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 #include <linux/kernel.h>
249 #include <linux/module.h>
250 #include <linux/types.h>
251 #include <linux/fcntl.h>
252 #include <linux/poll.h>
253 #include <linux/init.h>
254 #include <linux/fs.h>
255 #include <linux/skbuff.h>
256 #include <linux/scatterlist.h>
257 #include <linux/splice.h>
258 #include <linux/net.h>
259 #include <linux/socket.h>
260 #include <linux/random.h>
261 #include <linux/bootmem.h>
262 #include <linux/highmem.h>
263 #include <linux/swap.h>
264 #include <linux/cache.h>
265 #include <linux/err.h>
266 #include <linux/crypto.h>
267 #include <linux/time.h>
268 #include <linux/slab.h>
269 #include <linux/uid_stat.h>
271 #include <net/icmp.h>
273 #include <net/xfrm.h>
275 #include <net/ip6_route.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_mem[3] __read_mostly;
288 int sysctl_tcp_wmem[3] __read_mostly;
289 int sysctl_tcp_rmem[3] __read_mostly;
291 EXPORT_SYMBOL(sysctl_tcp_mem);
292 EXPORT_SYMBOL(sysctl_tcp_rmem);
293 EXPORT_SYMBOL(sysctl_tcp_wmem);
295 atomic_t tcp_memory_allocated; /* Current allocated memory. */
296 EXPORT_SYMBOL(tcp_memory_allocated);
299 * Current number of TCP sockets.
301 struct percpu_counter tcp_sockets_allocated;
302 EXPORT_SYMBOL(tcp_sockets_allocated);
307 struct tcp_splice_state {
308 struct pipe_inode_info *pipe;
314 * Pressure flag: try to collapse.
315 * Technical note: it is used by multiple contexts non atomically.
316 * All the __sk_mem_schedule() is of this nature: accounting
317 * is strict, actions are advisory and have some latency.
319 int tcp_memory_pressure __read_mostly;
320 EXPORT_SYMBOL(tcp_memory_pressure);
322 void tcp_enter_memory_pressure(struct sock *sk)
324 if (!tcp_memory_pressure) {
325 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMEMORYPRESSURES);
326 tcp_memory_pressure = 1;
329 EXPORT_SYMBOL(tcp_enter_memory_pressure);
331 /* Convert seconds to retransmits based on initial and max timeout */
332 static u8 secs_to_retrans(int seconds, int timeout, int rto_max)
337 int period = timeout;
340 while (seconds > period && res < 255) {
343 if (timeout > rto_max)
351 /* Convert retransmits to seconds based on initial and max timeout */
352 static int retrans_to_secs(u8 retrans, int timeout, int rto_max)
360 if (timeout > rto_max)
369 * Wait for a TCP event.
371 * Note that we don't need to lock the socket, as the upper poll layers
372 * take care of normal races (between the test and the event) and we don't
373 * go look at any of the socket buffers directly.
375 unsigned int tcp_poll(struct file *file, struct socket *sock, poll_table *wait)
378 struct sock *sk = sock->sk;
379 struct tcp_sock *tp = tcp_sk(sk);
381 sock_poll_wait(file, sk_sleep(sk), wait);
382 if (sk->sk_state == TCP_LISTEN)
383 return inet_csk_listen_poll(sk);
385 /* Socket is not locked. We are protected from async events
386 * by poll logic and correct handling of state changes
387 * made by other threads is impossible in any case.
393 * POLLHUP is certainly not done right. But poll() doesn't
394 * have a notion of HUP in just one direction, and for a
395 * socket the read side is more interesting.
397 * Some poll() documentation says that POLLHUP is incompatible
398 * with the POLLOUT/POLLWR flags, so somebody should check this
399 * all. But careful, it tends to be safer to return too many
400 * bits than too few, and you can easily break real applications
401 * if you don't tell them that something has hung up!
405 * Check number 1. POLLHUP is _UNMASKABLE_ event (see UNIX98 and
406 * our fs/select.c). It means that after we received EOF,
407 * poll always returns immediately, making impossible poll() on write()
408 * in state CLOSE_WAIT. One solution is evident --- to set POLLHUP
409 * if and only if shutdown has been made in both directions.
410 * Actually, it is interesting to look how Solaris and DUX
411 * solve this dilemma. I would prefer, if POLLHUP were maskable,
412 * then we could set it on SND_SHUTDOWN. BTW examples given
413 * in Stevens' books assume exactly this behaviour, it explains
414 * why POLLHUP is incompatible with POLLOUT. --ANK
416 * NOTE. Check for TCP_CLOSE is added. The goal is to prevent
417 * blocking on fresh not-connected or disconnected socket. --ANK
419 if (sk->sk_shutdown == SHUTDOWN_MASK || sk->sk_state == TCP_CLOSE)
421 if (sk->sk_shutdown & RCV_SHUTDOWN)
422 mask |= POLLIN | POLLRDNORM | POLLRDHUP;
425 if ((1 << sk->sk_state) & ~(TCPF_SYN_SENT | TCPF_SYN_RECV)) {
426 int target = sock_rcvlowat(sk, 0, INT_MAX);
428 if (tp->urg_seq == tp->copied_seq &&
429 !sock_flag(sk, SOCK_URGINLINE) &&
433 /* Potential race condition. If read of tp below will
434 * escape above sk->sk_state, we can be illegally awaken
435 * in SYN_* states. */
436 if (tp->rcv_nxt - tp->copied_seq >= target)
437 mask |= POLLIN | POLLRDNORM;
439 if (!(sk->sk_shutdown & SEND_SHUTDOWN)) {
440 if (sk_stream_wspace(sk) >= sk_stream_min_wspace(sk)) {
441 mask |= POLLOUT | POLLWRNORM;
442 } else { /* send SIGIO later */
443 set_bit(SOCK_ASYNC_NOSPACE,
444 &sk->sk_socket->flags);
445 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
447 /* Race breaker. If space is freed after
448 * wspace test but before the flags are set,
449 * IO signal will be lost.
451 if (sk_stream_wspace(sk) >= sk_stream_min_wspace(sk))
452 mask |= POLLOUT | POLLWRNORM;
455 mask |= POLLOUT | POLLWRNORM;
457 if (tp->urg_data & TCP_URG_VALID)
460 /* This barrier is coupled with smp_wmb() in tcp_reset() */
467 EXPORT_SYMBOL(tcp_poll);
469 int tcp_ioctl(struct sock *sk, int cmd, unsigned long arg)
471 struct tcp_sock *tp = tcp_sk(sk);
476 if (sk->sk_state == TCP_LISTEN)
480 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
482 else if (sock_flag(sk, SOCK_URGINLINE) ||
484 before(tp->urg_seq, tp->copied_seq) ||
485 !before(tp->urg_seq, tp->rcv_nxt)) {
488 answ = tp->rcv_nxt - tp->copied_seq;
490 /* Subtract 1, if FIN is in queue. */
491 skb = skb_peek_tail(&sk->sk_receive_queue);
493 answ -= tcp_hdr(skb)->fin;
495 answ = tp->urg_seq - tp->copied_seq;
499 answ = tp->urg_data && tp->urg_seq == tp->copied_seq;
502 if (sk->sk_state == TCP_LISTEN)
505 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
508 answ = tp->write_seq - tp->snd_una;
514 return put_user(answ, (int __user *)arg);
516 EXPORT_SYMBOL(tcp_ioctl);
518 static inline void tcp_mark_push(struct tcp_sock *tp, struct sk_buff *skb)
520 TCP_SKB_CB(skb)->flags |= TCPHDR_PSH;
521 tp->pushed_seq = tp->write_seq;
524 static inline int forced_push(struct tcp_sock *tp)
526 return after(tp->write_seq, tp->pushed_seq + (tp->max_window >> 1));
529 static inline void skb_entail(struct sock *sk, struct sk_buff *skb)
531 struct tcp_sock *tp = tcp_sk(sk);
532 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
535 tcb->seq = tcb->end_seq = tp->write_seq;
536 tcb->flags = TCPHDR_ACK;
538 skb_header_release(skb);
539 tcp_add_write_queue_tail(sk, skb);
540 sk->sk_wmem_queued += skb->truesize;
541 sk_mem_charge(sk, skb->truesize);
542 if (tp->nonagle & TCP_NAGLE_PUSH)
543 tp->nonagle &= ~TCP_NAGLE_PUSH;
546 static inline void tcp_mark_urg(struct tcp_sock *tp, int flags)
549 tp->snd_up = tp->write_seq;
552 static inline void tcp_push(struct sock *sk, int flags, int mss_now,
555 if (tcp_send_head(sk)) {
556 struct tcp_sock *tp = tcp_sk(sk);
558 if (!(flags & MSG_MORE) || forced_push(tp))
559 tcp_mark_push(tp, tcp_write_queue_tail(sk));
561 tcp_mark_urg(tp, flags);
562 __tcp_push_pending_frames(sk, mss_now,
563 (flags & MSG_MORE) ? TCP_NAGLE_CORK : nonagle);
567 static int tcp_splice_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb,
568 unsigned int offset, size_t len)
570 struct tcp_splice_state *tss = rd_desc->arg.data;
573 ret = skb_splice_bits(skb, offset, tss->pipe, min(rd_desc->count, len),
576 rd_desc->count -= ret;
580 static int __tcp_splice_read(struct sock *sk, struct tcp_splice_state *tss)
582 /* Store TCP splice context information in read_descriptor_t. */
583 read_descriptor_t rd_desc = {
588 return tcp_read_sock(sk, &rd_desc, tcp_splice_data_recv);
592 * tcp_splice_read - splice data from TCP socket to a pipe
593 * @sock: socket to splice from
594 * @ppos: position (not valid)
595 * @pipe: pipe to splice to
596 * @len: number of bytes to splice
597 * @flags: splice modifier flags
600 * Will read pages from given socket and fill them into a pipe.
603 ssize_t tcp_splice_read(struct socket *sock, loff_t *ppos,
604 struct pipe_inode_info *pipe, size_t len,
607 struct sock *sk = sock->sk;
608 struct tcp_splice_state tss = {
617 sock_rps_record_flow(sk);
619 * We can't seek on a socket input
628 timeo = sock_rcvtimeo(sk, sock->file->f_flags & O_NONBLOCK);
630 ret = __tcp_splice_read(sk, &tss);
636 if (sock_flag(sk, SOCK_DONE))
639 ret = sock_error(sk);
642 if (sk->sk_shutdown & RCV_SHUTDOWN)
644 if (sk->sk_state == TCP_CLOSE) {
646 * This occurs when user tries to read
647 * from never connected socket.
649 if (!sock_flag(sk, SOCK_DONE))
657 sk_wait_data(sk, &timeo);
658 if (signal_pending(current)) {
659 ret = sock_intr_errno(timeo);
672 if (sk->sk_err || sk->sk_state == TCP_CLOSE ||
673 (sk->sk_shutdown & RCV_SHUTDOWN) ||
674 signal_pending(current))
685 EXPORT_SYMBOL(tcp_splice_read);
687 struct sk_buff *sk_stream_alloc_skb(struct sock *sk, int size, gfp_t gfp)
691 /* The TCP header must be at least 32-bit aligned. */
692 size = ALIGN(size, 4);
694 skb = alloc_skb_fclone(size + sk->sk_prot->max_header, gfp);
696 if (sk_wmem_schedule(sk, skb->truesize)) {
698 * Make sure that we have exactly size bytes
699 * available to the caller, no more, no less.
701 skb_reserve(skb, skb_tailroom(skb) - size);
706 sk->sk_prot->enter_memory_pressure(sk);
707 sk_stream_moderate_sndbuf(sk);
712 static unsigned int tcp_xmit_size_goal(struct sock *sk, u32 mss_now,
715 struct tcp_sock *tp = tcp_sk(sk);
716 u32 xmit_size_goal, old_size_goal;
718 xmit_size_goal = mss_now;
720 if (large_allowed && sk_can_gso(sk)) {
721 xmit_size_goal = ((sk->sk_gso_max_size - 1) -
722 inet_csk(sk)->icsk_af_ops->net_header_len -
723 inet_csk(sk)->icsk_ext_hdr_len -
726 xmit_size_goal = tcp_bound_to_half_wnd(tp, xmit_size_goal);
728 /* We try hard to avoid divides here */
729 old_size_goal = tp->xmit_size_goal_segs * mss_now;
731 if (likely(old_size_goal <= xmit_size_goal &&
732 old_size_goal + mss_now > xmit_size_goal)) {
733 xmit_size_goal = old_size_goal;
735 tp->xmit_size_goal_segs = xmit_size_goal / mss_now;
736 xmit_size_goal = tp->xmit_size_goal_segs * mss_now;
740 return max(xmit_size_goal, mss_now);
743 static int tcp_send_mss(struct sock *sk, int *size_goal, int flags)
747 mss_now = tcp_current_mss(sk);
748 *size_goal = tcp_xmit_size_goal(sk, mss_now, !(flags & MSG_OOB));
753 static ssize_t do_tcp_sendpages(struct sock *sk, struct page **pages, int poffset,
754 size_t psize, int flags)
756 struct tcp_sock *tp = tcp_sk(sk);
757 int mss_now, size_goal;
760 long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
762 /* Wait for a connection to finish. */
763 if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
764 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
767 clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
769 mss_now = tcp_send_mss(sk, &size_goal, flags);
773 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
777 struct sk_buff *skb = tcp_write_queue_tail(sk);
778 struct page *page = pages[poffset / PAGE_SIZE];
779 int copy, i, can_coalesce;
780 int offset = poffset % PAGE_SIZE;
781 int size = min_t(size_t, psize, PAGE_SIZE - offset);
783 if (!tcp_send_head(sk) || (copy = size_goal - skb->len) <= 0) {
785 if (!sk_stream_memory_free(sk))
786 goto wait_for_sndbuf;
788 skb = sk_stream_alloc_skb(sk, 0, sk->sk_allocation);
790 goto wait_for_memory;
799 i = skb_shinfo(skb)->nr_frags;
800 can_coalesce = skb_can_coalesce(skb, i, page, offset);
801 if (!can_coalesce && i >= MAX_SKB_FRAGS) {
802 tcp_mark_push(tp, skb);
805 if (!sk_wmem_schedule(sk, copy))
806 goto wait_for_memory;
809 skb_shinfo(skb)->frags[i - 1].size += copy;
812 skb_fill_page_desc(skb, i, page, offset, copy);
816 skb->data_len += copy;
817 skb->truesize += copy;
818 sk->sk_wmem_queued += copy;
819 sk_mem_charge(sk, copy);
820 skb->ip_summed = CHECKSUM_PARTIAL;
821 tp->write_seq += copy;
822 TCP_SKB_CB(skb)->end_seq += copy;
823 skb_shinfo(skb)->gso_segs = 0;
826 TCP_SKB_CB(skb)->flags &= ~TCPHDR_PSH;
830 if (!(psize -= copy))
833 if (skb->len < size_goal || (flags & MSG_OOB))
836 if (forced_push(tp)) {
837 tcp_mark_push(tp, skb);
838 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
839 } else if (skb == tcp_send_head(sk))
840 tcp_push_one(sk, mss_now);
844 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
847 tcp_push(sk, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH);
849 if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
852 mss_now = tcp_send_mss(sk, &size_goal, flags);
857 tcp_push(sk, flags, mss_now, tp->nonagle);
864 return sk_stream_error(sk, flags, err);
867 int tcp_sendpage(struct sock *sk, struct page *page, int offset,
868 size_t size, int flags)
872 if (!(sk->sk_route_caps & NETIF_F_SG) ||
873 !(sk->sk_route_caps & NETIF_F_ALL_CSUM))
874 return sock_no_sendpage(sk->sk_socket, page, offset, size,
879 res = do_tcp_sendpages(sk, &page, offset, size, flags);
884 EXPORT_SYMBOL(tcp_sendpage);
886 #define TCP_PAGE(sk) (sk->sk_sndmsg_page)
887 #define TCP_OFF(sk) (sk->sk_sndmsg_off)
889 static inline int select_size(struct sock *sk, int sg)
891 struct tcp_sock *tp = tcp_sk(sk);
892 int tmp = tp->mss_cache;
898 int pgbreak = SKB_MAX_HEAD(MAX_TCP_HEADER);
900 if (tmp >= pgbreak &&
901 tmp <= pgbreak + (MAX_SKB_FRAGS - 1) * PAGE_SIZE)
909 int tcp_sendmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg,
913 struct tcp_sock *tp = tcp_sk(sk);
916 int mss_now, size_goal;
923 flags = msg->msg_flags;
924 timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
926 /* Wait for a connection to finish. */
927 if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
928 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
931 /* This should be in poll */
932 clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
934 mss_now = tcp_send_mss(sk, &size_goal, flags);
936 /* Ok commence sending. */
937 iovlen = msg->msg_iovlen;
942 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
945 sg = sk->sk_route_caps & NETIF_F_SG;
947 while (--iovlen >= 0) {
948 size_t seglen = iov->iov_len;
949 unsigned char __user *from = iov->iov_base;
957 skb = tcp_write_queue_tail(sk);
958 if (tcp_send_head(sk)) {
959 if (skb->ip_summed == CHECKSUM_NONE)
961 copy = max - skb->len;
966 /* Allocate new segment. If the interface is SG,
967 * allocate skb fitting to single page.
969 if (!sk_stream_memory_free(sk))
970 goto wait_for_sndbuf;
972 skb = sk_stream_alloc_skb(sk,
976 goto wait_for_memory;
979 * Check whether we can use HW checksum.
981 if (sk->sk_route_caps & NETIF_F_ALL_CSUM)
982 skb->ip_summed = CHECKSUM_PARTIAL;
989 /* Try to append data to the end of skb. */
993 /* Where to copy to? */
994 if (skb_tailroom(skb) > 0) {
995 /* We have some space in skb head. Superb! */
996 if (copy > skb_tailroom(skb))
997 copy = skb_tailroom(skb);
998 if ((err = skb_add_data(skb, from, copy)) != 0)
1002 int i = skb_shinfo(skb)->nr_frags;
1003 struct page *page = TCP_PAGE(sk);
1004 int off = TCP_OFF(sk);
1006 if (skb_can_coalesce(skb, i, page, off) &&
1008 /* We can extend the last page
1011 } else if (i == MAX_SKB_FRAGS || !sg) {
1012 /* Need to add new fragment and cannot
1013 * do this because interface is non-SG,
1014 * or because all the page slots are
1016 tcp_mark_push(tp, skb);
1019 if (off == PAGE_SIZE) {
1021 TCP_PAGE(sk) = page = NULL;
1027 if (copy > PAGE_SIZE - off)
1028 copy = PAGE_SIZE - off;
1030 if (!sk_wmem_schedule(sk, copy))
1031 goto wait_for_memory;
1034 /* Allocate new cache page. */
1035 if (!(page = sk_stream_alloc_page(sk)))
1036 goto wait_for_memory;
1039 /* Time to copy data. We are close to
1041 err = skb_copy_to_page(sk, from, skb, page,
1044 /* If this page was new, give it to the
1045 * socket so it does not get leaked.
1047 if (!TCP_PAGE(sk)) {
1048 TCP_PAGE(sk) = page;
1054 /* Update the skb. */
1056 skb_shinfo(skb)->frags[i - 1].size +=
1059 skb_fill_page_desc(skb, i, page, off, copy);
1062 } else if (off + copy < PAGE_SIZE) {
1064 TCP_PAGE(sk) = page;
1068 TCP_OFF(sk) = off + copy;
1072 TCP_SKB_CB(skb)->flags &= ~TCPHDR_PSH;
1074 tp->write_seq += copy;
1075 TCP_SKB_CB(skb)->end_seq += copy;
1076 skb_shinfo(skb)->gso_segs = 0;
1080 if ((seglen -= copy) == 0 && iovlen == 0)
1083 if (skb->len < max || (flags & MSG_OOB))
1086 if (forced_push(tp)) {
1087 tcp_mark_push(tp, skb);
1088 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
1089 } else if (skb == tcp_send_head(sk))
1090 tcp_push_one(sk, mss_now);
1094 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1097 tcp_push(sk, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH);
1099 if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
1102 mss_now = tcp_send_mss(sk, &size_goal, flags);
1108 tcp_push(sk, flags, mss_now, tp->nonagle);
1109 TCP_CHECK_TIMER(sk);
1113 uid_stat_tcp_snd(current_uid(), copied);
1118 tcp_unlink_write_queue(skb, sk);
1119 /* It is the one place in all of TCP, except connection
1120 * reset, where we can be unlinking the send_head.
1122 tcp_check_send_head(sk, skb);
1123 sk_wmem_free_skb(sk, skb);
1130 err = sk_stream_error(sk, flags, err);
1131 TCP_CHECK_TIMER(sk);
1135 EXPORT_SYMBOL(tcp_sendmsg);
1138 * Handle reading urgent data. BSD has very simple semantics for
1139 * this, no blocking and very strange errors 8)
1142 static int tcp_recv_urg(struct sock *sk, struct msghdr *msg, int len, int flags)
1144 struct tcp_sock *tp = tcp_sk(sk);
1146 /* No URG data to read. */
1147 if (sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data ||
1148 tp->urg_data == TCP_URG_READ)
1149 return -EINVAL; /* Yes this is right ! */
1151 if (sk->sk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DONE))
1154 if (tp->urg_data & TCP_URG_VALID) {
1156 char c = tp->urg_data;
1158 if (!(flags & MSG_PEEK))
1159 tp->urg_data = TCP_URG_READ;
1161 /* Read urgent data. */
1162 msg->msg_flags |= MSG_OOB;
1165 if (!(flags & MSG_TRUNC))
1166 err = memcpy_toiovec(msg->msg_iov, &c, 1);
1169 msg->msg_flags |= MSG_TRUNC;
1171 return err ? -EFAULT : len;
1174 if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
1177 /* Fixed the recv(..., MSG_OOB) behaviour. BSD docs and
1178 * the available implementations agree in this case:
1179 * this call should never block, independent of the
1180 * blocking state of the socket.
1181 * Mike <pall@rz.uni-karlsruhe.de>
1186 /* Clean up the receive buffer for full frames taken by the user,
1187 * then send an ACK if necessary. COPIED is the number of bytes
1188 * tcp_recvmsg has given to the user so far, it speeds up the
1189 * calculation of whether or not we must ACK for the sake of
1192 void tcp_cleanup_rbuf(struct sock *sk, int copied)
1194 struct tcp_sock *tp = tcp_sk(sk);
1195 int time_to_ack = 0;
1198 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
1200 WARN(skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq),
1201 KERN_INFO "cleanup rbuf bug: copied %X seq %X rcvnxt %X\n",
1202 tp->copied_seq, TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt);
1205 if (inet_csk_ack_scheduled(sk)) {
1206 const struct inet_connection_sock *icsk = inet_csk(sk);
1207 /* Delayed ACKs frequently hit locked sockets during bulk
1209 if (icsk->icsk_ack.blocked ||
1210 /* Once-per-two-segments ACK was not sent by tcp_input.c */
1211 tp->rcv_nxt - tp->rcv_wup > icsk->icsk_ack.rcv_mss ||
1213 * If this read emptied read buffer, we send ACK, if
1214 * connection is not bidirectional, user drained
1215 * receive buffer and there was a small segment
1219 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED2) ||
1220 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED) &&
1221 !icsk->icsk_ack.pingpong)) &&
1222 !atomic_read(&sk->sk_rmem_alloc)))
1226 /* We send an ACK if we can now advertise a non-zero window
1227 * which has been raised "significantly".
1229 * Even if window raised up to infinity, do not send window open ACK
1230 * in states, where we will not receive more. It is useless.
1232 if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) {
1233 __u32 rcv_window_now = tcp_receive_window(tp);
1235 /* Optimize, __tcp_select_window() is not cheap. */
1236 if (2*rcv_window_now <= tp->window_clamp) {
1237 __u32 new_window = __tcp_select_window(sk);
1239 /* Send ACK now, if this read freed lots of space
1240 * in our buffer. Certainly, new_window is new window.
1241 * We can advertise it now, if it is not less than current one.
1242 * "Lots" means "at least twice" here.
1244 if (new_window && new_window >= 2 * rcv_window_now)
1252 static void tcp_prequeue_process(struct sock *sk)
1254 struct sk_buff *skb;
1255 struct tcp_sock *tp = tcp_sk(sk);
1257 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPPREQUEUED);
1259 /* RX process wants to run with disabled BHs, though it is not
1262 while ((skb = __skb_dequeue(&tp->ucopy.prequeue)) != NULL)
1263 sk_backlog_rcv(sk, skb);
1266 /* Clear memory counter. */
1267 tp->ucopy.memory = 0;
1270 #ifdef CONFIG_NET_DMA
1271 static void tcp_service_net_dma(struct sock *sk, bool wait)
1273 dma_cookie_t done, used;
1274 dma_cookie_t last_issued;
1275 struct tcp_sock *tp = tcp_sk(sk);
1277 if (!tp->ucopy.dma_chan)
1280 last_issued = tp->ucopy.dma_cookie;
1281 dma_async_memcpy_issue_pending(tp->ucopy.dma_chan);
1284 if (dma_async_memcpy_complete(tp->ucopy.dma_chan,
1286 &used) == DMA_SUCCESS) {
1287 /* Safe to free early-copied skbs now */
1288 __skb_queue_purge(&sk->sk_async_wait_queue);
1291 struct sk_buff *skb;
1292 while ((skb = skb_peek(&sk->sk_async_wait_queue)) &&
1293 (dma_async_is_complete(skb->dma_cookie, done,
1294 used) == DMA_SUCCESS)) {
1295 __skb_dequeue(&sk->sk_async_wait_queue);
1303 static inline struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off)
1305 struct sk_buff *skb;
1308 skb_queue_walk(&sk->sk_receive_queue, skb) {
1309 offset = seq - TCP_SKB_CB(skb)->seq;
1310 if (tcp_hdr(skb)->syn)
1312 if (offset < skb->len || tcp_hdr(skb)->fin) {
1321 * This routine provides an alternative to tcp_recvmsg() for routines
1322 * that would like to handle copying from skbuffs directly in 'sendfile'
1325 * - It is assumed that the socket was locked by the caller.
1326 * - The routine does not block.
1327 * - At present, there is no support for reading OOB data
1328 * or for 'peeking' the socket using this routine
1329 * (although both would be easy to implement).
1331 int tcp_read_sock(struct sock *sk, read_descriptor_t *desc,
1332 sk_read_actor_t recv_actor)
1334 struct sk_buff *skb;
1335 struct tcp_sock *tp = tcp_sk(sk);
1336 u32 seq = tp->copied_seq;
1340 if (sk->sk_state == TCP_LISTEN)
1342 while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1343 if (offset < skb->len) {
1347 len = skb->len - offset;
1348 /* Stop reading if we hit a patch of urgent data */
1350 u32 urg_offset = tp->urg_seq - seq;
1351 if (urg_offset < len)
1356 used = recv_actor(desc, skb, offset, len);
1361 } else if (used <= len) {
1367 * If recv_actor drops the lock (e.g. TCP splice
1368 * receive) the skb pointer might be invalid when
1369 * getting here: tcp_collapse might have deleted it
1370 * while aggregating skbs from the socket queue.
1372 skb = tcp_recv_skb(sk, seq-1, &offset);
1373 if (!skb || (offset+1 != skb->len))
1376 if (tcp_hdr(skb)->fin) {
1377 sk_eat_skb(sk, skb, 0);
1381 sk_eat_skb(sk, skb, 0);
1384 tp->copied_seq = seq;
1386 tp->copied_seq = seq;
1388 tcp_rcv_space_adjust(sk);
1390 /* Clean up data we have read: This will do ACK frames. */
1392 tcp_cleanup_rbuf(sk, copied);
1393 uid_stat_tcp_rcv(current_uid(), copied);
1398 EXPORT_SYMBOL(tcp_read_sock);
1401 * This routine copies from a sock struct into the user buffer.
1403 * Technical note: in 2.3 we work on _locked_ socket, so that
1404 * tricks with *seq access order and skb->users are not required.
1405 * Probably, code can be easily improved even more.
1408 int tcp_recvmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg,
1409 size_t len, int nonblock, int flags, int *addr_len)
1411 struct tcp_sock *tp = tcp_sk(sk);
1417 int target; /* Read at least this many bytes */
1419 struct task_struct *user_recv = NULL;
1420 int copied_early = 0;
1421 struct sk_buff *skb;
1426 TCP_CHECK_TIMER(sk);
1429 if (sk->sk_state == TCP_LISTEN)
1432 timeo = sock_rcvtimeo(sk, nonblock);
1434 /* Urgent data needs to be handled specially. */
1435 if (flags & MSG_OOB)
1438 seq = &tp->copied_seq;
1439 if (flags & MSG_PEEK) {
1440 peek_seq = tp->copied_seq;
1444 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
1446 #ifdef CONFIG_NET_DMA
1447 tp->ucopy.dma_chan = NULL;
1449 skb = skb_peek_tail(&sk->sk_receive_queue);
1454 available = TCP_SKB_CB(skb)->seq + skb->len - (*seq);
1455 if ((available < target) &&
1456 (len > sysctl_tcp_dma_copybreak) && !(flags & MSG_PEEK) &&
1457 !sysctl_tcp_low_latency &&
1458 dma_find_channel(DMA_MEMCPY)) {
1459 preempt_enable_no_resched();
1460 tp->ucopy.pinned_list =
1461 dma_pin_iovec_pages(msg->msg_iov, len);
1463 preempt_enable_no_resched();
1471 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
1472 if (tp->urg_data && tp->urg_seq == *seq) {
1475 if (signal_pending(current)) {
1476 copied = timeo ? sock_intr_errno(timeo) : -EAGAIN;
1481 /* Next get a buffer. */
1483 skb_queue_walk(&sk->sk_receive_queue, skb) {
1484 /* Now that we have two receive queues this
1487 if (WARN(before(*seq, TCP_SKB_CB(skb)->seq),
1488 KERN_INFO "recvmsg bug: copied %X "
1489 "seq %X rcvnxt %X fl %X\n", *seq,
1490 TCP_SKB_CB(skb)->seq, tp->rcv_nxt,
1494 offset = *seq - TCP_SKB_CB(skb)->seq;
1495 if (tcp_hdr(skb)->syn)
1497 if (offset < skb->len)
1499 if (tcp_hdr(skb)->fin)
1501 WARN(!(flags & MSG_PEEK), KERN_INFO "recvmsg bug 2: "
1502 "copied %X seq %X rcvnxt %X fl %X\n",
1503 *seq, TCP_SKB_CB(skb)->seq,
1504 tp->rcv_nxt, flags);
1507 /* Well, if we have backlog, try to process it now yet. */
1509 if (copied >= target && !sk->sk_backlog.tail)
1514 sk->sk_state == TCP_CLOSE ||
1515 (sk->sk_shutdown & RCV_SHUTDOWN) ||
1517 signal_pending(current))
1520 if (sock_flag(sk, SOCK_DONE))
1524 copied = sock_error(sk);
1528 if (sk->sk_shutdown & RCV_SHUTDOWN)
1531 if (sk->sk_state == TCP_CLOSE) {
1532 if (!sock_flag(sk, SOCK_DONE)) {
1533 /* This occurs when user tries to read
1534 * from never connected socket.
1547 if (signal_pending(current)) {
1548 copied = sock_intr_errno(timeo);
1553 tcp_cleanup_rbuf(sk, copied);
1555 if (!sysctl_tcp_low_latency && tp->ucopy.task == user_recv) {
1556 /* Install new reader */
1557 if (!user_recv && !(flags & (MSG_TRUNC | MSG_PEEK))) {
1558 user_recv = current;
1559 tp->ucopy.task = user_recv;
1560 tp->ucopy.iov = msg->msg_iov;
1563 tp->ucopy.len = len;
1565 WARN_ON(tp->copied_seq != tp->rcv_nxt &&
1566 !(flags & (MSG_PEEK | MSG_TRUNC)));
1568 /* Ugly... If prequeue is not empty, we have to
1569 * process it before releasing socket, otherwise
1570 * order will be broken at second iteration.
1571 * More elegant solution is required!!!
1573 * Look: we have the following (pseudo)queues:
1575 * 1. packets in flight
1580 * Each queue can be processed only if the next ones
1581 * are empty. At this point we have empty receive_queue.
1582 * But prequeue _can_ be not empty after 2nd iteration,
1583 * when we jumped to start of loop because backlog
1584 * processing added something to receive_queue.
1585 * We cannot release_sock(), because backlog contains
1586 * packets arrived _after_ prequeued ones.
1588 * Shortly, algorithm is clear --- to process all
1589 * the queues in order. We could make it more directly,
1590 * requeueing packets from backlog to prequeue, if
1591 * is not empty. It is more elegant, but eats cycles,
1594 if (!skb_queue_empty(&tp->ucopy.prequeue))
1597 /* __ Set realtime policy in scheduler __ */
1600 #ifdef CONFIG_NET_DMA
1601 if (tp->ucopy.dma_chan)
1602 dma_async_memcpy_issue_pending(tp->ucopy.dma_chan);
1604 if (copied >= target) {
1605 /* Do not sleep, just process backlog. */
1609 sk_wait_data(sk, &timeo);
1611 #ifdef CONFIG_NET_DMA
1612 tcp_service_net_dma(sk, false); /* Don't block */
1613 tp->ucopy.wakeup = 0;
1619 /* __ Restore normal policy in scheduler __ */
1621 if ((chunk = len - tp->ucopy.len) != 0) {
1622 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMBACKLOG, chunk);
1627 if (tp->rcv_nxt == tp->copied_seq &&
1628 !skb_queue_empty(&tp->ucopy.prequeue)) {
1630 tcp_prequeue_process(sk);
1632 if ((chunk = len - tp->ucopy.len) != 0) {
1633 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1639 if ((flags & MSG_PEEK) &&
1640 (peek_seq - copied - urg_hole != tp->copied_seq)) {
1641 if (net_ratelimit())
1642 printk(KERN_DEBUG "TCP(%s:%d): Application bug, race in MSG_PEEK.\n",
1643 current->comm, task_pid_nr(current));
1644 peek_seq = tp->copied_seq;
1649 /* Ok so how much can we use? */
1650 used = skb->len - offset;
1654 /* Do we have urgent data here? */
1656 u32 urg_offset = tp->urg_seq - *seq;
1657 if (urg_offset < used) {
1659 if (!sock_flag(sk, SOCK_URGINLINE)) {
1672 if (!(flags & MSG_TRUNC)) {
1673 #ifdef CONFIG_NET_DMA
1674 if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list)
1675 tp->ucopy.dma_chan = dma_find_channel(DMA_MEMCPY);
1677 if (tp->ucopy.dma_chan) {
1678 tp->ucopy.dma_cookie = dma_skb_copy_datagram_iovec(
1679 tp->ucopy.dma_chan, skb, offset,
1681 tp->ucopy.pinned_list);
1683 if (tp->ucopy.dma_cookie < 0) {
1685 printk(KERN_ALERT "dma_cookie < 0\n");
1687 /* Exception. Bailout! */
1693 dma_async_memcpy_issue_pending(tp->ucopy.dma_chan);
1695 if ((offset + used) == skb->len)
1701 err = skb_copy_datagram_iovec(skb, offset,
1702 msg->msg_iov, used);
1704 /* Exception. Bailout! */
1716 tcp_rcv_space_adjust(sk);
1719 if (tp->urg_data && after(tp->copied_seq, tp->urg_seq)) {
1721 tcp_fast_path_check(sk);
1723 if (used + offset < skb->len)
1726 if (tcp_hdr(skb)->fin)
1728 if (!(flags & MSG_PEEK)) {
1729 sk_eat_skb(sk, skb, copied_early);
1735 /* Process the FIN. */
1737 if (!(flags & MSG_PEEK)) {
1738 sk_eat_skb(sk, skb, copied_early);
1745 if (!skb_queue_empty(&tp->ucopy.prequeue)) {
1748 tp->ucopy.len = copied > 0 ? len : 0;
1750 tcp_prequeue_process(sk);
1752 if (copied > 0 && (chunk = len - tp->ucopy.len) != 0) {
1753 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1759 tp->ucopy.task = NULL;
1763 #ifdef CONFIG_NET_DMA
1764 tcp_service_net_dma(sk, true); /* Wait for queue to drain */
1765 tp->ucopy.dma_chan = NULL;
1767 if (tp->ucopy.pinned_list) {
1768 dma_unpin_iovec_pages(tp->ucopy.pinned_list);
1769 tp->ucopy.pinned_list = NULL;
1773 /* According to UNIX98, msg_name/msg_namelen are ignored
1774 * on connected socket. I was just happy when found this 8) --ANK
1777 /* Clean up data we have read: This will do ACK frames. */
1778 tcp_cleanup_rbuf(sk, copied);
1780 TCP_CHECK_TIMER(sk);
1784 uid_stat_tcp_rcv(current_uid(), copied);
1788 TCP_CHECK_TIMER(sk);
1793 err = tcp_recv_urg(sk, msg, len, flags);
1795 uid_stat_tcp_rcv(current_uid(), err);
1798 EXPORT_SYMBOL(tcp_recvmsg);
1800 void tcp_set_state(struct sock *sk, int state)
1802 int oldstate = sk->sk_state;
1805 case TCP_ESTABLISHED:
1806 if (oldstate != TCP_ESTABLISHED)
1807 TCP_INC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
1811 if (oldstate == TCP_CLOSE_WAIT || oldstate == TCP_ESTABLISHED)
1812 TCP_INC_STATS(sock_net(sk), TCP_MIB_ESTABRESETS);
1814 sk->sk_prot->unhash(sk);
1815 if (inet_csk(sk)->icsk_bind_hash &&
1816 !(sk->sk_userlocks & SOCK_BINDPORT_LOCK))
1820 if (oldstate == TCP_ESTABLISHED)
1821 TCP_DEC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
1824 /* Change state AFTER socket is unhashed to avoid closed
1825 * socket sitting in hash tables.
1827 sk->sk_state = state;
1830 SOCK_DEBUG(sk, "TCP sk=%p, State %s -> %s\n", sk, statename[oldstate], statename[state]);
1833 EXPORT_SYMBOL_GPL(tcp_set_state);
1836 * State processing on a close. This implements the state shift for
1837 * sending our FIN frame. Note that we only send a FIN for some
1838 * states. A shutdown() may have already sent the FIN, or we may be
1842 static const unsigned char new_state[16] = {
1843 /* current state: new state: action: */
1844 /* (Invalid) */ TCP_CLOSE,
1845 /* TCP_ESTABLISHED */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1846 /* TCP_SYN_SENT */ TCP_CLOSE,
1847 /* TCP_SYN_RECV */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1848 /* TCP_FIN_WAIT1 */ TCP_FIN_WAIT1,
1849 /* TCP_FIN_WAIT2 */ TCP_FIN_WAIT2,
1850 /* TCP_TIME_WAIT */ TCP_CLOSE,
1851 /* TCP_CLOSE */ TCP_CLOSE,
1852 /* TCP_CLOSE_WAIT */ TCP_LAST_ACK | TCP_ACTION_FIN,
1853 /* TCP_LAST_ACK */ TCP_LAST_ACK,
1854 /* TCP_LISTEN */ TCP_CLOSE,
1855 /* TCP_CLOSING */ TCP_CLOSING,
1858 static int tcp_close_state(struct sock *sk)
1860 int next = (int)new_state[sk->sk_state];
1861 int ns = next & TCP_STATE_MASK;
1863 tcp_set_state(sk, ns);
1865 return next & TCP_ACTION_FIN;
1869 * Shutdown the sending side of a connection. Much like close except
1870 * that we don't receive shut down or sock_set_flag(sk, SOCK_DEAD).
1873 void tcp_shutdown(struct sock *sk, int how)
1875 /* We need to grab some memory, and put together a FIN,
1876 * and then put it into the queue to be sent.
1877 * Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
1879 if (!(how & SEND_SHUTDOWN))
1882 /* If we've already sent a FIN, or it's a closed state, skip this. */
1883 if ((1 << sk->sk_state) &
1884 (TCPF_ESTABLISHED | TCPF_SYN_SENT |
1885 TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) {
1886 /* Clear out any half completed packets. FIN if needed. */
1887 if (tcp_close_state(sk))
1891 EXPORT_SYMBOL(tcp_shutdown);
1893 void tcp_close(struct sock *sk, long timeout)
1895 struct sk_buff *skb;
1896 int data_was_unread = 0;
1900 sk->sk_shutdown = SHUTDOWN_MASK;
1902 if (sk->sk_state == TCP_LISTEN) {
1903 tcp_set_state(sk, TCP_CLOSE);
1906 inet_csk_listen_stop(sk);
1908 goto adjudge_to_death;
1911 /* We need to flush the recv. buffs. We do this only on the
1912 * descriptor close, not protocol-sourced closes, because the
1913 * reader process may not have drained the data yet!
1915 while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) {
1916 u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq -
1918 data_was_unread += len;
1924 /* If socket has been already reset (e.g. in tcp_reset()) - kill it. */
1925 if (sk->sk_state == TCP_CLOSE)
1926 goto adjudge_to_death;
1928 /* As outlined in RFC 2525, section 2.17, we send a RST here because
1929 * data was lost. To witness the awful effects of the old behavior of
1930 * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk
1931 * GET in an FTP client, suspend the process, wait for the client to
1932 * advertise a zero window, then kill -9 the FTP client, wheee...
1933 * Note: timeout is always zero in such a case.
1935 if (data_was_unread) {
1936 /* Unread data was tossed, zap the connection. */
1937 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONCLOSE);
1938 tcp_set_state(sk, TCP_CLOSE);
1939 tcp_send_active_reset(sk, sk->sk_allocation);
1940 } else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) {
1941 /* Check zero linger _after_ checking for unread data. */
1942 sk->sk_prot->disconnect(sk, 0);
1943 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
1944 } else if (tcp_close_state(sk)) {
1945 /* We FIN if the application ate all the data before
1946 * zapping the connection.
1949 /* RED-PEN. Formally speaking, we have broken TCP state
1950 * machine. State transitions:
1952 * TCP_ESTABLISHED -> TCP_FIN_WAIT1
1953 * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible)
1954 * TCP_CLOSE_WAIT -> TCP_LAST_ACK
1956 * are legal only when FIN has been sent (i.e. in window),
1957 * rather than queued out of window. Purists blame.
1959 * F.e. "RFC state" is ESTABLISHED,
1960 * if Linux state is FIN-WAIT-1, but FIN is still not sent.
1962 * The visible declinations are that sometimes
1963 * we enter time-wait state, when it is not required really
1964 * (harmless), do not send active resets, when they are
1965 * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
1966 * they look as CLOSING or LAST_ACK for Linux)
1967 * Probably, I missed some more holelets.
1973 sk_stream_wait_close(sk, timeout);
1976 state = sk->sk_state;
1980 /* It is the last release_sock in its life. It will remove backlog. */
1984 /* Now socket is owned by kernel and we acquire BH lock
1985 to finish close. No need to check for user refs.
1989 WARN_ON(sock_owned_by_user(sk));
1991 percpu_counter_inc(sk->sk_prot->orphan_count);
1993 /* Have we already been destroyed by a softirq or backlog? */
1994 if (state != TCP_CLOSE && sk->sk_state == TCP_CLOSE)
1997 /* This is a (useful) BSD violating of the RFC. There is a
1998 * problem with TCP as specified in that the other end could
1999 * keep a socket open forever with no application left this end.
2000 * We use a 3 minute timeout (about the same as BSD) then kill
2001 * our end. If they send after that then tough - BUT: long enough
2002 * that we won't make the old 4*rto = almost no time - whoops
2005 * Nope, it was not mistake. It is really desired behaviour
2006 * f.e. on http servers, when such sockets are useless, but
2007 * consume significant resources. Let's do it with special
2008 * linger2 option. --ANK
2011 if (sk->sk_state == TCP_FIN_WAIT2) {
2012 struct tcp_sock *tp = tcp_sk(sk);
2013 if (tp->linger2 < 0) {
2014 tcp_set_state(sk, TCP_CLOSE);
2015 tcp_send_active_reset(sk, GFP_ATOMIC);
2016 NET_INC_STATS_BH(sock_net(sk),
2017 LINUX_MIB_TCPABORTONLINGER);
2019 const int tmo = tcp_fin_time(sk);
2021 if (tmo > TCP_TIMEWAIT_LEN) {
2022 inet_csk_reset_keepalive_timer(sk,
2023 tmo - TCP_TIMEWAIT_LEN);
2025 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
2030 if (sk->sk_state != TCP_CLOSE) {
2032 if (tcp_too_many_orphans(sk, 0)) {
2033 if (net_ratelimit())
2034 printk(KERN_INFO "TCP: too many of orphaned "
2036 tcp_set_state(sk, TCP_CLOSE);
2037 tcp_send_active_reset(sk, GFP_ATOMIC);
2038 NET_INC_STATS_BH(sock_net(sk),
2039 LINUX_MIB_TCPABORTONMEMORY);
2043 if (sk->sk_state == TCP_CLOSE)
2044 inet_csk_destroy_sock(sk);
2045 /* Otherwise, socket is reprieved until protocol close. */
2052 EXPORT_SYMBOL(tcp_close);
2054 /* These states need RST on ABORT according to RFC793 */
2056 static inline int tcp_need_reset(int state)
2058 return (1 << state) &
2059 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 |
2060 TCPF_FIN_WAIT2 | TCPF_SYN_RECV);
2063 int tcp_disconnect(struct sock *sk, int flags)
2065 struct inet_sock *inet = inet_sk(sk);
2066 struct inet_connection_sock *icsk = inet_csk(sk);
2067 struct tcp_sock *tp = tcp_sk(sk);
2069 int old_state = sk->sk_state;
2071 if (old_state != TCP_CLOSE)
2072 tcp_set_state(sk, TCP_CLOSE);
2074 /* ABORT function of RFC793 */
2075 if (old_state == TCP_LISTEN) {
2076 inet_csk_listen_stop(sk);
2077 } else if (tcp_need_reset(old_state) ||
2078 (tp->snd_nxt != tp->write_seq &&
2079 (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK))) {
2080 /* The last check adjusts for discrepancy of Linux wrt. RFC
2083 tcp_send_active_reset(sk, gfp_any());
2084 sk->sk_err = ECONNRESET;
2085 } else if (old_state == TCP_SYN_SENT)
2086 sk->sk_err = ECONNRESET;
2088 tcp_clear_xmit_timers(sk);
2089 __skb_queue_purge(&sk->sk_receive_queue);
2090 tcp_write_queue_purge(sk);
2091 __skb_queue_purge(&tp->out_of_order_queue);
2092 #ifdef CONFIG_NET_DMA
2093 __skb_queue_purge(&sk->sk_async_wait_queue);
2096 inet->inet_dport = 0;
2098 if (!(sk->sk_userlocks & SOCK_BINDADDR_LOCK))
2099 inet_reset_saddr(sk);
2101 sk->sk_shutdown = 0;
2102 sock_reset_flag(sk, SOCK_DONE);
2104 if ((tp->write_seq += tp->max_window + 2) == 0)
2106 icsk->icsk_backoff = 0;
2108 icsk->icsk_probes_out = 0;
2109 tp->packets_out = 0;
2110 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
2111 tp->snd_cwnd_cnt = 0;
2112 tp->bytes_acked = 0;
2113 tp->window_clamp = 0;
2114 tcp_set_ca_state(sk, TCP_CA_Open);
2115 tcp_clear_retrans(tp);
2116 inet_csk_delack_init(sk);
2117 tcp_init_send_head(sk);
2118 memset(&tp->rx_opt, 0, sizeof(tp->rx_opt));
2121 WARN_ON(inet->inet_num && !icsk->icsk_bind_hash);
2123 sk->sk_error_report(sk);
2126 EXPORT_SYMBOL(tcp_disconnect);
2129 * Socket option code for TCP.
2131 static int do_tcp_setsockopt(struct sock *sk, int level,
2132 int optname, char __user *optval, unsigned int optlen)
2134 struct tcp_sock *tp = tcp_sk(sk);
2135 struct inet_connection_sock *icsk = inet_csk(sk);
2139 /* These are data/string values, all the others are ints */
2141 case TCP_CONGESTION: {
2142 char name[TCP_CA_NAME_MAX];
2147 val = strncpy_from_user(name, optval,
2148 min_t(long, TCP_CA_NAME_MAX-1, optlen));
2154 err = tcp_set_congestion_control(sk, name);
2158 case TCP_COOKIE_TRANSACTIONS: {
2159 struct tcp_cookie_transactions ctd;
2160 struct tcp_cookie_values *cvp = NULL;
2162 if (sizeof(ctd) > optlen)
2164 if (copy_from_user(&ctd, optval, sizeof(ctd)))
2167 if (ctd.tcpct_used > sizeof(ctd.tcpct_value) ||
2168 ctd.tcpct_s_data_desired > TCP_MSS_DESIRED)
2171 if (ctd.tcpct_cookie_desired == 0) {
2172 /* default to global value */
2173 } else if ((0x1 & ctd.tcpct_cookie_desired) ||
2174 ctd.tcpct_cookie_desired > TCP_COOKIE_MAX ||
2175 ctd.tcpct_cookie_desired < TCP_COOKIE_MIN) {
2179 if (TCP_COOKIE_OUT_NEVER & ctd.tcpct_flags) {
2180 /* Supercedes all other values */
2182 if (tp->cookie_values != NULL) {
2183 kref_put(&tp->cookie_values->kref,
2184 tcp_cookie_values_release);
2185 tp->cookie_values = NULL;
2187 tp->rx_opt.cookie_in_always = 0; /* false */
2188 tp->rx_opt.cookie_out_never = 1; /* true */
2193 /* Allocate ancillary memory before locking.
2195 if (ctd.tcpct_used > 0 ||
2196 (tp->cookie_values == NULL &&
2197 (sysctl_tcp_cookie_size > 0 ||
2198 ctd.tcpct_cookie_desired > 0 ||
2199 ctd.tcpct_s_data_desired > 0))) {
2200 cvp = kzalloc(sizeof(*cvp) + ctd.tcpct_used,
2205 kref_init(&cvp->kref);
2208 tp->rx_opt.cookie_in_always =
2209 (TCP_COOKIE_IN_ALWAYS & ctd.tcpct_flags);
2210 tp->rx_opt.cookie_out_never = 0; /* false */
2212 if (tp->cookie_values != NULL) {
2214 /* Changed values are recorded by a changed
2215 * pointer, ensuring the cookie will differ,
2216 * without separately hashing each value later.
2218 kref_put(&tp->cookie_values->kref,
2219 tcp_cookie_values_release);
2221 cvp = tp->cookie_values;
2226 cvp->cookie_desired = ctd.tcpct_cookie_desired;
2228 if (ctd.tcpct_used > 0) {
2229 memcpy(cvp->s_data_payload, ctd.tcpct_value,
2231 cvp->s_data_desired = ctd.tcpct_used;
2232 cvp->s_data_constant = 1; /* true */
2234 /* No constant payload data. */
2235 cvp->s_data_desired = ctd.tcpct_s_data_desired;
2236 cvp->s_data_constant = 0; /* false */
2239 tp->cookie_values = cvp;
2249 if (optlen < sizeof(int))
2252 if (get_user(val, (int __user *)optval))
2259 /* Values greater than interface MTU won't take effect. However
2260 * at the point when this call is done we typically don't yet
2261 * know which interface is going to be used */
2262 if (val < 8 || val > MAX_TCP_WINDOW) {
2266 tp->rx_opt.user_mss = val;
2271 /* TCP_NODELAY is weaker than TCP_CORK, so that
2272 * this option on corked socket is remembered, but
2273 * it is not activated until cork is cleared.
2275 * However, when TCP_NODELAY is set we make
2276 * an explicit push, which overrides even TCP_CORK
2277 * for currently queued segments.
2279 tp->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH;
2280 tcp_push_pending_frames(sk);
2282 tp->nonagle &= ~TCP_NAGLE_OFF;
2286 case TCP_THIN_LINEAR_TIMEOUTS:
2287 if (val < 0 || val > 1)
2293 case TCP_THIN_DUPACK:
2294 if (val < 0 || val > 1)
2297 tp->thin_dupack = val;
2301 /* When set indicates to always queue non-full frames.
2302 * Later the user clears this option and we transmit
2303 * any pending partial frames in the queue. This is
2304 * meant to be used alongside sendfile() to get properly
2305 * filled frames when the user (for example) must write
2306 * out headers with a write() call first and then use
2307 * sendfile to send out the data parts.
2309 * TCP_CORK can be set together with TCP_NODELAY and it is
2310 * stronger than TCP_NODELAY.
2313 tp->nonagle |= TCP_NAGLE_CORK;
2315 tp->nonagle &= ~TCP_NAGLE_CORK;
2316 if (tp->nonagle&TCP_NAGLE_OFF)
2317 tp->nonagle |= TCP_NAGLE_PUSH;
2318 tcp_push_pending_frames(sk);
2323 if (val < 1 || val > MAX_TCP_KEEPIDLE)
2326 tp->keepalive_time = val * HZ;
2327 if (sock_flag(sk, SOCK_KEEPOPEN) &&
2328 !((1 << sk->sk_state) &
2329 (TCPF_CLOSE | TCPF_LISTEN))) {
2330 u32 elapsed = keepalive_time_elapsed(tp);
2331 if (tp->keepalive_time > elapsed)
2332 elapsed = tp->keepalive_time - elapsed;
2335 inet_csk_reset_keepalive_timer(sk, elapsed);
2340 if (val < 1 || val > MAX_TCP_KEEPINTVL)
2343 tp->keepalive_intvl = val * HZ;
2346 if (val < 1 || val > MAX_TCP_KEEPCNT)
2349 tp->keepalive_probes = val;
2352 if (val < 1 || val > MAX_TCP_SYNCNT)
2355 icsk->icsk_syn_retries = val;
2361 else if (val > sysctl_tcp_fin_timeout / HZ)
2364 tp->linger2 = val * HZ;
2367 case TCP_DEFER_ACCEPT:
2368 /* Translate value in seconds to number of retransmits */
2369 icsk->icsk_accept_queue.rskq_defer_accept =
2370 secs_to_retrans(val, TCP_TIMEOUT_INIT / HZ,
2374 case TCP_WINDOW_CLAMP:
2376 if (sk->sk_state != TCP_CLOSE) {
2380 tp->window_clamp = 0;
2382 tp->window_clamp = val < SOCK_MIN_RCVBUF / 2 ?
2383 SOCK_MIN_RCVBUF / 2 : val;
2388 icsk->icsk_ack.pingpong = 1;
2390 icsk->icsk_ack.pingpong = 0;
2391 if ((1 << sk->sk_state) &
2392 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) &&
2393 inet_csk_ack_scheduled(sk)) {
2394 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
2395 tcp_cleanup_rbuf(sk, 1);
2397 icsk->icsk_ack.pingpong = 1;
2402 #ifdef CONFIG_TCP_MD5SIG
2404 /* Read the IP->Key mappings from userspace */
2405 err = tp->af_specific->md5_parse(sk, optval, optlen);
2418 int tcp_setsockopt(struct sock *sk, int level, int optname, char __user *optval,
2419 unsigned int optlen)
2421 struct inet_connection_sock *icsk = inet_csk(sk);
2423 if (level != SOL_TCP)
2424 return icsk->icsk_af_ops->setsockopt(sk, level, optname,
2426 return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2428 EXPORT_SYMBOL(tcp_setsockopt);
2430 #ifdef CONFIG_COMPAT
2431 int compat_tcp_setsockopt(struct sock *sk, int level, int optname,
2432 char __user *optval, unsigned int optlen)
2434 if (level != SOL_TCP)
2435 return inet_csk_compat_setsockopt(sk, level, optname,
2437 return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2439 EXPORT_SYMBOL(compat_tcp_setsockopt);
2442 /* Return information about state of tcp endpoint in API format. */
2443 void tcp_get_info(struct sock *sk, struct tcp_info *info)
2445 struct tcp_sock *tp = tcp_sk(sk);
2446 const struct inet_connection_sock *icsk = inet_csk(sk);
2447 u32 now = tcp_time_stamp;
2449 memset(info, 0, sizeof(*info));
2451 info->tcpi_state = sk->sk_state;
2452 info->tcpi_ca_state = icsk->icsk_ca_state;
2453 info->tcpi_retransmits = icsk->icsk_retransmits;
2454 info->tcpi_probes = icsk->icsk_probes_out;
2455 info->tcpi_backoff = icsk->icsk_backoff;
2457 if (tp->rx_opt.tstamp_ok)
2458 info->tcpi_options |= TCPI_OPT_TIMESTAMPS;
2459 if (tcp_is_sack(tp))
2460 info->tcpi_options |= TCPI_OPT_SACK;
2461 if (tp->rx_opt.wscale_ok) {
2462 info->tcpi_options |= TCPI_OPT_WSCALE;
2463 info->tcpi_snd_wscale = tp->rx_opt.snd_wscale;
2464 info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale;
2467 if (tp->ecn_flags&TCP_ECN_OK)
2468 info->tcpi_options |= TCPI_OPT_ECN;
2470 info->tcpi_rto = jiffies_to_usecs(icsk->icsk_rto);
2471 info->tcpi_ato = jiffies_to_usecs(icsk->icsk_ack.ato);
2472 info->tcpi_snd_mss = tp->mss_cache;
2473 info->tcpi_rcv_mss = icsk->icsk_ack.rcv_mss;
2475 if (sk->sk_state == TCP_LISTEN) {
2476 info->tcpi_unacked = sk->sk_ack_backlog;
2477 info->tcpi_sacked = sk->sk_max_ack_backlog;
2479 info->tcpi_unacked = tp->packets_out;
2480 info->tcpi_sacked = tp->sacked_out;
2482 info->tcpi_lost = tp->lost_out;
2483 info->tcpi_retrans = tp->retrans_out;
2484 info->tcpi_fackets = tp->fackets_out;
2486 info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime);
2487 info->tcpi_last_data_recv = jiffies_to_msecs(now - icsk->icsk_ack.lrcvtime);
2488 info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp);
2490 info->tcpi_pmtu = icsk->icsk_pmtu_cookie;
2491 info->tcpi_rcv_ssthresh = tp->rcv_ssthresh;
2492 info->tcpi_rtt = jiffies_to_usecs(tp->srtt)>>3;
2493 info->tcpi_rttvar = jiffies_to_usecs(tp->mdev)>>2;
2494 info->tcpi_snd_ssthresh = tp->snd_ssthresh;
2495 info->tcpi_snd_cwnd = tp->snd_cwnd;
2496 info->tcpi_advmss = tp->advmss;
2497 info->tcpi_reordering = tp->reordering;
2499 info->tcpi_rcv_rtt = jiffies_to_usecs(tp->rcv_rtt_est.rtt)>>3;
2500 info->tcpi_rcv_space = tp->rcvq_space.space;
2502 info->tcpi_total_retrans = tp->total_retrans;
2504 EXPORT_SYMBOL_GPL(tcp_get_info);
2506 static int do_tcp_getsockopt(struct sock *sk, int level,
2507 int optname, char __user *optval, int __user *optlen)
2509 struct inet_connection_sock *icsk = inet_csk(sk);
2510 struct tcp_sock *tp = tcp_sk(sk);
2513 if (get_user(len, optlen))
2516 len = min_t(unsigned int, len, sizeof(int));
2523 val = tp->mss_cache;
2524 if (!val && ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
2525 val = tp->rx_opt.user_mss;
2528 val = !!(tp->nonagle&TCP_NAGLE_OFF);
2531 val = !!(tp->nonagle&TCP_NAGLE_CORK);
2534 val = keepalive_time_when(tp) / HZ;
2537 val = keepalive_intvl_when(tp) / HZ;
2540 val = keepalive_probes(tp);
2543 val = icsk->icsk_syn_retries ? : sysctl_tcp_syn_retries;
2548 val = (val ? : sysctl_tcp_fin_timeout) / HZ;
2550 case TCP_DEFER_ACCEPT:
2551 val = retrans_to_secs(icsk->icsk_accept_queue.rskq_defer_accept,
2552 TCP_TIMEOUT_INIT / HZ, TCP_RTO_MAX / HZ);
2554 case TCP_WINDOW_CLAMP:
2555 val = tp->window_clamp;
2558 struct tcp_info info;
2560 if (get_user(len, optlen))
2563 tcp_get_info(sk, &info);
2565 len = min_t(unsigned int, len, sizeof(info));
2566 if (put_user(len, optlen))
2568 if (copy_to_user(optval, &info, len))
2573 val = !icsk->icsk_ack.pingpong;
2576 case TCP_CONGESTION:
2577 if (get_user(len, optlen))
2579 len = min_t(unsigned int, len, TCP_CA_NAME_MAX);
2580 if (put_user(len, optlen))
2582 if (copy_to_user(optval, icsk->icsk_ca_ops->name, len))
2586 case TCP_COOKIE_TRANSACTIONS: {
2587 struct tcp_cookie_transactions ctd;
2588 struct tcp_cookie_values *cvp = tp->cookie_values;
2590 if (get_user(len, optlen))
2592 if (len < sizeof(ctd))
2595 memset(&ctd, 0, sizeof(ctd));
2596 ctd.tcpct_flags = (tp->rx_opt.cookie_in_always ?
2597 TCP_COOKIE_IN_ALWAYS : 0)
2598 | (tp->rx_opt.cookie_out_never ?
2599 TCP_COOKIE_OUT_NEVER : 0);
2602 ctd.tcpct_flags |= (cvp->s_data_in ?
2604 | (cvp->s_data_out ?
2605 TCP_S_DATA_OUT : 0);
2607 ctd.tcpct_cookie_desired = cvp->cookie_desired;
2608 ctd.tcpct_s_data_desired = cvp->s_data_desired;
2610 memcpy(&ctd.tcpct_value[0], &cvp->cookie_pair[0],
2611 cvp->cookie_pair_size);
2612 ctd.tcpct_used = cvp->cookie_pair_size;
2615 if (put_user(sizeof(ctd), optlen))
2617 if (copy_to_user(optval, &ctd, sizeof(ctd)))
2621 case TCP_THIN_LINEAR_TIMEOUTS:
2624 case TCP_THIN_DUPACK:
2625 val = tp->thin_dupack;
2628 return -ENOPROTOOPT;
2631 if (put_user(len, optlen))
2633 if (copy_to_user(optval, &val, len))
2638 int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval,
2641 struct inet_connection_sock *icsk = inet_csk(sk);
2643 if (level != SOL_TCP)
2644 return icsk->icsk_af_ops->getsockopt(sk, level, optname,
2646 return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2648 EXPORT_SYMBOL(tcp_getsockopt);
2650 #ifdef CONFIG_COMPAT
2651 int compat_tcp_getsockopt(struct sock *sk, int level, int optname,
2652 char __user *optval, int __user *optlen)
2654 if (level != SOL_TCP)
2655 return inet_csk_compat_getsockopt(sk, level, optname,
2657 return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2659 EXPORT_SYMBOL(compat_tcp_getsockopt);
2662 struct sk_buff *tcp_tso_segment(struct sk_buff *skb, int features)
2664 struct sk_buff *segs = ERR_PTR(-EINVAL);
2669 unsigned int oldlen;
2672 if (!pskb_may_pull(skb, sizeof(*th)))
2676 thlen = th->doff * 4;
2677 if (thlen < sizeof(*th))
2680 if (!pskb_may_pull(skb, thlen))
2683 oldlen = (u16)~skb->len;
2684 __skb_pull(skb, thlen);
2686 mss = skb_shinfo(skb)->gso_size;
2687 if (unlikely(skb->len <= mss))
2690 if (skb_gso_ok(skb, features | NETIF_F_GSO_ROBUST)) {
2691 /* Packet is from an untrusted source, reset gso_segs. */
2692 int type = skb_shinfo(skb)->gso_type;
2700 !(type & (SKB_GSO_TCPV4 | SKB_GSO_TCPV6))))
2703 skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss);
2709 segs = skb_segment(skb, features);
2713 delta = htonl(oldlen + (thlen + mss));
2717 seq = ntohl(th->seq);
2720 th->fin = th->psh = 0;
2722 th->check = ~csum_fold((__force __wsum)((__force u32)th->check +
2723 (__force u32)delta));
2724 if (skb->ip_summed != CHECKSUM_PARTIAL)
2726 csum_fold(csum_partial(skb_transport_header(skb),
2733 th->seq = htonl(seq);
2735 } while (skb->next);
2737 delta = htonl(oldlen + (skb->tail - skb->transport_header) +
2739 th->check = ~csum_fold((__force __wsum)((__force u32)th->check +
2740 (__force u32)delta));
2741 if (skb->ip_summed != CHECKSUM_PARTIAL)
2742 th->check = csum_fold(csum_partial(skb_transport_header(skb),
2748 EXPORT_SYMBOL(tcp_tso_segment);
2750 struct sk_buff **tcp_gro_receive(struct sk_buff **head, struct sk_buff *skb)
2752 struct sk_buff **pp = NULL;
2759 unsigned int mss = 1;
2765 off = skb_gro_offset(skb);
2766 hlen = off + sizeof(*th);
2767 th = skb_gro_header_fast(skb, off);
2768 if (skb_gro_header_hard(skb, hlen)) {
2769 th = skb_gro_header_slow(skb, hlen, off);
2774 thlen = th->doff * 4;
2775 if (thlen < sizeof(*th))
2779 if (skb_gro_header_hard(skb, hlen)) {
2780 th = skb_gro_header_slow(skb, hlen, off);
2785 skb_gro_pull(skb, thlen);
2787 len = skb_gro_len(skb);
2788 flags = tcp_flag_word(th);
2790 for (; (p = *head); head = &p->next) {
2791 if (!NAPI_GRO_CB(p)->same_flow)
2796 if (*(u32 *)&th->source ^ *(u32 *)&th2->source) {
2797 NAPI_GRO_CB(p)->same_flow = 0;
2804 goto out_check_final;
2807 flush = NAPI_GRO_CB(p)->flush;
2808 flush |= (__force int)(flags & TCP_FLAG_CWR);
2809 flush |= (__force int)((flags ^ tcp_flag_word(th2)) &
2810 ~(TCP_FLAG_CWR | TCP_FLAG_FIN | TCP_FLAG_PSH));
2811 flush |= (__force int)(th->ack_seq ^ th2->ack_seq);
2812 for (i = sizeof(*th); i < thlen; i += 4)
2813 flush |= *(u32 *)((u8 *)th + i) ^
2814 *(u32 *)((u8 *)th2 + i);
2816 mss = skb_shinfo(p)->gso_size;
2818 flush |= (len - 1) >= mss;
2819 flush |= (ntohl(th2->seq) + skb_gro_len(p)) ^ ntohl(th->seq);
2821 if (flush || skb_gro_receive(head, skb)) {
2823 goto out_check_final;
2828 tcp_flag_word(th2) |= flags & (TCP_FLAG_FIN | TCP_FLAG_PSH);
2832 flush |= (__force int)(flags & (TCP_FLAG_URG | TCP_FLAG_PSH |
2833 TCP_FLAG_RST | TCP_FLAG_SYN |
2836 if (p && (!NAPI_GRO_CB(skb)->same_flow || flush))
2840 NAPI_GRO_CB(skb)->flush |= flush;
2844 EXPORT_SYMBOL(tcp_gro_receive);
2846 int tcp_gro_complete(struct sk_buff *skb)
2848 struct tcphdr *th = tcp_hdr(skb);
2850 skb->csum_start = skb_transport_header(skb) - skb->head;
2851 skb->csum_offset = offsetof(struct tcphdr, check);
2852 skb->ip_summed = CHECKSUM_PARTIAL;
2854 skb_shinfo(skb)->gso_segs = NAPI_GRO_CB(skb)->count;
2857 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
2861 EXPORT_SYMBOL(tcp_gro_complete);
2863 #ifdef CONFIG_TCP_MD5SIG
2864 static unsigned long tcp_md5sig_users;
2865 static struct tcp_md5sig_pool * __percpu *tcp_md5sig_pool;
2866 static DEFINE_SPINLOCK(tcp_md5sig_pool_lock);
2868 static void __tcp_free_md5sig_pool(struct tcp_md5sig_pool * __percpu *pool)
2871 for_each_possible_cpu(cpu) {
2872 struct tcp_md5sig_pool *p = *per_cpu_ptr(pool, cpu);
2874 if (p->md5_desc.tfm)
2875 crypto_free_hash(p->md5_desc.tfm);
2882 void tcp_free_md5sig_pool(void)
2884 struct tcp_md5sig_pool * __percpu *pool = NULL;
2886 spin_lock_bh(&tcp_md5sig_pool_lock);
2887 if (--tcp_md5sig_users == 0) {
2888 pool = tcp_md5sig_pool;
2889 tcp_md5sig_pool = NULL;
2891 spin_unlock_bh(&tcp_md5sig_pool_lock);
2893 __tcp_free_md5sig_pool(pool);
2895 EXPORT_SYMBOL(tcp_free_md5sig_pool);
2897 static struct tcp_md5sig_pool * __percpu *
2898 __tcp_alloc_md5sig_pool(struct sock *sk)
2901 struct tcp_md5sig_pool * __percpu *pool;
2903 pool = alloc_percpu(struct tcp_md5sig_pool *);
2907 for_each_possible_cpu(cpu) {
2908 struct tcp_md5sig_pool *p;
2909 struct crypto_hash *hash;
2911 p = kzalloc(sizeof(*p), sk->sk_allocation);
2914 *per_cpu_ptr(pool, cpu) = p;
2916 hash = crypto_alloc_hash("md5", 0, CRYPTO_ALG_ASYNC);
2917 if (!hash || IS_ERR(hash))
2920 p->md5_desc.tfm = hash;
2924 __tcp_free_md5sig_pool(pool);
2928 struct tcp_md5sig_pool * __percpu *tcp_alloc_md5sig_pool(struct sock *sk)
2930 struct tcp_md5sig_pool * __percpu *pool;
2934 spin_lock_bh(&tcp_md5sig_pool_lock);
2935 pool = tcp_md5sig_pool;
2936 if (tcp_md5sig_users++ == 0) {
2938 spin_unlock_bh(&tcp_md5sig_pool_lock);
2941 spin_unlock_bh(&tcp_md5sig_pool_lock);
2945 spin_unlock_bh(&tcp_md5sig_pool_lock);
2948 /* we cannot hold spinlock here because this may sleep. */
2949 struct tcp_md5sig_pool * __percpu *p;
2951 p = __tcp_alloc_md5sig_pool(sk);
2952 spin_lock_bh(&tcp_md5sig_pool_lock);
2955 spin_unlock_bh(&tcp_md5sig_pool_lock);
2958 pool = tcp_md5sig_pool;
2960 /* oops, it has already been assigned. */
2961 spin_unlock_bh(&tcp_md5sig_pool_lock);
2962 __tcp_free_md5sig_pool(p);
2964 tcp_md5sig_pool = pool = p;
2965 spin_unlock_bh(&tcp_md5sig_pool_lock);
2970 EXPORT_SYMBOL(tcp_alloc_md5sig_pool);
2974 * tcp_get_md5sig_pool - get md5sig_pool for this user
2976 * We use percpu structure, so if we succeed, we exit with preemption
2977 * and BH disabled, to make sure another thread or softirq handling
2978 * wont try to get same context.
2980 struct tcp_md5sig_pool *tcp_get_md5sig_pool(void)
2982 struct tcp_md5sig_pool * __percpu *p;
2986 spin_lock(&tcp_md5sig_pool_lock);
2987 p = tcp_md5sig_pool;
2990 spin_unlock(&tcp_md5sig_pool_lock);
2993 return *this_cpu_ptr(p);
2998 EXPORT_SYMBOL(tcp_get_md5sig_pool);
3000 void tcp_put_md5sig_pool(void)
3003 tcp_free_md5sig_pool();
3005 EXPORT_SYMBOL(tcp_put_md5sig_pool);
3007 int tcp_md5_hash_header(struct tcp_md5sig_pool *hp,
3010 struct scatterlist sg;
3013 __sum16 old_checksum = th->check;
3015 /* options aren't included in the hash */
3016 sg_init_one(&sg, th, sizeof(struct tcphdr));
3017 err = crypto_hash_update(&hp->md5_desc, &sg, sizeof(struct tcphdr));
3018 th->check = old_checksum;
3021 EXPORT_SYMBOL(tcp_md5_hash_header);
3023 int tcp_md5_hash_skb_data(struct tcp_md5sig_pool *hp,
3024 struct sk_buff *skb, unsigned header_len)
3026 struct scatterlist sg;
3027 const struct tcphdr *tp = tcp_hdr(skb);
3028 struct hash_desc *desc = &hp->md5_desc;
3030 const unsigned head_data_len = skb_headlen(skb) > header_len ?
3031 skb_headlen(skb) - header_len : 0;
3032 const struct skb_shared_info *shi = skb_shinfo(skb);
3033 struct sk_buff *frag_iter;
3035 sg_init_table(&sg, 1);
3037 sg_set_buf(&sg, ((u8 *) tp) + header_len, head_data_len);
3038 if (crypto_hash_update(desc, &sg, head_data_len))
3041 for (i = 0; i < shi->nr_frags; ++i) {
3042 const struct skb_frag_struct *f = &shi->frags[i];
3043 sg_set_page(&sg, f->page, f->size, f->page_offset);
3044 if (crypto_hash_update(desc, &sg, f->size))
3048 skb_walk_frags(skb, frag_iter)
3049 if (tcp_md5_hash_skb_data(hp, frag_iter, 0))
3054 EXPORT_SYMBOL(tcp_md5_hash_skb_data);
3056 int tcp_md5_hash_key(struct tcp_md5sig_pool *hp, struct tcp_md5sig_key *key)
3058 struct scatterlist sg;
3060 sg_init_one(&sg, key->key, key->keylen);
3061 return crypto_hash_update(&hp->md5_desc, &sg, key->keylen);
3063 EXPORT_SYMBOL(tcp_md5_hash_key);
3068 * Each Responder maintains up to two secret values concurrently for
3069 * efficient secret rollover. Each secret value has 4 states:
3071 * Generating. (tcp_secret_generating != tcp_secret_primary)
3072 * Generates new Responder-Cookies, but not yet used for primary
3073 * verification. This is a short-term state, typically lasting only
3074 * one round trip time (RTT).
3076 * Primary. (tcp_secret_generating == tcp_secret_primary)
3077 * Used both for generation and primary verification.
3079 * Retiring. (tcp_secret_retiring != tcp_secret_secondary)
3080 * Used for verification, until the first failure that can be
3081 * verified by the newer Generating secret. At that time, this
3082 * cookie's state is changed to Secondary, and the Generating
3083 * cookie's state is changed to Primary. This is a short-term state,
3084 * typically lasting only one round trip time (RTT).
3086 * Secondary. (tcp_secret_retiring == tcp_secret_secondary)
3087 * Used for secondary verification, after primary verification
3088 * failures. This state lasts no more than twice the Maximum Segment
3089 * Lifetime (2MSL). Then, the secret is discarded.
3091 struct tcp_cookie_secret {
3092 /* The secret is divided into two parts. The digest part is the
3093 * equivalent of previously hashing a secret and saving the state,
3094 * and serves as an initialization vector (IV). The message part
3095 * serves as the trailing secret.
3097 u32 secrets[COOKIE_WORKSPACE_WORDS];
3098 unsigned long expires;
3101 #define TCP_SECRET_1MSL (HZ * TCP_PAWS_MSL)
3102 #define TCP_SECRET_2MSL (HZ * TCP_PAWS_MSL * 2)
3103 #define TCP_SECRET_LIFE (HZ * 600)
3105 static struct tcp_cookie_secret tcp_secret_one;
3106 static struct tcp_cookie_secret tcp_secret_two;
3108 /* Essentially a circular list, without dynamic allocation. */
3109 static struct tcp_cookie_secret *tcp_secret_generating;
3110 static struct tcp_cookie_secret *tcp_secret_primary;
3111 static struct tcp_cookie_secret *tcp_secret_retiring;
3112 static struct tcp_cookie_secret *tcp_secret_secondary;
3114 static DEFINE_SPINLOCK(tcp_secret_locker);
3116 /* Select a pseudo-random word in the cookie workspace.
3118 static inline u32 tcp_cookie_work(const u32 *ws, const int n)
3120 return ws[COOKIE_DIGEST_WORDS + ((COOKIE_MESSAGE_WORDS-1) & ws[n])];
3123 /* Fill bakery[COOKIE_WORKSPACE_WORDS] with generator, updating as needed.
3124 * Called in softirq context.
3125 * Returns: 0 for success.
3127 int tcp_cookie_generator(u32 *bakery)
3129 unsigned long jiffy = jiffies;
3131 if (unlikely(time_after_eq(jiffy, tcp_secret_generating->expires))) {
3132 spin_lock_bh(&tcp_secret_locker);
3133 if (!time_after_eq(jiffy, tcp_secret_generating->expires)) {
3134 /* refreshed by another */
3136 &tcp_secret_generating->secrets[0],
3137 COOKIE_WORKSPACE_WORDS);
3139 /* still needs refreshing */
3140 get_random_bytes(bakery, COOKIE_WORKSPACE_WORDS);
3142 /* The first time, paranoia assumes that the
3143 * randomization function isn't as strong. But,
3144 * this secret initialization is delayed until
3145 * the last possible moment (packet arrival).
3146 * Although that time is observable, it is
3147 * unpredictably variable. Mash in the most
3148 * volatile clock bits available, and expire the
3149 * secret extra quickly.
3151 if (unlikely(tcp_secret_primary->expires ==
3152 tcp_secret_secondary->expires)) {
3155 getnstimeofday(&tv);
3156 bakery[COOKIE_DIGEST_WORDS+0] ^=
3159 tcp_secret_secondary->expires = jiffy
3161 + (0x0f & tcp_cookie_work(bakery, 0));
3163 tcp_secret_secondary->expires = jiffy
3165 + (0xff & tcp_cookie_work(bakery, 1));
3166 tcp_secret_primary->expires = jiffy
3168 + (0x1f & tcp_cookie_work(bakery, 2));
3170 memcpy(&tcp_secret_secondary->secrets[0],
3171 bakery, COOKIE_WORKSPACE_WORDS);
3173 rcu_assign_pointer(tcp_secret_generating,
3174 tcp_secret_secondary);
3175 rcu_assign_pointer(tcp_secret_retiring,
3176 tcp_secret_primary);
3178 * Neither call_rcu() nor synchronize_rcu() needed.
3179 * Retiring data is not freed. It is replaced after
3180 * further (locked) pointer updates, and a quiet time
3181 * (minimum 1MSL, maximum LIFE - 2MSL).
3184 spin_unlock_bh(&tcp_secret_locker);
3188 &rcu_dereference(tcp_secret_generating)->secrets[0],
3189 COOKIE_WORKSPACE_WORDS);
3190 rcu_read_unlock_bh();
3194 EXPORT_SYMBOL(tcp_cookie_generator);
3196 void tcp_done(struct sock *sk)
3198 if (sk->sk_state == TCP_SYN_SENT || sk->sk_state == TCP_SYN_RECV)
3199 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_ATTEMPTFAILS);
3201 tcp_set_state(sk, TCP_CLOSE);
3202 tcp_clear_xmit_timers(sk);
3204 sk->sk_shutdown = SHUTDOWN_MASK;
3206 if (!sock_flag(sk, SOCK_DEAD))
3207 sk->sk_state_change(sk);
3209 inet_csk_destroy_sock(sk);
3211 EXPORT_SYMBOL_GPL(tcp_done);
3213 extern struct tcp_congestion_ops tcp_reno;
3215 static __initdata unsigned long thash_entries;
3216 static int __init set_thash_entries(char *str)
3220 thash_entries = simple_strtoul(str, &str, 0);
3223 __setup("thash_entries=", set_thash_entries);
3225 void __init tcp_init(void)
3227 struct sk_buff *skb = NULL;
3228 unsigned long nr_pages, limit;
3229 int i, max_share, cnt;
3230 unsigned long jiffy = jiffies;
3232 BUILD_BUG_ON(sizeof(struct tcp_skb_cb) > sizeof(skb->cb));
3234 percpu_counter_init(&tcp_sockets_allocated, 0);
3235 percpu_counter_init(&tcp_orphan_count, 0);
3236 tcp_hashinfo.bind_bucket_cachep =
3237 kmem_cache_create("tcp_bind_bucket",
3238 sizeof(struct inet_bind_bucket), 0,
3239 SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL);
3241 /* Size and allocate the main established and bind bucket
3244 * The methodology is similar to that of the buffer cache.
3246 tcp_hashinfo.ehash =
3247 alloc_large_system_hash("TCP established",
3248 sizeof(struct inet_ehash_bucket),
3250 (totalram_pages >= 128 * 1024) ?
3254 &tcp_hashinfo.ehash_mask,
3255 thash_entries ? 0 : 512 * 1024);
3256 for (i = 0; i <= tcp_hashinfo.ehash_mask; i++) {
3257 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].chain, i);
3258 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].twchain, i);
3260 if (inet_ehash_locks_alloc(&tcp_hashinfo))
3261 panic("TCP: failed to alloc ehash_locks");
3262 tcp_hashinfo.bhash =
3263 alloc_large_system_hash("TCP bind",
3264 sizeof(struct inet_bind_hashbucket),
3265 tcp_hashinfo.ehash_mask + 1,
3266 (totalram_pages >= 128 * 1024) ?
3269 &tcp_hashinfo.bhash_size,
3272 tcp_hashinfo.bhash_size = 1 << tcp_hashinfo.bhash_size;
3273 for (i = 0; i < tcp_hashinfo.bhash_size; i++) {
3274 spin_lock_init(&tcp_hashinfo.bhash[i].lock);
3275 INIT_HLIST_HEAD(&tcp_hashinfo.bhash[i].chain);
3279 cnt = tcp_hashinfo.ehash_mask + 1;
3281 tcp_death_row.sysctl_max_tw_buckets = cnt / 2;
3282 sysctl_tcp_max_orphans = cnt / 2;
3283 sysctl_max_syn_backlog = max(128, cnt / 256);
3285 /* Set the pressure threshold to be a fraction of global memory that
3286 * is up to 1/2 at 256 MB, decreasing toward zero with the amount of
3287 * memory, with a floor of 128 pages.
3289 nr_pages = totalram_pages - totalhigh_pages;
3290 limit = min(nr_pages, 1UL<<(28-PAGE_SHIFT)) >> (20-PAGE_SHIFT);
3291 limit = (limit * (nr_pages >> (20-PAGE_SHIFT))) >> (PAGE_SHIFT-11);
3292 limit = max(limit, 128UL);
3293 sysctl_tcp_mem[0] = limit / 4 * 3;
3294 sysctl_tcp_mem[1] = limit;
3295 sysctl_tcp_mem[2] = sysctl_tcp_mem[0] * 2;
3297 /* Set per-socket limits to no more than 1/128 the pressure threshold */
3298 limit = ((unsigned long)sysctl_tcp_mem[1]) << (PAGE_SHIFT - 7);
3299 max_share = min(4UL*1024*1024, limit);
3301 sysctl_tcp_wmem[0] = SK_MEM_QUANTUM;
3302 sysctl_tcp_wmem[1] = 16*1024;
3303 sysctl_tcp_wmem[2] = max(64*1024, max_share);
3305 sysctl_tcp_rmem[0] = SK_MEM_QUANTUM;
3306 sysctl_tcp_rmem[1] = 87380;
3307 sysctl_tcp_rmem[2] = max(87380, max_share);
3309 printk(KERN_INFO "TCP: Hash tables configured "
3310 "(established %u bind %u)\n",
3311 tcp_hashinfo.ehash_mask + 1, tcp_hashinfo.bhash_size);
3313 tcp_register_congestion_control(&tcp_reno);
3315 memset(&tcp_secret_one.secrets[0], 0, sizeof(tcp_secret_one.secrets));
3316 memset(&tcp_secret_two.secrets[0], 0, sizeof(tcp_secret_two.secrets));
3317 tcp_secret_one.expires = jiffy; /* past due */
3318 tcp_secret_two.expires = jiffy; /* past due */
3319 tcp_secret_generating = &tcp_secret_one;
3320 tcp_secret_primary = &tcp_secret_one;
3321 tcp_secret_retiring = &tcp_secret_two;
3322 tcp_secret_secondary = &tcp_secret_two;
3325 static int tcp_is_local(struct net *net, __be32 addr) {
3327 struct flowi fl = { .nl_u = { .ip4_u = { .daddr = addr } } };
3328 if (ip_route_output_key(net, &rt, &fl) || !rt)
3330 return rt->dst.dev && (rt->dst.dev->flags & IFF_LOOPBACK);
3333 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
3334 static int tcp_is_local6(struct net *net, struct in6_addr *addr) {
3335 struct rt6_info *rt6 = rt6_lookup(net, addr, addr, 0, 0);
3336 return rt6 && rt6->rt6i_dev && (rt6->rt6i_dev->flags & IFF_LOOPBACK);
3341 * tcp_nuke_addr - destroy all sockets on the given local address
3342 * if local address is the unspecified address (0.0.0.0 or ::), destroy all
3343 * sockets with local addresses that are not configured.
3345 int tcp_nuke_addr(struct net *net, struct sockaddr *addr)
3347 int family = addr->sa_family;
3348 unsigned int bucket;
3351 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
3352 struct in6_addr *in6;
3354 if (family == AF_INET) {
3355 in = &((struct sockaddr_in *)addr)->sin_addr;
3356 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
3357 } else if (family == AF_INET6) {
3358 in6 = &((struct sockaddr_in6 *)addr)->sin6_addr;
3361 return -EAFNOSUPPORT;
3364 for (bucket = 0; bucket < tcp_hashinfo.ehash_mask; bucket++) {
3365 struct hlist_nulls_node *node;
3367 spinlock_t *lock = inet_ehash_lockp(&tcp_hashinfo, bucket);
3371 sk_nulls_for_each(sk, node, &tcp_hashinfo.ehash[bucket].chain) {
3372 struct inet_sock *inet = inet_sk(sk);
3374 if (family == AF_INET) {
3375 __be32 s4 = inet->inet_rcv_saddr;
3376 if (in->s_addr != s4 &&
3377 !(in->s_addr == INADDR_ANY &&
3378 !tcp_is_local(net, s4)))
3382 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
3383 if (family == AF_INET6) {
3384 struct in6_addr *s6;
3387 s6 = &inet->pinet6->rcv_saddr;
3388 if (!ipv6_addr_equal(in6, s6) &&
3389 !(ipv6_addr_equal(in6, &in6addr_any) &&
3390 !tcp_is_local6(net, s6)))
3395 if (sysctl_ip_dynaddr && sk->sk_state == TCP_SYN_SENT)
3397 if (sock_flag(sk, SOCK_DEAD))
3401 spin_unlock_bh(lock);
3405 sk->sk_err = ETIMEDOUT;
3406 sk->sk_error_report(sk);
3415 spin_unlock_bh(lock);