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>
271 #include <linux/uid_stat.h>
273 #include <net/icmp.h>
274 #include <net/inet_common.h>
276 #include <net/xfrm.h>
278 #include <net/ip6_route.h>
279 #include <net/ipv6.h>
280 #include <net/transp_v6.h>
281 #include <net/netdma.h>
282 #include <net/sock.h>
284 #include <asm/uaccess.h>
285 #include <asm/ioctls.h>
287 int sysctl_tcp_fin_timeout __read_mostly = TCP_FIN_TIMEOUT;
289 int sysctl_tcp_min_tso_segs __read_mostly = 2;
291 struct percpu_counter tcp_orphan_count;
292 EXPORT_SYMBOL_GPL(tcp_orphan_count);
294 int sysctl_tcp_wmem[3] __read_mostly;
295 int sysctl_tcp_rmem[3] __read_mostly;
297 EXPORT_SYMBOL(sysctl_tcp_rmem);
298 EXPORT_SYMBOL(sysctl_tcp_wmem);
300 atomic_long_t tcp_memory_allocated; /* Current allocated memory. */
301 EXPORT_SYMBOL(tcp_memory_allocated);
304 * Current number of TCP sockets.
306 struct percpu_counter tcp_sockets_allocated;
307 EXPORT_SYMBOL(tcp_sockets_allocated);
312 struct tcp_splice_state {
313 struct pipe_inode_info *pipe;
319 * Pressure flag: try to collapse.
320 * Technical note: it is used by multiple contexts non atomically.
321 * All the __sk_mem_schedule() is of this nature: accounting
322 * is strict, actions are advisory and have some latency.
324 int tcp_memory_pressure __read_mostly;
325 EXPORT_SYMBOL(tcp_memory_pressure);
327 void tcp_enter_memory_pressure(struct sock *sk)
329 if (!tcp_memory_pressure) {
330 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMEMORYPRESSURES);
331 tcp_memory_pressure = 1;
334 EXPORT_SYMBOL(tcp_enter_memory_pressure);
336 /* Convert seconds to retransmits based on initial and max timeout */
337 static u8 secs_to_retrans(int seconds, int timeout, int rto_max)
342 int period = timeout;
345 while (seconds > period && res < 255) {
348 if (timeout > rto_max)
356 /* Convert retransmits to seconds based on initial and max timeout */
357 static int retrans_to_secs(u8 retrans, int timeout, int rto_max)
365 if (timeout > rto_max)
373 /* Address-family independent initialization for a tcp_sock.
375 * NOTE: A lot of things set to zero explicitly by call to
376 * sk_alloc() so need not be done here.
378 void tcp_init_sock(struct sock *sk)
380 struct inet_connection_sock *icsk = inet_csk(sk);
381 struct tcp_sock *tp = tcp_sk(sk);
383 skb_queue_head_init(&tp->out_of_order_queue);
384 tcp_init_xmit_timers(sk);
385 tcp_prequeue_init(tp);
386 INIT_LIST_HEAD(&tp->tsq_node);
388 icsk->icsk_rto = TCP_TIMEOUT_INIT;
389 tp->mdev = TCP_TIMEOUT_INIT;
391 /* So many TCP implementations out there (incorrectly) count the
392 * initial SYN frame in their delayed-ACK and congestion control
393 * algorithms that we must have the following bandaid to talk
394 * efficiently to them. -DaveM
396 tp->snd_cwnd = TCP_INIT_CWND;
398 /* See draft-stevens-tcpca-spec-01 for discussion of the
399 * initialization of these values.
401 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
402 tp->snd_cwnd_clamp = ~0;
403 tp->mss_cache = TCP_MSS_DEFAULT;
405 tp->reordering = sysctl_tcp_reordering;
406 tcp_enable_early_retrans(tp);
407 icsk->icsk_ca_ops = &tcp_init_congestion_ops;
411 sk->sk_state = TCP_CLOSE;
413 sk->sk_write_space = sk_stream_write_space;
414 sock_set_flag(sk, SOCK_USE_WRITE_QUEUE);
416 icsk->icsk_sync_mss = tcp_sync_mss;
418 /* Presumed zeroed, in order of appearance:
419 * cookie_in_always, cookie_out_never,
420 * s_data_constant, s_data_in, s_data_out
422 sk->sk_sndbuf = sysctl_tcp_wmem[1];
423 sk->sk_rcvbuf = sysctl_tcp_rmem[1];
426 sock_update_memcg(sk);
427 sk_sockets_allocated_inc(sk);
430 EXPORT_SYMBOL(tcp_init_sock);
433 * Wait for a TCP event.
435 * Note that we don't need to lock the socket, as the upper poll layers
436 * take care of normal races (between the test and the event) and we don't
437 * go look at any of the socket buffers directly.
439 unsigned int tcp_poll(struct file *file, struct socket *sock, poll_table *wait)
442 struct sock *sk = sock->sk;
443 const struct tcp_sock *tp = tcp_sk(sk);
445 sock_poll_wait(file, sk_sleep(sk), wait);
446 if (sk->sk_state == TCP_LISTEN)
447 return inet_csk_listen_poll(sk);
449 /* Socket is not locked. We are protected from async events
450 * by poll logic and correct handling of state changes
451 * made by other threads is impossible in any case.
457 * POLLHUP is certainly not done right. But poll() doesn't
458 * have a notion of HUP in just one direction, and for a
459 * socket the read side is more interesting.
461 * Some poll() documentation says that POLLHUP is incompatible
462 * with the POLLOUT/POLLWR flags, so somebody should check this
463 * all. But careful, it tends to be safer to return too many
464 * bits than too few, and you can easily break real applications
465 * if you don't tell them that something has hung up!
469 * Check number 1. POLLHUP is _UNMASKABLE_ event (see UNIX98 and
470 * our fs/select.c). It means that after we received EOF,
471 * poll always returns immediately, making impossible poll() on write()
472 * in state CLOSE_WAIT. One solution is evident --- to set POLLHUP
473 * if and only if shutdown has been made in both directions.
474 * Actually, it is interesting to look how Solaris and DUX
475 * solve this dilemma. I would prefer, if POLLHUP were maskable,
476 * then we could set it on SND_SHUTDOWN. BTW examples given
477 * in Stevens' books assume exactly this behaviour, it explains
478 * why POLLHUP is incompatible with POLLOUT. --ANK
480 * NOTE. Check for TCP_CLOSE is added. The goal is to prevent
481 * blocking on fresh not-connected or disconnected socket. --ANK
483 if (sk->sk_shutdown == SHUTDOWN_MASK || sk->sk_state == TCP_CLOSE)
485 if (sk->sk_shutdown & RCV_SHUTDOWN)
486 mask |= POLLIN | POLLRDNORM | POLLRDHUP;
488 /* Connected or passive Fast Open socket? */
489 if (sk->sk_state != TCP_SYN_SENT &&
490 (sk->sk_state != TCP_SYN_RECV || tp->fastopen_rsk != NULL)) {
491 int target = sock_rcvlowat(sk, 0, INT_MAX);
493 if (tp->urg_seq == tp->copied_seq &&
494 !sock_flag(sk, SOCK_URGINLINE) &&
498 /* Potential race condition. If read of tp below will
499 * escape above sk->sk_state, we can be illegally awaken
500 * in SYN_* states. */
501 if (tp->rcv_nxt - tp->copied_seq >= target)
502 mask |= POLLIN | POLLRDNORM;
504 if (!(sk->sk_shutdown & SEND_SHUTDOWN)) {
505 if (sk_stream_wspace(sk) >= sk_stream_min_wspace(sk)) {
506 mask |= POLLOUT | POLLWRNORM;
507 } else { /* send SIGIO later */
508 set_bit(SOCK_ASYNC_NOSPACE,
509 &sk->sk_socket->flags);
510 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
512 /* Race breaker. If space is freed after
513 * wspace test but before the flags are set,
514 * IO signal will be lost.
516 if (sk_stream_wspace(sk) >= sk_stream_min_wspace(sk))
517 mask |= POLLOUT | POLLWRNORM;
520 mask |= POLLOUT | POLLWRNORM;
522 if (tp->urg_data & TCP_URG_VALID)
525 /* This barrier is coupled with smp_wmb() in tcp_reset() */
532 EXPORT_SYMBOL(tcp_poll);
534 int tcp_ioctl(struct sock *sk, int cmd, unsigned long arg)
536 struct tcp_sock *tp = tcp_sk(sk);
542 if (sk->sk_state == TCP_LISTEN)
545 slow = lock_sock_fast(sk);
546 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
548 else if (sock_flag(sk, SOCK_URGINLINE) ||
550 before(tp->urg_seq, tp->copied_seq) ||
551 !before(tp->urg_seq, tp->rcv_nxt)) {
553 answ = tp->rcv_nxt - tp->copied_seq;
555 /* Subtract 1, if FIN was received */
556 if (answ && sock_flag(sk, SOCK_DONE))
559 answ = tp->urg_seq - tp->copied_seq;
560 unlock_sock_fast(sk, slow);
563 answ = tp->urg_data && tp->urg_seq == tp->copied_seq;
566 if (sk->sk_state == TCP_LISTEN)
569 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
572 answ = tp->write_seq - tp->snd_una;
575 if (sk->sk_state == TCP_LISTEN)
578 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
581 answ = tp->write_seq - tp->snd_nxt;
587 return put_user(answ, (int __user *)arg);
589 EXPORT_SYMBOL(tcp_ioctl);
591 static inline void tcp_mark_push(struct tcp_sock *tp, struct sk_buff *skb)
593 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
594 tp->pushed_seq = tp->write_seq;
597 static inline bool forced_push(const struct tcp_sock *tp)
599 return after(tp->write_seq, tp->pushed_seq + (tp->max_window >> 1));
602 static inline void skb_entail(struct sock *sk, struct sk_buff *skb)
604 struct tcp_sock *tp = tcp_sk(sk);
605 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
608 tcb->seq = tcb->end_seq = tp->write_seq;
609 tcb->tcp_flags = TCPHDR_ACK;
611 skb_header_release(skb);
612 tcp_add_write_queue_tail(sk, skb);
613 sk->sk_wmem_queued += skb->truesize;
614 sk_mem_charge(sk, skb->truesize);
615 if (tp->nonagle & TCP_NAGLE_PUSH)
616 tp->nonagle &= ~TCP_NAGLE_PUSH;
619 static inline void tcp_mark_urg(struct tcp_sock *tp, int flags)
622 tp->snd_up = tp->write_seq;
625 static inline void tcp_push(struct sock *sk, int flags, int mss_now,
628 if (tcp_send_head(sk)) {
629 struct tcp_sock *tp = tcp_sk(sk);
631 if (!(flags & MSG_MORE) || forced_push(tp))
632 tcp_mark_push(tp, tcp_write_queue_tail(sk));
634 tcp_mark_urg(tp, flags);
635 __tcp_push_pending_frames(sk, mss_now,
636 (flags & MSG_MORE) ? TCP_NAGLE_CORK : nonagle);
640 static int tcp_splice_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb,
641 unsigned int offset, size_t len)
643 struct tcp_splice_state *tss = rd_desc->arg.data;
646 ret = skb_splice_bits(skb, offset, tss->pipe, min(rd_desc->count, len),
649 rd_desc->count -= ret;
653 static int __tcp_splice_read(struct sock *sk, struct tcp_splice_state *tss)
655 /* Store TCP splice context information in read_descriptor_t. */
656 read_descriptor_t rd_desc = {
661 return tcp_read_sock(sk, &rd_desc, tcp_splice_data_recv);
665 * tcp_splice_read - splice data from TCP socket to a pipe
666 * @sock: socket to splice from
667 * @ppos: position (not valid)
668 * @pipe: pipe to splice to
669 * @len: number of bytes to splice
670 * @flags: splice modifier flags
673 * Will read pages from given socket and fill them into a pipe.
676 ssize_t tcp_splice_read(struct socket *sock, loff_t *ppos,
677 struct pipe_inode_info *pipe, size_t len,
680 struct sock *sk = sock->sk;
681 struct tcp_splice_state tss = {
690 sock_rps_record_flow(sk);
692 * We can't seek on a socket input
701 timeo = sock_rcvtimeo(sk, sock->file->f_flags & O_NONBLOCK);
703 ret = __tcp_splice_read(sk, &tss);
709 if (sock_flag(sk, SOCK_DONE))
712 ret = sock_error(sk);
715 if (sk->sk_shutdown & RCV_SHUTDOWN)
717 if (sk->sk_state == TCP_CLOSE) {
719 * This occurs when user tries to read
720 * from never connected socket.
722 if (!sock_flag(sk, SOCK_DONE))
730 sk_wait_data(sk, &timeo);
731 if (signal_pending(current)) {
732 ret = sock_intr_errno(timeo);
745 if (sk->sk_err || sk->sk_state == TCP_CLOSE ||
746 (sk->sk_shutdown & RCV_SHUTDOWN) ||
747 signal_pending(current))
758 EXPORT_SYMBOL(tcp_splice_read);
760 struct sk_buff *sk_stream_alloc_skb(struct sock *sk, int size, gfp_t gfp)
764 /* The TCP header must be at least 32-bit aligned. */
765 size = ALIGN(size, 4);
767 skb = alloc_skb_fclone(size + sk->sk_prot->max_header, gfp);
769 if (sk_wmem_schedule(sk, skb->truesize)) {
770 skb_reserve(skb, sk->sk_prot->max_header);
772 * Make sure that we have exactly size bytes
773 * available to the caller, no more, no less.
775 skb->reserved_tailroom = skb->end - skb->tail - size;
780 sk->sk_prot->enter_memory_pressure(sk);
781 sk_stream_moderate_sndbuf(sk);
786 static unsigned int tcp_xmit_size_goal(struct sock *sk, u32 mss_now,
789 struct tcp_sock *tp = tcp_sk(sk);
790 u32 xmit_size_goal, old_size_goal;
792 xmit_size_goal = mss_now;
794 if (large_allowed && sk_can_gso(sk)) {
797 /* Maybe we should/could use sk->sk_prot->max_header here ? */
798 hlen = inet_csk(sk)->icsk_af_ops->net_header_len +
799 inet_csk(sk)->icsk_ext_hdr_len +
802 /* Goal is to send at least one packet per ms,
803 * not one big TSO packet every 100 ms.
804 * This preserves ACK clocking and is consistent
805 * with tcp_tso_should_defer() heuristic.
807 gso_size = sk->sk_pacing_rate / (2 * MSEC_PER_SEC);
808 gso_size = max_t(u32, gso_size,
809 sysctl_tcp_min_tso_segs * mss_now);
811 xmit_size_goal = min_t(u32, gso_size,
812 sk->sk_gso_max_size - 1 - hlen);
814 xmit_size_goal = tcp_bound_to_half_wnd(tp, xmit_size_goal);
816 /* We try hard to avoid divides here */
817 old_size_goal = tp->xmit_size_goal_segs * mss_now;
819 if (likely(old_size_goal <= xmit_size_goal &&
820 old_size_goal + mss_now > xmit_size_goal)) {
821 xmit_size_goal = old_size_goal;
823 tp->xmit_size_goal_segs =
824 min_t(u16, xmit_size_goal / mss_now,
825 sk->sk_gso_max_segs);
826 xmit_size_goal = tp->xmit_size_goal_segs * mss_now;
830 return max(xmit_size_goal, mss_now);
833 static int tcp_send_mss(struct sock *sk, int *size_goal, int flags)
837 mss_now = tcp_current_mss(sk);
838 *size_goal = tcp_xmit_size_goal(sk, mss_now, !(flags & MSG_OOB));
843 static ssize_t do_tcp_sendpages(struct sock *sk, struct page *page, int offset,
844 size_t size, int flags)
846 struct tcp_sock *tp = tcp_sk(sk);
847 int mss_now, size_goal;
850 long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
852 /* Wait for a connection to finish. One exception is TCP Fast Open
853 * (passive side) where data is allowed to be sent before a connection
854 * is fully established.
856 if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) &&
857 !tcp_passive_fastopen(sk)) {
858 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
862 clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
864 mss_now = tcp_send_mss(sk, &size_goal, flags);
868 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
872 struct sk_buff *skb = tcp_write_queue_tail(sk);
876 if (!tcp_send_head(sk) || (copy = size_goal - skb->len) <= 0) {
878 if (!sk_stream_memory_free(sk))
879 goto wait_for_sndbuf;
881 skb = sk_stream_alloc_skb(sk, 0, sk->sk_allocation);
883 goto wait_for_memory;
892 i = skb_shinfo(skb)->nr_frags;
893 can_coalesce = skb_can_coalesce(skb, i, page, offset);
894 if (!can_coalesce && i >= MAX_SKB_FRAGS) {
895 tcp_mark_push(tp, skb);
898 if (!sk_wmem_schedule(sk, copy))
899 goto wait_for_memory;
902 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
905 skb_fill_page_desc(skb, i, page, offset, copy);
907 skb_shinfo(skb)->tx_flags |= SKBTX_SHARED_FRAG;
910 skb->data_len += copy;
911 skb->truesize += copy;
912 sk->sk_wmem_queued += copy;
913 sk_mem_charge(sk, copy);
914 skb->ip_summed = CHECKSUM_PARTIAL;
915 tp->write_seq += copy;
916 TCP_SKB_CB(skb)->end_seq += copy;
917 skb_shinfo(skb)->gso_segs = 0;
920 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
927 if (skb->len < size_goal || (flags & MSG_OOB))
930 if (forced_push(tp)) {
931 tcp_mark_push(tp, skb);
932 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
933 } else if (skb == tcp_send_head(sk))
934 tcp_push_one(sk, mss_now);
938 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
940 tcp_push(sk, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH);
942 if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
945 mss_now = tcp_send_mss(sk, &size_goal, flags);
949 if (copied && !(flags & MSG_SENDPAGE_NOTLAST))
950 tcp_push(sk, flags, mss_now, tp->nonagle);
957 return sk_stream_error(sk, flags, err);
960 int tcp_sendpage(struct sock *sk, struct page *page, int offset,
961 size_t size, int flags)
965 if (!(sk->sk_route_caps & NETIF_F_SG) ||
966 !(sk->sk_route_caps & NETIF_F_ALL_CSUM))
967 return sock_no_sendpage(sk->sk_socket, page, offset, size,
971 res = do_tcp_sendpages(sk, page, offset, size, flags);
975 EXPORT_SYMBOL(tcp_sendpage);
977 static inline int select_size(const struct sock *sk, bool sg)
979 const struct tcp_sock *tp = tcp_sk(sk);
980 int tmp = tp->mss_cache;
983 if (sk_can_gso(sk)) {
984 /* Small frames wont use a full page:
985 * Payload will immediately follow tcp header.
987 tmp = SKB_WITH_OVERHEAD(2048 - MAX_TCP_HEADER);
989 int pgbreak = SKB_MAX_HEAD(MAX_TCP_HEADER);
991 if (tmp >= pgbreak &&
992 tmp <= pgbreak + (MAX_SKB_FRAGS - 1) * PAGE_SIZE)
1000 void tcp_free_fastopen_req(struct tcp_sock *tp)
1002 if (tp->fastopen_req != NULL) {
1003 kfree(tp->fastopen_req);
1004 tp->fastopen_req = NULL;
1008 static int tcp_sendmsg_fastopen(struct sock *sk, struct msghdr *msg,
1009 int *copied, size_t size)
1011 struct tcp_sock *tp = tcp_sk(sk);
1014 if (!(sysctl_tcp_fastopen & TFO_CLIENT_ENABLE))
1016 if (tp->fastopen_req != NULL)
1017 return -EALREADY; /* Another Fast Open is in progress */
1019 tp->fastopen_req = kzalloc(sizeof(struct tcp_fastopen_request),
1021 if (unlikely(tp->fastopen_req == NULL))
1023 tp->fastopen_req->data = msg;
1024 tp->fastopen_req->size = size;
1026 flags = (msg->msg_flags & MSG_DONTWAIT) ? O_NONBLOCK : 0;
1027 err = __inet_stream_connect(sk->sk_socket, msg->msg_name,
1028 msg->msg_namelen, flags);
1029 *copied = tp->fastopen_req->copied;
1030 tcp_free_fastopen_req(tp);
1034 int tcp_sendmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg,
1038 struct tcp_sock *tp = tcp_sk(sk);
1039 struct sk_buff *skb;
1040 int iovlen, flags, err, copied = 0;
1041 int mss_now = 0, size_goal, copied_syn = 0, offset = 0;
1047 flags = msg->msg_flags;
1048 if (flags & MSG_FASTOPEN) {
1049 err = tcp_sendmsg_fastopen(sk, msg, &copied_syn, size);
1050 if (err == -EINPROGRESS && copied_syn > 0)
1054 offset = copied_syn;
1057 timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
1059 /* Wait for a connection to finish. One exception is TCP Fast Open
1060 * (passive side) where data is allowed to be sent before a connection
1061 * is fully established.
1063 if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) &&
1064 !tcp_passive_fastopen(sk)) {
1065 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
1069 if (unlikely(tp->repair)) {
1070 if (tp->repair_queue == TCP_RECV_QUEUE) {
1071 copied = tcp_send_rcvq(sk, msg, size);
1076 if (tp->repair_queue == TCP_NO_QUEUE)
1079 /* 'common' sending to sendq */
1082 /* This should be in poll */
1083 clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
1085 mss_now = tcp_send_mss(sk, &size_goal, flags);
1087 /* Ok commence sending. */
1088 iovlen = msg->msg_iovlen;
1093 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
1096 sg = !!(sk->sk_route_caps & NETIF_F_SG);
1098 while (--iovlen >= 0) {
1099 size_t seglen = iov->iov_len;
1100 unsigned char __user *from = iov->iov_base;
1103 if (unlikely(offset > 0)) { /* Skip bytes copied in SYN */
1104 if (offset >= seglen) {
1113 while (seglen > 0) {
1115 int max = size_goal;
1117 skb = tcp_write_queue_tail(sk);
1118 if (tcp_send_head(sk)) {
1119 if (skb->ip_summed == CHECKSUM_NONE)
1121 copy = max - skb->len;
1126 /* Allocate new segment. If the interface is SG,
1127 * allocate skb fitting to single page.
1129 if (!sk_stream_memory_free(sk))
1130 goto wait_for_sndbuf;
1132 skb = sk_stream_alloc_skb(sk,
1133 select_size(sk, sg),
1136 goto wait_for_memory;
1139 * All packets are restored as if they have
1140 * already been sent.
1143 TCP_SKB_CB(skb)->when = tcp_time_stamp;
1146 * Check whether we can use HW checksum.
1148 if (sk->sk_route_caps & NETIF_F_ALL_CSUM)
1149 skb->ip_summed = CHECKSUM_PARTIAL;
1151 skb_entail(sk, skb);
1156 /* Try to append data to the end of skb. */
1160 /* Where to copy to? */
1161 if (skb_availroom(skb) > 0) {
1162 /* We have some space in skb head. Superb! */
1163 copy = min_t(int, copy, skb_availroom(skb));
1164 err = skb_add_data_nocache(sk, skb, from, copy);
1169 int i = skb_shinfo(skb)->nr_frags;
1170 struct page_frag *pfrag = sk_page_frag(sk);
1172 if (!sk_page_frag_refill(sk, pfrag))
1173 goto wait_for_memory;
1175 if (!skb_can_coalesce(skb, i, pfrag->page,
1177 if (i == MAX_SKB_FRAGS || !sg) {
1178 tcp_mark_push(tp, skb);
1184 copy = min_t(int, copy, pfrag->size - pfrag->offset);
1186 if (!sk_wmem_schedule(sk, copy))
1187 goto wait_for_memory;
1189 err = skb_copy_to_page_nocache(sk, from, skb,
1196 /* Update the skb. */
1198 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1200 skb_fill_page_desc(skb, i, pfrag->page,
1201 pfrag->offset, copy);
1202 get_page(pfrag->page);
1204 pfrag->offset += copy;
1208 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
1210 tp->write_seq += copy;
1211 TCP_SKB_CB(skb)->end_seq += copy;
1212 skb_shinfo(skb)->gso_segs = 0;
1216 if ((seglen -= copy) == 0 && iovlen == 0)
1219 if (skb->len < max || (flags & MSG_OOB) || unlikely(tp->repair))
1222 if (forced_push(tp)) {
1223 tcp_mark_push(tp, skb);
1224 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
1225 } else if (skb == tcp_send_head(sk))
1226 tcp_push_one(sk, mss_now);
1230 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1233 tcp_push(sk, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH);
1235 if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
1238 mss_now = tcp_send_mss(sk, &size_goal, flags);
1244 tcp_push(sk, flags, mss_now, tp->nonagle);
1247 if (copied + copied_syn)
1248 uid_stat_tcp_snd(current_uid(), copied + copied_syn);
1249 return copied + copied_syn;
1253 tcp_unlink_write_queue(skb, sk);
1254 /* It is the one place in all of TCP, except connection
1255 * reset, where we can be unlinking the send_head.
1257 tcp_check_send_head(sk, skb);
1258 sk_wmem_free_skb(sk, skb);
1262 if (copied + copied_syn)
1265 err = sk_stream_error(sk, flags, err);
1269 EXPORT_SYMBOL(tcp_sendmsg);
1272 * Handle reading urgent data. BSD has very simple semantics for
1273 * this, no blocking and very strange errors 8)
1276 static int tcp_recv_urg(struct sock *sk, struct msghdr *msg, int len, int flags)
1278 struct tcp_sock *tp = tcp_sk(sk);
1280 /* No URG data to read. */
1281 if (sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data ||
1282 tp->urg_data == TCP_URG_READ)
1283 return -EINVAL; /* Yes this is right ! */
1285 if (sk->sk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DONE))
1288 if (tp->urg_data & TCP_URG_VALID) {
1290 char c = tp->urg_data;
1292 if (!(flags & MSG_PEEK))
1293 tp->urg_data = TCP_URG_READ;
1295 /* Read urgent data. */
1296 msg->msg_flags |= MSG_OOB;
1299 if (!(flags & MSG_TRUNC))
1300 err = memcpy_toiovec(msg->msg_iov, &c, 1);
1303 msg->msg_flags |= MSG_TRUNC;
1305 return err ? -EFAULT : len;
1308 if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
1311 /* Fixed the recv(..., MSG_OOB) behaviour. BSD docs and
1312 * the available implementations agree in this case:
1313 * this call should never block, independent of the
1314 * blocking state of the socket.
1315 * Mike <pall@rz.uni-karlsruhe.de>
1320 static int tcp_peek_sndq(struct sock *sk, struct msghdr *msg, int len)
1322 struct sk_buff *skb;
1323 int copied = 0, err = 0;
1325 /* XXX -- need to support SO_PEEK_OFF */
1327 skb_queue_walk(&sk->sk_write_queue, skb) {
1328 err = skb_copy_datagram_iovec(skb, 0, msg->msg_iov, skb->len);
1335 return err ?: copied;
1338 /* Clean up the receive buffer for full frames taken by the user,
1339 * then send an ACK if necessary. COPIED is the number of bytes
1340 * tcp_recvmsg has given to the user so far, it speeds up the
1341 * calculation of whether or not we must ACK for the sake of
1344 void tcp_cleanup_rbuf(struct sock *sk, int copied)
1346 struct tcp_sock *tp = tcp_sk(sk);
1347 bool time_to_ack = false;
1349 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
1351 WARN(skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq),
1352 "cleanup rbuf bug: copied %X seq %X rcvnxt %X\n",
1353 tp->copied_seq, TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt);
1355 if (inet_csk_ack_scheduled(sk)) {
1356 const struct inet_connection_sock *icsk = inet_csk(sk);
1357 /* Delayed ACKs frequently hit locked sockets during bulk
1359 if (icsk->icsk_ack.blocked ||
1360 /* Once-per-two-segments ACK was not sent by tcp_input.c */
1361 tp->rcv_nxt - tp->rcv_wup > icsk->icsk_ack.rcv_mss ||
1363 * If this read emptied read buffer, we send ACK, if
1364 * connection is not bidirectional, user drained
1365 * receive buffer and there was a small segment
1369 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED2) ||
1370 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED) &&
1371 !icsk->icsk_ack.pingpong)) &&
1372 !atomic_read(&sk->sk_rmem_alloc)))
1376 /* We send an ACK if we can now advertise a non-zero window
1377 * which has been raised "significantly".
1379 * Even if window raised up to infinity, do not send window open ACK
1380 * in states, where we will not receive more. It is useless.
1382 if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) {
1383 __u32 rcv_window_now = tcp_receive_window(tp);
1385 /* Optimize, __tcp_select_window() is not cheap. */
1386 if (2*rcv_window_now <= tp->window_clamp) {
1387 __u32 new_window = __tcp_select_window(sk);
1389 /* Send ACK now, if this read freed lots of space
1390 * in our buffer. Certainly, new_window is new window.
1391 * We can advertise it now, if it is not less than current one.
1392 * "Lots" means "at least twice" here.
1394 if (new_window && new_window >= 2 * rcv_window_now)
1402 static void tcp_prequeue_process(struct sock *sk)
1404 struct sk_buff *skb;
1405 struct tcp_sock *tp = tcp_sk(sk);
1407 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPPREQUEUED);
1409 /* RX process wants to run with disabled BHs, though it is not
1412 while ((skb = __skb_dequeue(&tp->ucopy.prequeue)) != NULL)
1413 sk_backlog_rcv(sk, skb);
1416 /* Clear memory counter. */
1417 tp->ucopy.memory = 0;
1420 #ifdef CONFIG_NET_DMA
1421 static void tcp_service_net_dma(struct sock *sk, bool wait)
1423 dma_cookie_t done, used;
1424 dma_cookie_t last_issued;
1425 struct tcp_sock *tp = tcp_sk(sk);
1427 if (!tp->ucopy.dma_chan)
1430 last_issued = tp->ucopy.dma_cookie;
1431 dma_async_issue_pending(tp->ucopy.dma_chan);
1434 if (dma_async_is_tx_complete(tp->ucopy.dma_chan,
1436 &used) == DMA_SUCCESS) {
1437 /* Safe to free early-copied skbs now */
1438 __skb_queue_purge(&sk->sk_async_wait_queue);
1441 struct sk_buff *skb;
1442 while ((skb = skb_peek(&sk->sk_async_wait_queue)) &&
1443 (dma_async_is_complete(skb->dma_cookie, done,
1444 used) == DMA_SUCCESS)) {
1445 __skb_dequeue(&sk->sk_async_wait_queue);
1453 static struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off)
1455 struct sk_buff *skb;
1458 while ((skb = skb_peek(&sk->sk_receive_queue)) != NULL) {
1459 offset = seq - TCP_SKB_CB(skb)->seq;
1460 if (tcp_hdr(skb)->syn)
1462 if (offset < skb->len || tcp_hdr(skb)->fin) {
1466 /* This looks weird, but this can happen if TCP collapsing
1467 * splitted a fat GRO packet, while we released socket lock
1468 * in skb_splice_bits()
1470 sk_eat_skb(sk, skb, false);
1476 * This routine provides an alternative to tcp_recvmsg() for routines
1477 * that would like to handle copying from skbuffs directly in 'sendfile'
1480 * - It is assumed that the socket was locked by the caller.
1481 * - The routine does not block.
1482 * - At present, there is no support for reading OOB data
1483 * or for 'peeking' the socket using this routine
1484 * (although both would be easy to implement).
1486 int tcp_read_sock(struct sock *sk, read_descriptor_t *desc,
1487 sk_read_actor_t recv_actor)
1489 struct sk_buff *skb;
1490 struct tcp_sock *tp = tcp_sk(sk);
1491 u32 seq = tp->copied_seq;
1495 if (sk->sk_state == TCP_LISTEN)
1497 while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1498 if (offset < skb->len) {
1502 len = skb->len - offset;
1503 /* Stop reading if we hit a patch of urgent data */
1505 u32 urg_offset = tp->urg_seq - seq;
1506 if (urg_offset < len)
1511 used = recv_actor(desc, skb, offset, len);
1516 } else if (used <= len) {
1521 /* If recv_actor drops the lock (e.g. TCP splice
1522 * receive) the skb pointer might be invalid when
1523 * getting here: tcp_collapse might have deleted it
1524 * while aggregating skbs from the socket queue.
1526 skb = tcp_recv_skb(sk, seq - 1, &offset);
1529 /* TCP coalescing might have appended data to the skb.
1530 * Try to splice more frags
1532 if (offset + 1 != skb->len)
1535 if (tcp_hdr(skb)->fin) {
1536 sk_eat_skb(sk, skb, false);
1540 sk_eat_skb(sk, skb, false);
1543 tp->copied_seq = seq;
1545 tp->copied_seq = seq;
1547 tcp_rcv_space_adjust(sk);
1549 /* Clean up data we have read: This will do ACK frames. */
1551 tcp_recv_skb(sk, seq, &offset);
1552 tcp_cleanup_rbuf(sk, copied);
1553 uid_stat_tcp_rcv(current_uid(), copied);
1557 EXPORT_SYMBOL(tcp_read_sock);
1560 * This routine copies from a sock struct into the user buffer.
1562 * Technical note: in 2.3 we work on _locked_ socket, so that
1563 * tricks with *seq access order and skb->users are not required.
1564 * Probably, code can be easily improved even more.
1567 int tcp_recvmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg,
1568 size_t len, int nonblock, int flags, int *addr_len)
1570 struct tcp_sock *tp = tcp_sk(sk);
1576 int target; /* Read at least this many bytes */
1578 struct task_struct *user_recv = NULL;
1579 bool copied_early = false;
1580 struct sk_buff *skb;
1586 if (sk->sk_state == TCP_LISTEN)
1589 timeo = sock_rcvtimeo(sk, nonblock);
1591 /* Urgent data needs to be handled specially. */
1592 if (flags & MSG_OOB)
1595 if (unlikely(tp->repair)) {
1597 if (!(flags & MSG_PEEK))
1600 if (tp->repair_queue == TCP_SEND_QUEUE)
1604 if (tp->repair_queue == TCP_NO_QUEUE)
1607 /* 'common' recv queue MSG_PEEK-ing */
1610 seq = &tp->copied_seq;
1611 if (flags & MSG_PEEK) {
1612 peek_seq = tp->copied_seq;
1616 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
1618 #ifdef CONFIG_NET_DMA
1619 tp->ucopy.dma_chan = NULL;
1621 skb = skb_peek_tail(&sk->sk_receive_queue);
1626 available = TCP_SKB_CB(skb)->seq + skb->len - (*seq);
1627 if ((available < target) &&
1628 (len > sysctl_tcp_dma_copybreak) && !(flags & MSG_PEEK) &&
1629 !sysctl_tcp_low_latency &&
1630 net_dma_find_channel()) {
1631 preempt_enable_no_resched();
1632 tp->ucopy.pinned_list =
1633 dma_pin_iovec_pages(msg->msg_iov, len);
1635 preempt_enable_no_resched();
1643 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
1644 if (tp->urg_data && tp->urg_seq == *seq) {
1647 if (signal_pending(current)) {
1648 copied = timeo ? sock_intr_errno(timeo) : -EAGAIN;
1653 /* Next get a buffer. */
1655 skb_queue_walk(&sk->sk_receive_queue, skb) {
1656 /* Now that we have two receive queues this
1659 if (WARN(before(*seq, TCP_SKB_CB(skb)->seq),
1660 "recvmsg bug: copied %X seq %X rcvnxt %X fl %X\n",
1661 *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt,
1665 offset = *seq - TCP_SKB_CB(skb)->seq;
1666 if (tcp_hdr(skb)->syn)
1668 if (offset < skb->len)
1670 if (tcp_hdr(skb)->fin)
1672 WARN(!(flags & MSG_PEEK),
1673 "recvmsg bug 2: copied %X seq %X rcvnxt %X fl %X\n",
1674 *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt, flags);
1677 /* Well, if we have backlog, try to process it now yet. */
1679 if (copied >= target && !sk->sk_backlog.tail)
1684 sk->sk_state == TCP_CLOSE ||
1685 (sk->sk_shutdown & RCV_SHUTDOWN) ||
1687 signal_pending(current))
1690 if (sock_flag(sk, SOCK_DONE))
1694 copied = sock_error(sk);
1698 if (sk->sk_shutdown & RCV_SHUTDOWN)
1701 if (sk->sk_state == TCP_CLOSE) {
1702 if (!sock_flag(sk, SOCK_DONE)) {
1703 /* This occurs when user tries to read
1704 * from never connected socket.
1717 if (signal_pending(current)) {
1718 copied = sock_intr_errno(timeo);
1723 tcp_cleanup_rbuf(sk, copied);
1725 if (!sysctl_tcp_low_latency && tp->ucopy.task == user_recv) {
1726 /* Install new reader */
1727 if (!user_recv && !(flags & (MSG_TRUNC | MSG_PEEK))) {
1728 user_recv = current;
1729 tp->ucopy.task = user_recv;
1730 tp->ucopy.iov = msg->msg_iov;
1733 tp->ucopy.len = len;
1735 WARN_ON(tp->copied_seq != tp->rcv_nxt &&
1736 !(flags & (MSG_PEEK | MSG_TRUNC)));
1738 /* Ugly... If prequeue is not empty, we have to
1739 * process it before releasing socket, otherwise
1740 * order will be broken at second iteration.
1741 * More elegant solution is required!!!
1743 * Look: we have the following (pseudo)queues:
1745 * 1. packets in flight
1750 * Each queue can be processed only if the next ones
1751 * are empty. At this point we have empty receive_queue.
1752 * But prequeue _can_ be not empty after 2nd iteration,
1753 * when we jumped to start of loop because backlog
1754 * processing added something to receive_queue.
1755 * We cannot release_sock(), because backlog contains
1756 * packets arrived _after_ prequeued ones.
1758 * Shortly, algorithm is clear --- to process all
1759 * the queues in order. We could make it more directly,
1760 * requeueing packets from backlog to prequeue, if
1761 * is not empty. It is more elegant, but eats cycles,
1764 if (!skb_queue_empty(&tp->ucopy.prequeue))
1767 /* __ Set realtime policy in scheduler __ */
1770 #ifdef CONFIG_NET_DMA
1771 if (tp->ucopy.dma_chan) {
1772 if (tp->rcv_wnd == 0 &&
1773 !skb_queue_empty(&sk->sk_async_wait_queue)) {
1774 tcp_service_net_dma(sk, true);
1775 tcp_cleanup_rbuf(sk, copied);
1777 dma_async_issue_pending(tp->ucopy.dma_chan);
1780 if (copied >= target) {
1781 /* Do not sleep, just process backlog. */
1785 sk_wait_data(sk, &timeo);
1787 #ifdef CONFIG_NET_DMA
1788 tcp_service_net_dma(sk, false); /* Don't block */
1789 tp->ucopy.wakeup = 0;
1795 /* __ Restore normal policy in scheduler __ */
1797 if ((chunk = len - tp->ucopy.len) != 0) {
1798 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMBACKLOG, chunk);
1803 if (tp->rcv_nxt == tp->copied_seq &&
1804 !skb_queue_empty(&tp->ucopy.prequeue)) {
1806 tcp_prequeue_process(sk);
1808 if ((chunk = len - tp->ucopy.len) != 0) {
1809 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1815 if ((flags & MSG_PEEK) &&
1816 (peek_seq - copied - urg_hole != tp->copied_seq)) {
1817 net_dbg_ratelimited("TCP(%s:%d): Application bug, race in MSG_PEEK\n",
1819 task_pid_nr(current));
1820 peek_seq = tp->copied_seq;
1825 /* Ok so how much can we use? */
1826 used = skb->len - offset;
1830 /* Do we have urgent data here? */
1832 u32 urg_offset = tp->urg_seq - *seq;
1833 if (urg_offset < used) {
1835 if (!sock_flag(sk, SOCK_URGINLINE)) {
1848 if (!(flags & MSG_TRUNC)) {
1849 #ifdef CONFIG_NET_DMA
1850 if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list)
1851 tp->ucopy.dma_chan = net_dma_find_channel();
1853 if (tp->ucopy.dma_chan) {
1854 tp->ucopy.dma_cookie = dma_skb_copy_datagram_iovec(
1855 tp->ucopy.dma_chan, skb, offset,
1857 tp->ucopy.pinned_list);
1859 if (tp->ucopy.dma_cookie < 0) {
1861 pr_alert("%s: dma_cookie < 0\n",
1864 /* Exception. Bailout! */
1870 dma_async_issue_pending(tp->ucopy.dma_chan);
1872 if ((offset + used) == skb->len)
1873 copied_early = true;
1878 err = skb_copy_datagram_iovec(skb, offset,
1879 msg->msg_iov, used);
1881 /* Exception. Bailout! */
1893 tcp_rcv_space_adjust(sk);
1896 if (tp->urg_data && after(tp->copied_seq, tp->urg_seq)) {
1898 tcp_fast_path_check(sk);
1900 if (used + offset < skb->len)
1903 if (tcp_hdr(skb)->fin)
1905 if (!(flags & MSG_PEEK)) {
1906 sk_eat_skb(sk, skb, copied_early);
1907 copied_early = false;
1912 /* Process the FIN. */
1914 if (!(flags & MSG_PEEK)) {
1915 sk_eat_skb(sk, skb, copied_early);
1916 copied_early = false;
1922 if (!skb_queue_empty(&tp->ucopy.prequeue)) {
1925 tp->ucopy.len = copied > 0 ? len : 0;
1927 tcp_prequeue_process(sk);
1929 if (copied > 0 && (chunk = len - tp->ucopy.len) != 0) {
1930 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1936 tp->ucopy.task = NULL;
1940 #ifdef CONFIG_NET_DMA
1941 tcp_service_net_dma(sk, true); /* Wait for queue to drain */
1942 tp->ucopy.dma_chan = NULL;
1944 if (tp->ucopy.pinned_list) {
1945 dma_unpin_iovec_pages(tp->ucopy.pinned_list);
1946 tp->ucopy.pinned_list = NULL;
1950 /* According to UNIX98, msg_name/msg_namelen are ignored
1951 * on connected socket. I was just happy when found this 8) --ANK
1954 /* Clean up data we have read: This will do ACK frames. */
1955 tcp_cleanup_rbuf(sk, copied);
1960 uid_stat_tcp_rcv(current_uid(), copied);
1968 err = tcp_recv_urg(sk, msg, len, flags);
1970 uid_stat_tcp_rcv(current_uid(), err);
1974 err = tcp_peek_sndq(sk, msg, len);
1977 EXPORT_SYMBOL(tcp_recvmsg);
1979 void tcp_set_state(struct sock *sk, int state)
1981 int oldstate = sk->sk_state;
1984 case TCP_ESTABLISHED:
1985 if (oldstate != TCP_ESTABLISHED)
1986 TCP_INC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
1990 if (oldstate == TCP_CLOSE_WAIT || oldstate == TCP_ESTABLISHED)
1991 TCP_INC_STATS(sock_net(sk), TCP_MIB_ESTABRESETS);
1993 sk->sk_prot->unhash(sk);
1994 if (inet_csk(sk)->icsk_bind_hash &&
1995 !(sk->sk_userlocks & SOCK_BINDPORT_LOCK))
1999 if (oldstate == TCP_ESTABLISHED)
2000 TCP_DEC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
2003 /* Change state AFTER socket is unhashed to avoid closed
2004 * socket sitting in hash tables.
2006 sk->sk_state = state;
2009 SOCK_DEBUG(sk, "TCP sk=%p, State %s -> %s\n", sk, statename[oldstate], statename[state]);
2012 EXPORT_SYMBOL_GPL(tcp_set_state);
2015 * State processing on a close. This implements the state shift for
2016 * sending our FIN frame. Note that we only send a FIN for some
2017 * states. A shutdown() may have already sent the FIN, or we may be
2021 static const unsigned char new_state[16] = {
2022 /* current state: new state: action: */
2023 /* (Invalid) */ TCP_CLOSE,
2024 /* TCP_ESTABLISHED */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2025 /* TCP_SYN_SENT */ TCP_CLOSE,
2026 /* TCP_SYN_RECV */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2027 /* TCP_FIN_WAIT1 */ TCP_FIN_WAIT1,
2028 /* TCP_FIN_WAIT2 */ TCP_FIN_WAIT2,
2029 /* TCP_TIME_WAIT */ TCP_CLOSE,
2030 /* TCP_CLOSE */ TCP_CLOSE,
2031 /* TCP_CLOSE_WAIT */ TCP_LAST_ACK | TCP_ACTION_FIN,
2032 /* TCP_LAST_ACK */ TCP_LAST_ACK,
2033 /* TCP_LISTEN */ TCP_CLOSE,
2034 /* TCP_CLOSING */ TCP_CLOSING,
2037 static int tcp_close_state(struct sock *sk)
2039 int next = (int)new_state[sk->sk_state];
2040 int ns = next & TCP_STATE_MASK;
2042 tcp_set_state(sk, ns);
2044 return next & TCP_ACTION_FIN;
2048 * Shutdown the sending side of a connection. Much like close except
2049 * that we don't receive shut down or sock_set_flag(sk, SOCK_DEAD).
2052 void tcp_shutdown(struct sock *sk, int how)
2054 /* We need to grab some memory, and put together a FIN,
2055 * and then put it into the queue to be sent.
2056 * Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
2058 if (!(how & SEND_SHUTDOWN))
2061 /* If we've already sent a FIN, or it's a closed state, skip this. */
2062 if ((1 << sk->sk_state) &
2063 (TCPF_ESTABLISHED | TCPF_SYN_SENT |
2064 TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) {
2065 /* Clear out any half completed packets. FIN if needed. */
2066 if (tcp_close_state(sk))
2070 EXPORT_SYMBOL(tcp_shutdown);
2072 bool tcp_check_oom(struct sock *sk, int shift)
2074 bool too_many_orphans, out_of_socket_memory;
2076 too_many_orphans = tcp_too_many_orphans(sk, shift);
2077 out_of_socket_memory = tcp_out_of_memory(sk);
2079 if (too_many_orphans)
2080 net_info_ratelimited("too many orphaned sockets\n");
2081 if (out_of_socket_memory)
2082 net_info_ratelimited("out of memory -- consider tuning tcp_mem\n");
2083 return too_many_orphans || out_of_socket_memory;
2086 void tcp_close(struct sock *sk, long timeout)
2088 struct sk_buff *skb;
2089 int data_was_unread = 0;
2093 sk->sk_shutdown = SHUTDOWN_MASK;
2095 if (sk->sk_state == TCP_LISTEN) {
2096 tcp_set_state(sk, TCP_CLOSE);
2099 inet_csk_listen_stop(sk);
2101 goto adjudge_to_death;
2104 /* We need to flush the recv. buffs. We do this only on the
2105 * descriptor close, not protocol-sourced closes, because the
2106 * reader process may not have drained the data yet!
2108 while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) {
2109 u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq -
2111 data_was_unread += len;
2117 /* If socket has been already reset (e.g. in tcp_reset()) - kill it. */
2118 if (sk->sk_state == TCP_CLOSE)
2119 goto adjudge_to_death;
2121 /* As outlined in RFC 2525, section 2.17, we send a RST here because
2122 * data was lost. To witness the awful effects of the old behavior of
2123 * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk
2124 * GET in an FTP client, suspend the process, wait for the client to
2125 * advertise a zero window, then kill -9 the FTP client, wheee...
2126 * Note: timeout is always zero in such a case.
2128 if (unlikely(tcp_sk(sk)->repair)) {
2129 sk->sk_prot->disconnect(sk, 0);
2130 } else if (data_was_unread) {
2131 /* Unread data was tossed, zap the connection. */
2132 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONCLOSE);
2133 tcp_set_state(sk, TCP_CLOSE);
2134 tcp_send_active_reset(sk, sk->sk_allocation);
2135 } else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) {
2136 /* Check zero linger _after_ checking for unread data. */
2137 sk->sk_prot->disconnect(sk, 0);
2138 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
2139 } else if (tcp_close_state(sk)) {
2140 /* We FIN if the application ate all the data before
2141 * zapping the connection.
2144 /* RED-PEN. Formally speaking, we have broken TCP state
2145 * machine. State transitions:
2147 * TCP_ESTABLISHED -> TCP_FIN_WAIT1
2148 * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible)
2149 * TCP_CLOSE_WAIT -> TCP_LAST_ACK
2151 * are legal only when FIN has been sent (i.e. in window),
2152 * rather than queued out of window. Purists blame.
2154 * F.e. "RFC state" is ESTABLISHED,
2155 * if Linux state is FIN-WAIT-1, but FIN is still not sent.
2157 * The visible declinations are that sometimes
2158 * we enter time-wait state, when it is not required really
2159 * (harmless), do not send active resets, when they are
2160 * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
2161 * they look as CLOSING or LAST_ACK for Linux)
2162 * Probably, I missed some more holelets.
2164 * XXX (TFO) - To start off we don't support SYN+ACK+FIN
2165 * in a single packet! (May consider it later but will
2166 * probably need API support or TCP_CORK SYN-ACK until
2167 * data is written and socket is closed.)
2172 sk_stream_wait_close(sk, timeout);
2175 state = sk->sk_state;
2179 /* It is the last release_sock in its life. It will remove backlog. */
2183 /* Now socket is owned by kernel and we acquire BH lock
2184 to finish close. No need to check for user refs.
2188 WARN_ON(sock_owned_by_user(sk));
2190 percpu_counter_inc(sk->sk_prot->orphan_count);
2192 /* Have we already been destroyed by a softirq or backlog? */
2193 if (state != TCP_CLOSE && sk->sk_state == TCP_CLOSE)
2196 /* This is a (useful) BSD violating of the RFC. There is a
2197 * problem with TCP as specified in that the other end could
2198 * keep a socket open forever with no application left this end.
2199 * We use a 3 minute timeout (about the same as BSD) then kill
2200 * our end. If they send after that then tough - BUT: long enough
2201 * that we won't make the old 4*rto = almost no time - whoops
2204 * Nope, it was not mistake. It is really desired behaviour
2205 * f.e. on http servers, when such sockets are useless, but
2206 * consume significant resources. Let's do it with special
2207 * linger2 option. --ANK
2210 if (sk->sk_state == TCP_FIN_WAIT2) {
2211 struct tcp_sock *tp = tcp_sk(sk);
2212 if (tp->linger2 < 0) {
2213 tcp_set_state(sk, TCP_CLOSE);
2214 tcp_send_active_reset(sk, GFP_ATOMIC);
2215 NET_INC_STATS_BH(sock_net(sk),
2216 LINUX_MIB_TCPABORTONLINGER);
2218 const int tmo = tcp_fin_time(sk);
2220 if (tmo > TCP_TIMEWAIT_LEN) {
2221 inet_csk_reset_keepalive_timer(sk,
2222 tmo - TCP_TIMEWAIT_LEN);
2224 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
2229 if (sk->sk_state != TCP_CLOSE) {
2231 if (tcp_check_oom(sk, 0)) {
2232 tcp_set_state(sk, TCP_CLOSE);
2233 tcp_send_active_reset(sk, GFP_ATOMIC);
2234 NET_INC_STATS_BH(sock_net(sk),
2235 LINUX_MIB_TCPABORTONMEMORY);
2239 if (sk->sk_state == TCP_CLOSE) {
2240 struct request_sock *req = tcp_sk(sk)->fastopen_rsk;
2241 /* We could get here with a non-NULL req if the socket is
2242 * aborted (e.g., closed with unread data) before 3WHS
2246 reqsk_fastopen_remove(sk, req, false);
2247 inet_csk_destroy_sock(sk);
2249 /* Otherwise, socket is reprieved until protocol close. */
2256 EXPORT_SYMBOL(tcp_close);
2258 /* These states need RST on ABORT according to RFC793 */
2260 static inline bool tcp_need_reset(int state)
2262 return (1 << state) &
2263 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 |
2264 TCPF_FIN_WAIT2 | TCPF_SYN_RECV);
2267 int tcp_disconnect(struct sock *sk, int flags)
2269 struct inet_sock *inet = inet_sk(sk);
2270 struct inet_connection_sock *icsk = inet_csk(sk);
2271 struct tcp_sock *tp = tcp_sk(sk);
2273 int old_state = sk->sk_state;
2275 if (old_state != TCP_CLOSE)
2276 tcp_set_state(sk, TCP_CLOSE);
2278 /* ABORT function of RFC793 */
2279 if (old_state == TCP_LISTEN) {
2280 inet_csk_listen_stop(sk);
2281 } else if (unlikely(tp->repair)) {
2282 sk->sk_err = ECONNABORTED;
2283 } else if (tcp_need_reset(old_state) ||
2284 (tp->snd_nxt != tp->write_seq &&
2285 (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK))) {
2286 /* The last check adjusts for discrepancy of Linux wrt. RFC
2289 tcp_send_active_reset(sk, gfp_any());
2290 sk->sk_err = ECONNRESET;
2291 } else if (old_state == TCP_SYN_SENT)
2292 sk->sk_err = ECONNRESET;
2294 tcp_clear_xmit_timers(sk);
2295 __skb_queue_purge(&sk->sk_receive_queue);
2296 tcp_write_queue_purge(sk);
2297 __skb_queue_purge(&tp->out_of_order_queue);
2298 #ifdef CONFIG_NET_DMA
2299 __skb_queue_purge(&sk->sk_async_wait_queue);
2302 inet->inet_dport = 0;
2304 if (!(sk->sk_userlocks & SOCK_BINDADDR_LOCK))
2305 inet_reset_saddr(sk);
2307 sk->sk_shutdown = 0;
2308 sock_reset_flag(sk, SOCK_DONE);
2310 if ((tp->write_seq += tp->max_window + 2) == 0)
2312 icsk->icsk_backoff = 0;
2314 icsk->icsk_probes_out = 0;
2315 tp->packets_out = 0;
2316 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
2317 tp->snd_cwnd_cnt = 0;
2318 tp->window_clamp = 0;
2319 tcp_set_ca_state(sk, TCP_CA_Open);
2320 tcp_clear_retrans(tp);
2321 inet_csk_delack_init(sk);
2322 tcp_init_send_head(sk);
2323 memset(&tp->rx_opt, 0, sizeof(tp->rx_opt));
2326 WARN_ON(inet->inet_num && !icsk->icsk_bind_hash);
2328 sk->sk_error_report(sk);
2331 EXPORT_SYMBOL(tcp_disconnect);
2333 void tcp_sock_destruct(struct sock *sk)
2335 inet_sock_destruct(sk);
2337 kfree(inet_csk(sk)->icsk_accept_queue.fastopenq);
2340 static inline bool tcp_can_repair_sock(const struct sock *sk)
2342 return ns_capable(sock_net(sk)->user_ns, CAP_NET_ADMIN) &&
2343 ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_ESTABLISHED));
2346 static int tcp_repair_options_est(struct tcp_sock *tp,
2347 struct tcp_repair_opt __user *optbuf, unsigned int len)
2349 struct tcp_repair_opt opt;
2351 while (len >= sizeof(opt)) {
2352 if (copy_from_user(&opt, optbuf, sizeof(opt)))
2358 switch (opt.opt_code) {
2360 tp->rx_opt.mss_clamp = opt.opt_val;
2364 u16 snd_wscale = opt.opt_val & 0xFFFF;
2365 u16 rcv_wscale = opt.opt_val >> 16;
2367 if (snd_wscale > 14 || rcv_wscale > 14)
2370 tp->rx_opt.snd_wscale = snd_wscale;
2371 tp->rx_opt.rcv_wscale = rcv_wscale;
2372 tp->rx_opt.wscale_ok = 1;
2375 case TCPOPT_SACK_PERM:
2376 if (opt.opt_val != 0)
2379 tp->rx_opt.sack_ok |= TCP_SACK_SEEN;
2380 if (sysctl_tcp_fack)
2381 tcp_enable_fack(tp);
2383 case TCPOPT_TIMESTAMP:
2384 if (opt.opt_val != 0)
2387 tp->rx_opt.tstamp_ok = 1;
2396 * Socket option code for TCP.
2398 static int do_tcp_setsockopt(struct sock *sk, int level,
2399 int optname, char __user *optval, unsigned int optlen)
2401 struct tcp_sock *tp = tcp_sk(sk);
2402 struct inet_connection_sock *icsk = inet_csk(sk);
2406 /* These are data/string values, all the others are ints */
2408 case TCP_CONGESTION: {
2409 char name[TCP_CA_NAME_MAX];
2414 val = strncpy_from_user(name, optval,
2415 min_t(long, TCP_CA_NAME_MAX-1, optlen));
2421 err = tcp_set_congestion_control(sk, name);
2430 if (optlen < sizeof(int))
2433 if (get_user(val, (int __user *)optval))
2440 /* Values greater than interface MTU won't take effect. However
2441 * at the point when this call is done we typically don't yet
2442 * know which interface is going to be used */
2443 if (val < TCP_MIN_MSS || val > MAX_TCP_WINDOW) {
2447 tp->rx_opt.user_mss = val;
2452 /* TCP_NODELAY is weaker than TCP_CORK, so that
2453 * this option on corked socket is remembered, but
2454 * it is not activated until cork is cleared.
2456 * However, when TCP_NODELAY is set we make
2457 * an explicit push, which overrides even TCP_CORK
2458 * for currently queued segments.
2460 tp->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH;
2461 tcp_push_pending_frames(sk);
2463 tp->nonagle &= ~TCP_NAGLE_OFF;
2467 case TCP_THIN_LINEAR_TIMEOUTS:
2468 if (val < 0 || val > 1)
2474 case TCP_THIN_DUPACK:
2475 if (val < 0 || val > 1)
2478 tp->thin_dupack = val;
2479 if (tp->thin_dupack)
2480 tcp_disable_early_retrans(tp);
2485 if (!tcp_can_repair_sock(sk))
2487 else if (val == 1) {
2489 sk->sk_reuse = SK_FORCE_REUSE;
2490 tp->repair_queue = TCP_NO_QUEUE;
2491 } else if (val == 0) {
2493 sk->sk_reuse = SK_NO_REUSE;
2494 tcp_send_window_probe(sk);
2500 case TCP_REPAIR_QUEUE:
2503 else if (val < TCP_QUEUES_NR)
2504 tp->repair_queue = val;
2510 if (sk->sk_state != TCP_CLOSE)
2512 else if (tp->repair_queue == TCP_SEND_QUEUE)
2513 tp->write_seq = val;
2514 else if (tp->repair_queue == TCP_RECV_QUEUE)
2520 case TCP_REPAIR_OPTIONS:
2523 else if (sk->sk_state == TCP_ESTABLISHED)
2524 err = tcp_repair_options_est(tp,
2525 (struct tcp_repair_opt __user *)optval,
2532 /* When set indicates to always queue non-full frames.
2533 * Later the user clears this option and we transmit
2534 * any pending partial frames in the queue. This is
2535 * meant to be used alongside sendfile() to get properly
2536 * filled frames when the user (for example) must write
2537 * out headers with a write() call first and then use
2538 * sendfile to send out the data parts.
2540 * TCP_CORK can be set together with TCP_NODELAY and it is
2541 * stronger than TCP_NODELAY.
2544 tp->nonagle |= TCP_NAGLE_CORK;
2546 tp->nonagle &= ~TCP_NAGLE_CORK;
2547 if (tp->nonagle&TCP_NAGLE_OFF)
2548 tp->nonagle |= TCP_NAGLE_PUSH;
2549 tcp_push_pending_frames(sk);
2554 if (val < 1 || val > MAX_TCP_KEEPIDLE)
2557 tp->keepalive_time = val * HZ;
2558 if (sock_flag(sk, SOCK_KEEPOPEN) &&
2559 !((1 << sk->sk_state) &
2560 (TCPF_CLOSE | TCPF_LISTEN))) {
2561 u32 elapsed = keepalive_time_elapsed(tp);
2562 if (tp->keepalive_time > elapsed)
2563 elapsed = tp->keepalive_time - elapsed;
2566 inet_csk_reset_keepalive_timer(sk, elapsed);
2571 if (val < 1 || val > MAX_TCP_KEEPINTVL)
2574 tp->keepalive_intvl = val * HZ;
2577 if (val < 1 || val > MAX_TCP_KEEPCNT)
2580 tp->keepalive_probes = val;
2583 if (val < 1 || val > MAX_TCP_SYNCNT)
2586 icsk->icsk_syn_retries = val;
2592 else if (val > sysctl_tcp_fin_timeout / HZ)
2595 tp->linger2 = val * HZ;
2598 case TCP_DEFER_ACCEPT:
2599 /* Translate value in seconds to number of retransmits */
2600 icsk->icsk_accept_queue.rskq_defer_accept =
2601 secs_to_retrans(val, TCP_TIMEOUT_INIT / HZ,
2605 case TCP_WINDOW_CLAMP:
2607 if (sk->sk_state != TCP_CLOSE) {
2611 tp->window_clamp = 0;
2613 tp->window_clamp = val < SOCK_MIN_RCVBUF / 2 ?
2614 SOCK_MIN_RCVBUF / 2 : val;
2619 icsk->icsk_ack.pingpong = 1;
2621 icsk->icsk_ack.pingpong = 0;
2622 if ((1 << sk->sk_state) &
2623 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) &&
2624 inet_csk_ack_scheduled(sk)) {
2625 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
2626 tcp_cleanup_rbuf(sk, 1);
2628 icsk->icsk_ack.pingpong = 1;
2633 #ifdef CONFIG_TCP_MD5SIG
2635 /* Read the IP->Key mappings from userspace */
2636 err = tp->af_specific->md5_parse(sk, optval, optlen);
2639 case TCP_USER_TIMEOUT:
2640 /* Cap the max timeout in ms TCP will retry/retrans
2641 * before giving up and aborting (ETIMEDOUT) a connection.
2646 icsk->icsk_user_timeout = msecs_to_jiffies(val);
2650 if (val >= 0 && ((1 << sk->sk_state) & (TCPF_CLOSE |
2652 err = fastopen_init_queue(sk, val);
2660 tp->tsoffset = val - tcp_time_stamp;
2671 int tcp_setsockopt(struct sock *sk, int level, int optname, char __user *optval,
2672 unsigned int optlen)
2674 const struct inet_connection_sock *icsk = inet_csk(sk);
2676 if (level != SOL_TCP)
2677 return icsk->icsk_af_ops->setsockopt(sk, level, optname,
2679 return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2681 EXPORT_SYMBOL(tcp_setsockopt);
2683 #ifdef CONFIG_COMPAT
2684 int compat_tcp_setsockopt(struct sock *sk, int level, int optname,
2685 char __user *optval, unsigned int optlen)
2687 if (level != SOL_TCP)
2688 return inet_csk_compat_setsockopt(sk, level, optname,
2690 return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2692 EXPORT_SYMBOL(compat_tcp_setsockopt);
2695 /* Return information about state of tcp endpoint in API format. */
2696 void tcp_get_info(const struct sock *sk, struct tcp_info *info)
2698 const struct tcp_sock *tp = tcp_sk(sk);
2699 const struct inet_connection_sock *icsk = inet_csk(sk);
2700 u32 now = tcp_time_stamp;
2702 memset(info, 0, sizeof(*info));
2704 info->tcpi_state = sk->sk_state;
2705 info->tcpi_ca_state = icsk->icsk_ca_state;
2706 info->tcpi_retransmits = icsk->icsk_retransmits;
2707 info->tcpi_probes = icsk->icsk_probes_out;
2708 info->tcpi_backoff = icsk->icsk_backoff;
2710 if (tp->rx_opt.tstamp_ok)
2711 info->tcpi_options |= TCPI_OPT_TIMESTAMPS;
2712 if (tcp_is_sack(tp))
2713 info->tcpi_options |= TCPI_OPT_SACK;
2714 if (tp->rx_opt.wscale_ok) {
2715 info->tcpi_options |= TCPI_OPT_WSCALE;
2716 info->tcpi_snd_wscale = tp->rx_opt.snd_wscale;
2717 info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale;
2720 if (tp->ecn_flags & TCP_ECN_OK)
2721 info->tcpi_options |= TCPI_OPT_ECN;
2722 if (tp->ecn_flags & TCP_ECN_SEEN)
2723 info->tcpi_options |= TCPI_OPT_ECN_SEEN;
2724 if (tp->syn_data_acked)
2725 info->tcpi_options |= TCPI_OPT_SYN_DATA;
2727 info->tcpi_rto = jiffies_to_usecs(icsk->icsk_rto);
2728 info->tcpi_ato = jiffies_to_usecs(icsk->icsk_ack.ato);
2729 info->tcpi_snd_mss = tp->mss_cache;
2730 info->tcpi_rcv_mss = icsk->icsk_ack.rcv_mss;
2732 if (sk->sk_state == TCP_LISTEN) {
2733 info->tcpi_unacked = sk->sk_ack_backlog;
2734 info->tcpi_sacked = sk->sk_max_ack_backlog;
2736 info->tcpi_unacked = tp->packets_out;
2737 info->tcpi_sacked = tp->sacked_out;
2739 info->tcpi_lost = tp->lost_out;
2740 info->tcpi_retrans = tp->retrans_out;
2741 info->tcpi_fackets = tp->fackets_out;
2743 info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime);
2744 info->tcpi_last_data_recv = jiffies_to_msecs(now - icsk->icsk_ack.lrcvtime);
2745 info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp);
2747 info->tcpi_pmtu = icsk->icsk_pmtu_cookie;
2748 info->tcpi_rcv_ssthresh = tp->rcv_ssthresh;
2749 info->tcpi_rtt = jiffies_to_usecs(tp->srtt)>>3;
2750 info->tcpi_rttvar = jiffies_to_usecs(tp->mdev)>>2;
2751 info->tcpi_snd_ssthresh = tp->snd_ssthresh;
2752 info->tcpi_snd_cwnd = tp->snd_cwnd;
2753 info->tcpi_advmss = tp->advmss;
2754 info->tcpi_reordering = tp->reordering;
2756 info->tcpi_rcv_rtt = jiffies_to_usecs(tp->rcv_rtt_est.rtt)>>3;
2757 info->tcpi_rcv_space = tp->rcvq_space.space;
2759 info->tcpi_total_retrans = tp->total_retrans;
2761 EXPORT_SYMBOL_GPL(tcp_get_info);
2763 static int do_tcp_getsockopt(struct sock *sk, int level,
2764 int optname, char __user *optval, int __user *optlen)
2766 struct inet_connection_sock *icsk = inet_csk(sk);
2767 struct tcp_sock *tp = tcp_sk(sk);
2770 if (get_user(len, optlen))
2773 len = min_t(unsigned int, len, sizeof(int));
2780 val = tp->mss_cache;
2781 if (!val && ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
2782 val = tp->rx_opt.user_mss;
2784 val = tp->rx_opt.mss_clamp;
2787 val = !!(tp->nonagle&TCP_NAGLE_OFF);
2790 val = !!(tp->nonagle&TCP_NAGLE_CORK);
2793 val = keepalive_time_when(tp) / HZ;
2796 val = keepalive_intvl_when(tp) / HZ;
2799 val = keepalive_probes(tp);
2802 val = icsk->icsk_syn_retries ? : sysctl_tcp_syn_retries;
2807 val = (val ? : sysctl_tcp_fin_timeout) / HZ;
2809 case TCP_DEFER_ACCEPT:
2810 val = retrans_to_secs(icsk->icsk_accept_queue.rskq_defer_accept,
2811 TCP_TIMEOUT_INIT / HZ, TCP_RTO_MAX / HZ);
2813 case TCP_WINDOW_CLAMP:
2814 val = tp->window_clamp;
2817 struct tcp_info info;
2819 if (get_user(len, optlen))
2822 tcp_get_info(sk, &info);
2824 len = min_t(unsigned int, len, sizeof(info));
2825 if (put_user(len, optlen))
2827 if (copy_to_user(optval, &info, len))
2832 val = !icsk->icsk_ack.pingpong;
2835 case TCP_CONGESTION:
2836 if (get_user(len, optlen))
2838 len = min_t(unsigned int, len, TCP_CA_NAME_MAX);
2839 if (put_user(len, optlen))
2841 if (copy_to_user(optval, icsk->icsk_ca_ops->name, len))
2845 case TCP_THIN_LINEAR_TIMEOUTS:
2848 case TCP_THIN_DUPACK:
2849 val = tp->thin_dupack;
2856 case TCP_REPAIR_QUEUE:
2858 val = tp->repair_queue;
2864 if (tp->repair_queue == TCP_SEND_QUEUE)
2865 val = tp->write_seq;
2866 else if (tp->repair_queue == TCP_RECV_QUEUE)
2872 case TCP_USER_TIMEOUT:
2873 val = jiffies_to_msecs(icsk->icsk_user_timeout);
2876 val = tcp_time_stamp + tp->tsoffset;
2879 return -ENOPROTOOPT;
2882 if (put_user(len, optlen))
2884 if (copy_to_user(optval, &val, len))
2889 int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval,
2892 struct inet_connection_sock *icsk = inet_csk(sk);
2894 if (level != SOL_TCP)
2895 return icsk->icsk_af_ops->getsockopt(sk, level, optname,
2897 return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2899 EXPORT_SYMBOL(tcp_getsockopt);
2901 #ifdef CONFIG_COMPAT
2902 int compat_tcp_getsockopt(struct sock *sk, int level, int optname,
2903 char __user *optval, int __user *optlen)
2905 if (level != SOL_TCP)
2906 return inet_csk_compat_getsockopt(sk, level, optname,
2908 return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2910 EXPORT_SYMBOL(compat_tcp_getsockopt);
2913 struct sk_buff *tcp_tso_segment(struct sk_buff *skb,
2914 netdev_features_t features)
2916 struct sk_buff *segs = ERR_PTR(-EINVAL);
2917 unsigned int sum_truesize = 0;
2922 unsigned int oldlen;
2924 struct sk_buff *gso_skb = skb;
2926 bool ooo_okay, copy_destructor;
2928 if (!pskb_may_pull(skb, sizeof(*th)))
2932 thlen = th->doff * 4;
2933 if (thlen < sizeof(*th))
2936 if (!pskb_may_pull(skb, thlen))
2939 oldlen = (u16)~skb->len;
2940 __skb_pull(skb, thlen);
2942 mss = skb_shinfo(skb)->gso_size;
2943 if (unlikely(skb->len <= mss))
2946 if (skb_gso_ok(skb, features | NETIF_F_GSO_ROBUST)) {
2947 /* Packet is from an untrusted source, reset gso_segs. */
2948 int type = skb_shinfo(skb)->gso_type;
2956 SKB_GSO_UDP_TUNNEL |
2958 !(type & (SKB_GSO_TCPV4 | SKB_GSO_TCPV6))))
2961 skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss);
2967 copy_destructor = gso_skb->destructor == tcp_wfree;
2968 ooo_okay = gso_skb->ooo_okay;
2969 /* All segments but the first should have ooo_okay cleared */
2972 segs = skb_segment(skb, features);
2976 /* Only first segment might have ooo_okay set */
2977 segs->ooo_okay = ooo_okay;
2979 delta = htonl(oldlen + (thlen + mss));
2983 seq = ntohl(th->seq);
2985 newcheck = ~csum_fold((__force __wsum)((__force u32)th->check +
2986 (__force u32)delta));
2989 th->fin = th->psh = 0;
2990 th->check = newcheck;
2992 if (skb->ip_summed != CHECKSUM_PARTIAL)
2994 csum_fold(csum_partial(skb_transport_header(skb),
2998 if (copy_destructor) {
2999 skb->destructor = gso_skb->destructor;
3000 skb->sk = gso_skb->sk;
3001 sum_truesize += skb->truesize;
3006 th->seq = htonl(seq);
3008 } while (skb->next);
3010 /* Following permits TCP Small Queues to work well with GSO :
3011 * The callback to TCP stack will be called at the time last frag
3012 * is freed at TX completion, and not right now when gso_skb
3013 * is freed by GSO engine
3015 if (copy_destructor) {
3016 swap(gso_skb->sk, skb->sk);
3017 swap(gso_skb->destructor, skb->destructor);
3018 sum_truesize += skb->truesize;
3019 atomic_add(sum_truesize - gso_skb->truesize,
3020 &skb->sk->sk_wmem_alloc);
3023 delta = htonl(oldlen + (skb->tail - skb->transport_header) +
3025 th->check = ~csum_fold((__force __wsum)((__force u32)th->check +
3026 (__force u32)delta));
3027 if (skb->ip_summed != CHECKSUM_PARTIAL)
3028 th->check = csum_fold(csum_partial(skb_transport_header(skb),
3034 EXPORT_SYMBOL(tcp_tso_segment);
3036 struct sk_buff **tcp_gro_receive(struct sk_buff **head, struct sk_buff *skb)
3038 struct sk_buff **pp = NULL;
3045 unsigned int mss = 1;
3051 off = skb_gro_offset(skb);
3052 hlen = off + sizeof(*th);
3053 th = skb_gro_header_fast(skb, off);
3054 if (skb_gro_header_hard(skb, hlen)) {
3055 th = skb_gro_header_slow(skb, hlen, off);
3060 thlen = th->doff * 4;
3061 if (thlen < sizeof(*th))
3065 if (skb_gro_header_hard(skb, hlen)) {
3066 th = skb_gro_header_slow(skb, hlen, off);
3071 skb_gro_pull(skb, thlen);
3073 len = skb_gro_len(skb);
3074 flags = tcp_flag_word(th);
3076 for (; (p = *head); head = &p->next) {
3077 if (!NAPI_GRO_CB(p)->same_flow)
3082 if (*(u32 *)&th->source ^ *(u32 *)&th2->source) {
3083 NAPI_GRO_CB(p)->same_flow = 0;
3090 goto out_check_final;
3093 flush = NAPI_GRO_CB(p)->flush;
3094 flush |= (__force int)(flags & TCP_FLAG_CWR);
3095 flush |= (__force int)((flags ^ tcp_flag_word(th2)) &
3096 ~(TCP_FLAG_CWR | TCP_FLAG_FIN | TCP_FLAG_PSH));
3097 flush |= (__force int)(th->ack_seq ^ th2->ack_seq);
3098 for (i = sizeof(*th); i < thlen; i += 4)
3099 flush |= *(u32 *)((u8 *)th + i) ^
3100 *(u32 *)((u8 *)th2 + i);
3102 mss = skb_shinfo(p)->gso_size;
3104 flush |= (len - 1) >= mss;
3105 flush |= (ntohl(th2->seq) + skb_gro_len(p)) ^ ntohl(th->seq);
3107 if (flush || skb_gro_receive(head, skb)) {
3109 goto out_check_final;
3114 tcp_flag_word(th2) |= flags & (TCP_FLAG_FIN | TCP_FLAG_PSH);
3118 flush |= (__force int)(flags & (TCP_FLAG_URG | TCP_FLAG_PSH |
3119 TCP_FLAG_RST | TCP_FLAG_SYN |
3122 if (p && (!NAPI_GRO_CB(skb)->same_flow || flush))
3126 NAPI_GRO_CB(skb)->flush |= flush;
3130 EXPORT_SYMBOL(tcp_gro_receive);
3132 int tcp_gro_complete(struct sk_buff *skb)
3134 struct tcphdr *th = tcp_hdr(skb);
3136 skb->csum_start = skb_transport_header(skb) - skb->head;
3137 skb->csum_offset = offsetof(struct tcphdr, check);
3138 skb->ip_summed = CHECKSUM_PARTIAL;
3140 skb_shinfo(skb)->gso_segs = NAPI_GRO_CB(skb)->count;
3143 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
3147 EXPORT_SYMBOL(tcp_gro_complete);
3149 #ifdef CONFIG_TCP_MD5SIG
3150 static unsigned long tcp_md5sig_users;
3151 static struct tcp_md5sig_pool __percpu *tcp_md5sig_pool;
3152 static DEFINE_SPINLOCK(tcp_md5sig_pool_lock);
3154 static void __tcp_free_md5sig_pool(struct tcp_md5sig_pool __percpu *pool)
3158 for_each_possible_cpu(cpu) {
3159 struct tcp_md5sig_pool *p = per_cpu_ptr(pool, cpu);
3161 if (p->md5_desc.tfm)
3162 crypto_free_hash(p->md5_desc.tfm);
3167 void tcp_free_md5sig_pool(void)
3169 struct tcp_md5sig_pool __percpu *pool = NULL;
3171 spin_lock_bh(&tcp_md5sig_pool_lock);
3172 if (--tcp_md5sig_users == 0) {
3173 pool = tcp_md5sig_pool;
3174 tcp_md5sig_pool = NULL;
3176 spin_unlock_bh(&tcp_md5sig_pool_lock);
3178 __tcp_free_md5sig_pool(pool);
3180 EXPORT_SYMBOL(tcp_free_md5sig_pool);
3182 static struct tcp_md5sig_pool __percpu *
3183 __tcp_alloc_md5sig_pool(struct sock *sk)
3186 struct tcp_md5sig_pool __percpu *pool;
3188 pool = alloc_percpu(struct tcp_md5sig_pool);
3192 for_each_possible_cpu(cpu) {
3193 struct crypto_hash *hash;
3195 hash = crypto_alloc_hash("md5", 0, CRYPTO_ALG_ASYNC);
3196 if (IS_ERR_OR_NULL(hash))
3199 per_cpu_ptr(pool, cpu)->md5_desc.tfm = hash;
3203 __tcp_free_md5sig_pool(pool);
3207 struct tcp_md5sig_pool __percpu *tcp_alloc_md5sig_pool(struct sock *sk)
3209 struct tcp_md5sig_pool __percpu *pool;
3213 spin_lock_bh(&tcp_md5sig_pool_lock);
3214 pool = tcp_md5sig_pool;
3215 if (tcp_md5sig_users++ == 0) {
3217 spin_unlock_bh(&tcp_md5sig_pool_lock);
3220 spin_unlock_bh(&tcp_md5sig_pool_lock);
3224 spin_unlock_bh(&tcp_md5sig_pool_lock);
3227 /* we cannot hold spinlock here because this may sleep. */
3228 struct tcp_md5sig_pool __percpu *p;
3230 p = __tcp_alloc_md5sig_pool(sk);
3231 spin_lock_bh(&tcp_md5sig_pool_lock);
3234 spin_unlock_bh(&tcp_md5sig_pool_lock);
3237 pool = tcp_md5sig_pool;
3239 /* oops, it has already been assigned. */
3240 spin_unlock_bh(&tcp_md5sig_pool_lock);
3241 __tcp_free_md5sig_pool(p);
3243 tcp_md5sig_pool = pool = p;
3244 spin_unlock_bh(&tcp_md5sig_pool_lock);
3249 EXPORT_SYMBOL(tcp_alloc_md5sig_pool);
3253 * tcp_get_md5sig_pool - get md5sig_pool for this user
3255 * We use percpu structure, so if we succeed, we exit with preemption
3256 * and BH disabled, to make sure another thread or softirq handling
3257 * wont try to get same context.
3259 struct tcp_md5sig_pool *tcp_get_md5sig_pool(void)
3261 struct tcp_md5sig_pool __percpu *p;
3265 spin_lock(&tcp_md5sig_pool_lock);
3266 p = tcp_md5sig_pool;
3269 spin_unlock(&tcp_md5sig_pool_lock);
3272 return this_cpu_ptr(p);
3277 EXPORT_SYMBOL(tcp_get_md5sig_pool);
3279 void tcp_put_md5sig_pool(void)
3282 tcp_free_md5sig_pool();
3284 EXPORT_SYMBOL(tcp_put_md5sig_pool);
3286 int tcp_md5_hash_header(struct tcp_md5sig_pool *hp,
3287 const struct tcphdr *th)
3289 struct scatterlist sg;
3293 /* We are not allowed to change tcphdr, make a local copy */
3294 memcpy(&hdr, th, sizeof(hdr));
3297 /* options aren't included in the hash */
3298 sg_init_one(&sg, &hdr, sizeof(hdr));
3299 err = crypto_hash_update(&hp->md5_desc, &sg, sizeof(hdr));
3302 EXPORT_SYMBOL(tcp_md5_hash_header);
3304 int tcp_md5_hash_skb_data(struct tcp_md5sig_pool *hp,
3305 const struct sk_buff *skb, unsigned int header_len)
3307 struct scatterlist sg;
3308 const struct tcphdr *tp = tcp_hdr(skb);
3309 struct hash_desc *desc = &hp->md5_desc;
3311 const unsigned int head_data_len = skb_headlen(skb) > header_len ?
3312 skb_headlen(skb) - header_len : 0;
3313 const struct skb_shared_info *shi = skb_shinfo(skb);
3314 struct sk_buff *frag_iter;
3316 sg_init_table(&sg, 1);
3318 sg_set_buf(&sg, ((u8 *) tp) + header_len, head_data_len);
3319 if (crypto_hash_update(desc, &sg, head_data_len))
3322 for (i = 0; i < shi->nr_frags; ++i) {
3323 const struct skb_frag_struct *f = &shi->frags[i];
3324 unsigned int offset = f->page_offset;
3325 struct page *page = skb_frag_page(f) + (offset >> PAGE_SHIFT);
3327 sg_set_page(&sg, page, skb_frag_size(f),
3328 offset_in_page(offset));
3329 if (crypto_hash_update(desc, &sg, skb_frag_size(f)))
3333 skb_walk_frags(skb, frag_iter)
3334 if (tcp_md5_hash_skb_data(hp, frag_iter, 0))
3339 EXPORT_SYMBOL(tcp_md5_hash_skb_data);
3341 int tcp_md5_hash_key(struct tcp_md5sig_pool *hp, const struct tcp_md5sig_key *key)
3343 struct scatterlist sg;
3345 sg_init_one(&sg, key->key, key->keylen);
3346 return crypto_hash_update(&hp->md5_desc, &sg, key->keylen);
3348 EXPORT_SYMBOL(tcp_md5_hash_key);
3352 void tcp_done(struct sock *sk)
3354 struct request_sock *req = tcp_sk(sk)->fastopen_rsk;
3356 if (sk->sk_state == TCP_SYN_SENT || sk->sk_state == TCP_SYN_RECV)
3357 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_ATTEMPTFAILS);
3359 tcp_set_state(sk, TCP_CLOSE);
3360 tcp_clear_xmit_timers(sk);
3362 reqsk_fastopen_remove(sk, req, false);
3364 sk->sk_shutdown = SHUTDOWN_MASK;
3366 if (!sock_flag(sk, SOCK_DEAD))
3367 sk->sk_state_change(sk);
3369 inet_csk_destroy_sock(sk);
3371 EXPORT_SYMBOL_GPL(tcp_done);
3373 extern struct tcp_congestion_ops tcp_reno;
3375 static __initdata unsigned long thash_entries;
3376 static int __init set_thash_entries(char *str)
3383 ret = kstrtoul(str, 0, &thash_entries);
3389 __setup("thash_entries=", set_thash_entries);
3391 void tcp_init_mem(struct net *net)
3393 unsigned long limit = nr_free_buffer_pages() / 8;
3394 limit = max(limit, 128UL);
3395 net->ipv4.sysctl_tcp_mem[0] = limit / 4 * 3;
3396 net->ipv4.sysctl_tcp_mem[1] = limit;
3397 net->ipv4.sysctl_tcp_mem[2] = net->ipv4.sysctl_tcp_mem[0] * 2;
3400 void __init tcp_init(void)
3402 struct sk_buff *skb = NULL;
3403 unsigned long limit;
3404 int max_rshare, max_wshare, cnt;
3407 BUILD_BUG_ON(sizeof(struct tcp_skb_cb) > sizeof(skb->cb));
3409 percpu_counter_init(&tcp_sockets_allocated, 0);
3410 percpu_counter_init(&tcp_orphan_count, 0);
3411 tcp_hashinfo.bind_bucket_cachep =
3412 kmem_cache_create("tcp_bind_bucket",
3413 sizeof(struct inet_bind_bucket), 0,
3414 SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL);
3416 /* Size and allocate the main established and bind bucket
3419 * The methodology is similar to that of the buffer cache.
3421 tcp_hashinfo.ehash =
3422 alloc_large_system_hash("TCP established",
3423 sizeof(struct inet_ehash_bucket),
3425 17, /* one slot per 128 KB of memory */
3428 &tcp_hashinfo.ehash_mask,
3430 thash_entries ? 0 : 512 * 1024);
3431 for (i = 0; i <= tcp_hashinfo.ehash_mask; i++) {
3432 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].chain, i);
3433 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].twchain, i);
3435 if (inet_ehash_locks_alloc(&tcp_hashinfo))
3436 panic("TCP: failed to alloc ehash_locks");
3437 tcp_hashinfo.bhash =
3438 alloc_large_system_hash("TCP bind",
3439 sizeof(struct inet_bind_hashbucket),
3440 tcp_hashinfo.ehash_mask + 1,
3441 17, /* one slot per 128 KB of memory */
3443 &tcp_hashinfo.bhash_size,
3447 tcp_hashinfo.bhash_size = 1U << tcp_hashinfo.bhash_size;
3448 for (i = 0; i < tcp_hashinfo.bhash_size; i++) {
3449 spin_lock_init(&tcp_hashinfo.bhash[i].lock);
3450 INIT_HLIST_HEAD(&tcp_hashinfo.bhash[i].chain);
3454 cnt = tcp_hashinfo.ehash_mask + 1;
3456 tcp_death_row.sysctl_max_tw_buckets = cnt / 2;
3457 sysctl_tcp_max_orphans = cnt / 2;
3458 sysctl_max_syn_backlog = max(128, cnt / 256);
3460 tcp_init_mem(&init_net);
3461 /* Set per-socket limits to no more than 1/128 the pressure threshold */
3462 limit = nr_free_buffer_pages() << (PAGE_SHIFT - 7);
3463 max_wshare = min(4UL*1024*1024, limit);
3464 max_rshare = min(6UL*1024*1024, limit);
3466 sysctl_tcp_wmem[0] = SK_MEM_QUANTUM;
3467 sysctl_tcp_wmem[1] = 16*1024;
3468 sysctl_tcp_wmem[2] = max(64*1024, max_wshare);
3470 sysctl_tcp_rmem[0] = SK_MEM_QUANTUM;
3471 sysctl_tcp_rmem[1] = 87380;
3472 sysctl_tcp_rmem[2] = max(87380, max_rshare);
3474 pr_info("Hash tables configured (established %u bind %u)\n",
3475 tcp_hashinfo.ehash_mask + 1, tcp_hashinfo.bhash_size);
3479 tcp_register_congestion_control(&tcp_reno);
3484 static int tcp_is_local(struct net *net, __be32 addr) {
3486 struct flowi4 fl4 = { .daddr = addr };
3487 rt = ip_route_output_key(net, &fl4);
3488 if (IS_ERR_OR_NULL(rt))
3490 return rt->dst.dev && (rt->dst.dev->flags & IFF_LOOPBACK);
3493 #if defined(CONFIG_IPV6)
3494 static int tcp_is_local6(struct net *net, struct in6_addr *addr) {
3495 struct rt6_info *rt6 = rt6_lookup(net, addr, addr, 0, 0);
3496 return rt6 && rt6->dst.dev && (rt6->dst.dev->flags & IFF_LOOPBACK);
3501 * tcp_nuke_addr - destroy all sockets on the given local address
3502 * if local address is the unspecified address (0.0.0.0 or ::), destroy all
3503 * sockets with local addresses that are not configured.
3505 int tcp_nuke_addr(struct net *net, struct sockaddr *addr)
3507 int family = addr->sa_family;
3508 unsigned int bucket;
3511 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
3512 struct in6_addr *in6;
3514 if (family == AF_INET) {
3515 in = &((struct sockaddr_in *)addr)->sin_addr;
3516 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
3517 } else if (family == AF_INET6) {
3518 in6 = &((struct sockaddr_in6 *)addr)->sin6_addr;
3521 return -EAFNOSUPPORT;
3524 for (bucket = 0; bucket < tcp_hashinfo.ehash_mask; bucket++) {
3525 struct hlist_nulls_node *node;
3527 spinlock_t *lock = inet_ehash_lockp(&tcp_hashinfo, bucket);
3531 sk_nulls_for_each(sk, node, &tcp_hashinfo.ehash[bucket].chain) {
3532 struct inet_sock *inet = inet_sk(sk);
3534 if (sysctl_ip_dynaddr && sk->sk_state == TCP_SYN_SENT)
3536 if (sock_flag(sk, SOCK_DEAD))
3539 if (family == AF_INET) {
3540 __be32 s4 = inet->inet_rcv_saddr;
3541 if (s4 == LOOPBACK4_IPV6)
3544 if (in->s_addr != s4 &&
3545 !(in->s_addr == INADDR_ANY &&
3546 !tcp_is_local(net, s4)))
3550 #if defined(CONFIG_IPV6)
3551 if (family == AF_INET6) {
3552 struct in6_addr *s6;
3556 s6 = &inet->pinet6->rcv_saddr;
3557 if (ipv6_addr_type(s6) == IPV6_ADDR_MAPPED)
3560 if (!ipv6_addr_equal(in6, s6) &&
3561 !(ipv6_addr_equal(in6, &in6addr_any) &&
3562 !tcp_is_local6(net, s6)))
3568 spin_unlock_bh(lock);
3572 sk->sk_err = ETIMEDOUT;
3573 sk->sk_error_report(sk);
3582 spin_unlock_bh(lock);