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/cache.h>
264 #include <linux/err.h>
265 #include <linux/time.h>
266 #include <linux/slab.h>
267 #include <linux/errqueue.h>
268 #include <linux/static_key.h>
269 #include <linux/btf.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 /* Track pending CMSGs. */
289 DEFINE_PER_CPU(unsigned int, tcp_orphan_count);
290 EXPORT_PER_CPU_SYMBOL_GPL(tcp_orphan_count);
292 long sysctl_tcp_mem[3] __read_mostly;
293 EXPORT_SYMBOL(sysctl_tcp_mem);
295 atomic_long_t tcp_memory_allocated ____cacheline_aligned_in_smp; /* Current allocated memory. */
296 EXPORT_SYMBOL(tcp_memory_allocated);
297 DEFINE_PER_CPU(int, tcp_memory_per_cpu_fw_alloc);
298 EXPORT_PER_CPU_SYMBOL_GPL(tcp_memory_per_cpu_fw_alloc);
300 #if IS_ENABLED(CONFIG_SMC)
301 DEFINE_STATIC_KEY_FALSE(tcp_have_smc);
302 EXPORT_SYMBOL(tcp_have_smc);
306 * Current number of TCP sockets.
308 struct percpu_counter tcp_sockets_allocated ____cacheline_aligned_in_smp;
309 EXPORT_SYMBOL(tcp_sockets_allocated);
314 struct tcp_splice_state {
315 struct pipe_inode_info *pipe;
321 * Pressure flag: try to collapse.
322 * Technical note: it is used by multiple contexts non atomically.
323 * All the __sk_mem_schedule() is of this nature: accounting
324 * is strict, actions are advisory and have some latency.
326 unsigned long tcp_memory_pressure __read_mostly;
327 EXPORT_SYMBOL_GPL(tcp_memory_pressure);
329 void tcp_enter_memory_pressure(struct sock *sk)
333 if (READ_ONCE(tcp_memory_pressure))
339 if (!cmpxchg(&tcp_memory_pressure, 0, val))
340 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMEMORYPRESSURES);
342 EXPORT_SYMBOL_GPL(tcp_enter_memory_pressure);
344 void tcp_leave_memory_pressure(struct sock *sk)
348 if (!READ_ONCE(tcp_memory_pressure))
350 val = xchg(&tcp_memory_pressure, 0);
352 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPMEMORYPRESSURESCHRONO,
353 jiffies_to_msecs(jiffies - val));
355 EXPORT_SYMBOL_GPL(tcp_leave_memory_pressure);
357 /* Convert seconds to retransmits based on initial and max timeout */
358 static u8 secs_to_retrans(int seconds, int timeout, int rto_max)
363 int period = timeout;
366 while (seconds > period && res < 255) {
369 if (timeout > rto_max)
377 /* Convert retransmits to seconds based on initial and max timeout */
378 static int retrans_to_secs(u8 retrans, int timeout, int rto_max)
386 if (timeout > rto_max)
394 static u64 tcp_compute_delivery_rate(const struct tcp_sock *tp)
396 u32 rate = READ_ONCE(tp->rate_delivered);
397 u32 intv = READ_ONCE(tp->rate_interval_us);
401 rate64 = (u64)rate * tp->mss_cache * USEC_PER_SEC;
402 do_div(rate64, intv);
407 /* Address-family independent initialization for a tcp_sock.
409 * NOTE: A lot of things set to zero explicitly by call to
410 * sk_alloc() so need not be done here.
412 void tcp_init_sock(struct sock *sk)
414 struct inet_connection_sock *icsk = inet_csk(sk);
415 struct tcp_sock *tp = tcp_sk(sk);
417 tp->out_of_order_queue = RB_ROOT;
418 sk->tcp_rtx_queue = RB_ROOT;
419 tcp_init_xmit_timers(sk);
420 INIT_LIST_HEAD(&tp->tsq_node);
421 INIT_LIST_HEAD(&tp->tsorted_sent_queue);
423 icsk->icsk_rto = TCP_TIMEOUT_INIT;
424 icsk->icsk_rto_min = TCP_RTO_MIN;
425 icsk->icsk_delack_max = TCP_DELACK_MAX;
426 tp->mdev_us = jiffies_to_usecs(TCP_TIMEOUT_INIT);
427 minmax_reset(&tp->rtt_min, tcp_jiffies32, ~0U);
429 /* So many TCP implementations out there (incorrectly) count the
430 * initial SYN frame in their delayed-ACK and congestion control
431 * algorithms that we must have the following bandaid to talk
432 * efficiently to them. -DaveM
434 tcp_snd_cwnd_set(tp, TCP_INIT_CWND);
436 /* There's a bubble in the pipe until at least the first ACK. */
437 tp->app_limited = ~0U;
439 /* See draft-stevens-tcpca-spec-01 for discussion of the
440 * initialization of these values.
442 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
443 tp->snd_cwnd_clamp = ~0;
444 tp->mss_cache = TCP_MSS_DEFAULT;
446 tp->reordering = sock_net(sk)->ipv4.sysctl_tcp_reordering;
447 tcp_assign_congestion_control(sk);
450 tp->rack.reo_wnd_steps = 1;
452 sk->sk_write_space = sk_stream_write_space;
453 sock_set_flag(sk, SOCK_USE_WRITE_QUEUE);
455 icsk->icsk_sync_mss = tcp_sync_mss;
457 WRITE_ONCE(sk->sk_sndbuf, sock_net(sk)->ipv4.sysctl_tcp_wmem[1]);
458 WRITE_ONCE(sk->sk_rcvbuf, sock_net(sk)->ipv4.sysctl_tcp_rmem[1]);
460 sk_sockets_allocated_inc(sk);
462 EXPORT_SYMBOL(tcp_init_sock);
464 static void tcp_tx_timestamp(struct sock *sk, u16 tsflags)
466 struct sk_buff *skb = tcp_write_queue_tail(sk);
468 if (tsflags && skb) {
469 struct skb_shared_info *shinfo = skb_shinfo(skb);
470 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
472 sock_tx_timestamp(sk, tsflags, &shinfo->tx_flags);
473 if (tsflags & SOF_TIMESTAMPING_TX_ACK)
474 tcb->txstamp_ack = 1;
475 if (tsflags & SOF_TIMESTAMPING_TX_RECORD_MASK)
476 shinfo->tskey = TCP_SKB_CB(skb)->seq + skb->len - 1;
480 static bool tcp_stream_is_readable(struct sock *sk, int target)
482 if (tcp_epollin_ready(sk, target))
484 return sk_is_readable(sk);
488 * Wait for a TCP event.
490 * Note that we don't need to lock the socket, as the upper poll layers
491 * take care of normal races (between the test and the event) and we don't
492 * go look at any of the socket buffers directly.
494 __poll_t tcp_poll(struct file *file, struct socket *sock, poll_table *wait)
497 struct sock *sk = sock->sk;
498 const struct tcp_sock *tp = tcp_sk(sk);
501 sock_poll_wait(file, sock, wait);
503 state = inet_sk_state_load(sk);
504 if (state == TCP_LISTEN)
505 return inet_csk_listen_poll(sk);
507 /* Socket is not locked. We are protected from async events
508 * by poll logic and correct handling of state changes
509 * made by other threads is impossible in any case.
515 * EPOLLHUP is certainly not done right. But poll() doesn't
516 * have a notion of HUP in just one direction, and for a
517 * socket the read side is more interesting.
519 * Some poll() documentation says that EPOLLHUP is incompatible
520 * with the EPOLLOUT/POLLWR flags, so somebody should check this
521 * all. But careful, it tends to be safer to return too many
522 * bits than too few, and you can easily break real applications
523 * if you don't tell them that something has hung up!
527 * Check number 1. EPOLLHUP is _UNMASKABLE_ event (see UNIX98 and
528 * our fs/select.c). It means that after we received EOF,
529 * poll always returns immediately, making impossible poll() on write()
530 * in state CLOSE_WAIT. One solution is evident --- to set EPOLLHUP
531 * if and only if shutdown has been made in both directions.
532 * Actually, it is interesting to look how Solaris and DUX
533 * solve this dilemma. I would prefer, if EPOLLHUP were maskable,
534 * then we could set it on SND_SHUTDOWN. BTW examples given
535 * in Stevens' books assume exactly this behaviour, it explains
536 * why EPOLLHUP is incompatible with EPOLLOUT. --ANK
538 * NOTE. Check for TCP_CLOSE is added. The goal is to prevent
539 * blocking on fresh not-connected or disconnected socket. --ANK
541 if (sk->sk_shutdown == SHUTDOWN_MASK || state == TCP_CLOSE)
543 if (sk->sk_shutdown & RCV_SHUTDOWN)
544 mask |= EPOLLIN | EPOLLRDNORM | EPOLLRDHUP;
546 /* Connected or passive Fast Open socket? */
547 if (state != TCP_SYN_SENT &&
548 (state != TCP_SYN_RECV || rcu_access_pointer(tp->fastopen_rsk))) {
549 int target = sock_rcvlowat(sk, 0, INT_MAX);
550 u16 urg_data = READ_ONCE(tp->urg_data);
552 if (unlikely(urg_data) &&
553 READ_ONCE(tp->urg_seq) == READ_ONCE(tp->copied_seq) &&
554 !sock_flag(sk, SOCK_URGINLINE))
557 if (tcp_stream_is_readable(sk, target))
558 mask |= EPOLLIN | EPOLLRDNORM;
560 if (!(sk->sk_shutdown & SEND_SHUTDOWN)) {
561 if (__sk_stream_is_writeable(sk, 1)) {
562 mask |= EPOLLOUT | EPOLLWRNORM;
563 } else { /* send SIGIO later */
564 sk_set_bit(SOCKWQ_ASYNC_NOSPACE, sk);
565 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
567 /* Race breaker. If space is freed after
568 * wspace test but before the flags are set,
569 * IO signal will be lost. Memory barrier
570 * pairs with the input side.
572 smp_mb__after_atomic();
573 if (__sk_stream_is_writeable(sk, 1))
574 mask |= EPOLLOUT | EPOLLWRNORM;
577 mask |= EPOLLOUT | EPOLLWRNORM;
579 if (urg_data & TCP_URG_VALID)
581 } else if (state == TCP_SYN_SENT && inet_sk(sk)->defer_connect) {
582 /* Active TCP fastopen socket with defer_connect
583 * Return EPOLLOUT so application can call write()
584 * in order for kernel to generate SYN+data
586 mask |= EPOLLOUT | EPOLLWRNORM;
588 /* This barrier is coupled with smp_wmb() in tcp_reset() */
590 if (sk->sk_err || !skb_queue_empty_lockless(&sk->sk_error_queue))
595 EXPORT_SYMBOL(tcp_poll);
597 int tcp_ioctl(struct sock *sk, int cmd, unsigned long arg)
599 struct tcp_sock *tp = tcp_sk(sk);
605 if (sk->sk_state == TCP_LISTEN)
608 slow = lock_sock_fast(sk);
610 unlock_sock_fast(sk, slow);
613 answ = READ_ONCE(tp->urg_data) &&
614 READ_ONCE(tp->urg_seq) == READ_ONCE(tp->copied_seq);
617 if (sk->sk_state == TCP_LISTEN)
620 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
623 answ = READ_ONCE(tp->write_seq) - tp->snd_una;
626 if (sk->sk_state == TCP_LISTEN)
629 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
632 answ = READ_ONCE(tp->write_seq) -
633 READ_ONCE(tp->snd_nxt);
639 return put_user(answ, (int __user *)arg);
641 EXPORT_SYMBOL(tcp_ioctl);
643 void tcp_mark_push(struct tcp_sock *tp, struct sk_buff *skb)
645 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
646 tp->pushed_seq = tp->write_seq;
649 static inline bool forced_push(const struct tcp_sock *tp)
651 return after(tp->write_seq, tp->pushed_seq + (tp->max_window >> 1));
654 void tcp_skb_entail(struct sock *sk, struct sk_buff *skb)
656 struct tcp_sock *tp = tcp_sk(sk);
657 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
659 tcb->seq = tcb->end_seq = tp->write_seq;
660 tcb->tcp_flags = TCPHDR_ACK;
661 __skb_header_release(skb);
662 tcp_add_write_queue_tail(sk, skb);
663 sk_wmem_queued_add(sk, skb->truesize);
664 sk_mem_charge(sk, skb->truesize);
665 if (tp->nonagle & TCP_NAGLE_PUSH)
666 tp->nonagle &= ~TCP_NAGLE_PUSH;
668 tcp_slow_start_after_idle_check(sk);
671 static inline void tcp_mark_urg(struct tcp_sock *tp, int flags)
674 tp->snd_up = tp->write_seq;
677 /* If a not yet filled skb is pushed, do not send it if
678 * we have data packets in Qdisc or NIC queues :
679 * Because TX completion will happen shortly, it gives a chance
680 * to coalesce future sendmsg() payload into this skb, without
681 * need for a timer, and with no latency trade off.
682 * As packets containing data payload have a bigger truesize
683 * than pure acks (dataless) packets, the last checks prevent
684 * autocorking if we only have an ACK in Qdisc/NIC queues,
685 * or if TX completion was delayed after we processed ACK packet.
687 static bool tcp_should_autocork(struct sock *sk, struct sk_buff *skb,
690 return skb->len < size_goal &&
691 sock_net(sk)->ipv4.sysctl_tcp_autocorking &&
692 !tcp_rtx_queue_empty(sk) &&
693 refcount_read(&sk->sk_wmem_alloc) > skb->truesize &&
694 tcp_skb_can_collapse_to(skb);
697 void tcp_push(struct sock *sk, int flags, int mss_now,
698 int nonagle, int size_goal)
700 struct tcp_sock *tp = tcp_sk(sk);
703 skb = tcp_write_queue_tail(sk);
706 if (!(flags & MSG_MORE) || forced_push(tp))
707 tcp_mark_push(tp, skb);
709 tcp_mark_urg(tp, flags);
711 if (tcp_should_autocork(sk, skb, size_goal)) {
713 /* avoid atomic op if TSQ_THROTTLED bit is already set */
714 if (!test_bit(TSQ_THROTTLED, &sk->sk_tsq_flags)) {
715 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPAUTOCORKING);
716 set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags);
718 /* It is possible TX completion already happened
719 * before we set TSQ_THROTTLED.
721 if (refcount_read(&sk->sk_wmem_alloc) > skb->truesize)
725 if (flags & MSG_MORE)
726 nonagle = TCP_NAGLE_CORK;
728 __tcp_push_pending_frames(sk, mss_now, nonagle);
731 static int tcp_splice_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb,
732 unsigned int offset, size_t len)
734 struct tcp_splice_state *tss = rd_desc->arg.data;
737 ret = skb_splice_bits(skb, skb->sk, offset, tss->pipe,
738 min(rd_desc->count, len), tss->flags);
740 rd_desc->count -= ret;
744 static int __tcp_splice_read(struct sock *sk, struct tcp_splice_state *tss)
746 /* Store TCP splice context information in read_descriptor_t. */
747 read_descriptor_t rd_desc = {
752 return tcp_read_sock(sk, &rd_desc, tcp_splice_data_recv);
756 * tcp_splice_read - splice data from TCP socket to a pipe
757 * @sock: socket to splice from
758 * @ppos: position (not valid)
759 * @pipe: pipe to splice to
760 * @len: number of bytes to splice
761 * @flags: splice modifier flags
764 * Will read pages from given socket and fill them into a pipe.
767 ssize_t tcp_splice_read(struct socket *sock, loff_t *ppos,
768 struct pipe_inode_info *pipe, size_t len,
771 struct sock *sk = sock->sk;
772 struct tcp_splice_state tss = {
781 sock_rps_record_flow(sk);
783 * We can't seek on a socket input
792 timeo = sock_rcvtimeo(sk, sock->file->f_flags & O_NONBLOCK);
794 ret = __tcp_splice_read(sk, &tss);
800 if (sock_flag(sk, SOCK_DONE))
803 ret = sock_error(sk);
806 if (sk->sk_shutdown & RCV_SHUTDOWN)
808 if (sk->sk_state == TCP_CLOSE) {
810 * This occurs when user tries to read
811 * from never connected socket.
820 /* if __tcp_splice_read() got nothing while we have
821 * an skb in receive queue, we do not want to loop.
822 * This might happen with URG data.
824 if (!skb_queue_empty(&sk->sk_receive_queue))
826 sk_wait_data(sk, &timeo, NULL);
827 if (signal_pending(current)) {
828 ret = sock_intr_errno(timeo);
841 if (sk->sk_err || sk->sk_state == TCP_CLOSE ||
842 (sk->sk_shutdown & RCV_SHUTDOWN) ||
843 signal_pending(current))
854 EXPORT_SYMBOL(tcp_splice_read);
856 struct sk_buff *tcp_stream_alloc_skb(struct sock *sk, int size, gfp_t gfp,
861 skb = alloc_skb_fclone(size + MAX_TCP_HEADER, gfp);
865 skb->truesize = SKB_TRUESIZE(skb_end_offset(skb));
866 if (force_schedule) {
867 mem_scheduled = true;
868 sk_forced_mem_schedule(sk, skb->truesize);
870 mem_scheduled = sk_wmem_schedule(sk, skb->truesize);
872 if (likely(mem_scheduled)) {
873 skb_reserve(skb, MAX_TCP_HEADER);
874 skb->ip_summed = CHECKSUM_PARTIAL;
875 INIT_LIST_HEAD(&skb->tcp_tsorted_anchor);
880 sk->sk_prot->enter_memory_pressure(sk);
881 sk_stream_moderate_sndbuf(sk);
886 static unsigned int tcp_xmit_size_goal(struct sock *sk, u32 mss_now,
889 struct tcp_sock *tp = tcp_sk(sk);
890 u32 new_size_goal, size_goal;
895 /* Note : tcp_tso_autosize() will eventually split this later */
896 new_size_goal = tcp_bound_to_half_wnd(tp, sk->sk_gso_max_size);
898 /* We try hard to avoid divides here */
899 size_goal = tp->gso_segs * mss_now;
900 if (unlikely(new_size_goal < size_goal ||
901 new_size_goal >= size_goal + mss_now)) {
902 tp->gso_segs = min_t(u16, new_size_goal / mss_now,
903 sk->sk_gso_max_segs);
904 size_goal = tp->gso_segs * mss_now;
907 return max(size_goal, mss_now);
910 int tcp_send_mss(struct sock *sk, int *size_goal, int flags)
914 mss_now = tcp_current_mss(sk);
915 *size_goal = tcp_xmit_size_goal(sk, mss_now, !(flags & MSG_OOB));
920 /* In some cases, both sendpage() and sendmsg() could have added
921 * an skb to the write queue, but failed adding payload on it.
922 * We need to remove it to consume less memory, but more
923 * importantly be able to generate EPOLLOUT for Edge Trigger epoll()
926 void tcp_remove_empty_skb(struct sock *sk)
928 struct sk_buff *skb = tcp_write_queue_tail(sk);
930 if (skb && TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq) {
931 tcp_unlink_write_queue(skb, sk);
932 if (tcp_write_queue_empty(sk))
933 tcp_chrono_stop(sk, TCP_CHRONO_BUSY);
934 tcp_wmem_free_skb(sk, skb);
938 /* skb changing from pure zc to mixed, must charge zc */
939 static int tcp_downgrade_zcopy_pure(struct sock *sk, struct sk_buff *skb)
941 if (unlikely(skb_zcopy_pure(skb))) {
942 u32 extra = skb->truesize -
943 SKB_TRUESIZE(skb_end_offset(skb));
945 if (!sk_wmem_schedule(sk, extra))
948 sk_mem_charge(sk, extra);
949 skb_shinfo(skb)->flags &= ~SKBFL_PURE_ZEROCOPY;
955 static int tcp_wmem_schedule(struct sock *sk, int copy)
959 if (likely(sk_wmem_schedule(sk, copy)))
962 /* We could be in trouble if we have nothing queued.
963 * Use whatever is left in sk->sk_forward_alloc and tcp_wmem[0]
964 * to guarantee some progress.
966 left = sock_net(sk)->ipv4.sysctl_tcp_wmem[0] - sk->sk_wmem_queued;
968 sk_forced_mem_schedule(sk, min(left, copy));
969 return min(copy, sk->sk_forward_alloc);
972 static struct sk_buff *tcp_build_frag(struct sock *sk, int size_goal, int flags,
973 struct page *page, int offset, size_t *size)
975 struct sk_buff *skb = tcp_write_queue_tail(sk);
976 struct tcp_sock *tp = tcp_sk(sk);
980 if (!skb || (copy = size_goal - skb->len) <= 0 ||
981 !tcp_skb_can_collapse_to(skb)) {
983 if (!sk_stream_memory_free(sk))
986 skb = tcp_stream_alloc_skb(sk, 0, sk->sk_allocation,
987 tcp_rtx_and_write_queues_empty(sk));
991 #ifdef CONFIG_TLS_DEVICE
992 skb->decrypted = !!(flags & MSG_SENDPAGE_DECRYPTED);
994 tcp_skb_entail(sk, skb);
1001 i = skb_shinfo(skb)->nr_frags;
1002 can_coalesce = skb_can_coalesce(skb, i, page, offset);
1003 if (!can_coalesce && i >= sysctl_max_skb_frags) {
1004 tcp_mark_push(tp, skb);
1007 if (tcp_downgrade_zcopy_pure(sk, skb))
1010 copy = tcp_wmem_schedule(sk, copy);
1015 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1018 skb_fill_page_desc(skb, i, page, offset, copy);
1021 if (!(flags & MSG_NO_SHARED_FRAGS))
1022 skb_shinfo(skb)->flags |= SKBFL_SHARED_FRAG;
1025 skb->data_len += copy;
1026 skb->truesize += copy;
1027 sk_wmem_queued_add(sk, copy);
1028 sk_mem_charge(sk, copy);
1029 WRITE_ONCE(tp->write_seq, tp->write_seq + copy);
1030 TCP_SKB_CB(skb)->end_seq += copy;
1031 tcp_skb_pcount_set(skb, 0);
1037 ssize_t do_tcp_sendpages(struct sock *sk, struct page *page, int offset,
1038 size_t size, int flags)
1040 struct tcp_sock *tp = tcp_sk(sk);
1041 int mss_now, size_goal;
1044 long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
1046 if (IS_ENABLED(CONFIG_DEBUG_VM) &&
1047 WARN_ONCE(!sendpage_ok(page),
1048 "page must not be a Slab one and have page_count > 0"))
1051 /* Wait for a connection to finish. One exception is TCP Fast Open
1052 * (passive side) where data is allowed to be sent before a connection
1053 * is fully established.
1055 if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) &&
1056 !tcp_passive_fastopen(sk)) {
1057 err = sk_stream_wait_connect(sk, &timeo);
1062 sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
1064 mss_now = tcp_send_mss(sk, &size_goal, flags);
1068 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
1072 struct sk_buff *skb;
1075 skb = tcp_build_frag(sk, size_goal, flags, page, offset, ©);
1077 goto wait_for_space;
1080 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
1088 if (skb->len < size_goal || (flags & MSG_OOB))
1091 if (forced_push(tp)) {
1092 tcp_mark_push(tp, skb);
1093 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
1094 } else if (skb == tcp_send_head(sk))
1095 tcp_push_one(sk, mss_now);
1099 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1100 tcp_push(sk, flags & ~MSG_MORE, mss_now,
1101 TCP_NAGLE_PUSH, size_goal);
1103 err = sk_stream_wait_memory(sk, &timeo);
1107 mss_now = tcp_send_mss(sk, &size_goal, flags);
1112 tcp_tx_timestamp(sk, sk->sk_tsflags);
1113 if (!(flags & MSG_SENDPAGE_NOTLAST))
1114 tcp_push(sk, flags, mss_now, tp->nonagle, size_goal);
1119 tcp_remove_empty_skb(sk);
1123 /* make sure we wake any epoll edge trigger waiter */
1124 if (unlikely(tcp_rtx_and_write_queues_empty(sk) && err == -EAGAIN)) {
1125 sk->sk_write_space(sk);
1126 tcp_chrono_stop(sk, TCP_CHRONO_SNDBUF_LIMITED);
1128 return sk_stream_error(sk, flags, err);
1130 EXPORT_SYMBOL_GPL(do_tcp_sendpages);
1132 int tcp_sendpage_locked(struct sock *sk, struct page *page, int offset,
1133 size_t size, int flags)
1135 if (!(sk->sk_route_caps & NETIF_F_SG))
1136 return sock_no_sendpage_locked(sk, page, offset, size, flags);
1138 tcp_rate_check_app_limited(sk); /* is sending application-limited? */
1140 return do_tcp_sendpages(sk, page, offset, size, flags);
1142 EXPORT_SYMBOL_GPL(tcp_sendpage_locked);
1144 int tcp_sendpage(struct sock *sk, struct page *page, int offset,
1145 size_t size, int flags)
1150 ret = tcp_sendpage_locked(sk, page, offset, size, flags);
1155 EXPORT_SYMBOL(tcp_sendpage);
1157 void tcp_free_fastopen_req(struct tcp_sock *tp)
1159 if (tp->fastopen_req) {
1160 kfree(tp->fastopen_req);
1161 tp->fastopen_req = NULL;
1165 static int tcp_sendmsg_fastopen(struct sock *sk, struct msghdr *msg,
1166 int *copied, size_t size,
1167 struct ubuf_info *uarg)
1169 struct tcp_sock *tp = tcp_sk(sk);
1170 struct inet_sock *inet = inet_sk(sk);
1171 struct sockaddr *uaddr = msg->msg_name;
1174 if (!(sock_net(sk)->ipv4.sysctl_tcp_fastopen & TFO_CLIENT_ENABLE) ||
1175 (uaddr && msg->msg_namelen >= sizeof(uaddr->sa_family) &&
1176 uaddr->sa_family == AF_UNSPEC))
1178 if (tp->fastopen_req)
1179 return -EALREADY; /* Another Fast Open is in progress */
1181 tp->fastopen_req = kzalloc(sizeof(struct tcp_fastopen_request),
1183 if (unlikely(!tp->fastopen_req))
1185 tp->fastopen_req->data = msg;
1186 tp->fastopen_req->size = size;
1187 tp->fastopen_req->uarg = uarg;
1189 if (inet->defer_connect) {
1190 err = tcp_connect(sk);
1191 /* Same failure procedure as in tcp_v4/6_connect */
1193 tcp_set_state(sk, TCP_CLOSE);
1194 inet->inet_dport = 0;
1195 sk->sk_route_caps = 0;
1198 flags = (msg->msg_flags & MSG_DONTWAIT) ? O_NONBLOCK : 0;
1199 err = __inet_stream_connect(sk->sk_socket, uaddr,
1200 msg->msg_namelen, flags, 1);
1201 /* fastopen_req could already be freed in __inet_stream_connect
1202 * if the connection times out or gets rst
1204 if (tp->fastopen_req) {
1205 *copied = tp->fastopen_req->copied;
1206 tcp_free_fastopen_req(tp);
1207 inet->defer_connect = 0;
1212 int tcp_sendmsg_locked(struct sock *sk, struct msghdr *msg, size_t size)
1214 struct tcp_sock *tp = tcp_sk(sk);
1215 struct ubuf_info *uarg = NULL;
1216 struct sk_buff *skb;
1217 struct sockcm_cookie sockc;
1218 int flags, err, copied = 0;
1219 int mss_now = 0, size_goal, copied_syn = 0;
1220 int process_backlog = 0;
1224 flags = msg->msg_flags;
1226 if (flags & MSG_ZEROCOPY && size && sock_flag(sk, SOCK_ZEROCOPY)) {
1227 skb = tcp_write_queue_tail(sk);
1228 uarg = msg_zerocopy_realloc(sk, size, skb_zcopy(skb));
1234 zc = sk->sk_route_caps & NETIF_F_SG;
1239 if (unlikely(flags & MSG_FASTOPEN || inet_sk(sk)->defer_connect) &&
1241 err = tcp_sendmsg_fastopen(sk, msg, &copied_syn, size, uarg);
1242 if (err == -EINPROGRESS && copied_syn > 0)
1248 timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
1250 tcp_rate_check_app_limited(sk); /* is sending application-limited? */
1252 /* Wait for a connection to finish. One exception is TCP Fast Open
1253 * (passive side) where data is allowed to be sent before a connection
1254 * is fully established.
1256 if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) &&
1257 !tcp_passive_fastopen(sk)) {
1258 err = sk_stream_wait_connect(sk, &timeo);
1263 if (unlikely(tp->repair)) {
1264 if (tp->repair_queue == TCP_RECV_QUEUE) {
1265 copied = tcp_send_rcvq(sk, msg, size);
1270 if (tp->repair_queue == TCP_NO_QUEUE)
1273 /* 'common' sending to sendq */
1276 sockcm_init(&sockc, sk);
1277 if (msg->msg_controllen) {
1278 err = sock_cmsg_send(sk, msg, &sockc);
1279 if (unlikely(err)) {
1285 /* This should be in poll */
1286 sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
1288 /* Ok commence sending. */
1292 mss_now = tcp_send_mss(sk, &size_goal, flags);
1295 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
1298 while (msg_data_left(msg)) {
1301 skb = tcp_write_queue_tail(sk);
1303 copy = size_goal - skb->len;
1305 if (copy <= 0 || !tcp_skb_can_collapse_to(skb)) {
1309 if (!sk_stream_memory_free(sk))
1310 goto wait_for_space;
1312 if (unlikely(process_backlog >= 16)) {
1313 process_backlog = 0;
1314 if (sk_flush_backlog(sk))
1317 first_skb = tcp_rtx_and_write_queues_empty(sk);
1318 skb = tcp_stream_alloc_skb(sk, 0, sk->sk_allocation,
1321 goto wait_for_space;
1325 tcp_skb_entail(sk, skb);
1328 /* All packets are restored as if they have
1329 * already been sent. skb_mstamp_ns isn't set to
1330 * avoid wrong rtt estimation.
1333 TCP_SKB_CB(skb)->sacked |= TCPCB_REPAIRED;
1336 /* Try to append data to the end of skb. */
1337 if (copy > msg_data_left(msg))
1338 copy = msg_data_left(msg);
1342 int i = skb_shinfo(skb)->nr_frags;
1343 struct page_frag *pfrag = sk_page_frag(sk);
1345 if (!sk_page_frag_refill(sk, pfrag))
1346 goto wait_for_space;
1348 if (!skb_can_coalesce(skb, i, pfrag->page,
1350 if (i >= sysctl_max_skb_frags) {
1351 tcp_mark_push(tp, skb);
1357 copy = min_t(int, copy, pfrag->size - pfrag->offset);
1359 if (tcp_downgrade_zcopy_pure(sk, skb))
1360 goto wait_for_space;
1362 copy = tcp_wmem_schedule(sk, copy);
1364 goto wait_for_space;
1366 err = skb_copy_to_page_nocache(sk, &msg->msg_iter, skb,
1373 /* Update the skb. */
1375 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1377 skb_fill_page_desc(skb, i, pfrag->page,
1378 pfrag->offset, copy);
1379 page_ref_inc(pfrag->page);
1381 pfrag->offset += copy;
1383 /* First append to a fragless skb builds initial
1387 skb_shinfo(skb)->flags |= SKBFL_PURE_ZEROCOPY;
1389 if (!skb_zcopy_pure(skb)) {
1390 copy = tcp_wmem_schedule(sk, copy);
1392 goto wait_for_space;
1395 err = skb_zerocopy_iter_stream(sk, skb, msg, copy, uarg);
1396 if (err == -EMSGSIZE || err == -EEXIST) {
1397 tcp_mark_push(tp, skb);
1406 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
1408 WRITE_ONCE(tp->write_seq, tp->write_seq + copy);
1409 TCP_SKB_CB(skb)->end_seq += copy;
1410 tcp_skb_pcount_set(skb, 0);
1413 if (!msg_data_left(msg)) {
1414 if (unlikely(flags & MSG_EOR))
1415 TCP_SKB_CB(skb)->eor = 1;
1419 if (skb->len < size_goal || (flags & MSG_OOB) || unlikely(tp->repair))
1422 if (forced_push(tp)) {
1423 tcp_mark_push(tp, skb);
1424 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
1425 } else if (skb == tcp_send_head(sk))
1426 tcp_push_one(sk, mss_now);
1430 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1432 tcp_push(sk, flags & ~MSG_MORE, mss_now,
1433 TCP_NAGLE_PUSH, size_goal);
1435 err = sk_stream_wait_memory(sk, &timeo);
1439 mss_now = tcp_send_mss(sk, &size_goal, flags);
1444 tcp_tx_timestamp(sk, sockc.tsflags);
1445 tcp_push(sk, flags, mss_now, tp->nonagle, size_goal);
1448 net_zcopy_put(uarg);
1449 return copied + copied_syn;
1452 tcp_remove_empty_skb(sk);
1454 if (copied + copied_syn)
1457 net_zcopy_put_abort(uarg, true);
1458 err = sk_stream_error(sk, flags, err);
1459 /* make sure we wake any epoll edge trigger waiter */
1460 if (unlikely(tcp_rtx_and_write_queues_empty(sk) && err == -EAGAIN)) {
1461 sk->sk_write_space(sk);
1462 tcp_chrono_stop(sk, TCP_CHRONO_SNDBUF_LIMITED);
1466 EXPORT_SYMBOL_GPL(tcp_sendmsg_locked);
1468 int tcp_sendmsg(struct sock *sk, struct msghdr *msg, size_t size)
1473 ret = tcp_sendmsg_locked(sk, msg, size);
1478 EXPORT_SYMBOL(tcp_sendmsg);
1481 * Handle reading urgent data. BSD has very simple semantics for
1482 * this, no blocking and very strange errors 8)
1485 static int tcp_recv_urg(struct sock *sk, struct msghdr *msg, int len, int flags)
1487 struct tcp_sock *tp = tcp_sk(sk);
1489 /* No URG data to read. */
1490 if (sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data ||
1491 tp->urg_data == TCP_URG_READ)
1492 return -EINVAL; /* Yes this is right ! */
1494 if (sk->sk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DONE))
1497 if (tp->urg_data & TCP_URG_VALID) {
1499 char c = tp->urg_data;
1501 if (!(flags & MSG_PEEK))
1502 WRITE_ONCE(tp->urg_data, TCP_URG_READ);
1504 /* Read urgent data. */
1505 msg->msg_flags |= MSG_OOB;
1508 if (!(flags & MSG_TRUNC))
1509 err = memcpy_to_msg(msg, &c, 1);
1512 msg->msg_flags |= MSG_TRUNC;
1514 return err ? -EFAULT : len;
1517 if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
1520 /* Fixed the recv(..., MSG_OOB) behaviour. BSD docs and
1521 * the available implementations agree in this case:
1522 * this call should never block, independent of the
1523 * blocking state of the socket.
1524 * Mike <pall@rz.uni-karlsruhe.de>
1529 static int tcp_peek_sndq(struct sock *sk, struct msghdr *msg, int len)
1531 struct sk_buff *skb;
1532 int copied = 0, err = 0;
1534 /* XXX -- need to support SO_PEEK_OFF */
1536 skb_rbtree_walk(skb, &sk->tcp_rtx_queue) {
1537 err = skb_copy_datagram_msg(skb, 0, msg, skb->len);
1543 skb_queue_walk(&sk->sk_write_queue, skb) {
1544 err = skb_copy_datagram_msg(skb, 0, msg, skb->len);
1551 return err ?: copied;
1554 /* Clean up the receive buffer for full frames taken by the user,
1555 * then send an ACK if necessary. COPIED is the number of bytes
1556 * tcp_recvmsg has given to the user so far, it speeds up the
1557 * calculation of whether or not we must ACK for the sake of
1560 void tcp_cleanup_rbuf(struct sock *sk, int copied)
1562 struct tcp_sock *tp = tcp_sk(sk);
1563 bool time_to_ack = false;
1565 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
1567 WARN(skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq),
1568 "cleanup rbuf bug: copied %X seq %X rcvnxt %X\n",
1569 tp->copied_seq, TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt);
1571 if (inet_csk_ack_scheduled(sk)) {
1572 const struct inet_connection_sock *icsk = inet_csk(sk);
1574 if (/* Once-per-two-segments ACK was not sent by tcp_input.c */
1575 tp->rcv_nxt - tp->rcv_wup > icsk->icsk_ack.rcv_mss ||
1577 * If this read emptied read buffer, we send ACK, if
1578 * connection is not bidirectional, user drained
1579 * receive buffer and there was a small segment
1583 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED2) ||
1584 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED) &&
1585 !inet_csk_in_pingpong_mode(sk))) &&
1586 !atomic_read(&sk->sk_rmem_alloc)))
1590 /* We send an ACK if we can now advertise a non-zero window
1591 * which has been raised "significantly".
1593 * Even if window raised up to infinity, do not send window open ACK
1594 * in states, where we will not receive more. It is useless.
1596 if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) {
1597 __u32 rcv_window_now = tcp_receive_window(tp);
1599 /* Optimize, __tcp_select_window() is not cheap. */
1600 if (2*rcv_window_now <= tp->window_clamp) {
1601 __u32 new_window = __tcp_select_window(sk);
1603 /* Send ACK now, if this read freed lots of space
1604 * in our buffer. Certainly, new_window is new window.
1605 * We can advertise it now, if it is not less than current one.
1606 * "Lots" means "at least twice" here.
1608 if (new_window && new_window >= 2 * rcv_window_now)
1616 static void tcp_eat_recv_skb(struct sock *sk, struct sk_buff *skb)
1618 __skb_unlink(skb, &sk->sk_receive_queue);
1619 if (likely(skb->destructor == sock_rfree)) {
1621 skb->destructor = NULL;
1623 return skb_attempt_defer_free(skb);
1628 static struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off)
1630 struct sk_buff *skb;
1633 while ((skb = skb_peek(&sk->sk_receive_queue)) != NULL) {
1634 offset = seq - TCP_SKB_CB(skb)->seq;
1635 if (unlikely(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
1636 pr_err_once("%s: found a SYN, please report !\n", __func__);
1639 if (offset < skb->len || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)) {
1643 /* This looks weird, but this can happen if TCP collapsing
1644 * splitted a fat GRO packet, while we released socket lock
1645 * in skb_splice_bits()
1647 tcp_eat_recv_skb(sk, skb);
1653 * This routine provides an alternative to tcp_recvmsg() for routines
1654 * that would like to handle copying from skbuffs directly in 'sendfile'
1657 * - It is assumed that the socket was locked by the caller.
1658 * - The routine does not block.
1659 * - At present, there is no support for reading OOB data
1660 * or for 'peeking' the socket using this routine
1661 * (although both would be easy to implement).
1663 int tcp_read_sock(struct sock *sk, read_descriptor_t *desc,
1664 sk_read_actor_t recv_actor)
1666 struct sk_buff *skb;
1667 struct tcp_sock *tp = tcp_sk(sk);
1668 u32 seq = tp->copied_seq;
1672 if (sk->sk_state == TCP_LISTEN)
1674 while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1675 if (offset < skb->len) {
1679 len = skb->len - offset;
1680 /* Stop reading if we hit a patch of urgent data */
1681 if (unlikely(tp->urg_data)) {
1682 u32 urg_offset = tp->urg_seq - seq;
1683 if (urg_offset < len)
1688 used = recv_actor(desc, skb, offset, len);
1694 if (WARN_ON_ONCE(used > len))
1700 /* If recv_actor drops the lock (e.g. TCP splice
1701 * receive) the skb pointer might be invalid when
1702 * getting here: tcp_collapse might have deleted it
1703 * while aggregating skbs from the socket queue.
1705 skb = tcp_recv_skb(sk, seq - 1, &offset);
1708 /* TCP coalescing might have appended data to the skb.
1709 * Try to splice more frags
1711 if (offset + 1 != skb->len)
1714 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) {
1715 tcp_eat_recv_skb(sk, skb);
1719 tcp_eat_recv_skb(sk, skb);
1722 WRITE_ONCE(tp->copied_seq, seq);
1724 WRITE_ONCE(tp->copied_seq, seq);
1726 tcp_rcv_space_adjust(sk);
1728 /* Clean up data we have read: This will do ACK frames. */
1730 tcp_recv_skb(sk, seq, &offset);
1731 tcp_cleanup_rbuf(sk, copied);
1735 EXPORT_SYMBOL(tcp_read_sock);
1737 int tcp_read_skb(struct sock *sk, skb_read_actor_t recv_actor)
1739 struct tcp_sock *tp = tcp_sk(sk);
1740 u32 seq = tp->copied_seq;
1741 struct sk_buff *skb;
1745 if (sk->sk_state == TCP_LISTEN)
1748 while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1751 __skb_unlink(skb, &sk->sk_receive_queue);
1752 used = recv_actor(sk, skb);
1761 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) {
1769 WRITE_ONCE(tp->copied_seq, seq);
1771 tcp_rcv_space_adjust(sk);
1773 /* Clean up data we have read: This will do ACK frames. */
1775 tcp_cleanup_rbuf(sk, copied);
1779 EXPORT_SYMBOL(tcp_read_skb);
1781 int tcp_peek_len(struct socket *sock)
1783 return tcp_inq(sock->sk);
1785 EXPORT_SYMBOL(tcp_peek_len);
1787 /* Make sure sk_rcvbuf is big enough to satisfy SO_RCVLOWAT hint */
1788 int tcp_set_rcvlowat(struct sock *sk, int val)
1792 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK)
1793 cap = sk->sk_rcvbuf >> 1;
1795 cap = sock_net(sk)->ipv4.sysctl_tcp_rmem[2] >> 1;
1796 val = min(val, cap);
1797 WRITE_ONCE(sk->sk_rcvlowat, val ? : 1);
1799 /* Check if we need to signal EPOLLIN right now */
1802 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK)
1806 if (val > sk->sk_rcvbuf) {
1807 WRITE_ONCE(sk->sk_rcvbuf, val);
1808 tcp_sk(sk)->window_clamp = tcp_win_from_space(sk, val);
1812 EXPORT_SYMBOL(tcp_set_rcvlowat);
1814 void tcp_update_recv_tstamps(struct sk_buff *skb,
1815 struct scm_timestamping_internal *tss)
1818 tss->ts[0] = ktime_to_timespec64(skb->tstamp);
1820 tss->ts[0] = (struct timespec64) {0};
1822 if (skb_hwtstamps(skb)->hwtstamp)
1823 tss->ts[2] = ktime_to_timespec64(skb_hwtstamps(skb)->hwtstamp);
1825 tss->ts[2] = (struct timespec64) {0};
1829 static const struct vm_operations_struct tcp_vm_ops = {
1832 int tcp_mmap(struct file *file, struct socket *sock,
1833 struct vm_area_struct *vma)
1835 if (vma->vm_flags & (VM_WRITE | VM_EXEC))
1837 vma->vm_flags &= ~(VM_MAYWRITE | VM_MAYEXEC);
1839 /* Instruct vm_insert_page() to not mmap_read_lock(mm) */
1840 vma->vm_flags |= VM_MIXEDMAP;
1842 vma->vm_ops = &tcp_vm_ops;
1845 EXPORT_SYMBOL(tcp_mmap);
1847 static skb_frag_t *skb_advance_to_frag(struct sk_buff *skb, u32 offset_skb,
1852 if (unlikely(offset_skb >= skb->len))
1855 offset_skb -= skb_headlen(skb);
1856 if ((int)offset_skb < 0 || skb_has_frag_list(skb))
1859 frag = skb_shinfo(skb)->frags;
1860 while (offset_skb) {
1861 if (skb_frag_size(frag) > offset_skb) {
1862 *offset_frag = offset_skb;
1865 offset_skb -= skb_frag_size(frag);
1872 static bool can_map_frag(const skb_frag_t *frag)
1874 return skb_frag_size(frag) == PAGE_SIZE && !skb_frag_off(frag);
1877 static int find_next_mappable_frag(const skb_frag_t *frag,
1878 int remaining_in_skb)
1882 if (likely(can_map_frag(frag)))
1885 while (offset < remaining_in_skb && !can_map_frag(frag)) {
1886 offset += skb_frag_size(frag);
1892 static void tcp_zerocopy_set_hint_for_skb(struct sock *sk,
1893 struct tcp_zerocopy_receive *zc,
1894 struct sk_buff *skb, u32 offset)
1896 u32 frag_offset, partial_frag_remainder = 0;
1897 int mappable_offset;
1900 /* worst case: skip to next skb. try to improve on this case below */
1901 zc->recv_skip_hint = skb->len - offset;
1903 /* Find the frag containing this offset (and how far into that frag) */
1904 frag = skb_advance_to_frag(skb, offset, &frag_offset);
1909 struct skb_shared_info *info = skb_shinfo(skb);
1911 /* We read part of the last frag, must recvmsg() rest of skb. */
1912 if (frag == &info->frags[info->nr_frags - 1])
1915 /* Else, we must at least read the remainder in this frag. */
1916 partial_frag_remainder = skb_frag_size(frag) - frag_offset;
1917 zc->recv_skip_hint -= partial_frag_remainder;
1921 /* partial_frag_remainder: If part way through a frag, must read rest.
1922 * mappable_offset: Bytes till next mappable frag, *not* counting bytes
1923 * in partial_frag_remainder.
1925 mappable_offset = find_next_mappable_frag(frag, zc->recv_skip_hint);
1926 zc->recv_skip_hint = mappable_offset + partial_frag_remainder;
1929 static int tcp_recvmsg_locked(struct sock *sk, struct msghdr *msg, size_t len,
1930 int flags, struct scm_timestamping_internal *tss,
1932 static int receive_fallback_to_copy(struct sock *sk,
1933 struct tcp_zerocopy_receive *zc, int inq,
1934 struct scm_timestamping_internal *tss)
1936 unsigned long copy_address = (unsigned long)zc->copybuf_address;
1937 struct msghdr msg = {};
1942 zc->recv_skip_hint = 0;
1944 if (copy_address != zc->copybuf_address)
1947 err = import_single_range(READ, (void __user *)copy_address,
1948 inq, &iov, &msg.msg_iter);
1952 err = tcp_recvmsg_locked(sk, &msg, inq, MSG_DONTWAIT,
1953 tss, &zc->msg_flags);
1957 zc->copybuf_len = err;
1958 if (likely(zc->copybuf_len)) {
1959 struct sk_buff *skb;
1962 skb = tcp_recv_skb(sk, tcp_sk(sk)->copied_seq, &offset);
1964 tcp_zerocopy_set_hint_for_skb(sk, zc, skb, offset);
1969 static int tcp_copy_straggler_data(struct tcp_zerocopy_receive *zc,
1970 struct sk_buff *skb, u32 copylen,
1971 u32 *offset, u32 *seq)
1973 unsigned long copy_address = (unsigned long)zc->copybuf_address;
1974 struct msghdr msg = {};
1978 if (copy_address != zc->copybuf_address)
1981 err = import_single_range(READ, (void __user *)copy_address,
1982 copylen, &iov, &msg.msg_iter);
1985 err = skb_copy_datagram_msg(skb, *offset, &msg, copylen);
1988 zc->recv_skip_hint -= copylen;
1991 return (__s32)copylen;
1994 static int tcp_zc_handle_leftover(struct tcp_zerocopy_receive *zc,
1996 struct sk_buff *skb,
1999 struct scm_timestamping_internal *tss)
2001 u32 offset, copylen = min_t(u32, copybuf_len, zc->recv_skip_hint);
2005 /* skb is null if inq < PAGE_SIZE. */
2007 offset = *seq - TCP_SKB_CB(skb)->seq;
2009 skb = tcp_recv_skb(sk, *seq, &offset);
2010 if (TCP_SKB_CB(skb)->has_rxtstamp) {
2011 tcp_update_recv_tstamps(skb, tss);
2012 zc->msg_flags |= TCP_CMSG_TS;
2016 zc->copybuf_len = tcp_copy_straggler_data(zc, skb, copylen, &offset,
2018 return zc->copybuf_len < 0 ? 0 : copylen;
2021 static int tcp_zerocopy_vm_insert_batch_error(struct vm_area_struct *vma,
2022 struct page **pending_pages,
2023 unsigned long pages_remaining,
2024 unsigned long *address,
2027 struct tcp_zerocopy_receive *zc,
2028 u32 total_bytes_to_map,
2031 /* At least one page did not map. Try zapping if we skipped earlier. */
2032 if (err == -EBUSY &&
2033 zc->flags & TCP_RECEIVE_ZEROCOPY_FLAG_TLB_CLEAN_HINT) {
2036 maybe_zap_len = total_bytes_to_map - /* All bytes to map */
2037 *length + /* Mapped or pending */
2038 (pages_remaining * PAGE_SIZE); /* Failed map. */
2039 zap_page_range(vma, *address, maybe_zap_len);
2044 unsigned long leftover_pages = pages_remaining;
2047 /* We called zap_page_range, try to reinsert. */
2048 err = vm_insert_pages(vma, *address,
2051 bytes_mapped = PAGE_SIZE * (leftover_pages - pages_remaining);
2052 *seq += bytes_mapped;
2053 *address += bytes_mapped;
2056 /* Either we were unable to zap, OR we zapped, retried an
2057 * insert, and still had an issue. Either ways, pages_remaining
2058 * is the number of pages we were unable to map, and we unroll
2059 * some state we speculatively touched before.
2061 const int bytes_not_mapped = PAGE_SIZE * pages_remaining;
2063 *length -= bytes_not_mapped;
2064 zc->recv_skip_hint += bytes_not_mapped;
2069 static int tcp_zerocopy_vm_insert_batch(struct vm_area_struct *vma,
2070 struct page **pages,
2071 unsigned int pages_to_map,
2072 unsigned long *address,
2075 struct tcp_zerocopy_receive *zc,
2076 u32 total_bytes_to_map)
2078 unsigned long pages_remaining = pages_to_map;
2079 unsigned int pages_mapped;
2080 unsigned int bytes_mapped;
2083 err = vm_insert_pages(vma, *address, pages, &pages_remaining);
2084 pages_mapped = pages_to_map - (unsigned int)pages_remaining;
2085 bytes_mapped = PAGE_SIZE * pages_mapped;
2086 /* Even if vm_insert_pages fails, it may have partially succeeded in
2087 * mapping (some but not all of the pages).
2089 *seq += bytes_mapped;
2090 *address += bytes_mapped;
2095 /* Error: maybe zap and retry + rollback state for failed inserts. */
2096 return tcp_zerocopy_vm_insert_batch_error(vma, pages + pages_mapped,
2097 pages_remaining, address, length, seq, zc, total_bytes_to_map,
2101 #define TCP_VALID_ZC_MSG_FLAGS (TCP_CMSG_TS)
2102 static void tcp_zc_finalize_rx_tstamp(struct sock *sk,
2103 struct tcp_zerocopy_receive *zc,
2104 struct scm_timestamping_internal *tss)
2106 unsigned long msg_control_addr;
2107 struct msghdr cmsg_dummy;
2109 msg_control_addr = (unsigned long)zc->msg_control;
2110 cmsg_dummy.msg_control = (void *)msg_control_addr;
2111 cmsg_dummy.msg_controllen =
2112 (__kernel_size_t)zc->msg_controllen;
2113 cmsg_dummy.msg_flags = in_compat_syscall()
2114 ? MSG_CMSG_COMPAT : 0;
2115 cmsg_dummy.msg_control_is_user = true;
2117 if (zc->msg_control == msg_control_addr &&
2118 zc->msg_controllen == cmsg_dummy.msg_controllen) {
2119 tcp_recv_timestamp(&cmsg_dummy, sk, tss);
2120 zc->msg_control = (__u64)
2121 ((uintptr_t)cmsg_dummy.msg_control);
2122 zc->msg_controllen =
2123 (__u64)cmsg_dummy.msg_controllen;
2124 zc->msg_flags = (__u32)cmsg_dummy.msg_flags;
2128 #define TCP_ZEROCOPY_PAGE_BATCH_SIZE 32
2129 static int tcp_zerocopy_receive(struct sock *sk,
2130 struct tcp_zerocopy_receive *zc,
2131 struct scm_timestamping_internal *tss)
2133 u32 length = 0, offset, vma_len, avail_len, copylen = 0;
2134 unsigned long address = (unsigned long)zc->address;
2135 struct page *pages[TCP_ZEROCOPY_PAGE_BATCH_SIZE];
2136 s32 copybuf_len = zc->copybuf_len;
2137 struct tcp_sock *tp = tcp_sk(sk);
2138 const skb_frag_t *frags = NULL;
2139 unsigned int pages_to_map = 0;
2140 struct vm_area_struct *vma;
2141 struct sk_buff *skb = NULL;
2142 u32 seq = tp->copied_seq;
2143 u32 total_bytes_to_map;
2144 int inq = tcp_inq(sk);
2147 zc->copybuf_len = 0;
2150 if (address & (PAGE_SIZE - 1) || address != zc->address)
2153 if (sk->sk_state == TCP_LISTEN)
2156 sock_rps_record_flow(sk);
2158 if (inq && inq <= copybuf_len)
2159 return receive_fallback_to_copy(sk, zc, inq, tss);
2161 if (inq < PAGE_SIZE) {
2163 zc->recv_skip_hint = inq;
2164 if (!inq && sock_flag(sk, SOCK_DONE))
2169 mmap_read_lock(current->mm);
2171 vma = vma_lookup(current->mm, address);
2172 if (!vma || vma->vm_ops != &tcp_vm_ops) {
2173 mmap_read_unlock(current->mm);
2176 vma_len = min_t(unsigned long, zc->length, vma->vm_end - address);
2177 avail_len = min_t(u32, vma_len, inq);
2178 total_bytes_to_map = avail_len & ~(PAGE_SIZE - 1);
2179 if (total_bytes_to_map) {
2180 if (!(zc->flags & TCP_RECEIVE_ZEROCOPY_FLAG_TLB_CLEAN_HINT))
2181 zap_page_range(vma, address, total_bytes_to_map);
2182 zc->length = total_bytes_to_map;
2183 zc->recv_skip_hint = 0;
2185 zc->length = avail_len;
2186 zc->recv_skip_hint = avail_len;
2189 while (length + PAGE_SIZE <= zc->length) {
2190 int mappable_offset;
2193 if (zc->recv_skip_hint < PAGE_SIZE) {
2197 if (zc->recv_skip_hint > 0)
2200 offset = seq - TCP_SKB_CB(skb)->seq;
2202 skb = tcp_recv_skb(sk, seq, &offset);
2205 if (TCP_SKB_CB(skb)->has_rxtstamp) {
2206 tcp_update_recv_tstamps(skb, tss);
2207 zc->msg_flags |= TCP_CMSG_TS;
2209 zc->recv_skip_hint = skb->len - offset;
2210 frags = skb_advance_to_frag(skb, offset, &offset_frag);
2211 if (!frags || offset_frag)
2215 mappable_offset = find_next_mappable_frag(frags,
2216 zc->recv_skip_hint);
2217 if (mappable_offset) {
2218 zc->recv_skip_hint = mappable_offset;
2221 page = skb_frag_page(frags);
2223 pages[pages_to_map++] = page;
2224 length += PAGE_SIZE;
2225 zc->recv_skip_hint -= PAGE_SIZE;
2227 if (pages_to_map == TCP_ZEROCOPY_PAGE_BATCH_SIZE ||
2228 zc->recv_skip_hint < PAGE_SIZE) {
2229 /* Either full batch, or we're about to go to next skb
2230 * (and we cannot unroll failed ops across skbs).
2232 ret = tcp_zerocopy_vm_insert_batch(vma, pages,
2236 total_bytes_to_map);
2243 ret = tcp_zerocopy_vm_insert_batch(vma, pages, pages_to_map,
2244 &address, &length, &seq,
2245 zc, total_bytes_to_map);
2248 mmap_read_unlock(current->mm);
2249 /* Try to copy straggler data. */
2251 copylen = tcp_zc_handle_leftover(zc, sk, skb, &seq, copybuf_len, tss);
2253 if (length + copylen) {
2254 WRITE_ONCE(tp->copied_seq, seq);
2255 tcp_rcv_space_adjust(sk);
2257 /* Clean up data we have read: This will do ACK frames. */
2258 tcp_recv_skb(sk, seq, &offset);
2259 tcp_cleanup_rbuf(sk, length + copylen);
2261 if (length == zc->length)
2262 zc->recv_skip_hint = 0;
2264 if (!zc->recv_skip_hint && sock_flag(sk, SOCK_DONE))
2267 zc->length = length;
2272 /* Similar to __sock_recv_timestamp, but does not require an skb */
2273 void tcp_recv_timestamp(struct msghdr *msg, const struct sock *sk,
2274 struct scm_timestamping_internal *tss)
2276 int new_tstamp = sock_flag(sk, SOCK_TSTAMP_NEW);
2277 bool has_timestamping = false;
2279 if (tss->ts[0].tv_sec || tss->ts[0].tv_nsec) {
2280 if (sock_flag(sk, SOCK_RCVTSTAMP)) {
2281 if (sock_flag(sk, SOCK_RCVTSTAMPNS)) {
2283 struct __kernel_timespec kts = {
2284 .tv_sec = tss->ts[0].tv_sec,
2285 .tv_nsec = tss->ts[0].tv_nsec,
2287 put_cmsg(msg, SOL_SOCKET, SO_TIMESTAMPNS_NEW,
2290 struct __kernel_old_timespec ts_old = {
2291 .tv_sec = tss->ts[0].tv_sec,
2292 .tv_nsec = tss->ts[0].tv_nsec,
2294 put_cmsg(msg, SOL_SOCKET, SO_TIMESTAMPNS_OLD,
2295 sizeof(ts_old), &ts_old);
2299 struct __kernel_sock_timeval stv = {
2300 .tv_sec = tss->ts[0].tv_sec,
2301 .tv_usec = tss->ts[0].tv_nsec / 1000,
2303 put_cmsg(msg, SOL_SOCKET, SO_TIMESTAMP_NEW,
2306 struct __kernel_old_timeval tv = {
2307 .tv_sec = tss->ts[0].tv_sec,
2308 .tv_usec = tss->ts[0].tv_nsec / 1000,
2310 put_cmsg(msg, SOL_SOCKET, SO_TIMESTAMP_OLD,
2316 if (sk->sk_tsflags & SOF_TIMESTAMPING_SOFTWARE)
2317 has_timestamping = true;
2319 tss->ts[0] = (struct timespec64) {0};
2322 if (tss->ts[2].tv_sec || tss->ts[2].tv_nsec) {
2323 if (sk->sk_tsflags & SOF_TIMESTAMPING_RAW_HARDWARE)
2324 has_timestamping = true;
2326 tss->ts[2] = (struct timespec64) {0};
2329 if (has_timestamping) {
2330 tss->ts[1] = (struct timespec64) {0};
2331 if (sock_flag(sk, SOCK_TSTAMP_NEW))
2332 put_cmsg_scm_timestamping64(msg, tss);
2334 put_cmsg_scm_timestamping(msg, tss);
2338 static int tcp_inq_hint(struct sock *sk)
2340 const struct tcp_sock *tp = tcp_sk(sk);
2341 u32 copied_seq = READ_ONCE(tp->copied_seq);
2342 u32 rcv_nxt = READ_ONCE(tp->rcv_nxt);
2345 inq = rcv_nxt - copied_seq;
2346 if (unlikely(inq < 0 || copied_seq != READ_ONCE(tp->copied_seq))) {
2348 inq = tp->rcv_nxt - tp->copied_seq;
2351 /* After receiving a FIN, tell the user-space to continue reading
2352 * by returning a non-zero inq.
2354 if (inq == 0 && sock_flag(sk, SOCK_DONE))
2360 * This routine copies from a sock struct into the user buffer.
2362 * Technical note: in 2.3 we work on _locked_ socket, so that
2363 * tricks with *seq access order and skb->users are not required.
2364 * Probably, code can be easily improved even more.
2367 static int tcp_recvmsg_locked(struct sock *sk, struct msghdr *msg, size_t len,
2368 int flags, struct scm_timestamping_internal *tss,
2371 struct tcp_sock *tp = tcp_sk(sk);
2377 int target; /* Read at least this many bytes */
2379 struct sk_buff *skb, *last;
2383 if (sk->sk_state == TCP_LISTEN)
2386 if (tp->recvmsg_inq) {
2387 *cmsg_flags = TCP_CMSG_INQ;
2388 msg->msg_get_inq = 1;
2390 timeo = sock_rcvtimeo(sk, flags & MSG_DONTWAIT);
2392 /* Urgent data needs to be handled specially. */
2393 if (flags & MSG_OOB)
2396 if (unlikely(tp->repair)) {
2398 if (!(flags & MSG_PEEK))
2401 if (tp->repair_queue == TCP_SEND_QUEUE)
2405 if (tp->repair_queue == TCP_NO_QUEUE)
2408 /* 'common' recv queue MSG_PEEK-ing */
2411 seq = &tp->copied_seq;
2412 if (flags & MSG_PEEK) {
2413 peek_seq = tp->copied_seq;
2417 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
2422 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
2423 if (unlikely(tp->urg_data) && tp->urg_seq == *seq) {
2426 if (signal_pending(current)) {
2427 copied = timeo ? sock_intr_errno(timeo) : -EAGAIN;
2432 /* Next get a buffer. */
2434 last = skb_peek_tail(&sk->sk_receive_queue);
2435 skb_queue_walk(&sk->sk_receive_queue, skb) {
2437 /* Now that we have two receive queues this
2440 if (WARN(before(*seq, TCP_SKB_CB(skb)->seq),
2441 "TCP recvmsg seq # bug: copied %X, seq %X, rcvnxt %X, fl %X\n",
2442 *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt,
2446 offset = *seq - TCP_SKB_CB(skb)->seq;
2447 if (unlikely(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
2448 pr_err_once("%s: found a SYN, please report !\n", __func__);
2451 if (offset < skb->len)
2453 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
2455 WARN(!(flags & MSG_PEEK),
2456 "TCP recvmsg seq # bug 2: copied %X, seq %X, rcvnxt %X, fl %X\n",
2457 *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt, flags);
2460 /* Well, if we have backlog, try to process it now yet. */
2462 if (copied >= target && !READ_ONCE(sk->sk_backlog.tail))
2468 sk->sk_state == TCP_CLOSE ||
2469 (sk->sk_shutdown & RCV_SHUTDOWN) ||
2470 signal_pending(current))
2473 if (sock_flag(sk, SOCK_DONE))
2477 copied = sock_error(sk);
2481 if (sk->sk_shutdown & RCV_SHUTDOWN)
2484 if (sk->sk_state == TCP_CLOSE) {
2485 /* This occurs when user tries to read
2486 * from never connected socket.
2497 if (signal_pending(current)) {
2498 copied = sock_intr_errno(timeo);
2503 if (copied >= target) {
2504 /* Do not sleep, just process backlog. */
2505 __sk_flush_backlog(sk);
2507 tcp_cleanup_rbuf(sk, copied);
2508 sk_wait_data(sk, &timeo, last);
2511 if ((flags & MSG_PEEK) &&
2512 (peek_seq - copied - urg_hole != tp->copied_seq)) {
2513 net_dbg_ratelimited("TCP(%s:%d): Application bug, race in MSG_PEEK\n",
2515 task_pid_nr(current));
2516 peek_seq = tp->copied_seq;
2521 /* Ok so how much can we use? */
2522 used = skb->len - offset;
2526 /* Do we have urgent data here? */
2527 if (unlikely(tp->urg_data)) {
2528 u32 urg_offset = tp->urg_seq - *seq;
2529 if (urg_offset < used) {
2531 if (!sock_flag(sk, SOCK_URGINLINE)) {
2532 WRITE_ONCE(*seq, *seq + 1);
2544 if (!(flags & MSG_TRUNC)) {
2545 err = skb_copy_datagram_msg(skb, offset, msg, used);
2547 /* Exception. Bailout! */
2554 WRITE_ONCE(*seq, *seq + used);
2558 tcp_rcv_space_adjust(sk);
2561 if (unlikely(tp->urg_data) && after(tp->copied_seq, tp->urg_seq)) {
2562 WRITE_ONCE(tp->urg_data, 0);
2563 tcp_fast_path_check(sk);
2566 if (TCP_SKB_CB(skb)->has_rxtstamp) {
2567 tcp_update_recv_tstamps(skb, tss);
2568 *cmsg_flags |= TCP_CMSG_TS;
2571 if (used + offset < skb->len)
2574 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
2576 if (!(flags & MSG_PEEK))
2577 tcp_eat_recv_skb(sk, skb);
2581 /* Process the FIN. */
2582 WRITE_ONCE(*seq, *seq + 1);
2583 if (!(flags & MSG_PEEK))
2584 tcp_eat_recv_skb(sk, skb);
2588 /* According to UNIX98, msg_name/msg_namelen are ignored
2589 * on connected socket. I was just happy when found this 8) --ANK
2592 /* Clean up data we have read: This will do ACK frames. */
2593 tcp_cleanup_rbuf(sk, copied);
2600 err = tcp_recv_urg(sk, msg, len, flags);
2604 err = tcp_peek_sndq(sk, msg, len);
2608 int tcp_recvmsg(struct sock *sk, struct msghdr *msg, size_t len, int flags,
2611 int cmsg_flags = 0, ret;
2612 struct scm_timestamping_internal tss;
2614 if (unlikely(flags & MSG_ERRQUEUE))
2615 return inet_recv_error(sk, msg, len, addr_len);
2617 if (sk_can_busy_loop(sk) &&
2618 skb_queue_empty_lockless(&sk->sk_receive_queue) &&
2619 sk->sk_state == TCP_ESTABLISHED)
2620 sk_busy_loop(sk, flags & MSG_DONTWAIT);
2623 ret = tcp_recvmsg_locked(sk, msg, len, flags, &tss, &cmsg_flags);
2626 if ((cmsg_flags || msg->msg_get_inq) && ret >= 0) {
2627 if (cmsg_flags & TCP_CMSG_TS)
2628 tcp_recv_timestamp(msg, sk, &tss);
2629 if (msg->msg_get_inq) {
2630 msg->msg_inq = tcp_inq_hint(sk);
2631 if (cmsg_flags & TCP_CMSG_INQ)
2632 put_cmsg(msg, SOL_TCP, TCP_CM_INQ,
2633 sizeof(msg->msg_inq), &msg->msg_inq);
2638 EXPORT_SYMBOL(tcp_recvmsg);
2640 void tcp_set_state(struct sock *sk, int state)
2642 int oldstate = sk->sk_state;
2644 /* We defined a new enum for TCP states that are exported in BPF
2645 * so as not force the internal TCP states to be frozen. The
2646 * following checks will detect if an internal state value ever
2647 * differs from the BPF value. If this ever happens, then we will
2648 * need to remap the internal value to the BPF value before calling
2649 * tcp_call_bpf_2arg.
2651 BUILD_BUG_ON((int)BPF_TCP_ESTABLISHED != (int)TCP_ESTABLISHED);
2652 BUILD_BUG_ON((int)BPF_TCP_SYN_SENT != (int)TCP_SYN_SENT);
2653 BUILD_BUG_ON((int)BPF_TCP_SYN_RECV != (int)TCP_SYN_RECV);
2654 BUILD_BUG_ON((int)BPF_TCP_FIN_WAIT1 != (int)TCP_FIN_WAIT1);
2655 BUILD_BUG_ON((int)BPF_TCP_FIN_WAIT2 != (int)TCP_FIN_WAIT2);
2656 BUILD_BUG_ON((int)BPF_TCP_TIME_WAIT != (int)TCP_TIME_WAIT);
2657 BUILD_BUG_ON((int)BPF_TCP_CLOSE != (int)TCP_CLOSE);
2658 BUILD_BUG_ON((int)BPF_TCP_CLOSE_WAIT != (int)TCP_CLOSE_WAIT);
2659 BUILD_BUG_ON((int)BPF_TCP_LAST_ACK != (int)TCP_LAST_ACK);
2660 BUILD_BUG_ON((int)BPF_TCP_LISTEN != (int)TCP_LISTEN);
2661 BUILD_BUG_ON((int)BPF_TCP_CLOSING != (int)TCP_CLOSING);
2662 BUILD_BUG_ON((int)BPF_TCP_NEW_SYN_RECV != (int)TCP_NEW_SYN_RECV);
2663 BUILD_BUG_ON((int)BPF_TCP_MAX_STATES != (int)TCP_MAX_STATES);
2665 /* bpf uapi header bpf.h defines an anonymous enum with values
2666 * BPF_TCP_* used by bpf programs. Currently gcc built vmlinux
2667 * is able to emit this enum in DWARF due to the above BUILD_BUG_ON.
2668 * But clang built vmlinux does not have this enum in DWARF
2669 * since clang removes the above code before generating IR/debuginfo.
2670 * Let us explicitly emit the type debuginfo to ensure the
2671 * above-mentioned anonymous enum in the vmlinux DWARF and hence BTF
2672 * regardless of which compiler is used.
2674 BTF_TYPE_EMIT_ENUM(BPF_TCP_ESTABLISHED);
2676 if (BPF_SOCK_OPS_TEST_FLAG(tcp_sk(sk), BPF_SOCK_OPS_STATE_CB_FLAG))
2677 tcp_call_bpf_2arg(sk, BPF_SOCK_OPS_STATE_CB, oldstate, state);
2680 case TCP_ESTABLISHED:
2681 if (oldstate != TCP_ESTABLISHED)
2682 TCP_INC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
2686 if (oldstate == TCP_CLOSE_WAIT || oldstate == TCP_ESTABLISHED)
2687 TCP_INC_STATS(sock_net(sk), TCP_MIB_ESTABRESETS);
2689 sk->sk_prot->unhash(sk);
2690 if (inet_csk(sk)->icsk_bind_hash &&
2691 !(sk->sk_userlocks & SOCK_BINDPORT_LOCK))
2695 if (oldstate == TCP_ESTABLISHED)
2696 TCP_DEC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
2699 /* Change state AFTER socket is unhashed to avoid closed
2700 * socket sitting in hash tables.
2702 inet_sk_state_store(sk, state);
2704 EXPORT_SYMBOL_GPL(tcp_set_state);
2707 * State processing on a close. This implements the state shift for
2708 * sending our FIN frame. Note that we only send a FIN for some
2709 * states. A shutdown() may have already sent the FIN, or we may be
2713 static const unsigned char new_state[16] = {
2714 /* current state: new state: action: */
2715 [0 /* (Invalid) */] = TCP_CLOSE,
2716 [TCP_ESTABLISHED] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2717 [TCP_SYN_SENT] = TCP_CLOSE,
2718 [TCP_SYN_RECV] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2719 [TCP_FIN_WAIT1] = TCP_FIN_WAIT1,
2720 [TCP_FIN_WAIT2] = TCP_FIN_WAIT2,
2721 [TCP_TIME_WAIT] = TCP_CLOSE,
2722 [TCP_CLOSE] = TCP_CLOSE,
2723 [TCP_CLOSE_WAIT] = TCP_LAST_ACK | TCP_ACTION_FIN,
2724 [TCP_LAST_ACK] = TCP_LAST_ACK,
2725 [TCP_LISTEN] = TCP_CLOSE,
2726 [TCP_CLOSING] = TCP_CLOSING,
2727 [TCP_NEW_SYN_RECV] = TCP_CLOSE, /* should not happen ! */
2730 static int tcp_close_state(struct sock *sk)
2732 int next = (int)new_state[sk->sk_state];
2733 int ns = next & TCP_STATE_MASK;
2735 tcp_set_state(sk, ns);
2737 return next & TCP_ACTION_FIN;
2741 * Shutdown the sending side of a connection. Much like close except
2742 * that we don't receive shut down or sock_set_flag(sk, SOCK_DEAD).
2745 void tcp_shutdown(struct sock *sk, int how)
2747 /* We need to grab some memory, and put together a FIN,
2748 * and then put it into the queue to be sent.
2749 * Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
2751 if (!(how & SEND_SHUTDOWN))
2754 /* If we've already sent a FIN, or it's a closed state, skip this. */
2755 if ((1 << sk->sk_state) &
2756 (TCPF_ESTABLISHED | TCPF_SYN_SENT |
2757 TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) {
2758 /* Clear out any half completed packets. FIN if needed. */
2759 if (tcp_close_state(sk))
2763 EXPORT_SYMBOL(tcp_shutdown);
2765 int tcp_orphan_count_sum(void)
2769 for_each_possible_cpu(i)
2770 total += per_cpu(tcp_orphan_count, i);
2772 return max(total, 0);
2775 static int tcp_orphan_cache;
2776 static struct timer_list tcp_orphan_timer;
2777 #define TCP_ORPHAN_TIMER_PERIOD msecs_to_jiffies(100)
2779 static void tcp_orphan_update(struct timer_list *unused)
2781 WRITE_ONCE(tcp_orphan_cache, tcp_orphan_count_sum());
2782 mod_timer(&tcp_orphan_timer, jiffies + TCP_ORPHAN_TIMER_PERIOD);
2785 static bool tcp_too_many_orphans(int shift)
2787 return READ_ONCE(tcp_orphan_cache) << shift > sysctl_tcp_max_orphans;
2790 bool tcp_check_oom(struct sock *sk, int shift)
2792 bool too_many_orphans, out_of_socket_memory;
2794 too_many_orphans = tcp_too_many_orphans(shift);
2795 out_of_socket_memory = tcp_out_of_memory(sk);
2797 if (too_many_orphans)
2798 net_info_ratelimited("too many orphaned sockets\n");
2799 if (out_of_socket_memory)
2800 net_info_ratelimited("out of memory -- consider tuning tcp_mem\n");
2801 return too_many_orphans || out_of_socket_memory;
2804 void __tcp_close(struct sock *sk, long timeout)
2806 struct sk_buff *skb;
2807 int data_was_unread = 0;
2810 sk->sk_shutdown = SHUTDOWN_MASK;
2812 if (sk->sk_state == TCP_LISTEN) {
2813 tcp_set_state(sk, TCP_CLOSE);
2816 inet_csk_listen_stop(sk);
2818 goto adjudge_to_death;
2821 /* We need to flush the recv. buffs. We do this only on the
2822 * descriptor close, not protocol-sourced closes, because the
2823 * reader process may not have drained the data yet!
2825 while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) {
2826 u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq;
2828 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
2830 data_was_unread += len;
2834 /* If socket has been already reset (e.g. in tcp_reset()) - kill it. */
2835 if (sk->sk_state == TCP_CLOSE)
2836 goto adjudge_to_death;
2838 /* As outlined in RFC 2525, section 2.17, we send a RST here because
2839 * data was lost. To witness the awful effects of the old behavior of
2840 * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk
2841 * GET in an FTP client, suspend the process, wait for the client to
2842 * advertise a zero window, then kill -9 the FTP client, wheee...
2843 * Note: timeout is always zero in such a case.
2845 if (unlikely(tcp_sk(sk)->repair)) {
2846 sk->sk_prot->disconnect(sk, 0);
2847 } else if (data_was_unread) {
2848 /* Unread data was tossed, zap the connection. */
2849 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONCLOSE);
2850 tcp_set_state(sk, TCP_CLOSE);
2851 tcp_send_active_reset(sk, sk->sk_allocation);
2852 } else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) {
2853 /* Check zero linger _after_ checking for unread data. */
2854 sk->sk_prot->disconnect(sk, 0);
2855 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
2856 } else if (tcp_close_state(sk)) {
2857 /* We FIN if the application ate all the data before
2858 * zapping the connection.
2861 /* RED-PEN. Formally speaking, we have broken TCP state
2862 * machine. State transitions:
2864 * TCP_ESTABLISHED -> TCP_FIN_WAIT1
2865 * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible)
2866 * TCP_CLOSE_WAIT -> TCP_LAST_ACK
2868 * are legal only when FIN has been sent (i.e. in window),
2869 * rather than queued out of window. Purists blame.
2871 * F.e. "RFC state" is ESTABLISHED,
2872 * if Linux state is FIN-WAIT-1, but FIN is still not sent.
2874 * The visible declinations are that sometimes
2875 * we enter time-wait state, when it is not required really
2876 * (harmless), do not send active resets, when they are
2877 * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
2878 * they look as CLOSING or LAST_ACK for Linux)
2879 * Probably, I missed some more holelets.
2881 * XXX (TFO) - To start off we don't support SYN+ACK+FIN
2882 * in a single packet! (May consider it later but will
2883 * probably need API support or TCP_CORK SYN-ACK until
2884 * data is written and socket is closed.)
2889 sk_stream_wait_close(sk, timeout);
2892 state = sk->sk_state;
2898 /* remove backlog if any, without releasing ownership. */
2901 this_cpu_inc(tcp_orphan_count);
2903 /* Have we already been destroyed by a softirq or backlog? */
2904 if (state != TCP_CLOSE && sk->sk_state == TCP_CLOSE)
2907 /* This is a (useful) BSD violating of the RFC. There is a
2908 * problem with TCP as specified in that the other end could
2909 * keep a socket open forever with no application left this end.
2910 * We use a 1 minute timeout (about the same as BSD) then kill
2911 * our end. If they send after that then tough - BUT: long enough
2912 * that we won't make the old 4*rto = almost no time - whoops
2915 * Nope, it was not mistake. It is really desired behaviour
2916 * f.e. on http servers, when such sockets are useless, but
2917 * consume significant resources. Let's do it with special
2918 * linger2 option. --ANK
2921 if (sk->sk_state == TCP_FIN_WAIT2) {
2922 struct tcp_sock *tp = tcp_sk(sk);
2923 if (tp->linger2 < 0) {
2924 tcp_set_state(sk, TCP_CLOSE);
2925 tcp_send_active_reset(sk, GFP_ATOMIC);
2926 __NET_INC_STATS(sock_net(sk),
2927 LINUX_MIB_TCPABORTONLINGER);
2929 const int tmo = tcp_fin_time(sk);
2931 if (tmo > TCP_TIMEWAIT_LEN) {
2932 inet_csk_reset_keepalive_timer(sk,
2933 tmo - TCP_TIMEWAIT_LEN);
2935 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
2940 if (sk->sk_state != TCP_CLOSE) {
2941 if (tcp_check_oom(sk, 0)) {
2942 tcp_set_state(sk, TCP_CLOSE);
2943 tcp_send_active_reset(sk, GFP_ATOMIC);
2944 __NET_INC_STATS(sock_net(sk),
2945 LINUX_MIB_TCPABORTONMEMORY);
2946 } else if (!check_net(sock_net(sk))) {
2947 /* Not possible to send reset; just close */
2948 tcp_set_state(sk, TCP_CLOSE);
2952 if (sk->sk_state == TCP_CLOSE) {
2953 struct request_sock *req;
2955 req = rcu_dereference_protected(tcp_sk(sk)->fastopen_rsk,
2956 lockdep_sock_is_held(sk));
2957 /* We could get here with a non-NULL req if the socket is
2958 * aborted (e.g., closed with unread data) before 3WHS
2962 reqsk_fastopen_remove(sk, req, false);
2963 inet_csk_destroy_sock(sk);
2965 /* Otherwise, socket is reprieved until protocol close. */
2972 void tcp_close(struct sock *sk, long timeout)
2975 __tcp_close(sk, timeout);
2979 EXPORT_SYMBOL(tcp_close);
2981 /* These states need RST on ABORT according to RFC793 */
2983 static inline bool tcp_need_reset(int state)
2985 return (1 << state) &
2986 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 |
2987 TCPF_FIN_WAIT2 | TCPF_SYN_RECV);
2990 static void tcp_rtx_queue_purge(struct sock *sk)
2992 struct rb_node *p = rb_first(&sk->tcp_rtx_queue);
2994 tcp_sk(sk)->highest_sack = NULL;
2996 struct sk_buff *skb = rb_to_skb(p);
2999 /* Since we are deleting whole queue, no need to
3000 * list_del(&skb->tcp_tsorted_anchor)
3002 tcp_rtx_queue_unlink(skb, sk);
3003 tcp_wmem_free_skb(sk, skb);
3007 void tcp_write_queue_purge(struct sock *sk)
3009 struct sk_buff *skb;
3011 tcp_chrono_stop(sk, TCP_CHRONO_BUSY);
3012 while ((skb = __skb_dequeue(&sk->sk_write_queue)) != NULL) {
3013 tcp_skb_tsorted_anchor_cleanup(skb);
3014 tcp_wmem_free_skb(sk, skb);
3016 tcp_rtx_queue_purge(sk);
3017 INIT_LIST_HEAD(&tcp_sk(sk)->tsorted_sent_queue);
3018 tcp_clear_all_retrans_hints(tcp_sk(sk));
3019 tcp_sk(sk)->packets_out = 0;
3020 inet_csk(sk)->icsk_backoff = 0;
3023 int tcp_disconnect(struct sock *sk, int flags)
3025 struct inet_sock *inet = inet_sk(sk);
3026 struct inet_connection_sock *icsk = inet_csk(sk);
3027 struct tcp_sock *tp = tcp_sk(sk);
3028 int old_state = sk->sk_state;
3031 if (old_state != TCP_CLOSE)
3032 tcp_set_state(sk, TCP_CLOSE);
3034 /* ABORT function of RFC793 */
3035 if (old_state == TCP_LISTEN) {
3036 inet_csk_listen_stop(sk);
3037 } else if (unlikely(tp->repair)) {
3038 sk->sk_err = ECONNABORTED;
3039 } else if (tcp_need_reset(old_state) ||
3040 (tp->snd_nxt != tp->write_seq &&
3041 (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK))) {
3042 /* The last check adjusts for discrepancy of Linux wrt. RFC
3045 tcp_send_active_reset(sk, gfp_any());
3046 sk->sk_err = ECONNRESET;
3047 } else if (old_state == TCP_SYN_SENT)
3048 sk->sk_err = ECONNRESET;
3050 tcp_clear_xmit_timers(sk);
3051 __skb_queue_purge(&sk->sk_receive_queue);
3052 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt);
3053 WRITE_ONCE(tp->urg_data, 0);
3054 tcp_write_queue_purge(sk);
3055 tcp_fastopen_active_disable_ofo_check(sk);
3056 skb_rbtree_purge(&tp->out_of_order_queue);
3058 inet->inet_dport = 0;
3060 if (!(sk->sk_userlocks & SOCK_BINDADDR_LOCK))
3061 inet_reset_saddr(sk);
3063 sk->sk_shutdown = 0;
3064 sock_reset_flag(sk, SOCK_DONE);
3066 tp->mdev_us = jiffies_to_usecs(TCP_TIMEOUT_INIT);
3067 tp->rcv_rtt_last_tsecr = 0;
3069 seq = tp->write_seq + tp->max_window + 2;
3072 WRITE_ONCE(tp->write_seq, seq);
3074 icsk->icsk_backoff = 0;
3075 icsk->icsk_probes_out = 0;
3076 icsk->icsk_probes_tstamp = 0;
3077 icsk->icsk_rto = TCP_TIMEOUT_INIT;
3078 icsk->icsk_rto_min = TCP_RTO_MIN;
3079 icsk->icsk_delack_max = TCP_DELACK_MAX;
3080 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
3081 tcp_snd_cwnd_set(tp, TCP_INIT_CWND);
3082 tp->snd_cwnd_cnt = 0;
3083 tp->window_clamp = 0;
3085 tp->delivered_ce = 0;
3086 if (icsk->icsk_ca_ops->release)
3087 icsk->icsk_ca_ops->release(sk);
3088 memset(icsk->icsk_ca_priv, 0, sizeof(icsk->icsk_ca_priv));
3089 icsk->icsk_ca_initialized = 0;
3090 tcp_set_ca_state(sk, TCP_CA_Open);
3091 tp->is_sack_reneg = 0;
3092 tcp_clear_retrans(tp);
3093 tp->total_retrans = 0;
3094 inet_csk_delack_init(sk);
3095 /* Initialize rcv_mss to TCP_MIN_MSS to avoid division by 0
3096 * issue in __tcp_select_window()
3098 icsk->icsk_ack.rcv_mss = TCP_MIN_MSS;
3099 memset(&tp->rx_opt, 0, sizeof(tp->rx_opt));
3101 dst_release(xchg((__force struct dst_entry **)&sk->sk_rx_dst, NULL));
3102 tcp_saved_syn_free(tp);
3103 tp->compressed_ack = 0;
3107 tp->bytes_acked = 0;
3108 tp->bytes_received = 0;
3109 tp->bytes_retrans = 0;
3110 tp->data_segs_in = 0;
3111 tp->data_segs_out = 0;
3112 tp->duplicate_sack[0].start_seq = 0;
3113 tp->duplicate_sack[0].end_seq = 0;
3116 tp->retrans_out = 0;
3118 tp->tlp_high_seq = 0;
3119 tp->last_oow_ack_time = 0;
3120 /* There's a bubble in the pipe until at least the first ACK. */
3121 tp->app_limited = ~0U;
3122 tp->rack.mstamp = 0;
3123 tp->rack.advanced = 0;
3124 tp->rack.reo_wnd_steps = 1;
3125 tp->rack.last_delivered = 0;
3126 tp->rack.reo_wnd_persist = 0;
3127 tp->rack.dsack_seen = 0;
3128 tp->syn_data_acked = 0;
3129 tp->rx_opt.saw_tstamp = 0;
3130 tp->rx_opt.dsack = 0;
3131 tp->rx_opt.num_sacks = 0;
3132 tp->rcv_ooopack = 0;
3135 /* Clean up fastopen related fields */
3136 tcp_free_fastopen_req(tp);
3137 inet->defer_connect = 0;
3138 tp->fastopen_client_fail = 0;
3140 WARN_ON(inet->inet_num && !icsk->icsk_bind_hash);
3142 if (sk->sk_frag.page) {
3143 put_page(sk->sk_frag.page);
3144 sk->sk_frag.page = NULL;
3145 sk->sk_frag.offset = 0;
3147 sk_error_report(sk);
3150 EXPORT_SYMBOL(tcp_disconnect);
3152 static inline bool tcp_can_repair_sock(const struct sock *sk)
3154 return ns_capable(sock_net(sk)->user_ns, CAP_NET_ADMIN) &&
3155 (sk->sk_state != TCP_LISTEN);
3158 static int tcp_repair_set_window(struct tcp_sock *tp, sockptr_t optbuf, int len)
3160 struct tcp_repair_window opt;
3165 if (len != sizeof(opt))
3168 if (copy_from_sockptr(&opt, optbuf, sizeof(opt)))
3171 if (opt.max_window < opt.snd_wnd)
3174 if (after(opt.snd_wl1, tp->rcv_nxt + opt.rcv_wnd))
3177 if (after(opt.rcv_wup, tp->rcv_nxt))
3180 tp->snd_wl1 = opt.snd_wl1;
3181 tp->snd_wnd = opt.snd_wnd;
3182 tp->max_window = opt.max_window;
3184 tp->rcv_wnd = opt.rcv_wnd;
3185 tp->rcv_wup = opt.rcv_wup;
3190 static int tcp_repair_options_est(struct sock *sk, sockptr_t optbuf,
3193 struct tcp_sock *tp = tcp_sk(sk);
3194 struct tcp_repair_opt opt;
3197 while (len >= sizeof(opt)) {
3198 if (copy_from_sockptr_offset(&opt, optbuf, offset, sizeof(opt)))
3201 offset += sizeof(opt);
3204 switch (opt.opt_code) {
3206 tp->rx_opt.mss_clamp = opt.opt_val;
3211 u16 snd_wscale = opt.opt_val & 0xFFFF;
3212 u16 rcv_wscale = opt.opt_val >> 16;
3214 if (snd_wscale > TCP_MAX_WSCALE || rcv_wscale > TCP_MAX_WSCALE)
3217 tp->rx_opt.snd_wscale = snd_wscale;
3218 tp->rx_opt.rcv_wscale = rcv_wscale;
3219 tp->rx_opt.wscale_ok = 1;
3222 case TCPOPT_SACK_PERM:
3223 if (opt.opt_val != 0)
3226 tp->rx_opt.sack_ok |= TCP_SACK_SEEN;
3228 case TCPOPT_TIMESTAMP:
3229 if (opt.opt_val != 0)
3232 tp->rx_opt.tstamp_ok = 1;
3240 DEFINE_STATIC_KEY_FALSE(tcp_tx_delay_enabled);
3241 EXPORT_SYMBOL(tcp_tx_delay_enabled);
3243 static void tcp_enable_tx_delay(void)
3245 if (!static_branch_unlikely(&tcp_tx_delay_enabled)) {
3246 static int __tcp_tx_delay_enabled = 0;
3248 if (cmpxchg(&__tcp_tx_delay_enabled, 0, 1) == 0) {
3249 static_branch_enable(&tcp_tx_delay_enabled);
3250 pr_info("TCP_TX_DELAY enabled\n");
3255 /* When set indicates to always queue non-full frames. Later the user clears
3256 * this option and we transmit any pending partial frames in the queue. This is
3257 * meant to be used alongside sendfile() to get properly filled frames when the
3258 * user (for example) must write out headers with a write() call first and then
3259 * use sendfile to send out the data parts.
3261 * TCP_CORK can be set together with TCP_NODELAY and it is stronger than
3264 void __tcp_sock_set_cork(struct sock *sk, bool on)
3266 struct tcp_sock *tp = tcp_sk(sk);
3269 tp->nonagle |= TCP_NAGLE_CORK;
3271 tp->nonagle &= ~TCP_NAGLE_CORK;
3272 if (tp->nonagle & TCP_NAGLE_OFF)
3273 tp->nonagle |= TCP_NAGLE_PUSH;
3274 tcp_push_pending_frames(sk);
3278 void tcp_sock_set_cork(struct sock *sk, bool on)
3281 __tcp_sock_set_cork(sk, on);
3284 EXPORT_SYMBOL(tcp_sock_set_cork);
3286 /* TCP_NODELAY is weaker than TCP_CORK, so that this option on corked socket is
3287 * remembered, but it is not activated until cork is cleared.
3289 * However, when TCP_NODELAY is set we make an explicit push, which overrides
3290 * even TCP_CORK for currently queued segments.
3292 void __tcp_sock_set_nodelay(struct sock *sk, bool on)
3295 tcp_sk(sk)->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH;
3296 tcp_push_pending_frames(sk);
3298 tcp_sk(sk)->nonagle &= ~TCP_NAGLE_OFF;
3302 void tcp_sock_set_nodelay(struct sock *sk)
3305 __tcp_sock_set_nodelay(sk, true);
3308 EXPORT_SYMBOL(tcp_sock_set_nodelay);
3310 static void __tcp_sock_set_quickack(struct sock *sk, int val)
3313 inet_csk_enter_pingpong_mode(sk);
3317 inet_csk_exit_pingpong_mode(sk);
3318 if ((1 << sk->sk_state) & (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) &&
3319 inet_csk_ack_scheduled(sk)) {
3320 inet_csk(sk)->icsk_ack.pending |= ICSK_ACK_PUSHED;
3321 tcp_cleanup_rbuf(sk, 1);
3323 inet_csk_enter_pingpong_mode(sk);
3327 void tcp_sock_set_quickack(struct sock *sk, int val)
3330 __tcp_sock_set_quickack(sk, val);
3333 EXPORT_SYMBOL(tcp_sock_set_quickack);
3335 int tcp_sock_set_syncnt(struct sock *sk, int val)
3337 if (val < 1 || val > MAX_TCP_SYNCNT)
3341 inet_csk(sk)->icsk_syn_retries = val;
3345 EXPORT_SYMBOL(tcp_sock_set_syncnt);
3347 void tcp_sock_set_user_timeout(struct sock *sk, u32 val)
3350 inet_csk(sk)->icsk_user_timeout = val;
3353 EXPORT_SYMBOL(tcp_sock_set_user_timeout);
3355 int tcp_sock_set_keepidle_locked(struct sock *sk, int val)
3357 struct tcp_sock *tp = tcp_sk(sk);
3359 if (val < 1 || val > MAX_TCP_KEEPIDLE)
3362 tp->keepalive_time = val * HZ;
3363 if (sock_flag(sk, SOCK_KEEPOPEN) &&
3364 !((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN))) {
3365 u32 elapsed = keepalive_time_elapsed(tp);
3367 if (tp->keepalive_time > elapsed)
3368 elapsed = tp->keepalive_time - elapsed;
3371 inet_csk_reset_keepalive_timer(sk, elapsed);
3377 int tcp_sock_set_keepidle(struct sock *sk, int val)
3382 err = tcp_sock_set_keepidle_locked(sk, val);
3386 EXPORT_SYMBOL(tcp_sock_set_keepidle);
3388 int tcp_sock_set_keepintvl(struct sock *sk, int val)
3390 if (val < 1 || val > MAX_TCP_KEEPINTVL)
3394 tcp_sk(sk)->keepalive_intvl = val * HZ;
3398 EXPORT_SYMBOL(tcp_sock_set_keepintvl);
3400 int tcp_sock_set_keepcnt(struct sock *sk, int val)
3402 if (val < 1 || val > MAX_TCP_KEEPCNT)
3406 tcp_sk(sk)->keepalive_probes = val;
3410 EXPORT_SYMBOL(tcp_sock_set_keepcnt);
3412 int tcp_set_window_clamp(struct sock *sk, int val)
3414 struct tcp_sock *tp = tcp_sk(sk);
3417 if (sk->sk_state != TCP_CLOSE)
3419 tp->window_clamp = 0;
3421 tp->window_clamp = val < SOCK_MIN_RCVBUF / 2 ?
3422 SOCK_MIN_RCVBUF / 2 : val;
3423 tp->rcv_ssthresh = min(tp->rcv_wnd, tp->window_clamp);
3429 * Socket option code for TCP.
3431 static int do_tcp_setsockopt(struct sock *sk, int level, int optname,
3432 sockptr_t optval, unsigned int optlen)
3434 struct tcp_sock *tp = tcp_sk(sk);
3435 struct inet_connection_sock *icsk = inet_csk(sk);
3436 struct net *net = sock_net(sk);
3440 /* These are data/string values, all the others are ints */
3442 case TCP_CONGESTION: {
3443 char name[TCP_CA_NAME_MAX];
3448 val = strncpy_from_sockptr(name, optval,
3449 min_t(long, TCP_CA_NAME_MAX-1, optlen));
3455 err = tcp_set_congestion_control(sk, name, true,
3456 ns_capable(sock_net(sk)->user_ns,
3462 char name[TCP_ULP_NAME_MAX];
3467 val = strncpy_from_sockptr(name, optval,
3468 min_t(long, TCP_ULP_NAME_MAX - 1,
3475 err = tcp_set_ulp(sk, name);
3479 case TCP_FASTOPEN_KEY: {
3480 __u8 key[TCP_FASTOPEN_KEY_BUF_LENGTH];
3481 __u8 *backup_key = NULL;
3483 /* Allow a backup key as well to facilitate key rotation
3484 * First key is the active one.
3486 if (optlen != TCP_FASTOPEN_KEY_LENGTH &&
3487 optlen != TCP_FASTOPEN_KEY_BUF_LENGTH)
3490 if (copy_from_sockptr(key, optval, optlen))
3493 if (optlen == TCP_FASTOPEN_KEY_BUF_LENGTH)
3494 backup_key = key + TCP_FASTOPEN_KEY_LENGTH;
3496 return tcp_fastopen_reset_cipher(net, sk, key, backup_key);
3503 if (optlen < sizeof(int))
3506 if (copy_from_sockptr(&val, optval, sizeof(val)))
3513 /* Values greater than interface MTU won't take effect. However
3514 * at the point when this call is done we typically don't yet
3515 * know which interface is going to be used
3517 if (val && (val < TCP_MIN_MSS || val > MAX_TCP_WINDOW)) {
3521 tp->rx_opt.user_mss = val;
3525 __tcp_sock_set_nodelay(sk, val);
3528 case TCP_THIN_LINEAR_TIMEOUTS:
3529 if (val < 0 || val > 1)
3535 case TCP_THIN_DUPACK:
3536 if (val < 0 || val > 1)
3541 if (!tcp_can_repair_sock(sk))
3543 else if (val == TCP_REPAIR_ON) {
3545 sk->sk_reuse = SK_FORCE_REUSE;
3546 tp->repair_queue = TCP_NO_QUEUE;
3547 } else if (val == TCP_REPAIR_OFF) {
3549 sk->sk_reuse = SK_NO_REUSE;
3550 tcp_send_window_probe(sk);
3551 } else if (val == TCP_REPAIR_OFF_NO_WP) {
3553 sk->sk_reuse = SK_NO_REUSE;
3559 case TCP_REPAIR_QUEUE:
3562 else if ((unsigned int)val < TCP_QUEUES_NR)
3563 tp->repair_queue = val;
3569 if (sk->sk_state != TCP_CLOSE) {
3571 } else if (tp->repair_queue == TCP_SEND_QUEUE) {
3572 if (!tcp_rtx_queue_empty(sk))
3575 WRITE_ONCE(tp->write_seq, val);
3576 } else if (tp->repair_queue == TCP_RECV_QUEUE) {
3577 if (tp->rcv_nxt != tp->copied_seq) {
3580 WRITE_ONCE(tp->rcv_nxt, val);
3581 WRITE_ONCE(tp->copied_seq, val);
3588 case TCP_REPAIR_OPTIONS:
3591 else if (sk->sk_state == TCP_ESTABLISHED)
3592 err = tcp_repair_options_est(sk, optval, optlen);
3598 __tcp_sock_set_cork(sk, val);
3602 err = tcp_sock_set_keepidle_locked(sk, val);
3605 if (val < 1 || val > MAX_TCP_KEEPINTVL)
3608 tp->keepalive_intvl = val * HZ;
3611 if (val < 1 || val > MAX_TCP_KEEPCNT)
3614 tp->keepalive_probes = val;
3617 if (val < 1 || val > MAX_TCP_SYNCNT)
3620 icsk->icsk_syn_retries = val;
3624 /* 0: disable, 1: enable, 2: start from ether_header */
3625 if (val < 0 || val > 2)
3634 else if (val > TCP_FIN_TIMEOUT_MAX / HZ)
3635 tp->linger2 = TCP_FIN_TIMEOUT_MAX;
3637 tp->linger2 = val * HZ;
3640 case TCP_DEFER_ACCEPT:
3641 /* Translate value in seconds to number of retransmits */
3642 icsk->icsk_accept_queue.rskq_defer_accept =
3643 secs_to_retrans(val, TCP_TIMEOUT_INIT / HZ,
3647 case TCP_WINDOW_CLAMP:
3648 err = tcp_set_window_clamp(sk, val);
3652 __tcp_sock_set_quickack(sk, val);
3655 #ifdef CONFIG_TCP_MD5SIG
3657 case TCP_MD5SIG_EXT:
3658 err = tp->af_specific->md5_parse(sk, optname, optval, optlen);
3661 case TCP_USER_TIMEOUT:
3662 /* Cap the max time in ms TCP will retry or probe the window
3663 * before giving up and aborting (ETIMEDOUT) a connection.
3668 icsk->icsk_user_timeout = val;
3672 if (val >= 0 && ((1 << sk->sk_state) & (TCPF_CLOSE |
3674 tcp_fastopen_init_key_once(net);
3676 fastopen_queue_tune(sk, val);
3681 case TCP_FASTOPEN_CONNECT:
3682 if (val > 1 || val < 0) {
3684 } else if (net->ipv4.sysctl_tcp_fastopen & TFO_CLIENT_ENABLE) {
3685 if (sk->sk_state == TCP_CLOSE)
3686 tp->fastopen_connect = val;
3693 case TCP_FASTOPEN_NO_COOKIE:
3694 if (val > 1 || val < 0)
3696 else if (!((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
3699 tp->fastopen_no_cookie = val;
3705 tp->tsoffset = val - tcp_time_stamp_raw();
3707 case TCP_REPAIR_WINDOW:
3708 err = tcp_repair_set_window(tp, optval, optlen);
3710 case TCP_NOTSENT_LOWAT:
3711 tp->notsent_lowat = val;
3712 sk->sk_write_space(sk);
3715 if (val > 1 || val < 0)
3718 tp->recvmsg_inq = val;
3722 tcp_enable_tx_delay();
3723 tp->tcp_tx_delay = val;
3734 int tcp_setsockopt(struct sock *sk, int level, int optname, sockptr_t optval,
3735 unsigned int optlen)
3737 const struct inet_connection_sock *icsk = inet_csk(sk);
3739 if (level != SOL_TCP)
3740 return icsk->icsk_af_ops->setsockopt(sk, level, optname,
3742 return do_tcp_setsockopt(sk, level, optname, optval, optlen);
3744 EXPORT_SYMBOL(tcp_setsockopt);
3746 static void tcp_get_info_chrono_stats(const struct tcp_sock *tp,
3747 struct tcp_info *info)
3749 u64 stats[__TCP_CHRONO_MAX], total = 0;
3752 for (i = TCP_CHRONO_BUSY; i < __TCP_CHRONO_MAX; ++i) {
3753 stats[i] = tp->chrono_stat[i - 1];
3754 if (i == tp->chrono_type)
3755 stats[i] += tcp_jiffies32 - tp->chrono_start;
3756 stats[i] *= USEC_PER_SEC / HZ;
3760 info->tcpi_busy_time = total;
3761 info->tcpi_rwnd_limited = stats[TCP_CHRONO_RWND_LIMITED];
3762 info->tcpi_sndbuf_limited = stats[TCP_CHRONO_SNDBUF_LIMITED];
3765 /* Return information about state of tcp endpoint in API format. */
3766 void tcp_get_info(struct sock *sk, struct tcp_info *info)
3768 const struct tcp_sock *tp = tcp_sk(sk); /* iff sk_type == SOCK_STREAM */
3769 const struct inet_connection_sock *icsk = inet_csk(sk);
3775 memset(info, 0, sizeof(*info));
3776 if (sk->sk_type != SOCK_STREAM)
3779 info->tcpi_state = inet_sk_state_load(sk);
3781 /* Report meaningful fields for all TCP states, including listeners */
3782 rate = READ_ONCE(sk->sk_pacing_rate);
3783 rate64 = (rate != ~0UL) ? rate : ~0ULL;
3784 info->tcpi_pacing_rate = rate64;
3786 rate = READ_ONCE(sk->sk_max_pacing_rate);
3787 rate64 = (rate != ~0UL) ? rate : ~0ULL;
3788 info->tcpi_max_pacing_rate = rate64;
3790 info->tcpi_reordering = tp->reordering;
3791 info->tcpi_snd_cwnd = tcp_snd_cwnd(tp);
3793 if (info->tcpi_state == TCP_LISTEN) {
3794 /* listeners aliased fields :
3795 * tcpi_unacked -> Number of children ready for accept()
3796 * tcpi_sacked -> max backlog
3798 info->tcpi_unacked = READ_ONCE(sk->sk_ack_backlog);
3799 info->tcpi_sacked = READ_ONCE(sk->sk_max_ack_backlog);
3803 slow = lock_sock_fast(sk);
3805 info->tcpi_ca_state = icsk->icsk_ca_state;
3806 info->tcpi_retransmits = icsk->icsk_retransmits;
3807 info->tcpi_probes = icsk->icsk_probes_out;
3808 info->tcpi_backoff = icsk->icsk_backoff;
3810 if (tp->rx_opt.tstamp_ok)
3811 info->tcpi_options |= TCPI_OPT_TIMESTAMPS;
3812 if (tcp_is_sack(tp))
3813 info->tcpi_options |= TCPI_OPT_SACK;
3814 if (tp->rx_opt.wscale_ok) {
3815 info->tcpi_options |= TCPI_OPT_WSCALE;
3816 info->tcpi_snd_wscale = tp->rx_opt.snd_wscale;
3817 info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale;
3820 if (tp->ecn_flags & TCP_ECN_OK)
3821 info->tcpi_options |= TCPI_OPT_ECN;
3822 if (tp->ecn_flags & TCP_ECN_SEEN)
3823 info->tcpi_options |= TCPI_OPT_ECN_SEEN;
3824 if (tp->syn_data_acked)
3825 info->tcpi_options |= TCPI_OPT_SYN_DATA;
3827 info->tcpi_rto = jiffies_to_usecs(icsk->icsk_rto);
3828 info->tcpi_ato = jiffies_to_usecs(icsk->icsk_ack.ato);
3829 info->tcpi_snd_mss = tp->mss_cache;
3830 info->tcpi_rcv_mss = icsk->icsk_ack.rcv_mss;
3832 info->tcpi_unacked = tp->packets_out;
3833 info->tcpi_sacked = tp->sacked_out;
3835 info->tcpi_lost = tp->lost_out;
3836 info->tcpi_retrans = tp->retrans_out;
3838 now = tcp_jiffies32;
3839 info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime);
3840 info->tcpi_last_data_recv = jiffies_to_msecs(now - icsk->icsk_ack.lrcvtime);
3841 info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp);
3843 info->tcpi_pmtu = icsk->icsk_pmtu_cookie;
3844 info->tcpi_rcv_ssthresh = tp->rcv_ssthresh;
3845 info->tcpi_rtt = tp->srtt_us >> 3;
3846 info->tcpi_rttvar = tp->mdev_us >> 2;
3847 info->tcpi_snd_ssthresh = tp->snd_ssthresh;
3848 info->tcpi_advmss = tp->advmss;
3850 info->tcpi_rcv_rtt = tp->rcv_rtt_est.rtt_us >> 3;
3851 info->tcpi_rcv_space = tp->rcvq_space.space;
3853 info->tcpi_total_retrans = tp->total_retrans;
3855 info->tcpi_bytes_acked = tp->bytes_acked;
3856 info->tcpi_bytes_received = tp->bytes_received;
3857 info->tcpi_notsent_bytes = max_t(int, 0, tp->write_seq - tp->snd_nxt);
3858 tcp_get_info_chrono_stats(tp, info);
3860 info->tcpi_segs_out = tp->segs_out;
3862 /* segs_in and data_segs_in can be updated from tcp_segs_in() from BH */
3863 info->tcpi_segs_in = READ_ONCE(tp->segs_in);
3864 info->tcpi_data_segs_in = READ_ONCE(tp->data_segs_in);
3866 info->tcpi_min_rtt = tcp_min_rtt(tp);
3867 info->tcpi_data_segs_out = tp->data_segs_out;
3869 info->tcpi_delivery_rate_app_limited = tp->rate_app_limited ? 1 : 0;
3870 rate64 = tcp_compute_delivery_rate(tp);
3872 info->tcpi_delivery_rate = rate64;
3873 info->tcpi_delivered = tp->delivered;
3874 info->tcpi_delivered_ce = tp->delivered_ce;
3875 info->tcpi_bytes_sent = tp->bytes_sent;
3876 info->tcpi_bytes_retrans = tp->bytes_retrans;
3877 info->tcpi_dsack_dups = tp->dsack_dups;
3878 info->tcpi_reord_seen = tp->reord_seen;
3879 info->tcpi_rcv_ooopack = tp->rcv_ooopack;
3880 info->tcpi_snd_wnd = tp->snd_wnd;
3881 info->tcpi_fastopen_client_fail = tp->fastopen_client_fail;
3882 unlock_sock_fast(sk, slow);
3884 EXPORT_SYMBOL_GPL(tcp_get_info);
3886 static size_t tcp_opt_stats_get_size(void)
3889 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_BUSY */
3890 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_RWND_LIMITED */
3891 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_SNDBUF_LIMITED */
3892 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_DATA_SEGS_OUT */
3893 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_TOTAL_RETRANS */
3894 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_PACING_RATE */
3895 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_DELIVERY_RATE */
3896 nla_total_size(sizeof(u32)) + /* TCP_NLA_SND_CWND */
3897 nla_total_size(sizeof(u32)) + /* TCP_NLA_REORDERING */
3898 nla_total_size(sizeof(u32)) + /* TCP_NLA_MIN_RTT */
3899 nla_total_size(sizeof(u8)) + /* TCP_NLA_RECUR_RETRANS */
3900 nla_total_size(sizeof(u8)) + /* TCP_NLA_DELIVERY_RATE_APP_LMT */
3901 nla_total_size(sizeof(u32)) + /* TCP_NLA_SNDQ_SIZE */
3902 nla_total_size(sizeof(u8)) + /* TCP_NLA_CA_STATE */
3903 nla_total_size(sizeof(u32)) + /* TCP_NLA_SND_SSTHRESH */
3904 nla_total_size(sizeof(u32)) + /* TCP_NLA_DELIVERED */
3905 nla_total_size(sizeof(u32)) + /* TCP_NLA_DELIVERED_CE */
3906 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_BYTES_SENT */
3907 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_BYTES_RETRANS */
3908 nla_total_size(sizeof(u32)) + /* TCP_NLA_DSACK_DUPS */
3909 nla_total_size(sizeof(u32)) + /* TCP_NLA_REORD_SEEN */
3910 nla_total_size(sizeof(u32)) + /* TCP_NLA_SRTT */
3911 nla_total_size(sizeof(u16)) + /* TCP_NLA_TIMEOUT_REHASH */
3912 nla_total_size(sizeof(u32)) + /* TCP_NLA_BYTES_NOTSENT */
3913 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_EDT */
3914 nla_total_size(sizeof(u8)) + /* TCP_NLA_TTL */
3918 /* Returns TTL or hop limit of an incoming packet from skb. */
3919 static u8 tcp_skb_ttl_or_hop_limit(const struct sk_buff *skb)
3921 if (skb->protocol == htons(ETH_P_IP))
3922 return ip_hdr(skb)->ttl;
3923 else if (skb->protocol == htons(ETH_P_IPV6))
3924 return ipv6_hdr(skb)->hop_limit;
3929 struct sk_buff *tcp_get_timestamping_opt_stats(const struct sock *sk,
3930 const struct sk_buff *orig_skb,
3931 const struct sk_buff *ack_skb)
3933 const struct tcp_sock *tp = tcp_sk(sk);
3934 struct sk_buff *stats;
3935 struct tcp_info info;
3939 stats = alloc_skb(tcp_opt_stats_get_size(), GFP_ATOMIC);
3943 tcp_get_info_chrono_stats(tp, &info);
3944 nla_put_u64_64bit(stats, TCP_NLA_BUSY,
3945 info.tcpi_busy_time, TCP_NLA_PAD);
3946 nla_put_u64_64bit(stats, TCP_NLA_RWND_LIMITED,
3947 info.tcpi_rwnd_limited, TCP_NLA_PAD);
3948 nla_put_u64_64bit(stats, TCP_NLA_SNDBUF_LIMITED,
3949 info.tcpi_sndbuf_limited, TCP_NLA_PAD);
3950 nla_put_u64_64bit(stats, TCP_NLA_DATA_SEGS_OUT,
3951 tp->data_segs_out, TCP_NLA_PAD);
3952 nla_put_u64_64bit(stats, TCP_NLA_TOTAL_RETRANS,
3953 tp->total_retrans, TCP_NLA_PAD);
3955 rate = READ_ONCE(sk->sk_pacing_rate);
3956 rate64 = (rate != ~0UL) ? rate : ~0ULL;
3957 nla_put_u64_64bit(stats, TCP_NLA_PACING_RATE, rate64, TCP_NLA_PAD);
3959 rate64 = tcp_compute_delivery_rate(tp);
3960 nla_put_u64_64bit(stats, TCP_NLA_DELIVERY_RATE, rate64, TCP_NLA_PAD);
3962 nla_put_u32(stats, TCP_NLA_SND_CWND, tcp_snd_cwnd(tp));
3963 nla_put_u32(stats, TCP_NLA_REORDERING, tp->reordering);
3964 nla_put_u32(stats, TCP_NLA_MIN_RTT, tcp_min_rtt(tp));
3966 nla_put_u8(stats, TCP_NLA_RECUR_RETRANS, inet_csk(sk)->icsk_retransmits);
3967 nla_put_u8(stats, TCP_NLA_DELIVERY_RATE_APP_LMT, !!tp->rate_app_limited);
3968 nla_put_u32(stats, TCP_NLA_SND_SSTHRESH, tp->snd_ssthresh);
3969 nla_put_u32(stats, TCP_NLA_DELIVERED, tp->delivered);
3970 nla_put_u32(stats, TCP_NLA_DELIVERED_CE, tp->delivered_ce);
3972 nla_put_u32(stats, TCP_NLA_SNDQ_SIZE, tp->write_seq - tp->snd_una);
3973 nla_put_u8(stats, TCP_NLA_CA_STATE, inet_csk(sk)->icsk_ca_state);
3975 nla_put_u64_64bit(stats, TCP_NLA_BYTES_SENT, tp->bytes_sent,
3977 nla_put_u64_64bit(stats, TCP_NLA_BYTES_RETRANS, tp->bytes_retrans,
3979 nla_put_u32(stats, TCP_NLA_DSACK_DUPS, tp->dsack_dups);
3980 nla_put_u32(stats, TCP_NLA_REORD_SEEN, tp->reord_seen);
3981 nla_put_u32(stats, TCP_NLA_SRTT, tp->srtt_us >> 3);
3982 nla_put_u16(stats, TCP_NLA_TIMEOUT_REHASH, tp->timeout_rehash);
3983 nla_put_u32(stats, TCP_NLA_BYTES_NOTSENT,
3984 max_t(int, 0, tp->write_seq - tp->snd_nxt));
3985 nla_put_u64_64bit(stats, TCP_NLA_EDT, orig_skb->skb_mstamp_ns,
3988 nla_put_u8(stats, TCP_NLA_TTL,
3989 tcp_skb_ttl_or_hop_limit(ack_skb));
3994 static int do_tcp_getsockopt(struct sock *sk, int level,
3995 int optname, char __user *optval, int __user *optlen)
3997 struct inet_connection_sock *icsk = inet_csk(sk);
3998 struct tcp_sock *tp = tcp_sk(sk);
3999 struct net *net = sock_net(sk);
4002 if (get_user(len, optlen))
4005 len = min_t(unsigned int, len, sizeof(int));
4012 val = tp->mss_cache;
4013 if (!val && ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
4014 val = tp->rx_opt.user_mss;
4016 val = tp->rx_opt.mss_clamp;
4019 val = !!(tp->nonagle&TCP_NAGLE_OFF);
4022 val = !!(tp->nonagle&TCP_NAGLE_CORK);
4025 val = keepalive_time_when(tp) / HZ;
4028 val = keepalive_intvl_when(tp) / HZ;
4031 val = keepalive_probes(tp);
4034 val = icsk->icsk_syn_retries ? : net->ipv4.sysctl_tcp_syn_retries;
4039 val = (val ? : net->ipv4.sysctl_tcp_fin_timeout) / HZ;
4041 case TCP_DEFER_ACCEPT:
4042 val = retrans_to_secs(icsk->icsk_accept_queue.rskq_defer_accept,
4043 TCP_TIMEOUT_INIT / HZ, TCP_RTO_MAX / HZ);
4045 case TCP_WINDOW_CLAMP:
4046 val = tp->window_clamp;
4049 struct tcp_info info;
4051 if (get_user(len, optlen))
4054 tcp_get_info(sk, &info);
4056 len = min_t(unsigned int, len, sizeof(info));
4057 if (put_user(len, optlen))
4059 if (copy_to_user(optval, &info, len))
4064 const struct tcp_congestion_ops *ca_ops;
4065 union tcp_cc_info info;
4069 if (get_user(len, optlen))
4072 ca_ops = icsk->icsk_ca_ops;
4073 if (ca_ops && ca_ops->get_info)
4074 sz = ca_ops->get_info(sk, ~0U, &attr, &info);
4076 len = min_t(unsigned int, len, sz);
4077 if (put_user(len, optlen))
4079 if (copy_to_user(optval, &info, len))
4084 val = !inet_csk_in_pingpong_mode(sk);
4087 case TCP_CONGESTION:
4088 if (get_user(len, optlen))
4090 len = min_t(unsigned int, len, TCP_CA_NAME_MAX);
4091 if (put_user(len, optlen))
4093 if (copy_to_user(optval, icsk->icsk_ca_ops->name, len))
4098 if (get_user(len, optlen))
4100 len = min_t(unsigned int, len, TCP_ULP_NAME_MAX);
4101 if (!icsk->icsk_ulp_ops) {
4102 if (put_user(0, optlen))
4106 if (put_user(len, optlen))
4108 if (copy_to_user(optval, icsk->icsk_ulp_ops->name, len))
4112 case TCP_FASTOPEN_KEY: {
4113 u64 key[TCP_FASTOPEN_KEY_BUF_LENGTH / sizeof(u64)];
4114 unsigned int key_len;
4116 if (get_user(len, optlen))
4119 key_len = tcp_fastopen_get_cipher(net, icsk, key) *
4120 TCP_FASTOPEN_KEY_LENGTH;
4121 len = min_t(unsigned int, len, key_len);
4122 if (put_user(len, optlen))
4124 if (copy_to_user(optval, key, len))
4128 case TCP_THIN_LINEAR_TIMEOUTS:
4132 case TCP_THIN_DUPACK:
4140 case TCP_REPAIR_QUEUE:
4142 val = tp->repair_queue;
4147 case TCP_REPAIR_WINDOW: {
4148 struct tcp_repair_window opt;
4150 if (get_user(len, optlen))
4153 if (len != sizeof(opt))
4159 opt.snd_wl1 = tp->snd_wl1;
4160 opt.snd_wnd = tp->snd_wnd;
4161 opt.max_window = tp->max_window;
4162 opt.rcv_wnd = tp->rcv_wnd;
4163 opt.rcv_wup = tp->rcv_wup;
4165 if (copy_to_user(optval, &opt, len))
4170 if (tp->repair_queue == TCP_SEND_QUEUE)
4171 val = tp->write_seq;
4172 else if (tp->repair_queue == TCP_RECV_QUEUE)
4178 case TCP_USER_TIMEOUT:
4179 val = icsk->icsk_user_timeout;
4183 val = icsk->icsk_accept_queue.fastopenq.max_qlen;
4186 case TCP_FASTOPEN_CONNECT:
4187 val = tp->fastopen_connect;
4190 case TCP_FASTOPEN_NO_COOKIE:
4191 val = tp->fastopen_no_cookie;
4195 val = tp->tcp_tx_delay;
4199 val = tcp_time_stamp_raw() + tp->tsoffset;
4201 case TCP_NOTSENT_LOWAT:
4202 val = tp->notsent_lowat;
4205 val = tp->recvmsg_inq;
4210 case TCP_SAVED_SYN: {
4211 if (get_user(len, optlen))
4215 if (tp->saved_syn) {
4216 if (len < tcp_saved_syn_len(tp->saved_syn)) {
4217 if (put_user(tcp_saved_syn_len(tp->saved_syn),
4225 len = tcp_saved_syn_len(tp->saved_syn);
4226 if (put_user(len, optlen)) {
4230 if (copy_to_user(optval, tp->saved_syn->data, len)) {
4234 tcp_saved_syn_free(tp);
4239 if (put_user(len, optlen))
4245 case TCP_ZEROCOPY_RECEIVE: {
4246 struct scm_timestamping_internal tss;
4247 struct tcp_zerocopy_receive zc = {};
4250 if (get_user(len, optlen))
4253 len < offsetofend(struct tcp_zerocopy_receive, length))
4255 if (unlikely(len > sizeof(zc))) {
4256 err = check_zeroed_user(optval + sizeof(zc),
4259 return err == 0 ? -EINVAL : err;
4261 if (put_user(len, optlen))
4264 if (copy_from_user(&zc, optval, len))
4268 if (zc.msg_flags & ~(TCP_VALID_ZC_MSG_FLAGS))
4271 err = tcp_zerocopy_receive(sk, &zc, &tss);
4272 err = BPF_CGROUP_RUN_PROG_GETSOCKOPT_KERN(sk, level, optname,
4275 if (len >= offsetofend(struct tcp_zerocopy_receive, msg_flags))
4276 goto zerocopy_rcv_cmsg;
4278 case offsetofend(struct tcp_zerocopy_receive, msg_flags):
4279 goto zerocopy_rcv_cmsg;
4280 case offsetofend(struct tcp_zerocopy_receive, msg_controllen):
4281 case offsetofend(struct tcp_zerocopy_receive, msg_control):
4282 case offsetofend(struct tcp_zerocopy_receive, flags):
4283 case offsetofend(struct tcp_zerocopy_receive, copybuf_len):
4284 case offsetofend(struct tcp_zerocopy_receive, copybuf_address):
4285 case offsetofend(struct tcp_zerocopy_receive, err):
4286 goto zerocopy_rcv_sk_err;
4287 case offsetofend(struct tcp_zerocopy_receive, inq):
4288 goto zerocopy_rcv_inq;
4289 case offsetofend(struct tcp_zerocopy_receive, length):
4291 goto zerocopy_rcv_out;
4294 if (zc.msg_flags & TCP_CMSG_TS)
4295 tcp_zc_finalize_rx_tstamp(sk, &zc, &tss);
4298 zerocopy_rcv_sk_err:
4300 zc.err = sock_error(sk);
4302 zc.inq = tcp_inq_hint(sk);
4304 if (!err && copy_to_user(optval, &zc, len))
4310 return -ENOPROTOOPT;
4313 if (put_user(len, optlen))
4315 if (copy_to_user(optval, &val, len))
4320 bool tcp_bpf_bypass_getsockopt(int level, int optname)
4322 /* TCP do_tcp_getsockopt has optimized getsockopt implementation
4323 * to avoid extra socket lock for TCP_ZEROCOPY_RECEIVE.
4325 if (level == SOL_TCP && optname == TCP_ZEROCOPY_RECEIVE)
4330 EXPORT_SYMBOL(tcp_bpf_bypass_getsockopt);
4332 int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval,
4335 struct inet_connection_sock *icsk = inet_csk(sk);
4337 if (level != SOL_TCP)
4338 return icsk->icsk_af_ops->getsockopt(sk, level, optname,
4340 return do_tcp_getsockopt(sk, level, optname, optval, optlen);
4342 EXPORT_SYMBOL(tcp_getsockopt);
4344 #ifdef CONFIG_TCP_MD5SIG
4345 static DEFINE_PER_CPU(struct tcp_md5sig_pool, tcp_md5sig_pool);
4346 static DEFINE_MUTEX(tcp_md5sig_mutex);
4347 static bool tcp_md5sig_pool_populated = false;
4349 static void __tcp_alloc_md5sig_pool(void)
4351 struct crypto_ahash *hash;
4354 hash = crypto_alloc_ahash("md5", 0, CRYPTO_ALG_ASYNC);
4358 for_each_possible_cpu(cpu) {
4359 void *scratch = per_cpu(tcp_md5sig_pool, cpu).scratch;
4360 struct ahash_request *req;
4363 scratch = kmalloc_node(sizeof(union tcp_md5sum_block) +
4364 sizeof(struct tcphdr),
4369 per_cpu(tcp_md5sig_pool, cpu).scratch = scratch;
4371 if (per_cpu(tcp_md5sig_pool, cpu).md5_req)
4374 req = ahash_request_alloc(hash, GFP_KERNEL);
4378 ahash_request_set_callback(req, 0, NULL, NULL);
4380 per_cpu(tcp_md5sig_pool, cpu).md5_req = req;
4382 /* before setting tcp_md5sig_pool_populated, we must commit all writes
4383 * to memory. See smp_rmb() in tcp_get_md5sig_pool()
4386 tcp_md5sig_pool_populated = true;
4389 bool tcp_alloc_md5sig_pool(void)
4391 if (unlikely(!tcp_md5sig_pool_populated)) {
4392 mutex_lock(&tcp_md5sig_mutex);
4394 if (!tcp_md5sig_pool_populated) {
4395 __tcp_alloc_md5sig_pool();
4396 if (tcp_md5sig_pool_populated)
4397 static_branch_inc(&tcp_md5_needed);
4400 mutex_unlock(&tcp_md5sig_mutex);
4402 return tcp_md5sig_pool_populated;
4404 EXPORT_SYMBOL(tcp_alloc_md5sig_pool);
4408 * tcp_get_md5sig_pool - get md5sig_pool for this user
4410 * We use percpu structure, so if we succeed, we exit with preemption
4411 * and BH disabled, to make sure another thread or softirq handling
4412 * wont try to get same context.
4414 struct tcp_md5sig_pool *tcp_get_md5sig_pool(void)
4418 if (tcp_md5sig_pool_populated) {
4419 /* coupled with smp_wmb() in __tcp_alloc_md5sig_pool() */
4421 return this_cpu_ptr(&tcp_md5sig_pool);
4426 EXPORT_SYMBOL(tcp_get_md5sig_pool);
4428 int tcp_md5_hash_skb_data(struct tcp_md5sig_pool *hp,
4429 const struct sk_buff *skb, unsigned int header_len)
4431 struct scatterlist sg;
4432 const struct tcphdr *tp = tcp_hdr(skb);
4433 struct ahash_request *req = hp->md5_req;
4435 const unsigned int head_data_len = skb_headlen(skb) > header_len ?
4436 skb_headlen(skb) - header_len : 0;
4437 const struct skb_shared_info *shi = skb_shinfo(skb);
4438 struct sk_buff *frag_iter;
4440 sg_init_table(&sg, 1);
4442 sg_set_buf(&sg, ((u8 *) tp) + header_len, head_data_len);
4443 ahash_request_set_crypt(req, &sg, NULL, head_data_len);
4444 if (crypto_ahash_update(req))
4447 for (i = 0; i < shi->nr_frags; ++i) {
4448 const skb_frag_t *f = &shi->frags[i];
4449 unsigned int offset = skb_frag_off(f);
4450 struct page *page = skb_frag_page(f) + (offset >> PAGE_SHIFT);
4452 sg_set_page(&sg, page, skb_frag_size(f),
4453 offset_in_page(offset));
4454 ahash_request_set_crypt(req, &sg, NULL, skb_frag_size(f));
4455 if (crypto_ahash_update(req))
4459 skb_walk_frags(skb, frag_iter)
4460 if (tcp_md5_hash_skb_data(hp, frag_iter, 0))
4465 EXPORT_SYMBOL(tcp_md5_hash_skb_data);
4467 int tcp_md5_hash_key(struct tcp_md5sig_pool *hp, const struct tcp_md5sig_key *key)
4469 u8 keylen = READ_ONCE(key->keylen); /* paired with WRITE_ONCE() in tcp_md5_do_add */
4470 struct scatterlist sg;
4472 sg_init_one(&sg, key->key, keylen);
4473 ahash_request_set_crypt(hp->md5_req, &sg, NULL, keylen);
4475 /* We use data_race() because tcp_md5_do_add() might change key->key under us */
4476 return data_race(crypto_ahash_update(hp->md5_req));
4478 EXPORT_SYMBOL(tcp_md5_hash_key);
4480 /* Called with rcu_read_lock() */
4481 enum skb_drop_reason
4482 tcp_inbound_md5_hash(const struct sock *sk, const struct sk_buff *skb,
4483 const void *saddr, const void *daddr,
4484 int family, int dif, int sdif)
4487 * This gets called for each TCP segment that arrives
4488 * so we want to be efficient.
4489 * We have 3 drop cases:
4490 * o No MD5 hash and one expected.
4491 * o MD5 hash and we're not expecting one.
4492 * o MD5 hash and its wrong.
4494 const __u8 *hash_location = NULL;
4495 struct tcp_md5sig_key *hash_expected;
4496 const struct tcphdr *th = tcp_hdr(skb);
4497 struct tcp_sock *tp = tcp_sk(sk);
4498 int genhash, l3index;
4501 /* sdif set, means packet ingressed via a device
4502 * in an L3 domain and dif is set to the l3mdev
4504 l3index = sdif ? dif : 0;
4506 hash_expected = tcp_md5_do_lookup(sk, l3index, saddr, family);
4507 hash_location = tcp_parse_md5sig_option(th);
4509 /* We've parsed the options - do we have a hash? */
4510 if (!hash_expected && !hash_location)
4511 return SKB_NOT_DROPPED_YET;
4513 if (hash_expected && !hash_location) {
4514 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMD5NOTFOUND);
4515 return SKB_DROP_REASON_TCP_MD5NOTFOUND;
4518 if (!hash_expected && hash_location) {
4519 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMD5UNEXPECTED);
4520 return SKB_DROP_REASON_TCP_MD5UNEXPECTED;
4523 /* check the signature */
4524 genhash = tp->af_specific->calc_md5_hash(newhash, hash_expected,
4527 if (genhash || memcmp(hash_location, newhash, 16) != 0) {
4528 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMD5FAILURE);
4529 if (family == AF_INET) {
4530 net_info_ratelimited("MD5 Hash failed for (%pI4, %d)->(%pI4, %d)%s L3 index %d\n",
4531 saddr, ntohs(th->source),
4532 daddr, ntohs(th->dest),
4533 genhash ? " tcp_v4_calc_md5_hash failed"
4536 net_info_ratelimited("MD5 Hash %s for [%pI6c]:%u->[%pI6c]:%u L3 index %d\n",
4537 genhash ? "failed" : "mismatch",
4538 saddr, ntohs(th->source),
4539 daddr, ntohs(th->dest), l3index);
4541 return SKB_DROP_REASON_TCP_MD5FAILURE;
4543 return SKB_NOT_DROPPED_YET;
4545 EXPORT_SYMBOL(tcp_inbound_md5_hash);
4549 void tcp_done(struct sock *sk)
4551 struct request_sock *req;
4553 /* We might be called with a new socket, after
4554 * inet_csk_prepare_forced_close() has been called
4555 * so we can not use lockdep_sock_is_held(sk)
4557 req = rcu_dereference_protected(tcp_sk(sk)->fastopen_rsk, 1);
4559 if (sk->sk_state == TCP_SYN_SENT || sk->sk_state == TCP_SYN_RECV)
4560 TCP_INC_STATS(sock_net(sk), TCP_MIB_ATTEMPTFAILS);
4562 tcp_set_state(sk, TCP_CLOSE);
4563 tcp_clear_xmit_timers(sk);
4565 reqsk_fastopen_remove(sk, req, false);
4567 sk->sk_shutdown = SHUTDOWN_MASK;
4569 if (!sock_flag(sk, SOCK_DEAD))
4570 sk->sk_state_change(sk);
4572 inet_csk_destroy_sock(sk);
4574 EXPORT_SYMBOL_GPL(tcp_done);
4576 int tcp_abort(struct sock *sk, int err)
4578 int state = inet_sk_state_load(sk);
4580 if (state == TCP_NEW_SYN_RECV) {
4581 struct request_sock *req = inet_reqsk(sk);
4584 inet_csk_reqsk_queue_drop(req->rsk_listener, req);
4588 if (state == TCP_TIME_WAIT) {
4589 struct inet_timewait_sock *tw = inet_twsk(sk);
4591 refcount_inc(&tw->tw_refcnt);
4593 inet_twsk_deschedule_put(tw);
4598 /* Don't race with userspace socket closes such as tcp_close. */
4601 if (sk->sk_state == TCP_LISTEN) {
4602 tcp_set_state(sk, TCP_CLOSE);
4603 inet_csk_listen_stop(sk);
4606 /* Don't race with BH socket closes such as inet_csk_listen_stop. */
4610 if (!sock_flag(sk, SOCK_DEAD)) {
4612 /* This barrier is coupled with smp_rmb() in tcp_poll() */
4614 sk_error_report(sk);
4615 if (tcp_need_reset(sk->sk_state))
4616 tcp_send_active_reset(sk, GFP_ATOMIC);
4622 tcp_write_queue_purge(sk);
4626 EXPORT_SYMBOL_GPL(tcp_abort);
4628 extern struct tcp_congestion_ops tcp_reno;
4630 static __initdata unsigned long thash_entries;
4631 static int __init set_thash_entries(char *str)
4638 ret = kstrtoul(str, 0, &thash_entries);
4644 __setup("thash_entries=", set_thash_entries);
4646 static void __init tcp_init_mem(void)
4648 unsigned long limit = nr_free_buffer_pages() / 16;
4650 limit = max(limit, 128UL);
4651 sysctl_tcp_mem[0] = limit / 4 * 3; /* 4.68 % */
4652 sysctl_tcp_mem[1] = limit; /* 6.25 % */
4653 sysctl_tcp_mem[2] = sysctl_tcp_mem[0] * 2; /* 9.37 % */
4656 void __init tcp_init(void)
4658 int max_rshare, max_wshare, cnt;
4659 unsigned long limit;
4662 BUILD_BUG_ON(TCP_MIN_SND_MSS <= MAX_TCP_OPTION_SPACE);
4663 BUILD_BUG_ON(sizeof(struct tcp_skb_cb) >
4664 sizeof_field(struct sk_buff, cb));
4666 percpu_counter_init(&tcp_sockets_allocated, 0, GFP_KERNEL);
4668 timer_setup(&tcp_orphan_timer, tcp_orphan_update, TIMER_DEFERRABLE);
4669 mod_timer(&tcp_orphan_timer, jiffies + TCP_ORPHAN_TIMER_PERIOD);
4671 inet_hashinfo2_init(&tcp_hashinfo, "tcp_listen_portaddr_hash",
4672 thash_entries, 21, /* one slot per 2 MB*/
4674 tcp_hashinfo.bind_bucket_cachep =
4675 kmem_cache_create("tcp_bind_bucket",
4676 sizeof(struct inet_bind_bucket), 0,
4677 SLAB_HWCACHE_ALIGN | SLAB_PANIC |
4681 /* Size and allocate the main established and bind bucket
4684 * The methodology is similar to that of the buffer cache.
4686 tcp_hashinfo.ehash =
4687 alloc_large_system_hash("TCP established",
4688 sizeof(struct inet_ehash_bucket),
4690 17, /* one slot per 128 KB of memory */
4693 &tcp_hashinfo.ehash_mask,
4695 thash_entries ? 0 : 512 * 1024);
4696 for (i = 0; i <= tcp_hashinfo.ehash_mask; i++)
4697 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].chain, i);
4699 if (inet_ehash_locks_alloc(&tcp_hashinfo))
4700 panic("TCP: failed to alloc ehash_locks");
4701 tcp_hashinfo.bhash =
4702 alloc_large_system_hash("TCP bind",
4703 sizeof(struct inet_bind_hashbucket),
4704 tcp_hashinfo.ehash_mask + 1,
4705 17, /* one slot per 128 KB of memory */
4707 &tcp_hashinfo.bhash_size,
4711 tcp_hashinfo.bhash_size = 1U << tcp_hashinfo.bhash_size;
4712 for (i = 0; i < tcp_hashinfo.bhash_size; i++) {
4713 spin_lock_init(&tcp_hashinfo.bhash[i].lock);
4714 INIT_HLIST_HEAD(&tcp_hashinfo.bhash[i].chain);
4718 cnt = tcp_hashinfo.ehash_mask + 1;
4719 sysctl_tcp_max_orphans = cnt / 2;
4722 /* Set per-socket limits to no more than 1/128 the pressure threshold */
4723 limit = nr_free_buffer_pages() << (PAGE_SHIFT - 7);
4724 max_wshare = min(4UL*1024*1024, limit);
4725 max_rshare = min(6UL*1024*1024, limit);
4727 init_net.ipv4.sysctl_tcp_wmem[0] = PAGE_SIZE;
4728 init_net.ipv4.sysctl_tcp_wmem[1] = 16*1024;
4729 init_net.ipv4.sysctl_tcp_wmem[2] = max(64*1024, max_wshare);
4731 init_net.ipv4.sysctl_tcp_rmem[0] = PAGE_SIZE;
4732 init_net.ipv4.sysctl_tcp_rmem[1] = 131072;
4733 init_net.ipv4.sysctl_tcp_rmem[2] = max(131072, max_rshare);
4735 pr_info("Hash tables configured (established %u bind %u)\n",
4736 tcp_hashinfo.ehash_mask + 1, tcp_hashinfo.bhash_size);
4740 BUG_ON(tcp_register_congestion_control(&tcp_reno) != 0);