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);
298 #if IS_ENABLED(CONFIG_SMC)
299 DEFINE_STATIC_KEY_FALSE(tcp_have_smc);
300 EXPORT_SYMBOL(tcp_have_smc);
304 * Current number of TCP sockets.
306 struct percpu_counter tcp_sockets_allocated ____cacheline_aligned_in_smp;
307 EXPORT_SYMBOL(tcp_sockets_allocated);
312 struct tcp_splice_state {
313 struct pipe_inode_info *pipe;
319 * Pressure flag: try to collapse.
320 * Technical note: it is used by multiple contexts non atomically.
321 * All the __sk_mem_schedule() is of this nature: accounting
322 * is strict, actions are advisory and have some latency.
324 unsigned long tcp_memory_pressure __read_mostly;
325 EXPORT_SYMBOL_GPL(tcp_memory_pressure);
327 void tcp_enter_memory_pressure(struct sock *sk)
331 if (READ_ONCE(tcp_memory_pressure))
337 if (!cmpxchg(&tcp_memory_pressure, 0, val))
338 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMEMORYPRESSURES);
340 EXPORT_SYMBOL_GPL(tcp_enter_memory_pressure);
342 void tcp_leave_memory_pressure(struct sock *sk)
346 if (!READ_ONCE(tcp_memory_pressure))
348 val = xchg(&tcp_memory_pressure, 0);
350 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPMEMORYPRESSURESCHRONO,
351 jiffies_to_msecs(jiffies - val));
353 EXPORT_SYMBOL_GPL(tcp_leave_memory_pressure);
355 /* Convert seconds to retransmits based on initial and max timeout */
356 static u8 secs_to_retrans(int seconds, int timeout, int rto_max)
361 int period = timeout;
364 while (seconds > period && res < 255) {
367 if (timeout > rto_max)
375 /* Convert retransmits to seconds based on initial and max timeout */
376 static int retrans_to_secs(u8 retrans, int timeout, int rto_max)
384 if (timeout > rto_max)
392 static u64 tcp_compute_delivery_rate(const struct tcp_sock *tp)
394 u32 rate = READ_ONCE(tp->rate_delivered);
395 u32 intv = READ_ONCE(tp->rate_interval_us);
399 rate64 = (u64)rate * tp->mss_cache * USEC_PER_SEC;
400 do_div(rate64, intv);
405 /* Address-family independent initialization for a tcp_sock.
407 * NOTE: A lot of things set to zero explicitly by call to
408 * sk_alloc() so need not be done here.
410 void tcp_init_sock(struct sock *sk)
412 struct inet_connection_sock *icsk = inet_csk(sk);
413 struct tcp_sock *tp = tcp_sk(sk);
415 tp->out_of_order_queue = RB_ROOT;
416 sk->tcp_rtx_queue = RB_ROOT;
417 tcp_init_xmit_timers(sk);
418 INIT_LIST_HEAD(&tp->tsq_node);
419 INIT_LIST_HEAD(&tp->tsorted_sent_queue);
421 icsk->icsk_rto = TCP_TIMEOUT_INIT;
422 icsk->icsk_rto_min = TCP_RTO_MIN;
423 icsk->icsk_delack_max = TCP_DELACK_MAX;
424 tp->mdev_us = jiffies_to_usecs(TCP_TIMEOUT_INIT);
425 minmax_reset(&tp->rtt_min, tcp_jiffies32, ~0U);
427 /* So many TCP implementations out there (incorrectly) count the
428 * initial SYN frame in their delayed-ACK and congestion control
429 * algorithms that we must have the following bandaid to talk
430 * efficiently to them. -DaveM
432 tcp_snd_cwnd_set(tp, TCP_INIT_CWND);
434 /* There's a bubble in the pipe until at least the first ACK. */
435 tp->app_limited = ~0U;
437 /* See draft-stevens-tcpca-spec-01 for discussion of the
438 * initialization of these values.
440 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
441 tp->snd_cwnd_clamp = ~0;
442 tp->mss_cache = TCP_MSS_DEFAULT;
444 tp->reordering = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_reordering);
445 tcp_assign_congestion_control(sk);
448 tp->rack.reo_wnd_steps = 1;
450 sk->sk_write_space = sk_stream_write_space;
451 sock_set_flag(sk, SOCK_USE_WRITE_QUEUE);
453 icsk->icsk_sync_mss = tcp_sync_mss;
455 WRITE_ONCE(sk->sk_sndbuf, READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_wmem[1]));
456 WRITE_ONCE(sk->sk_rcvbuf, READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_rmem[1]));
458 sk_sockets_allocated_inc(sk);
460 EXPORT_SYMBOL(tcp_init_sock);
462 static void tcp_tx_timestamp(struct sock *sk, u16 tsflags)
464 struct sk_buff *skb = tcp_write_queue_tail(sk);
466 if (tsflags && skb) {
467 struct skb_shared_info *shinfo = skb_shinfo(skb);
468 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
470 sock_tx_timestamp(sk, tsflags, &shinfo->tx_flags);
471 if (tsflags & SOF_TIMESTAMPING_TX_ACK)
472 tcb->txstamp_ack = 1;
473 if (tsflags & SOF_TIMESTAMPING_TX_RECORD_MASK)
474 shinfo->tskey = TCP_SKB_CB(skb)->seq + skb->len - 1;
478 static bool tcp_stream_is_readable(struct sock *sk, int target)
480 if (tcp_epollin_ready(sk, target))
482 return sk_is_readable(sk);
486 * Wait for a TCP event.
488 * Note that we don't need to lock the socket, as the upper poll layers
489 * take care of normal races (between the test and the event) and we don't
490 * go look at any of the socket buffers directly.
492 __poll_t tcp_poll(struct file *file, struct socket *sock, poll_table *wait)
495 struct sock *sk = sock->sk;
496 const struct tcp_sock *tp = tcp_sk(sk);
499 sock_poll_wait(file, sock, wait);
501 state = inet_sk_state_load(sk);
502 if (state == TCP_LISTEN)
503 return inet_csk_listen_poll(sk);
505 /* Socket is not locked. We are protected from async events
506 * by poll logic and correct handling of state changes
507 * made by other threads is impossible in any case.
513 * EPOLLHUP is certainly not done right. But poll() doesn't
514 * have a notion of HUP in just one direction, and for a
515 * socket the read side is more interesting.
517 * Some poll() documentation says that EPOLLHUP is incompatible
518 * with the EPOLLOUT/POLLWR flags, so somebody should check this
519 * all. But careful, it tends to be safer to return too many
520 * bits than too few, and you can easily break real applications
521 * if you don't tell them that something has hung up!
525 * Check number 1. EPOLLHUP is _UNMASKABLE_ event (see UNIX98 and
526 * our fs/select.c). It means that after we received EOF,
527 * poll always returns immediately, making impossible poll() on write()
528 * in state CLOSE_WAIT. One solution is evident --- to set EPOLLHUP
529 * if and only if shutdown has been made in both directions.
530 * Actually, it is interesting to look how Solaris and DUX
531 * solve this dilemma. I would prefer, if EPOLLHUP were maskable,
532 * then we could set it on SND_SHUTDOWN. BTW examples given
533 * in Stevens' books assume exactly this behaviour, it explains
534 * why EPOLLHUP is incompatible with EPOLLOUT. --ANK
536 * NOTE. Check for TCP_CLOSE is added. The goal is to prevent
537 * blocking on fresh not-connected or disconnected socket. --ANK
539 if (sk->sk_shutdown == SHUTDOWN_MASK || state == TCP_CLOSE)
541 if (sk->sk_shutdown & RCV_SHUTDOWN)
542 mask |= EPOLLIN | EPOLLRDNORM | EPOLLRDHUP;
544 /* Connected or passive Fast Open socket? */
545 if (state != TCP_SYN_SENT &&
546 (state != TCP_SYN_RECV || rcu_access_pointer(tp->fastopen_rsk))) {
547 int target = sock_rcvlowat(sk, 0, INT_MAX);
548 u16 urg_data = READ_ONCE(tp->urg_data);
550 if (unlikely(urg_data) &&
551 READ_ONCE(tp->urg_seq) == READ_ONCE(tp->copied_seq) &&
552 !sock_flag(sk, SOCK_URGINLINE))
555 if (tcp_stream_is_readable(sk, target))
556 mask |= EPOLLIN | EPOLLRDNORM;
558 if (!(sk->sk_shutdown & SEND_SHUTDOWN)) {
559 if (__sk_stream_is_writeable(sk, 1)) {
560 mask |= EPOLLOUT | EPOLLWRNORM;
561 } else { /* send SIGIO later */
562 sk_set_bit(SOCKWQ_ASYNC_NOSPACE, sk);
563 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
565 /* Race breaker. If space is freed after
566 * wspace test but before the flags are set,
567 * IO signal will be lost. Memory barrier
568 * pairs with the input side.
570 smp_mb__after_atomic();
571 if (__sk_stream_is_writeable(sk, 1))
572 mask |= EPOLLOUT | EPOLLWRNORM;
575 mask |= EPOLLOUT | EPOLLWRNORM;
577 if (urg_data & TCP_URG_VALID)
579 } else if (state == TCP_SYN_SENT && inet_sk(sk)->defer_connect) {
580 /* Active TCP fastopen socket with defer_connect
581 * Return EPOLLOUT so application can call write()
582 * in order for kernel to generate SYN+data
584 mask |= EPOLLOUT | EPOLLWRNORM;
586 /* This barrier is coupled with smp_wmb() in tcp_reset() */
588 if (sk->sk_err || !skb_queue_empty_lockless(&sk->sk_error_queue))
593 EXPORT_SYMBOL(tcp_poll);
595 int tcp_ioctl(struct sock *sk, int cmd, unsigned long arg)
597 struct tcp_sock *tp = tcp_sk(sk);
603 if (sk->sk_state == TCP_LISTEN)
606 slow = lock_sock_fast(sk);
608 unlock_sock_fast(sk, slow);
611 answ = READ_ONCE(tp->urg_data) &&
612 READ_ONCE(tp->urg_seq) == READ_ONCE(tp->copied_seq);
615 if (sk->sk_state == TCP_LISTEN)
618 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
621 answ = READ_ONCE(tp->write_seq) - tp->snd_una;
624 if (sk->sk_state == TCP_LISTEN)
627 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
630 answ = READ_ONCE(tp->write_seq) -
631 READ_ONCE(tp->snd_nxt);
637 return put_user(answ, (int __user *)arg);
639 EXPORT_SYMBOL(tcp_ioctl);
641 void tcp_mark_push(struct tcp_sock *tp, struct sk_buff *skb)
643 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
644 tp->pushed_seq = tp->write_seq;
647 static inline bool forced_push(const struct tcp_sock *tp)
649 return after(tp->write_seq, tp->pushed_seq + (tp->max_window >> 1));
652 void tcp_skb_entail(struct sock *sk, struct sk_buff *skb)
654 struct tcp_sock *tp = tcp_sk(sk);
655 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
657 tcb->seq = tcb->end_seq = tp->write_seq;
658 tcb->tcp_flags = TCPHDR_ACK;
659 __skb_header_release(skb);
660 tcp_add_write_queue_tail(sk, skb);
661 sk_wmem_queued_add(sk, skb->truesize);
662 sk_mem_charge(sk, skb->truesize);
663 if (tp->nonagle & TCP_NAGLE_PUSH)
664 tp->nonagle &= ~TCP_NAGLE_PUSH;
666 tcp_slow_start_after_idle_check(sk);
669 static inline void tcp_mark_urg(struct tcp_sock *tp, int flags)
672 tp->snd_up = tp->write_seq;
675 /* If a not yet filled skb is pushed, do not send it if
676 * we have data packets in Qdisc or NIC queues :
677 * Because TX completion will happen shortly, it gives a chance
678 * to coalesce future sendmsg() payload into this skb, without
679 * need for a timer, and with no latency trade off.
680 * As packets containing data payload have a bigger truesize
681 * than pure acks (dataless) packets, the last checks prevent
682 * autocorking if we only have an ACK in Qdisc/NIC queues,
683 * or if TX completion was delayed after we processed ACK packet.
685 static bool tcp_should_autocork(struct sock *sk, struct sk_buff *skb,
688 return skb->len < size_goal &&
689 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_autocorking) &&
690 !tcp_rtx_queue_empty(sk) &&
691 refcount_read(&sk->sk_wmem_alloc) > skb->truesize &&
692 tcp_skb_can_collapse_to(skb);
695 void tcp_push(struct sock *sk, int flags, int mss_now,
696 int nonagle, int size_goal)
698 struct tcp_sock *tp = tcp_sk(sk);
701 skb = tcp_write_queue_tail(sk);
704 if (!(flags & MSG_MORE) || forced_push(tp))
705 tcp_mark_push(tp, skb);
707 tcp_mark_urg(tp, flags);
709 if (tcp_should_autocork(sk, skb, size_goal)) {
711 /* avoid atomic op if TSQ_THROTTLED bit is already set */
712 if (!test_bit(TSQ_THROTTLED, &sk->sk_tsq_flags)) {
713 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPAUTOCORKING);
714 set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags);
716 /* It is possible TX completion already happened
717 * before we set TSQ_THROTTLED.
719 if (refcount_read(&sk->sk_wmem_alloc) > skb->truesize)
723 if (flags & MSG_MORE)
724 nonagle = TCP_NAGLE_CORK;
726 __tcp_push_pending_frames(sk, mss_now, nonagle);
729 static int tcp_splice_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb,
730 unsigned int offset, size_t len)
732 struct tcp_splice_state *tss = rd_desc->arg.data;
735 ret = skb_splice_bits(skb, skb->sk, offset, tss->pipe,
736 min(rd_desc->count, len), tss->flags);
738 rd_desc->count -= ret;
742 static int __tcp_splice_read(struct sock *sk, struct tcp_splice_state *tss)
744 /* Store TCP splice context information in read_descriptor_t. */
745 read_descriptor_t rd_desc = {
750 return tcp_read_sock(sk, &rd_desc, tcp_splice_data_recv);
754 * tcp_splice_read - splice data from TCP socket to a pipe
755 * @sock: socket to splice from
756 * @ppos: position (not valid)
757 * @pipe: pipe to splice to
758 * @len: number of bytes to splice
759 * @flags: splice modifier flags
762 * Will read pages from given socket and fill them into a pipe.
765 ssize_t tcp_splice_read(struct socket *sock, loff_t *ppos,
766 struct pipe_inode_info *pipe, size_t len,
769 struct sock *sk = sock->sk;
770 struct tcp_splice_state tss = {
779 sock_rps_record_flow(sk);
781 * We can't seek on a socket input
790 timeo = sock_rcvtimeo(sk, sock->file->f_flags & O_NONBLOCK);
792 ret = __tcp_splice_read(sk, &tss);
798 if (sock_flag(sk, SOCK_DONE))
801 ret = sock_error(sk);
804 if (sk->sk_shutdown & RCV_SHUTDOWN)
806 if (sk->sk_state == TCP_CLOSE) {
808 * This occurs when user tries to read
809 * from never connected socket.
818 /* if __tcp_splice_read() got nothing while we have
819 * an skb in receive queue, we do not want to loop.
820 * This might happen with URG data.
822 if (!skb_queue_empty(&sk->sk_receive_queue))
824 sk_wait_data(sk, &timeo, NULL);
825 if (signal_pending(current)) {
826 ret = sock_intr_errno(timeo);
839 if (sk->sk_err || sk->sk_state == TCP_CLOSE ||
840 (sk->sk_shutdown & RCV_SHUTDOWN) ||
841 signal_pending(current))
852 EXPORT_SYMBOL(tcp_splice_read);
854 struct sk_buff *tcp_stream_alloc_skb(struct sock *sk, int size, gfp_t gfp,
859 if (unlikely(tcp_under_memory_pressure(sk)))
860 sk_mem_reclaim_partial(sk);
862 skb = alloc_skb_fclone(size + MAX_TCP_HEADER, gfp);
866 skb->truesize = SKB_TRUESIZE(skb_end_offset(skb));
867 if (force_schedule) {
868 mem_scheduled = true;
869 sk_forced_mem_schedule(sk, skb->truesize);
871 mem_scheduled = sk_wmem_schedule(sk, skb->truesize);
873 if (likely(mem_scheduled)) {
874 skb_reserve(skb, MAX_TCP_HEADER);
875 skb->ip_summed = CHECKSUM_PARTIAL;
876 INIT_LIST_HEAD(&skb->tcp_tsorted_anchor);
881 sk->sk_prot->enter_memory_pressure(sk);
882 sk_stream_moderate_sndbuf(sk);
887 static unsigned int tcp_xmit_size_goal(struct sock *sk, u32 mss_now,
890 struct tcp_sock *tp = tcp_sk(sk);
891 u32 new_size_goal, size_goal;
896 /* Note : tcp_tso_autosize() will eventually split this later */
897 new_size_goal = tcp_bound_to_half_wnd(tp, sk->sk_gso_max_size);
899 /* We try hard to avoid divides here */
900 size_goal = tp->gso_segs * mss_now;
901 if (unlikely(new_size_goal < size_goal ||
902 new_size_goal >= size_goal + mss_now)) {
903 tp->gso_segs = min_t(u16, new_size_goal / mss_now,
904 sk->sk_gso_max_segs);
905 size_goal = tp->gso_segs * mss_now;
908 return max(size_goal, mss_now);
911 int tcp_send_mss(struct sock *sk, int *size_goal, int flags)
915 mss_now = tcp_current_mss(sk);
916 *size_goal = tcp_xmit_size_goal(sk, mss_now, !(flags & MSG_OOB));
921 /* In some cases, both sendpage() and sendmsg() could have added
922 * an skb to the write queue, but failed adding payload on it.
923 * We need to remove it to consume less memory, but more
924 * importantly be able to generate EPOLLOUT for Edge Trigger epoll()
927 void tcp_remove_empty_skb(struct sock *sk)
929 struct sk_buff *skb = tcp_write_queue_tail(sk);
931 if (skb && TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq) {
932 tcp_unlink_write_queue(skb, sk);
933 if (tcp_write_queue_empty(sk))
934 tcp_chrono_stop(sk, TCP_CHRONO_BUSY);
935 tcp_wmem_free_skb(sk, skb);
939 /* skb changing from pure zc to mixed, must charge zc */
940 static int tcp_downgrade_zcopy_pure(struct sock *sk, struct sk_buff *skb)
942 if (unlikely(skb_zcopy_pure(skb))) {
943 u32 extra = skb->truesize -
944 SKB_TRUESIZE(skb_end_offset(skb));
946 if (!sk_wmem_schedule(sk, extra))
949 sk_mem_charge(sk, extra);
950 skb_shinfo(skb)->flags &= ~SKBFL_PURE_ZEROCOPY;
955 static struct sk_buff *tcp_build_frag(struct sock *sk, int size_goal, int flags,
956 struct page *page, int offset, size_t *size)
958 struct sk_buff *skb = tcp_write_queue_tail(sk);
959 struct tcp_sock *tp = tcp_sk(sk);
963 if (!skb || (copy = size_goal - skb->len) <= 0 ||
964 !tcp_skb_can_collapse_to(skb)) {
966 if (!sk_stream_memory_free(sk))
969 skb = tcp_stream_alloc_skb(sk, 0, sk->sk_allocation,
970 tcp_rtx_and_write_queues_empty(sk));
974 #ifdef CONFIG_TLS_DEVICE
975 skb->decrypted = !!(flags & MSG_SENDPAGE_DECRYPTED);
977 tcp_skb_entail(sk, skb);
984 i = skb_shinfo(skb)->nr_frags;
985 can_coalesce = skb_can_coalesce(skb, i, page, offset);
986 if (!can_coalesce && i >= sysctl_max_skb_frags) {
987 tcp_mark_push(tp, skb);
990 if (tcp_downgrade_zcopy_pure(sk, skb) || !sk_wmem_schedule(sk, copy))
994 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
997 skb_fill_page_desc(skb, i, page, offset, copy);
1000 if (!(flags & MSG_NO_SHARED_FRAGS))
1001 skb_shinfo(skb)->flags |= SKBFL_SHARED_FRAG;
1004 skb->data_len += copy;
1005 skb->truesize += copy;
1006 sk_wmem_queued_add(sk, copy);
1007 sk_mem_charge(sk, copy);
1008 WRITE_ONCE(tp->write_seq, tp->write_seq + copy);
1009 TCP_SKB_CB(skb)->end_seq += copy;
1010 tcp_skb_pcount_set(skb, 0);
1016 ssize_t do_tcp_sendpages(struct sock *sk, struct page *page, int offset,
1017 size_t size, int flags)
1019 struct tcp_sock *tp = tcp_sk(sk);
1020 int mss_now, size_goal;
1023 long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
1025 if (IS_ENABLED(CONFIG_DEBUG_VM) &&
1026 WARN_ONCE(!sendpage_ok(page),
1027 "page must not be a Slab one and have page_count > 0"))
1030 /* Wait for a connection to finish. One exception is TCP Fast Open
1031 * (passive side) where data is allowed to be sent before a connection
1032 * is fully established.
1034 if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) &&
1035 !tcp_passive_fastopen(sk)) {
1036 err = sk_stream_wait_connect(sk, &timeo);
1041 sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
1043 mss_now = tcp_send_mss(sk, &size_goal, flags);
1047 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
1051 struct sk_buff *skb;
1054 skb = tcp_build_frag(sk, size_goal, flags, page, offset, ©);
1056 goto wait_for_space;
1059 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
1067 if (skb->len < size_goal || (flags & MSG_OOB))
1070 if (forced_push(tp)) {
1071 tcp_mark_push(tp, skb);
1072 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
1073 } else if (skb == tcp_send_head(sk))
1074 tcp_push_one(sk, mss_now);
1078 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1079 tcp_push(sk, flags & ~MSG_MORE, mss_now,
1080 TCP_NAGLE_PUSH, size_goal);
1082 err = sk_stream_wait_memory(sk, &timeo);
1086 mss_now = tcp_send_mss(sk, &size_goal, flags);
1091 tcp_tx_timestamp(sk, sk->sk_tsflags);
1092 if (!(flags & MSG_SENDPAGE_NOTLAST))
1093 tcp_push(sk, flags, mss_now, tp->nonagle, size_goal);
1098 tcp_remove_empty_skb(sk);
1102 /* make sure we wake any epoll edge trigger waiter */
1103 if (unlikely(tcp_rtx_and_write_queues_empty(sk) && err == -EAGAIN)) {
1104 sk->sk_write_space(sk);
1105 tcp_chrono_stop(sk, TCP_CHRONO_SNDBUF_LIMITED);
1107 return sk_stream_error(sk, flags, err);
1109 EXPORT_SYMBOL_GPL(do_tcp_sendpages);
1111 int tcp_sendpage_locked(struct sock *sk, struct page *page, int offset,
1112 size_t size, int flags)
1114 if (!(sk->sk_route_caps & NETIF_F_SG))
1115 return sock_no_sendpage_locked(sk, page, offset, size, flags);
1117 tcp_rate_check_app_limited(sk); /* is sending application-limited? */
1119 return do_tcp_sendpages(sk, page, offset, size, flags);
1121 EXPORT_SYMBOL_GPL(tcp_sendpage_locked);
1123 int tcp_sendpage(struct sock *sk, struct page *page, int offset,
1124 size_t size, int flags)
1129 ret = tcp_sendpage_locked(sk, page, offset, size, flags);
1134 EXPORT_SYMBOL(tcp_sendpage);
1136 void tcp_free_fastopen_req(struct tcp_sock *tp)
1138 if (tp->fastopen_req) {
1139 kfree(tp->fastopen_req);
1140 tp->fastopen_req = NULL;
1144 static int tcp_sendmsg_fastopen(struct sock *sk, struct msghdr *msg,
1145 int *copied, size_t size,
1146 struct ubuf_info *uarg)
1148 struct tcp_sock *tp = tcp_sk(sk);
1149 struct inet_sock *inet = inet_sk(sk);
1150 struct sockaddr *uaddr = msg->msg_name;
1153 if (!(READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_fastopen) &
1154 TFO_CLIENT_ENABLE) ||
1155 (uaddr && msg->msg_namelen >= sizeof(uaddr->sa_family) &&
1156 uaddr->sa_family == AF_UNSPEC))
1158 if (tp->fastopen_req)
1159 return -EALREADY; /* Another Fast Open is in progress */
1161 tp->fastopen_req = kzalloc(sizeof(struct tcp_fastopen_request),
1163 if (unlikely(!tp->fastopen_req))
1165 tp->fastopen_req->data = msg;
1166 tp->fastopen_req->size = size;
1167 tp->fastopen_req->uarg = uarg;
1169 if (inet->defer_connect) {
1170 err = tcp_connect(sk);
1171 /* Same failure procedure as in tcp_v4/6_connect */
1173 tcp_set_state(sk, TCP_CLOSE);
1174 inet->inet_dport = 0;
1175 sk->sk_route_caps = 0;
1178 flags = (msg->msg_flags & MSG_DONTWAIT) ? O_NONBLOCK : 0;
1179 err = __inet_stream_connect(sk->sk_socket, uaddr,
1180 msg->msg_namelen, flags, 1);
1181 /* fastopen_req could already be freed in __inet_stream_connect
1182 * if the connection times out or gets rst
1184 if (tp->fastopen_req) {
1185 *copied = tp->fastopen_req->copied;
1186 tcp_free_fastopen_req(tp);
1187 inet->defer_connect = 0;
1192 int tcp_sendmsg_locked(struct sock *sk, struct msghdr *msg, size_t size)
1194 struct tcp_sock *tp = tcp_sk(sk);
1195 struct ubuf_info *uarg = NULL;
1196 struct sk_buff *skb;
1197 struct sockcm_cookie sockc;
1198 int flags, err, copied = 0;
1199 int mss_now = 0, size_goal, copied_syn = 0;
1200 int process_backlog = 0;
1204 flags = msg->msg_flags;
1206 if ((flags & MSG_ZEROCOPY) && size) {
1207 skb = tcp_write_queue_tail(sk);
1209 if (msg->msg_ubuf) {
1210 uarg = msg->msg_ubuf;
1211 net_zcopy_get(uarg);
1212 zc = sk->sk_route_caps & NETIF_F_SG;
1213 } else if (sock_flag(sk, SOCK_ZEROCOPY)) {
1214 uarg = msg_zerocopy_realloc(sk, size, skb_zcopy(skb));
1219 zc = sk->sk_route_caps & NETIF_F_SG;
1225 if (unlikely(flags & MSG_FASTOPEN || inet_sk(sk)->defer_connect) &&
1227 err = tcp_sendmsg_fastopen(sk, msg, &copied_syn, size, uarg);
1228 if (err == -EINPROGRESS && copied_syn > 0)
1234 timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
1236 tcp_rate_check_app_limited(sk); /* is sending application-limited? */
1238 /* Wait for a connection to finish. One exception is TCP Fast Open
1239 * (passive side) where data is allowed to be sent before a connection
1240 * is fully established.
1242 if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) &&
1243 !tcp_passive_fastopen(sk)) {
1244 err = sk_stream_wait_connect(sk, &timeo);
1249 if (unlikely(tp->repair)) {
1250 if (tp->repair_queue == TCP_RECV_QUEUE) {
1251 copied = tcp_send_rcvq(sk, msg, size);
1256 if (tp->repair_queue == TCP_NO_QUEUE)
1259 /* 'common' sending to sendq */
1262 sockcm_init(&sockc, sk);
1263 if (msg->msg_controllen) {
1264 err = sock_cmsg_send(sk, msg, &sockc);
1265 if (unlikely(err)) {
1271 /* This should be in poll */
1272 sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
1274 /* Ok commence sending. */
1278 mss_now = tcp_send_mss(sk, &size_goal, flags);
1281 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
1284 while (msg_data_left(msg)) {
1287 skb = tcp_write_queue_tail(sk);
1289 copy = size_goal - skb->len;
1291 if (copy <= 0 || !tcp_skb_can_collapse_to(skb)) {
1295 if (!sk_stream_memory_free(sk))
1296 goto wait_for_space;
1298 if (unlikely(process_backlog >= 16)) {
1299 process_backlog = 0;
1300 if (sk_flush_backlog(sk))
1303 first_skb = tcp_rtx_and_write_queues_empty(sk);
1304 skb = tcp_stream_alloc_skb(sk, 0, sk->sk_allocation,
1307 goto wait_for_space;
1311 tcp_skb_entail(sk, skb);
1314 /* All packets are restored as if they have
1315 * already been sent. skb_mstamp_ns isn't set to
1316 * avoid wrong rtt estimation.
1319 TCP_SKB_CB(skb)->sacked |= TCPCB_REPAIRED;
1322 /* Try to append data to the end of skb. */
1323 if (copy > msg_data_left(msg))
1324 copy = msg_data_left(msg);
1328 int i = skb_shinfo(skb)->nr_frags;
1329 struct page_frag *pfrag = sk_page_frag(sk);
1331 if (!sk_page_frag_refill(sk, pfrag))
1332 goto wait_for_space;
1334 if (!skb_can_coalesce(skb, i, pfrag->page,
1336 if (i >= sysctl_max_skb_frags) {
1337 tcp_mark_push(tp, skb);
1343 copy = min_t(int, copy, pfrag->size - pfrag->offset);
1345 if (unlikely(skb_zcopy_pure(skb) || skb_zcopy_managed(skb))) {
1346 if (tcp_downgrade_zcopy_pure(sk, skb))
1347 goto wait_for_space;
1348 skb_zcopy_downgrade_managed(skb);
1351 if (!sk_wmem_schedule(sk, copy))
1352 goto wait_for_space;
1354 err = skb_copy_to_page_nocache(sk, &msg->msg_iter, skb,
1361 /* Update the skb. */
1363 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1365 skb_fill_page_desc(skb, i, pfrag->page,
1366 pfrag->offset, copy);
1367 page_ref_inc(pfrag->page);
1369 pfrag->offset += copy;
1371 /* First append to a fragless skb builds initial
1375 skb_shinfo(skb)->flags |= SKBFL_PURE_ZEROCOPY;
1377 if (!skb_zcopy_pure(skb)) {
1378 if (!sk_wmem_schedule(sk, copy))
1379 goto wait_for_space;
1382 err = skb_zerocopy_iter_stream(sk, skb, msg, copy, uarg);
1383 if (err == -EMSGSIZE || err == -EEXIST) {
1384 tcp_mark_push(tp, skb);
1393 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
1395 WRITE_ONCE(tp->write_seq, tp->write_seq + copy);
1396 TCP_SKB_CB(skb)->end_seq += copy;
1397 tcp_skb_pcount_set(skb, 0);
1400 if (!msg_data_left(msg)) {
1401 if (unlikely(flags & MSG_EOR))
1402 TCP_SKB_CB(skb)->eor = 1;
1406 if (skb->len < size_goal || (flags & MSG_OOB) || unlikely(tp->repair))
1409 if (forced_push(tp)) {
1410 tcp_mark_push(tp, skb);
1411 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
1412 } else if (skb == tcp_send_head(sk))
1413 tcp_push_one(sk, mss_now);
1417 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1419 tcp_push(sk, flags & ~MSG_MORE, mss_now,
1420 TCP_NAGLE_PUSH, size_goal);
1422 err = sk_stream_wait_memory(sk, &timeo);
1426 mss_now = tcp_send_mss(sk, &size_goal, flags);
1431 tcp_tx_timestamp(sk, sockc.tsflags);
1432 tcp_push(sk, flags, mss_now, tp->nonagle, size_goal);
1435 net_zcopy_put(uarg);
1436 return copied + copied_syn;
1439 tcp_remove_empty_skb(sk);
1441 if (copied + copied_syn)
1444 net_zcopy_put_abort(uarg, true);
1445 err = sk_stream_error(sk, flags, err);
1446 /* make sure we wake any epoll edge trigger waiter */
1447 if (unlikely(tcp_rtx_and_write_queues_empty(sk) && err == -EAGAIN)) {
1448 sk->sk_write_space(sk);
1449 tcp_chrono_stop(sk, TCP_CHRONO_SNDBUF_LIMITED);
1453 EXPORT_SYMBOL_GPL(tcp_sendmsg_locked);
1455 int tcp_sendmsg(struct sock *sk, struct msghdr *msg, size_t size)
1460 ret = tcp_sendmsg_locked(sk, msg, size);
1465 EXPORT_SYMBOL(tcp_sendmsg);
1468 * Handle reading urgent data. BSD has very simple semantics for
1469 * this, no blocking and very strange errors 8)
1472 static int tcp_recv_urg(struct sock *sk, struct msghdr *msg, int len, int flags)
1474 struct tcp_sock *tp = tcp_sk(sk);
1476 /* No URG data to read. */
1477 if (sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data ||
1478 tp->urg_data == TCP_URG_READ)
1479 return -EINVAL; /* Yes this is right ! */
1481 if (sk->sk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DONE))
1484 if (tp->urg_data & TCP_URG_VALID) {
1486 char c = tp->urg_data;
1488 if (!(flags & MSG_PEEK))
1489 WRITE_ONCE(tp->urg_data, TCP_URG_READ);
1491 /* Read urgent data. */
1492 msg->msg_flags |= MSG_OOB;
1495 if (!(flags & MSG_TRUNC))
1496 err = memcpy_to_msg(msg, &c, 1);
1499 msg->msg_flags |= MSG_TRUNC;
1501 return err ? -EFAULT : len;
1504 if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
1507 /* Fixed the recv(..., MSG_OOB) behaviour. BSD docs and
1508 * the available implementations agree in this case:
1509 * this call should never block, independent of the
1510 * blocking state of the socket.
1511 * Mike <pall@rz.uni-karlsruhe.de>
1516 static int tcp_peek_sndq(struct sock *sk, struct msghdr *msg, int len)
1518 struct sk_buff *skb;
1519 int copied = 0, err = 0;
1521 /* XXX -- need to support SO_PEEK_OFF */
1523 skb_rbtree_walk(skb, &sk->tcp_rtx_queue) {
1524 err = skb_copy_datagram_msg(skb, 0, msg, skb->len);
1530 skb_queue_walk(&sk->sk_write_queue, skb) {
1531 err = skb_copy_datagram_msg(skb, 0, msg, skb->len);
1538 return err ?: copied;
1541 /* Clean up the receive buffer for full frames taken by the user,
1542 * then send an ACK if necessary. COPIED is the number of bytes
1543 * tcp_recvmsg has given to the user so far, it speeds up the
1544 * calculation of whether or not we must ACK for the sake of
1547 void tcp_cleanup_rbuf(struct sock *sk, int copied)
1549 struct tcp_sock *tp = tcp_sk(sk);
1550 bool time_to_ack = false;
1552 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
1554 WARN(skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq),
1555 "cleanup rbuf bug: copied %X seq %X rcvnxt %X\n",
1556 tp->copied_seq, TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt);
1558 if (inet_csk_ack_scheduled(sk)) {
1559 const struct inet_connection_sock *icsk = inet_csk(sk);
1561 if (/* Once-per-two-segments ACK was not sent by tcp_input.c */
1562 tp->rcv_nxt - tp->rcv_wup > icsk->icsk_ack.rcv_mss ||
1564 * If this read emptied read buffer, we send ACK, if
1565 * connection is not bidirectional, user drained
1566 * receive buffer and there was a small segment
1570 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED2) ||
1571 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED) &&
1572 !inet_csk_in_pingpong_mode(sk))) &&
1573 !atomic_read(&sk->sk_rmem_alloc)))
1577 /* We send an ACK if we can now advertise a non-zero window
1578 * which has been raised "significantly".
1580 * Even if window raised up to infinity, do not send window open ACK
1581 * in states, where we will not receive more. It is useless.
1583 if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) {
1584 __u32 rcv_window_now = tcp_receive_window(tp);
1586 /* Optimize, __tcp_select_window() is not cheap. */
1587 if (2*rcv_window_now <= tp->window_clamp) {
1588 __u32 new_window = __tcp_select_window(sk);
1590 /* Send ACK now, if this read freed lots of space
1591 * in our buffer. Certainly, new_window is new window.
1592 * We can advertise it now, if it is not less than current one.
1593 * "Lots" means "at least twice" here.
1595 if (new_window && new_window >= 2 * rcv_window_now)
1603 static void tcp_eat_recv_skb(struct sock *sk, struct sk_buff *skb)
1605 __skb_unlink(skb, &sk->sk_receive_queue);
1606 if (likely(skb->destructor == sock_rfree)) {
1608 skb->destructor = NULL;
1610 return skb_attempt_defer_free(skb);
1615 static struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off)
1617 struct sk_buff *skb;
1620 while ((skb = skb_peek(&sk->sk_receive_queue)) != NULL) {
1621 offset = seq - TCP_SKB_CB(skb)->seq;
1622 if (unlikely(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
1623 pr_err_once("%s: found a SYN, please report !\n", __func__);
1626 if (offset < skb->len || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)) {
1630 /* This looks weird, but this can happen if TCP collapsing
1631 * splitted a fat GRO packet, while we released socket lock
1632 * in skb_splice_bits()
1634 tcp_eat_recv_skb(sk, skb);
1640 * This routine provides an alternative to tcp_recvmsg() for routines
1641 * that would like to handle copying from skbuffs directly in 'sendfile'
1644 * - It is assumed that the socket was locked by the caller.
1645 * - The routine does not block.
1646 * - At present, there is no support for reading OOB data
1647 * or for 'peeking' the socket using this routine
1648 * (although both would be easy to implement).
1650 int tcp_read_sock(struct sock *sk, read_descriptor_t *desc,
1651 sk_read_actor_t recv_actor)
1653 struct sk_buff *skb;
1654 struct tcp_sock *tp = tcp_sk(sk);
1655 u32 seq = tp->copied_seq;
1659 if (sk->sk_state == TCP_LISTEN)
1661 while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1662 if (offset < skb->len) {
1666 len = skb->len - offset;
1667 /* Stop reading if we hit a patch of urgent data */
1668 if (unlikely(tp->urg_data)) {
1669 u32 urg_offset = tp->urg_seq - seq;
1670 if (urg_offset < len)
1675 used = recv_actor(desc, skb, offset, len);
1681 if (WARN_ON_ONCE(used > len))
1687 /* If recv_actor drops the lock (e.g. TCP splice
1688 * receive) the skb pointer might be invalid when
1689 * getting here: tcp_collapse might have deleted it
1690 * while aggregating skbs from the socket queue.
1692 skb = tcp_recv_skb(sk, seq - 1, &offset);
1695 /* TCP coalescing might have appended data to the skb.
1696 * Try to splice more frags
1698 if (offset + 1 != skb->len)
1701 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) {
1702 tcp_eat_recv_skb(sk, skb);
1706 tcp_eat_recv_skb(sk, skb);
1709 WRITE_ONCE(tp->copied_seq, seq);
1711 WRITE_ONCE(tp->copied_seq, seq);
1713 tcp_rcv_space_adjust(sk);
1715 /* Clean up data we have read: This will do ACK frames. */
1717 tcp_recv_skb(sk, seq, &offset);
1718 tcp_cleanup_rbuf(sk, copied);
1722 EXPORT_SYMBOL(tcp_read_sock);
1724 int tcp_peek_len(struct socket *sock)
1726 return tcp_inq(sock->sk);
1728 EXPORT_SYMBOL(tcp_peek_len);
1730 /* Make sure sk_rcvbuf is big enough to satisfy SO_RCVLOWAT hint */
1731 int tcp_set_rcvlowat(struct sock *sk, int val)
1735 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK)
1736 cap = sk->sk_rcvbuf >> 1;
1738 cap = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_rmem[2]) >> 1;
1739 val = min(val, cap);
1740 WRITE_ONCE(sk->sk_rcvlowat, val ? : 1);
1742 /* Check if we need to signal EPOLLIN right now */
1745 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK)
1749 if (val > sk->sk_rcvbuf) {
1750 WRITE_ONCE(sk->sk_rcvbuf, val);
1751 tcp_sk(sk)->window_clamp = tcp_win_from_space(sk, val);
1755 EXPORT_SYMBOL(tcp_set_rcvlowat);
1757 void tcp_update_recv_tstamps(struct sk_buff *skb,
1758 struct scm_timestamping_internal *tss)
1761 tss->ts[0] = ktime_to_timespec64(skb->tstamp);
1763 tss->ts[0] = (struct timespec64) {0};
1765 if (skb_hwtstamps(skb)->hwtstamp)
1766 tss->ts[2] = ktime_to_timespec64(skb_hwtstamps(skb)->hwtstamp);
1768 tss->ts[2] = (struct timespec64) {0};
1772 static const struct vm_operations_struct tcp_vm_ops = {
1775 int tcp_mmap(struct file *file, struct socket *sock,
1776 struct vm_area_struct *vma)
1778 if (vma->vm_flags & (VM_WRITE | VM_EXEC))
1780 vma->vm_flags &= ~(VM_MAYWRITE | VM_MAYEXEC);
1782 /* Instruct vm_insert_page() to not mmap_read_lock(mm) */
1783 vma->vm_flags |= VM_MIXEDMAP;
1785 vma->vm_ops = &tcp_vm_ops;
1788 EXPORT_SYMBOL(tcp_mmap);
1790 static skb_frag_t *skb_advance_to_frag(struct sk_buff *skb, u32 offset_skb,
1795 if (unlikely(offset_skb >= skb->len))
1798 offset_skb -= skb_headlen(skb);
1799 if ((int)offset_skb < 0 || skb_has_frag_list(skb))
1802 frag = skb_shinfo(skb)->frags;
1803 while (offset_skb) {
1804 if (skb_frag_size(frag) > offset_skb) {
1805 *offset_frag = offset_skb;
1808 offset_skb -= skb_frag_size(frag);
1815 static bool can_map_frag(const skb_frag_t *frag)
1817 return skb_frag_size(frag) == PAGE_SIZE && !skb_frag_off(frag);
1820 static int find_next_mappable_frag(const skb_frag_t *frag,
1821 int remaining_in_skb)
1825 if (likely(can_map_frag(frag)))
1828 while (offset < remaining_in_skb && !can_map_frag(frag)) {
1829 offset += skb_frag_size(frag);
1835 static void tcp_zerocopy_set_hint_for_skb(struct sock *sk,
1836 struct tcp_zerocopy_receive *zc,
1837 struct sk_buff *skb, u32 offset)
1839 u32 frag_offset, partial_frag_remainder = 0;
1840 int mappable_offset;
1843 /* worst case: skip to next skb. try to improve on this case below */
1844 zc->recv_skip_hint = skb->len - offset;
1846 /* Find the frag containing this offset (and how far into that frag) */
1847 frag = skb_advance_to_frag(skb, offset, &frag_offset);
1852 struct skb_shared_info *info = skb_shinfo(skb);
1854 /* We read part of the last frag, must recvmsg() rest of skb. */
1855 if (frag == &info->frags[info->nr_frags - 1])
1858 /* Else, we must at least read the remainder in this frag. */
1859 partial_frag_remainder = skb_frag_size(frag) - frag_offset;
1860 zc->recv_skip_hint -= partial_frag_remainder;
1864 /* partial_frag_remainder: If part way through a frag, must read rest.
1865 * mappable_offset: Bytes till next mappable frag, *not* counting bytes
1866 * in partial_frag_remainder.
1868 mappable_offset = find_next_mappable_frag(frag, zc->recv_skip_hint);
1869 zc->recv_skip_hint = mappable_offset + partial_frag_remainder;
1872 static int tcp_recvmsg_locked(struct sock *sk, struct msghdr *msg, size_t len,
1873 int flags, struct scm_timestamping_internal *tss,
1875 static int receive_fallback_to_copy(struct sock *sk,
1876 struct tcp_zerocopy_receive *zc, int inq,
1877 struct scm_timestamping_internal *tss)
1879 unsigned long copy_address = (unsigned long)zc->copybuf_address;
1880 struct msghdr msg = {};
1885 zc->recv_skip_hint = 0;
1887 if (copy_address != zc->copybuf_address)
1890 err = import_single_range(READ, (void __user *)copy_address,
1891 inq, &iov, &msg.msg_iter);
1895 err = tcp_recvmsg_locked(sk, &msg, inq, MSG_DONTWAIT,
1896 tss, &zc->msg_flags);
1900 zc->copybuf_len = err;
1901 if (likely(zc->copybuf_len)) {
1902 struct sk_buff *skb;
1905 skb = tcp_recv_skb(sk, tcp_sk(sk)->copied_seq, &offset);
1907 tcp_zerocopy_set_hint_for_skb(sk, zc, skb, offset);
1912 static int tcp_copy_straggler_data(struct tcp_zerocopy_receive *zc,
1913 struct sk_buff *skb, u32 copylen,
1914 u32 *offset, u32 *seq)
1916 unsigned long copy_address = (unsigned long)zc->copybuf_address;
1917 struct msghdr msg = {};
1921 if (copy_address != zc->copybuf_address)
1924 err = import_single_range(READ, (void __user *)copy_address,
1925 copylen, &iov, &msg.msg_iter);
1928 err = skb_copy_datagram_msg(skb, *offset, &msg, copylen);
1931 zc->recv_skip_hint -= copylen;
1934 return (__s32)copylen;
1937 static int tcp_zc_handle_leftover(struct tcp_zerocopy_receive *zc,
1939 struct sk_buff *skb,
1942 struct scm_timestamping_internal *tss)
1944 u32 offset, copylen = min_t(u32, copybuf_len, zc->recv_skip_hint);
1948 /* skb is null if inq < PAGE_SIZE. */
1950 offset = *seq - TCP_SKB_CB(skb)->seq;
1952 skb = tcp_recv_skb(sk, *seq, &offset);
1953 if (TCP_SKB_CB(skb)->has_rxtstamp) {
1954 tcp_update_recv_tstamps(skb, tss);
1955 zc->msg_flags |= TCP_CMSG_TS;
1959 zc->copybuf_len = tcp_copy_straggler_data(zc, skb, copylen, &offset,
1961 return zc->copybuf_len < 0 ? 0 : copylen;
1964 static int tcp_zerocopy_vm_insert_batch_error(struct vm_area_struct *vma,
1965 struct page **pending_pages,
1966 unsigned long pages_remaining,
1967 unsigned long *address,
1970 struct tcp_zerocopy_receive *zc,
1971 u32 total_bytes_to_map,
1974 /* At least one page did not map. Try zapping if we skipped earlier. */
1975 if (err == -EBUSY &&
1976 zc->flags & TCP_RECEIVE_ZEROCOPY_FLAG_TLB_CLEAN_HINT) {
1979 maybe_zap_len = total_bytes_to_map - /* All bytes to map */
1980 *length + /* Mapped or pending */
1981 (pages_remaining * PAGE_SIZE); /* Failed map. */
1982 zap_page_range(vma, *address, maybe_zap_len);
1987 unsigned long leftover_pages = pages_remaining;
1990 /* We called zap_page_range, try to reinsert. */
1991 err = vm_insert_pages(vma, *address,
1994 bytes_mapped = PAGE_SIZE * (leftover_pages - pages_remaining);
1995 *seq += bytes_mapped;
1996 *address += bytes_mapped;
1999 /* Either we were unable to zap, OR we zapped, retried an
2000 * insert, and still had an issue. Either ways, pages_remaining
2001 * is the number of pages we were unable to map, and we unroll
2002 * some state we speculatively touched before.
2004 const int bytes_not_mapped = PAGE_SIZE * pages_remaining;
2006 *length -= bytes_not_mapped;
2007 zc->recv_skip_hint += bytes_not_mapped;
2012 static int tcp_zerocopy_vm_insert_batch(struct vm_area_struct *vma,
2013 struct page **pages,
2014 unsigned int pages_to_map,
2015 unsigned long *address,
2018 struct tcp_zerocopy_receive *zc,
2019 u32 total_bytes_to_map)
2021 unsigned long pages_remaining = pages_to_map;
2022 unsigned int pages_mapped;
2023 unsigned int bytes_mapped;
2026 err = vm_insert_pages(vma, *address, pages, &pages_remaining);
2027 pages_mapped = pages_to_map - (unsigned int)pages_remaining;
2028 bytes_mapped = PAGE_SIZE * pages_mapped;
2029 /* Even if vm_insert_pages fails, it may have partially succeeded in
2030 * mapping (some but not all of the pages).
2032 *seq += bytes_mapped;
2033 *address += bytes_mapped;
2038 /* Error: maybe zap and retry + rollback state for failed inserts. */
2039 return tcp_zerocopy_vm_insert_batch_error(vma, pages + pages_mapped,
2040 pages_remaining, address, length, seq, zc, total_bytes_to_map,
2044 #define TCP_VALID_ZC_MSG_FLAGS (TCP_CMSG_TS)
2045 static void tcp_zc_finalize_rx_tstamp(struct sock *sk,
2046 struct tcp_zerocopy_receive *zc,
2047 struct scm_timestamping_internal *tss)
2049 unsigned long msg_control_addr;
2050 struct msghdr cmsg_dummy;
2052 msg_control_addr = (unsigned long)zc->msg_control;
2053 cmsg_dummy.msg_control = (void *)msg_control_addr;
2054 cmsg_dummy.msg_controllen =
2055 (__kernel_size_t)zc->msg_controllen;
2056 cmsg_dummy.msg_flags = in_compat_syscall()
2057 ? MSG_CMSG_COMPAT : 0;
2058 cmsg_dummy.msg_control_is_user = true;
2060 if (zc->msg_control == msg_control_addr &&
2061 zc->msg_controllen == cmsg_dummy.msg_controllen) {
2062 tcp_recv_timestamp(&cmsg_dummy, sk, tss);
2063 zc->msg_control = (__u64)
2064 ((uintptr_t)cmsg_dummy.msg_control);
2065 zc->msg_controllen =
2066 (__u64)cmsg_dummy.msg_controllen;
2067 zc->msg_flags = (__u32)cmsg_dummy.msg_flags;
2071 #define TCP_ZEROCOPY_PAGE_BATCH_SIZE 32
2072 static int tcp_zerocopy_receive(struct sock *sk,
2073 struct tcp_zerocopy_receive *zc,
2074 struct scm_timestamping_internal *tss)
2076 u32 length = 0, offset, vma_len, avail_len, copylen = 0;
2077 unsigned long address = (unsigned long)zc->address;
2078 struct page *pages[TCP_ZEROCOPY_PAGE_BATCH_SIZE];
2079 s32 copybuf_len = zc->copybuf_len;
2080 struct tcp_sock *tp = tcp_sk(sk);
2081 const skb_frag_t *frags = NULL;
2082 unsigned int pages_to_map = 0;
2083 struct vm_area_struct *vma;
2084 struct sk_buff *skb = NULL;
2085 u32 seq = tp->copied_seq;
2086 u32 total_bytes_to_map;
2087 int inq = tcp_inq(sk);
2090 zc->copybuf_len = 0;
2093 if (address & (PAGE_SIZE - 1) || address != zc->address)
2096 if (sk->sk_state == TCP_LISTEN)
2099 sock_rps_record_flow(sk);
2101 if (inq && inq <= copybuf_len)
2102 return receive_fallback_to_copy(sk, zc, inq, tss);
2104 if (inq < PAGE_SIZE) {
2106 zc->recv_skip_hint = inq;
2107 if (!inq && sock_flag(sk, SOCK_DONE))
2112 mmap_read_lock(current->mm);
2114 vma = vma_lookup(current->mm, address);
2115 if (!vma || vma->vm_ops != &tcp_vm_ops) {
2116 mmap_read_unlock(current->mm);
2119 vma_len = min_t(unsigned long, zc->length, vma->vm_end - address);
2120 avail_len = min_t(u32, vma_len, inq);
2121 total_bytes_to_map = avail_len & ~(PAGE_SIZE - 1);
2122 if (total_bytes_to_map) {
2123 if (!(zc->flags & TCP_RECEIVE_ZEROCOPY_FLAG_TLB_CLEAN_HINT))
2124 zap_page_range(vma, address, total_bytes_to_map);
2125 zc->length = total_bytes_to_map;
2126 zc->recv_skip_hint = 0;
2128 zc->length = avail_len;
2129 zc->recv_skip_hint = avail_len;
2132 while (length + PAGE_SIZE <= zc->length) {
2133 int mappable_offset;
2136 if (zc->recv_skip_hint < PAGE_SIZE) {
2140 if (zc->recv_skip_hint > 0)
2143 offset = seq - TCP_SKB_CB(skb)->seq;
2145 skb = tcp_recv_skb(sk, seq, &offset);
2148 if (TCP_SKB_CB(skb)->has_rxtstamp) {
2149 tcp_update_recv_tstamps(skb, tss);
2150 zc->msg_flags |= TCP_CMSG_TS;
2152 zc->recv_skip_hint = skb->len - offset;
2153 frags = skb_advance_to_frag(skb, offset, &offset_frag);
2154 if (!frags || offset_frag)
2158 mappable_offset = find_next_mappable_frag(frags,
2159 zc->recv_skip_hint);
2160 if (mappable_offset) {
2161 zc->recv_skip_hint = mappable_offset;
2164 page = skb_frag_page(frags);
2166 pages[pages_to_map++] = page;
2167 length += PAGE_SIZE;
2168 zc->recv_skip_hint -= PAGE_SIZE;
2170 if (pages_to_map == TCP_ZEROCOPY_PAGE_BATCH_SIZE ||
2171 zc->recv_skip_hint < PAGE_SIZE) {
2172 /* Either full batch, or we're about to go to next skb
2173 * (and we cannot unroll failed ops across skbs).
2175 ret = tcp_zerocopy_vm_insert_batch(vma, pages,
2179 total_bytes_to_map);
2186 ret = tcp_zerocopy_vm_insert_batch(vma, pages, pages_to_map,
2187 &address, &length, &seq,
2188 zc, total_bytes_to_map);
2191 mmap_read_unlock(current->mm);
2192 /* Try to copy straggler data. */
2194 copylen = tcp_zc_handle_leftover(zc, sk, skb, &seq, copybuf_len, tss);
2196 if (length + copylen) {
2197 WRITE_ONCE(tp->copied_seq, seq);
2198 tcp_rcv_space_adjust(sk);
2200 /* Clean up data we have read: This will do ACK frames. */
2201 tcp_recv_skb(sk, seq, &offset);
2202 tcp_cleanup_rbuf(sk, length + copylen);
2204 if (length == zc->length)
2205 zc->recv_skip_hint = 0;
2207 if (!zc->recv_skip_hint && sock_flag(sk, SOCK_DONE))
2210 zc->length = length;
2215 /* Similar to __sock_recv_timestamp, but does not require an skb */
2216 void tcp_recv_timestamp(struct msghdr *msg, const struct sock *sk,
2217 struct scm_timestamping_internal *tss)
2219 int new_tstamp = sock_flag(sk, SOCK_TSTAMP_NEW);
2220 bool has_timestamping = false;
2222 if (tss->ts[0].tv_sec || tss->ts[0].tv_nsec) {
2223 if (sock_flag(sk, SOCK_RCVTSTAMP)) {
2224 if (sock_flag(sk, SOCK_RCVTSTAMPNS)) {
2226 struct __kernel_timespec kts = {
2227 .tv_sec = tss->ts[0].tv_sec,
2228 .tv_nsec = tss->ts[0].tv_nsec,
2230 put_cmsg(msg, SOL_SOCKET, SO_TIMESTAMPNS_NEW,
2233 struct __kernel_old_timespec ts_old = {
2234 .tv_sec = tss->ts[0].tv_sec,
2235 .tv_nsec = tss->ts[0].tv_nsec,
2237 put_cmsg(msg, SOL_SOCKET, SO_TIMESTAMPNS_OLD,
2238 sizeof(ts_old), &ts_old);
2242 struct __kernel_sock_timeval stv = {
2243 .tv_sec = tss->ts[0].tv_sec,
2244 .tv_usec = tss->ts[0].tv_nsec / 1000,
2246 put_cmsg(msg, SOL_SOCKET, SO_TIMESTAMP_NEW,
2249 struct __kernel_old_timeval tv = {
2250 .tv_sec = tss->ts[0].tv_sec,
2251 .tv_usec = tss->ts[0].tv_nsec / 1000,
2253 put_cmsg(msg, SOL_SOCKET, SO_TIMESTAMP_OLD,
2259 if (sk->sk_tsflags & SOF_TIMESTAMPING_SOFTWARE)
2260 has_timestamping = true;
2262 tss->ts[0] = (struct timespec64) {0};
2265 if (tss->ts[2].tv_sec || tss->ts[2].tv_nsec) {
2266 if (sk->sk_tsflags & SOF_TIMESTAMPING_RAW_HARDWARE)
2267 has_timestamping = true;
2269 tss->ts[2] = (struct timespec64) {0};
2272 if (has_timestamping) {
2273 tss->ts[1] = (struct timespec64) {0};
2274 if (sock_flag(sk, SOCK_TSTAMP_NEW))
2275 put_cmsg_scm_timestamping64(msg, tss);
2277 put_cmsg_scm_timestamping(msg, tss);
2281 static int tcp_inq_hint(struct sock *sk)
2283 const struct tcp_sock *tp = tcp_sk(sk);
2284 u32 copied_seq = READ_ONCE(tp->copied_seq);
2285 u32 rcv_nxt = READ_ONCE(tp->rcv_nxt);
2288 inq = rcv_nxt - copied_seq;
2289 if (unlikely(inq < 0 || copied_seq != READ_ONCE(tp->copied_seq))) {
2291 inq = tp->rcv_nxt - tp->copied_seq;
2294 /* After receiving a FIN, tell the user-space to continue reading
2295 * by returning a non-zero inq.
2297 if (inq == 0 && sock_flag(sk, SOCK_DONE))
2303 * This routine copies from a sock struct into the user buffer.
2305 * Technical note: in 2.3 we work on _locked_ socket, so that
2306 * tricks with *seq access order and skb->users are not required.
2307 * Probably, code can be easily improved even more.
2310 static int tcp_recvmsg_locked(struct sock *sk, struct msghdr *msg, size_t len,
2311 int flags, struct scm_timestamping_internal *tss,
2314 struct tcp_sock *tp = tcp_sk(sk);
2320 int target; /* Read at least this many bytes */
2322 struct sk_buff *skb, *last;
2326 if (sk->sk_state == TCP_LISTEN)
2329 if (tp->recvmsg_inq) {
2330 *cmsg_flags = TCP_CMSG_INQ;
2331 msg->msg_get_inq = 1;
2333 timeo = sock_rcvtimeo(sk, flags & MSG_DONTWAIT);
2335 /* Urgent data needs to be handled specially. */
2336 if (flags & MSG_OOB)
2339 if (unlikely(tp->repair)) {
2341 if (!(flags & MSG_PEEK))
2344 if (tp->repair_queue == TCP_SEND_QUEUE)
2348 if (tp->repair_queue == TCP_NO_QUEUE)
2351 /* 'common' recv queue MSG_PEEK-ing */
2354 seq = &tp->copied_seq;
2355 if (flags & MSG_PEEK) {
2356 peek_seq = tp->copied_seq;
2360 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
2365 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
2366 if (unlikely(tp->urg_data) && tp->urg_seq == *seq) {
2369 if (signal_pending(current)) {
2370 copied = timeo ? sock_intr_errno(timeo) : -EAGAIN;
2375 /* Next get a buffer. */
2377 last = skb_peek_tail(&sk->sk_receive_queue);
2378 skb_queue_walk(&sk->sk_receive_queue, skb) {
2380 /* Now that we have two receive queues this
2383 if (WARN(before(*seq, TCP_SKB_CB(skb)->seq),
2384 "TCP recvmsg seq # bug: copied %X, seq %X, rcvnxt %X, fl %X\n",
2385 *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt,
2389 offset = *seq - TCP_SKB_CB(skb)->seq;
2390 if (unlikely(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
2391 pr_err_once("%s: found a SYN, please report !\n", __func__);
2394 if (offset < skb->len)
2396 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
2398 WARN(!(flags & MSG_PEEK),
2399 "TCP recvmsg seq # bug 2: copied %X, seq %X, rcvnxt %X, fl %X\n",
2400 *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt, flags);
2403 /* Well, if we have backlog, try to process it now yet. */
2405 if (copied >= target && !READ_ONCE(sk->sk_backlog.tail))
2411 sk->sk_state == TCP_CLOSE ||
2412 (sk->sk_shutdown & RCV_SHUTDOWN) ||
2413 signal_pending(current))
2416 if (sock_flag(sk, SOCK_DONE))
2420 copied = sock_error(sk);
2424 if (sk->sk_shutdown & RCV_SHUTDOWN)
2427 if (sk->sk_state == TCP_CLOSE) {
2428 /* This occurs when user tries to read
2429 * from never connected socket.
2440 if (signal_pending(current)) {
2441 copied = sock_intr_errno(timeo);
2446 if (copied >= target) {
2447 /* Do not sleep, just process backlog. */
2448 __sk_flush_backlog(sk);
2450 tcp_cleanup_rbuf(sk, copied);
2451 sk_wait_data(sk, &timeo, last);
2454 if ((flags & MSG_PEEK) &&
2455 (peek_seq - copied - urg_hole != tp->copied_seq)) {
2456 net_dbg_ratelimited("TCP(%s:%d): Application bug, race in MSG_PEEK\n",
2458 task_pid_nr(current));
2459 peek_seq = tp->copied_seq;
2464 /* Ok so how much can we use? */
2465 used = skb->len - offset;
2469 /* Do we have urgent data here? */
2470 if (unlikely(tp->urg_data)) {
2471 u32 urg_offset = tp->urg_seq - *seq;
2472 if (urg_offset < used) {
2474 if (!sock_flag(sk, SOCK_URGINLINE)) {
2475 WRITE_ONCE(*seq, *seq + 1);
2487 if (!(flags & MSG_TRUNC)) {
2488 err = skb_copy_datagram_msg(skb, offset, msg, used);
2490 /* Exception. Bailout! */
2497 WRITE_ONCE(*seq, *seq + used);
2501 tcp_rcv_space_adjust(sk);
2504 if (unlikely(tp->urg_data) && after(tp->copied_seq, tp->urg_seq)) {
2505 WRITE_ONCE(tp->urg_data, 0);
2506 tcp_fast_path_check(sk);
2509 if (TCP_SKB_CB(skb)->has_rxtstamp) {
2510 tcp_update_recv_tstamps(skb, tss);
2511 *cmsg_flags |= TCP_CMSG_TS;
2514 if (used + offset < skb->len)
2517 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
2519 if (!(flags & MSG_PEEK))
2520 tcp_eat_recv_skb(sk, skb);
2524 /* Process the FIN. */
2525 WRITE_ONCE(*seq, *seq + 1);
2526 if (!(flags & MSG_PEEK))
2527 tcp_eat_recv_skb(sk, skb);
2531 /* According to UNIX98, msg_name/msg_namelen are ignored
2532 * on connected socket. I was just happy when found this 8) --ANK
2535 /* Clean up data we have read: This will do ACK frames. */
2536 tcp_cleanup_rbuf(sk, copied);
2543 err = tcp_recv_urg(sk, msg, len, flags);
2547 err = tcp_peek_sndq(sk, msg, len);
2551 int tcp_recvmsg(struct sock *sk, struct msghdr *msg, size_t len, int flags,
2554 int cmsg_flags = 0, ret;
2555 struct scm_timestamping_internal tss;
2557 if (unlikely(flags & MSG_ERRQUEUE))
2558 return inet_recv_error(sk, msg, len, addr_len);
2560 if (sk_can_busy_loop(sk) &&
2561 skb_queue_empty_lockless(&sk->sk_receive_queue) &&
2562 sk->sk_state == TCP_ESTABLISHED)
2563 sk_busy_loop(sk, flags & MSG_DONTWAIT);
2566 ret = tcp_recvmsg_locked(sk, msg, len, flags, &tss, &cmsg_flags);
2569 if ((cmsg_flags || msg->msg_get_inq) && ret >= 0) {
2570 if (cmsg_flags & TCP_CMSG_TS)
2571 tcp_recv_timestamp(msg, sk, &tss);
2572 if (msg->msg_get_inq) {
2573 msg->msg_inq = tcp_inq_hint(sk);
2574 if (cmsg_flags & TCP_CMSG_INQ)
2575 put_cmsg(msg, SOL_TCP, TCP_CM_INQ,
2576 sizeof(msg->msg_inq), &msg->msg_inq);
2581 EXPORT_SYMBOL(tcp_recvmsg);
2583 void tcp_set_state(struct sock *sk, int state)
2585 int oldstate = sk->sk_state;
2587 /* We defined a new enum for TCP states that are exported in BPF
2588 * so as not force the internal TCP states to be frozen. The
2589 * following checks will detect if an internal state value ever
2590 * differs from the BPF value. If this ever happens, then we will
2591 * need to remap the internal value to the BPF value before calling
2592 * tcp_call_bpf_2arg.
2594 BUILD_BUG_ON((int)BPF_TCP_ESTABLISHED != (int)TCP_ESTABLISHED);
2595 BUILD_BUG_ON((int)BPF_TCP_SYN_SENT != (int)TCP_SYN_SENT);
2596 BUILD_BUG_ON((int)BPF_TCP_SYN_RECV != (int)TCP_SYN_RECV);
2597 BUILD_BUG_ON((int)BPF_TCP_FIN_WAIT1 != (int)TCP_FIN_WAIT1);
2598 BUILD_BUG_ON((int)BPF_TCP_FIN_WAIT2 != (int)TCP_FIN_WAIT2);
2599 BUILD_BUG_ON((int)BPF_TCP_TIME_WAIT != (int)TCP_TIME_WAIT);
2600 BUILD_BUG_ON((int)BPF_TCP_CLOSE != (int)TCP_CLOSE);
2601 BUILD_BUG_ON((int)BPF_TCP_CLOSE_WAIT != (int)TCP_CLOSE_WAIT);
2602 BUILD_BUG_ON((int)BPF_TCP_LAST_ACK != (int)TCP_LAST_ACK);
2603 BUILD_BUG_ON((int)BPF_TCP_LISTEN != (int)TCP_LISTEN);
2604 BUILD_BUG_ON((int)BPF_TCP_CLOSING != (int)TCP_CLOSING);
2605 BUILD_BUG_ON((int)BPF_TCP_NEW_SYN_RECV != (int)TCP_NEW_SYN_RECV);
2606 BUILD_BUG_ON((int)BPF_TCP_MAX_STATES != (int)TCP_MAX_STATES);
2608 /* bpf uapi header bpf.h defines an anonymous enum with values
2609 * BPF_TCP_* used by bpf programs. Currently gcc built vmlinux
2610 * is able to emit this enum in DWARF due to the above BUILD_BUG_ON.
2611 * But clang built vmlinux does not have this enum in DWARF
2612 * since clang removes the above code before generating IR/debuginfo.
2613 * Let us explicitly emit the type debuginfo to ensure the
2614 * above-mentioned anonymous enum in the vmlinux DWARF and hence BTF
2615 * regardless of which compiler is used.
2617 BTF_TYPE_EMIT_ENUM(BPF_TCP_ESTABLISHED);
2619 if (BPF_SOCK_OPS_TEST_FLAG(tcp_sk(sk), BPF_SOCK_OPS_STATE_CB_FLAG))
2620 tcp_call_bpf_2arg(sk, BPF_SOCK_OPS_STATE_CB, oldstate, state);
2623 case TCP_ESTABLISHED:
2624 if (oldstate != TCP_ESTABLISHED)
2625 TCP_INC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
2629 if (oldstate == TCP_CLOSE_WAIT || oldstate == TCP_ESTABLISHED)
2630 TCP_INC_STATS(sock_net(sk), TCP_MIB_ESTABRESETS);
2632 sk->sk_prot->unhash(sk);
2633 if (inet_csk(sk)->icsk_bind_hash &&
2634 !(sk->sk_userlocks & SOCK_BINDPORT_LOCK))
2638 if (oldstate == TCP_ESTABLISHED)
2639 TCP_DEC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
2642 /* Change state AFTER socket is unhashed to avoid closed
2643 * socket sitting in hash tables.
2645 inet_sk_state_store(sk, state);
2647 EXPORT_SYMBOL_GPL(tcp_set_state);
2650 * State processing on a close. This implements the state shift for
2651 * sending our FIN frame. Note that we only send a FIN for some
2652 * states. A shutdown() may have already sent the FIN, or we may be
2656 static const unsigned char new_state[16] = {
2657 /* current state: new state: action: */
2658 [0 /* (Invalid) */] = TCP_CLOSE,
2659 [TCP_ESTABLISHED] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2660 [TCP_SYN_SENT] = TCP_CLOSE,
2661 [TCP_SYN_RECV] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2662 [TCP_FIN_WAIT1] = TCP_FIN_WAIT1,
2663 [TCP_FIN_WAIT2] = TCP_FIN_WAIT2,
2664 [TCP_TIME_WAIT] = TCP_CLOSE,
2665 [TCP_CLOSE] = TCP_CLOSE,
2666 [TCP_CLOSE_WAIT] = TCP_LAST_ACK | TCP_ACTION_FIN,
2667 [TCP_LAST_ACK] = TCP_LAST_ACK,
2668 [TCP_LISTEN] = TCP_CLOSE,
2669 [TCP_CLOSING] = TCP_CLOSING,
2670 [TCP_NEW_SYN_RECV] = TCP_CLOSE, /* should not happen ! */
2673 static int tcp_close_state(struct sock *sk)
2675 int next = (int)new_state[sk->sk_state];
2676 int ns = next & TCP_STATE_MASK;
2678 tcp_set_state(sk, ns);
2680 return next & TCP_ACTION_FIN;
2684 * Shutdown the sending side of a connection. Much like close except
2685 * that we don't receive shut down or sock_set_flag(sk, SOCK_DEAD).
2688 void tcp_shutdown(struct sock *sk, int how)
2690 /* We need to grab some memory, and put together a FIN,
2691 * and then put it into the queue to be sent.
2692 * Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
2694 if (!(how & SEND_SHUTDOWN))
2697 /* If we've already sent a FIN, or it's a closed state, skip this. */
2698 if ((1 << sk->sk_state) &
2699 (TCPF_ESTABLISHED | TCPF_SYN_SENT |
2700 TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) {
2701 /* Clear out any half completed packets. FIN if needed. */
2702 if (tcp_close_state(sk))
2706 EXPORT_SYMBOL(tcp_shutdown);
2708 int tcp_orphan_count_sum(void)
2712 for_each_possible_cpu(i)
2713 total += per_cpu(tcp_orphan_count, i);
2715 return max(total, 0);
2718 static int tcp_orphan_cache;
2719 static struct timer_list tcp_orphan_timer;
2720 #define TCP_ORPHAN_TIMER_PERIOD msecs_to_jiffies(100)
2722 static void tcp_orphan_update(struct timer_list *unused)
2724 WRITE_ONCE(tcp_orphan_cache, tcp_orphan_count_sum());
2725 mod_timer(&tcp_orphan_timer, jiffies + TCP_ORPHAN_TIMER_PERIOD);
2728 static bool tcp_too_many_orphans(int shift)
2730 return READ_ONCE(tcp_orphan_cache) << shift >
2731 READ_ONCE(sysctl_tcp_max_orphans);
2734 bool tcp_check_oom(struct sock *sk, int shift)
2736 bool too_many_orphans, out_of_socket_memory;
2738 too_many_orphans = tcp_too_many_orphans(shift);
2739 out_of_socket_memory = tcp_out_of_memory(sk);
2741 if (too_many_orphans)
2742 net_info_ratelimited("too many orphaned sockets\n");
2743 if (out_of_socket_memory)
2744 net_info_ratelimited("out of memory -- consider tuning tcp_mem\n");
2745 return too_many_orphans || out_of_socket_memory;
2748 void __tcp_close(struct sock *sk, long timeout)
2750 struct sk_buff *skb;
2751 int data_was_unread = 0;
2754 sk->sk_shutdown = SHUTDOWN_MASK;
2756 if (sk->sk_state == TCP_LISTEN) {
2757 tcp_set_state(sk, TCP_CLOSE);
2760 inet_csk_listen_stop(sk);
2762 goto adjudge_to_death;
2765 /* We need to flush the recv. buffs. We do this only on the
2766 * descriptor close, not protocol-sourced closes, because the
2767 * reader process may not have drained the data yet!
2769 while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) {
2770 u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq;
2772 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
2774 data_was_unread += len;
2780 /* If socket has been already reset (e.g. in tcp_reset()) - kill it. */
2781 if (sk->sk_state == TCP_CLOSE)
2782 goto adjudge_to_death;
2784 /* As outlined in RFC 2525, section 2.17, we send a RST here because
2785 * data was lost. To witness the awful effects of the old behavior of
2786 * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk
2787 * GET in an FTP client, suspend the process, wait for the client to
2788 * advertise a zero window, then kill -9 the FTP client, wheee...
2789 * Note: timeout is always zero in such a case.
2791 if (unlikely(tcp_sk(sk)->repair)) {
2792 sk->sk_prot->disconnect(sk, 0);
2793 } else if (data_was_unread) {
2794 /* Unread data was tossed, zap the connection. */
2795 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONCLOSE);
2796 tcp_set_state(sk, TCP_CLOSE);
2797 tcp_send_active_reset(sk, sk->sk_allocation);
2798 } else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) {
2799 /* Check zero linger _after_ checking for unread data. */
2800 sk->sk_prot->disconnect(sk, 0);
2801 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
2802 } else if (tcp_close_state(sk)) {
2803 /* We FIN if the application ate all the data before
2804 * zapping the connection.
2807 /* RED-PEN. Formally speaking, we have broken TCP state
2808 * machine. State transitions:
2810 * TCP_ESTABLISHED -> TCP_FIN_WAIT1
2811 * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible)
2812 * TCP_CLOSE_WAIT -> TCP_LAST_ACK
2814 * are legal only when FIN has been sent (i.e. in window),
2815 * rather than queued out of window. Purists blame.
2817 * F.e. "RFC state" is ESTABLISHED,
2818 * if Linux state is FIN-WAIT-1, but FIN is still not sent.
2820 * The visible declinations are that sometimes
2821 * we enter time-wait state, when it is not required really
2822 * (harmless), do not send active resets, when they are
2823 * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
2824 * they look as CLOSING or LAST_ACK for Linux)
2825 * Probably, I missed some more holelets.
2827 * XXX (TFO) - To start off we don't support SYN+ACK+FIN
2828 * in a single packet! (May consider it later but will
2829 * probably need API support or TCP_CORK SYN-ACK until
2830 * data is written and socket is closed.)
2835 sk_stream_wait_close(sk, timeout);
2838 state = sk->sk_state;
2844 /* remove backlog if any, without releasing ownership. */
2847 this_cpu_inc(tcp_orphan_count);
2849 /* Have we already been destroyed by a softirq or backlog? */
2850 if (state != TCP_CLOSE && sk->sk_state == TCP_CLOSE)
2853 /* This is a (useful) BSD violating of the RFC. There is a
2854 * problem with TCP as specified in that the other end could
2855 * keep a socket open forever with no application left this end.
2856 * We use a 1 minute timeout (about the same as BSD) then kill
2857 * our end. If they send after that then tough - BUT: long enough
2858 * that we won't make the old 4*rto = almost no time - whoops
2861 * Nope, it was not mistake. It is really desired behaviour
2862 * f.e. on http servers, when such sockets are useless, but
2863 * consume significant resources. Let's do it with special
2864 * linger2 option. --ANK
2867 if (sk->sk_state == TCP_FIN_WAIT2) {
2868 struct tcp_sock *tp = tcp_sk(sk);
2869 if (tp->linger2 < 0) {
2870 tcp_set_state(sk, TCP_CLOSE);
2871 tcp_send_active_reset(sk, GFP_ATOMIC);
2872 __NET_INC_STATS(sock_net(sk),
2873 LINUX_MIB_TCPABORTONLINGER);
2875 const int tmo = tcp_fin_time(sk);
2877 if (tmo > TCP_TIMEWAIT_LEN) {
2878 inet_csk_reset_keepalive_timer(sk,
2879 tmo - TCP_TIMEWAIT_LEN);
2881 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
2886 if (sk->sk_state != TCP_CLOSE) {
2888 if (tcp_check_oom(sk, 0)) {
2889 tcp_set_state(sk, TCP_CLOSE);
2890 tcp_send_active_reset(sk, GFP_ATOMIC);
2891 __NET_INC_STATS(sock_net(sk),
2892 LINUX_MIB_TCPABORTONMEMORY);
2893 } else if (!check_net(sock_net(sk))) {
2894 /* Not possible to send reset; just close */
2895 tcp_set_state(sk, TCP_CLOSE);
2899 if (sk->sk_state == TCP_CLOSE) {
2900 struct request_sock *req;
2902 req = rcu_dereference_protected(tcp_sk(sk)->fastopen_rsk,
2903 lockdep_sock_is_held(sk));
2904 /* We could get here with a non-NULL req if the socket is
2905 * aborted (e.g., closed with unread data) before 3WHS
2909 reqsk_fastopen_remove(sk, req, false);
2910 inet_csk_destroy_sock(sk);
2912 /* Otherwise, socket is reprieved until protocol close. */
2919 void tcp_close(struct sock *sk, long timeout)
2922 __tcp_close(sk, timeout);
2926 EXPORT_SYMBOL(tcp_close);
2928 /* These states need RST on ABORT according to RFC793 */
2930 static inline bool tcp_need_reset(int state)
2932 return (1 << state) &
2933 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 |
2934 TCPF_FIN_WAIT2 | TCPF_SYN_RECV);
2937 static void tcp_rtx_queue_purge(struct sock *sk)
2939 struct rb_node *p = rb_first(&sk->tcp_rtx_queue);
2941 tcp_sk(sk)->highest_sack = NULL;
2943 struct sk_buff *skb = rb_to_skb(p);
2946 /* Since we are deleting whole queue, no need to
2947 * list_del(&skb->tcp_tsorted_anchor)
2949 tcp_rtx_queue_unlink(skb, sk);
2950 tcp_wmem_free_skb(sk, skb);
2954 void tcp_write_queue_purge(struct sock *sk)
2956 struct sk_buff *skb;
2958 tcp_chrono_stop(sk, TCP_CHRONO_BUSY);
2959 while ((skb = __skb_dequeue(&sk->sk_write_queue)) != NULL) {
2960 tcp_skb_tsorted_anchor_cleanup(skb);
2961 tcp_wmem_free_skb(sk, skb);
2963 tcp_rtx_queue_purge(sk);
2964 INIT_LIST_HEAD(&tcp_sk(sk)->tsorted_sent_queue);
2966 tcp_clear_all_retrans_hints(tcp_sk(sk));
2967 tcp_sk(sk)->packets_out = 0;
2968 inet_csk(sk)->icsk_backoff = 0;
2971 int tcp_disconnect(struct sock *sk, int flags)
2973 struct inet_sock *inet = inet_sk(sk);
2974 struct inet_connection_sock *icsk = inet_csk(sk);
2975 struct tcp_sock *tp = tcp_sk(sk);
2976 int old_state = sk->sk_state;
2979 if (old_state != TCP_CLOSE)
2980 tcp_set_state(sk, TCP_CLOSE);
2982 /* ABORT function of RFC793 */
2983 if (old_state == TCP_LISTEN) {
2984 inet_csk_listen_stop(sk);
2985 } else if (unlikely(tp->repair)) {
2986 sk->sk_err = ECONNABORTED;
2987 } else if (tcp_need_reset(old_state) ||
2988 (tp->snd_nxt != tp->write_seq &&
2989 (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK))) {
2990 /* The last check adjusts for discrepancy of Linux wrt. RFC
2993 tcp_send_active_reset(sk, gfp_any());
2994 sk->sk_err = ECONNRESET;
2995 } else if (old_state == TCP_SYN_SENT)
2996 sk->sk_err = ECONNRESET;
2998 tcp_clear_xmit_timers(sk);
2999 __skb_queue_purge(&sk->sk_receive_queue);
3000 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt);
3001 WRITE_ONCE(tp->urg_data, 0);
3002 tcp_write_queue_purge(sk);
3003 tcp_fastopen_active_disable_ofo_check(sk);
3004 skb_rbtree_purge(&tp->out_of_order_queue);
3006 inet->inet_dport = 0;
3008 if (!(sk->sk_userlocks & SOCK_BINDADDR_LOCK))
3009 inet_reset_saddr(sk);
3011 sk->sk_shutdown = 0;
3012 sock_reset_flag(sk, SOCK_DONE);
3014 tp->mdev_us = jiffies_to_usecs(TCP_TIMEOUT_INIT);
3015 tp->rcv_rtt_last_tsecr = 0;
3017 seq = tp->write_seq + tp->max_window + 2;
3020 WRITE_ONCE(tp->write_seq, seq);
3022 icsk->icsk_backoff = 0;
3023 icsk->icsk_probes_out = 0;
3024 icsk->icsk_probes_tstamp = 0;
3025 icsk->icsk_rto = TCP_TIMEOUT_INIT;
3026 icsk->icsk_rto_min = TCP_RTO_MIN;
3027 icsk->icsk_delack_max = TCP_DELACK_MAX;
3028 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
3029 tcp_snd_cwnd_set(tp, TCP_INIT_CWND);
3030 tp->snd_cwnd_cnt = 0;
3031 tp->window_clamp = 0;
3033 tp->delivered_ce = 0;
3034 if (icsk->icsk_ca_ops->release)
3035 icsk->icsk_ca_ops->release(sk);
3036 memset(icsk->icsk_ca_priv, 0, sizeof(icsk->icsk_ca_priv));
3037 icsk->icsk_ca_initialized = 0;
3038 tcp_set_ca_state(sk, TCP_CA_Open);
3039 tp->is_sack_reneg = 0;
3040 tcp_clear_retrans(tp);
3041 tp->total_retrans = 0;
3042 inet_csk_delack_init(sk);
3043 /* Initialize rcv_mss to TCP_MIN_MSS to avoid division by 0
3044 * issue in __tcp_select_window()
3046 icsk->icsk_ack.rcv_mss = TCP_MIN_MSS;
3047 memset(&tp->rx_opt, 0, sizeof(tp->rx_opt));
3049 dst_release(xchg((__force struct dst_entry **)&sk->sk_rx_dst, NULL));
3050 tcp_saved_syn_free(tp);
3051 tp->compressed_ack = 0;
3055 tp->bytes_acked = 0;
3056 tp->bytes_received = 0;
3057 tp->bytes_retrans = 0;
3058 tp->data_segs_in = 0;
3059 tp->data_segs_out = 0;
3060 tp->duplicate_sack[0].start_seq = 0;
3061 tp->duplicate_sack[0].end_seq = 0;
3064 tp->retrans_out = 0;
3066 tp->tlp_high_seq = 0;
3067 tp->last_oow_ack_time = 0;
3068 /* There's a bubble in the pipe until at least the first ACK. */
3069 tp->app_limited = ~0U;
3070 tp->rack.mstamp = 0;
3071 tp->rack.advanced = 0;
3072 tp->rack.reo_wnd_steps = 1;
3073 tp->rack.last_delivered = 0;
3074 tp->rack.reo_wnd_persist = 0;
3075 tp->rack.dsack_seen = 0;
3076 tp->syn_data_acked = 0;
3077 tp->rx_opt.saw_tstamp = 0;
3078 tp->rx_opt.dsack = 0;
3079 tp->rx_opt.num_sacks = 0;
3080 tp->rcv_ooopack = 0;
3083 /* Clean up fastopen related fields */
3084 tcp_free_fastopen_req(tp);
3085 inet->defer_connect = 0;
3086 tp->fastopen_client_fail = 0;
3088 WARN_ON(inet->inet_num && !icsk->icsk_bind_hash);
3090 if (sk->sk_frag.page) {
3091 put_page(sk->sk_frag.page);
3092 sk->sk_frag.page = NULL;
3093 sk->sk_frag.offset = 0;
3095 sk_error_report(sk);
3098 EXPORT_SYMBOL(tcp_disconnect);
3100 static inline bool tcp_can_repair_sock(const struct sock *sk)
3102 return ns_capable(sock_net(sk)->user_ns, CAP_NET_ADMIN) &&
3103 (sk->sk_state != TCP_LISTEN);
3106 static int tcp_repair_set_window(struct tcp_sock *tp, sockptr_t optbuf, int len)
3108 struct tcp_repair_window opt;
3113 if (len != sizeof(opt))
3116 if (copy_from_sockptr(&opt, optbuf, sizeof(opt)))
3119 if (opt.max_window < opt.snd_wnd)
3122 if (after(opt.snd_wl1, tp->rcv_nxt + opt.rcv_wnd))
3125 if (after(opt.rcv_wup, tp->rcv_nxt))
3128 tp->snd_wl1 = opt.snd_wl1;
3129 tp->snd_wnd = opt.snd_wnd;
3130 tp->max_window = opt.max_window;
3132 tp->rcv_wnd = opt.rcv_wnd;
3133 tp->rcv_wup = opt.rcv_wup;
3138 static int tcp_repair_options_est(struct sock *sk, sockptr_t optbuf,
3141 struct tcp_sock *tp = tcp_sk(sk);
3142 struct tcp_repair_opt opt;
3145 while (len >= sizeof(opt)) {
3146 if (copy_from_sockptr_offset(&opt, optbuf, offset, sizeof(opt)))
3149 offset += sizeof(opt);
3152 switch (opt.opt_code) {
3154 tp->rx_opt.mss_clamp = opt.opt_val;
3159 u16 snd_wscale = opt.opt_val & 0xFFFF;
3160 u16 rcv_wscale = opt.opt_val >> 16;
3162 if (snd_wscale > TCP_MAX_WSCALE || rcv_wscale > TCP_MAX_WSCALE)
3165 tp->rx_opt.snd_wscale = snd_wscale;
3166 tp->rx_opt.rcv_wscale = rcv_wscale;
3167 tp->rx_opt.wscale_ok = 1;
3170 case TCPOPT_SACK_PERM:
3171 if (opt.opt_val != 0)
3174 tp->rx_opt.sack_ok |= TCP_SACK_SEEN;
3176 case TCPOPT_TIMESTAMP:
3177 if (opt.opt_val != 0)
3180 tp->rx_opt.tstamp_ok = 1;
3188 DEFINE_STATIC_KEY_FALSE(tcp_tx_delay_enabled);
3189 EXPORT_SYMBOL(tcp_tx_delay_enabled);
3191 static void tcp_enable_tx_delay(void)
3193 if (!static_branch_unlikely(&tcp_tx_delay_enabled)) {
3194 static int __tcp_tx_delay_enabled = 0;
3196 if (cmpxchg(&__tcp_tx_delay_enabled, 0, 1) == 0) {
3197 static_branch_enable(&tcp_tx_delay_enabled);
3198 pr_info("TCP_TX_DELAY enabled\n");
3203 /* When set indicates to always queue non-full frames. Later the user clears
3204 * this option and we transmit any pending partial frames in the queue. This is
3205 * meant to be used alongside sendfile() to get properly filled frames when the
3206 * user (for example) must write out headers with a write() call first and then
3207 * use sendfile to send out the data parts.
3209 * TCP_CORK can be set together with TCP_NODELAY and it is stronger than
3212 void __tcp_sock_set_cork(struct sock *sk, bool on)
3214 struct tcp_sock *tp = tcp_sk(sk);
3217 tp->nonagle |= TCP_NAGLE_CORK;
3219 tp->nonagle &= ~TCP_NAGLE_CORK;
3220 if (tp->nonagle & TCP_NAGLE_OFF)
3221 tp->nonagle |= TCP_NAGLE_PUSH;
3222 tcp_push_pending_frames(sk);
3226 void tcp_sock_set_cork(struct sock *sk, bool on)
3229 __tcp_sock_set_cork(sk, on);
3232 EXPORT_SYMBOL(tcp_sock_set_cork);
3234 /* TCP_NODELAY is weaker than TCP_CORK, so that this option on corked socket is
3235 * remembered, but it is not activated until cork is cleared.
3237 * However, when TCP_NODELAY is set we make an explicit push, which overrides
3238 * even TCP_CORK for currently queued segments.
3240 void __tcp_sock_set_nodelay(struct sock *sk, bool on)
3243 tcp_sk(sk)->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH;
3244 tcp_push_pending_frames(sk);
3246 tcp_sk(sk)->nonagle &= ~TCP_NAGLE_OFF;
3250 void tcp_sock_set_nodelay(struct sock *sk)
3253 __tcp_sock_set_nodelay(sk, true);
3256 EXPORT_SYMBOL(tcp_sock_set_nodelay);
3258 static void __tcp_sock_set_quickack(struct sock *sk, int val)
3261 inet_csk_enter_pingpong_mode(sk);
3265 inet_csk_exit_pingpong_mode(sk);
3266 if ((1 << sk->sk_state) & (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) &&
3267 inet_csk_ack_scheduled(sk)) {
3268 inet_csk(sk)->icsk_ack.pending |= ICSK_ACK_PUSHED;
3269 tcp_cleanup_rbuf(sk, 1);
3271 inet_csk_enter_pingpong_mode(sk);
3275 void tcp_sock_set_quickack(struct sock *sk, int val)
3278 __tcp_sock_set_quickack(sk, val);
3281 EXPORT_SYMBOL(tcp_sock_set_quickack);
3283 int tcp_sock_set_syncnt(struct sock *sk, int val)
3285 if (val < 1 || val > MAX_TCP_SYNCNT)
3289 inet_csk(sk)->icsk_syn_retries = val;
3293 EXPORT_SYMBOL(tcp_sock_set_syncnt);
3295 void tcp_sock_set_user_timeout(struct sock *sk, u32 val)
3298 inet_csk(sk)->icsk_user_timeout = val;
3301 EXPORT_SYMBOL(tcp_sock_set_user_timeout);
3303 int tcp_sock_set_keepidle_locked(struct sock *sk, int val)
3305 struct tcp_sock *tp = tcp_sk(sk);
3307 if (val < 1 || val > MAX_TCP_KEEPIDLE)
3310 tp->keepalive_time = val * HZ;
3311 if (sock_flag(sk, SOCK_KEEPOPEN) &&
3312 !((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN))) {
3313 u32 elapsed = keepalive_time_elapsed(tp);
3315 if (tp->keepalive_time > elapsed)
3316 elapsed = tp->keepalive_time - elapsed;
3319 inet_csk_reset_keepalive_timer(sk, elapsed);
3325 int tcp_sock_set_keepidle(struct sock *sk, int val)
3330 err = tcp_sock_set_keepidle_locked(sk, val);
3334 EXPORT_SYMBOL(tcp_sock_set_keepidle);
3336 int tcp_sock_set_keepintvl(struct sock *sk, int val)
3338 if (val < 1 || val > MAX_TCP_KEEPINTVL)
3342 tcp_sk(sk)->keepalive_intvl = val * HZ;
3346 EXPORT_SYMBOL(tcp_sock_set_keepintvl);
3348 int tcp_sock_set_keepcnt(struct sock *sk, int val)
3350 if (val < 1 || val > MAX_TCP_KEEPCNT)
3354 tcp_sk(sk)->keepalive_probes = val;
3358 EXPORT_SYMBOL(tcp_sock_set_keepcnt);
3360 int tcp_set_window_clamp(struct sock *sk, int val)
3362 struct tcp_sock *tp = tcp_sk(sk);
3365 if (sk->sk_state != TCP_CLOSE)
3367 tp->window_clamp = 0;
3369 tp->window_clamp = val < SOCK_MIN_RCVBUF / 2 ?
3370 SOCK_MIN_RCVBUF / 2 : val;
3371 tp->rcv_ssthresh = min(tp->rcv_wnd, tp->window_clamp);
3377 * Socket option code for TCP.
3379 static int do_tcp_setsockopt(struct sock *sk, int level, int optname,
3380 sockptr_t optval, unsigned int optlen)
3382 struct tcp_sock *tp = tcp_sk(sk);
3383 struct inet_connection_sock *icsk = inet_csk(sk);
3384 struct net *net = sock_net(sk);
3388 /* These are data/string values, all the others are ints */
3390 case TCP_CONGESTION: {
3391 char name[TCP_CA_NAME_MAX];
3396 val = strncpy_from_sockptr(name, optval,
3397 min_t(long, TCP_CA_NAME_MAX-1, optlen));
3403 err = tcp_set_congestion_control(sk, name, true,
3404 ns_capable(sock_net(sk)->user_ns,
3410 char name[TCP_ULP_NAME_MAX];
3415 val = strncpy_from_sockptr(name, optval,
3416 min_t(long, TCP_ULP_NAME_MAX - 1,
3423 err = tcp_set_ulp(sk, name);
3427 case TCP_FASTOPEN_KEY: {
3428 __u8 key[TCP_FASTOPEN_KEY_BUF_LENGTH];
3429 __u8 *backup_key = NULL;
3431 /* Allow a backup key as well to facilitate key rotation
3432 * First key is the active one.
3434 if (optlen != TCP_FASTOPEN_KEY_LENGTH &&
3435 optlen != TCP_FASTOPEN_KEY_BUF_LENGTH)
3438 if (copy_from_sockptr(key, optval, optlen))
3441 if (optlen == TCP_FASTOPEN_KEY_BUF_LENGTH)
3442 backup_key = key + TCP_FASTOPEN_KEY_LENGTH;
3444 return tcp_fastopen_reset_cipher(net, sk, key, backup_key);
3451 if (optlen < sizeof(int))
3454 if (copy_from_sockptr(&val, optval, sizeof(val)))
3461 /* Values greater than interface MTU won't take effect. However
3462 * at the point when this call is done we typically don't yet
3463 * know which interface is going to be used
3465 if (val && (val < TCP_MIN_MSS || val > MAX_TCP_WINDOW)) {
3469 tp->rx_opt.user_mss = val;
3473 __tcp_sock_set_nodelay(sk, val);
3476 case TCP_THIN_LINEAR_TIMEOUTS:
3477 if (val < 0 || val > 1)
3483 case TCP_THIN_DUPACK:
3484 if (val < 0 || val > 1)
3489 if (!tcp_can_repair_sock(sk))
3491 else if (val == TCP_REPAIR_ON) {
3493 sk->sk_reuse = SK_FORCE_REUSE;
3494 tp->repair_queue = TCP_NO_QUEUE;
3495 } else if (val == TCP_REPAIR_OFF) {
3497 sk->sk_reuse = SK_NO_REUSE;
3498 tcp_send_window_probe(sk);
3499 } else if (val == TCP_REPAIR_OFF_NO_WP) {
3501 sk->sk_reuse = SK_NO_REUSE;
3507 case TCP_REPAIR_QUEUE:
3510 else if ((unsigned int)val < TCP_QUEUES_NR)
3511 tp->repair_queue = val;
3517 if (sk->sk_state != TCP_CLOSE) {
3519 } else if (tp->repair_queue == TCP_SEND_QUEUE) {
3520 if (!tcp_rtx_queue_empty(sk))
3523 WRITE_ONCE(tp->write_seq, val);
3524 } else if (tp->repair_queue == TCP_RECV_QUEUE) {
3525 if (tp->rcv_nxt != tp->copied_seq) {
3528 WRITE_ONCE(tp->rcv_nxt, val);
3529 WRITE_ONCE(tp->copied_seq, val);
3536 case TCP_REPAIR_OPTIONS:
3539 else if (sk->sk_state == TCP_ESTABLISHED)
3540 err = tcp_repair_options_est(sk, optval, optlen);
3546 __tcp_sock_set_cork(sk, val);
3550 err = tcp_sock_set_keepidle_locked(sk, val);
3553 if (val < 1 || val > MAX_TCP_KEEPINTVL)
3556 tp->keepalive_intvl = val * HZ;
3559 if (val < 1 || val > MAX_TCP_KEEPCNT)
3562 tp->keepalive_probes = val;
3565 if (val < 1 || val > MAX_TCP_SYNCNT)
3568 icsk->icsk_syn_retries = val;
3572 /* 0: disable, 1: enable, 2: start from ether_header */
3573 if (val < 0 || val > 2)
3582 else if (val > TCP_FIN_TIMEOUT_MAX / HZ)
3583 tp->linger2 = TCP_FIN_TIMEOUT_MAX;
3585 tp->linger2 = val * HZ;
3588 case TCP_DEFER_ACCEPT:
3589 /* Translate value in seconds to number of retransmits */
3590 icsk->icsk_accept_queue.rskq_defer_accept =
3591 secs_to_retrans(val, TCP_TIMEOUT_INIT / HZ,
3595 case TCP_WINDOW_CLAMP:
3596 err = tcp_set_window_clamp(sk, val);
3600 __tcp_sock_set_quickack(sk, val);
3603 #ifdef CONFIG_TCP_MD5SIG
3605 case TCP_MD5SIG_EXT:
3606 err = tp->af_specific->md5_parse(sk, optname, optval, optlen);
3609 case TCP_USER_TIMEOUT:
3610 /* Cap the max time in ms TCP will retry or probe the window
3611 * before giving up and aborting (ETIMEDOUT) a connection.
3616 icsk->icsk_user_timeout = val;
3620 if (val >= 0 && ((1 << sk->sk_state) & (TCPF_CLOSE |
3622 tcp_fastopen_init_key_once(net);
3624 fastopen_queue_tune(sk, val);
3629 case TCP_FASTOPEN_CONNECT:
3630 if (val > 1 || val < 0) {
3632 } else if (READ_ONCE(net->ipv4.sysctl_tcp_fastopen) &
3633 TFO_CLIENT_ENABLE) {
3634 if (sk->sk_state == TCP_CLOSE)
3635 tp->fastopen_connect = val;
3642 case TCP_FASTOPEN_NO_COOKIE:
3643 if (val > 1 || val < 0)
3645 else if (!((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
3648 tp->fastopen_no_cookie = val;
3654 tp->tsoffset = val - tcp_time_stamp_raw();
3656 case TCP_REPAIR_WINDOW:
3657 err = tcp_repair_set_window(tp, optval, optlen);
3659 case TCP_NOTSENT_LOWAT:
3660 tp->notsent_lowat = val;
3661 sk->sk_write_space(sk);
3664 if (val > 1 || val < 0)
3667 tp->recvmsg_inq = val;
3671 tcp_enable_tx_delay();
3672 tp->tcp_tx_delay = val;
3683 int tcp_setsockopt(struct sock *sk, int level, int optname, sockptr_t optval,
3684 unsigned int optlen)
3686 const struct inet_connection_sock *icsk = inet_csk(sk);
3688 if (level != SOL_TCP)
3689 return icsk->icsk_af_ops->setsockopt(sk, level, optname,
3691 return do_tcp_setsockopt(sk, level, optname, optval, optlen);
3693 EXPORT_SYMBOL(tcp_setsockopt);
3695 static void tcp_get_info_chrono_stats(const struct tcp_sock *tp,
3696 struct tcp_info *info)
3698 u64 stats[__TCP_CHRONO_MAX], total = 0;
3701 for (i = TCP_CHRONO_BUSY; i < __TCP_CHRONO_MAX; ++i) {
3702 stats[i] = tp->chrono_stat[i - 1];
3703 if (i == tp->chrono_type)
3704 stats[i] += tcp_jiffies32 - tp->chrono_start;
3705 stats[i] *= USEC_PER_SEC / HZ;
3709 info->tcpi_busy_time = total;
3710 info->tcpi_rwnd_limited = stats[TCP_CHRONO_RWND_LIMITED];
3711 info->tcpi_sndbuf_limited = stats[TCP_CHRONO_SNDBUF_LIMITED];
3714 /* Return information about state of tcp endpoint in API format. */
3715 void tcp_get_info(struct sock *sk, struct tcp_info *info)
3717 const struct tcp_sock *tp = tcp_sk(sk); /* iff sk_type == SOCK_STREAM */
3718 const struct inet_connection_sock *icsk = inet_csk(sk);
3724 memset(info, 0, sizeof(*info));
3725 if (sk->sk_type != SOCK_STREAM)
3728 info->tcpi_state = inet_sk_state_load(sk);
3730 /* Report meaningful fields for all TCP states, including listeners */
3731 rate = READ_ONCE(sk->sk_pacing_rate);
3732 rate64 = (rate != ~0UL) ? rate : ~0ULL;
3733 info->tcpi_pacing_rate = rate64;
3735 rate = READ_ONCE(sk->sk_max_pacing_rate);
3736 rate64 = (rate != ~0UL) ? rate : ~0ULL;
3737 info->tcpi_max_pacing_rate = rate64;
3739 info->tcpi_reordering = tp->reordering;
3740 info->tcpi_snd_cwnd = tcp_snd_cwnd(tp);
3742 if (info->tcpi_state == TCP_LISTEN) {
3743 /* listeners aliased fields :
3744 * tcpi_unacked -> Number of children ready for accept()
3745 * tcpi_sacked -> max backlog
3747 info->tcpi_unacked = READ_ONCE(sk->sk_ack_backlog);
3748 info->tcpi_sacked = READ_ONCE(sk->sk_max_ack_backlog);
3752 slow = lock_sock_fast(sk);
3754 info->tcpi_ca_state = icsk->icsk_ca_state;
3755 info->tcpi_retransmits = icsk->icsk_retransmits;
3756 info->tcpi_probes = icsk->icsk_probes_out;
3757 info->tcpi_backoff = icsk->icsk_backoff;
3759 if (tp->rx_opt.tstamp_ok)
3760 info->tcpi_options |= TCPI_OPT_TIMESTAMPS;
3761 if (tcp_is_sack(tp))
3762 info->tcpi_options |= TCPI_OPT_SACK;
3763 if (tp->rx_opt.wscale_ok) {
3764 info->tcpi_options |= TCPI_OPT_WSCALE;
3765 info->tcpi_snd_wscale = tp->rx_opt.snd_wscale;
3766 info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale;
3769 if (tp->ecn_flags & TCP_ECN_OK)
3770 info->tcpi_options |= TCPI_OPT_ECN;
3771 if (tp->ecn_flags & TCP_ECN_SEEN)
3772 info->tcpi_options |= TCPI_OPT_ECN_SEEN;
3773 if (tp->syn_data_acked)
3774 info->tcpi_options |= TCPI_OPT_SYN_DATA;
3776 info->tcpi_rto = jiffies_to_usecs(icsk->icsk_rto);
3777 info->tcpi_ato = jiffies_to_usecs(icsk->icsk_ack.ato);
3778 info->tcpi_snd_mss = tp->mss_cache;
3779 info->tcpi_rcv_mss = icsk->icsk_ack.rcv_mss;
3781 info->tcpi_unacked = tp->packets_out;
3782 info->tcpi_sacked = tp->sacked_out;
3784 info->tcpi_lost = tp->lost_out;
3785 info->tcpi_retrans = tp->retrans_out;
3787 now = tcp_jiffies32;
3788 info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime);
3789 info->tcpi_last_data_recv = jiffies_to_msecs(now - icsk->icsk_ack.lrcvtime);
3790 info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp);
3792 info->tcpi_pmtu = icsk->icsk_pmtu_cookie;
3793 info->tcpi_rcv_ssthresh = tp->rcv_ssthresh;
3794 info->tcpi_rtt = tp->srtt_us >> 3;
3795 info->tcpi_rttvar = tp->mdev_us >> 2;
3796 info->tcpi_snd_ssthresh = tp->snd_ssthresh;
3797 info->tcpi_advmss = tp->advmss;
3799 info->tcpi_rcv_rtt = tp->rcv_rtt_est.rtt_us >> 3;
3800 info->tcpi_rcv_space = tp->rcvq_space.space;
3802 info->tcpi_total_retrans = tp->total_retrans;
3804 info->tcpi_bytes_acked = tp->bytes_acked;
3805 info->tcpi_bytes_received = tp->bytes_received;
3806 info->tcpi_notsent_bytes = max_t(int, 0, tp->write_seq - tp->snd_nxt);
3807 tcp_get_info_chrono_stats(tp, info);
3809 info->tcpi_segs_out = tp->segs_out;
3811 /* segs_in and data_segs_in can be updated from tcp_segs_in() from BH */
3812 info->tcpi_segs_in = READ_ONCE(tp->segs_in);
3813 info->tcpi_data_segs_in = READ_ONCE(tp->data_segs_in);
3815 info->tcpi_min_rtt = tcp_min_rtt(tp);
3816 info->tcpi_data_segs_out = tp->data_segs_out;
3818 info->tcpi_delivery_rate_app_limited = tp->rate_app_limited ? 1 : 0;
3819 rate64 = tcp_compute_delivery_rate(tp);
3821 info->tcpi_delivery_rate = rate64;
3822 info->tcpi_delivered = tp->delivered;
3823 info->tcpi_delivered_ce = tp->delivered_ce;
3824 info->tcpi_bytes_sent = tp->bytes_sent;
3825 info->tcpi_bytes_retrans = tp->bytes_retrans;
3826 info->tcpi_dsack_dups = tp->dsack_dups;
3827 info->tcpi_reord_seen = tp->reord_seen;
3828 info->tcpi_rcv_ooopack = tp->rcv_ooopack;
3829 info->tcpi_snd_wnd = tp->snd_wnd;
3830 info->tcpi_fastopen_client_fail = tp->fastopen_client_fail;
3831 unlock_sock_fast(sk, slow);
3833 EXPORT_SYMBOL_GPL(tcp_get_info);
3835 static size_t tcp_opt_stats_get_size(void)
3838 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_BUSY */
3839 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_RWND_LIMITED */
3840 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_SNDBUF_LIMITED */
3841 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_DATA_SEGS_OUT */
3842 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_TOTAL_RETRANS */
3843 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_PACING_RATE */
3844 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_DELIVERY_RATE */
3845 nla_total_size(sizeof(u32)) + /* TCP_NLA_SND_CWND */
3846 nla_total_size(sizeof(u32)) + /* TCP_NLA_REORDERING */
3847 nla_total_size(sizeof(u32)) + /* TCP_NLA_MIN_RTT */
3848 nla_total_size(sizeof(u8)) + /* TCP_NLA_RECUR_RETRANS */
3849 nla_total_size(sizeof(u8)) + /* TCP_NLA_DELIVERY_RATE_APP_LMT */
3850 nla_total_size(sizeof(u32)) + /* TCP_NLA_SNDQ_SIZE */
3851 nla_total_size(sizeof(u8)) + /* TCP_NLA_CA_STATE */
3852 nla_total_size(sizeof(u32)) + /* TCP_NLA_SND_SSTHRESH */
3853 nla_total_size(sizeof(u32)) + /* TCP_NLA_DELIVERED */
3854 nla_total_size(sizeof(u32)) + /* TCP_NLA_DELIVERED_CE */
3855 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_BYTES_SENT */
3856 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_BYTES_RETRANS */
3857 nla_total_size(sizeof(u32)) + /* TCP_NLA_DSACK_DUPS */
3858 nla_total_size(sizeof(u32)) + /* TCP_NLA_REORD_SEEN */
3859 nla_total_size(sizeof(u32)) + /* TCP_NLA_SRTT */
3860 nla_total_size(sizeof(u16)) + /* TCP_NLA_TIMEOUT_REHASH */
3861 nla_total_size(sizeof(u32)) + /* TCP_NLA_BYTES_NOTSENT */
3862 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_EDT */
3863 nla_total_size(sizeof(u8)) + /* TCP_NLA_TTL */
3867 /* Returns TTL or hop limit of an incoming packet from skb. */
3868 static u8 tcp_skb_ttl_or_hop_limit(const struct sk_buff *skb)
3870 if (skb->protocol == htons(ETH_P_IP))
3871 return ip_hdr(skb)->ttl;
3872 else if (skb->protocol == htons(ETH_P_IPV6))
3873 return ipv6_hdr(skb)->hop_limit;
3878 struct sk_buff *tcp_get_timestamping_opt_stats(const struct sock *sk,
3879 const struct sk_buff *orig_skb,
3880 const struct sk_buff *ack_skb)
3882 const struct tcp_sock *tp = tcp_sk(sk);
3883 struct sk_buff *stats;
3884 struct tcp_info info;
3888 stats = alloc_skb(tcp_opt_stats_get_size(), GFP_ATOMIC);
3892 tcp_get_info_chrono_stats(tp, &info);
3893 nla_put_u64_64bit(stats, TCP_NLA_BUSY,
3894 info.tcpi_busy_time, TCP_NLA_PAD);
3895 nla_put_u64_64bit(stats, TCP_NLA_RWND_LIMITED,
3896 info.tcpi_rwnd_limited, TCP_NLA_PAD);
3897 nla_put_u64_64bit(stats, TCP_NLA_SNDBUF_LIMITED,
3898 info.tcpi_sndbuf_limited, TCP_NLA_PAD);
3899 nla_put_u64_64bit(stats, TCP_NLA_DATA_SEGS_OUT,
3900 tp->data_segs_out, TCP_NLA_PAD);
3901 nla_put_u64_64bit(stats, TCP_NLA_TOTAL_RETRANS,
3902 tp->total_retrans, TCP_NLA_PAD);
3904 rate = READ_ONCE(sk->sk_pacing_rate);
3905 rate64 = (rate != ~0UL) ? rate : ~0ULL;
3906 nla_put_u64_64bit(stats, TCP_NLA_PACING_RATE, rate64, TCP_NLA_PAD);
3908 rate64 = tcp_compute_delivery_rate(tp);
3909 nla_put_u64_64bit(stats, TCP_NLA_DELIVERY_RATE, rate64, TCP_NLA_PAD);
3911 nla_put_u32(stats, TCP_NLA_SND_CWND, tcp_snd_cwnd(tp));
3912 nla_put_u32(stats, TCP_NLA_REORDERING, tp->reordering);
3913 nla_put_u32(stats, TCP_NLA_MIN_RTT, tcp_min_rtt(tp));
3915 nla_put_u8(stats, TCP_NLA_RECUR_RETRANS, inet_csk(sk)->icsk_retransmits);
3916 nla_put_u8(stats, TCP_NLA_DELIVERY_RATE_APP_LMT, !!tp->rate_app_limited);
3917 nla_put_u32(stats, TCP_NLA_SND_SSTHRESH, tp->snd_ssthresh);
3918 nla_put_u32(stats, TCP_NLA_DELIVERED, tp->delivered);
3919 nla_put_u32(stats, TCP_NLA_DELIVERED_CE, tp->delivered_ce);
3921 nla_put_u32(stats, TCP_NLA_SNDQ_SIZE, tp->write_seq - tp->snd_una);
3922 nla_put_u8(stats, TCP_NLA_CA_STATE, inet_csk(sk)->icsk_ca_state);
3924 nla_put_u64_64bit(stats, TCP_NLA_BYTES_SENT, tp->bytes_sent,
3926 nla_put_u64_64bit(stats, TCP_NLA_BYTES_RETRANS, tp->bytes_retrans,
3928 nla_put_u32(stats, TCP_NLA_DSACK_DUPS, tp->dsack_dups);
3929 nla_put_u32(stats, TCP_NLA_REORD_SEEN, tp->reord_seen);
3930 nla_put_u32(stats, TCP_NLA_SRTT, tp->srtt_us >> 3);
3931 nla_put_u16(stats, TCP_NLA_TIMEOUT_REHASH, tp->timeout_rehash);
3932 nla_put_u32(stats, TCP_NLA_BYTES_NOTSENT,
3933 max_t(int, 0, tp->write_seq - tp->snd_nxt));
3934 nla_put_u64_64bit(stats, TCP_NLA_EDT, orig_skb->skb_mstamp_ns,
3937 nla_put_u8(stats, TCP_NLA_TTL,
3938 tcp_skb_ttl_or_hop_limit(ack_skb));
3943 static int do_tcp_getsockopt(struct sock *sk, int level,
3944 int optname, char __user *optval, int __user *optlen)
3946 struct inet_connection_sock *icsk = inet_csk(sk);
3947 struct tcp_sock *tp = tcp_sk(sk);
3948 struct net *net = sock_net(sk);
3951 if (get_user(len, optlen))
3954 len = min_t(unsigned int, len, sizeof(int));
3961 val = tp->mss_cache;
3962 if (!val && ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
3963 val = tp->rx_opt.user_mss;
3965 val = tp->rx_opt.mss_clamp;
3968 val = !!(tp->nonagle&TCP_NAGLE_OFF);
3971 val = !!(tp->nonagle&TCP_NAGLE_CORK);
3974 val = keepalive_time_when(tp) / HZ;
3977 val = keepalive_intvl_when(tp) / HZ;
3980 val = keepalive_probes(tp);
3983 val = icsk->icsk_syn_retries ? :
3984 READ_ONCE(net->ipv4.sysctl_tcp_syn_retries);
3989 val = (val ? : READ_ONCE(net->ipv4.sysctl_tcp_fin_timeout)) / HZ;
3991 case TCP_DEFER_ACCEPT:
3992 val = retrans_to_secs(icsk->icsk_accept_queue.rskq_defer_accept,
3993 TCP_TIMEOUT_INIT / HZ, TCP_RTO_MAX / HZ);
3995 case TCP_WINDOW_CLAMP:
3996 val = tp->window_clamp;
3999 struct tcp_info info;
4001 if (get_user(len, optlen))
4004 tcp_get_info(sk, &info);
4006 len = min_t(unsigned int, len, sizeof(info));
4007 if (put_user(len, optlen))
4009 if (copy_to_user(optval, &info, len))
4014 const struct tcp_congestion_ops *ca_ops;
4015 union tcp_cc_info info;
4019 if (get_user(len, optlen))
4022 ca_ops = icsk->icsk_ca_ops;
4023 if (ca_ops && ca_ops->get_info)
4024 sz = ca_ops->get_info(sk, ~0U, &attr, &info);
4026 len = min_t(unsigned int, len, sz);
4027 if (put_user(len, optlen))
4029 if (copy_to_user(optval, &info, len))
4034 val = !inet_csk_in_pingpong_mode(sk);
4037 case TCP_CONGESTION:
4038 if (get_user(len, optlen))
4040 len = min_t(unsigned int, len, TCP_CA_NAME_MAX);
4041 if (put_user(len, optlen))
4043 if (copy_to_user(optval, icsk->icsk_ca_ops->name, len))
4048 if (get_user(len, optlen))
4050 len = min_t(unsigned int, len, TCP_ULP_NAME_MAX);
4051 if (!icsk->icsk_ulp_ops) {
4052 if (put_user(0, optlen))
4056 if (put_user(len, optlen))
4058 if (copy_to_user(optval, icsk->icsk_ulp_ops->name, len))
4062 case TCP_FASTOPEN_KEY: {
4063 u64 key[TCP_FASTOPEN_KEY_BUF_LENGTH / sizeof(u64)];
4064 unsigned int key_len;
4066 if (get_user(len, optlen))
4069 key_len = tcp_fastopen_get_cipher(net, icsk, key) *
4070 TCP_FASTOPEN_KEY_LENGTH;
4071 len = min_t(unsigned int, len, key_len);
4072 if (put_user(len, optlen))
4074 if (copy_to_user(optval, key, len))
4078 case TCP_THIN_LINEAR_TIMEOUTS:
4082 case TCP_THIN_DUPACK:
4090 case TCP_REPAIR_QUEUE:
4092 val = tp->repair_queue;
4097 case TCP_REPAIR_WINDOW: {
4098 struct tcp_repair_window opt;
4100 if (get_user(len, optlen))
4103 if (len != sizeof(opt))
4109 opt.snd_wl1 = tp->snd_wl1;
4110 opt.snd_wnd = tp->snd_wnd;
4111 opt.max_window = tp->max_window;
4112 opt.rcv_wnd = tp->rcv_wnd;
4113 opt.rcv_wup = tp->rcv_wup;
4115 if (copy_to_user(optval, &opt, len))
4120 if (tp->repair_queue == TCP_SEND_QUEUE)
4121 val = tp->write_seq;
4122 else if (tp->repair_queue == TCP_RECV_QUEUE)
4128 case TCP_USER_TIMEOUT:
4129 val = icsk->icsk_user_timeout;
4133 val = icsk->icsk_accept_queue.fastopenq.max_qlen;
4136 case TCP_FASTOPEN_CONNECT:
4137 val = tp->fastopen_connect;
4140 case TCP_FASTOPEN_NO_COOKIE:
4141 val = tp->fastopen_no_cookie;
4145 val = tp->tcp_tx_delay;
4149 val = tcp_time_stamp_raw() + tp->tsoffset;
4151 case TCP_NOTSENT_LOWAT:
4152 val = tp->notsent_lowat;
4155 val = tp->recvmsg_inq;
4160 case TCP_SAVED_SYN: {
4161 if (get_user(len, optlen))
4165 if (tp->saved_syn) {
4166 if (len < tcp_saved_syn_len(tp->saved_syn)) {
4167 if (put_user(tcp_saved_syn_len(tp->saved_syn),
4175 len = tcp_saved_syn_len(tp->saved_syn);
4176 if (put_user(len, optlen)) {
4180 if (copy_to_user(optval, tp->saved_syn->data, len)) {
4184 tcp_saved_syn_free(tp);
4189 if (put_user(len, optlen))
4195 case TCP_ZEROCOPY_RECEIVE: {
4196 struct scm_timestamping_internal tss;
4197 struct tcp_zerocopy_receive zc = {};
4200 if (get_user(len, optlen))
4203 len < offsetofend(struct tcp_zerocopy_receive, length))
4205 if (unlikely(len > sizeof(zc))) {
4206 err = check_zeroed_user(optval + sizeof(zc),
4209 return err == 0 ? -EINVAL : err;
4211 if (put_user(len, optlen))
4214 if (copy_from_user(&zc, optval, len))
4218 if (zc.msg_flags & ~(TCP_VALID_ZC_MSG_FLAGS))
4221 err = tcp_zerocopy_receive(sk, &zc, &tss);
4222 err = BPF_CGROUP_RUN_PROG_GETSOCKOPT_KERN(sk, level, optname,
4225 if (len >= offsetofend(struct tcp_zerocopy_receive, msg_flags))
4226 goto zerocopy_rcv_cmsg;
4228 case offsetofend(struct tcp_zerocopy_receive, msg_flags):
4229 goto zerocopy_rcv_cmsg;
4230 case offsetofend(struct tcp_zerocopy_receive, msg_controllen):
4231 case offsetofend(struct tcp_zerocopy_receive, msg_control):
4232 case offsetofend(struct tcp_zerocopy_receive, flags):
4233 case offsetofend(struct tcp_zerocopy_receive, copybuf_len):
4234 case offsetofend(struct tcp_zerocopy_receive, copybuf_address):
4235 case offsetofend(struct tcp_zerocopy_receive, err):
4236 goto zerocopy_rcv_sk_err;
4237 case offsetofend(struct tcp_zerocopy_receive, inq):
4238 goto zerocopy_rcv_inq;
4239 case offsetofend(struct tcp_zerocopy_receive, length):
4241 goto zerocopy_rcv_out;
4244 if (zc.msg_flags & TCP_CMSG_TS)
4245 tcp_zc_finalize_rx_tstamp(sk, &zc, &tss);
4248 zerocopy_rcv_sk_err:
4250 zc.err = sock_error(sk);
4252 zc.inq = tcp_inq_hint(sk);
4254 if (!err && copy_to_user(optval, &zc, len))
4260 return -ENOPROTOOPT;
4263 if (put_user(len, optlen))
4265 if (copy_to_user(optval, &val, len))
4270 bool tcp_bpf_bypass_getsockopt(int level, int optname)
4272 /* TCP do_tcp_getsockopt has optimized getsockopt implementation
4273 * to avoid extra socket lock for TCP_ZEROCOPY_RECEIVE.
4275 if (level == SOL_TCP && optname == TCP_ZEROCOPY_RECEIVE)
4280 EXPORT_SYMBOL(tcp_bpf_bypass_getsockopt);
4282 int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval,
4285 struct inet_connection_sock *icsk = inet_csk(sk);
4287 if (level != SOL_TCP)
4288 return icsk->icsk_af_ops->getsockopt(sk, level, optname,
4290 return do_tcp_getsockopt(sk, level, optname, optval, optlen);
4292 EXPORT_SYMBOL(tcp_getsockopt);
4294 #ifdef CONFIG_TCP_MD5SIG
4295 static DEFINE_PER_CPU(struct tcp_md5sig_pool, tcp_md5sig_pool);
4296 static DEFINE_MUTEX(tcp_md5sig_mutex);
4297 static bool tcp_md5sig_pool_populated = false;
4299 static void __tcp_alloc_md5sig_pool(void)
4301 struct crypto_ahash *hash;
4304 hash = crypto_alloc_ahash("md5", 0, CRYPTO_ALG_ASYNC);
4308 for_each_possible_cpu(cpu) {
4309 void *scratch = per_cpu(tcp_md5sig_pool, cpu).scratch;
4310 struct ahash_request *req;
4313 scratch = kmalloc_node(sizeof(union tcp_md5sum_block) +
4314 sizeof(struct tcphdr),
4319 per_cpu(tcp_md5sig_pool, cpu).scratch = scratch;
4321 if (per_cpu(tcp_md5sig_pool, cpu).md5_req)
4324 req = ahash_request_alloc(hash, GFP_KERNEL);
4328 ahash_request_set_callback(req, 0, NULL, NULL);
4330 per_cpu(tcp_md5sig_pool, cpu).md5_req = req;
4332 /* before setting tcp_md5sig_pool_populated, we must commit all writes
4333 * to memory. See smp_rmb() in tcp_get_md5sig_pool()
4336 tcp_md5sig_pool_populated = true;
4339 bool tcp_alloc_md5sig_pool(void)
4341 if (unlikely(!tcp_md5sig_pool_populated)) {
4342 mutex_lock(&tcp_md5sig_mutex);
4344 if (!tcp_md5sig_pool_populated) {
4345 __tcp_alloc_md5sig_pool();
4346 if (tcp_md5sig_pool_populated)
4347 static_branch_inc(&tcp_md5_needed);
4350 mutex_unlock(&tcp_md5sig_mutex);
4352 return tcp_md5sig_pool_populated;
4354 EXPORT_SYMBOL(tcp_alloc_md5sig_pool);
4358 * tcp_get_md5sig_pool - get md5sig_pool for this user
4360 * We use percpu structure, so if we succeed, we exit with preemption
4361 * and BH disabled, to make sure another thread or softirq handling
4362 * wont try to get same context.
4364 struct tcp_md5sig_pool *tcp_get_md5sig_pool(void)
4368 if (tcp_md5sig_pool_populated) {
4369 /* coupled with smp_wmb() in __tcp_alloc_md5sig_pool() */
4371 return this_cpu_ptr(&tcp_md5sig_pool);
4376 EXPORT_SYMBOL(tcp_get_md5sig_pool);
4378 int tcp_md5_hash_skb_data(struct tcp_md5sig_pool *hp,
4379 const struct sk_buff *skb, unsigned int header_len)
4381 struct scatterlist sg;
4382 const struct tcphdr *tp = tcp_hdr(skb);
4383 struct ahash_request *req = hp->md5_req;
4385 const unsigned int head_data_len = skb_headlen(skb) > header_len ?
4386 skb_headlen(skb) - header_len : 0;
4387 const struct skb_shared_info *shi = skb_shinfo(skb);
4388 struct sk_buff *frag_iter;
4390 sg_init_table(&sg, 1);
4392 sg_set_buf(&sg, ((u8 *) tp) + header_len, head_data_len);
4393 ahash_request_set_crypt(req, &sg, NULL, head_data_len);
4394 if (crypto_ahash_update(req))
4397 for (i = 0; i < shi->nr_frags; ++i) {
4398 const skb_frag_t *f = &shi->frags[i];
4399 unsigned int offset = skb_frag_off(f);
4400 struct page *page = skb_frag_page(f) + (offset >> PAGE_SHIFT);
4402 sg_set_page(&sg, page, skb_frag_size(f),
4403 offset_in_page(offset));
4404 ahash_request_set_crypt(req, &sg, NULL, skb_frag_size(f));
4405 if (crypto_ahash_update(req))
4409 skb_walk_frags(skb, frag_iter)
4410 if (tcp_md5_hash_skb_data(hp, frag_iter, 0))
4415 EXPORT_SYMBOL(tcp_md5_hash_skb_data);
4417 int tcp_md5_hash_key(struct tcp_md5sig_pool *hp, const struct tcp_md5sig_key *key)
4419 u8 keylen = READ_ONCE(key->keylen); /* paired with WRITE_ONCE() in tcp_md5_do_add */
4420 struct scatterlist sg;
4422 sg_init_one(&sg, key->key, keylen);
4423 ahash_request_set_crypt(hp->md5_req, &sg, NULL, keylen);
4425 /* We use data_race() because tcp_md5_do_add() might change key->key under us */
4426 return data_race(crypto_ahash_update(hp->md5_req));
4428 EXPORT_SYMBOL(tcp_md5_hash_key);
4430 /* Called with rcu_read_lock() */
4431 enum skb_drop_reason
4432 tcp_inbound_md5_hash(const struct sock *sk, const struct sk_buff *skb,
4433 const void *saddr, const void *daddr,
4434 int family, int dif, int sdif)
4437 * This gets called for each TCP segment that arrives
4438 * so we want to be efficient.
4439 * We have 3 drop cases:
4440 * o No MD5 hash and one expected.
4441 * o MD5 hash and we're not expecting one.
4442 * o MD5 hash and its wrong.
4444 const __u8 *hash_location = NULL;
4445 struct tcp_md5sig_key *hash_expected;
4446 const struct tcphdr *th = tcp_hdr(skb);
4447 struct tcp_sock *tp = tcp_sk(sk);
4448 int genhash, l3index;
4451 /* sdif set, means packet ingressed via a device
4452 * in an L3 domain and dif is set to the l3mdev
4454 l3index = sdif ? dif : 0;
4456 hash_expected = tcp_md5_do_lookup(sk, l3index, saddr, family);
4457 hash_location = tcp_parse_md5sig_option(th);
4459 /* We've parsed the options - do we have a hash? */
4460 if (!hash_expected && !hash_location)
4461 return SKB_NOT_DROPPED_YET;
4463 if (hash_expected && !hash_location) {
4464 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMD5NOTFOUND);
4465 return SKB_DROP_REASON_TCP_MD5NOTFOUND;
4468 if (!hash_expected && hash_location) {
4469 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMD5UNEXPECTED);
4470 return SKB_DROP_REASON_TCP_MD5UNEXPECTED;
4473 /* Check the signature.
4474 * To support dual stack listeners, we need to handle
4477 if (family == AF_INET)
4478 genhash = tcp_v4_md5_hash_skb(newhash,
4482 genhash = tp->af_specific->calc_md5_hash(newhash,
4486 if (genhash || memcmp(hash_location, newhash, 16) != 0) {
4487 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMD5FAILURE);
4488 if (family == AF_INET) {
4489 net_info_ratelimited("MD5 Hash failed for (%pI4, %d)->(%pI4, %d)%s L3 index %d\n",
4490 saddr, ntohs(th->source),
4491 daddr, ntohs(th->dest),
4492 genhash ? " tcp_v4_calc_md5_hash failed"
4495 net_info_ratelimited("MD5 Hash %s for [%pI6c]:%u->[%pI6c]:%u L3 index %d\n",
4496 genhash ? "failed" : "mismatch",
4497 saddr, ntohs(th->source),
4498 daddr, ntohs(th->dest), l3index);
4500 return SKB_DROP_REASON_TCP_MD5FAILURE;
4502 return SKB_NOT_DROPPED_YET;
4504 EXPORT_SYMBOL(tcp_inbound_md5_hash);
4508 void tcp_done(struct sock *sk)
4510 struct request_sock *req;
4512 /* We might be called with a new socket, after
4513 * inet_csk_prepare_forced_close() has been called
4514 * so we can not use lockdep_sock_is_held(sk)
4516 req = rcu_dereference_protected(tcp_sk(sk)->fastopen_rsk, 1);
4518 if (sk->sk_state == TCP_SYN_SENT || sk->sk_state == TCP_SYN_RECV)
4519 TCP_INC_STATS(sock_net(sk), TCP_MIB_ATTEMPTFAILS);
4521 tcp_set_state(sk, TCP_CLOSE);
4522 tcp_clear_xmit_timers(sk);
4524 reqsk_fastopen_remove(sk, req, false);
4526 sk->sk_shutdown = SHUTDOWN_MASK;
4528 if (!sock_flag(sk, SOCK_DEAD))
4529 sk->sk_state_change(sk);
4531 inet_csk_destroy_sock(sk);
4533 EXPORT_SYMBOL_GPL(tcp_done);
4535 int tcp_abort(struct sock *sk, int err)
4537 if (!sk_fullsock(sk)) {
4538 if (sk->sk_state == TCP_NEW_SYN_RECV) {
4539 struct request_sock *req = inet_reqsk(sk);
4542 inet_csk_reqsk_queue_drop(req->rsk_listener, req);
4549 /* Don't race with userspace socket closes such as tcp_close. */
4552 if (sk->sk_state == TCP_LISTEN) {
4553 tcp_set_state(sk, TCP_CLOSE);
4554 inet_csk_listen_stop(sk);
4557 /* Don't race with BH socket closes such as inet_csk_listen_stop. */
4561 if (!sock_flag(sk, SOCK_DEAD)) {
4563 /* This barrier is coupled with smp_rmb() in tcp_poll() */
4565 sk_error_report(sk);
4566 if (tcp_need_reset(sk->sk_state))
4567 tcp_send_active_reset(sk, GFP_ATOMIC);
4573 tcp_write_queue_purge(sk);
4577 EXPORT_SYMBOL_GPL(tcp_abort);
4579 extern struct tcp_congestion_ops tcp_reno;
4581 static __initdata unsigned long thash_entries;
4582 static int __init set_thash_entries(char *str)
4589 ret = kstrtoul(str, 0, &thash_entries);
4595 __setup("thash_entries=", set_thash_entries);
4597 static void __init tcp_init_mem(void)
4599 unsigned long limit = nr_free_buffer_pages() / 16;
4601 limit = max(limit, 128UL);
4602 sysctl_tcp_mem[0] = limit / 4 * 3; /* 4.68 % */
4603 sysctl_tcp_mem[1] = limit; /* 6.25 % */
4604 sysctl_tcp_mem[2] = sysctl_tcp_mem[0] * 2; /* 9.37 % */
4607 void __init tcp_init(void)
4609 int max_rshare, max_wshare, cnt;
4610 unsigned long limit;
4613 BUILD_BUG_ON(TCP_MIN_SND_MSS <= MAX_TCP_OPTION_SPACE);
4614 BUILD_BUG_ON(sizeof(struct tcp_skb_cb) >
4615 sizeof_field(struct sk_buff, cb));
4617 percpu_counter_init(&tcp_sockets_allocated, 0, GFP_KERNEL);
4619 timer_setup(&tcp_orphan_timer, tcp_orphan_update, TIMER_DEFERRABLE);
4620 mod_timer(&tcp_orphan_timer, jiffies + TCP_ORPHAN_TIMER_PERIOD);
4622 inet_hashinfo2_init(&tcp_hashinfo, "tcp_listen_portaddr_hash",
4623 thash_entries, 21, /* one slot per 2 MB*/
4625 tcp_hashinfo.bind_bucket_cachep =
4626 kmem_cache_create("tcp_bind_bucket",
4627 sizeof(struct inet_bind_bucket), 0,
4628 SLAB_HWCACHE_ALIGN | SLAB_PANIC |
4632 /* Size and allocate the main established and bind bucket
4635 * The methodology is similar to that of the buffer cache.
4637 tcp_hashinfo.ehash =
4638 alloc_large_system_hash("TCP established",
4639 sizeof(struct inet_ehash_bucket),
4641 17, /* one slot per 128 KB of memory */
4644 &tcp_hashinfo.ehash_mask,
4646 thash_entries ? 0 : 512 * 1024);
4647 for (i = 0; i <= tcp_hashinfo.ehash_mask; i++)
4648 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].chain, i);
4650 if (inet_ehash_locks_alloc(&tcp_hashinfo))
4651 panic("TCP: failed to alloc ehash_locks");
4652 tcp_hashinfo.bhash =
4653 alloc_large_system_hash("TCP bind",
4654 sizeof(struct inet_bind_hashbucket),
4655 tcp_hashinfo.ehash_mask + 1,
4656 17, /* one slot per 128 KB of memory */
4658 &tcp_hashinfo.bhash_size,
4662 tcp_hashinfo.bhash_size = 1U << tcp_hashinfo.bhash_size;
4663 for (i = 0; i < tcp_hashinfo.bhash_size; i++) {
4664 spin_lock_init(&tcp_hashinfo.bhash[i].lock);
4665 INIT_HLIST_HEAD(&tcp_hashinfo.bhash[i].chain);
4669 cnt = tcp_hashinfo.ehash_mask + 1;
4670 sysctl_tcp_max_orphans = cnt / 2;
4673 /* Set per-socket limits to no more than 1/128 the pressure threshold */
4674 limit = nr_free_buffer_pages() << (PAGE_SHIFT - 7);
4675 max_wshare = min(4UL*1024*1024, limit);
4676 max_rshare = min(6UL*1024*1024, limit);
4678 init_net.ipv4.sysctl_tcp_wmem[0] = SK_MEM_QUANTUM;
4679 init_net.ipv4.sysctl_tcp_wmem[1] = 16*1024;
4680 init_net.ipv4.sysctl_tcp_wmem[2] = max(64*1024, max_wshare);
4682 init_net.ipv4.sysctl_tcp_rmem[0] = SK_MEM_QUANTUM;
4683 init_net.ipv4.sysctl_tcp_rmem[1] = 131072;
4684 init_net.ipv4.sysctl_tcp_rmem[2] = max(131072, max_rshare);
4686 pr_info("Hash tables configured (established %u bind %u)\n",
4687 tcp_hashinfo.ehash_mask + 1, tcp_hashinfo.bhash_size);
4691 BUG_ON(tcp_register_congestion_control(&tcp_reno) != 0);