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);
504 sock_poll_wait(file, sock, wait);
506 state = inet_sk_state_load(sk);
507 if (state == TCP_LISTEN)
508 return inet_csk_listen_poll(sk);
510 /* Socket is not locked. We are protected from async events
511 * by poll logic and correct handling of state changes
512 * made by other threads is impossible in any case.
518 * EPOLLHUP is certainly not done right. But poll() doesn't
519 * have a notion of HUP in just one direction, and for a
520 * socket the read side is more interesting.
522 * Some poll() documentation says that EPOLLHUP is incompatible
523 * with the EPOLLOUT/POLLWR flags, so somebody should check this
524 * all. But careful, it tends to be safer to return too many
525 * bits than too few, and you can easily break real applications
526 * if you don't tell them that something has hung up!
530 * Check number 1. EPOLLHUP is _UNMASKABLE_ event (see UNIX98 and
531 * our fs/select.c). It means that after we received EOF,
532 * poll always returns immediately, making impossible poll() on write()
533 * in state CLOSE_WAIT. One solution is evident --- to set EPOLLHUP
534 * if and only if shutdown has been made in both directions.
535 * Actually, it is interesting to look how Solaris and DUX
536 * solve this dilemma. I would prefer, if EPOLLHUP were maskable,
537 * then we could set it on SND_SHUTDOWN. BTW examples given
538 * in Stevens' books assume exactly this behaviour, it explains
539 * why EPOLLHUP is incompatible with EPOLLOUT. --ANK
541 * NOTE. Check for TCP_CLOSE is added. The goal is to prevent
542 * blocking on fresh not-connected or disconnected socket. --ANK
544 shutdown = READ_ONCE(sk->sk_shutdown);
545 if (shutdown == SHUTDOWN_MASK || state == TCP_CLOSE)
547 if (shutdown & RCV_SHUTDOWN)
548 mask |= EPOLLIN | EPOLLRDNORM | EPOLLRDHUP;
550 /* Connected or passive Fast Open socket? */
551 if (state != TCP_SYN_SENT &&
552 (state != TCP_SYN_RECV || rcu_access_pointer(tp->fastopen_rsk))) {
553 int target = sock_rcvlowat(sk, 0, INT_MAX);
554 u16 urg_data = READ_ONCE(tp->urg_data);
556 if (unlikely(urg_data) &&
557 READ_ONCE(tp->urg_seq) == READ_ONCE(tp->copied_seq) &&
558 !sock_flag(sk, SOCK_URGINLINE))
561 if (tcp_stream_is_readable(sk, target))
562 mask |= EPOLLIN | EPOLLRDNORM;
564 if (!(shutdown & SEND_SHUTDOWN)) {
565 if (__sk_stream_is_writeable(sk, 1)) {
566 mask |= EPOLLOUT | EPOLLWRNORM;
567 } else { /* send SIGIO later */
568 sk_set_bit(SOCKWQ_ASYNC_NOSPACE, sk);
569 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
571 /* Race breaker. If space is freed after
572 * wspace test but before the flags are set,
573 * IO signal will be lost. Memory barrier
574 * pairs with the input side.
576 smp_mb__after_atomic();
577 if (__sk_stream_is_writeable(sk, 1))
578 mask |= EPOLLOUT | EPOLLWRNORM;
581 mask |= EPOLLOUT | EPOLLWRNORM;
583 if (urg_data & TCP_URG_VALID)
585 } else if (state == TCP_SYN_SENT && inet_sk(sk)->defer_connect) {
586 /* Active TCP fastopen socket with defer_connect
587 * Return EPOLLOUT so application can call write()
588 * in order for kernel to generate SYN+data
590 mask |= EPOLLOUT | EPOLLWRNORM;
592 /* This barrier is coupled with smp_wmb() in tcp_reset() */
594 if (sk->sk_err || !skb_queue_empty_lockless(&sk->sk_error_queue))
599 EXPORT_SYMBOL(tcp_poll);
601 int tcp_ioctl(struct sock *sk, int cmd, unsigned long arg)
603 struct tcp_sock *tp = tcp_sk(sk);
609 if (sk->sk_state == TCP_LISTEN)
612 slow = lock_sock_fast(sk);
614 unlock_sock_fast(sk, slow);
617 answ = READ_ONCE(tp->urg_data) &&
618 READ_ONCE(tp->urg_seq) == READ_ONCE(tp->copied_seq);
621 if (sk->sk_state == TCP_LISTEN)
624 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
627 answ = READ_ONCE(tp->write_seq) - tp->snd_una;
630 if (sk->sk_state == TCP_LISTEN)
633 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
636 answ = READ_ONCE(tp->write_seq) -
637 READ_ONCE(tp->snd_nxt);
643 return put_user(answ, (int __user *)arg);
645 EXPORT_SYMBOL(tcp_ioctl);
647 void tcp_mark_push(struct tcp_sock *tp, struct sk_buff *skb)
649 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
650 tp->pushed_seq = tp->write_seq;
653 static inline bool forced_push(const struct tcp_sock *tp)
655 return after(tp->write_seq, tp->pushed_seq + (tp->max_window >> 1));
658 void tcp_skb_entail(struct sock *sk, struct sk_buff *skb)
660 struct tcp_sock *tp = tcp_sk(sk);
661 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
663 tcb->seq = tcb->end_seq = tp->write_seq;
664 tcb->tcp_flags = TCPHDR_ACK;
665 __skb_header_release(skb);
666 tcp_add_write_queue_tail(sk, skb);
667 sk_wmem_queued_add(sk, skb->truesize);
668 sk_mem_charge(sk, skb->truesize);
669 if (tp->nonagle & TCP_NAGLE_PUSH)
670 tp->nonagle &= ~TCP_NAGLE_PUSH;
672 tcp_slow_start_after_idle_check(sk);
675 static inline void tcp_mark_urg(struct tcp_sock *tp, int flags)
678 tp->snd_up = tp->write_seq;
681 /* If a not yet filled skb is pushed, do not send it if
682 * we have data packets in Qdisc or NIC queues :
683 * Because TX completion will happen shortly, it gives a chance
684 * to coalesce future sendmsg() payload into this skb, without
685 * need for a timer, and with no latency trade off.
686 * As packets containing data payload have a bigger truesize
687 * than pure acks (dataless) packets, the last checks prevent
688 * autocorking if we only have an ACK in Qdisc/NIC queues,
689 * or if TX completion was delayed after we processed ACK packet.
691 static bool tcp_should_autocork(struct sock *sk, struct sk_buff *skb,
694 return skb->len < size_goal &&
695 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_autocorking) &&
696 !tcp_rtx_queue_empty(sk) &&
697 refcount_read(&sk->sk_wmem_alloc) > skb->truesize &&
698 tcp_skb_can_collapse_to(skb);
701 void tcp_push(struct sock *sk, int flags, int mss_now,
702 int nonagle, int size_goal)
704 struct tcp_sock *tp = tcp_sk(sk);
707 skb = tcp_write_queue_tail(sk);
710 if (!(flags & MSG_MORE) || forced_push(tp))
711 tcp_mark_push(tp, skb);
713 tcp_mark_urg(tp, flags);
715 if (tcp_should_autocork(sk, skb, size_goal)) {
717 /* avoid atomic op if TSQ_THROTTLED bit is already set */
718 if (!test_bit(TSQ_THROTTLED, &sk->sk_tsq_flags)) {
719 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPAUTOCORKING);
720 set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags);
721 smp_mb__after_atomic();
723 /* It is possible TX completion already happened
724 * before we set TSQ_THROTTLED.
726 if (refcount_read(&sk->sk_wmem_alloc) > skb->truesize)
730 if (flags & MSG_MORE)
731 nonagle = TCP_NAGLE_CORK;
733 __tcp_push_pending_frames(sk, mss_now, nonagle);
736 static int tcp_splice_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb,
737 unsigned int offset, size_t len)
739 struct tcp_splice_state *tss = rd_desc->arg.data;
742 ret = skb_splice_bits(skb, skb->sk, offset, tss->pipe,
743 min(rd_desc->count, len), tss->flags);
745 rd_desc->count -= ret;
749 static int __tcp_splice_read(struct sock *sk, struct tcp_splice_state *tss)
751 /* Store TCP splice context information in read_descriptor_t. */
752 read_descriptor_t rd_desc = {
757 return tcp_read_sock(sk, &rd_desc, tcp_splice_data_recv);
761 * tcp_splice_read - splice data from TCP socket to a pipe
762 * @sock: socket to splice from
763 * @ppos: position (not valid)
764 * @pipe: pipe to splice to
765 * @len: number of bytes to splice
766 * @flags: splice modifier flags
769 * Will read pages from given socket and fill them into a pipe.
772 ssize_t tcp_splice_read(struct socket *sock, loff_t *ppos,
773 struct pipe_inode_info *pipe, size_t len,
776 struct sock *sk = sock->sk;
777 struct tcp_splice_state tss = {
786 sock_rps_record_flow(sk);
788 * We can't seek on a socket input
797 timeo = sock_rcvtimeo(sk, sock->file->f_flags & O_NONBLOCK);
799 ret = __tcp_splice_read(sk, &tss);
805 if (sock_flag(sk, SOCK_DONE))
808 ret = sock_error(sk);
811 if (sk->sk_shutdown & RCV_SHUTDOWN)
813 if (sk->sk_state == TCP_CLOSE) {
815 * This occurs when user tries to read
816 * from never connected socket.
825 /* if __tcp_splice_read() got nothing while we have
826 * an skb in receive queue, we do not want to loop.
827 * This might happen with URG data.
829 if (!skb_queue_empty(&sk->sk_receive_queue))
831 ret = sk_wait_data(sk, &timeo, NULL);
834 if (signal_pending(current)) {
835 ret = sock_intr_errno(timeo);
848 if (sk->sk_err || sk->sk_state == TCP_CLOSE ||
849 (sk->sk_shutdown & RCV_SHUTDOWN) ||
850 signal_pending(current))
861 EXPORT_SYMBOL(tcp_splice_read);
863 struct sk_buff *tcp_stream_alloc_skb(struct sock *sk, int size, gfp_t gfp,
868 skb = alloc_skb_fclone(size + MAX_TCP_HEADER, gfp);
872 skb->truesize = SKB_TRUESIZE(skb_end_offset(skb));
873 if (force_schedule) {
874 mem_scheduled = true;
875 sk_forced_mem_schedule(sk, skb->truesize);
877 mem_scheduled = sk_wmem_schedule(sk, skb->truesize);
879 if (likely(mem_scheduled)) {
880 skb_reserve(skb, MAX_TCP_HEADER);
881 skb->ip_summed = CHECKSUM_PARTIAL;
882 INIT_LIST_HEAD(&skb->tcp_tsorted_anchor);
887 sk->sk_prot->enter_memory_pressure(sk);
888 sk_stream_moderate_sndbuf(sk);
893 static unsigned int tcp_xmit_size_goal(struct sock *sk, u32 mss_now,
896 struct tcp_sock *tp = tcp_sk(sk);
897 u32 new_size_goal, size_goal;
902 /* Note : tcp_tso_autosize() will eventually split this later */
903 new_size_goal = tcp_bound_to_half_wnd(tp, sk->sk_gso_max_size);
905 /* We try hard to avoid divides here */
906 size_goal = tp->gso_segs * mss_now;
907 if (unlikely(new_size_goal < size_goal ||
908 new_size_goal >= size_goal + mss_now)) {
909 tp->gso_segs = min_t(u16, new_size_goal / mss_now,
910 sk->sk_gso_max_segs);
911 size_goal = tp->gso_segs * mss_now;
914 return max(size_goal, mss_now);
917 int tcp_send_mss(struct sock *sk, int *size_goal, int flags)
921 mss_now = tcp_current_mss(sk);
922 *size_goal = tcp_xmit_size_goal(sk, mss_now, !(flags & MSG_OOB));
927 /* In some cases, both sendpage() and sendmsg() could have added
928 * an skb to the write queue, but failed adding payload on it.
929 * We need to remove it to consume less memory, but more
930 * importantly be able to generate EPOLLOUT for Edge Trigger epoll()
933 void tcp_remove_empty_skb(struct sock *sk)
935 struct sk_buff *skb = tcp_write_queue_tail(sk);
937 if (skb && TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq) {
938 tcp_unlink_write_queue(skb, sk);
939 if (tcp_write_queue_empty(sk))
940 tcp_chrono_stop(sk, TCP_CHRONO_BUSY);
941 tcp_wmem_free_skb(sk, skb);
945 /* skb changing from pure zc to mixed, must charge zc */
946 static int tcp_downgrade_zcopy_pure(struct sock *sk, struct sk_buff *skb)
948 if (unlikely(skb_zcopy_pure(skb))) {
949 u32 extra = skb->truesize -
950 SKB_TRUESIZE(skb_end_offset(skb));
952 if (!sk_wmem_schedule(sk, extra))
955 sk_mem_charge(sk, extra);
956 skb_shinfo(skb)->flags &= ~SKBFL_PURE_ZEROCOPY;
962 static int tcp_wmem_schedule(struct sock *sk, int copy)
966 if (likely(sk_wmem_schedule(sk, copy)))
969 /* We could be in trouble if we have nothing queued.
970 * Use whatever is left in sk->sk_forward_alloc and tcp_wmem[0]
971 * to guarantee some progress.
973 left = sock_net(sk)->ipv4.sysctl_tcp_wmem[0] - sk->sk_wmem_queued;
975 sk_forced_mem_schedule(sk, min(left, copy));
976 return min(copy, sk->sk_forward_alloc);
979 static struct sk_buff *tcp_build_frag(struct sock *sk, int size_goal, int flags,
980 struct page *page, int offset, size_t *size)
982 struct sk_buff *skb = tcp_write_queue_tail(sk);
983 struct tcp_sock *tp = tcp_sk(sk);
987 if (!skb || (copy = size_goal - skb->len) <= 0 ||
988 !tcp_skb_can_collapse_to(skb)) {
990 if (!sk_stream_memory_free(sk))
993 skb = tcp_stream_alloc_skb(sk, 0, sk->sk_allocation,
994 tcp_rtx_and_write_queues_empty(sk));
998 #ifdef CONFIG_TLS_DEVICE
999 skb->decrypted = !!(flags & MSG_SENDPAGE_DECRYPTED);
1001 tcp_skb_entail(sk, skb);
1008 i = skb_shinfo(skb)->nr_frags;
1009 can_coalesce = skb_can_coalesce(skb, i, page, offset);
1010 if (!can_coalesce && i >= READ_ONCE(sysctl_max_skb_frags)) {
1011 tcp_mark_push(tp, skb);
1014 if (tcp_downgrade_zcopy_pure(sk, skb))
1017 copy = tcp_wmem_schedule(sk, copy);
1022 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1025 skb_fill_page_desc_noacc(skb, i, page, offset, copy);
1028 if (!(flags & MSG_NO_SHARED_FRAGS))
1029 skb_shinfo(skb)->flags |= SKBFL_SHARED_FRAG;
1032 skb->data_len += copy;
1033 skb->truesize += copy;
1034 sk_wmem_queued_add(sk, copy);
1035 sk_mem_charge(sk, copy);
1036 WRITE_ONCE(tp->write_seq, tp->write_seq + copy);
1037 TCP_SKB_CB(skb)->end_seq += copy;
1038 tcp_skb_pcount_set(skb, 0);
1044 ssize_t do_tcp_sendpages(struct sock *sk, struct page *page, int offset,
1045 size_t size, int flags)
1047 struct tcp_sock *tp = tcp_sk(sk);
1048 int mss_now, size_goal;
1051 long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
1053 if (IS_ENABLED(CONFIG_DEBUG_VM) &&
1054 WARN_ONCE(!sendpage_ok(page),
1055 "page must not be a Slab one and have page_count > 0"))
1058 /* Wait for a connection to finish. One exception is TCP Fast Open
1059 * (passive side) where data is allowed to be sent before a connection
1060 * is fully established.
1062 if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) &&
1063 !tcp_passive_fastopen(sk)) {
1064 err = sk_stream_wait_connect(sk, &timeo);
1069 sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
1071 mss_now = tcp_send_mss(sk, &size_goal, flags);
1075 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
1079 struct sk_buff *skb;
1082 skb = tcp_build_frag(sk, size_goal, flags, page, offset, ©);
1084 goto wait_for_space;
1087 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
1095 if (skb->len < size_goal || (flags & MSG_OOB))
1098 if (forced_push(tp)) {
1099 tcp_mark_push(tp, skb);
1100 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
1101 } else if (skb == tcp_send_head(sk))
1102 tcp_push_one(sk, mss_now);
1106 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1107 tcp_push(sk, flags & ~MSG_MORE, mss_now,
1108 TCP_NAGLE_PUSH, size_goal);
1110 err = sk_stream_wait_memory(sk, &timeo);
1114 mss_now = tcp_send_mss(sk, &size_goal, flags);
1119 tcp_tx_timestamp(sk, sk->sk_tsflags);
1120 if (!(flags & MSG_SENDPAGE_NOTLAST))
1121 tcp_push(sk, flags, mss_now, tp->nonagle, size_goal);
1126 tcp_remove_empty_skb(sk);
1130 /* make sure we wake any epoll edge trigger waiter */
1131 if (unlikely(tcp_rtx_and_write_queues_empty(sk) && err == -EAGAIN)) {
1132 sk->sk_write_space(sk);
1133 tcp_chrono_stop(sk, TCP_CHRONO_SNDBUF_LIMITED);
1135 return sk_stream_error(sk, flags, err);
1137 EXPORT_SYMBOL_GPL(do_tcp_sendpages);
1139 int tcp_sendpage_locked(struct sock *sk, struct page *page, int offset,
1140 size_t size, int flags)
1142 if (!(sk->sk_route_caps & NETIF_F_SG))
1143 return sock_no_sendpage_locked(sk, page, offset, size, flags);
1145 tcp_rate_check_app_limited(sk); /* is sending application-limited? */
1147 return do_tcp_sendpages(sk, page, offset, size, flags);
1149 EXPORT_SYMBOL_GPL(tcp_sendpage_locked);
1151 int tcp_sendpage(struct sock *sk, struct page *page, int offset,
1152 size_t size, int flags)
1157 ret = tcp_sendpage_locked(sk, page, offset, size, flags);
1162 EXPORT_SYMBOL(tcp_sendpage);
1164 void tcp_free_fastopen_req(struct tcp_sock *tp)
1166 if (tp->fastopen_req) {
1167 kfree(tp->fastopen_req);
1168 tp->fastopen_req = NULL;
1172 int tcp_sendmsg_fastopen(struct sock *sk, struct msghdr *msg, int *copied,
1173 size_t size, struct ubuf_info *uarg)
1175 struct tcp_sock *tp = tcp_sk(sk);
1176 struct inet_sock *inet = inet_sk(sk);
1177 struct sockaddr *uaddr = msg->msg_name;
1180 if (!(READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_fastopen) &
1181 TFO_CLIENT_ENABLE) ||
1182 (uaddr && msg->msg_namelen >= sizeof(uaddr->sa_family) &&
1183 uaddr->sa_family == AF_UNSPEC))
1185 if (tp->fastopen_req)
1186 return -EALREADY; /* Another Fast Open is in progress */
1188 tp->fastopen_req = kzalloc(sizeof(struct tcp_fastopen_request),
1190 if (unlikely(!tp->fastopen_req))
1192 tp->fastopen_req->data = msg;
1193 tp->fastopen_req->size = size;
1194 tp->fastopen_req->uarg = uarg;
1196 if (inet->defer_connect) {
1197 err = tcp_connect(sk);
1198 /* Same failure procedure as in tcp_v4/6_connect */
1200 tcp_set_state(sk, TCP_CLOSE);
1201 inet->inet_dport = 0;
1202 sk->sk_route_caps = 0;
1205 flags = (msg->msg_flags & MSG_DONTWAIT) ? O_NONBLOCK : 0;
1206 err = __inet_stream_connect(sk->sk_socket, uaddr,
1207 msg->msg_namelen, flags, 1);
1208 /* fastopen_req could already be freed in __inet_stream_connect
1209 * if the connection times out or gets rst
1211 if (tp->fastopen_req) {
1212 *copied = tp->fastopen_req->copied;
1213 tcp_free_fastopen_req(tp);
1214 inet->defer_connect = 0;
1219 int tcp_sendmsg_locked(struct sock *sk, struct msghdr *msg, size_t size)
1221 struct tcp_sock *tp = tcp_sk(sk);
1222 struct ubuf_info *uarg = NULL;
1223 struct sk_buff *skb;
1224 struct sockcm_cookie sockc;
1225 int flags, err, copied = 0;
1226 int mss_now = 0, size_goal, copied_syn = 0;
1227 int process_backlog = 0;
1231 flags = msg->msg_flags;
1233 if ((flags & MSG_ZEROCOPY) && size) {
1234 skb = tcp_write_queue_tail(sk);
1236 if (msg->msg_ubuf) {
1237 uarg = msg->msg_ubuf;
1238 net_zcopy_get(uarg);
1239 zc = sk->sk_route_caps & NETIF_F_SG;
1240 } else if (sock_flag(sk, SOCK_ZEROCOPY)) {
1241 uarg = msg_zerocopy_realloc(sk, size, skb_zcopy(skb));
1246 zc = sk->sk_route_caps & NETIF_F_SG;
1248 uarg_to_msgzc(uarg)->zerocopy = 0;
1252 if (unlikely(flags & MSG_FASTOPEN || inet_sk(sk)->defer_connect) &&
1254 err = tcp_sendmsg_fastopen(sk, msg, &copied_syn, size, uarg);
1255 if (err == -EINPROGRESS && copied_syn > 0)
1261 timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
1263 tcp_rate_check_app_limited(sk); /* is sending application-limited? */
1265 /* Wait for a connection to finish. One exception is TCP Fast Open
1266 * (passive side) where data is allowed to be sent before a connection
1267 * is fully established.
1269 if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) &&
1270 !tcp_passive_fastopen(sk)) {
1271 err = sk_stream_wait_connect(sk, &timeo);
1276 if (unlikely(tp->repair)) {
1277 if (tp->repair_queue == TCP_RECV_QUEUE) {
1278 copied = tcp_send_rcvq(sk, msg, size);
1283 if (tp->repair_queue == TCP_NO_QUEUE)
1286 /* 'common' sending to sendq */
1289 sockcm_init(&sockc, sk);
1290 if (msg->msg_controllen) {
1291 err = sock_cmsg_send(sk, msg, &sockc);
1292 if (unlikely(err)) {
1298 /* This should be in poll */
1299 sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
1301 /* Ok commence sending. */
1305 mss_now = tcp_send_mss(sk, &size_goal, flags);
1308 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
1311 while (msg_data_left(msg)) {
1314 skb = tcp_write_queue_tail(sk);
1316 copy = size_goal - skb->len;
1318 if (copy <= 0 || !tcp_skb_can_collapse_to(skb)) {
1322 if (!sk_stream_memory_free(sk))
1323 goto wait_for_space;
1325 if (unlikely(process_backlog >= 16)) {
1326 process_backlog = 0;
1327 if (sk_flush_backlog(sk))
1330 first_skb = tcp_rtx_and_write_queues_empty(sk);
1331 skb = tcp_stream_alloc_skb(sk, 0, sk->sk_allocation,
1334 goto wait_for_space;
1338 tcp_skb_entail(sk, skb);
1341 /* All packets are restored as if they have
1342 * already been sent. skb_mstamp_ns isn't set to
1343 * avoid wrong rtt estimation.
1346 TCP_SKB_CB(skb)->sacked |= TCPCB_REPAIRED;
1349 /* Try to append data to the end of skb. */
1350 if (copy > msg_data_left(msg))
1351 copy = msg_data_left(msg);
1355 int i = skb_shinfo(skb)->nr_frags;
1356 struct page_frag *pfrag = sk_page_frag(sk);
1358 if (!sk_page_frag_refill(sk, pfrag))
1359 goto wait_for_space;
1361 if (!skb_can_coalesce(skb, i, pfrag->page,
1363 if (i >= READ_ONCE(sysctl_max_skb_frags)) {
1364 tcp_mark_push(tp, skb);
1370 copy = min_t(int, copy, pfrag->size - pfrag->offset);
1372 if (unlikely(skb_zcopy_pure(skb) || skb_zcopy_managed(skb))) {
1373 if (tcp_downgrade_zcopy_pure(sk, skb))
1374 goto wait_for_space;
1375 skb_zcopy_downgrade_managed(skb);
1378 copy = tcp_wmem_schedule(sk, copy);
1380 goto wait_for_space;
1382 err = skb_copy_to_page_nocache(sk, &msg->msg_iter, skb,
1389 /* Update the skb. */
1391 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1393 skb_fill_page_desc(skb, i, pfrag->page,
1394 pfrag->offset, copy);
1395 page_ref_inc(pfrag->page);
1397 pfrag->offset += copy;
1399 /* First append to a fragless skb builds initial
1403 skb_shinfo(skb)->flags |= SKBFL_PURE_ZEROCOPY;
1405 if (!skb_zcopy_pure(skb)) {
1406 copy = tcp_wmem_schedule(sk, copy);
1408 goto wait_for_space;
1411 err = skb_zerocopy_iter_stream(sk, skb, msg, copy, uarg);
1412 if (err == -EMSGSIZE || err == -EEXIST) {
1413 tcp_mark_push(tp, skb);
1422 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
1424 WRITE_ONCE(tp->write_seq, tp->write_seq + copy);
1425 TCP_SKB_CB(skb)->end_seq += copy;
1426 tcp_skb_pcount_set(skb, 0);
1429 if (!msg_data_left(msg)) {
1430 if (unlikely(flags & MSG_EOR))
1431 TCP_SKB_CB(skb)->eor = 1;
1435 if (skb->len < size_goal || (flags & MSG_OOB) || unlikely(tp->repair))
1438 if (forced_push(tp)) {
1439 tcp_mark_push(tp, skb);
1440 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
1441 } else if (skb == tcp_send_head(sk))
1442 tcp_push_one(sk, mss_now);
1446 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1448 tcp_push(sk, flags & ~MSG_MORE, mss_now,
1449 TCP_NAGLE_PUSH, size_goal);
1451 err = sk_stream_wait_memory(sk, &timeo);
1455 mss_now = tcp_send_mss(sk, &size_goal, flags);
1460 tcp_tx_timestamp(sk, sockc.tsflags);
1461 tcp_push(sk, flags, mss_now, tp->nonagle, size_goal);
1464 net_zcopy_put(uarg);
1465 return copied + copied_syn;
1468 tcp_remove_empty_skb(sk);
1470 if (copied + copied_syn)
1473 net_zcopy_put_abort(uarg, true);
1474 err = sk_stream_error(sk, flags, err);
1475 /* make sure we wake any epoll edge trigger waiter */
1476 if (unlikely(tcp_rtx_and_write_queues_empty(sk) && err == -EAGAIN)) {
1477 sk->sk_write_space(sk);
1478 tcp_chrono_stop(sk, TCP_CHRONO_SNDBUF_LIMITED);
1482 EXPORT_SYMBOL_GPL(tcp_sendmsg_locked);
1484 int tcp_sendmsg(struct sock *sk, struct msghdr *msg, size_t size)
1489 ret = tcp_sendmsg_locked(sk, msg, size);
1494 EXPORT_SYMBOL(tcp_sendmsg);
1496 void tcp_splice_eof(struct socket *sock)
1498 struct sock *sk = sock->sk;
1499 struct tcp_sock *tp = tcp_sk(sk);
1500 int mss_now, size_goal;
1502 if (!tcp_write_queue_tail(sk))
1506 mss_now = tcp_send_mss(sk, &size_goal, 0);
1507 tcp_push(sk, 0, mss_now, tp->nonagle, size_goal);
1510 EXPORT_SYMBOL_GPL(tcp_splice_eof);
1513 * Handle reading urgent data. BSD has very simple semantics for
1514 * this, no blocking and very strange errors 8)
1517 static int tcp_recv_urg(struct sock *sk, struct msghdr *msg, int len, int flags)
1519 struct tcp_sock *tp = tcp_sk(sk);
1521 /* No URG data to read. */
1522 if (sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data ||
1523 tp->urg_data == TCP_URG_READ)
1524 return -EINVAL; /* Yes this is right ! */
1526 if (sk->sk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DONE))
1529 if (tp->urg_data & TCP_URG_VALID) {
1531 char c = tp->urg_data;
1533 if (!(flags & MSG_PEEK))
1534 WRITE_ONCE(tp->urg_data, TCP_URG_READ);
1536 /* Read urgent data. */
1537 msg->msg_flags |= MSG_OOB;
1540 if (!(flags & MSG_TRUNC))
1541 err = memcpy_to_msg(msg, &c, 1);
1544 msg->msg_flags |= MSG_TRUNC;
1546 return err ? -EFAULT : len;
1549 if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
1552 /* Fixed the recv(..., MSG_OOB) behaviour. BSD docs and
1553 * the available implementations agree in this case:
1554 * this call should never block, independent of the
1555 * blocking state of the socket.
1556 * Mike <pall@rz.uni-karlsruhe.de>
1561 static int tcp_peek_sndq(struct sock *sk, struct msghdr *msg, int len)
1563 struct sk_buff *skb;
1564 int copied = 0, err = 0;
1566 /* XXX -- need to support SO_PEEK_OFF */
1568 skb_rbtree_walk(skb, &sk->tcp_rtx_queue) {
1569 err = skb_copy_datagram_msg(skb, 0, msg, skb->len);
1575 skb_queue_walk(&sk->sk_write_queue, skb) {
1576 err = skb_copy_datagram_msg(skb, 0, msg, skb->len);
1583 return err ?: copied;
1586 /* Clean up the receive buffer for full frames taken by the user,
1587 * then send an ACK if necessary. COPIED is the number of bytes
1588 * tcp_recvmsg has given to the user so far, it speeds up the
1589 * calculation of whether or not we must ACK for the sake of
1592 void __tcp_cleanup_rbuf(struct sock *sk, int copied)
1594 struct tcp_sock *tp = tcp_sk(sk);
1595 bool time_to_ack = false;
1597 if (inet_csk_ack_scheduled(sk)) {
1598 const struct inet_connection_sock *icsk = inet_csk(sk);
1600 if (/* Once-per-two-segments ACK was not sent by tcp_input.c */
1601 tp->rcv_nxt - tp->rcv_wup > icsk->icsk_ack.rcv_mss ||
1603 * If this read emptied read buffer, we send ACK, if
1604 * connection is not bidirectional, user drained
1605 * receive buffer and there was a small segment
1609 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED2) ||
1610 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED) &&
1611 !inet_csk_in_pingpong_mode(sk))) &&
1612 !atomic_read(&sk->sk_rmem_alloc)))
1616 /* We send an ACK if we can now advertise a non-zero window
1617 * which has been raised "significantly".
1619 * Even if window raised up to infinity, do not send window open ACK
1620 * in states, where we will not receive more. It is useless.
1622 if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) {
1623 __u32 rcv_window_now = tcp_receive_window(tp);
1625 /* Optimize, __tcp_select_window() is not cheap. */
1626 if (2*rcv_window_now <= tp->window_clamp) {
1627 __u32 new_window = __tcp_select_window(sk);
1629 /* Send ACK now, if this read freed lots of space
1630 * in our buffer. Certainly, new_window is new window.
1631 * We can advertise it now, if it is not less than current one.
1632 * "Lots" means "at least twice" here.
1634 if (new_window && new_window >= 2 * rcv_window_now)
1642 void tcp_cleanup_rbuf(struct sock *sk, int copied)
1644 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
1645 struct tcp_sock *tp = tcp_sk(sk);
1647 WARN(skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq),
1648 "cleanup rbuf bug: copied %X seq %X rcvnxt %X\n",
1649 tp->copied_seq, TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt);
1650 __tcp_cleanup_rbuf(sk, copied);
1653 static void tcp_eat_recv_skb(struct sock *sk, struct sk_buff *skb)
1655 __skb_unlink(skb, &sk->sk_receive_queue);
1656 if (likely(skb->destructor == sock_rfree)) {
1658 skb->destructor = NULL;
1660 return skb_attempt_defer_free(skb);
1665 struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off)
1667 struct sk_buff *skb;
1670 while ((skb = skb_peek(&sk->sk_receive_queue)) != NULL) {
1671 offset = seq - TCP_SKB_CB(skb)->seq;
1672 if (unlikely(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
1673 pr_err_once("%s: found a SYN, please report !\n", __func__);
1676 if (offset < skb->len || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)) {
1680 /* This looks weird, but this can happen if TCP collapsing
1681 * splitted a fat GRO packet, while we released socket lock
1682 * in skb_splice_bits()
1684 tcp_eat_recv_skb(sk, skb);
1688 EXPORT_SYMBOL(tcp_recv_skb);
1691 * This routine provides an alternative to tcp_recvmsg() for routines
1692 * that would like to handle copying from skbuffs directly in 'sendfile'
1695 * - It is assumed that the socket was locked by the caller.
1696 * - The routine does not block.
1697 * - At present, there is no support for reading OOB data
1698 * or for 'peeking' the socket using this routine
1699 * (although both would be easy to implement).
1701 int tcp_read_sock(struct sock *sk, read_descriptor_t *desc,
1702 sk_read_actor_t recv_actor)
1704 struct sk_buff *skb;
1705 struct tcp_sock *tp = tcp_sk(sk);
1706 u32 seq = tp->copied_seq;
1710 if (sk->sk_state == TCP_LISTEN)
1712 while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1713 if (offset < skb->len) {
1717 len = skb->len - offset;
1718 /* Stop reading if we hit a patch of urgent data */
1719 if (unlikely(tp->urg_data)) {
1720 u32 urg_offset = tp->urg_seq - seq;
1721 if (urg_offset < len)
1726 used = recv_actor(desc, skb, offset, len);
1732 if (WARN_ON_ONCE(used > len))
1738 /* If recv_actor drops the lock (e.g. TCP splice
1739 * receive) the skb pointer might be invalid when
1740 * getting here: tcp_collapse might have deleted it
1741 * while aggregating skbs from the socket queue.
1743 skb = tcp_recv_skb(sk, seq - 1, &offset);
1746 /* TCP coalescing might have appended data to the skb.
1747 * Try to splice more frags
1749 if (offset + 1 != skb->len)
1752 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) {
1753 tcp_eat_recv_skb(sk, skb);
1757 tcp_eat_recv_skb(sk, skb);
1760 WRITE_ONCE(tp->copied_seq, seq);
1762 WRITE_ONCE(tp->copied_seq, seq);
1764 tcp_rcv_space_adjust(sk);
1766 /* Clean up data we have read: This will do ACK frames. */
1768 tcp_recv_skb(sk, seq, &offset);
1769 tcp_cleanup_rbuf(sk, copied);
1773 EXPORT_SYMBOL(tcp_read_sock);
1775 int tcp_read_skb(struct sock *sk, skb_read_actor_t recv_actor)
1777 struct sk_buff *skb;
1780 if (sk->sk_state == TCP_LISTEN)
1783 while ((skb = skb_peek(&sk->sk_receive_queue)) != NULL) {
1787 __skb_unlink(skb, &sk->sk_receive_queue);
1788 WARN_ON_ONCE(!skb_set_owner_sk_safe(skb, sk));
1789 tcp_flags = TCP_SKB_CB(skb)->tcp_flags;
1790 used = recv_actor(sk, skb);
1798 if (tcp_flags & TCPHDR_FIN)
1803 EXPORT_SYMBOL(tcp_read_skb);
1805 void tcp_read_done(struct sock *sk, size_t len)
1807 struct tcp_sock *tp = tcp_sk(sk);
1808 u32 seq = tp->copied_seq;
1809 struct sk_buff *skb;
1813 if (sk->sk_state == TCP_LISTEN)
1817 while (left && (skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1820 used = min_t(size_t, skb->len - offset, left);
1824 if (skb->len > offset + used)
1827 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) {
1828 tcp_eat_recv_skb(sk, skb);
1832 tcp_eat_recv_skb(sk, skb);
1834 WRITE_ONCE(tp->copied_seq, seq);
1836 tcp_rcv_space_adjust(sk);
1838 /* Clean up data we have read: This will do ACK frames. */
1840 tcp_cleanup_rbuf(sk, len - left);
1842 EXPORT_SYMBOL(tcp_read_done);
1844 int tcp_peek_len(struct socket *sock)
1846 return tcp_inq(sock->sk);
1848 EXPORT_SYMBOL(tcp_peek_len);
1850 /* Make sure sk_rcvbuf is big enough to satisfy SO_RCVLOWAT hint */
1851 int tcp_set_rcvlowat(struct sock *sk, int val)
1855 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK)
1856 cap = sk->sk_rcvbuf >> 1;
1858 cap = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_rmem[2]) >> 1;
1859 val = min(val, cap);
1860 WRITE_ONCE(sk->sk_rcvlowat, val ? : 1);
1862 /* Check if we need to signal EPOLLIN right now */
1865 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK)
1869 if (val > sk->sk_rcvbuf) {
1870 WRITE_ONCE(sk->sk_rcvbuf, val);
1871 tcp_sk(sk)->window_clamp = tcp_win_from_space(sk, val);
1875 EXPORT_SYMBOL(tcp_set_rcvlowat);
1877 void tcp_update_recv_tstamps(struct sk_buff *skb,
1878 struct scm_timestamping_internal *tss)
1881 tss->ts[0] = ktime_to_timespec64(skb->tstamp);
1883 tss->ts[0] = (struct timespec64) {0};
1885 if (skb_hwtstamps(skb)->hwtstamp)
1886 tss->ts[2] = ktime_to_timespec64(skb_hwtstamps(skb)->hwtstamp);
1888 tss->ts[2] = (struct timespec64) {0};
1892 static const struct vm_operations_struct tcp_vm_ops = {
1895 int tcp_mmap(struct file *file, struct socket *sock,
1896 struct vm_area_struct *vma)
1898 if (vma->vm_flags & (VM_WRITE | VM_EXEC))
1900 vma->vm_flags &= ~(VM_MAYWRITE | VM_MAYEXEC);
1902 /* Instruct vm_insert_page() to not mmap_read_lock(mm) */
1903 vma->vm_flags |= VM_MIXEDMAP;
1905 vma->vm_ops = &tcp_vm_ops;
1908 EXPORT_SYMBOL(tcp_mmap);
1910 static skb_frag_t *skb_advance_to_frag(struct sk_buff *skb, u32 offset_skb,
1915 if (unlikely(offset_skb >= skb->len))
1918 offset_skb -= skb_headlen(skb);
1919 if ((int)offset_skb < 0 || skb_has_frag_list(skb))
1922 frag = skb_shinfo(skb)->frags;
1923 while (offset_skb) {
1924 if (skb_frag_size(frag) > offset_skb) {
1925 *offset_frag = offset_skb;
1928 offset_skb -= skb_frag_size(frag);
1935 static bool can_map_frag(const skb_frag_t *frag)
1939 if (skb_frag_size(frag) != PAGE_SIZE || skb_frag_off(frag))
1942 page = skb_frag_page(frag);
1944 if (PageCompound(page) || page->mapping)
1950 static int find_next_mappable_frag(const skb_frag_t *frag,
1951 int remaining_in_skb)
1955 if (likely(can_map_frag(frag)))
1958 while (offset < remaining_in_skb && !can_map_frag(frag)) {
1959 offset += skb_frag_size(frag);
1965 static void tcp_zerocopy_set_hint_for_skb(struct sock *sk,
1966 struct tcp_zerocopy_receive *zc,
1967 struct sk_buff *skb, u32 offset)
1969 u32 frag_offset, partial_frag_remainder = 0;
1970 int mappable_offset;
1973 /* worst case: skip to next skb. try to improve on this case below */
1974 zc->recv_skip_hint = skb->len - offset;
1976 /* Find the frag containing this offset (and how far into that frag) */
1977 frag = skb_advance_to_frag(skb, offset, &frag_offset);
1982 struct skb_shared_info *info = skb_shinfo(skb);
1984 /* We read part of the last frag, must recvmsg() rest of skb. */
1985 if (frag == &info->frags[info->nr_frags - 1])
1988 /* Else, we must at least read the remainder in this frag. */
1989 partial_frag_remainder = skb_frag_size(frag) - frag_offset;
1990 zc->recv_skip_hint -= partial_frag_remainder;
1994 /* partial_frag_remainder: If part way through a frag, must read rest.
1995 * mappable_offset: Bytes till next mappable frag, *not* counting bytes
1996 * in partial_frag_remainder.
1998 mappable_offset = find_next_mappable_frag(frag, zc->recv_skip_hint);
1999 zc->recv_skip_hint = mappable_offset + partial_frag_remainder;
2002 static int tcp_recvmsg_locked(struct sock *sk, struct msghdr *msg, size_t len,
2003 int flags, struct scm_timestamping_internal *tss,
2005 static int receive_fallback_to_copy(struct sock *sk,
2006 struct tcp_zerocopy_receive *zc, int inq,
2007 struct scm_timestamping_internal *tss)
2009 unsigned long copy_address = (unsigned long)zc->copybuf_address;
2010 struct msghdr msg = {};
2015 zc->recv_skip_hint = 0;
2017 if (copy_address != zc->copybuf_address)
2020 err = import_single_range(ITER_DEST, (void __user *)copy_address,
2021 inq, &iov, &msg.msg_iter);
2025 err = tcp_recvmsg_locked(sk, &msg, inq, MSG_DONTWAIT,
2026 tss, &zc->msg_flags);
2030 zc->copybuf_len = err;
2031 if (likely(zc->copybuf_len)) {
2032 struct sk_buff *skb;
2035 skb = tcp_recv_skb(sk, tcp_sk(sk)->copied_seq, &offset);
2037 tcp_zerocopy_set_hint_for_skb(sk, zc, skb, offset);
2042 static int tcp_copy_straggler_data(struct tcp_zerocopy_receive *zc,
2043 struct sk_buff *skb, u32 copylen,
2044 u32 *offset, u32 *seq)
2046 unsigned long copy_address = (unsigned long)zc->copybuf_address;
2047 struct msghdr msg = {};
2051 if (copy_address != zc->copybuf_address)
2054 err = import_single_range(ITER_DEST, (void __user *)copy_address,
2055 copylen, &iov, &msg.msg_iter);
2058 err = skb_copy_datagram_msg(skb, *offset, &msg, copylen);
2061 zc->recv_skip_hint -= copylen;
2064 return (__s32)copylen;
2067 static int tcp_zc_handle_leftover(struct tcp_zerocopy_receive *zc,
2069 struct sk_buff *skb,
2072 struct scm_timestamping_internal *tss)
2074 u32 offset, copylen = min_t(u32, copybuf_len, zc->recv_skip_hint);
2078 /* skb is null if inq < PAGE_SIZE. */
2080 offset = *seq - TCP_SKB_CB(skb)->seq;
2082 skb = tcp_recv_skb(sk, *seq, &offset);
2083 if (TCP_SKB_CB(skb)->has_rxtstamp) {
2084 tcp_update_recv_tstamps(skb, tss);
2085 zc->msg_flags |= TCP_CMSG_TS;
2089 zc->copybuf_len = tcp_copy_straggler_data(zc, skb, copylen, &offset,
2091 return zc->copybuf_len < 0 ? 0 : copylen;
2094 static int tcp_zerocopy_vm_insert_batch_error(struct vm_area_struct *vma,
2095 struct page **pending_pages,
2096 unsigned long pages_remaining,
2097 unsigned long *address,
2100 struct tcp_zerocopy_receive *zc,
2101 u32 total_bytes_to_map,
2104 /* At least one page did not map. Try zapping if we skipped earlier. */
2105 if (err == -EBUSY &&
2106 zc->flags & TCP_RECEIVE_ZEROCOPY_FLAG_TLB_CLEAN_HINT) {
2109 maybe_zap_len = total_bytes_to_map - /* All bytes to map */
2110 *length + /* Mapped or pending */
2111 (pages_remaining * PAGE_SIZE); /* Failed map. */
2112 zap_page_range(vma, *address, maybe_zap_len);
2117 unsigned long leftover_pages = pages_remaining;
2120 /* We called zap_page_range, try to reinsert. */
2121 err = vm_insert_pages(vma, *address,
2124 bytes_mapped = PAGE_SIZE * (leftover_pages - pages_remaining);
2125 *seq += bytes_mapped;
2126 *address += bytes_mapped;
2129 /* Either we were unable to zap, OR we zapped, retried an
2130 * insert, and still had an issue. Either ways, pages_remaining
2131 * is the number of pages we were unable to map, and we unroll
2132 * some state we speculatively touched before.
2134 const int bytes_not_mapped = PAGE_SIZE * pages_remaining;
2136 *length -= bytes_not_mapped;
2137 zc->recv_skip_hint += bytes_not_mapped;
2142 static int tcp_zerocopy_vm_insert_batch(struct vm_area_struct *vma,
2143 struct page **pages,
2144 unsigned int pages_to_map,
2145 unsigned long *address,
2148 struct tcp_zerocopy_receive *zc,
2149 u32 total_bytes_to_map)
2151 unsigned long pages_remaining = pages_to_map;
2152 unsigned int pages_mapped;
2153 unsigned int bytes_mapped;
2156 err = vm_insert_pages(vma, *address, pages, &pages_remaining);
2157 pages_mapped = pages_to_map - (unsigned int)pages_remaining;
2158 bytes_mapped = PAGE_SIZE * pages_mapped;
2159 /* Even if vm_insert_pages fails, it may have partially succeeded in
2160 * mapping (some but not all of the pages).
2162 *seq += bytes_mapped;
2163 *address += bytes_mapped;
2168 /* Error: maybe zap and retry + rollback state for failed inserts. */
2169 return tcp_zerocopy_vm_insert_batch_error(vma, pages + pages_mapped,
2170 pages_remaining, address, length, seq, zc, total_bytes_to_map,
2174 #define TCP_VALID_ZC_MSG_FLAGS (TCP_CMSG_TS)
2175 static void tcp_zc_finalize_rx_tstamp(struct sock *sk,
2176 struct tcp_zerocopy_receive *zc,
2177 struct scm_timestamping_internal *tss)
2179 unsigned long msg_control_addr;
2180 struct msghdr cmsg_dummy;
2182 msg_control_addr = (unsigned long)zc->msg_control;
2183 cmsg_dummy.msg_control = (void *)msg_control_addr;
2184 cmsg_dummy.msg_controllen =
2185 (__kernel_size_t)zc->msg_controllen;
2186 cmsg_dummy.msg_flags = in_compat_syscall()
2187 ? MSG_CMSG_COMPAT : 0;
2188 cmsg_dummy.msg_control_is_user = true;
2190 if (zc->msg_control == msg_control_addr &&
2191 zc->msg_controllen == cmsg_dummy.msg_controllen) {
2192 tcp_recv_timestamp(&cmsg_dummy, sk, tss);
2193 zc->msg_control = (__u64)
2194 ((uintptr_t)cmsg_dummy.msg_control);
2195 zc->msg_controllen =
2196 (__u64)cmsg_dummy.msg_controllen;
2197 zc->msg_flags = (__u32)cmsg_dummy.msg_flags;
2201 #define TCP_ZEROCOPY_PAGE_BATCH_SIZE 32
2202 static int tcp_zerocopy_receive(struct sock *sk,
2203 struct tcp_zerocopy_receive *zc,
2204 struct scm_timestamping_internal *tss)
2206 u32 length = 0, offset, vma_len, avail_len, copylen = 0;
2207 unsigned long address = (unsigned long)zc->address;
2208 struct page *pages[TCP_ZEROCOPY_PAGE_BATCH_SIZE];
2209 s32 copybuf_len = zc->copybuf_len;
2210 struct tcp_sock *tp = tcp_sk(sk);
2211 const skb_frag_t *frags = NULL;
2212 unsigned int pages_to_map = 0;
2213 struct vm_area_struct *vma;
2214 struct sk_buff *skb = NULL;
2215 u32 seq = tp->copied_seq;
2216 u32 total_bytes_to_map;
2217 int inq = tcp_inq(sk);
2220 zc->copybuf_len = 0;
2223 if (address & (PAGE_SIZE - 1) || address != zc->address)
2226 if (sk->sk_state == TCP_LISTEN)
2229 sock_rps_record_flow(sk);
2231 if (inq && inq <= copybuf_len)
2232 return receive_fallback_to_copy(sk, zc, inq, tss);
2234 if (inq < PAGE_SIZE) {
2236 zc->recv_skip_hint = inq;
2237 if (!inq && sock_flag(sk, SOCK_DONE))
2242 mmap_read_lock(current->mm);
2244 vma = vma_lookup(current->mm, address);
2245 if (!vma || vma->vm_ops != &tcp_vm_ops) {
2246 mmap_read_unlock(current->mm);
2249 vma_len = min_t(unsigned long, zc->length, vma->vm_end - address);
2250 avail_len = min_t(u32, vma_len, inq);
2251 total_bytes_to_map = avail_len & ~(PAGE_SIZE - 1);
2252 if (total_bytes_to_map) {
2253 if (!(zc->flags & TCP_RECEIVE_ZEROCOPY_FLAG_TLB_CLEAN_HINT))
2254 zap_page_range(vma, address, total_bytes_to_map);
2255 zc->length = total_bytes_to_map;
2256 zc->recv_skip_hint = 0;
2258 zc->length = avail_len;
2259 zc->recv_skip_hint = avail_len;
2262 while (length + PAGE_SIZE <= zc->length) {
2263 int mappable_offset;
2266 if (zc->recv_skip_hint < PAGE_SIZE) {
2270 if (zc->recv_skip_hint > 0)
2273 offset = seq - TCP_SKB_CB(skb)->seq;
2275 skb = tcp_recv_skb(sk, seq, &offset);
2278 if (TCP_SKB_CB(skb)->has_rxtstamp) {
2279 tcp_update_recv_tstamps(skb, tss);
2280 zc->msg_flags |= TCP_CMSG_TS;
2282 zc->recv_skip_hint = skb->len - offset;
2283 frags = skb_advance_to_frag(skb, offset, &offset_frag);
2284 if (!frags || offset_frag)
2288 mappable_offset = find_next_mappable_frag(frags,
2289 zc->recv_skip_hint);
2290 if (mappable_offset) {
2291 zc->recv_skip_hint = mappable_offset;
2294 page = skb_frag_page(frags);
2296 pages[pages_to_map++] = page;
2297 length += PAGE_SIZE;
2298 zc->recv_skip_hint -= PAGE_SIZE;
2300 if (pages_to_map == TCP_ZEROCOPY_PAGE_BATCH_SIZE ||
2301 zc->recv_skip_hint < PAGE_SIZE) {
2302 /* Either full batch, or we're about to go to next skb
2303 * (and we cannot unroll failed ops across skbs).
2305 ret = tcp_zerocopy_vm_insert_batch(vma, pages,
2309 total_bytes_to_map);
2316 ret = tcp_zerocopy_vm_insert_batch(vma, pages, pages_to_map,
2317 &address, &length, &seq,
2318 zc, total_bytes_to_map);
2321 mmap_read_unlock(current->mm);
2322 /* Try to copy straggler data. */
2324 copylen = tcp_zc_handle_leftover(zc, sk, skb, &seq, copybuf_len, tss);
2326 if (length + copylen) {
2327 WRITE_ONCE(tp->copied_seq, seq);
2328 tcp_rcv_space_adjust(sk);
2330 /* Clean up data we have read: This will do ACK frames. */
2331 tcp_recv_skb(sk, seq, &offset);
2332 tcp_cleanup_rbuf(sk, length + copylen);
2334 if (length == zc->length)
2335 zc->recv_skip_hint = 0;
2337 if (!zc->recv_skip_hint && sock_flag(sk, SOCK_DONE))
2340 zc->length = length;
2345 /* Similar to __sock_recv_timestamp, but does not require an skb */
2346 void tcp_recv_timestamp(struct msghdr *msg, const struct sock *sk,
2347 struct scm_timestamping_internal *tss)
2349 int new_tstamp = sock_flag(sk, SOCK_TSTAMP_NEW);
2350 bool has_timestamping = false;
2352 if (tss->ts[0].tv_sec || tss->ts[0].tv_nsec) {
2353 if (sock_flag(sk, SOCK_RCVTSTAMP)) {
2354 if (sock_flag(sk, SOCK_RCVTSTAMPNS)) {
2356 struct __kernel_timespec kts = {
2357 .tv_sec = tss->ts[0].tv_sec,
2358 .tv_nsec = tss->ts[0].tv_nsec,
2360 put_cmsg(msg, SOL_SOCKET, SO_TIMESTAMPNS_NEW,
2363 struct __kernel_old_timespec ts_old = {
2364 .tv_sec = tss->ts[0].tv_sec,
2365 .tv_nsec = tss->ts[0].tv_nsec,
2367 put_cmsg(msg, SOL_SOCKET, SO_TIMESTAMPNS_OLD,
2368 sizeof(ts_old), &ts_old);
2372 struct __kernel_sock_timeval stv = {
2373 .tv_sec = tss->ts[0].tv_sec,
2374 .tv_usec = tss->ts[0].tv_nsec / 1000,
2376 put_cmsg(msg, SOL_SOCKET, SO_TIMESTAMP_NEW,
2379 struct __kernel_old_timeval tv = {
2380 .tv_sec = tss->ts[0].tv_sec,
2381 .tv_usec = tss->ts[0].tv_nsec / 1000,
2383 put_cmsg(msg, SOL_SOCKET, SO_TIMESTAMP_OLD,
2389 if (READ_ONCE(sk->sk_tsflags) & SOF_TIMESTAMPING_SOFTWARE)
2390 has_timestamping = true;
2392 tss->ts[0] = (struct timespec64) {0};
2395 if (tss->ts[2].tv_sec || tss->ts[2].tv_nsec) {
2396 if (READ_ONCE(sk->sk_tsflags) & SOF_TIMESTAMPING_RAW_HARDWARE)
2397 has_timestamping = true;
2399 tss->ts[2] = (struct timespec64) {0};
2402 if (has_timestamping) {
2403 tss->ts[1] = (struct timespec64) {0};
2404 if (sock_flag(sk, SOCK_TSTAMP_NEW))
2405 put_cmsg_scm_timestamping64(msg, tss);
2407 put_cmsg_scm_timestamping(msg, tss);
2411 static int tcp_inq_hint(struct sock *sk)
2413 const struct tcp_sock *tp = tcp_sk(sk);
2414 u32 copied_seq = READ_ONCE(tp->copied_seq);
2415 u32 rcv_nxt = READ_ONCE(tp->rcv_nxt);
2418 inq = rcv_nxt - copied_seq;
2419 if (unlikely(inq < 0 || copied_seq != READ_ONCE(tp->copied_seq))) {
2421 inq = tp->rcv_nxt - tp->copied_seq;
2424 /* After receiving a FIN, tell the user-space to continue reading
2425 * by returning a non-zero inq.
2427 if (inq == 0 && sock_flag(sk, SOCK_DONE))
2433 * This routine copies from a sock struct into the user buffer.
2435 * Technical note: in 2.3 we work on _locked_ socket, so that
2436 * tricks with *seq access order and skb->users are not required.
2437 * Probably, code can be easily improved even more.
2440 static int tcp_recvmsg_locked(struct sock *sk, struct msghdr *msg, size_t len,
2441 int flags, struct scm_timestamping_internal *tss,
2444 struct tcp_sock *tp = tcp_sk(sk);
2450 int target; /* Read at least this many bytes */
2452 struct sk_buff *skb, *last;
2456 if (sk->sk_state == TCP_LISTEN)
2459 if (tp->recvmsg_inq) {
2460 *cmsg_flags = TCP_CMSG_INQ;
2461 msg->msg_get_inq = 1;
2463 timeo = sock_rcvtimeo(sk, flags & MSG_DONTWAIT);
2465 /* Urgent data needs to be handled specially. */
2466 if (flags & MSG_OOB)
2469 if (unlikely(tp->repair)) {
2471 if (!(flags & MSG_PEEK))
2474 if (tp->repair_queue == TCP_SEND_QUEUE)
2478 if (tp->repair_queue == TCP_NO_QUEUE)
2481 /* 'common' recv queue MSG_PEEK-ing */
2484 seq = &tp->copied_seq;
2485 if (flags & MSG_PEEK) {
2486 peek_seq = tp->copied_seq;
2490 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
2495 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
2496 if (unlikely(tp->urg_data) && tp->urg_seq == *seq) {
2499 if (signal_pending(current)) {
2500 copied = timeo ? sock_intr_errno(timeo) : -EAGAIN;
2505 /* Next get a buffer. */
2507 last = skb_peek_tail(&sk->sk_receive_queue);
2508 skb_queue_walk(&sk->sk_receive_queue, skb) {
2510 /* Now that we have two receive queues this
2513 if (WARN(before(*seq, TCP_SKB_CB(skb)->seq),
2514 "TCP recvmsg seq # bug: copied %X, seq %X, rcvnxt %X, fl %X\n",
2515 *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt,
2519 offset = *seq - TCP_SKB_CB(skb)->seq;
2520 if (unlikely(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
2521 pr_err_once("%s: found a SYN, please report !\n", __func__);
2524 if (offset < skb->len)
2526 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
2528 WARN(!(flags & MSG_PEEK),
2529 "TCP recvmsg seq # bug 2: copied %X, seq %X, rcvnxt %X, fl %X\n",
2530 *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt, flags);
2533 /* Well, if we have backlog, try to process it now yet. */
2535 if (copied >= target && !READ_ONCE(sk->sk_backlog.tail))
2541 sk->sk_state == TCP_CLOSE ||
2542 (sk->sk_shutdown & RCV_SHUTDOWN) ||
2543 signal_pending(current))
2546 if (sock_flag(sk, SOCK_DONE))
2550 copied = sock_error(sk);
2554 if (sk->sk_shutdown & RCV_SHUTDOWN)
2557 if (sk->sk_state == TCP_CLOSE) {
2558 /* This occurs when user tries to read
2559 * from never connected socket.
2570 if (signal_pending(current)) {
2571 copied = sock_intr_errno(timeo);
2576 if (copied >= target) {
2577 /* Do not sleep, just process backlog. */
2578 __sk_flush_backlog(sk);
2580 tcp_cleanup_rbuf(sk, copied);
2581 err = sk_wait_data(sk, &timeo, last);
2583 err = copied ? : err;
2588 if ((flags & MSG_PEEK) &&
2589 (peek_seq - copied - urg_hole != tp->copied_seq)) {
2590 net_dbg_ratelimited("TCP(%s:%d): Application bug, race in MSG_PEEK\n",
2592 task_pid_nr(current));
2593 peek_seq = tp->copied_seq;
2598 /* Ok so how much can we use? */
2599 used = skb->len - offset;
2603 /* Do we have urgent data here? */
2604 if (unlikely(tp->urg_data)) {
2605 u32 urg_offset = tp->urg_seq - *seq;
2606 if (urg_offset < used) {
2608 if (!sock_flag(sk, SOCK_URGINLINE)) {
2609 WRITE_ONCE(*seq, *seq + 1);
2621 if (!(flags & MSG_TRUNC)) {
2622 err = skb_copy_datagram_msg(skb, offset, msg, used);
2624 /* Exception. Bailout! */
2631 WRITE_ONCE(*seq, *seq + used);
2635 tcp_rcv_space_adjust(sk);
2638 if (unlikely(tp->urg_data) && after(tp->copied_seq, tp->urg_seq)) {
2639 WRITE_ONCE(tp->urg_data, 0);
2640 tcp_fast_path_check(sk);
2643 if (TCP_SKB_CB(skb)->has_rxtstamp) {
2644 tcp_update_recv_tstamps(skb, tss);
2645 *cmsg_flags |= TCP_CMSG_TS;
2648 if (used + offset < skb->len)
2651 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
2653 if (!(flags & MSG_PEEK))
2654 tcp_eat_recv_skb(sk, skb);
2658 /* Process the FIN. */
2659 WRITE_ONCE(*seq, *seq + 1);
2660 if (!(flags & MSG_PEEK))
2661 tcp_eat_recv_skb(sk, skb);
2665 /* According to UNIX98, msg_name/msg_namelen are ignored
2666 * on connected socket. I was just happy when found this 8) --ANK
2669 /* Clean up data we have read: This will do ACK frames. */
2670 tcp_cleanup_rbuf(sk, copied);
2677 err = tcp_recv_urg(sk, msg, len, flags);
2681 err = tcp_peek_sndq(sk, msg, len);
2685 int tcp_recvmsg(struct sock *sk, struct msghdr *msg, size_t len, int flags,
2688 int cmsg_flags = 0, ret;
2689 struct scm_timestamping_internal tss;
2691 if (unlikely(flags & MSG_ERRQUEUE))
2692 return inet_recv_error(sk, msg, len, addr_len);
2694 if (sk_can_busy_loop(sk) &&
2695 skb_queue_empty_lockless(&sk->sk_receive_queue) &&
2696 sk->sk_state == TCP_ESTABLISHED)
2697 sk_busy_loop(sk, flags & MSG_DONTWAIT);
2700 ret = tcp_recvmsg_locked(sk, msg, len, flags, &tss, &cmsg_flags);
2703 if ((cmsg_flags || msg->msg_get_inq) && ret >= 0) {
2704 if (cmsg_flags & TCP_CMSG_TS)
2705 tcp_recv_timestamp(msg, sk, &tss);
2706 if (msg->msg_get_inq) {
2707 msg->msg_inq = tcp_inq_hint(sk);
2708 if (cmsg_flags & TCP_CMSG_INQ)
2709 put_cmsg(msg, SOL_TCP, TCP_CM_INQ,
2710 sizeof(msg->msg_inq), &msg->msg_inq);
2715 EXPORT_SYMBOL(tcp_recvmsg);
2717 void tcp_set_state(struct sock *sk, int state)
2719 int oldstate = sk->sk_state;
2721 /* We defined a new enum for TCP states that are exported in BPF
2722 * so as not force the internal TCP states to be frozen. The
2723 * following checks will detect if an internal state value ever
2724 * differs from the BPF value. If this ever happens, then we will
2725 * need to remap the internal value to the BPF value before calling
2726 * tcp_call_bpf_2arg.
2728 BUILD_BUG_ON((int)BPF_TCP_ESTABLISHED != (int)TCP_ESTABLISHED);
2729 BUILD_BUG_ON((int)BPF_TCP_SYN_SENT != (int)TCP_SYN_SENT);
2730 BUILD_BUG_ON((int)BPF_TCP_SYN_RECV != (int)TCP_SYN_RECV);
2731 BUILD_BUG_ON((int)BPF_TCP_FIN_WAIT1 != (int)TCP_FIN_WAIT1);
2732 BUILD_BUG_ON((int)BPF_TCP_FIN_WAIT2 != (int)TCP_FIN_WAIT2);
2733 BUILD_BUG_ON((int)BPF_TCP_TIME_WAIT != (int)TCP_TIME_WAIT);
2734 BUILD_BUG_ON((int)BPF_TCP_CLOSE != (int)TCP_CLOSE);
2735 BUILD_BUG_ON((int)BPF_TCP_CLOSE_WAIT != (int)TCP_CLOSE_WAIT);
2736 BUILD_BUG_ON((int)BPF_TCP_LAST_ACK != (int)TCP_LAST_ACK);
2737 BUILD_BUG_ON((int)BPF_TCP_LISTEN != (int)TCP_LISTEN);
2738 BUILD_BUG_ON((int)BPF_TCP_CLOSING != (int)TCP_CLOSING);
2739 BUILD_BUG_ON((int)BPF_TCP_NEW_SYN_RECV != (int)TCP_NEW_SYN_RECV);
2740 BUILD_BUG_ON((int)BPF_TCP_MAX_STATES != (int)TCP_MAX_STATES);
2742 /* bpf uapi header bpf.h defines an anonymous enum with values
2743 * BPF_TCP_* used by bpf programs. Currently gcc built vmlinux
2744 * is able to emit this enum in DWARF due to the above BUILD_BUG_ON.
2745 * But clang built vmlinux does not have this enum in DWARF
2746 * since clang removes the above code before generating IR/debuginfo.
2747 * Let us explicitly emit the type debuginfo to ensure the
2748 * above-mentioned anonymous enum in the vmlinux DWARF and hence BTF
2749 * regardless of which compiler is used.
2751 BTF_TYPE_EMIT_ENUM(BPF_TCP_ESTABLISHED);
2753 if (BPF_SOCK_OPS_TEST_FLAG(tcp_sk(sk), BPF_SOCK_OPS_STATE_CB_FLAG))
2754 tcp_call_bpf_2arg(sk, BPF_SOCK_OPS_STATE_CB, oldstate, state);
2757 case TCP_ESTABLISHED:
2758 if (oldstate != TCP_ESTABLISHED)
2759 TCP_INC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
2763 if (oldstate == TCP_CLOSE_WAIT || oldstate == TCP_ESTABLISHED)
2764 TCP_INC_STATS(sock_net(sk), TCP_MIB_ESTABRESETS);
2766 sk->sk_prot->unhash(sk);
2767 if (inet_csk(sk)->icsk_bind_hash &&
2768 !(sk->sk_userlocks & SOCK_BINDPORT_LOCK))
2772 if (oldstate == TCP_ESTABLISHED)
2773 TCP_DEC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
2776 /* Change state AFTER socket is unhashed to avoid closed
2777 * socket sitting in hash tables.
2779 inet_sk_state_store(sk, state);
2781 EXPORT_SYMBOL_GPL(tcp_set_state);
2784 * State processing on a close. This implements the state shift for
2785 * sending our FIN frame. Note that we only send a FIN for some
2786 * states. A shutdown() may have already sent the FIN, or we may be
2790 static const unsigned char new_state[16] = {
2791 /* current state: new state: action: */
2792 [0 /* (Invalid) */] = TCP_CLOSE,
2793 [TCP_ESTABLISHED] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2794 [TCP_SYN_SENT] = TCP_CLOSE,
2795 [TCP_SYN_RECV] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2796 [TCP_FIN_WAIT1] = TCP_FIN_WAIT1,
2797 [TCP_FIN_WAIT2] = TCP_FIN_WAIT2,
2798 [TCP_TIME_WAIT] = TCP_CLOSE,
2799 [TCP_CLOSE] = TCP_CLOSE,
2800 [TCP_CLOSE_WAIT] = TCP_LAST_ACK | TCP_ACTION_FIN,
2801 [TCP_LAST_ACK] = TCP_LAST_ACK,
2802 [TCP_LISTEN] = TCP_CLOSE,
2803 [TCP_CLOSING] = TCP_CLOSING,
2804 [TCP_NEW_SYN_RECV] = TCP_CLOSE, /* should not happen ! */
2807 static int tcp_close_state(struct sock *sk)
2809 int next = (int)new_state[sk->sk_state];
2810 int ns = next & TCP_STATE_MASK;
2812 tcp_set_state(sk, ns);
2814 return next & TCP_ACTION_FIN;
2818 * Shutdown the sending side of a connection. Much like close except
2819 * that we don't receive shut down or sock_set_flag(sk, SOCK_DEAD).
2822 void tcp_shutdown(struct sock *sk, int how)
2824 /* We need to grab some memory, and put together a FIN,
2825 * and then put it into the queue to be sent.
2826 * Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
2828 if (!(how & SEND_SHUTDOWN))
2831 /* If we've already sent a FIN, or it's a closed state, skip this. */
2832 if ((1 << sk->sk_state) &
2833 (TCPF_ESTABLISHED | TCPF_SYN_SENT |
2834 TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) {
2835 /* Clear out any half completed packets. FIN if needed. */
2836 if (tcp_close_state(sk))
2840 EXPORT_SYMBOL(tcp_shutdown);
2842 int tcp_orphan_count_sum(void)
2846 for_each_possible_cpu(i)
2847 total += per_cpu(tcp_orphan_count, i);
2849 return max(total, 0);
2852 static int tcp_orphan_cache;
2853 static struct timer_list tcp_orphan_timer;
2854 #define TCP_ORPHAN_TIMER_PERIOD msecs_to_jiffies(100)
2856 static void tcp_orphan_update(struct timer_list *unused)
2858 WRITE_ONCE(tcp_orphan_cache, tcp_orphan_count_sum());
2859 mod_timer(&tcp_orphan_timer, jiffies + TCP_ORPHAN_TIMER_PERIOD);
2862 static bool tcp_too_many_orphans(int shift)
2864 return READ_ONCE(tcp_orphan_cache) << shift >
2865 READ_ONCE(sysctl_tcp_max_orphans);
2868 bool tcp_check_oom(struct sock *sk, int shift)
2870 bool too_many_orphans, out_of_socket_memory;
2872 too_many_orphans = tcp_too_many_orphans(shift);
2873 out_of_socket_memory = tcp_out_of_memory(sk);
2875 if (too_many_orphans)
2876 net_info_ratelimited("too many orphaned sockets\n");
2877 if (out_of_socket_memory)
2878 net_info_ratelimited("out of memory -- consider tuning tcp_mem\n");
2879 return too_many_orphans || out_of_socket_memory;
2882 void __tcp_close(struct sock *sk, long timeout)
2884 struct sk_buff *skb;
2885 int data_was_unread = 0;
2888 WRITE_ONCE(sk->sk_shutdown, SHUTDOWN_MASK);
2890 if (sk->sk_state == TCP_LISTEN) {
2891 tcp_set_state(sk, TCP_CLOSE);
2894 inet_csk_listen_stop(sk);
2896 goto adjudge_to_death;
2899 /* We need to flush the recv. buffs. We do this only on the
2900 * descriptor close, not protocol-sourced closes, because the
2901 * reader process may not have drained the data yet!
2903 while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) {
2904 u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq;
2906 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
2908 data_was_unread += len;
2912 /* If socket has been already reset (e.g. in tcp_reset()) - kill it. */
2913 if (sk->sk_state == TCP_CLOSE)
2914 goto adjudge_to_death;
2916 /* As outlined in RFC 2525, section 2.17, we send a RST here because
2917 * data was lost. To witness the awful effects of the old behavior of
2918 * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk
2919 * GET in an FTP client, suspend the process, wait for the client to
2920 * advertise a zero window, then kill -9 the FTP client, wheee...
2921 * Note: timeout is always zero in such a case.
2923 if (unlikely(tcp_sk(sk)->repair)) {
2924 sk->sk_prot->disconnect(sk, 0);
2925 } else if (data_was_unread) {
2926 /* Unread data was tossed, zap the connection. */
2927 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONCLOSE);
2928 tcp_set_state(sk, TCP_CLOSE);
2929 tcp_send_active_reset(sk, sk->sk_allocation);
2930 } else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) {
2931 /* Check zero linger _after_ checking for unread data. */
2932 sk->sk_prot->disconnect(sk, 0);
2933 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
2934 } else if (tcp_close_state(sk)) {
2935 /* We FIN if the application ate all the data before
2936 * zapping the connection.
2939 /* RED-PEN. Formally speaking, we have broken TCP state
2940 * machine. State transitions:
2942 * TCP_ESTABLISHED -> TCP_FIN_WAIT1
2943 * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible)
2944 * TCP_CLOSE_WAIT -> TCP_LAST_ACK
2946 * are legal only when FIN has been sent (i.e. in window),
2947 * rather than queued out of window. Purists blame.
2949 * F.e. "RFC state" is ESTABLISHED,
2950 * if Linux state is FIN-WAIT-1, but FIN is still not sent.
2952 * The visible declinations are that sometimes
2953 * we enter time-wait state, when it is not required really
2954 * (harmless), do not send active resets, when they are
2955 * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
2956 * they look as CLOSING or LAST_ACK for Linux)
2957 * Probably, I missed some more holelets.
2959 * XXX (TFO) - To start off we don't support SYN+ACK+FIN
2960 * in a single packet! (May consider it later but will
2961 * probably need API support or TCP_CORK SYN-ACK until
2962 * data is written and socket is closed.)
2967 sk_stream_wait_close(sk, timeout);
2970 state = sk->sk_state;
2976 /* remove backlog if any, without releasing ownership. */
2979 this_cpu_inc(tcp_orphan_count);
2981 /* Have we already been destroyed by a softirq or backlog? */
2982 if (state != TCP_CLOSE && sk->sk_state == TCP_CLOSE)
2985 /* This is a (useful) BSD violating of the RFC. There is a
2986 * problem with TCP as specified in that the other end could
2987 * keep a socket open forever with no application left this end.
2988 * We use a 1 minute timeout (about the same as BSD) then kill
2989 * our end. If they send after that then tough - BUT: long enough
2990 * that we won't make the old 4*rto = almost no time - whoops
2993 * Nope, it was not mistake. It is really desired behaviour
2994 * f.e. on http servers, when such sockets are useless, but
2995 * consume significant resources. Let's do it with special
2996 * linger2 option. --ANK
2999 if (sk->sk_state == TCP_FIN_WAIT2) {
3000 struct tcp_sock *tp = tcp_sk(sk);
3001 if (tp->linger2 < 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_TCPABORTONLINGER);
3007 const int tmo = tcp_fin_time(sk);
3009 if (tmo > TCP_TIMEWAIT_LEN) {
3010 inet_csk_reset_keepalive_timer(sk,
3011 tmo - TCP_TIMEWAIT_LEN);
3013 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
3018 if (sk->sk_state != TCP_CLOSE) {
3019 if (tcp_check_oom(sk, 0)) {
3020 tcp_set_state(sk, TCP_CLOSE);
3021 tcp_send_active_reset(sk, GFP_ATOMIC);
3022 __NET_INC_STATS(sock_net(sk),
3023 LINUX_MIB_TCPABORTONMEMORY);
3024 } else if (!check_net(sock_net(sk))) {
3025 /* Not possible to send reset; just close */
3026 tcp_set_state(sk, TCP_CLOSE);
3030 if (sk->sk_state == TCP_CLOSE) {
3031 struct request_sock *req;
3033 req = rcu_dereference_protected(tcp_sk(sk)->fastopen_rsk,
3034 lockdep_sock_is_held(sk));
3035 /* We could get here with a non-NULL req if the socket is
3036 * aborted (e.g., closed with unread data) before 3WHS
3040 reqsk_fastopen_remove(sk, req, false);
3041 inet_csk_destroy_sock(sk);
3043 /* Otherwise, socket is reprieved until protocol close. */
3050 void tcp_close(struct sock *sk, long timeout)
3053 __tcp_close(sk, timeout);
3055 if (!sk->sk_net_refcnt)
3056 inet_csk_clear_xmit_timers_sync(sk);
3059 EXPORT_SYMBOL(tcp_close);
3061 /* These states need RST on ABORT according to RFC793 */
3063 static inline bool tcp_need_reset(int state)
3065 return (1 << state) &
3066 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 |
3067 TCPF_FIN_WAIT2 | TCPF_SYN_RECV);
3070 static void tcp_rtx_queue_purge(struct sock *sk)
3072 struct rb_node *p = rb_first(&sk->tcp_rtx_queue);
3074 tcp_sk(sk)->highest_sack = NULL;
3076 struct sk_buff *skb = rb_to_skb(p);
3079 /* Since we are deleting whole queue, no need to
3080 * list_del(&skb->tcp_tsorted_anchor)
3082 tcp_rtx_queue_unlink(skb, sk);
3083 tcp_wmem_free_skb(sk, skb);
3087 void tcp_write_queue_purge(struct sock *sk)
3089 struct sk_buff *skb;
3091 tcp_chrono_stop(sk, TCP_CHRONO_BUSY);
3092 while ((skb = __skb_dequeue(&sk->sk_write_queue)) != NULL) {
3093 tcp_skb_tsorted_anchor_cleanup(skb);
3094 tcp_wmem_free_skb(sk, skb);
3096 tcp_rtx_queue_purge(sk);
3097 INIT_LIST_HEAD(&tcp_sk(sk)->tsorted_sent_queue);
3098 tcp_clear_all_retrans_hints(tcp_sk(sk));
3099 tcp_sk(sk)->packets_out = 0;
3100 inet_csk(sk)->icsk_backoff = 0;
3103 int tcp_disconnect(struct sock *sk, int flags)
3105 struct inet_sock *inet = inet_sk(sk);
3106 struct inet_connection_sock *icsk = inet_csk(sk);
3107 struct tcp_sock *tp = tcp_sk(sk);
3108 int old_state = sk->sk_state;
3111 if (old_state != TCP_CLOSE)
3112 tcp_set_state(sk, TCP_CLOSE);
3114 /* ABORT function of RFC793 */
3115 if (old_state == TCP_LISTEN) {
3116 inet_csk_listen_stop(sk);
3117 } else if (unlikely(tp->repair)) {
3118 sk->sk_err = ECONNABORTED;
3119 } else if (tcp_need_reset(old_state) ||
3120 (tp->snd_nxt != tp->write_seq &&
3121 (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK))) {
3122 /* The last check adjusts for discrepancy of Linux wrt. RFC
3125 tcp_send_active_reset(sk, gfp_any());
3126 sk->sk_err = ECONNRESET;
3127 } else if (old_state == TCP_SYN_SENT)
3128 sk->sk_err = ECONNRESET;
3130 tcp_clear_xmit_timers(sk);
3131 __skb_queue_purge(&sk->sk_receive_queue);
3132 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt);
3133 WRITE_ONCE(tp->urg_data, 0);
3134 tcp_write_queue_purge(sk);
3135 tcp_fastopen_active_disable_ofo_check(sk);
3136 skb_rbtree_purge(&tp->out_of_order_queue);
3138 inet->inet_dport = 0;
3140 inet_bhash2_reset_saddr(sk);
3142 WRITE_ONCE(sk->sk_shutdown, 0);
3143 sock_reset_flag(sk, SOCK_DONE);
3145 tp->mdev_us = jiffies_to_usecs(TCP_TIMEOUT_INIT);
3146 tp->rcv_rtt_last_tsecr = 0;
3148 seq = tp->write_seq + tp->max_window + 2;
3151 WRITE_ONCE(tp->write_seq, seq);
3153 icsk->icsk_backoff = 0;
3154 icsk->icsk_probes_out = 0;
3155 icsk->icsk_probes_tstamp = 0;
3156 icsk->icsk_rto = TCP_TIMEOUT_INIT;
3157 icsk->icsk_rto_min = TCP_RTO_MIN;
3158 icsk->icsk_delack_max = TCP_DELACK_MAX;
3159 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
3160 tcp_snd_cwnd_set(tp, TCP_INIT_CWND);
3161 tp->snd_cwnd_cnt = 0;
3162 tp->is_cwnd_limited = 0;
3163 tp->max_packets_out = 0;
3164 tp->window_clamp = 0;
3166 tp->delivered_ce = 0;
3167 if (icsk->icsk_ca_ops->release)
3168 icsk->icsk_ca_ops->release(sk);
3169 memset(icsk->icsk_ca_priv, 0, sizeof(icsk->icsk_ca_priv));
3170 icsk->icsk_ca_initialized = 0;
3171 tcp_set_ca_state(sk, TCP_CA_Open);
3172 tp->is_sack_reneg = 0;
3173 tcp_clear_retrans(tp);
3174 tp->total_retrans = 0;
3175 inet_csk_delack_init(sk);
3176 /* Initialize rcv_mss to TCP_MIN_MSS to avoid division by 0
3177 * issue in __tcp_select_window()
3179 icsk->icsk_ack.rcv_mss = TCP_MIN_MSS;
3180 memset(&tp->rx_opt, 0, sizeof(tp->rx_opt));
3182 dst_release(xchg((__force struct dst_entry **)&sk->sk_rx_dst, NULL));
3183 tcp_saved_syn_free(tp);
3184 tp->compressed_ack = 0;
3188 tp->bytes_acked = 0;
3189 tp->bytes_received = 0;
3190 tp->bytes_retrans = 0;
3191 tp->data_segs_in = 0;
3192 tp->data_segs_out = 0;
3193 tp->duplicate_sack[0].start_seq = 0;
3194 tp->duplicate_sack[0].end_seq = 0;
3197 tp->retrans_out = 0;
3199 tp->tlp_high_seq = 0;
3200 tp->last_oow_ack_time = 0;
3201 /* There's a bubble in the pipe until at least the first ACK. */
3202 tp->app_limited = ~0U;
3203 tp->rate_app_limited = 1;
3204 tp->rack.mstamp = 0;
3205 tp->rack.advanced = 0;
3206 tp->rack.reo_wnd_steps = 1;
3207 tp->rack.last_delivered = 0;
3208 tp->rack.reo_wnd_persist = 0;
3209 tp->rack.dsack_seen = 0;
3210 tp->syn_data_acked = 0;
3211 tp->rx_opt.saw_tstamp = 0;
3212 tp->rx_opt.dsack = 0;
3213 tp->rx_opt.num_sacks = 0;
3214 tp->rcv_ooopack = 0;
3217 /* Clean up fastopen related fields */
3218 tcp_free_fastopen_req(tp);
3219 inet->defer_connect = 0;
3220 tp->fastopen_client_fail = 0;
3222 WARN_ON(inet->inet_num && !icsk->icsk_bind_hash);
3224 if (sk->sk_frag.page) {
3225 put_page(sk->sk_frag.page);
3226 sk->sk_frag.page = NULL;
3227 sk->sk_frag.offset = 0;
3229 sk_error_report(sk);
3232 EXPORT_SYMBOL(tcp_disconnect);
3234 static inline bool tcp_can_repair_sock(const struct sock *sk)
3236 return sockopt_ns_capable(sock_net(sk)->user_ns, CAP_NET_ADMIN) &&
3237 (sk->sk_state != TCP_LISTEN);
3240 static int tcp_repair_set_window(struct tcp_sock *tp, sockptr_t optbuf, int len)
3242 struct tcp_repair_window opt;
3247 if (len != sizeof(opt))
3250 if (copy_from_sockptr(&opt, optbuf, sizeof(opt)))
3253 if (opt.max_window < opt.snd_wnd)
3256 if (after(opt.snd_wl1, tp->rcv_nxt + opt.rcv_wnd))
3259 if (after(opt.rcv_wup, tp->rcv_nxt))
3262 tp->snd_wl1 = opt.snd_wl1;
3263 tp->snd_wnd = opt.snd_wnd;
3264 tp->max_window = opt.max_window;
3266 tp->rcv_wnd = opt.rcv_wnd;
3267 tp->rcv_wup = opt.rcv_wup;
3272 static int tcp_repair_options_est(struct sock *sk, sockptr_t optbuf,
3275 struct tcp_sock *tp = tcp_sk(sk);
3276 struct tcp_repair_opt opt;
3279 while (len >= sizeof(opt)) {
3280 if (copy_from_sockptr_offset(&opt, optbuf, offset, sizeof(opt)))
3283 offset += sizeof(opt);
3286 switch (opt.opt_code) {
3288 tp->rx_opt.mss_clamp = opt.opt_val;
3293 u16 snd_wscale = opt.opt_val & 0xFFFF;
3294 u16 rcv_wscale = opt.opt_val >> 16;
3296 if (snd_wscale > TCP_MAX_WSCALE || rcv_wscale > TCP_MAX_WSCALE)
3299 tp->rx_opt.snd_wscale = snd_wscale;
3300 tp->rx_opt.rcv_wscale = rcv_wscale;
3301 tp->rx_opt.wscale_ok = 1;
3304 case TCPOPT_SACK_PERM:
3305 if (opt.opt_val != 0)
3308 tp->rx_opt.sack_ok |= TCP_SACK_SEEN;
3310 case TCPOPT_TIMESTAMP:
3311 if (opt.opt_val != 0)
3314 tp->rx_opt.tstamp_ok = 1;
3322 DEFINE_STATIC_KEY_FALSE(tcp_tx_delay_enabled);
3323 EXPORT_SYMBOL(tcp_tx_delay_enabled);
3325 static void tcp_enable_tx_delay(void)
3327 if (!static_branch_unlikely(&tcp_tx_delay_enabled)) {
3328 static int __tcp_tx_delay_enabled = 0;
3330 if (cmpxchg(&__tcp_tx_delay_enabled, 0, 1) == 0) {
3331 static_branch_enable(&tcp_tx_delay_enabled);
3332 pr_info("TCP_TX_DELAY enabled\n");
3337 /* When set indicates to always queue non-full frames. Later the user clears
3338 * this option and we transmit any pending partial frames in the queue. This is
3339 * meant to be used alongside sendfile() to get properly filled frames when the
3340 * user (for example) must write out headers with a write() call first and then
3341 * use sendfile to send out the data parts.
3343 * TCP_CORK can be set together with TCP_NODELAY and it is stronger than
3346 void __tcp_sock_set_cork(struct sock *sk, bool on)
3348 struct tcp_sock *tp = tcp_sk(sk);
3351 tp->nonagle |= TCP_NAGLE_CORK;
3353 tp->nonagle &= ~TCP_NAGLE_CORK;
3354 if (tp->nonagle & TCP_NAGLE_OFF)
3355 tp->nonagle |= TCP_NAGLE_PUSH;
3356 tcp_push_pending_frames(sk);
3360 void tcp_sock_set_cork(struct sock *sk, bool on)
3363 __tcp_sock_set_cork(sk, on);
3366 EXPORT_SYMBOL(tcp_sock_set_cork);
3368 /* TCP_NODELAY is weaker than TCP_CORK, so that this option on corked socket is
3369 * remembered, but it is not activated until cork is cleared.
3371 * However, when TCP_NODELAY is set we make an explicit push, which overrides
3372 * even TCP_CORK for currently queued segments.
3374 void __tcp_sock_set_nodelay(struct sock *sk, bool on)
3377 tcp_sk(sk)->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH;
3378 tcp_push_pending_frames(sk);
3380 tcp_sk(sk)->nonagle &= ~TCP_NAGLE_OFF;
3384 void tcp_sock_set_nodelay(struct sock *sk)
3387 __tcp_sock_set_nodelay(sk, true);
3390 EXPORT_SYMBOL(tcp_sock_set_nodelay);
3392 static void __tcp_sock_set_quickack(struct sock *sk, int val)
3395 inet_csk_enter_pingpong_mode(sk);
3399 inet_csk_exit_pingpong_mode(sk);
3400 if ((1 << sk->sk_state) & (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) &&
3401 inet_csk_ack_scheduled(sk)) {
3402 inet_csk(sk)->icsk_ack.pending |= ICSK_ACK_PUSHED;
3403 tcp_cleanup_rbuf(sk, 1);
3405 inet_csk_enter_pingpong_mode(sk);
3409 void tcp_sock_set_quickack(struct sock *sk, int val)
3412 __tcp_sock_set_quickack(sk, val);
3415 EXPORT_SYMBOL(tcp_sock_set_quickack);
3417 int tcp_sock_set_syncnt(struct sock *sk, int val)
3419 if (val < 1 || val > MAX_TCP_SYNCNT)
3423 WRITE_ONCE(inet_csk(sk)->icsk_syn_retries, val);
3427 EXPORT_SYMBOL(tcp_sock_set_syncnt);
3429 void tcp_sock_set_user_timeout(struct sock *sk, u32 val)
3432 WRITE_ONCE(inet_csk(sk)->icsk_user_timeout, val);
3435 EXPORT_SYMBOL(tcp_sock_set_user_timeout);
3437 int tcp_sock_set_keepidle_locked(struct sock *sk, int val)
3439 struct tcp_sock *tp = tcp_sk(sk);
3441 if (val < 1 || val > MAX_TCP_KEEPIDLE)
3444 /* Paired with WRITE_ONCE() in keepalive_time_when() */
3445 WRITE_ONCE(tp->keepalive_time, val * HZ);
3446 if (sock_flag(sk, SOCK_KEEPOPEN) &&
3447 !((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN))) {
3448 u32 elapsed = keepalive_time_elapsed(tp);
3450 if (tp->keepalive_time > elapsed)
3451 elapsed = tp->keepalive_time - elapsed;
3454 inet_csk_reset_keepalive_timer(sk, elapsed);
3460 int tcp_sock_set_keepidle(struct sock *sk, int val)
3465 err = tcp_sock_set_keepidle_locked(sk, val);
3469 EXPORT_SYMBOL(tcp_sock_set_keepidle);
3471 int tcp_sock_set_keepintvl(struct sock *sk, int val)
3473 if (val < 1 || val > MAX_TCP_KEEPINTVL)
3477 WRITE_ONCE(tcp_sk(sk)->keepalive_intvl, val * HZ);
3481 EXPORT_SYMBOL(tcp_sock_set_keepintvl);
3483 int tcp_sock_set_keepcnt(struct sock *sk, int val)
3485 if (val < 1 || val > MAX_TCP_KEEPCNT)
3489 /* Paired with READ_ONCE() in keepalive_probes() */
3490 WRITE_ONCE(tcp_sk(sk)->keepalive_probes, val);
3494 EXPORT_SYMBOL(tcp_sock_set_keepcnt);
3496 int tcp_set_window_clamp(struct sock *sk, int val)
3498 struct tcp_sock *tp = tcp_sk(sk);
3501 if (sk->sk_state != TCP_CLOSE)
3503 tp->window_clamp = 0;
3505 u32 new_rcv_ssthresh, old_window_clamp = tp->window_clamp;
3506 u32 new_window_clamp = val < SOCK_MIN_RCVBUF / 2 ?
3507 SOCK_MIN_RCVBUF / 2 : val;
3509 if (new_window_clamp == old_window_clamp)
3512 tp->window_clamp = new_window_clamp;
3513 if (new_window_clamp < old_window_clamp) {
3514 /* need to apply the reserved mem provisioning only
3515 * when shrinking the window clamp
3517 __tcp_adjust_rcv_ssthresh(sk, tp->window_clamp);
3520 new_rcv_ssthresh = min(tp->rcv_wnd, tp->window_clamp);
3521 tp->rcv_ssthresh = max(new_rcv_ssthresh,
3529 * Socket option code for TCP.
3531 int do_tcp_setsockopt(struct sock *sk, int level, int optname,
3532 sockptr_t optval, unsigned int optlen)
3534 struct tcp_sock *tp = tcp_sk(sk);
3535 struct inet_connection_sock *icsk = inet_csk(sk);
3536 struct net *net = sock_net(sk);
3540 /* These are data/string values, all the others are ints */
3542 case TCP_CONGESTION: {
3543 char name[TCP_CA_NAME_MAX];
3548 val = strncpy_from_sockptr(name, optval,
3549 min_t(long, TCP_CA_NAME_MAX-1, optlen));
3554 sockopt_lock_sock(sk);
3555 err = tcp_set_congestion_control(sk, name, !has_current_bpf_ctx(),
3556 sockopt_ns_capable(sock_net(sk)->user_ns,
3558 sockopt_release_sock(sk);
3562 char name[TCP_ULP_NAME_MAX];
3567 val = strncpy_from_sockptr(name, optval,
3568 min_t(long, TCP_ULP_NAME_MAX - 1,
3574 sockopt_lock_sock(sk);
3575 err = tcp_set_ulp(sk, name);
3576 sockopt_release_sock(sk);
3579 case TCP_FASTOPEN_KEY: {
3580 __u8 key[TCP_FASTOPEN_KEY_BUF_LENGTH];
3581 __u8 *backup_key = NULL;
3583 /* Allow a backup key as well to facilitate key rotation
3584 * First key is the active one.
3586 if (optlen != TCP_FASTOPEN_KEY_LENGTH &&
3587 optlen != TCP_FASTOPEN_KEY_BUF_LENGTH)
3590 if (copy_from_sockptr(key, optval, optlen))
3593 if (optlen == TCP_FASTOPEN_KEY_BUF_LENGTH)
3594 backup_key = key + TCP_FASTOPEN_KEY_LENGTH;
3596 return tcp_fastopen_reset_cipher(net, sk, key, backup_key);
3603 if (optlen < sizeof(int))
3606 if (copy_from_sockptr(&val, optval, sizeof(val)))
3609 sockopt_lock_sock(sk);
3613 /* Values greater than interface MTU won't take effect. However
3614 * at the point when this call is done we typically don't yet
3615 * know which interface is going to be used
3617 if (val && (val < TCP_MIN_MSS || val > MAX_TCP_WINDOW)) {
3621 tp->rx_opt.user_mss = val;
3625 __tcp_sock_set_nodelay(sk, val);
3628 case TCP_THIN_LINEAR_TIMEOUTS:
3629 if (val < 0 || val > 1)
3635 case TCP_THIN_DUPACK:
3636 if (val < 0 || val > 1)
3641 if (!tcp_can_repair_sock(sk))
3643 else if (val == TCP_REPAIR_ON) {
3645 sk->sk_reuse = SK_FORCE_REUSE;
3646 tp->repair_queue = TCP_NO_QUEUE;
3647 } else if (val == TCP_REPAIR_OFF) {
3649 sk->sk_reuse = SK_NO_REUSE;
3650 tcp_send_window_probe(sk);
3651 } else if (val == TCP_REPAIR_OFF_NO_WP) {
3653 sk->sk_reuse = SK_NO_REUSE;
3659 case TCP_REPAIR_QUEUE:
3662 else if ((unsigned int)val < TCP_QUEUES_NR)
3663 tp->repair_queue = val;
3669 if (sk->sk_state != TCP_CLOSE) {
3671 } else if (tp->repair_queue == TCP_SEND_QUEUE) {
3672 if (!tcp_rtx_queue_empty(sk))
3675 WRITE_ONCE(tp->write_seq, val);
3676 } else if (tp->repair_queue == TCP_RECV_QUEUE) {
3677 if (tp->rcv_nxt != tp->copied_seq) {
3680 WRITE_ONCE(tp->rcv_nxt, val);
3681 WRITE_ONCE(tp->copied_seq, val);
3688 case TCP_REPAIR_OPTIONS:
3691 else if (sk->sk_state == TCP_ESTABLISHED && !tp->bytes_sent)
3692 err = tcp_repair_options_est(sk, optval, optlen);
3698 __tcp_sock_set_cork(sk, val);
3702 err = tcp_sock_set_keepidle_locked(sk, val);
3705 if (val < 1 || val > MAX_TCP_KEEPINTVL)
3708 WRITE_ONCE(tp->keepalive_intvl, val * HZ);
3711 if (val < 1 || val > MAX_TCP_KEEPCNT)
3714 WRITE_ONCE(tp->keepalive_probes, val);
3717 if (val < 1 || val > MAX_TCP_SYNCNT)
3720 WRITE_ONCE(icsk->icsk_syn_retries, val);
3724 /* 0: disable, 1: enable, 2: start from ether_header */
3725 if (val < 0 || val > 2)
3733 WRITE_ONCE(tp->linger2, -1);
3734 else if (val > TCP_FIN_TIMEOUT_MAX / HZ)
3735 WRITE_ONCE(tp->linger2, TCP_FIN_TIMEOUT_MAX);
3737 WRITE_ONCE(tp->linger2, val * HZ);
3740 case TCP_DEFER_ACCEPT:
3741 /* Translate value in seconds to number of retransmits */
3742 WRITE_ONCE(icsk->icsk_accept_queue.rskq_defer_accept,
3743 secs_to_retrans(val, TCP_TIMEOUT_INIT / HZ,
3747 case TCP_WINDOW_CLAMP:
3748 err = tcp_set_window_clamp(sk, val);
3752 __tcp_sock_set_quickack(sk, val);
3755 #ifdef CONFIG_TCP_MD5SIG
3757 case TCP_MD5SIG_EXT:
3758 err = tp->af_specific->md5_parse(sk, optname, optval, optlen);
3761 case TCP_USER_TIMEOUT:
3762 /* Cap the max time in ms TCP will retry or probe the window
3763 * before giving up and aborting (ETIMEDOUT) a connection.
3768 WRITE_ONCE(icsk->icsk_user_timeout, val);
3772 if (val >= 0 && ((1 << sk->sk_state) & (TCPF_CLOSE |
3774 tcp_fastopen_init_key_once(net);
3776 fastopen_queue_tune(sk, val);
3781 case TCP_FASTOPEN_CONNECT:
3782 if (val > 1 || val < 0) {
3784 } else if (READ_ONCE(net->ipv4.sysctl_tcp_fastopen) &
3785 TFO_CLIENT_ENABLE) {
3786 if (sk->sk_state == TCP_CLOSE)
3787 tp->fastopen_connect = val;
3794 case TCP_FASTOPEN_NO_COOKIE:
3795 if (val > 1 || val < 0)
3797 else if (!((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
3800 tp->fastopen_no_cookie = val;
3806 WRITE_ONCE(tp->tsoffset, val - tcp_time_stamp_raw());
3808 case TCP_REPAIR_WINDOW:
3809 err = tcp_repair_set_window(tp, optval, optlen);
3811 case TCP_NOTSENT_LOWAT:
3812 WRITE_ONCE(tp->notsent_lowat, val);
3813 sk->sk_write_space(sk);
3816 if (val > 1 || val < 0)
3819 tp->recvmsg_inq = val;
3823 tcp_enable_tx_delay();
3824 WRITE_ONCE(tp->tcp_tx_delay, val);
3831 sockopt_release_sock(sk);
3835 int tcp_setsockopt(struct sock *sk, int level, int optname, sockptr_t optval,
3836 unsigned int optlen)
3838 const struct inet_connection_sock *icsk = inet_csk(sk);
3840 if (level != SOL_TCP)
3841 /* Paired with WRITE_ONCE() in do_ipv6_setsockopt() and tcp_v6_connect() */
3842 return READ_ONCE(icsk->icsk_af_ops)->setsockopt(sk, level, optname,
3844 return do_tcp_setsockopt(sk, level, optname, optval, optlen);
3846 EXPORT_SYMBOL(tcp_setsockopt);
3848 static void tcp_get_info_chrono_stats(const struct tcp_sock *tp,
3849 struct tcp_info *info)
3851 u64 stats[__TCP_CHRONO_MAX], total = 0;
3854 for (i = TCP_CHRONO_BUSY; i < __TCP_CHRONO_MAX; ++i) {
3855 stats[i] = tp->chrono_stat[i - 1];
3856 if (i == tp->chrono_type)
3857 stats[i] += tcp_jiffies32 - tp->chrono_start;
3858 stats[i] *= USEC_PER_SEC / HZ;
3862 info->tcpi_busy_time = total;
3863 info->tcpi_rwnd_limited = stats[TCP_CHRONO_RWND_LIMITED];
3864 info->tcpi_sndbuf_limited = stats[TCP_CHRONO_SNDBUF_LIMITED];
3867 /* Return information about state of tcp endpoint in API format. */
3868 void tcp_get_info(struct sock *sk, struct tcp_info *info)
3870 const struct tcp_sock *tp = tcp_sk(sk); /* iff sk_type == SOCK_STREAM */
3871 const struct inet_connection_sock *icsk = inet_csk(sk);
3877 memset(info, 0, sizeof(*info));
3878 if (sk->sk_type != SOCK_STREAM)
3881 info->tcpi_state = inet_sk_state_load(sk);
3883 /* Report meaningful fields for all TCP states, including listeners */
3884 rate = READ_ONCE(sk->sk_pacing_rate);
3885 rate64 = (rate != ~0UL) ? rate : ~0ULL;
3886 info->tcpi_pacing_rate = rate64;
3888 rate = READ_ONCE(sk->sk_max_pacing_rate);
3889 rate64 = (rate != ~0UL) ? rate : ~0ULL;
3890 info->tcpi_max_pacing_rate = rate64;
3892 info->tcpi_reordering = tp->reordering;
3893 info->tcpi_snd_cwnd = tcp_snd_cwnd(tp);
3895 if (info->tcpi_state == TCP_LISTEN) {
3896 /* listeners aliased fields :
3897 * tcpi_unacked -> Number of children ready for accept()
3898 * tcpi_sacked -> max backlog
3900 info->tcpi_unacked = READ_ONCE(sk->sk_ack_backlog);
3901 info->tcpi_sacked = READ_ONCE(sk->sk_max_ack_backlog);
3905 slow = lock_sock_fast(sk);
3907 info->tcpi_ca_state = icsk->icsk_ca_state;
3908 info->tcpi_retransmits = icsk->icsk_retransmits;
3909 info->tcpi_probes = icsk->icsk_probes_out;
3910 info->tcpi_backoff = icsk->icsk_backoff;
3912 if (tp->rx_opt.tstamp_ok)
3913 info->tcpi_options |= TCPI_OPT_TIMESTAMPS;
3914 if (tcp_is_sack(tp))
3915 info->tcpi_options |= TCPI_OPT_SACK;
3916 if (tp->rx_opt.wscale_ok) {
3917 info->tcpi_options |= TCPI_OPT_WSCALE;
3918 info->tcpi_snd_wscale = tp->rx_opt.snd_wscale;
3919 info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale;
3922 if (tp->ecn_flags & TCP_ECN_OK)
3923 info->tcpi_options |= TCPI_OPT_ECN;
3924 if (tp->ecn_flags & TCP_ECN_SEEN)
3925 info->tcpi_options |= TCPI_OPT_ECN_SEEN;
3926 if (tp->syn_data_acked)
3927 info->tcpi_options |= TCPI_OPT_SYN_DATA;
3929 info->tcpi_rto = jiffies_to_usecs(icsk->icsk_rto);
3930 info->tcpi_ato = jiffies_to_usecs(icsk->icsk_ack.ato);
3931 info->tcpi_snd_mss = tp->mss_cache;
3932 info->tcpi_rcv_mss = icsk->icsk_ack.rcv_mss;
3934 info->tcpi_unacked = tp->packets_out;
3935 info->tcpi_sacked = tp->sacked_out;
3937 info->tcpi_lost = tp->lost_out;
3938 info->tcpi_retrans = tp->retrans_out;
3940 now = tcp_jiffies32;
3941 info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime);
3942 info->tcpi_last_data_recv = jiffies_to_msecs(now - icsk->icsk_ack.lrcvtime);
3943 info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp);
3945 info->tcpi_pmtu = icsk->icsk_pmtu_cookie;
3946 info->tcpi_rcv_ssthresh = tp->rcv_ssthresh;
3947 info->tcpi_rtt = tp->srtt_us >> 3;
3948 info->tcpi_rttvar = tp->mdev_us >> 2;
3949 info->tcpi_snd_ssthresh = tp->snd_ssthresh;
3950 info->tcpi_advmss = tp->advmss;
3952 info->tcpi_rcv_rtt = tp->rcv_rtt_est.rtt_us >> 3;
3953 info->tcpi_rcv_space = tp->rcvq_space.space;
3955 info->tcpi_total_retrans = tp->total_retrans;
3957 info->tcpi_bytes_acked = tp->bytes_acked;
3958 info->tcpi_bytes_received = tp->bytes_received;
3959 info->tcpi_notsent_bytes = max_t(int, 0, tp->write_seq - tp->snd_nxt);
3960 tcp_get_info_chrono_stats(tp, info);
3962 info->tcpi_segs_out = tp->segs_out;
3964 /* segs_in and data_segs_in can be updated from tcp_segs_in() from BH */
3965 info->tcpi_segs_in = READ_ONCE(tp->segs_in);
3966 info->tcpi_data_segs_in = READ_ONCE(tp->data_segs_in);
3968 info->tcpi_min_rtt = tcp_min_rtt(tp);
3969 info->tcpi_data_segs_out = tp->data_segs_out;
3971 info->tcpi_delivery_rate_app_limited = tp->rate_app_limited ? 1 : 0;
3972 rate64 = tcp_compute_delivery_rate(tp);
3974 info->tcpi_delivery_rate = rate64;
3975 info->tcpi_delivered = tp->delivered;
3976 info->tcpi_delivered_ce = tp->delivered_ce;
3977 info->tcpi_bytes_sent = tp->bytes_sent;
3978 info->tcpi_bytes_retrans = tp->bytes_retrans;
3979 info->tcpi_dsack_dups = tp->dsack_dups;
3980 info->tcpi_reord_seen = tp->reord_seen;
3981 info->tcpi_rcv_ooopack = tp->rcv_ooopack;
3982 info->tcpi_snd_wnd = tp->snd_wnd;
3983 info->tcpi_fastopen_client_fail = tp->fastopen_client_fail;
3984 unlock_sock_fast(sk, slow);
3986 EXPORT_SYMBOL_GPL(tcp_get_info);
3988 static size_t tcp_opt_stats_get_size(void)
3991 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_BUSY */
3992 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_RWND_LIMITED */
3993 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_SNDBUF_LIMITED */
3994 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_DATA_SEGS_OUT */
3995 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_TOTAL_RETRANS */
3996 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_PACING_RATE */
3997 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_DELIVERY_RATE */
3998 nla_total_size(sizeof(u32)) + /* TCP_NLA_SND_CWND */
3999 nla_total_size(sizeof(u32)) + /* TCP_NLA_REORDERING */
4000 nla_total_size(sizeof(u32)) + /* TCP_NLA_MIN_RTT */
4001 nla_total_size(sizeof(u8)) + /* TCP_NLA_RECUR_RETRANS */
4002 nla_total_size(sizeof(u8)) + /* TCP_NLA_DELIVERY_RATE_APP_LMT */
4003 nla_total_size(sizeof(u32)) + /* TCP_NLA_SNDQ_SIZE */
4004 nla_total_size(sizeof(u8)) + /* TCP_NLA_CA_STATE */
4005 nla_total_size(sizeof(u32)) + /* TCP_NLA_SND_SSTHRESH */
4006 nla_total_size(sizeof(u32)) + /* TCP_NLA_DELIVERED */
4007 nla_total_size(sizeof(u32)) + /* TCP_NLA_DELIVERED_CE */
4008 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_BYTES_SENT */
4009 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_BYTES_RETRANS */
4010 nla_total_size(sizeof(u32)) + /* TCP_NLA_DSACK_DUPS */
4011 nla_total_size(sizeof(u32)) + /* TCP_NLA_REORD_SEEN */
4012 nla_total_size(sizeof(u32)) + /* TCP_NLA_SRTT */
4013 nla_total_size(sizeof(u16)) + /* TCP_NLA_TIMEOUT_REHASH */
4014 nla_total_size(sizeof(u32)) + /* TCP_NLA_BYTES_NOTSENT */
4015 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_EDT */
4016 nla_total_size(sizeof(u8)) + /* TCP_NLA_TTL */
4020 /* Returns TTL or hop limit of an incoming packet from skb. */
4021 static u8 tcp_skb_ttl_or_hop_limit(const struct sk_buff *skb)
4023 if (skb->protocol == htons(ETH_P_IP))
4024 return ip_hdr(skb)->ttl;
4025 else if (skb->protocol == htons(ETH_P_IPV6))
4026 return ipv6_hdr(skb)->hop_limit;
4031 struct sk_buff *tcp_get_timestamping_opt_stats(const struct sock *sk,
4032 const struct sk_buff *orig_skb,
4033 const struct sk_buff *ack_skb)
4035 const struct tcp_sock *tp = tcp_sk(sk);
4036 struct sk_buff *stats;
4037 struct tcp_info info;
4041 stats = alloc_skb(tcp_opt_stats_get_size(), GFP_ATOMIC);
4045 tcp_get_info_chrono_stats(tp, &info);
4046 nla_put_u64_64bit(stats, TCP_NLA_BUSY,
4047 info.tcpi_busy_time, TCP_NLA_PAD);
4048 nla_put_u64_64bit(stats, TCP_NLA_RWND_LIMITED,
4049 info.tcpi_rwnd_limited, TCP_NLA_PAD);
4050 nla_put_u64_64bit(stats, TCP_NLA_SNDBUF_LIMITED,
4051 info.tcpi_sndbuf_limited, TCP_NLA_PAD);
4052 nla_put_u64_64bit(stats, TCP_NLA_DATA_SEGS_OUT,
4053 tp->data_segs_out, TCP_NLA_PAD);
4054 nla_put_u64_64bit(stats, TCP_NLA_TOTAL_RETRANS,
4055 tp->total_retrans, TCP_NLA_PAD);
4057 rate = READ_ONCE(sk->sk_pacing_rate);
4058 rate64 = (rate != ~0UL) ? rate : ~0ULL;
4059 nla_put_u64_64bit(stats, TCP_NLA_PACING_RATE, rate64, TCP_NLA_PAD);
4061 rate64 = tcp_compute_delivery_rate(tp);
4062 nla_put_u64_64bit(stats, TCP_NLA_DELIVERY_RATE, rate64, TCP_NLA_PAD);
4064 nla_put_u32(stats, TCP_NLA_SND_CWND, tcp_snd_cwnd(tp));
4065 nla_put_u32(stats, TCP_NLA_REORDERING, tp->reordering);
4066 nla_put_u32(stats, TCP_NLA_MIN_RTT, tcp_min_rtt(tp));
4068 nla_put_u8(stats, TCP_NLA_RECUR_RETRANS, inet_csk(sk)->icsk_retransmits);
4069 nla_put_u8(stats, TCP_NLA_DELIVERY_RATE_APP_LMT, !!tp->rate_app_limited);
4070 nla_put_u32(stats, TCP_NLA_SND_SSTHRESH, tp->snd_ssthresh);
4071 nla_put_u32(stats, TCP_NLA_DELIVERED, tp->delivered);
4072 nla_put_u32(stats, TCP_NLA_DELIVERED_CE, tp->delivered_ce);
4074 nla_put_u32(stats, TCP_NLA_SNDQ_SIZE, tp->write_seq - tp->snd_una);
4075 nla_put_u8(stats, TCP_NLA_CA_STATE, inet_csk(sk)->icsk_ca_state);
4077 nla_put_u64_64bit(stats, TCP_NLA_BYTES_SENT, tp->bytes_sent,
4079 nla_put_u64_64bit(stats, TCP_NLA_BYTES_RETRANS, tp->bytes_retrans,
4081 nla_put_u32(stats, TCP_NLA_DSACK_DUPS, tp->dsack_dups);
4082 nla_put_u32(stats, TCP_NLA_REORD_SEEN, tp->reord_seen);
4083 nla_put_u32(stats, TCP_NLA_SRTT, tp->srtt_us >> 3);
4084 nla_put_u16(stats, TCP_NLA_TIMEOUT_REHASH, tp->timeout_rehash);
4085 nla_put_u32(stats, TCP_NLA_BYTES_NOTSENT,
4086 max_t(int, 0, tp->write_seq - tp->snd_nxt));
4087 nla_put_u64_64bit(stats, TCP_NLA_EDT, orig_skb->skb_mstamp_ns,
4090 nla_put_u8(stats, TCP_NLA_TTL,
4091 tcp_skb_ttl_or_hop_limit(ack_skb));
4096 int do_tcp_getsockopt(struct sock *sk, int level,
4097 int optname, sockptr_t optval, sockptr_t optlen)
4099 struct inet_connection_sock *icsk = inet_csk(sk);
4100 struct tcp_sock *tp = tcp_sk(sk);
4101 struct net *net = sock_net(sk);
4104 if (copy_from_sockptr(&len, optlen, sizeof(int)))
4110 len = min_t(unsigned int, len, sizeof(int));
4114 val = tp->mss_cache;
4115 if (tp->rx_opt.user_mss &&
4116 ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
4117 val = tp->rx_opt.user_mss;
4119 val = tp->rx_opt.mss_clamp;
4122 val = !!(tp->nonagle&TCP_NAGLE_OFF);
4125 val = !!(tp->nonagle&TCP_NAGLE_CORK);
4128 val = keepalive_time_when(tp) / HZ;
4131 val = keepalive_intvl_when(tp) / HZ;
4134 val = keepalive_probes(tp);
4137 val = READ_ONCE(icsk->icsk_syn_retries) ? :
4138 READ_ONCE(net->ipv4.sysctl_tcp_syn_retries);
4141 val = READ_ONCE(tp->linger2);
4143 val = (val ? : READ_ONCE(net->ipv4.sysctl_tcp_fin_timeout)) / HZ;
4145 case TCP_DEFER_ACCEPT:
4146 val = READ_ONCE(icsk->icsk_accept_queue.rskq_defer_accept);
4147 val = retrans_to_secs(val, TCP_TIMEOUT_INIT / HZ,
4150 case TCP_WINDOW_CLAMP:
4151 val = tp->window_clamp;
4154 struct tcp_info info;
4156 if (copy_from_sockptr(&len, optlen, sizeof(int)))
4159 tcp_get_info(sk, &info);
4161 len = min_t(unsigned int, len, sizeof(info));
4162 if (copy_to_sockptr(optlen, &len, sizeof(int)))
4164 if (copy_to_sockptr(optval, &info, len))
4169 const struct tcp_congestion_ops *ca_ops;
4170 union tcp_cc_info info;
4174 if (copy_from_sockptr(&len, optlen, sizeof(int)))
4177 ca_ops = icsk->icsk_ca_ops;
4178 if (ca_ops && ca_ops->get_info)
4179 sz = ca_ops->get_info(sk, ~0U, &attr, &info);
4181 len = min_t(unsigned int, len, sz);
4182 if (copy_to_sockptr(optlen, &len, sizeof(int)))
4184 if (copy_to_sockptr(optval, &info, len))
4189 val = !inet_csk_in_pingpong_mode(sk);
4192 case TCP_CONGESTION:
4193 if (copy_from_sockptr(&len, optlen, sizeof(int)))
4195 len = min_t(unsigned int, len, TCP_CA_NAME_MAX);
4196 if (copy_to_sockptr(optlen, &len, sizeof(int)))
4198 if (copy_to_sockptr(optval, icsk->icsk_ca_ops->name, len))
4203 if (copy_from_sockptr(&len, optlen, sizeof(int)))
4205 len = min_t(unsigned int, len, TCP_ULP_NAME_MAX);
4206 if (!icsk->icsk_ulp_ops) {
4208 if (copy_to_sockptr(optlen, &len, sizeof(int)))
4212 if (copy_to_sockptr(optlen, &len, sizeof(int)))
4214 if (copy_to_sockptr(optval, icsk->icsk_ulp_ops->name, len))
4218 case TCP_FASTOPEN_KEY: {
4219 u64 key[TCP_FASTOPEN_KEY_BUF_LENGTH / sizeof(u64)];
4220 unsigned int key_len;
4222 if (copy_from_sockptr(&len, optlen, sizeof(int)))
4225 key_len = tcp_fastopen_get_cipher(net, icsk, key) *
4226 TCP_FASTOPEN_KEY_LENGTH;
4227 len = min_t(unsigned int, len, key_len);
4228 if (copy_to_sockptr(optlen, &len, sizeof(int)))
4230 if (copy_to_sockptr(optval, key, len))
4234 case TCP_THIN_LINEAR_TIMEOUTS:
4238 case TCP_THIN_DUPACK:
4246 case TCP_REPAIR_QUEUE:
4248 val = tp->repair_queue;
4253 case TCP_REPAIR_WINDOW: {
4254 struct tcp_repair_window opt;
4256 if (copy_from_sockptr(&len, optlen, sizeof(int)))
4259 if (len != sizeof(opt))
4265 opt.snd_wl1 = tp->snd_wl1;
4266 opt.snd_wnd = tp->snd_wnd;
4267 opt.max_window = tp->max_window;
4268 opt.rcv_wnd = tp->rcv_wnd;
4269 opt.rcv_wup = tp->rcv_wup;
4271 if (copy_to_sockptr(optval, &opt, len))
4276 if (tp->repair_queue == TCP_SEND_QUEUE)
4277 val = tp->write_seq;
4278 else if (tp->repair_queue == TCP_RECV_QUEUE)
4284 case TCP_USER_TIMEOUT:
4285 val = READ_ONCE(icsk->icsk_user_timeout);
4289 val = READ_ONCE(icsk->icsk_accept_queue.fastopenq.max_qlen);
4292 case TCP_FASTOPEN_CONNECT:
4293 val = tp->fastopen_connect;
4296 case TCP_FASTOPEN_NO_COOKIE:
4297 val = tp->fastopen_no_cookie;
4301 val = READ_ONCE(tp->tcp_tx_delay);
4305 val = tcp_time_stamp_raw() + READ_ONCE(tp->tsoffset);
4307 case TCP_NOTSENT_LOWAT:
4308 val = READ_ONCE(tp->notsent_lowat);
4311 val = tp->recvmsg_inq;
4316 case TCP_SAVED_SYN: {
4317 if (copy_from_sockptr(&len, optlen, sizeof(int)))
4320 sockopt_lock_sock(sk);
4321 if (tp->saved_syn) {
4322 if (len < tcp_saved_syn_len(tp->saved_syn)) {
4323 len = tcp_saved_syn_len(tp->saved_syn);
4324 if (copy_to_sockptr(optlen, &len, sizeof(int))) {
4325 sockopt_release_sock(sk);
4328 sockopt_release_sock(sk);
4331 len = tcp_saved_syn_len(tp->saved_syn);
4332 if (copy_to_sockptr(optlen, &len, sizeof(int))) {
4333 sockopt_release_sock(sk);
4336 if (copy_to_sockptr(optval, tp->saved_syn->data, len)) {
4337 sockopt_release_sock(sk);
4340 tcp_saved_syn_free(tp);
4341 sockopt_release_sock(sk);
4343 sockopt_release_sock(sk);
4345 if (copy_to_sockptr(optlen, &len, sizeof(int)))
4351 case TCP_ZEROCOPY_RECEIVE: {
4352 struct scm_timestamping_internal tss;
4353 struct tcp_zerocopy_receive zc = {};
4356 if (copy_from_sockptr(&len, optlen, sizeof(int)))
4359 len < offsetofend(struct tcp_zerocopy_receive, length))
4361 if (unlikely(len > sizeof(zc))) {
4362 err = check_zeroed_sockptr(optval, sizeof(zc),
4365 return err == 0 ? -EINVAL : err;
4367 if (copy_to_sockptr(optlen, &len, sizeof(int)))
4370 if (copy_from_sockptr(&zc, optval, len))
4374 if (zc.msg_flags & ~(TCP_VALID_ZC_MSG_FLAGS))
4376 sockopt_lock_sock(sk);
4377 err = tcp_zerocopy_receive(sk, &zc, &tss);
4378 err = BPF_CGROUP_RUN_PROG_GETSOCKOPT_KERN(sk, level, optname,
4380 sockopt_release_sock(sk);
4381 if (len >= offsetofend(struct tcp_zerocopy_receive, msg_flags))
4382 goto zerocopy_rcv_cmsg;
4384 case offsetofend(struct tcp_zerocopy_receive, msg_flags):
4385 goto zerocopy_rcv_cmsg;
4386 case offsetofend(struct tcp_zerocopy_receive, msg_controllen):
4387 case offsetofend(struct tcp_zerocopy_receive, msg_control):
4388 case offsetofend(struct tcp_zerocopy_receive, flags):
4389 case offsetofend(struct tcp_zerocopy_receive, copybuf_len):
4390 case offsetofend(struct tcp_zerocopy_receive, copybuf_address):
4391 case offsetofend(struct tcp_zerocopy_receive, err):
4392 goto zerocopy_rcv_sk_err;
4393 case offsetofend(struct tcp_zerocopy_receive, inq):
4394 goto zerocopy_rcv_inq;
4395 case offsetofend(struct tcp_zerocopy_receive, length):
4397 goto zerocopy_rcv_out;
4400 if (zc.msg_flags & TCP_CMSG_TS)
4401 tcp_zc_finalize_rx_tstamp(sk, &zc, &tss);
4404 zerocopy_rcv_sk_err:
4406 zc.err = sock_error(sk);
4408 zc.inq = tcp_inq_hint(sk);
4410 if (!err && copy_to_sockptr(optval, &zc, len))
4416 return -ENOPROTOOPT;
4419 if (copy_to_sockptr(optlen, &len, sizeof(int)))
4421 if (copy_to_sockptr(optval, &val, len))
4426 bool tcp_bpf_bypass_getsockopt(int level, int optname)
4428 /* TCP do_tcp_getsockopt has optimized getsockopt implementation
4429 * to avoid extra socket lock for TCP_ZEROCOPY_RECEIVE.
4431 if (level == SOL_TCP && optname == TCP_ZEROCOPY_RECEIVE)
4436 EXPORT_SYMBOL(tcp_bpf_bypass_getsockopt);
4438 int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval,
4441 struct inet_connection_sock *icsk = inet_csk(sk);
4443 if (level != SOL_TCP)
4444 /* Paired with WRITE_ONCE() in do_ipv6_setsockopt() and tcp_v6_connect() */
4445 return READ_ONCE(icsk->icsk_af_ops)->getsockopt(sk, level, optname,
4447 return do_tcp_getsockopt(sk, level, optname, USER_SOCKPTR(optval),
4448 USER_SOCKPTR(optlen));
4450 EXPORT_SYMBOL(tcp_getsockopt);
4452 #ifdef CONFIG_TCP_MD5SIG
4453 static DEFINE_PER_CPU(struct tcp_md5sig_pool, tcp_md5sig_pool);
4454 static DEFINE_MUTEX(tcp_md5sig_mutex);
4455 static bool tcp_md5sig_pool_populated = false;
4457 static void __tcp_alloc_md5sig_pool(void)
4459 struct crypto_ahash *hash;
4462 hash = crypto_alloc_ahash("md5", 0, CRYPTO_ALG_ASYNC);
4466 for_each_possible_cpu(cpu) {
4467 void *scratch = per_cpu(tcp_md5sig_pool, cpu).scratch;
4468 struct ahash_request *req;
4471 scratch = kmalloc_node(sizeof(union tcp_md5sum_block) +
4472 sizeof(struct tcphdr),
4477 per_cpu(tcp_md5sig_pool, cpu).scratch = scratch;
4479 if (per_cpu(tcp_md5sig_pool, cpu).md5_req)
4482 req = ahash_request_alloc(hash, GFP_KERNEL);
4486 ahash_request_set_callback(req, 0, NULL, NULL);
4488 per_cpu(tcp_md5sig_pool, cpu).md5_req = req;
4490 /* before setting tcp_md5sig_pool_populated, we must commit all writes
4491 * to memory. See smp_rmb() in tcp_get_md5sig_pool()
4494 /* Paired with READ_ONCE() from tcp_alloc_md5sig_pool()
4495 * and tcp_get_md5sig_pool().
4497 WRITE_ONCE(tcp_md5sig_pool_populated, true);
4500 bool tcp_alloc_md5sig_pool(void)
4502 /* Paired with WRITE_ONCE() from __tcp_alloc_md5sig_pool() */
4503 if (unlikely(!READ_ONCE(tcp_md5sig_pool_populated))) {
4504 mutex_lock(&tcp_md5sig_mutex);
4506 if (!tcp_md5sig_pool_populated) {
4507 __tcp_alloc_md5sig_pool();
4508 if (tcp_md5sig_pool_populated)
4509 static_branch_inc(&tcp_md5_needed);
4512 mutex_unlock(&tcp_md5sig_mutex);
4514 /* Paired with WRITE_ONCE() from __tcp_alloc_md5sig_pool() */
4515 return READ_ONCE(tcp_md5sig_pool_populated);
4517 EXPORT_SYMBOL(tcp_alloc_md5sig_pool);
4521 * tcp_get_md5sig_pool - get md5sig_pool for this user
4523 * We use percpu structure, so if we succeed, we exit with preemption
4524 * and BH disabled, to make sure another thread or softirq handling
4525 * wont try to get same context.
4527 struct tcp_md5sig_pool *tcp_get_md5sig_pool(void)
4531 /* Paired with WRITE_ONCE() from __tcp_alloc_md5sig_pool() */
4532 if (READ_ONCE(tcp_md5sig_pool_populated)) {
4533 /* coupled with smp_wmb() in __tcp_alloc_md5sig_pool() */
4535 return this_cpu_ptr(&tcp_md5sig_pool);
4540 EXPORT_SYMBOL(tcp_get_md5sig_pool);
4542 int tcp_md5_hash_skb_data(struct tcp_md5sig_pool *hp,
4543 const struct sk_buff *skb, unsigned int header_len)
4545 struct scatterlist sg;
4546 const struct tcphdr *tp = tcp_hdr(skb);
4547 struct ahash_request *req = hp->md5_req;
4549 const unsigned int head_data_len = skb_headlen(skb) > header_len ?
4550 skb_headlen(skb) - header_len : 0;
4551 const struct skb_shared_info *shi = skb_shinfo(skb);
4552 struct sk_buff *frag_iter;
4554 sg_init_table(&sg, 1);
4556 sg_set_buf(&sg, ((u8 *) tp) + header_len, head_data_len);
4557 ahash_request_set_crypt(req, &sg, NULL, head_data_len);
4558 if (crypto_ahash_update(req))
4561 for (i = 0; i < shi->nr_frags; ++i) {
4562 const skb_frag_t *f = &shi->frags[i];
4563 unsigned int offset = skb_frag_off(f);
4564 struct page *page = skb_frag_page(f) + (offset >> PAGE_SHIFT);
4566 sg_set_page(&sg, page, skb_frag_size(f),
4567 offset_in_page(offset));
4568 ahash_request_set_crypt(req, &sg, NULL, skb_frag_size(f));
4569 if (crypto_ahash_update(req))
4573 skb_walk_frags(skb, frag_iter)
4574 if (tcp_md5_hash_skb_data(hp, frag_iter, 0))
4579 EXPORT_SYMBOL(tcp_md5_hash_skb_data);
4581 int tcp_md5_hash_key(struct tcp_md5sig_pool *hp, const struct tcp_md5sig_key *key)
4583 u8 keylen = READ_ONCE(key->keylen); /* paired with WRITE_ONCE() in tcp_md5_do_add */
4584 struct scatterlist sg;
4586 sg_init_one(&sg, key->key, keylen);
4587 ahash_request_set_crypt(hp->md5_req, &sg, NULL, keylen);
4589 /* We use data_race() because tcp_md5_do_add() might change key->key under us */
4590 return data_race(crypto_ahash_update(hp->md5_req));
4592 EXPORT_SYMBOL(tcp_md5_hash_key);
4594 /* Called with rcu_read_lock() */
4595 enum skb_drop_reason
4596 tcp_inbound_md5_hash(const struct sock *sk, const struct sk_buff *skb,
4597 const void *saddr, const void *daddr,
4598 int family, int dif, int sdif)
4601 * This gets called for each TCP segment that arrives
4602 * so we want to be efficient.
4603 * We have 3 drop cases:
4604 * o No MD5 hash and one expected.
4605 * o MD5 hash and we're not expecting one.
4606 * o MD5 hash and its wrong.
4608 const __u8 *hash_location = NULL;
4609 struct tcp_md5sig_key *hash_expected;
4610 const struct tcphdr *th = tcp_hdr(skb);
4611 struct tcp_sock *tp = tcp_sk(sk);
4612 int genhash, l3index;
4615 /* sdif set, means packet ingressed via a device
4616 * in an L3 domain and dif is set to the l3mdev
4618 l3index = sdif ? dif : 0;
4620 hash_expected = tcp_md5_do_lookup(sk, l3index, saddr, family);
4621 hash_location = tcp_parse_md5sig_option(th);
4623 /* We've parsed the options - do we have a hash? */
4624 if (!hash_expected && !hash_location)
4625 return SKB_NOT_DROPPED_YET;
4627 if (hash_expected && !hash_location) {
4628 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMD5NOTFOUND);
4629 return SKB_DROP_REASON_TCP_MD5NOTFOUND;
4632 if (!hash_expected && hash_location) {
4633 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMD5UNEXPECTED);
4634 return SKB_DROP_REASON_TCP_MD5UNEXPECTED;
4637 /* Check the signature.
4638 * To support dual stack listeners, we need to handle
4641 if (family == AF_INET)
4642 genhash = tcp_v4_md5_hash_skb(newhash,
4646 genhash = tp->af_specific->calc_md5_hash(newhash,
4650 if (genhash || memcmp(hash_location, newhash, 16) != 0) {
4651 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMD5FAILURE);
4652 if (family == AF_INET) {
4653 net_info_ratelimited("MD5 Hash failed for (%pI4, %d)->(%pI4, %d)%s L3 index %d\n",
4654 saddr, ntohs(th->source),
4655 daddr, ntohs(th->dest),
4656 genhash ? " tcp_v4_calc_md5_hash failed"
4659 net_info_ratelimited("MD5 Hash %s for [%pI6c]:%u->[%pI6c]:%u L3 index %d\n",
4660 genhash ? "failed" : "mismatch",
4661 saddr, ntohs(th->source),
4662 daddr, ntohs(th->dest), l3index);
4664 return SKB_DROP_REASON_TCP_MD5FAILURE;
4666 return SKB_NOT_DROPPED_YET;
4668 EXPORT_SYMBOL(tcp_inbound_md5_hash);
4672 void tcp_done(struct sock *sk)
4674 struct request_sock *req;
4676 /* We might be called with a new socket, after
4677 * inet_csk_prepare_forced_close() has been called
4678 * so we can not use lockdep_sock_is_held(sk)
4680 req = rcu_dereference_protected(tcp_sk(sk)->fastopen_rsk, 1);
4682 if (sk->sk_state == TCP_SYN_SENT || sk->sk_state == TCP_SYN_RECV)
4683 TCP_INC_STATS(sock_net(sk), TCP_MIB_ATTEMPTFAILS);
4685 tcp_set_state(sk, TCP_CLOSE);
4686 tcp_clear_xmit_timers(sk);
4688 reqsk_fastopen_remove(sk, req, false);
4690 WRITE_ONCE(sk->sk_shutdown, SHUTDOWN_MASK);
4692 if (!sock_flag(sk, SOCK_DEAD))
4693 sk->sk_state_change(sk);
4695 inet_csk_destroy_sock(sk);
4697 EXPORT_SYMBOL_GPL(tcp_done);
4699 int tcp_abort(struct sock *sk, int err)
4701 int state = inet_sk_state_load(sk);
4703 if (state == TCP_NEW_SYN_RECV) {
4704 struct request_sock *req = inet_reqsk(sk);
4707 inet_csk_reqsk_queue_drop(req->rsk_listener, req);
4711 if (state == TCP_TIME_WAIT) {
4712 struct inet_timewait_sock *tw = inet_twsk(sk);
4714 refcount_inc(&tw->tw_refcnt);
4716 inet_twsk_deschedule_put(tw);
4721 /* Don't race with userspace socket closes such as tcp_close. */
4724 if (sk->sk_state == TCP_LISTEN) {
4725 tcp_set_state(sk, TCP_CLOSE);
4726 inet_csk_listen_stop(sk);
4729 /* Don't race with BH socket closes such as inet_csk_listen_stop. */
4733 if (!sock_flag(sk, SOCK_DEAD)) {
4735 /* This barrier is coupled with smp_rmb() in tcp_poll() */
4737 sk_error_report(sk);
4738 if (tcp_need_reset(sk->sk_state))
4739 tcp_send_active_reset(sk, GFP_ATOMIC);
4745 tcp_write_queue_purge(sk);
4749 EXPORT_SYMBOL_GPL(tcp_abort);
4751 extern struct tcp_congestion_ops tcp_reno;
4753 static __initdata unsigned long thash_entries;
4754 static int __init set_thash_entries(char *str)
4761 ret = kstrtoul(str, 0, &thash_entries);
4767 __setup("thash_entries=", set_thash_entries);
4769 static void __init tcp_init_mem(void)
4771 unsigned long limit = nr_free_buffer_pages() / 16;
4773 limit = max(limit, 128UL);
4774 sysctl_tcp_mem[0] = limit / 4 * 3; /* 4.68 % */
4775 sysctl_tcp_mem[1] = limit; /* 6.25 % */
4776 sysctl_tcp_mem[2] = sysctl_tcp_mem[0] * 2; /* 9.37 % */
4779 void __init tcp_init(void)
4781 int max_rshare, max_wshare, cnt;
4782 unsigned long limit;
4785 BUILD_BUG_ON(TCP_MIN_SND_MSS <= MAX_TCP_OPTION_SPACE);
4786 BUILD_BUG_ON(sizeof(struct tcp_skb_cb) >
4787 sizeof_field(struct sk_buff, cb));
4789 percpu_counter_init(&tcp_sockets_allocated, 0, GFP_KERNEL);
4791 timer_setup(&tcp_orphan_timer, tcp_orphan_update, TIMER_DEFERRABLE);
4792 mod_timer(&tcp_orphan_timer, jiffies + TCP_ORPHAN_TIMER_PERIOD);
4794 inet_hashinfo2_init(&tcp_hashinfo, "tcp_listen_portaddr_hash",
4795 thash_entries, 21, /* one slot per 2 MB*/
4797 tcp_hashinfo.bind_bucket_cachep =
4798 kmem_cache_create("tcp_bind_bucket",
4799 sizeof(struct inet_bind_bucket), 0,
4800 SLAB_HWCACHE_ALIGN | SLAB_PANIC |
4803 tcp_hashinfo.bind2_bucket_cachep =
4804 kmem_cache_create("tcp_bind2_bucket",
4805 sizeof(struct inet_bind2_bucket), 0,
4806 SLAB_HWCACHE_ALIGN | SLAB_PANIC |
4810 /* Size and allocate the main established and bind bucket
4813 * The methodology is similar to that of the buffer cache.
4815 tcp_hashinfo.ehash =
4816 alloc_large_system_hash("TCP established",
4817 sizeof(struct inet_ehash_bucket),
4819 17, /* one slot per 128 KB of memory */
4822 &tcp_hashinfo.ehash_mask,
4824 thash_entries ? 0 : 512 * 1024);
4825 for (i = 0; i <= tcp_hashinfo.ehash_mask; i++)
4826 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].chain, i);
4828 if (inet_ehash_locks_alloc(&tcp_hashinfo))
4829 panic("TCP: failed to alloc ehash_locks");
4830 tcp_hashinfo.bhash =
4831 alloc_large_system_hash("TCP bind",
4832 2 * sizeof(struct inet_bind_hashbucket),
4833 tcp_hashinfo.ehash_mask + 1,
4834 17, /* one slot per 128 KB of memory */
4836 &tcp_hashinfo.bhash_size,
4840 tcp_hashinfo.bhash_size = 1U << tcp_hashinfo.bhash_size;
4841 tcp_hashinfo.bhash2 = tcp_hashinfo.bhash + tcp_hashinfo.bhash_size;
4842 for (i = 0; i < tcp_hashinfo.bhash_size; i++) {
4843 spin_lock_init(&tcp_hashinfo.bhash[i].lock);
4844 INIT_HLIST_HEAD(&tcp_hashinfo.bhash[i].chain);
4845 spin_lock_init(&tcp_hashinfo.bhash2[i].lock);
4846 INIT_HLIST_HEAD(&tcp_hashinfo.bhash2[i].chain);
4849 tcp_hashinfo.pernet = false;
4851 cnt = tcp_hashinfo.ehash_mask + 1;
4852 sysctl_tcp_max_orphans = cnt / 2;
4855 /* Set per-socket limits to no more than 1/128 the pressure threshold */
4856 limit = nr_free_buffer_pages() << (PAGE_SHIFT - 7);
4857 max_wshare = min(4UL*1024*1024, limit);
4858 max_rshare = min(6UL*1024*1024, limit);
4860 init_net.ipv4.sysctl_tcp_wmem[0] = PAGE_SIZE;
4861 init_net.ipv4.sysctl_tcp_wmem[1] = 16*1024;
4862 init_net.ipv4.sysctl_tcp_wmem[2] = max(64*1024, max_wshare);
4864 init_net.ipv4.sysctl_tcp_rmem[0] = PAGE_SIZE;
4865 init_net.ipv4.sysctl_tcp_rmem[1] = 131072;
4866 init_net.ipv4.sysctl_tcp_rmem[2] = max(131072, max_rshare);
4868 pr_info("Hash tables configured (established %u bind %u)\n",
4869 tcp_hashinfo.ehash_mask + 1, tcp_hashinfo.bhash_size);
4873 BUG_ON(tcp_register_congestion_control(&tcp_reno) != 0);