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>
282 #include <net/busy_poll.h>
284 int sysctl_tcp_fin_timeout __read_mostly = TCP_FIN_TIMEOUT;
286 struct percpu_counter tcp_orphan_count;
287 EXPORT_SYMBOL_GPL(tcp_orphan_count);
289 int sysctl_tcp_wmem[3] __read_mostly;
290 int sysctl_tcp_rmem[3] __read_mostly;
292 EXPORT_SYMBOL(sysctl_tcp_rmem);
293 EXPORT_SYMBOL(sysctl_tcp_wmem);
295 atomic_long_t tcp_memory_allocated; /* Current allocated memory. */
296 EXPORT_SYMBOL(tcp_memory_allocated);
299 * Current number of TCP sockets.
301 struct percpu_counter tcp_sockets_allocated;
302 EXPORT_SYMBOL(tcp_sockets_allocated);
307 struct tcp_splice_state {
308 struct pipe_inode_info *pipe;
314 * Pressure flag: try to collapse.
315 * Technical note: it is used by multiple contexts non atomically.
316 * All the __sk_mem_schedule() is of this nature: accounting
317 * is strict, actions are advisory and have some latency.
319 int tcp_memory_pressure __read_mostly;
320 EXPORT_SYMBOL(tcp_memory_pressure);
322 void tcp_enter_memory_pressure(struct sock *sk)
324 if (!tcp_memory_pressure) {
325 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMEMORYPRESSURES);
326 tcp_memory_pressure = 1;
329 EXPORT_SYMBOL(tcp_enter_memory_pressure);
331 /* Convert seconds to retransmits based on initial and max timeout */
332 static u8 secs_to_retrans(int seconds, int timeout, int rto_max)
337 int period = timeout;
340 while (seconds > period && res < 255) {
343 if (timeout > rto_max)
351 /* Convert retransmits to seconds based on initial and max timeout */
352 static int retrans_to_secs(u8 retrans, int timeout, int rto_max)
360 if (timeout > rto_max)
368 /* Address-family independent initialization for a tcp_sock.
370 * NOTE: A lot of things set to zero explicitly by call to
371 * sk_alloc() so need not be done here.
373 void tcp_init_sock(struct sock *sk)
375 struct inet_connection_sock *icsk = inet_csk(sk);
376 struct tcp_sock *tp = tcp_sk(sk);
378 skb_queue_head_init(&tp->out_of_order_queue);
379 tcp_init_xmit_timers(sk);
380 tcp_prequeue_init(tp);
381 INIT_LIST_HEAD(&tp->tsq_node);
383 icsk->icsk_rto = TCP_TIMEOUT_INIT;
384 tp->mdev = TCP_TIMEOUT_INIT;
386 /* So many TCP implementations out there (incorrectly) count the
387 * initial SYN frame in their delayed-ACK and congestion control
388 * algorithms that we must have the following bandaid to talk
389 * efficiently to them. -DaveM
391 tp->snd_cwnd = TCP_INIT_CWND;
393 /* See draft-stevens-tcpca-spec-01 for discussion of the
394 * initialization of these values.
396 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
397 tp->snd_cwnd_clamp = ~0;
398 tp->mss_cache = TCP_MSS_DEFAULT;
400 tp->reordering = sysctl_tcp_reordering;
401 tcp_enable_early_retrans(tp);
402 icsk->icsk_ca_ops = &tcp_init_congestion_ops;
406 sk->sk_state = TCP_CLOSE;
408 sk->sk_write_space = sk_stream_write_space;
409 sock_set_flag(sk, SOCK_USE_WRITE_QUEUE);
411 icsk->icsk_sync_mss = tcp_sync_mss;
413 sk->sk_sndbuf = sysctl_tcp_wmem[1];
414 sk->sk_rcvbuf = sysctl_tcp_rmem[1];
417 sock_update_memcg(sk);
418 sk_sockets_allocated_inc(sk);
421 EXPORT_SYMBOL(tcp_init_sock);
424 * Wait for a TCP event.
426 * Note that we don't need to lock the socket, as the upper poll layers
427 * take care of normal races (between the test and the event) and we don't
428 * go look at any of the socket buffers directly.
430 unsigned int tcp_poll(struct file *file, struct socket *sock, poll_table *wait)
433 struct sock *sk = sock->sk;
434 const struct tcp_sock *tp = tcp_sk(sk);
436 sock_rps_record_flow(sk);
438 sock_poll_wait(file, sk_sleep(sk), wait);
439 if (sk->sk_state == TCP_LISTEN)
440 return inet_csk_listen_poll(sk);
442 /* Socket is not locked. We are protected from async events
443 * by poll logic and correct handling of state changes
444 * made by other threads is impossible in any case.
450 * POLLHUP is certainly not done right. But poll() doesn't
451 * have a notion of HUP in just one direction, and for a
452 * socket the read side is more interesting.
454 * Some poll() documentation says that POLLHUP is incompatible
455 * with the POLLOUT/POLLWR flags, so somebody should check this
456 * all. But careful, it tends to be safer to return too many
457 * bits than too few, and you can easily break real applications
458 * if you don't tell them that something has hung up!
462 * Check number 1. POLLHUP is _UNMASKABLE_ event (see UNIX98 and
463 * our fs/select.c). It means that after we received EOF,
464 * poll always returns immediately, making impossible poll() on write()
465 * in state CLOSE_WAIT. One solution is evident --- to set POLLHUP
466 * if and only if shutdown has been made in both directions.
467 * Actually, it is interesting to look how Solaris and DUX
468 * solve this dilemma. I would prefer, if POLLHUP were maskable,
469 * then we could set it on SND_SHUTDOWN. BTW examples given
470 * in Stevens' books assume exactly this behaviour, it explains
471 * why POLLHUP is incompatible with POLLOUT. --ANK
473 * NOTE. Check for TCP_CLOSE is added. The goal is to prevent
474 * blocking on fresh not-connected or disconnected socket. --ANK
476 if (sk->sk_shutdown == SHUTDOWN_MASK || sk->sk_state == TCP_CLOSE)
478 if (sk->sk_shutdown & RCV_SHUTDOWN)
479 mask |= POLLIN | POLLRDNORM | POLLRDHUP;
481 /* Connected or passive Fast Open socket? */
482 if (sk->sk_state != TCP_SYN_SENT &&
483 (sk->sk_state != TCP_SYN_RECV || tp->fastopen_rsk != NULL)) {
484 int target = sock_rcvlowat(sk, 0, INT_MAX);
486 if (tp->urg_seq == tp->copied_seq &&
487 !sock_flag(sk, SOCK_URGINLINE) &&
491 /* Potential race condition. If read of tp below will
492 * escape above sk->sk_state, we can be illegally awaken
493 * in SYN_* states. */
494 if (tp->rcv_nxt - tp->copied_seq >= target)
495 mask |= POLLIN | POLLRDNORM;
497 if (!(sk->sk_shutdown & SEND_SHUTDOWN)) {
498 if (sk_stream_is_writeable(sk)) {
499 mask |= POLLOUT | POLLWRNORM;
500 } else { /* send SIGIO later */
501 set_bit(SOCK_ASYNC_NOSPACE,
502 &sk->sk_socket->flags);
503 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
505 /* Race breaker. If space is freed after
506 * wspace test but before the flags are set,
507 * IO signal will be lost.
509 if (sk_stream_is_writeable(sk))
510 mask |= POLLOUT | POLLWRNORM;
513 mask |= POLLOUT | POLLWRNORM;
515 if (tp->urg_data & TCP_URG_VALID)
518 /* This barrier is coupled with smp_wmb() in tcp_reset() */
525 EXPORT_SYMBOL(tcp_poll);
527 int tcp_ioctl(struct sock *sk, int cmd, unsigned long arg)
529 struct tcp_sock *tp = tcp_sk(sk);
535 if (sk->sk_state == TCP_LISTEN)
538 slow = lock_sock_fast(sk);
539 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
541 else if (sock_flag(sk, SOCK_URGINLINE) ||
543 before(tp->urg_seq, tp->copied_seq) ||
544 !before(tp->urg_seq, tp->rcv_nxt)) {
546 answ = tp->rcv_nxt - tp->copied_seq;
548 /* Subtract 1, if FIN was received */
549 if (answ && sock_flag(sk, SOCK_DONE))
552 answ = tp->urg_seq - tp->copied_seq;
553 unlock_sock_fast(sk, slow);
556 answ = tp->urg_data && tp->urg_seq == tp->copied_seq;
559 if (sk->sk_state == TCP_LISTEN)
562 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
565 answ = tp->write_seq - tp->snd_una;
568 if (sk->sk_state == TCP_LISTEN)
571 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
574 answ = tp->write_seq - tp->snd_nxt;
580 return put_user(answ, (int __user *)arg);
582 EXPORT_SYMBOL(tcp_ioctl);
584 static inline void tcp_mark_push(struct tcp_sock *tp, struct sk_buff *skb)
586 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
587 tp->pushed_seq = tp->write_seq;
590 static inline bool forced_push(const struct tcp_sock *tp)
592 return after(tp->write_seq, tp->pushed_seq + (tp->max_window >> 1));
595 static inline void skb_entail(struct sock *sk, struct sk_buff *skb)
597 struct tcp_sock *tp = tcp_sk(sk);
598 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
601 tcb->seq = tcb->end_seq = tp->write_seq;
602 tcb->tcp_flags = TCPHDR_ACK;
604 skb_header_release(skb);
605 tcp_add_write_queue_tail(sk, skb);
606 sk->sk_wmem_queued += skb->truesize;
607 sk_mem_charge(sk, skb->truesize);
608 if (tp->nonagle & TCP_NAGLE_PUSH)
609 tp->nonagle &= ~TCP_NAGLE_PUSH;
612 static inline void tcp_mark_urg(struct tcp_sock *tp, int flags)
615 tp->snd_up = tp->write_seq;
618 static inline void tcp_push(struct sock *sk, int flags, int mss_now,
621 if (tcp_send_head(sk)) {
622 struct tcp_sock *tp = tcp_sk(sk);
624 if (!(flags & MSG_MORE) || forced_push(tp))
625 tcp_mark_push(tp, tcp_write_queue_tail(sk));
627 tcp_mark_urg(tp, flags);
628 __tcp_push_pending_frames(sk, mss_now,
629 (flags & MSG_MORE) ? TCP_NAGLE_CORK : nonagle);
633 static int tcp_splice_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb,
634 unsigned int offset, size_t len)
636 struct tcp_splice_state *tss = rd_desc->arg.data;
639 ret = skb_splice_bits(skb, offset, tss->pipe, min(rd_desc->count, len),
642 rd_desc->count -= ret;
646 static int __tcp_splice_read(struct sock *sk, struct tcp_splice_state *tss)
648 /* Store TCP splice context information in read_descriptor_t. */
649 read_descriptor_t rd_desc = {
654 return tcp_read_sock(sk, &rd_desc, tcp_splice_data_recv);
658 * tcp_splice_read - splice data from TCP socket to a pipe
659 * @sock: socket to splice from
660 * @ppos: position (not valid)
661 * @pipe: pipe to splice to
662 * @len: number of bytes to splice
663 * @flags: splice modifier flags
666 * Will read pages from given socket and fill them into a pipe.
669 ssize_t tcp_splice_read(struct socket *sock, loff_t *ppos,
670 struct pipe_inode_info *pipe, size_t len,
673 struct sock *sk = sock->sk;
674 struct tcp_splice_state tss = {
683 sock_rps_record_flow(sk);
685 * We can't seek on a socket input
694 timeo = sock_rcvtimeo(sk, sock->file->f_flags & O_NONBLOCK);
696 ret = __tcp_splice_read(sk, &tss);
702 if (sock_flag(sk, SOCK_DONE))
705 ret = sock_error(sk);
708 if (sk->sk_shutdown & RCV_SHUTDOWN)
710 if (sk->sk_state == TCP_CLOSE) {
712 * This occurs when user tries to read
713 * from never connected socket.
715 if (!sock_flag(sk, SOCK_DONE))
723 sk_wait_data(sk, &timeo);
724 if (signal_pending(current)) {
725 ret = sock_intr_errno(timeo);
738 if (sk->sk_err || sk->sk_state == TCP_CLOSE ||
739 (sk->sk_shutdown & RCV_SHUTDOWN) ||
740 signal_pending(current))
751 EXPORT_SYMBOL(tcp_splice_read);
753 struct sk_buff *sk_stream_alloc_skb(struct sock *sk, int size, gfp_t gfp)
757 /* The TCP header must be at least 32-bit aligned. */
758 size = ALIGN(size, 4);
760 skb = alloc_skb_fclone(size + sk->sk_prot->max_header, gfp);
762 if (sk_wmem_schedule(sk, skb->truesize)) {
763 skb_reserve(skb, sk->sk_prot->max_header);
765 * Make sure that we have exactly size bytes
766 * available to the caller, no more, no less.
768 skb->reserved_tailroom = skb->end - skb->tail - size;
773 sk->sk_prot->enter_memory_pressure(sk);
774 sk_stream_moderate_sndbuf(sk);
779 static unsigned int tcp_xmit_size_goal(struct sock *sk, u32 mss_now,
782 struct tcp_sock *tp = tcp_sk(sk);
783 u32 xmit_size_goal, old_size_goal;
785 xmit_size_goal = mss_now;
787 if (large_allowed && sk_can_gso(sk)) {
788 xmit_size_goal = ((sk->sk_gso_max_size - 1) -
789 inet_csk(sk)->icsk_af_ops->net_header_len -
790 inet_csk(sk)->icsk_ext_hdr_len -
793 /* TSQ : try to have two TSO segments in flight */
794 xmit_size_goal = min_t(u32, xmit_size_goal,
795 sysctl_tcp_limit_output_bytes >> 1);
797 xmit_size_goal = tcp_bound_to_half_wnd(tp, xmit_size_goal);
799 /* We try hard to avoid divides here */
800 old_size_goal = tp->xmit_size_goal_segs * mss_now;
802 if (likely(old_size_goal <= xmit_size_goal &&
803 old_size_goal + mss_now > xmit_size_goal)) {
804 xmit_size_goal = old_size_goal;
806 tp->xmit_size_goal_segs =
807 min_t(u16, xmit_size_goal / mss_now,
808 sk->sk_gso_max_segs);
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 *page, int offset,
827 size_t size, 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. One exception is TCP Fast Open
836 * (passive side) where data is allowed to be sent before a connection
837 * is fully established.
839 if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) &&
840 !tcp_passive_fastopen(sk)) {
841 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
845 clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
847 mss_now = tcp_send_mss(sk, &size_goal, flags);
851 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
855 struct sk_buff *skb = tcp_write_queue_tail(sk);
859 if (!tcp_send_head(sk) || (copy = size_goal - skb->len) <= 0) {
861 if (!sk_stream_memory_free(sk))
862 goto wait_for_sndbuf;
864 skb = sk_stream_alloc_skb(sk, 0, sk->sk_allocation);
866 goto wait_for_memory;
875 i = skb_shinfo(skb)->nr_frags;
876 can_coalesce = skb_can_coalesce(skb, i, page, offset);
877 if (!can_coalesce && i >= MAX_SKB_FRAGS) {
878 tcp_mark_push(tp, skb);
881 if (!sk_wmem_schedule(sk, copy))
882 goto wait_for_memory;
885 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
888 skb_fill_page_desc(skb, i, page, offset, copy);
890 skb_shinfo(skb)->tx_flags |= SKBTX_SHARED_FRAG;
893 skb->data_len += copy;
894 skb->truesize += copy;
895 sk->sk_wmem_queued += copy;
896 sk_mem_charge(sk, copy);
897 skb->ip_summed = CHECKSUM_PARTIAL;
898 tp->write_seq += copy;
899 TCP_SKB_CB(skb)->end_seq += copy;
900 skb_shinfo(skb)->gso_segs = 0;
903 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
910 if (skb->len < size_goal || (flags & MSG_OOB))
913 if (forced_push(tp)) {
914 tcp_mark_push(tp, skb);
915 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
916 } else if (skb == tcp_send_head(sk))
917 tcp_push_one(sk, mss_now);
921 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
923 tcp_push(sk, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH);
925 if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
928 mss_now = tcp_send_mss(sk, &size_goal, flags);
932 if (copied && !(flags & MSG_SENDPAGE_NOTLAST))
933 tcp_push(sk, flags, mss_now, tp->nonagle);
940 return sk_stream_error(sk, flags, err);
943 int tcp_sendpage(struct sock *sk, struct page *page, int offset,
944 size_t size, int flags)
948 if (!(sk->sk_route_caps & NETIF_F_SG) ||
949 !(sk->sk_route_caps & NETIF_F_ALL_CSUM))
950 return sock_no_sendpage(sk->sk_socket, page, offset, size,
954 res = do_tcp_sendpages(sk, page, offset, size, flags);
958 EXPORT_SYMBOL(tcp_sendpage);
960 static inline int select_size(const struct sock *sk, bool sg)
962 const struct tcp_sock *tp = tcp_sk(sk);
963 int tmp = tp->mss_cache;
966 if (sk_can_gso(sk)) {
967 /* Small frames wont use a full page:
968 * Payload will immediately follow tcp header.
970 tmp = SKB_WITH_OVERHEAD(2048 - MAX_TCP_HEADER);
972 int pgbreak = SKB_MAX_HEAD(MAX_TCP_HEADER);
974 if (tmp >= pgbreak &&
975 tmp <= pgbreak + (MAX_SKB_FRAGS - 1) * PAGE_SIZE)
983 void tcp_free_fastopen_req(struct tcp_sock *tp)
985 if (tp->fastopen_req != NULL) {
986 kfree(tp->fastopen_req);
987 tp->fastopen_req = NULL;
991 static int tcp_sendmsg_fastopen(struct sock *sk, struct msghdr *msg, int *size)
993 struct tcp_sock *tp = tcp_sk(sk);
996 if (!(sysctl_tcp_fastopen & TFO_CLIENT_ENABLE))
998 if (tp->fastopen_req != NULL)
999 return -EALREADY; /* Another Fast Open is in progress */
1001 tp->fastopen_req = kzalloc(sizeof(struct tcp_fastopen_request),
1003 if (unlikely(tp->fastopen_req == NULL))
1005 tp->fastopen_req->data = msg;
1007 flags = (msg->msg_flags & MSG_DONTWAIT) ? O_NONBLOCK : 0;
1008 err = __inet_stream_connect(sk->sk_socket, msg->msg_name,
1009 msg->msg_namelen, flags);
1010 *size = tp->fastopen_req->copied;
1011 tcp_free_fastopen_req(tp);
1015 int tcp_sendmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg,
1019 struct tcp_sock *tp = tcp_sk(sk);
1020 struct sk_buff *skb;
1021 int iovlen, flags, err, copied = 0;
1022 int mss_now = 0, size_goal, copied_syn = 0, offset = 0;
1028 flags = msg->msg_flags;
1029 if (flags & MSG_FASTOPEN) {
1030 err = tcp_sendmsg_fastopen(sk, msg, &copied_syn);
1031 if (err == -EINPROGRESS && copied_syn > 0)
1035 offset = copied_syn;
1038 timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
1040 /* Wait for a connection to finish. One exception is TCP Fast Open
1041 * (passive side) where data is allowed to be sent before a connection
1042 * is fully established.
1044 if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) &&
1045 !tcp_passive_fastopen(sk)) {
1046 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
1050 if (unlikely(tp->repair)) {
1051 if (tp->repair_queue == TCP_RECV_QUEUE) {
1052 copied = tcp_send_rcvq(sk, msg, size);
1057 if (tp->repair_queue == TCP_NO_QUEUE)
1060 /* 'common' sending to sendq */
1063 /* This should be in poll */
1064 clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
1066 mss_now = tcp_send_mss(sk, &size_goal, flags);
1068 /* Ok commence sending. */
1069 iovlen = msg->msg_iovlen;
1074 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
1077 sg = !!(sk->sk_route_caps & NETIF_F_SG);
1079 while (--iovlen >= 0) {
1080 size_t seglen = iov->iov_len;
1081 unsigned char __user *from = iov->iov_base;
1084 if (unlikely(offset > 0)) { /* Skip bytes copied in SYN */
1085 if (offset >= seglen) {
1094 while (seglen > 0) {
1096 int max = size_goal;
1098 skb = tcp_write_queue_tail(sk);
1099 if (tcp_send_head(sk)) {
1100 if (skb->ip_summed == CHECKSUM_NONE)
1102 copy = max - skb->len;
1107 /* Allocate new segment. If the interface is SG,
1108 * allocate skb fitting to single page.
1110 if (!sk_stream_memory_free(sk))
1111 goto wait_for_sndbuf;
1113 skb = sk_stream_alloc_skb(sk,
1114 select_size(sk, sg),
1117 goto wait_for_memory;
1120 * All packets are restored as if they have
1121 * already been sent.
1124 TCP_SKB_CB(skb)->when = tcp_time_stamp;
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_issue_pending(tp->ucopy.dma_chan);
1412 if (dma_async_is_tx_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 struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off)
1433 struct sk_buff *skb;
1436 while ((skb = skb_peek(&sk->sk_receive_queue)) != NULL) {
1437 offset = seq - TCP_SKB_CB(skb)->seq;
1438 if (tcp_hdr(skb)->syn)
1440 if (offset < skb->len || tcp_hdr(skb)->fin) {
1444 /* This looks weird, but this can happen if TCP collapsing
1445 * splitted a fat GRO packet, while we released socket lock
1446 * in skb_splice_bits()
1448 sk_eat_skb(sk, skb, false);
1454 * This routine provides an alternative to tcp_recvmsg() for routines
1455 * that would like to handle copying from skbuffs directly in 'sendfile'
1458 * - It is assumed that the socket was locked by the caller.
1459 * - The routine does not block.
1460 * - At present, there is no support for reading OOB data
1461 * or for 'peeking' the socket using this routine
1462 * (although both would be easy to implement).
1464 int tcp_read_sock(struct sock *sk, read_descriptor_t *desc,
1465 sk_read_actor_t recv_actor)
1467 struct sk_buff *skb;
1468 struct tcp_sock *tp = tcp_sk(sk);
1469 u32 seq = tp->copied_seq;
1473 if (sk->sk_state == TCP_LISTEN)
1475 while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1476 if (offset < skb->len) {
1480 len = skb->len - offset;
1481 /* Stop reading if we hit a patch of urgent data */
1483 u32 urg_offset = tp->urg_seq - seq;
1484 if (urg_offset < len)
1489 used = recv_actor(desc, skb, offset, len);
1494 } else if (used <= len) {
1499 /* If recv_actor drops the lock (e.g. TCP splice
1500 * receive) the skb pointer might be invalid when
1501 * getting here: tcp_collapse might have deleted it
1502 * while aggregating skbs from the socket queue.
1504 skb = tcp_recv_skb(sk, seq - 1, &offset);
1507 /* TCP coalescing might have appended data to the skb.
1508 * Try to splice more frags
1510 if (offset + 1 != skb->len)
1513 if (tcp_hdr(skb)->fin) {
1514 sk_eat_skb(sk, skb, false);
1518 sk_eat_skb(sk, skb, false);
1521 tp->copied_seq = seq;
1523 tp->copied_seq = seq;
1525 tcp_rcv_space_adjust(sk);
1527 /* Clean up data we have read: This will do ACK frames. */
1529 tcp_recv_skb(sk, seq, &offset);
1530 tcp_cleanup_rbuf(sk, copied);
1534 EXPORT_SYMBOL(tcp_read_sock);
1537 * This routine copies from a sock struct into the user buffer.
1539 * Technical note: in 2.3 we work on _locked_ socket, so that
1540 * tricks with *seq access order and skb->users are not required.
1541 * Probably, code can be easily improved even more.
1544 int tcp_recvmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg,
1545 size_t len, int nonblock, int flags, int *addr_len)
1547 struct tcp_sock *tp = tcp_sk(sk);
1553 int target; /* Read at least this many bytes */
1555 struct task_struct *user_recv = NULL;
1556 bool copied_early = false;
1557 struct sk_buff *skb;
1560 if (sk_can_busy_loop(sk) && skb_queue_empty(&sk->sk_receive_queue) &&
1561 (sk->sk_state == TCP_ESTABLISHED))
1562 sk_busy_loop(sk, nonblock);
1567 if (sk->sk_state == TCP_LISTEN)
1570 timeo = sock_rcvtimeo(sk, nonblock);
1572 /* Urgent data needs to be handled specially. */
1573 if (flags & MSG_OOB)
1576 if (unlikely(tp->repair)) {
1578 if (!(flags & MSG_PEEK))
1581 if (tp->repair_queue == TCP_SEND_QUEUE)
1585 if (tp->repair_queue == TCP_NO_QUEUE)
1588 /* 'common' recv queue MSG_PEEK-ing */
1591 seq = &tp->copied_seq;
1592 if (flags & MSG_PEEK) {
1593 peek_seq = tp->copied_seq;
1597 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
1599 #ifdef CONFIG_NET_DMA
1600 tp->ucopy.dma_chan = NULL;
1602 skb = skb_peek_tail(&sk->sk_receive_queue);
1607 available = TCP_SKB_CB(skb)->seq + skb->len - (*seq);
1608 if ((available < target) &&
1609 (len > sysctl_tcp_dma_copybreak) && !(flags & MSG_PEEK) &&
1610 !sysctl_tcp_low_latency &&
1611 net_dma_find_channel()) {
1612 preempt_enable_no_resched();
1613 tp->ucopy.pinned_list =
1614 dma_pin_iovec_pages(msg->msg_iov, len);
1616 preempt_enable_no_resched();
1624 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
1625 if (tp->urg_data && tp->urg_seq == *seq) {
1628 if (signal_pending(current)) {
1629 copied = timeo ? sock_intr_errno(timeo) : -EAGAIN;
1634 /* Next get a buffer. */
1636 skb_queue_walk(&sk->sk_receive_queue, skb) {
1637 /* Now that we have two receive queues this
1640 if (WARN(before(*seq, TCP_SKB_CB(skb)->seq),
1641 "recvmsg bug: copied %X seq %X rcvnxt %X fl %X\n",
1642 *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt,
1646 offset = *seq - TCP_SKB_CB(skb)->seq;
1647 if (tcp_hdr(skb)->syn)
1649 if (offset < skb->len)
1651 if (tcp_hdr(skb)->fin)
1653 WARN(!(flags & MSG_PEEK),
1654 "recvmsg bug 2: copied %X seq %X rcvnxt %X fl %X\n",
1655 *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt, flags);
1658 /* Well, if we have backlog, try to process it now yet. */
1660 if (copied >= target && !sk->sk_backlog.tail)
1665 sk->sk_state == TCP_CLOSE ||
1666 (sk->sk_shutdown & RCV_SHUTDOWN) ||
1668 signal_pending(current))
1671 if (sock_flag(sk, SOCK_DONE))
1675 copied = sock_error(sk);
1679 if (sk->sk_shutdown & RCV_SHUTDOWN)
1682 if (sk->sk_state == TCP_CLOSE) {
1683 if (!sock_flag(sk, SOCK_DONE)) {
1684 /* This occurs when user tries to read
1685 * from never connected socket.
1698 if (signal_pending(current)) {
1699 copied = sock_intr_errno(timeo);
1704 tcp_cleanup_rbuf(sk, copied);
1706 if (!sysctl_tcp_low_latency && tp->ucopy.task == user_recv) {
1707 /* Install new reader */
1708 if (!user_recv && !(flags & (MSG_TRUNC | MSG_PEEK))) {
1709 user_recv = current;
1710 tp->ucopy.task = user_recv;
1711 tp->ucopy.iov = msg->msg_iov;
1714 tp->ucopy.len = len;
1716 WARN_ON(tp->copied_seq != tp->rcv_nxt &&
1717 !(flags & (MSG_PEEK | MSG_TRUNC)));
1719 /* Ugly... If prequeue is not empty, we have to
1720 * process it before releasing socket, otherwise
1721 * order will be broken at second iteration.
1722 * More elegant solution is required!!!
1724 * Look: we have the following (pseudo)queues:
1726 * 1. packets in flight
1731 * Each queue can be processed only if the next ones
1732 * are empty. At this point we have empty receive_queue.
1733 * But prequeue _can_ be not empty after 2nd iteration,
1734 * when we jumped to start of loop because backlog
1735 * processing added something to receive_queue.
1736 * We cannot release_sock(), because backlog contains
1737 * packets arrived _after_ prequeued ones.
1739 * Shortly, algorithm is clear --- to process all
1740 * the queues in order. We could make it more directly,
1741 * requeueing packets from backlog to prequeue, if
1742 * is not empty. It is more elegant, but eats cycles,
1745 if (!skb_queue_empty(&tp->ucopy.prequeue))
1748 /* __ Set realtime policy in scheduler __ */
1751 #ifdef CONFIG_NET_DMA
1752 if (tp->ucopy.dma_chan) {
1753 if (tp->rcv_wnd == 0 &&
1754 !skb_queue_empty(&sk->sk_async_wait_queue)) {
1755 tcp_service_net_dma(sk, true);
1756 tcp_cleanup_rbuf(sk, copied);
1758 dma_async_issue_pending(tp->ucopy.dma_chan);
1761 if (copied >= target) {
1762 /* Do not sleep, just process backlog. */
1766 sk_wait_data(sk, &timeo);
1768 #ifdef CONFIG_NET_DMA
1769 tcp_service_net_dma(sk, false); /* Don't block */
1770 tp->ucopy.wakeup = 0;
1776 /* __ Restore normal policy in scheduler __ */
1778 if ((chunk = len - tp->ucopy.len) != 0) {
1779 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMBACKLOG, chunk);
1784 if (tp->rcv_nxt == tp->copied_seq &&
1785 !skb_queue_empty(&tp->ucopy.prequeue)) {
1787 tcp_prequeue_process(sk);
1789 if ((chunk = len - tp->ucopy.len) != 0) {
1790 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1796 if ((flags & MSG_PEEK) &&
1797 (peek_seq - copied - urg_hole != tp->copied_seq)) {
1798 net_dbg_ratelimited("TCP(%s:%d): Application bug, race in MSG_PEEK\n",
1800 task_pid_nr(current));
1801 peek_seq = tp->copied_seq;
1806 /* Ok so how much can we use? */
1807 used = skb->len - offset;
1811 /* Do we have urgent data here? */
1813 u32 urg_offset = tp->urg_seq - *seq;
1814 if (urg_offset < used) {
1816 if (!sock_flag(sk, SOCK_URGINLINE)) {
1829 if (!(flags & MSG_TRUNC)) {
1830 #ifdef CONFIG_NET_DMA
1831 if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list)
1832 tp->ucopy.dma_chan = net_dma_find_channel();
1834 if (tp->ucopy.dma_chan) {
1835 tp->ucopy.dma_cookie = dma_skb_copy_datagram_iovec(
1836 tp->ucopy.dma_chan, skb, offset,
1838 tp->ucopy.pinned_list);
1840 if (tp->ucopy.dma_cookie < 0) {
1842 pr_alert("%s: dma_cookie < 0\n",
1845 /* Exception. Bailout! */
1851 dma_async_issue_pending(tp->ucopy.dma_chan);
1853 if ((offset + used) == skb->len)
1854 copied_early = true;
1859 err = skb_copy_datagram_iovec(skb, offset,
1860 msg->msg_iov, used);
1862 /* Exception. Bailout! */
1874 tcp_rcv_space_adjust(sk);
1877 if (tp->urg_data && after(tp->copied_seq, tp->urg_seq)) {
1879 tcp_fast_path_check(sk);
1881 if (used + offset < skb->len)
1884 if (tcp_hdr(skb)->fin)
1886 if (!(flags & MSG_PEEK)) {
1887 sk_eat_skb(sk, skb, copied_early);
1888 copied_early = false;
1893 /* Process the FIN. */
1895 if (!(flags & MSG_PEEK)) {
1896 sk_eat_skb(sk, skb, copied_early);
1897 copied_early = false;
1903 if (!skb_queue_empty(&tp->ucopy.prequeue)) {
1906 tp->ucopy.len = copied > 0 ? len : 0;
1908 tcp_prequeue_process(sk);
1910 if (copied > 0 && (chunk = len - tp->ucopy.len) != 0) {
1911 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1917 tp->ucopy.task = NULL;
1921 #ifdef CONFIG_NET_DMA
1922 tcp_service_net_dma(sk, true); /* Wait for queue to drain */
1923 tp->ucopy.dma_chan = NULL;
1925 if (tp->ucopy.pinned_list) {
1926 dma_unpin_iovec_pages(tp->ucopy.pinned_list);
1927 tp->ucopy.pinned_list = NULL;
1931 /* According to UNIX98, msg_name/msg_namelen are ignored
1932 * on connected socket. I was just happy when found this 8) --ANK
1935 /* Clean up data we have read: This will do ACK frames. */
1936 tcp_cleanup_rbuf(sk, copied);
1946 err = tcp_recv_urg(sk, msg, len, flags);
1950 err = tcp_peek_sndq(sk, msg, len);
1953 EXPORT_SYMBOL(tcp_recvmsg);
1955 void tcp_set_state(struct sock *sk, int state)
1957 int oldstate = sk->sk_state;
1960 case TCP_ESTABLISHED:
1961 if (oldstate != TCP_ESTABLISHED)
1962 TCP_INC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
1966 if (oldstate == TCP_CLOSE_WAIT || oldstate == TCP_ESTABLISHED)
1967 TCP_INC_STATS(sock_net(sk), TCP_MIB_ESTABRESETS);
1969 sk->sk_prot->unhash(sk);
1970 if (inet_csk(sk)->icsk_bind_hash &&
1971 !(sk->sk_userlocks & SOCK_BINDPORT_LOCK))
1975 if (oldstate == TCP_ESTABLISHED)
1976 TCP_DEC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
1979 /* Change state AFTER socket is unhashed to avoid closed
1980 * socket sitting in hash tables.
1982 sk->sk_state = state;
1985 SOCK_DEBUG(sk, "TCP sk=%p, State %s -> %s\n", sk, statename[oldstate], statename[state]);
1988 EXPORT_SYMBOL_GPL(tcp_set_state);
1991 * State processing on a close. This implements the state shift for
1992 * sending our FIN frame. Note that we only send a FIN for some
1993 * states. A shutdown() may have already sent the FIN, or we may be
1997 static const unsigned char new_state[16] = {
1998 /* current state: new state: action: */
1999 /* (Invalid) */ TCP_CLOSE,
2000 /* TCP_ESTABLISHED */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2001 /* TCP_SYN_SENT */ TCP_CLOSE,
2002 /* TCP_SYN_RECV */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2003 /* TCP_FIN_WAIT1 */ TCP_FIN_WAIT1,
2004 /* TCP_FIN_WAIT2 */ TCP_FIN_WAIT2,
2005 /* TCP_TIME_WAIT */ TCP_CLOSE,
2006 /* TCP_CLOSE */ TCP_CLOSE,
2007 /* TCP_CLOSE_WAIT */ TCP_LAST_ACK | TCP_ACTION_FIN,
2008 /* TCP_LAST_ACK */ TCP_LAST_ACK,
2009 /* TCP_LISTEN */ TCP_CLOSE,
2010 /* TCP_CLOSING */ TCP_CLOSING,
2013 static int tcp_close_state(struct sock *sk)
2015 int next = (int)new_state[sk->sk_state];
2016 int ns = next & TCP_STATE_MASK;
2018 tcp_set_state(sk, ns);
2020 return next & TCP_ACTION_FIN;
2024 * Shutdown the sending side of a connection. Much like close except
2025 * that we don't receive shut down or sock_set_flag(sk, SOCK_DEAD).
2028 void tcp_shutdown(struct sock *sk, int how)
2030 /* We need to grab some memory, and put together a FIN,
2031 * and then put it into the queue to be sent.
2032 * Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
2034 if (!(how & SEND_SHUTDOWN))
2037 /* If we've already sent a FIN, or it's a closed state, skip this. */
2038 if ((1 << sk->sk_state) &
2039 (TCPF_ESTABLISHED | TCPF_SYN_SENT |
2040 TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) {
2041 /* Clear out any half completed packets. FIN if needed. */
2042 if (tcp_close_state(sk))
2046 EXPORT_SYMBOL(tcp_shutdown);
2048 bool tcp_check_oom(struct sock *sk, int shift)
2050 bool too_many_orphans, out_of_socket_memory;
2052 too_many_orphans = tcp_too_many_orphans(sk, shift);
2053 out_of_socket_memory = tcp_out_of_memory(sk);
2055 if (too_many_orphans)
2056 net_info_ratelimited("too many orphaned sockets\n");
2057 if (out_of_socket_memory)
2058 net_info_ratelimited("out of memory -- consider tuning tcp_mem\n");
2059 return too_many_orphans || out_of_socket_memory;
2062 void tcp_close(struct sock *sk, long timeout)
2064 struct sk_buff *skb;
2065 int data_was_unread = 0;
2069 sk->sk_shutdown = SHUTDOWN_MASK;
2071 if (sk->sk_state == TCP_LISTEN) {
2072 tcp_set_state(sk, TCP_CLOSE);
2075 inet_csk_listen_stop(sk);
2077 goto adjudge_to_death;
2080 /* We need to flush the recv. buffs. We do this only on the
2081 * descriptor close, not protocol-sourced closes, because the
2082 * reader process may not have drained the data yet!
2084 while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) {
2085 u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq -
2087 data_was_unread += len;
2093 /* If socket has been already reset (e.g. in tcp_reset()) - kill it. */
2094 if (sk->sk_state == TCP_CLOSE)
2095 goto adjudge_to_death;
2097 /* As outlined in RFC 2525, section 2.17, we send a RST here because
2098 * data was lost. To witness the awful effects of the old behavior of
2099 * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk
2100 * GET in an FTP client, suspend the process, wait for the client to
2101 * advertise a zero window, then kill -9 the FTP client, wheee...
2102 * Note: timeout is always zero in such a case.
2104 if (unlikely(tcp_sk(sk)->repair)) {
2105 sk->sk_prot->disconnect(sk, 0);
2106 } else if (data_was_unread) {
2107 /* Unread data was tossed, zap the connection. */
2108 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONCLOSE);
2109 tcp_set_state(sk, TCP_CLOSE);
2110 tcp_send_active_reset(sk, sk->sk_allocation);
2111 } else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) {
2112 /* Check zero linger _after_ checking for unread data. */
2113 sk->sk_prot->disconnect(sk, 0);
2114 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
2115 } else if (tcp_close_state(sk)) {
2116 /* We FIN if the application ate all the data before
2117 * zapping the connection.
2120 /* RED-PEN. Formally speaking, we have broken TCP state
2121 * machine. State transitions:
2123 * TCP_ESTABLISHED -> TCP_FIN_WAIT1
2124 * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible)
2125 * TCP_CLOSE_WAIT -> TCP_LAST_ACK
2127 * are legal only when FIN has been sent (i.e. in window),
2128 * rather than queued out of window. Purists blame.
2130 * F.e. "RFC state" is ESTABLISHED,
2131 * if Linux state is FIN-WAIT-1, but FIN is still not sent.
2133 * The visible declinations are that sometimes
2134 * we enter time-wait state, when it is not required really
2135 * (harmless), do not send active resets, when they are
2136 * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
2137 * they look as CLOSING or LAST_ACK for Linux)
2138 * Probably, I missed some more holelets.
2140 * XXX (TFO) - To start off we don't support SYN+ACK+FIN
2141 * in a single packet! (May consider it later but will
2142 * probably need API support or TCP_CORK SYN-ACK until
2143 * data is written and socket is closed.)
2148 sk_stream_wait_close(sk, timeout);
2151 state = sk->sk_state;
2155 /* It is the last release_sock in its life. It will remove backlog. */
2159 /* Now socket is owned by kernel and we acquire BH lock
2160 to finish close. No need to check for user refs.
2164 WARN_ON(sock_owned_by_user(sk));
2166 percpu_counter_inc(sk->sk_prot->orphan_count);
2168 /* Have we already been destroyed by a softirq or backlog? */
2169 if (state != TCP_CLOSE && sk->sk_state == TCP_CLOSE)
2172 /* This is a (useful) BSD violating of the RFC. There is a
2173 * problem with TCP as specified in that the other end could
2174 * keep a socket open forever with no application left this end.
2175 * We use a 3 minute timeout (about the same as BSD) then kill
2176 * our end. If they send after that then tough - BUT: long enough
2177 * that we won't make the old 4*rto = almost no time - whoops
2180 * Nope, it was not mistake. It is really desired behaviour
2181 * f.e. on http servers, when such sockets are useless, but
2182 * consume significant resources. Let's do it with special
2183 * linger2 option. --ANK
2186 if (sk->sk_state == TCP_FIN_WAIT2) {
2187 struct tcp_sock *tp = tcp_sk(sk);
2188 if (tp->linger2 < 0) {
2189 tcp_set_state(sk, TCP_CLOSE);
2190 tcp_send_active_reset(sk, GFP_ATOMIC);
2191 NET_INC_STATS_BH(sock_net(sk),
2192 LINUX_MIB_TCPABORTONLINGER);
2194 const int tmo = tcp_fin_time(sk);
2196 if (tmo > TCP_TIMEWAIT_LEN) {
2197 inet_csk_reset_keepalive_timer(sk,
2198 tmo - TCP_TIMEWAIT_LEN);
2200 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
2205 if (sk->sk_state != TCP_CLOSE) {
2207 if (tcp_check_oom(sk, 0)) {
2208 tcp_set_state(sk, TCP_CLOSE);
2209 tcp_send_active_reset(sk, GFP_ATOMIC);
2210 NET_INC_STATS_BH(sock_net(sk),
2211 LINUX_MIB_TCPABORTONMEMORY);
2215 if (sk->sk_state == TCP_CLOSE) {
2216 struct request_sock *req = tcp_sk(sk)->fastopen_rsk;
2217 /* We could get here with a non-NULL req if the socket is
2218 * aborted (e.g., closed with unread data) before 3WHS
2222 reqsk_fastopen_remove(sk, req, false);
2223 inet_csk_destroy_sock(sk);
2225 /* Otherwise, socket is reprieved until protocol close. */
2232 EXPORT_SYMBOL(tcp_close);
2234 /* These states need RST on ABORT according to RFC793 */
2236 static inline bool tcp_need_reset(int state)
2238 return (1 << state) &
2239 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 |
2240 TCPF_FIN_WAIT2 | TCPF_SYN_RECV);
2243 int tcp_disconnect(struct sock *sk, int flags)
2245 struct inet_sock *inet = inet_sk(sk);
2246 struct inet_connection_sock *icsk = inet_csk(sk);
2247 struct tcp_sock *tp = tcp_sk(sk);
2249 int old_state = sk->sk_state;
2251 if (old_state != TCP_CLOSE)
2252 tcp_set_state(sk, TCP_CLOSE);
2254 /* ABORT function of RFC793 */
2255 if (old_state == TCP_LISTEN) {
2256 inet_csk_listen_stop(sk);
2257 } else if (unlikely(tp->repair)) {
2258 sk->sk_err = ECONNABORTED;
2259 } else if (tcp_need_reset(old_state) ||
2260 (tp->snd_nxt != tp->write_seq &&
2261 (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK))) {
2262 /* The last check adjusts for discrepancy of Linux wrt. RFC
2265 tcp_send_active_reset(sk, gfp_any());
2266 sk->sk_err = ECONNRESET;
2267 } else if (old_state == TCP_SYN_SENT)
2268 sk->sk_err = ECONNRESET;
2270 tcp_clear_xmit_timers(sk);
2271 __skb_queue_purge(&sk->sk_receive_queue);
2272 tcp_write_queue_purge(sk);
2273 __skb_queue_purge(&tp->out_of_order_queue);
2274 #ifdef CONFIG_NET_DMA
2275 __skb_queue_purge(&sk->sk_async_wait_queue);
2278 inet->inet_dport = 0;
2280 if (!(sk->sk_userlocks & SOCK_BINDADDR_LOCK))
2281 inet_reset_saddr(sk);
2283 sk->sk_shutdown = 0;
2284 sock_reset_flag(sk, SOCK_DONE);
2286 if ((tp->write_seq += tp->max_window + 2) == 0)
2288 icsk->icsk_backoff = 0;
2290 icsk->icsk_probes_out = 0;
2291 tp->packets_out = 0;
2292 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
2293 tp->snd_cwnd_cnt = 0;
2294 tp->window_clamp = 0;
2295 tcp_set_ca_state(sk, TCP_CA_Open);
2296 tcp_clear_retrans(tp);
2297 inet_csk_delack_init(sk);
2298 tcp_init_send_head(sk);
2299 memset(&tp->rx_opt, 0, sizeof(tp->rx_opt));
2302 WARN_ON(inet->inet_num && !icsk->icsk_bind_hash);
2304 sk->sk_error_report(sk);
2307 EXPORT_SYMBOL(tcp_disconnect);
2309 void tcp_sock_destruct(struct sock *sk)
2311 inet_sock_destruct(sk);
2313 kfree(inet_csk(sk)->icsk_accept_queue.fastopenq);
2316 static inline bool tcp_can_repair_sock(const struct sock *sk)
2318 return ns_capable(sock_net(sk)->user_ns, CAP_NET_ADMIN) &&
2319 ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_ESTABLISHED));
2322 static int tcp_repair_options_est(struct tcp_sock *tp,
2323 struct tcp_repair_opt __user *optbuf, unsigned int len)
2325 struct tcp_repair_opt opt;
2327 while (len >= sizeof(opt)) {
2328 if (copy_from_user(&opt, optbuf, sizeof(opt)))
2334 switch (opt.opt_code) {
2336 tp->rx_opt.mss_clamp = opt.opt_val;
2340 u16 snd_wscale = opt.opt_val & 0xFFFF;
2341 u16 rcv_wscale = opt.opt_val >> 16;
2343 if (snd_wscale > 14 || rcv_wscale > 14)
2346 tp->rx_opt.snd_wscale = snd_wscale;
2347 tp->rx_opt.rcv_wscale = rcv_wscale;
2348 tp->rx_opt.wscale_ok = 1;
2351 case TCPOPT_SACK_PERM:
2352 if (opt.opt_val != 0)
2355 tp->rx_opt.sack_ok |= TCP_SACK_SEEN;
2356 if (sysctl_tcp_fack)
2357 tcp_enable_fack(tp);
2359 case TCPOPT_TIMESTAMP:
2360 if (opt.opt_val != 0)
2363 tp->rx_opt.tstamp_ok = 1;
2372 * Socket option code for TCP.
2374 static int do_tcp_setsockopt(struct sock *sk, int level,
2375 int optname, char __user *optval, unsigned int optlen)
2377 struct tcp_sock *tp = tcp_sk(sk);
2378 struct inet_connection_sock *icsk = inet_csk(sk);
2382 /* These are data/string values, all the others are ints */
2384 case TCP_CONGESTION: {
2385 char name[TCP_CA_NAME_MAX];
2390 val = strncpy_from_user(name, optval,
2391 min_t(long, TCP_CA_NAME_MAX-1, optlen));
2397 err = tcp_set_congestion_control(sk, name);
2406 if (optlen < sizeof(int))
2409 if (get_user(val, (int __user *)optval))
2416 /* Values greater than interface MTU won't take effect. However
2417 * at the point when this call is done we typically don't yet
2418 * know which interface is going to be used */
2419 if (val < TCP_MIN_MSS || val > MAX_TCP_WINDOW) {
2423 tp->rx_opt.user_mss = val;
2428 /* TCP_NODELAY is weaker than TCP_CORK, so that
2429 * this option on corked socket is remembered, but
2430 * it is not activated until cork is cleared.
2432 * However, when TCP_NODELAY is set we make
2433 * an explicit push, which overrides even TCP_CORK
2434 * for currently queued segments.
2436 tp->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH;
2437 tcp_push_pending_frames(sk);
2439 tp->nonagle &= ~TCP_NAGLE_OFF;
2443 case TCP_THIN_LINEAR_TIMEOUTS:
2444 if (val < 0 || val > 1)
2450 case TCP_THIN_DUPACK:
2451 if (val < 0 || val > 1)
2454 tp->thin_dupack = val;
2455 if (tp->thin_dupack)
2456 tcp_disable_early_retrans(tp);
2460 if (!tcp_can_repair_sock(sk))
2462 else if (val == 1) {
2464 sk->sk_reuse = SK_FORCE_REUSE;
2465 tp->repair_queue = TCP_NO_QUEUE;
2466 } else if (val == 0) {
2468 sk->sk_reuse = SK_NO_REUSE;
2469 tcp_send_window_probe(sk);
2475 case TCP_REPAIR_QUEUE:
2478 else if (val < TCP_QUEUES_NR)
2479 tp->repair_queue = val;
2485 if (sk->sk_state != TCP_CLOSE)
2487 else if (tp->repair_queue == TCP_SEND_QUEUE)
2488 tp->write_seq = val;
2489 else if (tp->repair_queue == TCP_RECV_QUEUE)
2495 case TCP_REPAIR_OPTIONS:
2498 else if (sk->sk_state == TCP_ESTABLISHED)
2499 err = tcp_repair_options_est(tp,
2500 (struct tcp_repair_opt __user *)optval,
2507 /* When set indicates to always queue non-full frames.
2508 * Later the user clears this option and we transmit
2509 * any pending partial frames in the queue. This is
2510 * meant to be used alongside sendfile() to get properly
2511 * filled frames when the user (for example) must write
2512 * out headers with a write() call first and then use
2513 * sendfile to send out the data parts.
2515 * TCP_CORK can be set together with TCP_NODELAY and it is
2516 * stronger than TCP_NODELAY.
2519 tp->nonagle |= TCP_NAGLE_CORK;
2521 tp->nonagle &= ~TCP_NAGLE_CORK;
2522 if (tp->nonagle&TCP_NAGLE_OFF)
2523 tp->nonagle |= TCP_NAGLE_PUSH;
2524 tcp_push_pending_frames(sk);
2529 if (val < 1 || val > MAX_TCP_KEEPIDLE)
2532 tp->keepalive_time = val * HZ;
2533 if (sock_flag(sk, SOCK_KEEPOPEN) &&
2534 !((1 << sk->sk_state) &
2535 (TCPF_CLOSE | TCPF_LISTEN))) {
2536 u32 elapsed = keepalive_time_elapsed(tp);
2537 if (tp->keepalive_time > elapsed)
2538 elapsed = tp->keepalive_time - elapsed;
2541 inet_csk_reset_keepalive_timer(sk, elapsed);
2546 if (val < 1 || val > MAX_TCP_KEEPINTVL)
2549 tp->keepalive_intvl = val * HZ;
2552 if (val < 1 || val > MAX_TCP_KEEPCNT)
2555 tp->keepalive_probes = val;
2558 if (val < 1 || val > MAX_TCP_SYNCNT)
2561 icsk->icsk_syn_retries = val;
2567 else if (val > sysctl_tcp_fin_timeout / HZ)
2570 tp->linger2 = val * HZ;
2573 case TCP_DEFER_ACCEPT:
2574 /* Translate value in seconds to number of retransmits */
2575 icsk->icsk_accept_queue.rskq_defer_accept =
2576 secs_to_retrans(val, TCP_TIMEOUT_INIT / HZ,
2580 case TCP_WINDOW_CLAMP:
2582 if (sk->sk_state != TCP_CLOSE) {
2586 tp->window_clamp = 0;
2588 tp->window_clamp = val < SOCK_MIN_RCVBUF / 2 ?
2589 SOCK_MIN_RCVBUF / 2 : val;
2594 icsk->icsk_ack.pingpong = 1;
2596 icsk->icsk_ack.pingpong = 0;
2597 if ((1 << sk->sk_state) &
2598 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) &&
2599 inet_csk_ack_scheduled(sk)) {
2600 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
2601 tcp_cleanup_rbuf(sk, 1);
2603 icsk->icsk_ack.pingpong = 1;
2608 #ifdef CONFIG_TCP_MD5SIG
2610 /* Read the IP->Key mappings from userspace */
2611 err = tp->af_specific->md5_parse(sk, optval, optlen);
2614 case TCP_USER_TIMEOUT:
2615 /* Cap the max timeout in ms TCP will retry/retrans
2616 * before giving up and aborting (ETIMEDOUT) a connection.
2621 icsk->icsk_user_timeout = msecs_to_jiffies(val);
2625 if (val >= 0 && ((1 << sk->sk_state) & (TCPF_CLOSE |
2627 err = fastopen_init_queue(sk, val);
2635 tp->tsoffset = val - tcp_time_stamp;
2637 case TCP_NOTSENT_LOWAT:
2638 tp->notsent_lowat = val;
2639 sk->sk_write_space(sk);
2650 int tcp_setsockopt(struct sock *sk, int level, int optname, char __user *optval,
2651 unsigned int optlen)
2653 const struct inet_connection_sock *icsk = inet_csk(sk);
2655 if (level != SOL_TCP)
2656 return icsk->icsk_af_ops->setsockopt(sk, level, optname,
2658 return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2660 EXPORT_SYMBOL(tcp_setsockopt);
2662 #ifdef CONFIG_COMPAT
2663 int compat_tcp_setsockopt(struct sock *sk, int level, int optname,
2664 char __user *optval, unsigned int optlen)
2666 if (level != SOL_TCP)
2667 return inet_csk_compat_setsockopt(sk, level, optname,
2669 return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2671 EXPORT_SYMBOL(compat_tcp_setsockopt);
2674 /* Return information about state of tcp endpoint in API format. */
2675 void tcp_get_info(const struct sock *sk, struct tcp_info *info)
2677 const struct tcp_sock *tp = tcp_sk(sk);
2678 const struct inet_connection_sock *icsk = inet_csk(sk);
2679 u32 now = tcp_time_stamp;
2681 memset(info, 0, sizeof(*info));
2683 info->tcpi_state = sk->sk_state;
2684 info->tcpi_ca_state = icsk->icsk_ca_state;
2685 info->tcpi_retransmits = icsk->icsk_retransmits;
2686 info->tcpi_probes = icsk->icsk_probes_out;
2687 info->tcpi_backoff = icsk->icsk_backoff;
2689 if (tp->rx_opt.tstamp_ok)
2690 info->tcpi_options |= TCPI_OPT_TIMESTAMPS;
2691 if (tcp_is_sack(tp))
2692 info->tcpi_options |= TCPI_OPT_SACK;
2693 if (tp->rx_opt.wscale_ok) {
2694 info->tcpi_options |= TCPI_OPT_WSCALE;
2695 info->tcpi_snd_wscale = tp->rx_opt.snd_wscale;
2696 info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale;
2699 if (tp->ecn_flags & TCP_ECN_OK)
2700 info->tcpi_options |= TCPI_OPT_ECN;
2701 if (tp->ecn_flags & TCP_ECN_SEEN)
2702 info->tcpi_options |= TCPI_OPT_ECN_SEEN;
2703 if (tp->syn_data_acked)
2704 info->tcpi_options |= TCPI_OPT_SYN_DATA;
2706 info->tcpi_rto = jiffies_to_usecs(icsk->icsk_rto);
2707 info->tcpi_ato = jiffies_to_usecs(icsk->icsk_ack.ato);
2708 info->tcpi_snd_mss = tp->mss_cache;
2709 info->tcpi_rcv_mss = icsk->icsk_ack.rcv_mss;
2711 if (sk->sk_state == TCP_LISTEN) {
2712 info->tcpi_unacked = sk->sk_ack_backlog;
2713 info->tcpi_sacked = sk->sk_max_ack_backlog;
2715 info->tcpi_unacked = tp->packets_out;
2716 info->tcpi_sacked = tp->sacked_out;
2718 info->tcpi_lost = tp->lost_out;
2719 info->tcpi_retrans = tp->retrans_out;
2720 info->tcpi_fackets = tp->fackets_out;
2722 info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime);
2723 info->tcpi_last_data_recv = jiffies_to_msecs(now - icsk->icsk_ack.lrcvtime);
2724 info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp);
2726 info->tcpi_pmtu = icsk->icsk_pmtu_cookie;
2727 info->tcpi_rcv_ssthresh = tp->rcv_ssthresh;
2728 info->tcpi_rtt = jiffies_to_usecs(tp->srtt)>>3;
2729 info->tcpi_rttvar = jiffies_to_usecs(tp->mdev)>>2;
2730 info->tcpi_snd_ssthresh = tp->snd_ssthresh;
2731 info->tcpi_snd_cwnd = tp->snd_cwnd;
2732 info->tcpi_advmss = tp->advmss;
2733 info->tcpi_reordering = tp->reordering;
2735 info->tcpi_rcv_rtt = jiffies_to_usecs(tp->rcv_rtt_est.rtt)>>3;
2736 info->tcpi_rcv_space = tp->rcvq_space.space;
2738 info->tcpi_total_retrans = tp->total_retrans;
2740 EXPORT_SYMBOL_GPL(tcp_get_info);
2742 static int do_tcp_getsockopt(struct sock *sk, int level,
2743 int optname, char __user *optval, int __user *optlen)
2745 struct inet_connection_sock *icsk = inet_csk(sk);
2746 struct tcp_sock *tp = tcp_sk(sk);
2749 if (get_user(len, optlen))
2752 len = min_t(unsigned int, len, sizeof(int));
2759 val = tp->mss_cache;
2760 if (!val && ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
2761 val = tp->rx_opt.user_mss;
2763 val = tp->rx_opt.mss_clamp;
2766 val = !!(tp->nonagle&TCP_NAGLE_OFF);
2769 val = !!(tp->nonagle&TCP_NAGLE_CORK);
2772 val = keepalive_time_when(tp) / HZ;
2775 val = keepalive_intvl_when(tp) / HZ;
2778 val = keepalive_probes(tp);
2781 val = icsk->icsk_syn_retries ? : sysctl_tcp_syn_retries;
2786 val = (val ? : sysctl_tcp_fin_timeout) / HZ;
2788 case TCP_DEFER_ACCEPT:
2789 val = retrans_to_secs(icsk->icsk_accept_queue.rskq_defer_accept,
2790 TCP_TIMEOUT_INIT / HZ, TCP_RTO_MAX / HZ);
2792 case TCP_WINDOW_CLAMP:
2793 val = tp->window_clamp;
2796 struct tcp_info info;
2798 if (get_user(len, optlen))
2801 tcp_get_info(sk, &info);
2803 len = min_t(unsigned int, len, sizeof(info));
2804 if (put_user(len, optlen))
2806 if (copy_to_user(optval, &info, len))
2811 val = !icsk->icsk_ack.pingpong;
2814 case TCP_CONGESTION:
2815 if (get_user(len, optlen))
2817 len = min_t(unsigned int, len, TCP_CA_NAME_MAX);
2818 if (put_user(len, optlen))
2820 if (copy_to_user(optval, icsk->icsk_ca_ops->name, len))
2824 case TCP_THIN_LINEAR_TIMEOUTS:
2827 case TCP_THIN_DUPACK:
2828 val = tp->thin_dupack;
2835 case TCP_REPAIR_QUEUE:
2837 val = tp->repair_queue;
2843 if (tp->repair_queue == TCP_SEND_QUEUE)
2844 val = tp->write_seq;
2845 else if (tp->repair_queue == TCP_RECV_QUEUE)
2851 case TCP_USER_TIMEOUT:
2852 val = jiffies_to_msecs(icsk->icsk_user_timeout);
2855 val = tcp_time_stamp + tp->tsoffset;
2857 case TCP_NOTSENT_LOWAT:
2858 val = tp->notsent_lowat;
2861 return -ENOPROTOOPT;
2864 if (put_user(len, optlen))
2866 if (copy_to_user(optval, &val, len))
2871 int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval,
2874 struct inet_connection_sock *icsk = inet_csk(sk);
2876 if (level != SOL_TCP)
2877 return icsk->icsk_af_ops->getsockopt(sk, level, optname,
2879 return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2881 EXPORT_SYMBOL(tcp_getsockopt);
2883 #ifdef CONFIG_COMPAT
2884 int compat_tcp_getsockopt(struct sock *sk, int level, int optname,
2885 char __user *optval, int __user *optlen)
2887 if (level != SOL_TCP)
2888 return inet_csk_compat_getsockopt(sk, level, optname,
2890 return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2892 EXPORT_SYMBOL(compat_tcp_getsockopt);
2895 #ifdef CONFIG_TCP_MD5SIG
2896 static struct tcp_md5sig_pool __percpu *tcp_md5sig_pool __read_mostly;
2897 static DEFINE_MUTEX(tcp_md5sig_mutex);
2899 static void __tcp_free_md5sig_pool(struct tcp_md5sig_pool __percpu *pool)
2903 for_each_possible_cpu(cpu) {
2904 struct tcp_md5sig_pool *p = per_cpu_ptr(pool, cpu);
2906 if (p->md5_desc.tfm)
2907 crypto_free_hash(p->md5_desc.tfm);
2912 static void __tcp_alloc_md5sig_pool(void)
2915 struct tcp_md5sig_pool __percpu *pool;
2917 pool = alloc_percpu(struct tcp_md5sig_pool);
2921 for_each_possible_cpu(cpu) {
2922 struct crypto_hash *hash;
2924 hash = crypto_alloc_hash("md5", 0, CRYPTO_ALG_ASYNC);
2925 if (IS_ERR_OR_NULL(hash))
2928 per_cpu_ptr(pool, cpu)->md5_desc.tfm = hash;
2930 /* before setting tcp_md5sig_pool, we must commit all writes
2931 * to memory. See ACCESS_ONCE() in tcp_get_md5sig_pool()
2934 tcp_md5sig_pool = pool;
2937 __tcp_free_md5sig_pool(pool);
2940 bool tcp_alloc_md5sig_pool(void)
2942 if (unlikely(!tcp_md5sig_pool)) {
2943 mutex_lock(&tcp_md5sig_mutex);
2945 if (!tcp_md5sig_pool)
2946 __tcp_alloc_md5sig_pool();
2948 mutex_unlock(&tcp_md5sig_mutex);
2950 return tcp_md5sig_pool != NULL;
2952 EXPORT_SYMBOL(tcp_alloc_md5sig_pool);
2956 * tcp_get_md5sig_pool - get md5sig_pool for this user
2958 * We use percpu structure, so if we succeed, we exit with preemption
2959 * and BH disabled, to make sure another thread or softirq handling
2960 * wont try to get same context.
2962 struct tcp_md5sig_pool *tcp_get_md5sig_pool(void)
2964 struct tcp_md5sig_pool __percpu *p;
2967 p = ACCESS_ONCE(tcp_md5sig_pool);
2969 return __this_cpu_ptr(p);
2974 EXPORT_SYMBOL(tcp_get_md5sig_pool);
2976 int tcp_md5_hash_header(struct tcp_md5sig_pool *hp,
2977 const struct tcphdr *th)
2979 struct scatterlist sg;
2983 /* We are not allowed to change tcphdr, make a local copy */
2984 memcpy(&hdr, th, sizeof(hdr));
2987 /* options aren't included in the hash */
2988 sg_init_one(&sg, &hdr, sizeof(hdr));
2989 err = crypto_hash_update(&hp->md5_desc, &sg, sizeof(hdr));
2992 EXPORT_SYMBOL(tcp_md5_hash_header);
2994 int tcp_md5_hash_skb_data(struct tcp_md5sig_pool *hp,
2995 const struct sk_buff *skb, unsigned int header_len)
2997 struct scatterlist sg;
2998 const struct tcphdr *tp = tcp_hdr(skb);
2999 struct hash_desc *desc = &hp->md5_desc;
3001 const unsigned int head_data_len = skb_headlen(skb) > header_len ?
3002 skb_headlen(skb) - header_len : 0;
3003 const struct skb_shared_info *shi = skb_shinfo(skb);
3004 struct sk_buff *frag_iter;
3006 sg_init_table(&sg, 1);
3008 sg_set_buf(&sg, ((u8 *) tp) + header_len, head_data_len);
3009 if (crypto_hash_update(desc, &sg, head_data_len))
3012 for (i = 0; i < shi->nr_frags; ++i) {
3013 const struct skb_frag_struct *f = &shi->frags[i];
3014 unsigned int offset = f->page_offset;
3015 struct page *page = skb_frag_page(f) + (offset >> PAGE_SHIFT);
3017 sg_set_page(&sg, page, skb_frag_size(f),
3018 offset_in_page(offset));
3019 if (crypto_hash_update(desc, &sg, skb_frag_size(f)))
3023 skb_walk_frags(skb, frag_iter)
3024 if (tcp_md5_hash_skb_data(hp, frag_iter, 0))
3029 EXPORT_SYMBOL(tcp_md5_hash_skb_data);
3031 int tcp_md5_hash_key(struct tcp_md5sig_pool *hp, const struct tcp_md5sig_key *key)
3033 struct scatterlist sg;
3035 sg_init_one(&sg, key->key, key->keylen);
3036 return crypto_hash_update(&hp->md5_desc, &sg, key->keylen);
3038 EXPORT_SYMBOL(tcp_md5_hash_key);
3042 void tcp_done(struct sock *sk)
3044 struct request_sock *req = tcp_sk(sk)->fastopen_rsk;
3046 if (sk->sk_state == TCP_SYN_SENT || sk->sk_state == TCP_SYN_RECV)
3047 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_ATTEMPTFAILS);
3049 tcp_set_state(sk, TCP_CLOSE);
3050 tcp_clear_xmit_timers(sk);
3052 reqsk_fastopen_remove(sk, req, false);
3054 sk->sk_shutdown = SHUTDOWN_MASK;
3056 if (!sock_flag(sk, SOCK_DEAD))
3057 sk->sk_state_change(sk);
3059 inet_csk_destroy_sock(sk);
3061 EXPORT_SYMBOL_GPL(tcp_done);
3063 extern struct tcp_congestion_ops tcp_reno;
3065 static __initdata unsigned long thash_entries;
3066 static int __init set_thash_entries(char *str)
3073 ret = kstrtoul(str, 0, &thash_entries);
3079 __setup("thash_entries=", set_thash_entries);
3081 void tcp_init_mem(struct net *net)
3083 unsigned long limit = nr_free_buffer_pages() / 8;
3084 limit = max(limit, 128UL);
3085 net->ipv4.sysctl_tcp_mem[0] = limit / 4 * 3;
3086 net->ipv4.sysctl_tcp_mem[1] = limit;
3087 net->ipv4.sysctl_tcp_mem[2] = net->ipv4.sysctl_tcp_mem[0] * 2;
3090 void __init tcp_init(void)
3092 struct sk_buff *skb = NULL;
3093 unsigned long limit;
3094 int max_rshare, max_wshare, cnt;
3097 BUILD_BUG_ON(sizeof(struct tcp_skb_cb) > sizeof(skb->cb));
3099 percpu_counter_init(&tcp_sockets_allocated, 0);
3100 percpu_counter_init(&tcp_orphan_count, 0);
3101 tcp_hashinfo.bind_bucket_cachep =
3102 kmem_cache_create("tcp_bind_bucket",
3103 sizeof(struct inet_bind_bucket), 0,
3104 SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL);
3106 /* Size and allocate the main established and bind bucket
3109 * The methodology is similar to that of the buffer cache.
3111 tcp_hashinfo.ehash =
3112 alloc_large_system_hash("TCP established",
3113 sizeof(struct inet_ehash_bucket),
3115 17, /* one slot per 128 KB of memory */
3118 &tcp_hashinfo.ehash_mask,
3120 thash_entries ? 0 : 512 * 1024);
3121 for (i = 0; i <= tcp_hashinfo.ehash_mask; i++) {
3122 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].chain, i);
3123 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].twchain, i);
3125 if (inet_ehash_locks_alloc(&tcp_hashinfo))
3126 panic("TCP: failed to alloc ehash_locks");
3127 tcp_hashinfo.bhash =
3128 alloc_large_system_hash("TCP bind",
3129 sizeof(struct inet_bind_hashbucket),
3130 tcp_hashinfo.ehash_mask + 1,
3131 17, /* one slot per 128 KB of memory */
3133 &tcp_hashinfo.bhash_size,
3137 tcp_hashinfo.bhash_size = 1U << tcp_hashinfo.bhash_size;
3138 for (i = 0; i < tcp_hashinfo.bhash_size; i++) {
3139 spin_lock_init(&tcp_hashinfo.bhash[i].lock);
3140 INIT_HLIST_HEAD(&tcp_hashinfo.bhash[i].chain);
3144 cnt = tcp_hashinfo.ehash_mask + 1;
3146 tcp_death_row.sysctl_max_tw_buckets = cnt / 2;
3147 sysctl_tcp_max_orphans = cnt / 2;
3148 sysctl_max_syn_backlog = max(128, cnt / 256);
3150 tcp_init_mem(&init_net);
3151 /* Set per-socket limits to no more than 1/128 the pressure threshold */
3152 limit = nr_free_buffer_pages() << (PAGE_SHIFT - 7);
3153 max_wshare = min(4UL*1024*1024, limit);
3154 max_rshare = min(6UL*1024*1024, limit);
3156 sysctl_tcp_wmem[0] = SK_MEM_QUANTUM;
3157 sysctl_tcp_wmem[1] = 16*1024;
3158 sysctl_tcp_wmem[2] = max(64*1024, max_wshare);
3160 sysctl_tcp_rmem[0] = SK_MEM_QUANTUM;
3161 sysctl_tcp_rmem[1] = 87380;
3162 sysctl_tcp_rmem[2] = max(87380, max_rshare);
3164 pr_info("Hash tables configured (established %u bind %u)\n",
3165 tcp_hashinfo.ehash_mask + 1, tcp_hashinfo.bhash_size);
3169 tcp_register_congestion_control(&tcp_reno);