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/smp_lock.h>
256 #include <linux/fs.h>
257 #include <linux/random.h>
258 #include <linux/bootmem.h>
259 #include <linux/cache.h>
260 #include <linux/err.h>
261 #include <linux/crypto.h>
263 #include <net/icmp.h>
265 #include <net/xfrm.h>
267 #include <net/netdma.h>
269 #include <asm/uaccess.h>
270 #include <asm/ioctls.h>
272 int sysctl_tcp_fin_timeout __read_mostly = TCP_FIN_TIMEOUT;
274 DEFINE_SNMP_STAT(struct tcp_mib, tcp_statistics) __read_mostly;
276 atomic_t tcp_orphan_count = ATOMIC_INIT(0);
278 EXPORT_SYMBOL_GPL(tcp_orphan_count);
280 int sysctl_tcp_mem[3] __read_mostly;
281 int sysctl_tcp_wmem[3] __read_mostly;
282 int sysctl_tcp_rmem[3] __read_mostly;
284 EXPORT_SYMBOL(sysctl_tcp_mem);
285 EXPORT_SYMBOL(sysctl_tcp_rmem);
286 EXPORT_SYMBOL(sysctl_tcp_wmem);
288 atomic_t tcp_memory_allocated; /* Current allocated memory. */
289 atomic_t tcp_sockets_allocated; /* Current number of TCP sockets. */
291 EXPORT_SYMBOL(tcp_memory_allocated);
292 EXPORT_SYMBOL(tcp_sockets_allocated);
295 * Pressure flag: try to collapse.
296 * Technical note: it is used by multiple contexts non atomically.
297 * All the sk_stream_mem_schedule() is of this nature: accounting
298 * is strict, actions are advisory and have some latency.
300 int tcp_memory_pressure __read_mostly;
302 EXPORT_SYMBOL(tcp_memory_pressure);
304 void tcp_enter_memory_pressure(void)
306 if (!tcp_memory_pressure) {
307 NET_INC_STATS(LINUX_MIB_TCPMEMORYPRESSURES);
308 tcp_memory_pressure = 1;
312 EXPORT_SYMBOL(tcp_enter_memory_pressure);
315 * Wait for a TCP event.
317 * Note that we don't need to lock the socket, as the upper poll layers
318 * take care of normal races (between the test and the event) and we don't
319 * go look at any of the socket buffers directly.
321 unsigned int tcp_poll(struct file *file, struct socket *sock, poll_table *wait)
324 struct sock *sk = sock->sk;
325 struct tcp_sock *tp = tcp_sk(sk);
327 poll_wait(file, sk->sk_sleep, wait);
328 if (sk->sk_state == TCP_LISTEN)
329 return inet_csk_listen_poll(sk);
331 /* Socket is not locked. We are protected from async events
332 by poll logic and correct handling of state changes
333 made by another threads is impossible in any case.
341 * POLLHUP is certainly not done right. But poll() doesn't
342 * have a notion of HUP in just one direction, and for a
343 * socket the read side is more interesting.
345 * Some poll() documentation says that POLLHUP is incompatible
346 * with the POLLOUT/POLLWR flags, so somebody should check this
347 * all. But careful, it tends to be safer to return too many
348 * bits than too few, and you can easily break real applications
349 * if you don't tell them that something has hung up!
353 * Check number 1. POLLHUP is _UNMASKABLE_ event (see UNIX98 and
354 * our fs/select.c). It means that after we received EOF,
355 * poll always returns immediately, making impossible poll() on write()
356 * in state CLOSE_WAIT. One solution is evident --- to set POLLHUP
357 * if and only if shutdown has been made in both directions.
358 * Actually, it is interesting to look how Solaris and DUX
359 * solve this dilemma. I would prefer, if PULLHUP were maskable,
360 * then we could set it on SND_SHUTDOWN. BTW examples given
361 * in Stevens' books assume exactly this behaviour, it explains
362 * why PULLHUP is incompatible with POLLOUT. --ANK
364 * NOTE. Check for TCP_CLOSE is added. The goal is to prevent
365 * blocking on fresh not-connected or disconnected socket. --ANK
367 if (sk->sk_shutdown == SHUTDOWN_MASK || sk->sk_state == TCP_CLOSE)
369 if (sk->sk_shutdown & RCV_SHUTDOWN)
370 mask |= POLLIN | POLLRDNORM | POLLRDHUP;
373 if ((1 << sk->sk_state) & ~(TCPF_SYN_SENT | TCPF_SYN_RECV)) {
374 /* Potential race condition. If read of tp below will
375 * escape above sk->sk_state, we can be illegally awaken
376 * in SYN_* states. */
377 if ((tp->rcv_nxt != tp->copied_seq) &&
378 (tp->urg_seq != tp->copied_seq ||
379 tp->rcv_nxt != tp->copied_seq + 1 ||
380 sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data))
381 mask |= POLLIN | POLLRDNORM;
383 if (!(sk->sk_shutdown & SEND_SHUTDOWN)) {
384 if (sk_stream_wspace(sk) >= sk_stream_min_wspace(sk)) {
385 mask |= POLLOUT | POLLWRNORM;
386 } else { /* send SIGIO later */
387 set_bit(SOCK_ASYNC_NOSPACE,
388 &sk->sk_socket->flags);
389 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
391 /* Race breaker. If space is freed after
392 * wspace test but before the flags are set,
393 * IO signal will be lost.
395 if (sk_stream_wspace(sk) >= sk_stream_min_wspace(sk))
396 mask |= POLLOUT | POLLWRNORM;
400 if (tp->urg_data & TCP_URG_VALID)
406 int tcp_ioctl(struct sock *sk, int cmd, unsigned long arg)
408 struct tcp_sock *tp = tcp_sk(sk);
413 if (sk->sk_state == TCP_LISTEN)
417 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
419 else if (sock_flag(sk, SOCK_URGINLINE) ||
421 before(tp->urg_seq, tp->copied_seq) ||
422 !before(tp->urg_seq, tp->rcv_nxt)) {
423 answ = tp->rcv_nxt - tp->copied_seq;
425 /* Subtract 1, if FIN is in queue. */
426 if (answ && !skb_queue_empty(&sk->sk_receive_queue))
428 tcp_hdr((struct sk_buff *)sk->sk_receive_queue.prev)->fin;
430 answ = tp->urg_seq - tp->copied_seq;
434 answ = tp->urg_data && tp->urg_seq == tp->copied_seq;
437 if (sk->sk_state == TCP_LISTEN)
440 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
443 answ = tp->write_seq - tp->snd_una;
449 return put_user(answ, (int __user *)arg);
452 static inline void tcp_mark_push(struct tcp_sock *tp, struct sk_buff *skb)
454 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH;
455 tp->pushed_seq = tp->write_seq;
458 static inline int forced_push(struct tcp_sock *tp)
460 return after(tp->write_seq, tp->pushed_seq + (tp->max_window >> 1));
463 static inline void skb_entail(struct sock *sk, struct sk_buff *skb)
465 struct tcp_sock *tp = tcp_sk(sk);
466 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
469 tcb->seq = tcb->end_seq = tp->write_seq;
470 tcb->flags = TCPCB_FLAG_ACK;
472 skb_header_release(skb);
473 tcp_add_write_queue_tail(sk, skb);
474 sk_charge_skb(sk, skb);
475 if (tp->nonagle & TCP_NAGLE_PUSH)
476 tp->nonagle &= ~TCP_NAGLE_PUSH;
479 static inline void tcp_mark_urg(struct tcp_sock *tp, int flags,
482 if (flags & MSG_OOB) {
484 tp->snd_up = tp->write_seq;
485 TCP_SKB_CB(skb)->sacked |= TCPCB_URG;
489 static inline void tcp_push(struct sock *sk, int flags, int mss_now,
492 struct tcp_sock *tp = tcp_sk(sk);
494 if (tcp_send_head(sk)) {
495 struct sk_buff *skb = tcp_write_queue_tail(sk);
496 if (!(flags & MSG_MORE) || forced_push(tp))
497 tcp_mark_push(tp, skb);
498 tcp_mark_urg(tp, flags, skb);
499 __tcp_push_pending_frames(sk, mss_now,
500 (flags & MSG_MORE) ? TCP_NAGLE_CORK : nonagle);
504 static ssize_t do_tcp_sendpages(struct sock *sk, struct page **pages, int poffset,
505 size_t psize, int flags)
507 struct tcp_sock *tp = tcp_sk(sk);
508 int mss_now, size_goal;
511 long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
513 /* Wait for a connection to finish. */
514 if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
515 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
518 clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
520 mss_now = tcp_current_mss(sk, !(flags&MSG_OOB));
521 size_goal = tp->xmit_size_goal;
525 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
529 struct sk_buff *skb = tcp_write_queue_tail(sk);
530 struct page *page = pages[poffset / PAGE_SIZE];
531 int copy, i, can_coalesce;
532 int offset = poffset % PAGE_SIZE;
533 int size = min_t(size_t, psize, PAGE_SIZE - offset);
535 if (!tcp_send_head(sk) || (copy = size_goal - skb->len) <= 0) {
537 if (!sk_stream_memory_free(sk))
538 goto wait_for_sndbuf;
540 skb = sk_stream_alloc_pskb(sk, 0, 0,
543 goto wait_for_memory;
552 i = skb_shinfo(skb)->nr_frags;
553 can_coalesce = skb_can_coalesce(skb, i, page, offset);
554 if (!can_coalesce && i >= MAX_SKB_FRAGS) {
555 tcp_mark_push(tp, skb);
558 if (!sk_stream_wmem_schedule(sk, copy))
559 goto wait_for_memory;
562 skb_shinfo(skb)->frags[i - 1].size += copy;
565 skb_fill_page_desc(skb, i, page, offset, copy);
569 skb->data_len += copy;
570 skb->truesize += copy;
571 sk->sk_wmem_queued += copy;
572 sk->sk_forward_alloc -= copy;
573 skb->ip_summed = CHECKSUM_PARTIAL;
574 tp->write_seq += copy;
575 TCP_SKB_CB(skb)->end_seq += copy;
576 skb_shinfo(skb)->gso_segs = 0;
579 TCP_SKB_CB(skb)->flags &= ~TCPCB_FLAG_PSH;
583 if (!(psize -= copy))
586 if (skb->len < mss_now || (flags & MSG_OOB))
589 if (forced_push(tp)) {
590 tcp_mark_push(tp, skb);
591 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
592 } else if (skb == tcp_send_head(sk))
593 tcp_push_one(sk, mss_now);
597 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
600 tcp_push(sk, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH);
602 if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
605 mss_now = tcp_current_mss(sk, !(flags&MSG_OOB));
606 size_goal = tp->xmit_size_goal;
611 tcp_push(sk, flags, mss_now, tp->nonagle);
618 return sk_stream_error(sk, flags, err);
621 ssize_t tcp_sendpage(struct socket *sock, struct page *page, int offset,
622 size_t size, int flags)
625 struct sock *sk = sock->sk;
627 if (!(sk->sk_route_caps & NETIF_F_SG) ||
628 !(sk->sk_route_caps & NETIF_F_ALL_CSUM))
629 return sock_no_sendpage(sock, page, offset, size, flags);
633 res = do_tcp_sendpages(sk, &page, offset, size, flags);
639 #define TCP_PAGE(sk) (sk->sk_sndmsg_page)
640 #define TCP_OFF(sk) (sk->sk_sndmsg_off)
642 static inline int select_size(struct sock *sk)
644 struct tcp_sock *tp = tcp_sk(sk);
645 int tmp = tp->mss_cache;
647 if (sk->sk_route_caps & NETIF_F_SG) {
651 int pgbreak = SKB_MAX_HEAD(MAX_TCP_HEADER);
653 if (tmp >= pgbreak &&
654 tmp <= pgbreak + (MAX_SKB_FRAGS - 1) * PAGE_SIZE)
662 int tcp_sendmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg,
666 struct tcp_sock *tp = tcp_sk(sk);
669 int mss_now, size_goal;
676 flags = msg->msg_flags;
677 timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
679 /* Wait for a connection to finish. */
680 if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
681 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
684 /* This should be in poll */
685 clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
687 mss_now = tcp_current_mss(sk, !(flags&MSG_OOB));
688 size_goal = tp->xmit_size_goal;
690 /* Ok commence sending. */
691 iovlen = msg->msg_iovlen;
696 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
699 while (--iovlen >= 0) {
700 int seglen = iov->iov_len;
701 unsigned char __user *from = iov->iov_base;
708 skb = tcp_write_queue_tail(sk);
710 if (!tcp_send_head(sk) ||
711 (copy = size_goal - skb->len) <= 0) {
714 /* Allocate new segment. If the interface is SG,
715 * allocate skb fitting to single page.
717 if (!sk_stream_memory_free(sk))
718 goto wait_for_sndbuf;
720 skb = sk_stream_alloc_pskb(sk, select_size(sk),
721 0, sk->sk_allocation);
723 goto wait_for_memory;
726 * Check whether we can use HW checksum.
728 if (sk->sk_route_caps & NETIF_F_ALL_CSUM)
729 skb->ip_summed = CHECKSUM_PARTIAL;
735 /* Try to append data to the end of skb. */
739 /* Where to copy to? */
740 if (skb_tailroom(skb) > 0) {
741 /* We have some space in skb head. Superb! */
742 if (copy > skb_tailroom(skb))
743 copy = skb_tailroom(skb);
744 if ((err = skb_add_data(skb, from, copy)) != 0)
748 int i = skb_shinfo(skb)->nr_frags;
749 struct page *page = TCP_PAGE(sk);
750 int off = TCP_OFF(sk);
752 if (skb_can_coalesce(skb, i, page, off) &&
754 /* We can extend the last page
757 } else if (i == MAX_SKB_FRAGS ||
759 !(sk->sk_route_caps & NETIF_F_SG))) {
760 /* Need to add new fragment and cannot
761 * do this because interface is non-SG,
762 * or because all the page slots are
764 tcp_mark_push(tp, skb);
767 if (off == PAGE_SIZE) {
769 TCP_PAGE(sk) = page = NULL;
775 if (copy > PAGE_SIZE - off)
776 copy = PAGE_SIZE - off;
778 if (!sk_stream_wmem_schedule(sk, copy))
779 goto wait_for_memory;
782 /* Allocate new cache page. */
783 if (!(page = sk_stream_alloc_page(sk)))
784 goto wait_for_memory;
787 /* Time to copy data. We are close to
789 err = skb_copy_to_page(sk, from, skb, page,
792 /* If this page was new, give it to the
793 * socket so it does not get leaked.
802 /* Update the skb. */
804 skb_shinfo(skb)->frags[i - 1].size +=
807 skb_fill_page_desc(skb, i, page, off, copy);
810 } else if (off + copy < PAGE_SIZE) {
816 TCP_OFF(sk) = off + copy;
820 TCP_SKB_CB(skb)->flags &= ~TCPCB_FLAG_PSH;
822 tp->write_seq += copy;
823 TCP_SKB_CB(skb)->end_seq += copy;
824 skb_shinfo(skb)->gso_segs = 0;
828 if ((seglen -= copy) == 0 && iovlen == 0)
831 if (skb->len < mss_now || (flags & MSG_OOB))
834 if (forced_push(tp)) {
835 tcp_mark_push(tp, skb);
836 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
837 } else if (skb == tcp_send_head(sk))
838 tcp_push_one(sk, mss_now);
842 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
845 tcp_push(sk, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH);
847 if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
850 mss_now = tcp_current_mss(sk, !(flags&MSG_OOB));
851 size_goal = tp->xmit_size_goal;
857 tcp_push(sk, flags, mss_now, tp->nonagle);
864 tcp_unlink_write_queue(skb, sk);
865 /* It is the one place in all of TCP, except connection
866 * reset, where we can be unlinking the send_head.
868 tcp_check_send_head(sk, skb);
869 sk_stream_free_skb(sk, skb);
876 err = sk_stream_error(sk, flags, err);
883 * Handle reading urgent data. BSD has very simple semantics for
884 * this, no blocking and very strange errors 8)
887 static int tcp_recv_urg(struct sock *sk, long timeo,
888 struct msghdr *msg, int len, int flags,
891 struct tcp_sock *tp = tcp_sk(sk);
893 /* No URG data to read. */
894 if (sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data ||
895 tp->urg_data == TCP_URG_READ)
896 return -EINVAL; /* Yes this is right ! */
898 if (sk->sk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DONE))
901 if (tp->urg_data & TCP_URG_VALID) {
903 char c = tp->urg_data;
905 if (!(flags & MSG_PEEK))
906 tp->urg_data = TCP_URG_READ;
908 /* Read urgent data. */
909 msg->msg_flags |= MSG_OOB;
912 if (!(flags & MSG_TRUNC))
913 err = memcpy_toiovec(msg->msg_iov, &c, 1);
916 msg->msg_flags |= MSG_TRUNC;
918 return err ? -EFAULT : len;
921 if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
924 /* Fixed the recv(..., MSG_OOB) behaviour. BSD docs and
925 * the available implementations agree in this case:
926 * this call should never block, independent of the
927 * blocking state of the socket.
928 * Mike <pall@rz.uni-karlsruhe.de>
933 /* Clean up the receive buffer for full frames taken by the user,
934 * then send an ACK if necessary. COPIED is the number of bytes
935 * tcp_recvmsg has given to the user so far, it speeds up the
936 * calculation of whether or not we must ACK for the sake of
939 void tcp_cleanup_rbuf(struct sock *sk, int copied)
941 struct tcp_sock *tp = tcp_sk(sk);
945 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
947 BUG_TRAP(!skb || before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq));
950 if (inet_csk_ack_scheduled(sk)) {
951 const struct inet_connection_sock *icsk = inet_csk(sk);
952 /* Delayed ACKs frequently hit locked sockets during bulk
954 if (icsk->icsk_ack.blocked ||
955 /* Once-per-two-segments ACK was not sent by tcp_input.c */
956 tp->rcv_nxt - tp->rcv_wup > icsk->icsk_ack.rcv_mss ||
958 * If this read emptied read buffer, we send ACK, if
959 * connection is not bidirectional, user drained
960 * receive buffer and there was a small segment
964 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED2) ||
965 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED) &&
966 !icsk->icsk_ack.pingpong)) &&
967 !atomic_read(&sk->sk_rmem_alloc)))
971 /* We send an ACK if we can now advertise a non-zero window
972 * which has been raised "significantly".
974 * Even if window raised up to infinity, do not send window open ACK
975 * in states, where we will not receive more. It is useless.
977 if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) {
978 __u32 rcv_window_now = tcp_receive_window(tp);
980 /* Optimize, __tcp_select_window() is not cheap. */
981 if (2*rcv_window_now <= tp->window_clamp) {
982 __u32 new_window = __tcp_select_window(sk);
984 /* Send ACK now, if this read freed lots of space
985 * in our buffer. Certainly, new_window is new window.
986 * We can advertise it now, if it is not less than current one.
987 * "Lots" means "at least twice" here.
989 if (new_window && new_window >= 2 * rcv_window_now)
997 static void tcp_prequeue_process(struct sock *sk)
1000 struct tcp_sock *tp = tcp_sk(sk);
1002 NET_INC_STATS_USER(LINUX_MIB_TCPPREQUEUED);
1004 /* RX process wants to run with disabled BHs, though it is not
1007 while ((skb = __skb_dequeue(&tp->ucopy.prequeue)) != NULL)
1008 sk->sk_backlog_rcv(sk, skb);
1011 /* Clear memory counter. */
1012 tp->ucopy.memory = 0;
1015 static inline struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off)
1017 struct sk_buff *skb;
1020 skb_queue_walk(&sk->sk_receive_queue, skb) {
1021 offset = seq - TCP_SKB_CB(skb)->seq;
1022 if (tcp_hdr(skb)->syn)
1024 if (offset < skb->len || tcp_hdr(skb)->fin) {
1033 * This routine provides an alternative to tcp_recvmsg() for routines
1034 * that would like to handle copying from skbuffs directly in 'sendfile'
1037 * - It is assumed that the socket was locked by the caller.
1038 * - The routine does not block.
1039 * - At present, there is no support for reading OOB data
1040 * or for 'peeking' the socket using this routine
1041 * (although both would be easy to implement).
1043 int tcp_read_sock(struct sock *sk, read_descriptor_t *desc,
1044 sk_read_actor_t recv_actor)
1046 struct sk_buff *skb;
1047 struct tcp_sock *tp = tcp_sk(sk);
1048 u32 seq = tp->copied_seq;
1052 if (sk->sk_state == TCP_LISTEN)
1054 while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1055 if (offset < skb->len) {
1058 len = skb->len - offset;
1059 /* Stop reading if we hit a patch of urgent data */
1061 u32 urg_offset = tp->urg_seq - seq;
1062 if (urg_offset < len)
1067 used = recv_actor(desc, skb, offset, len);
1073 if (offset != skb->len)
1076 if (tcp_hdr(skb)->fin) {
1077 sk_eat_skb(sk, skb, 0);
1081 sk_eat_skb(sk, skb, 0);
1085 tp->copied_seq = seq;
1087 tcp_rcv_space_adjust(sk);
1089 /* Clean up data we have read: This will do ACK frames. */
1091 tcp_cleanup_rbuf(sk, copied);
1096 * This routine copies from a sock struct into the user buffer.
1098 * Technical note: in 2.3 we work on _locked_ socket, so that
1099 * tricks with *seq access order and skb->users are not required.
1100 * Probably, code can be easily improved even more.
1103 int tcp_recvmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg,
1104 size_t len, int nonblock, int flags, int *addr_len)
1106 struct tcp_sock *tp = tcp_sk(sk);
1112 int target; /* Read at least this many bytes */
1114 struct task_struct *user_recv = NULL;
1115 int copied_early = 0;
1119 TCP_CHECK_TIMER(sk);
1122 if (sk->sk_state == TCP_LISTEN)
1125 timeo = sock_rcvtimeo(sk, nonblock);
1127 /* Urgent data needs to be handled specially. */
1128 if (flags & MSG_OOB)
1131 seq = &tp->copied_seq;
1132 if (flags & MSG_PEEK) {
1133 peek_seq = tp->copied_seq;
1137 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
1139 #ifdef CONFIG_NET_DMA
1140 tp->ucopy.dma_chan = NULL;
1142 if ((len > sysctl_tcp_dma_copybreak) && !(flags & MSG_PEEK) &&
1143 !sysctl_tcp_low_latency && __get_cpu_var(softnet_data).net_dma) {
1144 preempt_enable_no_resched();
1145 tp->ucopy.pinned_list = dma_pin_iovec_pages(msg->msg_iov, len);
1147 preempt_enable_no_resched();
1151 struct sk_buff *skb;
1154 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
1155 if (tp->urg_data && tp->urg_seq == *seq) {
1158 if (signal_pending(current)) {
1159 copied = timeo ? sock_intr_errno(timeo) : -EAGAIN;
1164 /* Next get a buffer. */
1166 skb = skb_peek(&sk->sk_receive_queue);
1171 /* Now that we have two receive queues this
1174 if (before(*seq, TCP_SKB_CB(skb)->seq)) {
1175 printk(KERN_INFO "recvmsg bug: copied %X "
1176 "seq %X\n", *seq, TCP_SKB_CB(skb)->seq);
1179 offset = *seq - TCP_SKB_CB(skb)->seq;
1180 if (tcp_hdr(skb)->syn)
1182 if (offset < skb->len)
1184 if (tcp_hdr(skb)->fin)
1186 BUG_TRAP(flags & MSG_PEEK);
1188 } while (skb != (struct sk_buff *)&sk->sk_receive_queue);
1190 /* Well, if we have backlog, try to process it now yet. */
1192 if (copied >= target && !sk->sk_backlog.tail)
1197 sk->sk_state == TCP_CLOSE ||
1198 (sk->sk_shutdown & RCV_SHUTDOWN) ||
1200 signal_pending(current) ||
1204 if (sock_flag(sk, SOCK_DONE))
1208 copied = sock_error(sk);
1212 if (sk->sk_shutdown & RCV_SHUTDOWN)
1215 if (sk->sk_state == TCP_CLOSE) {
1216 if (!sock_flag(sk, SOCK_DONE)) {
1217 /* This occurs when user tries to read
1218 * from never connected socket.
1231 if (signal_pending(current)) {
1232 copied = sock_intr_errno(timeo);
1237 tcp_cleanup_rbuf(sk, copied);
1239 if (!sysctl_tcp_low_latency && tp->ucopy.task == user_recv) {
1240 /* Install new reader */
1241 if (!user_recv && !(flags & (MSG_TRUNC | MSG_PEEK))) {
1242 user_recv = current;
1243 tp->ucopy.task = user_recv;
1244 tp->ucopy.iov = msg->msg_iov;
1247 tp->ucopy.len = len;
1249 BUG_TRAP(tp->copied_seq == tp->rcv_nxt ||
1250 (flags & (MSG_PEEK | MSG_TRUNC)));
1252 /* Ugly... If prequeue is not empty, we have to
1253 * process it before releasing socket, otherwise
1254 * order will be broken at second iteration.
1255 * More elegant solution is required!!!
1257 * Look: we have the following (pseudo)queues:
1259 * 1. packets in flight
1264 * Each queue can be processed only if the next ones
1265 * are empty. At this point we have empty receive_queue.
1266 * But prequeue _can_ be not empty after 2nd iteration,
1267 * when we jumped to start of loop because backlog
1268 * processing added something to receive_queue.
1269 * We cannot release_sock(), because backlog contains
1270 * packets arrived _after_ prequeued ones.
1272 * Shortly, algorithm is clear --- to process all
1273 * the queues in order. We could make it more directly,
1274 * requeueing packets from backlog to prequeue, if
1275 * is not empty. It is more elegant, but eats cycles,
1278 if (!skb_queue_empty(&tp->ucopy.prequeue))
1281 /* __ Set realtime policy in scheduler __ */
1284 if (copied >= target) {
1285 /* Do not sleep, just process backlog. */
1289 sk_wait_data(sk, &timeo);
1291 #ifdef CONFIG_NET_DMA
1292 tp->ucopy.wakeup = 0;
1298 /* __ Restore normal policy in scheduler __ */
1300 if ((chunk = len - tp->ucopy.len) != 0) {
1301 NET_ADD_STATS_USER(LINUX_MIB_TCPDIRECTCOPYFROMBACKLOG, chunk);
1306 if (tp->rcv_nxt == tp->copied_seq &&
1307 !skb_queue_empty(&tp->ucopy.prequeue)) {
1309 tcp_prequeue_process(sk);
1311 if ((chunk = len - tp->ucopy.len) != 0) {
1312 NET_ADD_STATS_USER(LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1318 if ((flags & MSG_PEEK) && peek_seq != tp->copied_seq) {
1319 if (net_ratelimit())
1320 printk(KERN_DEBUG "TCP(%s:%d): Application bug, race in MSG_PEEK.\n",
1321 current->comm, current->pid);
1322 peek_seq = tp->copied_seq;
1327 /* Ok so how much can we use? */
1328 used = skb->len - offset;
1332 /* Do we have urgent data here? */
1334 u32 urg_offset = tp->urg_seq - *seq;
1335 if (urg_offset < used) {
1337 if (!sock_flag(sk, SOCK_URGINLINE)) {
1349 if (!(flags & MSG_TRUNC)) {
1350 #ifdef CONFIG_NET_DMA
1351 if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list)
1352 tp->ucopy.dma_chan = get_softnet_dma();
1354 if (tp->ucopy.dma_chan) {
1355 tp->ucopy.dma_cookie = dma_skb_copy_datagram_iovec(
1356 tp->ucopy.dma_chan, skb, offset,
1358 tp->ucopy.pinned_list);
1360 if (tp->ucopy.dma_cookie < 0) {
1362 printk(KERN_ALERT "dma_cookie < 0\n");
1364 /* Exception. Bailout! */
1369 if ((offset + used) == skb->len)
1375 err = skb_copy_datagram_iovec(skb, offset,
1376 msg->msg_iov, used);
1378 /* Exception. Bailout! */
1390 tcp_rcv_space_adjust(sk);
1393 if (tp->urg_data && after(tp->copied_seq, tp->urg_seq)) {
1395 tcp_fast_path_check(sk);
1397 if (used + offset < skb->len)
1400 if (tcp_hdr(skb)->fin)
1402 if (!(flags & MSG_PEEK)) {
1403 sk_eat_skb(sk, skb, copied_early);
1409 /* Process the FIN. */
1411 if (!(flags & MSG_PEEK)) {
1412 sk_eat_skb(sk, skb, copied_early);
1419 if (!skb_queue_empty(&tp->ucopy.prequeue)) {
1422 tp->ucopy.len = copied > 0 ? len : 0;
1424 tcp_prequeue_process(sk);
1426 if (copied > 0 && (chunk = len - tp->ucopy.len) != 0) {
1427 NET_ADD_STATS_USER(LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1433 tp->ucopy.task = NULL;
1437 #ifdef CONFIG_NET_DMA
1438 if (tp->ucopy.dma_chan) {
1439 struct sk_buff *skb;
1440 dma_cookie_t done, used;
1442 dma_async_memcpy_issue_pending(tp->ucopy.dma_chan);
1444 while (dma_async_memcpy_complete(tp->ucopy.dma_chan,
1445 tp->ucopy.dma_cookie, &done,
1446 &used) == DMA_IN_PROGRESS) {
1447 /* do partial cleanup of sk_async_wait_queue */
1448 while ((skb = skb_peek(&sk->sk_async_wait_queue)) &&
1449 (dma_async_is_complete(skb->dma_cookie, done,
1450 used) == DMA_SUCCESS)) {
1451 __skb_dequeue(&sk->sk_async_wait_queue);
1456 /* Safe to free early-copied skbs now */
1457 __skb_queue_purge(&sk->sk_async_wait_queue);
1458 dma_chan_put(tp->ucopy.dma_chan);
1459 tp->ucopy.dma_chan = NULL;
1461 if (tp->ucopy.pinned_list) {
1462 dma_unpin_iovec_pages(tp->ucopy.pinned_list);
1463 tp->ucopy.pinned_list = NULL;
1467 /* According to UNIX98, msg_name/msg_namelen are ignored
1468 * on connected socket. I was just happy when found this 8) --ANK
1471 /* Clean up data we have read: This will do ACK frames. */
1472 tcp_cleanup_rbuf(sk, copied);
1474 TCP_CHECK_TIMER(sk);
1479 TCP_CHECK_TIMER(sk);
1484 err = tcp_recv_urg(sk, timeo, msg, len, flags, addr_len);
1489 * State processing on a close. This implements the state shift for
1490 * sending our FIN frame. Note that we only send a FIN for some
1491 * states. A shutdown() may have already sent the FIN, or we may be
1495 static const unsigned char new_state[16] = {
1496 /* current state: new state: action: */
1497 /* (Invalid) */ TCP_CLOSE,
1498 /* TCP_ESTABLISHED */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1499 /* TCP_SYN_SENT */ TCP_CLOSE,
1500 /* TCP_SYN_RECV */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1501 /* TCP_FIN_WAIT1 */ TCP_FIN_WAIT1,
1502 /* TCP_FIN_WAIT2 */ TCP_FIN_WAIT2,
1503 /* TCP_TIME_WAIT */ TCP_CLOSE,
1504 /* TCP_CLOSE */ TCP_CLOSE,
1505 /* TCP_CLOSE_WAIT */ TCP_LAST_ACK | TCP_ACTION_FIN,
1506 /* TCP_LAST_ACK */ TCP_LAST_ACK,
1507 /* TCP_LISTEN */ TCP_CLOSE,
1508 /* TCP_CLOSING */ TCP_CLOSING,
1511 static int tcp_close_state(struct sock *sk)
1513 int next = (int)new_state[sk->sk_state];
1514 int ns = next & TCP_STATE_MASK;
1516 tcp_set_state(sk, ns);
1518 return next & TCP_ACTION_FIN;
1522 * Shutdown the sending side of a connection. Much like close except
1523 * that we don't receive shut down or set_sock_flag(sk, SOCK_DEAD).
1526 void tcp_shutdown(struct sock *sk, int how)
1528 /* We need to grab some memory, and put together a FIN,
1529 * and then put it into the queue to be sent.
1530 * Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
1532 if (!(how & SEND_SHUTDOWN))
1535 /* If we've already sent a FIN, or it's a closed state, skip this. */
1536 if ((1 << sk->sk_state) &
1537 (TCPF_ESTABLISHED | TCPF_SYN_SENT |
1538 TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) {
1539 /* Clear out any half completed packets. FIN if needed. */
1540 if (tcp_close_state(sk))
1545 void tcp_close(struct sock *sk, long timeout)
1547 struct sk_buff *skb;
1548 int data_was_unread = 0;
1552 sk->sk_shutdown = SHUTDOWN_MASK;
1554 if (sk->sk_state == TCP_LISTEN) {
1555 tcp_set_state(sk, TCP_CLOSE);
1558 inet_csk_listen_stop(sk);
1560 goto adjudge_to_death;
1563 /* We need to flush the recv. buffs. We do this only on the
1564 * descriptor close, not protocol-sourced closes, because the
1565 * reader process may not have drained the data yet!
1567 while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) {
1568 u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq -
1570 data_was_unread += len;
1574 sk_stream_mem_reclaim(sk);
1576 /* As outlined in draft-ietf-tcpimpl-prob-03.txt, section
1577 * 3.10, we send a RST here because data was lost. To
1578 * witness the awful effects of the old behavior of always
1579 * doing a FIN, run an older 2.1.x kernel or 2.0.x, start
1580 * a bulk GET in an FTP client, suspend the process, wait
1581 * for the client to advertise a zero window, then kill -9
1582 * the FTP client, wheee... Note: timeout is always zero
1585 if (data_was_unread) {
1586 /* Unread data was tossed, zap the connection. */
1587 NET_INC_STATS_USER(LINUX_MIB_TCPABORTONCLOSE);
1588 tcp_set_state(sk, TCP_CLOSE);
1589 tcp_send_active_reset(sk, GFP_KERNEL);
1590 } else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) {
1591 /* Check zero linger _after_ checking for unread data. */
1592 sk->sk_prot->disconnect(sk, 0);
1593 NET_INC_STATS_USER(LINUX_MIB_TCPABORTONDATA);
1594 } else if (tcp_close_state(sk)) {
1595 /* We FIN if the application ate all the data before
1596 * zapping the connection.
1599 /* RED-PEN. Formally speaking, we have broken TCP state
1600 * machine. State transitions:
1602 * TCP_ESTABLISHED -> TCP_FIN_WAIT1
1603 * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible)
1604 * TCP_CLOSE_WAIT -> TCP_LAST_ACK
1606 * are legal only when FIN has been sent (i.e. in window),
1607 * rather than queued out of window. Purists blame.
1609 * F.e. "RFC state" is ESTABLISHED,
1610 * if Linux state is FIN-WAIT-1, but FIN is still not sent.
1612 * The visible declinations are that sometimes
1613 * we enter time-wait state, when it is not required really
1614 * (harmless), do not send active resets, when they are
1615 * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
1616 * they look as CLOSING or LAST_ACK for Linux)
1617 * Probably, I missed some more holelets.
1623 sk_stream_wait_close(sk, timeout);
1626 state = sk->sk_state;
1629 atomic_inc(sk->sk_prot->orphan_count);
1631 /* It is the last release_sock in its life. It will remove backlog. */
1635 /* Now socket is owned by kernel and we acquire BH lock
1636 to finish close. No need to check for user refs.
1640 BUG_TRAP(!sock_owned_by_user(sk));
1642 /* Have we already been destroyed by a softirq or backlog? */
1643 if (state != TCP_CLOSE && sk->sk_state == TCP_CLOSE)
1646 /* This is a (useful) BSD violating of the RFC. There is a
1647 * problem with TCP as specified in that the other end could
1648 * keep a socket open forever with no application left this end.
1649 * We use a 3 minute timeout (about the same as BSD) then kill
1650 * our end. If they send after that then tough - BUT: long enough
1651 * that we won't make the old 4*rto = almost no time - whoops
1654 * Nope, it was not mistake. It is really desired behaviour
1655 * f.e. on http servers, when such sockets are useless, but
1656 * consume significant resources. Let's do it with special
1657 * linger2 option. --ANK
1660 if (sk->sk_state == TCP_FIN_WAIT2) {
1661 struct tcp_sock *tp = tcp_sk(sk);
1662 if (tp->linger2 < 0) {
1663 tcp_set_state(sk, TCP_CLOSE);
1664 tcp_send_active_reset(sk, GFP_ATOMIC);
1665 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONLINGER);
1667 const int tmo = tcp_fin_time(sk);
1669 if (tmo > TCP_TIMEWAIT_LEN) {
1670 inet_csk_reset_keepalive_timer(sk,
1671 tmo - TCP_TIMEWAIT_LEN);
1673 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
1678 if (sk->sk_state != TCP_CLOSE) {
1679 sk_stream_mem_reclaim(sk);
1680 if (atomic_read(sk->sk_prot->orphan_count) > sysctl_tcp_max_orphans ||
1681 (sk->sk_wmem_queued > SOCK_MIN_SNDBUF &&
1682 atomic_read(&tcp_memory_allocated) > sysctl_tcp_mem[2])) {
1683 if (net_ratelimit())
1684 printk(KERN_INFO "TCP: too many of orphaned "
1686 tcp_set_state(sk, TCP_CLOSE);
1687 tcp_send_active_reset(sk, GFP_ATOMIC);
1688 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONMEMORY);
1692 if (sk->sk_state == TCP_CLOSE)
1693 inet_csk_destroy_sock(sk);
1694 /* Otherwise, socket is reprieved until protocol close. */
1702 /* These states need RST on ABORT according to RFC793 */
1704 static inline int tcp_need_reset(int state)
1706 return (1 << state) &
1707 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 |
1708 TCPF_FIN_WAIT2 | TCPF_SYN_RECV);
1711 int tcp_disconnect(struct sock *sk, int flags)
1713 struct inet_sock *inet = inet_sk(sk);
1714 struct inet_connection_sock *icsk = inet_csk(sk);
1715 struct tcp_sock *tp = tcp_sk(sk);
1717 int old_state = sk->sk_state;
1719 if (old_state != TCP_CLOSE)
1720 tcp_set_state(sk, TCP_CLOSE);
1722 /* ABORT function of RFC793 */
1723 if (old_state == TCP_LISTEN) {
1724 inet_csk_listen_stop(sk);
1725 } else if (tcp_need_reset(old_state) ||
1726 (tp->snd_nxt != tp->write_seq &&
1727 (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK))) {
1728 /* The last check adjusts for discrepancy of Linux wrt. RFC
1731 tcp_send_active_reset(sk, gfp_any());
1732 sk->sk_err = ECONNRESET;
1733 } else if (old_state == TCP_SYN_SENT)
1734 sk->sk_err = ECONNRESET;
1736 tcp_clear_xmit_timers(sk);
1737 __skb_queue_purge(&sk->sk_receive_queue);
1738 tcp_write_queue_purge(sk);
1739 __skb_queue_purge(&tp->out_of_order_queue);
1740 #ifdef CONFIG_NET_DMA
1741 __skb_queue_purge(&sk->sk_async_wait_queue);
1746 if (!(sk->sk_userlocks & SOCK_BINDADDR_LOCK))
1747 inet_reset_saddr(sk);
1749 sk->sk_shutdown = 0;
1750 sock_reset_flag(sk, SOCK_DONE);
1752 if ((tp->write_seq += tp->max_window + 2) == 0)
1754 icsk->icsk_backoff = 0;
1756 icsk->icsk_probes_out = 0;
1757 tp->packets_out = 0;
1758 tp->snd_ssthresh = 0x7fffffff;
1759 tp->snd_cwnd_cnt = 0;
1760 tp->bytes_acked = 0;
1761 tcp_set_ca_state(sk, TCP_CA_Open);
1762 tcp_clear_retrans(tp);
1763 inet_csk_delack_init(sk);
1764 tcp_init_send_head(sk);
1765 tp->rx_opt.saw_tstamp = 0;
1766 tcp_sack_reset(&tp->rx_opt);
1769 BUG_TRAP(!inet->num || icsk->icsk_bind_hash);
1771 sk->sk_error_report(sk);
1776 * Socket option code for TCP.
1778 static int do_tcp_setsockopt(struct sock *sk, int level,
1779 int optname, char __user *optval, int optlen)
1781 struct tcp_sock *tp = tcp_sk(sk);
1782 struct inet_connection_sock *icsk = inet_csk(sk);
1786 /* This is a string value all the others are int's */
1787 if (optname == TCP_CONGESTION) {
1788 char name[TCP_CA_NAME_MAX];
1793 val = strncpy_from_user(name, optval,
1794 min(TCP_CA_NAME_MAX-1, optlen));
1800 err = tcp_set_congestion_control(sk, name);
1805 if (optlen < sizeof(int))
1808 if (get_user(val, (int __user *)optval))
1815 /* Values greater than interface MTU won't take effect. However
1816 * at the point when this call is done we typically don't yet
1817 * know which interface is going to be used */
1818 if (val < 8 || val > MAX_TCP_WINDOW) {
1822 tp->rx_opt.user_mss = val;
1827 /* TCP_NODELAY is weaker than TCP_CORK, so that
1828 * this option on corked socket is remembered, but
1829 * it is not activated until cork is cleared.
1831 * However, when TCP_NODELAY is set we make
1832 * an explicit push, which overrides even TCP_CORK
1833 * for currently queued segments.
1835 tp->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH;
1836 tcp_push_pending_frames(sk);
1838 tp->nonagle &= ~TCP_NAGLE_OFF;
1843 /* When set indicates to always queue non-full frames.
1844 * Later the user clears this option and we transmit
1845 * any pending partial frames in the queue. This is
1846 * meant to be used alongside sendfile() to get properly
1847 * filled frames when the user (for example) must write
1848 * out headers with a write() call first and then use
1849 * sendfile to send out the data parts.
1851 * TCP_CORK can be set together with TCP_NODELAY and it is
1852 * stronger than TCP_NODELAY.
1855 tp->nonagle |= TCP_NAGLE_CORK;
1857 tp->nonagle &= ~TCP_NAGLE_CORK;
1858 if (tp->nonagle&TCP_NAGLE_OFF)
1859 tp->nonagle |= TCP_NAGLE_PUSH;
1860 tcp_push_pending_frames(sk);
1865 if (val < 1 || val > MAX_TCP_KEEPIDLE)
1868 tp->keepalive_time = val * HZ;
1869 if (sock_flag(sk, SOCK_KEEPOPEN) &&
1870 !((1 << sk->sk_state) &
1871 (TCPF_CLOSE | TCPF_LISTEN))) {
1872 __u32 elapsed = tcp_time_stamp - tp->rcv_tstamp;
1873 if (tp->keepalive_time > elapsed)
1874 elapsed = tp->keepalive_time - elapsed;
1877 inet_csk_reset_keepalive_timer(sk, elapsed);
1882 if (val < 1 || val > MAX_TCP_KEEPINTVL)
1885 tp->keepalive_intvl = val * HZ;
1888 if (val < 1 || val > MAX_TCP_KEEPCNT)
1891 tp->keepalive_probes = val;
1894 if (val < 1 || val > MAX_TCP_SYNCNT)
1897 icsk->icsk_syn_retries = val;
1903 else if (val > sysctl_tcp_fin_timeout / HZ)
1906 tp->linger2 = val * HZ;
1909 case TCP_DEFER_ACCEPT:
1910 icsk->icsk_accept_queue.rskq_defer_accept = 0;
1912 /* Translate value in seconds to number of
1914 while (icsk->icsk_accept_queue.rskq_defer_accept < 32 &&
1915 val > ((TCP_TIMEOUT_INIT / HZ) <<
1916 icsk->icsk_accept_queue.rskq_defer_accept))
1917 icsk->icsk_accept_queue.rskq_defer_accept++;
1918 icsk->icsk_accept_queue.rskq_defer_accept++;
1922 case TCP_WINDOW_CLAMP:
1924 if (sk->sk_state != TCP_CLOSE) {
1928 tp->window_clamp = 0;
1930 tp->window_clamp = val < SOCK_MIN_RCVBUF / 2 ?
1931 SOCK_MIN_RCVBUF / 2 : val;
1936 icsk->icsk_ack.pingpong = 1;
1938 icsk->icsk_ack.pingpong = 0;
1939 if ((1 << sk->sk_state) &
1940 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) &&
1941 inet_csk_ack_scheduled(sk)) {
1942 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
1943 tcp_cleanup_rbuf(sk, 1);
1945 icsk->icsk_ack.pingpong = 1;
1950 #ifdef CONFIG_TCP_MD5SIG
1952 /* Read the IP->Key mappings from userspace */
1953 err = tp->af_specific->md5_parse(sk, optval, optlen);
1966 int tcp_setsockopt(struct sock *sk, int level, int optname, char __user *optval,
1969 struct inet_connection_sock *icsk = inet_csk(sk);
1971 if (level != SOL_TCP)
1972 return icsk->icsk_af_ops->setsockopt(sk, level, optname,
1974 return do_tcp_setsockopt(sk, level, optname, optval, optlen);
1977 #ifdef CONFIG_COMPAT
1978 int compat_tcp_setsockopt(struct sock *sk, int level, int optname,
1979 char __user *optval, int optlen)
1981 if (level != SOL_TCP)
1982 return inet_csk_compat_setsockopt(sk, level, optname,
1984 return do_tcp_setsockopt(sk, level, optname, optval, optlen);
1987 EXPORT_SYMBOL(compat_tcp_setsockopt);
1990 /* Return information about state of tcp endpoint in API format. */
1991 void tcp_get_info(struct sock *sk, struct tcp_info *info)
1993 struct tcp_sock *tp = tcp_sk(sk);
1994 const struct inet_connection_sock *icsk = inet_csk(sk);
1995 u32 now = tcp_time_stamp;
1997 memset(info, 0, sizeof(*info));
1999 info->tcpi_state = sk->sk_state;
2000 info->tcpi_ca_state = icsk->icsk_ca_state;
2001 info->tcpi_retransmits = icsk->icsk_retransmits;
2002 info->tcpi_probes = icsk->icsk_probes_out;
2003 info->tcpi_backoff = icsk->icsk_backoff;
2005 if (tp->rx_opt.tstamp_ok)
2006 info->tcpi_options |= TCPI_OPT_TIMESTAMPS;
2007 if (tp->rx_opt.sack_ok)
2008 info->tcpi_options |= TCPI_OPT_SACK;
2009 if (tp->rx_opt.wscale_ok) {
2010 info->tcpi_options |= TCPI_OPT_WSCALE;
2011 info->tcpi_snd_wscale = tp->rx_opt.snd_wscale;
2012 info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale;
2015 if (tp->ecn_flags&TCP_ECN_OK)
2016 info->tcpi_options |= TCPI_OPT_ECN;
2018 info->tcpi_rto = jiffies_to_usecs(icsk->icsk_rto);
2019 info->tcpi_ato = jiffies_to_usecs(icsk->icsk_ack.ato);
2020 info->tcpi_snd_mss = tp->mss_cache;
2021 info->tcpi_rcv_mss = icsk->icsk_ack.rcv_mss;
2023 info->tcpi_unacked = tp->packets_out;
2024 info->tcpi_sacked = tp->sacked_out;
2025 info->tcpi_lost = tp->lost_out;
2026 info->tcpi_retrans = tp->retrans_out;
2027 info->tcpi_fackets = tp->fackets_out;
2029 info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime);
2030 info->tcpi_last_data_recv = jiffies_to_msecs(now - icsk->icsk_ack.lrcvtime);
2031 info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp);
2033 info->tcpi_pmtu = icsk->icsk_pmtu_cookie;
2034 info->tcpi_rcv_ssthresh = tp->rcv_ssthresh;
2035 info->tcpi_rtt = jiffies_to_usecs(tp->srtt)>>3;
2036 info->tcpi_rttvar = jiffies_to_usecs(tp->mdev)>>2;
2037 info->tcpi_snd_ssthresh = tp->snd_ssthresh;
2038 info->tcpi_snd_cwnd = tp->snd_cwnd;
2039 info->tcpi_advmss = tp->advmss;
2040 info->tcpi_reordering = tp->reordering;
2042 info->tcpi_rcv_rtt = jiffies_to_usecs(tp->rcv_rtt_est.rtt)>>3;
2043 info->tcpi_rcv_space = tp->rcvq_space.space;
2045 info->tcpi_total_retrans = tp->total_retrans;
2048 EXPORT_SYMBOL_GPL(tcp_get_info);
2050 static int do_tcp_getsockopt(struct sock *sk, int level,
2051 int optname, char __user *optval, int __user *optlen)
2053 struct inet_connection_sock *icsk = inet_csk(sk);
2054 struct tcp_sock *tp = tcp_sk(sk);
2057 if (get_user(len, optlen))
2060 len = min_t(unsigned int, len, sizeof(int));
2067 val = tp->mss_cache;
2068 if (!val && ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
2069 val = tp->rx_opt.user_mss;
2072 val = !!(tp->nonagle&TCP_NAGLE_OFF);
2075 val = !!(tp->nonagle&TCP_NAGLE_CORK);
2078 val = (tp->keepalive_time ? : sysctl_tcp_keepalive_time) / HZ;
2081 val = (tp->keepalive_intvl ? : sysctl_tcp_keepalive_intvl) / HZ;
2084 val = tp->keepalive_probes ? : sysctl_tcp_keepalive_probes;
2087 val = icsk->icsk_syn_retries ? : sysctl_tcp_syn_retries;
2092 val = (val ? : sysctl_tcp_fin_timeout) / HZ;
2094 case TCP_DEFER_ACCEPT:
2095 val = !icsk->icsk_accept_queue.rskq_defer_accept ? 0 :
2096 ((TCP_TIMEOUT_INIT / HZ) << (icsk->icsk_accept_queue.rskq_defer_accept - 1));
2098 case TCP_WINDOW_CLAMP:
2099 val = tp->window_clamp;
2102 struct tcp_info info;
2104 if (get_user(len, optlen))
2107 tcp_get_info(sk, &info);
2109 len = min_t(unsigned int, len, sizeof(info));
2110 if (put_user(len, optlen))
2112 if (copy_to_user(optval, &info, len))
2117 val = !icsk->icsk_ack.pingpong;
2120 case TCP_CONGESTION:
2121 if (get_user(len, optlen))
2123 len = min_t(unsigned int, len, TCP_CA_NAME_MAX);
2124 if (put_user(len, optlen))
2126 if (copy_to_user(optval, icsk->icsk_ca_ops->name, len))
2130 return -ENOPROTOOPT;
2133 if (put_user(len, optlen))
2135 if (copy_to_user(optval, &val, len))
2140 int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval,
2143 struct inet_connection_sock *icsk = inet_csk(sk);
2145 if (level != SOL_TCP)
2146 return icsk->icsk_af_ops->getsockopt(sk, level, optname,
2148 return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2151 #ifdef CONFIG_COMPAT
2152 int compat_tcp_getsockopt(struct sock *sk, int level, int optname,
2153 char __user *optval, int __user *optlen)
2155 if (level != SOL_TCP)
2156 return inet_csk_compat_getsockopt(sk, level, optname,
2158 return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2161 EXPORT_SYMBOL(compat_tcp_getsockopt);
2164 struct sk_buff *tcp_tso_segment(struct sk_buff *skb, int features)
2166 struct sk_buff *segs = ERR_PTR(-EINVAL);
2171 unsigned int oldlen;
2174 if (!pskb_may_pull(skb, sizeof(*th)))
2178 thlen = th->doff * 4;
2179 if (thlen < sizeof(*th))
2182 if (!pskb_may_pull(skb, thlen))
2185 oldlen = (u16)~skb->len;
2186 __skb_pull(skb, thlen);
2188 if (skb_gso_ok(skb, features | NETIF_F_GSO_ROBUST)) {
2189 /* Packet is from an untrusted source, reset gso_segs. */
2190 int type = skb_shinfo(skb)->gso_type;
2199 !(type & (SKB_GSO_TCPV4 | SKB_GSO_TCPV6))))
2202 mss = skb_shinfo(skb)->gso_size;
2203 skb_shinfo(skb)->gso_segs = (skb->len + mss - 1) / mss;
2209 segs = skb_segment(skb, features);
2213 len = skb_shinfo(skb)->gso_size;
2214 delta = htonl(oldlen + (thlen + len));
2218 seq = ntohl(th->seq);
2221 th->fin = th->psh = 0;
2223 th->check = ~csum_fold((__force __wsum)((__force u32)th->check +
2224 (__force u32)delta));
2225 if (skb->ip_summed != CHECKSUM_PARTIAL)
2227 csum_fold(csum_partial(skb_transport_header(skb),
2234 th->seq = htonl(seq);
2236 } while (skb->next);
2238 delta = htonl(oldlen + (skb->tail - skb->transport_header) +
2240 th->check = ~csum_fold((__force __wsum)((__force u32)th->check +
2241 (__force u32)delta));
2242 if (skb->ip_summed != CHECKSUM_PARTIAL)
2243 th->check = csum_fold(csum_partial(skb_transport_header(skb),
2249 EXPORT_SYMBOL(tcp_tso_segment);
2251 #ifdef CONFIG_TCP_MD5SIG
2252 static unsigned long tcp_md5sig_users;
2253 static struct tcp_md5sig_pool **tcp_md5sig_pool;
2254 static DEFINE_SPINLOCK(tcp_md5sig_pool_lock);
2256 static void __tcp_free_md5sig_pool(struct tcp_md5sig_pool **pool)
2259 for_each_possible_cpu(cpu) {
2260 struct tcp_md5sig_pool *p = *per_cpu_ptr(pool, cpu);
2262 if (p->md5_desc.tfm)
2263 crypto_free_hash(p->md5_desc.tfm);
2271 void tcp_free_md5sig_pool(void)
2273 struct tcp_md5sig_pool **pool = NULL;
2275 spin_lock_bh(&tcp_md5sig_pool_lock);
2276 if (--tcp_md5sig_users == 0) {
2277 pool = tcp_md5sig_pool;
2278 tcp_md5sig_pool = NULL;
2280 spin_unlock_bh(&tcp_md5sig_pool_lock);
2282 __tcp_free_md5sig_pool(pool);
2285 EXPORT_SYMBOL(tcp_free_md5sig_pool);
2287 static struct tcp_md5sig_pool **__tcp_alloc_md5sig_pool(void)
2290 struct tcp_md5sig_pool **pool;
2292 pool = alloc_percpu(struct tcp_md5sig_pool *);
2296 for_each_possible_cpu(cpu) {
2297 struct tcp_md5sig_pool *p;
2298 struct crypto_hash *hash;
2300 p = kzalloc(sizeof(*p), GFP_KERNEL);
2303 *per_cpu_ptr(pool, cpu) = p;
2305 hash = crypto_alloc_hash("md5", 0, CRYPTO_ALG_ASYNC);
2306 if (!hash || IS_ERR(hash))
2309 p->md5_desc.tfm = hash;
2313 __tcp_free_md5sig_pool(pool);
2317 struct tcp_md5sig_pool **tcp_alloc_md5sig_pool(void)
2319 struct tcp_md5sig_pool **pool;
2323 spin_lock_bh(&tcp_md5sig_pool_lock);
2324 pool = tcp_md5sig_pool;
2325 if (tcp_md5sig_users++ == 0) {
2327 spin_unlock_bh(&tcp_md5sig_pool_lock);
2330 spin_unlock_bh(&tcp_md5sig_pool_lock);
2334 spin_unlock_bh(&tcp_md5sig_pool_lock);
2337 /* we cannot hold spinlock here because this may sleep. */
2338 struct tcp_md5sig_pool **p = __tcp_alloc_md5sig_pool();
2339 spin_lock_bh(&tcp_md5sig_pool_lock);
2342 spin_unlock_bh(&tcp_md5sig_pool_lock);
2345 pool = tcp_md5sig_pool;
2347 /* oops, it has already been assigned. */
2348 spin_unlock_bh(&tcp_md5sig_pool_lock);
2349 __tcp_free_md5sig_pool(p);
2351 tcp_md5sig_pool = pool = p;
2352 spin_unlock_bh(&tcp_md5sig_pool_lock);
2358 EXPORT_SYMBOL(tcp_alloc_md5sig_pool);
2360 struct tcp_md5sig_pool *__tcp_get_md5sig_pool(int cpu)
2362 struct tcp_md5sig_pool **p;
2363 spin_lock_bh(&tcp_md5sig_pool_lock);
2364 p = tcp_md5sig_pool;
2367 spin_unlock_bh(&tcp_md5sig_pool_lock);
2368 return (p ? *per_cpu_ptr(p, cpu) : NULL);
2371 EXPORT_SYMBOL(__tcp_get_md5sig_pool);
2373 void __tcp_put_md5sig_pool(void)
2375 tcp_free_md5sig_pool();
2378 EXPORT_SYMBOL(__tcp_put_md5sig_pool);
2381 void tcp_done(struct sock *sk)
2383 if(sk->sk_state == TCP_SYN_SENT || sk->sk_state == TCP_SYN_RECV)
2384 TCP_INC_STATS_BH(TCP_MIB_ATTEMPTFAILS);
2386 tcp_set_state(sk, TCP_CLOSE);
2387 tcp_clear_xmit_timers(sk);
2389 sk->sk_shutdown = SHUTDOWN_MASK;
2391 if (!sock_flag(sk, SOCK_DEAD))
2392 sk->sk_state_change(sk);
2394 inet_csk_destroy_sock(sk);
2396 EXPORT_SYMBOL_GPL(tcp_done);
2398 extern void __skb_cb_too_small_for_tcp(int, int);
2399 extern struct tcp_congestion_ops tcp_reno;
2401 static __initdata unsigned long thash_entries;
2402 static int __init set_thash_entries(char *str)
2406 thash_entries = simple_strtoul(str, &str, 0);
2409 __setup("thash_entries=", set_thash_entries);
2411 void __init tcp_init(void)
2413 struct sk_buff *skb = NULL;
2414 unsigned long limit;
2415 int order, i, max_share;
2417 if (sizeof(struct tcp_skb_cb) > sizeof(skb->cb))
2418 __skb_cb_too_small_for_tcp(sizeof(struct tcp_skb_cb),
2421 tcp_hashinfo.bind_bucket_cachep =
2422 kmem_cache_create("tcp_bind_bucket",
2423 sizeof(struct inet_bind_bucket), 0,
2424 SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL, NULL);
2426 /* Size and allocate the main established and bind bucket
2429 * The methodology is similar to that of the buffer cache.
2431 tcp_hashinfo.ehash =
2432 alloc_large_system_hash("TCP established",
2433 sizeof(struct inet_ehash_bucket),
2435 (num_physpages >= 128 * 1024) ?
2438 &tcp_hashinfo.ehash_size,
2441 tcp_hashinfo.ehash_size = 1 << tcp_hashinfo.ehash_size;
2442 for (i = 0; i < tcp_hashinfo.ehash_size; i++) {
2443 rwlock_init(&tcp_hashinfo.ehash[i].lock);
2444 INIT_HLIST_HEAD(&tcp_hashinfo.ehash[i].chain);
2445 INIT_HLIST_HEAD(&tcp_hashinfo.ehash[i].twchain);
2448 tcp_hashinfo.bhash =
2449 alloc_large_system_hash("TCP bind",
2450 sizeof(struct inet_bind_hashbucket),
2451 tcp_hashinfo.ehash_size,
2452 (num_physpages >= 128 * 1024) ?
2455 &tcp_hashinfo.bhash_size,
2458 tcp_hashinfo.bhash_size = 1 << tcp_hashinfo.bhash_size;
2459 for (i = 0; i < tcp_hashinfo.bhash_size; i++) {
2460 spin_lock_init(&tcp_hashinfo.bhash[i].lock);
2461 INIT_HLIST_HEAD(&tcp_hashinfo.bhash[i].chain);
2464 /* Try to be a bit smarter and adjust defaults depending
2465 * on available memory.
2467 for (order = 0; ((1 << order) << PAGE_SHIFT) <
2468 (tcp_hashinfo.bhash_size * sizeof(struct inet_bind_hashbucket));
2472 sysctl_local_port_range[0] = 32768;
2473 sysctl_local_port_range[1] = 61000;
2474 tcp_death_row.sysctl_max_tw_buckets = 180000;
2475 sysctl_tcp_max_orphans = 4096 << (order - 4);
2476 sysctl_max_syn_backlog = 1024;
2477 } else if (order < 3) {
2478 sysctl_local_port_range[0] = 1024 * (3 - order);
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);