2 * INET An implementation of the TCP/IP protocol suite for the LINUX
3 * operating system. INET is implemented using the BSD Socket
4 * interface as the means of communication with the user level.
6 * Implementation of the Transmission Control Protocol(TCP).
9 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
10 * Mark Evans, <evansmp@uhura.aston.ac.uk>
11 * Corey Minyard <wf-rch!minyard@relay.EU.net>
12 * Florian La Roche, <flla@stud.uni-sb.de>
13 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
14 * Linus Torvalds, <torvalds@cs.helsinki.fi>
15 * Alan Cox, <gw4pts@gw4pts.ampr.org>
16 * Matthew Dillon, <dillon@apollo.west.oic.com>
17 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
18 * Jorge Cwik, <jorge@laser.satlink.net>
21 * Alan Cox : Numerous verify_area() calls
22 * Alan Cox : Set the ACK bit on a reset
23 * Alan Cox : Stopped it crashing if it closed while
24 * sk->inuse=1 and was trying to connect
26 * Alan Cox : All icmp error handling was broken
27 * pointers passed where wrong and the
28 * socket was looked up backwards. Nobody
29 * tested any icmp error code obviously.
30 * Alan Cox : tcp_err() now handled properly. It
31 * wakes people on errors. poll
32 * behaves and the icmp error race
33 * has gone by moving it into sock.c
34 * Alan Cox : tcp_send_reset() fixed to work for
35 * everything not just packets for
37 * Alan Cox : tcp option processing.
38 * Alan Cox : Reset tweaked (still not 100%) [Had
40 * Herp Rosmanith : More reset fixes
41 * Alan Cox : No longer acks invalid rst frames.
42 * Acking any kind of RST is right out.
43 * Alan Cox : Sets an ignore me flag on an rst
44 * receive otherwise odd bits of prattle
46 * Alan Cox : Fixed another acking RST frame bug.
47 * Should stop LAN workplace lockups.
48 * Alan Cox : Some tidyups using the new skb list
50 * Alan Cox : sk->keepopen now seems to work
51 * Alan Cox : Pulls options out correctly on accepts
52 * Alan Cox : Fixed assorted sk->rqueue->next errors
53 * Alan Cox : PSH doesn't end a TCP read. Switched a
55 * Alan Cox : Tidied tcp_data to avoid a potential
57 * Alan Cox : Added some better commenting, as the
58 * tcp is hard to follow
59 * Alan Cox : Removed incorrect check for 20 * psh
60 * Michael O'Reilly : ack < copied bug fix.
61 * Johannes Stille : Misc tcp fixes (not all in yet).
62 * Alan Cox : FIN with no memory -> CRASH
63 * Alan Cox : Added socket option proto entries.
64 * Also added awareness of them to accept.
65 * Alan Cox : Added TCP options (SOL_TCP)
66 * Alan Cox : Switched wakeup calls to callbacks,
67 * so the kernel can layer network
69 * Alan Cox : Use ip_tos/ip_ttl settings.
70 * Alan Cox : Handle FIN (more) properly (we hope).
71 * Alan Cox : RST frames sent on unsynchronised
73 * Alan Cox : Put in missing check for SYN bit.
74 * Alan Cox : Added tcp_select_window() aka NET2E
75 * window non shrink trick.
76 * Alan Cox : Added a couple of small NET2E timer
78 * Charles Hedrick : TCP fixes
79 * Toomas Tamm : TCP window fixes
80 * Alan Cox : Small URG fix to rlogin ^C ack fight
81 * Charles Hedrick : Rewrote most of it to actually work
82 * Linus : Rewrote tcp_read() and URG handling
84 * Gerhard Koerting: Fixed some missing timer handling
85 * Matthew Dillon : Reworked TCP machine states as per RFC
86 * Gerhard Koerting: PC/TCP workarounds
87 * Adam Caldwell : Assorted timer/timing errors
88 * Matthew Dillon : Fixed another RST bug
89 * Alan Cox : Move to kernel side addressing changes.
90 * Alan Cox : Beginning work on TCP fastpathing
92 * Arnt Gulbrandsen: Turbocharged tcp_check() routine.
93 * Alan Cox : TCP fast path debugging
94 * Alan Cox : Window clamping
95 * Michael Riepe : Bug in tcp_check()
96 * Matt Dillon : More TCP improvements and RST bug fixes
97 * Matt Dillon : Yet more small nasties remove from the
98 * TCP code (Be very nice to this man if
99 * tcp finally works 100%) 8)
100 * Alan Cox : BSD accept semantics.
101 * Alan Cox : Reset on closedown bug.
102 * Peter De Schrijver : ENOTCONN check missing in tcp_sendto().
103 * Michael Pall : Handle poll() after URG properly in
105 * Michael Pall : Undo the last fix in tcp_read_urg()
106 * (multi URG PUSH broke rlogin).
107 * Michael Pall : Fix the multi URG PUSH problem in
108 * tcp_readable(), poll() after URG
110 * Michael Pall : recv(...,MSG_OOB) never blocks in the
112 * Alan Cox : Changed the semantics of sk->socket to
113 * fix a race and a signal problem with
114 * accept() and async I/O.
115 * Alan Cox : Relaxed the rules on tcp_sendto().
116 * Yury Shevchuk : Really fixed accept() blocking problem.
117 * Craig I. Hagan : Allow for BSD compatible TIME_WAIT for
118 * clients/servers which listen in on
120 * Alan Cox : Cleaned the above up and shrank it to
121 * a sensible code size.
122 * Alan Cox : Self connect lockup fix.
123 * Alan Cox : No connect to multicast.
124 * Ross Biro : Close unaccepted children on master
126 * Alan Cox : Reset tracing code.
127 * Alan Cox : Spurious resets on shutdown.
128 * Alan Cox : Giant 15 minute/60 second timer error
129 * Alan Cox : Small whoops in polling before an
131 * Alan Cox : Kept the state trace facility since
132 * it's handy for debugging.
133 * Alan Cox : More reset handler fixes.
134 * Alan Cox : Started rewriting the code based on
135 * the RFC's for other useful protocol
136 * references see: Comer, KA9Q NOS, and
137 * for a reference on the difference
138 * between specifications and how BSD
139 * works see the 4.4lite source.
140 * A.N.Kuznetsov : Don't time wait on completion of tidy
142 * Linus Torvalds : Fin/Shutdown & copied_seq changes.
143 * Linus Torvalds : Fixed BSD port reuse to work first syn
144 * Alan Cox : Reimplemented timers as per the RFC
145 * and using multiple timers for sanity.
146 * Alan Cox : Small bug fixes, and a lot of new
148 * Alan Cox : Fixed dual reader crash by locking
149 * the buffers (much like datagram.c)
150 * Alan Cox : Fixed stuck sockets in probe. A probe
151 * now gets fed up of retrying without
152 * (even a no space) answer.
153 * Alan Cox : Extracted closing code better
154 * Alan Cox : Fixed the closing state machine to
156 * Alan Cox : More 'per spec' fixes.
157 * Jorge Cwik : Even faster checksumming.
158 * Alan Cox : tcp_data() doesn't ack illegal PSH
159 * only frames. At least one pc tcp stack
161 * Alan Cox : Cache last socket.
162 * Alan Cox : Per route irtt.
163 * Matt Day : poll()->select() match BSD precisely on error
164 * Alan Cox : New buffers
165 * Marc Tamsky : Various sk->prot->retransmits and
166 * sk->retransmits misupdating fixed.
167 * Fixed tcp_write_timeout: stuck close,
168 * and TCP syn retries gets used now.
169 * Mark Yarvis : In tcp_read_wakeup(), don't send an
170 * ack if state is TCP_CLOSED.
171 * Alan Cox : Look up device on a retransmit - routes may
172 * change. Doesn't yet cope with MSS shrink right
174 * Marc Tamsky : Closing in closing fixes.
175 * Mike Shaver : RFC1122 verifications.
176 * Alan Cox : rcv_saddr errors.
177 * Alan Cox : Block double connect().
178 * Alan Cox : Small hooks for enSKIP.
179 * Alexey Kuznetsov: Path MTU discovery.
180 * Alan Cox : Support soft errors.
181 * Alan Cox : Fix MTU discovery pathological case
182 * when the remote claims no mtu!
183 * Marc Tamsky : TCP_CLOSE fix.
184 * Colin (G3TNE) : Send a reset on syn ack replies in
185 * window but wrong (fixes NT lpd problems)
186 * Pedro Roque : Better TCP window handling, delayed ack.
187 * Joerg Reuter : No modification of locked buffers in
188 * tcp_do_retransmit()
189 * Eric Schenk : Changed receiver side silly window
190 * avoidance algorithm to BSD style
191 * algorithm. This doubles throughput
192 * against machines running Solaris,
193 * and seems to result in general
195 * Stefan Magdalinski : adjusted tcp_readable() to fix FIONREAD
196 * Willy Konynenberg : Transparent proxying support.
197 * Mike McLagan : Routing by source
198 * Keith Owens : Do proper merging with partial SKB's in
199 * tcp_do_sendmsg to avoid burstiness.
200 * Eric Schenk : Fix fast close down bug with
201 * shutdown() followed by close().
202 * Andi Kleen : Make poll agree with SIGIO
203 * Salvatore Sanfilippo : Support SO_LINGER with linger == 1 and
204 * lingertime == 0 (RFC 793 ABORT Call)
205 * Hirokazu Takahashi : Use copy_from_user() instead of
206 * csum_and_copy_from_user() if possible.
208 * This program is free software; you can redistribute it and/or
209 * modify it under the terms of the GNU General Public License
210 * as published by the Free Software Foundation; either version
211 * 2 of the License, or(at your option) any later version.
213 * Description of States:
215 * TCP_SYN_SENT sent a connection request, waiting for ack
217 * TCP_SYN_RECV received a connection request, sent ack,
218 * waiting for final ack in three-way handshake.
220 * TCP_ESTABLISHED connection established
222 * TCP_FIN_WAIT1 our side has shutdown, waiting to complete
223 * transmission of remaining buffered data
225 * TCP_FIN_WAIT2 all buffered data sent, waiting for remote
228 * TCP_CLOSING both sides have shutdown but we still have
229 * data we have to finish sending
231 * TCP_TIME_WAIT timeout to catch resent junk before entering
232 * closed, can only be entered from FIN_WAIT2
233 * or CLOSING. Required because the other end
234 * may not have gotten our last ACK causing it
235 * to retransmit the data packet (which we ignore)
237 * TCP_CLOSE_WAIT remote side has shutdown and is waiting for
238 * us to finish writing our data and to shutdown
239 * (we have to close() to move on to LAST_ACK)
241 * TCP_LAST_ACK out side has shutdown after remote has
242 * shutdown. There may still be data in our
243 * buffer that we have to finish sending
245 * TCP_CLOSE socket is finished
248 #define pr_fmt(fmt) "TCP: " fmt
250 #include <crypto/hash.h>
251 #include <linux/kernel.h>
252 #include <linux/module.h>
253 #include <linux/types.h>
254 #include <linux/fcntl.h>
255 #include <linux/poll.h>
256 #include <linux/inet_diag.h>
257 #include <linux/init.h>
258 #include <linux/fs.h>
259 #include <linux/skbuff.h>
260 #include <linux/scatterlist.h>
261 #include <linux/splice.h>
262 #include <linux/net.h>
263 #include <linux/socket.h>
264 #include <linux/random.h>
265 #include <linux/bootmem.h>
266 #include <linux/highmem.h>
267 #include <linux/swap.h>
268 #include <linux/cache.h>
269 #include <linux/err.h>
270 #include <linux/time.h>
271 #include <linux/slab.h>
272 #include <linux/errqueue.h>
273 #include <linux/static_key.h>
275 #include <net/icmp.h>
276 #include <net/inet_common.h>
278 #include <net/xfrm.h>
280 #include <net/sock.h>
282 #include <linux/uaccess.h>
283 #include <asm/ioctls.h>
284 #include <net/busy_poll.h>
286 struct percpu_counter tcp_orphan_count;
287 EXPORT_SYMBOL_GPL(tcp_orphan_count);
289 long sysctl_tcp_mem[3] __read_mostly;
290 EXPORT_SYMBOL(sysctl_tcp_mem);
292 atomic_long_t tcp_memory_allocated; /* Current allocated memory. */
293 EXPORT_SYMBOL(tcp_memory_allocated);
295 #if IS_ENABLED(CONFIG_SMC)
296 DEFINE_STATIC_KEY_FALSE(tcp_have_smc);
297 EXPORT_SYMBOL(tcp_have_smc);
301 * Current number of TCP sockets.
303 struct percpu_counter tcp_sockets_allocated;
304 EXPORT_SYMBOL(tcp_sockets_allocated);
309 struct tcp_splice_state {
310 struct pipe_inode_info *pipe;
316 * Pressure flag: try to collapse.
317 * Technical note: it is used by multiple contexts non atomically.
318 * All the __sk_mem_schedule() is of this nature: accounting
319 * is strict, actions are advisory and have some latency.
321 unsigned long tcp_memory_pressure __read_mostly;
322 EXPORT_SYMBOL_GPL(tcp_memory_pressure);
324 void tcp_enter_memory_pressure(struct sock *sk)
328 if (tcp_memory_pressure)
334 if (!cmpxchg(&tcp_memory_pressure, 0, val))
335 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMEMORYPRESSURES);
337 EXPORT_SYMBOL_GPL(tcp_enter_memory_pressure);
339 void tcp_leave_memory_pressure(struct sock *sk)
343 if (!tcp_memory_pressure)
345 val = xchg(&tcp_memory_pressure, 0);
347 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPMEMORYPRESSURESCHRONO,
348 jiffies_to_msecs(jiffies - val));
350 EXPORT_SYMBOL_GPL(tcp_leave_memory_pressure);
352 /* Convert seconds to retransmits based on initial and max timeout */
353 static u8 secs_to_retrans(int seconds, int timeout, int rto_max)
358 int period = timeout;
361 while (seconds > period && res < 255) {
364 if (timeout > rto_max)
372 /* Convert retransmits to seconds based on initial and max timeout */
373 static int retrans_to_secs(u8 retrans, int timeout, int rto_max)
381 if (timeout > rto_max)
389 static u64 tcp_compute_delivery_rate(const struct tcp_sock *tp)
391 u32 rate = READ_ONCE(tp->rate_delivered);
392 u32 intv = READ_ONCE(tp->rate_interval_us);
396 rate64 = (u64)rate * tp->mss_cache * USEC_PER_SEC;
397 do_div(rate64, intv);
402 /* Address-family independent initialization for a tcp_sock.
404 * NOTE: A lot of things set to zero explicitly by call to
405 * sk_alloc() so need not be done here.
407 void tcp_init_sock(struct sock *sk)
409 struct inet_connection_sock *icsk = inet_csk(sk);
410 struct tcp_sock *tp = tcp_sk(sk);
412 tp->out_of_order_queue = RB_ROOT;
413 sk->tcp_rtx_queue = RB_ROOT;
414 tcp_init_xmit_timers(sk);
415 INIT_LIST_HEAD(&tp->tsq_node);
416 INIT_LIST_HEAD(&tp->tsorted_sent_queue);
418 icsk->icsk_rto = TCP_TIMEOUT_INIT;
419 tp->mdev_us = jiffies_to_usecs(TCP_TIMEOUT_INIT);
420 minmax_reset(&tp->rtt_min, tcp_jiffies32, ~0U);
422 /* So many TCP implementations out there (incorrectly) count the
423 * initial SYN frame in their delayed-ACK and congestion control
424 * algorithms that we must have the following bandaid to talk
425 * efficiently to them. -DaveM
427 tp->snd_cwnd = TCP_INIT_CWND;
429 /* There's a bubble in the pipe until at least the first ACK. */
430 tp->app_limited = ~0U;
432 /* See draft-stevens-tcpca-spec-01 for discussion of the
433 * initialization of these values.
435 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
436 tp->snd_cwnd_clamp = ~0;
437 tp->mss_cache = TCP_MSS_DEFAULT;
439 tp->reordering = sock_net(sk)->ipv4.sysctl_tcp_reordering;
440 tcp_assign_congestion_control(sk);
443 tp->rack.reo_wnd_steps = 1;
445 sk->sk_state = TCP_CLOSE;
447 sk->sk_write_space = sk_stream_write_space;
448 sock_set_flag(sk, SOCK_USE_WRITE_QUEUE);
450 icsk->icsk_sync_mss = tcp_sync_mss;
452 sk->sk_sndbuf = sock_net(sk)->ipv4.sysctl_tcp_wmem[1];
453 sk->sk_rcvbuf = sock_net(sk)->ipv4.sysctl_tcp_rmem[1];
455 sk_sockets_allocated_inc(sk);
456 sk->sk_route_forced_caps = NETIF_F_GSO;
458 EXPORT_SYMBOL(tcp_init_sock);
460 void tcp_init_transfer(struct sock *sk, int bpf_op)
462 struct inet_connection_sock *icsk = inet_csk(sk);
465 icsk->icsk_af_ops->rebuild_header(sk);
466 tcp_init_metrics(sk);
467 tcp_call_bpf(sk, bpf_op, 0, NULL);
468 tcp_init_congestion_control(sk);
469 tcp_init_buffer_space(sk);
472 static void tcp_tx_timestamp(struct sock *sk, u16 tsflags)
474 struct sk_buff *skb = tcp_write_queue_tail(sk);
476 if (tsflags && skb) {
477 struct skb_shared_info *shinfo = skb_shinfo(skb);
478 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
480 sock_tx_timestamp(sk, tsflags, &shinfo->tx_flags);
481 if (tsflags & SOF_TIMESTAMPING_TX_ACK)
482 tcb->txstamp_ack = 1;
483 if (tsflags & SOF_TIMESTAMPING_TX_RECORD_MASK)
484 shinfo->tskey = TCP_SKB_CB(skb)->seq + skb->len - 1;
488 static inline bool tcp_stream_is_readable(const struct tcp_sock *tp,
489 int target, struct sock *sk)
491 return (READ_ONCE(tp->rcv_nxt) - tp->copied_seq >= target) ||
492 (sk->sk_prot->stream_memory_read ?
493 sk->sk_prot->stream_memory_read(sk) : false);
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 || tp->fastopen_rsk)) {
558 int target = sock_rcvlowat(sk, 0, INT_MAX);
560 if (tp->urg_seq == 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)) {
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))
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 && tp->urg_seq == tp->copied_seq;
624 if (sk->sk_state == TCP_LISTEN)
627 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
630 answ = tp->write_seq - tp->snd_una;
633 if (sk->sk_state == TCP_LISTEN)
636 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
639 answ = tp->write_seq - tp->snd_nxt;
645 return put_user(answ, (int __user *)arg);
647 EXPORT_SYMBOL(tcp_ioctl);
649 static inline void tcp_mark_push(struct tcp_sock *tp, struct sk_buff *skb)
651 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
652 tp->pushed_seq = tp->write_seq;
655 static inline bool forced_push(const struct tcp_sock *tp)
657 return after(tp->write_seq, tp->pushed_seq + (tp->max_window >> 1));
660 static void skb_entail(struct sock *sk, struct sk_buff *skb)
662 struct tcp_sock *tp = tcp_sk(sk);
663 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
666 tcb->seq = tcb->end_seq = tp->write_seq;
667 tcb->tcp_flags = TCPHDR_ACK;
669 __skb_header_release(skb);
670 tcp_add_write_queue_tail(sk, skb);
671 sk->sk_wmem_queued += skb->truesize;
672 sk_mem_charge(sk, skb->truesize);
673 if (tp->nonagle & TCP_NAGLE_PUSH)
674 tp->nonagle &= ~TCP_NAGLE_PUSH;
676 tcp_slow_start_after_idle_check(sk);
679 static inline void tcp_mark_urg(struct tcp_sock *tp, int flags)
682 tp->snd_up = tp->write_seq;
685 /* If a not yet filled skb is pushed, do not send it if
686 * we have data packets in Qdisc or NIC queues :
687 * Because TX completion will happen shortly, it gives a chance
688 * to coalesce future sendmsg() payload into this skb, without
689 * need for a timer, and with no latency trade off.
690 * As packets containing data payload have a bigger truesize
691 * than pure acks (dataless) packets, the last checks prevent
692 * autocorking if we only have an ACK in Qdisc/NIC queues,
693 * or if TX completion was delayed after we processed ACK packet.
695 static bool tcp_should_autocork(struct sock *sk, struct sk_buff *skb,
698 return skb->len < size_goal &&
699 sock_net(sk)->ipv4.sysctl_tcp_autocorking &&
700 !tcp_rtx_queue_empty(sk) &&
701 refcount_read(&sk->sk_wmem_alloc) > skb->truesize;
704 static void tcp_push(struct sock *sk, int flags, int mss_now,
705 int nonagle, int size_goal)
707 struct tcp_sock *tp = tcp_sk(sk);
710 skb = tcp_write_queue_tail(sk);
713 if (!(flags & MSG_MORE) || forced_push(tp))
714 tcp_mark_push(tp, skb);
716 tcp_mark_urg(tp, flags);
718 if (tcp_should_autocork(sk, skb, size_goal)) {
720 /* avoid atomic op if TSQ_THROTTLED bit is already set */
721 if (!test_bit(TSQ_THROTTLED, &sk->sk_tsq_flags)) {
722 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPAUTOCORKING);
723 set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags);
725 /* It is possible TX completion already happened
726 * before we set TSQ_THROTTLED.
728 if (refcount_read(&sk->sk_wmem_alloc) > skb->truesize)
732 if (flags & MSG_MORE)
733 nonagle = TCP_NAGLE_CORK;
735 __tcp_push_pending_frames(sk, mss_now, nonagle);
738 static int tcp_splice_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb,
739 unsigned int offset, size_t len)
741 struct tcp_splice_state *tss = rd_desc->arg.data;
744 ret = skb_splice_bits(skb, skb->sk, offset, tss->pipe,
745 min(rd_desc->count, len), tss->flags);
747 rd_desc->count -= ret;
751 static int __tcp_splice_read(struct sock *sk, struct tcp_splice_state *tss)
753 /* Store TCP splice context information in read_descriptor_t. */
754 read_descriptor_t rd_desc = {
759 return tcp_read_sock(sk, &rd_desc, tcp_splice_data_recv);
763 * tcp_splice_read - splice data from TCP socket to a pipe
764 * @sock: socket to splice from
765 * @ppos: position (not valid)
766 * @pipe: pipe to splice to
767 * @len: number of bytes to splice
768 * @flags: splice modifier flags
771 * Will read pages from given socket and fill them into a pipe.
774 ssize_t tcp_splice_read(struct socket *sock, loff_t *ppos,
775 struct pipe_inode_info *pipe, size_t len,
778 struct sock *sk = sock->sk;
779 struct tcp_splice_state tss = {
788 sock_rps_record_flow(sk);
790 * We can't seek on a socket input
799 timeo = sock_rcvtimeo(sk, sock->file->f_flags & O_NONBLOCK);
801 ret = __tcp_splice_read(sk, &tss);
807 if (sock_flag(sk, SOCK_DONE))
810 ret = sock_error(sk);
813 if (sk->sk_shutdown & RCV_SHUTDOWN)
815 if (sk->sk_state == TCP_CLOSE) {
817 * This occurs when user tries to read
818 * from never connected socket.
827 /* if __tcp_splice_read() got nothing while we have
828 * an skb in receive queue, we do not want to loop.
829 * This might happen with URG data.
831 if (!skb_queue_empty(&sk->sk_receive_queue))
833 sk_wait_data(sk, &timeo, NULL);
834 if (signal_pending(current)) {
835 ret = sock_intr_errno(timeo);
848 if (sk->sk_err || sk->sk_state == TCP_CLOSE ||
849 (sk->sk_shutdown & RCV_SHUTDOWN) ||
850 signal_pending(current))
861 EXPORT_SYMBOL(tcp_splice_read);
863 struct sk_buff *sk_stream_alloc_skb(struct sock *sk, int size, gfp_t gfp,
868 /* The TCP header must be at least 32-bit aligned. */
869 size = ALIGN(size, 4);
871 if (unlikely(tcp_under_memory_pressure(sk)))
872 sk_mem_reclaim_partial(sk);
874 skb = alloc_skb_fclone(size + sk->sk_prot->max_header, gfp);
878 if (force_schedule) {
879 mem_scheduled = true;
880 sk_forced_mem_schedule(sk, skb->truesize);
882 mem_scheduled = sk_wmem_schedule(sk, skb->truesize);
884 if (likely(mem_scheduled)) {
885 skb_reserve(skb, sk->sk_prot->max_header);
887 * Make sure that we have exactly size bytes
888 * available to the caller, no more, no less.
890 skb->reserved_tailroom = skb->end - skb->tail - size;
891 INIT_LIST_HEAD(&skb->tcp_tsorted_anchor);
896 sk->sk_prot->enter_memory_pressure(sk);
897 sk_stream_moderate_sndbuf(sk);
902 static unsigned int tcp_xmit_size_goal(struct sock *sk, u32 mss_now,
905 struct tcp_sock *tp = tcp_sk(sk);
906 u32 new_size_goal, size_goal;
911 /* Note : tcp_tso_autosize() will eventually split this later */
912 new_size_goal = sk->sk_gso_max_size - 1 - MAX_TCP_HEADER;
913 new_size_goal = tcp_bound_to_half_wnd(tp, new_size_goal);
915 /* We try hard to avoid divides here */
916 size_goal = tp->gso_segs * mss_now;
917 if (unlikely(new_size_goal < size_goal ||
918 new_size_goal >= size_goal + mss_now)) {
919 tp->gso_segs = min_t(u16, new_size_goal / mss_now,
920 sk->sk_gso_max_segs);
921 size_goal = tp->gso_segs * mss_now;
924 return max(size_goal, mss_now);
927 static int tcp_send_mss(struct sock *sk, int *size_goal, int flags)
931 mss_now = tcp_current_mss(sk);
932 *size_goal = tcp_xmit_size_goal(sk, mss_now, !(flags & MSG_OOB));
937 /* In some cases, both sendpage() and sendmsg() could have added
938 * an skb to the write queue, but failed adding payload on it.
939 * We need to remove it to consume less memory, but more
940 * importantly be able to generate EPOLLOUT for Edge Trigger epoll()
943 static void tcp_remove_empty_skb(struct sock *sk, struct sk_buff *skb)
945 if (skb && !skb->len) {
946 tcp_unlink_write_queue(skb, sk);
947 if (tcp_write_queue_empty(sk))
948 tcp_chrono_stop(sk, TCP_CHRONO_BUSY);
949 sk_wmem_free_skb(sk, skb);
953 ssize_t do_tcp_sendpages(struct sock *sk, struct page *page, int offset,
954 size_t size, int flags)
956 struct tcp_sock *tp = tcp_sk(sk);
957 int mss_now, size_goal;
960 long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
962 /* Wait for a connection to finish. One exception is TCP Fast Open
963 * (passive side) where data is allowed to be sent before a connection
964 * is fully established.
966 if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) &&
967 !tcp_passive_fastopen(sk)) {
968 err = sk_stream_wait_connect(sk, &timeo);
973 sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
975 mss_now = tcp_send_mss(sk, &size_goal, flags);
979 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
983 struct sk_buff *skb = tcp_write_queue_tail(sk);
987 if (!skb || (copy = size_goal - skb->len) <= 0 ||
988 !tcp_skb_can_collapse_to(skb)) {
990 if (!sk_stream_memory_free(sk))
991 goto wait_for_sndbuf;
993 skb = sk_stream_alloc_skb(sk, 0, sk->sk_allocation,
994 tcp_rtx_and_write_queues_empty(sk));
996 goto wait_for_memory;
1005 i = skb_shinfo(skb)->nr_frags;
1006 can_coalesce = skb_can_coalesce(skb, i, page, offset);
1007 if (!can_coalesce && i >= sysctl_max_skb_frags) {
1008 tcp_mark_push(tp, skb);
1011 if (!sk_wmem_schedule(sk, copy))
1012 goto wait_for_memory;
1015 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1018 skb_fill_page_desc(skb, i, page, offset, copy);
1021 if (!(flags & MSG_NO_SHARED_FRAGS))
1022 skb_shinfo(skb)->tx_flags |= SKBTX_SHARED_FRAG;
1025 skb->data_len += copy;
1026 skb->truesize += copy;
1027 sk->sk_wmem_queued += copy;
1028 sk_mem_charge(sk, copy);
1029 skb->ip_summed = CHECKSUM_PARTIAL;
1030 tp->write_seq += copy;
1031 TCP_SKB_CB(skb)->end_seq += copy;
1032 tcp_skb_pcount_set(skb, 0);
1035 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
1043 if (skb->len < size_goal || (flags & MSG_OOB))
1046 if (forced_push(tp)) {
1047 tcp_mark_push(tp, skb);
1048 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
1049 } else if (skb == tcp_send_head(sk))
1050 tcp_push_one(sk, mss_now);
1054 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1056 tcp_push(sk, flags & ~MSG_MORE, mss_now,
1057 TCP_NAGLE_PUSH, size_goal);
1059 err = sk_stream_wait_memory(sk, &timeo);
1063 mss_now = tcp_send_mss(sk, &size_goal, flags);
1068 tcp_tx_timestamp(sk, sk->sk_tsflags);
1069 if (!(flags & MSG_SENDPAGE_NOTLAST))
1070 tcp_push(sk, flags, mss_now, tp->nonagle, size_goal);
1075 tcp_remove_empty_skb(sk, tcp_write_queue_tail(sk));
1079 /* make sure we wake any epoll edge trigger waiter */
1080 if (unlikely(tcp_rtx_and_write_queues_empty(sk) && err == -EAGAIN)) {
1081 sk->sk_write_space(sk);
1082 tcp_chrono_stop(sk, TCP_CHRONO_SNDBUF_LIMITED);
1084 return sk_stream_error(sk, flags, err);
1086 EXPORT_SYMBOL_GPL(do_tcp_sendpages);
1088 int tcp_sendpage_locked(struct sock *sk, struct page *page, int offset,
1089 size_t size, int flags)
1091 if (!(sk->sk_route_caps & NETIF_F_SG))
1092 return sock_no_sendpage_locked(sk, page, offset, size, flags);
1094 tcp_rate_check_app_limited(sk); /* is sending application-limited? */
1096 return do_tcp_sendpages(sk, page, offset, size, flags);
1098 EXPORT_SYMBOL_GPL(tcp_sendpage_locked);
1100 int tcp_sendpage(struct sock *sk, struct page *page, int offset,
1101 size_t size, int flags)
1106 ret = tcp_sendpage_locked(sk, page, offset, size, flags);
1111 EXPORT_SYMBOL(tcp_sendpage);
1113 /* Do not bother using a page frag for very small frames.
1114 * But use this heuristic only for the first skb in write queue.
1116 * Having no payload in skb->head allows better SACK shifting
1117 * in tcp_shift_skb_data(), reducing sack/rack overhead, because
1118 * write queue has less skbs.
1119 * Each skb can hold up to MAX_SKB_FRAGS * 32Kbytes, or ~0.5 MB.
1120 * This also speeds up tso_fragment(), since it wont fallback
1121 * to tcp_fragment().
1123 static int linear_payload_sz(bool first_skb)
1126 return SKB_WITH_OVERHEAD(2048 - MAX_TCP_HEADER);
1130 static int select_size(bool first_skb, bool zc)
1134 return linear_payload_sz(first_skb);
1137 void tcp_free_fastopen_req(struct tcp_sock *tp)
1139 if (tp->fastopen_req) {
1140 kfree(tp->fastopen_req);
1141 tp->fastopen_req = NULL;
1145 static int tcp_sendmsg_fastopen(struct sock *sk, struct msghdr *msg,
1146 int *copied, size_t size)
1148 struct tcp_sock *tp = tcp_sk(sk);
1149 struct inet_sock *inet = inet_sk(sk);
1150 struct sockaddr *uaddr = msg->msg_name;
1153 if (!(sock_net(sk)->ipv4.sysctl_tcp_fastopen & TFO_CLIENT_ENABLE) ||
1154 (uaddr && msg->msg_namelen >= sizeof(uaddr->sa_family) &&
1155 uaddr->sa_family == AF_UNSPEC))
1157 if (tp->fastopen_req)
1158 return -EALREADY; /* Another Fast Open is in progress */
1160 tp->fastopen_req = kzalloc(sizeof(struct tcp_fastopen_request),
1162 if (unlikely(!tp->fastopen_req))
1164 tp->fastopen_req->data = msg;
1165 tp->fastopen_req->size = size;
1167 if (inet->defer_connect) {
1168 err = tcp_connect(sk);
1169 /* Same failure procedure as in tcp_v4/6_connect */
1171 tcp_set_state(sk, TCP_CLOSE);
1172 inet->inet_dport = 0;
1173 sk->sk_route_caps = 0;
1176 flags = (msg->msg_flags & MSG_DONTWAIT) ? O_NONBLOCK : 0;
1177 err = __inet_stream_connect(sk->sk_socket, uaddr,
1178 msg->msg_namelen, flags, 1);
1179 /* fastopen_req could already be freed in __inet_stream_connect
1180 * if the connection times out or gets rst
1182 if (tp->fastopen_req) {
1183 *copied = tp->fastopen_req->copied;
1184 tcp_free_fastopen_req(tp);
1185 inet->defer_connect = 0;
1190 int tcp_sendmsg_locked(struct sock *sk, struct msghdr *msg, size_t size)
1192 struct tcp_sock *tp = tcp_sk(sk);
1193 struct ubuf_info *uarg = NULL;
1194 struct sk_buff *skb;
1195 struct sockcm_cookie sockc;
1196 int flags, err, copied = 0;
1197 int mss_now = 0, size_goal, copied_syn = 0;
1198 bool process_backlog = false;
1202 flags = msg->msg_flags;
1204 if (flags & MSG_ZEROCOPY && size && sock_flag(sk, SOCK_ZEROCOPY)) {
1205 if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) {
1210 skb = tcp_write_queue_tail(sk);
1211 uarg = sock_zerocopy_realloc(sk, size, skb_zcopy(skb));
1217 zc = sk->sk_route_caps & NETIF_F_SG;
1222 if (unlikely(flags & MSG_FASTOPEN || inet_sk(sk)->defer_connect) &&
1224 err = tcp_sendmsg_fastopen(sk, msg, &copied_syn, size);
1225 if (err == -EINPROGRESS && copied_syn > 0)
1231 timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
1233 tcp_rate_check_app_limited(sk); /* is sending application-limited? */
1235 /* Wait for a connection to finish. One exception is TCP Fast Open
1236 * (passive side) where data is allowed to be sent before a connection
1237 * is fully established.
1239 if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) &&
1240 !tcp_passive_fastopen(sk)) {
1241 err = sk_stream_wait_connect(sk, &timeo);
1246 if (unlikely(tp->repair)) {
1247 if (tp->repair_queue == TCP_RECV_QUEUE) {
1248 copied = tcp_send_rcvq(sk, msg, size);
1253 if (tp->repair_queue == TCP_NO_QUEUE)
1256 /* 'common' sending to sendq */
1259 sockcm_init(&sockc, sk);
1260 if (msg->msg_controllen) {
1261 err = sock_cmsg_send(sk, msg, &sockc);
1262 if (unlikely(err)) {
1268 /* This should be in poll */
1269 sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
1271 /* Ok commence sending. */
1275 mss_now = tcp_send_mss(sk, &size_goal, flags);
1278 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
1281 while (msg_data_left(msg)) {
1284 skb = tcp_write_queue_tail(sk);
1286 copy = size_goal - skb->len;
1288 if (copy <= 0 || !tcp_skb_can_collapse_to(skb)) {
1293 if (!sk_stream_memory_free(sk))
1294 goto wait_for_sndbuf;
1296 if (process_backlog && sk_flush_backlog(sk)) {
1297 process_backlog = false;
1300 first_skb = tcp_rtx_and_write_queues_empty(sk);
1301 linear = select_size(first_skb, zc);
1302 skb = sk_stream_alloc_skb(sk, linear, sk->sk_allocation,
1305 goto wait_for_memory;
1307 process_backlog = true;
1308 skb->ip_summed = CHECKSUM_PARTIAL;
1310 skb_entail(sk, skb);
1313 /* All packets are restored as if they have
1314 * already been sent. skb_mstamp isn't set to
1315 * avoid wrong rtt estimation.
1318 TCP_SKB_CB(skb)->sacked |= TCPCB_REPAIRED;
1321 /* Try to append data to the end of skb. */
1322 if (copy > msg_data_left(msg))
1323 copy = msg_data_left(msg);
1325 /* Where to copy to? */
1326 if (skb_availroom(skb) > 0 && !zc) {
1327 /* We have some space in skb head. Superb! */
1328 copy = min_t(int, copy, skb_availroom(skb));
1329 err = skb_add_data_nocache(sk, skb, &msg->msg_iter, copy);
1334 int i = skb_shinfo(skb)->nr_frags;
1335 struct page_frag *pfrag = sk_page_frag(sk);
1337 if (!sk_page_frag_refill(sk, pfrag))
1338 goto wait_for_memory;
1340 if (!skb_can_coalesce(skb, i, pfrag->page,
1342 if (i >= sysctl_max_skb_frags) {
1343 tcp_mark_push(tp, skb);
1349 copy = min_t(int, copy, pfrag->size - pfrag->offset);
1351 if (!sk_wmem_schedule(sk, copy))
1352 goto wait_for_memory;
1354 err = skb_copy_to_page_nocache(sk, &msg->msg_iter, skb,
1361 /* Update the skb. */
1363 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1365 skb_fill_page_desc(skb, i, pfrag->page,
1366 pfrag->offset, copy);
1367 page_ref_inc(pfrag->page);
1369 pfrag->offset += copy;
1371 err = skb_zerocopy_iter_stream(sk, skb, msg, copy, uarg);
1372 if (err == -EMSGSIZE || err == -EEXIST) {
1373 tcp_mark_push(tp, skb);
1382 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
1384 tp->write_seq += copy;
1385 TCP_SKB_CB(skb)->end_seq += copy;
1386 tcp_skb_pcount_set(skb, 0);
1389 if (!msg_data_left(msg)) {
1390 if (unlikely(flags & MSG_EOR))
1391 TCP_SKB_CB(skb)->eor = 1;
1395 if (skb->len < size_goal || (flags & MSG_OOB) || unlikely(tp->repair))
1398 if (forced_push(tp)) {
1399 tcp_mark_push(tp, skb);
1400 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
1401 } else if (skb == tcp_send_head(sk))
1402 tcp_push_one(sk, mss_now);
1406 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1409 tcp_push(sk, flags & ~MSG_MORE, mss_now,
1410 TCP_NAGLE_PUSH, size_goal);
1412 err = sk_stream_wait_memory(sk, &timeo);
1416 mss_now = tcp_send_mss(sk, &size_goal, flags);
1421 tcp_tx_timestamp(sk, sockc.tsflags);
1422 tcp_push(sk, flags, mss_now, tp->nonagle, size_goal);
1425 sock_zerocopy_put(uarg);
1426 return copied + copied_syn;
1429 skb = tcp_write_queue_tail(sk);
1431 tcp_remove_empty_skb(sk, skb);
1433 if (copied + copied_syn)
1436 sock_zerocopy_put_abort(uarg);
1437 err = sk_stream_error(sk, flags, err);
1438 /* make sure we wake any epoll edge trigger waiter */
1439 if (unlikely(tcp_rtx_and_write_queues_empty(sk) && err == -EAGAIN)) {
1440 sk->sk_write_space(sk);
1441 tcp_chrono_stop(sk, TCP_CHRONO_SNDBUF_LIMITED);
1445 EXPORT_SYMBOL_GPL(tcp_sendmsg_locked);
1447 int tcp_sendmsg(struct sock *sk, struct msghdr *msg, size_t size)
1452 ret = tcp_sendmsg_locked(sk, msg, size);
1457 EXPORT_SYMBOL(tcp_sendmsg);
1460 * Handle reading urgent data. BSD has very simple semantics for
1461 * this, no blocking and very strange errors 8)
1464 static int tcp_recv_urg(struct sock *sk, struct msghdr *msg, int len, int flags)
1466 struct tcp_sock *tp = tcp_sk(sk);
1468 /* No URG data to read. */
1469 if (sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data ||
1470 tp->urg_data == TCP_URG_READ)
1471 return -EINVAL; /* Yes this is right ! */
1473 if (sk->sk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DONE))
1476 if (tp->urg_data & TCP_URG_VALID) {
1478 char c = tp->urg_data;
1480 if (!(flags & MSG_PEEK))
1481 tp->urg_data = TCP_URG_READ;
1483 /* Read urgent data. */
1484 msg->msg_flags |= MSG_OOB;
1487 if (!(flags & MSG_TRUNC))
1488 err = memcpy_to_msg(msg, &c, 1);
1491 msg->msg_flags |= MSG_TRUNC;
1493 return err ? -EFAULT : len;
1496 if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
1499 /* Fixed the recv(..., MSG_OOB) behaviour. BSD docs and
1500 * the available implementations agree in this case:
1501 * this call should never block, independent of the
1502 * blocking state of the socket.
1503 * Mike <pall@rz.uni-karlsruhe.de>
1508 static int tcp_peek_sndq(struct sock *sk, struct msghdr *msg, int len)
1510 struct sk_buff *skb;
1511 int copied = 0, err = 0;
1513 /* XXX -- need to support SO_PEEK_OFF */
1515 skb_rbtree_walk(skb, &sk->tcp_rtx_queue) {
1516 err = skb_copy_datagram_msg(skb, 0, msg, skb->len);
1522 skb_queue_walk(&sk->sk_write_queue, skb) {
1523 err = skb_copy_datagram_msg(skb, 0, msg, skb->len);
1530 return err ?: copied;
1533 /* Clean up the receive buffer for full frames taken by the user,
1534 * then send an ACK if necessary. COPIED is the number of bytes
1535 * tcp_recvmsg has given to the user so far, it speeds up the
1536 * calculation of whether or not we must ACK for the sake of
1539 static void tcp_cleanup_rbuf(struct sock *sk, int copied)
1541 struct tcp_sock *tp = tcp_sk(sk);
1542 bool time_to_ack = false;
1544 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
1546 WARN(skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq),
1547 "cleanup rbuf bug: copied %X seq %X rcvnxt %X\n",
1548 tp->copied_seq, TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt);
1550 if (inet_csk_ack_scheduled(sk)) {
1551 const struct inet_connection_sock *icsk = inet_csk(sk);
1552 /* Delayed ACKs frequently hit locked sockets during bulk
1554 if (icsk->icsk_ack.blocked ||
1555 /* Once-per-two-segments ACK was not sent by tcp_input.c */
1556 tp->rcv_nxt - tp->rcv_wup > icsk->icsk_ack.rcv_mss ||
1558 * If this read emptied read buffer, we send ACK, if
1559 * connection is not bidirectional, user drained
1560 * receive buffer and there was a small segment
1564 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED2) ||
1565 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED) &&
1566 !icsk->icsk_ack.pingpong)) &&
1567 !atomic_read(&sk->sk_rmem_alloc)))
1571 /* We send an ACK if we can now advertise a non-zero window
1572 * which has been raised "significantly".
1574 * Even if window raised up to infinity, do not send window open ACK
1575 * in states, where we will not receive more. It is useless.
1577 if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) {
1578 __u32 rcv_window_now = tcp_receive_window(tp);
1580 /* Optimize, __tcp_select_window() is not cheap. */
1581 if (2*rcv_window_now <= tp->window_clamp) {
1582 __u32 new_window = __tcp_select_window(sk);
1584 /* Send ACK now, if this read freed lots of space
1585 * in our buffer. Certainly, new_window is new window.
1586 * We can advertise it now, if it is not less than current one.
1587 * "Lots" means "at least twice" here.
1589 if (new_window && new_window >= 2 * rcv_window_now)
1597 static struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off)
1599 struct sk_buff *skb;
1602 while ((skb = skb_peek(&sk->sk_receive_queue)) != NULL) {
1603 offset = seq - TCP_SKB_CB(skb)->seq;
1604 if (unlikely(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
1605 pr_err_once("%s: found a SYN, please report !\n", __func__);
1608 if (offset < skb->len || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)) {
1612 /* This looks weird, but this can happen if TCP collapsing
1613 * splitted a fat GRO packet, while we released socket lock
1614 * in skb_splice_bits()
1616 sk_eat_skb(sk, skb);
1622 * This routine provides an alternative to tcp_recvmsg() for routines
1623 * that would like to handle copying from skbuffs directly in 'sendfile'
1626 * - It is assumed that the socket was locked by the caller.
1627 * - The routine does not block.
1628 * - At present, there is no support for reading OOB data
1629 * or for 'peeking' the socket using this routine
1630 * (although both would be easy to implement).
1632 int tcp_read_sock(struct sock *sk, read_descriptor_t *desc,
1633 sk_read_actor_t recv_actor)
1635 struct sk_buff *skb;
1636 struct tcp_sock *tp = tcp_sk(sk);
1637 u32 seq = tp->copied_seq;
1641 if (sk->sk_state == TCP_LISTEN)
1643 while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1644 if (offset < skb->len) {
1648 len = skb->len - offset;
1649 /* Stop reading if we hit a patch of urgent data */
1651 u32 urg_offset = tp->urg_seq - seq;
1652 if (urg_offset < len)
1657 used = recv_actor(desc, skb, offset, len);
1662 } else if (used <= len) {
1667 /* If recv_actor drops the lock (e.g. TCP splice
1668 * receive) the skb pointer might be invalid when
1669 * getting here: tcp_collapse might have deleted it
1670 * while aggregating skbs from the socket queue.
1672 skb = tcp_recv_skb(sk, seq - 1, &offset);
1675 /* TCP coalescing might have appended data to the skb.
1676 * Try to splice more frags
1678 if (offset + 1 != skb->len)
1681 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) {
1682 sk_eat_skb(sk, skb);
1686 sk_eat_skb(sk, skb);
1689 tp->copied_seq = seq;
1691 tp->copied_seq = seq;
1693 tcp_rcv_space_adjust(sk);
1695 /* Clean up data we have read: This will do ACK frames. */
1697 tcp_recv_skb(sk, seq, &offset);
1698 tcp_cleanup_rbuf(sk, copied);
1702 EXPORT_SYMBOL(tcp_read_sock);
1704 int tcp_peek_len(struct socket *sock)
1706 return tcp_inq(sock->sk);
1708 EXPORT_SYMBOL(tcp_peek_len);
1710 /* Make sure sk_rcvbuf is big enough to satisfy SO_RCVLOWAT hint */
1711 int tcp_set_rcvlowat(struct sock *sk, int val)
1715 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK)
1716 cap = sk->sk_rcvbuf >> 1;
1718 cap = sock_net(sk)->ipv4.sysctl_tcp_rmem[2] >> 1;
1719 val = min(val, cap);
1720 sk->sk_rcvlowat = val ? : 1;
1722 /* Check if we need to signal EPOLLIN right now */
1725 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK)
1729 if (val > sk->sk_rcvbuf) {
1730 sk->sk_rcvbuf = val;
1731 tcp_sk(sk)->window_clamp = tcp_win_from_space(sk, val);
1735 EXPORT_SYMBOL(tcp_set_rcvlowat);
1738 static const struct vm_operations_struct tcp_vm_ops = {
1741 int tcp_mmap(struct file *file, struct socket *sock,
1742 struct vm_area_struct *vma)
1744 if (vma->vm_flags & (VM_WRITE | VM_EXEC))
1746 vma->vm_flags &= ~(VM_MAYWRITE | VM_MAYEXEC);
1748 /* Instruct vm_insert_page() to not down_read(mmap_sem) */
1749 vma->vm_flags |= VM_MIXEDMAP;
1751 vma->vm_ops = &tcp_vm_ops;
1754 EXPORT_SYMBOL(tcp_mmap);
1756 static int tcp_zerocopy_receive(struct sock *sk,
1757 struct tcp_zerocopy_receive *zc)
1759 unsigned long address = (unsigned long)zc->address;
1760 const skb_frag_t *frags = NULL;
1761 u32 length = 0, seq, offset;
1762 struct vm_area_struct *vma;
1763 struct sk_buff *skb = NULL;
1764 struct tcp_sock *tp;
1767 if (address & (PAGE_SIZE - 1) || address != zc->address)
1770 if (sk->sk_state == TCP_LISTEN)
1773 sock_rps_record_flow(sk);
1775 down_read(¤t->mm->mmap_sem);
1778 vma = find_vma(current->mm, address);
1779 if (!vma || vma->vm_start > address || vma->vm_ops != &tcp_vm_ops)
1781 zc->length = min_t(unsigned long, zc->length, vma->vm_end - address);
1784 seq = tp->copied_seq;
1785 zc->length = min_t(u32, zc->length, tcp_inq(sk));
1786 zc->length &= ~(PAGE_SIZE - 1);
1788 zap_page_range(vma, address, zc->length);
1790 zc->recv_skip_hint = 0;
1792 while (length + PAGE_SIZE <= zc->length) {
1793 if (zc->recv_skip_hint < PAGE_SIZE) {
1796 offset = seq - TCP_SKB_CB(skb)->seq;
1798 skb = tcp_recv_skb(sk, seq, &offset);
1801 zc->recv_skip_hint = skb->len - offset;
1802 offset -= skb_headlen(skb);
1803 if ((int)offset < 0 || skb_has_frag_list(skb))
1805 frags = skb_shinfo(skb)->frags;
1807 if (frags->size > offset)
1809 offset -= frags->size;
1813 if (frags->size != PAGE_SIZE || frags->page_offset)
1815 ret = vm_insert_page(vma, address + length,
1816 skb_frag_page(frags));
1819 length += PAGE_SIZE;
1821 zc->recv_skip_hint -= PAGE_SIZE;
1825 up_read(¤t->mm->mmap_sem);
1827 tp->copied_seq = seq;
1828 tcp_rcv_space_adjust(sk);
1830 /* Clean up data we have read: This will do ACK frames. */
1831 tcp_recv_skb(sk, seq, &offset);
1832 tcp_cleanup_rbuf(sk, length);
1834 if (length == zc->length)
1835 zc->recv_skip_hint = 0;
1837 if (!zc->recv_skip_hint && sock_flag(sk, SOCK_DONE))
1840 zc->length = length;
1845 static void tcp_update_recv_tstamps(struct sk_buff *skb,
1846 struct scm_timestamping *tss)
1849 tss->ts[0] = ktime_to_timespec(skb->tstamp);
1851 tss->ts[0] = (struct timespec) {0};
1853 if (skb_hwtstamps(skb)->hwtstamp)
1854 tss->ts[2] = ktime_to_timespec(skb_hwtstamps(skb)->hwtstamp);
1856 tss->ts[2] = (struct timespec) {0};
1859 /* Similar to __sock_recv_timestamp, but does not require an skb */
1860 static void tcp_recv_timestamp(struct msghdr *msg, const struct sock *sk,
1861 struct scm_timestamping *tss)
1864 bool has_timestamping = false;
1866 if (tss->ts[0].tv_sec || tss->ts[0].tv_nsec) {
1867 if (sock_flag(sk, SOCK_RCVTSTAMP)) {
1868 if (sock_flag(sk, SOCK_RCVTSTAMPNS)) {
1869 put_cmsg(msg, SOL_SOCKET, SCM_TIMESTAMPNS,
1870 sizeof(tss->ts[0]), &tss->ts[0]);
1872 tv.tv_sec = tss->ts[0].tv_sec;
1873 tv.tv_usec = tss->ts[0].tv_nsec / 1000;
1875 put_cmsg(msg, SOL_SOCKET, SCM_TIMESTAMP,
1880 if (sk->sk_tsflags & SOF_TIMESTAMPING_SOFTWARE)
1881 has_timestamping = true;
1883 tss->ts[0] = (struct timespec) {0};
1886 if (tss->ts[2].tv_sec || tss->ts[2].tv_nsec) {
1887 if (sk->sk_tsflags & SOF_TIMESTAMPING_RAW_HARDWARE)
1888 has_timestamping = true;
1890 tss->ts[2] = (struct timespec) {0};
1893 if (has_timestamping) {
1894 tss->ts[1] = (struct timespec) {0};
1895 put_cmsg(msg, SOL_SOCKET, SCM_TIMESTAMPING,
1900 static int tcp_inq_hint(struct sock *sk)
1902 const struct tcp_sock *tp = tcp_sk(sk);
1903 u32 copied_seq = READ_ONCE(tp->copied_seq);
1904 u32 rcv_nxt = READ_ONCE(tp->rcv_nxt);
1907 inq = rcv_nxt - copied_seq;
1908 if (unlikely(inq < 0 || copied_seq != READ_ONCE(tp->copied_seq))) {
1910 inq = tp->rcv_nxt - tp->copied_seq;
1913 /* After receiving a FIN, tell the user-space to continue reading
1914 * by returning a non-zero inq.
1916 if (inq == 0 && sock_flag(sk, SOCK_DONE))
1922 * This routine copies from a sock struct into the user buffer.
1924 * Technical note: in 2.3 we work on _locked_ socket, so that
1925 * tricks with *seq access order and skb->users are not required.
1926 * Probably, code can be easily improved even more.
1929 int tcp_recvmsg(struct sock *sk, struct msghdr *msg, size_t len, int nonblock,
1930 int flags, int *addr_len)
1932 struct tcp_sock *tp = tcp_sk(sk);
1938 int target; /* Read at least this many bytes */
1940 struct sk_buff *skb, *last;
1942 struct scm_timestamping tss;
1943 bool has_tss = false;
1946 if (unlikely(flags & MSG_ERRQUEUE))
1947 return inet_recv_error(sk, msg, len, addr_len);
1949 if (sk_can_busy_loop(sk) && skb_queue_empty_lockless(&sk->sk_receive_queue) &&
1950 (sk->sk_state == TCP_ESTABLISHED))
1951 sk_busy_loop(sk, nonblock);
1956 if (sk->sk_state == TCP_LISTEN)
1959 has_cmsg = tp->recvmsg_inq;
1960 timeo = sock_rcvtimeo(sk, nonblock);
1962 /* Urgent data needs to be handled specially. */
1963 if (flags & MSG_OOB)
1966 if (unlikely(tp->repair)) {
1968 if (!(flags & MSG_PEEK))
1971 if (tp->repair_queue == TCP_SEND_QUEUE)
1975 if (tp->repair_queue == TCP_NO_QUEUE)
1978 /* 'common' recv queue MSG_PEEK-ing */
1981 seq = &tp->copied_seq;
1982 if (flags & MSG_PEEK) {
1983 peek_seq = tp->copied_seq;
1987 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
1992 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
1993 if (tp->urg_data && tp->urg_seq == *seq) {
1996 if (signal_pending(current)) {
1997 copied = timeo ? sock_intr_errno(timeo) : -EAGAIN;
2002 /* Next get a buffer. */
2004 last = skb_peek_tail(&sk->sk_receive_queue);
2005 skb_queue_walk(&sk->sk_receive_queue, skb) {
2007 /* Now that we have two receive queues this
2010 if (WARN(before(*seq, TCP_SKB_CB(skb)->seq),
2011 "TCP recvmsg seq # bug: copied %X, seq %X, rcvnxt %X, fl %X\n",
2012 *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt,
2016 offset = *seq - TCP_SKB_CB(skb)->seq;
2017 if (unlikely(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
2018 pr_err_once("%s: found a SYN, please report !\n", __func__);
2021 if (offset < skb->len)
2023 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
2025 WARN(!(flags & MSG_PEEK),
2026 "TCP recvmsg seq # bug 2: copied %X, seq %X, rcvnxt %X, fl %X\n",
2027 *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt, flags);
2030 /* Well, if we have backlog, try to process it now yet. */
2032 if (copied >= target && !sk->sk_backlog.tail)
2037 sk->sk_state == TCP_CLOSE ||
2038 (sk->sk_shutdown & RCV_SHUTDOWN) ||
2040 signal_pending(current))
2043 if (sock_flag(sk, SOCK_DONE))
2047 copied = sock_error(sk);
2051 if (sk->sk_shutdown & RCV_SHUTDOWN)
2054 if (sk->sk_state == TCP_CLOSE) {
2055 /* This occurs when user tries to read
2056 * from never connected socket.
2067 if (signal_pending(current)) {
2068 copied = sock_intr_errno(timeo);
2073 tcp_cleanup_rbuf(sk, copied);
2075 if (copied >= target) {
2076 /* Do not sleep, just process backlog. */
2080 sk_wait_data(sk, &timeo, last);
2083 if ((flags & MSG_PEEK) &&
2084 (peek_seq - copied - urg_hole != tp->copied_seq)) {
2085 net_dbg_ratelimited("TCP(%s:%d): Application bug, race in MSG_PEEK\n",
2087 task_pid_nr(current));
2088 peek_seq = tp->copied_seq;
2093 /* Ok so how much can we use? */
2094 used = skb->len - offset;
2098 /* Do we have urgent data here? */
2100 u32 urg_offset = tp->urg_seq - *seq;
2101 if (urg_offset < used) {
2103 if (!sock_flag(sk, SOCK_URGINLINE)) {
2116 if (!(flags & MSG_TRUNC)) {
2117 err = skb_copy_datagram_msg(skb, offset, msg, used);
2119 /* Exception. Bailout! */
2130 tcp_rcv_space_adjust(sk);
2133 if (tp->urg_data && after(tp->copied_seq, tp->urg_seq)) {
2135 tcp_fast_path_check(sk);
2137 if (used + offset < skb->len)
2140 if (TCP_SKB_CB(skb)->has_rxtstamp) {
2141 tcp_update_recv_tstamps(skb, &tss);
2145 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
2147 if (!(flags & MSG_PEEK))
2148 sk_eat_skb(sk, skb);
2152 /* Process the FIN. */
2154 if (!(flags & MSG_PEEK))
2155 sk_eat_skb(sk, skb);
2159 /* According to UNIX98, msg_name/msg_namelen are ignored
2160 * on connected socket. I was just happy when found this 8) --ANK
2163 /* Clean up data we have read: This will do ACK frames. */
2164 tcp_cleanup_rbuf(sk, copied);
2170 tcp_recv_timestamp(msg, sk, &tss);
2171 if (tp->recvmsg_inq) {
2172 inq = tcp_inq_hint(sk);
2173 put_cmsg(msg, SOL_TCP, TCP_CM_INQ, sizeof(inq), &inq);
2184 err = tcp_recv_urg(sk, msg, len, flags);
2188 err = tcp_peek_sndq(sk, msg, len);
2191 EXPORT_SYMBOL(tcp_recvmsg);
2193 void tcp_set_state(struct sock *sk, int state)
2195 int oldstate = sk->sk_state;
2197 /* We defined a new enum for TCP states that are exported in BPF
2198 * so as not force the internal TCP states to be frozen. The
2199 * following checks will detect if an internal state value ever
2200 * differs from the BPF value. If this ever happens, then we will
2201 * need to remap the internal value to the BPF value before calling
2202 * tcp_call_bpf_2arg.
2204 BUILD_BUG_ON((int)BPF_TCP_ESTABLISHED != (int)TCP_ESTABLISHED);
2205 BUILD_BUG_ON((int)BPF_TCP_SYN_SENT != (int)TCP_SYN_SENT);
2206 BUILD_BUG_ON((int)BPF_TCP_SYN_RECV != (int)TCP_SYN_RECV);
2207 BUILD_BUG_ON((int)BPF_TCP_FIN_WAIT1 != (int)TCP_FIN_WAIT1);
2208 BUILD_BUG_ON((int)BPF_TCP_FIN_WAIT2 != (int)TCP_FIN_WAIT2);
2209 BUILD_BUG_ON((int)BPF_TCP_TIME_WAIT != (int)TCP_TIME_WAIT);
2210 BUILD_BUG_ON((int)BPF_TCP_CLOSE != (int)TCP_CLOSE);
2211 BUILD_BUG_ON((int)BPF_TCP_CLOSE_WAIT != (int)TCP_CLOSE_WAIT);
2212 BUILD_BUG_ON((int)BPF_TCP_LAST_ACK != (int)TCP_LAST_ACK);
2213 BUILD_BUG_ON((int)BPF_TCP_LISTEN != (int)TCP_LISTEN);
2214 BUILD_BUG_ON((int)BPF_TCP_CLOSING != (int)TCP_CLOSING);
2215 BUILD_BUG_ON((int)BPF_TCP_NEW_SYN_RECV != (int)TCP_NEW_SYN_RECV);
2216 BUILD_BUG_ON((int)BPF_TCP_MAX_STATES != (int)TCP_MAX_STATES);
2218 if (BPF_SOCK_OPS_TEST_FLAG(tcp_sk(sk), BPF_SOCK_OPS_STATE_CB_FLAG))
2219 tcp_call_bpf_2arg(sk, BPF_SOCK_OPS_STATE_CB, oldstate, state);
2222 case TCP_ESTABLISHED:
2223 if (oldstate != TCP_ESTABLISHED)
2224 TCP_INC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
2228 if (oldstate == TCP_CLOSE_WAIT || oldstate == TCP_ESTABLISHED)
2229 TCP_INC_STATS(sock_net(sk), TCP_MIB_ESTABRESETS);
2231 sk->sk_prot->unhash(sk);
2232 if (inet_csk(sk)->icsk_bind_hash &&
2233 !(sk->sk_userlocks & SOCK_BINDPORT_LOCK))
2237 if (oldstate == TCP_ESTABLISHED)
2238 TCP_DEC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
2241 /* Change state AFTER socket is unhashed to avoid closed
2242 * socket sitting in hash tables.
2244 inet_sk_state_store(sk, state);
2247 SOCK_DEBUG(sk, "TCP sk=%p, State %s -> %s\n", sk, statename[oldstate], statename[state]);
2250 EXPORT_SYMBOL_GPL(tcp_set_state);
2253 * State processing on a close. This implements the state shift for
2254 * sending our FIN frame. Note that we only send a FIN for some
2255 * states. A shutdown() may have already sent the FIN, or we may be
2259 static const unsigned char new_state[16] = {
2260 /* current state: new state: action: */
2261 [0 /* (Invalid) */] = TCP_CLOSE,
2262 [TCP_ESTABLISHED] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2263 [TCP_SYN_SENT] = TCP_CLOSE,
2264 [TCP_SYN_RECV] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2265 [TCP_FIN_WAIT1] = TCP_FIN_WAIT1,
2266 [TCP_FIN_WAIT2] = TCP_FIN_WAIT2,
2267 [TCP_TIME_WAIT] = TCP_CLOSE,
2268 [TCP_CLOSE] = TCP_CLOSE,
2269 [TCP_CLOSE_WAIT] = TCP_LAST_ACK | TCP_ACTION_FIN,
2270 [TCP_LAST_ACK] = TCP_LAST_ACK,
2271 [TCP_LISTEN] = TCP_CLOSE,
2272 [TCP_CLOSING] = TCP_CLOSING,
2273 [TCP_NEW_SYN_RECV] = TCP_CLOSE, /* should not happen ! */
2276 static int tcp_close_state(struct sock *sk)
2278 int next = (int)new_state[sk->sk_state];
2279 int ns = next & TCP_STATE_MASK;
2281 tcp_set_state(sk, ns);
2283 return next & TCP_ACTION_FIN;
2287 * Shutdown the sending side of a connection. Much like close except
2288 * that we don't receive shut down or sock_set_flag(sk, SOCK_DEAD).
2291 void tcp_shutdown(struct sock *sk, int how)
2293 /* We need to grab some memory, and put together a FIN,
2294 * and then put it into the queue to be sent.
2295 * Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
2297 if (!(how & SEND_SHUTDOWN))
2300 /* If we've already sent a FIN, or it's a closed state, skip this. */
2301 if ((1 << sk->sk_state) &
2302 (TCPF_ESTABLISHED | TCPF_SYN_SENT |
2303 TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) {
2304 /* Clear out any half completed packets. FIN if needed. */
2305 if (tcp_close_state(sk))
2309 EXPORT_SYMBOL(tcp_shutdown);
2311 bool tcp_check_oom(struct sock *sk, int shift)
2313 bool too_many_orphans, out_of_socket_memory;
2315 too_many_orphans = tcp_too_many_orphans(sk, shift);
2316 out_of_socket_memory = tcp_out_of_memory(sk);
2318 if (too_many_orphans)
2319 net_info_ratelimited("too many orphaned sockets\n");
2320 if (out_of_socket_memory)
2321 net_info_ratelimited("out of memory -- consider tuning tcp_mem\n");
2322 return too_many_orphans || out_of_socket_memory;
2325 void tcp_close(struct sock *sk, long timeout)
2327 struct sk_buff *skb;
2328 int data_was_unread = 0;
2332 sk->sk_shutdown = SHUTDOWN_MASK;
2334 if (sk->sk_state == TCP_LISTEN) {
2335 tcp_set_state(sk, TCP_CLOSE);
2338 inet_csk_listen_stop(sk);
2340 goto adjudge_to_death;
2343 /* We need to flush the recv. buffs. We do this only on the
2344 * descriptor close, not protocol-sourced closes, because the
2345 * reader process may not have drained the data yet!
2347 while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) {
2348 u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq;
2350 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
2352 data_was_unread += len;
2358 /* If socket has been already reset (e.g. in tcp_reset()) - kill it. */
2359 if (sk->sk_state == TCP_CLOSE)
2360 goto adjudge_to_death;
2362 /* As outlined in RFC 2525, section 2.17, we send a RST here because
2363 * data was lost. To witness the awful effects of the old behavior of
2364 * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk
2365 * GET in an FTP client, suspend the process, wait for the client to
2366 * advertise a zero window, then kill -9 the FTP client, wheee...
2367 * Note: timeout is always zero in such a case.
2369 if (unlikely(tcp_sk(sk)->repair)) {
2370 sk->sk_prot->disconnect(sk, 0);
2371 } else if (data_was_unread) {
2372 /* Unread data was tossed, zap the connection. */
2373 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONCLOSE);
2374 tcp_set_state(sk, TCP_CLOSE);
2375 tcp_send_active_reset(sk, sk->sk_allocation);
2376 } else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) {
2377 /* Check zero linger _after_ checking for unread data. */
2378 sk->sk_prot->disconnect(sk, 0);
2379 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
2380 } else if (tcp_close_state(sk)) {
2381 /* We FIN if the application ate all the data before
2382 * zapping the connection.
2385 /* RED-PEN. Formally speaking, we have broken TCP state
2386 * machine. State transitions:
2388 * TCP_ESTABLISHED -> TCP_FIN_WAIT1
2389 * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible)
2390 * TCP_CLOSE_WAIT -> TCP_LAST_ACK
2392 * are legal only when FIN has been sent (i.e. in window),
2393 * rather than queued out of window. Purists blame.
2395 * F.e. "RFC state" is ESTABLISHED,
2396 * if Linux state is FIN-WAIT-1, but FIN is still not sent.
2398 * The visible declinations are that sometimes
2399 * we enter time-wait state, when it is not required really
2400 * (harmless), do not send active resets, when they are
2401 * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
2402 * they look as CLOSING or LAST_ACK for Linux)
2403 * Probably, I missed some more holelets.
2405 * XXX (TFO) - To start off we don't support SYN+ACK+FIN
2406 * in a single packet! (May consider it later but will
2407 * probably need API support or TCP_CORK SYN-ACK until
2408 * data is written and socket is closed.)
2413 sk_stream_wait_close(sk, timeout);
2416 state = sk->sk_state;
2422 /* remove backlog if any, without releasing ownership. */
2425 percpu_counter_inc(sk->sk_prot->orphan_count);
2427 /* Have we already been destroyed by a softirq or backlog? */
2428 if (state != TCP_CLOSE && sk->sk_state == TCP_CLOSE)
2431 /* This is a (useful) BSD violating of the RFC. There is a
2432 * problem with TCP as specified in that the other end could
2433 * keep a socket open forever with no application left this end.
2434 * We use a 1 minute timeout (about the same as BSD) then kill
2435 * our end. If they send after that then tough - BUT: long enough
2436 * that we won't make the old 4*rto = almost no time - whoops
2439 * Nope, it was not mistake. It is really desired behaviour
2440 * f.e. on http servers, when such sockets are useless, but
2441 * consume significant resources. Let's do it with special
2442 * linger2 option. --ANK
2445 if (sk->sk_state == TCP_FIN_WAIT2) {
2446 struct tcp_sock *tp = tcp_sk(sk);
2447 if (tp->linger2 < 0) {
2448 tcp_set_state(sk, TCP_CLOSE);
2449 tcp_send_active_reset(sk, GFP_ATOMIC);
2450 __NET_INC_STATS(sock_net(sk),
2451 LINUX_MIB_TCPABORTONLINGER);
2453 const int tmo = tcp_fin_time(sk);
2455 if (tmo > TCP_TIMEWAIT_LEN) {
2456 inet_csk_reset_keepalive_timer(sk,
2457 tmo - TCP_TIMEWAIT_LEN);
2459 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
2464 if (sk->sk_state != TCP_CLOSE) {
2466 if (tcp_check_oom(sk, 0)) {
2467 tcp_set_state(sk, TCP_CLOSE);
2468 tcp_send_active_reset(sk, GFP_ATOMIC);
2469 __NET_INC_STATS(sock_net(sk),
2470 LINUX_MIB_TCPABORTONMEMORY);
2471 } else if (!check_net(sock_net(sk))) {
2472 /* Not possible to send reset; just close */
2473 tcp_set_state(sk, TCP_CLOSE);
2477 if (sk->sk_state == TCP_CLOSE) {
2478 struct request_sock *req = tcp_sk(sk)->fastopen_rsk;
2479 /* We could get here with a non-NULL req if the socket is
2480 * aborted (e.g., closed with unread data) before 3WHS
2484 reqsk_fastopen_remove(sk, req, false);
2485 inet_csk_destroy_sock(sk);
2487 /* Otherwise, socket is reprieved until protocol close. */
2495 EXPORT_SYMBOL(tcp_close);
2497 /* These states need RST on ABORT according to RFC793 */
2499 static inline bool tcp_need_reset(int state)
2501 return (1 << state) &
2502 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 |
2503 TCPF_FIN_WAIT2 | TCPF_SYN_RECV);
2506 static void tcp_rtx_queue_purge(struct sock *sk)
2508 struct rb_node *p = rb_first(&sk->tcp_rtx_queue);
2511 struct sk_buff *skb = rb_to_skb(p);
2514 /* Since we are deleting whole queue, no need to
2515 * list_del(&skb->tcp_tsorted_anchor)
2517 tcp_rtx_queue_unlink(skb, sk);
2518 sk_wmem_free_skb(sk, skb);
2522 void tcp_write_queue_purge(struct sock *sk)
2524 struct sk_buff *skb;
2526 tcp_chrono_stop(sk, TCP_CHRONO_BUSY);
2527 while ((skb = __skb_dequeue(&sk->sk_write_queue)) != NULL) {
2528 tcp_skb_tsorted_anchor_cleanup(skb);
2529 sk_wmem_free_skb(sk, skb);
2531 tcp_rtx_queue_purge(sk);
2532 INIT_LIST_HEAD(&tcp_sk(sk)->tsorted_sent_queue);
2534 tcp_clear_all_retrans_hints(tcp_sk(sk));
2535 tcp_sk(sk)->packets_out = 0;
2536 inet_csk(sk)->icsk_backoff = 0;
2539 int tcp_disconnect(struct sock *sk, int flags)
2541 struct inet_sock *inet = inet_sk(sk);
2542 struct inet_connection_sock *icsk = inet_csk(sk);
2543 struct tcp_sock *tp = tcp_sk(sk);
2544 int old_state = sk->sk_state;
2546 if (old_state != TCP_CLOSE)
2547 tcp_set_state(sk, TCP_CLOSE);
2549 /* ABORT function of RFC793 */
2550 if (old_state == TCP_LISTEN) {
2551 inet_csk_listen_stop(sk);
2552 } else if (unlikely(tp->repair)) {
2553 sk->sk_err = ECONNABORTED;
2554 } else if (tcp_need_reset(old_state) ||
2555 (tp->snd_nxt != tp->write_seq &&
2556 (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK))) {
2557 /* The last check adjusts for discrepancy of Linux wrt. RFC
2560 tcp_send_active_reset(sk, gfp_any());
2561 sk->sk_err = ECONNRESET;
2562 } else if (old_state == TCP_SYN_SENT)
2563 sk->sk_err = ECONNRESET;
2565 tcp_clear_xmit_timers(sk);
2566 __skb_queue_purge(&sk->sk_receive_queue);
2567 tp->copied_seq = tp->rcv_nxt;
2569 tcp_write_queue_purge(sk);
2570 tcp_fastopen_active_disable_ofo_check(sk);
2571 skb_rbtree_purge(&tp->out_of_order_queue);
2573 inet->inet_dport = 0;
2575 if (!(sk->sk_userlocks & SOCK_BINDADDR_LOCK))
2576 inet_reset_saddr(sk);
2578 sk->sk_shutdown = 0;
2579 sock_reset_flag(sk, SOCK_DONE);
2581 tp->rcv_rtt_last_tsecr = 0;
2582 tp->write_seq += tp->max_window + 2;
2583 if (tp->write_seq == 0)
2586 icsk->icsk_probes_out = 0;
2587 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
2588 tp->snd_cwnd_cnt = 0;
2589 tp->window_clamp = 0;
2590 tp->delivered_ce = 0;
2591 tcp_set_ca_state(sk, TCP_CA_Open);
2592 tp->is_sack_reneg = 0;
2593 tcp_clear_retrans(tp);
2594 inet_csk_delack_init(sk);
2595 /* Initialize rcv_mss to TCP_MIN_MSS to avoid division by 0
2596 * issue in __tcp_select_window()
2598 icsk->icsk_ack.rcv_mss = TCP_MIN_MSS;
2599 memset(&tp->rx_opt, 0, sizeof(tp->rx_opt));
2601 dst_release(sk->sk_rx_dst);
2602 sk->sk_rx_dst = NULL;
2603 tcp_saved_syn_free(tp);
2604 tp->compressed_ack = 0;
2606 tp->bytes_acked = 0;
2607 tp->bytes_received = 0;
2608 tp->bytes_retrans = 0;
2612 /* Clean up fastopen related fields */
2613 tcp_free_fastopen_req(tp);
2614 inet->defer_connect = 0;
2616 WARN_ON(inet->inet_num && !icsk->icsk_bind_hash);
2618 if (sk->sk_frag.page) {
2619 put_page(sk->sk_frag.page);
2620 sk->sk_frag.page = NULL;
2621 sk->sk_frag.offset = 0;
2624 sk->sk_error_report(sk);
2627 EXPORT_SYMBOL(tcp_disconnect);
2629 static inline bool tcp_can_repair_sock(const struct sock *sk)
2631 return ns_capable(sock_net(sk)->user_ns, CAP_NET_ADMIN) &&
2632 (sk->sk_state != TCP_LISTEN);
2635 static int tcp_repair_set_window(struct tcp_sock *tp, char __user *optbuf, int len)
2637 struct tcp_repair_window opt;
2642 if (len != sizeof(opt))
2645 if (copy_from_user(&opt, optbuf, sizeof(opt)))
2648 if (opt.max_window < opt.snd_wnd)
2651 if (after(opt.snd_wl1, tp->rcv_nxt + opt.rcv_wnd))
2654 if (after(opt.rcv_wup, tp->rcv_nxt))
2657 tp->snd_wl1 = opt.snd_wl1;
2658 tp->snd_wnd = opt.snd_wnd;
2659 tp->max_window = opt.max_window;
2661 tp->rcv_wnd = opt.rcv_wnd;
2662 tp->rcv_wup = opt.rcv_wup;
2667 static int tcp_repair_options_est(struct sock *sk,
2668 struct tcp_repair_opt __user *optbuf, unsigned int len)
2670 struct tcp_sock *tp = tcp_sk(sk);
2671 struct tcp_repair_opt opt;
2673 while (len >= sizeof(opt)) {
2674 if (copy_from_user(&opt, optbuf, sizeof(opt)))
2680 switch (opt.opt_code) {
2682 tp->rx_opt.mss_clamp = opt.opt_val;
2687 u16 snd_wscale = opt.opt_val & 0xFFFF;
2688 u16 rcv_wscale = opt.opt_val >> 16;
2690 if (snd_wscale > TCP_MAX_WSCALE || rcv_wscale > TCP_MAX_WSCALE)
2693 tp->rx_opt.snd_wscale = snd_wscale;
2694 tp->rx_opt.rcv_wscale = rcv_wscale;
2695 tp->rx_opt.wscale_ok = 1;
2698 case TCPOPT_SACK_PERM:
2699 if (opt.opt_val != 0)
2702 tp->rx_opt.sack_ok |= TCP_SACK_SEEN;
2704 case TCPOPT_TIMESTAMP:
2705 if (opt.opt_val != 0)
2708 tp->rx_opt.tstamp_ok = 1;
2717 * Socket option code for TCP.
2719 static int do_tcp_setsockopt(struct sock *sk, int level,
2720 int optname, char __user *optval, unsigned int optlen)
2722 struct tcp_sock *tp = tcp_sk(sk);
2723 struct inet_connection_sock *icsk = inet_csk(sk);
2724 struct net *net = sock_net(sk);
2728 /* These are data/string values, all the others are ints */
2730 case TCP_CONGESTION: {
2731 char name[TCP_CA_NAME_MAX];
2736 val = strncpy_from_user(name, optval,
2737 min_t(long, TCP_CA_NAME_MAX-1, optlen));
2743 err = tcp_set_congestion_control(sk, name, true, true,
2744 ns_capable(sock_net(sk)->user_ns,
2750 char name[TCP_ULP_NAME_MAX];
2755 val = strncpy_from_user(name, optval,
2756 min_t(long, TCP_ULP_NAME_MAX - 1,
2763 err = tcp_set_ulp(sk, name);
2767 case TCP_FASTOPEN_KEY: {
2768 __u8 key[TCP_FASTOPEN_KEY_LENGTH];
2770 if (optlen != sizeof(key))
2773 if (copy_from_user(key, optval, optlen))
2776 return tcp_fastopen_reset_cipher(net, sk, key, sizeof(key));
2783 if (optlen < sizeof(int))
2786 if (get_user(val, (int __user *)optval))
2793 /* Values greater than interface MTU won't take effect. However
2794 * at the point when this call is done we typically don't yet
2795 * know which interface is going to be used
2797 if (val && (val < TCP_MIN_MSS || val > MAX_TCP_WINDOW)) {
2801 tp->rx_opt.user_mss = val;
2806 /* TCP_NODELAY is weaker than TCP_CORK, so that
2807 * this option on corked socket is remembered, but
2808 * it is not activated until cork is cleared.
2810 * However, when TCP_NODELAY is set we make
2811 * an explicit push, which overrides even TCP_CORK
2812 * for currently queued segments.
2814 tp->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH;
2815 tcp_push_pending_frames(sk);
2817 tp->nonagle &= ~TCP_NAGLE_OFF;
2821 case TCP_THIN_LINEAR_TIMEOUTS:
2822 if (val < 0 || val > 1)
2828 case TCP_THIN_DUPACK:
2829 if (val < 0 || val > 1)
2834 if (!tcp_can_repair_sock(sk))
2836 else if (val == TCP_REPAIR_ON) {
2838 sk->sk_reuse = SK_FORCE_REUSE;
2839 tp->repair_queue = TCP_NO_QUEUE;
2840 } else if (val == TCP_REPAIR_OFF) {
2842 sk->sk_reuse = SK_NO_REUSE;
2843 tcp_send_window_probe(sk);
2844 } else if (val == TCP_REPAIR_OFF_NO_WP) {
2846 sk->sk_reuse = SK_NO_REUSE;
2852 case TCP_REPAIR_QUEUE:
2855 else if ((unsigned int)val < TCP_QUEUES_NR)
2856 tp->repair_queue = val;
2862 if (sk->sk_state != TCP_CLOSE)
2864 else if (tp->repair_queue == TCP_SEND_QUEUE)
2865 tp->write_seq = val;
2866 else if (tp->repair_queue == TCP_RECV_QUEUE)
2867 WRITE_ONCE(tp->rcv_nxt, val);
2872 case TCP_REPAIR_OPTIONS:
2875 else if (sk->sk_state == TCP_ESTABLISHED)
2876 err = tcp_repair_options_est(sk,
2877 (struct tcp_repair_opt __user *)optval,
2884 /* When set indicates to always queue non-full frames.
2885 * Later the user clears this option and we transmit
2886 * any pending partial frames in the queue. This is
2887 * meant to be used alongside sendfile() to get properly
2888 * filled frames when the user (for example) must write
2889 * out headers with a write() call first and then use
2890 * sendfile to send out the data parts.
2892 * TCP_CORK can be set together with TCP_NODELAY and it is
2893 * stronger than TCP_NODELAY.
2896 tp->nonagle |= TCP_NAGLE_CORK;
2898 tp->nonagle &= ~TCP_NAGLE_CORK;
2899 if (tp->nonagle&TCP_NAGLE_OFF)
2900 tp->nonagle |= TCP_NAGLE_PUSH;
2901 tcp_push_pending_frames(sk);
2906 if (val < 1 || val > MAX_TCP_KEEPIDLE)
2909 tp->keepalive_time = val * HZ;
2910 if (sock_flag(sk, SOCK_KEEPOPEN) &&
2911 !((1 << sk->sk_state) &
2912 (TCPF_CLOSE | TCPF_LISTEN))) {
2913 u32 elapsed = keepalive_time_elapsed(tp);
2914 if (tp->keepalive_time > elapsed)
2915 elapsed = tp->keepalive_time - elapsed;
2918 inet_csk_reset_keepalive_timer(sk, elapsed);
2923 if (val < 1 || val > MAX_TCP_KEEPINTVL)
2926 tp->keepalive_intvl = val * HZ;
2929 if (val < 1 || val > MAX_TCP_KEEPCNT)
2932 tp->keepalive_probes = val;
2935 if (val < 1 || val > MAX_TCP_SYNCNT)
2938 icsk->icsk_syn_retries = val;
2942 if (val < 0 || val > 1)
2951 else if (val > net->ipv4.sysctl_tcp_fin_timeout / HZ)
2954 tp->linger2 = val * HZ;
2957 case TCP_DEFER_ACCEPT:
2958 /* Translate value in seconds to number of retransmits */
2959 icsk->icsk_accept_queue.rskq_defer_accept =
2960 secs_to_retrans(val, TCP_TIMEOUT_INIT / HZ,
2964 case TCP_WINDOW_CLAMP:
2966 if (sk->sk_state != TCP_CLOSE) {
2970 tp->window_clamp = 0;
2972 tp->window_clamp = val < SOCK_MIN_RCVBUF / 2 ?
2973 SOCK_MIN_RCVBUF / 2 : val;
2978 icsk->icsk_ack.pingpong = 1;
2980 icsk->icsk_ack.pingpong = 0;
2981 if ((1 << sk->sk_state) &
2982 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) &&
2983 inet_csk_ack_scheduled(sk)) {
2984 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
2985 tcp_cleanup_rbuf(sk, 1);
2987 icsk->icsk_ack.pingpong = 1;
2992 #ifdef CONFIG_TCP_MD5SIG
2994 case TCP_MD5SIG_EXT:
2995 if ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN))
2996 err = tp->af_specific->md5_parse(sk, optname, optval, optlen);
3001 case TCP_USER_TIMEOUT:
3002 /* Cap the max time in ms TCP will retry or probe the window
3003 * before giving up and aborting (ETIMEDOUT) a connection.
3008 icsk->icsk_user_timeout = val;
3012 if (val >= 0 && ((1 << sk->sk_state) & (TCPF_CLOSE |
3014 tcp_fastopen_init_key_once(net);
3016 fastopen_queue_tune(sk, val);
3021 case TCP_FASTOPEN_CONNECT:
3022 if (val > 1 || val < 0) {
3024 } else if (net->ipv4.sysctl_tcp_fastopen & TFO_CLIENT_ENABLE) {
3025 if (sk->sk_state == TCP_CLOSE)
3026 tp->fastopen_connect = val;
3033 case TCP_FASTOPEN_NO_COOKIE:
3034 if (val > 1 || val < 0)
3036 else if (!((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
3039 tp->fastopen_no_cookie = val;
3045 tp->tsoffset = val - tcp_time_stamp_raw();
3047 case TCP_REPAIR_WINDOW:
3048 err = tcp_repair_set_window(tp, optval, optlen);
3050 case TCP_NOTSENT_LOWAT:
3051 tp->notsent_lowat = val;
3052 sk->sk_write_space(sk);
3055 if (val > 1 || val < 0)
3058 tp->recvmsg_inq = val;
3069 int tcp_setsockopt(struct sock *sk, int level, int optname, char __user *optval,
3070 unsigned int optlen)
3072 const struct inet_connection_sock *icsk = inet_csk(sk);
3074 if (level != SOL_TCP)
3075 return icsk->icsk_af_ops->setsockopt(sk, level, optname,
3077 return do_tcp_setsockopt(sk, level, optname, optval, optlen);
3079 EXPORT_SYMBOL(tcp_setsockopt);
3081 #ifdef CONFIG_COMPAT
3082 int compat_tcp_setsockopt(struct sock *sk, int level, int optname,
3083 char __user *optval, unsigned int optlen)
3085 if (level != SOL_TCP)
3086 return inet_csk_compat_setsockopt(sk, level, optname,
3088 return do_tcp_setsockopt(sk, level, optname, optval, optlen);
3090 EXPORT_SYMBOL(compat_tcp_setsockopt);
3093 static void tcp_get_info_chrono_stats(const struct tcp_sock *tp,
3094 struct tcp_info *info)
3096 u64 stats[__TCP_CHRONO_MAX], total = 0;
3099 for (i = TCP_CHRONO_BUSY; i < __TCP_CHRONO_MAX; ++i) {
3100 stats[i] = tp->chrono_stat[i - 1];
3101 if (i == tp->chrono_type)
3102 stats[i] += tcp_jiffies32 - tp->chrono_start;
3103 stats[i] *= USEC_PER_SEC / HZ;
3107 info->tcpi_busy_time = total;
3108 info->tcpi_rwnd_limited = stats[TCP_CHRONO_RWND_LIMITED];
3109 info->tcpi_sndbuf_limited = stats[TCP_CHRONO_SNDBUF_LIMITED];
3112 /* Return information about state of tcp endpoint in API format. */
3113 void tcp_get_info(struct sock *sk, struct tcp_info *info)
3115 const struct tcp_sock *tp = tcp_sk(sk); /* iff sk_type == SOCK_STREAM */
3116 const struct inet_connection_sock *icsk = inet_csk(sk);
3122 memset(info, 0, sizeof(*info));
3123 if (sk->sk_type != SOCK_STREAM)
3126 info->tcpi_state = inet_sk_state_load(sk);
3128 /* Report meaningful fields for all TCP states, including listeners */
3129 rate = READ_ONCE(sk->sk_pacing_rate);
3130 rate64 = rate != ~0U ? rate : ~0ULL;
3131 info->tcpi_pacing_rate = rate64;
3133 rate = READ_ONCE(sk->sk_max_pacing_rate);
3134 rate64 = rate != ~0U ? rate : ~0ULL;
3135 info->tcpi_max_pacing_rate = rate64;
3137 info->tcpi_reordering = tp->reordering;
3138 info->tcpi_snd_cwnd = tp->snd_cwnd;
3140 if (info->tcpi_state == TCP_LISTEN) {
3141 /* listeners aliased fields :
3142 * tcpi_unacked -> Number of children ready for accept()
3143 * tcpi_sacked -> max backlog
3145 info->tcpi_unacked = sk->sk_ack_backlog;
3146 info->tcpi_sacked = sk->sk_max_ack_backlog;
3150 slow = lock_sock_fast(sk);
3152 info->tcpi_ca_state = icsk->icsk_ca_state;
3153 info->tcpi_retransmits = icsk->icsk_retransmits;
3154 info->tcpi_probes = icsk->icsk_probes_out;
3155 info->tcpi_backoff = icsk->icsk_backoff;
3157 if (tp->rx_opt.tstamp_ok)
3158 info->tcpi_options |= TCPI_OPT_TIMESTAMPS;
3159 if (tcp_is_sack(tp))
3160 info->tcpi_options |= TCPI_OPT_SACK;
3161 if (tp->rx_opt.wscale_ok) {
3162 info->tcpi_options |= TCPI_OPT_WSCALE;
3163 info->tcpi_snd_wscale = tp->rx_opt.snd_wscale;
3164 info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale;
3167 if (tp->ecn_flags & TCP_ECN_OK)
3168 info->tcpi_options |= TCPI_OPT_ECN;
3169 if (tp->ecn_flags & TCP_ECN_SEEN)
3170 info->tcpi_options |= TCPI_OPT_ECN_SEEN;
3171 if (tp->syn_data_acked)
3172 info->tcpi_options |= TCPI_OPT_SYN_DATA;
3174 info->tcpi_rto = jiffies_to_usecs(icsk->icsk_rto);
3175 info->tcpi_ato = jiffies_to_usecs(icsk->icsk_ack.ato);
3176 info->tcpi_snd_mss = tp->mss_cache;
3177 info->tcpi_rcv_mss = icsk->icsk_ack.rcv_mss;
3179 info->tcpi_unacked = tp->packets_out;
3180 info->tcpi_sacked = tp->sacked_out;
3182 info->tcpi_lost = tp->lost_out;
3183 info->tcpi_retrans = tp->retrans_out;
3185 now = tcp_jiffies32;
3186 info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime);
3187 info->tcpi_last_data_recv = jiffies_to_msecs(now - icsk->icsk_ack.lrcvtime);
3188 info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp);
3190 info->tcpi_pmtu = icsk->icsk_pmtu_cookie;
3191 info->tcpi_rcv_ssthresh = tp->rcv_ssthresh;
3192 info->tcpi_rtt = tp->srtt_us >> 3;
3193 info->tcpi_rttvar = tp->mdev_us >> 2;
3194 info->tcpi_snd_ssthresh = tp->snd_ssthresh;
3195 info->tcpi_advmss = tp->advmss;
3197 info->tcpi_rcv_rtt = tp->rcv_rtt_est.rtt_us >> 3;
3198 info->tcpi_rcv_space = tp->rcvq_space.space;
3200 info->tcpi_total_retrans = tp->total_retrans;
3202 info->tcpi_bytes_acked = tp->bytes_acked;
3203 info->tcpi_bytes_received = tp->bytes_received;
3204 info->tcpi_notsent_bytes = max_t(int, 0, tp->write_seq - tp->snd_nxt);
3205 tcp_get_info_chrono_stats(tp, info);
3207 info->tcpi_segs_out = tp->segs_out;
3208 info->tcpi_segs_in = tp->segs_in;
3210 info->tcpi_min_rtt = tcp_min_rtt(tp);
3211 info->tcpi_data_segs_in = tp->data_segs_in;
3212 info->tcpi_data_segs_out = tp->data_segs_out;
3214 info->tcpi_delivery_rate_app_limited = tp->rate_app_limited ? 1 : 0;
3215 rate64 = tcp_compute_delivery_rate(tp);
3217 info->tcpi_delivery_rate = rate64;
3218 info->tcpi_delivered = tp->delivered;
3219 info->tcpi_delivered_ce = tp->delivered_ce;
3220 info->tcpi_bytes_sent = tp->bytes_sent;
3221 info->tcpi_bytes_retrans = tp->bytes_retrans;
3222 info->tcpi_dsack_dups = tp->dsack_dups;
3223 info->tcpi_reord_seen = tp->reord_seen;
3224 unlock_sock_fast(sk, slow);
3226 EXPORT_SYMBOL_GPL(tcp_get_info);
3228 static size_t tcp_opt_stats_get_size(void)
3231 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_BUSY */
3232 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_RWND_LIMITED */
3233 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_SNDBUF_LIMITED */
3234 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_DATA_SEGS_OUT */
3235 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_TOTAL_RETRANS */
3236 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_PACING_RATE */
3237 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_DELIVERY_RATE */
3238 nla_total_size(sizeof(u32)) + /* TCP_NLA_SND_CWND */
3239 nla_total_size(sizeof(u32)) + /* TCP_NLA_REORDERING */
3240 nla_total_size(sizeof(u32)) + /* TCP_NLA_MIN_RTT */
3241 nla_total_size(sizeof(u8)) + /* TCP_NLA_RECUR_RETRANS */
3242 nla_total_size(sizeof(u8)) + /* TCP_NLA_DELIVERY_RATE_APP_LMT */
3243 nla_total_size(sizeof(u32)) + /* TCP_NLA_SNDQ_SIZE */
3244 nla_total_size(sizeof(u8)) + /* TCP_NLA_CA_STATE */
3245 nla_total_size(sizeof(u32)) + /* TCP_NLA_SND_SSTHRESH */
3246 nla_total_size(sizeof(u32)) + /* TCP_NLA_DELIVERED */
3247 nla_total_size(sizeof(u32)) + /* TCP_NLA_DELIVERED_CE */
3248 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_BYTES_SENT */
3249 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_BYTES_RETRANS */
3250 nla_total_size(sizeof(u32)) + /* TCP_NLA_DSACK_DUPS */
3251 nla_total_size(sizeof(u32)) + /* TCP_NLA_REORD_SEEN */
3255 struct sk_buff *tcp_get_timestamping_opt_stats(const struct sock *sk)
3257 const struct tcp_sock *tp = tcp_sk(sk);
3258 struct sk_buff *stats;
3259 struct tcp_info info;
3263 stats = alloc_skb(tcp_opt_stats_get_size(), GFP_ATOMIC);
3267 tcp_get_info_chrono_stats(tp, &info);
3268 nla_put_u64_64bit(stats, TCP_NLA_BUSY,
3269 info.tcpi_busy_time, TCP_NLA_PAD);
3270 nla_put_u64_64bit(stats, TCP_NLA_RWND_LIMITED,
3271 info.tcpi_rwnd_limited, TCP_NLA_PAD);
3272 nla_put_u64_64bit(stats, TCP_NLA_SNDBUF_LIMITED,
3273 info.tcpi_sndbuf_limited, TCP_NLA_PAD);
3274 nla_put_u64_64bit(stats, TCP_NLA_DATA_SEGS_OUT,
3275 tp->data_segs_out, TCP_NLA_PAD);
3276 nla_put_u64_64bit(stats, TCP_NLA_TOTAL_RETRANS,
3277 tp->total_retrans, TCP_NLA_PAD);
3279 rate = READ_ONCE(sk->sk_pacing_rate);
3280 rate64 = rate != ~0U ? rate : ~0ULL;
3281 nla_put_u64_64bit(stats, TCP_NLA_PACING_RATE, rate64, TCP_NLA_PAD);
3283 rate64 = tcp_compute_delivery_rate(tp);
3284 nla_put_u64_64bit(stats, TCP_NLA_DELIVERY_RATE, rate64, TCP_NLA_PAD);
3286 nla_put_u32(stats, TCP_NLA_SND_CWND, tp->snd_cwnd);
3287 nla_put_u32(stats, TCP_NLA_REORDERING, tp->reordering);
3288 nla_put_u32(stats, TCP_NLA_MIN_RTT, tcp_min_rtt(tp));
3290 nla_put_u8(stats, TCP_NLA_RECUR_RETRANS, inet_csk(sk)->icsk_retransmits);
3291 nla_put_u8(stats, TCP_NLA_DELIVERY_RATE_APP_LMT, !!tp->rate_app_limited);
3292 nla_put_u32(stats, TCP_NLA_SND_SSTHRESH, tp->snd_ssthresh);
3293 nla_put_u32(stats, TCP_NLA_DELIVERED, tp->delivered);
3294 nla_put_u32(stats, TCP_NLA_DELIVERED_CE, tp->delivered_ce);
3296 nla_put_u32(stats, TCP_NLA_SNDQ_SIZE, tp->write_seq - tp->snd_una);
3297 nla_put_u8(stats, TCP_NLA_CA_STATE, inet_csk(sk)->icsk_ca_state);
3299 nla_put_u64_64bit(stats, TCP_NLA_BYTES_SENT, tp->bytes_sent,
3301 nla_put_u64_64bit(stats, TCP_NLA_BYTES_RETRANS, tp->bytes_retrans,
3303 nla_put_u32(stats, TCP_NLA_DSACK_DUPS, tp->dsack_dups);
3304 nla_put_u32(stats, TCP_NLA_REORD_SEEN, tp->reord_seen);
3309 static int do_tcp_getsockopt(struct sock *sk, int level,
3310 int optname, char __user *optval, int __user *optlen)
3312 struct inet_connection_sock *icsk = inet_csk(sk);
3313 struct tcp_sock *tp = tcp_sk(sk);
3314 struct net *net = sock_net(sk);
3317 if (get_user(len, optlen))
3320 len = min_t(unsigned int, len, sizeof(int));
3327 val = tp->mss_cache;
3328 if (!val && ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
3329 val = tp->rx_opt.user_mss;
3331 val = tp->rx_opt.mss_clamp;
3334 val = !!(tp->nonagle&TCP_NAGLE_OFF);
3337 val = !!(tp->nonagle&TCP_NAGLE_CORK);
3340 val = keepalive_time_when(tp) / HZ;
3343 val = keepalive_intvl_when(tp) / HZ;
3346 val = keepalive_probes(tp);
3349 val = icsk->icsk_syn_retries ? : net->ipv4.sysctl_tcp_syn_retries;
3354 val = (val ? : net->ipv4.sysctl_tcp_fin_timeout) / HZ;
3356 case TCP_DEFER_ACCEPT:
3357 val = retrans_to_secs(icsk->icsk_accept_queue.rskq_defer_accept,
3358 TCP_TIMEOUT_INIT / HZ, TCP_RTO_MAX / HZ);
3360 case TCP_WINDOW_CLAMP:
3361 val = tp->window_clamp;
3364 struct tcp_info info;
3366 if (get_user(len, optlen))
3369 tcp_get_info(sk, &info);
3371 len = min_t(unsigned int, len, sizeof(info));
3372 if (put_user(len, optlen))
3374 if (copy_to_user(optval, &info, len))
3379 const struct tcp_congestion_ops *ca_ops;
3380 union tcp_cc_info info;
3384 if (get_user(len, optlen))
3387 ca_ops = icsk->icsk_ca_ops;
3388 if (ca_ops && ca_ops->get_info)
3389 sz = ca_ops->get_info(sk, ~0U, &attr, &info);
3391 len = min_t(unsigned int, len, sz);
3392 if (put_user(len, optlen))
3394 if (copy_to_user(optval, &info, len))
3399 val = !icsk->icsk_ack.pingpong;
3402 case TCP_CONGESTION:
3403 if (get_user(len, optlen))
3405 len = min_t(unsigned int, len, TCP_CA_NAME_MAX);
3406 if (put_user(len, optlen))
3408 if (copy_to_user(optval, icsk->icsk_ca_ops->name, len))
3413 if (get_user(len, optlen))
3415 len = min_t(unsigned int, len, TCP_ULP_NAME_MAX);
3416 if (!icsk->icsk_ulp_ops) {
3417 if (put_user(0, optlen))
3421 if (put_user(len, optlen))
3423 if (copy_to_user(optval, icsk->icsk_ulp_ops->name, len))
3427 case TCP_FASTOPEN_KEY: {
3428 __u8 key[TCP_FASTOPEN_KEY_LENGTH];
3429 struct tcp_fastopen_context *ctx;
3431 if (get_user(len, optlen))
3435 ctx = rcu_dereference(icsk->icsk_accept_queue.fastopenq.ctx);
3437 memcpy(key, ctx->key, sizeof(key));
3442 len = min_t(unsigned int, len, sizeof(key));
3443 if (put_user(len, optlen))
3445 if (copy_to_user(optval, key, len))
3449 case TCP_THIN_LINEAR_TIMEOUTS:
3453 case TCP_THIN_DUPACK:
3461 case TCP_REPAIR_QUEUE:
3463 val = tp->repair_queue;
3468 case TCP_REPAIR_WINDOW: {
3469 struct tcp_repair_window opt;
3471 if (get_user(len, optlen))
3474 if (len != sizeof(opt))
3480 opt.snd_wl1 = tp->snd_wl1;
3481 opt.snd_wnd = tp->snd_wnd;
3482 opt.max_window = tp->max_window;
3483 opt.rcv_wnd = tp->rcv_wnd;
3484 opt.rcv_wup = tp->rcv_wup;
3486 if (copy_to_user(optval, &opt, len))
3491 if (tp->repair_queue == TCP_SEND_QUEUE)
3492 val = tp->write_seq;
3493 else if (tp->repair_queue == TCP_RECV_QUEUE)
3499 case TCP_USER_TIMEOUT:
3500 val = icsk->icsk_user_timeout;
3504 val = icsk->icsk_accept_queue.fastopenq.max_qlen;
3507 case TCP_FASTOPEN_CONNECT:
3508 val = tp->fastopen_connect;
3511 case TCP_FASTOPEN_NO_COOKIE:
3512 val = tp->fastopen_no_cookie;
3516 val = tcp_time_stamp_raw() + tp->tsoffset;
3518 case TCP_NOTSENT_LOWAT:
3519 val = tp->notsent_lowat;
3522 val = tp->recvmsg_inq;
3527 case TCP_SAVED_SYN: {
3528 if (get_user(len, optlen))
3532 if (tp->saved_syn) {
3533 if (len < tp->saved_syn[0]) {
3534 if (put_user(tp->saved_syn[0], optlen)) {
3541 len = tp->saved_syn[0];
3542 if (put_user(len, optlen)) {
3546 if (copy_to_user(optval, tp->saved_syn + 1, len)) {
3550 tcp_saved_syn_free(tp);
3555 if (put_user(len, optlen))
3561 case TCP_ZEROCOPY_RECEIVE: {
3562 struct tcp_zerocopy_receive zc;
3565 if (get_user(len, optlen))
3567 if (len != sizeof(zc))
3569 if (copy_from_user(&zc, optval, len))
3572 err = tcp_zerocopy_receive(sk, &zc);
3574 if (!err && copy_to_user(optval, &zc, len))
3580 return -ENOPROTOOPT;
3583 if (put_user(len, optlen))
3585 if (copy_to_user(optval, &val, len))
3590 int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval,
3593 struct inet_connection_sock *icsk = inet_csk(sk);
3595 if (level != SOL_TCP)
3596 return icsk->icsk_af_ops->getsockopt(sk, level, optname,
3598 return do_tcp_getsockopt(sk, level, optname, optval, optlen);
3600 EXPORT_SYMBOL(tcp_getsockopt);
3602 #ifdef CONFIG_COMPAT
3603 int compat_tcp_getsockopt(struct sock *sk, int level, int optname,
3604 char __user *optval, int __user *optlen)
3606 if (level != SOL_TCP)
3607 return inet_csk_compat_getsockopt(sk, level, optname,
3609 return do_tcp_getsockopt(sk, level, optname, optval, optlen);
3611 EXPORT_SYMBOL(compat_tcp_getsockopt);
3614 #ifdef CONFIG_TCP_MD5SIG
3615 static DEFINE_PER_CPU(struct tcp_md5sig_pool, tcp_md5sig_pool);
3616 static DEFINE_MUTEX(tcp_md5sig_mutex);
3617 static bool tcp_md5sig_pool_populated = false;
3619 static void __tcp_alloc_md5sig_pool(void)
3621 struct crypto_ahash *hash;
3624 hash = crypto_alloc_ahash("md5", 0, CRYPTO_ALG_ASYNC);
3628 for_each_possible_cpu(cpu) {
3629 void *scratch = per_cpu(tcp_md5sig_pool, cpu).scratch;
3630 struct ahash_request *req;
3633 scratch = kmalloc_node(sizeof(union tcp_md5sum_block) +
3634 sizeof(struct tcphdr),
3639 per_cpu(tcp_md5sig_pool, cpu).scratch = scratch;
3641 if (per_cpu(tcp_md5sig_pool, cpu).md5_req)
3644 req = ahash_request_alloc(hash, GFP_KERNEL);
3648 ahash_request_set_callback(req, 0, NULL, NULL);
3650 per_cpu(tcp_md5sig_pool, cpu).md5_req = req;
3652 /* before setting tcp_md5sig_pool_populated, we must commit all writes
3653 * to memory. See smp_rmb() in tcp_get_md5sig_pool()
3656 tcp_md5sig_pool_populated = true;
3659 bool tcp_alloc_md5sig_pool(void)
3661 if (unlikely(!tcp_md5sig_pool_populated)) {
3662 mutex_lock(&tcp_md5sig_mutex);
3664 if (!tcp_md5sig_pool_populated)
3665 __tcp_alloc_md5sig_pool();
3667 mutex_unlock(&tcp_md5sig_mutex);
3669 return tcp_md5sig_pool_populated;
3671 EXPORT_SYMBOL(tcp_alloc_md5sig_pool);
3675 * tcp_get_md5sig_pool - get md5sig_pool for this user
3677 * We use percpu structure, so if we succeed, we exit with preemption
3678 * and BH disabled, to make sure another thread or softirq handling
3679 * wont try to get same context.
3681 struct tcp_md5sig_pool *tcp_get_md5sig_pool(void)
3685 if (tcp_md5sig_pool_populated) {
3686 /* coupled with smp_wmb() in __tcp_alloc_md5sig_pool() */
3688 return this_cpu_ptr(&tcp_md5sig_pool);
3693 EXPORT_SYMBOL(tcp_get_md5sig_pool);
3695 int tcp_md5_hash_skb_data(struct tcp_md5sig_pool *hp,
3696 const struct sk_buff *skb, unsigned int header_len)
3698 struct scatterlist sg;
3699 const struct tcphdr *tp = tcp_hdr(skb);
3700 struct ahash_request *req = hp->md5_req;
3702 const unsigned int head_data_len = skb_headlen(skb) > header_len ?
3703 skb_headlen(skb) - header_len : 0;
3704 const struct skb_shared_info *shi = skb_shinfo(skb);
3705 struct sk_buff *frag_iter;
3707 sg_init_table(&sg, 1);
3709 sg_set_buf(&sg, ((u8 *) tp) + header_len, head_data_len);
3710 ahash_request_set_crypt(req, &sg, NULL, head_data_len);
3711 if (crypto_ahash_update(req))
3714 for (i = 0; i < shi->nr_frags; ++i) {
3715 const struct skb_frag_struct *f = &shi->frags[i];
3716 unsigned int offset = f->page_offset;
3717 struct page *page = skb_frag_page(f) + (offset >> PAGE_SHIFT);
3719 sg_set_page(&sg, page, skb_frag_size(f),
3720 offset_in_page(offset));
3721 ahash_request_set_crypt(req, &sg, NULL, skb_frag_size(f));
3722 if (crypto_ahash_update(req))
3726 skb_walk_frags(skb, frag_iter)
3727 if (tcp_md5_hash_skb_data(hp, frag_iter, 0))
3732 EXPORT_SYMBOL(tcp_md5_hash_skb_data);
3734 int tcp_md5_hash_key(struct tcp_md5sig_pool *hp, const struct tcp_md5sig_key *key)
3736 struct scatterlist sg;
3738 sg_init_one(&sg, key->key, key->keylen);
3739 ahash_request_set_crypt(hp->md5_req, &sg, NULL, key->keylen);
3740 return crypto_ahash_update(hp->md5_req);
3742 EXPORT_SYMBOL(tcp_md5_hash_key);
3746 void tcp_done(struct sock *sk)
3748 struct request_sock *req = tcp_sk(sk)->fastopen_rsk;
3750 if (sk->sk_state == TCP_SYN_SENT || sk->sk_state == TCP_SYN_RECV)
3751 TCP_INC_STATS(sock_net(sk), TCP_MIB_ATTEMPTFAILS);
3753 tcp_set_state(sk, TCP_CLOSE);
3754 tcp_clear_xmit_timers(sk);
3756 reqsk_fastopen_remove(sk, req, false);
3758 sk->sk_shutdown = SHUTDOWN_MASK;
3760 if (!sock_flag(sk, SOCK_DEAD))
3761 sk->sk_state_change(sk);
3763 inet_csk_destroy_sock(sk);
3765 EXPORT_SYMBOL_GPL(tcp_done);
3767 int tcp_abort(struct sock *sk, int err)
3769 if (!sk_fullsock(sk)) {
3770 if (sk->sk_state == TCP_NEW_SYN_RECV) {
3771 struct request_sock *req = inet_reqsk(sk);
3774 inet_csk_reqsk_queue_drop(req->rsk_listener, req);
3781 /* Don't race with userspace socket closes such as tcp_close. */
3784 if (sk->sk_state == TCP_LISTEN) {
3785 tcp_set_state(sk, TCP_CLOSE);
3786 inet_csk_listen_stop(sk);
3789 /* Don't race with BH socket closes such as inet_csk_listen_stop. */
3793 if (!sock_flag(sk, SOCK_DEAD)) {
3795 /* This barrier is coupled with smp_rmb() in tcp_poll() */
3797 sk->sk_error_report(sk);
3798 if (tcp_need_reset(sk->sk_state))
3799 tcp_send_active_reset(sk, GFP_ATOMIC);
3805 tcp_write_queue_purge(sk);
3809 EXPORT_SYMBOL_GPL(tcp_abort);
3811 extern struct tcp_congestion_ops tcp_reno;
3813 static __initdata unsigned long thash_entries;
3814 static int __init set_thash_entries(char *str)
3821 ret = kstrtoul(str, 0, &thash_entries);
3827 __setup("thash_entries=", set_thash_entries);
3829 static void __init tcp_init_mem(void)
3831 unsigned long limit = nr_free_buffer_pages() / 16;
3833 limit = max(limit, 128UL);
3834 sysctl_tcp_mem[0] = limit / 4 * 3; /* 4.68 % */
3835 sysctl_tcp_mem[1] = limit; /* 6.25 % */
3836 sysctl_tcp_mem[2] = sysctl_tcp_mem[0] * 2; /* 9.37 % */
3839 void __init tcp_init(void)
3841 int max_rshare, max_wshare, cnt;
3842 unsigned long limit;
3845 BUILD_BUG_ON(TCP_MIN_SND_MSS <= MAX_TCP_OPTION_SPACE);
3846 BUILD_BUG_ON(sizeof(struct tcp_skb_cb) >
3847 FIELD_SIZEOF(struct sk_buff, cb));
3849 percpu_counter_init(&tcp_sockets_allocated, 0, GFP_KERNEL);
3850 percpu_counter_init(&tcp_orphan_count, 0, GFP_KERNEL);
3851 inet_hashinfo_init(&tcp_hashinfo);
3852 inet_hashinfo2_init(&tcp_hashinfo, "tcp_listen_portaddr_hash",
3853 thash_entries, 21, /* one slot per 2 MB*/
3855 tcp_hashinfo.bind_bucket_cachep =
3856 kmem_cache_create("tcp_bind_bucket",
3857 sizeof(struct inet_bind_bucket), 0,
3858 SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL);
3860 /* Size and allocate the main established and bind bucket
3863 * The methodology is similar to that of the buffer cache.
3865 tcp_hashinfo.ehash =
3866 alloc_large_system_hash("TCP established",
3867 sizeof(struct inet_ehash_bucket),
3869 17, /* one slot per 128 KB of memory */
3872 &tcp_hashinfo.ehash_mask,
3874 thash_entries ? 0 : 512 * 1024);
3875 for (i = 0; i <= tcp_hashinfo.ehash_mask; i++)
3876 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].chain, i);
3878 if (inet_ehash_locks_alloc(&tcp_hashinfo))
3879 panic("TCP: failed to alloc ehash_locks");
3880 tcp_hashinfo.bhash =
3881 alloc_large_system_hash("TCP bind",
3882 sizeof(struct inet_bind_hashbucket),
3883 tcp_hashinfo.ehash_mask + 1,
3884 17, /* one slot per 128 KB of memory */
3886 &tcp_hashinfo.bhash_size,
3890 tcp_hashinfo.bhash_size = 1U << tcp_hashinfo.bhash_size;
3891 for (i = 0; i < tcp_hashinfo.bhash_size; i++) {
3892 spin_lock_init(&tcp_hashinfo.bhash[i].lock);
3893 INIT_HLIST_HEAD(&tcp_hashinfo.bhash[i].chain);
3897 cnt = tcp_hashinfo.ehash_mask + 1;
3898 sysctl_tcp_max_orphans = cnt / 2;
3901 /* Set per-socket limits to no more than 1/128 the pressure threshold */
3902 limit = nr_free_buffer_pages() << (PAGE_SHIFT - 7);
3903 max_wshare = min(4UL*1024*1024, limit);
3904 max_rshare = min(6UL*1024*1024, limit);
3906 init_net.ipv4.sysctl_tcp_wmem[0] = SK_MEM_QUANTUM;
3907 init_net.ipv4.sysctl_tcp_wmem[1] = 16*1024;
3908 init_net.ipv4.sysctl_tcp_wmem[2] = max(64*1024, max_wshare);
3910 init_net.ipv4.sysctl_tcp_rmem[0] = SK_MEM_QUANTUM;
3911 init_net.ipv4.sysctl_tcp_rmem[1] = 131072;
3912 init_net.ipv4.sysctl_tcp_rmem[2] = max(131072, max_rshare);
3914 pr_info("Hash tables configured (established %u bind %u)\n",
3915 tcp_hashinfo.ehash_mask + 1, tcp_hashinfo.bhash_size);
3919 BUG_ON(tcp_register_congestion_control(&tcp_reno) != 0);