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 tcp_enable_early_retrans(tp);
399 icsk->icsk_ca_ops = &tcp_init_congestion_ops;
401 sk->sk_state = TCP_CLOSE;
403 sk->sk_write_space = sk_stream_write_space;
404 sock_set_flag(sk, SOCK_USE_WRITE_QUEUE);
406 icsk->icsk_sync_mss = tcp_sync_mss;
408 /* TCP Cookie Transactions */
409 if (sysctl_tcp_cookie_size > 0) {
410 /* Default, cookies without s_data_payload. */
412 kzalloc(sizeof(*tp->cookie_values),
414 if (tp->cookie_values != NULL)
415 kref_init(&tp->cookie_values->kref);
417 /* Presumed zeroed, in order of appearance:
418 * cookie_in_always, cookie_out_never,
419 * s_data_constant, s_data_in, s_data_out
421 sk->sk_sndbuf = sysctl_tcp_wmem[1];
422 sk->sk_rcvbuf = sysctl_tcp_rmem[1];
425 sock_update_memcg(sk);
426 sk_sockets_allocated_inc(sk);
429 EXPORT_SYMBOL(tcp_init_sock);
432 * Wait for a TCP event.
434 * Note that we don't need to lock the socket, as the upper poll layers
435 * take care of normal races (between the test and the event) and we don't
436 * go look at any of the socket buffers directly.
438 unsigned int tcp_poll(struct file *file, struct socket *sock, poll_table *wait)
441 struct sock *sk = sock->sk;
442 const struct tcp_sock *tp = tcp_sk(sk);
444 sock_poll_wait(file, sk_sleep(sk), wait);
445 if (sk->sk_state == TCP_LISTEN)
446 return inet_csk_listen_poll(sk);
448 /* Socket is not locked. We are protected from async events
449 * by poll logic and correct handling of state changes
450 * made by other threads is impossible in any case.
456 * POLLHUP is certainly not done right. But poll() doesn't
457 * have a notion of HUP in just one direction, and for a
458 * socket the read side is more interesting.
460 * Some poll() documentation says that POLLHUP is incompatible
461 * with the POLLOUT/POLLWR flags, so somebody should check this
462 * all. But careful, it tends to be safer to return too many
463 * bits than too few, and you can easily break real applications
464 * if you don't tell them that something has hung up!
468 * Check number 1. POLLHUP is _UNMASKABLE_ event (see UNIX98 and
469 * our fs/select.c). It means that after we received EOF,
470 * poll always returns immediately, making impossible poll() on write()
471 * in state CLOSE_WAIT. One solution is evident --- to set POLLHUP
472 * if and only if shutdown has been made in both directions.
473 * Actually, it is interesting to look how Solaris and DUX
474 * solve this dilemma. I would prefer, if POLLHUP were maskable,
475 * then we could set it on SND_SHUTDOWN. BTW examples given
476 * in Stevens' books assume exactly this behaviour, it explains
477 * why POLLHUP is incompatible with POLLOUT. --ANK
479 * NOTE. Check for TCP_CLOSE is added. The goal is to prevent
480 * blocking on fresh not-connected or disconnected socket. --ANK
482 if (sk->sk_shutdown == SHUTDOWN_MASK || sk->sk_state == TCP_CLOSE)
484 if (sk->sk_shutdown & RCV_SHUTDOWN)
485 mask |= POLLIN | POLLRDNORM | POLLRDHUP;
488 if ((1 << sk->sk_state) & ~(TCPF_SYN_SENT | TCPF_SYN_RECV)) {
489 int target = sock_rcvlowat(sk, 0, INT_MAX);
491 if (tp->urg_seq == tp->copied_seq &&
492 !sock_flag(sk, SOCK_URGINLINE) &&
496 /* Potential race condition. If read of tp below will
497 * escape above sk->sk_state, we can be illegally awaken
498 * in SYN_* states. */
499 if (tp->rcv_nxt - tp->copied_seq >= target)
500 mask |= POLLIN | POLLRDNORM;
502 if (!(sk->sk_shutdown & SEND_SHUTDOWN)) {
503 if (sk_stream_wspace(sk) >= sk_stream_min_wspace(sk)) {
504 mask |= POLLOUT | POLLWRNORM;
505 } else { /* send SIGIO later */
506 set_bit(SOCK_ASYNC_NOSPACE,
507 &sk->sk_socket->flags);
508 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
510 /* Race breaker. If space is freed after
511 * wspace test but before the flags are set,
512 * IO signal will be lost.
514 if (sk_stream_wspace(sk) >= sk_stream_min_wspace(sk))
515 mask |= POLLOUT | POLLWRNORM;
518 mask |= POLLOUT | POLLWRNORM;
520 if (tp->urg_data & TCP_URG_VALID)
523 /* This barrier is coupled with smp_wmb() in tcp_reset() */
530 EXPORT_SYMBOL(tcp_poll);
532 int tcp_ioctl(struct sock *sk, int cmd, unsigned long arg)
534 struct tcp_sock *tp = tcp_sk(sk);
539 if (sk->sk_state == TCP_LISTEN)
543 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
545 else if (sock_flag(sk, SOCK_URGINLINE) ||
547 before(tp->urg_seq, tp->copied_seq) ||
548 !before(tp->urg_seq, tp->rcv_nxt)) {
551 answ = tp->rcv_nxt - tp->copied_seq;
553 /* Subtract 1, if FIN is in queue. */
554 skb = skb_peek_tail(&sk->sk_receive_queue);
556 answ -= tcp_hdr(skb)->fin;
558 answ = tp->urg_seq - tp->copied_seq;
562 answ = tp->urg_data && tp->urg_seq == tp->copied_seq;
565 if (sk->sk_state == TCP_LISTEN)
568 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
571 answ = tp->write_seq - tp->snd_una;
574 if (sk->sk_state == TCP_LISTEN)
577 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
580 answ = tp->write_seq - tp->snd_nxt;
586 return put_user(answ, (int __user *)arg);
588 EXPORT_SYMBOL(tcp_ioctl);
590 static inline void tcp_mark_push(struct tcp_sock *tp, struct sk_buff *skb)
592 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
593 tp->pushed_seq = tp->write_seq;
596 static inline int forced_push(const struct tcp_sock *tp)
598 return after(tp->write_seq, tp->pushed_seq + (tp->max_window >> 1));
601 static inline void skb_entail(struct sock *sk, struct sk_buff *skb)
603 struct tcp_sock *tp = tcp_sk(sk);
604 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
607 tcb->seq = tcb->end_seq = tp->write_seq;
608 tcb->tcp_flags = TCPHDR_ACK;
610 skb_header_release(skb);
611 tcp_add_write_queue_tail(sk, skb);
612 sk->sk_wmem_queued += skb->truesize;
613 sk_mem_charge(sk, skb->truesize);
614 if (tp->nonagle & TCP_NAGLE_PUSH)
615 tp->nonagle &= ~TCP_NAGLE_PUSH;
618 static inline void tcp_mark_urg(struct tcp_sock *tp, int flags)
621 tp->snd_up = tp->write_seq;
624 static inline void tcp_push(struct sock *sk, int flags, int mss_now,
627 if (tcp_send_head(sk)) {
628 struct tcp_sock *tp = tcp_sk(sk);
630 if (!(flags & MSG_MORE) || forced_push(tp))
631 tcp_mark_push(tp, tcp_write_queue_tail(sk));
633 tcp_mark_urg(tp, flags);
634 __tcp_push_pending_frames(sk, mss_now,
635 (flags & MSG_MORE) ? TCP_NAGLE_CORK : nonagle);
639 static int tcp_splice_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb,
640 unsigned int offset, size_t len)
642 struct tcp_splice_state *tss = rd_desc->arg.data;
645 ret = skb_splice_bits(skb, offset, tss->pipe, min(rd_desc->count, len),
648 rd_desc->count -= ret;
652 static int __tcp_splice_read(struct sock *sk, struct tcp_splice_state *tss)
654 /* Store TCP splice context information in read_descriptor_t. */
655 read_descriptor_t rd_desc = {
660 return tcp_read_sock(sk, &rd_desc, tcp_splice_data_recv);
664 * tcp_splice_read - splice data from TCP socket to a pipe
665 * @sock: socket to splice from
666 * @ppos: position (not valid)
667 * @pipe: pipe to splice to
668 * @len: number of bytes to splice
669 * @flags: splice modifier flags
672 * Will read pages from given socket and fill them into a pipe.
675 ssize_t tcp_splice_read(struct socket *sock, loff_t *ppos,
676 struct pipe_inode_info *pipe, size_t len,
679 struct sock *sk = sock->sk;
680 struct tcp_splice_state tss = {
689 sock_rps_record_flow(sk);
691 * We can't seek on a socket input
700 timeo = sock_rcvtimeo(sk, sock->file->f_flags & O_NONBLOCK);
702 ret = __tcp_splice_read(sk, &tss);
708 if (sock_flag(sk, SOCK_DONE))
711 ret = sock_error(sk);
714 if (sk->sk_shutdown & RCV_SHUTDOWN)
716 if (sk->sk_state == TCP_CLOSE) {
718 * This occurs when user tries to read
719 * from never connected socket.
721 if (!sock_flag(sk, SOCK_DONE))
729 sk_wait_data(sk, &timeo);
730 if (signal_pending(current)) {
731 ret = sock_intr_errno(timeo);
744 if (sk->sk_err || sk->sk_state == TCP_CLOSE ||
745 (sk->sk_shutdown & RCV_SHUTDOWN) ||
746 signal_pending(current))
757 EXPORT_SYMBOL(tcp_splice_read);
759 struct sk_buff *sk_stream_alloc_skb(struct sock *sk, int size, gfp_t gfp)
763 /* The TCP header must be at least 32-bit aligned. */
764 size = ALIGN(size, 4);
766 skb = alloc_skb_fclone(size + sk->sk_prot->max_header, gfp);
768 if (sk_wmem_schedule(sk, skb->truesize)) {
769 skb_reserve(skb, sk->sk_prot->max_header);
771 * Make sure that we have exactly size bytes
772 * available to the caller, no more, no less.
774 skb->avail_size = size;
779 sk->sk_prot->enter_memory_pressure(sk);
780 sk_stream_moderate_sndbuf(sk);
785 static unsigned int tcp_xmit_size_goal(struct sock *sk, u32 mss_now,
788 struct tcp_sock *tp = tcp_sk(sk);
789 u32 xmit_size_goal, old_size_goal;
791 xmit_size_goal = mss_now;
793 if (large_allowed && sk_can_gso(sk)) {
794 xmit_size_goal = ((sk->sk_gso_max_size - 1) -
795 inet_csk(sk)->icsk_af_ops->net_header_len -
796 inet_csk(sk)->icsk_ext_hdr_len -
799 xmit_size_goal = tcp_bound_to_half_wnd(tp, xmit_size_goal);
801 /* We try hard to avoid divides here */
802 old_size_goal = tp->xmit_size_goal_segs * mss_now;
804 if (likely(old_size_goal <= xmit_size_goal &&
805 old_size_goal + mss_now > xmit_size_goal)) {
806 xmit_size_goal = old_size_goal;
808 tp->xmit_size_goal_segs = xmit_size_goal / mss_now;
809 xmit_size_goal = tp->xmit_size_goal_segs * mss_now;
813 return max(xmit_size_goal, mss_now);
816 static int tcp_send_mss(struct sock *sk, int *size_goal, int flags)
820 mss_now = tcp_current_mss(sk);
821 *size_goal = tcp_xmit_size_goal(sk, mss_now, !(flags & MSG_OOB));
826 static ssize_t do_tcp_sendpages(struct sock *sk, struct page **pages, int poffset,
827 size_t psize, int flags)
829 struct tcp_sock *tp = tcp_sk(sk);
830 int mss_now, size_goal;
833 long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
835 /* Wait for a connection to finish. */
836 if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
837 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
840 clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
842 mss_now = tcp_send_mss(sk, &size_goal, flags);
846 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
850 struct sk_buff *skb = tcp_write_queue_tail(sk);
851 struct page *page = pages[poffset / PAGE_SIZE];
853 int offset = poffset % PAGE_SIZE;
854 int size = min_t(size_t, psize, PAGE_SIZE - offset);
857 if (!tcp_send_head(sk) || (copy = size_goal - skb->len) <= 0) {
859 if (!sk_stream_memory_free(sk))
860 goto wait_for_sndbuf;
862 skb = sk_stream_alloc_skb(sk, 0, sk->sk_allocation);
864 goto wait_for_memory;
873 i = skb_shinfo(skb)->nr_frags;
874 can_coalesce = skb_can_coalesce(skb, i, page, offset);
875 if (!can_coalesce && i >= MAX_SKB_FRAGS) {
876 tcp_mark_push(tp, skb);
879 if (!sk_wmem_schedule(sk, copy))
880 goto wait_for_memory;
883 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
886 skb_fill_page_desc(skb, i, page, offset, copy);
890 skb->data_len += copy;
891 skb->truesize += copy;
892 sk->sk_wmem_queued += copy;
893 sk_mem_charge(sk, copy);
894 skb->ip_summed = CHECKSUM_PARTIAL;
895 tp->write_seq += copy;
896 TCP_SKB_CB(skb)->end_seq += copy;
897 skb_shinfo(skb)->gso_segs = 0;
900 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
904 if (!(psize -= copy))
907 if (skb->len < size_goal || (flags & MSG_OOB))
910 if (forced_push(tp)) {
911 tcp_mark_push(tp, skb);
912 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
913 } else if (skb == tcp_send_head(sk))
914 tcp_push_one(sk, mss_now);
918 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
921 tcp_push(sk, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH);
923 if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
926 mss_now = tcp_send_mss(sk, &size_goal, flags);
930 if (copied && !(flags & MSG_SENDPAGE_NOTLAST))
931 tcp_push(sk, flags, mss_now, tp->nonagle);
938 return sk_stream_error(sk, flags, err);
941 int tcp_sendpage(struct sock *sk, struct page *page, int offset,
942 size_t size, int flags)
946 if (!(sk->sk_route_caps & NETIF_F_SG) ||
947 !(sk->sk_route_caps & NETIF_F_ALL_CSUM))
948 return sock_no_sendpage(sk->sk_socket, page, offset, size,
952 res = do_tcp_sendpages(sk, &page, offset, size, flags);
956 EXPORT_SYMBOL(tcp_sendpage);
958 static inline int select_size(const struct sock *sk, bool sg)
960 const struct tcp_sock *tp = tcp_sk(sk);
961 int tmp = tp->mss_cache;
964 if (sk_can_gso(sk)) {
965 /* Small frames wont use a full page:
966 * Payload will immediately follow tcp header.
968 tmp = SKB_WITH_OVERHEAD(2048 - MAX_TCP_HEADER);
970 int pgbreak = SKB_MAX_HEAD(MAX_TCP_HEADER);
972 if (tmp >= pgbreak &&
973 tmp <= pgbreak + (MAX_SKB_FRAGS - 1) * PAGE_SIZE)
981 int tcp_sendmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg,
985 struct tcp_sock *tp = tcp_sk(sk);
987 int iovlen, flags, err, copied;
988 int mss_now = 0, size_goal;
994 flags = msg->msg_flags;
995 timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
997 /* Wait for a connection to finish. */
998 if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
999 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
1002 if (unlikely(tp->repair)) {
1003 if (tp->repair_queue == TCP_RECV_QUEUE) {
1004 copied = tcp_send_rcvq(sk, msg, size);
1009 if (tp->repair_queue == TCP_NO_QUEUE)
1012 /* 'common' sending to sendq */
1015 /* This should be in poll */
1016 clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
1018 mss_now = tcp_send_mss(sk, &size_goal, flags);
1020 /* Ok commence sending. */
1021 iovlen = msg->msg_iovlen;
1026 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
1029 sg = !!(sk->sk_route_caps & NETIF_F_SG);
1031 while (--iovlen >= 0) {
1032 size_t seglen = iov->iov_len;
1033 unsigned char __user *from = iov->iov_base;
1037 while (seglen > 0) {
1039 int max = size_goal;
1041 skb = tcp_write_queue_tail(sk);
1042 if (tcp_send_head(sk)) {
1043 if (skb->ip_summed == CHECKSUM_NONE)
1045 copy = max - skb->len;
1050 /* Allocate new segment. If the interface is SG,
1051 * allocate skb fitting to single page.
1053 if (!sk_stream_memory_free(sk))
1054 goto wait_for_sndbuf;
1056 skb = sk_stream_alloc_skb(sk,
1057 select_size(sk, sg),
1060 goto wait_for_memory;
1063 * Check whether we can use HW checksum.
1065 if (sk->sk_route_caps & NETIF_F_ALL_CSUM)
1066 skb->ip_summed = CHECKSUM_PARTIAL;
1068 skb_entail(sk, skb);
1073 /* Try to append data to the end of skb. */
1077 /* Where to copy to? */
1078 if (skb_availroom(skb) > 0) {
1079 /* We have some space in skb head. Superb! */
1080 copy = min_t(int, copy, skb_availroom(skb));
1081 err = skb_add_data_nocache(sk, skb, from, copy);
1086 int i = skb_shinfo(skb)->nr_frags;
1087 struct page *page = sk->sk_sndmsg_page;
1090 if (page && page_count(page) == 1)
1091 sk->sk_sndmsg_off = 0;
1093 off = sk->sk_sndmsg_off;
1095 if (skb_can_coalesce(skb, i, page, off) &&
1097 /* We can extend the last page
1100 } else if (i == MAX_SKB_FRAGS || !sg) {
1101 /* Need to add new fragment and cannot
1102 * do this because interface is non-SG,
1103 * or because all the page slots are
1105 tcp_mark_push(tp, skb);
1108 if (off == PAGE_SIZE) {
1110 sk->sk_sndmsg_page = page = NULL;
1116 if (copy > PAGE_SIZE - off)
1117 copy = PAGE_SIZE - off;
1119 if (!sk_wmem_schedule(sk, copy))
1120 goto wait_for_memory;
1123 /* Allocate new cache page. */
1124 if (!(page = sk_stream_alloc_page(sk)))
1125 goto wait_for_memory;
1128 /* Time to copy data. We are close to
1130 err = skb_copy_to_page_nocache(sk, from, skb,
1133 /* If this page was new, give it to the
1134 * socket so it does not get leaked.
1136 if (!sk->sk_sndmsg_page) {
1137 sk->sk_sndmsg_page = page;
1138 sk->sk_sndmsg_off = 0;
1143 /* Update the skb. */
1145 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1147 skb_fill_page_desc(skb, i, page, off, copy);
1148 if (sk->sk_sndmsg_page) {
1150 } else if (off + copy < PAGE_SIZE) {
1152 sk->sk_sndmsg_page = page;
1156 sk->sk_sndmsg_off = off + copy;
1160 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
1162 tp->write_seq += copy;
1163 TCP_SKB_CB(skb)->end_seq += copy;
1164 skb_shinfo(skb)->gso_segs = 0;
1168 if ((seglen -= copy) == 0 && iovlen == 0)
1171 if (skb->len < max || (flags & MSG_OOB) || unlikely(tp->repair))
1174 if (forced_push(tp)) {
1175 tcp_mark_push(tp, skb);
1176 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
1177 } else if (skb == tcp_send_head(sk))
1178 tcp_push_one(sk, mss_now);
1182 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1184 if (copied && likely(!tp->repair))
1185 tcp_push(sk, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH);
1187 if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
1190 mss_now = tcp_send_mss(sk, &size_goal, flags);
1195 if (copied && likely(!tp->repair))
1196 tcp_push(sk, flags, mss_now, tp->nonagle);
1202 tcp_unlink_write_queue(skb, sk);
1203 /* It is the one place in all of TCP, except connection
1204 * reset, where we can be unlinking the send_head.
1206 tcp_check_send_head(sk, skb);
1207 sk_wmem_free_skb(sk, skb);
1214 err = sk_stream_error(sk, flags, err);
1218 EXPORT_SYMBOL(tcp_sendmsg);
1221 * Handle reading urgent data. BSD has very simple semantics for
1222 * this, no blocking and very strange errors 8)
1225 static int tcp_recv_urg(struct sock *sk, struct msghdr *msg, int len, int flags)
1227 struct tcp_sock *tp = tcp_sk(sk);
1229 /* No URG data to read. */
1230 if (sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data ||
1231 tp->urg_data == TCP_URG_READ)
1232 return -EINVAL; /* Yes this is right ! */
1234 if (sk->sk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DONE))
1237 if (tp->urg_data & TCP_URG_VALID) {
1239 char c = tp->urg_data;
1241 if (!(flags & MSG_PEEK))
1242 tp->urg_data = TCP_URG_READ;
1244 /* Read urgent data. */
1245 msg->msg_flags |= MSG_OOB;
1248 if (!(flags & MSG_TRUNC))
1249 err = memcpy_toiovec(msg->msg_iov, &c, 1);
1252 msg->msg_flags |= MSG_TRUNC;
1254 return err ? -EFAULT : len;
1257 if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
1260 /* Fixed the recv(..., MSG_OOB) behaviour. BSD docs and
1261 * the available implementations agree in this case:
1262 * this call should never block, independent of the
1263 * blocking state of the socket.
1264 * Mike <pall@rz.uni-karlsruhe.de>
1269 static int tcp_peek_sndq(struct sock *sk, struct msghdr *msg, int len)
1271 struct sk_buff *skb;
1272 int copied = 0, err = 0;
1274 /* XXX -- need to support SO_PEEK_OFF */
1276 skb_queue_walk(&sk->sk_write_queue, skb) {
1277 err = skb_copy_datagram_iovec(skb, 0, msg->msg_iov, skb->len);
1284 return err ?: copied;
1287 /* Clean up the receive buffer for full frames taken by the user,
1288 * then send an ACK if necessary. COPIED is the number of bytes
1289 * tcp_recvmsg has given to the user so far, it speeds up the
1290 * calculation of whether or not we must ACK for the sake of
1293 void tcp_cleanup_rbuf(struct sock *sk, int copied)
1295 struct tcp_sock *tp = tcp_sk(sk);
1296 int time_to_ack = 0;
1298 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
1300 WARN(skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq),
1301 "cleanup rbuf bug: copied %X seq %X rcvnxt %X\n",
1302 tp->copied_seq, TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt);
1304 if (inet_csk_ack_scheduled(sk)) {
1305 const struct inet_connection_sock *icsk = inet_csk(sk);
1306 /* Delayed ACKs frequently hit locked sockets during bulk
1308 if (icsk->icsk_ack.blocked ||
1309 /* Once-per-two-segments ACK was not sent by tcp_input.c */
1310 tp->rcv_nxt - tp->rcv_wup > icsk->icsk_ack.rcv_mss ||
1312 * If this read emptied read buffer, we send ACK, if
1313 * connection is not bidirectional, user drained
1314 * receive buffer and there was a small segment
1318 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED2) ||
1319 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED) &&
1320 !icsk->icsk_ack.pingpong)) &&
1321 !atomic_read(&sk->sk_rmem_alloc)))
1325 /* We send an ACK if we can now advertise a non-zero window
1326 * which has been raised "significantly".
1328 * Even if window raised up to infinity, do not send window open ACK
1329 * in states, where we will not receive more. It is useless.
1331 if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) {
1332 __u32 rcv_window_now = tcp_receive_window(tp);
1334 /* Optimize, __tcp_select_window() is not cheap. */
1335 if (2*rcv_window_now <= tp->window_clamp) {
1336 __u32 new_window = __tcp_select_window(sk);
1338 /* Send ACK now, if this read freed lots of space
1339 * in our buffer. Certainly, new_window is new window.
1340 * We can advertise it now, if it is not less than current one.
1341 * "Lots" means "at least twice" here.
1343 if (new_window && new_window >= 2 * rcv_window_now)
1351 static void tcp_prequeue_process(struct sock *sk)
1353 struct sk_buff *skb;
1354 struct tcp_sock *tp = tcp_sk(sk);
1356 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPPREQUEUED);
1358 /* RX process wants to run with disabled BHs, though it is not
1361 while ((skb = __skb_dequeue(&tp->ucopy.prequeue)) != NULL)
1362 sk_backlog_rcv(sk, skb);
1365 /* Clear memory counter. */
1366 tp->ucopy.memory = 0;
1369 #ifdef CONFIG_NET_DMA
1370 static void tcp_service_net_dma(struct sock *sk, bool wait)
1372 dma_cookie_t done, used;
1373 dma_cookie_t last_issued;
1374 struct tcp_sock *tp = tcp_sk(sk);
1376 if (!tp->ucopy.dma_chan)
1379 last_issued = tp->ucopy.dma_cookie;
1380 dma_async_memcpy_issue_pending(tp->ucopy.dma_chan);
1383 if (dma_async_memcpy_complete(tp->ucopy.dma_chan,
1385 &used) == DMA_SUCCESS) {
1386 /* Safe to free early-copied skbs now */
1387 __skb_queue_purge(&sk->sk_async_wait_queue);
1390 struct sk_buff *skb;
1391 while ((skb = skb_peek(&sk->sk_async_wait_queue)) &&
1392 (dma_async_is_complete(skb->dma_cookie, done,
1393 used) == DMA_SUCCESS)) {
1394 __skb_dequeue(&sk->sk_async_wait_queue);
1402 static inline struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off)
1404 struct sk_buff *skb;
1407 skb_queue_walk(&sk->sk_receive_queue, skb) {
1408 offset = seq - TCP_SKB_CB(skb)->seq;
1409 if (tcp_hdr(skb)->syn)
1411 if (offset < skb->len || tcp_hdr(skb)->fin) {
1420 * This routine provides an alternative to tcp_recvmsg() for routines
1421 * that would like to handle copying from skbuffs directly in 'sendfile'
1424 * - It is assumed that the socket was locked by the caller.
1425 * - The routine does not block.
1426 * - At present, there is no support for reading OOB data
1427 * or for 'peeking' the socket using this routine
1428 * (although both would be easy to implement).
1430 int tcp_read_sock(struct sock *sk, read_descriptor_t *desc,
1431 sk_read_actor_t recv_actor)
1433 struct sk_buff *skb;
1434 struct tcp_sock *tp = tcp_sk(sk);
1435 u32 seq = tp->copied_seq;
1439 if (sk->sk_state == TCP_LISTEN)
1441 while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1442 if (offset < skb->len) {
1446 len = skb->len - offset;
1447 /* Stop reading if we hit a patch of urgent data */
1449 u32 urg_offset = tp->urg_seq - seq;
1450 if (urg_offset < len)
1455 used = recv_actor(desc, skb, offset, len);
1460 } else if (used <= len) {
1466 * If recv_actor drops the lock (e.g. TCP splice
1467 * receive) the skb pointer might be invalid when
1468 * getting here: tcp_collapse might have deleted it
1469 * while aggregating skbs from the socket queue.
1471 skb = tcp_recv_skb(sk, seq-1, &offset);
1472 if (!skb || (offset+1 != skb->len))
1475 if (tcp_hdr(skb)->fin) {
1476 sk_eat_skb(sk, skb, false);
1480 sk_eat_skb(sk, skb, false);
1483 tp->copied_seq = seq;
1485 tp->copied_seq = seq;
1487 tcp_rcv_space_adjust(sk);
1489 /* Clean up data we have read: This will do ACK frames. */
1491 tcp_cleanup_rbuf(sk, copied);
1494 EXPORT_SYMBOL(tcp_read_sock);
1497 * This routine copies from a sock struct into the user buffer.
1499 * Technical note: in 2.3 we work on _locked_ socket, so that
1500 * tricks with *seq access order and skb->users are not required.
1501 * Probably, code can be easily improved even more.
1504 int tcp_recvmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg,
1505 size_t len, int nonblock, int flags, int *addr_len)
1507 struct tcp_sock *tp = tcp_sk(sk);
1513 int target; /* Read at least this many bytes */
1515 struct task_struct *user_recv = NULL;
1516 bool copied_early = false;
1517 struct sk_buff *skb;
1523 if (sk->sk_state == TCP_LISTEN)
1526 timeo = sock_rcvtimeo(sk, nonblock);
1528 /* Urgent data needs to be handled specially. */
1529 if (flags & MSG_OOB)
1532 if (unlikely(tp->repair)) {
1534 if (!(flags & MSG_PEEK))
1537 if (tp->repair_queue == TCP_SEND_QUEUE)
1541 if (tp->repair_queue == TCP_NO_QUEUE)
1544 /* 'common' recv queue MSG_PEEK-ing */
1547 seq = &tp->copied_seq;
1548 if (flags & MSG_PEEK) {
1549 peek_seq = tp->copied_seq;
1553 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
1555 #ifdef CONFIG_NET_DMA
1556 tp->ucopy.dma_chan = NULL;
1558 skb = skb_peek_tail(&sk->sk_receive_queue);
1563 available = TCP_SKB_CB(skb)->seq + skb->len - (*seq);
1564 if ((available < target) &&
1565 (len > sysctl_tcp_dma_copybreak) && !(flags & MSG_PEEK) &&
1566 !sysctl_tcp_low_latency &&
1567 net_dma_find_channel()) {
1568 preempt_enable_no_resched();
1569 tp->ucopy.pinned_list =
1570 dma_pin_iovec_pages(msg->msg_iov, len);
1572 preempt_enable_no_resched();
1580 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
1581 if (tp->urg_data && tp->urg_seq == *seq) {
1584 if (signal_pending(current)) {
1585 copied = timeo ? sock_intr_errno(timeo) : -EAGAIN;
1590 /* Next get a buffer. */
1592 skb_queue_walk(&sk->sk_receive_queue, skb) {
1593 /* Now that we have two receive queues this
1596 if (WARN(before(*seq, TCP_SKB_CB(skb)->seq),
1597 "recvmsg bug: copied %X seq %X rcvnxt %X fl %X\n",
1598 *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt,
1602 offset = *seq - TCP_SKB_CB(skb)->seq;
1603 if (tcp_hdr(skb)->syn)
1605 if (offset < skb->len)
1607 if (tcp_hdr(skb)->fin)
1609 WARN(!(flags & MSG_PEEK),
1610 "recvmsg bug 2: copied %X seq %X rcvnxt %X fl %X\n",
1611 *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt, flags);
1614 /* Well, if we have backlog, try to process it now yet. */
1616 if (copied >= target && !sk->sk_backlog.tail)
1621 sk->sk_state == TCP_CLOSE ||
1622 (sk->sk_shutdown & RCV_SHUTDOWN) ||
1624 signal_pending(current))
1627 if (sock_flag(sk, SOCK_DONE))
1631 copied = sock_error(sk);
1635 if (sk->sk_shutdown & RCV_SHUTDOWN)
1638 if (sk->sk_state == TCP_CLOSE) {
1639 if (!sock_flag(sk, SOCK_DONE)) {
1640 /* This occurs when user tries to read
1641 * from never connected socket.
1654 if (signal_pending(current)) {
1655 copied = sock_intr_errno(timeo);
1660 tcp_cleanup_rbuf(sk, copied);
1662 if (!sysctl_tcp_low_latency && tp->ucopy.task == user_recv) {
1663 /* Install new reader */
1664 if (!user_recv && !(flags & (MSG_TRUNC | MSG_PEEK))) {
1665 user_recv = current;
1666 tp->ucopy.task = user_recv;
1667 tp->ucopy.iov = msg->msg_iov;
1670 tp->ucopy.len = len;
1672 WARN_ON(tp->copied_seq != tp->rcv_nxt &&
1673 !(flags & (MSG_PEEK | MSG_TRUNC)));
1675 /* Ugly... If prequeue is not empty, we have to
1676 * process it before releasing socket, otherwise
1677 * order will be broken at second iteration.
1678 * More elegant solution is required!!!
1680 * Look: we have the following (pseudo)queues:
1682 * 1. packets in flight
1687 * Each queue can be processed only if the next ones
1688 * are empty. At this point we have empty receive_queue.
1689 * But prequeue _can_ be not empty after 2nd iteration,
1690 * when we jumped to start of loop because backlog
1691 * processing added something to receive_queue.
1692 * We cannot release_sock(), because backlog contains
1693 * packets arrived _after_ prequeued ones.
1695 * Shortly, algorithm is clear --- to process all
1696 * the queues in order. We could make it more directly,
1697 * requeueing packets from backlog to prequeue, if
1698 * is not empty. It is more elegant, but eats cycles,
1701 if (!skb_queue_empty(&tp->ucopy.prequeue))
1704 /* __ Set realtime policy in scheduler __ */
1707 #ifdef CONFIG_NET_DMA
1708 if (tp->ucopy.dma_chan)
1709 dma_async_memcpy_issue_pending(tp->ucopy.dma_chan);
1711 if (copied >= target) {
1712 /* Do not sleep, just process backlog. */
1716 sk_wait_data(sk, &timeo);
1718 #ifdef CONFIG_NET_DMA
1719 tcp_service_net_dma(sk, false); /* Don't block */
1720 tp->ucopy.wakeup = 0;
1726 /* __ Restore normal policy in scheduler __ */
1728 if ((chunk = len - tp->ucopy.len) != 0) {
1729 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMBACKLOG, chunk);
1734 if (tp->rcv_nxt == tp->copied_seq &&
1735 !skb_queue_empty(&tp->ucopy.prequeue)) {
1737 tcp_prequeue_process(sk);
1739 if ((chunk = len - tp->ucopy.len) != 0) {
1740 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1746 if ((flags & MSG_PEEK) &&
1747 (peek_seq - copied - urg_hole != tp->copied_seq)) {
1748 net_dbg_ratelimited("TCP(%s:%d): Application bug, race in MSG_PEEK\n",
1750 task_pid_nr(current));
1751 peek_seq = tp->copied_seq;
1756 /* Ok so how much can we use? */
1757 used = skb->len - offset;
1761 /* Do we have urgent data here? */
1763 u32 urg_offset = tp->urg_seq - *seq;
1764 if (urg_offset < used) {
1766 if (!sock_flag(sk, SOCK_URGINLINE)) {
1779 if (!(flags & MSG_TRUNC)) {
1780 #ifdef CONFIG_NET_DMA
1781 if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list)
1782 tp->ucopy.dma_chan = net_dma_find_channel();
1784 if (tp->ucopy.dma_chan) {
1785 tp->ucopy.dma_cookie = dma_skb_copy_datagram_iovec(
1786 tp->ucopy.dma_chan, skb, offset,
1788 tp->ucopy.pinned_list);
1790 if (tp->ucopy.dma_cookie < 0) {
1792 pr_alert("%s: dma_cookie < 0\n",
1795 /* Exception. Bailout! */
1801 dma_async_memcpy_issue_pending(tp->ucopy.dma_chan);
1803 if ((offset + used) == skb->len)
1804 copied_early = true;
1809 err = skb_copy_datagram_iovec(skb, offset,
1810 msg->msg_iov, used);
1812 /* Exception. Bailout! */
1824 tcp_rcv_space_adjust(sk);
1827 if (tp->urg_data && after(tp->copied_seq, tp->urg_seq)) {
1829 tcp_fast_path_check(sk);
1831 if (used + offset < skb->len)
1834 if (tcp_hdr(skb)->fin)
1836 if (!(flags & MSG_PEEK)) {
1837 sk_eat_skb(sk, skb, copied_early);
1838 copied_early = false;
1843 /* Process the FIN. */
1845 if (!(flags & MSG_PEEK)) {
1846 sk_eat_skb(sk, skb, copied_early);
1847 copied_early = false;
1853 if (!skb_queue_empty(&tp->ucopy.prequeue)) {
1856 tp->ucopy.len = copied > 0 ? len : 0;
1858 tcp_prequeue_process(sk);
1860 if (copied > 0 && (chunk = len - tp->ucopy.len) != 0) {
1861 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1867 tp->ucopy.task = NULL;
1871 #ifdef CONFIG_NET_DMA
1872 tcp_service_net_dma(sk, true); /* Wait for queue to drain */
1873 tp->ucopy.dma_chan = NULL;
1875 if (tp->ucopy.pinned_list) {
1876 dma_unpin_iovec_pages(tp->ucopy.pinned_list);
1877 tp->ucopy.pinned_list = NULL;
1881 /* According to UNIX98, msg_name/msg_namelen are ignored
1882 * on connected socket. I was just happy when found this 8) --ANK
1885 /* Clean up data we have read: This will do ACK frames. */
1886 tcp_cleanup_rbuf(sk, copied);
1896 err = tcp_recv_urg(sk, msg, len, flags);
1900 err = tcp_peek_sndq(sk, msg, len);
1903 EXPORT_SYMBOL(tcp_recvmsg);
1905 void tcp_set_state(struct sock *sk, int state)
1907 int oldstate = sk->sk_state;
1910 case TCP_ESTABLISHED:
1911 if (oldstate != TCP_ESTABLISHED)
1912 TCP_INC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
1916 if (oldstate == TCP_CLOSE_WAIT || oldstate == TCP_ESTABLISHED)
1917 TCP_INC_STATS(sock_net(sk), TCP_MIB_ESTABRESETS);
1919 sk->sk_prot->unhash(sk);
1920 if (inet_csk(sk)->icsk_bind_hash &&
1921 !(sk->sk_userlocks & SOCK_BINDPORT_LOCK))
1925 if (oldstate == TCP_ESTABLISHED)
1926 TCP_DEC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
1929 /* Change state AFTER socket is unhashed to avoid closed
1930 * socket sitting in hash tables.
1932 sk->sk_state = state;
1935 SOCK_DEBUG(sk, "TCP sk=%p, State %s -> %s\n", sk, statename[oldstate], statename[state]);
1938 EXPORT_SYMBOL_GPL(tcp_set_state);
1941 * State processing on a close. This implements the state shift for
1942 * sending our FIN frame. Note that we only send a FIN for some
1943 * states. A shutdown() may have already sent the FIN, or we may be
1947 static const unsigned char new_state[16] = {
1948 /* current state: new state: action: */
1949 /* (Invalid) */ TCP_CLOSE,
1950 /* TCP_ESTABLISHED */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1951 /* TCP_SYN_SENT */ TCP_CLOSE,
1952 /* TCP_SYN_RECV */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1953 /* TCP_FIN_WAIT1 */ TCP_FIN_WAIT1,
1954 /* TCP_FIN_WAIT2 */ TCP_FIN_WAIT2,
1955 /* TCP_TIME_WAIT */ TCP_CLOSE,
1956 /* TCP_CLOSE */ TCP_CLOSE,
1957 /* TCP_CLOSE_WAIT */ TCP_LAST_ACK | TCP_ACTION_FIN,
1958 /* TCP_LAST_ACK */ TCP_LAST_ACK,
1959 /* TCP_LISTEN */ TCP_CLOSE,
1960 /* TCP_CLOSING */ TCP_CLOSING,
1963 static int tcp_close_state(struct sock *sk)
1965 int next = (int)new_state[sk->sk_state];
1966 int ns = next & TCP_STATE_MASK;
1968 tcp_set_state(sk, ns);
1970 return next & TCP_ACTION_FIN;
1974 * Shutdown the sending side of a connection. Much like close except
1975 * that we don't receive shut down or sock_set_flag(sk, SOCK_DEAD).
1978 void tcp_shutdown(struct sock *sk, int how)
1980 /* We need to grab some memory, and put together a FIN,
1981 * and then put it into the queue to be sent.
1982 * Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
1984 if (!(how & SEND_SHUTDOWN))
1987 /* If we've already sent a FIN, or it's a closed state, skip this. */
1988 if ((1 << sk->sk_state) &
1989 (TCPF_ESTABLISHED | TCPF_SYN_SENT |
1990 TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) {
1991 /* Clear out any half completed packets. FIN if needed. */
1992 if (tcp_close_state(sk))
1996 EXPORT_SYMBOL(tcp_shutdown);
1998 bool tcp_check_oom(struct sock *sk, int shift)
2000 bool too_many_orphans, out_of_socket_memory;
2002 too_many_orphans = tcp_too_many_orphans(sk, shift);
2003 out_of_socket_memory = tcp_out_of_memory(sk);
2005 if (too_many_orphans)
2006 net_info_ratelimited("too many orphaned sockets\n");
2007 if (out_of_socket_memory)
2008 net_info_ratelimited("out of memory -- consider tuning tcp_mem\n");
2009 return too_many_orphans || out_of_socket_memory;
2012 void tcp_close(struct sock *sk, long timeout)
2014 struct sk_buff *skb;
2015 int data_was_unread = 0;
2019 sk->sk_shutdown = SHUTDOWN_MASK;
2021 if (sk->sk_state == TCP_LISTEN) {
2022 tcp_set_state(sk, TCP_CLOSE);
2025 inet_csk_listen_stop(sk);
2027 goto adjudge_to_death;
2030 /* We need to flush the recv. buffs. We do this only on the
2031 * descriptor close, not protocol-sourced closes, because the
2032 * reader process may not have drained the data yet!
2034 while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) {
2035 u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq -
2037 data_was_unread += len;
2043 /* If socket has been already reset (e.g. in tcp_reset()) - kill it. */
2044 if (sk->sk_state == TCP_CLOSE)
2045 goto adjudge_to_death;
2047 /* As outlined in RFC 2525, section 2.17, we send a RST here because
2048 * data was lost. To witness the awful effects of the old behavior of
2049 * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk
2050 * GET in an FTP client, suspend the process, wait for the client to
2051 * advertise a zero window, then kill -9 the FTP client, wheee...
2052 * Note: timeout is always zero in such a case.
2054 if (unlikely(tcp_sk(sk)->repair)) {
2055 sk->sk_prot->disconnect(sk, 0);
2056 } else if (data_was_unread) {
2057 /* Unread data was tossed, zap the connection. */
2058 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONCLOSE);
2059 tcp_set_state(sk, TCP_CLOSE);
2060 tcp_send_active_reset(sk, sk->sk_allocation);
2061 } else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) {
2062 /* Check zero linger _after_ checking for unread data. */
2063 sk->sk_prot->disconnect(sk, 0);
2064 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
2065 } else if (tcp_close_state(sk)) {
2066 /* We FIN if the application ate all the data before
2067 * zapping the connection.
2070 /* RED-PEN. Formally speaking, we have broken TCP state
2071 * machine. State transitions:
2073 * TCP_ESTABLISHED -> TCP_FIN_WAIT1
2074 * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible)
2075 * TCP_CLOSE_WAIT -> TCP_LAST_ACK
2077 * are legal only when FIN has been sent (i.e. in window),
2078 * rather than queued out of window. Purists blame.
2080 * F.e. "RFC state" is ESTABLISHED,
2081 * if Linux state is FIN-WAIT-1, but FIN is still not sent.
2083 * The visible declinations are that sometimes
2084 * we enter time-wait state, when it is not required really
2085 * (harmless), do not send active resets, when they are
2086 * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
2087 * they look as CLOSING or LAST_ACK for Linux)
2088 * Probably, I missed some more holelets.
2094 sk_stream_wait_close(sk, timeout);
2097 state = sk->sk_state;
2101 /* It is the last release_sock in its life. It will remove backlog. */
2105 /* Now socket is owned by kernel and we acquire BH lock
2106 to finish close. No need to check for user refs.
2110 WARN_ON(sock_owned_by_user(sk));
2112 percpu_counter_inc(sk->sk_prot->orphan_count);
2114 /* Have we already been destroyed by a softirq or backlog? */
2115 if (state != TCP_CLOSE && sk->sk_state == TCP_CLOSE)
2118 /* This is a (useful) BSD violating of the RFC. There is a
2119 * problem with TCP as specified in that the other end could
2120 * keep a socket open forever with no application left this end.
2121 * We use a 3 minute timeout (about the same as BSD) then kill
2122 * our end. If they send after that then tough - BUT: long enough
2123 * that we won't make the old 4*rto = almost no time - whoops
2126 * Nope, it was not mistake. It is really desired behaviour
2127 * f.e. on http servers, when such sockets are useless, but
2128 * consume significant resources. Let's do it with special
2129 * linger2 option. --ANK
2132 if (sk->sk_state == TCP_FIN_WAIT2) {
2133 struct tcp_sock *tp = tcp_sk(sk);
2134 if (tp->linger2 < 0) {
2135 tcp_set_state(sk, TCP_CLOSE);
2136 tcp_send_active_reset(sk, GFP_ATOMIC);
2137 NET_INC_STATS_BH(sock_net(sk),
2138 LINUX_MIB_TCPABORTONLINGER);
2140 const int tmo = tcp_fin_time(sk);
2142 if (tmo > TCP_TIMEWAIT_LEN) {
2143 inet_csk_reset_keepalive_timer(sk,
2144 tmo - TCP_TIMEWAIT_LEN);
2146 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
2151 if (sk->sk_state != TCP_CLOSE) {
2153 if (tcp_check_oom(sk, 0)) {
2154 tcp_set_state(sk, TCP_CLOSE);
2155 tcp_send_active_reset(sk, GFP_ATOMIC);
2156 NET_INC_STATS_BH(sock_net(sk),
2157 LINUX_MIB_TCPABORTONMEMORY);
2161 if (sk->sk_state == TCP_CLOSE)
2162 inet_csk_destroy_sock(sk);
2163 /* Otherwise, socket is reprieved until protocol close. */
2170 EXPORT_SYMBOL(tcp_close);
2172 /* These states need RST on ABORT according to RFC793 */
2174 static inline int tcp_need_reset(int state)
2176 return (1 << state) &
2177 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 |
2178 TCPF_FIN_WAIT2 | TCPF_SYN_RECV);
2181 int tcp_disconnect(struct sock *sk, int flags)
2183 struct inet_sock *inet = inet_sk(sk);
2184 struct inet_connection_sock *icsk = inet_csk(sk);
2185 struct tcp_sock *tp = tcp_sk(sk);
2187 int old_state = sk->sk_state;
2189 if (old_state != TCP_CLOSE)
2190 tcp_set_state(sk, TCP_CLOSE);
2192 /* ABORT function of RFC793 */
2193 if (old_state == TCP_LISTEN) {
2194 inet_csk_listen_stop(sk);
2195 } else if (unlikely(tp->repair)) {
2196 sk->sk_err = ECONNABORTED;
2197 } else if (tcp_need_reset(old_state) ||
2198 (tp->snd_nxt != tp->write_seq &&
2199 (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK))) {
2200 /* The last check adjusts for discrepancy of Linux wrt. RFC
2203 tcp_send_active_reset(sk, gfp_any());
2204 sk->sk_err = ECONNRESET;
2205 } else if (old_state == TCP_SYN_SENT)
2206 sk->sk_err = ECONNRESET;
2208 tcp_clear_xmit_timers(sk);
2209 __skb_queue_purge(&sk->sk_receive_queue);
2210 tcp_write_queue_purge(sk);
2211 __skb_queue_purge(&tp->out_of_order_queue);
2212 #ifdef CONFIG_NET_DMA
2213 __skb_queue_purge(&sk->sk_async_wait_queue);
2216 inet->inet_dport = 0;
2218 if (!(sk->sk_userlocks & SOCK_BINDADDR_LOCK))
2219 inet_reset_saddr(sk);
2221 sk->sk_shutdown = 0;
2222 sock_reset_flag(sk, SOCK_DONE);
2224 if ((tp->write_seq += tp->max_window + 2) == 0)
2226 icsk->icsk_backoff = 0;
2228 icsk->icsk_probes_out = 0;
2229 tp->packets_out = 0;
2230 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
2231 tp->snd_cwnd_cnt = 0;
2232 tp->bytes_acked = 0;
2233 tp->window_clamp = 0;
2234 tcp_set_ca_state(sk, TCP_CA_Open);
2235 tcp_clear_retrans(tp);
2236 inet_csk_delack_init(sk);
2237 tcp_init_send_head(sk);
2238 memset(&tp->rx_opt, 0, sizeof(tp->rx_opt));
2241 WARN_ON(inet->inet_num && !icsk->icsk_bind_hash);
2243 sk->sk_error_report(sk);
2246 EXPORT_SYMBOL(tcp_disconnect);
2248 static inline int tcp_can_repair_sock(struct sock *sk)
2250 return capable(CAP_NET_ADMIN) &&
2251 ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_ESTABLISHED));
2254 static int tcp_repair_options_est(struct tcp_sock *tp,
2255 struct tcp_repair_opt __user *optbuf, unsigned int len)
2257 struct tcp_repair_opt opt;
2259 while (len >= sizeof(opt)) {
2260 if (copy_from_user(&opt, optbuf, sizeof(opt)))
2266 switch (opt.opt_code) {
2268 tp->rx_opt.mss_clamp = opt.opt_val;
2271 if (opt.opt_val > 14)
2274 tp->rx_opt.snd_wscale = opt.opt_val;
2276 case TCPOPT_SACK_PERM:
2277 if (opt.opt_val != 0)
2280 tp->rx_opt.sack_ok |= TCP_SACK_SEEN;
2281 if (sysctl_tcp_fack)
2282 tcp_enable_fack(tp);
2284 case TCPOPT_TIMESTAMP:
2285 if (opt.opt_val != 0)
2288 tp->rx_opt.tstamp_ok = 1;
2297 * Socket option code for TCP.
2299 static int do_tcp_setsockopt(struct sock *sk, int level,
2300 int optname, char __user *optval, unsigned int optlen)
2302 struct tcp_sock *tp = tcp_sk(sk);
2303 struct inet_connection_sock *icsk = inet_csk(sk);
2307 /* These are data/string values, all the others are ints */
2309 case TCP_CONGESTION: {
2310 char name[TCP_CA_NAME_MAX];
2315 val = strncpy_from_user(name, optval,
2316 min_t(long, TCP_CA_NAME_MAX-1, optlen));
2322 err = tcp_set_congestion_control(sk, name);
2326 case TCP_COOKIE_TRANSACTIONS: {
2327 struct tcp_cookie_transactions ctd;
2328 struct tcp_cookie_values *cvp = NULL;
2330 if (sizeof(ctd) > optlen)
2332 if (copy_from_user(&ctd, optval, sizeof(ctd)))
2335 if (ctd.tcpct_used > sizeof(ctd.tcpct_value) ||
2336 ctd.tcpct_s_data_desired > TCP_MSS_DESIRED)
2339 if (ctd.tcpct_cookie_desired == 0) {
2340 /* default to global value */
2341 } else if ((0x1 & ctd.tcpct_cookie_desired) ||
2342 ctd.tcpct_cookie_desired > TCP_COOKIE_MAX ||
2343 ctd.tcpct_cookie_desired < TCP_COOKIE_MIN) {
2347 if (TCP_COOKIE_OUT_NEVER & ctd.tcpct_flags) {
2348 /* Supercedes all other values */
2350 if (tp->cookie_values != NULL) {
2351 kref_put(&tp->cookie_values->kref,
2352 tcp_cookie_values_release);
2353 tp->cookie_values = NULL;
2355 tp->rx_opt.cookie_in_always = 0; /* false */
2356 tp->rx_opt.cookie_out_never = 1; /* true */
2361 /* Allocate ancillary memory before locking.
2363 if (ctd.tcpct_used > 0 ||
2364 (tp->cookie_values == NULL &&
2365 (sysctl_tcp_cookie_size > 0 ||
2366 ctd.tcpct_cookie_desired > 0 ||
2367 ctd.tcpct_s_data_desired > 0))) {
2368 cvp = kzalloc(sizeof(*cvp) + ctd.tcpct_used,
2373 kref_init(&cvp->kref);
2376 tp->rx_opt.cookie_in_always =
2377 (TCP_COOKIE_IN_ALWAYS & ctd.tcpct_flags);
2378 tp->rx_opt.cookie_out_never = 0; /* false */
2380 if (tp->cookie_values != NULL) {
2382 /* Changed values are recorded by a changed
2383 * pointer, ensuring the cookie will differ,
2384 * without separately hashing each value later.
2386 kref_put(&tp->cookie_values->kref,
2387 tcp_cookie_values_release);
2389 cvp = tp->cookie_values;
2394 cvp->cookie_desired = ctd.tcpct_cookie_desired;
2396 if (ctd.tcpct_used > 0) {
2397 memcpy(cvp->s_data_payload, ctd.tcpct_value,
2399 cvp->s_data_desired = ctd.tcpct_used;
2400 cvp->s_data_constant = 1; /* true */
2402 /* No constant payload data. */
2403 cvp->s_data_desired = ctd.tcpct_s_data_desired;
2404 cvp->s_data_constant = 0; /* false */
2407 tp->cookie_values = cvp;
2417 if (optlen < sizeof(int))
2420 if (get_user(val, (int __user *)optval))
2427 /* Values greater than interface MTU won't take effect. However
2428 * at the point when this call is done we typically don't yet
2429 * know which interface is going to be used */
2430 if (val < TCP_MIN_MSS || val > MAX_TCP_WINDOW) {
2434 tp->rx_opt.user_mss = val;
2439 /* TCP_NODELAY is weaker than TCP_CORK, so that
2440 * this option on corked socket is remembered, but
2441 * it is not activated until cork is cleared.
2443 * However, when TCP_NODELAY is set we make
2444 * an explicit push, which overrides even TCP_CORK
2445 * for currently queued segments.
2447 tp->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH;
2448 tcp_push_pending_frames(sk);
2450 tp->nonagle &= ~TCP_NAGLE_OFF;
2454 case TCP_THIN_LINEAR_TIMEOUTS:
2455 if (val < 0 || val > 1)
2461 case TCP_THIN_DUPACK:
2462 if (val < 0 || val > 1)
2465 tp->thin_dupack = val;
2466 if (tp->thin_dupack)
2467 tcp_disable_early_retrans(tp);
2471 if (!tcp_can_repair_sock(sk))
2473 else if (val == 1) {
2475 sk->sk_reuse = SK_FORCE_REUSE;
2476 tp->repair_queue = TCP_NO_QUEUE;
2477 } else if (val == 0) {
2479 sk->sk_reuse = SK_NO_REUSE;
2480 tcp_send_window_probe(sk);
2486 case TCP_REPAIR_QUEUE:
2489 else if (val < TCP_QUEUES_NR)
2490 tp->repair_queue = val;
2496 if (sk->sk_state != TCP_CLOSE)
2498 else if (tp->repair_queue == TCP_SEND_QUEUE)
2499 tp->write_seq = val;
2500 else if (tp->repair_queue == TCP_RECV_QUEUE)
2506 case TCP_REPAIR_OPTIONS:
2509 else if (sk->sk_state == TCP_ESTABLISHED)
2510 err = tcp_repair_options_est(tp,
2511 (struct tcp_repair_opt __user *)optval,
2518 /* When set indicates to always queue non-full frames.
2519 * Later the user clears this option and we transmit
2520 * any pending partial frames in the queue. This is
2521 * meant to be used alongside sendfile() to get properly
2522 * filled frames when the user (for example) must write
2523 * out headers with a write() call first and then use
2524 * sendfile to send out the data parts.
2526 * TCP_CORK can be set together with TCP_NODELAY and it is
2527 * stronger than TCP_NODELAY.
2530 tp->nonagle |= TCP_NAGLE_CORK;
2532 tp->nonagle &= ~TCP_NAGLE_CORK;
2533 if (tp->nonagle&TCP_NAGLE_OFF)
2534 tp->nonagle |= TCP_NAGLE_PUSH;
2535 tcp_push_pending_frames(sk);
2540 if (val < 1 || val > MAX_TCP_KEEPIDLE)
2543 tp->keepalive_time = val * HZ;
2544 if (sock_flag(sk, SOCK_KEEPOPEN) &&
2545 !((1 << sk->sk_state) &
2546 (TCPF_CLOSE | TCPF_LISTEN))) {
2547 u32 elapsed = keepalive_time_elapsed(tp);
2548 if (tp->keepalive_time > elapsed)
2549 elapsed = tp->keepalive_time - elapsed;
2552 inet_csk_reset_keepalive_timer(sk, elapsed);
2557 if (val < 1 || val > MAX_TCP_KEEPINTVL)
2560 tp->keepalive_intvl = val * HZ;
2563 if (val < 1 || val > MAX_TCP_KEEPCNT)
2566 tp->keepalive_probes = val;
2569 if (val < 1 || val > MAX_TCP_SYNCNT)
2572 icsk->icsk_syn_retries = val;
2578 else if (val > sysctl_tcp_fin_timeout / HZ)
2581 tp->linger2 = val * HZ;
2584 case TCP_DEFER_ACCEPT:
2585 /* Translate value in seconds to number of retransmits */
2586 icsk->icsk_accept_queue.rskq_defer_accept =
2587 secs_to_retrans(val, TCP_TIMEOUT_INIT / HZ,
2591 case TCP_WINDOW_CLAMP:
2593 if (sk->sk_state != TCP_CLOSE) {
2597 tp->window_clamp = 0;
2599 tp->window_clamp = val < SOCK_MIN_RCVBUF / 2 ?
2600 SOCK_MIN_RCVBUF / 2 : val;
2605 icsk->icsk_ack.pingpong = 1;
2607 icsk->icsk_ack.pingpong = 0;
2608 if ((1 << sk->sk_state) &
2609 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) &&
2610 inet_csk_ack_scheduled(sk)) {
2611 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
2612 tcp_cleanup_rbuf(sk, 1);
2614 icsk->icsk_ack.pingpong = 1;
2619 #ifdef CONFIG_TCP_MD5SIG
2621 /* Read the IP->Key mappings from userspace */
2622 err = tp->af_specific->md5_parse(sk, optval, optlen);
2625 case TCP_USER_TIMEOUT:
2626 /* Cap the max timeout in ms TCP will retry/retrans
2627 * before giving up and aborting (ETIMEDOUT) a connection.
2629 icsk->icsk_user_timeout = msecs_to_jiffies(val);
2640 int tcp_setsockopt(struct sock *sk, int level, int optname, char __user *optval,
2641 unsigned int optlen)
2643 const struct inet_connection_sock *icsk = inet_csk(sk);
2645 if (level != SOL_TCP)
2646 return icsk->icsk_af_ops->setsockopt(sk, level, optname,
2648 return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2650 EXPORT_SYMBOL(tcp_setsockopt);
2652 #ifdef CONFIG_COMPAT
2653 int compat_tcp_setsockopt(struct sock *sk, int level, int optname,
2654 char __user *optval, unsigned int optlen)
2656 if (level != SOL_TCP)
2657 return inet_csk_compat_setsockopt(sk, level, optname,
2659 return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2661 EXPORT_SYMBOL(compat_tcp_setsockopt);
2664 /* Return information about state of tcp endpoint in API format. */
2665 void tcp_get_info(const struct sock *sk, struct tcp_info *info)
2667 const struct tcp_sock *tp = tcp_sk(sk);
2668 const struct inet_connection_sock *icsk = inet_csk(sk);
2669 u32 now = tcp_time_stamp;
2671 memset(info, 0, sizeof(*info));
2673 info->tcpi_state = sk->sk_state;
2674 info->tcpi_ca_state = icsk->icsk_ca_state;
2675 info->tcpi_retransmits = icsk->icsk_retransmits;
2676 info->tcpi_probes = icsk->icsk_probes_out;
2677 info->tcpi_backoff = icsk->icsk_backoff;
2679 if (tp->rx_opt.tstamp_ok)
2680 info->tcpi_options |= TCPI_OPT_TIMESTAMPS;
2681 if (tcp_is_sack(tp))
2682 info->tcpi_options |= TCPI_OPT_SACK;
2683 if (tp->rx_opt.wscale_ok) {
2684 info->tcpi_options |= TCPI_OPT_WSCALE;
2685 info->tcpi_snd_wscale = tp->rx_opt.snd_wscale;
2686 info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale;
2689 if (tp->ecn_flags & TCP_ECN_OK)
2690 info->tcpi_options |= TCPI_OPT_ECN;
2691 if (tp->ecn_flags & TCP_ECN_SEEN)
2692 info->tcpi_options |= TCPI_OPT_ECN_SEEN;
2694 info->tcpi_rto = jiffies_to_usecs(icsk->icsk_rto);
2695 info->tcpi_ato = jiffies_to_usecs(icsk->icsk_ack.ato);
2696 info->tcpi_snd_mss = tp->mss_cache;
2697 info->tcpi_rcv_mss = icsk->icsk_ack.rcv_mss;
2699 if (sk->sk_state == TCP_LISTEN) {
2700 info->tcpi_unacked = sk->sk_ack_backlog;
2701 info->tcpi_sacked = sk->sk_max_ack_backlog;
2703 info->tcpi_unacked = tp->packets_out;
2704 info->tcpi_sacked = tp->sacked_out;
2706 info->tcpi_lost = tp->lost_out;
2707 info->tcpi_retrans = tp->retrans_out;
2708 info->tcpi_fackets = tp->fackets_out;
2710 info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime);
2711 info->tcpi_last_data_recv = jiffies_to_msecs(now - icsk->icsk_ack.lrcvtime);
2712 info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp);
2714 info->tcpi_pmtu = icsk->icsk_pmtu_cookie;
2715 info->tcpi_rcv_ssthresh = tp->rcv_ssthresh;
2716 info->tcpi_rtt = jiffies_to_usecs(tp->srtt)>>3;
2717 info->tcpi_rttvar = jiffies_to_usecs(tp->mdev)>>2;
2718 info->tcpi_snd_ssthresh = tp->snd_ssthresh;
2719 info->tcpi_snd_cwnd = tp->snd_cwnd;
2720 info->tcpi_advmss = tp->advmss;
2721 info->tcpi_reordering = tp->reordering;
2723 info->tcpi_rcv_rtt = jiffies_to_usecs(tp->rcv_rtt_est.rtt)>>3;
2724 info->tcpi_rcv_space = tp->rcvq_space.space;
2726 info->tcpi_total_retrans = tp->total_retrans;
2728 EXPORT_SYMBOL_GPL(tcp_get_info);
2730 static int do_tcp_getsockopt(struct sock *sk, int level,
2731 int optname, char __user *optval, int __user *optlen)
2733 struct inet_connection_sock *icsk = inet_csk(sk);
2734 struct tcp_sock *tp = tcp_sk(sk);
2737 if (get_user(len, optlen))
2740 len = min_t(unsigned int, len, sizeof(int));
2747 val = tp->mss_cache;
2748 if (!val && ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
2749 val = tp->rx_opt.user_mss;
2751 val = tp->rx_opt.mss_clamp;
2754 val = !!(tp->nonagle&TCP_NAGLE_OFF);
2757 val = !!(tp->nonagle&TCP_NAGLE_CORK);
2760 val = keepalive_time_when(tp) / HZ;
2763 val = keepalive_intvl_when(tp) / HZ;
2766 val = keepalive_probes(tp);
2769 val = icsk->icsk_syn_retries ? : sysctl_tcp_syn_retries;
2774 val = (val ? : sysctl_tcp_fin_timeout) / HZ;
2776 case TCP_DEFER_ACCEPT:
2777 val = retrans_to_secs(icsk->icsk_accept_queue.rskq_defer_accept,
2778 TCP_TIMEOUT_INIT / HZ, TCP_RTO_MAX / HZ);
2780 case TCP_WINDOW_CLAMP:
2781 val = tp->window_clamp;
2784 struct tcp_info info;
2786 if (get_user(len, optlen))
2789 tcp_get_info(sk, &info);
2791 len = min_t(unsigned int, len, sizeof(info));
2792 if (put_user(len, optlen))
2794 if (copy_to_user(optval, &info, len))
2799 val = !icsk->icsk_ack.pingpong;
2802 case TCP_CONGESTION:
2803 if (get_user(len, optlen))
2805 len = min_t(unsigned int, len, TCP_CA_NAME_MAX);
2806 if (put_user(len, optlen))
2808 if (copy_to_user(optval, icsk->icsk_ca_ops->name, len))
2812 case TCP_COOKIE_TRANSACTIONS: {
2813 struct tcp_cookie_transactions ctd;
2814 struct tcp_cookie_values *cvp = tp->cookie_values;
2816 if (get_user(len, optlen))
2818 if (len < sizeof(ctd))
2821 memset(&ctd, 0, sizeof(ctd));
2822 ctd.tcpct_flags = (tp->rx_opt.cookie_in_always ?
2823 TCP_COOKIE_IN_ALWAYS : 0)
2824 | (tp->rx_opt.cookie_out_never ?
2825 TCP_COOKIE_OUT_NEVER : 0);
2828 ctd.tcpct_flags |= (cvp->s_data_in ?
2830 | (cvp->s_data_out ?
2831 TCP_S_DATA_OUT : 0);
2833 ctd.tcpct_cookie_desired = cvp->cookie_desired;
2834 ctd.tcpct_s_data_desired = cvp->s_data_desired;
2836 memcpy(&ctd.tcpct_value[0], &cvp->cookie_pair[0],
2837 cvp->cookie_pair_size);
2838 ctd.tcpct_used = cvp->cookie_pair_size;
2841 if (put_user(sizeof(ctd), optlen))
2843 if (copy_to_user(optval, &ctd, sizeof(ctd)))
2847 case TCP_THIN_LINEAR_TIMEOUTS:
2850 case TCP_THIN_DUPACK:
2851 val = tp->thin_dupack;
2858 case TCP_REPAIR_QUEUE:
2860 val = tp->repair_queue;
2866 if (tp->repair_queue == TCP_SEND_QUEUE)
2867 val = tp->write_seq;
2868 else if (tp->repair_queue == TCP_RECV_QUEUE)
2874 case TCP_USER_TIMEOUT:
2875 val = jiffies_to_msecs(icsk->icsk_user_timeout);
2878 return -ENOPROTOOPT;
2881 if (put_user(len, optlen))
2883 if (copy_to_user(optval, &val, len))
2888 int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval,
2891 struct inet_connection_sock *icsk = inet_csk(sk);
2893 if (level != SOL_TCP)
2894 return icsk->icsk_af_ops->getsockopt(sk, level, optname,
2896 return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2898 EXPORT_SYMBOL(tcp_getsockopt);
2900 #ifdef CONFIG_COMPAT
2901 int compat_tcp_getsockopt(struct sock *sk, int level, int optname,
2902 char __user *optval, int __user *optlen)
2904 if (level != SOL_TCP)
2905 return inet_csk_compat_getsockopt(sk, level, optname,
2907 return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2909 EXPORT_SYMBOL(compat_tcp_getsockopt);
2912 struct sk_buff *tcp_tso_segment(struct sk_buff *skb,
2913 netdev_features_t features)
2915 struct sk_buff *segs = ERR_PTR(-EINVAL);
2920 unsigned int oldlen;
2923 if (!pskb_may_pull(skb, sizeof(*th)))
2927 thlen = th->doff * 4;
2928 if (thlen < sizeof(*th))
2931 if (!pskb_may_pull(skb, thlen))
2934 oldlen = (u16)~skb->len;
2935 __skb_pull(skb, thlen);
2937 mss = skb_shinfo(skb)->gso_size;
2938 if (unlikely(skb->len <= mss))
2941 if (skb_gso_ok(skb, features | NETIF_F_GSO_ROBUST)) {
2942 /* Packet is from an untrusted source, reset gso_segs. */
2943 int type = skb_shinfo(skb)->gso_type;
2951 !(type & (SKB_GSO_TCPV4 | SKB_GSO_TCPV6))))
2954 skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss);
2960 segs = skb_segment(skb, features);
2964 delta = htonl(oldlen + (thlen + mss));
2968 seq = ntohl(th->seq);
2971 th->fin = th->psh = 0;
2973 th->check = ~csum_fold((__force __wsum)((__force u32)th->check +
2974 (__force u32)delta));
2975 if (skb->ip_summed != CHECKSUM_PARTIAL)
2977 csum_fold(csum_partial(skb_transport_header(skb),
2984 th->seq = htonl(seq);
2986 } while (skb->next);
2988 delta = htonl(oldlen + (skb->tail - skb->transport_header) +
2990 th->check = ~csum_fold((__force __wsum)((__force u32)th->check +
2991 (__force u32)delta));
2992 if (skb->ip_summed != CHECKSUM_PARTIAL)
2993 th->check = csum_fold(csum_partial(skb_transport_header(skb),
2999 EXPORT_SYMBOL(tcp_tso_segment);
3001 struct sk_buff **tcp_gro_receive(struct sk_buff **head, struct sk_buff *skb)
3003 struct sk_buff **pp = NULL;
3010 unsigned int mss = 1;
3016 off = skb_gro_offset(skb);
3017 hlen = off + sizeof(*th);
3018 th = skb_gro_header_fast(skb, off);
3019 if (skb_gro_header_hard(skb, hlen)) {
3020 th = skb_gro_header_slow(skb, hlen, off);
3025 thlen = th->doff * 4;
3026 if (thlen < sizeof(*th))
3030 if (skb_gro_header_hard(skb, hlen)) {
3031 th = skb_gro_header_slow(skb, hlen, off);
3036 skb_gro_pull(skb, thlen);
3038 len = skb_gro_len(skb);
3039 flags = tcp_flag_word(th);
3041 for (; (p = *head); head = &p->next) {
3042 if (!NAPI_GRO_CB(p)->same_flow)
3047 if (*(u32 *)&th->source ^ *(u32 *)&th2->source) {
3048 NAPI_GRO_CB(p)->same_flow = 0;
3055 goto out_check_final;
3058 flush = NAPI_GRO_CB(p)->flush;
3059 flush |= (__force int)(flags & TCP_FLAG_CWR);
3060 flush |= (__force int)((flags ^ tcp_flag_word(th2)) &
3061 ~(TCP_FLAG_CWR | TCP_FLAG_FIN | TCP_FLAG_PSH));
3062 flush |= (__force int)(th->ack_seq ^ th2->ack_seq);
3063 for (i = sizeof(*th); i < thlen; i += 4)
3064 flush |= *(u32 *)((u8 *)th + i) ^
3065 *(u32 *)((u8 *)th2 + i);
3067 mss = skb_shinfo(p)->gso_size;
3069 flush |= (len - 1) >= mss;
3070 flush |= (ntohl(th2->seq) + skb_gro_len(p)) ^ ntohl(th->seq);
3072 if (flush || skb_gro_receive(head, skb)) {
3074 goto out_check_final;
3079 tcp_flag_word(th2) |= flags & (TCP_FLAG_FIN | TCP_FLAG_PSH);
3083 flush |= (__force int)(flags & (TCP_FLAG_URG | TCP_FLAG_PSH |
3084 TCP_FLAG_RST | TCP_FLAG_SYN |
3087 if (p && (!NAPI_GRO_CB(skb)->same_flow || flush))
3091 NAPI_GRO_CB(skb)->flush |= flush;
3095 EXPORT_SYMBOL(tcp_gro_receive);
3097 int tcp_gro_complete(struct sk_buff *skb)
3099 struct tcphdr *th = tcp_hdr(skb);
3101 skb->csum_start = skb_transport_header(skb) - skb->head;
3102 skb->csum_offset = offsetof(struct tcphdr, check);
3103 skb->ip_summed = CHECKSUM_PARTIAL;
3105 skb_shinfo(skb)->gso_segs = NAPI_GRO_CB(skb)->count;
3108 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
3112 EXPORT_SYMBOL(tcp_gro_complete);
3114 #ifdef CONFIG_TCP_MD5SIG
3115 static unsigned long tcp_md5sig_users;
3116 static struct tcp_md5sig_pool __percpu *tcp_md5sig_pool;
3117 static DEFINE_SPINLOCK(tcp_md5sig_pool_lock);
3119 static void __tcp_free_md5sig_pool(struct tcp_md5sig_pool __percpu *pool)
3123 for_each_possible_cpu(cpu) {
3124 struct tcp_md5sig_pool *p = per_cpu_ptr(pool, cpu);
3126 if (p->md5_desc.tfm)
3127 crypto_free_hash(p->md5_desc.tfm);
3132 void tcp_free_md5sig_pool(void)
3134 struct tcp_md5sig_pool __percpu *pool = NULL;
3136 spin_lock_bh(&tcp_md5sig_pool_lock);
3137 if (--tcp_md5sig_users == 0) {
3138 pool = tcp_md5sig_pool;
3139 tcp_md5sig_pool = NULL;
3141 spin_unlock_bh(&tcp_md5sig_pool_lock);
3143 __tcp_free_md5sig_pool(pool);
3145 EXPORT_SYMBOL(tcp_free_md5sig_pool);
3147 static struct tcp_md5sig_pool __percpu *
3148 __tcp_alloc_md5sig_pool(struct sock *sk)
3151 struct tcp_md5sig_pool __percpu *pool;
3153 pool = alloc_percpu(struct tcp_md5sig_pool);
3157 for_each_possible_cpu(cpu) {
3158 struct crypto_hash *hash;
3160 hash = crypto_alloc_hash("md5", 0, CRYPTO_ALG_ASYNC);
3161 if (!hash || IS_ERR(hash))
3164 per_cpu_ptr(pool, cpu)->md5_desc.tfm = hash;
3168 __tcp_free_md5sig_pool(pool);
3172 struct tcp_md5sig_pool __percpu *tcp_alloc_md5sig_pool(struct sock *sk)
3174 struct tcp_md5sig_pool __percpu *pool;
3178 spin_lock_bh(&tcp_md5sig_pool_lock);
3179 pool = tcp_md5sig_pool;
3180 if (tcp_md5sig_users++ == 0) {
3182 spin_unlock_bh(&tcp_md5sig_pool_lock);
3185 spin_unlock_bh(&tcp_md5sig_pool_lock);
3189 spin_unlock_bh(&tcp_md5sig_pool_lock);
3192 /* we cannot hold spinlock here because this may sleep. */
3193 struct tcp_md5sig_pool __percpu *p;
3195 p = __tcp_alloc_md5sig_pool(sk);
3196 spin_lock_bh(&tcp_md5sig_pool_lock);
3199 spin_unlock_bh(&tcp_md5sig_pool_lock);
3202 pool = tcp_md5sig_pool;
3204 /* oops, it has already been assigned. */
3205 spin_unlock_bh(&tcp_md5sig_pool_lock);
3206 __tcp_free_md5sig_pool(p);
3208 tcp_md5sig_pool = pool = p;
3209 spin_unlock_bh(&tcp_md5sig_pool_lock);
3214 EXPORT_SYMBOL(tcp_alloc_md5sig_pool);
3218 * tcp_get_md5sig_pool - get md5sig_pool for this user
3220 * We use percpu structure, so if we succeed, we exit with preemption
3221 * and BH disabled, to make sure another thread or softirq handling
3222 * wont try to get same context.
3224 struct tcp_md5sig_pool *tcp_get_md5sig_pool(void)
3226 struct tcp_md5sig_pool __percpu *p;
3230 spin_lock(&tcp_md5sig_pool_lock);
3231 p = tcp_md5sig_pool;
3234 spin_unlock(&tcp_md5sig_pool_lock);
3237 return this_cpu_ptr(p);
3242 EXPORT_SYMBOL(tcp_get_md5sig_pool);
3244 void tcp_put_md5sig_pool(void)
3247 tcp_free_md5sig_pool();
3249 EXPORT_SYMBOL(tcp_put_md5sig_pool);
3251 int tcp_md5_hash_header(struct tcp_md5sig_pool *hp,
3252 const struct tcphdr *th)
3254 struct scatterlist sg;
3258 /* We are not allowed to change tcphdr, make a local copy */
3259 memcpy(&hdr, th, sizeof(hdr));
3262 /* options aren't included in the hash */
3263 sg_init_one(&sg, &hdr, sizeof(hdr));
3264 err = crypto_hash_update(&hp->md5_desc, &sg, sizeof(hdr));
3267 EXPORT_SYMBOL(tcp_md5_hash_header);
3269 int tcp_md5_hash_skb_data(struct tcp_md5sig_pool *hp,
3270 const struct sk_buff *skb, unsigned int header_len)
3272 struct scatterlist sg;
3273 const struct tcphdr *tp = tcp_hdr(skb);
3274 struct hash_desc *desc = &hp->md5_desc;
3276 const unsigned int head_data_len = skb_headlen(skb) > header_len ?
3277 skb_headlen(skb) - header_len : 0;
3278 const struct skb_shared_info *shi = skb_shinfo(skb);
3279 struct sk_buff *frag_iter;
3281 sg_init_table(&sg, 1);
3283 sg_set_buf(&sg, ((u8 *) tp) + header_len, head_data_len);
3284 if (crypto_hash_update(desc, &sg, head_data_len))
3287 for (i = 0; i < shi->nr_frags; ++i) {
3288 const struct skb_frag_struct *f = &shi->frags[i];
3289 struct page *page = skb_frag_page(f);
3290 sg_set_page(&sg, page, skb_frag_size(f), f->page_offset);
3291 if (crypto_hash_update(desc, &sg, skb_frag_size(f)))
3295 skb_walk_frags(skb, frag_iter)
3296 if (tcp_md5_hash_skb_data(hp, frag_iter, 0))
3301 EXPORT_SYMBOL(tcp_md5_hash_skb_data);
3303 int tcp_md5_hash_key(struct tcp_md5sig_pool *hp, const struct tcp_md5sig_key *key)
3305 struct scatterlist sg;
3307 sg_init_one(&sg, key->key, key->keylen);
3308 return crypto_hash_update(&hp->md5_desc, &sg, key->keylen);
3310 EXPORT_SYMBOL(tcp_md5_hash_key);
3315 * Each Responder maintains up to two secret values concurrently for
3316 * efficient secret rollover. Each secret value has 4 states:
3318 * Generating. (tcp_secret_generating != tcp_secret_primary)
3319 * Generates new Responder-Cookies, but not yet used for primary
3320 * verification. This is a short-term state, typically lasting only
3321 * one round trip time (RTT).
3323 * Primary. (tcp_secret_generating == tcp_secret_primary)
3324 * Used both for generation and primary verification.
3326 * Retiring. (tcp_secret_retiring != tcp_secret_secondary)
3327 * Used for verification, until the first failure that can be
3328 * verified by the newer Generating secret. At that time, this
3329 * cookie's state is changed to Secondary, and the Generating
3330 * cookie's state is changed to Primary. This is a short-term state,
3331 * typically lasting only one round trip time (RTT).
3333 * Secondary. (tcp_secret_retiring == tcp_secret_secondary)
3334 * Used for secondary verification, after primary verification
3335 * failures. This state lasts no more than twice the Maximum Segment
3336 * Lifetime (2MSL). Then, the secret is discarded.
3338 struct tcp_cookie_secret {
3339 /* The secret is divided into two parts. The digest part is the
3340 * equivalent of previously hashing a secret and saving the state,
3341 * and serves as an initialization vector (IV). The message part
3342 * serves as the trailing secret.
3344 u32 secrets[COOKIE_WORKSPACE_WORDS];
3345 unsigned long expires;
3348 #define TCP_SECRET_1MSL (HZ * TCP_PAWS_MSL)
3349 #define TCP_SECRET_2MSL (HZ * TCP_PAWS_MSL * 2)
3350 #define TCP_SECRET_LIFE (HZ * 600)
3352 static struct tcp_cookie_secret tcp_secret_one;
3353 static struct tcp_cookie_secret tcp_secret_two;
3355 /* Essentially a circular list, without dynamic allocation. */
3356 static struct tcp_cookie_secret *tcp_secret_generating;
3357 static struct tcp_cookie_secret *tcp_secret_primary;
3358 static struct tcp_cookie_secret *tcp_secret_retiring;
3359 static struct tcp_cookie_secret *tcp_secret_secondary;
3361 static DEFINE_SPINLOCK(tcp_secret_locker);
3363 /* Select a pseudo-random word in the cookie workspace.
3365 static inline u32 tcp_cookie_work(const u32 *ws, const int n)
3367 return ws[COOKIE_DIGEST_WORDS + ((COOKIE_MESSAGE_WORDS-1) & ws[n])];
3370 /* Fill bakery[COOKIE_WORKSPACE_WORDS] with generator, updating as needed.
3371 * Called in softirq context.
3372 * Returns: 0 for success.
3374 int tcp_cookie_generator(u32 *bakery)
3376 unsigned long jiffy = jiffies;
3378 if (unlikely(time_after_eq(jiffy, tcp_secret_generating->expires))) {
3379 spin_lock_bh(&tcp_secret_locker);
3380 if (!time_after_eq(jiffy, tcp_secret_generating->expires)) {
3381 /* refreshed by another */
3383 &tcp_secret_generating->secrets[0],
3384 COOKIE_WORKSPACE_WORDS);
3386 /* still needs refreshing */
3387 get_random_bytes(bakery, COOKIE_WORKSPACE_WORDS);
3389 /* The first time, paranoia assumes that the
3390 * randomization function isn't as strong. But,
3391 * this secret initialization is delayed until
3392 * the last possible moment (packet arrival).
3393 * Although that time is observable, it is
3394 * unpredictably variable. Mash in the most
3395 * volatile clock bits available, and expire the
3396 * secret extra quickly.
3398 if (unlikely(tcp_secret_primary->expires ==
3399 tcp_secret_secondary->expires)) {
3402 getnstimeofday(&tv);
3403 bakery[COOKIE_DIGEST_WORDS+0] ^=
3406 tcp_secret_secondary->expires = jiffy
3408 + (0x0f & tcp_cookie_work(bakery, 0));
3410 tcp_secret_secondary->expires = jiffy
3412 + (0xff & tcp_cookie_work(bakery, 1));
3413 tcp_secret_primary->expires = jiffy
3415 + (0x1f & tcp_cookie_work(bakery, 2));
3417 memcpy(&tcp_secret_secondary->secrets[0],
3418 bakery, COOKIE_WORKSPACE_WORDS);
3420 rcu_assign_pointer(tcp_secret_generating,
3421 tcp_secret_secondary);
3422 rcu_assign_pointer(tcp_secret_retiring,
3423 tcp_secret_primary);
3425 * Neither call_rcu() nor synchronize_rcu() needed.
3426 * Retiring data is not freed. It is replaced after
3427 * further (locked) pointer updates, and a quiet time
3428 * (minimum 1MSL, maximum LIFE - 2MSL).
3431 spin_unlock_bh(&tcp_secret_locker);
3435 &rcu_dereference(tcp_secret_generating)->secrets[0],
3436 COOKIE_WORKSPACE_WORDS);
3437 rcu_read_unlock_bh();
3441 EXPORT_SYMBOL(tcp_cookie_generator);
3443 void tcp_done(struct sock *sk)
3445 if (sk->sk_state == TCP_SYN_SENT || sk->sk_state == TCP_SYN_RECV)
3446 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_ATTEMPTFAILS);
3448 tcp_set_state(sk, TCP_CLOSE);
3449 tcp_clear_xmit_timers(sk);
3451 sk->sk_shutdown = SHUTDOWN_MASK;
3453 if (!sock_flag(sk, SOCK_DEAD))
3454 sk->sk_state_change(sk);
3456 inet_csk_destroy_sock(sk);
3458 EXPORT_SYMBOL_GPL(tcp_done);
3460 extern struct tcp_congestion_ops tcp_reno;
3462 static __initdata unsigned long thash_entries;
3463 static int __init set_thash_entries(char *str)
3467 thash_entries = simple_strtoul(str, &str, 0);
3470 __setup("thash_entries=", set_thash_entries);
3472 void tcp_init_mem(struct net *net)
3474 unsigned long limit = nr_free_buffer_pages() / 8;
3475 limit = max(limit, 128UL);
3476 net->ipv4.sysctl_tcp_mem[0] = limit / 4 * 3;
3477 net->ipv4.sysctl_tcp_mem[1] = limit;
3478 net->ipv4.sysctl_tcp_mem[2] = net->ipv4.sysctl_tcp_mem[0] * 2;
3481 void __init tcp_init(void)
3483 struct sk_buff *skb = NULL;
3484 unsigned long limit;
3485 int max_rshare, max_wshare, cnt;
3487 unsigned long jiffy = jiffies;
3489 BUILD_BUG_ON(sizeof(struct tcp_skb_cb) > sizeof(skb->cb));
3491 percpu_counter_init(&tcp_sockets_allocated, 0);
3492 percpu_counter_init(&tcp_orphan_count, 0);
3493 tcp_hashinfo.bind_bucket_cachep =
3494 kmem_cache_create("tcp_bind_bucket",
3495 sizeof(struct inet_bind_bucket), 0,
3496 SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL);
3498 /* Size and allocate the main established and bind bucket
3501 * The methodology is similar to that of the buffer cache.
3503 tcp_hashinfo.ehash =
3504 alloc_large_system_hash("TCP established",
3505 sizeof(struct inet_ehash_bucket),
3507 (totalram_pages >= 128 * 1024) ?
3511 &tcp_hashinfo.ehash_mask,
3512 thash_entries ? 0 : 512 * 1024);
3513 for (i = 0; i <= tcp_hashinfo.ehash_mask; i++) {
3514 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].chain, i);
3515 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].twchain, i);
3517 if (inet_ehash_locks_alloc(&tcp_hashinfo))
3518 panic("TCP: failed to alloc ehash_locks");
3519 tcp_hashinfo.bhash =
3520 alloc_large_system_hash("TCP bind",
3521 sizeof(struct inet_bind_hashbucket),
3522 tcp_hashinfo.ehash_mask + 1,
3523 (totalram_pages >= 128 * 1024) ?
3526 &tcp_hashinfo.bhash_size,
3529 tcp_hashinfo.bhash_size = 1U << tcp_hashinfo.bhash_size;
3530 for (i = 0; i < tcp_hashinfo.bhash_size; i++) {
3531 spin_lock_init(&tcp_hashinfo.bhash[i].lock);
3532 INIT_HLIST_HEAD(&tcp_hashinfo.bhash[i].chain);
3536 cnt = tcp_hashinfo.ehash_mask + 1;
3538 tcp_death_row.sysctl_max_tw_buckets = cnt / 2;
3539 sysctl_tcp_max_orphans = cnt / 2;
3540 sysctl_max_syn_backlog = max(128, cnt / 256);
3542 tcp_init_mem(&init_net);
3543 /* Set per-socket limits to no more than 1/128 the pressure threshold */
3544 limit = nr_free_buffer_pages() << (PAGE_SHIFT - 7);
3545 max_wshare = min(4UL*1024*1024, limit);
3546 max_rshare = min(6UL*1024*1024, limit);
3548 sysctl_tcp_wmem[0] = SK_MEM_QUANTUM;
3549 sysctl_tcp_wmem[1] = 16*1024;
3550 sysctl_tcp_wmem[2] = max(64*1024, max_wshare);
3552 sysctl_tcp_rmem[0] = SK_MEM_QUANTUM;
3553 sysctl_tcp_rmem[1] = 87380;
3554 sysctl_tcp_rmem[2] = max(87380, max_rshare);
3556 pr_info("Hash tables configured (established %u bind %u)\n",
3557 tcp_hashinfo.ehash_mask + 1, tcp_hashinfo.bhash_size);
3559 tcp_register_congestion_control(&tcp_reno);
3561 memset(&tcp_secret_one.secrets[0], 0, sizeof(tcp_secret_one.secrets));
3562 memset(&tcp_secret_two.secrets[0], 0, sizeof(tcp_secret_two.secrets));
3563 tcp_secret_one.expires = jiffy; /* past due */
3564 tcp_secret_two.expires = jiffy; /* past due */
3565 tcp_secret_generating = &tcp_secret_one;
3566 tcp_secret_primary = &tcp_secret_one;
3567 tcp_secret_retiring = &tcp_secret_two;
3568 tcp_secret_secondary = &tcp_secret_two;