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 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 struct tcp_listen_opt *lopt;
467 sk->sk_max_ack_backlog = 0;
468 sk->sk_ack_backlog = 0;
469 tp->accept_queue = tp->accept_queue_tail = NULL;
470 rwlock_init(&tp->syn_wait_lock);
473 lopt = kmalloc(sizeof(struct tcp_listen_opt), GFP_KERNEL);
477 memset(lopt, 0, sizeof(struct tcp_listen_opt));
478 for (lopt->max_qlen_log = 6; ; lopt->max_qlen_log++)
479 if ((1 << lopt->max_qlen_log) >= sysctl_max_syn_backlog)
481 get_random_bytes(&lopt->hash_rnd, 4);
483 write_lock_bh(&tp->syn_wait_lock);
484 tp->listen_opt = lopt;
485 write_unlock_bh(&tp->syn_wait_lock);
487 /* There is race window here: we announce ourselves listening,
488 * but this transition is still not validated by get_port().
489 * It is OK, because this socket enters to hash table only
490 * after validation is complete.
492 sk->sk_state = TCP_LISTEN;
493 if (!sk->sk_prot->get_port(sk, inet->num)) {
494 inet->sport = htons(inet->num);
497 sk->sk_prot->hash(sk);
502 sk->sk_state = TCP_CLOSE;
503 write_lock_bh(&tp->syn_wait_lock);
504 tp->listen_opt = NULL;
505 write_unlock_bh(&tp->syn_wait_lock);
511 * This routine closes sockets which have been at least partially
512 * opened, but not yet accepted.
515 static void tcp_listen_stop (struct sock *sk)
517 struct tcp_sock *tp = tcp_sk(sk);
518 struct tcp_listen_opt *lopt = tp->listen_opt;
519 struct request_sock *acc_req = tp->accept_queue;
520 struct request_sock *req;
523 tcp_delete_keepalive_timer(sk);
525 /* make all the listen_opt local to us */
526 write_lock_bh(&tp->syn_wait_lock);
527 tp->listen_opt = NULL;
528 write_unlock_bh(&tp->syn_wait_lock);
529 tp->accept_queue = tp->accept_queue_tail = NULL;
532 for (i = 0; i < TCP_SYNQ_HSIZE; i++) {
533 while ((req = lopt->syn_table[i]) != NULL) {
534 lopt->syn_table[i] = req->dl_next;
538 /* Following specs, it would be better either to send FIN
539 * (and enter FIN-WAIT-1, it is normal close)
540 * or to send active reset (abort).
541 * Certainly, it is pretty dangerous while synflood, but it is
542 * bad justification for our negligence 8)
543 * To be honest, we are not able to make either
544 * of the variants now. --ANK
549 BUG_TRAP(!lopt->qlen);
553 while ((req = acc_req) != NULL) {
554 struct sock *child = req->sk;
556 acc_req = req->dl_next;
560 BUG_TRAP(!sock_owned_by_user(child));
563 tcp_disconnect(child, O_NONBLOCK);
567 atomic_inc(&tcp_orphan_count);
569 tcp_destroy_sock(child);
571 bh_unlock_sock(child);
575 sk_acceptq_removed(sk);
578 BUG_TRAP(!sk->sk_ack_backlog);
581 static inline void tcp_mark_push(struct tcp_sock *tp, struct sk_buff *skb)
583 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH;
584 tp->pushed_seq = tp->write_seq;
587 static inline int forced_push(struct tcp_sock *tp)
589 return after(tp->write_seq, tp->pushed_seq + (tp->max_window >> 1));
592 static inline void skb_entail(struct sock *sk, struct tcp_sock *tp,
596 TCP_SKB_CB(skb)->seq = tp->write_seq;
597 TCP_SKB_CB(skb)->end_seq = tp->write_seq;
598 TCP_SKB_CB(skb)->flags = TCPCB_FLAG_ACK;
599 TCP_SKB_CB(skb)->sacked = 0;
600 skb_header_release(skb);
601 __skb_queue_tail(&sk->sk_write_queue, skb);
602 sk_charge_skb(sk, skb);
603 if (!sk->sk_send_head)
604 sk->sk_send_head = skb;
605 else if (tp->nonagle&TCP_NAGLE_PUSH)
606 tp->nonagle &= ~TCP_NAGLE_PUSH;
609 static inline void tcp_mark_urg(struct tcp_sock *tp, int flags,
612 if (flags & MSG_OOB) {
614 tp->snd_up = tp->write_seq;
615 TCP_SKB_CB(skb)->sacked |= TCPCB_URG;
619 static inline void tcp_push(struct sock *sk, struct tcp_sock *tp, int flags,
620 int mss_now, int nonagle)
622 if (sk->sk_send_head) {
623 struct sk_buff *skb = sk->sk_write_queue.prev;
624 if (!(flags & MSG_MORE) || forced_push(tp))
625 tcp_mark_push(tp, skb);
626 tcp_mark_urg(tp, flags, skb);
627 __tcp_push_pending_frames(sk, tp, mss_now,
628 (flags & MSG_MORE) ? TCP_NAGLE_CORK : nonagle);
632 static ssize_t do_tcp_sendpages(struct sock *sk, struct page **pages, int poffset,
633 size_t psize, int flags)
635 struct tcp_sock *tp = tcp_sk(sk);
639 long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
641 /* Wait for a connection to finish. */
642 if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
643 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
646 clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
648 mss_now = tcp_current_mss(sk, !(flags&MSG_OOB));
652 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
656 struct sk_buff *skb = sk->sk_write_queue.prev;
657 struct page *page = pages[poffset / PAGE_SIZE];
658 int copy, i, can_coalesce;
659 int offset = poffset % PAGE_SIZE;
660 int size = min_t(size_t, psize, PAGE_SIZE - offset);
662 if (!sk->sk_send_head || (copy = mss_now - skb->len) <= 0) {
664 if (!sk_stream_memory_free(sk))
665 goto wait_for_sndbuf;
667 skb = sk_stream_alloc_pskb(sk, 0, 0,
670 goto wait_for_memory;
672 skb_entail(sk, tp, skb);
679 i = skb_shinfo(skb)->nr_frags;
680 can_coalesce = skb_can_coalesce(skb, i, page, offset);
681 if (!can_coalesce && i >= MAX_SKB_FRAGS) {
682 tcp_mark_push(tp, skb);
685 if (sk->sk_forward_alloc < copy &&
686 !sk_stream_mem_schedule(sk, copy, 0))
687 goto wait_for_memory;
690 skb_shinfo(skb)->frags[i - 1].size += copy;
693 skb_fill_page_desc(skb, i, page, offset, copy);
697 skb->data_len += copy;
698 skb->truesize += copy;
699 sk->sk_wmem_queued += copy;
700 sk->sk_forward_alloc -= copy;
701 skb->ip_summed = CHECKSUM_HW;
702 tp->write_seq += copy;
703 TCP_SKB_CB(skb)->end_seq += copy;
704 skb_shinfo(skb)->tso_segs = 0;
707 TCP_SKB_CB(skb)->flags &= ~TCPCB_FLAG_PSH;
711 if (!(psize -= copy))
714 if (skb->len != mss_now || (flags & MSG_OOB))
717 if (forced_push(tp)) {
718 tcp_mark_push(tp, skb);
719 __tcp_push_pending_frames(sk, tp, mss_now, TCP_NAGLE_PUSH);
720 } else if (skb == sk->sk_send_head)
721 tcp_push_one(sk, mss_now);
725 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
728 tcp_push(sk, tp, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH);
730 if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
733 mss_now = tcp_current_mss(sk, !(flags&MSG_OOB));
738 tcp_push(sk, tp, flags, mss_now, tp->nonagle);
745 return sk_stream_error(sk, flags, err);
748 ssize_t tcp_sendpage(struct socket *sock, struct page *page, int offset,
749 size_t size, int flags)
752 struct sock *sk = sock->sk;
754 #define TCP_ZC_CSUM_FLAGS (NETIF_F_IP_CSUM | NETIF_F_NO_CSUM | NETIF_F_HW_CSUM)
756 if (!(sk->sk_route_caps & NETIF_F_SG) ||
757 !(sk->sk_route_caps & TCP_ZC_CSUM_FLAGS))
758 return sock_no_sendpage(sock, page, offset, size, flags);
760 #undef TCP_ZC_CSUM_FLAGS
764 res = do_tcp_sendpages(sk, &page, offset, size, flags);
770 #define TCP_PAGE(sk) (sk->sk_sndmsg_page)
771 #define TCP_OFF(sk) (sk->sk_sndmsg_off)
773 static inline int select_size(struct sock *sk, struct tcp_sock *tp)
775 int tmp = tp->mss_cache_std;
777 if (sk->sk_route_caps & NETIF_F_SG) {
778 int pgbreak = SKB_MAX_HEAD(MAX_TCP_HEADER);
780 if (tmp >= pgbreak &&
781 tmp <= pgbreak + (MAX_SKB_FRAGS - 1) * PAGE_SIZE)
787 int tcp_sendmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg,
791 struct tcp_sock *tp = tcp_sk(sk);
801 flags = msg->msg_flags;
802 timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
804 /* Wait for a connection to finish. */
805 if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
806 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
809 /* This should be in poll */
810 clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
812 mss_now = tcp_current_mss(sk, !(flags&MSG_OOB));
814 /* Ok commence sending. */
815 iovlen = msg->msg_iovlen;
820 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
823 while (--iovlen >= 0) {
824 int seglen = iov->iov_len;
825 unsigned char __user *from = iov->iov_base;
832 skb = sk->sk_write_queue.prev;
834 if (!sk->sk_send_head ||
835 (copy = mss_now - skb->len) <= 0) {
838 /* Allocate new segment. If the interface is SG,
839 * allocate skb fitting to single page.
841 if (!sk_stream_memory_free(sk))
842 goto wait_for_sndbuf;
844 skb = sk_stream_alloc_pskb(sk, select_size(sk, tp),
845 0, sk->sk_allocation);
847 goto wait_for_memory;
850 * Check whether we can use HW checksum.
852 if (sk->sk_route_caps &
853 (NETIF_F_IP_CSUM | NETIF_F_NO_CSUM |
855 skb->ip_summed = CHECKSUM_HW;
857 skb_entail(sk, tp, skb);
861 /* Try to append data to the end of skb. */
865 /* Where to copy to? */
866 if (skb_tailroom(skb) > 0) {
867 /* We have some space in skb head. Superb! */
868 if (copy > skb_tailroom(skb))
869 copy = skb_tailroom(skb);
870 if ((err = skb_add_data(skb, from, copy)) != 0)
874 int i = skb_shinfo(skb)->nr_frags;
875 struct page *page = TCP_PAGE(sk);
876 int off = TCP_OFF(sk);
878 if (skb_can_coalesce(skb, i, page, off) &&
880 /* We can extend the last page
883 } else if (i == MAX_SKB_FRAGS ||
885 !(sk->sk_route_caps & NETIF_F_SG))) {
886 /* Need to add new fragment and cannot
887 * do this because interface is non-SG,
888 * or because all the page slots are
890 tcp_mark_push(tp, skb);
893 /* If page is cached, align
894 * offset to L1 cache boundary
896 off = (off + L1_CACHE_BYTES - 1) &
897 ~(L1_CACHE_BYTES - 1);
898 if (off == PAGE_SIZE) {
900 TCP_PAGE(sk) = page = NULL;
905 /* Allocate new cache page. */
906 if (!(page = sk_stream_alloc_page(sk)))
907 goto wait_for_memory;
911 if (copy > PAGE_SIZE - off)
912 copy = PAGE_SIZE - off;
914 /* Time to copy data. We are close to
916 err = skb_copy_to_page(sk, from, skb, page,
919 /* If this page was new, give it to the
920 * socket so it does not get leaked.
929 /* Update the skb. */
931 skb_shinfo(skb)->frags[i - 1].size +=
934 skb_fill_page_desc(skb, i, page, off, copy);
937 } else if (off + copy < PAGE_SIZE) {
943 TCP_OFF(sk) = off + copy;
947 TCP_SKB_CB(skb)->flags &= ~TCPCB_FLAG_PSH;
949 tp->write_seq += copy;
950 TCP_SKB_CB(skb)->end_seq += copy;
951 skb_shinfo(skb)->tso_segs = 0;
955 if ((seglen -= copy) == 0 && iovlen == 0)
958 if (skb->len != mss_now || (flags & MSG_OOB))
961 if (forced_push(tp)) {
962 tcp_mark_push(tp, skb);
963 __tcp_push_pending_frames(sk, tp, mss_now, TCP_NAGLE_PUSH);
964 } else if (skb == sk->sk_send_head)
965 tcp_push_one(sk, mss_now);
969 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
972 tcp_push(sk, tp, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH);
974 if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
977 mss_now = tcp_current_mss(sk, !(flags&MSG_OOB));
983 tcp_push(sk, tp, flags, mss_now, tp->nonagle);
990 if (sk->sk_send_head == skb)
991 sk->sk_send_head = NULL;
992 __skb_unlink(skb, skb->list);
993 sk_stream_free_skb(sk, skb);
1000 err = sk_stream_error(sk, flags, err);
1001 TCP_CHECK_TIMER(sk);
1007 * Handle reading urgent data. BSD has very simple semantics for
1008 * this, no blocking and very strange errors 8)
1011 static int tcp_recv_urg(struct sock *sk, long timeo,
1012 struct msghdr *msg, int len, int flags,
1015 struct tcp_sock *tp = tcp_sk(sk);
1017 /* No URG data to read. */
1018 if (sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data ||
1019 tp->urg_data == TCP_URG_READ)
1020 return -EINVAL; /* Yes this is right ! */
1022 if (sk->sk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DONE))
1025 if (tp->urg_data & TCP_URG_VALID) {
1027 char c = tp->urg_data;
1029 if (!(flags & MSG_PEEK))
1030 tp->urg_data = TCP_URG_READ;
1032 /* Read urgent data. */
1033 msg->msg_flags |= MSG_OOB;
1036 if (!(flags & MSG_TRUNC))
1037 err = memcpy_toiovec(msg->msg_iov, &c, 1);
1040 msg->msg_flags |= MSG_TRUNC;
1042 return err ? -EFAULT : len;
1045 if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
1048 /* Fixed the recv(..., MSG_OOB) behaviour. BSD docs and
1049 * the available implementations agree in this case:
1050 * this call should never block, independent of the
1051 * blocking state of the socket.
1052 * Mike <pall@rz.uni-karlsruhe.de>
1057 /* Clean up the receive buffer for full frames taken by the user,
1058 * then send an ACK if necessary. COPIED is the number of bytes
1059 * tcp_recvmsg has given to the user so far, it speeds up the
1060 * calculation of whether or not we must ACK for the sake of
1063 static void cleanup_rbuf(struct sock *sk, int copied)
1065 struct tcp_sock *tp = tcp_sk(sk);
1066 int time_to_ack = 0;
1069 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
1071 BUG_TRAP(!skb || before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq));
1074 if (tcp_ack_scheduled(tp)) {
1075 /* Delayed ACKs frequently hit locked sockets during bulk
1077 if (tp->ack.blocked ||
1078 /* Once-per-two-segments ACK was not sent by tcp_input.c */
1079 tp->rcv_nxt - tp->rcv_wup > tp->ack.rcv_mss ||
1081 * If this read emptied read buffer, we send ACK, if
1082 * connection is not bidirectional, user drained
1083 * receive buffer and there was a small segment
1086 (copied > 0 && (tp->ack.pending & TCP_ACK_PUSHED) &&
1087 !tp->ack.pingpong && !atomic_read(&sk->sk_rmem_alloc)))
1091 /* We send an ACK if we can now advertise a non-zero window
1092 * which has been raised "significantly".
1094 * Even if window raised up to infinity, do not send window open ACK
1095 * in states, where we will not receive more. It is useless.
1097 if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) {
1098 __u32 rcv_window_now = tcp_receive_window(tp);
1100 /* Optimize, __tcp_select_window() is not cheap. */
1101 if (2*rcv_window_now <= tp->window_clamp) {
1102 __u32 new_window = __tcp_select_window(sk);
1104 /* Send ACK now, if this read freed lots of space
1105 * in our buffer. Certainly, new_window is new window.
1106 * We can advertise it now, if it is not less than current one.
1107 * "Lots" means "at least twice" here.
1109 if (new_window && new_window >= 2 * rcv_window_now)
1117 static void tcp_prequeue_process(struct sock *sk)
1119 struct sk_buff *skb;
1120 struct tcp_sock *tp = tcp_sk(sk);
1122 NET_ADD_STATS_USER(LINUX_MIB_TCPPREQUEUED, skb_queue_len(&tp->ucopy.prequeue));
1124 /* RX process wants to run with disabled BHs, though it is not
1127 while ((skb = __skb_dequeue(&tp->ucopy.prequeue)) != NULL)
1128 sk->sk_backlog_rcv(sk, skb);
1131 /* Clear memory counter. */
1132 tp->ucopy.memory = 0;
1135 static inline struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off)
1137 struct sk_buff *skb;
1140 skb_queue_walk(&sk->sk_receive_queue, skb) {
1141 offset = seq - TCP_SKB_CB(skb)->seq;
1144 if (offset < skb->len || skb->h.th->fin) {
1153 * This routine provides an alternative to tcp_recvmsg() for routines
1154 * that would like to handle copying from skbuffs directly in 'sendfile'
1157 * - It is assumed that the socket was locked by the caller.
1158 * - The routine does not block.
1159 * - At present, there is no support for reading OOB data
1160 * or for 'peeking' the socket using this routine
1161 * (although both would be easy to implement).
1163 int tcp_read_sock(struct sock *sk, read_descriptor_t *desc,
1164 sk_read_actor_t recv_actor)
1166 struct sk_buff *skb;
1167 struct tcp_sock *tp = tcp_sk(sk);
1168 u32 seq = tp->copied_seq;
1172 if (sk->sk_state == TCP_LISTEN)
1174 while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1175 if (offset < skb->len) {
1178 len = skb->len - offset;
1179 /* Stop reading if we hit a patch of urgent data */
1181 u32 urg_offset = tp->urg_seq - seq;
1182 if (urg_offset < len)
1187 used = recv_actor(desc, skb, offset, len);
1193 if (offset != skb->len)
1196 if (skb->h.th->fin) {
1197 sk_eat_skb(sk, skb);
1201 sk_eat_skb(sk, skb);
1205 tp->copied_seq = seq;
1207 tcp_rcv_space_adjust(sk);
1209 /* Clean up data we have read: This will do ACK frames. */
1211 cleanup_rbuf(sk, copied);
1216 * This routine copies from a sock struct into the user buffer.
1218 * Technical note: in 2.3 we work on _locked_ socket, so that
1219 * tricks with *seq access order and skb->users are not required.
1220 * Probably, code can be easily improved even more.
1223 int tcp_recvmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg,
1224 size_t len, int nonblock, int flags, int *addr_len)
1226 struct tcp_sock *tp = tcp_sk(sk);
1232 int target; /* Read at least this many bytes */
1234 struct task_struct *user_recv = NULL;
1238 TCP_CHECK_TIMER(sk);
1241 if (sk->sk_state == TCP_LISTEN)
1244 timeo = sock_rcvtimeo(sk, nonblock);
1246 /* Urgent data needs to be handled specially. */
1247 if (flags & MSG_OOB)
1250 seq = &tp->copied_seq;
1251 if (flags & MSG_PEEK) {
1252 peek_seq = tp->copied_seq;
1256 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
1259 struct sk_buff *skb;
1262 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
1263 if (tp->urg_data && tp->urg_seq == *seq) {
1266 if (signal_pending(current)) {
1267 copied = timeo ? sock_intr_errno(timeo) : -EAGAIN;
1272 /* Next get a buffer. */
1274 skb = skb_peek(&sk->sk_receive_queue);
1279 /* Now that we have two receive queues this
1282 if (before(*seq, TCP_SKB_CB(skb)->seq)) {
1283 printk(KERN_INFO "recvmsg bug: copied %X "
1284 "seq %X\n", *seq, TCP_SKB_CB(skb)->seq);
1287 offset = *seq - TCP_SKB_CB(skb)->seq;
1290 if (offset < skb->len)
1294 BUG_TRAP(flags & MSG_PEEK);
1296 } while (skb != (struct sk_buff *)&sk->sk_receive_queue);
1298 /* Well, if we have backlog, try to process it now yet. */
1300 if (copied >= target && !sk->sk_backlog.tail)
1305 sk->sk_state == TCP_CLOSE ||
1306 (sk->sk_shutdown & RCV_SHUTDOWN) ||
1308 signal_pending(current) ||
1312 if (sock_flag(sk, SOCK_DONE))
1316 copied = sock_error(sk);
1320 if (sk->sk_shutdown & RCV_SHUTDOWN)
1323 if (sk->sk_state == TCP_CLOSE) {
1324 if (!sock_flag(sk, SOCK_DONE)) {
1325 /* This occurs when user tries to read
1326 * from never connected socket.
1339 if (signal_pending(current)) {
1340 copied = sock_intr_errno(timeo);
1345 cleanup_rbuf(sk, copied);
1347 if (tp->ucopy.task == user_recv) {
1348 /* Install new reader */
1349 if (!user_recv && !(flags & (MSG_TRUNC | MSG_PEEK))) {
1350 user_recv = current;
1351 tp->ucopy.task = user_recv;
1352 tp->ucopy.iov = msg->msg_iov;
1355 tp->ucopy.len = len;
1357 BUG_TRAP(tp->copied_seq == tp->rcv_nxt ||
1358 (flags & (MSG_PEEK | MSG_TRUNC)));
1360 /* Ugly... If prequeue is not empty, we have to
1361 * process it before releasing socket, otherwise
1362 * order will be broken at second iteration.
1363 * More elegant solution is required!!!
1365 * Look: we have the following (pseudo)queues:
1367 * 1. packets in flight
1372 * Each queue can be processed only if the next ones
1373 * are empty. At this point we have empty receive_queue.
1374 * But prequeue _can_ be not empty after 2nd iteration,
1375 * when we jumped to start of loop because backlog
1376 * processing added something to receive_queue.
1377 * We cannot release_sock(), because backlog contains
1378 * packets arrived _after_ prequeued ones.
1380 * Shortly, algorithm is clear --- to process all
1381 * the queues in order. We could make it more directly,
1382 * requeueing packets from backlog to prequeue, if
1383 * is not empty. It is more elegant, but eats cycles,
1386 if (skb_queue_len(&tp->ucopy.prequeue))
1389 /* __ Set realtime policy in scheduler __ */
1392 if (copied >= target) {
1393 /* Do not sleep, just process backlog. */
1397 sk_wait_data(sk, &timeo);
1402 /* __ Restore normal policy in scheduler __ */
1404 if ((chunk = len - tp->ucopy.len) != 0) {
1405 NET_ADD_STATS_USER(LINUX_MIB_TCPDIRECTCOPYFROMBACKLOG, chunk);
1410 if (tp->rcv_nxt == tp->copied_seq &&
1411 skb_queue_len(&tp->ucopy.prequeue)) {
1413 tcp_prequeue_process(sk);
1415 if ((chunk = len - tp->ucopy.len) != 0) {
1416 NET_ADD_STATS_USER(LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1422 if ((flags & MSG_PEEK) && peek_seq != tp->copied_seq) {
1423 if (net_ratelimit())
1424 printk(KERN_DEBUG "TCP(%s:%d): Application bug, race in MSG_PEEK.\n",
1425 current->comm, current->pid);
1426 peek_seq = tp->copied_seq;
1431 /* Ok so how much can we use? */
1432 used = skb->len - offset;
1436 /* Do we have urgent data here? */
1438 u32 urg_offset = tp->urg_seq - *seq;
1439 if (urg_offset < used) {
1441 if (!sock_flag(sk, SOCK_URGINLINE)) {
1453 if (!(flags & MSG_TRUNC)) {
1454 err = skb_copy_datagram_iovec(skb, offset,
1455 msg->msg_iov, used);
1457 /* Exception. Bailout! */
1468 tcp_rcv_space_adjust(sk);
1471 if (tp->urg_data && after(tp->copied_seq, tp->urg_seq)) {
1473 tcp_fast_path_check(sk, tp);
1475 if (used + offset < skb->len)
1480 if (!(flags & MSG_PEEK))
1481 sk_eat_skb(sk, skb);
1485 /* Process the FIN. */
1487 if (!(flags & MSG_PEEK))
1488 sk_eat_skb(sk, skb);
1493 if (skb_queue_len(&tp->ucopy.prequeue)) {
1496 tp->ucopy.len = copied > 0 ? len : 0;
1498 tcp_prequeue_process(sk);
1500 if (copied > 0 && (chunk = len - tp->ucopy.len) != 0) {
1501 NET_ADD_STATS_USER(LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1507 tp->ucopy.task = NULL;
1511 /* According to UNIX98, msg_name/msg_namelen are ignored
1512 * on connected socket. I was just happy when found this 8) --ANK
1515 /* Clean up data we have read: This will do ACK frames. */
1516 cleanup_rbuf(sk, copied);
1518 TCP_CHECK_TIMER(sk);
1523 TCP_CHECK_TIMER(sk);
1528 err = tcp_recv_urg(sk, timeo, msg, len, flags, addr_len);
1533 * State processing on a close. This implements the state shift for
1534 * sending our FIN frame. Note that we only send a FIN for some
1535 * states. A shutdown() may have already sent the FIN, or we may be
1539 static unsigned char new_state[16] = {
1540 /* current state: new state: action: */
1541 /* (Invalid) */ TCP_CLOSE,
1542 /* TCP_ESTABLISHED */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1543 /* TCP_SYN_SENT */ TCP_CLOSE,
1544 /* TCP_SYN_RECV */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1545 /* TCP_FIN_WAIT1 */ TCP_FIN_WAIT1,
1546 /* TCP_FIN_WAIT2 */ TCP_FIN_WAIT2,
1547 /* TCP_TIME_WAIT */ TCP_CLOSE,
1548 /* TCP_CLOSE */ TCP_CLOSE,
1549 /* TCP_CLOSE_WAIT */ TCP_LAST_ACK | TCP_ACTION_FIN,
1550 /* TCP_LAST_ACK */ TCP_LAST_ACK,
1551 /* TCP_LISTEN */ TCP_CLOSE,
1552 /* TCP_CLOSING */ TCP_CLOSING,
1555 static int tcp_close_state(struct sock *sk)
1557 int next = (int)new_state[sk->sk_state];
1558 int ns = next & TCP_STATE_MASK;
1560 tcp_set_state(sk, ns);
1562 return next & TCP_ACTION_FIN;
1566 * Shutdown the sending side of a connection. Much like close except
1567 * that we don't receive shut down or set_sock_flag(sk, SOCK_DEAD).
1570 void tcp_shutdown(struct sock *sk, int how)
1572 /* We need to grab some memory, and put together a FIN,
1573 * and then put it into the queue to be sent.
1574 * Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
1576 if (!(how & SEND_SHUTDOWN))
1579 /* If we've already sent a FIN, or it's a closed state, skip this. */
1580 if ((1 << sk->sk_state) &
1581 (TCPF_ESTABLISHED | TCPF_SYN_SENT |
1582 TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) {
1583 /* Clear out any half completed packets. FIN if needed. */
1584 if (tcp_close_state(sk))
1590 * At this point, there should be no process reference to this
1591 * socket, and thus no user references at all. Therefore we
1592 * can assume the socket waitqueue is inactive and nobody will
1593 * try to jump onto it.
1595 void tcp_destroy_sock(struct sock *sk)
1597 BUG_TRAP(sk->sk_state == TCP_CLOSE);
1598 BUG_TRAP(sock_flag(sk, SOCK_DEAD));
1600 /* It cannot be in hash table! */
1601 BUG_TRAP(sk_unhashed(sk));
1603 /* If it has not 0 inet_sk(sk)->num, it must be bound */
1604 BUG_TRAP(!inet_sk(sk)->num || tcp_sk(sk)->bind_hash);
1606 sk->sk_prot->destroy(sk);
1608 sk_stream_kill_queues(sk);
1610 xfrm_sk_free_policy(sk);
1612 #ifdef INET_REFCNT_DEBUG
1613 if (atomic_read(&sk->sk_refcnt) != 1) {
1614 printk(KERN_DEBUG "Destruction TCP %p delayed, c=%d\n",
1615 sk, atomic_read(&sk->sk_refcnt));
1619 atomic_dec(&tcp_orphan_count);
1623 void tcp_close(struct sock *sk, long timeout)
1625 struct sk_buff *skb;
1626 int data_was_unread = 0;
1629 sk->sk_shutdown = SHUTDOWN_MASK;
1631 if (sk->sk_state == TCP_LISTEN) {
1632 tcp_set_state(sk, TCP_CLOSE);
1635 tcp_listen_stop(sk);
1637 goto adjudge_to_death;
1640 /* We need to flush the recv. buffs. We do this only on the
1641 * descriptor close, not protocol-sourced closes, because the
1642 * reader process may not have drained the data yet!
1644 while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) {
1645 u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq -
1647 data_was_unread += len;
1651 sk_stream_mem_reclaim(sk);
1653 /* As outlined in draft-ietf-tcpimpl-prob-03.txt, section
1654 * 3.10, we send a RST here because data was lost. To
1655 * witness the awful effects of the old behavior of always
1656 * doing a FIN, run an older 2.1.x kernel or 2.0.x, start
1657 * a bulk GET in an FTP client, suspend the process, wait
1658 * for the client to advertise a zero window, then kill -9
1659 * the FTP client, wheee... Note: timeout is always zero
1662 if (data_was_unread) {
1663 /* Unread data was tossed, zap the connection. */
1664 NET_INC_STATS_USER(LINUX_MIB_TCPABORTONCLOSE);
1665 tcp_set_state(sk, TCP_CLOSE);
1666 tcp_send_active_reset(sk, GFP_KERNEL);
1667 } else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) {
1668 /* Check zero linger _after_ checking for unread data. */
1669 sk->sk_prot->disconnect(sk, 0);
1670 NET_INC_STATS_USER(LINUX_MIB_TCPABORTONDATA);
1671 } else if (tcp_close_state(sk)) {
1672 /* We FIN if the application ate all the data before
1673 * zapping the connection.
1676 /* RED-PEN. Formally speaking, we have broken TCP state
1677 * machine. State transitions:
1679 * TCP_ESTABLISHED -> TCP_FIN_WAIT1
1680 * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible)
1681 * TCP_CLOSE_WAIT -> TCP_LAST_ACK
1683 * are legal only when FIN has been sent (i.e. in window),
1684 * rather than queued out of window. Purists blame.
1686 * F.e. "RFC state" is ESTABLISHED,
1687 * if Linux state is FIN-WAIT-1, but FIN is still not sent.
1689 * The visible declinations are that sometimes
1690 * we enter time-wait state, when it is not required really
1691 * (harmless), do not send active resets, when they are
1692 * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
1693 * they look as CLOSING or LAST_ACK for Linux)
1694 * Probably, I missed some more holelets.
1700 sk_stream_wait_close(sk, timeout);
1703 /* It is the last release_sock in its life. It will remove backlog. */
1707 /* Now socket is owned by kernel and we acquire BH lock
1708 to finish close. No need to check for user refs.
1712 BUG_TRAP(!sock_owned_by_user(sk));
1717 /* This is a (useful) BSD violating of the RFC. There is a
1718 * problem with TCP as specified in that the other end could
1719 * keep a socket open forever with no application left this end.
1720 * We use a 3 minute timeout (about the same as BSD) then kill
1721 * our end. If they send after that then tough - BUT: long enough
1722 * that we won't make the old 4*rto = almost no time - whoops
1725 * Nope, it was not mistake. It is really desired behaviour
1726 * f.e. on http servers, when such sockets are useless, but
1727 * consume significant resources. Let's do it with special
1728 * linger2 option. --ANK
1731 if (sk->sk_state == TCP_FIN_WAIT2) {
1732 struct tcp_sock *tp = tcp_sk(sk);
1733 if (tp->linger2 < 0) {
1734 tcp_set_state(sk, TCP_CLOSE);
1735 tcp_send_active_reset(sk, GFP_ATOMIC);
1736 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONLINGER);
1738 int tmo = tcp_fin_time(tp);
1740 if (tmo > TCP_TIMEWAIT_LEN) {
1741 tcp_reset_keepalive_timer(sk, tcp_fin_time(tp));
1743 atomic_inc(&tcp_orphan_count);
1744 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
1749 if (sk->sk_state != TCP_CLOSE) {
1750 sk_stream_mem_reclaim(sk);
1751 if (atomic_read(&tcp_orphan_count) > sysctl_tcp_max_orphans ||
1752 (sk->sk_wmem_queued > SOCK_MIN_SNDBUF &&
1753 atomic_read(&tcp_memory_allocated) > sysctl_tcp_mem[2])) {
1754 if (net_ratelimit())
1755 printk(KERN_INFO "TCP: too many of orphaned "
1757 tcp_set_state(sk, TCP_CLOSE);
1758 tcp_send_active_reset(sk, GFP_ATOMIC);
1759 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONMEMORY);
1762 atomic_inc(&tcp_orphan_count);
1764 if (sk->sk_state == TCP_CLOSE)
1765 tcp_destroy_sock(sk);
1766 /* Otherwise, socket is reprieved until protocol close. */
1774 /* These states need RST on ABORT according to RFC793 */
1776 static inline int tcp_need_reset(int state)
1778 return (1 << state) &
1779 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 |
1780 TCPF_FIN_WAIT2 | TCPF_SYN_RECV);
1783 int tcp_disconnect(struct sock *sk, int flags)
1785 struct inet_sock *inet = inet_sk(sk);
1786 struct tcp_sock *tp = tcp_sk(sk);
1788 int old_state = sk->sk_state;
1790 if (old_state != TCP_CLOSE)
1791 tcp_set_state(sk, TCP_CLOSE);
1793 /* ABORT function of RFC793 */
1794 if (old_state == TCP_LISTEN) {
1795 tcp_listen_stop(sk);
1796 } else if (tcp_need_reset(old_state) ||
1797 (tp->snd_nxt != tp->write_seq &&
1798 (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK))) {
1799 /* The last check adjusts for discrepance of Linux wrt. RFC
1802 tcp_send_active_reset(sk, gfp_any());
1803 sk->sk_err = ECONNRESET;
1804 } else if (old_state == TCP_SYN_SENT)
1805 sk->sk_err = ECONNRESET;
1807 tcp_clear_xmit_timers(sk);
1808 __skb_queue_purge(&sk->sk_receive_queue);
1809 sk_stream_writequeue_purge(sk);
1810 __skb_queue_purge(&tp->out_of_order_queue);
1814 if (!(sk->sk_userlocks & SOCK_BINDADDR_LOCK))
1815 inet_reset_saddr(sk);
1817 sk->sk_shutdown = 0;
1818 sock_reset_flag(sk, SOCK_DONE);
1820 if ((tp->write_seq += tp->max_window + 2) == 0)
1825 tp->packets_out = 0;
1826 tp->snd_ssthresh = 0x7fffffff;
1827 tp->snd_cwnd_cnt = 0;
1828 tcp_set_ca_state(tp, TCP_CA_Open);
1829 tcp_clear_retrans(tp);
1830 tcp_delack_init(tp);
1831 sk->sk_send_head = NULL;
1832 tp->rx_opt.saw_tstamp = 0;
1833 tcp_sack_reset(&tp->rx_opt);
1836 BUG_TRAP(!inet->num || tp->bind_hash);
1838 sk->sk_error_report(sk);
1843 * Wait for an incoming connection, avoid race
1844 * conditions. This must be called with the socket locked.
1846 static int wait_for_connect(struct sock *sk, long timeo)
1848 struct tcp_sock *tp = tcp_sk(sk);
1853 * True wake-one mechanism for incoming connections: only
1854 * one process gets woken up, not the 'whole herd'.
1855 * Since we do not 'race & poll' for established sockets
1856 * anymore, the common case will execute the loop only once.
1858 * Subtle issue: "add_wait_queue_exclusive()" will be added
1859 * after any current non-exclusive waiters, and we know that
1860 * it will always _stay_ after any new non-exclusive waiters
1861 * because all non-exclusive waiters are added at the
1862 * beginning of the wait-queue. As such, it's ok to "drop"
1863 * our exclusiveness temporarily when we get woken up without
1864 * having to remove and re-insert us on the wait queue.
1867 prepare_to_wait_exclusive(sk->sk_sleep, &wait,
1868 TASK_INTERRUPTIBLE);
1870 if (!tp->accept_queue)
1871 timeo = schedule_timeout(timeo);
1874 if (tp->accept_queue)
1877 if (sk->sk_state != TCP_LISTEN)
1879 err = sock_intr_errno(timeo);
1880 if (signal_pending(current))
1886 finish_wait(sk->sk_sleep, &wait);
1891 * This will accept the next outstanding connection.
1894 struct sock *tcp_accept(struct sock *sk, int flags, int *err)
1896 struct tcp_sock *tp = tcp_sk(sk);
1897 struct request_sock *req;
1903 /* We need to make sure that this socket is listening,
1904 * and that it has something pending.
1907 if (sk->sk_state != TCP_LISTEN)
1910 /* Find already established connection */
1911 if (!tp->accept_queue) {
1912 long timeo = sock_rcvtimeo(sk, flags & O_NONBLOCK);
1914 /* If this is a non blocking socket don't sleep */
1919 error = wait_for_connect(sk, timeo);
1924 req = tp->accept_queue;
1925 if ((tp->accept_queue = req->dl_next) == NULL)
1926 tp->accept_queue_tail = NULL;
1929 sk_acceptq_removed(sk);
1931 BUG_TRAP(newsk->sk_state != TCP_SYN_RECV);
1942 * Socket option code for TCP.
1944 int tcp_setsockopt(struct sock *sk, int level, int optname, char __user *optval,
1947 struct tcp_sock *tp = tcp_sk(sk);
1951 if (level != SOL_TCP)
1952 return tp->af_specific->setsockopt(sk, level, optname,
1955 if (optlen < sizeof(int))
1958 if (get_user(val, (int __user *)optval))
1965 /* Values greater than interface MTU won't take effect. However
1966 * at the point when this call is done we typically don't yet
1967 * know which interface is going to be used */
1968 if (val < 8 || val > MAX_TCP_WINDOW) {
1972 tp->rx_opt.user_mss = val;
1977 /* TCP_NODELAY is weaker than TCP_CORK, so that
1978 * this option on corked socket is remembered, but
1979 * it is not activated until cork is cleared.
1981 * However, when TCP_NODELAY is set we make
1982 * an explicit push, which overrides even TCP_CORK
1983 * for currently queued segments.
1985 tp->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH;
1986 tcp_push_pending_frames(sk, tp);
1988 tp->nonagle &= ~TCP_NAGLE_OFF;
1993 /* When set indicates to always queue non-full frames.
1994 * Later the user clears this option and we transmit
1995 * any pending partial frames in the queue. This is
1996 * meant to be used alongside sendfile() to get properly
1997 * filled frames when the user (for example) must write
1998 * out headers with a write() call first and then use
1999 * sendfile to send out the data parts.
2001 * TCP_CORK can be set together with TCP_NODELAY and it is
2002 * stronger than TCP_NODELAY.
2005 tp->nonagle |= TCP_NAGLE_CORK;
2007 tp->nonagle &= ~TCP_NAGLE_CORK;
2008 if (tp->nonagle&TCP_NAGLE_OFF)
2009 tp->nonagle |= TCP_NAGLE_PUSH;
2010 tcp_push_pending_frames(sk, tp);
2015 if (val < 1 || val > MAX_TCP_KEEPIDLE)
2018 tp->keepalive_time = val * HZ;
2019 if (sock_flag(sk, SOCK_KEEPOPEN) &&
2020 !((1 << sk->sk_state) &
2021 (TCPF_CLOSE | TCPF_LISTEN))) {
2022 __u32 elapsed = tcp_time_stamp - tp->rcv_tstamp;
2023 if (tp->keepalive_time > elapsed)
2024 elapsed = tp->keepalive_time - elapsed;
2027 tcp_reset_keepalive_timer(sk, elapsed);
2032 if (val < 1 || val > MAX_TCP_KEEPINTVL)
2035 tp->keepalive_intvl = val * HZ;
2038 if (val < 1 || val > MAX_TCP_KEEPCNT)
2041 tp->keepalive_probes = val;
2044 if (val < 1 || val > MAX_TCP_SYNCNT)
2047 tp->syn_retries = val;
2053 else if (val > sysctl_tcp_fin_timeout / HZ)
2056 tp->linger2 = val * HZ;
2059 case TCP_DEFER_ACCEPT:
2060 tp->defer_accept = 0;
2062 /* Translate value in seconds to number of
2064 while (tp->defer_accept < 32 &&
2065 val > ((TCP_TIMEOUT_INIT / HZ) <<
2072 case TCP_WINDOW_CLAMP:
2074 if (sk->sk_state != TCP_CLOSE) {
2078 tp->window_clamp = 0;
2080 tp->window_clamp = val < SOCK_MIN_RCVBUF / 2 ?
2081 SOCK_MIN_RCVBUF / 2 : val;
2086 tp->ack.pingpong = 1;
2088 tp->ack.pingpong = 0;
2089 if ((1 << sk->sk_state) &
2090 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) &&
2091 tcp_ack_scheduled(tp)) {
2092 tp->ack.pending |= TCP_ACK_PUSHED;
2093 cleanup_rbuf(sk, 1);
2095 tp->ack.pingpong = 1;
2108 /* Return information about state of tcp endpoint in API format. */
2109 void tcp_get_info(struct sock *sk, struct tcp_info *info)
2111 struct tcp_sock *tp = tcp_sk(sk);
2112 u32 now = tcp_time_stamp;
2114 memset(info, 0, sizeof(*info));
2116 info->tcpi_state = sk->sk_state;
2117 info->tcpi_ca_state = tp->ca_state;
2118 info->tcpi_retransmits = tp->retransmits;
2119 info->tcpi_probes = tp->probes_out;
2120 info->tcpi_backoff = tp->backoff;
2122 if (tp->rx_opt.tstamp_ok)
2123 info->tcpi_options |= TCPI_OPT_TIMESTAMPS;
2124 if (tp->rx_opt.sack_ok)
2125 info->tcpi_options |= TCPI_OPT_SACK;
2126 if (tp->rx_opt.wscale_ok) {
2127 info->tcpi_options |= TCPI_OPT_WSCALE;
2128 info->tcpi_snd_wscale = tp->rx_opt.snd_wscale;
2129 info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale;
2132 if (tp->ecn_flags&TCP_ECN_OK)
2133 info->tcpi_options |= TCPI_OPT_ECN;
2135 info->tcpi_rto = jiffies_to_usecs(tp->rto);
2136 info->tcpi_ato = jiffies_to_usecs(tp->ack.ato);
2137 info->tcpi_snd_mss = tp->mss_cache_std;
2138 info->tcpi_rcv_mss = tp->ack.rcv_mss;
2140 info->tcpi_unacked = tp->packets_out;
2141 info->tcpi_sacked = tp->sacked_out;
2142 info->tcpi_lost = tp->lost_out;
2143 info->tcpi_retrans = tp->retrans_out;
2144 info->tcpi_fackets = tp->fackets_out;
2146 info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime);
2147 info->tcpi_last_data_recv = jiffies_to_msecs(now - tp->ack.lrcvtime);
2148 info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp);
2150 info->tcpi_pmtu = tp->pmtu_cookie;
2151 info->tcpi_rcv_ssthresh = tp->rcv_ssthresh;
2152 info->tcpi_rtt = jiffies_to_usecs(tp->srtt)>>3;
2153 info->tcpi_rttvar = jiffies_to_usecs(tp->mdev)>>2;
2154 info->tcpi_snd_ssthresh = tp->snd_ssthresh;
2155 info->tcpi_snd_cwnd = tp->snd_cwnd;
2156 info->tcpi_advmss = tp->advmss;
2157 info->tcpi_reordering = tp->reordering;
2159 info->tcpi_rcv_rtt = jiffies_to_usecs(tp->rcv_rtt_est.rtt)>>3;
2160 info->tcpi_rcv_space = tp->rcvq_space.space;
2162 info->tcpi_total_retrans = tp->total_retrans;
2165 EXPORT_SYMBOL_GPL(tcp_get_info);
2167 int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval,
2170 struct tcp_sock *tp = tcp_sk(sk);
2173 if (level != SOL_TCP)
2174 return tp->af_specific->getsockopt(sk, level, optname,
2177 if (get_user(len, optlen))
2180 len = min_t(unsigned int, len, sizeof(int));
2187 val = tp->mss_cache_std;
2188 if (!val && ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
2189 val = tp->rx_opt.user_mss;
2192 val = !!(tp->nonagle&TCP_NAGLE_OFF);
2195 val = !!(tp->nonagle&TCP_NAGLE_CORK);
2198 val = (tp->keepalive_time ? : sysctl_tcp_keepalive_time) / HZ;
2201 val = (tp->keepalive_intvl ? : sysctl_tcp_keepalive_intvl) / HZ;
2204 val = tp->keepalive_probes ? : sysctl_tcp_keepalive_probes;
2207 val = tp->syn_retries ? : sysctl_tcp_syn_retries;
2212 val = (val ? : sysctl_tcp_fin_timeout) / HZ;
2214 case TCP_DEFER_ACCEPT:
2215 val = !tp->defer_accept ? 0 : ((TCP_TIMEOUT_INIT / HZ) <<
2216 (tp->defer_accept - 1));
2218 case TCP_WINDOW_CLAMP:
2219 val = tp->window_clamp;
2222 struct tcp_info info;
2224 if (get_user(len, optlen))
2227 tcp_get_info(sk, &info);
2229 len = min_t(unsigned int, len, sizeof(info));
2230 if (put_user(len, optlen))
2232 if (copy_to_user(optval, &info, len))
2237 val = !tp->ack.pingpong;
2240 return -ENOPROTOOPT;
2243 if (put_user(len, optlen))
2245 if (copy_to_user(optval, &val, len))
2251 extern void __skb_cb_too_small_for_tcp(int, int);
2252 extern void tcpdiag_init(void);
2254 static __initdata unsigned long thash_entries;
2255 static int __init set_thash_entries(char *str)
2259 thash_entries = simple_strtoul(str, &str, 0);
2262 __setup("thash_entries=", set_thash_entries);
2264 void __init tcp_init(void)
2266 struct sk_buff *skb = NULL;
2269 if (sizeof(struct tcp_skb_cb) > sizeof(skb->cb))
2270 __skb_cb_too_small_for_tcp(sizeof(struct tcp_skb_cb),
2273 tcp_bucket_cachep = kmem_cache_create("tcp_bind_bucket",
2274 sizeof(struct tcp_bind_bucket),
2275 0, SLAB_HWCACHE_ALIGN,
2277 if (!tcp_bucket_cachep)
2278 panic("tcp_init: Cannot alloc tcp_bind_bucket cache.");
2280 tcp_timewait_cachep = kmem_cache_create("tcp_tw_bucket",
2281 sizeof(struct tcp_tw_bucket),
2282 0, SLAB_HWCACHE_ALIGN,
2284 if (!tcp_timewait_cachep)
2285 panic("tcp_init: Cannot alloc tcp_tw_bucket cache.");
2287 /* Size and allocate the main established and bind bucket
2290 * The methodology is similar to that of the buffer cache.
2292 tcp_ehash = (struct tcp_ehash_bucket *)
2293 alloc_large_system_hash("TCP established",
2294 sizeof(struct tcp_ehash_bucket),
2296 (num_physpages >= 128 * 1024) ?
2303 tcp_ehash_size = (1 << tcp_ehash_size) >> 1;
2304 for (i = 0; i < (tcp_ehash_size << 1); i++) {
2305 rwlock_init(&tcp_ehash[i].lock);
2306 INIT_HLIST_HEAD(&tcp_ehash[i].chain);
2309 tcp_bhash = (struct tcp_bind_hashbucket *)
2310 alloc_large_system_hash("TCP bind",
2311 sizeof(struct tcp_bind_hashbucket),
2313 (num_physpages >= 128 * 1024) ?
2320 tcp_bhash_size = 1 << tcp_bhash_size;
2321 for (i = 0; i < tcp_bhash_size; i++) {
2322 spin_lock_init(&tcp_bhash[i].lock);
2323 INIT_HLIST_HEAD(&tcp_bhash[i].chain);
2326 /* Try to be a bit smarter and adjust defaults depending
2327 * on available memory.
2329 for (order = 0; ((1 << order) << PAGE_SHIFT) <
2330 (tcp_bhash_size * sizeof(struct tcp_bind_hashbucket));
2334 sysctl_local_port_range[0] = 32768;
2335 sysctl_local_port_range[1] = 61000;
2336 sysctl_tcp_max_tw_buckets = 180000;
2337 sysctl_tcp_max_orphans = 4096 << (order - 4);
2338 sysctl_max_syn_backlog = 1024;
2339 } else if (order < 3) {
2340 sysctl_local_port_range[0] = 1024 * (3 - order);
2341 sysctl_tcp_max_tw_buckets >>= (3 - order);
2342 sysctl_tcp_max_orphans >>= (3 - order);
2343 sysctl_max_syn_backlog = 128;
2345 tcp_port_rover = sysctl_local_port_range[0] - 1;
2347 sysctl_tcp_mem[0] = 768 << order;
2348 sysctl_tcp_mem[1] = 1024 << order;
2349 sysctl_tcp_mem[2] = 1536 << order;
2352 sysctl_tcp_wmem[2] = 64 * 1024;
2353 sysctl_tcp_rmem[0] = PAGE_SIZE;
2354 sysctl_tcp_rmem[1] = 43689;
2355 sysctl_tcp_rmem[2] = 2 * 43689;
2358 printk(KERN_INFO "TCP: Hash tables configured "
2359 "(established %d bind %d)\n",
2360 tcp_ehash_size << 1, tcp_bhash_size);
2363 EXPORT_SYMBOL(tcp_accept);
2364 EXPORT_SYMBOL(tcp_close);
2365 EXPORT_SYMBOL(tcp_destroy_sock);
2366 EXPORT_SYMBOL(tcp_disconnect);
2367 EXPORT_SYMBOL(tcp_getsockopt);
2368 EXPORT_SYMBOL(tcp_ioctl);
2369 EXPORT_SYMBOL(tcp_poll);
2370 EXPORT_SYMBOL(tcp_read_sock);
2371 EXPORT_SYMBOL(tcp_recvmsg);
2372 EXPORT_SYMBOL(tcp_sendmsg);
2373 EXPORT_SYMBOL(tcp_sendpage);
2374 EXPORT_SYMBOL(tcp_setsockopt);
2375 EXPORT_SYMBOL(tcp_shutdown);
2376 EXPORT_SYMBOL(tcp_statistics);
2377 EXPORT_SYMBOL(tcp_timewait_cachep);