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 int sysctl_tcp_min_tso_segs __read_mostly = 2;
288 int sysctl_tcp_autocorking __read_mostly = 1;
290 struct percpu_counter tcp_orphan_count;
291 EXPORT_SYMBOL_GPL(tcp_orphan_count);
293 long sysctl_tcp_mem[3] __read_mostly;
294 int sysctl_tcp_wmem[3] __read_mostly;
295 int sysctl_tcp_rmem[3] __read_mostly;
297 EXPORT_SYMBOL(sysctl_tcp_mem);
298 EXPORT_SYMBOL(sysctl_tcp_rmem);
299 EXPORT_SYMBOL(sysctl_tcp_wmem);
301 atomic_long_t tcp_memory_allocated; /* Current allocated memory. */
302 EXPORT_SYMBOL(tcp_memory_allocated);
305 * Current number of TCP sockets.
307 struct percpu_counter tcp_sockets_allocated;
308 EXPORT_SYMBOL(tcp_sockets_allocated);
313 struct tcp_splice_state {
314 struct pipe_inode_info *pipe;
320 * Pressure flag: try to collapse.
321 * Technical note: it is used by multiple contexts non atomically.
322 * All the __sk_mem_schedule() is of this nature: accounting
323 * is strict, actions are advisory and have some latency.
325 int tcp_memory_pressure __read_mostly;
326 EXPORT_SYMBOL(tcp_memory_pressure);
328 void tcp_enter_memory_pressure(struct sock *sk)
330 if (!tcp_memory_pressure) {
331 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMEMORYPRESSURES);
332 tcp_memory_pressure = 1;
335 EXPORT_SYMBOL(tcp_enter_memory_pressure);
337 /* Convert seconds to retransmits based on initial and max timeout */
338 static u8 secs_to_retrans(int seconds, int timeout, int rto_max)
343 int period = timeout;
346 while (seconds > period && res < 255) {
349 if (timeout > rto_max)
357 /* Convert retransmits to seconds based on initial and max timeout */
358 static int retrans_to_secs(u8 retrans, int timeout, int rto_max)
366 if (timeout > rto_max)
374 /* Address-family independent initialization for a tcp_sock.
376 * NOTE: A lot of things set to zero explicitly by call to
377 * sk_alloc() so need not be done here.
379 void tcp_init_sock(struct sock *sk)
381 struct inet_connection_sock *icsk = inet_csk(sk);
382 struct tcp_sock *tp = tcp_sk(sk);
384 skb_queue_head_init(&tp->out_of_order_queue);
385 tcp_init_xmit_timers(sk);
386 tcp_prequeue_init(tp);
387 INIT_LIST_HEAD(&tp->tsq_node);
389 icsk->icsk_rto = TCP_TIMEOUT_INIT;
390 tp->mdev = TCP_TIMEOUT_INIT;
392 /* So many TCP implementations out there (incorrectly) count the
393 * initial SYN frame in their delayed-ACK and congestion control
394 * algorithms that we must have the following bandaid to talk
395 * efficiently to them. -DaveM
397 tp->snd_cwnd = TCP_INIT_CWND;
399 /* See draft-stevens-tcpca-spec-01 for discussion of the
400 * initialization of these values.
402 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
403 tp->snd_cwnd_clamp = ~0;
404 tp->mss_cache = TCP_MSS_DEFAULT;
406 tp->reordering = sysctl_tcp_reordering;
407 tcp_enable_early_retrans(tp);
408 icsk->icsk_ca_ops = &tcp_init_congestion_ops;
412 sk->sk_state = TCP_CLOSE;
414 sk->sk_write_space = sk_stream_write_space;
415 sock_set_flag(sk, SOCK_USE_WRITE_QUEUE);
417 icsk->icsk_sync_mss = tcp_sync_mss;
419 sk->sk_sndbuf = sysctl_tcp_wmem[1];
420 sk->sk_rcvbuf = sysctl_tcp_rmem[1];
423 sock_update_memcg(sk);
424 sk_sockets_allocated_inc(sk);
427 EXPORT_SYMBOL(tcp_init_sock);
430 * Wait for a TCP event.
432 * Note that we don't need to lock the socket, as the upper poll layers
433 * take care of normal races (between the test and the event) and we don't
434 * go look at any of the socket buffers directly.
436 unsigned int tcp_poll(struct file *file, struct socket *sock, poll_table *wait)
439 struct sock *sk = sock->sk;
440 const struct tcp_sock *tp = tcp_sk(sk);
442 sock_rps_record_flow(sk);
444 sock_poll_wait(file, sk_sleep(sk), wait);
445 if (sk->sk_state == TCP_LISTEN)
446 return inet_csk_listen_poll(sk);
448 /* Socket is not locked. We are protected from async events
449 * by poll logic and correct handling of state changes
450 * made by other threads is impossible in any case.
456 * POLLHUP is certainly not done right. But poll() doesn't
457 * have a notion of HUP in just one direction, and for a
458 * socket the read side is more interesting.
460 * Some poll() documentation says that POLLHUP is incompatible
461 * with the POLLOUT/POLLWR flags, so somebody should check this
462 * all. But careful, it tends to be safer to return too many
463 * bits than too few, and you can easily break real applications
464 * if you don't tell them that something has hung up!
468 * Check number 1. POLLHUP is _UNMASKABLE_ event (see UNIX98 and
469 * our fs/select.c). It means that after we received EOF,
470 * poll always returns immediately, making impossible poll() on write()
471 * in state CLOSE_WAIT. One solution is evident --- to set POLLHUP
472 * if and only if shutdown has been made in both directions.
473 * Actually, it is interesting to look how Solaris and DUX
474 * solve this dilemma. I would prefer, if POLLHUP were maskable,
475 * then we could set it on SND_SHUTDOWN. BTW examples given
476 * in Stevens' books assume exactly this behaviour, it explains
477 * why POLLHUP is incompatible with POLLOUT. --ANK
479 * NOTE. Check for TCP_CLOSE is added. The goal is to prevent
480 * blocking on fresh not-connected or disconnected socket. --ANK
482 if (sk->sk_shutdown == SHUTDOWN_MASK || sk->sk_state == TCP_CLOSE)
484 if (sk->sk_shutdown & RCV_SHUTDOWN)
485 mask |= POLLIN | POLLRDNORM | POLLRDHUP;
487 /* Connected or passive Fast Open socket? */
488 if (sk->sk_state != TCP_SYN_SENT &&
489 (sk->sk_state != TCP_SYN_RECV || tp->fastopen_rsk != NULL)) {
490 int target = sock_rcvlowat(sk, 0, INT_MAX);
492 if (tp->urg_seq == tp->copied_seq &&
493 !sock_flag(sk, SOCK_URGINLINE) &&
497 /* Potential race condition. If read of tp below will
498 * escape above sk->sk_state, we can be illegally awaken
499 * in SYN_* states. */
500 if (tp->rcv_nxt - tp->copied_seq >= target)
501 mask |= POLLIN | POLLRDNORM;
503 if (!(sk->sk_shutdown & SEND_SHUTDOWN)) {
504 if (sk_stream_is_writeable(sk)) {
505 mask |= POLLOUT | POLLWRNORM;
506 } else { /* send SIGIO later */
507 set_bit(SOCK_ASYNC_NOSPACE,
508 &sk->sk_socket->flags);
509 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
511 /* Race breaker. If space is freed after
512 * wspace test but before the flags are set,
513 * IO signal will be lost.
515 if (sk_stream_is_writeable(sk))
516 mask |= POLLOUT | POLLWRNORM;
519 mask |= POLLOUT | POLLWRNORM;
521 if (tp->urg_data & TCP_URG_VALID)
524 /* This barrier is coupled with smp_wmb() in tcp_reset() */
531 EXPORT_SYMBOL(tcp_poll);
533 int tcp_ioctl(struct sock *sk, int cmd, unsigned long arg)
535 struct tcp_sock *tp = tcp_sk(sk);
541 if (sk->sk_state == TCP_LISTEN)
544 slow = lock_sock_fast(sk);
545 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
547 else if (sock_flag(sk, SOCK_URGINLINE) ||
549 before(tp->urg_seq, tp->copied_seq) ||
550 !before(tp->urg_seq, tp->rcv_nxt)) {
552 answ = tp->rcv_nxt - tp->copied_seq;
554 /* Subtract 1, if FIN was received */
555 if (answ && sock_flag(sk, SOCK_DONE))
558 answ = tp->urg_seq - tp->copied_seq;
559 unlock_sock_fast(sk, slow);
562 answ = tp->urg_data && tp->urg_seq == tp->copied_seq;
565 if (sk->sk_state == TCP_LISTEN)
568 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
571 answ = tp->write_seq - tp->snd_una;
574 if (sk->sk_state == TCP_LISTEN)
577 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
580 answ = tp->write_seq - tp->snd_nxt;
586 return put_user(answ, (int __user *)arg);
588 EXPORT_SYMBOL(tcp_ioctl);
590 static inline void tcp_mark_push(struct tcp_sock *tp, struct sk_buff *skb)
592 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
593 tp->pushed_seq = tp->write_seq;
596 static inline bool forced_push(const struct tcp_sock *tp)
598 return after(tp->write_seq, tp->pushed_seq + (tp->max_window >> 1));
601 static inline void skb_entail(struct sock *sk, struct sk_buff *skb)
603 struct tcp_sock *tp = tcp_sk(sk);
604 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
607 tcb->seq = tcb->end_seq = tp->write_seq;
608 tcb->tcp_flags = TCPHDR_ACK;
610 skb_header_release(skb);
611 tcp_add_write_queue_tail(sk, skb);
612 sk->sk_wmem_queued += skb->truesize;
613 sk_mem_charge(sk, skb->truesize);
614 if (tp->nonagle & TCP_NAGLE_PUSH)
615 tp->nonagle &= ~TCP_NAGLE_PUSH;
618 static inline void tcp_mark_urg(struct tcp_sock *tp, int flags)
621 tp->snd_up = tp->write_seq;
624 /* If a not yet filled skb is pushed, do not send it if
625 * we have packets in Qdisc or NIC queues :
626 * Because TX completion will happen shortly, it gives a chance
627 * to coalesce future sendmsg() payload into this skb, without
628 * need for a timer, and with no latency trade off.
629 * As packets containing data payload have a bigger truesize
630 * than pure acks (dataless) packets, the last check prevents
631 * autocorking if we only have an ACK in Qdisc/NIC queues.
633 static bool tcp_should_autocork(struct sock *sk, struct sk_buff *skb,
636 return skb->len < size_goal &&
637 sysctl_tcp_autocorking &&
638 atomic_read(&sk->sk_wmem_alloc) > skb->truesize;
641 static void tcp_push(struct sock *sk, int flags, int mss_now,
642 int nonagle, int size_goal)
644 struct tcp_sock *tp = tcp_sk(sk);
647 if (!tcp_send_head(sk))
650 skb = tcp_write_queue_tail(sk);
651 if (!(flags & MSG_MORE) || forced_push(tp))
652 tcp_mark_push(tp, skb);
654 tcp_mark_urg(tp, flags);
656 if (tcp_should_autocork(sk, skb, size_goal)) {
658 /* avoid atomic op if TSQ_THROTTLED bit is already set */
659 if (!test_bit(TSQ_THROTTLED, &tp->tsq_flags)) {
660 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPAUTOCORKING);
661 set_bit(TSQ_THROTTLED, &tp->tsq_flags);
666 if (flags & MSG_MORE)
667 nonagle = TCP_NAGLE_CORK;
669 __tcp_push_pending_frames(sk, mss_now, nonagle);
672 static int tcp_splice_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb,
673 unsigned int offset, size_t len)
675 struct tcp_splice_state *tss = rd_desc->arg.data;
678 ret = skb_splice_bits(skb, offset, tss->pipe, min(rd_desc->count, len),
681 rd_desc->count -= ret;
685 static int __tcp_splice_read(struct sock *sk, struct tcp_splice_state *tss)
687 /* Store TCP splice context information in read_descriptor_t. */
688 read_descriptor_t rd_desc = {
693 return tcp_read_sock(sk, &rd_desc, tcp_splice_data_recv);
697 * tcp_splice_read - splice data from TCP socket to a pipe
698 * @sock: socket to splice from
699 * @ppos: position (not valid)
700 * @pipe: pipe to splice to
701 * @len: number of bytes to splice
702 * @flags: splice modifier flags
705 * Will read pages from given socket and fill them into a pipe.
708 ssize_t tcp_splice_read(struct socket *sock, loff_t *ppos,
709 struct pipe_inode_info *pipe, size_t len,
712 struct sock *sk = sock->sk;
713 struct tcp_splice_state tss = {
722 sock_rps_record_flow(sk);
724 * We can't seek on a socket input
733 timeo = sock_rcvtimeo(sk, sock->file->f_flags & O_NONBLOCK);
735 ret = __tcp_splice_read(sk, &tss);
741 if (sock_flag(sk, SOCK_DONE))
744 ret = sock_error(sk);
747 if (sk->sk_shutdown & RCV_SHUTDOWN)
749 if (sk->sk_state == TCP_CLOSE) {
751 * This occurs when user tries to read
752 * from never connected socket.
754 if (!sock_flag(sk, SOCK_DONE))
762 sk_wait_data(sk, &timeo);
763 if (signal_pending(current)) {
764 ret = sock_intr_errno(timeo);
777 if (sk->sk_err || sk->sk_state == TCP_CLOSE ||
778 (sk->sk_shutdown & RCV_SHUTDOWN) ||
779 signal_pending(current))
790 EXPORT_SYMBOL(tcp_splice_read);
792 struct sk_buff *sk_stream_alloc_skb(struct sock *sk, int size, gfp_t gfp)
796 /* The TCP header must be at least 32-bit aligned. */
797 size = ALIGN(size, 4);
799 skb = alloc_skb_fclone(size + sk->sk_prot->max_header, gfp);
801 if (sk_wmem_schedule(sk, skb->truesize)) {
802 skb_reserve(skb, sk->sk_prot->max_header);
804 * Make sure that we have exactly size bytes
805 * available to the caller, no more, no less.
807 skb->reserved_tailroom = skb->end - skb->tail - size;
812 sk->sk_prot->enter_memory_pressure(sk);
813 sk_stream_moderate_sndbuf(sk);
818 static unsigned int tcp_xmit_size_goal(struct sock *sk, u32 mss_now,
821 struct tcp_sock *tp = tcp_sk(sk);
822 u32 xmit_size_goal, old_size_goal;
824 xmit_size_goal = mss_now;
826 if (large_allowed && sk_can_gso(sk)) {
829 /* Maybe we should/could use sk->sk_prot->max_header here ? */
830 hlen = inet_csk(sk)->icsk_af_ops->net_header_len +
831 inet_csk(sk)->icsk_ext_hdr_len +
834 /* Goal is to send at least one packet per ms,
835 * not one big TSO packet every 100 ms.
836 * This preserves ACK clocking and is consistent
837 * with tcp_tso_should_defer() heuristic.
839 gso_size = sk->sk_pacing_rate / (2 * MSEC_PER_SEC);
840 gso_size = max_t(u32, gso_size,
841 sysctl_tcp_min_tso_segs * mss_now);
843 xmit_size_goal = min_t(u32, gso_size,
844 sk->sk_gso_max_size - 1 - hlen);
846 xmit_size_goal = tcp_bound_to_half_wnd(tp, xmit_size_goal);
848 /* We try hard to avoid divides here */
849 old_size_goal = tp->xmit_size_goal_segs * mss_now;
851 if (likely(old_size_goal <= xmit_size_goal &&
852 old_size_goal + mss_now > xmit_size_goal)) {
853 xmit_size_goal = old_size_goal;
855 tp->xmit_size_goal_segs =
856 min_t(u16, xmit_size_goal / mss_now,
857 sk->sk_gso_max_segs);
858 xmit_size_goal = tp->xmit_size_goal_segs * mss_now;
862 return max(xmit_size_goal, mss_now);
865 static int tcp_send_mss(struct sock *sk, int *size_goal, int flags)
869 mss_now = tcp_current_mss(sk);
870 *size_goal = tcp_xmit_size_goal(sk, mss_now, !(flags & MSG_OOB));
875 static ssize_t do_tcp_sendpages(struct sock *sk, struct page *page, int offset,
876 size_t size, int flags)
878 struct tcp_sock *tp = tcp_sk(sk);
879 int mss_now, size_goal;
882 long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
884 /* Wait for a connection to finish. One exception is TCP Fast Open
885 * (passive side) where data is allowed to be sent before a connection
886 * is fully established.
888 if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) &&
889 !tcp_passive_fastopen(sk)) {
890 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
894 clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
896 mss_now = tcp_send_mss(sk, &size_goal, flags);
900 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
904 struct sk_buff *skb = tcp_write_queue_tail(sk);
908 if (!tcp_send_head(sk) || (copy = size_goal - skb->len) <= 0) {
910 if (!sk_stream_memory_free(sk))
911 goto wait_for_sndbuf;
913 skb = sk_stream_alloc_skb(sk, 0, sk->sk_allocation);
915 goto wait_for_memory;
924 i = skb_shinfo(skb)->nr_frags;
925 can_coalesce = skb_can_coalesce(skb, i, page, offset);
926 if (!can_coalesce && i >= MAX_SKB_FRAGS) {
927 tcp_mark_push(tp, skb);
930 if (!sk_wmem_schedule(sk, copy))
931 goto wait_for_memory;
934 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
937 skb_fill_page_desc(skb, i, page, offset, copy);
939 skb_shinfo(skb)->tx_flags |= SKBTX_SHARED_FRAG;
942 skb->data_len += copy;
943 skb->truesize += copy;
944 sk->sk_wmem_queued += copy;
945 sk_mem_charge(sk, copy);
946 skb->ip_summed = CHECKSUM_PARTIAL;
947 tp->write_seq += copy;
948 TCP_SKB_CB(skb)->end_seq += copy;
949 skb_shinfo(skb)->gso_segs = 0;
952 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
959 if (skb->len < size_goal || (flags & MSG_OOB))
962 if (forced_push(tp)) {
963 tcp_mark_push(tp, skb);
964 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
965 } else if (skb == tcp_send_head(sk))
966 tcp_push_one(sk, mss_now);
970 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
972 tcp_push(sk, flags & ~MSG_MORE, mss_now,
973 TCP_NAGLE_PUSH, size_goal);
975 if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
978 mss_now = tcp_send_mss(sk, &size_goal, flags);
982 if (copied && !(flags & MSG_SENDPAGE_NOTLAST))
983 tcp_push(sk, flags, mss_now, tp->nonagle, size_goal);
990 return sk_stream_error(sk, flags, err);
993 int tcp_sendpage(struct sock *sk, struct page *page, int offset,
994 size_t size, int flags)
998 if (!(sk->sk_route_caps & NETIF_F_SG) ||
999 !(sk->sk_route_caps & NETIF_F_ALL_CSUM))
1000 return sock_no_sendpage(sk->sk_socket, page, offset, size,
1004 res = do_tcp_sendpages(sk, page, offset, size, flags);
1008 EXPORT_SYMBOL(tcp_sendpage);
1010 static inline int select_size(const struct sock *sk, bool sg)
1012 const struct tcp_sock *tp = tcp_sk(sk);
1013 int tmp = tp->mss_cache;
1016 if (sk_can_gso(sk)) {
1017 /* Small frames wont use a full page:
1018 * Payload will immediately follow tcp header.
1020 tmp = SKB_WITH_OVERHEAD(2048 - MAX_TCP_HEADER);
1022 int pgbreak = SKB_MAX_HEAD(MAX_TCP_HEADER);
1024 if (tmp >= pgbreak &&
1025 tmp <= pgbreak + (MAX_SKB_FRAGS - 1) * PAGE_SIZE)
1033 void tcp_free_fastopen_req(struct tcp_sock *tp)
1035 if (tp->fastopen_req != NULL) {
1036 kfree(tp->fastopen_req);
1037 tp->fastopen_req = NULL;
1041 static int tcp_sendmsg_fastopen(struct sock *sk, struct msghdr *msg, int *size)
1043 struct tcp_sock *tp = tcp_sk(sk);
1046 if (!(sysctl_tcp_fastopen & TFO_CLIENT_ENABLE))
1048 if (tp->fastopen_req != NULL)
1049 return -EALREADY; /* Another Fast Open is in progress */
1051 tp->fastopen_req = kzalloc(sizeof(struct tcp_fastopen_request),
1053 if (unlikely(tp->fastopen_req == NULL))
1055 tp->fastopen_req->data = msg;
1057 flags = (msg->msg_flags & MSG_DONTWAIT) ? O_NONBLOCK : 0;
1058 err = __inet_stream_connect(sk->sk_socket, msg->msg_name,
1059 msg->msg_namelen, flags);
1060 *size = tp->fastopen_req->copied;
1061 tcp_free_fastopen_req(tp);
1065 int tcp_sendmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg,
1069 struct tcp_sock *tp = tcp_sk(sk);
1070 struct sk_buff *skb;
1071 int iovlen, flags, err, copied = 0;
1072 int mss_now = 0, size_goal, copied_syn = 0, offset = 0;
1078 flags = msg->msg_flags;
1079 if (flags & MSG_FASTOPEN) {
1080 err = tcp_sendmsg_fastopen(sk, msg, &copied_syn);
1081 if (err == -EINPROGRESS && copied_syn > 0)
1085 offset = copied_syn;
1088 timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
1090 /* Wait for a connection to finish. One exception is TCP Fast Open
1091 * (passive side) where data is allowed to be sent before a connection
1092 * is fully established.
1094 if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) &&
1095 !tcp_passive_fastopen(sk)) {
1096 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
1100 if (unlikely(tp->repair)) {
1101 if (tp->repair_queue == TCP_RECV_QUEUE) {
1102 copied = tcp_send_rcvq(sk, msg, size);
1107 if (tp->repair_queue == TCP_NO_QUEUE)
1110 /* 'common' sending to sendq */
1113 /* This should be in poll */
1114 clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
1116 mss_now = tcp_send_mss(sk, &size_goal, flags);
1118 /* Ok commence sending. */
1119 iovlen = msg->msg_iovlen;
1124 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
1127 sg = !!(sk->sk_route_caps & NETIF_F_SG);
1129 while (--iovlen >= 0) {
1130 size_t seglen = iov->iov_len;
1131 unsigned char __user *from = iov->iov_base;
1134 if (unlikely(offset > 0)) { /* Skip bytes copied in SYN */
1135 if (offset >= seglen) {
1144 while (seglen > 0) {
1146 int max = size_goal;
1148 skb = tcp_write_queue_tail(sk);
1149 if (tcp_send_head(sk)) {
1150 if (skb->ip_summed == CHECKSUM_NONE)
1152 copy = max - skb->len;
1157 /* Allocate new segment. If the interface is SG,
1158 * allocate skb fitting to single page.
1160 if (!sk_stream_memory_free(sk))
1161 goto wait_for_sndbuf;
1163 skb = sk_stream_alloc_skb(sk,
1164 select_size(sk, sg),
1167 goto wait_for_memory;
1170 * All packets are restored as if they have
1171 * already been sent.
1174 TCP_SKB_CB(skb)->when = tcp_time_stamp;
1177 * Check whether we can use HW checksum.
1179 if (sk->sk_route_caps & NETIF_F_ALL_CSUM)
1180 skb->ip_summed = CHECKSUM_PARTIAL;
1182 skb_entail(sk, skb);
1187 /* Try to append data to the end of skb. */
1191 /* Where to copy to? */
1192 if (skb_availroom(skb) > 0) {
1193 /* We have some space in skb head. Superb! */
1194 copy = min_t(int, copy, skb_availroom(skb));
1195 err = skb_add_data_nocache(sk, skb, from, copy);
1200 int i = skb_shinfo(skb)->nr_frags;
1201 struct page_frag *pfrag = sk_page_frag(sk);
1203 if (!sk_page_frag_refill(sk, pfrag))
1204 goto wait_for_memory;
1206 if (!skb_can_coalesce(skb, i, pfrag->page,
1208 if (i == MAX_SKB_FRAGS || !sg) {
1209 tcp_mark_push(tp, skb);
1215 copy = min_t(int, copy, pfrag->size - pfrag->offset);
1217 if (!sk_wmem_schedule(sk, copy))
1218 goto wait_for_memory;
1220 err = skb_copy_to_page_nocache(sk, from, skb,
1227 /* Update the skb. */
1229 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1231 skb_fill_page_desc(skb, i, pfrag->page,
1232 pfrag->offset, copy);
1233 get_page(pfrag->page);
1235 pfrag->offset += copy;
1239 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
1241 tp->write_seq += copy;
1242 TCP_SKB_CB(skb)->end_seq += copy;
1243 skb_shinfo(skb)->gso_segs = 0;
1247 if ((seglen -= copy) == 0 && iovlen == 0)
1250 if (skb->len < max || (flags & MSG_OOB) || unlikely(tp->repair))
1253 if (forced_push(tp)) {
1254 tcp_mark_push(tp, skb);
1255 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
1256 } else if (skb == tcp_send_head(sk))
1257 tcp_push_one(sk, mss_now);
1261 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1264 tcp_push(sk, flags & ~MSG_MORE, mss_now,
1265 TCP_NAGLE_PUSH, size_goal);
1267 if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
1270 mss_now = tcp_send_mss(sk, &size_goal, flags);
1276 tcp_push(sk, flags, mss_now, tp->nonagle, size_goal);
1278 return copied + copied_syn;
1282 tcp_unlink_write_queue(skb, sk);
1283 /* It is the one place in all of TCP, except connection
1284 * reset, where we can be unlinking the send_head.
1286 tcp_check_send_head(sk, skb);
1287 sk_wmem_free_skb(sk, skb);
1291 if (copied + copied_syn)
1294 err = sk_stream_error(sk, flags, err);
1298 EXPORT_SYMBOL(tcp_sendmsg);
1301 * Handle reading urgent data. BSD has very simple semantics for
1302 * this, no blocking and very strange errors 8)
1305 static int tcp_recv_urg(struct sock *sk, struct msghdr *msg, int len, int flags)
1307 struct tcp_sock *tp = tcp_sk(sk);
1309 /* No URG data to read. */
1310 if (sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data ||
1311 tp->urg_data == TCP_URG_READ)
1312 return -EINVAL; /* Yes this is right ! */
1314 if (sk->sk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DONE))
1317 if (tp->urg_data & TCP_URG_VALID) {
1319 char c = tp->urg_data;
1321 if (!(flags & MSG_PEEK))
1322 tp->urg_data = TCP_URG_READ;
1324 /* Read urgent data. */
1325 msg->msg_flags |= MSG_OOB;
1328 if (!(flags & MSG_TRUNC))
1329 err = memcpy_toiovec(msg->msg_iov, &c, 1);
1332 msg->msg_flags |= MSG_TRUNC;
1334 return err ? -EFAULT : len;
1337 if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
1340 /* Fixed the recv(..., MSG_OOB) behaviour. BSD docs and
1341 * the available implementations agree in this case:
1342 * this call should never block, independent of the
1343 * blocking state of the socket.
1344 * Mike <pall@rz.uni-karlsruhe.de>
1349 static int tcp_peek_sndq(struct sock *sk, struct msghdr *msg, int len)
1351 struct sk_buff *skb;
1352 int copied = 0, err = 0;
1354 /* XXX -- need to support SO_PEEK_OFF */
1356 skb_queue_walk(&sk->sk_write_queue, skb) {
1357 err = skb_copy_datagram_iovec(skb, 0, msg->msg_iov, skb->len);
1364 return err ?: copied;
1367 /* Clean up the receive buffer for full frames taken by the user,
1368 * then send an ACK if necessary. COPIED is the number of bytes
1369 * tcp_recvmsg has given to the user so far, it speeds up the
1370 * calculation of whether or not we must ACK for the sake of
1373 void tcp_cleanup_rbuf(struct sock *sk, int copied)
1375 struct tcp_sock *tp = tcp_sk(sk);
1376 bool time_to_ack = false;
1378 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
1380 WARN(skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq),
1381 "cleanup rbuf bug: copied %X seq %X rcvnxt %X\n",
1382 tp->copied_seq, TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt);
1384 if (inet_csk_ack_scheduled(sk)) {
1385 const struct inet_connection_sock *icsk = inet_csk(sk);
1386 /* Delayed ACKs frequently hit locked sockets during bulk
1388 if (icsk->icsk_ack.blocked ||
1389 /* Once-per-two-segments ACK was not sent by tcp_input.c */
1390 tp->rcv_nxt - tp->rcv_wup > icsk->icsk_ack.rcv_mss ||
1392 * If this read emptied read buffer, we send ACK, if
1393 * connection is not bidirectional, user drained
1394 * receive buffer and there was a small segment
1398 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED2) ||
1399 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED) &&
1400 !icsk->icsk_ack.pingpong)) &&
1401 !atomic_read(&sk->sk_rmem_alloc)))
1405 /* We send an ACK if we can now advertise a non-zero window
1406 * which has been raised "significantly".
1408 * Even if window raised up to infinity, do not send window open ACK
1409 * in states, where we will not receive more. It is useless.
1411 if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) {
1412 __u32 rcv_window_now = tcp_receive_window(tp);
1414 /* Optimize, __tcp_select_window() is not cheap. */
1415 if (2*rcv_window_now <= tp->window_clamp) {
1416 __u32 new_window = __tcp_select_window(sk);
1418 /* Send ACK now, if this read freed lots of space
1419 * in our buffer. Certainly, new_window is new window.
1420 * We can advertise it now, if it is not less than current one.
1421 * "Lots" means "at least twice" here.
1423 if (new_window && new_window >= 2 * rcv_window_now)
1431 static void tcp_prequeue_process(struct sock *sk)
1433 struct sk_buff *skb;
1434 struct tcp_sock *tp = tcp_sk(sk);
1436 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPPREQUEUED);
1438 /* RX process wants to run with disabled BHs, though it is not
1441 while ((skb = __skb_dequeue(&tp->ucopy.prequeue)) != NULL)
1442 sk_backlog_rcv(sk, skb);
1445 /* Clear memory counter. */
1446 tp->ucopy.memory = 0;
1449 #ifdef CONFIG_NET_DMA
1450 static void tcp_service_net_dma(struct sock *sk, bool wait)
1452 dma_cookie_t done, used;
1453 dma_cookie_t last_issued;
1454 struct tcp_sock *tp = tcp_sk(sk);
1456 if (!tp->ucopy.dma_chan)
1459 last_issued = tp->ucopy.dma_cookie;
1460 dma_async_issue_pending(tp->ucopy.dma_chan);
1463 if (dma_async_is_tx_complete(tp->ucopy.dma_chan,
1465 &used) == DMA_COMPLETE) {
1466 /* Safe to free early-copied skbs now */
1467 __skb_queue_purge(&sk->sk_async_wait_queue);
1470 struct sk_buff *skb;
1471 while ((skb = skb_peek(&sk->sk_async_wait_queue)) &&
1472 (dma_async_is_complete(skb->dma_cookie, done,
1473 used) == DMA_COMPLETE)) {
1474 __skb_dequeue(&sk->sk_async_wait_queue);
1482 static struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off)
1484 struct sk_buff *skb;
1487 while ((skb = skb_peek(&sk->sk_receive_queue)) != NULL) {
1488 offset = seq - TCP_SKB_CB(skb)->seq;
1489 if (tcp_hdr(skb)->syn)
1491 if (offset < skb->len || tcp_hdr(skb)->fin) {
1495 /* This looks weird, but this can happen if TCP collapsing
1496 * splitted a fat GRO packet, while we released socket lock
1497 * in skb_splice_bits()
1499 sk_eat_skb(sk, skb, false);
1505 * This routine provides an alternative to tcp_recvmsg() for routines
1506 * that would like to handle copying from skbuffs directly in 'sendfile'
1509 * - It is assumed that the socket was locked by the caller.
1510 * - The routine does not block.
1511 * - At present, there is no support for reading OOB data
1512 * or for 'peeking' the socket using this routine
1513 * (although both would be easy to implement).
1515 int tcp_read_sock(struct sock *sk, read_descriptor_t *desc,
1516 sk_read_actor_t recv_actor)
1518 struct sk_buff *skb;
1519 struct tcp_sock *tp = tcp_sk(sk);
1520 u32 seq = tp->copied_seq;
1524 if (sk->sk_state == TCP_LISTEN)
1526 while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1527 if (offset < skb->len) {
1531 len = skb->len - offset;
1532 /* Stop reading if we hit a patch of urgent data */
1534 u32 urg_offset = tp->urg_seq - seq;
1535 if (urg_offset < len)
1540 used = recv_actor(desc, skb, offset, len);
1545 } else if (used <= len) {
1550 /* If recv_actor drops the lock (e.g. TCP splice
1551 * receive) the skb pointer might be invalid when
1552 * getting here: tcp_collapse might have deleted it
1553 * while aggregating skbs from the socket queue.
1555 skb = tcp_recv_skb(sk, seq - 1, &offset);
1558 /* TCP coalescing might have appended data to the skb.
1559 * Try to splice more frags
1561 if (offset + 1 != skb->len)
1564 if (tcp_hdr(skb)->fin) {
1565 sk_eat_skb(sk, skb, false);
1569 sk_eat_skb(sk, skb, false);
1572 tp->copied_seq = seq;
1574 tp->copied_seq = seq;
1576 tcp_rcv_space_adjust(sk);
1578 /* Clean up data we have read: This will do ACK frames. */
1580 tcp_recv_skb(sk, seq, &offset);
1581 tcp_cleanup_rbuf(sk, copied);
1585 EXPORT_SYMBOL(tcp_read_sock);
1588 * This routine copies from a sock struct into the user buffer.
1590 * Technical note: in 2.3 we work on _locked_ socket, so that
1591 * tricks with *seq access order and skb->users are not required.
1592 * Probably, code can be easily improved even more.
1595 int tcp_recvmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg,
1596 size_t len, int nonblock, int flags, int *addr_len)
1598 struct tcp_sock *tp = tcp_sk(sk);
1604 int target; /* Read at least this many bytes */
1606 struct task_struct *user_recv = NULL;
1607 bool copied_early = false;
1608 struct sk_buff *skb;
1611 if (sk_can_busy_loop(sk) && skb_queue_empty(&sk->sk_receive_queue) &&
1612 (sk->sk_state == TCP_ESTABLISHED))
1613 sk_busy_loop(sk, nonblock);
1618 if (sk->sk_state == TCP_LISTEN)
1621 timeo = sock_rcvtimeo(sk, nonblock);
1623 /* Urgent data needs to be handled specially. */
1624 if (flags & MSG_OOB)
1627 if (unlikely(tp->repair)) {
1629 if (!(flags & MSG_PEEK))
1632 if (tp->repair_queue == TCP_SEND_QUEUE)
1636 if (tp->repair_queue == TCP_NO_QUEUE)
1639 /* 'common' recv queue MSG_PEEK-ing */
1642 seq = &tp->copied_seq;
1643 if (flags & MSG_PEEK) {
1644 peek_seq = tp->copied_seq;
1648 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
1650 #ifdef CONFIG_NET_DMA
1651 tp->ucopy.dma_chan = NULL;
1653 skb = skb_peek_tail(&sk->sk_receive_queue);
1658 available = TCP_SKB_CB(skb)->seq + skb->len - (*seq);
1659 if ((available < target) &&
1660 (len > sysctl_tcp_dma_copybreak) && !(flags & MSG_PEEK) &&
1661 !sysctl_tcp_low_latency &&
1662 net_dma_find_channel()) {
1663 preempt_enable_no_resched();
1664 tp->ucopy.pinned_list =
1665 dma_pin_iovec_pages(msg->msg_iov, len);
1667 preempt_enable_no_resched();
1675 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
1676 if (tp->urg_data && tp->urg_seq == *seq) {
1679 if (signal_pending(current)) {
1680 copied = timeo ? sock_intr_errno(timeo) : -EAGAIN;
1685 /* Next get a buffer. */
1687 skb_queue_walk(&sk->sk_receive_queue, skb) {
1688 /* Now that we have two receive queues this
1691 if (WARN(before(*seq, TCP_SKB_CB(skb)->seq),
1692 "recvmsg bug: copied %X seq %X rcvnxt %X fl %X\n",
1693 *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt,
1697 offset = *seq - TCP_SKB_CB(skb)->seq;
1698 if (tcp_hdr(skb)->syn)
1700 if (offset < skb->len)
1702 if (tcp_hdr(skb)->fin)
1704 WARN(!(flags & MSG_PEEK),
1705 "recvmsg bug 2: copied %X seq %X rcvnxt %X fl %X\n",
1706 *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt, flags);
1709 /* Well, if we have backlog, try to process it now yet. */
1711 if (copied >= target && !sk->sk_backlog.tail)
1716 sk->sk_state == TCP_CLOSE ||
1717 (sk->sk_shutdown & RCV_SHUTDOWN) ||
1719 signal_pending(current))
1722 if (sock_flag(sk, SOCK_DONE))
1726 copied = sock_error(sk);
1730 if (sk->sk_shutdown & RCV_SHUTDOWN)
1733 if (sk->sk_state == TCP_CLOSE) {
1734 if (!sock_flag(sk, SOCK_DONE)) {
1735 /* This occurs when user tries to read
1736 * from never connected socket.
1749 if (signal_pending(current)) {
1750 copied = sock_intr_errno(timeo);
1755 tcp_cleanup_rbuf(sk, copied);
1757 if (!sysctl_tcp_low_latency && tp->ucopy.task == user_recv) {
1758 /* Install new reader */
1759 if (!user_recv && !(flags & (MSG_TRUNC | MSG_PEEK))) {
1760 user_recv = current;
1761 tp->ucopy.task = user_recv;
1762 tp->ucopy.iov = msg->msg_iov;
1765 tp->ucopy.len = len;
1767 WARN_ON(tp->copied_seq != tp->rcv_nxt &&
1768 !(flags & (MSG_PEEK | MSG_TRUNC)));
1770 /* Ugly... If prequeue is not empty, we have to
1771 * process it before releasing socket, otherwise
1772 * order will be broken at second iteration.
1773 * More elegant solution is required!!!
1775 * Look: we have the following (pseudo)queues:
1777 * 1. packets in flight
1782 * Each queue can be processed only if the next ones
1783 * are empty. At this point we have empty receive_queue.
1784 * But prequeue _can_ be not empty after 2nd iteration,
1785 * when we jumped to start of loop because backlog
1786 * processing added something to receive_queue.
1787 * We cannot release_sock(), because backlog contains
1788 * packets arrived _after_ prequeued ones.
1790 * Shortly, algorithm is clear --- to process all
1791 * the queues in order. We could make it more directly,
1792 * requeueing packets from backlog to prequeue, if
1793 * is not empty. It is more elegant, but eats cycles,
1796 if (!skb_queue_empty(&tp->ucopy.prequeue))
1799 /* __ Set realtime policy in scheduler __ */
1802 #ifdef CONFIG_NET_DMA
1803 if (tp->ucopy.dma_chan) {
1804 if (tp->rcv_wnd == 0 &&
1805 !skb_queue_empty(&sk->sk_async_wait_queue)) {
1806 tcp_service_net_dma(sk, true);
1807 tcp_cleanup_rbuf(sk, copied);
1809 dma_async_issue_pending(tp->ucopy.dma_chan);
1812 if (copied >= target) {
1813 /* Do not sleep, just process backlog. */
1817 sk_wait_data(sk, &timeo);
1819 #ifdef CONFIG_NET_DMA
1820 tcp_service_net_dma(sk, false); /* Don't block */
1821 tp->ucopy.wakeup = 0;
1827 /* __ Restore normal policy in scheduler __ */
1829 if ((chunk = len - tp->ucopy.len) != 0) {
1830 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMBACKLOG, chunk);
1835 if (tp->rcv_nxt == tp->copied_seq &&
1836 !skb_queue_empty(&tp->ucopy.prequeue)) {
1838 tcp_prequeue_process(sk);
1840 if ((chunk = len - tp->ucopy.len) != 0) {
1841 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1847 if ((flags & MSG_PEEK) &&
1848 (peek_seq - copied - urg_hole != tp->copied_seq)) {
1849 net_dbg_ratelimited("TCP(%s:%d): Application bug, race in MSG_PEEK\n",
1851 task_pid_nr(current));
1852 peek_seq = tp->copied_seq;
1857 /* Ok so how much can we use? */
1858 used = skb->len - offset;
1862 /* Do we have urgent data here? */
1864 u32 urg_offset = tp->urg_seq - *seq;
1865 if (urg_offset < used) {
1867 if (!sock_flag(sk, SOCK_URGINLINE)) {
1880 if (!(flags & MSG_TRUNC)) {
1881 #ifdef CONFIG_NET_DMA
1882 if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list)
1883 tp->ucopy.dma_chan = net_dma_find_channel();
1885 if (tp->ucopy.dma_chan) {
1886 tp->ucopy.dma_cookie = dma_skb_copy_datagram_iovec(
1887 tp->ucopy.dma_chan, skb, offset,
1889 tp->ucopy.pinned_list);
1891 if (tp->ucopy.dma_cookie < 0) {
1893 pr_alert("%s: dma_cookie < 0\n",
1896 /* Exception. Bailout! */
1902 dma_async_issue_pending(tp->ucopy.dma_chan);
1904 if ((offset + used) == skb->len)
1905 copied_early = true;
1910 err = skb_copy_datagram_iovec(skb, offset,
1911 msg->msg_iov, used);
1913 /* Exception. Bailout! */
1925 tcp_rcv_space_adjust(sk);
1928 if (tp->urg_data && after(tp->copied_seq, tp->urg_seq)) {
1930 tcp_fast_path_check(sk);
1932 if (used + offset < skb->len)
1935 if (tcp_hdr(skb)->fin)
1937 if (!(flags & MSG_PEEK)) {
1938 sk_eat_skb(sk, skb, copied_early);
1939 copied_early = false;
1944 /* Process the FIN. */
1946 if (!(flags & MSG_PEEK)) {
1947 sk_eat_skb(sk, skb, copied_early);
1948 copied_early = false;
1954 if (!skb_queue_empty(&tp->ucopy.prequeue)) {
1957 tp->ucopy.len = copied > 0 ? len : 0;
1959 tcp_prequeue_process(sk);
1961 if (copied > 0 && (chunk = len - tp->ucopy.len) != 0) {
1962 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1968 tp->ucopy.task = NULL;
1972 #ifdef CONFIG_NET_DMA
1973 tcp_service_net_dma(sk, true); /* Wait for queue to drain */
1974 tp->ucopy.dma_chan = NULL;
1976 if (tp->ucopy.pinned_list) {
1977 dma_unpin_iovec_pages(tp->ucopy.pinned_list);
1978 tp->ucopy.pinned_list = NULL;
1982 /* According to UNIX98, msg_name/msg_namelen are ignored
1983 * on connected socket. I was just happy when found this 8) --ANK
1986 /* Clean up data we have read: This will do ACK frames. */
1987 tcp_cleanup_rbuf(sk, copied);
1997 err = tcp_recv_urg(sk, msg, len, flags);
2001 err = tcp_peek_sndq(sk, msg, len);
2004 EXPORT_SYMBOL(tcp_recvmsg);
2006 void tcp_set_state(struct sock *sk, int state)
2008 int oldstate = sk->sk_state;
2011 case TCP_ESTABLISHED:
2012 if (oldstate != TCP_ESTABLISHED)
2013 TCP_INC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
2017 if (oldstate == TCP_CLOSE_WAIT || oldstate == TCP_ESTABLISHED)
2018 TCP_INC_STATS(sock_net(sk), TCP_MIB_ESTABRESETS);
2020 sk->sk_prot->unhash(sk);
2021 if (inet_csk(sk)->icsk_bind_hash &&
2022 !(sk->sk_userlocks & SOCK_BINDPORT_LOCK))
2026 if (oldstate == TCP_ESTABLISHED)
2027 TCP_DEC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
2030 /* Change state AFTER socket is unhashed to avoid closed
2031 * socket sitting in hash tables.
2033 sk->sk_state = state;
2036 SOCK_DEBUG(sk, "TCP sk=%p, State %s -> %s\n", sk, statename[oldstate], statename[state]);
2039 EXPORT_SYMBOL_GPL(tcp_set_state);
2042 * State processing on a close. This implements the state shift for
2043 * sending our FIN frame. Note that we only send a FIN for some
2044 * states. A shutdown() may have already sent the FIN, or we may be
2048 static const unsigned char new_state[16] = {
2049 /* current state: new state: action: */
2050 /* (Invalid) */ TCP_CLOSE,
2051 /* TCP_ESTABLISHED */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2052 /* TCP_SYN_SENT */ TCP_CLOSE,
2053 /* TCP_SYN_RECV */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2054 /* TCP_FIN_WAIT1 */ TCP_FIN_WAIT1,
2055 /* TCP_FIN_WAIT2 */ TCP_FIN_WAIT2,
2056 /* TCP_TIME_WAIT */ TCP_CLOSE,
2057 /* TCP_CLOSE */ TCP_CLOSE,
2058 /* TCP_CLOSE_WAIT */ TCP_LAST_ACK | TCP_ACTION_FIN,
2059 /* TCP_LAST_ACK */ TCP_LAST_ACK,
2060 /* TCP_LISTEN */ TCP_CLOSE,
2061 /* TCP_CLOSING */ TCP_CLOSING,
2064 static int tcp_close_state(struct sock *sk)
2066 int next = (int)new_state[sk->sk_state];
2067 int ns = next & TCP_STATE_MASK;
2069 tcp_set_state(sk, ns);
2071 return next & TCP_ACTION_FIN;
2075 * Shutdown the sending side of a connection. Much like close except
2076 * that we don't receive shut down or sock_set_flag(sk, SOCK_DEAD).
2079 void tcp_shutdown(struct sock *sk, int how)
2081 /* We need to grab some memory, and put together a FIN,
2082 * and then put it into the queue to be sent.
2083 * Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
2085 if (!(how & SEND_SHUTDOWN))
2088 /* If we've already sent a FIN, or it's a closed state, skip this. */
2089 if ((1 << sk->sk_state) &
2090 (TCPF_ESTABLISHED | TCPF_SYN_SENT |
2091 TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) {
2092 /* Clear out any half completed packets. FIN if needed. */
2093 if (tcp_close_state(sk))
2097 EXPORT_SYMBOL(tcp_shutdown);
2099 bool tcp_check_oom(struct sock *sk, int shift)
2101 bool too_many_orphans, out_of_socket_memory;
2103 too_many_orphans = tcp_too_many_orphans(sk, shift);
2104 out_of_socket_memory = tcp_out_of_memory(sk);
2106 if (too_many_orphans)
2107 net_info_ratelimited("too many orphaned sockets\n");
2108 if (out_of_socket_memory)
2109 net_info_ratelimited("out of memory -- consider tuning tcp_mem\n");
2110 return too_many_orphans || out_of_socket_memory;
2113 void tcp_close(struct sock *sk, long timeout)
2115 struct sk_buff *skb;
2116 int data_was_unread = 0;
2120 sk->sk_shutdown = SHUTDOWN_MASK;
2122 if (sk->sk_state == TCP_LISTEN) {
2123 tcp_set_state(sk, TCP_CLOSE);
2126 inet_csk_listen_stop(sk);
2128 goto adjudge_to_death;
2131 /* We need to flush the recv. buffs. We do this only on the
2132 * descriptor close, not protocol-sourced closes, because the
2133 * reader process may not have drained the data yet!
2135 while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) {
2136 u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq -
2138 data_was_unread += len;
2144 /* If socket has been already reset (e.g. in tcp_reset()) - kill it. */
2145 if (sk->sk_state == TCP_CLOSE)
2146 goto adjudge_to_death;
2148 /* As outlined in RFC 2525, section 2.17, we send a RST here because
2149 * data was lost. To witness the awful effects of the old behavior of
2150 * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk
2151 * GET in an FTP client, suspend the process, wait for the client to
2152 * advertise a zero window, then kill -9 the FTP client, wheee...
2153 * Note: timeout is always zero in such a case.
2155 if (unlikely(tcp_sk(sk)->repair)) {
2156 sk->sk_prot->disconnect(sk, 0);
2157 } else if (data_was_unread) {
2158 /* Unread data was tossed, zap the connection. */
2159 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONCLOSE);
2160 tcp_set_state(sk, TCP_CLOSE);
2161 tcp_send_active_reset(sk, sk->sk_allocation);
2162 } else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) {
2163 /* Check zero linger _after_ checking for unread data. */
2164 sk->sk_prot->disconnect(sk, 0);
2165 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
2166 } else if (tcp_close_state(sk)) {
2167 /* We FIN if the application ate all the data before
2168 * zapping the connection.
2171 /* RED-PEN. Formally speaking, we have broken TCP state
2172 * machine. State transitions:
2174 * TCP_ESTABLISHED -> TCP_FIN_WAIT1
2175 * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible)
2176 * TCP_CLOSE_WAIT -> TCP_LAST_ACK
2178 * are legal only when FIN has been sent (i.e. in window),
2179 * rather than queued out of window. Purists blame.
2181 * F.e. "RFC state" is ESTABLISHED,
2182 * if Linux state is FIN-WAIT-1, but FIN is still not sent.
2184 * The visible declinations are that sometimes
2185 * we enter time-wait state, when it is not required really
2186 * (harmless), do not send active resets, when they are
2187 * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
2188 * they look as CLOSING or LAST_ACK for Linux)
2189 * Probably, I missed some more holelets.
2191 * XXX (TFO) - To start off we don't support SYN+ACK+FIN
2192 * in a single packet! (May consider it later but will
2193 * probably need API support or TCP_CORK SYN-ACK until
2194 * data is written and socket is closed.)
2199 sk_stream_wait_close(sk, timeout);
2202 state = sk->sk_state;
2206 /* It is the last release_sock in its life. It will remove backlog. */
2210 /* Now socket is owned by kernel and we acquire BH lock
2211 to finish close. No need to check for user refs.
2215 WARN_ON(sock_owned_by_user(sk));
2217 percpu_counter_inc(sk->sk_prot->orphan_count);
2219 /* Have we already been destroyed by a softirq or backlog? */
2220 if (state != TCP_CLOSE && sk->sk_state == TCP_CLOSE)
2223 /* This is a (useful) BSD violating of the RFC. There is a
2224 * problem with TCP as specified in that the other end could
2225 * keep a socket open forever with no application left this end.
2226 * We use a 3 minute timeout (about the same as BSD) then kill
2227 * our end. If they send after that then tough - BUT: long enough
2228 * that we won't make the old 4*rto = almost no time - whoops
2231 * Nope, it was not mistake. It is really desired behaviour
2232 * f.e. on http servers, when such sockets are useless, but
2233 * consume significant resources. Let's do it with special
2234 * linger2 option. --ANK
2237 if (sk->sk_state == TCP_FIN_WAIT2) {
2238 struct tcp_sock *tp = tcp_sk(sk);
2239 if (tp->linger2 < 0) {
2240 tcp_set_state(sk, TCP_CLOSE);
2241 tcp_send_active_reset(sk, GFP_ATOMIC);
2242 NET_INC_STATS_BH(sock_net(sk),
2243 LINUX_MIB_TCPABORTONLINGER);
2245 const int tmo = tcp_fin_time(sk);
2247 if (tmo > TCP_TIMEWAIT_LEN) {
2248 inet_csk_reset_keepalive_timer(sk,
2249 tmo - TCP_TIMEWAIT_LEN);
2251 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
2256 if (sk->sk_state != TCP_CLOSE) {
2258 if (tcp_check_oom(sk, 0)) {
2259 tcp_set_state(sk, TCP_CLOSE);
2260 tcp_send_active_reset(sk, GFP_ATOMIC);
2261 NET_INC_STATS_BH(sock_net(sk),
2262 LINUX_MIB_TCPABORTONMEMORY);
2266 if (sk->sk_state == TCP_CLOSE) {
2267 struct request_sock *req = tcp_sk(sk)->fastopen_rsk;
2268 /* We could get here with a non-NULL req if the socket is
2269 * aborted (e.g., closed with unread data) before 3WHS
2273 reqsk_fastopen_remove(sk, req, false);
2274 inet_csk_destroy_sock(sk);
2276 /* Otherwise, socket is reprieved until protocol close. */
2283 EXPORT_SYMBOL(tcp_close);
2285 /* These states need RST on ABORT according to RFC793 */
2287 static inline bool tcp_need_reset(int state)
2289 return (1 << state) &
2290 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 |
2291 TCPF_FIN_WAIT2 | TCPF_SYN_RECV);
2294 int tcp_disconnect(struct sock *sk, int flags)
2296 struct inet_sock *inet = inet_sk(sk);
2297 struct inet_connection_sock *icsk = inet_csk(sk);
2298 struct tcp_sock *tp = tcp_sk(sk);
2300 int old_state = sk->sk_state;
2302 if (old_state != TCP_CLOSE)
2303 tcp_set_state(sk, TCP_CLOSE);
2305 /* ABORT function of RFC793 */
2306 if (old_state == TCP_LISTEN) {
2307 inet_csk_listen_stop(sk);
2308 } else if (unlikely(tp->repair)) {
2309 sk->sk_err = ECONNABORTED;
2310 } else if (tcp_need_reset(old_state) ||
2311 (tp->snd_nxt != tp->write_seq &&
2312 (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK))) {
2313 /* The last check adjusts for discrepancy of Linux wrt. RFC
2316 tcp_send_active_reset(sk, gfp_any());
2317 sk->sk_err = ECONNRESET;
2318 } else if (old_state == TCP_SYN_SENT)
2319 sk->sk_err = ECONNRESET;
2321 tcp_clear_xmit_timers(sk);
2322 __skb_queue_purge(&sk->sk_receive_queue);
2323 tcp_write_queue_purge(sk);
2324 __skb_queue_purge(&tp->out_of_order_queue);
2325 #ifdef CONFIG_NET_DMA
2326 __skb_queue_purge(&sk->sk_async_wait_queue);
2329 inet->inet_dport = 0;
2331 if (!(sk->sk_userlocks & SOCK_BINDADDR_LOCK))
2332 inet_reset_saddr(sk);
2334 sk->sk_shutdown = 0;
2335 sock_reset_flag(sk, SOCK_DONE);
2337 if ((tp->write_seq += tp->max_window + 2) == 0)
2339 icsk->icsk_backoff = 0;
2341 icsk->icsk_probes_out = 0;
2342 tp->packets_out = 0;
2343 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
2344 tp->snd_cwnd_cnt = 0;
2345 tp->window_clamp = 0;
2346 tcp_set_ca_state(sk, TCP_CA_Open);
2347 tcp_clear_retrans(tp);
2348 inet_csk_delack_init(sk);
2349 tcp_init_send_head(sk);
2350 memset(&tp->rx_opt, 0, sizeof(tp->rx_opt));
2353 WARN_ON(inet->inet_num && !icsk->icsk_bind_hash);
2355 sk->sk_error_report(sk);
2358 EXPORT_SYMBOL(tcp_disconnect);
2360 void tcp_sock_destruct(struct sock *sk)
2362 inet_sock_destruct(sk);
2364 kfree(inet_csk(sk)->icsk_accept_queue.fastopenq);
2367 static inline bool tcp_can_repair_sock(const struct sock *sk)
2369 return ns_capable(sock_net(sk)->user_ns, CAP_NET_ADMIN) &&
2370 ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_ESTABLISHED));
2373 static int tcp_repair_options_est(struct tcp_sock *tp,
2374 struct tcp_repair_opt __user *optbuf, unsigned int len)
2376 struct tcp_repair_opt opt;
2378 while (len >= sizeof(opt)) {
2379 if (copy_from_user(&opt, optbuf, sizeof(opt)))
2385 switch (opt.opt_code) {
2387 tp->rx_opt.mss_clamp = opt.opt_val;
2391 u16 snd_wscale = opt.opt_val & 0xFFFF;
2392 u16 rcv_wscale = opt.opt_val >> 16;
2394 if (snd_wscale > 14 || rcv_wscale > 14)
2397 tp->rx_opt.snd_wscale = snd_wscale;
2398 tp->rx_opt.rcv_wscale = rcv_wscale;
2399 tp->rx_opt.wscale_ok = 1;
2402 case TCPOPT_SACK_PERM:
2403 if (opt.opt_val != 0)
2406 tp->rx_opt.sack_ok |= TCP_SACK_SEEN;
2407 if (sysctl_tcp_fack)
2408 tcp_enable_fack(tp);
2410 case TCPOPT_TIMESTAMP:
2411 if (opt.opt_val != 0)
2414 tp->rx_opt.tstamp_ok = 1;
2423 * Socket option code for TCP.
2425 static int do_tcp_setsockopt(struct sock *sk, int level,
2426 int optname, char __user *optval, unsigned int optlen)
2428 struct tcp_sock *tp = tcp_sk(sk);
2429 struct inet_connection_sock *icsk = inet_csk(sk);
2433 /* These are data/string values, all the others are ints */
2435 case TCP_CONGESTION: {
2436 char name[TCP_CA_NAME_MAX];
2441 val = strncpy_from_user(name, optval,
2442 min_t(long, TCP_CA_NAME_MAX-1, optlen));
2448 err = tcp_set_congestion_control(sk, name);
2457 if (optlen < sizeof(int))
2460 if (get_user(val, (int __user *)optval))
2467 /* Values greater than interface MTU won't take effect. However
2468 * at the point when this call is done we typically don't yet
2469 * know which interface is going to be used */
2470 if (val < TCP_MIN_MSS || val > MAX_TCP_WINDOW) {
2474 tp->rx_opt.user_mss = val;
2479 /* TCP_NODELAY is weaker than TCP_CORK, so that
2480 * this option on corked socket is remembered, but
2481 * it is not activated until cork is cleared.
2483 * However, when TCP_NODELAY is set we make
2484 * an explicit push, which overrides even TCP_CORK
2485 * for currently queued segments.
2487 tp->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH;
2488 tcp_push_pending_frames(sk);
2490 tp->nonagle &= ~TCP_NAGLE_OFF;
2494 case TCP_THIN_LINEAR_TIMEOUTS:
2495 if (val < 0 || val > 1)
2501 case TCP_THIN_DUPACK:
2502 if (val < 0 || val > 1)
2505 tp->thin_dupack = val;
2506 if (tp->thin_dupack)
2507 tcp_disable_early_retrans(tp);
2512 if (!tcp_can_repair_sock(sk))
2514 else if (val == 1) {
2516 sk->sk_reuse = SK_FORCE_REUSE;
2517 tp->repair_queue = TCP_NO_QUEUE;
2518 } else if (val == 0) {
2520 sk->sk_reuse = SK_NO_REUSE;
2521 tcp_send_window_probe(sk);
2527 case TCP_REPAIR_QUEUE:
2530 else if (val < TCP_QUEUES_NR)
2531 tp->repair_queue = val;
2537 if (sk->sk_state != TCP_CLOSE)
2539 else if (tp->repair_queue == TCP_SEND_QUEUE)
2540 tp->write_seq = val;
2541 else if (tp->repair_queue == TCP_RECV_QUEUE)
2547 case TCP_REPAIR_OPTIONS:
2550 else if (sk->sk_state == TCP_ESTABLISHED)
2551 err = tcp_repair_options_est(tp,
2552 (struct tcp_repair_opt __user *)optval,
2559 /* When set indicates to always queue non-full frames.
2560 * Later the user clears this option and we transmit
2561 * any pending partial frames in the queue. This is
2562 * meant to be used alongside sendfile() to get properly
2563 * filled frames when the user (for example) must write
2564 * out headers with a write() call first and then use
2565 * sendfile to send out the data parts.
2567 * TCP_CORK can be set together with TCP_NODELAY and it is
2568 * stronger than TCP_NODELAY.
2571 tp->nonagle |= TCP_NAGLE_CORK;
2573 tp->nonagle &= ~TCP_NAGLE_CORK;
2574 if (tp->nonagle&TCP_NAGLE_OFF)
2575 tp->nonagle |= TCP_NAGLE_PUSH;
2576 tcp_push_pending_frames(sk);
2581 if (val < 1 || val > MAX_TCP_KEEPIDLE)
2584 tp->keepalive_time = val * HZ;
2585 if (sock_flag(sk, SOCK_KEEPOPEN) &&
2586 !((1 << sk->sk_state) &
2587 (TCPF_CLOSE | TCPF_LISTEN))) {
2588 u32 elapsed = keepalive_time_elapsed(tp);
2589 if (tp->keepalive_time > elapsed)
2590 elapsed = tp->keepalive_time - elapsed;
2593 inet_csk_reset_keepalive_timer(sk, elapsed);
2598 if (val < 1 || val > MAX_TCP_KEEPINTVL)
2601 tp->keepalive_intvl = val * HZ;
2604 if (val < 1 || val > MAX_TCP_KEEPCNT)
2607 tp->keepalive_probes = val;
2610 if (val < 1 || val > MAX_TCP_SYNCNT)
2613 icsk->icsk_syn_retries = val;
2619 else if (val > sysctl_tcp_fin_timeout / HZ)
2622 tp->linger2 = val * HZ;
2625 case TCP_DEFER_ACCEPT:
2626 /* Translate value in seconds to number of retransmits */
2627 icsk->icsk_accept_queue.rskq_defer_accept =
2628 secs_to_retrans(val, TCP_TIMEOUT_INIT / HZ,
2632 case TCP_WINDOW_CLAMP:
2634 if (sk->sk_state != TCP_CLOSE) {
2638 tp->window_clamp = 0;
2640 tp->window_clamp = val < SOCK_MIN_RCVBUF / 2 ?
2641 SOCK_MIN_RCVBUF / 2 : val;
2646 icsk->icsk_ack.pingpong = 1;
2648 icsk->icsk_ack.pingpong = 0;
2649 if ((1 << sk->sk_state) &
2650 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) &&
2651 inet_csk_ack_scheduled(sk)) {
2652 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
2653 tcp_cleanup_rbuf(sk, 1);
2655 icsk->icsk_ack.pingpong = 1;
2660 #ifdef CONFIG_TCP_MD5SIG
2662 /* Read the IP->Key mappings from userspace */
2663 err = tp->af_specific->md5_parse(sk, optval, optlen);
2666 case TCP_USER_TIMEOUT:
2667 /* Cap the max timeout in ms TCP will retry/retrans
2668 * before giving up and aborting (ETIMEDOUT) a connection.
2673 icsk->icsk_user_timeout = msecs_to_jiffies(val);
2677 if (val >= 0 && ((1 << sk->sk_state) & (TCPF_CLOSE |
2679 err = fastopen_init_queue(sk, val);
2687 tp->tsoffset = val - tcp_time_stamp;
2689 case TCP_NOTSENT_LOWAT:
2690 tp->notsent_lowat = val;
2691 sk->sk_write_space(sk);
2702 int tcp_setsockopt(struct sock *sk, int level, int optname, char __user *optval,
2703 unsigned int optlen)
2705 const struct inet_connection_sock *icsk = inet_csk(sk);
2707 if (level != SOL_TCP)
2708 return icsk->icsk_af_ops->setsockopt(sk, level, optname,
2710 return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2712 EXPORT_SYMBOL(tcp_setsockopt);
2714 #ifdef CONFIG_COMPAT
2715 int compat_tcp_setsockopt(struct sock *sk, int level, int optname,
2716 char __user *optval, unsigned int optlen)
2718 if (level != SOL_TCP)
2719 return inet_csk_compat_setsockopt(sk, level, optname,
2721 return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2723 EXPORT_SYMBOL(compat_tcp_setsockopt);
2726 /* Return information about state of tcp endpoint in API format. */
2727 void tcp_get_info(const struct sock *sk, struct tcp_info *info)
2729 const struct tcp_sock *tp = tcp_sk(sk);
2730 const struct inet_connection_sock *icsk = inet_csk(sk);
2731 u32 now = tcp_time_stamp;
2733 memset(info, 0, sizeof(*info));
2735 info->tcpi_state = sk->sk_state;
2736 info->tcpi_ca_state = icsk->icsk_ca_state;
2737 info->tcpi_retransmits = icsk->icsk_retransmits;
2738 info->tcpi_probes = icsk->icsk_probes_out;
2739 info->tcpi_backoff = icsk->icsk_backoff;
2741 if (tp->rx_opt.tstamp_ok)
2742 info->tcpi_options |= TCPI_OPT_TIMESTAMPS;
2743 if (tcp_is_sack(tp))
2744 info->tcpi_options |= TCPI_OPT_SACK;
2745 if (tp->rx_opt.wscale_ok) {
2746 info->tcpi_options |= TCPI_OPT_WSCALE;
2747 info->tcpi_snd_wscale = tp->rx_opt.snd_wscale;
2748 info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale;
2751 if (tp->ecn_flags & TCP_ECN_OK)
2752 info->tcpi_options |= TCPI_OPT_ECN;
2753 if (tp->ecn_flags & TCP_ECN_SEEN)
2754 info->tcpi_options |= TCPI_OPT_ECN_SEEN;
2755 if (tp->syn_data_acked)
2756 info->tcpi_options |= TCPI_OPT_SYN_DATA;
2758 info->tcpi_rto = jiffies_to_usecs(icsk->icsk_rto);
2759 info->tcpi_ato = jiffies_to_usecs(icsk->icsk_ack.ato);
2760 info->tcpi_snd_mss = tp->mss_cache;
2761 info->tcpi_rcv_mss = icsk->icsk_ack.rcv_mss;
2763 if (sk->sk_state == TCP_LISTEN) {
2764 info->tcpi_unacked = sk->sk_ack_backlog;
2765 info->tcpi_sacked = sk->sk_max_ack_backlog;
2767 info->tcpi_unacked = tp->packets_out;
2768 info->tcpi_sacked = tp->sacked_out;
2770 info->tcpi_lost = tp->lost_out;
2771 info->tcpi_retrans = tp->retrans_out;
2772 info->tcpi_fackets = tp->fackets_out;
2774 info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime);
2775 info->tcpi_last_data_recv = jiffies_to_msecs(now - icsk->icsk_ack.lrcvtime);
2776 info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp);
2778 info->tcpi_pmtu = icsk->icsk_pmtu_cookie;
2779 info->tcpi_rcv_ssthresh = tp->rcv_ssthresh;
2780 info->tcpi_rtt = jiffies_to_usecs(tp->srtt)>>3;
2781 info->tcpi_rttvar = jiffies_to_usecs(tp->mdev)>>2;
2782 info->tcpi_snd_ssthresh = tp->snd_ssthresh;
2783 info->tcpi_snd_cwnd = tp->snd_cwnd;
2784 info->tcpi_advmss = tp->advmss;
2785 info->tcpi_reordering = tp->reordering;
2787 info->tcpi_rcv_rtt = jiffies_to_usecs(tp->rcv_rtt_est.rtt)>>3;
2788 info->tcpi_rcv_space = tp->rcvq_space.space;
2790 info->tcpi_total_retrans = tp->total_retrans;
2792 EXPORT_SYMBOL_GPL(tcp_get_info);
2794 static int do_tcp_getsockopt(struct sock *sk, int level,
2795 int optname, char __user *optval, int __user *optlen)
2797 struct inet_connection_sock *icsk = inet_csk(sk);
2798 struct tcp_sock *tp = tcp_sk(sk);
2801 if (get_user(len, optlen))
2804 len = min_t(unsigned int, len, sizeof(int));
2811 val = tp->mss_cache;
2812 if (!val && ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
2813 val = tp->rx_opt.user_mss;
2815 val = tp->rx_opt.mss_clamp;
2818 val = !!(tp->nonagle&TCP_NAGLE_OFF);
2821 val = !!(tp->nonagle&TCP_NAGLE_CORK);
2824 val = keepalive_time_when(tp) / HZ;
2827 val = keepalive_intvl_when(tp) / HZ;
2830 val = keepalive_probes(tp);
2833 val = icsk->icsk_syn_retries ? : sysctl_tcp_syn_retries;
2838 val = (val ? : sysctl_tcp_fin_timeout) / HZ;
2840 case TCP_DEFER_ACCEPT:
2841 val = retrans_to_secs(icsk->icsk_accept_queue.rskq_defer_accept,
2842 TCP_TIMEOUT_INIT / HZ, TCP_RTO_MAX / HZ);
2844 case TCP_WINDOW_CLAMP:
2845 val = tp->window_clamp;
2848 struct tcp_info info;
2850 if (get_user(len, optlen))
2853 tcp_get_info(sk, &info);
2855 len = min_t(unsigned int, len, sizeof(info));
2856 if (put_user(len, optlen))
2858 if (copy_to_user(optval, &info, len))
2863 val = !icsk->icsk_ack.pingpong;
2866 case TCP_CONGESTION:
2867 if (get_user(len, optlen))
2869 len = min_t(unsigned int, len, TCP_CA_NAME_MAX);
2870 if (put_user(len, optlen))
2872 if (copy_to_user(optval, icsk->icsk_ca_ops->name, len))
2876 case TCP_THIN_LINEAR_TIMEOUTS:
2879 case TCP_THIN_DUPACK:
2880 val = tp->thin_dupack;
2887 case TCP_REPAIR_QUEUE:
2889 val = tp->repair_queue;
2895 if (tp->repair_queue == TCP_SEND_QUEUE)
2896 val = tp->write_seq;
2897 else if (tp->repair_queue == TCP_RECV_QUEUE)
2903 case TCP_USER_TIMEOUT:
2904 val = jiffies_to_msecs(icsk->icsk_user_timeout);
2907 val = tcp_time_stamp + tp->tsoffset;
2909 case TCP_NOTSENT_LOWAT:
2910 val = tp->notsent_lowat;
2913 return -ENOPROTOOPT;
2916 if (put_user(len, optlen))
2918 if (copy_to_user(optval, &val, len))
2923 int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval,
2926 struct inet_connection_sock *icsk = inet_csk(sk);
2928 if (level != SOL_TCP)
2929 return icsk->icsk_af_ops->getsockopt(sk, level, optname,
2931 return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2933 EXPORT_SYMBOL(tcp_getsockopt);
2935 #ifdef CONFIG_COMPAT
2936 int compat_tcp_getsockopt(struct sock *sk, int level, int optname,
2937 char __user *optval, int __user *optlen)
2939 if (level != SOL_TCP)
2940 return inet_csk_compat_getsockopt(sk, level, optname,
2942 return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2944 EXPORT_SYMBOL(compat_tcp_getsockopt);
2947 #ifdef CONFIG_TCP_MD5SIG
2948 static struct tcp_md5sig_pool __percpu *tcp_md5sig_pool __read_mostly;
2949 static DEFINE_MUTEX(tcp_md5sig_mutex);
2951 static void __tcp_free_md5sig_pool(struct tcp_md5sig_pool __percpu *pool)
2955 for_each_possible_cpu(cpu) {
2956 struct tcp_md5sig_pool *p = per_cpu_ptr(pool, cpu);
2958 if (p->md5_desc.tfm)
2959 crypto_free_hash(p->md5_desc.tfm);
2964 static void __tcp_alloc_md5sig_pool(void)
2967 struct tcp_md5sig_pool __percpu *pool;
2969 pool = alloc_percpu(struct tcp_md5sig_pool);
2973 for_each_possible_cpu(cpu) {
2974 struct crypto_hash *hash;
2976 hash = crypto_alloc_hash("md5", 0, CRYPTO_ALG_ASYNC);
2977 if (IS_ERR_OR_NULL(hash))
2980 per_cpu_ptr(pool, cpu)->md5_desc.tfm = hash;
2982 /* before setting tcp_md5sig_pool, we must commit all writes
2983 * to memory. See ACCESS_ONCE() in tcp_get_md5sig_pool()
2986 tcp_md5sig_pool = pool;
2989 __tcp_free_md5sig_pool(pool);
2992 bool tcp_alloc_md5sig_pool(void)
2994 if (unlikely(!tcp_md5sig_pool)) {
2995 mutex_lock(&tcp_md5sig_mutex);
2997 if (!tcp_md5sig_pool)
2998 __tcp_alloc_md5sig_pool();
3000 mutex_unlock(&tcp_md5sig_mutex);
3002 return tcp_md5sig_pool != NULL;
3004 EXPORT_SYMBOL(tcp_alloc_md5sig_pool);
3008 * tcp_get_md5sig_pool - get md5sig_pool for this user
3010 * We use percpu structure, so if we succeed, we exit with preemption
3011 * and BH disabled, to make sure another thread or softirq handling
3012 * wont try to get same context.
3014 struct tcp_md5sig_pool *tcp_get_md5sig_pool(void)
3016 struct tcp_md5sig_pool __percpu *p;
3019 p = ACCESS_ONCE(tcp_md5sig_pool);
3021 return __this_cpu_ptr(p);
3026 EXPORT_SYMBOL(tcp_get_md5sig_pool);
3028 int tcp_md5_hash_header(struct tcp_md5sig_pool *hp,
3029 const struct tcphdr *th)
3031 struct scatterlist sg;
3035 /* We are not allowed to change tcphdr, make a local copy */
3036 memcpy(&hdr, th, sizeof(hdr));
3039 /* options aren't included in the hash */
3040 sg_init_one(&sg, &hdr, sizeof(hdr));
3041 err = crypto_hash_update(&hp->md5_desc, &sg, sizeof(hdr));
3044 EXPORT_SYMBOL(tcp_md5_hash_header);
3046 int tcp_md5_hash_skb_data(struct tcp_md5sig_pool *hp,
3047 const struct sk_buff *skb, unsigned int header_len)
3049 struct scatterlist sg;
3050 const struct tcphdr *tp = tcp_hdr(skb);
3051 struct hash_desc *desc = &hp->md5_desc;
3053 const unsigned int head_data_len = skb_headlen(skb) > header_len ?
3054 skb_headlen(skb) - header_len : 0;
3055 const struct skb_shared_info *shi = skb_shinfo(skb);
3056 struct sk_buff *frag_iter;
3058 sg_init_table(&sg, 1);
3060 sg_set_buf(&sg, ((u8 *) tp) + header_len, head_data_len);
3061 if (crypto_hash_update(desc, &sg, head_data_len))
3064 for (i = 0; i < shi->nr_frags; ++i) {
3065 const struct skb_frag_struct *f = &shi->frags[i];
3066 unsigned int offset = f->page_offset;
3067 struct page *page = skb_frag_page(f) + (offset >> PAGE_SHIFT);
3069 sg_set_page(&sg, page, skb_frag_size(f),
3070 offset_in_page(offset));
3071 if (crypto_hash_update(desc, &sg, skb_frag_size(f)))
3075 skb_walk_frags(skb, frag_iter)
3076 if (tcp_md5_hash_skb_data(hp, frag_iter, 0))
3081 EXPORT_SYMBOL(tcp_md5_hash_skb_data);
3083 int tcp_md5_hash_key(struct tcp_md5sig_pool *hp, const struct tcp_md5sig_key *key)
3085 struct scatterlist sg;
3087 sg_init_one(&sg, key->key, key->keylen);
3088 return crypto_hash_update(&hp->md5_desc, &sg, key->keylen);
3090 EXPORT_SYMBOL(tcp_md5_hash_key);
3094 void tcp_done(struct sock *sk)
3096 struct request_sock *req = tcp_sk(sk)->fastopen_rsk;
3098 if (sk->sk_state == TCP_SYN_SENT || sk->sk_state == TCP_SYN_RECV)
3099 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_ATTEMPTFAILS);
3101 tcp_set_state(sk, TCP_CLOSE);
3102 tcp_clear_xmit_timers(sk);
3104 reqsk_fastopen_remove(sk, req, false);
3106 sk->sk_shutdown = SHUTDOWN_MASK;
3108 if (!sock_flag(sk, SOCK_DEAD))
3109 sk->sk_state_change(sk);
3111 inet_csk_destroy_sock(sk);
3113 EXPORT_SYMBOL_GPL(tcp_done);
3115 extern struct tcp_congestion_ops tcp_reno;
3117 static __initdata unsigned long thash_entries;
3118 static int __init set_thash_entries(char *str)
3125 ret = kstrtoul(str, 0, &thash_entries);
3131 __setup("thash_entries=", set_thash_entries);
3133 static void tcp_init_mem(void)
3135 unsigned long limit = nr_free_buffer_pages() / 8;
3136 limit = max(limit, 128UL);
3137 sysctl_tcp_mem[0] = limit / 4 * 3;
3138 sysctl_tcp_mem[1] = limit;
3139 sysctl_tcp_mem[2] = sysctl_tcp_mem[0] * 2;
3142 void __init tcp_init(void)
3144 struct sk_buff *skb = NULL;
3145 unsigned long limit;
3146 int max_rshare, max_wshare, cnt;
3149 BUILD_BUG_ON(sizeof(struct tcp_skb_cb) > sizeof(skb->cb));
3151 percpu_counter_init(&tcp_sockets_allocated, 0);
3152 percpu_counter_init(&tcp_orphan_count, 0);
3153 tcp_hashinfo.bind_bucket_cachep =
3154 kmem_cache_create("tcp_bind_bucket",
3155 sizeof(struct inet_bind_bucket), 0,
3156 SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL);
3158 /* Size and allocate the main established and bind bucket
3161 * The methodology is similar to that of the buffer cache.
3163 tcp_hashinfo.ehash =
3164 alloc_large_system_hash("TCP established",
3165 sizeof(struct inet_ehash_bucket),
3167 17, /* one slot per 128 KB of memory */
3170 &tcp_hashinfo.ehash_mask,
3172 thash_entries ? 0 : 512 * 1024);
3173 for (i = 0; i <= tcp_hashinfo.ehash_mask; i++)
3174 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].chain, i);
3176 if (inet_ehash_locks_alloc(&tcp_hashinfo))
3177 panic("TCP: failed to alloc ehash_locks");
3178 tcp_hashinfo.bhash =
3179 alloc_large_system_hash("TCP bind",
3180 sizeof(struct inet_bind_hashbucket),
3181 tcp_hashinfo.ehash_mask + 1,
3182 17, /* one slot per 128 KB of memory */
3184 &tcp_hashinfo.bhash_size,
3188 tcp_hashinfo.bhash_size = 1U << tcp_hashinfo.bhash_size;
3189 for (i = 0; i < tcp_hashinfo.bhash_size; i++) {
3190 spin_lock_init(&tcp_hashinfo.bhash[i].lock);
3191 INIT_HLIST_HEAD(&tcp_hashinfo.bhash[i].chain);
3195 cnt = tcp_hashinfo.ehash_mask + 1;
3197 tcp_death_row.sysctl_max_tw_buckets = cnt / 2;
3198 sysctl_tcp_max_orphans = cnt / 2;
3199 sysctl_max_syn_backlog = max(128, cnt / 256);
3202 /* Set per-socket limits to no more than 1/128 the pressure threshold */
3203 limit = nr_free_buffer_pages() << (PAGE_SHIFT - 7);
3204 max_wshare = min(4UL*1024*1024, limit);
3205 max_rshare = min(6UL*1024*1024, limit);
3207 sysctl_tcp_wmem[0] = SK_MEM_QUANTUM;
3208 sysctl_tcp_wmem[1] = 16*1024;
3209 sysctl_tcp_wmem[2] = max(64*1024, max_wshare);
3211 sysctl_tcp_rmem[0] = SK_MEM_QUANTUM;
3212 sysctl_tcp_rmem[1] = 87380;
3213 sysctl_tcp_rmem[2] = max(87380, max_rshare);
3215 pr_info("Hash tables configured (established %u bind %u)\n",
3216 tcp_hashinfo.ehash_mask + 1, tcp_hashinfo.bhash_size);
3220 tcp_register_congestion_control(&tcp_reno);