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 <crypto/hash.h>
251 #include <linux/kernel.h>
252 #include <linux/module.h>
253 #include <linux/types.h>
254 #include <linux/fcntl.h>
255 #include <linux/poll.h>
256 #include <linux/inet_diag.h>
257 #include <linux/init.h>
258 #include <linux/fs.h>
259 #include <linux/skbuff.h>
260 #include <linux/scatterlist.h>
261 #include <linux/splice.h>
262 #include <linux/net.h>
263 #include <linux/socket.h>
264 #include <linux/random.h>
265 #include <linux/bootmem.h>
266 #include <linux/highmem.h>
267 #include <linux/swap.h>
268 #include <linux/cache.h>
269 #include <linux/err.h>
270 #include <linux/time.h>
271 #include <linux/slab.h>
272 #include <linux/errqueue.h>
273 #include <linux/static_key.h>
275 #include <net/icmp.h>
276 #include <net/inet_common.h>
278 #include <net/xfrm.h>
280 #include <net/sock.h>
282 #include <linux/uaccess.h>
283 #include <asm/ioctls.h>
284 #include <net/busy_poll.h>
286 struct percpu_counter tcp_orphan_count;
287 EXPORT_SYMBOL_GPL(tcp_orphan_count);
289 long sysctl_tcp_mem[3] __read_mostly;
290 EXPORT_SYMBOL(sysctl_tcp_mem);
292 atomic_long_t tcp_memory_allocated; /* Current allocated memory. */
293 EXPORT_SYMBOL(tcp_memory_allocated);
295 #if IS_ENABLED(CONFIG_SMC)
296 DEFINE_STATIC_KEY_FALSE(tcp_have_smc);
297 EXPORT_SYMBOL(tcp_have_smc);
301 * Current number of TCP sockets.
303 struct percpu_counter tcp_sockets_allocated;
304 EXPORT_SYMBOL(tcp_sockets_allocated);
309 struct tcp_splice_state {
310 struct pipe_inode_info *pipe;
316 * Pressure flag: try to collapse.
317 * Technical note: it is used by multiple contexts non atomically.
318 * All the __sk_mem_schedule() is of this nature: accounting
319 * is strict, actions are advisory and have some latency.
321 unsigned long tcp_memory_pressure __read_mostly;
322 EXPORT_SYMBOL_GPL(tcp_memory_pressure);
324 void tcp_enter_memory_pressure(struct sock *sk)
328 if (READ_ONCE(tcp_memory_pressure))
334 if (!cmpxchg(&tcp_memory_pressure, 0, val))
335 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMEMORYPRESSURES);
337 EXPORT_SYMBOL_GPL(tcp_enter_memory_pressure);
339 void tcp_leave_memory_pressure(struct sock *sk)
343 if (!READ_ONCE(tcp_memory_pressure))
345 val = xchg(&tcp_memory_pressure, 0);
347 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPMEMORYPRESSURESCHRONO,
348 jiffies_to_msecs(jiffies - val));
350 EXPORT_SYMBOL_GPL(tcp_leave_memory_pressure);
352 /* Convert seconds to retransmits based on initial and max timeout */
353 static u8 secs_to_retrans(int seconds, int timeout, int rto_max)
358 int period = timeout;
361 while (seconds > period && res < 255) {
364 if (timeout > rto_max)
372 /* Convert retransmits to seconds based on initial and max timeout */
373 static int retrans_to_secs(u8 retrans, int timeout, int rto_max)
381 if (timeout > rto_max)
389 static u64 tcp_compute_delivery_rate(const struct tcp_sock *tp)
391 u32 rate = READ_ONCE(tp->rate_delivered);
392 u32 intv = READ_ONCE(tp->rate_interval_us);
396 rate64 = (u64)rate * tp->mss_cache * USEC_PER_SEC;
397 do_div(rate64, intv);
402 /* Address-family independent initialization for a tcp_sock.
404 * NOTE: A lot of things set to zero explicitly by call to
405 * sk_alloc() so need not be done here.
407 void tcp_init_sock(struct sock *sk)
409 struct inet_connection_sock *icsk = inet_csk(sk);
410 struct tcp_sock *tp = tcp_sk(sk);
412 tp->out_of_order_queue = RB_ROOT;
413 sk->tcp_rtx_queue = RB_ROOT;
414 tcp_init_xmit_timers(sk);
415 INIT_LIST_HEAD(&tp->tsq_node);
416 INIT_LIST_HEAD(&tp->tsorted_sent_queue);
418 icsk->icsk_rto = TCP_TIMEOUT_INIT;
419 tp->mdev_us = jiffies_to_usecs(TCP_TIMEOUT_INIT);
420 minmax_reset(&tp->rtt_min, tcp_jiffies32, ~0U);
422 /* So many TCP implementations out there (incorrectly) count the
423 * initial SYN frame in their delayed-ACK and congestion control
424 * algorithms that we must have the following bandaid to talk
425 * efficiently to them. -DaveM
427 tp->snd_cwnd = TCP_INIT_CWND;
429 /* There's a bubble in the pipe until at least the first ACK. */
430 tp->app_limited = ~0U;
432 /* See draft-stevens-tcpca-spec-01 for discussion of the
433 * initialization of these values.
435 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
436 tp->snd_cwnd_clamp = ~0;
437 tp->mss_cache = TCP_MSS_DEFAULT;
439 tp->reordering = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_reordering);
440 tcp_assign_congestion_control(sk);
443 tp->rack.reo_wnd_steps = 1;
445 sk->sk_state = TCP_CLOSE;
447 sk->sk_write_space = sk_stream_write_space;
448 sock_set_flag(sk, SOCK_USE_WRITE_QUEUE);
450 icsk->icsk_sync_mss = tcp_sync_mss;
452 sk->sk_sndbuf = sock_net(sk)->ipv4.sysctl_tcp_wmem[1];
453 sk->sk_rcvbuf = sock_net(sk)->ipv4.sysctl_tcp_rmem[1];
455 sk_sockets_allocated_inc(sk);
456 sk->sk_route_forced_caps = NETIF_F_GSO;
458 EXPORT_SYMBOL(tcp_init_sock);
460 void tcp_init_transfer(struct sock *sk, int bpf_op)
462 struct inet_connection_sock *icsk = inet_csk(sk);
465 icsk->icsk_af_ops->rebuild_header(sk);
466 tcp_init_metrics(sk);
467 tcp_call_bpf(sk, bpf_op, 0, NULL);
468 tcp_init_congestion_control(sk);
469 tcp_init_buffer_space(sk);
472 static void tcp_tx_timestamp(struct sock *sk, u16 tsflags)
474 struct sk_buff *skb = tcp_write_queue_tail(sk);
476 if (tsflags && skb) {
477 struct skb_shared_info *shinfo = skb_shinfo(skb);
478 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
480 sock_tx_timestamp(sk, tsflags, &shinfo->tx_flags);
481 if (tsflags & SOF_TIMESTAMPING_TX_ACK)
482 tcb->txstamp_ack = 1;
483 if (tsflags & SOF_TIMESTAMPING_TX_RECORD_MASK)
484 shinfo->tskey = TCP_SKB_CB(skb)->seq + skb->len - 1;
488 static inline bool tcp_stream_is_readable(const struct tcp_sock *tp,
489 int target, struct sock *sk)
491 int avail = READ_ONCE(tp->rcv_nxt) - READ_ONCE(tp->copied_seq);
496 if (tcp_rmem_pressure(sk))
498 if (tcp_receive_window(tp) <= inet_csk(sk)->icsk_ack.rcv_mss)
501 if (sk->sk_prot->stream_memory_read)
502 return sk->sk_prot->stream_memory_read(sk);
507 * Wait for a TCP event.
509 * Note that we don't need to lock the socket, as the upper poll layers
510 * take care of normal races (between the test and the event) and we don't
511 * go look at any of the socket buffers directly.
513 __poll_t tcp_poll(struct file *file, struct socket *sock, poll_table *wait)
516 struct sock *sk = sock->sk;
517 const struct tcp_sock *tp = tcp_sk(sk);
520 sock_poll_wait(file, sock, wait);
522 state = inet_sk_state_load(sk);
523 if (state == TCP_LISTEN)
524 return inet_csk_listen_poll(sk);
526 /* Socket is not locked. We are protected from async events
527 * by poll logic and correct handling of state changes
528 * made by other threads is impossible in any case.
534 * EPOLLHUP is certainly not done right. But poll() doesn't
535 * have a notion of HUP in just one direction, and for a
536 * socket the read side is more interesting.
538 * Some poll() documentation says that EPOLLHUP is incompatible
539 * with the EPOLLOUT/POLLWR flags, so somebody should check this
540 * all. But careful, it tends to be safer to return too many
541 * bits than too few, and you can easily break real applications
542 * if you don't tell them that something has hung up!
546 * Check number 1. EPOLLHUP is _UNMASKABLE_ event (see UNIX98 and
547 * our fs/select.c). It means that after we received EOF,
548 * poll always returns immediately, making impossible poll() on write()
549 * in state CLOSE_WAIT. One solution is evident --- to set EPOLLHUP
550 * if and only if shutdown has been made in both directions.
551 * Actually, it is interesting to look how Solaris and DUX
552 * solve this dilemma. I would prefer, if EPOLLHUP were maskable,
553 * then we could set it on SND_SHUTDOWN. BTW examples given
554 * in Stevens' books assume exactly this behaviour, it explains
555 * why EPOLLHUP is incompatible with EPOLLOUT. --ANK
557 * NOTE. Check for TCP_CLOSE is added. The goal is to prevent
558 * blocking on fresh not-connected or disconnected socket. --ANK
560 if (sk->sk_shutdown == SHUTDOWN_MASK || state == TCP_CLOSE)
562 if (sk->sk_shutdown & RCV_SHUTDOWN)
563 mask |= EPOLLIN | EPOLLRDNORM | EPOLLRDHUP;
565 /* Connected or passive Fast Open socket? */
566 if (state != TCP_SYN_SENT &&
567 (state != TCP_SYN_RECV || tp->fastopen_rsk)) {
568 int target = sock_rcvlowat(sk, 0, INT_MAX);
570 if (tp->urg_seq == READ_ONCE(tp->copied_seq) &&
571 !sock_flag(sk, SOCK_URGINLINE) &&
575 if (tcp_stream_is_readable(tp, target, sk))
576 mask |= EPOLLIN | EPOLLRDNORM;
578 if (!(sk->sk_shutdown & SEND_SHUTDOWN)) {
579 if (sk_stream_is_writeable(sk)) {
580 mask |= EPOLLOUT | EPOLLWRNORM;
581 } else { /* send SIGIO later */
582 sk_set_bit(SOCKWQ_ASYNC_NOSPACE, sk);
583 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
585 /* Race breaker. If space is freed after
586 * wspace test but before the flags are set,
587 * IO signal will be lost. Memory barrier
588 * pairs with the input side.
590 smp_mb__after_atomic();
591 if (sk_stream_is_writeable(sk))
592 mask |= EPOLLOUT | EPOLLWRNORM;
595 mask |= EPOLLOUT | EPOLLWRNORM;
597 if (tp->urg_data & TCP_URG_VALID)
599 } else if (state == TCP_SYN_SENT && inet_sk(sk)->defer_connect) {
600 /* Active TCP fastopen socket with defer_connect
601 * Return EPOLLOUT so application can call write()
602 * in order for kernel to generate SYN+data
604 mask |= EPOLLOUT | EPOLLWRNORM;
606 /* This barrier is coupled with smp_wmb() in tcp_reset() */
608 if (sk->sk_err || !skb_queue_empty_lockless(&sk->sk_error_queue))
613 EXPORT_SYMBOL(tcp_poll);
615 int tcp_ioctl(struct sock *sk, int cmd, unsigned long arg)
617 struct tcp_sock *tp = tcp_sk(sk);
623 if (sk->sk_state == TCP_LISTEN)
626 slow = lock_sock_fast(sk);
628 unlock_sock_fast(sk, slow);
631 answ = tp->urg_data && tp->urg_seq == READ_ONCE(tp->copied_seq);
634 if (sk->sk_state == TCP_LISTEN)
637 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
640 answ = READ_ONCE(tp->write_seq) - tp->snd_una;
643 if (sk->sk_state == TCP_LISTEN)
646 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
649 answ = READ_ONCE(tp->write_seq) - tp->snd_nxt;
655 return put_user(answ, (int __user *)arg);
657 EXPORT_SYMBOL(tcp_ioctl);
659 static inline void tcp_mark_push(struct tcp_sock *tp, struct sk_buff *skb)
661 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
662 tp->pushed_seq = tp->write_seq;
665 static inline bool forced_push(const struct tcp_sock *tp)
667 return after(tp->write_seq, tp->pushed_seq + (tp->max_window >> 1));
670 static void skb_entail(struct sock *sk, struct sk_buff *skb)
672 struct tcp_sock *tp = tcp_sk(sk);
673 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
676 tcb->seq = tcb->end_seq = tp->write_seq;
677 tcb->tcp_flags = TCPHDR_ACK;
679 __skb_header_release(skb);
680 tcp_add_write_queue_tail(sk, skb);
681 sk->sk_wmem_queued += skb->truesize;
682 sk_mem_charge(sk, skb->truesize);
683 if (tp->nonagle & TCP_NAGLE_PUSH)
684 tp->nonagle &= ~TCP_NAGLE_PUSH;
686 tcp_slow_start_after_idle_check(sk);
689 static inline void tcp_mark_urg(struct tcp_sock *tp, int flags)
692 tp->snd_up = tp->write_seq;
695 /* If a not yet filled skb is pushed, do not send it if
696 * we have data packets in Qdisc or NIC queues :
697 * Because TX completion will happen shortly, it gives a chance
698 * to coalesce future sendmsg() payload into this skb, without
699 * need for a timer, and with no latency trade off.
700 * As packets containing data payload have a bigger truesize
701 * than pure acks (dataless) packets, the last checks prevent
702 * autocorking if we only have an ACK in Qdisc/NIC queues,
703 * or if TX completion was delayed after we processed ACK packet.
705 static bool tcp_should_autocork(struct sock *sk, struct sk_buff *skb,
708 return skb->len < size_goal &&
709 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_autocorking) &&
710 !tcp_rtx_queue_empty(sk) &&
711 refcount_read(&sk->sk_wmem_alloc) > skb->truesize;
714 static void tcp_push(struct sock *sk, int flags, int mss_now,
715 int nonagle, int size_goal)
717 struct tcp_sock *tp = tcp_sk(sk);
720 skb = tcp_write_queue_tail(sk);
723 if (!(flags & MSG_MORE) || forced_push(tp))
724 tcp_mark_push(tp, skb);
726 tcp_mark_urg(tp, flags);
728 if (tcp_should_autocork(sk, skb, size_goal)) {
730 /* avoid atomic op if TSQ_THROTTLED bit is already set */
731 if (!test_bit(TSQ_THROTTLED, &sk->sk_tsq_flags)) {
732 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPAUTOCORKING);
733 set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags);
735 /* It is possible TX completion already happened
736 * before we set TSQ_THROTTLED.
738 if (refcount_read(&sk->sk_wmem_alloc) > skb->truesize)
742 if (flags & MSG_MORE)
743 nonagle = TCP_NAGLE_CORK;
745 __tcp_push_pending_frames(sk, mss_now, nonagle);
748 static int tcp_splice_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb,
749 unsigned int offset, size_t len)
751 struct tcp_splice_state *tss = rd_desc->arg.data;
754 ret = skb_splice_bits(skb, skb->sk, offset, tss->pipe,
755 min(rd_desc->count, len), tss->flags);
757 rd_desc->count -= ret;
761 static int __tcp_splice_read(struct sock *sk, struct tcp_splice_state *tss)
763 /* Store TCP splice context information in read_descriptor_t. */
764 read_descriptor_t rd_desc = {
769 return tcp_read_sock(sk, &rd_desc, tcp_splice_data_recv);
773 * tcp_splice_read - splice data from TCP socket to a pipe
774 * @sock: socket to splice from
775 * @ppos: position (not valid)
776 * @pipe: pipe to splice to
777 * @len: number of bytes to splice
778 * @flags: splice modifier flags
781 * Will read pages from given socket and fill them into a pipe.
784 ssize_t tcp_splice_read(struct socket *sock, loff_t *ppos,
785 struct pipe_inode_info *pipe, size_t len,
788 struct sock *sk = sock->sk;
789 struct tcp_splice_state tss = {
798 sock_rps_record_flow(sk);
800 * We can't seek on a socket input
809 timeo = sock_rcvtimeo(sk, sock->file->f_flags & O_NONBLOCK);
811 ret = __tcp_splice_read(sk, &tss);
817 if (sock_flag(sk, SOCK_DONE))
820 ret = sock_error(sk);
823 if (sk->sk_shutdown & RCV_SHUTDOWN)
825 if (sk->sk_state == TCP_CLOSE) {
827 * This occurs when user tries to read
828 * from never connected socket.
837 /* if __tcp_splice_read() got nothing while we have
838 * an skb in receive queue, we do not want to loop.
839 * This might happen with URG data.
841 if (!skb_queue_empty(&sk->sk_receive_queue))
843 sk_wait_data(sk, &timeo, NULL);
844 if (signal_pending(current)) {
845 ret = sock_intr_errno(timeo);
858 if (sk->sk_err || sk->sk_state == TCP_CLOSE ||
859 (sk->sk_shutdown & RCV_SHUTDOWN) ||
860 signal_pending(current))
871 EXPORT_SYMBOL(tcp_splice_read);
873 struct sk_buff *sk_stream_alloc_skb(struct sock *sk, int size, gfp_t gfp,
878 /* The TCP header must be at least 32-bit aligned. */
879 size = ALIGN(size, 4);
881 if (unlikely(tcp_under_memory_pressure(sk)))
882 sk_mem_reclaim_partial(sk);
884 skb = alloc_skb_fclone(size + sk->sk_prot->max_header, gfp);
888 if (force_schedule) {
889 mem_scheduled = true;
890 sk_forced_mem_schedule(sk, skb->truesize);
892 mem_scheduled = sk_wmem_schedule(sk, skb->truesize);
894 if (likely(mem_scheduled)) {
895 skb_reserve(skb, sk->sk_prot->max_header);
897 * Make sure that we have exactly size bytes
898 * available to the caller, no more, no less.
900 skb->reserved_tailroom = skb->end - skb->tail - size;
901 INIT_LIST_HEAD(&skb->tcp_tsorted_anchor);
906 sk->sk_prot->enter_memory_pressure(sk);
907 sk_stream_moderate_sndbuf(sk);
912 static unsigned int tcp_xmit_size_goal(struct sock *sk, u32 mss_now,
915 struct tcp_sock *tp = tcp_sk(sk);
916 u32 new_size_goal, size_goal;
921 /* Note : tcp_tso_autosize() will eventually split this later */
922 new_size_goal = sk->sk_gso_max_size - 1 - MAX_TCP_HEADER;
923 new_size_goal = tcp_bound_to_half_wnd(tp, new_size_goal);
925 /* We try hard to avoid divides here */
926 size_goal = tp->gso_segs * mss_now;
927 if (unlikely(new_size_goal < size_goal ||
928 new_size_goal >= size_goal + mss_now)) {
929 tp->gso_segs = min_t(u16, new_size_goal / mss_now,
930 sk->sk_gso_max_segs);
931 size_goal = tp->gso_segs * mss_now;
934 return max(size_goal, mss_now);
937 static int tcp_send_mss(struct sock *sk, int *size_goal, int flags)
941 mss_now = tcp_current_mss(sk);
942 *size_goal = tcp_xmit_size_goal(sk, mss_now, !(flags & MSG_OOB));
947 /* In some cases, both sendpage() and sendmsg() could have added
948 * an skb to the write queue, but failed adding payload on it.
949 * We need to remove it to consume less memory, but more
950 * importantly be able to generate EPOLLOUT for Edge Trigger epoll()
953 static void tcp_remove_empty_skb(struct sock *sk, struct sk_buff *skb)
955 if (skb && TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq) {
956 tcp_unlink_write_queue(skb, sk);
957 if (tcp_write_queue_empty(sk))
958 tcp_chrono_stop(sk, TCP_CHRONO_BUSY);
959 sk_wmem_free_skb(sk, skb);
963 ssize_t do_tcp_sendpages(struct sock *sk, struct page *page, int offset,
964 size_t size, int flags)
966 struct tcp_sock *tp = tcp_sk(sk);
967 int mss_now, size_goal;
970 long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
972 /* Wait for a connection to finish. One exception is TCP Fast Open
973 * (passive side) where data is allowed to be sent before a connection
974 * is fully established.
976 if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) &&
977 !tcp_passive_fastopen(sk)) {
978 err = sk_stream_wait_connect(sk, &timeo);
983 sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
985 mss_now = tcp_send_mss(sk, &size_goal, flags);
989 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
993 struct sk_buff *skb = tcp_write_queue_tail(sk);
997 if (!skb || (copy = size_goal - skb->len) <= 0 ||
998 !tcp_skb_can_collapse_to(skb)) {
1000 if (!sk_stream_memory_free(sk))
1001 goto wait_for_sndbuf;
1003 skb = sk_stream_alloc_skb(sk, 0, sk->sk_allocation,
1004 tcp_rtx_and_write_queues_empty(sk));
1006 goto wait_for_memory;
1008 skb_entail(sk, skb);
1015 i = skb_shinfo(skb)->nr_frags;
1016 can_coalesce = skb_can_coalesce(skb, i, page, offset);
1017 if (!can_coalesce && i >= sysctl_max_skb_frags) {
1018 tcp_mark_push(tp, skb);
1021 if (!sk_wmem_schedule(sk, copy))
1022 goto wait_for_memory;
1025 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1028 skb_fill_page_desc(skb, i, page, offset, copy);
1031 if (!(flags & MSG_NO_SHARED_FRAGS))
1032 skb_shinfo(skb)->tx_flags |= SKBTX_SHARED_FRAG;
1035 skb->data_len += copy;
1036 skb->truesize += copy;
1037 sk->sk_wmem_queued += copy;
1038 sk_mem_charge(sk, copy);
1039 skb->ip_summed = CHECKSUM_PARTIAL;
1040 WRITE_ONCE(tp->write_seq, tp->write_seq + copy);
1041 TCP_SKB_CB(skb)->end_seq += copy;
1042 tcp_skb_pcount_set(skb, 0);
1045 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
1053 if (skb->len < size_goal || (flags & MSG_OOB))
1056 if (forced_push(tp)) {
1057 tcp_mark_push(tp, skb);
1058 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
1059 } else if (skb == tcp_send_head(sk))
1060 tcp_push_one(sk, mss_now);
1064 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1066 tcp_push(sk, flags & ~MSG_MORE, mss_now,
1067 TCP_NAGLE_PUSH, size_goal);
1069 err = sk_stream_wait_memory(sk, &timeo);
1073 mss_now = tcp_send_mss(sk, &size_goal, flags);
1078 tcp_tx_timestamp(sk, sk->sk_tsflags);
1079 if (!(flags & MSG_SENDPAGE_NOTLAST))
1080 tcp_push(sk, flags, mss_now, tp->nonagle, size_goal);
1085 tcp_remove_empty_skb(sk, tcp_write_queue_tail(sk));
1089 /* make sure we wake any epoll edge trigger waiter */
1090 if (unlikely(tcp_rtx_and_write_queues_empty(sk) && err == -EAGAIN)) {
1091 sk->sk_write_space(sk);
1092 tcp_chrono_stop(sk, TCP_CHRONO_SNDBUF_LIMITED);
1094 return sk_stream_error(sk, flags, err);
1096 EXPORT_SYMBOL_GPL(do_tcp_sendpages);
1098 int tcp_sendpage_locked(struct sock *sk, struct page *page, int offset,
1099 size_t size, int flags)
1101 if (!(sk->sk_route_caps & NETIF_F_SG))
1102 return sock_no_sendpage_locked(sk, page, offset, size, flags);
1104 tcp_rate_check_app_limited(sk); /* is sending application-limited? */
1106 return do_tcp_sendpages(sk, page, offset, size, flags);
1108 EXPORT_SYMBOL_GPL(tcp_sendpage_locked);
1110 int tcp_sendpage(struct sock *sk, struct page *page, int offset,
1111 size_t size, int flags)
1116 ret = tcp_sendpage_locked(sk, page, offset, size, flags);
1121 EXPORT_SYMBOL(tcp_sendpage);
1123 /* Do not bother using a page frag for very small frames.
1124 * But use this heuristic only for the first skb in write queue.
1126 * Having no payload in skb->head allows better SACK shifting
1127 * in tcp_shift_skb_data(), reducing sack/rack overhead, because
1128 * write queue has less skbs.
1129 * Each skb can hold up to MAX_SKB_FRAGS * 32Kbytes, or ~0.5 MB.
1130 * This also speeds up tso_fragment(), since it wont fallback
1131 * to tcp_fragment().
1133 static int linear_payload_sz(bool first_skb)
1136 return SKB_WITH_OVERHEAD(2048 - MAX_TCP_HEADER);
1140 static int select_size(bool first_skb, bool zc)
1144 return linear_payload_sz(first_skb);
1147 void tcp_free_fastopen_req(struct tcp_sock *tp)
1149 if (tp->fastopen_req) {
1150 kfree(tp->fastopen_req);
1151 tp->fastopen_req = NULL;
1155 static int tcp_sendmsg_fastopen(struct sock *sk, struct msghdr *msg,
1156 int *copied, size_t size)
1158 struct tcp_sock *tp = tcp_sk(sk);
1159 struct inet_sock *inet = inet_sk(sk);
1160 struct sockaddr *uaddr = msg->msg_name;
1163 if (!(READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_fastopen) &
1164 TFO_CLIENT_ENABLE) ||
1165 (uaddr && msg->msg_namelen >= sizeof(uaddr->sa_family) &&
1166 uaddr->sa_family == AF_UNSPEC))
1168 if (tp->fastopen_req)
1169 return -EALREADY; /* Another Fast Open is in progress */
1171 tp->fastopen_req = kzalloc(sizeof(struct tcp_fastopen_request),
1173 if (unlikely(!tp->fastopen_req))
1175 tp->fastopen_req->data = msg;
1176 tp->fastopen_req->size = size;
1178 if (inet->defer_connect) {
1179 err = tcp_connect(sk);
1180 /* Same failure procedure as in tcp_v4/6_connect */
1182 tcp_set_state(sk, TCP_CLOSE);
1183 inet->inet_dport = 0;
1184 sk->sk_route_caps = 0;
1187 flags = (msg->msg_flags & MSG_DONTWAIT) ? O_NONBLOCK : 0;
1188 err = __inet_stream_connect(sk->sk_socket, uaddr,
1189 msg->msg_namelen, flags, 1);
1190 /* fastopen_req could already be freed in __inet_stream_connect
1191 * if the connection times out or gets rst
1193 if (tp->fastopen_req) {
1194 *copied = tp->fastopen_req->copied;
1195 tcp_free_fastopen_req(tp);
1196 inet->defer_connect = 0;
1201 int tcp_sendmsg_locked(struct sock *sk, struct msghdr *msg, size_t size)
1203 struct tcp_sock *tp = tcp_sk(sk);
1204 struct ubuf_info *uarg = NULL;
1205 struct sk_buff *skb;
1206 struct sockcm_cookie sockc;
1207 int flags, err, copied = 0;
1208 int mss_now = 0, size_goal, copied_syn = 0;
1209 bool process_backlog = false;
1213 flags = msg->msg_flags;
1215 if (flags & MSG_ZEROCOPY && size && sock_flag(sk, SOCK_ZEROCOPY)) {
1216 if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) {
1221 skb = tcp_write_queue_tail(sk);
1222 uarg = sock_zerocopy_realloc(sk, size, skb_zcopy(skb));
1228 zc = sk->sk_route_caps & NETIF_F_SG;
1233 if (unlikely(flags & MSG_FASTOPEN || inet_sk(sk)->defer_connect) &&
1235 err = tcp_sendmsg_fastopen(sk, msg, &copied_syn, size);
1236 if (err == -EINPROGRESS && copied_syn > 0)
1242 timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
1244 tcp_rate_check_app_limited(sk); /* is sending application-limited? */
1246 /* Wait for a connection to finish. One exception is TCP Fast Open
1247 * (passive side) where data is allowed to be sent before a connection
1248 * is fully established.
1250 if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) &&
1251 !tcp_passive_fastopen(sk)) {
1252 err = sk_stream_wait_connect(sk, &timeo);
1257 if (unlikely(tp->repair)) {
1258 if (tp->repair_queue == TCP_RECV_QUEUE) {
1259 copied = tcp_send_rcvq(sk, msg, size);
1264 if (tp->repair_queue == TCP_NO_QUEUE)
1267 /* 'common' sending to sendq */
1270 sockcm_init(&sockc, sk);
1271 if (msg->msg_controllen) {
1272 err = sock_cmsg_send(sk, msg, &sockc);
1273 if (unlikely(err)) {
1279 /* This should be in poll */
1280 sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
1282 /* Ok commence sending. */
1286 mss_now = tcp_send_mss(sk, &size_goal, flags);
1289 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
1292 while (msg_data_left(msg)) {
1295 skb = tcp_write_queue_tail(sk);
1297 copy = size_goal - skb->len;
1299 if (copy <= 0 || !tcp_skb_can_collapse_to(skb)) {
1304 if (!sk_stream_memory_free(sk))
1305 goto wait_for_sndbuf;
1307 if (process_backlog && sk_flush_backlog(sk)) {
1308 process_backlog = false;
1311 first_skb = tcp_rtx_and_write_queues_empty(sk);
1312 linear = select_size(first_skb, zc);
1313 skb = sk_stream_alloc_skb(sk, linear, sk->sk_allocation,
1316 goto wait_for_memory;
1318 process_backlog = true;
1319 skb->ip_summed = CHECKSUM_PARTIAL;
1321 skb_entail(sk, skb);
1324 /* All packets are restored as if they have
1325 * already been sent. skb_mstamp isn't set to
1326 * avoid wrong rtt estimation.
1329 TCP_SKB_CB(skb)->sacked |= TCPCB_REPAIRED;
1332 /* Try to append data to the end of skb. */
1333 if (copy > msg_data_left(msg))
1334 copy = msg_data_left(msg);
1336 /* Where to copy to? */
1337 if (skb_availroom(skb) > 0 && !zc) {
1338 /* We have some space in skb head. Superb! */
1339 copy = min_t(int, copy, skb_availroom(skb));
1340 err = skb_add_data_nocache(sk, skb, &msg->msg_iter, copy);
1345 int i = skb_shinfo(skb)->nr_frags;
1346 struct page_frag *pfrag = sk_page_frag(sk);
1348 if (!sk_page_frag_refill(sk, pfrag))
1349 goto wait_for_memory;
1351 if (!skb_can_coalesce(skb, i, pfrag->page,
1353 if (i >= sysctl_max_skb_frags) {
1354 tcp_mark_push(tp, skb);
1360 copy = min_t(int, copy, pfrag->size - pfrag->offset);
1362 if (!sk_wmem_schedule(sk, copy))
1363 goto wait_for_memory;
1365 err = skb_copy_to_page_nocache(sk, &msg->msg_iter, skb,
1372 /* Update the skb. */
1374 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1376 skb_fill_page_desc(skb, i, pfrag->page,
1377 pfrag->offset, copy);
1378 page_ref_inc(pfrag->page);
1380 pfrag->offset += copy;
1382 err = skb_zerocopy_iter_stream(sk, skb, msg, copy, uarg);
1383 if (err == -EMSGSIZE || err == -EEXIST) {
1384 tcp_mark_push(tp, skb);
1393 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
1395 WRITE_ONCE(tp->write_seq, tp->write_seq + copy);
1396 TCP_SKB_CB(skb)->end_seq += copy;
1397 tcp_skb_pcount_set(skb, 0);
1400 if (!msg_data_left(msg)) {
1401 if (unlikely(flags & MSG_EOR))
1402 TCP_SKB_CB(skb)->eor = 1;
1406 if (skb->len < size_goal || (flags & MSG_OOB) || unlikely(tp->repair))
1409 if (forced_push(tp)) {
1410 tcp_mark_push(tp, skb);
1411 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
1412 } else if (skb == tcp_send_head(sk))
1413 tcp_push_one(sk, mss_now);
1417 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1420 tcp_push(sk, flags & ~MSG_MORE, mss_now,
1421 TCP_NAGLE_PUSH, size_goal);
1423 err = sk_stream_wait_memory(sk, &timeo);
1427 mss_now = tcp_send_mss(sk, &size_goal, flags);
1432 tcp_tx_timestamp(sk, sockc.tsflags);
1433 tcp_push(sk, flags, mss_now, tp->nonagle, size_goal);
1436 sock_zerocopy_put(uarg);
1437 return copied + copied_syn;
1440 skb = tcp_write_queue_tail(sk);
1442 tcp_remove_empty_skb(sk, skb);
1444 if (copied + copied_syn)
1447 sock_zerocopy_put_abort(uarg);
1448 err = sk_stream_error(sk, flags, err);
1449 /* make sure we wake any epoll edge trigger waiter */
1450 if (unlikely(tcp_rtx_and_write_queues_empty(sk) && err == -EAGAIN)) {
1451 sk->sk_write_space(sk);
1452 tcp_chrono_stop(sk, TCP_CHRONO_SNDBUF_LIMITED);
1456 EXPORT_SYMBOL_GPL(tcp_sendmsg_locked);
1458 int tcp_sendmsg(struct sock *sk, struct msghdr *msg, size_t size)
1463 ret = tcp_sendmsg_locked(sk, msg, size);
1468 EXPORT_SYMBOL(tcp_sendmsg);
1471 * Handle reading urgent data. BSD has very simple semantics for
1472 * this, no blocking and very strange errors 8)
1475 static int tcp_recv_urg(struct sock *sk, struct msghdr *msg, int len, int flags)
1477 struct tcp_sock *tp = tcp_sk(sk);
1479 /* No URG data to read. */
1480 if (sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data ||
1481 tp->urg_data == TCP_URG_READ)
1482 return -EINVAL; /* Yes this is right ! */
1484 if (sk->sk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DONE))
1487 if (tp->urg_data & TCP_URG_VALID) {
1489 char c = tp->urg_data;
1491 if (!(flags & MSG_PEEK))
1492 tp->urg_data = TCP_URG_READ;
1494 /* Read urgent data. */
1495 msg->msg_flags |= MSG_OOB;
1498 if (!(flags & MSG_TRUNC))
1499 err = memcpy_to_msg(msg, &c, 1);
1502 msg->msg_flags |= MSG_TRUNC;
1504 return err ? -EFAULT : len;
1507 if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
1510 /* Fixed the recv(..., MSG_OOB) behaviour. BSD docs and
1511 * the available implementations agree in this case:
1512 * this call should never block, independent of the
1513 * blocking state of the socket.
1514 * Mike <pall@rz.uni-karlsruhe.de>
1519 static int tcp_peek_sndq(struct sock *sk, struct msghdr *msg, int len)
1521 struct sk_buff *skb;
1522 int copied = 0, err = 0;
1524 /* XXX -- need to support SO_PEEK_OFF */
1526 skb_rbtree_walk(skb, &sk->tcp_rtx_queue) {
1527 err = skb_copy_datagram_msg(skb, 0, msg, skb->len);
1533 skb_queue_walk(&sk->sk_write_queue, skb) {
1534 err = skb_copy_datagram_msg(skb, 0, msg, skb->len);
1541 return err ?: copied;
1544 /* Clean up the receive buffer for full frames taken by the user,
1545 * then send an ACK if necessary. COPIED is the number of bytes
1546 * tcp_recvmsg has given to the user so far, it speeds up the
1547 * calculation of whether or not we must ACK for the sake of
1550 static void tcp_cleanup_rbuf(struct sock *sk, int copied)
1552 struct tcp_sock *tp = tcp_sk(sk);
1553 bool time_to_ack = false;
1555 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
1557 WARN(skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq),
1558 "cleanup rbuf bug: copied %X seq %X rcvnxt %X\n",
1559 tp->copied_seq, TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt);
1561 if (inet_csk_ack_scheduled(sk)) {
1562 const struct inet_connection_sock *icsk = inet_csk(sk);
1563 /* Delayed ACKs frequently hit locked sockets during bulk
1565 if (icsk->icsk_ack.blocked ||
1566 /* Once-per-two-segments ACK was not sent by tcp_input.c */
1567 tp->rcv_nxt - tp->rcv_wup > icsk->icsk_ack.rcv_mss ||
1569 * If this read emptied read buffer, we send ACK, if
1570 * connection is not bidirectional, user drained
1571 * receive buffer and there was a small segment
1575 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED2) ||
1576 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED) &&
1577 !icsk->icsk_ack.pingpong)) &&
1578 !atomic_read(&sk->sk_rmem_alloc)))
1582 /* We send an ACK if we can now advertise a non-zero window
1583 * which has been raised "significantly".
1585 * Even if window raised up to infinity, do not send window open ACK
1586 * in states, where we will not receive more. It is useless.
1588 if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) {
1589 __u32 rcv_window_now = tcp_receive_window(tp);
1591 /* Optimize, __tcp_select_window() is not cheap. */
1592 if (2*rcv_window_now <= tp->window_clamp) {
1593 __u32 new_window = __tcp_select_window(sk);
1595 /* Send ACK now, if this read freed lots of space
1596 * in our buffer. Certainly, new_window is new window.
1597 * We can advertise it now, if it is not less than current one.
1598 * "Lots" means "at least twice" here.
1600 if (new_window && new_window >= 2 * rcv_window_now)
1608 static struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off)
1610 struct sk_buff *skb;
1613 while ((skb = skb_peek(&sk->sk_receive_queue)) != NULL) {
1614 offset = seq - TCP_SKB_CB(skb)->seq;
1615 if (unlikely(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
1616 pr_err_once("%s: found a SYN, please report !\n", __func__);
1619 if (offset < skb->len || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)) {
1623 /* This looks weird, but this can happen if TCP collapsing
1624 * splitted a fat GRO packet, while we released socket lock
1625 * in skb_splice_bits()
1627 sk_eat_skb(sk, skb);
1633 * This routine provides an alternative to tcp_recvmsg() for routines
1634 * that would like to handle copying from skbuffs directly in 'sendfile'
1637 * - It is assumed that the socket was locked by the caller.
1638 * - The routine does not block.
1639 * - At present, there is no support for reading OOB data
1640 * or for 'peeking' the socket using this routine
1641 * (although both would be easy to implement).
1643 int tcp_read_sock(struct sock *sk, read_descriptor_t *desc,
1644 sk_read_actor_t recv_actor)
1646 struct sk_buff *skb;
1647 struct tcp_sock *tp = tcp_sk(sk);
1648 u32 seq = tp->copied_seq;
1652 if (sk->sk_state == TCP_LISTEN)
1654 while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1655 if (offset < skb->len) {
1659 len = skb->len - offset;
1660 /* Stop reading if we hit a patch of urgent data */
1662 u32 urg_offset = tp->urg_seq - seq;
1663 if (urg_offset < len)
1668 used = recv_actor(desc, skb, offset, len);
1674 if (WARN_ON_ONCE(used > len))
1680 /* If recv_actor drops the lock (e.g. TCP splice
1681 * receive) the skb pointer might be invalid when
1682 * getting here: tcp_collapse might have deleted it
1683 * while aggregating skbs from the socket queue.
1685 skb = tcp_recv_skb(sk, seq - 1, &offset);
1688 /* TCP coalescing might have appended data to the skb.
1689 * Try to splice more frags
1691 if (offset + 1 != skb->len)
1694 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) {
1695 sk_eat_skb(sk, skb);
1699 sk_eat_skb(sk, skb);
1702 WRITE_ONCE(tp->copied_seq, seq);
1704 WRITE_ONCE(tp->copied_seq, seq);
1706 tcp_rcv_space_adjust(sk);
1708 /* Clean up data we have read: This will do ACK frames. */
1710 tcp_recv_skb(sk, seq, &offset);
1711 tcp_cleanup_rbuf(sk, copied);
1715 EXPORT_SYMBOL(tcp_read_sock);
1717 int tcp_peek_len(struct socket *sock)
1719 return tcp_inq(sock->sk);
1721 EXPORT_SYMBOL(tcp_peek_len);
1723 /* Make sure sk_rcvbuf is big enough to satisfy SO_RCVLOWAT hint */
1724 int tcp_set_rcvlowat(struct sock *sk, int val)
1728 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK)
1729 cap = sk->sk_rcvbuf >> 1;
1731 cap = sock_net(sk)->ipv4.sysctl_tcp_rmem[2] >> 1;
1732 val = min(val, cap);
1733 sk->sk_rcvlowat = val ? : 1;
1735 /* Check if we need to signal EPOLLIN right now */
1738 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK)
1742 if (val > sk->sk_rcvbuf) {
1743 sk->sk_rcvbuf = val;
1744 tcp_sk(sk)->window_clamp = tcp_win_from_space(sk, val);
1748 EXPORT_SYMBOL(tcp_set_rcvlowat);
1751 static const struct vm_operations_struct tcp_vm_ops = {
1754 int tcp_mmap(struct file *file, struct socket *sock,
1755 struct vm_area_struct *vma)
1757 if (vma->vm_flags & (VM_WRITE | VM_EXEC))
1759 vma->vm_flags &= ~(VM_MAYWRITE | VM_MAYEXEC);
1761 /* Instruct vm_insert_page() to not down_read(mmap_sem) */
1762 vma->vm_flags |= VM_MIXEDMAP;
1764 vma->vm_ops = &tcp_vm_ops;
1767 EXPORT_SYMBOL(tcp_mmap);
1769 static int tcp_zerocopy_receive(struct sock *sk,
1770 struct tcp_zerocopy_receive *zc)
1772 unsigned long address = (unsigned long)zc->address;
1773 const skb_frag_t *frags = NULL;
1774 u32 length = 0, seq, offset;
1775 struct vm_area_struct *vma;
1776 struct sk_buff *skb = NULL;
1777 struct tcp_sock *tp;
1780 if (address & (PAGE_SIZE - 1) || address != zc->address)
1783 if (sk->sk_state == TCP_LISTEN)
1786 sock_rps_record_flow(sk);
1788 down_read(¤t->mm->mmap_sem);
1790 vma = find_vma(current->mm, address);
1791 if (!vma || vma->vm_start > address || vma->vm_ops != &tcp_vm_ops) {
1792 up_read(¤t->mm->mmap_sem);
1795 zc->length = min_t(unsigned long, zc->length, vma->vm_end - address);
1798 seq = tp->copied_seq;
1799 zc->length = min_t(u32, zc->length, tcp_inq(sk));
1800 zc->length &= ~(PAGE_SIZE - 1);
1802 zap_page_range(vma, address, zc->length);
1804 zc->recv_skip_hint = 0;
1806 while (length + PAGE_SIZE <= zc->length) {
1807 if (zc->recv_skip_hint < PAGE_SIZE) {
1810 offset = seq - TCP_SKB_CB(skb)->seq;
1812 skb = tcp_recv_skb(sk, seq, &offset);
1815 zc->recv_skip_hint = skb->len - offset;
1816 offset -= skb_headlen(skb);
1817 if ((int)offset < 0 || skb_has_frag_list(skb))
1819 frags = skb_shinfo(skb)->frags;
1821 if (frags->size > offset)
1823 offset -= frags->size;
1827 if (frags->size != PAGE_SIZE || frags->page_offset)
1829 ret = vm_insert_page(vma, address + length,
1830 skb_frag_page(frags));
1833 length += PAGE_SIZE;
1835 zc->recv_skip_hint -= PAGE_SIZE;
1839 up_read(¤t->mm->mmap_sem);
1841 WRITE_ONCE(tp->copied_seq, seq);
1842 tcp_rcv_space_adjust(sk);
1844 /* Clean up data we have read: This will do ACK frames. */
1845 tcp_recv_skb(sk, seq, &offset);
1846 tcp_cleanup_rbuf(sk, length);
1848 if (length == zc->length)
1849 zc->recv_skip_hint = 0;
1851 if (!zc->recv_skip_hint && sock_flag(sk, SOCK_DONE))
1854 zc->length = length;
1859 static void tcp_update_recv_tstamps(struct sk_buff *skb,
1860 struct scm_timestamping *tss)
1863 tss->ts[0] = ktime_to_timespec(skb->tstamp);
1865 tss->ts[0] = (struct timespec) {0};
1867 if (skb_hwtstamps(skb)->hwtstamp)
1868 tss->ts[2] = ktime_to_timespec(skb_hwtstamps(skb)->hwtstamp);
1870 tss->ts[2] = (struct timespec) {0};
1873 /* Similar to __sock_recv_timestamp, but does not require an skb */
1874 static void tcp_recv_timestamp(struct msghdr *msg, const struct sock *sk,
1875 struct scm_timestamping *tss)
1878 bool has_timestamping = false;
1880 if (tss->ts[0].tv_sec || tss->ts[0].tv_nsec) {
1881 if (sock_flag(sk, SOCK_RCVTSTAMP)) {
1882 if (sock_flag(sk, SOCK_RCVTSTAMPNS)) {
1883 put_cmsg(msg, SOL_SOCKET, SCM_TIMESTAMPNS,
1884 sizeof(tss->ts[0]), &tss->ts[0]);
1886 tv.tv_sec = tss->ts[0].tv_sec;
1887 tv.tv_usec = tss->ts[0].tv_nsec / 1000;
1889 put_cmsg(msg, SOL_SOCKET, SCM_TIMESTAMP,
1894 if (sk->sk_tsflags & SOF_TIMESTAMPING_SOFTWARE)
1895 has_timestamping = true;
1897 tss->ts[0] = (struct timespec) {0};
1900 if (tss->ts[2].tv_sec || tss->ts[2].tv_nsec) {
1901 if (sk->sk_tsflags & SOF_TIMESTAMPING_RAW_HARDWARE)
1902 has_timestamping = true;
1904 tss->ts[2] = (struct timespec) {0};
1907 if (has_timestamping) {
1908 tss->ts[1] = (struct timespec) {0};
1909 put_cmsg(msg, SOL_SOCKET, SCM_TIMESTAMPING,
1914 static int tcp_inq_hint(struct sock *sk)
1916 const struct tcp_sock *tp = tcp_sk(sk);
1917 u32 copied_seq = READ_ONCE(tp->copied_seq);
1918 u32 rcv_nxt = READ_ONCE(tp->rcv_nxt);
1921 inq = rcv_nxt - copied_seq;
1922 if (unlikely(inq < 0 || copied_seq != READ_ONCE(tp->copied_seq))) {
1924 inq = tp->rcv_nxt - tp->copied_seq;
1927 /* After receiving a FIN, tell the user-space to continue reading
1928 * by returning a non-zero inq.
1930 if (inq == 0 && sock_flag(sk, SOCK_DONE))
1936 * This routine copies from a sock struct into the user buffer.
1938 * Technical note: in 2.3 we work on _locked_ socket, so that
1939 * tricks with *seq access order and skb->users are not required.
1940 * Probably, code can be easily improved even more.
1943 int tcp_recvmsg(struct sock *sk, struct msghdr *msg, size_t len, int nonblock,
1944 int flags, int *addr_len)
1946 struct tcp_sock *tp = tcp_sk(sk);
1952 int target; /* Read at least this many bytes */
1954 struct sk_buff *skb, *last;
1956 struct scm_timestamping tss;
1957 bool has_tss = false;
1960 if (unlikely(flags & MSG_ERRQUEUE))
1961 return inet_recv_error(sk, msg, len, addr_len);
1963 if (sk_can_busy_loop(sk) && skb_queue_empty_lockless(&sk->sk_receive_queue) &&
1964 (sk->sk_state == TCP_ESTABLISHED))
1965 sk_busy_loop(sk, nonblock);
1970 if (sk->sk_state == TCP_LISTEN)
1973 has_cmsg = tp->recvmsg_inq;
1974 timeo = sock_rcvtimeo(sk, nonblock);
1976 /* Urgent data needs to be handled specially. */
1977 if (flags & MSG_OOB)
1980 if (unlikely(tp->repair)) {
1982 if (!(flags & MSG_PEEK))
1985 if (tp->repair_queue == TCP_SEND_QUEUE)
1989 if (tp->repair_queue == TCP_NO_QUEUE)
1992 /* 'common' recv queue MSG_PEEK-ing */
1995 seq = &tp->copied_seq;
1996 if (flags & MSG_PEEK) {
1997 peek_seq = tp->copied_seq;
2001 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
2006 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
2007 if (tp->urg_data && tp->urg_seq == *seq) {
2010 if (signal_pending(current)) {
2011 copied = timeo ? sock_intr_errno(timeo) : -EAGAIN;
2016 /* Next get a buffer. */
2018 last = skb_peek_tail(&sk->sk_receive_queue);
2019 skb_queue_walk(&sk->sk_receive_queue, skb) {
2021 /* Now that we have two receive queues this
2024 if (WARN(before(*seq, TCP_SKB_CB(skb)->seq),
2025 "TCP recvmsg seq # bug: copied %X, seq %X, rcvnxt %X, fl %X\n",
2026 *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt,
2030 offset = *seq - TCP_SKB_CB(skb)->seq;
2031 if (unlikely(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
2032 pr_err_once("%s: found a SYN, please report !\n", __func__);
2035 if (offset < skb->len)
2037 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
2039 WARN(!(flags & MSG_PEEK),
2040 "TCP recvmsg seq # bug 2: copied %X, seq %X, rcvnxt %X, fl %X\n",
2041 *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt, flags);
2044 /* Well, if we have backlog, try to process it now yet. */
2046 if (copied >= target && !READ_ONCE(sk->sk_backlog.tail))
2051 sk->sk_state == TCP_CLOSE ||
2052 (sk->sk_shutdown & RCV_SHUTDOWN) ||
2054 signal_pending(current))
2057 if (sock_flag(sk, SOCK_DONE))
2061 copied = sock_error(sk);
2065 if (sk->sk_shutdown & RCV_SHUTDOWN)
2068 if (sk->sk_state == TCP_CLOSE) {
2069 /* This occurs when user tries to read
2070 * from never connected socket.
2081 if (signal_pending(current)) {
2082 copied = sock_intr_errno(timeo);
2087 tcp_cleanup_rbuf(sk, copied);
2089 if (copied >= target) {
2090 /* Do not sleep, just process backlog. */
2094 sk_wait_data(sk, &timeo, last);
2097 if ((flags & MSG_PEEK) &&
2098 (peek_seq - copied - urg_hole != tp->copied_seq)) {
2099 net_dbg_ratelimited("TCP(%s:%d): Application bug, race in MSG_PEEK\n",
2101 task_pid_nr(current));
2102 peek_seq = tp->copied_seq;
2107 /* Ok so how much can we use? */
2108 used = skb->len - offset;
2112 /* Do we have urgent data here? */
2114 u32 urg_offset = tp->urg_seq - *seq;
2115 if (urg_offset < used) {
2117 if (!sock_flag(sk, SOCK_URGINLINE)) {
2118 WRITE_ONCE(*seq, *seq + 1);
2130 if (!(flags & MSG_TRUNC)) {
2131 err = skb_copy_datagram_msg(skb, offset, msg, used);
2133 /* Exception. Bailout! */
2140 WRITE_ONCE(*seq, *seq + used);
2144 tcp_rcv_space_adjust(sk);
2147 if (tp->urg_data && after(tp->copied_seq, tp->urg_seq)) {
2149 tcp_fast_path_check(sk);
2152 if (TCP_SKB_CB(skb)->has_rxtstamp) {
2153 tcp_update_recv_tstamps(skb, &tss);
2158 if (used + offset < skb->len)
2161 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
2163 if (!(flags & MSG_PEEK))
2164 sk_eat_skb(sk, skb);
2168 /* Process the FIN. */
2169 WRITE_ONCE(*seq, *seq + 1);
2170 if (!(flags & MSG_PEEK))
2171 sk_eat_skb(sk, skb);
2175 /* According to UNIX98, msg_name/msg_namelen are ignored
2176 * on connected socket. I was just happy when found this 8) --ANK
2179 /* Clean up data we have read: This will do ACK frames. */
2180 tcp_cleanup_rbuf(sk, copied);
2186 tcp_recv_timestamp(msg, sk, &tss);
2187 if (tp->recvmsg_inq) {
2188 inq = tcp_inq_hint(sk);
2189 put_cmsg(msg, SOL_TCP, TCP_CM_INQ, sizeof(inq), &inq);
2200 err = tcp_recv_urg(sk, msg, len, flags);
2204 err = tcp_peek_sndq(sk, msg, len);
2207 EXPORT_SYMBOL(tcp_recvmsg);
2209 void tcp_set_state(struct sock *sk, int state)
2211 int oldstate = sk->sk_state;
2213 /* We defined a new enum for TCP states that are exported in BPF
2214 * so as not force the internal TCP states to be frozen. The
2215 * following checks will detect if an internal state value ever
2216 * differs from the BPF value. If this ever happens, then we will
2217 * need to remap the internal value to the BPF value before calling
2218 * tcp_call_bpf_2arg.
2220 BUILD_BUG_ON((int)BPF_TCP_ESTABLISHED != (int)TCP_ESTABLISHED);
2221 BUILD_BUG_ON((int)BPF_TCP_SYN_SENT != (int)TCP_SYN_SENT);
2222 BUILD_BUG_ON((int)BPF_TCP_SYN_RECV != (int)TCP_SYN_RECV);
2223 BUILD_BUG_ON((int)BPF_TCP_FIN_WAIT1 != (int)TCP_FIN_WAIT1);
2224 BUILD_BUG_ON((int)BPF_TCP_FIN_WAIT2 != (int)TCP_FIN_WAIT2);
2225 BUILD_BUG_ON((int)BPF_TCP_TIME_WAIT != (int)TCP_TIME_WAIT);
2226 BUILD_BUG_ON((int)BPF_TCP_CLOSE != (int)TCP_CLOSE);
2227 BUILD_BUG_ON((int)BPF_TCP_CLOSE_WAIT != (int)TCP_CLOSE_WAIT);
2228 BUILD_BUG_ON((int)BPF_TCP_LAST_ACK != (int)TCP_LAST_ACK);
2229 BUILD_BUG_ON((int)BPF_TCP_LISTEN != (int)TCP_LISTEN);
2230 BUILD_BUG_ON((int)BPF_TCP_CLOSING != (int)TCP_CLOSING);
2231 BUILD_BUG_ON((int)BPF_TCP_NEW_SYN_RECV != (int)TCP_NEW_SYN_RECV);
2232 BUILD_BUG_ON((int)BPF_TCP_MAX_STATES != (int)TCP_MAX_STATES);
2234 if (BPF_SOCK_OPS_TEST_FLAG(tcp_sk(sk), BPF_SOCK_OPS_STATE_CB_FLAG))
2235 tcp_call_bpf_2arg(sk, BPF_SOCK_OPS_STATE_CB, oldstate, state);
2238 case TCP_ESTABLISHED:
2239 if (oldstate != TCP_ESTABLISHED)
2240 TCP_INC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
2244 if (oldstate == TCP_CLOSE_WAIT || oldstate == TCP_ESTABLISHED)
2245 TCP_INC_STATS(sock_net(sk), TCP_MIB_ESTABRESETS);
2247 sk->sk_prot->unhash(sk);
2248 if (inet_csk(sk)->icsk_bind_hash &&
2249 !(sk->sk_userlocks & SOCK_BINDPORT_LOCK))
2253 if (oldstate == TCP_ESTABLISHED)
2254 TCP_DEC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
2257 /* Change state AFTER socket is unhashed to avoid closed
2258 * socket sitting in hash tables.
2260 inet_sk_state_store(sk, state);
2263 SOCK_DEBUG(sk, "TCP sk=%p, State %s -> %s\n", sk, statename[oldstate], statename[state]);
2266 EXPORT_SYMBOL_GPL(tcp_set_state);
2269 * State processing on a close. This implements the state shift for
2270 * sending our FIN frame. Note that we only send a FIN for some
2271 * states. A shutdown() may have already sent the FIN, or we may be
2275 static const unsigned char new_state[16] = {
2276 /* current state: new state: action: */
2277 [0 /* (Invalid) */] = TCP_CLOSE,
2278 [TCP_ESTABLISHED] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2279 [TCP_SYN_SENT] = TCP_CLOSE,
2280 [TCP_SYN_RECV] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2281 [TCP_FIN_WAIT1] = TCP_FIN_WAIT1,
2282 [TCP_FIN_WAIT2] = TCP_FIN_WAIT2,
2283 [TCP_TIME_WAIT] = TCP_CLOSE,
2284 [TCP_CLOSE] = TCP_CLOSE,
2285 [TCP_CLOSE_WAIT] = TCP_LAST_ACK | TCP_ACTION_FIN,
2286 [TCP_LAST_ACK] = TCP_LAST_ACK,
2287 [TCP_LISTEN] = TCP_CLOSE,
2288 [TCP_CLOSING] = TCP_CLOSING,
2289 [TCP_NEW_SYN_RECV] = TCP_CLOSE, /* should not happen ! */
2292 static int tcp_close_state(struct sock *sk)
2294 int next = (int)new_state[sk->sk_state];
2295 int ns = next & TCP_STATE_MASK;
2297 tcp_set_state(sk, ns);
2299 return next & TCP_ACTION_FIN;
2303 * Shutdown the sending side of a connection. Much like close except
2304 * that we don't receive shut down or sock_set_flag(sk, SOCK_DEAD).
2307 void tcp_shutdown(struct sock *sk, int how)
2309 /* We need to grab some memory, and put together a FIN,
2310 * and then put it into the queue to be sent.
2311 * Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
2313 if (!(how & SEND_SHUTDOWN))
2316 /* If we've already sent a FIN, or it's a closed state, skip this. */
2317 if ((1 << sk->sk_state) &
2318 (TCPF_ESTABLISHED | TCPF_SYN_SENT |
2319 TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) {
2320 /* Clear out any half completed packets. FIN if needed. */
2321 if (tcp_close_state(sk))
2325 EXPORT_SYMBOL(tcp_shutdown);
2327 bool tcp_check_oom(struct sock *sk, int shift)
2329 bool too_many_orphans, out_of_socket_memory;
2331 too_many_orphans = tcp_too_many_orphans(sk, shift);
2332 out_of_socket_memory = tcp_out_of_memory(sk);
2334 if (too_many_orphans)
2335 net_info_ratelimited("too many orphaned sockets\n");
2336 if (out_of_socket_memory)
2337 net_info_ratelimited("out of memory -- consider tuning tcp_mem\n");
2338 return too_many_orphans || out_of_socket_memory;
2341 void tcp_close(struct sock *sk, long timeout)
2343 struct sk_buff *skb;
2344 int data_was_unread = 0;
2348 sk->sk_shutdown = SHUTDOWN_MASK;
2350 if (sk->sk_state == TCP_LISTEN) {
2351 tcp_set_state(sk, TCP_CLOSE);
2354 inet_csk_listen_stop(sk);
2356 goto adjudge_to_death;
2359 /* We need to flush the recv. buffs. We do this only on the
2360 * descriptor close, not protocol-sourced closes, because the
2361 * reader process may not have drained the data yet!
2363 while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) {
2364 u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq;
2366 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
2368 data_was_unread += len;
2374 /* If socket has been already reset (e.g. in tcp_reset()) - kill it. */
2375 if (sk->sk_state == TCP_CLOSE)
2376 goto adjudge_to_death;
2378 /* As outlined in RFC 2525, section 2.17, we send a RST here because
2379 * data was lost. To witness the awful effects of the old behavior of
2380 * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk
2381 * GET in an FTP client, suspend the process, wait for the client to
2382 * advertise a zero window, then kill -9 the FTP client, wheee...
2383 * Note: timeout is always zero in such a case.
2385 if (unlikely(tcp_sk(sk)->repair)) {
2386 sk->sk_prot->disconnect(sk, 0);
2387 } else if (data_was_unread) {
2388 /* Unread data was tossed, zap the connection. */
2389 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONCLOSE);
2390 tcp_set_state(sk, TCP_CLOSE);
2391 tcp_send_active_reset(sk, sk->sk_allocation);
2392 } else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) {
2393 /* Check zero linger _after_ checking for unread data. */
2394 sk->sk_prot->disconnect(sk, 0);
2395 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
2396 } else if (tcp_close_state(sk)) {
2397 /* We FIN if the application ate all the data before
2398 * zapping the connection.
2401 /* RED-PEN. Formally speaking, we have broken TCP state
2402 * machine. State transitions:
2404 * TCP_ESTABLISHED -> TCP_FIN_WAIT1
2405 * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible)
2406 * TCP_CLOSE_WAIT -> TCP_LAST_ACK
2408 * are legal only when FIN has been sent (i.e. in window),
2409 * rather than queued out of window. Purists blame.
2411 * F.e. "RFC state" is ESTABLISHED,
2412 * if Linux state is FIN-WAIT-1, but FIN is still not sent.
2414 * The visible declinations are that sometimes
2415 * we enter time-wait state, when it is not required really
2416 * (harmless), do not send active resets, when they are
2417 * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
2418 * they look as CLOSING or LAST_ACK for Linux)
2419 * Probably, I missed some more holelets.
2421 * XXX (TFO) - To start off we don't support SYN+ACK+FIN
2422 * in a single packet! (May consider it later but will
2423 * probably need API support or TCP_CORK SYN-ACK until
2424 * data is written and socket is closed.)
2429 sk_stream_wait_close(sk, timeout);
2432 state = sk->sk_state;
2438 /* remove backlog if any, without releasing ownership. */
2441 percpu_counter_inc(sk->sk_prot->orphan_count);
2443 /* Have we already been destroyed by a softirq or backlog? */
2444 if (state != TCP_CLOSE && sk->sk_state == TCP_CLOSE)
2447 /* This is a (useful) BSD violating of the RFC. There is a
2448 * problem with TCP as specified in that the other end could
2449 * keep a socket open forever with no application left this end.
2450 * We use a 1 minute timeout (about the same as BSD) then kill
2451 * our end. If they send after that then tough - BUT: long enough
2452 * that we won't make the old 4*rto = almost no time - whoops
2455 * Nope, it was not mistake. It is really desired behaviour
2456 * f.e. on http servers, when such sockets are useless, but
2457 * consume significant resources. Let's do it with special
2458 * linger2 option. --ANK
2461 if (sk->sk_state == TCP_FIN_WAIT2) {
2462 struct tcp_sock *tp = tcp_sk(sk);
2463 if (tp->linger2 < 0) {
2464 tcp_set_state(sk, TCP_CLOSE);
2465 tcp_send_active_reset(sk, GFP_ATOMIC);
2466 __NET_INC_STATS(sock_net(sk),
2467 LINUX_MIB_TCPABORTONLINGER);
2469 const int tmo = tcp_fin_time(sk);
2471 if (tmo > TCP_TIMEWAIT_LEN) {
2472 inet_csk_reset_keepalive_timer(sk,
2473 tmo - TCP_TIMEWAIT_LEN);
2475 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
2480 if (sk->sk_state != TCP_CLOSE) {
2482 if (tcp_check_oom(sk, 0)) {
2483 tcp_set_state(sk, TCP_CLOSE);
2484 tcp_send_active_reset(sk, GFP_ATOMIC);
2485 __NET_INC_STATS(sock_net(sk),
2486 LINUX_MIB_TCPABORTONMEMORY);
2487 } else if (!check_net(sock_net(sk))) {
2488 /* Not possible to send reset; just close */
2489 tcp_set_state(sk, TCP_CLOSE);
2493 if (sk->sk_state == TCP_CLOSE) {
2494 struct request_sock *req = tcp_sk(sk)->fastopen_rsk;
2495 /* We could get here with a non-NULL req if the socket is
2496 * aborted (e.g., closed with unread data) before 3WHS
2500 reqsk_fastopen_remove(sk, req, false);
2501 inet_csk_destroy_sock(sk);
2503 /* Otherwise, socket is reprieved until protocol close. */
2511 EXPORT_SYMBOL(tcp_close);
2513 /* These states need RST on ABORT according to RFC793 */
2515 static inline bool tcp_need_reset(int state)
2517 return (1 << state) &
2518 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 |
2519 TCPF_FIN_WAIT2 | TCPF_SYN_RECV);
2522 static void tcp_rtx_queue_purge(struct sock *sk)
2524 struct rb_node *p = rb_first(&sk->tcp_rtx_queue);
2526 tcp_sk(sk)->highest_sack = NULL;
2528 struct sk_buff *skb = rb_to_skb(p);
2531 /* Since we are deleting whole queue, no need to
2532 * list_del(&skb->tcp_tsorted_anchor)
2534 tcp_rtx_queue_unlink(skb, sk);
2535 sk_wmem_free_skb(sk, skb);
2539 void tcp_write_queue_purge(struct sock *sk)
2541 struct sk_buff *skb;
2543 tcp_chrono_stop(sk, TCP_CHRONO_BUSY);
2544 while ((skb = __skb_dequeue(&sk->sk_write_queue)) != NULL) {
2545 tcp_skb_tsorted_anchor_cleanup(skb);
2546 sk_wmem_free_skb(sk, skb);
2548 tcp_rtx_queue_purge(sk);
2549 INIT_LIST_HEAD(&tcp_sk(sk)->tsorted_sent_queue);
2551 tcp_clear_all_retrans_hints(tcp_sk(sk));
2552 tcp_sk(sk)->packets_out = 0;
2553 inet_csk(sk)->icsk_backoff = 0;
2556 int tcp_disconnect(struct sock *sk, int flags)
2558 struct inet_sock *inet = inet_sk(sk);
2559 struct inet_connection_sock *icsk = inet_csk(sk);
2560 struct tcp_sock *tp = tcp_sk(sk);
2561 int old_state = sk->sk_state;
2564 if (old_state != TCP_CLOSE)
2565 tcp_set_state(sk, TCP_CLOSE);
2567 /* ABORT function of RFC793 */
2568 if (old_state == TCP_LISTEN) {
2569 inet_csk_listen_stop(sk);
2570 } else if (unlikely(tp->repair)) {
2571 sk->sk_err = ECONNABORTED;
2572 } else if (tcp_need_reset(old_state) ||
2573 (tp->snd_nxt != tp->write_seq &&
2574 (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK))) {
2575 /* The last check adjusts for discrepancy of Linux wrt. RFC
2578 tcp_send_active_reset(sk, gfp_any());
2579 sk->sk_err = ECONNRESET;
2580 } else if (old_state == TCP_SYN_SENT)
2581 sk->sk_err = ECONNRESET;
2583 tcp_clear_xmit_timers(sk);
2584 __skb_queue_purge(&sk->sk_receive_queue);
2585 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt);
2587 tcp_write_queue_purge(sk);
2588 tcp_fastopen_active_disable_ofo_check(sk);
2589 skb_rbtree_purge(&tp->out_of_order_queue);
2591 inet->inet_dport = 0;
2593 if (!(sk->sk_userlocks & SOCK_BINDADDR_LOCK))
2594 inet_reset_saddr(sk);
2596 sk->sk_shutdown = 0;
2597 sock_reset_flag(sk, SOCK_DONE);
2599 tp->rcv_rtt_last_tsecr = 0;
2601 seq = tp->write_seq + tp->max_window + 2;
2604 WRITE_ONCE(tp->write_seq, seq);
2607 icsk->icsk_probes_out = 0;
2608 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
2609 tp->snd_cwnd_cnt = 0;
2610 tp->is_cwnd_limited = 0;
2611 tp->max_packets_out = 0;
2612 tp->window_clamp = 0;
2614 tp->delivered_ce = 0;
2615 if (icsk->icsk_ca_ops->release)
2616 icsk->icsk_ca_ops->release(sk);
2617 memset(icsk->icsk_ca_priv, 0, sizeof(icsk->icsk_ca_priv));
2618 tcp_set_ca_state(sk, TCP_CA_Open);
2619 tp->is_sack_reneg = 0;
2620 tcp_clear_retrans(tp);
2621 tp->total_retrans = 0;
2622 inet_csk_delack_init(sk);
2623 /* Initialize rcv_mss to TCP_MIN_MSS to avoid division by 0
2624 * issue in __tcp_select_window()
2626 icsk->icsk_ack.rcv_mss = TCP_MIN_MSS;
2627 memset(&tp->rx_opt, 0, sizeof(tp->rx_opt));
2629 dst_release(xchg((__force struct dst_entry **)&sk->sk_rx_dst, NULL));
2630 tcp_saved_syn_free(tp);
2631 tp->compressed_ack = 0;
2635 tp->bytes_acked = 0;
2636 tp->bytes_received = 0;
2637 tp->bytes_retrans = 0;
2638 tp->data_segs_in = 0;
2639 tp->data_segs_out = 0;
2643 /* Clean up fastopen related fields */
2644 tcp_free_fastopen_req(tp);
2645 inet->defer_connect = 0;
2647 WARN_ON(inet->inet_num && !icsk->icsk_bind_hash);
2649 if (sk->sk_frag.page) {
2650 put_page(sk->sk_frag.page);
2651 sk->sk_frag.page = NULL;
2652 sk->sk_frag.offset = 0;
2655 sk->sk_error_report(sk);
2658 EXPORT_SYMBOL(tcp_disconnect);
2660 static inline bool tcp_can_repair_sock(const struct sock *sk)
2662 return ns_capable(sock_net(sk)->user_ns, CAP_NET_ADMIN) &&
2663 (sk->sk_state != TCP_LISTEN);
2666 static int tcp_repair_set_window(struct tcp_sock *tp, char __user *optbuf, int len)
2668 struct tcp_repair_window opt;
2673 if (len != sizeof(opt))
2676 if (copy_from_user(&opt, optbuf, sizeof(opt)))
2679 if (opt.max_window < opt.snd_wnd)
2682 if (after(opt.snd_wl1, tp->rcv_nxt + opt.rcv_wnd))
2685 if (after(opt.rcv_wup, tp->rcv_nxt))
2688 tp->snd_wl1 = opt.snd_wl1;
2689 tp->snd_wnd = opt.snd_wnd;
2690 tp->max_window = opt.max_window;
2692 tp->rcv_wnd = opt.rcv_wnd;
2693 tp->rcv_wup = opt.rcv_wup;
2698 static int tcp_repair_options_est(struct sock *sk,
2699 struct tcp_repair_opt __user *optbuf, unsigned int len)
2701 struct tcp_sock *tp = tcp_sk(sk);
2702 struct tcp_repair_opt opt;
2704 while (len >= sizeof(opt)) {
2705 if (copy_from_user(&opt, optbuf, sizeof(opt)))
2711 switch (opt.opt_code) {
2713 tp->rx_opt.mss_clamp = opt.opt_val;
2718 u16 snd_wscale = opt.opt_val & 0xFFFF;
2719 u16 rcv_wscale = opt.opt_val >> 16;
2721 if (snd_wscale > TCP_MAX_WSCALE || rcv_wscale > TCP_MAX_WSCALE)
2724 tp->rx_opt.snd_wscale = snd_wscale;
2725 tp->rx_opt.rcv_wscale = rcv_wscale;
2726 tp->rx_opt.wscale_ok = 1;
2729 case TCPOPT_SACK_PERM:
2730 if (opt.opt_val != 0)
2733 tp->rx_opt.sack_ok |= TCP_SACK_SEEN;
2735 case TCPOPT_TIMESTAMP:
2736 if (opt.opt_val != 0)
2739 tp->rx_opt.tstamp_ok = 1;
2748 * Socket option code for TCP.
2750 static int do_tcp_setsockopt(struct sock *sk, int level,
2751 int optname, char __user *optval, unsigned int optlen)
2753 struct tcp_sock *tp = tcp_sk(sk);
2754 struct inet_connection_sock *icsk = inet_csk(sk);
2755 struct net *net = sock_net(sk);
2759 /* These are data/string values, all the others are ints */
2761 case TCP_CONGESTION: {
2762 char name[TCP_CA_NAME_MAX];
2767 val = strncpy_from_user(name, optval,
2768 min_t(long, TCP_CA_NAME_MAX-1, optlen));
2774 err = tcp_set_congestion_control(sk, name, true, true,
2775 ns_capable(sock_net(sk)->user_ns,
2781 char name[TCP_ULP_NAME_MAX];
2786 val = strncpy_from_user(name, optval,
2787 min_t(long, TCP_ULP_NAME_MAX - 1,
2794 err = tcp_set_ulp(sk, name);
2798 case TCP_FASTOPEN_KEY: {
2799 __u8 key[TCP_FASTOPEN_KEY_LENGTH];
2801 if (optlen != sizeof(key))
2804 if (copy_from_user(key, optval, optlen))
2807 return tcp_fastopen_reset_cipher(net, sk, key, sizeof(key));
2814 if (optlen < sizeof(int))
2817 if (get_user(val, (int __user *)optval))
2824 /* Values greater than interface MTU won't take effect. However
2825 * at the point when this call is done we typically don't yet
2826 * know which interface is going to be used
2828 if (val && (val < TCP_MIN_MSS || val > MAX_TCP_WINDOW)) {
2832 tp->rx_opt.user_mss = val;
2837 /* TCP_NODELAY is weaker than TCP_CORK, so that
2838 * this option on corked socket is remembered, but
2839 * it is not activated until cork is cleared.
2841 * However, when TCP_NODELAY is set we make
2842 * an explicit push, which overrides even TCP_CORK
2843 * for currently queued segments.
2845 tp->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH;
2846 tcp_push_pending_frames(sk);
2848 tp->nonagle &= ~TCP_NAGLE_OFF;
2852 case TCP_THIN_LINEAR_TIMEOUTS:
2853 if (val < 0 || val > 1)
2859 case TCP_THIN_DUPACK:
2860 if (val < 0 || val > 1)
2865 if (!tcp_can_repair_sock(sk))
2867 else if (val == TCP_REPAIR_ON) {
2869 sk->sk_reuse = SK_FORCE_REUSE;
2870 tp->repair_queue = TCP_NO_QUEUE;
2871 } else if (val == TCP_REPAIR_OFF) {
2873 sk->sk_reuse = SK_NO_REUSE;
2874 tcp_send_window_probe(sk);
2875 } else if (val == TCP_REPAIR_OFF_NO_WP) {
2877 sk->sk_reuse = SK_NO_REUSE;
2883 case TCP_REPAIR_QUEUE:
2886 else if ((unsigned int)val < TCP_QUEUES_NR)
2887 tp->repair_queue = val;
2893 if (sk->sk_state != TCP_CLOSE) {
2895 } else if (tp->repair_queue == TCP_SEND_QUEUE) {
2896 if (!tcp_rtx_queue_empty(sk))
2899 WRITE_ONCE(tp->write_seq, val);
2900 } else if (tp->repair_queue == TCP_RECV_QUEUE) {
2901 if (tp->rcv_nxt != tp->copied_seq) {
2904 WRITE_ONCE(tp->rcv_nxt, val);
2905 WRITE_ONCE(tp->copied_seq, val);
2912 case TCP_REPAIR_OPTIONS:
2915 else if (sk->sk_state == TCP_ESTABLISHED)
2916 err = tcp_repair_options_est(sk,
2917 (struct tcp_repair_opt __user *)optval,
2924 /* When set indicates to always queue non-full frames.
2925 * Later the user clears this option and we transmit
2926 * any pending partial frames in the queue. This is
2927 * meant to be used alongside sendfile() to get properly
2928 * filled frames when the user (for example) must write
2929 * out headers with a write() call first and then use
2930 * sendfile to send out the data parts.
2932 * TCP_CORK can be set together with TCP_NODELAY and it is
2933 * stronger than TCP_NODELAY.
2936 tp->nonagle |= TCP_NAGLE_CORK;
2938 tp->nonagle &= ~TCP_NAGLE_CORK;
2939 if (tp->nonagle&TCP_NAGLE_OFF)
2940 tp->nonagle |= TCP_NAGLE_PUSH;
2941 tcp_push_pending_frames(sk);
2946 if (val < 1 || val > MAX_TCP_KEEPIDLE)
2949 tp->keepalive_time = val * HZ;
2950 if (sock_flag(sk, SOCK_KEEPOPEN) &&
2951 !((1 << sk->sk_state) &
2952 (TCPF_CLOSE | TCPF_LISTEN))) {
2953 u32 elapsed = keepalive_time_elapsed(tp);
2954 if (tp->keepalive_time > elapsed)
2955 elapsed = tp->keepalive_time - elapsed;
2958 inet_csk_reset_keepalive_timer(sk, elapsed);
2963 if (val < 1 || val > MAX_TCP_KEEPINTVL)
2966 tp->keepalive_intvl = val * HZ;
2969 if (val < 1 || val > MAX_TCP_KEEPCNT)
2972 tp->keepalive_probes = val;
2975 if (val < 1 || val > MAX_TCP_SYNCNT)
2978 icsk->icsk_syn_retries = val;
2982 if (val < 0 || val > 1)
2991 else if (val > net->ipv4.sysctl_tcp_fin_timeout / HZ)
2994 tp->linger2 = val * HZ;
2997 case TCP_DEFER_ACCEPT:
2998 /* Translate value in seconds to number of retransmits */
2999 icsk->icsk_accept_queue.rskq_defer_accept =
3000 secs_to_retrans(val, TCP_TIMEOUT_INIT / HZ,
3004 case TCP_WINDOW_CLAMP:
3006 if (sk->sk_state != TCP_CLOSE) {
3010 tp->window_clamp = 0;
3012 tp->window_clamp = val < SOCK_MIN_RCVBUF / 2 ?
3013 SOCK_MIN_RCVBUF / 2 : val;
3018 icsk->icsk_ack.pingpong = 1;
3020 icsk->icsk_ack.pingpong = 0;
3021 if ((1 << sk->sk_state) &
3022 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) &&
3023 inet_csk_ack_scheduled(sk)) {
3024 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
3025 tcp_cleanup_rbuf(sk, 1);
3027 icsk->icsk_ack.pingpong = 1;
3032 #ifdef CONFIG_TCP_MD5SIG
3034 case TCP_MD5SIG_EXT:
3035 err = tp->af_specific->md5_parse(sk, optname, optval, optlen);
3038 case TCP_USER_TIMEOUT:
3039 /* Cap the max time in ms TCP will retry or probe the window
3040 * before giving up and aborting (ETIMEDOUT) a connection.
3045 icsk->icsk_user_timeout = val;
3049 if (val >= 0 && ((1 << sk->sk_state) & (TCPF_CLOSE |
3051 tcp_fastopen_init_key_once(net);
3053 fastopen_queue_tune(sk, val);
3058 case TCP_FASTOPEN_CONNECT:
3059 if (val > 1 || val < 0) {
3061 } else if (READ_ONCE(net->ipv4.sysctl_tcp_fastopen) &
3062 TFO_CLIENT_ENABLE) {
3063 if (sk->sk_state == TCP_CLOSE)
3064 tp->fastopen_connect = val;
3071 case TCP_FASTOPEN_NO_COOKIE:
3072 if (val > 1 || val < 0)
3074 else if (!((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
3077 tp->fastopen_no_cookie = val;
3083 tp->tsoffset = val - tcp_time_stamp_raw();
3085 case TCP_REPAIR_WINDOW:
3086 err = tcp_repair_set_window(tp, optval, optlen);
3088 case TCP_NOTSENT_LOWAT:
3089 tp->notsent_lowat = val;
3090 sk->sk_write_space(sk);
3093 if (val > 1 || val < 0)
3096 tp->recvmsg_inq = val;
3107 int tcp_setsockopt(struct sock *sk, int level, int optname, char __user *optval,
3108 unsigned int optlen)
3110 const struct inet_connection_sock *icsk = inet_csk(sk);
3112 if (level != SOL_TCP)
3113 return icsk->icsk_af_ops->setsockopt(sk, level, optname,
3115 return do_tcp_setsockopt(sk, level, optname, optval, optlen);
3117 EXPORT_SYMBOL(tcp_setsockopt);
3119 #ifdef CONFIG_COMPAT
3120 int compat_tcp_setsockopt(struct sock *sk, int level, int optname,
3121 char __user *optval, unsigned int optlen)
3123 if (level != SOL_TCP)
3124 return inet_csk_compat_setsockopt(sk, level, optname,
3126 return do_tcp_setsockopt(sk, level, optname, optval, optlen);
3128 EXPORT_SYMBOL(compat_tcp_setsockopt);
3131 static void tcp_get_info_chrono_stats(const struct tcp_sock *tp,
3132 struct tcp_info *info)
3134 u64 stats[__TCP_CHRONO_MAX], total = 0;
3137 for (i = TCP_CHRONO_BUSY; i < __TCP_CHRONO_MAX; ++i) {
3138 stats[i] = tp->chrono_stat[i - 1];
3139 if (i == tp->chrono_type)
3140 stats[i] += tcp_jiffies32 - tp->chrono_start;
3141 stats[i] *= USEC_PER_SEC / HZ;
3145 info->tcpi_busy_time = total;
3146 info->tcpi_rwnd_limited = stats[TCP_CHRONO_RWND_LIMITED];
3147 info->tcpi_sndbuf_limited = stats[TCP_CHRONO_SNDBUF_LIMITED];
3150 /* Return information about state of tcp endpoint in API format. */
3151 void tcp_get_info(struct sock *sk, struct tcp_info *info)
3153 const struct tcp_sock *tp = tcp_sk(sk); /* iff sk_type == SOCK_STREAM */
3154 const struct inet_connection_sock *icsk = inet_csk(sk);
3160 memset(info, 0, sizeof(*info));
3161 if (sk->sk_type != SOCK_STREAM)
3164 info->tcpi_state = inet_sk_state_load(sk);
3166 /* Report meaningful fields for all TCP states, including listeners */
3167 rate = READ_ONCE(sk->sk_pacing_rate);
3168 rate64 = rate != ~0U ? rate : ~0ULL;
3169 info->tcpi_pacing_rate = rate64;
3171 rate = READ_ONCE(sk->sk_max_pacing_rate);
3172 rate64 = rate != ~0U ? rate : ~0ULL;
3173 info->tcpi_max_pacing_rate = rate64;
3175 info->tcpi_reordering = tp->reordering;
3176 info->tcpi_snd_cwnd = tp->snd_cwnd;
3178 if (info->tcpi_state == TCP_LISTEN) {
3179 /* listeners aliased fields :
3180 * tcpi_unacked -> Number of children ready for accept()
3181 * tcpi_sacked -> max backlog
3183 info->tcpi_unacked = sk->sk_ack_backlog;
3184 info->tcpi_sacked = sk->sk_max_ack_backlog;
3188 slow = lock_sock_fast(sk);
3190 info->tcpi_ca_state = icsk->icsk_ca_state;
3191 info->tcpi_retransmits = icsk->icsk_retransmits;
3192 info->tcpi_probes = icsk->icsk_probes_out;
3193 info->tcpi_backoff = icsk->icsk_backoff;
3195 if (tp->rx_opt.tstamp_ok)
3196 info->tcpi_options |= TCPI_OPT_TIMESTAMPS;
3197 if (tcp_is_sack(tp))
3198 info->tcpi_options |= TCPI_OPT_SACK;
3199 if (tp->rx_opt.wscale_ok) {
3200 info->tcpi_options |= TCPI_OPT_WSCALE;
3201 info->tcpi_snd_wscale = tp->rx_opt.snd_wscale;
3202 info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale;
3205 if (tp->ecn_flags & TCP_ECN_OK)
3206 info->tcpi_options |= TCPI_OPT_ECN;
3207 if (tp->ecn_flags & TCP_ECN_SEEN)
3208 info->tcpi_options |= TCPI_OPT_ECN_SEEN;
3209 if (tp->syn_data_acked)
3210 info->tcpi_options |= TCPI_OPT_SYN_DATA;
3212 info->tcpi_rto = jiffies_to_usecs(icsk->icsk_rto);
3213 info->tcpi_ato = jiffies_to_usecs(icsk->icsk_ack.ato);
3214 info->tcpi_snd_mss = tp->mss_cache;
3215 info->tcpi_rcv_mss = icsk->icsk_ack.rcv_mss;
3217 info->tcpi_unacked = tp->packets_out;
3218 info->tcpi_sacked = tp->sacked_out;
3220 info->tcpi_lost = tp->lost_out;
3221 info->tcpi_retrans = tp->retrans_out;
3223 now = tcp_jiffies32;
3224 info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime);
3225 info->tcpi_last_data_recv = jiffies_to_msecs(now - icsk->icsk_ack.lrcvtime);
3226 info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp);
3228 info->tcpi_pmtu = icsk->icsk_pmtu_cookie;
3229 info->tcpi_rcv_ssthresh = tp->rcv_ssthresh;
3230 info->tcpi_rtt = tp->srtt_us >> 3;
3231 info->tcpi_rttvar = tp->mdev_us >> 2;
3232 info->tcpi_snd_ssthresh = tp->snd_ssthresh;
3233 info->tcpi_advmss = tp->advmss;
3235 info->tcpi_rcv_rtt = tp->rcv_rtt_est.rtt_us >> 3;
3236 info->tcpi_rcv_space = tp->rcvq_space.space;
3238 info->tcpi_total_retrans = tp->total_retrans;
3240 info->tcpi_bytes_acked = tp->bytes_acked;
3241 info->tcpi_bytes_received = tp->bytes_received;
3242 info->tcpi_notsent_bytes = max_t(int, 0, tp->write_seq - tp->snd_nxt);
3243 tcp_get_info_chrono_stats(tp, info);
3245 info->tcpi_segs_out = tp->segs_out;
3246 info->tcpi_segs_in = tp->segs_in;
3248 info->tcpi_min_rtt = tcp_min_rtt(tp);
3249 info->tcpi_data_segs_in = tp->data_segs_in;
3250 info->tcpi_data_segs_out = tp->data_segs_out;
3252 info->tcpi_delivery_rate_app_limited = tp->rate_app_limited ? 1 : 0;
3253 rate64 = tcp_compute_delivery_rate(tp);
3255 info->tcpi_delivery_rate = rate64;
3256 info->tcpi_delivered = tp->delivered;
3257 info->tcpi_delivered_ce = tp->delivered_ce;
3258 info->tcpi_bytes_sent = tp->bytes_sent;
3259 info->tcpi_bytes_retrans = tp->bytes_retrans;
3260 info->tcpi_dsack_dups = tp->dsack_dups;
3261 info->tcpi_reord_seen = tp->reord_seen;
3262 unlock_sock_fast(sk, slow);
3264 EXPORT_SYMBOL_GPL(tcp_get_info);
3266 static size_t tcp_opt_stats_get_size(void)
3269 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_BUSY */
3270 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_RWND_LIMITED */
3271 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_SNDBUF_LIMITED */
3272 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_DATA_SEGS_OUT */
3273 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_TOTAL_RETRANS */
3274 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_PACING_RATE */
3275 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_DELIVERY_RATE */
3276 nla_total_size(sizeof(u32)) + /* TCP_NLA_SND_CWND */
3277 nla_total_size(sizeof(u32)) + /* TCP_NLA_REORDERING */
3278 nla_total_size(sizeof(u32)) + /* TCP_NLA_MIN_RTT */
3279 nla_total_size(sizeof(u8)) + /* TCP_NLA_RECUR_RETRANS */
3280 nla_total_size(sizeof(u8)) + /* TCP_NLA_DELIVERY_RATE_APP_LMT */
3281 nla_total_size(sizeof(u32)) + /* TCP_NLA_SNDQ_SIZE */
3282 nla_total_size(sizeof(u8)) + /* TCP_NLA_CA_STATE */
3283 nla_total_size(sizeof(u32)) + /* TCP_NLA_SND_SSTHRESH */
3284 nla_total_size(sizeof(u32)) + /* TCP_NLA_DELIVERED */
3285 nla_total_size(sizeof(u32)) + /* TCP_NLA_DELIVERED_CE */
3286 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_BYTES_SENT */
3287 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_BYTES_RETRANS */
3288 nla_total_size(sizeof(u32)) + /* TCP_NLA_DSACK_DUPS */
3289 nla_total_size(sizeof(u32)) + /* TCP_NLA_REORD_SEEN */
3293 struct sk_buff *tcp_get_timestamping_opt_stats(const struct sock *sk)
3295 const struct tcp_sock *tp = tcp_sk(sk);
3296 struct sk_buff *stats;
3297 struct tcp_info info;
3301 stats = alloc_skb(tcp_opt_stats_get_size(), GFP_ATOMIC);
3305 tcp_get_info_chrono_stats(tp, &info);
3306 nla_put_u64_64bit(stats, TCP_NLA_BUSY,
3307 info.tcpi_busy_time, TCP_NLA_PAD);
3308 nla_put_u64_64bit(stats, TCP_NLA_RWND_LIMITED,
3309 info.tcpi_rwnd_limited, TCP_NLA_PAD);
3310 nla_put_u64_64bit(stats, TCP_NLA_SNDBUF_LIMITED,
3311 info.tcpi_sndbuf_limited, TCP_NLA_PAD);
3312 nla_put_u64_64bit(stats, TCP_NLA_DATA_SEGS_OUT,
3313 tp->data_segs_out, TCP_NLA_PAD);
3314 nla_put_u64_64bit(stats, TCP_NLA_TOTAL_RETRANS,
3315 tp->total_retrans, TCP_NLA_PAD);
3317 rate = READ_ONCE(sk->sk_pacing_rate);
3318 rate64 = rate != ~0U ? rate : ~0ULL;
3319 nla_put_u64_64bit(stats, TCP_NLA_PACING_RATE, rate64, TCP_NLA_PAD);
3321 rate64 = tcp_compute_delivery_rate(tp);
3322 nla_put_u64_64bit(stats, TCP_NLA_DELIVERY_RATE, rate64, TCP_NLA_PAD);
3324 nla_put_u32(stats, TCP_NLA_SND_CWND, tp->snd_cwnd);
3325 nla_put_u32(stats, TCP_NLA_REORDERING, tp->reordering);
3326 nla_put_u32(stats, TCP_NLA_MIN_RTT, tcp_min_rtt(tp));
3328 nla_put_u8(stats, TCP_NLA_RECUR_RETRANS, inet_csk(sk)->icsk_retransmits);
3329 nla_put_u8(stats, TCP_NLA_DELIVERY_RATE_APP_LMT, !!tp->rate_app_limited);
3330 nla_put_u32(stats, TCP_NLA_SND_SSTHRESH, tp->snd_ssthresh);
3331 nla_put_u32(stats, TCP_NLA_DELIVERED, tp->delivered);
3332 nla_put_u32(stats, TCP_NLA_DELIVERED_CE, tp->delivered_ce);
3334 nla_put_u32(stats, TCP_NLA_SNDQ_SIZE, tp->write_seq - tp->snd_una);
3335 nla_put_u8(stats, TCP_NLA_CA_STATE, inet_csk(sk)->icsk_ca_state);
3337 nla_put_u64_64bit(stats, TCP_NLA_BYTES_SENT, tp->bytes_sent,
3339 nla_put_u64_64bit(stats, TCP_NLA_BYTES_RETRANS, tp->bytes_retrans,
3341 nla_put_u32(stats, TCP_NLA_DSACK_DUPS, tp->dsack_dups);
3342 nla_put_u32(stats, TCP_NLA_REORD_SEEN, tp->reord_seen);
3347 static int do_tcp_getsockopt(struct sock *sk, int level,
3348 int optname, char __user *optval, int __user *optlen)
3350 struct inet_connection_sock *icsk = inet_csk(sk);
3351 struct tcp_sock *tp = tcp_sk(sk);
3352 struct net *net = sock_net(sk);
3355 if (get_user(len, optlen))
3358 len = min_t(unsigned int, len, sizeof(int));
3365 val = tp->mss_cache;
3366 if (!val && ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
3367 val = tp->rx_opt.user_mss;
3369 val = tp->rx_opt.mss_clamp;
3372 val = !!(tp->nonagle&TCP_NAGLE_OFF);
3375 val = !!(tp->nonagle&TCP_NAGLE_CORK);
3378 val = keepalive_time_when(tp) / HZ;
3381 val = keepalive_intvl_when(tp) / HZ;
3384 val = keepalive_probes(tp);
3387 val = icsk->icsk_syn_retries ? : net->ipv4.sysctl_tcp_syn_retries;
3392 val = (val ? : READ_ONCE(net->ipv4.sysctl_tcp_fin_timeout)) / HZ;
3394 case TCP_DEFER_ACCEPT:
3395 val = retrans_to_secs(icsk->icsk_accept_queue.rskq_defer_accept,
3396 TCP_TIMEOUT_INIT / HZ, TCP_RTO_MAX / HZ);
3398 case TCP_WINDOW_CLAMP:
3399 val = tp->window_clamp;
3402 struct tcp_info info;
3404 if (get_user(len, optlen))
3407 tcp_get_info(sk, &info);
3409 len = min_t(unsigned int, len, sizeof(info));
3410 if (put_user(len, optlen))
3412 if (copy_to_user(optval, &info, len))
3417 const struct tcp_congestion_ops *ca_ops;
3418 union tcp_cc_info info;
3422 if (get_user(len, optlen))
3425 ca_ops = icsk->icsk_ca_ops;
3426 if (ca_ops && ca_ops->get_info)
3427 sz = ca_ops->get_info(sk, ~0U, &attr, &info);
3429 len = min_t(unsigned int, len, sz);
3430 if (put_user(len, optlen))
3432 if (copy_to_user(optval, &info, len))
3437 val = !icsk->icsk_ack.pingpong;
3440 case TCP_CONGESTION:
3441 if (get_user(len, optlen))
3443 len = min_t(unsigned int, len, TCP_CA_NAME_MAX);
3444 if (put_user(len, optlen))
3446 if (copy_to_user(optval, icsk->icsk_ca_ops->name, len))
3451 if (get_user(len, optlen))
3453 len = min_t(unsigned int, len, TCP_ULP_NAME_MAX);
3454 if (!icsk->icsk_ulp_ops) {
3455 if (put_user(0, optlen))
3459 if (put_user(len, optlen))
3461 if (copy_to_user(optval, icsk->icsk_ulp_ops->name, len))
3465 case TCP_FASTOPEN_KEY: {
3466 __u8 key[TCP_FASTOPEN_KEY_LENGTH];
3467 struct tcp_fastopen_context *ctx;
3469 if (get_user(len, optlen))
3473 ctx = rcu_dereference(icsk->icsk_accept_queue.fastopenq.ctx);
3475 memcpy(key, ctx->key, sizeof(key));
3480 len = min_t(unsigned int, len, sizeof(key));
3481 if (put_user(len, optlen))
3483 if (copy_to_user(optval, key, len))
3487 case TCP_THIN_LINEAR_TIMEOUTS:
3491 case TCP_THIN_DUPACK:
3499 case TCP_REPAIR_QUEUE:
3501 val = tp->repair_queue;
3506 case TCP_REPAIR_WINDOW: {
3507 struct tcp_repair_window opt;
3509 if (get_user(len, optlen))
3512 if (len != sizeof(opt))
3518 opt.snd_wl1 = tp->snd_wl1;
3519 opt.snd_wnd = tp->snd_wnd;
3520 opt.max_window = tp->max_window;
3521 opt.rcv_wnd = tp->rcv_wnd;
3522 opt.rcv_wup = tp->rcv_wup;
3524 if (copy_to_user(optval, &opt, len))
3529 if (tp->repair_queue == TCP_SEND_QUEUE)
3530 val = tp->write_seq;
3531 else if (tp->repair_queue == TCP_RECV_QUEUE)
3537 case TCP_USER_TIMEOUT:
3538 val = icsk->icsk_user_timeout;
3542 val = icsk->icsk_accept_queue.fastopenq.max_qlen;
3545 case TCP_FASTOPEN_CONNECT:
3546 val = tp->fastopen_connect;
3549 case TCP_FASTOPEN_NO_COOKIE:
3550 val = tp->fastopen_no_cookie;
3554 val = tcp_time_stamp_raw() + tp->tsoffset;
3556 case TCP_NOTSENT_LOWAT:
3557 val = tp->notsent_lowat;
3560 val = tp->recvmsg_inq;
3565 case TCP_SAVED_SYN: {
3566 if (get_user(len, optlen))
3570 if (tp->saved_syn) {
3571 if (len < tp->saved_syn[0]) {
3572 if (put_user(tp->saved_syn[0], optlen)) {
3579 len = tp->saved_syn[0];
3580 if (put_user(len, optlen)) {
3584 if (copy_to_user(optval, tp->saved_syn + 1, len)) {
3588 tcp_saved_syn_free(tp);
3593 if (put_user(len, optlen))
3599 case TCP_ZEROCOPY_RECEIVE: {
3600 struct tcp_zerocopy_receive zc;
3603 if (get_user(len, optlen))
3605 if (len != sizeof(zc))
3607 if (copy_from_user(&zc, optval, len))
3610 err = tcp_zerocopy_receive(sk, &zc);
3612 if (!err && copy_to_user(optval, &zc, len))
3618 return -ENOPROTOOPT;
3621 if (put_user(len, optlen))
3623 if (copy_to_user(optval, &val, len))
3628 int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval,
3631 struct inet_connection_sock *icsk = inet_csk(sk);
3633 if (level != SOL_TCP)
3634 return icsk->icsk_af_ops->getsockopt(sk, level, optname,
3636 return do_tcp_getsockopt(sk, level, optname, optval, optlen);
3638 EXPORT_SYMBOL(tcp_getsockopt);
3640 #ifdef CONFIG_COMPAT
3641 int compat_tcp_getsockopt(struct sock *sk, int level, int optname,
3642 char __user *optval, int __user *optlen)
3644 if (level != SOL_TCP)
3645 return inet_csk_compat_getsockopt(sk, level, optname,
3647 return do_tcp_getsockopt(sk, level, optname, optval, optlen);
3649 EXPORT_SYMBOL(compat_tcp_getsockopt);
3652 #ifdef CONFIG_TCP_MD5SIG
3653 static DEFINE_PER_CPU(struct tcp_md5sig_pool, tcp_md5sig_pool);
3654 static DEFINE_MUTEX(tcp_md5sig_mutex);
3655 static bool tcp_md5sig_pool_populated = false;
3657 static void __tcp_alloc_md5sig_pool(void)
3659 struct crypto_ahash *hash;
3662 hash = crypto_alloc_ahash("md5", 0, CRYPTO_ALG_ASYNC);
3666 for_each_possible_cpu(cpu) {
3667 void *scratch = per_cpu(tcp_md5sig_pool, cpu).scratch;
3668 struct ahash_request *req;
3671 scratch = kmalloc_node(sizeof(union tcp_md5sum_block) +
3672 sizeof(struct tcphdr),
3677 per_cpu(tcp_md5sig_pool, cpu).scratch = scratch;
3679 if (per_cpu(tcp_md5sig_pool, cpu).md5_req)
3682 req = ahash_request_alloc(hash, GFP_KERNEL);
3686 ahash_request_set_callback(req, 0, NULL, NULL);
3688 per_cpu(tcp_md5sig_pool, cpu).md5_req = req;
3690 /* before setting tcp_md5sig_pool_populated, we must commit all writes
3691 * to memory. See smp_rmb() in tcp_get_md5sig_pool()
3694 /* Paired with READ_ONCE() from tcp_alloc_md5sig_pool()
3695 * and tcp_get_md5sig_pool().
3697 WRITE_ONCE(tcp_md5sig_pool_populated, true);
3700 bool tcp_alloc_md5sig_pool(void)
3702 /* Paired with WRITE_ONCE() from __tcp_alloc_md5sig_pool() */
3703 if (unlikely(!READ_ONCE(tcp_md5sig_pool_populated))) {
3704 mutex_lock(&tcp_md5sig_mutex);
3706 if (!tcp_md5sig_pool_populated)
3707 __tcp_alloc_md5sig_pool();
3709 mutex_unlock(&tcp_md5sig_mutex);
3711 /* Paired with WRITE_ONCE() from __tcp_alloc_md5sig_pool() */
3712 return READ_ONCE(tcp_md5sig_pool_populated);
3714 EXPORT_SYMBOL(tcp_alloc_md5sig_pool);
3718 * tcp_get_md5sig_pool - get md5sig_pool for this user
3720 * We use percpu structure, so if we succeed, we exit with preemption
3721 * and BH disabled, to make sure another thread or softirq handling
3722 * wont try to get same context.
3724 struct tcp_md5sig_pool *tcp_get_md5sig_pool(void)
3728 /* Paired with WRITE_ONCE() from __tcp_alloc_md5sig_pool() */
3729 if (READ_ONCE(tcp_md5sig_pool_populated)) {
3730 /* coupled with smp_wmb() in __tcp_alloc_md5sig_pool() */
3732 return this_cpu_ptr(&tcp_md5sig_pool);
3737 EXPORT_SYMBOL(tcp_get_md5sig_pool);
3739 int tcp_md5_hash_skb_data(struct tcp_md5sig_pool *hp,
3740 const struct sk_buff *skb, unsigned int header_len)
3742 struct scatterlist sg;
3743 const struct tcphdr *tp = tcp_hdr(skb);
3744 struct ahash_request *req = hp->md5_req;
3746 const unsigned int head_data_len = skb_headlen(skb) > header_len ?
3747 skb_headlen(skb) - header_len : 0;
3748 const struct skb_shared_info *shi = skb_shinfo(skb);
3749 struct sk_buff *frag_iter;
3751 sg_init_table(&sg, 1);
3753 sg_set_buf(&sg, ((u8 *) tp) + header_len, head_data_len);
3754 ahash_request_set_crypt(req, &sg, NULL, head_data_len);
3755 if (crypto_ahash_update(req))
3758 for (i = 0; i < shi->nr_frags; ++i) {
3759 const struct skb_frag_struct *f = &shi->frags[i];
3760 unsigned int offset = f->page_offset;
3761 struct page *page = skb_frag_page(f) + (offset >> PAGE_SHIFT);
3763 sg_set_page(&sg, page, skb_frag_size(f),
3764 offset_in_page(offset));
3765 ahash_request_set_crypt(req, &sg, NULL, skb_frag_size(f));
3766 if (crypto_ahash_update(req))
3770 skb_walk_frags(skb, frag_iter)
3771 if (tcp_md5_hash_skb_data(hp, frag_iter, 0))
3776 EXPORT_SYMBOL(tcp_md5_hash_skb_data);
3778 int tcp_md5_hash_key(struct tcp_md5sig_pool *hp, const struct tcp_md5sig_key *key)
3780 u8 keylen = READ_ONCE(key->keylen); /* paired with WRITE_ONCE() in tcp_md5_do_add */
3781 struct scatterlist sg;
3783 sg_init_one(&sg, key->key, keylen);
3784 ahash_request_set_crypt(hp->md5_req, &sg, NULL, keylen);
3786 /* tcp_md5_do_add() might change key->key under us */
3787 return crypto_ahash_update(hp->md5_req);
3789 EXPORT_SYMBOL(tcp_md5_hash_key);
3793 void tcp_done(struct sock *sk)
3795 struct request_sock *req = tcp_sk(sk)->fastopen_rsk;
3797 if (sk->sk_state == TCP_SYN_SENT || sk->sk_state == TCP_SYN_RECV)
3798 TCP_INC_STATS(sock_net(sk), TCP_MIB_ATTEMPTFAILS);
3800 tcp_set_state(sk, TCP_CLOSE);
3801 tcp_clear_xmit_timers(sk);
3803 reqsk_fastopen_remove(sk, req, false);
3805 sk->sk_shutdown = SHUTDOWN_MASK;
3807 if (!sock_flag(sk, SOCK_DEAD))
3808 sk->sk_state_change(sk);
3810 inet_csk_destroy_sock(sk);
3812 EXPORT_SYMBOL_GPL(tcp_done);
3814 int tcp_abort(struct sock *sk, int err)
3816 if (!sk_fullsock(sk)) {
3817 if (sk->sk_state == TCP_NEW_SYN_RECV) {
3818 struct request_sock *req = inet_reqsk(sk);
3821 inet_csk_reqsk_queue_drop(req->rsk_listener, req);
3828 /* Don't race with userspace socket closes such as tcp_close. */
3831 if (sk->sk_state == TCP_LISTEN) {
3832 tcp_set_state(sk, TCP_CLOSE);
3833 inet_csk_listen_stop(sk);
3836 /* Don't race with BH socket closes such as inet_csk_listen_stop. */
3840 if (!sock_flag(sk, SOCK_DEAD)) {
3842 /* This barrier is coupled with smp_rmb() in tcp_poll() */
3844 sk->sk_error_report(sk);
3845 if (tcp_need_reset(sk->sk_state))
3846 tcp_send_active_reset(sk, GFP_ATOMIC);
3852 tcp_write_queue_purge(sk);
3856 EXPORT_SYMBOL_GPL(tcp_abort);
3858 extern struct tcp_congestion_ops tcp_reno;
3860 static __initdata unsigned long thash_entries;
3861 static int __init set_thash_entries(char *str)
3868 ret = kstrtoul(str, 0, &thash_entries);
3874 __setup("thash_entries=", set_thash_entries);
3876 static void __init tcp_init_mem(void)
3878 unsigned long limit = nr_free_buffer_pages() / 16;
3880 limit = max(limit, 128UL);
3881 sysctl_tcp_mem[0] = limit / 4 * 3; /* 4.68 % */
3882 sysctl_tcp_mem[1] = limit; /* 6.25 % */
3883 sysctl_tcp_mem[2] = sysctl_tcp_mem[0] * 2; /* 9.37 % */
3886 void __init tcp_init(void)
3888 int max_rshare, max_wshare, cnt;
3889 unsigned long limit;
3892 BUILD_BUG_ON(TCP_MIN_SND_MSS <= MAX_TCP_OPTION_SPACE);
3893 BUILD_BUG_ON(sizeof(struct tcp_skb_cb) >
3894 FIELD_SIZEOF(struct sk_buff, cb));
3896 percpu_counter_init(&tcp_sockets_allocated, 0, GFP_KERNEL);
3897 percpu_counter_init(&tcp_orphan_count, 0, GFP_KERNEL);
3898 inet_hashinfo_init(&tcp_hashinfo);
3899 inet_hashinfo2_init(&tcp_hashinfo, "tcp_listen_portaddr_hash",
3900 thash_entries, 21, /* one slot per 2 MB*/
3902 tcp_hashinfo.bind_bucket_cachep =
3903 kmem_cache_create("tcp_bind_bucket",
3904 sizeof(struct inet_bind_bucket), 0,
3905 SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL);
3907 /* Size and allocate the main established and bind bucket
3910 * The methodology is similar to that of the buffer cache.
3912 tcp_hashinfo.ehash =
3913 alloc_large_system_hash("TCP established",
3914 sizeof(struct inet_ehash_bucket),
3916 17, /* one slot per 128 KB of memory */
3919 &tcp_hashinfo.ehash_mask,
3921 thash_entries ? 0 : 512 * 1024);
3922 for (i = 0; i <= tcp_hashinfo.ehash_mask; i++)
3923 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].chain, i);
3925 if (inet_ehash_locks_alloc(&tcp_hashinfo))
3926 panic("TCP: failed to alloc ehash_locks");
3927 tcp_hashinfo.bhash =
3928 alloc_large_system_hash("TCP bind",
3929 sizeof(struct inet_bind_hashbucket),
3930 tcp_hashinfo.ehash_mask + 1,
3931 17, /* one slot per 128 KB of memory */
3933 &tcp_hashinfo.bhash_size,
3937 tcp_hashinfo.bhash_size = 1U << tcp_hashinfo.bhash_size;
3938 for (i = 0; i < tcp_hashinfo.bhash_size; i++) {
3939 spin_lock_init(&tcp_hashinfo.bhash[i].lock);
3940 INIT_HLIST_HEAD(&tcp_hashinfo.bhash[i].chain);
3944 cnt = tcp_hashinfo.ehash_mask + 1;
3945 sysctl_tcp_max_orphans = cnt / 2;
3948 /* Set per-socket limits to no more than 1/128 the pressure threshold */
3949 limit = nr_free_buffer_pages() << (PAGE_SHIFT - 7);
3950 max_wshare = min(4UL*1024*1024, limit);
3951 max_rshare = min(6UL*1024*1024, limit);
3953 init_net.ipv4.sysctl_tcp_wmem[0] = SK_MEM_QUANTUM;
3954 init_net.ipv4.sysctl_tcp_wmem[1] = 16*1024;
3955 init_net.ipv4.sysctl_tcp_wmem[2] = max(64*1024, max_wshare);
3957 init_net.ipv4.sysctl_tcp_rmem[0] = SK_MEM_QUANTUM;
3958 init_net.ipv4.sysctl_tcp_rmem[1] = 131072;
3959 init_net.ipv4.sysctl_tcp_rmem[2] = max(131072, max_rshare);
3961 pr_info("Hash tables configured (established %u bind %u)\n",
3962 tcp_hashinfo.ehash_mask + 1, tcp_hashinfo.bhash_size);
3966 BUG_ON(tcp_register_congestion_control(&tcp_reno) != 0);