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).
9 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
10 * Mark Evans, <evansmp@uhura.aston.ac.uk>
11 * Corey Minyard <wf-rch!minyard@relay.EU.net>
12 * Florian La Roche, <flla@stud.uni-sb.de>
13 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
14 * Linus Torvalds, <torvalds@cs.helsinki.fi>
15 * Alan Cox, <gw4pts@gw4pts.ampr.org>
16 * Matthew Dillon, <dillon@apollo.west.oic.com>
17 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
18 * Jorge Cwik, <jorge@laser.satlink.net>
21 * Alan Cox : Numerous verify_area() calls
22 * Alan Cox : Set the ACK bit on a reset
23 * Alan Cox : Stopped it crashing if it closed while
24 * sk->inuse=1 and was trying to connect
26 * Alan Cox : All icmp error handling was broken
27 * pointers passed where wrong and the
28 * socket was looked up backwards. Nobody
29 * tested any icmp error code obviously.
30 * Alan Cox : tcp_err() now handled properly. It
31 * wakes people on errors. poll
32 * behaves and the icmp error race
33 * has gone by moving it into sock.c
34 * Alan Cox : tcp_send_reset() fixed to work for
35 * everything not just packets for
37 * Alan Cox : tcp option processing.
38 * Alan Cox : Reset tweaked (still not 100%) [Had
40 * Herp Rosmanith : More reset fixes
41 * Alan Cox : No longer acks invalid rst frames.
42 * Acking any kind of RST is right out.
43 * Alan Cox : Sets an ignore me flag on an rst
44 * receive otherwise odd bits of prattle
46 * Alan Cox : Fixed another acking RST frame bug.
47 * Should stop LAN workplace lockups.
48 * Alan Cox : Some tidyups using the new skb list
50 * Alan Cox : sk->keepopen now seems to work
51 * Alan Cox : Pulls options out correctly on accepts
52 * Alan Cox : Fixed assorted sk->rqueue->next errors
53 * Alan Cox : PSH doesn't end a TCP read. Switched a
55 * Alan Cox : Tidied tcp_data to avoid a potential
57 * Alan Cox : Added some better commenting, as the
58 * tcp is hard to follow
59 * Alan Cox : Removed incorrect check for 20 * psh
60 * Michael O'Reilly : ack < copied bug fix.
61 * Johannes Stille : Misc tcp fixes (not all in yet).
62 * Alan Cox : FIN with no memory -> CRASH
63 * Alan Cox : Added socket option proto entries.
64 * Also added awareness of them to accept.
65 * Alan Cox : Added TCP options (SOL_TCP)
66 * Alan Cox : Switched wakeup calls to callbacks,
67 * so the kernel can layer network
69 * Alan Cox : Use ip_tos/ip_ttl settings.
70 * Alan Cox : Handle FIN (more) properly (we hope).
71 * Alan Cox : RST frames sent on unsynchronised
73 * Alan Cox : Put in missing check for SYN bit.
74 * Alan Cox : Added tcp_select_window() aka NET2E
75 * window non shrink trick.
76 * Alan Cox : Added a couple of small NET2E timer
78 * Charles Hedrick : TCP fixes
79 * Toomas Tamm : TCP window fixes
80 * Alan Cox : Small URG fix to rlogin ^C ack fight
81 * Charles Hedrick : Rewrote most of it to actually work
82 * Linus : Rewrote tcp_read() and URG handling
84 * Gerhard Koerting: Fixed some missing timer handling
85 * Matthew Dillon : Reworked TCP machine states as per RFC
86 * Gerhard Koerting: PC/TCP workarounds
87 * Adam Caldwell : Assorted timer/timing errors
88 * Matthew Dillon : Fixed another RST bug
89 * Alan Cox : Move to kernel side addressing changes.
90 * Alan Cox : Beginning work on TCP fastpathing
92 * Arnt Gulbrandsen: Turbocharged tcp_check() routine.
93 * Alan Cox : TCP fast path debugging
94 * Alan Cox : Window clamping
95 * Michael Riepe : Bug in tcp_check()
96 * Matt Dillon : More TCP improvements and RST bug fixes
97 * Matt Dillon : Yet more small nasties remove from the
98 * TCP code (Be very nice to this man if
99 * tcp finally works 100%) 8)
100 * Alan Cox : BSD accept semantics.
101 * Alan Cox : Reset on closedown bug.
102 * Peter De Schrijver : ENOTCONN check missing in tcp_sendto().
103 * Michael Pall : Handle poll() after URG properly in
105 * Michael Pall : Undo the last fix in tcp_read_urg()
106 * (multi URG PUSH broke rlogin).
107 * Michael Pall : Fix the multi URG PUSH problem in
108 * tcp_readable(), poll() after URG
110 * Michael Pall : recv(...,MSG_OOB) never blocks in the
112 * Alan Cox : Changed the semantics of sk->socket to
113 * fix a race and a signal problem with
114 * accept() and async I/O.
115 * Alan Cox : Relaxed the rules on tcp_sendto().
116 * Yury Shevchuk : Really fixed accept() blocking problem.
117 * Craig I. Hagan : Allow for BSD compatible TIME_WAIT for
118 * clients/servers which listen in on
120 * Alan Cox : Cleaned the above up and shrank it to
121 * a sensible code size.
122 * Alan Cox : Self connect lockup fix.
123 * Alan Cox : No connect to multicast.
124 * Ross Biro : Close unaccepted children on master
126 * Alan Cox : Reset tracing code.
127 * Alan Cox : Spurious resets on shutdown.
128 * Alan Cox : Giant 15 minute/60 second timer error
129 * Alan Cox : Small whoops in polling before an
131 * Alan Cox : Kept the state trace facility since
132 * it's handy for debugging.
133 * Alan Cox : More reset handler fixes.
134 * Alan Cox : Started rewriting the code based on
135 * the RFC's for other useful protocol
136 * references see: Comer, KA9Q NOS, and
137 * for a reference on the difference
138 * between specifications and how BSD
139 * works see the 4.4lite source.
140 * A.N.Kuznetsov : Don't time wait on completion of tidy
142 * Linus Torvalds : Fin/Shutdown & copied_seq changes.
143 * Linus Torvalds : Fixed BSD port reuse to work first syn
144 * Alan Cox : Reimplemented timers as per the RFC
145 * and using multiple timers for sanity.
146 * Alan Cox : Small bug fixes, and a lot of new
148 * Alan Cox : Fixed dual reader crash by locking
149 * the buffers (much like datagram.c)
150 * Alan Cox : Fixed stuck sockets in probe. A probe
151 * now gets fed up of retrying without
152 * (even a no space) answer.
153 * Alan Cox : Extracted closing code better
154 * Alan Cox : Fixed the closing state machine to
156 * Alan Cox : More 'per spec' fixes.
157 * Jorge Cwik : Even faster checksumming.
158 * Alan Cox : tcp_data() doesn't ack illegal PSH
159 * only frames. At least one pc tcp stack
161 * Alan Cox : Cache last socket.
162 * Alan Cox : Per route irtt.
163 * Matt Day : poll()->select() match BSD precisely on error
164 * Alan Cox : New buffers
165 * Marc Tamsky : Various sk->prot->retransmits and
166 * sk->retransmits misupdating fixed.
167 * Fixed tcp_write_timeout: stuck close,
168 * and TCP syn retries gets used now.
169 * Mark Yarvis : In tcp_read_wakeup(), don't send an
170 * ack if state is TCP_CLOSED.
171 * Alan Cox : Look up device on a retransmit - routes may
172 * change. Doesn't yet cope with MSS shrink right
174 * Marc Tamsky : Closing in closing fixes.
175 * Mike Shaver : RFC1122 verifications.
176 * Alan Cox : rcv_saddr errors.
177 * Alan Cox : Block double connect().
178 * Alan Cox : Small hooks for enSKIP.
179 * Alexey Kuznetsov: Path MTU discovery.
180 * Alan Cox : Support soft errors.
181 * Alan Cox : Fix MTU discovery pathological case
182 * when the remote claims no mtu!
183 * Marc Tamsky : TCP_CLOSE fix.
184 * Colin (G3TNE) : Send a reset on syn ack replies in
185 * window but wrong (fixes NT lpd problems)
186 * Pedro Roque : Better TCP window handling, delayed ack.
187 * Joerg Reuter : No modification of locked buffers in
188 * tcp_do_retransmit()
189 * Eric Schenk : Changed receiver side silly window
190 * avoidance algorithm to BSD style
191 * algorithm. This doubles throughput
192 * against machines running Solaris,
193 * and seems to result in general
195 * Stefan Magdalinski : adjusted tcp_readable() to fix FIONREAD
196 * Willy Konynenberg : Transparent proxying support.
197 * Mike McLagan : Routing by source
198 * Keith Owens : Do proper merging with partial SKB's in
199 * tcp_do_sendmsg to avoid burstiness.
200 * Eric Schenk : Fix fast close down bug with
201 * shutdown() followed by close().
202 * Andi Kleen : Make poll agree with SIGIO
203 * Salvatore Sanfilippo : Support SO_LINGER with linger == 1 and
204 * lingertime == 0 (RFC 793 ABORT Call)
205 * Hirokazu Takahashi : Use copy_from_user() instead of
206 * csum_and_copy_from_user() if possible.
208 * This program is free software; you can redistribute it and/or
209 * modify it under the terms of the GNU General Public License
210 * as published by the Free Software Foundation; either version
211 * 2 of the License, or(at your option) any later version.
213 * Description of States:
215 * TCP_SYN_SENT sent a connection request, waiting for ack
217 * TCP_SYN_RECV received a connection request, sent ack,
218 * waiting for final ack in three-way handshake.
220 * TCP_ESTABLISHED connection established
222 * TCP_FIN_WAIT1 our side has shutdown, waiting to complete
223 * transmission of remaining buffered data
225 * TCP_FIN_WAIT2 all buffered data sent, waiting for remote
228 * TCP_CLOSING both sides have shutdown but we still have
229 * data we have to finish sending
231 * TCP_TIME_WAIT timeout to catch resent junk before entering
232 * closed, can only be entered from FIN_WAIT2
233 * or CLOSING. Required because the other end
234 * may not have gotten our last ACK causing it
235 * to retransmit the data packet (which we ignore)
237 * TCP_CLOSE_WAIT remote side has shutdown and is waiting for
238 * us to finish writing our data and to shutdown
239 * (we have to close() to move on to LAST_ACK)
241 * TCP_LAST_ACK out side has shutdown after remote has
242 * shutdown. There may still be data in our
243 * buffer that we have to finish sending
245 * TCP_CLOSE socket is finished
248 #include <linux/kernel.h>
249 #include <linux/module.h>
250 #include <linux/types.h>
251 #include <linux/fcntl.h>
252 #include <linux/poll.h>
253 #include <linux/init.h>
254 #include <linux/fs.h>
255 #include <linux/skbuff.h>
256 #include <linux/scatterlist.h>
257 #include <linux/splice.h>
258 #include <linux/net.h>
259 #include <linux/socket.h>
260 #include <linux/random.h>
261 #include <linux/bootmem.h>
262 #include <linux/highmem.h>
263 #include <linux/swap.h>
264 #include <linux/cache.h>
265 #include <linux/err.h>
266 #include <linux/crypto.h>
268 #include <net/icmp.h>
270 #include <net/xfrm.h>
272 #include <net/netdma.h>
273 #include <net/sock.h>
275 #include <asm/uaccess.h>
276 #include <asm/ioctls.h>
278 int sysctl_tcp_fin_timeout __read_mostly = TCP_FIN_TIMEOUT;
280 struct percpu_counter tcp_orphan_count;
281 EXPORT_SYMBOL_GPL(tcp_orphan_count);
283 int sysctl_tcp_mem[3] __read_mostly;
284 int sysctl_tcp_wmem[3] __read_mostly;
285 int sysctl_tcp_rmem[3] __read_mostly;
287 EXPORT_SYMBOL(sysctl_tcp_mem);
288 EXPORT_SYMBOL(sysctl_tcp_rmem);
289 EXPORT_SYMBOL(sysctl_tcp_wmem);
291 atomic_t tcp_memory_allocated; /* Current allocated memory. */
292 EXPORT_SYMBOL(tcp_memory_allocated);
295 * Current number of TCP sockets.
297 struct percpu_counter tcp_sockets_allocated;
298 EXPORT_SYMBOL(tcp_sockets_allocated);
303 struct tcp_splice_state {
304 struct pipe_inode_info *pipe;
310 * Pressure flag: try to collapse.
311 * Technical note: it is used by multiple contexts non atomically.
312 * All the __sk_mem_schedule() is of this nature: accounting
313 * is strict, actions are advisory and have some latency.
315 int tcp_memory_pressure __read_mostly;
317 EXPORT_SYMBOL(tcp_memory_pressure);
319 void tcp_enter_memory_pressure(struct sock *sk)
321 if (!tcp_memory_pressure) {
322 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMEMORYPRESSURES);
323 tcp_memory_pressure = 1;
327 EXPORT_SYMBOL(tcp_enter_memory_pressure);
329 /* Convert seconds to retransmits based on initial and max timeout */
330 static u8 secs_to_retrans(int seconds, int timeout, int rto_max)
335 int period = timeout;
338 while (seconds > period && res < 255) {
341 if (timeout > rto_max)
349 /* Convert retransmits to seconds based on initial and max timeout */
350 static int retrans_to_secs(u8 retrans, int timeout, int rto_max)
358 if (timeout > rto_max)
367 * Wait for a TCP event.
369 * Note that we don't need to lock the socket, as the upper poll layers
370 * take care of normal races (between the test and the event) and we don't
371 * go look at any of the socket buffers directly.
373 unsigned int tcp_poll(struct file *file, struct socket *sock, poll_table *wait)
376 struct sock *sk = sock->sk;
377 struct tcp_sock *tp = tcp_sk(sk);
379 sock_poll_wait(file, sk->sk_sleep, wait);
380 if (sk->sk_state == TCP_LISTEN)
381 return inet_csk_listen_poll(sk);
383 /* Socket is not locked. We are protected from async events
384 * by poll logic and correct handling of state changes
385 * made by other threads is impossible in any case.
393 * POLLHUP is certainly not done right. But poll() doesn't
394 * have a notion of HUP in just one direction, and for a
395 * socket the read side is more interesting.
397 * Some poll() documentation says that POLLHUP is incompatible
398 * with the POLLOUT/POLLWR flags, so somebody should check this
399 * all. But careful, it tends to be safer to return too many
400 * bits than too few, and you can easily break real applications
401 * if you don't tell them that something has hung up!
405 * Check number 1. POLLHUP is _UNMASKABLE_ event (see UNIX98 and
406 * our fs/select.c). It means that after we received EOF,
407 * poll always returns immediately, making impossible poll() on write()
408 * in state CLOSE_WAIT. One solution is evident --- to set POLLHUP
409 * if and only if shutdown has been made in both directions.
410 * Actually, it is interesting to look how Solaris and DUX
411 * solve this dilemma. I would prefer, if POLLHUP were maskable,
412 * then we could set it on SND_SHUTDOWN. BTW examples given
413 * in Stevens' books assume exactly this behaviour, it explains
414 * why POLLHUP is incompatible with POLLOUT. --ANK
416 * NOTE. Check for TCP_CLOSE is added. The goal is to prevent
417 * blocking on fresh not-connected or disconnected socket. --ANK
419 if (sk->sk_shutdown == SHUTDOWN_MASK || sk->sk_state == TCP_CLOSE)
421 if (sk->sk_shutdown & RCV_SHUTDOWN)
422 mask |= POLLIN | POLLRDNORM | POLLRDHUP;
425 if ((1 << sk->sk_state) & ~(TCPF_SYN_SENT | TCPF_SYN_RECV)) {
426 int target = sock_rcvlowat(sk, 0, INT_MAX);
428 if (tp->urg_seq == tp->copied_seq &&
429 !sock_flag(sk, SOCK_URGINLINE) &&
433 /* Potential race condition. If read of tp below will
434 * escape above sk->sk_state, we can be illegally awaken
435 * in SYN_* states. */
436 if (tp->rcv_nxt - tp->copied_seq >= target)
437 mask |= POLLIN | POLLRDNORM;
439 if (!(sk->sk_shutdown & SEND_SHUTDOWN)) {
440 if (sk_stream_wspace(sk) >= sk_stream_min_wspace(sk)) {
441 mask |= POLLOUT | POLLWRNORM;
442 } else { /* send SIGIO later */
443 set_bit(SOCK_ASYNC_NOSPACE,
444 &sk->sk_socket->flags);
445 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
447 /* Race breaker. If space is freed after
448 * wspace test but before the flags are set,
449 * IO signal will be lost.
451 if (sk_stream_wspace(sk) >= sk_stream_min_wspace(sk))
452 mask |= POLLOUT | POLLWRNORM;
456 if (tp->urg_data & TCP_URG_VALID)
462 int tcp_ioctl(struct sock *sk, int cmd, unsigned long arg)
464 struct tcp_sock *tp = tcp_sk(sk);
469 if (sk->sk_state == TCP_LISTEN)
473 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
475 else if (sock_flag(sk, SOCK_URGINLINE) ||
477 before(tp->urg_seq, tp->copied_seq) ||
478 !before(tp->urg_seq, tp->rcv_nxt)) {
481 answ = tp->rcv_nxt - tp->copied_seq;
483 /* Subtract 1, if FIN is in queue. */
484 skb = skb_peek_tail(&sk->sk_receive_queue);
486 answ -= tcp_hdr(skb)->fin;
488 answ = tp->urg_seq - tp->copied_seq;
492 answ = tp->urg_data && tp->urg_seq == tp->copied_seq;
495 if (sk->sk_state == TCP_LISTEN)
498 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
501 answ = tp->write_seq - tp->snd_una;
507 return put_user(answ, (int __user *)arg);
510 static inline void tcp_mark_push(struct tcp_sock *tp, struct sk_buff *skb)
512 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH;
513 tp->pushed_seq = tp->write_seq;
516 static inline int forced_push(struct tcp_sock *tp)
518 return after(tp->write_seq, tp->pushed_seq + (tp->max_window >> 1));
521 static inline void skb_entail(struct sock *sk, struct sk_buff *skb)
523 struct tcp_sock *tp = tcp_sk(sk);
524 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
527 tcb->seq = tcb->end_seq = tp->write_seq;
528 tcb->flags = TCPCB_FLAG_ACK;
530 skb_header_release(skb);
531 tcp_add_write_queue_tail(sk, skb);
532 sk->sk_wmem_queued += skb->truesize;
533 sk_mem_charge(sk, skb->truesize);
534 if (tp->nonagle & TCP_NAGLE_PUSH)
535 tp->nonagle &= ~TCP_NAGLE_PUSH;
538 static inline void tcp_mark_urg(struct tcp_sock *tp, int flags,
542 tp->snd_up = tp->write_seq;
545 static inline void tcp_push(struct sock *sk, int flags, int mss_now,
548 struct tcp_sock *tp = tcp_sk(sk);
550 if (tcp_send_head(sk)) {
551 struct sk_buff *skb = tcp_write_queue_tail(sk);
552 if (!(flags & MSG_MORE) || forced_push(tp))
553 tcp_mark_push(tp, skb);
554 tcp_mark_urg(tp, flags, skb);
555 __tcp_push_pending_frames(sk, mss_now,
556 (flags & MSG_MORE) ? TCP_NAGLE_CORK : nonagle);
560 static int tcp_splice_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb,
561 unsigned int offset, size_t len)
563 struct tcp_splice_state *tss = rd_desc->arg.data;
566 ret = skb_splice_bits(skb, offset, tss->pipe, min(rd_desc->count, len),
569 rd_desc->count -= ret;
573 static int __tcp_splice_read(struct sock *sk, struct tcp_splice_state *tss)
575 /* Store TCP splice context information in read_descriptor_t. */
576 read_descriptor_t rd_desc = {
581 return tcp_read_sock(sk, &rd_desc, tcp_splice_data_recv);
585 * tcp_splice_read - splice data from TCP socket to a pipe
586 * @sock: socket to splice from
587 * @ppos: position (not valid)
588 * @pipe: pipe to splice to
589 * @len: number of bytes to splice
590 * @flags: splice modifier flags
593 * Will read pages from given socket and fill them into a pipe.
596 ssize_t tcp_splice_read(struct socket *sock, loff_t *ppos,
597 struct pipe_inode_info *pipe, size_t len,
600 struct sock *sk = sock->sk;
601 struct tcp_splice_state tss = {
611 * We can't seek on a socket input
620 timeo = sock_rcvtimeo(sk, sock->file->f_flags & O_NONBLOCK);
622 ret = __tcp_splice_read(sk, &tss);
628 if (sock_flag(sk, SOCK_DONE))
631 ret = sock_error(sk);
634 if (sk->sk_shutdown & RCV_SHUTDOWN)
636 if (sk->sk_state == TCP_CLOSE) {
638 * This occurs when user tries to read
639 * from never connected socket.
641 if (!sock_flag(sk, SOCK_DONE))
649 sk_wait_data(sk, &timeo);
650 if (signal_pending(current)) {
651 ret = sock_intr_errno(timeo);
664 if (sk->sk_err || sk->sk_state == TCP_CLOSE ||
665 (sk->sk_shutdown & RCV_SHUTDOWN) ||
666 signal_pending(current))
678 struct sk_buff *sk_stream_alloc_skb(struct sock *sk, int size, gfp_t gfp)
682 /* The TCP header must be at least 32-bit aligned. */
683 size = ALIGN(size, 4);
685 skb = alloc_skb_fclone(size + sk->sk_prot->max_header, gfp);
687 if (sk_wmem_schedule(sk, skb->truesize)) {
689 * Make sure that we have exactly size bytes
690 * available to the caller, no more, no less.
692 skb_reserve(skb, skb_tailroom(skb) - size);
697 sk->sk_prot->enter_memory_pressure(sk);
698 sk_stream_moderate_sndbuf(sk);
703 static unsigned int tcp_xmit_size_goal(struct sock *sk, u32 mss_now,
706 struct tcp_sock *tp = tcp_sk(sk);
707 u32 xmit_size_goal, old_size_goal;
709 xmit_size_goal = mss_now;
711 if (large_allowed && sk_can_gso(sk)) {
712 xmit_size_goal = ((sk->sk_gso_max_size - 1) -
713 inet_csk(sk)->icsk_af_ops->net_header_len -
714 inet_csk(sk)->icsk_ext_hdr_len -
717 xmit_size_goal = tcp_bound_to_half_wnd(tp, xmit_size_goal);
719 /* We try hard to avoid divides here */
720 old_size_goal = tp->xmit_size_goal_segs * mss_now;
722 if (likely(old_size_goal <= xmit_size_goal &&
723 old_size_goal + mss_now > xmit_size_goal)) {
724 xmit_size_goal = old_size_goal;
726 tp->xmit_size_goal_segs = xmit_size_goal / mss_now;
727 xmit_size_goal = tp->xmit_size_goal_segs * mss_now;
731 return max(xmit_size_goal, mss_now);
734 static int tcp_send_mss(struct sock *sk, int *size_goal, int flags)
738 mss_now = tcp_current_mss(sk);
739 *size_goal = tcp_xmit_size_goal(sk, mss_now, !(flags & MSG_OOB));
744 static ssize_t do_tcp_sendpages(struct sock *sk, struct page **pages, int poffset,
745 size_t psize, int flags)
747 struct tcp_sock *tp = tcp_sk(sk);
748 int mss_now, size_goal;
751 long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
753 /* Wait for a connection to finish. */
754 if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
755 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
758 clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
760 mss_now = tcp_send_mss(sk, &size_goal, flags);
764 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
768 struct sk_buff *skb = tcp_write_queue_tail(sk);
769 struct page *page = pages[poffset / PAGE_SIZE];
770 int copy, i, can_coalesce;
771 int offset = poffset % PAGE_SIZE;
772 int size = min_t(size_t, psize, PAGE_SIZE - offset);
774 if (!tcp_send_head(sk) || (copy = size_goal - skb->len) <= 0) {
776 if (!sk_stream_memory_free(sk))
777 goto wait_for_sndbuf;
779 skb = sk_stream_alloc_skb(sk, 0, sk->sk_allocation);
781 goto wait_for_memory;
790 i = skb_shinfo(skb)->nr_frags;
791 can_coalesce = skb_can_coalesce(skb, i, page, offset);
792 if (!can_coalesce && i >= MAX_SKB_FRAGS) {
793 tcp_mark_push(tp, skb);
796 if (!sk_wmem_schedule(sk, copy))
797 goto wait_for_memory;
800 skb_shinfo(skb)->frags[i - 1].size += copy;
803 skb_fill_page_desc(skb, i, page, offset, copy);
807 skb->data_len += copy;
808 skb->truesize += copy;
809 sk->sk_wmem_queued += copy;
810 sk_mem_charge(sk, copy);
811 skb->ip_summed = CHECKSUM_PARTIAL;
812 tp->write_seq += copy;
813 TCP_SKB_CB(skb)->end_seq += copy;
814 skb_shinfo(skb)->gso_segs = 0;
817 TCP_SKB_CB(skb)->flags &= ~TCPCB_FLAG_PSH;
821 if (!(psize -= copy))
824 if (skb->len < size_goal || (flags & MSG_OOB))
827 if (forced_push(tp)) {
828 tcp_mark_push(tp, skb);
829 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
830 } else if (skb == tcp_send_head(sk))
831 tcp_push_one(sk, mss_now);
835 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
838 tcp_push(sk, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH);
840 if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
843 mss_now = tcp_send_mss(sk, &size_goal, flags);
848 tcp_push(sk, flags, mss_now, tp->nonagle);
855 return sk_stream_error(sk, flags, err);
858 ssize_t tcp_sendpage(struct socket *sock, struct page *page, int offset,
859 size_t size, int flags)
862 struct sock *sk = sock->sk;
864 if (!(sk->sk_route_caps & NETIF_F_SG) ||
865 !(sk->sk_route_caps & NETIF_F_ALL_CSUM))
866 return sock_no_sendpage(sock, page, offset, size, flags);
870 res = do_tcp_sendpages(sk, &page, offset, size, flags);
876 #define TCP_PAGE(sk) (sk->sk_sndmsg_page)
877 #define TCP_OFF(sk) (sk->sk_sndmsg_off)
879 static inline int select_size(struct sock *sk)
881 struct tcp_sock *tp = tcp_sk(sk);
882 int tmp = tp->mss_cache;
884 if (sk->sk_route_caps & NETIF_F_SG) {
888 int pgbreak = SKB_MAX_HEAD(MAX_TCP_HEADER);
890 if (tmp >= pgbreak &&
891 tmp <= pgbreak + (MAX_SKB_FRAGS - 1) * PAGE_SIZE)
899 int tcp_sendmsg(struct kiocb *iocb, struct socket *sock, struct msghdr *msg,
902 struct sock *sk = sock->sk;
904 struct tcp_sock *tp = tcp_sk(sk);
907 int mss_now, size_goal;
914 flags = msg->msg_flags;
915 timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
917 /* Wait for a connection to finish. */
918 if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
919 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
922 /* This should be in poll */
923 clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
925 mss_now = tcp_send_mss(sk, &size_goal, flags);
927 /* Ok commence sending. */
928 iovlen = msg->msg_iovlen;
933 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
936 while (--iovlen >= 0) {
937 int seglen = iov->iov_len;
938 unsigned char __user *from = iov->iov_base;
946 skb = tcp_write_queue_tail(sk);
947 if (tcp_send_head(sk)) {
948 if (skb->ip_summed == CHECKSUM_NONE)
950 copy = max - skb->len;
955 /* Allocate new segment. If the interface is SG,
956 * allocate skb fitting to single page.
958 if (!sk_stream_memory_free(sk))
959 goto wait_for_sndbuf;
961 skb = sk_stream_alloc_skb(sk, select_size(sk),
964 goto wait_for_memory;
967 * Check whether we can use HW checksum.
969 if (sk->sk_route_caps & NETIF_F_ALL_CSUM)
970 skb->ip_summed = CHECKSUM_PARTIAL;
977 /* Try to append data to the end of skb. */
981 /* Where to copy to? */
982 if (skb_tailroom(skb) > 0) {
983 /* We have some space in skb head. Superb! */
984 if (copy > skb_tailroom(skb))
985 copy = skb_tailroom(skb);
986 if ((err = skb_add_data(skb, from, copy)) != 0)
990 int i = skb_shinfo(skb)->nr_frags;
991 struct page *page = TCP_PAGE(sk);
992 int off = TCP_OFF(sk);
994 if (skb_can_coalesce(skb, i, page, off) &&
996 /* We can extend the last page
999 } else if (i == MAX_SKB_FRAGS ||
1001 !(sk->sk_route_caps & NETIF_F_SG))) {
1002 /* Need to add new fragment and cannot
1003 * do this because interface is non-SG,
1004 * or because all the page slots are
1006 tcp_mark_push(tp, skb);
1009 if (off == PAGE_SIZE) {
1011 TCP_PAGE(sk) = page = NULL;
1017 if (copy > PAGE_SIZE - off)
1018 copy = PAGE_SIZE - off;
1020 if (!sk_wmem_schedule(sk, copy))
1021 goto wait_for_memory;
1024 /* Allocate new cache page. */
1025 if (!(page = sk_stream_alloc_page(sk)))
1026 goto wait_for_memory;
1029 /* Time to copy data. We are close to
1031 err = skb_copy_to_page(sk, from, skb, page,
1034 /* If this page was new, give it to the
1035 * socket so it does not get leaked.
1037 if (!TCP_PAGE(sk)) {
1038 TCP_PAGE(sk) = page;
1044 /* Update the skb. */
1046 skb_shinfo(skb)->frags[i - 1].size +=
1049 skb_fill_page_desc(skb, i, page, off, copy);
1052 } else if (off + copy < PAGE_SIZE) {
1054 TCP_PAGE(sk) = page;
1058 TCP_OFF(sk) = off + copy;
1062 TCP_SKB_CB(skb)->flags &= ~TCPCB_FLAG_PSH;
1064 tp->write_seq += copy;
1065 TCP_SKB_CB(skb)->end_seq += copy;
1066 skb_shinfo(skb)->gso_segs = 0;
1070 if ((seglen -= copy) == 0 && iovlen == 0)
1073 if (skb->len < max || (flags & MSG_OOB))
1076 if (forced_push(tp)) {
1077 tcp_mark_push(tp, skb);
1078 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
1079 } else if (skb == tcp_send_head(sk))
1080 tcp_push_one(sk, mss_now);
1084 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1087 tcp_push(sk, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH);
1089 if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
1092 mss_now = tcp_send_mss(sk, &size_goal, flags);
1098 tcp_push(sk, flags, mss_now, tp->nonagle);
1099 TCP_CHECK_TIMER(sk);
1105 tcp_unlink_write_queue(skb, sk);
1106 /* It is the one place in all of TCP, except connection
1107 * reset, where we can be unlinking the send_head.
1109 tcp_check_send_head(sk, skb);
1110 sk_wmem_free_skb(sk, skb);
1117 err = sk_stream_error(sk, flags, err);
1118 TCP_CHECK_TIMER(sk);
1124 * Handle reading urgent data. BSD has very simple semantics for
1125 * this, no blocking and very strange errors 8)
1128 static int tcp_recv_urg(struct sock *sk, struct msghdr *msg, int len, int flags)
1130 struct tcp_sock *tp = tcp_sk(sk);
1132 /* No URG data to read. */
1133 if (sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data ||
1134 tp->urg_data == TCP_URG_READ)
1135 return -EINVAL; /* Yes this is right ! */
1137 if (sk->sk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DONE))
1140 if (tp->urg_data & TCP_URG_VALID) {
1142 char c = tp->urg_data;
1144 if (!(flags & MSG_PEEK))
1145 tp->urg_data = TCP_URG_READ;
1147 /* Read urgent data. */
1148 msg->msg_flags |= MSG_OOB;
1151 if (!(flags & MSG_TRUNC))
1152 err = memcpy_toiovec(msg->msg_iov, &c, 1);
1155 msg->msg_flags |= MSG_TRUNC;
1157 return err ? -EFAULT : len;
1160 if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
1163 /* Fixed the recv(..., MSG_OOB) behaviour. BSD docs and
1164 * the available implementations agree in this case:
1165 * this call should never block, independent of the
1166 * blocking state of the socket.
1167 * Mike <pall@rz.uni-karlsruhe.de>
1172 /* Clean up the receive buffer for full frames taken by the user,
1173 * then send an ACK if necessary. COPIED is the number of bytes
1174 * tcp_recvmsg has given to the user so far, it speeds up the
1175 * calculation of whether or not we must ACK for the sake of
1178 void tcp_cleanup_rbuf(struct sock *sk, int copied)
1180 struct tcp_sock *tp = tcp_sk(sk);
1181 int time_to_ack = 0;
1184 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
1186 WARN_ON(skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq));
1189 if (inet_csk_ack_scheduled(sk)) {
1190 const struct inet_connection_sock *icsk = inet_csk(sk);
1191 /* Delayed ACKs frequently hit locked sockets during bulk
1193 if (icsk->icsk_ack.blocked ||
1194 /* Once-per-two-segments ACK was not sent by tcp_input.c */
1195 tp->rcv_nxt - tp->rcv_wup > icsk->icsk_ack.rcv_mss ||
1197 * If this read emptied read buffer, we send ACK, if
1198 * connection is not bidirectional, user drained
1199 * receive buffer and there was a small segment
1203 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED2) ||
1204 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED) &&
1205 !icsk->icsk_ack.pingpong)) &&
1206 !atomic_read(&sk->sk_rmem_alloc)))
1210 /* We send an ACK if we can now advertise a non-zero window
1211 * which has been raised "significantly".
1213 * Even if window raised up to infinity, do not send window open ACK
1214 * in states, where we will not receive more. It is useless.
1216 if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) {
1217 __u32 rcv_window_now = tcp_receive_window(tp);
1219 /* Optimize, __tcp_select_window() is not cheap. */
1220 if (2*rcv_window_now <= tp->window_clamp) {
1221 __u32 new_window = __tcp_select_window(sk);
1223 /* Send ACK now, if this read freed lots of space
1224 * in our buffer. Certainly, new_window is new window.
1225 * We can advertise it now, if it is not less than current one.
1226 * "Lots" means "at least twice" here.
1228 if (new_window && new_window >= 2 * rcv_window_now)
1236 static void tcp_prequeue_process(struct sock *sk)
1238 struct sk_buff *skb;
1239 struct tcp_sock *tp = tcp_sk(sk);
1241 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPPREQUEUED);
1243 /* RX process wants to run with disabled BHs, though it is not
1246 while ((skb = __skb_dequeue(&tp->ucopy.prequeue)) != NULL)
1247 sk_backlog_rcv(sk, skb);
1250 /* Clear memory counter. */
1251 tp->ucopy.memory = 0;
1254 static inline struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off)
1256 struct sk_buff *skb;
1259 skb_queue_walk(&sk->sk_receive_queue, skb) {
1260 offset = seq - TCP_SKB_CB(skb)->seq;
1261 if (tcp_hdr(skb)->syn)
1263 if (offset < skb->len || tcp_hdr(skb)->fin) {
1272 * This routine provides an alternative to tcp_recvmsg() for routines
1273 * that would like to handle copying from skbuffs directly in 'sendfile'
1276 * - It is assumed that the socket was locked by the caller.
1277 * - The routine does not block.
1278 * - At present, there is no support for reading OOB data
1279 * or for 'peeking' the socket using this routine
1280 * (although both would be easy to implement).
1282 int tcp_read_sock(struct sock *sk, read_descriptor_t *desc,
1283 sk_read_actor_t recv_actor)
1285 struct sk_buff *skb;
1286 struct tcp_sock *tp = tcp_sk(sk);
1287 u32 seq = tp->copied_seq;
1291 if (sk->sk_state == TCP_LISTEN)
1293 while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1294 if (offset < skb->len) {
1298 len = skb->len - offset;
1299 /* Stop reading if we hit a patch of urgent data */
1301 u32 urg_offset = tp->urg_seq - seq;
1302 if (urg_offset < len)
1307 used = recv_actor(desc, skb, offset, len);
1312 } else if (used <= len) {
1318 * If recv_actor drops the lock (e.g. TCP splice
1319 * receive) the skb pointer might be invalid when
1320 * getting here: tcp_collapse might have deleted it
1321 * while aggregating skbs from the socket queue.
1323 skb = tcp_recv_skb(sk, seq-1, &offset);
1324 if (!skb || (offset+1 != skb->len))
1327 if (tcp_hdr(skb)->fin) {
1328 sk_eat_skb(sk, skb, 0);
1332 sk_eat_skb(sk, skb, 0);
1336 tp->copied_seq = seq;
1338 tcp_rcv_space_adjust(sk);
1340 /* Clean up data we have read: This will do ACK frames. */
1342 tcp_cleanup_rbuf(sk, copied);
1347 * This routine copies from a sock struct into the user buffer.
1349 * Technical note: in 2.3 we work on _locked_ socket, so that
1350 * tricks with *seq access order and skb->users are not required.
1351 * Probably, code can be easily improved even more.
1354 int tcp_recvmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg,
1355 size_t len, int nonblock, int flags, int *addr_len)
1357 struct tcp_sock *tp = tcp_sk(sk);
1363 int target; /* Read at least this many bytes */
1365 struct task_struct *user_recv = NULL;
1366 int copied_early = 0;
1367 struct sk_buff *skb;
1372 TCP_CHECK_TIMER(sk);
1375 if (sk->sk_state == TCP_LISTEN)
1378 timeo = sock_rcvtimeo(sk, nonblock);
1380 /* Urgent data needs to be handled specially. */
1381 if (flags & MSG_OOB)
1384 seq = &tp->copied_seq;
1385 if (flags & MSG_PEEK) {
1386 peek_seq = tp->copied_seq;
1390 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
1392 #ifdef CONFIG_NET_DMA
1393 tp->ucopy.dma_chan = NULL;
1395 skb = skb_peek_tail(&sk->sk_receive_queue);
1400 available = TCP_SKB_CB(skb)->seq + skb->len - (*seq);
1401 if ((available < target) &&
1402 (len > sysctl_tcp_dma_copybreak) && !(flags & MSG_PEEK) &&
1403 !sysctl_tcp_low_latency &&
1404 dma_find_channel(DMA_MEMCPY)) {
1405 preempt_enable_no_resched();
1406 tp->ucopy.pinned_list =
1407 dma_pin_iovec_pages(msg->msg_iov, len);
1409 preempt_enable_no_resched();
1417 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
1418 if (tp->urg_data && tp->urg_seq == *seq) {
1421 if (signal_pending(current)) {
1422 copied = timeo ? sock_intr_errno(timeo) : -EAGAIN;
1427 /* Next get a buffer. */
1429 skb_queue_walk(&sk->sk_receive_queue, skb) {
1430 /* Now that we have two receive queues this
1433 if (before(*seq, TCP_SKB_CB(skb)->seq)) {
1434 printk(KERN_INFO "recvmsg bug: copied %X "
1435 "seq %X\n", *seq, TCP_SKB_CB(skb)->seq);
1438 offset = *seq - TCP_SKB_CB(skb)->seq;
1439 if (tcp_hdr(skb)->syn)
1441 if (offset < skb->len)
1443 if (tcp_hdr(skb)->fin)
1445 WARN_ON(!(flags & MSG_PEEK));
1448 /* Well, if we have backlog, try to process it now yet. */
1450 if (copied >= target && !sk->sk_backlog.tail)
1455 sk->sk_state == TCP_CLOSE ||
1456 (sk->sk_shutdown & RCV_SHUTDOWN) ||
1458 signal_pending(current))
1461 if (sock_flag(sk, SOCK_DONE))
1465 copied = sock_error(sk);
1469 if (sk->sk_shutdown & RCV_SHUTDOWN)
1472 if (sk->sk_state == TCP_CLOSE) {
1473 if (!sock_flag(sk, SOCK_DONE)) {
1474 /* This occurs when user tries to read
1475 * from never connected socket.
1488 if (signal_pending(current)) {
1489 copied = sock_intr_errno(timeo);
1494 tcp_cleanup_rbuf(sk, copied);
1496 if (!sysctl_tcp_low_latency && tp->ucopy.task == user_recv) {
1497 /* Install new reader */
1498 if (!user_recv && !(flags & (MSG_TRUNC | MSG_PEEK))) {
1499 user_recv = current;
1500 tp->ucopy.task = user_recv;
1501 tp->ucopy.iov = msg->msg_iov;
1504 tp->ucopy.len = len;
1506 WARN_ON(tp->copied_seq != tp->rcv_nxt &&
1507 !(flags & (MSG_PEEK | MSG_TRUNC)));
1509 /* Ugly... If prequeue is not empty, we have to
1510 * process it before releasing socket, otherwise
1511 * order will be broken at second iteration.
1512 * More elegant solution is required!!!
1514 * Look: we have the following (pseudo)queues:
1516 * 1. packets in flight
1521 * Each queue can be processed only if the next ones
1522 * are empty. At this point we have empty receive_queue.
1523 * But prequeue _can_ be not empty after 2nd iteration,
1524 * when we jumped to start of loop because backlog
1525 * processing added something to receive_queue.
1526 * We cannot release_sock(), because backlog contains
1527 * packets arrived _after_ prequeued ones.
1529 * Shortly, algorithm is clear --- to process all
1530 * the queues in order. We could make it more directly,
1531 * requeueing packets from backlog to prequeue, if
1532 * is not empty. It is more elegant, but eats cycles,
1535 if (!skb_queue_empty(&tp->ucopy.prequeue))
1538 /* __ Set realtime policy in scheduler __ */
1541 if (copied >= target) {
1542 /* Do not sleep, just process backlog. */
1546 sk_wait_data(sk, &timeo);
1548 #ifdef CONFIG_NET_DMA
1549 tp->ucopy.wakeup = 0;
1555 /* __ Restore normal policy in scheduler __ */
1557 if ((chunk = len - tp->ucopy.len) != 0) {
1558 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMBACKLOG, chunk);
1563 if (tp->rcv_nxt == tp->copied_seq &&
1564 !skb_queue_empty(&tp->ucopy.prequeue)) {
1566 tcp_prequeue_process(sk);
1568 if ((chunk = len - tp->ucopy.len) != 0) {
1569 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1575 if ((flags & MSG_PEEK) &&
1576 (peek_seq - copied - urg_hole != tp->copied_seq)) {
1577 if (net_ratelimit())
1578 printk(KERN_DEBUG "TCP(%s:%d): Application bug, race in MSG_PEEK.\n",
1579 current->comm, task_pid_nr(current));
1580 peek_seq = tp->copied_seq;
1585 /* Ok so how much can we use? */
1586 used = skb->len - offset;
1590 /* Do we have urgent data here? */
1592 u32 urg_offset = tp->urg_seq - *seq;
1593 if (urg_offset < used) {
1595 if (!sock_flag(sk, SOCK_URGINLINE)) {
1608 if (!(flags & MSG_TRUNC)) {
1609 #ifdef CONFIG_NET_DMA
1610 if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list)
1611 tp->ucopy.dma_chan = dma_find_channel(DMA_MEMCPY);
1613 if (tp->ucopy.dma_chan) {
1614 tp->ucopy.dma_cookie = dma_skb_copy_datagram_iovec(
1615 tp->ucopy.dma_chan, skb, offset,
1617 tp->ucopy.pinned_list);
1619 if (tp->ucopy.dma_cookie < 0) {
1621 printk(KERN_ALERT "dma_cookie < 0\n");
1623 /* Exception. Bailout! */
1628 if ((offset + used) == skb->len)
1634 err = skb_copy_datagram_iovec(skb, offset,
1635 msg->msg_iov, used);
1637 /* Exception. Bailout! */
1649 tcp_rcv_space_adjust(sk);
1652 if (tp->urg_data && after(tp->copied_seq, tp->urg_seq)) {
1654 tcp_fast_path_check(sk);
1656 if (used + offset < skb->len)
1659 if (tcp_hdr(skb)->fin)
1661 if (!(flags & MSG_PEEK)) {
1662 sk_eat_skb(sk, skb, copied_early);
1668 /* Process the FIN. */
1670 if (!(flags & MSG_PEEK)) {
1671 sk_eat_skb(sk, skb, copied_early);
1678 if (!skb_queue_empty(&tp->ucopy.prequeue)) {
1681 tp->ucopy.len = copied > 0 ? len : 0;
1683 tcp_prequeue_process(sk);
1685 if (copied > 0 && (chunk = len - tp->ucopy.len) != 0) {
1686 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1692 tp->ucopy.task = NULL;
1696 #ifdef CONFIG_NET_DMA
1697 if (tp->ucopy.dma_chan) {
1698 dma_cookie_t done, used;
1700 dma_async_memcpy_issue_pending(tp->ucopy.dma_chan);
1702 while (dma_async_memcpy_complete(tp->ucopy.dma_chan,
1703 tp->ucopy.dma_cookie, &done,
1704 &used) == DMA_IN_PROGRESS) {
1705 /* do partial cleanup of sk_async_wait_queue */
1706 while ((skb = skb_peek(&sk->sk_async_wait_queue)) &&
1707 (dma_async_is_complete(skb->dma_cookie, done,
1708 used) == DMA_SUCCESS)) {
1709 __skb_dequeue(&sk->sk_async_wait_queue);
1714 /* Safe to free early-copied skbs now */
1715 __skb_queue_purge(&sk->sk_async_wait_queue);
1716 tp->ucopy.dma_chan = NULL;
1718 if (tp->ucopy.pinned_list) {
1719 dma_unpin_iovec_pages(tp->ucopy.pinned_list);
1720 tp->ucopy.pinned_list = NULL;
1724 /* According to UNIX98, msg_name/msg_namelen are ignored
1725 * on connected socket. I was just happy when found this 8) --ANK
1728 /* Clean up data we have read: This will do ACK frames. */
1729 tcp_cleanup_rbuf(sk, copied);
1731 TCP_CHECK_TIMER(sk);
1736 TCP_CHECK_TIMER(sk);
1741 err = tcp_recv_urg(sk, msg, len, flags);
1745 void tcp_set_state(struct sock *sk, int state)
1747 int oldstate = sk->sk_state;
1750 case TCP_ESTABLISHED:
1751 if (oldstate != TCP_ESTABLISHED)
1752 TCP_INC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
1756 if (oldstate == TCP_CLOSE_WAIT || oldstate == TCP_ESTABLISHED)
1757 TCP_INC_STATS(sock_net(sk), TCP_MIB_ESTABRESETS);
1759 sk->sk_prot->unhash(sk);
1760 if (inet_csk(sk)->icsk_bind_hash &&
1761 !(sk->sk_userlocks & SOCK_BINDPORT_LOCK))
1765 if (oldstate == TCP_ESTABLISHED)
1766 TCP_DEC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
1769 /* Change state AFTER socket is unhashed to avoid closed
1770 * socket sitting in hash tables.
1772 sk->sk_state = state;
1775 SOCK_DEBUG(sk, "TCP sk=%p, State %s -> %s\n", sk, statename[oldstate], statename[state]);
1778 EXPORT_SYMBOL_GPL(tcp_set_state);
1781 * State processing on a close. This implements the state shift for
1782 * sending our FIN frame. Note that we only send a FIN for some
1783 * states. A shutdown() may have already sent the FIN, or we may be
1787 static const unsigned char new_state[16] = {
1788 /* current state: new state: action: */
1789 /* (Invalid) */ TCP_CLOSE,
1790 /* TCP_ESTABLISHED */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1791 /* TCP_SYN_SENT */ TCP_CLOSE,
1792 /* TCP_SYN_RECV */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1793 /* TCP_FIN_WAIT1 */ TCP_FIN_WAIT1,
1794 /* TCP_FIN_WAIT2 */ TCP_FIN_WAIT2,
1795 /* TCP_TIME_WAIT */ TCP_CLOSE,
1796 /* TCP_CLOSE */ TCP_CLOSE,
1797 /* TCP_CLOSE_WAIT */ TCP_LAST_ACK | TCP_ACTION_FIN,
1798 /* TCP_LAST_ACK */ TCP_LAST_ACK,
1799 /* TCP_LISTEN */ TCP_CLOSE,
1800 /* TCP_CLOSING */ TCP_CLOSING,
1803 static int tcp_close_state(struct sock *sk)
1805 int next = (int)new_state[sk->sk_state];
1806 int ns = next & TCP_STATE_MASK;
1808 tcp_set_state(sk, ns);
1810 return next & TCP_ACTION_FIN;
1814 * Shutdown the sending side of a connection. Much like close except
1815 * that we don't receive shut down or sock_set_flag(sk, SOCK_DEAD).
1818 void tcp_shutdown(struct sock *sk, int how)
1820 /* We need to grab some memory, and put together a FIN,
1821 * and then put it into the queue to be sent.
1822 * Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
1824 if (!(how & SEND_SHUTDOWN))
1827 /* If we've already sent a FIN, or it's a closed state, skip this. */
1828 if ((1 << sk->sk_state) &
1829 (TCPF_ESTABLISHED | TCPF_SYN_SENT |
1830 TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) {
1831 /* Clear out any half completed packets. FIN if needed. */
1832 if (tcp_close_state(sk))
1837 void tcp_close(struct sock *sk, long timeout)
1839 struct sk_buff *skb;
1840 int data_was_unread = 0;
1844 sk->sk_shutdown = SHUTDOWN_MASK;
1846 if (sk->sk_state == TCP_LISTEN) {
1847 tcp_set_state(sk, TCP_CLOSE);
1850 inet_csk_listen_stop(sk);
1852 goto adjudge_to_death;
1855 /* We need to flush the recv. buffs. We do this only on the
1856 * descriptor close, not protocol-sourced closes, because the
1857 * reader process may not have drained the data yet!
1859 while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) {
1860 u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq -
1862 data_was_unread += len;
1868 /* As outlined in RFC 2525, section 2.17, we send a RST here because
1869 * data was lost. To witness the awful effects of the old behavior of
1870 * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk
1871 * GET in an FTP client, suspend the process, wait for the client to
1872 * advertise a zero window, then kill -9 the FTP client, wheee...
1873 * Note: timeout is always zero in such a case.
1875 if (data_was_unread) {
1876 /* Unread data was tossed, zap the connection. */
1877 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONCLOSE);
1878 tcp_set_state(sk, TCP_CLOSE);
1879 tcp_send_active_reset(sk, sk->sk_allocation);
1880 } else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) {
1881 /* Check zero linger _after_ checking for unread data. */
1882 sk->sk_prot->disconnect(sk, 0);
1883 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
1884 } else if (tcp_close_state(sk)) {
1885 /* We FIN if the application ate all the data before
1886 * zapping the connection.
1889 /* RED-PEN. Formally speaking, we have broken TCP state
1890 * machine. State transitions:
1892 * TCP_ESTABLISHED -> TCP_FIN_WAIT1
1893 * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible)
1894 * TCP_CLOSE_WAIT -> TCP_LAST_ACK
1896 * are legal only when FIN has been sent (i.e. in window),
1897 * rather than queued out of window. Purists blame.
1899 * F.e. "RFC state" is ESTABLISHED,
1900 * if Linux state is FIN-WAIT-1, but FIN is still not sent.
1902 * The visible declinations are that sometimes
1903 * we enter time-wait state, when it is not required really
1904 * (harmless), do not send active resets, when they are
1905 * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
1906 * they look as CLOSING or LAST_ACK for Linux)
1907 * Probably, I missed some more holelets.
1913 sk_stream_wait_close(sk, timeout);
1916 state = sk->sk_state;
1920 /* It is the last release_sock in its life. It will remove backlog. */
1924 /* Now socket is owned by kernel and we acquire BH lock
1925 to finish close. No need to check for user refs.
1929 WARN_ON(sock_owned_by_user(sk));
1931 percpu_counter_inc(sk->sk_prot->orphan_count);
1933 /* Have we already been destroyed by a softirq or backlog? */
1934 if (state != TCP_CLOSE && sk->sk_state == TCP_CLOSE)
1937 /* This is a (useful) BSD violating of the RFC. There is a
1938 * problem with TCP as specified in that the other end could
1939 * keep a socket open forever with no application left this end.
1940 * We use a 3 minute timeout (about the same as BSD) then kill
1941 * our end. If they send after that then tough - BUT: long enough
1942 * that we won't make the old 4*rto = almost no time - whoops
1945 * Nope, it was not mistake. It is really desired behaviour
1946 * f.e. on http servers, when such sockets are useless, but
1947 * consume significant resources. Let's do it with special
1948 * linger2 option. --ANK
1951 if (sk->sk_state == TCP_FIN_WAIT2) {
1952 struct tcp_sock *tp = tcp_sk(sk);
1953 if (tp->linger2 < 0) {
1954 tcp_set_state(sk, TCP_CLOSE);
1955 tcp_send_active_reset(sk, GFP_ATOMIC);
1956 NET_INC_STATS_BH(sock_net(sk),
1957 LINUX_MIB_TCPABORTONLINGER);
1959 const int tmo = tcp_fin_time(sk);
1961 if (tmo > TCP_TIMEWAIT_LEN) {
1962 inet_csk_reset_keepalive_timer(sk,
1963 tmo - TCP_TIMEWAIT_LEN);
1965 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
1970 if (sk->sk_state != TCP_CLOSE) {
1971 int orphan_count = percpu_counter_read_positive(
1972 sk->sk_prot->orphan_count);
1975 if (tcp_too_many_orphans(sk, orphan_count)) {
1976 if (net_ratelimit())
1977 printk(KERN_INFO "TCP: too many of orphaned "
1979 tcp_set_state(sk, TCP_CLOSE);
1980 tcp_send_active_reset(sk, GFP_ATOMIC);
1981 NET_INC_STATS_BH(sock_net(sk),
1982 LINUX_MIB_TCPABORTONMEMORY);
1986 if (sk->sk_state == TCP_CLOSE)
1987 inet_csk_destroy_sock(sk);
1988 /* Otherwise, socket is reprieved until protocol close. */
1996 /* These states need RST on ABORT according to RFC793 */
1998 static inline int tcp_need_reset(int state)
2000 return (1 << state) &
2001 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 |
2002 TCPF_FIN_WAIT2 | TCPF_SYN_RECV);
2005 int tcp_disconnect(struct sock *sk, int flags)
2007 struct inet_sock *inet = inet_sk(sk);
2008 struct inet_connection_sock *icsk = inet_csk(sk);
2009 struct tcp_sock *tp = tcp_sk(sk);
2011 int old_state = sk->sk_state;
2013 if (old_state != TCP_CLOSE)
2014 tcp_set_state(sk, TCP_CLOSE);
2016 /* ABORT function of RFC793 */
2017 if (old_state == TCP_LISTEN) {
2018 inet_csk_listen_stop(sk);
2019 } else if (tcp_need_reset(old_state) ||
2020 (tp->snd_nxt != tp->write_seq &&
2021 (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK))) {
2022 /* The last check adjusts for discrepancy of Linux wrt. RFC
2025 tcp_send_active_reset(sk, gfp_any());
2026 sk->sk_err = ECONNRESET;
2027 } else if (old_state == TCP_SYN_SENT)
2028 sk->sk_err = ECONNRESET;
2030 tcp_clear_xmit_timers(sk);
2031 __skb_queue_purge(&sk->sk_receive_queue);
2032 tcp_write_queue_purge(sk);
2033 __skb_queue_purge(&tp->out_of_order_queue);
2034 #ifdef CONFIG_NET_DMA
2035 __skb_queue_purge(&sk->sk_async_wait_queue);
2040 if (!(sk->sk_userlocks & SOCK_BINDADDR_LOCK))
2041 inet_reset_saddr(sk);
2043 sk->sk_shutdown = 0;
2044 sock_reset_flag(sk, SOCK_DONE);
2046 if ((tp->write_seq += tp->max_window + 2) == 0)
2048 icsk->icsk_backoff = 0;
2050 icsk->icsk_probes_out = 0;
2051 tp->packets_out = 0;
2052 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
2053 tp->snd_cwnd_cnt = 0;
2054 tp->bytes_acked = 0;
2055 tcp_set_ca_state(sk, TCP_CA_Open);
2056 tcp_clear_retrans(tp);
2057 inet_csk_delack_init(sk);
2058 tcp_init_send_head(sk);
2059 memset(&tp->rx_opt, 0, sizeof(tp->rx_opt));
2062 WARN_ON(inet->num && !icsk->icsk_bind_hash);
2064 sk->sk_error_report(sk);
2069 * Socket option code for TCP.
2071 static int do_tcp_setsockopt(struct sock *sk, int level,
2072 int optname, char __user *optval, unsigned int optlen)
2074 struct tcp_sock *tp = tcp_sk(sk);
2075 struct inet_connection_sock *icsk = inet_csk(sk);
2079 /* This is a string value all the others are int's */
2080 if (optname == TCP_CONGESTION) {
2081 char name[TCP_CA_NAME_MAX];
2086 val = strncpy_from_user(name, optval,
2087 min_t(long, TCP_CA_NAME_MAX-1, optlen));
2093 err = tcp_set_congestion_control(sk, name);
2098 if (optlen < sizeof(int))
2101 if (get_user(val, (int __user *)optval))
2108 /* Values greater than interface MTU won't take effect. However
2109 * at the point when this call is done we typically don't yet
2110 * know which interface is going to be used */
2111 if (val < 8 || val > MAX_TCP_WINDOW) {
2115 tp->rx_opt.user_mss = val;
2120 /* TCP_NODELAY is weaker than TCP_CORK, so that
2121 * this option on corked socket is remembered, but
2122 * it is not activated until cork is cleared.
2124 * However, when TCP_NODELAY is set we make
2125 * an explicit push, which overrides even TCP_CORK
2126 * for currently queued segments.
2128 tp->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH;
2129 tcp_push_pending_frames(sk);
2131 tp->nonagle &= ~TCP_NAGLE_OFF;
2136 /* When set indicates to always queue non-full frames.
2137 * Later the user clears this option and we transmit
2138 * any pending partial frames in the queue. This is
2139 * meant to be used alongside sendfile() to get properly
2140 * filled frames when the user (for example) must write
2141 * out headers with a write() call first and then use
2142 * sendfile to send out the data parts.
2144 * TCP_CORK can be set together with TCP_NODELAY and it is
2145 * stronger than TCP_NODELAY.
2148 tp->nonagle |= TCP_NAGLE_CORK;
2150 tp->nonagle &= ~TCP_NAGLE_CORK;
2151 if (tp->nonagle&TCP_NAGLE_OFF)
2152 tp->nonagle |= TCP_NAGLE_PUSH;
2153 tcp_push_pending_frames(sk);
2158 if (val < 1 || val > MAX_TCP_KEEPIDLE)
2161 tp->keepalive_time = val * HZ;
2162 if (sock_flag(sk, SOCK_KEEPOPEN) &&
2163 !((1 << sk->sk_state) &
2164 (TCPF_CLOSE | TCPF_LISTEN))) {
2165 __u32 elapsed = tcp_time_stamp - tp->rcv_tstamp;
2166 if (tp->keepalive_time > elapsed)
2167 elapsed = tp->keepalive_time - elapsed;
2170 inet_csk_reset_keepalive_timer(sk, elapsed);
2175 if (val < 1 || val > MAX_TCP_KEEPINTVL)
2178 tp->keepalive_intvl = val * HZ;
2181 if (val < 1 || val > MAX_TCP_KEEPCNT)
2184 tp->keepalive_probes = val;
2187 if (val < 1 || val > MAX_TCP_SYNCNT)
2190 icsk->icsk_syn_retries = val;
2196 else if (val > sysctl_tcp_fin_timeout / HZ)
2199 tp->linger2 = val * HZ;
2202 case TCP_DEFER_ACCEPT:
2203 /* Translate value in seconds to number of retransmits */
2204 icsk->icsk_accept_queue.rskq_defer_accept =
2205 secs_to_retrans(val, TCP_TIMEOUT_INIT / HZ,
2209 case TCP_WINDOW_CLAMP:
2211 if (sk->sk_state != TCP_CLOSE) {
2215 tp->window_clamp = 0;
2217 tp->window_clamp = val < SOCK_MIN_RCVBUF / 2 ?
2218 SOCK_MIN_RCVBUF / 2 : val;
2223 icsk->icsk_ack.pingpong = 1;
2225 icsk->icsk_ack.pingpong = 0;
2226 if ((1 << sk->sk_state) &
2227 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) &&
2228 inet_csk_ack_scheduled(sk)) {
2229 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
2230 tcp_cleanup_rbuf(sk, 1);
2232 icsk->icsk_ack.pingpong = 1;
2237 #ifdef CONFIG_TCP_MD5SIG
2239 /* Read the IP->Key mappings from userspace */
2240 err = tp->af_specific->md5_parse(sk, optval, optlen);
2253 int tcp_setsockopt(struct sock *sk, int level, int optname, char __user *optval,
2254 unsigned int optlen)
2256 struct inet_connection_sock *icsk = inet_csk(sk);
2258 if (level != SOL_TCP)
2259 return icsk->icsk_af_ops->setsockopt(sk, level, optname,
2261 return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2264 #ifdef CONFIG_COMPAT
2265 int compat_tcp_setsockopt(struct sock *sk, int level, int optname,
2266 char __user *optval, unsigned int optlen)
2268 if (level != SOL_TCP)
2269 return inet_csk_compat_setsockopt(sk, level, optname,
2271 return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2274 EXPORT_SYMBOL(compat_tcp_setsockopt);
2277 /* Return information about state of tcp endpoint in API format. */
2278 void tcp_get_info(struct sock *sk, struct tcp_info *info)
2280 struct tcp_sock *tp = tcp_sk(sk);
2281 const struct inet_connection_sock *icsk = inet_csk(sk);
2282 u32 now = tcp_time_stamp;
2284 memset(info, 0, sizeof(*info));
2286 info->tcpi_state = sk->sk_state;
2287 info->tcpi_ca_state = icsk->icsk_ca_state;
2288 info->tcpi_retransmits = icsk->icsk_retransmits;
2289 info->tcpi_probes = icsk->icsk_probes_out;
2290 info->tcpi_backoff = icsk->icsk_backoff;
2292 if (tp->rx_opt.tstamp_ok)
2293 info->tcpi_options |= TCPI_OPT_TIMESTAMPS;
2294 if (tcp_is_sack(tp))
2295 info->tcpi_options |= TCPI_OPT_SACK;
2296 if (tp->rx_opt.wscale_ok) {
2297 info->tcpi_options |= TCPI_OPT_WSCALE;
2298 info->tcpi_snd_wscale = tp->rx_opt.snd_wscale;
2299 info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale;
2302 if (tp->ecn_flags&TCP_ECN_OK)
2303 info->tcpi_options |= TCPI_OPT_ECN;
2305 info->tcpi_rto = jiffies_to_usecs(icsk->icsk_rto);
2306 info->tcpi_ato = jiffies_to_usecs(icsk->icsk_ack.ato);
2307 info->tcpi_snd_mss = tp->mss_cache;
2308 info->tcpi_rcv_mss = icsk->icsk_ack.rcv_mss;
2310 if (sk->sk_state == TCP_LISTEN) {
2311 info->tcpi_unacked = sk->sk_ack_backlog;
2312 info->tcpi_sacked = sk->sk_max_ack_backlog;
2314 info->tcpi_unacked = tp->packets_out;
2315 info->tcpi_sacked = tp->sacked_out;
2317 info->tcpi_lost = tp->lost_out;
2318 info->tcpi_retrans = tp->retrans_out;
2319 info->tcpi_fackets = tp->fackets_out;
2321 info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime);
2322 info->tcpi_last_data_recv = jiffies_to_msecs(now - icsk->icsk_ack.lrcvtime);
2323 info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp);
2325 info->tcpi_pmtu = icsk->icsk_pmtu_cookie;
2326 info->tcpi_rcv_ssthresh = tp->rcv_ssthresh;
2327 info->tcpi_rtt = jiffies_to_usecs(tp->srtt)>>3;
2328 info->tcpi_rttvar = jiffies_to_usecs(tp->mdev)>>2;
2329 info->tcpi_snd_ssthresh = tp->snd_ssthresh;
2330 info->tcpi_snd_cwnd = tp->snd_cwnd;
2331 info->tcpi_advmss = tp->advmss;
2332 info->tcpi_reordering = tp->reordering;
2334 info->tcpi_rcv_rtt = jiffies_to_usecs(tp->rcv_rtt_est.rtt)>>3;
2335 info->tcpi_rcv_space = tp->rcvq_space.space;
2337 info->tcpi_total_retrans = tp->total_retrans;
2340 EXPORT_SYMBOL_GPL(tcp_get_info);
2342 static int do_tcp_getsockopt(struct sock *sk, int level,
2343 int optname, char __user *optval, int __user *optlen)
2345 struct inet_connection_sock *icsk = inet_csk(sk);
2346 struct tcp_sock *tp = tcp_sk(sk);
2349 if (get_user(len, optlen))
2352 len = min_t(unsigned int, len, sizeof(int));
2359 val = tp->mss_cache;
2360 if (!val && ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
2361 val = tp->rx_opt.user_mss;
2364 val = !!(tp->nonagle&TCP_NAGLE_OFF);
2367 val = !!(tp->nonagle&TCP_NAGLE_CORK);
2370 val = keepalive_time_when(tp) / HZ;
2373 val = keepalive_intvl_when(tp) / HZ;
2376 val = keepalive_probes(tp);
2379 val = icsk->icsk_syn_retries ? : sysctl_tcp_syn_retries;
2384 val = (val ? : sysctl_tcp_fin_timeout) / HZ;
2386 case TCP_DEFER_ACCEPT:
2387 val = retrans_to_secs(icsk->icsk_accept_queue.rskq_defer_accept,
2388 TCP_TIMEOUT_INIT / HZ, TCP_RTO_MAX / HZ);
2390 case TCP_WINDOW_CLAMP:
2391 val = tp->window_clamp;
2394 struct tcp_info info;
2396 if (get_user(len, optlen))
2399 tcp_get_info(sk, &info);
2401 len = min_t(unsigned int, len, sizeof(info));
2402 if (put_user(len, optlen))
2404 if (copy_to_user(optval, &info, len))
2409 val = !icsk->icsk_ack.pingpong;
2412 case TCP_CONGESTION:
2413 if (get_user(len, optlen))
2415 len = min_t(unsigned int, len, TCP_CA_NAME_MAX);
2416 if (put_user(len, optlen))
2418 if (copy_to_user(optval, icsk->icsk_ca_ops->name, len))
2422 return -ENOPROTOOPT;
2425 if (put_user(len, optlen))
2427 if (copy_to_user(optval, &val, len))
2432 int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval,
2435 struct inet_connection_sock *icsk = inet_csk(sk);
2437 if (level != SOL_TCP)
2438 return icsk->icsk_af_ops->getsockopt(sk, level, optname,
2440 return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2443 #ifdef CONFIG_COMPAT
2444 int compat_tcp_getsockopt(struct sock *sk, int level, int optname,
2445 char __user *optval, int __user *optlen)
2447 if (level != SOL_TCP)
2448 return inet_csk_compat_getsockopt(sk, level, optname,
2450 return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2453 EXPORT_SYMBOL(compat_tcp_getsockopt);
2456 struct sk_buff *tcp_tso_segment(struct sk_buff *skb, int features)
2458 struct sk_buff *segs = ERR_PTR(-EINVAL);
2463 unsigned int oldlen;
2466 if (!pskb_may_pull(skb, sizeof(*th)))
2470 thlen = th->doff * 4;
2471 if (thlen < sizeof(*th))
2474 if (!pskb_may_pull(skb, thlen))
2477 oldlen = (u16)~skb->len;
2478 __skb_pull(skb, thlen);
2480 mss = skb_shinfo(skb)->gso_size;
2481 if (unlikely(skb->len <= mss))
2484 if (skb_gso_ok(skb, features | NETIF_F_GSO_ROBUST)) {
2485 /* Packet is from an untrusted source, reset gso_segs. */
2486 int type = skb_shinfo(skb)->gso_type;
2494 !(type & (SKB_GSO_TCPV4 | SKB_GSO_TCPV6))))
2497 skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss);
2503 segs = skb_segment(skb, features);
2507 delta = htonl(oldlen + (thlen + mss));
2511 seq = ntohl(th->seq);
2514 th->fin = th->psh = 0;
2516 th->check = ~csum_fold((__force __wsum)((__force u32)th->check +
2517 (__force u32)delta));
2518 if (skb->ip_summed != CHECKSUM_PARTIAL)
2520 csum_fold(csum_partial(skb_transport_header(skb),
2527 th->seq = htonl(seq);
2529 } while (skb->next);
2531 delta = htonl(oldlen + (skb->tail - skb->transport_header) +
2533 th->check = ~csum_fold((__force __wsum)((__force u32)th->check +
2534 (__force u32)delta));
2535 if (skb->ip_summed != CHECKSUM_PARTIAL)
2536 th->check = csum_fold(csum_partial(skb_transport_header(skb),
2542 EXPORT_SYMBOL(tcp_tso_segment);
2544 struct sk_buff **tcp_gro_receive(struct sk_buff **head, struct sk_buff *skb)
2546 struct sk_buff **pp = NULL;
2553 unsigned int mss = 1;
2559 off = skb_gro_offset(skb);
2560 hlen = off + sizeof(*th);
2561 th = skb_gro_header_fast(skb, off);
2562 if (skb_gro_header_hard(skb, hlen)) {
2563 th = skb_gro_header_slow(skb, hlen, off);
2568 thlen = th->doff * 4;
2569 if (thlen < sizeof(*th))
2573 if (skb_gro_header_hard(skb, hlen)) {
2574 th = skb_gro_header_slow(skb, hlen, off);
2579 skb_gro_pull(skb, thlen);
2581 len = skb_gro_len(skb);
2582 flags = tcp_flag_word(th);
2584 for (; (p = *head); head = &p->next) {
2585 if (!NAPI_GRO_CB(p)->same_flow)
2590 if (*(u32 *)&th->source ^ *(u32 *)&th2->source) {
2591 NAPI_GRO_CB(p)->same_flow = 0;
2598 goto out_check_final;
2601 flush = NAPI_GRO_CB(p)->flush;
2602 flush |= flags & TCP_FLAG_CWR;
2603 flush |= (flags ^ tcp_flag_word(th2)) &
2604 ~(TCP_FLAG_CWR | TCP_FLAG_FIN | TCP_FLAG_PSH);
2605 flush |= th->ack_seq ^ th2->ack_seq;
2606 for (i = sizeof(*th); i < thlen; i += 4)
2607 flush |= *(u32 *)((u8 *)th + i) ^
2608 *(u32 *)((u8 *)th2 + i);
2610 mss = skb_shinfo(p)->gso_size;
2612 flush |= (len - 1) >= mss;
2613 flush |= (ntohl(th2->seq) + skb_gro_len(p)) ^ ntohl(th->seq);
2615 if (flush || skb_gro_receive(head, skb)) {
2617 goto out_check_final;
2622 tcp_flag_word(th2) |= flags & (TCP_FLAG_FIN | TCP_FLAG_PSH);
2626 flush |= flags & (TCP_FLAG_URG | TCP_FLAG_PSH | TCP_FLAG_RST |
2627 TCP_FLAG_SYN | TCP_FLAG_FIN);
2629 if (p && (!NAPI_GRO_CB(skb)->same_flow || flush))
2633 NAPI_GRO_CB(skb)->flush |= flush;
2637 EXPORT_SYMBOL(tcp_gro_receive);
2639 int tcp_gro_complete(struct sk_buff *skb)
2641 struct tcphdr *th = tcp_hdr(skb);
2643 skb->csum_start = skb_transport_header(skb) - skb->head;
2644 skb->csum_offset = offsetof(struct tcphdr, check);
2645 skb->ip_summed = CHECKSUM_PARTIAL;
2647 skb_shinfo(skb)->gso_segs = NAPI_GRO_CB(skb)->count;
2650 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
2654 EXPORT_SYMBOL(tcp_gro_complete);
2656 #ifdef CONFIG_TCP_MD5SIG
2657 static unsigned long tcp_md5sig_users;
2658 static struct tcp_md5sig_pool **tcp_md5sig_pool;
2659 static DEFINE_SPINLOCK(tcp_md5sig_pool_lock);
2661 static void __tcp_free_md5sig_pool(struct tcp_md5sig_pool **pool)
2664 for_each_possible_cpu(cpu) {
2665 struct tcp_md5sig_pool *p = *per_cpu_ptr(pool, cpu);
2667 if (p->md5_desc.tfm)
2668 crypto_free_hash(p->md5_desc.tfm);
2676 void tcp_free_md5sig_pool(void)
2678 struct tcp_md5sig_pool **pool = NULL;
2680 spin_lock_bh(&tcp_md5sig_pool_lock);
2681 if (--tcp_md5sig_users == 0) {
2682 pool = tcp_md5sig_pool;
2683 tcp_md5sig_pool = NULL;
2685 spin_unlock_bh(&tcp_md5sig_pool_lock);
2687 __tcp_free_md5sig_pool(pool);
2690 EXPORT_SYMBOL(tcp_free_md5sig_pool);
2692 static struct tcp_md5sig_pool **__tcp_alloc_md5sig_pool(struct sock *sk)
2695 struct tcp_md5sig_pool **pool;
2697 pool = alloc_percpu(struct tcp_md5sig_pool *);
2701 for_each_possible_cpu(cpu) {
2702 struct tcp_md5sig_pool *p;
2703 struct crypto_hash *hash;
2705 p = kzalloc(sizeof(*p), sk->sk_allocation);
2708 *per_cpu_ptr(pool, cpu) = p;
2710 hash = crypto_alloc_hash("md5", 0, CRYPTO_ALG_ASYNC);
2711 if (!hash || IS_ERR(hash))
2714 p->md5_desc.tfm = hash;
2718 __tcp_free_md5sig_pool(pool);
2722 struct tcp_md5sig_pool **tcp_alloc_md5sig_pool(struct sock *sk)
2724 struct tcp_md5sig_pool **pool;
2728 spin_lock_bh(&tcp_md5sig_pool_lock);
2729 pool = tcp_md5sig_pool;
2730 if (tcp_md5sig_users++ == 0) {
2732 spin_unlock_bh(&tcp_md5sig_pool_lock);
2735 spin_unlock_bh(&tcp_md5sig_pool_lock);
2739 spin_unlock_bh(&tcp_md5sig_pool_lock);
2742 /* we cannot hold spinlock here because this may sleep. */
2743 struct tcp_md5sig_pool **p = __tcp_alloc_md5sig_pool(sk);
2744 spin_lock_bh(&tcp_md5sig_pool_lock);
2747 spin_unlock_bh(&tcp_md5sig_pool_lock);
2750 pool = tcp_md5sig_pool;
2752 /* oops, it has already been assigned. */
2753 spin_unlock_bh(&tcp_md5sig_pool_lock);
2754 __tcp_free_md5sig_pool(p);
2756 tcp_md5sig_pool = pool = p;
2757 spin_unlock_bh(&tcp_md5sig_pool_lock);
2763 EXPORT_SYMBOL(tcp_alloc_md5sig_pool);
2765 struct tcp_md5sig_pool *__tcp_get_md5sig_pool(int cpu)
2767 struct tcp_md5sig_pool **p;
2768 spin_lock_bh(&tcp_md5sig_pool_lock);
2769 p = tcp_md5sig_pool;
2772 spin_unlock_bh(&tcp_md5sig_pool_lock);
2773 return (p ? *per_cpu_ptr(p, cpu) : NULL);
2776 EXPORT_SYMBOL(__tcp_get_md5sig_pool);
2778 void __tcp_put_md5sig_pool(void)
2780 tcp_free_md5sig_pool();
2783 EXPORT_SYMBOL(__tcp_put_md5sig_pool);
2785 int tcp_md5_hash_header(struct tcp_md5sig_pool *hp,
2788 struct scatterlist sg;
2791 __sum16 old_checksum = th->check;
2793 /* options aren't included in the hash */
2794 sg_init_one(&sg, th, sizeof(struct tcphdr));
2795 err = crypto_hash_update(&hp->md5_desc, &sg, sizeof(struct tcphdr));
2796 th->check = old_checksum;
2800 EXPORT_SYMBOL(tcp_md5_hash_header);
2802 int tcp_md5_hash_skb_data(struct tcp_md5sig_pool *hp,
2803 struct sk_buff *skb, unsigned header_len)
2805 struct scatterlist sg;
2806 const struct tcphdr *tp = tcp_hdr(skb);
2807 struct hash_desc *desc = &hp->md5_desc;
2809 const unsigned head_data_len = skb_headlen(skb) > header_len ?
2810 skb_headlen(skb) - header_len : 0;
2811 const struct skb_shared_info *shi = skb_shinfo(skb);
2813 sg_init_table(&sg, 1);
2815 sg_set_buf(&sg, ((u8 *) tp) + header_len, head_data_len);
2816 if (crypto_hash_update(desc, &sg, head_data_len))
2819 for (i = 0; i < shi->nr_frags; ++i) {
2820 const struct skb_frag_struct *f = &shi->frags[i];
2821 sg_set_page(&sg, f->page, f->size, f->page_offset);
2822 if (crypto_hash_update(desc, &sg, f->size))
2829 EXPORT_SYMBOL(tcp_md5_hash_skb_data);
2831 int tcp_md5_hash_key(struct tcp_md5sig_pool *hp, struct tcp_md5sig_key *key)
2833 struct scatterlist sg;
2835 sg_init_one(&sg, key->key, key->keylen);
2836 return crypto_hash_update(&hp->md5_desc, &sg, key->keylen);
2839 EXPORT_SYMBOL(tcp_md5_hash_key);
2843 void tcp_done(struct sock *sk)
2845 if (sk->sk_state == TCP_SYN_SENT || sk->sk_state == TCP_SYN_RECV)
2846 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_ATTEMPTFAILS);
2848 tcp_set_state(sk, TCP_CLOSE);
2849 tcp_clear_xmit_timers(sk);
2851 sk->sk_shutdown = SHUTDOWN_MASK;
2853 if (!sock_flag(sk, SOCK_DEAD))
2854 sk->sk_state_change(sk);
2856 inet_csk_destroy_sock(sk);
2858 EXPORT_SYMBOL_GPL(tcp_done);
2860 extern struct tcp_congestion_ops tcp_reno;
2862 static __initdata unsigned long thash_entries;
2863 static int __init set_thash_entries(char *str)
2867 thash_entries = simple_strtoul(str, &str, 0);
2870 __setup("thash_entries=", set_thash_entries);
2872 void __init tcp_init(void)
2874 struct sk_buff *skb = NULL;
2875 unsigned long nr_pages, limit;
2876 int order, i, max_share;
2878 BUILD_BUG_ON(sizeof(struct tcp_skb_cb) > sizeof(skb->cb));
2880 percpu_counter_init(&tcp_sockets_allocated, 0);
2881 percpu_counter_init(&tcp_orphan_count, 0);
2882 tcp_hashinfo.bind_bucket_cachep =
2883 kmem_cache_create("tcp_bind_bucket",
2884 sizeof(struct inet_bind_bucket), 0,
2885 SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL);
2887 /* Size and allocate the main established and bind bucket
2890 * The methodology is similar to that of the buffer cache.
2892 tcp_hashinfo.ehash =
2893 alloc_large_system_hash("TCP established",
2894 sizeof(struct inet_ehash_bucket),
2896 (totalram_pages >= 128 * 1024) ?
2899 &tcp_hashinfo.ehash_size,
2901 thash_entries ? 0 : 512 * 1024);
2902 tcp_hashinfo.ehash_size = 1 << tcp_hashinfo.ehash_size;
2903 for (i = 0; i < tcp_hashinfo.ehash_size; i++) {
2904 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].chain, i);
2905 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].twchain, i);
2907 if (inet_ehash_locks_alloc(&tcp_hashinfo))
2908 panic("TCP: failed to alloc ehash_locks");
2909 tcp_hashinfo.bhash =
2910 alloc_large_system_hash("TCP bind",
2911 sizeof(struct inet_bind_hashbucket),
2912 tcp_hashinfo.ehash_size,
2913 (totalram_pages >= 128 * 1024) ?
2916 &tcp_hashinfo.bhash_size,
2919 tcp_hashinfo.bhash_size = 1 << tcp_hashinfo.bhash_size;
2920 for (i = 0; i < tcp_hashinfo.bhash_size; i++) {
2921 spin_lock_init(&tcp_hashinfo.bhash[i].lock);
2922 INIT_HLIST_HEAD(&tcp_hashinfo.bhash[i].chain);
2925 /* Try to be a bit smarter and adjust defaults depending
2926 * on available memory.
2928 for (order = 0; ((1 << order) << PAGE_SHIFT) <
2929 (tcp_hashinfo.bhash_size * sizeof(struct inet_bind_hashbucket));
2933 tcp_death_row.sysctl_max_tw_buckets = 180000;
2934 sysctl_tcp_max_orphans = 4096 << (order - 4);
2935 sysctl_max_syn_backlog = 1024;
2936 } else if (order < 3) {
2937 tcp_death_row.sysctl_max_tw_buckets >>= (3 - order);
2938 sysctl_tcp_max_orphans >>= (3 - order);
2939 sysctl_max_syn_backlog = 128;
2942 /* Set the pressure threshold to be a fraction of global memory that
2943 * is up to 1/2 at 256 MB, decreasing toward zero with the amount of
2944 * memory, with a floor of 128 pages.
2946 nr_pages = totalram_pages - totalhigh_pages;
2947 limit = min(nr_pages, 1UL<<(28-PAGE_SHIFT)) >> (20-PAGE_SHIFT);
2948 limit = (limit * (nr_pages >> (20-PAGE_SHIFT))) >> (PAGE_SHIFT-11);
2949 limit = max(limit, 128UL);
2950 sysctl_tcp_mem[0] = limit / 4 * 3;
2951 sysctl_tcp_mem[1] = limit;
2952 sysctl_tcp_mem[2] = sysctl_tcp_mem[0] * 2;
2954 /* Set per-socket limits to no more than 1/128 the pressure threshold */
2955 limit = ((unsigned long)sysctl_tcp_mem[1]) << (PAGE_SHIFT - 7);
2956 max_share = min(4UL*1024*1024, limit);
2958 sysctl_tcp_wmem[0] = SK_MEM_QUANTUM;
2959 sysctl_tcp_wmem[1] = 16*1024;
2960 sysctl_tcp_wmem[2] = max(64*1024, max_share);
2962 sysctl_tcp_rmem[0] = SK_MEM_QUANTUM;
2963 sysctl_tcp_rmem[1] = 87380;
2964 sysctl_tcp_rmem[2] = max(87380, max_share);
2966 printk(KERN_INFO "TCP: Hash tables configured "
2967 "(established %d bind %d)\n",
2968 tcp_hashinfo.ehash_size, tcp_hashinfo.bhash_size);
2970 tcp_register_congestion_control(&tcp_reno);
2973 EXPORT_SYMBOL(tcp_close);
2974 EXPORT_SYMBOL(tcp_disconnect);
2975 EXPORT_SYMBOL(tcp_getsockopt);
2976 EXPORT_SYMBOL(tcp_ioctl);
2977 EXPORT_SYMBOL(tcp_poll);
2978 EXPORT_SYMBOL(tcp_read_sock);
2979 EXPORT_SYMBOL(tcp_recvmsg);
2980 EXPORT_SYMBOL(tcp_sendmsg);
2981 EXPORT_SYMBOL(tcp_splice_read);
2982 EXPORT_SYMBOL(tcp_sendpage);
2983 EXPORT_SYMBOL(tcp_setsockopt);
2984 EXPORT_SYMBOL(tcp_shutdown);