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>
273 #include <net/inet_common.h>
275 #include <net/xfrm.h>
277 #include <net/netdma.h>
278 #include <net/sock.h>
280 #include <asm/uaccess.h>
281 #include <asm/ioctls.h>
283 int sysctl_tcp_fin_timeout __read_mostly = TCP_FIN_TIMEOUT;
285 struct percpu_counter tcp_orphan_count;
286 EXPORT_SYMBOL_GPL(tcp_orphan_count);
288 int sysctl_tcp_wmem[3] __read_mostly;
289 int sysctl_tcp_rmem[3] __read_mostly;
291 EXPORT_SYMBOL(sysctl_tcp_rmem);
292 EXPORT_SYMBOL(sysctl_tcp_wmem);
294 atomic_long_t tcp_memory_allocated; /* Current allocated memory. */
295 EXPORT_SYMBOL(tcp_memory_allocated);
298 * Current number of TCP sockets.
300 struct percpu_counter tcp_sockets_allocated;
301 EXPORT_SYMBOL(tcp_sockets_allocated);
306 struct tcp_splice_state {
307 struct pipe_inode_info *pipe;
313 * Pressure flag: try to collapse.
314 * Technical note: it is used by multiple contexts non atomically.
315 * All the __sk_mem_schedule() is of this nature: accounting
316 * is strict, actions are advisory and have some latency.
318 int tcp_memory_pressure __read_mostly;
319 EXPORT_SYMBOL(tcp_memory_pressure);
321 void tcp_enter_memory_pressure(struct sock *sk)
323 if (!tcp_memory_pressure) {
324 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMEMORYPRESSURES);
325 tcp_memory_pressure = 1;
328 EXPORT_SYMBOL(tcp_enter_memory_pressure);
330 /* Convert seconds to retransmits based on initial and max timeout */
331 static u8 secs_to_retrans(int seconds, int timeout, int rto_max)
336 int period = timeout;
339 while (seconds > period && res < 255) {
342 if (timeout > rto_max)
350 /* Convert retransmits to seconds based on initial and max timeout */
351 static int retrans_to_secs(u8 retrans, int timeout, int rto_max)
359 if (timeout > rto_max)
367 /* Address-family independent initialization for a tcp_sock.
369 * NOTE: A lot of things set to zero explicitly by call to
370 * sk_alloc() so need not be done here.
372 void tcp_init_sock(struct sock *sk)
374 struct inet_connection_sock *icsk = inet_csk(sk);
375 struct tcp_sock *tp = tcp_sk(sk);
377 skb_queue_head_init(&tp->out_of_order_queue);
378 tcp_init_xmit_timers(sk);
379 tcp_prequeue_init(tp);
380 INIT_LIST_HEAD(&tp->tsq_node);
382 icsk->icsk_rto = TCP_TIMEOUT_INIT;
383 tp->mdev = TCP_TIMEOUT_INIT;
385 /* So many TCP implementations out there (incorrectly) count the
386 * initial SYN frame in their delayed-ACK and congestion control
387 * algorithms that we must have the following bandaid to talk
388 * efficiently to them. -DaveM
390 tp->snd_cwnd = TCP_INIT_CWND;
392 /* See draft-stevens-tcpca-spec-01 for discussion of the
393 * initialization of these values.
395 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
396 tp->snd_cwnd_clamp = ~0;
397 tp->mss_cache = TCP_MSS_DEFAULT;
399 tp->reordering = sysctl_tcp_reordering;
400 tcp_enable_early_retrans(tp);
401 icsk->icsk_ca_ops = &tcp_init_congestion_ops;
403 sk->sk_state = TCP_CLOSE;
405 sk->sk_write_space = sk_stream_write_space;
406 sock_set_flag(sk, SOCK_USE_WRITE_QUEUE);
408 icsk->icsk_sync_mss = tcp_sync_mss;
410 /* TCP Cookie Transactions */
411 if (sysctl_tcp_cookie_size > 0) {
412 /* Default, cookies without s_data_payload. */
414 kzalloc(sizeof(*tp->cookie_values),
416 if (tp->cookie_values != NULL)
417 kref_init(&tp->cookie_values->kref);
419 /* Presumed zeroed, in order of appearance:
420 * cookie_in_always, cookie_out_never,
421 * s_data_constant, s_data_in, s_data_out
423 sk->sk_sndbuf = sysctl_tcp_wmem[1];
424 sk->sk_rcvbuf = sysctl_tcp_rmem[1];
427 sock_update_memcg(sk);
428 sk_sockets_allocated_inc(sk);
431 EXPORT_SYMBOL(tcp_init_sock);
434 * Wait for a TCP event.
436 * Note that we don't need to lock the socket, as the upper poll layers
437 * take care of normal races (between the test and the event) and we don't
438 * go look at any of the socket buffers directly.
440 unsigned int tcp_poll(struct file *file, struct socket *sock, poll_table *wait)
443 struct sock *sk = sock->sk;
444 const struct tcp_sock *tp = tcp_sk(sk);
446 sock_poll_wait(file, sk_sleep(sk), wait);
447 if (sk->sk_state == TCP_LISTEN)
448 return inet_csk_listen_poll(sk);
450 /* Socket is not locked. We are protected from async events
451 * by poll logic and correct handling of state changes
452 * made by other threads is impossible in any case.
458 * POLLHUP is certainly not done right. But poll() doesn't
459 * have a notion of HUP in just one direction, and for a
460 * socket the read side is more interesting.
462 * Some poll() documentation says that POLLHUP is incompatible
463 * with the POLLOUT/POLLWR flags, so somebody should check this
464 * all. But careful, it tends to be safer to return too many
465 * bits than too few, and you can easily break real applications
466 * if you don't tell them that something has hung up!
470 * Check number 1. POLLHUP is _UNMASKABLE_ event (see UNIX98 and
471 * our fs/select.c). It means that after we received EOF,
472 * poll always returns immediately, making impossible poll() on write()
473 * in state CLOSE_WAIT. One solution is evident --- to set POLLHUP
474 * if and only if shutdown has been made in both directions.
475 * Actually, it is interesting to look how Solaris and DUX
476 * solve this dilemma. I would prefer, if POLLHUP were maskable,
477 * then we could set it on SND_SHUTDOWN. BTW examples given
478 * in Stevens' books assume exactly this behaviour, it explains
479 * why POLLHUP is incompatible with POLLOUT. --ANK
481 * NOTE. Check for TCP_CLOSE is added. The goal is to prevent
482 * blocking on fresh not-connected or disconnected socket. --ANK
484 if (sk->sk_shutdown == SHUTDOWN_MASK || sk->sk_state == TCP_CLOSE)
486 if (sk->sk_shutdown & RCV_SHUTDOWN)
487 mask |= POLLIN | POLLRDNORM | POLLRDHUP;
489 /* Connected or passive Fast Open socket? */
490 if (sk->sk_state != TCP_SYN_SENT &&
491 (sk->sk_state != TCP_SYN_RECV || tp->fastopen_rsk != NULL)) {
492 int target = sock_rcvlowat(sk, 0, INT_MAX);
494 if (tp->urg_seq == tp->copied_seq &&
495 !sock_flag(sk, SOCK_URGINLINE) &&
499 /* Potential race condition. If read of tp below will
500 * escape above sk->sk_state, we can be illegally awaken
501 * in SYN_* states. */
502 if (tp->rcv_nxt - tp->copied_seq >= target)
503 mask |= POLLIN | POLLRDNORM;
505 if (!(sk->sk_shutdown & SEND_SHUTDOWN)) {
506 if (sk_stream_wspace(sk) >= sk_stream_min_wspace(sk)) {
507 mask |= POLLOUT | POLLWRNORM;
508 } else { /* send SIGIO later */
509 set_bit(SOCK_ASYNC_NOSPACE,
510 &sk->sk_socket->flags);
511 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
513 /* Race breaker. If space is freed after
514 * wspace test but before the flags are set,
515 * IO signal will be lost.
517 if (sk_stream_wspace(sk) >= sk_stream_min_wspace(sk))
518 mask |= POLLOUT | POLLWRNORM;
521 mask |= POLLOUT | POLLWRNORM;
523 if (tp->urg_data & TCP_URG_VALID)
526 /* This barrier is coupled with smp_wmb() in tcp_reset() */
533 EXPORT_SYMBOL(tcp_poll);
535 int tcp_ioctl(struct sock *sk, int cmd, unsigned long arg)
537 struct tcp_sock *tp = tcp_sk(sk);
543 if (sk->sk_state == TCP_LISTEN)
546 slow = lock_sock_fast(sk);
547 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
549 else if (sock_flag(sk, SOCK_URGINLINE) ||
551 before(tp->urg_seq, tp->copied_seq) ||
552 !before(tp->urg_seq, tp->rcv_nxt)) {
554 answ = tp->rcv_nxt - tp->copied_seq;
556 /* Subtract 1, if FIN was received */
557 if (answ && sock_flag(sk, SOCK_DONE))
560 answ = tp->urg_seq - tp->copied_seq;
561 unlock_sock_fast(sk, slow);
564 answ = tp->urg_data && tp->urg_seq == tp->copied_seq;
567 if (sk->sk_state == TCP_LISTEN)
570 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
573 answ = tp->write_seq - tp->snd_una;
576 if (sk->sk_state == TCP_LISTEN)
579 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
582 answ = tp->write_seq - tp->snd_nxt;
588 return put_user(answ, (int __user *)arg);
590 EXPORT_SYMBOL(tcp_ioctl);
592 static inline void tcp_mark_push(struct tcp_sock *tp, struct sk_buff *skb)
594 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
595 tp->pushed_seq = tp->write_seq;
598 static inline bool forced_push(const struct tcp_sock *tp)
600 return after(tp->write_seq, tp->pushed_seq + (tp->max_window >> 1));
603 static inline void skb_entail(struct sock *sk, struct sk_buff *skb)
605 struct tcp_sock *tp = tcp_sk(sk);
606 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
609 tcb->seq = tcb->end_seq = tp->write_seq;
610 tcb->tcp_flags = TCPHDR_ACK;
612 skb_header_release(skb);
613 tcp_add_write_queue_tail(sk, skb);
614 sk->sk_wmem_queued += skb->truesize;
615 sk_mem_charge(sk, skb->truesize);
616 if (tp->nonagle & TCP_NAGLE_PUSH)
617 tp->nonagle &= ~TCP_NAGLE_PUSH;
620 static inline void tcp_mark_urg(struct tcp_sock *tp, int flags)
623 tp->snd_up = tp->write_seq;
626 static inline void tcp_push(struct sock *sk, int flags, int mss_now,
629 if (tcp_send_head(sk)) {
630 struct tcp_sock *tp = tcp_sk(sk);
632 if (!(flags & MSG_MORE) || forced_push(tp))
633 tcp_mark_push(tp, tcp_write_queue_tail(sk));
635 tcp_mark_urg(tp, flags);
636 __tcp_push_pending_frames(sk, mss_now,
637 (flags & MSG_MORE) ? TCP_NAGLE_CORK : nonagle);
641 static int tcp_splice_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb,
642 unsigned int offset, size_t len)
644 struct tcp_splice_state *tss = rd_desc->arg.data;
647 ret = skb_splice_bits(skb, offset, tss->pipe, min(rd_desc->count, len),
650 rd_desc->count -= ret;
654 static int __tcp_splice_read(struct sock *sk, struct tcp_splice_state *tss)
656 /* Store TCP splice context information in read_descriptor_t. */
657 read_descriptor_t rd_desc = {
662 return tcp_read_sock(sk, &rd_desc, tcp_splice_data_recv);
666 * tcp_splice_read - splice data from TCP socket to a pipe
667 * @sock: socket to splice from
668 * @ppos: position (not valid)
669 * @pipe: pipe to splice to
670 * @len: number of bytes to splice
671 * @flags: splice modifier flags
674 * Will read pages from given socket and fill them into a pipe.
677 ssize_t tcp_splice_read(struct socket *sock, loff_t *ppos,
678 struct pipe_inode_info *pipe, size_t len,
681 struct sock *sk = sock->sk;
682 struct tcp_splice_state tss = {
691 sock_rps_record_flow(sk);
693 * We can't seek on a socket input
702 timeo = sock_rcvtimeo(sk, sock->file->f_flags & O_NONBLOCK);
704 ret = __tcp_splice_read(sk, &tss);
710 if (sock_flag(sk, SOCK_DONE))
713 ret = sock_error(sk);
716 if (sk->sk_shutdown & RCV_SHUTDOWN)
718 if (sk->sk_state == TCP_CLOSE) {
720 * This occurs when user tries to read
721 * from never connected socket.
723 if (!sock_flag(sk, SOCK_DONE))
731 sk_wait_data(sk, &timeo);
732 if (signal_pending(current)) {
733 ret = sock_intr_errno(timeo);
746 if (sk->sk_err || sk->sk_state == TCP_CLOSE ||
747 (sk->sk_shutdown & RCV_SHUTDOWN) ||
748 signal_pending(current))
759 EXPORT_SYMBOL(tcp_splice_read);
761 struct sk_buff *sk_stream_alloc_skb(struct sock *sk, int size, gfp_t gfp)
765 /* The TCP header must be at least 32-bit aligned. */
766 size = ALIGN(size, 4);
768 skb = alloc_skb_fclone(size + sk->sk_prot->max_header, gfp);
770 if (sk_wmem_schedule(sk, skb->truesize)) {
771 skb_reserve(skb, sk->sk_prot->max_header);
773 * Make sure that we have exactly size bytes
774 * available to the caller, no more, no less.
776 skb->avail_size = size;
781 sk->sk_prot->enter_memory_pressure(sk);
782 sk_stream_moderate_sndbuf(sk);
787 static unsigned int tcp_xmit_size_goal(struct sock *sk, u32 mss_now,
790 struct tcp_sock *tp = tcp_sk(sk);
791 u32 xmit_size_goal, old_size_goal;
793 xmit_size_goal = mss_now;
795 if (large_allowed && sk_can_gso(sk)) {
796 xmit_size_goal = ((sk->sk_gso_max_size - 1) -
797 inet_csk(sk)->icsk_af_ops->net_header_len -
798 inet_csk(sk)->icsk_ext_hdr_len -
801 /* TSQ : try to have two TSO segments in flight */
802 xmit_size_goal = min_t(u32, xmit_size_goal,
803 sysctl_tcp_limit_output_bytes >> 1);
805 xmit_size_goal = tcp_bound_to_half_wnd(tp, xmit_size_goal);
807 /* We try hard to avoid divides here */
808 old_size_goal = tp->xmit_size_goal_segs * mss_now;
810 if (likely(old_size_goal <= xmit_size_goal &&
811 old_size_goal + mss_now > xmit_size_goal)) {
812 xmit_size_goal = old_size_goal;
814 tp->xmit_size_goal_segs =
815 min_t(u16, xmit_size_goal / mss_now,
816 sk->sk_gso_max_segs);
817 xmit_size_goal = tp->xmit_size_goal_segs * mss_now;
821 return max(xmit_size_goal, mss_now);
824 static int tcp_send_mss(struct sock *sk, int *size_goal, int flags)
828 mss_now = tcp_current_mss(sk);
829 *size_goal = tcp_xmit_size_goal(sk, mss_now, !(flags & MSG_OOB));
834 static ssize_t do_tcp_sendpages(struct sock *sk, struct page *page, int offset,
835 size_t size, int flags)
837 struct tcp_sock *tp = tcp_sk(sk);
838 int mss_now, size_goal;
841 long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
843 /* Wait for a connection to finish. One exception is TCP Fast Open
844 * (passive side) where data is allowed to be sent before a connection
845 * is fully established.
847 if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) &&
848 !tcp_passive_fastopen(sk)) {
849 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
853 clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
855 mss_now = tcp_send_mss(sk, &size_goal, flags);
859 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
863 struct sk_buff *skb = tcp_write_queue_tail(sk);
867 if (!tcp_send_head(sk) || (copy = size_goal - skb->len) <= 0) {
869 if (!sk_stream_memory_free(sk))
870 goto wait_for_sndbuf;
872 skb = sk_stream_alloc_skb(sk, 0, sk->sk_allocation);
874 goto wait_for_memory;
883 i = skb_shinfo(skb)->nr_frags;
884 can_coalesce = skb_can_coalesce(skb, i, page, offset);
885 if (!can_coalesce && i >= MAX_SKB_FRAGS) {
886 tcp_mark_push(tp, skb);
889 if (!sk_wmem_schedule(sk, copy))
890 goto wait_for_memory;
893 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
896 skb_fill_page_desc(skb, i, page, offset, copy);
900 skb->data_len += copy;
901 skb->truesize += copy;
902 sk->sk_wmem_queued += copy;
903 sk_mem_charge(sk, copy);
904 skb->ip_summed = CHECKSUM_PARTIAL;
905 tp->write_seq += copy;
906 TCP_SKB_CB(skb)->end_seq += copy;
907 skb_shinfo(skb)->gso_segs = 0;
910 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
917 if (skb->len < size_goal || (flags & MSG_OOB))
920 if (forced_push(tp)) {
921 tcp_mark_push(tp, skb);
922 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
923 } else if (skb == tcp_send_head(sk))
924 tcp_push_one(sk, mss_now);
928 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
930 tcp_push(sk, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH);
932 if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
935 mss_now = tcp_send_mss(sk, &size_goal, flags);
939 if (copied && !(flags & MSG_SENDPAGE_NOTLAST))
940 tcp_push(sk, flags, mss_now, tp->nonagle);
947 return sk_stream_error(sk, flags, err);
950 int tcp_sendpage(struct sock *sk, struct page *page, int offset,
951 size_t size, int flags)
955 if (!(sk->sk_route_caps & NETIF_F_SG) ||
956 !(sk->sk_route_caps & NETIF_F_ALL_CSUM))
957 return sock_no_sendpage(sk->sk_socket, page, offset, size,
961 res = do_tcp_sendpages(sk, page, offset, size, flags);
965 EXPORT_SYMBOL(tcp_sendpage);
967 static inline int select_size(const struct sock *sk, bool sg)
969 const struct tcp_sock *tp = tcp_sk(sk);
970 int tmp = tp->mss_cache;
973 if (sk_can_gso(sk)) {
974 /* Small frames wont use a full page:
975 * Payload will immediately follow tcp header.
977 tmp = SKB_WITH_OVERHEAD(2048 - MAX_TCP_HEADER);
979 int pgbreak = SKB_MAX_HEAD(MAX_TCP_HEADER);
981 if (tmp >= pgbreak &&
982 tmp <= pgbreak + (MAX_SKB_FRAGS - 1) * PAGE_SIZE)
990 void tcp_free_fastopen_req(struct tcp_sock *tp)
992 if (tp->fastopen_req != NULL) {
993 kfree(tp->fastopen_req);
994 tp->fastopen_req = NULL;
998 static int tcp_sendmsg_fastopen(struct sock *sk, struct msghdr *msg, int *size)
1000 struct tcp_sock *tp = tcp_sk(sk);
1003 if (!(sysctl_tcp_fastopen & TFO_CLIENT_ENABLE))
1005 if (tp->fastopen_req != NULL)
1006 return -EALREADY; /* Another Fast Open is in progress */
1008 tp->fastopen_req = kzalloc(sizeof(struct tcp_fastopen_request),
1010 if (unlikely(tp->fastopen_req == NULL))
1012 tp->fastopen_req->data = msg;
1014 flags = (msg->msg_flags & MSG_DONTWAIT) ? O_NONBLOCK : 0;
1015 err = __inet_stream_connect(sk->sk_socket, msg->msg_name,
1016 msg->msg_namelen, flags);
1017 *size = tp->fastopen_req->copied;
1018 tcp_free_fastopen_req(tp);
1022 int tcp_sendmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg,
1026 struct tcp_sock *tp = tcp_sk(sk);
1027 struct sk_buff *skb;
1028 int iovlen, flags, err, copied = 0;
1029 int mss_now = 0, size_goal, copied_syn = 0, offset = 0;
1035 flags = msg->msg_flags;
1036 if (flags & MSG_FASTOPEN) {
1037 err = tcp_sendmsg_fastopen(sk, msg, &copied_syn);
1038 if (err == -EINPROGRESS && copied_syn > 0)
1042 offset = copied_syn;
1045 timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
1047 /* Wait for a connection to finish. One exception is TCP Fast Open
1048 * (passive side) where data is allowed to be sent before a connection
1049 * is fully established.
1051 if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) &&
1052 !tcp_passive_fastopen(sk)) {
1053 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
1057 if (unlikely(tp->repair)) {
1058 if (tp->repair_queue == TCP_RECV_QUEUE) {
1059 copied = tcp_send_rcvq(sk, msg, size);
1064 if (tp->repair_queue == TCP_NO_QUEUE)
1067 /* 'common' sending to sendq */
1070 /* This should be in poll */
1071 clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
1073 mss_now = tcp_send_mss(sk, &size_goal, flags);
1075 /* Ok commence sending. */
1076 iovlen = msg->msg_iovlen;
1081 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
1084 sg = !!(sk->sk_route_caps & NETIF_F_SG);
1086 while (--iovlen >= 0) {
1087 size_t seglen = iov->iov_len;
1088 unsigned char __user *from = iov->iov_base;
1091 if (unlikely(offset > 0)) { /* Skip bytes copied in SYN */
1092 if (offset >= seglen) {
1101 while (seglen > 0) {
1103 int max = size_goal;
1105 skb = tcp_write_queue_tail(sk);
1106 if (tcp_send_head(sk)) {
1107 if (skb->ip_summed == CHECKSUM_NONE)
1109 copy = max - skb->len;
1114 /* Allocate new segment. If the interface is SG,
1115 * allocate skb fitting to single page.
1117 if (!sk_stream_memory_free(sk))
1118 goto wait_for_sndbuf;
1120 skb = sk_stream_alloc_skb(sk,
1121 select_size(sk, sg),
1124 goto wait_for_memory;
1127 * Check whether we can use HW checksum.
1129 if (sk->sk_route_caps & NETIF_F_ALL_CSUM)
1130 skb->ip_summed = CHECKSUM_PARTIAL;
1132 skb_entail(sk, skb);
1137 /* Try to append data to the end of skb. */
1141 /* Where to copy to? */
1142 if (skb_availroom(skb) > 0) {
1143 /* We have some space in skb head. Superb! */
1144 copy = min_t(int, copy, skb_availroom(skb));
1145 err = skb_add_data_nocache(sk, skb, from, copy);
1150 int i = skb_shinfo(skb)->nr_frags;
1151 struct page_frag *pfrag = sk_page_frag(sk);
1153 if (!sk_page_frag_refill(sk, pfrag))
1154 goto wait_for_memory;
1156 if (!skb_can_coalesce(skb, i, pfrag->page,
1158 if (i == MAX_SKB_FRAGS || !sg) {
1159 tcp_mark_push(tp, skb);
1165 copy = min_t(int, copy, pfrag->size - pfrag->offset);
1167 if (!sk_wmem_schedule(sk, copy))
1168 goto wait_for_memory;
1170 err = skb_copy_to_page_nocache(sk, from, skb,
1177 /* Update the skb. */
1179 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1181 skb_fill_page_desc(skb, i, pfrag->page,
1182 pfrag->offset, copy);
1183 get_page(pfrag->page);
1185 pfrag->offset += copy;
1189 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
1191 tp->write_seq += copy;
1192 TCP_SKB_CB(skb)->end_seq += copy;
1193 skb_shinfo(skb)->gso_segs = 0;
1197 if ((seglen -= copy) == 0 && iovlen == 0)
1200 if (skb->len < max || (flags & MSG_OOB) || unlikely(tp->repair))
1203 if (forced_push(tp)) {
1204 tcp_mark_push(tp, skb);
1205 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
1206 } else if (skb == tcp_send_head(sk))
1207 tcp_push_one(sk, mss_now);
1211 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1214 tcp_push(sk, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH);
1216 if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
1219 mss_now = tcp_send_mss(sk, &size_goal, flags);
1225 tcp_push(sk, flags, mss_now, tp->nonagle);
1227 return copied + copied_syn;
1231 tcp_unlink_write_queue(skb, sk);
1232 /* It is the one place in all of TCP, except connection
1233 * reset, where we can be unlinking the send_head.
1235 tcp_check_send_head(sk, skb);
1236 sk_wmem_free_skb(sk, skb);
1240 if (copied + copied_syn)
1243 err = sk_stream_error(sk, flags, err);
1247 EXPORT_SYMBOL(tcp_sendmsg);
1250 * Handle reading urgent data. BSD has very simple semantics for
1251 * this, no blocking and very strange errors 8)
1254 static int tcp_recv_urg(struct sock *sk, struct msghdr *msg, int len, int flags)
1256 struct tcp_sock *tp = tcp_sk(sk);
1258 /* No URG data to read. */
1259 if (sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data ||
1260 tp->urg_data == TCP_URG_READ)
1261 return -EINVAL; /* Yes this is right ! */
1263 if (sk->sk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DONE))
1266 if (tp->urg_data & TCP_URG_VALID) {
1268 char c = tp->urg_data;
1270 if (!(flags & MSG_PEEK))
1271 tp->urg_data = TCP_URG_READ;
1273 /* Read urgent data. */
1274 msg->msg_flags |= MSG_OOB;
1277 if (!(flags & MSG_TRUNC))
1278 err = memcpy_toiovec(msg->msg_iov, &c, 1);
1281 msg->msg_flags |= MSG_TRUNC;
1283 return err ? -EFAULT : len;
1286 if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
1289 /* Fixed the recv(..., MSG_OOB) behaviour. BSD docs and
1290 * the available implementations agree in this case:
1291 * this call should never block, independent of the
1292 * blocking state of the socket.
1293 * Mike <pall@rz.uni-karlsruhe.de>
1298 static int tcp_peek_sndq(struct sock *sk, struct msghdr *msg, int len)
1300 struct sk_buff *skb;
1301 int copied = 0, err = 0;
1303 /* XXX -- need to support SO_PEEK_OFF */
1305 skb_queue_walk(&sk->sk_write_queue, skb) {
1306 err = skb_copy_datagram_iovec(skb, 0, msg->msg_iov, skb->len);
1313 return err ?: copied;
1316 /* Clean up the receive buffer for full frames taken by the user,
1317 * then send an ACK if necessary. COPIED is the number of bytes
1318 * tcp_recvmsg has given to the user so far, it speeds up the
1319 * calculation of whether or not we must ACK for the sake of
1322 void tcp_cleanup_rbuf(struct sock *sk, int copied)
1324 struct tcp_sock *tp = tcp_sk(sk);
1325 bool time_to_ack = false;
1327 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
1329 WARN(skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq),
1330 "cleanup rbuf bug: copied %X seq %X rcvnxt %X\n",
1331 tp->copied_seq, TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt);
1333 if (inet_csk_ack_scheduled(sk)) {
1334 const struct inet_connection_sock *icsk = inet_csk(sk);
1335 /* Delayed ACKs frequently hit locked sockets during bulk
1337 if (icsk->icsk_ack.blocked ||
1338 /* Once-per-two-segments ACK was not sent by tcp_input.c */
1339 tp->rcv_nxt - tp->rcv_wup > icsk->icsk_ack.rcv_mss ||
1341 * If this read emptied read buffer, we send ACK, if
1342 * connection is not bidirectional, user drained
1343 * receive buffer and there was a small segment
1347 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED2) ||
1348 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED) &&
1349 !icsk->icsk_ack.pingpong)) &&
1350 !atomic_read(&sk->sk_rmem_alloc)))
1354 /* We send an ACK if we can now advertise a non-zero window
1355 * which has been raised "significantly".
1357 * Even if window raised up to infinity, do not send window open ACK
1358 * in states, where we will not receive more. It is useless.
1360 if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) {
1361 __u32 rcv_window_now = tcp_receive_window(tp);
1363 /* Optimize, __tcp_select_window() is not cheap. */
1364 if (2*rcv_window_now <= tp->window_clamp) {
1365 __u32 new_window = __tcp_select_window(sk);
1367 /* Send ACK now, if this read freed lots of space
1368 * in our buffer. Certainly, new_window is new window.
1369 * We can advertise it now, if it is not less than current one.
1370 * "Lots" means "at least twice" here.
1372 if (new_window && new_window >= 2 * rcv_window_now)
1380 static void tcp_prequeue_process(struct sock *sk)
1382 struct sk_buff *skb;
1383 struct tcp_sock *tp = tcp_sk(sk);
1385 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPPREQUEUED);
1387 /* RX process wants to run with disabled BHs, though it is not
1390 while ((skb = __skb_dequeue(&tp->ucopy.prequeue)) != NULL)
1391 sk_backlog_rcv(sk, skb);
1394 /* Clear memory counter. */
1395 tp->ucopy.memory = 0;
1398 #ifdef CONFIG_NET_DMA
1399 static void tcp_service_net_dma(struct sock *sk, bool wait)
1401 dma_cookie_t done, used;
1402 dma_cookie_t last_issued;
1403 struct tcp_sock *tp = tcp_sk(sk);
1405 if (!tp->ucopy.dma_chan)
1408 last_issued = tp->ucopy.dma_cookie;
1409 dma_async_memcpy_issue_pending(tp->ucopy.dma_chan);
1412 if (dma_async_memcpy_complete(tp->ucopy.dma_chan,
1414 &used) == DMA_SUCCESS) {
1415 /* Safe to free early-copied skbs now */
1416 __skb_queue_purge(&sk->sk_async_wait_queue);
1419 struct sk_buff *skb;
1420 while ((skb = skb_peek(&sk->sk_async_wait_queue)) &&
1421 (dma_async_is_complete(skb->dma_cookie, done,
1422 used) == DMA_SUCCESS)) {
1423 __skb_dequeue(&sk->sk_async_wait_queue);
1431 static inline struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off)
1433 struct sk_buff *skb;
1436 skb_queue_walk(&sk->sk_receive_queue, skb) {
1437 offset = seq - TCP_SKB_CB(skb)->seq;
1438 if (tcp_hdr(skb)->syn)
1440 if (offset < skb->len || tcp_hdr(skb)->fin) {
1449 * This routine provides an alternative to tcp_recvmsg() for routines
1450 * that would like to handle copying from skbuffs directly in 'sendfile'
1453 * - It is assumed that the socket was locked by the caller.
1454 * - The routine does not block.
1455 * - At present, there is no support for reading OOB data
1456 * or for 'peeking' the socket using this routine
1457 * (although both would be easy to implement).
1459 int tcp_read_sock(struct sock *sk, read_descriptor_t *desc,
1460 sk_read_actor_t recv_actor)
1462 struct sk_buff *skb;
1463 struct tcp_sock *tp = tcp_sk(sk);
1464 u32 seq = tp->copied_seq;
1468 if (sk->sk_state == TCP_LISTEN)
1470 while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1471 if (offset < skb->len) {
1475 len = skb->len - offset;
1476 /* Stop reading if we hit a patch of urgent data */
1478 u32 urg_offset = tp->urg_seq - seq;
1479 if (urg_offset < len)
1484 used = recv_actor(desc, skb, offset, len);
1489 } else if (used <= len) {
1494 /* If recv_actor drops the lock (e.g. TCP splice
1495 * receive) the skb pointer might be invalid when
1496 * getting here: tcp_collapse might have deleted it
1497 * while aggregating skbs from the socket queue.
1499 skb = tcp_recv_skb(sk, seq - 1, &offset);
1502 /* TCP coalescing might have appended data to the skb.
1503 * Try to splice more frags
1505 if (offset + 1 != skb->len)
1508 if (tcp_hdr(skb)->fin) {
1509 sk_eat_skb(sk, skb, false);
1513 sk_eat_skb(sk, skb, false);
1516 tp->copied_seq = seq;
1518 tp->copied_seq = seq;
1520 tcp_rcv_space_adjust(sk);
1522 /* Clean up data we have read: This will do ACK frames. */
1524 tcp_cleanup_rbuf(sk, copied);
1527 EXPORT_SYMBOL(tcp_read_sock);
1530 * This routine copies from a sock struct into the user buffer.
1532 * Technical note: in 2.3 we work on _locked_ socket, so that
1533 * tricks with *seq access order and skb->users are not required.
1534 * Probably, code can be easily improved even more.
1537 int tcp_recvmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg,
1538 size_t len, int nonblock, int flags, int *addr_len)
1540 struct tcp_sock *tp = tcp_sk(sk);
1546 int target; /* Read at least this many bytes */
1548 struct task_struct *user_recv = NULL;
1549 bool copied_early = false;
1550 struct sk_buff *skb;
1556 if (sk->sk_state == TCP_LISTEN)
1559 timeo = sock_rcvtimeo(sk, nonblock);
1561 /* Urgent data needs to be handled specially. */
1562 if (flags & MSG_OOB)
1565 if (unlikely(tp->repair)) {
1567 if (!(flags & MSG_PEEK))
1570 if (tp->repair_queue == TCP_SEND_QUEUE)
1574 if (tp->repair_queue == TCP_NO_QUEUE)
1577 /* 'common' recv queue MSG_PEEK-ing */
1580 seq = &tp->copied_seq;
1581 if (flags & MSG_PEEK) {
1582 peek_seq = tp->copied_seq;
1586 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
1588 #ifdef CONFIG_NET_DMA
1589 tp->ucopy.dma_chan = NULL;
1591 skb = skb_peek_tail(&sk->sk_receive_queue);
1596 available = TCP_SKB_CB(skb)->seq + skb->len - (*seq);
1597 if ((available < target) &&
1598 (len > sysctl_tcp_dma_copybreak) && !(flags & MSG_PEEK) &&
1599 !sysctl_tcp_low_latency &&
1600 net_dma_find_channel()) {
1601 preempt_enable_no_resched();
1602 tp->ucopy.pinned_list =
1603 dma_pin_iovec_pages(msg->msg_iov, len);
1605 preempt_enable_no_resched();
1613 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
1614 if (tp->urg_data && tp->urg_seq == *seq) {
1617 if (signal_pending(current)) {
1618 copied = timeo ? sock_intr_errno(timeo) : -EAGAIN;
1623 /* Next get a buffer. */
1625 skb_queue_walk(&sk->sk_receive_queue, skb) {
1626 /* Now that we have two receive queues this
1629 if (WARN(before(*seq, TCP_SKB_CB(skb)->seq),
1630 "recvmsg bug: copied %X seq %X rcvnxt %X fl %X\n",
1631 *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt,
1635 offset = *seq - TCP_SKB_CB(skb)->seq;
1636 if (tcp_hdr(skb)->syn)
1638 if (offset < skb->len)
1640 if (tcp_hdr(skb)->fin)
1642 WARN(!(flags & MSG_PEEK),
1643 "recvmsg bug 2: copied %X seq %X rcvnxt %X fl %X\n",
1644 *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt, flags);
1647 /* Well, if we have backlog, try to process it now yet. */
1649 if (copied >= target && !sk->sk_backlog.tail)
1654 sk->sk_state == TCP_CLOSE ||
1655 (sk->sk_shutdown & RCV_SHUTDOWN) ||
1657 signal_pending(current))
1660 if (sock_flag(sk, SOCK_DONE))
1664 copied = sock_error(sk);
1668 if (sk->sk_shutdown & RCV_SHUTDOWN)
1671 if (sk->sk_state == TCP_CLOSE) {
1672 if (!sock_flag(sk, SOCK_DONE)) {
1673 /* This occurs when user tries to read
1674 * from never connected socket.
1687 if (signal_pending(current)) {
1688 copied = sock_intr_errno(timeo);
1693 tcp_cleanup_rbuf(sk, copied);
1695 if (!sysctl_tcp_low_latency && tp->ucopy.task == user_recv) {
1696 /* Install new reader */
1697 if (!user_recv && !(flags & (MSG_TRUNC | MSG_PEEK))) {
1698 user_recv = current;
1699 tp->ucopy.task = user_recv;
1700 tp->ucopy.iov = msg->msg_iov;
1703 tp->ucopy.len = len;
1705 WARN_ON(tp->copied_seq != tp->rcv_nxt &&
1706 !(flags & (MSG_PEEK | MSG_TRUNC)));
1708 /* Ugly... If prequeue is not empty, we have to
1709 * process it before releasing socket, otherwise
1710 * order will be broken at second iteration.
1711 * More elegant solution is required!!!
1713 * Look: we have the following (pseudo)queues:
1715 * 1. packets in flight
1720 * Each queue can be processed only if the next ones
1721 * are empty. At this point we have empty receive_queue.
1722 * But prequeue _can_ be not empty after 2nd iteration,
1723 * when we jumped to start of loop because backlog
1724 * processing added something to receive_queue.
1725 * We cannot release_sock(), because backlog contains
1726 * packets arrived _after_ prequeued ones.
1728 * Shortly, algorithm is clear --- to process all
1729 * the queues in order. We could make it more directly,
1730 * requeueing packets from backlog to prequeue, if
1731 * is not empty. It is more elegant, but eats cycles,
1734 if (!skb_queue_empty(&tp->ucopy.prequeue))
1737 /* __ Set realtime policy in scheduler __ */
1740 #ifdef CONFIG_NET_DMA
1741 if (tp->ucopy.dma_chan) {
1742 if (tp->rcv_wnd == 0 &&
1743 !skb_queue_empty(&sk->sk_async_wait_queue)) {
1744 tcp_service_net_dma(sk, true);
1745 tcp_cleanup_rbuf(sk, copied);
1747 dma_async_memcpy_issue_pending(tp->ucopy.dma_chan);
1750 if (copied >= target) {
1751 /* Do not sleep, just process backlog. */
1755 sk_wait_data(sk, &timeo);
1757 #ifdef CONFIG_NET_DMA
1758 tcp_service_net_dma(sk, false); /* Don't block */
1759 tp->ucopy.wakeup = 0;
1765 /* __ Restore normal policy in scheduler __ */
1767 if ((chunk = len - tp->ucopy.len) != 0) {
1768 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMBACKLOG, chunk);
1773 if (tp->rcv_nxt == tp->copied_seq &&
1774 !skb_queue_empty(&tp->ucopy.prequeue)) {
1776 tcp_prequeue_process(sk);
1778 if ((chunk = len - tp->ucopy.len) != 0) {
1779 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1785 if ((flags & MSG_PEEK) &&
1786 (peek_seq - copied - urg_hole != tp->copied_seq)) {
1787 net_dbg_ratelimited("TCP(%s:%d): Application bug, race in MSG_PEEK\n",
1789 task_pid_nr(current));
1790 peek_seq = tp->copied_seq;
1795 /* Ok so how much can we use? */
1796 used = skb->len - offset;
1800 /* Do we have urgent data here? */
1802 u32 urg_offset = tp->urg_seq - *seq;
1803 if (urg_offset < used) {
1805 if (!sock_flag(sk, SOCK_URGINLINE)) {
1818 if (!(flags & MSG_TRUNC)) {
1819 #ifdef CONFIG_NET_DMA
1820 if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list)
1821 tp->ucopy.dma_chan = net_dma_find_channel();
1823 if (tp->ucopy.dma_chan) {
1824 tp->ucopy.dma_cookie = dma_skb_copy_datagram_iovec(
1825 tp->ucopy.dma_chan, skb, offset,
1827 tp->ucopy.pinned_list);
1829 if (tp->ucopy.dma_cookie < 0) {
1831 pr_alert("%s: dma_cookie < 0\n",
1834 /* Exception. Bailout! */
1840 dma_async_memcpy_issue_pending(tp->ucopy.dma_chan);
1842 if ((offset + used) == skb->len)
1843 copied_early = true;
1848 err = skb_copy_datagram_iovec(skb, offset,
1849 msg->msg_iov, used);
1851 /* Exception. Bailout! */
1863 tcp_rcv_space_adjust(sk);
1866 if (tp->urg_data && after(tp->copied_seq, tp->urg_seq)) {
1868 tcp_fast_path_check(sk);
1870 if (used + offset < skb->len)
1873 if (tcp_hdr(skb)->fin)
1875 if (!(flags & MSG_PEEK)) {
1876 sk_eat_skb(sk, skb, copied_early);
1877 copied_early = false;
1882 /* Process the FIN. */
1884 if (!(flags & MSG_PEEK)) {
1885 sk_eat_skb(sk, skb, copied_early);
1886 copied_early = false;
1892 if (!skb_queue_empty(&tp->ucopy.prequeue)) {
1895 tp->ucopy.len = copied > 0 ? len : 0;
1897 tcp_prequeue_process(sk);
1899 if (copied > 0 && (chunk = len - tp->ucopy.len) != 0) {
1900 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1906 tp->ucopy.task = NULL;
1910 #ifdef CONFIG_NET_DMA
1911 tcp_service_net_dma(sk, true); /* Wait for queue to drain */
1912 tp->ucopy.dma_chan = NULL;
1914 if (tp->ucopy.pinned_list) {
1915 dma_unpin_iovec_pages(tp->ucopy.pinned_list);
1916 tp->ucopy.pinned_list = NULL;
1920 /* According to UNIX98, msg_name/msg_namelen are ignored
1921 * on connected socket. I was just happy when found this 8) --ANK
1924 /* Clean up data we have read: This will do ACK frames. */
1925 tcp_cleanup_rbuf(sk, copied);
1935 err = tcp_recv_urg(sk, msg, len, flags);
1939 err = tcp_peek_sndq(sk, msg, len);
1942 EXPORT_SYMBOL(tcp_recvmsg);
1944 void tcp_set_state(struct sock *sk, int state)
1946 int oldstate = sk->sk_state;
1949 case TCP_ESTABLISHED:
1950 if (oldstate != TCP_ESTABLISHED)
1951 TCP_INC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
1955 if (oldstate == TCP_CLOSE_WAIT || oldstate == TCP_ESTABLISHED)
1956 TCP_INC_STATS(sock_net(sk), TCP_MIB_ESTABRESETS);
1958 sk->sk_prot->unhash(sk);
1959 if (inet_csk(sk)->icsk_bind_hash &&
1960 !(sk->sk_userlocks & SOCK_BINDPORT_LOCK))
1964 if (oldstate == TCP_ESTABLISHED)
1965 TCP_DEC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
1968 /* Change state AFTER socket is unhashed to avoid closed
1969 * socket sitting in hash tables.
1971 sk->sk_state = state;
1974 SOCK_DEBUG(sk, "TCP sk=%p, State %s -> %s\n", sk, statename[oldstate], statename[state]);
1977 EXPORT_SYMBOL_GPL(tcp_set_state);
1980 * State processing on a close. This implements the state shift for
1981 * sending our FIN frame. Note that we only send a FIN for some
1982 * states. A shutdown() may have already sent the FIN, or we may be
1986 static const unsigned char new_state[16] = {
1987 /* current state: new state: action: */
1988 /* (Invalid) */ TCP_CLOSE,
1989 /* TCP_ESTABLISHED */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1990 /* TCP_SYN_SENT */ TCP_CLOSE,
1991 /* TCP_SYN_RECV */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1992 /* TCP_FIN_WAIT1 */ TCP_FIN_WAIT1,
1993 /* TCP_FIN_WAIT2 */ TCP_FIN_WAIT2,
1994 /* TCP_TIME_WAIT */ TCP_CLOSE,
1995 /* TCP_CLOSE */ TCP_CLOSE,
1996 /* TCP_CLOSE_WAIT */ TCP_LAST_ACK | TCP_ACTION_FIN,
1997 /* TCP_LAST_ACK */ TCP_LAST_ACK,
1998 /* TCP_LISTEN */ TCP_CLOSE,
1999 /* TCP_CLOSING */ TCP_CLOSING,
2002 static int tcp_close_state(struct sock *sk)
2004 int next = (int)new_state[sk->sk_state];
2005 int ns = next & TCP_STATE_MASK;
2007 tcp_set_state(sk, ns);
2009 return next & TCP_ACTION_FIN;
2013 * Shutdown the sending side of a connection. Much like close except
2014 * that we don't receive shut down or sock_set_flag(sk, SOCK_DEAD).
2017 void tcp_shutdown(struct sock *sk, int how)
2019 /* We need to grab some memory, and put together a FIN,
2020 * and then put it into the queue to be sent.
2021 * Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
2023 if (!(how & SEND_SHUTDOWN))
2026 /* If we've already sent a FIN, or it's a closed state, skip this. */
2027 if ((1 << sk->sk_state) &
2028 (TCPF_ESTABLISHED | TCPF_SYN_SENT |
2029 TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) {
2030 /* Clear out any half completed packets. FIN if needed. */
2031 if (tcp_close_state(sk))
2035 EXPORT_SYMBOL(tcp_shutdown);
2037 bool tcp_check_oom(struct sock *sk, int shift)
2039 bool too_many_orphans, out_of_socket_memory;
2041 too_many_orphans = tcp_too_many_orphans(sk, shift);
2042 out_of_socket_memory = tcp_out_of_memory(sk);
2044 if (too_many_orphans)
2045 net_info_ratelimited("too many orphaned sockets\n");
2046 if (out_of_socket_memory)
2047 net_info_ratelimited("out of memory -- consider tuning tcp_mem\n");
2048 return too_many_orphans || out_of_socket_memory;
2051 void tcp_close(struct sock *sk, long timeout)
2053 struct sk_buff *skb;
2054 int data_was_unread = 0;
2058 sk->sk_shutdown = SHUTDOWN_MASK;
2060 if (sk->sk_state == TCP_LISTEN) {
2061 tcp_set_state(sk, TCP_CLOSE);
2064 inet_csk_listen_stop(sk);
2066 goto adjudge_to_death;
2069 /* We need to flush the recv. buffs. We do this only on the
2070 * descriptor close, not protocol-sourced closes, because the
2071 * reader process may not have drained the data yet!
2073 while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) {
2074 u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq -
2076 data_was_unread += len;
2082 /* If socket has been already reset (e.g. in tcp_reset()) - kill it. */
2083 if (sk->sk_state == TCP_CLOSE)
2084 goto adjudge_to_death;
2086 /* As outlined in RFC 2525, section 2.17, we send a RST here because
2087 * data was lost. To witness the awful effects of the old behavior of
2088 * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk
2089 * GET in an FTP client, suspend the process, wait for the client to
2090 * advertise a zero window, then kill -9 the FTP client, wheee...
2091 * Note: timeout is always zero in such a case.
2093 if (unlikely(tcp_sk(sk)->repair)) {
2094 sk->sk_prot->disconnect(sk, 0);
2095 } else if (data_was_unread) {
2096 /* Unread data was tossed, zap the connection. */
2097 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONCLOSE);
2098 tcp_set_state(sk, TCP_CLOSE);
2099 tcp_send_active_reset(sk, sk->sk_allocation);
2100 } else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) {
2101 /* Check zero linger _after_ checking for unread data. */
2102 sk->sk_prot->disconnect(sk, 0);
2103 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
2104 } else if (tcp_close_state(sk)) {
2105 /* We FIN if the application ate all the data before
2106 * zapping the connection.
2109 /* RED-PEN. Formally speaking, we have broken TCP state
2110 * machine. State transitions:
2112 * TCP_ESTABLISHED -> TCP_FIN_WAIT1
2113 * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible)
2114 * TCP_CLOSE_WAIT -> TCP_LAST_ACK
2116 * are legal only when FIN has been sent (i.e. in window),
2117 * rather than queued out of window. Purists blame.
2119 * F.e. "RFC state" is ESTABLISHED,
2120 * if Linux state is FIN-WAIT-1, but FIN is still not sent.
2122 * The visible declinations are that sometimes
2123 * we enter time-wait state, when it is not required really
2124 * (harmless), do not send active resets, when they are
2125 * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
2126 * they look as CLOSING or LAST_ACK for Linux)
2127 * Probably, I missed some more holelets.
2129 * XXX (TFO) - To start off we don't support SYN+ACK+FIN
2130 * in a single packet! (May consider it later but will
2131 * probably need API support or TCP_CORK SYN-ACK until
2132 * data is written and socket is closed.)
2137 sk_stream_wait_close(sk, timeout);
2140 state = sk->sk_state;
2144 /* It is the last release_sock in its life. It will remove backlog. */
2148 /* Now socket is owned by kernel and we acquire BH lock
2149 to finish close. No need to check for user refs.
2153 WARN_ON(sock_owned_by_user(sk));
2155 percpu_counter_inc(sk->sk_prot->orphan_count);
2157 /* Have we already been destroyed by a softirq or backlog? */
2158 if (state != TCP_CLOSE && sk->sk_state == TCP_CLOSE)
2161 /* This is a (useful) BSD violating of the RFC. There is a
2162 * problem with TCP as specified in that the other end could
2163 * keep a socket open forever with no application left this end.
2164 * We use a 3 minute timeout (about the same as BSD) then kill
2165 * our end. If they send after that then tough - BUT: long enough
2166 * that we won't make the old 4*rto = almost no time - whoops
2169 * Nope, it was not mistake. It is really desired behaviour
2170 * f.e. on http servers, when such sockets are useless, but
2171 * consume significant resources. Let's do it with special
2172 * linger2 option. --ANK
2175 if (sk->sk_state == TCP_FIN_WAIT2) {
2176 struct tcp_sock *tp = tcp_sk(sk);
2177 if (tp->linger2 < 0) {
2178 tcp_set_state(sk, TCP_CLOSE);
2179 tcp_send_active_reset(sk, GFP_ATOMIC);
2180 NET_INC_STATS_BH(sock_net(sk),
2181 LINUX_MIB_TCPABORTONLINGER);
2183 const int tmo = tcp_fin_time(sk);
2185 if (tmo > TCP_TIMEWAIT_LEN) {
2186 inet_csk_reset_keepalive_timer(sk,
2187 tmo - TCP_TIMEWAIT_LEN);
2189 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
2194 if (sk->sk_state != TCP_CLOSE) {
2196 if (tcp_check_oom(sk, 0)) {
2197 tcp_set_state(sk, TCP_CLOSE);
2198 tcp_send_active_reset(sk, GFP_ATOMIC);
2199 NET_INC_STATS_BH(sock_net(sk),
2200 LINUX_MIB_TCPABORTONMEMORY);
2204 if (sk->sk_state == TCP_CLOSE) {
2205 struct request_sock *req = tcp_sk(sk)->fastopen_rsk;
2206 /* We could get here with a non-NULL req if the socket is
2207 * aborted (e.g., closed with unread data) before 3WHS
2211 reqsk_fastopen_remove(sk, req, false);
2212 inet_csk_destroy_sock(sk);
2214 /* Otherwise, socket is reprieved until protocol close. */
2221 EXPORT_SYMBOL(tcp_close);
2223 /* These states need RST on ABORT according to RFC793 */
2225 static inline bool tcp_need_reset(int state)
2227 return (1 << state) &
2228 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 |
2229 TCPF_FIN_WAIT2 | TCPF_SYN_RECV);
2232 int tcp_disconnect(struct sock *sk, int flags)
2234 struct inet_sock *inet = inet_sk(sk);
2235 struct inet_connection_sock *icsk = inet_csk(sk);
2236 struct tcp_sock *tp = tcp_sk(sk);
2238 int old_state = sk->sk_state;
2240 if (old_state != TCP_CLOSE)
2241 tcp_set_state(sk, TCP_CLOSE);
2243 /* ABORT function of RFC793 */
2244 if (old_state == TCP_LISTEN) {
2245 inet_csk_listen_stop(sk);
2246 } else if (unlikely(tp->repair)) {
2247 sk->sk_err = ECONNABORTED;
2248 } else if (tcp_need_reset(old_state) ||
2249 (tp->snd_nxt != tp->write_seq &&
2250 (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK))) {
2251 /* The last check adjusts for discrepancy of Linux wrt. RFC
2254 tcp_send_active_reset(sk, gfp_any());
2255 sk->sk_err = ECONNRESET;
2256 } else if (old_state == TCP_SYN_SENT)
2257 sk->sk_err = ECONNRESET;
2259 tcp_clear_xmit_timers(sk);
2260 __skb_queue_purge(&sk->sk_receive_queue);
2261 tcp_write_queue_purge(sk);
2262 __skb_queue_purge(&tp->out_of_order_queue);
2263 #ifdef CONFIG_NET_DMA
2264 __skb_queue_purge(&sk->sk_async_wait_queue);
2267 inet->inet_dport = 0;
2269 if (!(sk->sk_userlocks & SOCK_BINDADDR_LOCK))
2270 inet_reset_saddr(sk);
2272 sk->sk_shutdown = 0;
2273 sock_reset_flag(sk, SOCK_DONE);
2275 if ((tp->write_seq += tp->max_window + 2) == 0)
2277 icsk->icsk_backoff = 0;
2279 icsk->icsk_probes_out = 0;
2280 tp->packets_out = 0;
2281 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
2282 tp->snd_cwnd_cnt = 0;
2283 tp->bytes_acked = 0;
2284 tp->window_clamp = 0;
2285 tcp_set_ca_state(sk, TCP_CA_Open);
2286 tcp_clear_retrans(tp);
2287 inet_csk_delack_init(sk);
2288 tcp_init_send_head(sk);
2289 memset(&tp->rx_opt, 0, sizeof(tp->rx_opt));
2292 WARN_ON(inet->inet_num && !icsk->icsk_bind_hash);
2294 sk->sk_error_report(sk);
2297 EXPORT_SYMBOL(tcp_disconnect);
2299 void tcp_sock_destruct(struct sock *sk)
2301 inet_sock_destruct(sk);
2303 kfree(inet_csk(sk)->icsk_accept_queue.fastopenq);
2306 static inline bool tcp_can_repair_sock(const struct sock *sk)
2308 return ns_capable(sock_net(sk)->user_ns, CAP_NET_ADMIN) &&
2309 ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_ESTABLISHED));
2312 static int tcp_repair_options_est(struct tcp_sock *tp,
2313 struct tcp_repair_opt __user *optbuf, unsigned int len)
2315 struct tcp_repair_opt opt;
2317 while (len >= sizeof(opt)) {
2318 if (copy_from_user(&opt, optbuf, sizeof(opt)))
2324 switch (opt.opt_code) {
2326 tp->rx_opt.mss_clamp = opt.opt_val;
2330 u16 snd_wscale = opt.opt_val & 0xFFFF;
2331 u16 rcv_wscale = opt.opt_val >> 16;
2333 if (snd_wscale > 14 || rcv_wscale > 14)
2336 tp->rx_opt.snd_wscale = snd_wscale;
2337 tp->rx_opt.rcv_wscale = rcv_wscale;
2338 tp->rx_opt.wscale_ok = 1;
2341 case TCPOPT_SACK_PERM:
2342 if (opt.opt_val != 0)
2345 tp->rx_opt.sack_ok |= TCP_SACK_SEEN;
2346 if (sysctl_tcp_fack)
2347 tcp_enable_fack(tp);
2349 case TCPOPT_TIMESTAMP:
2350 if (opt.opt_val != 0)
2353 tp->rx_opt.tstamp_ok = 1;
2362 * Socket option code for TCP.
2364 static int do_tcp_setsockopt(struct sock *sk, int level,
2365 int optname, char __user *optval, unsigned int optlen)
2367 struct tcp_sock *tp = tcp_sk(sk);
2368 struct inet_connection_sock *icsk = inet_csk(sk);
2372 /* These are data/string values, all the others are ints */
2374 case TCP_CONGESTION: {
2375 char name[TCP_CA_NAME_MAX];
2380 val = strncpy_from_user(name, optval,
2381 min_t(long, TCP_CA_NAME_MAX-1, optlen));
2387 err = tcp_set_congestion_control(sk, name);
2391 case TCP_COOKIE_TRANSACTIONS: {
2392 struct tcp_cookie_transactions ctd;
2393 struct tcp_cookie_values *cvp = NULL;
2395 if (sizeof(ctd) > optlen)
2397 if (copy_from_user(&ctd, optval, sizeof(ctd)))
2400 if (ctd.tcpct_used > sizeof(ctd.tcpct_value) ||
2401 ctd.tcpct_s_data_desired > TCP_MSS_DESIRED)
2404 if (ctd.tcpct_cookie_desired == 0) {
2405 /* default to global value */
2406 } else if ((0x1 & ctd.tcpct_cookie_desired) ||
2407 ctd.tcpct_cookie_desired > TCP_COOKIE_MAX ||
2408 ctd.tcpct_cookie_desired < TCP_COOKIE_MIN) {
2412 if (TCP_COOKIE_OUT_NEVER & ctd.tcpct_flags) {
2413 /* Supercedes all other values */
2415 if (tp->cookie_values != NULL) {
2416 kref_put(&tp->cookie_values->kref,
2417 tcp_cookie_values_release);
2418 tp->cookie_values = NULL;
2420 tp->rx_opt.cookie_in_always = 0; /* false */
2421 tp->rx_opt.cookie_out_never = 1; /* true */
2426 /* Allocate ancillary memory before locking.
2428 if (ctd.tcpct_used > 0 ||
2429 (tp->cookie_values == NULL &&
2430 (sysctl_tcp_cookie_size > 0 ||
2431 ctd.tcpct_cookie_desired > 0 ||
2432 ctd.tcpct_s_data_desired > 0))) {
2433 cvp = kzalloc(sizeof(*cvp) + ctd.tcpct_used,
2438 kref_init(&cvp->kref);
2441 tp->rx_opt.cookie_in_always =
2442 (TCP_COOKIE_IN_ALWAYS & ctd.tcpct_flags);
2443 tp->rx_opt.cookie_out_never = 0; /* false */
2445 if (tp->cookie_values != NULL) {
2447 /* Changed values are recorded by a changed
2448 * pointer, ensuring the cookie will differ,
2449 * without separately hashing each value later.
2451 kref_put(&tp->cookie_values->kref,
2452 tcp_cookie_values_release);
2454 cvp = tp->cookie_values;
2459 cvp->cookie_desired = ctd.tcpct_cookie_desired;
2461 if (ctd.tcpct_used > 0) {
2462 memcpy(cvp->s_data_payload, ctd.tcpct_value,
2464 cvp->s_data_desired = ctd.tcpct_used;
2465 cvp->s_data_constant = 1; /* true */
2467 /* No constant payload data. */
2468 cvp->s_data_desired = ctd.tcpct_s_data_desired;
2469 cvp->s_data_constant = 0; /* false */
2472 tp->cookie_values = cvp;
2482 if (optlen < sizeof(int))
2485 if (get_user(val, (int __user *)optval))
2492 /* Values greater than interface MTU won't take effect. However
2493 * at the point when this call is done we typically don't yet
2494 * know which interface is going to be used */
2495 if (val < TCP_MIN_MSS || val > MAX_TCP_WINDOW) {
2499 tp->rx_opt.user_mss = val;
2504 /* TCP_NODELAY is weaker than TCP_CORK, so that
2505 * this option on corked socket is remembered, but
2506 * it is not activated until cork is cleared.
2508 * However, when TCP_NODELAY is set we make
2509 * an explicit push, which overrides even TCP_CORK
2510 * for currently queued segments.
2512 tp->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH;
2513 tcp_push_pending_frames(sk);
2515 tp->nonagle &= ~TCP_NAGLE_OFF;
2519 case TCP_THIN_LINEAR_TIMEOUTS:
2520 if (val < 0 || val > 1)
2526 case TCP_THIN_DUPACK:
2527 if (val < 0 || val > 1)
2530 tp->thin_dupack = val;
2531 if (tp->thin_dupack)
2532 tcp_disable_early_retrans(tp);
2536 if (!tcp_can_repair_sock(sk))
2538 else if (val == 1) {
2540 sk->sk_reuse = SK_FORCE_REUSE;
2541 tp->repair_queue = TCP_NO_QUEUE;
2542 } else if (val == 0) {
2544 sk->sk_reuse = SK_NO_REUSE;
2545 tcp_send_window_probe(sk);
2551 case TCP_REPAIR_QUEUE:
2554 else if (val < TCP_QUEUES_NR)
2555 tp->repair_queue = val;
2561 if (sk->sk_state != TCP_CLOSE)
2563 else if (tp->repair_queue == TCP_SEND_QUEUE)
2564 tp->write_seq = val;
2565 else if (tp->repair_queue == TCP_RECV_QUEUE)
2571 case TCP_REPAIR_OPTIONS:
2574 else if (sk->sk_state == TCP_ESTABLISHED)
2575 err = tcp_repair_options_est(tp,
2576 (struct tcp_repair_opt __user *)optval,
2583 /* When set indicates to always queue non-full frames.
2584 * Later the user clears this option and we transmit
2585 * any pending partial frames in the queue. This is
2586 * meant to be used alongside sendfile() to get properly
2587 * filled frames when the user (for example) must write
2588 * out headers with a write() call first and then use
2589 * sendfile to send out the data parts.
2591 * TCP_CORK can be set together with TCP_NODELAY and it is
2592 * stronger than TCP_NODELAY.
2595 tp->nonagle |= TCP_NAGLE_CORK;
2597 tp->nonagle &= ~TCP_NAGLE_CORK;
2598 if (tp->nonagle&TCP_NAGLE_OFF)
2599 tp->nonagle |= TCP_NAGLE_PUSH;
2600 tcp_push_pending_frames(sk);
2605 if (val < 1 || val > MAX_TCP_KEEPIDLE)
2608 tp->keepalive_time = val * HZ;
2609 if (sock_flag(sk, SOCK_KEEPOPEN) &&
2610 !((1 << sk->sk_state) &
2611 (TCPF_CLOSE | TCPF_LISTEN))) {
2612 u32 elapsed = keepalive_time_elapsed(tp);
2613 if (tp->keepalive_time > elapsed)
2614 elapsed = tp->keepalive_time - elapsed;
2617 inet_csk_reset_keepalive_timer(sk, elapsed);
2622 if (val < 1 || val > MAX_TCP_KEEPINTVL)
2625 tp->keepalive_intvl = val * HZ;
2628 if (val < 1 || val > MAX_TCP_KEEPCNT)
2631 tp->keepalive_probes = val;
2634 if (val < 1 || val > MAX_TCP_SYNCNT)
2637 icsk->icsk_syn_retries = val;
2643 else if (val > sysctl_tcp_fin_timeout / HZ)
2646 tp->linger2 = val * HZ;
2649 case TCP_DEFER_ACCEPT:
2650 /* Translate value in seconds to number of retransmits */
2651 icsk->icsk_accept_queue.rskq_defer_accept =
2652 secs_to_retrans(val, TCP_TIMEOUT_INIT / HZ,
2656 case TCP_WINDOW_CLAMP:
2658 if (sk->sk_state != TCP_CLOSE) {
2662 tp->window_clamp = 0;
2664 tp->window_clamp = val < SOCK_MIN_RCVBUF / 2 ?
2665 SOCK_MIN_RCVBUF / 2 : val;
2670 icsk->icsk_ack.pingpong = 1;
2672 icsk->icsk_ack.pingpong = 0;
2673 if ((1 << sk->sk_state) &
2674 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) &&
2675 inet_csk_ack_scheduled(sk)) {
2676 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
2677 tcp_cleanup_rbuf(sk, 1);
2679 icsk->icsk_ack.pingpong = 1;
2684 #ifdef CONFIG_TCP_MD5SIG
2686 /* Read the IP->Key mappings from userspace */
2687 err = tp->af_specific->md5_parse(sk, optval, optlen);
2690 case TCP_USER_TIMEOUT:
2691 /* Cap the max timeout in ms TCP will retry/retrans
2692 * before giving up and aborting (ETIMEDOUT) a connection.
2697 icsk->icsk_user_timeout = msecs_to_jiffies(val);
2701 if (val >= 0 && ((1 << sk->sk_state) & (TCPF_CLOSE |
2703 err = fastopen_init_queue(sk, val);
2716 int tcp_setsockopt(struct sock *sk, int level, int optname, char __user *optval,
2717 unsigned int optlen)
2719 const struct inet_connection_sock *icsk = inet_csk(sk);
2721 if (level != SOL_TCP)
2722 return icsk->icsk_af_ops->setsockopt(sk, level, optname,
2724 return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2726 EXPORT_SYMBOL(tcp_setsockopt);
2728 #ifdef CONFIG_COMPAT
2729 int compat_tcp_setsockopt(struct sock *sk, int level, int optname,
2730 char __user *optval, unsigned int optlen)
2732 if (level != SOL_TCP)
2733 return inet_csk_compat_setsockopt(sk, level, optname,
2735 return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2737 EXPORT_SYMBOL(compat_tcp_setsockopt);
2740 /* Return information about state of tcp endpoint in API format. */
2741 void tcp_get_info(const struct sock *sk, struct tcp_info *info)
2743 const struct tcp_sock *tp = tcp_sk(sk);
2744 const struct inet_connection_sock *icsk = inet_csk(sk);
2745 u32 now = tcp_time_stamp;
2747 memset(info, 0, sizeof(*info));
2749 info->tcpi_state = sk->sk_state;
2750 info->tcpi_ca_state = icsk->icsk_ca_state;
2751 info->tcpi_retransmits = icsk->icsk_retransmits;
2752 info->tcpi_probes = icsk->icsk_probes_out;
2753 info->tcpi_backoff = icsk->icsk_backoff;
2755 if (tp->rx_opt.tstamp_ok)
2756 info->tcpi_options |= TCPI_OPT_TIMESTAMPS;
2757 if (tcp_is_sack(tp))
2758 info->tcpi_options |= TCPI_OPT_SACK;
2759 if (tp->rx_opt.wscale_ok) {
2760 info->tcpi_options |= TCPI_OPT_WSCALE;
2761 info->tcpi_snd_wscale = tp->rx_opt.snd_wscale;
2762 info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale;
2765 if (tp->ecn_flags & TCP_ECN_OK)
2766 info->tcpi_options |= TCPI_OPT_ECN;
2767 if (tp->ecn_flags & TCP_ECN_SEEN)
2768 info->tcpi_options |= TCPI_OPT_ECN_SEEN;
2769 if (tp->syn_data_acked)
2770 info->tcpi_options |= TCPI_OPT_SYN_DATA;
2772 info->tcpi_rto = jiffies_to_usecs(icsk->icsk_rto);
2773 info->tcpi_ato = jiffies_to_usecs(icsk->icsk_ack.ato);
2774 info->tcpi_snd_mss = tp->mss_cache;
2775 info->tcpi_rcv_mss = icsk->icsk_ack.rcv_mss;
2777 if (sk->sk_state == TCP_LISTEN) {
2778 info->tcpi_unacked = sk->sk_ack_backlog;
2779 info->tcpi_sacked = sk->sk_max_ack_backlog;
2781 info->tcpi_unacked = tp->packets_out;
2782 info->tcpi_sacked = tp->sacked_out;
2784 info->tcpi_lost = tp->lost_out;
2785 info->tcpi_retrans = tp->retrans_out;
2786 info->tcpi_fackets = tp->fackets_out;
2788 info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime);
2789 info->tcpi_last_data_recv = jiffies_to_msecs(now - icsk->icsk_ack.lrcvtime);
2790 info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp);
2792 info->tcpi_pmtu = icsk->icsk_pmtu_cookie;
2793 info->tcpi_rcv_ssthresh = tp->rcv_ssthresh;
2794 info->tcpi_rtt = jiffies_to_usecs(tp->srtt)>>3;
2795 info->tcpi_rttvar = jiffies_to_usecs(tp->mdev)>>2;
2796 info->tcpi_snd_ssthresh = tp->snd_ssthresh;
2797 info->tcpi_snd_cwnd = tp->snd_cwnd;
2798 info->tcpi_advmss = tp->advmss;
2799 info->tcpi_reordering = tp->reordering;
2801 info->tcpi_rcv_rtt = jiffies_to_usecs(tp->rcv_rtt_est.rtt)>>3;
2802 info->tcpi_rcv_space = tp->rcvq_space.space;
2804 info->tcpi_total_retrans = tp->total_retrans;
2806 EXPORT_SYMBOL_GPL(tcp_get_info);
2808 static int do_tcp_getsockopt(struct sock *sk, int level,
2809 int optname, char __user *optval, int __user *optlen)
2811 struct inet_connection_sock *icsk = inet_csk(sk);
2812 struct tcp_sock *tp = tcp_sk(sk);
2815 if (get_user(len, optlen))
2818 len = min_t(unsigned int, len, sizeof(int));
2825 val = tp->mss_cache;
2826 if (!val && ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
2827 val = tp->rx_opt.user_mss;
2829 val = tp->rx_opt.mss_clamp;
2832 val = !!(tp->nonagle&TCP_NAGLE_OFF);
2835 val = !!(tp->nonagle&TCP_NAGLE_CORK);
2838 val = keepalive_time_when(tp) / HZ;
2841 val = keepalive_intvl_when(tp) / HZ;
2844 val = keepalive_probes(tp);
2847 val = icsk->icsk_syn_retries ? : sysctl_tcp_syn_retries;
2852 val = (val ? : sysctl_tcp_fin_timeout) / HZ;
2854 case TCP_DEFER_ACCEPT:
2855 val = retrans_to_secs(icsk->icsk_accept_queue.rskq_defer_accept,
2856 TCP_TIMEOUT_INIT / HZ, TCP_RTO_MAX / HZ);
2858 case TCP_WINDOW_CLAMP:
2859 val = tp->window_clamp;
2862 struct tcp_info info;
2864 if (get_user(len, optlen))
2867 tcp_get_info(sk, &info);
2869 len = min_t(unsigned int, len, sizeof(info));
2870 if (put_user(len, optlen))
2872 if (copy_to_user(optval, &info, len))
2877 val = !icsk->icsk_ack.pingpong;
2880 case TCP_CONGESTION:
2881 if (get_user(len, optlen))
2883 len = min_t(unsigned int, len, TCP_CA_NAME_MAX);
2884 if (put_user(len, optlen))
2886 if (copy_to_user(optval, icsk->icsk_ca_ops->name, len))
2890 case TCP_COOKIE_TRANSACTIONS: {
2891 struct tcp_cookie_transactions ctd;
2892 struct tcp_cookie_values *cvp = tp->cookie_values;
2894 if (get_user(len, optlen))
2896 if (len < sizeof(ctd))
2899 memset(&ctd, 0, sizeof(ctd));
2900 ctd.tcpct_flags = (tp->rx_opt.cookie_in_always ?
2901 TCP_COOKIE_IN_ALWAYS : 0)
2902 | (tp->rx_opt.cookie_out_never ?
2903 TCP_COOKIE_OUT_NEVER : 0);
2906 ctd.tcpct_flags |= (cvp->s_data_in ?
2908 | (cvp->s_data_out ?
2909 TCP_S_DATA_OUT : 0);
2911 ctd.tcpct_cookie_desired = cvp->cookie_desired;
2912 ctd.tcpct_s_data_desired = cvp->s_data_desired;
2914 memcpy(&ctd.tcpct_value[0], &cvp->cookie_pair[0],
2915 cvp->cookie_pair_size);
2916 ctd.tcpct_used = cvp->cookie_pair_size;
2919 if (put_user(sizeof(ctd), optlen))
2921 if (copy_to_user(optval, &ctd, sizeof(ctd)))
2925 case TCP_THIN_LINEAR_TIMEOUTS:
2928 case TCP_THIN_DUPACK:
2929 val = tp->thin_dupack;
2936 case TCP_REPAIR_QUEUE:
2938 val = tp->repair_queue;
2944 if (tp->repair_queue == TCP_SEND_QUEUE)
2945 val = tp->write_seq;
2946 else if (tp->repair_queue == TCP_RECV_QUEUE)
2952 case TCP_USER_TIMEOUT:
2953 val = jiffies_to_msecs(icsk->icsk_user_timeout);
2956 return -ENOPROTOOPT;
2959 if (put_user(len, optlen))
2961 if (copy_to_user(optval, &val, len))
2966 int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval,
2969 struct inet_connection_sock *icsk = inet_csk(sk);
2971 if (level != SOL_TCP)
2972 return icsk->icsk_af_ops->getsockopt(sk, level, optname,
2974 return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2976 EXPORT_SYMBOL(tcp_getsockopt);
2978 #ifdef CONFIG_COMPAT
2979 int compat_tcp_getsockopt(struct sock *sk, int level, int optname,
2980 char __user *optval, int __user *optlen)
2982 if (level != SOL_TCP)
2983 return inet_csk_compat_getsockopt(sk, level, optname,
2985 return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2987 EXPORT_SYMBOL(compat_tcp_getsockopt);
2990 struct sk_buff *tcp_tso_segment(struct sk_buff *skb,
2991 netdev_features_t features)
2993 struct sk_buff *segs = ERR_PTR(-EINVAL);
2998 unsigned int oldlen;
3001 if (!pskb_may_pull(skb, sizeof(*th)))
3005 thlen = th->doff * 4;
3006 if (thlen < sizeof(*th))
3009 if (!pskb_may_pull(skb, thlen))
3012 oldlen = (u16)~skb->len;
3013 __skb_pull(skb, thlen);
3015 mss = skb_shinfo(skb)->gso_size;
3016 if (unlikely(skb->len <= mss))
3019 if (skb_gso_ok(skb, features | NETIF_F_GSO_ROBUST)) {
3020 /* Packet is from an untrusted source, reset gso_segs. */
3021 int type = skb_shinfo(skb)->gso_type;
3029 !(type & (SKB_GSO_TCPV4 | SKB_GSO_TCPV6))))
3032 skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss);
3038 segs = skb_segment(skb, features);
3042 delta = htonl(oldlen + (thlen + mss));
3046 seq = ntohl(th->seq);
3049 th->fin = th->psh = 0;
3051 th->check = ~csum_fold((__force __wsum)((__force u32)th->check +
3052 (__force u32)delta));
3053 if (skb->ip_summed != CHECKSUM_PARTIAL)
3055 csum_fold(csum_partial(skb_transport_header(skb),
3062 th->seq = htonl(seq);
3064 } while (skb->next);
3066 delta = htonl(oldlen + (skb->tail - skb->transport_header) +
3068 th->check = ~csum_fold((__force __wsum)((__force u32)th->check +
3069 (__force u32)delta));
3070 if (skb->ip_summed != CHECKSUM_PARTIAL)
3071 th->check = csum_fold(csum_partial(skb_transport_header(skb),
3077 EXPORT_SYMBOL(tcp_tso_segment);
3079 struct sk_buff **tcp_gro_receive(struct sk_buff **head, struct sk_buff *skb)
3081 struct sk_buff **pp = NULL;
3088 unsigned int mss = 1;
3094 off = skb_gro_offset(skb);
3095 hlen = off + sizeof(*th);
3096 th = skb_gro_header_fast(skb, off);
3097 if (skb_gro_header_hard(skb, hlen)) {
3098 th = skb_gro_header_slow(skb, hlen, off);
3103 thlen = th->doff * 4;
3104 if (thlen < sizeof(*th))
3108 if (skb_gro_header_hard(skb, hlen)) {
3109 th = skb_gro_header_slow(skb, hlen, off);
3114 skb_gro_pull(skb, thlen);
3116 len = skb_gro_len(skb);
3117 flags = tcp_flag_word(th);
3119 for (; (p = *head); head = &p->next) {
3120 if (!NAPI_GRO_CB(p)->same_flow)
3125 if (*(u32 *)&th->source ^ *(u32 *)&th2->source) {
3126 NAPI_GRO_CB(p)->same_flow = 0;
3133 goto out_check_final;
3136 flush = NAPI_GRO_CB(p)->flush;
3137 flush |= (__force int)(flags & TCP_FLAG_CWR);
3138 flush |= (__force int)((flags ^ tcp_flag_word(th2)) &
3139 ~(TCP_FLAG_CWR | TCP_FLAG_FIN | TCP_FLAG_PSH));
3140 flush |= (__force int)(th->ack_seq ^ th2->ack_seq);
3141 for (i = sizeof(*th); i < thlen; i += 4)
3142 flush |= *(u32 *)((u8 *)th + i) ^
3143 *(u32 *)((u8 *)th2 + i);
3145 mss = skb_shinfo(p)->gso_size;
3147 flush |= (len - 1) >= mss;
3148 flush |= (ntohl(th2->seq) + skb_gro_len(p)) ^ ntohl(th->seq);
3150 if (flush || skb_gro_receive(head, skb)) {
3152 goto out_check_final;
3157 tcp_flag_word(th2) |= flags & (TCP_FLAG_FIN | TCP_FLAG_PSH);
3161 flush |= (__force int)(flags & (TCP_FLAG_URG | TCP_FLAG_PSH |
3162 TCP_FLAG_RST | TCP_FLAG_SYN |
3165 if (p && (!NAPI_GRO_CB(skb)->same_flow || flush))
3169 NAPI_GRO_CB(skb)->flush |= flush;
3173 EXPORT_SYMBOL(tcp_gro_receive);
3175 int tcp_gro_complete(struct sk_buff *skb)
3177 struct tcphdr *th = tcp_hdr(skb);
3179 skb->csum_start = skb_transport_header(skb) - skb->head;
3180 skb->csum_offset = offsetof(struct tcphdr, check);
3181 skb->ip_summed = CHECKSUM_PARTIAL;
3183 skb_shinfo(skb)->gso_segs = NAPI_GRO_CB(skb)->count;
3186 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
3190 EXPORT_SYMBOL(tcp_gro_complete);
3192 #ifdef CONFIG_TCP_MD5SIG
3193 static unsigned long tcp_md5sig_users;
3194 static struct tcp_md5sig_pool __percpu *tcp_md5sig_pool;
3195 static DEFINE_SPINLOCK(tcp_md5sig_pool_lock);
3197 static void __tcp_free_md5sig_pool(struct tcp_md5sig_pool __percpu *pool)
3201 for_each_possible_cpu(cpu) {
3202 struct tcp_md5sig_pool *p = per_cpu_ptr(pool, cpu);
3204 if (p->md5_desc.tfm)
3205 crypto_free_hash(p->md5_desc.tfm);
3210 void tcp_free_md5sig_pool(void)
3212 struct tcp_md5sig_pool __percpu *pool = NULL;
3214 spin_lock_bh(&tcp_md5sig_pool_lock);
3215 if (--tcp_md5sig_users == 0) {
3216 pool = tcp_md5sig_pool;
3217 tcp_md5sig_pool = NULL;
3219 spin_unlock_bh(&tcp_md5sig_pool_lock);
3221 __tcp_free_md5sig_pool(pool);
3223 EXPORT_SYMBOL(tcp_free_md5sig_pool);
3225 static struct tcp_md5sig_pool __percpu *
3226 __tcp_alloc_md5sig_pool(struct sock *sk)
3229 struct tcp_md5sig_pool __percpu *pool;
3231 pool = alloc_percpu(struct tcp_md5sig_pool);
3235 for_each_possible_cpu(cpu) {
3236 struct crypto_hash *hash;
3238 hash = crypto_alloc_hash("md5", 0, CRYPTO_ALG_ASYNC);
3239 if (!hash || IS_ERR(hash))
3242 per_cpu_ptr(pool, cpu)->md5_desc.tfm = hash;
3246 __tcp_free_md5sig_pool(pool);
3250 struct tcp_md5sig_pool __percpu *tcp_alloc_md5sig_pool(struct sock *sk)
3252 struct tcp_md5sig_pool __percpu *pool;
3256 spin_lock_bh(&tcp_md5sig_pool_lock);
3257 pool = tcp_md5sig_pool;
3258 if (tcp_md5sig_users++ == 0) {
3260 spin_unlock_bh(&tcp_md5sig_pool_lock);
3263 spin_unlock_bh(&tcp_md5sig_pool_lock);
3267 spin_unlock_bh(&tcp_md5sig_pool_lock);
3270 /* we cannot hold spinlock here because this may sleep. */
3271 struct tcp_md5sig_pool __percpu *p;
3273 p = __tcp_alloc_md5sig_pool(sk);
3274 spin_lock_bh(&tcp_md5sig_pool_lock);
3277 spin_unlock_bh(&tcp_md5sig_pool_lock);
3280 pool = tcp_md5sig_pool;
3282 /* oops, it has already been assigned. */
3283 spin_unlock_bh(&tcp_md5sig_pool_lock);
3284 __tcp_free_md5sig_pool(p);
3286 tcp_md5sig_pool = pool = p;
3287 spin_unlock_bh(&tcp_md5sig_pool_lock);
3292 EXPORT_SYMBOL(tcp_alloc_md5sig_pool);
3296 * tcp_get_md5sig_pool - get md5sig_pool for this user
3298 * We use percpu structure, so if we succeed, we exit with preemption
3299 * and BH disabled, to make sure another thread or softirq handling
3300 * wont try to get same context.
3302 struct tcp_md5sig_pool *tcp_get_md5sig_pool(void)
3304 struct tcp_md5sig_pool __percpu *p;
3308 spin_lock(&tcp_md5sig_pool_lock);
3309 p = tcp_md5sig_pool;
3312 spin_unlock(&tcp_md5sig_pool_lock);
3315 return this_cpu_ptr(p);
3320 EXPORT_SYMBOL(tcp_get_md5sig_pool);
3322 void tcp_put_md5sig_pool(void)
3325 tcp_free_md5sig_pool();
3327 EXPORT_SYMBOL(tcp_put_md5sig_pool);
3329 int tcp_md5_hash_header(struct tcp_md5sig_pool *hp,
3330 const struct tcphdr *th)
3332 struct scatterlist sg;
3336 /* We are not allowed to change tcphdr, make a local copy */
3337 memcpy(&hdr, th, sizeof(hdr));
3340 /* options aren't included in the hash */
3341 sg_init_one(&sg, &hdr, sizeof(hdr));
3342 err = crypto_hash_update(&hp->md5_desc, &sg, sizeof(hdr));
3345 EXPORT_SYMBOL(tcp_md5_hash_header);
3347 int tcp_md5_hash_skb_data(struct tcp_md5sig_pool *hp,
3348 const struct sk_buff *skb, unsigned int header_len)
3350 struct scatterlist sg;
3351 const struct tcphdr *tp = tcp_hdr(skb);
3352 struct hash_desc *desc = &hp->md5_desc;
3354 const unsigned int head_data_len = skb_headlen(skb) > header_len ?
3355 skb_headlen(skb) - header_len : 0;
3356 const struct skb_shared_info *shi = skb_shinfo(skb);
3357 struct sk_buff *frag_iter;
3359 sg_init_table(&sg, 1);
3361 sg_set_buf(&sg, ((u8 *) tp) + header_len, head_data_len);
3362 if (crypto_hash_update(desc, &sg, head_data_len))
3365 for (i = 0; i < shi->nr_frags; ++i) {
3366 const struct skb_frag_struct *f = &shi->frags[i];
3367 struct page *page = skb_frag_page(f);
3368 sg_set_page(&sg, page, skb_frag_size(f), f->page_offset);
3369 if (crypto_hash_update(desc, &sg, skb_frag_size(f)))
3373 skb_walk_frags(skb, frag_iter)
3374 if (tcp_md5_hash_skb_data(hp, frag_iter, 0))
3379 EXPORT_SYMBOL(tcp_md5_hash_skb_data);
3381 int tcp_md5_hash_key(struct tcp_md5sig_pool *hp, const struct tcp_md5sig_key *key)
3383 struct scatterlist sg;
3385 sg_init_one(&sg, key->key, key->keylen);
3386 return crypto_hash_update(&hp->md5_desc, &sg, key->keylen);
3388 EXPORT_SYMBOL(tcp_md5_hash_key);
3392 /* Each Responder maintains up to two secret values concurrently for
3393 * efficient secret rollover. Each secret value has 4 states:
3395 * Generating. (tcp_secret_generating != tcp_secret_primary)
3396 * Generates new Responder-Cookies, but not yet used for primary
3397 * verification. This is a short-term state, typically lasting only
3398 * one round trip time (RTT).
3400 * Primary. (tcp_secret_generating == tcp_secret_primary)
3401 * Used both for generation and primary verification.
3403 * Retiring. (tcp_secret_retiring != tcp_secret_secondary)
3404 * Used for verification, until the first failure that can be
3405 * verified by the newer Generating secret. At that time, this
3406 * cookie's state is changed to Secondary, and the Generating
3407 * cookie's state is changed to Primary. This is a short-term state,
3408 * typically lasting only one round trip time (RTT).
3410 * Secondary. (tcp_secret_retiring == tcp_secret_secondary)
3411 * Used for secondary verification, after primary verification
3412 * failures. This state lasts no more than twice the Maximum Segment
3413 * Lifetime (2MSL). Then, the secret is discarded.
3415 struct tcp_cookie_secret {
3416 /* The secret is divided into two parts. The digest part is the
3417 * equivalent of previously hashing a secret and saving the state,
3418 * and serves as an initialization vector (IV). The message part
3419 * serves as the trailing secret.
3421 u32 secrets[COOKIE_WORKSPACE_WORDS];
3422 unsigned long expires;
3425 #define TCP_SECRET_1MSL (HZ * TCP_PAWS_MSL)
3426 #define TCP_SECRET_2MSL (HZ * TCP_PAWS_MSL * 2)
3427 #define TCP_SECRET_LIFE (HZ * 600)
3429 static struct tcp_cookie_secret tcp_secret_one;
3430 static struct tcp_cookie_secret tcp_secret_two;
3432 /* Essentially a circular list, without dynamic allocation. */
3433 static struct tcp_cookie_secret *tcp_secret_generating;
3434 static struct tcp_cookie_secret *tcp_secret_primary;
3435 static struct tcp_cookie_secret *tcp_secret_retiring;
3436 static struct tcp_cookie_secret *tcp_secret_secondary;
3438 static DEFINE_SPINLOCK(tcp_secret_locker);
3440 /* Select a pseudo-random word in the cookie workspace.
3442 static inline u32 tcp_cookie_work(const u32 *ws, const int n)
3444 return ws[COOKIE_DIGEST_WORDS + ((COOKIE_MESSAGE_WORDS-1) & ws[n])];
3447 /* Fill bakery[COOKIE_WORKSPACE_WORDS] with generator, updating as needed.
3448 * Called in softirq context.
3449 * Returns: 0 for success.
3451 int tcp_cookie_generator(u32 *bakery)
3453 unsigned long jiffy = jiffies;
3455 if (unlikely(time_after_eq(jiffy, tcp_secret_generating->expires))) {
3456 spin_lock_bh(&tcp_secret_locker);
3457 if (!time_after_eq(jiffy, tcp_secret_generating->expires)) {
3458 /* refreshed by another */
3460 &tcp_secret_generating->secrets[0],
3461 COOKIE_WORKSPACE_WORDS);
3463 /* still needs refreshing */
3464 get_random_bytes(bakery, COOKIE_WORKSPACE_WORDS);
3466 /* The first time, paranoia assumes that the
3467 * randomization function isn't as strong. But,
3468 * this secret initialization is delayed until
3469 * the last possible moment (packet arrival).
3470 * Although that time is observable, it is
3471 * unpredictably variable. Mash in the most
3472 * volatile clock bits available, and expire the
3473 * secret extra quickly.
3475 if (unlikely(tcp_secret_primary->expires ==
3476 tcp_secret_secondary->expires)) {
3479 getnstimeofday(&tv);
3480 bakery[COOKIE_DIGEST_WORDS+0] ^=
3483 tcp_secret_secondary->expires = jiffy
3485 + (0x0f & tcp_cookie_work(bakery, 0));
3487 tcp_secret_secondary->expires = jiffy
3489 + (0xff & tcp_cookie_work(bakery, 1));
3490 tcp_secret_primary->expires = jiffy
3492 + (0x1f & tcp_cookie_work(bakery, 2));
3494 memcpy(&tcp_secret_secondary->secrets[0],
3495 bakery, COOKIE_WORKSPACE_WORDS);
3497 rcu_assign_pointer(tcp_secret_generating,
3498 tcp_secret_secondary);
3499 rcu_assign_pointer(tcp_secret_retiring,
3500 tcp_secret_primary);
3502 * Neither call_rcu() nor synchronize_rcu() needed.
3503 * Retiring data is not freed. It is replaced after
3504 * further (locked) pointer updates, and a quiet time
3505 * (minimum 1MSL, maximum LIFE - 2MSL).
3508 spin_unlock_bh(&tcp_secret_locker);
3512 &rcu_dereference(tcp_secret_generating)->secrets[0],
3513 COOKIE_WORKSPACE_WORDS);
3514 rcu_read_unlock_bh();
3518 EXPORT_SYMBOL(tcp_cookie_generator);
3520 void tcp_done(struct sock *sk)
3522 struct request_sock *req = tcp_sk(sk)->fastopen_rsk;
3524 if (sk->sk_state == TCP_SYN_SENT || sk->sk_state == TCP_SYN_RECV)
3525 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_ATTEMPTFAILS);
3527 tcp_set_state(sk, TCP_CLOSE);
3528 tcp_clear_xmit_timers(sk);
3530 reqsk_fastopen_remove(sk, req, false);
3532 sk->sk_shutdown = SHUTDOWN_MASK;
3534 if (!sock_flag(sk, SOCK_DEAD))
3535 sk->sk_state_change(sk);
3537 inet_csk_destroy_sock(sk);
3539 EXPORT_SYMBOL_GPL(tcp_done);
3541 extern struct tcp_congestion_ops tcp_reno;
3543 static __initdata unsigned long thash_entries;
3544 static int __init set_thash_entries(char *str)
3551 ret = kstrtoul(str, 0, &thash_entries);
3557 __setup("thash_entries=", set_thash_entries);
3559 void tcp_init_mem(struct net *net)
3561 unsigned long limit = nr_free_buffer_pages() / 8;
3562 limit = max(limit, 128UL);
3563 net->ipv4.sysctl_tcp_mem[0] = limit / 4 * 3;
3564 net->ipv4.sysctl_tcp_mem[1] = limit;
3565 net->ipv4.sysctl_tcp_mem[2] = net->ipv4.sysctl_tcp_mem[0] * 2;
3568 void __init tcp_init(void)
3570 struct sk_buff *skb = NULL;
3571 unsigned long limit;
3572 int max_rshare, max_wshare, cnt;
3574 unsigned long jiffy = jiffies;
3576 BUILD_BUG_ON(sizeof(struct tcp_skb_cb) > sizeof(skb->cb));
3578 percpu_counter_init(&tcp_sockets_allocated, 0);
3579 percpu_counter_init(&tcp_orphan_count, 0);
3580 tcp_hashinfo.bind_bucket_cachep =
3581 kmem_cache_create("tcp_bind_bucket",
3582 sizeof(struct inet_bind_bucket), 0,
3583 SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL);
3585 /* Size and allocate the main established and bind bucket
3588 * The methodology is similar to that of the buffer cache.
3590 tcp_hashinfo.ehash =
3591 alloc_large_system_hash("TCP established",
3592 sizeof(struct inet_ehash_bucket),
3594 17, /* one slot per 128 KB of memory */
3597 &tcp_hashinfo.ehash_mask,
3599 thash_entries ? 0 : 512 * 1024);
3600 for (i = 0; i <= tcp_hashinfo.ehash_mask; i++) {
3601 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].chain, i);
3602 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].twchain, i);
3604 if (inet_ehash_locks_alloc(&tcp_hashinfo))
3605 panic("TCP: failed to alloc ehash_locks");
3606 tcp_hashinfo.bhash =
3607 alloc_large_system_hash("TCP bind",
3608 sizeof(struct inet_bind_hashbucket),
3609 tcp_hashinfo.ehash_mask + 1,
3610 17, /* one slot per 128 KB of memory */
3612 &tcp_hashinfo.bhash_size,
3616 tcp_hashinfo.bhash_size = 1U << tcp_hashinfo.bhash_size;
3617 for (i = 0; i < tcp_hashinfo.bhash_size; i++) {
3618 spin_lock_init(&tcp_hashinfo.bhash[i].lock);
3619 INIT_HLIST_HEAD(&tcp_hashinfo.bhash[i].chain);
3623 cnt = tcp_hashinfo.ehash_mask + 1;
3625 tcp_death_row.sysctl_max_tw_buckets = cnt / 2;
3626 sysctl_tcp_max_orphans = cnt / 2;
3627 sysctl_max_syn_backlog = max(128, cnt / 256);
3629 tcp_init_mem(&init_net);
3630 /* Set per-socket limits to no more than 1/128 the pressure threshold */
3631 limit = nr_free_buffer_pages() << (PAGE_SHIFT - 7);
3632 max_wshare = min(4UL*1024*1024, limit);
3633 max_rshare = min(6UL*1024*1024, limit);
3635 sysctl_tcp_wmem[0] = SK_MEM_QUANTUM;
3636 sysctl_tcp_wmem[1] = 16*1024;
3637 sysctl_tcp_wmem[2] = max(64*1024, max_wshare);
3639 sysctl_tcp_rmem[0] = SK_MEM_QUANTUM;
3640 sysctl_tcp_rmem[1] = 87380;
3641 sysctl_tcp_rmem[2] = max(87380, max_rshare);
3643 pr_info("Hash tables configured (established %u bind %u)\n",
3644 tcp_hashinfo.ehash_mask + 1, tcp_hashinfo.bhash_size);
3648 tcp_register_congestion_control(&tcp_reno);
3650 memset(&tcp_secret_one.secrets[0], 0, sizeof(tcp_secret_one.secrets));
3651 memset(&tcp_secret_two.secrets[0], 0, sizeof(tcp_secret_two.secrets));
3652 tcp_secret_one.expires = jiffy; /* past due */
3653 tcp_secret_two.expires = jiffy; /* past due */
3654 tcp_secret_generating = &tcp_secret_one;
3655 tcp_secret_primary = &tcp_secret_one;
3656 tcp_secret_retiring = &tcp_secret_two;
3657 tcp_secret_secondary = &tcp_secret_two;