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>
273 #include <net/icmp.h>
274 #include <net/inet_common.h>
276 #include <net/xfrm.h>
278 #include <net/sock.h>
280 #include <asm/uaccess.h>
281 #include <asm/ioctls.h>
282 #include <asm/unaligned.h>
283 #include <net/busy_poll.h>
285 int sysctl_tcp_min_tso_segs __read_mostly = 2;
287 int sysctl_tcp_autocorking __read_mostly = 1;
289 struct percpu_counter tcp_orphan_count;
290 EXPORT_SYMBOL_GPL(tcp_orphan_count);
292 long sysctl_tcp_mem[3] __read_mostly;
293 int sysctl_tcp_wmem[3] __read_mostly;
294 int sysctl_tcp_rmem[3] __read_mostly;
296 EXPORT_SYMBOL(sysctl_tcp_mem);
297 EXPORT_SYMBOL(sysctl_tcp_rmem);
298 EXPORT_SYMBOL(sysctl_tcp_wmem);
300 atomic_long_t tcp_memory_allocated; /* Current allocated memory. */
301 EXPORT_SYMBOL(tcp_memory_allocated);
304 * Current number of TCP sockets.
306 struct percpu_counter tcp_sockets_allocated;
307 EXPORT_SYMBOL(tcp_sockets_allocated);
312 struct tcp_splice_state {
313 struct pipe_inode_info *pipe;
319 * Pressure flag: try to collapse.
320 * Technical note: it is used by multiple contexts non atomically.
321 * All the __sk_mem_schedule() is of this nature: accounting
322 * is strict, actions are advisory and have some latency.
324 int tcp_memory_pressure __read_mostly;
325 EXPORT_SYMBOL(tcp_memory_pressure);
327 void tcp_enter_memory_pressure(struct sock *sk)
329 if (!tcp_memory_pressure) {
330 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMEMORYPRESSURES);
331 tcp_memory_pressure = 1;
334 EXPORT_SYMBOL(tcp_enter_memory_pressure);
336 /* Convert seconds to retransmits based on initial and max timeout */
337 static u8 secs_to_retrans(int seconds, int timeout, int rto_max)
342 int period = timeout;
345 while (seconds > period && res < 255) {
348 if (timeout > rto_max)
356 /* Convert retransmits to seconds based on initial and max timeout */
357 static int retrans_to_secs(u8 retrans, int timeout, int rto_max)
365 if (timeout > rto_max)
373 /* Address-family independent initialization for a tcp_sock.
375 * NOTE: A lot of things set to zero explicitly by call to
376 * sk_alloc() so need not be done here.
378 void tcp_init_sock(struct sock *sk)
380 struct inet_connection_sock *icsk = inet_csk(sk);
381 struct tcp_sock *tp = tcp_sk(sk);
383 __skb_queue_head_init(&tp->out_of_order_queue);
384 tcp_init_xmit_timers(sk);
385 tcp_prequeue_init(tp);
386 INIT_LIST_HEAD(&tp->tsq_node);
388 icsk->icsk_rto = TCP_TIMEOUT_INIT;
389 tp->mdev_us = jiffies_to_usecs(TCP_TIMEOUT_INIT);
390 tp->rtt_min[0].rtt = ~0U;
392 /* So many TCP implementations out there (incorrectly) count the
393 * initial SYN frame in their delayed-ACK and congestion control
394 * algorithms that we must have the following bandaid to talk
395 * efficiently to them. -DaveM
397 tp->snd_cwnd = TCP_INIT_CWND;
399 /* See draft-stevens-tcpca-spec-01 for discussion of the
400 * initialization of these values.
402 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
403 tp->snd_cwnd_clamp = ~0;
404 tp->mss_cache = TCP_MSS_DEFAULT;
405 u64_stats_init(&tp->syncp);
407 tp->reordering = sock_net(sk)->ipv4.sysctl_tcp_reordering;
408 tcp_enable_early_retrans(tp);
409 tcp_assign_congestion_control(sk);
413 sk->sk_state = TCP_CLOSE;
415 sk->sk_write_space = sk_stream_write_space;
416 sock_set_flag(sk, SOCK_USE_WRITE_QUEUE);
418 icsk->icsk_sync_mss = tcp_sync_mss;
420 sk->sk_sndbuf = sysctl_tcp_wmem[1];
421 sk->sk_rcvbuf = sysctl_tcp_rmem[1];
424 if (mem_cgroup_sockets_enabled)
425 sock_update_memcg(sk);
426 sk_sockets_allocated_inc(sk);
429 EXPORT_SYMBOL(tcp_init_sock);
431 static void tcp_tx_timestamp(struct sock *sk, u16 tsflags, struct sk_buff *skb)
434 struct skb_shared_info *shinfo = skb_shinfo(skb);
435 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
437 sock_tx_timestamp(sk, tsflags, &shinfo->tx_flags);
438 if (tsflags & SOF_TIMESTAMPING_TX_ACK)
439 tcb->txstamp_ack = 1;
440 if (tsflags & SOF_TIMESTAMPING_TX_RECORD_MASK)
441 shinfo->tskey = TCP_SKB_CB(skb)->seq + skb->len - 1;
446 * Wait for a TCP event.
448 * Note that we don't need to lock the socket, as the upper poll layers
449 * take care of normal races (between the test and the event) and we don't
450 * go look at any of the socket buffers directly.
452 unsigned int tcp_poll(struct file *file, struct socket *sock, poll_table *wait)
455 struct sock *sk = sock->sk;
456 const struct tcp_sock *tp = tcp_sk(sk);
459 sock_rps_record_flow(sk);
461 sock_poll_wait(file, sk_sleep(sk), wait);
463 state = sk_state_load(sk);
464 if (state == TCP_LISTEN)
465 return inet_csk_listen_poll(sk);
467 /* Socket is not locked. We are protected from async events
468 * by poll logic and correct handling of state changes
469 * made by other threads is impossible in any case.
475 * POLLHUP is certainly not done right. But poll() doesn't
476 * have a notion of HUP in just one direction, and for a
477 * socket the read side is more interesting.
479 * Some poll() documentation says that POLLHUP is incompatible
480 * with the POLLOUT/POLLWR flags, so somebody should check this
481 * all. But careful, it tends to be safer to return too many
482 * bits than too few, and you can easily break real applications
483 * if you don't tell them that something has hung up!
487 * Check number 1. POLLHUP is _UNMASKABLE_ event (see UNIX98 and
488 * our fs/select.c). It means that after we received EOF,
489 * poll always returns immediately, making impossible poll() on write()
490 * in state CLOSE_WAIT. One solution is evident --- to set POLLHUP
491 * if and only if shutdown has been made in both directions.
492 * Actually, it is interesting to look how Solaris and DUX
493 * solve this dilemma. I would prefer, if POLLHUP were maskable,
494 * then we could set it on SND_SHUTDOWN. BTW examples given
495 * in Stevens' books assume exactly this behaviour, it explains
496 * why POLLHUP is incompatible with POLLOUT. --ANK
498 * NOTE. Check for TCP_CLOSE is added. The goal is to prevent
499 * blocking on fresh not-connected or disconnected socket. --ANK
501 if (sk->sk_shutdown == SHUTDOWN_MASK || state == TCP_CLOSE)
503 if (sk->sk_shutdown & RCV_SHUTDOWN)
504 mask |= POLLIN | POLLRDNORM | POLLRDHUP;
506 /* Connected or passive Fast Open socket? */
507 if (state != TCP_SYN_SENT &&
508 (state != TCP_SYN_RECV || tp->fastopen_rsk)) {
509 int target = sock_rcvlowat(sk, 0, INT_MAX);
511 if (tp->urg_seq == tp->copied_seq &&
512 !sock_flag(sk, SOCK_URGINLINE) &&
516 if (tp->rcv_nxt - tp->copied_seq >= target)
517 mask |= POLLIN | POLLRDNORM;
519 if (!(sk->sk_shutdown & SEND_SHUTDOWN)) {
520 if (sk_stream_is_writeable(sk)) {
521 mask |= POLLOUT | POLLWRNORM;
522 } else { /* send SIGIO later */
523 sk_set_bit(SOCKWQ_ASYNC_NOSPACE, sk);
524 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
526 /* Race breaker. If space is freed after
527 * wspace test but before the flags are set,
528 * IO signal will be lost. Memory barrier
529 * pairs with the input side.
531 smp_mb__after_atomic();
532 if (sk_stream_is_writeable(sk))
533 mask |= POLLOUT | POLLWRNORM;
536 mask |= POLLOUT | POLLWRNORM;
538 if (tp->urg_data & TCP_URG_VALID)
541 /* This barrier is coupled with smp_wmb() in tcp_reset() */
543 if (sk->sk_err || !skb_queue_empty(&sk->sk_error_queue))
548 EXPORT_SYMBOL(tcp_poll);
550 int tcp_ioctl(struct sock *sk, int cmd, unsigned long arg)
552 struct tcp_sock *tp = tcp_sk(sk);
558 if (sk->sk_state == TCP_LISTEN)
561 slow = lock_sock_fast(sk);
563 unlock_sock_fast(sk, slow);
566 answ = tp->urg_data && tp->urg_seq == tp->copied_seq;
569 if (sk->sk_state == TCP_LISTEN)
572 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
575 answ = tp->write_seq - tp->snd_una;
578 if (sk->sk_state == TCP_LISTEN)
581 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
584 answ = tp->write_seq - tp->snd_nxt;
590 return put_user(answ, (int __user *)arg);
592 EXPORT_SYMBOL(tcp_ioctl);
594 static inline void tcp_mark_push(struct tcp_sock *tp, struct sk_buff *skb)
596 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
597 tp->pushed_seq = tp->write_seq;
600 static inline bool forced_push(const struct tcp_sock *tp)
602 return after(tp->write_seq, tp->pushed_seq + (tp->max_window >> 1));
605 static void skb_entail(struct sock *sk, struct sk_buff *skb)
607 struct tcp_sock *tp = tcp_sk(sk);
608 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
611 tcb->seq = tcb->end_seq = tp->write_seq;
612 tcb->tcp_flags = TCPHDR_ACK;
614 __skb_header_release(skb);
615 tcp_add_write_queue_tail(sk, skb);
616 sk->sk_wmem_queued += skb->truesize;
617 sk_mem_charge(sk, skb->truesize);
618 if (tp->nonagle & TCP_NAGLE_PUSH)
619 tp->nonagle &= ~TCP_NAGLE_PUSH;
621 tcp_slow_start_after_idle_check(sk);
624 static inline void tcp_mark_urg(struct tcp_sock *tp, int flags)
627 tp->snd_up = tp->write_seq;
630 /* If a not yet filled skb is pushed, do not send it if
631 * we have data packets in Qdisc or NIC queues :
632 * Because TX completion will happen shortly, it gives a chance
633 * to coalesce future sendmsg() payload into this skb, without
634 * need for a timer, and with no latency trade off.
635 * As packets containing data payload have a bigger truesize
636 * than pure acks (dataless) packets, the last checks prevent
637 * autocorking if we only have an ACK in Qdisc/NIC queues,
638 * or if TX completion was delayed after we processed ACK packet.
640 static bool tcp_should_autocork(struct sock *sk, struct sk_buff *skb,
643 return skb->len < size_goal &&
644 sysctl_tcp_autocorking &&
645 skb != tcp_write_queue_head(sk) &&
646 atomic_read(&sk->sk_wmem_alloc) > skb->truesize;
649 static void tcp_push(struct sock *sk, int flags, int mss_now,
650 int nonagle, int size_goal)
652 struct tcp_sock *tp = tcp_sk(sk);
655 if (!tcp_send_head(sk))
658 skb = tcp_write_queue_tail(sk);
659 if (!(flags & MSG_MORE) || forced_push(tp))
660 tcp_mark_push(tp, skb);
662 tcp_mark_urg(tp, flags);
664 if (tcp_should_autocork(sk, skb, size_goal)) {
666 /* avoid atomic op if TSQ_THROTTLED bit is already set */
667 if (!test_bit(TSQ_THROTTLED, &tp->tsq_flags)) {
668 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPAUTOCORKING);
669 set_bit(TSQ_THROTTLED, &tp->tsq_flags);
671 /* It is possible TX completion already happened
672 * before we set TSQ_THROTTLED.
674 if (atomic_read(&sk->sk_wmem_alloc) > skb->truesize)
678 if (flags & MSG_MORE)
679 nonagle = TCP_NAGLE_CORK;
681 __tcp_push_pending_frames(sk, mss_now, nonagle);
684 static int tcp_splice_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb,
685 unsigned int offset, size_t len)
687 struct tcp_splice_state *tss = rd_desc->arg.data;
690 ret = skb_splice_bits(skb, skb->sk, offset, tss->pipe,
691 min(rd_desc->count, len), tss->flags);
693 rd_desc->count -= ret;
697 static int __tcp_splice_read(struct sock *sk, struct tcp_splice_state *tss)
699 /* Store TCP splice context information in read_descriptor_t. */
700 read_descriptor_t rd_desc = {
705 return tcp_read_sock(sk, &rd_desc, tcp_splice_data_recv);
709 * tcp_splice_read - splice data from TCP socket to a pipe
710 * @sock: socket to splice from
711 * @ppos: position (not valid)
712 * @pipe: pipe to splice to
713 * @len: number of bytes to splice
714 * @flags: splice modifier flags
717 * Will read pages from given socket and fill them into a pipe.
720 ssize_t tcp_splice_read(struct socket *sock, loff_t *ppos,
721 struct pipe_inode_info *pipe, size_t len,
724 struct sock *sk = sock->sk;
725 struct tcp_splice_state tss = {
734 sock_rps_record_flow(sk);
736 * We can't seek on a socket input
745 timeo = sock_rcvtimeo(sk, sock->file->f_flags & O_NONBLOCK);
747 ret = __tcp_splice_read(sk, &tss);
753 if (sock_flag(sk, SOCK_DONE))
756 ret = sock_error(sk);
759 if (sk->sk_shutdown & RCV_SHUTDOWN)
761 if (sk->sk_state == TCP_CLOSE) {
763 * This occurs when user tries to read
764 * from never connected socket.
766 if (!sock_flag(sk, SOCK_DONE))
774 sk_wait_data(sk, &timeo, NULL);
775 if (signal_pending(current)) {
776 ret = sock_intr_errno(timeo);
789 if (sk->sk_err || sk->sk_state == TCP_CLOSE ||
790 (sk->sk_shutdown & RCV_SHUTDOWN) ||
791 signal_pending(current))
802 EXPORT_SYMBOL(tcp_splice_read);
804 struct sk_buff *sk_stream_alloc_skb(struct sock *sk, int size, gfp_t gfp,
809 /* The TCP header must be at least 32-bit aligned. */
810 size = ALIGN(size, 4);
812 if (unlikely(tcp_under_memory_pressure(sk)))
813 sk_mem_reclaim_partial(sk);
815 skb = alloc_skb_fclone(size + sk->sk_prot->max_header, gfp);
819 if (force_schedule) {
820 mem_scheduled = true;
821 sk_forced_mem_schedule(sk, skb->truesize);
823 mem_scheduled = sk_wmem_schedule(sk, skb->truesize);
825 if (likely(mem_scheduled)) {
826 skb_reserve(skb, sk->sk_prot->max_header);
828 * Make sure that we have exactly size bytes
829 * available to the caller, no more, no less.
831 skb->reserved_tailroom = skb->end - skb->tail - size;
836 sk->sk_prot->enter_memory_pressure(sk);
837 sk_stream_moderate_sndbuf(sk);
842 static unsigned int tcp_xmit_size_goal(struct sock *sk, u32 mss_now,
845 struct tcp_sock *tp = tcp_sk(sk);
846 u32 new_size_goal, size_goal;
848 if (!large_allowed || !sk_can_gso(sk))
851 /* Note : tcp_tso_autosize() will eventually split this later */
852 new_size_goal = sk->sk_gso_max_size - 1 - MAX_TCP_HEADER;
853 new_size_goal = tcp_bound_to_half_wnd(tp, new_size_goal);
855 /* We try hard to avoid divides here */
856 size_goal = tp->gso_segs * mss_now;
857 if (unlikely(new_size_goal < size_goal ||
858 new_size_goal >= size_goal + mss_now)) {
859 tp->gso_segs = min_t(u16, new_size_goal / mss_now,
860 sk->sk_gso_max_segs);
861 size_goal = tp->gso_segs * mss_now;
864 return max(size_goal, mss_now);
867 static int tcp_send_mss(struct sock *sk, int *size_goal, int flags)
871 mss_now = tcp_current_mss(sk);
872 *size_goal = tcp_xmit_size_goal(sk, mss_now, !(flags & MSG_OOB));
877 static ssize_t do_tcp_sendpages(struct sock *sk, struct page *page, int offset,
878 size_t size, int flags)
880 struct tcp_sock *tp = tcp_sk(sk);
881 int mss_now, size_goal;
884 long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
886 /* Wait for a connection to finish. One exception is TCP Fast Open
887 * (passive side) where data is allowed to be sent before a connection
888 * is fully established.
890 if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) &&
891 !tcp_passive_fastopen(sk)) {
892 err = sk_stream_wait_connect(sk, &timeo);
897 sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
899 mss_now = tcp_send_mss(sk, &size_goal, flags);
903 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
907 struct sk_buff *skb = tcp_write_queue_tail(sk);
911 if (!tcp_send_head(sk) || (copy = size_goal - skb->len) <= 0 ||
912 !tcp_skb_can_collapse_to(skb)) {
914 if (!sk_stream_memory_free(sk))
915 goto wait_for_sndbuf;
917 skb = sk_stream_alloc_skb(sk, 0, sk->sk_allocation,
918 skb_queue_empty(&sk->sk_write_queue));
920 goto wait_for_memory;
929 i = skb_shinfo(skb)->nr_frags;
930 can_coalesce = skb_can_coalesce(skb, i, page, offset);
931 if (!can_coalesce && i >= sysctl_max_skb_frags) {
932 tcp_mark_push(tp, skb);
935 if (!sk_wmem_schedule(sk, copy))
936 goto wait_for_memory;
939 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
942 skb_fill_page_desc(skb, i, page, offset, copy);
944 skb_shinfo(skb)->tx_flags |= SKBTX_SHARED_FRAG;
947 skb->data_len += copy;
948 skb->truesize += copy;
949 sk->sk_wmem_queued += copy;
950 sk_mem_charge(sk, copy);
951 skb->ip_summed = CHECKSUM_PARTIAL;
952 tp->write_seq += copy;
953 TCP_SKB_CB(skb)->end_seq += copy;
954 tcp_skb_pcount_set(skb, 0);
957 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
963 tcp_tx_timestamp(sk, sk->sk_tsflags, skb);
967 if (skb->len < size_goal || (flags & MSG_OOB))
970 if (forced_push(tp)) {
971 tcp_mark_push(tp, skb);
972 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
973 } else if (skb == tcp_send_head(sk))
974 tcp_push_one(sk, mss_now);
978 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
980 tcp_push(sk, flags & ~MSG_MORE, mss_now,
981 TCP_NAGLE_PUSH, size_goal);
983 err = sk_stream_wait_memory(sk, &timeo);
987 mss_now = tcp_send_mss(sk, &size_goal, flags);
991 if (copied && !(flags & MSG_SENDPAGE_NOTLAST))
992 tcp_push(sk, flags, mss_now, tp->nonagle, size_goal);
999 /* make sure we wake any epoll edge trigger waiter */
1000 if (unlikely(skb_queue_len(&sk->sk_write_queue) == 0 && err == -EAGAIN))
1001 sk->sk_write_space(sk);
1002 return sk_stream_error(sk, flags, err);
1005 int tcp_sendpage(struct sock *sk, struct page *page, int offset,
1006 size_t size, int flags)
1010 if (!(sk->sk_route_caps & NETIF_F_SG) ||
1011 !sk_check_csum_caps(sk))
1012 return sock_no_sendpage(sk->sk_socket, page, offset, size,
1016 res = do_tcp_sendpages(sk, page, offset, size, flags);
1020 EXPORT_SYMBOL(tcp_sendpage);
1022 static inline int select_size(const struct sock *sk, bool sg)
1024 const struct tcp_sock *tp = tcp_sk(sk);
1025 int tmp = tp->mss_cache;
1028 if (sk_can_gso(sk)) {
1029 /* Small frames wont use a full page:
1030 * Payload will immediately follow tcp header.
1032 tmp = SKB_WITH_OVERHEAD(2048 - MAX_TCP_HEADER);
1034 int pgbreak = SKB_MAX_HEAD(MAX_TCP_HEADER);
1036 if (tmp >= pgbreak &&
1037 tmp <= pgbreak + (MAX_SKB_FRAGS - 1) * PAGE_SIZE)
1045 void tcp_free_fastopen_req(struct tcp_sock *tp)
1047 if (tp->fastopen_req) {
1048 kfree(tp->fastopen_req);
1049 tp->fastopen_req = NULL;
1053 static int tcp_sendmsg_fastopen(struct sock *sk, struct msghdr *msg,
1054 int *copied, size_t size)
1056 struct tcp_sock *tp = tcp_sk(sk);
1059 if (!(sysctl_tcp_fastopen & TFO_CLIENT_ENABLE))
1061 if (tp->fastopen_req)
1062 return -EALREADY; /* Another Fast Open is in progress */
1064 tp->fastopen_req = kzalloc(sizeof(struct tcp_fastopen_request),
1066 if (unlikely(!tp->fastopen_req))
1068 tp->fastopen_req->data = msg;
1069 tp->fastopen_req->size = size;
1071 flags = (msg->msg_flags & MSG_DONTWAIT) ? O_NONBLOCK : 0;
1072 err = __inet_stream_connect(sk->sk_socket, msg->msg_name,
1073 msg->msg_namelen, flags);
1074 *copied = tp->fastopen_req->copied;
1075 tcp_free_fastopen_req(tp);
1079 int tcp_sendmsg(struct sock *sk, struct msghdr *msg, size_t size)
1081 struct tcp_sock *tp = tcp_sk(sk);
1082 struct sk_buff *skb;
1083 struct sockcm_cookie sockc;
1084 int flags, err, copied = 0;
1085 int mss_now = 0, size_goal, copied_syn = 0;
1086 bool process_backlog = false;
1092 flags = msg->msg_flags;
1093 if (flags & MSG_FASTOPEN) {
1094 err = tcp_sendmsg_fastopen(sk, msg, &copied_syn, size);
1095 if (err == -EINPROGRESS && copied_syn > 0)
1101 timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
1103 /* Wait for a connection to finish. One exception is TCP Fast Open
1104 * (passive side) where data is allowed to be sent before a connection
1105 * is fully established.
1107 if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) &&
1108 !tcp_passive_fastopen(sk)) {
1109 err = sk_stream_wait_connect(sk, &timeo);
1114 if (unlikely(tp->repair)) {
1115 if (tp->repair_queue == TCP_RECV_QUEUE) {
1116 copied = tcp_send_rcvq(sk, msg, size);
1121 if (tp->repair_queue == TCP_NO_QUEUE)
1124 /* 'common' sending to sendq */
1127 sockc.tsflags = sk->sk_tsflags;
1128 if (msg->msg_controllen) {
1129 err = sock_cmsg_send(sk, msg, &sockc);
1130 if (unlikely(err)) {
1136 /* This should be in poll */
1137 sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
1139 /* Ok commence sending. */
1143 mss_now = tcp_send_mss(sk, &size_goal, flags);
1146 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
1149 sg = !!(sk->sk_route_caps & NETIF_F_SG);
1151 while (msg_data_left(msg)) {
1153 int max = size_goal;
1155 skb = tcp_write_queue_tail(sk);
1156 if (tcp_send_head(sk)) {
1157 if (skb->ip_summed == CHECKSUM_NONE)
1159 copy = max - skb->len;
1162 if (copy <= 0 || !tcp_skb_can_collapse_to(skb)) {
1164 /* Allocate new segment. If the interface is SG,
1165 * allocate skb fitting to single page.
1167 if (!sk_stream_memory_free(sk))
1168 goto wait_for_sndbuf;
1170 if (process_backlog && sk_flush_backlog(sk)) {
1171 process_backlog = false;
1174 skb = sk_stream_alloc_skb(sk,
1175 select_size(sk, sg),
1177 skb_queue_empty(&sk->sk_write_queue));
1179 goto wait_for_memory;
1181 process_backlog = true;
1183 * Check whether we can use HW checksum.
1185 if (sk_check_csum_caps(sk))
1186 skb->ip_summed = CHECKSUM_PARTIAL;
1188 skb_entail(sk, skb);
1192 /* All packets are restored as if they have
1193 * already been sent. skb_mstamp isn't set to
1194 * avoid wrong rtt estimation.
1197 TCP_SKB_CB(skb)->sacked |= TCPCB_REPAIRED;
1200 /* Try to append data to the end of skb. */
1201 if (copy > msg_data_left(msg))
1202 copy = msg_data_left(msg);
1204 /* Where to copy to? */
1205 if (skb_availroom(skb) > 0) {
1206 /* We have some space in skb head. Superb! */
1207 copy = min_t(int, copy, skb_availroom(skb));
1208 err = skb_add_data_nocache(sk, skb, &msg->msg_iter, copy);
1213 int i = skb_shinfo(skb)->nr_frags;
1214 struct page_frag *pfrag = sk_page_frag(sk);
1216 if (!sk_page_frag_refill(sk, pfrag))
1217 goto wait_for_memory;
1219 if (!skb_can_coalesce(skb, i, pfrag->page,
1221 if (i == sysctl_max_skb_frags || !sg) {
1222 tcp_mark_push(tp, skb);
1228 copy = min_t(int, copy, pfrag->size - pfrag->offset);
1230 if (!sk_wmem_schedule(sk, copy))
1231 goto wait_for_memory;
1233 err = skb_copy_to_page_nocache(sk, &msg->msg_iter, skb,
1240 /* Update the skb. */
1242 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1244 skb_fill_page_desc(skb, i, pfrag->page,
1245 pfrag->offset, copy);
1246 get_page(pfrag->page);
1248 pfrag->offset += copy;
1252 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
1254 tp->write_seq += copy;
1255 TCP_SKB_CB(skb)->end_seq += copy;
1256 tcp_skb_pcount_set(skb, 0);
1259 if (!msg_data_left(msg)) {
1260 tcp_tx_timestamp(sk, sockc.tsflags, skb);
1261 if (unlikely(flags & MSG_EOR))
1262 TCP_SKB_CB(skb)->eor = 1;
1266 if (skb->len < max || (flags & MSG_OOB) || unlikely(tp->repair))
1269 if (forced_push(tp)) {
1270 tcp_mark_push(tp, skb);
1271 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
1272 } else if (skb == tcp_send_head(sk))
1273 tcp_push_one(sk, mss_now);
1277 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1280 tcp_push(sk, flags & ~MSG_MORE, mss_now,
1281 TCP_NAGLE_PUSH, size_goal);
1283 err = sk_stream_wait_memory(sk, &timeo);
1287 mss_now = tcp_send_mss(sk, &size_goal, flags);
1292 tcp_push(sk, flags, mss_now, tp->nonagle, size_goal);
1295 return copied + copied_syn;
1299 tcp_unlink_write_queue(skb, sk);
1300 /* It is the one place in all of TCP, except connection
1301 * reset, where we can be unlinking the send_head.
1303 tcp_check_send_head(sk, skb);
1304 sk_wmem_free_skb(sk, skb);
1308 if (copied + copied_syn)
1311 err = sk_stream_error(sk, flags, err);
1312 /* make sure we wake any epoll edge trigger waiter */
1313 if (unlikely(skb_queue_len(&sk->sk_write_queue) == 0 && err == -EAGAIN))
1314 sk->sk_write_space(sk);
1318 EXPORT_SYMBOL(tcp_sendmsg);
1321 * Handle reading urgent data. BSD has very simple semantics for
1322 * this, no blocking and very strange errors 8)
1325 static int tcp_recv_urg(struct sock *sk, struct msghdr *msg, int len, int flags)
1327 struct tcp_sock *tp = tcp_sk(sk);
1329 /* No URG data to read. */
1330 if (sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data ||
1331 tp->urg_data == TCP_URG_READ)
1332 return -EINVAL; /* Yes this is right ! */
1334 if (sk->sk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DONE))
1337 if (tp->urg_data & TCP_URG_VALID) {
1339 char c = tp->urg_data;
1341 if (!(flags & MSG_PEEK))
1342 tp->urg_data = TCP_URG_READ;
1344 /* Read urgent data. */
1345 msg->msg_flags |= MSG_OOB;
1348 if (!(flags & MSG_TRUNC))
1349 err = memcpy_to_msg(msg, &c, 1);
1352 msg->msg_flags |= MSG_TRUNC;
1354 return err ? -EFAULT : len;
1357 if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
1360 /* Fixed the recv(..., MSG_OOB) behaviour. BSD docs and
1361 * the available implementations agree in this case:
1362 * this call should never block, independent of the
1363 * blocking state of the socket.
1364 * Mike <pall@rz.uni-karlsruhe.de>
1369 static int tcp_peek_sndq(struct sock *sk, struct msghdr *msg, int len)
1371 struct sk_buff *skb;
1372 int copied = 0, err = 0;
1374 /* XXX -- need to support SO_PEEK_OFF */
1376 skb_queue_walk(&sk->sk_write_queue, skb) {
1377 err = skb_copy_datagram_msg(skb, 0, msg, skb->len);
1384 return err ?: copied;
1387 /* Clean up the receive buffer for full frames taken by the user,
1388 * then send an ACK if necessary. COPIED is the number of bytes
1389 * tcp_recvmsg has given to the user so far, it speeds up the
1390 * calculation of whether or not we must ACK for the sake of
1393 static void tcp_cleanup_rbuf(struct sock *sk, int copied)
1395 struct tcp_sock *tp = tcp_sk(sk);
1396 bool time_to_ack = false;
1398 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
1400 WARN(skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq),
1401 "cleanup rbuf bug: copied %X seq %X rcvnxt %X\n",
1402 tp->copied_seq, TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt);
1404 if (inet_csk_ack_scheduled(sk)) {
1405 const struct inet_connection_sock *icsk = inet_csk(sk);
1406 /* Delayed ACKs frequently hit locked sockets during bulk
1408 if (icsk->icsk_ack.blocked ||
1409 /* Once-per-two-segments ACK was not sent by tcp_input.c */
1410 tp->rcv_nxt - tp->rcv_wup > icsk->icsk_ack.rcv_mss ||
1412 * If this read emptied read buffer, we send ACK, if
1413 * connection is not bidirectional, user drained
1414 * receive buffer and there was a small segment
1418 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED2) ||
1419 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED) &&
1420 !icsk->icsk_ack.pingpong)) &&
1421 !atomic_read(&sk->sk_rmem_alloc)))
1425 /* We send an ACK if we can now advertise a non-zero window
1426 * which has been raised "significantly".
1428 * Even if window raised up to infinity, do not send window open ACK
1429 * in states, where we will not receive more. It is useless.
1431 if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) {
1432 __u32 rcv_window_now = tcp_receive_window(tp);
1434 /* Optimize, __tcp_select_window() is not cheap. */
1435 if (2*rcv_window_now <= tp->window_clamp) {
1436 __u32 new_window = __tcp_select_window(sk);
1438 /* Send ACK now, if this read freed lots of space
1439 * in our buffer. Certainly, new_window is new window.
1440 * We can advertise it now, if it is not less than current one.
1441 * "Lots" means "at least twice" here.
1443 if (new_window && new_window >= 2 * rcv_window_now)
1451 static void tcp_prequeue_process(struct sock *sk)
1453 struct sk_buff *skb;
1454 struct tcp_sock *tp = tcp_sk(sk);
1456 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPPREQUEUED);
1458 while ((skb = __skb_dequeue(&tp->ucopy.prequeue)) != NULL)
1459 sk_backlog_rcv(sk, skb);
1461 /* Clear memory counter. */
1462 tp->ucopy.memory = 0;
1465 static struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off)
1467 struct sk_buff *skb;
1470 while ((skb = skb_peek(&sk->sk_receive_queue)) != NULL) {
1471 offset = seq - TCP_SKB_CB(skb)->seq;
1472 if (unlikely(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
1473 pr_err_once("%s: found a SYN, please report !\n", __func__);
1476 if (offset < skb->len || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)) {
1480 /* This looks weird, but this can happen if TCP collapsing
1481 * splitted a fat GRO packet, while we released socket lock
1482 * in skb_splice_bits()
1484 sk_eat_skb(sk, skb);
1490 * This routine provides an alternative to tcp_recvmsg() for routines
1491 * that would like to handle copying from skbuffs directly in 'sendfile'
1494 * - It is assumed that the socket was locked by the caller.
1495 * - The routine does not block.
1496 * - At present, there is no support for reading OOB data
1497 * or for 'peeking' the socket using this routine
1498 * (although both would be easy to implement).
1500 int tcp_read_sock(struct sock *sk, read_descriptor_t *desc,
1501 sk_read_actor_t recv_actor)
1503 struct sk_buff *skb;
1504 struct tcp_sock *tp = tcp_sk(sk);
1505 u32 seq = tp->copied_seq;
1509 if (sk->sk_state == TCP_LISTEN)
1511 while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1512 if (offset < skb->len) {
1516 len = skb->len - offset;
1517 /* Stop reading if we hit a patch of urgent data */
1519 u32 urg_offset = tp->urg_seq - seq;
1520 if (urg_offset < len)
1525 used = recv_actor(desc, skb, offset, len);
1530 } else if (used <= len) {
1535 /* If recv_actor drops the lock (e.g. TCP splice
1536 * receive) the skb pointer might be invalid when
1537 * getting here: tcp_collapse might have deleted it
1538 * while aggregating skbs from the socket queue.
1540 skb = tcp_recv_skb(sk, seq - 1, &offset);
1543 /* TCP coalescing might have appended data to the skb.
1544 * Try to splice more frags
1546 if (offset + 1 != skb->len)
1549 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) {
1550 sk_eat_skb(sk, skb);
1554 sk_eat_skb(sk, skb);
1557 tp->copied_seq = seq;
1559 tp->copied_seq = seq;
1561 tcp_rcv_space_adjust(sk);
1563 /* Clean up data we have read: This will do ACK frames. */
1565 tcp_recv_skb(sk, seq, &offset);
1566 tcp_cleanup_rbuf(sk, copied);
1570 EXPORT_SYMBOL(tcp_read_sock);
1573 * This routine copies from a sock struct into the user buffer.
1575 * Technical note: in 2.3 we work on _locked_ socket, so that
1576 * tricks with *seq access order and skb->users are not required.
1577 * Probably, code can be easily improved even more.
1580 int tcp_recvmsg(struct sock *sk, struct msghdr *msg, size_t len, int nonblock,
1581 int flags, int *addr_len)
1583 struct tcp_sock *tp = tcp_sk(sk);
1589 int target; /* Read at least this many bytes */
1591 struct task_struct *user_recv = NULL;
1592 struct sk_buff *skb, *last;
1595 if (unlikely(flags & MSG_ERRQUEUE))
1596 return inet_recv_error(sk, msg, len, addr_len);
1598 if (sk_can_busy_loop(sk) && skb_queue_empty(&sk->sk_receive_queue) &&
1599 (sk->sk_state == TCP_ESTABLISHED))
1600 sk_busy_loop(sk, nonblock);
1605 if (sk->sk_state == TCP_LISTEN)
1608 timeo = sock_rcvtimeo(sk, nonblock);
1610 /* Urgent data needs to be handled specially. */
1611 if (flags & MSG_OOB)
1614 if (unlikely(tp->repair)) {
1616 if (!(flags & MSG_PEEK))
1619 if (tp->repair_queue == TCP_SEND_QUEUE)
1623 if (tp->repair_queue == TCP_NO_QUEUE)
1626 /* 'common' recv queue MSG_PEEK-ing */
1629 seq = &tp->copied_seq;
1630 if (flags & MSG_PEEK) {
1631 peek_seq = tp->copied_seq;
1635 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
1640 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
1641 if (tp->urg_data && tp->urg_seq == *seq) {
1644 if (signal_pending(current)) {
1645 copied = timeo ? sock_intr_errno(timeo) : -EAGAIN;
1650 /* Next get a buffer. */
1652 last = skb_peek_tail(&sk->sk_receive_queue);
1653 skb_queue_walk(&sk->sk_receive_queue, skb) {
1655 /* Now that we have two receive queues this
1658 if (WARN(before(*seq, TCP_SKB_CB(skb)->seq),
1659 "recvmsg bug: copied %X seq %X rcvnxt %X fl %X\n",
1660 *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt,
1664 offset = *seq - TCP_SKB_CB(skb)->seq;
1665 if (unlikely(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
1666 pr_err_once("%s: found a SYN, please report !\n", __func__);
1669 if (offset < skb->len)
1671 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1673 WARN(!(flags & MSG_PEEK),
1674 "recvmsg bug 2: copied %X seq %X rcvnxt %X fl %X\n",
1675 *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt, flags);
1678 /* Well, if we have backlog, try to process it now yet. */
1680 if (copied >= target && !sk->sk_backlog.tail)
1685 sk->sk_state == TCP_CLOSE ||
1686 (sk->sk_shutdown & RCV_SHUTDOWN) ||
1688 signal_pending(current))
1691 if (sock_flag(sk, SOCK_DONE))
1695 copied = sock_error(sk);
1699 if (sk->sk_shutdown & RCV_SHUTDOWN)
1702 if (sk->sk_state == TCP_CLOSE) {
1703 if (!sock_flag(sk, SOCK_DONE)) {
1704 /* This occurs when user tries to read
1705 * from never connected socket.
1718 if (signal_pending(current)) {
1719 copied = sock_intr_errno(timeo);
1724 tcp_cleanup_rbuf(sk, copied);
1726 if (!sysctl_tcp_low_latency && tp->ucopy.task == user_recv) {
1727 /* Install new reader */
1728 if (!user_recv && !(flags & (MSG_TRUNC | MSG_PEEK))) {
1729 user_recv = current;
1730 tp->ucopy.task = user_recv;
1731 tp->ucopy.msg = msg;
1734 tp->ucopy.len = len;
1736 WARN_ON(tp->copied_seq != tp->rcv_nxt &&
1737 !(flags & (MSG_PEEK | MSG_TRUNC)));
1739 /* Ugly... If prequeue is not empty, we have to
1740 * process it before releasing socket, otherwise
1741 * order will be broken at second iteration.
1742 * More elegant solution is required!!!
1744 * Look: we have the following (pseudo)queues:
1746 * 1. packets in flight
1751 * Each queue can be processed only if the next ones
1752 * are empty. At this point we have empty receive_queue.
1753 * But prequeue _can_ be not empty after 2nd iteration,
1754 * when we jumped to start of loop because backlog
1755 * processing added something to receive_queue.
1756 * We cannot release_sock(), because backlog contains
1757 * packets arrived _after_ prequeued ones.
1759 * Shortly, algorithm is clear --- to process all
1760 * the queues in order. We could make it more directly,
1761 * requeueing packets from backlog to prequeue, if
1762 * is not empty. It is more elegant, but eats cycles,
1765 if (!skb_queue_empty(&tp->ucopy.prequeue))
1768 /* __ Set realtime policy in scheduler __ */
1771 if (copied >= target) {
1772 /* Do not sleep, just process backlog. */
1776 sk_wait_data(sk, &timeo, last);
1782 /* __ Restore normal policy in scheduler __ */
1784 chunk = len - tp->ucopy.len;
1786 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMBACKLOG, chunk);
1791 if (tp->rcv_nxt == tp->copied_seq &&
1792 !skb_queue_empty(&tp->ucopy.prequeue)) {
1794 tcp_prequeue_process(sk);
1796 chunk = len - tp->ucopy.len;
1798 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1804 if ((flags & MSG_PEEK) &&
1805 (peek_seq - copied - urg_hole != tp->copied_seq)) {
1806 net_dbg_ratelimited("TCP(%s:%d): Application bug, race in MSG_PEEK\n",
1808 task_pid_nr(current));
1809 peek_seq = tp->copied_seq;
1814 /* Ok so how much can we use? */
1815 used = skb->len - offset;
1819 /* Do we have urgent data here? */
1821 u32 urg_offset = tp->urg_seq - *seq;
1822 if (urg_offset < used) {
1824 if (!sock_flag(sk, SOCK_URGINLINE)) {
1837 if (!(flags & MSG_TRUNC)) {
1838 err = skb_copy_datagram_msg(skb, offset, msg, used);
1840 /* Exception. Bailout! */
1851 tcp_rcv_space_adjust(sk);
1854 if (tp->urg_data && after(tp->copied_seq, tp->urg_seq)) {
1856 tcp_fast_path_check(sk);
1858 if (used + offset < skb->len)
1861 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1863 if (!(flags & MSG_PEEK))
1864 sk_eat_skb(sk, skb);
1868 /* Process the FIN. */
1870 if (!(flags & MSG_PEEK))
1871 sk_eat_skb(sk, skb);
1876 if (!skb_queue_empty(&tp->ucopy.prequeue)) {
1879 tp->ucopy.len = copied > 0 ? len : 0;
1881 tcp_prequeue_process(sk);
1883 if (copied > 0 && (chunk = len - tp->ucopy.len) != 0) {
1884 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1890 tp->ucopy.task = NULL;
1894 /* According to UNIX98, msg_name/msg_namelen are ignored
1895 * on connected socket. I was just happy when found this 8) --ANK
1898 /* Clean up data we have read: This will do ACK frames. */
1899 tcp_cleanup_rbuf(sk, copied);
1909 err = tcp_recv_urg(sk, msg, len, flags);
1913 err = tcp_peek_sndq(sk, msg, len);
1916 EXPORT_SYMBOL(tcp_recvmsg);
1918 void tcp_set_state(struct sock *sk, int state)
1920 int oldstate = sk->sk_state;
1923 case TCP_ESTABLISHED:
1924 if (oldstate != TCP_ESTABLISHED)
1925 TCP_INC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
1929 if (oldstate == TCP_CLOSE_WAIT || oldstate == TCP_ESTABLISHED)
1930 TCP_INC_STATS(sock_net(sk), TCP_MIB_ESTABRESETS);
1932 sk->sk_prot->unhash(sk);
1933 if (inet_csk(sk)->icsk_bind_hash &&
1934 !(sk->sk_userlocks & SOCK_BINDPORT_LOCK))
1938 if (oldstate == TCP_ESTABLISHED)
1939 TCP_DEC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
1942 /* Change state AFTER socket is unhashed to avoid closed
1943 * socket sitting in hash tables.
1945 sk_state_store(sk, state);
1948 SOCK_DEBUG(sk, "TCP sk=%p, State %s -> %s\n", sk, statename[oldstate], statename[state]);
1951 EXPORT_SYMBOL_GPL(tcp_set_state);
1954 * State processing on a close. This implements the state shift for
1955 * sending our FIN frame. Note that we only send a FIN for some
1956 * states. A shutdown() may have already sent the FIN, or we may be
1960 static const unsigned char new_state[16] = {
1961 /* current state: new state: action: */
1962 [0 /* (Invalid) */] = TCP_CLOSE,
1963 [TCP_ESTABLISHED] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1964 [TCP_SYN_SENT] = TCP_CLOSE,
1965 [TCP_SYN_RECV] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1966 [TCP_FIN_WAIT1] = TCP_FIN_WAIT1,
1967 [TCP_FIN_WAIT2] = TCP_FIN_WAIT2,
1968 [TCP_TIME_WAIT] = TCP_CLOSE,
1969 [TCP_CLOSE] = TCP_CLOSE,
1970 [TCP_CLOSE_WAIT] = TCP_LAST_ACK | TCP_ACTION_FIN,
1971 [TCP_LAST_ACK] = TCP_LAST_ACK,
1972 [TCP_LISTEN] = TCP_CLOSE,
1973 [TCP_CLOSING] = TCP_CLOSING,
1974 [TCP_NEW_SYN_RECV] = TCP_CLOSE, /* should not happen ! */
1977 static int tcp_close_state(struct sock *sk)
1979 int next = (int)new_state[sk->sk_state];
1980 int ns = next & TCP_STATE_MASK;
1982 tcp_set_state(sk, ns);
1984 return next & TCP_ACTION_FIN;
1988 * Shutdown the sending side of a connection. Much like close except
1989 * that we don't receive shut down or sock_set_flag(sk, SOCK_DEAD).
1992 void tcp_shutdown(struct sock *sk, int how)
1994 /* We need to grab some memory, and put together a FIN,
1995 * and then put it into the queue to be sent.
1996 * Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
1998 if (!(how & SEND_SHUTDOWN))
2001 /* If we've already sent a FIN, or it's a closed state, skip this. */
2002 if ((1 << sk->sk_state) &
2003 (TCPF_ESTABLISHED | TCPF_SYN_SENT |
2004 TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) {
2005 /* Clear out any half completed packets. FIN if needed. */
2006 if (tcp_close_state(sk))
2010 EXPORT_SYMBOL(tcp_shutdown);
2012 bool tcp_check_oom(struct sock *sk, int shift)
2014 bool too_many_orphans, out_of_socket_memory;
2016 too_many_orphans = tcp_too_many_orphans(sk, shift);
2017 out_of_socket_memory = tcp_out_of_memory(sk);
2019 if (too_many_orphans)
2020 net_info_ratelimited("too many orphaned sockets\n");
2021 if (out_of_socket_memory)
2022 net_info_ratelimited("out of memory -- consider tuning tcp_mem\n");
2023 return too_many_orphans || out_of_socket_memory;
2026 void tcp_close(struct sock *sk, long timeout)
2028 struct sk_buff *skb;
2029 int data_was_unread = 0;
2033 sk->sk_shutdown = SHUTDOWN_MASK;
2035 if (sk->sk_state == TCP_LISTEN) {
2036 tcp_set_state(sk, TCP_CLOSE);
2039 inet_csk_listen_stop(sk);
2041 goto adjudge_to_death;
2044 /* We need to flush the recv. buffs. We do this only on the
2045 * descriptor close, not protocol-sourced closes, because the
2046 * reader process may not have drained the data yet!
2048 while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) {
2049 u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq;
2051 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
2053 data_was_unread += len;
2059 /* If socket has been already reset (e.g. in tcp_reset()) - kill it. */
2060 if (sk->sk_state == TCP_CLOSE)
2061 goto adjudge_to_death;
2063 /* As outlined in RFC 2525, section 2.17, we send a RST here because
2064 * data was lost. To witness the awful effects of the old behavior of
2065 * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk
2066 * GET in an FTP client, suspend the process, wait for the client to
2067 * advertise a zero window, then kill -9 the FTP client, wheee...
2068 * Note: timeout is always zero in such a case.
2070 if (unlikely(tcp_sk(sk)->repair)) {
2071 sk->sk_prot->disconnect(sk, 0);
2072 } else if (data_was_unread) {
2073 /* Unread data was tossed, zap the connection. */
2074 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONCLOSE);
2075 tcp_set_state(sk, TCP_CLOSE);
2076 tcp_send_active_reset(sk, sk->sk_allocation);
2077 } else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) {
2078 /* Check zero linger _after_ checking for unread data. */
2079 sk->sk_prot->disconnect(sk, 0);
2080 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
2081 } else if (tcp_close_state(sk)) {
2082 /* We FIN if the application ate all the data before
2083 * zapping the connection.
2086 /* RED-PEN. Formally speaking, we have broken TCP state
2087 * machine. State transitions:
2089 * TCP_ESTABLISHED -> TCP_FIN_WAIT1
2090 * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible)
2091 * TCP_CLOSE_WAIT -> TCP_LAST_ACK
2093 * are legal only when FIN has been sent (i.e. in window),
2094 * rather than queued out of window. Purists blame.
2096 * F.e. "RFC state" is ESTABLISHED,
2097 * if Linux state is FIN-WAIT-1, but FIN is still not sent.
2099 * The visible declinations are that sometimes
2100 * we enter time-wait state, when it is not required really
2101 * (harmless), do not send active resets, when they are
2102 * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
2103 * they look as CLOSING or LAST_ACK for Linux)
2104 * Probably, I missed some more holelets.
2106 * XXX (TFO) - To start off we don't support SYN+ACK+FIN
2107 * in a single packet! (May consider it later but will
2108 * probably need API support or TCP_CORK SYN-ACK until
2109 * data is written and socket is closed.)
2114 sk_stream_wait_close(sk, timeout);
2117 state = sk->sk_state;
2121 /* It is the last release_sock in its life. It will remove backlog. */
2125 /* Now socket is owned by kernel and we acquire BH lock
2126 to finish close. No need to check for user refs.
2130 WARN_ON(sock_owned_by_user(sk));
2132 percpu_counter_inc(sk->sk_prot->orphan_count);
2134 /* Have we already been destroyed by a softirq or backlog? */
2135 if (state != TCP_CLOSE && sk->sk_state == TCP_CLOSE)
2138 /* This is a (useful) BSD violating of the RFC. There is a
2139 * problem with TCP as specified in that the other end could
2140 * keep a socket open forever with no application left this end.
2141 * We use a 1 minute timeout (about the same as BSD) then kill
2142 * our end. If they send after that then tough - BUT: long enough
2143 * that we won't make the old 4*rto = almost no time - whoops
2146 * Nope, it was not mistake. It is really desired behaviour
2147 * f.e. on http servers, when such sockets are useless, but
2148 * consume significant resources. Let's do it with special
2149 * linger2 option. --ANK
2152 if (sk->sk_state == TCP_FIN_WAIT2) {
2153 struct tcp_sock *tp = tcp_sk(sk);
2154 if (tp->linger2 < 0) {
2155 tcp_set_state(sk, TCP_CLOSE);
2156 tcp_send_active_reset(sk, GFP_ATOMIC);
2157 __NET_INC_STATS(sock_net(sk),
2158 LINUX_MIB_TCPABORTONLINGER);
2160 const int tmo = tcp_fin_time(sk);
2162 if (tmo > TCP_TIMEWAIT_LEN) {
2163 inet_csk_reset_keepalive_timer(sk,
2164 tmo - TCP_TIMEWAIT_LEN);
2166 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
2171 if (sk->sk_state != TCP_CLOSE) {
2173 if (tcp_check_oom(sk, 0)) {
2174 tcp_set_state(sk, TCP_CLOSE);
2175 tcp_send_active_reset(sk, GFP_ATOMIC);
2176 __NET_INC_STATS(sock_net(sk),
2177 LINUX_MIB_TCPABORTONMEMORY);
2181 if (sk->sk_state == TCP_CLOSE) {
2182 struct request_sock *req = tcp_sk(sk)->fastopen_rsk;
2183 /* We could get here with a non-NULL req if the socket is
2184 * aborted (e.g., closed with unread data) before 3WHS
2188 reqsk_fastopen_remove(sk, req, false);
2189 inet_csk_destroy_sock(sk);
2191 /* Otherwise, socket is reprieved until protocol close. */
2198 EXPORT_SYMBOL(tcp_close);
2200 /* These states need RST on ABORT according to RFC793 */
2202 static inline bool tcp_need_reset(int state)
2204 return (1 << state) &
2205 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 |
2206 TCPF_FIN_WAIT2 | TCPF_SYN_RECV);
2209 int tcp_disconnect(struct sock *sk, int flags)
2211 struct inet_sock *inet = inet_sk(sk);
2212 struct inet_connection_sock *icsk = inet_csk(sk);
2213 struct tcp_sock *tp = tcp_sk(sk);
2215 int old_state = sk->sk_state;
2217 if (old_state != TCP_CLOSE)
2218 tcp_set_state(sk, TCP_CLOSE);
2220 /* ABORT function of RFC793 */
2221 if (old_state == TCP_LISTEN) {
2222 inet_csk_listen_stop(sk);
2223 } else if (unlikely(tp->repair)) {
2224 sk->sk_err = ECONNABORTED;
2225 } else if (tcp_need_reset(old_state) ||
2226 (tp->snd_nxt != tp->write_seq &&
2227 (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK))) {
2228 /* The last check adjusts for discrepancy of Linux wrt. RFC
2231 tcp_send_active_reset(sk, gfp_any());
2232 sk->sk_err = ECONNRESET;
2233 } else if (old_state == TCP_SYN_SENT)
2234 sk->sk_err = ECONNRESET;
2236 tcp_clear_xmit_timers(sk);
2237 __skb_queue_purge(&sk->sk_receive_queue);
2238 tcp_write_queue_purge(sk);
2239 __skb_queue_purge(&tp->out_of_order_queue);
2241 inet->inet_dport = 0;
2243 if (!(sk->sk_userlocks & SOCK_BINDADDR_LOCK))
2244 inet_reset_saddr(sk);
2246 sk->sk_shutdown = 0;
2247 sock_reset_flag(sk, SOCK_DONE);
2249 tp->write_seq += tp->max_window + 2;
2250 if (tp->write_seq == 0)
2252 icsk->icsk_backoff = 0;
2254 icsk->icsk_probes_out = 0;
2255 tp->packets_out = 0;
2256 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
2257 tp->snd_cwnd_cnt = 0;
2258 tp->window_clamp = 0;
2259 tcp_set_ca_state(sk, TCP_CA_Open);
2260 tcp_clear_retrans(tp);
2261 inet_csk_delack_init(sk);
2262 tcp_init_send_head(sk);
2263 memset(&tp->rx_opt, 0, sizeof(tp->rx_opt));
2266 WARN_ON(inet->inet_num && !icsk->icsk_bind_hash);
2268 sk->sk_error_report(sk);
2271 EXPORT_SYMBOL(tcp_disconnect);
2273 static inline bool tcp_can_repair_sock(const struct sock *sk)
2275 return ns_capable(sock_net(sk)->user_ns, CAP_NET_ADMIN) &&
2276 ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_ESTABLISHED));
2279 static int tcp_repair_set_window(struct tcp_sock *tp, char __user *optbuf, int len)
2281 struct tcp_repair_window opt;
2286 if (len != sizeof(opt))
2289 if (copy_from_user(&opt, optbuf, sizeof(opt)))
2292 if (opt.max_window < opt.snd_wnd)
2295 if (after(opt.snd_wl1, tp->rcv_nxt + opt.rcv_wnd))
2298 if (after(opt.rcv_wup, tp->rcv_nxt))
2301 tp->snd_wl1 = opt.snd_wl1;
2302 tp->snd_wnd = opt.snd_wnd;
2303 tp->max_window = opt.max_window;
2305 tp->rcv_wnd = opt.rcv_wnd;
2306 tp->rcv_wup = opt.rcv_wup;
2311 static int tcp_repair_options_est(struct tcp_sock *tp,
2312 struct tcp_repair_opt __user *optbuf, unsigned int len)
2314 struct tcp_repair_opt opt;
2316 while (len >= sizeof(opt)) {
2317 if (copy_from_user(&opt, optbuf, sizeof(opt)))
2323 switch (opt.opt_code) {
2325 tp->rx_opt.mss_clamp = opt.opt_val;
2329 u16 snd_wscale = opt.opt_val & 0xFFFF;
2330 u16 rcv_wscale = opt.opt_val >> 16;
2332 if (snd_wscale > 14 || rcv_wscale > 14)
2335 tp->rx_opt.snd_wscale = snd_wscale;
2336 tp->rx_opt.rcv_wscale = rcv_wscale;
2337 tp->rx_opt.wscale_ok = 1;
2340 case TCPOPT_SACK_PERM:
2341 if (opt.opt_val != 0)
2344 tp->rx_opt.sack_ok |= TCP_SACK_SEEN;
2345 if (sysctl_tcp_fack)
2346 tcp_enable_fack(tp);
2348 case TCPOPT_TIMESTAMP:
2349 if (opt.opt_val != 0)
2352 tp->rx_opt.tstamp_ok = 1;
2361 * Socket option code for TCP.
2363 static int do_tcp_setsockopt(struct sock *sk, int level,
2364 int optname, char __user *optval, unsigned int optlen)
2366 struct tcp_sock *tp = tcp_sk(sk);
2367 struct inet_connection_sock *icsk = inet_csk(sk);
2368 struct net *net = sock_net(sk);
2372 /* These are data/string values, all the others are ints */
2374 case TCP_CONGESTION: {
2375 char name[TCP_CA_NAME_MAX];
2380 val = strncpy_from_user(name, optval,
2381 min_t(long, TCP_CA_NAME_MAX-1, optlen));
2387 err = tcp_set_congestion_control(sk, name);
2396 if (optlen < sizeof(int))
2399 if (get_user(val, (int __user *)optval))
2406 /* Values greater than interface MTU won't take effect. However
2407 * at the point when this call is done we typically don't yet
2408 * know which interface is going to be used */
2409 if (val < TCP_MIN_MSS || val > MAX_TCP_WINDOW) {
2413 tp->rx_opt.user_mss = val;
2418 /* TCP_NODELAY is weaker than TCP_CORK, so that
2419 * this option on corked socket is remembered, but
2420 * it is not activated until cork is cleared.
2422 * However, when TCP_NODELAY is set we make
2423 * an explicit push, which overrides even TCP_CORK
2424 * for currently queued segments.
2426 tp->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH;
2427 tcp_push_pending_frames(sk);
2429 tp->nonagle &= ~TCP_NAGLE_OFF;
2433 case TCP_THIN_LINEAR_TIMEOUTS:
2434 if (val < 0 || val > 1)
2440 case TCP_THIN_DUPACK:
2441 if (val < 0 || val > 1)
2444 tp->thin_dupack = val;
2445 if (tp->thin_dupack)
2446 tcp_disable_early_retrans(tp);
2451 if (!tcp_can_repair_sock(sk))
2453 else if (val == 1) {
2455 sk->sk_reuse = SK_FORCE_REUSE;
2456 tp->repair_queue = TCP_NO_QUEUE;
2457 } else if (val == 0) {
2459 sk->sk_reuse = SK_NO_REUSE;
2460 tcp_send_window_probe(sk);
2466 case TCP_REPAIR_QUEUE:
2469 else if (val < TCP_QUEUES_NR)
2470 tp->repair_queue = val;
2476 if (sk->sk_state != TCP_CLOSE)
2478 else if (tp->repair_queue == TCP_SEND_QUEUE)
2479 tp->write_seq = val;
2480 else if (tp->repair_queue == TCP_RECV_QUEUE)
2486 case TCP_REPAIR_OPTIONS:
2489 else if (sk->sk_state == TCP_ESTABLISHED)
2490 err = tcp_repair_options_est(tp,
2491 (struct tcp_repair_opt __user *)optval,
2498 /* When set indicates to always queue non-full frames.
2499 * Later the user clears this option and we transmit
2500 * any pending partial frames in the queue. This is
2501 * meant to be used alongside sendfile() to get properly
2502 * filled frames when the user (for example) must write
2503 * out headers with a write() call first and then use
2504 * sendfile to send out the data parts.
2506 * TCP_CORK can be set together with TCP_NODELAY and it is
2507 * stronger than TCP_NODELAY.
2510 tp->nonagle |= TCP_NAGLE_CORK;
2512 tp->nonagle &= ~TCP_NAGLE_CORK;
2513 if (tp->nonagle&TCP_NAGLE_OFF)
2514 tp->nonagle |= TCP_NAGLE_PUSH;
2515 tcp_push_pending_frames(sk);
2520 if (val < 1 || val > MAX_TCP_KEEPIDLE)
2523 tp->keepalive_time = val * HZ;
2524 if (sock_flag(sk, SOCK_KEEPOPEN) &&
2525 !((1 << sk->sk_state) &
2526 (TCPF_CLOSE | TCPF_LISTEN))) {
2527 u32 elapsed = keepalive_time_elapsed(tp);
2528 if (tp->keepalive_time > elapsed)
2529 elapsed = tp->keepalive_time - elapsed;
2532 inet_csk_reset_keepalive_timer(sk, elapsed);
2537 if (val < 1 || val > MAX_TCP_KEEPINTVL)
2540 tp->keepalive_intvl = val * HZ;
2543 if (val < 1 || val > MAX_TCP_KEEPCNT)
2546 tp->keepalive_probes = val;
2549 if (val < 1 || val > MAX_TCP_SYNCNT)
2552 icsk->icsk_syn_retries = val;
2556 if (val < 0 || val > 1)
2565 else if (val > net->ipv4.sysctl_tcp_fin_timeout / HZ)
2568 tp->linger2 = val * HZ;
2571 case TCP_DEFER_ACCEPT:
2572 /* Translate value in seconds to number of retransmits */
2573 icsk->icsk_accept_queue.rskq_defer_accept =
2574 secs_to_retrans(val, TCP_TIMEOUT_INIT / HZ,
2578 case TCP_WINDOW_CLAMP:
2580 if (sk->sk_state != TCP_CLOSE) {
2584 tp->window_clamp = 0;
2586 tp->window_clamp = val < SOCK_MIN_RCVBUF / 2 ?
2587 SOCK_MIN_RCVBUF / 2 : val;
2592 icsk->icsk_ack.pingpong = 1;
2594 icsk->icsk_ack.pingpong = 0;
2595 if ((1 << sk->sk_state) &
2596 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) &&
2597 inet_csk_ack_scheduled(sk)) {
2598 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
2599 tcp_cleanup_rbuf(sk, 1);
2601 icsk->icsk_ack.pingpong = 1;
2606 #ifdef CONFIG_TCP_MD5SIG
2608 /* Read the IP->Key mappings from userspace */
2609 err = tp->af_specific->md5_parse(sk, optval, optlen);
2612 case TCP_USER_TIMEOUT:
2613 /* Cap the max time in ms TCP will retry or probe the window
2614 * before giving up and aborting (ETIMEDOUT) a connection.
2619 icsk->icsk_user_timeout = msecs_to_jiffies(val);
2623 if (val >= 0 && ((1 << sk->sk_state) & (TCPF_CLOSE |
2625 tcp_fastopen_init_key_once(true);
2627 fastopen_queue_tune(sk, val);
2636 tp->tsoffset = val - tcp_time_stamp;
2638 case TCP_REPAIR_WINDOW:
2639 err = tcp_repair_set_window(tp, optval, optlen);
2641 case TCP_NOTSENT_LOWAT:
2642 tp->notsent_lowat = val;
2643 sk->sk_write_space(sk);
2654 int tcp_setsockopt(struct sock *sk, int level, int optname, char __user *optval,
2655 unsigned int optlen)
2657 const struct inet_connection_sock *icsk = inet_csk(sk);
2659 if (level != SOL_TCP)
2660 return icsk->icsk_af_ops->setsockopt(sk, level, optname,
2662 return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2664 EXPORT_SYMBOL(tcp_setsockopt);
2666 #ifdef CONFIG_COMPAT
2667 int compat_tcp_setsockopt(struct sock *sk, int level, int optname,
2668 char __user *optval, unsigned int optlen)
2670 if (level != SOL_TCP)
2671 return inet_csk_compat_setsockopt(sk, level, optname,
2673 return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2675 EXPORT_SYMBOL(compat_tcp_setsockopt);
2678 /* Return information about state of tcp endpoint in API format. */
2679 void tcp_get_info(struct sock *sk, struct tcp_info *info)
2681 const struct tcp_sock *tp = tcp_sk(sk); /* iff sk_type == SOCK_STREAM */
2682 const struct inet_connection_sock *icsk = inet_csk(sk);
2683 u32 now = tcp_time_stamp;
2689 memset(info, 0, sizeof(*info));
2690 if (sk->sk_type != SOCK_STREAM)
2693 info->tcpi_state = sk_state_load(sk);
2695 info->tcpi_ca_state = icsk->icsk_ca_state;
2696 info->tcpi_retransmits = icsk->icsk_retransmits;
2697 info->tcpi_probes = icsk->icsk_probes_out;
2698 info->tcpi_backoff = icsk->icsk_backoff;
2700 if (tp->rx_opt.tstamp_ok)
2701 info->tcpi_options |= TCPI_OPT_TIMESTAMPS;
2702 if (tcp_is_sack(tp))
2703 info->tcpi_options |= TCPI_OPT_SACK;
2704 if (tp->rx_opt.wscale_ok) {
2705 info->tcpi_options |= TCPI_OPT_WSCALE;
2706 info->tcpi_snd_wscale = tp->rx_opt.snd_wscale;
2707 info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale;
2710 if (tp->ecn_flags & TCP_ECN_OK)
2711 info->tcpi_options |= TCPI_OPT_ECN;
2712 if (tp->ecn_flags & TCP_ECN_SEEN)
2713 info->tcpi_options |= TCPI_OPT_ECN_SEEN;
2714 if (tp->syn_data_acked)
2715 info->tcpi_options |= TCPI_OPT_SYN_DATA;
2717 info->tcpi_rto = jiffies_to_usecs(icsk->icsk_rto);
2718 info->tcpi_ato = jiffies_to_usecs(icsk->icsk_ack.ato);
2719 info->tcpi_snd_mss = tp->mss_cache;
2720 info->tcpi_rcv_mss = icsk->icsk_ack.rcv_mss;
2722 if (info->tcpi_state == TCP_LISTEN) {
2723 info->tcpi_unacked = sk->sk_ack_backlog;
2724 info->tcpi_sacked = sk->sk_max_ack_backlog;
2726 info->tcpi_unacked = tp->packets_out;
2727 info->tcpi_sacked = tp->sacked_out;
2729 info->tcpi_lost = tp->lost_out;
2730 info->tcpi_retrans = tp->retrans_out;
2731 info->tcpi_fackets = tp->fackets_out;
2733 info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime);
2734 info->tcpi_last_data_recv = jiffies_to_msecs(now - icsk->icsk_ack.lrcvtime);
2735 info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp);
2737 info->tcpi_pmtu = icsk->icsk_pmtu_cookie;
2738 info->tcpi_rcv_ssthresh = tp->rcv_ssthresh;
2739 info->tcpi_rtt = tp->srtt_us >> 3;
2740 info->tcpi_rttvar = tp->mdev_us >> 2;
2741 info->tcpi_snd_ssthresh = tp->snd_ssthresh;
2742 info->tcpi_snd_cwnd = tp->snd_cwnd;
2743 info->tcpi_advmss = tp->advmss;
2744 info->tcpi_reordering = tp->reordering;
2746 info->tcpi_rcv_rtt = jiffies_to_usecs(tp->rcv_rtt_est.rtt)>>3;
2747 info->tcpi_rcv_space = tp->rcvq_space.space;
2749 info->tcpi_total_retrans = tp->total_retrans;
2751 rate = READ_ONCE(sk->sk_pacing_rate);
2752 rate64 = rate != ~0U ? rate : ~0ULL;
2753 put_unaligned(rate64, &info->tcpi_pacing_rate);
2755 rate = READ_ONCE(sk->sk_max_pacing_rate);
2756 rate64 = rate != ~0U ? rate : ~0ULL;
2757 put_unaligned(rate64, &info->tcpi_max_pacing_rate);
2760 start = u64_stats_fetch_begin_irq(&tp->syncp);
2761 put_unaligned(tp->bytes_acked, &info->tcpi_bytes_acked);
2762 put_unaligned(tp->bytes_received, &info->tcpi_bytes_received);
2763 } while (u64_stats_fetch_retry_irq(&tp->syncp, start));
2764 info->tcpi_segs_out = tp->segs_out;
2765 info->tcpi_segs_in = tp->segs_in;
2767 notsent_bytes = READ_ONCE(tp->write_seq) - READ_ONCE(tp->snd_nxt);
2768 info->tcpi_notsent_bytes = max(0, notsent_bytes);
2770 info->tcpi_min_rtt = tcp_min_rtt(tp);
2771 info->tcpi_data_segs_in = tp->data_segs_in;
2772 info->tcpi_data_segs_out = tp->data_segs_out;
2774 EXPORT_SYMBOL_GPL(tcp_get_info);
2776 static int do_tcp_getsockopt(struct sock *sk, int level,
2777 int optname, char __user *optval, int __user *optlen)
2779 struct inet_connection_sock *icsk = inet_csk(sk);
2780 struct tcp_sock *tp = tcp_sk(sk);
2781 struct net *net = sock_net(sk);
2784 if (get_user(len, optlen))
2787 len = min_t(unsigned int, len, sizeof(int));
2794 val = tp->mss_cache;
2795 if (!val && ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
2796 val = tp->rx_opt.user_mss;
2798 val = tp->rx_opt.mss_clamp;
2801 val = !!(tp->nonagle&TCP_NAGLE_OFF);
2804 val = !!(tp->nonagle&TCP_NAGLE_CORK);
2807 val = keepalive_time_when(tp) / HZ;
2810 val = keepalive_intvl_when(tp) / HZ;
2813 val = keepalive_probes(tp);
2816 val = icsk->icsk_syn_retries ? : net->ipv4.sysctl_tcp_syn_retries;
2821 val = (val ? : net->ipv4.sysctl_tcp_fin_timeout) / HZ;
2823 case TCP_DEFER_ACCEPT:
2824 val = retrans_to_secs(icsk->icsk_accept_queue.rskq_defer_accept,
2825 TCP_TIMEOUT_INIT / HZ, TCP_RTO_MAX / HZ);
2827 case TCP_WINDOW_CLAMP:
2828 val = tp->window_clamp;
2831 struct tcp_info info;
2833 if (get_user(len, optlen))
2836 tcp_get_info(sk, &info);
2838 len = min_t(unsigned int, len, sizeof(info));
2839 if (put_user(len, optlen))
2841 if (copy_to_user(optval, &info, len))
2846 const struct tcp_congestion_ops *ca_ops;
2847 union tcp_cc_info info;
2851 if (get_user(len, optlen))
2854 ca_ops = icsk->icsk_ca_ops;
2855 if (ca_ops && ca_ops->get_info)
2856 sz = ca_ops->get_info(sk, ~0U, &attr, &info);
2858 len = min_t(unsigned int, len, sz);
2859 if (put_user(len, optlen))
2861 if (copy_to_user(optval, &info, len))
2866 val = !icsk->icsk_ack.pingpong;
2869 case TCP_CONGESTION:
2870 if (get_user(len, optlen))
2872 len = min_t(unsigned int, len, TCP_CA_NAME_MAX);
2873 if (put_user(len, optlen))
2875 if (copy_to_user(optval, icsk->icsk_ca_ops->name, len))
2879 case TCP_THIN_LINEAR_TIMEOUTS:
2882 case TCP_THIN_DUPACK:
2883 val = tp->thin_dupack;
2890 case TCP_REPAIR_QUEUE:
2892 val = tp->repair_queue;
2897 case TCP_REPAIR_WINDOW: {
2898 struct tcp_repair_window opt;
2900 if (get_user(len, optlen))
2903 if (len != sizeof(opt))
2909 opt.snd_wl1 = tp->snd_wl1;
2910 opt.snd_wnd = tp->snd_wnd;
2911 opt.max_window = tp->max_window;
2912 opt.rcv_wnd = tp->rcv_wnd;
2913 opt.rcv_wup = tp->rcv_wup;
2915 if (copy_to_user(optval, &opt, len))
2920 if (tp->repair_queue == TCP_SEND_QUEUE)
2921 val = tp->write_seq;
2922 else if (tp->repair_queue == TCP_RECV_QUEUE)
2928 case TCP_USER_TIMEOUT:
2929 val = jiffies_to_msecs(icsk->icsk_user_timeout);
2933 val = icsk->icsk_accept_queue.fastopenq.max_qlen;
2937 val = tcp_time_stamp + tp->tsoffset;
2939 case TCP_NOTSENT_LOWAT:
2940 val = tp->notsent_lowat;
2945 case TCP_SAVED_SYN: {
2946 if (get_user(len, optlen))
2950 if (tp->saved_syn) {
2951 if (len < tp->saved_syn[0]) {
2952 if (put_user(tp->saved_syn[0], optlen)) {
2959 len = tp->saved_syn[0];
2960 if (put_user(len, optlen)) {
2964 if (copy_to_user(optval, tp->saved_syn + 1, len)) {
2968 tcp_saved_syn_free(tp);
2973 if (put_user(len, optlen))
2979 return -ENOPROTOOPT;
2982 if (put_user(len, optlen))
2984 if (copy_to_user(optval, &val, len))
2989 int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval,
2992 struct inet_connection_sock *icsk = inet_csk(sk);
2994 if (level != SOL_TCP)
2995 return icsk->icsk_af_ops->getsockopt(sk, level, optname,
2997 return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2999 EXPORT_SYMBOL(tcp_getsockopt);
3001 #ifdef CONFIG_COMPAT
3002 int compat_tcp_getsockopt(struct sock *sk, int level, int optname,
3003 char __user *optval, int __user *optlen)
3005 if (level != SOL_TCP)
3006 return inet_csk_compat_getsockopt(sk, level, optname,
3008 return do_tcp_getsockopt(sk, level, optname, optval, optlen);
3010 EXPORT_SYMBOL(compat_tcp_getsockopt);
3013 #ifdef CONFIG_TCP_MD5SIG
3014 static DEFINE_PER_CPU(struct tcp_md5sig_pool, tcp_md5sig_pool);
3015 static DEFINE_MUTEX(tcp_md5sig_mutex);
3016 static bool tcp_md5sig_pool_populated = false;
3018 static void __tcp_alloc_md5sig_pool(void)
3020 struct crypto_ahash *hash;
3023 hash = crypto_alloc_ahash("md5", 0, CRYPTO_ALG_ASYNC);
3027 for_each_possible_cpu(cpu) {
3028 void *scratch = per_cpu(tcp_md5sig_pool, cpu).scratch;
3029 struct ahash_request *req;
3032 scratch = kmalloc_node(sizeof(union tcp_md5sum_block) +
3033 sizeof(struct tcphdr),
3038 per_cpu(tcp_md5sig_pool, cpu).scratch = scratch;
3040 if (per_cpu(tcp_md5sig_pool, cpu).md5_req)
3043 req = ahash_request_alloc(hash, GFP_KERNEL);
3047 ahash_request_set_callback(req, 0, NULL, NULL);
3049 per_cpu(tcp_md5sig_pool, cpu).md5_req = req;
3051 /* before setting tcp_md5sig_pool_populated, we must commit all writes
3052 * to memory. See smp_rmb() in tcp_get_md5sig_pool()
3055 tcp_md5sig_pool_populated = true;
3058 bool tcp_alloc_md5sig_pool(void)
3060 if (unlikely(!tcp_md5sig_pool_populated)) {
3061 mutex_lock(&tcp_md5sig_mutex);
3063 if (!tcp_md5sig_pool_populated)
3064 __tcp_alloc_md5sig_pool();
3066 mutex_unlock(&tcp_md5sig_mutex);
3068 return tcp_md5sig_pool_populated;
3070 EXPORT_SYMBOL(tcp_alloc_md5sig_pool);
3074 * tcp_get_md5sig_pool - get md5sig_pool for this user
3076 * We use percpu structure, so if we succeed, we exit with preemption
3077 * and BH disabled, to make sure another thread or softirq handling
3078 * wont try to get same context.
3080 struct tcp_md5sig_pool *tcp_get_md5sig_pool(void)
3084 if (tcp_md5sig_pool_populated) {
3085 /* coupled with smp_wmb() in __tcp_alloc_md5sig_pool() */
3087 return this_cpu_ptr(&tcp_md5sig_pool);
3092 EXPORT_SYMBOL(tcp_get_md5sig_pool);
3094 int tcp_md5_hash_header(struct tcp_md5sig_pool *hp,
3095 const struct tcphdr *th)
3097 struct scatterlist sg;
3100 /* We are not allowed to change tcphdr, make a local copy */
3101 memcpy(&hdr, th, sizeof(hdr));
3104 /* options aren't included in the hash */
3105 sg_init_one(&sg, &hdr, sizeof(hdr));
3106 ahash_request_set_crypt(hp->md5_req, &sg, NULL, sizeof(hdr));
3107 return crypto_ahash_update(hp->md5_req);
3109 EXPORT_SYMBOL(tcp_md5_hash_header);
3111 int tcp_md5_hash_skb_data(struct tcp_md5sig_pool *hp,
3112 const struct sk_buff *skb, unsigned int header_len)
3114 struct scatterlist sg;
3115 const struct tcphdr *tp = tcp_hdr(skb);
3116 struct ahash_request *req = hp->md5_req;
3118 const unsigned int head_data_len = skb_headlen(skb) > header_len ?
3119 skb_headlen(skb) - header_len : 0;
3120 const struct skb_shared_info *shi = skb_shinfo(skb);
3121 struct sk_buff *frag_iter;
3123 sg_init_table(&sg, 1);
3125 sg_set_buf(&sg, ((u8 *) tp) + header_len, head_data_len);
3126 ahash_request_set_crypt(req, &sg, NULL, head_data_len);
3127 if (crypto_ahash_update(req))
3130 for (i = 0; i < shi->nr_frags; ++i) {
3131 const struct skb_frag_struct *f = &shi->frags[i];
3132 unsigned int offset = f->page_offset;
3133 struct page *page = skb_frag_page(f) + (offset >> PAGE_SHIFT);
3135 sg_set_page(&sg, page, skb_frag_size(f),
3136 offset_in_page(offset));
3137 ahash_request_set_crypt(req, &sg, NULL, skb_frag_size(f));
3138 if (crypto_ahash_update(req))
3142 skb_walk_frags(skb, frag_iter)
3143 if (tcp_md5_hash_skb_data(hp, frag_iter, 0))
3148 EXPORT_SYMBOL(tcp_md5_hash_skb_data);
3150 int tcp_md5_hash_key(struct tcp_md5sig_pool *hp, const struct tcp_md5sig_key *key)
3152 struct scatterlist sg;
3154 sg_init_one(&sg, key->key, key->keylen);
3155 ahash_request_set_crypt(hp->md5_req, &sg, NULL, key->keylen);
3156 return crypto_ahash_update(hp->md5_req);
3158 EXPORT_SYMBOL(tcp_md5_hash_key);
3162 void tcp_done(struct sock *sk)
3164 struct request_sock *req = tcp_sk(sk)->fastopen_rsk;
3166 if (sk->sk_state == TCP_SYN_SENT || sk->sk_state == TCP_SYN_RECV)
3167 TCP_INC_STATS(sock_net(sk), TCP_MIB_ATTEMPTFAILS);
3169 tcp_set_state(sk, TCP_CLOSE);
3170 tcp_clear_xmit_timers(sk);
3172 reqsk_fastopen_remove(sk, req, false);
3174 sk->sk_shutdown = SHUTDOWN_MASK;
3176 if (!sock_flag(sk, SOCK_DEAD))
3177 sk->sk_state_change(sk);
3179 inet_csk_destroy_sock(sk);
3181 EXPORT_SYMBOL_GPL(tcp_done);
3183 int tcp_abort(struct sock *sk, int err)
3185 if (!sk_fullsock(sk)) {
3186 if (sk->sk_state == TCP_NEW_SYN_RECV) {
3187 struct request_sock *req = inet_reqsk(sk);
3190 inet_csk_reqsk_queue_drop_and_put(req->rsk_listener,
3198 /* Don't race with userspace socket closes such as tcp_close. */
3201 if (sk->sk_state == TCP_LISTEN) {
3202 tcp_set_state(sk, TCP_CLOSE);
3203 inet_csk_listen_stop(sk);
3206 /* Don't race with BH socket closes such as inet_csk_listen_stop. */
3210 if (!sock_flag(sk, SOCK_DEAD)) {
3212 /* This barrier is coupled with smp_rmb() in tcp_poll() */
3214 sk->sk_error_report(sk);
3215 if (tcp_need_reset(sk->sk_state))
3216 tcp_send_active_reset(sk, GFP_ATOMIC);
3225 EXPORT_SYMBOL_GPL(tcp_abort);
3227 extern struct tcp_congestion_ops tcp_reno;
3229 static __initdata unsigned long thash_entries;
3230 static int __init set_thash_entries(char *str)
3237 ret = kstrtoul(str, 0, &thash_entries);
3243 __setup("thash_entries=", set_thash_entries);
3245 static void __init tcp_init_mem(void)
3247 unsigned long limit = nr_free_buffer_pages() / 16;
3249 limit = max(limit, 128UL);
3250 sysctl_tcp_mem[0] = limit / 4 * 3; /* 4.68 % */
3251 sysctl_tcp_mem[1] = limit; /* 6.25 % */
3252 sysctl_tcp_mem[2] = sysctl_tcp_mem[0] * 2; /* 9.37 % */
3255 void __init tcp_init(void)
3257 unsigned long limit;
3258 int max_rshare, max_wshare, cnt;
3261 sock_skb_cb_check_size(sizeof(struct tcp_skb_cb));
3263 percpu_counter_init(&tcp_sockets_allocated, 0, GFP_KERNEL);
3264 percpu_counter_init(&tcp_orphan_count, 0, GFP_KERNEL);
3265 tcp_hashinfo.bind_bucket_cachep =
3266 kmem_cache_create("tcp_bind_bucket",
3267 sizeof(struct inet_bind_bucket), 0,
3268 SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL);
3270 /* Size and allocate the main established and bind bucket
3273 * The methodology is similar to that of the buffer cache.
3275 tcp_hashinfo.ehash =
3276 alloc_large_system_hash("TCP established",
3277 sizeof(struct inet_ehash_bucket),
3279 17, /* one slot per 128 KB of memory */
3282 &tcp_hashinfo.ehash_mask,
3284 thash_entries ? 0 : 512 * 1024);
3285 for (i = 0; i <= tcp_hashinfo.ehash_mask; i++)
3286 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].chain, i);
3288 if (inet_ehash_locks_alloc(&tcp_hashinfo))
3289 panic("TCP: failed to alloc ehash_locks");
3290 tcp_hashinfo.bhash =
3291 alloc_large_system_hash("TCP bind",
3292 sizeof(struct inet_bind_hashbucket),
3293 tcp_hashinfo.ehash_mask + 1,
3294 17, /* one slot per 128 KB of memory */
3296 &tcp_hashinfo.bhash_size,
3300 tcp_hashinfo.bhash_size = 1U << tcp_hashinfo.bhash_size;
3301 for (i = 0; i < tcp_hashinfo.bhash_size; i++) {
3302 spin_lock_init(&tcp_hashinfo.bhash[i].lock);
3303 INIT_HLIST_HEAD(&tcp_hashinfo.bhash[i].chain);
3307 cnt = tcp_hashinfo.ehash_mask + 1;
3309 tcp_death_row.sysctl_max_tw_buckets = cnt / 2;
3310 sysctl_tcp_max_orphans = cnt / 2;
3311 sysctl_max_syn_backlog = max(128, cnt / 256);
3314 /* Set per-socket limits to no more than 1/128 the pressure threshold */
3315 limit = nr_free_buffer_pages() << (PAGE_SHIFT - 7);
3316 max_wshare = min(4UL*1024*1024, limit);
3317 max_rshare = min(6UL*1024*1024, limit);
3319 sysctl_tcp_wmem[0] = SK_MEM_QUANTUM;
3320 sysctl_tcp_wmem[1] = 16*1024;
3321 sysctl_tcp_wmem[2] = max(64*1024, max_wshare);
3323 sysctl_tcp_rmem[0] = SK_MEM_QUANTUM;
3324 sysctl_tcp_rmem[1] = 87380;
3325 sysctl_tcp_rmem[2] = max(87380, max_rshare);
3327 pr_info("Hash tables configured (established %u bind %u)\n",
3328 tcp_hashinfo.ehash_mask + 1, tcp_hashinfo.bhash_size);
3331 BUG_ON(tcp_register_congestion_control(&tcp_reno) != 0);