1 // SPDX-License-Identifier: GPL-2.0-or-later
3 * INET An implementation of the TCP/IP protocol suite for the LINUX
4 * operating system. INET is implemented using the BSD Socket
5 * interface as the means of communication with the user level.
7 * Implementation of the Transmission Control Protocol(TCP).
10 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
11 * Mark Evans, <evansmp@uhura.aston.ac.uk>
12 * Corey Minyard <wf-rch!minyard@relay.EU.net>
13 * Florian La Roche, <flla@stud.uni-sb.de>
14 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
15 * Linus Torvalds, <torvalds@cs.helsinki.fi>
16 * Alan Cox, <gw4pts@gw4pts.ampr.org>
17 * Matthew Dillon, <dillon@apollo.west.oic.com>
18 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
19 * Jorge Cwik, <jorge@laser.satlink.net>
22 * Alan Cox : Numerous verify_area() calls
23 * Alan Cox : Set the ACK bit on a reset
24 * Alan Cox : Stopped it crashing if it closed while
25 * sk->inuse=1 and was trying to connect
27 * Alan Cox : All icmp error handling was broken
28 * pointers passed where wrong and the
29 * socket was looked up backwards. Nobody
30 * tested any icmp error code obviously.
31 * Alan Cox : tcp_err() now handled properly. It
32 * wakes people on errors. poll
33 * behaves and the icmp error race
34 * has gone by moving it into sock.c
35 * Alan Cox : tcp_send_reset() fixed to work for
36 * everything not just packets for
38 * Alan Cox : tcp option processing.
39 * Alan Cox : Reset tweaked (still not 100%) [Had
41 * Herp Rosmanith : More reset fixes
42 * Alan Cox : No longer acks invalid rst frames.
43 * Acking any kind of RST is right out.
44 * Alan Cox : Sets an ignore me flag on an rst
45 * receive otherwise odd bits of prattle
47 * Alan Cox : Fixed another acking RST frame bug.
48 * Should stop LAN workplace lockups.
49 * Alan Cox : Some tidyups using the new skb list
51 * Alan Cox : sk->keepopen now seems to work
52 * Alan Cox : Pulls options out correctly on accepts
53 * Alan Cox : Fixed assorted sk->rqueue->next errors
54 * Alan Cox : PSH doesn't end a TCP read. Switched a
56 * Alan Cox : Tidied tcp_data to avoid a potential
58 * Alan Cox : Added some better commenting, as the
59 * tcp is hard to follow
60 * Alan Cox : Removed incorrect check for 20 * psh
61 * Michael O'Reilly : ack < copied bug fix.
62 * Johannes Stille : Misc tcp fixes (not all in yet).
63 * Alan Cox : FIN with no memory -> CRASH
64 * Alan Cox : Added socket option proto entries.
65 * Also added awareness of them to accept.
66 * Alan Cox : Added TCP options (SOL_TCP)
67 * Alan Cox : Switched wakeup calls to callbacks,
68 * so the kernel can layer network
70 * Alan Cox : Use ip_tos/ip_ttl settings.
71 * Alan Cox : Handle FIN (more) properly (we hope).
72 * Alan Cox : RST frames sent on unsynchronised
74 * Alan Cox : Put in missing check for SYN bit.
75 * Alan Cox : Added tcp_select_window() aka NET2E
76 * window non shrink trick.
77 * Alan Cox : Added a couple of small NET2E timer
79 * Charles Hedrick : TCP fixes
80 * Toomas Tamm : TCP window fixes
81 * Alan Cox : Small URG fix to rlogin ^C ack fight
82 * Charles Hedrick : Rewrote most of it to actually work
83 * Linus : Rewrote tcp_read() and URG handling
85 * Gerhard Koerting: Fixed some missing timer handling
86 * Matthew Dillon : Reworked TCP machine states as per RFC
87 * Gerhard Koerting: PC/TCP workarounds
88 * Adam Caldwell : Assorted timer/timing errors
89 * Matthew Dillon : Fixed another RST bug
90 * Alan Cox : Move to kernel side addressing changes.
91 * Alan Cox : Beginning work on TCP fastpathing
93 * Arnt Gulbrandsen: Turbocharged tcp_check() routine.
94 * Alan Cox : TCP fast path debugging
95 * Alan Cox : Window clamping
96 * Michael Riepe : Bug in tcp_check()
97 * Matt Dillon : More TCP improvements and RST bug fixes
98 * Matt Dillon : Yet more small nasties remove from the
99 * TCP code (Be very nice to this man if
100 * tcp finally works 100%) 8)
101 * Alan Cox : BSD accept semantics.
102 * Alan Cox : Reset on closedown bug.
103 * Peter De Schrijver : ENOTCONN check missing in tcp_sendto().
104 * Michael Pall : Handle poll() after URG properly in
106 * Michael Pall : Undo the last fix in tcp_read_urg()
107 * (multi URG PUSH broke rlogin).
108 * Michael Pall : Fix the multi URG PUSH problem in
109 * tcp_readable(), poll() after URG
111 * Michael Pall : recv(...,MSG_OOB) never blocks in the
113 * Alan Cox : Changed the semantics of sk->socket to
114 * fix a race and a signal problem with
115 * accept() and async I/O.
116 * Alan Cox : Relaxed the rules on tcp_sendto().
117 * Yury Shevchuk : Really fixed accept() blocking problem.
118 * Craig I. Hagan : Allow for BSD compatible TIME_WAIT for
119 * clients/servers which listen in on
121 * Alan Cox : Cleaned the above up and shrank it to
122 * a sensible code size.
123 * Alan Cox : Self connect lockup fix.
124 * Alan Cox : No connect to multicast.
125 * Ross Biro : Close unaccepted children on master
127 * Alan Cox : Reset tracing code.
128 * Alan Cox : Spurious resets on shutdown.
129 * Alan Cox : Giant 15 minute/60 second timer error
130 * Alan Cox : Small whoops in polling before an
132 * Alan Cox : Kept the state trace facility since
133 * it's handy for debugging.
134 * Alan Cox : More reset handler fixes.
135 * Alan Cox : Started rewriting the code based on
136 * the RFC's for other useful protocol
137 * references see: Comer, KA9Q NOS, and
138 * for a reference on the difference
139 * between specifications and how BSD
140 * works see the 4.4lite source.
141 * A.N.Kuznetsov : Don't time wait on completion of tidy
143 * Linus Torvalds : Fin/Shutdown & copied_seq changes.
144 * Linus Torvalds : Fixed BSD port reuse to work first syn
145 * Alan Cox : Reimplemented timers as per the RFC
146 * and using multiple timers for sanity.
147 * Alan Cox : Small bug fixes, and a lot of new
149 * Alan Cox : Fixed dual reader crash by locking
150 * the buffers (much like datagram.c)
151 * Alan Cox : Fixed stuck sockets in probe. A probe
152 * now gets fed up of retrying without
153 * (even a no space) answer.
154 * Alan Cox : Extracted closing code better
155 * Alan Cox : Fixed the closing state machine to
157 * Alan Cox : More 'per spec' fixes.
158 * Jorge Cwik : Even faster checksumming.
159 * Alan Cox : tcp_data() doesn't ack illegal PSH
160 * only frames. At least one pc tcp stack
162 * Alan Cox : Cache last socket.
163 * Alan Cox : Per route irtt.
164 * Matt Day : poll()->select() match BSD precisely on error
165 * Alan Cox : New buffers
166 * Marc Tamsky : Various sk->prot->retransmits and
167 * sk->retransmits misupdating fixed.
168 * Fixed tcp_write_timeout: stuck close,
169 * and TCP syn retries gets used now.
170 * Mark Yarvis : In tcp_read_wakeup(), don't send an
171 * ack if state is TCP_CLOSED.
172 * Alan Cox : Look up device on a retransmit - routes may
173 * change. Doesn't yet cope with MSS shrink right
175 * Marc Tamsky : Closing in closing fixes.
176 * Mike Shaver : RFC1122 verifications.
177 * Alan Cox : rcv_saddr errors.
178 * Alan Cox : Block double connect().
179 * Alan Cox : Small hooks for enSKIP.
180 * Alexey Kuznetsov: Path MTU discovery.
181 * Alan Cox : Support soft errors.
182 * Alan Cox : Fix MTU discovery pathological case
183 * when the remote claims no mtu!
184 * Marc Tamsky : TCP_CLOSE fix.
185 * Colin (G3TNE) : Send a reset on syn ack replies in
186 * window but wrong (fixes NT lpd problems)
187 * Pedro Roque : Better TCP window handling, delayed ack.
188 * Joerg Reuter : No modification of locked buffers in
189 * tcp_do_retransmit()
190 * Eric Schenk : Changed receiver side silly window
191 * avoidance algorithm to BSD style
192 * algorithm. This doubles throughput
193 * against machines running Solaris,
194 * and seems to result in general
196 * Stefan Magdalinski : adjusted tcp_readable() to fix FIONREAD
197 * Willy Konynenberg : Transparent proxying support.
198 * Mike McLagan : Routing by source
199 * Keith Owens : Do proper merging with partial SKB's in
200 * tcp_do_sendmsg to avoid burstiness.
201 * Eric Schenk : Fix fast close down bug with
202 * shutdown() followed by close().
203 * Andi Kleen : Make poll agree with SIGIO
204 * Salvatore Sanfilippo : Support SO_LINGER with linger == 1 and
205 * lingertime == 0 (RFC 793 ABORT Call)
206 * Hirokazu Takahashi : Use copy_from_user() instead of
207 * csum_and_copy_from_user() if possible.
209 * Description of States:
211 * TCP_SYN_SENT sent a connection request, waiting for ack
213 * TCP_SYN_RECV received a connection request, sent ack,
214 * waiting for final ack in three-way handshake.
216 * TCP_ESTABLISHED connection established
218 * TCP_FIN_WAIT1 our side has shutdown, waiting to complete
219 * transmission of remaining buffered data
221 * TCP_FIN_WAIT2 all buffered data sent, waiting for remote
224 * TCP_CLOSING both sides have shutdown but we still have
225 * data we have to finish sending
227 * TCP_TIME_WAIT timeout to catch resent junk before entering
228 * closed, can only be entered from FIN_WAIT2
229 * or CLOSING. Required because the other end
230 * may not have gotten our last ACK causing it
231 * to retransmit the data packet (which we ignore)
233 * TCP_CLOSE_WAIT remote side has shutdown and is waiting for
234 * us to finish writing our data and to shutdown
235 * (we have to close() to move on to LAST_ACK)
237 * TCP_LAST_ACK out side has shutdown after remote has
238 * shutdown. There may still be data in our
239 * buffer that we have to finish sending
241 * TCP_CLOSE socket is finished
244 #define pr_fmt(fmt) "TCP: " fmt
246 #include <crypto/hash.h>
247 #include <linux/kernel.h>
248 #include <linux/module.h>
249 #include <linux/types.h>
250 #include <linux/fcntl.h>
251 #include <linux/poll.h>
252 #include <linux/inet_diag.h>
253 #include <linux/init.h>
254 #include <linux/fs.h>
255 #include <linux/skbuff.h>
256 #include <linux/scatterlist.h>
257 #include <linux/splice.h>
258 #include <linux/net.h>
259 #include <linux/socket.h>
260 #include <linux/random.h>
261 #include <linux/memblock.h>
262 #include <linux/highmem.h>
263 #include <linux/cache.h>
264 #include <linux/err.h>
265 #include <linux/time.h>
266 #include <linux/slab.h>
267 #include <linux/errqueue.h>
268 #include <linux/static_key.h>
269 #include <linux/btf.h>
271 #include <net/icmp.h>
272 #include <net/inet_common.h>
274 #include <net/mptcp.h>
275 #include <net/xfrm.h>
277 #include <net/sock.h>
279 #include <linux/uaccess.h>
280 #include <asm/ioctls.h>
281 #include <net/busy_poll.h>
283 /* Track pending CMSGs. */
289 DEFINE_PER_CPU(unsigned int, tcp_orphan_count);
290 EXPORT_PER_CPU_SYMBOL_GPL(tcp_orphan_count);
292 long sysctl_tcp_mem[3] __read_mostly;
293 EXPORT_SYMBOL(sysctl_tcp_mem);
295 atomic_long_t tcp_memory_allocated ____cacheline_aligned_in_smp; /* Current allocated memory. */
296 EXPORT_SYMBOL(tcp_memory_allocated);
297 DEFINE_PER_CPU(int, tcp_memory_per_cpu_fw_alloc);
298 EXPORT_PER_CPU_SYMBOL_GPL(tcp_memory_per_cpu_fw_alloc);
300 #if IS_ENABLED(CONFIG_SMC)
301 DEFINE_STATIC_KEY_FALSE(tcp_have_smc);
302 EXPORT_SYMBOL(tcp_have_smc);
306 * Current number of TCP sockets.
308 struct percpu_counter tcp_sockets_allocated ____cacheline_aligned_in_smp;
309 EXPORT_SYMBOL(tcp_sockets_allocated);
314 struct tcp_splice_state {
315 struct pipe_inode_info *pipe;
321 * Pressure flag: try to collapse.
322 * Technical note: it is used by multiple contexts non atomically.
323 * All the __sk_mem_schedule() is of this nature: accounting
324 * is strict, actions are advisory and have some latency.
326 unsigned long tcp_memory_pressure __read_mostly;
327 EXPORT_SYMBOL_GPL(tcp_memory_pressure);
329 void tcp_enter_memory_pressure(struct sock *sk)
333 if (READ_ONCE(tcp_memory_pressure))
339 if (!cmpxchg(&tcp_memory_pressure, 0, val))
340 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMEMORYPRESSURES);
342 EXPORT_SYMBOL_GPL(tcp_enter_memory_pressure);
344 void tcp_leave_memory_pressure(struct sock *sk)
348 if (!READ_ONCE(tcp_memory_pressure))
350 val = xchg(&tcp_memory_pressure, 0);
352 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPMEMORYPRESSURESCHRONO,
353 jiffies_to_msecs(jiffies - val));
355 EXPORT_SYMBOL_GPL(tcp_leave_memory_pressure);
357 /* Convert seconds to retransmits based on initial and max timeout */
358 static u8 secs_to_retrans(int seconds, int timeout, int rto_max)
363 int period = timeout;
366 while (seconds > period && res < 255) {
369 if (timeout > rto_max)
377 /* Convert retransmits to seconds based on initial and max timeout */
378 static int retrans_to_secs(u8 retrans, int timeout, int rto_max)
386 if (timeout > rto_max)
394 static u64 tcp_compute_delivery_rate(const struct tcp_sock *tp)
396 u32 rate = READ_ONCE(tp->rate_delivered);
397 u32 intv = READ_ONCE(tp->rate_interval_us);
401 rate64 = (u64)rate * tp->mss_cache * USEC_PER_SEC;
402 do_div(rate64, intv);
407 /* Address-family independent initialization for a tcp_sock.
409 * NOTE: A lot of things set to zero explicitly by call to
410 * sk_alloc() so need not be done here.
412 void tcp_init_sock(struct sock *sk)
414 struct inet_connection_sock *icsk = inet_csk(sk);
415 struct tcp_sock *tp = tcp_sk(sk);
417 tp->out_of_order_queue = RB_ROOT;
418 sk->tcp_rtx_queue = RB_ROOT;
419 tcp_init_xmit_timers(sk);
420 INIT_LIST_HEAD(&tp->tsq_node);
421 INIT_LIST_HEAD(&tp->tsorted_sent_queue);
423 icsk->icsk_rto = TCP_TIMEOUT_INIT;
424 icsk->icsk_rto_min = TCP_RTO_MIN;
425 icsk->icsk_delack_max = TCP_DELACK_MAX;
426 tp->mdev_us = jiffies_to_usecs(TCP_TIMEOUT_INIT);
427 minmax_reset(&tp->rtt_min, tcp_jiffies32, ~0U);
429 /* So many TCP implementations out there (incorrectly) count the
430 * initial SYN frame in their delayed-ACK and congestion control
431 * algorithms that we must have the following bandaid to talk
432 * efficiently to them. -DaveM
434 tcp_snd_cwnd_set(tp, TCP_INIT_CWND);
436 /* There's a bubble in the pipe until at least the first ACK. */
437 tp->app_limited = ~0U;
438 tp->rate_app_limited = 1;
440 /* See draft-stevens-tcpca-spec-01 for discussion of the
441 * initialization of these values.
443 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
444 tp->snd_cwnd_clamp = ~0;
445 tp->mss_cache = TCP_MSS_DEFAULT;
447 tp->reordering = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_reordering);
448 tcp_assign_congestion_control(sk);
451 tp->rack.reo_wnd_steps = 1;
453 sk->sk_write_space = sk_stream_write_space;
454 sock_set_flag(sk, SOCK_USE_WRITE_QUEUE);
456 icsk->icsk_sync_mss = tcp_sync_mss;
458 WRITE_ONCE(sk->sk_sndbuf, READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_wmem[1]));
459 WRITE_ONCE(sk->sk_rcvbuf, READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_rmem[1]));
460 tcp_scaling_ratio_init(sk);
462 set_bit(SOCK_SUPPORT_ZC, &sk->sk_socket->flags);
463 sk_sockets_allocated_inc(sk);
465 EXPORT_SYMBOL(tcp_init_sock);
467 static void tcp_tx_timestamp(struct sock *sk, u16 tsflags)
469 struct sk_buff *skb = tcp_write_queue_tail(sk);
471 if (tsflags && skb) {
472 struct skb_shared_info *shinfo = skb_shinfo(skb);
473 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
475 sock_tx_timestamp(sk, tsflags, &shinfo->tx_flags);
476 if (tsflags & SOF_TIMESTAMPING_TX_ACK)
477 tcb->txstamp_ack = 1;
478 if (tsflags & SOF_TIMESTAMPING_TX_RECORD_MASK)
479 shinfo->tskey = TCP_SKB_CB(skb)->seq + skb->len - 1;
483 static bool tcp_stream_is_readable(struct sock *sk, int target)
485 if (tcp_epollin_ready(sk, target))
487 return sk_is_readable(sk);
491 * Wait for a TCP event.
493 * Note that we don't need to lock the socket, as the upper poll layers
494 * take care of normal races (between the test and the event) and we don't
495 * go look at any of the socket buffers directly.
497 __poll_t tcp_poll(struct file *file, struct socket *sock, poll_table *wait)
500 struct sock *sk = sock->sk;
501 const struct tcp_sock *tp = tcp_sk(sk);
505 sock_poll_wait(file, sock, wait);
507 state = inet_sk_state_load(sk);
508 if (state == TCP_LISTEN)
509 return inet_csk_listen_poll(sk);
511 /* Socket is not locked. We are protected from async events
512 * by poll logic and correct handling of state changes
513 * made by other threads is impossible in any case.
519 * EPOLLHUP is certainly not done right. But poll() doesn't
520 * have a notion of HUP in just one direction, and for a
521 * socket the read side is more interesting.
523 * Some poll() documentation says that EPOLLHUP is incompatible
524 * with the EPOLLOUT/POLLWR flags, so somebody should check this
525 * all. But careful, it tends to be safer to return too many
526 * bits than too few, and you can easily break real applications
527 * if you don't tell them that something has hung up!
531 * Check number 1. EPOLLHUP is _UNMASKABLE_ event (see UNIX98 and
532 * our fs/select.c). It means that after we received EOF,
533 * poll always returns immediately, making impossible poll() on write()
534 * in state CLOSE_WAIT. One solution is evident --- to set EPOLLHUP
535 * if and only if shutdown has been made in both directions.
536 * Actually, it is interesting to look how Solaris and DUX
537 * solve this dilemma. I would prefer, if EPOLLHUP were maskable,
538 * then we could set it on SND_SHUTDOWN. BTW examples given
539 * in Stevens' books assume exactly this behaviour, it explains
540 * why EPOLLHUP is incompatible with EPOLLOUT. --ANK
542 * NOTE. Check for TCP_CLOSE is added. The goal is to prevent
543 * blocking on fresh not-connected or disconnected socket. --ANK
545 shutdown = READ_ONCE(sk->sk_shutdown);
546 if (shutdown == SHUTDOWN_MASK || state == TCP_CLOSE)
548 if (shutdown & RCV_SHUTDOWN)
549 mask |= EPOLLIN | EPOLLRDNORM | EPOLLRDHUP;
551 /* Connected or passive Fast Open socket? */
552 if (state != TCP_SYN_SENT &&
553 (state != TCP_SYN_RECV || rcu_access_pointer(tp->fastopen_rsk))) {
554 int target = sock_rcvlowat(sk, 0, INT_MAX);
555 u16 urg_data = READ_ONCE(tp->urg_data);
557 if (unlikely(urg_data) &&
558 READ_ONCE(tp->urg_seq) == READ_ONCE(tp->copied_seq) &&
559 !sock_flag(sk, SOCK_URGINLINE))
562 if (tcp_stream_is_readable(sk, target))
563 mask |= EPOLLIN | EPOLLRDNORM;
565 if (!(shutdown & SEND_SHUTDOWN)) {
566 if (__sk_stream_is_writeable(sk, 1)) {
567 mask |= EPOLLOUT | EPOLLWRNORM;
568 } else { /* send SIGIO later */
569 sk_set_bit(SOCKWQ_ASYNC_NOSPACE, sk);
570 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
572 /* Race breaker. If space is freed after
573 * wspace test but before the flags are set,
574 * IO signal will be lost. Memory barrier
575 * pairs with the input side.
577 smp_mb__after_atomic();
578 if (__sk_stream_is_writeable(sk, 1))
579 mask |= EPOLLOUT | EPOLLWRNORM;
582 mask |= EPOLLOUT | EPOLLWRNORM;
584 if (urg_data & TCP_URG_VALID)
586 } else if (state == TCP_SYN_SENT &&
587 inet_test_bit(DEFER_CONNECT, sk)) {
588 /* Active TCP fastopen socket with defer_connect
589 * Return EPOLLOUT so application can call write()
590 * in order for kernel to generate SYN+data
592 mask |= EPOLLOUT | EPOLLWRNORM;
594 /* This barrier is coupled with smp_wmb() in tcp_reset() */
596 if (READ_ONCE(sk->sk_err) ||
597 !skb_queue_empty_lockless(&sk->sk_error_queue))
602 EXPORT_SYMBOL(tcp_poll);
604 int tcp_ioctl(struct sock *sk, int cmd, int *karg)
606 struct tcp_sock *tp = tcp_sk(sk);
612 if (sk->sk_state == TCP_LISTEN)
615 slow = lock_sock_fast(sk);
617 unlock_sock_fast(sk, slow);
620 answ = READ_ONCE(tp->urg_data) &&
621 READ_ONCE(tp->urg_seq) == READ_ONCE(tp->copied_seq);
624 if (sk->sk_state == TCP_LISTEN)
627 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
630 answ = READ_ONCE(tp->write_seq) - tp->snd_una;
633 if (sk->sk_state == TCP_LISTEN)
636 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
639 answ = READ_ONCE(tp->write_seq) -
640 READ_ONCE(tp->snd_nxt);
649 EXPORT_SYMBOL(tcp_ioctl);
651 void tcp_mark_push(struct tcp_sock *tp, struct sk_buff *skb)
653 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
654 tp->pushed_seq = tp->write_seq;
657 static inline bool forced_push(const struct tcp_sock *tp)
659 return after(tp->write_seq, tp->pushed_seq + (tp->max_window >> 1));
662 void tcp_skb_entail(struct sock *sk, struct sk_buff *skb)
664 struct tcp_sock *tp = tcp_sk(sk);
665 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
667 tcb->seq = tcb->end_seq = tp->write_seq;
668 tcb->tcp_flags = TCPHDR_ACK;
669 __skb_header_release(skb);
670 tcp_add_write_queue_tail(sk, skb);
671 sk_wmem_queued_add(sk, skb->truesize);
672 sk_mem_charge(sk, skb->truesize);
673 if (tp->nonagle & TCP_NAGLE_PUSH)
674 tp->nonagle &= ~TCP_NAGLE_PUSH;
676 tcp_slow_start_after_idle_check(sk);
679 static inline void tcp_mark_urg(struct tcp_sock *tp, int flags)
682 tp->snd_up = tp->write_seq;
685 /* If a not yet filled skb is pushed, do not send it if
686 * we have data packets in Qdisc or NIC queues :
687 * Because TX completion will happen shortly, it gives a chance
688 * to coalesce future sendmsg() payload into this skb, without
689 * need for a timer, and with no latency trade off.
690 * As packets containing data payload have a bigger truesize
691 * than pure acks (dataless) packets, the last checks prevent
692 * autocorking if we only have an ACK in Qdisc/NIC queues,
693 * or if TX completion was delayed after we processed ACK packet.
695 static bool tcp_should_autocork(struct sock *sk, struct sk_buff *skb,
698 return skb->len < size_goal &&
699 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_autocorking) &&
700 !tcp_rtx_queue_empty(sk) &&
701 refcount_read(&sk->sk_wmem_alloc) > skb->truesize &&
702 tcp_skb_can_collapse_to(skb);
705 void tcp_push(struct sock *sk, int flags, int mss_now,
706 int nonagle, int size_goal)
708 struct tcp_sock *tp = tcp_sk(sk);
711 skb = tcp_write_queue_tail(sk);
714 if (!(flags & MSG_MORE) || forced_push(tp))
715 tcp_mark_push(tp, skb);
717 tcp_mark_urg(tp, flags);
719 if (tcp_should_autocork(sk, skb, size_goal)) {
721 /* avoid atomic op if TSQ_THROTTLED bit is already set */
722 if (!test_bit(TSQ_THROTTLED, &sk->sk_tsq_flags)) {
723 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPAUTOCORKING);
724 set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags);
726 /* It is possible TX completion already happened
727 * before we set TSQ_THROTTLED.
729 if (refcount_read(&sk->sk_wmem_alloc) > skb->truesize)
733 if (flags & MSG_MORE)
734 nonagle = TCP_NAGLE_CORK;
736 __tcp_push_pending_frames(sk, mss_now, nonagle);
739 static int tcp_splice_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb,
740 unsigned int offset, size_t len)
742 struct tcp_splice_state *tss = rd_desc->arg.data;
745 ret = skb_splice_bits(skb, skb->sk, offset, tss->pipe,
746 min(rd_desc->count, len), tss->flags);
748 rd_desc->count -= ret;
752 static int __tcp_splice_read(struct sock *sk, struct tcp_splice_state *tss)
754 /* Store TCP splice context information in read_descriptor_t. */
755 read_descriptor_t rd_desc = {
760 return tcp_read_sock(sk, &rd_desc, tcp_splice_data_recv);
764 * tcp_splice_read - splice data from TCP socket to a pipe
765 * @sock: socket to splice from
766 * @ppos: position (not valid)
767 * @pipe: pipe to splice to
768 * @len: number of bytes to splice
769 * @flags: splice modifier flags
772 * Will read pages from given socket and fill them into a pipe.
775 ssize_t tcp_splice_read(struct socket *sock, loff_t *ppos,
776 struct pipe_inode_info *pipe, size_t len,
779 struct sock *sk = sock->sk;
780 struct tcp_splice_state tss = {
789 sock_rps_record_flow(sk);
791 * We can't seek on a socket input
800 timeo = sock_rcvtimeo(sk, sock->file->f_flags & O_NONBLOCK);
802 ret = __tcp_splice_read(sk, &tss);
808 if (sock_flag(sk, SOCK_DONE))
811 ret = sock_error(sk);
814 if (sk->sk_shutdown & RCV_SHUTDOWN)
816 if (sk->sk_state == TCP_CLOSE) {
818 * This occurs when user tries to read
819 * from never connected socket.
828 /* if __tcp_splice_read() got nothing while we have
829 * an skb in receive queue, we do not want to loop.
830 * This might happen with URG data.
832 if (!skb_queue_empty(&sk->sk_receive_queue))
834 ret = sk_wait_data(sk, &timeo, NULL);
837 if (signal_pending(current)) {
838 ret = sock_intr_errno(timeo);
846 if (!tss.len || !timeo)
851 if (sk->sk_err || sk->sk_state == TCP_CLOSE ||
852 (sk->sk_shutdown & RCV_SHUTDOWN) ||
853 signal_pending(current))
864 EXPORT_SYMBOL(tcp_splice_read);
866 struct sk_buff *tcp_stream_alloc_skb(struct sock *sk, gfp_t gfp,
871 skb = alloc_skb_fclone(MAX_TCP_HEADER, gfp);
875 skb->truesize = SKB_TRUESIZE(skb_end_offset(skb));
876 if (force_schedule) {
877 mem_scheduled = true;
878 sk_forced_mem_schedule(sk, skb->truesize);
880 mem_scheduled = sk_wmem_schedule(sk, skb->truesize);
882 if (likely(mem_scheduled)) {
883 skb_reserve(skb, MAX_TCP_HEADER);
884 skb->ip_summed = CHECKSUM_PARTIAL;
885 INIT_LIST_HEAD(&skb->tcp_tsorted_anchor);
890 sk->sk_prot->enter_memory_pressure(sk);
891 sk_stream_moderate_sndbuf(sk);
896 static unsigned int tcp_xmit_size_goal(struct sock *sk, u32 mss_now,
899 struct tcp_sock *tp = tcp_sk(sk);
900 u32 new_size_goal, size_goal;
905 /* Note : tcp_tso_autosize() will eventually split this later */
906 new_size_goal = tcp_bound_to_half_wnd(tp, sk->sk_gso_max_size);
908 /* We try hard to avoid divides here */
909 size_goal = tp->gso_segs * mss_now;
910 if (unlikely(new_size_goal < size_goal ||
911 new_size_goal >= size_goal + mss_now)) {
912 tp->gso_segs = min_t(u16, new_size_goal / mss_now,
913 sk->sk_gso_max_segs);
914 size_goal = tp->gso_segs * mss_now;
917 return max(size_goal, mss_now);
920 int tcp_send_mss(struct sock *sk, int *size_goal, int flags)
924 mss_now = tcp_current_mss(sk);
925 *size_goal = tcp_xmit_size_goal(sk, mss_now, !(flags & MSG_OOB));
930 /* In some cases, both sendmsg() could have added an skb to the write queue,
931 * but failed adding payload on it. We need to remove it to consume less
932 * memory, but more importantly be able to generate EPOLLOUT for Edge Trigger
935 void tcp_remove_empty_skb(struct sock *sk)
937 struct sk_buff *skb = tcp_write_queue_tail(sk);
939 if (skb && TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq) {
940 tcp_unlink_write_queue(skb, sk);
941 if (tcp_write_queue_empty(sk))
942 tcp_chrono_stop(sk, TCP_CHRONO_BUSY);
943 tcp_wmem_free_skb(sk, skb);
947 /* skb changing from pure zc to mixed, must charge zc */
948 static int tcp_downgrade_zcopy_pure(struct sock *sk, struct sk_buff *skb)
950 if (unlikely(skb_zcopy_pure(skb))) {
951 u32 extra = skb->truesize -
952 SKB_TRUESIZE(skb_end_offset(skb));
954 if (!sk_wmem_schedule(sk, extra))
957 sk_mem_charge(sk, extra);
958 skb_shinfo(skb)->flags &= ~SKBFL_PURE_ZEROCOPY;
964 int tcp_wmem_schedule(struct sock *sk, int copy)
968 if (likely(sk_wmem_schedule(sk, copy)))
971 /* We could be in trouble if we have nothing queued.
972 * Use whatever is left in sk->sk_forward_alloc and tcp_wmem[0]
973 * to guarantee some progress.
975 left = sock_net(sk)->ipv4.sysctl_tcp_wmem[0] - sk->sk_wmem_queued;
977 sk_forced_mem_schedule(sk, min(left, copy));
978 return min(copy, sk->sk_forward_alloc);
981 void tcp_free_fastopen_req(struct tcp_sock *tp)
983 if (tp->fastopen_req) {
984 kfree(tp->fastopen_req);
985 tp->fastopen_req = NULL;
989 int tcp_sendmsg_fastopen(struct sock *sk, struct msghdr *msg, int *copied,
990 size_t size, struct ubuf_info *uarg)
992 struct tcp_sock *tp = tcp_sk(sk);
993 struct inet_sock *inet = inet_sk(sk);
994 struct sockaddr *uaddr = msg->msg_name;
997 if (!(READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_fastopen) &
998 TFO_CLIENT_ENABLE) ||
999 (uaddr && msg->msg_namelen >= sizeof(uaddr->sa_family) &&
1000 uaddr->sa_family == AF_UNSPEC))
1002 if (tp->fastopen_req)
1003 return -EALREADY; /* Another Fast Open is in progress */
1005 tp->fastopen_req = kzalloc(sizeof(struct tcp_fastopen_request),
1007 if (unlikely(!tp->fastopen_req))
1009 tp->fastopen_req->data = msg;
1010 tp->fastopen_req->size = size;
1011 tp->fastopen_req->uarg = uarg;
1013 if (inet_test_bit(DEFER_CONNECT, sk)) {
1014 err = tcp_connect(sk);
1015 /* Same failure procedure as in tcp_v4/6_connect */
1017 tcp_set_state(sk, TCP_CLOSE);
1018 inet->inet_dport = 0;
1019 sk->sk_route_caps = 0;
1022 flags = (msg->msg_flags & MSG_DONTWAIT) ? O_NONBLOCK : 0;
1023 err = __inet_stream_connect(sk->sk_socket, uaddr,
1024 msg->msg_namelen, flags, 1);
1025 /* fastopen_req could already be freed in __inet_stream_connect
1026 * if the connection times out or gets rst
1028 if (tp->fastopen_req) {
1029 *copied = tp->fastopen_req->copied;
1030 tcp_free_fastopen_req(tp);
1031 inet_clear_bit(DEFER_CONNECT, sk);
1036 int tcp_sendmsg_locked(struct sock *sk, struct msghdr *msg, size_t size)
1038 struct tcp_sock *tp = tcp_sk(sk);
1039 struct ubuf_info *uarg = NULL;
1040 struct sk_buff *skb;
1041 struct sockcm_cookie sockc;
1042 int flags, err, copied = 0;
1043 int mss_now = 0, size_goal, copied_syn = 0;
1044 int process_backlog = 0;
1048 flags = msg->msg_flags;
1050 if ((flags & MSG_ZEROCOPY) && size) {
1051 if (msg->msg_ubuf) {
1052 uarg = msg->msg_ubuf;
1053 if (sk->sk_route_caps & NETIF_F_SG)
1055 } else if (sock_flag(sk, SOCK_ZEROCOPY)) {
1056 skb = tcp_write_queue_tail(sk);
1057 uarg = msg_zerocopy_realloc(sk, size, skb_zcopy(skb));
1062 if (sk->sk_route_caps & NETIF_F_SG)
1065 uarg_to_msgzc(uarg)->zerocopy = 0;
1067 } else if (unlikely(msg->msg_flags & MSG_SPLICE_PAGES) && size) {
1068 if (sk->sk_route_caps & NETIF_F_SG)
1069 zc = MSG_SPLICE_PAGES;
1072 if (unlikely(flags & MSG_FASTOPEN ||
1073 inet_test_bit(DEFER_CONNECT, sk)) &&
1075 err = tcp_sendmsg_fastopen(sk, msg, &copied_syn, size, uarg);
1076 if (err == -EINPROGRESS && copied_syn > 0)
1082 timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
1084 tcp_rate_check_app_limited(sk); /* is sending application-limited? */
1086 /* Wait for a connection to finish. One exception is TCP Fast Open
1087 * (passive side) where data is allowed to be sent before a connection
1088 * is fully established.
1090 if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) &&
1091 !tcp_passive_fastopen(sk)) {
1092 err = sk_stream_wait_connect(sk, &timeo);
1097 if (unlikely(tp->repair)) {
1098 if (tp->repair_queue == TCP_RECV_QUEUE) {
1099 copied = tcp_send_rcvq(sk, msg, size);
1104 if (tp->repair_queue == TCP_NO_QUEUE)
1107 /* 'common' sending to sendq */
1110 sockcm_init(&sockc, sk);
1111 if (msg->msg_controllen) {
1112 err = sock_cmsg_send(sk, msg, &sockc);
1113 if (unlikely(err)) {
1119 /* This should be in poll */
1120 sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
1122 /* Ok commence sending. */
1126 mss_now = tcp_send_mss(sk, &size_goal, flags);
1129 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
1132 while (msg_data_left(msg)) {
1135 skb = tcp_write_queue_tail(sk);
1137 copy = size_goal - skb->len;
1139 if (copy <= 0 || !tcp_skb_can_collapse_to(skb)) {
1143 if (!sk_stream_memory_free(sk))
1144 goto wait_for_space;
1146 if (unlikely(process_backlog >= 16)) {
1147 process_backlog = 0;
1148 if (sk_flush_backlog(sk))
1151 first_skb = tcp_rtx_and_write_queues_empty(sk);
1152 skb = tcp_stream_alloc_skb(sk, sk->sk_allocation,
1155 goto wait_for_space;
1159 tcp_skb_entail(sk, skb);
1162 /* All packets are restored as if they have
1163 * already been sent. skb_mstamp_ns isn't set to
1164 * avoid wrong rtt estimation.
1167 TCP_SKB_CB(skb)->sacked |= TCPCB_REPAIRED;
1170 /* Try to append data to the end of skb. */
1171 if (copy > msg_data_left(msg))
1172 copy = msg_data_left(msg);
1176 int i = skb_shinfo(skb)->nr_frags;
1177 struct page_frag *pfrag = sk_page_frag(sk);
1179 if (!sk_page_frag_refill(sk, pfrag))
1180 goto wait_for_space;
1182 if (!skb_can_coalesce(skb, i, pfrag->page,
1184 if (i >= READ_ONCE(sysctl_max_skb_frags)) {
1185 tcp_mark_push(tp, skb);
1191 copy = min_t(int, copy, pfrag->size - pfrag->offset);
1193 if (unlikely(skb_zcopy_pure(skb) || skb_zcopy_managed(skb))) {
1194 if (tcp_downgrade_zcopy_pure(sk, skb))
1195 goto wait_for_space;
1196 skb_zcopy_downgrade_managed(skb);
1199 copy = tcp_wmem_schedule(sk, copy);
1201 goto wait_for_space;
1203 err = skb_copy_to_page_nocache(sk, &msg->msg_iter, skb,
1210 /* Update the skb. */
1212 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1214 skb_fill_page_desc(skb, i, pfrag->page,
1215 pfrag->offset, copy);
1216 page_ref_inc(pfrag->page);
1218 pfrag->offset += copy;
1219 } else if (zc == MSG_ZEROCOPY) {
1220 /* First append to a fragless skb builds initial
1224 skb_shinfo(skb)->flags |= SKBFL_PURE_ZEROCOPY;
1226 if (!skb_zcopy_pure(skb)) {
1227 copy = tcp_wmem_schedule(sk, copy);
1229 goto wait_for_space;
1232 err = skb_zerocopy_iter_stream(sk, skb, msg, copy, uarg);
1233 if (err == -EMSGSIZE || err == -EEXIST) {
1234 tcp_mark_push(tp, skb);
1240 } else if (zc == MSG_SPLICE_PAGES) {
1241 /* Splice in data if we can; copy if we can't. */
1242 if (tcp_downgrade_zcopy_pure(sk, skb))
1243 goto wait_for_space;
1244 copy = tcp_wmem_schedule(sk, copy);
1246 goto wait_for_space;
1248 err = skb_splice_from_iter(skb, &msg->msg_iter, copy,
1251 if (err == -EMSGSIZE) {
1252 tcp_mark_push(tp, skb);
1259 if (!(flags & MSG_NO_SHARED_FRAGS))
1260 skb_shinfo(skb)->flags |= SKBFL_SHARED_FRAG;
1262 sk_wmem_queued_add(sk, copy);
1263 sk_mem_charge(sk, copy);
1267 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
1269 WRITE_ONCE(tp->write_seq, tp->write_seq + copy);
1270 TCP_SKB_CB(skb)->end_seq += copy;
1271 tcp_skb_pcount_set(skb, 0);
1274 if (!msg_data_left(msg)) {
1275 if (unlikely(flags & MSG_EOR))
1276 TCP_SKB_CB(skb)->eor = 1;
1280 if (skb->len < size_goal || (flags & MSG_OOB) || unlikely(tp->repair))
1283 if (forced_push(tp)) {
1284 tcp_mark_push(tp, skb);
1285 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
1286 } else if (skb == tcp_send_head(sk))
1287 tcp_push_one(sk, mss_now);
1291 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1293 tcp_push(sk, flags & ~MSG_MORE, mss_now,
1294 TCP_NAGLE_PUSH, size_goal);
1296 err = sk_stream_wait_memory(sk, &timeo);
1300 mss_now = tcp_send_mss(sk, &size_goal, flags);
1305 tcp_tx_timestamp(sk, sockc.tsflags);
1306 tcp_push(sk, flags, mss_now, tp->nonagle, size_goal);
1309 /* msg->msg_ubuf is pinned by the caller so we don't take extra refs */
1310 if (uarg && !msg->msg_ubuf)
1311 net_zcopy_put(uarg);
1312 return copied + copied_syn;
1315 tcp_remove_empty_skb(sk);
1317 if (copied + copied_syn)
1320 /* msg->msg_ubuf is pinned by the caller so we don't take extra refs */
1321 if (uarg && !msg->msg_ubuf)
1322 net_zcopy_put_abort(uarg, true);
1323 err = sk_stream_error(sk, flags, err);
1324 /* make sure we wake any epoll edge trigger waiter */
1325 if (unlikely(tcp_rtx_and_write_queues_empty(sk) && err == -EAGAIN)) {
1326 sk->sk_write_space(sk);
1327 tcp_chrono_stop(sk, TCP_CHRONO_SNDBUF_LIMITED);
1331 EXPORT_SYMBOL_GPL(tcp_sendmsg_locked);
1333 int tcp_sendmsg(struct sock *sk, struct msghdr *msg, size_t size)
1338 ret = tcp_sendmsg_locked(sk, msg, size);
1343 EXPORT_SYMBOL(tcp_sendmsg);
1345 void tcp_splice_eof(struct socket *sock)
1347 struct sock *sk = sock->sk;
1348 struct tcp_sock *tp = tcp_sk(sk);
1349 int mss_now, size_goal;
1351 if (!tcp_write_queue_tail(sk))
1355 mss_now = tcp_send_mss(sk, &size_goal, 0);
1356 tcp_push(sk, 0, mss_now, tp->nonagle, size_goal);
1359 EXPORT_SYMBOL_GPL(tcp_splice_eof);
1362 * Handle reading urgent data. BSD has very simple semantics for
1363 * this, no blocking and very strange errors 8)
1366 static int tcp_recv_urg(struct sock *sk, struct msghdr *msg, int len, int flags)
1368 struct tcp_sock *tp = tcp_sk(sk);
1370 /* No URG data to read. */
1371 if (sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data ||
1372 tp->urg_data == TCP_URG_READ)
1373 return -EINVAL; /* Yes this is right ! */
1375 if (sk->sk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DONE))
1378 if (tp->urg_data & TCP_URG_VALID) {
1380 char c = tp->urg_data;
1382 if (!(flags & MSG_PEEK))
1383 WRITE_ONCE(tp->urg_data, TCP_URG_READ);
1385 /* Read urgent data. */
1386 msg->msg_flags |= MSG_OOB;
1389 if (!(flags & MSG_TRUNC))
1390 err = memcpy_to_msg(msg, &c, 1);
1393 msg->msg_flags |= MSG_TRUNC;
1395 return err ? -EFAULT : len;
1398 if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
1401 /* Fixed the recv(..., MSG_OOB) behaviour. BSD docs and
1402 * the available implementations agree in this case:
1403 * this call should never block, independent of the
1404 * blocking state of the socket.
1405 * Mike <pall@rz.uni-karlsruhe.de>
1410 static int tcp_peek_sndq(struct sock *sk, struct msghdr *msg, int len)
1412 struct sk_buff *skb;
1413 int copied = 0, err = 0;
1415 /* XXX -- need to support SO_PEEK_OFF */
1417 skb_rbtree_walk(skb, &sk->tcp_rtx_queue) {
1418 err = skb_copy_datagram_msg(skb, 0, msg, skb->len);
1424 skb_queue_walk(&sk->sk_write_queue, skb) {
1425 err = skb_copy_datagram_msg(skb, 0, msg, skb->len);
1432 return err ?: copied;
1435 /* Clean up the receive buffer for full frames taken by the user,
1436 * then send an ACK if necessary. COPIED is the number of bytes
1437 * tcp_recvmsg has given to the user so far, it speeds up the
1438 * calculation of whether or not we must ACK for the sake of
1441 void __tcp_cleanup_rbuf(struct sock *sk, int copied)
1443 struct tcp_sock *tp = tcp_sk(sk);
1444 bool time_to_ack = false;
1446 if (inet_csk_ack_scheduled(sk)) {
1447 const struct inet_connection_sock *icsk = inet_csk(sk);
1449 if (/* Once-per-two-segments ACK was not sent by tcp_input.c */
1450 tp->rcv_nxt - tp->rcv_wup > icsk->icsk_ack.rcv_mss ||
1452 * If this read emptied read buffer, we send ACK, if
1453 * connection is not bidirectional, user drained
1454 * receive buffer and there was a small segment
1458 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED2) ||
1459 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED) &&
1460 !inet_csk_in_pingpong_mode(sk))) &&
1461 !atomic_read(&sk->sk_rmem_alloc)))
1465 /* We send an ACK if we can now advertise a non-zero window
1466 * which has been raised "significantly".
1468 * Even if window raised up to infinity, do not send window open ACK
1469 * in states, where we will not receive more. It is useless.
1471 if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) {
1472 __u32 rcv_window_now = tcp_receive_window(tp);
1474 /* Optimize, __tcp_select_window() is not cheap. */
1475 if (2*rcv_window_now <= tp->window_clamp) {
1476 __u32 new_window = __tcp_select_window(sk);
1478 /* Send ACK now, if this read freed lots of space
1479 * in our buffer. Certainly, new_window is new window.
1480 * We can advertise it now, if it is not less than current one.
1481 * "Lots" means "at least twice" here.
1483 if (new_window && new_window >= 2 * rcv_window_now)
1491 void tcp_cleanup_rbuf(struct sock *sk, int copied)
1493 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
1494 struct tcp_sock *tp = tcp_sk(sk);
1496 WARN(skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq),
1497 "cleanup rbuf bug: copied %X seq %X rcvnxt %X\n",
1498 tp->copied_seq, TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt);
1499 __tcp_cleanup_rbuf(sk, copied);
1502 static void tcp_eat_recv_skb(struct sock *sk, struct sk_buff *skb)
1504 __skb_unlink(skb, &sk->sk_receive_queue);
1505 if (likely(skb->destructor == sock_rfree)) {
1507 skb->destructor = NULL;
1509 return skb_attempt_defer_free(skb);
1514 struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off)
1516 struct sk_buff *skb;
1519 while ((skb = skb_peek(&sk->sk_receive_queue)) != NULL) {
1520 offset = seq - TCP_SKB_CB(skb)->seq;
1521 if (unlikely(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
1522 pr_err_once("%s: found a SYN, please report !\n", __func__);
1525 if (offset < skb->len || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)) {
1529 /* This looks weird, but this can happen if TCP collapsing
1530 * splitted a fat GRO packet, while we released socket lock
1531 * in skb_splice_bits()
1533 tcp_eat_recv_skb(sk, skb);
1537 EXPORT_SYMBOL(tcp_recv_skb);
1540 * This routine provides an alternative to tcp_recvmsg() for routines
1541 * that would like to handle copying from skbuffs directly in 'sendfile'
1544 * - It is assumed that the socket was locked by the caller.
1545 * - The routine does not block.
1546 * - At present, there is no support for reading OOB data
1547 * or for 'peeking' the socket using this routine
1548 * (although both would be easy to implement).
1550 int tcp_read_sock(struct sock *sk, read_descriptor_t *desc,
1551 sk_read_actor_t recv_actor)
1553 struct sk_buff *skb;
1554 struct tcp_sock *tp = tcp_sk(sk);
1555 u32 seq = tp->copied_seq;
1559 if (sk->sk_state == TCP_LISTEN)
1561 while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1562 if (offset < skb->len) {
1566 len = skb->len - offset;
1567 /* Stop reading if we hit a patch of urgent data */
1568 if (unlikely(tp->urg_data)) {
1569 u32 urg_offset = tp->urg_seq - seq;
1570 if (urg_offset < len)
1575 used = recv_actor(desc, skb, offset, len);
1581 if (WARN_ON_ONCE(used > len))
1587 /* If recv_actor drops the lock (e.g. TCP splice
1588 * receive) the skb pointer might be invalid when
1589 * getting here: tcp_collapse might have deleted it
1590 * while aggregating skbs from the socket queue.
1592 skb = tcp_recv_skb(sk, seq - 1, &offset);
1595 /* TCP coalescing might have appended data to the skb.
1596 * Try to splice more frags
1598 if (offset + 1 != skb->len)
1601 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) {
1602 tcp_eat_recv_skb(sk, skb);
1606 tcp_eat_recv_skb(sk, skb);
1609 WRITE_ONCE(tp->copied_seq, seq);
1611 WRITE_ONCE(tp->copied_seq, seq);
1613 tcp_rcv_space_adjust(sk);
1615 /* Clean up data we have read: This will do ACK frames. */
1617 tcp_recv_skb(sk, seq, &offset);
1618 tcp_cleanup_rbuf(sk, copied);
1622 EXPORT_SYMBOL(tcp_read_sock);
1624 int tcp_read_skb(struct sock *sk, skb_read_actor_t recv_actor)
1626 struct sk_buff *skb;
1629 if (sk->sk_state == TCP_LISTEN)
1632 while ((skb = skb_peek(&sk->sk_receive_queue)) != NULL) {
1636 __skb_unlink(skb, &sk->sk_receive_queue);
1637 WARN_ON_ONCE(!skb_set_owner_sk_safe(skb, sk));
1638 tcp_flags = TCP_SKB_CB(skb)->tcp_flags;
1639 used = recv_actor(sk, skb);
1647 if (tcp_flags & TCPHDR_FIN)
1652 EXPORT_SYMBOL(tcp_read_skb);
1654 void tcp_read_done(struct sock *sk, size_t len)
1656 struct tcp_sock *tp = tcp_sk(sk);
1657 u32 seq = tp->copied_seq;
1658 struct sk_buff *skb;
1662 if (sk->sk_state == TCP_LISTEN)
1666 while (left && (skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1669 used = min_t(size_t, skb->len - offset, left);
1673 if (skb->len > offset + used)
1676 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) {
1677 tcp_eat_recv_skb(sk, skb);
1681 tcp_eat_recv_skb(sk, skb);
1683 WRITE_ONCE(tp->copied_seq, seq);
1685 tcp_rcv_space_adjust(sk);
1687 /* Clean up data we have read: This will do ACK frames. */
1689 tcp_cleanup_rbuf(sk, len - left);
1691 EXPORT_SYMBOL(tcp_read_done);
1693 int tcp_peek_len(struct socket *sock)
1695 return tcp_inq(sock->sk);
1697 EXPORT_SYMBOL(tcp_peek_len);
1699 /* Make sure sk_rcvbuf is big enough to satisfy SO_RCVLOWAT hint */
1700 int tcp_set_rcvlowat(struct sock *sk, int val)
1704 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK)
1705 cap = sk->sk_rcvbuf >> 1;
1707 cap = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_rmem[2]) >> 1;
1708 val = min(val, cap);
1709 WRITE_ONCE(sk->sk_rcvlowat, val ? : 1);
1711 /* Check if we need to signal EPOLLIN right now */
1714 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK)
1717 space = tcp_space_from_win(sk, val);
1718 if (space > sk->sk_rcvbuf) {
1719 WRITE_ONCE(sk->sk_rcvbuf, space);
1720 tcp_sk(sk)->window_clamp = val;
1724 EXPORT_SYMBOL(tcp_set_rcvlowat);
1726 void tcp_update_recv_tstamps(struct sk_buff *skb,
1727 struct scm_timestamping_internal *tss)
1730 tss->ts[0] = ktime_to_timespec64(skb->tstamp);
1732 tss->ts[0] = (struct timespec64) {0};
1734 if (skb_hwtstamps(skb)->hwtstamp)
1735 tss->ts[2] = ktime_to_timespec64(skb_hwtstamps(skb)->hwtstamp);
1737 tss->ts[2] = (struct timespec64) {0};
1741 static const struct vm_operations_struct tcp_vm_ops = {
1744 int tcp_mmap(struct file *file, struct socket *sock,
1745 struct vm_area_struct *vma)
1747 if (vma->vm_flags & (VM_WRITE | VM_EXEC))
1749 vm_flags_clear(vma, VM_MAYWRITE | VM_MAYEXEC);
1751 /* Instruct vm_insert_page() to not mmap_read_lock(mm) */
1752 vm_flags_set(vma, VM_MIXEDMAP);
1754 vma->vm_ops = &tcp_vm_ops;
1757 EXPORT_SYMBOL(tcp_mmap);
1759 static skb_frag_t *skb_advance_to_frag(struct sk_buff *skb, u32 offset_skb,
1764 if (unlikely(offset_skb >= skb->len))
1767 offset_skb -= skb_headlen(skb);
1768 if ((int)offset_skb < 0 || skb_has_frag_list(skb))
1771 frag = skb_shinfo(skb)->frags;
1772 while (offset_skb) {
1773 if (skb_frag_size(frag) > offset_skb) {
1774 *offset_frag = offset_skb;
1777 offset_skb -= skb_frag_size(frag);
1784 static bool can_map_frag(const skb_frag_t *frag)
1786 return skb_frag_size(frag) == PAGE_SIZE && !skb_frag_off(frag);
1789 static int find_next_mappable_frag(const skb_frag_t *frag,
1790 int remaining_in_skb)
1794 if (likely(can_map_frag(frag)))
1797 while (offset < remaining_in_skb && !can_map_frag(frag)) {
1798 offset += skb_frag_size(frag);
1804 static void tcp_zerocopy_set_hint_for_skb(struct sock *sk,
1805 struct tcp_zerocopy_receive *zc,
1806 struct sk_buff *skb, u32 offset)
1808 u32 frag_offset, partial_frag_remainder = 0;
1809 int mappable_offset;
1812 /* worst case: skip to next skb. try to improve on this case below */
1813 zc->recv_skip_hint = skb->len - offset;
1815 /* Find the frag containing this offset (and how far into that frag) */
1816 frag = skb_advance_to_frag(skb, offset, &frag_offset);
1821 struct skb_shared_info *info = skb_shinfo(skb);
1823 /* We read part of the last frag, must recvmsg() rest of skb. */
1824 if (frag == &info->frags[info->nr_frags - 1])
1827 /* Else, we must at least read the remainder in this frag. */
1828 partial_frag_remainder = skb_frag_size(frag) - frag_offset;
1829 zc->recv_skip_hint -= partial_frag_remainder;
1833 /* partial_frag_remainder: If part way through a frag, must read rest.
1834 * mappable_offset: Bytes till next mappable frag, *not* counting bytes
1835 * in partial_frag_remainder.
1837 mappable_offset = find_next_mappable_frag(frag, zc->recv_skip_hint);
1838 zc->recv_skip_hint = mappable_offset + partial_frag_remainder;
1841 static int tcp_recvmsg_locked(struct sock *sk, struct msghdr *msg, size_t len,
1842 int flags, struct scm_timestamping_internal *tss,
1844 static int receive_fallback_to_copy(struct sock *sk,
1845 struct tcp_zerocopy_receive *zc, int inq,
1846 struct scm_timestamping_internal *tss)
1848 unsigned long copy_address = (unsigned long)zc->copybuf_address;
1849 struct msghdr msg = {};
1854 zc->recv_skip_hint = 0;
1856 if (copy_address != zc->copybuf_address)
1859 err = import_single_range(ITER_DEST, (void __user *)copy_address,
1860 inq, &iov, &msg.msg_iter);
1864 err = tcp_recvmsg_locked(sk, &msg, inq, MSG_DONTWAIT,
1865 tss, &zc->msg_flags);
1869 zc->copybuf_len = err;
1870 if (likely(zc->copybuf_len)) {
1871 struct sk_buff *skb;
1874 skb = tcp_recv_skb(sk, tcp_sk(sk)->copied_seq, &offset);
1876 tcp_zerocopy_set_hint_for_skb(sk, zc, skb, offset);
1881 static int tcp_copy_straggler_data(struct tcp_zerocopy_receive *zc,
1882 struct sk_buff *skb, u32 copylen,
1883 u32 *offset, u32 *seq)
1885 unsigned long copy_address = (unsigned long)zc->copybuf_address;
1886 struct msghdr msg = {};
1890 if (copy_address != zc->copybuf_address)
1893 err = import_single_range(ITER_DEST, (void __user *)copy_address,
1894 copylen, &iov, &msg.msg_iter);
1897 err = skb_copy_datagram_msg(skb, *offset, &msg, copylen);
1900 zc->recv_skip_hint -= copylen;
1903 return (__s32)copylen;
1906 static int tcp_zc_handle_leftover(struct tcp_zerocopy_receive *zc,
1908 struct sk_buff *skb,
1911 struct scm_timestamping_internal *tss)
1913 u32 offset, copylen = min_t(u32, copybuf_len, zc->recv_skip_hint);
1917 /* skb is null if inq < PAGE_SIZE. */
1919 offset = *seq - TCP_SKB_CB(skb)->seq;
1921 skb = tcp_recv_skb(sk, *seq, &offset);
1922 if (TCP_SKB_CB(skb)->has_rxtstamp) {
1923 tcp_update_recv_tstamps(skb, tss);
1924 zc->msg_flags |= TCP_CMSG_TS;
1928 zc->copybuf_len = tcp_copy_straggler_data(zc, skb, copylen, &offset,
1930 return zc->copybuf_len < 0 ? 0 : copylen;
1933 static int tcp_zerocopy_vm_insert_batch_error(struct vm_area_struct *vma,
1934 struct page **pending_pages,
1935 unsigned long pages_remaining,
1936 unsigned long *address,
1939 struct tcp_zerocopy_receive *zc,
1940 u32 total_bytes_to_map,
1943 /* At least one page did not map. Try zapping if we skipped earlier. */
1944 if (err == -EBUSY &&
1945 zc->flags & TCP_RECEIVE_ZEROCOPY_FLAG_TLB_CLEAN_HINT) {
1948 maybe_zap_len = total_bytes_to_map - /* All bytes to map */
1949 *length + /* Mapped or pending */
1950 (pages_remaining * PAGE_SIZE); /* Failed map. */
1951 zap_page_range_single(vma, *address, maybe_zap_len, NULL);
1956 unsigned long leftover_pages = pages_remaining;
1959 /* We called zap_page_range_single, try to reinsert. */
1960 err = vm_insert_pages(vma, *address,
1963 bytes_mapped = PAGE_SIZE * (leftover_pages - pages_remaining);
1964 *seq += bytes_mapped;
1965 *address += bytes_mapped;
1968 /* Either we were unable to zap, OR we zapped, retried an
1969 * insert, and still had an issue. Either ways, pages_remaining
1970 * is the number of pages we were unable to map, and we unroll
1971 * some state we speculatively touched before.
1973 const int bytes_not_mapped = PAGE_SIZE * pages_remaining;
1975 *length -= bytes_not_mapped;
1976 zc->recv_skip_hint += bytes_not_mapped;
1981 static int tcp_zerocopy_vm_insert_batch(struct vm_area_struct *vma,
1982 struct page **pages,
1983 unsigned int pages_to_map,
1984 unsigned long *address,
1987 struct tcp_zerocopy_receive *zc,
1988 u32 total_bytes_to_map)
1990 unsigned long pages_remaining = pages_to_map;
1991 unsigned int pages_mapped;
1992 unsigned int bytes_mapped;
1995 err = vm_insert_pages(vma, *address, pages, &pages_remaining);
1996 pages_mapped = pages_to_map - (unsigned int)pages_remaining;
1997 bytes_mapped = PAGE_SIZE * pages_mapped;
1998 /* Even if vm_insert_pages fails, it may have partially succeeded in
1999 * mapping (some but not all of the pages).
2001 *seq += bytes_mapped;
2002 *address += bytes_mapped;
2007 /* Error: maybe zap and retry + rollback state for failed inserts. */
2008 return tcp_zerocopy_vm_insert_batch_error(vma, pages + pages_mapped,
2009 pages_remaining, address, length, seq, zc, total_bytes_to_map,
2013 #define TCP_VALID_ZC_MSG_FLAGS (TCP_CMSG_TS)
2014 static void tcp_zc_finalize_rx_tstamp(struct sock *sk,
2015 struct tcp_zerocopy_receive *zc,
2016 struct scm_timestamping_internal *tss)
2018 unsigned long msg_control_addr;
2019 struct msghdr cmsg_dummy;
2021 msg_control_addr = (unsigned long)zc->msg_control;
2022 cmsg_dummy.msg_control_user = (void __user *)msg_control_addr;
2023 cmsg_dummy.msg_controllen =
2024 (__kernel_size_t)zc->msg_controllen;
2025 cmsg_dummy.msg_flags = in_compat_syscall()
2026 ? MSG_CMSG_COMPAT : 0;
2027 cmsg_dummy.msg_control_is_user = true;
2029 if (zc->msg_control == msg_control_addr &&
2030 zc->msg_controllen == cmsg_dummy.msg_controllen) {
2031 tcp_recv_timestamp(&cmsg_dummy, sk, tss);
2032 zc->msg_control = (__u64)
2033 ((uintptr_t)cmsg_dummy.msg_control_user);
2034 zc->msg_controllen =
2035 (__u64)cmsg_dummy.msg_controllen;
2036 zc->msg_flags = (__u32)cmsg_dummy.msg_flags;
2040 static struct vm_area_struct *find_tcp_vma(struct mm_struct *mm,
2041 unsigned long address,
2044 struct vm_area_struct *vma = lock_vma_under_rcu(mm, address);
2047 if (vma->vm_ops != &tcp_vm_ops) {
2051 *mmap_locked = false;
2056 vma = vma_lookup(mm, address);
2057 if (!vma || vma->vm_ops != &tcp_vm_ops) {
2058 mmap_read_unlock(mm);
2061 *mmap_locked = true;
2065 #define TCP_ZEROCOPY_PAGE_BATCH_SIZE 32
2066 static int tcp_zerocopy_receive(struct sock *sk,
2067 struct tcp_zerocopy_receive *zc,
2068 struct scm_timestamping_internal *tss)
2070 u32 length = 0, offset, vma_len, avail_len, copylen = 0;
2071 unsigned long address = (unsigned long)zc->address;
2072 struct page *pages[TCP_ZEROCOPY_PAGE_BATCH_SIZE];
2073 s32 copybuf_len = zc->copybuf_len;
2074 struct tcp_sock *tp = tcp_sk(sk);
2075 const skb_frag_t *frags = NULL;
2076 unsigned int pages_to_map = 0;
2077 struct vm_area_struct *vma;
2078 struct sk_buff *skb = NULL;
2079 u32 seq = tp->copied_seq;
2080 u32 total_bytes_to_map;
2081 int inq = tcp_inq(sk);
2085 zc->copybuf_len = 0;
2088 if (address & (PAGE_SIZE - 1) || address != zc->address)
2091 if (sk->sk_state == TCP_LISTEN)
2094 sock_rps_record_flow(sk);
2096 if (inq && inq <= copybuf_len)
2097 return receive_fallback_to_copy(sk, zc, inq, tss);
2099 if (inq < PAGE_SIZE) {
2101 zc->recv_skip_hint = inq;
2102 if (!inq && sock_flag(sk, SOCK_DONE))
2107 vma = find_tcp_vma(current->mm, address, &mmap_locked);
2111 vma_len = min_t(unsigned long, zc->length, vma->vm_end - address);
2112 avail_len = min_t(u32, vma_len, inq);
2113 total_bytes_to_map = avail_len & ~(PAGE_SIZE - 1);
2114 if (total_bytes_to_map) {
2115 if (!(zc->flags & TCP_RECEIVE_ZEROCOPY_FLAG_TLB_CLEAN_HINT))
2116 zap_page_range_single(vma, address, total_bytes_to_map,
2118 zc->length = total_bytes_to_map;
2119 zc->recv_skip_hint = 0;
2121 zc->length = avail_len;
2122 zc->recv_skip_hint = avail_len;
2125 while (length + PAGE_SIZE <= zc->length) {
2126 int mappable_offset;
2129 if (zc->recv_skip_hint < PAGE_SIZE) {
2133 if (zc->recv_skip_hint > 0)
2136 offset = seq - TCP_SKB_CB(skb)->seq;
2138 skb = tcp_recv_skb(sk, seq, &offset);
2141 if (TCP_SKB_CB(skb)->has_rxtstamp) {
2142 tcp_update_recv_tstamps(skb, tss);
2143 zc->msg_flags |= TCP_CMSG_TS;
2145 zc->recv_skip_hint = skb->len - offset;
2146 frags = skb_advance_to_frag(skb, offset, &offset_frag);
2147 if (!frags || offset_frag)
2151 mappable_offset = find_next_mappable_frag(frags,
2152 zc->recv_skip_hint);
2153 if (mappable_offset) {
2154 zc->recv_skip_hint = mappable_offset;
2157 page = skb_frag_page(frags);
2159 pages[pages_to_map++] = page;
2160 length += PAGE_SIZE;
2161 zc->recv_skip_hint -= PAGE_SIZE;
2163 if (pages_to_map == TCP_ZEROCOPY_PAGE_BATCH_SIZE ||
2164 zc->recv_skip_hint < PAGE_SIZE) {
2165 /* Either full batch, or we're about to go to next skb
2166 * (and we cannot unroll failed ops across skbs).
2168 ret = tcp_zerocopy_vm_insert_batch(vma, pages,
2172 total_bytes_to_map);
2179 ret = tcp_zerocopy_vm_insert_batch(vma, pages, pages_to_map,
2180 &address, &length, &seq,
2181 zc, total_bytes_to_map);
2185 mmap_read_unlock(current->mm);
2188 /* Try to copy straggler data. */
2190 copylen = tcp_zc_handle_leftover(zc, sk, skb, &seq, copybuf_len, tss);
2192 if (length + copylen) {
2193 WRITE_ONCE(tp->copied_seq, seq);
2194 tcp_rcv_space_adjust(sk);
2196 /* Clean up data we have read: This will do ACK frames. */
2197 tcp_recv_skb(sk, seq, &offset);
2198 tcp_cleanup_rbuf(sk, length + copylen);
2200 if (length == zc->length)
2201 zc->recv_skip_hint = 0;
2203 if (!zc->recv_skip_hint && sock_flag(sk, SOCK_DONE))
2206 zc->length = length;
2211 /* Similar to __sock_recv_timestamp, but does not require an skb */
2212 void tcp_recv_timestamp(struct msghdr *msg, const struct sock *sk,
2213 struct scm_timestamping_internal *tss)
2215 int new_tstamp = sock_flag(sk, SOCK_TSTAMP_NEW);
2216 bool has_timestamping = false;
2218 if (tss->ts[0].tv_sec || tss->ts[0].tv_nsec) {
2219 if (sock_flag(sk, SOCK_RCVTSTAMP)) {
2220 if (sock_flag(sk, SOCK_RCVTSTAMPNS)) {
2222 struct __kernel_timespec kts = {
2223 .tv_sec = tss->ts[0].tv_sec,
2224 .tv_nsec = tss->ts[0].tv_nsec,
2226 put_cmsg(msg, SOL_SOCKET, SO_TIMESTAMPNS_NEW,
2229 struct __kernel_old_timespec ts_old = {
2230 .tv_sec = tss->ts[0].tv_sec,
2231 .tv_nsec = tss->ts[0].tv_nsec,
2233 put_cmsg(msg, SOL_SOCKET, SO_TIMESTAMPNS_OLD,
2234 sizeof(ts_old), &ts_old);
2238 struct __kernel_sock_timeval stv = {
2239 .tv_sec = tss->ts[0].tv_sec,
2240 .tv_usec = tss->ts[0].tv_nsec / 1000,
2242 put_cmsg(msg, SOL_SOCKET, SO_TIMESTAMP_NEW,
2245 struct __kernel_old_timeval tv = {
2246 .tv_sec = tss->ts[0].tv_sec,
2247 .tv_usec = tss->ts[0].tv_nsec / 1000,
2249 put_cmsg(msg, SOL_SOCKET, SO_TIMESTAMP_OLD,
2255 if (READ_ONCE(sk->sk_tsflags) & SOF_TIMESTAMPING_SOFTWARE)
2256 has_timestamping = true;
2258 tss->ts[0] = (struct timespec64) {0};
2261 if (tss->ts[2].tv_sec || tss->ts[2].tv_nsec) {
2262 if (READ_ONCE(sk->sk_tsflags) & SOF_TIMESTAMPING_RAW_HARDWARE)
2263 has_timestamping = true;
2265 tss->ts[2] = (struct timespec64) {0};
2268 if (has_timestamping) {
2269 tss->ts[1] = (struct timespec64) {0};
2270 if (sock_flag(sk, SOCK_TSTAMP_NEW))
2271 put_cmsg_scm_timestamping64(msg, tss);
2273 put_cmsg_scm_timestamping(msg, tss);
2277 static int tcp_inq_hint(struct sock *sk)
2279 const struct tcp_sock *tp = tcp_sk(sk);
2280 u32 copied_seq = READ_ONCE(tp->copied_seq);
2281 u32 rcv_nxt = READ_ONCE(tp->rcv_nxt);
2284 inq = rcv_nxt - copied_seq;
2285 if (unlikely(inq < 0 || copied_seq != READ_ONCE(tp->copied_seq))) {
2287 inq = tp->rcv_nxt - tp->copied_seq;
2290 /* After receiving a FIN, tell the user-space to continue reading
2291 * by returning a non-zero inq.
2293 if (inq == 0 && sock_flag(sk, SOCK_DONE))
2299 * This routine copies from a sock struct into the user buffer.
2301 * Technical note: in 2.3 we work on _locked_ socket, so that
2302 * tricks with *seq access order and skb->users are not required.
2303 * Probably, code can be easily improved even more.
2306 static int tcp_recvmsg_locked(struct sock *sk, struct msghdr *msg, size_t len,
2307 int flags, struct scm_timestamping_internal *tss,
2310 struct tcp_sock *tp = tcp_sk(sk);
2316 int target; /* Read at least this many bytes */
2318 struct sk_buff *skb, *last;
2322 if (sk->sk_state == TCP_LISTEN)
2325 if (tp->recvmsg_inq) {
2326 *cmsg_flags = TCP_CMSG_INQ;
2327 msg->msg_get_inq = 1;
2329 timeo = sock_rcvtimeo(sk, flags & MSG_DONTWAIT);
2331 /* Urgent data needs to be handled specially. */
2332 if (flags & MSG_OOB)
2335 if (unlikely(tp->repair)) {
2337 if (!(flags & MSG_PEEK))
2340 if (tp->repair_queue == TCP_SEND_QUEUE)
2344 if (tp->repair_queue == TCP_NO_QUEUE)
2347 /* 'common' recv queue MSG_PEEK-ing */
2350 seq = &tp->copied_seq;
2351 if (flags & MSG_PEEK) {
2352 peek_seq = tp->copied_seq;
2356 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
2361 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
2362 if (unlikely(tp->urg_data) && tp->urg_seq == *seq) {
2365 if (signal_pending(current)) {
2366 copied = timeo ? sock_intr_errno(timeo) : -EAGAIN;
2371 /* Next get a buffer. */
2373 last = skb_peek_tail(&sk->sk_receive_queue);
2374 skb_queue_walk(&sk->sk_receive_queue, skb) {
2376 /* Now that we have two receive queues this
2379 if (WARN(before(*seq, TCP_SKB_CB(skb)->seq),
2380 "TCP recvmsg seq # bug: copied %X, seq %X, rcvnxt %X, fl %X\n",
2381 *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt,
2385 offset = *seq - TCP_SKB_CB(skb)->seq;
2386 if (unlikely(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
2387 pr_err_once("%s: found a SYN, please report !\n", __func__);
2390 if (offset < skb->len)
2392 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
2394 WARN(!(flags & MSG_PEEK),
2395 "TCP recvmsg seq # bug 2: copied %X, seq %X, rcvnxt %X, fl %X\n",
2396 *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt, flags);
2399 /* Well, if we have backlog, try to process it now yet. */
2401 if (copied >= target && !READ_ONCE(sk->sk_backlog.tail))
2407 sk->sk_state == TCP_CLOSE ||
2408 (sk->sk_shutdown & RCV_SHUTDOWN) ||
2409 signal_pending(current))
2412 if (sock_flag(sk, SOCK_DONE))
2416 copied = sock_error(sk);
2420 if (sk->sk_shutdown & RCV_SHUTDOWN)
2423 if (sk->sk_state == TCP_CLOSE) {
2424 /* This occurs when user tries to read
2425 * from never connected socket.
2436 if (signal_pending(current)) {
2437 copied = sock_intr_errno(timeo);
2442 if (copied >= target) {
2443 /* Do not sleep, just process backlog. */
2444 __sk_flush_backlog(sk);
2446 tcp_cleanup_rbuf(sk, copied);
2447 err = sk_wait_data(sk, &timeo, last);
2449 err = copied ? : err;
2454 if ((flags & MSG_PEEK) &&
2455 (peek_seq - copied - urg_hole != tp->copied_seq)) {
2456 net_dbg_ratelimited("TCP(%s:%d): Application bug, race in MSG_PEEK\n",
2458 task_pid_nr(current));
2459 peek_seq = tp->copied_seq;
2464 /* Ok so how much can we use? */
2465 used = skb->len - offset;
2469 /* Do we have urgent data here? */
2470 if (unlikely(tp->urg_data)) {
2471 u32 urg_offset = tp->urg_seq - *seq;
2472 if (urg_offset < used) {
2474 if (!sock_flag(sk, SOCK_URGINLINE)) {
2475 WRITE_ONCE(*seq, *seq + 1);
2487 if (!(flags & MSG_TRUNC)) {
2488 err = skb_copy_datagram_msg(skb, offset, msg, used);
2490 /* Exception. Bailout! */
2497 WRITE_ONCE(*seq, *seq + used);
2501 tcp_rcv_space_adjust(sk);
2504 if (unlikely(tp->urg_data) && after(tp->copied_seq, tp->urg_seq)) {
2505 WRITE_ONCE(tp->urg_data, 0);
2506 tcp_fast_path_check(sk);
2509 if (TCP_SKB_CB(skb)->has_rxtstamp) {
2510 tcp_update_recv_tstamps(skb, tss);
2511 *cmsg_flags |= TCP_CMSG_TS;
2514 if (used + offset < skb->len)
2517 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
2519 if (!(flags & MSG_PEEK))
2520 tcp_eat_recv_skb(sk, skb);
2524 /* Process the FIN. */
2525 WRITE_ONCE(*seq, *seq + 1);
2526 if (!(flags & MSG_PEEK))
2527 tcp_eat_recv_skb(sk, skb);
2531 /* According to UNIX98, msg_name/msg_namelen are ignored
2532 * on connected socket. I was just happy when found this 8) --ANK
2535 /* Clean up data we have read: This will do ACK frames. */
2536 tcp_cleanup_rbuf(sk, copied);
2543 err = tcp_recv_urg(sk, msg, len, flags);
2547 err = tcp_peek_sndq(sk, msg, len);
2551 int tcp_recvmsg(struct sock *sk, struct msghdr *msg, size_t len, int flags,
2554 int cmsg_flags = 0, ret;
2555 struct scm_timestamping_internal tss;
2557 if (unlikely(flags & MSG_ERRQUEUE))
2558 return inet_recv_error(sk, msg, len, addr_len);
2560 if (sk_can_busy_loop(sk) &&
2561 skb_queue_empty_lockless(&sk->sk_receive_queue) &&
2562 sk->sk_state == TCP_ESTABLISHED)
2563 sk_busy_loop(sk, flags & MSG_DONTWAIT);
2566 ret = tcp_recvmsg_locked(sk, msg, len, flags, &tss, &cmsg_flags);
2569 if ((cmsg_flags || msg->msg_get_inq) && ret >= 0) {
2570 if (cmsg_flags & TCP_CMSG_TS)
2571 tcp_recv_timestamp(msg, sk, &tss);
2572 if (msg->msg_get_inq) {
2573 msg->msg_inq = tcp_inq_hint(sk);
2574 if (cmsg_flags & TCP_CMSG_INQ)
2575 put_cmsg(msg, SOL_TCP, TCP_CM_INQ,
2576 sizeof(msg->msg_inq), &msg->msg_inq);
2581 EXPORT_SYMBOL(tcp_recvmsg);
2583 void tcp_set_state(struct sock *sk, int state)
2585 int oldstate = sk->sk_state;
2587 /* We defined a new enum for TCP states that are exported in BPF
2588 * so as not force the internal TCP states to be frozen. The
2589 * following checks will detect if an internal state value ever
2590 * differs from the BPF value. If this ever happens, then we will
2591 * need to remap the internal value to the BPF value before calling
2592 * tcp_call_bpf_2arg.
2594 BUILD_BUG_ON((int)BPF_TCP_ESTABLISHED != (int)TCP_ESTABLISHED);
2595 BUILD_BUG_ON((int)BPF_TCP_SYN_SENT != (int)TCP_SYN_SENT);
2596 BUILD_BUG_ON((int)BPF_TCP_SYN_RECV != (int)TCP_SYN_RECV);
2597 BUILD_BUG_ON((int)BPF_TCP_FIN_WAIT1 != (int)TCP_FIN_WAIT1);
2598 BUILD_BUG_ON((int)BPF_TCP_FIN_WAIT2 != (int)TCP_FIN_WAIT2);
2599 BUILD_BUG_ON((int)BPF_TCP_TIME_WAIT != (int)TCP_TIME_WAIT);
2600 BUILD_BUG_ON((int)BPF_TCP_CLOSE != (int)TCP_CLOSE);
2601 BUILD_BUG_ON((int)BPF_TCP_CLOSE_WAIT != (int)TCP_CLOSE_WAIT);
2602 BUILD_BUG_ON((int)BPF_TCP_LAST_ACK != (int)TCP_LAST_ACK);
2603 BUILD_BUG_ON((int)BPF_TCP_LISTEN != (int)TCP_LISTEN);
2604 BUILD_BUG_ON((int)BPF_TCP_CLOSING != (int)TCP_CLOSING);
2605 BUILD_BUG_ON((int)BPF_TCP_NEW_SYN_RECV != (int)TCP_NEW_SYN_RECV);
2606 BUILD_BUG_ON((int)BPF_TCP_MAX_STATES != (int)TCP_MAX_STATES);
2608 /* bpf uapi header bpf.h defines an anonymous enum with values
2609 * BPF_TCP_* used by bpf programs. Currently gcc built vmlinux
2610 * is able to emit this enum in DWARF due to the above BUILD_BUG_ON.
2611 * But clang built vmlinux does not have this enum in DWARF
2612 * since clang removes the above code before generating IR/debuginfo.
2613 * Let us explicitly emit the type debuginfo to ensure the
2614 * above-mentioned anonymous enum in the vmlinux DWARF and hence BTF
2615 * regardless of which compiler is used.
2617 BTF_TYPE_EMIT_ENUM(BPF_TCP_ESTABLISHED);
2619 if (BPF_SOCK_OPS_TEST_FLAG(tcp_sk(sk), BPF_SOCK_OPS_STATE_CB_FLAG))
2620 tcp_call_bpf_2arg(sk, BPF_SOCK_OPS_STATE_CB, oldstate, state);
2623 case TCP_ESTABLISHED:
2624 if (oldstate != TCP_ESTABLISHED)
2625 TCP_INC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
2629 if (oldstate == TCP_CLOSE_WAIT || oldstate == TCP_ESTABLISHED)
2630 TCP_INC_STATS(sock_net(sk), TCP_MIB_ESTABRESETS);
2632 sk->sk_prot->unhash(sk);
2633 if (inet_csk(sk)->icsk_bind_hash &&
2634 !(sk->sk_userlocks & SOCK_BINDPORT_LOCK))
2638 if (oldstate == TCP_ESTABLISHED)
2639 TCP_DEC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
2642 /* Change state AFTER socket is unhashed to avoid closed
2643 * socket sitting in hash tables.
2645 inet_sk_state_store(sk, state);
2647 EXPORT_SYMBOL_GPL(tcp_set_state);
2650 * State processing on a close. This implements the state shift for
2651 * sending our FIN frame. Note that we only send a FIN for some
2652 * states. A shutdown() may have already sent the FIN, or we may be
2656 static const unsigned char new_state[16] = {
2657 /* current state: new state: action: */
2658 [0 /* (Invalid) */] = TCP_CLOSE,
2659 [TCP_ESTABLISHED] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2660 [TCP_SYN_SENT] = TCP_CLOSE,
2661 [TCP_SYN_RECV] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2662 [TCP_FIN_WAIT1] = TCP_FIN_WAIT1,
2663 [TCP_FIN_WAIT2] = TCP_FIN_WAIT2,
2664 [TCP_TIME_WAIT] = TCP_CLOSE,
2665 [TCP_CLOSE] = TCP_CLOSE,
2666 [TCP_CLOSE_WAIT] = TCP_LAST_ACK | TCP_ACTION_FIN,
2667 [TCP_LAST_ACK] = TCP_LAST_ACK,
2668 [TCP_LISTEN] = TCP_CLOSE,
2669 [TCP_CLOSING] = TCP_CLOSING,
2670 [TCP_NEW_SYN_RECV] = TCP_CLOSE, /* should not happen ! */
2673 static int tcp_close_state(struct sock *sk)
2675 int next = (int)new_state[sk->sk_state];
2676 int ns = next & TCP_STATE_MASK;
2678 tcp_set_state(sk, ns);
2680 return next & TCP_ACTION_FIN;
2684 * Shutdown the sending side of a connection. Much like close except
2685 * that we don't receive shut down or sock_set_flag(sk, SOCK_DEAD).
2688 void tcp_shutdown(struct sock *sk, int how)
2690 /* We need to grab some memory, and put together a FIN,
2691 * and then put it into the queue to be sent.
2692 * Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
2694 if (!(how & SEND_SHUTDOWN))
2697 /* If we've already sent a FIN, or it's a closed state, skip this. */
2698 if ((1 << sk->sk_state) &
2699 (TCPF_ESTABLISHED | TCPF_SYN_SENT |
2700 TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) {
2701 /* Clear out any half completed packets. FIN if needed. */
2702 if (tcp_close_state(sk))
2706 EXPORT_SYMBOL(tcp_shutdown);
2708 int tcp_orphan_count_sum(void)
2712 for_each_possible_cpu(i)
2713 total += per_cpu(tcp_orphan_count, i);
2715 return max(total, 0);
2718 static int tcp_orphan_cache;
2719 static struct timer_list tcp_orphan_timer;
2720 #define TCP_ORPHAN_TIMER_PERIOD msecs_to_jiffies(100)
2722 static void tcp_orphan_update(struct timer_list *unused)
2724 WRITE_ONCE(tcp_orphan_cache, tcp_orphan_count_sum());
2725 mod_timer(&tcp_orphan_timer, jiffies + TCP_ORPHAN_TIMER_PERIOD);
2728 static bool tcp_too_many_orphans(int shift)
2730 return READ_ONCE(tcp_orphan_cache) << shift >
2731 READ_ONCE(sysctl_tcp_max_orphans);
2734 bool tcp_check_oom(struct sock *sk, int shift)
2736 bool too_many_orphans, out_of_socket_memory;
2738 too_many_orphans = tcp_too_many_orphans(shift);
2739 out_of_socket_memory = tcp_out_of_memory(sk);
2741 if (too_many_orphans)
2742 net_info_ratelimited("too many orphaned sockets\n");
2743 if (out_of_socket_memory)
2744 net_info_ratelimited("out of memory -- consider tuning tcp_mem\n");
2745 return too_many_orphans || out_of_socket_memory;
2748 void __tcp_close(struct sock *sk, long timeout)
2750 struct sk_buff *skb;
2751 int data_was_unread = 0;
2754 WRITE_ONCE(sk->sk_shutdown, SHUTDOWN_MASK);
2756 if (sk->sk_state == TCP_LISTEN) {
2757 tcp_set_state(sk, TCP_CLOSE);
2760 inet_csk_listen_stop(sk);
2762 goto adjudge_to_death;
2765 /* We need to flush the recv. buffs. We do this only on the
2766 * descriptor close, not protocol-sourced closes, because the
2767 * reader process may not have drained the data yet!
2769 while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) {
2770 u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq;
2772 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
2774 data_was_unread += len;
2778 /* If socket has been already reset (e.g. in tcp_reset()) - kill it. */
2779 if (sk->sk_state == TCP_CLOSE)
2780 goto adjudge_to_death;
2782 /* As outlined in RFC 2525, section 2.17, we send a RST here because
2783 * data was lost. To witness the awful effects of the old behavior of
2784 * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk
2785 * GET in an FTP client, suspend the process, wait for the client to
2786 * advertise a zero window, then kill -9 the FTP client, wheee...
2787 * Note: timeout is always zero in such a case.
2789 if (unlikely(tcp_sk(sk)->repair)) {
2790 sk->sk_prot->disconnect(sk, 0);
2791 } else if (data_was_unread) {
2792 /* Unread data was tossed, zap the connection. */
2793 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONCLOSE);
2794 tcp_set_state(sk, TCP_CLOSE);
2795 tcp_send_active_reset(sk, sk->sk_allocation);
2796 } else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) {
2797 /* Check zero linger _after_ checking for unread data. */
2798 sk->sk_prot->disconnect(sk, 0);
2799 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
2800 } else if (tcp_close_state(sk)) {
2801 /* We FIN if the application ate all the data before
2802 * zapping the connection.
2805 /* RED-PEN. Formally speaking, we have broken TCP state
2806 * machine. State transitions:
2808 * TCP_ESTABLISHED -> TCP_FIN_WAIT1
2809 * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible)
2810 * TCP_CLOSE_WAIT -> TCP_LAST_ACK
2812 * are legal only when FIN has been sent (i.e. in window),
2813 * rather than queued out of window. Purists blame.
2815 * F.e. "RFC state" is ESTABLISHED,
2816 * if Linux state is FIN-WAIT-1, but FIN is still not sent.
2818 * The visible declinations are that sometimes
2819 * we enter time-wait state, when it is not required really
2820 * (harmless), do not send active resets, when they are
2821 * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
2822 * they look as CLOSING or LAST_ACK for Linux)
2823 * Probably, I missed some more holelets.
2825 * XXX (TFO) - To start off we don't support SYN+ACK+FIN
2826 * in a single packet! (May consider it later but will
2827 * probably need API support or TCP_CORK SYN-ACK until
2828 * data is written and socket is closed.)
2833 sk_stream_wait_close(sk, timeout);
2836 state = sk->sk_state;
2842 /* remove backlog if any, without releasing ownership. */
2845 this_cpu_inc(tcp_orphan_count);
2847 /* Have we already been destroyed by a softirq or backlog? */
2848 if (state != TCP_CLOSE && sk->sk_state == TCP_CLOSE)
2851 /* This is a (useful) BSD violating of the RFC. There is a
2852 * problem with TCP as specified in that the other end could
2853 * keep a socket open forever with no application left this end.
2854 * We use a 1 minute timeout (about the same as BSD) then kill
2855 * our end. If they send after that then tough - BUT: long enough
2856 * that we won't make the old 4*rto = almost no time - whoops
2859 * Nope, it was not mistake. It is really desired behaviour
2860 * f.e. on http servers, when such sockets are useless, but
2861 * consume significant resources. Let's do it with special
2862 * linger2 option. --ANK
2865 if (sk->sk_state == TCP_FIN_WAIT2) {
2866 struct tcp_sock *tp = tcp_sk(sk);
2867 if (READ_ONCE(tp->linger2) < 0) {
2868 tcp_set_state(sk, TCP_CLOSE);
2869 tcp_send_active_reset(sk, GFP_ATOMIC);
2870 __NET_INC_STATS(sock_net(sk),
2871 LINUX_MIB_TCPABORTONLINGER);
2873 const int tmo = tcp_fin_time(sk);
2875 if (tmo > TCP_TIMEWAIT_LEN) {
2876 inet_csk_reset_keepalive_timer(sk,
2877 tmo - TCP_TIMEWAIT_LEN);
2879 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
2884 if (sk->sk_state != TCP_CLOSE) {
2885 if (tcp_check_oom(sk, 0)) {
2886 tcp_set_state(sk, TCP_CLOSE);
2887 tcp_send_active_reset(sk, GFP_ATOMIC);
2888 __NET_INC_STATS(sock_net(sk),
2889 LINUX_MIB_TCPABORTONMEMORY);
2890 } else if (!check_net(sock_net(sk))) {
2891 /* Not possible to send reset; just close */
2892 tcp_set_state(sk, TCP_CLOSE);
2896 if (sk->sk_state == TCP_CLOSE) {
2897 struct request_sock *req;
2899 req = rcu_dereference_protected(tcp_sk(sk)->fastopen_rsk,
2900 lockdep_sock_is_held(sk));
2901 /* We could get here with a non-NULL req if the socket is
2902 * aborted (e.g., closed with unread data) before 3WHS
2906 reqsk_fastopen_remove(sk, req, false);
2907 inet_csk_destroy_sock(sk);
2909 /* Otherwise, socket is reprieved until protocol close. */
2916 void tcp_close(struct sock *sk, long timeout)
2919 __tcp_close(sk, timeout);
2923 EXPORT_SYMBOL(tcp_close);
2925 /* These states need RST on ABORT according to RFC793 */
2927 static inline bool tcp_need_reset(int state)
2929 return (1 << state) &
2930 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 |
2931 TCPF_FIN_WAIT2 | TCPF_SYN_RECV);
2934 static void tcp_rtx_queue_purge(struct sock *sk)
2936 struct rb_node *p = rb_first(&sk->tcp_rtx_queue);
2938 tcp_sk(sk)->highest_sack = NULL;
2940 struct sk_buff *skb = rb_to_skb(p);
2943 /* Since we are deleting whole queue, no need to
2944 * list_del(&skb->tcp_tsorted_anchor)
2946 tcp_rtx_queue_unlink(skb, sk);
2947 tcp_wmem_free_skb(sk, skb);
2951 void tcp_write_queue_purge(struct sock *sk)
2953 struct sk_buff *skb;
2955 tcp_chrono_stop(sk, TCP_CHRONO_BUSY);
2956 while ((skb = __skb_dequeue(&sk->sk_write_queue)) != NULL) {
2957 tcp_skb_tsorted_anchor_cleanup(skb);
2958 tcp_wmem_free_skb(sk, skb);
2960 tcp_rtx_queue_purge(sk);
2961 INIT_LIST_HEAD(&tcp_sk(sk)->tsorted_sent_queue);
2962 tcp_clear_all_retrans_hints(tcp_sk(sk));
2963 tcp_sk(sk)->packets_out = 0;
2964 inet_csk(sk)->icsk_backoff = 0;
2967 int tcp_disconnect(struct sock *sk, int flags)
2969 struct inet_sock *inet = inet_sk(sk);
2970 struct inet_connection_sock *icsk = inet_csk(sk);
2971 struct tcp_sock *tp = tcp_sk(sk);
2972 int old_state = sk->sk_state;
2975 if (old_state != TCP_CLOSE)
2976 tcp_set_state(sk, TCP_CLOSE);
2978 /* ABORT function of RFC793 */
2979 if (old_state == TCP_LISTEN) {
2980 inet_csk_listen_stop(sk);
2981 } else if (unlikely(tp->repair)) {
2982 WRITE_ONCE(sk->sk_err, ECONNABORTED);
2983 } else if (tcp_need_reset(old_state) ||
2984 (tp->snd_nxt != tp->write_seq &&
2985 (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK))) {
2986 /* The last check adjusts for discrepancy of Linux wrt. RFC
2989 tcp_send_active_reset(sk, gfp_any());
2990 WRITE_ONCE(sk->sk_err, ECONNRESET);
2991 } else if (old_state == TCP_SYN_SENT)
2992 WRITE_ONCE(sk->sk_err, ECONNRESET);
2994 tcp_clear_xmit_timers(sk);
2995 __skb_queue_purge(&sk->sk_receive_queue);
2996 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt);
2997 WRITE_ONCE(tp->urg_data, 0);
2998 tcp_write_queue_purge(sk);
2999 tcp_fastopen_active_disable_ofo_check(sk);
3000 skb_rbtree_purge(&tp->out_of_order_queue);
3002 inet->inet_dport = 0;
3004 inet_bhash2_reset_saddr(sk);
3006 WRITE_ONCE(sk->sk_shutdown, 0);
3007 sock_reset_flag(sk, SOCK_DONE);
3009 tp->mdev_us = jiffies_to_usecs(TCP_TIMEOUT_INIT);
3010 tp->rcv_rtt_last_tsecr = 0;
3012 seq = tp->write_seq + tp->max_window + 2;
3015 WRITE_ONCE(tp->write_seq, seq);
3017 icsk->icsk_backoff = 0;
3018 icsk->icsk_probes_out = 0;
3019 icsk->icsk_probes_tstamp = 0;
3020 icsk->icsk_rto = TCP_TIMEOUT_INIT;
3021 icsk->icsk_rto_min = TCP_RTO_MIN;
3022 icsk->icsk_delack_max = TCP_DELACK_MAX;
3023 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
3024 tcp_snd_cwnd_set(tp, TCP_INIT_CWND);
3025 tp->snd_cwnd_cnt = 0;
3026 tp->is_cwnd_limited = 0;
3027 tp->max_packets_out = 0;
3028 tp->window_clamp = 0;
3030 tp->delivered_ce = 0;
3031 if (icsk->icsk_ca_ops->release)
3032 icsk->icsk_ca_ops->release(sk);
3033 memset(icsk->icsk_ca_priv, 0, sizeof(icsk->icsk_ca_priv));
3034 icsk->icsk_ca_initialized = 0;
3035 tcp_set_ca_state(sk, TCP_CA_Open);
3036 tp->is_sack_reneg = 0;
3037 tcp_clear_retrans(tp);
3038 tp->total_retrans = 0;
3039 inet_csk_delack_init(sk);
3040 /* Initialize rcv_mss to TCP_MIN_MSS to avoid division by 0
3041 * issue in __tcp_select_window()
3043 icsk->icsk_ack.rcv_mss = TCP_MIN_MSS;
3044 memset(&tp->rx_opt, 0, sizeof(tp->rx_opt));
3046 dst_release(xchg((__force struct dst_entry **)&sk->sk_rx_dst, NULL));
3047 tcp_saved_syn_free(tp);
3048 tp->compressed_ack = 0;
3052 tp->bytes_acked = 0;
3053 tp->bytes_received = 0;
3054 tp->bytes_retrans = 0;
3055 tp->data_segs_in = 0;
3056 tp->data_segs_out = 0;
3057 tp->duplicate_sack[0].start_seq = 0;
3058 tp->duplicate_sack[0].end_seq = 0;
3061 tp->retrans_out = 0;
3063 tp->tlp_high_seq = 0;
3064 tp->last_oow_ack_time = 0;
3066 /* There's a bubble in the pipe until at least the first ACK. */
3067 tp->app_limited = ~0U;
3068 tp->rate_app_limited = 1;
3069 tp->rack.mstamp = 0;
3070 tp->rack.advanced = 0;
3071 tp->rack.reo_wnd_steps = 1;
3072 tp->rack.last_delivered = 0;
3073 tp->rack.reo_wnd_persist = 0;
3074 tp->rack.dsack_seen = 0;
3075 tp->syn_data_acked = 0;
3076 tp->rx_opt.saw_tstamp = 0;
3077 tp->rx_opt.dsack = 0;
3078 tp->rx_opt.num_sacks = 0;
3079 tp->rcv_ooopack = 0;
3082 /* Clean up fastopen related fields */
3083 tcp_free_fastopen_req(tp);
3084 inet_clear_bit(DEFER_CONNECT, sk);
3085 tp->fastopen_client_fail = 0;
3087 WARN_ON(inet->inet_num && !icsk->icsk_bind_hash);
3089 if (sk->sk_frag.page) {
3090 put_page(sk->sk_frag.page);
3091 sk->sk_frag.page = NULL;
3092 sk->sk_frag.offset = 0;
3094 sk_error_report(sk);
3097 EXPORT_SYMBOL(tcp_disconnect);
3099 static inline bool tcp_can_repair_sock(const struct sock *sk)
3101 return sockopt_ns_capable(sock_net(sk)->user_ns, CAP_NET_ADMIN) &&
3102 (sk->sk_state != TCP_LISTEN);
3105 static int tcp_repair_set_window(struct tcp_sock *tp, sockptr_t optbuf, int len)
3107 struct tcp_repair_window opt;
3112 if (len != sizeof(opt))
3115 if (copy_from_sockptr(&opt, optbuf, sizeof(opt)))
3118 if (opt.max_window < opt.snd_wnd)
3121 if (after(opt.snd_wl1, tp->rcv_nxt + opt.rcv_wnd))
3124 if (after(opt.rcv_wup, tp->rcv_nxt))
3127 tp->snd_wl1 = opt.snd_wl1;
3128 tp->snd_wnd = opt.snd_wnd;
3129 tp->max_window = opt.max_window;
3131 tp->rcv_wnd = opt.rcv_wnd;
3132 tp->rcv_wup = opt.rcv_wup;
3137 static int tcp_repair_options_est(struct sock *sk, sockptr_t optbuf,
3140 struct tcp_sock *tp = tcp_sk(sk);
3141 struct tcp_repair_opt opt;
3144 while (len >= sizeof(opt)) {
3145 if (copy_from_sockptr_offset(&opt, optbuf, offset, sizeof(opt)))
3148 offset += sizeof(opt);
3151 switch (opt.opt_code) {
3153 tp->rx_opt.mss_clamp = opt.opt_val;
3158 u16 snd_wscale = opt.opt_val & 0xFFFF;
3159 u16 rcv_wscale = opt.opt_val >> 16;
3161 if (snd_wscale > TCP_MAX_WSCALE || rcv_wscale > TCP_MAX_WSCALE)
3164 tp->rx_opt.snd_wscale = snd_wscale;
3165 tp->rx_opt.rcv_wscale = rcv_wscale;
3166 tp->rx_opt.wscale_ok = 1;
3169 case TCPOPT_SACK_PERM:
3170 if (opt.opt_val != 0)
3173 tp->rx_opt.sack_ok |= TCP_SACK_SEEN;
3175 case TCPOPT_TIMESTAMP:
3176 if (opt.opt_val != 0)
3179 tp->rx_opt.tstamp_ok = 1;
3187 DEFINE_STATIC_KEY_FALSE(tcp_tx_delay_enabled);
3188 EXPORT_SYMBOL(tcp_tx_delay_enabled);
3190 static void tcp_enable_tx_delay(void)
3192 if (!static_branch_unlikely(&tcp_tx_delay_enabled)) {
3193 static int __tcp_tx_delay_enabled = 0;
3195 if (cmpxchg(&__tcp_tx_delay_enabled, 0, 1) == 0) {
3196 static_branch_enable(&tcp_tx_delay_enabled);
3197 pr_info("TCP_TX_DELAY enabled\n");
3202 /* When set indicates to always queue non-full frames. Later the user clears
3203 * this option and we transmit any pending partial frames in the queue. This is
3204 * meant to be used alongside sendfile() to get properly filled frames when the
3205 * user (for example) must write out headers with a write() call first and then
3206 * use sendfile to send out the data parts.
3208 * TCP_CORK can be set together with TCP_NODELAY and it is stronger than
3211 void __tcp_sock_set_cork(struct sock *sk, bool on)
3213 struct tcp_sock *tp = tcp_sk(sk);
3216 tp->nonagle |= TCP_NAGLE_CORK;
3218 tp->nonagle &= ~TCP_NAGLE_CORK;
3219 if (tp->nonagle & TCP_NAGLE_OFF)
3220 tp->nonagle |= TCP_NAGLE_PUSH;
3221 tcp_push_pending_frames(sk);
3225 void tcp_sock_set_cork(struct sock *sk, bool on)
3228 __tcp_sock_set_cork(sk, on);
3231 EXPORT_SYMBOL(tcp_sock_set_cork);
3233 /* TCP_NODELAY is weaker than TCP_CORK, so that this option on corked socket is
3234 * remembered, but it is not activated until cork is cleared.
3236 * However, when TCP_NODELAY is set we make an explicit push, which overrides
3237 * even TCP_CORK for currently queued segments.
3239 void __tcp_sock_set_nodelay(struct sock *sk, bool on)
3242 tcp_sk(sk)->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH;
3243 tcp_push_pending_frames(sk);
3245 tcp_sk(sk)->nonagle &= ~TCP_NAGLE_OFF;
3249 void tcp_sock_set_nodelay(struct sock *sk)
3252 __tcp_sock_set_nodelay(sk, true);
3255 EXPORT_SYMBOL(tcp_sock_set_nodelay);
3257 static void __tcp_sock_set_quickack(struct sock *sk, int val)
3260 inet_csk_enter_pingpong_mode(sk);
3264 inet_csk_exit_pingpong_mode(sk);
3265 if ((1 << sk->sk_state) & (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) &&
3266 inet_csk_ack_scheduled(sk)) {
3267 inet_csk(sk)->icsk_ack.pending |= ICSK_ACK_PUSHED;
3268 tcp_cleanup_rbuf(sk, 1);
3270 inet_csk_enter_pingpong_mode(sk);
3274 void tcp_sock_set_quickack(struct sock *sk, int val)
3277 __tcp_sock_set_quickack(sk, val);
3280 EXPORT_SYMBOL(tcp_sock_set_quickack);
3282 int tcp_sock_set_syncnt(struct sock *sk, int val)
3284 if (val < 1 || val > MAX_TCP_SYNCNT)
3287 WRITE_ONCE(inet_csk(sk)->icsk_syn_retries, val);
3290 EXPORT_SYMBOL(tcp_sock_set_syncnt);
3292 int tcp_sock_set_user_timeout(struct sock *sk, int val)
3294 /* Cap the max time in ms TCP will retry or probe the window
3295 * before giving up and aborting (ETIMEDOUT) a connection.
3300 WRITE_ONCE(inet_csk(sk)->icsk_user_timeout, val);
3303 EXPORT_SYMBOL(tcp_sock_set_user_timeout);
3305 int tcp_sock_set_keepidle_locked(struct sock *sk, int val)
3307 struct tcp_sock *tp = tcp_sk(sk);
3309 if (val < 1 || val > MAX_TCP_KEEPIDLE)
3312 /* Paired with WRITE_ONCE() in keepalive_time_when() */
3313 WRITE_ONCE(tp->keepalive_time, val * HZ);
3314 if (sock_flag(sk, SOCK_KEEPOPEN) &&
3315 !((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN))) {
3316 u32 elapsed = keepalive_time_elapsed(tp);
3318 if (tp->keepalive_time > elapsed)
3319 elapsed = tp->keepalive_time - elapsed;
3322 inet_csk_reset_keepalive_timer(sk, elapsed);
3328 int tcp_sock_set_keepidle(struct sock *sk, int val)
3333 err = tcp_sock_set_keepidle_locked(sk, val);
3337 EXPORT_SYMBOL(tcp_sock_set_keepidle);
3339 int tcp_sock_set_keepintvl(struct sock *sk, int val)
3341 if (val < 1 || val > MAX_TCP_KEEPINTVL)
3344 WRITE_ONCE(tcp_sk(sk)->keepalive_intvl, val * HZ);
3347 EXPORT_SYMBOL(tcp_sock_set_keepintvl);
3349 int tcp_sock_set_keepcnt(struct sock *sk, int val)
3351 if (val < 1 || val > MAX_TCP_KEEPCNT)
3354 /* Paired with READ_ONCE() in keepalive_probes() */
3355 WRITE_ONCE(tcp_sk(sk)->keepalive_probes, val);
3358 EXPORT_SYMBOL(tcp_sock_set_keepcnt);
3360 int tcp_set_window_clamp(struct sock *sk, int val)
3362 struct tcp_sock *tp = tcp_sk(sk);
3365 if (sk->sk_state != TCP_CLOSE)
3367 tp->window_clamp = 0;
3369 tp->window_clamp = val < SOCK_MIN_RCVBUF / 2 ?
3370 SOCK_MIN_RCVBUF / 2 : val;
3371 tp->rcv_ssthresh = min(tp->rcv_wnd, tp->window_clamp);
3377 * Socket option code for TCP.
3379 int do_tcp_setsockopt(struct sock *sk, int level, int optname,
3380 sockptr_t optval, unsigned int optlen)
3382 struct tcp_sock *tp = tcp_sk(sk);
3383 struct inet_connection_sock *icsk = inet_csk(sk);
3384 struct net *net = sock_net(sk);
3388 /* These are data/string values, all the others are ints */
3390 case TCP_CONGESTION: {
3391 char name[TCP_CA_NAME_MAX];
3396 val = strncpy_from_sockptr(name, optval,
3397 min_t(long, TCP_CA_NAME_MAX-1, optlen));
3402 sockopt_lock_sock(sk);
3403 err = tcp_set_congestion_control(sk, name, !has_current_bpf_ctx(),
3404 sockopt_ns_capable(sock_net(sk)->user_ns,
3406 sockopt_release_sock(sk);
3410 char name[TCP_ULP_NAME_MAX];
3415 val = strncpy_from_sockptr(name, optval,
3416 min_t(long, TCP_ULP_NAME_MAX - 1,
3422 sockopt_lock_sock(sk);
3423 err = tcp_set_ulp(sk, name);
3424 sockopt_release_sock(sk);
3427 case TCP_FASTOPEN_KEY: {
3428 __u8 key[TCP_FASTOPEN_KEY_BUF_LENGTH];
3429 __u8 *backup_key = NULL;
3431 /* Allow a backup key as well to facilitate key rotation
3432 * First key is the active one.
3434 if (optlen != TCP_FASTOPEN_KEY_LENGTH &&
3435 optlen != TCP_FASTOPEN_KEY_BUF_LENGTH)
3438 if (copy_from_sockptr(key, optval, optlen))
3441 if (optlen == TCP_FASTOPEN_KEY_BUF_LENGTH)
3442 backup_key = key + TCP_FASTOPEN_KEY_LENGTH;
3444 return tcp_fastopen_reset_cipher(net, sk, key, backup_key);
3451 if (optlen < sizeof(int))
3454 if (copy_from_sockptr(&val, optval, sizeof(val)))
3457 /* Handle options that can be set without locking the socket. */
3460 return tcp_sock_set_syncnt(sk, val);
3461 case TCP_USER_TIMEOUT:
3462 return tcp_sock_set_user_timeout(sk, val);
3464 return tcp_sock_set_keepintvl(sk, val);
3466 return tcp_sock_set_keepcnt(sk, val);
3469 WRITE_ONCE(tp->linger2, -1);
3470 else if (val > TCP_FIN_TIMEOUT_MAX / HZ)
3471 WRITE_ONCE(tp->linger2, TCP_FIN_TIMEOUT_MAX);
3473 WRITE_ONCE(tp->linger2, val * HZ);
3475 case TCP_DEFER_ACCEPT:
3476 /* Translate value in seconds to number of retransmits */
3477 WRITE_ONCE(icsk->icsk_accept_queue.rskq_defer_accept,
3478 secs_to_retrans(val, TCP_TIMEOUT_INIT / HZ,
3483 sockopt_lock_sock(sk);
3487 /* Values greater than interface MTU won't take effect. However
3488 * at the point when this call is done we typically don't yet
3489 * know which interface is going to be used
3491 if (val && (val < TCP_MIN_MSS || val > MAX_TCP_WINDOW)) {
3495 tp->rx_opt.user_mss = val;
3499 __tcp_sock_set_nodelay(sk, val);
3502 case TCP_THIN_LINEAR_TIMEOUTS:
3503 if (val < 0 || val > 1)
3509 case TCP_THIN_DUPACK:
3510 if (val < 0 || val > 1)
3515 if (!tcp_can_repair_sock(sk))
3517 else if (val == TCP_REPAIR_ON) {
3519 sk->sk_reuse = SK_FORCE_REUSE;
3520 tp->repair_queue = TCP_NO_QUEUE;
3521 } else if (val == TCP_REPAIR_OFF) {
3523 sk->sk_reuse = SK_NO_REUSE;
3524 tcp_send_window_probe(sk);
3525 } else if (val == TCP_REPAIR_OFF_NO_WP) {
3527 sk->sk_reuse = SK_NO_REUSE;
3533 case TCP_REPAIR_QUEUE:
3536 else if ((unsigned int)val < TCP_QUEUES_NR)
3537 tp->repair_queue = val;
3543 if (sk->sk_state != TCP_CLOSE) {
3545 } else if (tp->repair_queue == TCP_SEND_QUEUE) {
3546 if (!tcp_rtx_queue_empty(sk))
3549 WRITE_ONCE(tp->write_seq, val);
3550 } else if (tp->repair_queue == TCP_RECV_QUEUE) {
3551 if (tp->rcv_nxt != tp->copied_seq) {
3554 WRITE_ONCE(tp->rcv_nxt, val);
3555 WRITE_ONCE(tp->copied_seq, val);
3562 case TCP_REPAIR_OPTIONS:
3565 else if (sk->sk_state == TCP_ESTABLISHED && !tp->bytes_sent)
3566 err = tcp_repair_options_est(sk, optval, optlen);
3572 __tcp_sock_set_cork(sk, val);
3576 err = tcp_sock_set_keepidle_locked(sk, val);
3579 /* 0: disable, 1: enable, 2: start from ether_header */
3580 if (val < 0 || val > 2)
3586 case TCP_WINDOW_CLAMP:
3587 err = tcp_set_window_clamp(sk, val);
3591 __tcp_sock_set_quickack(sk, val);
3594 #ifdef CONFIG_TCP_MD5SIG
3596 case TCP_MD5SIG_EXT:
3597 err = tp->af_specific->md5_parse(sk, optname, optval, optlen);
3601 if (val >= 0 && ((1 << sk->sk_state) & (TCPF_CLOSE |
3603 tcp_fastopen_init_key_once(net);
3605 fastopen_queue_tune(sk, val);
3610 case TCP_FASTOPEN_CONNECT:
3611 if (val > 1 || val < 0) {
3613 } else if (READ_ONCE(net->ipv4.sysctl_tcp_fastopen) &
3614 TFO_CLIENT_ENABLE) {
3615 if (sk->sk_state == TCP_CLOSE)
3616 tp->fastopen_connect = val;
3623 case TCP_FASTOPEN_NO_COOKIE:
3624 if (val > 1 || val < 0)
3626 else if (!((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
3629 tp->fastopen_no_cookie = val;
3635 WRITE_ONCE(tp->tsoffset, val - tcp_time_stamp_raw());
3637 case TCP_REPAIR_WINDOW:
3638 err = tcp_repair_set_window(tp, optval, optlen);
3640 case TCP_NOTSENT_LOWAT:
3641 WRITE_ONCE(tp->notsent_lowat, val);
3642 sk->sk_write_space(sk);
3645 if (val > 1 || val < 0)
3648 tp->recvmsg_inq = val;
3652 tcp_enable_tx_delay();
3653 WRITE_ONCE(tp->tcp_tx_delay, val);
3660 sockopt_release_sock(sk);
3664 int tcp_setsockopt(struct sock *sk, int level, int optname, sockptr_t optval,
3665 unsigned int optlen)
3667 const struct inet_connection_sock *icsk = inet_csk(sk);
3669 if (level != SOL_TCP)
3670 /* Paired with WRITE_ONCE() in do_ipv6_setsockopt() and tcp_v6_connect() */
3671 return READ_ONCE(icsk->icsk_af_ops)->setsockopt(sk, level, optname,
3673 return do_tcp_setsockopt(sk, level, optname, optval, optlen);
3675 EXPORT_SYMBOL(tcp_setsockopt);
3677 static void tcp_get_info_chrono_stats(const struct tcp_sock *tp,
3678 struct tcp_info *info)
3680 u64 stats[__TCP_CHRONO_MAX], total = 0;
3683 for (i = TCP_CHRONO_BUSY; i < __TCP_CHRONO_MAX; ++i) {
3684 stats[i] = tp->chrono_stat[i - 1];
3685 if (i == tp->chrono_type)
3686 stats[i] += tcp_jiffies32 - tp->chrono_start;
3687 stats[i] *= USEC_PER_SEC / HZ;
3691 info->tcpi_busy_time = total;
3692 info->tcpi_rwnd_limited = stats[TCP_CHRONO_RWND_LIMITED];
3693 info->tcpi_sndbuf_limited = stats[TCP_CHRONO_SNDBUF_LIMITED];
3696 /* Return information about state of tcp endpoint in API format. */
3697 void tcp_get_info(struct sock *sk, struct tcp_info *info)
3699 const struct tcp_sock *tp = tcp_sk(sk); /* iff sk_type == SOCK_STREAM */
3700 const struct inet_connection_sock *icsk = inet_csk(sk);
3706 memset(info, 0, sizeof(*info));
3707 if (sk->sk_type != SOCK_STREAM)
3710 info->tcpi_state = inet_sk_state_load(sk);
3712 /* Report meaningful fields for all TCP states, including listeners */
3713 rate = READ_ONCE(sk->sk_pacing_rate);
3714 rate64 = (rate != ~0UL) ? rate : ~0ULL;
3715 info->tcpi_pacing_rate = rate64;
3717 rate = READ_ONCE(sk->sk_max_pacing_rate);
3718 rate64 = (rate != ~0UL) ? rate : ~0ULL;
3719 info->tcpi_max_pacing_rate = rate64;
3721 info->tcpi_reordering = tp->reordering;
3722 info->tcpi_snd_cwnd = tcp_snd_cwnd(tp);
3724 if (info->tcpi_state == TCP_LISTEN) {
3725 /* listeners aliased fields :
3726 * tcpi_unacked -> Number of children ready for accept()
3727 * tcpi_sacked -> max backlog
3729 info->tcpi_unacked = READ_ONCE(sk->sk_ack_backlog);
3730 info->tcpi_sacked = READ_ONCE(sk->sk_max_ack_backlog);
3734 slow = lock_sock_fast(sk);
3736 info->tcpi_ca_state = icsk->icsk_ca_state;
3737 info->tcpi_retransmits = icsk->icsk_retransmits;
3738 info->tcpi_probes = icsk->icsk_probes_out;
3739 info->tcpi_backoff = icsk->icsk_backoff;
3741 if (tp->rx_opt.tstamp_ok)
3742 info->tcpi_options |= TCPI_OPT_TIMESTAMPS;
3743 if (tcp_is_sack(tp))
3744 info->tcpi_options |= TCPI_OPT_SACK;
3745 if (tp->rx_opt.wscale_ok) {
3746 info->tcpi_options |= TCPI_OPT_WSCALE;
3747 info->tcpi_snd_wscale = tp->rx_opt.snd_wscale;
3748 info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale;
3751 if (tp->ecn_flags & TCP_ECN_OK)
3752 info->tcpi_options |= TCPI_OPT_ECN;
3753 if (tp->ecn_flags & TCP_ECN_SEEN)
3754 info->tcpi_options |= TCPI_OPT_ECN_SEEN;
3755 if (tp->syn_data_acked)
3756 info->tcpi_options |= TCPI_OPT_SYN_DATA;
3758 info->tcpi_rto = jiffies_to_usecs(icsk->icsk_rto);
3759 info->tcpi_ato = jiffies_to_usecs(icsk->icsk_ack.ato);
3760 info->tcpi_snd_mss = tp->mss_cache;
3761 info->tcpi_rcv_mss = icsk->icsk_ack.rcv_mss;
3763 info->tcpi_unacked = tp->packets_out;
3764 info->tcpi_sacked = tp->sacked_out;
3766 info->tcpi_lost = tp->lost_out;
3767 info->tcpi_retrans = tp->retrans_out;
3769 now = tcp_jiffies32;
3770 info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime);
3771 info->tcpi_last_data_recv = jiffies_to_msecs(now - icsk->icsk_ack.lrcvtime);
3772 info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp);
3774 info->tcpi_pmtu = icsk->icsk_pmtu_cookie;
3775 info->tcpi_rcv_ssthresh = tp->rcv_ssthresh;
3776 info->tcpi_rtt = tp->srtt_us >> 3;
3777 info->tcpi_rttvar = tp->mdev_us >> 2;
3778 info->tcpi_snd_ssthresh = tp->snd_ssthresh;
3779 info->tcpi_advmss = tp->advmss;
3781 info->tcpi_rcv_rtt = tp->rcv_rtt_est.rtt_us >> 3;
3782 info->tcpi_rcv_space = tp->rcvq_space.space;
3784 info->tcpi_total_retrans = tp->total_retrans;
3786 info->tcpi_bytes_acked = tp->bytes_acked;
3787 info->tcpi_bytes_received = tp->bytes_received;
3788 info->tcpi_notsent_bytes = max_t(int, 0, tp->write_seq - tp->snd_nxt);
3789 tcp_get_info_chrono_stats(tp, info);
3791 info->tcpi_segs_out = tp->segs_out;
3793 /* segs_in and data_segs_in can be updated from tcp_segs_in() from BH */
3794 info->tcpi_segs_in = READ_ONCE(tp->segs_in);
3795 info->tcpi_data_segs_in = READ_ONCE(tp->data_segs_in);
3797 info->tcpi_min_rtt = tcp_min_rtt(tp);
3798 info->tcpi_data_segs_out = tp->data_segs_out;
3800 info->tcpi_delivery_rate_app_limited = tp->rate_app_limited ? 1 : 0;
3801 rate64 = tcp_compute_delivery_rate(tp);
3803 info->tcpi_delivery_rate = rate64;
3804 info->tcpi_delivered = tp->delivered;
3805 info->tcpi_delivered_ce = tp->delivered_ce;
3806 info->tcpi_bytes_sent = tp->bytes_sent;
3807 info->tcpi_bytes_retrans = tp->bytes_retrans;
3808 info->tcpi_dsack_dups = tp->dsack_dups;
3809 info->tcpi_reord_seen = tp->reord_seen;
3810 info->tcpi_rcv_ooopack = tp->rcv_ooopack;
3811 info->tcpi_snd_wnd = tp->snd_wnd;
3812 info->tcpi_rcv_wnd = tp->rcv_wnd;
3813 info->tcpi_rehash = tp->plb_rehash + tp->timeout_rehash;
3814 info->tcpi_fastopen_client_fail = tp->fastopen_client_fail;
3815 unlock_sock_fast(sk, slow);
3817 EXPORT_SYMBOL_GPL(tcp_get_info);
3819 static size_t tcp_opt_stats_get_size(void)
3822 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_BUSY */
3823 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_RWND_LIMITED */
3824 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_SNDBUF_LIMITED */
3825 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_DATA_SEGS_OUT */
3826 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_TOTAL_RETRANS */
3827 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_PACING_RATE */
3828 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_DELIVERY_RATE */
3829 nla_total_size(sizeof(u32)) + /* TCP_NLA_SND_CWND */
3830 nla_total_size(sizeof(u32)) + /* TCP_NLA_REORDERING */
3831 nla_total_size(sizeof(u32)) + /* TCP_NLA_MIN_RTT */
3832 nla_total_size(sizeof(u8)) + /* TCP_NLA_RECUR_RETRANS */
3833 nla_total_size(sizeof(u8)) + /* TCP_NLA_DELIVERY_RATE_APP_LMT */
3834 nla_total_size(sizeof(u32)) + /* TCP_NLA_SNDQ_SIZE */
3835 nla_total_size(sizeof(u8)) + /* TCP_NLA_CA_STATE */
3836 nla_total_size(sizeof(u32)) + /* TCP_NLA_SND_SSTHRESH */
3837 nla_total_size(sizeof(u32)) + /* TCP_NLA_DELIVERED */
3838 nla_total_size(sizeof(u32)) + /* TCP_NLA_DELIVERED_CE */
3839 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_BYTES_SENT */
3840 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_BYTES_RETRANS */
3841 nla_total_size(sizeof(u32)) + /* TCP_NLA_DSACK_DUPS */
3842 nla_total_size(sizeof(u32)) + /* TCP_NLA_REORD_SEEN */
3843 nla_total_size(sizeof(u32)) + /* TCP_NLA_SRTT */
3844 nla_total_size(sizeof(u16)) + /* TCP_NLA_TIMEOUT_REHASH */
3845 nla_total_size(sizeof(u32)) + /* TCP_NLA_BYTES_NOTSENT */
3846 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_EDT */
3847 nla_total_size(sizeof(u8)) + /* TCP_NLA_TTL */
3848 nla_total_size(sizeof(u32)) + /* TCP_NLA_REHASH */
3852 /* Returns TTL or hop limit of an incoming packet from skb. */
3853 static u8 tcp_skb_ttl_or_hop_limit(const struct sk_buff *skb)
3855 if (skb->protocol == htons(ETH_P_IP))
3856 return ip_hdr(skb)->ttl;
3857 else if (skb->protocol == htons(ETH_P_IPV6))
3858 return ipv6_hdr(skb)->hop_limit;
3863 struct sk_buff *tcp_get_timestamping_opt_stats(const struct sock *sk,
3864 const struct sk_buff *orig_skb,
3865 const struct sk_buff *ack_skb)
3867 const struct tcp_sock *tp = tcp_sk(sk);
3868 struct sk_buff *stats;
3869 struct tcp_info info;
3873 stats = alloc_skb(tcp_opt_stats_get_size(), GFP_ATOMIC);
3877 tcp_get_info_chrono_stats(tp, &info);
3878 nla_put_u64_64bit(stats, TCP_NLA_BUSY,
3879 info.tcpi_busy_time, TCP_NLA_PAD);
3880 nla_put_u64_64bit(stats, TCP_NLA_RWND_LIMITED,
3881 info.tcpi_rwnd_limited, TCP_NLA_PAD);
3882 nla_put_u64_64bit(stats, TCP_NLA_SNDBUF_LIMITED,
3883 info.tcpi_sndbuf_limited, TCP_NLA_PAD);
3884 nla_put_u64_64bit(stats, TCP_NLA_DATA_SEGS_OUT,
3885 tp->data_segs_out, TCP_NLA_PAD);
3886 nla_put_u64_64bit(stats, TCP_NLA_TOTAL_RETRANS,
3887 tp->total_retrans, TCP_NLA_PAD);
3889 rate = READ_ONCE(sk->sk_pacing_rate);
3890 rate64 = (rate != ~0UL) ? rate : ~0ULL;
3891 nla_put_u64_64bit(stats, TCP_NLA_PACING_RATE, rate64, TCP_NLA_PAD);
3893 rate64 = tcp_compute_delivery_rate(tp);
3894 nla_put_u64_64bit(stats, TCP_NLA_DELIVERY_RATE, rate64, TCP_NLA_PAD);
3896 nla_put_u32(stats, TCP_NLA_SND_CWND, tcp_snd_cwnd(tp));
3897 nla_put_u32(stats, TCP_NLA_REORDERING, tp->reordering);
3898 nla_put_u32(stats, TCP_NLA_MIN_RTT, tcp_min_rtt(tp));
3900 nla_put_u8(stats, TCP_NLA_RECUR_RETRANS, inet_csk(sk)->icsk_retransmits);
3901 nla_put_u8(stats, TCP_NLA_DELIVERY_RATE_APP_LMT, !!tp->rate_app_limited);
3902 nla_put_u32(stats, TCP_NLA_SND_SSTHRESH, tp->snd_ssthresh);
3903 nla_put_u32(stats, TCP_NLA_DELIVERED, tp->delivered);
3904 nla_put_u32(stats, TCP_NLA_DELIVERED_CE, tp->delivered_ce);
3906 nla_put_u32(stats, TCP_NLA_SNDQ_SIZE, tp->write_seq - tp->snd_una);
3907 nla_put_u8(stats, TCP_NLA_CA_STATE, inet_csk(sk)->icsk_ca_state);
3909 nla_put_u64_64bit(stats, TCP_NLA_BYTES_SENT, tp->bytes_sent,
3911 nla_put_u64_64bit(stats, TCP_NLA_BYTES_RETRANS, tp->bytes_retrans,
3913 nla_put_u32(stats, TCP_NLA_DSACK_DUPS, tp->dsack_dups);
3914 nla_put_u32(stats, TCP_NLA_REORD_SEEN, tp->reord_seen);
3915 nla_put_u32(stats, TCP_NLA_SRTT, tp->srtt_us >> 3);
3916 nla_put_u16(stats, TCP_NLA_TIMEOUT_REHASH, tp->timeout_rehash);
3917 nla_put_u32(stats, TCP_NLA_BYTES_NOTSENT,
3918 max_t(int, 0, tp->write_seq - tp->snd_nxt));
3919 nla_put_u64_64bit(stats, TCP_NLA_EDT, orig_skb->skb_mstamp_ns,
3922 nla_put_u8(stats, TCP_NLA_TTL,
3923 tcp_skb_ttl_or_hop_limit(ack_skb));
3925 nla_put_u32(stats, TCP_NLA_REHASH, tp->plb_rehash + tp->timeout_rehash);
3929 int do_tcp_getsockopt(struct sock *sk, int level,
3930 int optname, sockptr_t optval, sockptr_t optlen)
3932 struct inet_connection_sock *icsk = inet_csk(sk);
3933 struct tcp_sock *tp = tcp_sk(sk);
3934 struct net *net = sock_net(sk);
3937 if (copy_from_sockptr(&len, optlen, sizeof(int)))
3940 len = min_t(unsigned int, len, sizeof(int));
3947 val = tp->mss_cache;
3948 if (tp->rx_opt.user_mss &&
3949 ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
3950 val = tp->rx_opt.user_mss;
3952 val = tp->rx_opt.mss_clamp;
3955 val = !!(tp->nonagle&TCP_NAGLE_OFF);
3958 val = !!(tp->nonagle&TCP_NAGLE_CORK);
3961 val = keepalive_time_when(tp) / HZ;
3964 val = keepalive_intvl_when(tp) / HZ;
3967 val = keepalive_probes(tp);
3970 val = READ_ONCE(icsk->icsk_syn_retries) ? :
3971 READ_ONCE(net->ipv4.sysctl_tcp_syn_retries);
3974 val = READ_ONCE(tp->linger2);
3976 val = (val ? : READ_ONCE(net->ipv4.sysctl_tcp_fin_timeout)) / HZ;
3978 case TCP_DEFER_ACCEPT:
3979 val = READ_ONCE(icsk->icsk_accept_queue.rskq_defer_accept);
3980 val = retrans_to_secs(val, TCP_TIMEOUT_INIT / HZ,
3983 case TCP_WINDOW_CLAMP:
3984 val = tp->window_clamp;
3987 struct tcp_info info;
3989 if (copy_from_sockptr(&len, optlen, sizeof(int)))
3992 tcp_get_info(sk, &info);
3994 len = min_t(unsigned int, len, sizeof(info));
3995 if (copy_to_sockptr(optlen, &len, sizeof(int)))
3997 if (copy_to_sockptr(optval, &info, len))
4002 const struct tcp_congestion_ops *ca_ops;
4003 union tcp_cc_info info;
4007 if (copy_from_sockptr(&len, optlen, sizeof(int)))
4010 ca_ops = icsk->icsk_ca_ops;
4011 if (ca_ops && ca_ops->get_info)
4012 sz = ca_ops->get_info(sk, ~0U, &attr, &info);
4014 len = min_t(unsigned int, len, sz);
4015 if (copy_to_sockptr(optlen, &len, sizeof(int)))
4017 if (copy_to_sockptr(optval, &info, len))
4022 val = !inet_csk_in_pingpong_mode(sk);
4025 case TCP_CONGESTION:
4026 if (copy_from_sockptr(&len, optlen, sizeof(int)))
4028 len = min_t(unsigned int, len, TCP_CA_NAME_MAX);
4029 if (copy_to_sockptr(optlen, &len, sizeof(int)))
4031 if (copy_to_sockptr(optval, icsk->icsk_ca_ops->name, len))
4036 if (copy_from_sockptr(&len, optlen, sizeof(int)))
4038 len = min_t(unsigned int, len, TCP_ULP_NAME_MAX);
4039 if (!icsk->icsk_ulp_ops) {
4041 if (copy_to_sockptr(optlen, &len, sizeof(int)))
4045 if (copy_to_sockptr(optlen, &len, sizeof(int)))
4047 if (copy_to_sockptr(optval, icsk->icsk_ulp_ops->name, len))
4051 case TCP_FASTOPEN_KEY: {
4052 u64 key[TCP_FASTOPEN_KEY_BUF_LENGTH / sizeof(u64)];
4053 unsigned int key_len;
4055 if (copy_from_sockptr(&len, optlen, sizeof(int)))
4058 key_len = tcp_fastopen_get_cipher(net, icsk, key) *
4059 TCP_FASTOPEN_KEY_LENGTH;
4060 len = min_t(unsigned int, len, key_len);
4061 if (copy_to_sockptr(optlen, &len, sizeof(int)))
4063 if (copy_to_sockptr(optval, key, len))
4067 case TCP_THIN_LINEAR_TIMEOUTS:
4071 case TCP_THIN_DUPACK:
4079 case TCP_REPAIR_QUEUE:
4081 val = tp->repair_queue;
4086 case TCP_REPAIR_WINDOW: {
4087 struct tcp_repair_window opt;
4089 if (copy_from_sockptr(&len, optlen, sizeof(int)))
4092 if (len != sizeof(opt))
4098 opt.snd_wl1 = tp->snd_wl1;
4099 opt.snd_wnd = tp->snd_wnd;
4100 opt.max_window = tp->max_window;
4101 opt.rcv_wnd = tp->rcv_wnd;
4102 opt.rcv_wup = tp->rcv_wup;
4104 if (copy_to_sockptr(optval, &opt, len))
4109 if (tp->repair_queue == TCP_SEND_QUEUE)
4110 val = tp->write_seq;
4111 else if (tp->repair_queue == TCP_RECV_QUEUE)
4117 case TCP_USER_TIMEOUT:
4118 val = READ_ONCE(icsk->icsk_user_timeout);
4122 val = READ_ONCE(icsk->icsk_accept_queue.fastopenq.max_qlen);
4125 case TCP_FASTOPEN_CONNECT:
4126 val = tp->fastopen_connect;
4129 case TCP_FASTOPEN_NO_COOKIE:
4130 val = tp->fastopen_no_cookie;
4134 val = READ_ONCE(tp->tcp_tx_delay);
4138 val = tcp_time_stamp_raw() + READ_ONCE(tp->tsoffset);
4140 case TCP_NOTSENT_LOWAT:
4141 val = READ_ONCE(tp->notsent_lowat);
4144 val = tp->recvmsg_inq;
4149 case TCP_SAVED_SYN: {
4150 if (copy_from_sockptr(&len, optlen, sizeof(int)))
4153 sockopt_lock_sock(sk);
4154 if (tp->saved_syn) {
4155 if (len < tcp_saved_syn_len(tp->saved_syn)) {
4156 len = tcp_saved_syn_len(tp->saved_syn);
4157 if (copy_to_sockptr(optlen, &len, sizeof(int))) {
4158 sockopt_release_sock(sk);
4161 sockopt_release_sock(sk);
4164 len = tcp_saved_syn_len(tp->saved_syn);
4165 if (copy_to_sockptr(optlen, &len, sizeof(int))) {
4166 sockopt_release_sock(sk);
4169 if (copy_to_sockptr(optval, tp->saved_syn->data, len)) {
4170 sockopt_release_sock(sk);
4173 tcp_saved_syn_free(tp);
4174 sockopt_release_sock(sk);
4176 sockopt_release_sock(sk);
4178 if (copy_to_sockptr(optlen, &len, sizeof(int)))
4184 case TCP_ZEROCOPY_RECEIVE: {
4185 struct scm_timestamping_internal tss;
4186 struct tcp_zerocopy_receive zc = {};
4189 if (copy_from_sockptr(&len, optlen, sizeof(int)))
4192 len < offsetofend(struct tcp_zerocopy_receive, length))
4194 if (unlikely(len > sizeof(zc))) {
4195 err = check_zeroed_sockptr(optval, sizeof(zc),
4198 return err == 0 ? -EINVAL : err;
4200 if (copy_to_sockptr(optlen, &len, sizeof(int)))
4203 if (copy_from_sockptr(&zc, optval, len))
4207 if (zc.msg_flags & ~(TCP_VALID_ZC_MSG_FLAGS))
4209 sockopt_lock_sock(sk);
4210 err = tcp_zerocopy_receive(sk, &zc, &tss);
4211 err = BPF_CGROUP_RUN_PROG_GETSOCKOPT_KERN(sk, level, optname,
4213 sockopt_release_sock(sk);
4214 if (len >= offsetofend(struct tcp_zerocopy_receive, msg_flags))
4215 goto zerocopy_rcv_cmsg;
4217 case offsetofend(struct tcp_zerocopy_receive, msg_flags):
4218 goto zerocopy_rcv_cmsg;
4219 case offsetofend(struct tcp_zerocopy_receive, msg_controllen):
4220 case offsetofend(struct tcp_zerocopy_receive, msg_control):
4221 case offsetofend(struct tcp_zerocopy_receive, flags):
4222 case offsetofend(struct tcp_zerocopy_receive, copybuf_len):
4223 case offsetofend(struct tcp_zerocopy_receive, copybuf_address):
4224 case offsetofend(struct tcp_zerocopy_receive, err):
4225 goto zerocopy_rcv_sk_err;
4226 case offsetofend(struct tcp_zerocopy_receive, inq):
4227 goto zerocopy_rcv_inq;
4228 case offsetofend(struct tcp_zerocopy_receive, length):
4230 goto zerocopy_rcv_out;
4233 if (zc.msg_flags & TCP_CMSG_TS)
4234 tcp_zc_finalize_rx_tstamp(sk, &zc, &tss);
4237 zerocopy_rcv_sk_err:
4239 zc.err = sock_error(sk);
4241 zc.inq = tcp_inq_hint(sk);
4243 if (!err && copy_to_sockptr(optval, &zc, len))
4249 return -ENOPROTOOPT;
4252 if (copy_to_sockptr(optlen, &len, sizeof(int)))
4254 if (copy_to_sockptr(optval, &val, len))
4259 bool tcp_bpf_bypass_getsockopt(int level, int optname)
4261 /* TCP do_tcp_getsockopt has optimized getsockopt implementation
4262 * to avoid extra socket lock for TCP_ZEROCOPY_RECEIVE.
4264 if (level == SOL_TCP && optname == TCP_ZEROCOPY_RECEIVE)
4269 EXPORT_SYMBOL(tcp_bpf_bypass_getsockopt);
4271 int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval,
4274 struct inet_connection_sock *icsk = inet_csk(sk);
4276 if (level != SOL_TCP)
4277 /* Paired with WRITE_ONCE() in do_ipv6_setsockopt() and tcp_v6_connect() */
4278 return READ_ONCE(icsk->icsk_af_ops)->getsockopt(sk, level, optname,
4280 return do_tcp_getsockopt(sk, level, optname, USER_SOCKPTR(optval),
4281 USER_SOCKPTR(optlen));
4283 EXPORT_SYMBOL(tcp_getsockopt);
4285 #ifdef CONFIG_TCP_MD5SIG
4286 static DEFINE_PER_CPU(struct tcp_md5sig_pool, tcp_md5sig_pool);
4287 static DEFINE_MUTEX(tcp_md5sig_mutex);
4288 static bool tcp_md5sig_pool_populated = false;
4290 static void __tcp_alloc_md5sig_pool(void)
4292 struct crypto_ahash *hash;
4295 hash = crypto_alloc_ahash("md5", 0, CRYPTO_ALG_ASYNC);
4299 for_each_possible_cpu(cpu) {
4300 void *scratch = per_cpu(tcp_md5sig_pool, cpu).scratch;
4301 struct ahash_request *req;
4304 scratch = kmalloc_node(sizeof(union tcp_md5sum_block) +
4305 sizeof(struct tcphdr),
4310 per_cpu(tcp_md5sig_pool, cpu).scratch = scratch;
4312 if (per_cpu(tcp_md5sig_pool, cpu).md5_req)
4315 req = ahash_request_alloc(hash, GFP_KERNEL);
4319 ahash_request_set_callback(req, 0, NULL, NULL);
4321 per_cpu(tcp_md5sig_pool, cpu).md5_req = req;
4323 /* before setting tcp_md5sig_pool_populated, we must commit all writes
4324 * to memory. See smp_rmb() in tcp_get_md5sig_pool()
4327 /* Paired with READ_ONCE() from tcp_alloc_md5sig_pool()
4328 * and tcp_get_md5sig_pool().
4330 WRITE_ONCE(tcp_md5sig_pool_populated, true);
4333 bool tcp_alloc_md5sig_pool(void)
4335 /* Paired with WRITE_ONCE() from __tcp_alloc_md5sig_pool() */
4336 if (unlikely(!READ_ONCE(tcp_md5sig_pool_populated))) {
4337 mutex_lock(&tcp_md5sig_mutex);
4339 if (!tcp_md5sig_pool_populated)
4340 __tcp_alloc_md5sig_pool();
4342 mutex_unlock(&tcp_md5sig_mutex);
4344 /* Paired with WRITE_ONCE() from __tcp_alloc_md5sig_pool() */
4345 return READ_ONCE(tcp_md5sig_pool_populated);
4347 EXPORT_SYMBOL(tcp_alloc_md5sig_pool);
4351 * tcp_get_md5sig_pool - get md5sig_pool for this user
4353 * We use percpu structure, so if we succeed, we exit with preemption
4354 * and BH disabled, to make sure another thread or softirq handling
4355 * wont try to get same context.
4357 struct tcp_md5sig_pool *tcp_get_md5sig_pool(void)
4361 /* Paired with WRITE_ONCE() from __tcp_alloc_md5sig_pool() */
4362 if (READ_ONCE(tcp_md5sig_pool_populated)) {
4363 /* coupled with smp_wmb() in __tcp_alloc_md5sig_pool() */
4365 return this_cpu_ptr(&tcp_md5sig_pool);
4370 EXPORT_SYMBOL(tcp_get_md5sig_pool);
4372 int tcp_md5_hash_skb_data(struct tcp_md5sig_pool *hp,
4373 const struct sk_buff *skb, unsigned int header_len)
4375 struct scatterlist sg;
4376 const struct tcphdr *tp = tcp_hdr(skb);
4377 struct ahash_request *req = hp->md5_req;
4379 const unsigned int head_data_len = skb_headlen(skb) > header_len ?
4380 skb_headlen(skb) - header_len : 0;
4381 const struct skb_shared_info *shi = skb_shinfo(skb);
4382 struct sk_buff *frag_iter;
4384 sg_init_table(&sg, 1);
4386 sg_set_buf(&sg, ((u8 *) tp) + header_len, head_data_len);
4387 ahash_request_set_crypt(req, &sg, NULL, head_data_len);
4388 if (crypto_ahash_update(req))
4391 for (i = 0; i < shi->nr_frags; ++i) {
4392 const skb_frag_t *f = &shi->frags[i];
4393 unsigned int offset = skb_frag_off(f);
4394 struct page *page = skb_frag_page(f) + (offset >> PAGE_SHIFT);
4396 sg_set_page(&sg, page, skb_frag_size(f),
4397 offset_in_page(offset));
4398 ahash_request_set_crypt(req, &sg, NULL, skb_frag_size(f));
4399 if (crypto_ahash_update(req))
4403 skb_walk_frags(skb, frag_iter)
4404 if (tcp_md5_hash_skb_data(hp, frag_iter, 0))
4409 EXPORT_SYMBOL(tcp_md5_hash_skb_data);
4411 int tcp_md5_hash_key(struct tcp_md5sig_pool *hp, const struct tcp_md5sig_key *key)
4413 u8 keylen = READ_ONCE(key->keylen); /* paired with WRITE_ONCE() in tcp_md5_do_add */
4414 struct scatterlist sg;
4416 sg_init_one(&sg, key->key, keylen);
4417 ahash_request_set_crypt(hp->md5_req, &sg, NULL, keylen);
4419 /* We use data_race() because tcp_md5_do_add() might change key->key under us */
4420 return data_race(crypto_ahash_update(hp->md5_req));
4422 EXPORT_SYMBOL(tcp_md5_hash_key);
4424 /* Called with rcu_read_lock() */
4425 enum skb_drop_reason
4426 tcp_inbound_md5_hash(const struct sock *sk, const struct sk_buff *skb,
4427 const void *saddr, const void *daddr,
4428 int family, int dif, int sdif)
4431 * This gets called for each TCP segment that arrives
4432 * so we want to be efficient.
4433 * We have 3 drop cases:
4434 * o No MD5 hash and one expected.
4435 * o MD5 hash and we're not expecting one.
4436 * o MD5 hash and its wrong.
4438 const __u8 *hash_location = NULL;
4439 struct tcp_md5sig_key *hash_expected;
4440 const struct tcphdr *th = tcp_hdr(skb);
4441 const struct tcp_sock *tp = tcp_sk(sk);
4442 int genhash, l3index;
4445 /* sdif set, means packet ingressed via a device
4446 * in an L3 domain and dif is set to the l3mdev
4448 l3index = sdif ? dif : 0;
4450 hash_expected = tcp_md5_do_lookup(sk, l3index, saddr, family);
4451 hash_location = tcp_parse_md5sig_option(th);
4453 /* We've parsed the options - do we have a hash? */
4454 if (!hash_expected && !hash_location)
4455 return SKB_NOT_DROPPED_YET;
4457 if (hash_expected && !hash_location) {
4458 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMD5NOTFOUND);
4459 return SKB_DROP_REASON_TCP_MD5NOTFOUND;
4462 if (!hash_expected && hash_location) {
4463 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMD5UNEXPECTED);
4464 return SKB_DROP_REASON_TCP_MD5UNEXPECTED;
4467 /* Check the signature.
4468 * To support dual stack listeners, we need to handle
4471 if (family == AF_INET)
4472 genhash = tcp_v4_md5_hash_skb(newhash,
4476 genhash = tp->af_specific->calc_md5_hash(newhash,
4480 if (genhash || memcmp(hash_location, newhash, 16) != 0) {
4481 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMD5FAILURE);
4482 if (family == AF_INET) {
4483 net_info_ratelimited("MD5 Hash failed for (%pI4, %d)->(%pI4, %d)%s L3 index %d\n",
4484 saddr, ntohs(th->source),
4485 daddr, ntohs(th->dest),
4486 genhash ? " tcp_v4_calc_md5_hash failed"
4489 net_info_ratelimited("MD5 Hash %s for [%pI6c]:%u->[%pI6c]:%u L3 index %d\n",
4490 genhash ? "failed" : "mismatch",
4491 saddr, ntohs(th->source),
4492 daddr, ntohs(th->dest), l3index);
4494 return SKB_DROP_REASON_TCP_MD5FAILURE;
4496 return SKB_NOT_DROPPED_YET;
4498 EXPORT_SYMBOL(tcp_inbound_md5_hash);
4502 void tcp_done(struct sock *sk)
4504 struct request_sock *req;
4506 /* We might be called with a new socket, after
4507 * inet_csk_prepare_forced_close() has been called
4508 * so we can not use lockdep_sock_is_held(sk)
4510 req = rcu_dereference_protected(tcp_sk(sk)->fastopen_rsk, 1);
4512 if (sk->sk_state == TCP_SYN_SENT || sk->sk_state == TCP_SYN_RECV)
4513 TCP_INC_STATS(sock_net(sk), TCP_MIB_ATTEMPTFAILS);
4515 tcp_set_state(sk, TCP_CLOSE);
4516 tcp_clear_xmit_timers(sk);
4518 reqsk_fastopen_remove(sk, req, false);
4520 WRITE_ONCE(sk->sk_shutdown, SHUTDOWN_MASK);
4522 if (!sock_flag(sk, SOCK_DEAD))
4523 sk->sk_state_change(sk);
4525 inet_csk_destroy_sock(sk);
4527 EXPORT_SYMBOL_GPL(tcp_done);
4529 int tcp_abort(struct sock *sk, int err)
4531 int state = inet_sk_state_load(sk);
4533 if (state == TCP_NEW_SYN_RECV) {
4534 struct request_sock *req = inet_reqsk(sk);
4537 inet_csk_reqsk_queue_drop(req->rsk_listener, req);
4541 if (state == TCP_TIME_WAIT) {
4542 struct inet_timewait_sock *tw = inet_twsk(sk);
4544 refcount_inc(&tw->tw_refcnt);
4546 inet_twsk_deschedule_put(tw);
4551 /* BPF context ensures sock locking. */
4552 if (!has_current_bpf_ctx())
4553 /* Don't race with userspace socket closes such as tcp_close. */
4556 if (sk->sk_state == TCP_LISTEN) {
4557 tcp_set_state(sk, TCP_CLOSE);
4558 inet_csk_listen_stop(sk);
4561 /* Don't race with BH socket closes such as inet_csk_listen_stop. */
4565 if (!sock_flag(sk, SOCK_DEAD)) {
4566 WRITE_ONCE(sk->sk_err, err);
4567 /* This barrier is coupled with smp_rmb() in tcp_poll() */
4569 sk_error_report(sk);
4570 if (tcp_need_reset(sk->sk_state))
4571 tcp_send_active_reset(sk, GFP_ATOMIC);
4577 tcp_write_queue_purge(sk);
4578 if (!has_current_bpf_ctx())
4582 EXPORT_SYMBOL_GPL(tcp_abort);
4584 extern struct tcp_congestion_ops tcp_reno;
4586 static __initdata unsigned long thash_entries;
4587 static int __init set_thash_entries(char *str)
4594 ret = kstrtoul(str, 0, &thash_entries);
4600 __setup("thash_entries=", set_thash_entries);
4602 static void __init tcp_init_mem(void)
4604 unsigned long limit = nr_free_buffer_pages() / 16;
4606 limit = max(limit, 128UL);
4607 sysctl_tcp_mem[0] = limit / 4 * 3; /* 4.68 % */
4608 sysctl_tcp_mem[1] = limit; /* 6.25 % */
4609 sysctl_tcp_mem[2] = sysctl_tcp_mem[0] * 2; /* 9.37 % */
4612 void __init tcp_init(void)
4614 int max_rshare, max_wshare, cnt;
4615 unsigned long limit;
4618 BUILD_BUG_ON(TCP_MIN_SND_MSS <= MAX_TCP_OPTION_SPACE);
4619 BUILD_BUG_ON(sizeof(struct tcp_skb_cb) >
4620 sizeof_field(struct sk_buff, cb));
4622 percpu_counter_init(&tcp_sockets_allocated, 0, GFP_KERNEL);
4624 timer_setup(&tcp_orphan_timer, tcp_orphan_update, TIMER_DEFERRABLE);
4625 mod_timer(&tcp_orphan_timer, jiffies + TCP_ORPHAN_TIMER_PERIOD);
4627 inet_hashinfo2_init(&tcp_hashinfo, "tcp_listen_portaddr_hash",
4628 thash_entries, 21, /* one slot per 2 MB*/
4630 tcp_hashinfo.bind_bucket_cachep =
4631 kmem_cache_create("tcp_bind_bucket",
4632 sizeof(struct inet_bind_bucket), 0,
4633 SLAB_HWCACHE_ALIGN | SLAB_PANIC |
4636 tcp_hashinfo.bind2_bucket_cachep =
4637 kmem_cache_create("tcp_bind2_bucket",
4638 sizeof(struct inet_bind2_bucket), 0,
4639 SLAB_HWCACHE_ALIGN | SLAB_PANIC |
4643 /* Size and allocate the main established and bind bucket
4646 * The methodology is similar to that of the buffer cache.
4648 tcp_hashinfo.ehash =
4649 alloc_large_system_hash("TCP established",
4650 sizeof(struct inet_ehash_bucket),
4652 17, /* one slot per 128 KB of memory */
4655 &tcp_hashinfo.ehash_mask,
4657 thash_entries ? 0 : 512 * 1024);
4658 for (i = 0; i <= tcp_hashinfo.ehash_mask; i++)
4659 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].chain, i);
4661 if (inet_ehash_locks_alloc(&tcp_hashinfo))
4662 panic("TCP: failed to alloc ehash_locks");
4663 tcp_hashinfo.bhash =
4664 alloc_large_system_hash("TCP bind",
4665 2 * sizeof(struct inet_bind_hashbucket),
4666 tcp_hashinfo.ehash_mask + 1,
4667 17, /* one slot per 128 KB of memory */
4669 &tcp_hashinfo.bhash_size,
4673 tcp_hashinfo.bhash_size = 1U << tcp_hashinfo.bhash_size;
4674 tcp_hashinfo.bhash2 = tcp_hashinfo.bhash + tcp_hashinfo.bhash_size;
4675 for (i = 0; i < tcp_hashinfo.bhash_size; i++) {
4676 spin_lock_init(&tcp_hashinfo.bhash[i].lock);
4677 INIT_HLIST_HEAD(&tcp_hashinfo.bhash[i].chain);
4678 spin_lock_init(&tcp_hashinfo.bhash2[i].lock);
4679 INIT_HLIST_HEAD(&tcp_hashinfo.bhash2[i].chain);
4682 tcp_hashinfo.pernet = false;
4684 cnt = tcp_hashinfo.ehash_mask + 1;
4685 sysctl_tcp_max_orphans = cnt / 2;
4688 /* Set per-socket limits to no more than 1/128 the pressure threshold */
4689 limit = nr_free_buffer_pages() << (PAGE_SHIFT - 7);
4690 max_wshare = min(4UL*1024*1024, limit);
4691 max_rshare = min(6UL*1024*1024, limit);
4693 init_net.ipv4.sysctl_tcp_wmem[0] = PAGE_SIZE;
4694 init_net.ipv4.sysctl_tcp_wmem[1] = 16*1024;
4695 init_net.ipv4.sysctl_tcp_wmem[2] = max(64*1024, max_wshare);
4697 init_net.ipv4.sysctl_tcp_rmem[0] = PAGE_SIZE;
4698 init_net.ipv4.sysctl_tcp_rmem[1] = 131072;
4699 init_net.ipv4.sysctl_tcp_rmem[2] = max(131072, max_rshare);
4701 pr_info("Hash tables configured (established %u bind %u)\n",
4702 tcp_hashinfo.ehash_mask + 1, tcp_hashinfo.bhash_size);
4706 BUG_ON(tcp_register_congestion_control(&tcp_reno) != 0);