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 (READ_ONCE(sk->sk_err) ||
595 !skb_queue_empty_lockless(&sk->sk_error_queue))
600 EXPORT_SYMBOL(tcp_poll);
602 int tcp_ioctl(struct sock *sk, int cmd, unsigned long arg)
604 struct tcp_sock *tp = tcp_sk(sk);
610 if (sk->sk_state == TCP_LISTEN)
613 slow = lock_sock_fast(sk);
615 unlock_sock_fast(sk, slow);
618 answ = READ_ONCE(tp->urg_data) &&
619 READ_ONCE(tp->urg_seq) == READ_ONCE(tp->copied_seq);
622 if (sk->sk_state == TCP_LISTEN)
625 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
628 answ = READ_ONCE(tp->write_seq) - tp->snd_una;
631 if (sk->sk_state == TCP_LISTEN)
634 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
637 answ = READ_ONCE(tp->write_seq) -
638 READ_ONCE(tp->snd_nxt);
644 return put_user(answ, (int __user *)arg);
646 EXPORT_SYMBOL(tcp_ioctl);
648 void tcp_mark_push(struct tcp_sock *tp, struct sk_buff *skb)
650 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
651 tp->pushed_seq = tp->write_seq;
654 static inline bool forced_push(const struct tcp_sock *tp)
656 return after(tp->write_seq, tp->pushed_seq + (tp->max_window >> 1));
659 void tcp_skb_entail(struct sock *sk, struct sk_buff *skb)
661 struct tcp_sock *tp = tcp_sk(sk);
662 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
664 tcb->seq = tcb->end_seq = tp->write_seq;
665 tcb->tcp_flags = TCPHDR_ACK;
666 __skb_header_release(skb);
667 tcp_add_write_queue_tail(sk, skb);
668 sk_wmem_queued_add(sk, skb->truesize);
669 sk_mem_charge(sk, skb->truesize);
670 if (tp->nonagle & TCP_NAGLE_PUSH)
671 tp->nonagle &= ~TCP_NAGLE_PUSH;
673 tcp_slow_start_after_idle_check(sk);
676 static inline void tcp_mark_urg(struct tcp_sock *tp, int flags)
679 tp->snd_up = tp->write_seq;
682 /* If a not yet filled skb is pushed, do not send it if
683 * we have data packets in Qdisc or NIC queues :
684 * Because TX completion will happen shortly, it gives a chance
685 * to coalesce future sendmsg() payload into this skb, without
686 * need for a timer, and with no latency trade off.
687 * As packets containing data payload have a bigger truesize
688 * than pure acks (dataless) packets, the last checks prevent
689 * autocorking if we only have an ACK in Qdisc/NIC queues,
690 * or if TX completion was delayed after we processed ACK packet.
692 static bool tcp_should_autocork(struct sock *sk, struct sk_buff *skb,
695 return skb->len < size_goal &&
696 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_autocorking) &&
697 !tcp_rtx_queue_empty(sk) &&
698 refcount_read(&sk->sk_wmem_alloc) > skb->truesize &&
699 tcp_skb_can_collapse_to(skb);
702 void tcp_push(struct sock *sk, int flags, int mss_now,
703 int nonagle, int size_goal)
705 struct tcp_sock *tp = tcp_sk(sk);
708 skb = tcp_write_queue_tail(sk);
711 if (!(flags & MSG_MORE) || forced_push(tp))
712 tcp_mark_push(tp, skb);
714 tcp_mark_urg(tp, flags);
716 if (tcp_should_autocork(sk, skb, size_goal)) {
718 /* avoid atomic op if TSQ_THROTTLED bit is already set */
719 if (!test_bit(TSQ_THROTTLED, &sk->sk_tsq_flags)) {
720 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPAUTOCORKING);
721 set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags);
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 sk_wait_data(sk, &timeo, NULL);
832 if (signal_pending(current)) {
833 ret = sock_intr_errno(timeo);
846 if (sk->sk_err || sk->sk_state == TCP_CLOSE ||
847 (sk->sk_shutdown & RCV_SHUTDOWN) ||
848 signal_pending(current))
859 EXPORT_SYMBOL(tcp_splice_read);
861 struct sk_buff *tcp_stream_alloc_skb(struct sock *sk, gfp_t gfp,
866 skb = alloc_skb_fclone(MAX_TCP_HEADER, gfp);
870 skb->truesize = SKB_TRUESIZE(skb_end_offset(skb));
871 if (force_schedule) {
872 mem_scheduled = true;
873 sk_forced_mem_schedule(sk, skb->truesize);
875 mem_scheduled = sk_wmem_schedule(sk, skb->truesize);
877 if (likely(mem_scheduled)) {
878 skb_reserve(skb, MAX_TCP_HEADER);
879 skb->ip_summed = CHECKSUM_PARTIAL;
880 INIT_LIST_HEAD(&skb->tcp_tsorted_anchor);
885 sk->sk_prot->enter_memory_pressure(sk);
886 sk_stream_moderate_sndbuf(sk);
891 static unsigned int tcp_xmit_size_goal(struct sock *sk, u32 mss_now,
894 struct tcp_sock *tp = tcp_sk(sk);
895 u32 new_size_goal, size_goal;
900 /* Note : tcp_tso_autosize() will eventually split this later */
901 new_size_goal = tcp_bound_to_half_wnd(tp, sk->sk_gso_max_size);
903 /* We try hard to avoid divides here */
904 size_goal = tp->gso_segs * mss_now;
905 if (unlikely(new_size_goal < size_goal ||
906 new_size_goal >= size_goal + mss_now)) {
907 tp->gso_segs = min_t(u16, new_size_goal / mss_now,
908 sk->sk_gso_max_segs);
909 size_goal = tp->gso_segs * mss_now;
912 return max(size_goal, mss_now);
915 int tcp_send_mss(struct sock *sk, int *size_goal, int flags)
919 mss_now = tcp_current_mss(sk);
920 *size_goal = tcp_xmit_size_goal(sk, mss_now, !(flags & MSG_OOB));
925 /* In some cases, both sendpage() and sendmsg() could have added
926 * an skb to the write queue, but failed adding payload on it.
927 * We need to remove it to consume less memory, but more
928 * importantly be able to generate EPOLLOUT for Edge Trigger epoll()
931 void tcp_remove_empty_skb(struct sock *sk)
933 struct sk_buff *skb = tcp_write_queue_tail(sk);
935 if (skb && TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq) {
936 tcp_unlink_write_queue(skb, sk);
937 if (tcp_write_queue_empty(sk))
938 tcp_chrono_stop(sk, TCP_CHRONO_BUSY);
939 tcp_wmem_free_skb(sk, skb);
943 /* skb changing from pure zc to mixed, must charge zc */
944 static int tcp_downgrade_zcopy_pure(struct sock *sk, struct sk_buff *skb)
946 if (unlikely(skb_zcopy_pure(skb))) {
947 u32 extra = skb->truesize -
948 SKB_TRUESIZE(skb_end_offset(skb));
950 if (!sk_wmem_schedule(sk, extra))
953 sk_mem_charge(sk, extra);
954 skb_shinfo(skb)->flags &= ~SKBFL_PURE_ZEROCOPY;
960 int tcp_wmem_schedule(struct sock *sk, int copy)
964 if (likely(sk_wmem_schedule(sk, copy)))
967 /* We could be in trouble if we have nothing queued.
968 * Use whatever is left in sk->sk_forward_alloc and tcp_wmem[0]
969 * to guarantee some progress.
971 left = sock_net(sk)->ipv4.sysctl_tcp_wmem[0] - sk->sk_wmem_queued;
973 sk_forced_mem_schedule(sk, min(left, copy));
974 return min(copy, sk->sk_forward_alloc);
977 int tcp_sendpage_locked(struct sock *sk, struct page *page, int offset,
978 size_t size, int flags)
981 struct msghdr msg = { .msg_flags = flags | MSG_SPLICE_PAGES, };
983 if (!(sk->sk_route_caps & NETIF_F_SG))
984 return sock_no_sendpage_locked(sk, page, offset, size, flags);
986 tcp_rate_check_app_limited(sk); /* is sending application-limited? */
988 bvec_set_page(&bvec, page, size, offset);
989 iov_iter_bvec(&msg.msg_iter, ITER_SOURCE, &bvec, 1, size);
991 if (flags & MSG_SENDPAGE_NOTLAST)
992 msg.msg_flags |= MSG_MORE;
994 return tcp_sendmsg_locked(sk, &msg, size);
996 EXPORT_SYMBOL_GPL(tcp_sendpage_locked);
998 int tcp_sendpage(struct sock *sk, struct page *page, int offset,
999 size_t size, int flags)
1004 ret = tcp_sendpage_locked(sk, page, offset, size, flags);
1009 EXPORT_SYMBOL(tcp_sendpage);
1011 void tcp_free_fastopen_req(struct tcp_sock *tp)
1013 if (tp->fastopen_req) {
1014 kfree(tp->fastopen_req);
1015 tp->fastopen_req = NULL;
1019 int tcp_sendmsg_fastopen(struct sock *sk, struct msghdr *msg, int *copied,
1020 size_t size, struct ubuf_info *uarg)
1022 struct tcp_sock *tp = tcp_sk(sk);
1023 struct inet_sock *inet = inet_sk(sk);
1024 struct sockaddr *uaddr = msg->msg_name;
1027 if (!(READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_fastopen) &
1028 TFO_CLIENT_ENABLE) ||
1029 (uaddr && msg->msg_namelen >= sizeof(uaddr->sa_family) &&
1030 uaddr->sa_family == AF_UNSPEC))
1032 if (tp->fastopen_req)
1033 return -EALREADY; /* Another Fast Open is in progress */
1035 tp->fastopen_req = kzalloc(sizeof(struct tcp_fastopen_request),
1037 if (unlikely(!tp->fastopen_req))
1039 tp->fastopen_req->data = msg;
1040 tp->fastopen_req->size = size;
1041 tp->fastopen_req->uarg = uarg;
1043 if (inet->defer_connect) {
1044 err = tcp_connect(sk);
1045 /* Same failure procedure as in tcp_v4/6_connect */
1047 tcp_set_state(sk, TCP_CLOSE);
1048 inet->inet_dport = 0;
1049 sk->sk_route_caps = 0;
1052 flags = (msg->msg_flags & MSG_DONTWAIT) ? O_NONBLOCK : 0;
1053 err = __inet_stream_connect(sk->sk_socket, uaddr,
1054 msg->msg_namelen, flags, 1);
1055 /* fastopen_req could already be freed in __inet_stream_connect
1056 * if the connection times out or gets rst
1058 if (tp->fastopen_req) {
1059 *copied = tp->fastopen_req->copied;
1060 tcp_free_fastopen_req(tp);
1061 inet->defer_connect = 0;
1066 int tcp_sendmsg_locked(struct sock *sk, struct msghdr *msg, size_t size)
1068 struct tcp_sock *tp = tcp_sk(sk);
1069 struct ubuf_info *uarg = NULL;
1070 struct sk_buff *skb;
1071 struct sockcm_cookie sockc;
1072 int flags, err, copied = 0;
1073 int mss_now = 0, size_goal, copied_syn = 0;
1074 int process_backlog = 0;
1078 flags = msg->msg_flags;
1080 if ((flags & MSG_ZEROCOPY) && size) {
1081 if (msg->msg_ubuf) {
1082 uarg = msg->msg_ubuf;
1083 if (sk->sk_route_caps & NETIF_F_SG)
1085 } else if (sock_flag(sk, SOCK_ZEROCOPY)) {
1086 skb = tcp_write_queue_tail(sk);
1087 uarg = msg_zerocopy_realloc(sk, size, skb_zcopy(skb));
1092 if (sk->sk_route_caps & NETIF_F_SG)
1095 uarg_to_msgzc(uarg)->zerocopy = 0;
1097 } else if (unlikely(msg->msg_flags & MSG_SPLICE_PAGES) && size) {
1098 if (sk->sk_route_caps & NETIF_F_SG)
1099 zc = MSG_SPLICE_PAGES;
1102 if (unlikely(flags & MSG_FASTOPEN || inet_sk(sk)->defer_connect) &&
1104 err = tcp_sendmsg_fastopen(sk, msg, &copied_syn, size, uarg);
1105 if (err == -EINPROGRESS && copied_syn > 0)
1111 timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
1113 tcp_rate_check_app_limited(sk); /* is sending application-limited? */
1115 /* Wait for a connection to finish. One exception is TCP Fast Open
1116 * (passive side) where data is allowed to be sent before a connection
1117 * is fully established.
1119 if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) &&
1120 !tcp_passive_fastopen(sk)) {
1121 err = sk_stream_wait_connect(sk, &timeo);
1126 if (unlikely(tp->repair)) {
1127 if (tp->repair_queue == TCP_RECV_QUEUE) {
1128 copied = tcp_send_rcvq(sk, msg, size);
1133 if (tp->repair_queue == TCP_NO_QUEUE)
1136 /* 'common' sending to sendq */
1139 sockcm_init(&sockc, sk);
1140 if (msg->msg_controllen) {
1141 err = sock_cmsg_send(sk, msg, &sockc);
1142 if (unlikely(err)) {
1148 /* This should be in poll */
1149 sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
1151 /* Ok commence sending. */
1155 mss_now = tcp_send_mss(sk, &size_goal, flags);
1158 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
1161 while (msg_data_left(msg)) {
1164 skb = tcp_write_queue_tail(sk);
1166 copy = size_goal - skb->len;
1168 if (copy <= 0 || !tcp_skb_can_collapse_to(skb)) {
1172 if (!sk_stream_memory_free(sk))
1173 goto wait_for_space;
1175 if (unlikely(process_backlog >= 16)) {
1176 process_backlog = 0;
1177 if (sk_flush_backlog(sk))
1180 first_skb = tcp_rtx_and_write_queues_empty(sk);
1181 skb = tcp_stream_alloc_skb(sk, sk->sk_allocation,
1184 goto wait_for_space;
1188 tcp_skb_entail(sk, skb);
1191 /* All packets are restored as if they have
1192 * already been sent. skb_mstamp_ns isn't set to
1193 * avoid wrong rtt estimation.
1196 TCP_SKB_CB(skb)->sacked |= TCPCB_REPAIRED;
1199 /* Try to append data to the end of skb. */
1200 if (copy > msg_data_left(msg))
1201 copy = msg_data_left(msg);
1205 int i = skb_shinfo(skb)->nr_frags;
1206 struct page_frag *pfrag = sk_page_frag(sk);
1208 if (!sk_page_frag_refill(sk, pfrag))
1209 goto wait_for_space;
1211 if (!skb_can_coalesce(skb, i, pfrag->page,
1213 if (i >= READ_ONCE(sysctl_max_skb_frags)) {
1214 tcp_mark_push(tp, skb);
1220 copy = min_t(int, copy, pfrag->size - pfrag->offset);
1222 if (unlikely(skb_zcopy_pure(skb) || skb_zcopy_managed(skb))) {
1223 if (tcp_downgrade_zcopy_pure(sk, skb))
1224 goto wait_for_space;
1225 skb_zcopy_downgrade_managed(skb);
1228 copy = tcp_wmem_schedule(sk, copy);
1230 goto wait_for_space;
1232 err = skb_copy_to_page_nocache(sk, &msg->msg_iter, skb,
1239 /* Update the skb. */
1241 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1243 skb_fill_page_desc(skb, i, pfrag->page,
1244 pfrag->offset, copy);
1245 page_ref_inc(pfrag->page);
1247 pfrag->offset += copy;
1248 } else if (zc == MSG_ZEROCOPY) {
1249 /* First append to a fragless skb builds initial
1253 skb_shinfo(skb)->flags |= SKBFL_PURE_ZEROCOPY;
1255 if (!skb_zcopy_pure(skb)) {
1256 copy = tcp_wmem_schedule(sk, copy);
1258 goto wait_for_space;
1261 err = skb_zerocopy_iter_stream(sk, skb, msg, copy, uarg);
1262 if (err == -EMSGSIZE || err == -EEXIST) {
1263 tcp_mark_push(tp, skb);
1269 } else if (zc == MSG_SPLICE_PAGES) {
1270 /* Splice in data if we can; copy if we can't. */
1271 if (tcp_downgrade_zcopy_pure(sk, skb))
1272 goto wait_for_space;
1273 copy = tcp_wmem_schedule(sk, copy);
1275 goto wait_for_space;
1277 err = skb_splice_from_iter(skb, &msg->msg_iter, copy,
1280 if (err == -EMSGSIZE) {
1281 tcp_mark_push(tp, skb);
1288 if (!(flags & MSG_NO_SHARED_FRAGS))
1289 skb_shinfo(skb)->flags |= SKBFL_SHARED_FRAG;
1291 sk_wmem_queued_add(sk, copy);
1292 sk_mem_charge(sk, copy);
1296 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
1298 WRITE_ONCE(tp->write_seq, tp->write_seq + copy);
1299 TCP_SKB_CB(skb)->end_seq += copy;
1300 tcp_skb_pcount_set(skb, 0);
1303 if (!msg_data_left(msg)) {
1304 if (unlikely(flags & MSG_EOR))
1305 TCP_SKB_CB(skb)->eor = 1;
1309 if (skb->len < size_goal || (flags & MSG_OOB) || unlikely(tp->repair))
1312 if (forced_push(tp)) {
1313 tcp_mark_push(tp, skb);
1314 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
1315 } else if (skb == tcp_send_head(sk))
1316 tcp_push_one(sk, mss_now);
1320 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1322 tcp_push(sk, flags & ~MSG_MORE, mss_now,
1323 TCP_NAGLE_PUSH, size_goal);
1325 err = sk_stream_wait_memory(sk, &timeo);
1329 mss_now = tcp_send_mss(sk, &size_goal, flags);
1334 tcp_tx_timestamp(sk, sockc.tsflags);
1335 tcp_push(sk, flags, mss_now, tp->nonagle, size_goal);
1338 /* msg->msg_ubuf is pinned by the caller so we don't take extra refs */
1339 if (uarg && !msg->msg_ubuf)
1340 net_zcopy_put(uarg);
1341 return copied + copied_syn;
1344 tcp_remove_empty_skb(sk);
1346 if (copied + copied_syn)
1349 /* msg->msg_ubuf is pinned by the caller so we don't take extra refs */
1350 if (uarg && !msg->msg_ubuf)
1351 net_zcopy_put_abort(uarg, true);
1352 err = sk_stream_error(sk, flags, err);
1353 /* make sure we wake any epoll edge trigger waiter */
1354 if (unlikely(tcp_rtx_and_write_queues_empty(sk) && err == -EAGAIN)) {
1355 sk->sk_write_space(sk);
1356 tcp_chrono_stop(sk, TCP_CHRONO_SNDBUF_LIMITED);
1360 EXPORT_SYMBOL_GPL(tcp_sendmsg_locked);
1362 int tcp_sendmsg(struct sock *sk, struct msghdr *msg, size_t size)
1367 ret = tcp_sendmsg_locked(sk, msg, size);
1372 EXPORT_SYMBOL(tcp_sendmsg);
1374 void tcp_splice_eof(struct socket *sock)
1376 struct sock *sk = sock->sk;
1377 struct tcp_sock *tp = tcp_sk(sk);
1378 int mss_now, size_goal;
1380 if (!tcp_write_queue_tail(sk))
1384 mss_now = tcp_send_mss(sk, &size_goal, 0);
1385 tcp_push(sk, 0, mss_now, tp->nonagle, size_goal);
1388 EXPORT_SYMBOL_GPL(tcp_splice_eof);
1391 * Handle reading urgent data. BSD has very simple semantics for
1392 * this, no blocking and very strange errors 8)
1395 static int tcp_recv_urg(struct sock *sk, struct msghdr *msg, int len, int flags)
1397 struct tcp_sock *tp = tcp_sk(sk);
1399 /* No URG data to read. */
1400 if (sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data ||
1401 tp->urg_data == TCP_URG_READ)
1402 return -EINVAL; /* Yes this is right ! */
1404 if (sk->sk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DONE))
1407 if (tp->urg_data & TCP_URG_VALID) {
1409 char c = tp->urg_data;
1411 if (!(flags & MSG_PEEK))
1412 WRITE_ONCE(tp->urg_data, TCP_URG_READ);
1414 /* Read urgent data. */
1415 msg->msg_flags |= MSG_OOB;
1418 if (!(flags & MSG_TRUNC))
1419 err = memcpy_to_msg(msg, &c, 1);
1422 msg->msg_flags |= MSG_TRUNC;
1424 return err ? -EFAULT : len;
1427 if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
1430 /* Fixed the recv(..., MSG_OOB) behaviour. BSD docs and
1431 * the available implementations agree in this case:
1432 * this call should never block, independent of the
1433 * blocking state of the socket.
1434 * Mike <pall@rz.uni-karlsruhe.de>
1439 static int tcp_peek_sndq(struct sock *sk, struct msghdr *msg, int len)
1441 struct sk_buff *skb;
1442 int copied = 0, err = 0;
1444 /* XXX -- need to support SO_PEEK_OFF */
1446 skb_rbtree_walk(skb, &sk->tcp_rtx_queue) {
1447 err = skb_copy_datagram_msg(skb, 0, msg, skb->len);
1453 skb_queue_walk(&sk->sk_write_queue, skb) {
1454 err = skb_copy_datagram_msg(skb, 0, msg, skb->len);
1461 return err ?: copied;
1464 /* Clean up the receive buffer for full frames taken by the user,
1465 * then send an ACK if necessary. COPIED is the number of bytes
1466 * tcp_recvmsg has given to the user so far, it speeds up the
1467 * calculation of whether or not we must ACK for the sake of
1470 void __tcp_cleanup_rbuf(struct sock *sk, int copied)
1472 struct tcp_sock *tp = tcp_sk(sk);
1473 bool time_to_ack = false;
1475 if (inet_csk_ack_scheduled(sk)) {
1476 const struct inet_connection_sock *icsk = inet_csk(sk);
1478 if (/* Once-per-two-segments ACK was not sent by tcp_input.c */
1479 tp->rcv_nxt - tp->rcv_wup > icsk->icsk_ack.rcv_mss ||
1481 * If this read emptied read buffer, we send ACK, if
1482 * connection is not bidirectional, user drained
1483 * receive buffer and there was a small segment
1487 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED2) ||
1488 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED) &&
1489 !inet_csk_in_pingpong_mode(sk))) &&
1490 !atomic_read(&sk->sk_rmem_alloc)))
1494 /* We send an ACK if we can now advertise a non-zero window
1495 * which has been raised "significantly".
1497 * Even if window raised up to infinity, do not send window open ACK
1498 * in states, where we will not receive more. It is useless.
1500 if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) {
1501 __u32 rcv_window_now = tcp_receive_window(tp);
1503 /* Optimize, __tcp_select_window() is not cheap. */
1504 if (2*rcv_window_now <= tp->window_clamp) {
1505 __u32 new_window = __tcp_select_window(sk);
1507 /* Send ACK now, if this read freed lots of space
1508 * in our buffer. Certainly, new_window is new window.
1509 * We can advertise it now, if it is not less than current one.
1510 * "Lots" means "at least twice" here.
1512 if (new_window && new_window >= 2 * rcv_window_now)
1520 void tcp_cleanup_rbuf(struct sock *sk, int copied)
1522 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
1523 struct tcp_sock *tp = tcp_sk(sk);
1525 WARN(skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq),
1526 "cleanup rbuf bug: copied %X seq %X rcvnxt %X\n",
1527 tp->copied_seq, TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt);
1528 __tcp_cleanup_rbuf(sk, copied);
1531 static void tcp_eat_recv_skb(struct sock *sk, struct sk_buff *skb)
1533 __skb_unlink(skb, &sk->sk_receive_queue);
1534 if (likely(skb->destructor == sock_rfree)) {
1536 skb->destructor = NULL;
1538 return skb_attempt_defer_free(skb);
1543 struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off)
1545 struct sk_buff *skb;
1548 while ((skb = skb_peek(&sk->sk_receive_queue)) != NULL) {
1549 offset = seq - TCP_SKB_CB(skb)->seq;
1550 if (unlikely(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
1551 pr_err_once("%s: found a SYN, please report !\n", __func__);
1554 if (offset < skb->len || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)) {
1558 /* This looks weird, but this can happen if TCP collapsing
1559 * splitted a fat GRO packet, while we released socket lock
1560 * in skb_splice_bits()
1562 tcp_eat_recv_skb(sk, skb);
1566 EXPORT_SYMBOL(tcp_recv_skb);
1569 * This routine provides an alternative to tcp_recvmsg() for routines
1570 * that would like to handle copying from skbuffs directly in 'sendfile'
1573 * - It is assumed that the socket was locked by the caller.
1574 * - The routine does not block.
1575 * - At present, there is no support for reading OOB data
1576 * or for 'peeking' the socket using this routine
1577 * (although both would be easy to implement).
1579 int tcp_read_sock(struct sock *sk, read_descriptor_t *desc,
1580 sk_read_actor_t recv_actor)
1582 struct sk_buff *skb;
1583 struct tcp_sock *tp = tcp_sk(sk);
1584 u32 seq = tp->copied_seq;
1588 if (sk->sk_state == TCP_LISTEN)
1590 while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1591 if (offset < skb->len) {
1595 len = skb->len - offset;
1596 /* Stop reading if we hit a patch of urgent data */
1597 if (unlikely(tp->urg_data)) {
1598 u32 urg_offset = tp->urg_seq - seq;
1599 if (urg_offset < len)
1604 used = recv_actor(desc, skb, offset, len);
1610 if (WARN_ON_ONCE(used > len))
1616 /* If recv_actor drops the lock (e.g. TCP splice
1617 * receive) the skb pointer might be invalid when
1618 * getting here: tcp_collapse might have deleted it
1619 * while aggregating skbs from the socket queue.
1621 skb = tcp_recv_skb(sk, seq - 1, &offset);
1624 /* TCP coalescing might have appended data to the skb.
1625 * Try to splice more frags
1627 if (offset + 1 != skb->len)
1630 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) {
1631 tcp_eat_recv_skb(sk, skb);
1635 tcp_eat_recv_skb(sk, skb);
1638 WRITE_ONCE(tp->copied_seq, seq);
1640 WRITE_ONCE(tp->copied_seq, seq);
1642 tcp_rcv_space_adjust(sk);
1644 /* Clean up data we have read: This will do ACK frames. */
1646 tcp_recv_skb(sk, seq, &offset);
1647 tcp_cleanup_rbuf(sk, copied);
1651 EXPORT_SYMBOL(tcp_read_sock);
1653 int tcp_read_skb(struct sock *sk, skb_read_actor_t recv_actor)
1655 struct tcp_sock *tp = tcp_sk(sk);
1656 u32 seq = tp->copied_seq;
1657 struct sk_buff *skb;
1661 if (sk->sk_state == TCP_LISTEN)
1664 while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1668 __skb_unlink(skb, &sk->sk_receive_queue);
1669 WARN_ON_ONCE(!skb_set_owner_sk_safe(skb, sk));
1670 tcp_flags = TCP_SKB_CB(skb)->tcp_flags;
1671 used = recv_actor(sk, skb);
1680 if (tcp_flags & TCPHDR_FIN) {
1687 EXPORT_SYMBOL(tcp_read_skb);
1689 void tcp_read_done(struct sock *sk, size_t len)
1691 struct tcp_sock *tp = tcp_sk(sk);
1692 u32 seq = tp->copied_seq;
1693 struct sk_buff *skb;
1697 if (sk->sk_state == TCP_LISTEN)
1701 while (left && (skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1704 used = min_t(size_t, skb->len - offset, left);
1708 if (skb->len > offset + used)
1711 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) {
1712 tcp_eat_recv_skb(sk, skb);
1716 tcp_eat_recv_skb(sk, skb);
1718 WRITE_ONCE(tp->copied_seq, seq);
1720 tcp_rcv_space_adjust(sk);
1722 /* Clean up data we have read: This will do ACK frames. */
1724 tcp_cleanup_rbuf(sk, len - left);
1726 EXPORT_SYMBOL(tcp_read_done);
1728 int tcp_peek_len(struct socket *sock)
1730 return tcp_inq(sock->sk);
1732 EXPORT_SYMBOL(tcp_peek_len);
1734 /* Make sure sk_rcvbuf is big enough to satisfy SO_RCVLOWAT hint */
1735 int tcp_set_rcvlowat(struct sock *sk, int val)
1739 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK)
1740 cap = sk->sk_rcvbuf >> 1;
1742 cap = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_rmem[2]) >> 1;
1743 val = min(val, cap);
1744 WRITE_ONCE(sk->sk_rcvlowat, val ? : 1);
1746 /* Check if we need to signal EPOLLIN right now */
1749 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK)
1753 if (val > sk->sk_rcvbuf) {
1754 WRITE_ONCE(sk->sk_rcvbuf, val);
1755 tcp_sk(sk)->window_clamp = tcp_win_from_space(sk, val);
1759 EXPORT_SYMBOL(tcp_set_rcvlowat);
1761 void tcp_update_recv_tstamps(struct sk_buff *skb,
1762 struct scm_timestamping_internal *tss)
1765 tss->ts[0] = ktime_to_timespec64(skb->tstamp);
1767 tss->ts[0] = (struct timespec64) {0};
1769 if (skb_hwtstamps(skb)->hwtstamp)
1770 tss->ts[2] = ktime_to_timespec64(skb_hwtstamps(skb)->hwtstamp);
1772 tss->ts[2] = (struct timespec64) {0};
1776 static const struct vm_operations_struct tcp_vm_ops = {
1779 int tcp_mmap(struct file *file, struct socket *sock,
1780 struct vm_area_struct *vma)
1782 if (vma->vm_flags & (VM_WRITE | VM_EXEC))
1784 vm_flags_clear(vma, VM_MAYWRITE | VM_MAYEXEC);
1786 /* Instruct vm_insert_page() to not mmap_read_lock(mm) */
1787 vm_flags_set(vma, VM_MIXEDMAP);
1789 vma->vm_ops = &tcp_vm_ops;
1792 EXPORT_SYMBOL(tcp_mmap);
1794 static skb_frag_t *skb_advance_to_frag(struct sk_buff *skb, u32 offset_skb,
1799 if (unlikely(offset_skb >= skb->len))
1802 offset_skb -= skb_headlen(skb);
1803 if ((int)offset_skb < 0 || skb_has_frag_list(skb))
1806 frag = skb_shinfo(skb)->frags;
1807 while (offset_skb) {
1808 if (skb_frag_size(frag) > offset_skb) {
1809 *offset_frag = offset_skb;
1812 offset_skb -= skb_frag_size(frag);
1819 static bool can_map_frag(const skb_frag_t *frag)
1821 return skb_frag_size(frag) == PAGE_SIZE && !skb_frag_off(frag);
1824 static int find_next_mappable_frag(const skb_frag_t *frag,
1825 int remaining_in_skb)
1829 if (likely(can_map_frag(frag)))
1832 while (offset < remaining_in_skb && !can_map_frag(frag)) {
1833 offset += skb_frag_size(frag);
1839 static void tcp_zerocopy_set_hint_for_skb(struct sock *sk,
1840 struct tcp_zerocopy_receive *zc,
1841 struct sk_buff *skb, u32 offset)
1843 u32 frag_offset, partial_frag_remainder = 0;
1844 int mappable_offset;
1847 /* worst case: skip to next skb. try to improve on this case below */
1848 zc->recv_skip_hint = skb->len - offset;
1850 /* Find the frag containing this offset (and how far into that frag) */
1851 frag = skb_advance_to_frag(skb, offset, &frag_offset);
1856 struct skb_shared_info *info = skb_shinfo(skb);
1858 /* We read part of the last frag, must recvmsg() rest of skb. */
1859 if (frag == &info->frags[info->nr_frags - 1])
1862 /* Else, we must at least read the remainder in this frag. */
1863 partial_frag_remainder = skb_frag_size(frag) - frag_offset;
1864 zc->recv_skip_hint -= partial_frag_remainder;
1868 /* partial_frag_remainder: If part way through a frag, must read rest.
1869 * mappable_offset: Bytes till next mappable frag, *not* counting bytes
1870 * in partial_frag_remainder.
1872 mappable_offset = find_next_mappable_frag(frag, zc->recv_skip_hint);
1873 zc->recv_skip_hint = mappable_offset + partial_frag_remainder;
1876 static int tcp_recvmsg_locked(struct sock *sk, struct msghdr *msg, size_t len,
1877 int flags, struct scm_timestamping_internal *tss,
1879 static int receive_fallback_to_copy(struct sock *sk,
1880 struct tcp_zerocopy_receive *zc, int inq,
1881 struct scm_timestamping_internal *tss)
1883 unsigned long copy_address = (unsigned long)zc->copybuf_address;
1884 struct msghdr msg = {};
1889 zc->recv_skip_hint = 0;
1891 if (copy_address != zc->copybuf_address)
1894 err = import_single_range(ITER_DEST, (void __user *)copy_address,
1895 inq, &iov, &msg.msg_iter);
1899 err = tcp_recvmsg_locked(sk, &msg, inq, MSG_DONTWAIT,
1900 tss, &zc->msg_flags);
1904 zc->copybuf_len = err;
1905 if (likely(zc->copybuf_len)) {
1906 struct sk_buff *skb;
1909 skb = tcp_recv_skb(sk, tcp_sk(sk)->copied_seq, &offset);
1911 tcp_zerocopy_set_hint_for_skb(sk, zc, skb, offset);
1916 static int tcp_copy_straggler_data(struct tcp_zerocopy_receive *zc,
1917 struct sk_buff *skb, u32 copylen,
1918 u32 *offset, u32 *seq)
1920 unsigned long copy_address = (unsigned long)zc->copybuf_address;
1921 struct msghdr msg = {};
1925 if (copy_address != zc->copybuf_address)
1928 err = import_single_range(ITER_DEST, (void __user *)copy_address,
1929 copylen, &iov, &msg.msg_iter);
1932 err = skb_copy_datagram_msg(skb, *offset, &msg, copylen);
1935 zc->recv_skip_hint -= copylen;
1938 return (__s32)copylen;
1941 static int tcp_zc_handle_leftover(struct tcp_zerocopy_receive *zc,
1943 struct sk_buff *skb,
1946 struct scm_timestamping_internal *tss)
1948 u32 offset, copylen = min_t(u32, copybuf_len, zc->recv_skip_hint);
1952 /* skb is null if inq < PAGE_SIZE. */
1954 offset = *seq - TCP_SKB_CB(skb)->seq;
1956 skb = tcp_recv_skb(sk, *seq, &offset);
1957 if (TCP_SKB_CB(skb)->has_rxtstamp) {
1958 tcp_update_recv_tstamps(skb, tss);
1959 zc->msg_flags |= TCP_CMSG_TS;
1963 zc->copybuf_len = tcp_copy_straggler_data(zc, skb, copylen, &offset,
1965 return zc->copybuf_len < 0 ? 0 : copylen;
1968 static int tcp_zerocopy_vm_insert_batch_error(struct vm_area_struct *vma,
1969 struct page **pending_pages,
1970 unsigned long pages_remaining,
1971 unsigned long *address,
1974 struct tcp_zerocopy_receive *zc,
1975 u32 total_bytes_to_map,
1978 /* At least one page did not map. Try zapping if we skipped earlier. */
1979 if (err == -EBUSY &&
1980 zc->flags & TCP_RECEIVE_ZEROCOPY_FLAG_TLB_CLEAN_HINT) {
1983 maybe_zap_len = total_bytes_to_map - /* All bytes to map */
1984 *length + /* Mapped or pending */
1985 (pages_remaining * PAGE_SIZE); /* Failed map. */
1986 zap_page_range_single(vma, *address, maybe_zap_len, NULL);
1991 unsigned long leftover_pages = pages_remaining;
1994 /* We called zap_page_range_single, try to reinsert. */
1995 err = vm_insert_pages(vma, *address,
1998 bytes_mapped = PAGE_SIZE * (leftover_pages - pages_remaining);
1999 *seq += bytes_mapped;
2000 *address += bytes_mapped;
2003 /* Either we were unable to zap, OR we zapped, retried an
2004 * insert, and still had an issue. Either ways, pages_remaining
2005 * is the number of pages we were unable to map, and we unroll
2006 * some state we speculatively touched before.
2008 const int bytes_not_mapped = PAGE_SIZE * pages_remaining;
2010 *length -= bytes_not_mapped;
2011 zc->recv_skip_hint += bytes_not_mapped;
2016 static int tcp_zerocopy_vm_insert_batch(struct vm_area_struct *vma,
2017 struct page **pages,
2018 unsigned int pages_to_map,
2019 unsigned long *address,
2022 struct tcp_zerocopy_receive *zc,
2023 u32 total_bytes_to_map)
2025 unsigned long pages_remaining = pages_to_map;
2026 unsigned int pages_mapped;
2027 unsigned int bytes_mapped;
2030 err = vm_insert_pages(vma, *address, pages, &pages_remaining);
2031 pages_mapped = pages_to_map - (unsigned int)pages_remaining;
2032 bytes_mapped = PAGE_SIZE * pages_mapped;
2033 /* Even if vm_insert_pages fails, it may have partially succeeded in
2034 * mapping (some but not all of the pages).
2036 *seq += bytes_mapped;
2037 *address += bytes_mapped;
2042 /* Error: maybe zap and retry + rollback state for failed inserts. */
2043 return tcp_zerocopy_vm_insert_batch_error(vma, pages + pages_mapped,
2044 pages_remaining, address, length, seq, zc, total_bytes_to_map,
2048 #define TCP_VALID_ZC_MSG_FLAGS (TCP_CMSG_TS)
2049 static void tcp_zc_finalize_rx_tstamp(struct sock *sk,
2050 struct tcp_zerocopy_receive *zc,
2051 struct scm_timestamping_internal *tss)
2053 unsigned long msg_control_addr;
2054 struct msghdr cmsg_dummy;
2056 msg_control_addr = (unsigned long)zc->msg_control;
2057 cmsg_dummy.msg_control_user = (void __user *)msg_control_addr;
2058 cmsg_dummy.msg_controllen =
2059 (__kernel_size_t)zc->msg_controllen;
2060 cmsg_dummy.msg_flags = in_compat_syscall()
2061 ? MSG_CMSG_COMPAT : 0;
2062 cmsg_dummy.msg_control_is_user = true;
2064 if (zc->msg_control == msg_control_addr &&
2065 zc->msg_controllen == cmsg_dummy.msg_controllen) {
2066 tcp_recv_timestamp(&cmsg_dummy, sk, tss);
2067 zc->msg_control = (__u64)
2068 ((uintptr_t)cmsg_dummy.msg_control_user);
2069 zc->msg_controllen =
2070 (__u64)cmsg_dummy.msg_controllen;
2071 zc->msg_flags = (__u32)cmsg_dummy.msg_flags;
2075 #define TCP_ZEROCOPY_PAGE_BATCH_SIZE 32
2076 static int tcp_zerocopy_receive(struct sock *sk,
2077 struct tcp_zerocopy_receive *zc,
2078 struct scm_timestamping_internal *tss)
2080 u32 length = 0, offset, vma_len, avail_len, copylen = 0;
2081 unsigned long address = (unsigned long)zc->address;
2082 struct page *pages[TCP_ZEROCOPY_PAGE_BATCH_SIZE];
2083 s32 copybuf_len = zc->copybuf_len;
2084 struct tcp_sock *tp = tcp_sk(sk);
2085 const skb_frag_t *frags = NULL;
2086 unsigned int pages_to_map = 0;
2087 struct vm_area_struct *vma;
2088 struct sk_buff *skb = NULL;
2089 u32 seq = tp->copied_seq;
2090 u32 total_bytes_to_map;
2091 int inq = tcp_inq(sk);
2094 zc->copybuf_len = 0;
2097 if (address & (PAGE_SIZE - 1) || address != zc->address)
2100 if (sk->sk_state == TCP_LISTEN)
2103 sock_rps_record_flow(sk);
2105 if (inq && inq <= copybuf_len)
2106 return receive_fallback_to_copy(sk, zc, inq, tss);
2108 if (inq < PAGE_SIZE) {
2110 zc->recv_skip_hint = inq;
2111 if (!inq && sock_flag(sk, SOCK_DONE))
2116 mmap_read_lock(current->mm);
2118 vma = vma_lookup(current->mm, address);
2119 if (!vma || vma->vm_ops != &tcp_vm_ops) {
2120 mmap_read_unlock(current->mm);
2123 vma_len = min_t(unsigned long, zc->length, vma->vm_end - address);
2124 avail_len = min_t(u32, vma_len, inq);
2125 total_bytes_to_map = avail_len & ~(PAGE_SIZE - 1);
2126 if (total_bytes_to_map) {
2127 if (!(zc->flags & TCP_RECEIVE_ZEROCOPY_FLAG_TLB_CLEAN_HINT))
2128 zap_page_range_single(vma, address, total_bytes_to_map,
2130 zc->length = total_bytes_to_map;
2131 zc->recv_skip_hint = 0;
2133 zc->length = avail_len;
2134 zc->recv_skip_hint = avail_len;
2137 while (length + PAGE_SIZE <= zc->length) {
2138 int mappable_offset;
2141 if (zc->recv_skip_hint < PAGE_SIZE) {
2145 if (zc->recv_skip_hint > 0)
2148 offset = seq - TCP_SKB_CB(skb)->seq;
2150 skb = tcp_recv_skb(sk, seq, &offset);
2153 if (TCP_SKB_CB(skb)->has_rxtstamp) {
2154 tcp_update_recv_tstamps(skb, tss);
2155 zc->msg_flags |= TCP_CMSG_TS;
2157 zc->recv_skip_hint = skb->len - offset;
2158 frags = skb_advance_to_frag(skb, offset, &offset_frag);
2159 if (!frags || offset_frag)
2163 mappable_offset = find_next_mappable_frag(frags,
2164 zc->recv_skip_hint);
2165 if (mappable_offset) {
2166 zc->recv_skip_hint = mappable_offset;
2169 page = skb_frag_page(frags);
2171 pages[pages_to_map++] = page;
2172 length += PAGE_SIZE;
2173 zc->recv_skip_hint -= PAGE_SIZE;
2175 if (pages_to_map == TCP_ZEROCOPY_PAGE_BATCH_SIZE ||
2176 zc->recv_skip_hint < PAGE_SIZE) {
2177 /* Either full batch, or we're about to go to next skb
2178 * (and we cannot unroll failed ops across skbs).
2180 ret = tcp_zerocopy_vm_insert_batch(vma, pages,
2184 total_bytes_to_map);
2191 ret = tcp_zerocopy_vm_insert_batch(vma, pages, pages_to_map,
2192 &address, &length, &seq,
2193 zc, total_bytes_to_map);
2196 mmap_read_unlock(current->mm);
2197 /* Try to copy straggler data. */
2199 copylen = tcp_zc_handle_leftover(zc, sk, skb, &seq, copybuf_len, tss);
2201 if (length + copylen) {
2202 WRITE_ONCE(tp->copied_seq, seq);
2203 tcp_rcv_space_adjust(sk);
2205 /* Clean up data we have read: This will do ACK frames. */
2206 tcp_recv_skb(sk, seq, &offset);
2207 tcp_cleanup_rbuf(sk, length + copylen);
2209 if (length == zc->length)
2210 zc->recv_skip_hint = 0;
2212 if (!zc->recv_skip_hint && sock_flag(sk, SOCK_DONE))
2215 zc->length = length;
2220 /* Similar to __sock_recv_timestamp, but does not require an skb */
2221 void tcp_recv_timestamp(struct msghdr *msg, const struct sock *sk,
2222 struct scm_timestamping_internal *tss)
2224 int new_tstamp = sock_flag(sk, SOCK_TSTAMP_NEW);
2225 bool has_timestamping = false;
2227 if (tss->ts[0].tv_sec || tss->ts[0].tv_nsec) {
2228 if (sock_flag(sk, SOCK_RCVTSTAMP)) {
2229 if (sock_flag(sk, SOCK_RCVTSTAMPNS)) {
2231 struct __kernel_timespec kts = {
2232 .tv_sec = tss->ts[0].tv_sec,
2233 .tv_nsec = tss->ts[0].tv_nsec,
2235 put_cmsg(msg, SOL_SOCKET, SO_TIMESTAMPNS_NEW,
2238 struct __kernel_old_timespec ts_old = {
2239 .tv_sec = tss->ts[0].tv_sec,
2240 .tv_nsec = tss->ts[0].tv_nsec,
2242 put_cmsg(msg, SOL_SOCKET, SO_TIMESTAMPNS_OLD,
2243 sizeof(ts_old), &ts_old);
2247 struct __kernel_sock_timeval stv = {
2248 .tv_sec = tss->ts[0].tv_sec,
2249 .tv_usec = tss->ts[0].tv_nsec / 1000,
2251 put_cmsg(msg, SOL_SOCKET, SO_TIMESTAMP_NEW,
2254 struct __kernel_old_timeval tv = {
2255 .tv_sec = tss->ts[0].tv_sec,
2256 .tv_usec = tss->ts[0].tv_nsec / 1000,
2258 put_cmsg(msg, SOL_SOCKET, SO_TIMESTAMP_OLD,
2264 if (sk->sk_tsflags & SOF_TIMESTAMPING_SOFTWARE)
2265 has_timestamping = true;
2267 tss->ts[0] = (struct timespec64) {0};
2270 if (tss->ts[2].tv_sec || tss->ts[2].tv_nsec) {
2271 if (sk->sk_tsflags & SOF_TIMESTAMPING_RAW_HARDWARE)
2272 has_timestamping = true;
2274 tss->ts[2] = (struct timespec64) {0};
2277 if (has_timestamping) {
2278 tss->ts[1] = (struct timespec64) {0};
2279 if (sock_flag(sk, SOCK_TSTAMP_NEW))
2280 put_cmsg_scm_timestamping64(msg, tss);
2282 put_cmsg_scm_timestamping(msg, tss);
2286 static int tcp_inq_hint(struct sock *sk)
2288 const struct tcp_sock *tp = tcp_sk(sk);
2289 u32 copied_seq = READ_ONCE(tp->copied_seq);
2290 u32 rcv_nxt = READ_ONCE(tp->rcv_nxt);
2293 inq = rcv_nxt - copied_seq;
2294 if (unlikely(inq < 0 || copied_seq != READ_ONCE(tp->copied_seq))) {
2296 inq = tp->rcv_nxt - tp->copied_seq;
2299 /* After receiving a FIN, tell the user-space to continue reading
2300 * by returning a non-zero inq.
2302 if (inq == 0 && sock_flag(sk, SOCK_DONE))
2308 * This routine copies from a sock struct into the user buffer.
2310 * Technical note: in 2.3 we work on _locked_ socket, so that
2311 * tricks with *seq access order and skb->users are not required.
2312 * Probably, code can be easily improved even more.
2315 static int tcp_recvmsg_locked(struct sock *sk, struct msghdr *msg, size_t len,
2316 int flags, struct scm_timestamping_internal *tss,
2319 struct tcp_sock *tp = tcp_sk(sk);
2325 int target; /* Read at least this many bytes */
2327 struct sk_buff *skb, *last;
2331 if (sk->sk_state == TCP_LISTEN)
2334 if (tp->recvmsg_inq) {
2335 *cmsg_flags = TCP_CMSG_INQ;
2336 msg->msg_get_inq = 1;
2338 timeo = sock_rcvtimeo(sk, flags & MSG_DONTWAIT);
2340 /* Urgent data needs to be handled specially. */
2341 if (flags & MSG_OOB)
2344 if (unlikely(tp->repair)) {
2346 if (!(flags & MSG_PEEK))
2349 if (tp->repair_queue == TCP_SEND_QUEUE)
2353 if (tp->repair_queue == TCP_NO_QUEUE)
2356 /* 'common' recv queue MSG_PEEK-ing */
2359 seq = &tp->copied_seq;
2360 if (flags & MSG_PEEK) {
2361 peek_seq = tp->copied_seq;
2365 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
2370 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
2371 if (unlikely(tp->urg_data) && tp->urg_seq == *seq) {
2374 if (signal_pending(current)) {
2375 copied = timeo ? sock_intr_errno(timeo) : -EAGAIN;
2380 /* Next get a buffer. */
2382 last = skb_peek_tail(&sk->sk_receive_queue);
2383 skb_queue_walk(&sk->sk_receive_queue, skb) {
2385 /* Now that we have two receive queues this
2388 if (WARN(before(*seq, TCP_SKB_CB(skb)->seq),
2389 "TCP recvmsg seq # bug: copied %X, seq %X, rcvnxt %X, fl %X\n",
2390 *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt,
2394 offset = *seq - TCP_SKB_CB(skb)->seq;
2395 if (unlikely(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
2396 pr_err_once("%s: found a SYN, please report !\n", __func__);
2399 if (offset < skb->len)
2401 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
2403 WARN(!(flags & MSG_PEEK),
2404 "TCP recvmsg seq # bug 2: copied %X, seq %X, rcvnxt %X, fl %X\n",
2405 *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt, flags);
2408 /* Well, if we have backlog, try to process it now yet. */
2410 if (copied >= target && !READ_ONCE(sk->sk_backlog.tail))
2416 sk->sk_state == TCP_CLOSE ||
2417 (sk->sk_shutdown & RCV_SHUTDOWN) ||
2418 signal_pending(current))
2421 if (sock_flag(sk, SOCK_DONE))
2425 copied = sock_error(sk);
2429 if (sk->sk_shutdown & RCV_SHUTDOWN)
2432 if (sk->sk_state == TCP_CLOSE) {
2433 /* This occurs when user tries to read
2434 * from never connected socket.
2445 if (signal_pending(current)) {
2446 copied = sock_intr_errno(timeo);
2451 if (copied >= target) {
2452 /* Do not sleep, just process backlog. */
2453 __sk_flush_backlog(sk);
2455 tcp_cleanup_rbuf(sk, copied);
2456 sk_wait_data(sk, &timeo, last);
2459 if ((flags & MSG_PEEK) &&
2460 (peek_seq - copied - urg_hole != tp->copied_seq)) {
2461 net_dbg_ratelimited("TCP(%s:%d): Application bug, race in MSG_PEEK\n",
2463 task_pid_nr(current));
2464 peek_seq = tp->copied_seq;
2469 /* Ok so how much can we use? */
2470 used = skb->len - offset;
2474 /* Do we have urgent data here? */
2475 if (unlikely(tp->urg_data)) {
2476 u32 urg_offset = tp->urg_seq - *seq;
2477 if (urg_offset < used) {
2479 if (!sock_flag(sk, SOCK_URGINLINE)) {
2480 WRITE_ONCE(*seq, *seq + 1);
2492 if (!(flags & MSG_TRUNC)) {
2493 err = skb_copy_datagram_msg(skb, offset, msg, used);
2495 /* Exception. Bailout! */
2502 WRITE_ONCE(*seq, *seq + used);
2506 tcp_rcv_space_adjust(sk);
2509 if (unlikely(tp->urg_data) && after(tp->copied_seq, tp->urg_seq)) {
2510 WRITE_ONCE(tp->urg_data, 0);
2511 tcp_fast_path_check(sk);
2514 if (TCP_SKB_CB(skb)->has_rxtstamp) {
2515 tcp_update_recv_tstamps(skb, tss);
2516 *cmsg_flags |= TCP_CMSG_TS;
2519 if (used + offset < skb->len)
2522 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
2524 if (!(flags & MSG_PEEK))
2525 tcp_eat_recv_skb(sk, skb);
2529 /* Process the FIN. */
2530 WRITE_ONCE(*seq, *seq + 1);
2531 if (!(flags & MSG_PEEK))
2532 tcp_eat_recv_skb(sk, skb);
2536 /* According to UNIX98, msg_name/msg_namelen are ignored
2537 * on connected socket. I was just happy when found this 8) --ANK
2540 /* Clean up data we have read: This will do ACK frames. */
2541 tcp_cleanup_rbuf(sk, copied);
2548 err = tcp_recv_urg(sk, msg, len, flags);
2552 err = tcp_peek_sndq(sk, msg, len);
2556 int tcp_recvmsg(struct sock *sk, struct msghdr *msg, size_t len, int flags,
2559 int cmsg_flags = 0, ret;
2560 struct scm_timestamping_internal tss;
2562 if (unlikely(flags & MSG_ERRQUEUE))
2563 return inet_recv_error(sk, msg, len, addr_len);
2565 if (sk_can_busy_loop(sk) &&
2566 skb_queue_empty_lockless(&sk->sk_receive_queue) &&
2567 sk->sk_state == TCP_ESTABLISHED)
2568 sk_busy_loop(sk, flags & MSG_DONTWAIT);
2571 ret = tcp_recvmsg_locked(sk, msg, len, flags, &tss, &cmsg_flags);
2574 if ((cmsg_flags || msg->msg_get_inq) && ret >= 0) {
2575 if (cmsg_flags & TCP_CMSG_TS)
2576 tcp_recv_timestamp(msg, sk, &tss);
2577 if (msg->msg_get_inq) {
2578 msg->msg_inq = tcp_inq_hint(sk);
2579 if (cmsg_flags & TCP_CMSG_INQ)
2580 put_cmsg(msg, SOL_TCP, TCP_CM_INQ,
2581 sizeof(msg->msg_inq), &msg->msg_inq);
2586 EXPORT_SYMBOL(tcp_recvmsg);
2588 void tcp_set_state(struct sock *sk, int state)
2590 int oldstate = sk->sk_state;
2592 /* We defined a new enum for TCP states that are exported in BPF
2593 * so as not force the internal TCP states to be frozen. The
2594 * following checks will detect if an internal state value ever
2595 * differs from the BPF value. If this ever happens, then we will
2596 * need to remap the internal value to the BPF value before calling
2597 * tcp_call_bpf_2arg.
2599 BUILD_BUG_ON((int)BPF_TCP_ESTABLISHED != (int)TCP_ESTABLISHED);
2600 BUILD_BUG_ON((int)BPF_TCP_SYN_SENT != (int)TCP_SYN_SENT);
2601 BUILD_BUG_ON((int)BPF_TCP_SYN_RECV != (int)TCP_SYN_RECV);
2602 BUILD_BUG_ON((int)BPF_TCP_FIN_WAIT1 != (int)TCP_FIN_WAIT1);
2603 BUILD_BUG_ON((int)BPF_TCP_FIN_WAIT2 != (int)TCP_FIN_WAIT2);
2604 BUILD_BUG_ON((int)BPF_TCP_TIME_WAIT != (int)TCP_TIME_WAIT);
2605 BUILD_BUG_ON((int)BPF_TCP_CLOSE != (int)TCP_CLOSE);
2606 BUILD_BUG_ON((int)BPF_TCP_CLOSE_WAIT != (int)TCP_CLOSE_WAIT);
2607 BUILD_BUG_ON((int)BPF_TCP_LAST_ACK != (int)TCP_LAST_ACK);
2608 BUILD_BUG_ON((int)BPF_TCP_LISTEN != (int)TCP_LISTEN);
2609 BUILD_BUG_ON((int)BPF_TCP_CLOSING != (int)TCP_CLOSING);
2610 BUILD_BUG_ON((int)BPF_TCP_NEW_SYN_RECV != (int)TCP_NEW_SYN_RECV);
2611 BUILD_BUG_ON((int)BPF_TCP_MAX_STATES != (int)TCP_MAX_STATES);
2613 /* bpf uapi header bpf.h defines an anonymous enum with values
2614 * BPF_TCP_* used by bpf programs. Currently gcc built vmlinux
2615 * is able to emit this enum in DWARF due to the above BUILD_BUG_ON.
2616 * But clang built vmlinux does not have this enum in DWARF
2617 * since clang removes the above code before generating IR/debuginfo.
2618 * Let us explicitly emit the type debuginfo to ensure the
2619 * above-mentioned anonymous enum in the vmlinux DWARF and hence BTF
2620 * regardless of which compiler is used.
2622 BTF_TYPE_EMIT_ENUM(BPF_TCP_ESTABLISHED);
2624 if (BPF_SOCK_OPS_TEST_FLAG(tcp_sk(sk), BPF_SOCK_OPS_STATE_CB_FLAG))
2625 tcp_call_bpf_2arg(sk, BPF_SOCK_OPS_STATE_CB, oldstate, state);
2628 case TCP_ESTABLISHED:
2629 if (oldstate != TCP_ESTABLISHED)
2630 TCP_INC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
2634 if (oldstate == TCP_CLOSE_WAIT || oldstate == TCP_ESTABLISHED)
2635 TCP_INC_STATS(sock_net(sk), TCP_MIB_ESTABRESETS);
2637 sk->sk_prot->unhash(sk);
2638 if (inet_csk(sk)->icsk_bind_hash &&
2639 !(sk->sk_userlocks & SOCK_BINDPORT_LOCK))
2643 if (oldstate == TCP_ESTABLISHED)
2644 TCP_DEC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
2647 /* Change state AFTER socket is unhashed to avoid closed
2648 * socket sitting in hash tables.
2650 inet_sk_state_store(sk, state);
2652 EXPORT_SYMBOL_GPL(tcp_set_state);
2655 * State processing on a close. This implements the state shift for
2656 * sending our FIN frame. Note that we only send a FIN for some
2657 * states. A shutdown() may have already sent the FIN, or we may be
2661 static const unsigned char new_state[16] = {
2662 /* current state: new state: action: */
2663 [0 /* (Invalid) */] = TCP_CLOSE,
2664 [TCP_ESTABLISHED] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2665 [TCP_SYN_SENT] = TCP_CLOSE,
2666 [TCP_SYN_RECV] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2667 [TCP_FIN_WAIT1] = TCP_FIN_WAIT1,
2668 [TCP_FIN_WAIT2] = TCP_FIN_WAIT2,
2669 [TCP_TIME_WAIT] = TCP_CLOSE,
2670 [TCP_CLOSE] = TCP_CLOSE,
2671 [TCP_CLOSE_WAIT] = TCP_LAST_ACK | TCP_ACTION_FIN,
2672 [TCP_LAST_ACK] = TCP_LAST_ACK,
2673 [TCP_LISTEN] = TCP_CLOSE,
2674 [TCP_CLOSING] = TCP_CLOSING,
2675 [TCP_NEW_SYN_RECV] = TCP_CLOSE, /* should not happen ! */
2678 static int tcp_close_state(struct sock *sk)
2680 int next = (int)new_state[sk->sk_state];
2681 int ns = next & TCP_STATE_MASK;
2683 tcp_set_state(sk, ns);
2685 return next & TCP_ACTION_FIN;
2689 * Shutdown the sending side of a connection. Much like close except
2690 * that we don't receive shut down or sock_set_flag(sk, SOCK_DEAD).
2693 void tcp_shutdown(struct sock *sk, int how)
2695 /* We need to grab some memory, and put together a FIN,
2696 * and then put it into the queue to be sent.
2697 * Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
2699 if (!(how & SEND_SHUTDOWN))
2702 /* If we've already sent a FIN, or it's a closed state, skip this. */
2703 if ((1 << sk->sk_state) &
2704 (TCPF_ESTABLISHED | TCPF_SYN_SENT |
2705 TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) {
2706 /* Clear out any half completed packets. FIN if needed. */
2707 if (tcp_close_state(sk))
2711 EXPORT_SYMBOL(tcp_shutdown);
2713 int tcp_orphan_count_sum(void)
2717 for_each_possible_cpu(i)
2718 total += per_cpu(tcp_orphan_count, i);
2720 return max(total, 0);
2723 static int tcp_orphan_cache;
2724 static struct timer_list tcp_orphan_timer;
2725 #define TCP_ORPHAN_TIMER_PERIOD msecs_to_jiffies(100)
2727 static void tcp_orphan_update(struct timer_list *unused)
2729 WRITE_ONCE(tcp_orphan_cache, tcp_orphan_count_sum());
2730 mod_timer(&tcp_orphan_timer, jiffies + TCP_ORPHAN_TIMER_PERIOD);
2733 static bool tcp_too_many_orphans(int shift)
2735 return READ_ONCE(tcp_orphan_cache) << shift >
2736 READ_ONCE(sysctl_tcp_max_orphans);
2739 bool tcp_check_oom(struct sock *sk, int shift)
2741 bool too_many_orphans, out_of_socket_memory;
2743 too_many_orphans = tcp_too_many_orphans(shift);
2744 out_of_socket_memory = tcp_out_of_memory(sk);
2746 if (too_many_orphans)
2747 net_info_ratelimited("too many orphaned sockets\n");
2748 if (out_of_socket_memory)
2749 net_info_ratelimited("out of memory -- consider tuning tcp_mem\n");
2750 return too_many_orphans || out_of_socket_memory;
2753 void __tcp_close(struct sock *sk, long timeout)
2755 struct sk_buff *skb;
2756 int data_was_unread = 0;
2759 WRITE_ONCE(sk->sk_shutdown, SHUTDOWN_MASK);
2761 if (sk->sk_state == TCP_LISTEN) {
2762 tcp_set_state(sk, TCP_CLOSE);
2765 inet_csk_listen_stop(sk);
2767 goto adjudge_to_death;
2770 /* We need to flush the recv. buffs. We do this only on the
2771 * descriptor close, not protocol-sourced closes, because the
2772 * reader process may not have drained the data yet!
2774 while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) {
2775 u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq;
2777 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
2779 data_was_unread += len;
2783 /* If socket has been already reset (e.g. in tcp_reset()) - kill it. */
2784 if (sk->sk_state == TCP_CLOSE)
2785 goto adjudge_to_death;
2787 /* As outlined in RFC 2525, section 2.17, we send a RST here because
2788 * data was lost. To witness the awful effects of the old behavior of
2789 * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk
2790 * GET in an FTP client, suspend the process, wait for the client to
2791 * advertise a zero window, then kill -9 the FTP client, wheee...
2792 * Note: timeout is always zero in such a case.
2794 if (unlikely(tcp_sk(sk)->repair)) {
2795 sk->sk_prot->disconnect(sk, 0);
2796 } else if (data_was_unread) {
2797 /* Unread data was tossed, zap the connection. */
2798 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONCLOSE);
2799 tcp_set_state(sk, TCP_CLOSE);
2800 tcp_send_active_reset(sk, sk->sk_allocation);
2801 } else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) {
2802 /* Check zero linger _after_ checking for unread data. */
2803 sk->sk_prot->disconnect(sk, 0);
2804 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
2805 } else if (tcp_close_state(sk)) {
2806 /* We FIN if the application ate all the data before
2807 * zapping the connection.
2810 /* RED-PEN. Formally speaking, we have broken TCP state
2811 * machine. State transitions:
2813 * TCP_ESTABLISHED -> TCP_FIN_WAIT1
2814 * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible)
2815 * TCP_CLOSE_WAIT -> TCP_LAST_ACK
2817 * are legal only when FIN has been sent (i.e. in window),
2818 * rather than queued out of window. Purists blame.
2820 * F.e. "RFC state" is ESTABLISHED,
2821 * if Linux state is FIN-WAIT-1, but FIN is still not sent.
2823 * The visible declinations are that sometimes
2824 * we enter time-wait state, when it is not required really
2825 * (harmless), do not send active resets, when they are
2826 * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
2827 * they look as CLOSING or LAST_ACK for Linux)
2828 * Probably, I missed some more holelets.
2830 * XXX (TFO) - To start off we don't support SYN+ACK+FIN
2831 * in a single packet! (May consider it later but will
2832 * probably need API support or TCP_CORK SYN-ACK until
2833 * data is written and socket is closed.)
2838 sk_stream_wait_close(sk, timeout);
2841 state = sk->sk_state;
2847 /* remove backlog if any, without releasing ownership. */
2850 this_cpu_inc(tcp_orphan_count);
2852 /* Have we already been destroyed by a softirq or backlog? */
2853 if (state != TCP_CLOSE && sk->sk_state == TCP_CLOSE)
2856 /* This is a (useful) BSD violating of the RFC. There is a
2857 * problem with TCP as specified in that the other end could
2858 * keep a socket open forever with no application left this end.
2859 * We use a 1 minute timeout (about the same as BSD) then kill
2860 * our end. If they send after that then tough - BUT: long enough
2861 * that we won't make the old 4*rto = almost no time - whoops
2864 * Nope, it was not mistake. It is really desired behaviour
2865 * f.e. on http servers, when such sockets are useless, but
2866 * consume significant resources. Let's do it with special
2867 * linger2 option. --ANK
2870 if (sk->sk_state == TCP_FIN_WAIT2) {
2871 struct tcp_sock *tp = tcp_sk(sk);
2872 if (tp->linger2 < 0) {
2873 tcp_set_state(sk, TCP_CLOSE);
2874 tcp_send_active_reset(sk, GFP_ATOMIC);
2875 __NET_INC_STATS(sock_net(sk),
2876 LINUX_MIB_TCPABORTONLINGER);
2878 const int tmo = tcp_fin_time(sk);
2880 if (tmo > TCP_TIMEWAIT_LEN) {
2881 inet_csk_reset_keepalive_timer(sk,
2882 tmo - TCP_TIMEWAIT_LEN);
2884 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
2889 if (sk->sk_state != TCP_CLOSE) {
2890 if (tcp_check_oom(sk, 0)) {
2891 tcp_set_state(sk, TCP_CLOSE);
2892 tcp_send_active_reset(sk, GFP_ATOMIC);
2893 __NET_INC_STATS(sock_net(sk),
2894 LINUX_MIB_TCPABORTONMEMORY);
2895 } else if (!check_net(sock_net(sk))) {
2896 /* Not possible to send reset; just close */
2897 tcp_set_state(sk, TCP_CLOSE);
2901 if (sk->sk_state == TCP_CLOSE) {
2902 struct request_sock *req;
2904 req = rcu_dereference_protected(tcp_sk(sk)->fastopen_rsk,
2905 lockdep_sock_is_held(sk));
2906 /* We could get here with a non-NULL req if the socket is
2907 * aborted (e.g., closed with unread data) before 3WHS
2911 reqsk_fastopen_remove(sk, req, false);
2912 inet_csk_destroy_sock(sk);
2914 /* Otherwise, socket is reprieved until protocol close. */
2921 void tcp_close(struct sock *sk, long timeout)
2924 __tcp_close(sk, timeout);
2928 EXPORT_SYMBOL(tcp_close);
2930 /* These states need RST on ABORT according to RFC793 */
2932 static inline bool tcp_need_reset(int state)
2934 return (1 << state) &
2935 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 |
2936 TCPF_FIN_WAIT2 | TCPF_SYN_RECV);
2939 static void tcp_rtx_queue_purge(struct sock *sk)
2941 struct rb_node *p = rb_first(&sk->tcp_rtx_queue);
2943 tcp_sk(sk)->highest_sack = NULL;
2945 struct sk_buff *skb = rb_to_skb(p);
2948 /* Since we are deleting whole queue, no need to
2949 * list_del(&skb->tcp_tsorted_anchor)
2951 tcp_rtx_queue_unlink(skb, sk);
2952 tcp_wmem_free_skb(sk, skb);
2956 void tcp_write_queue_purge(struct sock *sk)
2958 struct sk_buff *skb;
2960 tcp_chrono_stop(sk, TCP_CHRONO_BUSY);
2961 while ((skb = __skb_dequeue(&sk->sk_write_queue)) != NULL) {
2962 tcp_skb_tsorted_anchor_cleanup(skb);
2963 tcp_wmem_free_skb(sk, skb);
2965 tcp_rtx_queue_purge(sk);
2966 INIT_LIST_HEAD(&tcp_sk(sk)->tsorted_sent_queue);
2967 tcp_clear_all_retrans_hints(tcp_sk(sk));
2968 tcp_sk(sk)->packets_out = 0;
2969 inet_csk(sk)->icsk_backoff = 0;
2972 int tcp_disconnect(struct sock *sk, int flags)
2974 struct inet_sock *inet = inet_sk(sk);
2975 struct inet_connection_sock *icsk = inet_csk(sk);
2976 struct tcp_sock *tp = tcp_sk(sk);
2977 int old_state = sk->sk_state;
2980 /* Deny disconnect if other threads are blocked in sk_wait_event()
2981 * or inet_wait_for_connect().
2983 if (sk->sk_wait_pending)
2986 if (old_state != TCP_CLOSE)
2987 tcp_set_state(sk, TCP_CLOSE);
2989 /* ABORT function of RFC793 */
2990 if (old_state == TCP_LISTEN) {
2991 inet_csk_listen_stop(sk);
2992 } else if (unlikely(tp->repair)) {
2993 WRITE_ONCE(sk->sk_err, ECONNABORTED);
2994 } else if (tcp_need_reset(old_state) ||
2995 (tp->snd_nxt != tp->write_seq &&
2996 (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK))) {
2997 /* The last check adjusts for discrepancy of Linux wrt. RFC
3000 tcp_send_active_reset(sk, gfp_any());
3001 WRITE_ONCE(sk->sk_err, ECONNRESET);
3002 } else if (old_state == TCP_SYN_SENT)
3003 WRITE_ONCE(sk->sk_err, ECONNRESET);
3005 tcp_clear_xmit_timers(sk);
3006 __skb_queue_purge(&sk->sk_receive_queue);
3007 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt);
3008 WRITE_ONCE(tp->urg_data, 0);
3009 tcp_write_queue_purge(sk);
3010 tcp_fastopen_active_disable_ofo_check(sk);
3011 skb_rbtree_purge(&tp->out_of_order_queue);
3013 inet->inet_dport = 0;
3015 inet_bhash2_reset_saddr(sk);
3017 WRITE_ONCE(sk->sk_shutdown, 0);
3018 sock_reset_flag(sk, SOCK_DONE);
3020 tp->mdev_us = jiffies_to_usecs(TCP_TIMEOUT_INIT);
3021 tp->rcv_rtt_last_tsecr = 0;
3023 seq = tp->write_seq + tp->max_window + 2;
3026 WRITE_ONCE(tp->write_seq, seq);
3028 icsk->icsk_backoff = 0;
3029 icsk->icsk_probes_out = 0;
3030 icsk->icsk_probes_tstamp = 0;
3031 icsk->icsk_rto = TCP_TIMEOUT_INIT;
3032 icsk->icsk_rto_min = TCP_RTO_MIN;
3033 icsk->icsk_delack_max = TCP_DELACK_MAX;
3034 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
3035 tcp_snd_cwnd_set(tp, TCP_INIT_CWND);
3036 tp->snd_cwnd_cnt = 0;
3037 tp->is_cwnd_limited = 0;
3038 tp->max_packets_out = 0;
3039 tp->window_clamp = 0;
3041 tp->delivered_ce = 0;
3042 if (icsk->icsk_ca_ops->release)
3043 icsk->icsk_ca_ops->release(sk);
3044 memset(icsk->icsk_ca_priv, 0, sizeof(icsk->icsk_ca_priv));
3045 icsk->icsk_ca_initialized = 0;
3046 tcp_set_ca_state(sk, TCP_CA_Open);
3047 tp->is_sack_reneg = 0;
3048 tcp_clear_retrans(tp);
3049 tp->total_retrans = 0;
3050 inet_csk_delack_init(sk);
3051 /* Initialize rcv_mss to TCP_MIN_MSS to avoid division by 0
3052 * issue in __tcp_select_window()
3054 icsk->icsk_ack.rcv_mss = TCP_MIN_MSS;
3055 memset(&tp->rx_opt, 0, sizeof(tp->rx_opt));
3057 dst_release(xchg((__force struct dst_entry **)&sk->sk_rx_dst, NULL));
3058 tcp_saved_syn_free(tp);
3059 tp->compressed_ack = 0;
3063 tp->bytes_acked = 0;
3064 tp->bytes_received = 0;
3065 tp->bytes_retrans = 0;
3066 tp->data_segs_in = 0;
3067 tp->data_segs_out = 0;
3068 tp->duplicate_sack[0].start_seq = 0;
3069 tp->duplicate_sack[0].end_seq = 0;
3072 tp->retrans_out = 0;
3074 tp->tlp_high_seq = 0;
3075 tp->last_oow_ack_time = 0;
3077 /* There's a bubble in the pipe until at least the first ACK. */
3078 tp->app_limited = ~0U;
3079 tp->rate_app_limited = 1;
3080 tp->rack.mstamp = 0;
3081 tp->rack.advanced = 0;
3082 tp->rack.reo_wnd_steps = 1;
3083 tp->rack.last_delivered = 0;
3084 tp->rack.reo_wnd_persist = 0;
3085 tp->rack.dsack_seen = 0;
3086 tp->syn_data_acked = 0;
3087 tp->rx_opt.saw_tstamp = 0;
3088 tp->rx_opt.dsack = 0;
3089 tp->rx_opt.num_sacks = 0;
3090 tp->rcv_ooopack = 0;
3093 /* Clean up fastopen related fields */
3094 tcp_free_fastopen_req(tp);
3095 inet->defer_connect = 0;
3096 tp->fastopen_client_fail = 0;
3098 WARN_ON(inet->inet_num && !icsk->icsk_bind_hash);
3100 if (sk->sk_frag.page) {
3101 put_page(sk->sk_frag.page);
3102 sk->sk_frag.page = NULL;
3103 sk->sk_frag.offset = 0;
3105 sk_error_report(sk);
3108 EXPORT_SYMBOL(tcp_disconnect);
3110 static inline bool tcp_can_repair_sock(const struct sock *sk)
3112 return sockopt_ns_capable(sock_net(sk)->user_ns, CAP_NET_ADMIN) &&
3113 (sk->sk_state != TCP_LISTEN);
3116 static int tcp_repair_set_window(struct tcp_sock *tp, sockptr_t optbuf, int len)
3118 struct tcp_repair_window opt;
3123 if (len != sizeof(opt))
3126 if (copy_from_sockptr(&opt, optbuf, sizeof(opt)))
3129 if (opt.max_window < opt.snd_wnd)
3132 if (after(opt.snd_wl1, tp->rcv_nxt + opt.rcv_wnd))
3135 if (after(opt.rcv_wup, tp->rcv_nxt))
3138 tp->snd_wl1 = opt.snd_wl1;
3139 tp->snd_wnd = opt.snd_wnd;
3140 tp->max_window = opt.max_window;
3142 tp->rcv_wnd = opt.rcv_wnd;
3143 tp->rcv_wup = opt.rcv_wup;
3148 static int tcp_repair_options_est(struct sock *sk, sockptr_t optbuf,
3151 struct tcp_sock *tp = tcp_sk(sk);
3152 struct tcp_repair_opt opt;
3155 while (len >= sizeof(opt)) {
3156 if (copy_from_sockptr_offset(&opt, optbuf, offset, sizeof(opt)))
3159 offset += sizeof(opt);
3162 switch (opt.opt_code) {
3164 tp->rx_opt.mss_clamp = opt.opt_val;
3169 u16 snd_wscale = opt.opt_val & 0xFFFF;
3170 u16 rcv_wscale = opt.opt_val >> 16;
3172 if (snd_wscale > TCP_MAX_WSCALE || rcv_wscale > TCP_MAX_WSCALE)
3175 tp->rx_opt.snd_wscale = snd_wscale;
3176 tp->rx_opt.rcv_wscale = rcv_wscale;
3177 tp->rx_opt.wscale_ok = 1;
3180 case TCPOPT_SACK_PERM:
3181 if (opt.opt_val != 0)
3184 tp->rx_opt.sack_ok |= TCP_SACK_SEEN;
3186 case TCPOPT_TIMESTAMP:
3187 if (opt.opt_val != 0)
3190 tp->rx_opt.tstamp_ok = 1;
3198 DEFINE_STATIC_KEY_FALSE(tcp_tx_delay_enabled);
3199 EXPORT_SYMBOL(tcp_tx_delay_enabled);
3201 static void tcp_enable_tx_delay(void)
3203 if (!static_branch_unlikely(&tcp_tx_delay_enabled)) {
3204 static int __tcp_tx_delay_enabled = 0;
3206 if (cmpxchg(&__tcp_tx_delay_enabled, 0, 1) == 0) {
3207 static_branch_enable(&tcp_tx_delay_enabled);
3208 pr_info("TCP_TX_DELAY enabled\n");
3213 /* When set indicates to always queue non-full frames. Later the user clears
3214 * this option and we transmit any pending partial frames in the queue. This is
3215 * meant to be used alongside sendfile() to get properly filled frames when the
3216 * user (for example) must write out headers with a write() call first and then
3217 * use sendfile to send out the data parts.
3219 * TCP_CORK can be set together with TCP_NODELAY and it is stronger than
3222 void __tcp_sock_set_cork(struct sock *sk, bool on)
3224 struct tcp_sock *tp = tcp_sk(sk);
3227 tp->nonagle |= TCP_NAGLE_CORK;
3229 tp->nonagle &= ~TCP_NAGLE_CORK;
3230 if (tp->nonagle & TCP_NAGLE_OFF)
3231 tp->nonagle |= TCP_NAGLE_PUSH;
3232 tcp_push_pending_frames(sk);
3236 void tcp_sock_set_cork(struct sock *sk, bool on)
3239 __tcp_sock_set_cork(sk, on);
3242 EXPORT_SYMBOL(tcp_sock_set_cork);
3244 /* TCP_NODELAY is weaker than TCP_CORK, so that this option on corked socket is
3245 * remembered, but it is not activated until cork is cleared.
3247 * However, when TCP_NODELAY is set we make an explicit push, which overrides
3248 * even TCP_CORK for currently queued segments.
3250 void __tcp_sock_set_nodelay(struct sock *sk, bool on)
3253 tcp_sk(sk)->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH;
3254 tcp_push_pending_frames(sk);
3256 tcp_sk(sk)->nonagle &= ~TCP_NAGLE_OFF;
3260 void tcp_sock_set_nodelay(struct sock *sk)
3263 __tcp_sock_set_nodelay(sk, true);
3266 EXPORT_SYMBOL(tcp_sock_set_nodelay);
3268 static void __tcp_sock_set_quickack(struct sock *sk, int val)
3271 inet_csk_enter_pingpong_mode(sk);
3275 inet_csk_exit_pingpong_mode(sk);
3276 if ((1 << sk->sk_state) & (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) &&
3277 inet_csk_ack_scheduled(sk)) {
3278 inet_csk(sk)->icsk_ack.pending |= ICSK_ACK_PUSHED;
3279 tcp_cleanup_rbuf(sk, 1);
3281 inet_csk_enter_pingpong_mode(sk);
3285 void tcp_sock_set_quickack(struct sock *sk, int val)
3288 __tcp_sock_set_quickack(sk, val);
3291 EXPORT_SYMBOL(tcp_sock_set_quickack);
3293 int tcp_sock_set_syncnt(struct sock *sk, int val)
3295 if (val < 1 || val > MAX_TCP_SYNCNT)
3299 inet_csk(sk)->icsk_syn_retries = val;
3303 EXPORT_SYMBOL(tcp_sock_set_syncnt);
3305 void tcp_sock_set_user_timeout(struct sock *sk, u32 val)
3308 inet_csk(sk)->icsk_user_timeout = val;
3311 EXPORT_SYMBOL(tcp_sock_set_user_timeout);
3313 int tcp_sock_set_keepidle_locked(struct sock *sk, int val)
3315 struct tcp_sock *tp = tcp_sk(sk);
3317 if (val < 1 || val > MAX_TCP_KEEPIDLE)
3320 tp->keepalive_time = val * HZ;
3321 if (sock_flag(sk, SOCK_KEEPOPEN) &&
3322 !((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN))) {
3323 u32 elapsed = keepalive_time_elapsed(tp);
3325 if (tp->keepalive_time > elapsed)
3326 elapsed = tp->keepalive_time - elapsed;
3329 inet_csk_reset_keepalive_timer(sk, elapsed);
3335 int tcp_sock_set_keepidle(struct sock *sk, int val)
3340 err = tcp_sock_set_keepidle_locked(sk, val);
3344 EXPORT_SYMBOL(tcp_sock_set_keepidle);
3346 int tcp_sock_set_keepintvl(struct sock *sk, int val)
3348 if (val < 1 || val > MAX_TCP_KEEPINTVL)
3352 tcp_sk(sk)->keepalive_intvl = val * HZ;
3356 EXPORT_SYMBOL(tcp_sock_set_keepintvl);
3358 int tcp_sock_set_keepcnt(struct sock *sk, int val)
3360 if (val < 1 || val > MAX_TCP_KEEPCNT)
3364 tcp_sk(sk)->keepalive_probes = val;
3368 EXPORT_SYMBOL(tcp_sock_set_keepcnt);
3370 int tcp_set_window_clamp(struct sock *sk, int val)
3372 struct tcp_sock *tp = tcp_sk(sk);
3375 if (sk->sk_state != TCP_CLOSE)
3377 tp->window_clamp = 0;
3379 tp->window_clamp = val < SOCK_MIN_RCVBUF / 2 ?
3380 SOCK_MIN_RCVBUF / 2 : val;
3381 tp->rcv_ssthresh = min(tp->rcv_wnd, tp->window_clamp);
3387 * Socket option code for TCP.
3389 int do_tcp_setsockopt(struct sock *sk, int level, int optname,
3390 sockptr_t optval, unsigned int optlen)
3392 struct tcp_sock *tp = tcp_sk(sk);
3393 struct inet_connection_sock *icsk = inet_csk(sk);
3394 struct net *net = sock_net(sk);
3398 /* These are data/string values, all the others are ints */
3400 case TCP_CONGESTION: {
3401 char name[TCP_CA_NAME_MAX];
3406 val = strncpy_from_sockptr(name, optval,
3407 min_t(long, TCP_CA_NAME_MAX-1, optlen));
3412 sockopt_lock_sock(sk);
3413 err = tcp_set_congestion_control(sk, name, !has_current_bpf_ctx(),
3414 sockopt_ns_capable(sock_net(sk)->user_ns,
3416 sockopt_release_sock(sk);
3420 char name[TCP_ULP_NAME_MAX];
3425 val = strncpy_from_sockptr(name, optval,
3426 min_t(long, TCP_ULP_NAME_MAX - 1,
3432 sockopt_lock_sock(sk);
3433 err = tcp_set_ulp(sk, name);
3434 sockopt_release_sock(sk);
3437 case TCP_FASTOPEN_KEY: {
3438 __u8 key[TCP_FASTOPEN_KEY_BUF_LENGTH];
3439 __u8 *backup_key = NULL;
3441 /* Allow a backup key as well to facilitate key rotation
3442 * First key is the active one.
3444 if (optlen != TCP_FASTOPEN_KEY_LENGTH &&
3445 optlen != TCP_FASTOPEN_KEY_BUF_LENGTH)
3448 if (copy_from_sockptr(key, optval, optlen))
3451 if (optlen == TCP_FASTOPEN_KEY_BUF_LENGTH)
3452 backup_key = key + TCP_FASTOPEN_KEY_LENGTH;
3454 return tcp_fastopen_reset_cipher(net, sk, key, backup_key);
3461 if (optlen < sizeof(int))
3464 if (copy_from_sockptr(&val, optval, sizeof(val)))
3467 sockopt_lock_sock(sk);
3471 /* Values greater than interface MTU won't take effect. However
3472 * at the point when this call is done we typically don't yet
3473 * know which interface is going to be used
3475 if (val && (val < TCP_MIN_MSS || val > MAX_TCP_WINDOW)) {
3479 tp->rx_opt.user_mss = val;
3483 __tcp_sock_set_nodelay(sk, val);
3486 case TCP_THIN_LINEAR_TIMEOUTS:
3487 if (val < 0 || val > 1)
3493 case TCP_THIN_DUPACK:
3494 if (val < 0 || val > 1)
3499 if (!tcp_can_repair_sock(sk))
3501 else if (val == TCP_REPAIR_ON) {
3503 sk->sk_reuse = SK_FORCE_REUSE;
3504 tp->repair_queue = TCP_NO_QUEUE;
3505 } else if (val == TCP_REPAIR_OFF) {
3507 sk->sk_reuse = SK_NO_REUSE;
3508 tcp_send_window_probe(sk);
3509 } else if (val == TCP_REPAIR_OFF_NO_WP) {
3511 sk->sk_reuse = SK_NO_REUSE;
3517 case TCP_REPAIR_QUEUE:
3520 else if ((unsigned int)val < TCP_QUEUES_NR)
3521 tp->repair_queue = val;
3527 if (sk->sk_state != TCP_CLOSE) {
3529 } else if (tp->repair_queue == TCP_SEND_QUEUE) {
3530 if (!tcp_rtx_queue_empty(sk))
3533 WRITE_ONCE(tp->write_seq, val);
3534 } else if (tp->repair_queue == TCP_RECV_QUEUE) {
3535 if (tp->rcv_nxt != tp->copied_seq) {
3538 WRITE_ONCE(tp->rcv_nxt, val);
3539 WRITE_ONCE(tp->copied_seq, val);
3546 case TCP_REPAIR_OPTIONS:
3549 else if (sk->sk_state == TCP_ESTABLISHED && !tp->bytes_sent)
3550 err = tcp_repair_options_est(sk, optval, optlen);
3556 __tcp_sock_set_cork(sk, val);
3560 err = tcp_sock_set_keepidle_locked(sk, val);
3563 if (val < 1 || val > MAX_TCP_KEEPINTVL)
3566 tp->keepalive_intvl = val * HZ;
3569 if (val < 1 || val > MAX_TCP_KEEPCNT)
3572 tp->keepalive_probes = val;
3575 if (val < 1 || val > MAX_TCP_SYNCNT)
3578 icsk->icsk_syn_retries = val;
3582 /* 0: disable, 1: enable, 2: start from ether_header */
3583 if (val < 0 || val > 2)
3592 else if (val > TCP_FIN_TIMEOUT_MAX / HZ)
3593 tp->linger2 = TCP_FIN_TIMEOUT_MAX;
3595 tp->linger2 = val * HZ;
3598 case TCP_DEFER_ACCEPT:
3599 /* Translate value in seconds to number of retransmits */
3600 icsk->icsk_accept_queue.rskq_defer_accept =
3601 secs_to_retrans(val, TCP_TIMEOUT_INIT / HZ,
3605 case TCP_WINDOW_CLAMP:
3606 err = tcp_set_window_clamp(sk, val);
3610 __tcp_sock_set_quickack(sk, val);
3613 #ifdef CONFIG_TCP_MD5SIG
3615 case TCP_MD5SIG_EXT:
3616 err = tp->af_specific->md5_parse(sk, optname, optval, optlen);
3619 case TCP_USER_TIMEOUT:
3620 /* Cap the max time in ms TCP will retry or probe the window
3621 * before giving up and aborting (ETIMEDOUT) a connection.
3626 icsk->icsk_user_timeout = val;
3630 if (val >= 0 && ((1 << sk->sk_state) & (TCPF_CLOSE |
3632 tcp_fastopen_init_key_once(net);
3634 fastopen_queue_tune(sk, val);
3639 case TCP_FASTOPEN_CONNECT:
3640 if (val > 1 || val < 0) {
3642 } else if (READ_ONCE(net->ipv4.sysctl_tcp_fastopen) &
3643 TFO_CLIENT_ENABLE) {
3644 if (sk->sk_state == TCP_CLOSE)
3645 tp->fastopen_connect = val;
3652 case TCP_FASTOPEN_NO_COOKIE:
3653 if (val > 1 || val < 0)
3655 else if (!((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
3658 tp->fastopen_no_cookie = val;
3664 tp->tsoffset = val - tcp_time_stamp_raw();
3666 case TCP_REPAIR_WINDOW:
3667 err = tcp_repair_set_window(tp, optval, optlen);
3669 case TCP_NOTSENT_LOWAT:
3670 tp->notsent_lowat = val;
3671 sk->sk_write_space(sk);
3674 if (val > 1 || val < 0)
3677 tp->recvmsg_inq = val;
3681 tcp_enable_tx_delay();
3682 tp->tcp_tx_delay = val;
3689 sockopt_release_sock(sk);
3693 int tcp_setsockopt(struct sock *sk, int level, int optname, sockptr_t optval,
3694 unsigned int optlen)
3696 const struct inet_connection_sock *icsk = inet_csk(sk);
3698 if (level != SOL_TCP)
3699 /* Paired with WRITE_ONCE() in do_ipv6_setsockopt() and tcp_v6_connect() */
3700 return READ_ONCE(icsk->icsk_af_ops)->setsockopt(sk, level, optname,
3702 return do_tcp_setsockopt(sk, level, optname, optval, optlen);
3704 EXPORT_SYMBOL(tcp_setsockopt);
3706 static void tcp_get_info_chrono_stats(const struct tcp_sock *tp,
3707 struct tcp_info *info)
3709 u64 stats[__TCP_CHRONO_MAX], total = 0;
3712 for (i = TCP_CHRONO_BUSY; i < __TCP_CHRONO_MAX; ++i) {
3713 stats[i] = tp->chrono_stat[i - 1];
3714 if (i == tp->chrono_type)
3715 stats[i] += tcp_jiffies32 - tp->chrono_start;
3716 stats[i] *= USEC_PER_SEC / HZ;
3720 info->tcpi_busy_time = total;
3721 info->tcpi_rwnd_limited = stats[TCP_CHRONO_RWND_LIMITED];
3722 info->tcpi_sndbuf_limited = stats[TCP_CHRONO_SNDBUF_LIMITED];
3725 /* Return information about state of tcp endpoint in API format. */
3726 void tcp_get_info(struct sock *sk, struct tcp_info *info)
3728 const struct tcp_sock *tp = tcp_sk(sk); /* iff sk_type == SOCK_STREAM */
3729 const struct inet_connection_sock *icsk = inet_csk(sk);
3735 memset(info, 0, sizeof(*info));
3736 if (sk->sk_type != SOCK_STREAM)
3739 info->tcpi_state = inet_sk_state_load(sk);
3741 /* Report meaningful fields for all TCP states, including listeners */
3742 rate = READ_ONCE(sk->sk_pacing_rate);
3743 rate64 = (rate != ~0UL) ? rate : ~0ULL;
3744 info->tcpi_pacing_rate = rate64;
3746 rate = READ_ONCE(sk->sk_max_pacing_rate);
3747 rate64 = (rate != ~0UL) ? rate : ~0ULL;
3748 info->tcpi_max_pacing_rate = rate64;
3750 info->tcpi_reordering = tp->reordering;
3751 info->tcpi_snd_cwnd = tcp_snd_cwnd(tp);
3753 if (info->tcpi_state == TCP_LISTEN) {
3754 /* listeners aliased fields :
3755 * tcpi_unacked -> Number of children ready for accept()
3756 * tcpi_sacked -> max backlog
3758 info->tcpi_unacked = READ_ONCE(sk->sk_ack_backlog);
3759 info->tcpi_sacked = READ_ONCE(sk->sk_max_ack_backlog);
3763 slow = lock_sock_fast(sk);
3765 info->tcpi_ca_state = icsk->icsk_ca_state;
3766 info->tcpi_retransmits = icsk->icsk_retransmits;
3767 info->tcpi_probes = icsk->icsk_probes_out;
3768 info->tcpi_backoff = icsk->icsk_backoff;
3770 if (tp->rx_opt.tstamp_ok)
3771 info->tcpi_options |= TCPI_OPT_TIMESTAMPS;
3772 if (tcp_is_sack(tp))
3773 info->tcpi_options |= TCPI_OPT_SACK;
3774 if (tp->rx_opt.wscale_ok) {
3775 info->tcpi_options |= TCPI_OPT_WSCALE;
3776 info->tcpi_snd_wscale = tp->rx_opt.snd_wscale;
3777 info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale;
3780 if (tp->ecn_flags & TCP_ECN_OK)
3781 info->tcpi_options |= TCPI_OPT_ECN;
3782 if (tp->ecn_flags & TCP_ECN_SEEN)
3783 info->tcpi_options |= TCPI_OPT_ECN_SEEN;
3784 if (tp->syn_data_acked)
3785 info->tcpi_options |= TCPI_OPT_SYN_DATA;
3787 info->tcpi_rto = jiffies_to_usecs(icsk->icsk_rto);
3788 info->tcpi_ato = jiffies_to_usecs(icsk->icsk_ack.ato);
3789 info->tcpi_snd_mss = tp->mss_cache;
3790 info->tcpi_rcv_mss = icsk->icsk_ack.rcv_mss;
3792 info->tcpi_unacked = tp->packets_out;
3793 info->tcpi_sacked = tp->sacked_out;
3795 info->tcpi_lost = tp->lost_out;
3796 info->tcpi_retrans = tp->retrans_out;
3798 now = tcp_jiffies32;
3799 info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime);
3800 info->tcpi_last_data_recv = jiffies_to_msecs(now - icsk->icsk_ack.lrcvtime);
3801 info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp);
3803 info->tcpi_pmtu = icsk->icsk_pmtu_cookie;
3804 info->tcpi_rcv_ssthresh = tp->rcv_ssthresh;
3805 info->tcpi_rtt = tp->srtt_us >> 3;
3806 info->tcpi_rttvar = tp->mdev_us >> 2;
3807 info->tcpi_snd_ssthresh = tp->snd_ssthresh;
3808 info->tcpi_advmss = tp->advmss;
3810 info->tcpi_rcv_rtt = tp->rcv_rtt_est.rtt_us >> 3;
3811 info->tcpi_rcv_space = tp->rcvq_space.space;
3813 info->tcpi_total_retrans = tp->total_retrans;
3815 info->tcpi_bytes_acked = tp->bytes_acked;
3816 info->tcpi_bytes_received = tp->bytes_received;
3817 info->tcpi_notsent_bytes = max_t(int, 0, tp->write_seq - tp->snd_nxt);
3818 tcp_get_info_chrono_stats(tp, info);
3820 info->tcpi_segs_out = tp->segs_out;
3822 /* segs_in and data_segs_in can be updated from tcp_segs_in() from BH */
3823 info->tcpi_segs_in = READ_ONCE(tp->segs_in);
3824 info->tcpi_data_segs_in = READ_ONCE(tp->data_segs_in);
3826 info->tcpi_min_rtt = tcp_min_rtt(tp);
3827 info->tcpi_data_segs_out = tp->data_segs_out;
3829 info->tcpi_delivery_rate_app_limited = tp->rate_app_limited ? 1 : 0;
3830 rate64 = tcp_compute_delivery_rate(tp);
3832 info->tcpi_delivery_rate = rate64;
3833 info->tcpi_delivered = tp->delivered;
3834 info->tcpi_delivered_ce = tp->delivered_ce;
3835 info->tcpi_bytes_sent = tp->bytes_sent;
3836 info->tcpi_bytes_retrans = tp->bytes_retrans;
3837 info->tcpi_dsack_dups = tp->dsack_dups;
3838 info->tcpi_reord_seen = tp->reord_seen;
3839 info->tcpi_rcv_ooopack = tp->rcv_ooopack;
3840 info->tcpi_snd_wnd = tp->snd_wnd;
3841 info->tcpi_rcv_wnd = tp->rcv_wnd;
3842 info->tcpi_rehash = tp->plb_rehash + tp->timeout_rehash;
3843 info->tcpi_fastopen_client_fail = tp->fastopen_client_fail;
3844 unlock_sock_fast(sk, slow);
3846 EXPORT_SYMBOL_GPL(tcp_get_info);
3848 static size_t tcp_opt_stats_get_size(void)
3851 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_BUSY */
3852 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_RWND_LIMITED */
3853 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_SNDBUF_LIMITED */
3854 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_DATA_SEGS_OUT */
3855 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_TOTAL_RETRANS */
3856 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_PACING_RATE */
3857 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_DELIVERY_RATE */
3858 nla_total_size(sizeof(u32)) + /* TCP_NLA_SND_CWND */
3859 nla_total_size(sizeof(u32)) + /* TCP_NLA_REORDERING */
3860 nla_total_size(sizeof(u32)) + /* TCP_NLA_MIN_RTT */
3861 nla_total_size(sizeof(u8)) + /* TCP_NLA_RECUR_RETRANS */
3862 nla_total_size(sizeof(u8)) + /* TCP_NLA_DELIVERY_RATE_APP_LMT */
3863 nla_total_size(sizeof(u32)) + /* TCP_NLA_SNDQ_SIZE */
3864 nla_total_size(sizeof(u8)) + /* TCP_NLA_CA_STATE */
3865 nla_total_size(sizeof(u32)) + /* TCP_NLA_SND_SSTHRESH */
3866 nla_total_size(sizeof(u32)) + /* TCP_NLA_DELIVERED */
3867 nla_total_size(sizeof(u32)) + /* TCP_NLA_DELIVERED_CE */
3868 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_BYTES_SENT */
3869 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_BYTES_RETRANS */
3870 nla_total_size(sizeof(u32)) + /* TCP_NLA_DSACK_DUPS */
3871 nla_total_size(sizeof(u32)) + /* TCP_NLA_REORD_SEEN */
3872 nla_total_size(sizeof(u32)) + /* TCP_NLA_SRTT */
3873 nla_total_size(sizeof(u16)) + /* TCP_NLA_TIMEOUT_REHASH */
3874 nla_total_size(sizeof(u32)) + /* TCP_NLA_BYTES_NOTSENT */
3875 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_EDT */
3876 nla_total_size(sizeof(u8)) + /* TCP_NLA_TTL */
3877 nla_total_size(sizeof(u32)) + /* TCP_NLA_REHASH */
3881 /* Returns TTL or hop limit of an incoming packet from skb. */
3882 static u8 tcp_skb_ttl_or_hop_limit(const struct sk_buff *skb)
3884 if (skb->protocol == htons(ETH_P_IP))
3885 return ip_hdr(skb)->ttl;
3886 else if (skb->protocol == htons(ETH_P_IPV6))
3887 return ipv6_hdr(skb)->hop_limit;
3892 struct sk_buff *tcp_get_timestamping_opt_stats(const struct sock *sk,
3893 const struct sk_buff *orig_skb,
3894 const struct sk_buff *ack_skb)
3896 const struct tcp_sock *tp = tcp_sk(sk);
3897 struct sk_buff *stats;
3898 struct tcp_info info;
3902 stats = alloc_skb(tcp_opt_stats_get_size(), GFP_ATOMIC);
3906 tcp_get_info_chrono_stats(tp, &info);
3907 nla_put_u64_64bit(stats, TCP_NLA_BUSY,
3908 info.tcpi_busy_time, TCP_NLA_PAD);
3909 nla_put_u64_64bit(stats, TCP_NLA_RWND_LIMITED,
3910 info.tcpi_rwnd_limited, TCP_NLA_PAD);
3911 nla_put_u64_64bit(stats, TCP_NLA_SNDBUF_LIMITED,
3912 info.tcpi_sndbuf_limited, TCP_NLA_PAD);
3913 nla_put_u64_64bit(stats, TCP_NLA_DATA_SEGS_OUT,
3914 tp->data_segs_out, TCP_NLA_PAD);
3915 nla_put_u64_64bit(stats, TCP_NLA_TOTAL_RETRANS,
3916 tp->total_retrans, TCP_NLA_PAD);
3918 rate = READ_ONCE(sk->sk_pacing_rate);
3919 rate64 = (rate != ~0UL) ? rate : ~0ULL;
3920 nla_put_u64_64bit(stats, TCP_NLA_PACING_RATE, rate64, TCP_NLA_PAD);
3922 rate64 = tcp_compute_delivery_rate(tp);
3923 nla_put_u64_64bit(stats, TCP_NLA_DELIVERY_RATE, rate64, TCP_NLA_PAD);
3925 nla_put_u32(stats, TCP_NLA_SND_CWND, tcp_snd_cwnd(tp));
3926 nla_put_u32(stats, TCP_NLA_REORDERING, tp->reordering);
3927 nla_put_u32(stats, TCP_NLA_MIN_RTT, tcp_min_rtt(tp));
3929 nla_put_u8(stats, TCP_NLA_RECUR_RETRANS, inet_csk(sk)->icsk_retransmits);
3930 nla_put_u8(stats, TCP_NLA_DELIVERY_RATE_APP_LMT, !!tp->rate_app_limited);
3931 nla_put_u32(stats, TCP_NLA_SND_SSTHRESH, tp->snd_ssthresh);
3932 nla_put_u32(stats, TCP_NLA_DELIVERED, tp->delivered);
3933 nla_put_u32(stats, TCP_NLA_DELIVERED_CE, tp->delivered_ce);
3935 nla_put_u32(stats, TCP_NLA_SNDQ_SIZE, tp->write_seq - tp->snd_una);
3936 nla_put_u8(stats, TCP_NLA_CA_STATE, inet_csk(sk)->icsk_ca_state);
3938 nla_put_u64_64bit(stats, TCP_NLA_BYTES_SENT, tp->bytes_sent,
3940 nla_put_u64_64bit(stats, TCP_NLA_BYTES_RETRANS, tp->bytes_retrans,
3942 nla_put_u32(stats, TCP_NLA_DSACK_DUPS, tp->dsack_dups);
3943 nla_put_u32(stats, TCP_NLA_REORD_SEEN, tp->reord_seen);
3944 nla_put_u32(stats, TCP_NLA_SRTT, tp->srtt_us >> 3);
3945 nla_put_u16(stats, TCP_NLA_TIMEOUT_REHASH, tp->timeout_rehash);
3946 nla_put_u32(stats, TCP_NLA_BYTES_NOTSENT,
3947 max_t(int, 0, tp->write_seq - tp->snd_nxt));
3948 nla_put_u64_64bit(stats, TCP_NLA_EDT, orig_skb->skb_mstamp_ns,
3951 nla_put_u8(stats, TCP_NLA_TTL,
3952 tcp_skb_ttl_or_hop_limit(ack_skb));
3954 nla_put_u32(stats, TCP_NLA_REHASH, tp->plb_rehash + tp->timeout_rehash);
3958 int do_tcp_getsockopt(struct sock *sk, int level,
3959 int optname, sockptr_t optval, sockptr_t optlen)
3961 struct inet_connection_sock *icsk = inet_csk(sk);
3962 struct tcp_sock *tp = tcp_sk(sk);
3963 struct net *net = sock_net(sk);
3966 if (copy_from_sockptr(&len, optlen, sizeof(int)))
3969 len = min_t(unsigned int, len, sizeof(int));
3976 val = tp->mss_cache;
3977 if (tp->rx_opt.user_mss &&
3978 ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
3979 val = tp->rx_opt.user_mss;
3981 val = tp->rx_opt.mss_clamp;
3984 val = !!(tp->nonagle&TCP_NAGLE_OFF);
3987 val = !!(tp->nonagle&TCP_NAGLE_CORK);
3990 val = keepalive_time_when(tp) / HZ;
3993 val = keepalive_intvl_when(tp) / HZ;
3996 val = keepalive_probes(tp);
3999 val = icsk->icsk_syn_retries ? :
4000 READ_ONCE(net->ipv4.sysctl_tcp_syn_retries);
4005 val = (val ? : READ_ONCE(net->ipv4.sysctl_tcp_fin_timeout)) / HZ;
4007 case TCP_DEFER_ACCEPT:
4008 val = retrans_to_secs(icsk->icsk_accept_queue.rskq_defer_accept,
4009 TCP_TIMEOUT_INIT / HZ, TCP_RTO_MAX / HZ);
4011 case TCP_WINDOW_CLAMP:
4012 val = tp->window_clamp;
4015 struct tcp_info info;
4017 if (copy_from_sockptr(&len, optlen, sizeof(int)))
4020 tcp_get_info(sk, &info);
4022 len = min_t(unsigned int, len, sizeof(info));
4023 if (copy_to_sockptr(optlen, &len, sizeof(int)))
4025 if (copy_to_sockptr(optval, &info, len))
4030 const struct tcp_congestion_ops *ca_ops;
4031 union tcp_cc_info info;
4035 if (copy_from_sockptr(&len, optlen, sizeof(int)))
4038 ca_ops = icsk->icsk_ca_ops;
4039 if (ca_ops && ca_ops->get_info)
4040 sz = ca_ops->get_info(sk, ~0U, &attr, &info);
4042 len = min_t(unsigned int, len, sz);
4043 if (copy_to_sockptr(optlen, &len, sizeof(int)))
4045 if (copy_to_sockptr(optval, &info, len))
4050 val = !inet_csk_in_pingpong_mode(sk);
4053 case TCP_CONGESTION:
4054 if (copy_from_sockptr(&len, optlen, sizeof(int)))
4056 len = min_t(unsigned int, len, TCP_CA_NAME_MAX);
4057 if (copy_to_sockptr(optlen, &len, sizeof(int)))
4059 if (copy_to_sockptr(optval, icsk->icsk_ca_ops->name, len))
4064 if (copy_from_sockptr(&len, optlen, sizeof(int)))
4066 len = min_t(unsigned int, len, TCP_ULP_NAME_MAX);
4067 if (!icsk->icsk_ulp_ops) {
4069 if (copy_to_sockptr(optlen, &len, sizeof(int)))
4073 if (copy_to_sockptr(optlen, &len, sizeof(int)))
4075 if (copy_to_sockptr(optval, icsk->icsk_ulp_ops->name, len))
4079 case TCP_FASTOPEN_KEY: {
4080 u64 key[TCP_FASTOPEN_KEY_BUF_LENGTH / sizeof(u64)];
4081 unsigned int key_len;
4083 if (copy_from_sockptr(&len, optlen, sizeof(int)))
4086 key_len = tcp_fastopen_get_cipher(net, icsk, key) *
4087 TCP_FASTOPEN_KEY_LENGTH;
4088 len = min_t(unsigned int, len, key_len);
4089 if (copy_to_sockptr(optlen, &len, sizeof(int)))
4091 if (copy_to_sockptr(optval, key, len))
4095 case TCP_THIN_LINEAR_TIMEOUTS:
4099 case TCP_THIN_DUPACK:
4107 case TCP_REPAIR_QUEUE:
4109 val = tp->repair_queue;
4114 case TCP_REPAIR_WINDOW: {
4115 struct tcp_repair_window opt;
4117 if (copy_from_sockptr(&len, optlen, sizeof(int)))
4120 if (len != sizeof(opt))
4126 opt.snd_wl1 = tp->snd_wl1;
4127 opt.snd_wnd = tp->snd_wnd;
4128 opt.max_window = tp->max_window;
4129 opt.rcv_wnd = tp->rcv_wnd;
4130 opt.rcv_wup = tp->rcv_wup;
4132 if (copy_to_sockptr(optval, &opt, len))
4137 if (tp->repair_queue == TCP_SEND_QUEUE)
4138 val = tp->write_seq;
4139 else if (tp->repair_queue == TCP_RECV_QUEUE)
4145 case TCP_USER_TIMEOUT:
4146 val = icsk->icsk_user_timeout;
4150 val = icsk->icsk_accept_queue.fastopenq.max_qlen;
4153 case TCP_FASTOPEN_CONNECT:
4154 val = tp->fastopen_connect;
4157 case TCP_FASTOPEN_NO_COOKIE:
4158 val = tp->fastopen_no_cookie;
4162 val = tp->tcp_tx_delay;
4166 val = tcp_time_stamp_raw() + tp->tsoffset;
4168 case TCP_NOTSENT_LOWAT:
4169 val = tp->notsent_lowat;
4172 val = tp->recvmsg_inq;
4177 case TCP_SAVED_SYN: {
4178 if (copy_from_sockptr(&len, optlen, sizeof(int)))
4181 sockopt_lock_sock(sk);
4182 if (tp->saved_syn) {
4183 if (len < tcp_saved_syn_len(tp->saved_syn)) {
4184 len = tcp_saved_syn_len(tp->saved_syn);
4185 if (copy_to_sockptr(optlen, &len, sizeof(int))) {
4186 sockopt_release_sock(sk);
4189 sockopt_release_sock(sk);
4192 len = tcp_saved_syn_len(tp->saved_syn);
4193 if (copy_to_sockptr(optlen, &len, sizeof(int))) {
4194 sockopt_release_sock(sk);
4197 if (copy_to_sockptr(optval, tp->saved_syn->data, len)) {
4198 sockopt_release_sock(sk);
4201 tcp_saved_syn_free(tp);
4202 sockopt_release_sock(sk);
4204 sockopt_release_sock(sk);
4206 if (copy_to_sockptr(optlen, &len, sizeof(int)))
4212 case TCP_ZEROCOPY_RECEIVE: {
4213 struct scm_timestamping_internal tss;
4214 struct tcp_zerocopy_receive zc = {};
4217 if (copy_from_sockptr(&len, optlen, sizeof(int)))
4220 len < offsetofend(struct tcp_zerocopy_receive, length))
4222 if (unlikely(len > sizeof(zc))) {
4223 err = check_zeroed_sockptr(optval, sizeof(zc),
4226 return err == 0 ? -EINVAL : err;
4228 if (copy_to_sockptr(optlen, &len, sizeof(int)))
4231 if (copy_from_sockptr(&zc, optval, len))
4235 if (zc.msg_flags & ~(TCP_VALID_ZC_MSG_FLAGS))
4237 sockopt_lock_sock(sk);
4238 err = tcp_zerocopy_receive(sk, &zc, &tss);
4239 err = BPF_CGROUP_RUN_PROG_GETSOCKOPT_KERN(sk, level, optname,
4241 sockopt_release_sock(sk);
4242 if (len >= offsetofend(struct tcp_zerocopy_receive, msg_flags))
4243 goto zerocopy_rcv_cmsg;
4245 case offsetofend(struct tcp_zerocopy_receive, msg_flags):
4246 goto zerocopy_rcv_cmsg;
4247 case offsetofend(struct tcp_zerocopy_receive, msg_controllen):
4248 case offsetofend(struct tcp_zerocopy_receive, msg_control):
4249 case offsetofend(struct tcp_zerocopy_receive, flags):
4250 case offsetofend(struct tcp_zerocopy_receive, copybuf_len):
4251 case offsetofend(struct tcp_zerocopy_receive, copybuf_address):
4252 case offsetofend(struct tcp_zerocopy_receive, err):
4253 goto zerocopy_rcv_sk_err;
4254 case offsetofend(struct tcp_zerocopy_receive, inq):
4255 goto zerocopy_rcv_inq;
4256 case offsetofend(struct tcp_zerocopy_receive, length):
4258 goto zerocopy_rcv_out;
4261 if (zc.msg_flags & TCP_CMSG_TS)
4262 tcp_zc_finalize_rx_tstamp(sk, &zc, &tss);
4265 zerocopy_rcv_sk_err:
4267 zc.err = sock_error(sk);
4269 zc.inq = tcp_inq_hint(sk);
4271 if (!err && copy_to_sockptr(optval, &zc, len))
4277 return -ENOPROTOOPT;
4280 if (copy_to_sockptr(optlen, &len, sizeof(int)))
4282 if (copy_to_sockptr(optval, &val, len))
4287 bool tcp_bpf_bypass_getsockopt(int level, int optname)
4289 /* TCP do_tcp_getsockopt has optimized getsockopt implementation
4290 * to avoid extra socket lock for TCP_ZEROCOPY_RECEIVE.
4292 if (level == SOL_TCP && optname == TCP_ZEROCOPY_RECEIVE)
4297 EXPORT_SYMBOL(tcp_bpf_bypass_getsockopt);
4299 int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval,
4302 struct inet_connection_sock *icsk = inet_csk(sk);
4304 if (level != SOL_TCP)
4305 /* Paired with WRITE_ONCE() in do_ipv6_setsockopt() and tcp_v6_connect() */
4306 return READ_ONCE(icsk->icsk_af_ops)->getsockopt(sk, level, optname,
4308 return do_tcp_getsockopt(sk, level, optname, USER_SOCKPTR(optval),
4309 USER_SOCKPTR(optlen));
4311 EXPORT_SYMBOL(tcp_getsockopt);
4313 #ifdef CONFIG_TCP_MD5SIG
4314 static DEFINE_PER_CPU(struct tcp_md5sig_pool, tcp_md5sig_pool);
4315 static DEFINE_MUTEX(tcp_md5sig_mutex);
4316 static bool tcp_md5sig_pool_populated = false;
4318 static void __tcp_alloc_md5sig_pool(void)
4320 struct crypto_ahash *hash;
4323 hash = crypto_alloc_ahash("md5", 0, CRYPTO_ALG_ASYNC);
4327 for_each_possible_cpu(cpu) {
4328 void *scratch = per_cpu(tcp_md5sig_pool, cpu).scratch;
4329 struct ahash_request *req;
4332 scratch = kmalloc_node(sizeof(union tcp_md5sum_block) +
4333 sizeof(struct tcphdr),
4338 per_cpu(tcp_md5sig_pool, cpu).scratch = scratch;
4340 if (per_cpu(tcp_md5sig_pool, cpu).md5_req)
4343 req = ahash_request_alloc(hash, GFP_KERNEL);
4347 ahash_request_set_callback(req, 0, NULL, NULL);
4349 per_cpu(tcp_md5sig_pool, cpu).md5_req = req;
4351 /* before setting tcp_md5sig_pool_populated, we must commit all writes
4352 * to memory. See smp_rmb() in tcp_get_md5sig_pool()
4355 /* Paired with READ_ONCE() from tcp_alloc_md5sig_pool()
4356 * and tcp_get_md5sig_pool().
4358 WRITE_ONCE(tcp_md5sig_pool_populated, true);
4361 bool tcp_alloc_md5sig_pool(void)
4363 /* Paired with WRITE_ONCE() from __tcp_alloc_md5sig_pool() */
4364 if (unlikely(!READ_ONCE(tcp_md5sig_pool_populated))) {
4365 mutex_lock(&tcp_md5sig_mutex);
4367 if (!tcp_md5sig_pool_populated)
4368 __tcp_alloc_md5sig_pool();
4370 mutex_unlock(&tcp_md5sig_mutex);
4372 /* Paired with WRITE_ONCE() from __tcp_alloc_md5sig_pool() */
4373 return READ_ONCE(tcp_md5sig_pool_populated);
4375 EXPORT_SYMBOL(tcp_alloc_md5sig_pool);
4379 * tcp_get_md5sig_pool - get md5sig_pool for this user
4381 * We use percpu structure, so if we succeed, we exit with preemption
4382 * and BH disabled, to make sure another thread or softirq handling
4383 * wont try to get same context.
4385 struct tcp_md5sig_pool *tcp_get_md5sig_pool(void)
4389 /* Paired with WRITE_ONCE() from __tcp_alloc_md5sig_pool() */
4390 if (READ_ONCE(tcp_md5sig_pool_populated)) {
4391 /* coupled with smp_wmb() in __tcp_alloc_md5sig_pool() */
4393 return this_cpu_ptr(&tcp_md5sig_pool);
4398 EXPORT_SYMBOL(tcp_get_md5sig_pool);
4400 int tcp_md5_hash_skb_data(struct tcp_md5sig_pool *hp,
4401 const struct sk_buff *skb, unsigned int header_len)
4403 struct scatterlist sg;
4404 const struct tcphdr *tp = tcp_hdr(skb);
4405 struct ahash_request *req = hp->md5_req;
4407 const unsigned int head_data_len = skb_headlen(skb) > header_len ?
4408 skb_headlen(skb) - header_len : 0;
4409 const struct skb_shared_info *shi = skb_shinfo(skb);
4410 struct sk_buff *frag_iter;
4412 sg_init_table(&sg, 1);
4414 sg_set_buf(&sg, ((u8 *) tp) + header_len, head_data_len);
4415 ahash_request_set_crypt(req, &sg, NULL, head_data_len);
4416 if (crypto_ahash_update(req))
4419 for (i = 0; i < shi->nr_frags; ++i) {
4420 const skb_frag_t *f = &shi->frags[i];
4421 unsigned int offset = skb_frag_off(f);
4422 struct page *page = skb_frag_page(f) + (offset >> PAGE_SHIFT);
4424 sg_set_page(&sg, page, skb_frag_size(f),
4425 offset_in_page(offset));
4426 ahash_request_set_crypt(req, &sg, NULL, skb_frag_size(f));
4427 if (crypto_ahash_update(req))
4431 skb_walk_frags(skb, frag_iter)
4432 if (tcp_md5_hash_skb_data(hp, frag_iter, 0))
4437 EXPORT_SYMBOL(tcp_md5_hash_skb_data);
4439 int tcp_md5_hash_key(struct tcp_md5sig_pool *hp, const struct tcp_md5sig_key *key)
4441 u8 keylen = READ_ONCE(key->keylen); /* paired with WRITE_ONCE() in tcp_md5_do_add */
4442 struct scatterlist sg;
4444 sg_init_one(&sg, key->key, keylen);
4445 ahash_request_set_crypt(hp->md5_req, &sg, NULL, keylen);
4447 /* We use data_race() because tcp_md5_do_add() might change key->key under us */
4448 return data_race(crypto_ahash_update(hp->md5_req));
4450 EXPORT_SYMBOL(tcp_md5_hash_key);
4452 /* Called with rcu_read_lock() */
4453 enum skb_drop_reason
4454 tcp_inbound_md5_hash(const struct sock *sk, const struct sk_buff *skb,
4455 const void *saddr, const void *daddr,
4456 int family, int dif, int sdif)
4459 * This gets called for each TCP segment that arrives
4460 * so we want to be efficient.
4461 * We have 3 drop cases:
4462 * o No MD5 hash and one expected.
4463 * o MD5 hash and we're not expecting one.
4464 * o MD5 hash and its wrong.
4466 const __u8 *hash_location = NULL;
4467 struct tcp_md5sig_key *hash_expected;
4468 const struct tcphdr *th = tcp_hdr(skb);
4469 const struct tcp_sock *tp = tcp_sk(sk);
4470 int genhash, l3index;
4473 /* sdif set, means packet ingressed via a device
4474 * in an L3 domain and dif is set to the l3mdev
4476 l3index = sdif ? dif : 0;
4478 hash_expected = tcp_md5_do_lookup(sk, l3index, saddr, family);
4479 hash_location = tcp_parse_md5sig_option(th);
4481 /* We've parsed the options - do we have a hash? */
4482 if (!hash_expected && !hash_location)
4483 return SKB_NOT_DROPPED_YET;
4485 if (hash_expected && !hash_location) {
4486 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMD5NOTFOUND);
4487 return SKB_DROP_REASON_TCP_MD5NOTFOUND;
4490 if (!hash_expected && hash_location) {
4491 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMD5UNEXPECTED);
4492 return SKB_DROP_REASON_TCP_MD5UNEXPECTED;
4495 /* Check the signature.
4496 * To support dual stack listeners, we need to handle
4499 if (family == AF_INET)
4500 genhash = tcp_v4_md5_hash_skb(newhash,
4504 genhash = tp->af_specific->calc_md5_hash(newhash,
4508 if (genhash || memcmp(hash_location, newhash, 16) != 0) {
4509 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMD5FAILURE);
4510 if (family == AF_INET) {
4511 net_info_ratelimited("MD5 Hash failed for (%pI4, %d)->(%pI4, %d)%s L3 index %d\n",
4512 saddr, ntohs(th->source),
4513 daddr, ntohs(th->dest),
4514 genhash ? " tcp_v4_calc_md5_hash failed"
4517 net_info_ratelimited("MD5 Hash %s for [%pI6c]:%u->[%pI6c]:%u L3 index %d\n",
4518 genhash ? "failed" : "mismatch",
4519 saddr, ntohs(th->source),
4520 daddr, ntohs(th->dest), l3index);
4522 return SKB_DROP_REASON_TCP_MD5FAILURE;
4524 return SKB_NOT_DROPPED_YET;
4526 EXPORT_SYMBOL(tcp_inbound_md5_hash);
4530 void tcp_done(struct sock *sk)
4532 struct request_sock *req;
4534 /* We might be called with a new socket, after
4535 * inet_csk_prepare_forced_close() has been called
4536 * so we can not use lockdep_sock_is_held(sk)
4538 req = rcu_dereference_protected(tcp_sk(sk)->fastopen_rsk, 1);
4540 if (sk->sk_state == TCP_SYN_SENT || sk->sk_state == TCP_SYN_RECV)
4541 TCP_INC_STATS(sock_net(sk), TCP_MIB_ATTEMPTFAILS);
4543 tcp_set_state(sk, TCP_CLOSE);
4544 tcp_clear_xmit_timers(sk);
4546 reqsk_fastopen_remove(sk, req, false);
4548 WRITE_ONCE(sk->sk_shutdown, SHUTDOWN_MASK);
4550 if (!sock_flag(sk, SOCK_DEAD))
4551 sk->sk_state_change(sk);
4553 inet_csk_destroy_sock(sk);
4555 EXPORT_SYMBOL_GPL(tcp_done);
4557 int tcp_abort(struct sock *sk, int err)
4559 int state = inet_sk_state_load(sk);
4561 if (state == TCP_NEW_SYN_RECV) {
4562 struct request_sock *req = inet_reqsk(sk);
4565 inet_csk_reqsk_queue_drop(req->rsk_listener, req);
4569 if (state == TCP_TIME_WAIT) {
4570 struct inet_timewait_sock *tw = inet_twsk(sk);
4572 refcount_inc(&tw->tw_refcnt);
4574 inet_twsk_deschedule_put(tw);
4579 /* BPF context ensures sock locking. */
4580 if (!has_current_bpf_ctx())
4581 /* Don't race with userspace socket closes such as tcp_close. */
4584 if (sk->sk_state == TCP_LISTEN) {
4585 tcp_set_state(sk, TCP_CLOSE);
4586 inet_csk_listen_stop(sk);
4589 /* Don't race with BH socket closes such as inet_csk_listen_stop. */
4593 if (!sock_flag(sk, SOCK_DEAD)) {
4594 WRITE_ONCE(sk->sk_err, err);
4595 /* This barrier is coupled with smp_rmb() in tcp_poll() */
4597 sk_error_report(sk);
4598 if (tcp_need_reset(sk->sk_state))
4599 tcp_send_active_reset(sk, GFP_ATOMIC);
4605 tcp_write_queue_purge(sk);
4606 if (!has_current_bpf_ctx())
4610 EXPORT_SYMBOL_GPL(tcp_abort);
4612 extern struct tcp_congestion_ops tcp_reno;
4614 static __initdata unsigned long thash_entries;
4615 static int __init set_thash_entries(char *str)
4622 ret = kstrtoul(str, 0, &thash_entries);
4628 __setup("thash_entries=", set_thash_entries);
4630 static void __init tcp_init_mem(void)
4632 unsigned long limit = nr_free_buffer_pages() / 16;
4634 limit = max(limit, 128UL);
4635 sysctl_tcp_mem[0] = limit / 4 * 3; /* 4.68 % */
4636 sysctl_tcp_mem[1] = limit; /* 6.25 % */
4637 sysctl_tcp_mem[2] = sysctl_tcp_mem[0] * 2; /* 9.37 % */
4640 void __init tcp_init(void)
4642 int max_rshare, max_wshare, cnt;
4643 unsigned long limit;
4646 BUILD_BUG_ON(TCP_MIN_SND_MSS <= MAX_TCP_OPTION_SPACE);
4647 BUILD_BUG_ON(sizeof(struct tcp_skb_cb) >
4648 sizeof_field(struct sk_buff, cb));
4650 percpu_counter_init(&tcp_sockets_allocated, 0, GFP_KERNEL);
4652 timer_setup(&tcp_orphan_timer, tcp_orphan_update, TIMER_DEFERRABLE);
4653 mod_timer(&tcp_orphan_timer, jiffies + TCP_ORPHAN_TIMER_PERIOD);
4655 inet_hashinfo2_init(&tcp_hashinfo, "tcp_listen_portaddr_hash",
4656 thash_entries, 21, /* one slot per 2 MB*/
4658 tcp_hashinfo.bind_bucket_cachep =
4659 kmem_cache_create("tcp_bind_bucket",
4660 sizeof(struct inet_bind_bucket), 0,
4661 SLAB_HWCACHE_ALIGN | SLAB_PANIC |
4664 tcp_hashinfo.bind2_bucket_cachep =
4665 kmem_cache_create("tcp_bind2_bucket",
4666 sizeof(struct inet_bind2_bucket), 0,
4667 SLAB_HWCACHE_ALIGN | SLAB_PANIC |
4671 /* Size and allocate the main established and bind bucket
4674 * The methodology is similar to that of the buffer cache.
4676 tcp_hashinfo.ehash =
4677 alloc_large_system_hash("TCP established",
4678 sizeof(struct inet_ehash_bucket),
4680 17, /* one slot per 128 KB of memory */
4683 &tcp_hashinfo.ehash_mask,
4685 thash_entries ? 0 : 512 * 1024);
4686 for (i = 0; i <= tcp_hashinfo.ehash_mask; i++)
4687 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].chain, i);
4689 if (inet_ehash_locks_alloc(&tcp_hashinfo))
4690 panic("TCP: failed to alloc ehash_locks");
4691 tcp_hashinfo.bhash =
4692 alloc_large_system_hash("TCP bind",
4693 2 * sizeof(struct inet_bind_hashbucket),
4694 tcp_hashinfo.ehash_mask + 1,
4695 17, /* one slot per 128 KB of memory */
4697 &tcp_hashinfo.bhash_size,
4701 tcp_hashinfo.bhash_size = 1U << tcp_hashinfo.bhash_size;
4702 tcp_hashinfo.bhash2 = tcp_hashinfo.bhash + tcp_hashinfo.bhash_size;
4703 for (i = 0; i < tcp_hashinfo.bhash_size; i++) {
4704 spin_lock_init(&tcp_hashinfo.bhash[i].lock);
4705 INIT_HLIST_HEAD(&tcp_hashinfo.bhash[i].chain);
4706 spin_lock_init(&tcp_hashinfo.bhash2[i].lock);
4707 INIT_HLIST_HEAD(&tcp_hashinfo.bhash2[i].chain);
4710 tcp_hashinfo.pernet = false;
4712 cnt = tcp_hashinfo.ehash_mask + 1;
4713 sysctl_tcp_max_orphans = cnt / 2;
4716 /* Set per-socket limits to no more than 1/128 the pressure threshold */
4717 limit = nr_free_buffer_pages() << (PAGE_SHIFT - 7);
4718 max_wshare = min(4UL*1024*1024, limit);
4719 max_rshare = min(6UL*1024*1024, limit);
4721 init_net.ipv4.sysctl_tcp_wmem[0] = PAGE_SIZE;
4722 init_net.ipv4.sysctl_tcp_wmem[1] = 16*1024;
4723 init_net.ipv4.sysctl_tcp_wmem[2] = max(64*1024, max_wshare);
4725 init_net.ipv4.sysctl_tcp_rmem[0] = PAGE_SIZE;
4726 init_net.ipv4.sysctl_tcp_rmem[1] = 131072;
4727 init_net.ipv4.sysctl_tcp_rmem[2] = max(131072, max_rshare);
4729 pr_info("Hash tables configured (established %u bind %u)\n",
4730 tcp_hashinfo.ehash_mask + 1, tcp_hashinfo.bhash_size);
4734 BUG_ON(tcp_register_congestion_control(&tcp_reno) != 0);