1 // SPDX-License-Identifier: GPL-2.0-or-later
3 * INET An implementation of the TCP/IP protocol suite for the LINUX
4 * operating system. INET is implemented using the BSD Socket
5 * interface as the means of communication with the user level.
7 * Implementation of the Transmission Control Protocol(TCP).
10 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
11 * Mark Evans, <evansmp@uhura.aston.ac.uk>
12 * Corey Minyard <wf-rch!minyard@relay.EU.net>
13 * Florian La Roche, <flla@stud.uni-sb.de>
14 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
15 * Linus Torvalds, <torvalds@cs.helsinki.fi>
16 * Alan Cox, <gw4pts@gw4pts.ampr.org>
17 * Matthew Dillon, <dillon@apollo.west.oic.com>
18 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
19 * Jorge Cwik, <jorge@laser.satlink.net>
22 * Alan Cox : Numerous verify_area() calls
23 * Alan Cox : Set the ACK bit on a reset
24 * Alan Cox : Stopped it crashing if it closed while
25 * sk->inuse=1 and was trying to connect
27 * Alan Cox : All icmp error handling was broken
28 * pointers passed where wrong and the
29 * socket was looked up backwards. Nobody
30 * tested any icmp error code obviously.
31 * Alan Cox : tcp_err() now handled properly. It
32 * wakes people on errors. poll
33 * behaves and the icmp error race
34 * has gone by moving it into sock.c
35 * Alan Cox : tcp_send_reset() fixed to work for
36 * everything not just packets for
38 * Alan Cox : tcp option processing.
39 * Alan Cox : Reset tweaked (still not 100%) [Had
41 * Herp Rosmanith : More reset fixes
42 * Alan Cox : No longer acks invalid rst frames.
43 * Acking any kind of RST is right out.
44 * Alan Cox : Sets an ignore me flag on an rst
45 * receive otherwise odd bits of prattle
47 * Alan Cox : Fixed another acking RST frame bug.
48 * Should stop LAN workplace lockups.
49 * Alan Cox : Some tidyups using the new skb list
51 * Alan Cox : sk->keepopen now seems to work
52 * Alan Cox : Pulls options out correctly on accepts
53 * Alan Cox : Fixed assorted sk->rqueue->next errors
54 * Alan Cox : PSH doesn't end a TCP read. Switched a
56 * Alan Cox : Tidied tcp_data to avoid a potential
58 * Alan Cox : Added some better commenting, as the
59 * tcp is hard to follow
60 * Alan Cox : Removed incorrect check for 20 * psh
61 * Michael O'Reilly : ack < copied bug fix.
62 * Johannes Stille : Misc tcp fixes (not all in yet).
63 * Alan Cox : FIN with no memory -> CRASH
64 * Alan Cox : Added socket option proto entries.
65 * Also added awareness of them to accept.
66 * Alan Cox : Added TCP options (SOL_TCP)
67 * Alan Cox : Switched wakeup calls to callbacks,
68 * so the kernel can layer network
70 * Alan Cox : Use ip_tos/ip_ttl settings.
71 * Alan Cox : Handle FIN (more) properly (we hope).
72 * Alan Cox : RST frames sent on unsynchronised
74 * Alan Cox : Put in missing check for SYN bit.
75 * Alan Cox : Added tcp_select_window() aka NET2E
76 * window non shrink trick.
77 * Alan Cox : Added a couple of small NET2E timer
79 * Charles Hedrick : TCP fixes
80 * Toomas Tamm : TCP window fixes
81 * Alan Cox : Small URG fix to rlogin ^C ack fight
82 * Charles Hedrick : Rewrote most of it to actually work
83 * Linus : Rewrote tcp_read() and URG handling
85 * Gerhard Koerting: Fixed some missing timer handling
86 * Matthew Dillon : Reworked TCP machine states as per RFC
87 * Gerhard Koerting: PC/TCP workarounds
88 * Adam Caldwell : Assorted timer/timing errors
89 * Matthew Dillon : Fixed another RST bug
90 * Alan Cox : Move to kernel side addressing changes.
91 * Alan Cox : Beginning work on TCP fastpathing
93 * Arnt Gulbrandsen: Turbocharged tcp_check() routine.
94 * Alan Cox : TCP fast path debugging
95 * Alan Cox : Window clamping
96 * Michael Riepe : Bug in tcp_check()
97 * Matt Dillon : More TCP improvements and RST bug fixes
98 * Matt Dillon : Yet more small nasties remove from the
99 * TCP code (Be very nice to this man if
100 * tcp finally works 100%) 8)
101 * Alan Cox : BSD accept semantics.
102 * Alan Cox : Reset on closedown bug.
103 * Peter De Schrijver : ENOTCONN check missing in tcp_sendto().
104 * Michael Pall : Handle poll() after URG properly in
106 * Michael Pall : Undo the last fix in tcp_read_urg()
107 * (multi URG PUSH broke rlogin).
108 * Michael Pall : Fix the multi URG PUSH problem in
109 * tcp_readable(), poll() after URG
111 * Michael Pall : recv(...,MSG_OOB) never blocks in the
113 * Alan Cox : Changed the semantics of sk->socket to
114 * fix a race and a signal problem with
115 * accept() and async I/O.
116 * Alan Cox : Relaxed the rules on tcp_sendto().
117 * Yury Shevchuk : Really fixed accept() blocking problem.
118 * Craig I. Hagan : Allow for BSD compatible TIME_WAIT for
119 * clients/servers which listen in on
121 * Alan Cox : Cleaned the above up and shrank it to
122 * a sensible code size.
123 * Alan Cox : Self connect lockup fix.
124 * Alan Cox : No connect to multicast.
125 * Ross Biro : Close unaccepted children on master
127 * Alan Cox : Reset tracing code.
128 * Alan Cox : Spurious resets on shutdown.
129 * Alan Cox : Giant 15 minute/60 second timer error
130 * Alan Cox : Small whoops in polling before an
132 * Alan Cox : Kept the state trace facility since
133 * it's handy for debugging.
134 * Alan Cox : More reset handler fixes.
135 * Alan Cox : Started rewriting the code based on
136 * the RFC's for other useful protocol
137 * references see: Comer, KA9Q NOS, and
138 * for a reference on the difference
139 * between specifications and how BSD
140 * works see the 4.4lite source.
141 * A.N.Kuznetsov : Don't time wait on completion of tidy
143 * Linus Torvalds : Fin/Shutdown & copied_seq changes.
144 * Linus Torvalds : Fixed BSD port reuse to work first syn
145 * Alan Cox : Reimplemented timers as per the RFC
146 * and using multiple timers for sanity.
147 * Alan Cox : Small bug fixes, and a lot of new
149 * Alan Cox : Fixed dual reader crash by locking
150 * the buffers (much like datagram.c)
151 * Alan Cox : Fixed stuck sockets in probe. A probe
152 * now gets fed up of retrying without
153 * (even a no space) answer.
154 * Alan Cox : Extracted closing code better
155 * Alan Cox : Fixed the closing state machine to
157 * Alan Cox : More 'per spec' fixes.
158 * Jorge Cwik : Even faster checksumming.
159 * Alan Cox : tcp_data() doesn't ack illegal PSH
160 * only frames. At least one pc tcp stack
162 * Alan Cox : Cache last socket.
163 * Alan Cox : Per route irtt.
164 * Matt Day : poll()->select() match BSD precisely on error
165 * Alan Cox : New buffers
166 * Marc Tamsky : Various sk->prot->retransmits and
167 * sk->retransmits misupdating fixed.
168 * Fixed tcp_write_timeout: stuck close,
169 * and TCP syn retries gets used now.
170 * Mark Yarvis : In tcp_read_wakeup(), don't send an
171 * ack if state is TCP_CLOSED.
172 * Alan Cox : Look up device on a retransmit - routes may
173 * change. Doesn't yet cope with MSS shrink right
175 * Marc Tamsky : Closing in closing fixes.
176 * Mike Shaver : RFC1122 verifications.
177 * Alan Cox : rcv_saddr errors.
178 * Alan Cox : Block double connect().
179 * Alan Cox : Small hooks for enSKIP.
180 * Alexey Kuznetsov: Path MTU discovery.
181 * Alan Cox : Support soft errors.
182 * Alan Cox : Fix MTU discovery pathological case
183 * when the remote claims no mtu!
184 * Marc Tamsky : TCP_CLOSE fix.
185 * Colin (G3TNE) : Send a reset on syn ack replies in
186 * window but wrong (fixes NT lpd problems)
187 * Pedro Roque : Better TCP window handling, delayed ack.
188 * Joerg Reuter : No modification of locked buffers in
189 * tcp_do_retransmit()
190 * Eric Schenk : Changed receiver side silly window
191 * avoidance algorithm to BSD style
192 * algorithm. This doubles throughput
193 * against machines running Solaris,
194 * and seems to result in general
196 * Stefan Magdalinski : adjusted tcp_readable() to fix FIONREAD
197 * Willy Konynenberg : Transparent proxying support.
198 * Mike McLagan : Routing by source
199 * Keith Owens : Do proper merging with partial SKB's in
200 * tcp_do_sendmsg to avoid burstiness.
201 * Eric Schenk : Fix fast close down bug with
202 * shutdown() followed by close().
203 * Andi Kleen : Make poll agree with SIGIO
204 * Salvatore Sanfilippo : Support SO_LINGER with linger == 1 and
205 * lingertime == 0 (RFC 793 ABORT Call)
206 * Hirokazu Takahashi : Use copy_from_user() instead of
207 * csum_and_copy_from_user() if possible.
209 * Description of States:
211 * TCP_SYN_SENT sent a connection request, waiting for ack
213 * TCP_SYN_RECV received a connection request, sent ack,
214 * waiting for final ack in three-way handshake.
216 * TCP_ESTABLISHED connection established
218 * TCP_FIN_WAIT1 our side has shutdown, waiting to complete
219 * transmission of remaining buffered data
221 * TCP_FIN_WAIT2 all buffered data sent, waiting for remote
224 * TCP_CLOSING both sides have shutdown but we still have
225 * data we have to finish sending
227 * TCP_TIME_WAIT timeout to catch resent junk before entering
228 * closed, can only be entered from FIN_WAIT2
229 * or CLOSING. Required because the other end
230 * may not have gotten our last ACK causing it
231 * to retransmit the data packet (which we ignore)
233 * TCP_CLOSE_WAIT remote side has shutdown and is waiting for
234 * us to finish writing our data and to shutdown
235 * (we have to close() to move on to LAST_ACK)
237 * TCP_LAST_ACK out side has shutdown after remote has
238 * shutdown. There may still be data in our
239 * buffer that we have to finish sending
241 * TCP_CLOSE socket is finished
244 #define pr_fmt(fmt) "TCP: " fmt
246 #include <crypto/hash.h>
247 #include <linux/kernel.h>
248 #include <linux/module.h>
249 #include <linux/types.h>
250 #include <linux/fcntl.h>
251 #include <linux/poll.h>
252 #include <linux/inet_diag.h>
253 #include <linux/init.h>
254 #include <linux/fs.h>
255 #include <linux/skbuff.h>
256 #include <linux/scatterlist.h>
257 #include <linux/splice.h>
258 #include <linux/net.h>
259 #include <linux/socket.h>
260 #include <linux/random.h>
261 #include <linux/memblock.h>
262 #include <linux/highmem.h>
263 #include <linux/swap.h>
264 #include <linux/cache.h>
265 #include <linux/err.h>
266 #include <linux/time.h>
267 #include <linux/slab.h>
268 #include <linux/errqueue.h>
269 #include <linux/static_key.h>
270 #include <linux/btf.h>
272 #include <net/icmp.h>
273 #include <net/inet_common.h>
275 #include <net/mptcp.h>
276 #include <net/xfrm.h>
278 #include <net/sock.h>
280 #include <linux/uaccess.h>
281 #include <asm/ioctls.h>
282 #include <net/busy_poll.h>
284 /* Track pending CMSGs. */
290 struct percpu_counter tcp_orphan_count;
291 EXPORT_SYMBOL_GPL(tcp_orphan_count);
293 long sysctl_tcp_mem[3] __read_mostly;
294 EXPORT_SYMBOL(sysctl_tcp_mem);
296 atomic_long_t tcp_memory_allocated; /* Current allocated memory. */
297 EXPORT_SYMBOL(tcp_memory_allocated);
299 #if IS_ENABLED(CONFIG_SMC)
300 DEFINE_STATIC_KEY_FALSE(tcp_have_smc);
301 EXPORT_SYMBOL(tcp_have_smc);
305 * Current number of TCP sockets.
307 struct percpu_counter tcp_sockets_allocated;
308 EXPORT_SYMBOL(tcp_sockets_allocated);
313 struct tcp_splice_state {
314 struct pipe_inode_info *pipe;
320 * Pressure flag: try to collapse.
321 * Technical note: it is used by multiple contexts non atomically.
322 * All the __sk_mem_schedule() is of this nature: accounting
323 * is strict, actions are advisory and have some latency.
325 unsigned long tcp_memory_pressure __read_mostly;
326 EXPORT_SYMBOL_GPL(tcp_memory_pressure);
328 DEFINE_STATIC_KEY_FALSE(tcp_rx_skb_cache_key);
329 EXPORT_SYMBOL(tcp_rx_skb_cache_key);
331 DEFINE_STATIC_KEY_FALSE(tcp_tx_skb_cache_key);
333 void tcp_enter_memory_pressure(struct sock *sk)
337 if (READ_ONCE(tcp_memory_pressure))
343 if (!cmpxchg(&tcp_memory_pressure, 0, val))
344 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMEMORYPRESSURES);
346 EXPORT_SYMBOL_GPL(tcp_enter_memory_pressure);
348 void tcp_leave_memory_pressure(struct sock *sk)
352 if (!READ_ONCE(tcp_memory_pressure))
354 val = xchg(&tcp_memory_pressure, 0);
356 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPMEMORYPRESSURESCHRONO,
357 jiffies_to_msecs(jiffies - val));
359 EXPORT_SYMBOL_GPL(tcp_leave_memory_pressure);
361 /* Convert seconds to retransmits based on initial and max timeout */
362 static u8 secs_to_retrans(int seconds, int timeout, int rto_max)
367 int period = timeout;
370 while (seconds > period && res < 255) {
373 if (timeout > rto_max)
381 /* Convert retransmits to seconds based on initial and max timeout */
382 static int retrans_to_secs(u8 retrans, int timeout, int rto_max)
390 if (timeout > rto_max)
398 static u64 tcp_compute_delivery_rate(const struct tcp_sock *tp)
400 u32 rate = READ_ONCE(tp->rate_delivered);
401 u32 intv = READ_ONCE(tp->rate_interval_us);
405 rate64 = (u64)rate * tp->mss_cache * USEC_PER_SEC;
406 do_div(rate64, intv);
411 /* Address-family independent initialization for a tcp_sock.
413 * NOTE: A lot of things set to zero explicitly by call to
414 * sk_alloc() so need not be done here.
416 void tcp_init_sock(struct sock *sk)
418 struct inet_connection_sock *icsk = inet_csk(sk);
419 struct tcp_sock *tp = tcp_sk(sk);
421 tp->out_of_order_queue = RB_ROOT;
422 sk->tcp_rtx_queue = RB_ROOT;
423 tcp_init_xmit_timers(sk);
424 INIT_LIST_HEAD(&tp->tsq_node);
425 INIT_LIST_HEAD(&tp->tsorted_sent_queue);
427 icsk->icsk_rto = TCP_TIMEOUT_INIT;
428 icsk->icsk_rto_min = TCP_RTO_MIN;
429 icsk->icsk_delack_max = TCP_DELACK_MAX;
430 tp->mdev_us = jiffies_to_usecs(TCP_TIMEOUT_INIT);
431 minmax_reset(&tp->rtt_min, tcp_jiffies32, ~0U);
433 /* So many TCP implementations out there (incorrectly) count the
434 * initial SYN frame in their delayed-ACK and congestion control
435 * algorithms that we must have the following bandaid to talk
436 * efficiently to them. -DaveM
438 tp->snd_cwnd = TCP_INIT_CWND;
440 /* There's a bubble in the pipe until at least the first ACK. */
441 tp->app_limited = ~0U;
443 /* See draft-stevens-tcpca-spec-01 for discussion of the
444 * initialization of these values.
446 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
447 tp->snd_cwnd_clamp = ~0;
448 tp->mss_cache = TCP_MSS_DEFAULT;
450 tp->reordering = sock_net(sk)->ipv4.sysctl_tcp_reordering;
451 tcp_assign_congestion_control(sk);
454 tp->rack.reo_wnd_steps = 1;
456 sk->sk_write_space = sk_stream_write_space;
457 sock_set_flag(sk, SOCK_USE_WRITE_QUEUE);
459 icsk->icsk_sync_mss = tcp_sync_mss;
461 WRITE_ONCE(sk->sk_sndbuf, sock_net(sk)->ipv4.sysctl_tcp_wmem[1]);
462 WRITE_ONCE(sk->sk_rcvbuf, sock_net(sk)->ipv4.sysctl_tcp_rmem[1]);
464 sk_sockets_allocated_inc(sk);
465 sk->sk_route_forced_caps = NETIF_F_GSO;
467 EXPORT_SYMBOL(tcp_init_sock);
469 static void tcp_tx_timestamp(struct sock *sk, u16 tsflags)
471 struct sk_buff *skb = tcp_write_queue_tail(sk);
473 if (tsflags && skb) {
474 struct skb_shared_info *shinfo = skb_shinfo(skb);
475 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
477 sock_tx_timestamp(sk, tsflags, &shinfo->tx_flags);
478 if (tsflags & SOF_TIMESTAMPING_TX_ACK)
479 tcb->txstamp_ack = 1;
480 if (tsflags & SOF_TIMESTAMPING_TX_RECORD_MASK)
481 shinfo->tskey = TCP_SKB_CB(skb)->seq + skb->len - 1;
485 static bool tcp_stream_is_readable(struct sock *sk, int target)
487 if (tcp_epollin_ready(sk, target))
490 if (sk->sk_prot->stream_memory_read)
491 return sk->sk_prot->stream_memory_read(sk);
496 * Wait for a TCP event.
498 * Note that we don't need to lock the socket, as the upper poll layers
499 * take care of normal races (between the test and the event) and we don't
500 * go look at any of the socket buffers directly.
502 __poll_t tcp_poll(struct file *file, struct socket *sock, poll_table *wait)
505 struct sock *sk = sock->sk;
506 const struct tcp_sock *tp = tcp_sk(sk);
509 sock_poll_wait(file, sock, wait);
511 state = inet_sk_state_load(sk);
512 if (state == TCP_LISTEN)
513 return inet_csk_listen_poll(sk);
515 /* Socket is not locked. We are protected from async events
516 * by poll logic and correct handling of state changes
517 * made by other threads is impossible in any case.
523 * EPOLLHUP is certainly not done right. But poll() doesn't
524 * have a notion of HUP in just one direction, and for a
525 * socket the read side is more interesting.
527 * Some poll() documentation says that EPOLLHUP is incompatible
528 * with the EPOLLOUT/POLLWR flags, so somebody should check this
529 * all. But careful, it tends to be safer to return too many
530 * bits than too few, and you can easily break real applications
531 * if you don't tell them that something has hung up!
535 * Check number 1. EPOLLHUP is _UNMASKABLE_ event (see UNIX98 and
536 * our fs/select.c). It means that after we received EOF,
537 * poll always returns immediately, making impossible poll() on write()
538 * in state CLOSE_WAIT. One solution is evident --- to set EPOLLHUP
539 * if and only if shutdown has been made in both directions.
540 * Actually, it is interesting to look how Solaris and DUX
541 * solve this dilemma. I would prefer, if EPOLLHUP were maskable,
542 * then we could set it on SND_SHUTDOWN. BTW examples given
543 * in Stevens' books assume exactly this behaviour, it explains
544 * why EPOLLHUP is incompatible with EPOLLOUT. --ANK
546 * NOTE. Check for TCP_CLOSE is added. The goal is to prevent
547 * blocking on fresh not-connected or disconnected socket. --ANK
549 if (sk->sk_shutdown == SHUTDOWN_MASK || state == TCP_CLOSE)
551 if (sk->sk_shutdown & RCV_SHUTDOWN)
552 mask |= EPOLLIN | EPOLLRDNORM | EPOLLRDHUP;
554 /* Connected or passive Fast Open socket? */
555 if (state != TCP_SYN_SENT &&
556 (state != TCP_SYN_RECV || rcu_access_pointer(tp->fastopen_rsk))) {
557 int target = sock_rcvlowat(sk, 0, INT_MAX);
559 if (READ_ONCE(tp->urg_seq) == READ_ONCE(tp->copied_seq) &&
560 !sock_flag(sk, SOCK_URGINLINE) &&
564 if (tcp_stream_is_readable(sk, target))
565 mask |= EPOLLIN | EPOLLRDNORM;
567 if (!(sk->sk_shutdown & SEND_SHUTDOWN)) {
568 if (__sk_stream_is_writeable(sk, 1)) {
569 mask |= EPOLLOUT | EPOLLWRNORM;
570 } else { /* send SIGIO later */
571 sk_set_bit(SOCKWQ_ASYNC_NOSPACE, sk);
572 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
574 /* Race breaker. If space is freed after
575 * wspace test but before the flags are set,
576 * IO signal will be lost. Memory barrier
577 * pairs with the input side.
579 smp_mb__after_atomic();
580 if (__sk_stream_is_writeable(sk, 1))
581 mask |= EPOLLOUT | EPOLLWRNORM;
584 mask |= EPOLLOUT | EPOLLWRNORM;
586 if (tp->urg_data & TCP_URG_VALID)
588 } else if (state == TCP_SYN_SENT && inet_sk(sk)->defer_connect) {
589 /* Active TCP fastopen socket with defer_connect
590 * Return EPOLLOUT so application can call write()
591 * in order for kernel to generate SYN+data
593 mask |= EPOLLOUT | EPOLLWRNORM;
595 /* This barrier is coupled with smp_wmb() in tcp_reset() */
597 if (sk->sk_err || !skb_queue_empty_lockless(&sk->sk_error_queue))
602 EXPORT_SYMBOL(tcp_poll);
604 int tcp_ioctl(struct sock *sk, int cmd, unsigned long arg)
606 struct tcp_sock *tp = tcp_sk(sk);
612 if (sk->sk_state == TCP_LISTEN)
615 slow = lock_sock_fast(sk);
617 unlock_sock_fast(sk, slow);
620 answ = tp->urg_data &&
621 READ_ONCE(tp->urg_seq) == READ_ONCE(tp->copied_seq);
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) - tp->snd_una;
633 if (sk->sk_state == TCP_LISTEN)
636 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
639 answ = READ_ONCE(tp->write_seq) -
640 READ_ONCE(tp->snd_nxt);
646 return put_user(answ, (int __user *)arg);
648 EXPORT_SYMBOL(tcp_ioctl);
650 static inline void tcp_mark_push(struct tcp_sock *tp, struct sk_buff *skb)
652 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
653 tp->pushed_seq = tp->write_seq;
656 static inline bool forced_push(const struct tcp_sock *tp)
658 return after(tp->write_seq, tp->pushed_seq + (tp->max_window >> 1));
661 static void skb_entail(struct sock *sk, struct sk_buff *skb)
663 struct tcp_sock *tp = tcp_sk(sk);
664 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
667 tcb->seq = tcb->end_seq = tp->write_seq;
668 tcb->tcp_flags = TCPHDR_ACK;
670 __skb_header_release(skb);
671 tcp_add_write_queue_tail(sk, skb);
672 sk_wmem_queued_add(sk, skb->truesize);
673 sk_mem_charge(sk, skb->truesize);
674 if (tp->nonagle & TCP_NAGLE_PUSH)
675 tp->nonagle &= ~TCP_NAGLE_PUSH;
677 tcp_slow_start_after_idle_check(sk);
680 static inline void tcp_mark_urg(struct tcp_sock *tp, int flags)
683 tp->snd_up = tp->write_seq;
686 /* If a not yet filled skb is pushed, do not send it if
687 * we have data packets in Qdisc or NIC queues :
688 * Because TX completion will happen shortly, it gives a chance
689 * to coalesce future sendmsg() payload into this skb, without
690 * need for a timer, and with no latency trade off.
691 * As packets containing data payload have a bigger truesize
692 * than pure acks (dataless) packets, the last checks prevent
693 * autocorking if we only have an ACK in Qdisc/NIC queues,
694 * or if TX completion was delayed after we processed ACK packet.
696 static bool tcp_should_autocork(struct sock *sk, struct sk_buff *skb,
699 return skb->len < size_goal &&
700 sock_net(sk)->ipv4.sysctl_tcp_autocorking &&
701 !tcp_rtx_queue_empty(sk) &&
702 refcount_read(&sk->sk_wmem_alloc) > skb->truesize;
705 void tcp_push(struct sock *sk, int flags, int mss_now,
706 int nonagle, int size_goal)
708 struct tcp_sock *tp = tcp_sk(sk);
711 skb = tcp_write_queue_tail(sk);
714 if (!(flags & MSG_MORE) || forced_push(tp))
715 tcp_mark_push(tp, skb);
717 tcp_mark_urg(tp, flags);
719 if (tcp_should_autocork(sk, skb, size_goal)) {
721 /* avoid atomic op if TSQ_THROTTLED bit is already set */
722 if (!test_bit(TSQ_THROTTLED, &sk->sk_tsq_flags)) {
723 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPAUTOCORKING);
724 set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags);
726 /* It is possible TX completion already happened
727 * before we set TSQ_THROTTLED.
729 if (refcount_read(&sk->sk_wmem_alloc) > skb->truesize)
733 if (flags & MSG_MORE)
734 nonagle = TCP_NAGLE_CORK;
736 __tcp_push_pending_frames(sk, mss_now, nonagle);
739 static int tcp_splice_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb,
740 unsigned int offset, size_t len)
742 struct tcp_splice_state *tss = rd_desc->arg.data;
745 ret = skb_splice_bits(skb, skb->sk, offset, tss->pipe,
746 min(rd_desc->count, len), tss->flags);
748 rd_desc->count -= ret;
752 static int __tcp_splice_read(struct sock *sk, struct tcp_splice_state *tss)
754 /* Store TCP splice context information in read_descriptor_t. */
755 read_descriptor_t rd_desc = {
760 return tcp_read_sock(sk, &rd_desc, tcp_splice_data_recv);
764 * tcp_splice_read - splice data from TCP socket to a pipe
765 * @sock: socket to splice from
766 * @ppos: position (not valid)
767 * @pipe: pipe to splice to
768 * @len: number of bytes to splice
769 * @flags: splice modifier flags
772 * Will read pages from given socket and fill them into a pipe.
775 ssize_t tcp_splice_read(struct socket *sock, loff_t *ppos,
776 struct pipe_inode_info *pipe, size_t len,
779 struct sock *sk = sock->sk;
780 struct tcp_splice_state tss = {
789 sock_rps_record_flow(sk);
791 * We can't seek on a socket input
800 timeo = sock_rcvtimeo(sk, sock->file->f_flags & O_NONBLOCK);
802 ret = __tcp_splice_read(sk, &tss);
808 if (sock_flag(sk, SOCK_DONE))
811 ret = sock_error(sk);
814 if (sk->sk_shutdown & RCV_SHUTDOWN)
816 if (sk->sk_state == TCP_CLOSE) {
818 * This occurs when user tries to read
819 * from never connected socket.
828 /* if __tcp_splice_read() got nothing while we have
829 * an skb in receive queue, we do not want to loop.
830 * This might happen with URG data.
832 if (!skb_queue_empty(&sk->sk_receive_queue))
834 sk_wait_data(sk, &timeo, NULL);
835 if (signal_pending(current)) {
836 ret = sock_intr_errno(timeo);
849 if (sk->sk_err || sk->sk_state == TCP_CLOSE ||
850 (sk->sk_shutdown & RCV_SHUTDOWN) ||
851 signal_pending(current))
862 EXPORT_SYMBOL(tcp_splice_read);
864 struct sk_buff *sk_stream_alloc_skb(struct sock *sk, int size, gfp_t gfp,
870 skb = sk->sk_tx_skb_cache;
872 skb->truesize = SKB_TRUESIZE(skb_end_offset(skb));
873 sk->sk_tx_skb_cache = NULL;
875 INIT_LIST_HEAD(&skb->tcp_tsorted_anchor);
876 skb_shinfo(skb)->tx_flags = 0;
877 memset(TCP_SKB_CB(skb), 0, sizeof(struct tcp_skb_cb));
881 /* The TCP header must be at least 32-bit aligned. */
882 size = ALIGN(size, 4);
884 if (unlikely(tcp_under_memory_pressure(sk)))
885 sk_mem_reclaim_partial(sk);
887 skb = alloc_skb_fclone(size + sk->sk_prot->max_header, gfp);
891 if (force_schedule) {
892 mem_scheduled = true;
893 sk_forced_mem_schedule(sk, skb->truesize);
895 mem_scheduled = sk_wmem_schedule(sk, skb->truesize);
897 if (likely(mem_scheduled)) {
898 skb_reserve(skb, sk->sk_prot->max_header);
900 * Make sure that we have exactly size bytes
901 * available to the caller, no more, no less.
903 skb->reserved_tailroom = skb->end - skb->tail - size;
904 INIT_LIST_HEAD(&skb->tcp_tsorted_anchor);
909 sk->sk_prot->enter_memory_pressure(sk);
910 sk_stream_moderate_sndbuf(sk);
915 static unsigned int tcp_xmit_size_goal(struct sock *sk, u32 mss_now,
918 struct tcp_sock *tp = tcp_sk(sk);
919 u32 new_size_goal, size_goal;
924 /* Note : tcp_tso_autosize() will eventually split this later */
925 new_size_goal = sk->sk_gso_max_size - 1 - MAX_TCP_HEADER;
926 new_size_goal = tcp_bound_to_half_wnd(tp, new_size_goal);
928 /* We try hard to avoid divides here */
929 size_goal = tp->gso_segs * mss_now;
930 if (unlikely(new_size_goal < size_goal ||
931 new_size_goal >= size_goal + mss_now)) {
932 tp->gso_segs = min_t(u16, new_size_goal / mss_now,
933 sk->sk_gso_max_segs);
934 size_goal = tp->gso_segs * mss_now;
937 return max(size_goal, mss_now);
940 int tcp_send_mss(struct sock *sk, int *size_goal, int flags)
944 mss_now = tcp_current_mss(sk);
945 *size_goal = tcp_xmit_size_goal(sk, mss_now, !(flags & MSG_OOB));
950 /* In some cases, both sendpage() and sendmsg() could have added
951 * an skb to the write queue, but failed adding payload on it.
952 * We need to remove it to consume less memory, but more
953 * importantly be able to generate EPOLLOUT for Edge Trigger epoll()
956 void tcp_remove_empty_skb(struct sock *sk, struct sk_buff *skb)
958 if (skb && !skb->len) {
959 tcp_unlink_write_queue(skb, sk);
960 if (tcp_write_queue_empty(sk))
961 tcp_chrono_stop(sk, TCP_CHRONO_BUSY);
962 sk_wmem_free_skb(sk, skb);
966 struct sk_buff *tcp_build_frag(struct sock *sk, int size_goal, int flags,
967 struct page *page, int offset, size_t *size)
969 struct sk_buff *skb = tcp_write_queue_tail(sk);
970 struct tcp_sock *tp = tcp_sk(sk);
974 if (!skb || (copy = size_goal - skb->len) <= 0 ||
975 !tcp_skb_can_collapse_to(skb)) {
977 if (!sk_stream_memory_free(sk))
980 skb = sk_stream_alloc_skb(sk, 0, sk->sk_allocation,
981 tcp_rtx_and_write_queues_empty(sk));
985 #ifdef CONFIG_TLS_DEVICE
986 skb->decrypted = !!(flags & MSG_SENDPAGE_DECRYPTED);
995 i = skb_shinfo(skb)->nr_frags;
996 can_coalesce = skb_can_coalesce(skb, i, page, offset);
997 if (!can_coalesce && i >= sysctl_max_skb_frags) {
998 tcp_mark_push(tp, skb);
1001 if (!sk_wmem_schedule(sk, copy))
1005 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1008 skb_fill_page_desc(skb, i, page, offset, copy);
1011 if (!(flags & MSG_NO_SHARED_FRAGS))
1012 skb_shinfo(skb)->flags |= SKBFL_SHARED_FRAG;
1015 skb->data_len += copy;
1016 skb->truesize += copy;
1017 sk_wmem_queued_add(sk, copy);
1018 sk_mem_charge(sk, copy);
1019 skb->ip_summed = CHECKSUM_PARTIAL;
1020 WRITE_ONCE(tp->write_seq, tp->write_seq + copy);
1021 TCP_SKB_CB(skb)->end_seq += copy;
1022 tcp_skb_pcount_set(skb, 0);
1028 ssize_t do_tcp_sendpages(struct sock *sk, struct page *page, int offset,
1029 size_t size, int flags)
1031 struct tcp_sock *tp = tcp_sk(sk);
1032 int mss_now, size_goal;
1035 long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
1037 if (IS_ENABLED(CONFIG_DEBUG_VM) &&
1038 WARN_ONCE(!sendpage_ok(page),
1039 "page must not be a Slab one and have page_count > 0"))
1042 /* Wait for a connection to finish. One exception is TCP Fast Open
1043 * (passive side) where data is allowed to be sent before a connection
1044 * is fully established.
1046 if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) &&
1047 !tcp_passive_fastopen(sk)) {
1048 err = sk_stream_wait_connect(sk, &timeo);
1053 sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
1055 mss_now = tcp_send_mss(sk, &size_goal, flags);
1059 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
1063 struct sk_buff *skb;
1066 skb = tcp_build_frag(sk, size_goal, flags, page, offset, ©);
1068 goto wait_for_space;
1071 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
1079 if (skb->len < size_goal || (flags & MSG_OOB))
1082 if (forced_push(tp)) {
1083 tcp_mark_push(tp, skb);
1084 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
1085 } else if (skb == tcp_send_head(sk))
1086 tcp_push_one(sk, mss_now);
1090 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1091 tcp_push(sk, flags & ~MSG_MORE, mss_now,
1092 TCP_NAGLE_PUSH, size_goal);
1094 err = sk_stream_wait_memory(sk, &timeo);
1098 mss_now = tcp_send_mss(sk, &size_goal, flags);
1103 tcp_tx_timestamp(sk, sk->sk_tsflags);
1104 if (!(flags & MSG_SENDPAGE_NOTLAST))
1105 tcp_push(sk, flags, mss_now, tp->nonagle, size_goal);
1110 tcp_remove_empty_skb(sk, tcp_write_queue_tail(sk));
1114 /* make sure we wake any epoll edge trigger waiter */
1115 if (unlikely(tcp_rtx_and_write_queues_empty(sk) && err == -EAGAIN)) {
1116 sk->sk_write_space(sk);
1117 tcp_chrono_stop(sk, TCP_CHRONO_SNDBUF_LIMITED);
1119 return sk_stream_error(sk, flags, err);
1121 EXPORT_SYMBOL_GPL(do_tcp_sendpages);
1123 int tcp_sendpage_locked(struct sock *sk, struct page *page, int offset,
1124 size_t size, int flags)
1126 if (!(sk->sk_route_caps & NETIF_F_SG))
1127 return sock_no_sendpage_locked(sk, page, offset, size, flags);
1129 tcp_rate_check_app_limited(sk); /* is sending application-limited? */
1131 return do_tcp_sendpages(sk, page, offset, size, flags);
1133 EXPORT_SYMBOL_GPL(tcp_sendpage_locked);
1135 int tcp_sendpage(struct sock *sk, struct page *page, int offset,
1136 size_t size, int flags)
1141 ret = tcp_sendpage_locked(sk, page, offset, size, flags);
1146 EXPORT_SYMBOL(tcp_sendpage);
1148 void tcp_free_fastopen_req(struct tcp_sock *tp)
1150 if (tp->fastopen_req) {
1151 kfree(tp->fastopen_req);
1152 tp->fastopen_req = NULL;
1156 static int tcp_sendmsg_fastopen(struct sock *sk, struct msghdr *msg,
1157 int *copied, size_t size,
1158 struct ubuf_info *uarg)
1160 struct tcp_sock *tp = tcp_sk(sk);
1161 struct inet_sock *inet = inet_sk(sk);
1162 struct sockaddr *uaddr = msg->msg_name;
1165 if (!(sock_net(sk)->ipv4.sysctl_tcp_fastopen & TFO_CLIENT_ENABLE) ||
1166 (uaddr && msg->msg_namelen >= sizeof(uaddr->sa_family) &&
1167 uaddr->sa_family == AF_UNSPEC))
1169 if (tp->fastopen_req)
1170 return -EALREADY; /* Another Fast Open is in progress */
1172 tp->fastopen_req = kzalloc(sizeof(struct tcp_fastopen_request),
1174 if (unlikely(!tp->fastopen_req))
1176 tp->fastopen_req->data = msg;
1177 tp->fastopen_req->size = size;
1178 tp->fastopen_req->uarg = uarg;
1180 if (inet->defer_connect) {
1181 err = tcp_connect(sk);
1182 /* Same failure procedure as in tcp_v4/6_connect */
1184 tcp_set_state(sk, TCP_CLOSE);
1185 inet->inet_dport = 0;
1186 sk->sk_route_caps = 0;
1189 flags = (msg->msg_flags & MSG_DONTWAIT) ? O_NONBLOCK : 0;
1190 err = __inet_stream_connect(sk->sk_socket, uaddr,
1191 msg->msg_namelen, flags, 1);
1192 /* fastopen_req could already be freed in __inet_stream_connect
1193 * if the connection times out or gets rst
1195 if (tp->fastopen_req) {
1196 *copied = tp->fastopen_req->copied;
1197 tcp_free_fastopen_req(tp);
1198 inet->defer_connect = 0;
1203 int tcp_sendmsg_locked(struct sock *sk, struct msghdr *msg, size_t size)
1205 struct tcp_sock *tp = tcp_sk(sk);
1206 struct ubuf_info *uarg = NULL;
1207 struct sk_buff *skb;
1208 struct sockcm_cookie sockc;
1209 int flags, err, copied = 0;
1210 int mss_now = 0, size_goal, copied_syn = 0;
1211 int process_backlog = 0;
1215 flags = msg->msg_flags;
1217 if (flags & MSG_ZEROCOPY && size && sock_flag(sk, SOCK_ZEROCOPY)) {
1218 skb = tcp_write_queue_tail(sk);
1219 uarg = msg_zerocopy_realloc(sk, size, skb_zcopy(skb));
1225 zc = sk->sk_route_caps & NETIF_F_SG;
1230 if (unlikely(flags & MSG_FASTOPEN || inet_sk(sk)->defer_connect) &&
1232 err = tcp_sendmsg_fastopen(sk, msg, &copied_syn, size, uarg);
1233 if (err == -EINPROGRESS && copied_syn > 0)
1239 timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
1241 tcp_rate_check_app_limited(sk); /* is sending application-limited? */
1243 /* Wait for a connection to finish. One exception is TCP Fast Open
1244 * (passive side) where data is allowed to be sent before a connection
1245 * is fully established.
1247 if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) &&
1248 !tcp_passive_fastopen(sk)) {
1249 err = sk_stream_wait_connect(sk, &timeo);
1254 if (unlikely(tp->repair)) {
1255 if (tp->repair_queue == TCP_RECV_QUEUE) {
1256 copied = tcp_send_rcvq(sk, msg, size);
1261 if (tp->repair_queue == TCP_NO_QUEUE)
1264 /* 'common' sending to sendq */
1267 sockcm_init(&sockc, sk);
1268 if (msg->msg_controllen) {
1269 err = sock_cmsg_send(sk, msg, &sockc);
1270 if (unlikely(err)) {
1276 /* This should be in poll */
1277 sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
1279 /* Ok commence sending. */
1283 mss_now = tcp_send_mss(sk, &size_goal, flags);
1286 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
1289 while (msg_data_left(msg)) {
1292 skb = tcp_write_queue_tail(sk);
1294 copy = size_goal - skb->len;
1296 if (copy <= 0 || !tcp_skb_can_collapse_to(skb)) {
1300 if (!sk_stream_memory_free(sk))
1301 goto wait_for_space;
1303 if (unlikely(process_backlog >= 16)) {
1304 process_backlog = 0;
1305 if (sk_flush_backlog(sk))
1308 first_skb = tcp_rtx_and_write_queues_empty(sk);
1309 skb = sk_stream_alloc_skb(sk, 0, sk->sk_allocation,
1312 goto wait_for_space;
1315 skb->ip_summed = CHECKSUM_PARTIAL;
1317 skb_entail(sk, skb);
1320 /* All packets are restored as if they have
1321 * already been sent. skb_mstamp_ns isn't set to
1322 * avoid wrong rtt estimation.
1325 TCP_SKB_CB(skb)->sacked |= TCPCB_REPAIRED;
1328 /* Try to append data to the end of skb. */
1329 if (copy > msg_data_left(msg))
1330 copy = msg_data_left(msg);
1332 /* Where to copy to? */
1333 if (skb_availroom(skb) > 0 && !zc) {
1334 /* We have some space in skb head. Superb! */
1335 copy = min_t(int, copy, skb_availroom(skb));
1336 err = skb_add_data_nocache(sk, skb, &msg->msg_iter, copy);
1341 int i = skb_shinfo(skb)->nr_frags;
1342 struct page_frag *pfrag = sk_page_frag(sk);
1344 if (!sk_page_frag_refill(sk, pfrag))
1345 goto wait_for_space;
1347 if (!skb_can_coalesce(skb, i, pfrag->page,
1349 if (i >= sysctl_max_skb_frags) {
1350 tcp_mark_push(tp, skb);
1356 copy = min_t(int, copy, pfrag->size - pfrag->offset);
1358 if (!sk_wmem_schedule(sk, copy))
1359 goto wait_for_space;
1361 err = skb_copy_to_page_nocache(sk, &msg->msg_iter, skb,
1368 /* Update the skb. */
1370 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1372 skb_fill_page_desc(skb, i, pfrag->page,
1373 pfrag->offset, copy);
1374 page_ref_inc(pfrag->page);
1376 pfrag->offset += copy;
1378 err = skb_zerocopy_iter_stream(sk, skb, msg, copy, uarg);
1379 if (err == -EMSGSIZE || err == -EEXIST) {
1380 tcp_mark_push(tp, skb);
1389 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
1391 WRITE_ONCE(tp->write_seq, tp->write_seq + copy);
1392 TCP_SKB_CB(skb)->end_seq += copy;
1393 tcp_skb_pcount_set(skb, 0);
1396 if (!msg_data_left(msg)) {
1397 if (unlikely(flags & MSG_EOR))
1398 TCP_SKB_CB(skb)->eor = 1;
1402 if (skb->len < size_goal || (flags & MSG_OOB) || unlikely(tp->repair))
1405 if (forced_push(tp)) {
1406 tcp_mark_push(tp, skb);
1407 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
1408 } else if (skb == tcp_send_head(sk))
1409 tcp_push_one(sk, mss_now);
1413 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1415 tcp_push(sk, flags & ~MSG_MORE, mss_now,
1416 TCP_NAGLE_PUSH, size_goal);
1418 err = sk_stream_wait_memory(sk, &timeo);
1422 mss_now = tcp_send_mss(sk, &size_goal, flags);
1427 tcp_tx_timestamp(sk, sockc.tsflags);
1428 tcp_push(sk, flags, mss_now, tp->nonagle, size_goal);
1431 net_zcopy_put(uarg);
1432 return copied + copied_syn;
1435 skb = tcp_write_queue_tail(sk);
1437 tcp_remove_empty_skb(sk, skb);
1439 if (copied + copied_syn)
1442 net_zcopy_put_abort(uarg, true);
1443 err = sk_stream_error(sk, flags, err);
1444 /* make sure we wake any epoll edge trigger waiter */
1445 if (unlikely(tcp_rtx_and_write_queues_empty(sk) && err == -EAGAIN)) {
1446 sk->sk_write_space(sk);
1447 tcp_chrono_stop(sk, TCP_CHRONO_SNDBUF_LIMITED);
1451 EXPORT_SYMBOL_GPL(tcp_sendmsg_locked);
1453 int tcp_sendmsg(struct sock *sk, struct msghdr *msg, size_t size)
1458 ret = tcp_sendmsg_locked(sk, msg, size);
1463 EXPORT_SYMBOL(tcp_sendmsg);
1466 * Handle reading urgent data. BSD has very simple semantics for
1467 * this, no blocking and very strange errors 8)
1470 static int tcp_recv_urg(struct sock *sk, struct msghdr *msg, int len, int flags)
1472 struct tcp_sock *tp = tcp_sk(sk);
1474 /* No URG data to read. */
1475 if (sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data ||
1476 tp->urg_data == TCP_URG_READ)
1477 return -EINVAL; /* Yes this is right ! */
1479 if (sk->sk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DONE))
1482 if (tp->urg_data & TCP_URG_VALID) {
1484 char c = tp->urg_data;
1486 if (!(flags & MSG_PEEK))
1487 tp->urg_data = TCP_URG_READ;
1489 /* Read urgent data. */
1490 msg->msg_flags |= MSG_OOB;
1493 if (!(flags & MSG_TRUNC))
1494 err = memcpy_to_msg(msg, &c, 1);
1497 msg->msg_flags |= MSG_TRUNC;
1499 return err ? -EFAULT : len;
1502 if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
1505 /* Fixed the recv(..., MSG_OOB) behaviour. BSD docs and
1506 * the available implementations agree in this case:
1507 * this call should never block, independent of the
1508 * blocking state of the socket.
1509 * Mike <pall@rz.uni-karlsruhe.de>
1514 static int tcp_peek_sndq(struct sock *sk, struct msghdr *msg, int len)
1516 struct sk_buff *skb;
1517 int copied = 0, err = 0;
1519 /* XXX -- need to support SO_PEEK_OFF */
1521 skb_rbtree_walk(skb, &sk->tcp_rtx_queue) {
1522 err = skb_copy_datagram_msg(skb, 0, msg, skb->len);
1528 skb_queue_walk(&sk->sk_write_queue, skb) {
1529 err = skb_copy_datagram_msg(skb, 0, msg, skb->len);
1536 return err ?: copied;
1539 /* Clean up the receive buffer for full frames taken by the user,
1540 * then send an ACK if necessary. COPIED is the number of bytes
1541 * tcp_recvmsg has given to the user so far, it speeds up the
1542 * calculation of whether or not we must ACK for the sake of
1545 void tcp_cleanup_rbuf(struct sock *sk, int copied)
1547 struct tcp_sock *tp = tcp_sk(sk);
1548 bool time_to_ack = false;
1550 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
1552 WARN(skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq),
1553 "cleanup rbuf bug: copied %X seq %X rcvnxt %X\n",
1554 tp->copied_seq, TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt);
1556 if (inet_csk_ack_scheduled(sk)) {
1557 const struct inet_connection_sock *icsk = inet_csk(sk);
1559 if (/* Once-per-two-segments ACK was not sent by tcp_input.c */
1560 tp->rcv_nxt - tp->rcv_wup > icsk->icsk_ack.rcv_mss ||
1562 * If this read emptied read buffer, we send ACK, if
1563 * connection is not bidirectional, user drained
1564 * receive buffer and there was a small segment
1568 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED2) ||
1569 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED) &&
1570 !inet_csk_in_pingpong_mode(sk))) &&
1571 !atomic_read(&sk->sk_rmem_alloc)))
1575 /* We send an ACK if we can now advertise a non-zero window
1576 * which has been raised "significantly".
1578 * Even if window raised up to infinity, do not send window open ACK
1579 * in states, where we will not receive more. It is useless.
1581 if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) {
1582 __u32 rcv_window_now = tcp_receive_window(tp);
1584 /* Optimize, __tcp_select_window() is not cheap. */
1585 if (2*rcv_window_now <= tp->window_clamp) {
1586 __u32 new_window = __tcp_select_window(sk);
1588 /* Send ACK now, if this read freed lots of space
1589 * in our buffer. Certainly, new_window is new window.
1590 * We can advertise it now, if it is not less than current one.
1591 * "Lots" means "at least twice" here.
1593 if (new_window && new_window >= 2 * rcv_window_now)
1601 static struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off)
1603 struct sk_buff *skb;
1606 while ((skb = skb_peek(&sk->sk_receive_queue)) != NULL) {
1607 offset = seq - TCP_SKB_CB(skb)->seq;
1608 if (unlikely(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
1609 pr_err_once("%s: found a SYN, please report !\n", __func__);
1612 if (offset < skb->len || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)) {
1616 /* This looks weird, but this can happen if TCP collapsing
1617 * splitted a fat GRO packet, while we released socket lock
1618 * in skb_splice_bits()
1620 sk_eat_skb(sk, skb);
1626 * This routine provides an alternative to tcp_recvmsg() for routines
1627 * that would like to handle copying from skbuffs directly in 'sendfile'
1630 * - It is assumed that the socket was locked by the caller.
1631 * - The routine does not block.
1632 * - At present, there is no support for reading OOB data
1633 * or for 'peeking' the socket using this routine
1634 * (although both would be easy to implement).
1636 int tcp_read_sock(struct sock *sk, read_descriptor_t *desc,
1637 sk_read_actor_t recv_actor)
1639 struct sk_buff *skb;
1640 struct tcp_sock *tp = tcp_sk(sk);
1641 u32 seq = tp->copied_seq;
1645 if (sk->sk_state == TCP_LISTEN)
1647 while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1648 if (offset < skb->len) {
1652 len = skb->len - offset;
1653 /* Stop reading if we hit a patch of urgent data */
1655 u32 urg_offset = tp->urg_seq - seq;
1656 if (urg_offset < len)
1661 used = recv_actor(desc, skb, offset, len);
1666 } else if (used <= len) {
1671 /* If recv_actor drops the lock (e.g. TCP splice
1672 * receive) the skb pointer might be invalid when
1673 * getting here: tcp_collapse might have deleted it
1674 * while aggregating skbs from the socket queue.
1676 skb = tcp_recv_skb(sk, seq - 1, &offset);
1679 /* TCP coalescing might have appended data to the skb.
1680 * Try to splice more frags
1682 if (offset + 1 != skb->len)
1685 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) {
1686 sk_eat_skb(sk, skb);
1690 sk_eat_skb(sk, skb);
1693 WRITE_ONCE(tp->copied_seq, seq);
1695 WRITE_ONCE(tp->copied_seq, seq);
1697 tcp_rcv_space_adjust(sk);
1699 /* Clean up data we have read: This will do ACK frames. */
1701 tcp_recv_skb(sk, seq, &offset);
1702 tcp_cleanup_rbuf(sk, copied);
1706 EXPORT_SYMBOL(tcp_read_sock);
1708 int tcp_peek_len(struct socket *sock)
1710 return tcp_inq(sock->sk);
1712 EXPORT_SYMBOL(tcp_peek_len);
1714 /* Make sure sk_rcvbuf is big enough to satisfy SO_RCVLOWAT hint */
1715 int tcp_set_rcvlowat(struct sock *sk, int val)
1719 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK)
1720 cap = sk->sk_rcvbuf >> 1;
1722 cap = sock_net(sk)->ipv4.sysctl_tcp_rmem[2] >> 1;
1723 val = min(val, cap);
1724 WRITE_ONCE(sk->sk_rcvlowat, val ? : 1);
1726 /* Check if we need to signal EPOLLIN right now */
1729 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK)
1733 if (val > sk->sk_rcvbuf) {
1734 WRITE_ONCE(sk->sk_rcvbuf, val);
1735 tcp_sk(sk)->window_clamp = tcp_win_from_space(sk, val);
1739 EXPORT_SYMBOL(tcp_set_rcvlowat);
1741 void tcp_update_recv_tstamps(struct sk_buff *skb,
1742 struct scm_timestamping_internal *tss)
1745 tss->ts[0] = ktime_to_timespec64(skb->tstamp);
1747 tss->ts[0] = (struct timespec64) {0};
1749 if (skb_hwtstamps(skb)->hwtstamp)
1750 tss->ts[2] = ktime_to_timespec64(skb_hwtstamps(skb)->hwtstamp);
1752 tss->ts[2] = (struct timespec64) {0};
1756 static const struct vm_operations_struct tcp_vm_ops = {
1759 int tcp_mmap(struct file *file, struct socket *sock,
1760 struct vm_area_struct *vma)
1762 if (vma->vm_flags & (VM_WRITE | VM_EXEC))
1764 vma->vm_flags &= ~(VM_MAYWRITE | VM_MAYEXEC);
1766 /* Instruct vm_insert_page() to not mmap_read_lock(mm) */
1767 vma->vm_flags |= VM_MIXEDMAP;
1769 vma->vm_ops = &tcp_vm_ops;
1772 EXPORT_SYMBOL(tcp_mmap);
1774 static skb_frag_t *skb_advance_to_frag(struct sk_buff *skb, u32 offset_skb,
1779 offset_skb -= skb_headlen(skb);
1780 if ((int)offset_skb < 0 || skb_has_frag_list(skb))
1783 frag = skb_shinfo(skb)->frags;
1784 while (offset_skb) {
1785 if (skb_frag_size(frag) > offset_skb) {
1786 *offset_frag = offset_skb;
1789 offset_skb -= skb_frag_size(frag);
1796 static bool can_map_frag(const skb_frag_t *frag)
1798 return skb_frag_size(frag) == PAGE_SIZE && !skb_frag_off(frag);
1801 static int find_next_mappable_frag(const skb_frag_t *frag,
1802 int remaining_in_skb)
1806 if (likely(can_map_frag(frag)))
1809 while (offset < remaining_in_skb && !can_map_frag(frag)) {
1810 offset += skb_frag_size(frag);
1816 static void tcp_zerocopy_set_hint_for_skb(struct sock *sk,
1817 struct tcp_zerocopy_receive *zc,
1818 struct sk_buff *skb, u32 offset)
1820 u32 frag_offset, partial_frag_remainder = 0;
1821 int mappable_offset;
1824 /* worst case: skip to next skb. try to improve on this case below */
1825 zc->recv_skip_hint = skb->len - offset;
1827 /* Find the frag containing this offset (and how far into that frag) */
1828 frag = skb_advance_to_frag(skb, offset, &frag_offset);
1833 struct skb_shared_info *info = skb_shinfo(skb);
1835 /* We read part of the last frag, must recvmsg() rest of skb. */
1836 if (frag == &info->frags[info->nr_frags - 1])
1839 /* Else, we must at least read the remainder in this frag. */
1840 partial_frag_remainder = skb_frag_size(frag) - frag_offset;
1841 zc->recv_skip_hint -= partial_frag_remainder;
1845 /* partial_frag_remainder: If part way through a frag, must read rest.
1846 * mappable_offset: Bytes till next mappable frag, *not* counting bytes
1847 * in partial_frag_remainder.
1849 mappable_offset = find_next_mappable_frag(frag, zc->recv_skip_hint);
1850 zc->recv_skip_hint = mappable_offset + partial_frag_remainder;
1853 static int tcp_recvmsg_locked(struct sock *sk, struct msghdr *msg, size_t len,
1854 int nonblock, int flags,
1855 struct scm_timestamping_internal *tss,
1857 static int receive_fallback_to_copy(struct sock *sk,
1858 struct tcp_zerocopy_receive *zc, int inq,
1859 struct scm_timestamping_internal *tss)
1861 unsigned long copy_address = (unsigned long)zc->copybuf_address;
1862 struct msghdr msg = {};
1867 zc->recv_skip_hint = 0;
1869 if (copy_address != zc->copybuf_address)
1872 err = import_single_range(READ, (void __user *)copy_address,
1873 inq, &iov, &msg.msg_iter);
1877 err = tcp_recvmsg_locked(sk, &msg, inq, /*nonblock=*/1, /*flags=*/0,
1878 tss, &zc->msg_flags);
1882 zc->copybuf_len = err;
1883 if (likely(zc->copybuf_len)) {
1884 struct sk_buff *skb;
1887 skb = tcp_recv_skb(sk, tcp_sk(sk)->copied_seq, &offset);
1889 tcp_zerocopy_set_hint_for_skb(sk, zc, skb, offset);
1894 static int tcp_copy_straggler_data(struct tcp_zerocopy_receive *zc,
1895 struct sk_buff *skb, u32 copylen,
1896 u32 *offset, u32 *seq)
1898 unsigned long copy_address = (unsigned long)zc->copybuf_address;
1899 struct msghdr msg = {};
1903 if (copy_address != zc->copybuf_address)
1906 err = import_single_range(READ, (void __user *)copy_address,
1907 copylen, &iov, &msg.msg_iter);
1910 err = skb_copy_datagram_msg(skb, *offset, &msg, copylen);
1913 zc->recv_skip_hint -= copylen;
1916 return (__s32)copylen;
1919 static int tcp_zc_handle_leftover(struct tcp_zerocopy_receive *zc,
1921 struct sk_buff *skb,
1924 struct scm_timestamping_internal *tss)
1926 u32 offset, copylen = min_t(u32, copybuf_len, zc->recv_skip_hint);
1930 /* skb is null if inq < PAGE_SIZE. */
1932 offset = *seq - TCP_SKB_CB(skb)->seq;
1934 skb = tcp_recv_skb(sk, *seq, &offset);
1935 if (TCP_SKB_CB(skb)->has_rxtstamp) {
1936 tcp_update_recv_tstamps(skb, tss);
1937 zc->msg_flags |= TCP_CMSG_TS;
1941 zc->copybuf_len = tcp_copy_straggler_data(zc, skb, copylen, &offset,
1943 return zc->copybuf_len < 0 ? 0 : copylen;
1946 static int tcp_zerocopy_vm_insert_batch_error(struct vm_area_struct *vma,
1947 struct page **pending_pages,
1948 unsigned long pages_remaining,
1949 unsigned long *address,
1952 struct tcp_zerocopy_receive *zc,
1953 u32 total_bytes_to_map,
1956 /* At least one page did not map. Try zapping if we skipped earlier. */
1957 if (err == -EBUSY &&
1958 zc->flags & TCP_RECEIVE_ZEROCOPY_FLAG_TLB_CLEAN_HINT) {
1961 maybe_zap_len = total_bytes_to_map - /* All bytes to map */
1962 *length + /* Mapped or pending */
1963 (pages_remaining * PAGE_SIZE); /* Failed map. */
1964 zap_page_range(vma, *address, maybe_zap_len);
1969 unsigned long leftover_pages = pages_remaining;
1972 /* We called zap_page_range, try to reinsert. */
1973 err = vm_insert_pages(vma, *address,
1976 bytes_mapped = PAGE_SIZE * (leftover_pages - pages_remaining);
1977 *seq += bytes_mapped;
1978 *address += bytes_mapped;
1981 /* Either we were unable to zap, OR we zapped, retried an
1982 * insert, and still had an issue. Either ways, pages_remaining
1983 * is the number of pages we were unable to map, and we unroll
1984 * some state we speculatively touched before.
1986 const int bytes_not_mapped = PAGE_SIZE * pages_remaining;
1988 *length -= bytes_not_mapped;
1989 zc->recv_skip_hint += bytes_not_mapped;
1994 static int tcp_zerocopy_vm_insert_batch(struct vm_area_struct *vma,
1995 struct page **pages,
1996 unsigned int pages_to_map,
1997 unsigned long *address,
2000 struct tcp_zerocopy_receive *zc,
2001 u32 total_bytes_to_map)
2003 unsigned long pages_remaining = pages_to_map;
2004 unsigned int pages_mapped;
2005 unsigned int bytes_mapped;
2008 err = vm_insert_pages(vma, *address, pages, &pages_remaining);
2009 pages_mapped = pages_to_map - (unsigned int)pages_remaining;
2010 bytes_mapped = PAGE_SIZE * pages_mapped;
2011 /* Even if vm_insert_pages fails, it may have partially succeeded in
2012 * mapping (some but not all of the pages).
2014 *seq += bytes_mapped;
2015 *address += bytes_mapped;
2020 /* Error: maybe zap and retry + rollback state for failed inserts. */
2021 return tcp_zerocopy_vm_insert_batch_error(vma, pages + pages_mapped,
2022 pages_remaining, address, length, seq, zc, total_bytes_to_map,
2026 #define TCP_VALID_ZC_MSG_FLAGS (TCP_CMSG_TS)
2027 static void tcp_zc_finalize_rx_tstamp(struct sock *sk,
2028 struct tcp_zerocopy_receive *zc,
2029 struct scm_timestamping_internal *tss)
2031 unsigned long msg_control_addr;
2032 struct msghdr cmsg_dummy;
2034 msg_control_addr = (unsigned long)zc->msg_control;
2035 cmsg_dummy.msg_control = (void *)msg_control_addr;
2036 cmsg_dummy.msg_controllen =
2037 (__kernel_size_t)zc->msg_controllen;
2038 cmsg_dummy.msg_flags = in_compat_syscall()
2039 ? MSG_CMSG_COMPAT : 0;
2040 cmsg_dummy.msg_control_is_user = true;
2042 if (zc->msg_control == msg_control_addr &&
2043 zc->msg_controllen == cmsg_dummy.msg_controllen) {
2044 tcp_recv_timestamp(&cmsg_dummy, sk, tss);
2045 zc->msg_control = (__u64)
2046 ((uintptr_t)cmsg_dummy.msg_control);
2047 zc->msg_controllen =
2048 (__u64)cmsg_dummy.msg_controllen;
2049 zc->msg_flags = (__u32)cmsg_dummy.msg_flags;
2053 #define TCP_ZEROCOPY_PAGE_BATCH_SIZE 32
2054 static int tcp_zerocopy_receive(struct sock *sk,
2055 struct tcp_zerocopy_receive *zc,
2056 struct scm_timestamping_internal *tss)
2058 u32 length = 0, offset, vma_len, avail_len, copylen = 0;
2059 unsigned long address = (unsigned long)zc->address;
2060 struct page *pages[TCP_ZEROCOPY_PAGE_BATCH_SIZE];
2061 s32 copybuf_len = zc->copybuf_len;
2062 struct tcp_sock *tp = tcp_sk(sk);
2063 const skb_frag_t *frags = NULL;
2064 unsigned int pages_to_map = 0;
2065 struct vm_area_struct *vma;
2066 struct sk_buff *skb = NULL;
2067 u32 seq = tp->copied_seq;
2068 u32 total_bytes_to_map;
2069 int inq = tcp_inq(sk);
2072 zc->copybuf_len = 0;
2075 if (address & (PAGE_SIZE - 1) || address != zc->address)
2078 if (sk->sk_state == TCP_LISTEN)
2081 sock_rps_record_flow(sk);
2083 if (inq && inq <= copybuf_len)
2084 return receive_fallback_to_copy(sk, zc, inq, tss);
2086 if (inq < PAGE_SIZE) {
2088 zc->recv_skip_hint = inq;
2089 if (!inq && sock_flag(sk, SOCK_DONE))
2094 mmap_read_lock(current->mm);
2096 vma = vma_lookup(current->mm, address);
2097 if (!vma || vma->vm_ops != &tcp_vm_ops) {
2098 mmap_read_unlock(current->mm);
2101 vma_len = min_t(unsigned long, zc->length, vma->vm_end - address);
2102 avail_len = min_t(u32, vma_len, inq);
2103 total_bytes_to_map = avail_len & ~(PAGE_SIZE - 1);
2104 if (total_bytes_to_map) {
2105 if (!(zc->flags & TCP_RECEIVE_ZEROCOPY_FLAG_TLB_CLEAN_HINT))
2106 zap_page_range(vma, address, total_bytes_to_map);
2107 zc->length = total_bytes_to_map;
2108 zc->recv_skip_hint = 0;
2110 zc->length = avail_len;
2111 zc->recv_skip_hint = avail_len;
2114 while (length + PAGE_SIZE <= zc->length) {
2115 int mappable_offset;
2118 if (zc->recv_skip_hint < PAGE_SIZE) {
2122 if (zc->recv_skip_hint > 0)
2125 offset = seq - TCP_SKB_CB(skb)->seq;
2127 skb = tcp_recv_skb(sk, seq, &offset);
2130 if (TCP_SKB_CB(skb)->has_rxtstamp) {
2131 tcp_update_recv_tstamps(skb, tss);
2132 zc->msg_flags |= TCP_CMSG_TS;
2134 zc->recv_skip_hint = skb->len - offset;
2135 frags = skb_advance_to_frag(skb, offset, &offset_frag);
2136 if (!frags || offset_frag)
2140 mappable_offset = find_next_mappable_frag(frags,
2141 zc->recv_skip_hint);
2142 if (mappable_offset) {
2143 zc->recv_skip_hint = mappable_offset;
2146 page = skb_frag_page(frags);
2148 pages[pages_to_map++] = page;
2149 length += PAGE_SIZE;
2150 zc->recv_skip_hint -= PAGE_SIZE;
2152 if (pages_to_map == TCP_ZEROCOPY_PAGE_BATCH_SIZE ||
2153 zc->recv_skip_hint < PAGE_SIZE) {
2154 /* Either full batch, or we're about to go to next skb
2155 * (and we cannot unroll failed ops across skbs).
2157 ret = tcp_zerocopy_vm_insert_batch(vma, pages,
2161 total_bytes_to_map);
2168 ret = tcp_zerocopy_vm_insert_batch(vma, pages, pages_to_map,
2169 &address, &length, &seq,
2170 zc, total_bytes_to_map);
2173 mmap_read_unlock(current->mm);
2174 /* Try to copy straggler data. */
2176 copylen = tcp_zc_handle_leftover(zc, sk, skb, &seq, copybuf_len, tss);
2178 if (length + copylen) {
2179 WRITE_ONCE(tp->copied_seq, seq);
2180 tcp_rcv_space_adjust(sk);
2182 /* Clean up data we have read: This will do ACK frames. */
2183 tcp_recv_skb(sk, seq, &offset);
2184 tcp_cleanup_rbuf(sk, length + copylen);
2186 if (length == zc->length)
2187 zc->recv_skip_hint = 0;
2189 if (!zc->recv_skip_hint && sock_flag(sk, SOCK_DONE))
2192 zc->length = length;
2197 /* Similar to __sock_recv_timestamp, but does not require an skb */
2198 void tcp_recv_timestamp(struct msghdr *msg, const struct sock *sk,
2199 struct scm_timestamping_internal *tss)
2201 int new_tstamp = sock_flag(sk, SOCK_TSTAMP_NEW);
2202 bool has_timestamping = false;
2204 if (tss->ts[0].tv_sec || tss->ts[0].tv_nsec) {
2205 if (sock_flag(sk, SOCK_RCVTSTAMP)) {
2206 if (sock_flag(sk, SOCK_RCVTSTAMPNS)) {
2208 struct __kernel_timespec kts = {
2209 .tv_sec = tss->ts[0].tv_sec,
2210 .tv_nsec = tss->ts[0].tv_nsec,
2212 put_cmsg(msg, SOL_SOCKET, SO_TIMESTAMPNS_NEW,
2215 struct __kernel_old_timespec ts_old = {
2216 .tv_sec = tss->ts[0].tv_sec,
2217 .tv_nsec = tss->ts[0].tv_nsec,
2219 put_cmsg(msg, SOL_SOCKET, SO_TIMESTAMPNS_OLD,
2220 sizeof(ts_old), &ts_old);
2224 struct __kernel_sock_timeval stv = {
2225 .tv_sec = tss->ts[0].tv_sec,
2226 .tv_usec = tss->ts[0].tv_nsec / 1000,
2228 put_cmsg(msg, SOL_SOCKET, SO_TIMESTAMP_NEW,
2231 struct __kernel_old_timeval tv = {
2232 .tv_sec = tss->ts[0].tv_sec,
2233 .tv_usec = tss->ts[0].tv_nsec / 1000,
2235 put_cmsg(msg, SOL_SOCKET, SO_TIMESTAMP_OLD,
2241 if (sk->sk_tsflags & SOF_TIMESTAMPING_SOFTWARE)
2242 has_timestamping = true;
2244 tss->ts[0] = (struct timespec64) {0};
2247 if (tss->ts[2].tv_sec || tss->ts[2].tv_nsec) {
2248 if (sk->sk_tsflags & SOF_TIMESTAMPING_RAW_HARDWARE)
2249 has_timestamping = true;
2251 tss->ts[2] = (struct timespec64) {0};
2254 if (has_timestamping) {
2255 tss->ts[1] = (struct timespec64) {0};
2256 if (sock_flag(sk, SOCK_TSTAMP_NEW))
2257 put_cmsg_scm_timestamping64(msg, tss);
2259 put_cmsg_scm_timestamping(msg, tss);
2263 static int tcp_inq_hint(struct sock *sk)
2265 const struct tcp_sock *tp = tcp_sk(sk);
2266 u32 copied_seq = READ_ONCE(tp->copied_seq);
2267 u32 rcv_nxt = READ_ONCE(tp->rcv_nxt);
2270 inq = rcv_nxt - copied_seq;
2271 if (unlikely(inq < 0 || copied_seq != READ_ONCE(tp->copied_seq))) {
2273 inq = tp->rcv_nxt - tp->copied_seq;
2276 /* After receiving a FIN, tell the user-space to continue reading
2277 * by returning a non-zero inq.
2279 if (inq == 0 && sock_flag(sk, SOCK_DONE))
2285 * This routine copies from a sock struct into the user buffer.
2287 * Technical note: in 2.3 we work on _locked_ socket, so that
2288 * tricks with *seq access order and skb->users are not required.
2289 * Probably, code can be easily improved even more.
2292 static int tcp_recvmsg_locked(struct sock *sk, struct msghdr *msg, size_t len,
2293 int nonblock, int flags,
2294 struct scm_timestamping_internal *tss,
2297 struct tcp_sock *tp = tcp_sk(sk);
2303 int target; /* Read at least this many bytes */
2305 struct sk_buff *skb, *last;
2309 if (sk->sk_state == TCP_LISTEN)
2312 if (tp->recvmsg_inq)
2313 *cmsg_flags = TCP_CMSG_INQ;
2314 timeo = sock_rcvtimeo(sk, nonblock);
2316 /* Urgent data needs to be handled specially. */
2317 if (flags & MSG_OOB)
2320 if (unlikely(tp->repair)) {
2322 if (!(flags & MSG_PEEK))
2325 if (tp->repair_queue == TCP_SEND_QUEUE)
2329 if (tp->repair_queue == TCP_NO_QUEUE)
2332 /* 'common' recv queue MSG_PEEK-ing */
2335 seq = &tp->copied_seq;
2336 if (flags & MSG_PEEK) {
2337 peek_seq = tp->copied_seq;
2341 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
2346 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
2347 if (tp->urg_data && tp->urg_seq == *seq) {
2350 if (signal_pending(current)) {
2351 copied = timeo ? sock_intr_errno(timeo) : -EAGAIN;
2356 /* Next get a buffer. */
2358 last = skb_peek_tail(&sk->sk_receive_queue);
2359 skb_queue_walk(&sk->sk_receive_queue, skb) {
2361 /* Now that we have two receive queues this
2364 if (WARN(before(*seq, TCP_SKB_CB(skb)->seq),
2365 "TCP recvmsg seq # bug: copied %X, seq %X, rcvnxt %X, fl %X\n",
2366 *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt,
2370 offset = *seq - TCP_SKB_CB(skb)->seq;
2371 if (unlikely(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
2372 pr_err_once("%s: found a SYN, please report !\n", __func__);
2375 if (offset < skb->len)
2377 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
2379 WARN(!(flags & MSG_PEEK),
2380 "TCP recvmsg seq # bug 2: copied %X, seq %X, rcvnxt %X, fl %X\n",
2381 *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt, flags);
2384 /* Well, if we have backlog, try to process it now yet. */
2386 if (copied >= target && !READ_ONCE(sk->sk_backlog.tail))
2391 sk->sk_state == TCP_CLOSE ||
2392 (sk->sk_shutdown & RCV_SHUTDOWN) ||
2394 signal_pending(current))
2397 if (sock_flag(sk, SOCK_DONE))
2401 copied = sock_error(sk);
2405 if (sk->sk_shutdown & RCV_SHUTDOWN)
2408 if (sk->sk_state == TCP_CLOSE) {
2409 /* This occurs when user tries to read
2410 * from never connected socket.
2421 if (signal_pending(current)) {
2422 copied = sock_intr_errno(timeo);
2427 tcp_cleanup_rbuf(sk, copied);
2429 if (copied >= target) {
2430 /* Do not sleep, just process backlog. */
2434 sk_wait_data(sk, &timeo, last);
2437 if ((flags & MSG_PEEK) &&
2438 (peek_seq - copied - urg_hole != tp->copied_seq)) {
2439 net_dbg_ratelimited("TCP(%s:%d): Application bug, race in MSG_PEEK\n",
2441 task_pid_nr(current));
2442 peek_seq = tp->copied_seq;
2447 /* Ok so how much can we use? */
2448 used = skb->len - offset;
2452 /* Do we have urgent data here? */
2454 u32 urg_offset = tp->urg_seq - *seq;
2455 if (urg_offset < used) {
2457 if (!sock_flag(sk, SOCK_URGINLINE)) {
2458 WRITE_ONCE(*seq, *seq + 1);
2470 if (!(flags & MSG_TRUNC)) {
2471 err = skb_copy_datagram_msg(skb, offset, msg, used);
2473 /* Exception. Bailout! */
2480 WRITE_ONCE(*seq, *seq + used);
2484 tcp_rcv_space_adjust(sk);
2487 if (tp->urg_data && after(tp->copied_seq, tp->urg_seq)) {
2489 tcp_fast_path_check(sk);
2492 if (TCP_SKB_CB(skb)->has_rxtstamp) {
2493 tcp_update_recv_tstamps(skb, tss);
2494 *cmsg_flags |= TCP_CMSG_TS;
2497 if (used + offset < skb->len)
2500 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
2502 if (!(flags & MSG_PEEK))
2503 sk_eat_skb(sk, skb);
2507 /* Process the FIN. */
2508 WRITE_ONCE(*seq, *seq + 1);
2509 if (!(flags & MSG_PEEK))
2510 sk_eat_skb(sk, skb);
2514 /* According to UNIX98, msg_name/msg_namelen are ignored
2515 * on connected socket. I was just happy when found this 8) --ANK
2518 /* Clean up data we have read: This will do ACK frames. */
2519 tcp_cleanup_rbuf(sk, copied);
2526 err = tcp_recv_urg(sk, msg, len, flags);
2530 err = tcp_peek_sndq(sk, msg, len);
2534 int tcp_recvmsg(struct sock *sk, struct msghdr *msg, size_t len, int nonblock,
2535 int flags, int *addr_len)
2537 int cmsg_flags = 0, ret, inq;
2538 struct scm_timestamping_internal tss;
2540 if (unlikely(flags & MSG_ERRQUEUE))
2541 return inet_recv_error(sk, msg, len, addr_len);
2543 if (sk_can_busy_loop(sk) &&
2544 skb_queue_empty_lockless(&sk->sk_receive_queue) &&
2545 sk->sk_state == TCP_ESTABLISHED)
2546 sk_busy_loop(sk, nonblock);
2549 ret = tcp_recvmsg_locked(sk, msg, len, nonblock, flags, &tss,
2553 if (cmsg_flags && ret >= 0) {
2554 if (cmsg_flags & TCP_CMSG_TS)
2555 tcp_recv_timestamp(msg, sk, &tss);
2556 if (cmsg_flags & TCP_CMSG_INQ) {
2557 inq = tcp_inq_hint(sk);
2558 put_cmsg(msg, SOL_TCP, TCP_CM_INQ, sizeof(inq), &inq);
2563 EXPORT_SYMBOL(tcp_recvmsg);
2565 void tcp_set_state(struct sock *sk, int state)
2567 int oldstate = sk->sk_state;
2569 /* We defined a new enum for TCP states that are exported in BPF
2570 * so as not force the internal TCP states to be frozen. The
2571 * following checks will detect if an internal state value ever
2572 * differs from the BPF value. If this ever happens, then we will
2573 * need to remap the internal value to the BPF value before calling
2574 * tcp_call_bpf_2arg.
2576 BUILD_BUG_ON((int)BPF_TCP_ESTABLISHED != (int)TCP_ESTABLISHED);
2577 BUILD_BUG_ON((int)BPF_TCP_SYN_SENT != (int)TCP_SYN_SENT);
2578 BUILD_BUG_ON((int)BPF_TCP_SYN_RECV != (int)TCP_SYN_RECV);
2579 BUILD_BUG_ON((int)BPF_TCP_FIN_WAIT1 != (int)TCP_FIN_WAIT1);
2580 BUILD_BUG_ON((int)BPF_TCP_FIN_WAIT2 != (int)TCP_FIN_WAIT2);
2581 BUILD_BUG_ON((int)BPF_TCP_TIME_WAIT != (int)TCP_TIME_WAIT);
2582 BUILD_BUG_ON((int)BPF_TCP_CLOSE != (int)TCP_CLOSE);
2583 BUILD_BUG_ON((int)BPF_TCP_CLOSE_WAIT != (int)TCP_CLOSE_WAIT);
2584 BUILD_BUG_ON((int)BPF_TCP_LAST_ACK != (int)TCP_LAST_ACK);
2585 BUILD_BUG_ON((int)BPF_TCP_LISTEN != (int)TCP_LISTEN);
2586 BUILD_BUG_ON((int)BPF_TCP_CLOSING != (int)TCP_CLOSING);
2587 BUILD_BUG_ON((int)BPF_TCP_NEW_SYN_RECV != (int)TCP_NEW_SYN_RECV);
2588 BUILD_BUG_ON((int)BPF_TCP_MAX_STATES != (int)TCP_MAX_STATES);
2590 /* bpf uapi header bpf.h defines an anonymous enum with values
2591 * BPF_TCP_* used by bpf programs. Currently gcc built vmlinux
2592 * is able to emit this enum in DWARF due to the above BUILD_BUG_ON.
2593 * But clang built vmlinux does not have this enum in DWARF
2594 * since clang removes the above code before generating IR/debuginfo.
2595 * Let us explicitly emit the type debuginfo to ensure the
2596 * above-mentioned anonymous enum in the vmlinux DWARF and hence BTF
2597 * regardless of which compiler is used.
2599 BTF_TYPE_EMIT_ENUM(BPF_TCP_ESTABLISHED);
2601 if (BPF_SOCK_OPS_TEST_FLAG(tcp_sk(sk), BPF_SOCK_OPS_STATE_CB_FLAG))
2602 tcp_call_bpf_2arg(sk, BPF_SOCK_OPS_STATE_CB, oldstate, state);
2605 case TCP_ESTABLISHED:
2606 if (oldstate != TCP_ESTABLISHED)
2607 TCP_INC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
2611 if (oldstate == TCP_CLOSE_WAIT || oldstate == TCP_ESTABLISHED)
2612 TCP_INC_STATS(sock_net(sk), TCP_MIB_ESTABRESETS);
2614 sk->sk_prot->unhash(sk);
2615 if (inet_csk(sk)->icsk_bind_hash &&
2616 !(sk->sk_userlocks & SOCK_BINDPORT_LOCK))
2620 if (oldstate == TCP_ESTABLISHED)
2621 TCP_DEC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
2624 /* Change state AFTER socket is unhashed to avoid closed
2625 * socket sitting in hash tables.
2627 inet_sk_state_store(sk, state);
2629 EXPORT_SYMBOL_GPL(tcp_set_state);
2632 * State processing on a close. This implements the state shift for
2633 * sending our FIN frame. Note that we only send a FIN for some
2634 * states. A shutdown() may have already sent the FIN, or we may be
2638 static const unsigned char new_state[16] = {
2639 /* current state: new state: action: */
2640 [0 /* (Invalid) */] = TCP_CLOSE,
2641 [TCP_ESTABLISHED] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2642 [TCP_SYN_SENT] = TCP_CLOSE,
2643 [TCP_SYN_RECV] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2644 [TCP_FIN_WAIT1] = TCP_FIN_WAIT1,
2645 [TCP_FIN_WAIT2] = TCP_FIN_WAIT2,
2646 [TCP_TIME_WAIT] = TCP_CLOSE,
2647 [TCP_CLOSE] = TCP_CLOSE,
2648 [TCP_CLOSE_WAIT] = TCP_LAST_ACK | TCP_ACTION_FIN,
2649 [TCP_LAST_ACK] = TCP_LAST_ACK,
2650 [TCP_LISTEN] = TCP_CLOSE,
2651 [TCP_CLOSING] = TCP_CLOSING,
2652 [TCP_NEW_SYN_RECV] = TCP_CLOSE, /* should not happen ! */
2655 static int tcp_close_state(struct sock *sk)
2657 int next = (int)new_state[sk->sk_state];
2658 int ns = next & TCP_STATE_MASK;
2660 tcp_set_state(sk, ns);
2662 return next & TCP_ACTION_FIN;
2666 * Shutdown the sending side of a connection. Much like close except
2667 * that we don't receive shut down or sock_set_flag(sk, SOCK_DEAD).
2670 void tcp_shutdown(struct sock *sk, int how)
2672 /* We need to grab some memory, and put together a FIN,
2673 * and then put it into the queue to be sent.
2674 * Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
2676 if (!(how & SEND_SHUTDOWN))
2679 /* If we've already sent a FIN, or it's a closed state, skip this. */
2680 if ((1 << sk->sk_state) &
2681 (TCPF_ESTABLISHED | TCPF_SYN_SENT |
2682 TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) {
2683 /* Clear out any half completed packets. FIN if needed. */
2684 if (tcp_close_state(sk))
2688 EXPORT_SYMBOL(tcp_shutdown);
2690 bool tcp_check_oom(struct sock *sk, int shift)
2692 bool too_many_orphans, out_of_socket_memory;
2694 too_many_orphans = tcp_too_many_orphans(sk, shift);
2695 out_of_socket_memory = tcp_out_of_memory(sk);
2697 if (too_many_orphans)
2698 net_info_ratelimited("too many orphaned sockets\n");
2699 if (out_of_socket_memory)
2700 net_info_ratelimited("out of memory -- consider tuning tcp_mem\n");
2701 return too_many_orphans || out_of_socket_memory;
2704 void __tcp_close(struct sock *sk, long timeout)
2706 struct sk_buff *skb;
2707 int data_was_unread = 0;
2710 sk->sk_shutdown = SHUTDOWN_MASK;
2712 if (sk->sk_state == TCP_LISTEN) {
2713 tcp_set_state(sk, TCP_CLOSE);
2716 inet_csk_listen_stop(sk);
2718 goto adjudge_to_death;
2721 /* We need to flush the recv. buffs. We do this only on the
2722 * descriptor close, not protocol-sourced closes, because the
2723 * reader process may not have drained the data yet!
2725 while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) {
2726 u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq;
2728 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
2730 data_was_unread += len;
2736 /* If socket has been already reset (e.g. in tcp_reset()) - kill it. */
2737 if (sk->sk_state == TCP_CLOSE)
2738 goto adjudge_to_death;
2740 /* As outlined in RFC 2525, section 2.17, we send a RST here because
2741 * data was lost. To witness the awful effects of the old behavior of
2742 * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk
2743 * GET in an FTP client, suspend the process, wait for the client to
2744 * advertise a zero window, then kill -9 the FTP client, wheee...
2745 * Note: timeout is always zero in such a case.
2747 if (unlikely(tcp_sk(sk)->repair)) {
2748 sk->sk_prot->disconnect(sk, 0);
2749 } else if (data_was_unread) {
2750 /* Unread data was tossed, zap the connection. */
2751 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONCLOSE);
2752 tcp_set_state(sk, TCP_CLOSE);
2753 tcp_send_active_reset(sk, sk->sk_allocation);
2754 } else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) {
2755 /* Check zero linger _after_ checking for unread data. */
2756 sk->sk_prot->disconnect(sk, 0);
2757 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
2758 } else if (tcp_close_state(sk)) {
2759 /* We FIN if the application ate all the data before
2760 * zapping the connection.
2763 /* RED-PEN. Formally speaking, we have broken TCP state
2764 * machine. State transitions:
2766 * TCP_ESTABLISHED -> TCP_FIN_WAIT1
2767 * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible)
2768 * TCP_CLOSE_WAIT -> TCP_LAST_ACK
2770 * are legal only when FIN has been sent (i.e. in window),
2771 * rather than queued out of window. Purists blame.
2773 * F.e. "RFC state" is ESTABLISHED,
2774 * if Linux state is FIN-WAIT-1, but FIN is still not sent.
2776 * The visible declinations are that sometimes
2777 * we enter time-wait state, when it is not required really
2778 * (harmless), do not send active resets, when they are
2779 * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
2780 * they look as CLOSING or LAST_ACK for Linux)
2781 * Probably, I missed some more holelets.
2783 * XXX (TFO) - To start off we don't support SYN+ACK+FIN
2784 * in a single packet! (May consider it later but will
2785 * probably need API support or TCP_CORK SYN-ACK until
2786 * data is written and socket is closed.)
2791 sk_stream_wait_close(sk, timeout);
2794 state = sk->sk_state;
2800 /* remove backlog if any, without releasing ownership. */
2803 percpu_counter_inc(sk->sk_prot->orphan_count);
2805 /* Have we already been destroyed by a softirq or backlog? */
2806 if (state != TCP_CLOSE && sk->sk_state == TCP_CLOSE)
2809 /* This is a (useful) BSD violating of the RFC. There is a
2810 * problem with TCP as specified in that the other end could
2811 * keep a socket open forever with no application left this end.
2812 * We use a 1 minute timeout (about the same as BSD) then kill
2813 * our end. If they send after that then tough - BUT: long enough
2814 * that we won't make the old 4*rto = almost no time - whoops
2817 * Nope, it was not mistake. It is really desired behaviour
2818 * f.e. on http servers, when such sockets are useless, but
2819 * consume significant resources. Let's do it with special
2820 * linger2 option. --ANK
2823 if (sk->sk_state == TCP_FIN_WAIT2) {
2824 struct tcp_sock *tp = tcp_sk(sk);
2825 if (tp->linger2 < 0) {
2826 tcp_set_state(sk, TCP_CLOSE);
2827 tcp_send_active_reset(sk, GFP_ATOMIC);
2828 __NET_INC_STATS(sock_net(sk),
2829 LINUX_MIB_TCPABORTONLINGER);
2831 const int tmo = tcp_fin_time(sk);
2833 if (tmo > TCP_TIMEWAIT_LEN) {
2834 inet_csk_reset_keepalive_timer(sk,
2835 tmo - TCP_TIMEWAIT_LEN);
2837 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
2842 if (sk->sk_state != TCP_CLOSE) {
2844 if (tcp_check_oom(sk, 0)) {
2845 tcp_set_state(sk, TCP_CLOSE);
2846 tcp_send_active_reset(sk, GFP_ATOMIC);
2847 __NET_INC_STATS(sock_net(sk),
2848 LINUX_MIB_TCPABORTONMEMORY);
2849 } else if (!check_net(sock_net(sk))) {
2850 /* Not possible to send reset; just close */
2851 tcp_set_state(sk, TCP_CLOSE);
2855 if (sk->sk_state == TCP_CLOSE) {
2856 struct request_sock *req;
2858 req = rcu_dereference_protected(tcp_sk(sk)->fastopen_rsk,
2859 lockdep_sock_is_held(sk));
2860 /* We could get here with a non-NULL req if the socket is
2861 * aborted (e.g., closed with unread data) before 3WHS
2865 reqsk_fastopen_remove(sk, req, false);
2866 inet_csk_destroy_sock(sk);
2868 /* Otherwise, socket is reprieved until protocol close. */
2875 void tcp_close(struct sock *sk, long timeout)
2878 __tcp_close(sk, timeout);
2882 EXPORT_SYMBOL(tcp_close);
2884 /* These states need RST on ABORT according to RFC793 */
2886 static inline bool tcp_need_reset(int state)
2888 return (1 << state) &
2889 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 |
2890 TCPF_FIN_WAIT2 | TCPF_SYN_RECV);
2893 static void tcp_rtx_queue_purge(struct sock *sk)
2895 struct rb_node *p = rb_first(&sk->tcp_rtx_queue);
2897 tcp_sk(sk)->highest_sack = NULL;
2899 struct sk_buff *skb = rb_to_skb(p);
2902 /* Since we are deleting whole queue, no need to
2903 * list_del(&skb->tcp_tsorted_anchor)
2905 tcp_rtx_queue_unlink(skb, sk);
2906 sk_wmem_free_skb(sk, skb);
2910 void tcp_write_queue_purge(struct sock *sk)
2912 struct sk_buff *skb;
2914 tcp_chrono_stop(sk, TCP_CHRONO_BUSY);
2915 while ((skb = __skb_dequeue(&sk->sk_write_queue)) != NULL) {
2916 tcp_skb_tsorted_anchor_cleanup(skb);
2917 sk_wmem_free_skb(sk, skb);
2919 tcp_rtx_queue_purge(sk);
2920 skb = sk->sk_tx_skb_cache;
2923 sk->sk_tx_skb_cache = NULL;
2925 INIT_LIST_HEAD(&tcp_sk(sk)->tsorted_sent_queue);
2927 tcp_clear_all_retrans_hints(tcp_sk(sk));
2928 tcp_sk(sk)->packets_out = 0;
2929 inet_csk(sk)->icsk_backoff = 0;
2932 int tcp_disconnect(struct sock *sk, int flags)
2934 struct inet_sock *inet = inet_sk(sk);
2935 struct inet_connection_sock *icsk = inet_csk(sk);
2936 struct tcp_sock *tp = tcp_sk(sk);
2937 int old_state = sk->sk_state;
2940 if (old_state != TCP_CLOSE)
2941 tcp_set_state(sk, TCP_CLOSE);
2943 /* ABORT function of RFC793 */
2944 if (old_state == TCP_LISTEN) {
2945 inet_csk_listen_stop(sk);
2946 } else if (unlikely(tp->repair)) {
2947 sk->sk_err = ECONNABORTED;
2948 } else if (tcp_need_reset(old_state) ||
2949 (tp->snd_nxt != tp->write_seq &&
2950 (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK))) {
2951 /* The last check adjusts for discrepancy of Linux wrt. RFC
2954 tcp_send_active_reset(sk, gfp_any());
2955 sk->sk_err = ECONNRESET;
2956 } else if (old_state == TCP_SYN_SENT)
2957 sk->sk_err = ECONNRESET;
2959 tcp_clear_xmit_timers(sk);
2960 __skb_queue_purge(&sk->sk_receive_queue);
2961 if (sk->sk_rx_skb_cache) {
2962 __kfree_skb(sk->sk_rx_skb_cache);
2963 sk->sk_rx_skb_cache = NULL;
2965 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt);
2967 tcp_write_queue_purge(sk);
2968 tcp_fastopen_active_disable_ofo_check(sk);
2969 skb_rbtree_purge(&tp->out_of_order_queue);
2971 inet->inet_dport = 0;
2973 if (!(sk->sk_userlocks & SOCK_BINDADDR_LOCK))
2974 inet_reset_saddr(sk);
2976 sk->sk_shutdown = 0;
2977 sock_reset_flag(sk, SOCK_DONE);
2979 tp->mdev_us = jiffies_to_usecs(TCP_TIMEOUT_INIT);
2980 tp->rcv_rtt_last_tsecr = 0;
2982 seq = tp->write_seq + tp->max_window + 2;
2985 WRITE_ONCE(tp->write_seq, seq);
2987 icsk->icsk_backoff = 0;
2988 icsk->icsk_probes_out = 0;
2989 icsk->icsk_probes_tstamp = 0;
2990 icsk->icsk_rto = TCP_TIMEOUT_INIT;
2991 icsk->icsk_rto_min = TCP_RTO_MIN;
2992 icsk->icsk_delack_max = TCP_DELACK_MAX;
2993 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
2994 tp->snd_cwnd = TCP_INIT_CWND;
2995 tp->snd_cwnd_cnt = 0;
2996 tp->window_clamp = 0;
2998 tp->delivered_ce = 0;
2999 if (icsk->icsk_ca_ops->release)
3000 icsk->icsk_ca_ops->release(sk);
3001 memset(icsk->icsk_ca_priv, 0, sizeof(icsk->icsk_ca_priv));
3002 icsk->icsk_ca_initialized = 0;
3003 tcp_set_ca_state(sk, TCP_CA_Open);
3004 tp->is_sack_reneg = 0;
3005 tcp_clear_retrans(tp);
3006 tp->total_retrans = 0;
3007 inet_csk_delack_init(sk);
3008 /* Initialize rcv_mss to TCP_MIN_MSS to avoid division by 0
3009 * issue in __tcp_select_window()
3011 icsk->icsk_ack.rcv_mss = TCP_MIN_MSS;
3012 memset(&tp->rx_opt, 0, sizeof(tp->rx_opt));
3014 dst_release(sk->sk_rx_dst);
3015 sk->sk_rx_dst = NULL;
3016 tcp_saved_syn_free(tp);
3017 tp->compressed_ack = 0;
3021 tp->bytes_acked = 0;
3022 tp->bytes_received = 0;
3023 tp->bytes_retrans = 0;
3024 tp->data_segs_in = 0;
3025 tp->data_segs_out = 0;
3026 tp->duplicate_sack[0].start_seq = 0;
3027 tp->duplicate_sack[0].end_seq = 0;
3030 tp->retrans_out = 0;
3032 tp->tlp_high_seq = 0;
3033 tp->last_oow_ack_time = 0;
3034 /* There's a bubble in the pipe until at least the first ACK. */
3035 tp->app_limited = ~0U;
3036 tp->rack.mstamp = 0;
3037 tp->rack.advanced = 0;
3038 tp->rack.reo_wnd_steps = 1;
3039 tp->rack.last_delivered = 0;
3040 tp->rack.reo_wnd_persist = 0;
3041 tp->rack.dsack_seen = 0;
3042 tp->syn_data_acked = 0;
3043 tp->rx_opt.saw_tstamp = 0;
3044 tp->rx_opt.dsack = 0;
3045 tp->rx_opt.num_sacks = 0;
3046 tp->rcv_ooopack = 0;
3049 /* Clean up fastopen related fields */
3050 tcp_free_fastopen_req(tp);
3051 inet->defer_connect = 0;
3052 tp->fastopen_client_fail = 0;
3054 WARN_ON(inet->inet_num && !icsk->icsk_bind_hash);
3056 if (sk->sk_frag.page) {
3057 put_page(sk->sk_frag.page);
3058 sk->sk_frag.page = NULL;
3059 sk->sk_frag.offset = 0;
3062 sk_error_report(sk);
3065 EXPORT_SYMBOL(tcp_disconnect);
3067 static inline bool tcp_can_repair_sock(const struct sock *sk)
3069 return ns_capable(sock_net(sk)->user_ns, CAP_NET_ADMIN) &&
3070 (sk->sk_state != TCP_LISTEN);
3073 static int tcp_repair_set_window(struct tcp_sock *tp, sockptr_t optbuf, int len)
3075 struct tcp_repair_window opt;
3080 if (len != sizeof(opt))
3083 if (copy_from_sockptr(&opt, optbuf, sizeof(opt)))
3086 if (opt.max_window < opt.snd_wnd)
3089 if (after(opt.snd_wl1, tp->rcv_nxt + opt.rcv_wnd))
3092 if (after(opt.rcv_wup, tp->rcv_nxt))
3095 tp->snd_wl1 = opt.snd_wl1;
3096 tp->snd_wnd = opt.snd_wnd;
3097 tp->max_window = opt.max_window;
3099 tp->rcv_wnd = opt.rcv_wnd;
3100 tp->rcv_wup = opt.rcv_wup;
3105 static int tcp_repair_options_est(struct sock *sk, sockptr_t optbuf,
3108 struct tcp_sock *tp = tcp_sk(sk);
3109 struct tcp_repair_opt opt;
3112 while (len >= sizeof(opt)) {
3113 if (copy_from_sockptr_offset(&opt, optbuf, offset, sizeof(opt)))
3116 offset += sizeof(opt);
3119 switch (opt.opt_code) {
3121 tp->rx_opt.mss_clamp = opt.opt_val;
3126 u16 snd_wscale = opt.opt_val & 0xFFFF;
3127 u16 rcv_wscale = opt.opt_val >> 16;
3129 if (snd_wscale > TCP_MAX_WSCALE || rcv_wscale > TCP_MAX_WSCALE)
3132 tp->rx_opt.snd_wscale = snd_wscale;
3133 tp->rx_opt.rcv_wscale = rcv_wscale;
3134 tp->rx_opt.wscale_ok = 1;
3137 case TCPOPT_SACK_PERM:
3138 if (opt.opt_val != 0)
3141 tp->rx_opt.sack_ok |= TCP_SACK_SEEN;
3143 case TCPOPT_TIMESTAMP:
3144 if (opt.opt_val != 0)
3147 tp->rx_opt.tstamp_ok = 1;
3155 DEFINE_STATIC_KEY_FALSE(tcp_tx_delay_enabled);
3156 EXPORT_SYMBOL(tcp_tx_delay_enabled);
3158 static void tcp_enable_tx_delay(void)
3160 if (!static_branch_unlikely(&tcp_tx_delay_enabled)) {
3161 static int __tcp_tx_delay_enabled = 0;
3163 if (cmpxchg(&__tcp_tx_delay_enabled, 0, 1) == 0) {
3164 static_branch_enable(&tcp_tx_delay_enabled);
3165 pr_info("TCP_TX_DELAY enabled\n");
3170 /* When set indicates to always queue non-full frames. Later the user clears
3171 * this option and we transmit any pending partial frames in the queue. This is
3172 * meant to be used alongside sendfile() to get properly filled frames when the
3173 * user (for example) must write out headers with a write() call first and then
3174 * use sendfile to send out the data parts.
3176 * TCP_CORK can be set together with TCP_NODELAY and it is stronger than
3179 static void __tcp_sock_set_cork(struct sock *sk, bool on)
3181 struct tcp_sock *tp = tcp_sk(sk);
3184 tp->nonagle |= TCP_NAGLE_CORK;
3186 tp->nonagle &= ~TCP_NAGLE_CORK;
3187 if (tp->nonagle & TCP_NAGLE_OFF)
3188 tp->nonagle |= TCP_NAGLE_PUSH;
3189 tcp_push_pending_frames(sk);
3193 void tcp_sock_set_cork(struct sock *sk, bool on)
3196 __tcp_sock_set_cork(sk, on);
3199 EXPORT_SYMBOL(tcp_sock_set_cork);
3201 /* TCP_NODELAY is weaker than TCP_CORK, so that this option on corked socket is
3202 * remembered, but it is not activated until cork is cleared.
3204 * However, when TCP_NODELAY is set we make an explicit push, which overrides
3205 * even TCP_CORK for currently queued segments.
3207 static void __tcp_sock_set_nodelay(struct sock *sk, bool on)
3210 tcp_sk(sk)->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH;
3211 tcp_push_pending_frames(sk);
3213 tcp_sk(sk)->nonagle &= ~TCP_NAGLE_OFF;
3217 void tcp_sock_set_nodelay(struct sock *sk)
3220 __tcp_sock_set_nodelay(sk, true);
3223 EXPORT_SYMBOL(tcp_sock_set_nodelay);
3225 static void __tcp_sock_set_quickack(struct sock *sk, int val)
3228 inet_csk_enter_pingpong_mode(sk);
3232 inet_csk_exit_pingpong_mode(sk);
3233 if ((1 << sk->sk_state) & (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) &&
3234 inet_csk_ack_scheduled(sk)) {
3235 inet_csk(sk)->icsk_ack.pending |= ICSK_ACK_PUSHED;
3236 tcp_cleanup_rbuf(sk, 1);
3238 inet_csk_enter_pingpong_mode(sk);
3242 void tcp_sock_set_quickack(struct sock *sk, int val)
3245 __tcp_sock_set_quickack(sk, val);
3248 EXPORT_SYMBOL(tcp_sock_set_quickack);
3250 int tcp_sock_set_syncnt(struct sock *sk, int val)
3252 if (val < 1 || val > MAX_TCP_SYNCNT)
3256 inet_csk(sk)->icsk_syn_retries = val;
3260 EXPORT_SYMBOL(tcp_sock_set_syncnt);
3262 void tcp_sock_set_user_timeout(struct sock *sk, u32 val)
3265 inet_csk(sk)->icsk_user_timeout = val;
3268 EXPORT_SYMBOL(tcp_sock_set_user_timeout);
3270 int tcp_sock_set_keepidle_locked(struct sock *sk, int val)
3272 struct tcp_sock *tp = tcp_sk(sk);
3274 if (val < 1 || val > MAX_TCP_KEEPIDLE)
3277 tp->keepalive_time = val * HZ;
3278 if (sock_flag(sk, SOCK_KEEPOPEN) &&
3279 !((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN))) {
3280 u32 elapsed = keepalive_time_elapsed(tp);
3282 if (tp->keepalive_time > elapsed)
3283 elapsed = tp->keepalive_time - elapsed;
3286 inet_csk_reset_keepalive_timer(sk, elapsed);
3292 int tcp_sock_set_keepidle(struct sock *sk, int val)
3297 err = tcp_sock_set_keepidle_locked(sk, val);
3301 EXPORT_SYMBOL(tcp_sock_set_keepidle);
3303 int tcp_sock_set_keepintvl(struct sock *sk, int val)
3305 if (val < 1 || val > MAX_TCP_KEEPINTVL)
3309 tcp_sk(sk)->keepalive_intvl = val * HZ;
3313 EXPORT_SYMBOL(tcp_sock_set_keepintvl);
3315 int tcp_sock_set_keepcnt(struct sock *sk, int val)
3317 if (val < 1 || val > MAX_TCP_KEEPCNT)
3321 tcp_sk(sk)->keepalive_probes = val;
3325 EXPORT_SYMBOL(tcp_sock_set_keepcnt);
3327 int tcp_set_window_clamp(struct sock *sk, int val)
3329 struct tcp_sock *tp = tcp_sk(sk);
3332 if (sk->sk_state != TCP_CLOSE)
3334 tp->window_clamp = 0;
3336 tp->window_clamp = val < SOCK_MIN_RCVBUF / 2 ?
3337 SOCK_MIN_RCVBUF / 2 : val;
3343 * Socket option code for TCP.
3345 static int do_tcp_setsockopt(struct sock *sk, int level, int optname,
3346 sockptr_t optval, unsigned int optlen)
3348 struct tcp_sock *tp = tcp_sk(sk);
3349 struct inet_connection_sock *icsk = inet_csk(sk);
3350 struct net *net = sock_net(sk);
3354 /* These are data/string values, all the others are ints */
3356 case TCP_CONGESTION: {
3357 char name[TCP_CA_NAME_MAX];
3362 val = strncpy_from_sockptr(name, optval,
3363 min_t(long, TCP_CA_NAME_MAX-1, optlen));
3369 err = tcp_set_congestion_control(sk, name, true,
3370 ns_capable(sock_net(sk)->user_ns,
3376 char name[TCP_ULP_NAME_MAX];
3381 val = strncpy_from_sockptr(name, optval,
3382 min_t(long, TCP_ULP_NAME_MAX - 1,
3389 err = tcp_set_ulp(sk, name);
3393 case TCP_FASTOPEN_KEY: {
3394 __u8 key[TCP_FASTOPEN_KEY_BUF_LENGTH];
3395 __u8 *backup_key = NULL;
3397 /* Allow a backup key as well to facilitate key rotation
3398 * First key is the active one.
3400 if (optlen != TCP_FASTOPEN_KEY_LENGTH &&
3401 optlen != TCP_FASTOPEN_KEY_BUF_LENGTH)
3404 if (copy_from_sockptr(key, optval, optlen))
3407 if (optlen == TCP_FASTOPEN_KEY_BUF_LENGTH)
3408 backup_key = key + TCP_FASTOPEN_KEY_LENGTH;
3410 return tcp_fastopen_reset_cipher(net, sk, key, backup_key);
3417 if (optlen < sizeof(int))
3420 if (copy_from_sockptr(&val, optval, sizeof(val)))
3427 /* Values greater than interface MTU won't take effect. However
3428 * at the point when this call is done we typically don't yet
3429 * know which interface is going to be used
3431 if (val && (val < TCP_MIN_MSS || val > MAX_TCP_WINDOW)) {
3435 tp->rx_opt.user_mss = val;
3439 __tcp_sock_set_nodelay(sk, val);
3442 case TCP_THIN_LINEAR_TIMEOUTS:
3443 if (val < 0 || val > 1)
3449 case TCP_THIN_DUPACK:
3450 if (val < 0 || val > 1)
3455 if (!tcp_can_repair_sock(sk))
3457 else if (val == TCP_REPAIR_ON) {
3459 sk->sk_reuse = SK_FORCE_REUSE;
3460 tp->repair_queue = TCP_NO_QUEUE;
3461 } else if (val == TCP_REPAIR_OFF) {
3463 sk->sk_reuse = SK_NO_REUSE;
3464 tcp_send_window_probe(sk);
3465 } else if (val == TCP_REPAIR_OFF_NO_WP) {
3467 sk->sk_reuse = SK_NO_REUSE;
3473 case TCP_REPAIR_QUEUE:
3476 else if ((unsigned int)val < TCP_QUEUES_NR)
3477 tp->repair_queue = val;
3483 if (sk->sk_state != TCP_CLOSE) {
3485 } else if (tp->repair_queue == TCP_SEND_QUEUE) {
3486 if (!tcp_rtx_queue_empty(sk))
3489 WRITE_ONCE(tp->write_seq, val);
3490 } else if (tp->repair_queue == TCP_RECV_QUEUE) {
3491 if (tp->rcv_nxt != tp->copied_seq) {
3494 WRITE_ONCE(tp->rcv_nxt, val);
3495 WRITE_ONCE(tp->copied_seq, val);
3502 case TCP_REPAIR_OPTIONS:
3505 else if (sk->sk_state == TCP_ESTABLISHED)
3506 err = tcp_repair_options_est(sk, optval, optlen);
3512 __tcp_sock_set_cork(sk, val);
3516 err = tcp_sock_set_keepidle_locked(sk, val);
3519 if (val < 1 || val > MAX_TCP_KEEPINTVL)
3522 tp->keepalive_intvl = val * HZ;
3525 if (val < 1 || val > MAX_TCP_KEEPCNT)
3528 tp->keepalive_probes = val;
3531 if (val < 1 || val > MAX_TCP_SYNCNT)
3534 icsk->icsk_syn_retries = val;
3538 /* 0: disable, 1: enable, 2: start from ether_header */
3539 if (val < 0 || val > 2)
3548 else if (val > TCP_FIN_TIMEOUT_MAX / HZ)
3549 tp->linger2 = TCP_FIN_TIMEOUT_MAX;
3551 tp->linger2 = val * HZ;
3554 case TCP_DEFER_ACCEPT:
3555 /* Translate value in seconds to number of retransmits */
3556 icsk->icsk_accept_queue.rskq_defer_accept =
3557 secs_to_retrans(val, TCP_TIMEOUT_INIT / HZ,
3561 case TCP_WINDOW_CLAMP:
3562 err = tcp_set_window_clamp(sk, val);
3566 __tcp_sock_set_quickack(sk, val);
3569 #ifdef CONFIG_TCP_MD5SIG
3571 case TCP_MD5SIG_EXT:
3572 err = tp->af_specific->md5_parse(sk, optname, optval, optlen);
3575 case TCP_USER_TIMEOUT:
3576 /* Cap the max time in ms TCP will retry or probe the window
3577 * before giving up and aborting (ETIMEDOUT) a connection.
3582 icsk->icsk_user_timeout = val;
3586 if (val >= 0 && ((1 << sk->sk_state) & (TCPF_CLOSE |
3588 tcp_fastopen_init_key_once(net);
3590 fastopen_queue_tune(sk, val);
3595 case TCP_FASTOPEN_CONNECT:
3596 if (val > 1 || val < 0) {
3598 } else if (net->ipv4.sysctl_tcp_fastopen & TFO_CLIENT_ENABLE) {
3599 if (sk->sk_state == TCP_CLOSE)
3600 tp->fastopen_connect = val;
3607 case TCP_FASTOPEN_NO_COOKIE:
3608 if (val > 1 || val < 0)
3610 else if (!((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
3613 tp->fastopen_no_cookie = val;
3619 tp->tsoffset = val - tcp_time_stamp_raw();
3621 case TCP_REPAIR_WINDOW:
3622 err = tcp_repair_set_window(tp, optval, optlen);
3624 case TCP_NOTSENT_LOWAT:
3625 tp->notsent_lowat = val;
3626 sk->sk_write_space(sk);
3629 if (val > 1 || val < 0)
3632 tp->recvmsg_inq = val;
3636 tcp_enable_tx_delay();
3637 tp->tcp_tx_delay = val;
3648 int tcp_setsockopt(struct sock *sk, int level, int optname, sockptr_t optval,
3649 unsigned int optlen)
3651 const struct inet_connection_sock *icsk = inet_csk(sk);
3653 if (level != SOL_TCP)
3654 return icsk->icsk_af_ops->setsockopt(sk, level, optname,
3656 return do_tcp_setsockopt(sk, level, optname, optval, optlen);
3658 EXPORT_SYMBOL(tcp_setsockopt);
3660 static void tcp_get_info_chrono_stats(const struct tcp_sock *tp,
3661 struct tcp_info *info)
3663 u64 stats[__TCP_CHRONO_MAX], total = 0;
3666 for (i = TCP_CHRONO_BUSY; i < __TCP_CHRONO_MAX; ++i) {
3667 stats[i] = tp->chrono_stat[i - 1];
3668 if (i == tp->chrono_type)
3669 stats[i] += tcp_jiffies32 - tp->chrono_start;
3670 stats[i] *= USEC_PER_SEC / HZ;
3674 info->tcpi_busy_time = total;
3675 info->tcpi_rwnd_limited = stats[TCP_CHRONO_RWND_LIMITED];
3676 info->tcpi_sndbuf_limited = stats[TCP_CHRONO_SNDBUF_LIMITED];
3679 /* Return information about state of tcp endpoint in API format. */
3680 void tcp_get_info(struct sock *sk, struct tcp_info *info)
3682 const struct tcp_sock *tp = tcp_sk(sk); /* iff sk_type == SOCK_STREAM */
3683 const struct inet_connection_sock *icsk = inet_csk(sk);
3689 memset(info, 0, sizeof(*info));
3690 if (sk->sk_type != SOCK_STREAM)
3693 info->tcpi_state = inet_sk_state_load(sk);
3695 /* Report meaningful fields for all TCP states, including listeners */
3696 rate = READ_ONCE(sk->sk_pacing_rate);
3697 rate64 = (rate != ~0UL) ? rate : ~0ULL;
3698 info->tcpi_pacing_rate = rate64;
3700 rate = READ_ONCE(sk->sk_max_pacing_rate);
3701 rate64 = (rate != ~0UL) ? rate : ~0ULL;
3702 info->tcpi_max_pacing_rate = rate64;
3704 info->tcpi_reordering = tp->reordering;
3705 info->tcpi_snd_cwnd = tp->snd_cwnd;
3707 if (info->tcpi_state == TCP_LISTEN) {
3708 /* listeners aliased fields :
3709 * tcpi_unacked -> Number of children ready for accept()
3710 * tcpi_sacked -> max backlog
3712 info->tcpi_unacked = READ_ONCE(sk->sk_ack_backlog);
3713 info->tcpi_sacked = READ_ONCE(sk->sk_max_ack_backlog);
3717 slow = lock_sock_fast(sk);
3719 info->tcpi_ca_state = icsk->icsk_ca_state;
3720 info->tcpi_retransmits = icsk->icsk_retransmits;
3721 info->tcpi_probes = icsk->icsk_probes_out;
3722 info->tcpi_backoff = icsk->icsk_backoff;
3724 if (tp->rx_opt.tstamp_ok)
3725 info->tcpi_options |= TCPI_OPT_TIMESTAMPS;
3726 if (tcp_is_sack(tp))
3727 info->tcpi_options |= TCPI_OPT_SACK;
3728 if (tp->rx_opt.wscale_ok) {
3729 info->tcpi_options |= TCPI_OPT_WSCALE;
3730 info->tcpi_snd_wscale = tp->rx_opt.snd_wscale;
3731 info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale;
3734 if (tp->ecn_flags & TCP_ECN_OK)
3735 info->tcpi_options |= TCPI_OPT_ECN;
3736 if (tp->ecn_flags & TCP_ECN_SEEN)
3737 info->tcpi_options |= TCPI_OPT_ECN_SEEN;
3738 if (tp->syn_data_acked)
3739 info->tcpi_options |= TCPI_OPT_SYN_DATA;
3741 info->tcpi_rto = jiffies_to_usecs(icsk->icsk_rto);
3742 info->tcpi_ato = jiffies_to_usecs(icsk->icsk_ack.ato);
3743 info->tcpi_snd_mss = tp->mss_cache;
3744 info->tcpi_rcv_mss = icsk->icsk_ack.rcv_mss;
3746 info->tcpi_unacked = tp->packets_out;
3747 info->tcpi_sacked = tp->sacked_out;
3749 info->tcpi_lost = tp->lost_out;
3750 info->tcpi_retrans = tp->retrans_out;
3752 now = tcp_jiffies32;
3753 info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime);
3754 info->tcpi_last_data_recv = jiffies_to_msecs(now - icsk->icsk_ack.lrcvtime);
3755 info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp);
3757 info->tcpi_pmtu = icsk->icsk_pmtu_cookie;
3758 info->tcpi_rcv_ssthresh = tp->rcv_ssthresh;
3759 info->tcpi_rtt = tp->srtt_us >> 3;
3760 info->tcpi_rttvar = tp->mdev_us >> 2;
3761 info->tcpi_snd_ssthresh = tp->snd_ssthresh;
3762 info->tcpi_advmss = tp->advmss;
3764 info->tcpi_rcv_rtt = tp->rcv_rtt_est.rtt_us >> 3;
3765 info->tcpi_rcv_space = tp->rcvq_space.space;
3767 info->tcpi_total_retrans = tp->total_retrans;
3769 info->tcpi_bytes_acked = tp->bytes_acked;
3770 info->tcpi_bytes_received = tp->bytes_received;
3771 info->tcpi_notsent_bytes = max_t(int, 0, tp->write_seq - tp->snd_nxt);
3772 tcp_get_info_chrono_stats(tp, info);
3774 info->tcpi_segs_out = tp->segs_out;
3775 info->tcpi_segs_in = tp->segs_in;
3777 info->tcpi_min_rtt = tcp_min_rtt(tp);
3778 info->tcpi_data_segs_in = tp->data_segs_in;
3779 info->tcpi_data_segs_out = tp->data_segs_out;
3781 info->tcpi_delivery_rate_app_limited = tp->rate_app_limited ? 1 : 0;
3782 rate64 = tcp_compute_delivery_rate(tp);
3784 info->tcpi_delivery_rate = rate64;
3785 info->tcpi_delivered = tp->delivered;
3786 info->tcpi_delivered_ce = tp->delivered_ce;
3787 info->tcpi_bytes_sent = tp->bytes_sent;
3788 info->tcpi_bytes_retrans = tp->bytes_retrans;
3789 info->tcpi_dsack_dups = tp->dsack_dups;
3790 info->tcpi_reord_seen = tp->reord_seen;
3791 info->tcpi_rcv_ooopack = tp->rcv_ooopack;
3792 info->tcpi_snd_wnd = tp->snd_wnd;
3793 info->tcpi_fastopen_client_fail = tp->fastopen_client_fail;
3794 unlock_sock_fast(sk, slow);
3796 EXPORT_SYMBOL_GPL(tcp_get_info);
3798 static size_t tcp_opt_stats_get_size(void)
3801 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_BUSY */
3802 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_RWND_LIMITED */
3803 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_SNDBUF_LIMITED */
3804 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_DATA_SEGS_OUT */
3805 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_TOTAL_RETRANS */
3806 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_PACING_RATE */
3807 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_DELIVERY_RATE */
3808 nla_total_size(sizeof(u32)) + /* TCP_NLA_SND_CWND */
3809 nla_total_size(sizeof(u32)) + /* TCP_NLA_REORDERING */
3810 nla_total_size(sizeof(u32)) + /* TCP_NLA_MIN_RTT */
3811 nla_total_size(sizeof(u8)) + /* TCP_NLA_RECUR_RETRANS */
3812 nla_total_size(sizeof(u8)) + /* TCP_NLA_DELIVERY_RATE_APP_LMT */
3813 nla_total_size(sizeof(u32)) + /* TCP_NLA_SNDQ_SIZE */
3814 nla_total_size(sizeof(u8)) + /* TCP_NLA_CA_STATE */
3815 nla_total_size(sizeof(u32)) + /* TCP_NLA_SND_SSTHRESH */
3816 nla_total_size(sizeof(u32)) + /* TCP_NLA_DELIVERED */
3817 nla_total_size(sizeof(u32)) + /* TCP_NLA_DELIVERED_CE */
3818 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_BYTES_SENT */
3819 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_BYTES_RETRANS */
3820 nla_total_size(sizeof(u32)) + /* TCP_NLA_DSACK_DUPS */
3821 nla_total_size(sizeof(u32)) + /* TCP_NLA_REORD_SEEN */
3822 nla_total_size(sizeof(u32)) + /* TCP_NLA_SRTT */
3823 nla_total_size(sizeof(u16)) + /* TCP_NLA_TIMEOUT_REHASH */
3824 nla_total_size(sizeof(u32)) + /* TCP_NLA_BYTES_NOTSENT */
3825 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_EDT */
3826 nla_total_size(sizeof(u8)) + /* TCP_NLA_TTL */
3830 /* Returns TTL or hop limit of an incoming packet from skb. */
3831 static u8 tcp_skb_ttl_or_hop_limit(const struct sk_buff *skb)
3833 if (skb->protocol == htons(ETH_P_IP))
3834 return ip_hdr(skb)->ttl;
3835 else if (skb->protocol == htons(ETH_P_IPV6))
3836 return ipv6_hdr(skb)->hop_limit;
3841 struct sk_buff *tcp_get_timestamping_opt_stats(const struct sock *sk,
3842 const struct sk_buff *orig_skb,
3843 const struct sk_buff *ack_skb)
3845 const struct tcp_sock *tp = tcp_sk(sk);
3846 struct sk_buff *stats;
3847 struct tcp_info info;
3851 stats = alloc_skb(tcp_opt_stats_get_size(), GFP_ATOMIC);
3855 tcp_get_info_chrono_stats(tp, &info);
3856 nla_put_u64_64bit(stats, TCP_NLA_BUSY,
3857 info.tcpi_busy_time, TCP_NLA_PAD);
3858 nla_put_u64_64bit(stats, TCP_NLA_RWND_LIMITED,
3859 info.tcpi_rwnd_limited, TCP_NLA_PAD);
3860 nla_put_u64_64bit(stats, TCP_NLA_SNDBUF_LIMITED,
3861 info.tcpi_sndbuf_limited, TCP_NLA_PAD);
3862 nla_put_u64_64bit(stats, TCP_NLA_DATA_SEGS_OUT,
3863 tp->data_segs_out, TCP_NLA_PAD);
3864 nla_put_u64_64bit(stats, TCP_NLA_TOTAL_RETRANS,
3865 tp->total_retrans, TCP_NLA_PAD);
3867 rate = READ_ONCE(sk->sk_pacing_rate);
3868 rate64 = (rate != ~0UL) ? rate : ~0ULL;
3869 nla_put_u64_64bit(stats, TCP_NLA_PACING_RATE, rate64, TCP_NLA_PAD);
3871 rate64 = tcp_compute_delivery_rate(tp);
3872 nla_put_u64_64bit(stats, TCP_NLA_DELIVERY_RATE, rate64, TCP_NLA_PAD);
3874 nla_put_u32(stats, TCP_NLA_SND_CWND, tp->snd_cwnd);
3875 nla_put_u32(stats, TCP_NLA_REORDERING, tp->reordering);
3876 nla_put_u32(stats, TCP_NLA_MIN_RTT, tcp_min_rtt(tp));
3878 nla_put_u8(stats, TCP_NLA_RECUR_RETRANS, inet_csk(sk)->icsk_retransmits);
3879 nla_put_u8(stats, TCP_NLA_DELIVERY_RATE_APP_LMT, !!tp->rate_app_limited);
3880 nla_put_u32(stats, TCP_NLA_SND_SSTHRESH, tp->snd_ssthresh);
3881 nla_put_u32(stats, TCP_NLA_DELIVERED, tp->delivered);
3882 nla_put_u32(stats, TCP_NLA_DELIVERED_CE, tp->delivered_ce);
3884 nla_put_u32(stats, TCP_NLA_SNDQ_SIZE, tp->write_seq - tp->snd_una);
3885 nla_put_u8(stats, TCP_NLA_CA_STATE, inet_csk(sk)->icsk_ca_state);
3887 nla_put_u64_64bit(stats, TCP_NLA_BYTES_SENT, tp->bytes_sent,
3889 nla_put_u64_64bit(stats, TCP_NLA_BYTES_RETRANS, tp->bytes_retrans,
3891 nla_put_u32(stats, TCP_NLA_DSACK_DUPS, tp->dsack_dups);
3892 nla_put_u32(stats, TCP_NLA_REORD_SEEN, tp->reord_seen);
3893 nla_put_u32(stats, TCP_NLA_SRTT, tp->srtt_us >> 3);
3894 nla_put_u16(stats, TCP_NLA_TIMEOUT_REHASH, tp->timeout_rehash);
3895 nla_put_u32(stats, TCP_NLA_BYTES_NOTSENT,
3896 max_t(int, 0, tp->write_seq - tp->snd_nxt));
3897 nla_put_u64_64bit(stats, TCP_NLA_EDT, orig_skb->skb_mstamp_ns,
3900 nla_put_u8(stats, TCP_NLA_TTL,
3901 tcp_skb_ttl_or_hop_limit(ack_skb));
3906 static int do_tcp_getsockopt(struct sock *sk, int level,
3907 int optname, char __user *optval, int __user *optlen)
3909 struct inet_connection_sock *icsk = inet_csk(sk);
3910 struct tcp_sock *tp = tcp_sk(sk);
3911 struct net *net = sock_net(sk);
3914 if (get_user(len, optlen))
3917 len = min_t(unsigned int, len, sizeof(int));
3924 val = tp->mss_cache;
3925 if (!val && ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
3926 val = tp->rx_opt.user_mss;
3928 val = tp->rx_opt.mss_clamp;
3931 val = !!(tp->nonagle&TCP_NAGLE_OFF);
3934 val = !!(tp->nonagle&TCP_NAGLE_CORK);
3937 val = keepalive_time_when(tp) / HZ;
3940 val = keepalive_intvl_when(tp) / HZ;
3943 val = keepalive_probes(tp);
3946 val = icsk->icsk_syn_retries ? : net->ipv4.sysctl_tcp_syn_retries;
3951 val = (val ? : net->ipv4.sysctl_tcp_fin_timeout) / HZ;
3953 case TCP_DEFER_ACCEPT:
3954 val = retrans_to_secs(icsk->icsk_accept_queue.rskq_defer_accept,
3955 TCP_TIMEOUT_INIT / HZ, TCP_RTO_MAX / HZ);
3957 case TCP_WINDOW_CLAMP:
3958 val = tp->window_clamp;
3961 struct tcp_info info;
3963 if (get_user(len, optlen))
3966 tcp_get_info(sk, &info);
3968 len = min_t(unsigned int, len, sizeof(info));
3969 if (put_user(len, optlen))
3971 if (copy_to_user(optval, &info, len))
3976 const struct tcp_congestion_ops *ca_ops;
3977 union tcp_cc_info info;
3981 if (get_user(len, optlen))
3984 ca_ops = icsk->icsk_ca_ops;
3985 if (ca_ops && ca_ops->get_info)
3986 sz = ca_ops->get_info(sk, ~0U, &attr, &info);
3988 len = min_t(unsigned int, len, sz);
3989 if (put_user(len, optlen))
3991 if (copy_to_user(optval, &info, len))
3996 val = !inet_csk_in_pingpong_mode(sk);
3999 case TCP_CONGESTION:
4000 if (get_user(len, optlen))
4002 len = min_t(unsigned int, len, TCP_CA_NAME_MAX);
4003 if (put_user(len, optlen))
4005 if (copy_to_user(optval, icsk->icsk_ca_ops->name, len))
4010 if (get_user(len, optlen))
4012 len = min_t(unsigned int, len, TCP_ULP_NAME_MAX);
4013 if (!icsk->icsk_ulp_ops) {
4014 if (put_user(0, optlen))
4018 if (put_user(len, optlen))
4020 if (copy_to_user(optval, icsk->icsk_ulp_ops->name, len))
4024 case TCP_FASTOPEN_KEY: {
4025 u64 key[TCP_FASTOPEN_KEY_BUF_LENGTH / sizeof(u64)];
4026 unsigned int key_len;
4028 if (get_user(len, optlen))
4031 key_len = tcp_fastopen_get_cipher(net, icsk, key) *
4032 TCP_FASTOPEN_KEY_LENGTH;
4033 len = min_t(unsigned int, len, key_len);
4034 if (put_user(len, optlen))
4036 if (copy_to_user(optval, key, len))
4040 case TCP_THIN_LINEAR_TIMEOUTS:
4044 case TCP_THIN_DUPACK:
4052 case TCP_REPAIR_QUEUE:
4054 val = tp->repair_queue;
4059 case TCP_REPAIR_WINDOW: {
4060 struct tcp_repair_window opt;
4062 if (get_user(len, optlen))
4065 if (len != sizeof(opt))
4071 opt.snd_wl1 = tp->snd_wl1;
4072 opt.snd_wnd = tp->snd_wnd;
4073 opt.max_window = tp->max_window;
4074 opt.rcv_wnd = tp->rcv_wnd;
4075 opt.rcv_wup = tp->rcv_wup;
4077 if (copy_to_user(optval, &opt, len))
4082 if (tp->repair_queue == TCP_SEND_QUEUE)
4083 val = tp->write_seq;
4084 else if (tp->repair_queue == TCP_RECV_QUEUE)
4090 case TCP_USER_TIMEOUT:
4091 val = icsk->icsk_user_timeout;
4095 val = icsk->icsk_accept_queue.fastopenq.max_qlen;
4098 case TCP_FASTOPEN_CONNECT:
4099 val = tp->fastopen_connect;
4102 case TCP_FASTOPEN_NO_COOKIE:
4103 val = tp->fastopen_no_cookie;
4107 val = tp->tcp_tx_delay;
4111 val = tcp_time_stamp_raw() + tp->tsoffset;
4113 case TCP_NOTSENT_LOWAT:
4114 val = tp->notsent_lowat;
4117 val = tp->recvmsg_inq;
4122 case TCP_SAVED_SYN: {
4123 if (get_user(len, optlen))
4127 if (tp->saved_syn) {
4128 if (len < tcp_saved_syn_len(tp->saved_syn)) {
4129 if (put_user(tcp_saved_syn_len(tp->saved_syn),
4137 len = tcp_saved_syn_len(tp->saved_syn);
4138 if (put_user(len, optlen)) {
4142 if (copy_to_user(optval, tp->saved_syn->data, len)) {
4146 tcp_saved_syn_free(tp);
4151 if (put_user(len, optlen))
4157 case TCP_ZEROCOPY_RECEIVE: {
4158 struct scm_timestamping_internal tss;
4159 struct tcp_zerocopy_receive zc = {};
4162 if (get_user(len, optlen))
4165 len < offsetofend(struct tcp_zerocopy_receive, length))
4167 if (unlikely(len > sizeof(zc))) {
4168 err = check_zeroed_user(optval + sizeof(zc),
4171 return err == 0 ? -EINVAL : err;
4173 if (put_user(len, optlen))
4176 if (copy_from_user(&zc, optval, len))
4180 if (zc.msg_flags & ~(TCP_VALID_ZC_MSG_FLAGS))
4183 err = tcp_zerocopy_receive(sk, &zc, &tss);
4184 err = BPF_CGROUP_RUN_PROG_GETSOCKOPT_KERN(sk, level, optname,
4187 if (len >= offsetofend(struct tcp_zerocopy_receive, msg_flags))
4188 goto zerocopy_rcv_cmsg;
4190 case offsetofend(struct tcp_zerocopy_receive, msg_flags):
4191 goto zerocopy_rcv_cmsg;
4192 case offsetofend(struct tcp_zerocopy_receive, msg_controllen):
4193 case offsetofend(struct tcp_zerocopy_receive, msg_control):
4194 case offsetofend(struct tcp_zerocopy_receive, flags):
4195 case offsetofend(struct tcp_zerocopy_receive, copybuf_len):
4196 case offsetofend(struct tcp_zerocopy_receive, copybuf_address):
4197 case offsetofend(struct tcp_zerocopy_receive, err):
4198 goto zerocopy_rcv_sk_err;
4199 case offsetofend(struct tcp_zerocopy_receive, inq):
4200 goto zerocopy_rcv_inq;
4201 case offsetofend(struct tcp_zerocopy_receive, length):
4203 goto zerocopy_rcv_out;
4206 if (zc.msg_flags & TCP_CMSG_TS)
4207 tcp_zc_finalize_rx_tstamp(sk, &zc, &tss);
4210 zerocopy_rcv_sk_err:
4212 zc.err = sock_error(sk);
4214 zc.inq = tcp_inq_hint(sk);
4216 if (!err && copy_to_user(optval, &zc, len))
4222 return -ENOPROTOOPT;
4225 if (put_user(len, optlen))
4227 if (copy_to_user(optval, &val, len))
4232 bool tcp_bpf_bypass_getsockopt(int level, int optname)
4234 /* TCP do_tcp_getsockopt has optimized getsockopt implementation
4235 * to avoid extra socket lock for TCP_ZEROCOPY_RECEIVE.
4237 if (level == SOL_TCP && optname == TCP_ZEROCOPY_RECEIVE)
4242 EXPORT_SYMBOL(tcp_bpf_bypass_getsockopt);
4244 int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval,
4247 struct inet_connection_sock *icsk = inet_csk(sk);
4249 if (level != SOL_TCP)
4250 return icsk->icsk_af_ops->getsockopt(sk, level, optname,
4252 return do_tcp_getsockopt(sk, level, optname, optval, optlen);
4254 EXPORT_SYMBOL(tcp_getsockopt);
4256 #ifdef CONFIG_TCP_MD5SIG
4257 static DEFINE_PER_CPU(struct tcp_md5sig_pool, tcp_md5sig_pool);
4258 static DEFINE_MUTEX(tcp_md5sig_mutex);
4259 static bool tcp_md5sig_pool_populated = false;
4261 static void __tcp_alloc_md5sig_pool(void)
4263 struct crypto_ahash *hash;
4266 hash = crypto_alloc_ahash("md5", 0, CRYPTO_ALG_ASYNC);
4270 for_each_possible_cpu(cpu) {
4271 void *scratch = per_cpu(tcp_md5sig_pool, cpu).scratch;
4272 struct ahash_request *req;
4275 scratch = kmalloc_node(sizeof(union tcp_md5sum_block) +
4276 sizeof(struct tcphdr),
4281 per_cpu(tcp_md5sig_pool, cpu).scratch = scratch;
4283 if (per_cpu(tcp_md5sig_pool, cpu).md5_req)
4286 req = ahash_request_alloc(hash, GFP_KERNEL);
4290 ahash_request_set_callback(req, 0, NULL, NULL);
4292 per_cpu(tcp_md5sig_pool, cpu).md5_req = req;
4294 /* before setting tcp_md5sig_pool_populated, we must commit all writes
4295 * to memory. See smp_rmb() in tcp_get_md5sig_pool()
4298 tcp_md5sig_pool_populated = true;
4301 bool tcp_alloc_md5sig_pool(void)
4303 if (unlikely(!tcp_md5sig_pool_populated)) {
4304 mutex_lock(&tcp_md5sig_mutex);
4306 if (!tcp_md5sig_pool_populated) {
4307 __tcp_alloc_md5sig_pool();
4308 if (tcp_md5sig_pool_populated)
4309 static_branch_inc(&tcp_md5_needed);
4312 mutex_unlock(&tcp_md5sig_mutex);
4314 return tcp_md5sig_pool_populated;
4316 EXPORT_SYMBOL(tcp_alloc_md5sig_pool);
4320 * tcp_get_md5sig_pool - get md5sig_pool for this user
4322 * We use percpu structure, so if we succeed, we exit with preemption
4323 * and BH disabled, to make sure another thread or softirq handling
4324 * wont try to get same context.
4326 struct tcp_md5sig_pool *tcp_get_md5sig_pool(void)
4330 if (tcp_md5sig_pool_populated) {
4331 /* coupled with smp_wmb() in __tcp_alloc_md5sig_pool() */
4333 return this_cpu_ptr(&tcp_md5sig_pool);
4338 EXPORT_SYMBOL(tcp_get_md5sig_pool);
4340 int tcp_md5_hash_skb_data(struct tcp_md5sig_pool *hp,
4341 const struct sk_buff *skb, unsigned int header_len)
4343 struct scatterlist sg;
4344 const struct tcphdr *tp = tcp_hdr(skb);
4345 struct ahash_request *req = hp->md5_req;
4347 const unsigned int head_data_len = skb_headlen(skb) > header_len ?
4348 skb_headlen(skb) - header_len : 0;
4349 const struct skb_shared_info *shi = skb_shinfo(skb);
4350 struct sk_buff *frag_iter;
4352 sg_init_table(&sg, 1);
4354 sg_set_buf(&sg, ((u8 *) tp) + header_len, head_data_len);
4355 ahash_request_set_crypt(req, &sg, NULL, head_data_len);
4356 if (crypto_ahash_update(req))
4359 for (i = 0; i < shi->nr_frags; ++i) {
4360 const skb_frag_t *f = &shi->frags[i];
4361 unsigned int offset = skb_frag_off(f);
4362 struct page *page = skb_frag_page(f) + (offset >> PAGE_SHIFT);
4364 sg_set_page(&sg, page, skb_frag_size(f),
4365 offset_in_page(offset));
4366 ahash_request_set_crypt(req, &sg, NULL, skb_frag_size(f));
4367 if (crypto_ahash_update(req))
4371 skb_walk_frags(skb, frag_iter)
4372 if (tcp_md5_hash_skb_data(hp, frag_iter, 0))
4377 EXPORT_SYMBOL(tcp_md5_hash_skb_data);
4379 int tcp_md5_hash_key(struct tcp_md5sig_pool *hp, const struct tcp_md5sig_key *key)
4381 u8 keylen = READ_ONCE(key->keylen); /* paired with WRITE_ONCE() in tcp_md5_do_add */
4382 struct scatterlist sg;
4384 sg_init_one(&sg, key->key, keylen);
4385 ahash_request_set_crypt(hp->md5_req, &sg, NULL, keylen);
4387 /* We use data_race() because tcp_md5_do_add() might change key->key under us */
4388 return data_race(crypto_ahash_update(hp->md5_req));
4390 EXPORT_SYMBOL(tcp_md5_hash_key);
4394 void tcp_done(struct sock *sk)
4396 struct request_sock *req;
4398 /* We might be called with a new socket, after
4399 * inet_csk_prepare_forced_close() has been called
4400 * so we can not use lockdep_sock_is_held(sk)
4402 req = rcu_dereference_protected(tcp_sk(sk)->fastopen_rsk, 1);
4404 if (sk->sk_state == TCP_SYN_SENT || sk->sk_state == TCP_SYN_RECV)
4405 TCP_INC_STATS(sock_net(sk), TCP_MIB_ATTEMPTFAILS);
4407 tcp_set_state(sk, TCP_CLOSE);
4408 tcp_clear_xmit_timers(sk);
4410 reqsk_fastopen_remove(sk, req, false);
4412 sk->sk_shutdown = SHUTDOWN_MASK;
4414 if (!sock_flag(sk, SOCK_DEAD))
4415 sk->sk_state_change(sk);
4417 inet_csk_destroy_sock(sk);
4419 EXPORT_SYMBOL_GPL(tcp_done);
4421 int tcp_abort(struct sock *sk, int err)
4423 if (!sk_fullsock(sk)) {
4424 if (sk->sk_state == TCP_NEW_SYN_RECV) {
4425 struct request_sock *req = inet_reqsk(sk);
4428 inet_csk_reqsk_queue_drop(req->rsk_listener, req);
4435 /* Don't race with userspace socket closes such as tcp_close. */
4438 if (sk->sk_state == TCP_LISTEN) {
4439 tcp_set_state(sk, TCP_CLOSE);
4440 inet_csk_listen_stop(sk);
4443 /* Don't race with BH socket closes such as inet_csk_listen_stop. */
4447 if (!sock_flag(sk, SOCK_DEAD)) {
4449 /* This barrier is coupled with smp_rmb() in tcp_poll() */
4451 sk_error_report(sk);
4452 if (tcp_need_reset(sk->sk_state))
4453 tcp_send_active_reset(sk, GFP_ATOMIC);
4459 tcp_write_queue_purge(sk);
4463 EXPORT_SYMBOL_GPL(tcp_abort);
4465 extern struct tcp_congestion_ops tcp_reno;
4467 static __initdata unsigned long thash_entries;
4468 static int __init set_thash_entries(char *str)
4475 ret = kstrtoul(str, 0, &thash_entries);
4481 __setup("thash_entries=", set_thash_entries);
4483 static void __init tcp_init_mem(void)
4485 unsigned long limit = nr_free_buffer_pages() / 16;
4487 limit = max(limit, 128UL);
4488 sysctl_tcp_mem[0] = limit / 4 * 3; /* 4.68 % */
4489 sysctl_tcp_mem[1] = limit; /* 6.25 % */
4490 sysctl_tcp_mem[2] = sysctl_tcp_mem[0] * 2; /* 9.37 % */
4493 void __init tcp_init(void)
4495 int max_rshare, max_wshare, cnt;
4496 unsigned long limit;
4499 BUILD_BUG_ON(TCP_MIN_SND_MSS <= MAX_TCP_OPTION_SPACE);
4500 BUILD_BUG_ON(sizeof(struct tcp_skb_cb) >
4501 sizeof_field(struct sk_buff, cb));
4503 percpu_counter_init(&tcp_sockets_allocated, 0, GFP_KERNEL);
4504 percpu_counter_init(&tcp_orphan_count, 0, GFP_KERNEL);
4505 inet_hashinfo_init(&tcp_hashinfo);
4506 inet_hashinfo2_init(&tcp_hashinfo, "tcp_listen_portaddr_hash",
4507 thash_entries, 21, /* one slot per 2 MB*/
4509 tcp_hashinfo.bind_bucket_cachep =
4510 kmem_cache_create("tcp_bind_bucket",
4511 sizeof(struct inet_bind_bucket), 0,
4512 SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL);
4514 /* Size and allocate the main established and bind bucket
4517 * The methodology is similar to that of the buffer cache.
4519 tcp_hashinfo.ehash =
4520 alloc_large_system_hash("TCP established",
4521 sizeof(struct inet_ehash_bucket),
4523 17, /* one slot per 128 KB of memory */
4526 &tcp_hashinfo.ehash_mask,
4528 thash_entries ? 0 : 512 * 1024);
4529 for (i = 0; i <= tcp_hashinfo.ehash_mask; i++)
4530 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].chain, i);
4532 if (inet_ehash_locks_alloc(&tcp_hashinfo))
4533 panic("TCP: failed to alloc ehash_locks");
4534 tcp_hashinfo.bhash =
4535 alloc_large_system_hash("TCP bind",
4536 sizeof(struct inet_bind_hashbucket),
4537 tcp_hashinfo.ehash_mask + 1,
4538 17, /* one slot per 128 KB of memory */
4540 &tcp_hashinfo.bhash_size,
4544 tcp_hashinfo.bhash_size = 1U << tcp_hashinfo.bhash_size;
4545 for (i = 0; i < tcp_hashinfo.bhash_size; i++) {
4546 spin_lock_init(&tcp_hashinfo.bhash[i].lock);
4547 INIT_HLIST_HEAD(&tcp_hashinfo.bhash[i].chain);
4551 cnt = tcp_hashinfo.ehash_mask + 1;
4552 sysctl_tcp_max_orphans = cnt / 2;
4555 /* Set per-socket limits to no more than 1/128 the pressure threshold */
4556 limit = nr_free_buffer_pages() << (PAGE_SHIFT - 7);
4557 max_wshare = min(4UL*1024*1024, limit);
4558 max_rshare = min(6UL*1024*1024, limit);
4560 init_net.ipv4.sysctl_tcp_wmem[0] = SK_MEM_QUANTUM;
4561 init_net.ipv4.sysctl_tcp_wmem[1] = 16*1024;
4562 init_net.ipv4.sysctl_tcp_wmem[2] = max(64*1024, max_wshare);
4564 init_net.ipv4.sysctl_tcp_rmem[0] = SK_MEM_QUANTUM;
4565 init_net.ipv4.sysctl_tcp_rmem[1] = 131072;
4566 init_net.ipv4.sysctl_tcp_rmem[2] = max(131072, max_rshare);
4568 pr_info("Hash tables configured (established %u bind %u)\n",
4569 tcp_hashinfo.ehash_mask + 1, tcp_hashinfo.bhash_size);
4573 BUG_ON(tcp_register_congestion_control(&tcp_reno) != 0);