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 request_sock *acc_req;
503 struct request_sock *req;
505 tcp_delete_keepalive_timer(sk);
507 /* make all the listen_opt local to us */
508 acc_req = reqsk_queue_yank_acceptq(&tp->accept_queue);
510 /* Following specs, it would be better either to send FIN
511 * (and enter FIN-WAIT-1, it is normal close)
512 * or to send active reset (abort).
513 * Certainly, it is pretty dangerous while synflood, but it is
514 * bad justification for our negligence 8)
515 * To be honest, we are not able to make either
516 * of the variants now. --ANK
518 reqsk_queue_destroy(&tp->accept_queue);
520 while ((req = acc_req) != NULL) {
521 struct sock *child = req->sk;
523 acc_req = req->dl_next;
527 BUG_TRAP(!sock_owned_by_user(child));
530 tcp_disconnect(child, O_NONBLOCK);
534 atomic_inc(&tcp_orphan_count);
536 tcp_destroy_sock(child);
538 bh_unlock_sock(child);
542 sk_acceptq_removed(sk);
545 BUG_TRAP(!sk->sk_ack_backlog);
548 static inline void tcp_mark_push(struct tcp_sock *tp, struct sk_buff *skb)
550 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH;
551 tp->pushed_seq = tp->write_seq;
554 static inline int forced_push(struct tcp_sock *tp)
556 return after(tp->write_seq, tp->pushed_seq + (tp->max_window >> 1));
559 static inline void skb_entail(struct sock *sk, struct tcp_sock *tp,
563 TCP_SKB_CB(skb)->seq = tp->write_seq;
564 TCP_SKB_CB(skb)->end_seq = tp->write_seq;
565 TCP_SKB_CB(skb)->flags = TCPCB_FLAG_ACK;
566 TCP_SKB_CB(skb)->sacked = 0;
567 skb_header_release(skb);
568 __skb_queue_tail(&sk->sk_write_queue, skb);
569 sk_charge_skb(sk, skb);
570 if (!sk->sk_send_head)
571 sk->sk_send_head = skb;
572 if (tp->nonagle & TCP_NAGLE_PUSH)
573 tp->nonagle &= ~TCP_NAGLE_PUSH;
576 static inline void tcp_mark_urg(struct tcp_sock *tp, int flags,
579 if (flags & MSG_OOB) {
581 tp->snd_up = tp->write_seq;
582 TCP_SKB_CB(skb)->sacked |= TCPCB_URG;
586 static inline void tcp_push(struct sock *sk, struct tcp_sock *tp, int flags,
587 int mss_now, int nonagle)
589 if (sk->sk_send_head) {
590 struct sk_buff *skb = sk->sk_write_queue.prev;
591 if (!(flags & MSG_MORE) || forced_push(tp))
592 tcp_mark_push(tp, skb);
593 tcp_mark_urg(tp, flags, skb);
594 __tcp_push_pending_frames(sk, tp, mss_now,
595 (flags & MSG_MORE) ? TCP_NAGLE_CORK : nonagle);
599 static ssize_t do_tcp_sendpages(struct sock *sk, struct page **pages, int poffset,
600 size_t psize, int flags)
602 struct tcp_sock *tp = tcp_sk(sk);
603 int mss_now, size_goal;
606 long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
608 /* Wait for a connection to finish. */
609 if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
610 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
613 clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
615 mss_now = tcp_current_mss(sk, !(flags&MSG_OOB));
616 size_goal = tp->xmit_size_goal;
620 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
624 struct sk_buff *skb = sk->sk_write_queue.prev;
625 struct page *page = pages[poffset / PAGE_SIZE];
626 int copy, i, can_coalesce;
627 int offset = poffset % PAGE_SIZE;
628 int size = min_t(size_t, psize, PAGE_SIZE - offset);
630 if (!sk->sk_send_head || (copy = size_goal - skb->len) <= 0) {
632 if (!sk_stream_memory_free(sk))
633 goto wait_for_sndbuf;
635 skb = sk_stream_alloc_pskb(sk, 0, 0,
638 goto wait_for_memory;
640 skb_entail(sk, tp, skb);
647 i = skb_shinfo(skb)->nr_frags;
648 can_coalesce = skb_can_coalesce(skb, i, page, offset);
649 if (!can_coalesce && i >= MAX_SKB_FRAGS) {
650 tcp_mark_push(tp, skb);
653 if (sk->sk_forward_alloc < copy &&
654 !sk_stream_mem_schedule(sk, copy, 0))
655 goto wait_for_memory;
658 skb_shinfo(skb)->frags[i - 1].size += copy;
661 skb_fill_page_desc(skb, i, page, offset, copy);
665 skb->data_len += copy;
666 skb->truesize += copy;
667 sk->sk_wmem_queued += copy;
668 sk->sk_forward_alloc -= copy;
669 skb->ip_summed = CHECKSUM_HW;
670 tp->write_seq += copy;
671 TCP_SKB_CB(skb)->end_seq += copy;
672 skb_shinfo(skb)->tso_segs = 0;
675 TCP_SKB_CB(skb)->flags &= ~TCPCB_FLAG_PSH;
679 if (!(psize -= copy))
682 if (skb->len < mss_now || (flags & MSG_OOB))
685 if (forced_push(tp)) {
686 tcp_mark_push(tp, skb);
687 __tcp_push_pending_frames(sk, tp, mss_now, TCP_NAGLE_PUSH);
688 } else if (skb == sk->sk_send_head)
689 tcp_push_one(sk, mss_now);
693 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
696 tcp_push(sk, tp, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH);
698 if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
701 mss_now = tcp_current_mss(sk, !(flags&MSG_OOB));
702 size_goal = tp->xmit_size_goal;
707 tcp_push(sk, tp, flags, mss_now, tp->nonagle);
714 return sk_stream_error(sk, flags, err);
717 ssize_t tcp_sendpage(struct socket *sock, struct page *page, int offset,
718 size_t size, int flags)
721 struct sock *sk = sock->sk;
723 #define TCP_ZC_CSUM_FLAGS (NETIF_F_IP_CSUM | NETIF_F_NO_CSUM | NETIF_F_HW_CSUM)
725 if (!(sk->sk_route_caps & NETIF_F_SG) ||
726 !(sk->sk_route_caps & TCP_ZC_CSUM_FLAGS))
727 return sock_no_sendpage(sock, page, offset, size, flags);
729 #undef TCP_ZC_CSUM_FLAGS
733 res = do_tcp_sendpages(sk, &page, offset, size, flags);
739 #define TCP_PAGE(sk) (sk->sk_sndmsg_page)
740 #define TCP_OFF(sk) (sk->sk_sndmsg_off)
742 static inline int select_size(struct sock *sk, struct tcp_sock *tp)
744 int tmp = tp->mss_cache;
746 if (sk->sk_route_caps & NETIF_F_SG) {
747 if (sk->sk_route_caps & NETIF_F_TSO)
750 int pgbreak = SKB_MAX_HEAD(MAX_TCP_HEADER);
752 if (tmp >= pgbreak &&
753 tmp <= pgbreak + (MAX_SKB_FRAGS - 1) * PAGE_SIZE)
761 int tcp_sendmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg,
765 struct tcp_sock *tp = tcp_sk(sk);
768 int mss_now, size_goal;
775 flags = msg->msg_flags;
776 timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
778 /* Wait for a connection to finish. */
779 if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
780 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
783 /* This should be in poll */
784 clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
786 mss_now = tcp_current_mss(sk, !(flags&MSG_OOB));
787 size_goal = tp->xmit_size_goal;
789 /* Ok commence sending. */
790 iovlen = msg->msg_iovlen;
795 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
798 while (--iovlen >= 0) {
799 int seglen = iov->iov_len;
800 unsigned char __user *from = iov->iov_base;
807 skb = sk->sk_write_queue.prev;
809 if (!sk->sk_send_head ||
810 (copy = size_goal - skb->len) <= 0) {
813 /* Allocate new segment. If the interface is SG,
814 * allocate skb fitting to single page.
816 if (!sk_stream_memory_free(sk))
817 goto wait_for_sndbuf;
819 skb = sk_stream_alloc_pskb(sk, select_size(sk, tp),
820 0, sk->sk_allocation);
822 goto wait_for_memory;
825 * Check whether we can use HW checksum.
827 if (sk->sk_route_caps &
828 (NETIF_F_IP_CSUM | NETIF_F_NO_CSUM |
830 skb->ip_summed = CHECKSUM_HW;
832 skb_entail(sk, tp, skb);
836 /* Try to append data to the end of skb. */
840 /* Where to copy to? */
841 if (skb_tailroom(skb) > 0) {
842 /* We have some space in skb head. Superb! */
843 if (copy > skb_tailroom(skb))
844 copy = skb_tailroom(skb);
845 if ((err = skb_add_data(skb, from, copy)) != 0)
849 int i = skb_shinfo(skb)->nr_frags;
850 struct page *page = TCP_PAGE(sk);
851 int off = TCP_OFF(sk);
853 if (skb_can_coalesce(skb, i, page, off) &&
855 /* We can extend the last page
858 } else if (i == MAX_SKB_FRAGS ||
860 !(sk->sk_route_caps & NETIF_F_SG))) {
861 /* Need to add new fragment and cannot
862 * do this because interface is non-SG,
863 * or because all the page slots are
865 tcp_mark_push(tp, skb);
868 if (off == PAGE_SIZE) {
870 TCP_PAGE(sk) = page = NULL;
875 /* Allocate new cache page. */
876 if (!(page = sk_stream_alloc_page(sk)))
877 goto wait_for_memory;
881 if (copy > PAGE_SIZE - off)
882 copy = PAGE_SIZE - off;
884 /* Time to copy data. We are close to
886 err = skb_copy_to_page(sk, from, skb, page,
889 /* If this page was new, give it to the
890 * socket so it does not get leaked.
899 /* Update the skb. */
901 skb_shinfo(skb)->frags[i - 1].size +=
904 skb_fill_page_desc(skb, i, page, off, copy);
907 } else if (off + copy < PAGE_SIZE) {
913 TCP_OFF(sk) = off + copy;
917 TCP_SKB_CB(skb)->flags &= ~TCPCB_FLAG_PSH;
919 tp->write_seq += copy;
920 TCP_SKB_CB(skb)->end_seq += copy;
921 skb_shinfo(skb)->tso_segs = 0;
925 if ((seglen -= copy) == 0 && iovlen == 0)
928 if (skb->len < mss_now || (flags & MSG_OOB))
931 if (forced_push(tp)) {
932 tcp_mark_push(tp, skb);
933 __tcp_push_pending_frames(sk, tp, mss_now, TCP_NAGLE_PUSH);
934 } else if (skb == sk->sk_send_head)
935 tcp_push_one(sk, mss_now);
939 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
942 tcp_push(sk, tp, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH);
944 if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
947 mss_now = tcp_current_mss(sk, !(flags&MSG_OOB));
948 size_goal = tp->xmit_size_goal;
954 tcp_push(sk, tp, flags, mss_now, tp->nonagle);
961 if (sk->sk_send_head == skb)
962 sk->sk_send_head = NULL;
963 __skb_unlink(skb, &sk->sk_write_queue);
964 sk_stream_free_skb(sk, skb);
971 err = sk_stream_error(sk, flags, err);
978 * Handle reading urgent data. BSD has very simple semantics for
979 * this, no blocking and very strange errors 8)
982 static int tcp_recv_urg(struct sock *sk, long timeo,
983 struct msghdr *msg, int len, int flags,
986 struct tcp_sock *tp = tcp_sk(sk);
988 /* No URG data to read. */
989 if (sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data ||
990 tp->urg_data == TCP_URG_READ)
991 return -EINVAL; /* Yes this is right ! */
993 if (sk->sk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DONE))
996 if (tp->urg_data & TCP_URG_VALID) {
998 char c = tp->urg_data;
1000 if (!(flags & MSG_PEEK))
1001 tp->urg_data = TCP_URG_READ;
1003 /* Read urgent data. */
1004 msg->msg_flags |= MSG_OOB;
1007 if (!(flags & MSG_TRUNC))
1008 err = memcpy_toiovec(msg->msg_iov, &c, 1);
1011 msg->msg_flags |= MSG_TRUNC;
1013 return err ? -EFAULT : len;
1016 if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
1019 /* Fixed the recv(..., MSG_OOB) behaviour. BSD docs and
1020 * the available implementations agree in this case:
1021 * this call should never block, independent of the
1022 * blocking state of the socket.
1023 * Mike <pall@rz.uni-karlsruhe.de>
1028 /* Clean up the receive buffer for full frames taken by the user,
1029 * then send an ACK if necessary. COPIED is the number of bytes
1030 * tcp_recvmsg has given to the user so far, it speeds up the
1031 * calculation of whether or not we must ACK for the sake of
1034 static void cleanup_rbuf(struct sock *sk, int copied)
1036 struct tcp_sock *tp = tcp_sk(sk);
1037 int time_to_ack = 0;
1040 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
1042 BUG_TRAP(!skb || before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq));
1045 if (tcp_ack_scheduled(tp)) {
1046 /* Delayed ACKs frequently hit locked sockets during bulk
1048 if (tp->ack.blocked ||
1049 /* Once-per-two-segments ACK was not sent by tcp_input.c */
1050 tp->rcv_nxt - tp->rcv_wup > tp->ack.rcv_mss ||
1052 * If this read emptied read buffer, we send ACK, if
1053 * connection is not bidirectional, user drained
1054 * receive buffer and there was a small segment
1057 (copied > 0 && (tp->ack.pending & TCP_ACK_PUSHED) &&
1058 !tp->ack.pingpong && !atomic_read(&sk->sk_rmem_alloc)))
1062 /* We send an ACK if we can now advertise a non-zero window
1063 * which has been raised "significantly".
1065 * Even if window raised up to infinity, do not send window open ACK
1066 * in states, where we will not receive more. It is useless.
1068 if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) {
1069 __u32 rcv_window_now = tcp_receive_window(tp);
1071 /* Optimize, __tcp_select_window() is not cheap. */
1072 if (2*rcv_window_now <= tp->window_clamp) {
1073 __u32 new_window = __tcp_select_window(sk);
1075 /* Send ACK now, if this read freed lots of space
1076 * in our buffer. Certainly, new_window is new window.
1077 * We can advertise it now, if it is not less than current one.
1078 * "Lots" means "at least twice" here.
1080 if (new_window && new_window >= 2 * rcv_window_now)
1088 static void tcp_prequeue_process(struct sock *sk)
1090 struct sk_buff *skb;
1091 struct tcp_sock *tp = tcp_sk(sk);
1093 NET_INC_STATS_USER(LINUX_MIB_TCPPREQUEUED);
1095 /* RX process wants to run with disabled BHs, though it is not
1098 while ((skb = __skb_dequeue(&tp->ucopy.prequeue)) != NULL)
1099 sk->sk_backlog_rcv(sk, skb);
1102 /* Clear memory counter. */
1103 tp->ucopy.memory = 0;
1106 static inline struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off)
1108 struct sk_buff *skb;
1111 skb_queue_walk(&sk->sk_receive_queue, skb) {
1112 offset = seq - TCP_SKB_CB(skb)->seq;
1115 if (offset < skb->len || skb->h.th->fin) {
1124 * This routine provides an alternative to tcp_recvmsg() for routines
1125 * that would like to handle copying from skbuffs directly in 'sendfile'
1128 * - It is assumed that the socket was locked by the caller.
1129 * - The routine does not block.
1130 * - At present, there is no support for reading OOB data
1131 * or for 'peeking' the socket using this routine
1132 * (although both would be easy to implement).
1134 int tcp_read_sock(struct sock *sk, read_descriptor_t *desc,
1135 sk_read_actor_t recv_actor)
1137 struct sk_buff *skb;
1138 struct tcp_sock *tp = tcp_sk(sk);
1139 u32 seq = tp->copied_seq;
1143 if (sk->sk_state == TCP_LISTEN)
1145 while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1146 if (offset < skb->len) {
1149 len = skb->len - offset;
1150 /* Stop reading if we hit a patch of urgent data */
1152 u32 urg_offset = tp->urg_seq - seq;
1153 if (urg_offset < len)
1158 used = recv_actor(desc, skb, offset, len);
1164 if (offset != skb->len)
1167 if (skb->h.th->fin) {
1168 sk_eat_skb(sk, skb);
1172 sk_eat_skb(sk, skb);
1176 tp->copied_seq = seq;
1178 tcp_rcv_space_adjust(sk);
1180 /* Clean up data we have read: This will do ACK frames. */
1182 cleanup_rbuf(sk, copied);
1187 * This routine copies from a sock struct into the user buffer.
1189 * Technical note: in 2.3 we work on _locked_ socket, so that
1190 * tricks with *seq access order and skb->users are not required.
1191 * Probably, code can be easily improved even more.
1194 int tcp_recvmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg,
1195 size_t len, int nonblock, int flags, int *addr_len)
1197 struct tcp_sock *tp = tcp_sk(sk);
1203 int target; /* Read at least this many bytes */
1205 struct task_struct *user_recv = NULL;
1209 TCP_CHECK_TIMER(sk);
1212 if (sk->sk_state == TCP_LISTEN)
1215 timeo = sock_rcvtimeo(sk, nonblock);
1217 /* Urgent data needs to be handled specially. */
1218 if (flags & MSG_OOB)
1221 seq = &tp->copied_seq;
1222 if (flags & MSG_PEEK) {
1223 peek_seq = tp->copied_seq;
1227 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
1230 struct sk_buff *skb;
1233 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
1234 if (tp->urg_data && tp->urg_seq == *seq) {
1237 if (signal_pending(current)) {
1238 copied = timeo ? sock_intr_errno(timeo) : -EAGAIN;
1243 /* Next get a buffer. */
1245 skb = skb_peek(&sk->sk_receive_queue);
1250 /* Now that we have two receive queues this
1253 if (before(*seq, TCP_SKB_CB(skb)->seq)) {
1254 printk(KERN_INFO "recvmsg bug: copied %X "
1255 "seq %X\n", *seq, TCP_SKB_CB(skb)->seq);
1258 offset = *seq - TCP_SKB_CB(skb)->seq;
1261 if (offset < skb->len)
1265 BUG_TRAP(flags & MSG_PEEK);
1267 } while (skb != (struct sk_buff *)&sk->sk_receive_queue);
1269 /* Well, if we have backlog, try to process it now yet. */
1271 if (copied >= target && !sk->sk_backlog.tail)
1276 sk->sk_state == TCP_CLOSE ||
1277 (sk->sk_shutdown & RCV_SHUTDOWN) ||
1279 signal_pending(current) ||
1283 if (sock_flag(sk, SOCK_DONE))
1287 copied = sock_error(sk);
1291 if (sk->sk_shutdown & RCV_SHUTDOWN)
1294 if (sk->sk_state == TCP_CLOSE) {
1295 if (!sock_flag(sk, SOCK_DONE)) {
1296 /* This occurs when user tries to read
1297 * from never connected socket.
1310 if (signal_pending(current)) {
1311 copied = sock_intr_errno(timeo);
1316 cleanup_rbuf(sk, copied);
1318 if (!sysctl_tcp_low_latency && tp->ucopy.task == user_recv) {
1319 /* Install new reader */
1320 if (!user_recv && !(flags & (MSG_TRUNC | MSG_PEEK))) {
1321 user_recv = current;
1322 tp->ucopy.task = user_recv;
1323 tp->ucopy.iov = msg->msg_iov;
1326 tp->ucopy.len = len;
1328 BUG_TRAP(tp->copied_seq == tp->rcv_nxt ||
1329 (flags & (MSG_PEEK | MSG_TRUNC)));
1331 /* Ugly... If prequeue is not empty, we have to
1332 * process it before releasing socket, otherwise
1333 * order will be broken at second iteration.
1334 * More elegant solution is required!!!
1336 * Look: we have the following (pseudo)queues:
1338 * 1. packets in flight
1343 * Each queue can be processed only if the next ones
1344 * are empty. At this point we have empty receive_queue.
1345 * But prequeue _can_ be not empty after 2nd iteration,
1346 * when we jumped to start of loop because backlog
1347 * processing added something to receive_queue.
1348 * We cannot release_sock(), because backlog contains
1349 * packets arrived _after_ prequeued ones.
1351 * Shortly, algorithm is clear --- to process all
1352 * the queues in order. We could make it more directly,
1353 * requeueing packets from backlog to prequeue, if
1354 * is not empty. It is more elegant, but eats cycles,
1357 if (!skb_queue_empty(&tp->ucopy.prequeue))
1360 /* __ Set realtime policy in scheduler __ */
1363 if (copied >= target) {
1364 /* Do not sleep, just process backlog. */
1368 sk_wait_data(sk, &timeo);
1373 /* __ Restore normal policy in scheduler __ */
1375 if ((chunk = len - tp->ucopy.len) != 0) {
1376 NET_ADD_STATS_USER(LINUX_MIB_TCPDIRECTCOPYFROMBACKLOG, chunk);
1381 if (tp->rcv_nxt == tp->copied_seq &&
1382 !skb_queue_empty(&tp->ucopy.prequeue)) {
1384 tcp_prequeue_process(sk);
1386 if ((chunk = len - tp->ucopy.len) != 0) {
1387 NET_ADD_STATS_USER(LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1393 if ((flags & MSG_PEEK) && peek_seq != tp->copied_seq) {
1394 if (net_ratelimit())
1395 printk(KERN_DEBUG "TCP(%s:%d): Application bug, race in MSG_PEEK.\n",
1396 current->comm, current->pid);
1397 peek_seq = tp->copied_seq;
1402 /* Ok so how much can we use? */
1403 used = skb->len - offset;
1407 /* Do we have urgent data here? */
1409 u32 urg_offset = tp->urg_seq - *seq;
1410 if (urg_offset < used) {
1412 if (!sock_flag(sk, SOCK_URGINLINE)) {
1424 if (!(flags & MSG_TRUNC)) {
1425 err = skb_copy_datagram_iovec(skb, offset,
1426 msg->msg_iov, used);
1428 /* Exception. Bailout! */
1439 tcp_rcv_space_adjust(sk);
1442 if (tp->urg_data && after(tp->copied_seq, tp->urg_seq)) {
1444 tcp_fast_path_check(sk, tp);
1446 if (used + offset < skb->len)
1451 if (!(flags & MSG_PEEK))
1452 sk_eat_skb(sk, skb);
1456 /* Process the FIN. */
1458 if (!(flags & MSG_PEEK))
1459 sk_eat_skb(sk, skb);
1464 if (!skb_queue_empty(&tp->ucopy.prequeue)) {
1467 tp->ucopy.len = copied > 0 ? len : 0;
1469 tcp_prequeue_process(sk);
1471 if (copied > 0 && (chunk = len - tp->ucopy.len) != 0) {
1472 NET_ADD_STATS_USER(LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1478 tp->ucopy.task = NULL;
1482 /* According to UNIX98, msg_name/msg_namelen are ignored
1483 * on connected socket. I was just happy when found this 8) --ANK
1486 /* Clean up data we have read: This will do ACK frames. */
1487 cleanup_rbuf(sk, copied);
1489 TCP_CHECK_TIMER(sk);
1494 TCP_CHECK_TIMER(sk);
1499 err = tcp_recv_urg(sk, timeo, msg, len, flags, addr_len);
1504 * State processing on a close. This implements the state shift for
1505 * sending our FIN frame. Note that we only send a FIN for some
1506 * states. A shutdown() may have already sent the FIN, or we may be
1510 static unsigned char new_state[16] = {
1511 /* current state: new state: action: */
1512 /* (Invalid) */ TCP_CLOSE,
1513 /* TCP_ESTABLISHED */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1514 /* TCP_SYN_SENT */ TCP_CLOSE,
1515 /* TCP_SYN_RECV */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1516 /* TCP_FIN_WAIT1 */ TCP_FIN_WAIT1,
1517 /* TCP_FIN_WAIT2 */ TCP_FIN_WAIT2,
1518 /* TCP_TIME_WAIT */ TCP_CLOSE,
1519 /* TCP_CLOSE */ TCP_CLOSE,
1520 /* TCP_CLOSE_WAIT */ TCP_LAST_ACK | TCP_ACTION_FIN,
1521 /* TCP_LAST_ACK */ TCP_LAST_ACK,
1522 /* TCP_LISTEN */ TCP_CLOSE,
1523 /* TCP_CLOSING */ TCP_CLOSING,
1526 static int tcp_close_state(struct sock *sk)
1528 int next = (int)new_state[sk->sk_state];
1529 int ns = next & TCP_STATE_MASK;
1531 tcp_set_state(sk, ns);
1533 return next & TCP_ACTION_FIN;
1537 * Shutdown the sending side of a connection. Much like close except
1538 * that we don't receive shut down or set_sock_flag(sk, SOCK_DEAD).
1541 void tcp_shutdown(struct sock *sk, int how)
1543 /* We need to grab some memory, and put together a FIN,
1544 * and then put it into the queue to be sent.
1545 * Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
1547 if (!(how & SEND_SHUTDOWN))
1550 /* If we've already sent a FIN, or it's a closed state, skip this. */
1551 if ((1 << sk->sk_state) &
1552 (TCPF_ESTABLISHED | TCPF_SYN_SENT |
1553 TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) {
1554 /* Clear out any half completed packets. FIN if needed. */
1555 if (tcp_close_state(sk))
1561 * At this point, there should be no process reference to this
1562 * socket, and thus no user references at all. Therefore we
1563 * can assume the socket waitqueue is inactive and nobody will
1564 * try to jump onto it.
1566 void tcp_destroy_sock(struct sock *sk)
1568 BUG_TRAP(sk->sk_state == TCP_CLOSE);
1569 BUG_TRAP(sock_flag(sk, SOCK_DEAD));
1571 /* It cannot be in hash table! */
1572 BUG_TRAP(sk_unhashed(sk));
1574 /* If it has not 0 inet_sk(sk)->num, it must be bound */
1575 BUG_TRAP(!inet_sk(sk)->num || tcp_sk(sk)->bind_hash);
1577 sk->sk_prot->destroy(sk);
1579 sk_stream_kill_queues(sk);
1581 xfrm_sk_free_policy(sk);
1583 sk_refcnt_debug_release(sk);
1585 atomic_dec(&tcp_orphan_count);
1589 void tcp_close(struct sock *sk, long timeout)
1591 struct sk_buff *skb;
1592 int data_was_unread = 0;
1595 sk->sk_shutdown = SHUTDOWN_MASK;
1597 if (sk->sk_state == TCP_LISTEN) {
1598 tcp_set_state(sk, TCP_CLOSE);
1601 tcp_listen_stop(sk);
1603 goto adjudge_to_death;
1606 /* We need to flush the recv. buffs. We do this only on the
1607 * descriptor close, not protocol-sourced closes, because the
1608 * reader process may not have drained the data yet!
1610 while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) {
1611 u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq -
1613 data_was_unread += len;
1617 sk_stream_mem_reclaim(sk);
1619 /* As outlined in draft-ietf-tcpimpl-prob-03.txt, section
1620 * 3.10, we send a RST here because data was lost. To
1621 * witness the awful effects of the old behavior of always
1622 * doing a FIN, run an older 2.1.x kernel or 2.0.x, start
1623 * a bulk GET in an FTP client, suspend the process, wait
1624 * for the client to advertise a zero window, then kill -9
1625 * the FTP client, wheee... Note: timeout is always zero
1628 if (data_was_unread) {
1629 /* Unread data was tossed, zap the connection. */
1630 NET_INC_STATS_USER(LINUX_MIB_TCPABORTONCLOSE);
1631 tcp_set_state(sk, TCP_CLOSE);
1632 tcp_send_active_reset(sk, GFP_KERNEL);
1633 } else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) {
1634 /* Check zero linger _after_ checking for unread data. */
1635 sk->sk_prot->disconnect(sk, 0);
1636 NET_INC_STATS_USER(LINUX_MIB_TCPABORTONDATA);
1637 } else if (tcp_close_state(sk)) {
1638 /* We FIN if the application ate all the data before
1639 * zapping the connection.
1642 /* RED-PEN. Formally speaking, we have broken TCP state
1643 * machine. State transitions:
1645 * TCP_ESTABLISHED -> TCP_FIN_WAIT1
1646 * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible)
1647 * TCP_CLOSE_WAIT -> TCP_LAST_ACK
1649 * are legal only when FIN has been sent (i.e. in window),
1650 * rather than queued out of window. Purists blame.
1652 * F.e. "RFC state" is ESTABLISHED,
1653 * if Linux state is FIN-WAIT-1, but FIN is still not sent.
1655 * The visible declinations are that sometimes
1656 * we enter time-wait state, when it is not required really
1657 * (harmless), do not send active resets, when they are
1658 * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
1659 * they look as CLOSING or LAST_ACK for Linux)
1660 * Probably, I missed some more holelets.
1666 sk_stream_wait_close(sk, timeout);
1669 /* It is the last release_sock in its life. It will remove backlog. */
1673 /* Now socket is owned by kernel and we acquire BH lock
1674 to finish close. No need to check for user refs.
1678 BUG_TRAP(!sock_owned_by_user(sk));
1683 /* This is a (useful) BSD violating of the RFC. There is a
1684 * problem with TCP as specified in that the other end could
1685 * keep a socket open forever with no application left this end.
1686 * We use a 3 minute timeout (about the same as BSD) then kill
1687 * our end. If they send after that then tough - BUT: long enough
1688 * that we won't make the old 4*rto = almost no time - whoops
1691 * Nope, it was not mistake. It is really desired behaviour
1692 * f.e. on http servers, when such sockets are useless, but
1693 * consume significant resources. Let's do it with special
1694 * linger2 option. --ANK
1697 if (sk->sk_state == TCP_FIN_WAIT2) {
1698 struct tcp_sock *tp = tcp_sk(sk);
1699 if (tp->linger2 < 0) {
1700 tcp_set_state(sk, TCP_CLOSE);
1701 tcp_send_active_reset(sk, GFP_ATOMIC);
1702 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONLINGER);
1704 int tmo = tcp_fin_time(tp);
1706 if (tmo > TCP_TIMEWAIT_LEN) {
1707 tcp_reset_keepalive_timer(sk, tcp_fin_time(tp));
1709 atomic_inc(&tcp_orphan_count);
1710 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
1715 if (sk->sk_state != TCP_CLOSE) {
1716 sk_stream_mem_reclaim(sk);
1717 if (atomic_read(&tcp_orphan_count) > sysctl_tcp_max_orphans ||
1718 (sk->sk_wmem_queued > SOCK_MIN_SNDBUF &&
1719 atomic_read(&tcp_memory_allocated) > sysctl_tcp_mem[2])) {
1720 if (net_ratelimit())
1721 printk(KERN_INFO "TCP: too many of orphaned "
1723 tcp_set_state(sk, TCP_CLOSE);
1724 tcp_send_active_reset(sk, GFP_ATOMIC);
1725 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONMEMORY);
1728 atomic_inc(&tcp_orphan_count);
1730 if (sk->sk_state == TCP_CLOSE)
1731 tcp_destroy_sock(sk);
1732 /* Otherwise, socket is reprieved until protocol close. */
1740 /* These states need RST on ABORT according to RFC793 */
1742 static inline int tcp_need_reset(int state)
1744 return (1 << state) &
1745 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 |
1746 TCPF_FIN_WAIT2 | TCPF_SYN_RECV);
1749 int tcp_disconnect(struct sock *sk, int flags)
1751 struct inet_sock *inet = inet_sk(sk);
1752 struct tcp_sock *tp = tcp_sk(sk);
1754 int old_state = sk->sk_state;
1756 if (old_state != TCP_CLOSE)
1757 tcp_set_state(sk, TCP_CLOSE);
1759 /* ABORT function of RFC793 */
1760 if (old_state == TCP_LISTEN) {
1761 tcp_listen_stop(sk);
1762 } else if (tcp_need_reset(old_state) ||
1763 (tp->snd_nxt != tp->write_seq &&
1764 (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK))) {
1765 /* The last check adjusts for discrepance of Linux wrt. RFC
1768 tcp_send_active_reset(sk, gfp_any());
1769 sk->sk_err = ECONNRESET;
1770 } else if (old_state == TCP_SYN_SENT)
1771 sk->sk_err = ECONNRESET;
1773 tcp_clear_xmit_timers(sk);
1774 __skb_queue_purge(&sk->sk_receive_queue);
1775 sk_stream_writequeue_purge(sk);
1776 __skb_queue_purge(&tp->out_of_order_queue);
1780 if (!(sk->sk_userlocks & SOCK_BINDADDR_LOCK))
1781 inet_reset_saddr(sk);
1783 sk->sk_shutdown = 0;
1784 sock_reset_flag(sk, SOCK_DONE);
1786 if ((tp->write_seq += tp->max_window + 2) == 0)
1791 tp->packets_out = 0;
1792 tp->snd_ssthresh = 0x7fffffff;
1793 tp->snd_cwnd_cnt = 0;
1794 tcp_set_ca_state(tp, TCP_CA_Open);
1795 tcp_clear_retrans(tp);
1796 tcp_delack_init(tp);
1797 sk->sk_send_head = NULL;
1798 tp->rx_opt.saw_tstamp = 0;
1799 tcp_sack_reset(&tp->rx_opt);
1802 BUG_TRAP(!inet->num || tp->bind_hash);
1804 sk->sk_error_report(sk);
1809 * Wait for an incoming connection, avoid race
1810 * conditions. This must be called with the socket locked.
1812 static int wait_for_connect(struct sock *sk, long timeo)
1814 struct tcp_sock *tp = tcp_sk(sk);
1819 * True wake-one mechanism for incoming connections: only
1820 * one process gets woken up, not the 'whole herd'.
1821 * Since we do not 'race & poll' for established sockets
1822 * anymore, the common case will execute the loop only once.
1824 * Subtle issue: "add_wait_queue_exclusive()" will be added
1825 * after any current non-exclusive waiters, and we know that
1826 * it will always _stay_ after any new non-exclusive waiters
1827 * because all non-exclusive waiters are added at the
1828 * beginning of the wait-queue. As such, it's ok to "drop"
1829 * our exclusiveness temporarily when we get woken up without
1830 * having to remove and re-insert us on the wait queue.
1833 prepare_to_wait_exclusive(sk->sk_sleep, &wait,
1834 TASK_INTERRUPTIBLE);
1836 if (reqsk_queue_empty(&tp->accept_queue))
1837 timeo = schedule_timeout(timeo);
1840 if (!reqsk_queue_empty(&tp->accept_queue))
1843 if (sk->sk_state != TCP_LISTEN)
1845 err = sock_intr_errno(timeo);
1846 if (signal_pending(current))
1852 finish_wait(sk->sk_sleep, &wait);
1857 * This will accept the next outstanding connection.
1860 struct sock *tcp_accept(struct sock *sk, int flags, int *err)
1862 struct tcp_sock *tp = tcp_sk(sk);
1868 /* We need to make sure that this socket is listening,
1869 * and that it has something pending.
1872 if (sk->sk_state != TCP_LISTEN)
1875 /* Find already established connection */
1876 if (reqsk_queue_empty(&tp->accept_queue)) {
1877 long timeo = sock_rcvtimeo(sk, flags & O_NONBLOCK);
1879 /* If this is a non blocking socket don't sleep */
1884 error = wait_for_connect(sk, timeo);
1889 newsk = reqsk_queue_get_child(&tp->accept_queue, sk);
1890 BUG_TRAP(newsk->sk_state != TCP_SYN_RECV);
1901 * Socket option code for TCP.
1903 int tcp_setsockopt(struct sock *sk, int level, int optname, char __user *optval,
1906 struct tcp_sock *tp = tcp_sk(sk);
1910 if (level != SOL_TCP)
1911 return tp->af_specific->setsockopt(sk, level, optname,
1914 /* This is a string value all the others are int's */
1915 if (optname == TCP_CONGESTION) {
1916 char name[TCP_CA_NAME_MAX];
1921 val = strncpy_from_user(name, optval,
1922 min(TCP_CA_NAME_MAX-1, optlen));
1928 err = tcp_set_congestion_control(tp, name);
1933 if (optlen < sizeof(int))
1936 if (get_user(val, (int __user *)optval))
1943 /* Values greater than interface MTU won't take effect. However
1944 * at the point when this call is done we typically don't yet
1945 * know which interface is going to be used */
1946 if (val < 8 || val > MAX_TCP_WINDOW) {
1950 tp->rx_opt.user_mss = val;
1955 /* TCP_NODELAY is weaker than TCP_CORK, so that
1956 * this option on corked socket is remembered, but
1957 * it is not activated until cork is cleared.
1959 * However, when TCP_NODELAY is set we make
1960 * an explicit push, which overrides even TCP_CORK
1961 * for currently queued segments.
1963 tp->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH;
1964 tcp_push_pending_frames(sk, tp);
1966 tp->nonagle &= ~TCP_NAGLE_OFF;
1971 /* When set indicates to always queue non-full frames.
1972 * Later the user clears this option and we transmit
1973 * any pending partial frames in the queue. This is
1974 * meant to be used alongside sendfile() to get properly
1975 * filled frames when the user (for example) must write
1976 * out headers with a write() call first and then use
1977 * sendfile to send out the data parts.
1979 * TCP_CORK can be set together with TCP_NODELAY and it is
1980 * stronger than TCP_NODELAY.
1983 tp->nonagle |= TCP_NAGLE_CORK;
1985 tp->nonagle &= ~TCP_NAGLE_CORK;
1986 if (tp->nonagle&TCP_NAGLE_OFF)
1987 tp->nonagle |= TCP_NAGLE_PUSH;
1988 tcp_push_pending_frames(sk, tp);
1993 if (val < 1 || val > MAX_TCP_KEEPIDLE)
1996 tp->keepalive_time = val * HZ;
1997 if (sock_flag(sk, SOCK_KEEPOPEN) &&
1998 !((1 << sk->sk_state) &
1999 (TCPF_CLOSE | TCPF_LISTEN))) {
2000 __u32 elapsed = tcp_time_stamp - tp->rcv_tstamp;
2001 if (tp->keepalive_time > elapsed)
2002 elapsed = tp->keepalive_time - elapsed;
2005 tcp_reset_keepalive_timer(sk, elapsed);
2010 if (val < 1 || val > MAX_TCP_KEEPINTVL)
2013 tp->keepalive_intvl = val * HZ;
2016 if (val < 1 || val > MAX_TCP_KEEPCNT)
2019 tp->keepalive_probes = val;
2022 if (val < 1 || val > MAX_TCP_SYNCNT)
2025 tp->syn_retries = val;
2031 else if (val > sysctl_tcp_fin_timeout / HZ)
2034 tp->linger2 = val * HZ;
2037 case TCP_DEFER_ACCEPT:
2038 tp->defer_accept = 0;
2040 /* Translate value in seconds to number of
2042 while (tp->defer_accept < 32 &&
2043 val > ((TCP_TIMEOUT_INIT / HZ) <<
2050 case TCP_WINDOW_CLAMP:
2052 if (sk->sk_state != TCP_CLOSE) {
2056 tp->window_clamp = 0;
2058 tp->window_clamp = val < SOCK_MIN_RCVBUF / 2 ?
2059 SOCK_MIN_RCVBUF / 2 : val;
2064 tp->ack.pingpong = 1;
2066 tp->ack.pingpong = 0;
2067 if ((1 << sk->sk_state) &
2068 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) &&
2069 tcp_ack_scheduled(tp)) {
2070 tp->ack.pending |= TCP_ACK_PUSHED;
2071 cleanup_rbuf(sk, 1);
2073 tp->ack.pingpong = 1;
2086 /* Return information about state of tcp endpoint in API format. */
2087 void tcp_get_info(struct sock *sk, struct tcp_info *info)
2089 struct tcp_sock *tp = tcp_sk(sk);
2090 u32 now = tcp_time_stamp;
2092 memset(info, 0, sizeof(*info));
2094 info->tcpi_state = sk->sk_state;
2095 info->tcpi_ca_state = tp->ca_state;
2096 info->tcpi_retransmits = tp->retransmits;
2097 info->tcpi_probes = tp->probes_out;
2098 info->tcpi_backoff = tp->backoff;
2100 if (tp->rx_opt.tstamp_ok)
2101 info->tcpi_options |= TCPI_OPT_TIMESTAMPS;
2102 if (tp->rx_opt.sack_ok)
2103 info->tcpi_options |= TCPI_OPT_SACK;
2104 if (tp->rx_opt.wscale_ok) {
2105 info->tcpi_options |= TCPI_OPT_WSCALE;
2106 info->tcpi_snd_wscale = tp->rx_opt.snd_wscale;
2107 info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale;
2110 if (tp->ecn_flags&TCP_ECN_OK)
2111 info->tcpi_options |= TCPI_OPT_ECN;
2113 info->tcpi_rto = jiffies_to_usecs(tp->rto);
2114 info->tcpi_ato = jiffies_to_usecs(tp->ack.ato);
2115 info->tcpi_snd_mss = tp->mss_cache;
2116 info->tcpi_rcv_mss = tp->ack.rcv_mss;
2118 info->tcpi_unacked = tp->packets_out;
2119 info->tcpi_sacked = tp->sacked_out;
2120 info->tcpi_lost = tp->lost_out;
2121 info->tcpi_retrans = tp->retrans_out;
2122 info->tcpi_fackets = tp->fackets_out;
2124 info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime);
2125 info->tcpi_last_data_recv = jiffies_to_msecs(now - tp->ack.lrcvtime);
2126 info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp);
2128 info->tcpi_pmtu = tp->pmtu_cookie;
2129 info->tcpi_rcv_ssthresh = tp->rcv_ssthresh;
2130 info->tcpi_rtt = jiffies_to_usecs(tp->srtt)>>3;
2131 info->tcpi_rttvar = jiffies_to_usecs(tp->mdev)>>2;
2132 info->tcpi_snd_ssthresh = tp->snd_ssthresh;
2133 info->tcpi_snd_cwnd = tp->snd_cwnd;
2134 info->tcpi_advmss = tp->advmss;
2135 info->tcpi_reordering = tp->reordering;
2137 info->tcpi_rcv_rtt = jiffies_to_usecs(tp->rcv_rtt_est.rtt)>>3;
2138 info->tcpi_rcv_space = tp->rcvq_space.space;
2140 info->tcpi_total_retrans = tp->total_retrans;
2143 EXPORT_SYMBOL_GPL(tcp_get_info);
2145 int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval,
2148 struct tcp_sock *tp = tcp_sk(sk);
2151 if (level != SOL_TCP)
2152 return tp->af_specific->getsockopt(sk, level, optname,
2155 if (get_user(len, optlen))
2158 len = min_t(unsigned int, len, sizeof(int));
2165 val = tp->mss_cache;
2166 if (!val && ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
2167 val = tp->rx_opt.user_mss;
2170 val = !!(tp->nonagle&TCP_NAGLE_OFF);
2173 val = !!(tp->nonagle&TCP_NAGLE_CORK);
2176 val = (tp->keepalive_time ? : sysctl_tcp_keepalive_time) / HZ;
2179 val = (tp->keepalive_intvl ? : sysctl_tcp_keepalive_intvl) / HZ;
2182 val = tp->keepalive_probes ? : sysctl_tcp_keepalive_probes;
2185 val = tp->syn_retries ? : sysctl_tcp_syn_retries;
2190 val = (val ? : sysctl_tcp_fin_timeout) / HZ;
2192 case TCP_DEFER_ACCEPT:
2193 val = !tp->defer_accept ? 0 : ((TCP_TIMEOUT_INIT / HZ) <<
2194 (tp->defer_accept - 1));
2196 case TCP_WINDOW_CLAMP:
2197 val = tp->window_clamp;
2200 struct tcp_info info;
2202 if (get_user(len, optlen))
2205 tcp_get_info(sk, &info);
2207 len = min_t(unsigned int, len, sizeof(info));
2208 if (put_user(len, optlen))
2210 if (copy_to_user(optval, &info, len))
2215 val = !tp->ack.pingpong;
2218 case TCP_CONGESTION:
2219 if (get_user(len, optlen))
2221 len = min_t(unsigned int, len, TCP_CA_NAME_MAX);
2222 if (put_user(len, optlen))
2224 if (copy_to_user(optval, tp->ca_ops->name, len))
2228 return -ENOPROTOOPT;
2231 if (put_user(len, optlen))
2233 if (copy_to_user(optval, &val, len))
2239 extern void __skb_cb_too_small_for_tcp(int, int);
2240 extern struct tcp_congestion_ops tcp_reno;
2242 static __initdata unsigned long thash_entries;
2243 static int __init set_thash_entries(char *str)
2247 thash_entries = simple_strtoul(str, &str, 0);
2250 __setup("thash_entries=", set_thash_entries);
2252 void __init tcp_init(void)
2254 struct sk_buff *skb = NULL;
2257 if (sizeof(struct tcp_skb_cb) > sizeof(skb->cb))
2258 __skb_cb_too_small_for_tcp(sizeof(struct tcp_skb_cb),
2261 tcp_bucket_cachep = kmem_cache_create("tcp_bind_bucket",
2262 sizeof(struct tcp_bind_bucket),
2263 0, SLAB_HWCACHE_ALIGN,
2265 if (!tcp_bucket_cachep)
2266 panic("tcp_init: Cannot alloc tcp_bind_bucket cache.");
2268 tcp_timewait_cachep = kmem_cache_create("tcp_tw_bucket",
2269 sizeof(struct tcp_tw_bucket),
2270 0, SLAB_HWCACHE_ALIGN,
2272 if (!tcp_timewait_cachep)
2273 panic("tcp_init: Cannot alloc tcp_tw_bucket cache.");
2275 /* Size and allocate the main established and bind bucket
2278 * The methodology is similar to that of the buffer cache.
2280 tcp_ehash = (struct tcp_ehash_bucket *)
2281 alloc_large_system_hash("TCP established",
2282 sizeof(struct tcp_ehash_bucket),
2284 (num_physpages >= 128 * 1024) ?
2291 tcp_ehash_size = (1 << tcp_ehash_size) >> 1;
2292 for (i = 0; i < (tcp_ehash_size << 1); i++) {
2293 rwlock_init(&tcp_ehash[i].lock);
2294 INIT_HLIST_HEAD(&tcp_ehash[i].chain);
2297 tcp_bhash = (struct tcp_bind_hashbucket *)
2298 alloc_large_system_hash("TCP bind",
2299 sizeof(struct tcp_bind_hashbucket),
2301 (num_physpages >= 128 * 1024) ?
2308 tcp_bhash_size = 1 << tcp_bhash_size;
2309 for (i = 0; i < tcp_bhash_size; i++) {
2310 spin_lock_init(&tcp_bhash[i].lock);
2311 INIT_HLIST_HEAD(&tcp_bhash[i].chain);
2314 /* Try to be a bit smarter and adjust defaults depending
2315 * on available memory.
2317 for (order = 0; ((1 << order) << PAGE_SHIFT) <
2318 (tcp_bhash_size * sizeof(struct tcp_bind_hashbucket));
2322 sysctl_local_port_range[0] = 32768;
2323 sysctl_local_port_range[1] = 61000;
2324 sysctl_tcp_max_tw_buckets = 180000;
2325 sysctl_tcp_max_orphans = 4096 << (order - 4);
2326 sysctl_max_syn_backlog = 1024;
2327 } else if (order < 3) {
2328 sysctl_local_port_range[0] = 1024 * (3 - order);
2329 sysctl_tcp_max_tw_buckets >>= (3 - order);
2330 sysctl_tcp_max_orphans >>= (3 - order);
2331 sysctl_max_syn_backlog = 128;
2333 tcp_port_rover = sysctl_local_port_range[0] - 1;
2335 sysctl_tcp_mem[0] = 768 << order;
2336 sysctl_tcp_mem[1] = 1024 << order;
2337 sysctl_tcp_mem[2] = 1536 << order;
2340 sysctl_tcp_wmem[2] = 64 * 1024;
2341 sysctl_tcp_rmem[0] = PAGE_SIZE;
2342 sysctl_tcp_rmem[1] = 43689;
2343 sysctl_tcp_rmem[2] = 2 * 43689;
2346 printk(KERN_INFO "TCP: Hash tables configured "
2347 "(established %d bind %d)\n",
2348 tcp_ehash_size << 1, tcp_bhash_size);
2350 tcp_register_congestion_control(&tcp_reno);
2353 EXPORT_SYMBOL(tcp_accept);
2354 EXPORT_SYMBOL(tcp_close);
2355 EXPORT_SYMBOL(tcp_destroy_sock);
2356 EXPORT_SYMBOL(tcp_disconnect);
2357 EXPORT_SYMBOL(tcp_getsockopt);
2358 EXPORT_SYMBOL(tcp_ioctl);
2359 EXPORT_SYMBOL(tcp_poll);
2360 EXPORT_SYMBOL(tcp_read_sock);
2361 EXPORT_SYMBOL(tcp_recvmsg);
2362 EXPORT_SYMBOL(tcp_sendmsg);
2363 EXPORT_SYMBOL(tcp_sendpage);
2364 EXPORT_SYMBOL(tcp_setsockopt);
2365 EXPORT_SYMBOL(tcp_shutdown);
2366 EXPORT_SYMBOL(tcp_statistics);
2367 EXPORT_SYMBOL(tcp_timewait_cachep);