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 $
10 * Authors: Ross Biro, <bir7@leland.Stanford.Edu>
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 <net/icmp.h>
261 #include <net/xfrm.h>
265 #include <asm/uaccess.h>
266 #include <asm/ioctls.h>
268 int sysctl_tcp_fin_timeout = TCP_FIN_TIMEOUT;
270 DEFINE_SNMP_STAT(struct tcp_mib, tcp_statistics);
272 kmem_cache_t *tcp_openreq_cachep;
273 kmem_cache_t *tcp_bucket_cachep;
274 kmem_cache_t *tcp_timewait_cachep;
276 atomic_t tcp_orphan_count = ATOMIC_INIT(0);
278 int sysctl_tcp_mem[3];
279 int sysctl_tcp_wmem[3] = { 4 * 1024, 16 * 1024, 128 * 1024 };
280 int sysctl_tcp_rmem[3] = { 4 * 1024, 87380, 87380 * 2 };
282 EXPORT_SYMBOL(sysctl_tcp_mem);
283 EXPORT_SYMBOL(sysctl_tcp_rmem);
284 EXPORT_SYMBOL(sysctl_tcp_wmem);
286 atomic_t tcp_memory_allocated; /* Current allocated memory. */
287 atomic_t tcp_sockets_allocated; /* Current number of TCP sockets. */
289 EXPORT_SYMBOL(tcp_memory_allocated);
290 EXPORT_SYMBOL(tcp_sockets_allocated);
293 * Pressure flag: try to collapse.
294 * Technical note: it is used by multiple contexts non atomically.
295 * All the sk_stream_mem_schedule() is of this nature: accounting
296 * is strict, actions are advisory and have some latency.
298 int tcp_memory_pressure;
300 EXPORT_SYMBOL(tcp_memory_pressure);
302 void tcp_enter_memory_pressure(void)
304 if (!tcp_memory_pressure) {
305 NET_INC_STATS(LINUX_MIB_TCPMEMORYPRESSURES);
306 tcp_memory_pressure = 1;
310 EXPORT_SYMBOL(tcp_enter_memory_pressure);
313 * LISTEN is a special case for poll..
315 static __inline__ unsigned int tcp_listen_poll(struct sock *sk,
318 return tcp_sk(sk)->accept_queue ? (POLLIN | POLLRDNORM) : 0;
322 * Wait for a TCP event.
324 * Note that we don't need to lock the socket, as the upper poll layers
325 * take care of normal races (between the test and the event) and we don't
326 * go look at any of the socket buffers directly.
328 unsigned int tcp_poll(struct file *file, struct socket *sock, poll_table *wait)
331 struct sock *sk = sock->sk;
332 struct tcp_opt *tp = tcp_sk(sk);
334 poll_wait(file, sk->sk_sleep, wait);
335 if (sk->sk_state == TCP_LISTEN)
336 return tcp_listen_poll(sk, wait);
338 /* Socket is not locked. We are protected from async events
339 by poll logic and correct handling of state changes
340 made by another threads is impossible in any case.
348 * POLLHUP is certainly not done right. But poll() doesn't
349 * have a notion of HUP in just one direction, and for a
350 * socket the read side is more interesting.
352 * Some poll() documentation says that POLLHUP is incompatible
353 * with the POLLOUT/POLLWR flags, so somebody should check this
354 * all. But careful, it tends to be safer to return too many
355 * bits than too few, and you can easily break real applications
356 * if you don't tell them that something has hung up!
360 * Check number 1. POLLHUP is _UNMASKABLE_ event (see UNIX98 and
361 * our fs/select.c). It means that after we received EOF,
362 * poll always returns immediately, making impossible poll() on write()
363 * in state CLOSE_WAIT. One solution is evident --- to set POLLHUP
364 * if and only if shutdown has been made in both directions.
365 * Actually, it is interesting to look how Solaris and DUX
366 * solve this dilemma. I would prefer, if PULLHUP were maskable,
367 * then we could set it on SND_SHUTDOWN. BTW examples given
368 * in Stevens' books assume exactly this behaviour, it explains
369 * why PULLHUP is incompatible with POLLOUT. --ANK
371 * NOTE. Check for TCP_CLOSE is added. The goal is to prevent
372 * blocking on fresh not-connected or disconnected socket. --ANK
374 if (sk->sk_shutdown == SHUTDOWN_MASK || sk->sk_state == TCP_CLOSE)
376 if (sk->sk_shutdown & RCV_SHUTDOWN)
377 mask |= POLLIN | POLLRDNORM;
380 if ((1 << sk->sk_state) & ~(TCPF_SYN_SENT | TCPF_SYN_RECV)) {
381 /* Potential race condition. If read of tp below will
382 * escape above sk->sk_state, we can be illegally awaken
383 * in SYN_* states. */
384 if ((tp->rcv_nxt != tp->copied_seq) &&
385 (tp->urg_seq != tp->copied_seq ||
386 tp->rcv_nxt != tp->copied_seq + 1 ||
387 sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data))
388 mask |= POLLIN | POLLRDNORM;
390 if (!(sk->sk_shutdown & SEND_SHUTDOWN)) {
391 if (sk_stream_wspace(sk) >= sk_stream_min_wspace(sk)) {
392 mask |= POLLOUT | POLLWRNORM;
393 } else { /* send SIGIO later */
394 set_bit(SOCK_ASYNC_NOSPACE,
395 &sk->sk_socket->flags);
396 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
398 /* Race breaker. If space is freed after
399 * wspace test but before the flags are set,
400 * IO signal will be lost.
402 if (sk_stream_wspace(sk) >= sk_stream_min_wspace(sk))
403 mask |= POLLOUT | POLLWRNORM;
407 if (tp->urg_data & TCP_URG_VALID)
413 int tcp_ioctl(struct sock *sk, int cmd, unsigned long arg)
415 struct tcp_opt *tp = tcp_sk(sk);
420 if (sk->sk_state == TCP_LISTEN)
424 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
426 else if (sock_flag(sk, SOCK_URGINLINE) ||
428 before(tp->urg_seq, tp->copied_seq) ||
429 !before(tp->urg_seq, tp->rcv_nxt)) {
430 answ = tp->rcv_nxt - tp->copied_seq;
432 /* Subtract 1, if FIN is in queue. */
433 if (answ && !skb_queue_empty(&sk->sk_receive_queue))
435 ((struct sk_buff *)sk->sk_receive_queue.prev)->h.th->fin;
437 answ = tp->urg_seq - tp->copied_seq;
441 answ = tp->urg_data && tp->urg_seq == tp->copied_seq;
444 if (sk->sk_state == TCP_LISTEN)
447 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
450 answ = tp->write_seq - tp->snd_una;
456 return put_user(answ, (int __user *)arg);
460 int tcp_listen_start(struct sock *sk)
462 struct inet_opt *inet = inet_sk(sk);
463 struct tcp_opt *tp = tcp_sk(sk);
464 struct tcp_listen_opt *lopt;
466 sk->sk_max_ack_backlog = 0;
467 sk->sk_ack_backlog = 0;
468 tp->accept_queue = tp->accept_queue_tail = NULL;
469 rwlock_init(&tp->syn_wait_lock);
472 lopt = kmalloc(sizeof(struct tcp_listen_opt), GFP_KERNEL);
476 memset(lopt, 0, sizeof(struct tcp_listen_opt));
477 for (lopt->max_qlen_log = 6; ; lopt->max_qlen_log++)
478 if ((1 << lopt->max_qlen_log) >= sysctl_max_syn_backlog)
480 get_random_bytes(&lopt->hash_rnd, 4);
482 write_lock_bh(&tp->syn_wait_lock);
483 tp->listen_opt = lopt;
484 write_unlock_bh(&tp->syn_wait_lock);
486 /* There is race window here: we announce ourselves listening,
487 * but this transition is still not validated by get_port().
488 * It is OK, because this socket enters to hash table only
489 * after validation is complete.
491 sk->sk_state = TCP_LISTEN;
492 if (!sk->sk_prot->get_port(sk, inet->num)) {
493 inet->sport = htons(inet->num);
496 sk->sk_prot->hash(sk);
500 sk->sk_state = TCP_CLOSE;
501 write_lock_bh(&tp->syn_wait_lock);
502 tp->listen_opt = NULL;
503 write_unlock_bh(&tp->syn_wait_lock);
509 * This routine closes sockets which have been at least partially
510 * opened, but not yet accepted.
513 static void tcp_listen_stop (struct sock *sk)
515 struct tcp_opt *tp = tcp_sk(sk);
516 struct tcp_listen_opt *lopt = tp->listen_opt;
517 struct open_request *acc_req = tp->accept_queue;
518 struct open_request *req;
521 tcp_delete_keepalive_timer(sk);
523 /* make all the listen_opt local to us */
524 write_lock_bh(&tp->syn_wait_lock);
525 tp->listen_opt = NULL;
526 write_unlock_bh(&tp->syn_wait_lock);
528 tp->accept_queue_tail = NULL;
529 tp->accept_queue = 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;
536 tcp_openreq_free(req);
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);
576 tcp_openreq_fastfree(req);
578 BUG_TRAP(!sk->sk_ack_backlog);
581 static inline void tcp_mark_push(struct tcp_opt *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_opt *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_opt *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_queue_tail(&sk->sk_write_queue, skb);
601 sk_charge_skb(sk, skb);
602 if (!sk->sk_send_head)
603 sk->sk_send_head = skb;
604 else if (tp->nonagle&TCP_NAGLE_PUSH)
605 tp->nonagle &= ~TCP_NAGLE_PUSH;
608 static inline void tcp_mark_urg(struct tcp_opt *tp, int flags,
611 if (flags & MSG_OOB) {
613 tp->snd_up = tp->write_seq;
614 TCP_SKB_CB(skb)->sacked |= TCPCB_URG;
618 static inline void tcp_push(struct sock *sk, struct tcp_opt *tp, int flags,
619 int mss_now, int nonagle)
621 if (sk->sk_send_head) {
622 struct sk_buff *skb = sk->sk_write_queue.prev;
623 if (!(flags & MSG_MORE) || forced_push(tp))
624 tcp_mark_push(tp, skb);
625 tcp_mark_urg(tp, flags, skb);
626 __tcp_push_pending_frames(sk, tp, mss_now,
627 (flags & MSG_MORE) ? TCP_NAGLE_CORK : nonagle);
631 static ssize_t do_tcp_sendpages(struct sock *sk, struct page **pages, int poffset,
632 size_t psize, int flags)
634 struct tcp_opt *tp = tcp_sk(sk);
638 long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
640 /* Wait for a connection to finish. */
641 if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
642 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
645 clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
647 mss_now = tcp_current_mss(sk, !(flags&MSG_OOB));
651 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
655 struct sk_buff *skb = sk->sk_write_queue.prev;
656 struct page *page = pages[poffset / PAGE_SIZE];
658 int offset = poffset % PAGE_SIZE;
659 int size = min_t(size_t, psize, PAGE_SIZE - offset);
661 if (!sk->sk_send_head || (copy = mss_now - skb->len) <= 0) {
663 if (!sk_stream_memory_free(sk))
664 goto wait_for_sndbuf;
666 skb = sk_stream_alloc_pskb(sk, 0, tp->mss_cache,
669 goto wait_for_memory;
671 skb_entail(sk, tp, skb);
678 i = skb_shinfo(skb)->nr_frags;
679 if (skb_can_coalesce(skb, i, page, offset)) {
680 skb_shinfo(skb)->frags[i - 1].size += copy;
681 } else if (i < MAX_SKB_FRAGS) {
683 skb_fill_page_desc(skb, i, page, offset, copy);
685 tcp_mark_push(tp, skb);
690 skb->data_len += copy;
691 skb->ip_summed = CHECKSUM_HW;
692 tp->write_seq += copy;
693 TCP_SKB_CB(skb)->end_seq += copy;
694 skb_shinfo(skb)->tso_segs = 0;
697 TCP_SKB_CB(skb)->flags &= ~TCPCB_FLAG_PSH;
701 if (!(psize -= copy))
704 if (skb->len != mss_now || (flags & MSG_OOB))
707 if (forced_push(tp)) {
708 tcp_mark_push(tp, skb);
709 __tcp_push_pending_frames(sk, tp, mss_now, TCP_NAGLE_PUSH);
710 } else if (skb == sk->sk_send_head)
711 tcp_push_one(sk, mss_now);
715 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
718 tcp_push(sk, tp, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH);
720 if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
723 mss_now = tcp_current_mss(sk, !(flags&MSG_OOB));
728 tcp_push(sk, tp, flags, mss_now, tp->nonagle);
735 return sk_stream_error(sk, flags, err);
738 ssize_t tcp_sendpage(struct socket *sock, struct page *page, int offset,
739 size_t size, int flags)
742 struct sock *sk = sock->sk;
744 #define TCP_ZC_CSUM_FLAGS (NETIF_F_IP_CSUM | NETIF_F_NO_CSUM | NETIF_F_HW_CSUM)
746 if (!(sk->sk_route_caps & NETIF_F_SG) ||
747 !(sk->sk_route_caps & TCP_ZC_CSUM_FLAGS))
748 return sock_no_sendpage(sock, page, offset, size, flags);
750 #undef TCP_ZC_CSUM_FLAGS
754 res = do_tcp_sendpages(sk, &page, offset, size, flags);
760 #define TCP_PAGE(sk) (sk->sk_sndmsg_page)
761 #define TCP_OFF(sk) (sk->sk_sndmsg_off)
763 static inline int select_size(struct sock *sk, struct tcp_opt *tp)
765 int tmp = tp->mss_cache_std;
767 if (sk->sk_route_caps & NETIF_F_SG) {
768 int pgbreak = SKB_MAX_HEAD(MAX_TCP_HEADER);
770 if (tmp >= pgbreak &&
771 tmp <= pgbreak + (MAX_SKB_FRAGS - 1) * PAGE_SIZE)
777 int tcp_sendmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg,
781 struct tcp_opt *tp = tcp_sk(sk);
791 flags = msg->msg_flags;
792 timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
794 /* Wait for a connection to finish. */
795 if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
796 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
799 /* This should be in poll */
800 clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
802 mss_now = tcp_current_mss(sk, !(flags&MSG_OOB));
804 /* Ok commence sending. */
805 iovlen = msg->msg_iovlen;
810 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
813 while (--iovlen >= 0) {
814 int seglen = iov->iov_len;
815 unsigned char __user *from = iov->iov_base;
822 skb = sk->sk_write_queue.prev;
824 if (!sk->sk_send_head ||
825 (copy = mss_now - skb->len) <= 0) {
828 /* Allocate new segment. If the interface is SG,
829 * allocate skb fitting to single page.
831 if (!sk_stream_memory_free(sk))
832 goto wait_for_sndbuf;
834 skb = sk_stream_alloc_pskb(sk, select_size(sk, tp),
835 0, sk->sk_allocation);
837 goto wait_for_memory;
840 * Check whether we can use HW checksum.
842 if (sk->sk_route_caps &
843 (NETIF_F_IP_CSUM | NETIF_F_NO_CSUM |
845 skb->ip_summed = CHECKSUM_HW;
847 skb_entail(sk, tp, skb);
851 /* Try to append data to the end of skb. */
855 /* Where to copy to? */
856 if (skb_tailroom(skb) > 0) {
857 /* We have some space in skb head. Superb! */
858 if (copy > skb_tailroom(skb))
859 copy = skb_tailroom(skb);
860 if ((err = skb_add_data(skb, from, copy)) != 0)
864 int i = skb_shinfo(skb)->nr_frags;
865 struct page *page = TCP_PAGE(sk);
866 int off = TCP_OFF(sk);
868 if (skb_can_coalesce(skb, i, page, off) &&
870 /* We can extend the last page
873 } else if (i == MAX_SKB_FRAGS ||
875 !(sk->sk_route_caps & NETIF_F_SG))) {
876 /* Need to add new fragment and cannot
877 * do this because interface is non-SG,
878 * or because all the page slots are
880 tcp_mark_push(tp, skb);
883 /* If page is cached, align
884 * offset to L1 cache boundary
886 off = (off + L1_CACHE_BYTES - 1) &
887 ~(L1_CACHE_BYTES - 1);
888 if (off == PAGE_SIZE) {
890 TCP_PAGE(sk) = page = NULL;
895 /* Allocate new cache page. */
896 if (!(page = sk_stream_alloc_page(sk)))
897 goto wait_for_memory;
901 if (copy > PAGE_SIZE - off)
902 copy = PAGE_SIZE - off;
904 /* Time to copy data. We are close to
906 err = skb_copy_to_page(sk, from, skb, page,
909 /* If this page was new, give it to the
910 * socket so it does not get leaked.
919 /* Update the skb. */
921 skb_shinfo(skb)->frags[i - 1].size +=
924 skb_fill_page_desc(skb, i, page, off, copy);
927 } else if (off + copy < PAGE_SIZE) {
933 TCP_OFF(sk) = off + copy;
937 TCP_SKB_CB(skb)->flags &= ~TCPCB_FLAG_PSH;
939 tp->write_seq += copy;
940 TCP_SKB_CB(skb)->end_seq += copy;
941 skb_shinfo(skb)->tso_segs = 0;
945 if ((seglen -= copy) == 0 && iovlen == 0)
948 if (skb->len != mss_now || (flags & MSG_OOB))
951 if (forced_push(tp)) {
952 tcp_mark_push(tp, skb);
953 __tcp_push_pending_frames(sk, tp, mss_now, TCP_NAGLE_PUSH);
954 } else if (skb == sk->sk_send_head)
955 tcp_push_one(sk, mss_now);
959 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
962 tcp_push(sk, tp, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH);
964 if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
967 mss_now = tcp_current_mss(sk, !(flags&MSG_OOB));
973 tcp_push(sk, tp, flags, mss_now, tp->nonagle);
980 if (sk->sk_send_head == skb)
981 sk->sk_send_head = NULL;
982 __skb_unlink(skb, skb->list);
983 sk_stream_free_skb(sk, skb);
990 err = sk_stream_error(sk, flags, err);
997 * Handle reading urgent data. BSD has very simple semantics for
998 * this, no blocking and very strange errors 8)
1001 static int tcp_recv_urg(struct sock *sk, long timeo,
1002 struct msghdr *msg, int len, int flags,
1005 struct tcp_opt *tp = tcp_sk(sk);
1007 /* No URG data to read. */
1008 if (sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data ||
1009 tp->urg_data == TCP_URG_READ)
1010 return -EINVAL; /* Yes this is right ! */
1012 if (sk->sk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DONE))
1015 if (tp->urg_data & TCP_URG_VALID) {
1017 char c = tp->urg_data;
1019 if (!(flags & MSG_PEEK))
1020 tp->urg_data = TCP_URG_READ;
1022 /* Read urgent data. */
1023 msg->msg_flags |= MSG_OOB;
1026 if (!(flags & MSG_TRUNC))
1027 err = memcpy_toiovec(msg->msg_iov, &c, 1);
1030 msg->msg_flags |= MSG_TRUNC;
1032 return err ? -EFAULT : len;
1035 if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
1038 /* Fixed the recv(..., MSG_OOB) behaviour. BSD docs and
1039 * the available implementations agree in this case:
1040 * this call should never block, independent of the
1041 * blocking state of the socket.
1042 * Mike <pall@rz.uni-karlsruhe.de>
1047 /* Clean up the receive buffer for full frames taken by the user,
1048 * then send an ACK if necessary. COPIED is the number of bytes
1049 * tcp_recvmsg has given to the user so far, it speeds up the
1050 * calculation of whether or not we must ACK for the sake of
1053 void cleanup_rbuf(struct sock *sk, int copied)
1055 struct tcp_opt *tp = tcp_sk(sk);
1056 int time_to_ack = 0;
1059 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
1061 BUG_TRAP(!skb || before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq));
1064 if (tcp_ack_scheduled(tp)) {
1065 /* Delayed ACKs frequently hit locked sockets during bulk
1067 if (tp->ack.blocked ||
1068 /* Once-per-two-segments ACK was not sent by tcp_input.c */
1069 tp->rcv_nxt - tp->rcv_wup > tp->ack.rcv_mss ||
1071 * If this read emptied read buffer, we send ACK, if
1072 * connection is not bidirectional, user drained
1073 * receive buffer and there was a small segment
1076 (copied > 0 && (tp->ack.pending & TCP_ACK_PUSHED) &&
1077 !tp->ack.pingpong && !atomic_read(&sk->sk_rmem_alloc)))
1081 /* We send an ACK if we can now advertise a non-zero window
1082 * which has been raised "significantly".
1084 * Even if window raised up to infinity, do not send window open ACK
1085 * in states, where we will not receive more. It is useless.
1087 if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) {
1088 __u32 rcv_window_now = tcp_receive_window(tp);
1090 /* Optimize, __tcp_select_window() is not cheap. */
1091 if (2*rcv_window_now <= tp->window_clamp) {
1092 __u32 new_window = __tcp_select_window(sk);
1094 /* Send ACK now, if this read freed lots of space
1095 * in our buffer. Certainly, new_window is new window.
1096 * We can advertise it now, if it is not less than current one.
1097 * "Lots" means "at least twice" here.
1099 if (new_window && new_window >= 2 * rcv_window_now)
1107 static void tcp_prequeue_process(struct sock *sk)
1109 struct sk_buff *skb;
1110 struct tcp_opt *tp = tcp_sk(sk);
1112 NET_ADD_STATS_USER(LINUX_MIB_TCPPREQUEUED, skb_queue_len(&tp->ucopy.prequeue));
1114 /* RX process wants to run with disabled BHs, though it is not
1117 while ((skb = __skb_dequeue(&tp->ucopy.prequeue)) != NULL)
1118 sk->sk_backlog_rcv(sk, skb);
1121 /* Clear memory counter. */
1122 tp->ucopy.memory = 0;
1125 static inline struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off)
1127 struct sk_buff *skb;
1130 skb_queue_walk(&sk->sk_receive_queue, skb) {
1131 offset = seq - TCP_SKB_CB(skb)->seq;
1134 if (offset < skb->len || skb->h.th->fin) {
1143 * This routine provides an alternative to tcp_recvmsg() for routines
1144 * that would like to handle copying from skbuffs directly in 'sendfile'
1147 * - It is assumed that the socket was locked by the caller.
1148 * - The routine does not block.
1149 * - At present, there is no support for reading OOB data
1150 * or for 'peeking' the socket using this routine
1151 * (although both would be easy to implement).
1153 int tcp_read_sock(struct sock *sk, read_descriptor_t *desc,
1154 sk_read_actor_t recv_actor)
1156 struct sk_buff *skb;
1157 struct tcp_opt *tp = tcp_sk(sk);
1158 u32 seq = tp->copied_seq;
1162 if (sk->sk_state == TCP_LISTEN)
1164 while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1165 if (offset < skb->len) {
1168 len = skb->len - offset;
1169 /* Stop reading if we hit a patch of urgent data */
1171 u32 urg_offset = tp->urg_seq - seq;
1172 if (urg_offset < len)
1177 used = recv_actor(desc, skb, offset, len);
1183 if (offset != skb->len)
1186 if (skb->h.th->fin) {
1187 sk_eat_skb(sk, skb);
1191 sk_eat_skb(sk, skb);
1195 tp->copied_seq = seq;
1197 tcp_rcv_space_adjust(sk);
1199 /* Clean up data we have read: This will do ACK frames. */
1201 cleanup_rbuf(sk, copied);
1206 * This routine copies from a sock struct into the user buffer.
1208 * Technical note: in 2.3 we work on _locked_ socket, so that
1209 * tricks with *seq access order and skb->users are not required.
1210 * Probably, code can be easily improved even more.
1213 int tcp_recvmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg,
1214 size_t len, int nonblock, int flags, int *addr_len)
1216 struct tcp_opt *tp = tcp_sk(sk);
1222 int target; /* Read at least this many bytes */
1224 struct task_struct *user_recv = NULL;
1228 TCP_CHECK_TIMER(sk);
1231 if (sk->sk_state == TCP_LISTEN)
1234 timeo = sock_rcvtimeo(sk, nonblock);
1236 /* Urgent data needs to be handled specially. */
1237 if (flags & MSG_OOB)
1240 seq = &tp->copied_seq;
1241 if (flags & MSG_PEEK) {
1242 peek_seq = tp->copied_seq;
1246 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
1249 struct sk_buff *skb;
1252 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
1253 if (tp->urg_data && tp->urg_seq == *seq) {
1256 if (signal_pending(current)) {
1257 copied = timeo ? sock_intr_errno(timeo) : -EAGAIN;
1262 /* Next get a buffer. */
1264 skb = skb_peek(&sk->sk_receive_queue);
1269 /* Now that we have two receive queues this
1272 if (before(*seq, TCP_SKB_CB(skb)->seq)) {
1273 printk(KERN_INFO "recvmsg bug: copied %X "
1274 "seq %X\n", *seq, TCP_SKB_CB(skb)->seq);
1277 offset = *seq - TCP_SKB_CB(skb)->seq;
1280 if (offset < skb->len)
1284 BUG_TRAP(flags & MSG_PEEK);
1286 } while (skb != (struct sk_buff *)&sk->sk_receive_queue);
1288 /* Well, if we have backlog, try to process it now yet. */
1290 if (copied >= target && !sk->sk_backlog.tail)
1295 sk->sk_state == TCP_CLOSE ||
1296 (sk->sk_shutdown & RCV_SHUTDOWN) ||
1298 signal_pending(current) ||
1302 if (sock_flag(sk, SOCK_DONE))
1306 copied = sock_error(sk);
1310 if (sk->sk_shutdown & RCV_SHUTDOWN)
1313 if (sk->sk_state == TCP_CLOSE) {
1314 if (!sock_flag(sk, SOCK_DONE)) {
1315 /* This occurs when user tries to read
1316 * from never connected socket.
1329 if (signal_pending(current)) {
1330 copied = sock_intr_errno(timeo);
1335 cleanup_rbuf(sk, copied);
1337 if (tp->ucopy.task == user_recv) {
1338 /* Install new reader */
1339 if (!user_recv && !(flags & (MSG_TRUNC | MSG_PEEK))) {
1340 user_recv = current;
1341 tp->ucopy.task = user_recv;
1342 tp->ucopy.iov = msg->msg_iov;
1345 tp->ucopy.len = len;
1347 BUG_TRAP(tp->copied_seq == tp->rcv_nxt ||
1348 (flags & (MSG_PEEK | MSG_TRUNC)));
1350 /* Ugly... If prequeue is not empty, we have to
1351 * process it before releasing socket, otherwise
1352 * order will be broken at second iteration.
1353 * More elegant solution is required!!!
1355 * Look: we have the following (pseudo)queues:
1357 * 1. packets in flight
1362 * Each queue can be processed only if the next ones
1363 * are empty. At this point we have empty receive_queue.
1364 * But prequeue _can_ be not empty after 2nd iteration,
1365 * when we jumped to start of loop because backlog
1366 * processing added something to receive_queue.
1367 * We cannot release_sock(), because backlog contains
1368 * packets arrived _after_ prequeued ones.
1370 * Shortly, algorithm is clear --- to process all
1371 * the queues in order. We could make it more directly,
1372 * requeueing packets from backlog to prequeue, if
1373 * is not empty. It is more elegant, but eats cycles,
1376 if (skb_queue_len(&tp->ucopy.prequeue))
1379 /* __ Set realtime policy in scheduler __ */
1382 if (copied >= target) {
1383 /* Do not sleep, just process backlog. */
1387 sk_wait_data(sk, &timeo);
1392 /* __ Restore normal policy in scheduler __ */
1394 if ((chunk = len - tp->ucopy.len) != 0) {
1395 NET_ADD_STATS_USER(LINUX_MIB_TCPDIRECTCOPYFROMBACKLOG, chunk);
1400 if (tp->rcv_nxt == tp->copied_seq &&
1401 skb_queue_len(&tp->ucopy.prequeue)) {
1403 tcp_prequeue_process(sk);
1405 if ((chunk = len - tp->ucopy.len) != 0) {
1406 NET_ADD_STATS_USER(LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1412 if ((flags & MSG_PEEK) && peek_seq != tp->copied_seq) {
1413 if (net_ratelimit())
1414 printk(KERN_DEBUG "TCP(%s:%d): Application bug, race in MSG_PEEK.\n",
1415 current->comm, current->pid);
1416 peek_seq = tp->copied_seq;
1421 /* Ok so how much can we use? */
1422 used = skb->len - offset;
1426 /* Do we have urgent data here? */
1428 u32 urg_offset = tp->urg_seq - *seq;
1429 if (urg_offset < used) {
1431 if (!sock_flag(sk, SOCK_URGINLINE)) {
1443 if (!(flags & MSG_TRUNC)) {
1444 err = skb_copy_datagram_iovec(skb, offset,
1445 msg->msg_iov, used);
1447 /* Exception. Bailout! */
1458 tcp_rcv_space_adjust(sk);
1461 if (tp->urg_data && after(tp->copied_seq, tp->urg_seq)) {
1463 tcp_fast_path_check(sk, tp);
1465 if (used + offset < skb->len)
1470 if (!(flags & MSG_PEEK))
1471 sk_eat_skb(sk, skb);
1475 /* Process the FIN. */
1477 if (!(flags & MSG_PEEK))
1478 sk_eat_skb(sk, skb);
1483 if (skb_queue_len(&tp->ucopy.prequeue)) {
1486 tp->ucopy.len = copied > 0 ? len : 0;
1488 tcp_prequeue_process(sk);
1490 if (copied > 0 && (chunk = len - tp->ucopy.len) != 0) {
1491 NET_ADD_STATS_USER(LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1497 tp->ucopy.task = NULL;
1501 /* According to UNIX98, msg_name/msg_namelen are ignored
1502 * on connected socket. I was just happy when found this 8) --ANK
1505 /* Clean up data we have read: This will do ACK frames. */
1506 cleanup_rbuf(sk, copied);
1508 TCP_CHECK_TIMER(sk);
1513 TCP_CHECK_TIMER(sk);
1518 err = tcp_recv_urg(sk, timeo, msg, len, flags, addr_len);
1523 * State processing on a close. This implements the state shift for
1524 * sending our FIN frame. Note that we only send a FIN for some
1525 * states. A shutdown() may have already sent the FIN, or we may be
1529 static unsigned char new_state[16] = {
1530 /* current state: new state: action: */
1531 /* (Invalid) */ TCP_CLOSE,
1532 /* TCP_ESTABLISHED */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1533 /* TCP_SYN_SENT */ TCP_CLOSE,
1534 /* TCP_SYN_RECV */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1535 /* TCP_FIN_WAIT1 */ TCP_FIN_WAIT1,
1536 /* TCP_FIN_WAIT2 */ TCP_FIN_WAIT2,
1537 /* TCP_TIME_WAIT */ TCP_CLOSE,
1538 /* TCP_CLOSE */ TCP_CLOSE,
1539 /* TCP_CLOSE_WAIT */ TCP_LAST_ACK | TCP_ACTION_FIN,
1540 /* TCP_LAST_ACK */ TCP_LAST_ACK,
1541 /* TCP_LISTEN */ TCP_CLOSE,
1542 /* TCP_CLOSING */ TCP_CLOSING,
1545 static int tcp_close_state(struct sock *sk)
1547 int next = (int)new_state[sk->sk_state];
1548 int ns = next & TCP_STATE_MASK;
1550 tcp_set_state(sk, ns);
1552 return next & TCP_ACTION_FIN;
1556 * Shutdown the sending side of a connection. Much like close except
1557 * that we don't receive shut down or set_sock_flag(sk, SOCK_DEAD).
1560 void tcp_shutdown(struct sock *sk, int how)
1562 /* We need to grab some memory, and put together a FIN,
1563 * and then put it into the queue to be sent.
1564 * Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
1566 if (!(how & SEND_SHUTDOWN))
1569 /* If we've already sent a FIN, or it's a closed state, skip this. */
1570 if ((1 << sk->sk_state) &
1571 (TCPF_ESTABLISHED | TCPF_SYN_SENT |
1572 TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) {
1573 /* Clear out any half completed packets. FIN if needed. */
1574 if (tcp_close_state(sk))
1580 * At this point, there should be no process reference to this
1581 * socket, and thus no user references at all. Therefore we
1582 * can assume the socket waitqueue is inactive and nobody will
1583 * try to jump onto it.
1585 void tcp_destroy_sock(struct sock *sk)
1587 BUG_TRAP(sk->sk_state == TCP_CLOSE);
1588 BUG_TRAP(sock_flag(sk, SOCK_DEAD));
1590 /* It cannot be in hash table! */
1591 BUG_TRAP(sk_unhashed(sk));
1593 /* If it has not 0 inet_sk(sk)->num, it must be bound */
1594 BUG_TRAP(!inet_sk(sk)->num || tcp_sk(sk)->bind_hash);
1596 sk->sk_prot->destroy(sk);
1598 sk_stream_kill_queues(sk);
1600 xfrm_sk_free_policy(sk);
1602 #ifdef INET_REFCNT_DEBUG
1603 if (atomic_read(&sk->sk_refcnt) != 1) {
1604 printk(KERN_DEBUG "Destruction TCP %p delayed, c=%d\n",
1605 sk, atomic_read(&sk->sk_refcnt));
1609 atomic_dec(&tcp_orphan_count);
1613 void tcp_close(struct sock *sk, long timeout)
1615 struct sk_buff *skb;
1616 int data_was_unread = 0;
1619 sk->sk_shutdown = SHUTDOWN_MASK;
1621 if (sk->sk_state == TCP_LISTEN) {
1622 tcp_set_state(sk, TCP_CLOSE);
1625 tcp_listen_stop(sk);
1627 goto adjudge_to_death;
1630 /* We need to flush the recv. buffs. We do this only on the
1631 * descriptor close, not protocol-sourced closes, because the
1632 * reader process may not have drained the data yet!
1634 while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) {
1635 u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq -
1637 data_was_unread += len;
1641 sk_stream_mem_reclaim(sk);
1643 /* As outlined in draft-ietf-tcpimpl-prob-03.txt, section
1644 * 3.10, we send a RST here because data was lost. To
1645 * witness the awful effects of the old behavior of always
1646 * doing a FIN, run an older 2.1.x kernel or 2.0.x, start
1647 * a bulk GET in an FTP client, suspend the process, wait
1648 * for the client to advertise a zero window, then kill -9
1649 * the FTP client, wheee... Note: timeout is always zero
1652 if (data_was_unread) {
1653 /* Unread data was tossed, zap the connection. */
1654 NET_INC_STATS_USER(LINUX_MIB_TCPABORTONCLOSE);
1655 tcp_set_state(sk, TCP_CLOSE);
1656 tcp_send_active_reset(sk, GFP_KERNEL);
1657 } else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) {
1658 /* Check zero linger _after_ checking for unread data. */
1659 sk->sk_prot->disconnect(sk, 0);
1660 NET_INC_STATS_USER(LINUX_MIB_TCPABORTONDATA);
1661 } else if (tcp_close_state(sk)) {
1662 /* We FIN if the application ate all the data before
1663 * zapping the connection.
1666 /* RED-PEN. Formally speaking, we have broken TCP state
1667 * machine. State transitions:
1669 * TCP_ESTABLISHED -> TCP_FIN_WAIT1
1670 * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible)
1671 * TCP_CLOSE_WAIT -> TCP_LAST_ACK
1673 * are legal only when FIN has been sent (i.e. in window),
1674 * rather than queued out of window. Purists blame.
1676 * F.e. "RFC state" is ESTABLISHED,
1677 * if Linux state is FIN-WAIT-1, but FIN is still not sent.
1679 * The visible declinations are that sometimes
1680 * we enter time-wait state, when it is not required really
1681 * (harmless), do not send active resets, when they are
1682 * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
1683 * they look as CLOSING or LAST_ACK for Linux)
1684 * Probably, I missed some more holelets.
1690 sk_stream_wait_close(sk, timeout);
1693 /* It is the last release_sock in its life. It will remove backlog. */
1697 /* Now socket is owned by kernel and we acquire BH lock
1698 to finish close. No need to check for user refs.
1702 BUG_TRAP(!sock_owned_by_user(sk));
1707 /* This is a (useful) BSD violating of the RFC. There is a
1708 * problem with TCP as specified in that the other end could
1709 * keep a socket open forever with no application left this end.
1710 * We use a 3 minute timeout (about the same as BSD) then kill
1711 * our end. If they send after that then tough - BUT: long enough
1712 * that we won't make the old 4*rto = almost no time - whoops
1715 * Nope, it was not mistake. It is really desired behaviour
1716 * f.e. on http servers, when such sockets are useless, but
1717 * consume significant resources. Let's do it with special
1718 * linger2 option. --ANK
1721 if (sk->sk_state == TCP_FIN_WAIT2) {
1722 struct tcp_opt *tp = tcp_sk(sk);
1723 if (tp->linger2 < 0) {
1724 tcp_set_state(sk, TCP_CLOSE);
1725 tcp_send_active_reset(sk, GFP_ATOMIC);
1726 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONLINGER);
1728 int tmo = tcp_fin_time(tp);
1730 if (tmo > TCP_TIMEWAIT_LEN) {
1731 tcp_reset_keepalive_timer(sk, tcp_fin_time(tp));
1733 atomic_inc(&tcp_orphan_count);
1734 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
1739 if (sk->sk_state != TCP_CLOSE) {
1740 sk_stream_mem_reclaim(sk);
1741 if (atomic_read(&tcp_orphan_count) > sysctl_tcp_max_orphans ||
1742 (sk->sk_wmem_queued > SOCK_MIN_SNDBUF &&
1743 atomic_read(&tcp_memory_allocated) > sysctl_tcp_mem[2])) {
1744 if (net_ratelimit())
1745 printk(KERN_INFO "TCP: too many of orphaned "
1747 tcp_set_state(sk, TCP_CLOSE);
1748 tcp_send_active_reset(sk, GFP_ATOMIC);
1749 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONMEMORY);
1752 atomic_inc(&tcp_orphan_count);
1754 if (sk->sk_state == TCP_CLOSE)
1755 tcp_destroy_sock(sk);
1756 /* Otherwise, socket is reprieved until protocol close. */
1764 /* These states need RST on ABORT according to RFC793 */
1766 static inline int tcp_need_reset(int state)
1768 return (1 << state) &
1769 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 |
1770 TCPF_FIN_WAIT2 | TCPF_SYN_RECV);
1773 int tcp_disconnect(struct sock *sk, int flags)
1775 struct inet_opt *inet = inet_sk(sk);
1776 struct tcp_opt *tp = tcp_sk(sk);
1778 int old_state = sk->sk_state;
1780 if (old_state != TCP_CLOSE)
1781 tcp_set_state(sk, TCP_CLOSE);
1783 /* ABORT function of RFC793 */
1784 if (old_state == TCP_LISTEN) {
1785 tcp_listen_stop(sk);
1786 } else if (tcp_need_reset(old_state) ||
1787 (tp->snd_nxt != tp->write_seq &&
1788 (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK))) {
1789 /* The last check adjusts for discrepance of Linux wrt. RFC
1792 tcp_send_active_reset(sk, gfp_any());
1793 sk->sk_err = ECONNRESET;
1794 } else if (old_state == TCP_SYN_SENT)
1795 sk->sk_err = ECONNRESET;
1797 tcp_clear_xmit_timers(sk);
1798 __skb_queue_purge(&sk->sk_receive_queue);
1799 sk_stream_writequeue_purge(sk);
1800 __skb_queue_purge(&tp->out_of_order_queue);
1804 if (!(sk->sk_userlocks & SOCK_BINDADDR_LOCK))
1805 inet_reset_saddr(sk);
1807 sk->sk_shutdown = 0;
1808 sock_reset_flag(sk, SOCK_DONE);
1810 if ((tp->write_seq += tp->max_window + 2) == 0)
1815 tcp_set_pcount(&tp->packets_out, 0);
1816 tp->snd_ssthresh = 0x7fffffff;
1817 tp->snd_cwnd_cnt = 0;
1818 tcp_set_ca_state(tp, TCP_CA_Open);
1819 tcp_clear_retrans(tp);
1820 tcp_delack_init(tp);
1821 sk->sk_send_head = NULL;
1826 BUG_TRAP(!inet->num || tp->bind_hash);
1828 sk->sk_error_report(sk);
1833 * Wait for an incoming connection, avoid race
1834 * conditions. This must be called with the socket locked.
1836 static int wait_for_connect(struct sock *sk, long timeo)
1838 struct tcp_opt *tp = tcp_sk(sk);
1843 * True wake-one mechanism for incoming connections: only
1844 * one process gets woken up, not the 'whole herd'.
1845 * Since we do not 'race & poll' for established sockets
1846 * anymore, the common case will execute the loop only once.
1848 * Subtle issue: "add_wait_queue_exclusive()" will be added
1849 * after any current non-exclusive waiters, and we know that
1850 * it will always _stay_ after any new non-exclusive waiters
1851 * because all non-exclusive waiters are added at the
1852 * beginning of the wait-queue. As such, it's ok to "drop"
1853 * our exclusiveness temporarily when we get woken up without
1854 * having to remove and re-insert us on the wait queue.
1857 prepare_to_wait_exclusive(sk->sk_sleep, &wait,
1858 TASK_INTERRUPTIBLE);
1860 if (!tp->accept_queue)
1861 timeo = schedule_timeout(timeo);
1864 if (tp->accept_queue)
1867 if (sk->sk_state != TCP_LISTEN)
1869 err = sock_intr_errno(timeo);
1870 if (signal_pending(current))
1876 finish_wait(sk->sk_sleep, &wait);
1881 * This will accept the next outstanding connection.
1884 struct sock *tcp_accept(struct sock *sk, int flags, int *err)
1886 struct tcp_opt *tp = tcp_sk(sk);
1887 struct open_request *req;
1893 /* We need to make sure that this socket is listening,
1894 * and that it has something pending.
1897 if (sk->sk_state != TCP_LISTEN)
1900 /* Find already established connection */
1901 if (!tp->accept_queue) {
1902 long timeo = sock_rcvtimeo(sk, flags & O_NONBLOCK);
1903 /* If this is a non blocking socket don't sleep */
1908 error = wait_for_connect(sk, timeo);
1913 req = tp->accept_queue;
1914 if ((tp->accept_queue = req->dl_next) == NULL)
1915 tp->accept_queue_tail = NULL;
1917 sk_acceptq_removed(sk);
1918 tcp_openreq_fastfree(req);
1919 BUG_TRAP(newsk->sk_state != TCP_SYN_RECV);
1931 * Socket option code for TCP.
1933 int tcp_setsockopt(struct sock *sk, int level, int optname, char __user *optval,
1936 struct tcp_opt *tp = tcp_sk(sk);
1940 if (level != SOL_TCP)
1941 return tp->af_specific->setsockopt(sk, level, optname,
1944 if (optlen < sizeof(int))
1947 if (get_user(val, (int __user *)optval))
1954 /* Values greater than interface MTU won't take effect. However
1955 * at the point when this call is done we typically don't yet
1956 * know which interface is going to be used */
1957 if (val < 8 || val > MAX_TCP_WINDOW) {
1966 /* TCP_NODELAY is weaker than TCP_CORK, so that
1967 * this option on corked socket is remembered, but
1968 * it is not activated until cork is cleared.
1970 * However, when TCP_NODELAY is set we make
1971 * an explicit push, which overrides even TCP_CORK
1972 * for currently queued segments.
1974 tp->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH;
1975 tcp_push_pending_frames(sk, tp);
1977 tp->nonagle &= ~TCP_NAGLE_OFF;
1982 /* When set indicates to always queue non-full frames.
1983 * Later the user clears this option and we transmit
1984 * any pending partial frames in the queue. This is
1985 * meant to be used alongside sendfile() to get properly
1986 * filled frames when the user (for example) must write
1987 * out headers with a write() call first and then use
1988 * sendfile to send out the data parts.
1990 * TCP_CORK can be set together with TCP_NODELAY and it is
1991 * stronger than TCP_NODELAY.
1994 tp->nonagle |= TCP_NAGLE_CORK;
1996 tp->nonagle &= ~TCP_NAGLE_CORK;
1997 if (tp->nonagle&TCP_NAGLE_OFF)
1998 tp->nonagle |= TCP_NAGLE_PUSH;
1999 tcp_push_pending_frames(sk, tp);
2004 if (val < 1 || val > MAX_TCP_KEEPIDLE)
2007 tp->keepalive_time = val * HZ;
2008 if (sock_flag(sk, SOCK_KEEPOPEN) &&
2009 !((1 << sk->sk_state) &
2010 (TCPF_CLOSE | TCPF_LISTEN))) {
2011 __u32 elapsed = tcp_time_stamp - tp->rcv_tstamp;
2012 if (tp->keepalive_time > elapsed)
2013 elapsed = tp->keepalive_time - elapsed;
2016 tcp_reset_keepalive_timer(sk, elapsed);
2021 if (val < 1 || val > MAX_TCP_KEEPINTVL)
2024 tp->keepalive_intvl = val * HZ;
2027 if (val < 1 || val > MAX_TCP_KEEPCNT)
2030 tp->keepalive_probes = val;
2033 if (val < 1 || val > MAX_TCP_SYNCNT)
2036 tp->syn_retries = val;
2042 else if (val > sysctl_tcp_fin_timeout / HZ)
2045 tp->linger2 = val * HZ;
2048 case TCP_DEFER_ACCEPT:
2049 tp->defer_accept = 0;
2051 /* Translate value in seconds to number of
2053 while (tp->defer_accept < 32 &&
2054 val > ((TCP_TIMEOUT_INIT / HZ) <<
2061 case TCP_WINDOW_CLAMP:
2063 if (sk->sk_state != TCP_CLOSE) {
2067 tp->window_clamp = 0;
2069 tp->window_clamp = val < SOCK_MIN_RCVBUF / 2 ?
2070 SOCK_MIN_RCVBUF / 2 : val;
2075 tp->ack.pingpong = 1;
2077 tp->ack.pingpong = 0;
2078 if ((1 << sk->sk_state) &
2079 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) &&
2080 tcp_ack_scheduled(tp)) {
2081 tp->ack.pending |= TCP_ACK_PUSHED;
2082 cleanup_rbuf(sk, 1);
2084 tp->ack.pingpong = 1;
2097 /* Return information about state of tcp endpoint in API format. */
2098 void tcp_get_info(struct sock *sk, struct tcp_info *info)
2100 struct tcp_opt *tp = tcp_sk(sk);
2101 u32 now = tcp_time_stamp;
2103 memset(info, 0, sizeof(*info));
2105 info->tcpi_state = sk->sk_state;
2106 info->tcpi_ca_state = tp->ca_state;
2107 info->tcpi_retransmits = tp->retransmits;
2108 info->tcpi_probes = tp->probes_out;
2109 info->tcpi_backoff = tp->backoff;
2112 info->tcpi_options |= TCPI_OPT_TIMESTAMPS;
2114 info->tcpi_options |= TCPI_OPT_SACK;
2115 if (tp->wscale_ok) {
2116 info->tcpi_options |= TCPI_OPT_WSCALE;
2117 info->tcpi_snd_wscale = tp->snd_wscale;
2118 info->tcpi_rcv_wscale = tp->rcv_wscale;
2121 if (tp->ecn_flags&TCP_ECN_OK)
2122 info->tcpi_options |= TCPI_OPT_ECN;
2124 info->tcpi_rto = jiffies_to_usecs(tp->rto);
2125 info->tcpi_ato = jiffies_to_usecs(tp->ack.ato);
2126 info->tcpi_snd_mss = tp->mss_cache_std;
2127 info->tcpi_rcv_mss = tp->ack.rcv_mss;
2129 info->tcpi_unacked = tcp_get_pcount(&tp->packets_out);
2130 info->tcpi_sacked = tcp_get_pcount(&tp->sacked_out);
2131 info->tcpi_lost = tcp_get_pcount(&tp->lost_out);
2132 info->tcpi_retrans = tcp_get_pcount(&tp->retrans_out);
2133 info->tcpi_fackets = tcp_get_pcount(&tp->fackets_out);
2135 info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime);
2136 info->tcpi_last_data_recv = jiffies_to_msecs(now - tp->ack.lrcvtime);
2137 info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp);
2139 info->tcpi_pmtu = tp->pmtu_cookie;
2140 info->tcpi_rcv_ssthresh = tp->rcv_ssthresh;
2141 info->tcpi_rtt = jiffies_to_usecs(tp->srtt)>>3;
2142 info->tcpi_rttvar = jiffies_to_usecs(tp->mdev)>>2;
2143 info->tcpi_snd_ssthresh = tp->snd_ssthresh;
2144 info->tcpi_snd_cwnd = tp->snd_cwnd;
2145 info->tcpi_advmss = tp->advmss;
2146 info->tcpi_reordering = tp->reordering;
2148 info->tcpi_rcv_rtt = jiffies_to_usecs(tp->rcv_rtt_est.rtt)>>3;
2149 info->tcpi_rcv_space = tp->rcvq_space.space;
2151 info->tcpi_total_retrans = tp->total_retrans;
2154 EXPORT_SYMBOL_GPL(tcp_get_info);
2156 int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval,
2159 struct tcp_opt *tp = tcp_sk(sk);
2162 if (level != SOL_TCP)
2163 return tp->af_specific->getsockopt(sk, level, optname,
2166 if (get_user(len, optlen))
2169 len = min_t(unsigned int, len, sizeof(int));
2176 val = tp->mss_cache_std;
2177 if (!val && ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
2181 val = !!(tp->nonagle&TCP_NAGLE_OFF);
2184 val = !!(tp->nonagle&TCP_NAGLE_CORK);
2187 val = (tp->keepalive_time ? : sysctl_tcp_keepalive_time) / HZ;
2190 val = (tp->keepalive_intvl ? : sysctl_tcp_keepalive_intvl) / HZ;
2193 val = tp->keepalive_probes ? : sysctl_tcp_keepalive_probes;
2196 val = tp->syn_retries ? : sysctl_tcp_syn_retries;
2201 val = (val ? : sysctl_tcp_fin_timeout) / HZ;
2203 case TCP_DEFER_ACCEPT:
2204 val = !tp->defer_accept ? 0 : ((TCP_TIMEOUT_INIT / HZ) <<
2205 (tp->defer_accept - 1));
2207 case TCP_WINDOW_CLAMP:
2208 val = tp->window_clamp;
2211 struct tcp_info info;
2213 if (get_user(len, optlen))
2216 tcp_get_info(sk, &info);
2218 len = min_t(unsigned int, len, sizeof(info));
2219 if (put_user(len, optlen))
2221 if (copy_to_user(optval, &info, len))
2226 val = !tp->ack.pingpong;
2230 return -ENOPROTOOPT;
2233 if (put_user(len, optlen))
2235 if (copy_to_user(optval, &val, len))
2241 extern void __skb_cb_too_small_for_tcp(int, int);
2242 extern void tcpdiag_init(void);
2244 static __initdata unsigned long thash_entries;
2245 static int __init set_thash_entries(char *str)
2249 thash_entries = simple_strtoul(str, &str, 0);
2252 __setup("thash_entries=", set_thash_entries);
2254 void __init tcp_init(void)
2256 struct sk_buff *skb = NULL;
2260 if (sizeof(struct tcp_skb_cb) > sizeof(skb->cb))
2261 __skb_cb_too_small_for_tcp(sizeof(struct tcp_skb_cb),
2264 tcp_openreq_cachep = kmem_cache_create("tcp_open_request",
2265 sizeof(struct open_request),
2266 0, SLAB_HWCACHE_ALIGN,
2268 if (!tcp_openreq_cachep)
2269 panic("tcp_init: Cannot alloc open_request cache.");
2271 tcp_bucket_cachep = kmem_cache_create("tcp_bind_bucket",
2272 sizeof(struct tcp_bind_bucket),
2273 0, SLAB_HWCACHE_ALIGN,
2275 if (!tcp_bucket_cachep)
2276 panic("tcp_init: Cannot alloc tcp_bind_bucket cache.");
2278 tcp_timewait_cachep = kmem_cache_create("tcp_tw_bucket",
2279 sizeof(struct tcp_tw_bucket),
2280 0, SLAB_HWCACHE_ALIGN,
2282 if (!tcp_timewait_cachep)
2283 panic("tcp_init: Cannot alloc tcp_tw_bucket cache.");
2285 /* Size and allocate the main established and bind bucket
2288 * The methodology is similar to that of the buffer cache.
2290 if (num_physpages >= (128 * 1024))
2291 goal = num_physpages >> (21 - PAGE_SHIFT);
2293 goal = num_physpages >> (23 - PAGE_SHIFT);
2296 goal = (thash_entries * sizeof(struct tcp_ehash_bucket)) >> PAGE_SHIFT;
2297 for (order = 0; (1UL << order) < goal; order++)
2300 tcp_ehash_size = (1UL << order) * PAGE_SIZE /
2301 sizeof(struct tcp_ehash_bucket);
2302 tcp_ehash_size >>= 1;
2303 while (tcp_ehash_size & (tcp_ehash_size - 1))
2305 tcp_ehash = (struct tcp_ehash_bucket *)
2306 __get_free_pages(GFP_ATOMIC, order);
2307 } while (!tcp_ehash && --order > 0);
2310 panic("Failed to allocate TCP established hash table\n");
2311 for (i = 0; i < (tcp_ehash_size << 1); i++) {
2312 rwlock_init(&tcp_ehash[i].lock);
2313 INIT_HLIST_HEAD(&tcp_ehash[i].chain);
2317 tcp_bhash_size = (1UL << order) * PAGE_SIZE /
2318 sizeof(struct tcp_bind_hashbucket);
2319 if ((tcp_bhash_size > (64 * 1024)) && order > 0)
2321 tcp_bhash = (struct tcp_bind_hashbucket *)
2322 __get_free_pages(GFP_ATOMIC, order);
2323 } while (!tcp_bhash && --order >= 0);
2326 panic("Failed to allocate TCP bind hash table\n");
2327 for (i = 0; i < tcp_bhash_size; i++) {
2328 spin_lock_init(&tcp_bhash[i].lock);
2329 INIT_HLIST_HEAD(&tcp_bhash[i].chain);
2332 /* Try to be a bit smarter and adjust defaults depending
2333 * on available memory.
2336 sysctl_local_port_range[0] = 32768;
2337 sysctl_local_port_range[1] = 61000;
2338 sysctl_tcp_max_tw_buckets = 180000;
2339 sysctl_tcp_max_orphans = 4096 << (order - 4);
2340 sysctl_max_syn_backlog = 1024;
2341 } else if (order < 3) {
2342 sysctl_local_port_range[0] = 1024 * (3 - order);
2343 sysctl_tcp_max_tw_buckets >>= (3 - order);
2344 sysctl_tcp_max_orphans >>= (3 - order);
2345 sysctl_max_syn_backlog = 128;
2347 tcp_port_rover = sysctl_local_port_range[0] - 1;
2349 sysctl_tcp_mem[0] = 768 << order;
2350 sysctl_tcp_mem[1] = 1024 << order;
2351 sysctl_tcp_mem[2] = 1536 << order;
2354 sysctl_tcp_wmem[2] = 64 * 1024;
2355 sysctl_tcp_rmem[0] = PAGE_SIZE;
2356 sysctl_tcp_rmem[1] = 43689;
2357 sysctl_tcp_rmem[2] = 2 * 43689;
2360 printk(KERN_INFO "TCP: Hash tables configured "
2361 "(established %d bind %d)\n",
2362 tcp_ehash_size << 1, tcp_bhash_size);
2365 EXPORT_SYMBOL(tcp_accept);
2366 EXPORT_SYMBOL(tcp_close);
2367 EXPORT_SYMBOL(tcp_destroy_sock);
2368 EXPORT_SYMBOL(tcp_disconnect);
2369 EXPORT_SYMBOL(tcp_getsockopt);
2370 EXPORT_SYMBOL(tcp_ioctl);
2371 EXPORT_SYMBOL(tcp_openreq_cachep);
2372 EXPORT_SYMBOL(tcp_poll);
2373 EXPORT_SYMBOL(tcp_read_sock);
2374 EXPORT_SYMBOL(tcp_recvmsg);
2375 EXPORT_SYMBOL(tcp_sendmsg);
2376 EXPORT_SYMBOL(tcp_sendpage);
2377 EXPORT_SYMBOL(tcp_setsockopt);
2378 EXPORT_SYMBOL(tcp_shutdown);
2379 EXPORT_SYMBOL(tcp_statistics);
2380 EXPORT_SYMBOL(tcp_timewait_cachep);
2381 EXPORT_SYMBOL_GPL(cleanup_rbuf);