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 tp->snd_cwnd = 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 = 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, sock_net(sk)->ipv4.sysctl_tcp_wmem[1]);
456 WRITE_ONCE(sk->sk_rcvbuf, 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 sock_net(sk)->ipv4.sysctl_tcp_autocorking &&
690 !tcp_rtx_queue_empty(sk) &&
691 refcount_read(&sk->sk_wmem_alloc) > skb->truesize;
694 void tcp_push(struct sock *sk, int flags, int mss_now,
695 int nonagle, int size_goal)
697 struct tcp_sock *tp = tcp_sk(sk);
700 skb = tcp_write_queue_tail(sk);
703 if (!(flags & MSG_MORE) || forced_push(tp))
704 tcp_mark_push(tp, skb);
706 tcp_mark_urg(tp, flags);
708 if (tcp_should_autocork(sk, skb, size_goal)) {
710 /* avoid atomic op if TSQ_THROTTLED bit is already set */
711 if (!test_bit(TSQ_THROTTLED, &sk->sk_tsq_flags)) {
712 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPAUTOCORKING);
713 set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags);
715 /* It is possible TX completion already happened
716 * before we set TSQ_THROTTLED.
718 if (refcount_read(&sk->sk_wmem_alloc) > skb->truesize)
722 if (flags & MSG_MORE)
723 nonagle = TCP_NAGLE_CORK;
725 __tcp_push_pending_frames(sk, mss_now, nonagle);
728 static int tcp_splice_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb,
729 unsigned int offset, size_t len)
731 struct tcp_splice_state *tss = rd_desc->arg.data;
734 ret = skb_splice_bits(skb, skb->sk, offset, tss->pipe,
735 min(rd_desc->count, len), tss->flags);
737 rd_desc->count -= ret;
741 static int __tcp_splice_read(struct sock *sk, struct tcp_splice_state *tss)
743 /* Store TCP splice context information in read_descriptor_t. */
744 read_descriptor_t rd_desc = {
749 return tcp_read_sock(sk, &rd_desc, tcp_splice_data_recv);
753 * tcp_splice_read - splice data from TCP socket to a pipe
754 * @sock: socket to splice from
755 * @ppos: position (not valid)
756 * @pipe: pipe to splice to
757 * @len: number of bytes to splice
758 * @flags: splice modifier flags
761 * Will read pages from given socket and fill them into a pipe.
764 ssize_t tcp_splice_read(struct socket *sock, loff_t *ppos,
765 struct pipe_inode_info *pipe, size_t len,
768 struct sock *sk = sock->sk;
769 struct tcp_splice_state tss = {
778 sock_rps_record_flow(sk);
780 * We can't seek on a socket input
789 timeo = sock_rcvtimeo(sk, sock->file->f_flags & O_NONBLOCK);
791 ret = __tcp_splice_read(sk, &tss);
797 if (sock_flag(sk, SOCK_DONE))
800 ret = sock_error(sk);
803 if (sk->sk_shutdown & RCV_SHUTDOWN)
805 if (sk->sk_state == TCP_CLOSE) {
807 * This occurs when user tries to read
808 * from never connected socket.
817 /* if __tcp_splice_read() got nothing while we have
818 * an skb in receive queue, we do not want to loop.
819 * This might happen with URG data.
821 if (!skb_queue_empty(&sk->sk_receive_queue))
823 sk_wait_data(sk, &timeo, NULL);
824 if (signal_pending(current)) {
825 ret = sock_intr_errno(timeo);
838 if (sk->sk_err || sk->sk_state == TCP_CLOSE ||
839 (sk->sk_shutdown & RCV_SHUTDOWN) ||
840 signal_pending(current))
851 EXPORT_SYMBOL(tcp_splice_read);
853 struct sk_buff *tcp_stream_alloc_skb(struct sock *sk, int size, gfp_t gfp,
858 if (unlikely(tcp_under_memory_pressure(sk)))
859 sk_mem_reclaim_partial(sk);
861 skb = alloc_skb_fclone(size + MAX_TCP_HEADER, gfp);
865 skb->truesize = SKB_TRUESIZE(skb_end_offset(skb));
866 if (force_schedule) {
867 mem_scheduled = true;
868 sk_forced_mem_schedule(sk, skb->truesize);
870 mem_scheduled = sk_wmem_schedule(sk, skb->truesize);
872 if (likely(mem_scheduled)) {
873 skb_reserve(skb, MAX_TCP_HEADER);
874 skb->ip_summed = CHECKSUM_PARTIAL;
875 INIT_LIST_HEAD(&skb->tcp_tsorted_anchor);
880 sk->sk_prot->enter_memory_pressure(sk);
881 sk_stream_moderate_sndbuf(sk);
886 static unsigned int tcp_xmit_size_goal(struct sock *sk, u32 mss_now,
889 struct tcp_sock *tp = tcp_sk(sk);
890 u32 new_size_goal, size_goal;
895 /* Note : tcp_tso_autosize() will eventually split this later */
896 new_size_goal = sk->sk_gso_max_size - 1 - MAX_TCP_HEADER;
897 new_size_goal = tcp_bound_to_half_wnd(tp, new_size_goal);
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 static struct sk_buff *tcp_build_frag(struct sock *sk, int size_goal, int flags,
940 struct page *page, int offset, size_t *size)
942 struct sk_buff *skb = tcp_write_queue_tail(sk);
943 struct tcp_sock *tp = tcp_sk(sk);
947 if (!skb || (copy = size_goal - skb->len) <= 0 ||
948 !tcp_skb_can_collapse_to(skb)) {
950 if (!sk_stream_memory_free(sk))
953 skb = tcp_stream_alloc_skb(sk, 0, sk->sk_allocation,
954 tcp_rtx_and_write_queues_empty(sk));
958 #ifdef CONFIG_TLS_DEVICE
959 skb->decrypted = !!(flags & MSG_SENDPAGE_DECRYPTED);
961 tcp_skb_entail(sk, skb);
968 i = skb_shinfo(skb)->nr_frags;
969 can_coalesce = skb_can_coalesce(skb, i, page, offset);
970 if (!can_coalesce && i >= sysctl_max_skb_frags) {
971 tcp_mark_push(tp, skb);
974 if (!sk_wmem_schedule(sk, copy))
978 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
981 skb_fill_page_desc(skb, i, page, offset, copy);
984 if (!(flags & MSG_NO_SHARED_FRAGS))
985 skb_shinfo(skb)->flags |= SKBFL_SHARED_FRAG;
988 skb->data_len += copy;
989 skb->truesize += copy;
990 sk_wmem_queued_add(sk, copy);
991 sk_mem_charge(sk, copy);
992 WRITE_ONCE(tp->write_seq, tp->write_seq + copy);
993 TCP_SKB_CB(skb)->end_seq += copy;
994 tcp_skb_pcount_set(skb, 0);
1000 ssize_t do_tcp_sendpages(struct sock *sk, struct page *page, int offset,
1001 size_t size, int flags)
1003 struct tcp_sock *tp = tcp_sk(sk);
1004 int mss_now, size_goal;
1007 long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
1009 if (IS_ENABLED(CONFIG_DEBUG_VM) &&
1010 WARN_ONCE(!sendpage_ok(page),
1011 "page must not be a Slab one and have page_count > 0"))
1014 /* Wait for a connection to finish. One exception is TCP Fast Open
1015 * (passive side) where data is allowed to be sent before a connection
1016 * is fully established.
1018 if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) &&
1019 !tcp_passive_fastopen(sk)) {
1020 err = sk_stream_wait_connect(sk, &timeo);
1025 sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
1027 mss_now = tcp_send_mss(sk, &size_goal, flags);
1031 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
1035 struct sk_buff *skb;
1038 skb = tcp_build_frag(sk, size_goal, flags, page, offset, ©);
1040 goto wait_for_space;
1043 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
1051 if (skb->len < size_goal || (flags & MSG_OOB))
1054 if (forced_push(tp)) {
1055 tcp_mark_push(tp, skb);
1056 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
1057 } else if (skb == tcp_send_head(sk))
1058 tcp_push_one(sk, mss_now);
1062 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1063 tcp_push(sk, flags & ~MSG_MORE, mss_now,
1064 TCP_NAGLE_PUSH, size_goal);
1066 err = sk_stream_wait_memory(sk, &timeo);
1070 mss_now = tcp_send_mss(sk, &size_goal, flags);
1075 tcp_tx_timestamp(sk, sk->sk_tsflags);
1076 if (!(flags & MSG_SENDPAGE_NOTLAST))
1077 tcp_push(sk, flags, mss_now, tp->nonagle, size_goal);
1082 tcp_remove_empty_skb(sk);
1086 /* make sure we wake any epoll edge trigger waiter */
1087 if (unlikely(tcp_rtx_and_write_queues_empty(sk) && err == -EAGAIN)) {
1088 sk->sk_write_space(sk);
1089 tcp_chrono_stop(sk, TCP_CHRONO_SNDBUF_LIMITED);
1091 return sk_stream_error(sk, flags, err);
1093 EXPORT_SYMBOL_GPL(do_tcp_sendpages);
1095 int tcp_sendpage_locked(struct sock *sk, struct page *page, int offset,
1096 size_t size, int flags)
1098 if (!(sk->sk_route_caps & NETIF_F_SG))
1099 return sock_no_sendpage_locked(sk, page, offset, size, flags);
1101 tcp_rate_check_app_limited(sk); /* is sending application-limited? */
1103 return do_tcp_sendpages(sk, page, offset, size, flags);
1105 EXPORT_SYMBOL_GPL(tcp_sendpage_locked);
1107 int tcp_sendpage(struct sock *sk, struct page *page, int offset,
1108 size_t size, int flags)
1113 ret = tcp_sendpage_locked(sk, page, offset, size, flags);
1118 EXPORT_SYMBOL(tcp_sendpage);
1120 void tcp_free_fastopen_req(struct tcp_sock *tp)
1122 if (tp->fastopen_req) {
1123 kfree(tp->fastopen_req);
1124 tp->fastopen_req = NULL;
1128 static int tcp_sendmsg_fastopen(struct sock *sk, struct msghdr *msg,
1129 int *copied, size_t size,
1130 struct ubuf_info *uarg)
1132 struct tcp_sock *tp = tcp_sk(sk);
1133 struct inet_sock *inet = inet_sk(sk);
1134 struct sockaddr *uaddr = msg->msg_name;
1137 if (!(sock_net(sk)->ipv4.sysctl_tcp_fastopen & TFO_CLIENT_ENABLE) ||
1138 (uaddr && msg->msg_namelen >= sizeof(uaddr->sa_family) &&
1139 uaddr->sa_family == AF_UNSPEC))
1141 if (tp->fastopen_req)
1142 return -EALREADY; /* Another Fast Open is in progress */
1144 tp->fastopen_req = kzalloc(sizeof(struct tcp_fastopen_request),
1146 if (unlikely(!tp->fastopen_req))
1148 tp->fastopen_req->data = msg;
1149 tp->fastopen_req->size = size;
1150 tp->fastopen_req->uarg = uarg;
1152 if (inet->defer_connect) {
1153 err = tcp_connect(sk);
1154 /* Same failure procedure as in tcp_v4/6_connect */
1156 tcp_set_state(sk, TCP_CLOSE);
1157 inet->inet_dport = 0;
1158 sk->sk_route_caps = 0;
1161 flags = (msg->msg_flags & MSG_DONTWAIT) ? O_NONBLOCK : 0;
1162 err = __inet_stream_connect(sk->sk_socket, uaddr,
1163 msg->msg_namelen, flags, 1);
1164 /* fastopen_req could already be freed in __inet_stream_connect
1165 * if the connection times out or gets rst
1167 if (tp->fastopen_req) {
1168 *copied = tp->fastopen_req->copied;
1169 tcp_free_fastopen_req(tp);
1170 inet->defer_connect = 0;
1175 int tcp_sendmsg_locked(struct sock *sk, struct msghdr *msg, size_t size)
1177 struct tcp_sock *tp = tcp_sk(sk);
1178 struct ubuf_info *uarg = NULL;
1179 struct sk_buff *skb;
1180 struct sockcm_cookie sockc;
1181 int flags, err, copied = 0;
1182 int mss_now = 0, size_goal, copied_syn = 0;
1183 int process_backlog = 0;
1187 flags = msg->msg_flags;
1189 if (flags & MSG_ZEROCOPY && size && sock_flag(sk, SOCK_ZEROCOPY)) {
1190 skb = tcp_write_queue_tail(sk);
1191 uarg = msg_zerocopy_realloc(sk, size, skb_zcopy(skb));
1197 zc = sk->sk_route_caps & NETIF_F_SG;
1202 if (unlikely(flags & MSG_FASTOPEN || inet_sk(sk)->defer_connect) &&
1204 err = tcp_sendmsg_fastopen(sk, msg, &copied_syn, size, uarg);
1205 if (err == -EINPROGRESS && copied_syn > 0)
1211 timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
1213 tcp_rate_check_app_limited(sk); /* is sending application-limited? */
1215 /* Wait for a connection to finish. One exception is TCP Fast Open
1216 * (passive side) where data is allowed to be sent before a connection
1217 * is fully established.
1219 if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) &&
1220 !tcp_passive_fastopen(sk)) {
1221 err = sk_stream_wait_connect(sk, &timeo);
1226 if (unlikely(tp->repair)) {
1227 if (tp->repair_queue == TCP_RECV_QUEUE) {
1228 copied = tcp_send_rcvq(sk, msg, size);
1233 if (tp->repair_queue == TCP_NO_QUEUE)
1236 /* 'common' sending to sendq */
1239 sockcm_init(&sockc, sk);
1240 if (msg->msg_controllen) {
1241 err = sock_cmsg_send(sk, msg, &sockc);
1242 if (unlikely(err)) {
1248 /* This should be in poll */
1249 sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
1251 /* Ok commence sending. */
1255 mss_now = tcp_send_mss(sk, &size_goal, flags);
1258 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
1261 while (msg_data_left(msg)) {
1264 skb = tcp_write_queue_tail(sk);
1266 copy = size_goal - skb->len;
1268 if (copy <= 0 || !tcp_skb_can_collapse_to(skb)) {
1272 if (!sk_stream_memory_free(sk))
1273 goto wait_for_space;
1275 if (unlikely(process_backlog >= 16)) {
1276 process_backlog = 0;
1277 if (sk_flush_backlog(sk))
1280 first_skb = tcp_rtx_and_write_queues_empty(sk);
1281 skb = tcp_stream_alloc_skb(sk, 0, sk->sk_allocation,
1284 goto wait_for_space;
1288 tcp_skb_entail(sk, skb);
1291 /* All packets are restored as if they have
1292 * already been sent. skb_mstamp_ns isn't set to
1293 * avoid wrong rtt estimation.
1296 TCP_SKB_CB(skb)->sacked |= TCPCB_REPAIRED;
1299 /* Try to append data to the end of skb. */
1300 if (copy > msg_data_left(msg))
1301 copy = msg_data_left(msg);
1305 int i = skb_shinfo(skb)->nr_frags;
1306 struct page_frag *pfrag = sk_page_frag(sk);
1308 if (!sk_page_frag_refill(sk, pfrag))
1309 goto wait_for_space;
1311 if (!skb_can_coalesce(skb, i, pfrag->page,
1313 if (i >= sysctl_max_skb_frags) {
1314 tcp_mark_push(tp, skb);
1320 copy = min_t(int, copy, pfrag->size - pfrag->offset);
1322 /* skb changing from pure zc to mixed, must charge zc */
1323 if (unlikely(skb_zcopy_pure(skb))) {
1324 if (!sk_wmem_schedule(sk, skb->data_len))
1325 goto wait_for_space;
1327 sk_mem_charge(sk, skb->data_len);
1328 skb_shinfo(skb)->flags &= ~SKBFL_PURE_ZEROCOPY;
1331 if (!sk_wmem_schedule(sk, copy))
1332 goto wait_for_space;
1334 err = skb_copy_to_page_nocache(sk, &msg->msg_iter, skb,
1341 /* Update the skb. */
1343 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1345 skb_fill_page_desc(skb, i, pfrag->page,
1346 pfrag->offset, copy);
1347 page_ref_inc(pfrag->page);
1349 pfrag->offset += copy;
1351 /* First append to a fragless skb builds initial
1355 skb_shinfo(skb)->flags |= SKBFL_PURE_ZEROCOPY;
1357 if (!skb_zcopy_pure(skb)) {
1358 if (!sk_wmem_schedule(sk, copy))
1359 goto wait_for_space;
1362 err = skb_zerocopy_iter_stream(sk, skb, msg, copy, uarg);
1363 if (err == -EMSGSIZE || err == -EEXIST) {
1364 tcp_mark_push(tp, skb);
1373 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
1375 WRITE_ONCE(tp->write_seq, tp->write_seq + copy);
1376 TCP_SKB_CB(skb)->end_seq += copy;
1377 tcp_skb_pcount_set(skb, 0);
1380 if (!msg_data_left(msg)) {
1381 if (unlikely(flags & MSG_EOR))
1382 TCP_SKB_CB(skb)->eor = 1;
1386 if (skb->len < size_goal || (flags & MSG_OOB) || unlikely(tp->repair))
1389 if (forced_push(tp)) {
1390 tcp_mark_push(tp, skb);
1391 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
1392 } else if (skb == tcp_send_head(sk))
1393 tcp_push_one(sk, mss_now);
1397 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1399 tcp_push(sk, flags & ~MSG_MORE, mss_now,
1400 TCP_NAGLE_PUSH, size_goal);
1402 err = sk_stream_wait_memory(sk, &timeo);
1406 mss_now = tcp_send_mss(sk, &size_goal, flags);
1411 tcp_tx_timestamp(sk, sockc.tsflags);
1412 tcp_push(sk, flags, mss_now, tp->nonagle, size_goal);
1415 net_zcopy_put(uarg);
1416 return copied + copied_syn;
1419 tcp_remove_empty_skb(sk);
1421 if (copied + copied_syn)
1424 net_zcopy_put_abort(uarg, true);
1425 err = sk_stream_error(sk, flags, err);
1426 /* make sure we wake any epoll edge trigger waiter */
1427 if (unlikely(tcp_rtx_and_write_queues_empty(sk) && err == -EAGAIN)) {
1428 sk->sk_write_space(sk);
1429 tcp_chrono_stop(sk, TCP_CHRONO_SNDBUF_LIMITED);
1433 EXPORT_SYMBOL_GPL(tcp_sendmsg_locked);
1435 int tcp_sendmsg(struct sock *sk, struct msghdr *msg, size_t size)
1440 ret = tcp_sendmsg_locked(sk, msg, size);
1445 EXPORT_SYMBOL(tcp_sendmsg);
1448 * Handle reading urgent data. BSD has very simple semantics for
1449 * this, no blocking and very strange errors 8)
1452 static int tcp_recv_urg(struct sock *sk, struct msghdr *msg, int len, int flags)
1454 struct tcp_sock *tp = tcp_sk(sk);
1456 /* No URG data to read. */
1457 if (sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data ||
1458 tp->urg_data == TCP_URG_READ)
1459 return -EINVAL; /* Yes this is right ! */
1461 if (sk->sk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DONE))
1464 if (tp->urg_data & TCP_URG_VALID) {
1466 char c = tp->urg_data;
1468 if (!(flags & MSG_PEEK))
1469 WRITE_ONCE(tp->urg_data, TCP_URG_READ);
1471 /* Read urgent data. */
1472 msg->msg_flags |= MSG_OOB;
1475 if (!(flags & MSG_TRUNC))
1476 err = memcpy_to_msg(msg, &c, 1);
1479 msg->msg_flags |= MSG_TRUNC;
1481 return err ? -EFAULT : len;
1484 if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
1487 /* Fixed the recv(..., MSG_OOB) behaviour. BSD docs and
1488 * the available implementations agree in this case:
1489 * this call should never block, independent of the
1490 * blocking state of the socket.
1491 * Mike <pall@rz.uni-karlsruhe.de>
1496 static int tcp_peek_sndq(struct sock *sk, struct msghdr *msg, int len)
1498 struct sk_buff *skb;
1499 int copied = 0, err = 0;
1501 /* XXX -- need to support SO_PEEK_OFF */
1503 skb_rbtree_walk(skb, &sk->tcp_rtx_queue) {
1504 err = skb_copy_datagram_msg(skb, 0, msg, skb->len);
1510 skb_queue_walk(&sk->sk_write_queue, skb) {
1511 err = skb_copy_datagram_msg(skb, 0, msg, skb->len);
1518 return err ?: copied;
1521 /* Clean up the receive buffer for full frames taken by the user,
1522 * then send an ACK if necessary. COPIED is the number of bytes
1523 * tcp_recvmsg has given to the user so far, it speeds up the
1524 * calculation of whether or not we must ACK for the sake of
1527 void tcp_cleanup_rbuf(struct sock *sk, int copied)
1529 struct tcp_sock *tp = tcp_sk(sk);
1530 bool time_to_ack = false;
1532 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
1534 WARN(skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq),
1535 "cleanup rbuf bug: copied %X seq %X rcvnxt %X\n",
1536 tp->copied_seq, TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt);
1538 if (inet_csk_ack_scheduled(sk)) {
1539 const struct inet_connection_sock *icsk = inet_csk(sk);
1541 if (/* Once-per-two-segments ACK was not sent by tcp_input.c */
1542 tp->rcv_nxt - tp->rcv_wup > icsk->icsk_ack.rcv_mss ||
1544 * If this read emptied read buffer, we send ACK, if
1545 * connection is not bidirectional, user drained
1546 * receive buffer and there was a small segment
1550 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED2) ||
1551 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED) &&
1552 !inet_csk_in_pingpong_mode(sk))) &&
1553 !atomic_read(&sk->sk_rmem_alloc)))
1557 /* We send an ACK if we can now advertise a non-zero window
1558 * which has been raised "significantly".
1560 * Even if window raised up to infinity, do not send window open ACK
1561 * in states, where we will not receive more. It is useless.
1563 if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) {
1564 __u32 rcv_window_now = tcp_receive_window(tp);
1566 /* Optimize, __tcp_select_window() is not cheap. */
1567 if (2*rcv_window_now <= tp->window_clamp) {
1568 __u32 new_window = __tcp_select_window(sk);
1570 /* Send ACK now, if this read freed lots of space
1571 * in our buffer. Certainly, new_window is new window.
1572 * We can advertise it now, if it is not less than current one.
1573 * "Lots" means "at least twice" here.
1575 if (new_window && new_window >= 2 * rcv_window_now)
1583 void __sk_defer_free_flush(struct sock *sk)
1585 struct llist_node *head;
1586 struct sk_buff *skb, *n;
1588 head = llist_del_all(&sk->defer_list);
1589 llist_for_each_entry_safe(skb, n, head, ll_node) {
1591 skb_mark_not_on_list(skb);
1595 EXPORT_SYMBOL(__sk_defer_free_flush);
1597 static void tcp_eat_recv_skb(struct sock *sk, struct sk_buff *skb)
1599 __skb_unlink(skb, &sk->sk_receive_queue);
1600 if (likely(skb->destructor == sock_rfree)) {
1602 skb->destructor = NULL;
1604 if (!skb_queue_empty(&sk->sk_receive_queue) ||
1605 !llist_empty(&sk->defer_list)) {
1606 llist_add(&skb->ll_node, &sk->defer_list);
1613 static struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off)
1615 struct sk_buff *skb;
1618 while ((skb = skb_peek(&sk->sk_receive_queue)) != NULL) {
1619 offset = seq - TCP_SKB_CB(skb)->seq;
1620 if (unlikely(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
1621 pr_err_once("%s: found a SYN, please report !\n", __func__);
1624 if (offset < skb->len || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)) {
1628 /* This looks weird, but this can happen if TCP collapsing
1629 * splitted a fat GRO packet, while we released socket lock
1630 * in skb_splice_bits()
1632 tcp_eat_recv_skb(sk, skb);
1638 * This routine provides an alternative to tcp_recvmsg() for routines
1639 * that would like to handle copying from skbuffs directly in 'sendfile'
1642 * - It is assumed that the socket was locked by the caller.
1643 * - The routine does not block.
1644 * - At present, there is no support for reading OOB data
1645 * or for 'peeking' the socket using this routine
1646 * (although both would be easy to implement).
1648 int tcp_read_sock(struct sock *sk, read_descriptor_t *desc,
1649 sk_read_actor_t recv_actor)
1651 struct sk_buff *skb;
1652 struct tcp_sock *tp = tcp_sk(sk);
1653 u32 seq = tp->copied_seq;
1657 if (sk->sk_state == TCP_LISTEN)
1659 while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1660 if (offset < skb->len) {
1664 len = skb->len - offset;
1665 /* Stop reading if we hit a patch of urgent data */
1666 if (unlikely(tp->urg_data)) {
1667 u32 urg_offset = tp->urg_seq - seq;
1668 if (urg_offset < len)
1673 used = recv_actor(desc, skb, offset, len);
1678 } else if (used <= len) {
1683 /* If recv_actor drops the lock (e.g. TCP splice
1684 * receive) the skb pointer might be invalid when
1685 * getting here: tcp_collapse might have deleted it
1686 * while aggregating skbs from the socket queue.
1688 skb = tcp_recv_skb(sk, seq - 1, &offset);
1691 /* TCP coalescing might have appended data to the skb.
1692 * Try to splice more frags
1694 if (offset + 1 != skb->len)
1697 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) {
1698 tcp_eat_recv_skb(sk, skb);
1702 tcp_eat_recv_skb(sk, skb);
1705 WRITE_ONCE(tp->copied_seq, seq);
1707 WRITE_ONCE(tp->copied_seq, seq);
1709 tcp_rcv_space_adjust(sk);
1711 /* Clean up data we have read: This will do ACK frames. */
1713 tcp_recv_skb(sk, seq, &offset);
1714 tcp_cleanup_rbuf(sk, copied);
1718 EXPORT_SYMBOL(tcp_read_sock);
1720 int tcp_peek_len(struct socket *sock)
1722 return tcp_inq(sock->sk);
1724 EXPORT_SYMBOL(tcp_peek_len);
1726 /* Make sure sk_rcvbuf is big enough to satisfy SO_RCVLOWAT hint */
1727 int tcp_set_rcvlowat(struct sock *sk, int val)
1731 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK)
1732 cap = sk->sk_rcvbuf >> 1;
1734 cap = sock_net(sk)->ipv4.sysctl_tcp_rmem[2] >> 1;
1735 val = min(val, cap);
1736 WRITE_ONCE(sk->sk_rcvlowat, val ? : 1);
1738 /* Check if we need to signal EPOLLIN right now */
1741 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK)
1745 if (val > sk->sk_rcvbuf) {
1746 WRITE_ONCE(sk->sk_rcvbuf, val);
1747 tcp_sk(sk)->window_clamp = tcp_win_from_space(sk, val);
1751 EXPORT_SYMBOL(tcp_set_rcvlowat);
1753 void tcp_update_recv_tstamps(struct sk_buff *skb,
1754 struct scm_timestamping_internal *tss)
1757 tss->ts[0] = ktime_to_timespec64(skb->tstamp);
1759 tss->ts[0] = (struct timespec64) {0};
1761 if (skb_hwtstamps(skb)->hwtstamp)
1762 tss->ts[2] = ktime_to_timespec64(skb_hwtstamps(skb)->hwtstamp);
1764 tss->ts[2] = (struct timespec64) {0};
1768 static const struct vm_operations_struct tcp_vm_ops = {
1771 int tcp_mmap(struct file *file, struct socket *sock,
1772 struct vm_area_struct *vma)
1774 if (vma->vm_flags & (VM_WRITE | VM_EXEC))
1776 vma->vm_flags &= ~(VM_MAYWRITE | VM_MAYEXEC);
1778 /* Instruct vm_insert_page() to not mmap_read_lock(mm) */
1779 vma->vm_flags |= VM_MIXEDMAP;
1781 vma->vm_ops = &tcp_vm_ops;
1784 EXPORT_SYMBOL(tcp_mmap);
1786 static skb_frag_t *skb_advance_to_frag(struct sk_buff *skb, u32 offset_skb,
1791 offset_skb -= skb_headlen(skb);
1792 if ((int)offset_skb < 0 || skb_has_frag_list(skb))
1795 frag = skb_shinfo(skb)->frags;
1796 while (offset_skb) {
1797 if (skb_frag_size(frag) > offset_skb) {
1798 *offset_frag = offset_skb;
1801 offset_skb -= skb_frag_size(frag);
1808 static bool can_map_frag(const skb_frag_t *frag)
1810 return skb_frag_size(frag) == PAGE_SIZE && !skb_frag_off(frag);
1813 static int find_next_mappable_frag(const skb_frag_t *frag,
1814 int remaining_in_skb)
1818 if (likely(can_map_frag(frag)))
1821 while (offset < remaining_in_skb && !can_map_frag(frag)) {
1822 offset += skb_frag_size(frag);
1828 static void tcp_zerocopy_set_hint_for_skb(struct sock *sk,
1829 struct tcp_zerocopy_receive *zc,
1830 struct sk_buff *skb, u32 offset)
1832 u32 frag_offset, partial_frag_remainder = 0;
1833 int mappable_offset;
1836 /* worst case: skip to next skb. try to improve on this case below */
1837 zc->recv_skip_hint = skb->len - offset;
1839 /* Find the frag containing this offset (and how far into that frag) */
1840 frag = skb_advance_to_frag(skb, offset, &frag_offset);
1845 struct skb_shared_info *info = skb_shinfo(skb);
1847 /* We read part of the last frag, must recvmsg() rest of skb. */
1848 if (frag == &info->frags[info->nr_frags - 1])
1851 /* Else, we must at least read the remainder in this frag. */
1852 partial_frag_remainder = skb_frag_size(frag) - frag_offset;
1853 zc->recv_skip_hint -= partial_frag_remainder;
1857 /* partial_frag_remainder: If part way through a frag, must read rest.
1858 * mappable_offset: Bytes till next mappable frag, *not* counting bytes
1859 * in partial_frag_remainder.
1861 mappable_offset = find_next_mappable_frag(frag, zc->recv_skip_hint);
1862 zc->recv_skip_hint = mappable_offset + partial_frag_remainder;
1865 static int tcp_recvmsg_locked(struct sock *sk, struct msghdr *msg, size_t len,
1866 int nonblock, int flags,
1867 struct scm_timestamping_internal *tss,
1869 static int receive_fallback_to_copy(struct sock *sk,
1870 struct tcp_zerocopy_receive *zc, int inq,
1871 struct scm_timestamping_internal *tss)
1873 unsigned long copy_address = (unsigned long)zc->copybuf_address;
1874 struct msghdr msg = {};
1879 zc->recv_skip_hint = 0;
1881 if (copy_address != zc->copybuf_address)
1884 err = import_single_range(READ, (void __user *)copy_address,
1885 inq, &iov, &msg.msg_iter);
1889 err = tcp_recvmsg_locked(sk, &msg, inq, /*nonblock=*/1, /*flags=*/0,
1890 tss, &zc->msg_flags);
1894 zc->copybuf_len = err;
1895 if (likely(zc->copybuf_len)) {
1896 struct sk_buff *skb;
1899 skb = tcp_recv_skb(sk, tcp_sk(sk)->copied_seq, &offset);
1901 tcp_zerocopy_set_hint_for_skb(sk, zc, skb, offset);
1906 static int tcp_copy_straggler_data(struct tcp_zerocopy_receive *zc,
1907 struct sk_buff *skb, u32 copylen,
1908 u32 *offset, u32 *seq)
1910 unsigned long copy_address = (unsigned long)zc->copybuf_address;
1911 struct msghdr msg = {};
1915 if (copy_address != zc->copybuf_address)
1918 err = import_single_range(READ, (void __user *)copy_address,
1919 copylen, &iov, &msg.msg_iter);
1922 err = skb_copy_datagram_msg(skb, *offset, &msg, copylen);
1925 zc->recv_skip_hint -= copylen;
1928 return (__s32)copylen;
1931 static int tcp_zc_handle_leftover(struct tcp_zerocopy_receive *zc,
1933 struct sk_buff *skb,
1936 struct scm_timestamping_internal *tss)
1938 u32 offset, copylen = min_t(u32, copybuf_len, zc->recv_skip_hint);
1942 /* skb is null if inq < PAGE_SIZE. */
1944 offset = *seq - TCP_SKB_CB(skb)->seq;
1946 skb = tcp_recv_skb(sk, *seq, &offset);
1947 if (TCP_SKB_CB(skb)->has_rxtstamp) {
1948 tcp_update_recv_tstamps(skb, tss);
1949 zc->msg_flags |= TCP_CMSG_TS;
1953 zc->copybuf_len = tcp_copy_straggler_data(zc, skb, copylen, &offset,
1955 return zc->copybuf_len < 0 ? 0 : copylen;
1958 static int tcp_zerocopy_vm_insert_batch_error(struct vm_area_struct *vma,
1959 struct page **pending_pages,
1960 unsigned long pages_remaining,
1961 unsigned long *address,
1964 struct tcp_zerocopy_receive *zc,
1965 u32 total_bytes_to_map,
1968 /* At least one page did not map. Try zapping if we skipped earlier. */
1969 if (err == -EBUSY &&
1970 zc->flags & TCP_RECEIVE_ZEROCOPY_FLAG_TLB_CLEAN_HINT) {
1973 maybe_zap_len = total_bytes_to_map - /* All bytes to map */
1974 *length + /* Mapped or pending */
1975 (pages_remaining * PAGE_SIZE); /* Failed map. */
1976 zap_page_range(vma, *address, maybe_zap_len);
1981 unsigned long leftover_pages = pages_remaining;
1984 /* We called zap_page_range, try to reinsert. */
1985 err = vm_insert_pages(vma, *address,
1988 bytes_mapped = PAGE_SIZE * (leftover_pages - pages_remaining);
1989 *seq += bytes_mapped;
1990 *address += bytes_mapped;
1993 /* Either we were unable to zap, OR we zapped, retried an
1994 * insert, and still had an issue. Either ways, pages_remaining
1995 * is the number of pages we were unable to map, and we unroll
1996 * some state we speculatively touched before.
1998 const int bytes_not_mapped = PAGE_SIZE * pages_remaining;
2000 *length -= bytes_not_mapped;
2001 zc->recv_skip_hint += bytes_not_mapped;
2006 static int tcp_zerocopy_vm_insert_batch(struct vm_area_struct *vma,
2007 struct page **pages,
2008 unsigned int pages_to_map,
2009 unsigned long *address,
2012 struct tcp_zerocopy_receive *zc,
2013 u32 total_bytes_to_map)
2015 unsigned long pages_remaining = pages_to_map;
2016 unsigned int pages_mapped;
2017 unsigned int bytes_mapped;
2020 err = vm_insert_pages(vma, *address, pages, &pages_remaining);
2021 pages_mapped = pages_to_map - (unsigned int)pages_remaining;
2022 bytes_mapped = PAGE_SIZE * pages_mapped;
2023 /* Even if vm_insert_pages fails, it may have partially succeeded in
2024 * mapping (some but not all of the pages).
2026 *seq += bytes_mapped;
2027 *address += bytes_mapped;
2032 /* Error: maybe zap and retry + rollback state for failed inserts. */
2033 return tcp_zerocopy_vm_insert_batch_error(vma, pages + pages_mapped,
2034 pages_remaining, address, length, seq, zc, total_bytes_to_map,
2038 #define TCP_VALID_ZC_MSG_FLAGS (TCP_CMSG_TS)
2039 static void tcp_zc_finalize_rx_tstamp(struct sock *sk,
2040 struct tcp_zerocopy_receive *zc,
2041 struct scm_timestamping_internal *tss)
2043 unsigned long msg_control_addr;
2044 struct msghdr cmsg_dummy;
2046 msg_control_addr = (unsigned long)zc->msg_control;
2047 cmsg_dummy.msg_control = (void *)msg_control_addr;
2048 cmsg_dummy.msg_controllen =
2049 (__kernel_size_t)zc->msg_controllen;
2050 cmsg_dummy.msg_flags = in_compat_syscall()
2051 ? MSG_CMSG_COMPAT : 0;
2052 cmsg_dummy.msg_control_is_user = true;
2054 if (zc->msg_control == msg_control_addr &&
2055 zc->msg_controllen == cmsg_dummy.msg_controllen) {
2056 tcp_recv_timestamp(&cmsg_dummy, sk, tss);
2057 zc->msg_control = (__u64)
2058 ((uintptr_t)cmsg_dummy.msg_control);
2059 zc->msg_controllen =
2060 (__u64)cmsg_dummy.msg_controllen;
2061 zc->msg_flags = (__u32)cmsg_dummy.msg_flags;
2065 #define TCP_ZEROCOPY_PAGE_BATCH_SIZE 32
2066 static int tcp_zerocopy_receive(struct sock *sk,
2067 struct tcp_zerocopy_receive *zc,
2068 struct scm_timestamping_internal *tss)
2070 u32 length = 0, offset, vma_len, avail_len, copylen = 0;
2071 unsigned long address = (unsigned long)zc->address;
2072 struct page *pages[TCP_ZEROCOPY_PAGE_BATCH_SIZE];
2073 s32 copybuf_len = zc->copybuf_len;
2074 struct tcp_sock *tp = tcp_sk(sk);
2075 const skb_frag_t *frags = NULL;
2076 unsigned int pages_to_map = 0;
2077 struct vm_area_struct *vma;
2078 struct sk_buff *skb = NULL;
2079 u32 seq = tp->copied_seq;
2080 u32 total_bytes_to_map;
2081 int inq = tcp_inq(sk);
2084 zc->copybuf_len = 0;
2087 if (address & (PAGE_SIZE - 1) || address != zc->address)
2090 if (sk->sk_state == TCP_LISTEN)
2093 sock_rps_record_flow(sk);
2095 if (inq && inq <= copybuf_len)
2096 return receive_fallback_to_copy(sk, zc, inq, tss);
2098 if (inq < PAGE_SIZE) {
2100 zc->recv_skip_hint = inq;
2101 if (!inq && sock_flag(sk, SOCK_DONE))
2106 mmap_read_lock(current->mm);
2108 vma = vma_lookup(current->mm, address);
2109 if (!vma || vma->vm_ops != &tcp_vm_ops) {
2110 mmap_read_unlock(current->mm);
2113 vma_len = min_t(unsigned long, zc->length, vma->vm_end - address);
2114 avail_len = min_t(u32, vma_len, inq);
2115 total_bytes_to_map = avail_len & ~(PAGE_SIZE - 1);
2116 if (total_bytes_to_map) {
2117 if (!(zc->flags & TCP_RECEIVE_ZEROCOPY_FLAG_TLB_CLEAN_HINT))
2118 zap_page_range(vma, address, total_bytes_to_map);
2119 zc->length = total_bytes_to_map;
2120 zc->recv_skip_hint = 0;
2122 zc->length = avail_len;
2123 zc->recv_skip_hint = avail_len;
2126 while (length + PAGE_SIZE <= zc->length) {
2127 int mappable_offset;
2130 if (zc->recv_skip_hint < PAGE_SIZE) {
2134 if (zc->recv_skip_hint > 0)
2137 offset = seq - TCP_SKB_CB(skb)->seq;
2139 skb = tcp_recv_skb(sk, seq, &offset);
2142 if (TCP_SKB_CB(skb)->has_rxtstamp) {
2143 tcp_update_recv_tstamps(skb, tss);
2144 zc->msg_flags |= TCP_CMSG_TS;
2146 zc->recv_skip_hint = skb->len - offset;
2147 frags = skb_advance_to_frag(skb, offset, &offset_frag);
2148 if (!frags || offset_frag)
2152 mappable_offset = find_next_mappable_frag(frags,
2153 zc->recv_skip_hint);
2154 if (mappable_offset) {
2155 zc->recv_skip_hint = mappable_offset;
2158 page = skb_frag_page(frags);
2160 pages[pages_to_map++] = page;
2161 length += PAGE_SIZE;
2162 zc->recv_skip_hint -= PAGE_SIZE;
2164 if (pages_to_map == TCP_ZEROCOPY_PAGE_BATCH_SIZE ||
2165 zc->recv_skip_hint < PAGE_SIZE) {
2166 /* Either full batch, or we're about to go to next skb
2167 * (and we cannot unroll failed ops across skbs).
2169 ret = tcp_zerocopy_vm_insert_batch(vma, pages,
2173 total_bytes_to_map);
2180 ret = tcp_zerocopy_vm_insert_batch(vma, pages, pages_to_map,
2181 &address, &length, &seq,
2182 zc, total_bytes_to_map);
2185 mmap_read_unlock(current->mm);
2186 /* Try to copy straggler data. */
2188 copylen = tcp_zc_handle_leftover(zc, sk, skb, &seq, copybuf_len, tss);
2190 if (length + copylen) {
2191 WRITE_ONCE(tp->copied_seq, seq);
2192 tcp_rcv_space_adjust(sk);
2194 /* Clean up data we have read: This will do ACK frames. */
2195 tcp_recv_skb(sk, seq, &offset);
2196 tcp_cleanup_rbuf(sk, length + copylen);
2198 if (length == zc->length)
2199 zc->recv_skip_hint = 0;
2201 if (!zc->recv_skip_hint && sock_flag(sk, SOCK_DONE))
2204 zc->length = length;
2209 /* Similar to __sock_recv_timestamp, but does not require an skb */
2210 void tcp_recv_timestamp(struct msghdr *msg, const struct sock *sk,
2211 struct scm_timestamping_internal *tss)
2213 int new_tstamp = sock_flag(sk, SOCK_TSTAMP_NEW);
2214 bool has_timestamping = false;
2216 if (tss->ts[0].tv_sec || tss->ts[0].tv_nsec) {
2217 if (sock_flag(sk, SOCK_RCVTSTAMP)) {
2218 if (sock_flag(sk, SOCK_RCVTSTAMPNS)) {
2220 struct __kernel_timespec kts = {
2221 .tv_sec = tss->ts[0].tv_sec,
2222 .tv_nsec = tss->ts[0].tv_nsec,
2224 put_cmsg(msg, SOL_SOCKET, SO_TIMESTAMPNS_NEW,
2227 struct __kernel_old_timespec ts_old = {
2228 .tv_sec = tss->ts[0].tv_sec,
2229 .tv_nsec = tss->ts[0].tv_nsec,
2231 put_cmsg(msg, SOL_SOCKET, SO_TIMESTAMPNS_OLD,
2232 sizeof(ts_old), &ts_old);
2236 struct __kernel_sock_timeval stv = {
2237 .tv_sec = tss->ts[0].tv_sec,
2238 .tv_usec = tss->ts[0].tv_nsec / 1000,
2240 put_cmsg(msg, SOL_SOCKET, SO_TIMESTAMP_NEW,
2243 struct __kernel_old_timeval tv = {
2244 .tv_sec = tss->ts[0].tv_sec,
2245 .tv_usec = tss->ts[0].tv_nsec / 1000,
2247 put_cmsg(msg, SOL_SOCKET, SO_TIMESTAMP_OLD,
2253 if (sk->sk_tsflags & SOF_TIMESTAMPING_SOFTWARE)
2254 has_timestamping = true;
2256 tss->ts[0] = (struct timespec64) {0};
2259 if (tss->ts[2].tv_sec || tss->ts[2].tv_nsec) {
2260 if (sk->sk_tsflags & SOF_TIMESTAMPING_RAW_HARDWARE)
2261 has_timestamping = true;
2263 tss->ts[2] = (struct timespec64) {0};
2266 if (has_timestamping) {
2267 tss->ts[1] = (struct timespec64) {0};
2268 if (sock_flag(sk, SOCK_TSTAMP_NEW))
2269 put_cmsg_scm_timestamping64(msg, tss);
2271 put_cmsg_scm_timestamping(msg, tss);
2275 static int tcp_inq_hint(struct sock *sk)
2277 const struct tcp_sock *tp = tcp_sk(sk);
2278 u32 copied_seq = READ_ONCE(tp->copied_seq);
2279 u32 rcv_nxt = READ_ONCE(tp->rcv_nxt);
2282 inq = rcv_nxt - copied_seq;
2283 if (unlikely(inq < 0 || copied_seq != READ_ONCE(tp->copied_seq))) {
2285 inq = tp->rcv_nxt - tp->copied_seq;
2288 /* After receiving a FIN, tell the user-space to continue reading
2289 * by returning a non-zero inq.
2291 if (inq == 0 && sock_flag(sk, SOCK_DONE))
2297 * This routine copies from a sock struct into the user buffer.
2299 * Technical note: in 2.3 we work on _locked_ socket, so that
2300 * tricks with *seq access order and skb->users are not required.
2301 * Probably, code can be easily improved even more.
2304 static int tcp_recvmsg_locked(struct sock *sk, struct msghdr *msg, size_t len,
2305 int nonblock, int flags,
2306 struct scm_timestamping_internal *tss,
2309 struct tcp_sock *tp = tcp_sk(sk);
2315 int target; /* Read at least this many bytes */
2317 struct sk_buff *skb, *last;
2321 if (sk->sk_state == TCP_LISTEN)
2324 if (tp->recvmsg_inq)
2325 *cmsg_flags = TCP_CMSG_INQ;
2326 timeo = sock_rcvtimeo(sk, nonblock);
2328 /* Urgent data needs to be handled specially. */
2329 if (flags & MSG_OOB)
2332 if (unlikely(tp->repair)) {
2334 if (!(flags & MSG_PEEK))
2337 if (tp->repair_queue == TCP_SEND_QUEUE)
2341 if (tp->repair_queue == TCP_NO_QUEUE)
2344 /* 'common' recv queue MSG_PEEK-ing */
2347 seq = &tp->copied_seq;
2348 if (flags & MSG_PEEK) {
2349 peek_seq = tp->copied_seq;
2353 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
2358 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
2359 if (unlikely(tp->urg_data) && tp->urg_seq == *seq) {
2362 if (signal_pending(current)) {
2363 copied = timeo ? sock_intr_errno(timeo) : -EAGAIN;
2368 /* Next get a buffer. */
2370 last = skb_peek_tail(&sk->sk_receive_queue);
2371 skb_queue_walk(&sk->sk_receive_queue, skb) {
2373 /* Now that we have two receive queues this
2376 if (WARN(before(*seq, TCP_SKB_CB(skb)->seq),
2377 "TCP recvmsg seq # bug: copied %X, seq %X, rcvnxt %X, fl %X\n",
2378 *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt,
2382 offset = *seq - TCP_SKB_CB(skb)->seq;
2383 if (unlikely(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
2384 pr_err_once("%s: found a SYN, please report !\n", __func__);
2387 if (offset < skb->len)
2389 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
2391 WARN(!(flags & MSG_PEEK),
2392 "TCP recvmsg seq # bug 2: copied %X, seq %X, rcvnxt %X, fl %X\n",
2393 *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt, flags);
2396 /* Well, if we have backlog, try to process it now yet. */
2398 if (copied >= target && !READ_ONCE(sk->sk_backlog.tail))
2403 sk->sk_state == TCP_CLOSE ||
2404 (sk->sk_shutdown & RCV_SHUTDOWN) ||
2406 signal_pending(current))
2409 if (sock_flag(sk, SOCK_DONE))
2413 copied = sock_error(sk);
2417 if (sk->sk_shutdown & RCV_SHUTDOWN)
2420 if (sk->sk_state == TCP_CLOSE) {
2421 /* This occurs when user tries to read
2422 * from never connected socket.
2433 if (signal_pending(current)) {
2434 copied = sock_intr_errno(timeo);
2439 tcp_cleanup_rbuf(sk, copied);
2441 if (copied >= target) {
2442 /* Do not sleep, just process backlog. */
2443 __sk_flush_backlog(sk);
2445 sk_defer_free_flush(sk);
2446 sk_wait_data(sk, &timeo, last);
2449 if ((flags & MSG_PEEK) &&
2450 (peek_seq - copied - urg_hole != tp->copied_seq)) {
2451 net_dbg_ratelimited("TCP(%s:%d): Application bug, race in MSG_PEEK\n",
2453 task_pid_nr(current));
2454 peek_seq = tp->copied_seq;
2459 /* Ok so how much can we use? */
2460 used = skb->len - offset;
2464 /* Do we have urgent data here? */
2465 if (unlikely(tp->urg_data)) {
2466 u32 urg_offset = tp->urg_seq - *seq;
2467 if (urg_offset < used) {
2469 if (!sock_flag(sk, SOCK_URGINLINE)) {
2470 WRITE_ONCE(*seq, *seq + 1);
2482 if (!(flags & MSG_TRUNC)) {
2483 err = skb_copy_datagram_msg(skb, offset, msg, used);
2485 /* Exception. Bailout! */
2492 WRITE_ONCE(*seq, *seq + used);
2496 tcp_rcv_space_adjust(sk);
2499 if (unlikely(tp->urg_data) && after(tp->copied_seq, tp->urg_seq)) {
2500 WRITE_ONCE(tp->urg_data, 0);
2501 tcp_fast_path_check(sk);
2504 if (TCP_SKB_CB(skb)->has_rxtstamp) {
2505 tcp_update_recv_tstamps(skb, tss);
2506 *cmsg_flags |= TCP_CMSG_TS;
2509 if (used + offset < skb->len)
2512 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
2514 if (!(flags & MSG_PEEK))
2515 tcp_eat_recv_skb(sk, skb);
2519 /* Process the FIN. */
2520 WRITE_ONCE(*seq, *seq + 1);
2521 if (!(flags & MSG_PEEK))
2522 tcp_eat_recv_skb(sk, skb);
2526 /* According to UNIX98, msg_name/msg_namelen are ignored
2527 * on connected socket. I was just happy when found this 8) --ANK
2530 /* Clean up data we have read: This will do ACK frames. */
2531 tcp_cleanup_rbuf(sk, copied);
2538 err = tcp_recv_urg(sk, msg, len, flags);
2542 err = tcp_peek_sndq(sk, msg, len);
2546 int tcp_recvmsg(struct sock *sk, struct msghdr *msg, size_t len, int nonblock,
2547 int flags, int *addr_len)
2549 int cmsg_flags = 0, ret, inq;
2550 struct scm_timestamping_internal tss;
2552 if (unlikely(flags & MSG_ERRQUEUE))
2553 return inet_recv_error(sk, msg, len, addr_len);
2555 if (sk_can_busy_loop(sk) &&
2556 skb_queue_empty_lockless(&sk->sk_receive_queue) &&
2557 sk->sk_state == TCP_ESTABLISHED)
2558 sk_busy_loop(sk, nonblock);
2561 ret = tcp_recvmsg_locked(sk, msg, len, nonblock, flags, &tss,
2564 sk_defer_free_flush(sk);
2566 if (cmsg_flags && ret >= 0) {
2567 if (cmsg_flags & TCP_CMSG_TS)
2568 tcp_recv_timestamp(msg, sk, &tss);
2569 if (cmsg_flags & TCP_CMSG_INQ) {
2570 inq = tcp_inq_hint(sk);
2571 put_cmsg(msg, SOL_TCP, TCP_CM_INQ, sizeof(inq), &inq);
2576 EXPORT_SYMBOL(tcp_recvmsg);
2578 void tcp_set_state(struct sock *sk, int state)
2580 int oldstate = sk->sk_state;
2582 /* We defined a new enum for TCP states that are exported in BPF
2583 * so as not force the internal TCP states to be frozen. The
2584 * following checks will detect if an internal state value ever
2585 * differs from the BPF value. If this ever happens, then we will
2586 * need to remap the internal value to the BPF value before calling
2587 * tcp_call_bpf_2arg.
2589 BUILD_BUG_ON((int)BPF_TCP_ESTABLISHED != (int)TCP_ESTABLISHED);
2590 BUILD_BUG_ON((int)BPF_TCP_SYN_SENT != (int)TCP_SYN_SENT);
2591 BUILD_BUG_ON((int)BPF_TCP_SYN_RECV != (int)TCP_SYN_RECV);
2592 BUILD_BUG_ON((int)BPF_TCP_FIN_WAIT1 != (int)TCP_FIN_WAIT1);
2593 BUILD_BUG_ON((int)BPF_TCP_FIN_WAIT2 != (int)TCP_FIN_WAIT2);
2594 BUILD_BUG_ON((int)BPF_TCP_TIME_WAIT != (int)TCP_TIME_WAIT);
2595 BUILD_BUG_ON((int)BPF_TCP_CLOSE != (int)TCP_CLOSE);
2596 BUILD_BUG_ON((int)BPF_TCP_CLOSE_WAIT != (int)TCP_CLOSE_WAIT);
2597 BUILD_BUG_ON((int)BPF_TCP_LAST_ACK != (int)TCP_LAST_ACK);
2598 BUILD_BUG_ON((int)BPF_TCP_LISTEN != (int)TCP_LISTEN);
2599 BUILD_BUG_ON((int)BPF_TCP_CLOSING != (int)TCP_CLOSING);
2600 BUILD_BUG_ON((int)BPF_TCP_NEW_SYN_RECV != (int)TCP_NEW_SYN_RECV);
2601 BUILD_BUG_ON((int)BPF_TCP_MAX_STATES != (int)TCP_MAX_STATES);
2603 /* bpf uapi header bpf.h defines an anonymous enum with values
2604 * BPF_TCP_* used by bpf programs. Currently gcc built vmlinux
2605 * is able to emit this enum in DWARF due to the above BUILD_BUG_ON.
2606 * But clang built vmlinux does not have this enum in DWARF
2607 * since clang removes the above code before generating IR/debuginfo.
2608 * Let us explicitly emit the type debuginfo to ensure the
2609 * above-mentioned anonymous enum in the vmlinux DWARF and hence BTF
2610 * regardless of which compiler is used.
2612 BTF_TYPE_EMIT_ENUM(BPF_TCP_ESTABLISHED);
2614 if (BPF_SOCK_OPS_TEST_FLAG(tcp_sk(sk), BPF_SOCK_OPS_STATE_CB_FLAG))
2615 tcp_call_bpf_2arg(sk, BPF_SOCK_OPS_STATE_CB, oldstate, state);
2618 case TCP_ESTABLISHED:
2619 if (oldstate != TCP_ESTABLISHED)
2620 TCP_INC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
2624 if (oldstate == TCP_CLOSE_WAIT || oldstate == TCP_ESTABLISHED)
2625 TCP_INC_STATS(sock_net(sk), TCP_MIB_ESTABRESETS);
2627 sk->sk_prot->unhash(sk);
2628 if (inet_csk(sk)->icsk_bind_hash &&
2629 !(sk->sk_userlocks & SOCK_BINDPORT_LOCK))
2633 if (oldstate == TCP_ESTABLISHED)
2634 TCP_DEC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
2637 /* Change state AFTER socket is unhashed to avoid closed
2638 * socket sitting in hash tables.
2640 inet_sk_state_store(sk, state);
2642 EXPORT_SYMBOL_GPL(tcp_set_state);
2645 * State processing on a close. This implements the state shift for
2646 * sending our FIN frame. Note that we only send a FIN for some
2647 * states. A shutdown() may have already sent the FIN, or we may be
2651 static const unsigned char new_state[16] = {
2652 /* current state: new state: action: */
2653 [0 /* (Invalid) */] = TCP_CLOSE,
2654 [TCP_ESTABLISHED] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2655 [TCP_SYN_SENT] = TCP_CLOSE,
2656 [TCP_SYN_RECV] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2657 [TCP_FIN_WAIT1] = TCP_FIN_WAIT1,
2658 [TCP_FIN_WAIT2] = TCP_FIN_WAIT2,
2659 [TCP_TIME_WAIT] = TCP_CLOSE,
2660 [TCP_CLOSE] = TCP_CLOSE,
2661 [TCP_CLOSE_WAIT] = TCP_LAST_ACK | TCP_ACTION_FIN,
2662 [TCP_LAST_ACK] = TCP_LAST_ACK,
2663 [TCP_LISTEN] = TCP_CLOSE,
2664 [TCP_CLOSING] = TCP_CLOSING,
2665 [TCP_NEW_SYN_RECV] = TCP_CLOSE, /* should not happen ! */
2668 static int tcp_close_state(struct sock *sk)
2670 int next = (int)new_state[sk->sk_state];
2671 int ns = next & TCP_STATE_MASK;
2673 tcp_set_state(sk, ns);
2675 return next & TCP_ACTION_FIN;
2679 * Shutdown the sending side of a connection. Much like close except
2680 * that we don't receive shut down or sock_set_flag(sk, SOCK_DEAD).
2683 void tcp_shutdown(struct sock *sk, int how)
2685 /* We need to grab some memory, and put together a FIN,
2686 * and then put it into the queue to be sent.
2687 * Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
2689 if (!(how & SEND_SHUTDOWN))
2692 /* If we've already sent a FIN, or it's a closed state, skip this. */
2693 if ((1 << sk->sk_state) &
2694 (TCPF_ESTABLISHED | TCPF_SYN_SENT |
2695 TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) {
2696 /* Clear out any half completed packets. FIN if needed. */
2697 if (tcp_close_state(sk))
2701 EXPORT_SYMBOL(tcp_shutdown);
2703 int tcp_orphan_count_sum(void)
2707 for_each_possible_cpu(i)
2708 total += per_cpu(tcp_orphan_count, i);
2710 return max(total, 0);
2713 static int tcp_orphan_cache;
2714 static struct timer_list tcp_orphan_timer;
2715 #define TCP_ORPHAN_TIMER_PERIOD msecs_to_jiffies(100)
2717 static void tcp_orphan_update(struct timer_list *unused)
2719 WRITE_ONCE(tcp_orphan_cache, tcp_orphan_count_sum());
2720 mod_timer(&tcp_orphan_timer, jiffies + TCP_ORPHAN_TIMER_PERIOD);
2723 static bool tcp_too_many_orphans(int shift)
2725 return READ_ONCE(tcp_orphan_cache) << shift > sysctl_tcp_max_orphans;
2728 bool tcp_check_oom(struct sock *sk, int shift)
2730 bool too_many_orphans, out_of_socket_memory;
2732 too_many_orphans = tcp_too_many_orphans(shift);
2733 out_of_socket_memory = tcp_out_of_memory(sk);
2735 if (too_many_orphans)
2736 net_info_ratelimited("too many orphaned sockets\n");
2737 if (out_of_socket_memory)
2738 net_info_ratelimited("out of memory -- consider tuning tcp_mem\n");
2739 return too_many_orphans || out_of_socket_memory;
2742 void __tcp_close(struct sock *sk, long timeout)
2744 struct sk_buff *skb;
2745 int data_was_unread = 0;
2748 sk->sk_shutdown = SHUTDOWN_MASK;
2750 if (sk->sk_state == TCP_LISTEN) {
2751 tcp_set_state(sk, TCP_CLOSE);
2754 inet_csk_listen_stop(sk);
2756 goto adjudge_to_death;
2759 /* We need to flush the recv. buffs. We do this only on the
2760 * descriptor close, not protocol-sourced closes, because the
2761 * reader process may not have drained the data yet!
2763 while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) {
2764 u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq;
2766 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
2768 data_was_unread += len;
2774 /* If socket has been already reset (e.g. in tcp_reset()) - kill it. */
2775 if (sk->sk_state == TCP_CLOSE)
2776 goto adjudge_to_death;
2778 /* As outlined in RFC 2525, section 2.17, we send a RST here because
2779 * data was lost. To witness the awful effects of the old behavior of
2780 * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk
2781 * GET in an FTP client, suspend the process, wait for the client to
2782 * advertise a zero window, then kill -9 the FTP client, wheee...
2783 * Note: timeout is always zero in such a case.
2785 if (unlikely(tcp_sk(sk)->repair)) {
2786 sk->sk_prot->disconnect(sk, 0);
2787 } else if (data_was_unread) {
2788 /* Unread data was tossed, zap the connection. */
2789 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONCLOSE);
2790 tcp_set_state(sk, TCP_CLOSE);
2791 tcp_send_active_reset(sk, sk->sk_allocation);
2792 } else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) {
2793 /* Check zero linger _after_ checking for unread data. */
2794 sk->sk_prot->disconnect(sk, 0);
2795 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
2796 } else if (tcp_close_state(sk)) {
2797 /* We FIN if the application ate all the data before
2798 * zapping the connection.
2801 /* RED-PEN. Formally speaking, we have broken TCP state
2802 * machine. State transitions:
2804 * TCP_ESTABLISHED -> TCP_FIN_WAIT1
2805 * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible)
2806 * TCP_CLOSE_WAIT -> TCP_LAST_ACK
2808 * are legal only when FIN has been sent (i.e. in window),
2809 * rather than queued out of window. Purists blame.
2811 * F.e. "RFC state" is ESTABLISHED,
2812 * if Linux state is FIN-WAIT-1, but FIN is still not sent.
2814 * The visible declinations are that sometimes
2815 * we enter time-wait state, when it is not required really
2816 * (harmless), do not send active resets, when they are
2817 * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
2818 * they look as CLOSING or LAST_ACK for Linux)
2819 * Probably, I missed some more holelets.
2821 * XXX (TFO) - To start off we don't support SYN+ACK+FIN
2822 * in a single packet! (May consider it later but will
2823 * probably need API support or TCP_CORK SYN-ACK until
2824 * data is written and socket is closed.)
2829 sk_stream_wait_close(sk, timeout);
2832 state = sk->sk_state;
2838 /* remove backlog if any, without releasing ownership. */
2841 this_cpu_inc(tcp_orphan_count);
2843 /* Have we already been destroyed by a softirq or backlog? */
2844 if (state != TCP_CLOSE && sk->sk_state == TCP_CLOSE)
2847 /* This is a (useful) BSD violating of the RFC. There is a
2848 * problem with TCP as specified in that the other end could
2849 * keep a socket open forever with no application left this end.
2850 * We use a 1 minute timeout (about the same as BSD) then kill
2851 * our end. If they send after that then tough - BUT: long enough
2852 * that we won't make the old 4*rto = almost no time - whoops
2855 * Nope, it was not mistake. It is really desired behaviour
2856 * f.e. on http servers, when such sockets are useless, but
2857 * consume significant resources. Let's do it with special
2858 * linger2 option. --ANK
2861 if (sk->sk_state == TCP_FIN_WAIT2) {
2862 struct tcp_sock *tp = tcp_sk(sk);
2863 if (tp->linger2 < 0) {
2864 tcp_set_state(sk, TCP_CLOSE);
2865 tcp_send_active_reset(sk, GFP_ATOMIC);
2866 __NET_INC_STATS(sock_net(sk),
2867 LINUX_MIB_TCPABORTONLINGER);
2869 const int tmo = tcp_fin_time(sk);
2871 if (tmo > TCP_TIMEWAIT_LEN) {
2872 inet_csk_reset_keepalive_timer(sk,
2873 tmo - TCP_TIMEWAIT_LEN);
2875 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
2880 if (sk->sk_state != TCP_CLOSE) {
2882 if (tcp_check_oom(sk, 0)) {
2883 tcp_set_state(sk, TCP_CLOSE);
2884 tcp_send_active_reset(sk, GFP_ATOMIC);
2885 __NET_INC_STATS(sock_net(sk),
2886 LINUX_MIB_TCPABORTONMEMORY);
2887 } else if (!check_net(sock_net(sk))) {
2888 /* Not possible to send reset; just close */
2889 tcp_set_state(sk, TCP_CLOSE);
2893 if (sk->sk_state == TCP_CLOSE) {
2894 struct request_sock *req;
2896 req = rcu_dereference_protected(tcp_sk(sk)->fastopen_rsk,
2897 lockdep_sock_is_held(sk));
2898 /* We could get here with a non-NULL req if the socket is
2899 * aborted (e.g., closed with unread data) before 3WHS
2903 reqsk_fastopen_remove(sk, req, false);
2904 inet_csk_destroy_sock(sk);
2906 /* Otherwise, socket is reprieved until protocol close. */
2913 void tcp_close(struct sock *sk, long timeout)
2916 __tcp_close(sk, timeout);
2920 EXPORT_SYMBOL(tcp_close);
2922 /* These states need RST on ABORT according to RFC793 */
2924 static inline bool tcp_need_reset(int state)
2926 return (1 << state) &
2927 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 |
2928 TCPF_FIN_WAIT2 | TCPF_SYN_RECV);
2931 static void tcp_rtx_queue_purge(struct sock *sk)
2933 struct rb_node *p = rb_first(&sk->tcp_rtx_queue);
2935 tcp_sk(sk)->highest_sack = NULL;
2937 struct sk_buff *skb = rb_to_skb(p);
2940 /* Since we are deleting whole queue, no need to
2941 * list_del(&skb->tcp_tsorted_anchor)
2943 tcp_rtx_queue_unlink(skb, sk);
2944 tcp_wmem_free_skb(sk, skb);
2948 void tcp_write_queue_purge(struct sock *sk)
2950 struct sk_buff *skb;
2952 tcp_chrono_stop(sk, TCP_CHRONO_BUSY);
2953 while ((skb = __skb_dequeue(&sk->sk_write_queue)) != NULL) {
2954 tcp_skb_tsorted_anchor_cleanup(skb);
2955 tcp_wmem_free_skb(sk, skb);
2957 tcp_rtx_queue_purge(sk);
2958 INIT_LIST_HEAD(&tcp_sk(sk)->tsorted_sent_queue);
2960 tcp_clear_all_retrans_hints(tcp_sk(sk));
2961 tcp_sk(sk)->packets_out = 0;
2962 inet_csk(sk)->icsk_backoff = 0;
2965 int tcp_disconnect(struct sock *sk, int flags)
2967 struct inet_sock *inet = inet_sk(sk);
2968 struct inet_connection_sock *icsk = inet_csk(sk);
2969 struct tcp_sock *tp = tcp_sk(sk);
2970 int old_state = sk->sk_state;
2973 if (old_state != TCP_CLOSE)
2974 tcp_set_state(sk, TCP_CLOSE);
2976 /* ABORT function of RFC793 */
2977 if (old_state == TCP_LISTEN) {
2978 inet_csk_listen_stop(sk);
2979 } else if (unlikely(tp->repair)) {
2980 sk->sk_err = ECONNABORTED;
2981 } else if (tcp_need_reset(old_state) ||
2982 (tp->snd_nxt != tp->write_seq &&
2983 (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK))) {
2984 /* The last check adjusts for discrepancy of Linux wrt. RFC
2987 tcp_send_active_reset(sk, gfp_any());
2988 sk->sk_err = ECONNRESET;
2989 } else if (old_state == TCP_SYN_SENT)
2990 sk->sk_err = ECONNRESET;
2992 tcp_clear_xmit_timers(sk);
2993 __skb_queue_purge(&sk->sk_receive_queue);
2994 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt);
2995 WRITE_ONCE(tp->urg_data, 0);
2996 tcp_write_queue_purge(sk);
2997 tcp_fastopen_active_disable_ofo_check(sk);
2998 skb_rbtree_purge(&tp->out_of_order_queue);
3000 inet->inet_dport = 0;
3002 if (!(sk->sk_userlocks & SOCK_BINDADDR_LOCK))
3003 inet_reset_saddr(sk);
3005 sk->sk_shutdown = 0;
3006 sock_reset_flag(sk, SOCK_DONE);
3008 tp->mdev_us = jiffies_to_usecs(TCP_TIMEOUT_INIT);
3009 tp->rcv_rtt_last_tsecr = 0;
3011 seq = tp->write_seq + tp->max_window + 2;
3014 WRITE_ONCE(tp->write_seq, seq);
3016 icsk->icsk_backoff = 0;
3017 icsk->icsk_probes_out = 0;
3018 icsk->icsk_probes_tstamp = 0;
3019 icsk->icsk_rto = TCP_TIMEOUT_INIT;
3020 icsk->icsk_rto_min = TCP_RTO_MIN;
3021 icsk->icsk_delack_max = TCP_DELACK_MAX;
3022 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
3023 tp->snd_cwnd = TCP_INIT_CWND;
3024 tp->snd_cwnd_cnt = 0;
3025 tp->window_clamp = 0;
3027 tp->delivered_ce = 0;
3028 if (icsk->icsk_ca_ops->release)
3029 icsk->icsk_ca_ops->release(sk);
3030 memset(icsk->icsk_ca_priv, 0, sizeof(icsk->icsk_ca_priv));
3031 icsk->icsk_ca_initialized = 0;
3032 tcp_set_ca_state(sk, TCP_CA_Open);
3033 tp->is_sack_reneg = 0;
3034 tcp_clear_retrans(tp);
3035 tp->total_retrans = 0;
3036 inet_csk_delack_init(sk);
3037 /* Initialize rcv_mss to TCP_MIN_MSS to avoid division by 0
3038 * issue in __tcp_select_window()
3040 icsk->icsk_ack.rcv_mss = TCP_MIN_MSS;
3041 memset(&tp->rx_opt, 0, sizeof(tp->rx_opt));
3043 dst_release(sk->sk_rx_dst);
3044 sk->sk_rx_dst = NULL;
3045 tcp_saved_syn_free(tp);
3046 tp->compressed_ack = 0;
3050 tp->bytes_acked = 0;
3051 tp->bytes_received = 0;
3052 tp->bytes_retrans = 0;
3053 tp->data_segs_in = 0;
3054 tp->data_segs_out = 0;
3055 tp->duplicate_sack[0].start_seq = 0;
3056 tp->duplicate_sack[0].end_seq = 0;
3059 tp->retrans_out = 0;
3061 tp->tlp_high_seq = 0;
3062 tp->last_oow_ack_time = 0;
3063 /* There's a bubble in the pipe until at least the first ACK. */
3064 tp->app_limited = ~0U;
3065 tp->rack.mstamp = 0;
3066 tp->rack.advanced = 0;
3067 tp->rack.reo_wnd_steps = 1;
3068 tp->rack.last_delivered = 0;
3069 tp->rack.reo_wnd_persist = 0;
3070 tp->rack.dsack_seen = 0;
3071 tp->syn_data_acked = 0;
3072 tp->rx_opt.saw_tstamp = 0;
3073 tp->rx_opt.dsack = 0;
3074 tp->rx_opt.num_sacks = 0;
3075 tp->rcv_ooopack = 0;
3078 /* Clean up fastopen related fields */
3079 tcp_free_fastopen_req(tp);
3080 inet->defer_connect = 0;
3081 tp->fastopen_client_fail = 0;
3083 WARN_ON(inet->inet_num && !icsk->icsk_bind_hash);
3085 if (sk->sk_frag.page) {
3086 put_page(sk->sk_frag.page);
3087 sk->sk_frag.page = NULL;
3088 sk->sk_frag.offset = 0;
3090 sk_defer_free_flush(sk);
3091 sk_error_report(sk);
3094 EXPORT_SYMBOL(tcp_disconnect);
3096 static inline bool tcp_can_repair_sock(const struct sock *sk)
3098 return ns_capable(sock_net(sk)->user_ns, CAP_NET_ADMIN) &&
3099 (sk->sk_state != TCP_LISTEN);
3102 static int tcp_repair_set_window(struct tcp_sock *tp, sockptr_t optbuf, int len)
3104 struct tcp_repair_window opt;
3109 if (len != sizeof(opt))
3112 if (copy_from_sockptr(&opt, optbuf, sizeof(opt)))
3115 if (opt.max_window < opt.snd_wnd)
3118 if (after(opt.snd_wl1, tp->rcv_nxt + opt.rcv_wnd))
3121 if (after(opt.rcv_wup, tp->rcv_nxt))
3124 tp->snd_wl1 = opt.snd_wl1;
3125 tp->snd_wnd = opt.snd_wnd;
3126 tp->max_window = opt.max_window;
3128 tp->rcv_wnd = opt.rcv_wnd;
3129 tp->rcv_wup = opt.rcv_wup;
3134 static int tcp_repair_options_est(struct sock *sk, sockptr_t optbuf,
3137 struct tcp_sock *tp = tcp_sk(sk);
3138 struct tcp_repair_opt opt;
3141 while (len >= sizeof(opt)) {
3142 if (copy_from_sockptr_offset(&opt, optbuf, offset, sizeof(opt)))
3145 offset += sizeof(opt);
3148 switch (opt.opt_code) {
3150 tp->rx_opt.mss_clamp = opt.opt_val;
3155 u16 snd_wscale = opt.opt_val & 0xFFFF;
3156 u16 rcv_wscale = opt.opt_val >> 16;
3158 if (snd_wscale > TCP_MAX_WSCALE || rcv_wscale > TCP_MAX_WSCALE)
3161 tp->rx_opt.snd_wscale = snd_wscale;
3162 tp->rx_opt.rcv_wscale = rcv_wscale;
3163 tp->rx_opt.wscale_ok = 1;
3166 case TCPOPT_SACK_PERM:
3167 if (opt.opt_val != 0)
3170 tp->rx_opt.sack_ok |= TCP_SACK_SEEN;
3172 case TCPOPT_TIMESTAMP:
3173 if (opt.opt_val != 0)
3176 tp->rx_opt.tstamp_ok = 1;
3184 DEFINE_STATIC_KEY_FALSE(tcp_tx_delay_enabled);
3185 EXPORT_SYMBOL(tcp_tx_delay_enabled);
3187 static void tcp_enable_tx_delay(void)
3189 if (!static_branch_unlikely(&tcp_tx_delay_enabled)) {
3190 static int __tcp_tx_delay_enabled = 0;
3192 if (cmpxchg(&__tcp_tx_delay_enabled, 0, 1) == 0) {
3193 static_branch_enable(&tcp_tx_delay_enabled);
3194 pr_info("TCP_TX_DELAY enabled\n");
3199 /* When set indicates to always queue non-full frames. Later the user clears
3200 * this option and we transmit any pending partial frames in the queue. This is
3201 * meant to be used alongside sendfile() to get properly filled frames when the
3202 * user (for example) must write out headers with a write() call first and then
3203 * use sendfile to send out the data parts.
3205 * TCP_CORK can be set together with TCP_NODELAY and it is stronger than
3208 static void __tcp_sock_set_cork(struct sock *sk, bool on)
3210 struct tcp_sock *tp = tcp_sk(sk);
3213 tp->nonagle |= TCP_NAGLE_CORK;
3215 tp->nonagle &= ~TCP_NAGLE_CORK;
3216 if (tp->nonagle & TCP_NAGLE_OFF)
3217 tp->nonagle |= TCP_NAGLE_PUSH;
3218 tcp_push_pending_frames(sk);
3222 void tcp_sock_set_cork(struct sock *sk, bool on)
3225 __tcp_sock_set_cork(sk, on);
3228 EXPORT_SYMBOL(tcp_sock_set_cork);
3230 /* TCP_NODELAY is weaker than TCP_CORK, so that this option on corked socket is
3231 * remembered, but it is not activated until cork is cleared.
3233 * However, when TCP_NODELAY is set we make an explicit push, which overrides
3234 * even TCP_CORK for currently queued segments.
3236 static void __tcp_sock_set_nodelay(struct sock *sk, bool on)
3239 tcp_sk(sk)->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH;
3240 tcp_push_pending_frames(sk);
3242 tcp_sk(sk)->nonagle &= ~TCP_NAGLE_OFF;
3246 void tcp_sock_set_nodelay(struct sock *sk)
3249 __tcp_sock_set_nodelay(sk, true);
3252 EXPORT_SYMBOL(tcp_sock_set_nodelay);
3254 static void __tcp_sock_set_quickack(struct sock *sk, int val)
3257 inet_csk_enter_pingpong_mode(sk);
3261 inet_csk_exit_pingpong_mode(sk);
3262 if ((1 << sk->sk_state) & (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) &&
3263 inet_csk_ack_scheduled(sk)) {
3264 inet_csk(sk)->icsk_ack.pending |= ICSK_ACK_PUSHED;
3265 tcp_cleanup_rbuf(sk, 1);
3267 inet_csk_enter_pingpong_mode(sk);
3271 void tcp_sock_set_quickack(struct sock *sk, int val)
3274 __tcp_sock_set_quickack(sk, val);
3277 EXPORT_SYMBOL(tcp_sock_set_quickack);
3279 int tcp_sock_set_syncnt(struct sock *sk, int val)
3281 if (val < 1 || val > MAX_TCP_SYNCNT)
3285 inet_csk(sk)->icsk_syn_retries = val;
3289 EXPORT_SYMBOL(tcp_sock_set_syncnt);
3291 void tcp_sock_set_user_timeout(struct sock *sk, u32 val)
3294 inet_csk(sk)->icsk_user_timeout = val;
3297 EXPORT_SYMBOL(tcp_sock_set_user_timeout);
3299 int tcp_sock_set_keepidle_locked(struct sock *sk, int val)
3301 struct tcp_sock *tp = tcp_sk(sk);
3303 if (val < 1 || val > MAX_TCP_KEEPIDLE)
3306 tp->keepalive_time = val * HZ;
3307 if (sock_flag(sk, SOCK_KEEPOPEN) &&
3308 !((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN))) {
3309 u32 elapsed = keepalive_time_elapsed(tp);
3311 if (tp->keepalive_time > elapsed)
3312 elapsed = tp->keepalive_time - elapsed;
3315 inet_csk_reset_keepalive_timer(sk, elapsed);
3321 int tcp_sock_set_keepidle(struct sock *sk, int val)
3326 err = tcp_sock_set_keepidle_locked(sk, val);
3330 EXPORT_SYMBOL(tcp_sock_set_keepidle);
3332 int tcp_sock_set_keepintvl(struct sock *sk, int val)
3334 if (val < 1 || val > MAX_TCP_KEEPINTVL)
3338 tcp_sk(sk)->keepalive_intvl = val * HZ;
3342 EXPORT_SYMBOL(tcp_sock_set_keepintvl);
3344 int tcp_sock_set_keepcnt(struct sock *sk, int val)
3346 if (val < 1 || val > MAX_TCP_KEEPCNT)
3350 tcp_sk(sk)->keepalive_probes = val;
3354 EXPORT_SYMBOL(tcp_sock_set_keepcnt);
3356 int tcp_set_window_clamp(struct sock *sk, int val)
3358 struct tcp_sock *tp = tcp_sk(sk);
3361 if (sk->sk_state != TCP_CLOSE)
3363 tp->window_clamp = 0;
3365 tp->window_clamp = val < SOCK_MIN_RCVBUF / 2 ?
3366 SOCK_MIN_RCVBUF / 2 : val;
3367 tp->rcv_ssthresh = min(tp->rcv_wnd, tp->window_clamp);
3373 * Socket option code for TCP.
3375 static int do_tcp_setsockopt(struct sock *sk, int level, int optname,
3376 sockptr_t optval, unsigned int optlen)
3378 struct tcp_sock *tp = tcp_sk(sk);
3379 struct inet_connection_sock *icsk = inet_csk(sk);
3380 struct net *net = sock_net(sk);
3384 /* These are data/string values, all the others are ints */
3386 case TCP_CONGESTION: {
3387 char name[TCP_CA_NAME_MAX];
3392 val = strncpy_from_sockptr(name, optval,
3393 min_t(long, TCP_CA_NAME_MAX-1, optlen));
3399 err = tcp_set_congestion_control(sk, name, true,
3400 ns_capable(sock_net(sk)->user_ns,
3406 char name[TCP_ULP_NAME_MAX];
3411 val = strncpy_from_sockptr(name, optval,
3412 min_t(long, TCP_ULP_NAME_MAX - 1,
3419 err = tcp_set_ulp(sk, name);
3423 case TCP_FASTOPEN_KEY: {
3424 __u8 key[TCP_FASTOPEN_KEY_BUF_LENGTH];
3425 __u8 *backup_key = NULL;
3427 /* Allow a backup key as well to facilitate key rotation
3428 * First key is the active one.
3430 if (optlen != TCP_FASTOPEN_KEY_LENGTH &&
3431 optlen != TCP_FASTOPEN_KEY_BUF_LENGTH)
3434 if (copy_from_sockptr(key, optval, optlen))
3437 if (optlen == TCP_FASTOPEN_KEY_BUF_LENGTH)
3438 backup_key = key + TCP_FASTOPEN_KEY_LENGTH;
3440 return tcp_fastopen_reset_cipher(net, sk, key, backup_key);
3447 if (optlen < sizeof(int))
3450 if (copy_from_sockptr(&val, optval, sizeof(val)))
3457 /* Values greater than interface MTU won't take effect. However
3458 * at the point when this call is done we typically don't yet
3459 * know which interface is going to be used
3461 if (val && (val < TCP_MIN_MSS || val > MAX_TCP_WINDOW)) {
3465 tp->rx_opt.user_mss = val;
3469 __tcp_sock_set_nodelay(sk, val);
3472 case TCP_THIN_LINEAR_TIMEOUTS:
3473 if (val < 0 || val > 1)
3479 case TCP_THIN_DUPACK:
3480 if (val < 0 || val > 1)
3485 if (!tcp_can_repair_sock(sk))
3487 else if (val == TCP_REPAIR_ON) {
3489 sk->sk_reuse = SK_FORCE_REUSE;
3490 tp->repair_queue = TCP_NO_QUEUE;
3491 } else if (val == TCP_REPAIR_OFF) {
3493 sk->sk_reuse = SK_NO_REUSE;
3494 tcp_send_window_probe(sk);
3495 } else if (val == TCP_REPAIR_OFF_NO_WP) {
3497 sk->sk_reuse = SK_NO_REUSE;
3503 case TCP_REPAIR_QUEUE:
3506 else if ((unsigned int)val < TCP_QUEUES_NR)
3507 tp->repair_queue = val;
3513 if (sk->sk_state != TCP_CLOSE) {
3515 } else if (tp->repair_queue == TCP_SEND_QUEUE) {
3516 if (!tcp_rtx_queue_empty(sk))
3519 WRITE_ONCE(tp->write_seq, val);
3520 } else if (tp->repair_queue == TCP_RECV_QUEUE) {
3521 if (tp->rcv_nxt != tp->copied_seq) {
3524 WRITE_ONCE(tp->rcv_nxt, val);
3525 WRITE_ONCE(tp->copied_seq, val);
3532 case TCP_REPAIR_OPTIONS:
3535 else if (sk->sk_state == TCP_ESTABLISHED)
3536 err = tcp_repair_options_est(sk, optval, optlen);
3542 __tcp_sock_set_cork(sk, val);
3546 err = tcp_sock_set_keepidle_locked(sk, val);
3549 if (val < 1 || val > MAX_TCP_KEEPINTVL)
3552 tp->keepalive_intvl = val * HZ;
3555 if (val < 1 || val > MAX_TCP_KEEPCNT)
3558 tp->keepalive_probes = val;
3561 if (val < 1 || val > MAX_TCP_SYNCNT)
3564 icsk->icsk_syn_retries = val;
3568 /* 0: disable, 1: enable, 2: start from ether_header */
3569 if (val < 0 || val > 2)
3578 else if (val > TCP_FIN_TIMEOUT_MAX / HZ)
3579 tp->linger2 = TCP_FIN_TIMEOUT_MAX;
3581 tp->linger2 = val * HZ;
3584 case TCP_DEFER_ACCEPT:
3585 /* Translate value in seconds to number of retransmits */
3586 icsk->icsk_accept_queue.rskq_defer_accept =
3587 secs_to_retrans(val, TCP_TIMEOUT_INIT / HZ,
3591 case TCP_WINDOW_CLAMP:
3592 err = tcp_set_window_clamp(sk, val);
3596 __tcp_sock_set_quickack(sk, val);
3599 #ifdef CONFIG_TCP_MD5SIG
3601 case TCP_MD5SIG_EXT:
3602 err = tp->af_specific->md5_parse(sk, optname, optval, optlen);
3605 case TCP_USER_TIMEOUT:
3606 /* Cap the max time in ms TCP will retry or probe the window
3607 * before giving up and aborting (ETIMEDOUT) a connection.
3612 icsk->icsk_user_timeout = val;
3616 if (val >= 0 && ((1 << sk->sk_state) & (TCPF_CLOSE |
3618 tcp_fastopen_init_key_once(net);
3620 fastopen_queue_tune(sk, val);
3625 case TCP_FASTOPEN_CONNECT:
3626 if (val > 1 || val < 0) {
3628 } else if (net->ipv4.sysctl_tcp_fastopen & TFO_CLIENT_ENABLE) {
3629 if (sk->sk_state == TCP_CLOSE)
3630 tp->fastopen_connect = val;
3637 case TCP_FASTOPEN_NO_COOKIE:
3638 if (val > 1 || val < 0)
3640 else if (!((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
3643 tp->fastopen_no_cookie = val;
3649 tp->tsoffset = val - tcp_time_stamp_raw();
3651 case TCP_REPAIR_WINDOW:
3652 err = tcp_repair_set_window(tp, optval, optlen);
3654 case TCP_NOTSENT_LOWAT:
3655 tp->notsent_lowat = val;
3656 sk->sk_write_space(sk);
3659 if (val > 1 || val < 0)
3662 tp->recvmsg_inq = val;
3666 tcp_enable_tx_delay();
3667 tp->tcp_tx_delay = val;
3678 int tcp_setsockopt(struct sock *sk, int level, int optname, sockptr_t optval,
3679 unsigned int optlen)
3681 const struct inet_connection_sock *icsk = inet_csk(sk);
3683 if (level != SOL_TCP)
3684 return icsk->icsk_af_ops->setsockopt(sk, level, optname,
3686 return do_tcp_setsockopt(sk, level, optname, optval, optlen);
3688 EXPORT_SYMBOL(tcp_setsockopt);
3690 static void tcp_get_info_chrono_stats(const struct tcp_sock *tp,
3691 struct tcp_info *info)
3693 u64 stats[__TCP_CHRONO_MAX], total = 0;
3696 for (i = TCP_CHRONO_BUSY; i < __TCP_CHRONO_MAX; ++i) {
3697 stats[i] = tp->chrono_stat[i - 1];
3698 if (i == tp->chrono_type)
3699 stats[i] += tcp_jiffies32 - tp->chrono_start;
3700 stats[i] *= USEC_PER_SEC / HZ;
3704 info->tcpi_busy_time = total;
3705 info->tcpi_rwnd_limited = stats[TCP_CHRONO_RWND_LIMITED];
3706 info->tcpi_sndbuf_limited = stats[TCP_CHRONO_SNDBUF_LIMITED];
3709 /* Return information about state of tcp endpoint in API format. */
3710 void tcp_get_info(struct sock *sk, struct tcp_info *info)
3712 const struct tcp_sock *tp = tcp_sk(sk); /* iff sk_type == SOCK_STREAM */
3713 const struct inet_connection_sock *icsk = inet_csk(sk);
3719 memset(info, 0, sizeof(*info));
3720 if (sk->sk_type != SOCK_STREAM)
3723 info->tcpi_state = inet_sk_state_load(sk);
3725 /* Report meaningful fields for all TCP states, including listeners */
3726 rate = READ_ONCE(sk->sk_pacing_rate);
3727 rate64 = (rate != ~0UL) ? rate : ~0ULL;
3728 info->tcpi_pacing_rate = rate64;
3730 rate = READ_ONCE(sk->sk_max_pacing_rate);
3731 rate64 = (rate != ~0UL) ? rate : ~0ULL;
3732 info->tcpi_max_pacing_rate = rate64;
3734 info->tcpi_reordering = tp->reordering;
3735 info->tcpi_snd_cwnd = tp->snd_cwnd;
3737 if (info->tcpi_state == TCP_LISTEN) {
3738 /* listeners aliased fields :
3739 * tcpi_unacked -> Number of children ready for accept()
3740 * tcpi_sacked -> max backlog
3742 info->tcpi_unacked = READ_ONCE(sk->sk_ack_backlog);
3743 info->tcpi_sacked = READ_ONCE(sk->sk_max_ack_backlog);
3747 slow = lock_sock_fast(sk);
3749 info->tcpi_ca_state = icsk->icsk_ca_state;
3750 info->tcpi_retransmits = icsk->icsk_retransmits;
3751 info->tcpi_probes = icsk->icsk_probes_out;
3752 info->tcpi_backoff = icsk->icsk_backoff;
3754 if (tp->rx_opt.tstamp_ok)
3755 info->tcpi_options |= TCPI_OPT_TIMESTAMPS;
3756 if (tcp_is_sack(tp))
3757 info->tcpi_options |= TCPI_OPT_SACK;
3758 if (tp->rx_opt.wscale_ok) {
3759 info->tcpi_options |= TCPI_OPT_WSCALE;
3760 info->tcpi_snd_wscale = tp->rx_opt.snd_wscale;
3761 info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale;
3764 if (tp->ecn_flags & TCP_ECN_OK)
3765 info->tcpi_options |= TCPI_OPT_ECN;
3766 if (tp->ecn_flags & TCP_ECN_SEEN)
3767 info->tcpi_options |= TCPI_OPT_ECN_SEEN;
3768 if (tp->syn_data_acked)
3769 info->tcpi_options |= TCPI_OPT_SYN_DATA;
3771 info->tcpi_rto = jiffies_to_usecs(icsk->icsk_rto);
3772 info->tcpi_ato = jiffies_to_usecs(icsk->icsk_ack.ato);
3773 info->tcpi_snd_mss = tp->mss_cache;
3774 info->tcpi_rcv_mss = icsk->icsk_ack.rcv_mss;
3776 info->tcpi_unacked = tp->packets_out;
3777 info->tcpi_sacked = tp->sacked_out;
3779 info->tcpi_lost = tp->lost_out;
3780 info->tcpi_retrans = tp->retrans_out;
3782 now = tcp_jiffies32;
3783 info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime);
3784 info->tcpi_last_data_recv = jiffies_to_msecs(now - icsk->icsk_ack.lrcvtime);
3785 info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp);
3787 info->tcpi_pmtu = icsk->icsk_pmtu_cookie;
3788 info->tcpi_rcv_ssthresh = tp->rcv_ssthresh;
3789 info->tcpi_rtt = tp->srtt_us >> 3;
3790 info->tcpi_rttvar = tp->mdev_us >> 2;
3791 info->tcpi_snd_ssthresh = tp->snd_ssthresh;
3792 info->tcpi_advmss = tp->advmss;
3794 info->tcpi_rcv_rtt = tp->rcv_rtt_est.rtt_us >> 3;
3795 info->tcpi_rcv_space = tp->rcvq_space.space;
3797 info->tcpi_total_retrans = tp->total_retrans;
3799 info->tcpi_bytes_acked = tp->bytes_acked;
3800 info->tcpi_bytes_received = tp->bytes_received;
3801 info->tcpi_notsent_bytes = max_t(int, 0, tp->write_seq - tp->snd_nxt);
3802 tcp_get_info_chrono_stats(tp, info);
3804 info->tcpi_segs_out = tp->segs_out;
3806 /* segs_in and data_segs_in can be updated from tcp_segs_in() from BH */
3807 info->tcpi_segs_in = READ_ONCE(tp->segs_in);
3808 info->tcpi_data_segs_in = READ_ONCE(tp->data_segs_in);
3810 info->tcpi_min_rtt = tcp_min_rtt(tp);
3811 info->tcpi_data_segs_out = tp->data_segs_out;
3813 info->tcpi_delivery_rate_app_limited = tp->rate_app_limited ? 1 : 0;
3814 rate64 = tcp_compute_delivery_rate(tp);
3816 info->tcpi_delivery_rate = rate64;
3817 info->tcpi_delivered = tp->delivered;
3818 info->tcpi_delivered_ce = tp->delivered_ce;
3819 info->tcpi_bytes_sent = tp->bytes_sent;
3820 info->tcpi_bytes_retrans = tp->bytes_retrans;
3821 info->tcpi_dsack_dups = tp->dsack_dups;
3822 info->tcpi_reord_seen = tp->reord_seen;
3823 info->tcpi_rcv_ooopack = tp->rcv_ooopack;
3824 info->tcpi_snd_wnd = tp->snd_wnd;
3825 info->tcpi_fastopen_client_fail = tp->fastopen_client_fail;
3826 unlock_sock_fast(sk, slow);
3828 EXPORT_SYMBOL_GPL(tcp_get_info);
3830 static size_t tcp_opt_stats_get_size(void)
3833 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_BUSY */
3834 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_RWND_LIMITED */
3835 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_SNDBUF_LIMITED */
3836 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_DATA_SEGS_OUT */
3837 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_TOTAL_RETRANS */
3838 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_PACING_RATE */
3839 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_DELIVERY_RATE */
3840 nla_total_size(sizeof(u32)) + /* TCP_NLA_SND_CWND */
3841 nla_total_size(sizeof(u32)) + /* TCP_NLA_REORDERING */
3842 nla_total_size(sizeof(u32)) + /* TCP_NLA_MIN_RTT */
3843 nla_total_size(sizeof(u8)) + /* TCP_NLA_RECUR_RETRANS */
3844 nla_total_size(sizeof(u8)) + /* TCP_NLA_DELIVERY_RATE_APP_LMT */
3845 nla_total_size(sizeof(u32)) + /* TCP_NLA_SNDQ_SIZE */
3846 nla_total_size(sizeof(u8)) + /* TCP_NLA_CA_STATE */
3847 nla_total_size(sizeof(u32)) + /* TCP_NLA_SND_SSTHRESH */
3848 nla_total_size(sizeof(u32)) + /* TCP_NLA_DELIVERED */
3849 nla_total_size(sizeof(u32)) + /* TCP_NLA_DELIVERED_CE */
3850 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_BYTES_SENT */
3851 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_BYTES_RETRANS */
3852 nla_total_size(sizeof(u32)) + /* TCP_NLA_DSACK_DUPS */
3853 nla_total_size(sizeof(u32)) + /* TCP_NLA_REORD_SEEN */
3854 nla_total_size(sizeof(u32)) + /* TCP_NLA_SRTT */
3855 nla_total_size(sizeof(u16)) + /* TCP_NLA_TIMEOUT_REHASH */
3856 nla_total_size(sizeof(u32)) + /* TCP_NLA_BYTES_NOTSENT */
3857 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_EDT */
3858 nla_total_size(sizeof(u8)) + /* TCP_NLA_TTL */
3862 /* Returns TTL or hop limit of an incoming packet from skb. */
3863 static u8 tcp_skb_ttl_or_hop_limit(const struct sk_buff *skb)
3865 if (skb->protocol == htons(ETH_P_IP))
3866 return ip_hdr(skb)->ttl;
3867 else if (skb->protocol == htons(ETH_P_IPV6))
3868 return ipv6_hdr(skb)->hop_limit;
3873 struct sk_buff *tcp_get_timestamping_opt_stats(const struct sock *sk,
3874 const struct sk_buff *orig_skb,
3875 const struct sk_buff *ack_skb)
3877 const struct tcp_sock *tp = tcp_sk(sk);
3878 struct sk_buff *stats;
3879 struct tcp_info info;
3883 stats = alloc_skb(tcp_opt_stats_get_size(), GFP_ATOMIC);
3887 tcp_get_info_chrono_stats(tp, &info);
3888 nla_put_u64_64bit(stats, TCP_NLA_BUSY,
3889 info.tcpi_busy_time, TCP_NLA_PAD);
3890 nla_put_u64_64bit(stats, TCP_NLA_RWND_LIMITED,
3891 info.tcpi_rwnd_limited, TCP_NLA_PAD);
3892 nla_put_u64_64bit(stats, TCP_NLA_SNDBUF_LIMITED,
3893 info.tcpi_sndbuf_limited, TCP_NLA_PAD);
3894 nla_put_u64_64bit(stats, TCP_NLA_DATA_SEGS_OUT,
3895 tp->data_segs_out, TCP_NLA_PAD);
3896 nla_put_u64_64bit(stats, TCP_NLA_TOTAL_RETRANS,
3897 tp->total_retrans, TCP_NLA_PAD);
3899 rate = READ_ONCE(sk->sk_pacing_rate);
3900 rate64 = (rate != ~0UL) ? rate : ~0ULL;
3901 nla_put_u64_64bit(stats, TCP_NLA_PACING_RATE, rate64, TCP_NLA_PAD);
3903 rate64 = tcp_compute_delivery_rate(tp);
3904 nla_put_u64_64bit(stats, TCP_NLA_DELIVERY_RATE, rate64, TCP_NLA_PAD);
3906 nla_put_u32(stats, TCP_NLA_SND_CWND, tp->snd_cwnd);
3907 nla_put_u32(stats, TCP_NLA_REORDERING, tp->reordering);
3908 nla_put_u32(stats, TCP_NLA_MIN_RTT, tcp_min_rtt(tp));
3910 nla_put_u8(stats, TCP_NLA_RECUR_RETRANS, inet_csk(sk)->icsk_retransmits);
3911 nla_put_u8(stats, TCP_NLA_DELIVERY_RATE_APP_LMT, !!tp->rate_app_limited);
3912 nla_put_u32(stats, TCP_NLA_SND_SSTHRESH, tp->snd_ssthresh);
3913 nla_put_u32(stats, TCP_NLA_DELIVERED, tp->delivered);
3914 nla_put_u32(stats, TCP_NLA_DELIVERED_CE, tp->delivered_ce);
3916 nla_put_u32(stats, TCP_NLA_SNDQ_SIZE, tp->write_seq - tp->snd_una);
3917 nla_put_u8(stats, TCP_NLA_CA_STATE, inet_csk(sk)->icsk_ca_state);
3919 nla_put_u64_64bit(stats, TCP_NLA_BYTES_SENT, tp->bytes_sent,
3921 nla_put_u64_64bit(stats, TCP_NLA_BYTES_RETRANS, tp->bytes_retrans,
3923 nla_put_u32(stats, TCP_NLA_DSACK_DUPS, tp->dsack_dups);
3924 nla_put_u32(stats, TCP_NLA_REORD_SEEN, tp->reord_seen);
3925 nla_put_u32(stats, TCP_NLA_SRTT, tp->srtt_us >> 3);
3926 nla_put_u16(stats, TCP_NLA_TIMEOUT_REHASH, tp->timeout_rehash);
3927 nla_put_u32(stats, TCP_NLA_BYTES_NOTSENT,
3928 max_t(int, 0, tp->write_seq - tp->snd_nxt));
3929 nla_put_u64_64bit(stats, TCP_NLA_EDT, orig_skb->skb_mstamp_ns,
3932 nla_put_u8(stats, TCP_NLA_TTL,
3933 tcp_skb_ttl_or_hop_limit(ack_skb));
3938 static int do_tcp_getsockopt(struct sock *sk, int level,
3939 int optname, char __user *optval, int __user *optlen)
3941 struct inet_connection_sock *icsk = inet_csk(sk);
3942 struct tcp_sock *tp = tcp_sk(sk);
3943 struct net *net = sock_net(sk);
3946 if (get_user(len, optlen))
3949 len = min_t(unsigned int, len, sizeof(int));
3956 val = tp->mss_cache;
3957 if (!val && ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
3958 val = tp->rx_opt.user_mss;
3960 val = tp->rx_opt.mss_clamp;
3963 val = !!(tp->nonagle&TCP_NAGLE_OFF);
3966 val = !!(tp->nonagle&TCP_NAGLE_CORK);
3969 val = keepalive_time_when(tp) / HZ;
3972 val = keepalive_intvl_when(tp) / HZ;
3975 val = keepalive_probes(tp);
3978 val = icsk->icsk_syn_retries ? : net->ipv4.sysctl_tcp_syn_retries;
3983 val = (val ? : net->ipv4.sysctl_tcp_fin_timeout) / HZ;
3985 case TCP_DEFER_ACCEPT:
3986 val = retrans_to_secs(icsk->icsk_accept_queue.rskq_defer_accept,
3987 TCP_TIMEOUT_INIT / HZ, TCP_RTO_MAX / HZ);
3989 case TCP_WINDOW_CLAMP:
3990 val = tp->window_clamp;
3993 struct tcp_info info;
3995 if (get_user(len, optlen))
3998 tcp_get_info(sk, &info);
4000 len = min_t(unsigned int, len, sizeof(info));
4001 if (put_user(len, optlen))
4003 if (copy_to_user(optval, &info, len))
4008 const struct tcp_congestion_ops *ca_ops;
4009 union tcp_cc_info info;
4013 if (get_user(len, optlen))
4016 ca_ops = icsk->icsk_ca_ops;
4017 if (ca_ops && ca_ops->get_info)
4018 sz = ca_ops->get_info(sk, ~0U, &attr, &info);
4020 len = min_t(unsigned int, len, sz);
4021 if (put_user(len, optlen))
4023 if (copy_to_user(optval, &info, len))
4028 val = !inet_csk_in_pingpong_mode(sk);
4031 case TCP_CONGESTION:
4032 if (get_user(len, optlen))
4034 len = min_t(unsigned int, len, TCP_CA_NAME_MAX);
4035 if (put_user(len, optlen))
4037 if (copy_to_user(optval, icsk->icsk_ca_ops->name, len))
4042 if (get_user(len, optlen))
4044 len = min_t(unsigned int, len, TCP_ULP_NAME_MAX);
4045 if (!icsk->icsk_ulp_ops) {
4046 if (put_user(0, optlen))
4050 if (put_user(len, optlen))
4052 if (copy_to_user(optval, icsk->icsk_ulp_ops->name, len))
4056 case TCP_FASTOPEN_KEY: {
4057 u64 key[TCP_FASTOPEN_KEY_BUF_LENGTH / sizeof(u64)];
4058 unsigned int key_len;
4060 if (get_user(len, optlen))
4063 key_len = tcp_fastopen_get_cipher(net, icsk, key) *
4064 TCP_FASTOPEN_KEY_LENGTH;
4065 len = min_t(unsigned int, len, key_len);
4066 if (put_user(len, optlen))
4068 if (copy_to_user(optval, key, len))
4072 case TCP_THIN_LINEAR_TIMEOUTS:
4076 case TCP_THIN_DUPACK:
4084 case TCP_REPAIR_QUEUE:
4086 val = tp->repair_queue;
4091 case TCP_REPAIR_WINDOW: {
4092 struct tcp_repair_window opt;
4094 if (get_user(len, optlen))
4097 if (len != sizeof(opt))
4103 opt.snd_wl1 = tp->snd_wl1;
4104 opt.snd_wnd = tp->snd_wnd;
4105 opt.max_window = tp->max_window;
4106 opt.rcv_wnd = tp->rcv_wnd;
4107 opt.rcv_wup = tp->rcv_wup;
4109 if (copy_to_user(optval, &opt, len))
4114 if (tp->repair_queue == TCP_SEND_QUEUE)
4115 val = tp->write_seq;
4116 else if (tp->repair_queue == TCP_RECV_QUEUE)
4122 case TCP_USER_TIMEOUT:
4123 val = icsk->icsk_user_timeout;
4127 val = icsk->icsk_accept_queue.fastopenq.max_qlen;
4130 case TCP_FASTOPEN_CONNECT:
4131 val = tp->fastopen_connect;
4134 case TCP_FASTOPEN_NO_COOKIE:
4135 val = tp->fastopen_no_cookie;
4139 val = tp->tcp_tx_delay;
4143 val = tcp_time_stamp_raw() + tp->tsoffset;
4145 case TCP_NOTSENT_LOWAT:
4146 val = tp->notsent_lowat;
4149 val = tp->recvmsg_inq;
4154 case TCP_SAVED_SYN: {
4155 if (get_user(len, optlen))
4159 if (tp->saved_syn) {
4160 if (len < tcp_saved_syn_len(tp->saved_syn)) {
4161 if (put_user(tcp_saved_syn_len(tp->saved_syn),
4169 len = tcp_saved_syn_len(tp->saved_syn);
4170 if (put_user(len, optlen)) {
4174 if (copy_to_user(optval, tp->saved_syn->data, len)) {
4178 tcp_saved_syn_free(tp);
4183 if (put_user(len, optlen))
4189 case TCP_ZEROCOPY_RECEIVE: {
4190 struct scm_timestamping_internal tss;
4191 struct tcp_zerocopy_receive zc = {};
4194 if (get_user(len, optlen))
4197 len < offsetofend(struct tcp_zerocopy_receive, length))
4199 if (unlikely(len > sizeof(zc))) {
4200 err = check_zeroed_user(optval + sizeof(zc),
4203 return err == 0 ? -EINVAL : err;
4205 if (put_user(len, optlen))
4208 if (copy_from_user(&zc, optval, len))
4212 if (zc.msg_flags & ~(TCP_VALID_ZC_MSG_FLAGS))
4215 err = tcp_zerocopy_receive(sk, &zc, &tss);
4216 err = BPF_CGROUP_RUN_PROG_GETSOCKOPT_KERN(sk, level, optname,
4219 sk_defer_free_flush(sk);
4220 if (len >= offsetofend(struct tcp_zerocopy_receive, msg_flags))
4221 goto zerocopy_rcv_cmsg;
4223 case offsetofend(struct tcp_zerocopy_receive, msg_flags):
4224 goto zerocopy_rcv_cmsg;
4225 case offsetofend(struct tcp_zerocopy_receive, msg_controllen):
4226 case offsetofend(struct tcp_zerocopy_receive, msg_control):
4227 case offsetofend(struct tcp_zerocopy_receive, flags):
4228 case offsetofend(struct tcp_zerocopy_receive, copybuf_len):
4229 case offsetofend(struct tcp_zerocopy_receive, copybuf_address):
4230 case offsetofend(struct tcp_zerocopy_receive, err):
4231 goto zerocopy_rcv_sk_err;
4232 case offsetofend(struct tcp_zerocopy_receive, inq):
4233 goto zerocopy_rcv_inq;
4234 case offsetofend(struct tcp_zerocopy_receive, length):
4236 goto zerocopy_rcv_out;
4239 if (zc.msg_flags & TCP_CMSG_TS)
4240 tcp_zc_finalize_rx_tstamp(sk, &zc, &tss);
4243 zerocopy_rcv_sk_err:
4245 zc.err = sock_error(sk);
4247 zc.inq = tcp_inq_hint(sk);
4249 if (!err && copy_to_user(optval, &zc, len))
4255 return -ENOPROTOOPT;
4258 if (put_user(len, optlen))
4260 if (copy_to_user(optval, &val, len))
4265 bool tcp_bpf_bypass_getsockopt(int level, int optname)
4267 /* TCP do_tcp_getsockopt has optimized getsockopt implementation
4268 * to avoid extra socket lock for TCP_ZEROCOPY_RECEIVE.
4270 if (level == SOL_TCP && optname == TCP_ZEROCOPY_RECEIVE)
4275 EXPORT_SYMBOL(tcp_bpf_bypass_getsockopt);
4277 int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval,
4280 struct inet_connection_sock *icsk = inet_csk(sk);
4282 if (level != SOL_TCP)
4283 return icsk->icsk_af_ops->getsockopt(sk, level, optname,
4285 return do_tcp_getsockopt(sk, level, optname, optval, optlen);
4287 EXPORT_SYMBOL(tcp_getsockopt);
4289 #ifdef CONFIG_TCP_MD5SIG
4290 static DEFINE_PER_CPU(struct tcp_md5sig_pool, tcp_md5sig_pool);
4291 static DEFINE_MUTEX(tcp_md5sig_mutex);
4292 static bool tcp_md5sig_pool_populated = false;
4294 static void __tcp_alloc_md5sig_pool(void)
4296 struct crypto_ahash *hash;
4299 hash = crypto_alloc_ahash("md5", 0, CRYPTO_ALG_ASYNC);
4303 for_each_possible_cpu(cpu) {
4304 void *scratch = per_cpu(tcp_md5sig_pool, cpu).scratch;
4305 struct ahash_request *req;
4308 scratch = kmalloc_node(sizeof(union tcp_md5sum_block) +
4309 sizeof(struct tcphdr),
4314 per_cpu(tcp_md5sig_pool, cpu).scratch = scratch;
4316 if (per_cpu(tcp_md5sig_pool, cpu).md5_req)
4319 req = ahash_request_alloc(hash, GFP_KERNEL);
4323 ahash_request_set_callback(req, 0, NULL, NULL);
4325 per_cpu(tcp_md5sig_pool, cpu).md5_req = req;
4327 /* before setting tcp_md5sig_pool_populated, we must commit all writes
4328 * to memory. See smp_rmb() in tcp_get_md5sig_pool()
4331 tcp_md5sig_pool_populated = true;
4334 bool tcp_alloc_md5sig_pool(void)
4336 if (unlikely(!tcp_md5sig_pool_populated)) {
4337 mutex_lock(&tcp_md5sig_mutex);
4339 if (!tcp_md5sig_pool_populated) {
4340 __tcp_alloc_md5sig_pool();
4341 if (tcp_md5sig_pool_populated)
4342 static_branch_inc(&tcp_md5_needed);
4345 mutex_unlock(&tcp_md5sig_mutex);
4347 return tcp_md5sig_pool_populated;
4349 EXPORT_SYMBOL(tcp_alloc_md5sig_pool);
4353 * tcp_get_md5sig_pool - get md5sig_pool for this user
4355 * We use percpu structure, so if we succeed, we exit with preemption
4356 * and BH disabled, to make sure another thread or softirq handling
4357 * wont try to get same context.
4359 struct tcp_md5sig_pool *tcp_get_md5sig_pool(void)
4363 if (tcp_md5sig_pool_populated) {
4364 /* coupled with smp_wmb() in __tcp_alloc_md5sig_pool() */
4366 return this_cpu_ptr(&tcp_md5sig_pool);
4371 EXPORT_SYMBOL(tcp_get_md5sig_pool);
4373 int tcp_md5_hash_skb_data(struct tcp_md5sig_pool *hp,
4374 const struct sk_buff *skb, unsigned int header_len)
4376 struct scatterlist sg;
4377 const struct tcphdr *tp = tcp_hdr(skb);
4378 struct ahash_request *req = hp->md5_req;
4380 const unsigned int head_data_len = skb_headlen(skb) > header_len ?
4381 skb_headlen(skb) - header_len : 0;
4382 const struct skb_shared_info *shi = skb_shinfo(skb);
4383 struct sk_buff *frag_iter;
4385 sg_init_table(&sg, 1);
4387 sg_set_buf(&sg, ((u8 *) tp) + header_len, head_data_len);
4388 ahash_request_set_crypt(req, &sg, NULL, head_data_len);
4389 if (crypto_ahash_update(req))
4392 for (i = 0; i < shi->nr_frags; ++i) {
4393 const skb_frag_t *f = &shi->frags[i];
4394 unsigned int offset = skb_frag_off(f);
4395 struct page *page = skb_frag_page(f) + (offset >> PAGE_SHIFT);
4397 sg_set_page(&sg, page, skb_frag_size(f),
4398 offset_in_page(offset));
4399 ahash_request_set_crypt(req, &sg, NULL, skb_frag_size(f));
4400 if (crypto_ahash_update(req))
4404 skb_walk_frags(skb, frag_iter)
4405 if (tcp_md5_hash_skb_data(hp, frag_iter, 0))
4410 EXPORT_SYMBOL(tcp_md5_hash_skb_data);
4412 int tcp_md5_hash_key(struct tcp_md5sig_pool *hp, const struct tcp_md5sig_key *key)
4414 u8 keylen = READ_ONCE(key->keylen); /* paired with WRITE_ONCE() in tcp_md5_do_add */
4415 struct scatterlist sg;
4417 sg_init_one(&sg, key->key, keylen);
4418 ahash_request_set_crypt(hp->md5_req, &sg, NULL, keylen);
4420 /* We use data_race() because tcp_md5_do_add() might change key->key under us */
4421 return data_race(crypto_ahash_update(hp->md5_req));
4423 EXPORT_SYMBOL(tcp_md5_hash_key);
4427 void tcp_done(struct sock *sk)
4429 struct request_sock *req;
4431 /* We might be called with a new socket, after
4432 * inet_csk_prepare_forced_close() has been called
4433 * so we can not use lockdep_sock_is_held(sk)
4435 req = rcu_dereference_protected(tcp_sk(sk)->fastopen_rsk, 1);
4437 if (sk->sk_state == TCP_SYN_SENT || sk->sk_state == TCP_SYN_RECV)
4438 TCP_INC_STATS(sock_net(sk), TCP_MIB_ATTEMPTFAILS);
4440 tcp_set_state(sk, TCP_CLOSE);
4441 tcp_clear_xmit_timers(sk);
4443 reqsk_fastopen_remove(sk, req, false);
4445 sk->sk_shutdown = SHUTDOWN_MASK;
4447 if (!sock_flag(sk, SOCK_DEAD))
4448 sk->sk_state_change(sk);
4450 inet_csk_destroy_sock(sk);
4452 EXPORT_SYMBOL_GPL(tcp_done);
4454 int tcp_abort(struct sock *sk, int err)
4456 if (!sk_fullsock(sk)) {
4457 if (sk->sk_state == TCP_NEW_SYN_RECV) {
4458 struct request_sock *req = inet_reqsk(sk);
4461 inet_csk_reqsk_queue_drop(req->rsk_listener, req);
4468 /* Don't race with userspace socket closes such as tcp_close. */
4471 if (sk->sk_state == TCP_LISTEN) {
4472 tcp_set_state(sk, TCP_CLOSE);
4473 inet_csk_listen_stop(sk);
4476 /* Don't race with BH socket closes such as inet_csk_listen_stop. */
4480 if (!sock_flag(sk, SOCK_DEAD)) {
4482 /* This barrier is coupled with smp_rmb() in tcp_poll() */
4484 sk_error_report(sk);
4485 if (tcp_need_reset(sk->sk_state))
4486 tcp_send_active_reset(sk, GFP_ATOMIC);
4492 tcp_write_queue_purge(sk);
4496 EXPORT_SYMBOL_GPL(tcp_abort);
4498 extern struct tcp_congestion_ops tcp_reno;
4500 static __initdata unsigned long thash_entries;
4501 static int __init set_thash_entries(char *str)
4508 ret = kstrtoul(str, 0, &thash_entries);
4514 __setup("thash_entries=", set_thash_entries);
4516 static void __init tcp_init_mem(void)
4518 unsigned long limit = nr_free_buffer_pages() / 16;
4520 limit = max(limit, 128UL);
4521 sysctl_tcp_mem[0] = limit / 4 * 3; /* 4.68 % */
4522 sysctl_tcp_mem[1] = limit; /* 6.25 % */
4523 sysctl_tcp_mem[2] = sysctl_tcp_mem[0] * 2; /* 9.37 % */
4526 void __init tcp_init(void)
4528 int max_rshare, max_wshare, cnt;
4529 unsigned long limit;
4532 BUILD_BUG_ON(TCP_MIN_SND_MSS <= MAX_TCP_OPTION_SPACE);
4533 BUILD_BUG_ON(sizeof(struct tcp_skb_cb) >
4534 sizeof_field(struct sk_buff, cb));
4536 percpu_counter_init(&tcp_sockets_allocated, 0, GFP_KERNEL);
4538 timer_setup(&tcp_orphan_timer, tcp_orphan_update, TIMER_DEFERRABLE);
4539 mod_timer(&tcp_orphan_timer, jiffies + TCP_ORPHAN_TIMER_PERIOD);
4541 inet_hashinfo_init(&tcp_hashinfo);
4542 inet_hashinfo2_init(&tcp_hashinfo, "tcp_listen_portaddr_hash",
4543 thash_entries, 21, /* one slot per 2 MB*/
4545 tcp_hashinfo.bind_bucket_cachep =
4546 kmem_cache_create("tcp_bind_bucket",
4547 sizeof(struct inet_bind_bucket), 0,
4548 SLAB_HWCACHE_ALIGN | SLAB_PANIC |
4552 /* Size and allocate the main established and bind bucket
4555 * The methodology is similar to that of the buffer cache.
4557 tcp_hashinfo.ehash =
4558 alloc_large_system_hash("TCP established",
4559 sizeof(struct inet_ehash_bucket),
4561 17, /* one slot per 128 KB of memory */
4564 &tcp_hashinfo.ehash_mask,
4566 thash_entries ? 0 : 512 * 1024);
4567 for (i = 0; i <= tcp_hashinfo.ehash_mask; i++)
4568 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].chain, i);
4570 if (inet_ehash_locks_alloc(&tcp_hashinfo))
4571 panic("TCP: failed to alloc ehash_locks");
4572 tcp_hashinfo.bhash =
4573 alloc_large_system_hash("TCP bind",
4574 sizeof(struct inet_bind_hashbucket),
4575 tcp_hashinfo.ehash_mask + 1,
4576 17, /* one slot per 128 KB of memory */
4578 &tcp_hashinfo.bhash_size,
4582 tcp_hashinfo.bhash_size = 1U << tcp_hashinfo.bhash_size;
4583 for (i = 0; i < tcp_hashinfo.bhash_size; i++) {
4584 spin_lock_init(&tcp_hashinfo.bhash[i].lock);
4585 INIT_HLIST_HEAD(&tcp_hashinfo.bhash[i].chain);
4589 cnt = tcp_hashinfo.ehash_mask + 1;
4590 sysctl_tcp_max_orphans = cnt / 2;
4593 /* Set per-socket limits to no more than 1/128 the pressure threshold */
4594 limit = nr_free_buffer_pages() << (PAGE_SHIFT - 7);
4595 max_wshare = min(4UL*1024*1024, limit);
4596 max_rshare = min(6UL*1024*1024, limit);
4598 init_net.ipv4.sysctl_tcp_wmem[0] = SK_MEM_QUANTUM;
4599 init_net.ipv4.sysctl_tcp_wmem[1] = 16*1024;
4600 init_net.ipv4.sysctl_tcp_wmem[2] = max(64*1024, max_wshare);
4602 init_net.ipv4.sysctl_tcp_rmem[0] = SK_MEM_QUANTUM;
4603 init_net.ipv4.sysctl_tcp_rmem[1] = 131072;
4604 init_net.ipv4.sysctl_tcp_rmem[2] = max(131072, max_rshare);
4606 pr_info("Hash tables configured (established %u bind %u)\n",
4607 tcp_hashinfo.ehash_mask + 1, tcp_hashinfo.bhash_size);
4611 BUG_ON(tcp_register_congestion_control(&tcp_reno) != 0);