) properly (we hope).
73 * Alan Cox : RST frames sent on unsynchronised
75 * Alan Cox : Put in missing check for SYN bit.
76 * Alan Cox : Added tcp_select_window() aka NET2E
77 * window non shrink trick.
78 * Alan Cox : Added a couple of small NET2E timer
80 * Charles Hedrick : TCP fixes
81 * Toomas Tamm : TCP window fixes
82 * Alan Cox : Small URG fix to rlogin ^C ack fight
83 * Charles Hedrick : Rewrote most of it to actually work
84 * Linus : Rewrote tcp_read() and URG handling
86 * Gerhard Koerting: Fixed some missing timer handling
87 * Matthew Dillon : Reworked TCP machine states as per RFC
88 * Gerhard Koerting: PC/TCP workarounds
89 * Adam Caldwell : Assorted timer/timing errors
90 * Matthew Dillon : Fixed another RST bug
91 * Alan Cox : Move to kernel side addressing changes.
92 * Alan Cox : Beginning work on TCP fastpathing
94 * Arnt Gulbrandsen: Turbocharged tcp_check() routine.
95 * Alan Cox : TCP fast path debugging
96 * Alan Cox : Window clamping
97 * Michael Riepe : Bug in tcp_check()
98 * Matt Dillon : More TCP improvements and RST bug fixes
99 * Matt Dillon : Yet more small nasties remove from the
100 * TCP code (Be very nice to this man if
101 * tcp finally works 100%) 8)
102 * Alan Cox : BSD accept semantics.
103 * Alan Cox : Reset on closedown bug.
104 * Peter De Schrijver : ENOTCONN check missing in tcp_sendto().
105 * Michael Pall : Handle poll() after URG properly in
107 * Michael Pall : Undo the last fix in tcp_read_urg()
108 * (multi URG PUSH broke rlogin).
109 * Michael Pall : Fix the multi URG PUSH problem in
110 * tcp_readable(), poll() after URG
112 * Michael Pall : recv(...,MSG_OOB) never blocks in the
114 * Alan Cox : Changed the semantics of sk->socket to
115 * fix a race and a signal problem with
116 * accept() and async I/O.
117 * Alan Cox : Relaxed the rules on tcp_sendto().
118 * Yury Shevchuk : Really fixed accept() blocking problem.
119 * Craig I. Hagan : Allow for BSD compatible TIME_WAIT for
120 * clients/servers which listen in on
122 * Alan Cox : Cleaned the above up and shrank it to
123 * a sensible code size.
124 * Alan Cox : Self connect lockup fix.
125 * Alan Cox : No connect to multicast.
126 * Ross Biro : Close unaccepted children on master
128 * Alan Cox : Reset tracing code.
129 * Alan Cox : Spurious resets on shutdown.
130 * Alan Cox : Giant 15 minute/60 second timer error
131 * Alan Cox : Small whoops in polling before an
133 * Alan Cox : Kept the state trace facility since
134 * it's handy for debugging.
135 * Alan Cox : More reset handler fixes.
136 * Alan Cox : Started rewriting the code based on
137 * the RFC's for other useful protocol
138 * references see: Comer, KA9Q NOS, and
139 * for a reference on the difference
140 * between specifications and how BSD
141 * works see the 4.4lite source.
142 * A.N.Kuznetsov : Don't time wait on completion of tidy
144 * Linus Torvalds : Fin/Shutdown & copied_seq changes.
145 * Linus Torvalds : Fixed BSD port reuse to work first syn
146 * Alan Cox : Reimplemented timers as per the RFC
147 * and using multiple timers for sanity.
148 * Alan Cox : Small bug fixes, and a lot of new
150 * Alan Cox : Fixed dual reader crash by locking
151 * the buffers (much like datagram.c)
152 * Alan Cox : Fixed stuck sockets in probe. A probe
153 * now gets fed up of retrying without
154 * (even a no space) answer.
155 * Alan Cox : Extracted closing code better
156 * Alan Cox : Fixed the closing state machine to
158 * Alan Cox : More 'per spec' fixes.
159 * Jorge Cwik : Even faster checksumming.
160 * Alan Cox : tcp_data() doesn't ack illegal PSH
161 * only frames. At least one pc tcp stack
163 * Alan Cox : Cache last socket.
164 * Alan Cox : Per route irtt.
165 * Matt Day : poll()->select() match BSD precisely on error
166 * Alan Cox : New buffers
167 * Marc Tamsky : Various sk->prot->retransmits and
168 * sk->retransmits misupdating fixed.
169 * Fixed tcp_write_timeout: stuck close,
170 * and TCP syn retries gets used now.
171 * Mark Yarvis : In tcp_read_wakeup(), don't send an
172 * ack if state is TCP_CLOSED.
173 * Alan Cox : Look up device on a retransmit - routes may
174 * change. Doesn't yet cope with MSS shrink right
176 * Marc Tamsky : Closing in closing fixes.
177 * Mike Shaver : RFC1122 verifications.
178 * Alan Cox : rcv_saddr errors.
179 * Alan Cox : Block double connect().
180 * Alan Cox : Small hooks for enSKIP.
181 * Alexey Kuznetsov: Path MTU discovery.
182 * Alan Cox : Support soft errors.
183 * Alan Cox : Fix MTU discovery pathological case
184 * when the remote claims no mtu!
185 * Marc Tamsky : TCP_CLOSE fix.
186 * Colin (G3TNE) : Send a reset on syn ack replies in
187 * window but wrong (fixes NT lpd problems)
188 * Pedro Roque : Better TCP window handling, delayed ack.
189 * Joerg Reuter : No modification of locked buffers in
190 * tcp_do_retransmit()
191 * Eric Schenk : Changed receiver side silly window
192 * avoidance algorithm to BSD style
193 * algorithm. This doubles throughput
194 * against machines running Solaris,
195 * and seems to result in general
197 * Stefan Magdalinski : adjusted tcp_readable() to fix FIONREAD
198 * Willy Konynenberg : Transparent proxying support.
199 * Mike McLagan : Routing by source
200 * Keith Owens : Do proper merging with partial SKB's in
201 * tcp_do_sendmsg to avoid burstiness.
202 * Eric Schenk : Fix fast close down bug with
203 * shutdown() followed by close().
204 * Andi Kleen : Make poll agree with SIGIO
205 * Salvatore Sanfilippo : Support SO_LINGER with linger == 1 and
206 * lingertime == 0 (RFC 793 ABORT Call)
207 * Hirokazu Takahashi : Use copy_from_user() instead of
208 * csum_and_copy_from_user() if possible.
210 * This program is free software; you can redistribute it and/or
211 * modify it under the terms of the GNU General Public License
212 * as published by the Free Software Foundation; either version
213 * 2 of the License, or(at your option) any later version.
215 * Description of States:
217 * TCP_SYN_SENT sent a connection request, waiting for ack
219 * TCP_SYN_RECV received a connection request, sent ack,
220 * waiting for final ack in three-way handshake.
222 * TCP_ESTABLISHED connection established
224 * TCP_FIN_WAIT1 our side has shutdown, waiting to complete
225 * transmission of remaining buffered data
227 * TCP_FIN_WAIT2 all buffered data sent, waiting for remote
230 * TCP_CLOSING both sides have shutdown but we still have
231 * data we have to finish sending
233 * TCP_TIME_WAIT timeout to catch resent junk before entering
234 * closed, can only be entered from FIN_WAIT2
235 * or CLOSING. Required because the other end
236 * may not have gotten our last ACK causing it
237 * to retransmit the data packet (which we ignore)
239 * TCP_CLOSE_WAIT remote side has shutdown and is waiting for
240 * us to finish writing our data and to shutdown
241 * (we have to close() to move on to LAST_ACK)
243 * TCP_LAST_ACK out side has shutdown after remote has
244 * shutdown. There may still be data in our
245 * buffer that we have to finish sending
247 * TCP_CLOSE socket is finished
250 #include <linux/config.h>
251 #include <linux/module.h>
252 #include <linux/types.h>
253 #include <linux/fcntl.h>
254 #include <linux/poll.h>
255 #include <linux/init.h>
256 #include <linux/smp_lock.h>
257 #include <linux/fs.h>
258 #include <linux/random.h>
259 #include <linux/bootmem.h>
261 #include <net/icmp.h>
263 #include <net/xfrm.h>
267 #include <asm/uaccess.h>
268 #include <asm/ioctls.h>
270 int sysctl_tcp_fin_timeout = TCP_FIN_TIMEOUT;
272 DEFINE_SNMP_STAT(struct tcp_mib, tcp_statistics);
274 kmem_cache_t *tcp_openreq_cachep;
275 kmem_cache_t *tcp_bucket_cachep;
276 kmem_cache_t *tcp_timewait_cachep;
278 atomic_t tcp_orphan_count = ATOMIC_INIT(0);
280 int sysctl_tcp_mem[3];
281 int sysctl_tcp_wmem[3] = { 4 * 1024, 16 * 1024, 128 * 1024 };
282 int sysctl_tcp_rmem[3] = { 4 * 1024, 87380, 87380 * 2 };
284 EXPORT_SYMBOL(sysctl_tcp_mem);
285 EXPORT_SYMBOL(sysctl_tcp_rmem);
286 EXPORT_SYMBOL(sysctl_tcp_wmem);
288 atomic_t tcp_memory_allocated; /* Current allocated memory. */
289 atomic_t tcp_sockets_allocated; /* Current number of TCP sockets. */
291 EXPORT_SYMBOL(tcp_memory_allocated);
292 EXPORT_SYMBOL(tcp_sockets_allocated);
295 * Pressure flag: try to collapse.
296 * Technical note: it is used by multiple contexts non atomically.
297 * All the sk_stream_mem_schedule() is of this nature: accounting
298 * is strict, actions are advisory and have some latency.
300 int tcp_memory_pressure;
302 EXPORT_SYMBOL(tcp_memory_pressure);
304 void tcp_enter_memory_pressure(void)
306 if (!tcp_memory_pressure) {
307 NET_INC_STATS(LINUX_MIB_TCPMEMORYPRESSURES);
308 tcp_memory_pressure = 1;
312 EXPORT_SYMBOL(tcp_enter_memory_pressure);
315 * LISTEN is a special case for poll..
317 static __inline__ unsigned int tcp_listen_poll(struct sock *sk,
320 return tcp_sk(sk)->accept_queue ? (POLLIN | POLLRDNORM) : 0;
324 * Wait for a TCP event.
326 * Note that we don't need to lock the socket, as the upper poll layers
327 * take care of normal races (between the test and the event) and we don't
328 * go look at any of the socket buffers directly.
330 unsigned int tcp_poll(struct file *file, struct socket *sock, poll_table *wait)
333 struct sock *sk = sock->sk;
334 struct tcp_sock *tp = tcp_sk(sk);
336 poll_wait(file, sk->sk_sleep, wait);
337 if (sk->sk_state == TCP_LISTEN)
338 return tcp_listen_poll(sk, wait);
340 /* Socket is not locked. We are protected from async events
341 by poll logic and correct handling of state changes
342 made by another threads is impossible in any case.
350 * POLLHUP is certainly not done right. But poll() doesn't
351 * have a notion of HUP in just one direction, and for a
352 * socket the read side is more interesting.
354 * Some poll() documentation says that POLLHUP is incompatible
355 * with the POLLOUT/POLLWR flags, so somebody should check this
356 * all. But careful, it tends to be safer to return too many
357 * bits than too few, and you can easily break real applications
358 * if you don't tell them that something has hung up!
362 * Check number 1. POLLHUP is _UNMASKABLE_ event (see UNIX98 and
363 * our fs/select.c). It means that after we received EOF,
364 * poll always returns immediately, making impossible poll() on write()
365 * in state CLOSE_WAIT. One solution is evident --- to set POLLHUP
366 * if and only if shutdown has been made in both directions.
367 * Actually, it is interesting to look how Solaris and DUX
368 * solve this dilemma. I would prefer, if PULLHUP were maskable,
369 * then we could set it on SND_SHUTDOWN. BTW examples given
370 * in Stevens' books assume exactly this behaviour, it explains
371 * why PULLHUP is incompatible with POLLOUT. --ANK
373 * NOTE. Check for TCP_CLOSE is added. The goal is to prevent
374 * blocking on fresh not-connected or disconnected socket. --ANK
376 if (sk->sk_shutdown == SHUTDOWN_MASK || sk->sk_state == TCP_CLOSE)
378 if (sk->sk_shutdown & RCV_SHUTDOWN)
379 mask |= POLLIN | POLLRDNORM;
382 if ((1 << sk->sk_state) & ~(TCPF_SYN_SENT | TCPF_SYN_RECV)) {
383 /* Potential race condition. If read of tp below will
384 * escape above sk->sk_state, we can be illegally awaken
385 * in SYN_* states. */
386 if ((tp->rcv_nxt != tp->copied_seq) &&
387 (tp->urg_seq != tp->copied_seq ||
388 tp->rcv_nxt != tp->copied_seq + 1 ||
389 sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data))
390 mask |= POLLIN | POLLRDNORM;
392 if (!(sk->sk_shutdown & SEND_SHUTDOWN)) {
393 if (sk_stream_wspace(sk) >= sk_stream_min_wspace(sk)) {
394 mask |= POLLOUT | POLLWRNORM;
395 } else { /* send SIGIO later */
396 set_bit(SOCK_ASYNC_NOSPACE,
397 &sk->sk_socket->flags);
398 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
400 /* Race breaker. If space is freed after
401 * wspace test but before the flags are set,
402 * IO signal will be lost.
404 if (sk_stream_wspace(sk) >= sk_stream_min_wspace(sk))
405 mask |= POLLOUT | POLLWRNORM;
409 if (tp->urg_data & TCP_URG_VALID)
415 int tcp_ioctl(struct sock *sk, int cmd, unsigned long arg)
417 struct tcp_sock *tp = tcp_sk(sk);
422 if (sk->sk_state == TCP_LISTEN)
426 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
428 else if (sock_flag(sk, SOCK_URGINLINE) ||
430 before(tp->urg_seq, tp->copied_seq) ||
431 !before(tp->urg_seq, tp->rcv_nxt)) {
432 answ = tp->rcv_nxt - tp->copied_seq;
434 /* Subtract 1, if FIN is in queue. */
435 if (answ && !skb_queue_empty(&sk->sk_receive_queue))
437 ((struct sk_buff *)sk->sk_receive_queue.prev)->h.th->fin;
439 answ = tp->urg_seq - tp->copied_seq;
443 answ = tp->urg_data && tp->urg_seq == tp->copied_seq;
446 if (sk->sk_state == TCP_LISTEN)
449 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
452 answ = tp->write_seq - tp->snd_una;
458 return put_user(answ, (int __user *)arg);
462 int tcp_listen_start(struct sock *sk)
464 struct inet_sock *inet = inet_sk(sk);
465 struct tcp_sock *tp = tcp_sk(sk);
466 struct tcp_listen_opt *lopt;
468 sk->sk_max_ack_backlog = 0;
469 sk->sk_ack_backlog = 0;
470 tp->accept_queue = tp->accept_queue_tail = NULL;
471 rwlock_init(&tp->syn_wait_lock);
474 lopt = kmalloc(sizeof(struct tcp_listen_opt), GFP_KERNEL);
478 memset(lopt, 0, sizeof(struct tcp_listen_opt));
479 for (lopt->max_qlen_log = 6; ; lopt->max_qlen_log++)
480 if ((1 << lopt->max_qlen_log) >= sysctl_max_syn_backlog)
482 get_random_bytes(&lopt->hash_rnd, 4);
484 write_lock_bh(&tp->syn_wait_lock);
485 tp->listen_opt = lopt;
486 write_unlock_bh(&tp->syn_wait_lock);
488 /* There is race window here: we announce ourselves listening,
489 * but this transition is still not validated by get_port().
490 * It is OK, because this socket enters to hash table only
491 * after validation is complete.
493 sk->sk_state = TCP_LISTEN;
494 if (!sk->sk_prot->get_port(sk, inet->num)) {
495 inet->sport = htons(inet->num);
498 sk->sk_prot->hash(sk);
502 sk->sk_state = TCP_CLOSE;
503 write_lock_bh(&tp->syn_wait_lock);
504 tp->listen_opt = NULL;
505 write_unlock_bh(&tp->syn_wait_lock);
511 * This routine closes sockets which have been at least partially
512 * opened, but not yet accepted.
515 static void tcp_listen_stop (struct sock *sk)
517 struct tcp_sock *tp = tcp_sk(sk);
518 struct tcp_listen_opt *lopt = tp->listen_opt;
519 struct open_request *acc_req = tp->accept_queue;
520 struct open_request *req;
523 tcp_delete_keepalive_timer(sk);
525 /* make all the listen_opt local to us */
526 write_lock_bh(&tp->syn_wait_lock);
527 tp->listen_opt = NULL;
528 write_unlock_bh(&tp->syn_wait_lock);
530 tp->accept_queue_tail = NULL;
531 tp->accept_queue = NULL;
534 for (i = 0; i < TCP_SYNQ_HSIZE; i++) {
535 while ((req = lopt->syn_table[i]) != NULL) {
536 lopt->syn_table[i] = req->dl_next;
538 tcp_openreq_free(req);
540 /* Following specs, it would be better either to send FIN
541 * (and enter FIN-WAIT-1, it is normal close)
542 * or to send active reset (abort).
543 * Certainly, it is pretty dangerous while synflood, but it is
544 * bad justification for our negligence 8)
545 * To be honest, we are not able to make either
546 * of the variants now. --ANK
551 BUG_TRAP(!lopt->qlen);
555 while ((req = acc_req) != NULL) {
556 struct sock *child = req->sk;
558 acc_req = req->dl_next;
562 BUG_TRAP(!sock_owned_by_user(child));
565 tcp_disconnect(child, O_NONBLOCK);
569 atomic_inc(&tcp_orphan_count);
571 tcp_destroy_sock(child);
573 bh_unlock_sock(child);
577 sk_acceptq_removed(sk);
578 tcp_openreq_fastfree(req);
580 BUG_TRAP(!sk->sk_ack_backlog);
583 static inline void tcp_mark_push(struct tcp_sock *tp, struct sk_buff *skb)
585 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH;
586 tp->pushed_seq = tp->write_seq;
589 static inline int forced_push(struct tcp_sock *tp)
591 return after(tp->write_seq, tp->pushed_seq + (tp->max_window >> 1));
594 static inline void skb_entail(struct sock *sk, struct tcp_sock *tp,
598 TCP_SKB_CB(skb)->seq = tp->write_seq;
599 TCP_SKB_CB(skb)->end_seq = tp->write_seq;
600 TCP_SKB_CB(skb)->flags = TCPCB_FLAG_ACK;
601 TCP_SKB_CB(skb)->sacked = 0;
602 skb_header_release(skb);
603 __skb_queue_tail(&sk->sk_write_queue, skb);
604 sk_charge_skb(sk, skb);
605 if (!sk->sk_send_head)
606 sk->sk_send_head = skb;
607 else if (tp->nonagle&TCP_NAGLE_PUSH)
608 tp->nonagle &= ~TCP_NAGLE_PUSH;
611 static inline void tcp_mark_urg(struct tcp_sock *tp, int flags,
614 if (flags & MSG_OOB) {
616 tp->snd_up = tp->write_seq;
617 TCP_SKB_CB(skb)->sacked |= TCPCB_URG;
621 static inline void tcp_push(struct sock *sk, struct tcp_sock *tp, int flags,
622 int mss_now, int nonagle)
624 if (sk->sk_send_head) {
625 struct sk_buff *skb = sk->sk_write_queue.prev;
626 if (!(flags & MSG_MORE) || forced_push(tp))
627 tcp_mark_push(tp, skb);
628 tcp_mark_urg(tp, flags, skb);
629 __tcp_push_pending_frames(sk, tp, mss_now,
630 (flags & MSG_MORE) ? TCP_NAGLE_CORK : nonagle);
634 static ssize_t do_tcp_sendpages(struct sock *sk, struct page **pages, int poffset,
635 size_t psize, int flags)
637 struct tcp_sock *tp = tcp_sk(sk);
641 long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
643 /* Wait for a connection to finish. */
644 if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
645 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
648 clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
650 mss_now = tcp_current_mss(sk, !(flags&MSG_OOB));
654 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
658 struct sk_buff *skb = sk->sk_write_queue.prev;
659 struct page *page = pages[poffset / PAGE_SIZE];
660 int copy, i, can_coalesce;
661 int offset = poffset % PAGE_SIZE;
662 int size = min_t(size_t, psize, PAGE_SIZE - offset);
664 if (!sk->sk_send_head || (copy = mss_now - skb->len) <= 0) {
666 if (!sk_stream_memory_free(sk))
667 goto wait_for_sndbuf;
669 skb = sk_stream_alloc_pskb(sk, 0, 0,
672 goto wait_for_memory;
674 skb_entail(sk, tp, skb);
681 i = skb_shinfo(skb)->nr_frags;
682 can_coalesce = skb_can_coalesce(skb, i, page, offset);
683 if (!can_coalesce && i >= MAX_SKB_FRAGS) {
684 tcp_mark_push(tp, skb);
687 if (sk->sk_forward_alloc < copy &&
688 !sk_stream_mem_schedule(sk, copy, 0))
689 goto wait_for_memory;
692 skb_shinfo(skb)->frags[i - 1].size += copy;
695 skb_fill_page_desc(skb, i, page, offset, copy);
699 skb->data_len += copy;
700 skb->truesize += copy;
701 sk->sk_wmem_queued += copy;
702 sk->sk_forward_alloc -= copy;
703 skb->ip_summed = CHECKSUM_HW;
704 tp->write_seq += copy;
705 TCP_SKB_CB(skb)->end_seq += copy;
706 skb_shinfo(skb)->tso_segs = 0;
709 TCP_SKB_CB(skb)->flags &= ~TCPCB_FLAG_PSH;
713 if (!(psize -= copy))
716 if (skb->len != mss_now || (flags & MSG_OOB))
719 if (forced_push(tp)) {
720 tcp_mark_push(tp, skb);
721 __tcp_push_pending_frames(sk, tp, mss_now, TCP_NAGLE_PUSH);
722 } else if (skb == sk->sk_send_head)
723 tcp_push_one(sk, mss_now);
727 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
730 tcp_push(sk, tp, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH);
732 if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
735 mss_now = tcp_current_mss(sk, !(flags&MSG_OOB));
740 tcp_push(sk, tp, flags, mss_now, tp->nonagle);
747 return sk_stream_error(sk, flags, err);
750 ssize_t tcp_sendpage(struct socket *sock, struct page *page, int offset,
751 size_t size, int flags)
754 struct sock *sk = sock->sk;
756 #define TCP_ZC_CSUM_FLAGS (NETIF_F_IP_CSUM | NETIF_F_NO_CSUM | NETIF_F_HW_CSUM)
758 if (!(sk->sk_route_caps & NETIF_F_SG) ||
759 !(sk->sk_route_caps & TCP_ZC_CSUM_FLAGS))
760 return sock_no_sendpage(sock, page, offset, size, flags);
762 #undef TCP_ZC_CSUM_FLAGS
766 res = do_tcp_sendpages(sk, &page, offset, size, flags);
772 #define TCP_PAGE(sk) (sk->sk_sndmsg_page)
773 #define TCP_OFF(sk) (sk->sk_sndmsg_off)
775 static inline int select_size(struct sock *sk, struct tcp_sock *tp)
777 int tmp = tp->mss_cache_std;
779 if (sk->sk_route_caps & NETIF_F_SG) {
780 int pgbreak = SKB_MAX_HEAD(MAX_TCP_HEADER);
782 if (tmp >= pgbreak &&
783 tmp <= pgbreak + (MAX_SKB_FRAGS - 1) * PAGE_SIZE)
789 int tcp_sendmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg,
793 struct tcp_sock *tp = tcp_sk(sk);
803 flags = msg->msg_flags;
804 timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
806 /* Wait for a connection to finish. */
807 if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
808 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
811 /* This should be in poll */
812 clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
814 mss_now = tcp_current_mss(sk, !(flags&MSG_OOB));
816 /* Ok commence sending. */
817 iovlen = msg->msg_iovlen;
822 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
825 while (--iovlen >= 0) {
826 int seglen = iov->iov_len;
827 unsigned char __user *from = iov->iov_base;
834 skb = sk->sk_write_queue.prev;
836 if (!sk->sk_send_head ||
837 (copy = mss_now - skb->len) <= 0) {
840 /* Allocate new segment. If the interface is SG,
841 * allocate skb fitting to single page.
843 if (!sk_stream_memory_free(sk))
844 goto wait_for_sndbuf;
846 skb = sk_stream_alloc_pskb(sk, select_size(sk, tp),
847 0, sk->sk_allocation);
849 goto wait_for_memory;
852 * Check whether we can use HW checksum.
854 if (sk->sk_route_caps &
855 (NETIF_F_IP_CSUM | NETIF_F_NO_CSUM |
857 skb->ip_summed = CHECKSUM_HW;
859 skb_entail(sk, tp, skb);
863 /* Try to append data to the end of skb. */
867 /* Where to copy to? */
868 if (skb_tailroom(skb) > 0) {
869 /* We have some space in skb head. Superb! */
870 if (copy > skb_tailroom(skb))
871 copy = skb_tailroom(skb);
872 if ((err = skb_add_data(skb, from, copy)) != 0)
876 int i = skb_shinfo(skb)->nr_frags;
877 struct page *page = TCP_PAGE(sk);
878 int off = TCP_OFF(sk);
880 if (skb_can_coalesce(skb, i, page, off) &&
882 /* We can extend the last page
885 } else if (i == MAX_SKB_FRAGS ||
887 !(sk->sk_route_caps & NETIF_F_SG))) {
888 /* Need to add new fragment and cannot
889 * do this because interface is non-SG,
890 * or because all the page slots are
892 tcp_mark_push(tp, skb);
895 /* If page is cached, align
896 * offset to L1 cache boundary
898 off = (off + L1_CACHE_BYTES - 1) &
899 ~(L1_CACHE_BYTES - 1);
900 if (off == PAGE_SIZE) {
902 TCP_PAGE(sk) = page = NULL;
907 /* Allocate new cache page. */
908 if (!(page = sk_stream_alloc_page(sk)))
909 goto wait_for_memory;
913 if (copy > PAGE_SIZE - off)
914 copy = PAGE_SIZE - off;
916 /* Time to copy data. We are close to
918 err = skb_copy_to_page(sk, from, skb, page,
921 /* If this page was new, give it to the
922 * socket so it does not get leaked.
931 /* Update the skb. */
933 skb_shinfo(skb)->frags[i - 1].size +=
936 skb_fill_page_desc(skb, i, page, off, copy);
939 } else if (off + copy < PAGE_SIZE) {
945 TCP_OFF(sk) = off + copy;
949 TCP_SKB_CB(skb)->flags &= ~TCPCB_FLAG_PSH;
951 tp->write_seq += copy;
952 TCP_SKB_CB(skb)->end_seq += copy;
953 skb_shinfo(skb)->tso_segs = 0;
957 if ((seglen -= copy) == 0 && iovlen == 0)
960 if (skb->len != mss_now || (flags & MSG_OOB))
963 if (forced_push(tp)) {
964 tcp_mark_push(tp, skb);
965 __tcp_push_pending_frames(sk, tp, mss_now, TCP_NAGLE_PUSH);
966 } else if (skb == sk->sk_send_head)
967 tcp_push_one(sk, mss_now);
971 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
974 tcp_push(sk, tp, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH);
976 if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
979 mss_now = tcp_current_mss(sk, !(flags&MSG_OOB));
985 tcp_push(sk, tp, flags, mss_now, tp->nonagle);
992 if (sk->sk_send_head == skb)
993 sk->sk_send_head = NULL;
994 __skb_unlink(skb, skb->list);
995 sk_stream_free_skb(sk, skb);
1002 err = sk_stream_error(sk, flags, err);
1003 TCP_CHECK_TIMER(sk);
1009 * Handle reading urgent data. BSD has very simple semantics for
1010 * this, no blocking and very strange errors 8)
1013 static int tcp_recv_urg(struct sock *sk, long timeo,
1014 struct msghdr *msg, int len, int flags,
1017 struct tcp_sock *tp = tcp_sk(sk);
1019 /* No URG data to read. */
1020 if (sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data ||
1021 tp->urg_data == TCP_URG_READ)
1022 return -EINVAL; /* Yes this is right ! */
1024 if (sk->sk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DONE))
1027 if (tp->urg_data & TCP_URG_VALID) {
1029 char c = tp->urg_data;
1031 if (!(flags & MSG_PEEK))
1032 tp->urg_data = TCP_URG_READ;
1034 /* Read urgent data. */
1035 msg->msg_flags |= MSG_OOB;
1038 if (!(flags & MSG_TRUNC))
1039 err = memcpy_toiovec(msg->msg_iov, &c, 1);
1042 msg->msg_flags |= MSG_TRUNC;
1044 return err ? -EFAULT : len;
1047 if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
1050 /* Fixed the recv(..., MSG_OOB) behaviour. BSD docs and
1051 * the available implementations agree in this case:
1052 * this call should never block, independent of the
1053 * blocking state of the socket.
1054 * Mike <pall@rz.uni-karlsruhe.de>
1059 /* Clean up the receive buffer for full frames taken by the user,
1060 * then send an ACK if necessary. COPIED is the number of bytes
1061 * tcp_recvmsg has given to the user so far, it speeds up the
1062 * calculation of whether or not we must ACK for the sake of
1065 void cleanup_rbuf(struct sock *sk, int copied)
1067 struct tcp_sock *tp = tcp_sk(sk);
1068 int time_to_ack = 0;
1071 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
1073 BUG_TRAP(!skb || before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq));
1076 if (tcp_ack_scheduled(tp)) {
1077 /* Delayed ACKs frequently hit locked sockets during bulk
1079 if (tp->ack.blocked ||
1080 /* Once-per-two-segments ACK was not sent by tcp_input.c */
1081 tp->rcv_nxt - tp->rcv_wup > tp->ack.rcv_mss ||
1083 * If this read emptied read buffer, we send ACK, if
1084 * connection is not bidirectional, user drained
1085 * receive buffer and there was a small segment
1088 (copied > 0 && (tp->ack.pending & TCP_ACK_PUSHED) &&
1089 !tp->ack.pingpong && !atomic_read(&sk->sk_rmem_alloc)))
1093 /* We send an ACK if we can now advertise a non-zero window
1094 * which has been raised "significantly".
1096 * Even if window raised up to infinity, do not send window open ACK
1097 * in states, where we will not receive more. It is useless.
1099 if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) {
1100 __u32 rcv_window_now = tcp_receive_window(tp);
1102 /* Optimize, __tcp_select_window() is not cheap. */
1103 if (2*rcv_window_now <= tp->window_clamp) {
1104 __u32 new_window = __tcp_select_window(sk);
1106 /* Send ACK now, if this read freed lots of space
1107 * in our buffer. Certainly, new_window is new window.
1108 * We can advertise it now, if it is not less than current one.
1109 * "Lots" means "at least twice" here.
1111 if (new_window && new_window >= 2 * rcv_window_now)
1119 static void tcp_prequeue_process(struct sock *sk)
1121 struct sk_buff *skb;
1122 struct tcp_sock *tp = tcp_sk(sk);
1124 NET_ADD_STATS_USER(LINUX_MIB_TCPPREQUEUED, skb_queue_len(&tp->ucopy.prequeue));
1126 /* RX process wants to run with disabled BHs, though it is not
1129 while ((skb = __skb_dequeue(&tp->ucopy.prequeue)) != NULL)
1130 sk->sk_backlog_rcv(sk, skb);
1133 /* Clear memory counter. */
1134 tp->ucopy.memory = 0;
1137 static inline struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off)
1139 struct sk_buff *skb;
1142 skb_queue_walk(&sk->sk_receive_queue, skb) {
1143 offset = seq - TCP_SKB_CB(skb)->seq;
1146 if (offset < skb->len || skb->h.th->fin) {
1155 * This routine provides an alternative to tcp_recvmsg() for routines
1156 * that would like to handle copying from skbuffs directly in 'sendfile'
1159 * - It is assumed that the socket was locked by the caller.
1160 * - The routine does not block.
1161 * - At present, there is no support for reading OOB data
1162 * or for 'peeking' the socket using this routine
1163 * (although both would be easy to implement).
1165 int tcp_read_sock(struct sock *sk, read_descriptor_t *desc,
1166 sk_read_actor_t recv_actor)
1168 struct sk_buff *skb;
1169 struct tcp_sock *tp = tcp_sk(sk);
1170 u32 seq = tp->copied_seq;
1174 if (sk->sk_state == TCP_LISTEN)
1176 while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1177 if (offset < skb->len) {
1180 len = skb->len - offset;
1181 /* Stop reading if we hit a patch of urgent data */
1183 u32 urg_offset = tp->urg_seq - seq;
1184 if (urg_offset < len)
1189 used = recv_actor(desc, skb, offset, len);
1195 if (offset != skb->len)
1198 if (skb->h.th->fin) {
1199 sk_eat_skb(sk, skb);
1203 sk_eat_skb(sk, skb);
1207 tp->copied_seq = seq;
1209 tcp_rcv_space_adjust(sk);
1211 /* Clean up data we have read: This will do ACK frames. */
1213 cleanup_rbuf(sk, copied);
1218 * This routine copies from a sock struct into the user buffer.
1220 * Technical note: in 2.3 we work on _locked_ socket, so that
1221 * tricks with *seq access order and skb->users are not required.
1222 * Probably, code can be easily improved even more.
1225 int tcp_recvmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg,
1226 size_t len, int nonblock, int flags, int *addr_len)
1228 struct tcp_sock *tp = tcp_sk(sk);
1234 int target; /* Read at least this many bytes */
1236 struct task_struct *user_recv = NULL;
1240 TCP_CHECK_TIMER(sk);
1243 if (sk->sk_state == TCP_LISTEN)
1246 timeo = sock_rcvtimeo(sk, nonblock);
1248 /* Urgent data needs to be handled specially. */
1249 if (flags & MSG_OOB)
1252 seq = &tp->copied_seq;
1253 if (flags & MSG_PEEK) {
1254 peek_seq = tp->copied_seq;
1258 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
1261 struct sk_buff *skb;
1264 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
1265 if (tp->urg_data && tp->urg_seq == *seq) {
1268 if (signal_pending(current)) {
1269 copied = timeo ? sock_intr_errno(timeo) : -EAGAIN;
1274 /* Next get a buffer. */
1276 skb = skb_peek(&sk->sk_receive_queue);
1281 /* Now that we have two receive queues this
1284 if (before(*seq, TCP_SKB_CB(skb)->seq)) {
1285 printk(KERN_INFO "recvmsg bug: copied %X "
1286 "seq %X\n", *seq, TCP_SKB_CB(skb)->seq);
1289 offset = *seq - TCP_SKB_CB(skb)->seq;
1292 if (offset < skb->len)
1296 BUG_TRAP(flags & MSG_PEEK);
1298 } while (skb != (struct sk_buff *)&sk->sk_receive_queue);
1300 /* Well, if we have backlog, try to process it now yet. */
1302 if (copied >= target && !sk->sk_backlog.tail)
1307 sk->sk_state == TCP_CLOSE ||
1308 (sk->sk_shutdown & RCV_SHUTDOWN) ||
1310 signal_pending(current) ||
1314 if (sock_flag(sk, SOCK_DONE))
1318 copied = sock_error(sk);
1322 if (sk->sk_shutdown & RCV_SHUTDOWN)
1325 if (sk->sk_state == TCP_CLOSE) {
1326 if (!sock_flag(sk, SOCK_DONE)) {
1327 /* This occurs when user tries to read
1328 * from never connected socket.
1341 if (signal_pending(current)) {
1342 copied = sock_intr_errno(timeo);
1347 cleanup_rbuf(sk, copied);
1349 if (tp->ucopy.task == user_recv) {
1350 /* Install new reader */
1351 if (!user_recv && !(flags & (MSG_TRUNC | MSG_PEEK))) {
1352 user_recv = current;
1353 tp->ucopy.task = user_recv;
1354 tp->ucopy.iov = msg->msg_iov;
1357 tp->ucopy.len = len;
1359 BUG_TRAP(tp->copied_seq == tp->rcv_nxt ||
1360 (flags & (MSG_PEEK | MSG_TRUNC)));
1362 /* Ugly... If prequeue is not empty, we have to
1363 * process it before releasing socket, otherwise
1364 * order will be broken at second iteration.
1365 * More elegant solution is required!!!
1367 * Look: we have the following (pseudo)queues:
1369 * 1. packets in flight
1374 * Each queue can be processed only if the next ones
1375 * are empty. At this point we have empty receive_queue.
1376 * But prequeue _can_ be not empty after 2nd iteration,
1377 * when we jumped to start of loop because backlog
1378 * processing added something to receive_queue.
1379 * We cannot release_sock(), because backlog contains
1380 * packets arrived _after_ prequeued ones.
1382 * Shortly, algorithm is clear --- to process all
1383 * the queues in order. We could make it more directly,
1384 * requeueing packets from backlog to prequeue, if
1385 * is not empty. It is more elegant, but eats cycles,
1388 if (skb_queue_len(&tp->ucopy.prequeue))
1391 /* __ Set realtime policy in scheduler __ */
1394 if (copied >= target) {
1395 /* Do not sleep, just process backlog. */
1399 sk_wait_data(sk, &timeo);
1404 /* __ Restore normal policy in scheduler __ */
1406 if ((chunk = len - tp->ucopy.len) != 0) {
1407 NET_ADD_STATS_USER(LINUX_MIB_TCPDIRECTCOPYFROMBACKLOG, chunk);
1412 if (tp->rcv_nxt == tp->copied_seq &&
1413 skb_queue_len(&tp->ucopy.prequeue)) {
1415 tcp_prequeue_process(sk);
1417 if ((chunk = len - tp->ucopy.len) != 0) {
1418 NET_ADD_STATS_USER(LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1424 if ((flags & MSG_PEEK) && peek_seq != tp->copied_seq) {
1425 if (net_ratelimit())
1426 printk(KERN_DEBUG "TCP(%s:%d): Application bug, race in MSG_PEEK.\n",
1427 current->comm, current->pid);
1428 peek_seq = tp->copied_seq;
1433 /* Ok so how much can we use? */
1434 used = skb->len - offset;
1438 /* Do we have urgent data here? */
1440 u32 urg_offset = tp->urg_seq - *seq;
1441 if (urg_offset < used) {
1443 if (!sock_flag(sk, SOCK_URGINLINE)) {
1455 if (!(flags & MSG_TRUNC)) {
1456 err = skb_copy_datagram_iovec(skb, offset,
1457 msg->msg_iov, used);
1459 /* Exception. Bailout! */
1470 tcp_rcv_space_adjust(sk);
1473 if (tp->urg_data && after(tp->copied_seq, tp->urg_seq)) {
1475 tcp_fast_path_check(sk, tp);
1477 if (used + offset < skb->len)
1482 if (!(flags & MSG_PEEK))
1483 sk_eat_skb(sk, skb);
1487 /* Process the FIN. */
1489 if (!(flags & MSG_PEEK))
1490 sk_eat_skb(sk, skb);
1495 if (skb_queue_len(&tp->ucopy.prequeue)) {
1498 tp->ucopy.len = copied > 0 ? len : 0;
1500 tcp_prequeue_process(sk);
1502 if (copied > 0 && (chunk = len - tp->ucopy.len) != 0) {
1503 NET_ADD_STATS_USER(LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1509 tp->ucopy.task = NULL;
1513 /* According to UNIX98, msg_name/msg_namelen are ignored
1514 * on connected socket. I was just happy when found this 8) --ANK
1517 /* Clean up data we have read: This will do ACK frames. */
1518 cleanup_rbuf(sk, copied);
1520 TCP_CHECK_TIMER(sk);
1525 TCP_CHECK_TIMER(sk);
1530 err = tcp_recv_urg(sk, timeo, msg, len, flags, addr_len);
1535 * State processing on a close. This implements the state shift for
1536 * sending our FIN frame. Note that we only send a FIN for some
1537 * states. A shutdown() may have already sent the FIN, or we may be
1541 static unsigned char new_state[16] = {
1542 /* current state: new state: action: */
1543 /* (Invalid) */ TCP_CLOSE,
1544 /* TCP_ESTABLISHED */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1545 /* TCP_SYN_SENT */ TCP_CLOSE,
1546 /* TCP_SYN_RECV */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1547 /* TCP_FIN_WAIT1 */ TCP_FIN_WAIT1,
1548 /* TCP_FIN_WAIT2 */ TCP_FIN_WAIT2,
1549 /* TCP_TIME_WAIT */ TCP_CLOSE,
1550 /* TCP_CLOSE */ TCP_CLOSE,
1551 /* TCP_CLOSE_WAIT */ TCP_LAST_ACK | TCP_ACTION_FIN,
1552 /* TCP_LAST_ACK */ TCP_LAST_ACK,
1553 /* TCP_LISTEN */ TCP_CLOSE,
1554 /* TCP_CLOSING */ TCP_CLOSING,
1557 static int tcp_close_state(struct sock *sk)
1559 int next = (int)new_state[sk->sk_state];
1560 int ns = next & TCP_STATE_MASK;
1562 tcp_set_state(sk, ns);
1564 return next & TCP_ACTION_FIN;
1568 * Shutdown the sending side of a connection. Much like close except
1569 * that we don't receive shut down or set_sock_flag(sk, SOCK_DEAD).
1572 void tcp_shutdown(struct sock *sk, int how)
1574 /* We need to grab some memory, and put together a FIN,
1575 * and then put it into the queue to be sent.
1576 * Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
1578 if (!(how & SEND_SHUTDOWN))
1581 /* If we've already sent a FIN, or it's a closed state, skip this. */
1582 if ((1 << sk->sk_state) &
1583 (TCPF_ESTABLISHED | TCPF_SYN_SENT |
1584 TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) {
1585 /* Clear out any half completed packets. FIN if needed. */
1586 if (tcp_close_state(sk))
1592 * At this point, there should be no process reference to this
1593 * socket, and thus no user references at all. Therefore we
1594 * can assume the socket waitqueue is inactive and nobody will
1595 * try to jump onto it.
1597 void tcp_destroy_sock(struct sock *sk)
1599 BUG_TRAP(sk->sk_state == TCP_CLOSE);
1600 BUG_TRAP(sock_flag(sk, SOCK_DEAD));
1602 /* It cannot be in hash table! */
1603 BUG_TRAP(sk_unhashed(sk));
1605 /* If it has not 0 inet_sk(sk)->num, it must be bound */
1606 BUG_TRAP(!inet_sk(sk)->num || tcp_sk(sk)->bind_hash);
1608 sk->sk_prot->destroy(sk);
1610 sk_stream_kill_queues(sk);
1612 xfrm_sk_free_policy(sk);
1614 #ifdef INET_REFCNT_DEBUG
1615 if (atomic_read(&sk->sk_refcnt) != 1) {
1616 printk(KERN_DEBUG "Destruction TCP %p delayed, c=%d\n",
1617 sk, atomic_read(&sk->sk_refcnt));
1621 atomic_dec(&tcp_orphan_count);
1625 void tcp_close(struct sock *sk, long timeout)
1627 struct sk_buff *skb;
1628 int data_was_unread = 0;
1631 sk->sk_shutdown = SHUTDOWN_MASK;
1633 if (sk->sk_state == TCP_LISTEN) {
1634 tcp_set_state(sk, TCP_CLOSE);
1637 tcp_listen_stop(sk);
1639 goto adjudge_to_death;
1642 /* We need to flush the recv. buffs. We do this only on the
1643 * descriptor close, not protocol-sourced closes, because the
1644 * reader process may not have drained the data yet!
1646 while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) {
1647 u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq -
1649 data_was_unread += len;
1653 sk_stream_mem_reclaim(sk);
1655 /* As outlined in draft-ietf-tcpimpl-prob-03.txt, section
1656 * 3.10, we send a RST here because data was lost. To
1657 * witness the awful effects of the old behavior of always
1658 * doing a FIN, run an older 2.1.x kernel or 2.0.x, start
1659 * a bulk GET in an FTP client, suspend the process, wait
1660 * for the client to advertise a zero window, then kill -9
1661 * the FTP client, wheee... Note: timeout is always zero
1664 if (data_was_unread) {
1665 /* Unread data was tossed, zap the connection. */
1666 NET_INC_STATS_USER(LINUX_MIB_TCPABORTONCLOSE);
1667 tcp_set_state(sk, TCP_CLOSE);
1668 tcp_send_active_reset(sk, GFP_KERNEL);
1669 } else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) {
1670 /* Check zero linger _after_ checking for unread data. */
1671 sk->sk_prot->disconnect(sk, 0);
1672 NET_INC_STATS_USER(LINUX_MIB_TCPABORTONDATA);
1673 } else if (tcp_close_state(sk)) {
1674 /* We FIN if the application ate all the data before
1675 * zapping the connection.
1678 /* RED-PEN. Formally speaking, we have broken TCP state
1679 * machine. State transitions:
1681 * TCP_ESTABLISHED -> TCP_FIN_WAIT1
1682 * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible)
1683 * TCP_CLOSE_WAIT -> TCP_LAST_ACK
1685 * are legal only when FIN has been sent (i.e. in window),
1686 * rather than queued out of window. Purists blame.
1688 * F.e. "RFC state" is ESTABLISHED,
1689 * if Linux state is FIN-WAIT-1, but FIN is still not sent.
1691 * The visible declinations are that sometimes
1692 * we enter time-wait state, when it is not required really
1693 * (harmless), do not send active resets, when they are
1694 * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
1695 * they look as CLOSING or LAST_ACK for Linux)
1696 * Probably, I missed some more holelets.
1702 sk_stream_wait_close(sk, timeout);
1705 /* It is the last release_sock in its life. It will remove backlog. */
1709 /* Now socket is owned by kernel and we acquire BH lock
1710 to finish close. No need to check for user refs.
1714 BUG_TRAP(!sock_owned_by_user(sk));
1719 /* This is a (useful) BSD violating of the RFC. There is a
1720 * problem with TCP as specified in that the other end could
1721 * keep a socket open forever with no application left this end.
1722 * We use a 3 minute timeout (about the same as BSD) then kill
1723 * our end. If they send after that then tough - BUT: long enough
1724 * that we won't make the old 4*rto = almost no time - whoops
1727 * Nope, it was not mistake. It is really desired behaviour
1728 * f.e. on http servers, when such sockets are useless, but
1729 * consume significant resources. Let's do it with special
1730 * linger2 option. --ANK
1733 if (sk->sk_state == TCP_FIN_WAIT2) {
1734 struct tcp_sock *tp = tcp_sk(sk);
1735 if (tp->linger2 < 0) {
1736 tcp_set_state(sk, TCP_CLOSE);
1737 tcp_send_active_reset(sk, GFP_ATOMIC);
1738 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONLINGER);
1740 int tmo = tcp_fin_time(tp);
1742 if (tmo > TCP_TIMEWAIT_LEN) {
1743 tcp_reset_keepalive_timer(sk, tcp_fin_time(tp));
1745 atomic_inc(&tcp_orphan_count);
1746 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
1751 if (sk->sk_state != TCP_CLOSE) {
1752 sk_stream_mem_reclaim(sk);
1753 if (atomic_read(&tcp_orphan_count) > sysctl_tcp_max_orphans ||
1754 (sk->sk_wmem_queued > SOCK_MIN_SNDBUF &&
1755 atomic_read(&tcp_memory_allocated) > sysctl_tcp_mem[2])) {
1756 if (net_ratelimit())
1757 printk(KERN_INFO "TCP: too many of orphaned "
1759 tcp_set_state(sk, TCP_CLOSE);
1760 tcp_send_active_reset(sk, GFP_ATOMIC);
1761 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONMEMORY);
1764 atomic_inc(&tcp_orphan_count);
1766 if (sk->sk_state == TCP_CLOSE)
1767 tcp_destroy_sock(sk);
1768 /* Otherwise, socket is reprieved until protocol close. */
1776 /* These states need RST on ABORT according to RFC793 */
1778 static inline int tcp_need_reset(int state)
1780 return (1 << state) &
1781 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 |
1782 TCPF_FIN_WAIT2 | TCPF_SYN_RECV);
1785 int tcp_disconnect(struct sock *sk, int flags)
1787 struct inet_sock *inet = inet_sk(sk);
1788 struct tcp_sock *tp = tcp_sk(sk);
1790 int old_state = sk->sk_state;
1792 if (old_state != TCP_CLOSE)
1793 tcp_set_state(sk, TCP_CLOSE);
1795 /* ABORT function of RFC793 */
1796 if (old_state == TCP_LISTEN) {
1797 tcp_listen_stop(sk);
1798 } else if (tcp_need_reset(old_state) ||
1799 (tp->snd_nxt != tp->write_seq &&
1800 (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK))) {
1801 /* The last check adjusts for discrepance of Linux wrt. RFC
1804 tcp_send_active_reset(sk, gfp_any());
1805 sk->sk_err = ECONNRESET;
1806 } else if (old_state == TCP_SYN_SENT)
1807 sk->sk_err = ECONNRESET;
1809 tcp_clear_xmit_timers(sk);
1810 __skb_queue_purge(&sk->sk_receive_queue);
1811 sk_stream_writequeue_purge(sk);
1812 __skb_queue_purge(&tp->out_of_order_queue);
1816 if (!(sk->sk_userlocks & SOCK_BINDADDR_LOCK))
1817 inet_reset_saddr(sk);
1819 sk->sk_shutdown = 0;
1820 sock_reset_flag(sk, SOCK_DONE);
1822 if ((tp->write_seq += tp->max_window + 2) == 0)
1827 tp->packets_out = 0;
1828 tp->snd_ssthresh = 0x7fffffff;
1829 tp->snd_cwnd_cnt = 0;
1830 tcp_set_ca_state(tp, TCP_CA_Open);
1831 tcp_clear_retrans(tp);
1832 tcp_delack_init(tp);
1833 sk->sk_send_head = NULL;
1834 tp->rx_opt.saw_tstamp = 0;
1835 tcp_sack_reset(&tp->rx_opt);
1838 BUG_TRAP(!inet->num || tp->bind_hash);
1840 sk->sk_error_report(sk);
1845 * Wait for an incoming connection, avoid race
1846 * conditions. This must be called with the socket locked.
1848 static int wait_for_connect(struct sock *sk, long timeo)
1850 struct tcp_sock *tp = tcp_sk(sk);
1855 * True wake-one mechanism for incoming connections: only
1856 * one process gets woken up, not the 'whole herd'.
1857 * Since we do not 'race & poll' for established sockets
1858 * anymore, the common case will execute the loop only once.
1860 * Subtle issue: "add_wait_queue_exclusive()" will be added
1861 * after any current non-exclusive waiters, and we know that
1862 * it will always _stay_ after any new non-exclusive waiters
1863 * because all non-exclusive waiters are added at the
1864 * beginning of the wait-queue. As such, it's ok to "drop"
1865 * our exclusiveness temporarily when we get woken up without
1866 * having to remove and re-insert us on the wait queue.
1869 prepare_to_wait_exclusive(sk->sk_sleep, &wait,
1870 TASK_INTERRUPTIBLE);
1872 if (!tp->accept_queue)
1873 timeo = schedule_timeout(timeo);
1876 if (tp->accept_queue)
1879 if (sk->sk_state != TCP_LISTEN)
1881 err = sock_intr_errno(timeo);
1882 if (signal_pending(current))
1888 finish_wait(sk->sk_sleep, &wait);
1893 * This will accept the next outstanding connection.
1896 struct sock *tcp_accept(struct sock *sk, int flags, int *err)
1898 struct tcp_sock *tp = tcp_sk(sk);
1899 struct open_request *req;
1905 /* We need to make sure that this socket is listening,
1906 * and that it has something pending.
1909 if (sk->sk_state != TCP_LISTEN)
1912 /* Find already established connection */
1913 if (!tp->accept_queue) {
1914 long timeo = sock_rcvtimeo(sk, flags & O_NONBLOCK);
1915 /* If this is a non blocking socket don't sleep */
1920 error = wait_for_connect(sk, timeo);
1925 req = tp->accept_queue;
1926 if ((tp->accept_queue = req->dl_next) == NULL)
1927 tp->accept_queue_tail = NULL;
1929 sk_acceptq_removed(sk);
1930 tcp_openreq_fastfree(req);
1931 BUG_TRAP(newsk->sk_state != TCP_SYN_RECV);
1943 * Socket option code for TCP.
1945 int tcp_setsockopt(struct sock *sk, int level, int optname, char __user *optval,
1948 struct tcp_sock *tp = tcp_sk(sk);
1952 if (level != SOL_TCP)
1953 return tp->af_specific->setsockopt(sk, level, optname,
1956 if (optlen < sizeof(int))
1959 if (get_user(val, (int __user *)optval))
1966 /* Values greater than interface MTU won't take effect. However
1967 * at the point when this call is done we typically don't yet
1968 * know which interface is going to be used */
1969 if (val < 8 || val > MAX_TCP_WINDOW) {
1973 tp->rx_opt.user_mss = val;
1978 /* TCP_NODELAY is weaker than TCP_CORK, so that
1979 * this option on corked socket is remembered, but
1980 * it is not activated until cork is cleared.
1982 * However, when TCP_NODELAY is set we make
1983 * an explicit push, which overrides even TCP_CORK
1984 * for currently queued segments.
1986 tp->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH;
1987 tcp_push_pending_frames(sk, tp);
1989 tp->nonagle &= ~TCP_NAGLE_OFF;
1994 /* When set indicates to always queue non-full frames.
1995 * Later the user clears this option and we transmit
1996 * any pending partial frames in the queue. This is
1997 * meant to be used alongside sendfile() to get properly
1998 * filled frames when the user (for example) must write
1999 * out headers with a write() call first and then use
2000 * sendfile to send out the data parts.
2002 * TCP_CORK can be set together with TCP_NODELAY and it is
2003 * stronger than TCP_NODELAY.
2006 tp->nonagle |= TCP_NAGLE_CORK;
2008 tp->nonagle &= ~TCP_NAGLE_CORK;
2009 if (tp->nonagle&TCP_NAGLE_OFF)
2010 tp->nonagle |= TCP_NAGLE_PUSH;
2011 tcp_push_pending_frames(sk, tp);
2016 if (val < 1 || val > MAX_TCP_KEEPIDLE)
2019 tp->keepalive_time = val * HZ;
2020 if (sock_flag(sk, SOCK_KEEPOPEN) &&
2021 !((1 << sk->sk_state) &
2022 (TCPF_CLOSE | TCPF_LISTEN))) {
2023 __u32 elapsed = tcp_time_stamp - tp->rcv_tstamp;
2024 if (tp->keepalive_time > elapsed)
2025 elapsed = tp->keepalive_time - elapsed;
2028 tcp_reset_keepalive_timer(sk, elapsed);
2033 if (val < 1 || val > MAX_TCP_KEEPINTVL)
2036 tp->keepalive_intvl = val * HZ;
2039 if (val < 1 || val > MAX_TCP_KEEPCNT)
2042 tp->keepalive_probes = val;
2045 if (val < 1 || val > MAX_TCP_SYNCNT)
2048 tp->syn_retries = val;
2054 else if (val > sysctl_tcp_fin_timeout / HZ)
2057 tp->linger2 = val * HZ;
2060 case TCP_DEFER_ACCEPT:
2061 tp->defer_accept = 0;
2063 /* Translate value in seconds to number of
2065 while (tp->defer_accept < 32 &&
2066 val > ((TCP_TIMEOUT_INIT / HZ) <<
2073 case TCP_WINDOW_CLAMP:
2075 if (sk->sk_state != TCP_CLOSE) {
2079 tp->window_clamp = 0;
2081 tp->window_clamp = val < SOCK_MIN_RCVBUF / 2 ?
2082 SOCK_MIN_RCVBUF / 2 : val;
2087 tp->ack.pingpong = 1;
2089 tp->ack.pingpong = 0;
2090 if ((1 << sk->sk_state) &
2091 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) &&
2092 tcp_ack_scheduled(tp)) {
2093 tp->ack.pending |= TCP_ACK_PUSHED;
2094 cleanup_rbuf(sk, 1);
2096 tp->ack.pingpong = 1;
2109 /* Return information about state of tcp endpoint in API format. */
2110 void tcp_get_info(struct sock *sk, struct tcp_info *info)
2112 struct tcp_sock *tp = tcp_sk(sk);
2113 u32 now = tcp_time_stamp;
2115 memset(info, 0, sizeof(*info));
2117 info->tcpi_state = sk->sk_state;
2118 info->tcpi_ca_state = tp->ca_state;
2119 info->tcpi_retransmits = tp->retransmits;
2120 info->tcpi_probes = tp->probes_out;
2121 info->tcpi_backoff = tp->backoff;
2123 if (tp->rx_opt.tstamp_ok)
2124 info->tcpi_options |= TCPI_OPT_TIMESTAMPS;
2125 if (tp->rx_opt.sack_ok)
2126 info->tcpi_options |= TCPI_OPT_SACK;
2127 if (tp->rx_opt.wscale_ok) {
2128 info->tcpi_options |= TCPI_OPT_WSCALE;
2129 info->tcpi_snd_wscale = tp->rx_opt.snd_wscale;
2130 info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale;
2133 if (tp->ecn_flags&TCP_ECN_OK)
2134 info->tcpi_options |= TCPI_OPT_ECN;
2136 info->tcpi_rto = jiffies_to_usecs(tp->rto);
2137 info->tcpi_ato = jiffies_to_usecs(tp->ack.ato);
2138 info->tcpi_snd_mss = tp->mss_cache_std;
2139 info->tcpi_rcv_mss = tp->ack.rcv_mss;
2141 info->tcpi_unacked = tp->packets_out;
2142 info->tcpi_sacked = tp->sacked_out;
2143 info->tcpi_lost = tp->lost_out;
2144 info->tcpi_retrans = tp->retrans_out;
2145 info->tcpi_fackets = tp->fackets_out;
2147 info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime);
2148 info->tcpi_last_data_recv = jiffies_to_msecs(now - tp->ack.lrcvtime);
2149 info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp);
2151 info->tcpi_pmtu = tp->pmtu_cookie;
2152 info->tcpi_rcv_ssthresh = tp->rcv_ssthresh;
2153 info->tcpi_rtt = jiffies_to_usecs(tp->srtt)>>3;
2154 info->tcpi_rttvar = jiffies_to_usecs(tp->mdev)>>2;
2155 info->tcpi_snd_ssthresh = tp->snd_ssthresh;
2156 info->tcpi_snd_cwnd = tp->snd_cwnd;
2157 info->tcpi_advmss = tp->advmss;
2158 info->tcpi_reordering = tp->reordering;
2160 info->tcpi_rcv_rtt = jiffies_to_usecs(tp->rcv_rtt_est.rtt)>>3;
2161 info->tcpi_rcv_space = tp->rcvq_space.space;
2163 info->tcpi_total_retrans = tp->total_retrans;
2166 EXPORT_SYMBOL_GPL(tcp_get_info);
2168 int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval,
2171 struct tcp_sock *tp = tcp_sk(sk);
2174 if (level != SOL_TCP)
2175 return tp->af_specific->getsockopt(sk, level, optname,
2178 if (get_user(len, optlen))
2181 len = min_t(unsigned int, len, sizeof(int));
2188 val = tp->mss_cache_std;
2189 if (!val && ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
2190 val = tp->rx_opt.user_mss;
2193 val = !!(tp->nonagle&TCP_NAGLE_OFF);
2196 val = !!(tp->nonagle&TCP_NAGLE_CORK);
2199 val = (tp->keepalive_time ? : sysctl_tcp_keepalive_time) / HZ;
2202 val = (tp->keepalive_intvl ? : sysctl_tcp_keepalive_intvl) / HZ;
2205 val = tp->keepalive_probes ? : sysctl_tcp_keepalive_probes;
2208 val = tp->syn_retries ? : sysctl_tcp_syn_retries;
2213 val = (val ? : sysctl_tcp_fin_timeout) / HZ;
2215 case TCP_DEFER_ACCEPT:
2216 val = !tp->defer_accept ? 0 : ((TCP_TIMEOUT_INIT / HZ) <<
2217 (tp->defer_accept - 1));
2219 case TCP_WINDOW_CLAMP:
2220 val = tp->window_clamp;
2223 struct tcp_info info;
2225 if (get_user(len, optlen))
2228 tcp_get_info(sk, &info);
2230 len = min_t(unsigned int, len, sizeof(info));
2231 if (put_user(len, optlen))
2233 if (copy_to_user(optval, &info, len))
2238 val = !tp->ack.pingpong;
2242 return -ENOPROTOOPT;
2245 if (put_user(len, optlen))
2247 if (copy_to_user(optval, &val, len))
2253 extern void __skb_cb_too_small_for_tcp(int, int);
2254 extern void tcpdiag_init(void);
2256 static __initdata unsigned long thash_entries;
2257 static int __init set_thash_entries(char *str)
2261 thash_entries = simple_strtoul(str, &str, 0);
2264 __setup("thash_entries=", set_thash_entries);
2266 void __init tcp_init(void)
2268 struct sk_buff *skb = NULL;
2271 if (sizeof(struct tcp_skb_cb) > sizeof(skb->cb))
2272 __skb_cb_too_small_for_tcp(sizeof(struct tcp_skb_cb),
2275 tcp_openreq_cachep = kmem_cache_create("tcp_open_request",
2276 sizeof(struct open_request),
2277 0, SLAB_HWCACHE_ALIGN,
2279 if (!tcp_openreq_cachep)
2280 panic("tcp_init: Cannot alloc open_request cache.");
2282 tcp_bucket_cachep = kmem_cache_create("tcp_bind_bucket",
2283 sizeof(struct tcp_bind_bucket),
2284 0, SLAB_HWCACHE_ALIGN,
2286 if (!tcp_bucket_cachep)
2287 panic("tcp_init: Cannot alloc tcp_bind_bucket cache.");
2289 tcp_timewait_cachep = kmem_cache_create("tcp_tw_bucket",
2290 sizeof(struct tcp_tw_bucket),
2291 0, SLAB_HWCACHE_ALIGN,
2293 if (!tcp_timewait_cachep)
2294 panic("tcp_init: Cannot alloc tcp_tw_bucket cache.");
2296 /* Size and allocate the main established and bind bucket
2299 * The methodology is similar to that of the buffer cache.
2301 tcp_ehash = (struct tcp_ehash_bucket *)
2302 alloc_large_system_hash("TCP established",
2303 sizeof(struct tcp_ehash_bucket),
2305 (num_physpages >= 128 * 1024) ?
2312 tcp_ehash_size = (1 << tcp_ehash_size) >> 1;
2313 for (i = 0; i < (tcp_ehash_size << 1); i++) {
2314 rwlock_init(&tcp_ehash[i].lock);
2315 INIT_HLIST_HEAD(&tcp_ehash[i].chain);
2318 tcp_bhash = (struct tcp_bind_hashbucket *)
2319 alloc_large_system_hash("TCP bind",
2320 sizeof(struct tcp_bind_hashbucket),
2322 (num_physpages >= 128 * 1024) ?
2329 tcp_bhash_size = 1 << tcp_bhash_size;
2330 for (i = 0; i < tcp_bhash_size; i++) {
2331 spin_lock_init(&tcp_bhash[i].lock);
2332 INIT_HLIST_HEAD(&tcp_bhash[i].chain);
2335 /* Try to be a bit smarter and adjust defaults depending
2336 * on available memory.
2338 for (order = 0; ((1 << order) << PAGE_SHIFT) <
2339 (tcp_bhash_size * sizeof(struct tcp_bind_hashbucket));
2343 sysctl_local_port_range[0] = 32768;
2344 sysctl_local_port_range[1] = 61000;
2345 sysctl_tcp_max_tw_buckets = 180000;
2346 sysctl_tcp_max_orphans = 4096 << (order - 4);
2347 sysctl_max_syn_backlog = 1024;
2348 } else if (order < 3) {
2349 sysctl_local_port_range[0] = 1024 * (3 - order);
2350 sysctl_tcp_max_tw_buckets >>= (3 - order);
2351 sysctl_tcp_max_orphans >>= (3 - order);
2352 sysctl_max_syn_backlog = 128;
2354 tcp_port_rover = sysctl_local_port_range[0] - 1;
2356 sysctl_tcp_mem[0] = 768 << order;
2357 sysctl_tcp_mem[1] = 1024 << order;
2358 sysctl_tcp_mem[2] = 1536 << order;
2361 sysctl_tcp_wmem[2] = 64 * 1024;
2362 sysctl_tcp_rmem[0] = PAGE_SIZE;
2363 sysctl_tcp_rmem[1] = 43689;
2364 sysctl_tcp_rmem[2] = 2 * 43689;
2367 printk(KERN_INFO "TCP: Hash tables configured "
2368 "(established %d bind %d)\n",
2369 tcp_ehash_size << 1, tcp_bhash_size);
2372 EXPORT_SYMBOL(tcp_accept);
2373 EXPORT_SYMBOL(tcp_close);
2374 EXPORT_SYMBOL(tcp_destroy_sock);
2375 EXPORT_SYMBOL(tcp_disconnect);
2376 EXPORT_SYMBOL(tcp_getsockopt);
2377 EXPORT_SYMBOL(tcp_ioctl);
2378 EXPORT_SYMBOL(tcp_openreq_cachep);
2379 EXPORT_SYMBOL(tcp_poll);
2380 EXPORT_SYMBOL(tcp_read_sock);
2381 EXPORT_SYMBOL(tcp_recvmsg);
2382 EXPORT_SYMBOL(tcp_sendmsg);
2383 EXPORT_SYMBOL(tcp_sendpage);
2384 EXPORT_SYMBOL(tcp_setsockopt);
2385 EXPORT_SYMBOL(tcp_shutdown);
2386 EXPORT_SYMBOL(tcp_statistics);
2387 EXPORT_SYMBOL(tcp_timewait_cachep);
2388 EXPORT_SYMBOL_GPL(cleanup_rbuf);