1 // SPDX-License-Identifier: GPL-2.0-or-later
3 * INET An implementation of the TCP/IP protocol suite for the LINUX
4 * operating system. INET is implemented using the BSD Socket
5 * interface as the means of communication with the user level.
7 * Implementation of the Transmission Control Protocol(TCP).
10 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
11 * Mark Evans, <evansmp@uhura.aston.ac.uk>
12 * Corey Minyard <wf-rch!minyard@relay.EU.net>
13 * Florian La Roche, <flla@stud.uni-sb.de>
14 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
15 * Linus Torvalds, <torvalds@cs.helsinki.fi>
16 * Alan Cox, <gw4pts@gw4pts.ampr.org>
17 * Matthew Dillon, <dillon@apollo.west.oic.com>
18 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
19 * Jorge Cwik, <jorge@laser.satlink.net>
22 * Alan Cox : Numerous verify_area() calls
23 * Alan Cox : Set the ACK bit on a reset
24 * Alan Cox : Stopped it crashing if it closed while
25 * sk->inuse=1 and was trying to connect
27 * Alan Cox : All icmp error handling was broken
28 * pointers passed where wrong and the
29 * socket was looked up backwards. Nobody
30 * tested any icmp error code obviously.
31 * Alan Cox : tcp_err() now handled properly. It
32 * wakes people on errors. poll
33 * behaves and the icmp error race
34 * has gone by moving it into sock.c
35 * Alan Cox : tcp_send_reset() fixed to work for
36 * everything not just packets for
38 * Alan Cox : tcp option processing.
39 * Alan Cox : Reset tweaked (still not 100%) [Had
41 * Herp Rosmanith : More reset fixes
42 * Alan Cox : No longer acks invalid rst frames.
43 * Acking any kind of RST is right out.
44 * Alan Cox : Sets an ignore me flag on an rst
45 * receive otherwise odd bits of prattle
47 * Alan Cox : Fixed another acking RST frame bug.
48 * Should stop LAN workplace lockups.
49 * Alan Cox : Some tidyups using the new skb list
51 * Alan Cox : sk->keepopen now seems to work
52 * Alan Cox : Pulls options out correctly on accepts
53 * Alan Cox : Fixed assorted sk->rqueue->next errors
54 * Alan Cox : PSH doesn't end a TCP read. Switched a
56 * Alan Cox : Tidied tcp_data to avoid a potential
58 * Alan Cox : Added some better commenting, as the
59 * tcp is hard to follow
60 * Alan Cox : Removed incorrect check for 20 * psh
61 * Michael O'Reilly : ack < copied bug fix.
62 * Johannes Stille : Misc tcp fixes (not all in yet).
63 * Alan Cox : FIN with no memory -> CRASH
64 * Alan Cox : Added socket option proto entries.
65 * Also added awareness of them to accept.
66 * Alan Cox : Added TCP options (SOL_TCP)
67 * Alan Cox : Switched wakeup calls to callbacks,
68 * so the kernel can layer network
70 * Alan Cox : Use ip_tos/ip_ttl settings.
71 * Alan Cox : Handle FIN (more) properly (we hope).
72 * Alan Cox : RST frames sent on unsynchronised
74 * Alan Cox : Put in missing check for SYN bit.
75 * Alan Cox : Added tcp_select_window() aka NET2E
76 * window non shrink trick.
77 * Alan Cox : Added a couple of small NET2E timer
79 * Charles Hedrick : TCP fixes
80 * Toomas Tamm : TCP window fixes
81 * Alan Cox : Small URG fix to rlogin ^C ack fight
82 * Charles Hedrick : Rewrote most of it to actually work
83 * Linus : Rewrote tcp_read() and URG handling
85 * Gerhard Koerting: Fixed some missing timer handling
86 * Matthew Dillon : Reworked TCP machine states as per RFC
87 * Gerhard Koerting: PC/TCP workarounds
88 * Adam Caldwell : Assorted timer/timing errors
89 * Matthew Dillon : Fixed another RST bug
90 * Alan Cox : Move to kernel side addressing changes.
91 * Alan Cox : Beginning work on TCP fastpathing
93 * Arnt Gulbrandsen: Turbocharged tcp_check() routine.
94 * Alan Cox : TCP fast path debugging
95 * Alan Cox : Window clamping
96 * Michael Riepe : Bug in tcp_check()
97 * Matt Dillon : More TCP improvements and RST bug fixes
98 * Matt Dillon : Yet more small nasties remove from the
99 * TCP code (Be very nice to this man if
100 * tcp finally works 100%) 8)
101 * Alan Cox : BSD accept semantics.
102 * Alan Cox : Reset on closedown bug.
103 * Peter De Schrijver : ENOTCONN check missing in tcp_sendto().
104 * Michael Pall : Handle poll() after URG properly in
106 * Michael Pall : Undo the last fix in tcp_read_urg()
107 * (multi URG PUSH broke rlogin).
108 * Michael Pall : Fix the multi URG PUSH problem in
109 * tcp_readable(), poll() after URG
111 * Michael Pall : recv(...,MSG_OOB) never blocks in the
113 * Alan Cox : Changed the semantics of sk->socket to
114 * fix a race and a signal problem with
115 * accept() and async I/O.
116 * Alan Cox : Relaxed the rules on tcp_sendto().
117 * Yury Shevchuk : Really fixed accept() blocking problem.
118 * Craig I. Hagan : Allow for BSD compatible TIME_WAIT for
119 * clients/servers which listen in on
121 * Alan Cox : Cleaned the above up and shrank it to
122 * a sensible code size.
123 * Alan Cox : Self connect lockup fix.
124 * Alan Cox : No connect to multicast.
125 * Ross Biro : Close unaccepted children on master
127 * Alan Cox : Reset tracing code.
128 * Alan Cox : Spurious resets on shutdown.
129 * Alan Cox : Giant 15 minute/60 second timer error
130 * Alan Cox : Small whoops in polling before an
132 * Alan Cox : Kept the state trace facility since
133 * it's handy for debugging.
134 * Alan Cox : More reset handler fixes.
135 * Alan Cox : Started rewriting the code based on
136 * the RFC's for other useful protocol
137 * references see: Comer, KA9Q NOS, and
138 * for a reference on the difference
139 * between specifications and how BSD
140 * works see the 4.4lite source.
141 * A.N.Kuznetsov : Don't time wait on completion of tidy
143 * Linus Torvalds : Fin/Shutdown & copied_seq changes.
144 * Linus Torvalds : Fixed BSD port reuse to work first syn
145 * Alan Cox : Reimplemented timers as per the RFC
146 * and using multiple timers for sanity.
147 * Alan Cox : Small bug fixes, and a lot of new
149 * Alan Cox : Fixed dual reader crash by locking
150 * the buffers (much like datagram.c)
151 * Alan Cox : Fixed stuck sockets in probe. A probe
152 * now gets fed up of retrying without
153 * (even a no space) answer.
154 * Alan Cox : Extracted closing code better
155 * Alan Cox : Fixed the closing state machine to
157 * Alan Cox : More 'per spec' fixes.
158 * Jorge Cwik : Even faster checksumming.
159 * Alan Cox : tcp_data() doesn't ack illegal PSH
160 * only frames. At least one pc tcp stack
162 * Alan Cox : Cache last socket.
163 * Alan Cox : Per route irtt.
164 * Matt Day : poll()->select() match BSD precisely on error
165 * Alan Cox : New buffers
166 * Marc Tamsky : Various sk->prot->retransmits and
167 * sk->retransmits misupdating fixed.
168 * Fixed tcp_write_timeout: stuck close,
169 * and TCP syn retries gets used now.
170 * Mark Yarvis : In tcp_read_wakeup(), don't send an
171 * ack if state is TCP_CLOSED.
172 * Alan Cox : Look up device on a retransmit - routes may
173 * change. Doesn't yet cope with MSS shrink right
175 * Marc Tamsky : Closing in closing fixes.
176 * Mike Shaver : RFC1122 verifications.
177 * Alan Cox : rcv_saddr errors.
178 * Alan Cox : Block double connect().
179 * Alan Cox : Small hooks for enSKIP.
180 * Alexey Kuznetsov: Path MTU discovery.
181 * Alan Cox : Support soft errors.
182 * Alan Cox : Fix MTU discovery pathological case
183 * when the remote claims no mtu!
184 * Marc Tamsky : TCP_CLOSE fix.
185 * Colin (G3TNE) : Send a reset on syn ack replies in
186 * window but wrong (fixes NT lpd problems)
187 * Pedro Roque : Better TCP window handling, delayed ack.
188 * Joerg Reuter : No modification of locked buffers in
189 * tcp_do_retransmit()
190 * Eric Schenk : Changed receiver side silly window
191 * avoidance algorithm to BSD style
192 * algorithm. This doubles throughput
193 * against machines running Solaris,
194 * and seems to result in general
196 * Stefan Magdalinski : adjusted tcp_readable() to fix FIONREAD
197 * Willy Konynenberg : Transparent proxying support.
198 * Mike McLagan : Routing by source
199 * Keith Owens : Do proper merging with partial SKB's in
200 * tcp_do_sendmsg to avoid burstiness.
201 * Eric Schenk : Fix fast close down bug with
202 * shutdown() followed by close().
203 * Andi Kleen : Make poll agree with SIGIO
204 * Salvatore Sanfilippo : Support SO_LINGER with linger == 1 and
205 * lingertime == 0 (RFC 793 ABORT Call)
206 * Hirokazu Takahashi : Use copy_from_user() instead of
207 * csum_and_copy_from_user() if possible.
209 * Description of States:
211 * TCP_SYN_SENT sent a connection request, waiting for ack
213 * TCP_SYN_RECV received a connection request, sent ack,
214 * waiting for final ack in three-way handshake.
216 * TCP_ESTABLISHED connection established
218 * TCP_FIN_WAIT1 our side has shutdown, waiting to complete
219 * transmission of remaining buffered data
221 * TCP_FIN_WAIT2 all buffered data sent, waiting for remote
224 * TCP_CLOSING both sides have shutdown but we still have
225 * data we have to finish sending
227 * TCP_TIME_WAIT timeout to catch resent junk before entering
228 * closed, can only be entered from FIN_WAIT2
229 * or CLOSING. Required because the other end
230 * may not have gotten our last ACK causing it
231 * to retransmit the data packet (which we ignore)
233 * TCP_CLOSE_WAIT remote side has shutdown and is waiting for
234 * us to finish writing our data and to shutdown
235 * (we have to close() to move on to LAST_ACK)
237 * TCP_LAST_ACK out side has shutdown after remote has
238 * shutdown. There may still be data in our
239 * buffer that we have to finish sending
241 * TCP_CLOSE socket is finished
244 #define pr_fmt(fmt) "TCP: " fmt
246 #include <crypto/hash.h>
247 #include <linux/kernel.h>
248 #include <linux/module.h>
249 #include <linux/types.h>
250 #include <linux/fcntl.h>
251 #include <linux/poll.h>
252 #include <linux/inet_diag.h>
253 #include <linux/init.h>
254 #include <linux/fs.h>
255 #include <linux/skbuff.h>
256 #include <linux/scatterlist.h>
257 #include <linux/splice.h>
258 #include <linux/net.h>
259 #include <linux/socket.h>
260 #include <linux/random.h>
261 #include <linux/memblock.h>
262 #include <linux/highmem.h>
263 #include <linux/swap.h>
264 #include <linux/cache.h>
265 #include <linux/err.h>
266 #include <linux/time.h>
267 #include <linux/slab.h>
268 #include <linux/errqueue.h>
269 #include <linux/static_key.h>
271 #include <net/icmp.h>
272 #include <net/inet_common.h>
274 #include <net/xfrm.h>
276 #include <net/sock.h>
278 #include <linux/uaccess.h>
279 #include <asm/ioctls.h>
280 #include <net/busy_poll.h>
282 struct percpu_counter tcp_orphan_count;
283 EXPORT_SYMBOL_GPL(tcp_orphan_count);
285 long sysctl_tcp_mem[3] __read_mostly;
286 EXPORT_SYMBOL(sysctl_tcp_mem);
288 atomic_long_t tcp_memory_allocated; /* Current allocated memory. */
289 EXPORT_SYMBOL(tcp_memory_allocated);
291 #if IS_ENABLED(CONFIG_SMC)
292 DEFINE_STATIC_KEY_FALSE(tcp_have_smc);
293 EXPORT_SYMBOL(tcp_have_smc);
297 * Current number of TCP sockets.
299 struct percpu_counter tcp_sockets_allocated;
300 EXPORT_SYMBOL(tcp_sockets_allocated);
305 struct tcp_splice_state {
306 struct pipe_inode_info *pipe;
312 * Pressure flag: try to collapse.
313 * Technical note: it is used by multiple contexts non atomically.
314 * All the __sk_mem_schedule() is of this nature: accounting
315 * is strict, actions are advisory and have some latency.
317 unsigned long tcp_memory_pressure __read_mostly;
318 EXPORT_SYMBOL_GPL(tcp_memory_pressure);
320 DEFINE_STATIC_KEY_FALSE(tcp_rx_skb_cache_key);
321 EXPORT_SYMBOL(tcp_rx_skb_cache_key);
323 DEFINE_STATIC_KEY_FALSE(tcp_tx_skb_cache_key);
325 void tcp_enter_memory_pressure(struct sock *sk)
329 if (READ_ONCE(tcp_memory_pressure))
335 if (!cmpxchg(&tcp_memory_pressure, 0, val))
336 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMEMORYPRESSURES);
338 EXPORT_SYMBOL_GPL(tcp_enter_memory_pressure);
340 void tcp_leave_memory_pressure(struct sock *sk)
344 if (!READ_ONCE(tcp_memory_pressure))
346 val = xchg(&tcp_memory_pressure, 0);
348 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPMEMORYPRESSURESCHRONO,
349 jiffies_to_msecs(jiffies - val));
351 EXPORT_SYMBOL_GPL(tcp_leave_memory_pressure);
353 /* Convert seconds to retransmits based on initial and max timeout */
354 static u8 secs_to_retrans(int seconds, int timeout, int rto_max)
359 int period = timeout;
362 while (seconds > period && res < 255) {
365 if (timeout > rto_max)
373 /* Convert retransmits to seconds based on initial and max timeout */
374 static int retrans_to_secs(u8 retrans, int timeout, int rto_max)
382 if (timeout > rto_max)
390 static u64 tcp_compute_delivery_rate(const struct tcp_sock *tp)
392 u32 rate = READ_ONCE(tp->rate_delivered);
393 u32 intv = READ_ONCE(tp->rate_interval_us);
397 rate64 = (u64)rate * tp->mss_cache * USEC_PER_SEC;
398 do_div(rate64, intv);
403 /* Address-family independent initialization for a tcp_sock.
405 * NOTE: A lot of things set to zero explicitly by call to
406 * sk_alloc() so need not be done here.
408 void tcp_init_sock(struct sock *sk)
410 struct inet_connection_sock *icsk = inet_csk(sk);
411 struct tcp_sock *tp = tcp_sk(sk);
413 tp->out_of_order_queue = RB_ROOT;
414 sk->tcp_rtx_queue = RB_ROOT;
415 tcp_init_xmit_timers(sk);
416 INIT_LIST_HEAD(&tp->tsq_node);
417 INIT_LIST_HEAD(&tp->tsorted_sent_queue);
419 icsk->icsk_rto = TCP_TIMEOUT_INIT;
420 tp->mdev_us = jiffies_to_usecs(TCP_TIMEOUT_INIT);
421 minmax_reset(&tp->rtt_min, tcp_jiffies32, ~0U);
423 /* So many TCP implementations out there (incorrectly) count the
424 * initial SYN frame in their delayed-ACK and congestion control
425 * algorithms that we must have the following bandaid to talk
426 * efficiently to them. -DaveM
428 tp->snd_cwnd = TCP_INIT_CWND;
430 /* There's a bubble in the pipe until at least the first ACK. */
431 tp->app_limited = ~0U;
433 /* See draft-stevens-tcpca-spec-01 for discussion of the
434 * initialization of these values.
436 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
437 tp->snd_cwnd_clamp = ~0;
438 tp->mss_cache = TCP_MSS_DEFAULT;
440 tp->reordering = sock_net(sk)->ipv4.sysctl_tcp_reordering;
441 tcp_assign_congestion_control(sk);
444 tp->rack.reo_wnd_steps = 1;
446 sk->sk_state = TCP_CLOSE;
448 sk->sk_write_space = sk_stream_write_space;
449 sock_set_flag(sk, SOCK_USE_WRITE_QUEUE);
451 icsk->icsk_sync_mss = tcp_sync_mss;
453 WRITE_ONCE(sk->sk_sndbuf, sock_net(sk)->ipv4.sysctl_tcp_wmem[1]);
454 WRITE_ONCE(sk->sk_rcvbuf, sock_net(sk)->ipv4.sysctl_tcp_rmem[1]);
456 sk_sockets_allocated_inc(sk);
457 sk->sk_route_forced_caps = NETIF_F_GSO;
459 EXPORT_SYMBOL(tcp_init_sock);
461 static void tcp_tx_timestamp(struct sock *sk, u16 tsflags)
463 struct sk_buff *skb = tcp_write_queue_tail(sk);
465 if (tsflags && skb) {
466 struct skb_shared_info *shinfo = skb_shinfo(skb);
467 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
469 sock_tx_timestamp(sk, tsflags, &shinfo->tx_flags);
470 if (tsflags & SOF_TIMESTAMPING_TX_ACK)
471 tcb->txstamp_ack = 1;
472 if (tsflags & SOF_TIMESTAMPING_TX_RECORD_MASK)
473 shinfo->tskey = TCP_SKB_CB(skb)->seq + skb->len - 1;
477 static inline bool tcp_stream_is_readable(const struct tcp_sock *tp,
478 int target, struct sock *sk)
480 return (READ_ONCE(tp->rcv_nxt) - READ_ONCE(tp->copied_seq) >= target) ||
481 (sk->sk_prot->stream_memory_read ?
482 sk->sk_prot->stream_memory_read(sk) : false);
486 * Wait for a TCP event.
488 * Note that we don't need to lock the socket, as the upper poll layers
489 * take care of normal races (between the test and the event) and we don't
490 * go look at any of the socket buffers directly.
492 __poll_t tcp_poll(struct file *file, struct socket *sock, poll_table *wait)
495 struct sock *sk = sock->sk;
496 const struct tcp_sock *tp = tcp_sk(sk);
499 sock_poll_wait(file, sock, wait);
501 state = inet_sk_state_load(sk);
502 if (state == TCP_LISTEN)
503 return inet_csk_listen_poll(sk);
505 /* Socket is not locked. We are protected from async events
506 * by poll logic and correct handling of state changes
507 * made by other threads is impossible in any case.
513 * EPOLLHUP is certainly not done right. But poll() doesn't
514 * have a notion of HUP in just one direction, and for a
515 * socket the read side is more interesting.
517 * Some poll() documentation says that EPOLLHUP is incompatible
518 * with the EPOLLOUT/POLLWR flags, so somebody should check this
519 * all. But careful, it tends to be safer to return too many
520 * bits than too few, and you can easily break real applications
521 * if you don't tell them that something has hung up!
525 * Check number 1. EPOLLHUP is _UNMASKABLE_ event (see UNIX98 and
526 * our fs/select.c). It means that after we received EOF,
527 * poll always returns immediately, making impossible poll() on write()
528 * in state CLOSE_WAIT. One solution is evident --- to set EPOLLHUP
529 * if and only if shutdown has been made in both directions.
530 * Actually, it is interesting to look how Solaris and DUX
531 * solve this dilemma. I would prefer, if EPOLLHUP were maskable,
532 * then we could set it on SND_SHUTDOWN. BTW examples given
533 * in Stevens' books assume exactly this behaviour, it explains
534 * why EPOLLHUP is incompatible with EPOLLOUT. --ANK
536 * NOTE. Check for TCP_CLOSE is added. The goal is to prevent
537 * blocking on fresh not-connected or disconnected socket. --ANK
539 if (sk->sk_shutdown == SHUTDOWN_MASK || state == TCP_CLOSE)
541 if (sk->sk_shutdown & RCV_SHUTDOWN)
542 mask |= EPOLLIN | EPOLLRDNORM | EPOLLRDHUP;
544 /* Connected or passive Fast Open socket? */
545 if (state != TCP_SYN_SENT &&
546 (state != TCP_SYN_RECV || rcu_access_pointer(tp->fastopen_rsk))) {
547 int target = sock_rcvlowat(sk, 0, INT_MAX);
549 if (READ_ONCE(tp->urg_seq) == READ_ONCE(tp->copied_seq) &&
550 !sock_flag(sk, SOCK_URGINLINE) &&
554 if (tcp_stream_is_readable(tp, target, sk))
555 mask |= EPOLLIN | EPOLLRDNORM;
557 if (!(sk->sk_shutdown & SEND_SHUTDOWN)) {
558 if (sk_stream_is_writeable(sk)) {
559 mask |= EPOLLOUT | EPOLLWRNORM;
560 } else { /* send SIGIO later */
561 sk_set_bit(SOCKWQ_ASYNC_NOSPACE, sk);
562 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
564 /* Race breaker. If space is freed after
565 * wspace test but before the flags are set,
566 * IO signal will be lost. Memory barrier
567 * pairs with the input side.
569 smp_mb__after_atomic();
570 if (sk_stream_is_writeable(sk))
571 mask |= EPOLLOUT | EPOLLWRNORM;
574 mask |= EPOLLOUT | EPOLLWRNORM;
576 if (tp->urg_data & TCP_URG_VALID)
578 } else if (state == TCP_SYN_SENT && inet_sk(sk)->defer_connect) {
579 /* Active TCP fastopen socket with defer_connect
580 * Return EPOLLOUT so application can call write()
581 * in order for kernel to generate SYN+data
583 mask |= EPOLLOUT | EPOLLWRNORM;
585 /* This barrier is coupled with smp_wmb() in tcp_reset() */
587 if (sk->sk_err || !skb_queue_empty_lockless(&sk->sk_error_queue))
592 EXPORT_SYMBOL(tcp_poll);
594 int tcp_ioctl(struct sock *sk, int cmd, unsigned long arg)
596 struct tcp_sock *tp = tcp_sk(sk);
602 if (sk->sk_state == TCP_LISTEN)
605 slow = lock_sock_fast(sk);
607 unlock_sock_fast(sk, slow);
610 answ = tp->urg_data &&
611 READ_ONCE(tp->urg_seq) == READ_ONCE(tp->copied_seq);
614 if (sk->sk_state == TCP_LISTEN)
617 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
620 answ = READ_ONCE(tp->write_seq) - tp->snd_una;
623 if (sk->sk_state == TCP_LISTEN)
626 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
629 answ = READ_ONCE(tp->write_seq) -
630 READ_ONCE(tp->snd_nxt);
636 return put_user(answ, (int __user *)arg);
638 EXPORT_SYMBOL(tcp_ioctl);
640 static inline void tcp_mark_push(struct tcp_sock *tp, struct sk_buff *skb)
642 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
643 tp->pushed_seq = tp->write_seq;
646 static inline bool forced_push(const struct tcp_sock *tp)
648 return after(tp->write_seq, tp->pushed_seq + (tp->max_window >> 1));
651 static void skb_entail(struct sock *sk, struct sk_buff *skb)
653 struct tcp_sock *tp = tcp_sk(sk);
654 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
657 tcb->seq = tcb->end_seq = tp->write_seq;
658 tcb->tcp_flags = TCPHDR_ACK;
660 __skb_header_release(skb);
661 tcp_add_write_queue_tail(sk, skb);
662 sk_wmem_queued_add(sk, skb->truesize);
663 sk_mem_charge(sk, skb->truesize);
664 if (tp->nonagle & TCP_NAGLE_PUSH)
665 tp->nonagle &= ~TCP_NAGLE_PUSH;
667 tcp_slow_start_after_idle_check(sk);
670 static inline void tcp_mark_urg(struct tcp_sock *tp, int flags)
673 tp->snd_up = tp->write_seq;
676 /* If a not yet filled skb is pushed, do not send it if
677 * we have data packets in Qdisc or NIC queues :
678 * Because TX completion will happen shortly, it gives a chance
679 * to coalesce future sendmsg() payload into this skb, without
680 * need for a timer, and with no latency trade off.
681 * As packets containing data payload have a bigger truesize
682 * than pure acks (dataless) packets, the last checks prevent
683 * autocorking if we only have an ACK in Qdisc/NIC queues,
684 * or if TX completion was delayed after we processed ACK packet.
686 static bool tcp_should_autocork(struct sock *sk, struct sk_buff *skb,
689 return skb->len < size_goal &&
690 sock_net(sk)->ipv4.sysctl_tcp_autocorking &&
691 !tcp_rtx_queue_empty(sk) &&
692 refcount_read(&sk->sk_wmem_alloc) > skb->truesize;
695 static void tcp_push(struct sock *sk, int flags, int mss_now,
696 int nonagle, int size_goal)
698 struct tcp_sock *tp = tcp_sk(sk);
701 skb = tcp_write_queue_tail(sk);
704 if (!(flags & MSG_MORE) || forced_push(tp))
705 tcp_mark_push(tp, skb);
707 tcp_mark_urg(tp, flags);
709 if (tcp_should_autocork(sk, skb, size_goal)) {
711 /* avoid atomic op if TSQ_THROTTLED bit is already set */
712 if (!test_bit(TSQ_THROTTLED, &sk->sk_tsq_flags)) {
713 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPAUTOCORKING);
714 set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags);
716 /* It is possible TX completion already happened
717 * before we set TSQ_THROTTLED.
719 if (refcount_read(&sk->sk_wmem_alloc) > skb->truesize)
723 if (flags & MSG_MORE)
724 nonagle = TCP_NAGLE_CORK;
726 __tcp_push_pending_frames(sk, mss_now, nonagle);
729 static int tcp_splice_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb,
730 unsigned int offset, size_t len)
732 struct tcp_splice_state *tss = rd_desc->arg.data;
735 ret = skb_splice_bits(skb, skb->sk, offset, tss->pipe,
736 min(rd_desc->count, len), tss->flags);
738 rd_desc->count -= ret;
742 static int __tcp_splice_read(struct sock *sk, struct tcp_splice_state *tss)
744 /* Store TCP splice context information in read_descriptor_t. */
745 read_descriptor_t rd_desc = {
750 return tcp_read_sock(sk, &rd_desc, tcp_splice_data_recv);
754 * tcp_splice_read - splice data from TCP socket to a pipe
755 * @sock: socket to splice from
756 * @ppos: position (not valid)
757 * @pipe: pipe to splice to
758 * @len: number of bytes to splice
759 * @flags: splice modifier flags
762 * Will read pages from given socket and fill them into a pipe.
765 ssize_t tcp_splice_read(struct socket *sock, loff_t *ppos,
766 struct pipe_inode_info *pipe, size_t len,
769 struct sock *sk = sock->sk;
770 struct tcp_splice_state tss = {
779 sock_rps_record_flow(sk);
781 * We can't seek on a socket input
790 timeo = sock_rcvtimeo(sk, sock->file->f_flags & O_NONBLOCK);
792 ret = __tcp_splice_read(sk, &tss);
798 if (sock_flag(sk, SOCK_DONE))
801 ret = sock_error(sk);
804 if (sk->sk_shutdown & RCV_SHUTDOWN)
806 if (sk->sk_state == TCP_CLOSE) {
808 * This occurs when user tries to read
809 * from never connected socket.
818 /* if __tcp_splice_read() got nothing while we have
819 * an skb in receive queue, we do not want to loop.
820 * This might happen with URG data.
822 if (!skb_queue_empty(&sk->sk_receive_queue))
824 sk_wait_data(sk, &timeo, NULL);
825 if (signal_pending(current)) {
826 ret = sock_intr_errno(timeo);
839 if (sk->sk_err || sk->sk_state == TCP_CLOSE ||
840 (sk->sk_shutdown & RCV_SHUTDOWN) ||
841 signal_pending(current))
852 EXPORT_SYMBOL(tcp_splice_read);
854 struct sk_buff *sk_stream_alloc_skb(struct sock *sk, int size, gfp_t gfp,
860 skb = sk->sk_tx_skb_cache;
862 skb->truesize = SKB_TRUESIZE(skb_end_offset(skb));
863 sk->sk_tx_skb_cache = NULL;
865 INIT_LIST_HEAD(&skb->tcp_tsorted_anchor);
866 skb_shinfo(skb)->tx_flags = 0;
867 memset(TCP_SKB_CB(skb), 0, sizeof(struct tcp_skb_cb));
871 /* The TCP header must be at least 32-bit aligned. */
872 size = ALIGN(size, 4);
874 if (unlikely(tcp_under_memory_pressure(sk)))
875 sk_mem_reclaim_partial(sk);
877 skb = alloc_skb_fclone(size + sk->sk_prot->max_header, gfp);
881 if (force_schedule) {
882 mem_scheduled = true;
883 sk_forced_mem_schedule(sk, skb->truesize);
885 mem_scheduled = sk_wmem_schedule(sk, skb->truesize);
887 if (likely(mem_scheduled)) {
888 skb_reserve(skb, sk->sk_prot->max_header);
890 * Make sure that we have exactly size bytes
891 * available to the caller, no more, no less.
893 skb->reserved_tailroom = skb->end - skb->tail - size;
894 INIT_LIST_HEAD(&skb->tcp_tsorted_anchor);
899 sk->sk_prot->enter_memory_pressure(sk);
900 sk_stream_moderate_sndbuf(sk);
905 static unsigned int tcp_xmit_size_goal(struct sock *sk, u32 mss_now,
908 struct tcp_sock *tp = tcp_sk(sk);
909 u32 new_size_goal, size_goal;
914 /* Note : tcp_tso_autosize() will eventually split this later */
915 new_size_goal = sk->sk_gso_max_size - 1 - MAX_TCP_HEADER;
916 new_size_goal = tcp_bound_to_half_wnd(tp, new_size_goal);
918 /* We try hard to avoid divides here */
919 size_goal = tp->gso_segs * mss_now;
920 if (unlikely(new_size_goal < size_goal ||
921 new_size_goal >= size_goal + mss_now)) {
922 tp->gso_segs = min_t(u16, new_size_goal / mss_now,
923 sk->sk_gso_max_segs);
924 size_goal = tp->gso_segs * mss_now;
927 return max(size_goal, mss_now);
930 static int tcp_send_mss(struct sock *sk, int *size_goal, int flags)
934 mss_now = tcp_current_mss(sk);
935 *size_goal = tcp_xmit_size_goal(sk, mss_now, !(flags & MSG_OOB));
940 /* In some cases, both sendpage() and sendmsg() could have added
941 * an skb to the write queue, but failed adding payload on it.
942 * We need to remove it to consume less memory, but more
943 * importantly be able to generate EPOLLOUT for Edge Trigger epoll()
946 static void tcp_remove_empty_skb(struct sock *sk, struct sk_buff *skb)
948 if (skb && !skb->len) {
949 tcp_unlink_write_queue(skb, sk);
950 if (tcp_write_queue_empty(sk))
951 tcp_chrono_stop(sk, TCP_CHRONO_BUSY);
952 sk_wmem_free_skb(sk, skb);
956 ssize_t do_tcp_sendpages(struct sock *sk, struct page *page, int offset,
957 size_t size, int flags)
959 struct tcp_sock *tp = tcp_sk(sk);
960 int mss_now, size_goal;
963 long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
965 if (IS_ENABLED(CONFIG_DEBUG_VM) &&
966 WARN_ONCE(PageSlab(page), "page must not be a Slab one"))
969 /* Wait for a connection to finish. One exception is TCP Fast Open
970 * (passive side) where data is allowed to be sent before a connection
971 * is fully established.
973 if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) &&
974 !tcp_passive_fastopen(sk)) {
975 err = sk_stream_wait_connect(sk, &timeo);
980 sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
982 mss_now = tcp_send_mss(sk, &size_goal, flags);
986 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
990 struct sk_buff *skb = tcp_write_queue_tail(sk);
994 if (!skb || (copy = size_goal - skb->len) <= 0 ||
995 !tcp_skb_can_collapse_to(skb)) {
997 if (!sk_stream_memory_free(sk))
998 goto wait_for_sndbuf;
1000 skb = sk_stream_alloc_skb(sk, 0, sk->sk_allocation,
1001 tcp_rtx_and_write_queues_empty(sk));
1003 goto wait_for_memory;
1005 #ifdef CONFIG_TLS_DEVICE
1006 skb->decrypted = !!(flags & MSG_SENDPAGE_DECRYPTED);
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_wmem_queued_add(sk, 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(skb_queue_len(&sk->sk_write_queue) == 0 &&
1092 sk->sk_write_space(sk);
1093 tcp_chrono_stop(sk, TCP_CHRONO_SNDBUF_LIMITED);
1095 return sk_stream_error(sk, flags, err);
1097 EXPORT_SYMBOL_GPL(do_tcp_sendpages);
1099 int tcp_sendpage_locked(struct sock *sk, struct page *page, int offset,
1100 size_t size, int flags)
1102 if (!(sk->sk_route_caps & NETIF_F_SG))
1103 return sock_no_sendpage_locked(sk, page, offset, size, flags);
1105 tcp_rate_check_app_limited(sk); /* is sending application-limited? */
1107 return do_tcp_sendpages(sk, page, offset, size, flags);
1109 EXPORT_SYMBOL_GPL(tcp_sendpage_locked);
1111 int tcp_sendpage(struct sock *sk, struct page *page, int offset,
1112 size_t size, int flags)
1117 ret = tcp_sendpage_locked(sk, page, offset, size, flags);
1122 EXPORT_SYMBOL(tcp_sendpage);
1124 void tcp_free_fastopen_req(struct tcp_sock *tp)
1126 if (tp->fastopen_req) {
1127 kfree(tp->fastopen_req);
1128 tp->fastopen_req = NULL;
1132 static int tcp_sendmsg_fastopen(struct sock *sk, struct msghdr *msg,
1133 int *copied, size_t size,
1134 struct ubuf_info *uarg)
1136 struct tcp_sock *tp = tcp_sk(sk);
1137 struct inet_sock *inet = inet_sk(sk);
1138 struct sockaddr *uaddr = msg->msg_name;
1141 if (!(sock_net(sk)->ipv4.sysctl_tcp_fastopen & TFO_CLIENT_ENABLE) ||
1142 (uaddr && msg->msg_namelen >= sizeof(uaddr->sa_family) &&
1143 uaddr->sa_family == AF_UNSPEC))
1145 if (tp->fastopen_req)
1146 return -EALREADY; /* Another Fast Open is in progress */
1148 tp->fastopen_req = kzalloc(sizeof(struct tcp_fastopen_request),
1150 if (unlikely(!tp->fastopen_req))
1152 tp->fastopen_req->data = msg;
1153 tp->fastopen_req->size = size;
1154 tp->fastopen_req->uarg = uarg;
1156 if (inet->defer_connect) {
1157 err = tcp_connect(sk);
1158 /* Same failure procedure as in tcp_v4/6_connect */
1160 tcp_set_state(sk, TCP_CLOSE);
1161 inet->inet_dport = 0;
1162 sk->sk_route_caps = 0;
1165 flags = (msg->msg_flags & MSG_DONTWAIT) ? O_NONBLOCK : 0;
1166 err = __inet_stream_connect(sk->sk_socket, uaddr,
1167 msg->msg_namelen, flags, 1);
1168 /* fastopen_req could already be freed in __inet_stream_connect
1169 * if the connection times out or gets rst
1171 if (tp->fastopen_req) {
1172 *copied = tp->fastopen_req->copied;
1173 tcp_free_fastopen_req(tp);
1174 inet->defer_connect = 0;
1179 int tcp_sendmsg_locked(struct sock *sk, struct msghdr *msg, size_t size)
1181 struct tcp_sock *tp = tcp_sk(sk);
1182 struct ubuf_info *uarg = NULL;
1183 struct sk_buff *skb;
1184 struct sockcm_cookie sockc;
1185 int flags, err, copied = 0;
1186 int mss_now = 0, size_goal, copied_syn = 0;
1187 int process_backlog = 0;
1191 flags = msg->msg_flags;
1193 if (flags & MSG_ZEROCOPY && size && sock_flag(sk, SOCK_ZEROCOPY)) {
1194 skb = tcp_write_queue_tail(sk);
1195 uarg = sock_zerocopy_realloc(sk, size, skb_zcopy(skb));
1201 zc = sk->sk_route_caps & NETIF_F_SG;
1206 if (unlikely(flags & MSG_FASTOPEN || inet_sk(sk)->defer_connect) &&
1208 err = tcp_sendmsg_fastopen(sk, msg, &copied_syn, size, uarg);
1209 if (err == -EINPROGRESS && copied_syn > 0)
1215 timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
1217 tcp_rate_check_app_limited(sk); /* is sending application-limited? */
1219 /* Wait for a connection to finish. One exception is TCP Fast Open
1220 * (passive side) where data is allowed to be sent before a connection
1221 * is fully established.
1223 if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) &&
1224 !tcp_passive_fastopen(sk)) {
1225 err = sk_stream_wait_connect(sk, &timeo);
1230 if (unlikely(tp->repair)) {
1231 if (tp->repair_queue == TCP_RECV_QUEUE) {
1232 copied = tcp_send_rcvq(sk, msg, size);
1237 if (tp->repair_queue == TCP_NO_QUEUE)
1240 /* 'common' sending to sendq */
1243 sockcm_init(&sockc, sk);
1244 if (msg->msg_controllen) {
1245 err = sock_cmsg_send(sk, msg, &sockc);
1246 if (unlikely(err)) {
1252 /* This should be in poll */
1253 sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
1255 /* Ok commence sending. */
1259 mss_now = tcp_send_mss(sk, &size_goal, flags);
1262 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
1265 while (msg_data_left(msg)) {
1268 skb = tcp_write_queue_tail(sk);
1270 copy = size_goal - skb->len;
1272 if (copy <= 0 || !tcp_skb_can_collapse_to(skb)) {
1276 if (!sk_stream_memory_free(sk))
1277 goto wait_for_sndbuf;
1279 if (unlikely(process_backlog >= 16)) {
1280 process_backlog = 0;
1281 if (sk_flush_backlog(sk))
1284 first_skb = tcp_rtx_and_write_queues_empty(sk);
1285 skb = sk_stream_alloc_skb(sk, 0, sk->sk_allocation,
1288 goto wait_for_memory;
1291 skb->ip_summed = CHECKSUM_PARTIAL;
1293 skb_entail(sk, skb);
1296 /* All packets are restored as if they have
1297 * already been sent. skb_mstamp_ns isn't set to
1298 * avoid wrong rtt estimation.
1301 TCP_SKB_CB(skb)->sacked |= TCPCB_REPAIRED;
1304 /* Try to append data to the end of skb. */
1305 if (copy > msg_data_left(msg))
1306 copy = msg_data_left(msg);
1308 /* Where to copy to? */
1309 if (skb_availroom(skb) > 0 && !zc) {
1310 /* We have some space in skb head. Superb! */
1311 copy = min_t(int, copy, skb_availroom(skb));
1312 err = skb_add_data_nocache(sk, skb, &msg->msg_iter, copy);
1317 int i = skb_shinfo(skb)->nr_frags;
1318 struct page_frag *pfrag = sk_page_frag(sk);
1320 if (!sk_page_frag_refill(sk, pfrag))
1321 goto wait_for_memory;
1323 if (!skb_can_coalesce(skb, i, pfrag->page,
1325 if (i >= sysctl_max_skb_frags) {
1326 tcp_mark_push(tp, skb);
1332 copy = min_t(int, copy, pfrag->size - pfrag->offset);
1334 if (!sk_wmem_schedule(sk, copy))
1335 goto wait_for_memory;
1337 err = skb_copy_to_page_nocache(sk, &msg->msg_iter, skb,
1344 /* Update the skb. */
1346 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1348 skb_fill_page_desc(skb, i, pfrag->page,
1349 pfrag->offset, copy);
1350 page_ref_inc(pfrag->page);
1352 pfrag->offset += copy;
1354 err = skb_zerocopy_iter_stream(sk, skb, msg, copy, uarg);
1355 if (err == -EMSGSIZE || err == -EEXIST) {
1356 tcp_mark_push(tp, skb);
1365 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
1367 WRITE_ONCE(tp->write_seq, tp->write_seq + copy);
1368 TCP_SKB_CB(skb)->end_seq += copy;
1369 tcp_skb_pcount_set(skb, 0);
1372 if (!msg_data_left(msg)) {
1373 if (unlikely(flags & MSG_EOR))
1374 TCP_SKB_CB(skb)->eor = 1;
1378 if (skb->len < size_goal || (flags & MSG_OOB) || unlikely(tp->repair))
1381 if (forced_push(tp)) {
1382 tcp_mark_push(tp, skb);
1383 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
1384 } else if (skb == tcp_send_head(sk))
1385 tcp_push_one(sk, mss_now);
1389 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1392 tcp_push(sk, flags & ~MSG_MORE, mss_now,
1393 TCP_NAGLE_PUSH, size_goal);
1395 err = sk_stream_wait_memory(sk, &timeo);
1399 mss_now = tcp_send_mss(sk, &size_goal, flags);
1404 tcp_tx_timestamp(sk, sockc.tsflags);
1405 tcp_push(sk, flags, mss_now, tp->nonagle, size_goal);
1408 sock_zerocopy_put(uarg);
1409 return copied + copied_syn;
1412 skb = tcp_write_queue_tail(sk);
1414 tcp_remove_empty_skb(sk, skb);
1416 if (copied + copied_syn)
1419 sock_zerocopy_put_abort(uarg, true);
1420 err = sk_stream_error(sk, flags, err);
1421 /* make sure we wake any epoll edge trigger waiter */
1422 if (unlikely(skb_queue_len(&sk->sk_write_queue) == 0 &&
1424 sk->sk_write_space(sk);
1425 tcp_chrono_stop(sk, TCP_CHRONO_SNDBUF_LIMITED);
1429 EXPORT_SYMBOL_GPL(tcp_sendmsg_locked);
1431 int tcp_sendmsg(struct sock *sk, struct msghdr *msg, size_t size)
1436 ret = tcp_sendmsg_locked(sk, msg, size);
1441 EXPORT_SYMBOL(tcp_sendmsg);
1444 * Handle reading urgent data. BSD has very simple semantics for
1445 * this, no blocking and very strange errors 8)
1448 static int tcp_recv_urg(struct sock *sk, struct msghdr *msg, int len, int flags)
1450 struct tcp_sock *tp = tcp_sk(sk);
1452 /* No URG data to read. */
1453 if (sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data ||
1454 tp->urg_data == TCP_URG_READ)
1455 return -EINVAL; /* Yes this is right ! */
1457 if (sk->sk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DONE))
1460 if (tp->urg_data & TCP_URG_VALID) {
1462 char c = tp->urg_data;
1464 if (!(flags & MSG_PEEK))
1465 tp->urg_data = TCP_URG_READ;
1467 /* Read urgent data. */
1468 msg->msg_flags |= MSG_OOB;
1471 if (!(flags & MSG_TRUNC))
1472 err = memcpy_to_msg(msg, &c, 1);
1475 msg->msg_flags |= MSG_TRUNC;
1477 return err ? -EFAULT : len;
1480 if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
1483 /* Fixed the recv(..., MSG_OOB) behaviour. BSD docs and
1484 * the available implementations agree in this case:
1485 * this call should never block, independent of the
1486 * blocking state of the socket.
1487 * Mike <pall@rz.uni-karlsruhe.de>
1492 static int tcp_peek_sndq(struct sock *sk, struct msghdr *msg, int len)
1494 struct sk_buff *skb;
1495 int copied = 0, err = 0;
1497 /* XXX -- need to support SO_PEEK_OFF */
1499 skb_rbtree_walk(skb, &sk->tcp_rtx_queue) {
1500 err = skb_copy_datagram_msg(skb, 0, msg, skb->len);
1506 skb_queue_walk(&sk->sk_write_queue, skb) {
1507 err = skb_copy_datagram_msg(skb, 0, msg, skb->len);
1514 return err ?: copied;
1517 /* Clean up the receive buffer for full frames taken by the user,
1518 * then send an ACK if necessary. COPIED is the number of bytes
1519 * tcp_recvmsg has given to the user so far, it speeds up the
1520 * calculation of whether or not we must ACK for the sake of
1523 static void tcp_cleanup_rbuf(struct sock *sk, int copied)
1525 struct tcp_sock *tp = tcp_sk(sk);
1526 bool time_to_ack = false;
1528 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
1530 WARN(skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq),
1531 "cleanup rbuf bug: copied %X seq %X rcvnxt %X\n",
1532 tp->copied_seq, TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt);
1534 if (inet_csk_ack_scheduled(sk)) {
1535 const struct inet_connection_sock *icsk = inet_csk(sk);
1536 /* Delayed ACKs frequently hit locked sockets during bulk
1538 if (icsk->icsk_ack.blocked ||
1539 /* Once-per-two-segments ACK was not sent by tcp_input.c */
1540 tp->rcv_nxt - tp->rcv_wup > icsk->icsk_ack.rcv_mss ||
1542 * If this read emptied read buffer, we send ACK, if
1543 * connection is not bidirectional, user drained
1544 * receive buffer and there was a small segment
1548 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED2) ||
1549 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED) &&
1550 !inet_csk_in_pingpong_mode(sk))) &&
1551 !atomic_read(&sk->sk_rmem_alloc)))
1555 /* We send an ACK if we can now advertise a non-zero window
1556 * which has been raised "significantly".
1558 * Even if window raised up to infinity, do not send window open ACK
1559 * in states, where we will not receive more. It is useless.
1561 if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) {
1562 __u32 rcv_window_now = tcp_receive_window(tp);
1564 /* Optimize, __tcp_select_window() is not cheap. */
1565 if (2*rcv_window_now <= tp->window_clamp) {
1566 __u32 new_window = __tcp_select_window(sk);
1568 /* Send ACK now, if this read freed lots of space
1569 * in our buffer. Certainly, new_window is new window.
1570 * We can advertise it now, if it is not less than current one.
1571 * "Lots" means "at least twice" here.
1573 if (new_window && new_window >= 2 * rcv_window_now)
1581 static struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off)
1583 struct sk_buff *skb;
1586 while ((skb = skb_peek(&sk->sk_receive_queue)) != NULL) {
1587 offset = seq - TCP_SKB_CB(skb)->seq;
1588 if (unlikely(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
1589 pr_err_once("%s: found a SYN, please report !\n", __func__);
1592 if (offset < skb->len || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)) {
1596 /* This looks weird, but this can happen if TCP collapsing
1597 * splitted a fat GRO packet, while we released socket lock
1598 * in skb_splice_bits()
1600 sk_eat_skb(sk, skb);
1606 * This routine provides an alternative to tcp_recvmsg() for routines
1607 * that would like to handle copying from skbuffs directly in 'sendfile'
1610 * - It is assumed that the socket was locked by the caller.
1611 * - The routine does not block.
1612 * - At present, there is no support for reading OOB data
1613 * or for 'peeking' the socket using this routine
1614 * (although both would be easy to implement).
1616 int tcp_read_sock(struct sock *sk, read_descriptor_t *desc,
1617 sk_read_actor_t recv_actor)
1619 struct sk_buff *skb;
1620 struct tcp_sock *tp = tcp_sk(sk);
1621 u32 seq = tp->copied_seq;
1625 if (sk->sk_state == TCP_LISTEN)
1627 while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1628 if (offset < skb->len) {
1632 len = skb->len - offset;
1633 /* Stop reading if we hit a patch of urgent data */
1635 u32 urg_offset = tp->urg_seq - seq;
1636 if (urg_offset < len)
1641 used = recv_actor(desc, skb, offset, len);
1646 } else if (used <= len) {
1651 /* If recv_actor drops the lock (e.g. TCP splice
1652 * receive) the skb pointer might be invalid when
1653 * getting here: tcp_collapse might have deleted it
1654 * while aggregating skbs from the socket queue.
1656 skb = tcp_recv_skb(sk, seq - 1, &offset);
1659 /* TCP coalescing might have appended data to the skb.
1660 * Try to splice more frags
1662 if (offset + 1 != skb->len)
1665 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) {
1666 sk_eat_skb(sk, skb);
1670 sk_eat_skb(sk, skb);
1673 WRITE_ONCE(tp->copied_seq, seq);
1675 WRITE_ONCE(tp->copied_seq, seq);
1677 tcp_rcv_space_adjust(sk);
1679 /* Clean up data we have read: This will do ACK frames. */
1681 tcp_recv_skb(sk, seq, &offset);
1682 tcp_cleanup_rbuf(sk, copied);
1686 EXPORT_SYMBOL(tcp_read_sock);
1688 int tcp_peek_len(struct socket *sock)
1690 return tcp_inq(sock->sk);
1692 EXPORT_SYMBOL(tcp_peek_len);
1694 /* Make sure sk_rcvbuf is big enough to satisfy SO_RCVLOWAT hint */
1695 int tcp_set_rcvlowat(struct sock *sk, int val)
1699 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK)
1700 cap = sk->sk_rcvbuf >> 1;
1702 cap = sock_net(sk)->ipv4.sysctl_tcp_rmem[2] >> 1;
1703 val = min(val, cap);
1704 WRITE_ONCE(sk->sk_rcvlowat, val ? : 1);
1706 /* Check if we need to signal EPOLLIN right now */
1709 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK)
1713 if (val > sk->sk_rcvbuf) {
1714 WRITE_ONCE(sk->sk_rcvbuf, val);
1715 tcp_sk(sk)->window_clamp = tcp_win_from_space(sk, val);
1719 EXPORT_SYMBOL(tcp_set_rcvlowat);
1722 static const struct vm_operations_struct tcp_vm_ops = {
1725 int tcp_mmap(struct file *file, struct socket *sock,
1726 struct vm_area_struct *vma)
1728 if (vma->vm_flags & (VM_WRITE | VM_EXEC))
1730 vma->vm_flags &= ~(VM_MAYWRITE | VM_MAYEXEC);
1732 /* Instruct vm_insert_page() to not down_read(mmap_sem) */
1733 vma->vm_flags |= VM_MIXEDMAP;
1735 vma->vm_ops = &tcp_vm_ops;
1738 EXPORT_SYMBOL(tcp_mmap);
1740 static int tcp_zerocopy_receive(struct sock *sk,
1741 struct tcp_zerocopy_receive *zc)
1743 unsigned long address = (unsigned long)zc->address;
1744 u32 length = 0, seq, offset, zap_len;
1745 const skb_frag_t *frags = NULL;
1746 struct vm_area_struct *vma;
1747 struct sk_buff *skb = NULL;
1748 struct tcp_sock *tp;
1752 if (address & (PAGE_SIZE - 1) || address != zc->address)
1755 if (sk->sk_state == TCP_LISTEN)
1758 sock_rps_record_flow(sk);
1760 down_read(¤t->mm->mmap_sem);
1763 vma = find_vma(current->mm, address);
1764 if (!vma || vma->vm_start > address || vma->vm_ops != &tcp_vm_ops)
1766 zc->length = min_t(unsigned long, zc->length, vma->vm_end - address);
1769 seq = tp->copied_seq;
1771 zc->length = min_t(u32, zc->length, inq);
1772 zap_len = zc->length & ~(PAGE_SIZE - 1);
1774 zap_page_range(vma, address, zap_len);
1775 zc->recv_skip_hint = 0;
1777 zc->recv_skip_hint = zc->length;
1780 while (length + PAGE_SIZE <= zc->length) {
1781 if (zc->recv_skip_hint < PAGE_SIZE) {
1784 offset = seq - TCP_SKB_CB(skb)->seq;
1786 skb = tcp_recv_skb(sk, seq, &offset);
1789 zc->recv_skip_hint = skb->len - offset;
1790 offset -= skb_headlen(skb);
1791 if ((int)offset < 0 || skb_has_frag_list(skb))
1793 frags = skb_shinfo(skb)->frags;
1795 if (skb_frag_size(frags) > offset)
1797 offset -= skb_frag_size(frags);
1801 if (skb_frag_size(frags) != PAGE_SIZE || skb_frag_off(frags)) {
1802 int remaining = zc->recv_skip_hint;
1804 while (remaining && (skb_frag_size(frags) != PAGE_SIZE ||
1805 skb_frag_off(frags))) {
1806 remaining -= skb_frag_size(frags);
1809 zc->recv_skip_hint -= remaining;
1812 ret = vm_insert_page(vma, address + length,
1813 skb_frag_page(frags));
1816 length += PAGE_SIZE;
1818 zc->recv_skip_hint -= PAGE_SIZE;
1822 up_read(¤t->mm->mmap_sem);
1824 WRITE_ONCE(tp->copied_seq, seq);
1825 tcp_rcv_space_adjust(sk);
1827 /* Clean up data we have read: This will do ACK frames. */
1828 tcp_recv_skb(sk, seq, &offset);
1829 tcp_cleanup_rbuf(sk, length);
1831 if (length == zc->length)
1832 zc->recv_skip_hint = 0;
1834 if (!zc->recv_skip_hint && sock_flag(sk, SOCK_DONE))
1837 zc->length = length;
1842 static void tcp_update_recv_tstamps(struct sk_buff *skb,
1843 struct scm_timestamping_internal *tss)
1846 tss->ts[0] = ktime_to_timespec64(skb->tstamp);
1848 tss->ts[0] = (struct timespec64) {0};
1850 if (skb_hwtstamps(skb)->hwtstamp)
1851 tss->ts[2] = ktime_to_timespec64(skb_hwtstamps(skb)->hwtstamp);
1853 tss->ts[2] = (struct timespec64) {0};
1856 /* Similar to __sock_recv_timestamp, but does not require an skb */
1857 static void tcp_recv_timestamp(struct msghdr *msg, const struct sock *sk,
1858 struct scm_timestamping_internal *tss)
1860 int new_tstamp = sock_flag(sk, SOCK_TSTAMP_NEW);
1861 bool has_timestamping = false;
1863 if (tss->ts[0].tv_sec || tss->ts[0].tv_nsec) {
1864 if (sock_flag(sk, SOCK_RCVTSTAMP)) {
1865 if (sock_flag(sk, SOCK_RCVTSTAMPNS)) {
1867 struct __kernel_timespec kts = {tss->ts[0].tv_sec, tss->ts[0].tv_nsec};
1869 put_cmsg(msg, SOL_SOCKET, SO_TIMESTAMPNS_NEW,
1872 struct timespec ts_old = timespec64_to_timespec(tss->ts[0]);
1874 put_cmsg(msg, SOL_SOCKET, SO_TIMESTAMPNS_OLD,
1875 sizeof(ts_old), &ts_old);
1879 struct __kernel_sock_timeval stv;
1881 stv.tv_sec = tss->ts[0].tv_sec;
1882 stv.tv_usec = tss->ts[0].tv_nsec / 1000;
1883 put_cmsg(msg, SOL_SOCKET, SO_TIMESTAMP_NEW,
1886 struct __kernel_old_timeval tv;
1888 tv.tv_sec = tss->ts[0].tv_sec;
1889 tv.tv_usec = tss->ts[0].tv_nsec / 1000;
1890 put_cmsg(msg, SOL_SOCKET, SO_TIMESTAMP_OLD,
1896 if (sk->sk_tsflags & SOF_TIMESTAMPING_SOFTWARE)
1897 has_timestamping = true;
1899 tss->ts[0] = (struct timespec64) {0};
1902 if (tss->ts[2].tv_sec || tss->ts[2].tv_nsec) {
1903 if (sk->sk_tsflags & SOF_TIMESTAMPING_RAW_HARDWARE)
1904 has_timestamping = true;
1906 tss->ts[2] = (struct timespec64) {0};
1909 if (has_timestamping) {
1910 tss->ts[1] = (struct timespec64) {0};
1911 if (sock_flag(sk, SOCK_TSTAMP_NEW))
1912 put_cmsg_scm_timestamping64(msg, tss);
1914 put_cmsg_scm_timestamping(msg, tss);
1918 static int tcp_inq_hint(struct sock *sk)
1920 const struct tcp_sock *tp = tcp_sk(sk);
1921 u32 copied_seq = READ_ONCE(tp->copied_seq);
1922 u32 rcv_nxt = READ_ONCE(tp->rcv_nxt);
1925 inq = rcv_nxt - copied_seq;
1926 if (unlikely(inq < 0 || copied_seq != READ_ONCE(tp->copied_seq))) {
1928 inq = tp->rcv_nxt - tp->copied_seq;
1931 /* After receiving a FIN, tell the user-space to continue reading
1932 * by returning a non-zero inq.
1934 if (inq == 0 && sock_flag(sk, SOCK_DONE))
1940 * This routine copies from a sock struct into the user buffer.
1942 * Technical note: in 2.3 we work on _locked_ socket, so that
1943 * tricks with *seq access order and skb->users are not required.
1944 * Probably, code can be easily improved even more.
1947 int tcp_recvmsg(struct sock *sk, struct msghdr *msg, size_t len, int nonblock,
1948 int flags, int *addr_len)
1950 struct tcp_sock *tp = tcp_sk(sk);
1956 int target; /* Read at least this many bytes */
1958 struct sk_buff *skb, *last;
1960 struct scm_timestamping_internal tss;
1963 if (unlikely(flags & MSG_ERRQUEUE))
1964 return inet_recv_error(sk, msg, len, addr_len);
1966 if (sk_can_busy_loop(sk) && skb_queue_empty_lockless(&sk->sk_receive_queue) &&
1967 (sk->sk_state == TCP_ESTABLISHED))
1968 sk_busy_loop(sk, nonblock);
1973 if (sk->sk_state == TCP_LISTEN)
1976 cmsg_flags = tp->recvmsg_inq ? 1 : 0;
1977 timeo = sock_rcvtimeo(sk, nonblock);
1979 /* Urgent data needs to be handled specially. */
1980 if (flags & MSG_OOB)
1983 if (unlikely(tp->repair)) {
1985 if (!(flags & MSG_PEEK))
1988 if (tp->repair_queue == TCP_SEND_QUEUE)
1992 if (tp->repair_queue == TCP_NO_QUEUE)
1995 /* 'common' recv queue MSG_PEEK-ing */
1998 seq = &tp->copied_seq;
1999 if (flags & MSG_PEEK) {
2000 peek_seq = tp->copied_seq;
2004 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
2009 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
2010 if (tp->urg_data && tp->urg_seq == *seq) {
2013 if (signal_pending(current)) {
2014 copied = timeo ? sock_intr_errno(timeo) : -EAGAIN;
2019 /* Next get a buffer. */
2021 last = skb_peek_tail(&sk->sk_receive_queue);
2022 skb_queue_walk(&sk->sk_receive_queue, skb) {
2024 /* Now that we have two receive queues this
2027 if (WARN(before(*seq, TCP_SKB_CB(skb)->seq),
2028 "TCP recvmsg seq # bug: copied %X, seq %X, rcvnxt %X, fl %X\n",
2029 *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt,
2033 offset = *seq - TCP_SKB_CB(skb)->seq;
2034 if (unlikely(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
2035 pr_err_once("%s: found a SYN, please report !\n", __func__);
2038 if (offset < skb->len)
2040 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
2042 WARN(!(flags & MSG_PEEK),
2043 "TCP recvmsg seq # bug 2: copied %X, seq %X, rcvnxt %X, fl %X\n",
2044 *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt, flags);
2047 /* Well, if we have backlog, try to process it now yet. */
2049 if (copied >= target && !READ_ONCE(sk->sk_backlog.tail))
2054 sk->sk_state == TCP_CLOSE ||
2055 (sk->sk_shutdown & RCV_SHUTDOWN) ||
2057 signal_pending(current))
2060 if (sock_flag(sk, SOCK_DONE))
2064 copied = sock_error(sk);
2068 if (sk->sk_shutdown & RCV_SHUTDOWN)
2071 if (sk->sk_state == TCP_CLOSE) {
2072 /* This occurs when user tries to read
2073 * from never connected socket.
2084 if (signal_pending(current)) {
2085 copied = sock_intr_errno(timeo);
2090 tcp_cleanup_rbuf(sk, copied);
2092 if (copied >= target) {
2093 /* Do not sleep, just process backlog. */
2097 sk_wait_data(sk, &timeo, last);
2100 if ((flags & MSG_PEEK) &&
2101 (peek_seq - copied - urg_hole != tp->copied_seq)) {
2102 net_dbg_ratelimited("TCP(%s:%d): Application bug, race in MSG_PEEK\n",
2104 task_pid_nr(current));
2105 peek_seq = tp->copied_seq;
2110 /* Ok so how much can we use? */
2111 used = skb->len - offset;
2115 /* Do we have urgent data here? */
2117 u32 urg_offset = tp->urg_seq - *seq;
2118 if (urg_offset < used) {
2120 if (!sock_flag(sk, SOCK_URGINLINE)) {
2121 WRITE_ONCE(*seq, *seq + 1);
2133 if (!(flags & MSG_TRUNC)) {
2134 err = skb_copy_datagram_msg(skb, offset, msg, used);
2136 /* Exception. Bailout! */
2143 WRITE_ONCE(*seq, *seq + used);
2147 tcp_rcv_space_adjust(sk);
2150 if (tp->urg_data && after(tp->copied_seq, tp->urg_seq)) {
2152 tcp_fast_path_check(sk);
2154 if (used + offset < skb->len)
2157 if (TCP_SKB_CB(skb)->has_rxtstamp) {
2158 tcp_update_recv_tstamps(skb, &tss);
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 (cmsg_flags & 1) {
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);
2262 EXPORT_SYMBOL_GPL(tcp_set_state);
2265 * State processing on a close. This implements the state shift for
2266 * sending our FIN frame. Note that we only send a FIN for some
2267 * states. A shutdown() may have already sent the FIN, or we may be
2271 static const unsigned char new_state[16] = {
2272 /* current state: new state: action: */
2273 [0 /* (Invalid) */] = TCP_CLOSE,
2274 [TCP_ESTABLISHED] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2275 [TCP_SYN_SENT] = TCP_CLOSE,
2276 [TCP_SYN_RECV] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2277 [TCP_FIN_WAIT1] = TCP_FIN_WAIT1,
2278 [TCP_FIN_WAIT2] = TCP_FIN_WAIT2,
2279 [TCP_TIME_WAIT] = TCP_CLOSE,
2280 [TCP_CLOSE] = TCP_CLOSE,
2281 [TCP_CLOSE_WAIT] = TCP_LAST_ACK | TCP_ACTION_FIN,
2282 [TCP_LAST_ACK] = TCP_LAST_ACK,
2283 [TCP_LISTEN] = TCP_CLOSE,
2284 [TCP_CLOSING] = TCP_CLOSING,
2285 [TCP_NEW_SYN_RECV] = TCP_CLOSE, /* should not happen ! */
2288 static int tcp_close_state(struct sock *sk)
2290 int next = (int)new_state[sk->sk_state];
2291 int ns = next & TCP_STATE_MASK;
2293 tcp_set_state(sk, ns);
2295 return next & TCP_ACTION_FIN;
2299 * Shutdown the sending side of a connection. Much like close except
2300 * that we don't receive shut down or sock_set_flag(sk, SOCK_DEAD).
2303 void tcp_shutdown(struct sock *sk, int how)
2305 /* We need to grab some memory, and put together a FIN,
2306 * and then put it into the queue to be sent.
2307 * Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
2309 if (!(how & SEND_SHUTDOWN))
2312 /* If we've already sent a FIN, or it's a closed state, skip this. */
2313 if ((1 << sk->sk_state) &
2314 (TCPF_ESTABLISHED | TCPF_SYN_SENT |
2315 TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) {
2316 /* Clear out any half completed packets. FIN if needed. */
2317 if (tcp_close_state(sk))
2321 EXPORT_SYMBOL(tcp_shutdown);
2323 bool tcp_check_oom(struct sock *sk, int shift)
2325 bool too_many_orphans, out_of_socket_memory;
2327 too_many_orphans = tcp_too_many_orphans(sk, shift);
2328 out_of_socket_memory = tcp_out_of_memory(sk);
2330 if (too_many_orphans)
2331 net_info_ratelimited("too many orphaned sockets\n");
2332 if (out_of_socket_memory)
2333 net_info_ratelimited("out of memory -- consider tuning tcp_mem\n");
2334 return too_many_orphans || out_of_socket_memory;
2337 void tcp_close(struct sock *sk, long timeout)
2339 struct sk_buff *skb;
2340 int data_was_unread = 0;
2344 sk->sk_shutdown = SHUTDOWN_MASK;
2346 if (sk->sk_state == TCP_LISTEN) {
2347 tcp_set_state(sk, TCP_CLOSE);
2350 inet_csk_listen_stop(sk);
2352 goto adjudge_to_death;
2355 /* We need to flush the recv. buffs. We do this only on the
2356 * descriptor close, not protocol-sourced closes, because the
2357 * reader process may not have drained the data yet!
2359 while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) {
2360 u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq;
2362 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
2364 data_was_unread += len;
2370 /* If socket has been already reset (e.g. in tcp_reset()) - kill it. */
2371 if (sk->sk_state == TCP_CLOSE)
2372 goto adjudge_to_death;
2374 /* As outlined in RFC 2525, section 2.17, we send a RST here because
2375 * data was lost. To witness the awful effects of the old behavior of
2376 * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk
2377 * GET in an FTP client, suspend the process, wait for the client to
2378 * advertise a zero window, then kill -9 the FTP client, wheee...
2379 * Note: timeout is always zero in such a case.
2381 if (unlikely(tcp_sk(sk)->repair)) {
2382 sk->sk_prot->disconnect(sk, 0);
2383 } else if (data_was_unread) {
2384 /* Unread data was tossed, zap the connection. */
2385 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONCLOSE);
2386 tcp_set_state(sk, TCP_CLOSE);
2387 tcp_send_active_reset(sk, sk->sk_allocation);
2388 } else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) {
2389 /* Check zero linger _after_ checking for unread data. */
2390 sk->sk_prot->disconnect(sk, 0);
2391 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
2392 } else if (tcp_close_state(sk)) {
2393 /* We FIN if the application ate all the data before
2394 * zapping the connection.
2397 /* RED-PEN. Formally speaking, we have broken TCP state
2398 * machine. State transitions:
2400 * TCP_ESTABLISHED -> TCP_FIN_WAIT1
2401 * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible)
2402 * TCP_CLOSE_WAIT -> TCP_LAST_ACK
2404 * are legal only when FIN has been sent (i.e. in window),
2405 * rather than queued out of window. Purists blame.
2407 * F.e. "RFC state" is ESTABLISHED,
2408 * if Linux state is FIN-WAIT-1, but FIN is still not sent.
2410 * The visible declinations are that sometimes
2411 * we enter time-wait state, when it is not required really
2412 * (harmless), do not send active resets, when they are
2413 * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
2414 * they look as CLOSING or LAST_ACK for Linux)
2415 * Probably, I missed some more holelets.
2417 * XXX (TFO) - To start off we don't support SYN+ACK+FIN
2418 * in a single packet! (May consider it later but will
2419 * probably need API support or TCP_CORK SYN-ACK until
2420 * data is written and socket is closed.)
2425 sk_stream_wait_close(sk, timeout);
2428 state = sk->sk_state;
2434 /* remove backlog if any, without releasing ownership. */
2437 percpu_counter_inc(sk->sk_prot->orphan_count);
2439 /* Have we already been destroyed by a softirq or backlog? */
2440 if (state != TCP_CLOSE && sk->sk_state == TCP_CLOSE)
2443 /* This is a (useful) BSD violating of the RFC. There is a
2444 * problem with TCP as specified in that the other end could
2445 * keep a socket open forever with no application left this end.
2446 * We use a 1 minute timeout (about the same as BSD) then kill
2447 * our end. If they send after that then tough - BUT: long enough
2448 * that we won't make the old 4*rto = almost no time - whoops
2451 * Nope, it was not mistake. It is really desired behaviour
2452 * f.e. on http servers, when such sockets are useless, but
2453 * consume significant resources. Let's do it with special
2454 * linger2 option. --ANK
2457 if (sk->sk_state == TCP_FIN_WAIT2) {
2458 struct tcp_sock *tp = tcp_sk(sk);
2459 if (tp->linger2 < 0) {
2460 tcp_set_state(sk, TCP_CLOSE);
2461 tcp_send_active_reset(sk, GFP_ATOMIC);
2462 __NET_INC_STATS(sock_net(sk),
2463 LINUX_MIB_TCPABORTONLINGER);
2465 const int tmo = tcp_fin_time(sk);
2467 if (tmo > TCP_TIMEWAIT_LEN) {
2468 inet_csk_reset_keepalive_timer(sk,
2469 tmo - TCP_TIMEWAIT_LEN);
2471 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
2476 if (sk->sk_state != TCP_CLOSE) {
2478 if (tcp_check_oom(sk, 0)) {
2479 tcp_set_state(sk, TCP_CLOSE);
2480 tcp_send_active_reset(sk, GFP_ATOMIC);
2481 __NET_INC_STATS(sock_net(sk),
2482 LINUX_MIB_TCPABORTONMEMORY);
2483 } else if (!check_net(sock_net(sk))) {
2484 /* Not possible to send reset; just close */
2485 tcp_set_state(sk, TCP_CLOSE);
2489 if (sk->sk_state == TCP_CLOSE) {
2490 struct request_sock *req;
2492 req = rcu_dereference_protected(tcp_sk(sk)->fastopen_rsk,
2493 lockdep_sock_is_held(sk));
2494 /* We could get here with a non-NULL req if the socket is
2495 * aborted (e.g., closed with unread data) before 3WHS
2499 reqsk_fastopen_remove(sk, req, false);
2500 inet_csk_destroy_sock(sk);
2502 /* Otherwise, socket is reprieved until protocol close. */
2510 EXPORT_SYMBOL(tcp_close);
2512 /* These states need RST on ABORT according to RFC793 */
2514 static inline bool tcp_need_reset(int state)
2516 return (1 << state) &
2517 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 |
2518 TCPF_FIN_WAIT2 | TCPF_SYN_RECV);
2521 static void tcp_rtx_queue_purge(struct sock *sk)
2523 struct rb_node *p = rb_first(&sk->tcp_rtx_queue);
2526 struct sk_buff *skb = rb_to_skb(p);
2529 /* Since we are deleting whole queue, no need to
2530 * list_del(&skb->tcp_tsorted_anchor)
2532 tcp_rtx_queue_unlink(skb, sk);
2533 sk_wmem_free_skb(sk, skb);
2537 void tcp_write_queue_purge(struct sock *sk)
2539 struct sk_buff *skb;
2541 tcp_chrono_stop(sk, TCP_CHRONO_BUSY);
2542 while ((skb = __skb_dequeue(&sk->sk_write_queue)) != NULL) {
2543 tcp_skb_tsorted_anchor_cleanup(skb);
2544 sk_wmem_free_skb(sk, skb);
2546 tcp_rtx_queue_purge(sk);
2547 skb = sk->sk_tx_skb_cache;
2550 sk->sk_tx_skb_cache = NULL;
2552 INIT_LIST_HEAD(&tcp_sk(sk)->tsorted_sent_queue);
2554 tcp_clear_all_retrans_hints(tcp_sk(sk));
2555 tcp_sk(sk)->packets_out = 0;
2556 inet_csk(sk)->icsk_backoff = 0;
2559 int tcp_disconnect(struct sock *sk, int flags)
2561 struct inet_sock *inet = inet_sk(sk);
2562 struct inet_connection_sock *icsk = inet_csk(sk);
2563 struct tcp_sock *tp = tcp_sk(sk);
2564 int old_state = sk->sk_state;
2567 if (old_state != TCP_CLOSE)
2568 tcp_set_state(sk, TCP_CLOSE);
2570 /* ABORT function of RFC793 */
2571 if (old_state == TCP_LISTEN) {
2572 inet_csk_listen_stop(sk);
2573 } else if (unlikely(tp->repair)) {
2574 sk->sk_err = ECONNABORTED;
2575 } else if (tcp_need_reset(old_state) ||
2576 (tp->snd_nxt != tp->write_seq &&
2577 (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK))) {
2578 /* The last check adjusts for discrepancy of Linux wrt. RFC
2581 tcp_send_active_reset(sk, gfp_any());
2582 sk->sk_err = ECONNRESET;
2583 } else if (old_state == TCP_SYN_SENT)
2584 sk->sk_err = ECONNRESET;
2586 tcp_clear_xmit_timers(sk);
2587 __skb_queue_purge(&sk->sk_receive_queue);
2588 if (sk->sk_rx_skb_cache) {
2589 __kfree_skb(sk->sk_rx_skb_cache);
2590 sk->sk_rx_skb_cache = NULL;
2592 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt);
2594 tcp_write_queue_purge(sk);
2595 tcp_fastopen_active_disable_ofo_check(sk);
2596 skb_rbtree_purge(&tp->out_of_order_queue);
2598 inet->inet_dport = 0;
2600 if (!(sk->sk_userlocks & SOCK_BINDADDR_LOCK))
2601 inet_reset_saddr(sk);
2603 sk->sk_shutdown = 0;
2604 sock_reset_flag(sk, SOCK_DONE);
2606 tp->mdev_us = jiffies_to_usecs(TCP_TIMEOUT_INIT);
2607 tp->rcv_rtt_last_tsecr = 0;
2609 seq = tp->write_seq + tp->max_window + 2;
2612 WRITE_ONCE(tp->write_seq, seq);
2614 icsk->icsk_backoff = 0;
2616 icsk->icsk_probes_out = 0;
2617 icsk->icsk_rto = TCP_TIMEOUT_INIT;
2618 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
2619 tp->snd_cwnd = TCP_INIT_CWND;
2620 tp->snd_cwnd_cnt = 0;
2621 tp->window_clamp = 0;
2622 tp->delivered_ce = 0;
2623 tcp_set_ca_state(sk, TCP_CA_Open);
2624 tp->is_sack_reneg = 0;
2625 tcp_clear_retrans(tp);
2626 inet_csk_delack_init(sk);
2627 /* Initialize rcv_mss to TCP_MIN_MSS to avoid division by 0
2628 * issue in __tcp_select_window()
2630 icsk->icsk_ack.rcv_mss = TCP_MIN_MSS;
2631 memset(&tp->rx_opt, 0, sizeof(tp->rx_opt));
2633 dst_release(sk->sk_rx_dst);
2634 sk->sk_rx_dst = NULL;
2635 tcp_saved_syn_free(tp);
2636 tp->compressed_ack = 0;
2638 tp->bytes_acked = 0;
2639 tp->bytes_received = 0;
2640 tp->bytes_retrans = 0;
2641 tp->duplicate_sack[0].start_seq = 0;
2642 tp->duplicate_sack[0].end_seq = 0;
2645 tp->retrans_out = 0;
2647 tp->tlp_high_seq = 0;
2648 tp->last_oow_ack_time = 0;
2649 /* There's a bubble in the pipe until at least the first ACK. */
2650 tp->app_limited = ~0U;
2651 tp->rack.mstamp = 0;
2652 tp->rack.advanced = 0;
2653 tp->rack.reo_wnd_steps = 1;
2654 tp->rack.last_delivered = 0;
2655 tp->rack.reo_wnd_persist = 0;
2656 tp->rack.dsack_seen = 0;
2657 tp->syn_data_acked = 0;
2658 tp->rx_opt.saw_tstamp = 0;
2659 tp->rx_opt.dsack = 0;
2660 tp->rx_opt.num_sacks = 0;
2661 tp->rcv_ooopack = 0;
2664 /* Clean up fastopen related fields */
2665 tcp_free_fastopen_req(tp);
2666 inet->defer_connect = 0;
2667 tp->fastopen_client_fail = 0;
2669 WARN_ON(inet->inet_num && !icsk->icsk_bind_hash);
2671 if (sk->sk_frag.page) {
2672 put_page(sk->sk_frag.page);
2673 sk->sk_frag.page = NULL;
2674 sk->sk_frag.offset = 0;
2677 sk->sk_error_report(sk);
2680 EXPORT_SYMBOL(tcp_disconnect);
2682 static inline bool tcp_can_repair_sock(const struct sock *sk)
2684 return ns_capable(sock_net(sk)->user_ns, CAP_NET_ADMIN) &&
2685 (sk->sk_state != TCP_LISTEN);
2688 static int tcp_repair_set_window(struct tcp_sock *tp, char __user *optbuf, int len)
2690 struct tcp_repair_window opt;
2695 if (len != sizeof(opt))
2698 if (copy_from_user(&opt, optbuf, sizeof(opt)))
2701 if (opt.max_window < opt.snd_wnd)
2704 if (after(opt.snd_wl1, tp->rcv_nxt + opt.rcv_wnd))
2707 if (after(opt.rcv_wup, tp->rcv_nxt))
2710 tp->snd_wl1 = opt.snd_wl1;
2711 tp->snd_wnd = opt.snd_wnd;
2712 tp->max_window = opt.max_window;
2714 tp->rcv_wnd = opt.rcv_wnd;
2715 tp->rcv_wup = opt.rcv_wup;
2720 static int tcp_repair_options_est(struct sock *sk,
2721 struct tcp_repair_opt __user *optbuf, unsigned int len)
2723 struct tcp_sock *tp = tcp_sk(sk);
2724 struct tcp_repair_opt opt;
2726 while (len >= sizeof(opt)) {
2727 if (copy_from_user(&opt, optbuf, sizeof(opt)))
2733 switch (opt.opt_code) {
2735 tp->rx_opt.mss_clamp = opt.opt_val;
2740 u16 snd_wscale = opt.opt_val & 0xFFFF;
2741 u16 rcv_wscale = opt.opt_val >> 16;
2743 if (snd_wscale > TCP_MAX_WSCALE || rcv_wscale > TCP_MAX_WSCALE)
2746 tp->rx_opt.snd_wscale = snd_wscale;
2747 tp->rx_opt.rcv_wscale = rcv_wscale;
2748 tp->rx_opt.wscale_ok = 1;
2751 case TCPOPT_SACK_PERM:
2752 if (opt.opt_val != 0)
2755 tp->rx_opt.sack_ok |= TCP_SACK_SEEN;
2757 case TCPOPT_TIMESTAMP:
2758 if (opt.opt_val != 0)
2761 tp->rx_opt.tstamp_ok = 1;
2769 DEFINE_STATIC_KEY_FALSE(tcp_tx_delay_enabled);
2770 EXPORT_SYMBOL(tcp_tx_delay_enabled);
2772 static void tcp_enable_tx_delay(void)
2774 if (!static_branch_unlikely(&tcp_tx_delay_enabled)) {
2775 static int __tcp_tx_delay_enabled = 0;
2777 if (cmpxchg(&__tcp_tx_delay_enabled, 0, 1) == 0) {
2778 static_branch_enable(&tcp_tx_delay_enabled);
2779 pr_info("TCP_TX_DELAY enabled\n");
2785 * Socket option code for TCP.
2787 static int do_tcp_setsockopt(struct sock *sk, int level,
2788 int optname, char __user *optval, unsigned int optlen)
2790 struct tcp_sock *tp = tcp_sk(sk);
2791 struct inet_connection_sock *icsk = inet_csk(sk);
2792 struct net *net = sock_net(sk);
2796 /* These are data/string values, all the others are ints */
2798 case TCP_CONGESTION: {
2799 char name[TCP_CA_NAME_MAX];
2804 val = strncpy_from_user(name, optval,
2805 min_t(long, TCP_CA_NAME_MAX-1, optlen));
2811 err = tcp_set_congestion_control(sk, name, true, true,
2812 ns_capable(sock_net(sk)->user_ns,
2818 char name[TCP_ULP_NAME_MAX];
2823 val = strncpy_from_user(name, optval,
2824 min_t(long, TCP_ULP_NAME_MAX - 1,
2831 err = tcp_set_ulp(sk, name);
2835 case TCP_FASTOPEN_KEY: {
2836 __u8 key[TCP_FASTOPEN_KEY_BUF_LENGTH];
2837 __u8 *backup_key = NULL;
2839 /* Allow a backup key as well to facilitate key rotation
2840 * First key is the active one.
2842 if (optlen != TCP_FASTOPEN_KEY_LENGTH &&
2843 optlen != TCP_FASTOPEN_KEY_BUF_LENGTH)
2846 if (copy_from_user(key, optval, optlen))
2849 if (optlen == TCP_FASTOPEN_KEY_BUF_LENGTH)
2850 backup_key = key + TCP_FASTOPEN_KEY_LENGTH;
2852 return tcp_fastopen_reset_cipher(net, sk, key, backup_key);
2859 if (optlen < sizeof(int))
2862 if (get_user(val, (int __user *)optval))
2869 /* Values greater than interface MTU won't take effect. However
2870 * at the point when this call is done we typically don't yet
2871 * know which interface is going to be used
2873 if (val && (val < TCP_MIN_MSS || val > MAX_TCP_WINDOW)) {
2877 tp->rx_opt.user_mss = val;
2882 /* TCP_NODELAY is weaker than TCP_CORK, so that
2883 * this option on corked socket is remembered, but
2884 * it is not activated until cork is cleared.
2886 * However, when TCP_NODELAY is set we make
2887 * an explicit push, which overrides even TCP_CORK
2888 * for currently queued segments.
2890 tp->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH;
2891 tcp_push_pending_frames(sk);
2893 tp->nonagle &= ~TCP_NAGLE_OFF;
2897 case TCP_THIN_LINEAR_TIMEOUTS:
2898 if (val < 0 || val > 1)
2904 case TCP_THIN_DUPACK:
2905 if (val < 0 || val > 1)
2910 if (!tcp_can_repair_sock(sk))
2912 else if (val == TCP_REPAIR_ON) {
2914 sk->sk_reuse = SK_FORCE_REUSE;
2915 tp->repair_queue = TCP_NO_QUEUE;
2916 } else if (val == TCP_REPAIR_OFF) {
2918 sk->sk_reuse = SK_NO_REUSE;
2919 tcp_send_window_probe(sk);
2920 } else if (val == TCP_REPAIR_OFF_NO_WP) {
2922 sk->sk_reuse = SK_NO_REUSE;
2928 case TCP_REPAIR_QUEUE:
2931 else if ((unsigned int)val < TCP_QUEUES_NR)
2932 tp->repair_queue = val;
2938 if (sk->sk_state != TCP_CLOSE)
2940 else if (tp->repair_queue == TCP_SEND_QUEUE)
2941 WRITE_ONCE(tp->write_seq, val);
2942 else if (tp->repair_queue == TCP_RECV_QUEUE)
2943 WRITE_ONCE(tp->rcv_nxt, val);
2948 case TCP_REPAIR_OPTIONS:
2951 else if (sk->sk_state == TCP_ESTABLISHED)
2952 err = tcp_repair_options_est(sk,
2953 (struct tcp_repair_opt __user *)optval,
2960 /* When set indicates to always queue non-full frames.
2961 * Later the user clears this option and we transmit
2962 * any pending partial frames in the queue. This is
2963 * meant to be used alongside sendfile() to get properly
2964 * filled frames when the user (for example) must write
2965 * out headers with a write() call first and then use
2966 * sendfile to send out the data parts.
2968 * TCP_CORK can be set together with TCP_NODELAY and it is
2969 * stronger than TCP_NODELAY.
2972 tp->nonagle |= TCP_NAGLE_CORK;
2974 tp->nonagle &= ~TCP_NAGLE_CORK;
2975 if (tp->nonagle&TCP_NAGLE_OFF)
2976 tp->nonagle |= TCP_NAGLE_PUSH;
2977 tcp_push_pending_frames(sk);
2982 if (val < 1 || val > MAX_TCP_KEEPIDLE)
2985 tp->keepalive_time = val * HZ;
2986 if (sock_flag(sk, SOCK_KEEPOPEN) &&
2987 !((1 << sk->sk_state) &
2988 (TCPF_CLOSE | TCPF_LISTEN))) {
2989 u32 elapsed = keepalive_time_elapsed(tp);
2990 if (tp->keepalive_time > elapsed)
2991 elapsed = tp->keepalive_time - elapsed;
2994 inet_csk_reset_keepalive_timer(sk, elapsed);
2999 if (val < 1 || val > MAX_TCP_KEEPINTVL)
3002 tp->keepalive_intvl = val * HZ;
3005 if (val < 1 || val > MAX_TCP_KEEPCNT)
3008 tp->keepalive_probes = val;
3011 if (val < 1 || val > MAX_TCP_SYNCNT)
3014 icsk->icsk_syn_retries = val;
3018 if (val < 0 || val > 1)
3027 else if (val > net->ipv4.sysctl_tcp_fin_timeout / HZ)
3030 tp->linger2 = val * HZ;
3033 case TCP_DEFER_ACCEPT:
3034 /* Translate value in seconds to number of retransmits */
3035 icsk->icsk_accept_queue.rskq_defer_accept =
3036 secs_to_retrans(val, TCP_TIMEOUT_INIT / HZ,
3040 case TCP_WINDOW_CLAMP:
3042 if (sk->sk_state != TCP_CLOSE) {
3046 tp->window_clamp = 0;
3048 tp->window_clamp = val < SOCK_MIN_RCVBUF / 2 ?
3049 SOCK_MIN_RCVBUF / 2 : val;
3054 inet_csk_enter_pingpong_mode(sk);
3056 inet_csk_exit_pingpong_mode(sk);
3057 if ((1 << sk->sk_state) &
3058 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) &&
3059 inet_csk_ack_scheduled(sk)) {
3060 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
3061 tcp_cleanup_rbuf(sk, 1);
3063 inet_csk_enter_pingpong_mode(sk);
3068 #ifdef CONFIG_TCP_MD5SIG
3070 case TCP_MD5SIG_EXT:
3071 if ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN))
3072 err = tp->af_specific->md5_parse(sk, optname, optval, optlen);
3077 case TCP_USER_TIMEOUT:
3078 /* Cap the max time in ms TCP will retry or probe the window
3079 * before giving up and aborting (ETIMEDOUT) a connection.
3084 icsk->icsk_user_timeout = val;
3088 if (val >= 0 && ((1 << sk->sk_state) & (TCPF_CLOSE |
3090 tcp_fastopen_init_key_once(net);
3092 fastopen_queue_tune(sk, val);
3097 case TCP_FASTOPEN_CONNECT:
3098 if (val > 1 || val < 0) {
3100 } else if (net->ipv4.sysctl_tcp_fastopen & TFO_CLIENT_ENABLE) {
3101 if (sk->sk_state == TCP_CLOSE)
3102 tp->fastopen_connect = val;
3109 case TCP_FASTOPEN_NO_COOKIE:
3110 if (val > 1 || val < 0)
3112 else if (!((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
3115 tp->fastopen_no_cookie = val;
3121 tp->tsoffset = val - tcp_time_stamp_raw();
3123 case TCP_REPAIR_WINDOW:
3124 err = tcp_repair_set_window(tp, optval, optlen);
3126 case TCP_NOTSENT_LOWAT:
3127 tp->notsent_lowat = val;
3128 sk->sk_write_space(sk);
3131 if (val > 1 || val < 0)
3134 tp->recvmsg_inq = val;
3138 tcp_enable_tx_delay();
3139 tp->tcp_tx_delay = val;
3150 int tcp_setsockopt(struct sock *sk, int level, int optname, char __user *optval,
3151 unsigned int optlen)
3153 const struct inet_connection_sock *icsk = inet_csk(sk);
3155 if (level != SOL_TCP)
3156 return icsk->icsk_af_ops->setsockopt(sk, level, optname,
3158 return do_tcp_setsockopt(sk, level, optname, optval, optlen);
3160 EXPORT_SYMBOL(tcp_setsockopt);
3162 #ifdef CONFIG_COMPAT
3163 int compat_tcp_setsockopt(struct sock *sk, int level, int optname,
3164 char __user *optval, unsigned int optlen)
3166 if (level != SOL_TCP)
3167 return inet_csk_compat_setsockopt(sk, level, optname,
3169 return do_tcp_setsockopt(sk, level, optname, optval, optlen);
3171 EXPORT_SYMBOL(compat_tcp_setsockopt);
3174 static void tcp_get_info_chrono_stats(const struct tcp_sock *tp,
3175 struct tcp_info *info)
3177 u64 stats[__TCP_CHRONO_MAX], total = 0;
3180 for (i = TCP_CHRONO_BUSY; i < __TCP_CHRONO_MAX; ++i) {
3181 stats[i] = tp->chrono_stat[i - 1];
3182 if (i == tp->chrono_type)
3183 stats[i] += tcp_jiffies32 - tp->chrono_start;
3184 stats[i] *= USEC_PER_SEC / HZ;
3188 info->tcpi_busy_time = total;
3189 info->tcpi_rwnd_limited = stats[TCP_CHRONO_RWND_LIMITED];
3190 info->tcpi_sndbuf_limited = stats[TCP_CHRONO_SNDBUF_LIMITED];
3193 /* Return information about state of tcp endpoint in API format. */
3194 void tcp_get_info(struct sock *sk, struct tcp_info *info)
3196 const struct tcp_sock *tp = tcp_sk(sk); /* iff sk_type == SOCK_STREAM */
3197 const struct inet_connection_sock *icsk = inet_csk(sk);
3203 memset(info, 0, sizeof(*info));
3204 if (sk->sk_type != SOCK_STREAM)
3207 info->tcpi_state = inet_sk_state_load(sk);
3209 /* Report meaningful fields for all TCP states, including listeners */
3210 rate = READ_ONCE(sk->sk_pacing_rate);
3211 rate64 = (rate != ~0UL) ? rate : ~0ULL;
3212 info->tcpi_pacing_rate = rate64;
3214 rate = READ_ONCE(sk->sk_max_pacing_rate);
3215 rate64 = (rate != ~0UL) ? rate : ~0ULL;
3216 info->tcpi_max_pacing_rate = rate64;
3218 info->tcpi_reordering = tp->reordering;
3219 info->tcpi_snd_cwnd = tp->snd_cwnd;
3221 if (info->tcpi_state == TCP_LISTEN) {
3222 /* listeners aliased fields :
3223 * tcpi_unacked -> Number of children ready for accept()
3224 * tcpi_sacked -> max backlog
3226 info->tcpi_unacked = READ_ONCE(sk->sk_ack_backlog);
3227 info->tcpi_sacked = READ_ONCE(sk->sk_max_ack_backlog);
3231 slow = lock_sock_fast(sk);
3233 info->tcpi_ca_state = icsk->icsk_ca_state;
3234 info->tcpi_retransmits = icsk->icsk_retransmits;
3235 info->tcpi_probes = icsk->icsk_probes_out;
3236 info->tcpi_backoff = icsk->icsk_backoff;
3238 if (tp->rx_opt.tstamp_ok)
3239 info->tcpi_options |= TCPI_OPT_TIMESTAMPS;
3240 if (tcp_is_sack(tp))
3241 info->tcpi_options |= TCPI_OPT_SACK;
3242 if (tp->rx_opt.wscale_ok) {
3243 info->tcpi_options |= TCPI_OPT_WSCALE;
3244 info->tcpi_snd_wscale = tp->rx_opt.snd_wscale;
3245 info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale;
3248 if (tp->ecn_flags & TCP_ECN_OK)
3249 info->tcpi_options |= TCPI_OPT_ECN;
3250 if (tp->ecn_flags & TCP_ECN_SEEN)
3251 info->tcpi_options |= TCPI_OPT_ECN_SEEN;
3252 if (tp->syn_data_acked)
3253 info->tcpi_options |= TCPI_OPT_SYN_DATA;
3255 info->tcpi_rto = jiffies_to_usecs(icsk->icsk_rto);
3256 info->tcpi_ato = jiffies_to_usecs(icsk->icsk_ack.ato);
3257 info->tcpi_snd_mss = tp->mss_cache;
3258 info->tcpi_rcv_mss = icsk->icsk_ack.rcv_mss;
3260 info->tcpi_unacked = tp->packets_out;
3261 info->tcpi_sacked = tp->sacked_out;
3263 info->tcpi_lost = tp->lost_out;
3264 info->tcpi_retrans = tp->retrans_out;
3266 now = tcp_jiffies32;
3267 info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime);
3268 info->tcpi_last_data_recv = jiffies_to_msecs(now - icsk->icsk_ack.lrcvtime);
3269 info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp);
3271 info->tcpi_pmtu = icsk->icsk_pmtu_cookie;
3272 info->tcpi_rcv_ssthresh = tp->rcv_ssthresh;
3273 info->tcpi_rtt = tp->srtt_us >> 3;
3274 info->tcpi_rttvar = tp->mdev_us >> 2;
3275 info->tcpi_snd_ssthresh = tp->snd_ssthresh;
3276 info->tcpi_advmss = tp->advmss;
3278 info->tcpi_rcv_rtt = tp->rcv_rtt_est.rtt_us >> 3;
3279 info->tcpi_rcv_space = tp->rcvq_space.space;
3281 info->tcpi_total_retrans = tp->total_retrans;
3283 info->tcpi_bytes_acked = tp->bytes_acked;
3284 info->tcpi_bytes_received = tp->bytes_received;
3285 info->tcpi_notsent_bytes = max_t(int, 0, tp->write_seq - tp->snd_nxt);
3286 tcp_get_info_chrono_stats(tp, info);
3288 info->tcpi_segs_out = tp->segs_out;
3289 info->tcpi_segs_in = tp->segs_in;
3291 info->tcpi_min_rtt = tcp_min_rtt(tp);
3292 info->tcpi_data_segs_in = tp->data_segs_in;
3293 info->tcpi_data_segs_out = tp->data_segs_out;
3295 info->tcpi_delivery_rate_app_limited = tp->rate_app_limited ? 1 : 0;
3296 rate64 = tcp_compute_delivery_rate(tp);
3298 info->tcpi_delivery_rate = rate64;
3299 info->tcpi_delivered = tp->delivered;
3300 info->tcpi_delivered_ce = tp->delivered_ce;
3301 info->tcpi_bytes_sent = tp->bytes_sent;
3302 info->tcpi_bytes_retrans = tp->bytes_retrans;
3303 info->tcpi_dsack_dups = tp->dsack_dups;
3304 info->tcpi_reord_seen = tp->reord_seen;
3305 info->tcpi_rcv_ooopack = tp->rcv_ooopack;
3306 info->tcpi_snd_wnd = tp->snd_wnd;
3307 info->tcpi_fastopen_client_fail = tp->fastopen_client_fail;
3308 unlock_sock_fast(sk, slow);
3310 EXPORT_SYMBOL_GPL(tcp_get_info);
3312 static size_t tcp_opt_stats_get_size(void)
3315 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_BUSY */
3316 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_RWND_LIMITED */
3317 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_SNDBUF_LIMITED */
3318 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_DATA_SEGS_OUT */
3319 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_TOTAL_RETRANS */
3320 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_PACING_RATE */
3321 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_DELIVERY_RATE */
3322 nla_total_size(sizeof(u32)) + /* TCP_NLA_SND_CWND */
3323 nla_total_size(sizeof(u32)) + /* TCP_NLA_REORDERING */
3324 nla_total_size(sizeof(u32)) + /* TCP_NLA_MIN_RTT */
3325 nla_total_size(sizeof(u8)) + /* TCP_NLA_RECUR_RETRANS */
3326 nla_total_size(sizeof(u8)) + /* TCP_NLA_DELIVERY_RATE_APP_LMT */
3327 nla_total_size(sizeof(u32)) + /* TCP_NLA_SNDQ_SIZE */
3328 nla_total_size(sizeof(u8)) + /* TCP_NLA_CA_STATE */
3329 nla_total_size(sizeof(u32)) + /* TCP_NLA_SND_SSTHRESH */
3330 nla_total_size(sizeof(u32)) + /* TCP_NLA_DELIVERED */
3331 nla_total_size(sizeof(u32)) + /* TCP_NLA_DELIVERED_CE */
3332 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_BYTES_SENT */
3333 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_BYTES_RETRANS */
3334 nla_total_size(sizeof(u32)) + /* TCP_NLA_DSACK_DUPS */
3335 nla_total_size(sizeof(u32)) + /* TCP_NLA_REORD_SEEN */
3336 nla_total_size(sizeof(u32)) + /* TCP_NLA_SRTT */
3340 struct sk_buff *tcp_get_timestamping_opt_stats(const struct sock *sk)
3342 const struct tcp_sock *tp = tcp_sk(sk);
3343 struct sk_buff *stats;
3344 struct tcp_info info;
3348 stats = alloc_skb(tcp_opt_stats_get_size(), GFP_ATOMIC);
3352 tcp_get_info_chrono_stats(tp, &info);
3353 nla_put_u64_64bit(stats, TCP_NLA_BUSY,
3354 info.tcpi_busy_time, TCP_NLA_PAD);
3355 nla_put_u64_64bit(stats, TCP_NLA_RWND_LIMITED,
3356 info.tcpi_rwnd_limited, TCP_NLA_PAD);
3357 nla_put_u64_64bit(stats, TCP_NLA_SNDBUF_LIMITED,
3358 info.tcpi_sndbuf_limited, TCP_NLA_PAD);
3359 nla_put_u64_64bit(stats, TCP_NLA_DATA_SEGS_OUT,
3360 tp->data_segs_out, TCP_NLA_PAD);
3361 nla_put_u64_64bit(stats, TCP_NLA_TOTAL_RETRANS,
3362 tp->total_retrans, TCP_NLA_PAD);
3364 rate = READ_ONCE(sk->sk_pacing_rate);
3365 rate64 = (rate != ~0UL) ? rate : ~0ULL;
3366 nla_put_u64_64bit(stats, TCP_NLA_PACING_RATE, rate64, TCP_NLA_PAD);
3368 rate64 = tcp_compute_delivery_rate(tp);
3369 nla_put_u64_64bit(stats, TCP_NLA_DELIVERY_RATE, rate64, TCP_NLA_PAD);
3371 nla_put_u32(stats, TCP_NLA_SND_CWND, tp->snd_cwnd);
3372 nla_put_u32(stats, TCP_NLA_REORDERING, tp->reordering);
3373 nla_put_u32(stats, TCP_NLA_MIN_RTT, tcp_min_rtt(tp));
3375 nla_put_u8(stats, TCP_NLA_RECUR_RETRANS, inet_csk(sk)->icsk_retransmits);
3376 nla_put_u8(stats, TCP_NLA_DELIVERY_RATE_APP_LMT, !!tp->rate_app_limited);
3377 nla_put_u32(stats, TCP_NLA_SND_SSTHRESH, tp->snd_ssthresh);
3378 nla_put_u32(stats, TCP_NLA_DELIVERED, tp->delivered);
3379 nla_put_u32(stats, TCP_NLA_DELIVERED_CE, tp->delivered_ce);
3381 nla_put_u32(stats, TCP_NLA_SNDQ_SIZE, tp->write_seq - tp->snd_una);
3382 nla_put_u8(stats, TCP_NLA_CA_STATE, inet_csk(sk)->icsk_ca_state);
3384 nla_put_u64_64bit(stats, TCP_NLA_BYTES_SENT, tp->bytes_sent,
3386 nla_put_u64_64bit(stats, TCP_NLA_BYTES_RETRANS, tp->bytes_retrans,
3388 nla_put_u32(stats, TCP_NLA_DSACK_DUPS, tp->dsack_dups);
3389 nla_put_u32(stats, TCP_NLA_REORD_SEEN, tp->reord_seen);
3390 nla_put_u32(stats, TCP_NLA_SRTT, tp->srtt_us >> 3);
3395 static int do_tcp_getsockopt(struct sock *sk, int level,
3396 int optname, char __user *optval, int __user *optlen)
3398 struct inet_connection_sock *icsk = inet_csk(sk);
3399 struct tcp_sock *tp = tcp_sk(sk);
3400 struct net *net = sock_net(sk);
3403 if (get_user(len, optlen))
3406 len = min_t(unsigned int, len, sizeof(int));
3413 val = tp->mss_cache;
3414 if (!val && ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
3415 val = tp->rx_opt.user_mss;
3417 val = tp->rx_opt.mss_clamp;
3420 val = !!(tp->nonagle&TCP_NAGLE_OFF);
3423 val = !!(tp->nonagle&TCP_NAGLE_CORK);
3426 val = keepalive_time_when(tp) / HZ;
3429 val = keepalive_intvl_when(tp) / HZ;
3432 val = keepalive_probes(tp);
3435 val = icsk->icsk_syn_retries ? : net->ipv4.sysctl_tcp_syn_retries;
3440 val = (val ? : net->ipv4.sysctl_tcp_fin_timeout) / HZ;
3442 case TCP_DEFER_ACCEPT:
3443 val = retrans_to_secs(icsk->icsk_accept_queue.rskq_defer_accept,
3444 TCP_TIMEOUT_INIT / HZ, TCP_RTO_MAX / HZ);
3446 case TCP_WINDOW_CLAMP:
3447 val = tp->window_clamp;
3450 struct tcp_info info;
3452 if (get_user(len, optlen))
3455 tcp_get_info(sk, &info);
3457 len = min_t(unsigned int, len, sizeof(info));
3458 if (put_user(len, optlen))
3460 if (copy_to_user(optval, &info, len))
3465 const struct tcp_congestion_ops *ca_ops;
3466 union tcp_cc_info info;
3470 if (get_user(len, optlen))
3473 ca_ops = icsk->icsk_ca_ops;
3474 if (ca_ops && ca_ops->get_info)
3475 sz = ca_ops->get_info(sk, ~0U, &attr, &info);
3477 len = min_t(unsigned int, len, sz);
3478 if (put_user(len, optlen))
3480 if (copy_to_user(optval, &info, len))
3485 val = !inet_csk_in_pingpong_mode(sk);
3488 case TCP_CONGESTION:
3489 if (get_user(len, optlen))
3491 len = min_t(unsigned int, len, TCP_CA_NAME_MAX);
3492 if (put_user(len, optlen))
3494 if (copy_to_user(optval, icsk->icsk_ca_ops->name, len))
3499 if (get_user(len, optlen))
3501 len = min_t(unsigned int, len, TCP_ULP_NAME_MAX);
3502 if (!icsk->icsk_ulp_ops) {
3503 if (put_user(0, optlen))
3507 if (put_user(len, optlen))
3509 if (copy_to_user(optval, icsk->icsk_ulp_ops->name, len))
3513 case TCP_FASTOPEN_KEY: {
3514 __u8 key[TCP_FASTOPEN_KEY_BUF_LENGTH];
3515 struct tcp_fastopen_context *ctx;
3516 unsigned int key_len = 0;
3518 if (get_user(len, optlen))
3522 ctx = rcu_dereference(icsk->icsk_accept_queue.fastopenq.ctx);
3524 key_len = tcp_fastopen_context_len(ctx) *
3525 TCP_FASTOPEN_KEY_LENGTH;
3526 memcpy(&key[0], &ctx->key[0], key_len);
3530 len = min_t(unsigned int, len, key_len);
3531 if (put_user(len, optlen))
3533 if (copy_to_user(optval, key, len))
3537 case TCP_THIN_LINEAR_TIMEOUTS:
3541 case TCP_THIN_DUPACK:
3549 case TCP_REPAIR_QUEUE:
3551 val = tp->repair_queue;
3556 case TCP_REPAIR_WINDOW: {
3557 struct tcp_repair_window opt;
3559 if (get_user(len, optlen))
3562 if (len != sizeof(opt))
3568 opt.snd_wl1 = tp->snd_wl1;
3569 opt.snd_wnd = tp->snd_wnd;
3570 opt.max_window = tp->max_window;
3571 opt.rcv_wnd = tp->rcv_wnd;
3572 opt.rcv_wup = tp->rcv_wup;
3574 if (copy_to_user(optval, &opt, len))
3579 if (tp->repair_queue == TCP_SEND_QUEUE)
3580 val = tp->write_seq;
3581 else if (tp->repair_queue == TCP_RECV_QUEUE)
3587 case TCP_USER_TIMEOUT:
3588 val = icsk->icsk_user_timeout;
3592 val = icsk->icsk_accept_queue.fastopenq.max_qlen;
3595 case TCP_FASTOPEN_CONNECT:
3596 val = tp->fastopen_connect;
3599 case TCP_FASTOPEN_NO_COOKIE:
3600 val = tp->fastopen_no_cookie;
3604 val = tp->tcp_tx_delay;
3608 val = tcp_time_stamp_raw() + tp->tsoffset;
3610 case TCP_NOTSENT_LOWAT:
3611 val = tp->notsent_lowat;
3614 val = tp->recvmsg_inq;
3619 case TCP_SAVED_SYN: {
3620 if (get_user(len, optlen))
3624 if (tp->saved_syn) {
3625 if (len < tp->saved_syn[0]) {
3626 if (put_user(tp->saved_syn[0], optlen)) {
3633 len = tp->saved_syn[0];
3634 if (put_user(len, optlen)) {
3638 if (copy_to_user(optval, tp->saved_syn + 1, len)) {
3642 tcp_saved_syn_free(tp);
3647 if (put_user(len, optlen))
3653 case TCP_ZEROCOPY_RECEIVE: {
3654 struct tcp_zerocopy_receive zc;
3657 if (get_user(len, optlen))
3659 if (len != sizeof(zc))
3661 if (copy_from_user(&zc, optval, len))
3664 err = tcp_zerocopy_receive(sk, &zc);
3666 if (!err && copy_to_user(optval, &zc, len))
3672 return -ENOPROTOOPT;
3675 if (put_user(len, optlen))
3677 if (copy_to_user(optval, &val, len))
3682 int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval,
3685 struct inet_connection_sock *icsk = inet_csk(sk);
3687 if (level != SOL_TCP)
3688 return icsk->icsk_af_ops->getsockopt(sk, level, optname,
3690 return do_tcp_getsockopt(sk, level, optname, optval, optlen);
3692 EXPORT_SYMBOL(tcp_getsockopt);
3694 #ifdef CONFIG_COMPAT
3695 int compat_tcp_getsockopt(struct sock *sk, int level, int optname,
3696 char __user *optval, int __user *optlen)
3698 if (level != SOL_TCP)
3699 return inet_csk_compat_getsockopt(sk, level, optname,
3701 return do_tcp_getsockopt(sk, level, optname, optval, optlen);
3703 EXPORT_SYMBOL(compat_tcp_getsockopt);
3706 #ifdef CONFIG_TCP_MD5SIG
3707 static DEFINE_PER_CPU(struct tcp_md5sig_pool, tcp_md5sig_pool);
3708 static DEFINE_MUTEX(tcp_md5sig_mutex);
3709 static bool tcp_md5sig_pool_populated = false;
3711 static void __tcp_alloc_md5sig_pool(void)
3713 struct crypto_ahash *hash;
3716 hash = crypto_alloc_ahash("md5", 0, CRYPTO_ALG_ASYNC);
3720 for_each_possible_cpu(cpu) {
3721 void *scratch = per_cpu(tcp_md5sig_pool, cpu).scratch;
3722 struct ahash_request *req;
3725 scratch = kmalloc_node(sizeof(union tcp_md5sum_block) +
3726 sizeof(struct tcphdr),
3731 per_cpu(tcp_md5sig_pool, cpu).scratch = scratch;
3733 if (per_cpu(tcp_md5sig_pool, cpu).md5_req)
3736 req = ahash_request_alloc(hash, GFP_KERNEL);
3740 ahash_request_set_callback(req, 0, NULL, NULL);
3742 per_cpu(tcp_md5sig_pool, cpu).md5_req = req;
3744 /* before setting tcp_md5sig_pool_populated, we must commit all writes
3745 * to memory. See smp_rmb() in tcp_get_md5sig_pool()
3748 tcp_md5sig_pool_populated = true;
3751 bool tcp_alloc_md5sig_pool(void)
3753 if (unlikely(!tcp_md5sig_pool_populated)) {
3754 mutex_lock(&tcp_md5sig_mutex);
3756 if (!tcp_md5sig_pool_populated) {
3757 __tcp_alloc_md5sig_pool();
3758 if (tcp_md5sig_pool_populated)
3759 static_branch_inc(&tcp_md5_needed);
3762 mutex_unlock(&tcp_md5sig_mutex);
3764 return tcp_md5sig_pool_populated;
3766 EXPORT_SYMBOL(tcp_alloc_md5sig_pool);
3770 * tcp_get_md5sig_pool - get md5sig_pool for this user
3772 * We use percpu structure, so if we succeed, we exit with preemption
3773 * and BH disabled, to make sure another thread or softirq handling
3774 * wont try to get same context.
3776 struct tcp_md5sig_pool *tcp_get_md5sig_pool(void)
3780 if (tcp_md5sig_pool_populated) {
3781 /* coupled with smp_wmb() in __tcp_alloc_md5sig_pool() */
3783 return this_cpu_ptr(&tcp_md5sig_pool);
3788 EXPORT_SYMBOL(tcp_get_md5sig_pool);
3790 int tcp_md5_hash_skb_data(struct tcp_md5sig_pool *hp,
3791 const struct sk_buff *skb, unsigned int header_len)
3793 struct scatterlist sg;
3794 const struct tcphdr *tp = tcp_hdr(skb);
3795 struct ahash_request *req = hp->md5_req;
3797 const unsigned int head_data_len = skb_headlen(skb) > header_len ?
3798 skb_headlen(skb) - header_len : 0;
3799 const struct skb_shared_info *shi = skb_shinfo(skb);
3800 struct sk_buff *frag_iter;
3802 sg_init_table(&sg, 1);
3804 sg_set_buf(&sg, ((u8 *) tp) + header_len, head_data_len);
3805 ahash_request_set_crypt(req, &sg, NULL, head_data_len);
3806 if (crypto_ahash_update(req))
3809 for (i = 0; i < shi->nr_frags; ++i) {
3810 const skb_frag_t *f = &shi->frags[i];
3811 unsigned int offset = skb_frag_off(f);
3812 struct page *page = skb_frag_page(f) + (offset >> PAGE_SHIFT);
3814 sg_set_page(&sg, page, skb_frag_size(f),
3815 offset_in_page(offset));
3816 ahash_request_set_crypt(req, &sg, NULL, skb_frag_size(f));
3817 if (crypto_ahash_update(req))
3821 skb_walk_frags(skb, frag_iter)
3822 if (tcp_md5_hash_skb_data(hp, frag_iter, 0))
3827 EXPORT_SYMBOL(tcp_md5_hash_skb_data);
3829 int tcp_md5_hash_key(struct tcp_md5sig_pool *hp, const struct tcp_md5sig_key *key)
3831 struct scatterlist sg;
3833 sg_init_one(&sg, key->key, key->keylen);
3834 ahash_request_set_crypt(hp->md5_req, &sg, NULL, key->keylen);
3835 return crypto_ahash_update(hp->md5_req);
3837 EXPORT_SYMBOL(tcp_md5_hash_key);
3841 void tcp_done(struct sock *sk)
3843 struct request_sock *req;
3845 /* We might be called with a new socket, after
3846 * inet_csk_prepare_forced_close() has been called
3847 * so we can not use lockdep_sock_is_held(sk)
3849 req = rcu_dereference_protected(tcp_sk(sk)->fastopen_rsk, 1);
3851 if (sk->sk_state == TCP_SYN_SENT || sk->sk_state == TCP_SYN_RECV)
3852 TCP_INC_STATS(sock_net(sk), TCP_MIB_ATTEMPTFAILS);
3854 tcp_set_state(sk, TCP_CLOSE);
3855 tcp_clear_xmit_timers(sk);
3857 reqsk_fastopen_remove(sk, req, false);
3859 sk->sk_shutdown = SHUTDOWN_MASK;
3861 if (!sock_flag(sk, SOCK_DEAD))
3862 sk->sk_state_change(sk);
3864 inet_csk_destroy_sock(sk);
3866 EXPORT_SYMBOL_GPL(tcp_done);
3868 int tcp_abort(struct sock *sk, int err)
3870 if (!sk_fullsock(sk)) {
3871 if (sk->sk_state == TCP_NEW_SYN_RECV) {
3872 struct request_sock *req = inet_reqsk(sk);
3875 inet_csk_reqsk_queue_drop(req->rsk_listener, req);
3882 /* Don't race with userspace socket closes such as tcp_close. */
3885 if (sk->sk_state == TCP_LISTEN) {
3886 tcp_set_state(sk, TCP_CLOSE);
3887 inet_csk_listen_stop(sk);
3890 /* Don't race with BH socket closes such as inet_csk_listen_stop. */
3894 if (!sock_flag(sk, SOCK_DEAD)) {
3896 /* This barrier is coupled with smp_rmb() in tcp_poll() */
3898 sk->sk_error_report(sk);
3899 if (tcp_need_reset(sk->sk_state))
3900 tcp_send_active_reset(sk, GFP_ATOMIC);
3906 tcp_write_queue_purge(sk);
3910 EXPORT_SYMBOL_GPL(tcp_abort);
3912 extern struct tcp_congestion_ops tcp_reno;
3914 static __initdata unsigned long thash_entries;
3915 static int __init set_thash_entries(char *str)
3922 ret = kstrtoul(str, 0, &thash_entries);
3928 __setup("thash_entries=", set_thash_entries);
3930 static void __init tcp_init_mem(void)
3932 unsigned long limit = nr_free_buffer_pages() / 16;
3934 limit = max(limit, 128UL);
3935 sysctl_tcp_mem[0] = limit / 4 * 3; /* 4.68 % */
3936 sysctl_tcp_mem[1] = limit; /* 6.25 % */
3937 sysctl_tcp_mem[2] = sysctl_tcp_mem[0] * 2; /* 9.37 % */
3940 void __init tcp_init(void)
3942 int max_rshare, max_wshare, cnt;
3943 unsigned long limit;
3946 BUILD_BUG_ON(TCP_MIN_SND_MSS <= MAX_TCP_OPTION_SPACE);
3947 BUILD_BUG_ON(sizeof(struct tcp_skb_cb) >
3948 FIELD_SIZEOF(struct sk_buff, cb));
3950 percpu_counter_init(&tcp_sockets_allocated, 0, GFP_KERNEL);
3951 percpu_counter_init(&tcp_orphan_count, 0, GFP_KERNEL);
3952 inet_hashinfo_init(&tcp_hashinfo);
3953 inet_hashinfo2_init(&tcp_hashinfo, "tcp_listen_portaddr_hash",
3954 thash_entries, 21, /* one slot per 2 MB*/
3956 tcp_hashinfo.bind_bucket_cachep =
3957 kmem_cache_create("tcp_bind_bucket",
3958 sizeof(struct inet_bind_bucket), 0,
3959 SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL);
3961 /* Size and allocate the main established and bind bucket
3964 * The methodology is similar to that of the buffer cache.
3966 tcp_hashinfo.ehash =
3967 alloc_large_system_hash("TCP established",
3968 sizeof(struct inet_ehash_bucket),
3970 17, /* one slot per 128 KB of memory */
3973 &tcp_hashinfo.ehash_mask,
3975 thash_entries ? 0 : 512 * 1024);
3976 for (i = 0; i <= tcp_hashinfo.ehash_mask; i++)
3977 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].chain, i);
3979 if (inet_ehash_locks_alloc(&tcp_hashinfo))
3980 panic("TCP: failed to alloc ehash_locks");
3981 tcp_hashinfo.bhash =
3982 alloc_large_system_hash("TCP bind",
3983 sizeof(struct inet_bind_hashbucket),
3984 tcp_hashinfo.ehash_mask + 1,
3985 17, /* one slot per 128 KB of memory */
3987 &tcp_hashinfo.bhash_size,
3991 tcp_hashinfo.bhash_size = 1U << tcp_hashinfo.bhash_size;
3992 for (i = 0; i < tcp_hashinfo.bhash_size; i++) {
3993 spin_lock_init(&tcp_hashinfo.bhash[i].lock);
3994 INIT_HLIST_HEAD(&tcp_hashinfo.bhash[i].chain);
3998 cnt = tcp_hashinfo.ehash_mask + 1;
3999 sysctl_tcp_max_orphans = cnt / 2;
4002 /* Set per-socket limits to no more than 1/128 the pressure threshold */
4003 limit = nr_free_buffer_pages() << (PAGE_SHIFT - 7);
4004 max_wshare = min(4UL*1024*1024, limit);
4005 max_rshare = min(6UL*1024*1024, limit);
4007 init_net.ipv4.sysctl_tcp_wmem[0] = SK_MEM_QUANTUM;
4008 init_net.ipv4.sysctl_tcp_wmem[1] = 16*1024;
4009 init_net.ipv4.sysctl_tcp_wmem[2] = max(64*1024, max_wshare);
4011 init_net.ipv4.sysctl_tcp_rmem[0] = SK_MEM_QUANTUM;
4012 init_net.ipv4.sysctl_tcp_rmem[1] = 131072;
4013 init_net.ipv4.sysctl_tcp_rmem[2] = max(131072, max_rshare);
4015 pr_info("Hash tables configured (established %u bind %u)\n",
4016 tcp_hashinfo.ehash_mask + 1, tcp_hashinfo.bhash_size);
4020 BUG_ON(tcp_register_congestion_control(&tcp_reno) != 0);
projects / linux.git / blob