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_bucket_cachep;
275 kmem_cache_t *tcp_timewait_cachep;
277 atomic_t tcp_orphan_count = ATOMIC_INIT(0);
279 int sysctl_tcp_mem[3];
280 int sysctl_tcp_wmem[3] = { 4 * 1024, 16 * 1024, 128 * 1024 };
281 int sysctl_tcp_rmem[3] = { 4 * 1024, 87380, 87380 * 2 };
283 EXPORT_SYMBOL(sysctl_tcp_mem);
284 EXPORT_SYMBOL(sysctl_tcp_rmem);
285 EXPORT_SYMBOL(sysctl_tcp_wmem);
287 atomic_t tcp_memory_allocated; /* Current allocated memory. */
288 atomic_t tcp_sockets_allocated; /* Current number of TCP sockets. */
290 EXPORT_SYMBOL(tcp_memory_allocated);
291 EXPORT_SYMBOL(tcp_sockets_allocated);
294 * Pressure flag: try to collapse.
295 * Technical note: it is used by multiple contexts non atomically.
296 * All the sk_stream_mem_schedule() is of this nature: accounting
297 * is strict, actions are advisory and have some latency.
299 int tcp_memory_pressure;
301 EXPORT_SYMBOL(tcp_memory_pressure);
303 void tcp_enter_memory_pressure(void)
305 if (!tcp_memory_pressure) {
306 NET_INC_STATS(LINUX_MIB_TCPMEMORYPRESSURES);
307 tcp_memory_pressure = 1;
311 EXPORT_SYMBOL(tcp_enter_memory_pressure);
314 * LISTEN is a special case for poll..
316 static __inline__ unsigned int tcp_listen_poll(struct sock *sk,
319 return !reqsk_queue_empty(&tcp_sk(sk)->accept_queue) ? (POLLIN | POLLRDNORM) : 0;
323 * Wait for a TCP event.
325 * Note that we don't need to lock the socket, as the upper poll layers
326 * take care of normal races (between the test and the event) and we don't
327 * go look at any of the socket buffers directly.
329 unsigned int tcp_poll(struct file *file, struct socket *sock, poll_table *wait)
332 struct sock *sk = sock->sk;
333 struct tcp_sock *tp = tcp_sk(sk);
335 poll_wait(file, sk->sk_sleep, wait);
336 if (sk->sk_state == TCP_LISTEN)
337 return tcp_listen_poll(sk, wait);
339 /* Socket is not locked. We are protected from async events
340 by poll logic and correct handling of state changes
341 made by another threads is impossible in any case.
349 * POLLHUP is certainly not done right. But poll() doesn't
350 * have a notion of HUP in just one direction, and for a
351 * socket the read side is more interesting.
353 * Some poll() documentation says that POLLHUP is incompatible
354 * with the POLLOUT/POLLWR flags, so somebody should check this
355 * all. But careful, it tends to be safer to return too many
356 * bits than too few, and you can easily break real applications
357 * if you don't tell them that something has hung up!
361 * Check number 1. POLLHUP is _UNMASKABLE_ event (see UNIX98 and
362 * our fs/select.c). It means that after we received EOF,
363 * poll always returns immediately, making impossible poll() on write()
364 * in state CLOSE_WAIT. One solution is evident --- to set POLLHUP
365 * if and only if shutdown has been made in both directions.
366 * Actually, it is interesting to look how Solaris and DUX
367 * solve this dilemma. I would prefer, if PULLHUP were maskable,
368 * then we could set it on SND_SHUTDOWN. BTW examples given
369 * in Stevens' books assume exactly this behaviour, it explains
370 * why PULLHUP is incompatible with POLLOUT. --ANK
372 * NOTE. Check for TCP_CLOSE is added. The goal is to prevent
373 * blocking on fresh not-connected or disconnected socket. --ANK
375 if (sk->sk_shutdown == SHUTDOWN_MASK || sk->sk_state == TCP_CLOSE)
377 if (sk->sk_shutdown & RCV_SHUTDOWN)
378 mask |= POLLIN | POLLRDNORM;
381 if ((1 << sk->sk_state) & ~(TCPF_SYN_SENT | TCPF_SYN_RECV)) {
382 /* Potential race condition. If read of tp below will
383 * escape above sk->sk_state, we can be illegally awaken
384 * in SYN_* states. */
385 if ((tp->rcv_nxt != tp->copied_seq) &&
386 (tp->urg_seq != tp->copied_seq ||
387 tp->rcv_nxt != tp->copied_seq + 1 ||
388 sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data))
389 mask |= POLLIN | POLLRDNORM;
391 if (!(sk->sk_shutdown & SEND_SHUTDOWN)) {
392 if (sk_stream_wspace(sk) >= sk_stream_min_wspace(sk)) {
393 mask |= POLLOUT | POLLWRNORM;
394 } else { /* send SIGIO later */
395 set_bit(SOCK_ASYNC_NOSPACE,
396 &sk->sk_socket->flags);
397 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
399 /* Race breaker. If space is freed after
400 * wspace test but before the flags are set,
401 * IO signal will be lost.
403 if (sk_stream_wspace(sk) >= sk_stream_min_wspace(sk))
404 mask |= POLLOUT | POLLWRNORM;
408 if (tp->urg_data & TCP_URG_VALID)
414 int tcp_ioctl(struct sock *sk, int cmd, unsigned long arg)
416 struct tcp_sock *tp = tcp_sk(sk);
421 if (sk->sk_state == TCP_LISTEN)
425 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
427 else if (sock_flag(sk, SOCK_URGINLINE) ||
429 before(tp->urg_seq, tp->copied_seq) ||
430 !before(tp->urg_seq, tp->rcv_nxt)) {
431 answ = tp->rcv_nxt - tp->copied_seq;
433 /* Subtract 1, if FIN is in queue. */
434 if (answ && !skb_queue_empty(&sk->sk_receive_queue))
436 ((struct sk_buff *)sk->sk_receive_queue.prev)->h.th->fin;
438 answ = tp->urg_seq - tp->copied_seq;
442 answ = tp->urg_data && tp->urg_seq == tp->copied_seq;
445 if (sk->sk_state == TCP_LISTEN)
448 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
451 answ = tp->write_seq - tp->snd_una;
457 return put_user(answ, (int __user *)arg);
461 int tcp_listen_start(struct sock *sk)
463 struct inet_sock *inet = inet_sk(sk);
464 struct tcp_sock *tp = tcp_sk(sk);
465 int rc = reqsk_queue_alloc(&tp->accept_queue, TCP_SYNQ_HSIZE);
470 sk->sk_max_ack_backlog = 0;
471 sk->sk_ack_backlog = 0;
474 /* There is race window here: we announce ourselves listening,
475 * but this transition is still not validated by get_port().
476 * It is OK, because this socket enters to hash table only
477 * after validation is complete.
479 sk->sk_state = TCP_LISTEN;
480 if (!sk->sk_prot->get_port(sk, inet->num)) {
481 inet->sport = htons(inet->num);
484 sk->sk_prot->hash(sk);
489 sk->sk_state = TCP_CLOSE;
490 reqsk_queue_destroy(&tp->accept_queue);
495 * This routine closes sockets which have been at least partially
496 * opened, but not yet accepted.
499 static void tcp_listen_stop (struct sock *sk)
501 struct tcp_sock *tp = tcp_sk(sk);
502 struct listen_sock *lopt;
503 struct request_sock *acc_req;
504 struct request_sock *req;
507 tcp_delete_keepalive_timer(sk);
509 /* make all the listen_opt local to us */
510 lopt = reqsk_queue_yank_listen_sk(&tp->accept_queue);
511 acc_req = reqsk_queue_yank_acceptq(&tp->accept_queue);
514 for (i = 0; i < TCP_SYNQ_HSIZE; i++) {
515 while ((req = lopt->syn_table[i]) != NULL) {
516 lopt->syn_table[i] = req->dl_next;
520 /* Following specs, it would be better either to send FIN
521 * (and enter FIN-WAIT-1, it is normal close)
522 * or to send active reset (abort).
523 * Certainly, it is pretty dangerous while synflood, but it is
524 * bad justification for our negligence 8)
525 * To be honest, we are not able to make either
526 * of the variants now. --ANK
531 BUG_TRAP(!lopt->qlen);
535 while ((req = acc_req) != NULL) {
536 struct sock *child = req->sk;
538 acc_req = req->dl_next;
542 BUG_TRAP(!sock_owned_by_user(child));
545 tcp_disconnect(child, O_NONBLOCK);
549 atomic_inc(&tcp_orphan_count);
551 tcp_destroy_sock(child);
553 bh_unlock_sock(child);
557 sk_acceptq_removed(sk);
560 BUG_TRAP(!sk->sk_ack_backlog);
563 static inline void tcp_mark_push(struct tcp_sock *tp, struct sk_buff *skb)
565 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH;
566 tp->pushed_seq = tp->write_seq;
569 static inline int forced_push(struct tcp_sock *tp)
571 return after(tp->write_seq, tp->pushed_seq + (tp->max_window >> 1));
574 static inline void skb_entail(struct sock *sk, struct tcp_sock *tp,
578 TCP_SKB_CB(skb)->seq = tp->write_seq;
579 TCP_SKB_CB(skb)->end_seq = tp->write_seq;
580 TCP_SKB_CB(skb)->flags = TCPCB_FLAG_ACK;
581 TCP_SKB_CB(skb)->sacked = 0;
582 skb_header_release(skb);
583 __skb_queue_tail(&sk->sk_write_queue, skb);
584 sk_charge_skb(sk, skb);
585 if (!sk->sk_send_head)
586 sk->sk_send_head = skb;
587 else if (tp->nonagle&TCP_NAGLE_PUSH)
588 tp->nonagle &= ~TCP_NAGLE_PUSH;
591 static inline void tcp_mark_urg(struct tcp_sock *tp, int flags,
594 if (flags & MSG_OOB) {
596 tp->snd_up = tp->write_seq;
597 TCP_SKB_CB(skb)->sacked |= TCPCB_URG;
601 static inline void tcp_push(struct sock *sk, struct tcp_sock *tp, int flags,
602 int mss_now, int nonagle)
604 if (sk->sk_send_head) {
605 struct sk_buff *skb = sk->sk_write_queue.prev;
606 if (!(flags & MSG_MORE) || forced_push(tp))
607 tcp_mark_push(tp, skb);
608 tcp_mark_urg(tp, flags, skb);
609 __tcp_push_pending_frames(sk, tp, mss_now,
610 (flags & MSG_MORE) ? TCP_NAGLE_CORK : nonagle);
614 static ssize_t do_tcp_sendpages(struct sock *sk, struct page **pages, int poffset,
615 size_t psize, int flags)
617 struct tcp_sock *tp = tcp_sk(sk);
621 long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
623 /* Wait for a connection to finish. */
624 if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
625 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
628 clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
630 mss_now = tcp_current_mss(sk, !(flags&MSG_OOB));
634 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
638 struct sk_buff *skb = sk->sk_write_queue.prev;
639 struct page *page = pages[poffset / PAGE_SIZE];
640 int copy, i, can_coalesce;
641 int offset = poffset % PAGE_SIZE;
642 int size = min_t(size_t, psize, PAGE_SIZE - offset);
644 if (!sk->sk_send_head || (copy = mss_now - skb->len) <= 0) {
646 if (!sk_stream_memory_free(sk))
647 goto wait_for_sndbuf;
649 skb = sk_stream_alloc_pskb(sk, 0, 0,
652 goto wait_for_memory;
654 skb_entail(sk, tp, skb);
661 i = skb_shinfo(skb)->nr_frags;
662 can_coalesce = skb_can_coalesce(skb, i, page, offset);
663 if (!can_coalesce && i >= MAX_SKB_FRAGS) {
664 tcp_mark_push(tp, skb);
667 if (sk->sk_forward_alloc < copy &&
668 !sk_stream_mem_schedule(sk, copy, 0))
669 goto wait_for_memory;
672 skb_shinfo(skb)->frags[i - 1].size += copy;
675 skb_fill_page_desc(skb, i, page, offset, copy);
679 skb->data_len += copy;
680 skb->truesize += copy;
681 sk->sk_wmem_queued += copy;
682 sk->sk_forward_alloc -= copy;
683 skb->ip_summed = CHECKSUM_HW;
684 tp->write_seq += copy;
685 TCP_SKB_CB(skb)->end_seq += copy;
686 skb_shinfo(skb)->tso_segs = 0;
689 TCP_SKB_CB(skb)->flags &= ~TCPCB_FLAG_PSH;
693 if (!(psize -= copy))
696 if (skb->len != mss_now || (flags & MSG_OOB))
699 if (forced_push(tp)) {
700 tcp_mark_push(tp, skb);
701 __tcp_push_pending_frames(sk, tp, mss_now, TCP_NAGLE_PUSH);
702 } else if (skb == sk->sk_send_head)
703 tcp_push_one(sk, mss_now);
707 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
710 tcp_push(sk, tp, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH);
712 if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
715 mss_now = tcp_current_mss(sk, !(flags&MSG_OOB));
720 tcp_push(sk, tp, flags, mss_now, tp->nonagle);
727 return sk_stream_error(sk, flags, err);
730 ssize_t tcp_sendpage(struct socket *sock, struct page *page, int offset,
731 size_t size, int flags)
734 struct sock *sk = sock->sk;
736 #define TCP_ZC_CSUM_FLAGS (NETIF_F_IP_CSUM | NETIF_F_NO_CSUM | NETIF_F_HW_CSUM)
738 if (!(sk->sk_route_caps & NETIF_F_SG) ||
739 !(sk->sk_route_caps & TCP_ZC_CSUM_FLAGS))
740 return sock_no_sendpage(sock, page, offset, size, flags);
742 #undef TCP_ZC_CSUM_FLAGS
746 res = do_tcp_sendpages(sk, &page, offset, size, flags);
752 #define TCP_PAGE(sk) (sk->sk_sndmsg_page)
753 #define TCP_OFF(sk) (sk->sk_sndmsg_off)
755 static inline int select_size(struct sock *sk, struct tcp_sock *tp)
757 int tmp = tp->mss_cache_std;
759 if (sk->sk_route_caps & NETIF_F_SG)
765 int tcp_sendmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg,
769 struct tcp_sock *tp = tcp_sk(sk);
779 flags = msg->msg_flags;
780 timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
782 /* Wait for a connection to finish. */
783 if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
784 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
787 /* This should be in poll */
788 clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
790 mss_now = tcp_current_mss(sk, !(flags&MSG_OOB));
792 /* Ok commence sending. */
793 iovlen = msg->msg_iovlen;
798 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
801 while (--iovlen >= 0) {
802 int seglen = iov->iov_len;
803 unsigned char __user *from = iov->iov_base;
810 skb = sk->sk_write_queue.prev;
812 if (!sk->sk_send_head ||
813 (copy = mss_now - skb->len) <= 0) {
816 /* Allocate new segment. If the interface is SG,
817 * allocate skb fitting to single page.
819 if (!sk_stream_memory_free(sk))
820 goto wait_for_sndbuf;
822 skb = sk_stream_alloc_pskb(sk, select_size(sk, tp),
823 0, sk->sk_allocation);
825 goto wait_for_memory;
828 * Check whether we can use HW checksum.
830 if (sk->sk_route_caps &
831 (NETIF_F_IP_CSUM | NETIF_F_NO_CSUM |
833 skb->ip_summed = CHECKSUM_HW;
835 skb_entail(sk, tp, skb);
839 /* Try to append data to the end of skb. */
843 /* Where to copy to? */
844 if (skb_tailroom(skb) > 0) {
845 /* We have some space in skb head. Superb! */
846 if (copy > skb_tailroom(skb))
847 copy = skb_tailroom(skb);
848 if ((err = skb_add_data(skb, from, copy)) != 0)
852 int i = skb_shinfo(skb)->nr_frags;
853 struct page *page = TCP_PAGE(sk);
854 int off = TCP_OFF(sk);
856 if (skb_can_coalesce(skb, i, page, off) &&
858 /* We can extend the last page
861 } else if (i == MAX_SKB_FRAGS ||
863 !(sk->sk_route_caps & NETIF_F_SG))) {
864 /* Need to add new fragment and cannot
865 * do this because interface is non-SG,
866 * or because all the page slots are
868 tcp_mark_push(tp, skb);
871 if (off == PAGE_SIZE) {
873 TCP_PAGE(sk) = page = NULL;
878 /* Allocate new cache page. */
879 if (!(page = sk_stream_alloc_page(sk)))
880 goto wait_for_memory;
884 if (copy > PAGE_SIZE - off)
885 copy = PAGE_SIZE - off;
887 /* Time to copy data. We are close to
889 err = skb_copy_to_page(sk, from, skb, page,
892 /* If this page was new, give it to the
893 * socket so it does not get leaked.
902 /* Update the skb. */
904 skb_shinfo(skb)->frags[i - 1].size +=
907 skb_fill_page_desc(skb, i, page, off, copy);
910 } else if (off + copy < PAGE_SIZE) {
916 TCP_OFF(sk) = off + copy;
920 TCP_SKB_CB(skb)->flags &= ~TCPCB_FLAG_PSH;
922 tp->write_seq += copy;
923 TCP_SKB_CB(skb)->end_seq += copy;
924 skb_shinfo(skb)->tso_segs = 0;
928 if ((seglen -= copy) == 0 && iovlen == 0)
931 if (skb->len != mss_now || (flags & MSG_OOB))
934 if (forced_push(tp)) {
935 tcp_mark_push(tp, skb);
936 __tcp_push_pending_frames(sk, tp, mss_now, TCP_NAGLE_PUSH);
937 } else if (skb == sk->sk_send_head)
938 tcp_push_one(sk, mss_now);
942 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
945 tcp_push(sk, tp, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH);
947 if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
950 mss_now = tcp_current_mss(sk, !(flags&MSG_OOB));
956 tcp_push(sk, tp, flags, mss_now, tp->nonagle);
963 if (sk->sk_send_head == skb)
964 sk->sk_send_head = NULL;
965 __skb_unlink(skb, skb->list);
966 sk_stream_free_skb(sk, skb);
973 err = sk_stream_error(sk, flags, err);
980 * Handle reading urgent data. BSD has very simple semantics for
981 * this, no blocking and very strange errors 8)
984 static int tcp_recv_urg(struct sock *sk, long timeo,
985 struct msghdr *msg, int len, int flags,
988 struct tcp_sock *tp = tcp_sk(sk);
990 /* No URG data to read. */
991 if (sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data ||
992 tp->urg_data == TCP_URG_READ)
993 return -EINVAL; /* Yes this is right ! */
995 if (sk->sk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DONE))
998 if (tp->urg_data & TCP_URG_VALID) {
1000 char c = tp->urg_data;
1002 if (!(flags & MSG_PEEK))
1003 tp->urg_data = TCP_URG_READ;
1005 /* Read urgent data. */
1006 msg->msg_flags |= MSG_OOB;
1009 if (!(flags & MSG_TRUNC))
1010 err = memcpy_toiovec(msg->msg_iov, &c, 1);
1013 msg->msg_flags |= MSG_TRUNC;
1015 return err ? -EFAULT : len;
1018 if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
1021 /* Fixed the recv(..., MSG_OOB) behaviour. BSD docs and
1022 * the available implementations agree in this case:
1023 * this call should never block, independent of the
1024 * blocking state of the socket.
1025 * Mike <pall@rz.uni-karlsruhe.de>
1030 /* Clean up the receive buffer for full frames taken by the user,
1031 * then send an ACK if necessary. COPIED is the number of bytes
1032 * tcp_recvmsg has given to the user so far, it speeds up the
1033 * calculation of whether or not we must ACK for the sake of
1036 static void cleanup_rbuf(struct sock *sk, int copied)
1038 struct tcp_sock *tp = tcp_sk(sk);
1039 int time_to_ack = 0;
1042 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
1044 BUG_TRAP(!skb || before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq));
1047 if (tcp_ack_scheduled(tp)) {
1048 /* Delayed ACKs frequently hit locked sockets during bulk
1050 if (tp->ack.blocked ||
1051 /* Once-per-two-segments ACK was not sent by tcp_input.c */
1052 tp->rcv_nxt - tp->rcv_wup > tp->ack.rcv_mss ||
1054 * If this read emptied read buffer, we send ACK, if
1055 * connection is not bidirectional, user drained
1056 * receive buffer and there was a small segment
1059 (copied > 0 && (tp->ack.pending & TCP_ACK_PUSHED) &&
1060 !tp->ack.pingpong && !atomic_read(&sk->sk_rmem_alloc)))
1064 /* We send an ACK if we can now advertise a non-zero window
1065 * which has been raised "significantly".
1067 * Even if window raised up to infinity, do not send window open ACK
1068 * in states, where we will not receive more. It is useless.
1070 if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) {
1071 __u32 rcv_window_now = tcp_receive_window(tp);
1073 /* Optimize, __tcp_select_window() is not cheap. */
1074 if (2*rcv_window_now <= tp->window_clamp) {
1075 __u32 new_window = __tcp_select_window(sk);
1077 /* Send ACK now, if this read freed lots of space
1078 * in our buffer. Certainly, new_window is new window.
1079 * We can advertise it now, if it is not less than current one.
1080 * "Lots" means "at least twice" here.
1082 if (new_window && new_window >= 2 * rcv_window_now)
1090 static void tcp_prequeue_process(struct sock *sk)
1092 struct sk_buff *skb;
1093 struct tcp_sock *tp = tcp_sk(sk);
1095 NET_ADD_STATS_USER(LINUX_MIB_TCPPREQUEUED, skb_queue_len(&tp->ucopy.prequeue));
1097 /* RX process wants to run with disabled BHs, though it is not
1100 while ((skb = __skb_dequeue(&tp->ucopy.prequeue)) != NULL)
1101 sk->sk_backlog_rcv(sk, skb);
1104 /* Clear memory counter. */
1105 tp->ucopy.memory = 0;
1108 static inline struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off)
1110 struct sk_buff *skb;
1113 skb_queue_walk(&sk->sk_receive_queue, skb) {
1114 offset = seq - TCP_SKB_CB(skb)->seq;
1117 if (offset < skb->len || skb->h.th->fin) {
1126 * This routine provides an alternative to tcp_recvmsg() for routines
1127 * that would like to handle copying from skbuffs directly in 'sendfile'
1130 * - It is assumed that the socket was locked by the caller.
1131 * - The routine does not block.
1132 * - At present, there is no support for reading OOB data
1133 * or for 'peeking' the socket using this routine
1134 * (although both would be easy to implement).
1136 int tcp_read_sock(struct sock *sk, read_descriptor_t *desc,
1137 sk_read_actor_t recv_actor)
1139 struct sk_buff *skb;
1140 struct tcp_sock *tp = tcp_sk(sk);
1141 u32 seq = tp->copied_seq;
1145 if (sk->sk_state == TCP_LISTEN)
1147 while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1148 if (offset < skb->len) {
1151 len = skb->len - offset;
1152 /* Stop reading if we hit a patch of urgent data */
1154 u32 urg_offset = tp->urg_seq - seq;
1155 if (urg_offset < len)
1160 used = recv_actor(desc, skb, offset, len);
1166 if (offset != skb->len)
1169 if (skb->h.th->fin) {
1170 sk_eat_skb(sk, skb);
1174 sk_eat_skb(sk, skb);
1178 tp->copied_seq = seq;
1180 tcp_rcv_space_adjust(sk);
1182 /* Clean up data we have read: This will do ACK frames. */
1184 cleanup_rbuf(sk, copied);
1189 * This routine copies from a sock struct into the user buffer.
1191 * Technical note: in 2.3 we work on _locked_ socket, so that
1192 * tricks with *seq access order and skb->users are not required.
1193 * Probably, code can be easily improved even more.
1196 int tcp_recvmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg,
1197 size_t len, int nonblock, int flags, int *addr_len)
1199 struct tcp_sock *tp = tcp_sk(sk);
1205 int target; /* Read at least this many bytes */
1207 struct task_struct *user_recv = NULL;
1211 TCP_CHECK_TIMER(sk);
1214 if (sk->sk_state == TCP_LISTEN)
1217 timeo = sock_rcvtimeo(sk, nonblock);
1219 /* Urgent data needs to be handled specially. */
1220 if (flags & MSG_OOB)
1223 seq = &tp->copied_seq;
1224 if (flags & MSG_PEEK) {
1225 peek_seq = tp->copied_seq;
1229 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
1232 struct sk_buff *skb;
1235 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
1236 if (tp->urg_data && tp->urg_seq == *seq) {
1239 if (signal_pending(current)) {
1240 copied = timeo ? sock_intr_errno(timeo) : -EAGAIN;
1245 /* Next get a buffer. */
1247 skb = skb_peek(&sk->sk_receive_queue);
1252 /* Now that we have two receive queues this
1255 if (before(*seq, TCP_SKB_CB(skb)->seq)) {
1256 printk(KERN_INFO "recvmsg bug: copied %X "
1257 "seq %X\n", *seq, TCP_SKB_CB(skb)->seq);
1260 offset = *seq - TCP_SKB_CB(skb)->seq;
1263 if (offset < skb->len)
1267 BUG_TRAP(flags & MSG_PEEK);
1269 } while (skb != (struct sk_buff *)&sk->sk_receive_queue);
1271 /* Well, if we have backlog, try to process it now yet. */
1273 if (copied >= target && !sk->sk_backlog.tail)
1278 sk->sk_state == TCP_CLOSE ||
1279 (sk->sk_shutdown & RCV_SHUTDOWN) ||
1281 signal_pending(current) ||
1285 if (sock_flag(sk, SOCK_DONE))
1289 copied = sock_error(sk);
1293 if (sk->sk_shutdown & RCV_SHUTDOWN)
1296 if (sk->sk_state == TCP_CLOSE) {
1297 if (!sock_flag(sk, SOCK_DONE)) {
1298 /* This occurs when user tries to read
1299 * from never connected socket.
1312 if (signal_pending(current)) {
1313 copied = sock_intr_errno(timeo);
1318 cleanup_rbuf(sk, copied);
1320 if (!sysctl_tcp_low_latency && tp->ucopy.task == user_recv) {
1321 /* Install new reader */
1322 if (!user_recv && !(flags & (MSG_TRUNC | MSG_PEEK))) {
1323 user_recv = current;
1324 tp->ucopy.task = user_recv;
1325 tp->ucopy.iov = msg->msg_iov;
1328 tp->ucopy.len = len;
1330 BUG_TRAP(tp->copied_seq == tp->rcv_nxt ||
1331 (flags & (MSG_PEEK | MSG_TRUNC)));
1333 /* Ugly... If prequeue is not empty, we have to
1334 * process it before releasing socket, otherwise
1335 * order will be broken at second iteration.
1336 * More elegant solution is required!!!
1338 * Look: we have the following (pseudo)queues:
1340 * 1. packets in flight
1345 * Each queue can be processed only if the next ones
1346 * are empty. At this point we have empty receive_queue.
1347 * But prequeue _can_ be not empty after 2nd iteration,
1348 * when we jumped to start of loop because backlog
1349 * processing added something to receive_queue.
1350 * We cannot release_sock(), because backlog contains
1351 * packets arrived _after_ prequeued ones.
1353 * Shortly, algorithm is clear --- to process all
1354 * the queues in order. We could make it more directly,
1355 * requeueing packets from backlog to prequeue, if
1356 * is not empty. It is more elegant, but eats cycles,
1359 if (skb_queue_len(&tp->ucopy.prequeue))
1362 /* __ Set realtime policy in scheduler __ */
1365 if (copied >= target) {
1366 /* Do not sleep, just process backlog. */
1370 sk_wait_data(sk, &timeo);
1375 /* __ Restore normal policy in scheduler __ */
1377 if ((chunk = len - tp->ucopy.len) != 0) {
1378 NET_ADD_STATS_USER(LINUX_MIB_TCPDIRECTCOPYFROMBACKLOG, chunk);
1383 if (tp->rcv_nxt == tp->copied_seq &&
1384 skb_queue_len(&tp->ucopy.prequeue)) {
1386 tcp_prequeue_process(sk);
1388 if ((chunk = len - tp->ucopy.len) != 0) {
1389 NET_ADD_STATS_USER(LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1395 if ((flags & MSG_PEEK) && peek_seq != tp->copied_seq) {
1396 if (net_ratelimit())
1397 printk(KERN_DEBUG "TCP(%s:%d): Application bug, race in MSG_PEEK.\n",
1398 current->comm, current->pid);
1399 peek_seq = tp->copied_seq;
1404 /* Ok so how much can we use? */
1405 used = skb->len - offset;
1409 /* Do we have urgent data here? */
1411 u32 urg_offset = tp->urg_seq - *seq;
1412 if (urg_offset < used) {
1414 if (!sock_flag(sk, SOCK_URGINLINE)) {
1426 if (!(flags & MSG_TRUNC)) {
1427 err = skb_copy_datagram_iovec(skb, offset,
1428 msg->msg_iov, used);
1430 /* Exception. Bailout! */
1441 tcp_rcv_space_adjust(sk);
1444 if (tp->urg_data && after(tp->copied_seq, tp->urg_seq)) {
1446 tcp_fast_path_check(sk, tp);
1448 if (used + offset < skb->len)
1453 if (!(flags & MSG_PEEK))
1454 sk_eat_skb(sk, skb);
1458 /* Process the FIN. */
1460 if (!(flags & MSG_PEEK))
1461 sk_eat_skb(sk, skb);
1466 if (skb_queue_len(&tp->ucopy.prequeue)) {
1469 tp->ucopy.len = copied > 0 ? len : 0;
1471 tcp_prequeue_process(sk);
1473 if (copied > 0 && (chunk = len - tp->ucopy.len) != 0) {
1474 NET_ADD_STATS_USER(LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1480 tp->ucopy.task = NULL;
1484 /* According to UNIX98, msg_name/msg_namelen are ignored
1485 * on connected socket. I was just happy when found this 8) --ANK
1488 /* Clean up data we have read: This will do ACK frames. */
1489 cleanup_rbuf(sk, copied);
1491 TCP_CHECK_TIMER(sk);
1496 TCP_CHECK_TIMER(sk);
1501 err = tcp_recv_urg(sk, timeo, msg, len, flags, addr_len);
1506 * State processing on a close. This implements the state shift for
1507 * sending our FIN frame. Note that we only send a FIN for some
1508 * states. A shutdown() may have already sent the FIN, or we may be
1512 static unsigned char new_state[16] = {
1513 /* current state: new state: action: */
1514 /* (Invalid) */ TCP_CLOSE,
1515 /* TCP_ESTABLISHED */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1516 /* TCP_SYN_SENT */ TCP_CLOSE,
1517 /* TCP_SYN_RECV */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1518 /* TCP_FIN_WAIT1 */ TCP_FIN_WAIT1,
1519 /* TCP_FIN_WAIT2 */ TCP_FIN_WAIT2,
1520 /* TCP_TIME_WAIT */ TCP_CLOSE,
1521 /* TCP_CLOSE */ TCP_CLOSE,
1522 /* TCP_CLOSE_WAIT */ TCP_LAST_ACK | TCP_ACTION_FIN,
1523 /* TCP_LAST_ACK */ TCP_LAST_ACK,
1524 /* TCP_LISTEN */ TCP_CLOSE,
1525 /* TCP_CLOSING */ TCP_CLOSING,
1528 static int tcp_close_state(struct sock *sk)
1530 int next = (int)new_state[sk->sk_state];
1531 int ns = next & TCP_STATE_MASK;
1533 tcp_set_state(sk, ns);
1535 return next & TCP_ACTION_FIN;
1539 * Shutdown the sending side of a connection. Much like close except
1540 * that we don't receive shut down or set_sock_flag(sk, SOCK_DEAD).
1543 void tcp_shutdown(struct sock *sk, int how)
1545 /* We need to grab some memory, and put together a FIN,
1546 * and then put it into the queue to be sent.
1547 * Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
1549 if (!(how & SEND_SHUTDOWN))
1552 /* If we've already sent a FIN, or it's a closed state, skip this. */
1553 if ((1 << sk->sk_state) &
1554 (TCPF_ESTABLISHED | TCPF_SYN_SENT |
1555 TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) {
1556 /* Clear out any half completed packets. FIN if needed. */
1557 if (tcp_close_state(sk))
1563 * At this point, there should be no process reference to this
1564 * socket, and thus no user references at all. Therefore we
1565 * can assume the socket waitqueue is inactive and nobody will
1566 * try to jump onto it.
1568 void tcp_destroy_sock(struct sock *sk)
1570 BUG_TRAP(sk->sk_state == TCP_CLOSE);
1571 BUG_TRAP(sock_flag(sk, SOCK_DEAD));
1573 /* It cannot be in hash table! */
1574 BUG_TRAP(sk_unhashed(sk));
1576 /* If it has not 0 inet_sk(sk)->num, it must be bound */
1577 BUG_TRAP(!inet_sk(sk)->num || tcp_sk(sk)->bind_hash);
1579 sk->sk_prot->destroy(sk);
1581 sk_stream_kill_queues(sk);
1583 xfrm_sk_free_policy(sk);
1585 #ifdef INET_REFCNT_DEBUG
1586 if (atomic_read(&sk->sk_refcnt) != 1) {
1587 printk(KERN_DEBUG "Destruction TCP %p delayed, c=%d\n",
1588 sk, atomic_read(&sk->sk_refcnt));
1592 atomic_dec(&tcp_orphan_count);
1596 void tcp_close(struct sock *sk, long timeout)
1598 struct sk_buff *skb;
1599 int data_was_unread = 0;
1602 sk->sk_shutdown = SHUTDOWN_MASK;
1604 if (sk->sk_state == TCP_LISTEN) {
1605 tcp_set_state(sk, TCP_CLOSE);
1608 tcp_listen_stop(sk);
1610 goto adjudge_to_death;
1613 /* We need to flush the recv. buffs. We do this only on the
1614 * descriptor close, not protocol-sourced closes, because the
1615 * reader process may not have drained the data yet!
1617 while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) {
1618 u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq -
1620 data_was_unread += len;
1624 sk_stream_mem_reclaim(sk);
1626 /* As outlined in draft-ietf-tcpimpl-prob-03.txt, section
1627 * 3.10, we send a RST here because data was lost. To
1628 * witness the awful effects of the old behavior of always
1629 * doing a FIN, run an older 2.1.x kernel or 2.0.x, start
1630 * a bulk GET in an FTP client, suspend the process, wait
1631 * for the client to advertise a zero window, then kill -9
1632 * the FTP client, wheee... Note: timeout is always zero
1635 if (data_was_unread) {
1636 /* Unread data was tossed, zap the connection. */
1637 NET_INC_STATS_USER(LINUX_MIB_TCPABORTONCLOSE);
1638 tcp_set_state(sk, TCP_CLOSE);
1639 tcp_send_active_reset(sk, GFP_KERNEL);
1640 } else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) {
1641 /* Check zero linger _after_ checking for unread data. */
1642 sk->sk_prot->disconnect(sk, 0);
1643 NET_INC_STATS_USER(LINUX_MIB_TCPABORTONDATA);
1644 } else if (tcp_close_state(sk)) {
1645 /* We FIN if the application ate all the data before
1646 * zapping the connection.
1649 /* RED-PEN. Formally speaking, we have broken TCP state
1650 * machine. State transitions:
1652 * TCP_ESTABLISHED -> TCP_FIN_WAIT1
1653 * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible)
1654 * TCP_CLOSE_WAIT -> TCP_LAST_ACK
1656 * are legal only when FIN has been sent (i.e. in window),
1657 * rather than queued out of window. Purists blame.
1659 * F.e. "RFC state" is ESTABLISHED,
1660 * if Linux state is FIN-WAIT-1, but FIN is still not sent.
1662 * The visible declinations are that sometimes
1663 * we enter time-wait state, when it is not required really
1664 * (harmless), do not send active resets, when they are
1665 * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
1666 * they look as CLOSING or LAST_ACK for Linux)
1667 * Probably, I missed some more holelets.
1673 sk_stream_wait_close(sk, timeout);
1676 /* It is the last release_sock in its life. It will remove backlog. */
1680 /* Now socket is owned by kernel and we acquire BH lock
1681 to finish close. No need to check for user refs.
1685 BUG_TRAP(!sock_owned_by_user(sk));
1690 /* This is a (useful) BSD violating of the RFC. There is a
1691 * problem with TCP as specified in that the other end could
1692 * keep a socket open forever with no application left this end.
1693 * We use a 3 minute timeout (about the same as BSD) then kill
1694 * our end. If they send after that then tough - BUT: long enough
1695 * that we won't make the old 4*rto = almost no time - whoops
1698 * Nope, it was not mistake. It is really desired behaviour
1699 * f.e. on http servers, when such sockets are useless, but
1700 * consume significant resources. Let's do it with special
1701 * linger2 option. --ANK
1704 if (sk->sk_state == TCP_FIN_WAIT2) {
1705 struct tcp_sock *tp = tcp_sk(sk);
1706 if (tp->linger2 < 0) {
1707 tcp_set_state(sk, TCP_CLOSE);
1708 tcp_send_active_reset(sk, GFP_ATOMIC);
1709 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONLINGER);
1711 int tmo = tcp_fin_time(tp);
1713 if (tmo > TCP_TIMEWAIT_LEN) {
1714 tcp_reset_keepalive_timer(sk, tcp_fin_time(tp));
1716 atomic_inc(&tcp_orphan_count);
1717 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
1722 if (sk->sk_state != TCP_CLOSE) {
1723 sk_stream_mem_reclaim(sk);
1724 if (atomic_read(&tcp_orphan_count) > sysctl_tcp_max_orphans ||
1725 (sk->sk_wmem_queued > SOCK_MIN_SNDBUF &&
1726 atomic_read(&tcp_memory_allocated) > sysctl_tcp_mem[2])) {
1727 if (net_ratelimit())
1728 printk(KERN_INFO "TCP: too many of orphaned "
1730 tcp_set_state(sk, TCP_CLOSE);
1731 tcp_send_active_reset(sk, GFP_ATOMIC);
1732 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONMEMORY);
1735 atomic_inc(&tcp_orphan_count);
1737 if (sk->sk_state == TCP_CLOSE)
1738 tcp_destroy_sock(sk);
1739 /* Otherwise, socket is reprieved until protocol close. */
1747 /* These states need RST on ABORT according to RFC793 */
1749 static inline int tcp_need_reset(int state)
1751 return (1 << state) &
1752 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 |
1753 TCPF_FIN_WAIT2 | TCPF_SYN_RECV);
1756 int tcp_disconnect(struct sock *sk, int flags)
1758 struct inet_sock *inet = inet_sk(sk);
1759 struct tcp_sock *tp = tcp_sk(sk);
1761 int old_state = sk->sk_state;
1763 if (old_state != TCP_CLOSE)
1764 tcp_set_state(sk, TCP_CLOSE);
1766 /* ABORT function of RFC793 */
1767 if (old_state == TCP_LISTEN) {
1768 tcp_listen_stop(sk);
1769 } else if (tcp_need_reset(old_state) ||
1770 (tp->snd_nxt != tp->write_seq &&
1771 (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK))) {
1772 /* The last check adjusts for discrepance of Linux wrt. RFC
1775 tcp_send_active_reset(sk, gfp_any());
1776 sk->sk_err = ECONNRESET;
1777 } else if (old_state == TCP_SYN_SENT)
1778 sk->sk_err = ECONNRESET;
1780 tcp_clear_xmit_timers(sk);
1781 __skb_queue_purge(&sk->sk_receive_queue);
1782 sk_stream_writequeue_purge(sk);
1783 __skb_queue_purge(&tp->out_of_order_queue);
1787 if (!(sk->sk_userlocks & SOCK_BINDADDR_LOCK))
1788 inet_reset_saddr(sk);
1790 sk->sk_shutdown = 0;
1791 sock_reset_flag(sk, SOCK_DONE);
1793 if ((tp->write_seq += tp->max_window + 2) == 0)
1798 tp->packets_out = 0;
1799 tp->snd_ssthresh = 0x7fffffff;
1800 tp->snd_cwnd_cnt = 0;
1801 tcp_set_ca_state(tp, TCP_CA_Open);
1802 tcp_clear_retrans(tp);
1803 tcp_delack_init(tp);
1804 sk->sk_send_head = NULL;
1805 tp->rx_opt.saw_tstamp = 0;
1806 tcp_sack_reset(&tp->rx_opt);
1809 BUG_TRAP(!inet->num || tp->bind_hash);
1811 sk->sk_error_report(sk);
1816 * Wait for an incoming connection, avoid race
1817 * conditions. This must be called with the socket locked.
1819 static int wait_for_connect(struct sock *sk, long timeo)
1821 struct tcp_sock *tp = tcp_sk(sk);
1826 * True wake-one mechanism for incoming connections: only
1827 * one process gets woken up, not the 'whole herd'.
1828 * Since we do not 'race & poll' for established sockets
1829 * anymore, the common case will execute the loop only once.
1831 * Subtle issue: "add_wait_queue_exclusive()" will be added
1832 * after any current non-exclusive waiters, and we know that
1833 * it will always _stay_ after any new non-exclusive waiters
1834 * because all non-exclusive waiters are added at the
1835 * beginning of the wait-queue. As such, it's ok to "drop"
1836 * our exclusiveness temporarily when we get woken up without
1837 * having to remove and re-insert us on the wait queue.
1840 prepare_to_wait_exclusive(sk->sk_sleep, &wait,
1841 TASK_INTERRUPTIBLE);
1843 if (reqsk_queue_empty(&tp->accept_queue))
1844 timeo = schedule_timeout(timeo);
1847 if (!reqsk_queue_empty(&tp->accept_queue))
1850 if (sk->sk_state != TCP_LISTEN)
1852 err = sock_intr_errno(timeo);
1853 if (signal_pending(current))
1859 finish_wait(sk->sk_sleep, &wait);
1864 * This will accept the next outstanding connection.
1867 struct sock *tcp_accept(struct sock *sk, int flags, int *err)
1869 struct tcp_sock *tp = tcp_sk(sk);
1875 /* We need to make sure that this socket is listening,
1876 * and that it has something pending.
1879 if (sk->sk_state != TCP_LISTEN)
1882 /* Find already established connection */
1883 if (reqsk_queue_empty(&tp->accept_queue)) {
1884 long timeo = sock_rcvtimeo(sk, flags & O_NONBLOCK);
1886 /* If this is a non blocking socket don't sleep */
1891 error = wait_for_connect(sk, timeo);
1896 newsk = reqsk_queue_get_child(&tp->accept_queue, sk);
1897 BUG_TRAP(newsk->sk_state != TCP_SYN_RECV);
1908 * Socket option code for TCP.
1910 int tcp_setsockopt(struct sock *sk, int level, int optname, char __user *optval,
1913 struct tcp_sock *tp = tcp_sk(sk);
1917 if (level != SOL_TCP)
1918 return tp->af_specific->setsockopt(sk, level, optname,
1921 /* This is a string value all the others are int's */
1922 if (optname == TCP_CONGESTION) {
1923 char name[TCP_CA_NAME_MAX];
1928 val = strncpy_from_user(name, optval,
1929 min(TCP_CA_NAME_MAX-1, optlen));
1935 err = tcp_set_congestion_control(tp, name);
1940 if (optlen < sizeof(int))
1943 if (get_user(val, (int __user *)optval))
1950 /* Values greater than interface MTU won't take effect. However
1951 * at the point when this call is done we typically don't yet
1952 * know which interface is going to be used */
1953 if (val < 8 || val > MAX_TCP_WINDOW) {
1957 tp->rx_opt.user_mss = val;
1962 /* TCP_NODELAY is weaker than TCP_CORK, so that
1963 * this option on corked socket is remembered, but
1964 * it is not activated until cork is cleared.
1966 * However, when TCP_NODELAY is set we make
1967 * an explicit push, which overrides even TCP_CORK
1968 * for currently queued segments.
1970 tp->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH;
1971 tcp_push_pending_frames(sk, tp);
1973 tp->nonagle &= ~TCP_NAGLE_OFF;
1978 /* When set indicates to always queue non-full frames.
1979 * Later the user clears this option and we transmit
1980 * any pending partial frames in the queue. This is
1981 * meant to be used alongside sendfile() to get properly
1982 * filled frames when the user (for example) must write
1983 * out headers with a write() call first and then use
1984 * sendfile to send out the data parts.
1986 * TCP_CORK can be set together with TCP_NODELAY and it is
1987 * stronger than TCP_NODELAY.
1990 tp->nonagle |= TCP_NAGLE_CORK;
1992 tp->nonagle &= ~TCP_NAGLE_CORK;
1993 if (tp->nonagle&TCP_NAGLE_OFF)
1994 tp->nonagle |= TCP_NAGLE_PUSH;
1995 tcp_push_pending_frames(sk, tp);
2000 if (val < 1 || val > MAX_TCP_KEEPIDLE)
2003 tp->keepalive_time = val * HZ;
2004 if (sock_flag(sk, SOCK_KEEPOPEN) &&
2005 !((1 << sk->sk_state) &
2006 (TCPF_CLOSE | TCPF_LISTEN))) {
2007 __u32 elapsed = tcp_time_stamp - tp->rcv_tstamp;
2008 if (tp->keepalive_time > elapsed)
2009 elapsed = tp->keepalive_time - elapsed;
2012 tcp_reset_keepalive_timer(sk, elapsed);
2017 if (val < 1 || val > MAX_TCP_KEEPINTVL)
2020 tp->keepalive_intvl = val * HZ;
2023 if (val < 1 || val > MAX_TCP_KEEPCNT)
2026 tp->keepalive_probes = val;
2029 if (val < 1 || val > MAX_TCP_SYNCNT)
2032 tp->syn_retries = val;
2038 else if (val > sysctl_tcp_fin_timeout / HZ)
2041 tp->linger2 = val * HZ;
2044 case TCP_DEFER_ACCEPT:
2045 tp->defer_accept = 0;
2047 /* Translate value in seconds to number of
2049 while (tp->defer_accept < 32 &&
2050 val > ((TCP_TIMEOUT_INIT / HZ) <<
2057 case TCP_WINDOW_CLAMP:
2059 if (sk->sk_state != TCP_CLOSE) {
2063 tp->window_clamp = 0;
2065 tp->window_clamp = val < SOCK_MIN_RCVBUF / 2 ?
2066 SOCK_MIN_RCVBUF / 2 : val;
2071 tp->ack.pingpong = 1;
2073 tp->ack.pingpong = 0;
2074 if ((1 << sk->sk_state) &
2075 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) &&
2076 tcp_ack_scheduled(tp)) {
2077 tp->ack.pending |= TCP_ACK_PUSHED;
2078 cleanup_rbuf(sk, 1);
2080 tp->ack.pingpong = 1;
2093 /* Return information about state of tcp endpoint in API format. */
2094 void tcp_get_info(struct sock *sk, struct tcp_info *info)
2096 struct tcp_sock *tp = tcp_sk(sk);
2097 u32 now = tcp_time_stamp;
2099 memset(info, 0, sizeof(*info));
2101 info->tcpi_state = sk->sk_state;
2102 info->tcpi_ca_state = tp->ca_state;
2103 info->tcpi_retransmits = tp->retransmits;
2104 info->tcpi_probes = tp->probes_out;
2105 info->tcpi_backoff = tp->backoff;
2107 if (tp->rx_opt.tstamp_ok)
2108 info->tcpi_options |= TCPI_OPT_TIMESTAMPS;
2109 if (tp->rx_opt.sack_ok)
2110 info->tcpi_options |= TCPI_OPT_SACK;
2111 if (tp->rx_opt.wscale_ok) {
2112 info->tcpi_options |= TCPI_OPT_WSCALE;
2113 info->tcpi_snd_wscale = tp->rx_opt.snd_wscale;
2114 info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale;
2117 if (tp->ecn_flags&TCP_ECN_OK)
2118 info->tcpi_options |= TCPI_OPT_ECN;
2120 info->tcpi_rto = jiffies_to_usecs(tp->rto);
2121 info->tcpi_ato = jiffies_to_usecs(tp->ack.ato);
2122 info->tcpi_snd_mss = tp->mss_cache_std;
2123 info->tcpi_rcv_mss = tp->ack.rcv_mss;
2125 info->tcpi_unacked = tp->packets_out;
2126 info->tcpi_sacked = tp->sacked_out;
2127 info->tcpi_lost = tp->lost_out;
2128 info->tcpi_retrans = tp->retrans_out;
2129 info->tcpi_fackets = tp->fackets_out;
2131 info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime);
2132 info->tcpi_last_data_recv = jiffies_to_msecs(now - tp->ack.lrcvtime);
2133 info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp);
2135 info->tcpi_pmtu = tp->pmtu_cookie;
2136 info->tcpi_rcv_ssthresh = tp->rcv_ssthresh;
2137 info->tcpi_rtt = jiffies_to_usecs(tp->srtt)>>3;
2138 info->tcpi_rttvar = jiffies_to_usecs(tp->mdev)>>2;
2139 info->tcpi_snd_ssthresh = tp->snd_ssthresh;
2140 info->tcpi_snd_cwnd = tp->snd_cwnd;
2141 info->tcpi_advmss = tp->advmss;
2142 info->tcpi_reordering = tp->reordering;
2144 info->tcpi_rcv_rtt = jiffies_to_usecs(tp->rcv_rtt_est.rtt)>>3;
2145 info->tcpi_rcv_space = tp->rcvq_space.space;
2147 info->tcpi_total_retrans = tp->total_retrans;
2150 EXPORT_SYMBOL_GPL(tcp_get_info);
2152 int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval,
2155 struct tcp_sock *tp = tcp_sk(sk);
2158 if (level != SOL_TCP)
2159 return tp->af_specific->getsockopt(sk, level, optname,
2162 if (get_user(len, optlen))
2165 len = min_t(unsigned int, len, sizeof(int));
2172 val = tp->mss_cache_std;
2173 if (!val && ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
2174 val = tp->rx_opt.user_mss;
2177 val = !!(tp->nonagle&TCP_NAGLE_OFF);
2180 val = !!(tp->nonagle&TCP_NAGLE_CORK);
2183 val = (tp->keepalive_time ? : sysctl_tcp_keepalive_time) / HZ;
2186 val = (tp->keepalive_intvl ? : sysctl_tcp_keepalive_intvl) / HZ;
2189 val = tp->keepalive_probes ? : sysctl_tcp_keepalive_probes;
2192 val = tp->syn_retries ? : sysctl_tcp_syn_retries;
2197 val = (val ? : sysctl_tcp_fin_timeout) / HZ;
2199 case TCP_DEFER_ACCEPT:
2200 val = !tp->defer_accept ? 0 : ((TCP_TIMEOUT_INIT / HZ) <<
2201 (tp->defer_accept - 1));
2203 case TCP_WINDOW_CLAMP:
2204 val = tp->window_clamp;
2207 struct tcp_info info;
2209 if (get_user(len, optlen))
2212 tcp_get_info(sk, &info);
2214 len = min_t(unsigned int, len, sizeof(info));
2215 if (put_user(len, optlen))
2217 if (copy_to_user(optval, &info, len))
2222 val = !tp->ack.pingpong;
2225 case TCP_CONGESTION:
2226 if (get_user(len, optlen))
2228 len = min_t(unsigned int, len, TCP_CA_NAME_MAX);
2229 if (put_user(len, optlen))
2231 if (copy_to_user(optval, tp->ca_ops->name, len))
2235 return -ENOPROTOOPT;
2238 if (put_user(len, optlen))
2240 if (copy_to_user(optval, &val, len))
2246 extern void __skb_cb_too_small_for_tcp(int, int);
2247 extern struct tcp_congestion_ops tcp_reno;
2249 static __initdata unsigned long thash_entries;
2250 static int __init set_thash_entries(char *str)
2254 thash_entries = simple_strtoul(str, &str, 0);
2257 __setup("thash_entries=", set_thash_entries);
2259 void __init tcp_init(void)
2261 struct sk_buff *skb = NULL;
2264 if (sizeof(struct tcp_skb_cb) > sizeof(skb->cb))
2265 __skb_cb_too_small_for_tcp(sizeof(struct tcp_skb_cb),
2268 tcp_bucket_cachep = kmem_cache_create("tcp_bind_bucket",
2269 sizeof(struct tcp_bind_bucket),
2270 0, SLAB_HWCACHE_ALIGN,
2272 if (!tcp_bucket_cachep)
2273 panic("tcp_init: Cannot alloc tcp_bind_bucket cache.");
2275 tcp_timewait_cachep = kmem_cache_create("tcp_tw_bucket",
2276 sizeof(struct tcp_tw_bucket),
2277 0, SLAB_HWCACHE_ALIGN,
2279 if (!tcp_timewait_cachep)
2280 panic("tcp_init: Cannot alloc tcp_tw_bucket cache.");
2282 /* Size and allocate the main established and bind bucket
2285 * The methodology is similar to that of the buffer cache.
2287 tcp_ehash = (struct tcp_ehash_bucket *)
2288 alloc_large_system_hash("TCP established",
2289 sizeof(struct tcp_ehash_bucket),
2291 (num_physpages >= 128 * 1024) ?
2298 tcp_ehash_size = (1 << tcp_ehash_size) >> 1;
2299 for (i = 0; i < (tcp_ehash_size << 1); i++) {
2300 rwlock_init(&tcp_ehash[i].lock);
2301 INIT_HLIST_HEAD(&tcp_ehash[i].chain);
2304 tcp_bhash = (struct tcp_bind_hashbucket *)
2305 alloc_large_system_hash("TCP bind",
2306 sizeof(struct tcp_bind_hashbucket),
2308 (num_physpages >= 128 * 1024) ?
2315 tcp_bhash_size = 1 << tcp_bhash_size;
2316 for (i = 0; i < tcp_bhash_size; i++) {
2317 spin_lock_init(&tcp_bhash[i].lock);
2318 INIT_HLIST_HEAD(&tcp_bhash[i].chain);
2321 /* Try to be a bit smarter and adjust defaults depending
2322 * on available memory.
2324 for (order = 0; ((1 << order) << PAGE_SHIFT) <
2325 (tcp_bhash_size * sizeof(struct tcp_bind_hashbucket));
2329 sysctl_local_port_range[0] = 32768;
2330 sysctl_local_port_range[1] = 61000;
2331 sysctl_tcp_max_tw_buckets = 180000;
2332 sysctl_tcp_max_orphans = 4096 << (order - 4);
2333 sysctl_max_syn_backlog = 1024;
2334 } else if (order < 3) {
2335 sysctl_local_port_range[0] = 1024 * (3 - order);
2336 sysctl_tcp_max_tw_buckets >>= (3 - order);
2337 sysctl_tcp_max_orphans >>= (3 - order);
2338 sysctl_max_syn_backlog = 128;
2340 tcp_port_rover = sysctl_local_port_range[0] - 1;
2342 sysctl_tcp_mem[0] = 768 << order;
2343 sysctl_tcp_mem[1] = 1024 << order;
2344 sysctl_tcp_mem[2] = 1536 << order;
2347 sysctl_tcp_wmem[2] = 64 * 1024;
2348 sysctl_tcp_rmem[0] = PAGE_SIZE;
2349 sysctl_tcp_rmem[1] = 43689;
2350 sysctl_tcp_rmem[2] = 2 * 43689;
2353 printk(KERN_INFO "TCP: Hash tables configured "
2354 "(established %d bind %d)\n",
2355 tcp_ehash_size << 1, tcp_bhash_size);
2357 tcp_register_congestion_control(&tcp_reno);
2360 EXPORT_SYMBOL(tcp_accept);
2361 EXPORT_SYMBOL(tcp_close);
2362 EXPORT_SYMBOL(tcp_destroy_sock);
2363 EXPORT_SYMBOL(tcp_disconnect);
2364 EXPORT_SYMBOL(tcp_getsockopt);
2365 EXPORT_SYMBOL(tcp_ioctl);
2366 EXPORT_SYMBOL(tcp_poll);
2367 EXPORT_SYMBOL(tcp_read_sock);
2368 EXPORT_SYMBOL(tcp_recvmsg);
2369 EXPORT_SYMBOL(tcp_sendmsg);
2370 EXPORT_SYMBOL(tcp_sendpage);
2371 EXPORT_SYMBOL(tcp_setsockopt);
2372 EXPORT_SYMBOL(tcp_shutdown);
2373 EXPORT_SYMBOL(tcp_statistics);
2374 EXPORT_SYMBOL(tcp_timewait_cachep);