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>
284 /* Track pending CMSGs. */
290 DEFINE_PER_CPU(unsigned int, tcp_orphan_count);
291 EXPORT_PER_CPU_SYMBOL_GPL(tcp_orphan_count);
293 long sysctl_tcp_mem[3] __read_mostly;
294 EXPORT_SYMBOL(sysctl_tcp_mem);
296 atomic_long_t tcp_memory_allocated ____cacheline_aligned_in_smp; /* Current allocated memory. */
297 EXPORT_SYMBOL(tcp_memory_allocated);
298 DEFINE_PER_CPU(int, tcp_memory_per_cpu_fw_alloc);
299 EXPORT_PER_CPU_SYMBOL_GPL(tcp_memory_per_cpu_fw_alloc);
301 #if IS_ENABLED(CONFIG_SMC)
302 DEFINE_STATIC_KEY_FALSE(tcp_have_smc);
303 EXPORT_SYMBOL(tcp_have_smc);
307 * Current number of TCP sockets.
309 struct percpu_counter tcp_sockets_allocated ____cacheline_aligned_in_smp;
310 EXPORT_SYMBOL(tcp_sockets_allocated);
315 struct tcp_splice_state {
316 struct pipe_inode_info *pipe;
322 * Pressure flag: try to collapse.
323 * Technical note: it is used by multiple contexts non atomically.
324 * All the __sk_mem_schedule() is of this nature: accounting
325 * is strict, actions are advisory and have some latency.
327 unsigned long tcp_memory_pressure __read_mostly;
328 EXPORT_SYMBOL_GPL(tcp_memory_pressure);
330 void tcp_enter_memory_pressure(struct sock *sk)
334 if (READ_ONCE(tcp_memory_pressure))
340 if (!cmpxchg(&tcp_memory_pressure, 0, val))
341 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMEMORYPRESSURES);
343 EXPORT_SYMBOL_GPL(tcp_enter_memory_pressure);
345 void tcp_leave_memory_pressure(struct sock *sk)
349 if (!READ_ONCE(tcp_memory_pressure))
351 val = xchg(&tcp_memory_pressure, 0);
353 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPMEMORYPRESSURESCHRONO,
354 jiffies_to_msecs(jiffies - val));
356 EXPORT_SYMBOL_GPL(tcp_leave_memory_pressure);
358 /* Convert seconds to retransmits based on initial and max timeout */
359 static u8 secs_to_retrans(int seconds, int timeout, int rto_max)
364 int period = timeout;
367 while (seconds > period && res < 255) {
370 if (timeout > rto_max)
378 /* Convert retransmits to seconds based on initial and max timeout */
379 static int retrans_to_secs(u8 retrans, int timeout, int rto_max)
387 if (timeout > rto_max)
395 static u64 tcp_compute_delivery_rate(const struct tcp_sock *tp)
397 u32 rate = READ_ONCE(tp->rate_delivered);
398 u32 intv = READ_ONCE(tp->rate_interval_us);
402 rate64 = (u64)rate * tp->mss_cache * USEC_PER_SEC;
403 do_div(rate64, intv);
408 /* Address-family independent initialization for a tcp_sock.
410 * NOTE: A lot of things set to zero explicitly by call to
411 * sk_alloc() so need not be done here.
413 void tcp_init_sock(struct sock *sk)
415 struct inet_connection_sock *icsk = inet_csk(sk);
416 struct tcp_sock *tp = tcp_sk(sk);
418 tp->out_of_order_queue = RB_ROOT;
419 sk->tcp_rtx_queue = RB_ROOT;
420 tcp_init_xmit_timers(sk);
421 INIT_LIST_HEAD(&tp->tsq_node);
422 INIT_LIST_HEAD(&tp->tsorted_sent_queue);
424 icsk->icsk_rto = TCP_TIMEOUT_INIT;
425 icsk->icsk_rto_min = TCP_RTO_MIN;
426 icsk->icsk_delack_max = TCP_DELACK_MAX;
427 tp->mdev_us = jiffies_to_usecs(TCP_TIMEOUT_INIT);
428 minmax_reset(&tp->rtt_min, tcp_jiffies32, ~0U);
430 /* So many TCP implementations out there (incorrectly) count the
431 * initial SYN frame in their delayed-ACK and congestion control
432 * algorithms that we must have the following bandaid to talk
433 * efficiently to them. -DaveM
435 tcp_snd_cwnd_set(tp, TCP_INIT_CWND);
437 /* There's a bubble in the pipe until at least the first ACK. */
438 tp->app_limited = ~0U;
439 tp->rate_app_limited = 1;
441 /* See draft-stevens-tcpca-spec-01 for discussion of the
442 * initialization of these values.
444 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
445 tp->snd_cwnd_clamp = ~0;
446 tp->mss_cache = TCP_MSS_DEFAULT;
448 tp->reordering = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_reordering);
449 tcp_assign_congestion_control(sk);
452 tp->rack.reo_wnd_steps = 1;
454 sk->sk_write_space = sk_stream_write_space;
455 sock_set_flag(sk, SOCK_USE_WRITE_QUEUE);
457 icsk->icsk_sync_mss = tcp_sync_mss;
459 WRITE_ONCE(sk->sk_sndbuf, READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_wmem[1]));
460 WRITE_ONCE(sk->sk_rcvbuf, READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_rmem[1]));
461 tcp_scaling_ratio_init(sk);
463 set_bit(SOCK_SUPPORT_ZC, &sk->sk_socket->flags);
464 sk_sockets_allocated_inc(sk);
466 EXPORT_SYMBOL(tcp_init_sock);
468 static void tcp_tx_timestamp(struct sock *sk, u16 tsflags)
470 struct sk_buff *skb = tcp_write_queue_tail(sk);
472 if (tsflags && skb) {
473 struct skb_shared_info *shinfo = skb_shinfo(skb);
474 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
476 sock_tx_timestamp(sk, tsflags, &shinfo->tx_flags);
477 if (tsflags & SOF_TIMESTAMPING_TX_ACK)
478 tcb->txstamp_ack = 1;
479 if (tsflags & SOF_TIMESTAMPING_TX_RECORD_MASK)
480 shinfo->tskey = TCP_SKB_CB(skb)->seq + skb->len - 1;
484 static bool tcp_stream_is_readable(struct sock *sk, int target)
486 if (tcp_epollin_ready(sk, target))
488 return sk_is_readable(sk);
492 * Wait for a TCP event.
494 * Note that we don't need to lock the socket, as the upper poll layers
495 * take care of normal races (between the test and the event) and we don't
496 * go look at any of the socket buffers directly.
498 __poll_t tcp_poll(struct file *file, struct socket *sock, poll_table *wait)
501 struct sock *sk = sock->sk;
502 const struct tcp_sock *tp = tcp_sk(sk);
506 sock_poll_wait(file, sock, wait);
508 state = inet_sk_state_load(sk);
509 if (state == TCP_LISTEN)
510 return inet_csk_listen_poll(sk);
512 /* Socket is not locked. We are protected from async events
513 * by poll logic and correct handling of state changes
514 * made by other threads is impossible in any case.
520 * EPOLLHUP is certainly not done right. But poll() doesn't
521 * have a notion of HUP in just one direction, and for a
522 * socket the read side is more interesting.
524 * Some poll() documentation says that EPOLLHUP is incompatible
525 * with the EPOLLOUT/POLLWR flags, so somebody should check this
526 * all. But careful, it tends to be safer to return too many
527 * bits than too few, and you can easily break real applications
528 * if you don't tell them that something has hung up!
532 * Check number 1. EPOLLHUP is _UNMASKABLE_ event (see UNIX98 and
533 * our fs/select.c). It means that after we received EOF,
534 * poll always returns immediately, making impossible poll() on write()
535 * in state CLOSE_WAIT. One solution is evident --- to set EPOLLHUP
536 * if and only if shutdown has been made in both directions.
537 * Actually, it is interesting to look how Solaris and DUX
538 * solve this dilemma. I would prefer, if EPOLLHUP were maskable,
539 * then we could set it on SND_SHUTDOWN. BTW examples given
540 * in Stevens' books assume exactly this behaviour, it explains
541 * why EPOLLHUP is incompatible with EPOLLOUT. --ANK
543 * NOTE. Check for TCP_CLOSE is added. The goal is to prevent
544 * blocking on fresh not-connected or disconnected socket. --ANK
546 shutdown = READ_ONCE(sk->sk_shutdown);
547 if (shutdown == SHUTDOWN_MASK || state == TCP_CLOSE)
549 if (shutdown & RCV_SHUTDOWN)
550 mask |= EPOLLIN | EPOLLRDNORM | EPOLLRDHUP;
552 /* Connected or passive Fast Open socket? */
553 if (state != TCP_SYN_SENT &&
554 (state != TCP_SYN_RECV || rcu_access_pointer(tp->fastopen_rsk))) {
555 int target = sock_rcvlowat(sk, 0, INT_MAX);
556 u16 urg_data = READ_ONCE(tp->urg_data);
558 if (unlikely(urg_data) &&
559 READ_ONCE(tp->urg_seq) == READ_ONCE(tp->copied_seq) &&
560 !sock_flag(sk, SOCK_URGINLINE))
563 if (tcp_stream_is_readable(sk, target))
564 mask |= EPOLLIN | EPOLLRDNORM;
566 if (!(shutdown & SEND_SHUTDOWN)) {
567 if (__sk_stream_is_writeable(sk, 1)) {
568 mask |= EPOLLOUT | EPOLLWRNORM;
569 } else { /* send SIGIO later */
570 sk_set_bit(SOCKWQ_ASYNC_NOSPACE, sk);
571 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
573 /* Race breaker. If space is freed after
574 * wspace test but before the flags are set,
575 * IO signal will be lost. Memory barrier
576 * pairs with the input side.
578 smp_mb__after_atomic();
579 if (__sk_stream_is_writeable(sk, 1))
580 mask |= EPOLLOUT | EPOLLWRNORM;
583 mask |= EPOLLOUT | EPOLLWRNORM;
585 if (urg_data & TCP_URG_VALID)
587 } else if (state == TCP_SYN_SENT &&
588 inet_test_bit(DEFER_CONNECT, sk)) {
589 /* Active TCP fastopen socket with defer_connect
590 * Return EPOLLOUT so application can call write()
591 * in order for kernel to generate SYN+data
593 mask |= EPOLLOUT | EPOLLWRNORM;
595 /* This barrier is coupled with smp_wmb() in tcp_reset() */
597 if (READ_ONCE(sk->sk_err) ||
598 !skb_queue_empty_lockless(&sk->sk_error_queue))
603 EXPORT_SYMBOL(tcp_poll);
605 int tcp_ioctl(struct sock *sk, int cmd, int *karg)
607 struct tcp_sock *tp = tcp_sk(sk);
613 if (sk->sk_state == TCP_LISTEN)
616 slow = lock_sock_fast(sk);
618 unlock_sock_fast(sk, slow);
621 answ = READ_ONCE(tp->urg_data) &&
622 READ_ONCE(tp->urg_seq) == READ_ONCE(tp->copied_seq);
625 if (sk->sk_state == TCP_LISTEN)
628 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
631 answ = READ_ONCE(tp->write_seq) - tp->snd_una;
634 if (sk->sk_state == TCP_LISTEN)
637 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
640 answ = READ_ONCE(tp->write_seq) -
641 READ_ONCE(tp->snd_nxt);
650 EXPORT_SYMBOL(tcp_ioctl);
652 void tcp_mark_push(struct tcp_sock *tp, struct sk_buff *skb)
654 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
655 tp->pushed_seq = tp->write_seq;
658 static inline bool forced_push(const struct tcp_sock *tp)
660 return after(tp->write_seq, tp->pushed_seq + (tp->max_window >> 1));
663 void tcp_skb_entail(struct sock *sk, struct sk_buff *skb)
665 struct tcp_sock *tp = tcp_sk(sk);
666 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
668 tcb->seq = tcb->end_seq = tp->write_seq;
669 tcb->tcp_flags = TCPHDR_ACK;
670 __skb_header_release(skb);
671 tcp_add_write_queue_tail(sk, skb);
672 sk_wmem_queued_add(sk, skb->truesize);
673 sk_mem_charge(sk, skb->truesize);
674 if (tp->nonagle & TCP_NAGLE_PUSH)
675 tp->nonagle &= ~TCP_NAGLE_PUSH;
677 tcp_slow_start_after_idle_check(sk);
680 static inline void tcp_mark_urg(struct tcp_sock *tp, int flags)
683 tp->snd_up = tp->write_seq;
686 /* If a not yet filled skb is pushed, do not send it if
687 * we have data packets in Qdisc or NIC queues :
688 * Because TX completion will happen shortly, it gives a chance
689 * to coalesce future sendmsg() payload into this skb, without
690 * need for a timer, and with no latency trade off.
691 * As packets containing data payload have a bigger truesize
692 * than pure acks (dataless) packets, the last checks prevent
693 * autocorking if we only have an ACK in Qdisc/NIC queues,
694 * or if TX completion was delayed after we processed ACK packet.
696 static bool tcp_should_autocork(struct sock *sk, struct sk_buff *skb,
699 return skb->len < size_goal &&
700 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_autocorking) &&
701 !tcp_rtx_queue_empty(sk) &&
702 refcount_read(&sk->sk_wmem_alloc) > skb->truesize &&
703 tcp_skb_can_collapse_to(skb);
706 void tcp_push(struct sock *sk, int flags, int mss_now,
707 int nonagle, int size_goal)
709 struct tcp_sock *tp = tcp_sk(sk);
712 skb = tcp_write_queue_tail(sk);
715 if (!(flags & MSG_MORE) || forced_push(tp))
716 tcp_mark_push(tp, skb);
718 tcp_mark_urg(tp, flags);
720 if (tcp_should_autocork(sk, skb, size_goal)) {
722 /* avoid atomic op if TSQ_THROTTLED bit is already set */
723 if (!test_bit(TSQ_THROTTLED, &sk->sk_tsq_flags)) {
724 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPAUTOCORKING);
725 set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags);
726 smp_mb__after_atomic();
728 /* It is possible TX completion already happened
729 * before we set TSQ_THROTTLED.
731 if (refcount_read(&sk->sk_wmem_alloc) > skb->truesize)
735 if (flags & MSG_MORE)
736 nonagle = TCP_NAGLE_CORK;
738 __tcp_push_pending_frames(sk, mss_now, nonagle);
741 static int tcp_splice_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb,
742 unsigned int offset, size_t len)
744 struct tcp_splice_state *tss = rd_desc->arg.data;
747 ret = skb_splice_bits(skb, skb->sk, offset, tss->pipe,
748 min(rd_desc->count, len), tss->flags);
750 rd_desc->count -= ret;
754 static int __tcp_splice_read(struct sock *sk, struct tcp_splice_state *tss)
756 /* Store TCP splice context information in read_descriptor_t. */
757 read_descriptor_t rd_desc = {
762 return tcp_read_sock(sk, &rd_desc, tcp_splice_data_recv);
766 * tcp_splice_read - splice data from TCP socket to a pipe
767 * @sock: socket to splice from
768 * @ppos: position (not valid)
769 * @pipe: pipe to splice to
770 * @len: number of bytes to splice
771 * @flags: splice modifier flags
774 * Will read pages from given socket and fill them into a pipe.
777 ssize_t tcp_splice_read(struct socket *sock, loff_t *ppos,
778 struct pipe_inode_info *pipe, size_t len,
781 struct sock *sk = sock->sk;
782 struct tcp_splice_state tss = {
791 sock_rps_record_flow(sk);
793 * We can't seek on a socket input
802 timeo = sock_rcvtimeo(sk, sock->file->f_flags & O_NONBLOCK);
804 ret = __tcp_splice_read(sk, &tss);
810 if (sock_flag(sk, SOCK_DONE))
813 ret = sock_error(sk);
816 if (sk->sk_shutdown & RCV_SHUTDOWN)
818 if (sk->sk_state == TCP_CLOSE) {
820 * This occurs when user tries to read
821 * from never connected socket.
830 /* if __tcp_splice_read() got nothing while we have
831 * an skb in receive queue, we do not want to loop.
832 * This might happen with URG data.
834 if (!skb_queue_empty(&sk->sk_receive_queue))
836 ret = sk_wait_data(sk, &timeo, NULL);
839 if (signal_pending(current)) {
840 ret = sock_intr_errno(timeo);
848 if (!tss.len || !timeo)
853 if (sk->sk_err || sk->sk_state == TCP_CLOSE ||
854 (sk->sk_shutdown & RCV_SHUTDOWN) ||
855 signal_pending(current))
866 EXPORT_SYMBOL(tcp_splice_read);
868 struct sk_buff *tcp_stream_alloc_skb(struct sock *sk, gfp_t gfp,
873 skb = alloc_skb_fclone(MAX_TCP_HEADER, gfp);
877 skb->truesize = SKB_TRUESIZE(skb_end_offset(skb));
878 if (force_schedule) {
879 mem_scheduled = true;
880 sk_forced_mem_schedule(sk, skb->truesize);
882 mem_scheduled = sk_wmem_schedule(sk, skb->truesize);
884 if (likely(mem_scheduled)) {
885 skb_reserve(skb, MAX_TCP_HEADER);
886 skb->ip_summed = CHECKSUM_PARTIAL;
887 INIT_LIST_HEAD(&skb->tcp_tsorted_anchor);
892 sk->sk_prot->enter_memory_pressure(sk);
893 sk_stream_moderate_sndbuf(sk);
898 static unsigned int tcp_xmit_size_goal(struct sock *sk, u32 mss_now,
901 struct tcp_sock *tp = tcp_sk(sk);
902 u32 new_size_goal, size_goal;
907 /* Note : tcp_tso_autosize() will eventually split this later */
908 new_size_goal = tcp_bound_to_half_wnd(tp, sk->sk_gso_max_size);
910 /* We try hard to avoid divides here */
911 size_goal = tp->gso_segs * mss_now;
912 if (unlikely(new_size_goal < size_goal ||
913 new_size_goal >= size_goal + mss_now)) {
914 tp->gso_segs = min_t(u16, new_size_goal / mss_now,
915 sk->sk_gso_max_segs);
916 size_goal = tp->gso_segs * mss_now;
919 return max(size_goal, mss_now);
922 int tcp_send_mss(struct sock *sk, int *size_goal, int flags)
926 mss_now = tcp_current_mss(sk);
927 *size_goal = tcp_xmit_size_goal(sk, mss_now, !(flags & MSG_OOB));
932 /* In some cases, sendmsg() could have added an skb to the write queue,
933 * but failed adding payload on it. We need to remove it to consume less
934 * memory, but more importantly be able to generate EPOLLOUT for Edge Trigger
935 * epoll() users. Another reason is that tcp_write_xmit() does not like
936 * finding an empty skb in the write queue.
938 void tcp_remove_empty_skb(struct sock *sk)
940 struct sk_buff *skb = tcp_write_queue_tail(sk);
942 if (skb && TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq) {
943 tcp_unlink_write_queue(skb, sk);
944 if (tcp_write_queue_empty(sk))
945 tcp_chrono_stop(sk, TCP_CHRONO_BUSY);
946 tcp_wmem_free_skb(sk, skb);
950 /* skb changing from pure zc to mixed, must charge zc */
951 static int tcp_downgrade_zcopy_pure(struct sock *sk, struct sk_buff *skb)
953 if (unlikely(skb_zcopy_pure(skb))) {
954 u32 extra = skb->truesize -
955 SKB_TRUESIZE(skb_end_offset(skb));
957 if (!sk_wmem_schedule(sk, extra))
960 sk_mem_charge(sk, extra);
961 skb_shinfo(skb)->flags &= ~SKBFL_PURE_ZEROCOPY;
967 int tcp_wmem_schedule(struct sock *sk, int copy)
971 if (likely(sk_wmem_schedule(sk, copy)))
974 /* We could be in trouble if we have nothing queued.
975 * Use whatever is left in sk->sk_forward_alloc and tcp_wmem[0]
976 * to guarantee some progress.
978 left = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_wmem[0]) - sk->sk_wmem_queued;
980 sk_forced_mem_schedule(sk, min(left, copy));
981 return min(copy, sk->sk_forward_alloc);
984 void tcp_free_fastopen_req(struct tcp_sock *tp)
986 if (tp->fastopen_req) {
987 kfree(tp->fastopen_req);
988 tp->fastopen_req = NULL;
992 int tcp_sendmsg_fastopen(struct sock *sk, struct msghdr *msg, int *copied,
993 size_t size, struct ubuf_info *uarg)
995 struct tcp_sock *tp = tcp_sk(sk);
996 struct inet_sock *inet = inet_sk(sk);
997 struct sockaddr *uaddr = msg->msg_name;
1000 if (!(READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_fastopen) &
1001 TFO_CLIENT_ENABLE) ||
1002 (uaddr && msg->msg_namelen >= sizeof(uaddr->sa_family) &&
1003 uaddr->sa_family == AF_UNSPEC))
1005 if (tp->fastopen_req)
1006 return -EALREADY; /* Another Fast Open is in progress */
1008 tp->fastopen_req = kzalloc(sizeof(struct tcp_fastopen_request),
1010 if (unlikely(!tp->fastopen_req))
1012 tp->fastopen_req->data = msg;
1013 tp->fastopen_req->size = size;
1014 tp->fastopen_req->uarg = uarg;
1016 if (inet_test_bit(DEFER_CONNECT, sk)) {
1017 err = tcp_connect(sk);
1018 /* Same failure procedure as in tcp_v4/6_connect */
1020 tcp_set_state(sk, TCP_CLOSE);
1021 inet->inet_dport = 0;
1022 sk->sk_route_caps = 0;
1025 flags = (msg->msg_flags & MSG_DONTWAIT) ? O_NONBLOCK : 0;
1026 err = __inet_stream_connect(sk->sk_socket, uaddr,
1027 msg->msg_namelen, flags, 1);
1028 /* fastopen_req could already be freed in __inet_stream_connect
1029 * if the connection times out or gets rst
1031 if (tp->fastopen_req) {
1032 *copied = tp->fastopen_req->copied;
1033 tcp_free_fastopen_req(tp);
1034 inet_clear_bit(DEFER_CONNECT, sk);
1039 int tcp_sendmsg_locked(struct sock *sk, struct msghdr *msg, size_t size)
1041 struct tcp_sock *tp = tcp_sk(sk);
1042 struct ubuf_info *uarg = NULL;
1043 struct sk_buff *skb;
1044 struct sockcm_cookie sockc;
1045 int flags, err, copied = 0;
1046 int mss_now = 0, size_goal, copied_syn = 0;
1047 int process_backlog = 0;
1051 flags = msg->msg_flags;
1053 if ((flags & MSG_ZEROCOPY) && size) {
1054 if (msg->msg_ubuf) {
1055 uarg = msg->msg_ubuf;
1056 if (sk->sk_route_caps & NETIF_F_SG)
1058 } else if (sock_flag(sk, SOCK_ZEROCOPY)) {
1059 skb = tcp_write_queue_tail(sk);
1060 uarg = msg_zerocopy_realloc(sk, size, skb_zcopy(skb));
1065 if (sk->sk_route_caps & NETIF_F_SG)
1068 uarg_to_msgzc(uarg)->zerocopy = 0;
1070 } else if (unlikely(msg->msg_flags & MSG_SPLICE_PAGES) && size) {
1071 if (sk->sk_route_caps & NETIF_F_SG)
1072 zc = MSG_SPLICE_PAGES;
1075 if (unlikely(flags & MSG_FASTOPEN ||
1076 inet_test_bit(DEFER_CONNECT, sk)) &&
1078 err = tcp_sendmsg_fastopen(sk, msg, &copied_syn, size, uarg);
1079 if (err == -EINPROGRESS && copied_syn > 0)
1085 timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
1087 tcp_rate_check_app_limited(sk); /* is sending application-limited? */
1089 /* Wait for a connection to finish. One exception is TCP Fast Open
1090 * (passive side) where data is allowed to be sent before a connection
1091 * is fully established.
1093 if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) &&
1094 !tcp_passive_fastopen(sk)) {
1095 err = sk_stream_wait_connect(sk, &timeo);
1100 if (unlikely(tp->repair)) {
1101 if (tp->repair_queue == TCP_RECV_QUEUE) {
1102 copied = tcp_send_rcvq(sk, msg, size);
1107 if (tp->repair_queue == TCP_NO_QUEUE)
1110 /* 'common' sending to sendq */
1113 sockcm_init(&sockc, sk);
1114 if (msg->msg_controllen) {
1115 err = sock_cmsg_send(sk, msg, &sockc);
1116 if (unlikely(err)) {
1122 /* This should be in poll */
1123 sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
1125 /* Ok commence sending. */
1129 mss_now = tcp_send_mss(sk, &size_goal, flags);
1132 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
1135 while (msg_data_left(msg)) {
1138 skb = tcp_write_queue_tail(sk);
1140 copy = size_goal - skb->len;
1142 if (copy <= 0 || !tcp_skb_can_collapse_to(skb)) {
1146 if (!sk_stream_memory_free(sk))
1147 goto wait_for_space;
1149 if (unlikely(process_backlog >= 16)) {
1150 process_backlog = 0;
1151 if (sk_flush_backlog(sk))
1154 first_skb = tcp_rtx_and_write_queues_empty(sk);
1155 skb = tcp_stream_alloc_skb(sk, sk->sk_allocation,
1158 goto wait_for_space;
1162 tcp_skb_entail(sk, skb);
1165 /* All packets are restored as if they have
1166 * already been sent. skb_mstamp_ns isn't set to
1167 * avoid wrong rtt estimation.
1170 TCP_SKB_CB(skb)->sacked |= TCPCB_REPAIRED;
1173 /* Try to append data to the end of skb. */
1174 if (copy > msg_data_left(msg))
1175 copy = msg_data_left(msg);
1179 int i = skb_shinfo(skb)->nr_frags;
1180 struct page_frag *pfrag = sk_page_frag(sk);
1182 if (!sk_page_frag_refill(sk, pfrag))
1183 goto wait_for_space;
1185 if (!skb_can_coalesce(skb, i, pfrag->page,
1187 if (i >= READ_ONCE(sysctl_max_skb_frags)) {
1188 tcp_mark_push(tp, skb);
1194 copy = min_t(int, copy, pfrag->size - pfrag->offset);
1196 if (unlikely(skb_zcopy_pure(skb) || skb_zcopy_managed(skb))) {
1197 if (tcp_downgrade_zcopy_pure(sk, skb))
1198 goto wait_for_space;
1199 skb_zcopy_downgrade_managed(skb);
1202 copy = tcp_wmem_schedule(sk, copy);
1204 goto wait_for_space;
1206 err = skb_copy_to_page_nocache(sk, &msg->msg_iter, skb,
1213 /* Update the skb. */
1215 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1217 skb_fill_page_desc(skb, i, pfrag->page,
1218 pfrag->offset, copy);
1219 page_ref_inc(pfrag->page);
1221 pfrag->offset += copy;
1222 } else if (zc == MSG_ZEROCOPY) {
1223 /* First append to a fragless skb builds initial
1227 skb_shinfo(skb)->flags |= SKBFL_PURE_ZEROCOPY;
1229 if (!skb_zcopy_pure(skb)) {
1230 copy = tcp_wmem_schedule(sk, copy);
1232 goto wait_for_space;
1235 err = skb_zerocopy_iter_stream(sk, skb, msg, copy, uarg);
1236 if (err == -EMSGSIZE || err == -EEXIST) {
1237 tcp_mark_push(tp, skb);
1243 } else if (zc == MSG_SPLICE_PAGES) {
1244 /* Splice in data if we can; copy if we can't. */
1245 if (tcp_downgrade_zcopy_pure(sk, skb))
1246 goto wait_for_space;
1247 copy = tcp_wmem_schedule(sk, copy);
1249 goto wait_for_space;
1251 err = skb_splice_from_iter(skb, &msg->msg_iter, copy,
1254 if (err == -EMSGSIZE) {
1255 tcp_mark_push(tp, skb);
1262 if (!(flags & MSG_NO_SHARED_FRAGS))
1263 skb_shinfo(skb)->flags |= SKBFL_SHARED_FRAG;
1265 sk_wmem_queued_add(sk, copy);
1266 sk_mem_charge(sk, copy);
1270 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
1272 WRITE_ONCE(tp->write_seq, tp->write_seq + copy);
1273 TCP_SKB_CB(skb)->end_seq += copy;
1274 tcp_skb_pcount_set(skb, 0);
1277 if (!msg_data_left(msg)) {
1278 if (unlikely(flags & MSG_EOR))
1279 TCP_SKB_CB(skb)->eor = 1;
1283 if (skb->len < size_goal || (flags & MSG_OOB) || unlikely(tp->repair))
1286 if (forced_push(tp)) {
1287 tcp_mark_push(tp, skb);
1288 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
1289 } else if (skb == tcp_send_head(sk))
1290 tcp_push_one(sk, mss_now);
1294 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1295 tcp_remove_empty_skb(sk);
1297 tcp_push(sk, flags & ~MSG_MORE, mss_now,
1298 TCP_NAGLE_PUSH, size_goal);
1300 err = sk_stream_wait_memory(sk, &timeo);
1304 mss_now = tcp_send_mss(sk, &size_goal, flags);
1309 tcp_tx_timestamp(sk, sockc.tsflags);
1310 tcp_push(sk, flags, mss_now, tp->nonagle, size_goal);
1313 /* msg->msg_ubuf is pinned by the caller so we don't take extra refs */
1314 if (uarg && !msg->msg_ubuf)
1315 net_zcopy_put(uarg);
1316 return copied + copied_syn;
1319 tcp_remove_empty_skb(sk);
1321 if (copied + copied_syn)
1324 /* msg->msg_ubuf is pinned by the caller so we don't take extra refs */
1325 if (uarg && !msg->msg_ubuf)
1326 net_zcopy_put_abort(uarg, true);
1327 err = sk_stream_error(sk, flags, err);
1328 /* make sure we wake any epoll edge trigger waiter */
1329 if (unlikely(tcp_rtx_and_write_queues_empty(sk) && err == -EAGAIN)) {
1330 sk->sk_write_space(sk);
1331 tcp_chrono_stop(sk, TCP_CHRONO_SNDBUF_LIMITED);
1335 EXPORT_SYMBOL_GPL(tcp_sendmsg_locked);
1337 int tcp_sendmsg(struct sock *sk, struct msghdr *msg, size_t size)
1342 ret = tcp_sendmsg_locked(sk, msg, size);
1347 EXPORT_SYMBOL(tcp_sendmsg);
1349 void tcp_splice_eof(struct socket *sock)
1351 struct sock *sk = sock->sk;
1352 struct tcp_sock *tp = tcp_sk(sk);
1353 int mss_now, size_goal;
1355 if (!tcp_write_queue_tail(sk))
1359 mss_now = tcp_send_mss(sk, &size_goal, 0);
1360 tcp_push(sk, 0, mss_now, tp->nonagle, size_goal);
1363 EXPORT_SYMBOL_GPL(tcp_splice_eof);
1366 * Handle reading urgent data. BSD has very simple semantics for
1367 * this, no blocking and very strange errors 8)
1370 static int tcp_recv_urg(struct sock *sk, struct msghdr *msg, int len, int flags)
1372 struct tcp_sock *tp = tcp_sk(sk);
1374 /* No URG data to read. */
1375 if (sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data ||
1376 tp->urg_data == TCP_URG_READ)
1377 return -EINVAL; /* Yes this is right ! */
1379 if (sk->sk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DONE))
1382 if (tp->urg_data & TCP_URG_VALID) {
1384 char c = tp->urg_data;
1386 if (!(flags & MSG_PEEK))
1387 WRITE_ONCE(tp->urg_data, TCP_URG_READ);
1389 /* Read urgent data. */
1390 msg->msg_flags |= MSG_OOB;
1393 if (!(flags & MSG_TRUNC))
1394 err = memcpy_to_msg(msg, &c, 1);
1397 msg->msg_flags |= MSG_TRUNC;
1399 return err ? -EFAULT : len;
1402 if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
1405 /* Fixed the recv(..., MSG_OOB) behaviour. BSD docs and
1406 * the available implementations agree in this case:
1407 * this call should never block, independent of the
1408 * blocking state of the socket.
1409 * Mike <pall@rz.uni-karlsruhe.de>
1414 static int tcp_peek_sndq(struct sock *sk, struct msghdr *msg, int len)
1416 struct sk_buff *skb;
1417 int copied = 0, err = 0;
1419 /* XXX -- need to support SO_PEEK_OFF */
1421 skb_rbtree_walk(skb, &sk->tcp_rtx_queue) {
1422 err = skb_copy_datagram_msg(skb, 0, msg, skb->len);
1428 skb_queue_walk(&sk->sk_write_queue, skb) {
1429 err = skb_copy_datagram_msg(skb, 0, msg, skb->len);
1436 return err ?: copied;
1439 /* Clean up the receive buffer for full frames taken by the user,
1440 * then send an ACK if necessary. COPIED is the number of bytes
1441 * tcp_recvmsg has given to the user so far, it speeds up the
1442 * calculation of whether or not we must ACK for the sake of
1445 void __tcp_cleanup_rbuf(struct sock *sk, int copied)
1447 struct tcp_sock *tp = tcp_sk(sk);
1448 bool time_to_ack = false;
1450 if (inet_csk_ack_scheduled(sk)) {
1451 const struct inet_connection_sock *icsk = inet_csk(sk);
1453 if (/* Once-per-two-segments ACK was not sent by tcp_input.c */
1454 tp->rcv_nxt - tp->rcv_wup > icsk->icsk_ack.rcv_mss ||
1456 * If this read emptied read buffer, we send ACK, if
1457 * connection is not bidirectional, user drained
1458 * receive buffer and there was a small segment
1462 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED2) ||
1463 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED) &&
1464 !inet_csk_in_pingpong_mode(sk))) &&
1465 !atomic_read(&sk->sk_rmem_alloc)))
1469 /* We send an ACK if we can now advertise a non-zero window
1470 * which has been raised "significantly".
1472 * Even if window raised up to infinity, do not send window open ACK
1473 * in states, where we will not receive more. It is useless.
1475 if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) {
1476 __u32 rcv_window_now = tcp_receive_window(tp);
1478 /* Optimize, __tcp_select_window() is not cheap. */
1479 if (2*rcv_window_now <= tp->window_clamp) {
1480 __u32 new_window = __tcp_select_window(sk);
1482 /* Send ACK now, if this read freed lots of space
1483 * in our buffer. Certainly, new_window is new window.
1484 * We can advertise it now, if it is not less than current one.
1485 * "Lots" means "at least twice" here.
1487 if (new_window && new_window >= 2 * rcv_window_now)
1495 void tcp_cleanup_rbuf(struct sock *sk, int copied)
1497 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
1498 struct tcp_sock *tp = tcp_sk(sk);
1500 WARN(skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq),
1501 "cleanup rbuf bug: copied %X seq %X rcvnxt %X\n",
1502 tp->copied_seq, TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt);
1503 __tcp_cleanup_rbuf(sk, copied);
1506 static void tcp_eat_recv_skb(struct sock *sk, struct sk_buff *skb)
1508 __skb_unlink(skb, &sk->sk_receive_queue);
1509 if (likely(skb->destructor == sock_rfree)) {
1511 skb->destructor = NULL;
1513 return skb_attempt_defer_free(skb);
1518 struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off)
1520 struct sk_buff *skb;
1523 while ((skb = skb_peek(&sk->sk_receive_queue)) != NULL) {
1524 offset = seq - TCP_SKB_CB(skb)->seq;
1525 if (unlikely(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
1526 pr_err_once("%s: found a SYN, please report !\n", __func__);
1529 if (offset < skb->len || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)) {
1533 /* This looks weird, but this can happen if TCP collapsing
1534 * splitted a fat GRO packet, while we released socket lock
1535 * in skb_splice_bits()
1537 tcp_eat_recv_skb(sk, skb);
1541 EXPORT_SYMBOL(tcp_recv_skb);
1544 * This routine provides an alternative to tcp_recvmsg() for routines
1545 * that would like to handle copying from skbuffs directly in 'sendfile'
1548 * - It is assumed that the socket was locked by the caller.
1549 * - The routine does not block.
1550 * - At present, there is no support for reading OOB data
1551 * or for 'peeking' the socket using this routine
1552 * (although both would be easy to implement).
1554 int tcp_read_sock(struct sock *sk, read_descriptor_t *desc,
1555 sk_read_actor_t recv_actor)
1557 struct sk_buff *skb;
1558 struct tcp_sock *tp = tcp_sk(sk);
1559 u32 seq = tp->copied_seq;
1563 if (sk->sk_state == TCP_LISTEN)
1565 while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1566 if (offset < skb->len) {
1570 len = skb->len - offset;
1571 /* Stop reading if we hit a patch of urgent data */
1572 if (unlikely(tp->urg_data)) {
1573 u32 urg_offset = tp->urg_seq - seq;
1574 if (urg_offset < len)
1579 used = recv_actor(desc, skb, offset, len);
1585 if (WARN_ON_ONCE(used > len))
1591 /* If recv_actor drops the lock (e.g. TCP splice
1592 * receive) the skb pointer might be invalid when
1593 * getting here: tcp_collapse might have deleted it
1594 * while aggregating skbs from the socket queue.
1596 skb = tcp_recv_skb(sk, seq - 1, &offset);
1599 /* TCP coalescing might have appended data to the skb.
1600 * Try to splice more frags
1602 if (offset + 1 != skb->len)
1605 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) {
1606 tcp_eat_recv_skb(sk, skb);
1610 tcp_eat_recv_skb(sk, skb);
1613 WRITE_ONCE(tp->copied_seq, seq);
1615 WRITE_ONCE(tp->copied_seq, seq);
1617 tcp_rcv_space_adjust(sk);
1619 /* Clean up data we have read: This will do ACK frames. */
1621 tcp_recv_skb(sk, seq, &offset);
1622 tcp_cleanup_rbuf(sk, copied);
1626 EXPORT_SYMBOL(tcp_read_sock);
1628 int tcp_read_skb(struct sock *sk, skb_read_actor_t recv_actor)
1630 struct sk_buff *skb;
1633 if (sk->sk_state == TCP_LISTEN)
1636 while ((skb = skb_peek(&sk->sk_receive_queue)) != NULL) {
1640 __skb_unlink(skb, &sk->sk_receive_queue);
1641 WARN_ON_ONCE(!skb_set_owner_sk_safe(skb, sk));
1642 tcp_flags = TCP_SKB_CB(skb)->tcp_flags;
1643 used = recv_actor(sk, skb);
1651 if (tcp_flags & TCPHDR_FIN)
1656 EXPORT_SYMBOL(tcp_read_skb);
1658 void tcp_read_done(struct sock *sk, size_t len)
1660 struct tcp_sock *tp = tcp_sk(sk);
1661 u32 seq = tp->copied_seq;
1662 struct sk_buff *skb;
1666 if (sk->sk_state == TCP_LISTEN)
1670 while (left && (skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1673 used = min_t(size_t, skb->len - offset, left);
1677 if (skb->len > offset + used)
1680 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) {
1681 tcp_eat_recv_skb(sk, skb);
1685 tcp_eat_recv_skb(sk, skb);
1687 WRITE_ONCE(tp->copied_seq, seq);
1689 tcp_rcv_space_adjust(sk);
1691 /* Clean up data we have read: This will do ACK frames. */
1693 tcp_cleanup_rbuf(sk, len - left);
1695 EXPORT_SYMBOL(tcp_read_done);
1697 int tcp_peek_len(struct socket *sock)
1699 return tcp_inq(sock->sk);
1701 EXPORT_SYMBOL(tcp_peek_len);
1703 /* Make sure sk_rcvbuf is big enough to satisfy SO_RCVLOWAT hint */
1704 int tcp_set_rcvlowat(struct sock *sk, int val)
1708 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK)
1709 cap = sk->sk_rcvbuf >> 1;
1711 cap = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_rmem[2]) >> 1;
1712 val = min(val, cap);
1713 WRITE_ONCE(sk->sk_rcvlowat, val ? : 1);
1715 /* Check if we need to signal EPOLLIN right now */
1718 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK)
1721 space = tcp_space_from_win(sk, val);
1722 if (space > sk->sk_rcvbuf) {
1723 WRITE_ONCE(sk->sk_rcvbuf, space);
1724 tcp_sk(sk)->window_clamp = val;
1728 EXPORT_SYMBOL(tcp_set_rcvlowat);
1730 void tcp_update_recv_tstamps(struct sk_buff *skb,
1731 struct scm_timestamping_internal *tss)
1734 tss->ts[0] = ktime_to_timespec64(skb->tstamp);
1736 tss->ts[0] = (struct timespec64) {0};
1738 if (skb_hwtstamps(skb)->hwtstamp)
1739 tss->ts[2] = ktime_to_timespec64(skb_hwtstamps(skb)->hwtstamp);
1741 tss->ts[2] = (struct timespec64) {0};
1745 static const struct vm_operations_struct tcp_vm_ops = {
1748 int tcp_mmap(struct file *file, struct socket *sock,
1749 struct vm_area_struct *vma)
1751 if (vma->vm_flags & (VM_WRITE | VM_EXEC))
1753 vm_flags_clear(vma, VM_MAYWRITE | VM_MAYEXEC);
1755 /* Instruct vm_insert_page() to not mmap_read_lock(mm) */
1756 vm_flags_set(vma, VM_MIXEDMAP);
1758 vma->vm_ops = &tcp_vm_ops;
1761 EXPORT_SYMBOL(tcp_mmap);
1763 static skb_frag_t *skb_advance_to_frag(struct sk_buff *skb, u32 offset_skb,
1768 if (unlikely(offset_skb >= skb->len))
1771 offset_skb -= skb_headlen(skb);
1772 if ((int)offset_skb < 0 || skb_has_frag_list(skb))
1775 frag = skb_shinfo(skb)->frags;
1776 while (offset_skb) {
1777 if (skb_frag_size(frag) > offset_skb) {
1778 *offset_frag = offset_skb;
1781 offset_skb -= skb_frag_size(frag);
1788 static bool can_map_frag(const skb_frag_t *frag)
1792 if (skb_frag_size(frag) != PAGE_SIZE || skb_frag_off(frag))
1795 page = skb_frag_page(frag);
1797 if (PageCompound(page) || page->mapping)
1803 static int find_next_mappable_frag(const skb_frag_t *frag,
1804 int remaining_in_skb)
1808 if (likely(can_map_frag(frag)))
1811 while (offset < remaining_in_skb && !can_map_frag(frag)) {
1812 offset += skb_frag_size(frag);
1818 static void tcp_zerocopy_set_hint_for_skb(struct sock *sk,
1819 struct tcp_zerocopy_receive *zc,
1820 struct sk_buff *skb, u32 offset)
1822 u32 frag_offset, partial_frag_remainder = 0;
1823 int mappable_offset;
1826 /* worst case: skip to next skb. try to improve on this case below */
1827 zc->recv_skip_hint = skb->len - offset;
1829 /* Find the frag containing this offset (and how far into that frag) */
1830 frag = skb_advance_to_frag(skb, offset, &frag_offset);
1835 struct skb_shared_info *info = skb_shinfo(skb);
1837 /* We read part of the last frag, must recvmsg() rest of skb. */
1838 if (frag == &info->frags[info->nr_frags - 1])
1841 /* Else, we must at least read the remainder in this frag. */
1842 partial_frag_remainder = skb_frag_size(frag) - frag_offset;
1843 zc->recv_skip_hint -= partial_frag_remainder;
1847 /* partial_frag_remainder: If part way through a frag, must read rest.
1848 * mappable_offset: Bytes till next mappable frag, *not* counting bytes
1849 * in partial_frag_remainder.
1851 mappable_offset = find_next_mappable_frag(frag, zc->recv_skip_hint);
1852 zc->recv_skip_hint = mappable_offset + partial_frag_remainder;
1855 static int tcp_recvmsg_locked(struct sock *sk, struct msghdr *msg, size_t len,
1856 int flags, struct scm_timestamping_internal *tss,
1858 static int receive_fallback_to_copy(struct sock *sk,
1859 struct tcp_zerocopy_receive *zc, int inq,
1860 struct scm_timestamping_internal *tss)
1862 unsigned long copy_address = (unsigned long)zc->copybuf_address;
1863 struct msghdr msg = {};
1867 zc->recv_skip_hint = 0;
1869 if (copy_address != zc->copybuf_address)
1872 err = import_ubuf(ITER_DEST, (void __user *)copy_address, inq,
1877 err = tcp_recvmsg_locked(sk, &msg, inq, MSG_DONTWAIT,
1878 tss, &zc->msg_flags);
1882 zc->copybuf_len = err;
1883 if (likely(zc->copybuf_len)) {
1884 struct sk_buff *skb;
1887 skb = tcp_recv_skb(sk, tcp_sk(sk)->copied_seq, &offset);
1889 tcp_zerocopy_set_hint_for_skb(sk, zc, skb, offset);
1894 static int tcp_copy_straggler_data(struct tcp_zerocopy_receive *zc,
1895 struct sk_buff *skb, u32 copylen,
1896 u32 *offset, u32 *seq)
1898 unsigned long copy_address = (unsigned long)zc->copybuf_address;
1899 struct msghdr msg = {};
1902 if (copy_address != zc->copybuf_address)
1905 err = import_ubuf(ITER_DEST, (void __user *)copy_address, copylen,
1909 err = skb_copy_datagram_msg(skb, *offset, &msg, copylen);
1912 zc->recv_skip_hint -= copylen;
1915 return (__s32)copylen;
1918 static int tcp_zc_handle_leftover(struct tcp_zerocopy_receive *zc,
1920 struct sk_buff *skb,
1923 struct scm_timestamping_internal *tss)
1925 u32 offset, copylen = min_t(u32, copybuf_len, zc->recv_skip_hint);
1929 /* skb is null if inq < PAGE_SIZE. */
1931 offset = *seq - TCP_SKB_CB(skb)->seq;
1933 skb = tcp_recv_skb(sk, *seq, &offset);
1934 if (TCP_SKB_CB(skb)->has_rxtstamp) {
1935 tcp_update_recv_tstamps(skb, tss);
1936 zc->msg_flags |= TCP_CMSG_TS;
1940 zc->copybuf_len = tcp_copy_straggler_data(zc, skb, copylen, &offset,
1942 return zc->copybuf_len < 0 ? 0 : copylen;
1945 static int tcp_zerocopy_vm_insert_batch_error(struct vm_area_struct *vma,
1946 struct page **pending_pages,
1947 unsigned long pages_remaining,
1948 unsigned long *address,
1951 struct tcp_zerocopy_receive *zc,
1952 u32 total_bytes_to_map,
1955 /* At least one page did not map. Try zapping if we skipped earlier. */
1956 if (err == -EBUSY &&
1957 zc->flags & TCP_RECEIVE_ZEROCOPY_FLAG_TLB_CLEAN_HINT) {
1960 maybe_zap_len = total_bytes_to_map - /* All bytes to map */
1961 *length + /* Mapped or pending */
1962 (pages_remaining * PAGE_SIZE); /* Failed map. */
1963 zap_page_range_single(vma, *address, maybe_zap_len, NULL);
1968 unsigned long leftover_pages = pages_remaining;
1971 /* We called zap_page_range_single, try to reinsert. */
1972 err = vm_insert_pages(vma, *address,
1975 bytes_mapped = PAGE_SIZE * (leftover_pages - pages_remaining);
1976 *seq += bytes_mapped;
1977 *address += bytes_mapped;
1980 /* Either we were unable to zap, OR we zapped, retried an
1981 * insert, and still had an issue. Either ways, pages_remaining
1982 * is the number of pages we were unable to map, and we unroll
1983 * some state we speculatively touched before.
1985 const int bytes_not_mapped = PAGE_SIZE * pages_remaining;
1987 *length -= bytes_not_mapped;
1988 zc->recv_skip_hint += bytes_not_mapped;
1993 static int tcp_zerocopy_vm_insert_batch(struct vm_area_struct *vma,
1994 struct page **pages,
1995 unsigned int pages_to_map,
1996 unsigned long *address,
1999 struct tcp_zerocopy_receive *zc,
2000 u32 total_bytes_to_map)
2002 unsigned long pages_remaining = pages_to_map;
2003 unsigned int pages_mapped;
2004 unsigned int bytes_mapped;
2007 err = vm_insert_pages(vma, *address, pages, &pages_remaining);
2008 pages_mapped = pages_to_map - (unsigned int)pages_remaining;
2009 bytes_mapped = PAGE_SIZE * pages_mapped;
2010 /* Even if vm_insert_pages fails, it may have partially succeeded in
2011 * mapping (some but not all of the pages).
2013 *seq += bytes_mapped;
2014 *address += bytes_mapped;
2019 /* Error: maybe zap and retry + rollback state for failed inserts. */
2020 return tcp_zerocopy_vm_insert_batch_error(vma, pages + pages_mapped,
2021 pages_remaining, address, length, seq, zc, total_bytes_to_map,
2025 #define TCP_VALID_ZC_MSG_FLAGS (TCP_CMSG_TS)
2026 static void tcp_zc_finalize_rx_tstamp(struct sock *sk,
2027 struct tcp_zerocopy_receive *zc,
2028 struct scm_timestamping_internal *tss)
2030 unsigned long msg_control_addr;
2031 struct msghdr cmsg_dummy;
2033 msg_control_addr = (unsigned long)zc->msg_control;
2034 cmsg_dummy.msg_control_user = (void __user *)msg_control_addr;
2035 cmsg_dummy.msg_controllen =
2036 (__kernel_size_t)zc->msg_controllen;
2037 cmsg_dummy.msg_flags = in_compat_syscall()
2038 ? MSG_CMSG_COMPAT : 0;
2039 cmsg_dummy.msg_control_is_user = true;
2041 if (zc->msg_control == msg_control_addr &&
2042 zc->msg_controllen == cmsg_dummy.msg_controllen) {
2043 tcp_recv_timestamp(&cmsg_dummy, sk, tss);
2044 zc->msg_control = (__u64)
2045 ((uintptr_t)cmsg_dummy.msg_control_user);
2046 zc->msg_controllen =
2047 (__u64)cmsg_dummy.msg_controllen;
2048 zc->msg_flags = (__u32)cmsg_dummy.msg_flags;
2052 static struct vm_area_struct *find_tcp_vma(struct mm_struct *mm,
2053 unsigned long address,
2056 struct vm_area_struct *vma = lock_vma_under_rcu(mm, address);
2059 if (vma->vm_ops != &tcp_vm_ops) {
2063 *mmap_locked = false;
2068 vma = vma_lookup(mm, address);
2069 if (!vma || vma->vm_ops != &tcp_vm_ops) {
2070 mmap_read_unlock(mm);
2073 *mmap_locked = true;
2077 #define TCP_ZEROCOPY_PAGE_BATCH_SIZE 32
2078 static int tcp_zerocopy_receive(struct sock *sk,
2079 struct tcp_zerocopy_receive *zc,
2080 struct scm_timestamping_internal *tss)
2082 u32 length = 0, offset, vma_len, avail_len, copylen = 0;
2083 unsigned long address = (unsigned long)zc->address;
2084 struct page *pages[TCP_ZEROCOPY_PAGE_BATCH_SIZE];
2085 s32 copybuf_len = zc->copybuf_len;
2086 struct tcp_sock *tp = tcp_sk(sk);
2087 const skb_frag_t *frags = NULL;
2088 unsigned int pages_to_map = 0;
2089 struct vm_area_struct *vma;
2090 struct sk_buff *skb = NULL;
2091 u32 seq = tp->copied_seq;
2092 u32 total_bytes_to_map;
2093 int inq = tcp_inq(sk);
2097 zc->copybuf_len = 0;
2100 if (address & (PAGE_SIZE - 1) || address != zc->address)
2103 if (sk->sk_state == TCP_LISTEN)
2106 sock_rps_record_flow(sk);
2108 if (inq && inq <= copybuf_len)
2109 return receive_fallback_to_copy(sk, zc, inq, tss);
2111 if (inq < PAGE_SIZE) {
2113 zc->recv_skip_hint = inq;
2114 if (!inq && sock_flag(sk, SOCK_DONE))
2119 vma = find_tcp_vma(current->mm, address, &mmap_locked);
2123 vma_len = min_t(unsigned long, zc->length, vma->vm_end - address);
2124 avail_len = min_t(u32, vma_len, inq);
2125 total_bytes_to_map = avail_len & ~(PAGE_SIZE - 1);
2126 if (total_bytes_to_map) {
2127 if (!(zc->flags & TCP_RECEIVE_ZEROCOPY_FLAG_TLB_CLEAN_HINT))
2128 zap_page_range_single(vma, address, total_bytes_to_map,
2130 zc->length = total_bytes_to_map;
2131 zc->recv_skip_hint = 0;
2133 zc->length = avail_len;
2134 zc->recv_skip_hint = avail_len;
2137 while (length + PAGE_SIZE <= zc->length) {
2138 int mappable_offset;
2141 if (zc->recv_skip_hint < PAGE_SIZE) {
2145 if (zc->recv_skip_hint > 0)
2148 offset = seq - TCP_SKB_CB(skb)->seq;
2150 skb = tcp_recv_skb(sk, seq, &offset);
2153 if (TCP_SKB_CB(skb)->has_rxtstamp) {
2154 tcp_update_recv_tstamps(skb, tss);
2155 zc->msg_flags |= TCP_CMSG_TS;
2157 zc->recv_skip_hint = skb->len - offset;
2158 frags = skb_advance_to_frag(skb, offset, &offset_frag);
2159 if (!frags || offset_frag)
2163 mappable_offset = find_next_mappable_frag(frags,
2164 zc->recv_skip_hint);
2165 if (mappable_offset) {
2166 zc->recv_skip_hint = mappable_offset;
2169 page = skb_frag_page(frags);
2171 pages[pages_to_map++] = page;
2172 length += PAGE_SIZE;
2173 zc->recv_skip_hint -= PAGE_SIZE;
2175 if (pages_to_map == TCP_ZEROCOPY_PAGE_BATCH_SIZE ||
2176 zc->recv_skip_hint < PAGE_SIZE) {
2177 /* Either full batch, or we're about to go to next skb
2178 * (and we cannot unroll failed ops across skbs).
2180 ret = tcp_zerocopy_vm_insert_batch(vma, pages,
2184 total_bytes_to_map);
2191 ret = tcp_zerocopy_vm_insert_batch(vma, pages, pages_to_map,
2192 &address, &length, &seq,
2193 zc, total_bytes_to_map);
2197 mmap_read_unlock(current->mm);
2200 /* Try to copy straggler data. */
2202 copylen = tcp_zc_handle_leftover(zc, sk, skb, &seq, copybuf_len, tss);
2204 if (length + copylen) {
2205 WRITE_ONCE(tp->copied_seq, seq);
2206 tcp_rcv_space_adjust(sk);
2208 /* Clean up data we have read: This will do ACK frames. */
2209 tcp_recv_skb(sk, seq, &offset);
2210 tcp_cleanup_rbuf(sk, length + copylen);
2212 if (length == zc->length)
2213 zc->recv_skip_hint = 0;
2215 if (!zc->recv_skip_hint && sock_flag(sk, SOCK_DONE))
2218 zc->length = length;
2223 /* Similar to __sock_recv_timestamp, but does not require an skb */
2224 void tcp_recv_timestamp(struct msghdr *msg, const struct sock *sk,
2225 struct scm_timestamping_internal *tss)
2227 int new_tstamp = sock_flag(sk, SOCK_TSTAMP_NEW);
2228 bool has_timestamping = false;
2230 if (tss->ts[0].tv_sec || tss->ts[0].tv_nsec) {
2231 if (sock_flag(sk, SOCK_RCVTSTAMP)) {
2232 if (sock_flag(sk, SOCK_RCVTSTAMPNS)) {
2234 struct __kernel_timespec kts = {
2235 .tv_sec = tss->ts[0].tv_sec,
2236 .tv_nsec = tss->ts[0].tv_nsec,
2238 put_cmsg(msg, SOL_SOCKET, SO_TIMESTAMPNS_NEW,
2241 struct __kernel_old_timespec ts_old = {
2242 .tv_sec = tss->ts[0].tv_sec,
2243 .tv_nsec = tss->ts[0].tv_nsec,
2245 put_cmsg(msg, SOL_SOCKET, SO_TIMESTAMPNS_OLD,
2246 sizeof(ts_old), &ts_old);
2250 struct __kernel_sock_timeval stv = {
2251 .tv_sec = tss->ts[0].tv_sec,
2252 .tv_usec = tss->ts[0].tv_nsec / 1000,
2254 put_cmsg(msg, SOL_SOCKET, SO_TIMESTAMP_NEW,
2257 struct __kernel_old_timeval tv = {
2258 .tv_sec = tss->ts[0].tv_sec,
2259 .tv_usec = tss->ts[0].tv_nsec / 1000,
2261 put_cmsg(msg, SOL_SOCKET, SO_TIMESTAMP_OLD,
2267 if (READ_ONCE(sk->sk_tsflags) & SOF_TIMESTAMPING_SOFTWARE)
2268 has_timestamping = true;
2270 tss->ts[0] = (struct timespec64) {0};
2273 if (tss->ts[2].tv_sec || tss->ts[2].tv_nsec) {
2274 if (READ_ONCE(sk->sk_tsflags) & SOF_TIMESTAMPING_RAW_HARDWARE)
2275 has_timestamping = true;
2277 tss->ts[2] = (struct timespec64) {0};
2280 if (has_timestamping) {
2281 tss->ts[1] = (struct timespec64) {0};
2282 if (sock_flag(sk, SOCK_TSTAMP_NEW))
2283 put_cmsg_scm_timestamping64(msg, tss);
2285 put_cmsg_scm_timestamping(msg, tss);
2289 static int tcp_inq_hint(struct sock *sk)
2291 const struct tcp_sock *tp = tcp_sk(sk);
2292 u32 copied_seq = READ_ONCE(tp->copied_seq);
2293 u32 rcv_nxt = READ_ONCE(tp->rcv_nxt);
2296 inq = rcv_nxt - copied_seq;
2297 if (unlikely(inq < 0 || copied_seq != READ_ONCE(tp->copied_seq))) {
2299 inq = tp->rcv_nxt - tp->copied_seq;
2302 /* After receiving a FIN, tell the user-space to continue reading
2303 * by returning a non-zero inq.
2305 if (inq == 0 && sock_flag(sk, SOCK_DONE))
2311 * This routine copies from a sock struct into the user buffer.
2313 * Technical note: in 2.3 we work on _locked_ socket, so that
2314 * tricks with *seq access order and skb->users are not required.
2315 * Probably, code can be easily improved even more.
2318 static int tcp_recvmsg_locked(struct sock *sk, struct msghdr *msg, size_t len,
2319 int flags, struct scm_timestamping_internal *tss,
2322 struct tcp_sock *tp = tcp_sk(sk);
2328 int target; /* Read at least this many bytes */
2330 struct sk_buff *skb, *last;
2334 if (sk->sk_state == TCP_LISTEN)
2337 if (tp->recvmsg_inq) {
2338 *cmsg_flags = TCP_CMSG_INQ;
2339 msg->msg_get_inq = 1;
2341 timeo = sock_rcvtimeo(sk, flags & MSG_DONTWAIT);
2343 /* Urgent data needs to be handled specially. */
2344 if (flags & MSG_OOB)
2347 if (unlikely(tp->repair)) {
2349 if (!(flags & MSG_PEEK))
2352 if (tp->repair_queue == TCP_SEND_QUEUE)
2356 if (tp->repair_queue == TCP_NO_QUEUE)
2359 /* 'common' recv queue MSG_PEEK-ing */
2362 seq = &tp->copied_seq;
2363 if (flags & MSG_PEEK) {
2364 peek_seq = tp->copied_seq;
2368 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
2373 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
2374 if (unlikely(tp->urg_data) && tp->urg_seq == *seq) {
2377 if (signal_pending(current)) {
2378 copied = timeo ? sock_intr_errno(timeo) : -EAGAIN;
2383 /* Next get a buffer. */
2385 last = skb_peek_tail(&sk->sk_receive_queue);
2386 skb_queue_walk(&sk->sk_receive_queue, skb) {
2388 /* Now that we have two receive queues this
2391 if (WARN(before(*seq, TCP_SKB_CB(skb)->seq),
2392 "TCP recvmsg seq # bug: copied %X, seq %X, rcvnxt %X, fl %X\n",
2393 *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt,
2397 offset = *seq - TCP_SKB_CB(skb)->seq;
2398 if (unlikely(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
2399 pr_err_once("%s: found a SYN, please report !\n", __func__);
2402 if (offset < skb->len)
2404 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
2406 WARN(!(flags & MSG_PEEK),
2407 "TCP recvmsg seq # bug 2: copied %X, seq %X, rcvnxt %X, fl %X\n",
2408 *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt, flags);
2411 /* Well, if we have backlog, try to process it now yet. */
2413 if (copied >= target && !READ_ONCE(sk->sk_backlog.tail))
2419 sk->sk_state == TCP_CLOSE ||
2420 (sk->sk_shutdown & RCV_SHUTDOWN) ||
2421 signal_pending(current))
2424 if (sock_flag(sk, SOCK_DONE))
2428 copied = sock_error(sk);
2432 if (sk->sk_shutdown & RCV_SHUTDOWN)
2435 if (sk->sk_state == TCP_CLOSE) {
2436 /* This occurs when user tries to read
2437 * from never connected socket.
2448 if (signal_pending(current)) {
2449 copied = sock_intr_errno(timeo);
2454 if (copied >= target) {
2455 /* Do not sleep, just process backlog. */
2456 __sk_flush_backlog(sk);
2458 tcp_cleanup_rbuf(sk, copied);
2459 err = sk_wait_data(sk, &timeo, last);
2461 err = copied ? : err;
2466 if ((flags & MSG_PEEK) &&
2467 (peek_seq - copied - urg_hole != tp->copied_seq)) {
2468 net_dbg_ratelimited("TCP(%s:%d): Application bug, race in MSG_PEEK\n",
2470 task_pid_nr(current));
2471 peek_seq = tp->copied_seq;
2476 /* Ok so how much can we use? */
2477 used = skb->len - offset;
2481 /* Do we have urgent data here? */
2482 if (unlikely(tp->urg_data)) {
2483 u32 urg_offset = tp->urg_seq - *seq;
2484 if (urg_offset < used) {
2486 if (!sock_flag(sk, SOCK_URGINLINE)) {
2487 WRITE_ONCE(*seq, *seq + 1);
2499 if (!(flags & MSG_TRUNC)) {
2500 err = skb_copy_datagram_msg(skb, offset, msg, used);
2502 /* Exception. Bailout! */
2509 WRITE_ONCE(*seq, *seq + used);
2513 tcp_rcv_space_adjust(sk);
2516 if (unlikely(tp->urg_data) && after(tp->copied_seq, tp->urg_seq)) {
2517 WRITE_ONCE(tp->urg_data, 0);
2518 tcp_fast_path_check(sk);
2521 if (TCP_SKB_CB(skb)->has_rxtstamp) {
2522 tcp_update_recv_tstamps(skb, tss);
2523 *cmsg_flags |= TCP_CMSG_TS;
2526 if (used + offset < skb->len)
2529 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
2531 if (!(flags & MSG_PEEK))
2532 tcp_eat_recv_skb(sk, skb);
2536 /* Process the FIN. */
2537 WRITE_ONCE(*seq, *seq + 1);
2538 if (!(flags & MSG_PEEK))
2539 tcp_eat_recv_skb(sk, skb);
2543 /* According to UNIX98, msg_name/msg_namelen are ignored
2544 * on connected socket. I was just happy when found this 8) --ANK
2547 /* Clean up data we have read: This will do ACK frames. */
2548 tcp_cleanup_rbuf(sk, copied);
2555 err = tcp_recv_urg(sk, msg, len, flags);
2559 err = tcp_peek_sndq(sk, msg, len);
2563 int tcp_recvmsg(struct sock *sk, struct msghdr *msg, size_t len, int flags,
2566 int cmsg_flags = 0, ret;
2567 struct scm_timestamping_internal tss;
2569 if (unlikely(flags & MSG_ERRQUEUE))
2570 return inet_recv_error(sk, msg, len, addr_len);
2572 if (sk_can_busy_loop(sk) &&
2573 skb_queue_empty_lockless(&sk->sk_receive_queue) &&
2574 sk->sk_state == TCP_ESTABLISHED)
2575 sk_busy_loop(sk, flags & MSG_DONTWAIT);
2578 ret = tcp_recvmsg_locked(sk, msg, len, flags, &tss, &cmsg_flags);
2581 if ((cmsg_flags || msg->msg_get_inq) && ret >= 0) {
2582 if (cmsg_flags & TCP_CMSG_TS)
2583 tcp_recv_timestamp(msg, sk, &tss);
2584 if (msg->msg_get_inq) {
2585 msg->msg_inq = tcp_inq_hint(sk);
2586 if (cmsg_flags & TCP_CMSG_INQ)
2587 put_cmsg(msg, SOL_TCP, TCP_CM_INQ,
2588 sizeof(msg->msg_inq), &msg->msg_inq);
2593 EXPORT_SYMBOL(tcp_recvmsg);
2595 void tcp_set_state(struct sock *sk, int state)
2597 int oldstate = sk->sk_state;
2599 /* We defined a new enum for TCP states that are exported in BPF
2600 * so as not force the internal TCP states to be frozen. The
2601 * following checks will detect if an internal state value ever
2602 * differs from the BPF value. If this ever happens, then we will
2603 * need to remap the internal value to the BPF value before calling
2604 * tcp_call_bpf_2arg.
2606 BUILD_BUG_ON((int)BPF_TCP_ESTABLISHED != (int)TCP_ESTABLISHED);
2607 BUILD_BUG_ON((int)BPF_TCP_SYN_SENT != (int)TCP_SYN_SENT);
2608 BUILD_BUG_ON((int)BPF_TCP_SYN_RECV != (int)TCP_SYN_RECV);
2609 BUILD_BUG_ON((int)BPF_TCP_FIN_WAIT1 != (int)TCP_FIN_WAIT1);
2610 BUILD_BUG_ON((int)BPF_TCP_FIN_WAIT2 != (int)TCP_FIN_WAIT2);
2611 BUILD_BUG_ON((int)BPF_TCP_TIME_WAIT != (int)TCP_TIME_WAIT);
2612 BUILD_BUG_ON((int)BPF_TCP_CLOSE != (int)TCP_CLOSE);
2613 BUILD_BUG_ON((int)BPF_TCP_CLOSE_WAIT != (int)TCP_CLOSE_WAIT);
2614 BUILD_BUG_ON((int)BPF_TCP_LAST_ACK != (int)TCP_LAST_ACK);
2615 BUILD_BUG_ON((int)BPF_TCP_LISTEN != (int)TCP_LISTEN);
2616 BUILD_BUG_ON((int)BPF_TCP_CLOSING != (int)TCP_CLOSING);
2617 BUILD_BUG_ON((int)BPF_TCP_NEW_SYN_RECV != (int)TCP_NEW_SYN_RECV);
2618 BUILD_BUG_ON((int)BPF_TCP_BOUND_INACTIVE != (int)TCP_BOUND_INACTIVE);
2619 BUILD_BUG_ON((int)BPF_TCP_MAX_STATES != (int)TCP_MAX_STATES);
2621 /* bpf uapi header bpf.h defines an anonymous enum with values
2622 * BPF_TCP_* used by bpf programs. Currently gcc built vmlinux
2623 * is able to emit this enum in DWARF due to the above BUILD_BUG_ON.
2624 * But clang built vmlinux does not have this enum in DWARF
2625 * since clang removes the above code before generating IR/debuginfo.
2626 * Let us explicitly emit the type debuginfo to ensure the
2627 * above-mentioned anonymous enum in the vmlinux DWARF and hence BTF
2628 * regardless of which compiler is used.
2630 BTF_TYPE_EMIT_ENUM(BPF_TCP_ESTABLISHED);
2632 if (BPF_SOCK_OPS_TEST_FLAG(tcp_sk(sk), BPF_SOCK_OPS_STATE_CB_FLAG))
2633 tcp_call_bpf_2arg(sk, BPF_SOCK_OPS_STATE_CB, oldstate, state);
2636 case TCP_ESTABLISHED:
2637 if (oldstate != TCP_ESTABLISHED)
2638 TCP_INC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
2642 if (oldstate == TCP_CLOSE_WAIT || oldstate == TCP_ESTABLISHED)
2643 TCP_INC_STATS(sock_net(sk), TCP_MIB_ESTABRESETS);
2645 sk->sk_prot->unhash(sk);
2646 if (inet_csk(sk)->icsk_bind_hash &&
2647 !(sk->sk_userlocks & SOCK_BINDPORT_LOCK))
2651 if (oldstate == TCP_ESTABLISHED)
2652 TCP_DEC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
2655 /* Change state AFTER socket is unhashed to avoid closed
2656 * socket sitting in hash tables.
2658 inet_sk_state_store(sk, state);
2660 EXPORT_SYMBOL_GPL(tcp_set_state);
2663 * State processing on a close. This implements the state shift for
2664 * sending our FIN frame. Note that we only send a FIN for some
2665 * states. A shutdown() may have already sent the FIN, or we may be
2669 static const unsigned char new_state[16] = {
2670 /* current state: new state: action: */
2671 [0 /* (Invalid) */] = TCP_CLOSE,
2672 [TCP_ESTABLISHED] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2673 [TCP_SYN_SENT] = TCP_CLOSE,
2674 [TCP_SYN_RECV] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2675 [TCP_FIN_WAIT1] = TCP_FIN_WAIT1,
2676 [TCP_FIN_WAIT2] = TCP_FIN_WAIT2,
2677 [TCP_TIME_WAIT] = TCP_CLOSE,
2678 [TCP_CLOSE] = TCP_CLOSE,
2679 [TCP_CLOSE_WAIT] = TCP_LAST_ACK | TCP_ACTION_FIN,
2680 [TCP_LAST_ACK] = TCP_LAST_ACK,
2681 [TCP_LISTEN] = TCP_CLOSE,
2682 [TCP_CLOSING] = TCP_CLOSING,
2683 [TCP_NEW_SYN_RECV] = TCP_CLOSE, /* should not happen ! */
2686 static int tcp_close_state(struct sock *sk)
2688 int next = (int)new_state[sk->sk_state];
2689 int ns = next & TCP_STATE_MASK;
2691 tcp_set_state(sk, ns);
2693 return next & TCP_ACTION_FIN;
2697 * Shutdown the sending side of a connection. Much like close except
2698 * that we don't receive shut down or sock_set_flag(sk, SOCK_DEAD).
2701 void tcp_shutdown(struct sock *sk, int how)
2703 /* We need to grab some memory, and put together a FIN,
2704 * and then put it into the queue to be sent.
2705 * Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
2707 if (!(how & SEND_SHUTDOWN))
2710 /* If we've already sent a FIN, or it's a closed state, skip this. */
2711 if ((1 << sk->sk_state) &
2712 (TCPF_ESTABLISHED | TCPF_SYN_SENT |
2713 TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) {
2714 /* Clear out any half completed packets. FIN if needed. */
2715 if (tcp_close_state(sk))
2719 EXPORT_SYMBOL(tcp_shutdown);
2721 int tcp_orphan_count_sum(void)
2725 for_each_possible_cpu(i)
2726 total += per_cpu(tcp_orphan_count, i);
2728 return max(total, 0);
2731 static int tcp_orphan_cache;
2732 static struct timer_list tcp_orphan_timer;
2733 #define TCP_ORPHAN_TIMER_PERIOD msecs_to_jiffies(100)
2735 static void tcp_orphan_update(struct timer_list *unused)
2737 WRITE_ONCE(tcp_orphan_cache, tcp_orphan_count_sum());
2738 mod_timer(&tcp_orphan_timer, jiffies + TCP_ORPHAN_TIMER_PERIOD);
2741 static bool tcp_too_many_orphans(int shift)
2743 return READ_ONCE(tcp_orphan_cache) << shift >
2744 READ_ONCE(sysctl_tcp_max_orphans);
2747 bool tcp_check_oom(struct sock *sk, int shift)
2749 bool too_many_orphans, out_of_socket_memory;
2751 too_many_orphans = tcp_too_many_orphans(shift);
2752 out_of_socket_memory = tcp_out_of_memory(sk);
2754 if (too_many_orphans)
2755 net_info_ratelimited("too many orphaned sockets\n");
2756 if (out_of_socket_memory)
2757 net_info_ratelimited("out of memory -- consider tuning tcp_mem\n");
2758 return too_many_orphans || out_of_socket_memory;
2761 void __tcp_close(struct sock *sk, long timeout)
2763 struct sk_buff *skb;
2764 int data_was_unread = 0;
2767 WRITE_ONCE(sk->sk_shutdown, SHUTDOWN_MASK);
2769 if (sk->sk_state == TCP_LISTEN) {
2770 tcp_set_state(sk, TCP_CLOSE);
2773 inet_csk_listen_stop(sk);
2775 goto adjudge_to_death;
2778 /* We need to flush the recv. buffs. We do this only on the
2779 * descriptor close, not protocol-sourced closes, because the
2780 * reader process may not have drained the data yet!
2782 while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) {
2783 u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq;
2785 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
2787 data_was_unread += len;
2791 /* If socket has been already reset (e.g. in tcp_reset()) - kill it. */
2792 if (sk->sk_state == TCP_CLOSE)
2793 goto adjudge_to_death;
2795 /* As outlined in RFC 2525, section 2.17, we send a RST here because
2796 * data was lost. To witness the awful effects of the old behavior of
2797 * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk
2798 * GET in an FTP client, suspend the process, wait for the client to
2799 * advertise a zero window, then kill -9 the FTP client, wheee...
2800 * Note: timeout is always zero in such a case.
2802 if (unlikely(tcp_sk(sk)->repair)) {
2803 sk->sk_prot->disconnect(sk, 0);
2804 } else if (data_was_unread) {
2805 /* Unread data was tossed, zap the connection. */
2806 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONCLOSE);
2807 tcp_set_state(sk, TCP_CLOSE);
2808 tcp_send_active_reset(sk, sk->sk_allocation);
2809 } else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) {
2810 /* Check zero linger _after_ checking for unread data. */
2811 sk->sk_prot->disconnect(sk, 0);
2812 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
2813 } else if (tcp_close_state(sk)) {
2814 /* We FIN if the application ate all the data before
2815 * zapping the connection.
2818 /* RED-PEN. Formally speaking, we have broken TCP state
2819 * machine. State transitions:
2821 * TCP_ESTABLISHED -> TCP_FIN_WAIT1
2822 * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible)
2823 * TCP_CLOSE_WAIT -> TCP_LAST_ACK
2825 * are legal only when FIN has been sent (i.e. in window),
2826 * rather than queued out of window. Purists blame.
2828 * F.e. "RFC state" is ESTABLISHED,
2829 * if Linux state is FIN-WAIT-1, but FIN is still not sent.
2831 * The visible declinations are that sometimes
2832 * we enter time-wait state, when it is not required really
2833 * (harmless), do not send active resets, when they are
2834 * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
2835 * they look as CLOSING or LAST_ACK for Linux)
2836 * Probably, I missed some more holelets.
2838 * XXX (TFO) - To start off we don't support SYN+ACK+FIN
2839 * in a single packet! (May consider it later but will
2840 * probably need API support or TCP_CORK SYN-ACK until
2841 * data is written and socket is closed.)
2846 sk_stream_wait_close(sk, timeout);
2849 state = sk->sk_state;
2855 /* remove backlog if any, without releasing ownership. */
2858 this_cpu_inc(tcp_orphan_count);
2860 /* Have we already been destroyed by a softirq or backlog? */
2861 if (state != TCP_CLOSE && sk->sk_state == TCP_CLOSE)
2864 /* This is a (useful) BSD violating of the RFC. There is a
2865 * problem with TCP as specified in that the other end could
2866 * keep a socket open forever with no application left this end.
2867 * We use a 1 minute timeout (about the same as BSD) then kill
2868 * our end. If they send after that then tough - BUT: long enough
2869 * that we won't make the old 4*rto = almost no time - whoops
2872 * Nope, it was not mistake. It is really desired behaviour
2873 * f.e. on http servers, when such sockets are useless, but
2874 * consume significant resources. Let's do it with special
2875 * linger2 option. --ANK
2878 if (sk->sk_state == TCP_FIN_WAIT2) {
2879 struct tcp_sock *tp = tcp_sk(sk);
2880 if (READ_ONCE(tp->linger2) < 0) {
2881 tcp_set_state(sk, TCP_CLOSE);
2882 tcp_send_active_reset(sk, GFP_ATOMIC);
2883 __NET_INC_STATS(sock_net(sk),
2884 LINUX_MIB_TCPABORTONLINGER);
2886 const int tmo = tcp_fin_time(sk);
2888 if (tmo > TCP_TIMEWAIT_LEN) {
2889 inet_csk_reset_keepalive_timer(sk,
2890 tmo - TCP_TIMEWAIT_LEN);
2892 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
2897 if (sk->sk_state != TCP_CLOSE) {
2898 if (tcp_check_oom(sk, 0)) {
2899 tcp_set_state(sk, TCP_CLOSE);
2900 tcp_send_active_reset(sk, GFP_ATOMIC);
2901 __NET_INC_STATS(sock_net(sk),
2902 LINUX_MIB_TCPABORTONMEMORY);
2903 } else if (!check_net(sock_net(sk))) {
2904 /* Not possible to send reset; just close */
2905 tcp_set_state(sk, TCP_CLOSE);
2909 if (sk->sk_state == TCP_CLOSE) {
2910 struct request_sock *req;
2912 req = rcu_dereference_protected(tcp_sk(sk)->fastopen_rsk,
2913 lockdep_sock_is_held(sk));
2914 /* We could get here with a non-NULL req if the socket is
2915 * aborted (e.g., closed with unread data) before 3WHS
2919 reqsk_fastopen_remove(sk, req, false);
2920 inet_csk_destroy_sock(sk);
2922 /* Otherwise, socket is reprieved until protocol close. */
2929 void tcp_close(struct sock *sk, long timeout)
2932 __tcp_close(sk, timeout);
2936 EXPORT_SYMBOL(tcp_close);
2938 /* These states need RST on ABORT according to RFC793 */
2940 static inline bool tcp_need_reset(int state)
2942 return (1 << state) &
2943 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 |
2944 TCPF_FIN_WAIT2 | TCPF_SYN_RECV);
2947 static void tcp_rtx_queue_purge(struct sock *sk)
2949 struct rb_node *p = rb_first(&sk->tcp_rtx_queue);
2951 tcp_sk(sk)->highest_sack = NULL;
2953 struct sk_buff *skb = rb_to_skb(p);
2956 /* Since we are deleting whole queue, no need to
2957 * list_del(&skb->tcp_tsorted_anchor)
2959 tcp_rtx_queue_unlink(skb, sk);
2960 tcp_wmem_free_skb(sk, skb);
2964 void tcp_write_queue_purge(struct sock *sk)
2966 struct sk_buff *skb;
2968 tcp_chrono_stop(sk, TCP_CHRONO_BUSY);
2969 while ((skb = __skb_dequeue(&sk->sk_write_queue)) != NULL) {
2970 tcp_skb_tsorted_anchor_cleanup(skb);
2971 tcp_wmem_free_skb(sk, skb);
2973 tcp_rtx_queue_purge(sk);
2974 INIT_LIST_HEAD(&tcp_sk(sk)->tsorted_sent_queue);
2975 tcp_clear_all_retrans_hints(tcp_sk(sk));
2976 tcp_sk(sk)->packets_out = 0;
2977 inet_csk(sk)->icsk_backoff = 0;
2980 int tcp_disconnect(struct sock *sk, int flags)
2982 struct inet_sock *inet = inet_sk(sk);
2983 struct inet_connection_sock *icsk = inet_csk(sk);
2984 struct tcp_sock *tp = tcp_sk(sk);
2985 int old_state = sk->sk_state;
2988 if (old_state != TCP_CLOSE)
2989 tcp_set_state(sk, TCP_CLOSE);
2991 /* ABORT function of RFC793 */
2992 if (old_state == TCP_LISTEN) {
2993 inet_csk_listen_stop(sk);
2994 } else if (unlikely(tp->repair)) {
2995 WRITE_ONCE(sk->sk_err, ECONNABORTED);
2996 } else if (tcp_need_reset(old_state) ||
2997 (tp->snd_nxt != tp->write_seq &&
2998 (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK))) {
2999 /* The last check adjusts for discrepancy of Linux wrt. RFC
3002 tcp_send_active_reset(sk, gfp_any());
3003 WRITE_ONCE(sk->sk_err, ECONNRESET);
3004 } else if (old_state == TCP_SYN_SENT)
3005 WRITE_ONCE(sk->sk_err, ECONNRESET);
3007 tcp_clear_xmit_timers(sk);
3008 __skb_queue_purge(&sk->sk_receive_queue);
3009 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt);
3010 WRITE_ONCE(tp->urg_data, 0);
3011 tcp_write_queue_purge(sk);
3012 tcp_fastopen_active_disable_ofo_check(sk);
3013 skb_rbtree_purge(&tp->out_of_order_queue);
3015 inet->inet_dport = 0;
3017 inet_bhash2_reset_saddr(sk);
3019 WRITE_ONCE(sk->sk_shutdown, 0);
3020 sock_reset_flag(sk, SOCK_DONE);
3022 tp->mdev_us = jiffies_to_usecs(TCP_TIMEOUT_INIT);
3023 tp->rcv_rtt_last_tsecr = 0;
3025 seq = tp->write_seq + tp->max_window + 2;
3028 WRITE_ONCE(tp->write_seq, seq);
3030 icsk->icsk_backoff = 0;
3031 icsk->icsk_probes_out = 0;
3032 icsk->icsk_probes_tstamp = 0;
3033 icsk->icsk_rto = TCP_TIMEOUT_INIT;
3034 icsk->icsk_rto_min = TCP_RTO_MIN;
3035 icsk->icsk_delack_max = TCP_DELACK_MAX;
3036 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
3037 tcp_snd_cwnd_set(tp, TCP_INIT_CWND);
3038 tp->snd_cwnd_cnt = 0;
3039 tp->is_cwnd_limited = 0;
3040 tp->max_packets_out = 0;
3041 tp->window_clamp = 0;
3043 tp->delivered_ce = 0;
3044 if (icsk->icsk_ca_ops->release)
3045 icsk->icsk_ca_ops->release(sk);
3046 memset(icsk->icsk_ca_priv, 0, sizeof(icsk->icsk_ca_priv));
3047 icsk->icsk_ca_initialized = 0;
3048 tcp_set_ca_state(sk, TCP_CA_Open);
3049 tp->is_sack_reneg = 0;
3050 tcp_clear_retrans(tp);
3051 tp->total_retrans = 0;
3052 inet_csk_delack_init(sk);
3053 /* Initialize rcv_mss to TCP_MIN_MSS to avoid division by 0
3054 * issue in __tcp_select_window()
3056 icsk->icsk_ack.rcv_mss = TCP_MIN_MSS;
3057 memset(&tp->rx_opt, 0, sizeof(tp->rx_opt));
3059 dst_release(xchg((__force struct dst_entry **)&sk->sk_rx_dst, NULL));
3060 tcp_saved_syn_free(tp);
3061 tp->compressed_ack = 0;
3065 tp->bytes_acked = 0;
3066 tp->bytes_received = 0;
3067 tp->bytes_retrans = 0;
3068 tp->data_segs_in = 0;
3069 tp->data_segs_out = 0;
3070 tp->duplicate_sack[0].start_seq = 0;
3071 tp->duplicate_sack[0].end_seq = 0;
3074 tp->retrans_out = 0;
3076 tp->tlp_high_seq = 0;
3077 tp->last_oow_ack_time = 0;
3079 /* There's a bubble in the pipe until at least the first ACK. */
3080 tp->app_limited = ~0U;
3081 tp->rate_app_limited = 1;
3082 tp->rack.mstamp = 0;
3083 tp->rack.advanced = 0;
3084 tp->rack.reo_wnd_steps = 1;
3085 tp->rack.last_delivered = 0;
3086 tp->rack.reo_wnd_persist = 0;
3087 tp->rack.dsack_seen = 0;
3088 tp->syn_data_acked = 0;
3089 tp->rx_opt.saw_tstamp = 0;
3090 tp->rx_opt.dsack = 0;
3091 tp->rx_opt.num_sacks = 0;
3092 tp->rcv_ooopack = 0;
3095 /* Clean up fastopen related fields */
3096 tcp_free_fastopen_req(tp);
3097 inet_clear_bit(DEFER_CONNECT, sk);
3098 tp->fastopen_client_fail = 0;
3100 WARN_ON(inet->inet_num && !icsk->icsk_bind_hash);
3102 if (sk->sk_frag.page) {
3103 put_page(sk->sk_frag.page);
3104 sk->sk_frag.page = NULL;
3105 sk->sk_frag.offset = 0;
3107 sk_error_report(sk);
3110 EXPORT_SYMBOL(tcp_disconnect);
3112 static inline bool tcp_can_repair_sock(const struct sock *sk)
3114 return sockopt_ns_capable(sock_net(sk)->user_ns, CAP_NET_ADMIN) &&
3115 (sk->sk_state != TCP_LISTEN);
3118 static int tcp_repair_set_window(struct tcp_sock *tp, sockptr_t optbuf, int len)
3120 struct tcp_repair_window opt;
3125 if (len != sizeof(opt))
3128 if (copy_from_sockptr(&opt, optbuf, sizeof(opt)))
3131 if (opt.max_window < opt.snd_wnd)
3134 if (after(opt.snd_wl1, tp->rcv_nxt + opt.rcv_wnd))
3137 if (after(opt.rcv_wup, tp->rcv_nxt))
3140 tp->snd_wl1 = opt.snd_wl1;
3141 tp->snd_wnd = opt.snd_wnd;
3142 tp->max_window = opt.max_window;
3144 tp->rcv_wnd = opt.rcv_wnd;
3145 tp->rcv_wup = opt.rcv_wup;
3150 static int tcp_repair_options_est(struct sock *sk, sockptr_t optbuf,
3153 struct tcp_sock *tp = tcp_sk(sk);
3154 struct tcp_repair_opt opt;
3157 while (len >= sizeof(opt)) {
3158 if (copy_from_sockptr_offset(&opt, optbuf, offset, sizeof(opt)))
3161 offset += sizeof(opt);
3164 switch (opt.opt_code) {
3166 tp->rx_opt.mss_clamp = opt.opt_val;
3171 u16 snd_wscale = opt.opt_val & 0xFFFF;
3172 u16 rcv_wscale = opt.opt_val >> 16;
3174 if (snd_wscale > TCP_MAX_WSCALE || rcv_wscale > TCP_MAX_WSCALE)
3177 tp->rx_opt.snd_wscale = snd_wscale;
3178 tp->rx_opt.rcv_wscale = rcv_wscale;
3179 tp->rx_opt.wscale_ok = 1;
3182 case TCPOPT_SACK_PERM:
3183 if (opt.opt_val != 0)
3186 tp->rx_opt.sack_ok |= TCP_SACK_SEEN;
3188 case TCPOPT_TIMESTAMP:
3189 if (opt.opt_val != 0)
3192 tp->rx_opt.tstamp_ok = 1;
3200 DEFINE_STATIC_KEY_FALSE(tcp_tx_delay_enabled);
3201 EXPORT_SYMBOL(tcp_tx_delay_enabled);
3203 static void tcp_enable_tx_delay(void)
3205 if (!static_branch_unlikely(&tcp_tx_delay_enabled)) {
3206 static int __tcp_tx_delay_enabled = 0;
3208 if (cmpxchg(&__tcp_tx_delay_enabled, 0, 1) == 0) {
3209 static_branch_enable(&tcp_tx_delay_enabled);
3210 pr_info("TCP_TX_DELAY enabled\n");
3215 /* When set indicates to always queue non-full frames. Later the user clears
3216 * this option and we transmit any pending partial frames in the queue. This is
3217 * meant to be used alongside sendfile() to get properly filled frames when the
3218 * user (for example) must write out headers with a write() call first and then
3219 * use sendfile to send out the data parts.
3221 * TCP_CORK can be set together with TCP_NODELAY and it is stronger than
3224 void __tcp_sock_set_cork(struct sock *sk, bool on)
3226 struct tcp_sock *tp = tcp_sk(sk);
3229 tp->nonagle |= TCP_NAGLE_CORK;
3231 tp->nonagle &= ~TCP_NAGLE_CORK;
3232 if (tp->nonagle & TCP_NAGLE_OFF)
3233 tp->nonagle |= TCP_NAGLE_PUSH;
3234 tcp_push_pending_frames(sk);
3238 void tcp_sock_set_cork(struct sock *sk, bool on)
3241 __tcp_sock_set_cork(sk, on);
3244 EXPORT_SYMBOL(tcp_sock_set_cork);
3246 /* TCP_NODELAY is weaker than TCP_CORK, so that this option on corked socket is
3247 * remembered, but it is not activated until cork is cleared.
3249 * However, when TCP_NODELAY is set we make an explicit push, which overrides
3250 * even TCP_CORK for currently queued segments.
3252 void __tcp_sock_set_nodelay(struct sock *sk, bool on)
3255 tcp_sk(sk)->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH;
3256 tcp_push_pending_frames(sk);
3258 tcp_sk(sk)->nonagle &= ~TCP_NAGLE_OFF;
3262 void tcp_sock_set_nodelay(struct sock *sk)
3265 __tcp_sock_set_nodelay(sk, true);
3268 EXPORT_SYMBOL(tcp_sock_set_nodelay);
3270 static void __tcp_sock_set_quickack(struct sock *sk, int val)
3273 inet_csk_enter_pingpong_mode(sk);
3277 inet_csk_exit_pingpong_mode(sk);
3278 if ((1 << sk->sk_state) & (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) &&
3279 inet_csk_ack_scheduled(sk)) {
3280 inet_csk(sk)->icsk_ack.pending |= ICSK_ACK_PUSHED;
3281 tcp_cleanup_rbuf(sk, 1);
3283 inet_csk_enter_pingpong_mode(sk);
3287 void tcp_sock_set_quickack(struct sock *sk, int val)
3290 __tcp_sock_set_quickack(sk, val);
3293 EXPORT_SYMBOL(tcp_sock_set_quickack);
3295 int tcp_sock_set_syncnt(struct sock *sk, int val)
3297 if (val < 1 || val > MAX_TCP_SYNCNT)
3300 WRITE_ONCE(inet_csk(sk)->icsk_syn_retries, val);
3303 EXPORT_SYMBOL(tcp_sock_set_syncnt);
3305 int tcp_sock_set_user_timeout(struct sock *sk, int val)
3307 /* Cap the max time in ms TCP will retry or probe the window
3308 * before giving up and aborting (ETIMEDOUT) a connection.
3313 WRITE_ONCE(inet_csk(sk)->icsk_user_timeout, val);
3316 EXPORT_SYMBOL(tcp_sock_set_user_timeout);
3318 int tcp_sock_set_keepidle_locked(struct sock *sk, int val)
3320 struct tcp_sock *tp = tcp_sk(sk);
3322 if (val < 1 || val > MAX_TCP_KEEPIDLE)
3325 /* Paired with WRITE_ONCE() in keepalive_time_when() */
3326 WRITE_ONCE(tp->keepalive_time, val * HZ);
3327 if (sock_flag(sk, SOCK_KEEPOPEN) &&
3328 !((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN))) {
3329 u32 elapsed = keepalive_time_elapsed(tp);
3331 if (tp->keepalive_time > elapsed)
3332 elapsed = tp->keepalive_time - elapsed;
3335 inet_csk_reset_keepalive_timer(sk, elapsed);
3341 int tcp_sock_set_keepidle(struct sock *sk, int val)
3346 err = tcp_sock_set_keepidle_locked(sk, val);
3350 EXPORT_SYMBOL(tcp_sock_set_keepidle);
3352 int tcp_sock_set_keepintvl(struct sock *sk, int val)
3354 if (val < 1 || val > MAX_TCP_KEEPINTVL)
3357 WRITE_ONCE(tcp_sk(sk)->keepalive_intvl, val * HZ);
3360 EXPORT_SYMBOL(tcp_sock_set_keepintvl);
3362 int tcp_sock_set_keepcnt(struct sock *sk, int val)
3364 if (val < 1 || val > MAX_TCP_KEEPCNT)
3367 /* Paired with READ_ONCE() in keepalive_probes() */
3368 WRITE_ONCE(tcp_sk(sk)->keepalive_probes, val);
3371 EXPORT_SYMBOL(tcp_sock_set_keepcnt);
3373 int tcp_set_window_clamp(struct sock *sk, int val)
3375 struct tcp_sock *tp = tcp_sk(sk);
3378 if (sk->sk_state != TCP_CLOSE)
3380 tp->window_clamp = 0;
3382 u32 new_rcv_ssthresh, old_window_clamp = tp->window_clamp;
3383 u32 new_window_clamp = val < SOCK_MIN_RCVBUF / 2 ?
3384 SOCK_MIN_RCVBUF / 2 : val;
3386 if (new_window_clamp == old_window_clamp)
3389 tp->window_clamp = new_window_clamp;
3390 if (new_window_clamp < old_window_clamp) {
3391 /* need to apply the reserved mem provisioning only
3392 * when shrinking the window clamp
3394 __tcp_adjust_rcv_ssthresh(sk, tp->window_clamp);
3397 new_rcv_ssthresh = min(tp->rcv_wnd, tp->window_clamp);
3398 tp->rcv_ssthresh = max(new_rcv_ssthresh,
3406 * Socket option code for TCP.
3408 int do_tcp_setsockopt(struct sock *sk, int level, int optname,
3409 sockptr_t optval, unsigned int optlen)
3411 struct tcp_sock *tp = tcp_sk(sk);
3412 struct inet_connection_sock *icsk = inet_csk(sk);
3413 struct net *net = sock_net(sk);
3417 /* These are data/string values, all the others are ints */
3419 case TCP_CONGESTION: {
3420 char name[TCP_CA_NAME_MAX];
3425 val = strncpy_from_sockptr(name, optval,
3426 min_t(long, TCP_CA_NAME_MAX-1, optlen));
3431 sockopt_lock_sock(sk);
3432 err = tcp_set_congestion_control(sk, name, !has_current_bpf_ctx(),
3433 sockopt_ns_capable(sock_net(sk)->user_ns,
3435 sockopt_release_sock(sk);
3439 char name[TCP_ULP_NAME_MAX];
3444 val = strncpy_from_sockptr(name, optval,
3445 min_t(long, TCP_ULP_NAME_MAX - 1,
3451 sockopt_lock_sock(sk);
3452 err = tcp_set_ulp(sk, name);
3453 sockopt_release_sock(sk);
3456 case TCP_FASTOPEN_KEY: {
3457 __u8 key[TCP_FASTOPEN_KEY_BUF_LENGTH];
3458 __u8 *backup_key = NULL;
3460 /* Allow a backup key as well to facilitate key rotation
3461 * First key is the active one.
3463 if (optlen != TCP_FASTOPEN_KEY_LENGTH &&
3464 optlen != TCP_FASTOPEN_KEY_BUF_LENGTH)
3467 if (copy_from_sockptr(key, optval, optlen))
3470 if (optlen == TCP_FASTOPEN_KEY_BUF_LENGTH)
3471 backup_key = key + TCP_FASTOPEN_KEY_LENGTH;
3473 return tcp_fastopen_reset_cipher(net, sk, key, backup_key);
3480 if (optlen < sizeof(int))
3483 if (copy_from_sockptr(&val, optval, sizeof(val)))
3486 /* Handle options that can be set without locking the socket. */
3489 return tcp_sock_set_syncnt(sk, val);
3490 case TCP_USER_TIMEOUT:
3491 return tcp_sock_set_user_timeout(sk, val);
3493 return tcp_sock_set_keepintvl(sk, val);
3495 return tcp_sock_set_keepcnt(sk, val);
3498 WRITE_ONCE(tp->linger2, -1);
3499 else if (val > TCP_FIN_TIMEOUT_MAX / HZ)
3500 WRITE_ONCE(tp->linger2, TCP_FIN_TIMEOUT_MAX);
3502 WRITE_ONCE(tp->linger2, val * HZ);
3504 case TCP_DEFER_ACCEPT:
3505 /* Translate value in seconds to number of retransmits */
3506 WRITE_ONCE(icsk->icsk_accept_queue.rskq_defer_accept,
3507 secs_to_retrans(val, TCP_TIMEOUT_INIT / HZ,
3512 sockopt_lock_sock(sk);
3516 /* Values greater than interface MTU won't take effect. However
3517 * at the point when this call is done we typically don't yet
3518 * know which interface is going to be used
3520 if (val && (val < TCP_MIN_MSS || val > MAX_TCP_WINDOW)) {
3524 tp->rx_opt.user_mss = val;
3528 __tcp_sock_set_nodelay(sk, val);
3531 case TCP_THIN_LINEAR_TIMEOUTS:
3532 if (val < 0 || val > 1)
3538 case TCP_THIN_DUPACK:
3539 if (val < 0 || val > 1)
3544 if (!tcp_can_repair_sock(sk))
3546 else if (val == TCP_REPAIR_ON) {
3548 sk->sk_reuse = SK_FORCE_REUSE;
3549 tp->repair_queue = TCP_NO_QUEUE;
3550 } else if (val == TCP_REPAIR_OFF) {
3552 sk->sk_reuse = SK_NO_REUSE;
3553 tcp_send_window_probe(sk);
3554 } else if (val == TCP_REPAIR_OFF_NO_WP) {
3556 sk->sk_reuse = SK_NO_REUSE;
3562 case TCP_REPAIR_QUEUE:
3565 else if ((unsigned int)val < TCP_QUEUES_NR)
3566 tp->repair_queue = val;
3572 if (sk->sk_state != TCP_CLOSE) {
3574 } else if (tp->repair_queue == TCP_SEND_QUEUE) {
3575 if (!tcp_rtx_queue_empty(sk))
3578 WRITE_ONCE(tp->write_seq, val);
3579 } else if (tp->repair_queue == TCP_RECV_QUEUE) {
3580 if (tp->rcv_nxt != tp->copied_seq) {
3583 WRITE_ONCE(tp->rcv_nxt, val);
3584 WRITE_ONCE(tp->copied_seq, val);
3591 case TCP_REPAIR_OPTIONS:
3594 else if (sk->sk_state == TCP_ESTABLISHED && !tp->bytes_sent)
3595 err = tcp_repair_options_est(sk, optval, optlen);
3601 __tcp_sock_set_cork(sk, val);
3605 err = tcp_sock_set_keepidle_locked(sk, val);
3608 /* 0: disable, 1: enable, 2: start from ether_header */
3609 if (val < 0 || val > 2)
3615 case TCP_WINDOW_CLAMP:
3616 err = tcp_set_window_clamp(sk, val);
3620 __tcp_sock_set_quickack(sk, val);
3624 if (!tcp_can_repair_sock(sk)) {
3628 err = tcp_ao_set_repair(sk, optval, optlen);
3630 #ifdef CONFIG_TCP_AO
3631 case TCP_AO_ADD_KEY:
3632 case TCP_AO_DEL_KEY:
3634 /* If this is the first TCP-AO setsockopt() on the socket,
3635 * sk_state has to be LISTEN or CLOSE. Allow TCP_REPAIR
3638 if ((1 << sk->sk_state) & (TCPF_LISTEN | TCPF_CLOSE))
3640 if (rcu_dereference_protected(tcp_sk(sk)->ao_info,
3641 lockdep_sock_is_held(sk)))
3648 err = tp->af_specific->ao_parse(sk, optname, optval, optlen);
3652 #ifdef CONFIG_TCP_MD5SIG
3654 case TCP_MD5SIG_EXT:
3655 err = tp->af_specific->md5_parse(sk, optname, optval, optlen);
3659 if (val >= 0 && ((1 << sk->sk_state) & (TCPF_CLOSE |
3661 tcp_fastopen_init_key_once(net);
3663 fastopen_queue_tune(sk, val);
3668 case TCP_FASTOPEN_CONNECT:
3669 if (val > 1 || val < 0) {
3671 } else if (READ_ONCE(net->ipv4.sysctl_tcp_fastopen) &
3672 TFO_CLIENT_ENABLE) {
3673 if (sk->sk_state == TCP_CLOSE)
3674 tp->fastopen_connect = val;
3681 case TCP_FASTOPEN_NO_COOKIE:
3682 if (val > 1 || val < 0)
3684 else if (!((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
3687 tp->fastopen_no_cookie = val;
3694 /* val is an opaque field,
3695 * and low order bit contains usec_ts enable bit.
3696 * Its a best effort, and we do not care if user makes an error.
3698 tp->tcp_usec_ts = val & 1;
3699 WRITE_ONCE(tp->tsoffset, val - tcp_clock_ts(tp->tcp_usec_ts));
3701 case TCP_REPAIR_WINDOW:
3702 err = tcp_repair_set_window(tp, optval, optlen);
3704 case TCP_NOTSENT_LOWAT:
3705 WRITE_ONCE(tp->notsent_lowat, val);
3706 sk->sk_write_space(sk);
3709 if (val > 1 || val < 0)
3712 tp->recvmsg_inq = val;
3716 tcp_enable_tx_delay();
3717 WRITE_ONCE(tp->tcp_tx_delay, val);
3724 sockopt_release_sock(sk);
3728 int tcp_setsockopt(struct sock *sk, int level, int optname, sockptr_t optval,
3729 unsigned int optlen)
3731 const struct inet_connection_sock *icsk = inet_csk(sk);
3733 if (level != SOL_TCP)
3734 /* Paired with WRITE_ONCE() in do_ipv6_setsockopt() and tcp_v6_connect() */
3735 return READ_ONCE(icsk->icsk_af_ops)->setsockopt(sk, level, optname,
3737 return do_tcp_setsockopt(sk, level, optname, optval, optlen);
3739 EXPORT_SYMBOL(tcp_setsockopt);
3741 static void tcp_get_info_chrono_stats(const struct tcp_sock *tp,
3742 struct tcp_info *info)
3744 u64 stats[__TCP_CHRONO_MAX], total = 0;
3747 for (i = TCP_CHRONO_BUSY; i < __TCP_CHRONO_MAX; ++i) {
3748 stats[i] = tp->chrono_stat[i - 1];
3749 if (i == tp->chrono_type)
3750 stats[i] += tcp_jiffies32 - tp->chrono_start;
3751 stats[i] *= USEC_PER_SEC / HZ;
3755 info->tcpi_busy_time = total;
3756 info->tcpi_rwnd_limited = stats[TCP_CHRONO_RWND_LIMITED];
3757 info->tcpi_sndbuf_limited = stats[TCP_CHRONO_SNDBUF_LIMITED];
3760 /* Return information about state of tcp endpoint in API format. */
3761 void tcp_get_info(struct sock *sk, struct tcp_info *info)
3763 const struct tcp_sock *tp = tcp_sk(sk); /* iff sk_type == SOCK_STREAM */
3764 const struct inet_connection_sock *icsk = inet_csk(sk);
3770 memset(info, 0, sizeof(*info));
3771 if (sk->sk_type != SOCK_STREAM)
3774 info->tcpi_state = inet_sk_state_load(sk);
3776 /* Report meaningful fields for all TCP states, including listeners */
3777 rate = READ_ONCE(sk->sk_pacing_rate);
3778 rate64 = (rate != ~0UL) ? rate : ~0ULL;
3779 info->tcpi_pacing_rate = rate64;
3781 rate = READ_ONCE(sk->sk_max_pacing_rate);
3782 rate64 = (rate != ~0UL) ? rate : ~0ULL;
3783 info->tcpi_max_pacing_rate = rate64;
3785 info->tcpi_reordering = tp->reordering;
3786 info->tcpi_snd_cwnd = tcp_snd_cwnd(tp);
3788 if (info->tcpi_state == TCP_LISTEN) {
3789 /* listeners aliased fields :
3790 * tcpi_unacked -> Number of children ready for accept()
3791 * tcpi_sacked -> max backlog
3793 info->tcpi_unacked = READ_ONCE(sk->sk_ack_backlog);
3794 info->tcpi_sacked = READ_ONCE(sk->sk_max_ack_backlog);
3798 slow = lock_sock_fast(sk);
3800 info->tcpi_ca_state = icsk->icsk_ca_state;
3801 info->tcpi_retransmits = icsk->icsk_retransmits;
3802 info->tcpi_probes = icsk->icsk_probes_out;
3803 info->tcpi_backoff = icsk->icsk_backoff;
3805 if (tp->rx_opt.tstamp_ok)
3806 info->tcpi_options |= TCPI_OPT_TIMESTAMPS;
3807 if (tcp_is_sack(tp))
3808 info->tcpi_options |= TCPI_OPT_SACK;
3809 if (tp->rx_opt.wscale_ok) {
3810 info->tcpi_options |= TCPI_OPT_WSCALE;
3811 info->tcpi_snd_wscale = tp->rx_opt.snd_wscale;
3812 info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale;
3815 if (tp->ecn_flags & TCP_ECN_OK)
3816 info->tcpi_options |= TCPI_OPT_ECN;
3817 if (tp->ecn_flags & TCP_ECN_SEEN)
3818 info->tcpi_options |= TCPI_OPT_ECN_SEEN;
3819 if (tp->syn_data_acked)
3820 info->tcpi_options |= TCPI_OPT_SYN_DATA;
3821 if (tp->tcp_usec_ts)
3822 info->tcpi_options |= TCPI_OPT_USEC_TS;
3824 info->tcpi_rto = jiffies_to_usecs(icsk->icsk_rto);
3825 info->tcpi_ato = jiffies_to_usecs(min_t(u32, icsk->icsk_ack.ato,
3826 tcp_delack_max(sk)));
3827 info->tcpi_snd_mss = tp->mss_cache;
3828 info->tcpi_rcv_mss = icsk->icsk_ack.rcv_mss;
3830 info->tcpi_unacked = tp->packets_out;
3831 info->tcpi_sacked = tp->sacked_out;
3833 info->tcpi_lost = tp->lost_out;
3834 info->tcpi_retrans = tp->retrans_out;
3836 now = tcp_jiffies32;
3837 info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime);
3838 info->tcpi_last_data_recv = jiffies_to_msecs(now - icsk->icsk_ack.lrcvtime);
3839 info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp);
3841 info->tcpi_pmtu = icsk->icsk_pmtu_cookie;
3842 info->tcpi_rcv_ssthresh = tp->rcv_ssthresh;
3843 info->tcpi_rtt = tp->srtt_us >> 3;
3844 info->tcpi_rttvar = tp->mdev_us >> 2;
3845 info->tcpi_snd_ssthresh = tp->snd_ssthresh;
3846 info->tcpi_advmss = tp->advmss;
3848 info->tcpi_rcv_rtt = tp->rcv_rtt_est.rtt_us >> 3;
3849 info->tcpi_rcv_space = tp->rcvq_space.space;
3851 info->tcpi_total_retrans = tp->total_retrans;
3853 info->tcpi_bytes_acked = tp->bytes_acked;
3854 info->tcpi_bytes_received = tp->bytes_received;
3855 info->tcpi_notsent_bytes = max_t(int, 0, tp->write_seq - tp->snd_nxt);
3856 tcp_get_info_chrono_stats(tp, info);
3858 info->tcpi_segs_out = tp->segs_out;
3860 /* segs_in and data_segs_in can be updated from tcp_segs_in() from BH */
3861 info->tcpi_segs_in = READ_ONCE(tp->segs_in);
3862 info->tcpi_data_segs_in = READ_ONCE(tp->data_segs_in);
3864 info->tcpi_min_rtt = tcp_min_rtt(tp);
3865 info->tcpi_data_segs_out = tp->data_segs_out;
3867 info->tcpi_delivery_rate_app_limited = tp->rate_app_limited ? 1 : 0;
3868 rate64 = tcp_compute_delivery_rate(tp);
3870 info->tcpi_delivery_rate = rate64;
3871 info->tcpi_delivered = tp->delivered;
3872 info->tcpi_delivered_ce = tp->delivered_ce;
3873 info->tcpi_bytes_sent = tp->bytes_sent;
3874 info->tcpi_bytes_retrans = tp->bytes_retrans;
3875 info->tcpi_dsack_dups = tp->dsack_dups;
3876 info->tcpi_reord_seen = tp->reord_seen;
3877 info->tcpi_rcv_ooopack = tp->rcv_ooopack;
3878 info->tcpi_snd_wnd = tp->snd_wnd;
3879 info->tcpi_rcv_wnd = tp->rcv_wnd;
3880 info->tcpi_rehash = tp->plb_rehash + tp->timeout_rehash;
3881 info->tcpi_fastopen_client_fail = tp->fastopen_client_fail;
3883 info->tcpi_total_rto = tp->total_rto;
3884 info->tcpi_total_rto_recoveries = tp->total_rto_recoveries;
3885 info->tcpi_total_rto_time = tp->total_rto_time;
3887 info->tcpi_total_rto_time += tcp_clock_ms() - tp->rto_stamp;
3889 unlock_sock_fast(sk, slow);
3891 EXPORT_SYMBOL_GPL(tcp_get_info);
3893 static size_t tcp_opt_stats_get_size(void)
3896 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_BUSY */
3897 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_RWND_LIMITED */
3898 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_SNDBUF_LIMITED */
3899 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_DATA_SEGS_OUT */
3900 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_TOTAL_RETRANS */
3901 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_PACING_RATE */
3902 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_DELIVERY_RATE */
3903 nla_total_size(sizeof(u32)) + /* TCP_NLA_SND_CWND */
3904 nla_total_size(sizeof(u32)) + /* TCP_NLA_REORDERING */
3905 nla_total_size(sizeof(u32)) + /* TCP_NLA_MIN_RTT */
3906 nla_total_size(sizeof(u8)) + /* TCP_NLA_RECUR_RETRANS */
3907 nla_total_size(sizeof(u8)) + /* TCP_NLA_DELIVERY_RATE_APP_LMT */
3908 nla_total_size(sizeof(u32)) + /* TCP_NLA_SNDQ_SIZE */
3909 nla_total_size(sizeof(u8)) + /* TCP_NLA_CA_STATE */
3910 nla_total_size(sizeof(u32)) + /* TCP_NLA_SND_SSTHRESH */
3911 nla_total_size(sizeof(u32)) + /* TCP_NLA_DELIVERED */
3912 nla_total_size(sizeof(u32)) + /* TCP_NLA_DELIVERED_CE */
3913 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_BYTES_SENT */
3914 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_BYTES_RETRANS */
3915 nla_total_size(sizeof(u32)) + /* TCP_NLA_DSACK_DUPS */
3916 nla_total_size(sizeof(u32)) + /* TCP_NLA_REORD_SEEN */
3917 nla_total_size(sizeof(u32)) + /* TCP_NLA_SRTT */
3918 nla_total_size(sizeof(u16)) + /* TCP_NLA_TIMEOUT_REHASH */
3919 nla_total_size(sizeof(u32)) + /* TCP_NLA_BYTES_NOTSENT */
3920 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_EDT */
3921 nla_total_size(sizeof(u8)) + /* TCP_NLA_TTL */
3922 nla_total_size(sizeof(u32)) + /* TCP_NLA_REHASH */
3926 /* Returns TTL or hop limit of an incoming packet from skb. */
3927 static u8 tcp_skb_ttl_or_hop_limit(const struct sk_buff *skb)
3929 if (skb->protocol == htons(ETH_P_IP))
3930 return ip_hdr(skb)->ttl;
3931 else if (skb->protocol == htons(ETH_P_IPV6))
3932 return ipv6_hdr(skb)->hop_limit;
3937 struct sk_buff *tcp_get_timestamping_opt_stats(const struct sock *sk,
3938 const struct sk_buff *orig_skb,
3939 const struct sk_buff *ack_skb)
3941 const struct tcp_sock *tp = tcp_sk(sk);
3942 struct sk_buff *stats;
3943 struct tcp_info info;
3947 stats = alloc_skb(tcp_opt_stats_get_size(), GFP_ATOMIC);
3951 tcp_get_info_chrono_stats(tp, &info);
3952 nla_put_u64_64bit(stats, TCP_NLA_BUSY,
3953 info.tcpi_busy_time, TCP_NLA_PAD);
3954 nla_put_u64_64bit(stats, TCP_NLA_RWND_LIMITED,
3955 info.tcpi_rwnd_limited, TCP_NLA_PAD);
3956 nla_put_u64_64bit(stats, TCP_NLA_SNDBUF_LIMITED,
3957 info.tcpi_sndbuf_limited, TCP_NLA_PAD);
3958 nla_put_u64_64bit(stats, TCP_NLA_DATA_SEGS_OUT,
3959 tp->data_segs_out, TCP_NLA_PAD);
3960 nla_put_u64_64bit(stats, TCP_NLA_TOTAL_RETRANS,
3961 tp->total_retrans, TCP_NLA_PAD);
3963 rate = READ_ONCE(sk->sk_pacing_rate);
3964 rate64 = (rate != ~0UL) ? rate : ~0ULL;
3965 nla_put_u64_64bit(stats, TCP_NLA_PACING_RATE, rate64, TCP_NLA_PAD);
3967 rate64 = tcp_compute_delivery_rate(tp);
3968 nla_put_u64_64bit(stats, TCP_NLA_DELIVERY_RATE, rate64, TCP_NLA_PAD);
3970 nla_put_u32(stats, TCP_NLA_SND_CWND, tcp_snd_cwnd(tp));
3971 nla_put_u32(stats, TCP_NLA_REORDERING, tp->reordering);
3972 nla_put_u32(stats, TCP_NLA_MIN_RTT, tcp_min_rtt(tp));
3974 nla_put_u8(stats, TCP_NLA_RECUR_RETRANS, inet_csk(sk)->icsk_retransmits);
3975 nla_put_u8(stats, TCP_NLA_DELIVERY_RATE_APP_LMT, !!tp->rate_app_limited);
3976 nla_put_u32(stats, TCP_NLA_SND_SSTHRESH, tp->snd_ssthresh);
3977 nla_put_u32(stats, TCP_NLA_DELIVERED, tp->delivered);
3978 nla_put_u32(stats, TCP_NLA_DELIVERED_CE, tp->delivered_ce);
3980 nla_put_u32(stats, TCP_NLA_SNDQ_SIZE, tp->write_seq - tp->snd_una);
3981 nla_put_u8(stats, TCP_NLA_CA_STATE, inet_csk(sk)->icsk_ca_state);
3983 nla_put_u64_64bit(stats, TCP_NLA_BYTES_SENT, tp->bytes_sent,
3985 nla_put_u64_64bit(stats, TCP_NLA_BYTES_RETRANS, tp->bytes_retrans,
3987 nla_put_u32(stats, TCP_NLA_DSACK_DUPS, tp->dsack_dups);
3988 nla_put_u32(stats, TCP_NLA_REORD_SEEN, tp->reord_seen);
3989 nla_put_u32(stats, TCP_NLA_SRTT, tp->srtt_us >> 3);
3990 nla_put_u16(stats, TCP_NLA_TIMEOUT_REHASH, tp->timeout_rehash);
3991 nla_put_u32(stats, TCP_NLA_BYTES_NOTSENT,
3992 max_t(int, 0, tp->write_seq - tp->snd_nxt));
3993 nla_put_u64_64bit(stats, TCP_NLA_EDT, orig_skb->skb_mstamp_ns,
3996 nla_put_u8(stats, TCP_NLA_TTL,
3997 tcp_skb_ttl_or_hop_limit(ack_skb));
3999 nla_put_u32(stats, TCP_NLA_REHASH, tp->plb_rehash + tp->timeout_rehash);
4003 int do_tcp_getsockopt(struct sock *sk, int level,
4004 int optname, sockptr_t optval, sockptr_t optlen)
4006 struct inet_connection_sock *icsk = inet_csk(sk);
4007 struct tcp_sock *tp = tcp_sk(sk);
4008 struct net *net = sock_net(sk);
4011 if (copy_from_sockptr(&len, optlen, sizeof(int)))
4017 len = min_t(unsigned int, len, sizeof(int));
4021 val = tp->mss_cache;
4022 if (tp->rx_opt.user_mss &&
4023 ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
4024 val = tp->rx_opt.user_mss;
4026 val = tp->rx_opt.mss_clamp;
4029 val = !!(tp->nonagle&TCP_NAGLE_OFF);
4032 val = !!(tp->nonagle&TCP_NAGLE_CORK);
4035 val = keepalive_time_when(tp) / HZ;
4038 val = keepalive_intvl_when(tp) / HZ;
4041 val = keepalive_probes(tp);
4044 val = READ_ONCE(icsk->icsk_syn_retries) ? :
4045 READ_ONCE(net->ipv4.sysctl_tcp_syn_retries);
4048 val = READ_ONCE(tp->linger2);
4050 val = (val ? : READ_ONCE(net->ipv4.sysctl_tcp_fin_timeout)) / HZ;
4052 case TCP_DEFER_ACCEPT:
4053 val = READ_ONCE(icsk->icsk_accept_queue.rskq_defer_accept);
4054 val = retrans_to_secs(val, TCP_TIMEOUT_INIT / HZ,
4057 case TCP_WINDOW_CLAMP:
4058 val = tp->window_clamp;
4061 struct tcp_info info;
4063 if (copy_from_sockptr(&len, optlen, sizeof(int)))
4066 tcp_get_info(sk, &info);
4068 len = min_t(unsigned int, len, sizeof(info));
4069 if (copy_to_sockptr(optlen, &len, sizeof(int)))
4071 if (copy_to_sockptr(optval, &info, len))
4076 const struct tcp_congestion_ops *ca_ops;
4077 union tcp_cc_info info;
4081 if (copy_from_sockptr(&len, optlen, sizeof(int)))
4084 ca_ops = icsk->icsk_ca_ops;
4085 if (ca_ops && ca_ops->get_info)
4086 sz = ca_ops->get_info(sk, ~0U, &attr, &info);
4088 len = min_t(unsigned int, len, sz);
4089 if (copy_to_sockptr(optlen, &len, sizeof(int)))
4091 if (copy_to_sockptr(optval, &info, len))
4096 val = !inet_csk_in_pingpong_mode(sk);
4099 case TCP_CONGESTION:
4100 if (copy_from_sockptr(&len, optlen, sizeof(int)))
4102 len = min_t(unsigned int, len, TCP_CA_NAME_MAX);
4103 if (copy_to_sockptr(optlen, &len, sizeof(int)))
4105 if (copy_to_sockptr(optval, icsk->icsk_ca_ops->name, len))
4110 if (copy_from_sockptr(&len, optlen, sizeof(int)))
4112 len = min_t(unsigned int, len, TCP_ULP_NAME_MAX);
4113 if (!icsk->icsk_ulp_ops) {
4115 if (copy_to_sockptr(optlen, &len, sizeof(int)))
4119 if (copy_to_sockptr(optlen, &len, sizeof(int)))
4121 if (copy_to_sockptr(optval, icsk->icsk_ulp_ops->name, len))
4125 case TCP_FASTOPEN_KEY: {
4126 u64 key[TCP_FASTOPEN_KEY_BUF_LENGTH / sizeof(u64)];
4127 unsigned int key_len;
4129 if (copy_from_sockptr(&len, optlen, sizeof(int)))
4132 key_len = tcp_fastopen_get_cipher(net, icsk, key) *
4133 TCP_FASTOPEN_KEY_LENGTH;
4134 len = min_t(unsigned int, len, key_len);
4135 if (copy_to_sockptr(optlen, &len, sizeof(int)))
4137 if (copy_to_sockptr(optval, key, len))
4141 case TCP_THIN_LINEAR_TIMEOUTS:
4145 case TCP_THIN_DUPACK:
4153 case TCP_REPAIR_QUEUE:
4155 val = tp->repair_queue;
4160 case TCP_REPAIR_WINDOW: {
4161 struct tcp_repair_window opt;
4163 if (copy_from_sockptr(&len, optlen, sizeof(int)))
4166 if (len != sizeof(opt))
4172 opt.snd_wl1 = tp->snd_wl1;
4173 opt.snd_wnd = tp->snd_wnd;
4174 opt.max_window = tp->max_window;
4175 opt.rcv_wnd = tp->rcv_wnd;
4176 opt.rcv_wup = tp->rcv_wup;
4178 if (copy_to_sockptr(optval, &opt, len))
4183 if (tp->repair_queue == TCP_SEND_QUEUE)
4184 val = tp->write_seq;
4185 else if (tp->repair_queue == TCP_RECV_QUEUE)
4191 case TCP_USER_TIMEOUT:
4192 val = READ_ONCE(icsk->icsk_user_timeout);
4196 val = READ_ONCE(icsk->icsk_accept_queue.fastopenq.max_qlen);
4199 case TCP_FASTOPEN_CONNECT:
4200 val = tp->fastopen_connect;
4203 case TCP_FASTOPEN_NO_COOKIE:
4204 val = tp->fastopen_no_cookie;
4208 val = READ_ONCE(tp->tcp_tx_delay);
4212 val = tcp_clock_ts(tp->tcp_usec_ts) + READ_ONCE(tp->tsoffset);
4213 if (tp->tcp_usec_ts)
4218 case TCP_NOTSENT_LOWAT:
4219 val = READ_ONCE(tp->notsent_lowat);
4222 val = tp->recvmsg_inq;
4227 case TCP_SAVED_SYN: {
4228 if (copy_from_sockptr(&len, optlen, sizeof(int)))
4231 sockopt_lock_sock(sk);
4232 if (tp->saved_syn) {
4233 if (len < tcp_saved_syn_len(tp->saved_syn)) {
4234 len = tcp_saved_syn_len(tp->saved_syn);
4235 if (copy_to_sockptr(optlen, &len, sizeof(int))) {
4236 sockopt_release_sock(sk);
4239 sockopt_release_sock(sk);
4242 len = tcp_saved_syn_len(tp->saved_syn);
4243 if (copy_to_sockptr(optlen, &len, sizeof(int))) {
4244 sockopt_release_sock(sk);
4247 if (copy_to_sockptr(optval, tp->saved_syn->data, len)) {
4248 sockopt_release_sock(sk);
4251 tcp_saved_syn_free(tp);
4252 sockopt_release_sock(sk);
4254 sockopt_release_sock(sk);
4256 if (copy_to_sockptr(optlen, &len, sizeof(int)))
4262 case TCP_ZEROCOPY_RECEIVE: {
4263 struct scm_timestamping_internal tss;
4264 struct tcp_zerocopy_receive zc = {};
4267 if (copy_from_sockptr(&len, optlen, sizeof(int)))
4270 len < offsetofend(struct tcp_zerocopy_receive, length))
4272 if (unlikely(len > sizeof(zc))) {
4273 err = check_zeroed_sockptr(optval, sizeof(zc),
4276 return err == 0 ? -EINVAL : err;
4278 if (copy_to_sockptr(optlen, &len, sizeof(int)))
4281 if (copy_from_sockptr(&zc, optval, len))
4285 if (zc.msg_flags & ~(TCP_VALID_ZC_MSG_FLAGS))
4287 sockopt_lock_sock(sk);
4288 err = tcp_zerocopy_receive(sk, &zc, &tss);
4289 err = BPF_CGROUP_RUN_PROG_GETSOCKOPT_KERN(sk, level, optname,
4291 sockopt_release_sock(sk);
4292 if (len >= offsetofend(struct tcp_zerocopy_receive, msg_flags))
4293 goto zerocopy_rcv_cmsg;
4295 case offsetofend(struct tcp_zerocopy_receive, msg_flags):
4296 goto zerocopy_rcv_cmsg;
4297 case offsetofend(struct tcp_zerocopy_receive, msg_controllen):
4298 case offsetofend(struct tcp_zerocopy_receive, msg_control):
4299 case offsetofend(struct tcp_zerocopy_receive, flags):
4300 case offsetofend(struct tcp_zerocopy_receive, copybuf_len):
4301 case offsetofend(struct tcp_zerocopy_receive, copybuf_address):
4302 case offsetofend(struct tcp_zerocopy_receive, err):
4303 goto zerocopy_rcv_sk_err;
4304 case offsetofend(struct tcp_zerocopy_receive, inq):
4305 goto zerocopy_rcv_inq;
4306 case offsetofend(struct tcp_zerocopy_receive, length):
4308 goto zerocopy_rcv_out;
4311 if (zc.msg_flags & TCP_CMSG_TS)
4312 tcp_zc_finalize_rx_tstamp(sk, &zc, &tss);
4315 zerocopy_rcv_sk_err:
4317 zc.err = sock_error(sk);
4319 zc.inq = tcp_inq_hint(sk);
4321 if (!err && copy_to_sockptr(optval, &zc, len))
4327 if (!tcp_can_repair_sock(sk))
4329 return tcp_ao_get_repair(sk, optval, optlen);
4330 case TCP_AO_GET_KEYS:
4334 sockopt_lock_sock(sk);
4335 if (optname == TCP_AO_GET_KEYS)
4336 err = tcp_ao_get_mkts(sk, optval, optlen);
4338 err = tcp_ao_get_sock_info(sk, optval, optlen);
4339 sockopt_release_sock(sk);
4344 return -ENOPROTOOPT;
4347 if (copy_to_sockptr(optlen, &len, sizeof(int)))
4349 if (copy_to_sockptr(optval, &val, len))
4354 bool tcp_bpf_bypass_getsockopt(int level, int optname)
4356 /* TCP do_tcp_getsockopt has optimized getsockopt implementation
4357 * to avoid extra socket lock for TCP_ZEROCOPY_RECEIVE.
4359 if (level == SOL_TCP && optname == TCP_ZEROCOPY_RECEIVE)
4364 EXPORT_SYMBOL(tcp_bpf_bypass_getsockopt);
4366 int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval,
4369 struct inet_connection_sock *icsk = inet_csk(sk);
4371 if (level != SOL_TCP)
4372 /* Paired with WRITE_ONCE() in do_ipv6_setsockopt() and tcp_v6_connect() */
4373 return READ_ONCE(icsk->icsk_af_ops)->getsockopt(sk, level, optname,
4375 return do_tcp_getsockopt(sk, level, optname, USER_SOCKPTR(optval),
4376 USER_SOCKPTR(optlen));
4378 EXPORT_SYMBOL(tcp_getsockopt);
4380 #ifdef CONFIG_TCP_MD5SIG
4381 int tcp_md5_sigpool_id = -1;
4382 EXPORT_SYMBOL_GPL(tcp_md5_sigpool_id);
4384 int tcp_md5_alloc_sigpool(void)
4386 size_t scratch_size;
4389 scratch_size = sizeof(union tcp_md5sum_block) + sizeof(struct tcphdr);
4390 ret = tcp_sigpool_alloc_ahash("md5", scratch_size);
4392 /* As long as any md5 sigpool was allocated, the return
4393 * id would stay the same. Re-write the id only for the case
4394 * when previously all MD5 keys were deleted and this call
4395 * allocates the first MD5 key, which may return a different
4396 * sigpool id than was used previously.
4398 WRITE_ONCE(tcp_md5_sigpool_id, ret); /* Avoids the compiler potentially being smart here */
4404 void tcp_md5_release_sigpool(void)
4406 tcp_sigpool_release(READ_ONCE(tcp_md5_sigpool_id));
4409 void tcp_md5_add_sigpool(void)
4411 tcp_sigpool_get(READ_ONCE(tcp_md5_sigpool_id));
4414 int tcp_md5_hash_key(struct tcp_sigpool *hp,
4415 const struct tcp_md5sig_key *key)
4417 u8 keylen = READ_ONCE(key->keylen); /* paired with WRITE_ONCE() in tcp_md5_do_add */
4418 struct scatterlist sg;
4420 sg_init_one(&sg, key->key, keylen);
4421 ahash_request_set_crypt(hp->req, &sg, NULL, keylen);
4423 /* We use data_race() because tcp_md5_do_add() might change
4426 return data_race(crypto_ahash_update(hp->req));
4428 EXPORT_SYMBOL(tcp_md5_hash_key);
4430 /* Called with rcu_read_lock() */
4431 enum skb_drop_reason
4432 tcp_inbound_md5_hash(const struct sock *sk, const struct sk_buff *skb,
4433 const void *saddr, const void *daddr,
4434 int family, int l3index, const __u8 *hash_location)
4436 /* This gets called for each TCP segment that has TCP-MD5 option.
4437 * We have 3 drop cases:
4438 * o No MD5 hash and one expected.
4439 * o MD5 hash and we're not expecting one.
4440 * o MD5 hash and its wrong.
4442 const struct tcp_sock *tp = tcp_sk(sk);
4443 struct tcp_md5sig_key *key;
4447 key = tcp_md5_do_lookup(sk, l3index, saddr, family);
4449 if (!key && hash_location) {
4450 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMD5UNEXPECTED);
4451 tcp_hash_fail("Unexpected MD5 Hash found", family, skb, "");
4452 return SKB_DROP_REASON_TCP_MD5UNEXPECTED;
4455 /* Check the signature.
4456 * To support dual stack listeners, we need to handle
4459 if (family == AF_INET)
4460 genhash = tcp_v4_md5_hash_skb(newhash, key, NULL, skb);
4462 genhash = tp->af_specific->calc_md5_hash(newhash, key,
4464 if (genhash || memcmp(hash_location, newhash, 16) != 0) {
4465 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMD5FAILURE);
4466 if (family == AF_INET) {
4467 tcp_hash_fail("MD5 Hash failed", AF_INET, skb, "%s L3 index %d",
4468 genhash ? "tcp_v4_calc_md5_hash failed"
4472 tcp_hash_fail("MD5 Hash failed",
4473 AF_INET6, skb, "L3 index %d",
4476 tcp_hash_fail("MD5 Hash mismatch",
4477 AF_INET6, skb, "L3 index %d",
4481 return SKB_DROP_REASON_TCP_MD5FAILURE;
4483 return SKB_NOT_DROPPED_YET;
4485 EXPORT_SYMBOL(tcp_inbound_md5_hash);
4489 void tcp_done(struct sock *sk)
4491 struct request_sock *req;
4493 /* We might be called with a new socket, after
4494 * inet_csk_prepare_forced_close() has been called
4495 * so we can not use lockdep_sock_is_held(sk)
4497 req = rcu_dereference_protected(tcp_sk(sk)->fastopen_rsk, 1);
4499 if (sk->sk_state == TCP_SYN_SENT || sk->sk_state == TCP_SYN_RECV)
4500 TCP_INC_STATS(sock_net(sk), TCP_MIB_ATTEMPTFAILS);
4502 tcp_set_state(sk, TCP_CLOSE);
4503 tcp_clear_xmit_timers(sk);
4505 reqsk_fastopen_remove(sk, req, false);
4507 WRITE_ONCE(sk->sk_shutdown, SHUTDOWN_MASK);
4509 if (!sock_flag(sk, SOCK_DEAD))
4510 sk->sk_state_change(sk);
4512 inet_csk_destroy_sock(sk);
4514 EXPORT_SYMBOL_GPL(tcp_done);
4516 int tcp_abort(struct sock *sk, int err)
4518 int state = inet_sk_state_load(sk);
4520 if (state == TCP_NEW_SYN_RECV) {
4521 struct request_sock *req = inet_reqsk(sk);
4524 inet_csk_reqsk_queue_drop(req->rsk_listener, req);
4528 if (state == TCP_TIME_WAIT) {
4529 struct inet_timewait_sock *tw = inet_twsk(sk);
4531 refcount_inc(&tw->tw_refcnt);
4533 inet_twsk_deschedule_put(tw);
4538 /* BPF context ensures sock locking. */
4539 if (!has_current_bpf_ctx())
4540 /* Don't race with userspace socket closes such as tcp_close. */
4543 if (sk->sk_state == TCP_LISTEN) {
4544 tcp_set_state(sk, TCP_CLOSE);
4545 inet_csk_listen_stop(sk);
4548 /* Don't race with BH socket closes such as inet_csk_listen_stop. */
4552 if (!sock_flag(sk, SOCK_DEAD)) {
4553 WRITE_ONCE(sk->sk_err, err);
4554 /* This barrier is coupled with smp_rmb() in tcp_poll() */
4556 sk_error_report(sk);
4557 if (tcp_need_reset(sk->sk_state))
4558 tcp_send_active_reset(sk, GFP_ATOMIC);
4564 tcp_write_queue_purge(sk);
4565 if (!has_current_bpf_ctx())
4569 EXPORT_SYMBOL_GPL(tcp_abort);
4571 extern struct tcp_congestion_ops tcp_reno;
4573 static __initdata unsigned long thash_entries;
4574 static int __init set_thash_entries(char *str)
4581 ret = kstrtoul(str, 0, &thash_entries);
4587 __setup("thash_entries=", set_thash_entries);
4589 static void __init tcp_init_mem(void)
4591 unsigned long limit = nr_free_buffer_pages() / 16;
4593 limit = max(limit, 128UL);
4594 sysctl_tcp_mem[0] = limit / 4 * 3; /* 4.68 % */
4595 sysctl_tcp_mem[1] = limit; /* 6.25 % */
4596 sysctl_tcp_mem[2] = sysctl_tcp_mem[0] * 2; /* 9.37 % */
4599 static void __init tcp_struct_check(void)
4601 /* TX read-mostly hotpath cache lines */
4602 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_tx, max_window);
4603 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_tx, rcv_ssthresh);
4604 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_tx, reordering);
4605 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_tx, notsent_lowat);
4606 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_tx, gso_segs);
4607 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_tx, lost_skb_hint);
4608 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_tx, retransmit_skb_hint);
4609 CACHELINE_ASSERT_GROUP_SIZE(struct tcp_sock, tcp_sock_read_tx, 40);
4611 /* TXRX read-mostly hotpath cache lines */
4612 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_txrx, tsoffset);
4613 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_txrx, snd_wnd);
4614 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_txrx, mss_cache);
4615 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_txrx, snd_cwnd);
4616 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_txrx, prr_out);
4617 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_txrx, lost_out);
4618 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_txrx, sacked_out);
4619 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_txrx, scaling_ratio);
4620 CACHELINE_ASSERT_GROUP_SIZE(struct tcp_sock, tcp_sock_read_txrx, 32);
4622 /* RX read-mostly hotpath cache lines */
4623 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_rx, copied_seq);
4624 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_rx, rcv_tstamp);
4625 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_rx, snd_wl1);
4626 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_rx, tlp_high_seq);
4627 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_rx, rttvar_us);
4628 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_rx, retrans_out);
4629 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_rx, advmss);
4630 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_rx, urg_data);
4631 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_rx, lost);
4632 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_rx, rtt_min);
4633 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_rx, out_of_order_queue);
4634 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_rx, snd_ssthresh);
4635 CACHELINE_ASSERT_GROUP_SIZE(struct tcp_sock, tcp_sock_read_rx, 69);
4637 /* TX read-write hotpath cache lines */
4638 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, segs_out);
4639 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, data_segs_out);
4640 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, bytes_sent);
4641 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, snd_sml);
4642 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, chrono_start);
4643 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, chrono_stat);
4644 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, write_seq);
4645 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, pushed_seq);
4646 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, lsndtime);
4647 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, mdev_us);
4648 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, tcp_wstamp_ns);
4649 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, tcp_clock_cache);
4650 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, tcp_mstamp);
4651 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, rtt_seq);
4652 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, tsorted_sent_queue);
4653 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, highest_sack);
4654 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, ecn_flags);
4655 CACHELINE_ASSERT_GROUP_SIZE(struct tcp_sock, tcp_sock_write_tx, 105);
4657 /* TXRX read-write hotpath cache lines */
4658 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_txrx, pred_flags);
4659 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_txrx, rcv_nxt);
4660 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_txrx, snd_nxt);
4661 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_txrx, snd_una);
4662 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_txrx, window_clamp);
4663 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_txrx, srtt_us);
4664 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_txrx, packets_out);
4665 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_txrx, snd_up);
4666 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_txrx, delivered);
4667 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_txrx, delivered_ce);
4668 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_txrx, app_limited);
4669 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_txrx, rcv_wnd);
4670 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_txrx, rx_opt);
4671 CACHELINE_ASSERT_GROUP_SIZE(struct tcp_sock, tcp_sock_write_txrx, 76);
4673 /* RX read-write hotpath cache lines */
4674 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_rx, bytes_received);
4675 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_rx, segs_in);
4676 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_rx, data_segs_in);
4677 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_rx, rcv_wup);
4678 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_rx, max_packets_out);
4679 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_rx, cwnd_usage_seq);
4680 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_rx, rate_delivered);
4681 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_rx, rate_interval_us);
4682 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_rx, rcv_rtt_last_tsecr);
4683 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_rx, first_tx_mstamp);
4684 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_rx, delivered_mstamp);
4685 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_rx, bytes_acked);
4686 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_rx, rcv_rtt_est);
4687 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_rx, rcvq_space);
4688 CACHELINE_ASSERT_GROUP_SIZE(struct tcp_sock, tcp_sock_write_rx, 99);
4691 void __init tcp_init(void)
4693 int max_rshare, max_wshare, cnt;
4694 unsigned long limit;
4697 BUILD_BUG_ON(TCP_MIN_SND_MSS <= MAX_TCP_OPTION_SPACE);
4698 BUILD_BUG_ON(sizeof(struct tcp_skb_cb) >
4699 sizeof_field(struct sk_buff, cb));
4703 percpu_counter_init(&tcp_sockets_allocated, 0, GFP_KERNEL);
4705 timer_setup(&tcp_orphan_timer, tcp_orphan_update, TIMER_DEFERRABLE);
4706 mod_timer(&tcp_orphan_timer, jiffies + TCP_ORPHAN_TIMER_PERIOD);
4708 inet_hashinfo2_init(&tcp_hashinfo, "tcp_listen_portaddr_hash",
4709 thash_entries, 21, /* one slot per 2 MB*/
4711 tcp_hashinfo.bind_bucket_cachep =
4712 kmem_cache_create("tcp_bind_bucket",
4713 sizeof(struct inet_bind_bucket), 0,
4714 SLAB_HWCACHE_ALIGN | SLAB_PANIC |
4717 tcp_hashinfo.bind2_bucket_cachep =
4718 kmem_cache_create("tcp_bind2_bucket",
4719 sizeof(struct inet_bind2_bucket), 0,
4720 SLAB_HWCACHE_ALIGN | SLAB_PANIC |
4724 /* Size and allocate the main established and bind bucket
4727 * The methodology is similar to that of the buffer cache.
4729 tcp_hashinfo.ehash =
4730 alloc_large_system_hash("TCP established",
4731 sizeof(struct inet_ehash_bucket),
4733 17, /* one slot per 128 KB of memory */
4736 &tcp_hashinfo.ehash_mask,
4738 thash_entries ? 0 : 512 * 1024);
4739 for (i = 0; i <= tcp_hashinfo.ehash_mask; i++)
4740 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].chain, i);
4742 if (inet_ehash_locks_alloc(&tcp_hashinfo))
4743 panic("TCP: failed to alloc ehash_locks");
4744 tcp_hashinfo.bhash =
4745 alloc_large_system_hash("TCP bind",
4746 2 * sizeof(struct inet_bind_hashbucket),
4747 tcp_hashinfo.ehash_mask + 1,
4748 17, /* one slot per 128 KB of memory */
4750 &tcp_hashinfo.bhash_size,
4754 tcp_hashinfo.bhash_size = 1U << tcp_hashinfo.bhash_size;
4755 tcp_hashinfo.bhash2 = tcp_hashinfo.bhash + tcp_hashinfo.bhash_size;
4756 for (i = 0; i < tcp_hashinfo.bhash_size; i++) {
4757 spin_lock_init(&tcp_hashinfo.bhash[i].lock);
4758 INIT_HLIST_HEAD(&tcp_hashinfo.bhash[i].chain);
4759 spin_lock_init(&tcp_hashinfo.bhash2[i].lock);
4760 INIT_HLIST_HEAD(&tcp_hashinfo.bhash2[i].chain);
4763 tcp_hashinfo.pernet = false;
4765 cnt = tcp_hashinfo.ehash_mask + 1;
4766 sysctl_tcp_max_orphans = cnt / 2;
4769 /* Set per-socket limits to no more than 1/128 the pressure threshold */
4770 limit = nr_free_buffer_pages() << (PAGE_SHIFT - 7);
4771 max_wshare = min(4UL*1024*1024, limit);
4772 max_rshare = min(6UL*1024*1024, limit);
4774 init_net.ipv4.sysctl_tcp_wmem[0] = PAGE_SIZE;
4775 init_net.ipv4.sysctl_tcp_wmem[1] = 16*1024;
4776 init_net.ipv4.sysctl_tcp_wmem[2] = max(64*1024, max_wshare);
4778 init_net.ipv4.sysctl_tcp_rmem[0] = PAGE_SIZE;
4779 init_net.ipv4.sysctl_tcp_rmem[1] = 131072;
4780 init_net.ipv4.sysctl_tcp_rmem[2] = max(131072, max_rshare);
4782 pr_info("Hash tables configured (established %u bind %u)\n",
4783 tcp_hashinfo.ehash_mask + 1, tcp_hashinfo.bhash_size);
4787 BUG_ON(tcp_register_congestion_control(&tcp_reno) != 0);