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]));
461 set_bit(SOCK_SUPPORT_ZC, &sk->sk_socket->flags);
462 sk_sockets_allocated_inc(sk);
464 EXPORT_SYMBOL(tcp_init_sock);
466 static void tcp_tx_timestamp(struct sock *sk, u16 tsflags)
468 struct sk_buff *skb = tcp_write_queue_tail(sk);
470 if (tsflags && skb) {
471 struct skb_shared_info *shinfo = skb_shinfo(skb);
472 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
474 sock_tx_timestamp(sk, tsflags, &shinfo->tx_flags);
475 if (tsflags & SOF_TIMESTAMPING_TX_ACK)
476 tcb->txstamp_ack = 1;
477 if (tsflags & SOF_TIMESTAMPING_TX_RECORD_MASK)
478 shinfo->tskey = TCP_SKB_CB(skb)->seq + skb->len - 1;
482 static bool tcp_stream_is_readable(struct sock *sk, int target)
484 if (tcp_epollin_ready(sk, target))
486 return sk_is_readable(sk);
490 * Wait for a TCP event.
492 * Note that we don't need to lock the socket, as the upper poll layers
493 * take care of normal races (between the test and the event) and we don't
494 * go look at any of the socket buffers directly.
496 __poll_t tcp_poll(struct file *file, struct socket *sock, poll_table *wait)
499 struct sock *sk = sock->sk;
500 const struct tcp_sock *tp = tcp_sk(sk);
503 sock_poll_wait(file, sock, wait);
505 state = inet_sk_state_load(sk);
506 if (state == TCP_LISTEN)
507 return inet_csk_listen_poll(sk);
509 /* Socket is not locked. We are protected from async events
510 * by poll logic and correct handling of state changes
511 * made by other threads is impossible in any case.
517 * EPOLLHUP is certainly not done right. But poll() doesn't
518 * have a notion of HUP in just one direction, and for a
519 * socket the read side is more interesting.
521 * Some poll() documentation says that EPOLLHUP is incompatible
522 * with the EPOLLOUT/POLLWR flags, so somebody should check this
523 * all. But careful, it tends to be safer to return too many
524 * bits than too few, and you can easily break real applications
525 * if you don't tell them that something has hung up!
529 * Check number 1. EPOLLHUP is _UNMASKABLE_ event (see UNIX98 and
530 * our fs/select.c). It means that after we received EOF,
531 * poll always returns immediately, making impossible poll() on write()
532 * in state CLOSE_WAIT. One solution is evident --- to set EPOLLHUP
533 * if and only if shutdown has been made in both directions.
534 * Actually, it is interesting to look how Solaris and DUX
535 * solve this dilemma. I would prefer, if EPOLLHUP were maskable,
536 * then we could set it on SND_SHUTDOWN. BTW examples given
537 * in Stevens' books assume exactly this behaviour, it explains
538 * why EPOLLHUP is incompatible with EPOLLOUT. --ANK
540 * NOTE. Check for TCP_CLOSE is added. The goal is to prevent
541 * blocking on fresh not-connected or disconnected socket. --ANK
543 if (sk->sk_shutdown == SHUTDOWN_MASK || state == TCP_CLOSE)
545 if (sk->sk_shutdown & RCV_SHUTDOWN)
546 mask |= EPOLLIN | EPOLLRDNORM | EPOLLRDHUP;
548 /* Connected or passive Fast Open socket? */
549 if (state != TCP_SYN_SENT &&
550 (state != TCP_SYN_RECV || rcu_access_pointer(tp->fastopen_rsk))) {
551 int target = sock_rcvlowat(sk, 0, INT_MAX);
552 u16 urg_data = READ_ONCE(tp->urg_data);
554 if (unlikely(urg_data) &&
555 READ_ONCE(tp->urg_seq) == READ_ONCE(tp->copied_seq) &&
556 !sock_flag(sk, SOCK_URGINLINE))
559 if (tcp_stream_is_readable(sk, target))
560 mask |= EPOLLIN | EPOLLRDNORM;
562 if (!(sk->sk_shutdown & SEND_SHUTDOWN)) {
563 if (__sk_stream_is_writeable(sk, 1)) {
564 mask |= EPOLLOUT | EPOLLWRNORM;
565 } else { /* send SIGIO later */
566 sk_set_bit(SOCKWQ_ASYNC_NOSPACE, sk);
567 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
569 /* Race breaker. If space is freed after
570 * wspace test but before the flags are set,
571 * IO signal will be lost. Memory barrier
572 * pairs with the input side.
574 smp_mb__after_atomic();
575 if (__sk_stream_is_writeable(sk, 1))
576 mask |= EPOLLOUT | EPOLLWRNORM;
579 mask |= EPOLLOUT | EPOLLWRNORM;
581 if (urg_data & TCP_URG_VALID)
583 } else if (state == TCP_SYN_SENT && inet_sk(sk)->defer_connect) {
584 /* Active TCP fastopen socket with defer_connect
585 * Return EPOLLOUT so application can call write()
586 * in order for kernel to generate SYN+data
588 mask |= EPOLLOUT | EPOLLWRNORM;
590 /* This barrier is coupled with smp_wmb() in tcp_reset() */
592 if (READ_ONCE(sk->sk_err) ||
593 !skb_queue_empty_lockless(&sk->sk_error_queue))
598 EXPORT_SYMBOL(tcp_poll);
600 int tcp_ioctl(struct sock *sk, int cmd, unsigned long arg)
602 struct tcp_sock *tp = tcp_sk(sk);
608 if (sk->sk_state == TCP_LISTEN)
611 slow = lock_sock_fast(sk);
613 unlock_sock_fast(sk, slow);
616 answ = READ_ONCE(tp->urg_data) &&
617 READ_ONCE(tp->urg_seq) == READ_ONCE(tp->copied_seq);
620 if (sk->sk_state == TCP_LISTEN)
623 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
626 answ = READ_ONCE(tp->write_seq) - tp->snd_una;
629 if (sk->sk_state == TCP_LISTEN)
632 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
635 answ = READ_ONCE(tp->write_seq) -
636 READ_ONCE(tp->snd_nxt);
642 return put_user(answ, (int __user *)arg);
644 EXPORT_SYMBOL(tcp_ioctl);
646 void tcp_mark_push(struct tcp_sock *tp, struct sk_buff *skb)
648 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
649 tp->pushed_seq = tp->write_seq;
652 static inline bool forced_push(const struct tcp_sock *tp)
654 return after(tp->write_seq, tp->pushed_seq + (tp->max_window >> 1));
657 void tcp_skb_entail(struct sock *sk, struct sk_buff *skb)
659 struct tcp_sock *tp = tcp_sk(sk);
660 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
662 tcb->seq = tcb->end_seq = tp->write_seq;
663 tcb->tcp_flags = TCPHDR_ACK;
664 __skb_header_release(skb);
665 tcp_add_write_queue_tail(sk, skb);
666 sk_wmem_queued_add(sk, skb->truesize);
667 sk_mem_charge(sk, skb->truesize);
668 if (tp->nonagle & TCP_NAGLE_PUSH)
669 tp->nonagle &= ~TCP_NAGLE_PUSH;
671 tcp_slow_start_after_idle_check(sk);
674 static inline void tcp_mark_urg(struct tcp_sock *tp, int flags)
677 tp->snd_up = tp->write_seq;
680 /* If a not yet filled skb is pushed, do not send it if
681 * we have data packets in Qdisc or NIC queues :
682 * Because TX completion will happen shortly, it gives a chance
683 * to coalesce future sendmsg() payload into this skb, without
684 * need for a timer, and with no latency trade off.
685 * As packets containing data payload have a bigger truesize
686 * than pure acks (dataless) packets, the last checks prevent
687 * autocorking if we only have an ACK in Qdisc/NIC queues,
688 * or if TX completion was delayed after we processed ACK packet.
690 static bool tcp_should_autocork(struct sock *sk, struct sk_buff *skb,
693 return skb->len < size_goal &&
694 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_autocorking) &&
695 !tcp_rtx_queue_empty(sk) &&
696 refcount_read(&sk->sk_wmem_alloc) > skb->truesize &&
697 tcp_skb_can_collapse_to(skb);
700 void tcp_push(struct sock *sk, int flags, int mss_now,
701 int nonagle, int size_goal)
703 struct tcp_sock *tp = tcp_sk(sk);
706 skb = tcp_write_queue_tail(sk);
709 if (!(flags & MSG_MORE) || forced_push(tp))
710 tcp_mark_push(tp, skb);
712 tcp_mark_urg(tp, flags);
714 if (tcp_should_autocork(sk, skb, size_goal)) {
716 /* avoid atomic op if TSQ_THROTTLED bit is already set */
717 if (!test_bit(TSQ_THROTTLED, &sk->sk_tsq_flags)) {
718 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPAUTOCORKING);
719 set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags);
721 /* It is possible TX completion already happened
722 * before we set TSQ_THROTTLED.
724 if (refcount_read(&sk->sk_wmem_alloc) > skb->truesize)
728 if (flags & MSG_MORE)
729 nonagle = TCP_NAGLE_CORK;
731 __tcp_push_pending_frames(sk, mss_now, nonagle);
734 static int tcp_splice_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb,
735 unsigned int offset, size_t len)
737 struct tcp_splice_state *tss = rd_desc->arg.data;
740 ret = skb_splice_bits(skb, skb->sk, offset, tss->pipe,
741 min(rd_desc->count, len), tss->flags);
743 rd_desc->count -= ret;
747 static int __tcp_splice_read(struct sock *sk, struct tcp_splice_state *tss)
749 /* Store TCP splice context information in read_descriptor_t. */
750 read_descriptor_t rd_desc = {
755 return tcp_read_sock(sk, &rd_desc, tcp_splice_data_recv);
759 * tcp_splice_read - splice data from TCP socket to a pipe
760 * @sock: socket to splice from
761 * @ppos: position (not valid)
762 * @pipe: pipe to splice to
763 * @len: number of bytes to splice
764 * @flags: splice modifier flags
767 * Will read pages from given socket and fill them into a pipe.
770 ssize_t tcp_splice_read(struct socket *sock, loff_t *ppos,
771 struct pipe_inode_info *pipe, size_t len,
774 struct sock *sk = sock->sk;
775 struct tcp_splice_state tss = {
784 sock_rps_record_flow(sk);
786 * We can't seek on a socket input
795 timeo = sock_rcvtimeo(sk, sock->file->f_flags & O_NONBLOCK);
797 ret = __tcp_splice_read(sk, &tss);
803 if (sock_flag(sk, SOCK_DONE))
806 ret = sock_error(sk);
809 if (sk->sk_shutdown & RCV_SHUTDOWN)
811 if (sk->sk_state == TCP_CLOSE) {
813 * This occurs when user tries to read
814 * from never connected socket.
823 /* if __tcp_splice_read() got nothing while we have
824 * an skb in receive queue, we do not want to loop.
825 * This might happen with URG data.
827 if (!skb_queue_empty(&sk->sk_receive_queue))
829 sk_wait_data(sk, &timeo, NULL);
830 if (signal_pending(current)) {
831 ret = sock_intr_errno(timeo);
844 if (sk->sk_err || sk->sk_state == TCP_CLOSE ||
845 (sk->sk_shutdown & RCV_SHUTDOWN) ||
846 signal_pending(current))
857 EXPORT_SYMBOL(tcp_splice_read);
859 struct sk_buff *tcp_stream_alloc_skb(struct sock *sk, int size, gfp_t gfp,
864 skb = alloc_skb_fclone(size + MAX_TCP_HEADER, gfp);
868 skb->truesize = SKB_TRUESIZE(skb_end_offset(skb));
869 if (force_schedule) {
870 mem_scheduled = true;
871 sk_forced_mem_schedule(sk, skb->truesize);
873 mem_scheduled = sk_wmem_schedule(sk, skb->truesize);
875 if (likely(mem_scheduled)) {
876 skb_reserve(skb, MAX_TCP_HEADER);
877 skb->ip_summed = CHECKSUM_PARTIAL;
878 INIT_LIST_HEAD(&skb->tcp_tsorted_anchor);
883 sk->sk_prot->enter_memory_pressure(sk);
884 sk_stream_moderate_sndbuf(sk);
889 static unsigned int tcp_xmit_size_goal(struct sock *sk, u32 mss_now,
892 struct tcp_sock *tp = tcp_sk(sk);
893 u32 new_size_goal, size_goal;
898 /* Note : tcp_tso_autosize() will eventually split this later */
899 new_size_goal = tcp_bound_to_half_wnd(tp, sk->sk_gso_max_size);
901 /* We try hard to avoid divides here */
902 size_goal = tp->gso_segs * mss_now;
903 if (unlikely(new_size_goal < size_goal ||
904 new_size_goal >= size_goal + mss_now)) {
905 tp->gso_segs = min_t(u16, new_size_goal / mss_now,
906 sk->sk_gso_max_segs);
907 size_goal = tp->gso_segs * mss_now;
910 return max(size_goal, mss_now);
913 int tcp_send_mss(struct sock *sk, int *size_goal, int flags)
917 mss_now = tcp_current_mss(sk);
918 *size_goal = tcp_xmit_size_goal(sk, mss_now, !(flags & MSG_OOB));
923 /* In some cases, both sendpage() and sendmsg() could have added
924 * an skb to the write queue, but failed adding payload on it.
925 * We need to remove it to consume less memory, but more
926 * importantly be able to generate EPOLLOUT for Edge Trigger epoll()
929 void tcp_remove_empty_skb(struct sock *sk)
931 struct sk_buff *skb = tcp_write_queue_tail(sk);
933 if (skb && TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq) {
934 tcp_unlink_write_queue(skb, sk);
935 if (tcp_write_queue_empty(sk))
936 tcp_chrono_stop(sk, TCP_CHRONO_BUSY);
937 tcp_wmem_free_skb(sk, skb);
941 /* skb changing from pure zc to mixed, must charge zc */
942 static int tcp_downgrade_zcopy_pure(struct sock *sk, struct sk_buff *skb)
944 if (unlikely(skb_zcopy_pure(skb))) {
945 u32 extra = skb->truesize -
946 SKB_TRUESIZE(skb_end_offset(skb));
948 if (!sk_wmem_schedule(sk, extra))
951 sk_mem_charge(sk, extra);
952 skb_shinfo(skb)->flags &= ~SKBFL_PURE_ZEROCOPY;
958 static int tcp_wmem_schedule(struct sock *sk, int copy)
962 if (likely(sk_wmem_schedule(sk, copy)))
965 /* We could be in trouble if we have nothing queued.
966 * Use whatever is left in sk->sk_forward_alloc and tcp_wmem[0]
967 * to guarantee some progress.
969 left = sock_net(sk)->ipv4.sysctl_tcp_wmem[0] - sk->sk_wmem_queued;
971 sk_forced_mem_schedule(sk, min(left, copy));
972 return min(copy, sk->sk_forward_alloc);
975 static struct sk_buff *tcp_build_frag(struct sock *sk, int size_goal, int flags,
976 struct page *page, int offset, size_t *size)
978 struct sk_buff *skb = tcp_write_queue_tail(sk);
979 struct tcp_sock *tp = tcp_sk(sk);
983 if (!skb || (copy = size_goal - skb->len) <= 0 ||
984 !tcp_skb_can_collapse_to(skb)) {
986 if (!sk_stream_memory_free(sk))
989 skb = tcp_stream_alloc_skb(sk, 0, sk->sk_allocation,
990 tcp_rtx_and_write_queues_empty(sk));
994 #ifdef CONFIG_TLS_DEVICE
995 skb->decrypted = !!(flags & MSG_SENDPAGE_DECRYPTED);
997 tcp_skb_entail(sk, skb);
1004 i = skb_shinfo(skb)->nr_frags;
1005 can_coalesce = skb_can_coalesce(skb, i, page, offset);
1006 if (!can_coalesce && i >= READ_ONCE(sysctl_max_skb_frags)) {
1007 tcp_mark_push(tp, skb);
1010 if (tcp_downgrade_zcopy_pure(sk, skb))
1013 copy = tcp_wmem_schedule(sk, copy);
1018 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1021 skb_fill_page_desc_noacc(skb, i, page, offset, copy);
1024 if (!(flags & MSG_NO_SHARED_FRAGS))
1025 skb_shinfo(skb)->flags |= SKBFL_SHARED_FRAG;
1028 skb->data_len += copy;
1029 skb->truesize += copy;
1030 sk_wmem_queued_add(sk, copy);
1031 sk_mem_charge(sk, copy);
1032 WRITE_ONCE(tp->write_seq, tp->write_seq + copy);
1033 TCP_SKB_CB(skb)->end_seq += copy;
1034 tcp_skb_pcount_set(skb, 0);
1040 ssize_t do_tcp_sendpages(struct sock *sk, struct page *page, int offset,
1041 size_t size, int flags)
1043 struct tcp_sock *tp = tcp_sk(sk);
1044 int mss_now, size_goal;
1047 long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
1049 if (IS_ENABLED(CONFIG_DEBUG_VM) &&
1050 WARN_ONCE(!sendpage_ok(page),
1051 "page must not be a Slab one and have page_count > 0"))
1054 /* Wait for a connection to finish. One exception is TCP Fast Open
1055 * (passive side) where data is allowed to be sent before a connection
1056 * is fully established.
1058 if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) &&
1059 !tcp_passive_fastopen(sk)) {
1060 err = sk_stream_wait_connect(sk, &timeo);
1065 sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
1067 mss_now = tcp_send_mss(sk, &size_goal, flags);
1071 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
1075 struct sk_buff *skb;
1078 skb = tcp_build_frag(sk, size_goal, flags, page, offset, ©);
1080 goto wait_for_space;
1083 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
1091 if (skb->len < size_goal || (flags & MSG_OOB))
1094 if (forced_push(tp)) {
1095 tcp_mark_push(tp, skb);
1096 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
1097 } else if (skb == tcp_send_head(sk))
1098 tcp_push_one(sk, mss_now);
1102 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1103 tcp_push(sk, flags & ~MSG_MORE, mss_now,
1104 TCP_NAGLE_PUSH, size_goal);
1106 err = sk_stream_wait_memory(sk, &timeo);
1110 mss_now = tcp_send_mss(sk, &size_goal, flags);
1115 tcp_tx_timestamp(sk, sk->sk_tsflags);
1116 if (!(flags & MSG_SENDPAGE_NOTLAST))
1117 tcp_push(sk, flags, mss_now, tp->nonagle, size_goal);
1122 tcp_remove_empty_skb(sk);
1126 /* make sure we wake any epoll edge trigger waiter */
1127 if (unlikely(tcp_rtx_and_write_queues_empty(sk) && err == -EAGAIN)) {
1128 sk->sk_write_space(sk);
1129 tcp_chrono_stop(sk, TCP_CHRONO_SNDBUF_LIMITED);
1131 return sk_stream_error(sk, flags, err);
1133 EXPORT_SYMBOL_GPL(do_tcp_sendpages);
1135 int tcp_sendpage_locked(struct sock *sk, struct page *page, int offset,
1136 size_t size, int flags)
1138 if (!(sk->sk_route_caps & NETIF_F_SG))
1139 return sock_no_sendpage_locked(sk, page, offset, size, flags);
1141 tcp_rate_check_app_limited(sk); /* is sending application-limited? */
1143 return do_tcp_sendpages(sk, page, offset, size, flags);
1145 EXPORT_SYMBOL_GPL(tcp_sendpage_locked);
1147 int tcp_sendpage(struct sock *sk, struct page *page, int offset,
1148 size_t size, int flags)
1153 ret = tcp_sendpage_locked(sk, page, offset, size, flags);
1158 EXPORT_SYMBOL(tcp_sendpage);
1160 void tcp_free_fastopen_req(struct tcp_sock *tp)
1162 if (tp->fastopen_req) {
1163 kfree(tp->fastopen_req);
1164 tp->fastopen_req = NULL;
1168 int tcp_sendmsg_fastopen(struct sock *sk, struct msghdr *msg, int *copied,
1169 size_t size, struct ubuf_info *uarg)
1171 struct tcp_sock *tp = tcp_sk(sk);
1172 struct inet_sock *inet = inet_sk(sk);
1173 struct sockaddr *uaddr = msg->msg_name;
1176 if (!(READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_fastopen) &
1177 TFO_CLIENT_ENABLE) ||
1178 (uaddr && msg->msg_namelen >= sizeof(uaddr->sa_family) &&
1179 uaddr->sa_family == AF_UNSPEC))
1181 if (tp->fastopen_req)
1182 return -EALREADY; /* Another Fast Open is in progress */
1184 tp->fastopen_req = kzalloc(sizeof(struct tcp_fastopen_request),
1186 if (unlikely(!tp->fastopen_req))
1188 tp->fastopen_req->data = msg;
1189 tp->fastopen_req->size = size;
1190 tp->fastopen_req->uarg = uarg;
1192 if (inet->defer_connect) {
1193 err = tcp_connect(sk);
1194 /* Same failure procedure as in tcp_v4/6_connect */
1196 tcp_set_state(sk, TCP_CLOSE);
1197 inet->inet_dport = 0;
1198 sk->sk_route_caps = 0;
1201 flags = (msg->msg_flags & MSG_DONTWAIT) ? O_NONBLOCK : 0;
1202 err = __inet_stream_connect(sk->sk_socket, uaddr,
1203 msg->msg_namelen, flags, 1);
1204 /* fastopen_req could already be freed in __inet_stream_connect
1205 * if the connection times out or gets rst
1207 if (tp->fastopen_req) {
1208 *copied = tp->fastopen_req->copied;
1209 tcp_free_fastopen_req(tp);
1210 inet->defer_connect = 0;
1215 int tcp_sendmsg_locked(struct sock *sk, struct msghdr *msg, size_t size)
1217 struct tcp_sock *tp = tcp_sk(sk);
1218 struct ubuf_info *uarg = NULL;
1219 struct sk_buff *skb;
1220 struct sockcm_cookie sockc;
1221 int flags, err, copied = 0;
1222 int mss_now = 0, size_goal, copied_syn = 0;
1223 int process_backlog = 0;
1227 flags = msg->msg_flags;
1229 if ((flags & MSG_ZEROCOPY) && size) {
1230 skb = tcp_write_queue_tail(sk);
1232 if (msg->msg_ubuf) {
1233 uarg = msg->msg_ubuf;
1234 net_zcopy_get(uarg);
1235 zc = sk->sk_route_caps & NETIF_F_SG;
1236 } else if (sock_flag(sk, SOCK_ZEROCOPY)) {
1237 uarg = msg_zerocopy_realloc(sk, size, skb_zcopy(skb));
1242 zc = sk->sk_route_caps & NETIF_F_SG;
1244 uarg_to_msgzc(uarg)->zerocopy = 0;
1248 if (unlikely(flags & MSG_FASTOPEN || inet_sk(sk)->defer_connect) &&
1250 err = tcp_sendmsg_fastopen(sk, msg, &copied_syn, size, uarg);
1251 if (err == -EINPROGRESS && copied_syn > 0)
1257 timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
1259 tcp_rate_check_app_limited(sk); /* is sending application-limited? */
1261 /* Wait for a connection to finish. One exception is TCP Fast Open
1262 * (passive side) where data is allowed to be sent before a connection
1263 * is fully established.
1265 if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) &&
1266 !tcp_passive_fastopen(sk)) {
1267 err = sk_stream_wait_connect(sk, &timeo);
1272 if (unlikely(tp->repair)) {
1273 if (tp->repair_queue == TCP_RECV_QUEUE) {
1274 copied = tcp_send_rcvq(sk, msg, size);
1279 if (tp->repair_queue == TCP_NO_QUEUE)
1282 /* 'common' sending to sendq */
1285 sockcm_init(&sockc, sk);
1286 if (msg->msg_controllen) {
1287 err = sock_cmsg_send(sk, msg, &sockc);
1288 if (unlikely(err)) {
1294 /* This should be in poll */
1295 sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
1297 /* Ok commence sending. */
1301 mss_now = tcp_send_mss(sk, &size_goal, flags);
1304 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
1307 while (msg_data_left(msg)) {
1310 skb = tcp_write_queue_tail(sk);
1312 copy = size_goal - skb->len;
1314 if (copy <= 0 || !tcp_skb_can_collapse_to(skb)) {
1318 if (!sk_stream_memory_free(sk))
1319 goto wait_for_space;
1321 if (unlikely(process_backlog >= 16)) {
1322 process_backlog = 0;
1323 if (sk_flush_backlog(sk))
1326 first_skb = tcp_rtx_and_write_queues_empty(sk);
1327 skb = tcp_stream_alloc_skb(sk, 0, sk->sk_allocation,
1330 goto wait_for_space;
1334 tcp_skb_entail(sk, skb);
1337 /* All packets are restored as if they have
1338 * already been sent. skb_mstamp_ns isn't set to
1339 * avoid wrong rtt estimation.
1342 TCP_SKB_CB(skb)->sacked |= TCPCB_REPAIRED;
1345 /* Try to append data to the end of skb. */
1346 if (copy > msg_data_left(msg))
1347 copy = msg_data_left(msg);
1351 int i = skb_shinfo(skb)->nr_frags;
1352 struct page_frag *pfrag = sk_page_frag(sk);
1354 if (!sk_page_frag_refill(sk, pfrag))
1355 goto wait_for_space;
1357 if (!skb_can_coalesce(skb, i, pfrag->page,
1359 if (i >= READ_ONCE(sysctl_max_skb_frags)) {
1360 tcp_mark_push(tp, skb);
1366 copy = min_t(int, copy, pfrag->size - pfrag->offset);
1368 if (unlikely(skb_zcopy_pure(skb) || skb_zcopy_managed(skb))) {
1369 if (tcp_downgrade_zcopy_pure(sk, skb))
1370 goto wait_for_space;
1371 skb_zcopy_downgrade_managed(skb);
1374 copy = tcp_wmem_schedule(sk, copy);
1376 goto wait_for_space;
1378 err = skb_copy_to_page_nocache(sk, &msg->msg_iter, skb,
1385 /* Update the skb. */
1387 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1389 skb_fill_page_desc(skb, i, pfrag->page,
1390 pfrag->offset, copy);
1391 page_ref_inc(pfrag->page);
1393 pfrag->offset += copy;
1395 /* First append to a fragless skb builds initial
1399 skb_shinfo(skb)->flags |= SKBFL_PURE_ZEROCOPY;
1401 if (!skb_zcopy_pure(skb)) {
1402 copy = tcp_wmem_schedule(sk, copy);
1404 goto wait_for_space;
1407 err = skb_zerocopy_iter_stream(sk, skb, msg, copy, uarg);
1408 if (err == -EMSGSIZE || err == -EEXIST) {
1409 tcp_mark_push(tp, skb);
1418 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
1420 WRITE_ONCE(tp->write_seq, tp->write_seq + copy);
1421 TCP_SKB_CB(skb)->end_seq += copy;
1422 tcp_skb_pcount_set(skb, 0);
1425 if (!msg_data_left(msg)) {
1426 if (unlikely(flags & MSG_EOR))
1427 TCP_SKB_CB(skb)->eor = 1;
1431 if (skb->len < size_goal || (flags & MSG_OOB) || unlikely(tp->repair))
1434 if (forced_push(tp)) {
1435 tcp_mark_push(tp, skb);
1436 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
1437 } else if (skb == tcp_send_head(sk))
1438 tcp_push_one(sk, mss_now);
1442 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1444 tcp_push(sk, flags & ~MSG_MORE, mss_now,
1445 TCP_NAGLE_PUSH, size_goal);
1447 err = sk_stream_wait_memory(sk, &timeo);
1451 mss_now = tcp_send_mss(sk, &size_goal, flags);
1456 tcp_tx_timestamp(sk, sockc.tsflags);
1457 tcp_push(sk, flags, mss_now, tp->nonagle, size_goal);
1460 net_zcopy_put(uarg);
1461 return copied + copied_syn;
1464 tcp_remove_empty_skb(sk);
1466 if (copied + copied_syn)
1469 net_zcopy_put_abort(uarg, true);
1470 err = sk_stream_error(sk, flags, err);
1471 /* make sure we wake any epoll edge trigger waiter */
1472 if (unlikely(tcp_rtx_and_write_queues_empty(sk) && err == -EAGAIN)) {
1473 sk->sk_write_space(sk);
1474 tcp_chrono_stop(sk, TCP_CHRONO_SNDBUF_LIMITED);
1478 EXPORT_SYMBOL_GPL(tcp_sendmsg_locked);
1480 int tcp_sendmsg(struct sock *sk, struct msghdr *msg, size_t size)
1485 ret = tcp_sendmsg_locked(sk, msg, size);
1490 EXPORT_SYMBOL(tcp_sendmsg);
1493 * Handle reading urgent data. BSD has very simple semantics for
1494 * this, no blocking and very strange errors 8)
1497 static int tcp_recv_urg(struct sock *sk, struct msghdr *msg, int len, int flags)
1499 struct tcp_sock *tp = tcp_sk(sk);
1501 /* No URG data to read. */
1502 if (sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data ||
1503 tp->urg_data == TCP_URG_READ)
1504 return -EINVAL; /* Yes this is right ! */
1506 if (sk->sk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DONE))
1509 if (tp->urg_data & TCP_URG_VALID) {
1511 char c = tp->urg_data;
1513 if (!(flags & MSG_PEEK))
1514 WRITE_ONCE(tp->urg_data, TCP_URG_READ);
1516 /* Read urgent data. */
1517 msg->msg_flags |= MSG_OOB;
1520 if (!(flags & MSG_TRUNC))
1521 err = memcpy_to_msg(msg, &c, 1);
1524 msg->msg_flags |= MSG_TRUNC;
1526 return err ? -EFAULT : len;
1529 if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
1532 /* Fixed the recv(..., MSG_OOB) behaviour. BSD docs and
1533 * the available implementations agree in this case:
1534 * this call should never block, independent of the
1535 * blocking state of the socket.
1536 * Mike <pall@rz.uni-karlsruhe.de>
1541 static int tcp_peek_sndq(struct sock *sk, struct msghdr *msg, int len)
1543 struct sk_buff *skb;
1544 int copied = 0, err = 0;
1546 /* XXX -- need to support SO_PEEK_OFF */
1548 skb_rbtree_walk(skb, &sk->tcp_rtx_queue) {
1549 err = skb_copy_datagram_msg(skb, 0, msg, skb->len);
1555 skb_queue_walk(&sk->sk_write_queue, skb) {
1556 err = skb_copy_datagram_msg(skb, 0, msg, skb->len);
1563 return err ?: copied;
1566 /* Clean up the receive buffer for full frames taken by the user,
1567 * then send an ACK if necessary. COPIED is the number of bytes
1568 * tcp_recvmsg has given to the user so far, it speeds up the
1569 * calculation of whether or not we must ACK for the sake of
1572 static void __tcp_cleanup_rbuf(struct sock *sk, int copied)
1574 struct tcp_sock *tp = tcp_sk(sk);
1575 bool time_to_ack = false;
1577 if (inet_csk_ack_scheduled(sk)) {
1578 const struct inet_connection_sock *icsk = inet_csk(sk);
1580 if (/* Once-per-two-segments ACK was not sent by tcp_input.c */
1581 tp->rcv_nxt - tp->rcv_wup > icsk->icsk_ack.rcv_mss ||
1583 * If this read emptied read buffer, we send ACK, if
1584 * connection is not bidirectional, user drained
1585 * receive buffer and there was a small segment
1589 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED2) ||
1590 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED) &&
1591 !inet_csk_in_pingpong_mode(sk))) &&
1592 !atomic_read(&sk->sk_rmem_alloc)))
1596 /* We send an ACK if we can now advertise a non-zero window
1597 * which has been raised "significantly".
1599 * Even if window raised up to infinity, do not send window open ACK
1600 * in states, where we will not receive more. It is useless.
1602 if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) {
1603 __u32 rcv_window_now = tcp_receive_window(tp);
1605 /* Optimize, __tcp_select_window() is not cheap. */
1606 if (2*rcv_window_now <= tp->window_clamp) {
1607 __u32 new_window = __tcp_select_window(sk);
1609 /* Send ACK now, if this read freed lots of space
1610 * in our buffer. Certainly, new_window is new window.
1611 * We can advertise it now, if it is not less than current one.
1612 * "Lots" means "at least twice" here.
1614 if (new_window && new_window >= 2 * rcv_window_now)
1622 void tcp_cleanup_rbuf(struct sock *sk, int copied)
1624 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
1625 struct tcp_sock *tp = tcp_sk(sk);
1627 WARN(skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq),
1628 "cleanup rbuf bug: copied %X seq %X rcvnxt %X\n",
1629 tp->copied_seq, TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt);
1630 __tcp_cleanup_rbuf(sk, copied);
1633 static void tcp_eat_recv_skb(struct sock *sk, struct sk_buff *skb)
1635 __skb_unlink(skb, &sk->sk_receive_queue);
1636 if (likely(skb->destructor == sock_rfree)) {
1638 skb->destructor = NULL;
1640 return skb_attempt_defer_free(skb);
1645 struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off)
1647 struct sk_buff *skb;
1650 while ((skb = skb_peek(&sk->sk_receive_queue)) != NULL) {
1651 offset = seq - TCP_SKB_CB(skb)->seq;
1652 if (unlikely(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
1653 pr_err_once("%s: found a SYN, please report !\n", __func__);
1656 if (offset < skb->len || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)) {
1660 /* This looks weird, but this can happen if TCP collapsing
1661 * splitted a fat GRO packet, while we released socket lock
1662 * in skb_splice_bits()
1664 tcp_eat_recv_skb(sk, skb);
1668 EXPORT_SYMBOL(tcp_recv_skb);
1671 * This routine provides an alternative to tcp_recvmsg() for routines
1672 * that would like to handle copying from skbuffs directly in 'sendfile'
1675 * - It is assumed that the socket was locked by the caller.
1676 * - The routine does not block.
1677 * - At present, there is no support for reading OOB data
1678 * or for 'peeking' the socket using this routine
1679 * (although both would be easy to implement).
1681 int tcp_read_sock(struct sock *sk, read_descriptor_t *desc,
1682 sk_read_actor_t recv_actor)
1684 struct sk_buff *skb;
1685 struct tcp_sock *tp = tcp_sk(sk);
1686 u32 seq = tp->copied_seq;
1690 if (sk->sk_state == TCP_LISTEN)
1692 while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1693 if (offset < skb->len) {
1697 len = skb->len - offset;
1698 /* Stop reading if we hit a patch of urgent data */
1699 if (unlikely(tp->urg_data)) {
1700 u32 urg_offset = tp->urg_seq - seq;
1701 if (urg_offset < len)
1706 used = recv_actor(desc, skb, offset, len);
1712 if (WARN_ON_ONCE(used > len))
1718 /* If recv_actor drops the lock (e.g. TCP splice
1719 * receive) the skb pointer might be invalid when
1720 * getting here: tcp_collapse might have deleted it
1721 * while aggregating skbs from the socket queue.
1723 skb = tcp_recv_skb(sk, seq - 1, &offset);
1726 /* TCP coalescing might have appended data to the skb.
1727 * Try to splice more frags
1729 if (offset + 1 != skb->len)
1732 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) {
1733 tcp_eat_recv_skb(sk, skb);
1737 tcp_eat_recv_skb(sk, skb);
1740 WRITE_ONCE(tp->copied_seq, seq);
1742 WRITE_ONCE(tp->copied_seq, seq);
1744 tcp_rcv_space_adjust(sk);
1746 /* Clean up data we have read: This will do ACK frames. */
1748 tcp_recv_skb(sk, seq, &offset);
1749 tcp_cleanup_rbuf(sk, copied);
1753 EXPORT_SYMBOL(tcp_read_sock);
1755 int tcp_read_skb(struct sock *sk, skb_read_actor_t recv_actor)
1757 struct tcp_sock *tp = tcp_sk(sk);
1758 u32 seq = tp->copied_seq;
1759 struct sk_buff *skb;
1763 if (sk->sk_state == TCP_LISTEN)
1766 while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1770 __skb_unlink(skb, &sk->sk_receive_queue);
1771 WARN_ON_ONCE(!skb_set_owner_sk_safe(skb, sk));
1772 tcp_flags = TCP_SKB_CB(skb)->tcp_flags;
1773 used = recv_actor(sk, skb);
1783 if (tcp_flags & TCPHDR_FIN) {
1788 WRITE_ONCE(tp->copied_seq, seq);
1790 tcp_rcv_space_adjust(sk);
1792 /* Clean up data we have read: This will do ACK frames. */
1794 __tcp_cleanup_rbuf(sk, copied);
1798 EXPORT_SYMBOL(tcp_read_skb);
1800 void tcp_read_done(struct sock *sk, size_t len)
1802 struct tcp_sock *tp = tcp_sk(sk);
1803 u32 seq = tp->copied_seq;
1804 struct sk_buff *skb;
1808 if (sk->sk_state == TCP_LISTEN)
1812 while (left && (skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1815 used = min_t(size_t, skb->len - offset, left);
1819 if (skb->len > offset + used)
1822 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) {
1823 tcp_eat_recv_skb(sk, skb);
1827 tcp_eat_recv_skb(sk, skb);
1829 WRITE_ONCE(tp->copied_seq, seq);
1831 tcp_rcv_space_adjust(sk);
1833 /* Clean up data we have read: This will do ACK frames. */
1835 tcp_cleanup_rbuf(sk, len - left);
1837 EXPORT_SYMBOL(tcp_read_done);
1839 int tcp_peek_len(struct socket *sock)
1841 return tcp_inq(sock->sk);
1843 EXPORT_SYMBOL(tcp_peek_len);
1845 /* Make sure sk_rcvbuf is big enough to satisfy SO_RCVLOWAT hint */
1846 int tcp_set_rcvlowat(struct sock *sk, int val)
1850 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK)
1851 cap = sk->sk_rcvbuf >> 1;
1853 cap = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_rmem[2]) >> 1;
1854 val = min(val, cap);
1855 WRITE_ONCE(sk->sk_rcvlowat, val ? : 1);
1857 /* Check if we need to signal EPOLLIN right now */
1860 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK)
1864 if (val > sk->sk_rcvbuf) {
1865 WRITE_ONCE(sk->sk_rcvbuf, val);
1866 tcp_sk(sk)->window_clamp = tcp_win_from_space(sk, val);
1870 EXPORT_SYMBOL(tcp_set_rcvlowat);
1872 void tcp_update_recv_tstamps(struct sk_buff *skb,
1873 struct scm_timestamping_internal *tss)
1876 tss->ts[0] = ktime_to_timespec64(skb->tstamp);
1878 tss->ts[0] = (struct timespec64) {0};
1880 if (skb_hwtstamps(skb)->hwtstamp)
1881 tss->ts[2] = ktime_to_timespec64(skb_hwtstamps(skb)->hwtstamp);
1883 tss->ts[2] = (struct timespec64) {0};
1887 static const struct vm_operations_struct tcp_vm_ops = {
1890 int tcp_mmap(struct file *file, struct socket *sock,
1891 struct vm_area_struct *vma)
1893 if (vma->vm_flags & (VM_WRITE | VM_EXEC))
1895 vm_flags_clear(vma, VM_MAYWRITE | VM_MAYEXEC);
1897 /* Instruct vm_insert_page() to not mmap_read_lock(mm) */
1898 vm_flags_set(vma, VM_MIXEDMAP);
1900 vma->vm_ops = &tcp_vm_ops;
1903 EXPORT_SYMBOL(tcp_mmap);
1905 static skb_frag_t *skb_advance_to_frag(struct sk_buff *skb, u32 offset_skb,
1910 if (unlikely(offset_skb >= skb->len))
1913 offset_skb -= skb_headlen(skb);
1914 if ((int)offset_skb < 0 || skb_has_frag_list(skb))
1917 frag = skb_shinfo(skb)->frags;
1918 while (offset_skb) {
1919 if (skb_frag_size(frag) > offset_skb) {
1920 *offset_frag = offset_skb;
1923 offset_skb -= skb_frag_size(frag);
1930 static bool can_map_frag(const skb_frag_t *frag)
1932 return skb_frag_size(frag) == PAGE_SIZE && !skb_frag_off(frag);
1935 static int find_next_mappable_frag(const skb_frag_t *frag,
1936 int remaining_in_skb)
1940 if (likely(can_map_frag(frag)))
1943 while (offset < remaining_in_skb && !can_map_frag(frag)) {
1944 offset += skb_frag_size(frag);
1950 static void tcp_zerocopy_set_hint_for_skb(struct sock *sk,
1951 struct tcp_zerocopy_receive *zc,
1952 struct sk_buff *skb, u32 offset)
1954 u32 frag_offset, partial_frag_remainder = 0;
1955 int mappable_offset;
1958 /* worst case: skip to next skb. try to improve on this case below */
1959 zc->recv_skip_hint = skb->len - offset;
1961 /* Find the frag containing this offset (and how far into that frag) */
1962 frag = skb_advance_to_frag(skb, offset, &frag_offset);
1967 struct skb_shared_info *info = skb_shinfo(skb);
1969 /* We read part of the last frag, must recvmsg() rest of skb. */
1970 if (frag == &info->frags[info->nr_frags - 1])
1973 /* Else, we must at least read the remainder in this frag. */
1974 partial_frag_remainder = skb_frag_size(frag) - frag_offset;
1975 zc->recv_skip_hint -= partial_frag_remainder;
1979 /* partial_frag_remainder: If part way through a frag, must read rest.
1980 * mappable_offset: Bytes till next mappable frag, *not* counting bytes
1981 * in partial_frag_remainder.
1983 mappable_offset = find_next_mappable_frag(frag, zc->recv_skip_hint);
1984 zc->recv_skip_hint = mappable_offset + partial_frag_remainder;
1987 static int tcp_recvmsg_locked(struct sock *sk, struct msghdr *msg, size_t len,
1988 int flags, struct scm_timestamping_internal *tss,
1990 static int receive_fallback_to_copy(struct sock *sk,
1991 struct tcp_zerocopy_receive *zc, int inq,
1992 struct scm_timestamping_internal *tss)
1994 unsigned long copy_address = (unsigned long)zc->copybuf_address;
1995 struct msghdr msg = {};
2000 zc->recv_skip_hint = 0;
2002 if (copy_address != zc->copybuf_address)
2005 err = import_single_range(ITER_DEST, (void __user *)copy_address,
2006 inq, &iov, &msg.msg_iter);
2010 err = tcp_recvmsg_locked(sk, &msg, inq, MSG_DONTWAIT,
2011 tss, &zc->msg_flags);
2015 zc->copybuf_len = err;
2016 if (likely(zc->copybuf_len)) {
2017 struct sk_buff *skb;
2020 skb = tcp_recv_skb(sk, tcp_sk(sk)->copied_seq, &offset);
2022 tcp_zerocopy_set_hint_for_skb(sk, zc, skb, offset);
2027 static int tcp_copy_straggler_data(struct tcp_zerocopy_receive *zc,
2028 struct sk_buff *skb, u32 copylen,
2029 u32 *offset, u32 *seq)
2031 unsigned long copy_address = (unsigned long)zc->copybuf_address;
2032 struct msghdr msg = {};
2036 if (copy_address != zc->copybuf_address)
2039 err = import_single_range(ITER_DEST, (void __user *)copy_address,
2040 copylen, &iov, &msg.msg_iter);
2043 err = skb_copy_datagram_msg(skb, *offset, &msg, copylen);
2046 zc->recv_skip_hint -= copylen;
2049 return (__s32)copylen;
2052 static int tcp_zc_handle_leftover(struct tcp_zerocopy_receive *zc,
2054 struct sk_buff *skb,
2057 struct scm_timestamping_internal *tss)
2059 u32 offset, copylen = min_t(u32, copybuf_len, zc->recv_skip_hint);
2063 /* skb is null if inq < PAGE_SIZE. */
2065 offset = *seq - TCP_SKB_CB(skb)->seq;
2067 skb = tcp_recv_skb(sk, *seq, &offset);
2068 if (TCP_SKB_CB(skb)->has_rxtstamp) {
2069 tcp_update_recv_tstamps(skb, tss);
2070 zc->msg_flags |= TCP_CMSG_TS;
2074 zc->copybuf_len = tcp_copy_straggler_data(zc, skb, copylen, &offset,
2076 return zc->copybuf_len < 0 ? 0 : copylen;
2079 static int tcp_zerocopy_vm_insert_batch_error(struct vm_area_struct *vma,
2080 struct page **pending_pages,
2081 unsigned long pages_remaining,
2082 unsigned long *address,
2085 struct tcp_zerocopy_receive *zc,
2086 u32 total_bytes_to_map,
2089 /* At least one page did not map. Try zapping if we skipped earlier. */
2090 if (err == -EBUSY &&
2091 zc->flags & TCP_RECEIVE_ZEROCOPY_FLAG_TLB_CLEAN_HINT) {
2094 maybe_zap_len = total_bytes_to_map - /* All bytes to map */
2095 *length + /* Mapped or pending */
2096 (pages_remaining * PAGE_SIZE); /* Failed map. */
2097 zap_page_range_single(vma, *address, maybe_zap_len, NULL);
2102 unsigned long leftover_pages = pages_remaining;
2105 /* We called zap_page_range_single, try to reinsert. */
2106 err = vm_insert_pages(vma, *address,
2109 bytes_mapped = PAGE_SIZE * (leftover_pages - pages_remaining);
2110 *seq += bytes_mapped;
2111 *address += bytes_mapped;
2114 /* Either we were unable to zap, OR we zapped, retried an
2115 * insert, and still had an issue. Either ways, pages_remaining
2116 * is the number of pages we were unable to map, and we unroll
2117 * some state we speculatively touched before.
2119 const int bytes_not_mapped = PAGE_SIZE * pages_remaining;
2121 *length -= bytes_not_mapped;
2122 zc->recv_skip_hint += bytes_not_mapped;
2127 static int tcp_zerocopy_vm_insert_batch(struct vm_area_struct *vma,
2128 struct page **pages,
2129 unsigned int pages_to_map,
2130 unsigned long *address,
2133 struct tcp_zerocopy_receive *zc,
2134 u32 total_bytes_to_map)
2136 unsigned long pages_remaining = pages_to_map;
2137 unsigned int pages_mapped;
2138 unsigned int bytes_mapped;
2141 err = vm_insert_pages(vma, *address, pages, &pages_remaining);
2142 pages_mapped = pages_to_map - (unsigned int)pages_remaining;
2143 bytes_mapped = PAGE_SIZE * pages_mapped;
2144 /* Even if vm_insert_pages fails, it may have partially succeeded in
2145 * mapping (some but not all of the pages).
2147 *seq += bytes_mapped;
2148 *address += bytes_mapped;
2153 /* Error: maybe zap and retry + rollback state for failed inserts. */
2154 return tcp_zerocopy_vm_insert_batch_error(vma, pages + pages_mapped,
2155 pages_remaining, address, length, seq, zc, total_bytes_to_map,
2159 #define TCP_VALID_ZC_MSG_FLAGS (TCP_CMSG_TS)
2160 static void tcp_zc_finalize_rx_tstamp(struct sock *sk,
2161 struct tcp_zerocopy_receive *zc,
2162 struct scm_timestamping_internal *tss)
2164 unsigned long msg_control_addr;
2165 struct msghdr cmsg_dummy;
2167 msg_control_addr = (unsigned long)zc->msg_control;
2168 cmsg_dummy.msg_control_user = (void __user *)msg_control_addr;
2169 cmsg_dummy.msg_controllen =
2170 (__kernel_size_t)zc->msg_controllen;
2171 cmsg_dummy.msg_flags = in_compat_syscall()
2172 ? MSG_CMSG_COMPAT : 0;
2173 cmsg_dummy.msg_control_is_user = true;
2175 if (zc->msg_control == msg_control_addr &&
2176 zc->msg_controllen == cmsg_dummy.msg_controllen) {
2177 tcp_recv_timestamp(&cmsg_dummy, sk, tss);
2178 zc->msg_control = (__u64)
2179 ((uintptr_t)cmsg_dummy.msg_control_user);
2180 zc->msg_controllen =
2181 (__u64)cmsg_dummy.msg_controllen;
2182 zc->msg_flags = (__u32)cmsg_dummy.msg_flags;
2186 #define TCP_ZEROCOPY_PAGE_BATCH_SIZE 32
2187 static int tcp_zerocopy_receive(struct sock *sk,
2188 struct tcp_zerocopy_receive *zc,
2189 struct scm_timestamping_internal *tss)
2191 u32 length = 0, offset, vma_len, avail_len, copylen = 0;
2192 unsigned long address = (unsigned long)zc->address;
2193 struct page *pages[TCP_ZEROCOPY_PAGE_BATCH_SIZE];
2194 s32 copybuf_len = zc->copybuf_len;
2195 struct tcp_sock *tp = tcp_sk(sk);
2196 const skb_frag_t *frags = NULL;
2197 unsigned int pages_to_map = 0;
2198 struct vm_area_struct *vma;
2199 struct sk_buff *skb = NULL;
2200 u32 seq = tp->copied_seq;
2201 u32 total_bytes_to_map;
2202 int inq = tcp_inq(sk);
2205 zc->copybuf_len = 0;
2208 if (address & (PAGE_SIZE - 1) || address != zc->address)
2211 if (sk->sk_state == TCP_LISTEN)
2214 sock_rps_record_flow(sk);
2216 if (inq && inq <= copybuf_len)
2217 return receive_fallback_to_copy(sk, zc, inq, tss);
2219 if (inq < PAGE_SIZE) {
2221 zc->recv_skip_hint = inq;
2222 if (!inq && sock_flag(sk, SOCK_DONE))
2227 mmap_read_lock(current->mm);
2229 vma = vma_lookup(current->mm, address);
2230 if (!vma || vma->vm_ops != &tcp_vm_ops) {
2231 mmap_read_unlock(current->mm);
2234 vma_len = min_t(unsigned long, zc->length, vma->vm_end - address);
2235 avail_len = min_t(u32, vma_len, inq);
2236 total_bytes_to_map = avail_len & ~(PAGE_SIZE - 1);
2237 if (total_bytes_to_map) {
2238 if (!(zc->flags & TCP_RECEIVE_ZEROCOPY_FLAG_TLB_CLEAN_HINT))
2239 zap_page_range_single(vma, address, total_bytes_to_map,
2241 zc->length = total_bytes_to_map;
2242 zc->recv_skip_hint = 0;
2244 zc->length = avail_len;
2245 zc->recv_skip_hint = avail_len;
2248 while (length + PAGE_SIZE <= zc->length) {
2249 int mappable_offset;
2252 if (zc->recv_skip_hint < PAGE_SIZE) {
2256 if (zc->recv_skip_hint > 0)
2259 offset = seq - TCP_SKB_CB(skb)->seq;
2261 skb = tcp_recv_skb(sk, seq, &offset);
2264 if (TCP_SKB_CB(skb)->has_rxtstamp) {
2265 tcp_update_recv_tstamps(skb, tss);
2266 zc->msg_flags |= TCP_CMSG_TS;
2268 zc->recv_skip_hint = skb->len - offset;
2269 frags = skb_advance_to_frag(skb, offset, &offset_frag);
2270 if (!frags || offset_frag)
2274 mappable_offset = find_next_mappable_frag(frags,
2275 zc->recv_skip_hint);
2276 if (mappable_offset) {
2277 zc->recv_skip_hint = mappable_offset;
2280 page = skb_frag_page(frags);
2282 pages[pages_to_map++] = page;
2283 length += PAGE_SIZE;
2284 zc->recv_skip_hint -= PAGE_SIZE;
2286 if (pages_to_map == TCP_ZEROCOPY_PAGE_BATCH_SIZE ||
2287 zc->recv_skip_hint < PAGE_SIZE) {
2288 /* Either full batch, or we're about to go to next skb
2289 * (and we cannot unroll failed ops across skbs).
2291 ret = tcp_zerocopy_vm_insert_batch(vma, pages,
2295 total_bytes_to_map);
2302 ret = tcp_zerocopy_vm_insert_batch(vma, pages, pages_to_map,
2303 &address, &length, &seq,
2304 zc, total_bytes_to_map);
2307 mmap_read_unlock(current->mm);
2308 /* Try to copy straggler data. */
2310 copylen = tcp_zc_handle_leftover(zc, sk, skb, &seq, copybuf_len, tss);
2312 if (length + copylen) {
2313 WRITE_ONCE(tp->copied_seq, seq);
2314 tcp_rcv_space_adjust(sk);
2316 /* Clean up data we have read: This will do ACK frames. */
2317 tcp_recv_skb(sk, seq, &offset);
2318 tcp_cleanup_rbuf(sk, length + copylen);
2320 if (length == zc->length)
2321 zc->recv_skip_hint = 0;
2323 if (!zc->recv_skip_hint && sock_flag(sk, SOCK_DONE))
2326 zc->length = length;
2331 /* Similar to __sock_recv_timestamp, but does not require an skb */
2332 void tcp_recv_timestamp(struct msghdr *msg, const struct sock *sk,
2333 struct scm_timestamping_internal *tss)
2335 int new_tstamp = sock_flag(sk, SOCK_TSTAMP_NEW);
2336 bool has_timestamping = false;
2338 if (tss->ts[0].tv_sec || tss->ts[0].tv_nsec) {
2339 if (sock_flag(sk, SOCK_RCVTSTAMP)) {
2340 if (sock_flag(sk, SOCK_RCVTSTAMPNS)) {
2342 struct __kernel_timespec kts = {
2343 .tv_sec = tss->ts[0].tv_sec,
2344 .tv_nsec = tss->ts[0].tv_nsec,
2346 put_cmsg(msg, SOL_SOCKET, SO_TIMESTAMPNS_NEW,
2349 struct __kernel_old_timespec ts_old = {
2350 .tv_sec = tss->ts[0].tv_sec,
2351 .tv_nsec = tss->ts[0].tv_nsec,
2353 put_cmsg(msg, SOL_SOCKET, SO_TIMESTAMPNS_OLD,
2354 sizeof(ts_old), &ts_old);
2358 struct __kernel_sock_timeval stv = {
2359 .tv_sec = tss->ts[0].tv_sec,
2360 .tv_usec = tss->ts[0].tv_nsec / 1000,
2362 put_cmsg(msg, SOL_SOCKET, SO_TIMESTAMP_NEW,
2365 struct __kernel_old_timeval tv = {
2366 .tv_sec = tss->ts[0].tv_sec,
2367 .tv_usec = tss->ts[0].tv_nsec / 1000,
2369 put_cmsg(msg, SOL_SOCKET, SO_TIMESTAMP_OLD,
2375 if (sk->sk_tsflags & SOF_TIMESTAMPING_SOFTWARE)
2376 has_timestamping = true;
2378 tss->ts[0] = (struct timespec64) {0};
2381 if (tss->ts[2].tv_sec || tss->ts[2].tv_nsec) {
2382 if (sk->sk_tsflags & SOF_TIMESTAMPING_RAW_HARDWARE)
2383 has_timestamping = true;
2385 tss->ts[2] = (struct timespec64) {0};
2388 if (has_timestamping) {
2389 tss->ts[1] = (struct timespec64) {0};
2390 if (sock_flag(sk, SOCK_TSTAMP_NEW))
2391 put_cmsg_scm_timestamping64(msg, tss);
2393 put_cmsg_scm_timestamping(msg, tss);
2397 static int tcp_inq_hint(struct sock *sk)
2399 const struct tcp_sock *tp = tcp_sk(sk);
2400 u32 copied_seq = READ_ONCE(tp->copied_seq);
2401 u32 rcv_nxt = READ_ONCE(tp->rcv_nxt);
2404 inq = rcv_nxt - copied_seq;
2405 if (unlikely(inq < 0 || copied_seq != READ_ONCE(tp->copied_seq))) {
2407 inq = tp->rcv_nxt - tp->copied_seq;
2410 /* After receiving a FIN, tell the user-space to continue reading
2411 * by returning a non-zero inq.
2413 if (inq == 0 && sock_flag(sk, SOCK_DONE))
2419 * This routine copies from a sock struct into the user buffer.
2421 * Technical note: in 2.3 we work on _locked_ socket, so that
2422 * tricks with *seq access order and skb->users are not required.
2423 * Probably, code can be easily improved even more.
2426 static int tcp_recvmsg_locked(struct sock *sk, struct msghdr *msg, size_t len,
2427 int flags, struct scm_timestamping_internal *tss,
2430 struct tcp_sock *tp = tcp_sk(sk);
2436 int target; /* Read at least this many bytes */
2438 struct sk_buff *skb, *last;
2442 if (sk->sk_state == TCP_LISTEN)
2445 if (tp->recvmsg_inq) {
2446 *cmsg_flags = TCP_CMSG_INQ;
2447 msg->msg_get_inq = 1;
2449 timeo = sock_rcvtimeo(sk, flags & MSG_DONTWAIT);
2451 /* Urgent data needs to be handled specially. */
2452 if (flags & MSG_OOB)
2455 if (unlikely(tp->repair)) {
2457 if (!(flags & MSG_PEEK))
2460 if (tp->repair_queue == TCP_SEND_QUEUE)
2464 if (tp->repair_queue == TCP_NO_QUEUE)
2467 /* 'common' recv queue MSG_PEEK-ing */
2470 seq = &tp->copied_seq;
2471 if (flags & MSG_PEEK) {
2472 peek_seq = tp->copied_seq;
2476 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
2481 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
2482 if (unlikely(tp->urg_data) && tp->urg_seq == *seq) {
2485 if (signal_pending(current)) {
2486 copied = timeo ? sock_intr_errno(timeo) : -EAGAIN;
2491 /* Next get a buffer. */
2493 last = skb_peek_tail(&sk->sk_receive_queue);
2494 skb_queue_walk(&sk->sk_receive_queue, skb) {
2496 /* Now that we have two receive queues this
2499 if (WARN(before(*seq, TCP_SKB_CB(skb)->seq),
2500 "TCP recvmsg seq # bug: copied %X, seq %X, rcvnxt %X, fl %X\n",
2501 *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt,
2505 offset = *seq - TCP_SKB_CB(skb)->seq;
2506 if (unlikely(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
2507 pr_err_once("%s: found a SYN, please report !\n", __func__);
2510 if (offset < skb->len)
2512 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
2514 WARN(!(flags & MSG_PEEK),
2515 "TCP recvmsg seq # bug 2: copied %X, seq %X, rcvnxt %X, fl %X\n",
2516 *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt, flags);
2519 /* Well, if we have backlog, try to process it now yet. */
2521 if (copied >= target && !READ_ONCE(sk->sk_backlog.tail))
2527 sk->sk_state == TCP_CLOSE ||
2528 (sk->sk_shutdown & RCV_SHUTDOWN) ||
2529 signal_pending(current))
2532 if (sock_flag(sk, SOCK_DONE))
2536 copied = sock_error(sk);
2540 if (sk->sk_shutdown & RCV_SHUTDOWN)
2543 if (sk->sk_state == TCP_CLOSE) {
2544 /* This occurs when user tries to read
2545 * from never connected socket.
2556 if (signal_pending(current)) {
2557 copied = sock_intr_errno(timeo);
2562 if (copied >= target) {
2563 /* Do not sleep, just process backlog. */
2564 __sk_flush_backlog(sk);
2566 tcp_cleanup_rbuf(sk, copied);
2567 sk_wait_data(sk, &timeo, last);
2570 if ((flags & MSG_PEEK) &&
2571 (peek_seq - copied - urg_hole != tp->copied_seq)) {
2572 net_dbg_ratelimited("TCP(%s:%d): Application bug, race in MSG_PEEK\n",
2574 task_pid_nr(current));
2575 peek_seq = tp->copied_seq;
2580 /* Ok so how much can we use? */
2581 used = skb->len - offset;
2585 /* Do we have urgent data here? */
2586 if (unlikely(tp->urg_data)) {
2587 u32 urg_offset = tp->urg_seq - *seq;
2588 if (urg_offset < used) {
2590 if (!sock_flag(sk, SOCK_URGINLINE)) {
2591 WRITE_ONCE(*seq, *seq + 1);
2603 if (!(flags & MSG_TRUNC)) {
2604 err = skb_copy_datagram_msg(skb, offset, msg, used);
2606 /* Exception. Bailout! */
2613 WRITE_ONCE(*seq, *seq + used);
2617 tcp_rcv_space_adjust(sk);
2620 if (unlikely(tp->urg_data) && after(tp->copied_seq, tp->urg_seq)) {
2621 WRITE_ONCE(tp->urg_data, 0);
2622 tcp_fast_path_check(sk);
2625 if (TCP_SKB_CB(skb)->has_rxtstamp) {
2626 tcp_update_recv_tstamps(skb, tss);
2627 *cmsg_flags |= TCP_CMSG_TS;
2630 if (used + offset < skb->len)
2633 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
2635 if (!(flags & MSG_PEEK))
2636 tcp_eat_recv_skb(sk, skb);
2640 /* Process the FIN. */
2641 WRITE_ONCE(*seq, *seq + 1);
2642 if (!(flags & MSG_PEEK))
2643 tcp_eat_recv_skb(sk, skb);
2647 /* According to UNIX98, msg_name/msg_namelen are ignored
2648 * on connected socket. I was just happy when found this 8) --ANK
2651 /* Clean up data we have read: This will do ACK frames. */
2652 tcp_cleanup_rbuf(sk, copied);
2659 err = tcp_recv_urg(sk, msg, len, flags);
2663 err = tcp_peek_sndq(sk, msg, len);
2667 int tcp_recvmsg(struct sock *sk, struct msghdr *msg, size_t len, int flags,
2670 int cmsg_flags = 0, ret;
2671 struct scm_timestamping_internal tss;
2673 if (unlikely(flags & MSG_ERRQUEUE))
2674 return inet_recv_error(sk, msg, len, addr_len);
2676 if (sk_can_busy_loop(sk) &&
2677 skb_queue_empty_lockless(&sk->sk_receive_queue) &&
2678 sk->sk_state == TCP_ESTABLISHED)
2679 sk_busy_loop(sk, flags & MSG_DONTWAIT);
2682 ret = tcp_recvmsg_locked(sk, msg, len, flags, &tss, &cmsg_flags);
2685 if ((cmsg_flags || msg->msg_get_inq) && ret >= 0) {
2686 if (cmsg_flags & TCP_CMSG_TS)
2687 tcp_recv_timestamp(msg, sk, &tss);
2688 if (msg->msg_get_inq) {
2689 msg->msg_inq = tcp_inq_hint(sk);
2690 if (cmsg_flags & TCP_CMSG_INQ)
2691 put_cmsg(msg, SOL_TCP, TCP_CM_INQ,
2692 sizeof(msg->msg_inq), &msg->msg_inq);
2697 EXPORT_SYMBOL(tcp_recvmsg);
2699 void tcp_set_state(struct sock *sk, int state)
2701 int oldstate = sk->sk_state;
2703 /* We defined a new enum for TCP states that are exported in BPF
2704 * so as not force the internal TCP states to be frozen. The
2705 * following checks will detect if an internal state value ever
2706 * differs from the BPF value. If this ever happens, then we will
2707 * need to remap the internal value to the BPF value before calling
2708 * tcp_call_bpf_2arg.
2710 BUILD_BUG_ON((int)BPF_TCP_ESTABLISHED != (int)TCP_ESTABLISHED);
2711 BUILD_BUG_ON((int)BPF_TCP_SYN_SENT != (int)TCP_SYN_SENT);
2712 BUILD_BUG_ON((int)BPF_TCP_SYN_RECV != (int)TCP_SYN_RECV);
2713 BUILD_BUG_ON((int)BPF_TCP_FIN_WAIT1 != (int)TCP_FIN_WAIT1);
2714 BUILD_BUG_ON((int)BPF_TCP_FIN_WAIT2 != (int)TCP_FIN_WAIT2);
2715 BUILD_BUG_ON((int)BPF_TCP_TIME_WAIT != (int)TCP_TIME_WAIT);
2716 BUILD_BUG_ON((int)BPF_TCP_CLOSE != (int)TCP_CLOSE);
2717 BUILD_BUG_ON((int)BPF_TCP_CLOSE_WAIT != (int)TCP_CLOSE_WAIT);
2718 BUILD_BUG_ON((int)BPF_TCP_LAST_ACK != (int)TCP_LAST_ACK);
2719 BUILD_BUG_ON((int)BPF_TCP_LISTEN != (int)TCP_LISTEN);
2720 BUILD_BUG_ON((int)BPF_TCP_CLOSING != (int)TCP_CLOSING);
2721 BUILD_BUG_ON((int)BPF_TCP_NEW_SYN_RECV != (int)TCP_NEW_SYN_RECV);
2722 BUILD_BUG_ON((int)BPF_TCP_MAX_STATES != (int)TCP_MAX_STATES);
2724 /* bpf uapi header bpf.h defines an anonymous enum with values
2725 * BPF_TCP_* used by bpf programs. Currently gcc built vmlinux
2726 * is able to emit this enum in DWARF due to the above BUILD_BUG_ON.
2727 * But clang built vmlinux does not have this enum in DWARF
2728 * since clang removes the above code before generating IR/debuginfo.
2729 * Let us explicitly emit the type debuginfo to ensure the
2730 * above-mentioned anonymous enum in the vmlinux DWARF and hence BTF
2731 * regardless of which compiler is used.
2733 BTF_TYPE_EMIT_ENUM(BPF_TCP_ESTABLISHED);
2735 if (BPF_SOCK_OPS_TEST_FLAG(tcp_sk(sk), BPF_SOCK_OPS_STATE_CB_FLAG))
2736 tcp_call_bpf_2arg(sk, BPF_SOCK_OPS_STATE_CB, oldstate, state);
2739 case TCP_ESTABLISHED:
2740 if (oldstate != TCP_ESTABLISHED)
2741 TCP_INC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
2745 if (oldstate == TCP_CLOSE_WAIT || oldstate == TCP_ESTABLISHED)
2746 TCP_INC_STATS(sock_net(sk), TCP_MIB_ESTABRESETS);
2748 sk->sk_prot->unhash(sk);
2749 if (inet_csk(sk)->icsk_bind_hash &&
2750 !(sk->sk_userlocks & SOCK_BINDPORT_LOCK))
2754 if (oldstate == TCP_ESTABLISHED)
2755 TCP_DEC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
2758 /* Change state AFTER socket is unhashed to avoid closed
2759 * socket sitting in hash tables.
2761 inet_sk_state_store(sk, state);
2763 EXPORT_SYMBOL_GPL(tcp_set_state);
2766 * State processing on a close. This implements the state shift for
2767 * sending our FIN frame. Note that we only send a FIN for some
2768 * states. A shutdown() may have already sent the FIN, or we may be
2772 static const unsigned char new_state[16] = {
2773 /* current state: new state: action: */
2774 [0 /* (Invalid) */] = TCP_CLOSE,
2775 [TCP_ESTABLISHED] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2776 [TCP_SYN_SENT] = TCP_CLOSE,
2777 [TCP_SYN_RECV] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2778 [TCP_FIN_WAIT1] = TCP_FIN_WAIT1,
2779 [TCP_FIN_WAIT2] = TCP_FIN_WAIT2,
2780 [TCP_TIME_WAIT] = TCP_CLOSE,
2781 [TCP_CLOSE] = TCP_CLOSE,
2782 [TCP_CLOSE_WAIT] = TCP_LAST_ACK | TCP_ACTION_FIN,
2783 [TCP_LAST_ACK] = TCP_LAST_ACK,
2784 [TCP_LISTEN] = TCP_CLOSE,
2785 [TCP_CLOSING] = TCP_CLOSING,
2786 [TCP_NEW_SYN_RECV] = TCP_CLOSE, /* should not happen ! */
2789 static int tcp_close_state(struct sock *sk)
2791 int next = (int)new_state[sk->sk_state];
2792 int ns = next & TCP_STATE_MASK;
2794 tcp_set_state(sk, ns);
2796 return next & TCP_ACTION_FIN;
2800 * Shutdown the sending side of a connection. Much like close except
2801 * that we don't receive shut down or sock_set_flag(sk, SOCK_DEAD).
2804 void tcp_shutdown(struct sock *sk, int how)
2806 /* We need to grab some memory, and put together a FIN,
2807 * and then put it into the queue to be sent.
2808 * Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
2810 if (!(how & SEND_SHUTDOWN))
2813 /* If we've already sent a FIN, or it's a closed state, skip this. */
2814 if ((1 << sk->sk_state) &
2815 (TCPF_ESTABLISHED | TCPF_SYN_SENT |
2816 TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) {
2817 /* Clear out any half completed packets. FIN if needed. */
2818 if (tcp_close_state(sk))
2822 EXPORT_SYMBOL(tcp_shutdown);
2824 int tcp_orphan_count_sum(void)
2828 for_each_possible_cpu(i)
2829 total += per_cpu(tcp_orphan_count, i);
2831 return max(total, 0);
2834 static int tcp_orphan_cache;
2835 static struct timer_list tcp_orphan_timer;
2836 #define TCP_ORPHAN_TIMER_PERIOD msecs_to_jiffies(100)
2838 static void tcp_orphan_update(struct timer_list *unused)
2840 WRITE_ONCE(tcp_orphan_cache, tcp_orphan_count_sum());
2841 mod_timer(&tcp_orphan_timer, jiffies + TCP_ORPHAN_TIMER_PERIOD);
2844 static bool tcp_too_many_orphans(int shift)
2846 return READ_ONCE(tcp_orphan_cache) << shift >
2847 READ_ONCE(sysctl_tcp_max_orphans);
2850 bool tcp_check_oom(struct sock *sk, int shift)
2852 bool too_many_orphans, out_of_socket_memory;
2854 too_many_orphans = tcp_too_many_orphans(shift);
2855 out_of_socket_memory = tcp_out_of_memory(sk);
2857 if (too_many_orphans)
2858 net_info_ratelimited("too many orphaned sockets\n");
2859 if (out_of_socket_memory)
2860 net_info_ratelimited("out of memory -- consider tuning tcp_mem\n");
2861 return too_many_orphans || out_of_socket_memory;
2864 void __tcp_close(struct sock *sk, long timeout)
2866 struct sk_buff *skb;
2867 int data_was_unread = 0;
2870 sk->sk_shutdown = SHUTDOWN_MASK;
2872 if (sk->sk_state == TCP_LISTEN) {
2873 tcp_set_state(sk, TCP_CLOSE);
2876 inet_csk_listen_stop(sk);
2878 goto adjudge_to_death;
2881 /* We need to flush the recv. buffs. We do this only on the
2882 * descriptor close, not protocol-sourced closes, because the
2883 * reader process may not have drained the data yet!
2885 while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) {
2886 u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq;
2888 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
2890 data_was_unread += len;
2894 /* If socket has been already reset (e.g. in tcp_reset()) - kill it. */
2895 if (sk->sk_state == TCP_CLOSE)
2896 goto adjudge_to_death;
2898 /* As outlined in RFC 2525, section 2.17, we send a RST here because
2899 * data was lost. To witness the awful effects of the old behavior of
2900 * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk
2901 * GET in an FTP client, suspend the process, wait for the client to
2902 * advertise a zero window, then kill -9 the FTP client, wheee...
2903 * Note: timeout is always zero in such a case.
2905 if (unlikely(tcp_sk(sk)->repair)) {
2906 sk->sk_prot->disconnect(sk, 0);
2907 } else if (data_was_unread) {
2908 /* Unread data was tossed, zap the connection. */
2909 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONCLOSE);
2910 tcp_set_state(sk, TCP_CLOSE);
2911 tcp_send_active_reset(sk, sk->sk_allocation);
2912 } else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) {
2913 /* Check zero linger _after_ checking for unread data. */
2914 sk->sk_prot->disconnect(sk, 0);
2915 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
2916 } else if (tcp_close_state(sk)) {
2917 /* We FIN if the application ate all the data before
2918 * zapping the connection.
2921 /* RED-PEN. Formally speaking, we have broken TCP state
2922 * machine. State transitions:
2924 * TCP_ESTABLISHED -> TCP_FIN_WAIT1
2925 * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible)
2926 * TCP_CLOSE_WAIT -> TCP_LAST_ACK
2928 * are legal only when FIN has been sent (i.e. in window),
2929 * rather than queued out of window. Purists blame.
2931 * F.e. "RFC state" is ESTABLISHED,
2932 * if Linux state is FIN-WAIT-1, but FIN is still not sent.
2934 * The visible declinations are that sometimes
2935 * we enter time-wait state, when it is not required really
2936 * (harmless), do not send active resets, when they are
2937 * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
2938 * they look as CLOSING or LAST_ACK for Linux)
2939 * Probably, I missed some more holelets.
2941 * XXX (TFO) - To start off we don't support SYN+ACK+FIN
2942 * in a single packet! (May consider it later but will
2943 * probably need API support or TCP_CORK SYN-ACK until
2944 * data is written and socket is closed.)
2949 sk_stream_wait_close(sk, timeout);
2952 state = sk->sk_state;
2958 /* remove backlog if any, without releasing ownership. */
2961 this_cpu_inc(tcp_orphan_count);
2963 /* Have we already been destroyed by a softirq or backlog? */
2964 if (state != TCP_CLOSE && sk->sk_state == TCP_CLOSE)
2967 /* This is a (useful) BSD violating of the RFC. There is a
2968 * problem with TCP as specified in that the other end could
2969 * keep a socket open forever with no application left this end.
2970 * We use a 1 minute timeout (about the same as BSD) then kill
2971 * our end. If they send after that then tough - BUT: long enough
2972 * that we won't make the old 4*rto = almost no time - whoops
2975 * Nope, it was not mistake. It is really desired behaviour
2976 * f.e. on http servers, when such sockets are useless, but
2977 * consume significant resources. Let's do it with special
2978 * linger2 option. --ANK
2981 if (sk->sk_state == TCP_FIN_WAIT2) {
2982 struct tcp_sock *tp = tcp_sk(sk);
2983 if (tp->linger2 < 0) {
2984 tcp_set_state(sk, TCP_CLOSE);
2985 tcp_send_active_reset(sk, GFP_ATOMIC);
2986 __NET_INC_STATS(sock_net(sk),
2987 LINUX_MIB_TCPABORTONLINGER);
2989 const int tmo = tcp_fin_time(sk);
2991 if (tmo > TCP_TIMEWAIT_LEN) {
2992 inet_csk_reset_keepalive_timer(sk,
2993 tmo - TCP_TIMEWAIT_LEN);
2995 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
3000 if (sk->sk_state != TCP_CLOSE) {
3001 if (tcp_check_oom(sk, 0)) {
3002 tcp_set_state(sk, TCP_CLOSE);
3003 tcp_send_active_reset(sk, GFP_ATOMIC);
3004 __NET_INC_STATS(sock_net(sk),
3005 LINUX_MIB_TCPABORTONMEMORY);
3006 } else if (!check_net(sock_net(sk))) {
3007 /* Not possible to send reset; just close */
3008 tcp_set_state(sk, TCP_CLOSE);
3012 if (sk->sk_state == TCP_CLOSE) {
3013 struct request_sock *req;
3015 req = rcu_dereference_protected(tcp_sk(sk)->fastopen_rsk,
3016 lockdep_sock_is_held(sk));
3017 /* We could get here with a non-NULL req if the socket is
3018 * aborted (e.g., closed with unread data) before 3WHS
3022 reqsk_fastopen_remove(sk, req, false);
3023 inet_csk_destroy_sock(sk);
3025 /* Otherwise, socket is reprieved until protocol close. */
3032 void tcp_close(struct sock *sk, long timeout)
3035 __tcp_close(sk, timeout);
3039 EXPORT_SYMBOL(tcp_close);
3041 /* These states need RST on ABORT according to RFC793 */
3043 static inline bool tcp_need_reset(int state)
3045 return (1 << state) &
3046 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 |
3047 TCPF_FIN_WAIT2 | TCPF_SYN_RECV);
3050 static void tcp_rtx_queue_purge(struct sock *sk)
3052 struct rb_node *p = rb_first(&sk->tcp_rtx_queue);
3054 tcp_sk(sk)->highest_sack = NULL;
3056 struct sk_buff *skb = rb_to_skb(p);
3059 /* Since we are deleting whole queue, no need to
3060 * list_del(&skb->tcp_tsorted_anchor)
3062 tcp_rtx_queue_unlink(skb, sk);
3063 tcp_wmem_free_skb(sk, skb);
3067 void tcp_write_queue_purge(struct sock *sk)
3069 struct sk_buff *skb;
3071 tcp_chrono_stop(sk, TCP_CHRONO_BUSY);
3072 while ((skb = __skb_dequeue(&sk->sk_write_queue)) != NULL) {
3073 tcp_skb_tsorted_anchor_cleanup(skb);
3074 tcp_wmem_free_skb(sk, skb);
3076 tcp_rtx_queue_purge(sk);
3077 INIT_LIST_HEAD(&tcp_sk(sk)->tsorted_sent_queue);
3078 tcp_clear_all_retrans_hints(tcp_sk(sk));
3079 tcp_sk(sk)->packets_out = 0;
3080 inet_csk(sk)->icsk_backoff = 0;
3083 int tcp_disconnect(struct sock *sk, int flags)
3085 struct inet_sock *inet = inet_sk(sk);
3086 struct inet_connection_sock *icsk = inet_csk(sk);
3087 struct tcp_sock *tp = tcp_sk(sk);
3088 int old_state = sk->sk_state;
3091 if (old_state != TCP_CLOSE)
3092 tcp_set_state(sk, TCP_CLOSE);
3094 /* ABORT function of RFC793 */
3095 if (old_state == TCP_LISTEN) {
3096 inet_csk_listen_stop(sk);
3097 } else if (unlikely(tp->repair)) {
3098 WRITE_ONCE(sk->sk_err, ECONNABORTED);
3099 } else if (tcp_need_reset(old_state) ||
3100 (tp->snd_nxt != tp->write_seq &&
3101 (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK))) {
3102 /* The last check adjusts for discrepancy of Linux wrt. RFC
3105 tcp_send_active_reset(sk, gfp_any());
3106 WRITE_ONCE(sk->sk_err, ECONNRESET);
3107 } else if (old_state == TCP_SYN_SENT)
3108 WRITE_ONCE(sk->sk_err, ECONNRESET);
3110 tcp_clear_xmit_timers(sk);
3111 __skb_queue_purge(&sk->sk_receive_queue);
3112 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt);
3113 WRITE_ONCE(tp->urg_data, 0);
3114 tcp_write_queue_purge(sk);
3115 tcp_fastopen_active_disable_ofo_check(sk);
3116 skb_rbtree_purge(&tp->out_of_order_queue);
3118 inet->inet_dport = 0;
3120 inet_bhash2_reset_saddr(sk);
3122 sk->sk_shutdown = 0;
3123 sock_reset_flag(sk, SOCK_DONE);
3125 tp->mdev_us = jiffies_to_usecs(TCP_TIMEOUT_INIT);
3126 tp->rcv_rtt_last_tsecr = 0;
3128 seq = tp->write_seq + tp->max_window + 2;
3131 WRITE_ONCE(tp->write_seq, seq);
3133 icsk->icsk_backoff = 0;
3134 icsk->icsk_probes_out = 0;
3135 icsk->icsk_probes_tstamp = 0;
3136 icsk->icsk_rto = TCP_TIMEOUT_INIT;
3137 icsk->icsk_rto_min = TCP_RTO_MIN;
3138 icsk->icsk_delack_max = TCP_DELACK_MAX;
3139 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
3140 tcp_snd_cwnd_set(tp, TCP_INIT_CWND);
3141 tp->snd_cwnd_cnt = 0;
3142 tp->is_cwnd_limited = 0;
3143 tp->max_packets_out = 0;
3144 tp->window_clamp = 0;
3146 tp->delivered_ce = 0;
3147 if (icsk->icsk_ca_ops->release)
3148 icsk->icsk_ca_ops->release(sk);
3149 memset(icsk->icsk_ca_priv, 0, sizeof(icsk->icsk_ca_priv));
3150 icsk->icsk_ca_initialized = 0;
3151 tcp_set_ca_state(sk, TCP_CA_Open);
3152 tp->is_sack_reneg = 0;
3153 tcp_clear_retrans(tp);
3154 tp->total_retrans = 0;
3155 inet_csk_delack_init(sk);
3156 /* Initialize rcv_mss to TCP_MIN_MSS to avoid division by 0
3157 * issue in __tcp_select_window()
3159 icsk->icsk_ack.rcv_mss = TCP_MIN_MSS;
3160 memset(&tp->rx_opt, 0, sizeof(tp->rx_opt));
3162 dst_release(xchg((__force struct dst_entry **)&sk->sk_rx_dst, NULL));
3163 tcp_saved_syn_free(tp);
3164 tp->compressed_ack = 0;
3168 tp->bytes_acked = 0;
3169 tp->bytes_received = 0;
3170 tp->bytes_retrans = 0;
3171 tp->data_segs_in = 0;
3172 tp->data_segs_out = 0;
3173 tp->duplicate_sack[0].start_seq = 0;
3174 tp->duplicate_sack[0].end_seq = 0;
3177 tp->retrans_out = 0;
3179 tp->tlp_high_seq = 0;
3180 tp->last_oow_ack_time = 0;
3182 /* There's a bubble in the pipe until at least the first ACK. */
3183 tp->app_limited = ~0U;
3184 tp->rate_app_limited = 1;
3185 tp->rack.mstamp = 0;
3186 tp->rack.advanced = 0;
3187 tp->rack.reo_wnd_steps = 1;
3188 tp->rack.last_delivered = 0;
3189 tp->rack.reo_wnd_persist = 0;
3190 tp->rack.dsack_seen = 0;
3191 tp->syn_data_acked = 0;
3192 tp->rx_opt.saw_tstamp = 0;
3193 tp->rx_opt.dsack = 0;
3194 tp->rx_opt.num_sacks = 0;
3195 tp->rcv_ooopack = 0;
3198 /* Clean up fastopen related fields */
3199 tcp_free_fastopen_req(tp);
3200 inet->defer_connect = 0;
3201 tp->fastopen_client_fail = 0;
3203 WARN_ON(inet->inet_num && !icsk->icsk_bind_hash);
3205 if (sk->sk_frag.page) {
3206 put_page(sk->sk_frag.page);
3207 sk->sk_frag.page = NULL;
3208 sk->sk_frag.offset = 0;
3210 sk_error_report(sk);
3213 EXPORT_SYMBOL(tcp_disconnect);
3215 static inline bool tcp_can_repair_sock(const struct sock *sk)
3217 return sockopt_ns_capable(sock_net(sk)->user_ns, CAP_NET_ADMIN) &&
3218 (sk->sk_state != TCP_LISTEN);
3221 static int tcp_repair_set_window(struct tcp_sock *tp, sockptr_t optbuf, int len)
3223 struct tcp_repair_window opt;
3228 if (len != sizeof(opt))
3231 if (copy_from_sockptr(&opt, optbuf, sizeof(opt)))
3234 if (opt.max_window < opt.snd_wnd)
3237 if (after(opt.snd_wl1, tp->rcv_nxt + opt.rcv_wnd))
3240 if (after(opt.rcv_wup, tp->rcv_nxt))
3243 tp->snd_wl1 = opt.snd_wl1;
3244 tp->snd_wnd = opt.snd_wnd;
3245 tp->max_window = opt.max_window;
3247 tp->rcv_wnd = opt.rcv_wnd;
3248 tp->rcv_wup = opt.rcv_wup;
3253 static int tcp_repair_options_est(struct sock *sk, sockptr_t optbuf,
3256 struct tcp_sock *tp = tcp_sk(sk);
3257 struct tcp_repair_opt opt;
3260 while (len >= sizeof(opt)) {
3261 if (copy_from_sockptr_offset(&opt, optbuf, offset, sizeof(opt)))
3264 offset += sizeof(opt);
3267 switch (opt.opt_code) {
3269 tp->rx_opt.mss_clamp = opt.opt_val;
3274 u16 snd_wscale = opt.opt_val & 0xFFFF;
3275 u16 rcv_wscale = opt.opt_val >> 16;
3277 if (snd_wscale > TCP_MAX_WSCALE || rcv_wscale > TCP_MAX_WSCALE)
3280 tp->rx_opt.snd_wscale = snd_wscale;
3281 tp->rx_opt.rcv_wscale = rcv_wscale;
3282 tp->rx_opt.wscale_ok = 1;
3285 case TCPOPT_SACK_PERM:
3286 if (opt.opt_val != 0)
3289 tp->rx_opt.sack_ok |= TCP_SACK_SEEN;
3291 case TCPOPT_TIMESTAMP:
3292 if (opt.opt_val != 0)
3295 tp->rx_opt.tstamp_ok = 1;
3303 DEFINE_STATIC_KEY_FALSE(tcp_tx_delay_enabled);
3304 EXPORT_SYMBOL(tcp_tx_delay_enabled);
3306 static void tcp_enable_tx_delay(void)
3308 if (!static_branch_unlikely(&tcp_tx_delay_enabled)) {
3309 static int __tcp_tx_delay_enabled = 0;
3311 if (cmpxchg(&__tcp_tx_delay_enabled, 0, 1) == 0) {
3312 static_branch_enable(&tcp_tx_delay_enabled);
3313 pr_info("TCP_TX_DELAY enabled\n");
3318 /* When set indicates to always queue non-full frames. Later the user clears
3319 * this option and we transmit any pending partial frames in the queue. This is
3320 * meant to be used alongside sendfile() to get properly filled frames when the
3321 * user (for example) must write out headers with a write() call first and then
3322 * use sendfile to send out the data parts.
3324 * TCP_CORK can be set together with TCP_NODELAY and it is stronger than
3327 void __tcp_sock_set_cork(struct sock *sk, bool on)
3329 struct tcp_sock *tp = tcp_sk(sk);
3332 tp->nonagle |= TCP_NAGLE_CORK;
3334 tp->nonagle &= ~TCP_NAGLE_CORK;
3335 if (tp->nonagle & TCP_NAGLE_OFF)
3336 tp->nonagle |= TCP_NAGLE_PUSH;
3337 tcp_push_pending_frames(sk);
3341 void tcp_sock_set_cork(struct sock *sk, bool on)
3344 __tcp_sock_set_cork(sk, on);
3347 EXPORT_SYMBOL(tcp_sock_set_cork);
3349 /* TCP_NODELAY is weaker than TCP_CORK, so that this option on corked socket is
3350 * remembered, but it is not activated until cork is cleared.
3352 * However, when TCP_NODELAY is set we make an explicit push, which overrides
3353 * even TCP_CORK for currently queued segments.
3355 void __tcp_sock_set_nodelay(struct sock *sk, bool on)
3358 tcp_sk(sk)->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH;
3359 tcp_push_pending_frames(sk);
3361 tcp_sk(sk)->nonagle &= ~TCP_NAGLE_OFF;
3365 void tcp_sock_set_nodelay(struct sock *sk)
3368 __tcp_sock_set_nodelay(sk, true);
3371 EXPORT_SYMBOL(tcp_sock_set_nodelay);
3373 static void __tcp_sock_set_quickack(struct sock *sk, int val)
3376 inet_csk_enter_pingpong_mode(sk);
3380 inet_csk_exit_pingpong_mode(sk);
3381 if ((1 << sk->sk_state) & (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) &&
3382 inet_csk_ack_scheduled(sk)) {
3383 inet_csk(sk)->icsk_ack.pending |= ICSK_ACK_PUSHED;
3384 tcp_cleanup_rbuf(sk, 1);
3386 inet_csk_enter_pingpong_mode(sk);
3390 void tcp_sock_set_quickack(struct sock *sk, int val)
3393 __tcp_sock_set_quickack(sk, val);
3396 EXPORT_SYMBOL(tcp_sock_set_quickack);
3398 int tcp_sock_set_syncnt(struct sock *sk, int val)
3400 if (val < 1 || val > MAX_TCP_SYNCNT)
3404 inet_csk(sk)->icsk_syn_retries = val;
3408 EXPORT_SYMBOL(tcp_sock_set_syncnt);
3410 void tcp_sock_set_user_timeout(struct sock *sk, u32 val)
3413 inet_csk(sk)->icsk_user_timeout = val;
3416 EXPORT_SYMBOL(tcp_sock_set_user_timeout);
3418 int tcp_sock_set_keepidle_locked(struct sock *sk, int val)
3420 struct tcp_sock *tp = tcp_sk(sk);
3422 if (val < 1 || val > MAX_TCP_KEEPIDLE)
3425 tp->keepalive_time = val * HZ;
3426 if (sock_flag(sk, SOCK_KEEPOPEN) &&
3427 !((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN))) {
3428 u32 elapsed = keepalive_time_elapsed(tp);
3430 if (tp->keepalive_time > elapsed)
3431 elapsed = tp->keepalive_time - elapsed;
3434 inet_csk_reset_keepalive_timer(sk, elapsed);
3440 int tcp_sock_set_keepidle(struct sock *sk, int val)
3445 err = tcp_sock_set_keepidle_locked(sk, val);
3449 EXPORT_SYMBOL(tcp_sock_set_keepidle);
3451 int tcp_sock_set_keepintvl(struct sock *sk, int val)
3453 if (val < 1 || val > MAX_TCP_KEEPINTVL)
3457 tcp_sk(sk)->keepalive_intvl = val * HZ;
3461 EXPORT_SYMBOL(tcp_sock_set_keepintvl);
3463 int tcp_sock_set_keepcnt(struct sock *sk, int val)
3465 if (val < 1 || val > MAX_TCP_KEEPCNT)
3469 tcp_sk(sk)->keepalive_probes = val;
3473 EXPORT_SYMBOL(tcp_sock_set_keepcnt);
3475 int tcp_set_window_clamp(struct sock *sk, int val)
3477 struct tcp_sock *tp = tcp_sk(sk);
3480 if (sk->sk_state != TCP_CLOSE)
3482 tp->window_clamp = 0;
3484 tp->window_clamp = val < SOCK_MIN_RCVBUF / 2 ?
3485 SOCK_MIN_RCVBUF / 2 : val;
3486 tp->rcv_ssthresh = min(tp->rcv_wnd, tp->window_clamp);
3492 * Socket option code for TCP.
3494 int do_tcp_setsockopt(struct sock *sk, int level, int optname,
3495 sockptr_t optval, unsigned int optlen)
3497 struct tcp_sock *tp = tcp_sk(sk);
3498 struct inet_connection_sock *icsk = inet_csk(sk);
3499 struct net *net = sock_net(sk);
3503 /* These are data/string values, all the others are ints */
3505 case TCP_CONGESTION: {
3506 char name[TCP_CA_NAME_MAX];
3511 val = strncpy_from_sockptr(name, optval,
3512 min_t(long, TCP_CA_NAME_MAX-1, optlen));
3517 sockopt_lock_sock(sk);
3518 err = tcp_set_congestion_control(sk, name, !has_current_bpf_ctx(),
3519 sockopt_ns_capable(sock_net(sk)->user_ns,
3521 sockopt_release_sock(sk);
3525 char name[TCP_ULP_NAME_MAX];
3530 val = strncpy_from_sockptr(name, optval,
3531 min_t(long, TCP_ULP_NAME_MAX - 1,
3537 sockopt_lock_sock(sk);
3538 err = tcp_set_ulp(sk, name);
3539 sockopt_release_sock(sk);
3542 case TCP_FASTOPEN_KEY: {
3543 __u8 key[TCP_FASTOPEN_KEY_BUF_LENGTH];
3544 __u8 *backup_key = NULL;
3546 /* Allow a backup key as well to facilitate key rotation
3547 * First key is the active one.
3549 if (optlen != TCP_FASTOPEN_KEY_LENGTH &&
3550 optlen != TCP_FASTOPEN_KEY_BUF_LENGTH)
3553 if (copy_from_sockptr(key, optval, optlen))
3556 if (optlen == TCP_FASTOPEN_KEY_BUF_LENGTH)
3557 backup_key = key + TCP_FASTOPEN_KEY_LENGTH;
3559 return tcp_fastopen_reset_cipher(net, sk, key, backup_key);
3566 if (optlen < sizeof(int))
3569 if (copy_from_sockptr(&val, optval, sizeof(val)))
3572 sockopt_lock_sock(sk);
3576 /* Values greater than interface MTU won't take effect. However
3577 * at the point when this call is done we typically don't yet
3578 * know which interface is going to be used
3580 if (val && (val < TCP_MIN_MSS || val > MAX_TCP_WINDOW)) {
3584 tp->rx_opt.user_mss = val;
3588 __tcp_sock_set_nodelay(sk, val);
3591 case TCP_THIN_LINEAR_TIMEOUTS:
3592 if (val < 0 || val > 1)
3598 case TCP_THIN_DUPACK:
3599 if (val < 0 || val > 1)
3604 if (!tcp_can_repair_sock(sk))
3606 else if (val == TCP_REPAIR_ON) {
3608 sk->sk_reuse = SK_FORCE_REUSE;
3609 tp->repair_queue = TCP_NO_QUEUE;
3610 } else if (val == TCP_REPAIR_OFF) {
3612 sk->sk_reuse = SK_NO_REUSE;
3613 tcp_send_window_probe(sk);
3614 } else if (val == TCP_REPAIR_OFF_NO_WP) {
3616 sk->sk_reuse = SK_NO_REUSE;
3622 case TCP_REPAIR_QUEUE:
3625 else if ((unsigned int)val < TCP_QUEUES_NR)
3626 tp->repair_queue = val;
3632 if (sk->sk_state != TCP_CLOSE) {
3634 } else if (tp->repair_queue == TCP_SEND_QUEUE) {
3635 if (!tcp_rtx_queue_empty(sk))
3638 WRITE_ONCE(tp->write_seq, val);
3639 } else if (tp->repair_queue == TCP_RECV_QUEUE) {
3640 if (tp->rcv_nxt != tp->copied_seq) {
3643 WRITE_ONCE(tp->rcv_nxt, val);
3644 WRITE_ONCE(tp->copied_seq, val);
3651 case TCP_REPAIR_OPTIONS:
3654 else if (sk->sk_state == TCP_ESTABLISHED && !tp->bytes_sent)
3655 err = tcp_repair_options_est(sk, optval, optlen);
3661 __tcp_sock_set_cork(sk, val);
3665 err = tcp_sock_set_keepidle_locked(sk, val);
3668 if (val < 1 || val > MAX_TCP_KEEPINTVL)
3671 tp->keepalive_intvl = val * HZ;
3674 if (val < 1 || val > MAX_TCP_KEEPCNT)
3677 tp->keepalive_probes = val;
3680 if (val < 1 || val > MAX_TCP_SYNCNT)
3683 icsk->icsk_syn_retries = val;
3687 /* 0: disable, 1: enable, 2: start from ether_header */
3688 if (val < 0 || val > 2)
3697 else if (val > TCP_FIN_TIMEOUT_MAX / HZ)
3698 tp->linger2 = TCP_FIN_TIMEOUT_MAX;
3700 tp->linger2 = val * HZ;
3703 case TCP_DEFER_ACCEPT:
3704 /* Translate value in seconds to number of retransmits */
3705 icsk->icsk_accept_queue.rskq_defer_accept =
3706 secs_to_retrans(val, TCP_TIMEOUT_INIT / HZ,
3710 case TCP_WINDOW_CLAMP:
3711 err = tcp_set_window_clamp(sk, val);
3715 __tcp_sock_set_quickack(sk, val);
3718 #ifdef CONFIG_TCP_MD5SIG
3720 case TCP_MD5SIG_EXT:
3721 err = tp->af_specific->md5_parse(sk, optname, optval, optlen);
3724 case TCP_USER_TIMEOUT:
3725 /* Cap the max time in ms TCP will retry or probe the window
3726 * before giving up and aborting (ETIMEDOUT) a connection.
3731 icsk->icsk_user_timeout = val;
3735 if (val >= 0 && ((1 << sk->sk_state) & (TCPF_CLOSE |
3737 tcp_fastopen_init_key_once(net);
3739 fastopen_queue_tune(sk, val);
3744 case TCP_FASTOPEN_CONNECT:
3745 if (val > 1 || val < 0) {
3747 } else if (READ_ONCE(net->ipv4.sysctl_tcp_fastopen) &
3748 TFO_CLIENT_ENABLE) {
3749 if (sk->sk_state == TCP_CLOSE)
3750 tp->fastopen_connect = val;
3757 case TCP_FASTOPEN_NO_COOKIE:
3758 if (val > 1 || val < 0)
3760 else if (!((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
3763 tp->fastopen_no_cookie = val;
3769 tp->tsoffset = val - tcp_time_stamp_raw();
3771 case TCP_REPAIR_WINDOW:
3772 err = tcp_repair_set_window(tp, optval, optlen);
3774 case TCP_NOTSENT_LOWAT:
3775 tp->notsent_lowat = val;
3776 sk->sk_write_space(sk);
3779 if (val > 1 || val < 0)
3782 tp->recvmsg_inq = val;
3786 tcp_enable_tx_delay();
3787 tp->tcp_tx_delay = val;
3794 sockopt_release_sock(sk);
3798 int tcp_setsockopt(struct sock *sk, int level, int optname, sockptr_t optval,
3799 unsigned int optlen)
3801 const struct inet_connection_sock *icsk = inet_csk(sk);
3803 if (level != SOL_TCP)
3804 /* Paired with WRITE_ONCE() in do_ipv6_setsockopt() and tcp_v6_connect() */
3805 return READ_ONCE(icsk->icsk_af_ops)->setsockopt(sk, level, optname,
3807 return do_tcp_setsockopt(sk, level, optname, optval, optlen);
3809 EXPORT_SYMBOL(tcp_setsockopt);
3811 static void tcp_get_info_chrono_stats(const struct tcp_sock *tp,
3812 struct tcp_info *info)
3814 u64 stats[__TCP_CHRONO_MAX], total = 0;
3817 for (i = TCP_CHRONO_BUSY; i < __TCP_CHRONO_MAX; ++i) {
3818 stats[i] = tp->chrono_stat[i - 1];
3819 if (i == tp->chrono_type)
3820 stats[i] += tcp_jiffies32 - tp->chrono_start;
3821 stats[i] *= USEC_PER_SEC / HZ;
3825 info->tcpi_busy_time = total;
3826 info->tcpi_rwnd_limited = stats[TCP_CHRONO_RWND_LIMITED];
3827 info->tcpi_sndbuf_limited = stats[TCP_CHRONO_SNDBUF_LIMITED];
3830 /* Return information about state of tcp endpoint in API format. */
3831 void tcp_get_info(struct sock *sk, struct tcp_info *info)
3833 const struct tcp_sock *tp = tcp_sk(sk); /* iff sk_type == SOCK_STREAM */
3834 const struct inet_connection_sock *icsk = inet_csk(sk);
3840 memset(info, 0, sizeof(*info));
3841 if (sk->sk_type != SOCK_STREAM)
3844 info->tcpi_state = inet_sk_state_load(sk);
3846 /* Report meaningful fields for all TCP states, including listeners */
3847 rate = READ_ONCE(sk->sk_pacing_rate);
3848 rate64 = (rate != ~0UL) ? rate : ~0ULL;
3849 info->tcpi_pacing_rate = rate64;
3851 rate = READ_ONCE(sk->sk_max_pacing_rate);
3852 rate64 = (rate != ~0UL) ? rate : ~0ULL;
3853 info->tcpi_max_pacing_rate = rate64;
3855 info->tcpi_reordering = tp->reordering;
3856 info->tcpi_snd_cwnd = tcp_snd_cwnd(tp);
3858 if (info->tcpi_state == TCP_LISTEN) {
3859 /* listeners aliased fields :
3860 * tcpi_unacked -> Number of children ready for accept()
3861 * tcpi_sacked -> max backlog
3863 info->tcpi_unacked = READ_ONCE(sk->sk_ack_backlog);
3864 info->tcpi_sacked = READ_ONCE(sk->sk_max_ack_backlog);
3868 slow = lock_sock_fast(sk);
3870 info->tcpi_ca_state = icsk->icsk_ca_state;
3871 info->tcpi_retransmits = icsk->icsk_retransmits;
3872 info->tcpi_probes = icsk->icsk_probes_out;
3873 info->tcpi_backoff = icsk->icsk_backoff;
3875 if (tp->rx_opt.tstamp_ok)
3876 info->tcpi_options |= TCPI_OPT_TIMESTAMPS;
3877 if (tcp_is_sack(tp))
3878 info->tcpi_options |= TCPI_OPT_SACK;
3879 if (tp->rx_opt.wscale_ok) {
3880 info->tcpi_options |= TCPI_OPT_WSCALE;
3881 info->tcpi_snd_wscale = tp->rx_opt.snd_wscale;
3882 info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale;
3885 if (tp->ecn_flags & TCP_ECN_OK)
3886 info->tcpi_options |= TCPI_OPT_ECN;
3887 if (tp->ecn_flags & TCP_ECN_SEEN)
3888 info->tcpi_options |= TCPI_OPT_ECN_SEEN;
3889 if (tp->syn_data_acked)
3890 info->tcpi_options |= TCPI_OPT_SYN_DATA;
3892 info->tcpi_rto = jiffies_to_usecs(icsk->icsk_rto);
3893 info->tcpi_ato = jiffies_to_usecs(icsk->icsk_ack.ato);
3894 info->tcpi_snd_mss = tp->mss_cache;
3895 info->tcpi_rcv_mss = icsk->icsk_ack.rcv_mss;
3897 info->tcpi_unacked = tp->packets_out;
3898 info->tcpi_sacked = tp->sacked_out;
3900 info->tcpi_lost = tp->lost_out;
3901 info->tcpi_retrans = tp->retrans_out;
3903 now = tcp_jiffies32;
3904 info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime);
3905 info->tcpi_last_data_recv = jiffies_to_msecs(now - icsk->icsk_ack.lrcvtime);
3906 info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp);
3908 info->tcpi_pmtu = icsk->icsk_pmtu_cookie;
3909 info->tcpi_rcv_ssthresh = tp->rcv_ssthresh;
3910 info->tcpi_rtt = tp->srtt_us >> 3;
3911 info->tcpi_rttvar = tp->mdev_us >> 2;
3912 info->tcpi_snd_ssthresh = tp->snd_ssthresh;
3913 info->tcpi_advmss = tp->advmss;
3915 info->tcpi_rcv_rtt = tp->rcv_rtt_est.rtt_us >> 3;
3916 info->tcpi_rcv_space = tp->rcvq_space.space;
3918 info->tcpi_total_retrans = tp->total_retrans;
3920 info->tcpi_bytes_acked = tp->bytes_acked;
3921 info->tcpi_bytes_received = tp->bytes_received;
3922 info->tcpi_notsent_bytes = max_t(int, 0, tp->write_seq - tp->snd_nxt);
3923 tcp_get_info_chrono_stats(tp, info);
3925 info->tcpi_segs_out = tp->segs_out;
3927 /* segs_in and data_segs_in can be updated from tcp_segs_in() from BH */
3928 info->tcpi_segs_in = READ_ONCE(tp->segs_in);
3929 info->tcpi_data_segs_in = READ_ONCE(tp->data_segs_in);
3931 info->tcpi_min_rtt = tcp_min_rtt(tp);
3932 info->tcpi_data_segs_out = tp->data_segs_out;
3934 info->tcpi_delivery_rate_app_limited = tp->rate_app_limited ? 1 : 0;
3935 rate64 = tcp_compute_delivery_rate(tp);
3937 info->tcpi_delivery_rate = rate64;
3938 info->tcpi_delivered = tp->delivered;
3939 info->tcpi_delivered_ce = tp->delivered_ce;
3940 info->tcpi_bytes_sent = tp->bytes_sent;
3941 info->tcpi_bytes_retrans = tp->bytes_retrans;
3942 info->tcpi_dsack_dups = tp->dsack_dups;
3943 info->tcpi_reord_seen = tp->reord_seen;
3944 info->tcpi_rcv_ooopack = tp->rcv_ooopack;
3945 info->tcpi_snd_wnd = tp->snd_wnd;
3946 info->tcpi_rcv_wnd = tp->rcv_wnd;
3947 info->tcpi_rehash = tp->plb_rehash + tp->timeout_rehash;
3948 info->tcpi_fastopen_client_fail = tp->fastopen_client_fail;
3949 unlock_sock_fast(sk, slow);
3951 EXPORT_SYMBOL_GPL(tcp_get_info);
3953 static size_t tcp_opt_stats_get_size(void)
3956 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_BUSY */
3957 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_RWND_LIMITED */
3958 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_SNDBUF_LIMITED */
3959 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_DATA_SEGS_OUT */
3960 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_TOTAL_RETRANS */
3961 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_PACING_RATE */
3962 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_DELIVERY_RATE */
3963 nla_total_size(sizeof(u32)) + /* TCP_NLA_SND_CWND */
3964 nla_total_size(sizeof(u32)) + /* TCP_NLA_REORDERING */
3965 nla_total_size(sizeof(u32)) + /* TCP_NLA_MIN_RTT */
3966 nla_total_size(sizeof(u8)) + /* TCP_NLA_RECUR_RETRANS */
3967 nla_total_size(sizeof(u8)) + /* TCP_NLA_DELIVERY_RATE_APP_LMT */
3968 nla_total_size(sizeof(u32)) + /* TCP_NLA_SNDQ_SIZE */
3969 nla_total_size(sizeof(u8)) + /* TCP_NLA_CA_STATE */
3970 nla_total_size(sizeof(u32)) + /* TCP_NLA_SND_SSTHRESH */
3971 nla_total_size(sizeof(u32)) + /* TCP_NLA_DELIVERED */
3972 nla_total_size(sizeof(u32)) + /* TCP_NLA_DELIVERED_CE */
3973 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_BYTES_SENT */
3974 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_BYTES_RETRANS */
3975 nla_total_size(sizeof(u32)) + /* TCP_NLA_DSACK_DUPS */
3976 nla_total_size(sizeof(u32)) + /* TCP_NLA_REORD_SEEN */
3977 nla_total_size(sizeof(u32)) + /* TCP_NLA_SRTT */
3978 nla_total_size(sizeof(u16)) + /* TCP_NLA_TIMEOUT_REHASH */
3979 nla_total_size(sizeof(u32)) + /* TCP_NLA_BYTES_NOTSENT */
3980 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_EDT */
3981 nla_total_size(sizeof(u8)) + /* TCP_NLA_TTL */
3982 nla_total_size(sizeof(u32)) + /* TCP_NLA_REHASH */
3986 /* Returns TTL or hop limit of an incoming packet from skb. */
3987 static u8 tcp_skb_ttl_or_hop_limit(const struct sk_buff *skb)
3989 if (skb->protocol == htons(ETH_P_IP))
3990 return ip_hdr(skb)->ttl;
3991 else if (skb->protocol == htons(ETH_P_IPV6))
3992 return ipv6_hdr(skb)->hop_limit;
3997 struct sk_buff *tcp_get_timestamping_opt_stats(const struct sock *sk,
3998 const struct sk_buff *orig_skb,
3999 const struct sk_buff *ack_skb)
4001 const struct tcp_sock *tp = tcp_sk(sk);
4002 struct sk_buff *stats;
4003 struct tcp_info info;
4007 stats = alloc_skb(tcp_opt_stats_get_size(), GFP_ATOMIC);
4011 tcp_get_info_chrono_stats(tp, &info);
4012 nla_put_u64_64bit(stats, TCP_NLA_BUSY,
4013 info.tcpi_busy_time, TCP_NLA_PAD);
4014 nla_put_u64_64bit(stats, TCP_NLA_RWND_LIMITED,
4015 info.tcpi_rwnd_limited, TCP_NLA_PAD);
4016 nla_put_u64_64bit(stats, TCP_NLA_SNDBUF_LIMITED,
4017 info.tcpi_sndbuf_limited, TCP_NLA_PAD);
4018 nla_put_u64_64bit(stats, TCP_NLA_DATA_SEGS_OUT,
4019 tp->data_segs_out, TCP_NLA_PAD);
4020 nla_put_u64_64bit(stats, TCP_NLA_TOTAL_RETRANS,
4021 tp->total_retrans, TCP_NLA_PAD);
4023 rate = READ_ONCE(sk->sk_pacing_rate);
4024 rate64 = (rate != ~0UL) ? rate : ~0ULL;
4025 nla_put_u64_64bit(stats, TCP_NLA_PACING_RATE, rate64, TCP_NLA_PAD);
4027 rate64 = tcp_compute_delivery_rate(tp);
4028 nla_put_u64_64bit(stats, TCP_NLA_DELIVERY_RATE, rate64, TCP_NLA_PAD);
4030 nla_put_u32(stats, TCP_NLA_SND_CWND, tcp_snd_cwnd(tp));
4031 nla_put_u32(stats, TCP_NLA_REORDERING, tp->reordering);
4032 nla_put_u32(stats, TCP_NLA_MIN_RTT, tcp_min_rtt(tp));
4034 nla_put_u8(stats, TCP_NLA_RECUR_RETRANS, inet_csk(sk)->icsk_retransmits);
4035 nla_put_u8(stats, TCP_NLA_DELIVERY_RATE_APP_LMT, !!tp->rate_app_limited);
4036 nla_put_u32(stats, TCP_NLA_SND_SSTHRESH, tp->snd_ssthresh);
4037 nla_put_u32(stats, TCP_NLA_DELIVERED, tp->delivered);
4038 nla_put_u32(stats, TCP_NLA_DELIVERED_CE, tp->delivered_ce);
4040 nla_put_u32(stats, TCP_NLA_SNDQ_SIZE, tp->write_seq - tp->snd_una);
4041 nla_put_u8(stats, TCP_NLA_CA_STATE, inet_csk(sk)->icsk_ca_state);
4043 nla_put_u64_64bit(stats, TCP_NLA_BYTES_SENT, tp->bytes_sent,
4045 nla_put_u64_64bit(stats, TCP_NLA_BYTES_RETRANS, tp->bytes_retrans,
4047 nla_put_u32(stats, TCP_NLA_DSACK_DUPS, tp->dsack_dups);
4048 nla_put_u32(stats, TCP_NLA_REORD_SEEN, tp->reord_seen);
4049 nla_put_u32(stats, TCP_NLA_SRTT, tp->srtt_us >> 3);
4050 nla_put_u16(stats, TCP_NLA_TIMEOUT_REHASH, tp->timeout_rehash);
4051 nla_put_u32(stats, TCP_NLA_BYTES_NOTSENT,
4052 max_t(int, 0, tp->write_seq - tp->snd_nxt));
4053 nla_put_u64_64bit(stats, TCP_NLA_EDT, orig_skb->skb_mstamp_ns,
4056 nla_put_u8(stats, TCP_NLA_TTL,
4057 tcp_skb_ttl_or_hop_limit(ack_skb));
4059 nla_put_u32(stats, TCP_NLA_REHASH, tp->plb_rehash + tp->timeout_rehash);
4063 int do_tcp_getsockopt(struct sock *sk, int level,
4064 int optname, sockptr_t optval, sockptr_t optlen)
4066 struct inet_connection_sock *icsk = inet_csk(sk);
4067 struct tcp_sock *tp = tcp_sk(sk);
4068 struct net *net = sock_net(sk);
4071 if (copy_from_sockptr(&len, optlen, sizeof(int)))
4074 len = min_t(unsigned int, len, sizeof(int));
4081 val = tp->mss_cache;
4082 if (!val && ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
4083 val = tp->rx_opt.user_mss;
4085 val = tp->rx_opt.mss_clamp;
4088 val = !!(tp->nonagle&TCP_NAGLE_OFF);
4091 val = !!(tp->nonagle&TCP_NAGLE_CORK);
4094 val = keepalive_time_when(tp) / HZ;
4097 val = keepalive_intvl_when(tp) / HZ;
4100 val = keepalive_probes(tp);
4103 val = icsk->icsk_syn_retries ? :
4104 READ_ONCE(net->ipv4.sysctl_tcp_syn_retries);
4109 val = (val ? : READ_ONCE(net->ipv4.sysctl_tcp_fin_timeout)) / HZ;
4111 case TCP_DEFER_ACCEPT:
4112 val = retrans_to_secs(icsk->icsk_accept_queue.rskq_defer_accept,
4113 TCP_TIMEOUT_INIT / HZ, TCP_RTO_MAX / HZ);
4115 case TCP_WINDOW_CLAMP:
4116 val = tp->window_clamp;
4119 struct tcp_info info;
4121 if (copy_from_sockptr(&len, optlen, sizeof(int)))
4124 tcp_get_info(sk, &info);
4126 len = min_t(unsigned int, len, sizeof(info));
4127 if (copy_to_sockptr(optlen, &len, sizeof(int)))
4129 if (copy_to_sockptr(optval, &info, len))
4134 const struct tcp_congestion_ops *ca_ops;
4135 union tcp_cc_info info;
4139 if (copy_from_sockptr(&len, optlen, sizeof(int)))
4142 ca_ops = icsk->icsk_ca_ops;
4143 if (ca_ops && ca_ops->get_info)
4144 sz = ca_ops->get_info(sk, ~0U, &attr, &info);
4146 len = min_t(unsigned int, len, sz);
4147 if (copy_to_sockptr(optlen, &len, sizeof(int)))
4149 if (copy_to_sockptr(optval, &info, len))
4154 val = !inet_csk_in_pingpong_mode(sk);
4157 case TCP_CONGESTION:
4158 if (copy_from_sockptr(&len, optlen, sizeof(int)))
4160 len = min_t(unsigned int, len, TCP_CA_NAME_MAX);
4161 if (copy_to_sockptr(optlen, &len, sizeof(int)))
4163 if (copy_to_sockptr(optval, icsk->icsk_ca_ops->name, len))
4168 if (copy_from_sockptr(&len, optlen, sizeof(int)))
4170 len = min_t(unsigned int, len, TCP_ULP_NAME_MAX);
4171 if (!icsk->icsk_ulp_ops) {
4173 if (copy_to_sockptr(optlen, &len, sizeof(int)))
4177 if (copy_to_sockptr(optlen, &len, sizeof(int)))
4179 if (copy_to_sockptr(optval, icsk->icsk_ulp_ops->name, len))
4183 case TCP_FASTOPEN_KEY: {
4184 u64 key[TCP_FASTOPEN_KEY_BUF_LENGTH / sizeof(u64)];
4185 unsigned int key_len;
4187 if (copy_from_sockptr(&len, optlen, sizeof(int)))
4190 key_len = tcp_fastopen_get_cipher(net, icsk, key) *
4191 TCP_FASTOPEN_KEY_LENGTH;
4192 len = min_t(unsigned int, len, key_len);
4193 if (copy_to_sockptr(optlen, &len, sizeof(int)))
4195 if (copy_to_sockptr(optval, key, len))
4199 case TCP_THIN_LINEAR_TIMEOUTS:
4203 case TCP_THIN_DUPACK:
4211 case TCP_REPAIR_QUEUE:
4213 val = tp->repair_queue;
4218 case TCP_REPAIR_WINDOW: {
4219 struct tcp_repair_window opt;
4221 if (copy_from_sockptr(&len, optlen, sizeof(int)))
4224 if (len != sizeof(opt))
4230 opt.snd_wl1 = tp->snd_wl1;
4231 opt.snd_wnd = tp->snd_wnd;
4232 opt.max_window = tp->max_window;
4233 opt.rcv_wnd = tp->rcv_wnd;
4234 opt.rcv_wup = tp->rcv_wup;
4236 if (copy_to_sockptr(optval, &opt, len))
4241 if (tp->repair_queue == TCP_SEND_QUEUE)
4242 val = tp->write_seq;
4243 else if (tp->repair_queue == TCP_RECV_QUEUE)
4249 case TCP_USER_TIMEOUT:
4250 val = icsk->icsk_user_timeout;
4254 val = icsk->icsk_accept_queue.fastopenq.max_qlen;
4257 case TCP_FASTOPEN_CONNECT:
4258 val = tp->fastopen_connect;
4261 case TCP_FASTOPEN_NO_COOKIE:
4262 val = tp->fastopen_no_cookie;
4266 val = tp->tcp_tx_delay;
4270 val = tcp_time_stamp_raw() + tp->tsoffset;
4272 case TCP_NOTSENT_LOWAT:
4273 val = tp->notsent_lowat;
4276 val = tp->recvmsg_inq;
4281 case TCP_SAVED_SYN: {
4282 if (copy_from_sockptr(&len, optlen, sizeof(int)))
4285 sockopt_lock_sock(sk);
4286 if (tp->saved_syn) {
4287 if (len < tcp_saved_syn_len(tp->saved_syn)) {
4288 len = tcp_saved_syn_len(tp->saved_syn);
4289 if (copy_to_sockptr(optlen, &len, sizeof(int))) {
4290 sockopt_release_sock(sk);
4293 sockopt_release_sock(sk);
4296 len = tcp_saved_syn_len(tp->saved_syn);
4297 if (copy_to_sockptr(optlen, &len, sizeof(int))) {
4298 sockopt_release_sock(sk);
4301 if (copy_to_sockptr(optval, tp->saved_syn->data, len)) {
4302 sockopt_release_sock(sk);
4305 tcp_saved_syn_free(tp);
4306 sockopt_release_sock(sk);
4308 sockopt_release_sock(sk);
4310 if (copy_to_sockptr(optlen, &len, sizeof(int)))
4316 case TCP_ZEROCOPY_RECEIVE: {
4317 struct scm_timestamping_internal tss;
4318 struct tcp_zerocopy_receive zc = {};
4321 if (copy_from_sockptr(&len, optlen, sizeof(int)))
4324 len < offsetofend(struct tcp_zerocopy_receive, length))
4326 if (unlikely(len > sizeof(zc))) {
4327 err = check_zeroed_sockptr(optval, sizeof(zc),
4330 return err == 0 ? -EINVAL : err;
4332 if (copy_to_sockptr(optlen, &len, sizeof(int)))
4335 if (copy_from_sockptr(&zc, optval, len))
4339 if (zc.msg_flags & ~(TCP_VALID_ZC_MSG_FLAGS))
4341 sockopt_lock_sock(sk);
4342 err = tcp_zerocopy_receive(sk, &zc, &tss);
4343 err = BPF_CGROUP_RUN_PROG_GETSOCKOPT_KERN(sk, level, optname,
4345 sockopt_release_sock(sk);
4346 if (len >= offsetofend(struct tcp_zerocopy_receive, msg_flags))
4347 goto zerocopy_rcv_cmsg;
4349 case offsetofend(struct tcp_zerocopy_receive, msg_flags):
4350 goto zerocopy_rcv_cmsg;
4351 case offsetofend(struct tcp_zerocopy_receive, msg_controllen):
4352 case offsetofend(struct tcp_zerocopy_receive, msg_control):
4353 case offsetofend(struct tcp_zerocopy_receive, flags):
4354 case offsetofend(struct tcp_zerocopy_receive, copybuf_len):
4355 case offsetofend(struct tcp_zerocopy_receive, copybuf_address):
4356 case offsetofend(struct tcp_zerocopy_receive, err):
4357 goto zerocopy_rcv_sk_err;
4358 case offsetofend(struct tcp_zerocopy_receive, inq):
4359 goto zerocopy_rcv_inq;
4360 case offsetofend(struct tcp_zerocopy_receive, length):
4362 goto zerocopy_rcv_out;
4365 if (zc.msg_flags & TCP_CMSG_TS)
4366 tcp_zc_finalize_rx_tstamp(sk, &zc, &tss);
4369 zerocopy_rcv_sk_err:
4371 zc.err = sock_error(sk);
4373 zc.inq = tcp_inq_hint(sk);
4375 if (!err && copy_to_sockptr(optval, &zc, len))
4381 return -ENOPROTOOPT;
4384 if (copy_to_sockptr(optlen, &len, sizeof(int)))
4386 if (copy_to_sockptr(optval, &val, len))
4391 bool tcp_bpf_bypass_getsockopt(int level, int optname)
4393 /* TCP do_tcp_getsockopt has optimized getsockopt implementation
4394 * to avoid extra socket lock for TCP_ZEROCOPY_RECEIVE.
4396 if (level == SOL_TCP && optname == TCP_ZEROCOPY_RECEIVE)
4401 EXPORT_SYMBOL(tcp_bpf_bypass_getsockopt);
4403 int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval,
4406 struct inet_connection_sock *icsk = inet_csk(sk);
4408 if (level != SOL_TCP)
4409 /* Paired with WRITE_ONCE() in do_ipv6_setsockopt() and tcp_v6_connect() */
4410 return READ_ONCE(icsk->icsk_af_ops)->getsockopt(sk, level, optname,
4412 return do_tcp_getsockopt(sk, level, optname, USER_SOCKPTR(optval),
4413 USER_SOCKPTR(optlen));
4415 EXPORT_SYMBOL(tcp_getsockopt);
4417 #ifdef CONFIG_TCP_MD5SIG
4418 static DEFINE_PER_CPU(struct tcp_md5sig_pool, tcp_md5sig_pool);
4419 static DEFINE_MUTEX(tcp_md5sig_mutex);
4420 static bool tcp_md5sig_pool_populated = false;
4422 static void __tcp_alloc_md5sig_pool(void)
4424 struct crypto_ahash *hash;
4427 hash = crypto_alloc_ahash("md5", 0, CRYPTO_ALG_ASYNC);
4431 for_each_possible_cpu(cpu) {
4432 void *scratch = per_cpu(tcp_md5sig_pool, cpu).scratch;
4433 struct ahash_request *req;
4436 scratch = kmalloc_node(sizeof(union tcp_md5sum_block) +
4437 sizeof(struct tcphdr),
4442 per_cpu(tcp_md5sig_pool, cpu).scratch = scratch;
4444 if (per_cpu(tcp_md5sig_pool, cpu).md5_req)
4447 req = ahash_request_alloc(hash, GFP_KERNEL);
4451 ahash_request_set_callback(req, 0, NULL, NULL);
4453 per_cpu(tcp_md5sig_pool, cpu).md5_req = req;
4455 /* before setting tcp_md5sig_pool_populated, we must commit all writes
4456 * to memory. See smp_rmb() in tcp_get_md5sig_pool()
4459 /* Paired with READ_ONCE() from tcp_alloc_md5sig_pool()
4460 * and tcp_get_md5sig_pool().
4462 WRITE_ONCE(tcp_md5sig_pool_populated, true);
4465 bool tcp_alloc_md5sig_pool(void)
4467 /* Paired with WRITE_ONCE() from __tcp_alloc_md5sig_pool() */
4468 if (unlikely(!READ_ONCE(tcp_md5sig_pool_populated))) {
4469 mutex_lock(&tcp_md5sig_mutex);
4471 if (!tcp_md5sig_pool_populated)
4472 __tcp_alloc_md5sig_pool();
4474 mutex_unlock(&tcp_md5sig_mutex);
4476 /* Paired with WRITE_ONCE() from __tcp_alloc_md5sig_pool() */
4477 return READ_ONCE(tcp_md5sig_pool_populated);
4479 EXPORT_SYMBOL(tcp_alloc_md5sig_pool);
4483 * tcp_get_md5sig_pool - get md5sig_pool for this user
4485 * We use percpu structure, so if we succeed, we exit with preemption
4486 * and BH disabled, to make sure another thread or softirq handling
4487 * wont try to get same context.
4489 struct tcp_md5sig_pool *tcp_get_md5sig_pool(void)
4493 /* Paired with WRITE_ONCE() from __tcp_alloc_md5sig_pool() */
4494 if (READ_ONCE(tcp_md5sig_pool_populated)) {
4495 /* coupled with smp_wmb() in __tcp_alloc_md5sig_pool() */
4497 return this_cpu_ptr(&tcp_md5sig_pool);
4502 EXPORT_SYMBOL(tcp_get_md5sig_pool);
4504 int tcp_md5_hash_skb_data(struct tcp_md5sig_pool *hp,
4505 const struct sk_buff *skb, unsigned int header_len)
4507 struct scatterlist sg;
4508 const struct tcphdr *tp = tcp_hdr(skb);
4509 struct ahash_request *req = hp->md5_req;
4511 const unsigned int head_data_len = skb_headlen(skb) > header_len ?
4512 skb_headlen(skb) - header_len : 0;
4513 const struct skb_shared_info *shi = skb_shinfo(skb);
4514 struct sk_buff *frag_iter;
4516 sg_init_table(&sg, 1);
4518 sg_set_buf(&sg, ((u8 *) tp) + header_len, head_data_len);
4519 ahash_request_set_crypt(req, &sg, NULL, head_data_len);
4520 if (crypto_ahash_update(req))
4523 for (i = 0; i < shi->nr_frags; ++i) {
4524 const skb_frag_t *f = &shi->frags[i];
4525 unsigned int offset = skb_frag_off(f);
4526 struct page *page = skb_frag_page(f) + (offset >> PAGE_SHIFT);
4528 sg_set_page(&sg, page, skb_frag_size(f),
4529 offset_in_page(offset));
4530 ahash_request_set_crypt(req, &sg, NULL, skb_frag_size(f));
4531 if (crypto_ahash_update(req))
4535 skb_walk_frags(skb, frag_iter)
4536 if (tcp_md5_hash_skb_data(hp, frag_iter, 0))
4541 EXPORT_SYMBOL(tcp_md5_hash_skb_data);
4543 int tcp_md5_hash_key(struct tcp_md5sig_pool *hp, const struct tcp_md5sig_key *key)
4545 u8 keylen = READ_ONCE(key->keylen); /* paired with WRITE_ONCE() in tcp_md5_do_add */
4546 struct scatterlist sg;
4548 sg_init_one(&sg, key->key, keylen);
4549 ahash_request_set_crypt(hp->md5_req, &sg, NULL, keylen);
4551 /* We use data_race() because tcp_md5_do_add() might change key->key under us */
4552 return data_race(crypto_ahash_update(hp->md5_req));
4554 EXPORT_SYMBOL(tcp_md5_hash_key);
4556 /* Called with rcu_read_lock() */
4557 enum skb_drop_reason
4558 tcp_inbound_md5_hash(const struct sock *sk, const struct sk_buff *skb,
4559 const void *saddr, const void *daddr,
4560 int family, int dif, int sdif)
4563 * This gets called for each TCP segment that arrives
4564 * so we want to be efficient.
4565 * We have 3 drop cases:
4566 * o No MD5 hash and one expected.
4567 * o MD5 hash and we're not expecting one.
4568 * o MD5 hash and its wrong.
4570 const __u8 *hash_location = NULL;
4571 struct tcp_md5sig_key *hash_expected;
4572 const struct tcphdr *th = tcp_hdr(skb);
4573 const struct tcp_sock *tp = tcp_sk(sk);
4574 int genhash, l3index;
4577 /* sdif set, means packet ingressed via a device
4578 * in an L3 domain and dif is set to the l3mdev
4580 l3index = sdif ? dif : 0;
4582 hash_expected = tcp_md5_do_lookup(sk, l3index, saddr, family);
4583 hash_location = tcp_parse_md5sig_option(th);
4585 /* We've parsed the options - do we have a hash? */
4586 if (!hash_expected && !hash_location)
4587 return SKB_NOT_DROPPED_YET;
4589 if (hash_expected && !hash_location) {
4590 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMD5NOTFOUND);
4591 return SKB_DROP_REASON_TCP_MD5NOTFOUND;
4594 if (!hash_expected && hash_location) {
4595 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMD5UNEXPECTED);
4596 return SKB_DROP_REASON_TCP_MD5UNEXPECTED;
4599 /* Check the signature.
4600 * To support dual stack listeners, we need to handle
4603 if (family == AF_INET)
4604 genhash = tcp_v4_md5_hash_skb(newhash,
4608 genhash = tp->af_specific->calc_md5_hash(newhash,
4612 if (genhash || memcmp(hash_location, newhash, 16) != 0) {
4613 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMD5FAILURE);
4614 if (family == AF_INET) {
4615 net_info_ratelimited("MD5 Hash failed for (%pI4, %d)->(%pI4, %d)%s L3 index %d\n",
4616 saddr, ntohs(th->source),
4617 daddr, ntohs(th->dest),
4618 genhash ? " tcp_v4_calc_md5_hash failed"
4621 net_info_ratelimited("MD5 Hash %s for [%pI6c]:%u->[%pI6c]:%u L3 index %d\n",
4622 genhash ? "failed" : "mismatch",
4623 saddr, ntohs(th->source),
4624 daddr, ntohs(th->dest), l3index);
4626 return SKB_DROP_REASON_TCP_MD5FAILURE;
4628 return SKB_NOT_DROPPED_YET;
4630 EXPORT_SYMBOL(tcp_inbound_md5_hash);
4634 void tcp_done(struct sock *sk)
4636 struct request_sock *req;
4638 /* We might be called with a new socket, after
4639 * inet_csk_prepare_forced_close() has been called
4640 * so we can not use lockdep_sock_is_held(sk)
4642 req = rcu_dereference_protected(tcp_sk(sk)->fastopen_rsk, 1);
4644 if (sk->sk_state == TCP_SYN_SENT || sk->sk_state == TCP_SYN_RECV)
4645 TCP_INC_STATS(sock_net(sk), TCP_MIB_ATTEMPTFAILS);
4647 tcp_set_state(sk, TCP_CLOSE);
4648 tcp_clear_xmit_timers(sk);
4650 reqsk_fastopen_remove(sk, req, false);
4652 sk->sk_shutdown = SHUTDOWN_MASK;
4654 if (!sock_flag(sk, SOCK_DEAD))
4655 sk->sk_state_change(sk);
4657 inet_csk_destroy_sock(sk);
4659 EXPORT_SYMBOL_GPL(tcp_done);
4661 int tcp_abort(struct sock *sk, int err)
4663 int state = inet_sk_state_load(sk);
4665 if (state == TCP_NEW_SYN_RECV) {
4666 struct request_sock *req = inet_reqsk(sk);
4669 inet_csk_reqsk_queue_drop(req->rsk_listener, req);
4673 if (state == TCP_TIME_WAIT) {
4674 struct inet_timewait_sock *tw = inet_twsk(sk);
4676 refcount_inc(&tw->tw_refcnt);
4678 inet_twsk_deschedule_put(tw);
4683 /* Don't race with userspace socket closes such as tcp_close. */
4686 if (sk->sk_state == TCP_LISTEN) {
4687 tcp_set_state(sk, TCP_CLOSE);
4688 inet_csk_listen_stop(sk);
4691 /* Don't race with BH socket closes such as inet_csk_listen_stop. */
4695 if (!sock_flag(sk, SOCK_DEAD)) {
4696 WRITE_ONCE(sk->sk_err, err);
4697 /* This barrier is coupled with smp_rmb() in tcp_poll() */
4699 sk_error_report(sk);
4700 if (tcp_need_reset(sk->sk_state))
4701 tcp_send_active_reset(sk, GFP_ATOMIC);
4707 tcp_write_queue_purge(sk);
4711 EXPORT_SYMBOL_GPL(tcp_abort);
4713 extern struct tcp_congestion_ops tcp_reno;
4715 static __initdata unsigned long thash_entries;
4716 static int __init set_thash_entries(char *str)
4723 ret = kstrtoul(str, 0, &thash_entries);
4729 __setup("thash_entries=", set_thash_entries);
4731 static void __init tcp_init_mem(void)
4733 unsigned long limit = nr_free_buffer_pages() / 16;
4735 limit = max(limit, 128UL);
4736 sysctl_tcp_mem[0] = limit / 4 * 3; /* 4.68 % */
4737 sysctl_tcp_mem[1] = limit; /* 6.25 % */
4738 sysctl_tcp_mem[2] = sysctl_tcp_mem[0] * 2; /* 9.37 % */
4741 void __init tcp_init(void)
4743 int max_rshare, max_wshare, cnt;
4744 unsigned long limit;
4747 BUILD_BUG_ON(TCP_MIN_SND_MSS <= MAX_TCP_OPTION_SPACE);
4748 BUILD_BUG_ON(sizeof(struct tcp_skb_cb) >
4749 sizeof_field(struct sk_buff, cb));
4751 percpu_counter_init(&tcp_sockets_allocated, 0, GFP_KERNEL);
4753 timer_setup(&tcp_orphan_timer, tcp_orphan_update, TIMER_DEFERRABLE);
4754 mod_timer(&tcp_orphan_timer, jiffies + TCP_ORPHAN_TIMER_PERIOD);
4756 inet_hashinfo2_init(&tcp_hashinfo, "tcp_listen_portaddr_hash",
4757 thash_entries, 21, /* one slot per 2 MB*/
4759 tcp_hashinfo.bind_bucket_cachep =
4760 kmem_cache_create("tcp_bind_bucket",
4761 sizeof(struct inet_bind_bucket), 0,
4762 SLAB_HWCACHE_ALIGN | SLAB_PANIC |
4765 tcp_hashinfo.bind2_bucket_cachep =
4766 kmem_cache_create("tcp_bind2_bucket",
4767 sizeof(struct inet_bind2_bucket), 0,
4768 SLAB_HWCACHE_ALIGN | SLAB_PANIC |
4772 /* Size and allocate the main established and bind bucket
4775 * The methodology is similar to that of the buffer cache.
4777 tcp_hashinfo.ehash =
4778 alloc_large_system_hash("TCP established",
4779 sizeof(struct inet_ehash_bucket),
4781 17, /* one slot per 128 KB of memory */
4784 &tcp_hashinfo.ehash_mask,
4786 thash_entries ? 0 : 512 * 1024);
4787 for (i = 0; i <= tcp_hashinfo.ehash_mask; i++)
4788 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].chain, i);
4790 if (inet_ehash_locks_alloc(&tcp_hashinfo))
4791 panic("TCP: failed to alloc ehash_locks");
4792 tcp_hashinfo.bhash =
4793 alloc_large_system_hash("TCP bind",
4794 2 * sizeof(struct inet_bind_hashbucket),
4795 tcp_hashinfo.ehash_mask + 1,
4796 17, /* one slot per 128 KB of memory */
4798 &tcp_hashinfo.bhash_size,
4802 tcp_hashinfo.bhash_size = 1U << tcp_hashinfo.bhash_size;
4803 tcp_hashinfo.bhash2 = tcp_hashinfo.bhash + tcp_hashinfo.bhash_size;
4804 for (i = 0; i < tcp_hashinfo.bhash_size; i++) {
4805 spin_lock_init(&tcp_hashinfo.bhash[i].lock);
4806 INIT_HLIST_HEAD(&tcp_hashinfo.bhash[i].chain);
4807 spin_lock_init(&tcp_hashinfo.bhash2[i].lock);
4808 INIT_HLIST_HEAD(&tcp_hashinfo.bhash2[i].chain);
4811 tcp_hashinfo.pernet = false;
4813 cnt = tcp_hashinfo.ehash_mask + 1;
4814 sysctl_tcp_max_orphans = cnt / 2;
4817 /* Set per-socket limits to no more than 1/128 the pressure threshold */
4818 limit = nr_free_buffer_pages() << (PAGE_SHIFT - 7);
4819 max_wshare = min(4UL*1024*1024, limit);
4820 max_rshare = min(6UL*1024*1024, limit);
4822 init_net.ipv4.sysctl_tcp_wmem[0] = PAGE_SIZE;
4823 init_net.ipv4.sysctl_tcp_wmem[1] = 16*1024;
4824 init_net.ipv4.sysctl_tcp_wmem[2] = max(64*1024, max_wshare);
4826 init_net.ipv4.sysctl_tcp_rmem[0] = PAGE_SIZE;
4827 init_net.ipv4.sysctl_tcp_rmem[1] = 131072;
4828 init_net.ipv4.sysctl_tcp_rmem[2] = max(131072, max_rshare);
4830 pr_info("Hash tables configured (established %u bind %u)\n",
4831 tcp_hashinfo.ehash_mask + 1, tcp_hashinfo.bhash_size);
4835 BUG_ON(tcp_register_congestion_control(&tcp_reno) != 0);