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/module.h>
251 #include <linux/types.h>
252 #include <linux/fcntl.h>
253 #include <linux/poll.h>
254 #include <linux/init.h>
255 #include <linux/fs.h>
256 #include <linux/random.h>
257 #include <linux/bootmem.h>
258 #include <linux/cache.h>
259 #include <linux/err.h>
260 #include <linux/crypto.h>
262 #include <net/icmp.h>
264 #include <net/xfrm.h>
266 #include <net/netdma.h>
268 #include <asm/uaccess.h>
269 #include <asm/ioctls.h>
271 int sysctl_tcp_fin_timeout __read_mostly = TCP_FIN_TIMEOUT;
273 DEFINE_SNMP_STAT(struct tcp_mib, tcp_statistics) __read_mostly;
275 atomic_t tcp_orphan_count = ATOMIC_INIT(0);
277 EXPORT_SYMBOL_GPL(tcp_orphan_count);
279 int sysctl_tcp_mem[3] __read_mostly;
280 int sysctl_tcp_wmem[3] __read_mostly;
281 int sysctl_tcp_rmem[3] __read_mostly;
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 __read_mostly;
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 * Wait for a TCP event.
316 * Note that we don't need to lock the socket, as the upper poll layers
317 * take care of normal races (between the test and the event) and we don't
318 * go look at any of the socket buffers directly.
320 unsigned int tcp_poll(struct file *file, struct socket *sock, poll_table *wait)
323 struct sock *sk = sock->sk;
324 struct tcp_sock *tp = tcp_sk(sk);
326 poll_wait(file, sk->sk_sleep, wait);
327 if (sk->sk_state == TCP_LISTEN)
328 return inet_csk_listen_poll(sk);
330 /* Socket is not locked. We are protected from async events
331 by poll logic and correct handling of state changes
332 made by another threads is impossible in any case.
340 * POLLHUP is certainly not done right. But poll() doesn't
341 * have a notion of HUP in just one direction, and for a
342 * socket the read side is more interesting.
344 * Some poll() documentation says that POLLHUP is incompatible
345 * with the POLLOUT/POLLWR flags, so somebody should check this
346 * all. But careful, it tends to be safer to return too many
347 * bits than too few, and you can easily break real applications
348 * if you don't tell them that something has hung up!
352 * Check number 1. POLLHUP is _UNMASKABLE_ event (see UNIX98 and
353 * our fs/select.c). It means that after we received EOF,
354 * poll always returns immediately, making impossible poll() on write()
355 * in state CLOSE_WAIT. One solution is evident --- to set POLLHUP
356 * if and only if shutdown has been made in both directions.
357 * Actually, it is interesting to look how Solaris and DUX
358 * solve this dilemma. I would prefer, if PULLHUP were maskable,
359 * then we could set it on SND_SHUTDOWN. BTW examples given
360 * in Stevens' books assume exactly this behaviour, it explains
361 * why PULLHUP is incompatible with POLLOUT. --ANK
363 * NOTE. Check for TCP_CLOSE is added. The goal is to prevent
364 * blocking on fresh not-connected or disconnected socket. --ANK
366 if (sk->sk_shutdown == SHUTDOWN_MASK || sk->sk_state == TCP_CLOSE)
368 if (sk->sk_shutdown & RCV_SHUTDOWN)
369 mask |= POLLIN | POLLRDNORM | POLLRDHUP;
372 if ((1 << sk->sk_state) & ~(TCPF_SYN_SENT | TCPF_SYN_RECV)) {
373 /* Potential race condition. If read of tp below will
374 * escape above sk->sk_state, we can be illegally awaken
375 * in SYN_* states. */
376 if ((tp->rcv_nxt != tp->copied_seq) &&
377 (tp->urg_seq != tp->copied_seq ||
378 tp->rcv_nxt != tp->copied_seq + 1 ||
379 sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data))
380 mask |= POLLIN | POLLRDNORM;
382 if (!(sk->sk_shutdown & SEND_SHUTDOWN)) {
383 if (sk_stream_wspace(sk) >= sk_stream_min_wspace(sk)) {
384 mask |= POLLOUT | POLLWRNORM;
385 } else { /* send SIGIO later */
386 set_bit(SOCK_ASYNC_NOSPACE,
387 &sk->sk_socket->flags);
388 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
390 /* Race breaker. If space is freed after
391 * wspace test but before the flags are set,
392 * IO signal will be lost.
394 if (sk_stream_wspace(sk) >= sk_stream_min_wspace(sk))
395 mask |= POLLOUT | POLLWRNORM;
399 if (tp->urg_data & TCP_URG_VALID)
405 int tcp_ioctl(struct sock *sk, int cmd, unsigned long arg)
407 struct tcp_sock *tp = tcp_sk(sk);
412 if (sk->sk_state == TCP_LISTEN)
416 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
418 else if (sock_flag(sk, SOCK_URGINLINE) ||
420 before(tp->urg_seq, tp->copied_seq) ||
421 !before(tp->urg_seq, tp->rcv_nxt)) {
422 answ = tp->rcv_nxt - tp->copied_seq;
424 /* Subtract 1, if FIN is in queue. */
425 if (answ && !skb_queue_empty(&sk->sk_receive_queue))
427 tcp_hdr((struct sk_buff *)sk->sk_receive_queue.prev)->fin;
429 answ = tp->urg_seq - tp->copied_seq;
433 answ = tp->urg_data && tp->urg_seq == tp->copied_seq;
436 if (sk->sk_state == TCP_LISTEN)
439 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
442 answ = tp->write_seq - tp->snd_una;
448 return put_user(answ, (int __user *)arg);
451 static inline void tcp_mark_push(struct tcp_sock *tp, struct sk_buff *skb)
453 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH;
454 tp->pushed_seq = tp->write_seq;
457 static inline int forced_push(struct tcp_sock *tp)
459 return after(tp->write_seq, tp->pushed_seq + (tp->max_window >> 1));
462 static inline void skb_entail(struct sock *sk, struct sk_buff *skb)
464 struct tcp_sock *tp = tcp_sk(sk);
465 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
468 tcb->seq = tcb->end_seq = tp->write_seq;
469 tcb->flags = TCPCB_FLAG_ACK;
471 skb_header_release(skb);
472 tcp_add_write_queue_tail(sk, skb);
473 sk_charge_skb(sk, skb);
474 if (tp->nonagle & TCP_NAGLE_PUSH)
475 tp->nonagle &= ~TCP_NAGLE_PUSH;
478 static inline void tcp_mark_urg(struct tcp_sock *tp, int flags,
481 if (flags & MSG_OOB) {
483 tp->snd_up = tp->write_seq;
484 TCP_SKB_CB(skb)->sacked |= TCPCB_URG;
488 static inline void tcp_push(struct sock *sk, int flags, int mss_now,
491 struct tcp_sock *tp = tcp_sk(sk);
493 if (tcp_send_head(sk)) {
494 struct sk_buff *skb = tcp_write_queue_tail(sk);
495 if (!(flags & MSG_MORE) || forced_push(tp))
496 tcp_mark_push(tp, skb);
497 tcp_mark_urg(tp, flags, skb);
498 __tcp_push_pending_frames(sk, mss_now,
499 (flags & MSG_MORE) ? TCP_NAGLE_CORK : nonagle);
503 static ssize_t do_tcp_sendpages(struct sock *sk, struct page **pages, int poffset,
504 size_t psize, int flags)
506 struct tcp_sock *tp = tcp_sk(sk);
507 int mss_now, size_goal;
510 long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
512 /* Wait for a connection to finish. */
513 if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
514 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
517 clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
519 mss_now = tcp_current_mss(sk, !(flags&MSG_OOB));
520 size_goal = tp->xmit_size_goal;
524 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
528 struct sk_buff *skb = tcp_write_queue_tail(sk);
529 struct page *page = pages[poffset / PAGE_SIZE];
530 int copy, i, can_coalesce;
531 int offset = poffset % PAGE_SIZE;
532 int size = min_t(size_t, psize, PAGE_SIZE - offset);
534 if (!tcp_send_head(sk) || (copy = size_goal - skb->len) <= 0) {
536 if (!sk_stream_memory_free(sk))
537 goto wait_for_sndbuf;
539 skb = sk_stream_alloc_pskb(sk, 0, 0,
542 goto wait_for_memory;
551 i = skb_shinfo(skb)->nr_frags;
552 can_coalesce = skb_can_coalesce(skb, i, page, offset);
553 if (!can_coalesce && i >= MAX_SKB_FRAGS) {
554 tcp_mark_push(tp, skb);
557 if (!sk_stream_wmem_schedule(sk, copy))
558 goto wait_for_memory;
561 skb_shinfo(skb)->frags[i - 1].size += copy;
564 skb_fill_page_desc(skb, i, page, offset, copy);
568 skb->data_len += copy;
569 skb->truesize += copy;
570 sk->sk_wmem_queued += copy;
571 sk->sk_forward_alloc -= copy;
572 skb->ip_summed = CHECKSUM_PARTIAL;
573 tp->write_seq += copy;
574 TCP_SKB_CB(skb)->end_seq += copy;
575 skb_shinfo(skb)->gso_segs = 0;
578 TCP_SKB_CB(skb)->flags &= ~TCPCB_FLAG_PSH;
582 if (!(psize -= copy))
585 if (skb->len < mss_now || (flags & MSG_OOB))
588 if (forced_push(tp)) {
589 tcp_mark_push(tp, skb);
590 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
591 } else if (skb == tcp_send_head(sk))
592 tcp_push_one(sk, mss_now);
596 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
599 tcp_push(sk, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH);
601 if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
604 mss_now = tcp_current_mss(sk, !(flags&MSG_OOB));
605 size_goal = tp->xmit_size_goal;
610 tcp_push(sk, flags, mss_now, tp->nonagle);
617 return sk_stream_error(sk, flags, err);
620 ssize_t tcp_sendpage(struct socket *sock, struct page *page, int offset,
621 size_t size, int flags)
624 struct sock *sk = sock->sk;
626 if (!(sk->sk_route_caps & NETIF_F_SG) ||
627 !(sk->sk_route_caps & NETIF_F_ALL_CSUM))
628 return sock_no_sendpage(sock, page, offset, size, flags);
632 res = do_tcp_sendpages(sk, &page, offset, size, flags);
638 #define TCP_PAGE(sk) (sk->sk_sndmsg_page)
639 #define TCP_OFF(sk) (sk->sk_sndmsg_off)
641 static inline int select_size(struct sock *sk)
643 struct tcp_sock *tp = tcp_sk(sk);
644 int tmp = tp->mss_cache;
646 if (sk->sk_route_caps & NETIF_F_SG) {
650 int pgbreak = SKB_MAX_HEAD(MAX_TCP_HEADER);
652 if (tmp >= pgbreak &&
653 tmp <= pgbreak + (MAX_SKB_FRAGS - 1) * PAGE_SIZE)
661 int tcp_sendmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg,
665 struct tcp_sock *tp = tcp_sk(sk);
668 int mss_now, size_goal;
675 flags = msg->msg_flags;
676 timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
678 /* Wait for a connection to finish. */
679 if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
680 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
683 /* This should be in poll */
684 clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
686 mss_now = tcp_current_mss(sk, !(flags&MSG_OOB));
687 size_goal = tp->xmit_size_goal;
689 /* Ok commence sending. */
690 iovlen = msg->msg_iovlen;
695 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
698 while (--iovlen >= 0) {
699 int seglen = iov->iov_len;
700 unsigned char __user *from = iov->iov_base;
707 skb = tcp_write_queue_tail(sk);
709 if (!tcp_send_head(sk) ||
710 (copy = size_goal - skb->len) <= 0) {
713 /* Allocate new segment. If the interface is SG,
714 * allocate skb fitting to single page.
716 if (!sk_stream_memory_free(sk))
717 goto wait_for_sndbuf;
719 skb = sk_stream_alloc_pskb(sk, select_size(sk),
720 0, sk->sk_allocation);
722 goto wait_for_memory;
725 * Check whether we can use HW checksum.
727 if (sk->sk_route_caps & NETIF_F_ALL_CSUM)
728 skb->ip_summed = CHECKSUM_PARTIAL;
734 /* Try to append data to the end of skb. */
738 /* Where to copy to? */
739 if (skb_tailroom(skb) > 0) {
740 /* We have some space in skb head. Superb! */
741 if (copy > skb_tailroom(skb))
742 copy = skb_tailroom(skb);
743 if ((err = skb_add_data(skb, from, copy)) != 0)
747 int i = skb_shinfo(skb)->nr_frags;
748 struct page *page = TCP_PAGE(sk);
749 int off = TCP_OFF(sk);
751 if (skb_can_coalesce(skb, i, page, off) &&
753 /* We can extend the last page
756 } else if (i == MAX_SKB_FRAGS ||
758 !(sk->sk_route_caps & NETIF_F_SG))) {
759 /* Need to add new fragment and cannot
760 * do this because interface is non-SG,
761 * or because all the page slots are
763 tcp_mark_push(tp, skb);
766 if (off == PAGE_SIZE) {
768 TCP_PAGE(sk) = page = NULL;
774 if (copy > PAGE_SIZE - off)
775 copy = PAGE_SIZE - off;
777 if (!sk_stream_wmem_schedule(sk, copy))
778 goto wait_for_memory;
781 /* Allocate new cache page. */
782 if (!(page = sk_stream_alloc_page(sk)))
783 goto wait_for_memory;
786 /* Time to copy data. We are close to
788 err = skb_copy_to_page(sk, from, skb, page,
791 /* If this page was new, give it to the
792 * socket so it does not get leaked.
801 /* Update the skb. */
803 skb_shinfo(skb)->frags[i - 1].size +=
806 skb_fill_page_desc(skb, i, page, off, copy);
809 } else if (off + copy < PAGE_SIZE) {
815 TCP_OFF(sk) = off + copy;
819 TCP_SKB_CB(skb)->flags &= ~TCPCB_FLAG_PSH;
821 tp->write_seq += copy;
822 TCP_SKB_CB(skb)->end_seq += copy;
823 skb_shinfo(skb)->gso_segs = 0;
827 if ((seglen -= copy) == 0 && iovlen == 0)
830 if (skb->len < mss_now || (flags & MSG_OOB))
833 if (forced_push(tp)) {
834 tcp_mark_push(tp, skb);
835 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
836 } else if (skb == tcp_send_head(sk))
837 tcp_push_one(sk, mss_now);
841 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
844 tcp_push(sk, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH);
846 if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
849 mss_now = tcp_current_mss(sk, !(flags&MSG_OOB));
850 size_goal = tp->xmit_size_goal;
856 tcp_push(sk, flags, mss_now, tp->nonagle);
863 tcp_unlink_write_queue(skb, sk);
864 /* It is the one place in all of TCP, except connection
865 * reset, where we can be unlinking the send_head.
867 tcp_check_send_head(sk, skb);
868 sk_stream_free_skb(sk, skb);
875 err = sk_stream_error(sk, flags, err);
882 * Handle reading urgent data. BSD has very simple semantics for
883 * this, no blocking and very strange errors 8)
886 static int tcp_recv_urg(struct sock *sk, long timeo,
887 struct msghdr *msg, int len, int flags,
890 struct tcp_sock *tp = tcp_sk(sk);
892 /* No URG data to read. */
893 if (sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data ||
894 tp->urg_data == TCP_URG_READ)
895 return -EINVAL; /* Yes this is right ! */
897 if (sk->sk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DONE))
900 if (tp->urg_data & TCP_URG_VALID) {
902 char c = tp->urg_data;
904 if (!(flags & MSG_PEEK))
905 tp->urg_data = TCP_URG_READ;
907 /* Read urgent data. */
908 msg->msg_flags |= MSG_OOB;
911 if (!(flags & MSG_TRUNC))
912 err = memcpy_toiovec(msg->msg_iov, &c, 1);
915 msg->msg_flags |= MSG_TRUNC;
917 return err ? -EFAULT : len;
920 if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
923 /* Fixed the recv(..., MSG_OOB) behaviour. BSD docs and
924 * the available implementations agree in this case:
925 * this call should never block, independent of the
926 * blocking state of the socket.
927 * Mike <pall@rz.uni-karlsruhe.de>
932 /* Clean up the receive buffer for full frames taken by the user,
933 * then send an ACK if necessary. COPIED is the number of bytes
934 * tcp_recvmsg has given to the user so far, it speeds up the
935 * calculation of whether or not we must ACK for the sake of
938 void tcp_cleanup_rbuf(struct sock *sk, int copied)
940 struct tcp_sock *tp = tcp_sk(sk);
944 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
946 BUG_TRAP(!skb || before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq));
949 if (inet_csk_ack_scheduled(sk)) {
950 const struct inet_connection_sock *icsk = inet_csk(sk);
951 /* Delayed ACKs frequently hit locked sockets during bulk
953 if (icsk->icsk_ack.blocked ||
954 /* Once-per-two-segments ACK was not sent by tcp_input.c */
955 tp->rcv_nxt - tp->rcv_wup > icsk->icsk_ack.rcv_mss ||
957 * If this read emptied read buffer, we send ACK, if
958 * connection is not bidirectional, user drained
959 * receive buffer and there was a small segment
963 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED2) ||
964 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED) &&
965 !icsk->icsk_ack.pingpong)) &&
966 !atomic_read(&sk->sk_rmem_alloc)))
970 /* We send an ACK if we can now advertise a non-zero window
971 * which has been raised "significantly".
973 * Even if window raised up to infinity, do not send window open ACK
974 * in states, where we will not receive more. It is useless.
976 if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) {
977 __u32 rcv_window_now = tcp_receive_window(tp);
979 /* Optimize, __tcp_select_window() is not cheap. */
980 if (2*rcv_window_now <= tp->window_clamp) {
981 __u32 new_window = __tcp_select_window(sk);
983 /* Send ACK now, if this read freed lots of space
984 * in our buffer. Certainly, new_window is new window.
985 * We can advertise it now, if it is not less than current one.
986 * "Lots" means "at least twice" here.
988 if (new_window && new_window >= 2 * rcv_window_now)
996 static void tcp_prequeue_process(struct sock *sk)
999 struct tcp_sock *tp = tcp_sk(sk);
1001 NET_INC_STATS_USER(LINUX_MIB_TCPPREQUEUED);
1003 /* RX process wants to run with disabled BHs, though it is not
1006 while ((skb = __skb_dequeue(&tp->ucopy.prequeue)) != NULL)
1007 sk->sk_backlog_rcv(sk, skb);
1010 /* Clear memory counter. */
1011 tp->ucopy.memory = 0;
1014 static inline struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off)
1016 struct sk_buff *skb;
1019 skb_queue_walk(&sk->sk_receive_queue, skb) {
1020 offset = seq - TCP_SKB_CB(skb)->seq;
1021 if (tcp_hdr(skb)->syn)
1023 if (offset < skb->len || tcp_hdr(skb)->fin) {
1032 * This routine provides an alternative to tcp_recvmsg() for routines
1033 * that would like to handle copying from skbuffs directly in 'sendfile'
1036 * - It is assumed that the socket was locked by the caller.
1037 * - The routine does not block.
1038 * - At present, there is no support for reading OOB data
1039 * or for 'peeking' the socket using this routine
1040 * (although both would be easy to implement).
1042 int tcp_read_sock(struct sock *sk, read_descriptor_t *desc,
1043 sk_read_actor_t recv_actor)
1045 struct sk_buff *skb;
1046 struct tcp_sock *tp = tcp_sk(sk);
1047 u32 seq = tp->copied_seq;
1051 if (sk->sk_state == TCP_LISTEN)
1053 while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1054 if (offset < skb->len) {
1057 len = skb->len - offset;
1058 /* Stop reading if we hit a patch of urgent data */
1060 u32 urg_offset = tp->urg_seq - seq;
1061 if (urg_offset < len)
1066 used = recv_actor(desc, skb, offset, len);
1071 } else if (used <= len) {
1076 if (offset != skb->len)
1079 if (tcp_hdr(skb)->fin) {
1080 sk_eat_skb(sk, skb, 0);
1084 sk_eat_skb(sk, skb, 0);
1088 tp->copied_seq = seq;
1090 tcp_rcv_space_adjust(sk);
1092 /* Clean up data we have read: This will do ACK frames. */
1094 tcp_cleanup_rbuf(sk, copied);
1099 * This routine copies from a sock struct into the user buffer.
1101 * Technical note: in 2.3 we work on _locked_ socket, so that
1102 * tricks with *seq access order and skb->users are not required.
1103 * Probably, code can be easily improved even more.
1106 int tcp_recvmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg,
1107 size_t len, int nonblock, int flags, int *addr_len)
1109 struct tcp_sock *tp = tcp_sk(sk);
1115 int target; /* Read at least this many bytes */
1117 struct task_struct *user_recv = NULL;
1118 int copied_early = 0;
1120 struct sk_buff *skb;
1124 TCP_CHECK_TIMER(sk);
1127 if (sk->sk_state == TCP_LISTEN)
1130 timeo = sock_rcvtimeo(sk, nonblock);
1132 /* Urgent data needs to be handled specially. */
1133 if (flags & MSG_OOB)
1136 seq = &tp->copied_seq;
1137 if (flags & MSG_PEEK) {
1138 peek_seq = tp->copied_seq;
1142 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
1144 #ifdef CONFIG_NET_DMA
1145 tp->ucopy.dma_chan = NULL;
1147 skb = skb_peek_tail(&sk->sk_receive_queue);
1149 available = TCP_SKB_CB(skb)->seq + skb->len - (*seq);
1150 if ((available < target) &&
1151 (len > sysctl_tcp_dma_copybreak) && !(flags & MSG_PEEK) &&
1152 !sysctl_tcp_low_latency && __get_cpu_var(softnet_data).net_dma) {
1153 preempt_enable_no_resched();
1154 tp->ucopy.pinned_list = dma_pin_iovec_pages(msg->msg_iov, len);
1156 preempt_enable_no_resched();
1162 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
1163 if (tp->urg_data && tp->urg_seq == *seq) {
1166 if (signal_pending(current)) {
1167 copied = timeo ? sock_intr_errno(timeo) : -EAGAIN;
1172 /* Next get a buffer. */
1174 skb = skb_peek(&sk->sk_receive_queue);
1179 /* Now that we have two receive queues this
1182 if (before(*seq, TCP_SKB_CB(skb)->seq)) {
1183 printk(KERN_INFO "recvmsg bug: copied %X "
1184 "seq %X\n", *seq, TCP_SKB_CB(skb)->seq);
1187 offset = *seq - TCP_SKB_CB(skb)->seq;
1188 if (tcp_hdr(skb)->syn)
1190 if (offset < skb->len)
1192 if (tcp_hdr(skb)->fin)
1194 BUG_TRAP(flags & MSG_PEEK);
1196 } while (skb != (struct sk_buff *)&sk->sk_receive_queue);
1198 /* Well, if we have backlog, try to process it now yet. */
1200 if (copied >= target && !sk->sk_backlog.tail)
1205 sk->sk_state == TCP_CLOSE ||
1206 (sk->sk_shutdown & RCV_SHUTDOWN) ||
1208 signal_pending(current) ||
1212 if (sock_flag(sk, SOCK_DONE))
1216 copied = sock_error(sk);
1220 if (sk->sk_shutdown & RCV_SHUTDOWN)
1223 if (sk->sk_state == TCP_CLOSE) {
1224 if (!sock_flag(sk, SOCK_DONE)) {
1225 /* This occurs when user tries to read
1226 * from never connected socket.
1239 if (signal_pending(current)) {
1240 copied = sock_intr_errno(timeo);
1245 tcp_cleanup_rbuf(sk, copied);
1247 if (!sysctl_tcp_low_latency && tp->ucopy.task == user_recv) {
1248 /* Install new reader */
1249 if (!user_recv && !(flags & (MSG_TRUNC | MSG_PEEK))) {
1250 user_recv = current;
1251 tp->ucopy.task = user_recv;
1252 tp->ucopy.iov = msg->msg_iov;
1255 tp->ucopy.len = len;
1257 BUG_TRAP(tp->copied_seq == tp->rcv_nxt ||
1258 (flags & (MSG_PEEK | MSG_TRUNC)));
1260 /* Ugly... If prequeue is not empty, we have to
1261 * process it before releasing socket, otherwise
1262 * order will be broken at second iteration.
1263 * More elegant solution is required!!!
1265 * Look: we have the following (pseudo)queues:
1267 * 1. packets in flight
1272 * Each queue can be processed only if the next ones
1273 * are empty. At this point we have empty receive_queue.
1274 * But prequeue _can_ be not empty after 2nd iteration,
1275 * when we jumped to start of loop because backlog
1276 * processing added something to receive_queue.
1277 * We cannot release_sock(), because backlog contains
1278 * packets arrived _after_ prequeued ones.
1280 * Shortly, algorithm is clear --- to process all
1281 * the queues in order. We could make it more directly,
1282 * requeueing packets from backlog to prequeue, if
1283 * is not empty. It is more elegant, but eats cycles,
1286 if (!skb_queue_empty(&tp->ucopy.prequeue))
1289 /* __ Set realtime policy in scheduler __ */
1292 if (copied >= target) {
1293 /* Do not sleep, just process backlog. */
1297 sk_wait_data(sk, &timeo);
1299 #ifdef CONFIG_NET_DMA
1300 tp->ucopy.wakeup = 0;
1306 /* __ Restore normal policy in scheduler __ */
1308 if ((chunk = len - tp->ucopy.len) != 0) {
1309 NET_ADD_STATS_USER(LINUX_MIB_TCPDIRECTCOPYFROMBACKLOG, chunk);
1314 if (tp->rcv_nxt == tp->copied_seq &&
1315 !skb_queue_empty(&tp->ucopy.prequeue)) {
1317 tcp_prequeue_process(sk);
1319 if ((chunk = len - tp->ucopy.len) != 0) {
1320 NET_ADD_STATS_USER(LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1326 if ((flags & MSG_PEEK) && peek_seq != tp->copied_seq) {
1327 if (net_ratelimit())
1328 printk(KERN_DEBUG "TCP(%s:%d): Application bug, race in MSG_PEEK.\n",
1329 current->comm, current->pid);
1330 peek_seq = tp->copied_seq;
1335 /* Ok so how much can we use? */
1336 used = skb->len - offset;
1340 /* Do we have urgent data here? */
1342 u32 urg_offset = tp->urg_seq - *seq;
1343 if (urg_offset < used) {
1345 if (!sock_flag(sk, SOCK_URGINLINE)) {
1357 if (!(flags & MSG_TRUNC)) {
1358 #ifdef CONFIG_NET_DMA
1359 if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list)
1360 tp->ucopy.dma_chan = get_softnet_dma();
1362 if (tp->ucopy.dma_chan) {
1363 tp->ucopy.dma_cookie = dma_skb_copy_datagram_iovec(
1364 tp->ucopy.dma_chan, skb, offset,
1366 tp->ucopy.pinned_list);
1368 if (tp->ucopy.dma_cookie < 0) {
1370 printk(KERN_ALERT "dma_cookie < 0\n");
1372 /* Exception. Bailout! */
1377 if ((offset + used) == skb->len)
1383 err = skb_copy_datagram_iovec(skb, offset,
1384 msg->msg_iov, used);
1386 /* Exception. Bailout! */
1398 tcp_rcv_space_adjust(sk);
1401 if (tp->urg_data && after(tp->copied_seq, tp->urg_seq)) {
1403 tcp_fast_path_check(sk);
1405 if (used + offset < skb->len)
1408 if (tcp_hdr(skb)->fin)
1410 if (!(flags & MSG_PEEK)) {
1411 sk_eat_skb(sk, skb, copied_early);
1417 /* Process the FIN. */
1419 if (!(flags & MSG_PEEK)) {
1420 sk_eat_skb(sk, skb, copied_early);
1427 if (!skb_queue_empty(&tp->ucopy.prequeue)) {
1430 tp->ucopy.len = copied > 0 ? len : 0;
1432 tcp_prequeue_process(sk);
1434 if (copied > 0 && (chunk = len - tp->ucopy.len) != 0) {
1435 NET_ADD_STATS_USER(LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1441 tp->ucopy.task = NULL;
1445 #ifdef CONFIG_NET_DMA
1446 if (tp->ucopy.dma_chan) {
1447 dma_cookie_t done, used;
1449 dma_async_memcpy_issue_pending(tp->ucopy.dma_chan);
1451 while (dma_async_memcpy_complete(tp->ucopy.dma_chan,
1452 tp->ucopy.dma_cookie, &done,
1453 &used) == DMA_IN_PROGRESS) {
1454 /* do partial cleanup of sk_async_wait_queue */
1455 while ((skb = skb_peek(&sk->sk_async_wait_queue)) &&
1456 (dma_async_is_complete(skb->dma_cookie, done,
1457 used) == DMA_SUCCESS)) {
1458 __skb_dequeue(&sk->sk_async_wait_queue);
1463 /* Safe to free early-copied skbs now */
1464 __skb_queue_purge(&sk->sk_async_wait_queue);
1465 dma_chan_put(tp->ucopy.dma_chan);
1466 tp->ucopy.dma_chan = NULL;
1468 if (tp->ucopy.pinned_list) {
1469 dma_unpin_iovec_pages(tp->ucopy.pinned_list);
1470 tp->ucopy.pinned_list = NULL;
1474 /* According to UNIX98, msg_name/msg_namelen are ignored
1475 * on connected socket. I was just happy when found this 8) --ANK
1478 /* Clean up data we have read: This will do ACK frames. */
1479 tcp_cleanup_rbuf(sk, copied);
1481 TCP_CHECK_TIMER(sk);
1486 TCP_CHECK_TIMER(sk);
1491 err = tcp_recv_urg(sk, timeo, msg, len, flags, addr_len);
1496 * State processing on a close. This implements the state shift for
1497 * sending our FIN frame. Note that we only send a FIN for some
1498 * states. A shutdown() may have already sent the FIN, or we may be
1502 static const unsigned char new_state[16] = {
1503 /* current state: new state: action: */
1504 /* (Invalid) */ TCP_CLOSE,
1505 /* TCP_ESTABLISHED */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1506 /* TCP_SYN_SENT */ TCP_CLOSE,
1507 /* TCP_SYN_RECV */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1508 /* TCP_FIN_WAIT1 */ TCP_FIN_WAIT1,
1509 /* TCP_FIN_WAIT2 */ TCP_FIN_WAIT2,
1510 /* TCP_TIME_WAIT */ TCP_CLOSE,
1511 /* TCP_CLOSE */ TCP_CLOSE,
1512 /* TCP_CLOSE_WAIT */ TCP_LAST_ACK | TCP_ACTION_FIN,
1513 /* TCP_LAST_ACK */ TCP_LAST_ACK,
1514 /* TCP_LISTEN */ TCP_CLOSE,
1515 /* TCP_CLOSING */ TCP_CLOSING,
1518 static int tcp_close_state(struct sock *sk)
1520 int next = (int)new_state[sk->sk_state];
1521 int ns = next & TCP_STATE_MASK;
1523 tcp_set_state(sk, ns);
1525 return next & TCP_ACTION_FIN;
1529 * Shutdown the sending side of a connection. Much like close except
1530 * that we don't receive shut down or set_sock_flag(sk, SOCK_DEAD).
1533 void tcp_shutdown(struct sock *sk, int how)
1535 /* We need to grab some memory, and put together a FIN,
1536 * and then put it into the queue to be sent.
1537 * Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
1539 if (!(how & SEND_SHUTDOWN))
1542 /* If we've already sent a FIN, or it's a closed state, skip this. */
1543 if ((1 << sk->sk_state) &
1544 (TCPF_ESTABLISHED | TCPF_SYN_SENT |
1545 TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) {
1546 /* Clear out any half completed packets. FIN if needed. */
1547 if (tcp_close_state(sk))
1552 void tcp_close(struct sock *sk, long timeout)
1554 struct sk_buff *skb;
1555 int data_was_unread = 0;
1559 sk->sk_shutdown = SHUTDOWN_MASK;
1561 if (sk->sk_state == TCP_LISTEN) {
1562 tcp_set_state(sk, TCP_CLOSE);
1565 inet_csk_listen_stop(sk);
1567 goto adjudge_to_death;
1570 /* We need to flush the recv. buffs. We do this only on the
1571 * descriptor close, not protocol-sourced closes, because the
1572 * reader process may not have drained the data yet!
1574 while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) {
1575 u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq -
1577 data_was_unread += len;
1581 sk_stream_mem_reclaim(sk);
1583 /* As outlined in RFC 2525, section 2.17, we send a RST here because
1584 * data was lost. To witness the awful effects of the old behavior of
1585 * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk
1586 * GET in an FTP client, suspend the process, wait for the client to
1587 * advertise a zero window, then kill -9 the FTP client, wheee...
1588 * Note: timeout is always zero in such a case.
1590 if (data_was_unread) {
1591 /* Unread data was tossed, zap the connection. */
1592 NET_INC_STATS_USER(LINUX_MIB_TCPABORTONCLOSE);
1593 tcp_set_state(sk, TCP_CLOSE);
1594 tcp_send_active_reset(sk, GFP_KERNEL);
1595 } else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) {
1596 /* Check zero linger _after_ checking for unread data. */
1597 sk->sk_prot->disconnect(sk, 0);
1598 NET_INC_STATS_USER(LINUX_MIB_TCPABORTONDATA);
1599 } else if (tcp_close_state(sk)) {
1600 /* We FIN if the application ate all the data before
1601 * zapping the connection.
1604 /* RED-PEN. Formally speaking, we have broken TCP state
1605 * machine. State transitions:
1607 * TCP_ESTABLISHED -> TCP_FIN_WAIT1
1608 * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible)
1609 * TCP_CLOSE_WAIT -> TCP_LAST_ACK
1611 * are legal only when FIN has been sent (i.e. in window),
1612 * rather than queued out of window. Purists blame.
1614 * F.e. "RFC state" is ESTABLISHED,
1615 * if Linux state is FIN-WAIT-1, but FIN is still not sent.
1617 * The visible declinations are that sometimes
1618 * we enter time-wait state, when it is not required really
1619 * (harmless), do not send active resets, when they are
1620 * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
1621 * they look as CLOSING or LAST_ACK for Linux)
1622 * Probably, I missed some more holelets.
1628 sk_stream_wait_close(sk, timeout);
1631 state = sk->sk_state;
1634 atomic_inc(sk->sk_prot->orphan_count);
1636 /* It is the last release_sock in its life. It will remove backlog. */
1640 /* Now socket is owned by kernel and we acquire BH lock
1641 to finish close. No need to check for user refs.
1645 BUG_TRAP(!sock_owned_by_user(sk));
1647 /* Have we already been destroyed by a softirq or backlog? */
1648 if (state != TCP_CLOSE && sk->sk_state == TCP_CLOSE)
1651 /* This is a (useful) BSD violating of the RFC. There is a
1652 * problem with TCP as specified in that the other end could
1653 * keep a socket open forever with no application left this end.
1654 * We use a 3 minute timeout (about the same as BSD) then kill
1655 * our end. If they send after that then tough - BUT: long enough
1656 * that we won't make the old 4*rto = almost no time - whoops
1659 * Nope, it was not mistake. It is really desired behaviour
1660 * f.e. on http servers, when such sockets are useless, but
1661 * consume significant resources. Let's do it with special
1662 * linger2 option. --ANK
1665 if (sk->sk_state == TCP_FIN_WAIT2) {
1666 struct tcp_sock *tp = tcp_sk(sk);
1667 if (tp->linger2 < 0) {
1668 tcp_set_state(sk, TCP_CLOSE);
1669 tcp_send_active_reset(sk, GFP_ATOMIC);
1670 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONLINGER);
1672 const int tmo = tcp_fin_time(sk);
1674 if (tmo > TCP_TIMEWAIT_LEN) {
1675 inet_csk_reset_keepalive_timer(sk,
1676 tmo - TCP_TIMEWAIT_LEN);
1678 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
1683 if (sk->sk_state != TCP_CLOSE) {
1684 sk_stream_mem_reclaim(sk);
1685 if (tcp_too_many_orphans(sk,
1686 atomic_read(sk->sk_prot->orphan_count))) {
1687 if (net_ratelimit())
1688 printk(KERN_INFO "TCP: too many of orphaned "
1690 tcp_set_state(sk, TCP_CLOSE);
1691 tcp_send_active_reset(sk, GFP_ATOMIC);
1692 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONMEMORY);
1696 if (sk->sk_state == TCP_CLOSE)
1697 inet_csk_destroy_sock(sk);
1698 /* Otherwise, socket is reprieved until protocol close. */
1706 /* These states need RST on ABORT according to RFC793 */
1708 static inline int tcp_need_reset(int state)
1710 return (1 << state) &
1711 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 |
1712 TCPF_FIN_WAIT2 | TCPF_SYN_RECV);
1715 int tcp_disconnect(struct sock *sk, int flags)
1717 struct inet_sock *inet = inet_sk(sk);
1718 struct inet_connection_sock *icsk = inet_csk(sk);
1719 struct tcp_sock *tp = tcp_sk(sk);
1721 int old_state = sk->sk_state;
1723 if (old_state != TCP_CLOSE)
1724 tcp_set_state(sk, TCP_CLOSE);
1726 /* ABORT function of RFC793 */
1727 if (old_state == TCP_LISTEN) {
1728 inet_csk_listen_stop(sk);
1729 } else if (tcp_need_reset(old_state) ||
1730 (tp->snd_nxt != tp->write_seq &&
1731 (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK))) {
1732 /* The last check adjusts for discrepancy of Linux wrt. RFC
1735 tcp_send_active_reset(sk, gfp_any());
1736 sk->sk_err = ECONNRESET;
1737 } else if (old_state == TCP_SYN_SENT)
1738 sk->sk_err = ECONNRESET;
1740 tcp_clear_xmit_timers(sk);
1741 __skb_queue_purge(&sk->sk_receive_queue);
1742 tcp_write_queue_purge(sk);
1743 __skb_queue_purge(&tp->out_of_order_queue);
1744 #ifdef CONFIG_NET_DMA
1745 __skb_queue_purge(&sk->sk_async_wait_queue);
1750 if (!(sk->sk_userlocks & SOCK_BINDADDR_LOCK))
1751 inet_reset_saddr(sk);
1753 sk->sk_shutdown = 0;
1754 sock_reset_flag(sk, SOCK_DONE);
1756 if ((tp->write_seq += tp->max_window + 2) == 0)
1758 icsk->icsk_backoff = 0;
1760 icsk->icsk_probes_out = 0;
1761 tp->packets_out = 0;
1762 tp->snd_ssthresh = 0x7fffffff;
1763 tp->snd_cwnd_cnt = 0;
1764 tp->bytes_acked = 0;
1765 tcp_set_ca_state(sk, TCP_CA_Open);
1766 tcp_clear_retrans(tp);
1767 inet_csk_delack_init(sk);
1768 tcp_init_send_head(sk);
1769 memset(&tp->rx_opt, 0, sizeof(tp->rx_opt));
1772 BUG_TRAP(!inet->num || icsk->icsk_bind_hash);
1774 sk->sk_error_report(sk);
1779 * Socket option code for TCP.
1781 static int do_tcp_setsockopt(struct sock *sk, int level,
1782 int optname, char __user *optval, int optlen)
1784 struct tcp_sock *tp = tcp_sk(sk);
1785 struct inet_connection_sock *icsk = inet_csk(sk);
1789 /* This is a string value all the others are int's */
1790 if (optname == TCP_CONGESTION) {
1791 char name[TCP_CA_NAME_MAX];
1796 val = strncpy_from_user(name, optval,
1797 min(TCP_CA_NAME_MAX-1, optlen));
1803 err = tcp_set_congestion_control(sk, name);
1808 if (optlen < sizeof(int))
1811 if (get_user(val, (int __user *)optval))
1818 /* Values greater than interface MTU won't take effect. However
1819 * at the point when this call is done we typically don't yet
1820 * know which interface is going to be used */
1821 if (val < 8 || val > MAX_TCP_WINDOW) {
1825 tp->rx_opt.user_mss = val;
1830 /* TCP_NODELAY is weaker than TCP_CORK, so that
1831 * this option on corked socket is remembered, but
1832 * it is not activated until cork is cleared.
1834 * However, when TCP_NODELAY is set we make
1835 * an explicit push, which overrides even TCP_CORK
1836 * for currently queued segments.
1838 tp->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH;
1839 tcp_push_pending_frames(sk);
1841 tp->nonagle &= ~TCP_NAGLE_OFF;
1846 /* When set indicates to always queue non-full frames.
1847 * Later the user clears this option and we transmit
1848 * any pending partial frames in the queue. This is
1849 * meant to be used alongside sendfile() to get properly
1850 * filled frames when the user (for example) must write
1851 * out headers with a write() call first and then use
1852 * sendfile to send out the data parts.
1854 * TCP_CORK can be set together with TCP_NODELAY and it is
1855 * stronger than TCP_NODELAY.
1858 tp->nonagle |= TCP_NAGLE_CORK;
1860 tp->nonagle &= ~TCP_NAGLE_CORK;
1861 if (tp->nonagle&TCP_NAGLE_OFF)
1862 tp->nonagle |= TCP_NAGLE_PUSH;
1863 tcp_push_pending_frames(sk);
1868 if (val < 1 || val > MAX_TCP_KEEPIDLE)
1871 tp->keepalive_time = val * HZ;
1872 if (sock_flag(sk, SOCK_KEEPOPEN) &&
1873 !((1 << sk->sk_state) &
1874 (TCPF_CLOSE | TCPF_LISTEN))) {
1875 __u32 elapsed = tcp_time_stamp - tp->rcv_tstamp;
1876 if (tp->keepalive_time > elapsed)
1877 elapsed = tp->keepalive_time - elapsed;
1880 inet_csk_reset_keepalive_timer(sk, elapsed);
1885 if (val < 1 || val > MAX_TCP_KEEPINTVL)
1888 tp->keepalive_intvl = val * HZ;
1891 if (val < 1 || val > MAX_TCP_KEEPCNT)
1894 tp->keepalive_probes = val;
1897 if (val < 1 || val > MAX_TCP_SYNCNT)
1900 icsk->icsk_syn_retries = val;
1906 else if (val > sysctl_tcp_fin_timeout / HZ)
1909 tp->linger2 = val * HZ;
1912 case TCP_DEFER_ACCEPT:
1913 icsk->icsk_accept_queue.rskq_defer_accept = 0;
1915 /* Translate value in seconds to number of
1917 while (icsk->icsk_accept_queue.rskq_defer_accept < 32 &&
1918 val > ((TCP_TIMEOUT_INIT / HZ) <<
1919 icsk->icsk_accept_queue.rskq_defer_accept))
1920 icsk->icsk_accept_queue.rskq_defer_accept++;
1921 icsk->icsk_accept_queue.rskq_defer_accept++;
1925 case TCP_WINDOW_CLAMP:
1927 if (sk->sk_state != TCP_CLOSE) {
1931 tp->window_clamp = 0;
1933 tp->window_clamp = val < SOCK_MIN_RCVBUF / 2 ?
1934 SOCK_MIN_RCVBUF / 2 : val;
1939 icsk->icsk_ack.pingpong = 1;
1941 icsk->icsk_ack.pingpong = 0;
1942 if ((1 << sk->sk_state) &
1943 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) &&
1944 inet_csk_ack_scheduled(sk)) {
1945 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
1946 tcp_cleanup_rbuf(sk, 1);
1948 icsk->icsk_ack.pingpong = 1;
1953 #ifdef CONFIG_TCP_MD5SIG
1955 /* Read the IP->Key mappings from userspace */
1956 err = tp->af_specific->md5_parse(sk, optval, optlen);
1969 int tcp_setsockopt(struct sock *sk, int level, int optname, char __user *optval,
1972 struct inet_connection_sock *icsk = inet_csk(sk);
1974 if (level != SOL_TCP)
1975 return icsk->icsk_af_ops->setsockopt(sk, level, optname,
1977 return do_tcp_setsockopt(sk, level, optname, optval, optlen);
1980 #ifdef CONFIG_COMPAT
1981 int compat_tcp_setsockopt(struct sock *sk, int level, int optname,
1982 char __user *optval, int optlen)
1984 if (level != SOL_TCP)
1985 return inet_csk_compat_setsockopt(sk, level, optname,
1987 return do_tcp_setsockopt(sk, level, optname, optval, optlen);
1990 EXPORT_SYMBOL(compat_tcp_setsockopt);
1993 /* Return information about state of tcp endpoint in API format. */
1994 void tcp_get_info(struct sock *sk, struct tcp_info *info)
1996 struct tcp_sock *tp = tcp_sk(sk);
1997 const struct inet_connection_sock *icsk = inet_csk(sk);
1998 u32 now = tcp_time_stamp;
2000 memset(info, 0, sizeof(*info));
2002 info->tcpi_state = sk->sk_state;
2003 info->tcpi_ca_state = icsk->icsk_ca_state;
2004 info->tcpi_retransmits = icsk->icsk_retransmits;
2005 info->tcpi_probes = icsk->icsk_probes_out;
2006 info->tcpi_backoff = icsk->icsk_backoff;
2008 if (tp->rx_opt.tstamp_ok)
2009 info->tcpi_options |= TCPI_OPT_TIMESTAMPS;
2010 if (tp->rx_opt.sack_ok)
2011 info->tcpi_options |= TCPI_OPT_SACK;
2012 if (tp->rx_opt.wscale_ok) {
2013 info->tcpi_options |= TCPI_OPT_WSCALE;
2014 info->tcpi_snd_wscale = tp->rx_opt.snd_wscale;
2015 info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale;
2018 if (tp->ecn_flags&TCP_ECN_OK)
2019 info->tcpi_options |= TCPI_OPT_ECN;
2021 info->tcpi_rto = jiffies_to_usecs(icsk->icsk_rto);
2022 info->tcpi_ato = jiffies_to_usecs(icsk->icsk_ack.ato);
2023 info->tcpi_snd_mss = tp->mss_cache;
2024 info->tcpi_rcv_mss = icsk->icsk_ack.rcv_mss;
2026 info->tcpi_unacked = tp->packets_out;
2027 info->tcpi_sacked = tp->sacked_out;
2028 info->tcpi_lost = tp->lost_out;
2029 info->tcpi_retrans = tp->retrans_out;
2030 info->tcpi_fackets = tp->fackets_out;
2032 info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime);
2033 info->tcpi_last_data_recv = jiffies_to_msecs(now - icsk->icsk_ack.lrcvtime);
2034 info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp);
2036 info->tcpi_pmtu = icsk->icsk_pmtu_cookie;
2037 info->tcpi_rcv_ssthresh = tp->rcv_ssthresh;
2038 info->tcpi_rtt = jiffies_to_usecs(tp->srtt)>>3;
2039 info->tcpi_rttvar = jiffies_to_usecs(tp->mdev)>>2;
2040 info->tcpi_snd_ssthresh = tp->snd_ssthresh;
2041 info->tcpi_snd_cwnd = tp->snd_cwnd;
2042 info->tcpi_advmss = tp->advmss;
2043 info->tcpi_reordering = tp->reordering;
2045 info->tcpi_rcv_rtt = jiffies_to_usecs(tp->rcv_rtt_est.rtt)>>3;
2046 info->tcpi_rcv_space = tp->rcvq_space.space;
2048 info->tcpi_total_retrans = tp->total_retrans;
2051 EXPORT_SYMBOL_GPL(tcp_get_info);
2053 static int do_tcp_getsockopt(struct sock *sk, int level,
2054 int optname, char __user *optval, int __user *optlen)
2056 struct inet_connection_sock *icsk = inet_csk(sk);
2057 struct tcp_sock *tp = tcp_sk(sk);
2060 if (get_user(len, optlen))
2063 len = min_t(unsigned int, len, sizeof(int));
2070 val = tp->mss_cache;
2071 if (!val && ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
2072 val = tp->rx_opt.user_mss;
2075 val = !!(tp->nonagle&TCP_NAGLE_OFF);
2078 val = !!(tp->nonagle&TCP_NAGLE_CORK);
2081 val = (tp->keepalive_time ? : sysctl_tcp_keepalive_time) / HZ;
2084 val = (tp->keepalive_intvl ? : sysctl_tcp_keepalive_intvl) / HZ;
2087 val = tp->keepalive_probes ? : sysctl_tcp_keepalive_probes;
2090 val = icsk->icsk_syn_retries ? : sysctl_tcp_syn_retries;
2095 val = (val ? : sysctl_tcp_fin_timeout) / HZ;
2097 case TCP_DEFER_ACCEPT:
2098 val = !icsk->icsk_accept_queue.rskq_defer_accept ? 0 :
2099 ((TCP_TIMEOUT_INIT / HZ) << (icsk->icsk_accept_queue.rskq_defer_accept - 1));
2101 case TCP_WINDOW_CLAMP:
2102 val = tp->window_clamp;
2105 struct tcp_info info;
2107 if (get_user(len, optlen))
2110 tcp_get_info(sk, &info);
2112 len = min_t(unsigned int, len, sizeof(info));
2113 if (put_user(len, optlen))
2115 if (copy_to_user(optval, &info, len))
2120 val = !icsk->icsk_ack.pingpong;
2123 case TCP_CONGESTION:
2124 if (get_user(len, optlen))
2126 len = min_t(unsigned int, len, TCP_CA_NAME_MAX);
2127 if (put_user(len, optlen))
2129 if (copy_to_user(optval, icsk->icsk_ca_ops->name, len))
2133 return -ENOPROTOOPT;
2136 if (put_user(len, optlen))
2138 if (copy_to_user(optval, &val, len))
2143 int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval,
2146 struct inet_connection_sock *icsk = inet_csk(sk);
2148 if (level != SOL_TCP)
2149 return icsk->icsk_af_ops->getsockopt(sk, level, optname,
2151 return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2154 #ifdef CONFIG_COMPAT
2155 int compat_tcp_getsockopt(struct sock *sk, int level, int optname,
2156 char __user *optval, int __user *optlen)
2158 if (level != SOL_TCP)
2159 return inet_csk_compat_getsockopt(sk, level, optname,
2161 return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2164 EXPORT_SYMBOL(compat_tcp_getsockopt);
2167 struct sk_buff *tcp_tso_segment(struct sk_buff *skb, int features)
2169 struct sk_buff *segs = ERR_PTR(-EINVAL);
2174 unsigned int oldlen;
2177 if (!pskb_may_pull(skb, sizeof(*th)))
2181 thlen = th->doff * 4;
2182 if (thlen < sizeof(*th))
2185 if (!pskb_may_pull(skb, thlen))
2188 oldlen = (u16)~skb->len;
2189 __skb_pull(skb, thlen);
2191 if (skb_gso_ok(skb, features | NETIF_F_GSO_ROBUST)) {
2192 /* Packet is from an untrusted source, reset gso_segs. */
2193 int type = skb_shinfo(skb)->gso_type;
2202 !(type & (SKB_GSO_TCPV4 | SKB_GSO_TCPV6))))
2205 mss = skb_shinfo(skb)->gso_size;
2206 skb_shinfo(skb)->gso_segs = (skb->len + mss - 1) / mss;
2212 segs = skb_segment(skb, features);
2216 len = skb_shinfo(skb)->gso_size;
2217 delta = htonl(oldlen + (thlen + len));
2221 seq = ntohl(th->seq);
2224 th->fin = th->psh = 0;
2226 th->check = ~csum_fold((__force __wsum)((__force u32)th->check +
2227 (__force u32)delta));
2228 if (skb->ip_summed != CHECKSUM_PARTIAL)
2230 csum_fold(csum_partial(skb_transport_header(skb),
2237 th->seq = htonl(seq);
2239 } while (skb->next);
2241 delta = htonl(oldlen + (skb->tail - skb->transport_header) +
2243 th->check = ~csum_fold((__force __wsum)((__force u32)th->check +
2244 (__force u32)delta));
2245 if (skb->ip_summed != CHECKSUM_PARTIAL)
2246 th->check = csum_fold(csum_partial(skb_transport_header(skb),
2252 EXPORT_SYMBOL(tcp_tso_segment);
2254 #ifdef CONFIG_TCP_MD5SIG
2255 static unsigned long tcp_md5sig_users;
2256 static struct tcp_md5sig_pool **tcp_md5sig_pool;
2257 static DEFINE_SPINLOCK(tcp_md5sig_pool_lock);
2259 static void __tcp_free_md5sig_pool(struct tcp_md5sig_pool **pool)
2262 for_each_possible_cpu(cpu) {
2263 struct tcp_md5sig_pool *p = *per_cpu_ptr(pool, cpu);
2265 if (p->md5_desc.tfm)
2266 crypto_free_hash(p->md5_desc.tfm);
2274 void tcp_free_md5sig_pool(void)
2276 struct tcp_md5sig_pool **pool = NULL;
2278 spin_lock_bh(&tcp_md5sig_pool_lock);
2279 if (--tcp_md5sig_users == 0) {
2280 pool = tcp_md5sig_pool;
2281 tcp_md5sig_pool = NULL;
2283 spin_unlock_bh(&tcp_md5sig_pool_lock);
2285 __tcp_free_md5sig_pool(pool);
2288 EXPORT_SYMBOL(tcp_free_md5sig_pool);
2290 static struct tcp_md5sig_pool **__tcp_alloc_md5sig_pool(void)
2293 struct tcp_md5sig_pool **pool;
2295 pool = alloc_percpu(struct tcp_md5sig_pool *);
2299 for_each_possible_cpu(cpu) {
2300 struct tcp_md5sig_pool *p;
2301 struct crypto_hash *hash;
2303 p = kzalloc(sizeof(*p), GFP_KERNEL);
2306 *per_cpu_ptr(pool, cpu) = p;
2308 hash = crypto_alloc_hash("md5", 0, CRYPTO_ALG_ASYNC);
2309 if (!hash || IS_ERR(hash))
2312 p->md5_desc.tfm = hash;
2316 __tcp_free_md5sig_pool(pool);
2320 struct tcp_md5sig_pool **tcp_alloc_md5sig_pool(void)
2322 struct tcp_md5sig_pool **pool;
2326 spin_lock_bh(&tcp_md5sig_pool_lock);
2327 pool = tcp_md5sig_pool;
2328 if (tcp_md5sig_users++ == 0) {
2330 spin_unlock_bh(&tcp_md5sig_pool_lock);
2333 spin_unlock_bh(&tcp_md5sig_pool_lock);
2337 spin_unlock_bh(&tcp_md5sig_pool_lock);
2340 /* we cannot hold spinlock here because this may sleep. */
2341 struct tcp_md5sig_pool **p = __tcp_alloc_md5sig_pool();
2342 spin_lock_bh(&tcp_md5sig_pool_lock);
2345 spin_unlock_bh(&tcp_md5sig_pool_lock);
2348 pool = tcp_md5sig_pool;
2350 /* oops, it has already been assigned. */
2351 spin_unlock_bh(&tcp_md5sig_pool_lock);
2352 __tcp_free_md5sig_pool(p);
2354 tcp_md5sig_pool = pool = p;
2355 spin_unlock_bh(&tcp_md5sig_pool_lock);
2361 EXPORT_SYMBOL(tcp_alloc_md5sig_pool);
2363 struct tcp_md5sig_pool *__tcp_get_md5sig_pool(int cpu)
2365 struct tcp_md5sig_pool **p;
2366 spin_lock_bh(&tcp_md5sig_pool_lock);
2367 p = tcp_md5sig_pool;
2370 spin_unlock_bh(&tcp_md5sig_pool_lock);
2371 return (p ? *per_cpu_ptr(p, cpu) : NULL);
2374 EXPORT_SYMBOL(__tcp_get_md5sig_pool);
2376 void __tcp_put_md5sig_pool(void)
2378 tcp_free_md5sig_pool();
2381 EXPORT_SYMBOL(__tcp_put_md5sig_pool);
2384 void tcp_done(struct sock *sk)
2386 if(sk->sk_state == TCP_SYN_SENT || sk->sk_state == TCP_SYN_RECV)
2387 TCP_INC_STATS_BH(TCP_MIB_ATTEMPTFAILS);
2389 tcp_set_state(sk, TCP_CLOSE);
2390 tcp_clear_xmit_timers(sk);
2392 sk->sk_shutdown = SHUTDOWN_MASK;
2394 if (!sock_flag(sk, SOCK_DEAD))
2395 sk->sk_state_change(sk);
2397 inet_csk_destroy_sock(sk);
2399 EXPORT_SYMBOL_GPL(tcp_done);
2401 extern void __skb_cb_too_small_for_tcp(int, int);
2402 extern struct tcp_congestion_ops tcp_reno;
2404 static __initdata unsigned long thash_entries;
2405 static int __init set_thash_entries(char *str)
2409 thash_entries = simple_strtoul(str, &str, 0);
2412 __setup("thash_entries=", set_thash_entries);
2414 void __init tcp_init(void)
2416 struct sk_buff *skb = NULL;
2417 unsigned long limit;
2418 int order, i, max_share;
2420 if (sizeof(struct tcp_skb_cb) > sizeof(skb->cb))
2421 __skb_cb_too_small_for_tcp(sizeof(struct tcp_skb_cb),
2424 tcp_hashinfo.bind_bucket_cachep =
2425 kmem_cache_create("tcp_bind_bucket",
2426 sizeof(struct inet_bind_bucket), 0,
2427 SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL, NULL);
2429 /* Size and allocate the main established and bind bucket
2432 * The methodology is similar to that of the buffer cache.
2434 tcp_hashinfo.ehash =
2435 alloc_large_system_hash("TCP established",
2436 sizeof(struct inet_ehash_bucket),
2438 (num_physpages >= 128 * 1024) ?
2441 &tcp_hashinfo.ehash_size,
2444 tcp_hashinfo.ehash_size = 1 << tcp_hashinfo.ehash_size;
2445 for (i = 0; i < tcp_hashinfo.ehash_size; i++) {
2446 rwlock_init(&tcp_hashinfo.ehash[i].lock);
2447 INIT_HLIST_HEAD(&tcp_hashinfo.ehash[i].chain);
2448 INIT_HLIST_HEAD(&tcp_hashinfo.ehash[i].twchain);
2451 tcp_hashinfo.bhash =
2452 alloc_large_system_hash("TCP bind",
2453 sizeof(struct inet_bind_hashbucket),
2454 tcp_hashinfo.ehash_size,
2455 (num_physpages >= 128 * 1024) ?
2458 &tcp_hashinfo.bhash_size,
2461 tcp_hashinfo.bhash_size = 1 << tcp_hashinfo.bhash_size;
2462 for (i = 0; i < tcp_hashinfo.bhash_size; i++) {
2463 spin_lock_init(&tcp_hashinfo.bhash[i].lock);
2464 INIT_HLIST_HEAD(&tcp_hashinfo.bhash[i].chain);
2467 /* Try to be a bit smarter and adjust defaults depending
2468 * on available memory.
2470 for (order = 0; ((1 << order) << PAGE_SHIFT) <
2471 (tcp_hashinfo.bhash_size * sizeof(struct inet_bind_hashbucket));
2475 tcp_death_row.sysctl_max_tw_buckets = 180000;
2476 sysctl_tcp_max_orphans = 4096 << (order - 4);
2477 sysctl_max_syn_backlog = 1024;
2478 } else if (order < 3) {
2479 tcp_death_row.sysctl_max_tw_buckets >>= (3 - order);
2480 sysctl_tcp_max_orphans >>= (3 - order);
2481 sysctl_max_syn_backlog = 128;
2484 /* Set the pressure threshold to be a fraction of global memory that
2485 * is up to 1/2 at 256 MB, decreasing toward zero with the amount of
2486 * memory, with a floor of 128 pages.
2488 limit = min(nr_all_pages, 1UL<<(28-PAGE_SHIFT)) >> (20-PAGE_SHIFT);
2489 limit = (limit * (nr_all_pages >> (20-PAGE_SHIFT))) >> (PAGE_SHIFT-11);
2490 limit = max(limit, 128UL);
2491 sysctl_tcp_mem[0] = limit / 4 * 3;
2492 sysctl_tcp_mem[1] = limit;
2493 sysctl_tcp_mem[2] = sysctl_tcp_mem[0] * 2;
2495 /* Set per-socket limits to no more than 1/128 the pressure threshold */
2496 limit = ((unsigned long)sysctl_tcp_mem[1]) << (PAGE_SHIFT - 7);
2497 max_share = min(4UL*1024*1024, limit);
2499 sysctl_tcp_wmem[0] = SK_STREAM_MEM_QUANTUM;
2500 sysctl_tcp_wmem[1] = 16*1024;
2501 sysctl_tcp_wmem[2] = max(64*1024, max_share);
2503 sysctl_tcp_rmem[0] = SK_STREAM_MEM_QUANTUM;
2504 sysctl_tcp_rmem[1] = 87380;
2505 sysctl_tcp_rmem[2] = max(87380, max_share);
2507 printk(KERN_INFO "TCP: Hash tables configured "
2508 "(established %d bind %d)\n",
2509 tcp_hashinfo.ehash_size, tcp_hashinfo.bhash_size);
2511 tcp_register_congestion_control(&tcp_reno);
2514 EXPORT_SYMBOL(tcp_close);
2515 EXPORT_SYMBOL(tcp_disconnect);
2516 EXPORT_SYMBOL(tcp_getsockopt);
2517 EXPORT_SYMBOL(tcp_ioctl);
2518 EXPORT_SYMBOL(tcp_poll);
2519 EXPORT_SYMBOL(tcp_read_sock);
2520 EXPORT_SYMBOL(tcp_recvmsg);
2521 EXPORT_SYMBOL(tcp_sendmsg);
2522 EXPORT_SYMBOL(tcp_sendpage);
2523 EXPORT_SYMBOL(tcp_setsockopt);
2524 EXPORT_SYMBOL(tcp_shutdown);
2525 EXPORT_SYMBOL(tcp_statistics);