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/mptcp.h>
275 #include <net/xfrm.h>
277 #include <net/sock.h>
279 #include <linux/uaccess.h>
280 #include <asm/ioctls.h>
281 #include <net/busy_poll.h>
283 DEFINE_PER_CPU(unsigned int, tcp_orphan_count);
284 EXPORT_PER_CPU_SYMBOL_GPL(tcp_orphan_count);
286 long sysctl_tcp_mem[3] __read_mostly;
287 EXPORT_SYMBOL(sysctl_tcp_mem);
289 atomic_long_t tcp_memory_allocated; /* Current allocated memory. */
290 EXPORT_SYMBOL(tcp_memory_allocated);
292 #if IS_ENABLED(CONFIG_SMC)
293 DEFINE_STATIC_KEY_FALSE(tcp_have_smc);
294 EXPORT_SYMBOL(tcp_have_smc);
298 * Current number of TCP sockets.
300 struct percpu_counter tcp_sockets_allocated;
301 EXPORT_SYMBOL(tcp_sockets_allocated);
306 struct tcp_splice_state {
307 struct pipe_inode_info *pipe;
313 * Pressure flag: try to collapse.
314 * Technical note: it is used by multiple contexts non atomically.
315 * All the __sk_mem_schedule() is of this nature: accounting
316 * is strict, actions are advisory and have some latency.
318 unsigned long tcp_memory_pressure __read_mostly;
319 EXPORT_SYMBOL_GPL(tcp_memory_pressure);
321 DEFINE_STATIC_KEY_FALSE(tcp_rx_skb_cache_key);
322 EXPORT_SYMBOL(tcp_rx_skb_cache_key);
324 DEFINE_STATIC_KEY_FALSE(tcp_tx_skb_cache_key);
326 void tcp_enter_memory_pressure(struct sock *sk)
330 if (READ_ONCE(tcp_memory_pressure))
336 if (!cmpxchg(&tcp_memory_pressure, 0, val))
337 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMEMORYPRESSURES);
339 EXPORT_SYMBOL_GPL(tcp_enter_memory_pressure);
341 void tcp_leave_memory_pressure(struct sock *sk)
345 if (!READ_ONCE(tcp_memory_pressure))
347 val = xchg(&tcp_memory_pressure, 0);
349 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPMEMORYPRESSURESCHRONO,
350 jiffies_to_msecs(jiffies - val));
352 EXPORT_SYMBOL_GPL(tcp_leave_memory_pressure);
354 /* Convert seconds to retransmits based on initial and max timeout */
355 static u8 secs_to_retrans(int seconds, int timeout, int rto_max)
360 int period = timeout;
363 while (seconds > period && res < 255) {
366 if (timeout > rto_max)
374 /* Convert retransmits to seconds based on initial and max timeout */
375 static int retrans_to_secs(u8 retrans, int timeout, int rto_max)
383 if (timeout > rto_max)
391 static u64 tcp_compute_delivery_rate(const struct tcp_sock *tp)
393 u32 rate = READ_ONCE(tp->rate_delivered);
394 u32 intv = READ_ONCE(tp->rate_interval_us);
398 rate64 = (u64)rate * tp->mss_cache * USEC_PER_SEC;
399 do_div(rate64, intv);
404 /* Address-family independent initialization for a tcp_sock.
406 * NOTE: A lot of things set to zero explicitly by call to
407 * sk_alloc() so need not be done here.
409 void tcp_init_sock(struct sock *sk)
411 struct inet_connection_sock *icsk = inet_csk(sk);
412 struct tcp_sock *tp = tcp_sk(sk);
414 tp->out_of_order_queue = RB_ROOT;
415 sk->tcp_rtx_queue = RB_ROOT;
416 tcp_init_xmit_timers(sk);
417 INIT_LIST_HEAD(&tp->tsq_node);
418 INIT_LIST_HEAD(&tp->tsorted_sent_queue);
420 icsk->icsk_rto = TCP_TIMEOUT_INIT;
421 icsk->icsk_rto_min = TCP_RTO_MIN;
422 icsk->icsk_delack_max = TCP_DELACK_MAX;
423 tp->mdev_us = jiffies_to_usecs(TCP_TIMEOUT_INIT);
424 minmax_reset(&tp->rtt_min, tcp_jiffies32, ~0U);
426 /* So many TCP implementations out there (incorrectly) count the
427 * initial SYN frame in their delayed-ACK and congestion control
428 * algorithms that we must have the following bandaid to talk
429 * efficiently to them. -DaveM
431 tp->snd_cwnd = TCP_INIT_CWND;
433 /* There's a bubble in the pipe until at least the first ACK. */
434 tp->app_limited = ~0U;
436 /* See draft-stevens-tcpca-spec-01 for discussion of the
437 * initialization of these values.
439 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
440 tp->snd_cwnd_clamp = ~0;
441 tp->mss_cache = TCP_MSS_DEFAULT;
443 tp->reordering = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_reordering);
444 tcp_assign_congestion_control(sk);
447 tp->rack.reo_wnd_steps = 1;
449 sk->sk_write_space = sk_stream_write_space;
450 sock_set_flag(sk, SOCK_USE_WRITE_QUEUE);
452 icsk->icsk_sync_mss = tcp_sync_mss;
454 WRITE_ONCE(sk->sk_sndbuf, READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_wmem[1]));
455 WRITE_ONCE(sk->sk_rcvbuf, READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_rmem[1]));
457 sk_sockets_allocated_inc(sk);
458 sk->sk_route_forced_caps = NETIF_F_GSO;
460 EXPORT_SYMBOL(tcp_init_sock);
462 static void tcp_tx_timestamp(struct sock *sk, u16 tsflags)
464 struct sk_buff *skb = tcp_write_queue_tail(sk);
466 if (tsflags && skb) {
467 struct skb_shared_info *shinfo = skb_shinfo(skb);
468 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
470 sock_tx_timestamp(sk, tsflags, &shinfo->tx_flags);
471 if (tsflags & SOF_TIMESTAMPING_TX_ACK)
472 tcb->txstamp_ack = 1;
473 if (tsflags & SOF_TIMESTAMPING_TX_RECORD_MASK)
474 shinfo->tskey = TCP_SKB_CB(skb)->seq + skb->len - 1;
478 static inline bool tcp_stream_is_readable(const struct tcp_sock *tp,
479 int target, struct sock *sk)
481 int avail = READ_ONCE(tp->rcv_nxt) - READ_ONCE(tp->copied_seq);
486 if (tcp_rmem_pressure(sk))
488 if (tcp_receive_window(tp) <= inet_csk(sk)->icsk_ack.rcv_mss)
491 if (sk->sk_prot->stream_memory_read)
492 return sk->sk_prot->stream_memory_read(sk);
497 * Wait for a TCP event.
499 * Note that we don't need to lock the socket, as the upper poll layers
500 * take care of normal races (between the test and the event) and we don't
501 * go look at any of the socket buffers directly.
503 __poll_t tcp_poll(struct file *file, struct socket *sock, poll_table *wait)
506 struct sock *sk = sock->sk;
507 const struct tcp_sock *tp = tcp_sk(sk);
510 sock_poll_wait(file, sock, wait);
512 state = inet_sk_state_load(sk);
513 if (state == TCP_LISTEN)
514 return inet_csk_listen_poll(sk);
516 /* Socket is not locked. We are protected from async events
517 * by poll logic and correct handling of state changes
518 * made by other threads is impossible in any case.
524 * EPOLLHUP is certainly not done right. But poll() doesn't
525 * have a notion of HUP in just one direction, and for a
526 * socket the read side is more interesting.
528 * Some poll() documentation says that EPOLLHUP is incompatible
529 * with the EPOLLOUT/POLLWR flags, so somebody should check this
530 * all. But careful, it tends to be safer to return too many
531 * bits than too few, and you can easily break real applications
532 * if you don't tell them that something has hung up!
536 * Check number 1. EPOLLHUP is _UNMASKABLE_ event (see UNIX98 and
537 * our fs/select.c). It means that after we received EOF,
538 * poll always returns immediately, making impossible poll() on write()
539 * in state CLOSE_WAIT. One solution is evident --- to set EPOLLHUP
540 * if and only if shutdown has been made in both directions.
541 * Actually, it is interesting to look how Solaris and DUX
542 * solve this dilemma. I would prefer, if EPOLLHUP were maskable,
543 * then we could set it on SND_SHUTDOWN. BTW examples given
544 * in Stevens' books assume exactly this behaviour, it explains
545 * why EPOLLHUP is incompatible with EPOLLOUT. --ANK
547 * NOTE. Check for TCP_CLOSE is added. The goal is to prevent
548 * blocking on fresh not-connected or disconnected socket. --ANK
550 if (sk->sk_shutdown == SHUTDOWN_MASK || state == TCP_CLOSE)
552 if (sk->sk_shutdown & RCV_SHUTDOWN)
553 mask |= EPOLLIN | EPOLLRDNORM | EPOLLRDHUP;
555 /* Connected or passive Fast Open socket? */
556 if (state != TCP_SYN_SENT &&
557 (state != TCP_SYN_RECV || rcu_access_pointer(tp->fastopen_rsk))) {
558 int target = sock_rcvlowat(sk, 0, INT_MAX);
560 if (READ_ONCE(tp->urg_seq) == READ_ONCE(tp->copied_seq) &&
561 !sock_flag(sk, SOCK_URGINLINE) &&
565 if (tcp_stream_is_readable(tp, target, sk))
566 mask |= EPOLLIN | EPOLLRDNORM;
568 if (!(sk->sk_shutdown & SEND_SHUTDOWN)) {
569 if (__sk_stream_is_writeable(sk, 1)) {
570 mask |= EPOLLOUT | EPOLLWRNORM;
571 } else { /* send SIGIO later */
572 sk_set_bit(SOCKWQ_ASYNC_NOSPACE, sk);
573 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
575 /* Race breaker. If space is freed after
576 * wspace test but before the flags are set,
577 * IO signal will be lost. Memory barrier
578 * pairs with the input side.
580 smp_mb__after_atomic();
581 if (__sk_stream_is_writeable(sk, 1))
582 mask |= EPOLLOUT | EPOLLWRNORM;
585 mask |= EPOLLOUT | EPOLLWRNORM;
587 if (tp->urg_data & TCP_URG_VALID)
589 } else if (state == TCP_SYN_SENT && inet_sk(sk)->defer_connect) {
590 /* Active TCP fastopen socket with defer_connect
591 * Return EPOLLOUT so application can call write()
592 * in order for kernel to generate SYN+data
594 mask |= EPOLLOUT | EPOLLWRNORM;
596 /* This barrier is coupled with smp_wmb() in tcp_reset() */
598 if (sk->sk_err || !skb_queue_empty_lockless(&sk->sk_error_queue))
603 EXPORT_SYMBOL(tcp_poll);
605 int tcp_ioctl(struct sock *sk, int cmd, unsigned long arg)
607 struct tcp_sock *tp = tcp_sk(sk);
613 if (sk->sk_state == TCP_LISTEN)
616 slow = lock_sock_fast(sk);
618 unlock_sock_fast(sk, slow);
621 answ = tp->urg_data &&
622 READ_ONCE(tp->urg_seq) == READ_ONCE(tp->copied_seq);
625 if (sk->sk_state == TCP_LISTEN)
628 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
631 answ = READ_ONCE(tp->write_seq) - tp->snd_una;
634 if (sk->sk_state == TCP_LISTEN)
637 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
640 answ = READ_ONCE(tp->write_seq) -
641 READ_ONCE(tp->snd_nxt);
647 return put_user(answ, (int __user *)arg);
649 EXPORT_SYMBOL(tcp_ioctl);
651 static inline void tcp_mark_push(struct tcp_sock *tp, struct sk_buff *skb)
653 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
654 tp->pushed_seq = tp->write_seq;
657 static inline bool forced_push(const struct tcp_sock *tp)
659 return after(tp->write_seq, tp->pushed_seq + (tp->max_window >> 1));
662 static void skb_entail(struct sock *sk, struct sk_buff *skb)
664 struct tcp_sock *tp = tcp_sk(sk);
665 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
668 tcb->seq = tcb->end_seq = tp->write_seq;
669 tcb->tcp_flags = TCPHDR_ACK;
671 __skb_header_release(skb);
672 tcp_add_write_queue_tail(sk, skb);
673 sk_wmem_queued_add(sk, skb->truesize);
674 sk_mem_charge(sk, skb->truesize);
675 if (tp->nonagle & TCP_NAGLE_PUSH)
676 tp->nonagle &= ~TCP_NAGLE_PUSH;
678 tcp_slow_start_after_idle_check(sk);
681 static inline void tcp_mark_urg(struct tcp_sock *tp, int flags)
684 tp->snd_up = tp->write_seq;
687 /* If a not yet filled skb is pushed, do not send it if
688 * we have data packets in Qdisc or NIC queues :
689 * Because TX completion will happen shortly, it gives a chance
690 * to coalesce future sendmsg() payload into this skb, without
691 * need for a timer, and with no latency trade off.
692 * As packets containing data payload have a bigger truesize
693 * than pure acks (dataless) packets, the last checks prevent
694 * autocorking if we only have an ACK in Qdisc/NIC queues,
695 * or if TX completion was delayed after we processed ACK packet.
697 static bool tcp_should_autocork(struct sock *sk, struct sk_buff *skb,
700 return skb->len < size_goal &&
701 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_autocorking) &&
702 !tcp_rtx_queue_empty(sk) &&
703 refcount_read(&sk->sk_wmem_alloc) > skb->truesize;
706 void tcp_push(struct sock *sk, int flags, int mss_now,
707 int nonagle, int size_goal)
709 struct tcp_sock *tp = tcp_sk(sk);
712 skb = tcp_write_queue_tail(sk);
715 if (!(flags & MSG_MORE) || forced_push(tp))
716 tcp_mark_push(tp, skb);
718 tcp_mark_urg(tp, flags);
720 if (tcp_should_autocork(sk, skb, size_goal)) {
722 /* avoid atomic op if TSQ_THROTTLED bit is already set */
723 if (!test_bit(TSQ_THROTTLED, &sk->sk_tsq_flags)) {
724 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPAUTOCORKING);
725 set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags);
727 /* It is possible TX completion already happened
728 * before we set TSQ_THROTTLED.
730 if (refcount_read(&sk->sk_wmem_alloc) > skb->truesize)
734 if (flags & MSG_MORE)
735 nonagle = TCP_NAGLE_CORK;
737 __tcp_push_pending_frames(sk, mss_now, nonagle);
740 static int tcp_splice_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb,
741 unsigned int offset, size_t len)
743 struct tcp_splice_state *tss = rd_desc->arg.data;
746 ret = skb_splice_bits(skb, skb->sk, offset, tss->pipe,
747 min(rd_desc->count, len), tss->flags);
749 rd_desc->count -= ret;
753 static int __tcp_splice_read(struct sock *sk, struct tcp_splice_state *tss)
755 /* Store TCP splice context information in read_descriptor_t. */
756 read_descriptor_t rd_desc = {
761 return tcp_read_sock(sk, &rd_desc, tcp_splice_data_recv);
765 * tcp_splice_read - splice data from TCP socket to a pipe
766 * @sock: socket to splice from
767 * @ppos: position (not valid)
768 * @pipe: pipe to splice to
769 * @len: number of bytes to splice
770 * @flags: splice modifier flags
773 * Will read pages from given socket and fill them into a pipe.
776 ssize_t tcp_splice_read(struct socket *sock, loff_t *ppos,
777 struct pipe_inode_info *pipe, size_t len,
780 struct sock *sk = sock->sk;
781 struct tcp_splice_state tss = {
790 sock_rps_record_flow(sk);
792 * We can't seek on a socket input
801 timeo = sock_rcvtimeo(sk, sock->file->f_flags & O_NONBLOCK);
803 ret = __tcp_splice_read(sk, &tss);
809 if (sock_flag(sk, SOCK_DONE))
812 ret = sock_error(sk);
815 if (sk->sk_shutdown & RCV_SHUTDOWN)
817 if (sk->sk_state == TCP_CLOSE) {
819 * This occurs when user tries to read
820 * from never connected socket.
829 /* if __tcp_splice_read() got nothing while we have
830 * an skb in receive queue, we do not want to loop.
831 * This might happen with URG data.
833 if (!skb_queue_empty(&sk->sk_receive_queue))
835 sk_wait_data(sk, &timeo, NULL);
836 if (signal_pending(current)) {
837 ret = sock_intr_errno(timeo);
850 if (sk->sk_err || sk->sk_state == TCP_CLOSE ||
851 (sk->sk_shutdown & RCV_SHUTDOWN) ||
852 signal_pending(current))
863 EXPORT_SYMBOL(tcp_splice_read);
865 struct sk_buff *sk_stream_alloc_skb(struct sock *sk, int size, gfp_t gfp,
871 skb = sk->sk_tx_skb_cache;
873 skb->truesize = SKB_TRUESIZE(skb_end_offset(skb));
874 sk->sk_tx_skb_cache = NULL;
876 INIT_LIST_HEAD(&skb->tcp_tsorted_anchor);
877 skb_shinfo(skb)->tx_flags = 0;
878 memset(TCP_SKB_CB(skb), 0, sizeof(struct tcp_skb_cb));
882 /* The TCP header must be at least 32-bit aligned. */
883 size = ALIGN(size, 4);
885 if (unlikely(tcp_under_memory_pressure(sk)))
886 sk_mem_reclaim_partial(sk);
888 skb = alloc_skb_fclone(size + sk->sk_prot->max_header, gfp);
892 if (force_schedule) {
893 mem_scheduled = true;
894 sk_forced_mem_schedule(sk, skb->truesize);
896 mem_scheduled = sk_wmem_schedule(sk, skb->truesize);
898 if (likely(mem_scheduled)) {
899 skb_reserve(skb, sk->sk_prot->max_header);
901 * Make sure that we have exactly size bytes
902 * available to the caller, no more, no less.
904 skb->reserved_tailroom = skb->end - skb->tail - size;
905 INIT_LIST_HEAD(&skb->tcp_tsorted_anchor);
910 sk->sk_prot->enter_memory_pressure(sk);
911 sk_stream_moderate_sndbuf(sk);
916 static unsigned int tcp_xmit_size_goal(struct sock *sk, u32 mss_now,
919 struct tcp_sock *tp = tcp_sk(sk);
920 u32 new_size_goal, size_goal;
925 /* Note : tcp_tso_autosize() will eventually split this later */
926 new_size_goal = sk->sk_gso_max_size - 1 - MAX_TCP_HEADER;
927 new_size_goal = tcp_bound_to_half_wnd(tp, new_size_goal);
929 /* We try hard to avoid divides here */
930 size_goal = tp->gso_segs * mss_now;
931 if (unlikely(new_size_goal < size_goal ||
932 new_size_goal >= size_goal + mss_now)) {
933 tp->gso_segs = min_t(u16, new_size_goal / mss_now,
934 sk->sk_gso_max_segs);
935 size_goal = tp->gso_segs * mss_now;
938 return max(size_goal, mss_now);
941 int tcp_send_mss(struct sock *sk, int *size_goal, int flags)
945 mss_now = tcp_current_mss(sk);
946 *size_goal = tcp_xmit_size_goal(sk, mss_now, !(flags & MSG_OOB));
951 /* In some cases, both sendpage() and sendmsg() could have added
952 * an skb to the write queue, but failed adding payload on it.
953 * We need to remove it to consume less memory, but more
954 * importantly be able to generate EPOLLOUT for Edge Trigger epoll()
957 static void tcp_remove_empty_skb(struct sock *sk, struct sk_buff *skb)
959 if (skb && TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq) {
960 tcp_unlink_write_queue(skb, sk);
961 if (tcp_write_queue_empty(sk))
962 tcp_chrono_stop(sk, TCP_CHRONO_BUSY);
963 sk_wmem_free_skb(sk, skb);
967 ssize_t do_tcp_sendpages(struct sock *sk, struct page *page, int offset,
968 size_t size, int flags)
970 struct tcp_sock *tp = tcp_sk(sk);
971 int mss_now, size_goal;
974 long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
976 if (IS_ENABLED(CONFIG_DEBUG_VM) &&
977 WARN_ONCE(!sendpage_ok(page),
978 "page must not be a Slab one and have page_count > 0"))
981 /* Wait for a connection to finish. One exception is TCP Fast Open
982 * (passive side) where data is allowed to be sent before a connection
983 * is fully established.
985 if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) &&
986 !tcp_passive_fastopen(sk)) {
987 err = sk_stream_wait_connect(sk, &timeo);
992 sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
994 mss_now = tcp_send_mss(sk, &size_goal, flags);
998 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
1002 struct sk_buff *skb = tcp_write_queue_tail(sk);
1006 if (!skb || (copy = size_goal - skb->len) <= 0 ||
1007 !tcp_skb_can_collapse_to(skb)) {
1009 if (!sk_stream_memory_free(sk))
1010 goto wait_for_space;
1012 skb = sk_stream_alloc_skb(sk, 0, sk->sk_allocation,
1013 tcp_rtx_and_write_queues_empty(sk));
1015 goto wait_for_space;
1017 #ifdef CONFIG_TLS_DEVICE
1018 skb->decrypted = !!(flags & MSG_SENDPAGE_DECRYPTED);
1020 skb_entail(sk, skb);
1027 i = skb_shinfo(skb)->nr_frags;
1028 can_coalesce = skb_can_coalesce(skb, i, page, offset);
1029 if (!can_coalesce && i >= sysctl_max_skb_frags) {
1030 tcp_mark_push(tp, skb);
1033 if (!sk_wmem_schedule(sk, copy))
1034 goto wait_for_space;
1037 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1040 skb_fill_page_desc(skb, i, page, offset, copy);
1043 if (!(flags & MSG_NO_SHARED_FRAGS))
1044 skb_shinfo(skb)->tx_flags |= SKBTX_SHARED_FRAG;
1047 skb->data_len += copy;
1048 skb->truesize += copy;
1049 sk_wmem_queued_add(sk, copy);
1050 sk_mem_charge(sk, copy);
1051 skb->ip_summed = CHECKSUM_PARTIAL;
1052 WRITE_ONCE(tp->write_seq, tp->write_seq + copy);
1053 TCP_SKB_CB(skb)->end_seq += copy;
1054 tcp_skb_pcount_set(skb, 0);
1057 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
1065 if (skb->len < size_goal || (flags & MSG_OOB))
1068 if (forced_push(tp)) {
1069 tcp_mark_push(tp, skb);
1070 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
1071 } else if (skb == tcp_send_head(sk))
1072 tcp_push_one(sk, mss_now);
1076 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1077 tcp_push(sk, flags & ~MSG_MORE, mss_now,
1078 TCP_NAGLE_PUSH, size_goal);
1080 err = sk_stream_wait_memory(sk, &timeo);
1084 mss_now = tcp_send_mss(sk, &size_goal, flags);
1089 tcp_tx_timestamp(sk, sk->sk_tsflags);
1090 if (!(flags & MSG_SENDPAGE_NOTLAST))
1091 tcp_push(sk, flags, mss_now, tp->nonagle, size_goal);
1096 tcp_remove_empty_skb(sk, tcp_write_queue_tail(sk));
1100 /* make sure we wake any epoll edge trigger waiter */
1101 if (unlikely(tcp_rtx_and_write_queues_empty(sk) && err == -EAGAIN)) {
1102 sk->sk_write_space(sk);
1103 tcp_chrono_stop(sk, TCP_CHRONO_SNDBUF_LIMITED);
1105 return sk_stream_error(sk, flags, err);
1107 EXPORT_SYMBOL_GPL(do_tcp_sendpages);
1109 int tcp_sendpage_locked(struct sock *sk, struct page *page, int offset,
1110 size_t size, int flags)
1112 if (!(sk->sk_route_caps & NETIF_F_SG))
1113 return sock_no_sendpage_locked(sk, page, offset, size, flags);
1115 tcp_rate_check_app_limited(sk); /* is sending application-limited? */
1117 return do_tcp_sendpages(sk, page, offset, size, flags);
1119 EXPORT_SYMBOL_GPL(tcp_sendpage_locked);
1121 int tcp_sendpage(struct sock *sk, struct page *page, int offset,
1122 size_t size, int flags)
1127 ret = tcp_sendpage_locked(sk, page, offset, size, flags);
1132 EXPORT_SYMBOL(tcp_sendpage);
1134 void tcp_free_fastopen_req(struct tcp_sock *tp)
1136 if (tp->fastopen_req) {
1137 kfree(tp->fastopen_req);
1138 tp->fastopen_req = NULL;
1142 static int tcp_sendmsg_fastopen(struct sock *sk, struct msghdr *msg,
1143 int *copied, size_t size,
1144 struct ubuf_info *uarg)
1146 struct tcp_sock *tp = tcp_sk(sk);
1147 struct inet_sock *inet = inet_sk(sk);
1148 struct sockaddr *uaddr = msg->msg_name;
1151 if (!(READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_fastopen) &
1152 TFO_CLIENT_ENABLE) ||
1153 (uaddr && msg->msg_namelen >= sizeof(uaddr->sa_family) &&
1154 uaddr->sa_family == AF_UNSPEC))
1156 if (tp->fastopen_req)
1157 return -EALREADY; /* Another Fast Open is in progress */
1159 tp->fastopen_req = kzalloc(sizeof(struct tcp_fastopen_request),
1161 if (unlikely(!tp->fastopen_req))
1163 tp->fastopen_req->data = msg;
1164 tp->fastopen_req->size = size;
1165 tp->fastopen_req->uarg = uarg;
1167 if (inet->defer_connect) {
1168 err = tcp_connect(sk);
1169 /* Same failure procedure as in tcp_v4/6_connect */
1171 tcp_set_state(sk, TCP_CLOSE);
1172 inet->inet_dport = 0;
1173 sk->sk_route_caps = 0;
1176 flags = (msg->msg_flags & MSG_DONTWAIT) ? O_NONBLOCK : 0;
1177 err = __inet_stream_connect(sk->sk_socket, uaddr,
1178 msg->msg_namelen, flags, 1);
1179 /* fastopen_req could already be freed in __inet_stream_connect
1180 * if the connection times out or gets rst
1182 if (tp->fastopen_req) {
1183 *copied = tp->fastopen_req->copied;
1184 tcp_free_fastopen_req(tp);
1185 inet->defer_connect = 0;
1190 int tcp_sendmsg_locked(struct sock *sk, struct msghdr *msg, size_t size)
1192 struct tcp_sock *tp = tcp_sk(sk);
1193 struct ubuf_info *uarg = NULL;
1194 struct sk_buff *skb;
1195 struct sockcm_cookie sockc;
1196 int flags, err, copied = 0;
1197 int mss_now = 0, size_goal, copied_syn = 0;
1198 int process_backlog = 0;
1202 flags = msg->msg_flags;
1204 if (flags & MSG_ZEROCOPY && size && sock_flag(sk, SOCK_ZEROCOPY)) {
1205 skb = tcp_write_queue_tail(sk);
1206 uarg = sock_zerocopy_realloc(sk, size, skb_zcopy(skb));
1212 zc = sk->sk_route_caps & NETIF_F_SG;
1217 if (unlikely(flags & MSG_FASTOPEN || inet_sk(sk)->defer_connect) &&
1219 err = tcp_sendmsg_fastopen(sk, msg, &copied_syn, size, uarg);
1220 if (err == -EINPROGRESS && copied_syn > 0)
1226 timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
1228 tcp_rate_check_app_limited(sk); /* is sending application-limited? */
1230 /* Wait for a connection to finish. One exception is TCP Fast Open
1231 * (passive side) where data is allowed to be sent before a connection
1232 * is fully established.
1234 if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) &&
1235 !tcp_passive_fastopen(sk)) {
1236 err = sk_stream_wait_connect(sk, &timeo);
1241 if (unlikely(tp->repair)) {
1242 if (tp->repair_queue == TCP_RECV_QUEUE) {
1243 copied = tcp_send_rcvq(sk, msg, size);
1248 if (tp->repair_queue == TCP_NO_QUEUE)
1251 /* 'common' sending to sendq */
1254 sockcm_init(&sockc, sk);
1255 if (msg->msg_controllen) {
1256 err = sock_cmsg_send(sk, msg, &sockc);
1257 if (unlikely(err)) {
1263 /* This should be in poll */
1264 sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
1266 /* Ok commence sending. */
1270 mss_now = tcp_send_mss(sk, &size_goal, flags);
1273 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
1276 while (msg_data_left(msg)) {
1279 skb = tcp_write_queue_tail(sk);
1281 copy = size_goal - skb->len;
1283 if (copy <= 0 || !tcp_skb_can_collapse_to(skb)) {
1287 if (!sk_stream_memory_free(sk))
1288 goto wait_for_space;
1290 if (unlikely(process_backlog >= 16)) {
1291 process_backlog = 0;
1292 if (sk_flush_backlog(sk))
1295 first_skb = tcp_rtx_and_write_queues_empty(sk);
1296 skb = sk_stream_alloc_skb(sk, 0, sk->sk_allocation,
1299 goto wait_for_space;
1302 skb->ip_summed = CHECKSUM_PARTIAL;
1304 skb_entail(sk, skb);
1307 /* All packets are restored as if they have
1308 * already been sent. skb_mstamp_ns isn't set to
1309 * avoid wrong rtt estimation.
1312 TCP_SKB_CB(skb)->sacked |= TCPCB_REPAIRED;
1315 /* Try to append data to the end of skb. */
1316 if (copy > msg_data_left(msg))
1317 copy = msg_data_left(msg);
1319 /* Where to copy to? */
1320 if (skb_availroom(skb) > 0 && !zc) {
1321 /* We have some space in skb head. Superb! */
1322 copy = min_t(int, copy, skb_availroom(skb));
1323 err = skb_add_data_nocache(sk, skb, &msg->msg_iter, copy);
1328 int i = skb_shinfo(skb)->nr_frags;
1329 struct page_frag *pfrag = sk_page_frag(sk);
1331 if (!sk_page_frag_refill(sk, pfrag))
1332 goto wait_for_space;
1334 if (!skb_can_coalesce(skb, i, pfrag->page,
1336 if (i >= sysctl_max_skb_frags) {
1337 tcp_mark_push(tp, skb);
1343 copy = min_t(int, copy, pfrag->size - pfrag->offset);
1345 if (!sk_wmem_schedule(sk, copy))
1346 goto wait_for_space;
1348 err = skb_copy_to_page_nocache(sk, &msg->msg_iter, skb,
1355 /* Update the skb. */
1357 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1359 skb_fill_page_desc(skb, i, pfrag->page,
1360 pfrag->offset, copy);
1361 page_ref_inc(pfrag->page);
1363 pfrag->offset += copy;
1365 if (!sk_wmem_schedule(sk, copy))
1366 goto wait_for_space;
1368 err = skb_zerocopy_iter_stream(sk, skb, msg, copy, uarg);
1369 if (err == -EMSGSIZE || err == -EEXIST) {
1370 tcp_mark_push(tp, skb);
1379 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
1381 WRITE_ONCE(tp->write_seq, tp->write_seq + copy);
1382 TCP_SKB_CB(skb)->end_seq += copy;
1383 tcp_skb_pcount_set(skb, 0);
1386 if (!msg_data_left(msg)) {
1387 if (unlikely(flags & MSG_EOR))
1388 TCP_SKB_CB(skb)->eor = 1;
1392 if (skb->len < size_goal || (flags & MSG_OOB) || unlikely(tp->repair))
1395 if (forced_push(tp)) {
1396 tcp_mark_push(tp, skb);
1397 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
1398 } else if (skb == tcp_send_head(sk))
1399 tcp_push_one(sk, mss_now);
1403 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1405 tcp_push(sk, flags & ~MSG_MORE, mss_now,
1406 TCP_NAGLE_PUSH, size_goal);
1408 err = sk_stream_wait_memory(sk, &timeo);
1412 mss_now = tcp_send_mss(sk, &size_goal, flags);
1417 tcp_tx_timestamp(sk, sockc.tsflags);
1418 tcp_push(sk, flags, mss_now, tp->nonagle, size_goal);
1421 sock_zerocopy_put(uarg);
1422 return copied + copied_syn;
1425 skb = tcp_write_queue_tail(sk);
1427 tcp_remove_empty_skb(sk, skb);
1429 if (copied + copied_syn)
1432 sock_zerocopy_put_abort(uarg, true);
1433 err = sk_stream_error(sk, flags, err);
1434 /* make sure we wake any epoll edge trigger waiter */
1435 if (unlikely(tcp_rtx_and_write_queues_empty(sk) && err == -EAGAIN)) {
1436 sk->sk_write_space(sk);
1437 tcp_chrono_stop(sk, TCP_CHRONO_SNDBUF_LIMITED);
1441 EXPORT_SYMBOL_GPL(tcp_sendmsg_locked);
1443 int tcp_sendmsg(struct sock *sk, struct msghdr *msg, size_t size)
1448 ret = tcp_sendmsg_locked(sk, msg, size);
1453 EXPORT_SYMBOL(tcp_sendmsg);
1456 * Handle reading urgent data. BSD has very simple semantics for
1457 * this, no blocking and very strange errors 8)
1460 static int tcp_recv_urg(struct sock *sk, struct msghdr *msg, int len, int flags)
1462 struct tcp_sock *tp = tcp_sk(sk);
1464 /* No URG data to read. */
1465 if (sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data ||
1466 tp->urg_data == TCP_URG_READ)
1467 return -EINVAL; /* Yes this is right ! */
1469 if (sk->sk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DONE))
1472 if (tp->urg_data & TCP_URG_VALID) {
1474 char c = tp->urg_data;
1476 if (!(flags & MSG_PEEK))
1477 tp->urg_data = TCP_URG_READ;
1479 /* Read urgent data. */
1480 msg->msg_flags |= MSG_OOB;
1483 if (!(flags & MSG_TRUNC))
1484 err = memcpy_to_msg(msg, &c, 1);
1487 msg->msg_flags |= MSG_TRUNC;
1489 return err ? -EFAULT : len;
1492 if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
1495 /* Fixed the recv(..., MSG_OOB) behaviour. BSD docs and
1496 * the available implementations agree in this case:
1497 * this call should never block, independent of the
1498 * blocking state of the socket.
1499 * Mike <pall@rz.uni-karlsruhe.de>
1504 static int tcp_peek_sndq(struct sock *sk, struct msghdr *msg, int len)
1506 struct sk_buff *skb;
1507 int copied = 0, err = 0;
1509 /* XXX -- need to support SO_PEEK_OFF */
1511 skb_rbtree_walk(skb, &sk->tcp_rtx_queue) {
1512 err = skb_copy_datagram_msg(skb, 0, msg, skb->len);
1518 skb_queue_walk(&sk->sk_write_queue, skb) {
1519 err = skb_copy_datagram_msg(skb, 0, msg, skb->len);
1526 return err ?: copied;
1529 /* Clean up the receive buffer for full frames taken by the user,
1530 * then send an ACK if necessary. COPIED is the number of bytes
1531 * tcp_recvmsg has given to the user so far, it speeds up the
1532 * calculation of whether or not we must ACK for the sake of
1535 void tcp_cleanup_rbuf(struct sock *sk, int copied)
1537 struct tcp_sock *tp = tcp_sk(sk);
1538 bool time_to_ack = false;
1540 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
1542 WARN(skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq),
1543 "cleanup rbuf bug: copied %X seq %X rcvnxt %X\n",
1544 tp->copied_seq, TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt);
1546 if (inet_csk_ack_scheduled(sk)) {
1547 const struct inet_connection_sock *icsk = inet_csk(sk);
1549 if (/* Once-per-two-segments ACK was not sent by tcp_input.c */
1550 tp->rcv_nxt - tp->rcv_wup > icsk->icsk_ack.rcv_mss ||
1552 * If this read emptied read buffer, we send ACK, if
1553 * connection is not bidirectional, user drained
1554 * receive buffer and there was a small segment
1558 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED2) ||
1559 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED) &&
1560 !inet_csk_in_pingpong_mode(sk))) &&
1561 !atomic_read(&sk->sk_rmem_alloc)))
1565 /* We send an ACK if we can now advertise a non-zero window
1566 * which has been raised "significantly".
1568 * Even if window raised up to infinity, do not send window open ACK
1569 * in states, where we will not receive more. It is useless.
1571 if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) {
1572 __u32 rcv_window_now = tcp_receive_window(tp);
1574 /* Optimize, __tcp_select_window() is not cheap. */
1575 if (2*rcv_window_now <= tp->window_clamp) {
1576 __u32 new_window = __tcp_select_window(sk);
1578 /* Send ACK now, if this read freed lots of space
1579 * in our buffer. Certainly, new_window is new window.
1580 * We can advertise it now, if it is not less than current one.
1581 * "Lots" means "at least twice" here.
1583 if (new_window && new_window >= 2 * rcv_window_now)
1591 static struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off)
1593 struct sk_buff *skb;
1596 while ((skb = skb_peek(&sk->sk_receive_queue)) != NULL) {
1597 offset = seq - TCP_SKB_CB(skb)->seq;
1598 if (unlikely(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
1599 pr_err_once("%s: found a SYN, please report !\n", __func__);
1602 if (offset < skb->len || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)) {
1606 /* This looks weird, but this can happen if TCP collapsing
1607 * splitted a fat GRO packet, while we released socket lock
1608 * in skb_splice_bits()
1610 sk_eat_skb(sk, skb);
1616 * This routine provides an alternative to tcp_recvmsg() for routines
1617 * that would like to handle copying from skbuffs directly in 'sendfile'
1620 * - It is assumed that the socket was locked by the caller.
1621 * - The routine does not block.
1622 * - At present, there is no support for reading OOB data
1623 * or for 'peeking' the socket using this routine
1624 * (although both would be easy to implement).
1626 int tcp_read_sock(struct sock *sk, read_descriptor_t *desc,
1627 sk_read_actor_t recv_actor)
1629 struct sk_buff *skb;
1630 struct tcp_sock *tp = tcp_sk(sk);
1631 u32 seq = tp->copied_seq;
1635 if (sk->sk_state == TCP_LISTEN)
1637 while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1638 if (offset < skb->len) {
1642 len = skb->len - offset;
1643 /* Stop reading if we hit a patch of urgent data */
1645 u32 urg_offset = tp->urg_seq - seq;
1646 if (urg_offset < len)
1651 used = recv_actor(desc, skb, offset, len);
1657 if (WARN_ON_ONCE(used > len))
1663 /* If recv_actor drops the lock (e.g. TCP splice
1664 * receive) the skb pointer might be invalid when
1665 * getting here: tcp_collapse might have deleted it
1666 * while aggregating skbs from the socket queue.
1668 skb = tcp_recv_skb(sk, seq - 1, &offset);
1671 /* TCP coalescing might have appended data to the skb.
1672 * Try to splice more frags
1674 if (offset + 1 != skb->len)
1677 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) {
1678 sk_eat_skb(sk, skb);
1682 sk_eat_skb(sk, skb);
1685 WRITE_ONCE(tp->copied_seq, seq);
1687 WRITE_ONCE(tp->copied_seq, seq);
1689 tcp_rcv_space_adjust(sk);
1691 /* Clean up data we have read: This will do ACK frames. */
1693 tcp_recv_skb(sk, seq, &offset);
1694 tcp_cleanup_rbuf(sk, copied);
1698 EXPORT_SYMBOL(tcp_read_sock);
1700 int tcp_peek_len(struct socket *sock)
1702 return tcp_inq(sock->sk);
1704 EXPORT_SYMBOL(tcp_peek_len);
1706 /* Make sure sk_rcvbuf is big enough to satisfy SO_RCVLOWAT hint */
1707 int tcp_set_rcvlowat(struct sock *sk, int val)
1711 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK)
1712 cap = sk->sk_rcvbuf >> 1;
1714 cap = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_rmem[2]) >> 1;
1715 val = min(val, cap);
1716 WRITE_ONCE(sk->sk_rcvlowat, val ? : 1);
1718 /* Check if we need to signal EPOLLIN right now */
1721 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK)
1725 if (val > sk->sk_rcvbuf) {
1726 WRITE_ONCE(sk->sk_rcvbuf, val);
1727 tcp_sk(sk)->window_clamp = tcp_win_from_space(sk, val);
1731 EXPORT_SYMBOL(tcp_set_rcvlowat);
1734 static const struct vm_operations_struct tcp_vm_ops = {
1737 int tcp_mmap(struct file *file, struct socket *sock,
1738 struct vm_area_struct *vma)
1740 if (vma->vm_flags & (VM_WRITE | VM_EXEC))
1742 vma->vm_flags &= ~(VM_MAYWRITE | VM_MAYEXEC);
1744 /* Instruct vm_insert_page() to not mmap_read_lock(mm) */
1745 vma->vm_flags |= VM_MIXEDMAP;
1747 vma->vm_ops = &tcp_vm_ops;
1750 EXPORT_SYMBOL(tcp_mmap);
1752 static skb_frag_t *skb_advance_to_frag(struct sk_buff *skb, u32 offset_skb,
1757 if (unlikely(offset_skb >= skb->len))
1760 offset_skb -= skb_headlen(skb);
1761 if ((int)offset_skb < 0 || skb_has_frag_list(skb))
1764 frag = skb_shinfo(skb)->frags;
1765 while (offset_skb) {
1766 if (skb_frag_size(frag) > offset_skb) {
1767 *offset_frag = offset_skb;
1770 offset_skb -= skb_frag_size(frag);
1777 static int tcp_copy_straggler_data(struct tcp_zerocopy_receive *zc,
1778 struct sk_buff *skb, u32 copylen,
1779 u32 *offset, u32 *seq)
1781 unsigned long copy_address = (unsigned long)zc->copybuf_address;
1782 struct msghdr msg = {};
1786 if (copy_address != zc->copybuf_address)
1789 err = import_single_range(READ, (void __user *)copy_address,
1790 copylen, &iov, &msg.msg_iter);
1793 err = skb_copy_datagram_msg(skb, *offset, &msg, copylen);
1796 zc->recv_skip_hint -= copylen;
1799 return (__s32)copylen;
1802 static int tcp_zerocopy_handle_leftover_data(struct tcp_zerocopy_receive *zc,
1804 struct sk_buff *skb,
1808 u32 offset, copylen = min_t(u32, copybuf_len, zc->recv_skip_hint);
1812 /* skb is null if inq < PAGE_SIZE. */
1814 offset = *seq - TCP_SKB_CB(skb)->seq;
1816 skb = tcp_recv_skb(sk, *seq, &offset);
1818 zc->copybuf_len = tcp_copy_straggler_data(zc, skb, copylen, &offset,
1820 return zc->copybuf_len < 0 ? 0 : copylen;
1823 static int tcp_zerocopy_vm_insert_batch(struct vm_area_struct *vma,
1824 struct page **pages,
1825 unsigned long pages_to_map,
1826 unsigned long *insert_addr,
1827 u32 *length_with_pending,
1829 struct tcp_zerocopy_receive *zc)
1831 unsigned long pages_remaining = pages_to_map;
1835 ret = vm_insert_pages(vma, *insert_addr, pages, &pages_remaining);
1836 bytes_mapped = PAGE_SIZE * (pages_to_map - pages_remaining);
1837 /* Even if vm_insert_pages fails, it may have partially succeeded in
1838 * mapping (some but not all of the pages).
1840 *seq += bytes_mapped;
1841 *insert_addr += bytes_mapped;
1843 /* But if vm_insert_pages did fail, we have to unroll some state
1844 * we speculatively touched before.
1846 const int bytes_not_mapped = PAGE_SIZE * pages_remaining;
1847 *length_with_pending -= bytes_not_mapped;
1848 zc->recv_skip_hint += bytes_not_mapped;
1853 static int tcp_zerocopy_receive(struct sock *sk,
1854 struct tcp_zerocopy_receive *zc)
1856 u32 length = 0, offset, vma_len, avail_len, aligned_len, copylen = 0;
1857 unsigned long address = (unsigned long)zc->address;
1858 s32 copybuf_len = zc->copybuf_len;
1859 struct tcp_sock *tp = tcp_sk(sk);
1860 #define PAGE_BATCH_SIZE 8
1861 struct page *pages[PAGE_BATCH_SIZE];
1862 const skb_frag_t *frags = NULL;
1863 struct vm_area_struct *vma;
1864 struct sk_buff *skb = NULL;
1865 unsigned long pg_idx = 0;
1866 unsigned long curr_addr;
1867 u32 seq = tp->copied_seq;
1868 int inq = tcp_inq(sk);
1871 zc->copybuf_len = 0;
1873 if (address & (PAGE_SIZE - 1) || address != zc->address)
1876 if (sk->sk_state == TCP_LISTEN)
1879 sock_rps_record_flow(sk);
1881 mmap_read_lock(current->mm);
1883 vma = find_vma(current->mm, address);
1884 if (!vma || vma->vm_start > address || vma->vm_ops != &tcp_vm_ops) {
1885 mmap_read_unlock(current->mm);
1888 vma_len = min_t(unsigned long, zc->length, vma->vm_end - address);
1889 avail_len = min_t(u32, vma_len, inq);
1890 aligned_len = avail_len & ~(PAGE_SIZE - 1);
1892 zap_page_range(vma, address, aligned_len);
1893 zc->length = aligned_len;
1894 zc->recv_skip_hint = 0;
1896 zc->length = avail_len;
1897 zc->recv_skip_hint = avail_len;
1900 curr_addr = address;
1901 while (length + PAGE_SIZE <= zc->length) {
1902 if (zc->recv_skip_hint < PAGE_SIZE) {
1905 /* If we're here, finish the current batch. */
1907 ret = tcp_zerocopy_vm_insert_batch(vma, pages,
1917 if (zc->recv_skip_hint > 0)
1920 offset = seq - TCP_SKB_CB(skb)->seq;
1922 skb = tcp_recv_skb(sk, seq, &offset);
1924 zc->recv_skip_hint = skb->len - offset;
1925 frags = skb_advance_to_frag(skb, offset, &offset_frag);
1926 if (!frags || offset_frag)
1929 if (skb_frag_size(frags) != PAGE_SIZE || skb_frag_off(frags)) {
1930 int remaining = zc->recv_skip_hint;
1932 while (remaining && (skb_frag_size(frags) != PAGE_SIZE ||
1933 skb_frag_off(frags))) {
1934 remaining -= skb_frag_size(frags);
1937 zc->recv_skip_hint -= remaining;
1940 pages[pg_idx] = skb_frag_page(frags);
1942 length += PAGE_SIZE;
1943 zc->recv_skip_hint -= PAGE_SIZE;
1945 if (pg_idx == PAGE_BATCH_SIZE) {
1946 ret = tcp_zerocopy_vm_insert_batch(vma, pages, pg_idx,
1947 &curr_addr, &length,
1955 ret = tcp_zerocopy_vm_insert_batch(vma, pages, pg_idx,
1956 &curr_addr, &length, &seq,
1960 mmap_read_unlock(current->mm);
1961 /* Try to copy straggler data. */
1963 copylen = tcp_zerocopy_handle_leftover_data(zc, sk, skb, &seq,
1966 if (length + copylen) {
1967 WRITE_ONCE(tp->copied_seq, seq);
1968 tcp_rcv_space_adjust(sk);
1970 /* Clean up data we have read: This will do ACK frames. */
1971 tcp_recv_skb(sk, seq, &offset);
1972 tcp_cleanup_rbuf(sk, length + copylen);
1974 if (length == zc->length)
1975 zc->recv_skip_hint = 0;
1977 if (!zc->recv_skip_hint && sock_flag(sk, SOCK_DONE))
1980 zc->length = length;
1985 static void tcp_update_recv_tstamps(struct sk_buff *skb,
1986 struct scm_timestamping_internal *tss)
1989 tss->ts[0] = ktime_to_timespec64(skb->tstamp);
1991 tss->ts[0] = (struct timespec64) {0};
1993 if (skb_hwtstamps(skb)->hwtstamp)
1994 tss->ts[2] = ktime_to_timespec64(skb_hwtstamps(skb)->hwtstamp);
1996 tss->ts[2] = (struct timespec64) {0};
1999 /* Similar to __sock_recv_timestamp, but does not require an skb */
2000 static void tcp_recv_timestamp(struct msghdr *msg, const struct sock *sk,
2001 struct scm_timestamping_internal *tss)
2003 int new_tstamp = sock_flag(sk, SOCK_TSTAMP_NEW);
2004 bool has_timestamping = false;
2006 if (tss->ts[0].tv_sec || tss->ts[0].tv_nsec) {
2007 if (sock_flag(sk, SOCK_RCVTSTAMP)) {
2008 if (sock_flag(sk, SOCK_RCVTSTAMPNS)) {
2010 struct __kernel_timespec kts = {
2011 .tv_sec = tss->ts[0].tv_sec,
2012 .tv_nsec = tss->ts[0].tv_nsec,
2014 put_cmsg(msg, SOL_SOCKET, SO_TIMESTAMPNS_NEW,
2017 struct __kernel_old_timespec ts_old = {
2018 .tv_sec = tss->ts[0].tv_sec,
2019 .tv_nsec = tss->ts[0].tv_nsec,
2021 put_cmsg(msg, SOL_SOCKET, SO_TIMESTAMPNS_OLD,
2022 sizeof(ts_old), &ts_old);
2026 struct __kernel_sock_timeval stv = {
2027 .tv_sec = tss->ts[0].tv_sec,
2028 .tv_usec = tss->ts[0].tv_nsec / 1000,
2030 put_cmsg(msg, SOL_SOCKET, SO_TIMESTAMP_NEW,
2033 struct __kernel_old_timeval tv = {
2034 .tv_sec = tss->ts[0].tv_sec,
2035 .tv_usec = tss->ts[0].tv_nsec / 1000,
2037 put_cmsg(msg, SOL_SOCKET, SO_TIMESTAMP_OLD,
2043 if (sk->sk_tsflags & SOF_TIMESTAMPING_SOFTWARE)
2044 has_timestamping = true;
2046 tss->ts[0] = (struct timespec64) {0};
2049 if (tss->ts[2].tv_sec || tss->ts[2].tv_nsec) {
2050 if (sk->sk_tsflags & SOF_TIMESTAMPING_RAW_HARDWARE)
2051 has_timestamping = true;
2053 tss->ts[2] = (struct timespec64) {0};
2056 if (has_timestamping) {
2057 tss->ts[1] = (struct timespec64) {0};
2058 if (sock_flag(sk, SOCK_TSTAMP_NEW))
2059 put_cmsg_scm_timestamping64(msg, tss);
2061 put_cmsg_scm_timestamping(msg, tss);
2065 static int tcp_inq_hint(struct sock *sk)
2067 const struct tcp_sock *tp = tcp_sk(sk);
2068 u32 copied_seq = READ_ONCE(tp->copied_seq);
2069 u32 rcv_nxt = READ_ONCE(tp->rcv_nxt);
2072 inq = rcv_nxt - copied_seq;
2073 if (unlikely(inq < 0 || copied_seq != READ_ONCE(tp->copied_seq))) {
2075 inq = tp->rcv_nxt - tp->copied_seq;
2078 /* After receiving a FIN, tell the user-space to continue reading
2079 * by returning a non-zero inq.
2081 if (inq == 0 && sock_flag(sk, SOCK_DONE))
2087 * This routine copies from a sock struct into the user buffer.
2089 * Technical note: in 2.3 we work on _locked_ socket, so that
2090 * tricks with *seq access order and skb->users are not required.
2091 * Probably, code can be easily improved even more.
2094 int tcp_recvmsg(struct sock *sk, struct msghdr *msg, size_t len, int nonblock,
2095 int flags, int *addr_len)
2097 struct tcp_sock *tp = tcp_sk(sk);
2103 int target; /* Read at least this many bytes */
2105 struct sk_buff *skb, *last;
2107 struct scm_timestamping_internal tss;
2110 if (unlikely(flags & MSG_ERRQUEUE))
2111 return inet_recv_error(sk, msg, len, addr_len);
2113 if (sk_can_busy_loop(sk) && skb_queue_empty_lockless(&sk->sk_receive_queue) &&
2114 (sk->sk_state == TCP_ESTABLISHED))
2115 sk_busy_loop(sk, nonblock);
2120 if (sk->sk_state == TCP_LISTEN)
2123 cmsg_flags = tp->recvmsg_inq ? 1 : 0;
2124 timeo = sock_rcvtimeo(sk, nonblock);
2126 /* Urgent data needs to be handled specially. */
2127 if (flags & MSG_OOB)
2130 if (unlikely(tp->repair)) {
2132 if (!(flags & MSG_PEEK))
2135 if (tp->repair_queue == TCP_SEND_QUEUE)
2139 if (tp->repair_queue == TCP_NO_QUEUE)
2142 /* 'common' recv queue MSG_PEEK-ing */
2145 seq = &tp->copied_seq;
2146 if (flags & MSG_PEEK) {
2147 peek_seq = tp->copied_seq;
2151 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
2156 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
2157 if (tp->urg_data && tp->urg_seq == *seq) {
2160 if (signal_pending(current)) {
2161 copied = timeo ? sock_intr_errno(timeo) : -EAGAIN;
2166 /* Next get a buffer. */
2168 last = skb_peek_tail(&sk->sk_receive_queue);
2169 skb_queue_walk(&sk->sk_receive_queue, skb) {
2171 /* Now that we have two receive queues this
2174 if (WARN(before(*seq, TCP_SKB_CB(skb)->seq),
2175 "TCP recvmsg seq # bug: copied %X, seq %X, rcvnxt %X, fl %X\n",
2176 *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt,
2180 offset = *seq - TCP_SKB_CB(skb)->seq;
2181 if (unlikely(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
2182 pr_err_once("%s: found a SYN, please report !\n", __func__);
2185 if (offset < skb->len)
2187 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
2189 WARN(!(flags & MSG_PEEK),
2190 "TCP recvmsg seq # bug 2: copied %X, seq %X, rcvnxt %X, fl %X\n",
2191 *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt, flags);
2194 /* Well, if we have backlog, try to process it now yet. */
2196 if (copied >= target && !READ_ONCE(sk->sk_backlog.tail))
2201 sk->sk_state == TCP_CLOSE ||
2202 (sk->sk_shutdown & RCV_SHUTDOWN) ||
2204 signal_pending(current))
2207 if (sock_flag(sk, SOCK_DONE))
2211 copied = sock_error(sk);
2215 if (sk->sk_shutdown & RCV_SHUTDOWN)
2218 if (sk->sk_state == TCP_CLOSE) {
2219 /* This occurs when user tries to read
2220 * from never connected socket.
2231 if (signal_pending(current)) {
2232 copied = sock_intr_errno(timeo);
2237 tcp_cleanup_rbuf(sk, copied);
2239 if (copied >= target) {
2240 /* Do not sleep, just process backlog. */
2244 sk_wait_data(sk, &timeo, last);
2247 if ((flags & MSG_PEEK) &&
2248 (peek_seq - copied - urg_hole != tp->copied_seq)) {
2249 net_dbg_ratelimited("TCP(%s:%d): Application bug, race in MSG_PEEK\n",
2251 task_pid_nr(current));
2252 peek_seq = tp->copied_seq;
2257 /* Ok so how much can we use? */
2258 used = skb->len - offset;
2262 /* Do we have urgent data here? */
2264 u32 urg_offset = tp->urg_seq - *seq;
2265 if (urg_offset < used) {
2267 if (!sock_flag(sk, SOCK_URGINLINE)) {
2268 WRITE_ONCE(*seq, *seq + 1);
2280 if (!(flags & MSG_TRUNC)) {
2281 err = skb_copy_datagram_msg(skb, offset, msg, used);
2283 /* Exception. Bailout! */
2290 WRITE_ONCE(*seq, *seq + used);
2294 tcp_rcv_space_adjust(sk);
2297 if (tp->urg_data && after(tp->copied_seq, tp->urg_seq)) {
2299 tcp_fast_path_check(sk);
2302 if (TCP_SKB_CB(skb)->has_rxtstamp) {
2303 tcp_update_recv_tstamps(skb, &tss);
2307 if (used + offset < skb->len)
2310 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
2312 if (!(flags & MSG_PEEK))
2313 sk_eat_skb(sk, skb);
2317 /* Process the FIN. */
2318 WRITE_ONCE(*seq, *seq + 1);
2319 if (!(flags & MSG_PEEK))
2320 sk_eat_skb(sk, skb);
2324 /* According to UNIX98, msg_name/msg_namelen are ignored
2325 * on connected socket. I was just happy when found this 8) --ANK
2328 /* Clean up data we have read: This will do ACK frames. */
2329 tcp_cleanup_rbuf(sk, copied);
2335 tcp_recv_timestamp(msg, sk, &tss);
2336 if (cmsg_flags & 1) {
2337 inq = tcp_inq_hint(sk);
2338 put_cmsg(msg, SOL_TCP, TCP_CM_INQ, sizeof(inq), &inq);
2349 err = tcp_recv_urg(sk, msg, len, flags);
2353 err = tcp_peek_sndq(sk, msg, len);
2356 EXPORT_SYMBOL(tcp_recvmsg);
2358 void tcp_set_state(struct sock *sk, int state)
2360 int oldstate = sk->sk_state;
2362 /* We defined a new enum for TCP states that are exported in BPF
2363 * so as not force the internal TCP states to be frozen. The
2364 * following checks will detect if an internal state value ever
2365 * differs from the BPF value. If this ever happens, then we will
2366 * need to remap the internal value to the BPF value before calling
2367 * tcp_call_bpf_2arg.
2369 BUILD_BUG_ON((int)BPF_TCP_ESTABLISHED != (int)TCP_ESTABLISHED);
2370 BUILD_BUG_ON((int)BPF_TCP_SYN_SENT != (int)TCP_SYN_SENT);
2371 BUILD_BUG_ON((int)BPF_TCP_SYN_RECV != (int)TCP_SYN_RECV);
2372 BUILD_BUG_ON((int)BPF_TCP_FIN_WAIT1 != (int)TCP_FIN_WAIT1);
2373 BUILD_BUG_ON((int)BPF_TCP_FIN_WAIT2 != (int)TCP_FIN_WAIT2);
2374 BUILD_BUG_ON((int)BPF_TCP_TIME_WAIT != (int)TCP_TIME_WAIT);
2375 BUILD_BUG_ON((int)BPF_TCP_CLOSE != (int)TCP_CLOSE);
2376 BUILD_BUG_ON((int)BPF_TCP_CLOSE_WAIT != (int)TCP_CLOSE_WAIT);
2377 BUILD_BUG_ON((int)BPF_TCP_LAST_ACK != (int)TCP_LAST_ACK);
2378 BUILD_BUG_ON((int)BPF_TCP_LISTEN != (int)TCP_LISTEN);
2379 BUILD_BUG_ON((int)BPF_TCP_CLOSING != (int)TCP_CLOSING);
2380 BUILD_BUG_ON((int)BPF_TCP_NEW_SYN_RECV != (int)TCP_NEW_SYN_RECV);
2381 BUILD_BUG_ON((int)BPF_TCP_MAX_STATES != (int)TCP_MAX_STATES);
2383 if (BPF_SOCK_OPS_TEST_FLAG(tcp_sk(sk), BPF_SOCK_OPS_STATE_CB_FLAG))
2384 tcp_call_bpf_2arg(sk, BPF_SOCK_OPS_STATE_CB, oldstate, state);
2387 case TCP_ESTABLISHED:
2388 if (oldstate != TCP_ESTABLISHED)
2389 TCP_INC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
2393 if (oldstate == TCP_CLOSE_WAIT || oldstate == TCP_ESTABLISHED)
2394 TCP_INC_STATS(sock_net(sk), TCP_MIB_ESTABRESETS);
2396 sk->sk_prot->unhash(sk);
2397 if (inet_csk(sk)->icsk_bind_hash &&
2398 !(sk->sk_userlocks & SOCK_BINDPORT_LOCK))
2402 if (oldstate == TCP_ESTABLISHED)
2403 TCP_DEC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
2406 /* Change state AFTER socket is unhashed to avoid closed
2407 * socket sitting in hash tables.
2409 inet_sk_state_store(sk, state);
2411 EXPORT_SYMBOL_GPL(tcp_set_state);
2414 * State processing on a close. This implements the state shift for
2415 * sending our FIN frame. Note that we only send a FIN for some
2416 * states. A shutdown() may have already sent the FIN, or we may be
2420 static const unsigned char new_state[16] = {
2421 /* current state: new state: action: */
2422 [0 /* (Invalid) */] = TCP_CLOSE,
2423 [TCP_ESTABLISHED] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2424 [TCP_SYN_SENT] = TCP_CLOSE,
2425 [TCP_SYN_RECV] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2426 [TCP_FIN_WAIT1] = TCP_FIN_WAIT1,
2427 [TCP_FIN_WAIT2] = TCP_FIN_WAIT2,
2428 [TCP_TIME_WAIT] = TCP_CLOSE,
2429 [TCP_CLOSE] = TCP_CLOSE,
2430 [TCP_CLOSE_WAIT] = TCP_LAST_ACK | TCP_ACTION_FIN,
2431 [TCP_LAST_ACK] = TCP_LAST_ACK,
2432 [TCP_LISTEN] = TCP_CLOSE,
2433 [TCP_CLOSING] = TCP_CLOSING,
2434 [TCP_NEW_SYN_RECV] = TCP_CLOSE, /* should not happen ! */
2437 static int tcp_close_state(struct sock *sk)
2439 int next = (int)new_state[sk->sk_state];
2440 int ns = next & TCP_STATE_MASK;
2442 tcp_set_state(sk, ns);
2444 return next & TCP_ACTION_FIN;
2448 * Shutdown the sending side of a connection. Much like close except
2449 * that we don't receive shut down or sock_set_flag(sk, SOCK_DEAD).
2452 void tcp_shutdown(struct sock *sk, int how)
2454 /* We need to grab some memory, and put together a FIN,
2455 * and then put it into the queue to be sent.
2456 * Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
2458 if (!(how & SEND_SHUTDOWN))
2461 /* If we've already sent a FIN, or it's a closed state, skip this. */
2462 if ((1 << sk->sk_state) &
2463 (TCPF_ESTABLISHED | TCPF_SYN_SENT |
2464 TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) {
2465 /* Clear out any half completed packets. FIN if needed. */
2466 if (tcp_close_state(sk))
2470 EXPORT_SYMBOL(tcp_shutdown);
2472 int tcp_orphan_count_sum(void)
2476 for_each_possible_cpu(i)
2477 total += per_cpu(tcp_orphan_count, i);
2479 return max(total, 0);
2482 static int tcp_orphan_cache;
2483 static struct timer_list tcp_orphan_timer;
2484 #define TCP_ORPHAN_TIMER_PERIOD msecs_to_jiffies(100)
2486 static void tcp_orphan_update(struct timer_list *unused)
2488 WRITE_ONCE(tcp_orphan_cache, tcp_orphan_count_sum());
2489 mod_timer(&tcp_orphan_timer, jiffies + TCP_ORPHAN_TIMER_PERIOD);
2492 static bool tcp_too_many_orphans(int shift)
2494 return READ_ONCE(tcp_orphan_cache) << shift >
2495 READ_ONCE(sysctl_tcp_max_orphans);
2498 bool tcp_check_oom(struct sock *sk, int shift)
2500 bool too_many_orphans, out_of_socket_memory;
2502 too_many_orphans = tcp_too_many_orphans(shift);
2503 out_of_socket_memory = tcp_out_of_memory(sk);
2505 if (too_many_orphans)
2506 net_info_ratelimited("too many orphaned sockets\n");
2507 if (out_of_socket_memory)
2508 net_info_ratelimited("out of memory -- consider tuning tcp_mem\n");
2509 return too_many_orphans || out_of_socket_memory;
2512 void tcp_close(struct sock *sk, long timeout)
2514 struct sk_buff *skb;
2515 int data_was_unread = 0;
2519 sk->sk_shutdown = SHUTDOWN_MASK;
2521 if (sk->sk_state == TCP_LISTEN) {
2522 tcp_set_state(sk, TCP_CLOSE);
2525 inet_csk_listen_stop(sk);
2527 goto adjudge_to_death;
2530 /* We need to flush the recv. buffs. We do this only on the
2531 * descriptor close, not protocol-sourced closes, because the
2532 * reader process may not have drained the data yet!
2534 while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) {
2535 u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq;
2537 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
2539 data_was_unread += len;
2545 /* If socket has been already reset (e.g. in tcp_reset()) - kill it. */
2546 if (sk->sk_state == TCP_CLOSE)
2547 goto adjudge_to_death;
2549 /* As outlined in RFC 2525, section 2.17, we send a RST here because
2550 * data was lost. To witness the awful effects of the old behavior of
2551 * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk
2552 * GET in an FTP client, suspend the process, wait for the client to
2553 * advertise a zero window, then kill -9 the FTP client, wheee...
2554 * Note: timeout is always zero in such a case.
2556 if (unlikely(tcp_sk(sk)->repair)) {
2557 sk->sk_prot->disconnect(sk, 0);
2558 } else if (data_was_unread) {
2559 /* Unread data was tossed, zap the connection. */
2560 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONCLOSE);
2561 tcp_set_state(sk, TCP_CLOSE);
2562 tcp_send_active_reset(sk, sk->sk_allocation);
2563 } else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) {
2564 /* Check zero linger _after_ checking for unread data. */
2565 sk->sk_prot->disconnect(sk, 0);
2566 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
2567 } else if (tcp_close_state(sk)) {
2568 /* We FIN if the application ate all the data before
2569 * zapping the connection.
2572 /* RED-PEN. Formally speaking, we have broken TCP state
2573 * machine. State transitions:
2575 * TCP_ESTABLISHED -> TCP_FIN_WAIT1
2576 * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible)
2577 * TCP_CLOSE_WAIT -> TCP_LAST_ACK
2579 * are legal only when FIN has been sent (i.e. in window),
2580 * rather than queued out of window. Purists blame.
2582 * F.e. "RFC state" is ESTABLISHED,
2583 * if Linux state is FIN-WAIT-1, but FIN is still not sent.
2585 * The visible declinations are that sometimes
2586 * we enter time-wait state, when it is not required really
2587 * (harmless), do not send active resets, when they are
2588 * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
2589 * they look as CLOSING or LAST_ACK for Linux)
2590 * Probably, I missed some more holelets.
2592 * XXX (TFO) - To start off we don't support SYN+ACK+FIN
2593 * in a single packet! (May consider it later but will
2594 * probably need API support or TCP_CORK SYN-ACK until
2595 * data is written and socket is closed.)
2600 sk_stream_wait_close(sk, timeout);
2603 state = sk->sk_state;
2609 /* remove backlog if any, without releasing ownership. */
2612 this_cpu_inc(tcp_orphan_count);
2614 /* Have we already been destroyed by a softirq or backlog? */
2615 if (state != TCP_CLOSE && sk->sk_state == TCP_CLOSE)
2618 /* This is a (useful) BSD violating of the RFC. There is a
2619 * problem with TCP as specified in that the other end could
2620 * keep a socket open forever with no application left this end.
2621 * We use a 1 minute timeout (about the same as BSD) then kill
2622 * our end. If they send after that then tough - BUT: long enough
2623 * that we won't make the old 4*rto = almost no time - whoops
2626 * Nope, it was not mistake. It is really desired behaviour
2627 * f.e. on http servers, when such sockets are useless, but
2628 * consume significant resources. Let's do it with special
2629 * linger2 option. --ANK
2632 if (sk->sk_state == TCP_FIN_WAIT2) {
2633 struct tcp_sock *tp = tcp_sk(sk);
2634 if (tp->linger2 < 0) {
2635 tcp_set_state(sk, TCP_CLOSE);
2636 tcp_send_active_reset(sk, GFP_ATOMIC);
2637 __NET_INC_STATS(sock_net(sk),
2638 LINUX_MIB_TCPABORTONLINGER);
2640 const int tmo = tcp_fin_time(sk);
2642 if (tmo > TCP_TIMEWAIT_LEN) {
2643 inet_csk_reset_keepalive_timer(sk,
2644 tmo - TCP_TIMEWAIT_LEN);
2646 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
2651 if (sk->sk_state != TCP_CLOSE) {
2653 if (tcp_check_oom(sk, 0)) {
2654 tcp_set_state(sk, TCP_CLOSE);
2655 tcp_send_active_reset(sk, GFP_ATOMIC);
2656 __NET_INC_STATS(sock_net(sk),
2657 LINUX_MIB_TCPABORTONMEMORY);
2658 } else if (!check_net(sock_net(sk))) {
2659 /* Not possible to send reset; just close */
2660 tcp_set_state(sk, TCP_CLOSE);
2664 if (sk->sk_state == TCP_CLOSE) {
2665 struct request_sock *req;
2667 req = rcu_dereference_protected(tcp_sk(sk)->fastopen_rsk,
2668 lockdep_sock_is_held(sk));
2669 /* We could get here with a non-NULL req if the socket is
2670 * aborted (e.g., closed with unread data) before 3WHS
2674 reqsk_fastopen_remove(sk, req, false);
2675 inet_csk_destroy_sock(sk);
2677 /* Otherwise, socket is reprieved until protocol close. */
2685 EXPORT_SYMBOL(tcp_close);
2687 /* These states need RST on ABORT according to RFC793 */
2689 static inline bool tcp_need_reset(int state)
2691 return (1 << state) &
2692 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 |
2693 TCPF_FIN_WAIT2 | TCPF_SYN_RECV);
2696 static void tcp_rtx_queue_purge(struct sock *sk)
2698 struct rb_node *p = rb_first(&sk->tcp_rtx_queue);
2700 tcp_sk(sk)->highest_sack = NULL;
2702 struct sk_buff *skb = rb_to_skb(p);
2705 /* Since we are deleting whole queue, no need to
2706 * list_del(&skb->tcp_tsorted_anchor)
2708 tcp_rtx_queue_unlink(skb, sk);
2709 sk_wmem_free_skb(sk, skb);
2713 void tcp_write_queue_purge(struct sock *sk)
2715 struct sk_buff *skb;
2717 tcp_chrono_stop(sk, TCP_CHRONO_BUSY);
2718 while ((skb = __skb_dequeue(&sk->sk_write_queue)) != NULL) {
2719 tcp_skb_tsorted_anchor_cleanup(skb);
2720 sk_wmem_free_skb(sk, skb);
2722 tcp_rtx_queue_purge(sk);
2723 skb = sk->sk_tx_skb_cache;
2726 sk->sk_tx_skb_cache = NULL;
2728 INIT_LIST_HEAD(&tcp_sk(sk)->tsorted_sent_queue);
2730 tcp_clear_all_retrans_hints(tcp_sk(sk));
2731 tcp_sk(sk)->packets_out = 0;
2732 inet_csk(sk)->icsk_backoff = 0;
2735 int tcp_disconnect(struct sock *sk, int flags)
2737 struct inet_sock *inet = inet_sk(sk);
2738 struct inet_connection_sock *icsk = inet_csk(sk);
2739 struct tcp_sock *tp = tcp_sk(sk);
2740 int old_state = sk->sk_state;
2743 if (old_state != TCP_CLOSE)
2744 tcp_set_state(sk, TCP_CLOSE);
2746 /* ABORT function of RFC793 */
2747 if (old_state == TCP_LISTEN) {
2748 inet_csk_listen_stop(sk);
2749 } else if (unlikely(tp->repair)) {
2750 sk->sk_err = ECONNABORTED;
2751 } else if (tcp_need_reset(old_state) ||
2752 (tp->snd_nxt != tp->write_seq &&
2753 (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK))) {
2754 /* The last check adjusts for discrepancy of Linux wrt. RFC
2757 tcp_send_active_reset(sk, gfp_any());
2758 sk->sk_err = ECONNRESET;
2759 } else if (old_state == TCP_SYN_SENT)
2760 sk->sk_err = ECONNRESET;
2762 tcp_clear_xmit_timers(sk);
2763 __skb_queue_purge(&sk->sk_receive_queue);
2764 if (sk->sk_rx_skb_cache) {
2765 __kfree_skb(sk->sk_rx_skb_cache);
2766 sk->sk_rx_skb_cache = NULL;
2768 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt);
2770 tcp_write_queue_purge(sk);
2771 tcp_fastopen_active_disable_ofo_check(sk);
2772 skb_rbtree_purge(&tp->out_of_order_queue);
2774 inet->inet_dport = 0;
2776 if (!(sk->sk_userlocks & SOCK_BINDADDR_LOCK))
2777 inet_reset_saddr(sk);
2779 sk->sk_shutdown = 0;
2780 sock_reset_flag(sk, SOCK_DONE);
2782 tp->mdev_us = jiffies_to_usecs(TCP_TIMEOUT_INIT);
2783 tp->rcv_rtt_last_tsecr = 0;
2785 seq = tp->write_seq + tp->max_window + 2;
2788 WRITE_ONCE(tp->write_seq, seq);
2790 icsk->icsk_backoff = 0;
2791 icsk->icsk_probes_out = 0;
2792 icsk->icsk_probes_tstamp = 0;
2793 icsk->icsk_rto = TCP_TIMEOUT_INIT;
2794 icsk->icsk_rto_min = TCP_RTO_MIN;
2795 icsk->icsk_delack_max = TCP_DELACK_MAX;
2796 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
2797 tp->snd_cwnd = TCP_INIT_CWND;
2798 tp->snd_cwnd_cnt = 0;
2799 tp->is_cwnd_limited = 0;
2800 tp->max_packets_out = 0;
2801 tp->window_clamp = 0;
2803 tp->delivered_ce = 0;
2804 if (icsk->icsk_ca_ops->release)
2805 icsk->icsk_ca_ops->release(sk);
2806 memset(icsk->icsk_ca_priv, 0, sizeof(icsk->icsk_ca_priv));
2807 icsk->icsk_ca_initialized = 0;
2808 tcp_set_ca_state(sk, TCP_CA_Open);
2809 tp->is_sack_reneg = 0;
2810 tcp_clear_retrans(tp);
2811 tp->total_retrans = 0;
2812 inet_csk_delack_init(sk);
2813 /* Initialize rcv_mss to TCP_MIN_MSS to avoid division by 0
2814 * issue in __tcp_select_window()
2816 icsk->icsk_ack.rcv_mss = TCP_MIN_MSS;
2817 memset(&tp->rx_opt, 0, sizeof(tp->rx_opt));
2819 dst_release(xchg((__force struct dst_entry **)&sk->sk_rx_dst, NULL));
2820 tcp_saved_syn_free(tp);
2821 tp->compressed_ack = 0;
2825 tp->bytes_acked = 0;
2826 tp->bytes_received = 0;
2827 tp->bytes_retrans = 0;
2828 tp->data_segs_in = 0;
2829 tp->data_segs_out = 0;
2830 tp->duplicate_sack[0].start_seq = 0;
2831 tp->duplicate_sack[0].end_seq = 0;
2834 tp->retrans_out = 0;
2836 tp->tlp_high_seq = 0;
2837 tp->last_oow_ack_time = 0;
2838 /* There's a bubble in the pipe until at least the first ACK. */
2839 tp->app_limited = ~0U;
2840 tp->rack.mstamp = 0;
2841 tp->rack.advanced = 0;
2842 tp->rack.reo_wnd_steps = 1;
2843 tp->rack.last_delivered = 0;
2844 tp->rack.reo_wnd_persist = 0;
2845 tp->rack.dsack_seen = 0;
2846 tp->syn_data_acked = 0;
2847 tp->rx_opt.saw_tstamp = 0;
2848 tp->rx_opt.dsack = 0;
2849 tp->rx_opt.num_sacks = 0;
2850 tp->rcv_ooopack = 0;
2853 /* Clean up fastopen related fields */
2854 tcp_free_fastopen_req(tp);
2855 inet->defer_connect = 0;
2856 tp->fastopen_client_fail = 0;
2858 WARN_ON(inet->inet_num && !icsk->icsk_bind_hash);
2860 if (sk->sk_frag.page) {
2861 put_page(sk->sk_frag.page);
2862 sk->sk_frag.page = NULL;
2863 sk->sk_frag.offset = 0;
2866 sk->sk_error_report(sk);
2869 EXPORT_SYMBOL(tcp_disconnect);
2871 static inline bool tcp_can_repair_sock(const struct sock *sk)
2873 return ns_capable(sock_net(sk)->user_ns, CAP_NET_ADMIN) &&
2874 (sk->sk_state != TCP_LISTEN);
2877 static int tcp_repair_set_window(struct tcp_sock *tp, sockptr_t optbuf, int len)
2879 struct tcp_repair_window opt;
2884 if (len != sizeof(opt))
2887 if (copy_from_sockptr(&opt, optbuf, sizeof(opt)))
2890 if (opt.max_window < opt.snd_wnd)
2893 if (after(opt.snd_wl1, tp->rcv_nxt + opt.rcv_wnd))
2896 if (after(opt.rcv_wup, tp->rcv_nxt))
2899 tp->snd_wl1 = opt.snd_wl1;
2900 tp->snd_wnd = opt.snd_wnd;
2901 tp->max_window = opt.max_window;
2903 tp->rcv_wnd = opt.rcv_wnd;
2904 tp->rcv_wup = opt.rcv_wup;
2909 static int tcp_repair_options_est(struct sock *sk, sockptr_t optbuf,
2912 struct tcp_sock *tp = tcp_sk(sk);
2913 struct tcp_repair_opt opt;
2916 while (len >= sizeof(opt)) {
2917 if (copy_from_sockptr_offset(&opt, optbuf, offset, sizeof(opt)))
2920 offset += sizeof(opt);
2923 switch (opt.opt_code) {
2925 tp->rx_opt.mss_clamp = opt.opt_val;
2930 u16 snd_wscale = opt.opt_val & 0xFFFF;
2931 u16 rcv_wscale = opt.opt_val >> 16;
2933 if (snd_wscale > TCP_MAX_WSCALE || rcv_wscale > TCP_MAX_WSCALE)
2936 tp->rx_opt.snd_wscale = snd_wscale;
2937 tp->rx_opt.rcv_wscale = rcv_wscale;
2938 tp->rx_opt.wscale_ok = 1;
2941 case TCPOPT_SACK_PERM:
2942 if (opt.opt_val != 0)
2945 tp->rx_opt.sack_ok |= TCP_SACK_SEEN;
2947 case TCPOPT_TIMESTAMP:
2948 if (opt.opt_val != 0)
2951 tp->rx_opt.tstamp_ok = 1;
2959 DEFINE_STATIC_KEY_FALSE(tcp_tx_delay_enabled);
2960 EXPORT_SYMBOL(tcp_tx_delay_enabled);
2962 static void tcp_enable_tx_delay(void)
2964 if (!static_branch_unlikely(&tcp_tx_delay_enabled)) {
2965 static int __tcp_tx_delay_enabled = 0;
2967 if (cmpxchg(&__tcp_tx_delay_enabled, 0, 1) == 0) {
2968 static_branch_enable(&tcp_tx_delay_enabled);
2969 pr_info("TCP_TX_DELAY enabled\n");
2974 /* When set indicates to always queue non-full frames. Later the user clears
2975 * this option and we transmit any pending partial frames in the queue. This is
2976 * meant to be used alongside sendfile() to get properly filled frames when the
2977 * user (for example) must write out headers with a write() call first and then
2978 * use sendfile to send out the data parts.
2980 * TCP_CORK can be set together with TCP_NODELAY and it is stronger than
2983 static void __tcp_sock_set_cork(struct sock *sk, bool on)
2985 struct tcp_sock *tp = tcp_sk(sk);
2988 tp->nonagle |= TCP_NAGLE_CORK;
2990 tp->nonagle &= ~TCP_NAGLE_CORK;
2991 if (tp->nonagle & TCP_NAGLE_OFF)
2992 tp->nonagle |= TCP_NAGLE_PUSH;
2993 tcp_push_pending_frames(sk);
2997 void tcp_sock_set_cork(struct sock *sk, bool on)
3000 __tcp_sock_set_cork(sk, on);
3003 EXPORT_SYMBOL(tcp_sock_set_cork);
3005 /* TCP_NODELAY is weaker than TCP_CORK, so that this option on corked socket is
3006 * remembered, but it is not activated until cork is cleared.
3008 * However, when TCP_NODELAY is set we make an explicit push, which overrides
3009 * even TCP_CORK for currently queued segments.
3011 static void __tcp_sock_set_nodelay(struct sock *sk, bool on)
3014 tcp_sk(sk)->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH;
3015 tcp_push_pending_frames(sk);
3017 tcp_sk(sk)->nonagle &= ~TCP_NAGLE_OFF;
3021 void tcp_sock_set_nodelay(struct sock *sk)
3024 __tcp_sock_set_nodelay(sk, true);
3027 EXPORT_SYMBOL(tcp_sock_set_nodelay);
3029 static void __tcp_sock_set_quickack(struct sock *sk, int val)
3032 inet_csk_enter_pingpong_mode(sk);
3036 inet_csk_exit_pingpong_mode(sk);
3037 if ((1 << sk->sk_state) & (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) &&
3038 inet_csk_ack_scheduled(sk)) {
3039 inet_csk(sk)->icsk_ack.pending |= ICSK_ACK_PUSHED;
3040 tcp_cleanup_rbuf(sk, 1);
3042 inet_csk_enter_pingpong_mode(sk);
3046 void tcp_sock_set_quickack(struct sock *sk, int val)
3049 __tcp_sock_set_quickack(sk, val);
3052 EXPORT_SYMBOL(tcp_sock_set_quickack);
3054 int tcp_sock_set_syncnt(struct sock *sk, int val)
3056 if (val < 1 || val > MAX_TCP_SYNCNT)
3060 inet_csk(sk)->icsk_syn_retries = val;
3064 EXPORT_SYMBOL(tcp_sock_set_syncnt);
3066 void tcp_sock_set_user_timeout(struct sock *sk, u32 val)
3069 inet_csk(sk)->icsk_user_timeout = val;
3072 EXPORT_SYMBOL(tcp_sock_set_user_timeout);
3074 int tcp_sock_set_keepidle_locked(struct sock *sk, int val)
3076 struct tcp_sock *tp = tcp_sk(sk);
3078 if (val < 1 || val > MAX_TCP_KEEPIDLE)
3081 tp->keepalive_time = val * HZ;
3082 if (sock_flag(sk, SOCK_KEEPOPEN) &&
3083 !((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN))) {
3084 u32 elapsed = keepalive_time_elapsed(tp);
3086 if (tp->keepalive_time > elapsed)
3087 elapsed = tp->keepalive_time - elapsed;
3090 inet_csk_reset_keepalive_timer(sk, elapsed);
3096 int tcp_sock_set_keepidle(struct sock *sk, int val)
3101 err = tcp_sock_set_keepidle_locked(sk, val);
3105 EXPORT_SYMBOL(tcp_sock_set_keepidle);
3107 int tcp_sock_set_keepintvl(struct sock *sk, int val)
3109 if (val < 1 || val > MAX_TCP_KEEPINTVL)
3113 tcp_sk(sk)->keepalive_intvl = val * HZ;
3117 EXPORT_SYMBOL(tcp_sock_set_keepintvl);
3119 int tcp_sock_set_keepcnt(struct sock *sk, int val)
3121 if (val < 1 || val > MAX_TCP_KEEPCNT)
3125 tcp_sk(sk)->keepalive_probes = val;
3129 EXPORT_SYMBOL(tcp_sock_set_keepcnt);
3132 * Socket option code for TCP.
3134 static int do_tcp_setsockopt(struct sock *sk, int level, int optname,
3135 sockptr_t optval, unsigned int optlen)
3137 struct tcp_sock *tp = tcp_sk(sk);
3138 struct inet_connection_sock *icsk = inet_csk(sk);
3139 struct net *net = sock_net(sk);
3143 /* These are data/string values, all the others are ints */
3145 case TCP_CONGESTION: {
3146 char name[TCP_CA_NAME_MAX];
3151 val = strncpy_from_sockptr(name, optval,
3152 min_t(long, TCP_CA_NAME_MAX-1, optlen));
3158 err = tcp_set_congestion_control(sk, name, true,
3159 ns_capable(sock_net(sk)->user_ns,
3165 char name[TCP_ULP_NAME_MAX];
3170 val = strncpy_from_sockptr(name, optval,
3171 min_t(long, TCP_ULP_NAME_MAX - 1,
3178 err = tcp_set_ulp(sk, name);
3182 case TCP_FASTOPEN_KEY: {
3183 __u8 key[TCP_FASTOPEN_KEY_BUF_LENGTH];
3184 __u8 *backup_key = NULL;
3186 /* Allow a backup key as well to facilitate key rotation
3187 * First key is the active one.
3189 if (optlen != TCP_FASTOPEN_KEY_LENGTH &&
3190 optlen != TCP_FASTOPEN_KEY_BUF_LENGTH)
3193 if (copy_from_sockptr(key, optval, optlen))
3196 if (optlen == TCP_FASTOPEN_KEY_BUF_LENGTH)
3197 backup_key = key + TCP_FASTOPEN_KEY_LENGTH;
3199 return tcp_fastopen_reset_cipher(net, sk, key, backup_key);
3206 if (optlen < sizeof(int))
3209 if (copy_from_sockptr(&val, optval, sizeof(val)))
3216 /* Values greater than interface MTU won't take effect. However
3217 * at the point when this call is done we typically don't yet
3218 * know which interface is going to be used
3220 if (val && (val < TCP_MIN_MSS || val > MAX_TCP_WINDOW)) {
3224 tp->rx_opt.user_mss = val;
3228 __tcp_sock_set_nodelay(sk, val);
3231 case TCP_THIN_LINEAR_TIMEOUTS:
3232 if (val < 0 || val > 1)
3238 case TCP_THIN_DUPACK:
3239 if (val < 0 || val > 1)
3244 if (!tcp_can_repair_sock(sk))
3246 else if (val == TCP_REPAIR_ON) {
3248 sk->sk_reuse = SK_FORCE_REUSE;
3249 tp->repair_queue = TCP_NO_QUEUE;
3250 } else if (val == TCP_REPAIR_OFF) {
3252 sk->sk_reuse = SK_NO_REUSE;
3253 tcp_send_window_probe(sk);
3254 } else if (val == TCP_REPAIR_OFF_NO_WP) {
3256 sk->sk_reuse = SK_NO_REUSE;
3262 case TCP_REPAIR_QUEUE:
3265 else if ((unsigned int)val < TCP_QUEUES_NR)
3266 tp->repair_queue = val;
3272 if (sk->sk_state != TCP_CLOSE) {
3274 } else if (tp->repair_queue == TCP_SEND_QUEUE) {
3275 if (!tcp_rtx_queue_empty(sk))
3278 WRITE_ONCE(tp->write_seq, val);
3279 } else if (tp->repair_queue == TCP_RECV_QUEUE) {
3280 if (tp->rcv_nxt != tp->copied_seq) {
3283 WRITE_ONCE(tp->rcv_nxt, val);
3284 WRITE_ONCE(tp->copied_seq, val);
3291 case TCP_REPAIR_OPTIONS:
3294 else if (sk->sk_state == TCP_ESTABLISHED)
3295 err = tcp_repair_options_est(sk, optval, optlen);
3301 __tcp_sock_set_cork(sk, val);
3305 err = tcp_sock_set_keepidle_locked(sk, val);
3308 if (val < 1 || val > MAX_TCP_KEEPINTVL)
3311 tp->keepalive_intvl = val * HZ;
3314 if (val < 1 || val > MAX_TCP_KEEPCNT)
3317 tp->keepalive_probes = val;
3320 if (val < 1 || val > MAX_TCP_SYNCNT)
3323 icsk->icsk_syn_retries = val;
3327 /* 0: disable, 1: enable, 2: start from ether_header */
3328 if (val < 0 || val > 2)
3337 else if (val > TCP_FIN_TIMEOUT_MAX / HZ)
3338 tp->linger2 = TCP_FIN_TIMEOUT_MAX;
3340 tp->linger2 = val * HZ;
3343 case TCP_DEFER_ACCEPT:
3344 /* Translate value in seconds to number of retransmits */
3345 icsk->icsk_accept_queue.rskq_defer_accept =
3346 secs_to_retrans(val, TCP_TIMEOUT_INIT / HZ,
3350 case TCP_WINDOW_CLAMP:
3352 if (sk->sk_state != TCP_CLOSE) {
3356 tp->window_clamp = 0;
3358 tp->window_clamp = val < SOCK_MIN_RCVBUF / 2 ?
3359 SOCK_MIN_RCVBUF / 2 : val;
3363 __tcp_sock_set_quickack(sk, val);
3366 #ifdef CONFIG_TCP_MD5SIG
3368 case TCP_MD5SIG_EXT:
3369 err = tp->af_specific->md5_parse(sk, optname, optval, optlen);
3372 case TCP_USER_TIMEOUT:
3373 /* Cap the max time in ms TCP will retry or probe the window
3374 * before giving up and aborting (ETIMEDOUT) a connection.
3379 icsk->icsk_user_timeout = val;
3383 if (val >= 0 && ((1 << sk->sk_state) & (TCPF_CLOSE |
3385 tcp_fastopen_init_key_once(net);
3387 fastopen_queue_tune(sk, val);
3392 case TCP_FASTOPEN_CONNECT:
3393 if (val > 1 || val < 0) {
3395 } else if (READ_ONCE(net->ipv4.sysctl_tcp_fastopen) &
3396 TFO_CLIENT_ENABLE) {
3397 if (sk->sk_state == TCP_CLOSE)
3398 tp->fastopen_connect = val;
3405 case TCP_FASTOPEN_NO_COOKIE:
3406 if (val > 1 || val < 0)
3408 else if (!((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
3411 tp->fastopen_no_cookie = val;
3417 tp->tsoffset = val - tcp_time_stamp_raw();
3419 case TCP_REPAIR_WINDOW:
3420 err = tcp_repair_set_window(tp, optval, optlen);
3422 case TCP_NOTSENT_LOWAT:
3423 tp->notsent_lowat = val;
3424 sk->sk_write_space(sk);
3427 if (val > 1 || val < 0)
3430 tp->recvmsg_inq = val;
3434 tcp_enable_tx_delay();
3435 tp->tcp_tx_delay = val;
3446 int tcp_setsockopt(struct sock *sk, int level, int optname, sockptr_t optval,
3447 unsigned int optlen)
3449 const struct inet_connection_sock *icsk = inet_csk(sk);
3451 if (level != SOL_TCP)
3452 return icsk->icsk_af_ops->setsockopt(sk, level, optname,
3454 return do_tcp_setsockopt(sk, level, optname, optval, optlen);
3456 EXPORT_SYMBOL(tcp_setsockopt);
3458 static void tcp_get_info_chrono_stats(const struct tcp_sock *tp,
3459 struct tcp_info *info)
3461 u64 stats[__TCP_CHRONO_MAX], total = 0;
3464 for (i = TCP_CHRONO_BUSY; i < __TCP_CHRONO_MAX; ++i) {
3465 stats[i] = tp->chrono_stat[i - 1];
3466 if (i == tp->chrono_type)
3467 stats[i] += tcp_jiffies32 - tp->chrono_start;
3468 stats[i] *= USEC_PER_SEC / HZ;
3472 info->tcpi_busy_time = total;
3473 info->tcpi_rwnd_limited = stats[TCP_CHRONO_RWND_LIMITED];
3474 info->tcpi_sndbuf_limited = stats[TCP_CHRONO_SNDBUF_LIMITED];
3477 /* Return information about state of tcp endpoint in API format. */
3478 void tcp_get_info(struct sock *sk, struct tcp_info *info)
3480 const struct tcp_sock *tp = tcp_sk(sk); /* iff sk_type == SOCK_STREAM */
3481 const struct inet_connection_sock *icsk = inet_csk(sk);
3487 memset(info, 0, sizeof(*info));
3488 if (sk->sk_type != SOCK_STREAM)
3491 info->tcpi_state = inet_sk_state_load(sk);
3493 /* Report meaningful fields for all TCP states, including listeners */
3494 rate = READ_ONCE(sk->sk_pacing_rate);
3495 rate64 = (rate != ~0UL) ? rate : ~0ULL;
3496 info->tcpi_pacing_rate = rate64;
3498 rate = READ_ONCE(sk->sk_max_pacing_rate);
3499 rate64 = (rate != ~0UL) ? rate : ~0ULL;
3500 info->tcpi_max_pacing_rate = rate64;
3502 info->tcpi_reordering = tp->reordering;
3503 info->tcpi_snd_cwnd = tp->snd_cwnd;
3505 if (info->tcpi_state == TCP_LISTEN) {
3506 /* listeners aliased fields :
3507 * tcpi_unacked -> Number of children ready for accept()
3508 * tcpi_sacked -> max backlog
3510 info->tcpi_unacked = READ_ONCE(sk->sk_ack_backlog);
3511 info->tcpi_sacked = READ_ONCE(sk->sk_max_ack_backlog);
3515 slow = lock_sock_fast(sk);
3517 info->tcpi_ca_state = icsk->icsk_ca_state;
3518 info->tcpi_retransmits = icsk->icsk_retransmits;
3519 info->tcpi_probes = icsk->icsk_probes_out;
3520 info->tcpi_backoff = icsk->icsk_backoff;
3522 if (tp->rx_opt.tstamp_ok)
3523 info->tcpi_options |= TCPI_OPT_TIMESTAMPS;
3524 if (tcp_is_sack(tp))
3525 info->tcpi_options |= TCPI_OPT_SACK;
3526 if (tp->rx_opt.wscale_ok) {
3527 info->tcpi_options |= TCPI_OPT_WSCALE;
3528 info->tcpi_snd_wscale = tp->rx_opt.snd_wscale;
3529 info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale;
3532 if (tp->ecn_flags & TCP_ECN_OK)
3533 info->tcpi_options |= TCPI_OPT_ECN;
3534 if (tp->ecn_flags & TCP_ECN_SEEN)
3535 info->tcpi_options |= TCPI_OPT_ECN_SEEN;
3536 if (tp->syn_data_acked)
3537 info->tcpi_options |= TCPI_OPT_SYN_DATA;
3539 info->tcpi_rto = jiffies_to_usecs(icsk->icsk_rto);
3540 info->tcpi_ato = jiffies_to_usecs(icsk->icsk_ack.ato);
3541 info->tcpi_snd_mss = tp->mss_cache;
3542 info->tcpi_rcv_mss = icsk->icsk_ack.rcv_mss;
3544 info->tcpi_unacked = tp->packets_out;
3545 info->tcpi_sacked = tp->sacked_out;
3547 info->tcpi_lost = tp->lost_out;
3548 info->tcpi_retrans = tp->retrans_out;
3550 now = tcp_jiffies32;
3551 info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime);
3552 info->tcpi_last_data_recv = jiffies_to_msecs(now - icsk->icsk_ack.lrcvtime);
3553 info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp);
3555 info->tcpi_pmtu = icsk->icsk_pmtu_cookie;
3556 info->tcpi_rcv_ssthresh = tp->rcv_ssthresh;
3557 info->tcpi_rtt = tp->srtt_us >> 3;
3558 info->tcpi_rttvar = tp->mdev_us >> 2;
3559 info->tcpi_snd_ssthresh = tp->snd_ssthresh;
3560 info->tcpi_advmss = tp->advmss;
3562 info->tcpi_rcv_rtt = tp->rcv_rtt_est.rtt_us >> 3;
3563 info->tcpi_rcv_space = tp->rcvq_space.space;
3565 info->tcpi_total_retrans = tp->total_retrans;
3567 info->tcpi_bytes_acked = tp->bytes_acked;
3568 info->tcpi_bytes_received = tp->bytes_received;
3569 info->tcpi_notsent_bytes = max_t(int, 0, tp->write_seq - tp->snd_nxt);
3570 tcp_get_info_chrono_stats(tp, info);
3572 info->tcpi_segs_out = tp->segs_out;
3573 info->tcpi_segs_in = tp->segs_in;
3575 info->tcpi_min_rtt = tcp_min_rtt(tp);
3576 info->tcpi_data_segs_in = tp->data_segs_in;
3577 info->tcpi_data_segs_out = tp->data_segs_out;
3579 info->tcpi_delivery_rate_app_limited = tp->rate_app_limited ? 1 : 0;
3580 rate64 = tcp_compute_delivery_rate(tp);
3582 info->tcpi_delivery_rate = rate64;
3583 info->tcpi_delivered = tp->delivered;
3584 info->tcpi_delivered_ce = tp->delivered_ce;
3585 info->tcpi_bytes_sent = tp->bytes_sent;
3586 info->tcpi_bytes_retrans = tp->bytes_retrans;
3587 info->tcpi_dsack_dups = tp->dsack_dups;
3588 info->tcpi_reord_seen = tp->reord_seen;
3589 info->tcpi_rcv_ooopack = tp->rcv_ooopack;
3590 info->tcpi_snd_wnd = tp->snd_wnd;
3591 info->tcpi_fastopen_client_fail = tp->fastopen_client_fail;
3592 unlock_sock_fast(sk, slow);
3594 EXPORT_SYMBOL_GPL(tcp_get_info);
3596 static size_t tcp_opt_stats_get_size(void)
3599 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_BUSY */
3600 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_RWND_LIMITED */
3601 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_SNDBUF_LIMITED */
3602 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_DATA_SEGS_OUT */
3603 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_TOTAL_RETRANS */
3604 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_PACING_RATE */
3605 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_DELIVERY_RATE */
3606 nla_total_size(sizeof(u32)) + /* TCP_NLA_SND_CWND */
3607 nla_total_size(sizeof(u32)) + /* TCP_NLA_REORDERING */
3608 nla_total_size(sizeof(u32)) + /* TCP_NLA_MIN_RTT */
3609 nla_total_size(sizeof(u8)) + /* TCP_NLA_RECUR_RETRANS */
3610 nla_total_size(sizeof(u8)) + /* TCP_NLA_DELIVERY_RATE_APP_LMT */
3611 nla_total_size(sizeof(u32)) + /* TCP_NLA_SNDQ_SIZE */
3612 nla_total_size(sizeof(u8)) + /* TCP_NLA_CA_STATE */
3613 nla_total_size(sizeof(u32)) + /* TCP_NLA_SND_SSTHRESH */
3614 nla_total_size(sizeof(u32)) + /* TCP_NLA_DELIVERED */
3615 nla_total_size(sizeof(u32)) + /* TCP_NLA_DELIVERED_CE */
3616 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_BYTES_SENT */
3617 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_BYTES_RETRANS */
3618 nla_total_size(sizeof(u32)) + /* TCP_NLA_DSACK_DUPS */
3619 nla_total_size(sizeof(u32)) + /* TCP_NLA_REORD_SEEN */
3620 nla_total_size(sizeof(u32)) + /* TCP_NLA_SRTT */
3621 nla_total_size(sizeof(u16)) + /* TCP_NLA_TIMEOUT_REHASH */
3622 nla_total_size(sizeof(u32)) + /* TCP_NLA_BYTES_NOTSENT */
3623 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_EDT */
3627 struct sk_buff *tcp_get_timestamping_opt_stats(const struct sock *sk,
3628 const struct sk_buff *orig_skb)
3630 const struct tcp_sock *tp = tcp_sk(sk);
3631 struct sk_buff *stats;
3632 struct tcp_info info;
3636 stats = alloc_skb(tcp_opt_stats_get_size(), GFP_ATOMIC);
3640 tcp_get_info_chrono_stats(tp, &info);
3641 nla_put_u64_64bit(stats, TCP_NLA_BUSY,
3642 info.tcpi_busy_time, TCP_NLA_PAD);
3643 nla_put_u64_64bit(stats, TCP_NLA_RWND_LIMITED,
3644 info.tcpi_rwnd_limited, TCP_NLA_PAD);
3645 nla_put_u64_64bit(stats, TCP_NLA_SNDBUF_LIMITED,
3646 info.tcpi_sndbuf_limited, TCP_NLA_PAD);
3647 nla_put_u64_64bit(stats, TCP_NLA_DATA_SEGS_OUT,
3648 tp->data_segs_out, TCP_NLA_PAD);
3649 nla_put_u64_64bit(stats, TCP_NLA_TOTAL_RETRANS,
3650 tp->total_retrans, TCP_NLA_PAD);
3652 rate = READ_ONCE(sk->sk_pacing_rate);
3653 rate64 = (rate != ~0UL) ? rate : ~0ULL;
3654 nla_put_u64_64bit(stats, TCP_NLA_PACING_RATE, rate64, TCP_NLA_PAD);
3656 rate64 = tcp_compute_delivery_rate(tp);
3657 nla_put_u64_64bit(stats, TCP_NLA_DELIVERY_RATE, rate64, TCP_NLA_PAD);
3659 nla_put_u32(stats, TCP_NLA_SND_CWND, tp->snd_cwnd);
3660 nla_put_u32(stats, TCP_NLA_REORDERING, tp->reordering);
3661 nla_put_u32(stats, TCP_NLA_MIN_RTT, tcp_min_rtt(tp));
3663 nla_put_u8(stats, TCP_NLA_RECUR_RETRANS, inet_csk(sk)->icsk_retransmits);
3664 nla_put_u8(stats, TCP_NLA_DELIVERY_RATE_APP_LMT, !!tp->rate_app_limited);
3665 nla_put_u32(stats, TCP_NLA_SND_SSTHRESH, tp->snd_ssthresh);
3666 nla_put_u32(stats, TCP_NLA_DELIVERED, tp->delivered);
3667 nla_put_u32(stats, TCP_NLA_DELIVERED_CE, tp->delivered_ce);
3669 nla_put_u32(stats, TCP_NLA_SNDQ_SIZE, tp->write_seq - tp->snd_una);
3670 nla_put_u8(stats, TCP_NLA_CA_STATE, inet_csk(sk)->icsk_ca_state);
3672 nla_put_u64_64bit(stats, TCP_NLA_BYTES_SENT, tp->bytes_sent,
3674 nla_put_u64_64bit(stats, TCP_NLA_BYTES_RETRANS, tp->bytes_retrans,
3676 nla_put_u32(stats, TCP_NLA_DSACK_DUPS, tp->dsack_dups);
3677 nla_put_u32(stats, TCP_NLA_REORD_SEEN, tp->reord_seen);
3678 nla_put_u32(stats, TCP_NLA_SRTT, tp->srtt_us >> 3);
3679 nla_put_u16(stats, TCP_NLA_TIMEOUT_REHASH, tp->timeout_rehash);
3680 nla_put_u32(stats, TCP_NLA_BYTES_NOTSENT,
3681 max_t(int, 0, tp->write_seq - tp->snd_nxt));
3682 nla_put_u64_64bit(stats, TCP_NLA_EDT, orig_skb->skb_mstamp_ns,
3688 static int do_tcp_getsockopt(struct sock *sk, int level,
3689 int optname, char __user *optval, int __user *optlen)
3691 struct inet_connection_sock *icsk = inet_csk(sk);
3692 struct tcp_sock *tp = tcp_sk(sk);
3693 struct net *net = sock_net(sk);
3696 if (get_user(len, optlen))
3699 len = min_t(unsigned int, len, sizeof(int));
3706 val = tp->mss_cache;
3707 if (!val && ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
3708 val = tp->rx_opt.user_mss;
3710 val = tp->rx_opt.mss_clamp;
3713 val = !!(tp->nonagle&TCP_NAGLE_OFF);
3716 val = !!(tp->nonagle&TCP_NAGLE_CORK);
3719 val = keepalive_time_when(tp) / HZ;
3722 val = keepalive_intvl_when(tp) / HZ;
3725 val = keepalive_probes(tp);
3728 val = icsk->icsk_syn_retries ? : net->ipv4.sysctl_tcp_syn_retries;
3733 val = (val ? : READ_ONCE(net->ipv4.sysctl_tcp_fin_timeout)) / HZ;
3735 case TCP_DEFER_ACCEPT:
3736 val = retrans_to_secs(icsk->icsk_accept_queue.rskq_defer_accept,
3737 TCP_TIMEOUT_INIT / HZ, TCP_RTO_MAX / HZ);
3739 case TCP_WINDOW_CLAMP:
3740 val = tp->window_clamp;
3743 struct tcp_info info;
3745 if (get_user(len, optlen))
3748 tcp_get_info(sk, &info);
3750 len = min_t(unsigned int, len, sizeof(info));
3751 if (put_user(len, optlen))
3753 if (copy_to_user(optval, &info, len))
3758 const struct tcp_congestion_ops *ca_ops;
3759 union tcp_cc_info info;
3763 if (get_user(len, optlen))
3766 ca_ops = icsk->icsk_ca_ops;
3767 if (ca_ops && ca_ops->get_info)
3768 sz = ca_ops->get_info(sk, ~0U, &attr, &info);
3770 len = min_t(unsigned int, len, sz);
3771 if (put_user(len, optlen))
3773 if (copy_to_user(optval, &info, len))
3778 val = !inet_csk_in_pingpong_mode(sk);
3781 case TCP_CONGESTION:
3782 if (get_user(len, optlen))
3784 len = min_t(unsigned int, len, TCP_CA_NAME_MAX);
3785 if (put_user(len, optlen))
3787 if (copy_to_user(optval, icsk->icsk_ca_ops->name, len))
3792 if (get_user(len, optlen))
3794 len = min_t(unsigned int, len, TCP_ULP_NAME_MAX);
3795 if (!icsk->icsk_ulp_ops) {
3796 if (put_user(0, optlen))
3800 if (put_user(len, optlen))
3802 if (copy_to_user(optval, icsk->icsk_ulp_ops->name, len))
3806 case TCP_FASTOPEN_KEY: {
3807 u64 key[TCP_FASTOPEN_KEY_BUF_LENGTH / sizeof(u64)];
3808 unsigned int key_len;
3810 if (get_user(len, optlen))
3813 key_len = tcp_fastopen_get_cipher(net, icsk, key) *
3814 TCP_FASTOPEN_KEY_LENGTH;
3815 len = min_t(unsigned int, len, key_len);
3816 if (put_user(len, optlen))
3818 if (copy_to_user(optval, key, len))
3822 case TCP_THIN_LINEAR_TIMEOUTS:
3826 case TCP_THIN_DUPACK:
3834 case TCP_REPAIR_QUEUE:
3836 val = tp->repair_queue;
3841 case TCP_REPAIR_WINDOW: {
3842 struct tcp_repair_window opt;
3844 if (get_user(len, optlen))
3847 if (len != sizeof(opt))
3853 opt.snd_wl1 = tp->snd_wl1;
3854 opt.snd_wnd = tp->snd_wnd;
3855 opt.max_window = tp->max_window;
3856 opt.rcv_wnd = tp->rcv_wnd;
3857 opt.rcv_wup = tp->rcv_wup;
3859 if (copy_to_user(optval, &opt, len))
3864 if (tp->repair_queue == TCP_SEND_QUEUE)
3865 val = tp->write_seq;
3866 else if (tp->repair_queue == TCP_RECV_QUEUE)
3872 case TCP_USER_TIMEOUT:
3873 val = icsk->icsk_user_timeout;
3877 val = icsk->icsk_accept_queue.fastopenq.max_qlen;
3880 case TCP_FASTOPEN_CONNECT:
3881 val = tp->fastopen_connect;
3884 case TCP_FASTOPEN_NO_COOKIE:
3885 val = tp->fastopen_no_cookie;
3889 val = tp->tcp_tx_delay;
3893 val = tcp_time_stamp_raw() + tp->tsoffset;
3895 case TCP_NOTSENT_LOWAT:
3896 val = tp->notsent_lowat;
3899 val = tp->recvmsg_inq;
3904 case TCP_SAVED_SYN: {
3905 if (get_user(len, optlen))
3909 if (tp->saved_syn) {
3910 if (len < tcp_saved_syn_len(tp->saved_syn)) {
3911 if (put_user(tcp_saved_syn_len(tp->saved_syn),
3919 len = tcp_saved_syn_len(tp->saved_syn);
3920 if (put_user(len, optlen)) {
3924 if (copy_to_user(optval, tp->saved_syn->data, len)) {
3928 tcp_saved_syn_free(tp);
3933 if (put_user(len, optlen))
3939 case TCP_ZEROCOPY_RECEIVE: {
3940 struct tcp_zerocopy_receive zc = {};
3943 if (get_user(len, optlen))
3946 len < offsetofend(struct tcp_zerocopy_receive, length))
3948 if (len > sizeof(zc)) {
3950 if (put_user(len, optlen))
3953 if (copy_from_user(&zc, optval, len))
3956 err = tcp_zerocopy_receive(sk, &zc);
3958 if (len >= offsetofend(struct tcp_zerocopy_receive, err))
3959 goto zerocopy_rcv_sk_err;
3961 case offsetofend(struct tcp_zerocopy_receive, err):
3962 goto zerocopy_rcv_sk_err;
3963 case offsetofend(struct tcp_zerocopy_receive, inq):
3964 goto zerocopy_rcv_inq;
3965 case offsetofend(struct tcp_zerocopy_receive, length):
3967 goto zerocopy_rcv_out;
3969 zerocopy_rcv_sk_err:
3971 zc.err = sock_error(sk);
3973 zc.inq = tcp_inq_hint(sk);
3975 if (!err && copy_to_user(optval, &zc, len))
3981 return -ENOPROTOOPT;
3984 if (put_user(len, optlen))
3986 if (copy_to_user(optval, &val, len))
3991 int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval,
3994 struct inet_connection_sock *icsk = inet_csk(sk);
3996 if (level != SOL_TCP)
3997 return icsk->icsk_af_ops->getsockopt(sk, level, optname,
3999 return do_tcp_getsockopt(sk, level, optname, optval, optlen);
4001 EXPORT_SYMBOL(tcp_getsockopt);
4003 #ifdef CONFIG_TCP_MD5SIG
4004 static DEFINE_PER_CPU(struct tcp_md5sig_pool, tcp_md5sig_pool);
4005 static DEFINE_MUTEX(tcp_md5sig_mutex);
4006 static bool tcp_md5sig_pool_populated = false;
4008 static void __tcp_alloc_md5sig_pool(void)
4010 struct crypto_ahash *hash;
4013 hash = crypto_alloc_ahash("md5", 0, CRYPTO_ALG_ASYNC);
4017 for_each_possible_cpu(cpu) {
4018 void *scratch = per_cpu(tcp_md5sig_pool, cpu).scratch;
4019 struct ahash_request *req;
4022 scratch = kmalloc_node(sizeof(union tcp_md5sum_block) +
4023 sizeof(struct tcphdr),
4028 per_cpu(tcp_md5sig_pool, cpu).scratch = scratch;
4030 if (per_cpu(tcp_md5sig_pool, cpu).md5_req)
4033 req = ahash_request_alloc(hash, GFP_KERNEL);
4037 ahash_request_set_callback(req, 0, NULL, NULL);
4039 per_cpu(tcp_md5sig_pool, cpu).md5_req = req;
4041 /* before setting tcp_md5sig_pool_populated, we must commit all writes
4042 * to memory. See smp_rmb() in tcp_get_md5sig_pool()
4045 /* Paired with READ_ONCE() from tcp_alloc_md5sig_pool()
4046 * and tcp_get_md5sig_pool().
4048 WRITE_ONCE(tcp_md5sig_pool_populated, true);
4051 bool tcp_alloc_md5sig_pool(void)
4053 /* Paired with WRITE_ONCE() from __tcp_alloc_md5sig_pool() */
4054 if (unlikely(!READ_ONCE(tcp_md5sig_pool_populated))) {
4055 mutex_lock(&tcp_md5sig_mutex);
4057 if (!tcp_md5sig_pool_populated) {
4058 __tcp_alloc_md5sig_pool();
4059 if (tcp_md5sig_pool_populated)
4060 static_branch_inc(&tcp_md5_needed);
4063 mutex_unlock(&tcp_md5sig_mutex);
4065 /* Paired with WRITE_ONCE() from __tcp_alloc_md5sig_pool() */
4066 return READ_ONCE(tcp_md5sig_pool_populated);
4068 EXPORT_SYMBOL(tcp_alloc_md5sig_pool);
4072 * tcp_get_md5sig_pool - get md5sig_pool for this user
4074 * We use percpu structure, so if we succeed, we exit with preemption
4075 * and BH disabled, to make sure another thread or softirq handling
4076 * wont try to get same context.
4078 struct tcp_md5sig_pool *tcp_get_md5sig_pool(void)
4082 /* Paired with WRITE_ONCE() from __tcp_alloc_md5sig_pool() */
4083 if (READ_ONCE(tcp_md5sig_pool_populated)) {
4084 /* coupled with smp_wmb() in __tcp_alloc_md5sig_pool() */
4086 return this_cpu_ptr(&tcp_md5sig_pool);
4091 EXPORT_SYMBOL(tcp_get_md5sig_pool);
4093 int tcp_md5_hash_skb_data(struct tcp_md5sig_pool *hp,
4094 const struct sk_buff *skb, unsigned int header_len)
4096 struct scatterlist sg;
4097 const struct tcphdr *tp = tcp_hdr(skb);
4098 struct ahash_request *req = hp->md5_req;
4100 const unsigned int head_data_len = skb_headlen(skb) > header_len ?
4101 skb_headlen(skb) - header_len : 0;
4102 const struct skb_shared_info *shi = skb_shinfo(skb);
4103 struct sk_buff *frag_iter;
4105 sg_init_table(&sg, 1);
4107 sg_set_buf(&sg, ((u8 *) tp) + header_len, head_data_len);
4108 ahash_request_set_crypt(req, &sg, NULL, head_data_len);
4109 if (crypto_ahash_update(req))
4112 for (i = 0; i < shi->nr_frags; ++i) {
4113 const skb_frag_t *f = &shi->frags[i];
4114 unsigned int offset = skb_frag_off(f);
4115 struct page *page = skb_frag_page(f) + (offset >> PAGE_SHIFT);
4117 sg_set_page(&sg, page, skb_frag_size(f),
4118 offset_in_page(offset));
4119 ahash_request_set_crypt(req, &sg, NULL, skb_frag_size(f));
4120 if (crypto_ahash_update(req))
4124 skb_walk_frags(skb, frag_iter)
4125 if (tcp_md5_hash_skb_data(hp, frag_iter, 0))
4130 EXPORT_SYMBOL(tcp_md5_hash_skb_data);
4132 int tcp_md5_hash_key(struct tcp_md5sig_pool *hp, const struct tcp_md5sig_key *key)
4134 u8 keylen = READ_ONCE(key->keylen); /* paired with WRITE_ONCE() in tcp_md5_do_add */
4135 struct scatterlist sg;
4137 sg_init_one(&sg, key->key, keylen);
4138 ahash_request_set_crypt(hp->md5_req, &sg, NULL, keylen);
4140 /* We use data_race() because tcp_md5_do_add() might change key->key under us */
4141 return data_race(crypto_ahash_update(hp->md5_req));
4143 EXPORT_SYMBOL(tcp_md5_hash_key);
4147 void tcp_done(struct sock *sk)
4149 struct request_sock *req;
4151 /* We might be called with a new socket, after
4152 * inet_csk_prepare_forced_close() has been called
4153 * so we can not use lockdep_sock_is_held(sk)
4155 req = rcu_dereference_protected(tcp_sk(sk)->fastopen_rsk, 1);
4157 if (sk->sk_state == TCP_SYN_SENT || sk->sk_state == TCP_SYN_RECV)
4158 TCP_INC_STATS(sock_net(sk), TCP_MIB_ATTEMPTFAILS);
4160 tcp_set_state(sk, TCP_CLOSE);
4161 tcp_clear_xmit_timers(sk);
4163 reqsk_fastopen_remove(sk, req, false);
4165 sk->sk_shutdown = SHUTDOWN_MASK;
4167 if (!sock_flag(sk, SOCK_DEAD))
4168 sk->sk_state_change(sk);
4170 inet_csk_destroy_sock(sk);
4172 EXPORT_SYMBOL_GPL(tcp_done);
4174 int tcp_abort(struct sock *sk, int err)
4176 if (!sk_fullsock(sk)) {
4177 if (sk->sk_state == TCP_NEW_SYN_RECV) {
4178 struct request_sock *req = inet_reqsk(sk);
4181 inet_csk_reqsk_queue_drop(req->rsk_listener, req);
4188 /* Don't race with userspace socket closes such as tcp_close. */
4191 if (sk->sk_state == TCP_LISTEN) {
4192 tcp_set_state(sk, TCP_CLOSE);
4193 inet_csk_listen_stop(sk);
4196 /* Don't race with BH socket closes such as inet_csk_listen_stop. */
4200 if (!sock_flag(sk, SOCK_DEAD)) {
4202 /* This barrier is coupled with smp_rmb() in tcp_poll() */
4204 sk->sk_error_report(sk);
4205 if (tcp_need_reset(sk->sk_state))
4206 tcp_send_active_reset(sk, GFP_ATOMIC);
4212 tcp_write_queue_purge(sk);
4216 EXPORT_SYMBOL_GPL(tcp_abort);
4218 extern struct tcp_congestion_ops tcp_reno;
4220 static __initdata unsigned long thash_entries;
4221 static int __init set_thash_entries(char *str)
4228 ret = kstrtoul(str, 0, &thash_entries);
4234 __setup("thash_entries=", set_thash_entries);
4236 static void __init tcp_init_mem(void)
4238 unsigned long limit = nr_free_buffer_pages() / 16;
4240 limit = max(limit, 128UL);
4241 sysctl_tcp_mem[0] = limit / 4 * 3; /* 4.68 % */
4242 sysctl_tcp_mem[1] = limit; /* 6.25 % */
4243 sysctl_tcp_mem[2] = sysctl_tcp_mem[0] * 2; /* 9.37 % */
4246 void __init tcp_init(void)
4248 int max_rshare, max_wshare, cnt;
4249 unsigned long limit;
4252 BUILD_BUG_ON(TCP_MIN_SND_MSS <= MAX_TCP_OPTION_SPACE);
4253 BUILD_BUG_ON(sizeof(struct tcp_skb_cb) >
4254 sizeof_field(struct sk_buff, cb));
4256 percpu_counter_init(&tcp_sockets_allocated, 0, GFP_KERNEL);
4258 timer_setup(&tcp_orphan_timer, tcp_orphan_update, TIMER_DEFERRABLE);
4259 mod_timer(&tcp_orphan_timer, jiffies + TCP_ORPHAN_TIMER_PERIOD);
4261 inet_hashinfo_init(&tcp_hashinfo);
4262 inet_hashinfo2_init(&tcp_hashinfo, "tcp_listen_portaddr_hash",
4263 thash_entries, 21, /* one slot per 2 MB*/
4265 tcp_hashinfo.bind_bucket_cachep =
4266 kmem_cache_create("tcp_bind_bucket",
4267 sizeof(struct inet_bind_bucket), 0,
4268 SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL);
4270 /* Size and allocate the main established and bind bucket
4273 * The methodology is similar to that of the buffer cache.
4275 tcp_hashinfo.ehash =
4276 alloc_large_system_hash("TCP established",
4277 sizeof(struct inet_ehash_bucket),
4279 17, /* one slot per 128 KB of memory */
4282 &tcp_hashinfo.ehash_mask,
4284 thash_entries ? 0 : 512 * 1024);
4285 for (i = 0; i <= tcp_hashinfo.ehash_mask; i++)
4286 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].chain, i);
4288 if (inet_ehash_locks_alloc(&tcp_hashinfo))
4289 panic("TCP: failed to alloc ehash_locks");
4290 tcp_hashinfo.bhash =
4291 alloc_large_system_hash("TCP bind",
4292 sizeof(struct inet_bind_hashbucket),
4293 tcp_hashinfo.ehash_mask + 1,
4294 17, /* one slot per 128 KB of memory */
4296 &tcp_hashinfo.bhash_size,
4300 tcp_hashinfo.bhash_size = 1U << tcp_hashinfo.bhash_size;
4301 for (i = 0; i < tcp_hashinfo.bhash_size; i++) {
4302 spin_lock_init(&tcp_hashinfo.bhash[i].lock);
4303 INIT_HLIST_HEAD(&tcp_hashinfo.bhash[i].chain);
4307 cnt = tcp_hashinfo.ehash_mask + 1;
4308 sysctl_tcp_max_orphans = cnt / 2;
4311 /* Set per-socket limits to no more than 1/128 the pressure threshold */
4312 limit = nr_free_buffer_pages() << (PAGE_SHIFT - 7);
4313 max_wshare = min(4UL*1024*1024, limit);
4314 max_rshare = min(6UL*1024*1024, limit);
4316 init_net.ipv4.sysctl_tcp_wmem[0] = SK_MEM_QUANTUM;
4317 init_net.ipv4.sysctl_tcp_wmem[1] = 16*1024;
4318 init_net.ipv4.sysctl_tcp_wmem[2] = max(64*1024, max_wshare);
4320 init_net.ipv4.sysctl_tcp_rmem[0] = SK_MEM_QUANTUM;
4321 init_net.ipv4.sysctl_tcp_rmem[1] = 131072;
4322 init_net.ipv4.sysctl_tcp_rmem[2] = max(131072, max_rshare);
4324 pr_info("Hash tables configured (established %u bind %u)\n",
4325 tcp_hashinfo.ehash_mask + 1, tcp_hashinfo.bhash_size);
4329 BUG_ON(tcp_register_congestion_control(&tcp_reno) != 0);