2 * INET An implementation of the TCP/IP protocol suite for the LINUX
3 * operating system. INET is implemented using the BSD Socket
4 * interface as the means of communication with the user level.
6 * Implementation of the Transmission Control Protocol(TCP).
9 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
10 * Mark Evans, <evansmp@uhura.aston.ac.uk>
11 * Corey Minyard <wf-rch!minyard@relay.EU.net>
12 * Florian La Roche, <flla@stud.uni-sb.de>
13 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
14 * Linus Torvalds, <torvalds@cs.helsinki.fi>
15 * Alan Cox, <gw4pts@gw4pts.ampr.org>
16 * Matthew Dillon, <dillon@apollo.west.oic.com>
17 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
18 * Jorge Cwik, <jorge@laser.satlink.net>
21 * Alan Cox : Numerous verify_area() calls
22 * Alan Cox : Set the ACK bit on a reset
23 * Alan Cox : Stopped it crashing if it closed while
24 * sk->inuse=1 and was trying to connect
26 * Alan Cox : All icmp error handling was broken
27 * pointers passed where wrong and the
28 * socket was looked up backwards. Nobody
29 * tested any icmp error code obviously.
30 * Alan Cox : tcp_err() now handled properly. It
31 * wakes people on errors. poll
32 * behaves and the icmp error race
33 * has gone by moving it into sock.c
34 * Alan Cox : tcp_send_reset() fixed to work for
35 * everything not just packets for
37 * Alan Cox : tcp option processing.
38 * Alan Cox : Reset tweaked (still not 100%) [Had
40 * Herp Rosmanith : More reset fixes
41 * Alan Cox : No longer acks invalid rst frames.
42 * Acking any kind of RST is right out.
43 * Alan Cox : Sets an ignore me flag on an rst
44 * receive otherwise odd bits of prattle
46 * Alan Cox : Fixed another acking RST frame bug.
47 * Should stop LAN workplace lockups.
48 * Alan Cox : Some tidyups using the new skb list
50 * Alan Cox : sk->keepopen now seems to work
51 * Alan Cox : Pulls options out correctly on accepts
52 * Alan Cox : Fixed assorted sk->rqueue->next errors
53 * Alan Cox : PSH doesn't end a TCP read. Switched a
55 * Alan Cox : Tidied tcp_data to avoid a potential
57 * Alan Cox : Added some better commenting, as the
58 * tcp is hard to follow
59 * Alan Cox : Removed incorrect check for 20 * psh
60 * Michael O'Reilly : ack < copied bug fix.
61 * Johannes Stille : Misc tcp fixes (not all in yet).
62 * Alan Cox : FIN with no memory -> CRASH
63 * Alan Cox : Added socket option proto entries.
64 * Also added awareness of them to accept.
65 * Alan Cox : Added TCP options (SOL_TCP)
66 * Alan Cox : Switched wakeup calls to callbacks,
67 * so the kernel can layer network
69 * Alan Cox : Use ip_tos/ip_ttl settings.
70 * Alan Cox : Handle FIN (more) properly (we hope).
71 * Alan Cox : RST frames sent on unsynchronised
73 * Alan Cox : Put in missing check for SYN bit.
74 * Alan Cox : Added tcp_select_window() aka NET2E
75 * window non shrink trick.
76 * Alan Cox : Added a couple of small NET2E timer
78 * Charles Hedrick : TCP fixes
79 * Toomas Tamm : TCP window fixes
80 * Alan Cox : Small URG fix to rlogin ^C ack fight
81 * Charles Hedrick : Rewrote most of it to actually work
82 * Linus : Rewrote tcp_read() and URG handling
84 * Gerhard Koerting: Fixed some missing timer handling
85 * Matthew Dillon : Reworked TCP machine states as per RFC
86 * Gerhard Koerting: PC/TCP workarounds
87 * Adam Caldwell : Assorted timer/timing errors
88 * Matthew Dillon : Fixed another RST bug
89 * Alan Cox : Move to kernel side addressing changes.
90 * Alan Cox : Beginning work on TCP fastpathing
92 * Arnt Gulbrandsen: Turbocharged tcp_check() routine.
93 * Alan Cox : TCP fast path debugging
94 * Alan Cox : Window clamping
95 * Michael Riepe : Bug in tcp_check()
96 * Matt Dillon : More TCP improvements and RST bug fixes
97 * Matt Dillon : Yet more small nasties remove from the
98 * TCP code (Be very nice to this man if
99 * tcp finally works 100%) 8)
100 * Alan Cox : BSD accept semantics.
101 * Alan Cox : Reset on closedown bug.
102 * Peter De Schrijver : ENOTCONN check missing in tcp_sendto().
103 * Michael Pall : Handle poll() after URG properly in
105 * Michael Pall : Undo the last fix in tcp_read_urg()
106 * (multi URG PUSH broke rlogin).
107 * Michael Pall : Fix the multi URG PUSH problem in
108 * tcp_readable(), poll() after URG
110 * Michael Pall : recv(...,MSG_OOB) never blocks in the
112 * Alan Cox : Changed the semantics of sk->socket to
113 * fix a race and a signal problem with
114 * accept() and async I/O.
115 * Alan Cox : Relaxed the rules on tcp_sendto().
116 * Yury Shevchuk : Really fixed accept() blocking problem.
117 * Craig I. Hagan : Allow for BSD compatible TIME_WAIT for
118 * clients/servers which listen in on
120 * Alan Cox : Cleaned the above up and shrank it to
121 * a sensible code size.
122 * Alan Cox : Self connect lockup fix.
123 * Alan Cox : No connect to multicast.
124 * Ross Biro : Close unaccepted children on master
126 * Alan Cox : Reset tracing code.
127 * Alan Cox : Spurious resets on shutdown.
128 * Alan Cox : Giant 15 minute/60 second timer error
129 * Alan Cox : Small whoops in polling before an
131 * Alan Cox : Kept the state trace facility since
132 * it's handy for debugging.
133 * Alan Cox : More reset handler fixes.
134 * Alan Cox : Started rewriting the code based on
135 * the RFC's for other useful protocol
136 * references see: Comer, KA9Q NOS, and
137 * for a reference on the difference
138 * between specifications and how BSD
139 * works see the 4.4lite source.
140 * A.N.Kuznetsov : Don't time wait on completion of tidy
142 * Linus Torvalds : Fin/Shutdown & copied_seq changes.
143 * Linus Torvalds : Fixed BSD port reuse to work first syn
144 * Alan Cox : Reimplemented timers as per the RFC
145 * and using multiple timers for sanity.
146 * Alan Cox : Small bug fixes, and a lot of new
148 * Alan Cox : Fixed dual reader crash by locking
149 * the buffers (much like datagram.c)
150 * Alan Cox : Fixed stuck sockets in probe. A probe
151 * now gets fed up of retrying without
152 * (even a no space) answer.
153 * Alan Cox : Extracted closing code better
154 * Alan Cox : Fixed the closing state machine to
156 * Alan Cox : More 'per spec' fixes.
157 * Jorge Cwik : Even faster checksumming.
158 * Alan Cox : tcp_data() doesn't ack illegal PSH
159 * only frames. At least one pc tcp stack
161 * Alan Cox : Cache last socket.
162 * Alan Cox : Per route irtt.
163 * Matt Day : poll()->select() match BSD precisely on error
164 * Alan Cox : New buffers
165 * Marc Tamsky : Various sk->prot->retransmits and
166 * sk->retransmits misupdating fixed.
167 * Fixed tcp_write_timeout: stuck close,
168 * and TCP syn retries gets used now.
169 * Mark Yarvis : In tcp_read_wakeup(), don't send an
170 * ack if state is TCP_CLOSED.
171 * Alan Cox : Look up device on a retransmit - routes may
172 * change. Doesn't yet cope with MSS shrink right
174 * Marc Tamsky : Closing in closing fixes.
175 * Mike Shaver : RFC1122 verifications.
176 * Alan Cox : rcv_saddr errors.
177 * Alan Cox : Block double connect().
178 * Alan Cox : Small hooks for enSKIP.
179 * Alexey Kuznetsov: Path MTU discovery.
180 * Alan Cox : Support soft errors.
181 * Alan Cox : Fix MTU discovery pathological case
182 * when the remote claims no mtu!
183 * Marc Tamsky : TCP_CLOSE fix.
184 * Colin (G3TNE) : Send a reset on syn ack replies in
185 * window but wrong (fixes NT lpd problems)
186 * Pedro Roque : Better TCP window handling, delayed ack.
187 * Joerg Reuter : No modification of locked buffers in
188 * tcp_do_retransmit()
189 * Eric Schenk : Changed receiver side silly window
190 * avoidance algorithm to BSD style
191 * algorithm. This doubles throughput
192 * against machines running Solaris,
193 * and seems to result in general
195 * Stefan Magdalinski : adjusted tcp_readable() to fix FIONREAD
196 * Willy Konynenberg : Transparent proxying support.
197 * Mike McLagan : Routing by source
198 * Keith Owens : Do proper merging with partial SKB's in
199 * tcp_do_sendmsg to avoid burstiness.
200 * Eric Schenk : Fix fast close down bug with
201 * shutdown() followed by close().
202 * Andi Kleen : Make poll agree with SIGIO
203 * Salvatore Sanfilippo : Support SO_LINGER with linger == 1 and
204 * lingertime == 0 (RFC 793 ABORT Call)
205 * Hirokazu Takahashi : Use copy_from_user() instead of
206 * csum_and_copy_from_user() if possible.
208 * This program is free software; you can redistribute it and/or
209 * modify it under the terms of the GNU General Public License
210 * as published by the Free Software Foundation; either version
211 * 2 of the License, or(at your option) any later version.
213 * Description of States:
215 * TCP_SYN_SENT sent a connection request, waiting for ack
217 * TCP_SYN_RECV received a connection request, sent ack,
218 * waiting for final ack in three-way handshake.
220 * TCP_ESTABLISHED connection established
222 * TCP_FIN_WAIT1 our side has shutdown, waiting to complete
223 * transmission of remaining buffered data
225 * TCP_FIN_WAIT2 all buffered data sent, waiting for remote
228 * TCP_CLOSING both sides have shutdown but we still have
229 * data we have to finish sending
231 * TCP_TIME_WAIT timeout to catch resent junk before entering
232 * closed, can only be entered from FIN_WAIT2
233 * or CLOSING. Required because the other end
234 * may not have gotten our last ACK causing it
235 * to retransmit the data packet (which we ignore)
237 * TCP_CLOSE_WAIT remote side has shutdown and is waiting for
238 * us to finish writing our data and to shutdown
239 * (we have to close() to move on to LAST_ACK)
241 * TCP_LAST_ACK out side has shutdown after remote has
242 * shutdown. There may still be data in our
243 * buffer that we have to finish sending
245 * TCP_CLOSE socket is finished
248 #define pr_fmt(fmt) "TCP: " fmt
250 #include <linux/kernel.h>
251 #include <linux/module.h>
252 #include <linux/types.h>
253 #include <linux/fcntl.h>
254 #include <linux/poll.h>
255 #include <linux/init.h>
256 #include <linux/fs.h>
257 #include <linux/skbuff.h>
258 #include <linux/scatterlist.h>
259 #include <linux/splice.h>
260 #include <linux/net.h>
261 #include <linux/socket.h>
262 #include <linux/random.h>
263 #include <linux/bootmem.h>
264 #include <linux/highmem.h>
265 #include <linux/swap.h>
266 #include <linux/cache.h>
267 #include <linux/err.h>
268 #include <linux/crypto.h>
269 #include <linux/time.h>
270 #include <linux/slab.h>
272 #include <net/icmp.h>
273 #include <net/inet_common.h>
275 #include <net/xfrm.h>
277 #include <net/netdma.h>
278 #include <net/sock.h>
280 #include <asm/uaccess.h>
281 #include <asm/ioctls.h>
283 int sysctl_tcp_fin_timeout __read_mostly = TCP_FIN_TIMEOUT;
285 struct percpu_counter tcp_orphan_count;
286 EXPORT_SYMBOL_GPL(tcp_orphan_count);
288 int sysctl_tcp_wmem[3] __read_mostly;
289 int sysctl_tcp_rmem[3] __read_mostly;
291 EXPORT_SYMBOL(sysctl_tcp_rmem);
292 EXPORT_SYMBOL(sysctl_tcp_wmem);
294 atomic_long_t tcp_memory_allocated; /* Current allocated memory. */
295 EXPORT_SYMBOL(tcp_memory_allocated);
298 * Current number of TCP sockets.
300 struct percpu_counter tcp_sockets_allocated;
301 EXPORT_SYMBOL(tcp_sockets_allocated);
306 struct tcp_splice_state {
307 struct pipe_inode_info *pipe;
313 * Pressure flag: try to collapse.
314 * Technical note: it is used by multiple contexts non atomically.
315 * All the __sk_mem_schedule() is of this nature: accounting
316 * is strict, actions are advisory and have some latency.
318 int tcp_memory_pressure __read_mostly;
319 EXPORT_SYMBOL(tcp_memory_pressure);
321 void tcp_enter_memory_pressure(struct sock *sk)
323 if (!tcp_memory_pressure) {
324 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMEMORYPRESSURES);
325 tcp_memory_pressure = 1;
328 EXPORT_SYMBOL(tcp_enter_memory_pressure);
330 /* Convert seconds to retransmits based on initial and max timeout */
331 static u8 secs_to_retrans(int seconds, int timeout, int rto_max)
336 int period = timeout;
339 while (seconds > period && res < 255) {
342 if (timeout > rto_max)
350 /* Convert retransmits to seconds based on initial and max timeout */
351 static int retrans_to_secs(u8 retrans, int timeout, int rto_max)
359 if (timeout > rto_max)
367 /* Address-family independent initialization for a tcp_sock.
369 * NOTE: A lot of things set to zero explicitly by call to
370 * sk_alloc() so need not be done here.
372 void tcp_init_sock(struct sock *sk)
374 struct inet_connection_sock *icsk = inet_csk(sk);
375 struct tcp_sock *tp = tcp_sk(sk);
377 skb_queue_head_init(&tp->out_of_order_queue);
378 tcp_init_xmit_timers(sk);
379 tcp_prequeue_init(tp);
380 INIT_LIST_HEAD(&tp->tsq_node);
382 icsk->icsk_rto = TCP_TIMEOUT_INIT;
383 tp->mdev = TCP_TIMEOUT_INIT;
385 /* So many TCP implementations out there (incorrectly) count the
386 * initial SYN frame in their delayed-ACK and congestion control
387 * algorithms that we must have the following bandaid to talk
388 * efficiently to them. -DaveM
390 tp->snd_cwnd = TCP_INIT_CWND;
392 /* See draft-stevens-tcpca-spec-01 for discussion of the
393 * initialization of these values.
395 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
396 tp->snd_cwnd_clamp = ~0;
397 tp->mss_cache = TCP_MSS_DEFAULT;
399 tp->reordering = sysctl_tcp_reordering;
400 tcp_enable_early_retrans(tp);
401 icsk->icsk_ca_ops = &tcp_init_congestion_ops;
405 sk->sk_state = TCP_CLOSE;
407 sk->sk_write_space = sk_stream_write_space;
408 sock_set_flag(sk, SOCK_USE_WRITE_QUEUE);
410 icsk->icsk_sync_mss = tcp_sync_mss;
412 /* Presumed zeroed, in order of appearance:
413 * cookie_in_always, cookie_out_never,
414 * s_data_constant, s_data_in, s_data_out
416 sk->sk_sndbuf = sysctl_tcp_wmem[1];
417 sk->sk_rcvbuf = sysctl_tcp_rmem[1];
420 sock_update_memcg(sk);
421 sk_sockets_allocated_inc(sk);
424 EXPORT_SYMBOL(tcp_init_sock);
427 * Wait for a TCP event.
429 * Note that we don't need to lock the socket, as the upper poll layers
430 * take care of normal races (between the test and the event) and we don't
431 * go look at any of the socket buffers directly.
433 unsigned int tcp_poll(struct file *file, struct socket *sock, poll_table *wait)
436 struct sock *sk = sock->sk;
437 const struct tcp_sock *tp = tcp_sk(sk);
439 sock_rps_record_flow(sk);
441 sock_poll_wait(file, sk_sleep(sk), wait);
442 if (sk->sk_state == TCP_LISTEN)
443 return inet_csk_listen_poll(sk);
445 /* Socket is not locked. We are protected from async events
446 * by poll logic and correct handling of state changes
447 * made by other threads is impossible in any case.
453 * POLLHUP is certainly not done right. But poll() doesn't
454 * have a notion of HUP in just one direction, and for a
455 * socket the read side is more interesting.
457 * Some poll() documentation says that POLLHUP is incompatible
458 * with the POLLOUT/POLLWR flags, so somebody should check this
459 * all. But careful, it tends to be safer to return too many
460 * bits than too few, and you can easily break real applications
461 * if you don't tell them that something has hung up!
465 * Check number 1. POLLHUP is _UNMASKABLE_ event (see UNIX98 and
466 * our fs/select.c). It means that after we received EOF,
467 * poll always returns immediately, making impossible poll() on write()
468 * in state CLOSE_WAIT. One solution is evident --- to set POLLHUP
469 * if and only if shutdown has been made in both directions.
470 * Actually, it is interesting to look how Solaris and DUX
471 * solve this dilemma. I would prefer, if POLLHUP were maskable,
472 * then we could set it on SND_SHUTDOWN. BTW examples given
473 * in Stevens' books assume exactly this behaviour, it explains
474 * why POLLHUP is incompatible with POLLOUT. --ANK
476 * NOTE. Check for TCP_CLOSE is added. The goal is to prevent
477 * blocking on fresh not-connected or disconnected socket. --ANK
479 if (sk->sk_shutdown == SHUTDOWN_MASK || sk->sk_state == TCP_CLOSE)
481 if (sk->sk_shutdown & RCV_SHUTDOWN)
482 mask |= POLLIN | POLLRDNORM | POLLRDHUP;
484 /* Connected or passive Fast Open socket? */
485 if (sk->sk_state != TCP_SYN_SENT &&
486 (sk->sk_state != TCP_SYN_RECV || tp->fastopen_rsk != NULL)) {
487 int target = sock_rcvlowat(sk, 0, INT_MAX);
489 if (tp->urg_seq == tp->copied_seq &&
490 !sock_flag(sk, SOCK_URGINLINE) &&
494 /* Potential race condition. If read of tp below will
495 * escape above sk->sk_state, we can be illegally awaken
496 * in SYN_* states. */
497 if (tp->rcv_nxt - tp->copied_seq >= target)
498 mask |= POLLIN | POLLRDNORM;
500 if (!(sk->sk_shutdown & SEND_SHUTDOWN)) {
501 if (sk_stream_wspace(sk) >= sk_stream_min_wspace(sk)) {
502 mask |= POLLOUT | POLLWRNORM;
503 } else { /* send SIGIO later */
504 set_bit(SOCK_ASYNC_NOSPACE,
505 &sk->sk_socket->flags);
506 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
508 /* Race breaker. If space is freed after
509 * wspace test but before the flags are set,
510 * IO signal will be lost.
512 if (sk_stream_wspace(sk) >= sk_stream_min_wspace(sk))
513 mask |= POLLOUT | POLLWRNORM;
516 mask |= POLLOUT | POLLWRNORM;
518 if (tp->urg_data & TCP_URG_VALID)
521 /* This barrier is coupled with smp_wmb() in tcp_reset() */
528 EXPORT_SYMBOL(tcp_poll);
530 int tcp_ioctl(struct sock *sk, int cmd, unsigned long arg)
532 struct tcp_sock *tp = tcp_sk(sk);
538 if (sk->sk_state == TCP_LISTEN)
541 slow = lock_sock_fast(sk);
542 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
544 else if (sock_flag(sk, SOCK_URGINLINE) ||
546 before(tp->urg_seq, tp->copied_seq) ||
547 !before(tp->urg_seq, tp->rcv_nxt)) {
549 answ = tp->rcv_nxt - tp->copied_seq;
551 /* Subtract 1, if FIN was received */
552 if (answ && sock_flag(sk, SOCK_DONE))
555 answ = tp->urg_seq - tp->copied_seq;
556 unlock_sock_fast(sk, slow);
559 answ = tp->urg_data && tp->urg_seq == tp->copied_seq;
562 if (sk->sk_state == TCP_LISTEN)
565 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
568 answ = tp->write_seq - tp->snd_una;
571 if (sk->sk_state == TCP_LISTEN)
574 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
577 answ = tp->write_seq - tp->snd_nxt;
583 return put_user(answ, (int __user *)arg);
585 EXPORT_SYMBOL(tcp_ioctl);
587 static inline void tcp_mark_push(struct tcp_sock *tp, struct sk_buff *skb)
589 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
590 tp->pushed_seq = tp->write_seq;
593 static inline bool forced_push(const struct tcp_sock *tp)
595 return after(tp->write_seq, tp->pushed_seq + (tp->max_window >> 1));
598 static inline void skb_entail(struct sock *sk, struct sk_buff *skb)
600 struct tcp_sock *tp = tcp_sk(sk);
601 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
604 tcb->seq = tcb->end_seq = tp->write_seq;
605 tcb->tcp_flags = TCPHDR_ACK;
607 skb_header_release(skb);
608 tcp_add_write_queue_tail(sk, skb);
609 sk->sk_wmem_queued += skb->truesize;
610 sk_mem_charge(sk, skb->truesize);
611 if (tp->nonagle & TCP_NAGLE_PUSH)
612 tp->nonagle &= ~TCP_NAGLE_PUSH;
615 static inline void tcp_mark_urg(struct tcp_sock *tp, int flags)
618 tp->snd_up = tp->write_seq;
621 static inline void tcp_push(struct sock *sk, int flags, int mss_now,
624 if (tcp_send_head(sk)) {
625 struct tcp_sock *tp = tcp_sk(sk);
627 if (!(flags & MSG_MORE) || forced_push(tp))
628 tcp_mark_push(tp, tcp_write_queue_tail(sk));
630 tcp_mark_urg(tp, flags);
631 __tcp_push_pending_frames(sk, mss_now,
632 (flags & MSG_MORE) ? TCP_NAGLE_CORK : nonagle);
636 static int tcp_splice_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb,
637 unsigned int offset, size_t len)
639 struct tcp_splice_state *tss = rd_desc->arg.data;
642 ret = skb_splice_bits(skb, offset, tss->pipe, min(rd_desc->count, len),
645 rd_desc->count -= ret;
649 static int __tcp_splice_read(struct sock *sk, struct tcp_splice_state *tss)
651 /* Store TCP splice context information in read_descriptor_t. */
652 read_descriptor_t rd_desc = {
657 return tcp_read_sock(sk, &rd_desc, tcp_splice_data_recv);
661 * tcp_splice_read - splice data from TCP socket to a pipe
662 * @sock: socket to splice from
663 * @ppos: position (not valid)
664 * @pipe: pipe to splice to
665 * @len: number of bytes to splice
666 * @flags: splice modifier flags
669 * Will read pages from given socket and fill them into a pipe.
672 ssize_t tcp_splice_read(struct socket *sock, loff_t *ppos,
673 struct pipe_inode_info *pipe, size_t len,
676 struct sock *sk = sock->sk;
677 struct tcp_splice_state tss = {
686 sock_rps_record_flow(sk);
688 * We can't seek on a socket input
697 timeo = sock_rcvtimeo(sk, sock->file->f_flags & O_NONBLOCK);
699 ret = __tcp_splice_read(sk, &tss);
705 if (sock_flag(sk, SOCK_DONE))
708 ret = sock_error(sk);
711 if (sk->sk_shutdown & RCV_SHUTDOWN)
713 if (sk->sk_state == TCP_CLOSE) {
715 * This occurs when user tries to read
716 * from never connected socket.
718 if (!sock_flag(sk, SOCK_DONE))
726 sk_wait_data(sk, &timeo);
727 if (signal_pending(current)) {
728 ret = sock_intr_errno(timeo);
741 if (sk->sk_err || sk->sk_state == TCP_CLOSE ||
742 (sk->sk_shutdown & RCV_SHUTDOWN) ||
743 signal_pending(current))
754 EXPORT_SYMBOL(tcp_splice_read);
756 struct sk_buff *sk_stream_alloc_skb(struct sock *sk, int size, gfp_t gfp)
760 /* The TCP header must be at least 32-bit aligned. */
761 size = ALIGN(size, 4);
763 skb = alloc_skb_fclone(size + sk->sk_prot->max_header, gfp);
765 if (sk_wmem_schedule(sk, skb->truesize)) {
766 skb_reserve(skb, sk->sk_prot->max_header);
768 * Make sure that we have exactly size bytes
769 * available to the caller, no more, no less.
771 skb->reserved_tailroom = skb->end - skb->tail - size;
776 sk->sk_prot->enter_memory_pressure(sk);
777 sk_stream_moderate_sndbuf(sk);
782 static unsigned int tcp_xmit_size_goal(struct sock *sk, u32 mss_now,
785 struct tcp_sock *tp = tcp_sk(sk);
786 u32 xmit_size_goal, old_size_goal;
788 xmit_size_goal = mss_now;
790 if (large_allowed && sk_can_gso(sk)) {
791 xmit_size_goal = ((sk->sk_gso_max_size - 1) -
792 inet_csk(sk)->icsk_af_ops->net_header_len -
793 inet_csk(sk)->icsk_ext_hdr_len -
796 /* TSQ : try to have two TSO segments in flight */
797 xmit_size_goal = min_t(u32, xmit_size_goal,
798 sysctl_tcp_limit_output_bytes >> 1);
800 xmit_size_goal = tcp_bound_to_half_wnd(tp, xmit_size_goal);
802 /* We try hard to avoid divides here */
803 old_size_goal = tp->xmit_size_goal_segs * mss_now;
805 if (likely(old_size_goal <= xmit_size_goal &&
806 old_size_goal + mss_now > xmit_size_goal)) {
807 xmit_size_goal = old_size_goal;
809 tp->xmit_size_goal_segs =
810 min_t(u16, xmit_size_goal / mss_now,
811 sk->sk_gso_max_segs);
812 xmit_size_goal = tp->xmit_size_goal_segs * mss_now;
816 return max(xmit_size_goal, mss_now);
819 static int tcp_send_mss(struct sock *sk, int *size_goal, int flags)
823 mss_now = tcp_current_mss(sk);
824 *size_goal = tcp_xmit_size_goal(sk, mss_now, !(flags & MSG_OOB));
829 static ssize_t do_tcp_sendpages(struct sock *sk, struct page *page, int offset,
830 size_t size, int flags)
832 struct tcp_sock *tp = tcp_sk(sk);
833 int mss_now, size_goal;
836 long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
838 /* Wait for a connection to finish. One exception is TCP Fast Open
839 * (passive side) where data is allowed to be sent before a connection
840 * is fully established.
842 if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) &&
843 !tcp_passive_fastopen(sk)) {
844 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
848 clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
850 mss_now = tcp_send_mss(sk, &size_goal, flags);
854 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
858 struct sk_buff *skb = tcp_write_queue_tail(sk);
862 if (!tcp_send_head(sk) || (copy = size_goal - skb->len) <= 0) {
864 if (!sk_stream_memory_free(sk))
865 goto wait_for_sndbuf;
867 skb = sk_stream_alloc_skb(sk, 0, sk->sk_allocation);
869 goto wait_for_memory;
878 i = skb_shinfo(skb)->nr_frags;
879 can_coalesce = skb_can_coalesce(skb, i, page, offset);
880 if (!can_coalesce && i >= MAX_SKB_FRAGS) {
881 tcp_mark_push(tp, skb);
884 if (!sk_wmem_schedule(sk, copy))
885 goto wait_for_memory;
888 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
891 skb_fill_page_desc(skb, i, page, offset, copy);
893 skb_shinfo(skb)->tx_flags |= SKBTX_SHARED_FRAG;
896 skb->data_len += copy;
897 skb->truesize += copy;
898 sk->sk_wmem_queued += copy;
899 sk_mem_charge(sk, copy);
900 skb->ip_summed = CHECKSUM_PARTIAL;
901 tp->write_seq += copy;
902 TCP_SKB_CB(skb)->end_seq += copy;
903 skb_shinfo(skb)->gso_segs = 0;
906 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
913 if (skb->len < size_goal || (flags & MSG_OOB))
916 if (forced_push(tp)) {
917 tcp_mark_push(tp, skb);
918 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
919 } else if (skb == tcp_send_head(sk))
920 tcp_push_one(sk, mss_now);
924 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
926 tcp_push(sk, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH);
928 if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
931 mss_now = tcp_send_mss(sk, &size_goal, flags);
935 if (copied && !(flags & MSG_SENDPAGE_NOTLAST))
936 tcp_push(sk, flags, mss_now, tp->nonagle);
943 return sk_stream_error(sk, flags, err);
946 int tcp_sendpage(struct sock *sk, struct page *page, int offset,
947 size_t size, int flags)
951 if (!(sk->sk_route_caps & NETIF_F_SG) ||
952 !(sk->sk_route_caps & NETIF_F_ALL_CSUM))
953 return sock_no_sendpage(sk->sk_socket, page, offset, size,
957 res = do_tcp_sendpages(sk, page, offset, size, flags);
961 EXPORT_SYMBOL(tcp_sendpage);
963 static inline int select_size(const struct sock *sk, bool sg)
965 const struct tcp_sock *tp = tcp_sk(sk);
966 int tmp = tp->mss_cache;
969 if (sk_can_gso(sk)) {
970 /* Small frames wont use a full page:
971 * Payload will immediately follow tcp header.
973 tmp = SKB_WITH_OVERHEAD(2048 - MAX_TCP_HEADER);
975 int pgbreak = SKB_MAX_HEAD(MAX_TCP_HEADER);
977 if (tmp >= pgbreak &&
978 tmp <= pgbreak + (MAX_SKB_FRAGS - 1) * PAGE_SIZE)
986 void tcp_free_fastopen_req(struct tcp_sock *tp)
988 if (tp->fastopen_req != NULL) {
989 kfree(tp->fastopen_req);
990 tp->fastopen_req = NULL;
994 static int tcp_sendmsg_fastopen(struct sock *sk, struct msghdr *msg, int *size)
996 struct tcp_sock *tp = tcp_sk(sk);
999 if (!(sysctl_tcp_fastopen & TFO_CLIENT_ENABLE))
1001 if (tp->fastopen_req != NULL)
1002 return -EALREADY; /* Another Fast Open is in progress */
1004 tp->fastopen_req = kzalloc(sizeof(struct tcp_fastopen_request),
1006 if (unlikely(tp->fastopen_req == NULL))
1008 tp->fastopen_req->data = msg;
1010 flags = (msg->msg_flags & MSG_DONTWAIT) ? O_NONBLOCK : 0;
1011 err = __inet_stream_connect(sk->sk_socket, msg->msg_name,
1012 msg->msg_namelen, flags);
1013 *size = tp->fastopen_req->copied;
1014 tcp_free_fastopen_req(tp);
1018 int tcp_sendmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg,
1022 struct tcp_sock *tp = tcp_sk(sk);
1023 struct sk_buff *skb;
1024 int iovlen, flags, err, copied = 0;
1025 int mss_now = 0, size_goal, copied_syn = 0, offset = 0;
1031 flags = msg->msg_flags;
1032 if (flags & MSG_FASTOPEN) {
1033 err = tcp_sendmsg_fastopen(sk, msg, &copied_syn);
1034 if (err == -EINPROGRESS && copied_syn > 0)
1038 offset = copied_syn;
1041 timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
1043 /* Wait for a connection to finish. One exception is TCP Fast Open
1044 * (passive side) where data is allowed to be sent before a connection
1045 * is fully established.
1047 if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) &&
1048 !tcp_passive_fastopen(sk)) {
1049 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
1053 if (unlikely(tp->repair)) {
1054 if (tp->repair_queue == TCP_RECV_QUEUE) {
1055 copied = tcp_send_rcvq(sk, msg, size);
1060 if (tp->repair_queue == TCP_NO_QUEUE)
1063 /* 'common' sending to sendq */
1066 /* This should be in poll */
1067 clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
1069 mss_now = tcp_send_mss(sk, &size_goal, flags);
1071 /* Ok commence sending. */
1072 iovlen = msg->msg_iovlen;
1077 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
1080 sg = !!(sk->sk_route_caps & NETIF_F_SG);
1082 while (--iovlen >= 0) {
1083 size_t seglen = iov->iov_len;
1084 unsigned char __user *from = iov->iov_base;
1087 if (unlikely(offset > 0)) { /* Skip bytes copied in SYN */
1088 if (offset >= seglen) {
1097 while (seglen > 0) {
1099 int max = size_goal;
1101 skb = tcp_write_queue_tail(sk);
1102 if (tcp_send_head(sk)) {
1103 if (skb->ip_summed == CHECKSUM_NONE)
1105 copy = max - skb->len;
1110 /* Allocate new segment. If the interface is SG,
1111 * allocate skb fitting to single page.
1113 if (!sk_stream_memory_free(sk))
1114 goto wait_for_sndbuf;
1116 skb = sk_stream_alloc_skb(sk,
1117 select_size(sk, sg),
1120 goto wait_for_memory;
1123 * Check whether we can use HW checksum.
1125 if (sk->sk_route_caps & NETIF_F_ALL_CSUM)
1126 skb->ip_summed = CHECKSUM_PARTIAL;
1128 skb_entail(sk, skb);
1133 /* Try to append data to the end of skb. */
1137 /* Where to copy to? */
1138 if (skb_availroom(skb) > 0) {
1139 /* We have some space in skb head. Superb! */
1140 copy = min_t(int, copy, skb_availroom(skb));
1141 err = skb_add_data_nocache(sk, skb, from, copy);
1146 int i = skb_shinfo(skb)->nr_frags;
1147 struct page_frag *pfrag = sk_page_frag(sk);
1149 if (!sk_page_frag_refill(sk, pfrag))
1150 goto wait_for_memory;
1152 if (!skb_can_coalesce(skb, i, pfrag->page,
1154 if (i == MAX_SKB_FRAGS || !sg) {
1155 tcp_mark_push(tp, skb);
1161 copy = min_t(int, copy, pfrag->size - pfrag->offset);
1163 if (!sk_wmem_schedule(sk, copy))
1164 goto wait_for_memory;
1166 err = skb_copy_to_page_nocache(sk, from, skb,
1173 /* Update the skb. */
1175 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1177 skb_fill_page_desc(skb, i, pfrag->page,
1178 pfrag->offset, copy);
1179 get_page(pfrag->page);
1181 pfrag->offset += copy;
1185 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
1187 tp->write_seq += copy;
1188 TCP_SKB_CB(skb)->end_seq += copy;
1189 skb_shinfo(skb)->gso_segs = 0;
1193 if ((seglen -= copy) == 0 && iovlen == 0)
1196 if (skb->len < max || (flags & MSG_OOB) || unlikely(tp->repair))
1199 if (forced_push(tp)) {
1200 tcp_mark_push(tp, skb);
1201 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
1202 } else if (skb == tcp_send_head(sk))
1203 tcp_push_one(sk, mss_now);
1207 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1210 tcp_push(sk, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH);
1212 if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
1215 mss_now = tcp_send_mss(sk, &size_goal, flags);
1221 tcp_push(sk, flags, mss_now, tp->nonagle);
1223 return copied + copied_syn;
1227 tcp_unlink_write_queue(skb, sk);
1228 /* It is the one place in all of TCP, except connection
1229 * reset, where we can be unlinking the send_head.
1231 tcp_check_send_head(sk, skb);
1232 sk_wmem_free_skb(sk, skb);
1236 if (copied + copied_syn)
1239 err = sk_stream_error(sk, flags, err);
1243 EXPORT_SYMBOL(tcp_sendmsg);
1246 * Handle reading urgent data. BSD has very simple semantics for
1247 * this, no blocking and very strange errors 8)
1250 static int tcp_recv_urg(struct sock *sk, struct msghdr *msg, int len, int flags)
1252 struct tcp_sock *tp = tcp_sk(sk);
1254 /* No URG data to read. */
1255 if (sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data ||
1256 tp->urg_data == TCP_URG_READ)
1257 return -EINVAL; /* Yes this is right ! */
1259 if (sk->sk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DONE))
1262 if (tp->urg_data & TCP_URG_VALID) {
1264 char c = tp->urg_data;
1266 if (!(flags & MSG_PEEK))
1267 tp->urg_data = TCP_URG_READ;
1269 /* Read urgent data. */
1270 msg->msg_flags |= MSG_OOB;
1273 if (!(flags & MSG_TRUNC))
1274 err = memcpy_toiovec(msg->msg_iov, &c, 1);
1277 msg->msg_flags |= MSG_TRUNC;
1279 return err ? -EFAULT : len;
1282 if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
1285 /* Fixed the recv(..., MSG_OOB) behaviour. BSD docs and
1286 * the available implementations agree in this case:
1287 * this call should never block, independent of the
1288 * blocking state of the socket.
1289 * Mike <pall@rz.uni-karlsruhe.de>
1294 static int tcp_peek_sndq(struct sock *sk, struct msghdr *msg, int len)
1296 struct sk_buff *skb;
1297 int copied = 0, err = 0;
1299 /* XXX -- need to support SO_PEEK_OFF */
1301 skb_queue_walk(&sk->sk_write_queue, skb) {
1302 err = skb_copy_datagram_iovec(skb, 0, msg->msg_iov, skb->len);
1309 return err ?: copied;
1312 /* Clean up the receive buffer for full frames taken by the user,
1313 * then send an ACK if necessary. COPIED is the number of bytes
1314 * tcp_recvmsg has given to the user so far, it speeds up the
1315 * calculation of whether or not we must ACK for the sake of
1318 void tcp_cleanup_rbuf(struct sock *sk, int copied)
1320 struct tcp_sock *tp = tcp_sk(sk);
1321 bool time_to_ack = false;
1323 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
1325 WARN(skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq),
1326 "cleanup rbuf bug: copied %X seq %X rcvnxt %X\n",
1327 tp->copied_seq, TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt);
1329 if (inet_csk_ack_scheduled(sk)) {
1330 const struct inet_connection_sock *icsk = inet_csk(sk);
1331 /* Delayed ACKs frequently hit locked sockets during bulk
1333 if (icsk->icsk_ack.blocked ||
1334 /* Once-per-two-segments ACK was not sent by tcp_input.c */
1335 tp->rcv_nxt - tp->rcv_wup > icsk->icsk_ack.rcv_mss ||
1337 * If this read emptied read buffer, we send ACK, if
1338 * connection is not bidirectional, user drained
1339 * receive buffer and there was a small segment
1343 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED2) ||
1344 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED) &&
1345 !icsk->icsk_ack.pingpong)) &&
1346 !atomic_read(&sk->sk_rmem_alloc)))
1350 /* We send an ACK if we can now advertise a non-zero window
1351 * which has been raised "significantly".
1353 * Even if window raised up to infinity, do not send window open ACK
1354 * in states, where we will not receive more. It is useless.
1356 if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) {
1357 __u32 rcv_window_now = tcp_receive_window(tp);
1359 /* Optimize, __tcp_select_window() is not cheap. */
1360 if (2*rcv_window_now <= tp->window_clamp) {
1361 __u32 new_window = __tcp_select_window(sk);
1363 /* Send ACK now, if this read freed lots of space
1364 * in our buffer. Certainly, new_window is new window.
1365 * We can advertise it now, if it is not less than current one.
1366 * "Lots" means "at least twice" here.
1368 if (new_window && new_window >= 2 * rcv_window_now)
1376 static void tcp_prequeue_process(struct sock *sk)
1378 struct sk_buff *skb;
1379 struct tcp_sock *tp = tcp_sk(sk);
1381 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPPREQUEUED);
1383 /* RX process wants to run with disabled BHs, though it is not
1386 while ((skb = __skb_dequeue(&tp->ucopy.prequeue)) != NULL)
1387 sk_backlog_rcv(sk, skb);
1390 /* Clear memory counter. */
1391 tp->ucopy.memory = 0;
1394 #ifdef CONFIG_NET_DMA
1395 static void tcp_service_net_dma(struct sock *sk, bool wait)
1397 dma_cookie_t done, used;
1398 dma_cookie_t last_issued;
1399 struct tcp_sock *tp = tcp_sk(sk);
1401 if (!tp->ucopy.dma_chan)
1404 last_issued = tp->ucopy.dma_cookie;
1405 dma_async_issue_pending(tp->ucopy.dma_chan);
1408 if (dma_async_is_tx_complete(tp->ucopy.dma_chan,
1410 &used) == DMA_SUCCESS) {
1411 /* Safe to free early-copied skbs now */
1412 __skb_queue_purge(&sk->sk_async_wait_queue);
1415 struct sk_buff *skb;
1416 while ((skb = skb_peek(&sk->sk_async_wait_queue)) &&
1417 (dma_async_is_complete(skb->dma_cookie, done,
1418 used) == DMA_SUCCESS)) {
1419 __skb_dequeue(&sk->sk_async_wait_queue);
1427 static struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off)
1429 struct sk_buff *skb;
1432 while ((skb = skb_peek(&sk->sk_receive_queue)) != NULL) {
1433 offset = seq - TCP_SKB_CB(skb)->seq;
1434 if (tcp_hdr(skb)->syn)
1436 if (offset < skb->len || tcp_hdr(skb)->fin) {
1440 /* This looks weird, but this can happen if TCP collapsing
1441 * splitted a fat GRO packet, while we released socket lock
1442 * in skb_splice_bits()
1444 sk_eat_skb(sk, skb, false);
1450 * This routine provides an alternative to tcp_recvmsg() for routines
1451 * that would like to handle copying from skbuffs directly in 'sendfile'
1454 * - It is assumed that the socket was locked by the caller.
1455 * - The routine does not block.
1456 * - At present, there is no support for reading OOB data
1457 * or for 'peeking' the socket using this routine
1458 * (although both would be easy to implement).
1460 int tcp_read_sock(struct sock *sk, read_descriptor_t *desc,
1461 sk_read_actor_t recv_actor)
1463 struct sk_buff *skb;
1464 struct tcp_sock *tp = tcp_sk(sk);
1465 u32 seq = tp->copied_seq;
1469 if (sk->sk_state == TCP_LISTEN)
1471 while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1472 if (offset < skb->len) {
1476 len = skb->len - offset;
1477 /* Stop reading if we hit a patch of urgent data */
1479 u32 urg_offset = tp->urg_seq - seq;
1480 if (urg_offset < len)
1485 used = recv_actor(desc, skb, offset, len);
1490 } else if (used <= len) {
1495 /* If recv_actor drops the lock (e.g. TCP splice
1496 * receive) the skb pointer might be invalid when
1497 * getting here: tcp_collapse might have deleted it
1498 * while aggregating skbs from the socket queue.
1500 skb = tcp_recv_skb(sk, seq - 1, &offset);
1503 /* TCP coalescing might have appended data to the skb.
1504 * Try to splice more frags
1506 if (offset + 1 != skb->len)
1509 if (tcp_hdr(skb)->fin) {
1510 sk_eat_skb(sk, skb, false);
1514 sk_eat_skb(sk, skb, false);
1517 tp->copied_seq = seq;
1519 tp->copied_seq = seq;
1521 tcp_rcv_space_adjust(sk);
1523 /* Clean up data we have read: This will do ACK frames. */
1525 tcp_recv_skb(sk, seq, &offset);
1526 tcp_cleanup_rbuf(sk, copied);
1530 EXPORT_SYMBOL(tcp_read_sock);
1533 * This routine copies from a sock struct into the user buffer.
1535 * Technical note: in 2.3 we work on _locked_ socket, so that
1536 * tricks with *seq access order and skb->users are not required.
1537 * Probably, code can be easily improved even more.
1540 int tcp_recvmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg,
1541 size_t len, int nonblock, int flags, int *addr_len)
1543 struct tcp_sock *tp = tcp_sk(sk);
1549 int target; /* Read at least this many bytes */
1551 struct task_struct *user_recv = NULL;
1552 bool copied_early = false;
1553 struct sk_buff *skb;
1559 if (sk->sk_state == TCP_LISTEN)
1562 timeo = sock_rcvtimeo(sk, nonblock);
1564 /* Urgent data needs to be handled specially. */
1565 if (flags & MSG_OOB)
1568 if (unlikely(tp->repair)) {
1570 if (!(flags & MSG_PEEK))
1573 if (tp->repair_queue == TCP_SEND_QUEUE)
1577 if (tp->repair_queue == TCP_NO_QUEUE)
1580 /* 'common' recv queue MSG_PEEK-ing */
1583 seq = &tp->copied_seq;
1584 if (flags & MSG_PEEK) {
1585 peek_seq = tp->copied_seq;
1589 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
1591 #ifdef CONFIG_NET_DMA
1592 tp->ucopy.dma_chan = NULL;
1594 skb = skb_peek_tail(&sk->sk_receive_queue);
1599 available = TCP_SKB_CB(skb)->seq + skb->len - (*seq);
1600 if ((available < target) &&
1601 (len > sysctl_tcp_dma_copybreak) && !(flags & MSG_PEEK) &&
1602 !sysctl_tcp_low_latency &&
1603 net_dma_find_channel()) {
1604 preempt_enable_no_resched();
1605 tp->ucopy.pinned_list =
1606 dma_pin_iovec_pages(msg->msg_iov, len);
1608 preempt_enable_no_resched();
1616 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
1617 if (tp->urg_data && tp->urg_seq == *seq) {
1620 if (signal_pending(current)) {
1621 copied = timeo ? sock_intr_errno(timeo) : -EAGAIN;
1626 /* Next get a buffer. */
1628 skb_queue_walk(&sk->sk_receive_queue, skb) {
1629 /* Now that we have two receive queues this
1632 if (WARN(before(*seq, TCP_SKB_CB(skb)->seq),
1633 "recvmsg bug: copied %X seq %X rcvnxt %X fl %X\n",
1634 *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt,
1638 offset = *seq - TCP_SKB_CB(skb)->seq;
1639 if (tcp_hdr(skb)->syn)
1641 if (offset < skb->len)
1643 if (tcp_hdr(skb)->fin)
1645 WARN(!(flags & MSG_PEEK),
1646 "recvmsg bug 2: copied %X seq %X rcvnxt %X fl %X\n",
1647 *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt, flags);
1650 /* Well, if we have backlog, try to process it now yet. */
1652 if (copied >= target && !sk->sk_backlog.tail)
1657 sk->sk_state == TCP_CLOSE ||
1658 (sk->sk_shutdown & RCV_SHUTDOWN) ||
1660 signal_pending(current))
1663 if (sock_flag(sk, SOCK_DONE))
1667 copied = sock_error(sk);
1671 if (sk->sk_shutdown & RCV_SHUTDOWN)
1674 if (sk->sk_state == TCP_CLOSE) {
1675 if (!sock_flag(sk, SOCK_DONE)) {
1676 /* This occurs when user tries to read
1677 * from never connected socket.
1690 if (signal_pending(current)) {
1691 copied = sock_intr_errno(timeo);
1696 tcp_cleanup_rbuf(sk, copied);
1698 if (!sysctl_tcp_low_latency && tp->ucopy.task == user_recv) {
1699 /* Install new reader */
1700 if (!user_recv && !(flags & (MSG_TRUNC | MSG_PEEK))) {
1701 user_recv = current;
1702 tp->ucopy.task = user_recv;
1703 tp->ucopy.iov = msg->msg_iov;
1706 tp->ucopy.len = len;
1708 WARN_ON(tp->copied_seq != tp->rcv_nxt &&
1709 !(flags & (MSG_PEEK | MSG_TRUNC)));
1711 /* Ugly... If prequeue is not empty, we have to
1712 * process it before releasing socket, otherwise
1713 * order will be broken at second iteration.
1714 * More elegant solution is required!!!
1716 * Look: we have the following (pseudo)queues:
1718 * 1. packets in flight
1723 * Each queue can be processed only if the next ones
1724 * are empty. At this point we have empty receive_queue.
1725 * But prequeue _can_ be not empty after 2nd iteration,
1726 * when we jumped to start of loop because backlog
1727 * processing added something to receive_queue.
1728 * We cannot release_sock(), because backlog contains
1729 * packets arrived _after_ prequeued ones.
1731 * Shortly, algorithm is clear --- to process all
1732 * the queues in order. We could make it more directly,
1733 * requeueing packets from backlog to prequeue, if
1734 * is not empty. It is more elegant, but eats cycles,
1737 if (!skb_queue_empty(&tp->ucopy.prequeue))
1740 /* __ Set realtime policy in scheduler __ */
1743 #ifdef CONFIG_NET_DMA
1744 if (tp->ucopy.dma_chan) {
1745 if (tp->rcv_wnd == 0 &&
1746 !skb_queue_empty(&sk->sk_async_wait_queue)) {
1747 tcp_service_net_dma(sk, true);
1748 tcp_cleanup_rbuf(sk, copied);
1750 dma_async_issue_pending(tp->ucopy.dma_chan);
1753 if (copied >= target) {
1754 /* Do not sleep, just process backlog. */
1758 sk_wait_data(sk, &timeo);
1760 #ifdef CONFIG_NET_DMA
1761 tcp_service_net_dma(sk, false); /* Don't block */
1762 tp->ucopy.wakeup = 0;
1768 /* __ Restore normal policy in scheduler __ */
1770 if ((chunk = len - tp->ucopy.len) != 0) {
1771 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMBACKLOG, chunk);
1776 if (tp->rcv_nxt == tp->copied_seq &&
1777 !skb_queue_empty(&tp->ucopy.prequeue)) {
1779 tcp_prequeue_process(sk);
1781 if ((chunk = len - tp->ucopy.len) != 0) {
1782 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1788 if ((flags & MSG_PEEK) &&
1789 (peek_seq - copied - urg_hole != tp->copied_seq)) {
1790 net_dbg_ratelimited("TCP(%s:%d): Application bug, race in MSG_PEEK\n",
1792 task_pid_nr(current));
1793 peek_seq = tp->copied_seq;
1798 /* Ok so how much can we use? */
1799 used = skb->len - offset;
1803 /* Do we have urgent data here? */
1805 u32 urg_offset = tp->urg_seq - *seq;
1806 if (urg_offset < used) {
1808 if (!sock_flag(sk, SOCK_URGINLINE)) {
1821 if (!(flags & MSG_TRUNC)) {
1822 #ifdef CONFIG_NET_DMA
1823 if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list)
1824 tp->ucopy.dma_chan = net_dma_find_channel();
1826 if (tp->ucopy.dma_chan) {
1827 tp->ucopy.dma_cookie = dma_skb_copy_datagram_iovec(
1828 tp->ucopy.dma_chan, skb, offset,
1830 tp->ucopy.pinned_list);
1832 if (tp->ucopy.dma_cookie < 0) {
1834 pr_alert("%s: dma_cookie < 0\n",
1837 /* Exception. Bailout! */
1843 dma_async_issue_pending(tp->ucopy.dma_chan);
1845 if ((offset + used) == skb->len)
1846 copied_early = true;
1851 err = skb_copy_datagram_iovec(skb, offset,
1852 msg->msg_iov, used);
1854 /* Exception. Bailout! */
1866 tcp_rcv_space_adjust(sk);
1869 if (tp->urg_data && after(tp->copied_seq, tp->urg_seq)) {
1871 tcp_fast_path_check(sk);
1873 if (used + offset < skb->len)
1876 if (tcp_hdr(skb)->fin)
1878 if (!(flags & MSG_PEEK)) {
1879 sk_eat_skb(sk, skb, copied_early);
1880 copied_early = false;
1885 /* Process the FIN. */
1887 if (!(flags & MSG_PEEK)) {
1888 sk_eat_skb(sk, skb, copied_early);
1889 copied_early = false;
1895 if (!skb_queue_empty(&tp->ucopy.prequeue)) {
1898 tp->ucopy.len = copied > 0 ? len : 0;
1900 tcp_prequeue_process(sk);
1902 if (copied > 0 && (chunk = len - tp->ucopy.len) != 0) {
1903 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1909 tp->ucopy.task = NULL;
1913 #ifdef CONFIG_NET_DMA
1914 tcp_service_net_dma(sk, true); /* Wait for queue to drain */
1915 tp->ucopy.dma_chan = NULL;
1917 if (tp->ucopy.pinned_list) {
1918 dma_unpin_iovec_pages(tp->ucopy.pinned_list);
1919 tp->ucopy.pinned_list = NULL;
1923 /* According to UNIX98, msg_name/msg_namelen are ignored
1924 * on connected socket. I was just happy when found this 8) --ANK
1927 /* Clean up data we have read: This will do ACK frames. */
1928 tcp_cleanup_rbuf(sk, copied);
1938 err = tcp_recv_urg(sk, msg, len, flags);
1942 err = tcp_peek_sndq(sk, msg, len);
1945 EXPORT_SYMBOL(tcp_recvmsg);
1947 void tcp_set_state(struct sock *sk, int state)
1949 int oldstate = sk->sk_state;
1952 case TCP_ESTABLISHED:
1953 if (oldstate != TCP_ESTABLISHED)
1954 TCP_INC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
1958 if (oldstate == TCP_CLOSE_WAIT || oldstate == TCP_ESTABLISHED)
1959 TCP_INC_STATS(sock_net(sk), TCP_MIB_ESTABRESETS);
1961 sk->sk_prot->unhash(sk);
1962 if (inet_csk(sk)->icsk_bind_hash &&
1963 !(sk->sk_userlocks & SOCK_BINDPORT_LOCK))
1967 if (oldstate == TCP_ESTABLISHED)
1968 TCP_DEC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
1971 /* Change state AFTER socket is unhashed to avoid closed
1972 * socket sitting in hash tables.
1974 sk->sk_state = state;
1977 SOCK_DEBUG(sk, "TCP sk=%p, State %s -> %s\n", sk, statename[oldstate], statename[state]);
1980 EXPORT_SYMBOL_GPL(tcp_set_state);
1983 * State processing on a close. This implements the state shift for
1984 * sending our FIN frame. Note that we only send a FIN for some
1985 * states. A shutdown() may have already sent the FIN, or we may be
1989 static const unsigned char new_state[16] = {
1990 /* current state: new state: action: */
1991 /* (Invalid) */ TCP_CLOSE,
1992 /* TCP_ESTABLISHED */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1993 /* TCP_SYN_SENT */ TCP_CLOSE,
1994 /* TCP_SYN_RECV */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1995 /* TCP_FIN_WAIT1 */ TCP_FIN_WAIT1,
1996 /* TCP_FIN_WAIT2 */ TCP_FIN_WAIT2,
1997 /* TCP_TIME_WAIT */ TCP_CLOSE,
1998 /* TCP_CLOSE */ TCP_CLOSE,
1999 /* TCP_CLOSE_WAIT */ TCP_LAST_ACK | TCP_ACTION_FIN,
2000 /* TCP_LAST_ACK */ TCP_LAST_ACK,
2001 /* TCP_LISTEN */ TCP_CLOSE,
2002 /* TCP_CLOSING */ TCP_CLOSING,
2005 static int tcp_close_state(struct sock *sk)
2007 int next = (int)new_state[sk->sk_state];
2008 int ns = next & TCP_STATE_MASK;
2010 tcp_set_state(sk, ns);
2012 return next & TCP_ACTION_FIN;
2016 * Shutdown the sending side of a connection. Much like close except
2017 * that we don't receive shut down or sock_set_flag(sk, SOCK_DEAD).
2020 void tcp_shutdown(struct sock *sk, int how)
2022 /* We need to grab some memory, and put together a FIN,
2023 * and then put it into the queue to be sent.
2024 * Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
2026 if (!(how & SEND_SHUTDOWN))
2029 /* If we've already sent a FIN, or it's a closed state, skip this. */
2030 if ((1 << sk->sk_state) &
2031 (TCPF_ESTABLISHED | TCPF_SYN_SENT |
2032 TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) {
2033 /* Clear out any half completed packets. FIN if needed. */
2034 if (tcp_close_state(sk))
2038 EXPORT_SYMBOL(tcp_shutdown);
2040 bool tcp_check_oom(struct sock *sk, int shift)
2042 bool too_many_orphans, out_of_socket_memory;
2044 too_many_orphans = tcp_too_many_orphans(sk, shift);
2045 out_of_socket_memory = tcp_out_of_memory(sk);
2047 if (too_many_orphans)
2048 net_info_ratelimited("too many orphaned sockets\n");
2049 if (out_of_socket_memory)
2050 net_info_ratelimited("out of memory -- consider tuning tcp_mem\n");
2051 return too_many_orphans || out_of_socket_memory;
2054 void tcp_close(struct sock *sk, long timeout)
2056 struct sk_buff *skb;
2057 int data_was_unread = 0;
2061 sk->sk_shutdown = SHUTDOWN_MASK;
2063 if (sk->sk_state == TCP_LISTEN) {
2064 tcp_set_state(sk, TCP_CLOSE);
2067 inet_csk_listen_stop(sk);
2069 goto adjudge_to_death;
2072 /* We need to flush the recv. buffs. We do this only on the
2073 * descriptor close, not protocol-sourced closes, because the
2074 * reader process may not have drained the data yet!
2076 while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) {
2077 u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq -
2079 data_was_unread += len;
2085 /* If socket has been already reset (e.g. in tcp_reset()) - kill it. */
2086 if (sk->sk_state == TCP_CLOSE)
2087 goto adjudge_to_death;
2089 /* As outlined in RFC 2525, section 2.17, we send a RST here because
2090 * data was lost. To witness the awful effects of the old behavior of
2091 * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk
2092 * GET in an FTP client, suspend the process, wait for the client to
2093 * advertise a zero window, then kill -9 the FTP client, wheee...
2094 * Note: timeout is always zero in such a case.
2096 if (unlikely(tcp_sk(sk)->repair)) {
2097 sk->sk_prot->disconnect(sk, 0);
2098 } else if (data_was_unread) {
2099 /* Unread data was tossed, zap the connection. */
2100 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONCLOSE);
2101 tcp_set_state(sk, TCP_CLOSE);
2102 tcp_send_active_reset(sk, sk->sk_allocation);
2103 } else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) {
2104 /* Check zero linger _after_ checking for unread data. */
2105 sk->sk_prot->disconnect(sk, 0);
2106 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
2107 } else if (tcp_close_state(sk)) {
2108 /* We FIN if the application ate all the data before
2109 * zapping the connection.
2112 /* RED-PEN. Formally speaking, we have broken TCP state
2113 * machine. State transitions:
2115 * TCP_ESTABLISHED -> TCP_FIN_WAIT1
2116 * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible)
2117 * TCP_CLOSE_WAIT -> TCP_LAST_ACK
2119 * are legal only when FIN has been sent (i.e. in window),
2120 * rather than queued out of window. Purists blame.
2122 * F.e. "RFC state" is ESTABLISHED,
2123 * if Linux state is FIN-WAIT-1, but FIN is still not sent.
2125 * The visible declinations are that sometimes
2126 * we enter time-wait state, when it is not required really
2127 * (harmless), do not send active resets, when they are
2128 * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
2129 * they look as CLOSING or LAST_ACK for Linux)
2130 * Probably, I missed some more holelets.
2132 * XXX (TFO) - To start off we don't support SYN+ACK+FIN
2133 * in a single packet! (May consider it later but will
2134 * probably need API support or TCP_CORK SYN-ACK until
2135 * data is written and socket is closed.)
2140 sk_stream_wait_close(sk, timeout);
2143 state = sk->sk_state;
2147 /* It is the last release_sock in its life. It will remove backlog. */
2151 /* Now socket is owned by kernel and we acquire BH lock
2152 to finish close. No need to check for user refs.
2156 WARN_ON(sock_owned_by_user(sk));
2158 percpu_counter_inc(sk->sk_prot->orphan_count);
2160 /* Have we already been destroyed by a softirq or backlog? */
2161 if (state != TCP_CLOSE && sk->sk_state == TCP_CLOSE)
2164 /* This is a (useful) BSD violating of the RFC. There is a
2165 * problem with TCP as specified in that the other end could
2166 * keep a socket open forever with no application left this end.
2167 * We use a 3 minute timeout (about the same as BSD) then kill
2168 * our end. If they send after that then tough - BUT: long enough
2169 * that we won't make the old 4*rto = almost no time - whoops
2172 * Nope, it was not mistake. It is really desired behaviour
2173 * f.e. on http servers, when such sockets are useless, but
2174 * consume significant resources. Let's do it with special
2175 * linger2 option. --ANK
2178 if (sk->sk_state == TCP_FIN_WAIT2) {
2179 struct tcp_sock *tp = tcp_sk(sk);
2180 if (tp->linger2 < 0) {
2181 tcp_set_state(sk, TCP_CLOSE);
2182 tcp_send_active_reset(sk, GFP_ATOMIC);
2183 NET_INC_STATS_BH(sock_net(sk),
2184 LINUX_MIB_TCPABORTONLINGER);
2186 const int tmo = tcp_fin_time(sk);
2188 if (tmo > TCP_TIMEWAIT_LEN) {
2189 inet_csk_reset_keepalive_timer(sk,
2190 tmo - TCP_TIMEWAIT_LEN);
2192 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
2197 if (sk->sk_state != TCP_CLOSE) {
2199 if (tcp_check_oom(sk, 0)) {
2200 tcp_set_state(sk, TCP_CLOSE);
2201 tcp_send_active_reset(sk, GFP_ATOMIC);
2202 NET_INC_STATS_BH(sock_net(sk),
2203 LINUX_MIB_TCPABORTONMEMORY);
2207 if (sk->sk_state == TCP_CLOSE) {
2208 struct request_sock *req = tcp_sk(sk)->fastopen_rsk;
2209 /* We could get here with a non-NULL req if the socket is
2210 * aborted (e.g., closed with unread data) before 3WHS
2214 reqsk_fastopen_remove(sk, req, false);
2215 inet_csk_destroy_sock(sk);
2217 /* Otherwise, socket is reprieved until protocol close. */
2224 EXPORT_SYMBOL(tcp_close);
2226 /* These states need RST on ABORT according to RFC793 */
2228 static inline bool tcp_need_reset(int state)
2230 return (1 << state) &
2231 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 |
2232 TCPF_FIN_WAIT2 | TCPF_SYN_RECV);
2235 int tcp_disconnect(struct sock *sk, int flags)
2237 struct inet_sock *inet = inet_sk(sk);
2238 struct inet_connection_sock *icsk = inet_csk(sk);
2239 struct tcp_sock *tp = tcp_sk(sk);
2241 int old_state = sk->sk_state;
2243 if (old_state != TCP_CLOSE)
2244 tcp_set_state(sk, TCP_CLOSE);
2246 /* ABORT function of RFC793 */
2247 if (old_state == TCP_LISTEN) {
2248 inet_csk_listen_stop(sk);
2249 } else if (unlikely(tp->repair)) {
2250 sk->sk_err = ECONNABORTED;
2251 } else if (tcp_need_reset(old_state) ||
2252 (tp->snd_nxt != tp->write_seq &&
2253 (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK))) {
2254 /* The last check adjusts for discrepancy of Linux wrt. RFC
2257 tcp_send_active_reset(sk, gfp_any());
2258 sk->sk_err = ECONNRESET;
2259 } else if (old_state == TCP_SYN_SENT)
2260 sk->sk_err = ECONNRESET;
2262 tcp_clear_xmit_timers(sk);
2263 __skb_queue_purge(&sk->sk_receive_queue);
2264 tcp_write_queue_purge(sk);
2265 __skb_queue_purge(&tp->out_of_order_queue);
2266 #ifdef CONFIG_NET_DMA
2267 __skb_queue_purge(&sk->sk_async_wait_queue);
2270 inet->inet_dport = 0;
2272 if (!(sk->sk_userlocks & SOCK_BINDADDR_LOCK))
2273 inet_reset_saddr(sk);
2275 sk->sk_shutdown = 0;
2276 sock_reset_flag(sk, SOCK_DONE);
2278 if ((tp->write_seq += tp->max_window + 2) == 0)
2280 icsk->icsk_backoff = 0;
2282 icsk->icsk_probes_out = 0;
2283 tp->packets_out = 0;
2284 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
2285 tp->snd_cwnd_cnt = 0;
2286 tp->window_clamp = 0;
2287 tcp_set_ca_state(sk, TCP_CA_Open);
2288 tcp_clear_retrans(tp);
2289 inet_csk_delack_init(sk);
2290 tcp_init_send_head(sk);
2291 memset(&tp->rx_opt, 0, sizeof(tp->rx_opt));
2294 WARN_ON(inet->inet_num && !icsk->icsk_bind_hash);
2296 sk->sk_error_report(sk);
2299 EXPORT_SYMBOL(tcp_disconnect);
2301 void tcp_sock_destruct(struct sock *sk)
2303 inet_sock_destruct(sk);
2305 kfree(inet_csk(sk)->icsk_accept_queue.fastopenq);
2308 static inline bool tcp_can_repair_sock(const struct sock *sk)
2310 return ns_capable(sock_net(sk)->user_ns, CAP_NET_ADMIN) &&
2311 ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_ESTABLISHED));
2314 static int tcp_repair_options_est(struct tcp_sock *tp,
2315 struct tcp_repair_opt __user *optbuf, unsigned int len)
2317 struct tcp_repair_opt opt;
2319 while (len >= sizeof(opt)) {
2320 if (copy_from_user(&opt, optbuf, sizeof(opt)))
2326 switch (opt.opt_code) {
2328 tp->rx_opt.mss_clamp = opt.opt_val;
2332 u16 snd_wscale = opt.opt_val & 0xFFFF;
2333 u16 rcv_wscale = opt.opt_val >> 16;
2335 if (snd_wscale > 14 || rcv_wscale > 14)
2338 tp->rx_opt.snd_wscale = snd_wscale;
2339 tp->rx_opt.rcv_wscale = rcv_wscale;
2340 tp->rx_opt.wscale_ok = 1;
2343 case TCPOPT_SACK_PERM:
2344 if (opt.opt_val != 0)
2347 tp->rx_opt.sack_ok |= TCP_SACK_SEEN;
2348 if (sysctl_tcp_fack)
2349 tcp_enable_fack(tp);
2351 case TCPOPT_TIMESTAMP:
2352 if (opt.opt_val != 0)
2355 tp->rx_opt.tstamp_ok = 1;
2364 * Socket option code for TCP.
2366 static int do_tcp_setsockopt(struct sock *sk, int level,
2367 int optname, char __user *optval, unsigned int optlen)
2369 struct tcp_sock *tp = tcp_sk(sk);
2370 struct inet_connection_sock *icsk = inet_csk(sk);
2374 /* These are data/string values, all the others are ints */
2376 case TCP_CONGESTION: {
2377 char name[TCP_CA_NAME_MAX];
2382 val = strncpy_from_user(name, optval,
2383 min_t(long, TCP_CA_NAME_MAX-1, optlen));
2389 err = tcp_set_congestion_control(sk, name);
2398 if (optlen < sizeof(int))
2401 if (get_user(val, (int __user *)optval))
2408 /* Values greater than interface MTU won't take effect. However
2409 * at the point when this call is done we typically don't yet
2410 * know which interface is going to be used */
2411 if (val < TCP_MIN_MSS || val > MAX_TCP_WINDOW) {
2415 tp->rx_opt.user_mss = val;
2420 /* TCP_NODELAY is weaker than TCP_CORK, so that
2421 * this option on corked socket is remembered, but
2422 * it is not activated until cork is cleared.
2424 * However, when TCP_NODELAY is set we make
2425 * an explicit push, which overrides even TCP_CORK
2426 * for currently queued segments.
2428 tp->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH;
2429 tcp_push_pending_frames(sk);
2431 tp->nonagle &= ~TCP_NAGLE_OFF;
2435 case TCP_THIN_LINEAR_TIMEOUTS:
2436 if (val < 0 || val > 1)
2442 case TCP_THIN_DUPACK:
2443 if (val < 0 || val > 1)
2446 tp->thin_dupack = val;
2447 if (tp->thin_dupack)
2448 tcp_disable_early_retrans(tp);
2452 if (!tcp_can_repair_sock(sk))
2454 else if (val == 1) {
2456 sk->sk_reuse = SK_FORCE_REUSE;
2457 tp->repair_queue = TCP_NO_QUEUE;
2458 } else if (val == 0) {
2460 sk->sk_reuse = SK_NO_REUSE;
2461 tcp_send_window_probe(sk);
2467 case TCP_REPAIR_QUEUE:
2470 else if (val < TCP_QUEUES_NR)
2471 tp->repair_queue = val;
2477 if (sk->sk_state != TCP_CLOSE)
2479 else if (tp->repair_queue == TCP_SEND_QUEUE)
2480 tp->write_seq = val;
2481 else if (tp->repair_queue == TCP_RECV_QUEUE)
2487 case TCP_REPAIR_OPTIONS:
2490 else if (sk->sk_state == TCP_ESTABLISHED)
2491 err = tcp_repair_options_est(tp,
2492 (struct tcp_repair_opt __user *)optval,
2499 /* When set indicates to always queue non-full frames.
2500 * Later the user clears this option and we transmit
2501 * any pending partial frames in the queue. This is
2502 * meant to be used alongside sendfile() to get properly
2503 * filled frames when the user (for example) must write
2504 * out headers with a write() call first and then use
2505 * sendfile to send out the data parts.
2507 * TCP_CORK can be set together with TCP_NODELAY and it is
2508 * stronger than TCP_NODELAY.
2511 tp->nonagle |= TCP_NAGLE_CORK;
2513 tp->nonagle &= ~TCP_NAGLE_CORK;
2514 if (tp->nonagle&TCP_NAGLE_OFF)
2515 tp->nonagle |= TCP_NAGLE_PUSH;
2516 tcp_push_pending_frames(sk);
2521 if (val < 1 || val > MAX_TCP_KEEPIDLE)
2524 tp->keepalive_time = val * HZ;
2525 if (sock_flag(sk, SOCK_KEEPOPEN) &&
2526 !((1 << sk->sk_state) &
2527 (TCPF_CLOSE | TCPF_LISTEN))) {
2528 u32 elapsed = keepalive_time_elapsed(tp);
2529 if (tp->keepalive_time > elapsed)
2530 elapsed = tp->keepalive_time - elapsed;
2533 inet_csk_reset_keepalive_timer(sk, elapsed);
2538 if (val < 1 || val > MAX_TCP_KEEPINTVL)
2541 tp->keepalive_intvl = val * HZ;
2544 if (val < 1 || val > MAX_TCP_KEEPCNT)
2547 tp->keepalive_probes = val;
2550 if (val < 1 || val > MAX_TCP_SYNCNT)
2553 icsk->icsk_syn_retries = val;
2559 else if (val > sysctl_tcp_fin_timeout / HZ)
2562 tp->linger2 = val * HZ;
2565 case TCP_DEFER_ACCEPT:
2566 /* Translate value in seconds to number of retransmits */
2567 icsk->icsk_accept_queue.rskq_defer_accept =
2568 secs_to_retrans(val, TCP_TIMEOUT_INIT / HZ,
2572 case TCP_WINDOW_CLAMP:
2574 if (sk->sk_state != TCP_CLOSE) {
2578 tp->window_clamp = 0;
2580 tp->window_clamp = val < SOCK_MIN_RCVBUF / 2 ?
2581 SOCK_MIN_RCVBUF / 2 : val;
2586 icsk->icsk_ack.pingpong = 1;
2588 icsk->icsk_ack.pingpong = 0;
2589 if ((1 << sk->sk_state) &
2590 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) &&
2591 inet_csk_ack_scheduled(sk)) {
2592 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
2593 tcp_cleanup_rbuf(sk, 1);
2595 icsk->icsk_ack.pingpong = 1;
2600 #ifdef CONFIG_TCP_MD5SIG
2602 /* Read the IP->Key mappings from userspace */
2603 err = tp->af_specific->md5_parse(sk, optval, optlen);
2606 case TCP_USER_TIMEOUT:
2607 /* Cap the max timeout in ms TCP will retry/retrans
2608 * before giving up and aborting (ETIMEDOUT) a connection.
2613 icsk->icsk_user_timeout = msecs_to_jiffies(val);
2617 if (val >= 0 && ((1 << sk->sk_state) & (TCPF_CLOSE |
2619 err = fastopen_init_queue(sk, val);
2627 tp->tsoffset = val - tcp_time_stamp;
2638 int tcp_setsockopt(struct sock *sk, int level, int optname, char __user *optval,
2639 unsigned int optlen)
2641 const struct inet_connection_sock *icsk = inet_csk(sk);
2643 if (level != SOL_TCP)
2644 return icsk->icsk_af_ops->setsockopt(sk, level, optname,
2646 return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2648 EXPORT_SYMBOL(tcp_setsockopt);
2650 #ifdef CONFIG_COMPAT
2651 int compat_tcp_setsockopt(struct sock *sk, int level, int optname,
2652 char __user *optval, unsigned int optlen)
2654 if (level != SOL_TCP)
2655 return inet_csk_compat_setsockopt(sk, level, optname,
2657 return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2659 EXPORT_SYMBOL(compat_tcp_setsockopt);
2662 /* Return information about state of tcp endpoint in API format. */
2663 void tcp_get_info(const struct sock *sk, struct tcp_info *info)
2665 const struct tcp_sock *tp = tcp_sk(sk);
2666 const struct inet_connection_sock *icsk = inet_csk(sk);
2667 u32 now = tcp_time_stamp;
2669 memset(info, 0, sizeof(*info));
2671 info->tcpi_state = sk->sk_state;
2672 info->tcpi_ca_state = icsk->icsk_ca_state;
2673 info->tcpi_retransmits = icsk->icsk_retransmits;
2674 info->tcpi_probes = icsk->icsk_probes_out;
2675 info->tcpi_backoff = icsk->icsk_backoff;
2677 if (tp->rx_opt.tstamp_ok)
2678 info->tcpi_options |= TCPI_OPT_TIMESTAMPS;
2679 if (tcp_is_sack(tp))
2680 info->tcpi_options |= TCPI_OPT_SACK;
2681 if (tp->rx_opt.wscale_ok) {
2682 info->tcpi_options |= TCPI_OPT_WSCALE;
2683 info->tcpi_snd_wscale = tp->rx_opt.snd_wscale;
2684 info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale;
2687 if (tp->ecn_flags & TCP_ECN_OK)
2688 info->tcpi_options |= TCPI_OPT_ECN;
2689 if (tp->ecn_flags & TCP_ECN_SEEN)
2690 info->tcpi_options |= TCPI_OPT_ECN_SEEN;
2691 if (tp->syn_data_acked)
2692 info->tcpi_options |= TCPI_OPT_SYN_DATA;
2694 info->tcpi_rto = jiffies_to_usecs(icsk->icsk_rto);
2695 info->tcpi_ato = jiffies_to_usecs(icsk->icsk_ack.ato);
2696 info->tcpi_snd_mss = tp->mss_cache;
2697 info->tcpi_rcv_mss = icsk->icsk_ack.rcv_mss;
2699 if (sk->sk_state == TCP_LISTEN) {
2700 info->tcpi_unacked = sk->sk_ack_backlog;
2701 info->tcpi_sacked = sk->sk_max_ack_backlog;
2703 info->tcpi_unacked = tp->packets_out;
2704 info->tcpi_sacked = tp->sacked_out;
2706 info->tcpi_lost = tp->lost_out;
2707 info->tcpi_retrans = tp->retrans_out;
2708 info->tcpi_fackets = tp->fackets_out;
2710 info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime);
2711 info->tcpi_last_data_recv = jiffies_to_msecs(now - icsk->icsk_ack.lrcvtime);
2712 info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp);
2714 info->tcpi_pmtu = icsk->icsk_pmtu_cookie;
2715 info->tcpi_rcv_ssthresh = tp->rcv_ssthresh;
2716 info->tcpi_rtt = jiffies_to_usecs(tp->srtt)>>3;
2717 info->tcpi_rttvar = jiffies_to_usecs(tp->mdev)>>2;
2718 info->tcpi_snd_ssthresh = tp->snd_ssthresh;
2719 info->tcpi_snd_cwnd = tp->snd_cwnd;
2720 info->tcpi_advmss = tp->advmss;
2721 info->tcpi_reordering = tp->reordering;
2723 info->tcpi_rcv_rtt = jiffies_to_usecs(tp->rcv_rtt_est.rtt)>>3;
2724 info->tcpi_rcv_space = tp->rcvq_space.space;
2726 info->tcpi_total_retrans = tp->total_retrans;
2728 EXPORT_SYMBOL_GPL(tcp_get_info);
2730 static int do_tcp_getsockopt(struct sock *sk, int level,
2731 int optname, char __user *optval, int __user *optlen)
2733 struct inet_connection_sock *icsk = inet_csk(sk);
2734 struct tcp_sock *tp = tcp_sk(sk);
2737 if (get_user(len, optlen))
2740 len = min_t(unsigned int, len, sizeof(int));
2747 val = tp->mss_cache;
2748 if (!val && ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
2749 val = tp->rx_opt.user_mss;
2751 val = tp->rx_opt.mss_clamp;
2754 val = !!(tp->nonagle&TCP_NAGLE_OFF);
2757 val = !!(tp->nonagle&TCP_NAGLE_CORK);
2760 val = keepalive_time_when(tp) / HZ;
2763 val = keepalive_intvl_when(tp) / HZ;
2766 val = keepalive_probes(tp);
2769 val = icsk->icsk_syn_retries ? : sysctl_tcp_syn_retries;
2774 val = (val ? : sysctl_tcp_fin_timeout) / HZ;
2776 case TCP_DEFER_ACCEPT:
2777 val = retrans_to_secs(icsk->icsk_accept_queue.rskq_defer_accept,
2778 TCP_TIMEOUT_INIT / HZ, TCP_RTO_MAX / HZ);
2780 case TCP_WINDOW_CLAMP:
2781 val = tp->window_clamp;
2784 struct tcp_info info;
2786 if (get_user(len, optlen))
2789 tcp_get_info(sk, &info);
2791 len = min_t(unsigned int, len, sizeof(info));
2792 if (put_user(len, optlen))
2794 if (copy_to_user(optval, &info, len))
2799 val = !icsk->icsk_ack.pingpong;
2802 case TCP_CONGESTION:
2803 if (get_user(len, optlen))
2805 len = min_t(unsigned int, len, TCP_CA_NAME_MAX);
2806 if (put_user(len, optlen))
2808 if (copy_to_user(optval, icsk->icsk_ca_ops->name, len))
2812 case TCP_THIN_LINEAR_TIMEOUTS:
2815 case TCP_THIN_DUPACK:
2816 val = tp->thin_dupack;
2823 case TCP_REPAIR_QUEUE:
2825 val = tp->repair_queue;
2831 if (tp->repair_queue == TCP_SEND_QUEUE)
2832 val = tp->write_seq;
2833 else if (tp->repair_queue == TCP_RECV_QUEUE)
2839 case TCP_USER_TIMEOUT:
2840 val = jiffies_to_msecs(icsk->icsk_user_timeout);
2843 val = tcp_time_stamp + tp->tsoffset;
2846 return -ENOPROTOOPT;
2849 if (put_user(len, optlen))
2851 if (copy_to_user(optval, &val, len))
2856 int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval,
2859 struct inet_connection_sock *icsk = inet_csk(sk);
2861 if (level != SOL_TCP)
2862 return icsk->icsk_af_ops->getsockopt(sk, level, optname,
2864 return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2866 EXPORT_SYMBOL(tcp_getsockopt);
2868 #ifdef CONFIG_COMPAT
2869 int compat_tcp_getsockopt(struct sock *sk, int level, int optname,
2870 char __user *optval, int __user *optlen)
2872 if (level != SOL_TCP)
2873 return inet_csk_compat_getsockopt(sk, level, optname,
2875 return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2877 EXPORT_SYMBOL(compat_tcp_getsockopt);
2880 struct sk_buff *tcp_tso_segment(struct sk_buff *skb,
2881 netdev_features_t features)
2883 struct sk_buff *segs = ERR_PTR(-EINVAL);
2888 unsigned int oldlen;
2890 struct sk_buff *gso_skb = skb;
2892 bool ooo_okay, copy_destructor;
2894 if (!pskb_may_pull(skb, sizeof(*th)))
2898 thlen = th->doff * 4;
2899 if (thlen < sizeof(*th))
2902 if (!pskb_may_pull(skb, thlen))
2905 oldlen = (u16)~skb->len;
2906 __skb_pull(skb, thlen);
2908 mss = skb_shinfo(skb)->gso_size;
2909 if (unlikely(skb->len <= mss))
2912 if (skb_gso_ok(skb, features | NETIF_F_GSO_ROBUST)) {
2913 /* Packet is from an untrusted source, reset gso_segs. */
2914 int type = skb_shinfo(skb)->gso_type;
2923 SKB_GSO_UDP_TUNNEL |
2925 !(type & (SKB_GSO_TCPV4 | SKB_GSO_TCPV6))))
2928 skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss);
2934 copy_destructor = gso_skb->destructor == tcp_wfree;
2935 ooo_okay = gso_skb->ooo_okay;
2936 /* All segments but the first should have ooo_okay cleared */
2939 segs = skb_segment(skb, features);
2943 /* Only first segment might have ooo_okay set */
2944 segs->ooo_okay = ooo_okay;
2946 delta = htonl(oldlen + (thlen + mss));
2950 seq = ntohl(th->seq);
2952 newcheck = ~csum_fold((__force __wsum)((__force u32)th->check +
2953 (__force u32)delta));
2956 th->fin = th->psh = 0;
2957 th->check = newcheck;
2959 if (skb->ip_summed != CHECKSUM_PARTIAL)
2961 csum_fold(csum_partial(skb_transport_header(skb),
2965 if (copy_destructor) {
2966 skb->destructor = gso_skb->destructor;
2967 skb->sk = gso_skb->sk;
2968 /* {tcp|sock}_wfree() use exact truesize accounting :
2969 * sum(skb->truesize) MUST be exactly be gso_skb->truesize
2970 * So we account mss bytes of 'true size' for each segment.
2971 * The last segment will contain the remaining.
2973 skb->truesize = mss;
2974 gso_skb->truesize -= mss;
2979 th->seq = htonl(seq);
2981 } while (skb->next);
2983 /* Following permits TCP Small Queues to work well with GSO :
2984 * The callback to TCP stack will be called at the time last frag
2985 * is freed at TX completion, and not right now when gso_skb
2986 * is freed by GSO engine
2988 if (copy_destructor) {
2989 swap(gso_skb->sk, skb->sk);
2990 swap(gso_skb->destructor, skb->destructor);
2991 swap(gso_skb->truesize, skb->truesize);
2994 delta = htonl(oldlen + (skb_tail_pointer(skb) -
2995 skb_transport_header(skb)) +
2997 th->check = ~csum_fold((__force __wsum)((__force u32)th->check +
2998 (__force u32)delta));
2999 if (skb->ip_summed != CHECKSUM_PARTIAL)
3000 th->check = csum_fold(csum_partial(skb_transport_header(skb),
3006 EXPORT_SYMBOL(tcp_tso_segment);
3008 struct sk_buff **tcp_gro_receive(struct sk_buff **head, struct sk_buff *skb)
3010 struct sk_buff **pp = NULL;
3017 unsigned int mss = 1;
3023 off = skb_gro_offset(skb);
3024 hlen = off + sizeof(*th);
3025 th = skb_gro_header_fast(skb, off);
3026 if (skb_gro_header_hard(skb, hlen)) {
3027 th = skb_gro_header_slow(skb, hlen, off);
3032 thlen = th->doff * 4;
3033 if (thlen < sizeof(*th))
3037 if (skb_gro_header_hard(skb, hlen)) {
3038 th = skb_gro_header_slow(skb, hlen, off);
3043 skb_gro_pull(skb, thlen);
3045 len = skb_gro_len(skb);
3046 flags = tcp_flag_word(th);
3048 for (; (p = *head); head = &p->next) {
3049 if (!NAPI_GRO_CB(p)->same_flow)
3054 if (*(u32 *)&th->source ^ *(u32 *)&th2->source) {
3055 NAPI_GRO_CB(p)->same_flow = 0;
3062 goto out_check_final;
3065 flush = NAPI_GRO_CB(p)->flush;
3066 flush |= (__force int)(flags & TCP_FLAG_CWR);
3067 flush |= (__force int)((flags ^ tcp_flag_word(th2)) &
3068 ~(TCP_FLAG_CWR | TCP_FLAG_FIN | TCP_FLAG_PSH));
3069 flush |= (__force int)(th->ack_seq ^ th2->ack_seq);
3070 for (i = sizeof(*th); i < thlen; i += 4)
3071 flush |= *(u32 *)((u8 *)th + i) ^
3072 *(u32 *)((u8 *)th2 + i);
3074 mss = skb_shinfo(p)->gso_size;
3076 flush |= (len - 1) >= mss;
3077 flush |= (ntohl(th2->seq) + skb_gro_len(p)) ^ ntohl(th->seq);
3079 if (flush || skb_gro_receive(head, skb)) {
3081 goto out_check_final;
3086 tcp_flag_word(th2) |= flags & (TCP_FLAG_FIN | TCP_FLAG_PSH);
3090 flush |= (__force int)(flags & (TCP_FLAG_URG | TCP_FLAG_PSH |
3091 TCP_FLAG_RST | TCP_FLAG_SYN |
3094 if (p && (!NAPI_GRO_CB(skb)->same_flow || flush))
3098 NAPI_GRO_CB(skb)->flush |= flush;
3102 EXPORT_SYMBOL(tcp_gro_receive);
3104 int tcp_gro_complete(struct sk_buff *skb)
3106 struct tcphdr *th = tcp_hdr(skb);
3108 skb->csum_start = skb_transport_header(skb) - skb->head;
3109 skb->csum_offset = offsetof(struct tcphdr, check);
3110 skb->ip_summed = CHECKSUM_PARTIAL;
3112 skb_shinfo(skb)->gso_segs = NAPI_GRO_CB(skb)->count;
3115 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
3119 EXPORT_SYMBOL(tcp_gro_complete);
3121 #ifdef CONFIG_TCP_MD5SIG
3122 static struct tcp_md5sig_pool __percpu *tcp_md5sig_pool __read_mostly;
3123 static DEFINE_MUTEX(tcp_md5sig_mutex);
3125 static void __tcp_free_md5sig_pool(struct tcp_md5sig_pool __percpu *pool)
3129 for_each_possible_cpu(cpu) {
3130 struct tcp_md5sig_pool *p = per_cpu_ptr(pool, cpu);
3132 if (p->md5_desc.tfm)
3133 crypto_free_hash(p->md5_desc.tfm);
3138 static void __tcp_alloc_md5sig_pool(void)
3141 struct tcp_md5sig_pool __percpu *pool;
3143 pool = alloc_percpu(struct tcp_md5sig_pool);
3147 for_each_possible_cpu(cpu) {
3148 struct crypto_hash *hash;
3150 hash = crypto_alloc_hash("md5", 0, CRYPTO_ALG_ASYNC);
3151 if (IS_ERR_OR_NULL(hash))
3154 per_cpu_ptr(pool, cpu)->md5_desc.tfm = hash;
3156 /* before setting tcp_md5sig_pool, we must commit all writes
3157 * to memory. See ACCESS_ONCE() in tcp_get_md5sig_pool()
3160 tcp_md5sig_pool = pool;
3163 __tcp_free_md5sig_pool(pool);
3166 bool tcp_alloc_md5sig_pool(void)
3168 if (unlikely(!tcp_md5sig_pool)) {
3169 mutex_lock(&tcp_md5sig_mutex);
3171 if (!tcp_md5sig_pool)
3172 __tcp_alloc_md5sig_pool();
3174 mutex_unlock(&tcp_md5sig_mutex);
3176 return tcp_md5sig_pool != NULL;
3178 EXPORT_SYMBOL(tcp_alloc_md5sig_pool);
3182 * tcp_get_md5sig_pool - get md5sig_pool for this user
3184 * We use percpu structure, so if we succeed, we exit with preemption
3185 * and BH disabled, to make sure another thread or softirq handling
3186 * wont try to get same context.
3188 struct tcp_md5sig_pool *tcp_get_md5sig_pool(void)
3190 struct tcp_md5sig_pool __percpu *p;
3193 p = ACCESS_ONCE(tcp_md5sig_pool);
3195 return __this_cpu_ptr(p);
3200 EXPORT_SYMBOL(tcp_get_md5sig_pool);
3202 int tcp_md5_hash_header(struct tcp_md5sig_pool *hp,
3203 const struct tcphdr *th)
3205 struct scatterlist sg;
3209 /* We are not allowed to change tcphdr, make a local copy */
3210 memcpy(&hdr, th, sizeof(hdr));
3213 /* options aren't included in the hash */
3214 sg_init_one(&sg, &hdr, sizeof(hdr));
3215 err = crypto_hash_update(&hp->md5_desc, &sg, sizeof(hdr));
3218 EXPORT_SYMBOL(tcp_md5_hash_header);
3220 int tcp_md5_hash_skb_data(struct tcp_md5sig_pool *hp,
3221 const struct sk_buff *skb, unsigned int header_len)
3223 struct scatterlist sg;
3224 const struct tcphdr *tp = tcp_hdr(skb);
3225 struct hash_desc *desc = &hp->md5_desc;
3227 const unsigned int head_data_len = skb_headlen(skb) > header_len ?
3228 skb_headlen(skb) - header_len : 0;
3229 const struct skb_shared_info *shi = skb_shinfo(skb);
3230 struct sk_buff *frag_iter;
3232 sg_init_table(&sg, 1);
3234 sg_set_buf(&sg, ((u8 *) tp) + header_len, head_data_len);
3235 if (crypto_hash_update(desc, &sg, head_data_len))
3238 for (i = 0; i < shi->nr_frags; ++i) {
3239 const struct skb_frag_struct *f = &shi->frags[i];
3240 unsigned int offset = f->page_offset;
3241 struct page *page = skb_frag_page(f) + (offset >> PAGE_SHIFT);
3243 sg_set_page(&sg, page, skb_frag_size(f),
3244 offset_in_page(offset));
3245 if (crypto_hash_update(desc, &sg, skb_frag_size(f)))
3249 skb_walk_frags(skb, frag_iter)
3250 if (tcp_md5_hash_skb_data(hp, frag_iter, 0))
3255 EXPORT_SYMBOL(tcp_md5_hash_skb_data);
3257 int tcp_md5_hash_key(struct tcp_md5sig_pool *hp, const struct tcp_md5sig_key *key)
3259 struct scatterlist sg;
3261 sg_init_one(&sg, key->key, key->keylen);
3262 return crypto_hash_update(&hp->md5_desc, &sg, key->keylen);
3264 EXPORT_SYMBOL(tcp_md5_hash_key);
3268 void tcp_done(struct sock *sk)
3270 struct request_sock *req = tcp_sk(sk)->fastopen_rsk;
3272 if (sk->sk_state == TCP_SYN_SENT || sk->sk_state == TCP_SYN_RECV)
3273 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_ATTEMPTFAILS);
3275 tcp_set_state(sk, TCP_CLOSE);
3276 tcp_clear_xmit_timers(sk);
3278 reqsk_fastopen_remove(sk, req, false);
3280 sk->sk_shutdown = SHUTDOWN_MASK;
3282 if (!sock_flag(sk, SOCK_DEAD))
3283 sk->sk_state_change(sk);
3285 inet_csk_destroy_sock(sk);
3287 EXPORT_SYMBOL_GPL(tcp_done);
3289 extern struct tcp_congestion_ops tcp_reno;
3291 static __initdata unsigned long thash_entries;
3292 static int __init set_thash_entries(char *str)
3299 ret = kstrtoul(str, 0, &thash_entries);
3305 __setup("thash_entries=", set_thash_entries);
3307 void tcp_init_mem(struct net *net)
3309 unsigned long limit = nr_free_buffer_pages() / 8;
3310 limit = max(limit, 128UL);
3311 net->ipv4.sysctl_tcp_mem[0] = limit / 4 * 3;
3312 net->ipv4.sysctl_tcp_mem[1] = limit;
3313 net->ipv4.sysctl_tcp_mem[2] = net->ipv4.sysctl_tcp_mem[0] * 2;
3316 void __init tcp_init(void)
3318 struct sk_buff *skb = NULL;
3319 unsigned long limit;
3320 int max_rshare, max_wshare, cnt;
3323 BUILD_BUG_ON(sizeof(struct tcp_skb_cb) > sizeof(skb->cb));
3325 percpu_counter_init(&tcp_sockets_allocated, 0);
3326 percpu_counter_init(&tcp_orphan_count, 0);
3327 tcp_hashinfo.bind_bucket_cachep =
3328 kmem_cache_create("tcp_bind_bucket",
3329 sizeof(struct inet_bind_bucket), 0,
3330 SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL);
3332 /* Size and allocate the main established and bind bucket
3335 * The methodology is similar to that of the buffer cache.
3337 tcp_hashinfo.ehash =
3338 alloc_large_system_hash("TCP established",
3339 sizeof(struct inet_ehash_bucket),
3341 17, /* one slot per 128 KB of memory */
3344 &tcp_hashinfo.ehash_mask,
3346 thash_entries ? 0 : 512 * 1024);
3347 for (i = 0; i <= tcp_hashinfo.ehash_mask; i++) {
3348 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].chain, i);
3349 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].twchain, i);
3351 if (inet_ehash_locks_alloc(&tcp_hashinfo))
3352 panic("TCP: failed to alloc ehash_locks");
3353 tcp_hashinfo.bhash =
3354 alloc_large_system_hash("TCP bind",
3355 sizeof(struct inet_bind_hashbucket),
3356 tcp_hashinfo.ehash_mask + 1,
3357 17, /* one slot per 128 KB of memory */
3359 &tcp_hashinfo.bhash_size,
3363 tcp_hashinfo.bhash_size = 1U << tcp_hashinfo.bhash_size;
3364 for (i = 0; i < tcp_hashinfo.bhash_size; i++) {
3365 spin_lock_init(&tcp_hashinfo.bhash[i].lock);
3366 INIT_HLIST_HEAD(&tcp_hashinfo.bhash[i].chain);
3370 cnt = tcp_hashinfo.ehash_mask + 1;
3372 tcp_death_row.sysctl_max_tw_buckets = cnt / 2;
3373 sysctl_tcp_max_orphans = cnt / 2;
3374 sysctl_max_syn_backlog = max(128, cnt / 256);
3376 tcp_init_mem(&init_net);
3377 /* Set per-socket limits to no more than 1/128 the pressure threshold */
3378 limit = nr_free_buffer_pages() << (PAGE_SHIFT - 7);
3379 max_wshare = min(4UL*1024*1024, limit);
3380 max_rshare = min(6UL*1024*1024, limit);
3382 sysctl_tcp_wmem[0] = SK_MEM_QUANTUM;
3383 sysctl_tcp_wmem[1] = 16*1024;
3384 sysctl_tcp_wmem[2] = max(64*1024, max_wshare);
3386 sysctl_tcp_rmem[0] = SK_MEM_QUANTUM;
3387 sysctl_tcp_rmem[1] = 87380;
3388 sysctl_tcp_rmem[2] = max(87380, max_rshare);
3390 pr_info("Hash tables configured (established %u bind %u)\n",
3391 tcp_hashinfo.ehash_mask + 1, tcp_hashinfo.bhash_size);
3395 tcp_register_congestion_control(&tcp_reno);