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 #define pr_fmt(fmt) "TCP: " fmt
250 #include <linux/kernel.h>
251 #include <linux/module.h>
252 #include <linux/types.h>
253 #include <linux/fcntl.h>
254 #include <linux/poll.h>
255 #include <linux/init.h>
256 #include <linux/fs.h>
257 #include <linux/skbuff.h>
258 #include <linux/scatterlist.h>
259 #include <linux/splice.h>
260 #include <linux/net.h>
261 #include <linux/socket.h>
262 #include <linux/random.h>
263 #include <linux/bootmem.h>
264 #include <linux/highmem.h>
265 #include <linux/swap.h>
266 #include <linux/cache.h>
267 #include <linux/err.h>
268 #include <linux/crypto.h>
269 #include <linux/time.h>
270 #include <linux/slab.h>
272 #include <net/icmp.h>
274 #include <net/xfrm.h>
276 #include <net/netdma.h>
277 #include <net/sock.h>
279 #include <asm/uaccess.h>
280 #include <asm/ioctls.h>
282 int sysctl_tcp_fin_timeout __read_mostly = TCP_FIN_TIMEOUT;
284 struct percpu_counter tcp_orphan_count;
285 EXPORT_SYMBOL_GPL(tcp_orphan_count);
287 int sysctl_tcp_wmem[3] __read_mostly;
288 int sysctl_tcp_rmem[3] __read_mostly;
290 EXPORT_SYMBOL(sysctl_tcp_rmem);
291 EXPORT_SYMBOL(sysctl_tcp_wmem);
293 atomic_long_t tcp_memory_allocated; /* Current allocated memory. */
294 EXPORT_SYMBOL(tcp_memory_allocated);
297 * Current number of TCP sockets.
299 struct percpu_counter tcp_sockets_allocated;
300 EXPORT_SYMBOL(tcp_sockets_allocated);
305 struct tcp_splice_state {
306 struct pipe_inode_info *pipe;
312 * Pressure flag: try to collapse.
313 * Technical note: it is used by multiple contexts non atomically.
314 * All the __sk_mem_schedule() is of this nature: accounting
315 * is strict, actions are advisory and have some latency.
317 int tcp_memory_pressure __read_mostly;
318 EXPORT_SYMBOL(tcp_memory_pressure);
320 void tcp_enter_memory_pressure(struct sock *sk)
322 if (!tcp_memory_pressure) {
323 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMEMORYPRESSURES);
324 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)
366 /* Address-family independent initialization for a tcp_sock.
368 * NOTE: A lot of things set to zero explicitly by call to
369 * sk_alloc() so need not be done here.
371 void tcp_init_sock(struct sock *sk)
373 struct inet_connection_sock *icsk = inet_csk(sk);
374 struct tcp_sock *tp = tcp_sk(sk);
376 skb_queue_head_init(&tp->out_of_order_queue);
377 tcp_init_xmit_timers(sk);
378 tcp_prequeue_init(tp);
380 icsk->icsk_rto = TCP_TIMEOUT_INIT;
381 tp->mdev = TCP_TIMEOUT_INIT;
383 /* So many TCP implementations out there (incorrectly) count the
384 * initial SYN frame in their delayed-ACK and congestion control
385 * algorithms that we must have the following bandaid to talk
386 * efficiently to them. -DaveM
388 tp->snd_cwnd = TCP_INIT_CWND;
390 /* See draft-stevens-tcpca-spec-01 for discussion of the
391 * initialization of these values.
393 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
394 tp->snd_cwnd_clamp = ~0;
395 tp->mss_cache = TCP_MSS_DEFAULT;
397 tp->reordering = sysctl_tcp_reordering;
398 icsk->icsk_ca_ops = &tcp_init_congestion_ops;
400 sk->sk_state = TCP_CLOSE;
402 sk->sk_write_space = sk_stream_write_space;
403 sock_set_flag(sk, SOCK_USE_WRITE_QUEUE);
405 icsk->icsk_sync_mss = tcp_sync_mss;
407 /* TCP Cookie Transactions */
408 if (sysctl_tcp_cookie_size > 0) {
409 /* Default, cookies without s_data_payload. */
411 kzalloc(sizeof(*tp->cookie_values),
413 if (tp->cookie_values != NULL)
414 kref_init(&tp->cookie_values->kref);
416 /* Presumed zeroed, in order of appearance:
417 * cookie_in_always, cookie_out_never,
418 * s_data_constant, s_data_in, s_data_out
420 sk->sk_sndbuf = sysctl_tcp_wmem[1];
421 sk->sk_rcvbuf = sysctl_tcp_rmem[1];
424 sock_update_memcg(sk);
425 sk_sockets_allocated_inc(sk);
428 EXPORT_SYMBOL(tcp_init_sock);
431 * Wait for a TCP event.
433 * Note that we don't need to lock the socket, as the upper poll layers
434 * take care of normal races (between the test and the event) and we don't
435 * go look at any of the socket buffers directly.
437 unsigned int tcp_poll(struct file *file, struct socket *sock, poll_table *wait)
440 struct sock *sk = sock->sk;
441 const struct tcp_sock *tp = tcp_sk(sk);
443 sock_poll_wait(file, sk_sleep(sk), wait);
444 if (sk->sk_state == TCP_LISTEN)
445 return inet_csk_listen_poll(sk);
447 /* Socket is not locked. We are protected from async events
448 * by poll logic and correct handling of state changes
449 * made by other threads is impossible in any case.
455 * POLLHUP is certainly not done right. But poll() doesn't
456 * have a notion of HUP in just one direction, and for a
457 * socket the read side is more interesting.
459 * Some poll() documentation says that POLLHUP is incompatible
460 * with the POLLOUT/POLLWR flags, so somebody should check this
461 * all. But careful, it tends to be safer to return too many
462 * bits than too few, and you can easily break real applications
463 * if you don't tell them that something has hung up!
467 * Check number 1. POLLHUP is _UNMASKABLE_ event (see UNIX98 and
468 * our fs/select.c). It means that after we received EOF,
469 * poll always returns immediately, making impossible poll() on write()
470 * in state CLOSE_WAIT. One solution is evident --- to set POLLHUP
471 * if and only if shutdown has been made in both directions.
472 * Actually, it is interesting to look how Solaris and DUX
473 * solve this dilemma. I would prefer, if POLLHUP were maskable,
474 * then we could set it on SND_SHUTDOWN. BTW examples given
475 * in Stevens' books assume exactly this behaviour, it explains
476 * why POLLHUP is incompatible with POLLOUT. --ANK
478 * NOTE. Check for TCP_CLOSE is added. The goal is to prevent
479 * blocking on fresh not-connected or disconnected socket. --ANK
481 if (sk->sk_shutdown == SHUTDOWN_MASK || sk->sk_state == TCP_CLOSE)
483 if (sk->sk_shutdown & RCV_SHUTDOWN)
484 mask |= POLLIN | POLLRDNORM | POLLRDHUP;
487 if ((1 << sk->sk_state) & ~(TCPF_SYN_SENT | TCPF_SYN_RECV)) {
488 int target = sock_rcvlowat(sk, 0, INT_MAX);
490 if (tp->urg_seq == tp->copied_seq &&
491 !sock_flag(sk, SOCK_URGINLINE) &&
495 /* Potential race condition. If read of tp below will
496 * escape above sk->sk_state, we can be illegally awaken
497 * in SYN_* states. */
498 if (tp->rcv_nxt - tp->copied_seq >= target)
499 mask |= POLLIN | POLLRDNORM;
501 if (!(sk->sk_shutdown & SEND_SHUTDOWN)) {
502 if (sk_stream_wspace(sk) >= sk_stream_min_wspace(sk)) {
503 mask |= POLLOUT | POLLWRNORM;
504 } else { /* send SIGIO later */
505 set_bit(SOCK_ASYNC_NOSPACE,
506 &sk->sk_socket->flags);
507 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
509 /* Race breaker. If space is freed after
510 * wspace test but before the flags are set,
511 * IO signal will be lost.
513 if (sk_stream_wspace(sk) >= sk_stream_min_wspace(sk))
514 mask |= POLLOUT | POLLWRNORM;
517 mask |= POLLOUT | POLLWRNORM;
519 if (tp->urg_data & TCP_URG_VALID)
522 /* This barrier is coupled with smp_wmb() in tcp_reset() */
529 EXPORT_SYMBOL(tcp_poll);
531 int tcp_ioctl(struct sock *sk, int cmd, unsigned long arg)
533 struct tcp_sock *tp = tcp_sk(sk);
538 if (sk->sk_state == TCP_LISTEN)
542 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
544 else if (sock_flag(sk, SOCK_URGINLINE) ||
546 before(tp->urg_seq, tp->copied_seq) ||
547 !before(tp->urg_seq, tp->rcv_nxt)) {
550 answ = tp->rcv_nxt - tp->copied_seq;
552 /* Subtract 1, if FIN is in queue. */
553 skb = skb_peek_tail(&sk->sk_receive_queue);
555 answ -= tcp_hdr(skb)->fin;
557 answ = tp->urg_seq - tp->copied_seq;
561 answ = tp->urg_data && tp->urg_seq == tp->copied_seq;
564 if (sk->sk_state == TCP_LISTEN)
567 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
570 answ = tp->write_seq - tp->snd_una;
573 if (sk->sk_state == TCP_LISTEN)
576 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
579 answ = tp->write_seq - tp->snd_nxt;
585 return put_user(answ, (int __user *)arg);
587 EXPORT_SYMBOL(tcp_ioctl);
589 static inline void tcp_mark_push(struct tcp_sock *tp, struct sk_buff *skb)
591 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
592 tp->pushed_seq = tp->write_seq;
595 static inline int forced_push(const struct tcp_sock *tp)
597 return after(tp->write_seq, tp->pushed_seq + (tp->max_window >> 1));
600 static inline void skb_entail(struct sock *sk, struct sk_buff *skb)
602 struct tcp_sock *tp = tcp_sk(sk);
603 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
606 tcb->seq = tcb->end_seq = tp->write_seq;
607 tcb->tcp_flags = TCPHDR_ACK;
609 skb_header_release(skb);
610 tcp_add_write_queue_tail(sk, skb);
611 sk->sk_wmem_queued += skb->truesize;
612 sk_mem_charge(sk, skb->truesize);
613 if (tp->nonagle & TCP_NAGLE_PUSH)
614 tp->nonagle &= ~TCP_NAGLE_PUSH;
617 static inline void tcp_mark_urg(struct tcp_sock *tp, int flags)
620 tp->snd_up = tp->write_seq;
623 static inline void tcp_push(struct sock *sk, int flags, int mss_now,
626 if (tcp_send_head(sk)) {
627 struct tcp_sock *tp = tcp_sk(sk);
629 if (!(flags & MSG_MORE) || forced_push(tp))
630 tcp_mark_push(tp, tcp_write_queue_tail(sk));
632 tcp_mark_urg(tp, flags);
633 __tcp_push_pending_frames(sk, mss_now,
634 (flags & MSG_MORE) ? TCP_NAGLE_CORK : nonagle);
638 static int tcp_splice_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb,
639 unsigned int offset, size_t len)
641 struct tcp_splice_state *tss = rd_desc->arg.data;
644 ret = skb_splice_bits(skb, offset, tss->pipe, min(rd_desc->count, len),
647 rd_desc->count -= ret;
651 static int __tcp_splice_read(struct sock *sk, struct tcp_splice_state *tss)
653 /* Store TCP splice context information in read_descriptor_t. */
654 read_descriptor_t rd_desc = {
659 return tcp_read_sock(sk, &rd_desc, tcp_splice_data_recv);
663 * tcp_splice_read - splice data from TCP socket to a pipe
664 * @sock: socket to splice from
665 * @ppos: position (not valid)
666 * @pipe: pipe to splice to
667 * @len: number of bytes to splice
668 * @flags: splice modifier flags
671 * Will read pages from given socket and fill them into a pipe.
674 ssize_t tcp_splice_read(struct socket *sock, loff_t *ppos,
675 struct pipe_inode_info *pipe, size_t len,
678 struct sock *sk = sock->sk;
679 struct tcp_splice_state tss = {
688 sock_rps_record_flow(sk);
690 * We can't seek on a socket input
699 timeo = sock_rcvtimeo(sk, sock->file->f_flags & O_NONBLOCK);
701 ret = __tcp_splice_read(sk, &tss);
707 if (sock_flag(sk, SOCK_DONE))
710 ret = sock_error(sk);
713 if (sk->sk_shutdown & RCV_SHUTDOWN)
715 if (sk->sk_state == TCP_CLOSE) {
717 * This occurs when user tries to read
718 * from never connected socket.
720 if (!sock_flag(sk, SOCK_DONE))
728 sk_wait_data(sk, &timeo);
729 if (signal_pending(current)) {
730 ret = sock_intr_errno(timeo);
743 if (sk->sk_err || sk->sk_state == TCP_CLOSE ||
744 (sk->sk_shutdown & RCV_SHUTDOWN) ||
745 signal_pending(current))
756 EXPORT_SYMBOL(tcp_splice_read);
758 struct sk_buff *sk_stream_alloc_skb(struct sock *sk, int size, gfp_t gfp)
762 /* The TCP header must be at least 32-bit aligned. */
763 size = ALIGN(size, 4);
765 skb = alloc_skb_fclone(size + sk->sk_prot->max_header, gfp);
767 if (sk_wmem_schedule(sk, skb->truesize)) {
768 skb_reserve(skb, sk->sk_prot->max_header);
770 * Make sure that we have exactly size bytes
771 * available to the caller, no more, no less.
773 skb->avail_size = size;
778 sk->sk_prot->enter_memory_pressure(sk);
779 sk_stream_moderate_sndbuf(sk);
784 static unsigned int tcp_xmit_size_goal(struct sock *sk, u32 mss_now,
787 struct tcp_sock *tp = tcp_sk(sk);
788 u32 xmit_size_goal, old_size_goal;
790 xmit_size_goal = mss_now;
792 if (large_allowed && sk_can_gso(sk)) {
793 xmit_size_goal = ((sk->sk_gso_max_size - 1) -
794 inet_csk(sk)->icsk_af_ops->net_header_len -
795 inet_csk(sk)->icsk_ext_hdr_len -
798 xmit_size_goal = tcp_bound_to_half_wnd(tp, xmit_size_goal);
800 /* We try hard to avoid divides here */
801 old_size_goal = tp->xmit_size_goal_segs * mss_now;
803 if (likely(old_size_goal <= xmit_size_goal &&
804 old_size_goal + mss_now > xmit_size_goal)) {
805 xmit_size_goal = old_size_goal;
807 tp->xmit_size_goal_segs = xmit_size_goal / mss_now;
808 xmit_size_goal = tp->xmit_size_goal_segs * mss_now;
812 return max(xmit_size_goal, mss_now);
815 static int tcp_send_mss(struct sock *sk, int *size_goal, int flags)
819 mss_now = tcp_current_mss(sk);
820 *size_goal = tcp_xmit_size_goal(sk, mss_now, !(flags & MSG_OOB));
825 static ssize_t do_tcp_sendpages(struct sock *sk, struct page **pages, int poffset,
826 size_t psize, int flags)
828 struct tcp_sock *tp = tcp_sk(sk);
829 int mss_now, size_goal;
832 long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
834 /* Wait for a connection to finish. */
835 if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
836 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
839 clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
841 mss_now = tcp_send_mss(sk, &size_goal, flags);
845 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
849 struct sk_buff *skb = tcp_write_queue_tail(sk);
850 struct page *page = pages[poffset / PAGE_SIZE];
852 int offset = poffset % PAGE_SIZE;
853 int size = min_t(size_t, psize, PAGE_SIZE - offset);
856 if (!tcp_send_head(sk) || (copy = size_goal - skb->len) <= 0) {
858 if (!sk_stream_memory_free(sk))
859 goto wait_for_sndbuf;
861 skb = sk_stream_alloc_skb(sk, 0, sk->sk_allocation);
863 goto wait_for_memory;
872 i = skb_shinfo(skb)->nr_frags;
873 can_coalesce = skb_can_coalesce(skb, i, page, offset);
874 if (!can_coalesce && i >= MAX_SKB_FRAGS) {
875 tcp_mark_push(tp, skb);
878 if (!sk_wmem_schedule(sk, copy))
879 goto wait_for_memory;
882 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
885 skb_fill_page_desc(skb, i, page, offset, copy);
889 skb->data_len += copy;
890 skb->truesize += copy;
891 sk->sk_wmem_queued += copy;
892 sk_mem_charge(sk, copy);
893 skb->ip_summed = CHECKSUM_PARTIAL;
894 tp->write_seq += copy;
895 TCP_SKB_CB(skb)->end_seq += copy;
896 skb_shinfo(skb)->gso_segs = 0;
899 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
903 if (!(psize -= copy))
906 if (skb->len < size_goal || (flags & MSG_OOB))
909 if (forced_push(tp)) {
910 tcp_mark_push(tp, skb);
911 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
912 } else if (skb == tcp_send_head(sk))
913 tcp_push_one(sk, mss_now);
917 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
920 tcp_push(sk, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH);
922 if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
925 mss_now = tcp_send_mss(sk, &size_goal, flags);
929 if (copied && !(flags & MSG_SENDPAGE_NOTLAST))
930 tcp_push(sk, flags, mss_now, tp->nonagle);
937 return sk_stream_error(sk, flags, err);
940 int tcp_sendpage(struct sock *sk, struct page *page, int offset,
941 size_t size, int flags)
945 if (!(sk->sk_route_caps & NETIF_F_SG) ||
946 !(sk->sk_route_caps & NETIF_F_ALL_CSUM))
947 return sock_no_sendpage(sk->sk_socket, page, offset, size,
951 res = do_tcp_sendpages(sk, &page, offset, size, flags);
955 EXPORT_SYMBOL(tcp_sendpage);
957 static inline int select_size(const struct sock *sk, bool sg)
959 const struct tcp_sock *tp = tcp_sk(sk);
960 int tmp = tp->mss_cache;
963 if (sk_can_gso(sk)) {
964 /* Small frames wont use a full page:
965 * Payload will immediately follow tcp header.
967 tmp = SKB_WITH_OVERHEAD(2048 - MAX_TCP_HEADER);
969 int pgbreak = SKB_MAX_HEAD(MAX_TCP_HEADER);
971 if (tmp >= pgbreak &&
972 tmp <= pgbreak + (MAX_SKB_FRAGS - 1) * PAGE_SIZE)
980 static int tcp_send_rcvq(struct sock *sk, struct msghdr *msg, size_t size)
983 struct tcp_skb_cb *cb;
986 skb = alloc_skb(size + sizeof(*th), sk->sk_allocation);
990 th = (struct tcphdr *)skb_put(skb, sizeof(*th));
991 skb_reset_transport_header(skb);
992 memset(th, 0, sizeof(*th));
994 if (memcpy_fromiovec(skb_put(skb, size), msg->msg_iov, size))
997 cb = TCP_SKB_CB(skb);
999 TCP_SKB_CB(skb)->seq = tcp_sk(sk)->rcv_nxt;
1000 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + size;
1001 TCP_SKB_CB(skb)->ack_seq = tcp_sk(sk)->snd_una - 1;
1003 tcp_queue_rcv(sk, skb, sizeof(*th));
1013 int tcp_sendmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg,
1017 struct tcp_sock *tp = tcp_sk(sk);
1018 struct sk_buff *skb;
1019 int iovlen, flags, err, copied;
1020 int mss_now = 0, size_goal;
1026 flags = msg->msg_flags;
1027 timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
1029 /* Wait for a connection to finish. */
1030 if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
1031 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
1034 if (unlikely(tp->repair)) {
1035 if (tp->repair_queue == TCP_RECV_QUEUE) {
1036 copied = tcp_send_rcvq(sk, msg, size);
1041 if (tp->repair_queue == TCP_NO_QUEUE)
1044 /* 'common' sending to sendq */
1047 /* This should be in poll */
1048 clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
1050 mss_now = tcp_send_mss(sk, &size_goal, flags);
1052 /* Ok commence sending. */
1053 iovlen = msg->msg_iovlen;
1058 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
1061 sg = !!(sk->sk_route_caps & NETIF_F_SG);
1063 while (--iovlen >= 0) {
1064 size_t seglen = iov->iov_len;
1065 unsigned char __user *from = iov->iov_base;
1069 while (seglen > 0) {
1071 int max = size_goal;
1073 skb = tcp_write_queue_tail(sk);
1074 if (tcp_send_head(sk)) {
1075 if (skb->ip_summed == CHECKSUM_NONE)
1077 copy = max - skb->len;
1082 /* Allocate new segment. If the interface is SG,
1083 * allocate skb fitting to single page.
1085 if (!sk_stream_memory_free(sk))
1086 goto wait_for_sndbuf;
1088 skb = sk_stream_alloc_skb(sk,
1089 select_size(sk, sg),
1092 goto wait_for_memory;
1095 * Check whether we can use HW checksum.
1097 if (sk->sk_route_caps & NETIF_F_ALL_CSUM)
1098 skb->ip_summed = CHECKSUM_PARTIAL;
1100 skb_entail(sk, skb);
1105 /* Try to append data to the end of skb. */
1109 /* Where to copy to? */
1110 if (skb_availroom(skb) > 0) {
1111 /* We have some space in skb head. Superb! */
1112 copy = min_t(int, copy, skb_availroom(skb));
1113 err = skb_add_data_nocache(sk, skb, from, copy);
1118 int i = skb_shinfo(skb)->nr_frags;
1119 struct page *page = sk->sk_sndmsg_page;
1122 if (page && page_count(page) == 1)
1123 sk->sk_sndmsg_off = 0;
1125 off = sk->sk_sndmsg_off;
1127 if (skb_can_coalesce(skb, i, page, off) &&
1129 /* We can extend the last page
1132 } else if (i == MAX_SKB_FRAGS || !sg) {
1133 /* Need to add new fragment and cannot
1134 * do this because interface is non-SG,
1135 * or because all the page slots are
1137 tcp_mark_push(tp, skb);
1140 if (off == PAGE_SIZE) {
1142 sk->sk_sndmsg_page = page = NULL;
1148 if (copy > PAGE_SIZE - off)
1149 copy = PAGE_SIZE - off;
1151 if (!sk_wmem_schedule(sk, copy))
1152 goto wait_for_memory;
1155 /* Allocate new cache page. */
1156 if (!(page = sk_stream_alloc_page(sk)))
1157 goto wait_for_memory;
1160 /* Time to copy data. We are close to
1162 err = skb_copy_to_page_nocache(sk, from, skb,
1165 /* If this page was new, give it to the
1166 * socket so it does not get leaked.
1168 if (!sk->sk_sndmsg_page) {
1169 sk->sk_sndmsg_page = page;
1170 sk->sk_sndmsg_off = 0;
1175 /* Update the skb. */
1177 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1179 skb_fill_page_desc(skb, i, page, off, copy);
1180 if (sk->sk_sndmsg_page) {
1182 } else if (off + copy < PAGE_SIZE) {
1184 sk->sk_sndmsg_page = page;
1188 sk->sk_sndmsg_off = off + copy;
1192 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
1194 tp->write_seq += copy;
1195 TCP_SKB_CB(skb)->end_seq += copy;
1196 skb_shinfo(skb)->gso_segs = 0;
1200 if ((seglen -= copy) == 0 && iovlen == 0)
1203 if (skb->len < max || (flags & MSG_OOB) || unlikely(tp->repair))
1206 if (forced_push(tp)) {
1207 tcp_mark_push(tp, skb);
1208 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
1209 } else if (skb == tcp_send_head(sk))
1210 tcp_push_one(sk, mss_now);
1214 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1216 if (copied && likely(!tp->repair))
1217 tcp_push(sk, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH);
1219 if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
1222 mss_now = tcp_send_mss(sk, &size_goal, flags);
1227 if (copied && likely(!tp->repair))
1228 tcp_push(sk, flags, mss_now, tp->nonagle);
1234 tcp_unlink_write_queue(skb, sk);
1235 /* It is the one place in all of TCP, except connection
1236 * reset, where we can be unlinking the send_head.
1238 tcp_check_send_head(sk, skb);
1239 sk_wmem_free_skb(sk, skb);
1246 err = sk_stream_error(sk, flags, err);
1250 EXPORT_SYMBOL(tcp_sendmsg);
1253 * Handle reading urgent data. BSD has very simple semantics for
1254 * this, no blocking and very strange errors 8)
1257 static int tcp_recv_urg(struct sock *sk, struct msghdr *msg, int len, int flags)
1259 struct tcp_sock *tp = tcp_sk(sk);
1261 /* No URG data to read. */
1262 if (sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data ||
1263 tp->urg_data == TCP_URG_READ)
1264 return -EINVAL; /* Yes this is right ! */
1266 if (sk->sk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DONE))
1269 if (tp->urg_data & TCP_URG_VALID) {
1271 char c = tp->urg_data;
1273 if (!(flags & MSG_PEEK))
1274 tp->urg_data = TCP_URG_READ;
1276 /* Read urgent data. */
1277 msg->msg_flags |= MSG_OOB;
1280 if (!(flags & MSG_TRUNC))
1281 err = memcpy_toiovec(msg->msg_iov, &c, 1);
1284 msg->msg_flags |= MSG_TRUNC;
1286 return err ? -EFAULT : len;
1289 if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
1292 /* Fixed the recv(..., MSG_OOB) behaviour. BSD docs and
1293 * the available implementations agree in this case:
1294 * this call should never block, independent of the
1295 * blocking state of the socket.
1296 * Mike <pall@rz.uni-karlsruhe.de>
1301 static int tcp_peek_sndq(struct sock *sk, struct msghdr *msg, int len)
1303 struct sk_buff *skb;
1304 int copied = 0, err = 0;
1306 /* XXX -- need to support SO_PEEK_OFF */
1308 skb_queue_walk(&sk->sk_write_queue, skb) {
1309 err = skb_copy_datagram_iovec(skb, 0, msg->msg_iov, skb->len);
1316 return err ?: copied;
1319 /* Clean up the receive buffer for full frames taken by the user,
1320 * then send an ACK if necessary. COPIED is the number of bytes
1321 * tcp_recvmsg has given to the user so far, it speeds up the
1322 * calculation of whether or not we must ACK for the sake of
1325 void tcp_cleanup_rbuf(struct sock *sk, int copied)
1327 struct tcp_sock *tp = tcp_sk(sk);
1328 int time_to_ack = 0;
1330 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
1332 WARN(skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq),
1333 "cleanup rbuf bug: copied %X seq %X rcvnxt %X\n",
1334 tp->copied_seq, TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt);
1336 if (inet_csk_ack_scheduled(sk)) {
1337 const struct inet_connection_sock *icsk = inet_csk(sk);
1338 /* Delayed ACKs frequently hit locked sockets during bulk
1340 if (icsk->icsk_ack.blocked ||
1341 /* Once-per-two-segments ACK was not sent by tcp_input.c */
1342 tp->rcv_nxt - tp->rcv_wup > icsk->icsk_ack.rcv_mss ||
1344 * If this read emptied read buffer, we send ACK, if
1345 * connection is not bidirectional, user drained
1346 * receive buffer and there was a small segment
1350 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED2) ||
1351 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED) &&
1352 !icsk->icsk_ack.pingpong)) &&
1353 !atomic_read(&sk->sk_rmem_alloc)))
1357 /* We send an ACK if we can now advertise a non-zero window
1358 * which has been raised "significantly".
1360 * Even if window raised up to infinity, do not send window open ACK
1361 * in states, where we will not receive more. It is useless.
1363 if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) {
1364 __u32 rcv_window_now = tcp_receive_window(tp);
1366 /* Optimize, __tcp_select_window() is not cheap. */
1367 if (2*rcv_window_now <= tp->window_clamp) {
1368 __u32 new_window = __tcp_select_window(sk);
1370 /* Send ACK now, if this read freed lots of space
1371 * in our buffer. Certainly, new_window is new window.
1372 * We can advertise it now, if it is not less than current one.
1373 * "Lots" means "at least twice" here.
1375 if (new_window && new_window >= 2 * rcv_window_now)
1383 static void tcp_prequeue_process(struct sock *sk)
1385 struct sk_buff *skb;
1386 struct tcp_sock *tp = tcp_sk(sk);
1388 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPPREQUEUED);
1390 /* RX process wants to run with disabled BHs, though it is not
1393 while ((skb = __skb_dequeue(&tp->ucopy.prequeue)) != NULL)
1394 sk_backlog_rcv(sk, skb);
1397 /* Clear memory counter. */
1398 tp->ucopy.memory = 0;
1401 #ifdef CONFIG_NET_DMA
1402 static void tcp_service_net_dma(struct sock *sk, bool wait)
1404 dma_cookie_t done, used;
1405 dma_cookie_t last_issued;
1406 struct tcp_sock *tp = tcp_sk(sk);
1408 if (!tp->ucopy.dma_chan)
1411 last_issued = tp->ucopy.dma_cookie;
1412 dma_async_memcpy_issue_pending(tp->ucopy.dma_chan);
1415 if (dma_async_memcpy_complete(tp->ucopy.dma_chan,
1417 &used) == DMA_SUCCESS) {
1418 /* Safe to free early-copied skbs now */
1419 __skb_queue_purge(&sk->sk_async_wait_queue);
1422 struct sk_buff *skb;
1423 while ((skb = skb_peek(&sk->sk_async_wait_queue)) &&
1424 (dma_async_is_complete(skb->dma_cookie, done,
1425 used) == DMA_SUCCESS)) {
1426 __skb_dequeue(&sk->sk_async_wait_queue);
1434 static inline struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off)
1436 struct sk_buff *skb;
1439 skb_queue_walk(&sk->sk_receive_queue, skb) {
1440 offset = seq - TCP_SKB_CB(skb)->seq;
1441 if (tcp_hdr(skb)->syn)
1443 if (offset < skb->len || tcp_hdr(skb)->fin) {
1452 * This routine provides an alternative to tcp_recvmsg() for routines
1453 * that would like to handle copying from skbuffs directly in 'sendfile'
1456 * - It is assumed that the socket was locked by the caller.
1457 * - The routine does not block.
1458 * - At present, there is no support for reading OOB data
1459 * or for 'peeking' the socket using this routine
1460 * (although both would be easy to implement).
1462 int tcp_read_sock(struct sock *sk, read_descriptor_t *desc,
1463 sk_read_actor_t recv_actor)
1465 struct sk_buff *skb;
1466 struct tcp_sock *tp = tcp_sk(sk);
1467 u32 seq = tp->copied_seq;
1471 if (sk->sk_state == TCP_LISTEN)
1473 while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1474 if (offset < skb->len) {
1478 len = skb->len - offset;
1479 /* Stop reading if we hit a patch of urgent data */
1481 u32 urg_offset = tp->urg_seq - seq;
1482 if (urg_offset < len)
1487 used = recv_actor(desc, skb, offset, len);
1492 } else if (used <= len) {
1498 * If recv_actor drops the lock (e.g. TCP splice
1499 * receive) the skb pointer might be invalid when
1500 * getting here: tcp_collapse might have deleted it
1501 * while aggregating skbs from the socket queue.
1503 skb = tcp_recv_skb(sk, seq-1, &offset);
1504 if (!skb || (offset+1 != skb->len))
1507 if (tcp_hdr(skb)->fin) {
1508 sk_eat_skb(sk, skb, 0);
1512 sk_eat_skb(sk, skb, 0);
1515 tp->copied_seq = seq;
1517 tp->copied_seq = seq;
1519 tcp_rcv_space_adjust(sk);
1521 /* Clean up data we have read: This will do ACK frames. */
1523 tcp_cleanup_rbuf(sk, copied);
1526 EXPORT_SYMBOL(tcp_read_sock);
1529 * This routine copies from a sock struct into the user buffer.
1531 * Technical note: in 2.3 we work on _locked_ socket, so that
1532 * tricks with *seq access order and skb->users are not required.
1533 * Probably, code can be easily improved even more.
1536 int tcp_recvmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg,
1537 size_t len, int nonblock, int flags, int *addr_len)
1539 struct tcp_sock *tp = tcp_sk(sk);
1545 int target; /* Read at least this many bytes */
1547 struct task_struct *user_recv = NULL;
1548 int copied_early = 0;
1549 struct sk_buff *skb;
1555 if (sk->sk_state == TCP_LISTEN)
1558 timeo = sock_rcvtimeo(sk, nonblock);
1560 /* Urgent data needs to be handled specially. */
1561 if (flags & MSG_OOB)
1564 if (unlikely(tp->repair)) {
1566 if (!(flags & MSG_PEEK))
1569 if (tp->repair_queue == TCP_SEND_QUEUE)
1573 if (tp->repair_queue == TCP_NO_QUEUE)
1576 /* 'common' recv queue MSG_PEEK-ing */
1579 seq = &tp->copied_seq;
1580 if (flags & MSG_PEEK) {
1581 peek_seq = tp->copied_seq;
1585 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
1587 #ifdef CONFIG_NET_DMA
1588 tp->ucopy.dma_chan = NULL;
1590 skb = skb_peek_tail(&sk->sk_receive_queue);
1595 available = TCP_SKB_CB(skb)->seq + skb->len - (*seq);
1596 if ((available < target) &&
1597 (len > sysctl_tcp_dma_copybreak) && !(flags & MSG_PEEK) &&
1598 !sysctl_tcp_low_latency &&
1599 net_dma_find_channel()) {
1600 preempt_enable_no_resched();
1601 tp->ucopy.pinned_list =
1602 dma_pin_iovec_pages(msg->msg_iov, len);
1604 preempt_enable_no_resched();
1612 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
1613 if (tp->urg_data && tp->urg_seq == *seq) {
1616 if (signal_pending(current)) {
1617 copied = timeo ? sock_intr_errno(timeo) : -EAGAIN;
1622 /* Next get a buffer. */
1624 skb_queue_walk(&sk->sk_receive_queue, skb) {
1625 /* Now that we have two receive queues this
1628 if (WARN(before(*seq, TCP_SKB_CB(skb)->seq),
1629 "recvmsg bug: copied %X seq %X rcvnxt %X fl %X\n",
1630 *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt,
1634 offset = *seq - TCP_SKB_CB(skb)->seq;
1635 if (tcp_hdr(skb)->syn)
1637 if (offset < skb->len)
1639 if (tcp_hdr(skb)->fin)
1641 WARN(!(flags & MSG_PEEK),
1642 "recvmsg bug 2: copied %X seq %X rcvnxt %X fl %X\n",
1643 *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt, flags);
1646 /* Well, if we have backlog, try to process it now yet. */
1648 if (copied >= target && !sk->sk_backlog.tail)
1653 sk->sk_state == TCP_CLOSE ||
1654 (sk->sk_shutdown & RCV_SHUTDOWN) ||
1656 signal_pending(current))
1659 if (sock_flag(sk, SOCK_DONE))
1663 copied = sock_error(sk);
1667 if (sk->sk_shutdown & RCV_SHUTDOWN)
1670 if (sk->sk_state == TCP_CLOSE) {
1671 if (!sock_flag(sk, SOCK_DONE)) {
1672 /* This occurs when user tries to read
1673 * from never connected socket.
1686 if (signal_pending(current)) {
1687 copied = sock_intr_errno(timeo);
1692 tcp_cleanup_rbuf(sk, copied);
1694 if (!sysctl_tcp_low_latency && tp->ucopy.task == user_recv) {
1695 /* Install new reader */
1696 if (!user_recv && !(flags & (MSG_TRUNC | MSG_PEEK))) {
1697 user_recv = current;
1698 tp->ucopy.task = user_recv;
1699 tp->ucopy.iov = msg->msg_iov;
1702 tp->ucopy.len = len;
1704 WARN_ON(tp->copied_seq != tp->rcv_nxt &&
1705 !(flags & (MSG_PEEK | MSG_TRUNC)));
1707 /* Ugly... If prequeue is not empty, we have to
1708 * process it before releasing socket, otherwise
1709 * order will be broken at second iteration.
1710 * More elegant solution is required!!!
1712 * Look: we have the following (pseudo)queues:
1714 * 1. packets in flight
1719 * Each queue can be processed only if the next ones
1720 * are empty. At this point we have empty receive_queue.
1721 * But prequeue _can_ be not empty after 2nd iteration,
1722 * when we jumped to start of loop because backlog
1723 * processing added something to receive_queue.
1724 * We cannot release_sock(), because backlog contains
1725 * packets arrived _after_ prequeued ones.
1727 * Shortly, algorithm is clear --- to process all
1728 * the queues in order. We could make it more directly,
1729 * requeueing packets from backlog to prequeue, if
1730 * is not empty. It is more elegant, but eats cycles,
1733 if (!skb_queue_empty(&tp->ucopy.prequeue))
1736 /* __ Set realtime policy in scheduler __ */
1739 #ifdef CONFIG_NET_DMA
1740 if (tp->ucopy.dma_chan)
1741 dma_async_memcpy_issue_pending(tp->ucopy.dma_chan);
1743 if (copied >= target) {
1744 /* Do not sleep, just process backlog. */
1748 sk_wait_data(sk, &timeo);
1750 #ifdef CONFIG_NET_DMA
1751 tcp_service_net_dma(sk, false); /* Don't block */
1752 tp->ucopy.wakeup = 0;
1758 /* __ Restore normal policy in scheduler __ */
1760 if ((chunk = len - tp->ucopy.len) != 0) {
1761 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMBACKLOG, chunk);
1766 if (tp->rcv_nxt == tp->copied_seq &&
1767 !skb_queue_empty(&tp->ucopy.prequeue)) {
1769 tcp_prequeue_process(sk);
1771 if ((chunk = len - tp->ucopy.len) != 0) {
1772 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1778 if ((flags & MSG_PEEK) &&
1779 (peek_seq - copied - urg_hole != tp->copied_seq)) {
1780 if (net_ratelimit())
1781 printk(KERN_DEBUG "TCP(%s:%d): Application bug, race in MSG_PEEK.\n",
1782 current->comm, task_pid_nr(current));
1783 peek_seq = tp->copied_seq;
1788 /* Ok so how much can we use? */
1789 used = skb->len - offset;
1793 /* Do we have urgent data here? */
1795 u32 urg_offset = tp->urg_seq - *seq;
1796 if (urg_offset < used) {
1798 if (!sock_flag(sk, SOCK_URGINLINE)) {
1811 if (!(flags & MSG_TRUNC)) {
1812 #ifdef CONFIG_NET_DMA
1813 if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list)
1814 tp->ucopy.dma_chan = net_dma_find_channel();
1816 if (tp->ucopy.dma_chan) {
1817 tp->ucopy.dma_cookie = dma_skb_copy_datagram_iovec(
1818 tp->ucopy.dma_chan, skb, offset,
1820 tp->ucopy.pinned_list);
1822 if (tp->ucopy.dma_cookie < 0) {
1824 pr_alert("%s: dma_cookie < 0\n",
1827 /* Exception. Bailout! */
1833 dma_async_memcpy_issue_pending(tp->ucopy.dma_chan);
1835 if ((offset + used) == skb->len)
1841 err = skb_copy_datagram_iovec(skb, offset,
1842 msg->msg_iov, used);
1844 /* Exception. Bailout! */
1856 tcp_rcv_space_adjust(sk);
1859 if (tp->urg_data && after(tp->copied_seq, tp->urg_seq)) {
1861 tcp_fast_path_check(sk);
1863 if (used + offset < skb->len)
1866 if (tcp_hdr(skb)->fin)
1868 if (!(flags & MSG_PEEK)) {
1869 sk_eat_skb(sk, skb, copied_early);
1875 /* Process the FIN. */
1877 if (!(flags & MSG_PEEK)) {
1878 sk_eat_skb(sk, skb, copied_early);
1885 if (!skb_queue_empty(&tp->ucopy.prequeue)) {
1888 tp->ucopy.len = copied > 0 ? len : 0;
1890 tcp_prequeue_process(sk);
1892 if (copied > 0 && (chunk = len - tp->ucopy.len) != 0) {
1893 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1899 tp->ucopy.task = NULL;
1903 #ifdef CONFIG_NET_DMA
1904 tcp_service_net_dma(sk, true); /* Wait for queue to drain */
1905 tp->ucopy.dma_chan = NULL;
1907 if (tp->ucopy.pinned_list) {
1908 dma_unpin_iovec_pages(tp->ucopy.pinned_list);
1909 tp->ucopy.pinned_list = NULL;
1913 /* According to UNIX98, msg_name/msg_namelen are ignored
1914 * on connected socket. I was just happy when found this 8) --ANK
1917 /* Clean up data we have read: This will do ACK frames. */
1918 tcp_cleanup_rbuf(sk, copied);
1928 err = tcp_recv_urg(sk, msg, len, flags);
1932 err = tcp_peek_sndq(sk, msg, len);
1935 EXPORT_SYMBOL(tcp_recvmsg);
1937 void tcp_set_state(struct sock *sk, int state)
1939 int oldstate = sk->sk_state;
1942 case TCP_ESTABLISHED:
1943 if (oldstate != TCP_ESTABLISHED)
1944 TCP_INC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
1948 if (oldstate == TCP_CLOSE_WAIT || oldstate == TCP_ESTABLISHED)
1949 TCP_INC_STATS(sock_net(sk), TCP_MIB_ESTABRESETS);
1951 sk->sk_prot->unhash(sk);
1952 if (inet_csk(sk)->icsk_bind_hash &&
1953 !(sk->sk_userlocks & SOCK_BINDPORT_LOCK))
1957 if (oldstate == TCP_ESTABLISHED)
1958 TCP_DEC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
1961 /* Change state AFTER socket is unhashed to avoid closed
1962 * socket sitting in hash tables.
1964 sk->sk_state = state;
1967 SOCK_DEBUG(sk, "TCP sk=%p, State %s -> %s\n", sk, statename[oldstate], statename[state]);
1970 EXPORT_SYMBOL_GPL(tcp_set_state);
1973 * State processing on a close. This implements the state shift for
1974 * sending our FIN frame. Note that we only send a FIN for some
1975 * states. A shutdown() may have already sent the FIN, or we may be
1979 static const unsigned char new_state[16] = {
1980 /* current state: new state: action: */
1981 /* (Invalid) */ TCP_CLOSE,
1982 /* TCP_ESTABLISHED */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1983 /* TCP_SYN_SENT */ TCP_CLOSE,
1984 /* TCP_SYN_RECV */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1985 /* TCP_FIN_WAIT1 */ TCP_FIN_WAIT1,
1986 /* TCP_FIN_WAIT2 */ TCP_FIN_WAIT2,
1987 /* TCP_TIME_WAIT */ TCP_CLOSE,
1988 /* TCP_CLOSE */ TCP_CLOSE,
1989 /* TCP_CLOSE_WAIT */ TCP_LAST_ACK | TCP_ACTION_FIN,
1990 /* TCP_LAST_ACK */ TCP_LAST_ACK,
1991 /* TCP_LISTEN */ TCP_CLOSE,
1992 /* TCP_CLOSING */ TCP_CLOSING,
1995 static int tcp_close_state(struct sock *sk)
1997 int next = (int)new_state[sk->sk_state];
1998 int ns = next & TCP_STATE_MASK;
2000 tcp_set_state(sk, ns);
2002 return next & TCP_ACTION_FIN;
2006 * Shutdown the sending side of a connection. Much like close except
2007 * that we don't receive shut down or sock_set_flag(sk, SOCK_DEAD).
2010 void tcp_shutdown(struct sock *sk, int how)
2012 /* We need to grab some memory, and put together a FIN,
2013 * and then put it into the queue to be sent.
2014 * Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
2016 if (!(how & SEND_SHUTDOWN))
2019 /* If we've already sent a FIN, or it's a closed state, skip this. */
2020 if ((1 << sk->sk_state) &
2021 (TCPF_ESTABLISHED | TCPF_SYN_SENT |
2022 TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) {
2023 /* Clear out any half completed packets. FIN if needed. */
2024 if (tcp_close_state(sk))
2028 EXPORT_SYMBOL(tcp_shutdown);
2030 bool tcp_check_oom(struct sock *sk, int shift)
2032 bool too_many_orphans, out_of_socket_memory;
2034 too_many_orphans = tcp_too_many_orphans(sk, shift);
2035 out_of_socket_memory = tcp_out_of_memory(sk);
2037 if (too_many_orphans && net_ratelimit())
2038 pr_info("too many orphaned sockets\n");
2039 if (out_of_socket_memory && net_ratelimit())
2040 pr_info("out of memory -- consider tuning tcp_mem\n");
2041 return too_many_orphans || out_of_socket_memory;
2044 void tcp_close(struct sock *sk, long timeout)
2046 struct sk_buff *skb;
2047 int data_was_unread = 0;
2051 sk->sk_shutdown = SHUTDOWN_MASK;
2053 if (sk->sk_state == TCP_LISTEN) {
2054 tcp_set_state(sk, TCP_CLOSE);
2057 inet_csk_listen_stop(sk);
2059 goto adjudge_to_death;
2062 /* We need to flush the recv. buffs. We do this only on the
2063 * descriptor close, not protocol-sourced closes, because the
2064 * reader process may not have drained the data yet!
2066 while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) {
2067 u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq -
2069 data_was_unread += len;
2075 /* If socket has been already reset (e.g. in tcp_reset()) - kill it. */
2076 if (sk->sk_state == TCP_CLOSE)
2077 goto adjudge_to_death;
2079 /* As outlined in RFC 2525, section 2.17, we send a RST here because
2080 * data was lost. To witness the awful effects of the old behavior of
2081 * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk
2082 * GET in an FTP client, suspend the process, wait for the client to
2083 * advertise a zero window, then kill -9 the FTP client, wheee...
2084 * Note: timeout is always zero in such a case.
2086 if (unlikely(tcp_sk(sk)->repair)) {
2087 sk->sk_prot->disconnect(sk, 0);
2088 } else if (data_was_unread) {
2089 /* Unread data was tossed, zap the connection. */
2090 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONCLOSE);
2091 tcp_set_state(sk, TCP_CLOSE);
2092 tcp_send_active_reset(sk, sk->sk_allocation);
2093 } else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) {
2094 /* Check zero linger _after_ checking for unread data. */
2095 sk->sk_prot->disconnect(sk, 0);
2096 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
2097 } else if (tcp_close_state(sk)) {
2098 /* We FIN if the application ate all the data before
2099 * zapping the connection.
2102 /* RED-PEN. Formally speaking, we have broken TCP state
2103 * machine. State transitions:
2105 * TCP_ESTABLISHED -> TCP_FIN_WAIT1
2106 * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible)
2107 * TCP_CLOSE_WAIT -> TCP_LAST_ACK
2109 * are legal only when FIN has been sent (i.e. in window),
2110 * rather than queued out of window. Purists blame.
2112 * F.e. "RFC state" is ESTABLISHED,
2113 * if Linux state is FIN-WAIT-1, but FIN is still not sent.
2115 * The visible declinations are that sometimes
2116 * we enter time-wait state, when it is not required really
2117 * (harmless), do not send active resets, when they are
2118 * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
2119 * they look as CLOSING or LAST_ACK for Linux)
2120 * Probably, I missed some more holelets.
2126 sk_stream_wait_close(sk, timeout);
2129 state = sk->sk_state;
2133 /* It is the last release_sock in its life. It will remove backlog. */
2137 /* Now socket is owned by kernel and we acquire BH lock
2138 to finish close. No need to check for user refs.
2142 WARN_ON(sock_owned_by_user(sk));
2144 percpu_counter_inc(sk->sk_prot->orphan_count);
2146 /* Have we already been destroyed by a softirq or backlog? */
2147 if (state != TCP_CLOSE && sk->sk_state == TCP_CLOSE)
2150 /* This is a (useful) BSD violating of the RFC. There is a
2151 * problem with TCP as specified in that the other end could
2152 * keep a socket open forever with no application left this end.
2153 * We use a 3 minute timeout (about the same as BSD) then kill
2154 * our end. If they send after that then tough - BUT: long enough
2155 * that we won't make the old 4*rto = almost no time - whoops
2158 * Nope, it was not mistake. It is really desired behaviour
2159 * f.e. on http servers, when such sockets are useless, but
2160 * consume significant resources. Let's do it with special
2161 * linger2 option. --ANK
2164 if (sk->sk_state == TCP_FIN_WAIT2) {
2165 struct tcp_sock *tp = tcp_sk(sk);
2166 if (tp->linger2 < 0) {
2167 tcp_set_state(sk, TCP_CLOSE);
2168 tcp_send_active_reset(sk, GFP_ATOMIC);
2169 NET_INC_STATS_BH(sock_net(sk),
2170 LINUX_MIB_TCPABORTONLINGER);
2172 const int tmo = tcp_fin_time(sk);
2174 if (tmo > TCP_TIMEWAIT_LEN) {
2175 inet_csk_reset_keepalive_timer(sk,
2176 tmo - TCP_TIMEWAIT_LEN);
2178 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
2183 if (sk->sk_state != TCP_CLOSE) {
2185 if (tcp_check_oom(sk, 0)) {
2186 tcp_set_state(sk, TCP_CLOSE);
2187 tcp_send_active_reset(sk, GFP_ATOMIC);
2188 NET_INC_STATS_BH(sock_net(sk),
2189 LINUX_MIB_TCPABORTONMEMORY);
2193 if (sk->sk_state == TCP_CLOSE)
2194 inet_csk_destroy_sock(sk);
2195 /* Otherwise, socket is reprieved until protocol close. */
2202 EXPORT_SYMBOL(tcp_close);
2204 /* These states need RST on ABORT according to RFC793 */
2206 static inline int tcp_need_reset(int state)
2208 return (1 << state) &
2209 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 |
2210 TCPF_FIN_WAIT2 | TCPF_SYN_RECV);
2213 int tcp_disconnect(struct sock *sk, int flags)
2215 struct inet_sock *inet = inet_sk(sk);
2216 struct inet_connection_sock *icsk = inet_csk(sk);
2217 struct tcp_sock *tp = tcp_sk(sk);
2219 int old_state = sk->sk_state;
2221 if (old_state != TCP_CLOSE)
2222 tcp_set_state(sk, TCP_CLOSE);
2224 /* ABORT function of RFC793 */
2225 if (old_state == TCP_LISTEN) {
2226 inet_csk_listen_stop(sk);
2227 } else if (unlikely(tp->repair)) {
2228 sk->sk_err = ECONNABORTED;
2229 } else if (tcp_need_reset(old_state) ||
2230 (tp->snd_nxt != tp->write_seq &&
2231 (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK))) {
2232 /* The last check adjusts for discrepancy of Linux wrt. RFC
2235 tcp_send_active_reset(sk, gfp_any());
2236 sk->sk_err = ECONNRESET;
2237 } else if (old_state == TCP_SYN_SENT)
2238 sk->sk_err = ECONNRESET;
2240 tcp_clear_xmit_timers(sk);
2241 __skb_queue_purge(&sk->sk_receive_queue);
2242 tcp_write_queue_purge(sk);
2243 __skb_queue_purge(&tp->out_of_order_queue);
2244 #ifdef CONFIG_NET_DMA
2245 __skb_queue_purge(&sk->sk_async_wait_queue);
2248 inet->inet_dport = 0;
2250 if (!(sk->sk_userlocks & SOCK_BINDADDR_LOCK))
2251 inet_reset_saddr(sk);
2253 sk->sk_shutdown = 0;
2254 sock_reset_flag(sk, SOCK_DONE);
2256 if ((tp->write_seq += tp->max_window + 2) == 0)
2258 icsk->icsk_backoff = 0;
2260 icsk->icsk_probes_out = 0;
2261 tp->packets_out = 0;
2262 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
2263 tp->snd_cwnd_cnt = 0;
2264 tp->bytes_acked = 0;
2265 tp->window_clamp = 0;
2266 tcp_set_ca_state(sk, TCP_CA_Open);
2267 tcp_clear_retrans(tp);
2268 inet_csk_delack_init(sk);
2269 tcp_init_send_head(sk);
2270 memset(&tp->rx_opt, 0, sizeof(tp->rx_opt));
2273 WARN_ON(inet->inet_num && !icsk->icsk_bind_hash);
2275 sk->sk_error_report(sk);
2278 EXPORT_SYMBOL(tcp_disconnect);
2280 static inline int tcp_can_repair_sock(struct sock *sk)
2282 return capable(CAP_NET_ADMIN) &&
2283 ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_ESTABLISHED));
2286 static int tcp_repair_options_est(struct tcp_sock *tp,
2287 struct tcp_repair_opt __user *optbuf, unsigned int len)
2289 struct tcp_repair_opt opt;
2291 while (len >= sizeof(opt)) {
2292 if (copy_from_user(&opt, optbuf, sizeof(opt)))
2298 switch (opt.opt_code) {
2300 tp->rx_opt.mss_clamp = opt.opt_val;
2303 if (opt.opt_val > 14)
2306 tp->rx_opt.snd_wscale = opt.opt_val;
2308 case TCPOPT_SACK_PERM:
2309 if (opt.opt_val != 0)
2312 tp->rx_opt.sack_ok |= TCP_SACK_SEEN;
2313 if (sysctl_tcp_fack)
2314 tcp_enable_fack(tp);
2316 case TCPOPT_TIMESTAMP:
2317 if (opt.opt_val != 0)
2320 tp->rx_opt.tstamp_ok = 1;
2329 * Socket option code for TCP.
2331 static int do_tcp_setsockopt(struct sock *sk, int level,
2332 int optname, char __user *optval, unsigned int optlen)
2334 struct tcp_sock *tp = tcp_sk(sk);
2335 struct inet_connection_sock *icsk = inet_csk(sk);
2339 /* These are data/string values, all the others are ints */
2341 case TCP_CONGESTION: {
2342 char name[TCP_CA_NAME_MAX];
2347 val = strncpy_from_user(name, optval,
2348 min_t(long, TCP_CA_NAME_MAX-1, optlen));
2354 err = tcp_set_congestion_control(sk, name);
2358 case TCP_COOKIE_TRANSACTIONS: {
2359 struct tcp_cookie_transactions ctd;
2360 struct tcp_cookie_values *cvp = NULL;
2362 if (sizeof(ctd) > optlen)
2364 if (copy_from_user(&ctd, optval, sizeof(ctd)))
2367 if (ctd.tcpct_used > sizeof(ctd.tcpct_value) ||
2368 ctd.tcpct_s_data_desired > TCP_MSS_DESIRED)
2371 if (ctd.tcpct_cookie_desired == 0) {
2372 /* default to global value */
2373 } else if ((0x1 & ctd.tcpct_cookie_desired) ||
2374 ctd.tcpct_cookie_desired > TCP_COOKIE_MAX ||
2375 ctd.tcpct_cookie_desired < TCP_COOKIE_MIN) {
2379 if (TCP_COOKIE_OUT_NEVER & ctd.tcpct_flags) {
2380 /* Supercedes all other values */
2382 if (tp->cookie_values != NULL) {
2383 kref_put(&tp->cookie_values->kref,
2384 tcp_cookie_values_release);
2385 tp->cookie_values = NULL;
2387 tp->rx_opt.cookie_in_always = 0; /* false */
2388 tp->rx_opt.cookie_out_never = 1; /* true */
2393 /* Allocate ancillary memory before locking.
2395 if (ctd.tcpct_used > 0 ||
2396 (tp->cookie_values == NULL &&
2397 (sysctl_tcp_cookie_size > 0 ||
2398 ctd.tcpct_cookie_desired > 0 ||
2399 ctd.tcpct_s_data_desired > 0))) {
2400 cvp = kzalloc(sizeof(*cvp) + ctd.tcpct_used,
2405 kref_init(&cvp->kref);
2408 tp->rx_opt.cookie_in_always =
2409 (TCP_COOKIE_IN_ALWAYS & ctd.tcpct_flags);
2410 tp->rx_opt.cookie_out_never = 0; /* false */
2412 if (tp->cookie_values != NULL) {
2414 /* Changed values are recorded by a changed
2415 * pointer, ensuring the cookie will differ,
2416 * without separately hashing each value later.
2418 kref_put(&tp->cookie_values->kref,
2419 tcp_cookie_values_release);
2421 cvp = tp->cookie_values;
2426 cvp->cookie_desired = ctd.tcpct_cookie_desired;
2428 if (ctd.tcpct_used > 0) {
2429 memcpy(cvp->s_data_payload, ctd.tcpct_value,
2431 cvp->s_data_desired = ctd.tcpct_used;
2432 cvp->s_data_constant = 1; /* true */
2434 /* No constant payload data. */
2435 cvp->s_data_desired = ctd.tcpct_s_data_desired;
2436 cvp->s_data_constant = 0; /* false */
2439 tp->cookie_values = cvp;
2449 if (optlen < sizeof(int))
2452 if (get_user(val, (int __user *)optval))
2459 /* Values greater than interface MTU won't take effect. However
2460 * at the point when this call is done we typically don't yet
2461 * know which interface is going to be used */
2462 if (val < TCP_MIN_MSS || val > MAX_TCP_WINDOW) {
2466 tp->rx_opt.user_mss = val;
2471 /* TCP_NODELAY is weaker than TCP_CORK, so that
2472 * this option on corked socket is remembered, but
2473 * it is not activated until cork is cleared.
2475 * However, when TCP_NODELAY is set we make
2476 * an explicit push, which overrides even TCP_CORK
2477 * for currently queued segments.
2479 tp->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH;
2480 tcp_push_pending_frames(sk);
2482 tp->nonagle &= ~TCP_NAGLE_OFF;
2486 case TCP_THIN_LINEAR_TIMEOUTS:
2487 if (val < 0 || val > 1)
2493 case TCP_THIN_DUPACK:
2494 if (val < 0 || val > 1)
2497 tp->thin_dupack = val;
2501 if (!tcp_can_repair_sock(sk))
2503 else if (val == 1) {
2505 sk->sk_reuse = SK_FORCE_REUSE;
2506 tp->repair_queue = TCP_NO_QUEUE;
2507 } else if (val == 0) {
2509 sk->sk_reuse = SK_NO_REUSE;
2510 tcp_send_window_probe(sk);
2516 case TCP_REPAIR_QUEUE:
2519 else if (val < TCP_QUEUES_NR)
2520 tp->repair_queue = val;
2526 if (sk->sk_state != TCP_CLOSE)
2528 else if (tp->repair_queue == TCP_SEND_QUEUE)
2529 tp->write_seq = val;
2530 else if (tp->repair_queue == TCP_RECV_QUEUE)
2536 case TCP_REPAIR_OPTIONS:
2539 else if (sk->sk_state == TCP_ESTABLISHED)
2540 err = tcp_repair_options_est(tp,
2541 (struct tcp_repair_opt __user *)optval,
2548 /* When set indicates to always queue non-full frames.
2549 * Later the user clears this option and we transmit
2550 * any pending partial frames in the queue. This is
2551 * meant to be used alongside sendfile() to get properly
2552 * filled frames when the user (for example) must write
2553 * out headers with a write() call first and then use
2554 * sendfile to send out the data parts.
2556 * TCP_CORK can be set together with TCP_NODELAY and it is
2557 * stronger than TCP_NODELAY.
2560 tp->nonagle |= TCP_NAGLE_CORK;
2562 tp->nonagle &= ~TCP_NAGLE_CORK;
2563 if (tp->nonagle&TCP_NAGLE_OFF)
2564 tp->nonagle |= TCP_NAGLE_PUSH;
2565 tcp_push_pending_frames(sk);
2570 if (val < 1 || val > MAX_TCP_KEEPIDLE)
2573 tp->keepalive_time = val * HZ;
2574 if (sock_flag(sk, SOCK_KEEPOPEN) &&
2575 !((1 << sk->sk_state) &
2576 (TCPF_CLOSE | TCPF_LISTEN))) {
2577 u32 elapsed = keepalive_time_elapsed(tp);
2578 if (tp->keepalive_time > elapsed)
2579 elapsed = tp->keepalive_time - elapsed;
2582 inet_csk_reset_keepalive_timer(sk, elapsed);
2587 if (val < 1 || val > MAX_TCP_KEEPINTVL)
2590 tp->keepalive_intvl = val * HZ;
2593 if (val < 1 || val > MAX_TCP_KEEPCNT)
2596 tp->keepalive_probes = val;
2599 if (val < 1 || val > MAX_TCP_SYNCNT)
2602 icsk->icsk_syn_retries = val;
2608 else if (val > sysctl_tcp_fin_timeout / HZ)
2611 tp->linger2 = val * HZ;
2614 case TCP_DEFER_ACCEPT:
2615 /* Translate value in seconds to number of retransmits */
2616 icsk->icsk_accept_queue.rskq_defer_accept =
2617 secs_to_retrans(val, TCP_TIMEOUT_INIT / HZ,
2621 case TCP_WINDOW_CLAMP:
2623 if (sk->sk_state != TCP_CLOSE) {
2627 tp->window_clamp = 0;
2629 tp->window_clamp = val < SOCK_MIN_RCVBUF / 2 ?
2630 SOCK_MIN_RCVBUF / 2 : val;
2635 icsk->icsk_ack.pingpong = 1;
2637 icsk->icsk_ack.pingpong = 0;
2638 if ((1 << sk->sk_state) &
2639 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) &&
2640 inet_csk_ack_scheduled(sk)) {
2641 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
2642 tcp_cleanup_rbuf(sk, 1);
2644 icsk->icsk_ack.pingpong = 1;
2649 #ifdef CONFIG_TCP_MD5SIG
2651 /* Read the IP->Key mappings from userspace */
2652 err = tp->af_specific->md5_parse(sk, optval, optlen);
2655 case TCP_USER_TIMEOUT:
2656 /* Cap the max timeout in ms TCP will retry/retrans
2657 * before giving up and aborting (ETIMEDOUT) a connection.
2659 icsk->icsk_user_timeout = msecs_to_jiffies(val);
2670 int tcp_setsockopt(struct sock *sk, int level, int optname, char __user *optval,
2671 unsigned int optlen)
2673 const struct inet_connection_sock *icsk = inet_csk(sk);
2675 if (level != SOL_TCP)
2676 return icsk->icsk_af_ops->setsockopt(sk, level, optname,
2678 return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2680 EXPORT_SYMBOL(tcp_setsockopt);
2682 #ifdef CONFIG_COMPAT
2683 int compat_tcp_setsockopt(struct sock *sk, int level, int optname,
2684 char __user *optval, unsigned int optlen)
2686 if (level != SOL_TCP)
2687 return inet_csk_compat_setsockopt(sk, level, optname,
2689 return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2691 EXPORT_SYMBOL(compat_tcp_setsockopt);
2694 /* Return information about state of tcp endpoint in API format. */
2695 void tcp_get_info(const struct sock *sk, struct tcp_info *info)
2697 const struct tcp_sock *tp = tcp_sk(sk);
2698 const struct inet_connection_sock *icsk = inet_csk(sk);
2699 u32 now = tcp_time_stamp;
2701 memset(info, 0, sizeof(*info));
2703 info->tcpi_state = sk->sk_state;
2704 info->tcpi_ca_state = icsk->icsk_ca_state;
2705 info->tcpi_retransmits = icsk->icsk_retransmits;
2706 info->tcpi_probes = icsk->icsk_probes_out;
2707 info->tcpi_backoff = icsk->icsk_backoff;
2709 if (tp->rx_opt.tstamp_ok)
2710 info->tcpi_options |= TCPI_OPT_TIMESTAMPS;
2711 if (tcp_is_sack(tp))
2712 info->tcpi_options |= TCPI_OPT_SACK;
2713 if (tp->rx_opt.wscale_ok) {
2714 info->tcpi_options |= TCPI_OPT_WSCALE;
2715 info->tcpi_snd_wscale = tp->rx_opt.snd_wscale;
2716 info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale;
2719 if (tp->ecn_flags & TCP_ECN_OK)
2720 info->tcpi_options |= TCPI_OPT_ECN;
2721 if (tp->ecn_flags & TCP_ECN_SEEN)
2722 info->tcpi_options |= TCPI_OPT_ECN_SEEN;
2724 info->tcpi_rto = jiffies_to_usecs(icsk->icsk_rto);
2725 info->tcpi_ato = jiffies_to_usecs(icsk->icsk_ack.ato);
2726 info->tcpi_snd_mss = tp->mss_cache;
2727 info->tcpi_rcv_mss = icsk->icsk_ack.rcv_mss;
2729 if (sk->sk_state == TCP_LISTEN) {
2730 info->tcpi_unacked = sk->sk_ack_backlog;
2731 info->tcpi_sacked = sk->sk_max_ack_backlog;
2733 info->tcpi_unacked = tp->packets_out;
2734 info->tcpi_sacked = tp->sacked_out;
2736 info->tcpi_lost = tp->lost_out;
2737 info->tcpi_retrans = tp->retrans_out;
2738 info->tcpi_fackets = tp->fackets_out;
2740 info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime);
2741 info->tcpi_last_data_recv = jiffies_to_msecs(now - icsk->icsk_ack.lrcvtime);
2742 info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp);
2744 info->tcpi_pmtu = icsk->icsk_pmtu_cookie;
2745 info->tcpi_rcv_ssthresh = tp->rcv_ssthresh;
2746 info->tcpi_rtt = jiffies_to_usecs(tp->srtt)>>3;
2747 info->tcpi_rttvar = jiffies_to_usecs(tp->mdev)>>2;
2748 info->tcpi_snd_ssthresh = tp->snd_ssthresh;
2749 info->tcpi_snd_cwnd = tp->snd_cwnd;
2750 info->tcpi_advmss = tp->advmss;
2751 info->tcpi_reordering = tp->reordering;
2753 info->tcpi_rcv_rtt = jiffies_to_usecs(tp->rcv_rtt_est.rtt)>>3;
2754 info->tcpi_rcv_space = tp->rcvq_space.space;
2756 info->tcpi_total_retrans = tp->total_retrans;
2758 EXPORT_SYMBOL_GPL(tcp_get_info);
2760 static int do_tcp_getsockopt(struct sock *sk, int level,
2761 int optname, char __user *optval, int __user *optlen)
2763 struct inet_connection_sock *icsk = inet_csk(sk);
2764 struct tcp_sock *tp = tcp_sk(sk);
2767 if (get_user(len, optlen))
2770 len = min_t(unsigned int, len, sizeof(int));
2777 val = tp->mss_cache;
2778 if (!val && ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
2779 val = tp->rx_opt.user_mss;
2781 val = tp->rx_opt.mss_clamp;
2784 val = !!(tp->nonagle&TCP_NAGLE_OFF);
2787 val = !!(tp->nonagle&TCP_NAGLE_CORK);
2790 val = keepalive_time_when(tp) / HZ;
2793 val = keepalive_intvl_when(tp) / HZ;
2796 val = keepalive_probes(tp);
2799 val = icsk->icsk_syn_retries ? : sysctl_tcp_syn_retries;
2804 val = (val ? : sysctl_tcp_fin_timeout) / HZ;
2806 case TCP_DEFER_ACCEPT:
2807 val = retrans_to_secs(icsk->icsk_accept_queue.rskq_defer_accept,
2808 TCP_TIMEOUT_INIT / HZ, TCP_RTO_MAX / HZ);
2810 case TCP_WINDOW_CLAMP:
2811 val = tp->window_clamp;
2814 struct tcp_info info;
2816 if (get_user(len, optlen))
2819 tcp_get_info(sk, &info);
2821 len = min_t(unsigned int, len, sizeof(info));
2822 if (put_user(len, optlen))
2824 if (copy_to_user(optval, &info, len))
2829 val = !icsk->icsk_ack.pingpong;
2832 case TCP_CONGESTION:
2833 if (get_user(len, optlen))
2835 len = min_t(unsigned int, len, TCP_CA_NAME_MAX);
2836 if (put_user(len, optlen))
2838 if (copy_to_user(optval, icsk->icsk_ca_ops->name, len))
2842 case TCP_COOKIE_TRANSACTIONS: {
2843 struct tcp_cookie_transactions ctd;
2844 struct tcp_cookie_values *cvp = tp->cookie_values;
2846 if (get_user(len, optlen))
2848 if (len < sizeof(ctd))
2851 memset(&ctd, 0, sizeof(ctd));
2852 ctd.tcpct_flags = (tp->rx_opt.cookie_in_always ?
2853 TCP_COOKIE_IN_ALWAYS : 0)
2854 | (tp->rx_opt.cookie_out_never ?
2855 TCP_COOKIE_OUT_NEVER : 0);
2858 ctd.tcpct_flags |= (cvp->s_data_in ?
2860 | (cvp->s_data_out ?
2861 TCP_S_DATA_OUT : 0);
2863 ctd.tcpct_cookie_desired = cvp->cookie_desired;
2864 ctd.tcpct_s_data_desired = cvp->s_data_desired;
2866 memcpy(&ctd.tcpct_value[0], &cvp->cookie_pair[0],
2867 cvp->cookie_pair_size);
2868 ctd.tcpct_used = cvp->cookie_pair_size;
2871 if (put_user(sizeof(ctd), optlen))
2873 if (copy_to_user(optval, &ctd, sizeof(ctd)))
2877 case TCP_THIN_LINEAR_TIMEOUTS:
2880 case TCP_THIN_DUPACK:
2881 val = tp->thin_dupack;
2888 case TCP_REPAIR_QUEUE:
2890 val = tp->repair_queue;
2896 if (tp->repair_queue == TCP_SEND_QUEUE)
2897 val = tp->write_seq;
2898 else if (tp->repair_queue == TCP_RECV_QUEUE)
2904 case TCP_USER_TIMEOUT:
2905 val = jiffies_to_msecs(icsk->icsk_user_timeout);
2908 return -ENOPROTOOPT;
2911 if (put_user(len, optlen))
2913 if (copy_to_user(optval, &val, len))
2918 int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval,
2921 struct inet_connection_sock *icsk = inet_csk(sk);
2923 if (level != SOL_TCP)
2924 return icsk->icsk_af_ops->getsockopt(sk, level, optname,
2926 return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2928 EXPORT_SYMBOL(tcp_getsockopt);
2930 #ifdef CONFIG_COMPAT
2931 int compat_tcp_getsockopt(struct sock *sk, int level, int optname,
2932 char __user *optval, int __user *optlen)
2934 if (level != SOL_TCP)
2935 return inet_csk_compat_getsockopt(sk, level, optname,
2937 return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2939 EXPORT_SYMBOL(compat_tcp_getsockopt);
2942 struct sk_buff *tcp_tso_segment(struct sk_buff *skb,
2943 netdev_features_t features)
2945 struct sk_buff *segs = ERR_PTR(-EINVAL);
2950 unsigned int oldlen;
2953 if (!pskb_may_pull(skb, sizeof(*th)))
2957 thlen = th->doff * 4;
2958 if (thlen < sizeof(*th))
2961 if (!pskb_may_pull(skb, thlen))
2964 oldlen = (u16)~skb->len;
2965 __skb_pull(skb, thlen);
2967 mss = skb_shinfo(skb)->gso_size;
2968 if (unlikely(skb->len <= mss))
2971 if (skb_gso_ok(skb, features | NETIF_F_GSO_ROBUST)) {
2972 /* Packet is from an untrusted source, reset gso_segs. */
2973 int type = skb_shinfo(skb)->gso_type;
2981 !(type & (SKB_GSO_TCPV4 | SKB_GSO_TCPV6))))
2984 skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss);
2990 segs = skb_segment(skb, features);
2994 delta = htonl(oldlen + (thlen + mss));
2998 seq = ntohl(th->seq);
3001 th->fin = th->psh = 0;
3003 th->check = ~csum_fold((__force __wsum)((__force u32)th->check +
3004 (__force u32)delta));
3005 if (skb->ip_summed != CHECKSUM_PARTIAL)
3007 csum_fold(csum_partial(skb_transport_header(skb),
3014 th->seq = htonl(seq);
3016 } while (skb->next);
3018 delta = htonl(oldlen + (skb->tail - skb->transport_header) +
3020 th->check = ~csum_fold((__force __wsum)((__force u32)th->check +
3021 (__force u32)delta));
3022 if (skb->ip_summed != CHECKSUM_PARTIAL)
3023 th->check = csum_fold(csum_partial(skb_transport_header(skb),
3029 EXPORT_SYMBOL(tcp_tso_segment);
3031 struct sk_buff **tcp_gro_receive(struct sk_buff **head, struct sk_buff *skb)
3033 struct sk_buff **pp = NULL;
3040 unsigned int mss = 1;
3046 off = skb_gro_offset(skb);
3047 hlen = off + sizeof(*th);
3048 th = skb_gro_header_fast(skb, off);
3049 if (skb_gro_header_hard(skb, hlen)) {
3050 th = skb_gro_header_slow(skb, hlen, off);
3055 thlen = th->doff * 4;
3056 if (thlen < sizeof(*th))
3060 if (skb_gro_header_hard(skb, hlen)) {
3061 th = skb_gro_header_slow(skb, hlen, off);
3066 skb_gro_pull(skb, thlen);
3068 len = skb_gro_len(skb);
3069 flags = tcp_flag_word(th);
3071 for (; (p = *head); head = &p->next) {
3072 if (!NAPI_GRO_CB(p)->same_flow)
3077 if (*(u32 *)&th->source ^ *(u32 *)&th2->source) {
3078 NAPI_GRO_CB(p)->same_flow = 0;
3085 goto out_check_final;
3088 flush = NAPI_GRO_CB(p)->flush;
3089 flush |= (__force int)(flags & TCP_FLAG_CWR);
3090 flush |= (__force int)((flags ^ tcp_flag_word(th2)) &
3091 ~(TCP_FLAG_CWR | TCP_FLAG_FIN | TCP_FLAG_PSH));
3092 flush |= (__force int)(th->ack_seq ^ th2->ack_seq);
3093 for (i = sizeof(*th); i < thlen; i += 4)
3094 flush |= *(u32 *)((u8 *)th + i) ^
3095 *(u32 *)((u8 *)th2 + i);
3097 mss = skb_shinfo(p)->gso_size;
3099 flush |= (len - 1) >= mss;
3100 flush |= (ntohl(th2->seq) + skb_gro_len(p)) ^ ntohl(th->seq);
3102 if (flush || skb_gro_receive(head, skb)) {
3104 goto out_check_final;
3109 tcp_flag_word(th2) |= flags & (TCP_FLAG_FIN | TCP_FLAG_PSH);
3113 flush |= (__force int)(flags & (TCP_FLAG_URG | TCP_FLAG_PSH |
3114 TCP_FLAG_RST | TCP_FLAG_SYN |
3117 if (p && (!NAPI_GRO_CB(skb)->same_flow || flush))
3121 NAPI_GRO_CB(skb)->flush |= flush;
3125 EXPORT_SYMBOL(tcp_gro_receive);
3127 int tcp_gro_complete(struct sk_buff *skb)
3129 struct tcphdr *th = tcp_hdr(skb);
3131 skb->csum_start = skb_transport_header(skb) - skb->head;
3132 skb->csum_offset = offsetof(struct tcphdr, check);
3133 skb->ip_summed = CHECKSUM_PARTIAL;
3135 skb_shinfo(skb)->gso_segs = NAPI_GRO_CB(skb)->count;
3138 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
3142 EXPORT_SYMBOL(tcp_gro_complete);
3144 #ifdef CONFIG_TCP_MD5SIG
3145 static unsigned long tcp_md5sig_users;
3146 static struct tcp_md5sig_pool __percpu *tcp_md5sig_pool;
3147 static DEFINE_SPINLOCK(tcp_md5sig_pool_lock);
3149 static void __tcp_free_md5sig_pool(struct tcp_md5sig_pool __percpu *pool)
3153 for_each_possible_cpu(cpu) {
3154 struct tcp_md5sig_pool *p = per_cpu_ptr(pool, cpu);
3156 if (p->md5_desc.tfm)
3157 crypto_free_hash(p->md5_desc.tfm);
3162 void tcp_free_md5sig_pool(void)
3164 struct tcp_md5sig_pool __percpu *pool = NULL;
3166 spin_lock_bh(&tcp_md5sig_pool_lock);
3167 if (--tcp_md5sig_users == 0) {
3168 pool = tcp_md5sig_pool;
3169 tcp_md5sig_pool = NULL;
3171 spin_unlock_bh(&tcp_md5sig_pool_lock);
3173 __tcp_free_md5sig_pool(pool);
3175 EXPORT_SYMBOL(tcp_free_md5sig_pool);
3177 static struct tcp_md5sig_pool __percpu *
3178 __tcp_alloc_md5sig_pool(struct sock *sk)
3181 struct tcp_md5sig_pool __percpu *pool;
3183 pool = alloc_percpu(struct tcp_md5sig_pool);
3187 for_each_possible_cpu(cpu) {
3188 struct crypto_hash *hash;
3190 hash = crypto_alloc_hash("md5", 0, CRYPTO_ALG_ASYNC);
3191 if (!hash || IS_ERR(hash))
3194 per_cpu_ptr(pool, cpu)->md5_desc.tfm = hash;
3198 __tcp_free_md5sig_pool(pool);
3202 struct tcp_md5sig_pool __percpu *tcp_alloc_md5sig_pool(struct sock *sk)
3204 struct tcp_md5sig_pool __percpu *pool;
3208 spin_lock_bh(&tcp_md5sig_pool_lock);
3209 pool = tcp_md5sig_pool;
3210 if (tcp_md5sig_users++ == 0) {
3212 spin_unlock_bh(&tcp_md5sig_pool_lock);
3215 spin_unlock_bh(&tcp_md5sig_pool_lock);
3219 spin_unlock_bh(&tcp_md5sig_pool_lock);
3222 /* we cannot hold spinlock here because this may sleep. */
3223 struct tcp_md5sig_pool __percpu *p;
3225 p = __tcp_alloc_md5sig_pool(sk);
3226 spin_lock_bh(&tcp_md5sig_pool_lock);
3229 spin_unlock_bh(&tcp_md5sig_pool_lock);
3232 pool = tcp_md5sig_pool;
3234 /* oops, it has already been assigned. */
3235 spin_unlock_bh(&tcp_md5sig_pool_lock);
3236 __tcp_free_md5sig_pool(p);
3238 tcp_md5sig_pool = pool = p;
3239 spin_unlock_bh(&tcp_md5sig_pool_lock);
3244 EXPORT_SYMBOL(tcp_alloc_md5sig_pool);
3248 * tcp_get_md5sig_pool - get md5sig_pool for this user
3250 * We use percpu structure, so if we succeed, we exit with preemption
3251 * and BH disabled, to make sure another thread or softirq handling
3252 * wont try to get same context.
3254 struct tcp_md5sig_pool *tcp_get_md5sig_pool(void)
3256 struct tcp_md5sig_pool __percpu *p;
3260 spin_lock(&tcp_md5sig_pool_lock);
3261 p = tcp_md5sig_pool;
3264 spin_unlock(&tcp_md5sig_pool_lock);
3267 return this_cpu_ptr(p);
3272 EXPORT_SYMBOL(tcp_get_md5sig_pool);
3274 void tcp_put_md5sig_pool(void)
3277 tcp_free_md5sig_pool();
3279 EXPORT_SYMBOL(tcp_put_md5sig_pool);
3281 int tcp_md5_hash_header(struct tcp_md5sig_pool *hp,
3282 const struct tcphdr *th)
3284 struct scatterlist sg;
3288 /* We are not allowed to change tcphdr, make a local copy */
3289 memcpy(&hdr, th, sizeof(hdr));
3292 /* options aren't included in the hash */
3293 sg_init_one(&sg, &hdr, sizeof(hdr));
3294 err = crypto_hash_update(&hp->md5_desc, &sg, sizeof(hdr));
3297 EXPORT_SYMBOL(tcp_md5_hash_header);
3299 int tcp_md5_hash_skb_data(struct tcp_md5sig_pool *hp,
3300 const struct sk_buff *skb, unsigned int header_len)
3302 struct scatterlist sg;
3303 const struct tcphdr *tp = tcp_hdr(skb);
3304 struct hash_desc *desc = &hp->md5_desc;
3306 const unsigned int head_data_len = skb_headlen(skb) > header_len ?
3307 skb_headlen(skb) - header_len : 0;
3308 const struct skb_shared_info *shi = skb_shinfo(skb);
3309 struct sk_buff *frag_iter;
3311 sg_init_table(&sg, 1);
3313 sg_set_buf(&sg, ((u8 *) tp) + header_len, head_data_len);
3314 if (crypto_hash_update(desc, &sg, head_data_len))
3317 for (i = 0; i < shi->nr_frags; ++i) {
3318 const struct skb_frag_struct *f = &shi->frags[i];
3319 struct page *page = skb_frag_page(f);
3320 sg_set_page(&sg, page, skb_frag_size(f), f->page_offset);
3321 if (crypto_hash_update(desc, &sg, skb_frag_size(f)))
3325 skb_walk_frags(skb, frag_iter)
3326 if (tcp_md5_hash_skb_data(hp, frag_iter, 0))
3331 EXPORT_SYMBOL(tcp_md5_hash_skb_data);
3333 int tcp_md5_hash_key(struct tcp_md5sig_pool *hp, const struct tcp_md5sig_key *key)
3335 struct scatterlist sg;
3337 sg_init_one(&sg, key->key, key->keylen);
3338 return crypto_hash_update(&hp->md5_desc, &sg, key->keylen);
3340 EXPORT_SYMBOL(tcp_md5_hash_key);
3345 * Each Responder maintains up to two secret values concurrently for
3346 * efficient secret rollover. Each secret value has 4 states:
3348 * Generating. (tcp_secret_generating != tcp_secret_primary)
3349 * Generates new Responder-Cookies, but not yet used for primary
3350 * verification. This is a short-term state, typically lasting only
3351 * one round trip time (RTT).
3353 * Primary. (tcp_secret_generating == tcp_secret_primary)
3354 * Used both for generation and primary verification.
3356 * Retiring. (tcp_secret_retiring != tcp_secret_secondary)
3357 * Used for verification, until the first failure that can be
3358 * verified by the newer Generating secret. At that time, this
3359 * cookie's state is changed to Secondary, and the Generating
3360 * cookie's state is changed to Primary. This is a short-term state,
3361 * typically lasting only one round trip time (RTT).
3363 * Secondary. (tcp_secret_retiring == tcp_secret_secondary)
3364 * Used for secondary verification, after primary verification
3365 * failures. This state lasts no more than twice the Maximum Segment
3366 * Lifetime (2MSL). Then, the secret is discarded.
3368 struct tcp_cookie_secret {
3369 /* The secret is divided into two parts. The digest part is the
3370 * equivalent of previously hashing a secret and saving the state,
3371 * and serves as an initialization vector (IV). The message part
3372 * serves as the trailing secret.
3374 u32 secrets[COOKIE_WORKSPACE_WORDS];
3375 unsigned long expires;
3378 #define TCP_SECRET_1MSL (HZ * TCP_PAWS_MSL)
3379 #define TCP_SECRET_2MSL (HZ * TCP_PAWS_MSL * 2)
3380 #define TCP_SECRET_LIFE (HZ * 600)
3382 static struct tcp_cookie_secret tcp_secret_one;
3383 static struct tcp_cookie_secret tcp_secret_two;
3385 /* Essentially a circular list, without dynamic allocation. */
3386 static struct tcp_cookie_secret *tcp_secret_generating;
3387 static struct tcp_cookie_secret *tcp_secret_primary;
3388 static struct tcp_cookie_secret *tcp_secret_retiring;
3389 static struct tcp_cookie_secret *tcp_secret_secondary;
3391 static DEFINE_SPINLOCK(tcp_secret_locker);
3393 /* Select a pseudo-random word in the cookie workspace.
3395 static inline u32 tcp_cookie_work(const u32 *ws, const int n)
3397 return ws[COOKIE_DIGEST_WORDS + ((COOKIE_MESSAGE_WORDS-1) & ws[n])];
3400 /* Fill bakery[COOKIE_WORKSPACE_WORDS] with generator, updating as needed.
3401 * Called in softirq context.
3402 * Returns: 0 for success.
3404 int tcp_cookie_generator(u32 *bakery)
3406 unsigned long jiffy = jiffies;
3408 if (unlikely(time_after_eq(jiffy, tcp_secret_generating->expires))) {
3409 spin_lock_bh(&tcp_secret_locker);
3410 if (!time_after_eq(jiffy, tcp_secret_generating->expires)) {
3411 /* refreshed by another */
3413 &tcp_secret_generating->secrets[0],
3414 COOKIE_WORKSPACE_WORDS);
3416 /* still needs refreshing */
3417 get_random_bytes(bakery, COOKIE_WORKSPACE_WORDS);
3419 /* The first time, paranoia assumes that the
3420 * randomization function isn't as strong. But,
3421 * this secret initialization is delayed until
3422 * the last possible moment (packet arrival).
3423 * Although that time is observable, it is
3424 * unpredictably variable. Mash in the most
3425 * volatile clock bits available, and expire the
3426 * secret extra quickly.
3428 if (unlikely(tcp_secret_primary->expires ==
3429 tcp_secret_secondary->expires)) {
3432 getnstimeofday(&tv);
3433 bakery[COOKIE_DIGEST_WORDS+0] ^=
3436 tcp_secret_secondary->expires = jiffy
3438 + (0x0f & tcp_cookie_work(bakery, 0));
3440 tcp_secret_secondary->expires = jiffy
3442 + (0xff & tcp_cookie_work(bakery, 1));
3443 tcp_secret_primary->expires = jiffy
3445 + (0x1f & tcp_cookie_work(bakery, 2));
3447 memcpy(&tcp_secret_secondary->secrets[0],
3448 bakery, COOKIE_WORKSPACE_WORDS);
3450 rcu_assign_pointer(tcp_secret_generating,
3451 tcp_secret_secondary);
3452 rcu_assign_pointer(tcp_secret_retiring,
3453 tcp_secret_primary);
3455 * Neither call_rcu() nor synchronize_rcu() needed.
3456 * Retiring data is not freed. It is replaced after
3457 * further (locked) pointer updates, and a quiet time
3458 * (minimum 1MSL, maximum LIFE - 2MSL).
3461 spin_unlock_bh(&tcp_secret_locker);
3465 &rcu_dereference(tcp_secret_generating)->secrets[0],
3466 COOKIE_WORKSPACE_WORDS);
3467 rcu_read_unlock_bh();
3471 EXPORT_SYMBOL(tcp_cookie_generator);
3473 void tcp_done(struct sock *sk)
3475 if (sk->sk_state == TCP_SYN_SENT || sk->sk_state == TCP_SYN_RECV)
3476 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_ATTEMPTFAILS);
3478 tcp_set_state(sk, TCP_CLOSE);
3479 tcp_clear_xmit_timers(sk);
3481 sk->sk_shutdown = SHUTDOWN_MASK;
3483 if (!sock_flag(sk, SOCK_DEAD))
3484 sk->sk_state_change(sk);
3486 inet_csk_destroy_sock(sk);
3488 EXPORT_SYMBOL_GPL(tcp_done);
3490 extern struct tcp_congestion_ops tcp_reno;
3492 static __initdata unsigned long thash_entries;
3493 static int __init set_thash_entries(char *str)
3497 thash_entries = simple_strtoul(str, &str, 0);
3500 __setup("thash_entries=", set_thash_entries);
3502 void tcp_init_mem(struct net *net)
3504 unsigned long limit = nr_free_buffer_pages() / 8;
3505 limit = max(limit, 128UL);
3506 net->ipv4.sysctl_tcp_mem[0] = limit / 4 * 3;
3507 net->ipv4.sysctl_tcp_mem[1] = limit;
3508 net->ipv4.sysctl_tcp_mem[2] = net->ipv4.sysctl_tcp_mem[0] * 2;
3511 void __init tcp_init(void)
3513 struct sk_buff *skb = NULL;
3514 unsigned long limit;
3517 unsigned long jiffy = jiffies;
3519 BUILD_BUG_ON(sizeof(struct tcp_skb_cb) > sizeof(skb->cb));
3521 percpu_counter_init(&tcp_sockets_allocated, 0);
3522 percpu_counter_init(&tcp_orphan_count, 0);
3523 tcp_hashinfo.bind_bucket_cachep =
3524 kmem_cache_create("tcp_bind_bucket",
3525 sizeof(struct inet_bind_bucket), 0,
3526 SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL);
3528 /* Size and allocate the main established and bind bucket
3531 * The methodology is similar to that of the buffer cache.
3533 tcp_hashinfo.ehash =
3534 alloc_large_system_hash("TCP established",
3535 sizeof(struct inet_ehash_bucket),
3537 (totalram_pages >= 128 * 1024) ?
3541 &tcp_hashinfo.ehash_mask,
3542 thash_entries ? 0 : 512 * 1024);
3543 for (i = 0; i <= tcp_hashinfo.ehash_mask; i++) {
3544 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].chain, i);
3545 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].twchain, i);
3547 if (inet_ehash_locks_alloc(&tcp_hashinfo))
3548 panic("TCP: failed to alloc ehash_locks");
3549 tcp_hashinfo.bhash =
3550 alloc_large_system_hash("TCP bind",
3551 sizeof(struct inet_bind_hashbucket),
3552 tcp_hashinfo.ehash_mask + 1,
3553 (totalram_pages >= 128 * 1024) ?
3556 &tcp_hashinfo.bhash_size,
3559 tcp_hashinfo.bhash_size = 1U << tcp_hashinfo.bhash_size;
3560 for (i = 0; i < tcp_hashinfo.bhash_size; i++) {
3561 spin_lock_init(&tcp_hashinfo.bhash[i].lock);
3562 INIT_HLIST_HEAD(&tcp_hashinfo.bhash[i].chain);
3566 cnt = tcp_hashinfo.ehash_mask + 1;
3568 tcp_death_row.sysctl_max_tw_buckets = cnt / 2;
3569 sysctl_tcp_max_orphans = cnt / 2;
3570 sysctl_max_syn_backlog = max(128, cnt / 256);
3572 tcp_init_mem(&init_net);
3573 /* Set per-socket limits to no more than 1/128 the pressure threshold */
3574 limit = nr_free_buffer_pages() << (PAGE_SHIFT - 7);
3575 max_share = min(4UL*1024*1024, limit);
3577 sysctl_tcp_wmem[0] = SK_MEM_QUANTUM;
3578 sysctl_tcp_wmem[1] = 16*1024;
3579 sysctl_tcp_wmem[2] = max(64*1024, max_share);
3581 sysctl_tcp_rmem[0] = SK_MEM_QUANTUM;
3582 sysctl_tcp_rmem[1] = 87380;
3583 sysctl_tcp_rmem[2] = max(87380, max_share);
3585 pr_info("Hash tables configured (established %u bind %u)\n",
3586 tcp_hashinfo.ehash_mask + 1, tcp_hashinfo.bhash_size);
3588 tcp_register_congestion_control(&tcp_reno);
3590 memset(&tcp_secret_one.secrets[0], 0, sizeof(tcp_secret_one.secrets));
3591 memset(&tcp_secret_two.secrets[0], 0, sizeof(tcp_secret_two.secrets));
3592 tcp_secret_one.expires = jiffy; /* past due */
3593 tcp_secret_two.expires = jiffy; /* past due */
3594 tcp_secret_generating = &tcp_secret_one;
3595 tcp_secret_primary = &tcp_secret_one;
3596 tcp_secret_retiring = &tcp_secret_two;
3597 tcp_secret_secondary = &tcp_secret_two;