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/swap.h>
264 #include <linux/cache.h>
265 #include <linux/err.h>
266 #include <linux/time.h>
267 #include <linux/slab.h>
268 #include <linux/errqueue.h>
269 #include <linux/static_key.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 struct percpu_counter tcp_orphan_count;
284 EXPORT_SYMBOL_GPL(tcp_orphan_count);
286 long sysctl_tcp_mem[3] __read_mostly;
287 EXPORT_SYMBOL(sysctl_tcp_mem);
289 atomic_long_t tcp_memory_allocated; /* Current allocated memory. */
290 EXPORT_SYMBOL(tcp_memory_allocated);
292 #if IS_ENABLED(CONFIG_SMC)
293 DEFINE_STATIC_KEY_FALSE(tcp_have_smc);
294 EXPORT_SYMBOL(tcp_have_smc);
298 * Current number of TCP sockets.
300 struct percpu_counter tcp_sockets_allocated;
301 EXPORT_SYMBOL(tcp_sockets_allocated);
306 struct tcp_splice_state {
307 struct pipe_inode_info *pipe;
313 * Pressure flag: try to collapse.
314 * Technical note: it is used by multiple contexts non atomically.
315 * All the __sk_mem_schedule() is of this nature: accounting
316 * is strict, actions are advisory and have some latency.
318 unsigned long tcp_memory_pressure __read_mostly;
319 EXPORT_SYMBOL_GPL(tcp_memory_pressure);
321 DEFINE_STATIC_KEY_FALSE(tcp_rx_skb_cache_key);
322 EXPORT_SYMBOL(tcp_rx_skb_cache_key);
324 DEFINE_STATIC_KEY_FALSE(tcp_tx_skb_cache_key);
326 void tcp_enter_memory_pressure(struct sock *sk)
330 if (READ_ONCE(tcp_memory_pressure))
336 if (!cmpxchg(&tcp_memory_pressure, 0, val))
337 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMEMORYPRESSURES);
339 EXPORT_SYMBOL_GPL(tcp_enter_memory_pressure);
341 void tcp_leave_memory_pressure(struct sock *sk)
345 if (!READ_ONCE(tcp_memory_pressure))
347 val = xchg(&tcp_memory_pressure, 0);
349 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPMEMORYPRESSURESCHRONO,
350 jiffies_to_msecs(jiffies - val));
352 EXPORT_SYMBOL_GPL(tcp_leave_memory_pressure);
354 /* Convert seconds to retransmits based on initial and max timeout */
355 static u8 secs_to_retrans(int seconds, int timeout, int rto_max)
360 int period = timeout;
363 while (seconds > period && res < 255) {
366 if (timeout > rto_max)
374 /* Convert retransmits to seconds based on initial and max timeout */
375 static int retrans_to_secs(u8 retrans, int timeout, int rto_max)
383 if (timeout > rto_max)
391 static u64 tcp_compute_delivery_rate(const struct tcp_sock *tp)
393 u32 rate = READ_ONCE(tp->rate_delivered);
394 u32 intv = READ_ONCE(tp->rate_interval_us);
398 rate64 = (u64)rate * tp->mss_cache * USEC_PER_SEC;
399 do_div(rate64, intv);
404 /* Address-family independent initialization for a tcp_sock.
406 * NOTE: A lot of things set to zero explicitly by call to
407 * sk_alloc() so need not be done here.
409 void tcp_init_sock(struct sock *sk)
411 struct inet_connection_sock *icsk = inet_csk(sk);
412 struct tcp_sock *tp = tcp_sk(sk);
414 tp->out_of_order_queue = RB_ROOT;
415 sk->tcp_rtx_queue = RB_ROOT;
416 tcp_init_xmit_timers(sk);
417 INIT_LIST_HEAD(&tp->tsq_node);
418 INIT_LIST_HEAD(&tp->tsorted_sent_queue);
420 icsk->icsk_rto = TCP_TIMEOUT_INIT;
421 icsk->icsk_rto_min = TCP_RTO_MIN;
422 icsk->icsk_delack_max = TCP_DELACK_MAX;
423 tp->mdev_us = jiffies_to_usecs(TCP_TIMEOUT_INIT);
424 minmax_reset(&tp->rtt_min, tcp_jiffies32, ~0U);
426 /* So many TCP implementations out there (incorrectly) count the
427 * initial SYN frame in their delayed-ACK and congestion control
428 * algorithms that we must have the following bandaid to talk
429 * efficiently to them. -DaveM
431 tp->snd_cwnd = TCP_INIT_CWND;
433 /* There's a bubble in the pipe until at least the first ACK. */
434 tp->app_limited = ~0U;
436 /* See draft-stevens-tcpca-spec-01 for discussion of the
437 * initialization of these values.
439 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
440 tp->snd_cwnd_clamp = ~0;
441 tp->mss_cache = TCP_MSS_DEFAULT;
443 tp->reordering = sock_net(sk)->ipv4.sysctl_tcp_reordering;
444 tcp_assign_congestion_control(sk);
447 tp->rack.reo_wnd_steps = 1;
449 sk->sk_write_space = sk_stream_write_space;
450 sock_set_flag(sk, SOCK_USE_WRITE_QUEUE);
452 icsk->icsk_sync_mss = tcp_sync_mss;
454 WRITE_ONCE(sk->sk_sndbuf, sock_net(sk)->ipv4.sysctl_tcp_wmem[1]);
455 WRITE_ONCE(sk->sk_rcvbuf, sock_net(sk)->ipv4.sysctl_tcp_rmem[1]);
457 sk_sockets_allocated_inc(sk);
458 sk->sk_route_forced_caps = NETIF_F_GSO;
460 EXPORT_SYMBOL(tcp_init_sock);
462 static void tcp_tx_timestamp(struct sock *sk, u16 tsflags)
464 struct sk_buff *skb = tcp_write_queue_tail(sk);
466 if (tsflags && skb) {
467 struct skb_shared_info *shinfo = skb_shinfo(skb);
468 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
470 sock_tx_timestamp(sk, tsflags, &shinfo->tx_flags);
471 if (tsflags & SOF_TIMESTAMPING_TX_ACK)
472 tcb->txstamp_ack = 1;
473 if (tsflags & SOF_TIMESTAMPING_TX_RECORD_MASK)
474 shinfo->tskey = TCP_SKB_CB(skb)->seq + skb->len - 1;
478 static inline bool tcp_stream_is_readable(const struct tcp_sock *tp,
479 int target, struct sock *sk)
481 int avail = READ_ONCE(tp->rcv_nxt) - READ_ONCE(tp->copied_seq);
486 if (tcp_rmem_pressure(sk))
488 if (tcp_receive_window(tp) <= inet_csk(sk)->icsk_ack.rcv_mss)
491 if (sk->sk_prot->stream_memory_read)
492 return sk->sk_prot->stream_memory_read(sk);
497 * Wait for a TCP event.
499 * Note that we don't need to lock the socket, as the upper poll layers
500 * take care of normal races (between the test and the event) and we don't
501 * go look at any of the socket buffers directly.
503 __poll_t tcp_poll(struct file *file, struct socket *sock, poll_table *wait)
506 struct sock *sk = sock->sk;
507 const struct tcp_sock *tp = tcp_sk(sk);
510 sock_poll_wait(file, sock, wait);
512 state = inet_sk_state_load(sk);
513 if (state == TCP_LISTEN)
514 return inet_csk_listen_poll(sk);
516 /* Socket is not locked. We are protected from async events
517 * by poll logic and correct handling of state changes
518 * made by other threads is impossible in any case.
524 * EPOLLHUP is certainly not done right. But poll() doesn't
525 * have a notion of HUP in just one direction, and for a
526 * socket the read side is more interesting.
528 * Some poll() documentation says that EPOLLHUP is incompatible
529 * with the EPOLLOUT/POLLWR flags, so somebody should check this
530 * all. But careful, it tends to be safer to return too many
531 * bits than too few, and you can easily break real applications
532 * if you don't tell them that something has hung up!
536 * Check number 1. EPOLLHUP is _UNMASKABLE_ event (see UNIX98 and
537 * our fs/select.c). It means that after we received EOF,
538 * poll always returns immediately, making impossible poll() on write()
539 * in state CLOSE_WAIT. One solution is evident --- to set EPOLLHUP
540 * if and only if shutdown has been made in both directions.
541 * Actually, it is interesting to look how Solaris and DUX
542 * solve this dilemma. I would prefer, if EPOLLHUP were maskable,
543 * then we could set it on SND_SHUTDOWN. BTW examples given
544 * in Stevens' books assume exactly this behaviour, it explains
545 * why EPOLLHUP is incompatible with EPOLLOUT. --ANK
547 * NOTE. Check for TCP_CLOSE is added. The goal is to prevent
548 * blocking on fresh not-connected or disconnected socket. --ANK
550 if (sk->sk_shutdown == SHUTDOWN_MASK || state == TCP_CLOSE)
552 if (sk->sk_shutdown & RCV_SHUTDOWN)
553 mask |= EPOLLIN | EPOLLRDNORM | EPOLLRDHUP;
555 /* Connected or passive Fast Open socket? */
556 if (state != TCP_SYN_SENT &&
557 (state != TCP_SYN_RECV || rcu_access_pointer(tp->fastopen_rsk))) {
558 int target = sock_rcvlowat(sk, 0, INT_MAX);
560 if (READ_ONCE(tp->urg_seq) == READ_ONCE(tp->copied_seq) &&
561 !sock_flag(sk, SOCK_URGINLINE) &&
565 if (tcp_stream_is_readable(tp, target, sk))
566 mask |= EPOLLIN | EPOLLRDNORM;
568 if (!(sk->sk_shutdown & SEND_SHUTDOWN)) {
569 if (__sk_stream_is_writeable(sk, 1)) {
570 mask |= EPOLLOUT | EPOLLWRNORM;
571 } else { /* send SIGIO later */
572 sk_set_bit(SOCKWQ_ASYNC_NOSPACE, sk);
573 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
575 /* Race breaker. If space is freed after
576 * wspace test but before the flags are set,
577 * IO signal will be lost. Memory barrier
578 * pairs with the input side.
580 smp_mb__after_atomic();
581 if (__sk_stream_is_writeable(sk, 1))
582 mask |= EPOLLOUT | EPOLLWRNORM;
585 mask |= EPOLLOUT | EPOLLWRNORM;
587 if (tp->urg_data & TCP_URG_VALID)
589 } else if (state == TCP_SYN_SENT && inet_sk(sk)->defer_connect) {
590 /* Active TCP fastopen socket with defer_connect
591 * Return EPOLLOUT so application can call write()
592 * in order for kernel to generate SYN+data
594 mask |= EPOLLOUT | EPOLLWRNORM;
596 /* This barrier is coupled with smp_wmb() in tcp_reset() */
598 if (sk->sk_err || !skb_queue_empty_lockless(&sk->sk_error_queue))
603 EXPORT_SYMBOL(tcp_poll);
605 int tcp_ioctl(struct sock *sk, int cmd, unsigned long arg)
607 struct tcp_sock *tp = tcp_sk(sk);
613 if (sk->sk_state == TCP_LISTEN)
616 slow = lock_sock_fast(sk);
618 unlock_sock_fast(sk, slow);
621 answ = tp->urg_data &&
622 READ_ONCE(tp->urg_seq) == READ_ONCE(tp->copied_seq);
625 if (sk->sk_state == TCP_LISTEN)
628 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
631 answ = READ_ONCE(tp->write_seq) - tp->snd_una;
634 if (sk->sk_state == TCP_LISTEN)
637 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
640 answ = READ_ONCE(tp->write_seq) -
641 READ_ONCE(tp->snd_nxt);
647 return put_user(answ, (int __user *)arg);
649 EXPORT_SYMBOL(tcp_ioctl);
651 static inline void tcp_mark_push(struct tcp_sock *tp, struct sk_buff *skb)
653 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
654 tp->pushed_seq = tp->write_seq;
657 static inline bool forced_push(const struct tcp_sock *tp)
659 return after(tp->write_seq, tp->pushed_seq + (tp->max_window >> 1));
662 static void skb_entail(struct sock *sk, struct sk_buff *skb)
664 struct tcp_sock *tp = tcp_sk(sk);
665 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
668 tcb->seq = tcb->end_seq = tp->write_seq;
669 tcb->tcp_flags = TCPHDR_ACK;
671 __skb_header_release(skb);
672 tcp_add_write_queue_tail(sk, skb);
673 sk_wmem_queued_add(sk, skb->truesize);
674 sk_mem_charge(sk, skb->truesize);
675 if (tp->nonagle & TCP_NAGLE_PUSH)
676 tp->nonagle &= ~TCP_NAGLE_PUSH;
678 tcp_slow_start_after_idle_check(sk);
681 static inline void tcp_mark_urg(struct tcp_sock *tp, int flags)
684 tp->snd_up = tp->write_seq;
687 /* If a not yet filled skb is pushed, do not send it if
688 * we have data packets in Qdisc or NIC queues :
689 * Because TX completion will happen shortly, it gives a chance
690 * to coalesce future sendmsg() payload into this skb, without
691 * need for a timer, and with no latency trade off.
692 * As packets containing data payload have a bigger truesize
693 * than pure acks (dataless) packets, the last checks prevent
694 * autocorking if we only have an ACK in Qdisc/NIC queues,
695 * or if TX completion was delayed after we processed ACK packet.
697 static bool tcp_should_autocork(struct sock *sk, struct sk_buff *skb,
700 return skb->len < size_goal &&
701 sock_net(sk)->ipv4.sysctl_tcp_autocorking &&
702 !tcp_rtx_queue_empty(sk) &&
703 refcount_read(&sk->sk_wmem_alloc) > skb->truesize;
706 void tcp_push(struct sock *sk, int flags, int mss_now,
707 int nonagle, int size_goal)
709 struct tcp_sock *tp = tcp_sk(sk);
712 skb = tcp_write_queue_tail(sk);
715 if (!(flags & MSG_MORE) || forced_push(tp))
716 tcp_mark_push(tp, skb);
718 tcp_mark_urg(tp, flags);
720 if (tcp_should_autocork(sk, skb, size_goal)) {
722 /* avoid atomic op if TSQ_THROTTLED bit is already set */
723 if (!test_bit(TSQ_THROTTLED, &sk->sk_tsq_flags)) {
724 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPAUTOCORKING);
725 set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags);
727 /* It is possible TX completion already happened
728 * before we set TSQ_THROTTLED.
730 if (refcount_read(&sk->sk_wmem_alloc) > skb->truesize)
734 if (flags & MSG_MORE)
735 nonagle = TCP_NAGLE_CORK;
737 __tcp_push_pending_frames(sk, mss_now, nonagle);
740 static int tcp_splice_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb,
741 unsigned int offset, size_t len)
743 struct tcp_splice_state *tss = rd_desc->arg.data;
746 ret = skb_splice_bits(skb, skb->sk, offset, tss->pipe,
747 min(rd_desc->count, len), tss->flags);
749 rd_desc->count -= ret;
753 static int __tcp_splice_read(struct sock *sk, struct tcp_splice_state *tss)
755 /* Store TCP splice context information in read_descriptor_t. */
756 read_descriptor_t rd_desc = {
761 return tcp_read_sock(sk, &rd_desc, tcp_splice_data_recv);
765 * tcp_splice_read - splice data from TCP socket to a pipe
766 * @sock: socket to splice from
767 * @ppos: position (not valid)
768 * @pipe: pipe to splice to
769 * @len: number of bytes to splice
770 * @flags: splice modifier flags
773 * Will read pages from given socket and fill them into a pipe.
776 ssize_t tcp_splice_read(struct socket *sock, loff_t *ppos,
777 struct pipe_inode_info *pipe, size_t len,
780 struct sock *sk = sock->sk;
781 struct tcp_splice_state tss = {
790 sock_rps_record_flow(sk);
792 * We can't seek on a socket input
801 timeo = sock_rcvtimeo(sk, sock->file->f_flags & O_NONBLOCK);
803 ret = __tcp_splice_read(sk, &tss);
809 if (sock_flag(sk, SOCK_DONE))
812 ret = sock_error(sk);
815 if (sk->sk_shutdown & RCV_SHUTDOWN)
817 if (sk->sk_state == TCP_CLOSE) {
819 * This occurs when user tries to read
820 * from never connected socket.
829 /* if __tcp_splice_read() got nothing while we have
830 * an skb in receive queue, we do not want to loop.
831 * This might happen with URG data.
833 if (!skb_queue_empty(&sk->sk_receive_queue))
835 sk_wait_data(sk, &timeo, NULL);
836 if (signal_pending(current)) {
837 ret = sock_intr_errno(timeo);
850 if (sk->sk_err || sk->sk_state == TCP_CLOSE ||
851 (sk->sk_shutdown & RCV_SHUTDOWN) ||
852 signal_pending(current))
863 EXPORT_SYMBOL(tcp_splice_read);
865 struct sk_buff *sk_stream_alloc_skb(struct sock *sk, int size, gfp_t gfp,
871 skb = sk->sk_tx_skb_cache;
873 skb->truesize = SKB_TRUESIZE(skb_end_offset(skb));
874 sk->sk_tx_skb_cache = NULL;
876 INIT_LIST_HEAD(&skb->tcp_tsorted_anchor);
877 skb_shinfo(skb)->tx_flags = 0;
878 memset(TCP_SKB_CB(skb), 0, sizeof(struct tcp_skb_cb));
882 /* The TCP header must be at least 32-bit aligned. */
883 size = ALIGN(size, 4);
885 if (unlikely(tcp_under_memory_pressure(sk)))
886 sk_mem_reclaim_partial(sk);
888 skb = alloc_skb_fclone(size + sk->sk_prot->max_header, gfp);
892 if (force_schedule) {
893 mem_scheduled = true;
894 sk_forced_mem_schedule(sk, skb->truesize);
896 mem_scheduled = sk_wmem_schedule(sk, skb->truesize);
898 if (likely(mem_scheduled)) {
899 skb_reserve(skb, sk->sk_prot->max_header);
901 * Make sure that we have exactly size bytes
902 * available to the caller, no more, no less.
904 skb->reserved_tailroom = skb->end - skb->tail - size;
905 INIT_LIST_HEAD(&skb->tcp_tsorted_anchor);
910 sk->sk_prot->enter_memory_pressure(sk);
911 sk_stream_moderate_sndbuf(sk);
916 static unsigned int tcp_xmit_size_goal(struct sock *sk, u32 mss_now,
919 struct tcp_sock *tp = tcp_sk(sk);
920 u32 new_size_goal, size_goal;
925 /* Note : tcp_tso_autosize() will eventually split this later */
926 new_size_goal = sk->sk_gso_max_size - 1 - MAX_TCP_HEADER;
927 new_size_goal = tcp_bound_to_half_wnd(tp, new_size_goal);
929 /* We try hard to avoid divides here */
930 size_goal = tp->gso_segs * mss_now;
931 if (unlikely(new_size_goal < size_goal ||
932 new_size_goal >= size_goal + mss_now)) {
933 tp->gso_segs = min_t(u16, new_size_goal / mss_now,
934 sk->sk_gso_max_segs);
935 size_goal = tp->gso_segs * mss_now;
938 return max(size_goal, mss_now);
941 int tcp_send_mss(struct sock *sk, int *size_goal, int flags)
945 mss_now = tcp_current_mss(sk);
946 *size_goal = tcp_xmit_size_goal(sk, mss_now, !(flags & MSG_OOB));
951 /* In some cases, both sendpage() and sendmsg() could have added
952 * an skb to the write queue, but failed adding payload on it.
953 * We need to remove it to consume less memory, but more
954 * importantly be able to generate EPOLLOUT for Edge Trigger epoll()
957 void tcp_remove_empty_skb(struct sock *sk, struct sk_buff *skb)
959 if (skb && !skb->len) {
960 tcp_unlink_write_queue(skb, sk);
961 if (tcp_write_queue_empty(sk))
962 tcp_chrono_stop(sk, TCP_CHRONO_BUSY);
963 sk_wmem_free_skb(sk, skb);
967 struct sk_buff *tcp_build_frag(struct sock *sk, int size_goal, int flags,
968 struct page *page, int offset, size_t *size)
970 struct sk_buff *skb = tcp_write_queue_tail(sk);
971 struct tcp_sock *tp = tcp_sk(sk);
975 if (!skb || (copy = size_goal - skb->len) <= 0 ||
976 !tcp_skb_can_collapse_to(skb)) {
978 if (!sk_stream_memory_free(sk))
981 skb = sk_stream_alloc_skb(sk, 0, sk->sk_allocation,
982 tcp_rtx_and_write_queues_empty(sk));
986 #ifdef CONFIG_TLS_DEVICE
987 skb->decrypted = !!(flags & MSG_SENDPAGE_DECRYPTED);
996 i = skb_shinfo(skb)->nr_frags;
997 can_coalesce = skb_can_coalesce(skb, i, page, offset);
998 if (!can_coalesce && i >= sysctl_max_skb_frags) {
999 tcp_mark_push(tp, skb);
1002 if (!sk_wmem_schedule(sk, copy))
1006 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1009 skb_fill_page_desc(skb, i, page, offset, copy);
1012 if (!(flags & MSG_NO_SHARED_FRAGS))
1013 skb_shinfo(skb)->tx_flags |= SKBTX_SHARED_FRAG;
1016 skb->data_len += copy;
1017 skb->truesize += copy;
1018 sk_wmem_queued_add(sk, copy);
1019 sk_mem_charge(sk, copy);
1020 skb->ip_summed = CHECKSUM_PARTIAL;
1021 WRITE_ONCE(tp->write_seq, tp->write_seq + copy);
1022 TCP_SKB_CB(skb)->end_seq += copy;
1023 tcp_skb_pcount_set(skb, 0);
1029 ssize_t do_tcp_sendpages(struct sock *sk, struct page *page, int offset,
1030 size_t size, int flags)
1032 struct tcp_sock *tp = tcp_sk(sk);
1033 int mss_now, size_goal;
1036 long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
1038 if (IS_ENABLED(CONFIG_DEBUG_VM) &&
1039 WARN_ONCE(!sendpage_ok(page),
1040 "page must not be a Slab one and have page_count > 0"))
1043 /* Wait for a connection to finish. One exception is TCP Fast Open
1044 * (passive side) where data is allowed to be sent before a connection
1045 * is fully established.
1047 if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) &&
1048 !tcp_passive_fastopen(sk)) {
1049 err = sk_stream_wait_connect(sk, &timeo);
1054 sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
1056 mss_now = tcp_send_mss(sk, &size_goal, flags);
1060 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
1064 struct sk_buff *skb;
1067 skb = tcp_build_frag(sk, size_goal, flags, page, offset, ©);
1069 goto wait_for_space;
1072 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
1080 if (skb->len < size_goal || (flags & MSG_OOB))
1083 if (forced_push(tp)) {
1084 tcp_mark_push(tp, skb);
1085 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
1086 } else if (skb == tcp_send_head(sk))
1087 tcp_push_one(sk, mss_now);
1091 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1092 tcp_push(sk, flags & ~MSG_MORE, mss_now,
1093 TCP_NAGLE_PUSH, size_goal);
1095 err = sk_stream_wait_memory(sk, &timeo);
1099 mss_now = tcp_send_mss(sk, &size_goal, flags);
1104 tcp_tx_timestamp(sk, sk->sk_tsflags);
1105 if (!(flags & MSG_SENDPAGE_NOTLAST))
1106 tcp_push(sk, flags, mss_now, tp->nonagle, size_goal);
1111 tcp_remove_empty_skb(sk, tcp_write_queue_tail(sk));
1115 /* make sure we wake any epoll edge trigger waiter */
1116 if (unlikely(tcp_rtx_and_write_queues_empty(sk) && err == -EAGAIN)) {
1117 sk->sk_write_space(sk);
1118 tcp_chrono_stop(sk, TCP_CHRONO_SNDBUF_LIMITED);
1120 return sk_stream_error(sk, flags, err);
1122 EXPORT_SYMBOL_GPL(do_tcp_sendpages);
1124 int tcp_sendpage_locked(struct sock *sk, struct page *page, int offset,
1125 size_t size, int flags)
1127 if (!(sk->sk_route_caps & NETIF_F_SG))
1128 return sock_no_sendpage_locked(sk, page, offset, size, flags);
1130 tcp_rate_check_app_limited(sk); /* is sending application-limited? */
1132 return do_tcp_sendpages(sk, page, offset, size, flags);
1134 EXPORT_SYMBOL_GPL(tcp_sendpage_locked);
1136 int tcp_sendpage(struct sock *sk, struct page *page, int offset,
1137 size_t size, int flags)
1142 ret = tcp_sendpage_locked(sk, page, offset, size, flags);
1147 EXPORT_SYMBOL(tcp_sendpage);
1149 void tcp_free_fastopen_req(struct tcp_sock *tp)
1151 if (tp->fastopen_req) {
1152 kfree(tp->fastopen_req);
1153 tp->fastopen_req = NULL;
1157 static int tcp_sendmsg_fastopen(struct sock *sk, struct msghdr *msg,
1158 int *copied, size_t size,
1159 struct ubuf_info *uarg)
1161 struct tcp_sock *tp = tcp_sk(sk);
1162 struct inet_sock *inet = inet_sk(sk);
1163 struct sockaddr *uaddr = msg->msg_name;
1166 if (!(sock_net(sk)->ipv4.sysctl_tcp_fastopen & TFO_CLIENT_ENABLE) ||
1167 (uaddr && msg->msg_namelen >= sizeof(uaddr->sa_family) &&
1168 uaddr->sa_family == AF_UNSPEC))
1170 if (tp->fastopen_req)
1171 return -EALREADY; /* Another Fast Open is in progress */
1173 tp->fastopen_req = kzalloc(sizeof(struct tcp_fastopen_request),
1175 if (unlikely(!tp->fastopen_req))
1177 tp->fastopen_req->data = msg;
1178 tp->fastopen_req->size = size;
1179 tp->fastopen_req->uarg = uarg;
1181 if (inet->defer_connect) {
1182 err = tcp_connect(sk);
1183 /* Same failure procedure as in tcp_v4/6_connect */
1185 tcp_set_state(sk, TCP_CLOSE);
1186 inet->inet_dport = 0;
1187 sk->sk_route_caps = 0;
1190 flags = (msg->msg_flags & MSG_DONTWAIT) ? O_NONBLOCK : 0;
1191 err = __inet_stream_connect(sk->sk_socket, uaddr,
1192 msg->msg_namelen, flags, 1);
1193 /* fastopen_req could already be freed in __inet_stream_connect
1194 * if the connection times out or gets rst
1196 if (tp->fastopen_req) {
1197 *copied = tp->fastopen_req->copied;
1198 tcp_free_fastopen_req(tp);
1199 inet->defer_connect = 0;
1204 int tcp_sendmsg_locked(struct sock *sk, struct msghdr *msg, size_t size)
1206 struct tcp_sock *tp = tcp_sk(sk);
1207 struct ubuf_info *uarg = NULL;
1208 struct sk_buff *skb;
1209 struct sockcm_cookie sockc;
1210 int flags, err, copied = 0;
1211 int mss_now = 0, size_goal, copied_syn = 0;
1212 int process_backlog = 0;
1216 flags = msg->msg_flags;
1218 if (flags & MSG_ZEROCOPY && size && sock_flag(sk, SOCK_ZEROCOPY)) {
1219 skb = tcp_write_queue_tail(sk);
1220 uarg = sock_zerocopy_realloc(sk, size, skb_zcopy(skb));
1226 zc = sk->sk_route_caps & NETIF_F_SG;
1231 if (unlikely(flags & MSG_FASTOPEN || inet_sk(sk)->defer_connect) &&
1233 err = tcp_sendmsg_fastopen(sk, msg, &copied_syn, size, uarg);
1234 if (err == -EINPROGRESS && copied_syn > 0)
1240 timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
1242 tcp_rate_check_app_limited(sk); /* is sending application-limited? */
1244 /* Wait for a connection to finish. One exception is TCP Fast Open
1245 * (passive side) where data is allowed to be sent before a connection
1246 * is fully established.
1248 if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) &&
1249 !tcp_passive_fastopen(sk)) {
1250 err = sk_stream_wait_connect(sk, &timeo);
1255 if (unlikely(tp->repair)) {
1256 if (tp->repair_queue == TCP_RECV_QUEUE) {
1257 copied = tcp_send_rcvq(sk, msg, size);
1262 if (tp->repair_queue == TCP_NO_QUEUE)
1265 /* 'common' sending to sendq */
1268 sockcm_init(&sockc, sk);
1269 if (msg->msg_controllen) {
1270 err = sock_cmsg_send(sk, msg, &sockc);
1271 if (unlikely(err)) {
1277 /* This should be in poll */
1278 sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
1280 /* Ok commence sending. */
1284 mss_now = tcp_send_mss(sk, &size_goal, flags);
1287 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
1290 while (msg_data_left(msg)) {
1293 skb = tcp_write_queue_tail(sk);
1295 copy = size_goal - skb->len;
1297 if (copy <= 0 || !tcp_skb_can_collapse_to(skb)) {
1301 if (!sk_stream_memory_free(sk))
1302 goto wait_for_space;
1304 if (unlikely(process_backlog >= 16)) {
1305 process_backlog = 0;
1306 if (sk_flush_backlog(sk))
1309 first_skb = tcp_rtx_and_write_queues_empty(sk);
1310 skb = sk_stream_alloc_skb(sk, 0, sk->sk_allocation,
1313 goto wait_for_space;
1316 skb->ip_summed = CHECKSUM_PARTIAL;
1318 skb_entail(sk, skb);
1321 /* All packets are restored as if they have
1322 * already been sent. skb_mstamp_ns isn't set to
1323 * avoid wrong rtt estimation.
1326 TCP_SKB_CB(skb)->sacked |= TCPCB_REPAIRED;
1329 /* Try to append data to the end of skb. */
1330 if (copy > msg_data_left(msg))
1331 copy = msg_data_left(msg);
1333 /* Where to copy to? */
1334 if (skb_availroom(skb) > 0 && !zc) {
1335 /* We have some space in skb head. Superb! */
1336 copy = min_t(int, copy, skb_availroom(skb));
1337 err = skb_add_data_nocache(sk, skb, &msg->msg_iter, copy);
1342 int i = skb_shinfo(skb)->nr_frags;
1343 struct page_frag *pfrag = sk_page_frag(sk);
1345 if (!sk_page_frag_refill(sk, pfrag))
1346 goto wait_for_space;
1348 if (!skb_can_coalesce(skb, i, pfrag->page,
1350 if (i >= sysctl_max_skb_frags) {
1351 tcp_mark_push(tp, skb);
1357 copy = min_t(int, copy, pfrag->size - pfrag->offset);
1359 if (!sk_wmem_schedule(sk, copy))
1360 goto wait_for_space;
1362 err = skb_copy_to_page_nocache(sk, &msg->msg_iter, skb,
1369 /* Update the skb. */
1371 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1373 skb_fill_page_desc(skb, i, pfrag->page,
1374 pfrag->offset, copy);
1375 page_ref_inc(pfrag->page);
1377 pfrag->offset += copy;
1379 err = skb_zerocopy_iter_stream(sk, skb, msg, copy, uarg);
1380 if (err == -EMSGSIZE || err == -EEXIST) {
1381 tcp_mark_push(tp, skb);
1390 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
1392 WRITE_ONCE(tp->write_seq, tp->write_seq + copy);
1393 TCP_SKB_CB(skb)->end_seq += copy;
1394 tcp_skb_pcount_set(skb, 0);
1397 if (!msg_data_left(msg)) {
1398 if (unlikely(flags & MSG_EOR))
1399 TCP_SKB_CB(skb)->eor = 1;
1403 if (skb->len < size_goal || (flags & MSG_OOB) || unlikely(tp->repair))
1406 if (forced_push(tp)) {
1407 tcp_mark_push(tp, skb);
1408 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
1409 } else if (skb == tcp_send_head(sk))
1410 tcp_push_one(sk, mss_now);
1414 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1416 tcp_push(sk, flags & ~MSG_MORE, mss_now,
1417 TCP_NAGLE_PUSH, size_goal);
1419 err = sk_stream_wait_memory(sk, &timeo);
1423 mss_now = tcp_send_mss(sk, &size_goal, flags);
1428 tcp_tx_timestamp(sk, sockc.tsflags);
1429 tcp_push(sk, flags, mss_now, tp->nonagle, size_goal);
1432 sock_zerocopy_put(uarg);
1433 return copied + copied_syn;
1436 skb = tcp_write_queue_tail(sk);
1438 tcp_remove_empty_skb(sk, skb);
1440 if (copied + copied_syn)
1443 sock_zerocopy_put_abort(uarg, true);
1444 err = sk_stream_error(sk, flags, err);
1445 /* make sure we wake any epoll edge trigger waiter */
1446 if (unlikely(tcp_rtx_and_write_queues_empty(sk) && err == -EAGAIN)) {
1447 sk->sk_write_space(sk);
1448 tcp_chrono_stop(sk, TCP_CHRONO_SNDBUF_LIMITED);
1452 EXPORT_SYMBOL_GPL(tcp_sendmsg_locked);
1454 int tcp_sendmsg(struct sock *sk, struct msghdr *msg, size_t size)
1459 ret = tcp_sendmsg_locked(sk, msg, size);
1464 EXPORT_SYMBOL(tcp_sendmsg);
1467 * Handle reading urgent data. BSD has very simple semantics for
1468 * this, no blocking and very strange errors 8)
1471 static int tcp_recv_urg(struct sock *sk, struct msghdr *msg, int len, int flags)
1473 struct tcp_sock *tp = tcp_sk(sk);
1475 /* No URG data to read. */
1476 if (sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data ||
1477 tp->urg_data == TCP_URG_READ)
1478 return -EINVAL; /* Yes this is right ! */
1480 if (sk->sk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DONE))
1483 if (tp->urg_data & TCP_URG_VALID) {
1485 char c = tp->urg_data;
1487 if (!(flags & MSG_PEEK))
1488 tp->urg_data = TCP_URG_READ;
1490 /* Read urgent data. */
1491 msg->msg_flags |= MSG_OOB;
1494 if (!(flags & MSG_TRUNC))
1495 err = memcpy_to_msg(msg, &c, 1);
1498 msg->msg_flags |= MSG_TRUNC;
1500 return err ? -EFAULT : len;
1503 if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
1506 /* Fixed the recv(..., MSG_OOB) behaviour. BSD docs and
1507 * the available implementations agree in this case:
1508 * this call should never block, independent of the
1509 * blocking state of the socket.
1510 * Mike <pall@rz.uni-karlsruhe.de>
1515 static int tcp_peek_sndq(struct sock *sk, struct msghdr *msg, int len)
1517 struct sk_buff *skb;
1518 int copied = 0, err = 0;
1520 /* XXX -- need to support SO_PEEK_OFF */
1522 skb_rbtree_walk(skb, &sk->tcp_rtx_queue) {
1523 err = skb_copy_datagram_msg(skb, 0, msg, skb->len);
1529 skb_queue_walk(&sk->sk_write_queue, skb) {
1530 err = skb_copy_datagram_msg(skb, 0, msg, skb->len);
1537 return err ?: copied;
1540 /* Clean up the receive buffer for full frames taken by the user,
1541 * then send an ACK if necessary. COPIED is the number of bytes
1542 * tcp_recvmsg has given to the user so far, it speeds up the
1543 * calculation of whether or not we must ACK for the sake of
1546 void tcp_cleanup_rbuf(struct sock *sk, int copied)
1548 struct tcp_sock *tp = tcp_sk(sk);
1549 bool time_to_ack = false;
1551 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
1553 WARN(skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq),
1554 "cleanup rbuf bug: copied %X seq %X rcvnxt %X\n",
1555 tp->copied_seq, TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt);
1557 if (inet_csk_ack_scheduled(sk)) {
1558 const struct inet_connection_sock *icsk = inet_csk(sk);
1560 if (/* Once-per-two-segments ACK was not sent by tcp_input.c */
1561 tp->rcv_nxt - tp->rcv_wup > icsk->icsk_ack.rcv_mss ||
1563 * If this read emptied read buffer, we send ACK, if
1564 * connection is not bidirectional, user drained
1565 * receive buffer and there was a small segment
1569 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED2) ||
1570 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED) &&
1571 !inet_csk_in_pingpong_mode(sk))) &&
1572 !atomic_read(&sk->sk_rmem_alloc)))
1576 /* We send an ACK if we can now advertise a non-zero window
1577 * which has been raised "significantly".
1579 * Even if window raised up to infinity, do not send window open ACK
1580 * in states, where we will not receive more. It is useless.
1582 if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) {
1583 __u32 rcv_window_now = tcp_receive_window(tp);
1585 /* Optimize, __tcp_select_window() is not cheap. */
1586 if (2*rcv_window_now <= tp->window_clamp) {
1587 __u32 new_window = __tcp_select_window(sk);
1589 /* Send ACK now, if this read freed lots of space
1590 * in our buffer. Certainly, new_window is new window.
1591 * We can advertise it now, if it is not less than current one.
1592 * "Lots" means "at least twice" here.
1594 if (new_window && new_window >= 2 * rcv_window_now)
1602 static struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off)
1604 struct sk_buff *skb;
1607 while ((skb = skb_peek(&sk->sk_receive_queue)) != NULL) {
1608 offset = seq - TCP_SKB_CB(skb)->seq;
1609 if (unlikely(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
1610 pr_err_once("%s: found a SYN, please report !\n", __func__);
1613 if (offset < skb->len || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)) {
1617 /* This looks weird, but this can happen if TCP collapsing
1618 * splitted a fat GRO packet, while we released socket lock
1619 * in skb_splice_bits()
1621 sk_eat_skb(sk, skb);
1627 * This routine provides an alternative to tcp_recvmsg() for routines
1628 * that would like to handle copying from skbuffs directly in 'sendfile'
1631 * - It is assumed that the socket was locked by the caller.
1632 * - The routine does not block.
1633 * - At present, there is no support for reading OOB data
1634 * or for 'peeking' the socket using this routine
1635 * (although both would be easy to implement).
1637 int tcp_read_sock(struct sock *sk, read_descriptor_t *desc,
1638 sk_read_actor_t recv_actor)
1640 struct sk_buff *skb;
1641 struct tcp_sock *tp = tcp_sk(sk);
1642 u32 seq = tp->copied_seq;
1646 if (sk->sk_state == TCP_LISTEN)
1648 while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1649 if (offset < skb->len) {
1653 len = skb->len - offset;
1654 /* Stop reading if we hit a patch of urgent data */
1656 u32 urg_offset = tp->urg_seq - seq;
1657 if (urg_offset < len)
1662 used = recv_actor(desc, skb, offset, len);
1667 } else if (used <= len) {
1672 /* If recv_actor drops the lock (e.g. TCP splice
1673 * receive) the skb pointer might be invalid when
1674 * getting here: tcp_collapse might have deleted it
1675 * while aggregating skbs from the socket queue.
1677 skb = tcp_recv_skb(sk, seq - 1, &offset);
1680 /* TCP coalescing might have appended data to the skb.
1681 * Try to splice more frags
1683 if (offset + 1 != skb->len)
1686 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) {
1687 sk_eat_skb(sk, skb);
1691 sk_eat_skb(sk, skb);
1694 WRITE_ONCE(tp->copied_seq, seq);
1696 WRITE_ONCE(tp->copied_seq, seq);
1698 tcp_rcv_space_adjust(sk);
1700 /* Clean up data we have read: This will do ACK frames. */
1702 tcp_recv_skb(sk, seq, &offset);
1703 tcp_cleanup_rbuf(sk, copied);
1707 EXPORT_SYMBOL(tcp_read_sock);
1709 int tcp_peek_len(struct socket *sock)
1711 return tcp_inq(sock->sk);
1713 EXPORT_SYMBOL(tcp_peek_len);
1715 /* Make sure sk_rcvbuf is big enough to satisfy SO_RCVLOWAT hint */
1716 int tcp_set_rcvlowat(struct sock *sk, int val)
1720 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK)
1721 cap = sk->sk_rcvbuf >> 1;
1723 cap = sock_net(sk)->ipv4.sysctl_tcp_rmem[2] >> 1;
1724 val = min(val, cap);
1725 WRITE_ONCE(sk->sk_rcvlowat, val ? : 1);
1727 /* Check if we need to signal EPOLLIN right now */
1730 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK)
1734 if (val > sk->sk_rcvbuf) {
1735 WRITE_ONCE(sk->sk_rcvbuf, val);
1736 tcp_sk(sk)->window_clamp = tcp_win_from_space(sk, val);
1740 EXPORT_SYMBOL(tcp_set_rcvlowat);
1743 static const struct vm_operations_struct tcp_vm_ops = {
1746 int tcp_mmap(struct file *file, struct socket *sock,
1747 struct vm_area_struct *vma)
1749 if (vma->vm_flags & (VM_WRITE | VM_EXEC))
1751 vma->vm_flags &= ~(VM_MAYWRITE | VM_MAYEXEC);
1753 /* Instruct vm_insert_page() to not mmap_read_lock(mm) */
1754 vma->vm_flags |= VM_MIXEDMAP;
1756 vma->vm_ops = &tcp_vm_ops;
1759 EXPORT_SYMBOL(tcp_mmap);
1761 static skb_frag_t *skb_advance_to_frag(struct sk_buff *skb, u32 offset_skb,
1766 offset_skb -= skb_headlen(skb);
1767 if ((int)offset_skb < 0 || skb_has_frag_list(skb))
1770 frag = skb_shinfo(skb)->frags;
1771 while (offset_skb) {
1772 if (skb_frag_size(frag) > offset_skb) {
1773 *offset_frag = offset_skb;
1776 offset_skb -= skb_frag_size(frag);
1783 static bool can_map_frag(const skb_frag_t *frag)
1785 return skb_frag_size(frag) == PAGE_SIZE && !skb_frag_off(frag);
1788 static int find_next_mappable_frag(const skb_frag_t *frag,
1789 int remaining_in_skb)
1793 if (likely(can_map_frag(frag)))
1796 while (offset < remaining_in_skb && !can_map_frag(frag)) {
1797 offset += skb_frag_size(frag);
1803 static void tcp_zerocopy_set_hint_for_skb(struct sock *sk,
1804 struct tcp_zerocopy_receive *zc,
1805 struct sk_buff *skb, u32 offset)
1807 u32 frag_offset, partial_frag_remainder = 0;
1808 int mappable_offset;
1811 /* worst case: skip to next skb. try to improve on this case below */
1812 zc->recv_skip_hint = skb->len - offset;
1814 /* Find the frag containing this offset (and how far into that frag) */
1815 frag = skb_advance_to_frag(skb, offset, &frag_offset);
1820 struct skb_shared_info *info = skb_shinfo(skb);
1822 /* We read part of the last frag, must recvmsg() rest of skb. */
1823 if (frag == &info->frags[info->nr_frags - 1])
1826 /* Else, we must at least read the remainder in this frag. */
1827 partial_frag_remainder = skb_frag_size(frag) - frag_offset;
1828 zc->recv_skip_hint -= partial_frag_remainder;
1832 /* partial_frag_remainder: If part way through a frag, must read rest.
1833 * mappable_offset: Bytes till next mappable frag, *not* counting bytes
1834 * in partial_frag_remainder.
1836 mappable_offset = find_next_mappable_frag(frag, zc->recv_skip_hint);
1837 zc->recv_skip_hint = mappable_offset + partial_frag_remainder;
1840 static int tcp_recvmsg_locked(struct sock *sk, struct msghdr *msg, size_t len,
1841 int nonblock, int flags,
1842 struct scm_timestamping_internal *tss,
1844 static int receive_fallback_to_copy(struct sock *sk,
1845 struct tcp_zerocopy_receive *zc, int inq)
1847 unsigned long copy_address = (unsigned long)zc->copybuf_address;
1848 struct scm_timestamping_internal tss_unused;
1849 int err, cmsg_flags_unused;
1850 struct msghdr msg = {};
1854 zc->recv_skip_hint = 0;
1856 if (copy_address != zc->copybuf_address)
1859 err = import_single_range(READ, (void __user *)copy_address,
1860 inq, &iov, &msg.msg_iter);
1864 err = tcp_recvmsg_locked(sk, &msg, inq, /*nonblock=*/1, /*flags=*/0,
1865 &tss_unused, &cmsg_flags_unused);
1869 zc->copybuf_len = err;
1870 if (likely(zc->copybuf_len)) {
1871 struct sk_buff *skb;
1874 skb = tcp_recv_skb(sk, tcp_sk(sk)->copied_seq, &offset);
1876 tcp_zerocopy_set_hint_for_skb(sk, zc, skb, offset);
1881 static int tcp_copy_straggler_data(struct tcp_zerocopy_receive *zc,
1882 struct sk_buff *skb, u32 copylen,
1883 u32 *offset, u32 *seq)
1885 unsigned long copy_address = (unsigned long)zc->copybuf_address;
1886 struct msghdr msg = {};
1890 if (copy_address != zc->copybuf_address)
1893 err = import_single_range(READ, (void __user *)copy_address,
1894 copylen, &iov, &msg.msg_iter);
1897 err = skb_copy_datagram_msg(skb, *offset, &msg, copylen);
1900 zc->recv_skip_hint -= copylen;
1903 return (__s32)copylen;
1906 static int tcp_zerocopy_handle_leftover_data(struct tcp_zerocopy_receive *zc,
1908 struct sk_buff *skb,
1912 u32 offset, copylen = min_t(u32, copybuf_len, zc->recv_skip_hint);
1916 /* skb is null if inq < PAGE_SIZE. */
1918 offset = *seq - TCP_SKB_CB(skb)->seq;
1920 skb = tcp_recv_skb(sk, *seq, &offset);
1922 zc->copybuf_len = tcp_copy_straggler_data(zc, skb, copylen, &offset,
1924 return zc->copybuf_len < 0 ? 0 : copylen;
1927 static int tcp_zerocopy_vm_insert_batch_error(struct vm_area_struct *vma,
1928 struct page **pending_pages,
1929 unsigned long pages_remaining,
1930 unsigned long *address,
1933 struct tcp_zerocopy_receive *zc,
1934 u32 total_bytes_to_map,
1937 /* At least one page did not map. Try zapping if we skipped earlier. */
1938 if (err == -EBUSY &&
1939 zc->flags & TCP_RECEIVE_ZEROCOPY_FLAG_TLB_CLEAN_HINT) {
1942 maybe_zap_len = total_bytes_to_map - /* All bytes to map */
1943 *length + /* Mapped or pending */
1944 (pages_remaining * PAGE_SIZE); /* Failed map. */
1945 zap_page_range(vma, *address, maybe_zap_len);
1950 unsigned long leftover_pages = pages_remaining;
1953 /* We called zap_page_range, try to reinsert. */
1954 err = vm_insert_pages(vma, *address,
1957 bytes_mapped = PAGE_SIZE * (leftover_pages - pages_remaining);
1958 *seq += bytes_mapped;
1959 *address += bytes_mapped;
1962 /* Either we were unable to zap, OR we zapped, retried an
1963 * insert, and still had an issue. Either ways, pages_remaining
1964 * is the number of pages we were unable to map, and we unroll
1965 * some state we speculatively touched before.
1967 const int bytes_not_mapped = PAGE_SIZE * pages_remaining;
1969 *length -= bytes_not_mapped;
1970 zc->recv_skip_hint += bytes_not_mapped;
1975 static int tcp_zerocopy_vm_insert_batch(struct vm_area_struct *vma,
1976 struct page **pages,
1977 unsigned int pages_to_map,
1978 unsigned long *address,
1981 struct tcp_zerocopy_receive *zc,
1982 u32 total_bytes_to_map)
1984 unsigned long pages_remaining = pages_to_map;
1985 unsigned int pages_mapped;
1986 unsigned int bytes_mapped;
1989 err = vm_insert_pages(vma, *address, pages, &pages_remaining);
1990 pages_mapped = pages_to_map - (unsigned int)pages_remaining;
1991 bytes_mapped = PAGE_SIZE * pages_mapped;
1992 /* Even if vm_insert_pages fails, it may have partially succeeded in
1993 * mapping (some but not all of the pages).
1995 *seq += bytes_mapped;
1996 *address += bytes_mapped;
2001 /* Error: maybe zap and retry + rollback state for failed inserts. */
2002 return tcp_zerocopy_vm_insert_batch_error(vma, pages + pages_mapped,
2003 pages_remaining, address, length, seq, zc, total_bytes_to_map,
2007 #define TCP_ZEROCOPY_PAGE_BATCH_SIZE 32
2008 static int tcp_zerocopy_receive(struct sock *sk,
2009 struct tcp_zerocopy_receive *zc)
2011 u32 length = 0, offset, vma_len, avail_len, copylen = 0;
2012 unsigned long address = (unsigned long)zc->address;
2013 struct page *pages[TCP_ZEROCOPY_PAGE_BATCH_SIZE];
2014 s32 copybuf_len = zc->copybuf_len;
2015 struct tcp_sock *tp = tcp_sk(sk);
2016 const skb_frag_t *frags = NULL;
2017 unsigned int pages_to_map = 0;
2018 struct vm_area_struct *vma;
2019 struct sk_buff *skb = NULL;
2020 u32 seq = tp->copied_seq;
2021 u32 total_bytes_to_map;
2022 int inq = tcp_inq(sk);
2025 zc->copybuf_len = 0;
2027 if (address & (PAGE_SIZE - 1) || address != zc->address)
2030 if (sk->sk_state == TCP_LISTEN)
2033 sock_rps_record_flow(sk);
2035 if (inq && inq <= copybuf_len)
2036 return receive_fallback_to_copy(sk, zc, inq);
2038 if (inq < PAGE_SIZE) {
2040 zc->recv_skip_hint = inq;
2041 if (!inq && sock_flag(sk, SOCK_DONE))
2046 mmap_read_lock(current->mm);
2048 vma = find_vma(current->mm, address);
2049 if (!vma || vma->vm_start > address || vma->vm_ops != &tcp_vm_ops) {
2050 mmap_read_unlock(current->mm);
2053 vma_len = min_t(unsigned long, zc->length, vma->vm_end - address);
2054 avail_len = min_t(u32, vma_len, inq);
2055 total_bytes_to_map = avail_len & ~(PAGE_SIZE - 1);
2056 if (total_bytes_to_map) {
2057 if (!(zc->flags & TCP_RECEIVE_ZEROCOPY_FLAG_TLB_CLEAN_HINT))
2058 zap_page_range(vma, address, total_bytes_to_map);
2059 zc->length = total_bytes_to_map;
2060 zc->recv_skip_hint = 0;
2062 zc->length = avail_len;
2063 zc->recv_skip_hint = avail_len;
2066 while (length + PAGE_SIZE <= zc->length) {
2067 int mappable_offset;
2070 if (zc->recv_skip_hint < PAGE_SIZE) {
2074 if (zc->recv_skip_hint > 0)
2077 offset = seq - TCP_SKB_CB(skb)->seq;
2079 skb = tcp_recv_skb(sk, seq, &offset);
2081 zc->recv_skip_hint = skb->len - offset;
2082 frags = skb_advance_to_frag(skb, offset, &offset_frag);
2083 if (!frags || offset_frag)
2087 mappable_offset = find_next_mappable_frag(frags,
2088 zc->recv_skip_hint);
2089 if (mappable_offset) {
2090 zc->recv_skip_hint = mappable_offset;
2093 page = skb_frag_page(frags);
2095 pages[pages_to_map++] = page;
2096 length += PAGE_SIZE;
2097 zc->recv_skip_hint -= PAGE_SIZE;
2099 if (pages_to_map == TCP_ZEROCOPY_PAGE_BATCH_SIZE ||
2100 zc->recv_skip_hint < PAGE_SIZE) {
2101 /* Either full batch, or we're about to go to next skb
2102 * (and we cannot unroll failed ops across skbs).
2104 ret = tcp_zerocopy_vm_insert_batch(vma, pages,
2108 total_bytes_to_map);
2115 ret = tcp_zerocopy_vm_insert_batch(vma, pages, pages_to_map,
2116 &address, &length, &seq,
2117 zc, total_bytes_to_map);
2120 mmap_read_unlock(current->mm);
2121 /* Try to copy straggler data. */
2123 copylen = tcp_zerocopy_handle_leftover_data(zc, sk, skb, &seq,
2126 if (length + copylen) {
2127 WRITE_ONCE(tp->copied_seq, seq);
2128 tcp_rcv_space_adjust(sk);
2130 /* Clean up data we have read: This will do ACK frames. */
2131 tcp_recv_skb(sk, seq, &offset);
2132 tcp_cleanup_rbuf(sk, length + copylen);
2134 if (length == zc->length)
2135 zc->recv_skip_hint = 0;
2137 if (!zc->recv_skip_hint && sock_flag(sk, SOCK_DONE))
2140 zc->length = length;
2145 static void tcp_update_recv_tstamps(struct sk_buff *skb,
2146 struct scm_timestamping_internal *tss)
2149 tss->ts[0] = ktime_to_timespec64(skb->tstamp);
2151 tss->ts[0] = (struct timespec64) {0};
2153 if (skb_hwtstamps(skb)->hwtstamp)
2154 tss->ts[2] = ktime_to_timespec64(skb_hwtstamps(skb)->hwtstamp);
2156 tss->ts[2] = (struct timespec64) {0};
2159 /* Similar to __sock_recv_timestamp, but does not require an skb */
2160 static void tcp_recv_timestamp(struct msghdr *msg, const struct sock *sk,
2161 struct scm_timestamping_internal *tss)
2163 int new_tstamp = sock_flag(sk, SOCK_TSTAMP_NEW);
2164 bool has_timestamping = false;
2166 if (tss->ts[0].tv_sec || tss->ts[0].tv_nsec) {
2167 if (sock_flag(sk, SOCK_RCVTSTAMP)) {
2168 if (sock_flag(sk, SOCK_RCVTSTAMPNS)) {
2170 struct __kernel_timespec kts = {
2171 .tv_sec = tss->ts[0].tv_sec,
2172 .tv_nsec = tss->ts[0].tv_nsec,
2174 put_cmsg(msg, SOL_SOCKET, SO_TIMESTAMPNS_NEW,
2177 struct __kernel_old_timespec ts_old = {
2178 .tv_sec = tss->ts[0].tv_sec,
2179 .tv_nsec = tss->ts[0].tv_nsec,
2181 put_cmsg(msg, SOL_SOCKET, SO_TIMESTAMPNS_OLD,
2182 sizeof(ts_old), &ts_old);
2186 struct __kernel_sock_timeval stv = {
2187 .tv_sec = tss->ts[0].tv_sec,
2188 .tv_usec = tss->ts[0].tv_nsec / 1000,
2190 put_cmsg(msg, SOL_SOCKET, SO_TIMESTAMP_NEW,
2193 struct __kernel_old_timeval tv = {
2194 .tv_sec = tss->ts[0].tv_sec,
2195 .tv_usec = tss->ts[0].tv_nsec / 1000,
2197 put_cmsg(msg, SOL_SOCKET, SO_TIMESTAMP_OLD,
2203 if (sk->sk_tsflags & SOF_TIMESTAMPING_SOFTWARE)
2204 has_timestamping = true;
2206 tss->ts[0] = (struct timespec64) {0};
2209 if (tss->ts[2].tv_sec || tss->ts[2].tv_nsec) {
2210 if (sk->sk_tsflags & SOF_TIMESTAMPING_RAW_HARDWARE)
2211 has_timestamping = true;
2213 tss->ts[2] = (struct timespec64) {0};
2216 if (has_timestamping) {
2217 tss->ts[1] = (struct timespec64) {0};
2218 if (sock_flag(sk, SOCK_TSTAMP_NEW))
2219 put_cmsg_scm_timestamping64(msg, tss);
2221 put_cmsg_scm_timestamping(msg, tss);
2225 static int tcp_inq_hint(struct sock *sk)
2227 const struct tcp_sock *tp = tcp_sk(sk);
2228 u32 copied_seq = READ_ONCE(tp->copied_seq);
2229 u32 rcv_nxt = READ_ONCE(tp->rcv_nxt);
2232 inq = rcv_nxt - copied_seq;
2233 if (unlikely(inq < 0 || copied_seq != READ_ONCE(tp->copied_seq))) {
2235 inq = tp->rcv_nxt - tp->copied_seq;
2238 /* After receiving a FIN, tell the user-space to continue reading
2239 * by returning a non-zero inq.
2241 if (inq == 0 && sock_flag(sk, SOCK_DONE))
2247 * This routine copies from a sock struct into the user buffer.
2249 * Technical note: in 2.3 we work on _locked_ socket, so that
2250 * tricks with *seq access order and skb->users are not required.
2251 * Probably, code can be easily improved even more.
2254 static int tcp_recvmsg_locked(struct sock *sk, struct msghdr *msg, size_t len,
2255 int nonblock, int flags,
2256 struct scm_timestamping_internal *tss,
2259 struct tcp_sock *tp = tcp_sk(sk);
2265 int target; /* Read at least this many bytes */
2267 struct sk_buff *skb, *last;
2271 if (sk->sk_state == TCP_LISTEN)
2274 if (tp->recvmsg_inq)
2276 timeo = sock_rcvtimeo(sk, nonblock);
2278 /* Urgent data needs to be handled specially. */
2279 if (flags & MSG_OOB)
2282 if (unlikely(tp->repair)) {
2284 if (!(flags & MSG_PEEK))
2287 if (tp->repair_queue == TCP_SEND_QUEUE)
2291 if (tp->repair_queue == TCP_NO_QUEUE)
2294 /* 'common' recv queue MSG_PEEK-ing */
2297 seq = &tp->copied_seq;
2298 if (flags & MSG_PEEK) {
2299 peek_seq = tp->copied_seq;
2303 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
2308 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
2309 if (tp->urg_data && tp->urg_seq == *seq) {
2312 if (signal_pending(current)) {
2313 copied = timeo ? sock_intr_errno(timeo) : -EAGAIN;
2318 /* Next get a buffer. */
2320 last = skb_peek_tail(&sk->sk_receive_queue);
2321 skb_queue_walk(&sk->sk_receive_queue, skb) {
2323 /* Now that we have two receive queues this
2326 if (WARN(before(*seq, TCP_SKB_CB(skb)->seq),
2327 "TCP recvmsg seq # bug: copied %X, seq %X, rcvnxt %X, fl %X\n",
2328 *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt,
2332 offset = *seq - TCP_SKB_CB(skb)->seq;
2333 if (unlikely(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
2334 pr_err_once("%s: found a SYN, please report !\n", __func__);
2337 if (offset < skb->len)
2339 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
2341 WARN(!(flags & MSG_PEEK),
2342 "TCP recvmsg seq # bug 2: copied %X, seq %X, rcvnxt %X, fl %X\n",
2343 *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt, flags);
2346 /* Well, if we have backlog, try to process it now yet. */
2348 if (copied >= target && !READ_ONCE(sk->sk_backlog.tail))
2353 sk->sk_state == TCP_CLOSE ||
2354 (sk->sk_shutdown & RCV_SHUTDOWN) ||
2356 signal_pending(current))
2359 if (sock_flag(sk, SOCK_DONE))
2363 copied = sock_error(sk);
2367 if (sk->sk_shutdown & RCV_SHUTDOWN)
2370 if (sk->sk_state == TCP_CLOSE) {
2371 /* This occurs when user tries to read
2372 * from never connected socket.
2383 if (signal_pending(current)) {
2384 copied = sock_intr_errno(timeo);
2389 tcp_cleanup_rbuf(sk, copied);
2391 if (copied >= target) {
2392 /* Do not sleep, just process backlog. */
2396 sk_wait_data(sk, &timeo, last);
2399 if ((flags & MSG_PEEK) &&
2400 (peek_seq - copied - urg_hole != tp->copied_seq)) {
2401 net_dbg_ratelimited("TCP(%s:%d): Application bug, race in MSG_PEEK\n",
2403 task_pid_nr(current));
2404 peek_seq = tp->copied_seq;
2409 /* Ok so how much can we use? */
2410 used = skb->len - offset;
2414 /* Do we have urgent data here? */
2416 u32 urg_offset = tp->urg_seq - *seq;
2417 if (urg_offset < used) {
2419 if (!sock_flag(sk, SOCK_URGINLINE)) {
2420 WRITE_ONCE(*seq, *seq + 1);
2432 if (!(flags & MSG_TRUNC)) {
2433 err = skb_copy_datagram_msg(skb, offset, msg, used);
2435 /* Exception. Bailout! */
2442 WRITE_ONCE(*seq, *seq + used);
2446 tcp_rcv_space_adjust(sk);
2449 if (tp->urg_data && after(tp->copied_seq, tp->urg_seq)) {
2451 tcp_fast_path_check(sk);
2454 if (TCP_SKB_CB(skb)->has_rxtstamp) {
2455 tcp_update_recv_tstamps(skb, tss);
2459 if (used + offset < skb->len)
2462 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
2464 if (!(flags & MSG_PEEK))
2465 sk_eat_skb(sk, skb);
2469 /* Process the FIN. */
2470 WRITE_ONCE(*seq, *seq + 1);
2471 if (!(flags & MSG_PEEK))
2472 sk_eat_skb(sk, skb);
2476 /* According to UNIX98, msg_name/msg_namelen are ignored
2477 * on connected socket. I was just happy when found this 8) --ANK
2480 /* Clean up data we have read: This will do ACK frames. */
2481 tcp_cleanup_rbuf(sk, copied);
2488 err = tcp_recv_urg(sk, msg, len, flags);
2492 err = tcp_peek_sndq(sk, msg, len);
2496 int tcp_recvmsg(struct sock *sk, struct msghdr *msg, size_t len, int nonblock,
2497 int flags, int *addr_len)
2499 int cmsg_flags = 0, ret, inq;
2500 struct scm_timestamping_internal tss;
2502 if (unlikely(flags & MSG_ERRQUEUE))
2503 return inet_recv_error(sk, msg, len, addr_len);
2505 if (sk_can_busy_loop(sk) &&
2506 skb_queue_empty_lockless(&sk->sk_receive_queue) &&
2507 sk->sk_state == TCP_ESTABLISHED)
2508 sk_busy_loop(sk, nonblock);
2511 ret = tcp_recvmsg_locked(sk, msg, len, nonblock, flags, &tss,
2515 if (cmsg_flags && ret >= 0) {
2517 tcp_recv_timestamp(msg, sk, &tss);
2518 if (cmsg_flags & 1) {
2519 inq = tcp_inq_hint(sk);
2520 put_cmsg(msg, SOL_TCP, TCP_CM_INQ, sizeof(inq), &inq);
2525 EXPORT_SYMBOL(tcp_recvmsg);
2527 void tcp_set_state(struct sock *sk, int state)
2529 int oldstate = sk->sk_state;
2531 /* We defined a new enum for TCP states that are exported in BPF
2532 * so as not force the internal TCP states to be frozen. The
2533 * following checks will detect if an internal state value ever
2534 * differs from the BPF value. If this ever happens, then we will
2535 * need to remap the internal value to the BPF value before calling
2536 * tcp_call_bpf_2arg.
2538 BUILD_BUG_ON((int)BPF_TCP_ESTABLISHED != (int)TCP_ESTABLISHED);
2539 BUILD_BUG_ON((int)BPF_TCP_SYN_SENT != (int)TCP_SYN_SENT);
2540 BUILD_BUG_ON((int)BPF_TCP_SYN_RECV != (int)TCP_SYN_RECV);
2541 BUILD_BUG_ON((int)BPF_TCP_FIN_WAIT1 != (int)TCP_FIN_WAIT1);
2542 BUILD_BUG_ON((int)BPF_TCP_FIN_WAIT2 != (int)TCP_FIN_WAIT2);
2543 BUILD_BUG_ON((int)BPF_TCP_TIME_WAIT != (int)TCP_TIME_WAIT);
2544 BUILD_BUG_ON((int)BPF_TCP_CLOSE != (int)TCP_CLOSE);
2545 BUILD_BUG_ON((int)BPF_TCP_CLOSE_WAIT != (int)TCP_CLOSE_WAIT);
2546 BUILD_BUG_ON((int)BPF_TCP_LAST_ACK != (int)TCP_LAST_ACK);
2547 BUILD_BUG_ON((int)BPF_TCP_LISTEN != (int)TCP_LISTEN);
2548 BUILD_BUG_ON((int)BPF_TCP_CLOSING != (int)TCP_CLOSING);
2549 BUILD_BUG_ON((int)BPF_TCP_NEW_SYN_RECV != (int)TCP_NEW_SYN_RECV);
2550 BUILD_BUG_ON((int)BPF_TCP_MAX_STATES != (int)TCP_MAX_STATES);
2552 if (BPF_SOCK_OPS_TEST_FLAG(tcp_sk(sk), BPF_SOCK_OPS_STATE_CB_FLAG))
2553 tcp_call_bpf_2arg(sk, BPF_SOCK_OPS_STATE_CB, oldstate, state);
2556 case TCP_ESTABLISHED:
2557 if (oldstate != TCP_ESTABLISHED)
2558 TCP_INC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
2562 if (oldstate == TCP_CLOSE_WAIT || oldstate == TCP_ESTABLISHED)
2563 TCP_INC_STATS(sock_net(sk), TCP_MIB_ESTABRESETS);
2565 sk->sk_prot->unhash(sk);
2566 if (inet_csk(sk)->icsk_bind_hash &&
2567 !(sk->sk_userlocks & SOCK_BINDPORT_LOCK))
2571 if (oldstate == TCP_ESTABLISHED)
2572 TCP_DEC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
2575 /* Change state AFTER socket is unhashed to avoid closed
2576 * socket sitting in hash tables.
2578 inet_sk_state_store(sk, state);
2580 EXPORT_SYMBOL_GPL(tcp_set_state);
2583 * State processing on a close. This implements the state shift for
2584 * sending our FIN frame. Note that we only send a FIN for some
2585 * states. A shutdown() may have already sent the FIN, or we may be
2589 static const unsigned char new_state[16] = {
2590 /* current state: new state: action: */
2591 [0 /* (Invalid) */] = TCP_CLOSE,
2592 [TCP_ESTABLISHED] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2593 [TCP_SYN_SENT] = TCP_CLOSE,
2594 [TCP_SYN_RECV] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2595 [TCP_FIN_WAIT1] = TCP_FIN_WAIT1,
2596 [TCP_FIN_WAIT2] = TCP_FIN_WAIT2,
2597 [TCP_TIME_WAIT] = TCP_CLOSE,
2598 [TCP_CLOSE] = TCP_CLOSE,
2599 [TCP_CLOSE_WAIT] = TCP_LAST_ACK | TCP_ACTION_FIN,
2600 [TCP_LAST_ACK] = TCP_LAST_ACK,
2601 [TCP_LISTEN] = TCP_CLOSE,
2602 [TCP_CLOSING] = TCP_CLOSING,
2603 [TCP_NEW_SYN_RECV] = TCP_CLOSE, /* should not happen ! */
2606 static int tcp_close_state(struct sock *sk)
2608 int next = (int)new_state[sk->sk_state];
2609 int ns = next & TCP_STATE_MASK;
2611 tcp_set_state(sk, ns);
2613 return next & TCP_ACTION_FIN;
2617 * Shutdown the sending side of a connection. Much like close except
2618 * that we don't receive shut down or sock_set_flag(sk, SOCK_DEAD).
2621 void tcp_shutdown(struct sock *sk, int how)
2623 /* We need to grab some memory, and put together a FIN,
2624 * and then put it into the queue to be sent.
2625 * Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
2627 if (!(how & SEND_SHUTDOWN))
2630 /* If we've already sent a FIN, or it's a closed state, skip this. */
2631 if ((1 << sk->sk_state) &
2632 (TCPF_ESTABLISHED | TCPF_SYN_SENT |
2633 TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) {
2634 /* Clear out any half completed packets. FIN if needed. */
2635 if (tcp_close_state(sk))
2639 EXPORT_SYMBOL(tcp_shutdown);
2641 bool tcp_check_oom(struct sock *sk, int shift)
2643 bool too_many_orphans, out_of_socket_memory;
2645 too_many_orphans = tcp_too_many_orphans(sk, shift);
2646 out_of_socket_memory = tcp_out_of_memory(sk);
2648 if (too_many_orphans)
2649 net_info_ratelimited("too many orphaned sockets\n");
2650 if (out_of_socket_memory)
2651 net_info_ratelimited("out of memory -- consider tuning tcp_mem\n");
2652 return too_many_orphans || out_of_socket_memory;
2655 void __tcp_close(struct sock *sk, long timeout)
2657 struct sk_buff *skb;
2658 int data_was_unread = 0;
2661 sk->sk_shutdown = SHUTDOWN_MASK;
2663 if (sk->sk_state == TCP_LISTEN) {
2664 tcp_set_state(sk, TCP_CLOSE);
2667 inet_csk_listen_stop(sk);
2669 goto adjudge_to_death;
2672 /* We need to flush the recv. buffs. We do this only on the
2673 * descriptor close, not protocol-sourced closes, because the
2674 * reader process may not have drained the data yet!
2676 while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) {
2677 u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq;
2679 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
2681 data_was_unread += len;
2687 /* If socket has been already reset (e.g. in tcp_reset()) - kill it. */
2688 if (sk->sk_state == TCP_CLOSE)
2689 goto adjudge_to_death;
2691 /* As outlined in RFC 2525, section 2.17, we send a RST here because
2692 * data was lost. To witness the awful effects of the old behavior of
2693 * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk
2694 * GET in an FTP client, suspend the process, wait for the client to
2695 * advertise a zero window, then kill -9 the FTP client, wheee...
2696 * Note: timeout is always zero in such a case.
2698 if (unlikely(tcp_sk(sk)->repair)) {
2699 sk->sk_prot->disconnect(sk, 0);
2700 } else if (data_was_unread) {
2701 /* Unread data was tossed, zap the connection. */
2702 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONCLOSE);
2703 tcp_set_state(sk, TCP_CLOSE);
2704 tcp_send_active_reset(sk, sk->sk_allocation);
2705 } else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) {
2706 /* Check zero linger _after_ checking for unread data. */
2707 sk->sk_prot->disconnect(sk, 0);
2708 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
2709 } else if (tcp_close_state(sk)) {
2710 /* We FIN if the application ate all the data before
2711 * zapping the connection.
2714 /* RED-PEN. Formally speaking, we have broken TCP state
2715 * machine. State transitions:
2717 * TCP_ESTABLISHED -> TCP_FIN_WAIT1
2718 * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible)
2719 * TCP_CLOSE_WAIT -> TCP_LAST_ACK
2721 * are legal only when FIN has been sent (i.e. in window),
2722 * rather than queued out of window. Purists blame.
2724 * F.e. "RFC state" is ESTABLISHED,
2725 * if Linux state is FIN-WAIT-1, but FIN is still not sent.
2727 * The visible declinations are that sometimes
2728 * we enter time-wait state, when it is not required really
2729 * (harmless), do not send active resets, when they are
2730 * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
2731 * they look as CLOSING or LAST_ACK for Linux)
2732 * Probably, I missed some more holelets.
2734 * XXX (TFO) - To start off we don't support SYN+ACK+FIN
2735 * in a single packet! (May consider it later but will
2736 * probably need API support or TCP_CORK SYN-ACK until
2737 * data is written and socket is closed.)
2742 sk_stream_wait_close(sk, timeout);
2745 state = sk->sk_state;
2751 /* remove backlog if any, without releasing ownership. */
2754 percpu_counter_inc(sk->sk_prot->orphan_count);
2756 /* Have we already been destroyed by a softirq or backlog? */
2757 if (state != TCP_CLOSE && sk->sk_state == TCP_CLOSE)
2760 /* This is a (useful) BSD violating of the RFC. There is a
2761 * problem with TCP as specified in that the other end could
2762 * keep a socket open forever with no application left this end.
2763 * We use a 1 minute timeout (about the same as BSD) then kill
2764 * our end. If they send after that then tough - BUT: long enough
2765 * that we won't make the old 4*rto = almost no time - whoops
2768 * Nope, it was not mistake. It is really desired behaviour
2769 * f.e. on http servers, when such sockets are useless, but
2770 * consume significant resources. Let's do it with special
2771 * linger2 option. --ANK
2774 if (sk->sk_state == TCP_FIN_WAIT2) {
2775 struct tcp_sock *tp = tcp_sk(sk);
2776 if (tp->linger2 < 0) {
2777 tcp_set_state(sk, TCP_CLOSE);
2778 tcp_send_active_reset(sk, GFP_ATOMIC);
2779 __NET_INC_STATS(sock_net(sk),
2780 LINUX_MIB_TCPABORTONLINGER);
2782 const int tmo = tcp_fin_time(sk);
2784 if (tmo > TCP_TIMEWAIT_LEN) {
2785 inet_csk_reset_keepalive_timer(sk,
2786 tmo - TCP_TIMEWAIT_LEN);
2788 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
2793 if (sk->sk_state != TCP_CLOSE) {
2795 if (tcp_check_oom(sk, 0)) {
2796 tcp_set_state(sk, TCP_CLOSE);
2797 tcp_send_active_reset(sk, GFP_ATOMIC);
2798 __NET_INC_STATS(sock_net(sk),
2799 LINUX_MIB_TCPABORTONMEMORY);
2800 } else if (!check_net(sock_net(sk))) {
2801 /* Not possible to send reset; just close */
2802 tcp_set_state(sk, TCP_CLOSE);
2806 if (sk->sk_state == TCP_CLOSE) {
2807 struct request_sock *req;
2809 req = rcu_dereference_protected(tcp_sk(sk)->fastopen_rsk,
2810 lockdep_sock_is_held(sk));
2811 /* We could get here with a non-NULL req if the socket is
2812 * aborted (e.g., closed with unread data) before 3WHS
2816 reqsk_fastopen_remove(sk, req, false);
2817 inet_csk_destroy_sock(sk);
2819 /* Otherwise, socket is reprieved until protocol close. */
2826 void tcp_close(struct sock *sk, long timeout)
2829 __tcp_close(sk, timeout);
2833 EXPORT_SYMBOL(tcp_close);
2835 /* These states need RST on ABORT according to RFC793 */
2837 static inline bool tcp_need_reset(int state)
2839 return (1 << state) &
2840 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 |
2841 TCPF_FIN_WAIT2 | TCPF_SYN_RECV);
2844 static void tcp_rtx_queue_purge(struct sock *sk)
2846 struct rb_node *p = rb_first(&sk->tcp_rtx_queue);
2848 tcp_sk(sk)->highest_sack = NULL;
2850 struct sk_buff *skb = rb_to_skb(p);
2853 /* Since we are deleting whole queue, no need to
2854 * list_del(&skb->tcp_tsorted_anchor)
2856 tcp_rtx_queue_unlink(skb, sk);
2857 sk_wmem_free_skb(sk, skb);
2861 void tcp_write_queue_purge(struct sock *sk)
2863 struct sk_buff *skb;
2865 tcp_chrono_stop(sk, TCP_CHRONO_BUSY);
2866 while ((skb = __skb_dequeue(&sk->sk_write_queue)) != NULL) {
2867 tcp_skb_tsorted_anchor_cleanup(skb);
2868 sk_wmem_free_skb(sk, skb);
2870 tcp_rtx_queue_purge(sk);
2871 skb = sk->sk_tx_skb_cache;
2874 sk->sk_tx_skb_cache = NULL;
2876 INIT_LIST_HEAD(&tcp_sk(sk)->tsorted_sent_queue);
2878 tcp_clear_all_retrans_hints(tcp_sk(sk));
2879 tcp_sk(sk)->packets_out = 0;
2880 inet_csk(sk)->icsk_backoff = 0;
2883 int tcp_disconnect(struct sock *sk, int flags)
2885 struct inet_sock *inet = inet_sk(sk);
2886 struct inet_connection_sock *icsk = inet_csk(sk);
2887 struct tcp_sock *tp = tcp_sk(sk);
2888 int old_state = sk->sk_state;
2891 if (old_state != TCP_CLOSE)
2892 tcp_set_state(sk, TCP_CLOSE);
2894 /* ABORT function of RFC793 */
2895 if (old_state == TCP_LISTEN) {
2896 inet_csk_listen_stop(sk);
2897 } else if (unlikely(tp->repair)) {
2898 sk->sk_err = ECONNABORTED;
2899 } else if (tcp_need_reset(old_state) ||
2900 (tp->snd_nxt != tp->write_seq &&
2901 (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK))) {
2902 /* The last check adjusts for discrepancy of Linux wrt. RFC
2905 tcp_send_active_reset(sk, gfp_any());
2906 sk->sk_err = ECONNRESET;
2907 } else if (old_state == TCP_SYN_SENT)
2908 sk->sk_err = ECONNRESET;
2910 tcp_clear_xmit_timers(sk);
2911 __skb_queue_purge(&sk->sk_receive_queue);
2912 if (sk->sk_rx_skb_cache) {
2913 __kfree_skb(sk->sk_rx_skb_cache);
2914 sk->sk_rx_skb_cache = NULL;
2916 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt);
2918 tcp_write_queue_purge(sk);
2919 tcp_fastopen_active_disable_ofo_check(sk);
2920 skb_rbtree_purge(&tp->out_of_order_queue);
2922 inet->inet_dport = 0;
2924 if (!(sk->sk_userlocks & SOCK_BINDADDR_LOCK))
2925 inet_reset_saddr(sk);
2927 sk->sk_shutdown = 0;
2928 sock_reset_flag(sk, SOCK_DONE);
2930 tp->mdev_us = jiffies_to_usecs(TCP_TIMEOUT_INIT);
2931 tp->rcv_rtt_last_tsecr = 0;
2933 seq = tp->write_seq + tp->max_window + 2;
2936 WRITE_ONCE(tp->write_seq, seq);
2938 icsk->icsk_backoff = 0;
2939 icsk->icsk_probes_out = 0;
2940 icsk->icsk_probes_tstamp = 0;
2941 icsk->icsk_rto = TCP_TIMEOUT_INIT;
2942 icsk->icsk_rto_min = TCP_RTO_MIN;
2943 icsk->icsk_delack_max = TCP_DELACK_MAX;
2944 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
2945 tp->snd_cwnd = TCP_INIT_CWND;
2946 tp->snd_cwnd_cnt = 0;
2947 tp->window_clamp = 0;
2949 tp->delivered_ce = 0;
2950 if (icsk->icsk_ca_ops->release)
2951 icsk->icsk_ca_ops->release(sk);
2952 memset(icsk->icsk_ca_priv, 0, sizeof(icsk->icsk_ca_priv));
2953 icsk->icsk_ca_initialized = 0;
2954 tcp_set_ca_state(sk, TCP_CA_Open);
2955 tp->is_sack_reneg = 0;
2956 tcp_clear_retrans(tp);
2957 tp->total_retrans = 0;
2958 inet_csk_delack_init(sk);
2959 /* Initialize rcv_mss to TCP_MIN_MSS to avoid division by 0
2960 * issue in __tcp_select_window()
2962 icsk->icsk_ack.rcv_mss = TCP_MIN_MSS;
2963 memset(&tp->rx_opt, 0, sizeof(tp->rx_opt));
2965 dst_release(sk->sk_rx_dst);
2966 sk->sk_rx_dst = NULL;
2967 tcp_saved_syn_free(tp);
2968 tp->compressed_ack = 0;
2972 tp->bytes_acked = 0;
2973 tp->bytes_received = 0;
2974 tp->bytes_retrans = 0;
2975 tp->data_segs_in = 0;
2976 tp->data_segs_out = 0;
2977 tp->duplicate_sack[0].start_seq = 0;
2978 tp->duplicate_sack[0].end_seq = 0;
2981 tp->retrans_out = 0;
2983 tp->tlp_high_seq = 0;
2984 tp->last_oow_ack_time = 0;
2985 /* There's a bubble in the pipe until at least the first ACK. */
2986 tp->app_limited = ~0U;
2987 tp->rack.mstamp = 0;
2988 tp->rack.advanced = 0;
2989 tp->rack.reo_wnd_steps = 1;
2990 tp->rack.last_delivered = 0;
2991 tp->rack.reo_wnd_persist = 0;
2992 tp->rack.dsack_seen = 0;
2993 tp->syn_data_acked = 0;
2994 tp->rx_opt.saw_tstamp = 0;
2995 tp->rx_opt.dsack = 0;
2996 tp->rx_opt.num_sacks = 0;
2997 tp->rcv_ooopack = 0;
3000 /* Clean up fastopen related fields */
3001 tcp_free_fastopen_req(tp);
3002 inet->defer_connect = 0;
3003 tp->fastopen_client_fail = 0;
3005 WARN_ON(inet->inet_num && !icsk->icsk_bind_hash);
3007 if (sk->sk_frag.page) {
3008 put_page(sk->sk_frag.page);
3009 sk->sk_frag.page = NULL;
3010 sk->sk_frag.offset = 0;
3013 sk->sk_error_report(sk);
3016 EXPORT_SYMBOL(tcp_disconnect);
3018 static inline bool tcp_can_repair_sock(const struct sock *sk)
3020 return ns_capable(sock_net(sk)->user_ns, CAP_NET_ADMIN) &&
3021 (sk->sk_state != TCP_LISTEN);
3024 static int tcp_repair_set_window(struct tcp_sock *tp, sockptr_t optbuf, int len)
3026 struct tcp_repair_window opt;
3031 if (len != sizeof(opt))
3034 if (copy_from_sockptr(&opt, optbuf, sizeof(opt)))
3037 if (opt.max_window < opt.snd_wnd)
3040 if (after(opt.snd_wl1, tp->rcv_nxt + opt.rcv_wnd))
3043 if (after(opt.rcv_wup, tp->rcv_nxt))
3046 tp->snd_wl1 = opt.snd_wl1;
3047 tp->snd_wnd = opt.snd_wnd;
3048 tp->max_window = opt.max_window;
3050 tp->rcv_wnd = opt.rcv_wnd;
3051 tp->rcv_wup = opt.rcv_wup;
3056 static int tcp_repair_options_est(struct sock *sk, sockptr_t optbuf,
3059 struct tcp_sock *tp = tcp_sk(sk);
3060 struct tcp_repair_opt opt;
3063 while (len >= sizeof(opt)) {
3064 if (copy_from_sockptr_offset(&opt, optbuf, offset, sizeof(opt)))
3067 offset += sizeof(opt);
3070 switch (opt.opt_code) {
3072 tp->rx_opt.mss_clamp = opt.opt_val;
3077 u16 snd_wscale = opt.opt_val & 0xFFFF;
3078 u16 rcv_wscale = opt.opt_val >> 16;
3080 if (snd_wscale > TCP_MAX_WSCALE || rcv_wscale > TCP_MAX_WSCALE)
3083 tp->rx_opt.snd_wscale = snd_wscale;
3084 tp->rx_opt.rcv_wscale = rcv_wscale;
3085 tp->rx_opt.wscale_ok = 1;
3088 case TCPOPT_SACK_PERM:
3089 if (opt.opt_val != 0)
3092 tp->rx_opt.sack_ok |= TCP_SACK_SEEN;
3094 case TCPOPT_TIMESTAMP:
3095 if (opt.opt_val != 0)
3098 tp->rx_opt.tstamp_ok = 1;
3106 DEFINE_STATIC_KEY_FALSE(tcp_tx_delay_enabled);
3107 EXPORT_SYMBOL(tcp_tx_delay_enabled);
3109 static void tcp_enable_tx_delay(void)
3111 if (!static_branch_unlikely(&tcp_tx_delay_enabled)) {
3112 static int __tcp_tx_delay_enabled = 0;
3114 if (cmpxchg(&__tcp_tx_delay_enabled, 0, 1) == 0) {
3115 static_branch_enable(&tcp_tx_delay_enabled);
3116 pr_info("TCP_TX_DELAY enabled\n");
3121 /* When set indicates to always queue non-full frames. Later the user clears
3122 * this option and we transmit any pending partial frames in the queue. This is
3123 * meant to be used alongside sendfile() to get properly filled frames when the
3124 * user (for example) must write out headers with a write() call first and then
3125 * use sendfile to send out the data parts.
3127 * TCP_CORK can be set together with TCP_NODELAY and it is stronger than
3130 static void __tcp_sock_set_cork(struct sock *sk, bool on)
3132 struct tcp_sock *tp = tcp_sk(sk);
3135 tp->nonagle |= TCP_NAGLE_CORK;
3137 tp->nonagle &= ~TCP_NAGLE_CORK;
3138 if (tp->nonagle & TCP_NAGLE_OFF)
3139 tp->nonagle |= TCP_NAGLE_PUSH;
3140 tcp_push_pending_frames(sk);
3144 void tcp_sock_set_cork(struct sock *sk, bool on)
3147 __tcp_sock_set_cork(sk, on);
3150 EXPORT_SYMBOL(tcp_sock_set_cork);
3152 /* TCP_NODELAY is weaker than TCP_CORK, so that this option on corked socket is
3153 * remembered, but it is not activated until cork is cleared.
3155 * However, when TCP_NODELAY is set we make an explicit push, which overrides
3156 * even TCP_CORK for currently queued segments.
3158 static void __tcp_sock_set_nodelay(struct sock *sk, bool on)
3161 tcp_sk(sk)->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH;
3162 tcp_push_pending_frames(sk);
3164 tcp_sk(sk)->nonagle &= ~TCP_NAGLE_OFF;
3168 void tcp_sock_set_nodelay(struct sock *sk)
3171 __tcp_sock_set_nodelay(sk, true);
3174 EXPORT_SYMBOL(tcp_sock_set_nodelay);
3176 static void __tcp_sock_set_quickack(struct sock *sk, int val)
3179 inet_csk_enter_pingpong_mode(sk);
3183 inet_csk_exit_pingpong_mode(sk);
3184 if ((1 << sk->sk_state) & (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) &&
3185 inet_csk_ack_scheduled(sk)) {
3186 inet_csk(sk)->icsk_ack.pending |= ICSK_ACK_PUSHED;
3187 tcp_cleanup_rbuf(sk, 1);
3189 inet_csk_enter_pingpong_mode(sk);
3193 void tcp_sock_set_quickack(struct sock *sk, int val)
3196 __tcp_sock_set_quickack(sk, val);
3199 EXPORT_SYMBOL(tcp_sock_set_quickack);
3201 int tcp_sock_set_syncnt(struct sock *sk, int val)
3203 if (val < 1 || val > MAX_TCP_SYNCNT)
3207 inet_csk(sk)->icsk_syn_retries = val;
3211 EXPORT_SYMBOL(tcp_sock_set_syncnt);
3213 void tcp_sock_set_user_timeout(struct sock *sk, u32 val)
3216 inet_csk(sk)->icsk_user_timeout = val;
3219 EXPORT_SYMBOL(tcp_sock_set_user_timeout);
3221 int tcp_sock_set_keepidle_locked(struct sock *sk, int val)
3223 struct tcp_sock *tp = tcp_sk(sk);
3225 if (val < 1 || val > MAX_TCP_KEEPIDLE)
3228 tp->keepalive_time = val * HZ;
3229 if (sock_flag(sk, SOCK_KEEPOPEN) &&
3230 !((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN))) {
3231 u32 elapsed = keepalive_time_elapsed(tp);
3233 if (tp->keepalive_time > elapsed)
3234 elapsed = tp->keepalive_time - elapsed;
3237 inet_csk_reset_keepalive_timer(sk, elapsed);
3243 int tcp_sock_set_keepidle(struct sock *sk, int val)
3248 err = tcp_sock_set_keepidle_locked(sk, val);
3252 EXPORT_SYMBOL(tcp_sock_set_keepidle);
3254 int tcp_sock_set_keepintvl(struct sock *sk, int val)
3256 if (val < 1 || val > MAX_TCP_KEEPINTVL)
3260 tcp_sk(sk)->keepalive_intvl = val * HZ;
3264 EXPORT_SYMBOL(tcp_sock_set_keepintvl);
3266 int tcp_sock_set_keepcnt(struct sock *sk, int val)
3268 if (val < 1 || val > MAX_TCP_KEEPCNT)
3272 tcp_sk(sk)->keepalive_probes = val;
3276 EXPORT_SYMBOL(tcp_sock_set_keepcnt);
3278 int tcp_set_window_clamp(struct sock *sk, int val)
3280 struct tcp_sock *tp = tcp_sk(sk);
3283 if (sk->sk_state != TCP_CLOSE)
3285 tp->window_clamp = 0;
3287 tp->window_clamp = val < SOCK_MIN_RCVBUF / 2 ?
3288 SOCK_MIN_RCVBUF / 2 : val;
3294 * Socket option code for TCP.
3296 static int do_tcp_setsockopt(struct sock *sk, int level, int optname,
3297 sockptr_t optval, unsigned int optlen)
3299 struct tcp_sock *tp = tcp_sk(sk);
3300 struct inet_connection_sock *icsk = inet_csk(sk);
3301 struct net *net = sock_net(sk);
3305 /* These are data/string values, all the others are ints */
3307 case TCP_CONGESTION: {
3308 char name[TCP_CA_NAME_MAX];
3313 val = strncpy_from_sockptr(name, optval,
3314 min_t(long, TCP_CA_NAME_MAX-1, optlen));
3320 err = tcp_set_congestion_control(sk, name, true,
3321 ns_capable(sock_net(sk)->user_ns,
3327 char name[TCP_ULP_NAME_MAX];
3332 val = strncpy_from_sockptr(name, optval,
3333 min_t(long, TCP_ULP_NAME_MAX - 1,
3340 err = tcp_set_ulp(sk, name);
3344 case TCP_FASTOPEN_KEY: {
3345 __u8 key[TCP_FASTOPEN_KEY_BUF_LENGTH];
3346 __u8 *backup_key = NULL;
3348 /* Allow a backup key as well to facilitate key rotation
3349 * First key is the active one.
3351 if (optlen != TCP_FASTOPEN_KEY_LENGTH &&
3352 optlen != TCP_FASTOPEN_KEY_BUF_LENGTH)
3355 if (copy_from_sockptr(key, optval, optlen))
3358 if (optlen == TCP_FASTOPEN_KEY_BUF_LENGTH)
3359 backup_key = key + TCP_FASTOPEN_KEY_LENGTH;
3361 return tcp_fastopen_reset_cipher(net, sk, key, backup_key);
3368 if (optlen < sizeof(int))
3371 if (copy_from_sockptr(&val, optval, sizeof(val)))
3378 /* Values greater than interface MTU won't take effect. However
3379 * at the point when this call is done we typically don't yet
3380 * know which interface is going to be used
3382 if (val && (val < TCP_MIN_MSS || val > MAX_TCP_WINDOW)) {
3386 tp->rx_opt.user_mss = val;
3390 __tcp_sock_set_nodelay(sk, val);
3393 case TCP_THIN_LINEAR_TIMEOUTS:
3394 if (val < 0 || val > 1)
3400 case TCP_THIN_DUPACK:
3401 if (val < 0 || val > 1)
3406 if (!tcp_can_repair_sock(sk))
3408 else if (val == TCP_REPAIR_ON) {
3410 sk->sk_reuse = SK_FORCE_REUSE;
3411 tp->repair_queue = TCP_NO_QUEUE;
3412 } else if (val == TCP_REPAIR_OFF) {
3414 sk->sk_reuse = SK_NO_REUSE;
3415 tcp_send_window_probe(sk);
3416 } else if (val == TCP_REPAIR_OFF_NO_WP) {
3418 sk->sk_reuse = SK_NO_REUSE;
3424 case TCP_REPAIR_QUEUE:
3427 else if ((unsigned int)val < TCP_QUEUES_NR)
3428 tp->repair_queue = val;
3434 if (sk->sk_state != TCP_CLOSE)
3436 else if (tp->repair_queue == TCP_SEND_QUEUE)
3437 WRITE_ONCE(tp->write_seq, val);
3438 else if (tp->repair_queue == TCP_RECV_QUEUE) {
3439 WRITE_ONCE(tp->rcv_nxt, val);
3440 WRITE_ONCE(tp->copied_seq, val);
3446 case TCP_REPAIR_OPTIONS:
3449 else if (sk->sk_state == TCP_ESTABLISHED)
3450 err = tcp_repair_options_est(sk, optval, optlen);
3456 __tcp_sock_set_cork(sk, val);
3460 err = tcp_sock_set_keepidle_locked(sk, val);
3463 if (val < 1 || val > MAX_TCP_KEEPINTVL)
3466 tp->keepalive_intvl = val * HZ;
3469 if (val < 1 || val > MAX_TCP_KEEPCNT)
3472 tp->keepalive_probes = val;
3475 if (val < 1 || val > MAX_TCP_SYNCNT)
3478 icsk->icsk_syn_retries = val;
3482 /* 0: disable, 1: enable, 2: start from ether_header */
3483 if (val < 0 || val > 2)
3492 else if (val > TCP_FIN_TIMEOUT_MAX / HZ)
3493 tp->linger2 = TCP_FIN_TIMEOUT_MAX;
3495 tp->linger2 = val * HZ;
3498 case TCP_DEFER_ACCEPT:
3499 /* Translate value in seconds to number of retransmits */
3500 icsk->icsk_accept_queue.rskq_defer_accept =
3501 secs_to_retrans(val, TCP_TIMEOUT_INIT / HZ,
3505 case TCP_WINDOW_CLAMP:
3506 err = tcp_set_window_clamp(sk, val);
3510 __tcp_sock_set_quickack(sk, val);
3513 #ifdef CONFIG_TCP_MD5SIG
3515 case TCP_MD5SIG_EXT:
3516 err = tp->af_specific->md5_parse(sk, optname, optval, optlen);
3519 case TCP_USER_TIMEOUT:
3520 /* Cap the max time in ms TCP will retry or probe the window
3521 * before giving up and aborting (ETIMEDOUT) a connection.
3526 icsk->icsk_user_timeout = val;
3530 if (val >= 0 && ((1 << sk->sk_state) & (TCPF_CLOSE |
3532 tcp_fastopen_init_key_once(net);
3534 fastopen_queue_tune(sk, val);
3539 case TCP_FASTOPEN_CONNECT:
3540 if (val > 1 || val < 0) {
3542 } else if (net->ipv4.sysctl_tcp_fastopen & TFO_CLIENT_ENABLE) {
3543 if (sk->sk_state == TCP_CLOSE)
3544 tp->fastopen_connect = val;
3551 case TCP_FASTOPEN_NO_COOKIE:
3552 if (val > 1 || val < 0)
3554 else if (!((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
3557 tp->fastopen_no_cookie = val;
3563 tp->tsoffset = val - tcp_time_stamp_raw();
3565 case TCP_REPAIR_WINDOW:
3566 err = tcp_repair_set_window(tp, optval, optlen);
3568 case TCP_NOTSENT_LOWAT:
3569 tp->notsent_lowat = val;
3570 sk->sk_write_space(sk);
3573 if (val > 1 || val < 0)
3576 tp->recvmsg_inq = val;
3580 tcp_enable_tx_delay();
3581 tp->tcp_tx_delay = val;
3592 int tcp_setsockopt(struct sock *sk, int level, int optname, sockptr_t optval,
3593 unsigned int optlen)
3595 const struct inet_connection_sock *icsk = inet_csk(sk);
3597 if (level != SOL_TCP)
3598 return icsk->icsk_af_ops->setsockopt(sk, level, optname,
3600 return do_tcp_setsockopt(sk, level, optname, optval, optlen);
3602 EXPORT_SYMBOL(tcp_setsockopt);
3604 static void tcp_get_info_chrono_stats(const struct tcp_sock *tp,
3605 struct tcp_info *info)
3607 u64 stats[__TCP_CHRONO_MAX], total = 0;
3610 for (i = TCP_CHRONO_BUSY; i < __TCP_CHRONO_MAX; ++i) {
3611 stats[i] = tp->chrono_stat[i - 1];
3612 if (i == tp->chrono_type)
3613 stats[i] += tcp_jiffies32 - tp->chrono_start;
3614 stats[i] *= USEC_PER_SEC / HZ;
3618 info->tcpi_busy_time = total;
3619 info->tcpi_rwnd_limited = stats[TCP_CHRONO_RWND_LIMITED];
3620 info->tcpi_sndbuf_limited = stats[TCP_CHRONO_SNDBUF_LIMITED];
3623 /* Return information about state of tcp endpoint in API format. */
3624 void tcp_get_info(struct sock *sk, struct tcp_info *info)
3626 const struct tcp_sock *tp = tcp_sk(sk); /* iff sk_type == SOCK_STREAM */
3627 const struct inet_connection_sock *icsk = inet_csk(sk);
3633 memset(info, 0, sizeof(*info));
3634 if (sk->sk_type != SOCK_STREAM)
3637 info->tcpi_state = inet_sk_state_load(sk);
3639 /* Report meaningful fields for all TCP states, including listeners */
3640 rate = READ_ONCE(sk->sk_pacing_rate);
3641 rate64 = (rate != ~0UL) ? rate : ~0ULL;
3642 info->tcpi_pacing_rate = rate64;
3644 rate = READ_ONCE(sk->sk_max_pacing_rate);
3645 rate64 = (rate != ~0UL) ? rate : ~0ULL;
3646 info->tcpi_max_pacing_rate = rate64;
3648 info->tcpi_reordering = tp->reordering;
3649 info->tcpi_snd_cwnd = tp->snd_cwnd;
3651 if (info->tcpi_state == TCP_LISTEN) {
3652 /* listeners aliased fields :
3653 * tcpi_unacked -> Number of children ready for accept()
3654 * tcpi_sacked -> max backlog
3656 info->tcpi_unacked = READ_ONCE(sk->sk_ack_backlog);
3657 info->tcpi_sacked = READ_ONCE(sk->sk_max_ack_backlog);
3661 slow = lock_sock_fast(sk);
3663 info->tcpi_ca_state = icsk->icsk_ca_state;
3664 info->tcpi_retransmits = icsk->icsk_retransmits;
3665 info->tcpi_probes = icsk->icsk_probes_out;
3666 info->tcpi_backoff = icsk->icsk_backoff;
3668 if (tp->rx_opt.tstamp_ok)
3669 info->tcpi_options |= TCPI_OPT_TIMESTAMPS;
3670 if (tcp_is_sack(tp))
3671 info->tcpi_options |= TCPI_OPT_SACK;
3672 if (tp->rx_opt.wscale_ok) {
3673 info->tcpi_options |= TCPI_OPT_WSCALE;
3674 info->tcpi_snd_wscale = tp->rx_opt.snd_wscale;
3675 info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale;
3678 if (tp->ecn_flags & TCP_ECN_OK)
3679 info->tcpi_options |= TCPI_OPT_ECN;
3680 if (tp->ecn_flags & TCP_ECN_SEEN)
3681 info->tcpi_options |= TCPI_OPT_ECN_SEEN;
3682 if (tp->syn_data_acked)
3683 info->tcpi_options |= TCPI_OPT_SYN_DATA;
3685 info->tcpi_rto = jiffies_to_usecs(icsk->icsk_rto);
3686 info->tcpi_ato = jiffies_to_usecs(icsk->icsk_ack.ato);
3687 info->tcpi_snd_mss = tp->mss_cache;
3688 info->tcpi_rcv_mss = icsk->icsk_ack.rcv_mss;
3690 info->tcpi_unacked = tp->packets_out;
3691 info->tcpi_sacked = tp->sacked_out;
3693 info->tcpi_lost = tp->lost_out;
3694 info->tcpi_retrans = tp->retrans_out;
3696 now = tcp_jiffies32;
3697 info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime);
3698 info->tcpi_last_data_recv = jiffies_to_msecs(now - icsk->icsk_ack.lrcvtime);
3699 info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp);
3701 info->tcpi_pmtu = icsk->icsk_pmtu_cookie;
3702 info->tcpi_rcv_ssthresh = tp->rcv_ssthresh;
3703 info->tcpi_rtt = tp->srtt_us >> 3;
3704 info->tcpi_rttvar = tp->mdev_us >> 2;
3705 info->tcpi_snd_ssthresh = tp->snd_ssthresh;
3706 info->tcpi_advmss = tp->advmss;
3708 info->tcpi_rcv_rtt = tp->rcv_rtt_est.rtt_us >> 3;
3709 info->tcpi_rcv_space = tp->rcvq_space.space;
3711 info->tcpi_total_retrans = tp->total_retrans;
3713 info->tcpi_bytes_acked = tp->bytes_acked;
3714 info->tcpi_bytes_received = tp->bytes_received;
3715 info->tcpi_notsent_bytes = max_t(int, 0, tp->write_seq - tp->snd_nxt);
3716 tcp_get_info_chrono_stats(tp, info);
3718 info->tcpi_segs_out = tp->segs_out;
3719 info->tcpi_segs_in = tp->segs_in;
3721 info->tcpi_min_rtt = tcp_min_rtt(tp);
3722 info->tcpi_data_segs_in = tp->data_segs_in;
3723 info->tcpi_data_segs_out = tp->data_segs_out;
3725 info->tcpi_delivery_rate_app_limited = tp->rate_app_limited ? 1 : 0;
3726 rate64 = tcp_compute_delivery_rate(tp);
3728 info->tcpi_delivery_rate = rate64;
3729 info->tcpi_delivered = tp->delivered;
3730 info->tcpi_delivered_ce = tp->delivered_ce;
3731 info->tcpi_bytes_sent = tp->bytes_sent;
3732 info->tcpi_bytes_retrans = tp->bytes_retrans;
3733 info->tcpi_dsack_dups = tp->dsack_dups;
3734 info->tcpi_reord_seen = tp->reord_seen;
3735 info->tcpi_rcv_ooopack = tp->rcv_ooopack;
3736 info->tcpi_snd_wnd = tp->snd_wnd;
3737 info->tcpi_fastopen_client_fail = tp->fastopen_client_fail;
3738 unlock_sock_fast(sk, slow);
3740 EXPORT_SYMBOL_GPL(tcp_get_info);
3742 static size_t tcp_opt_stats_get_size(void)
3745 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_BUSY */
3746 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_RWND_LIMITED */
3747 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_SNDBUF_LIMITED */
3748 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_DATA_SEGS_OUT */
3749 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_TOTAL_RETRANS */
3750 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_PACING_RATE */
3751 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_DELIVERY_RATE */
3752 nla_total_size(sizeof(u32)) + /* TCP_NLA_SND_CWND */
3753 nla_total_size(sizeof(u32)) + /* TCP_NLA_REORDERING */
3754 nla_total_size(sizeof(u32)) + /* TCP_NLA_MIN_RTT */
3755 nla_total_size(sizeof(u8)) + /* TCP_NLA_RECUR_RETRANS */
3756 nla_total_size(sizeof(u8)) + /* TCP_NLA_DELIVERY_RATE_APP_LMT */
3757 nla_total_size(sizeof(u32)) + /* TCP_NLA_SNDQ_SIZE */
3758 nla_total_size(sizeof(u8)) + /* TCP_NLA_CA_STATE */
3759 nla_total_size(sizeof(u32)) + /* TCP_NLA_SND_SSTHRESH */
3760 nla_total_size(sizeof(u32)) + /* TCP_NLA_DELIVERED */
3761 nla_total_size(sizeof(u32)) + /* TCP_NLA_DELIVERED_CE */
3762 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_BYTES_SENT */
3763 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_BYTES_RETRANS */
3764 nla_total_size(sizeof(u32)) + /* TCP_NLA_DSACK_DUPS */
3765 nla_total_size(sizeof(u32)) + /* TCP_NLA_REORD_SEEN */
3766 nla_total_size(sizeof(u32)) + /* TCP_NLA_SRTT */
3767 nla_total_size(sizeof(u16)) + /* TCP_NLA_TIMEOUT_REHASH */
3768 nla_total_size(sizeof(u32)) + /* TCP_NLA_BYTES_NOTSENT */
3769 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_EDT */
3773 struct sk_buff *tcp_get_timestamping_opt_stats(const struct sock *sk,
3774 const struct sk_buff *orig_skb)
3776 const struct tcp_sock *tp = tcp_sk(sk);
3777 struct sk_buff *stats;
3778 struct tcp_info info;
3782 stats = alloc_skb(tcp_opt_stats_get_size(), GFP_ATOMIC);
3786 tcp_get_info_chrono_stats(tp, &info);
3787 nla_put_u64_64bit(stats, TCP_NLA_BUSY,
3788 info.tcpi_busy_time, TCP_NLA_PAD);
3789 nla_put_u64_64bit(stats, TCP_NLA_RWND_LIMITED,
3790 info.tcpi_rwnd_limited, TCP_NLA_PAD);
3791 nla_put_u64_64bit(stats, TCP_NLA_SNDBUF_LIMITED,
3792 info.tcpi_sndbuf_limited, TCP_NLA_PAD);
3793 nla_put_u64_64bit(stats, TCP_NLA_DATA_SEGS_OUT,
3794 tp->data_segs_out, TCP_NLA_PAD);
3795 nla_put_u64_64bit(stats, TCP_NLA_TOTAL_RETRANS,
3796 tp->total_retrans, TCP_NLA_PAD);
3798 rate = READ_ONCE(sk->sk_pacing_rate);
3799 rate64 = (rate != ~0UL) ? rate : ~0ULL;
3800 nla_put_u64_64bit(stats, TCP_NLA_PACING_RATE, rate64, TCP_NLA_PAD);
3802 rate64 = tcp_compute_delivery_rate(tp);
3803 nla_put_u64_64bit(stats, TCP_NLA_DELIVERY_RATE, rate64, TCP_NLA_PAD);
3805 nla_put_u32(stats, TCP_NLA_SND_CWND, tp->snd_cwnd);
3806 nla_put_u32(stats, TCP_NLA_REORDERING, tp->reordering);
3807 nla_put_u32(stats, TCP_NLA_MIN_RTT, tcp_min_rtt(tp));
3809 nla_put_u8(stats, TCP_NLA_RECUR_RETRANS, inet_csk(sk)->icsk_retransmits);
3810 nla_put_u8(stats, TCP_NLA_DELIVERY_RATE_APP_LMT, !!tp->rate_app_limited);
3811 nla_put_u32(stats, TCP_NLA_SND_SSTHRESH, tp->snd_ssthresh);
3812 nla_put_u32(stats, TCP_NLA_DELIVERED, tp->delivered);
3813 nla_put_u32(stats, TCP_NLA_DELIVERED_CE, tp->delivered_ce);
3815 nla_put_u32(stats, TCP_NLA_SNDQ_SIZE, tp->write_seq - tp->snd_una);
3816 nla_put_u8(stats, TCP_NLA_CA_STATE, inet_csk(sk)->icsk_ca_state);
3818 nla_put_u64_64bit(stats, TCP_NLA_BYTES_SENT, tp->bytes_sent,
3820 nla_put_u64_64bit(stats, TCP_NLA_BYTES_RETRANS, tp->bytes_retrans,
3822 nla_put_u32(stats, TCP_NLA_DSACK_DUPS, tp->dsack_dups);
3823 nla_put_u32(stats, TCP_NLA_REORD_SEEN, tp->reord_seen);
3824 nla_put_u32(stats, TCP_NLA_SRTT, tp->srtt_us >> 3);
3825 nla_put_u16(stats, TCP_NLA_TIMEOUT_REHASH, tp->timeout_rehash);
3826 nla_put_u32(stats, TCP_NLA_BYTES_NOTSENT,
3827 max_t(int, 0, tp->write_seq - tp->snd_nxt));
3828 nla_put_u64_64bit(stats, TCP_NLA_EDT, orig_skb->skb_mstamp_ns,
3834 static int do_tcp_getsockopt(struct sock *sk, int level,
3835 int optname, char __user *optval, int __user *optlen)
3837 struct inet_connection_sock *icsk = inet_csk(sk);
3838 struct tcp_sock *tp = tcp_sk(sk);
3839 struct net *net = sock_net(sk);
3842 if (get_user(len, optlen))
3845 len = min_t(unsigned int, len, sizeof(int));
3852 val = tp->mss_cache;
3853 if (!val && ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
3854 val = tp->rx_opt.user_mss;
3856 val = tp->rx_opt.mss_clamp;
3859 val = !!(tp->nonagle&TCP_NAGLE_OFF);
3862 val = !!(tp->nonagle&TCP_NAGLE_CORK);
3865 val = keepalive_time_when(tp) / HZ;
3868 val = keepalive_intvl_when(tp) / HZ;
3871 val = keepalive_probes(tp);
3874 val = icsk->icsk_syn_retries ? : net->ipv4.sysctl_tcp_syn_retries;
3879 val = (val ? : net->ipv4.sysctl_tcp_fin_timeout) / HZ;
3881 case TCP_DEFER_ACCEPT:
3882 val = retrans_to_secs(icsk->icsk_accept_queue.rskq_defer_accept,
3883 TCP_TIMEOUT_INIT / HZ, TCP_RTO_MAX / HZ);
3885 case TCP_WINDOW_CLAMP:
3886 val = tp->window_clamp;
3889 struct tcp_info info;
3891 if (get_user(len, optlen))
3894 tcp_get_info(sk, &info);
3896 len = min_t(unsigned int, len, sizeof(info));
3897 if (put_user(len, optlen))
3899 if (copy_to_user(optval, &info, len))
3904 const struct tcp_congestion_ops *ca_ops;
3905 union tcp_cc_info info;
3909 if (get_user(len, optlen))
3912 ca_ops = icsk->icsk_ca_ops;
3913 if (ca_ops && ca_ops->get_info)
3914 sz = ca_ops->get_info(sk, ~0U, &attr, &info);
3916 len = min_t(unsigned int, len, sz);
3917 if (put_user(len, optlen))
3919 if (copy_to_user(optval, &info, len))
3924 val = !inet_csk_in_pingpong_mode(sk);
3927 case TCP_CONGESTION:
3928 if (get_user(len, optlen))
3930 len = min_t(unsigned int, len, TCP_CA_NAME_MAX);
3931 if (put_user(len, optlen))
3933 if (copy_to_user(optval, icsk->icsk_ca_ops->name, len))
3938 if (get_user(len, optlen))
3940 len = min_t(unsigned int, len, TCP_ULP_NAME_MAX);
3941 if (!icsk->icsk_ulp_ops) {
3942 if (put_user(0, optlen))
3946 if (put_user(len, optlen))
3948 if (copy_to_user(optval, icsk->icsk_ulp_ops->name, len))
3952 case TCP_FASTOPEN_KEY: {
3953 u64 key[TCP_FASTOPEN_KEY_BUF_LENGTH / sizeof(u64)];
3954 unsigned int key_len;
3956 if (get_user(len, optlen))
3959 key_len = tcp_fastopen_get_cipher(net, icsk, key) *
3960 TCP_FASTOPEN_KEY_LENGTH;
3961 len = min_t(unsigned int, len, key_len);
3962 if (put_user(len, optlen))
3964 if (copy_to_user(optval, key, len))
3968 case TCP_THIN_LINEAR_TIMEOUTS:
3972 case TCP_THIN_DUPACK:
3980 case TCP_REPAIR_QUEUE:
3982 val = tp->repair_queue;
3987 case TCP_REPAIR_WINDOW: {
3988 struct tcp_repair_window opt;
3990 if (get_user(len, optlen))
3993 if (len != sizeof(opt))
3999 opt.snd_wl1 = tp->snd_wl1;
4000 opt.snd_wnd = tp->snd_wnd;
4001 opt.max_window = tp->max_window;
4002 opt.rcv_wnd = tp->rcv_wnd;
4003 opt.rcv_wup = tp->rcv_wup;
4005 if (copy_to_user(optval, &opt, len))
4010 if (tp->repair_queue == TCP_SEND_QUEUE)
4011 val = tp->write_seq;
4012 else if (tp->repair_queue == TCP_RECV_QUEUE)
4018 case TCP_USER_TIMEOUT:
4019 val = icsk->icsk_user_timeout;
4023 val = icsk->icsk_accept_queue.fastopenq.max_qlen;
4026 case TCP_FASTOPEN_CONNECT:
4027 val = tp->fastopen_connect;
4030 case TCP_FASTOPEN_NO_COOKIE:
4031 val = tp->fastopen_no_cookie;
4035 val = tp->tcp_tx_delay;
4039 val = tcp_time_stamp_raw() + tp->tsoffset;
4041 case TCP_NOTSENT_LOWAT:
4042 val = tp->notsent_lowat;
4045 val = tp->recvmsg_inq;
4050 case TCP_SAVED_SYN: {
4051 if (get_user(len, optlen))
4055 if (tp->saved_syn) {
4056 if (len < tcp_saved_syn_len(tp->saved_syn)) {
4057 if (put_user(tcp_saved_syn_len(tp->saved_syn),
4065 len = tcp_saved_syn_len(tp->saved_syn);
4066 if (put_user(len, optlen)) {
4070 if (copy_to_user(optval, tp->saved_syn->data, len)) {
4074 tcp_saved_syn_free(tp);
4079 if (put_user(len, optlen))
4085 case TCP_ZEROCOPY_RECEIVE: {
4086 struct tcp_zerocopy_receive zc = {};
4089 if (get_user(len, optlen))
4091 if (len < offsetofend(struct tcp_zerocopy_receive, length))
4093 if (len > sizeof(zc)) {
4095 if (put_user(len, optlen))
4098 if (copy_from_user(&zc, optval, len))
4101 err = tcp_zerocopy_receive(sk, &zc);
4103 if (len >= offsetofend(struct tcp_zerocopy_receive, err))
4104 goto zerocopy_rcv_sk_err;
4106 case offsetofend(struct tcp_zerocopy_receive, err):
4107 goto zerocopy_rcv_sk_err;
4108 case offsetofend(struct tcp_zerocopy_receive, inq):
4109 goto zerocopy_rcv_inq;
4110 case offsetofend(struct tcp_zerocopy_receive, length):
4112 goto zerocopy_rcv_out;
4114 zerocopy_rcv_sk_err:
4116 zc.err = sock_error(sk);
4118 zc.inq = tcp_inq_hint(sk);
4120 if (!err && copy_to_user(optval, &zc, len))
4126 return -ENOPROTOOPT;
4129 if (put_user(len, optlen))
4131 if (copy_to_user(optval, &val, len))
4136 int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval,
4139 struct inet_connection_sock *icsk = inet_csk(sk);
4141 if (level != SOL_TCP)
4142 return icsk->icsk_af_ops->getsockopt(sk, level, optname,
4144 return do_tcp_getsockopt(sk, level, optname, optval, optlen);
4146 EXPORT_SYMBOL(tcp_getsockopt);
4148 #ifdef CONFIG_TCP_MD5SIG
4149 static DEFINE_PER_CPU(struct tcp_md5sig_pool, tcp_md5sig_pool);
4150 static DEFINE_MUTEX(tcp_md5sig_mutex);
4151 static bool tcp_md5sig_pool_populated = false;
4153 static void __tcp_alloc_md5sig_pool(void)
4155 struct crypto_ahash *hash;
4158 hash = crypto_alloc_ahash("md5", 0, CRYPTO_ALG_ASYNC);
4162 for_each_possible_cpu(cpu) {
4163 void *scratch = per_cpu(tcp_md5sig_pool, cpu).scratch;
4164 struct ahash_request *req;
4167 scratch = kmalloc_node(sizeof(union tcp_md5sum_block) +
4168 sizeof(struct tcphdr),
4173 per_cpu(tcp_md5sig_pool, cpu).scratch = scratch;
4175 if (per_cpu(tcp_md5sig_pool, cpu).md5_req)
4178 req = ahash_request_alloc(hash, GFP_KERNEL);
4182 ahash_request_set_callback(req, 0, NULL, NULL);
4184 per_cpu(tcp_md5sig_pool, cpu).md5_req = req;
4186 /* before setting tcp_md5sig_pool_populated, we must commit all writes
4187 * to memory. See smp_rmb() in tcp_get_md5sig_pool()
4190 tcp_md5sig_pool_populated = true;
4193 bool tcp_alloc_md5sig_pool(void)
4195 if (unlikely(!tcp_md5sig_pool_populated)) {
4196 mutex_lock(&tcp_md5sig_mutex);
4198 if (!tcp_md5sig_pool_populated) {
4199 __tcp_alloc_md5sig_pool();
4200 if (tcp_md5sig_pool_populated)
4201 static_branch_inc(&tcp_md5_needed);
4204 mutex_unlock(&tcp_md5sig_mutex);
4206 return tcp_md5sig_pool_populated;
4208 EXPORT_SYMBOL(tcp_alloc_md5sig_pool);
4212 * tcp_get_md5sig_pool - get md5sig_pool for this user
4214 * We use percpu structure, so if we succeed, we exit with preemption
4215 * and BH disabled, to make sure another thread or softirq handling
4216 * wont try to get same context.
4218 struct tcp_md5sig_pool *tcp_get_md5sig_pool(void)
4222 if (tcp_md5sig_pool_populated) {
4223 /* coupled with smp_wmb() in __tcp_alloc_md5sig_pool() */
4225 return this_cpu_ptr(&tcp_md5sig_pool);
4230 EXPORT_SYMBOL(tcp_get_md5sig_pool);
4232 int tcp_md5_hash_skb_data(struct tcp_md5sig_pool *hp,
4233 const struct sk_buff *skb, unsigned int header_len)
4235 struct scatterlist sg;
4236 const struct tcphdr *tp = tcp_hdr(skb);
4237 struct ahash_request *req = hp->md5_req;
4239 const unsigned int head_data_len = skb_headlen(skb) > header_len ?
4240 skb_headlen(skb) - header_len : 0;
4241 const struct skb_shared_info *shi = skb_shinfo(skb);
4242 struct sk_buff *frag_iter;
4244 sg_init_table(&sg, 1);
4246 sg_set_buf(&sg, ((u8 *) tp) + header_len, head_data_len);
4247 ahash_request_set_crypt(req, &sg, NULL, head_data_len);
4248 if (crypto_ahash_update(req))
4251 for (i = 0; i < shi->nr_frags; ++i) {
4252 const skb_frag_t *f = &shi->frags[i];
4253 unsigned int offset = skb_frag_off(f);
4254 struct page *page = skb_frag_page(f) + (offset >> PAGE_SHIFT);
4256 sg_set_page(&sg, page, skb_frag_size(f),
4257 offset_in_page(offset));
4258 ahash_request_set_crypt(req, &sg, NULL, skb_frag_size(f));
4259 if (crypto_ahash_update(req))
4263 skb_walk_frags(skb, frag_iter)
4264 if (tcp_md5_hash_skb_data(hp, frag_iter, 0))
4269 EXPORT_SYMBOL(tcp_md5_hash_skb_data);
4271 int tcp_md5_hash_key(struct tcp_md5sig_pool *hp, const struct tcp_md5sig_key *key)
4273 u8 keylen = READ_ONCE(key->keylen); /* paired with WRITE_ONCE() in tcp_md5_do_add */
4274 struct scatterlist sg;
4276 sg_init_one(&sg, key->key, keylen);
4277 ahash_request_set_crypt(hp->md5_req, &sg, NULL, keylen);
4279 /* We use data_race() because tcp_md5_do_add() might change key->key under us */
4280 return data_race(crypto_ahash_update(hp->md5_req));
4282 EXPORT_SYMBOL(tcp_md5_hash_key);
4286 void tcp_done(struct sock *sk)
4288 struct request_sock *req;
4290 /* We might be called with a new socket, after
4291 * inet_csk_prepare_forced_close() has been called
4292 * so we can not use lockdep_sock_is_held(sk)
4294 req = rcu_dereference_protected(tcp_sk(sk)->fastopen_rsk, 1);
4296 if (sk->sk_state == TCP_SYN_SENT || sk->sk_state == TCP_SYN_RECV)
4297 TCP_INC_STATS(sock_net(sk), TCP_MIB_ATTEMPTFAILS);
4299 tcp_set_state(sk, TCP_CLOSE);
4300 tcp_clear_xmit_timers(sk);
4302 reqsk_fastopen_remove(sk, req, false);
4304 sk->sk_shutdown = SHUTDOWN_MASK;
4306 if (!sock_flag(sk, SOCK_DEAD))
4307 sk->sk_state_change(sk);
4309 inet_csk_destroy_sock(sk);
4311 EXPORT_SYMBOL_GPL(tcp_done);
4313 int tcp_abort(struct sock *sk, int err)
4315 if (!sk_fullsock(sk)) {
4316 if (sk->sk_state == TCP_NEW_SYN_RECV) {
4317 struct request_sock *req = inet_reqsk(sk);
4320 inet_csk_reqsk_queue_drop(req->rsk_listener, req);
4327 /* Don't race with userspace socket closes such as tcp_close. */
4330 if (sk->sk_state == TCP_LISTEN) {
4331 tcp_set_state(sk, TCP_CLOSE);
4332 inet_csk_listen_stop(sk);
4335 /* Don't race with BH socket closes such as inet_csk_listen_stop. */
4339 if (!sock_flag(sk, SOCK_DEAD)) {
4341 /* This barrier is coupled with smp_rmb() in tcp_poll() */
4343 sk->sk_error_report(sk);
4344 if (tcp_need_reset(sk->sk_state))
4345 tcp_send_active_reset(sk, GFP_ATOMIC);
4351 tcp_write_queue_purge(sk);
4355 EXPORT_SYMBOL_GPL(tcp_abort);
4357 extern struct tcp_congestion_ops tcp_reno;
4359 static __initdata unsigned long thash_entries;
4360 static int __init set_thash_entries(char *str)
4367 ret = kstrtoul(str, 0, &thash_entries);
4373 __setup("thash_entries=", set_thash_entries);
4375 static void __init tcp_init_mem(void)
4377 unsigned long limit = nr_free_buffer_pages() / 16;
4379 limit = max(limit, 128UL);
4380 sysctl_tcp_mem[0] = limit / 4 * 3; /* 4.68 % */
4381 sysctl_tcp_mem[1] = limit; /* 6.25 % */
4382 sysctl_tcp_mem[2] = sysctl_tcp_mem[0] * 2; /* 9.37 % */
4385 void __init tcp_init(void)
4387 int max_rshare, max_wshare, cnt;
4388 unsigned long limit;
4391 BUILD_BUG_ON(TCP_MIN_SND_MSS <= MAX_TCP_OPTION_SPACE);
4392 BUILD_BUG_ON(sizeof(struct tcp_skb_cb) >
4393 sizeof_field(struct sk_buff, cb));
4395 percpu_counter_init(&tcp_sockets_allocated, 0, GFP_KERNEL);
4396 percpu_counter_init(&tcp_orphan_count, 0, GFP_KERNEL);
4397 inet_hashinfo_init(&tcp_hashinfo);
4398 inet_hashinfo2_init(&tcp_hashinfo, "tcp_listen_portaddr_hash",
4399 thash_entries, 21, /* one slot per 2 MB*/
4401 tcp_hashinfo.bind_bucket_cachep =
4402 kmem_cache_create("tcp_bind_bucket",
4403 sizeof(struct inet_bind_bucket), 0,
4404 SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL);
4406 /* Size and allocate the main established and bind bucket
4409 * The methodology is similar to that of the buffer cache.
4411 tcp_hashinfo.ehash =
4412 alloc_large_system_hash("TCP established",
4413 sizeof(struct inet_ehash_bucket),
4415 17, /* one slot per 128 KB of memory */
4418 &tcp_hashinfo.ehash_mask,
4420 thash_entries ? 0 : 512 * 1024);
4421 for (i = 0; i <= tcp_hashinfo.ehash_mask; i++)
4422 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].chain, i);
4424 if (inet_ehash_locks_alloc(&tcp_hashinfo))
4425 panic("TCP: failed to alloc ehash_locks");
4426 tcp_hashinfo.bhash =
4427 alloc_large_system_hash("TCP bind",
4428 sizeof(struct inet_bind_hashbucket),
4429 tcp_hashinfo.ehash_mask + 1,
4430 17, /* one slot per 128 KB of memory */
4432 &tcp_hashinfo.bhash_size,
4436 tcp_hashinfo.bhash_size = 1U << tcp_hashinfo.bhash_size;
4437 for (i = 0; i < tcp_hashinfo.bhash_size; i++) {
4438 spin_lock_init(&tcp_hashinfo.bhash[i].lock);
4439 INIT_HLIST_HEAD(&tcp_hashinfo.bhash[i].chain);
4443 cnt = tcp_hashinfo.ehash_mask + 1;
4444 sysctl_tcp_max_orphans = cnt / 2;
4447 /* Set per-socket limits to no more than 1/128 the pressure threshold */
4448 limit = nr_free_buffer_pages() << (PAGE_SHIFT - 7);
4449 max_wshare = min(4UL*1024*1024, limit);
4450 max_rshare = min(6UL*1024*1024, limit);
4452 init_net.ipv4.sysctl_tcp_wmem[0] = SK_MEM_QUANTUM;
4453 init_net.ipv4.sysctl_tcp_wmem[1] = 16*1024;
4454 init_net.ipv4.sysctl_tcp_wmem[2] = max(64*1024, max_wshare);
4456 init_net.ipv4.sysctl_tcp_rmem[0] = SK_MEM_QUANTUM;
4457 init_net.ipv4.sysctl_tcp_rmem[1] = 131072;
4458 init_net.ipv4.sysctl_tcp_rmem[2] = max(131072, max_rshare);
4460 pr_info("Hash tables configured (established %u bind %u)\n",
4461 tcp_hashinfo.ehash_mask + 1, tcp_hashinfo.bhash_size);
4465 BUG_ON(tcp_register_congestion_control(&tcp_reno) != 0);