1 // SPDX-License-Identifier: GPL-2.0-or-later
3 * INET An implementation of the TCP/IP protocol suite for the LINUX
4 * operating system. INET is implemented using the BSD Socket
5 * interface as the means of communication with the user level.
7 * Implementation of the Transmission Control Protocol(TCP).
10 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
11 * Mark Evans, <evansmp@uhura.aston.ac.uk>
12 * Corey Minyard <wf-rch!minyard@relay.EU.net>
13 * Florian La Roche, <flla@stud.uni-sb.de>
14 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
15 * Linus Torvalds, <torvalds@cs.helsinki.fi>
16 * Alan Cox, <gw4pts@gw4pts.ampr.org>
17 * Matthew Dillon, <dillon@apollo.west.oic.com>
18 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
19 * Jorge Cwik, <jorge@laser.satlink.net>
22 * Alan Cox : Numerous verify_area() calls
23 * Alan Cox : Set the ACK bit on a reset
24 * Alan Cox : Stopped it crashing if it closed while
25 * sk->inuse=1 and was trying to connect
27 * Alan Cox : All icmp error handling was broken
28 * pointers passed where wrong and the
29 * socket was looked up backwards. Nobody
30 * tested any icmp error code obviously.
31 * Alan Cox : tcp_err() now handled properly. It
32 * wakes people on errors. poll
33 * behaves and the icmp error race
34 * has gone by moving it into sock.c
35 * Alan Cox : tcp_send_reset() fixed to work for
36 * everything not just packets for
38 * Alan Cox : tcp option processing.
39 * Alan Cox : Reset tweaked (still not 100%) [Had
41 * Herp Rosmanith : More reset fixes
42 * Alan Cox : No longer acks invalid rst frames.
43 * Acking any kind of RST is right out.
44 * Alan Cox : Sets an ignore me flag on an rst
45 * receive otherwise odd bits of prattle
47 * Alan Cox : Fixed another acking RST frame bug.
48 * Should stop LAN workplace lockups.
49 * Alan Cox : Some tidyups using the new skb list
51 * Alan Cox : sk->keepopen now seems to work
52 * Alan Cox : Pulls options out correctly on accepts
53 * Alan Cox : Fixed assorted sk->rqueue->next errors
54 * Alan Cox : PSH doesn't end a TCP read. Switched a
56 * Alan Cox : Tidied tcp_data to avoid a potential
58 * Alan Cox : Added some better commenting, as the
59 * tcp is hard to follow
60 * Alan Cox : Removed incorrect check for 20 * psh
61 * Michael O'Reilly : ack < copied bug fix.
62 * Johannes Stille : Misc tcp fixes (not all in yet).
63 * Alan Cox : FIN with no memory -> CRASH
64 * Alan Cox : Added socket option proto entries.
65 * Also added awareness of them to accept.
66 * Alan Cox : Added TCP options (SOL_TCP)
67 * Alan Cox : Switched wakeup calls to callbacks,
68 * so the kernel can layer network
70 * Alan Cox : Use ip_tos/ip_ttl settings.
71 * Alan Cox : Handle FIN (more) properly (we hope).
72 * Alan Cox : RST frames sent on unsynchronised
74 * Alan Cox : Put in missing check for SYN bit.
75 * Alan Cox : Added tcp_select_window() aka NET2E
76 * window non shrink trick.
77 * Alan Cox : Added a couple of small NET2E timer
79 * Charles Hedrick : TCP fixes
80 * Toomas Tamm : TCP window fixes
81 * Alan Cox : Small URG fix to rlogin ^C ack fight
82 * Charles Hedrick : Rewrote most of it to actually work
83 * Linus : Rewrote tcp_read() and URG handling
85 * Gerhard Koerting: Fixed some missing timer handling
86 * Matthew Dillon : Reworked TCP machine states as per RFC
87 * Gerhard Koerting: PC/TCP workarounds
88 * Adam Caldwell : Assorted timer/timing errors
89 * Matthew Dillon : Fixed another RST bug
90 * Alan Cox : Move to kernel side addressing changes.
91 * Alan Cox : Beginning work on TCP fastpathing
93 * Arnt Gulbrandsen: Turbocharged tcp_check() routine.
94 * Alan Cox : TCP fast path debugging
95 * Alan Cox : Window clamping
96 * Michael Riepe : Bug in tcp_check()
97 * Matt Dillon : More TCP improvements and RST bug fixes
98 * Matt Dillon : Yet more small nasties remove from the
99 * TCP code (Be very nice to this man if
100 * tcp finally works 100%) 8)
101 * Alan Cox : BSD accept semantics.
102 * Alan Cox : Reset on closedown bug.
103 * Peter De Schrijver : ENOTCONN check missing in tcp_sendto().
104 * Michael Pall : Handle poll() after URG properly in
106 * Michael Pall : Undo the last fix in tcp_read_urg()
107 * (multi URG PUSH broke rlogin).
108 * Michael Pall : Fix the multi URG PUSH problem in
109 * tcp_readable(), poll() after URG
111 * Michael Pall : recv(...,MSG_OOB) never blocks in the
113 * Alan Cox : Changed the semantics of sk->socket to
114 * fix a race and a signal problem with
115 * accept() and async I/O.
116 * Alan Cox : Relaxed the rules on tcp_sendto().
117 * Yury Shevchuk : Really fixed accept() blocking problem.
118 * Craig I. Hagan : Allow for BSD compatible TIME_WAIT for
119 * clients/servers which listen in on
121 * Alan Cox : Cleaned the above up and shrank it to
122 * a sensible code size.
123 * Alan Cox : Self connect lockup fix.
124 * Alan Cox : No connect to multicast.
125 * Ross Biro : Close unaccepted children on master
127 * Alan Cox : Reset tracing code.
128 * Alan Cox : Spurious resets on shutdown.
129 * Alan Cox : Giant 15 minute/60 second timer error
130 * Alan Cox : Small whoops in polling before an
132 * Alan Cox : Kept the state trace facility since
133 * it's handy for debugging.
134 * Alan Cox : More reset handler fixes.
135 * Alan Cox : Started rewriting the code based on
136 * the RFC's for other useful protocol
137 * references see: Comer, KA9Q NOS, and
138 * for a reference on the difference
139 * between specifications and how BSD
140 * works see the 4.4lite source.
141 * A.N.Kuznetsov : Don't time wait on completion of tidy
143 * Linus Torvalds : Fin/Shutdown & copied_seq changes.
144 * Linus Torvalds : Fixed BSD port reuse to work first syn
145 * Alan Cox : Reimplemented timers as per the RFC
146 * and using multiple timers for sanity.
147 * Alan Cox : Small bug fixes, and a lot of new
149 * Alan Cox : Fixed dual reader crash by locking
150 * the buffers (much like datagram.c)
151 * Alan Cox : Fixed stuck sockets in probe. A probe
152 * now gets fed up of retrying without
153 * (even a no space) answer.
154 * Alan Cox : Extracted closing code better
155 * Alan Cox : Fixed the closing state machine to
157 * Alan Cox : More 'per spec' fixes.
158 * Jorge Cwik : Even faster checksumming.
159 * Alan Cox : tcp_data() doesn't ack illegal PSH
160 * only frames. At least one pc tcp stack
162 * Alan Cox : Cache last socket.
163 * Alan Cox : Per route irtt.
164 * Matt Day : poll()->select() match BSD precisely on error
165 * Alan Cox : New buffers
166 * Marc Tamsky : Various sk->prot->retransmits and
167 * sk->retransmits misupdating fixed.
168 * Fixed tcp_write_timeout: stuck close,
169 * and TCP syn retries gets used now.
170 * Mark Yarvis : In tcp_read_wakeup(), don't send an
171 * ack if state is TCP_CLOSED.
172 * Alan Cox : Look up device on a retransmit - routes may
173 * change. Doesn't yet cope with MSS shrink right
175 * Marc Tamsky : Closing in closing fixes.
176 * Mike Shaver : RFC1122 verifications.
177 * Alan Cox : rcv_saddr errors.
178 * Alan Cox : Block double connect().
179 * Alan Cox : Small hooks for enSKIP.
180 * Alexey Kuznetsov: Path MTU discovery.
181 * Alan Cox : Support soft errors.
182 * Alan Cox : Fix MTU discovery pathological case
183 * when the remote claims no mtu!
184 * Marc Tamsky : TCP_CLOSE fix.
185 * Colin (G3TNE) : Send a reset on syn ack replies in
186 * window but wrong (fixes NT lpd problems)
187 * Pedro Roque : Better TCP window handling, delayed ack.
188 * Joerg Reuter : No modification of locked buffers in
189 * tcp_do_retransmit()
190 * Eric Schenk : Changed receiver side silly window
191 * avoidance algorithm to BSD style
192 * algorithm. This doubles throughput
193 * against machines running Solaris,
194 * and seems to result in general
196 * Stefan Magdalinski : adjusted tcp_readable() to fix FIONREAD
197 * Willy Konynenberg : Transparent proxying support.
198 * Mike McLagan : Routing by source
199 * Keith Owens : Do proper merging with partial SKB's in
200 * tcp_do_sendmsg to avoid burstiness.
201 * Eric Schenk : Fix fast close down bug with
202 * shutdown() followed by close().
203 * Andi Kleen : Make poll agree with SIGIO
204 * Salvatore Sanfilippo : Support SO_LINGER with linger == 1 and
205 * lingertime == 0 (RFC 793 ABORT Call)
206 * Hirokazu Takahashi : Use copy_from_user() instead of
207 * csum_and_copy_from_user() if possible.
209 * Description of States:
211 * TCP_SYN_SENT sent a connection request, waiting for ack
213 * TCP_SYN_RECV received a connection request, sent ack,
214 * waiting for final ack in three-way handshake.
216 * TCP_ESTABLISHED connection established
218 * TCP_FIN_WAIT1 our side has shutdown, waiting to complete
219 * transmission of remaining buffered data
221 * TCP_FIN_WAIT2 all buffered data sent, waiting for remote
224 * TCP_CLOSING both sides have shutdown but we still have
225 * data we have to finish sending
227 * TCP_TIME_WAIT timeout to catch resent junk before entering
228 * closed, can only be entered from FIN_WAIT2
229 * or CLOSING. Required because the other end
230 * may not have gotten our last ACK causing it
231 * to retransmit the data packet (which we ignore)
233 * TCP_CLOSE_WAIT remote side has shutdown and is waiting for
234 * us to finish writing our data and to shutdown
235 * (we have to close() to move on to LAST_ACK)
237 * TCP_LAST_ACK out side has shutdown after remote has
238 * shutdown. There may still be data in our
239 * buffer that we have to finish sending
241 * TCP_CLOSE socket is finished
244 #define pr_fmt(fmt) "TCP: " fmt
246 #include <crypto/hash.h>
247 #include <linux/kernel.h>
248 #include <linux/module.h>
249 #include <linux/types.h>
250 #include <linux/fcntl.h>
251 #include <linux/poll.h>
252 #include <linux/inet_diag.h>
253 #include <linux/init.h>
254 #include <linux/fs.h>
255 #include <linux/skbuff.h>
256 #include <linux/scatterlist.h>
257 #include <linux/splice.h>
258 #include <linux/net.h>
259 #include <linux/socket.h>
260 #include <linux/random.h>
261 #include <linux/memblock.h>
262 #include <linux/highmem.h>
263 #include <linux/cache.h>
264 #include <linux/err.h>
265 #include <linux/time.h>
266 #include <linux/slab.h>
267 #include <linux/errqueue.h>
268 #include <linux/static_key.h>
269 #include <linux/btf.h>
271 #include <net/icmp.h>
272 #include <net/inet_common.h>
274 #include <net/mptcp.h>
275 #include <net/xfrm.h>
277 #include <net/sock.h>
279 #include <linux/uaccess.h>
280 #include <asm/ioctls.h>
281 #include <net/busy_poll.h>
283 /* Track pending CMSGs. */
289 DEFINE_PER_CPU(unsigned int, tcp_orphan_count);
290 EXPORT_PER_CPU_SYMBOL_GPL(tcp_orphan_count);
292 long sysctl_tcp_mem[3] __read_mostly;
293 EXPORT_SYMBOL(sysctl_tcp_mem);
295 atomic_long_t tcp_memory_allocated ____cacheline_aligned_in_smp; /* Current allocated memory. */
296 EXPORT_SYMBOL(tcp_memory_allocated);
297 DEFINE_PER_CPU(int, tcp_memory_per_cpu_fw_alloc);
298 EXPORT_PER_CPU_SYMBOL_GPL(tcp_memory_per_cpu_fw_alloc);
300 #if IS_ENABLED(CONFIG_SMC)
301 DEFINE_STATIC_KEY_FALSE(tcp_have_smc);
302 EXPORT_SYMBOL(tcp_have_smc);
306 * Current number of TCP sockets.
308 struct percpu_counter tcp_sockets_allocated ____cacheline_aligned_in_smp;
309 EXPORT_SYMBOL(tcp_sockets_allocated);
314 struct tcp_splice_state {
315 struct pipe_inode_info *pipe;
321 * Pressure flag: try to collapse.
322 * Technical note: it is used by multiple contexts non atomically.
323 * All the __sk_mem_schedule() is of this nature: accounting
324 * is strict, actions are advisory and have some latency.
326 unsigned long tcp_memory_pressure __read_mostly;
327 EXPORT_SYMBOL_GPL(tcp_memory_pressure);
329 void tcp_enter_memory_pressure(struct sock *sk)
333 if (READ_ONCE(tcp_memory_pressure))
339 if (!cmpxchg(&tcp_memory_pressure, 0, val))
340 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMEMORYPRESSURES);
342 EXPORT_SYMBOL_GPL(tcp_enter_memory_pressure);
344 void tcp_leave_memory_pressure(struct sock *sk)
348 if (!READ_ONCE(tcp_memory_pressure))
350 val = xchg(&tcp_memory_pressure, 0);
352 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPMEMORYPRESSURESCHRONO,
353 jiffies_to_msecs(jiffies - val));
355 EXPORT_SYMBOL_GPL(tcp_leave_memory_pressure);
357 /* Convert seconds to retransmits based on initial and max timeout */
358 static u8 secs_to_retrans(int seconds, int timeout, int rto_max)
363 int period = timeout;
366 while (seconds > period && res < 255) {
369 if (timeout > rto_max)
377 /* Convert retransmits to seconds based on initial and max timeout */
378 static int retrans_to_secs(u8 retrans, int timeout, int rto_max)
386 if (timeout > rto_max)
394 static u64 tcp_compute_delivery_rate(const struct tcp_sock *tp)
396 u32 rate = READ_ONCE(tp->rate_delivered);
397 u32 intv = READ_ONCE(tp->rate_interval_us);
401 rate64 = (u64)rate * tp->mss_cache * USEC_PER_SEC;
402 do_div(rate64, intv);
407 /* Address-family independent initialization for a tcp_sock.
409 * NOTE: A lot of things set to zero explicitly by call to
410 * sk_alloc() so need not be done here.
412 void tcp_init_sock(struct sock *sk)
414 struct inet_connection_sock *icsk = inet_csk(sk);
415 struct tcp_sock *tp = tcp_sk(sk);
417 tp->out_of_order_queue = RB_ROOT;
418 sk->tcp_rtx_queue = RB_ROOT;
419 tcp_init_xmit_timers(sk);
420 INIT_LIST_HEAD(&tp->tsq_node);
421 INIT_LIST_HEAD(&tp->tsorted_sent_queue);
423 icsk->icsk_rto = TCP_TIMEOUT_INIT;
424 icsk->icsk_rto_min = TCP_RTO_MIN;
425 icsk->icsk_delack_max = TCP_DELACK_MAX;
426 tp->mdev_us = jiffies_to_usecs(TCP_TIMEOUT_INIT);
427 minmax_reset(&tp->rtt_min, tcp_jiffies32, ~0U);
429 /* So many TCP implementations out there (incorrectly) count the
430 * initial SYN frame in their delayed-ACK and congestion control
431 * algorithms that we must have the following bandaid to talk
432 * efficiently to them. -DaveM
434 tcp_snd_cwnd_set(tp, TCP_INIT_CWND);
436 /* There's a bubble in the pipe until at least the first ACK. */
437 tp->app_limited = ~0U;
438 tp->rate_app_limited = 1;
440 /* See draft-stevens-tcpca-spec-01 for discussion of the
441 * initialization of these values.
443 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
444 tp->snd_cwnd_clamp = ~0;
445 tp->mss_cache = TCP_MSS_DEFAULT;
447 tp->reordering = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_reordering);
448 tcp_assign_congestion_control(sk);
451 tp->rack.reo_wnd_steps = 1;
453 sk->sk_write_space = sk_stream_write_space;
454 sock_set_flag(sk, SOCK_USE_WRITE_QUEUE);
456 icsk->icsk_sync_mss = tcp_sync_mss;
458 WRITE_ONCE(sk->sk_sndbuf, READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_wmem[1]));
459 WRITE_ONCE(sk->sk_rcvbuf, READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_rmem[1]));
460 tcp_scaling_ratio_init(sk);
462 set_bit(SOCK_SUPPORT_ZC, &sk->sk_socket->flags);
463 sk_sockets_allocated_inc(sk);
465 EXPORT_SYMBOL(tcp_init_sock);
467 static void tcp_tx_timestamp(struct sock *sk, u16 tsflags)
469 struct sk_buff *skb = tcp_write_queue_tail(sk);
471 if (tsflags && skb) {
472 struct skb_shared_info *shinfo = skb_shinfo(skb);
473 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
475 sock_tx_timestamp(sk, tsflags, &shinfo->tx_flags);
476 if (tsflags & SOF_TIMESTAMPING_TX_ACK)
477 tcb->txstamp_ack = 1;
478 if (tsflags & SOF_TIMESTAMPING_TX_RECORD_MASK)
479 shinfo->tskey = TCP_SKB_CB(skb)->seq + skb->len - 1;
483 static bool tcp_stream_is_readable(struct sock *sk, int target)
485 if (tcp_epollin_ready(sk, target))
487 return sk_is_readable(sk);
491 * Wait for a TCP event.
493 * Note that we don't need to lock the socket, as the upper poll layers
494 * take care of normal races (between the test and the event) and we don't
495 * go look at any of the socket buffers directly.
497 __poll_t tcp_poll(struct file *file, struct socket *sock, poll_table *wait)
500 struct sock *sk = sock->sk;
501 const struct tcp_sock *tp = tcp_sk(sk);
505 sock_poll_wait(file, sock, wait);
507 state = inet_sk_state_load(sk);
508 if (state == TCP_LISTEN)
509 return inet_csk_listen_poll(sk);
511 /* Socket is not locked. We are protected from async events
512 * by poll logic and correct handling of state changes
513 * made by other threads is impossible in any case.
519 * EPOLLHUP is certainly not done right. But poll() doesn't
520 * have a notion of HUP in just one direction, and for a
521 * socket the read side is more interesting.
523 * Some poll() documentation says that EPOLLHUP is incompatible
524 * with the EPOLLOUT/POLLWR flags, so somebody should check this
525 * all. But careful, it tends to be safer to return too many
526 * bits than too few, and you can easily break real applications
527 * if you don't tell them that something has hung up!
531 * Check number 1. EPOLLHUP is _UNMASKABLE_ event (see UNIX98 and
532 * our fs/select.c). It means that after we received EOF,
533 * poll always returns immediately, making impossible poll() on write()
534 * in state CLOSE_WAIT. One solution is evident --- to set EPOLLHUP
535 * if and only if shutdown has been made in both directions.
536 * Actually, it is interesting to look how Solaris and DUX
537 * solve this dilemma. I would prefer, if EPOLLHUP were maskable,
538 * then we could set it on SND_SHUTDOWN. BTW examples given
539 * in Stevens' books assume exactly this behaviour, it explains
540 * why EPOLLHUP is incompatible with EPOLLOUT. --ANK
542 * NOTE. Check for TCP_CLOSE is added. The goal is to prevent
543 * blocking on fresh not-connected or disconnected socket. --ANK
545 shutdown = READ_ONCE(sk->sk_shutdown);
546 if (shutdown == SHUTDOWN_MASK || state == TCP_CLOSE)
548 if (shutdown & RCV_SHUTDOWN)
549 mask |= EPOLLIN | EPOLLRDNORM | EPOLLRDHUP;
551 /* Connected or passive Fast Open socket? */
552 if (state != TCP_SYN_SENT &&
553 (state != TCP_SYN_RECV || rcu_access_pointer(tp->fastopen_rsk))) {
554 int target = sock_rcvlowat(sk, 0, INT_MAX);
555 u16 urg_data = READ_ONCE(tp->urg_data);
557 if (unlikely(urg_data) &&
558 READ_ONCE(tp->urg_seq) == READ_ONCE(tp->copied_seq) &&
559 !sock_flag(sk, SOCK_URGINLINE))
562 if (tcp_stream_is_readable(sk, target))
563 mask |= EPOLLIN | EPOLLRDNORM;
565 if (!(shutdown & SEND_SHUTDOWN)) {
566 if (__sk_stream_is_writeable(sk, 1)) {
567 mask |= EPOLLOUT | EPOLLWRNORM;
568 } else { /* send SIGIO later */
569 sk_set_bit(SOCKWQ_ASYNC_NOSPACE, sk);
570 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
572 /* Race breaker. If space is freed after
573 * wspace test but before the flags are set,
574 * IO signal will be lost. Memory barrier
575 * pairs with the input side.
577 smp_mb__after_atomic();
578 if (__sk_stream_is_writeable(sk, 1))
579 mask |= EPOLLOUT | EPOLLWRNORM;
582 mask |= EPOLLOUT | EPOLLWRNORM;
584 if (urg_data & TCP_URG_VALID)
586 } else if (state == TCP_SYN_SENT &&
587 inet_test_bit(DEFER_CONNECT, sk)) {
588 /* Active TCP fastopen socket with defer_connect
589 * Return EPOLLOUT so application can call write()
590 * in order for kernel to generate SYN+data
592 mask |= EPOLLOUT | EPOLLWRNORM;
594 /* This barrier is coupled with smp_wmb() in tcp_reset() */
596 if (READ_ONCE(sk->sk_err) ||
597 !skb_queue_empty_lockless(&sk->sk_error_queue))
602 EXPORT_SYMBOL(tcp_poll);
604 int tcp_ioctl(struct sock *sk, int cmd, int *karg)
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 = READ_ONCE(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);
649 EXPORT_SYMBOL(tcp_ioctl);
651 void tcp_mark_push(struct tcp_sock *tp, struct sk_buff *skb)
653 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
654 tp->pushed_seq = tp->write_seq;
657 static inline bool forced_push(const struct tcp_sock *tp)
659 return after(tp->write_seq, tp->pushed_seq + (tp->max_window >> 1));
662 void tcp_skb_entail(struct sock *sk, struct sk_buff *skb)
664 struct tcp_sock *tp = tcp_sk(sk);
665 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
667 tcb->seq = tcb->end_seq = tp->write_seq;
668 tcb->tcp_flags = TCPHDR_ACK;
669 __skb_header_release(skb);
670 tcp_add_write_queue_tail(sk, skb);
671 sk_wmem_queued_add(sk, skb->truesize);
672 sk_mem_charge(sk, skb->truesize);
673 if (tp->nonagle & TCP_NAGLE_PUSH)
674 tp->nonagle &= ~TCP_NAGLE_PUSH;
676 tcp_slow_start_after_idle_check(sk);
679 static inline void tcp_mark_urg(struct tcp_sock *tp, int flags)
682 tp->snd_up = tp->write_seq;
685 /* If a not yet filled skb is pushed, do not send it if
686 * we have data packets in Qdisc or NIC queues :
687 * Because TX completion will happen shortly, it gives a chance
688 * to coalesce future sendmsg() payload into this skb, without
689 * need for a timer, and with no latency trade off.
690 * As packets containing data payload have a bigger truesize
691 * than pure acks (dataless) packets, the last checks prevent
692 * autocorking if we only have an ACK in Qdisc/NIC queues,
693 * or if TX completion was delayed after we processed ACK packet.
695 static bool tcp_should_autocork(struct sock *sk, struct sk_buff *skb,
698 return skb->len < size_goal &&
699 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_autocorking) &&
700 !tcp_rtx_queue_empty(sk) &&
701 refcount_read(&sk->sk_wmem_alloc) > skb->truesize &&
702 tcp_skb_can_collapse_to(skb);
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 ret = sk_wait_data(sk, &timeo, NULL);
837 if (signal_pending(current)) {
838 ret = sock_intr_errno(timeo);
846 if (!tss.len || !timeo)
851 if (sk->sk_err || sk->sk_state == TCP_CLOSE ||
852 (sk->sk_shutdown & RCV_SHUTDOWN) ||
853 signal_pending(current))
864 EXPORT_SYMBOL(tcp_splice_read);
866 struct sk_buff *tcp_stream_alloc_skb(struct sock *sk, gfp_t gfp,
871 skb = alloc_skb_fclone(MAX_TCP_HEADER, gfp);
875 skb->truesize = SKB_TRUESIZE(skb_end_offset(skb));
876 if (force_schedule) {
877 mem_scheduled = true;
878 sk_forced_mem_schedule(sk, skb->truesize);
880 mem_scheduled = sk_wmem_schedule(sk, skb->truesize);
882 if (likely(mem_scheduled)) {
883 skb_reserve(skb, MAX_TCP_HEADER);
884 skb->ip_summed = CHECKSUM_PARTIAL;
885 INIT_LIST_HEAD(&skb->tcp_tsorted_anchor);
890 sk->sk_prot->enter_memory_pressure(sk);
891 sk_stream_moderate_sndbuf(sk);
896 static unsigned int tcp_xmit_size_goal(struct sock *sk, u32 mss_now,
899 struct tcp_sock *tp = tcp_sk(sk);
900 u32 new_size_goal, size_goal;
905 /* Note : tcp_tso_autosize() will eventually split this later */
906 new_size_goal = tcp_bound_to_half_wnd(tp, sk->sk_gso_max_size);
908 /* We try hard to avoid divides here */
909 size_goal = tp->gso_segs * mss_now;
910 if (unlikely(new_size_goal < size_goal ||
911 new_size_goal >= size_goal + mss_now)) {
912 tp->gso_segs = min_t(u16, new_size_goal / mss_now,
913 sk->sk_gso_max_segs);
914 size_goal = tp->gso_segs * mss_now;
917 return max(size_goal, mss_now);
920 int tcp_send_mss(struct sock *sk, int *size_goal, int flags)
924 mss_now = tcp_current_mss(sk);
925 *size_goal = tcp_xmit_size_goal(sk, mss_now, !(flags & MSG_OOB));
930 /* In some cases, sendmsg() could have added an skb to the write queue,
931 * but failed adding payload on it. We need to remove it to consume less
932 * memory, but more importantly be able to generate EPOLLOUT for Edge Trigger
933 * epoll() users. Another reason is that tcp_write_xmit() does not like
934 * finding an empty skb in the write queue.
936 void tcp_remove_empty_skb(struct sock *sk)
938 struct sk_buff *skb = tcp_write_queue_tail(sk);
940 if (skb && TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq) {
941 tcp_unlink_write_queue(skb, sk);
942 if (tcp_write_queue_empty(sk))
943 tcp_chrono_stop(sk, TCP_CHRONO_BUSY);
944 tcp_wmem_free_skb(sk, skb);
948 /* skb changing from pure zc to mixed, must charge zc */
949 static int tcp_downgrade_zcopy_pure(struct sock *sk, struct sk_buff *skb)
951 if (unlikely(skb_zcopy_pure(skb))) {
952 u32 extra = skb->truesize -
953 SKB_TRUESIZE(skb_end_offset(skb));
955 if (!sk_wmem_schedule(sk, extra))
958 sk_mem_charge(sk, extra);
959 skb_shinfo(skb)->flags &= ~SKBFL_PURE_ZEROCOPY;
965 int tcp_wmem_schedule(struct sock *sk, int copy)
969 if (likely(sk_wmem_schedule(sk, copy)))
972 /* We could be in trouble if we have nothing queued.
973 * Use whatever is left in sk->sk_forward_alloc and tcp_wmem[0]
974 * to guarantee some progress.
976 left = sock_net(sk)->ipv4.sysctl_tcp_wmem[0] - sk->sk_wmem_queued;
978 sk_forced_mem_schedule(sk, min(left, copy));
979 return min(copy, sk->sk_forward_alloc);
982 void tcp_free_fastopen_req(struct tcp_sock *tp)
984 if (tp->fastopen_req) {
985 kfree(tp->fastopen_req);
986 tp->fastopen_req = NULL;
990 int tcp_sendmsg_fastopen(struct sock *sk, struct msghdr *msg, int *copied,
991 size_t size, struct ubuf_info *uarg)
993 struct tcp_sock *tp = tcp_sk(sk);
994 struct inet_sock *inet = inet_sk(sk);
995 struct sockaddr *uaddr = msg->msg_name;
998 if (!(READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_fastopen) &
999 TFO_CLIENT_ENABLE) ||
1000 (uaddr && msg->msg_namelen >= sizeof(uaddr->sa_family) &&
1001 uaddr->sa_family == AF_UNSPEC))
1003 if (tp->fastopen_req)
1004 return -EALREADY; /* Another Fast Open is in progress */
1006 tp->fastopen_req = kzalloc(sizeof(struct tcp_fastopen_request),
1008 if (unlikely(!tp->fastopen_req))
1010 tp->fastopen_req->data = msg;
1011 tp->fastopen_req->size = size;
1012 tp->fastopen_req->uarg = uarg;
1014 if (inet_test_bit(DEFER_CONNECT, sk)) {
1015 err = tcp_connect(sk);
1016 /* Same failure procedure as in tcp_v4/6_connect */
1018 tcp_set_state(sk, TCP_CLOSE);
1019 inet->inet_dport = 0;
1020 sk->sk_route_caps = 0;
1023 flags = (msg->msg_flags & MSG_DONTWAIT) ? O_NONBLOCK : 0;
1024 err = __inet_stream_connect(sk->sk_socket, uaddr,
1025 msg->msg_namelen, flags, 1);
1026 /* fastopen_req could already be freed in __inet_stream_connect
1027 * if the connection times out or gets rst
1029 if (tp->fastopen_req) {
1030 *copied = tp->fastopen_req->copied;
1031 tcp_free_fastopen_req(tp);
1032 inet_clear_bit(DEFER_CONNECT, sk);
1037 int tcp_sendmsg_locked(struct sock *sk, struct msghdr *msg, size_t size)
1039 struct tcp_sock *tp = tcp_sk(sk);
1040 struct ubuf_info *uarg = NULL;
1041 struct sk_buff *skb;
1042 struct sockcm_cookie sockc;
1043 int flags, err, copied = 0;
1044 int mss_now = 0, size_goal, copied_syn = 0;
1045 int process_backlog = 0;
1049 flags = msg->msg_flags;
1051 if ((flags & MSG_ZEROCOPY) && size) {
1052 if (msg->msg_ubuf) {
1053 uarg = msg->msg_ubuf;
1054 if (sk->sk_route_caps & NETIF_F_SG)
1056 } else if (sock_flag(sk, SOCK_ZEROCOPY)) {
1057 skb = tcp_write_queue_tail(sk);
1058 uarg = msg_zerocopy_realloc(sk, size, skb_zcopy(skb));
1063 if (sk->sk_route_caps & NETIF_F_SG)
1066 uarg_to_msgzc(uarg)->zerocopy = 0;
1068 } else if (unlikely(msg->msg_flags & MSG_SPLICE_PAGES) && size) {
1069 if (sk->sk_route_caps & NETIF_F_SG)
1070 zc = MSG_SPLICE_PAGES;
1073 if (unlikely(flags & MSG_FASTOPEN ||
1074 inet_test_bit(DEFER_CONNECT, sk)) &&
1076 err = tcp_sendmsg_fastopen(sk, msg, &copied_syn, size, uarg);
1077 if (err == -EINPROGRESS && copied_syn > 0)
1083 timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
1085 tcp_rate_check_app_limited(sk); /* is sending application-limited? */
1087 /* Wait for a connection to finish. One exception is TCP Fast Open
1088 * (passive side) where data is allowed to be sent before a connection
1089 * is fully established.
1091 if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) &&
1092 !tcp_passive_fastopen(sk)) {
1093 err = sk_stream_wait_connect(sk, &timeo);
1098 if (unlikely(tp->repair)) {
1099 if (tp->repair_queue == TCP_RECV_QUEUE) {
1100 copied = tcp_send_rcvq(sk, msg, size);
1105 if (tp->repair_queue == TCP_NO_QUEUE)
1108 /* 'common' sending to sendq */
1111 sockcm_init(&sockc, sk);
1112 if (msg->msg_controllen) {
1113 err = sock_cmsg_send(sk, msg, &sockc);
1114 if (unlikely(err)) {
1120 /* This should be in poll */
1121 sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
1123 /* Ok commence sending. */
1127 mss_now = tcp_send_mss(sk, &size_goal, flags);
1130 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
1133 while (msg_data_left(msg)) {
1136 skb = tcp_write_queue_tail(sk);
1138 copy = size_goal - skb->len;
1140 if (copy <= 0 || !tcp_skb_can_collapse_to(skb)) {
1144 if (!sk_stream_memory_free(sk))
1145 goto wait_for_space;
1147 if (unlikely(process_backlog >= 16)) {
1148 process_backlog = 0;
1149 if (sk_flush_backlog(sk))
1152 first_skb = tcp_rtx_and_write_queues_empty(sk);
1153 skb = tcp_stream_alloc_skb(sk, sk->sk_allocation,
1156 goto wait_for_space;
1160 tcp_skb_entail(sk, skb);
1163 /* All packets are restored as if they have
1164 * already been sent. skb_mstamp_ns isn't set to
1165 * avoid wrong rtt estimation.
1168 TCP_SKB_CB(skb)->sacked |= TCPCB_REPAIRED;
1171 /* Try to append data to the end of skb. */
1172 if (copy > msg_data_left(msg))
1173 copy = msg_data_left(msg);
1177 int i = skb_shinfo(skb)->nr_frags;
1178 struct page_frag *pfrag = sk_page_frag(sk);
1180 if (!sk_page_frag_refill(sk, pfrag))
1181 goto wait_for_space;
1183 if (!skb_can_coalesce(skb, i, pfrag->page,
1185 if (i >= READ_ONCE(sysctl_max_skb_frags)) {
1186 tcp_mark_push(tp, skb);
1192 copy = min_t(int, copy, pfrag->size - pfrag->offset);
1194 if (unlikely(skb_zcopy_pure(skb) || skb_zcopy_managed(skb))) {
1195 if (tcp_downgrade_zcopy_pure(sk, skb))
1196 goto wait_for_space;
1197 skb_zcopy_downgrade_managed(skb);
1200 copy = tcp_wmem_schedule(sk, copy);
1202 goto wait_for_space;
1204 err = skb_copy_to_page_nocache(sk, &msg->msg_iter, skb,
1211 /* Update the skb. */
1213 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1215 skb_fill_page_desc(skb, i, pfrag->page,
1216 pfrag->offset, copy);
1217 page_ref_inc(pfrag->page);
1219 pfrag->offset += copy;
1220 } else if (zc == MSG_ZEROCOPY) {
1221 /* First append to a fragless skb builds initial
1225 skb_shinfo(skb)->flags |= SKBFL_PURE_ZEROCOPY;
1227 if (!skb_zcopy_pure(skb)) {
1228 copy = tcp_wmem_schedule(sk, copy);
1230 goto wait_for_space;
1233 err = skb_zerocopy_iter_stream(sk, skb, msg, copy, uarg);
1234 if (err == -EMSGSIZE || err == -EEXIST) {
1235 tcp_mark_push(tp, skb);
1241 } else if (zc == MSG_SPLICE_PAGES) {
1242 /* Splice in data if we can; copy if we can't. */
1243 if (tcp_downgrade_zcopy_pure(sk, skb))
1244 goto wait_for_space;
1245 copy = tcp_wmem_schedule(sk, copy);
1247 goto wait_for_space;
1249 err = skb_splice_from_iter(skb, &msg->msg_iter, copy,
1252 if (err == -EMSGSIZE) {
1253 tcp_mark_push(tp, skb);
1260 if (!(flags & MSG_NO_SHARED_FRAGS))
1261 skb_shinfo(skb)->flags |= SKBFL_SHARED_FRAG;
1263 sk_wmem_queued_add(sk, copy);
1264 sk_mem_charge(sk, copy);
1268 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
1270 WRITE_ONCE(tp->write_seq, tp->write_seq + copy);
1271 TCP_SKB_CB(skb)->end_seq += copy;
1272 tcp_skb_pcount_set(skb, 0);
1275 if (!msg_data_left(msg)) {
1276 if (unlikely(flags & MSG_EOR))
1277 TCP_SKB_CB(skb)->eor = 1;
1281 if (skb->len < size_goal || (flags & MSG_OOB) || unlikely(tp->repair))
1284 if (forced_push(tp)) {
1285 tcp_mark_push(tp, skb);
1286 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
1287 } else if (skb == tcp_send_head(sk))
1288 tcp_push_one(sk, mss_now);
1292 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1293 tcp_remove_empty_skb(sk);
1295 tcp_push(sk, flags & ~MSG_MORE, mss_now,
1296 TCP_NAGLE_PUSH, size_goal);
1298 err = sk_stream_wait_memory(sk, &timeo);
1302 mss_now = tcp_send_mss(sk, &size_goal, flags);
1307 tcp_tx_timestamp(sk, sockc.tsflags);
1308 tcp_push(sk, flags, mss_now, tp->nonagle, size_goal);
1311 /* msg->msg_ubuf is pinned by the caller so we don't take extra refs */
1312 if (uarg && !msg->msg_ubuf)
1313 net_zcopy_put(uarg);
1314 return copied + copied_syn;
1317 tcp_remove_empty_skb(sk);
1319 if (copied + copied_syn)
1322 /* msg->msg_ubuf is pinned by the caller so we don't take extra refs */
1323 if (uarg && !msg->msg_ubuf)
1324 net_zcopy_put_abort(uarg, true);
1325 err = sk_stream_error(sk, flags, err);
1326 /* make sure we wake any epoll edge trigger waiter */
1327 if (unlikely(tcp_rtx_and_write_queues_empty(sk) && err == -EAGAIN)) {
1328 sk->sk_write_space(sk);
1329 tcp_chrono_stop(sk, TCP_CHRONO_SNDBUF_LIMITED);
1333 EXPORT_SYMBOL_GPL(tcp_sendmsg_locked);
1335 int tcp_sendmsg(struct sock *sk, struct msghdr *msg, size_t size)
1340 ret = tcp_sendmsg_locked(sk, msg, size);
1345 EXPORT_SYMBOL(tcp_sendmsg);
1347 void tcp_splice_eof(struct socket *sock)
1349 struct sock *sk = sock->sk;
1350 struct tcp_sock *tp = tcp_sk(sk);
1351 int mss_now, size_goal;
1353 if (!tcp_write_queue_tail(sk))
1357 mss_now = tcp_send_mss(sk, &size_goal, 0);
1358 tcp_push(sk, 0, mss_now, tp->nonagle, size_goal);
1361 EXPORT_SYMBOL_GPL(tcp_splice_eof);
1364 * Handle reading urgent data. BSD has very simple semantics for
1365 * this, no blocking and very strange errors 8)
1368 static int tcp_recv_urg(struct sock *sk, struct msghdr *msg, int len, int flags)
1370 struct tcp_sock *tp = tcp_sk(sk);
1372 /* No URG data to read. */
1373 if (sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data ||
1374 tp->urg_data == TCP_URG_READ)
1375 return -EINVAL; /* Yes this is right ! */
1377 if (sk->sk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DONE))
1380 if (tp->urg_data & TCP_URG_VALID) {
1382 char c = tp->urg_data;
1384 if (!(flags & MSG_PEEK))
1385 WRITE_ONCE(tp->urg_data, TCP_URG_READ);
1387 /* Read urgent data. */
1388 msg->msg_flags |= MSG_OOB;
1391 if (!(flags & MSG_TRUNC))
1392 err = memcpy_to_msg(msg, &c, 1);
1395 msg->msg_flags |= MSG_TRUNC;
1397 return err ? -EFAULT : len;
1400 if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
1403 /* Fixed the recv(..., MSG_OOB) behaviour. BSD docs and
1404 * the available implementations agree in this case:
1405 * this call should never block, independent of the
1406 * blocking state of the socket.
1407 * Mike <pall@rz.uni-karlsruhe.de>
1412 static int tcp_peek_sndq(struct sock *sk, struct msghdr *msg, int len)
1414 struct sk_buff *skb;
1415 int copied = 0, err = 0;
1417 /* XXX -- need to support SO_PEEK_OFF */
1419 skb_rbtree_walk(skb, &sk->tcp_rtx_queue) {
1420 err = skb_copy_datagram_msg(skb, 0, msg, skb->len);
1426 skb_queue_walk(&sk->sk_write_queue, skb) {
1427 err = skb_copy_datagram_msg(skb, 0, msg, skb->len);
1434 return err ?: copied;
1437 /* Clean up the receive buffer for full frames taken by the user,
1438 * then send an ACK if necessary. COPIED is the number of bytes
1439 * tcp_recvmsg has given to the user so far, it speeds up the
1440 * calculation of whether or not we must ACK for the sake of
1443 void __tcp_cleanup_rbuf(struct sock *sk, int copied)
1445 struct tcp_sock *tp = tcp_sk(sk);
1446 bool time_to_ack = false;
1448 if (inet_csk_ack_scheduled(sk)) {
1449 const struct inet_connection_sock *icsk = inet_csk(sk);
1451 if (/* Once-per-two-segments ACK was not sent by tcp_input.c */
1452 tp->rcv_nxt - tp->rcv_wup > icsk->icsk_ack.rcv_mss ||
1454 * If this read emptied read buffer, we send ACK, if
1455 * connection is not bidirectional, user drained
1456 * receive buffer and there was a small segment
1460 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED2) ||
1461 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED) &&
1462 !inet_csk_in_pingpong_mode(sk))) &&
1463 !atomic_read(&sk->sk_rmem_alloc)))
1467 /* We send an ACK if we can now advertise a non-zero window
1468 * which has been raised "significantly".
1470 * Even if window raised up to infinity, do not send window open ACK
1471 * in states, where we will not receive more. It is useless.
1473 if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) {
1474 __u32 rcv_window_now = tcp_receive_window(tp);
1476 /* Optimize, __tcp_select_window() is not cheap. */
1477 if (2*rcv_window_now <= tp->window_clamp) {
1478 __u32 new_window = __tcp_select_window(sk);
1480 /* Send ACK now, if this read freed lots of space
1481 * in our buffer. Certainly, new_window is new window.
1482 * We can advertise it now, if it is not less than current one.
1483 * "Lots" means "at least twice" here.
1485 if (new_window && new_window >= 2 * rcv_window_now)
1493 void tcp_cleanup_rbuf(struct sock *sk, int copied)
1495 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
1496 struct tcp_sock *tp = tcp_sk(sk);
1498 WARN(skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq),
1499 "cleanup rbuf bug: copied %X seq %X rcvnxt %X\n",
1500 tp->copied_seq, TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt);
1501 __tcp_cleanup_rbuf(sk, copied);
1504 static void tcp_eat_recv_skb(struct sock *sk, struct sk_buff *skb)
1506 __skb_unlink(skb, &sk->sk_receive_queue);
1507 if (likely(skb->destructor == sock_rfree)) {
1509 skb->destructor = NULL;
1511 return skb_attempt_defer_free(skb);
1516 struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off)
1518 struct sk_buff *skb;
1521 while ((skb = skb_peek(&sk->sk_receive_queue)) != NULL) {
1522 offset = seq - TCP_SKB_CB(skb)->seq;
1523 if (unlikely(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
1524 pr_err_once("%s: found a SYN, please report !\n", __func__);
1527 if (offset < skb->len || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)) {
1531 /* This looks weird, but this can happen if TCP collapsing
1532 * splitted a fat GRO packet, while we released socket lock
1533 * in skb_splice_bits()
1535 tcp_eat_recv_skb(sk, skb);
1539 EXPORT_SYMBOL(tcp_recv_skb);
1542 * This routine provides an alternative to tcp_recvmsg() for routines
1543 * that would like to handle copying from skbuffs directly in 'sendfile'
1546 * - It is assumed that the socket was locked by the caller.
1547 * - The routine does not block.
1548 * - At present, there is no support for reading OOB data
1549 * or for 'peeking' the socket using this routine
1550 * (although both would be easy to implement).
1552 int tcp_read_sock(struct sock *sk, read_descriptor_t *desc,
1553 sk_read_actor_t recv_actor)
1555 struct sk_buff *skb;
1556 struct tcp_sock *tp = tcp_sk(sk);
1557 u32 seq = tp->copied_seq;
1561 if (sk->sk_state == TCP_LISTEN)
1563 while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1564 if (offset < skb->len) {
1568 len = skb->len - offset;
1569 /* Stop reading if we hit a patch of urgent data */
1570 if (unlikely(tp->urg_data)) {
1571 u32 urg_offset = tp->urg_seq - seq;
1572 if (urg_offset < len)
1577 used = recv_actor(desc, skb, offset, len);
1583 if (WARN_ON_ONCE(used > len))
1589 /* If recv_actor drops the lock (e.g. TCP splice
1590 * receive) the skb pointer might be invalid when
1591 * getting here: tcp_collapse might have deleted it
1592 * while aggregating skbs from the socket queue.
1594 skb = tcp_recv_skb(sk, seq - 1, &offset);
1597 /* TCP coalescing might have appended data to the skb.
1598 * Try to splice more frags
1600 if (offset + 1 != skb->len)
1603 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) {
1604 tcp_eat_recv_skb(sk, skb);
1608 tcp_eat_recv_skb(sk, skb);
1611 WRITE_ONCE(tp->copied_seq, seq);
1613 WRITE_ONCE(tp->copied_seq, seq);
1615 tcp_rcv_space_adjust(sk);
1617 /* Clean up data we have read: This will do ACK frames. */
1619 tcp_recv_skb(sk, seq, &offset);
1620 tcp_cleanup_rbuf(sk, copied);
1624 EXPORT_SYMBOL(tcp_read_sock);
1626 int tcp_read_skb(struct sock *sk, skb_read_actor_t recv_actor)
1628 struct sk_buff *skb;
1631 if (sk->sk_state == TCP_LISTEN)
1634 while ((skb = skb_peek(&sk->sk_receive_queue)) != NULL) {
1638 __skb_unlink(skb, &sk->sk_receive_queue);
1639 WARN_ON_ONCE(!skb_set_owner_sk_safe(skb, sk));
1640 tcp_flags = TCP_SKB_CB(skb)->tcp_flags;
1641 used = recv_actor(sk, skb);
1649 if (tcp_flags & TCPHDR_FIN)
1654 EXPORT_SYMBOL(tcp_read_skb);
1656 void tcp_read_done(struct sock *sk, size_t len)
1658 struct tcp_sock *tp = tcp_sk(sk);
1659 u32 seq = tp->copied_seq;
1660 struct sk_buff *skb;
1664 if (sk->sk_state == TCP_LISTEN)
1668 while (left && (skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1671 used = min_t(size_t, skb->len - offset, left);
1675 if (skb->len > offset + used)
1678 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) {
1679 tcp_eat_recv_skb(sk, skb);
1683 tcp_eat_recv_skb(sk, skb);
1685 WRITE_ONCE(tp->copied_seq, seq);
1687 tcp_rcv_space_adjust(sk);
1689 /* Clean up data we have read: This will do ACK frames. */
1691 tcp_cleanup_rbuf(sk, len - left);
1693 EXPORT_SYMBOL(tcp_read_done);
1695 int tcp_peek_len(struct socket *sock)
1697 return tcp_inq(sock->sk);
1699 EXPORT_SYMBOL(tcp_peek_len);
1701 /* Make sure sk_rcvbuf is big enough to satisfy SO_RCVLOWAT hint */
1702 int tcp_set_rcvlowat(struct sock *sk, int val)
1706 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK)
1707 cap = sk->sk_rcvbuf >> 1;
1709 cap = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_rmem[2]) >> 1;
1710 val = min(val, cap);
1711 WRITE_ONCE(sk->sk_rcvlowat, val ? : 1);
1713 /* Check if we need to signal EPOLLIN right now */
1716 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK)
1719 space = tcp_space_from_win(sk, val);
1720 if (space > sk->sk_rcvbuf) {
1721 WRITE_ONCE(sk->sk_rcvbuf, space);
1722 tcp_sk(sk)->window_clamp = val;
1726 EXPORT_SYMBOL(tcp_set_rcvlowat);
1728 void tcp_update_recv_tstamps(struct sk_buff *skb,
1729 struct scm_timestamping_internal *tss)
1732 tss->ts[0] = ktime_to_timespec64(skb->tstamp);
1734 tss->ts[0] = (struct timespec64) {0};
1736 if (skb_hwtstamps(skb)->hwtstamp)
1737 tss->ts[2] = ktime_to_timespec64(skb_hwtstamps(skb)->hwtstamp);
1739 tss->ts[2] = (struct timespec64) {0};
1743 static const struct vm_operations_struct tcp_vm_ops = {
1746 int tcp_mmap(struct file *file, struct socket *sock,
1747 struct vm_area_struct *vma)
1749 if (vma->vm_flags & (VM_WRITE | VM_EXEC))
1751 vm_flags_clear(vma, VM_MAYWRITE | VM_MAYEXEC);
1753 /* Instruct vm_insert_page() to not mmap_read_lock(mm) */
1754 vm_flags_set(vma, VM_MIXEDMAP);
1756 vma->vm_ops = &tcp_vm_ops;
1759 EXPORT_SYMBOL(tcp_mmap);
1761 static skb_frag_t *skb_advance_to_frag(struct sk_buff *skb, u32 offset_skb,
1766 if (unlikely(offset_skb >= skb->len))
1769 offset_skb -= skb_headlen(skb);
1770 if ((int)offset_skb < 0 || skb_has_frag_list(skb))
1773 frag = skb_shinfo(skb)->frags;
1774 while (offset_skb) {
1775 if (skb_frag_size(frag) > offset_skb) {
1776 *offset_frag = offset_skb;
1779 offset_skb -= skb_frag_size(frag);
1786 static bool can_map_frag(const skb_frag_t *frag)
1788 return skb_frag_size(frag) == PAGE_SIZE && !skb_frag_off(frag);
1791 static int find_next_mappable_frag(const skb_frag_t *frag,
1792 int remaining_in_skb)
1796 if (likely(can_map_frag(frag)))
1799 while (offset < remaining_in_skb && !can_map_frag(frag)) {
1800 offset += skb_frag_size(frag);
1806 static void tcp_zerocopy_set_hint_for_skb(struct sock *sk,
1807 struct tcp_zerocopy_receive *zc,
1808 struct sk_buff *skb, u32 offset)
1810 u32 frag_offset, partial_frag_remainder = 0;
1811 int mappable_offset;
1814 /* worst case: skip to next skb. try to improve on this case below */
1815 zc->recv_skip_hint = skb->len - offset;
1817 /* Find the frag containing this offset (and how far into that frag) */
1818 frag = skb_advance_to_frag(skb, offset, &frag_offset);
1823 struct skb_shared_info *info = skb_shinfo(skb);
1825 /* We read part of the last frag, must recvmsg() rest of skb. */
1826 if (frag == &info->frags[info->nr_frags - 1])
1829 /* Else, we must at least read the remainder in this frag. */
1830 partial_frag_remainder = skb_frag_size(frag) - frag_offset;
1831 zc->recv_skip_hint -= partial_frag_remainder;
1835 /* partial_frag_remainder: If part way through a frag, must read rest.
1836 * mappable_offset: Bytes till next mappable frag, *not* counting bytes
1837 * in partial_frag_remainder.
1839 mappable_offset = find_next_mappable_frag(frag, zc->recv_skip_hint);
1840 zc->recv_skip_hint = mappable_offset + partial_frag_remainder;
1843 static int tcp_recvmsg_locked(struct sock *sk, struct msghdr *msg, size_t len,
1844 int flags, struct scm_timestamping_internal *tss,
1846 static int receive_fallback_to_copy(struct sock *sk,
1847 struct tcp_zerocopy_receive *zc, int inq,
1848 struct scm_timestamping_internal *tss)
1850 unsigned long copy_address = (unsigned long)zc->copybuf_address;
1851 struct msghdr msg = {};
1856 zc->recv_skip_hint = 0;
1858 if (copy_address != zc->copybuf_address)
1861 err = import_single_range(ITER_DEST, (void __user *)copy_address,
1862 inq, &iov, &msg.msg_iter);
1866 err = tcp_recvmsg_locked(sk, &msg, inq, MSG_DONTWAIT,
1867 tss, &zc->msg_flags);
1871 zc->copybuf_len = err;
1872 if (likely(zc->copybuf_len)) {
1873 struct sk_buff *skb;
1876 skb = tcp_recv_skb(sk, tcp_sk(sk)->copied_seq, &offset);
1878 tcp_zerocopy_set_hint_for_skb(sk, zc, skb, offset);
1883 static int tcp_copy_straggler_data(struct tcp_zerocopy_receive *zc,
1884 struct sk_buff *skb, u32 copylen,
1885 u32 *offset, u32 *seq)
1887 unsigned long copy_address = (unsigned long)zc->copybuf_address;
1888 struct msghdr msg = {};
1892 if (copy_address != zc->copybuf_address)
1895 err = import_single_range(ITER_DEST, (void __user *)copy_address,
1896 copylen, &iov, &msg.msg_iter);
1899 err = skb_copy_datagram_msg(skb, *offset, &msg, copylen);
1902 zc->recv_skip_hint -= copylen;
1905 return (__s32)copylen;
1908 static int tcp_zc_handle_leftover(struct tcp_zerocopy_receive *zc,
1910 struct sk_buff *skb,
1913 struct scm_timestamping_internal *tss)
1915 u32 offset, copylen = min_t(u32, copybuf_len, zc->recv_skip_hint);
1919 /* skb is null if inq < PAGE_SIZE. */
1921 offset = *seq - TCP_SKB_CB(skb)->seq;
1923 skb = tcp_recv_skb(sk, *seq, &offset);
1924 if (TCP_SKB_CB(skb)->has_rxtstamp) {
1925 tcp_update_recv_tstamps(skb, tss);
1926 zc->msg_flags |= TCP_CMSG_TS;
1930 zc->copybuf_len = tcp_copy_straggler_data(zc, skb, copylen, &offset,
1932 return zc->copybuf_len < 0 ? 0 : copylen;
1935 static int tcp_zerocopy_vm_insert_batch_error(struct vm_area_struct *vma,
1936 struct page **pending_pages,
1937 unsigned long pages_remaining,
1938 unsigned long *address,
1941 struct tcp_zerocopy_receive *zc,
1942 u32 total_bytes_to_map,
1945 /* At least one page did not map. Try zapping if we skipped earlier. */
1946 if (err == -EBUSY &&
1947 zc->flags & TCP_RECEIVE_ZEROCOPY_FLAG_TLB_CLEAN_HINT) {
1950 maybe_zap_len = total_bytes_to_map - /* All bytes to map */
1951 *length + /* Mapped or pending */
1952 (pages_remaining * PAGE_SIZE); /* Failed map. */
1953 zap_page_range_single(vma, *address, maybe_zap_len, NULL);
1958 unsigned long leftover_pages = pages_remaining;
1961 /* We called zap_page_range_single, try to reinsert. */
1962 err = vm_insert_pages(vma, *address,
1965 bytes_mapped = PAGE_SIZE * (leftover_pages - pages_remaining);
1966 *seq += bytes_mapped;
1967 *address += bytes_mapped;
1970 /* Either we were unable to zap, OR we zapped, retried an
1971 * insert, and still had an issue. Either ways, pages_remaining
1972 * is the number of pages we were unable to map, and we unroll
1973 * some state we speculatively touched before.
1975 const int bytes_not_mapped = PAGE_SIZE * pages_remaining;
1977 *length -= bytes_not_mapped;
1978 zc->recv_skip_hint += bytes_not_mapped;
1983 static int tcp_zerocopy_vm_insert_batch(struct vm_area_struct *vma,
1984 struct page **pages,
1985 unsigned int pages_to_map,
1986 unsigned long *address,
1989 struct tcp_zerocopy_receive *zc,
1990 u32 total_bytes_to_map)
1992 unsigned long pages_remaining = pages_to_map;
1993 unsigned int pages_mapped;
1994 unsigned int bytes_mapped;
1997 err = vm_insert_pages(vma, *address, pages, &pages_remaining);
1998 pages_mapped = pages_to_map - (unsigned int)pages_remaining;
1999 bytes_mapped = PAGE_SIZE * pages_mapped;
2000 /* Even if vm_insert_pages fails, it may have partially succeeded in
2001 * mapping (some but not all of the pages).
2003 *seq += bytes_mapped;
2004 *address += bytes_mapped;
2009 /* Error: maybe zap and retry + rollback state for failed inserts. */
2010 return tcp_zerocopy_vm_insert_batch_error(vma, pages + pages_mapped,
2011 pages_remaining, address, length, seq, zc, total_bytes_to_map,
2015 #define TCP_VALID_ZC_MSG_FLAGS (TCP_CMSG_TS)
2016 static void tcp_zc_finalize_rx_tstamp(struct sock *sk,
2017 struct tcp_zerocopy_receive *zc,
2018 struct scm_timestamping_internal *tss)
2020 unsigned long msg_control_addr;
2021 struct msghdr cmsg_dummy;
2023 msg_control_addr = (unsigned long)zc->msg_control;
2024 cmsg_dummy.msg_control_user = (void __user *)msg_control_addr;
2025 cmsg_dummy.msg_controllen =
2026 (__kernel_size_t)zc->msg_controllen;
2027 cmsg_dummy.msg_flags = in_compat_syscall()
2028 ? MSG_CMSG_COMPAT : 0;
2029 cmsg_dummy.msg_control_is_user = true;
2031 if (zc->msg_control == msg_control_addr &&
2032 zc->msg_controllen == cmsg_dummy.msg_controllen) {
2033 tcp_recv_timestamp(&cmsg_dummy, sk, tss);
2034 zc->msg_control = (__u64)
2035 ((uintptr_t)cmsg_dummy.msg_control_user);
2036 zc->msg_controllen =
2037 (__u64)cmsg_dummy.msg_controllen;
2038 zc->msg_flags = (__u32)cmsg_dummy.msg_flags;
2042 static struct vm_area_struct *find_tcp_vma(struct mm_struct *mm,
2043 unsigned long address,
2046 struct vm_area_struct *vma = lock_vma_under_rcu(mm, address);
2049 if (vma->vm_ops != &tcp_vm_ops) {
2053 *mmap_locked = false;
2058 vma = vma_lookup(mm, address);
2059 if (!vma || vma->vm_ops != &tcp_vm_ops) {
2060 mmap_read_unlock(mm);
2063 *mmap_locked = true;
2067 #define TCP_ZEROCOPY_PAGE_BATCH_SIZE 32
2068 static int tcp_zerocopy_receive(struct sock *sk,
2069 struct tcp_zerocopy_receive *zc,
2070 struct scm_timestamping_internal *tss)
2072 u32 length = 0, offset, vma_len, avail_len, copylen = 0;
2073 unsigned long address = (unsigned long)zc->address;
2074 struct page *pages[TCP_ZEROCOPY_PAGE_BATCH_SIZE];
2075 s32 copybuf_len = zc->copybuf_len;
2076 struct tcp_sock *tp = tcp_sk(sk);
2077 const skb_frag_t *frags = NULL;
2078 unsigned int pages_to_map = 0;
2079 struct vm_area_struct *vma;
2080 struct sk_buff *skb = NULL;
2081 u32 seq = tp->copied_seq;
2082 u32 total_bytes_to_map;
2083 int inq = tcp_inq(sk);
2087 zc->copybuf_len = 0;
2090 if (address & (PAGE_SIZE - 1) || address != zc->address)
2093 if (sk->sk_state == TCP_LISTEN)
2096 sock_rps_record_flow(sk);
2098 if (inq && inq <= copybuf_len)
2099 return receive_fallback_to_copy(sk, zc, inq, tss);
2101 if (inq < PAGE_SIZE) {
2103 zc->recv_skip_hint = inq;
2104 if (!inq && sock_flag(sk, SOCK_DONE))
2109 vma = find_tcp_vma(current->mm, address, &mmap_locked);
2113 vma_len = min_t(unsigned long, zc->length, vma->vm_end - address);
2114 avail_len = min_t(u32, vma_len, inq);
2115 total_bytes_to_map = avail_len & ~(PAGE_SIZE - 1);
2116 if (total_bytes_to_map) {
2117 if (!(zc->flags & TCP_RECEIVE_ZEROCOPY_FLAG_TLB_CLEAN_HINT))
2118 zap_page_range_single(vma, address, total_bytes_to_map,
2120 zc->length = total_bytes_to_map;
2121 zc->recv_skip_hint = 0;
2123 zc->length = avail_len;
2124 zc->recv_skip_hint = avail_len;
2127 while (length + PAGE_SIZE <= zc->length) {
2128 int mappable_offset;
2131 if (zc->recv_skip_hint < PAGE_SIZE) {
2135 if (zc->recv_skip_hint > 0)
2138 offset = seq - TCP_SKB_CB(skb)->seq;
2140 skb = tcp_recv_skb(sk, seq, &offset);
2143 if (TCP_SKB_CB(skb)->has_rxtstamp) {
2144 tcp_update_recv_tstamps(skb, tss);
2145 zc->msg_flags |= TCP_CMSG_TS;
2147 zc->recv_skip_hint = skb->len - offset;
2148 frags = skb_advance_to_frag(skb, offset, &offset_frag);
2149 if (!frags || offset_frag)
2153 mappable_offset = find_next_mappable_frag(frags,
2154 zc->recv_skip_hint);
2155 if (mappable_offset) {
2156 zc->recv_skip_hint = mappable_offset;
2159 page = skb_frag_page(frags);
2161 pages[pages_to_map++] = page;
2162 length += PAGE_SIZE;
2163 zc->recv_skip_hint -= PAGE_SIZE;
2165 if (pages_to_map == TCP_ZEROCOPY_PAGE_BATCH_SIZE ||
2166 zc->recv_skip_hint < PAGE_SIZE) {
2167 /* Either full batch, or we're about to go to next skb
2168 * (and we cannot unroll failed ops across skbs).
2170 ret = tcp_zerocopy_vm_insert_batch(vma, pages,
2174 total_bytes_to_map);
2181 ret = tcp_zerocopy_vm_insert_batch(vma, pages, pages_to_map,
2182 &address, &length, &seq,
2183 zc, total_bytes_to_map);
2187 mmap_read_unlock(current->mm);
2190 /* Try to copy straggler data. */
2192 copylen = tcp_zc_handle_leftover(zc, sk, skb, &seq, copybuf_len, tss);
2194 if (length + copylen) {
2195 WRITE_ONCE(tp->copied_seq, seq);
2196 tcp_rcv_space_adjust(sk);
2198 /* Clean up data we have read: This will do ACK frames. */
2199 tcp_recv_skb(sk, seq, &offset);
2200 tcp_cleanup_rbuf(sk, length + copylen);
2202 if (length == zc->length)
2203 zc->recv_skip_hint = 0;
2205 if (!zc->recv_skip_hint && sock_flag(sk, SOCK_DONE))
2208 zc->length = length;
2213 /* Similar to __sock_recv_timestamp, but does not require an skb */
2214 void tcp_recv_timestamp(struct msghdr *msg, const struct sock *sk,
2215 struct scm_timestamping_internal *tss)
2217 int new_tstamp = sock_flag(sk, SOCK_TSTAMP_NEW);
2218 bool has_timestamping = false;
2220 if (tss->ts[0].tv_sec || tss->ts[0].tv_nsec) {
2221 if (sock_flag(sk, SOCK_RCVTSTAMP)) {
2222 if (sock_flag(sk, SOCK_RCVTSTAMPNS)) {
2224 struct __kernel_timespec kts = {
2225 .tv_sec = tss->ts[0].tv_sec,
2226 .tv_nsec = tss->ts[0].tv_nsec,
2228 put_cmsg(msg, SOL_SOCKET, SO_TIMESTAMPNS_NEW,
2231 struct __kernel_old_timespec ts_old = {
2232 .tv_sec = tss->ts[0].tv_sec,
2233 .tv_nsec = tss->ts[0].tv_nsec,
2235 put_cmsg(msg, SOL_SOCKET, SO_TIMESTAMPNS_OLD,
2236 sizeof(ts_old), &ts_old);
2240 struct __kernel_sock_timeval stv = {
2241 .tv_sec = tss->ts[0].tv_sec,
2242 .tv_usec = tss->ts[0].tv_nsec / 1000,
2244 put_cmsg(msg, SOL_SOCKET, SO_TIMESTAMP_NEW,
2247 struct __kernel_old_timeval tv = {
2248 .tv_sec = tss->ts[0].tv_sec,
2249 .tv_usec = tss->ts[0].tv_nsec / 1000,
2251 put_cmsg(msg, SOL_SOCKET, SO_TIMESTAMP_OLD,
2257 if (READ_ONCE(sk->sk_tsflags) & SOF_TIMESTAMPING_SOFTWARE)
2258 has_timestamping = true;
2260 tss->ts[0] = (struct timespec64) {0};
2263 if (tss->ts[2].tv_sec || tss->ts[2].tv_nsec) {
2264 if (READ_ONCE(sk->sk_tsflags) & SOF_TIMESTAMPING_RAW_HARDWARE)
2265 has_timestamping = true;
2267 tss->ts[2] = (struct timespec64) {0};
2270 if (has_timestamping) {
2271 tss->ts[1] = (struct timespec64) {0};
2272 if (sock_flag(sk, SOCK_TSTAMP_NEW))
2273 put_cmsg_scm_timestamping64(msg, tss);
2275 put_cmsg_scm_timestamping(msg, tss);
2279 static int tcp_inq_hint(struct sock *sk)
2281 const struct tcp_sock *tp = tcp_sk(sk);
2282 u32 copied_seq = READ_ONCE(tp->copied_seq);
2283 u32 rcv_nxt = READ_ONCE(tp->rcv_nxt);
2286 inq = rcv_nxt - copied_seq;
2287 if (unlikely(inq < 0 || copied_seq != READ_ONCE(tp->copied_seq))) {
2289 inq = tp->rcv_nxt - tp->copied_seq;
2292 /* After receiving a FIN, tell the user-space to continue reading
2293 * by returning a non-zero inq.
2295 if (inq == 0 && sock_flag(sk, SOCK_DONE))
2301 * This routine copies from a sock struct into the user buffer.
2303 * Technical note: in 2.3 we work on _locked_ socket, so that
2304 * tricks with *seq access order and skb->users are not required.
2305 * Probably, code can be easily improved even more.
2308 static int tcp_recvmsg_locked(struct sock *sk, struct msghdr *msg, size_t len,
2309 int flags, struct scm_timestamping_internal *tss,
2312 struct tcp_sock *tp = tcp_sk(sk);
2318 int target; /* Read at least this many bytes */
2320 struct sk_buff *skb, *last;
2324 if (sk->sk_state == TCP_LISTEN)
2327 if (tp->recvmsg_inq) {
2328 *cmsg_flags = TCP_CMSG_INQ;
2329 msg->msg_get_inq = 1;
2331 timeo = sock_rcvtimeo(sk, flags & MSG_DONTWAIT);
2333 /* Urgent data needs to be handled specially. */
2334 if (flags & MSG_OOB)
2337 if (unlikely(tp->repair)) {
2339 if (!(flags & MSG_PEEK))
2342 if (tp->repair_queue == TCP_SEND_QUEUE)
2346 if (tp->repair_queue == TCP_NO_QUEUE)
2349 /* 'common' recv queue MSG_PEEK-ing */
2352 seq = &tp->copied_seq;
2353 if (flags & MSG_PEEK) {
2354 peek_seq = tp->copied_seq;
2358 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
2363 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
2364 if (unlikely(tp->urg_data) && tp->urg_seq == *seq) {
2367 if (signal_pending(current)) {
2368 copied = timeo ? sock_intr_errno(timeo) : -EAGAIN;
2373 /* Next get a buffer. */
2375 last = skb_peek_tail(&sk->sk_receive_queue);
2376 skb_queue_walk(&sk->sk_receive_queue, skb) {
2378 /* Now that we have two receive queues this
2381 if (WARN(before(*seq, TCP_SKB_CB(skb)->seq),
2382 "TCP recvmsg seq # bug: copied %X, seq %X, rcvnxt %X, fl %X\n",
2383 *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt,
2387 offset = *seq - TCP_SKB_CB(skb)->seq;
2388 if (unlikely(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
2389 pr_err_once("%s: found a SYN, please report !\n", __func__);
2392 if (offset < skb->len)
2394 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
2396 WARN(!(flags & MSG_PEEK),
2397 "TCP recvmsg seq # bug 2: copied %X, seq %X, rcvnxt %X, fl %X\n",
2398 *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt, flags);
2401 /* Well, if we have backlog, try to process it now yet. */
2403 if (copied >= target && !READ_ONCE(sk->sk_backlog.tail))
2409 sk->sk_state == TCP_CLOSE ||
2410 (sk->sk_shutdown & RCV_SHUTDOWN) ||
2411 signal_pending(current))
2414 if (sock_flag(sk, SOCK_DONE))
2418 copied = sock_error(sk);
2422 if (sk->sk_shutdown & RCV_SHUTDOWN)
2425 if (sk->sk_state == TCP_CLOSE) {
2426 /* This occurs when user tries to read
2427 * from never connected socket.
2438 if (signal_pending(current)) {
2439 copied = sock_intr_errno(timeo);
2444 if (copied >= target) {
2445 /* Do not sleep, just process backlog. */
2446 __sk_flush_backlog(sk);
2448 tcp_cleanup_rbuf(sk, copied);
2449 err = sk_wait_data(sk, &timeo, last);
2451 err = copied ? : err;
2456 if ((flags & MSG_PEEK) &&
2457 (peek_seq - copied - urg_hole != tp->copied_seq)) {
2458 net_dbg_ratelimited("TCP(%s:%d): Application bug, race in MSG_PEEK\n",
2460 task_pid_nr(current));
2461 peek_seq = tp->copied_seq;
2466 /* Ok so how much can we use? */
2467 used = skb->len - offset;
2471 /* Do we have urgent data here? */
2472 if (unlikely(tp->urg_data)) {
2473 u32 urg_offset = tp->urg_seq - *seq;
2474 if (urg_offset < used) {
2476 if (!sock_flag(sk, SOCK_URGINLINE)) {
2477 WRITE_ONCE(*seq, *seq + 1);
2489 if (!(flags & MSG_TRUNC)) {
2490 err = skb_copy_datagram_msg(skb, offset, msg, used);
2492 /* Exception. Bailout! */
2499 WRITE_ONCE(*seq, *seq + used);
2503 tcp_rcv_space_adjust(sk);
2506 if (unlikely(tp->urg_data) && after(tp->copied_seq, tp->urg_seq)) {
2507 WRITE_ONCE(tp->urg_data, 0);
2508 tcp_fast_path_check(sk);
2511 if (TCP_SKB_CB(skb)->has_rxtstamp) {
2512 tcp_update_recv_tstamps(skb, tss);
2513 *cmsg_flags |= TCP_CMSG_TS;
2516 if (used + offset < skb->len)
2519 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
2521 if (!(flags & MSG_PEEK))
2522 tcp_eat_recv_skb(sk, skb);
2526 /* Process the FIN. */
2527 WRITE_ONCE(*seq, *seq + 1);
2528 if (!(flags & MSG_PEEK))
2529 tcp_eat_recv_skb(sk, skb);
2533 /* According to UNIX98, msg_name/msg_namelen are ignored
2534 * on connected socket. I was just happy when found this 8) --ANK
2537 /* Clean up data we have read: This will do ACK frames. */
2538 tcp_cleanup_rbuf(sk, copied);
2545 err = tcp_recv_urg(sk, msg, len, flags);
2549 err = tcp_peek_sndq(sk, msg, len);
2553 int tcp_recvmsg(struct sock *sk, struct msghdr *msg, size_t len, int flags,
2556 int cmsg_flags = 0, ret;
2557 struct scm_timestamping_internal tss;
2559 if (unlikely(flags & MSG_ERRQUEUE))
2560 return inet_recv_error(sk, msg, len, addr_len);
2562 if (sk_can_busy_loop(sk) &&
2563 skb_queue_empty_lockless(&sk->sk_receive_queue) &&
2564 sk->sk_state == TCP_ESTABLISHED)
2565 sk_busy_loop(sk, flags & MSG_DONTWAIT);
2568 ret = tcp_recvmsg_locked(sk, msg, len, flags, &tss, &cmsg_flags);
2571 if ((cmsg_flags || msg->msg_get_inq) && ret >= 0) {
2572 if (cmsg_flags & TCP_CMSG_TS)
2573 tcp_recv_timestamp(msg, sk, &tss);
2574 if (msg->msg_get_inq) {
2575 msg->msg_inq = tcp_inq_hint(sk);
2576 if (cmsg_flags & TCP_CMSG_INQ)
2577 put_cmsg(msg, SOL_TCP, TCP_CM_INQ,
2578 sizeof(msg->msg_inq), &msg->msg_inq);
2583 EXPORT_SYMBOL(tcp_recvmsg);
2585 void tcp_set_state(struct sock *sk, int state)
2587 int oldstate = sk->sk_state;
2589 /* We defined a new enum for TCP states that are exported in BPF
2590 * so as not force the internal TCP states to be frozen. The
2591 * following checks will detect if an internal state value ever
2592 * differs from the BPF value. If this ever happens, then we will
2593 * need to remap the internal value to the BPF value before calling
2594 * tcp_call_bpf_2arg.
2596 BUILD_BUG_ON((int)BPF_TCP_ESTABLISHED != (int)TCP_ESTABLISHED);
2597 BUILD_BUG_ON((int)BPF_TCP_SYN_SENT != (int)TCP_SYN_SENT);
2598 BUILD_BUG_ON((int)BPF_TCP_SYN_RECV != (int)TCP_SYN_RECV);
2599 BUILD_BUG_ON((int)BPF_TCP_FIN_WAIT1 != (int)TCP_FIN_WAIT1);
2600 BUILD_BUG_ON((int)BPF_TCP_FIN_WAIT2 != (int)TCP_FIN_WAIT2);
2601 BUILD_BUG_ON((int)BPF_TCP_TIME_WAIT != (int)TCP_TIME_WAIT);
2602 BUILD_BUG_ON((int)BPF_TCP_CLOSE != (int)TCP_CLOSE);
2603 BUILD_BUG_ON((int)BPF_TCP_CLOSE_WAIT != (int)TCP_CLOSE_WAIT);
2604 BUILD_BUG_ON((int)BPF_TCP_LAST_ACK != (int)TCP_LAST_ACK);
2605 BUILD_BUG_ON((int)BPF_TCP_LISTEN != (int)TCP_LISTEN);
2606 BUILD_BUG_ON((int)BPF_TCP_CLOSING != (int)TCP_CLOSING);
2607 BUILD_BUG_ON((int)BPF_TCP_NEW_SYN_RECV != (int)TCP_NEW_SYN_RECV);
2608 BUILD_BUG_ON((int)BPF_TCP_MAX_STATES != (int)TCP_MAX_STATES);
2610 /* bpf uapi header bpf.h defines an anonymous enum with values
2611 * BPF_TCP_* used by bpf programs. Currently gcc built vmlinux
2612 * is able to emit this enum in DWARF due to the above BUILD_BUG_ON.
2613 * But clang built vmlinux does not have this enum in DWARF
2614 * since clang removes the above code before generating IR/debuginfo.
2615 * Let us explicitly emit the type debuginfo to ensure the
2616 * above-mentioned anonymous enum in the vmlinux DWARF and hence BTF
2617 * regardless of which compiler is used.
2619 BTF_TYPE_EMIT_ENUM(BPF_TCP_ESTABLISHED);
2621 if (BPF_SOCK_OPS_TEST_FLAG(tcp_sk(sk), BPF_SOCK_OPS_STATE_CB_FLAG))
2622 tcp_call_bpf_2arg(sk, BPF_SOCK_OPS_STATE_CB, oldstate, state);
2625 case TCP_ESTABLISHED:
2626 if (oldstate != TCP_ESTABLISHED)
2627 TCP_INC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
2631 if (oldstate == TCP_CLOSE_WAIT || oldstate == TCP_ESTABLISHED)
2632 TCP_INC_STATS(sock_net(sk), TCP_MIB_ESTABRESETS);
2634 sk->sk_prot->unhash(sk);
2635 if (inet_csk(sk)->icsk_bind_hash &&
2636 !(sk->sk_userlocks & SOCK_BINDPORT_LOCK))
2640 if (oldstate == TCP_ESTABLISHED)
2641 TCP_DEC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
2644 /* Change state AFTER socket is unhashed to avoid closed
2645 * socket sitting in hash tables.
2647 inet_sk_state_store(sk, state);
2649 EXPORT_SYMBOL_GPL(tcp_set_state);
2652 * State processing on a close. This implements the state shift for
2653 * sending our FIN frame. Note that we only send a FIN for some
2654 * states. A shutdown() may have already sent the FIN, or we may be
2658 static const unsigned char new_state[16] = {
2659 /* current state: new state: action: */
2660 [0 /* (Invalid) */] = TCP_CLOSE,
2661 [TCP_ESTABLISHED] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2662 [TCP_SYN_SENT] = TCP_CLOSE,
2663 [TCP_SYN_RECV] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2664 [TCP_FIN_WAIT1] = TCP_FIN_WAIT1,
2665 [TCP_FIN_WAIT2] = TCP_FIN_WAIT2,
2666 [TCP_TIME_WAIT] = TCP_CLOSE,
2667 [TCP_CLOSE] = TCP_CLOSE,
2668 [TCP_CLOSE_WAIT] = TCP_LAST_ACK | TCP_ACTION_FIN,
2669 [TCP_LAST_ACK] = TCP_LAST_ACK,
2670 [TCP_LISTEN] = TCP_CLOSE,
2671 [TCP_CLOSING] = TCP_CLOSING,
2672 [TCP_NEW_SYN_RECV] = TCP_CLOSE, /* should not happen ! */
2675 static int tcp_close_state(struct sock *sk)
2677 int next = (int)new_state[sk->sk_state];
2678 int ns = next & TCP_STATE_MASK;
2680 tcp_set_state(sk, ns);
2682 return next & TCP_ACTION_FIN;
2686 * Shutdown the sending side of a connection. Much like close except
2687 * that we don't receive shut down or sock_set_flag(sk, SOCK_DEAD).
2690 void tcp_shutdown(struct sock *sk, int how)
2692 /* We need to grab some memory, and put together a FIN,
2693 * and then put it into the queue to be sent.
2694 * Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
2696 if (!(how & SEND_SHUTDOWN))
2699 /* If we've already sent a FIN, or it's a closed state, skip this. */
2700 if ((1 << sk->sk_state) &
2701 (TCPF_ESTABLISHED | TCPF_SYN_SENT |
2702 TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) {
2703 /* Clear out any half completed packets. FIN if needed. */
2704 if (tcp_close_state(sk))
2708 EXPORT_SYMBOL(tcp_shutdown);
2710 int tcp_orphan_count_sum(void)
2714 for_each_possible_cpu(i)
2715 total += per_cpu(tcp_orphan_count, i);
2717 return max(total, 0);
2720 static int tcp_orphan_cache;
2721 static struct timer_list tcp_orphan_timer;
2722 #define TCP_ORPHAN_TIMER_PERIOD msecs_to_jiffies(100)
2724 static void tcp_orphan_update(struct timer_list *unused)
2726 WRITE_ONCE(tcp_orphan_cache, tcp_orphan_count_sum());
2727 mod_timer(&tcp_orphan_timer, jiffies + TCP_ORPHAN_TIMER_PERIOD);
2730 static bool tcp_too_many_orphans(int shift)
2732 return READ_ONCE(tcp_orphan_cache) << shift >
2733 READ_ONCE(sysctl_tcp_max_orphans);
2736 bool tcp_check_oom(struct sock *sk, int shift)
2738 bool too_many_orphans, out_of_socket_memory;
2740 too_many_orphans = tcp_too_many_orphans(shift);
2741 out_of_socket_memory = tcp_out_of_memory(sk);
2743 if (too_many_orphans)
2744 net_info_ratelimited("too many orphaned sockets\n");
2745 if (out_of_socket_memory)
2746 net_info_ratelimited("out of memory -- consider tuning tcp_mem\n");
2747 return too_many_orphans || out_of_socket_memory;
2750 void __tcp_close(struct sock *sk, long timeout)
2752 struct sk_buff *skb;
2753 int data_was_unread = 0;
2756 WRITE_ONCE(sk->sk_shutdown, SHUTDOWN_MASK);
2758 if (sk->sk_state == TCP_LISTEN) {
2759 tcp_set_state(sk, TCP_CLOSE);
2762 inet_csk_listen_stop(sk);
2764 goto adjudge_to_death;
2767 /* We need to flush the recv. buffs. We do this only on the
2768 * descriptor close, not protocol-sourced closes, because the
2769 * reader process may not have drained the data yet!
2771 while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) {
2772 u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq;
2774 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
2776 data_was_unread += len;
2780 /* If socket has been already reset (e.g. in tcp_reset()) - kill it. */
2781 if (sk->sk_state == TCP_CLOSE)
2782 goto adjudge_to_death;
2784 /* As outlined in RFC 2525, section 2.17, we send a RST here because
2785 * data was lost. To witness the awful effects of the old behavior of
2786 * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk
2787 * GET in an FTP client, suspend the process, wait for the client to
2788 * advertise a zero window, then kill -9 the FTP client, wheee...
2789 * Note: timeout is always zero in such a case.
2791 if (unlikely(tcp_sk(sk)->repair)) {
2792 sk->sk_prot->disconnect(sk, 0);
2793 } else if (data_was_unread) {
2794 /* Unread data was tossed, zap the connection. */
2795 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONCLOSE);
2796 tcp_set_state(sk, TCP_CLOSE);
2797 tcp_send_active_reset(sk, sk->sk_allocation);
2798 } else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) {
2799 /* Check zero linger _after_ checking for unread data. */
2800 sk->sk_prot->disconnect(sk, 0);
2801 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
2802 } else if (tcp_close_state(sk)) {
2803 /* We FIN if the application ate all the data before
2804 * zapping the connection.
2807 /* RED-PEN. Formally speaking, we have broken TCP state
2808 * machine. State transitions:
2810 * TCP_ESTABLISHED -> TCP_FIN_WAIT1
2811 * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible)
2812 * TCP_CLOSE_WAIT -> TCP_LAST_ACK
2814 * are legal only when FIN has been sent (i.e. in window),
2815 * rather than queued out of window. Purists blame.
2817 * F.e. "RFC state" is ESTABLISHED,
2818 * if Linux state is FIN-WAIT-1, but FIN is still not sent.
2820 * The visible declinations are that sometimes
2821 * we enter time-wait state, when it is not required really
2822 * (harmless), do not send active resets, when they are
2823 * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
2824 * they look as CLOSING or LAST_ACK for Linux)
2825 * Probably, I missed some more holelets.
2827 * XXX (TFO) - To start off we don't support SYN+ACK+FIN
2828 * in a single packet! (May consider it later but will
2829 * probably need API support or TCP_CORK SYN-ACK until
2830 * data is written and socket is closed.)
2835 sk_stream_wait_close(sk, timeout);
2838 state = sk->sk_state;
2844 /* remove backlog if any, without releasing ownership. */
2847 this_cpu_inc(tcp_orphan_count);
2849 /* Have we already been destroyed by a softirq or backlog? */
2850 if (state != TCP_CLOSE && sk->sk_state == TCP_CLOSE)
2853 /* This is a (useful) BSD violating of the RFC. There is a
2854 * problem with TCP as specified in that the other end could
2855 * keep a socket open forever with no application left this end.
2856 * We use a 1 minute timeout (about the same as BSD) then kill
2857 * our end. If they send after that then tough - BUT: long enough
2858 * that we won't make the old 4*rto = almost no time - whoops
2861 * Nope, it was not mistake. It is really desired behaviour
2862 * f.e. on http servers, when such sockets are useless, but
2863 * consume significant resources. Let's do it with special
2864 * linger2 option. --ANK
2867 if (sk->sk_state == TCP_FIN_WAIT2) {
2868 struct tcp_sock *tp = tcp_sk(sk);
2869 if (READ_ONCE(tp->linger2) < 0) {
2870 tcp_set_state(sk, TCP_CLOSE);
2871 tcp_send_active_reset(sk, GFP_ATOMIC);
2872 __NET_INC_STATS(sock_net(sk),
2873 LINUX_MIB_TCPABORTONLINGER);
2875 const int tmo = tcp_fin_time(sk);
2877 if (tmo > TCP_TIMEWAIT_LEN) {
2878 inet_csk_reset_keepalive_timer(sk,
2879 tmo - TCP_TIMEWAIT_LEN);
2881 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
2886 if (sk->sk_state != TCP_CLOSE) {
2887 if (tcp_check_oom(sk, 0)) {
2888 tcp_set_state(sk, TCP_CLOSE);
2889 tcp_send_active_reset(sk, GFP_ATOMIC);
2890 __NET_INC_STATS(sock_net(sk),
2891 LINUX_MIB_TCPABORTONMEMORY);
2892 } else if (!check_net(sock_net(sk))) {
2893 /* Not possible to send reset; just close */
2894 tcp_set_state(sk, TCP_CLOSE);
2898 if (sk->sk_state == TCP_CLOSE) {
2899 struct request_sock *req;
2901 req = rcu_dereference_protected(tcp_sk(sk)->fastopen_rsk,
2902 lockdep_sock_is_held(sk));
2903 /* We could get here with a non-NULL req if the socket is
2904 * aborted (e.g., closed with unread data) before 3WHS
2908 reqsk_fastopen_remove(sk, req, false);
2909 inet_csk_destroy_sock(sk);
2911 /* Otherwise, socket is reprieved until protocol close. */
2918 void tcp_close(struct sock *sk, long timeout)
2921 __tcp_close(sk, timeout);
2925 EXPORT_SYMBOL(tcp_close);
2927 /* These states need RST on ABORT according to RFC793 */
2929 static inline bool tcp_need_reset(int state)
2931 return (1 << state) &
2932 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 |
2933 TCPF_FIN_WAIT2 | TCPF_SYN_RECV);
2936 static void tcp_rtx_queue_purge(struct sock *sk)
2938 struct rb_node *p = rb_first(&sk->tcp_rtx_queue);
2940 tcp_sk(sk)->highest_sack = NULL;
2942 struct sk_buff *skb = rb_to_skb(p);
2945 /* Since we are deleting whole queue, no need to
2946 * list_del(&skb->tcp_tsorted_anchor)
2948 tcp_rtx_queue_unlink(skb, sk);
2949 tcp_wmem_free_skb(sk, skb);
2953 void tcp_write_queue_purge(struct sock *sk)
2955 struct sk_buff *skb;
2957 tcp_chrono_stop(sk, TCP_CHRONO_BUSY);
2958 while ((skb = __skb_dequeue(&sk->sk_write_queue)) != NULL) {
2959 tcp_skb_tsorted_anchor_cleanup(skb);
2960 tcp_wmem_free_skb(sk, skb);
2962 tcp_rtx_queue_purge(sk);
2963 INIT_LIST_HEAD(&tcp_sk(sk)->tsorted_sent_queue);
2964 tcp_clear_all_retrans_hints(tcp_sk(sk));
2965 tcp_sk(sk)->packets_out = 0;
2966 inet_csk(sk)->icsk_backoff = 0;
2969 int tcp_disconnect(struct sock *sk, int flags)
2971 struct inet_sock *inet = inet_sk(sk);
2972 struct inet_connection_sock *icsk = inet_csk(sk);
2973 struct tcp_sock *tp = tcp_sk(sk);
2974 int old_state = sk->sk_state;
2977 if (old_state != TCP_CLOSE)
2978 tcp_set_state(sk, TCP_CLOSE);
2980 /* ABORT function of RFC793 */
2981 if (old_state == TCP_LISTEN) {
2982 inet_csk_listen_stop(sk);
2983 } else if (unlikely(tp->repair)) {
2984 WRITE_ONCE(sk->sk_err, ECONNABORTED);
2985 } else if (tcp_need_reset(old_state) ||
2986 (tp->snd_nxt != tp->write_seq &&
2987 (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK))) {
2988 /* The last check adjusts for discrepancy of Linux wrt. RFC
2991 tcp_send_active_reset(sk, gfp_any());
2992 WRITE_ONCE(sk->sk_err, ECONNRESET);
2993 } else if (old_state == TCP_SYN_SENT)
2994 WRITE_ONCE(sk->sk_err, ECONNRESET);
2996 tcp_clear_xmit_timers(sk);
2997 __skb_queue_purge(&sk->sk_receive_queue);
2998 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt);
2999 WRITE_ONCE(tp->urg_data, 0);
3000 tcp_write_queue_purge(sk);
3001 tcp_fastopen_active_disable_ofo_check(sk);
3002 skb_rbtree_purge(&tp->out_of_order_queue);
3004 inet->inet_dport = 0;
3006 inet_bhash2_reset_saddr(sk);
3008 WRITE_ONCE(sk->sk_shutdown, 0);
3009 sock_reset_flag(sk, SOCK_DONE);
3011 tp->mdev_us = jiffies_to_usecs(TCP_TIMEOUT_INIT);
3012 tp->rcv_rtt_last_tsecr = 0;
3014 seq = tp->write_seq + tp->max_window + 2;
3017 WRITE_ONCE(tp->write_seq, seq);
3019 icsk->icsk_backoff = 0;
3020 icsk->icsk_probes_out = 0;
3021 icsk->icsk_probes_tstamp = 0;
3022 icsk->icsk_rto = TCP_TIMEOUT_INIT;
3023 icsk->icsk_rto_min = TCP_RTO_MIN;
3024 icsk->icsk_delack_max = TCP_DELACK_MAX;
3025 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
3026 tcp_snd_cwnd_set(tp, TCP_INIT_CWND);
3027 tp->snd_cwnd_cnt = 0;
3028 tp->is_cwnd_limited = 0;
3029 tp->max_packets_out = 0;
3030 tp->window_clamp = 0;
3032 tp->delivered_ce = 0;
3033 if (icsk->icsk_ca_ops->release)
3034 icsk->icsk_ca_ops->release(sk);
3035 memset(icsk->icsk_ca_priv, 0, sizeof(icsk->icsk_ca_priv));
3036 icsk->icsk_ca_initialized = 0;
3037 tcp_set_ca_state(sk, TCP_CA_Open);
3038 tp->is_sack_reneg = 0;
3039 tcp_clear_retrans(tp);
3040 tp->total_retrans = 0;
3041 inet_csk_delack_init(sk);
3042 /* Initialize rcv_mss to TCP_MIN_MSS to avoid division by 0
3043 * issue in __tcp_select_window()
3045 icsk->icsk_ack.rcv_mss = TCP_MIN_MSS;
3046 memset(&tp->rx_opt, 0, sizeof(tp->rx_opt));
3048 dst_release(xchg((__force struct dst_entry **)&sk->sk_rx_dst, NULL));
3049 tcp_saved_syn_free(tp);
3050 tp->compressed_ack = 0;
3054 tp->bytes_acked = 0;
3055 tp->bytes_received = 0;
3056 tp->bytes_retrans = 0;
3057 tp->data_segs_in = 0;
3058 tp->data_segs_out = 0;
3059 tp->duplicate_sack[0].start_seq = 0;
3060 tp->duplicate_sack[0].end_seq = 0;
3063 tp->retrans_out = 0;
3065 tp->tlp_high_seq = 0;
3066 tp->last_oow_ack_time = 0;
3068 /* There's a bubble in the pipe until at least the first ACK. */
3069 tp->app_limited = ~0U;
3070 tp->rate_app_limited = 1;
3071 tp->rack.mstamp = 0;
3072 tp->rack.advanced = 0;
3073 tp->rack.reo_wnd_steps = 1;
3074 tp->rack.last_delivered = 0;
3075 tp->rack.reo_wnd_persist = 0;
3076 tp->rack.dsack_seen = 0;
3077 tp->syn_data_acked = 0;
3078 tp->rx_opt.saw_tstamp = 0;
3079 tp->rx_opt.dsack = 0;
3080 tp->rx_opt.num_sacks = 0;
3081 tp->rcv_ooopack = 0;
3084 /* Clean up fastopen related fields */
3085 tcp_free_fastopen_req(tp);
3086 inet_clear_bit(DEFER_CONNECT, sk);
3087 tp->fastopen_client_fail = 0;
3089 WARN_ON(inet->inet_num && !icsk->icsk_bind_hash);
3091 if (sk->sk_frag.page) {
3092 put_page(sk->sk_frag.page);
3093 sk->sk_frag.page = NULL;
3094 sk->sk_frag.offset = 0;
3096 sk_error_report(sk);
3099 EXPORT_SYMBOL(tcp_disconnect);
3101 static inline bool tcp_can_repair_sock(const struct sock *sk)
3103 return sockopt_ns_capable(sock_net(sk)->user_ns, CAP_NET_ADMIN) &&
3104 (sk->sk_state != TCP_LISTEN);
3107 static int tcp_repair_set_window(struct tcp_sock *tp, sockptr_t optbuf, int len)
3109 struct tcp_repair_window opt;
3114 if (len != sizeof(opt))
3117 if (copy_from_sockptr(&opt, optbuf, sizeof(opt)))
3120 if (opt.max_window < opt.snd_wnd)
3123 if (after(opt.snd_wl1, tp->rcv_nxt + opt.rcv_wnd))
3126 if (after(opt.rcv_wup, tp->rcv_nxt))
3129 tp->snd_wl1 = opt.snd_wl1;
3130 tp->snd_wnd = opt.snd_wnd;
3131 tp->max_window = opt.max_window;
3133 tp->rcv_wnd = opt.rcv_wnd;
3134 tp->rcv_wup = opt.rcv_wup;
3139 static int tcp_repair_options_est(struct sock *sk, sockptr_t optbuf,
3142 struct tcp_sock *tp = tcp_sk(sk);
3143 struct tcp_repair_opt opt;
3146 while (len >= sizeof(opt)) {
3147 if (copy_from_sockptr_offset(&opt, optbuf, offset, sizeof(opt)))
3150 offset += sizeof(opt);
3153 switch (opt.opt_code) {
3155 tp->rx_opt.mss_clamp = opt.opt_val;
3160 u16 snd_wscale = opt.opt_val & 0xFFFF;
3161 u16 rcv_wscale = opt.opt_val >> 16;
3163 if (snd_wscale > TCP_MAX_WSCALE || rcv_wscale > TCP_MAX_WSCALE)
3166 tp->rx_opt.snd_wscale = snd_wscale;
3167 tp->rx_opt.rcv_wscale = rcv_wscale;
3168 tp->rx_opt.wscale_ok = 1;
3171 case TCPOPT_SACK_PERM:
3172 if (opt.opt_val != 0)
3175 tp->rx_opt.sack_ok |= TCP_SACK_SEEN;
3177 case TCPOPT_TIMESTAMP:
3178 if (opt.opt_val != 0)
3181 tp->rx_opt.tstamp_ok = 1;
3189 DEFINE_STATIC_KEY_FALSE(tcp_tx_delay_enabled);
3190 EXPORT_SYMBOL(tcp_tx_delay_enabled);
3192 static void tcp_enable_tx_delay(void)
3194 if (!static_branch_unlikely(&tcp_tx_delay_enabled)) {
3195 static int __tcp_tx_delay_enabled = 0;
3197 if (cmpxchg(&__tcp_tx_delay_enabled, 0, 1) == 0) {
3198 static_branch_enable(&tcp_tx_delay_enabled);
3199 pr_info("TCP_TX_DELAY enabled\n");
3204 /* When set indicates to always queue non-full frames. Later the user clears
3205 * this option and we transmit any pending partial frames in the queue. This is
3206 * meant to be used alongside sendfile() to get properly filled frames when the
3207 * user (for example) must write out headers with a write() call first and then
3208 * use sendfile to send out the data parts.
3210 * TCP_CORK can be set together with TCP_NODELAY and it is stronger than
3213 void __tcp_sock_set_cork(struct sock *sk, bool on)
3215 struct tcp_sock *tp = tcp_sk(sk);
3218 tp->nonagle |= TCP_NAGLE_CORK;
3220 tp->nonagle &= ~TCP_NAGLE_CORK;
3221 if (tp->nonagle & TCP_NAGLE_OFF)
3222 tp->nonagle |= TCP_NAGLE_PUSH;
3223 tcp_push_pending_frames(sk);
3227 void tcp_sock_set_cork(struct sock *sk, bool on)
3230 __tcp_sock_set_cork(sk, on);
3233 EXPORT_SYMBOL(tcp_sock_set_cork);
3235 /* TCP_NODELAY is weaker than TCP_CORK, so that this option on corked socket is
3236 * remembered, but it is not activated until cork is cleared.
3238 * However, when TCP_NODELAY is set we make an explicit push, which overrides
3239 * even TCP_CORK for currently queued segments.
3241 void __tcp_sock_set_nodelay(struct sock *sk, bool on)
3244 tcp_sk(sk)->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH;
3245 tcp_push_pending_frames(sk);
3247 tcp_sk(sk)->nonagle &= ~TCP_NAGLE_OFF;
3251 void tcp_sock_set_nodelay(struct sock *sk)
3254 __tcp_sock_set_nodelay(sk, true);
3257 EXPORT_SYMBOL(tcp_sock_set_nodelay);
3259 static void __tcp_sock_set_quickack(struct sock *sk, int val)
3262 inet_csk_enter_pingpong_mode(sk);
3266 inet_csk_exit_pingpong_mode(sk);
3267 if ((1 << sk->sk_state) & (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) &&
3268 inet_csk_ack_scheduled(sk)) {
3269 inet_csk(sk)->icsk_ack.pending |= ICSK_ACK_PUSHED;
3270 tcp_cleanup_rbuf(sk, 1);
3272 inet_csk_enter_pingpong_mode(sk);
3276 void tcp_sock_set_quickack(struct sock *sk, int val)
3279 __tcp_sock_set_quickack(sk, val);
3282 EXPORT_SYMBOL(tcp_sock_set_quickack);
3284 int tcp_sock_set_syncnt(struct sock *sk, int val)
3286 if (val < 1 || val > MAX_TCP_SYNCNT)
3289 WRITE_ONCE(inet_csk(sk)->icsk_syn_retries, val);
3292 EXPORT_SYMBOL(tcp_sock_set_syncnt);
3294 int tcp_sock_set_user_timeout(struct sock *sk, int val)
3296 /* Cap the max time in ms TCP will retry or probe the window
3297 * before giving up and aborting (ETIMEDOUT) a connection.
3302 WRITE_ONCE(inet_csk(sk)->icsk_user_timeout, val);
3305 EXPORT_SYMBOL(tcp_sock_set_user_timeout);
3307 int tcp_sock_set_keepidle_locked(struct sock *sk, int val)
3309 struct tcp_sock *tp = tcp_sk(sk);
3311 if (val < 1 || val > MAX_TCP_KEEPIDLE)
3314 /* Paired with WRITE_ONCE() in keepalive_time_when() */
3315 WRITE_ONCE(tp->keepalive_time, val * HZ);
3316 if (sock_flag(sk, SOCK_KEEPOPEN) &&
3317 !((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN))) {
3318 u32 elapsed = keepalive_time_elapsed(tp);
3320 if (tp->keepalive_time > elapsed)
3321 elapsed = tp->keepalive_time - elapsed;
3324 inet_csk_reset_keepalive_timer(sk, elapsed);
3330 int tcp_sock_set_keepidle(struct sock *sk, int val)
3335 err = tcp_sock_set_keepidle_locked(sk, val);
3339 EXPORT_SYMBOL(tcp_sock_set_keepidle);
3341 int tcp_sock_set_keepintvl(struct sock *sk, int val)
3343 if (val < 1 || val > MAX_TCP_KEEPINTVL)
3346 WRITE_ONCE(tcp_sk(sk)->keepalive_intvl, val * HZ);
3349 EXPORT_SYMBOL(tcp_sock_set_keepintvl);
3351 int tcp_sock_set_keepcnt(struct sock *sk, int val)
3353 if (val < 1 || val > MAX_TCP_KEEPCNT)
3356 /* Paired with READ_ONCE() in keepalive_probes() */
3357 WRITE_ONCE(tcp_sk(sk)->keepalive_probes, val);
3360 EXPORT_SYMBOL(tcp_sock_set_keepcnt);
3362 int tcp_set_window_clamp(struct sock *sk, int val)
3364 struct tcp_sock *tp = tcp_sk(sk);
3367 if (sk->sk_state != TCP_CLOSE)
3369 tp->window_clamp = 0;
3371 tp->window_clamp = val < SOCK_MIN_RCVBUF / 2 ?
3372 SOCK_MIN_RCVBUF / 2 : val;
3373 tp->rcv_ssthresh = min(tp->rcv_wnd, tp->window_clamp);
3379 * Socket option code for TCP.
3381 int do_tcp_setsockopt(struct sock *sk, int level, int optname,
3382 sockptr_t optval, unsigned int optlen)
3384 struct tcp_sock *tp = tcp_sk(sk);
3385 struct inet_connection_sock *icsk = inet_csk(sk);
3386 struct net *net = sock_net(sk);
3390 /* These are data/string values, all the others are ints */
3392 case TCP_CONGESTION: {
3393 char name[TCP_CA_NAME_MAX];
3398 val = strncpy_from_sockptr(name, optval,
3399 min_t(long, TCP_CA_NAME_MAX-1, optlen));
3404 sockopt_lock_sock(sk);
3405 err = tcp_set_congestion_control(sk, name, !has_current_bpf_ctx(),
3406 sockopt_ns_capable(sock_net(sk)->user_ns,
3408 sockopt_release_sock(sk);
3412 char name[TCP_ULP_NAME_MAX];
3417 val = strncpy_from_sockptr(name, optval,
3418 min_t(long, TCP_ULP_NAME_MAX - 1,
3424 sockopt_lock_sock(sk);
3425 err = tcp_set_ulp(sk, name);
3426 sockopt_release_sock(sk);
3429 case TCP_FASTOPEN_KEY: {
3430 __u8 key[TCP_FASTOPEN_KEY_BUF_LENGTH];
3431 __u8 *backup_key = NULL;
3433 /* Allow a backup key as well to facilitate key rotation
3434 * First key is the active one.
3436 if (optlen != TCP_FASTOPEN_KEY_LENGTH &&
3437 optlen != TCP_FASTOPEN_KEY_BUF_LENGTH)
3440 if (copy_from_sockptr(key, optval, optlen))
3443 if (optlen == TCP_FASTOPEN_KEY_BUF_LENGTH)
3444 backup_key = key + TCP_FASTOPEN_KEY_LENGTH;
3446 return tcp_fastopen_reset_cipher(net, sk, key, backup_key);
3453 if (optlen < sizeof(int))
3456 if (copy_from_sockptr(&val, optval, sizeof(val)))
3459 /* Handle options that can be set without locking the socket. */
3462 return tcp_sock_set_syncnt(sk, val);
3463 case TCP_USER_TIMEOUT:
3464 return tcp_sock_set_user_timeout(sk, val);
3466 return tcp_sock_set_keepintvl(sk, val);
3468 return tcp_sock_set_keepcnt(sk, val);
3471 WRITE_ONCE(tp->linger2, -1);
3472 else if (val > TCP_FIN_TIMEOUT_MAX / HZ)
3473 WRITE_ONCE(tp->linger2, TCP_FIN_TIMEOUT_MAX);
3475 WRITE_ONCE(tp->linger2, val * HZ);
3477 case TCP_DEFER_ACCEPT:
3478 /* Translate value in seconds to number of retransmits */
3479 WRITE_ONCE(icsk->icsk_accept_queue.rskq_defer_accept,
3480 secs_to_retrans(val, TCP_TIMEOUT_INIT / HZ,
3485 sockopt_lock_sock(sk);
3489 /* Values greater than interface MTU won't take effect. However
3490 * at the point when this call is done we typically don't yet
3491 * know which interface is going to be used
3493 if (val && (val < TCP_MIN_MSS || val > MAX_TCP_WINDOW)) {
3497 tp->rx_opt.user_mss = val;
3501 __tcp_sock_set_nodelay(sk, val);
3504 case TCP_THIN_LINEAR_TIMEOUTS:
3505 if (val < 0 || val > 1)
3511 case TCP_THIN_DUPACK:
3512 if (val < 0 || val > 1)
3517 if (!tcp_can_repair_sock(sk))
3519 else if (val == TCP_REPAIR_ON) {
3521 sk->sk_reuse = SK_FORCE_REUSE;
3522 tp->repair_queue = TCP_NO_QUEUE;
3523 } else if (val == TCP_REPAIR_OFF) {
3525 sk->sk_reuse = SK_NO_REUSE;
3526 tcp_send_window_probe(sk);
3527 } else if (val == TCP_REPAIR_OFF_NO_WP) {
3529 sk->sk_reuse = SK_NO_REUSE;
3535 case TCP_REPAIR_QUEUE:
3538 else if ((unsigned int)val < TCP_QUEUES_NR)
3539 tp->repair_queue = val;
3545 if (sk->sk_state != TCP_CLOSE) {
3547 } else if (tp->repair_queue == TCP_SEND_QUEUE) {
3548 if (!tcp_rtx_queue_empty(sk))
3551 WRITE_ONCE(tp->write_seq, val);
3552 } else if (tp->repair_queue == TCP_RECV_QUEUE) {
3553 if (tp->rcv_nxt != tp->copied_seq) {
3556 WRITE_ONCE(tp->rcv_nxt, val);
3557 WRITE_ONCE(tp->copied_seq, val);
3564 case TCP_REPAIR_OPTIONS:
3567 else if (sk->sk_state == TCP_ESTABLISHED && !tp->bytes_sent)
3568 err = tcp_repair_options_est(sk, optval, optlen);
3574 __tcp_sock_set_cork(sk, val);
3578 err = tcp_sock_set_keepidle_locked(sk, val);
3581 /* 0: disable, 1: enable, 2: start from ether_header */
3582 if (val < 0 || val > 2)
3588 case TCP_WINDOW_CLAMP:
3589 err = tcp_set_window_clamp(sk, val);
3593 __tcp_sock_set_quickack(sk, val);
3596 #ifdef CONFIG_TCP_MD5SIG
3598 case TCP_MD5SIG_EXT:
3599 err = tp->af_specific->md5_parse(sk, optname, optval, optlen);
3603 if (val >= 0 && ((1 << sk->sk_state) & (TCPF_CLOSE |
3605 tcp_fastopen_init_key_once(net);
3607 fastopen_queue_tune(sk, val);
3612 case TCP_FASTOPEN_CONNECT:
3613 if (val > 1 || val < 0) {
3615 } else if (READ_ONCE(net->ipv4.sysctl_tcp_fastopen) &
3616 TFO_CLIENT_ENABLE) {
3617 if (sk->sk_state == TCP_CLOSE)
3618 tp->fastopen_connect = val;
3625 case TCP_FASTOPEN_NO_COOKIE:
3626 if (val > 1 || val < 0)
3628 else if (!((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
3631 tp->fastopen_no_cookie = val;
3637 WRITE_ONCE(tp->tsoffset, val - tcp_time_stamp_raw());
3639 case TCP_REPAIR_WINDOW:
3640 err = tcp_repair_set_window(tp, optval, optlen);
3642 case TCP_NOTSENT_LOWAT:
3643 WRITE_ONCE(tp->notsent_lowat, val);
3644 sk->sk_write_space(sk);
3647 if (val > 1 || val < 0)
3650 tp->recvmsg_inq = val;
3654 tcp_enable_tx_delay();
3655 WRITE_ONCE(tp->tcp_tx_delay, val);
3662 sockopt_release_sock(sk);
3666 int tcp_setsockopt(struct sock *sk, int level, int optname, sockptr_t optval,
3667 unsigned int optlen)
3669 const struct inet_connection_sock *icsk = inet_csk(sk);
3671 if (level != SOL_TCP)
3672 /* Paired with WRITE_ONCE() in do_ipv6_setsockopt() and tcp_v6_connect() */
3673 return READ_ONCE(icsk->icsk_af_ops)->setsockopt(sk, level, optname,
3675 return do_tcp_setsockopt(sk, level, optname, optval, optlen);
3677 EXPORT_SYMBOL(tcp_setsockopt);
3679 static void tcp_get_info_chrono_stats(const struct tcp_sock *tp,
3680 struct tcp_info *info)
3682 u64 stats[__TCP_CHRONO_MAX], total = 0;
3685 for (i = TCP_CHRONO_BUSY; i < __TCP_CHRONO_MAX; ++i) {
3686 stats[i] = tp->chrono_stat[i - 1];
3687 if (i == tp->chrono_type)
3688 stats[i] += tcp_jiffies32 - tp->chrono_start;
3689 stats[i] *= USEC_PER_SEC / HZ;
3693 info->tcpi_busy_time = total;
3694 info->tcpi_rwnd_limited = stats[TCP_CHRONO_RWND_LIMITED];
3695 info->tcpi_sndbuf_limited = stats[TCP_CHRONO_SNDBUF_LIMITED];
3698 /* Return information about state of tcp endpoint in API format. */
3699 void tcp_get_info(struct sock *sk, struct tcp_info *info)
3701 const struct tcp_sock *tp = tcp_sk(sk); /* iff sk_type == SOCK_STREAM */
3702 const struct inet_connection_sock *icsk = inet_csk(sk);
3708 memset(info, 0, sizeof(*info));
3709 if (sk->sk_type != SOCK_STREAM)
3712 info->tcpi_state = inet_sk_state_load(sk);
3714 /* Report meaningful fields for all TCP states, including listeners */
3715 rate = READ_ONCE(sk->sk_pacing_rate);
3716 rate64 = (rate != ~0UL) ? rate : ~0ULL;
3717 info->tcpi_pacing_rate = rate64;
3719 rate = READ_ONCE(sk->sk_max_pacing_rate);
3720 rate64 = (rate != ~0UL) ? rate : ~0ULL;
3721 info->tcpi_max_pacing_rate = rate64;
3723 info->tcpi_reordering = tp->reordering;
3724 info->tcpi_snd_cwnd = tcp_snd_cwnd(tp);
3726 if (info->tcpi_state == TCP_LISTEN) {
3727 /* listeners aliased fields :
3728 * tcpi_unacked -> Number of children ready for accept()
3729 * tcpi_sacked -> max backlog
3731 info->tcpi_unacked = READ_ONCE(sk->sk_ack_backlog);
3732 info->tcpi_sacked = READ_ONCE(sk->sk_max_ack_backlog);
3736 slow = lock_sock_fast(sk);
3738 info->tcpi_ca_state = icsk->icsk_ca_state;
3739 info->tcpi_retransmits = icsk->icsk_retransmits;
3740 info->tcpi_probes = icsk->icsk_probes_out;
3741 info->tcpi_backoff = icsk->icsk_backoff;
3743 if (tp->rx_opt.tstamp_ok)
3744 info->tcpi_options |= TCPI_OPT_TIMESTAMPS;
3745 if (tcp_is_sack(tp))
3746 info->tcpi_options |= TCPI_OPT_SACK;
3747 if (tp->rx_opt.wscale_ok) {
3748 info->tcpi_options |= TCPI_OPT_WSCALE;
3749 info->tcpi_snd_wscale = tp->rx_opt.snd_wscale;
3750 info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale;
3753 if (tp->ecn_flags & TCP_ECN_OK)
3754 info->tcpi_options |= TCPI_OPT_ECN;
3755 if (tp->ecn_flags & TCP_ECN_SEEN)
3756 info->tcpi_options |= TCPI_OPT_ECN_SEEN;
3757 if (tp->syn_data_acked)
3758 info->tcpi_options |= TCPI_OPT_SYN_DATA;
3760 info->tcpi_rto = jiffies_to_usecs(icsk->icsk_rto);
3761 info->tcpi_ato = jiffies_to_usecs(icsk->icsk_ack.ato);
3762 info->tcpi_snd_mss = tp->mss_cache;
3763 info->tcpi_rcv_mss = icsk->icsk_ack.rcv_mss;
3765 info->tcpi_unacked = tp->packets_out;
3766 info->tcpi_sacked = tp->sacked_out;
3768 info->tcpi_lost = tp->lost_out;
3769 info->tcpi_retrans = tp->retrans_out;
3771 now = tcp_jiffies32;
3772 info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime);
3773 info->tcpi_last_data_recv = jiffies_to_msecs(now - icsk->icsk_ack.lrcvtime);
3774 info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp);
3776 info->tcpi_pmtu = icsk->icsk_pmtu_cookie;
3777 info->tcpi_rcv_ssthresh = tp->rcv_ssthresh;
3778 info->tcpi_rtt = tp->srtt_us >> 3;
3779 info->tcpi_rttvar = tp->mdev_us >> 2;
3780 info->tcpi_snd_ssthresh = tp->snd_ssthresh;
3781 info->tcpi_advmss = tp->advmss;
3783 info->tcpi_rcv_rtt = tp->rcv_rtt_est.rtt_us >> 3;
3784 info->tcpi_rcv_space = tp->rcvq_space.space;
3786 info->tcpi_total_retrans = tp->total_retrans;
3788 info->tcpi_bytes_acked = tp->bytes_acked;
3789 info->tcpi_bytes_received = tp->bytes_received;
3790 info->tcpi_notsent_bytes = max_t(int, 0, tp->write_seq - tp->snd_nxt);
3791 tcp_get_info_chrono_stats(tp, info);
3793 info->tcpi_segs_out = tp->segs_out;
3795 /* segs_in and data_segs_in can be updated from tcp_segs_in() from BH */
3796 info->tcpi_segs_in = READ_ONCE(tp->segs_in);
3797 info->tcpi_data_segs_in = READ_ONCE(tp->data_segs_in);
3799 info->tcpi_min_rtt = tcp_min_rtt(tp);
3800 info->tcpi_data_segs_out = tp->data_segs_out;
3802 info->tcpi_delivery_rate_app_limited = tp->rate_app_limited ? 1 : 0;
3803 rate64 = tcp_compute_delivery_rate(tp);
3805 info->tcpi_delivery_rate = rate64;
3806 info->tcpi_delivered = tp->delivered;
3807 info->tcpi_delivered_ce = tp->delivered_ce;
3808 info->tcpi_bytes_sent = tp->bytes_sent;
3809 info->tcpi_bytes_retrans = tp->bytes_retrans;
3810 info->tcpi_dsack_dups = tp->dsack_dups;
3811 info->tcpi_reord_seen = tp->reord_seen;
3812 info->tcpi_rcv_ooopack = tp->rcv_ooopack;
3813 info->tcpi_snd_wnd = tp->snd_wnd;
3814 info->tcpi_rcv_wnd = tp->rcv_wnd;
3815 info->tcpi_rehash = tp->plb_rehash + tp->timeout_rehash;
3816 info->tcpi_fastopen_client_fail = tp->fastopen_client_fail;
3817 unlock_sock_fast(sk, slow);
3819 EXPORT_SYMBOL_GPL(tcp_get_info);
3821 static size_t tcp_opt_stats_get_size(void)
3824 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_BUSY */
3825 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_RWND_LIMITED */
3826 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_SNDBUF_LIMITED */
3827 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_DATA_SEGS_OUT */
3828 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_TOTAL_RETRANS */
3829 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_PACING_RATE */
3830 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_DELIVERY_RATE */
3831 nla_total_size(sizeof(u32)) + /* TCP_NLA_SND_CWND */
3832 nla_total_size(sizeof(u32)) + /* TCP_NLA_REORDERING */
3833 nla_total_size(sizeof(u32)) + /* TCP_NLA_MIN_RTT */
3834 nla_total_size(sizeof(u8)) + /* TCP_NLA_RECUR_RETRANS */
3835 nla_total_size(sizeof(u8)) + /* TCP_NLA_DELIVERY_RATE_APP_LMT */
3836 nla_total_size(sizeof(u32)) + /* TCP_NLA_SNDQ_SIZE */
3837 nla_total_size(sizeof(u8)) + /* TCP_NLA_CA_STATE */
3838 nla_total_size(sizeof(u32)) + /* TCP_NLA_SND_SSTHRESH */
3839 nla_total_size(sizeof(u32)) + /* TCP_NLA_DELIVERED */
3840 nla_total_size(sizeof(u32)) + /* TCP_NLA_DELIVERED_CE */
3841 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_BYTES_SENT */
3842 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_BYTES_RETRANS */
3843 nla_total_size(sizeof(u32)) + /* TCP_NLA_DSACK_DUPS */
3844 nla_total_size(sizeof(u32)) + /* TCP_NLA_REORD_SEEN */
3845 nla_total_size(sizeof(u32)) + /* TCP_NLA_SRTT */
3846 nla_total_size(sizeof(u16)) + /* TCP_NLA_TIMEOUT_REHASH */
3847 nla_total_size(sizeof(u32)) + /* TCP_NLA_BYTES_NOTSENT */
3848 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_EDT */
3849 nla_total_size(sizeof(u8)) + /* TCP_NLA_TTL */
3850 nla_total_size(sizeof(u32)) + /* TCP_NLA_REHASH */
3854 /* Returns TTL or hop limit of an incoming packet from skb. */
3855 static u8 tcp_skb_ttl_or_hop_limit(const struct sk_buff *skb)
3857 if (skb->protocol == htons(ETH_P_IP))
3858 return ip_hdr(skb)->ttl;
3859 else if (skb->protocol == htons(ETH_P_IPV6))
3860 return ipv6_hdr(skb)->hop_limit;
3865 struct sk_buff *tcp_get_timestamping_opt_stats(const struct sock *sk,
3866 const struct sk_buff *orig_skb,
3867 const struct sk_buff *ack_skb)
3869 const struct tcp_sock *tp = tcp_sk(sk);
3870 struct sk_buff *stats;
3871 struct tcp_info info;
3875 stats = alloc_skb(tcp_opt_stats_get_size(), GFP_ATOMIC);
3879 tcp_get_info_chrono_stats(tp, &info);
3880 nla_put_u64_64bit(stats, TCP_NLA_BUSY,
3881 info.tcpi_busy_time, TCP_NLA_PAD);
3882 nla_put_u64_64bit(stats, TCP_NLA_RWND_LIMITED,
3883 info.tcpi_rwnd_limited, TCP_NLA_PAD);
3884 nla_put_u64_64bit(stats, TCP_NLA_SNDBUF_LIMITED,
3885 info.tcpi_sndbuf_limited, TCP_NLA_PAD);
3886 nla_put_u64_64bit(stats, TCP_NLA_DATA_SEGS_OUT,
3887 tp->data_segs_out, TCP_NLA_PAD);
3888 nla_put_u64_64bit(stats, TCP_NLA_TOTAL_RETRANS,
3889 tp->total_retrans, TCP_NLA_PAD);
3891 rate = READ_ONCE(sk->sk_pacing_rate);
3892 rate64 = (rate != ~0UL) ? rate : ~0ULL;
3893 nla_put_u64_64bit(stats, TCP_NLA_PACING_RATE, rate64, TCP_NLA_PAD);
3895 rate64 = tcp_compute_delivery_rate(tp);
3896 nla_put_u64_64bit(stats, TCP_NLA_DELIVERY_RATE, rate64, TCP_NLA_PAD);
3898 nla_put_u32(stats, TCP_NLA_SND_CWND, tcp_snd_cwnd(tp));
3899 nla_put_u32(stats, TCP_NLA_REORDERING, tp->reordering);
3900 nla_put_u32(stats, TCP_NLA_MIN_RTT, tcp_min_rtt(tp));
3902 nla_put_u8(stats, TCP_NLA_RECUR_RETRANS, inet_csk(sk)->icsk_retransmits);
3903 nla_put_u8(stats, TCP_NLA_DELIVERY_RATE_APP_LMT, !!tp->rate_app_limited);
3904 nla_put_u32(stats, TCP_NLA_SND_SSTHRESH, tp->snd_ssthresh);
3905 nla_put_u32(stats, TCP_NLA_DELIVERED, tp->delivered);
3906 nla_put_u32(stats, TCP_NLA_DELIVERED_CE, tp->delivered_ce);
3908 nla_put_u32(stats, TCP_NLA_SNDQ_SIZE, tp->write_seq - tp->snd_una);
3909 nla_put_u8(stats, TCP_NLA_CA_STATE, inet_csk(sk)->icsk_ca_state);
3911 nla_put_u64_64bit(stats, TCP_NLA_BYTES_SENT, tp->bytes_sent,
3913 nla_put_u64_64bit(stats, TCP_NLA_BYTES_RETRANS, tp->bytes_retrans,
3915 nla_put_u32(stats, TCP_NLA_DSACK_DUPS, tp->dsack_dups);
3916 nla_put_u32(stats, TCP_NLA_REORD_SEEN, tp->reord_seen);
3917 nla_put_u32(stats, TCP_NLA_SRTT, tp->srtt_us >> 3);
3918 nla_put_u16(stats, TCP_NLA_TIMEOUT_REHASH, tp->timeout_rehash);
3919 nla_put_u32(stats, TCP_NLA_BYTES_NOTSENT,
3920 max_t(int, 0, tp->write_seq - tp->snd_nxt));
3921 nla_put_u64_64bit(stats, TCP_NLA_EDT, orig_skb->skb_mstamp_ns,
3924 nla_put_u8(stats, TCP_NLA_TTL,
3925 tcp_skb_ttl_or_hop_limit(ack_skb));
3927 nla_put_u32(stats, TCP_NLA_REHASH, tp->plb_rehash + tp->timeout_rehash);
3931 int do_tcp_getsockopt(struct sock *sk, int level,
3932 int optname, sockptr_t optval, sockptr_t optlen)
3934 struct inet_connection_sock *icsk = inet_csk(sk);
3935 struct tcp_sock *tp = tcp_sk(sk);
3936 struct net *net = sock_net(sk);
3939 if (copy_from_sockptr(&len, optlen, sizeof(int)))
3942 len = min_t(unsigned int, len, sizeof(int));
3949 val = tp->mss_cache;
3950 if (tp->rx_opt.user_mss &&
3951 ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
3952 val = tp->rx_opt.user_mss;
3954 val = tp->rx_opt.mss_clamp;
3957 val = !!(tp->nonagle&TCP_NAGLE_OFF);
3960 val = !!(tp->nonagle&TCP_NAGLE_CORK);
3963 val = keepalive_time_when(tp) / HZ;
3966 val = keepalive_intvl_when(tp) / HZ;
3969 val = keepalive_probes(tp);
3972 val = READ_ONCE(icsk->icsk_syn_retries) ? :
3973 READ_ONCE(net->ipv4.sysctl_tcp_syn_retries);
3976 val = READ_ONCE(tp->linger2);
3978 val = (val ? : READ_ONCE(net->ipv4.sysctl_tcp_fin_timeout)) / HZ;
3980 case TCP_DEFER_ACCEPT:
3981 val = READ_ONCE(icsk->icsk_accept_queue.rskq_defer_accept);
3982 val = retrans_to_secs(val, TCP_TIMEOUT_INIT / HZ,
3985 case TCP_WINDOW_CLAMP:
3986 val = tp->window_clamp;
3989 struct tcp_info info;
3991 if (copy_from_sockptr(&len, optlen, sizeof(int)))
3994 tcp_get_info(sk, &info);
3996 len = min_t(unsigned int, len, sizeof(info));
3997 if (copy_to_sockptr(optlen, &len, sizeof(int)))
3999 if (copy_to_sockptr(optval, &info, len))
4004 const struct tcp_congestion_ops *ca_ops;
4005 union tcp_cc_info info;
4009 if (copy_from_sockptr(&len, optlen, sizeof(int)))
4012 ca_ops = icsk->icsk_ca_ops;
4013 if (ca_ops && ca_ops->get_info)
4014 sz = ca_ops->get_info(sk, ~0U, &attr, &info);
4016 len = min_t(unsigned int, len, sz);
4017 if (copy_to_sockptr(optlen, &len, sizeof(int)))
4019 if (copy_to_sockptr(optval, &info, len))
4024 val = !inet_csk_in_pingpong_mode(sk);
4027 case TCP_CONGESTION:
4028 if (copy_from_sockptr(&len, optlen, sizeof(int)))
4030 len = min_t(unsigned int, len, TCP_CA_NAME_MAX);
4031 if (copy_to_sockptr(optlen, &len, sizeof(int)))
4033 if (copy_to_sockptr(optval, icsk->icsk_ca_ops->name, len))
4038 if (copy_from_sockptr(&len, optlen, sizeof(int)))
4040 len = min_t(unsigned int, len, TCP_ULP_NAME_MAX);
4041 if (!icsk->icsk_ulp_ops) {
4043 if (copy_to_sockptr(optlen, &len, sizeof(int)))
4047 if (copy_to_sockptr(optlen, &len, sizeof(int)))
4049 if (copy_to_sockptr(optval, icsk->icsk_ulp_ops->name, len))
4053 case TCP_FASTOPEN_KEY: {
4054 u64 key[TCP_FASTOPEN_KEY_BUF_LENGTH / sizeof(u64)];
4055 unsigned int key_len;
4057 if (copy_from_sockptr(&len, optlen, sizeof(int)))
4060 key_len = tcp_fastopen_get_cipher(net, icsk, key) *
4061 TCP_FASTOPEN_KEY_LENGTH;
4062 len = min_t(unsigned int, len, key_len);
4063 if (copy_to_sockptr(optlen, &len, sizeof(int)))
4065 if (copy_to_sockptr(optval, key, len))
4069 case TCP_THIN_LINEAR_TIMEOUTS:
4073 case TCP_THIN_DUPACK:
4081 case TCP_REPAIR_QUEUE:
4083 val = tp->repair_queue;
4088 case TCP_REPAIR_WINDOW: {
4089 struct tcp_repair_window opt;
4091 if (copy_from_sockptr(&len, optlen, sizeof(int)))
4094 if (len != sizeof(opt))
4100 opt.snd_wl1 = tp->snd_wl1;
4101 opt.snd_wnd = tp->snd_wnd;
4102 opt.max_window = tp->max_window;
4103 opt.rcv_wnd = tp->rcv_wnd;
4104 opt.rcv_wup = tp->rcv_wup;
4106 if (copy_to_sockptr(optval, &opt, len))
4111 if (tp->repair_queue == TCP_SEND_QUEUE)
4112 val = tp->write_seq;
4113 else if (tp->repair_queue == TCP_RECV_QUEUE)
4119 case TCP_USER_TIMEOUT:
4120 val = READ_ONCE(icsk->icsk_user_timeout);
4124 val = READ_ONCE(icsk->icsk_accept_queue.fastopenq.max_qlen);
4127 case TCP_FASTOPEN_CONNECT:
4128 val = tp->fastopen_connect;
4131 case TCP_FASTOPEN_NO_COOKIE:
4132 val = tp->fastopen_no_cookie;
4136 val = READ_ONCE(tp->tcp_tx_delay);
4140 val = tcp_time_stamp_raw() + READ_ONCE(tp->tsoffset);
4142 case TCP_NOTSENT_LOWAT:
4143 val = READ_ONCE(tp->notsent_lowat);
4146 val = tp->recvmsg_inq;
4151 case TCP_SAVED_SYN: {
4152 if (copy_from_sockptr(&len, optlen, sizeof(int)))
4155 sockopt_lock_sock(sk);
4156 if (tp->saved_syn) {
4157 if (len < tcp_saved_syn_len(tp->saved_syn)) {
4158 len = tcp_saved_syn_len(tp->saved_syn);
4159 if (copy_to_sockptr(optlen, &len, sizeof(int))) {
4160 sockopt_release_sock(sk);
4163 sockopt_release_sock(sk);
4166 len = tcp_saved_syn_len(tp->saved_syn);
4167 if (copy_to_sockptr(optlen, &len, sizeof(int))) {
4168 sockopt_release_sock(sk);
4171 if (copy_to_sockptr(optval, tp->saved_syn->data, len)) {
4172 sockopt_release_sock(sk);
4175 tcp_saved_syn_free(tp);
4176 sockopt_release_sock(sk);
4178 sockopt_release_sock(sk);
4180 if (copy_to_sockptr(optlen, &len, sizeof(int)))
4186 case TCP_ZEROCOPY_RECEIVE: {
4187 struct scm_timestamping_internal tss;
4188 struct tcp_zerocopy_receive zc = {};
4191 if (copy_from_sockptr(&len, optlen, sizeof(int)))
4194 len < offsetofend(struct tcp_zerocopy_receive, length))
4196 if (unlikely(len > sizeof(zc))) {
4197 err = check_zeroed_sockptr(optval, sizeof(zc),
4200 return err == 0 ? -EINVAL : err;
4202 if (copy_to_sockptr(optlen, &len, sizeof(int)))
4205 if (copy_from_sockptr(&zc, optval, len))
4209 if (zc.msg_flags & ~(TCP_VALID_ZC_MSG_FLAGS))
4211 sockopt_lock_sock(sk);
4212 err = tcp_zerocopy_receive(sk, &zc, &tss);
4213 err = BPF_CGROUP_RUN_PROG_GETSOCKOPT_KERN(sk, level, optname,
4215 sockopt_release_sock(sk);
4216 if (len >= offsetofend(struct tcp_zerocopy_receive, msg_flags))
4217 goto zerocopy_rcv_cmsg;
4219 case offsetofend(struct tcp_zerocopy_receive, msg_flags):
4220 goto zerocopy_rcv_cmsg;
4221 case offsetofend(struct tcp_zerocopy_receive, msg_controllen):
4222 case offsetofend(struct tcp_zerocopy_receive, msg_control):
4223 case offsetofend(struct tcp_zerocopy_receive, flags):
4224 case offsetofend(struct tcp_zerocopy_receive, copybuf_len):
4225 case offsetofend(struct tcp_zerocopy_receive, copybuf_address):
4226 case offsetofend(struct tcp_zerocopy_receive, err):
4227 goto zerocopy_rcv_sk_err;
4228 case offsetofend(struct tcp_zerocopy_receive, inq):
4229 goto zerocopy_rcv_inq;
4230 case offsetofend(struct tcp_zerocopy_receive, length):
4232 goto zerocopy_rcv_out;
4235 if (zc.msg_flags & TCP_CMSG_TS)
4236 tcp_zc_finalize_rx_tstamp(sk, &zc, &tss);
4239 zerocopy_rcv_sk_err:
4241 zc.err = sock_error(sk);
4243 zc.inq = tcp_inq_hint(sk);
4245 if (!err && copy_to_sockptr(optval, &zc, len))
4251 return -ENOPROTOOPT;
4254 if (copy_to_sockptr(optlen, &len, sizeof(int)))
4256 if (copy_to_sockptr(optval, &val, len))
4261 bool tcp_bpf_bypass_getsockopt(int level, int optname)
4263 /* TCP do_tcp_getsockopt has optimized getsockopt implementation
4264 * to avoid extra socket lock for TCP_ZEROCOPY_RECEIVE.
4266 if (level == SOL_TCP && optname == TCP_ZEROCOPY_RECEIVE)
4271 EXPORT_SYMBOL(tcp_bpf_bypass_getsockopt);
4273 int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval,
4276 struct inet_connection_sock *icsk = inet_csk(sk);
4278 if (level != SOL_TCP)
4279 /* Paired with WRITE_ONCE() in do_ipv6_setsockopt() and tcp_v6_connect() */
4280 return READ_ONCE(icsk->icsk_af_ops)->getsockopt(sk, level, optname,
4282 return do_tcp_getsockopt(sk, level, optname, USER_SOCKPTR(optval),
4283 USER_SOCKPTR(optlen));
4285 EXPORT_SYMBOL(tcp_getsockopt);
4287 #ifdef CONFIG_TCP_MD5SIG
4288 static DEFINE_PER_CPU(struct tcp_md5sig_pool, tcp_md5sig_pool);
4289 static DEFINE_MUTEX(tcp_md5sig_mutex);
4290 static bool tcp_md5sig_pool_populated = false;
4292 static void __tcp_alloc_md5sig_pool(void)
4294 struct crypto_ahash *hash;
4297 hash = crypto_alloc_ahash("md5", 0, CRYPTO_ALG_ASYNC);
4301 for_each_possible_cpu(cpu) {
4302 void *scratch = per_cpu(tcp_md5sig_pool, cpu).scratch;
4303 struct ahash_request *req;
4306 scratch = kmalloc_node(sizeof(union tcp_md5sum_block) +
4307 sizeof(struct tcphdr),
4312 per_cpu(tcp_md5sig_pool, cpu).scratch = scratch;
4314 if (per_cpu(tcp_md5sig_pool, cpu).md5_req)
4317 req = ahash_request_alloc(hash, GFP_KERNEL);
4321 ahash_request_set_callback(req, 0, NULL, NULL);
4323 per_cpu(tcp_md5sig_pool, cpu).md5_req = req;
4325 /* before setting tcp_md5sig_pool_populated, we must commit all writes
4326 * to memory. See smp_rmb() in tcp_get_md5sig_pool()
4329 /* Paired with READ_ONCE() from tcp_alloc_md5sig_pool()
4330 * and tcp_get_md5sig_pool().
4332 WRITE_ONCE(tcp_md5sig_pool_populated, true);
4335 bool tcp_alloc_md5sig_pool(void)
4337 /* Paired with WRITE_ONCE() from __tcp_alloc_md5sig_pool() */
4338 if (unlikely(!READ_ONCE(tcp_md5sig_pool_populated))) {
4339 mutex_lock(&tcp_md5sig_mutex);
4341 if (!tcp_md5sig_pool_populated)
4342 __tcp_alloc_md5sig_pool();
4344 mutex_unlock(&tcp_md5sig_mutex);
4346 /* Paired with WRITE_ONCE() from __tcp_alloc_md5sig_pool() */
4347 return READ_ONCE(tcp_md5sig_pool_populated);
4349 EXPORT_SYMBOL(tcp_alloc_md5sig_pool);
4353 * tcp_get_md5sig_pool - get md5sig_pool for this user
4355 * We use percpu structure, so if we succeed, we exit with preemption
4356 * and BH disabled, to make sure another thread or softirq handling
4357 * wont try to get same context.
4359 struct tcp_md5sig_pool *tcp_get_md5sig_pool(void)
4363 /* Paired with WRITE_ONCE() from __tcp_alloc_md5sig_pool() */
4364 if (READ_ONCE(tcp_md5sig_pool_populated)) {
4365 /* coupled with smp_wmb() in __tcp_alloc_md5sig_pool() */
4367 return this_cpu_ptr(&tcp_md5sig_pool);
4372 EXPORT_SYMBOL(tcp_get_md5sig_pool);
4374 int tcp_md5_hash_skb_data(struct tcp_md5sig_pool *hp,
4375 const struct sk_buff *skb, unsigned int header_len)
4377 struct scatterlist sg;
4378 const struct tcphdr *tp = tcp_hdr(skb);
4379 struct ahash_request *req = hp->md5_req;
4381 const unsigned int head_data_len = skb_headlen(skb) > header_len ?
4382 skb_headlen(skb) - header_len : 0;
4383 const struct skb_shared_info *shi = skb_shinfo(skb);
4384 struct sk_buff *frag_iter;
4386 sg_init_table(&sg, 1);
4388 sg_set_buf(&sg, ((u8 *) tp) + header_len, head_data_len);
4389 ahash_request_set_crypt(req, &sg, NULL, head_data_len);
4390 if (crypto_ahash_update(req))
4393 for (i = 0; i < shi->nr_frags; ++i) {
4394 const skb_frag_t *f = &shi->frags[i];
4395 unsigned int offset = skb_frag_off(f);
4396 struct page *page = skb_frag_page(f) + (offset >> PAGE_SHIFT);
4398 sg_set_page(&sg, page, skb_frag_size(f),
4399 offset_in_page(offset));
4400 ahash_request_set_crypt(req, &sg, NULL, skb_frag_size(f));
4401 if (crypto_ahash_update(req))
4405 skb_walk_frags(skb, frag_iter)
4406 if (tcp_md5_hash_skb_data(hp, frag_iter, 0))
4411 EXPORT_SYMBOL(tcp_md5_hash_skb_data);
4413 int tcp_md5_hash_key(struct tcp_md5sig_pool *hp, const struct tcp_md5sig_key *key)
4415 u8 keylen = READ_ONCE(key->keylen); /* paired with WRITE_ONCE() in tcp_md5_do_add */
4416 struct scatterlist sg;
4418 sg_init_one(&sg, key->key, keylen);
4419 ahash_request_set_crypt(hp->md5_req, &sg, NULL, keylen);
4421 /* We use data_race() because tcp_md5_do_add() might change key->key under us */
4422 return data_race(crypto_ahash_update(hp->md5_req));
4424 EXPORT_SYMBOL(tcp_md5_hash_key);
4426 /* Called with rcu_read_lock() */
4427 enum skb_drop_reason
4428 tcp_inbound_md5_hash(const struct sock *sk, const struct sk_buff *skb,
4429 const void *saddr, const void *daddr,
4430 int family, int dif, int sdif)
4433 * This gets called for each TCP segment that arrives
4434 * so we want to be efficient.
4435 * We have 3 drop cases:
4436 * o No MD5 hash and one expected.
4437 * o MD5 hash and we're not expecting one.
4438 * o MD5 hash and its wrong.
4440 const __u8 *hash_location = NULL;
4441 struct tcp_md5sig_key *hash_expected;
4442 const struct tcphdr *th = tcp_hdr(skb);
4443 const struct tcp_sock *tp = tcp_sk(sk);
4444 int genhash, l3index;
4447 /* sdif set, means packet ingressed via a device
4448 * in an L3 domain and dif is set to the l3mdev
4450 l3index = sdif ? dif : 0;
4452 hash_expected = tcp_md5_do_lookup(sk, l3index, saddr, family);
4453 hash_location = tcp_parse_md5sig_option(th);
4455 /* We've parsed the options - do we have a hash? */
4456 if (!hash_expected && !hash_location)
4457 return SKB_NOT_DROPPED_YET;
4459 if (hash_expected && !hash_location) {
4460 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMD5NOTFOUND);
4461 return SKB_DROP_REASON_TCP_MD5NOTFOUND;
4464 if (!hash_expected && hash_location) {
4465 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMD5UNEXPECTED);
4466 return SKB_DROP_REASON_TCP_MD5UNEXPECTED;
4469 /* Check the signature.
4470 * To support dual stack listeners, we need to handle
4473 if (family == AF_INET)
4474 genhash = tcp_v4_md5_hash_skb(newhash,
4478 genhash = tp->af_specific->calc_md5_hash(newhash,
4482 if (genhash || memcmp(hash_location, newhash, 16) != 0) {
4483 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMD5FAILURE);
4484 if (family == AF_INET) {
4485 net_info_ratelimited("MD5 Hash failed for (%pI4, %d)->(%pI4, %d)%s L3 index %d\n",
4486 saddr, ntohs(th->source),
4487 daddr, ntohs(th->dest),
4488 genhash ? " tcp_v4_calc_md5_hash failed"
4491 net_info_ratelimited("MD5 Hash %s for [%pI6c]:%u->[%pI6c]:%u L3 index %d\n",
4492 genhash ? "failed" : "mismatch",
4493 saddr, ntohs(th->source),
4494 daddr, ntohs(th->dest), l3index);
4496 return SKB_DROP_REASON_TCP_MD5FAILURE;
4498 return SKB_NOT_DROPPED_YET;
4500 EXPORT_SYMBOL(tcp_inbound_md5_hash);
4504 void tcp_done(struct sock *sk)
4506 struct request_sock *req;
4508 /* We might be called with a new socket, after
4509 * inet_csk_prepare_forced_close() has been called
4510 * so we can not use lockdep_sock_is_held(sk)
4512 req = rcu_dereference_protected(tcp_sk(sk)->fastopen_rsk, 1);
4514 if (sk->sk_state == TCP_SYN_SENT || sk->sk_state == TCP_SYN_RECV)
4515 TCP_INC_STATS(sock_net(sk), TCP_MIB_ATTEMPTFAILS);
4517 tcp_set_state(sk, TCP_CLOSE);
4518 tcp_clear_xmit_timers(sk);
4520 reqsk_fastopen_remove(sk, req, false);
4522 WRITE_ONCE(sk->sk_shutdown, SHUTDOWN_MASK);
4524 if (!sock_flag(sk, SOCK_DEAD))
4525 sk->sk_state_change(sk);
4527 inet_csk_destroy_sock(sk);
4529 EXPORT_SYMBOL_GPL(tcp_done);
4531 int tcp_abort(struct sock *sk, int err)
4533 int state = inet_sk_state_load(sk);
4535 if (state == TCP_NEW_SYN_RECV) {
4536 struct request_sock *req = inet_reqsk(sk);
4539 inet_csk_reqsk_queue_drop(req->rsk_listener, req);
4543 if (state == TCP_TIME_WAIT) {
4544 struct inet_timewait_sock *tw = inet_twsk(sk);
4546 refcount_inc(&tw->tw_refcnt);
4548 inet_twsk_deschedule_put(tw);
4553 /* BPF context ensures sock locking. */
4554 if (!has_current_bpf_ctx())
4555 /* Don't race with userspace socket closes such as tcp_close. */
4558 if (sk->sk_state == TCP_LISTEN) {
4559 tcp_set_state(sk, TCP_CLOSE);
4560 inet_csk_listen_stop(sk);
4563 /* Don't race with BH socket closes such as inet_csk_listen_stop. */
4567 if (!sock_flag(sk, SOCK_DEAD)) {
4568 WRITE_ONCE(sk->sk_err, err);
4569 /* This barrier is coupled with smp_rmb() in tcp_poll() */
4571 sk_error_report(sk);
4572 if (tcp_need_reset(sk->sk_state))
4573 tcp_send_active_reset(sk, GFP_ATOMIC);
4579 tcp_write_queue_purge(sk);
4580 if (!has_current_bpf_ctx())
4584 EXPORT_SYMBOL_GPL(tcp_abort);
4586 extern struct tcp_congestion_ops tcp_reno;
4588 static __initdata unsigned long thash_entries;
4589 static int __init set_thash_entries(char *str)
4596 ret = kstrtoul(str, 0, &thash_entries);
4602 __setup("thash_entries=", set_thash_entries);
4604 static void __init tcp_init_mem(void)
4606 unsigned long limit = nr_free_buffer_pages() / 16;
4608 limit = max(limit, 128UL);
4609 sysctl_tcp_mem[0] = limit / 4 * 3; /* 4.68 % */
4610 sysctl_tcp_mem[1] = limit; /* 6.25 % */
4611 sysctl_tcp_mem[2] = sysctl_tcp_mem[0] * 2; /* 9.37 % */
4614 void __init tcp_init(void)
4616 int max_rshare, max_wshare, cnt;
4617 unsigned long limit;
4620 BUILD_BUG_ON(TCP_MIN_SND_MSS <= MAX_TCP_OPTION_SPACE);
4621 BUILD_BUG_ON(sizeof(struct tcp_skb_cb) >
4622 sizeof_field(struct sk_buff, cb));
4624 percpu_counter_init(&tcp_sockets_allocated, 0, GFP_KERNEL);
4626 timer_setup(&tcp_orphan_timer, tcp_orphan_update, TIMER_DEFERRABLE);
4627 mod_timer(&tcp_orphan_timer, jiffies + TCP_ORPHAN_TIMER_PERIOD);
4629 inet_hashinfo2_init(&tcp_hashinfo, "tcp_listen_portaddr_hash",
4630 thash_entries, 21, /* one slot per 2 MB*/
4632 tcp_hashinfo.bind_bucket_cachep =
4633 kmem_cache_create("tcp_bind_bucket",
4634 sizeof(struct inet_bind_bucket), 0,
4635 SLAB_HWCACHE_ALIGN | SLAB_PANIC |
4638 tcp_hashinfo.bind2_bucket_cachep =
4639 kmem_cache_create("tcp_bind2_bucket",
4640 sizeof(struct inet_bind2_bucket), 0,
4641 SLAB_HWCACHE_ALIGN | SLAB_PANIC |
4645 /* Size and allocate the main established and bind bucket
4648 * The methodology is similar to that of the buffer cache.
4650 tcp_hashinfo.ehash =
4651 alloc_large_system_hash("TCP established",
4652 sizeof(struct inet_ehash_bucket),
4654 17, /* one slot per 128 KB of memory */
4657 &tcp_hashinfo.ehash_mask,
4659 thash_entries ? 0 : 512 * 1024);
4660 for (i = 0; i <= tcp_hashinfo.ehash_mask; i++)
4661 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].chain, i);
4663 if (inet_ehash_locks_alloc(&tcp_hashinfo))
4664 panic("TCP: failed to alloc ehash_locks");
4665 tcp_hashinfo.bhash =
4666 alloc_large_system_hash("TCP bind",
4667 2 * sizeof(struct inet_bind_hashbucket),
4668 tcp_hashinfo.ehash_mask + 1,
4669 17, /* one slot per 128 KB of memory */
4671 &tcp_hashinfo.bhash_size,
4675 tcp_hashinfo.bhash_size = 1U << tcp_hashinfo.bhash_size;
4676 tcp_hashinfo.bhash2 = tcp_hashinfo.bhash + tcp_hashinfo.bhash_size;
4677 for (i = 0; i < tcp_hashinfo.bhash_size; i++) {
4678 spin_lock_init(&tcp_hashinfo.bhash[i].lock);
4679 INIT_HLIST_HEAD(&tcp_hashinfo.bhash[i].chain);
4680 spin_lock_init(&tcp_hashinfo.bhash2[i].lock);
4681 INIT_HLIST_HEAD(&tcp_hashinfo.bhash2[i].chain);
4684 tcp_hashinfo.pernet = false;
4686 cnt = tcp_hashinfo.ehash_mask + 1;
4687 sysctl_tcp_max_orphans = cnt / 2;
4690 /* Set per-socket limits to no more than 1/128 the pressure threshold */
4691 limit = nr_free_buffer_pages() << (PAGE_SHIFT - 7);
4692 max_wshare = min(4UL*1024*1024, limit);
4693 max_rshare = min(6UL*1024*1024, limit);
4695 init_net.ipv4.sysctl_tcp_wmem[0] = PAGE_SIZE;
4696 init_net.ipv4.sysctl_tcp_wmem[1] = 16*1024;
4697 init_net.ipv4.sysctl_tcp_wmem[2] = max(64*1024, max_wshare);
4699 init_net.ipv4.sysctl_tcp_rmem[0] = PAGE_SIZE;
4700 init_net.ipv4.sysctl_tcp_rmem[1] = 131072;
4701 init_net.ipv4.sysctl_tcp_rmem[2] = max(131072, max_rshare);
4703 pr_info("Hash tables configured (established %u bind %u)\n",
4704 tcp_hashinfo.ehash_mask + 1, tcp_hashinfo.bhash_size);
4708 BUG_ON(tcp_register_congestion_control(&tcp_reno) != 0);