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 int sysctl_tcp_min_tso_segs __read_mostly = 2;
287 struct percpu_counter tcp_orphan_count;
288 EXPORT_SYMBOL_GPL(tcp_orphan_count);
290 int sysctl_tcp_wmem[3] __read_mostly;
291 int sysctl_tcp_rmem[3] __read_mostly;
293 EXPORT_SYMBOL(sysctl_tcp_rmem);
294 EXPORT_SYMBOL(sysctl_tcp_wmem);
296 atomic_long_t tcp_memory_allocated; /* Current allocated memory. */
297 EXPORT_SYMBOL(tcp_memory_allocated);
300 * Current number of TCP sockets.
302 struct percpu_counter tcp_sockets_allocated;
303 EXPORT_SYMBOL(tcp_sockets_allocated);
308 struct tcp_splice_state {
309 struct pipe_inode_info *pipe;
315 * Pressure flag: try to collapse.
316 * Technical note: it is used by multiple contexts non atomically.
317 * All the __sk_mem_schedule() is of this nature: accounting
318 * is strict, actions are advisory and have some latency.
320 int tcp_memory_pressure __read_mostly;
321 EXPORT_SYMBOL(tcp_memory_pressure);
323 void tcp_enter_memory_pressure(struct sock *sk)
325 if (!tcp_memory_pressure) {
326 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMEMORYPRESSURES);
327 tcp_memory_pressure = 1;
330 EXPORT_SYMBOL(tcp_enter_memory_pressure);
332 /* Convert seconds to retransmits based on initial and max timeout */
333 static u8 secs_to_retrans(int seconds, int timeout, int rto_max)
338 int period = timeout;
341 while (seconds > period && res < 255) {
344 if (timeout > rto_max)
352 /* Convert retransmits to seconds based on initial and max timeout */
353 static int retrans_to_secs(u8 retrans, int timeout, int rto_max)
361 if (timeout > rto_max)
369 /* Address-family independent initialization for a tcp_sock.
371 * NOTE: A lot of things set to zero explicitly by call to
372 * sk_alloc() so need not be done here.
374 void tcp_init_sock(struct sock *sk)
376 struct inet_connection_sock *icsk = inet_csk(sk);
377 struct tcp_sock *tp = tcp_sk(sk);
379 skb_queue_head_init(&tp->out_of_order_queue);
380 tcp_init_xmit_timers(sk);
381 tcp_prequeue_init(tp);
382 INIT_LIST_HEAD(&tp->tsq_node);
384 icsk->icsk_rto = TCP_TIMEOUT_INIT;
385 tp->mdev = TCP_TIMEOUT_INIT;
387 /* So many TCP implementations out there (incorrectly) count the
388 * initial SYN frame in their delayed-ACK and congestion control
389 * algorithms that we must have the following bandaid to talk
390 * efficiently to them. -DaveM
392 tp->snd_cwnd = TCP_INIT_CWND;
394 /* See draft-stevens-tcpca-spec-01 for discussion of the
395 * initialization of these values.
397 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
398 tp->snd_cwnd_clamp = ~0;
399 tp->mss_cache = TCP_MSS_DEFAULT;
401 tp->reordering = sysctl_tcp_reordering;
402 tcp_enable_early_retrans(tp);
403 icsk->icsk_ca_ops = &tcp_init_congestion_ops;
407 sk->sk_state = TCP_CLOSE;
409 sk->sk_write_space = sk_stream_write_space;
410 sock_set_flag(sk, SOCK_USE_WRITE_QUEUE);
412 icsk->icsk_sync_mss = tcp_sync_mss;
414 /* Presumed zeroed, in order of appearance:
415 * cookie_in_always, cookie_out_never,
416 * s_data_constant, s_data_in, s_data_out
418 sk->sk_sndbuf = sysctl_tcp_wmem[1];
419 sk->sk_rcvbuf = sysctl_tcp_rmem[1];
422 sock_update_memcg(sk);
423 sk_sockets_allocated_inc(sk);
426 EXPORT_SYMBOL(tcp_init_sock);
429 * Wait for a TCP event.
431 * Note that we don't need to lock the socket, as the upper poll layers
432 * take care of normal races (between the test and the event) and we don't
433 * go look at any of the socket buffers directly.
435 unsigned int tcp_poll(struct file *file, struct socket *sock, poll_table *wait)
438 struct sock *sk = sock->sk;
439 const struct tcp_sock *tp = tcp_sk(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)) {
793 /* Maybe we should/could use sk->sk_prot->max_header here ? */
794 hlen = inet_csk(sk)->icsk_af_ops->net_header_len +
795 inet_csk(sk)->icsk_ext_hdr_len +
798 /* Goal is to send at least one packet per ms,
799 * not one big TSO packet every 100 ms.
800 * This preserves ACK clocking and is consistent
801 * with tcp_tso_should_defer() heuristic.
803 gso_size = sk->sk_pacing_rate / (2 * MSEC_PER_SEC);
804 gso_size = max_t(u32, gso_size,
805 sysctl_tcp_min_tso_segs * mss_now);
807 xmit_size_goal = min_t(u32, gso_size,
808 sk->sk_gso_max_size - 1 - hlen);
810 xmit_size_goal = tcp_bound_to_half_wnd(tp, xmit_size_goal);
812 /* We try hard to avoid divides here */
813 old_size_goal = tp->xmit_size_goal_segs * mss_now;
815 if (likely(old_size_goal <= xmit_size_goal &&
816 old_size_goal + mss_now > xmit_size_goal)) {
817 xmit_size_goal = old_size_goal;
819 tp->xmit_size_goal_segs =
820 min_t(u16, xmit_size_goal / mss_now,
821 sk->sk_gso_max_segs);
822 xmit_size_goal = tp->xmit_size_goal_segs * mss_now;
826 return max(xmit_size_goal, mss_now);
829 static int tcp_send_mss(struct sock *sk, int *size_goal, int flags)
833 mss_now = tcp_current_mss(sk);
834 *size_goal = tcp_xmit_size_goal(sk, mss_now, !(flags & MSG_OOB));
839 static ssize_t do_tcp_sendpages(struct sock *sk, struct page *page, int offset,
840 size_t size, int flags)
842 struct tcp_sock *tp = tcp_sk(sk);
843 int mss_now, size_goal;
846 long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
848 /* Wait for a connection to finish. One exception is TCP Fast Open
849 * (passive side) where data is allowed to be sent before a connection
850 * is fully established.
852 if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) &&
853 !tcp_passive_fastopen(sk)) {
854 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
858 clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
860 mss_now = tcp_send_mss(sk, &size_goal, flags);
864 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
868 struct sk_buff *skb = tcp_write_queue_tail(sk);
872 if (!tcp_send_head(sk) || (copy = size_goal - skb->len) <= 0) {
874 if (!sk_stream_memory_free(sk))
875 goto wait_for_sndbuf;
877 skb = sk_stream_alloc_skb(sk, 0, sk->sk_allocation);
879 goto wait_for_memory;
888 i = skb_shinfo(skb)->nr_frags;
889 can_coalesce = skb_can_coalesce(skb, i, page, offset);
890 if (!can_coalesce && i >= MAX_SKB_FRAGS) {
891 tcp_mark_push(tp, skb);
894 if (!sk_wmem_schedule(sk, copy))
895 goto wait_for_memory;
898 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
901 skb_fill_page_desc(skb, i, page, offset, copy);
903 skb_shinfo(skb)->tx_flags |= SKBTX_SHARED_FRAG;
906 skb->data_len += copy;
907 skb->truesize += copy;
908 sk->sk_wmem_queued += copy;
909 sk_mem_charge(sk, copy);
910 skb->ip_summed = CHECKSUM_PARTIAL;
911 tp->write_seq += copy;
912 TCP_SKB_CB(skb)->end_seq += copy;
913 skb_shinfo(skb)->gso_segs = 0;
916 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
923 if (skb->len < size_goal || (flags & MSG_OOB))
926 if (forced_push(tp)) {
927 tcp_mark_push(tp, skb);
928 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
929 } else if (skb == tcp_send_head(sk))
930 tcp_push_one(sk, mss_now);
934 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
936 tcp_push(sk, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH);
938 if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
941 mss_now = tcp_send_mss(sk, &size_goal, flags);
945 if (copied && !(flags & MSG_SENDPAGE_NOTLAST))
946 tcp_push(sk, flags, mss_now, tp->nonagle);
953 return sk_stream_error(sk, flags, err);
956 int tcp_sendpage(struct sock *sk, struct page *page, int offset,
957 size_t size, int flags)
961 if (!(sk->sk_route_caps & NETIF_F_SG) ||
962 !(sk->sk_route_caps & NETIF_F_ALL_CSUM))
963 return sock_no_sendpage(sk->sk_socket, page, offset, size,
967 res = do_tcp_sendpages(sk, page, offset, size, flags);
971 EXPORT_SYMBOL(tcp_sendpage);
973 static inline int select_size(const struct sock *sk, bool sg)
975 const struct tcp_sock *tp = tcp_sk(sk);
976 int tmp = tp->mss_cache;
979 if (sk_can_gso(sk)) {
980 /* Small frames wont use a full page:
981 * Payload will immediately follow tcp header.
983 tmp = SKB_WITH_OVERHEAD(2048 - MAX_TCP_HEADER);
985 int pgbreak = SKB_MAX_HEAD(MAX_TCP_HEADER);
987 if (tmp >= pgbreak &&
988 tmp <= pgbreak + (MAX_SKB_FRAGS - 1) * PAGE_SIZE)
996 void tcp_free_fastopen_req(struct tcp_sock *tp)
998 if (tp->fastopen_req != NULL) {
999 kfree(tp->fastopen_req);
1000 tp->fastopen_req = NULL;
1004 static int tcp_sendmsg_fastopen(struct sock *sk, struct msghdr *msg,
1005 int *copied, size_t size)
1007 struct tcp_sock *tp = tcp_sk(sk);
1010 if (!(sysctl_tcp_fastopen & TFO_CLIENT_ENABLE))
1012 if (tp->fastopen_req != NULL)
1013 return -EALREADY; /* Another Fast Open is in progress */
1015 tp->fastopen_req = kzalloc(sizeof(struct tcp_fastopen_request),
1017 if (unlikely(tp->fastopen_req == NULL))
1019 tp->fastopen_req->data = msg;
1020 tp->fastopen_req->size = size;
1022 flags = (msg->msg_flags & MSG_DONTWAIT) ? O_NONBLOCK : 0;
1023 err = __inet_stream_connect(sk->sk_socket, msg->msg_name,
1024 msg->msg_namelen, flags);
1025 *copied = tp->fastopen_req->copied;
1026 tcp_free_fastopen_req(tp);
1030 int tcp_sendmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg,
1034 struct tcp_sock *tp = tcp_sk(sk);
1035 struct sk_buff *skb;
1036 int iovlen, flags, err, copied = 0;
1037 int mss_now = 0, size_goal, copied_syn = 0, offset = 0;
1043 flags = msg->msg_flags;
1044 if (flags & MSG_FASTOPEN) {
1045 err = tcp_sendmsg_fastopen(sk, msg, &copied_syn, size);
1046 if (err == -EINPROGRESS && copied_syn > 0)
1050 offset = copied_syn;
1053 timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
1055 /* Wait for a connection to finish. One exception is TCP Fast Open
1056 * (passive side) where data is allowed to be sent before a connection
1057 * is fully established.
1059 if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) &&
1060 !tcp_passive_fastopen(sk)) {
1061 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
1065 if (unlikely(tp->repair)) {
1066 if (tp->repair_queue == TCP_RECV_QUEUE) {
1067 copied = tcp_send_rcvq(sk, msg, size);
1072 if (tp->repair_queue == TCP_NO_QUEUE)
1075 /* 'common' sending to sendq */
1078 /* This should be in poll */
1079 clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
1081 mss_now = tcp_send_mss(sk, &size_goal, flags);
1083 /* Ok commence sending. */
1084 iovlen = msg->msg_iovlen;
1089 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
1092 sg = !!(sk->sk_route_caps & NETIF_F_SG);
1094 while (--iovlen >= 0) {
1095 size_t seglen = iov->iov_len;
1096 unsigned char __user *from = iov->iov_base;
1099 if (unlikely(offset > 0)) { /* Skip bytes copied in SYN */
1100 if (offset >= seglen) {
1109 while (seglen > 0) {
1111 int max = size_goal;
1113 skb = tcp_write_queue_tail(sk);
1114 if (tcp_send_head(sk)) {
1115 if (skb->ip_summed == CHECKSUM_NONE)
1117 copy = max - skb->len;
1122 /* Allocate new segment. If the interface is SG,
1123 * allocate skb fitting to single page.
1125 if (!sk_stream_memory_free(sk))
1126 goto wait_for_sndbuf;
1128 skb = sk_stream_alloc_skb(sk,
1129 select_size(sk, sg),
1132 goto wait_for_memory;
1135 * All packets are restored as if they have
1136 * already been sent.
1139 TCP_SKB_CB(skb)->when = tcp_time_stamp;
1142 * Check whether we can use HW checksum.
1144 if (sk->sk_route_caps & NETIF_F_ALL_CSUM)
1145 skb->ip_summed = CHECKSUM_PARTIAL;
1147 skb_entail(sk, skb);
1152 /* Try to append data to the end of skb. */
1156 /* Where to copy to? */
1157 if (skb_availroom(skb) > 0) {
1158 /* We have some space in skb head. Superb! */
1159 copy = min_t(int, copy, skb_availroom(skb));
1160 err = skb_add_data_nocache(sk, skb, from, copy);
1165 int i = skb_shinfo(skb)->nr_frags;
1166 struct page_frag *pfrag = sk_page_frag(sk);
1168 if (!sk_page_frag_refill(sk, pfrag))
1169 goto wait_for_memory;
1171 if (!skb_can_coalesce(skb, i, pfrag->page,
1173 if (i == MAX_SKB_FRAGS || !sg) {
1174 tcp_mark_push(tp, skb);
1180 copy = min_t(int, copy, pfrag->size - pfrag->offset);
1182 if (!sk_wmem_schedule(sk, copy))
1183 goto wait_for_memory;
1185 err = skb_copy_to_page_nocache(sk, from, skb,
1192 /* Update the skb. */
1194 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1196 skb_fill_page_desc(skb, i, pfrag->page,
1197 pfrag->offset, copy);
1198 get_page(pfrag->page);
1200 pfrag->offset += copy;
1204 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
1206 tp->write_seq += copy;
1207 TCP_SKB_CB(skb)->end_seq += copy;
1208 skb_shinfo(skb)->gso_segs = 0;
1212 if ((seglen -= copy) == 0 && iovlen == 0)
1215 if (skb->len < max || (flags & MSG_OOB) || unlikely(tp->repair))
1218 if (forced_push(tp)) {
1219 tcp_mark_push(tp, skb);
1220 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
1221 } else if (skb == tcp_send_head(sk))
1222 tcp_push_one(sk, mss_now);
1226 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1229 tcp_push(sk, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH);
1231 if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
1234 mss_now = tcp_send_mss(sk, &size_goal, flags);
1240 tcp_push(sk, flags, mss_now, tp->nonagle);
1242 return copied + copied_syn;
1246 tcp_unlink_write_queue(skb, sk);
1247 /* It is the one place in all of TCP, except connection
1248 * reset, where we can be unlinking the send_head.
1250 tcp_check_send_head(sk, skb);
1251 sk_wmem_free_skb(sk, skb);
1255 if (copied + copied_syn)
1258 err = sk_stream_error(sk, flags, err);
1262 EXPORT_SYMBOL(tcp_sendmsg);
1265 * Handle reading urgent data. BSD has very simple semantics for
1266 * this, no blocking and very strange errors 8)
1269 static int tcp_recv_urg(struct sock *sk, struct msghdr *msg, int len, int flags)
1271 struct tcp_sock *tp = tcp_sk(sk);
1273 /* No URG data to read. */
1274 if (sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data ||
1275 tp->urg_data == TCP_URG_READ)
1276 return -EINVAL; /* Yes this is right ! */
1278 if (sk->sk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DONE))
1281 if (tp->urg_data & TCP_URG_VALID) {
1283 char c = tp->urg_data;
1285 if (!(flags & MSG_PEEK))
1286 tp->urg_data = TCP_URG_READ;
1288 /* Read urgent data. */
1289 msg->msg_flags |= MSG_OOB;
1292 if (!(flags & MSG_TRUNC))
1293 err = memcpy_toiovec(msg->msg_iov, &c, 1);
1296 msg->msg_flags |= MSG_TRUNC;
1298 return err ? -EFAULT : len;
1301 if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
1304 /* Fixed the recv(..., MSG_OOB) behaviour. BSD docs and
1305 * the available implementations agree in this case:
1306 * this call should never block, independent of the
1307 * blocking state of the socket.
1308 * Mike <pall@rz.uni-karlsruhe.de>
1313 static int tcp_peek_sndq(struct sock *sk, struct msghdr *msg, int len)
1315 struct sk_buff *skb;
1316 int copied = 0, err = 0;
1318 /* XXX -- need to support SO_PEEK_OFF */
1320 skb_queue_walk(&sk->sk_write_queue, skb) {
1321 err = skb_copy_datagram_iovec(skb, 0, msg->msg_iov, skb->len);
1328 return err ?: copied;
1331 /* Clean up the receive buffer for full frames taken by the user,
1332 * then send an ACK if necessary. COPIED is the number of bytes
1333 * tcp_recvmsg has given to the user so far, it speeds up the
1334 * calculation of whether or not we must ACK for the sake of
1337 void tcp_cleanup_rbuf(struct sock *sk, int copied)
1339 struct tcp_sock *tp = tcp_sk(sk);
1340 bool time_to_ack = false;
1342 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
1344 WARN(skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq),
1345 "cleanup rbuf bug: copied %X seq %X rcvnxt %X\n",
1346 tp->copied_seq, TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt);
1348 if (inet_csk_ack_scheduled(sk)) {
1349 const struct inet_connection_sock *icsk = inet_csk(sk);
1350 /* Delayed ACKs frequently hit locked sockets during bulk
1352 if (icsk->icsk_ack.blocked ||
1353 /* Once-per-two-segments ACK was not sent by tcp_input.c */
1354 tp->rcv_nxt - tp->rcv_wup > icsk->icsk_ack.rcv_mss ||
1356 * If this read emptied read buffer, we send ACK, if
1357 * connection is not bidirectional, user drained
1358 * receive buffer and there was a small segment
1362 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED2) ||
1363 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED) &&
1364 !icsk->icsk_ack.pingpong)) &&
1365 !atomic_read(&sk->sk_rmem_alloc)))
1369 /* We send an ACK if we can now advertise a non-zero window
1370 * which has been raised "significantly".
1372 * Even if window raised up to infinity, do not send window open ACK
1373 * in states, where we will not receive more. It is useless.
1375 if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) {
1376 __u32 rcv_window_now = tcp_receive_window(tp);
1378 /* Optimize, __tcp_select_window() is not cheap. */
1379 if (2*rcv_window_now <= tp->window_clamp) {
1380 __u32 new_window = __tcp_select_window(sk);
1382 /* Send ACK now, if this read freed lots of space
1383 * in our buffer. Certainly, new_window is new window.
1384 * We can advertise it now, if it is not less than current one.
1385 * "Lots" means "at least twice" here.
1387 if (new_window && new_window >= 2 * rcv_window_now)
1395 static void tcp_prequeue_process(struct sock *sk)
1397 struct sk_buff *skb;
1398 struct tcp_sock *tp = tcp_sk(sk);
1400 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPPREQUEUED);
1402 /* RX process wants to run with disabled BHs, though it is not
1405 while ((skb = __skb_dequeue(&tp->ucopy.prequeue)) != NULL)
1406 sk_backlog_rcv(sk, skb);
1409 /* Clear memory counter. */
1410 tp->ucopy.memory = 0;
1413 #ifdef CONFIG_NET_DMA
1414 static void tcp_service_net_dma(struct sock *sk, bool wait)
1416 dma_cookie_t done, used;
1417 dma_cookie_t last_issued;
1418 struct tcp_sock *tp = tcp_sk(sk);
1420 if (!tp->ucopy.dma_chan)
1423 last_issued = tp->ucopy.dma_cookie;
1424 dma_async_issue_pending(tp->ucopy.dma_chan);
1427 if (dma_async_is_tx_complete(tp->ucopy.dma_chan,
1429 &used) == DMA_SUCCESS) {
1430 /* Safe to free early-copied skbs now */
1431 __skb_queue_purge(&sk->sk_async_wait_queue);
1434 struct sk_buff *skb;
1435 while ((skb = skb_peek(&sk->sk_async_wait_queue)) &&
1436 (dma_async_is_complete(skb->dma_cookie, done,
1437 used) == DMA_SUCCESS)) {
1438 __skb_dequeue(&sk->sk_async_wait_queue);
1446 static struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off)
1448 struct sk_buff *skb;
1451 while ((skb = skb_peek(&sk->sk_receive_queue)) != NULL) {
1452 offset = seq - TCP_SKB_CB(skb)->seq;
1453 if (tcp_hdr(skb)->syn)
1455 if (offset < skb->len || tcp_hdr(skb)->fin) {
1459 /* This looks weird, but this can happen if TCP collapsing
1460 * splitted a fat GRO packet, while we released socket lock
1461 * in skb_splice_bits()
1463 sk_eat_skb(sk, skb, false);
1469 * This routine provides an alternative to tcp_recvmsg() for routines
1470 * that would like to handle copying from skbuffs directly in 'sendfile'
1473 * - It is assumed that the socket was locked by the caller.
1474 * - The routine does not block.
1475 * - At present, there is no support for reading OOB data
1476 * or for 'peeking' the socket using this routine
1477 * (although both would be easy to implement).
1479 int tcp_read_sock(struct sock *sk, read_descriptor_t *desc,
1480 sk_read_actor_t recv_actor)
1482 struct sk_buff *skb;
1483 struct tcp_sock *tp = tcp_sk(sk);
1484 u32 seq = tp->copied_seq;
1488 if (sk->sk_state == TCP_LISTEN)
1490 while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1491 if (offset < skb->len) {
1495 len = skb->len - offset;
1496 /* Stop reading if we hit a patch of urgent data */
1498 u32 urg_offset = tp->urg_seq - seq;
1499 if (urg_offset < len)
1504 used = recv_actor(desc, skb, offset, len);
1509 } else if (used <= len) {
1514 /* If recv_actor drops the lock (e.g. TCP splice
1515 * receive) the skb pointer might be invalid when
1516 * getting here: tcp_collapse might have deleted it
1517 * while aggregating skbs from the socket queue.
1519 skb = tcp_recv_skb(sk, seq - 1, &offset);
1522 /* TCP coalescing might have appended data to the skb.
1523 * Try to splice more frags
1525 if (offset + 1 != skb->len)
1528 if (tcp_hdr(skb)->fin) {
1529 sk_eat_skb(sk, skb, false);
1533 sk_eat_skb(sk, skb, false);
1536 tp->copied_seq = seq;
1538 tp->copied_seq = seq;
1540 tcp_rcv_space_adjust(sk);
1542 /* Clean up data we have read: This will do ACK frames. */
1544 tcp_recv_skb(sk, seq, &offset);
1545 tcp_cleanup_rbuf(sk, copied);
1549 EXPORT_SYMBOL(tcp_read_sock);
1552 * This routine copies from a sock struct into the user buffer.
1554 * Technical note: in 2.3 we work on _locked_ socket, so that
1555 * tricks with *seq access order and skb->users are not required.
1556 * Probably, code can be easily improved even more.
1559 int tcp_recvmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg,
1560 size_t len, int nonblock, int flags, int *addr_len)
1562 struct tcp_sock *tp = tcp_sk(sk);
1568 int target; /* Read at least this many bytes */
1570 struct task_struct *user_recv = NULL;
1571 bool copied_early = false;
1572 struct sk_buff *skb;
1578 if (sk->sk_state == TCP_LISTEN)
1581 timeo = sock_rcvtimeo(sk, nonblock);
1583 /* Urgent data needs to be handled specially. */
1584 if (flags & MSG_OOB)
1587 if (unlikely(tp->repair)) {
1589 if (!(flags & MSG_PEEK))
1592 if (tp->repair_queue == TCP_SEND_QUEUE)
1596 if (tp->repair_queue == TCP_NO_QUEUE)
1599 /* 'common' recv queue MSG_PEEK-ing */
1602 seq = &tp->copied_seq;
1603 if (flags & MSG_PEEK) {
1604 peek_seq = tp->copied_seq;
1608 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
1610 #ifdef CONFIG_NET_DMA
1611 tp->ucopy.dma_chan = NULL;
1613 skb = skb_peek_tail(&sk->sk_receive_queue);
1618 available = TCP_SKB_CB(skb)->seq + skb->len - (*seq);
1619 if ((available < target) &&
1620 (len > sysctl_tcp_dma_copybreak) && !(flags & MSG_PEEK) &&
1621 !sysctl_tcp_low_latency &&
1622 net_dma_find_channel()) {
1623 preempt_enable_no_resched();
1624 tp->ucopy.pinned_list =
1625 dma_pin_iovec_pages(msg->msg_iov, len);
1627 preempt_enable_no_resched();
1635 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
1636 if (tp->urg_data && tp->urg_seq == *seq) {
1639 if (signal_pending(current)) {
1640 copied = timeo ? sock_intr_errno(timeo) : -EAGAIN;
1645 /* Next get a buffer. */
1647 skb_queue_walk(&sk->sk_receive_queue, skb) {
1648 /* Now that we have two receive queues this
1651 if (WARN(before(*seq, TCP_SKB_CB(skb)->seq),
1652 "recvmsg bug: copied %X seq %X rcvnxt %X fl %X\n",
1653 *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt,
1657 offset = *seq - TCP_SKB_CB(skb)->seq;
1658 if (tcp_hdr(skb)->syn)
1660 if (offset < skb->len)
1662 if (tcp_hdr(skb)->fin)
1664 WARN(!(flags & MSG_PEEK),
1665 "recvmsg bug 2: copied %X seq %X rcvnxt %X fl %X\n",
1666 *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt, flags);
1669 /* Well, if we have backlog, try to process it now yet. */
1671 if (copied >= target && !sk->sk_backlog.tail)
1676 sk->sk_state == TCP_CLOSE ||
1677 (sk->sk_shutdown & RCV_SHUTDOWN) ||
1679 signal_pending(current))
1682 if (sock_flag(sk, SOCK_DONE))
1686 copied = sock_error(sk);
1690 if (sk->sk_shutdown & RCV_SHUTDOWN)
1693 if (sk->sk_state == TCP_CLOSE) {
1694 if (!sock_flag(sk, SOCK_DONE)) {
1695 /* This occurs when user tries to read
1696 * from never connected socket.
1709 if (signal_pending(current)) {
1710 copied = sock_intr_errno(timeo);
1715 tcp_cleanup_rbuf(sk, copied);
1717 if (!sysctl_tcp_low_latency && tp->ucopy.task == user_recv) {
1718 /* Install new reader */
1719 if (!user_recv && !(flags & (MSG_TRUNC | MSG_PEEK))) {
1720 user_recv = current;
1721 tp->ucopy.task = user_recv;
1722 tp->ucopy.iov = msg->msg_iov;
1725 tp->ucopy.len = len;
1727 WARN_ON(tp->copied_seq != tp->rcv_nxt &&
1728 !(flags & (MSG_PEEK | MSG_TRUNC)));
1730 /* Ugly... If prequeue is not empty, we have to
1731 * process it before releasing socket, otherwise
1732 * order will be broken at second iteration.
1733 * More elegant solution is required!!!
1735 * Look: we have the following (pseudo)queues:
1737 * 1. packets in flight
1742 * Each queue can be processed only if the next ones
1743 * are empty. At this point we have empty receive_queue.
1744 * But prequeue _can_ be not empty after 2nd iteration,
1745 * when we jumped to start of loop because backlog
1746 * processing added something to receive_queue.
1747 * We cannot release_sock(), because backlog contains
1748 * packets arrived _after_ prequeued ones.
1750 * Shortly, algorithm is clear --- to process all
1751 * the queues in order. We could make it more directly,
1752 * requeueing packets from backlog to prequeue, if
1753 * is not empty. It is more elegant, but eats cycles,
1756 if (!skb_queue_empty(&tp->ucopy.prequeue))
1759 /* __ Set realtime policy in scheduler __ */
1762 #ifdef CONFIG_NET_DMA
1763 if (tp->ucopy.dma_chan) {
1764 if (tp->rcv_wnd == 0 &&
1765 !skb_queue_empty(&sk->sk_async_wait_queue)) {
1766 tcp_service_net_dma(sk, true);
1767 tcp_cleanup_rbuf(sk, copied);
1769 dma_async_issue_pending(tp->ucopy.dma_chan);
1772 if (copied >= target) {
1773 /* Do not sleep, just process backlog. */
1777 sk_wait_data(sk, &timeo);
1779 #ifdef CONFIG_NET_DMA
1780 tcp_service_net_dma(sk, false); /* Don't block */
1781 tp->ucopy.wakeup = 0;
1787 /* __ Restore normal policy in scheduler __ */
1789 if ((chunk = len - tp->ucopy.len) != 0) {
1790 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMBACKLOG, chunk);
1795 if (tp->rcv_nxt == tp->copied_seq &&
1796 !skb_queue_empty(&tp->ucopy.prequeue)) {
1798 tcp_prequeue_process(sk);
1800 if ((chunk = len - tp->ucopy.len) != 0) {
1801 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1807 if ((flags & MSG_PEEK) &&
1808 (peek_seq - copied - urg_hole != tp->copied_seq)) {
1809 net_dbg_ratelimited("TCP(%s:%d): Application bug, race in MSG_PEEK\n",
1811 task_pid_nr(current));
1812 peek_seq = tp->copied_seq;
1817 /* Ok so how much can we use? */
1818 used = skb->len - offset;
1822 /* Do we have urgent data here? */
1824 u32 urg_offset = tp->urg_seq - *seq;
1825 if (urg_offset < used) {
1827 if (!sock_flag(sk, SOCK_URGINLINE)) {
1840 if (!(flags & MSG_TRUNC)) {
1841 #ifdef CONFIG_NET_DMA
1842 if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list)
1843 tp->ucopy.dma_chan = net_dma_find_channel();
1845 if (tp->ucopy.dma_chan) {
1846 tp->ucopy.dma_cookie = dma_skb_copy_datagram_iovec(
1847 tp->ucopy.dma_chan, skb, offset,
1849 tp->ucopy.pinned_list);
1851 if (tp->ucopy.dma_cookie < 0) {
1853 pr_alert("%s: dma_cookie < 0\n",
1856 /* Exception. Bailout! */
1862 dma_async_issue_pending(tp->ucopy.dma_chan);
1864 if ((offset + used) == skb->len)
1865 copied_early = true;
1870 err = skb_copy_datagram_iovec(skb, offset,
1871 msg->msg_iov, used);
1873 /* Exception. Bailout! */
1885 tcp_rcv_space_adjust(sk);
1888 if (tp->urg_data && after(tp->copied_seq, tp->urg_seq)) {
1890 tcp_fast_path_check(sk);
1892 if (used + offset < skb->len)
1895 if (tcp_hdr(skb)->fin)
1897 if (!(flags & MSG_PEEK)) {
1898 sk_eat_skb(sk, skb, copied_early);
1899 copied_early = false;
1904 /* Process the FIN. */
1906 if (!(flags & MSG_PEEK)) {
1907 sk_eat_skb(sk, skb, copied_early);
1908 copied_early = false;
1914 if (!skb_queue_empty(&tp->ucopy.prequeue)) {
1917 tp->ucopy.len = copied > 0 ? len : 0;
1919 tcp_prequeue_process(sk);
1921 if (copied > 0 && (chunk = len - tp->ucopy.len) != 0) {
1922 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1928 tp->ucopy.task = NULL;
1932 #ifdef CONFIG_NET_DMA
1933 tcp_service_net_dma(sk, true); /* Wait for queue to drain */
1934 tp->ucopy.dma_chan = NULL;
1936 if (tp->ucopy.pinned_list) {
1937 dma_unpin_iovec_pages(tp->ucopy.pinned_list);
1938 tp->ucopy.pinned_list = NULL;
1942 /* According to UNIX98, msg_name/msg_namelen are ignored
1943 * on connected socket. I was just happy when found this 8) --ANK
1946 /* Clean up data we have read: This will do ACK frames. */
1947 tcp_cleanup_rbuf(sk, copied);
1957 err = tcp_recv_urg(sk, msg, len, flags);
1961 err = tcp_peek_sndq(sk, msg, len);
1964 EXPORT_SYMBOL(tcp_recvmsg);
1966 void tcp_set_state(struct sock *sk, int state)
1968 int oldstate = sk->sk_state;
1971 case TCP_ESTABLISHED:
1972 if (oldstate != TCP_ESTABLISHED)
1973 TCP_INC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
1977 if (oldstate == TCP_CLOSE_WAIT || oldstate == TCP_ESTABLISHED)
1978 TCP_INC_STATS(sock_net(sk), TCP_MIB_ESTABRESETS);
1980 sk->sk_prot->unhash(sk);
1981 if (inet_csk(sk)->icsk_bind_hash &&
1982 !(sk->sk_userlocks & SOCK_BINDPORT_LOCK))
1986 if (oldstate == TCP_ESTABLISHED)
1987 TCP_DEC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
1990 /* Change state AFTER socket is unhashed to avoid closed
1991 * socket sitting in hash tables.
1993 sk->sk_state = state;
1996 SOCK_DEBUG(sk, "TCP sk=%p, State %s -> %s\n", sk, statename[oldstate], statename[state]);
1999 EXPORT_SYMBOL_GPL(tcp_set_state);
2002 * State processing on a close. This implements the state shift for
2003 * sending our FIN frame. Note that we only send a FIN for some
2004 * states. A shutdown() may have already sent the FIN, or we may be
2008 static const unsigned char new_state[16] = {
2009 /* current state: new state: action: */
2010 /* (Invalid) */ TCP_CLOSE,
2011 /* TCP_ESTABLISHED */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2012 /* TCP_SYN_SENT */ TCP_CLOSE,
2013 /* TCP_SYN_RECV */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2014 /* TCP_FIN_WAIT1 */ TCP_FIN_WAIT1,
2015 /* TCP_FIN_WAIT2 */ TCP_FIN_WAIT2,
2016 /* TCP_TIME_WAIT */ TCP_CLOSE,
2017 /* TCP_CLOSE */ TCP_CLOSE,
2018 /* TCP_CLOSE_WAIT */ TCP_LAST_ACK | TCP_ACTION_FIN,
2019 /* TCP_LAST_ACK */ TCP_LAST_ACK,
2020 /* TCP_LISTEN */ TCP_CLOSE,
2021 /* TCP_CLOSING */ TCP_CLOSING,
2024 static int tcp_close_state(struct sock *sk)
2026 int next = (int)new_state[sk->sk_state];
2027 int ns = next & TCP_STATE_MASK;
2029 tcp_set_state(sk, ns);
2031 return next & TCP_ACTION_FIN;
2035 * Shutdown the sending side of a connection. Much like close except
2036 * that we don't receive shut down or sock_set_flag(sk, SOCK_DEAD).
2039 void tcp_shutdown(struct sock *sk, int how)
2041 /* We need to grab some memory, and put together a FIN,
2042 * and then put it into the queue to be sent.
2043 * Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
2045 if (!(how & SEND_SHUTDOWN))
2048 /* If we've already sent a FIN, or it's a closed state, skip this. */
2049 if ((1 << sk->sk_state) &
2050 (TCPF_ESTABLISHED | TCPF_SYN_SENT |
2051 TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) {
2052 /* Clear out any half completed packets. FIN if needed. */
2053 if (tcp_close_state(sk))
2057 EXPORT_SYMBOL(tcp_shutdown);
2059 bool tcp_check_oom(struct sock *sk, int shift)
2061 bool too_many_orphans, out_of_socket_memory;
2063 too_many_orphans = tcp_too_many_orphans(sk, shift);
2064 out_of_socket_memory = tcp_out_of_memory(sk);
2066 if (too_many_orphans)
2067 net_info_ratelimited("too many orphaned sockets\n");
2068 if (out_of_socket_memory)
2069 net_info_ratelimited("out of memory -- consider tuning tcp_mem\n");
2070 return too_many_orphans || out_of_socket_memory;
2073 void tcp_close(struct sock *sk, long timeout)
2075 struct sk_buff *skb;
2076 int data_was_unread = 0;
2080 sk->sk_shutdown = SHUTDOWN_MASK;
2082 if (sk->sk_state == TCP_LISTEN) {
2083 tcp_set_state(sk, TCP_CLOSE);
2086 inet_csk_listen_stop(sk);
2088 goto adjudge_to_death;
2091 /* We need to flush the recv. buffs. We do this only on the
2092 * descriptor close, not protocol-sourced closes, because the
2093 * reader process may not have drained the data yet!
2095 while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) {
2096 u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq -
2098 data_was_unread += len;
2104 /* If socket has been already reset (e.g. in tcp_reset()) - kill it. */
2105 if (sk->sk_state == TCP_CLOSE)
2106 goto adjudge_to_death;
2108 /* As outlined in RFC 2525, section 2.17, we send a RST here because
2109 * data was lost. To witness the awful effects of the old behavior of
2110 * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk
2111 * GET in an FTP client, suspend the process, wait for the client to
2112 * advertise a zero window, then kill -9 the FTP client, wheee...
2113 * Note: timeout is always zero in such a case.
2115 if (unlikely(tcp_sk(sk)->repair)) {
2116 sk->sk_prot->disconnect(sk, 0);
2117 } else if (data_was_unread) {
2118 /* Unread data was tossed, zap the connection. */
2119 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONCLOSE);
2120 tcp_set_state(sk, TCP_CLOSE);
2121 tcp_send_active_reset(sk, sk->sk_allocation);
2122 } else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) {
2123 /* Check zero linger _after_ checking for unread data. */
2124 sk->sk_prot->disconnect(sk, 0);
2125 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
2126 } else if (tcp_close_state(sk)) {
2127 /* We FIN if the application ate all the data before
2128 * zapping the connection.
2131 /* RED-PEN. Formally speaking, we have broken TCP state
2132 * machine. State transitions:
2134 * TCP_ESTABLISHED -> TCP_FIN_WAIT1
2135 * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible)
2136 * TCP_CLOSE_WAIT -> TCP_LAST_ACK
2138 * are legal only when FIN has been sent (i.e. in window),
2139 * rather than queued out of window. Purists blame.
2141 * F.e. "RFC state" is ESTABLISHED,
2142 * if Linux state is FIN-WAIT-1, but FIN is still not sent.
2144 * The visible declinations are that sometimes
2145 * we enter time-wait state, when it is not required really
2146 * (harmless), do not send active resets, when they are
2147 * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
2148 * they look as CLOSING or LAST_ACK for Linux)
2149 * Probably, I missed some more holelets.
2151 * XXX (TFO) - To start off we don't support SYN+ACK+FIN
2152 * in a single packet! (May consider it later but will
2153 * probably need API support or TCP_CORK SYN-ACK until
2154 * data is written and socket is closed.)
2159 sk_stream_wait_close(sk, timeout);
2162 state = sk->sk_state;
2166 /* It is the last release_sock in its life. It will remove backlog. */
2170 /* Now socket is owned by kernel and we acquire BH lock
2171 to finish close. No need to check for user refs.
2175 WARN_ON(sock_owned_by_user(sk));
2177 percpu_counter_inc(sk->sk_prot->orphan_count);
2179 /* Have we already been destroyed by a softirq or backlog? */
2180 if (state != TCP_CLOSE && sk->sk_state == TCP_CLOSE)
2183 /* This is a (useful) BSD violating of the RFC. There is a
2184 * problem with TCP as specified in that the other end could
2185 * keep a socket open forever with no application left this end.
2186 * We use a 3 minute timeout (about the same as BSD) then kill
2187 * our end. If they send after that then tough - BUT: long enough
2188 * that we won't make the old 4*rto = almost no time - whoops
2191 * Nope, it was not mistake. It is really desired behaviour
2192 * f.e. on http servers, when such sockets are useless, but
2193 * consume significant resources. Let's do it with special
2194 * linger2 option. --ANK
2197 if (sk->sk_state == TCP_FIN_WAIT2) {
2198 struct tcp_sock *tp = tcp_sk(sk);
2199 if (tp->linger2 < 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_TCPABORTONLINGER);
2205 const int tmo = tcp_fin_time(sk);
2207 if (tmo > TCP_TIMEWAIT_LEN) {
2208 inet_csk_reset_keepalive_timer(sk,
2209 tmo - TCP_TIMEWAIT_LEN);
2211 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
2216 if (sk->sk_state != TCP_CLOSE) {
2218 if (tcp_check_oom(sk, 0)) {
2219 tcp_set_state(sk, TCP_CLOSE);
2220 tcp_send_active_reset(sk, GFP_ATOMIC);
2221 NET_INC_STATS_BH(sock_net(sk),
2222 LINUX_MIB_TCPABORTONMEMORY);
2226 if (sk->sk_state == TCP_CLOSE) {
2227 struct request_sock *req = tcp_sk(sk)->fastopen_rsk;
2228 /* We could get here with a non-NULL req if the socket is
2229 * aborted (e.g., closed with unread data) before 3WHS
2233 reqsk_fastopen_remove(sk, req, false);
2234 inet_csk_destroy_sock(sk);
2236 /* Otherwise, socket is reprieved until protocol close. */
2243 EXPORT_SYMBOL(tcp_close);
2245 /* These states need RST on ABORT according to RFC793 */
2247 static inline bool tcp_need_reset(int state)
2249 return (1 << state) &
2250 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 |
2251 TCPF_FIN_WAIT2 | TCPF_SYN_RECV);
2254 int tcp_disconnect(struct sock *sk, int flags)
2256 struct inet_sock *inet = inet_sk(sk);
2257 struct inet_connection_sock *icsk = inet_csk(sk);
2258 struct tcp_sock *tp = tcp_sk(sk);
2260 int old_state = sk->sk_state;
2262 if (old_state != TCP_CLOSE)
2263 tcp_set_state(sk, TCP_CLOSE);
2265 /* ABORT function of RFC793 */
2266 if (old_state == TCP_LISTEN) {
2267 inet_csk_listen_stop(sk);
2268 } else if (unlikely(tp->repair)) {
2269 sk->sk_err = ECONNABORTED;
2270 } else if (tcp_need_reset(old_state) ||
2271 (tp->snd_nxt != tp->write_seq &&
2272 (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK))) {
2273 /* The last check adjusts for discrepancy of Linux wrt. RFC
2276 tcp_send_active_reset(sk, gfp_any());
2277 sk->sk_err = ECONNRESET;
2278 } else if (old_state == TCP_SYN_SENT)
2279 sk->sk_err = ECONNRESET;
2281 tcp_clear_xmit_timers(sk);
2282 __skb_queue_purge(&sk->sk_receive_queue);
2283 tcp_write_queue_purge(sk);
2284 __skb_queue_purge(&tp->out_of_order_queue);
2285 #ifdef CONFIG_NET_DMA
2286 __skb_queue_purge(&sk->sk_async_wait_queue);
2289 inet->inet_dport = 0;
2291 if (!(sk->sk_userlocks & SOCK_BINDADDR_LOCK))
2292 inet_reset_saddr(sk);
2294 sk->sk_shutdown = 0;
2295 sock_reset_flag(sk, SOCK_DONE);
2297 if ((tp->write_seq += tp->max_window + 2) == 0)
2299 icsk->icsk_backoff = 0;
2301 icsk->icsk_probes_out = 0;
2302 tp->packets_out = 0;
2303 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
2304 tp->snd_cwnd_cnt = 0;
2305 tp->window_clamp = 0;
2306 tcp_set_ca_state(sk, TCP_CA_Open);
2307 tcp_clear_retrans(tp);
2308 inet_csk_delack_init(sk);
2309 tcp_init_send_head(sk);
2310 memset(&tp->rx_opt, 0, sizeof(tp->rx_opt));
2313 WARN_ON(inet->inet_num && !icsk->icsk_bind_hash);
2315 sk->sk_error_report(sk);
2318 EXPORT_SYMBOL(tcp_disconnect);
2320 void tcp_sock_destruct(struct sock *sk)
2322 inet_sock_destruct(sk);
2324 kfree(inet_csk(sk)->icsk_accept_queue.fastopenq);
2327 static inline bool tcp_can_repair_sock(const struct sock *sk)
2329 return ns_capable(sock_net(sk)->user_ns, CAP_NET_ADMIN) &&
2330 ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_ESTABLISHED));
2333 static int tcp_repair_options_est(struct tcp_sock *tp,
2334 struct tcp_repair_opt __user *optbuf, unsigned int len)
2336 struct tcp_repair_opt opt;
2338 while (len >= sizeof(opt)) {
2339 if (copy_from_user(&opt, optbuf, sizeof(opt)))
2345 switch (opt.opt_code) {
2347 tp->rx_opt.mss_clamp = opt.opt_val;
2351 u16 snd_wscale = opt.opt_val & 0xFFFF;
2352 u16 rcv_wscale = opt.opt_val >> 16;
2354 if (snd_wscale > 14 || rcv_wscale > 14)
2357 tp->rx_opt.snd_wscale = snd_wscale;
2358 tp->rx_opt.rcv_wscale = rcv_wscale;
2359 tp->rx_opt.wscale_ok = 1;
2362 case TCPOPT_SACK_PERM:
2363 if (opt.opt_val != 0)
2366 tp->rx_opt.sack_ok |= TCP_SACK_SEEN;
2367 if (sysctl_tcp_fack)
2368 tcp_enable_fack(tp);
2370 case TCPOPT_TIMESTAMP:
2371 if (opt.opt_val != 0)
2374 tp->rx_opt.tstamp_ok = 1;
2383 * Socket option code for TCP.
2385 static int do_tcp_setsockopt(struct sock *sk, int level,
2386 int optname, char __user *optval, unsigned int optlen)
2388 struct tcp_sock *tp = tcp_sk(sk);
2389 struct inet_connection_sock *icsk = inet_csk(sk);
2393 /* These are data/string values, all the others are ints */
2395 case TCP_CONGESTION: {
2396 char name[TCP_CA_NAME_MAX];
2401 val = strncpy_from_user(name, optval,
2402 min_t(long, TCP_CA_NAME_MAX-1, optlen));
2408 err = tcp_set_congestion_control(sk, name);
2417 if (optlen < sizeof(int))
2420 if (get_user(val, (int __user *)optval))
2427 /* Values greater than interface MTU won't take effect. However
2428 * at the point when this call is done we typically don't yet
2429 * know which interface is going to be used */
2430 if (val < TCP_MIN_MSS || val > MAX_TCP_WINDOW) {
2434 tp->rx_opt.user_mss = val;
2439 /* TCP_NODELAY is weaker than TCP_CORK, so that
2440 * this option on corked socket is remembered, but
2441 * it is not activated until cork is cleared.
2443 * However, when TCP_NODELAY is set we make
2444 * an explicit push, which overrides even TCP_CORK
2445 * for currently queued segments.
2447 tp->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH;
2448 tcp_push_pending_frames(sk);
2450 tp->nonagle &= ~TCP_NAGLE_OFF;
2454 case TCP_THIN_LINEAR_TIMEOUTS:
2455 if (val < 0 || val > 1)
2461 case TCP_THIN_DUPACK:
2462 if (val < 0 || val > 1)
2465 tp->thin_dupack = val;
2466 if (tp->thin_dupack)
2467 tcp_disable_early_retrans(tp);
2472 if (!tcp_can_repair_sock(sk))
2474 else if (val == 1) {
2476 sk->sk_reuse = SK_FORCE_REUSE;
2477 tp->repair_queue = TCP_NO_QUEUE;
2478 } else if (val == 0) {
2480 sk->sk_reuse = SK_NO_REUSE;
2481 tcp_send_window_probe(sk);
2487 case TCP_REPAIR_QUEUE:
2490 else if (val < TCP_QUEUES_NR)
2491 tp->repair_queue = val;
2497 if (sk->sk_state != TCP_CLOSE)
2499 else if (tp->repair_queue == TCP_SEND_QUEUE)
2500 tp->write_seq = val;
2501 else if (tp->repair_queue == TCP_RECV_QUEUE)
2507 case TCP_REPAIR_OPTIONS:
2510 else if (sk->sk_state == TCP_ESTABLISHED)
2511 err = tcp_repair_options_est(tp,
2512 (struct tcp_repair_opt __user *)optval,
2519 /* When set indicates to always queue non-full frames.
2520 * Later the user clears this option and we transmit
2521 * any pending partial frames in the queue. This is
2522 * meant to be used alongside sendfile() to get properly
2523 * filled frames when the user (for example) must write
2524 * out headers with a write() call first and then use
2525 * sendfile to send out the data parts.
2527 * TCP_CORK can be set together with TCP_NODELAY and it is
2528 * stronger than TCP_NODELAY.
2531 tp->nonagle |= TCP_NAGLE_CORK;
2533 tp->nonagle &= ~TCP_NAGLE_CORK;
2534 if (tp->nonagle&TCP_NAGLE_OFF)
2535 tp->nonagle |= TCP_NAGLE_PUSH;
2536 tcp_push_pending_frames(sk);
2541 if (val < 1 || val > MAX_TCP_KEEPIDLE)
2544 tp->keepalive_time = val * HZ;
2545 if (sock_flag(sk, SOCK_KEEPOPEN) &&
2546 !((1 << sk->sk_state) &
2547 (TCPF_CLOSE | TCPF_LISTEN))) {
2548 u32 elapsed = keepalive_time_elapsed(tp);
2549 if (tp->keepalive_time > elapsed)
2550 elapsed = tp->keepalive_time - elapsed;
2553 inet_csk_reset_keepalive_timer(sk, elapsed);
2558 if (val < 1 || val > MAX_TCP_KEEPINTVL)
2561 tp->keepalive_intvl = val * HZ;
2564 if (val < 1 || val > MAX_TCP_KEEPCNT)
2567 tp->keepalive_probes = val;
2570 if (val < 1 || val > MAX_TCP_SYNCNT)
2573 icsk->icsk_syn_retries = val;
2579 else if (val > sysctl_tcp_fin_timeout / HZ)
2582 tp->linger2 = val * HZ;
2585 case TCP_DEFER_ACCEPT:
2586 /* Translate value in seconds to number of retransmits */
2587 icsk->icsk_accept_queue.rskq_defer_accept =
2588 secs_to_retrans(val, TCP_TIMEOUT_INIT / HZ,
2592 case TCP_WINDOW_CLAMP:
2594 if (sk->sk_state != TCP_CLOSE) {
2598 tp->window_clamp = 0;
2600 tp->window_clamp = val < SOCK_MIN_RCVBUF / 2 ?
2601 SOCK_MIN_RCVBUF / 2 : val;
2606 icsk->icsk_ack.pingpong = 1;
2608 icsk->icsk_ack.pingpong = 0;
2609 if ((1 << sk->sk_state) &
2610 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) &&
2611 inet_csk_ack_scheduled(sk)) {
2612 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
2613 tcp_cleanup_rbuf(sk, 1);
2615 icsk->icsk_ack.pingpong = 1;
2620 #ifdef CONFIG_TCP_MD5SIG
2622 /* Read the IP->Key mappings from userspace */
2623 err = tp->af_specific->md5_parse(sk, optval, optlen);
2626 case TCP_USER_TIMEOUT:
2627 /* Cap the max timeout in ms TCP will retry/retrans
2628 * before giving up and aborting (ETIMEDOUT) a connection.
2633 icsk->icsk_user_timeout = msecs_to_jiffies(val);
2637 if (val >= 0 && ((1 << sk->sk_state) & (TCPF_CLOSE |
2639 err = fastopen_init_queue(sk, val);
2647 tp->tsoffset = val - tcp_time_stamp;
2658 int tcp_setsockopt(struct sock *sk, int level, int optname, char __user *optval,
2659 unsigned int optlen)
2661 const struct inet_connection_sock *icsk = inet_csk(sk);
2663 if (level != SOL_TCP)
2664 return icsk->icsk_af_ops->setsockopt(sk, level, optname,
2666 return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2668 EXPORT_SYMBOL(tcp_setsockopt);
2670 #ifdef CONFIG_COMPAT
2671 int compat_tcp_setsockopt(struct sock *sk, int level, int optname,
2672 char __user *optval, unsigned int optlen)
2674 if (level != SOL_TCP)
2675 return inet_csk_compat_setsockopt(sk, level, optname,
2677 return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2679 EXPORT_SYMBOL(compat_tcp_setsockopt);
2682 /* Return information about state of tcp endpoint in API format. */
2683 void tcp_get_info(const struct sock *sk, struct tcp_info *info)
2685 const struct tcp_sock *tp = tcp_sk(sk);
2686 const struct inet_connection_sock *icsk = inet_csk(sk);
2687 u32 now = tcp_time_stamp;
2689 memset(info, 0, sizeof(*info));
2691 info->tcpi_state = sk->sk_state;
2692 info->tcpi_ca_state = icsk->icsk_ca_state;
2693 info->tcpi_retransmits = icsk->icsk_retransmits;
2694 info->tcpi_probes = icsk->icsk_probes_out;
2695 info->tcpi_backoff = icsk->icsk_backoff;
2697 if (tp->rx_opt.tstamp_ok)
2698 info->tcpi_options |= TCPI_OPT_TIMESTAMPS;
2699 if (tcp_is_sack(tp))
2700 info->tcpi_options |= TCPI_OPT_SACK;
2701 if (tp->rx_opt.wscale_ok) {
2702 info->tcpi_options |= TCPI_OPT_WSCALE;
2703 info->tcpi_snd_wscale = tp->rx_opt.snd_wscale;
2704 info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale;
2707 if (tp->ecn_flags & TCP_ECN_OK)
2708 info->tcpi_options |= TCPI_OPT_ECN;
2709 if (tp->ecn_flags & TCP_ECN_SEEN)
2710 info->tcpi_options |= TCPI_OPT_ECN_SEEN;
2711 if (tp->syn_data_acked)
2712 info->tcpi_options |= TCPI_OPT_SYN_DATA;
2714 info->tcpi_rto = jiffies_to_usecs(icsk->icsk_rto);
2715 info->tcpi_ato = jiffies_to_usecs(icsk->icsk_ack.ato);
2716 info->tcpi_snd_mss = tp->mss_cache;
2717 info->tcpi_rcv_mss = icsk->icsk_ack.rcv_mss;
2719 if (sk->sk_state == TCP_LISTEN) {
2720 info->tcpi_unacked = sk->sk_ack_backlog;
2721 info->tcpi_sacked = sk->sk_max_ack_backlog;
2723 info->tcpi_unacked = tp->packets_out;
2724 info->tcpi_sacked = tp->sacked_out;
2726 info->tcpi_lost = tp->lost_out;
2727 info->tcpi_retrans = tp->retrans_out;
2728 info->tcpi_fackets = tp->fackets_out;
2730 info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime);
2731 info->tcpi_last_data_recv = jiffies_to_msecs(now - icsk->icsk_ack.lrcvtime);
2732 info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp);
2734 info->tcpi_pmtu = icsk->icsk_pmtu_cookie;
2735 info->tcpi_rcv_ssthresh = tp->rcv_ssthresh;
2736 info->tcpi_rtt = jiffies_to_usecs(tp->srtt)>>3;
2737 info->tcpi_rttvar = jiffies_to_usecs(tp->mdev)>>2;
2738 info->tcpi_snd_ssthresh = tp->snd_ssthresh;
2739 info->tcpi_snd_cwnd = tp->snd_cwnd;
2740 info->tcpi_advmss = tp->advmss;
2741 info->tcpi_reordering = tp->reordering;
2743 info->tcpi_rcv_rtt = jiffies_to_usecs(tp->rcv_rtt_est.rtt)>>3;
2744 info->tcpi_rcv_space = tp->rcvq_space.space;
2746 info->tcpi_total_retrans = tp->total_retrans;
2748 EXPORT_SYMBOL_GPL(tcp_get_info);
2750 static int do_tcp_getsockopt(struct sock *sk, int level,
2751 int optname, char __user *optval, int __user *optlen)
2753 struct inet_connection_sock *icsk = inet_csk(sk);
2754 struct tcp_sock *tp = tcp_sk(sk);
2757 if (get_user(len, optlen))
2760 len = min_t(unsigned int, len, sizeof(int));
2767 val = tp->mss_cache;
2768 if (!val && ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
2769 val = tp->rx_opt.user_mss;
2771 val = tp->rx_opt.mss_clamp;
2774 val = !!(tp->nonagle&TCP_NAGLE_OFF);
2777 val = !!(tp->nonagle&TCP_NAGLE_CORK);
2780 val = keepalive_time_when(tp) / HZ;
2783 val = keepalive_intvl_when(tp) / HZ;
2786 val = keepalive_probes(tp);
2789 val = icsk->icsk_syn_retries ? : sysctl_tcp_syn_retries;
2794 val = (val ? : sysctl_tcp_fin_timeout) / HZ;
2796 case TCP_DEFER_ACCEPT:
2797 val = retrans_to_secs(icsk->icsk_accept_queue.rskq_defer_accept,
2798 TCP_TIMEOUT_INIT / HZ, TCP_RTO_MAX / HZ);
2800 case TCP_WINDOW_CLAMP:
2801 val = tp->window_clamp;
2804 struct tcp_info info;
2806 if (get_user(len, optlen))
2809 tcp_get_info(sk, &info);
2811 len = min_t(unsigned int, len, sizeof(info));
2812 if (put_user(len, optlen))
2814 if (copy_to_user(optval, &info, len))
2819 val = !icsk->icsk_ack.pingpong;
2822 case TCP_CONGESTION:
2823 if (get_user(len, optlen))
2825 len = min_t(unsigned int, len, TCP_CA_NAME_MAX);
2826 if (put_user(len, optlen))
2828 if (copy_to_user(optval, icsk->icsk_ca_ops->name, len))
2832 case TCP_THIN_LINEAR_TIMEOUTS:
2835 case TCP_THIN_DUPACK:
2836 val = tp->thin_dupack;
2843 case TCP_REPAIR_QUEUE:
2845 val = tp->repair_queue;
2851 if (tp->repair_queue == TCP_SEND_QUEUE)
2852 val = tp->write_seq;
2853 else if (tp->repair_queue == TCP_RECV_QUEUE)
2859 case TCP_USER_TIMEOUT:
2860 val = jiffies_to_msecs(icsk->icsk_user_timeout);
2863 val = tcp_time_stamp + tp->tsoffset;
2866 return -ENOPROTOOPT;
2869 if (put_user(len, optlen))
2871 if (copy_to_user(optval, &val, len))
2876 int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval,
2879 struct inet_connection_sock *icsk = inet_csk(sk);
2881 if (level != SOL_TCP)
2882 return icsk->icsk_af_ops->getsockopt(sk, level, optname,
2884 return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2886 EXPORT_SYMBOL(tcp_getsockopt);
2888 #ifdef CONFIG_COMPAT
2889 int compat_tcp_getsockopt(struct sock *sk, int level, int optname,
2890 char __user *optval, int __user *optlen)
2892 if (level != SOL_TCP)
2893 return inet_csk_compat_getsockopt(sk, level, optname,
2895 return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2897 EXPORT_SYMBOL(compat_tcp_getsockopt);
2900 struct sk_buff *tcp_tso_segment(struct sk_buff *skb,
2901 netdev_features_t features)
2903 struct sk_buff *segs = ERR_PTR(-EINVAL);
2904 unsigned int sum_truesize = 0;
2909 unsigned int oldlen;
2911 struct sk_buff *gso_skb = skb;
2913 bool ooo_okay, copy_destructor;
2915 if (!pskb_may_pull(skb, sizeof(*th)))
2919 thlen = th->doff * 4;
2920 if (thlen < sizeof(*th))
2923 if (!pskb_may_pull(skb, thlen))
2926 oldlen = (u16)~skb->len;
2927 __skb_pull(skb, thlen);
2929 mss = skb_shinfo(skb)->gso_size;
2930 if (unlikely(skb->len <= mss))
2933 if (skb_gso_ok(skb, features | NETIF_F_GSO_ROBUST)) {
2934 /* Packet is from an untrusted source, reset gso_segs. */
2935 int type = skb_shinfo(skb)->gso_type;
2943 SKB_GSO_UDP_TUNNEL |
2945 !(type & (SKB_GSO_TCPV4 | SKB_GSO_TCPV6))))
2948 skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss);
2954 copy_destructor = gso_skb->destructor == tcp_wfree;
2955 ooo_okay = gso_skb->ooo_okay;
2956 /* All segments but the first should have ooo_okay cleared */
2959 segs = skb_segment(skb, features);
2963 /* Only first segment might have ooo_okay set */
2964 segs->ooo_okay = ooo_okay;
2966 delta = htonl(oldlen + (thlen + mss));
2970 seq = ntohl(th->seq);
2972 newcheck = ~csum_fold((__force __wsum)((__force u32)th->check +
2973 (__force u32)delta));
2976 th->fin = th->psh = 0;
2977 th->check = newcheck;
2979 if (skb->ip_summed != CHECKSUM_PARTIAL)
2981 csum_fold(csum_partial(skb_transport_header(skb),
2985 if (copy_destructor) {
2986 skb->destructor = gso_skb->destructor;
2987 skb->sk = gso_skb->sk;
2988 sum_truesize += skb->truesize;
2993 th->seq = htonl(seq);
2995 } while (skb->next);
2997 /* Following permits TCP Small Queues to work well with GSO :
2998 * The callback to TCP stack will be called at the time last frag
2999 * is freed at TX completion, and not right now when gso_skb
3000 * is freed by GSO engine
3002 if (copy_destructor) {
3003 swap(gso_skb->sk, skb->sk);
3004 swap(gso_skb->destructor, skb->destructor);
3005 sum_truesize += skb->truesize;
3006 atomic_add(sum_truesize - gso_skb->truesize,
3007 &skb->sk->sk_wmem_alloc);
3010 delta = htonl(oldlen + (skb->tail - skb->transport_header) +
3012 th->check = ~csum_fold((__force __wsum)((__force u32)th->check +
3013 (__force u32)delta));
3014 if (skb->ip_summed != CHECKSUM_PARTIAL)
3015 th->check = csum_fold(csum_partial(skb_transport_header(skb),
3021 EXPORT_SYMBOL(tcp_tso_segment);
3023 struct sk_buff **tcp_gro_receive(struct sk_buff **head, struct sk_buff *skb)
3025 struct sk_buff **pp = NULL;
3032 unsigned int mss = 1;
3038 off = skb_gro_offset(skb);
3039 hlen = off + sizeof(*th);
3040 th = skb_gro_header_fast(skb, off);
3041 if (skb_gro_header_hard(skb, hlen)) {
3042 th = skb_gro_header_slow(skb, hlen, off);
3047 thlen = th->doff * 4;
3048 if (thlen < sizeof(*th))
3052 if (skb_gro_header_hard(skb, hlen)) {
3053 th = skb_gro_header_slow(skb, hlen, off);
3058 skb_gro_pull(skb, thlen);
3060 len = skb_gro_len(skb);
3061 flags = tcp_flag_word(th);
3063 for (; (p = *head); head = &p->next) {
3064 if (!NAPI_GRO_CB(p)->same_flow)
3069 if (*(u32 *)&th->source ^ *(u32 *)&th2->source) {
3070 NAPI_GRO_CB(p)->same_flow = 0;
3077 goto out_check_final;
3080 flush = NAPI_GRO_CB(p)->flush;
3081 flush |= (__force int)(flags & TCP_FLAG_CWR);
3082 flush |= (__force int)((flags ^ tcp_flag_word(th2)) &
3083 ~(TCP_FLAG_CWR | TCP_FLAG_FIN | TCP_FLAG_PSH));
3084 flush |= (__force int)(th->ack_seq ^ th2->ack_seq);
3085 for (i = sizeof(*th); i < thlen; i += 4)
3086 flush |= *(u32 *)((u8 *)th + i) ^
3087 *(u32 *)((u8 *)th2 + i);
3089 mss = skb_shinfo(p)->gso_size;
3091 flush |= (len - 1) >= mss;
3092 flush |= (ntohl(th2->seq) + skb_gro_len(p)) ^ ntohl(th->seq);
3094 if (flush || skb_gro_receive(head, skb)) {
3096 goto out_check_final;
3101 tcp_flag_word(th2) |= flags & (TCP_FLAG_FIN | TCP_FLAG_PSH);
3105 flush |= (__force int)(flags & (TCP_FLAG_URG | TCP_FLAG_PSH |
3106 TCP_FLAG_RST | TCP_FLAG_SYN |
3109 if (p && (!NAPI_GRO_CB(skb)->same_flow || flush))
3113 NAPI_GRO_CB(skb)->flush |= flush;
3117 EXPORT_SYMBOL(tcp_gro_receive);
3119 int tcp_gro_complete(struct sk_buff *skb)
3121 struct tcphdr *th = tcp_hdr(skb);
3123 skb->csum_start = skb_transport_header(skb) - skb->head;
3124 skb->csum_offset = offsetof(struct tcphdr, check);
3125 skb->ip_summed = CHECKSUM_PARTIAL;
3127 skb_shinfo(skb)->gso_segs = NAPI_GRO_CB(skb)->count;
3130 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
3134 EXPORT_SYMBOL(tcp_gro_complete);
3136 #ifdef CONFIG_TCP_MD5SIG
3137 static unsigned long tcp_md5sig_users;
3138 static struct tcp_md5sig_pool __percpu *tcp_md5sig_pool;
3139 static DEFINE_SPINLOCK(tcp_md5sig_pool_lock);
3141 static void __tcp_free_md5sig_pool(struct tcp_md5sig_pool __percpu *pool)
3145 for_each_possible_cpu(cpu) {
3146 struct tcp_md5sig_pool *p = per_cpu_ptr(pool, cpu);
3148 if (p->md5_desc.tfm)
3149 crypto_free_hash(p->md5_desc.tfm);
3154 void tcp_free_md5sig_pool(void)
3156 struct tcp_md5sig_pool __percpu *pool = NULL;
3158 spin_lock_bh(&tcp_md5sig_pool_lock);
3159 if (--tcp_md5sig_users == 0) {
3160 pool = tcp_md5sig_pool;
3161 tcp_md5sig_pool = NULL;
3163 spin_unlock_bh(&tcp_md5sig_pool_lock);
3165 __tcp_free_md5sig_pool(pool);
3167 EXPORT_SYMBOL(tcp_free_md5sig_pool);
3169 static struct tcp_md5sig_pool __percpu *
3170 __tcp_alloc_md5sig_pool(struct sock *sk)
3173 struct tcp_md5sig_pool __percpu *pool;
3175 pool = alloc_percpu(struct tcp_md5sig_pool);
3179 for_each_possible_cpu(cpu) {
3180 struct crypto_hash *hash;
3182 hash = crypto_alloc_hash("md5", 0, CRYPTO_ALG_ASYNC);
3183 if (IS_ERR_OR_NULL(hash))
3186 per_cpu_ptr(pool, cpu)->md5_desc.tfm = hash;
3190 __tcp_free_md5sig_pool(pool);
3194 struct tcp_md5sig_pool __percpu *tcp_alloc_md5sig_pool(struct sock *sk)
3196 struct tcp_md5sig_pool __percpu *pool;
3200 spin_lock_bh(&tcp_md5sig_pool_lock);
3201 pool = tcp_md5sig_pool;
3202 if (tcp_md5sig_users++ == 0) {
3204 spin_unlock_bh(&tcp_md5sig_pool_lock);
3207 spin_unlock_bh(&tcp_md5sig_pool_lock);
3211 spin_unlock_bh(&tcp_md5sig_pool_lock);
3214 /* we cannot hold spinlock here because this may sleep. */
3215 struct tcp_md5sig_pool __percpu *p;
3217 p = __tcp_alloc_md5sig_pool(sk);
3218 spin_lock_bh(&tcp_md5sig_pool_lock);
3221 spin_unlock_bh(&tcp_md5sig_pool_lock);
3224 pool = tcp_md5sig_pool;
3226 /* oops, it has already been assigned. */
3227 spin_unlock_bh(&tcp_md5sig_pool_lock);
3228 __tcp_free_md5sig_pool(p);
3230 tcp_md5sig_pool = pool = p;
3231 spin_unlock_bh(&tcp_md5sig_pool_lock);
3236 EXPORT_SYMBOL(tcp_alloc_md5sig_pool);
3240 * tcp_get_md5sig_pool - get md5sig_pool for this user
3242 * We use percpu structure, so if we succeed, we exit with preemption
3243 * and BH disabled, to make sure another thread or softirq handling
3244 * wont try to get same context.
3246 struct tcp_md5sig_pool *tcp_get_md5sig_pool(void)
3248 struct tcp_md5sig_pool __percpu *p;
3252 spin_lock(&tcp_md5sig_pool_lock);
3253 p = tcp_md5sig_pool;
3256 spin_unlock(&tcp_md5sig_pool_lock);
3259 return this_cpu_ptr(p);
3264 EXPORT_SYMBOL(tcp_get_md5sig_pool);
3266 void tcp_put_md5sig_pool(void)
3269 tcp_free_md5sig_pool();
3271 EXPORT_SYMBOL(tcp_put_md5sig_pool);
3273 int tcp_md5_hash_header(struct tcp_md5sig_pool *hp,
3274 const struct tcphdr *th)
3276 struct scatterlist sg;
3280 /* We are not allowed to change tcphdr, make a local copy */
3281 memcpy(&hdr, th, sizeof(hdr));
3284 /* options aren't included in the hash */
3285 sg_init_one(&sg, &hdr, sizeof(hdr));
3286 err = crypto_hash_update(&hp->md5_desc, &sg, sizeof(hdr));
3289 EXPORT_SYMBOL(tcp_md5_hash_header);
3291 int tcp_md5_hash_skb_data(struct tcp_md5sig_pool *hp,
3292 const struct sk_buff *skb, unsigned int header_len)
3294 struct scatterlist sg;
3295 const struct tcphdr *tp = tcp_hdr(skb);
3296 struct hash_desc *desc = &hp->md5_desc;
3298 const unsigned int head_data_len = skb_headlen(skb) > header_len ?
3299 skb_headlen(skb) - header_len : 0;
3300 const struct skb_shared_info *shi = skb_shinfo(skb);
3301 struct sk_buff *frag_iter;
3303 sg_init_table(&sg, 1);
3305 sg_set_buf(&sg, ((u8 *) tp) + header_len, head_data_len);
3306 if (crypto_hash_update(desc, &sg, head_data_len))
3309 for (i = 0; i < shi->nr_frags; ++i) {
3310 const struct skb_frag_struct *f = &shi->frags[i];
3311 unsigned int offset = f->page_offset;
3312 struct page *page = skb_frag_page(f) + (offset >> PAGE_SHIFT);
3314 sg_set_page(&sg, page, skb_frag_size(f),
3315 offset_in_page(offset));
3316 if (crypto_hash_update(desc, &sg, skb_frag_size(f)))
3320 skb_walk_frags(skb, frag_iter)
3321 if (tcp_md5_hash_skb_data(hp, frag_iter, 0))
3326 EXPORT_SYMBOL(tcp_md5_hash_skb_data);
3328 int tcp_md5_hash_key(struct tcp_md5sig_pool *hp, const struct tcp_md5sig_key *key)
3330 struct scatterlist sg;
3332 sg_init_one(&sg, key->key, key->keylen);
3333 return crypto_hash_update(&hp->md5_desc, &sg, key->keylen);
3335 EXPORT_SYMBOL(tcp_md5_hash_key);
3339 void tcp_done(struct sock *sk)
3341 struct request_sock *req = tcp_sk(sk)->fastopen_rsk;
3343 if (sk->sk_state == TCP_SYN_SENT || sk->sk_state == TCP_SYN_RECV)
3344 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_ATTEMPTFAILS);
3346 tcp_set_state(sk, TCP_CLOSE);
3347 tcp_clear_xmit_timers(sk);
3349 reqsk_fastopen_remove(sk, req, false);
3351 sk->sk_shutdown = SHUTDOWN_MASK;
3353 if (!sock_flag(sk, SOCK_DEAD))
3354 sk->sk_state_change(sk);
3356 inet_csk_destroy_sock(sk);
3358 EXPORT_SYMBOL_GPL(tcp_done);
3360 extern struct tcp_congestion_ops tcp_reno;
3362 static __initdata unsigned long thash_entries;
3363 static int __init set_thash_entries(char *str)
3370 ret = kstrtoul(str, 0, &thash_entries);
3376 __setup("thash_entries=", set_thash_entries);
3378 void tcp_init_mem(struct net *net)
3380 unsigned long limit = nr_free_buffer_pages() / 8;
3381 limit = max(limit, 128UL);
3382 net->ipv4.sysctl_tcp_mem[0] = limit / 4 * 3;
3383 net->ipv4.sysctl_tcp_mem[1] = limit;
3384 net->ipv4.sysctl_tcp_mem[2] = net->ipv4.sysctl_tcp_mem[0] * 2;
3387 void __init tcp_init(void)
3389 struct sk_buff *skb = NULL;
3390 unsigned long limit;
3391 int max_rshare, max_wshare, cnt;
3394 BUILD_BUG_ON(sizeof(struct tcp_skb_cb) > sizeof(skb->cb));
3396 percpu_counter_init(&tcp_sockets_allocated, 0);
3397 percpu_counter_init(&tcp_orphan_count, 0);
3398 tcp_hashinfo.bind_bucket_cachep =
3399 kmem_cache_create("tcp_bind_bucket",
3400 sizeof(struct inet_bind_bucket), 0,
3401 SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL);
3403 /* Size and allocate the main established and bind bucket
3406 * The methodology is similar to that of the buffer cache.
3408 tcp_hashinfo.ehash =
3409 alloc_large_system_hash("TCP established",
3410 sizeof(struct inet_ehash_bucket),
3412 17, /* one slot per 128 KB of memory */
3415 &tcp_hashinfo.ehash_mask,
3417 thash_entries ? 0 : 512 * 1024);
3418 for (i = 0; i <= tcp_hashinfo.ehash_mask; i++) {
3419 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].chain, i);
3420 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].twchain, i);
3422 if (inet_ehash_locks_alloc(&tcp_hashinfo))
3423 panic("TCP: failed to alloc ehash_locks");
3424 tcp_hashinfo.bhash =
3425 alloc_large_system_hash("TCP bind",
3426 sizeof(struct inet_bind_hashbucket),
3427 tcp_hashinfo.ehash_mask + 1,
3428 17, /* one slot per 128 KB of memory */
3430 &tcp_hashinfo.bhash_size,
3434 tcp_hashinfo.bhash_size = 1U << tcp_hashinfo.bhash_size;
3435 for (i = 0; i < tcp_hashinfo.bhash_size; i++) {
3436 spin_lock_init(&tcp_hashinfo.bhash[i].lock);
3437 INIT_HLIST_HEAD(&tcp_hashinfo.bhash[i].chain);
3441 cnt = tcp_hashinfo.ehash_mask + 1;
3443 tcp_death_row.sysctl_max_tw_buckets = cnt / 2;
3444 sysctl_tcp_max_orphans = cnt / 2;
3445 sysctl_max_syn_backlog = max(128, cnt / 256);
3447 tcp_init_mem(&init_net);
3448 /* Set per-socket limits to no more than 1/128 the pressure threshold */
3449 limit = nr_free_buffer_pages() << (PAGE_SHIFT - 7);
3450 max_wshare = min(4UL*1024*1024, limit);
3451 max_rshare = min(6UL*1024*1024, limit);
3453 sysctl_tcp_wmem[0] = SK_MEM_QUANTUM;
3454 sysctl_tcp_wmem[1] = 16*1024;
3455 sysctl_tcp_wmem[2] = max(64*1024, max_wshare);
3457 sysctl_tcp_rmem[0] = SK_MEM_QUANTUM;
3458 sysctl_tcp_rmem[1] = 87380;
3459 sysctl_tcp_rmem[2] = max(87380, max_rshare);
3461 pr_info("Hash tables configured (established %u bind %u)\n",
3462 tcp_hashinfo.ehash_mask + 1, tcp_hashinfo.bhash_size);
3466 tcp_register_congestion_control(&tcp_reno);