2 * INET An implementation of the TCP/IP protocol suite for the LINUX
3 * operating system. INET is implemented using the BSD Socket
4 * interface as the means of communication with the user level.
6 * Implementation of the Transmission Control Protocol(TCP).
9 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
10 * Mark Evans, <evansmp@uhura.aston.ac.uk>
11 * Corey Minyard <wf-rch!minyard@relay.EU.net>
12 * Florian La Roche, <flla@stud.uni-sb.de>
13 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
14 * Linus Torvalds, <torvalds@cs.helsinki.fi>
15 * Alan Cox, <gw4pts@gw4pts.ampr.org>
16 * Matthew Dillon, <dillon@apollo.west.oic.com>
17 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
18 * Jorge Cwik, <jorge@laser.satlink.net>
21 * Alan Cox : Numerous verify_area() calls
22 * Alan Cox : Set the ACK bit on a reset
23 * Alan Cox : Stopped it crashing if it closed while
24 * sk->inuse=1 and was trying to connect
26 * Alan Cox : All icmp error handling was broken
27 * pointers passed where wrong and the
28 * socket was looked up backwards. Nobody
29 * tested any icmp error code obviously.
30 * Alan Cox : tcp_err() now handled properly. It
31 * wakes people on errors. poll
32 * behaves and the icmp error race
33 * has gone by moving it into sock.c
34 * Alan Cox : tcp_send_reset() fixed to work for
35 * everything not just packets for
37 * Alan Cox : tcp option processing.
38 * Alan Cox : Reset tweaked (still not 100%) [Had
40 * Herp Rosmanith : More reset fixes
41 * Alan Cox : No longer acks invalid rst frames.
42 * Acking any kind of RST is right out.
43 * Alan Cox : Sets an ignore me flag on an rst
44 * receive otherwise odd bits of prattle
46 * Alan Cox : Fixed another acking RST frame bug.
47 * Should stop LAN workplace lockups.
48 * Alan Cox : Some tidyups using the new skb list
50 * Alan Cox : sk->keepopen now seems to work
51 * Alan Cox : Pulls options out correctly on accepts
52 * Alan Cox : Fixed assorted sk->rqueue->next errors
53 * Alan Cox : PSH doesn't end a TCP read. Switched a
55 * Alan Cox : Tidied tcp_data to avoid a potential
57 * Alan Cox : Added some better commenting, as the
58 * tcp is hard to follow
59 * Alan Cox : Removed incorrect check for 20 * psh
60 * Michael O'Reilly : ack < copied bug fix.
61 * Johannes Stille : Misc tcp fixes (not all in yet).
62 * Alan Cox : FIN with no memory -> CRASH
63 * Alan Cox : Added socket option proto entries.
64 * Also added awareness of them to accept.
65 * Alan Cox : Added TCP options (SOL_TCP)
66 * Alan Cox : Switched wakeup calls to callbacks,
67 * so the kernel can layer network
69 * Alan Cox : Use ip_tos/ip_ttl settings.
70 * Alan Cox : Handle FIN (more) properly (we hope).
71 * Alan Cox : RST frames sent on unsynchronised
73 * Alan Cox : Put in missing check for SYN bit.
74 * Alan Cox : Added tcp_select_window() aka NET2E
75 * window non shrink trick.
76 * Alan Cox : Added a couple of small NET2E timer
78 * Charles Hedrick : TCP fixes
79 * Toomas Tamm : TCP window fixes
80 * Alan Cox : Small URG fix to rlogin ^C ack fight
81 * Charles Hedrick : Rewrote most of it to actually work
82 * Linus : Rewrote tcp_read() and URG handling
84 * Gerhard Koerting: Fixed some missing timer handling
85 * Matthew Dillon : Reworked TCP machine states as per RFC
86 * Gerhard Koerting: PC/TCP workarounds
87 * Adam Caldwell : Assorted timer/timing errors
88 * Matthew Dillon : Fixed another RST bug
89 * Alan Cox : Move to kernel side addressing changes.
90 * Alan Cox : Beginning work on TCP fastpathing
92 * Arnt Gulbrandsen: Turbocharged tcp_check() routine.
93 * Alan Cox : TCP fast path debugging
94 * Alan Cox : Window clamping
95 * Michael Riepe : Bug in tcp_check()
96 * Matt Dillon : More TCP improvements and RST bug fixes
97 * Matt Dillon : Yet more small nasties remove from the
98 * TCP code (Be very nice to this man if
99 * tcp finally works 100%) 8)
100 * Alan Cox : BSD accept semantics.
101 * Alan Cox : Reset on closedown bug.
102 * Peter De Schrijver : ENOTCONN check missing in tcp_sendto().
103 * Michael Pall : Handle poll() after URG properly in
105 * Michael Pall : Undo the last fix in tcp_read_urg()
106 * (multi URG PUSH broke rlogin).
107 * Michael Pall : Fix the multi URG PUSH problem in
108 * tcp_readable(), poll() after URG
110 * Michael Pall : recv(...,MSG_OOB) never blocks in the
112 * Alan Cox : Changed the semantics of sk->socket to
113 * fix a race and a signal problem with
114 * accept() and async I/O.
115 * Alan Cox : Relaxed the rules on tcp_sendto().
116 * Yury Shevchuk : Really fixed accept() blocking problem.
117 * Craig I. Hagan : Allow for BSD compatible TIME_WAIT for
118 * clients/servers which listen in on
120 * Alan Cox : Cleaned the above up and shrank it to
121 * a sensible code size.
122 * Alan Cox : Self connect lockup fix.
123 * Alan Cox : No connect to multicast.
124 * Ross Biro : Close unaccepted children on master
126 * Alan Cox : Reset tracing code.
127 * Alan Cox : Spurious resets on shutdown.
128 * Alan Cox : Giant 15 minute/60 second timer error
129 * Alan Cox : Small whoops in polling before an
131 * Alan Cox : Kept the state trace facility since
132 * it's handy for debugging.
133 * Alan Cox : More reset handler fixes.
134 * Alan Cox : Started rewriting the code based on
135 * the RFC's for other useful protocol
136 * references see: Comer, KA9Q NOS, and
137 * for a reference on the difference
138 * between specifications and how BSD
139 * works see the 4.4lite source.
140 * A.N.Kuznetsov : Don't time wait on completion of tidy
142 * Linus Torvalds : Fin/Shutdown & copied_seq changes.
143 * Linus Torvalds : Fixed BSD port reuse to work first syn
144 * Alan Cox : Reimplemented timers as per the RFC
145 * and using multiple timers for sanity.
146 * Alan Cox : Small bug fixes, and a lot of new
148 * Alan Cox : Fixed dual reader crash by locking
149 * the buffers (much like datagram.c)
150 * Alan Cox : Fixed stuck sockets in probe. A probe
151 * now gets fed up of retrying without
152 * (even a no space) answer.
153 * Alan Cox : Extracted closing code better
154 * Alan Cox : Fixed the closing state machine to
156 * Alan Cox : More 'per spec' fixes.
157 * Jorge Cwik : Even faster checksumming.
158 * Alan Cox : tcp_data() doesn't ack illegal PSH
159 * only frames. At least one pc tcp stack
161 * Alan Cox : Cache last socket.
162 * Alan Cox : Per route irtt.
163 * Matt Day : poll()->select() match BSD precisely on error
164 * Alan Cox : New buffers
165 * Marc Tamsky : Various sk->prot->retransmits and
166 * sk->retransmits misupdating fixed.
167 * Fixed tcp_write_timeout: stuck close,
168 * and TCP syn retries gets used now.
169 * Mark Yarvis : In tcp_read_wakeup(), don't send an
170 * ack if state is TCP_CLOSED.
171 * Alan Cox : Look up device on a retransmit - routes may
172 * change. Doesn't yet cope with MSS shrink right
174 * Marc Tamsky : Closing in closing fixes.
175 * Mike Shaver : RFC1122 verifications.
176 * Alan Cox : rcv_saddr errors.
177 * Alan Cox : Block double connect().
178 * Alan Cox : Small hooks for enSKIP.
179 * Alexey Kuznetsov: Path MTU discovery.
180 * Alan Cox : Support soft errors.
181 * Alan Cox : Fix MTU discovery pathological case
182 * when the remote claims no mtu!
183 * Marc Tamsky : TCP_CLOSE fix.
184 * Colin (G3TNE) : Send a reset on syn ack replies in
185 * window but wrong (fixes NT lpd problems)
186 * Pedro Roque : Better TCP window handling, delayed ack.
187 * Joerg Reuter : No modification of locked buffers in
188 * tcp_do_retransmit()
189 * Eric Schenk : Changed receiver side silly window
190 * avoidance algorithm to BSD style
191 * algorithm. This doubles throughput
192 * against machines running Solaris,
193 * and seems to result in general
195 * Stefan Magdalinski : adjusted tcp_readable() to fix FIONREAD
196 * Willy Konynenberg : Transparent proxying support.
197 * Mike McLagan : Routing by source
198 * Keith Owens : Do proper merging with partial SKB's in
199 * tcp_do_sendmsg to avoid burstiness.
200 * Eric Schenk : Fix fast close down bug with
201 * shutdown() followed by close().
202 * Andi Kleen : Make poll agree with SIGIO
203 * Salvatore Sanfilippo : Support SO_LINGER with linger == 1 and
204 * lingertime == 0 (RFC 793 ABORT Call)
205 * Hirokazu Takahashi : Use copy_from_user() instead of
206 * csum_and_copy_from_user() if possible.
208 * This program is free software; you can redistribute it and/or
209 * modify it under the terms of the GNU General Public License
210 * as published by the Free Software Foundation; either version
211 * 2 of the License, or(at your option) any later version.
213 * Description of States:
215 * TCP_SYN_SENT sent a connection request, waiting for ack
217 * TCP_SYN_RECV received a connection request, sent ack,
218 * waiting for final ack in three-way handshake.
220 * TCP_ESTABLISHED connection established
222 * TCP_FIN_WAIT1 our side has shutdown, waiting to complete
223 * transmission of remaining buffered data
225 * TCP_FIN_WAIT2 all buffered data sent, waiting for remote
228 * TCP_CLOSING both sides have shutdown but we still have
229 * data we have to finish sending
231 * TCP_TIME_WAIT timeout to catch resent junk before entering
232 * closed, can only be entered from FIN_WAIT2
233 * or CLOSING. Required because the other end
234 * may not have gotten our last ACK causing it
235 * to retransmit the data packet (which we ignore)
237 * TCP_CLOSE_WAIT remote side has shutdown and is waiting for
238 * us to finish writing our data and to shutdown
239 * (we have to close() to move on to LAST_ACK)
241 * TCP_LAST_ACK out side has shutdown after remote has
242 * shutdown. There may still be data in our
243 * buffer that we have to finish sending
245 * TCP_CLOSE socket is finished
248 #define pr_fmt(fmt) "TCP: " fmt
250 #include <linux/kernel.h>
251 #include <linux/module.h>
252 #include <linux/types.h>
253 #include <linux/fcntl.h>
254 #include <linux/poll.h>
255 #include <linux/init.h>
256 #include <linux/fs.h>
257 #include <linux/skbuff.h>
258 #include <linux/scatterlist.h>
259 #include <linux/splice.h>
260 #include <linux/net.h>
261 #include <linux/socket.h>
262 #include <linux/random.h>
263 #include <linux/bootmem.h>
264 #include <linux/highmem.h>
265 #include <linux/swap.h>
266 #include <linux/cache.h>
267 #include <linux/err.h>
268 #include <linux/crypto.h>
269 #include <linux/time.h>
270 #include <linux/slab.h>
272 #include <net/icmp.h>
274 #include <net/xfrm.h>
276 #include <net/netdma.h>
277 #include <net/sock.h>
279 #include <asm/uaccess.h>
280 #include <asm/ioctls.h>
282 int sysctl_tcp_fin_timeout __read_mostly = TCP_FIN_TIMEOUT;
284 struct percpu_counter tcp_orphan_count;
285 EXPORT_SYMBOL_GPL(tcp_orphan_count);
287 int sysctl_tcp_wmem[3] __read_mostly;
288 int sysctl_tcp_rmem[3] __read_mostly;
290 EXPORT_SYMBOL(sysctl_tcp_rmem);
291 EXPORT_SYMBOL(sysctl_tcp_wmem);
293 atomic_long_t tcp_memory_allocated; /* Current allocated memory. */
294 EXPORT_SYMBOL(tcp_memory_allocated);
297 * Current number of TCP sockets.
299 struct percpu_counter tcp_sockets_allocated;
300 EXPORT_SYMBOL(tcp_sockets_allocated);
305 struct tcp_splice_state {
306 struct pipe_inode_info *pipe;
312 * Pressure flag: try to collapse.
313 * Technical note: it is used by multiple contexts non atomically.
314 * All the __sk_mem_schedule() is of this nature: accounting
315 * is strict, actions are advisory and have some latency.
317 int tcp_memory_pressure __read_mostly;
318 EXPORT_SYMBOL(tcp_memory_pressure);
320 void tcp_enter_memory_pressure(struct sock *sk)
322 if (!tcp_memory_pressure) {
323 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMEMORYPRESSURES);
324 tcp_memory_pressure = 1;
327 EXPORT_SYMBOL(tcp_enter_memory_pressure);
329 /* Convert seconds to retransmits based on initial and max timeout */
330 static u8 secs_to_retrans(int seconds, int timeout, int rto_max)
335 int period = timeout;
338 while (seconds > period && res < 255) {
341 if (timeout > rto_max)
349 /* Convert retransmits to seconds based on initial and max timeout */
350 static int retrans_to_secs(u8 retrans, int timeout, int rto_max)
358 if (timeout > rto_max)
366 /* Address-family independent initialization for a tcp_sock.
368 * NOTE: A lot of things set to zero explicitly by call to
369 * sk_alloc() so need not be done here.
371 void tcp_init_sock(struct sock *sk)
373 struct inet_connection_sock *icsk = inet_csk(sk);
374 struct tcp_sock *tp = tcp_sk(sk);
376 skb_queue_head_init(&tp->out_of_order_queue);
377 tcp_init_xmit_timers(sk);
378 tcp_prequeue_init(tp);
380 icsk->icsk_rto = TCP_TIMEOUT_INIT;
381 tp->mdev = TCP_TIMEOUT_INIT;
383 /* So many TCP implementations out there (incorrectly) count the
384 * initial SYN frame in their delayed-ACK and congestion control
385 * algorithms that we must have the following bandaid to talk
386 * efficiently to them. -DaveM
388 tp->snd_cwnd = TCP_INIT_CWND;
390 /* See draft-stevens-tcpca-spec-01 for discussion of the
391 * initialization of these values.
393 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
394 tp->snd_cwnd_clamp = ~0;
395 tp->mss_cache = TCP_MSS_DEFAULT;
397 tp->reordering = sysctl_tcp_reordering;
398 tcp_enable_early_retrans(tp);
399 icsk->icsk_ca_ops = &tcp_init_congestion_ops;
401 sk->sk_state = TCP_CLOSE;
403 sk->sk_write_space = sk_stream_write_space;
404 sock_set_flag(sk, SOCK_USE_WRITE_QUEUE);
406 icsk->icsk_sync_mss = tcp_sync_mss;
408 /* TCP Cookie Transactions */
409 if (sysctl_tcp_cookie_size > 0) {
410 /* Default, cookies without s_data_payload. */
412 kzalloc(sizeof(*tp->cookie_values),
414 if (tp->cookie_values != NULL)
415 kref_init(&tp->cookie_values->kref);
417 /* Presumed zeroed, in order of appearance:
418 * cookie_in_always, cookie_out_never,
419 * s_data_constant, s_data_in, s_data_out
421 sk->sk_sndbuf = sysctl_tcp_wmem[1];
422 sk->sk_rcvbuf = sysctl_tcp_rmem[1];
425 sock_update_memcg(sk);
426 sk_sockets_allocated_inc(sk);
429 EXPORT_SYMBOL(tcp_init_sock);
432 * Wait for a TCP event.
434 * Note that we don't need to lock the socket, as the upper poll layers
435 * take care of normal races (between the test and the event) and we don't
436 * go look at any of the socket buffers directly.
438 unsigned int tcp_poll(struct file *file, struct socket *sock, poll_table *wait)
441 struct sock *sk = sock->sk;
442 const struct tcp_sock *tp = tcp_sk(sk);
444 sock_poll_wait(file, sk_sleep(sk), wait);
445 if (sk->sk_state == TCP_LISTEN)
446 return inet_csk_listen_poll(sk);
448 /* Socket is not locked. We are protected from async events
449 * by poll logic and correct handling of state changes
450 * made by other threads is impossible in any case.
456 * POLLHUP is certainly not done right. But poll() doesn't
457 * have a notion of HUP in just one direction, and for a
458 * socket the read side is more interesting.
460 * Some poll() documentation says that POLLHUP is incompatible
461 * with the POLLOUT/POLLWR flags, so somebody should check this
462 * all. But careful, it tends to be safer to return too many
463 * bits than too few, and you can easily break real applications
464 * if you don't tell them that something has hung up!
468 * Check number 1. POLLHUP is _UNMASKABLE_ event (see UNIX98 and
469 * our fs/select.c). It means that after we received EOF,
470 * poll always returns immediately, making impossible poll() on write()
471 * in state CLOSE_WAIT. One solution is evident --- to set POLLHUP
472 * if and only if shutdown has been made in both directions.
473 * Actually, it is interesting to look how Solaris and DUX
474 * solve this dilemma. I would prefer, if POLLHUP were maskable,
475 * then we could set it on SND_SHUTDOWN. BTW examples given
476 * in Stevens' books assume exactly this behaviour, it explains
477 * why POLLHUP is incompatible with POLLOUT. --ANK
479 * NOTE. Check for TCP_CLOSE is added. The goal is to prevent
480 * blocking on fresh not-connected or disconnected socket. --ANK
482 if (sk->sk_shutdown == SHUTDOWN_MASK || sk->sk_state == TCP_CLOSE)
484 if (sk->sk_shutdown & RCV_SHUTDOWN)
485 mask |= POLLIN | POLLRDNORM | POLLRDHUP;
488 if ((1 << sk->sk_state) & ~(TCPF_SYN_SENT | TCPF_SYN_RECV)) {
489 int target = sock_rcvlowat(sk, 0, INT_MAX);
491 if (tp->urg_seq == tp->copied_seq &&
492 !sock_flag(sk, SOCK_URGINLINE) &&
496 /* Potential race condition. If read of tp below will
497 * escape above sk->sk_state, we can be illegally awaken
498 * in SYN_* states. */
499 if (tp->rcv_nxt - tp->copied_seq >= target)
500 mask |= POLLIN | POLLRDNORM;
502 if (!(sk->sk_shutdown & SEND_SHUTDOWN)) {
503 if (sk_stream_wspace(sk) >= sk_stream_min_wspace(sk)) {
504 mask |= POLLOUT | POLLWRNORM;
505 } else { /* send SIGIO later */
506 set_bit(SOCK_ASYNC_NOSPACE,
507 &sk->sk_socket->flags);
508 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
510 /* Race breaker. If space is freed after
511 * wspace test but before the flags are set,
512 * IO signal will be lost.
514 if (sk_stream_wspace(sk) >= sk_stream_min_wspace(sk))
515 mask |= POLLOUT | POLLWRNORM;
518 mask |= POLLOUT | POLLWRNORM;
520 if (tp->urg_data & TCP_URG_VALID)
523 /* This barrier is coupled with smp_wmb() in tcp_reset() */
530 EXPORT_SYMBOL(tcp_poll);
532 int tcp_ioctl(struct sock *sk, int cmd, unsigned long arg)
534 struct tcp_sock *tp = tcp_sk(sk);
539 if (sk->sk_state == TCP_LISTEN)
543 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
545 else if (sock_flag(sk, SOCK_URGINLINE) ||
547 before(tp->urg_seq, tp->copied_seq) ||
548 !before(tp->urg_seq, tp->rcv_nxt)) {
551 answ = tp->rcv_nxt - tp->copied_seq;
553 /* Subtract 1, if FIN is in queue. */
554 skb = skb_peek_tail(&sk->sk_receive_queue);
556 answ -= tcp_hdr(skb)->fin;
558 answ = tp->urg_seq - tp->copied_seq;
562 answ = tp->urg_data && tp->urg_seq == tp->copied_seq;
565 if (sk->sk_state == TCP_LISTEN)
568 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
571 answ = tp->write_seq - tp->snd_una;
574 if (sk->sk_state == TCP_LISTEN)
577 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
580 answ = tp->write_seq - tp->snd_nxt;
586 return put_user(answ, (int __user *)arg);
588 EXPORT_SYMBOL(tcp_ioctl);
590 static inline void tcp_mark_push(struct tcp_sock *tp, struct sk_buff *skb)
592 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
593 tp->pushed_seq = tp->write_seq;
596 static inline int forced_push(const struct tcp_sock *tp)
598 return after(tp->write_seq, tp->pushed_seq + (tp->max_window >> 1));
601 static inline void skb_entail(struct sock *sk, struct sk_buff *skb)
603 struct tcp_sock *tp = tcp_sk(sk);
604 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
607 tcb->seq = tcb->end_seq = tp->write_seq;
608 tcb->tcp_flags = TCPHDR_ACK;
610 skb_header_release(skb);
611 tcp_add_write_queue_tail(sk, skb);
612 sk->sk_wmem_queued += skb->truesize;
613 sk_mem_charge(sk, skb->truesize);
614 if (tp->nonagle & TCP_NAGLE_PUSH)
615 tp->nonagle &= ~TCP_NAGLE_PUSH;
618 static inline void tcp_mark_urg(struct tcp_sock *tp, int flags)
621 tp->snd_up = tp->write_seq;
624 static inline void tcp_push(struct sock *sk, int flags, int mss_now,
627 if (tcp_send_head(sk)) {
628 struct tcp_sock *tp = tcp_sk(sk);
630 if (!(flags & MSG_MORE) || forced_push(tp))
631 tcp_mark_push(tp, tcp_write_queue_tail(sk));
633 tcp_mark_urg(tp, flags);
634 __tcp_push_pending_frames(sk, mss_now,
635 (flags & MSG_MORE) ? TCP_NAGLE_CORK : nonagle);
639 static int tcp_splice_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb,
640 unsigned int offset, size_t len)
642 struct tcp_splice_state *tss = rd_desc->arg.data;
645 ret = skb_splice_bits(skb, offset, tss->pipe, min(rd_desc->count, len),
648 rd_desc->count -= ret;
652 static int __tcp_splice_read(struct sock *sk, struct tcp_splice_state *tss)
654 /* Store TCP splice context information in read_descriptor_t. */
655 read_descriptor_t rd_desc = {
660 return tcp_read_sock(sk, &rd_desc, tcp_splice_data_recv);
664 * tcp_splice_read - splice data from TCP socket to a pipe
665 * @sock: socket to splice from
666 * @ppos: position (not valid)
667 * @pipe: pipe to splice to
668 * @len: number of bytes to splice
669 * @flags: splice modifier flags
672 * Will read pages from given socket and fill them into a pipe.
675 ssize_t tcp_splice_read(struct socket *sock, loff_t *ppos,
676 struct pipe_inode_info *pipe, size_t len,
679 struct sock *sk = sock->sk;
680 struct tcp_splice_state tss = {
689 sock_rps_record_flow(sk);
691 * We can't seek on a socket input
700 timeo = sock_rcvtimeo(sk, sock->file->f_flags & O_NONBLOCK);
702 ret = __tcp_splice_read(sk, &tss);
708 if (sock_flag(sk, SOCK_DONE))
711 ret = sock_error(sk);
714 if (sk->sk_shutdown & RCV_SHUTDOWN)
716 if (sk->sk_state == TCP_CLOSE) {
718 * This occurs when user tries to read
719 * from never connected socket.
721 if (!sock_flag(sk, SOCK_DONE))
729 sk_wait_data(sk, &timeo);
730 if (signal_pending(current)) {
731 ret = sock_intr_errno(timeo);
744 if (sk->sk_err || sk->sk_state == TCP_CLOSE ||
745 (sk->sk_shutdown & RCV_SHUTDOWN) ||
746 signal_pending(current))
757 EXPORT_SYMBOL(tcp_splice_read);
759 struct sk_buff *sk_stream_alloc_skb(struct sock *sk, int size, gfp_t gfp)
763 /* The TCP header must be at least 32-bit aligned. */
764 size = ALIGN(size, 4);
766 skb = alloc_skb_fclone(size + sk->sk_prot->max_header, gfp);
768 if (sk_wmem_schedule(sk, skb->truesize)) {
769 skb_reserve(skb, sk->sk_prot->max_header);
771 * Make sure that we have exactly size bytes
772 * available to the caller, no more, no less.
774 skb->avail_size = size;
779 sk->sk_prot->enter_memory_pressure(sk);
780 sk_stream_moderate_sndbuf(sk);
785 static unsigned int tcp_xmit_size_goal(struct sock *sk, u32 mss_now,
788 struct tcp_sock *tp = tcp_sk(sk);
789 u32 xmit_size_goal, old_size_goal;
791 xmit_size_goal = mss_now;
793 if (large_allowed && sk_can_gso(sk)) {
794 xmit_size_goal = ((sk->sk_gso_max_size - 1) -
795 inet_csk(sk)->icsk_af_ops->net_header_len -
796 inet_csk(sk)->icsk_ext_hdr_len -
799 xmit_size_goal = tcp_bound_to_half_wnd(tp, xmit_size_goal);
801 /* We try hard to avoid divides here */
802 old_size_goal = tp->xmit_size_goal_segs * mss_now;
804 if (likely(old_size_goal <= xmit_size_goal &&
805 old_size_goal + mss_now > xmit_size_goal)) {
806 xmit_size_goal = old_size_goal;
808 tp->xmit_size_goal_segs = xmit_size_goal / mss_now;
809 xmit_size_goal = tp->xmit_size_goal_segs * mss_now;
813 return max(xmit_size_goal, mss_now);
816 static int tcp_send_mss(struct sock *sk, int *size_goal, int flags)
820 mss_now = tcp_current_mss(sk);
821 *size_goal = tcp_xmit_size_goal(sk, mss_now, !(flags & MSG_OOB));
826 static ssize_t do_tcp_sendpages(struct sock *sk, struct page **pages, int poffset,
827 size_t psize, int flags)
829 struct tcp_sock *tp = tcp_sk(sk);
830 int mss_now, size_goal;
833 long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
835 /* Wait for a connection to finish. */
836 if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
837 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
840 clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
842 mss_now = tcp_send_mss(sk, &size_goal, flags);
846 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
850 struct sk_buff *skb = tcp_write_queue_tail(sk);
851 struct page *page = pages[poffset / PAGE_SIZE];
853 int offset = poffset % PAGE_SIZE;
854 int size = min_t(size_t, psize, PAGE_SIZE - offset);
857 if (!tcp_send_head(sk) || (copy = size_goal - skb->len) <= 0) {
859 if (!sk_stream_memory_free(sk))
860 goto wait_for_sndbuf;
862 skb = sk_stream_alloc_skb(sk, 0, sk->sk_allocation);
864 goto wait_for_memory;
873 i = skb_shinfo(skb)->nr_frags;
874 can_coalesce = skb_can_coalesce(skb, i, page, offset);
875 if (!can_coalesce && i >= MAX_SKB_FRAGS) {
876 tcp_mark_push(tp, skb);
879 if (!sk_wmem_schedule(sk, copy))
880 goto wait_for_memory;
883 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
886 skb_fill_page_desc(skb, i, page, offset, copy);
890 skb->data_len += copy;
891 skb->truesize += copy;
892 sk->sk_wmem_queued += copy;
893 sk_mem_charge(sk, copy);
894 skb->ip_summed = CHECKSUM_PARTIAL;
895 tp->write_seq += copy;
896 TCP_SKB_CB(skb)->end_seq += copy;
897 skb_shinfo(skb)->gso_segs = 0;
900 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
904 if (!(psize -= copy))
907 if (skb->len < size_goal || (flags & MSG_OOB))
910 if (forced_push(tp)) {
911 tcp_mark_push(tp, skb);
912 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
913 } else if (skb == tcp_send_head(sk))
914 tcp_push_one(sk, mss_now);
918 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
921 tcp_push(sk, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH);
923 if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
926 mss_now = tcp_send_mss(sk, &size_goal, flags);
930 if (copied && !(flags & MSG_SENDPAGE_NOTLAST))
931 tcp_push(sk, flags, mss_now, tp->nonagle);
938 return sk_stream_error(sk, flags, err);
941 int tcp_sendpage(struct sock *sk, struct page *page, int offset,
942 size_t size, int flags)
946 if (!(sk->sk_route_caps & NETIF_F_SG) ||
947 !(sk->sk_route_caps & NETIF_F_ALL_CSUM))
948 return sock_no_sendpage(sk->sk_socket, page, offset, size,
952 res = do_tcp_sendpages(sk, &page, offset, size, flags);
956 EXPORT_SYMBOL(tcp_sendpage);
958 static inline int select_size(const struct sock *sk, bool sg)
960 const struct tcp_sock *tp = tcp_sk(sk);
961 int tmp = tp->mss_cache;
964 if (sk_can_gso(sk)) {
965 /* Small frames wont use a full page:
966 * Payload will immediately follow tcp header.
968 tmp = SKB_WITH_OVERHEAD(2048 - MAX_TCP_HEADER);
970 int pgbreak = SKB_MAX_HEAD(MAX_TCP_HEADER);
972 if (tmp >= pgbreak &&
973 tmp <= pgbreak + (MAX_SKB_FRAGS - 1) * PAGE_SIZE)
981 static int tcp_send_rcvq(struct sock *sk, struct msghdr *msg, size_t size)
987 skb = alloc_skb(size + sizeof(*th), sk->sk_allocation);
991 th = (struct tcphdr *)skb_put(skb, sizeof(*th));
992 skb_reset_transport_header(skb);
993 memset(th, 0, sizeof(*th));
995 if (memcpy_fromiovec(skb_put(skb, size), msg->msg_iov, size))
998 TCP_SKB_CB(skb)->seq = tcp_sk(sk)->rcv_nxt;
999 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + size;
1000 TCP_SKB_CB(skb)->ack_seq = tcp_sk(sk)->snd_una - 1;
1002 if (tcp_queue_rcv(sk, skb, sizeof(*th), &fragstolen)) {
1003 WARN_ON_ONCE(fragstolen); /* should not happen */
1014 int tcp_sendmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg,
1018 struct tcp_sock *tp = tcp_sk(sk);
1019 struct sk_buff *skb;
1020 int iovlen, flags, err, copied;
1021 int mss_now = 0, size_goal;
1027 flags = msg->msg_flags;
1028 timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
1030 /* Wait for a connection to finish. */
1031 if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
1032 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
1035 if (unlikely(tp->repair)) {
1036 if (tp->repair_queue == TCP_RECV_QUEUE) {
1037 copied = tcp_send_rcvq(sk, msg, size);
1042 if (tp->repair_queue == TCP_NO_QUEUE)
1045 /* 'common' sending to sendq */
1048 /* This should be in poll */
1049 clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
1051 mss_now = tcp_send_mss(sk, &size_goal, flags);
1053 /* Ok commence sending. */
1054 iovlen = msg->msg_iovlen;
1059 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
1062 sg = !!(sk->sk_route_caps & NETIF_F_SG);
1064 while (--iovlen >= 0) {
1065 size_t seglen = iov->iov_len;
1066 unsigned char __user *from = iov->iov_base;
1070 while (seglen > 0) {
1072 int max = size_goal;
1074 skb = tcp_write_queue_tail(sk);
1075 if (tcp_send_head(sk)) {
1076 if (skb->ip_summed == CHECKSUM_NONE)
1078 copy = max - skb->len;
1083 /* Allocate new segment. If the interface is SG,
1084 * allocate skb fitting to single page.
1086 if (!sk_stream_memory_free(sk))
1087 goto wait_for_sndbuf;
1089 skb = sk_stream_alloc_skb(sk,
1090 select_size(sk, sg),
1093 goto wait_for_memory;
1096 * Check whether we can use HW checksum.
1098 if (sk->sk_route_caps & NETIF_F_ALL_CSUM)
1099 skb->ip_summed = CHECKSUM_PARTIAL;
1101 skb_entail(sk, skb);
1106 /* Try to append data to the end of skb. */
1110 /* Where to copy to? */
1111 if (skb_availroom(skb) > 0) {
1112 /* We have some space in skb head. Superb! */
1113 copy = min_t(int, copy, skb_availroom(skb));
1114 err = skb_add_data_nocache(sk, skb, from, copy);
1119 int i = skb_shinfo(skb)->nr_frags;
1120 struct page *page = sk->sk_sndmsg_page;
1123 if (page && page_count(page) == 1)
1124 sk->sk_sndmsg_off = 0;
1126 off = sk->sk_sndmsg_off;
1128 if (skb_can_coalesce(skb, i, page, off) &&
1130 /* We can extend the last page
1133 } else if (i == MAX_SKB_FRAGS || !sg) {
1134 /* Need to add new fragment and cannot
1135 * do this because interface is non-SG,
1136 * or because all the page slots are
1138 tcp_mark_push(tp, skb);
1141 if (off == PAGE_SIZE) {
1143 sk->sk_sndmsg_page = page = NULL;
1149 if (copy > PAGE_SIZE - off)
1150 copy = PAGE_SIZE - off;
1152 if (!sk_wmem_schedule(sk, copy))
1153 goto wait_for_memory;
1156 /* Allocate new cache page. */
1157 if (!(page = sk_stream_alloc_page(sk)))
1158 goto wait_for_memory;
1161 /* Time to copy data. We are close to
1163 err = skb_copy_to_page_nocache(sk, from, skb,
1166 /* If this page was new, give it to the
1167 * socket so it does not get leaked.
1169 if (!sk->sk_sndmsg_page) {
1170 sk->sk_sndmsg_page = page;
1171 sk->sk_sndmsg_off = 0;
1176 /* Update the skb. */
1178 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1180 skb_fill_page_desc(skb, i, page, off, copy);
1181 if (sk->sk_sndmsg_page) {
1183 } else if (off + copy < PAGE_SIZE) {
1185 sk->sk_sndmsg_page = page;
1189 sk->sk_sndmsg_off = off + copy;
1193 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
1195 tp->write_seq += copy;
1196 TCP_SKB_CB(skb)->end_seq += copy;
1197 skb_shinfo(skb)->gso_segs = 0;
1201 if ((seglen -= copy) == 0 && iovlen == 0)
1204 if (skb->len < max || (flags & MSG_OOB) || unlikely(tp->repair))
1207 if (forced_push(tp)) {
1208 tcp_mark_push(tp, skb);
1209 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
1210 } else if (skb == tcp_send_head(sk))
1211 tcp_push_one(sk, mss_now);
1215 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1217 if (copied && likely(!tp->repair))
1218 tcp_push(sk, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH);
1220 if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
1223 mss_now = tcp_send_mss(sk, &size_goal, flags);
1228 if (copied && likely(!tp->repair))
1229 tcp_push(sk, flags, mss_now, tp->nonagle);
1235 tcp_unlink_write_queue(skb, sk);
1236 /* It is the one place in all of TCP, except connection
1237 * reset, where we can be unlinking the send_head.
1239 tcp_check_send_head(sk, skb);
1240 sk_wmem_free_skb(sk, skb);
1247 err = sk_stream_error(sk, flags, err);
1251 EXPORT_SYMBOL(tcp_sendmsg);
1254 * Handle reading urgent data. BSD has very simple semantics for
1255 * this, no blocking and very strange errors 8)
1258 static int tcp_recv_urg(struct sock *sk, struct msghdr *msg, int len, int flags)
1260 struct tcp_sock *tp = tcp_sk(sk);
1262 /* No URG data to read. */
1263 if (sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data ||
1264 tp->urg_data == TCP_URG_READ)
1265 return -EINVAL; /* Yes this is right ! */
1267 if (sk->sk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DONE))
1270 if (tp->urg_data & TCP_URG_VALID) {
1272 char c = tp->urg_data;
1274 if (!(flags & MSG_PEEK))
1275 tp->urg_data = TCP_URG_READ;
1277 /* Read urgent data. */
1278 msg->msg_flags |= MSG_OOB;
1281 if (!(flags & MSG_TRUNC))
1282 err = memcpy_toiovec(msg->msg_iov, &c, 1);
1285 msg->msg_flags |= MSG_TRUNC;
1287 return err ? -EFAULT : len;
1290 if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
1293 /* Fixed the recv(..., MSG_OOB) behaviour. BSD docs and
1294 * the available implementations agree in this case:
1295 * this call should never block, independent of the
1296 * blocking state of the socket.
1297 * Mike <pall@rz.uni-karlsruhe.de>
1302 static int tcp_peek_sndq(struct sock *sk, struct msghdr *msg, int len)
1304 struct sk_buff *skb;
1305 int copied = 0, err = 0;
1307 /* XXX -- need to support SO_PEEK_OFF */
1309 skb_queue_walk(&sk->sk_write_queue, skb) {
1310 err = skb_copy_datagram_iovec(skb, 0, msg->msg_iov, skb->len);
1317 return err ?: copied;
1320 /* Clean up the receive buffer for full frames taken by the user,
1321 * then send an ACK if necessary. COPIED is the number of bytes
1322 * tcp_recvmsg has given to the user so far, it speeds up the
1323 * calculation of whether or not we must ACK for the sake of
1326 void tcp_cleanup_rbuf(struct sock *sk, int copied)
1328 struct tcp_sock *tp = tcp_sk(sk);
1329 int time_to_ack = 0;
1331 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
1333 WARN(skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq),
1334 "cleanup rbuf bug: copied %X seq %X rcvnxt %X\n",
1335 tp->copied_seq, TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt);
1337 if (inet_csk_ack_scheduled(sk)) {
1338 const struct inet_connection_sock *icsk = inet_csk(sk);
1339 /* Delayed ACKs frequently hit locked sockets during bulk
1341 if (icsk->icsk_ack.blocked ||
1342 /* Once-per-two-segments ACK was not sent by tcp_input.c */
1343 tp->rcv_nxt - tp->rcv_wup > icsk->icsk_ack.rcv_mss ||
1345 * If this read emptied read buffer, we send ACK, if
1346 * connection is not bidirectional, user drained
1347 * receive buffer and there was a small segment
1351 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED2) ||
1352 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED) &&
1353 !icsk->icsk_ack.pingpong)) &&
1354 !atomic_read(&sk->sk_rmem_alloc)))
1358 /* We send an ACK if we can now advertise a non-zero window
1359 * which has been raised "significantly".
1361 * Even if window raised up to infinity, do not send window open ACK
1362 * in states, where we will not receive more. It is useless.
1364 if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) {
1365 __u32 rcv_window_now = tcp_receive_window(tp);
1367 /* Optimize, __tcp_select_window() is not cheap. */
1368 if (2*rcv_window_now <= tp->window_clamp) {
1369 __u32 new_window = __tcp_select_window(sk);
1371 /* Send ACK now, if this read freed lots of space
1372 * in our buffer. Certainly, new_window is new window.
1373 * We can advertise it now, if it is not less than current one.
1374 * "Lots" means "at least twice" here.
1376 if (new_window && new_window >= 2 * rcv_window_now)
1384 static void tcp_prequeue_process(struct sock *sk)
1386 struct sk_buff *skb;
1387 struct tcp_sock *tp = tcp_sk(sk);
1389 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPPREQUEUED);
1391 /* RX process wants to run with disabled BHs, though it is not
1394 while ((skb = __skb_dequeue(&tp->ucopy.prequeue)) != NULL)
1395 sk_backlog_rcv(sk, skb);
1398 /* Clear memory counter. */
1399 tp->ucopy.memory = 0;
1402 #ifdef CONFIG_NET_DMA
1403 static void tcp_service_net_dma(struct sock *sk, bool wait)
1405 dma_cookie_t done, used;
1406 dma_cookie_t last_issued;
1407 struct tcp_sock *tp = tcp_sk(sk);
1409 if (!tp->ucopy.dma_chan)
1412 last_issued = tp->ucopy.dma_cookie;
1413 dma_async_memcpy_issue_pending(tp->ucopy.dma_chan);
1416 if (dma_async_memcpy_complete(tp->ucopy.dma_chan,
1418 &used) == DMA_SUCCESS) {
1419 /* Safe to free early-copied skbs now */
1420 __skb_queue_purge(&sk->sk_async_wait_queue);
1423 struct sk_buff *skb;
1424 while ((skb = skb_peek(&sk->sk_async_wait_queue)) &&
1425 (dma_async_is_complete(skb->dma_cookie, done,
1426 used) == DMA_SUCCESS)) {
1427 __skb_dequeue(&sk->sk_async_wait_queue);
1435 static inline struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off)
1437 struct sk_buff *skb;
1440 skb_queue_walk(&sk->sk_receive_queue, skb) {
1441 offset = seq - TCP_SKB_CB(skb)->seq;
1442 if (tcp_hdr(skb)->syn)
1444 if (offset < skb->len || tcp_hdr(skb)->fin) {
1453 * This routine provides an alternative to tcp_recvmsg() for routines
1454 * that would like to handle copying from skbuffs directly in 'sendfile'
1457 * - It is assumed that the socket was locked by the caller.
1458 * - The routine does not block.
1459 * - At present, there is no support for reading OOB data
1460 * or for 'peeking' the socket using this routine
1461 * (although both would be easy to implement).
1463 int tcp_read_sock(struct sock *sk, read_descriptor_t *desc,
1464 sk_read_actor_t recv_actor)
1466 struct sk_buff *skb;
1467 struct tcp_sock *tp = tcp_sk(sk);
1468 u32 seq = tp->copied_seq;
1472 if (sk->sk_state == TCP_LISTEN)
1474 while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1475 if (offset < skb->len) {
1479 len = skb->len - offset;
1480 /* Stop reading if we hit a patch of urgent data */
1482 u32 urg_offset = tp->urg_seq - seq;
1483 if (urg_offset < len)
1488 used = recv_actor(desc, skb, offset, len);
1493 } else if (used <= len) {
1499 * If recv_actor drops the lock (e.g. TCP splice
1500 * receive) the skb pointer might be invalid when
1501 * getting here: tcp_collapse might have deleted it
1502 * while aggregating skbs from the socket queue.
1504 skb = tcp_recv_skb(sk, seq-1, &offset);
1505 if (!skb || (offset+1 != skb->len))
1508 if (tcp_hdr(skb)->fin) {
1509 sk_eat_skb(sk, skb, 0);
1513 sk_eat_skb(sk, skb, 0);
1516 tp->copied_seq = seq;
1518 tp->copied_seq = seq;
1520 tcp_rcv_space_adjust(sk);
1522 /* Clean up data we have read: This will do ACK frames. */
1524 tcp_cleanup_rbuf(sk, copied);
1527 EXPORT_SYMBOL(tcp_read_sock);
1530 * This routine copies from a sock struct into the user buffer.
1532 * Technical note: in 2.3 we work on _locked_ socket, so that
1533 * tricks with *seq access order and skb->users are not required.
1534 * Probably, code can be easily improved even more.
1537 int tcp_recvmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg,
1538 size_t len, int nonblock, int flags, int *addr_len)
1540 struct tcp_sock *tp = tcp_sk(sk);
1546 int target; /* Read at least this many bytes */
1548 struct task_struct *user_recv = NULL;
1549 int copied_early = 0;
1550 struct sk_buff *skb;
1556 if (sk->sk_state == TCP_LISTEN)
1559 timeo = sock_rcvtimeo(sk, nonblock);
1561 /* Urgent data needs to be handled specially. */
1562 if (flags & MSG_OOB)
1565 if (unlikely(tp->repair)) {
1567 if (!(flags & MSG_PEEK))
1570 if (tp->repair_queue == TCP_SEND_QUEUE)
1574 if (tp->repair_queue == TCP_NO_QUEUE)
1577 /* 'common' recv queue MSG_PEEK-ing */
1580 seq = &tp->copied_seq;
1581 if (flags & MSG_PEEK) {
1582 peek_seq = tp->copied_seq;
1586 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
1588 #ifdef CONFIG_NET_DMA
1589 tp->ucopy.dma_chan = NULL;
1591 skb = skb_peek_tail(&sk->sk_receive_queue);
1596 available = TCP_SKB_CB(skb)->seq + skb->len - (*seq);
1597 if ((available < target) &&
1598 (len > sysctl_tcp_dma_copybreak) && !(flags & MSG_PEEK) &&
1599 !sysctl_tcp_low_latency &&
1600 net_dma_find_channel()) {
1601 preempt_enable_no_resched();
1602 tp->ucopy.pinned_list =
1603 dma_pin_iovec_pages(msg->msg_iov, len);
1605 preempt_enable_no_resched();
1613 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
1614 if (tp->urg_data && tp->urg_seq == *seq) {
1617 if (signal_pending(current)) {
1618 copied = timeo ? sock_intr_errno(timeo) : -EAGAIN;
1623 /* Next get a buffer. */
1625 skb_queue_walk(&sk->sk_receive_queue, skb) {
1626 /* Now that we have two receive queues this
1629 if (WARN(before(*seq, TCP_SKB_CB(skb)->seq),
1630 "recvmsg bug: copied %X seq %X rcvnxt %X fl %X\n",
1631 *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt,
1635 offset = *seq - TCP_SKB_CB(skb)->seq;
1636 if (tcp_hdr(skb)->syn)
1638 if (offset < skb->len)
1640 if (tcp_hdr(skb)->fin)
1642 WARN(!(flags & MSG_PEEK),
1643 "recvmsg bug 2: copied %X seq %X rcvnxt %X fl %X\n",
1644 *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt, flags);
1647 /* Well, if we have backlog, try to process it now yet. */
1649 if (copied >= target && !sk->sk_backlog.tail)
1654 sk->sk_state == TCP_CLOSE ||
1655 (sk->sk_shutdown & RCV_SHUTDOWN) ||
1657 signal_pending(current))
1660 if (sock_flag(sk, SOCK_DONE))
1664 copied = sock_error(sk);
1668 if (sk->sk_shutdown & RCV_SHUTDOWN)
1671 if (sk->sk_state == TCP_CLOSE) {
1672 if (!sock_flag(sk, SOCK_DONE)) {
1673 /* This occurs when user tries to read
1674 * from never connected socket.
1687 if (signal_pending(current)) {
1688 copied = sock_intr_errno(timeo);
1693 tcp_cleanup_rbuf(sk, copied);
1695 if (!sysctl_tcp_low_latency && tp->ucopy.task == user_recv) {
1696 /* Install new reader */
1697 if (!user_recv && !(flags & (MSG_TRUNC | MSG_PEEK))) {
1698 user_recv = current;
1699 tp->ucopy.task = user_recv;
1700 tp->ucopy.iov = msg->msg_iov;
1703 tp->ucopy.len = len;
1705 WARN_ON(tp->copied_seq != tp->rcv_nxt &&
1706 !(flags & (MSG_PEEK | MSG_TRUNC)));
1708 /* Ugly... If prequeue is not empty, we have to
1709 * process it before releasing socket, otherwise
1710 * order will be broken at second iteration.
1711 * More elegant solution is required!!!
1713 * Look: we have the following (pseudo)queues:
1715 * 1. packets in flight
1720 * Each queue can be processed only if the next ones
1721 * are empty. At this point we have empty receive_queue.
1722 * But prequeue _can_ be not empty after 2nd iteration,
1723 * when we jumped to start of loop because backlog
1724 * processing added something to receive_queue.
1725 * We cannot release_sock(), because backlog contains
1726 * packets arrived _after_ prequeued ones.
1728 * Shortly, algorithm is clear --- to process all
1729 * the queues in order. We could make it more directly,
1730 * requeueing packets from backlog to prequeue, if
1731 * is not empty. It is more elegant, but eats cycles,
1734 if (!skb_queue_empty(&tp->ucopy.prequeue))
1737 /* __ Set realtime policy in scheduler __ */
1740 #ifdef CONFIG_NET_DMA
1741 if (tp->ucopy.dma_chan)
1742 dma_async_memcpy_issue_pending(tp->ucopy.dma_chan);
1744 if (copied >= target) {
1745 /* Do not sleep, just process backlog. */
1749 sk_wait_data(sk, &timeo);
1751 #ifdef CONFIG_NET_DMA
1752 tcp_service_net_dma(sk, false); /* Don't block */
1753 tp->ucopy.wakeup = 0;
1759 /* __ Restore normal policy in scheduler __ */
1761 if ((chunk = len - tp->ucopy.len) != 0) {
1762 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMBACKLOG, chunk);
1767 if (tp->rcv_nxt == tp->copied_seq &&
1768 !skb_queue_empty(&tp->ucopy.prequeue)) {
1770 tcp_prequeue_process(sk);
1772 if ((chunk = len - tp->ucopy.len) != 0) {
1773 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1779 if ((flags & MSG_PEEK) &&
1780 (peek_seq - copied - urg_hole != tp->copied_seq)) {
1781 if (net_ratelimit())
1782 printk(KERN_DEBUG "TCP(%s:%d): Application bug, race in MSG_PEEK.\n",
1783 current->comm, task_pid_nr(current));
1784 peek_seq = tp->copied_seq;
1789 /* Ok so how much can we use? */
1790 used = skb->len - offset;
1794 /* Do we have urgent data here? */
1796 u32 urg_offset = tp->urg_seq - *seq;
1797 if (urg_offset < used) {
1799 if (!sock_flag(sk, SOCK_URGINLINE)) {
1812 if (!(flags & MSG_TRUNC)) {
1813 #ifdef CONFIG_NET_DMA
1814 if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list)
1815 tp->ucopy.dma_chan = net_dma_find_channel();
1817 if (tp->ucopy.dma_chan) {
1818 tp->ucopy.dma_cookie = dma_skb_copy_datagram_iovec(
1819 tp->ucopy.dma_chan, skb, offset,
1821 tp->ucopy.pinned_list);
1823 if (tp->ucopy.dma_cookie < 0) {
1825 pr_alert("%s: dma_cookie < 0\n",
1828 /* Exception. Bailout! */
1834 dma_async_memcpy_issue_pending(tp->ucopy.dma_chan);
1836 if ((offset + used) == skb->len)
1842 err = skb_copy_datagram_iovec(skb, offset,
1843 msg->msg_iov, used);
1845 /* Exception. Bailout! */
1857 tcp_rcv_space_adjust(sk);
1860 if (tp->urg_data && after(tp->copied_seq, tp->urg_seq)) {
1862 tcp_fast_path_check(sk);
1864 if (used + offset < skb->len)
1867 if (tcp_hdr(skb)->fin)
1869 if (!(flags & MSG_PEEK)) {
1870 sk_eat_skb(sk, skb, copied_early);
1876 /* Process the FIN. */
1878 if (!(flags & MSG_PEEK)) {
1879 sk_eat_skb(sk, skb, copied_early);
1886 if (!skb_queue_empty(&tp->ucopy.prequeue)) {
1889 tp->ucopy.len = copied > 0 ? len : 0;
1891 tcp_prequeue_process(sk);
1893 if (copied > 0 && (chunk = len - tp->ucopy.len) != 0) {
1894 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1900 tp->ucopy.task = NULL;
1904 #ifdef CONFIG_NET_DMA
1905 tcp_service_net_dma(sk, true); /* Wait for queue to drain */
1906 tp->ucopy.dma_chan = NULL;
1908 if (tp->ucopy.pinned_list) {
1909 dma_unpin_iovec_pages(tp->ucopy.pinned_list);
1910 tp->ucopy.pinned_list = NULL;
1914 /* According to UNIX98, msg_name/msg_namelen are ignored
1915 * on connected socket. I was just happy when found this 8) --ANK
1918 /* Clean up data we have read: This will do ACK frames. */
1919 tcp_cleanup_rbuf(sk, copied);
1929 err = tcp_recv_urg(sk, msg, len, flags);
1933 err = tcp_peek_sndq(sk, msg, len);
1936 EXPORT_SYMBOL(tcp_recvmsg);
1938 void tcp_set_state(struct sock *sk, int state)
1940 int oldstate = sk->sk_state;
1943 case TCP_ESTABLISHED:
1944 if (oldstate != TCP_ESTABLISHED)
1945 TCP_INC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
1949 if (oldstate == TCP_CLOSE_WAIT || oldstate == TCP_ESTABLISHED)
1950 TCP_INC_STATS(sock_net(sk), TCP_MIB_ESTABRESETS);
1952 sk->sk_prot->unhash(sk);
1953 if (inet_csk(sk)->icsk_bind_hash &&
1954 !(sk->sk_userlocks & SOCK_BINDPORT_LOCK))
1958 if (oldstate == TCP_ESTABLISHED)
1959 TCP_DEC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
1962 /* Change state AFTER socket is unhashed to avoid closed
1963 * socket sitting in hash tables.
1965 sk->sk_state = state;
1968 SOCK_DEBUG(sk, "TCP sk=%p, State %s -> %s\n", sk, statename[oldstate], statename[state]);
1971 EXPORT_SYMBOL_GPL(tcp_set_state);
1974 * State processing on a close. This implements the state shift for
1975 * sending our FIN frame. Note that we only send a FIN for some
1976 * states. A shutdown() may have already sent the FIN, or we may be
1980 static const unsigned char new_state[16] = {
1981 /* current state: new state: action: */
1982 /* (Invalid) */ TCP_CLOSE,
1983 /* TCP_ESTABLISHED */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1984 /* TCP_SYN_SENT */ TCP_CLOSE,
1985 /* TCP_SYN_RECV */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1986 /* TCP_FIN_WAIT1 */ TCP_FIN_WAIT1,
1987 /* TCP_FIN_WAIT2 */ TCP_FIN_WAIT2,
1988 /* TCP_TIME_WAIT */ TCP_CLOSE,
1989 /* TCP_CLOSE */ TCP_CLOSE,
1990 /* TCP_CLOSE_WAIT */ TCP_LAST_ACK | TCP_ACTION_FIN,
1991 /* TCP_LAST_ACK */ TCP_LAST_ACK,
1992 /* TCP_LISTEN */ TCP_CLOSE,
1993 /* TCP_CLOSING */ TCP_CLOSING,
1996 static int tcp_close_state(struct sock *sk)
1998 int next = (int)new_state[sk->sk_state];
1999 int ns = next & TCP_STATE_MASK;
2001 tcp_set_state(sk, ns);
2003 return next & TCP_ACTION_FIN;
2007 * Shutdown the sending side of a connection. Much like close except
2008 * that we don't receive shut down or sock_set_flag(sk, SOCK_DEAD).
2011 void tcp_shutdown(struct sock *sk, int how)
2013 /* We need to grab some memory, and put together a FIN,
2014 * and then put it into the queue to be sent.
2015 * Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
2017 if (!(how & SEND_SHUTDOWN))
2020 /* If we've already sent a FIN, or it's a closed state, skip this. */
2021 if ((1 << sk->sk_state) &
2022 (TCPF_ESTABLISHED | TCPF_SYN_SENT |
2023 TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) {
2024 /* Clear out any half completed packets. FIN if needed. */
2025 if (tcp_close_state(sk))
2029 EXPORT_SYMBOL(tcp_shutdown);
2031 bool tcp_check_oom(struct sock *sk, int shift)
2033 bool too_many_orphans, out_of_socket_memory;
2035 too_many_orphans = tcp_too_many_orphans(sk, shift);
2036 out_of_socket_memory = tcp_out_of_memory(sk);
2038 if (too_many_orphans && net_ratelimit())
2039 pr_info("too many orphaned sockets\n");
2040 if (out_of_socket_memory && net_ratelimit())
2041 pr_info("out of memory -- consider tuning tcp_mem\n");
2042 return too_many_orphans || out_of_socket_memory;
2045 void tcp_close(struct sock *sk, long timeout)
2047 struct sk_buff *skb;
2048 int data_was_unread = 0;
2052 sk->sk_shutdown = SHUTDOWN_MASK;
2054 if (sk->sk_state == TCP_LISTEN) {
2055 tcp_set_state(sk, TCP_CLOSE);
2058 inet_csk_listen_stop(sk);
2060 goto adjudge_to_death;
2063 /* We need to flush the recv. buffs. We do this only on the
2064 * descriptor close, not protocol-sourced closes, because the
2065 * reader process may not have drained the data yet!
2067 while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) {
2068 u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq -
2070 data_was_unread += len;
2076 /* If socket has been already reset (e.g. in tcp_reset()) - kill it. */
2077 if (sk->sk_state == TCP_CLOSE)
2078 goto adjudge_to_death;
2080 /* As outlined in RFC 2525, section 2.17, we send a RST here because
2081 * data was lost. To witness the awful effects of the old behavior of
2082 * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk
2083 * GET in an FTP client, suspend the process, wait for the client to
2084 * advertise a zero window, then kill -9 the FTP client, wheee...
2085 * Note: timeout is always zero in such a case.
2087 if (unlikely(tcp_sk(sk)->repair)) {
2088 sk->sk_prot->disconnect(sk, 0);
2089 } else if (data_was_unread) {
2090 /* Unread data was tossed, zap the connection. */
2091 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONCLOSE);
2092 tcp_set_state(sk, TCP_CLOSE);
2093 tcp_send_active_reset(sk, sk->sk_allocation);
2094 } else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) {
2095 /* Check zero linger _after_ checking for unread data. */
2096 sk->sk_prot->disconnect(sk, 0);
2097 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
2098 } else if (tcp_close_state(sk)) {
2099 /* We FIN if the application ate all the data before
2100 * zapping the connection.
2103 /* RED-PEN. Formally speaking, we have broken TCP state
2104 * machine. State transitions:
2106 * TCP_ESTABLISHED -> TCP_FIN_WAIT1
2107 * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible)
2108 * TCP_CLOSE_WAIT -> TCP_LAST_ACK
2110 * are legal only when FIN has been sent (i.e. in window),
2111 * rather than queued out of window. Purists blame.
2113 * F.e. "RFC state" is ESTABLISHED,
2114 * if Linux state is FIN-WAIT-1, but FIN is still not sent.
2116 * The visible declinations are that sometimes
2117 * we enter time-wait state, when it is not required really
2118 * (harmless), do not send active resets, when they are
2119 * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
2120 * they look as CLOSING or LAST_ACK for Linux)
2121 * Probably, I missed some more holelets.
2127 sk_stream_wait_close(sk, timeout);
2130 state = sk->sk_state;
2134 /* It is the last release_sock in its life. It will remove backlog. */
2138 /* Now socket is owned by kernel and we acquire BH lock
2139 to finish close. No need to check for user refs.
2143 WARN_ON(sock_owned_by_user(sk));
2145 percpu_counter_inc(sk->sk_prot->orphan_count);
2147 /* Have we already been destroyed by a softirq or backlog? */
2148 if (state != TCP_CLOSE && sk->sk_state == TCP_CLOSE)
2151 /* This is a (useful) BSD violating of the RFC. There is a
2152 * problem with TCP as specified in that the other end could
2153 * keep a socket open forever with no application left this end.
2154 * We use a 3 minute timeout (about the same as BSD) then kill
2155 * our end. If they send after that then tough - BUT: long enough
2156 * that we won't make the old 4*rto = almost no time - whoops
2159 * Nope, it was not mistake. It is really desired behaviour
2160 * f.e. on http servers, when such sockets are useless, but
2161 * consume significant resources. Let's do it with special
2162 * linger2 option. --ANK
2165 if (sk->sk_state == TCP_FIN_WAIT2) {
2166 struct tcp_sock *tp = tcp_sk(sk);
2167 if (tp->linger2 < 0) {
2168 tcp_set_state(sk, TCP_CLOSE);
2169 tcp_send_active_reset(sk, GFP_ATOMIC);
2170 NET_INC_STATS_BH(sock_net(sk),
2171 LINUX_MIB_TCPABORTONLINGER);
2173 const int tmo = tcp_fin_time(sk);
2175 if (tmo > TCP_TIMEWAIT_LEN) {
2176 inet_csk_reset_keepalive_timer(sk,
2177 tmo - TCP_TIMEWAIT_LEN);
2179 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
2184 if (sk->sk_state != TCP_CLOSE) {
2186 if (tcp_check_oom(sk, 0)) {
2187 tcp_set_state(sk, TCP_CLOSE);
2188 tcp_send_active_reset(sk, GFP_ATOMIC);
2189 NET_INC_STATS_BH(sock_net(sk),
2190 LINUX_MIB_TCPABORTONMEMORY);
2194 if (sk->sk_state == TCP_CLOSE)
2195 inet_csk_destroy_sock(sk);
2196 /* Otherwise, socket is reprieved until protocol close. */
2203 EXPORT_SYMBOL(tcp_close);
2205 /* These states need RST on ABORT according to RFC793 */
2207 static inline int tcp_need_reset(int state)
2209 return (1 << state) &
2210 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 |
2211 TCPF_FIN_WAIT2 | TCPF_SYN_RECV);
2214 int tcp_disconnect(struct sock *sk, int flags)
2216 struct inet_sock *inet = inet_sk(sk);
2217 struct inet_connection_sock *icsk = inet_csk(sk);
2218 struct tcp_sock *tp = tcp_sk(sk);
2220 int old_state = sk->sk_state;
2222 if (old_state != TCP_CLOSE)
2223 tcp_set_state(sk, TCP_CLOSE);
2225 /* ABORT function of RFC793 */
2226 if (old_state == TCP_LISTEN) {
2227 inet_csk_listen_stop(sk);
2228 } else if (unlikely(tp->repair)) {
2229 sk->sk_err = ECONNABORTED;
2230 } else if (tcp_need_reset(old_state) ||
2231 (tp->snd_nxt != tp->write_seq &&
2232 (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK))) {
2233 /* The last check adjusts for discrepancy of Linux wrt. RFC
2236 tcp_send_active_reset(sk, gfp_any());
2237 sk->sk_err = ECONNRESET;
2238 } else if (old_state == TCP_SYN_SENT)
2239 sk->sk_err = ECONNRESET;
2241 tcp_clear_xmit_timers(sk);
2242 __skb_queue_purge(&sk->sk_receive_queue);
2243 tcp_write_queue_purge(sk);
2244 __skb_queue_purge(&tp->out_of_order_queue);
2245 #ifdef CONFIG_NET_DMA
2246 __skb_queue_purge(&sk->sk_async_wait_queue);
2249 inet->inet_dport = 0;
2251 if (!(sk->sk_userlocks & SOCK_BINDADDR_LOCK))
2252 inet_reset_saddr(sk);
2254 sk->sk_shutdown = 0;
2255 sock_reset_flag(sk, SOCK_DONE);
2257 if ((tp->write_seq += tp->max_window + 2) == 0)
2259 icsk->icsk_backoff = 0;
2261 icsk->icsk_probes_out = 0;
2262 tp->packets_out = 0;
2263 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
2264 tp->snd_cwnd_cnt = 0;
2265 tp->bytes_acked = 0;
2266 tp->window_clamp = 0;
2267 tcp_set_ca_state(sk, TCP_CA_Open);
2268 tcp_clear_retrans(tp);
2269 inet_csk_delack_init(sk);
2270 tcp_init_send_head(sk);
2271 memset(&tp->rx_opt, 0, sizeof(tp->rx_opt));
2274 WARN_ON(inet->inet_num && !icsk->icsk_bind_hash);
2276 sk->sk_error_report(sk);
2279 EXPORT_SYMBOL(tcp_disconnect);
2281 static inline int tcp_can_repair_sock(struct sock *sk)
2283 return capable(CAP_NET_ADMIN) &&
2284 ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_ESTABLISHED));
2287 static int tcp_repair_options_est(struct tcp_sock *tp,
2288 struct tcp_repair_opt __user *optbuf, unsigned int len)
2290 struct tcp_repair_opt opt;
2292 while (len >= sizeof(opt)) {
2293 if (copy_from_user(&opt, optbuf, sizeof(opt)))
2299 switch (opt.opt_code) {
2301 tp->rx_opt.mss_clamp = opt.opt_val;
2304 if (opt.opt_val > 14)
2307 tp->rx_opt.snd_wscale = opt.opt_val;
2309 case TCPOPT_SACK_PERM:
2310 if (opt.opt_val != 0)
2313 tp->rx_opt.sack_ok |= TCP_SACK_SEEN;
2314 if (sysctl_tcp_fack)
2315 tcp_enable_fack(tp);
2317 case TCPOPT_TIMESTAMP:
2318 if (opt.opt_val != 0)
2321 tp->rx_opt.tstamp_ok = 1;
2330 * Socket option code for TCP.
2332 static int do_tcp_setsockopt(struct sock *sk, int level,
2333 int optname, char __user *optval, unsigned int optlen)
2335 struct tcp_sock *tp = tcp_sk(sk);
2336 struct inet_connection_sock *icsk = inet_csk(sk);
2340 /* These are data/string values, all the others are ints */
2342 case TCP_CONGESTION: {
2343 char name[TCP_CA_NAME_MAX];
2348 val = strncpy_from_user(name, optval,
2349 min_t(long, TCP_CA_NAME_MAX-1, optlen));
2355 err = tcp_set_congestion_control(sk, name);
2359 case TCP_COOKIE_TRANSACTIONS: {
2360 struct tcp_cookie_transactions ctd;
2361 struct tcp_cookie_values *cvp = NULL;
2363 if (sizeof(ctd) > optlen)
2365 if (copy_from_user(&ctd, optval, sizeof(ctd)))
2368 if (ctd.tcpct_used > sizeof(ctd.tcpct_value) ||
2369 ctd.tcpct_s_data_desired > TCP_MSS_DESIRED)
2372 if (ctd.tcpct_cookie_desired == 0) {
2373 /* default to global value */
2374 } else if ((0x1 & ctd.tcpct_cookie_desired) ||
2375 ctd.tcpct_cookie_desired > TCP_COOKIE_MAX ||
2376 ctd.tcpct_cookie_desired < TCP_COOKIE_MIN) {
2380 if (TCP_COOKIE_OUT_NEVER & ctd.tcpct_flags) {
2381 /* Supercedes all other values */
2383 if (tp->cookie_values != NULL) {
2384 kref_put(&tp->cookie_values->kref,
2385 tcp_cookie_values_release);
2386 tp->cookie_values = NULL;
2388 tp->rx_opt.cookie_in_always = 0; /* false */
2389 tp->rx_opt.cookie_out_never = 1; /* true */
2394 /* Allocate ancillary memory before locking.
2396 if (ctd.tcpct_used > 0 ||
2397 (tp->cookie_values == NULL &&
2398 (sysctl_tcp_cookie_size > 0 ||
2399 ctd.tcpct_cookie_desired > 0 ||
2400 ctd.tcpct_s_data_desired > 0))) {
2401 cvp = kzalloc(sizeof(*cvp) + ctd.tcpct_used,
2406 kref_init(&cvp->kref);
2409 tp->rx_opt.cookie_in_always =
2410 (TCP_COOKIE_IN_ALWAYS & ctd.tcpct_flags);
2411 tp->rx_opt.cookie_out_never = 0; /* false */
2413 if (tp->cookie_values != NULL) {
2415 /* Changed values are recorded by a changed
2416 * pointer, ensuring the cookie will differ,
2417 * without separately hashing each value later.
2419 kref_put(&tp->cookie_values->kref,
2420 tcp_cookie_values_release);
2422 cvp = tp->cookie_values;
2427 cvp->cookie_desired = ctd.tcpct_cookie_desired;
2429 if (ctd.tcpct_used > 0) {
2430 memcpy(cvp->s_data_payload, ctd.tcpct_value,
2432 cvp->s_data_desired = ctd.tcpct_used;
2433 cvp->s_data_constant = 1; /* true */
2435 /* No constant payload data. */
2436 cvp->s_data_desired = ctd.tcpct_s_data_desired;
2437 cvp->s_data_constant = 0; /* false */
2440 tp->cookie_values = cvp;
2450 if (optlen < sizeof(int))
2453 if (get_user(val, (int __user *)optval))
2460 /* Values greater than interface MTU won't take effect. However
2461 * at the point when this call is done we typically don't yet
2462 * know which interface is going to be used */
2463 if (val < TCP_MIN_MSS || val > MAX_TCP_WINDOW) {
2467 tp->rx_opt.user_mss = val;
2472 /* TCP_NODELAY is weaker than TCP_CORK, so that
2473 * this option on corked socket is remembered, but
2474 * it is not activated until cork is cleared.
2476 * However, when TCP_NODELAY is set we make
2477 * an explicit push, which overrides even TCP_CORK
2478 * for currently queued segments.
2480 tp->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH;
2481 tcp_push_pending_frames(sk);
2483 tp->nonagle &= ~TCP_NAGLE_OFF;
2487 case TCP_THIN_LINEAR_TIMEOUTS:
2488 if (val < 0 || val > 1)
2494 case TCP_THIN_DUPACK:
2495 if (val < 0 || val > 1)
2498 tp->thin_dupack = val;
2499 if (tp->thin_dupack)
2500 tcp_disable_early_retrans(tp);
2504 if (!tcp_can_repair_sock(sk))
2506 else if (val == 1) {
2508 sk->sk_reuse = SK_FORCE_REUSE;
2509 tp->repair_queue = TCP_NO_QUEUE;
2510 } else if (val == 0) {
2512 sk->sk_reuse = SK_NO_REUSE;
2513 tcp_send_window_probe(sk);
2519 case TCP_REPAIR_QUEUE:
2522 else if (val < TCP_QUEUES_NR)
2523 tp->repair_queue = val;
2529 if (sk->sk_state != TCP_CLOSE)
2531 else if (tp->repair_queue == TCP_SEND_QUEUE)
2532 tp->write_seq = val;
2533 else if (tp->repair_queue == TCP_RECV_QUEUE)
2539 case TCP_REPAIR_OPTIONS:
2542 else if (sk->sk_state == TCP_ESTABLISHED)
2543 err = tcp_repair_options_est(tp,
2544 (struct tcp_repair_opt __user *)optval,
2551 /* When set indicates to always queue non-full frames.
2552 * Later the user clears this option and we transmit
2553 * any pending partial frames in the queue. This is
2554 * meant to be used alongside sendfile() to get properly
2555 * filled frames when the user (for example) must write
2556 * out headers with a write() call first and then use
2557 * sendfile to send out the data parts.
2559 * TCP_CORK can be set together with TCP_NODELAY and it is
2560 * stronger than TCP_NODELAY.
2563 tp->nonagle |= TCP_NAGLE_CORK;
2565 tp->nonagle &= ~TCP_NAGLE_CORK;
2566 if (tp->nonagle&TCP_NAGLE_OFF)
2567 tp->nonagle |= TCP_NAGLE_PUSH;
2568 tcp_push_pending_frames(sk);
2573 if (val < 1 || val > MAX_TCP_KEEPIDLE)
2576 tp->keepalive_time = val * HZ;
2577 if (sock_flag(sk, SOCK_KEEPOPEN) &&
2578 !((1 << sk->sk_state) &
2579 (TCPF_CLOSE | TCPF_LISTEN))) {
2580 u32 elapsed = keepalive_time_elapsed(tp);
2581 if (tp->keepalive_time > elapsed)
2582 elapsed = tp->keepalive_time - elapsed;
2585 inet_csk_reset_keepalive_timer(sk, elapsed);
2590 if (val < 1 || val > MAX_TCP_KEEPINTVL)
2593 tp->keepalive_intvl = val * HZ;
2596 if (val < 1 || val > MAX_TCP_KEEPCNT)
2599 tp->keepalive_probes = val;
2602 if (val < 1 || val > MAX_TCP_SYNCNT)
2605 icsk->icsk_syn_retries = val;
2611 else if (val > sysctl_tcp_fin_timeout / HZ)
2614 tp->linger2 = val * HZ;
2617 case TCP_DEFER_ACCEPT:
2618 /* Translate value in seconds to number of retransmits */
2619 icsk->icsk_accept_queue.rskq_defer_accept =
2620 secs_to_retrans(val, TCP_TIMEOUT_INIT / HZ,
2624 case TCP_WINDOW_CLAMP:
2626 if (sk->sk_state != TCP_CLOSE) {
2630 tp->window_clamp = 0;
2632 tp->window_clamp = val < SOCK_MIN_RCVBUF / 2 ?
2633 SOCK_MIN_RCVBUF / 2 : val;
2638 icsk->icsk_ack.pingpong = 1;
2640 icsk->icsk_ack.pingpong = 0;
2641 if ((1 << sk->sk_state) &
2642 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) &&
2643 inet_csk_ack_scheduled(sk)) {
2644 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
2645 tcp_cleanup_rbuf(sk, 1);
2647 icsk->icsk_ack.pingpong = 1;
2652 #ifdef CONFIG_TCP_MD5SIG
2654 /* Read the IP->Key mappings from userspace */
2655 err = tp->af_specific->md5_parse(sk, optval, optlen);
2658 case TCP_USER_TIMEOUT:
2659 /* Cap the max timeout in ms TCP will retry/retrans
2660 * before giving up and aborting (ETIMEDOUT) a connection.
2662 icsk->icsk_user_timeout = msecs_to_jiffies(val);
2673 int tcp_setsockopt(struct sock *sk, int level, int optname, char __user *optval,
2674 unsigned int optlen)
2676 const struct inet_connection_sock *icsk = inet_csk(sk);
2678 if (level != SOL_TCP)
2679 return icsk->icsk_af_ops->setsockopt(sk, level, optname,
2681 return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2683 EXPORT_SYMBOL(tcp_setsockopt);
2685 #ifdef CONFIG_COMPAT
2686 int compat_tcp_setsockopt(struct sock *sk, int level, int optname,
2687 char __user *optval, unsigned int optlen)
2689 if (level != SOL_TCP)
2690 return inet_csk_compat_setsockopt(sk, level, optname,
2692 return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2694 EXPORT_SYMBOL(compat_tcp_setsockopt);
2697 /* Return information about state of tcp endpoint in API format. */
2698 void tcp_get_info(const struct sock *sk, struct tcp_info *info)
2700 const struct tcp_sock *tp = tcp_sk(sk);
2701 const struct inet_connection_sock *icsk = inet_csk(sk);
2702 u32 now = tcp_time_stamp;
2704 memset(info, 0, sizeof(*info));
2706 info->tcpi_state = sk->sk_state;
2707 info->tcpi_ca_state = icsk->icsk_ca_state;
2708 info->tcpi_retransmits = icsk->icsk_retransmits;
2709 info->tcpi_probes = icsk->icsk_probes_out;
2710 info->tcpi_backoff = icsk->icsk_backoff;
2712 if (tp->rx_opt.tstamp_ok)
2713 info->tcpi_options |= TCPI_OPT_TIMESTAMPS;
2714 if (tcp_is_sack(tp))
2715 info->tcpi_options |= TCPI_OPT_SACK;
2716 if (tp->rx_opt.wscale_ok) {
2717 info->tcpi_options |= TCPI_OPT_WSCALE;
2718 info->tcpi_snd_wscale = tp->rx_opt.snd_wscale;
2719 info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale;
2722 if (tp->ecn_flags & TCP_ECN_OK)
2723 info->tcpi_options |= TCPI_OPT_ECN;
2724 if (tp->ecn_flags & TCP_ECN_SEEN)
2725 info->tcpi_options |= TCPI_OPT_ECN_SEEN;
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_COOKIE_TRANSACTIONS: {
2846 struct tcp_cookie_transactions ctd;
2847 struct tcp_cookie_values *cvp = tp->cookie_values;
2849 if (get_user(len, optlen))
2851 if (len < sizeof(ctd))
2854 memset(&ctd, 0, sizeof(ctd));
2855 ctd.tcpct_flags = (tp->rx_opt.cookie_in_always ?
2856 TCP_COOKIE_IN_ALWAYS : 0)
2857 | (tp->rx_opt.cookie_out_never ?
2858 TCP_COOKIE_OUT_NEVER : 0);
2861 ctd.tcpct_flags |= (cvp->s_data_in ?
2863 | (cvp->s_data_out ?
2864 TCP_S_DATA_OUT : 0);
2866 ctd.tcpct_cookie_desired = cvp->cookie_desired;
2867 ctd.tcpct_s_data_desired = cvp->s_data_desired;
2869 memcpy(&ctd.tcpct_value[0], &cvp->cookie_pair[0],
2870 cvp->cookie_pair_size);
2871 ctd.tcpct_used = cvp->cookie_pair_size;
2874 if (put_user(sizeof(ctd), optlen))
2876 if (copy_to_user(optval, &ctd, sizeof(ctd)))
2880 case TCP_THIN_LINEAR_TIMEOUTS:
2883 case TCP_THIN_DUPACK:
2884 val = tp->thin_dupack;
2891 case TCP_REPAIR_QUEUE:
2893 val = tp->repair_queue;
2899 if (tp->repair_queue == TCP_SEND_QUEUE)
2900 val = tp->write_seq;
2901 else if (tp->repair_queue == TCP_RECV_QUEUE)
2907 case TCP_USER_TIMEOUT:
2908 val = jiffies_to_msecs(icsk->icsk_user_timeout);
2911 return -ENOPROTOOPT;
2914 if (put_user(len, optlen))
2916 if (copy_to_user(optval, &val, len))
2921 int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval,
2924 struct inet_connection_sock *icsk = inet_csk(sk);
2926 if (level != SOL_TCP)
2927 return icsk->icsk_af_ops->getsockopt(sk, level, optname,
2929 return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2931 EXPORT_SYMBOL(tcp_getsockopt);
2933 #ifdef CONFIG_COMPAT
2934 int compat_tcp_getsockopt(struct sock *sk, int level, int optname,
2935 char __user *optval, int __user *optlen)
2937 if (level != SOL_TCP)
2938 return inet_csk_compat_getsockopt(sk, level, optname,
2940 return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2942 EXPORT_SYMBOL(compat_tcp_getsockopt);
2945 struct sk_buff *tcp_tso_segment(struct sk_buff *skb,
2946 netdev_features_t features)
2948 struct sk_buff *segs = ERR_PTR(-EINVAL);
2953 unsigned int oldlen;
2956 if (!pskb_may_pull(skb, sizeof(*th)))
2960 thlen = th->doff * 4;
2961 if (thlen < sizeof(*th))
2964 if (!pskb_may_pull(skb, thlen))
2967 oldlen = (u16)~skb->len;
2968 __skb_pull(skb, thlen);
2970 mss = skb_shinfo(skb)->gso_size;
2971 if (unlikely(skb->len <= mss))
2974 if (skb_gso_ok(skb, features | NETIF_F_GSO_ROBUST)) {
2975 /* Packet is from an untrusted source, reset gso_segs. */
2976 int type = skb_shinfo(skb)->gso_type;
2984 !(type & (SKB_GSO_TCPV4 | SKB_GSO_TCPV6))))
2987 skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss);
2993 segs = skb_segment(skb, features);
2997 delta = htonl(oldlen + (thlen + mss));
3001 seq = ntohl(th->seq);
3004 th->fin = th->psh = 0;
3006 th->check = ~csum_fold((__force __wsum)((__force u32)th->check +
3007 (__force u32)delta));
3008 if (skb->ip_summed != CHECKSUM_PARTIAL)
3010 csum_fold(csum_partial(skb_transport_header(skb),
3017 th->seq = htonl(seq);
3019 } while (skb->next);
3021 delta = htonl(oldlen + (skb->tail - skb->transport_header) +
3023 th->check = ~csum_fold((__force __wsum)((__force u32)th->check +
3024 (__force u32)delta));
3025 if (skb->ip_summed != CHECKSUM_PARTIAL)
3026 th->check = csum_fold(csum_partial(skb_transport_header(skb),
3032 EXPORT_SYMBOL(tcp_tso_segment);
3034 struct sk_buff **tcp_gro_receive(struct sk_buff **head, struct sk_buff *skb)
3036 struct sk_buff **pp = NULL;
3043 unsigned int mss = 1;
3049 off = skb_gro_offset(skb);
3050 hlen = off + sizeof(*th);
3051 th = skb_gro_header_fast(skb, off);
3052 if (skb_gro_header_hard(skb, hlen)) {
3053 th = skb_gro_header_slow(skb, hlen, off);
3058 thlen = th->doff * 4;
3059 if (thlen < sizeof(*th))
3063 if (skb_gro_header_hard(skb, hlen)) {
3064 th = skb_gro_header_slow(skb, hlen, off);
3069 skb_gro_pull(skb, thlen);
3071 len = skb_gro_len(skb);
3072 flags = tcp_flag_word(th);
3074 for (; (p = *head); head = &p->next) {
3075 if (!NAPI_GRO_CB(p)->same_flow)
3080 if (*(u32 *)&th->source ^ *(u32 *)&th2->source) {
3081 NAPI_GRO_CB(p)->same_flow = 0;
3088 goto out_check_final;
3091 flush = NAPI_GRO_CB(p)->flush;
3092 flush |= (__force int)(flags & TCP_FLAG_CWR);
3093 flush |= (__force int)((flags ^ tcp_flag_word(th2)) &
3094 ~(TCP_FLAG_CWR | TCP_FLAG_FIN | TCP_FLAG_PSH));
3095 flush |= (__force int)(th->ack_seq ^ th2->ack_seq);
3096 for (i = sizeof(*th); i < thlen; i += 4)
3097 flush |= *(u32 *)((u8 *)th + i) ^
3098 *(u32 *)((u8 *)th2 + i);
3100 mss = skb_shinfo(p)->gso_size;
3102 flush |= (len - 1) >= mss;
3103 flush |= (ntohl(th2->seq) + skb_gro_len(p)) ^ ntohl(th->seq);
3105 if (flush || skb_gro_receive(head, skb)) {
3107 goto out_check_final;
3112 tcp_flag_word(th2) |= flags & (TCP_FLAG_FIN | TCP_FLAG_PSH);
3116 flush |= (__force int)(flags & (TCP_FLAG_URG | TCP_FLAG_PSH |
3117 TCP_FLAG_RST | TCP_FLAG_SYN |
3120 if (p && (!NAPI_GRO_CB(skb)->same_flow || flush))
3124 NAPI_GRO_CB(skb)->flush |= flush;
3128 EXPORT_SYMBOL(tcp_gro_receive);
3130 int tcp_gro_complete(struct sk_buff *skb)
3132 struct tcphdr *th = tcp_hdr(skb);
3134 skb->csum_start = skb_transport_header(skb) - skb->head;
3135 skb->csum_offset = offsetof(struct tcphdr, check);
3136 skb->ip_summed = CHECKSUM_PARTIAL;
3138 skb_shinfo(skb)->gso_segs = NAPI_GRO_CB(skb)->count;
3141 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
3145 EXPORT_SYMBOL(tcp_gro_complete);
3147 #ifdef CONFIG_TCP_MD5SIG
3148 static unsigned long tcp_md5sig_users;
3149 static struct tcp_md5sig_pool __percpu *tcp_md5sig_pool;
3150 static DEFINE_SPINLOCK(tcp_md5sig_pool_lock);
3152 static void __tcp_free_md5sig_pool(struct tcp_md5sig_pool __percpu *pool)
3156 for_each_possible_cpu(cpu) {
3157 struct tcp_md5sig_pool *p = per_cpu_ptr(pool, cpu);
3159 if (p->md5_desc.tfm)
3160 crypto_free_hash(p->md5_desc.tfm);
3165 void tcp_free_md5sig_pool(void)
3167 struct tcp_md5sig_pool __percpu *pool = NULL;
3169 spin_lock_bh(&tcp_md5sig_pool_lock);
3170 if (--tcp_md5sig_users == 0) {
3171 pool = tcp_md5sig_pool;
3172 tcp_md5sig_pool = NULL;
3174 spin_unlock_bh(&tcp_md5sig_pool_lock);
3176 __tcp_free_md5sig_pool(pool);
3178 EXPORT_SYMBOL(tcp_free_md5sig_pool);
3180 static struct tcp_md5sig_pool __percpu *
3181 __tcp_alloc_md5sig_pool(struct sock *sk)
3184 struct tcp_md5sig_pool __percpu *pool;
3186 pool = alloc_percpu(struct tcp_md5sig_pool);
3190 for_each_possible_cpu(cpu) {
3191 struct crypto_hash *hash;
3193 hash = crypto_alloc_hash("md5", 0, CRYPTO_ALG_ASYNC);
3194 if (!hash || IS_ERR(hash))
3197 per_cpu_ptr(pool, cpu)->md5_desc.tfm = hash;
3201 __tcp_free_md5sig_pool(pool);
3205 struct tcp_md5sig_pool __percpu *tcp_alloc_md5sig_pool(struct sock *sk)
3207 struct tcp_md5sig_pool __percpu *pool;
3211 spin_lock_bh(&tcp_md5sig_pool_lock);
3212 pool = tcp_md5sig_pool;
3213 if (tcp_md5sig_users++ == 0) {
3215 spin_unlock_bh(&tcp_md5sig_pool_lock);
3218 spin_unlock_bh(&tcp_md5sig_pool_lock);
3222 spin_unlock_bh(&tcp_md5sig_pool_lock);
3225 /* we cannot hold spinlock here because this may sleep. */
3226 struct tcp_md5sig_pool __percpu *p;
3228 p = __tcp_alloc_md5sig_pool(sk);
3229 spin_lock_bh(&tcp_md5sig_pool_lock);
3232 spin_unlock_bh(&tcp_md5sig_pool_lock);
3235 pool = tcp_md5sig_pool;
3237 /* oops, it has already been assigned. */
3238 spin_unlock_bh(&tcp_md5sig_pool_lock);
3239 __tcp_free_md5sig_pool(p);
3241 tcp_md5sig_pool = pool = p;
3242 spin_unlock_bh(&tcp_md5sig_pool_lock);
3247 EXPORT_SYMBOL(tcp_alloc_md5sig_pool);
3251 * tcp_get_md5sig_pool - get md5sig_pool for this user
3253 * We use percpu structure, so if we succeed, we exit with preemption
3254 * and BH disabled, to make sure another thread or softirq handling
3255 * wont try to get same context.
3257 struct tcp_md5sig_pool *tcp_get_md5sig_pool(void)
3259 struct tcp_md5sig_pool __percpu *p;
3263 spin_lock(&tcp_md5sig_pool_lock);
3264 p = tcp_md5sig_pool;
3267 spin_unlock(&tcp_md5sig_pool_lock);
3270 return this_cpu_ptr(p);
3275 EXPORT_SYMBOL(tcp_get_md5sig_pool);
3277 void tcp_put_md5sig_pool(void)
3280 tcp_free_md5sig_pool();
3282 EXPORT_SYMBOL(tcp_put_md5sig_pool);
3284 int tcp_md5_hash_header(struct tcp_md5sig_pool *hp,
3285 const struct tcphdr *th)
3287 struct scatterlist sg;
3291 /* We are not allowed to change tcphdr, make a local copy */
3292 memcpy(&hdr, th, sizeof(hdr));
3295 /* options aren't included in the hash */
3296 sg_init_one(&sg, &hdr, sizeof(hdr));
3297 err = crypto_hash_update(&hp->md5_desc, &sg, sizeof(hdr));
3300 EXPORT_SYMBOL(tcp_md5_hash_header);
3302 int tcp_md5_hash_skb_data(struct tcp_md5sig_pool *hp,
3303 const struct sk_buff *skb, unsigned int header_len)
3305 struct scatterlist sg;
3306 const struct tcphdr *tp = tcp_hdr(skb);
3307 struct hash_desc *desc = &hp->md5_desc;
3309 const unsigned int head_data_len = skb_headlen(skb) > header_len ?
3310 skb_headlen(skb) - header_len : 0;
3311 const struct skb_shared_info *shi = skb_shinfo(skb);
3312 struct sk_buff *frag_iter;
3314 sg_init_table(&sg, 1);
3316 sg_set_buf(&sg, ((u8 *) tp) + header_len, head_data_len);
3317 if (crypto_hash_update(desc, &sg, head_data_len))
3320 for (i = 0; i < shi->nr_frags; ++i) {
3321 const struct skb_frag_struct *f = &shi->frags[i];
3322 struct page *page = skb_frag_page(f);
3323 sg_set_page(&sg, page, skb_frag_size(f), f->page_offset);
3324 if (crypto_hash_update(desc, &sg, skb_frag_size(f)))
3328 skb_walk_frags(skb, frag_iter)
3329 if (tcp_md5_hash_skb_data(hp, frag_iter, 0))
3334 EXPORT_SYMBOL(tcp_md5_hash_skb_data);
3336 int tcp_md5_hash_key(struct tcp_md5sig_pool *hp, const struct tcp_md5sig_key *key)
3338 struct scatterlist sg;
3340 sg_init_one(&sg, key->key, key->keylen);
3341 return crypto_hash_update(&hp->md5_desc, &sg, key->keylen);
3343 EXPORT_SYMBOL(tcp_md5_hash_key);
3348 * Each Responder maintains up to two secret values concurrently for
3349 * efficient secret rollover. Each secret value has 4 states:
3351 * Generating. (tcp_secret_generating != tcp_secret_primary)
3352 * Generates new Responder-Cookies, but not yet used for primary
3353 * verification. This is a short-term state, typically lasting only
3354 * one round trip time (RTT).
3356 * Primary. (tcp_secret_generating == tcp_secret_primary)
3357 * Used both for generation and primary verification.
3359 * Retiring. (tcp_secret_retiring != tcp_secret_secondary)
3360 * Used for verification, until the first failure that can be
3361 * verified by the newer Generating secret. At that time, this
3362 * cookie's state is changed to Secondary, and the Generating
3363 * cookie's state is changed to Primary. This is a short-term state,
3364 * typically lasting only one round trip time (RTT).
3366 * Secondary. (tcp_secret_retiring == tcp_secret_secondary)
3367 * Used for secondary verification, after primary verification
3368 * failures. This state lasts no more than twice the Maximum Segment
3369 * Lifetime (2MSL). Then, the secret is discarded.
3371 struct tcp_cookie_secret {
3372 /* The secret is divided into two parts. The digest part is the
3373 * equivalent of previously hashing a secret and saving the state,
3374 * and serves as an initialization vector (IV). The message part
3375 * serves as the trailing secret.
3377 u32 secrets[COOKIE_WORKSPACE_WORDS];
3378 unsigned long expires;
3381 #define TCP_SECRET_1MSL (HZ * TCP_PAWS_MSL)
3382 #define TCP_SECRET_2MSL (HZ * TCP_PAWS_MSL * 2)
3383 #define TCP_SECRET_LIFE (HZ * 600)
3385 static struct tcp_cookie_secret tcp_secret_one;
3386 static struct tcp_cookie_secret tcp_secret_two;
3388 /* Essentially a circular list, without dynamic allocation. */
3389 static struct tcp_cookie_secret *tcp_secret_generating;
3390 static struct tcp_cookie_secret *tcp_secret_primary;
3391 static struct tcp_cookie_secret *tcp_secret_retiring;
3392 static struct tcp_cookie_secret *tcp_secret_secondary;
3394 static DEFINE_SPINLOCK(tcp_secret_locker);
3396 /* Select a pseudo-random word in the cookie workspace.
3398 static inline u32 tcp_cookie_work(const u32 *ws, const int n)
3400 return ws[COOKIE_DIGEST_WORDS + ((COOKIE_MESSAGE_WORDS-1) & ws[n])];
3403 /* Fill bakery[COOKIE_WORKSPACE_WORDS] with generator, updating as needed.
3404 * Called in softirq context.
3405 * Returns: 0 for success.
3407 int tcp_cookie_generator(u32 *bakery)
3409 unsigned long jiffy = jiffies;
3411 if (unlikely(time_after_eq(jiffy, tcp_secret_generating->expires))) {
3412 spin_lock_bh(&tcp_secret_locker);
3413 if (!time_after_eq(jiffy, tcp_secret_generating->expires)) {
3414 /* refreshed by another */
3416 &tcp_secret_generating->secrets[0],
3417 COOKIE_WORKSPACE_WORDS);
3419 /* still needs refreshing */
3420 get_random_bytes(bakery, COOKIE_WORKSPACE_WORDS);
3422 /* The first time, paranoia assumes that the
3423 * randomization function isn't as strong. But,
3424 * this secret initialization is delayed until
3425 * the last possible moment (packet arrival).
3426 * Although that time is observable, it is
3427 * unpredictably variable. Mash in the most
3428 * volatile clock bits available, and expire the
3429 * secret extra quickly.
3431 if (unlikely(tcp_secret_primary->expires ==
3432 tcp_secret_secondary->expires)) {
3435 getnstimeofday(&tv);
3436 bakery[COOKIE_DIGEST_WORDS+0] ^=
3439 tcp_secret_secondary->expires = jiffy
3441 + (0x0f & tcp_cookie_work(bakery, 0));
3443 tcp_secret_secondary->expires = jiffy
3445 + (0xff & tcp_cookie_work(bakery, 1));
3446 tcp_secret_primary->expires = jiffy
3448 + (0x1f & tcp_cookie_work(bakery, 2));
3450 memcpy(&tcp_secret_secondary->secrets[0],
3451 bakery, COOKIE_WORKSPACE_WORDS);
3453 rcu_assign_pointer(tcp_secret_generating,
3454 tcp_secret_secondary);
3455 rcu_assign_pointer(tcp_secret_retiring,
3456 tcp_secret_primary);
3458 * Neither call_rcu() nor synchronize_rcu() needed.
3459 * Retiring data is not freed. It is replaced after
3460 * further (locked) pointer updates, and a quiet time
3461 * (minimum 1MSL, maximum LIFE - 2MSL).
3464 spin_unlock_bh(&tcp_secret_locker);
3468 &rcu_dereference(tcp_secret_generating)->secrets[0],
3469 COOKIE_WORKSPACE_WORDS);
3470 rcu_read_unlock_bh();
3474 EXPORT_SYMBOL(tcp_cookie_generator);
3476 void tcp_done(struct sock *sk)
3478 if (sk->sk_state == TCP_SYN_SENT || sk->sk_state == TCP_SYN_RECV)
3479 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_ATTEMPTFAILS);
3481 tcp_set_state(sk, TCP_CLOSE);
3482 tcp_clear_xmit_timers(sk);
3484 sk->sk_shutdown = SHUTDOWN_MASK;
3486 if (!sock_flag(sk, SOCK_DEAD))
3487 sk->sk_state_change(sk);
3489 inet_csk_destroy_sock(sk);
3491 EXPORT_SYMBOL_GPL(tcp_done);
3493 extern struct tcp_congestion_ops tcp_reno;
3495 static __initdata unsigned long thash_entries;
3496 static int __init set_thash_entries(char *str)
3500 thash_entries = simple_strtoul(str, &str, 0);
3503 __setup("thash_entries=", set_thash_entries);
3505 void tcp_init_mem(struct net *net)
3507 unsigned long limit = nr_free_buffer_pages() / 8;
3508 limit = max(limit, 128UL);
3509 net->ipv4.sysctl_tcp_mem[0] = limit / 4 * 3;
3510 net->ipv4.sysctl_tcp_mem[1] = limit;
3511 net->ipv4.sysctl_tcp_mem[2] = net->ipv4.sysctl_tcp_mem[0] * 2;
3514 void __init tcp_init(void)
3516 struct sk_buff *skb = NULL;
3517 unsigned long limit;
3520 unsigned long jiffy = jiffies;
3522 BUILD_BUG_ON(sizeof(struct tcp_skb_cb) > sizeof(skb->cb));
3524 percpu_counter_init(&tcp_sockets_allocated, 0);
3525 percpu_counter_init(&tcp_orphan_count, 0);
3526 tcp_hashinfo.bind_bucket_cachep =
3527 kmem_cache_create("tcp_bind_bucket",
3528 sizeof(struct inet_bind_bucket), 0,
3529 SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL);
3531 /* Size and allocate the main established and bind bucket
3534 * The methodology is similar to that of the buffer cache.
3536 tcp_hashinfo.ehash =
3537 alloc_large_system_hash("TCP established",
3538 sizeof(struct inet_ehash_bucket),
3540 (totalram_pages >= 128 * 1024) ?
3544 &tcp_hashinfo.ehash_mask,
3545 thash_entries ? 0 : 512 * 1024);
3546 for (i = 0; i <= tcp_hashinfo.ehash_mask; i++) {
3547 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].chain, i);
3548 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].twchain, i);
3550 if (inet_ehash_locks_alloc(&tcp_hashinfo))
3551 panic("TCP: failed to alloc ehash_locks");
3552 tcp_hashinfo.bhash =
3553 alloc_large_system_hash("TCP bind",
3554 sizeof(struct inet_bind_hashbucket),
3555 tcp_hashinfo.ehash_mask + 1,
3556 (totalram_pages >= 128 * 1024) ?
3559 &tcp_hashinfo.bhash_size,
3562 tcp_hashinfo.bhash_size = 1U << tcp_hashinfo.bhash_size;
3563 for (i = 0; i < tcp_hashinfo.bhash_size; i++) {
3564 spin_lock_init(&tcp_hashinfo.bhash[i].lock);
3565 INIT_HLIST_HEAD(&tcp_hashinfo.bhash[i].chain);
3569 cnt = tcp_hashinfo.ehash_mask + 1;
3571 tcp_death_row.sysctl_max_tw_buckets = cnt / 2;
3572 sysctl_tcp_max_orphans = cnt / 2;
3573 sysctl_max_syn_backlog = max(128, cnt / 256);
3575 tcp_init_mem(&init_net);
3576 /* Set per-socket limits to no more than 1/128 the pressure threshold */
3577 limit = nr_free_buffer_pages() << (PAGE_SHIFT - 7);
3578 max_share = min(4UL*1024*1024, limit);
3580 sysctl_tcp_wmem[0] = SK_MEM_QUANTUM;
3581 sysctl_tcp_wmem[1] = 16*1024;
3582 sysctl_tcp_wmem[2] = max(64*1024, max_share);
3584 sysctl_tcp_rmem[0] = SK_MEM_QUANTUM;
3585 sysctl_tcp_rmem[1] = 87380;
3586 sysctl_tcp_rmem[2] = max(87380, max_share);
3588 pr_info("Hash tables configured (established %u bind %u)\n",
3589 tcp_hashinfo.ehash_mask + 1, tcp_hashinfo.bhash_size);
3591 tcp_register_congestion_control(&tcp_reno);
3593 memset(&tcp_secret_one.secrets[0], 0, sizeof(tcp_secret_one.secrets));
3594 memset(&tcp_secret_two.secrets[0], 0, sizeof(tcp_secret_two.secrets));
3595 tcp_secret_one.expires = jiffy; /* past due */
3596 tcp_secret_two.expires = jiffy; /* past due */
3597 tcp_secret_generating = &tcp_secret_one;
3598 tcp_secret_primary = &tcp_secret_one;
3599 tcp_secret_retiring = &tcp_secret_two;
3600 tcp_secret_secondary = &tcp_secret_two;