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 tp->out_of_order_queue = RB_ROOT;
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 minmax_reset(&tp->rtt_min, tcp_time_stamp, ~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 /* There's a bubble in the pipe until at least the first ACK. */
400 tp->app_limited = ~0U;
402 /* See draft-stevens-tcpca-spec-01 for discussion of the
403 * initialization of these values.
405 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
406 tp->snd_cwnd_clamp = ~0;
407 tp->mss_cache = TCP_MSS_DEFAULT;
408 u64_stats_init(&tp->syncp);
410 tp->reordering = sock_net(sk)->ipv4.sysctl_tcp_reordering;
411 tcp_enable_early_retrans(tp);
412 tcp_assign_congestion_control(sk);
416 sk->sk_state = TCP_CLOSE;
418 sk->sk_write_space = sk_stream_write_space;
419 sock_set_flag(sk, SOCK_USE_WRITE_QUEUE);
421 icsk->icsk_sync_mss = tcp_sync_mss;
423 sk->sk_sndbuf = sysctl_tcp_wmem[1];
424 sk->sk_rcvbuf = sysctl_tcp_rmem[1];
427 sk_sockets_allocated_inc(sk);
430 EXPORT_SYMBOL(tcp_init_sock);
432 static void tcp_tx_timestamp(struct sock *sk, u16 tsflags, struct sk_buff *skb)
435 struct skb_shared_info *shinfo = skb_shinfo(skb);
436 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
438 sock_tx_timestamp(sk, tsflags, &shinfo->tx_flags);
439 if (tsflags & SOF_TIMESTAMPING_TX_ACK)
440 tcb->txstamp_ack = 1;
441 if (tsflags & SOF_TIMESTAMPING_TX_RECORD_MASK)
442 shinfo->tskey = TCP_SKB_CB(skb)->seq + skb->len - 1;
447 * Wait for a TCP event.
449 * Note that we don't need to lock the socket, as the upper poll layers
450 * take care of normal races (between the test and the event) and we don't
451 * go look at any of the socket buffers directly.
453 unsigned int tcp_poll(struct file *file, struct socket *sock, poll_table *wait)
456 struct sock *sk = sock->sk;
457 const struct tcp_sock *tp = tcp_sk(sk);
460 sock_rps_record_flow(sk);
462 sock_poll_wait(file, sk_sleep(sk), wait);
464 state = sk_state_load(sk);
465 if (state == TCP_LISTEN)
466 return inet_csk_listen_poll(sk);
468 /* Socket is not locked. We are protected from async events
469 * by poll logic and correct handling of state changes
470 * made by other threads is impossible in any case.
476 * POLLHUP is certainly not done right. But poll() doesn't
477 * have a notion of HUP in just one direction, and for a
478 * socket the read side is more interesting.
480 * Some poll() documentation says that POLLHUP is incompatible
481 * with the POLLOUT/POLLWR flags, so somebody should check this
482 * all. But careful, it tends to be safer to return too many
483 * bits than too few, and you can easily break real applications
484 * if you don't tell them that something has hung up!
488 * Check number 1. POLLHUP is _UNMASKABLE_ event (see UNIX98 and
489 * our fs/select.c). It means that after we received EOF,
490 * poll always returns immediately, making impossible poll() on write()
491 * in state CLOSE_WAIT. One solution is evident --- to set POLLHUP
492 * if and only if shutdown has been made in both directions.
493 * Actually, it is interesting to look how Solaris and DUX
494 * solve this dilemma. I would prefer, if POLLHUP were maskable,
495 * then we could set it on SND_SHUTDOWN. BTW examples given
496 * in Stevens' books assume exactly this behaviour, it explains
497 * why POLLHUP is incompatible with POLLOUT. --ANK
499 * NOTE. Check for TCP_CLOSE is added. The goal is to prevent
500 * blocking on fresh not-connected or disconnected socket. --ANK
502 if (sk->sk_shutdown == SHUTDOWN_MASK || state == TCP_CLOSE)
504 if (sk->sk_shutdown & RCV_SHUTDOWN)
505 mask |= POLLIN | POLLRDNORM | POLLRDHUP;
507 /* Connected or passive Fast Open socket? */
508 if (state != TCP_SYN_SENT &&
509 (state != TCP_SYN_RECV || tp->fastopen_rsk)) {
510 int target = sock_rcvlowat(sk, 0, INT_MAX);
512 if (tp->urg_seq == tp->copied_seq &&
513 !sock_flag(sk, SOCK_URGINLINE) &&
517 if (tp->rcv_nxt - tp->copied_seq >= target)
518 mask |= POLLIN | POLLRDNORM;
520 if (!(sk->sk_shutdown & SEND_SHUTDOWN)) {
521 if (sk_stream_is_writeable(sk)) {
522 mask |= POLLOUT | POLLWRNORM;
523 } else { /* send SIGIO later */
524 sk_set_bit(SOCKWQ_ASYNC_NOSPACE, sk);
525 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
527 /* Race breaker. If space is freed after
528 * wspace test but before the flags are set,
529 * IO signal will be lost. Memory barrier
530 * pairs with the input side.
532 smp_mb__after_atomic();
533 if (sk_stream_is_writeable(sk))
534 mask |= POLLOUT | POLLWRNORM;
537 mask |= POLLOUT | POLLWRNORM;
539 if (tp->urg_data & TCP_URG_VALID)
542 /* This barrier is coupled with smp_wmb() in tcp_reset() */
544 if (sk->sk_err || !skb_queue_empty(&sk->sk_error_queue))
549 EXPORT_SYMBOL(tcp_poll);
551 int tcp_ioctl(struct sock *sk, int cmd, unsigned long arg)
553 struct tcp_sock *tp = tcp_sk(sk);
559 if (sk->sk_state == TCP_LISTEN)
562 slow = lock_sock_fast(sk);
564 unlock_sock_fast(sk, slow);
567 answ = tp->urg_data && tp->urg_seq == tp->copied_seq;
570 if (sk->sk_state == TCP_LISTEN)
573 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
576 answ = tp->write_seq - tp->snd_una;
579 if (sk->sk_state == TCP_LISTEN)
582 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
585 answ = tp->write_seq - tp->snd_nxt;
591 return put_user(answ, (int __user *)arg);
593 EXPORT_SYMBOL(tcp_ioctl);
595 static inline void tcp_mark_push(struct tcp_sock *tp, struct sk_buff *skb)
597 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
598 tp->pushed_seq = tp->write_seq;
601 static inline bool forced_push(const struct tcp_sock *tp)
603 return after(tp->write_seq, tp->pushed_seq + (tp->max_window >> 1));
606 static void skb_entail(struct sock *sk, struct sk_buff *skb)
608 struct tcp_sock *tp = tcp_sk(sk);
609 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
612 tcb->seq = tcb->end_seq = tp->write_seq;
613 tcb->tcp_flags = TCPHDR_ACK;
615 __skb_header_release(skb);
616 tcp_add_write_queue_tail(sk, skb);
617 sk->sk_wmem_queued += skb->truesize;
618 sk_mem_charge(sk, skb->truesize);
619 if (tp->nonagle & TCP_NAGLE_PUSH)
620 tp->nonagle &= ~TCP_NAGLE_PUSH;
622 tcp_slow_start_after_idle_check(sk);
625 static inline void tcp_mark_urg(struct tcp_sock *tp, int flags)
628 tp->snd_up = tp->write_seq;
631 /* If a not yet filled skb is pushed, do not send it if
632 * we have data packets in Qdisc or NIC queues :
633 * Because TX completion will happen shortly, it gives a chance
634 * to coalesce future sendmsg() payload into this skb, without
635 * need for a timer, and with no latency trade off.
636 * As packets containing data payload have a bigger truesize
637 * than pure acks (dataless) packets, the last checks prevent
638 * autocorking if we only have an ACK in Qdisc/NIC queues,
639 * or if TX completion was delayed after we processed ACK packet.
641 static bool tcp_should_autocork(struct sock *sk, struct sk_buff *skb,
644 return skb->len < size_goal &&
645 sysctl_tcp_autocorking &&
646 skb != tcp_write_queue_head(sk) &&
647 atomic_read(&sk->sk_wmem_alloc) > skb->truesize;
650 static void tcp_push(struct sock *sk, int flags, int mss_now,
651 int nonagle, int size_goal)
653 struct tcp_sock *tp = tcp_sk(sk);
656 if (!tcp_send_head(sk))
659 skb = tcp_write_queue_tail(sk);
660 if (!(flags & MSG_MORE) || forced_push(tp))
661 tcp_mark_push(tp, skb);
663 tcp_mark_urg(tp, flags);
665 if (tcp_should_autocork(sk, skb, size_goal)) {
667 /* avoid atomic op if TSQ_THROTTLED bit is already set */
668 if (!test_bit(TSQ_THROTTLED, &tp->tsq_flags)) {
669 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPAUTOCORKING);
670 set_bit(TSQ_THROTTLED, &tp->tsq_flags);
672 /* It is possible TX completion already happened
673 * before we set TSQ_THROTTLED.
675 if (atomic_read(&sk->sk_wmem_alloc) > skb->truesize)
679 if (flags & MSG_MORE)
680 nonagle = TCP_NAGLE_CORK;
682 __tcp_push_pending_frames(sk, mss_now, nonagle);
685 static int tcp_splice_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb,
686 unsigned int offset, size_t len)
688 struct tcp_splice_state *tss = rd_desc->arg.data;
691 ret = skb_splice_bits(skb, skb->sk, offset, tss->pipe,
692 min(rd_desc->count, len), tss->flags,
695 rd_desc->count -= ret;
699 static int __tcp_splice_read(struct sock *sk, struct tcp_splice_state *tss)
701 /* Store TCP splice context information in read_descriptor_t. */
702 read_descriptor_t rd_desc = {
707 return tcp_read_sock(sk, &rd_desc, tcp_splice_data_recv);
711 * tcp_splice_read - splice data from TCP socket to a pipe
712 * @sock: socket to splice from
713 * @ppos: position (not valid)
714 * @pipe: pipe to splice to
715 * @len: number of bytes to splice
716 * @flags: splice modifier flags
719 * Will read pages from given socket and fill them into a pipe.
722 ssize_t tcp_splice_read(struct socket *sock, loff_t *ppos,
723 struct pipe_inode_info *pipe, size_t len,
726 struct sock *sk = sock->sk;
727 struct tcp_splice_state tss = {
736 sock_rps_record_flow(sk);
738 * We can't seek on a socket input
747 timeo = sock_rcvtimeo(sk, sock->file->f_flags & O_NONBLOCK);
749 ret = __tcp_splice_read(sk, &tss);
755 if (sock_flag(sk, SOCK_DONE))
758 ret = sock_error(sk);
761 if (sk->sk_shutdown & RCV_SHUTDOWN)
763 if (sk->sk_state == TCP_CLOSE) {
765 * This occurs when user tries to read
766 * from never connected socket.
768 if (!sock_flag(sk, SOCK_DONE))
776 sk_wait_data(sk, &timeo, NULL);
777 if (signal_pending(current)) {
778 ret = sock_intr_errno(timeo);
791 if (sk->sk_err || sk->sk_state == TCP_CLOSE ||
792 (sk->sk_shutdown & RCV_SHUTDOWN) ||
793 signal_pending(current))
804 EXPORT_SYMBOL(tcp_splice_read);
806 struct sk_buff *sk_stream_alloc_skb(struct sock *sk, int size, gfp_t gfp,
811 /* The TCP header must be at least 32-bit aligned. */
812 size = ALIGN(size, 4);
814 if (unlikely(tcp_under_memory_pressure(sk)))
815 sk_mem_reclaim_partial(sk);
817 skb = alloc_skb_fclone(size + sk->sk_prot->max_header, gfp);
821 if (force_schedule) {
822 mem_scheduled = true;
823 sk_forced_mem_schedule(sk, skb->truesize);
825 mem_scheduled = sk_wmem_schedule(sk, skb->truesize);
827 if (likely(mem_scheduled)) {
828 skb_reserve(skb, sk->sk_prot->max_header);
830 * Make sure that we have exactly size bytes
831 * available to the caller, no more, no less.
833 skb->reserved_tailroom = skb->end - skb->tail - size;
838 sk->sk_prot->enter_memory_pressure(sk);
839 sk_stream_moderate_sndbuf(sk);
844 static unsigned int tcp_xmit_size_goal(struct sock *sk, u32 mss_now,
847 struct tcp_sock *tp = tcp_sk(sk);
848 u32 new_size_goal, size_goal;
850 if (!large_allowed || !sk_can_gso(sk))
853 /* Note : tcp_tso_autosize() will eventually split this later */
854 new_size_goal = sk->sk_gso_max_size - 1 - MAX_TCP_HEADER;
855 new_size_goal = tcp_bound_to_half_wnd(tp, new_size_goal);
857 /* We try hard to avoid divides here */
858 size_goal = tp->gso_segs * mss_now;
859 if (unlikely(new_size_goal < size_goal ||
860 new_size_goal >= size_goal + mss_now)) {
861 tp->gso_segs = min_t(u16, new_size_goal / mss_now,
862 sk->sk_gso_max_segs);
863 size_goal = tp->gso_segs * mss_now;
866 return max(size_goal, mss_now);
869 static int tcp_send_mss(struct sock *sk, int *size_goal, int flags)
873 mss_now = tcp_current_mss(sk);
874 *size_goal = tcp_xmit_size_goal(sk, mss_now, !(flags & MSG_OOB));
879 static ssize_t do_tcp_sendpages(struct sock *sk, struct page *page, int offset,
880 size_t size, int flags)
882 struct tcp_sock *tp = tcp_sk(sk);
883 int mss_now, size_goal;
886 long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
888 /* Wait for a connection to finish. One exception is TCP Fast Open
889 * (passive side) where data is allowed to be sent before a connection
890 * is fully established.
892 if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) &&
893 !tcp_passive_fastopen(sk)) {
894 err = sk_stream_wait_connect(sk, &timeo);
899 sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
901 mss_now = tcp_send_mss(sk, &size_goal, flags);
905 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
909 struct sk_buff *skb = tcp_write_queue_tail(sk);
913 if (!tcp_send_head(sk) || (copy = size_goal - skb->len) <= 0 ||
914 !tcp_skb_can_collapse_to(skb)) {
916 if (!sk_stream_memory_free(sk))
917 goto wait_for_sndbuf;
919 skb = sk_stream_alloc_skb(sk, 0, sk->sk_allocation,
920 skb_queue_empty(&sk->sk_write_queue));
922 goto wait_for_memory;
931 i = skb_shinfo(skb)->nr_frags;
932 can_coalesce = skb_can_coalesce(skb, i, page, offset);
933 if (!can_coalesce && i >= sysctl_max_skb_frags) {
934 tcp_mark_push(tp, skb);
937 if (!sk_wmem_schedule(sk, copy))
938 goto wait_for_memory;
941 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
944 skb_fill_page_desc(skb, i, page, offset, copy);
946 skb_shinfo(skb)->tx_flags |= SKBTX_SHARED_FRAG;
949 skb->data_len += copy;
950 skb->truesize += copy;
951 sk->sk_wmem_queued += copy;
952 sk_mem_charge(sk, copy);
953 skb->ip_summed = CHECKSUM_PARTIAL;
954 tp->write_seq += copy;
955 TCP_SKB_CB(skb)->end_seq += copy;
956 tcp_skb_pcount_set(skb, 0);
959 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
965 tcp_tx_timestamp(sk, sk->sk_tsflags, skb);
969 if (skb->len < size_goal || (flags & MSG_OOB))
972 if (forced_push(tp)) {
973 tcp_mark_push(tp, skb);
974 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
975 } else if (skb == tcp_send_head(sk))
976 tcp_push_one(sk, mss_now);
980 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
982 tcp_push(sk, flags & ~MSG_MORE, mss_now,
983 TCP_NAGLE_PUSH, size_goal);
985 err = sk_stream_wait_memory(sk, &timeo);
989 mss_now = tcp_send_mss(sk, &size_goal, flags);
993 if (copied && !(flags & MSG_SENDPAGE_NOTLAST))
994 tcp_push(sk, flags, mss_now, tp->nonagle, size_goal);
1001 /* make sure we wake any epoll edge trigger waiter */
1002 if (unlikely(skb_queue_len(&sk->sk_write_queue) == 0 && err == -EAGAIN))
1003 sk->sk_write_space(sk);
1004 return sk_stream_error(sk, flags, err);
1007 int tcp_sendpage(struct sock *sk, struct page *page, int offset,
1008 size_t size, int flags)
1012 if (!(sk->sk_route_caps & NETIF_F_SG) ||
1013 !sk_check_csum_caps(sk))
1014 return sock_no_sendpage(sk->sk_socket, page, offset, size,
1019 tcp_rate_check_app_limited(sk); /* is sending application-limited? */
1021 res = do_tcp_sendpages(sk, page, offset, size, flags);
1025 EXPORT_SYMBOL(tcp_sendpage);
1027 /* Do not bother using a page frag for very small frames.
1028 * But use this heuristic only for the first skb in write queue.
1030 * Having no payload in skb->head allows better SACK shifting
1031 * in tcp_shift_skb_data(), reducing sack/rack overhead, because
1032 * write queue has less skbs.
1033 * Each skb can hold up to MAX_SKB_FRAGS * 32Kbytes, or ~0.5 MB.
1034 * This also speeds up tso_fragment(), since it wont fallback
1035 * to tcp_fragment().
1037 static int linear_payload_sz(bool first_skb)
1040 return SKB_WITH_OVERHEAD(2048 - MAX_TCP_HEADER);
1044 static int select_size(const struct sock *sk, bool sg, bool first_skb)
1046 const struct tcp_sock *tp = tcp_sk(sk);
1047 int tmp = tp->mss_cache;
1050 if (sk_can_gso(sk)) {
1051 tmp = linear_payload_sz(first_skb);
1053 int pgbreak = SKB_MAX_HEAD(MAX_TCP_HEADER);
1055 if (tmp >= pgbreak &&
1056 tmp <= pgbreak + (MAX_SKB_FRAGS - 1) * PAGE_SIZE)
1064 void tcp_free_fastopen_req(struct tcp_sock *tp)
1066 if (tp->fastopen_req) {
1067 kfree(tp->fastopen_req);
1068 tp->fastopen_req = NULL;
1072 static int tcp_sendmsg_fastopen(struct sock *sk, struct msghdr *msg,
1073 int *copied, size_t size)
1075 struct tcp_sock *tp = tcp_sk(sk);
1078 if (!(sysctl_tcp_fastopen & TFO_CLIENT_ENABLE))
1080 if (tp->fastopen_req)
1081 return -EALREADY; /* Another Fast Open is in progress */
1083 tp->fastopen_req = kzalloc(sizeof(struct tcp_fastopen_request),
1085 if (unlikely(!tp->fastopen_req))
1087 tp->fastopen_req->data = msg;
1088 tp->fastopen_req->size = size;
1090 flags = (msg->msg_flags & MSG_DONTWAIT) ? O_NONBLOCK : 0;
1091 err = __inet_stream_connect(sk->sk_socket, msg->msg_name,
1092 msg->msg_namelen, flags);
1093 *copied = tp->fastopen_req->copied;
1094 tcp_free_fastopen_req(tp);
1098 int tcp_sendmsg(struct sock *sk, struct msghdr *msg, size_t size)
1100 struct tcp_sock *tp = tcp_sk(sk);
1101 struct sk_buff *skb;
1102 struct sockcm_cookie sockc;
1103 int flags, err, copied = 0;
1104 int mss_now = 0, size_goal, copied_syn = 0;
1105 bool process_backlog = false;
1111 flags = msg->msg_flags;
1112 if (flags & MSG_FASTOPEN) {
1113 err = tcp_sendmsg_fastopen(sk, msg, &copied_syn, size);
1114 if (err == -EINPROGRESS && copied_syn > 0)
1120 timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
1122 tcp_rate_check_app_limited(sk); /* is sending application-limited? */
1124 /* Wait for a connection to finish. One exception is TCP Fast Open
1125 * (passive side) where data is allowed to be sent before a connection
1126 * is fully established.
1128 if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) &&
1129 !tcp_passive_fastopen(sk)) {
1130 err = sk_stream_wait_connect(sk, &timeo);
1135 if (unlikely(tp->repair)) {
1136 if (tp->repair_queue == TCP_RECV_QUEUE) {
1137 copied = tcp_send_rcvq(sk, msg, size);
1142 if (tp->repair_queue == TCP_NO_QUEUE)
1145 /* 'common' sending to sendq */
1148 sockc.tsflags = sk->sk_tsflags;
1149 if (msg->msg_controllen) {
1150 err = sock_cmsg_send(sk, msg, &sockc);
1151 if (unlikely(err)) {
1157 /* This should be in poll */
1158 sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
1160 /* Ok commence sending. */
1164 mss_now = tcp_send_mss(sk, &size_goal, flags);
1167 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
1170 sg = !!(sk->sk_route_caps & NETIF_F_SG);
1172 while (msg_data_left(msg)) {
1174 int max = size_goal;
1176 skb = tcp_write_queue_tail(sk);
1177 if (tcp_send_head(sk)) {
1178 if (skb->ip_summed == CHECKSUM_NONE)
1180 copy = max - skb->len;
1183 if (copy <= 0 || !tcp_skb_can_collapse_to(skb)) {
1187 /* Allocate new segment. If the interface is SG,
1188 * allocate skb fitting to single page.
1190 if (!sk_stream_memory_free(sk))
1191 goto wait_for_sndbuf;
1193 if (process_backlog && sk_flush_backlog(sk)) {
1194 process_backlog = false;
1197 first_skb = skb_queue_empty(&sk->sk_write_queue);
1198 skb = sk_stream_alloc_skb(sk,
1199 select_size(sk, sg, first_skb),
1203 goto wait_for_memory;
1205 process_backlog = true;
1207 * Check whether we can use HW checksum.
1209 if (sk_check_csum_caps(sk))
1210 skb->ip_summed = CHECKSUM_PARTIAL;
1212 skb_entail(sk, skb);
1216 /* All packets are restored as if they have
1217 * already been sent. skb_mstamp isn't set to
1218 * avoid wrong rtt estimation.
1221 TCP_SKB_CB(skb)->sacked |= TCPCB_REPAIRED;
1224 /* Try to append data to the end of skb. */
1225 if (copy > msg_data_left(msg))
1226 copy = msg_data_left(msg);
1228 /* Where to copy to? */
1229 if (skb_availroom(skb) > 0) {
1230 /* We have some space in skb head. Superb! */
1231 copy = min_t(int, copy, skb_availroom(skb));
1232 err = skb_add_data_nocache(sk, skb, &msg->msg_iter, copy);
1237 int i = skb_shinfo(skb)->nr_frags;
1238 struct page_frag *pfrag = sk_page_frag(sk);
1240 if (!sk_page_frag_refill(sk, pfrag))
1241 goto wait_for_memory;
1243 if (!skb_can_coalesce(skb, i, pfrag->page,
1245 if (i == sysctl_max_skb_frags || !sg) {
1246 tcp_mark_push(tp, skb);
1252 copy = min_t(int, copy, pfrag->size - pfrag->offset);
1254 if (!sk_wmem_schedule(sk, copy))
1255 goto wait_for_memory;
1257 err = skb_copy_to_page_nocache(sk, &msg->msg_iter, skb,
1264 /* Update the skb. */
1266 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1268 skb_fill_page_desc(skb, i, pfrag->page,
1269 pfrag->offset, copy);
1270 get_page(pfrag->page);
1272 pfrag->offset += copy;
1276 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
1278 tp->write_seq += copy;
1279 TCP_SKB_CB(skb)->end_seq += copy;
1280 tcp_skb_pcount_set(skb, 0);
1283 if (!msg_data_left(msg)) {
1284 tcp_tx_timestamp(sk, sockc.tsflags, skb);
1285 if (unlikely(flags & MSG_EOR))
1286 TCP_SKB_CB(skb)->eor = 1;
1290 if (skb->len < max || (flags & MSG_OOB) || unlikely(tp->repair))
1293 if (forced_push(tp)) {
1294 tcp_mark_push(tp, skb);
1295 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
1296 } else if (skb == tcp_send_head(sk))
1297 tcp_push_one(sk, mss_now);
1301 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1304 tcp_push(sk, flags & ~MSG_MORE, mss_now,
1305 TCP_NAGLE_PUSH, size_goal);
1307 err = sk_stream_wait_memory(sk, &timeo);
1311 mss_now = tcp_send_mss(sk, &size_goal, flags);
1316 tcp_push(sk, flags, mss_now, tp->nonagle, size_goal);
1319 return copied + copied_syn;
1323 tcp_unlink_write_queue(skb, sk);
1324 /* It is the one place in all of TCP, except connection
1325 * reset, where we can be unlinking the send_head.
1327 tcp_check_send_head(sk, skb);
1328 sk_wmem_free_skb(sk, skb);
1332 if (copied + copied_syn)
1335 err = sk_stream_error(sk, flags, err);
1336 /* make sure we wake any epoll edge trigger waiter */
1337 if (unlikely(skb_queue_len(&sk->sk_write_queue) == 0 && err == -EAGAIN))
1338 sk->sk_write_space(sk);
1342 EXPORT_SYMBOL(tcp_sendmsg);
1345 * Handle reading urgent data. BSD has very simple semantics for
1346 * this, no blocking and very strange errors 8)
1349 static int tcp_recv_urg(struct sock *sk, struct msghdr *msg, int len, int flags)
1351 struct tcp_sock *tp = tcp_sk(sk);
1353 /* No URG data to read. */
1354 if (sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data ||
1355 tp->urg_data == TCP_URG_READ)
1356 return -EINVAL; /* Yes this is right ! */
1358 if (sk->sk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DONE))
1361 if (tp->urg_data & TCP_URG_VALID) {
1363 char c = tp->urg_data;
1365 if (!(flags & MSG_PEEK))
1366 tp->urg_data = TCP_URG_READ;
1368 /* Read urgent data. */
1369 msg->msg_flags |= MSG_OOB;
1372 if (!(flags & MSG_TRUNC))
1373 err = memcpy_to_msg(msg, &c, 1);
1376 msg->msg_flags |= MSG_TRUNC;
1378 return err ? -EFAULT : len;
1381 if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
1384 /* Fixed the recv(..., MSG_OOB) behaviour. BSD docs and
1385 * the available implementations agree in this case:
1386 * this call should never block, independent of the
1387 * blocking state of the socket.
1388 * Mike <pall@rz.uni-karlsruhe.de>
1393 static int tcp_peek_sndq(struct sock *sk, struct msghdr *msg, int len)
1395 struct sk_buff *skb;
1396 int copied = 0, err = 0;
1398 /* XXX -- need to support SO_PEEK_OFF */
1400 skb_queue_walk(&sk->sk_write_queue, skb) {
1401 err = skb_copy_datagram_msg(skb, 0, msg, skb->len);
1408 return err ?: copied;
1411 /* Clean up the receive buffer for full frames taken by the user,
1412 * then send an ACK if necessary. COPIED is the number of bytes
1413 * tcp_recvmsg has given to the user so far, it speeds up the
1414 * calculation of whether or not we must ACK for the sake of
1417 static void tcp_cleanup_rbuf(struct sock *sk, int copied)
1419 struct tcp_sock *tp = tcp_sk(sk);
1420 bool time_to_ack = false;
1422 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
1424 WARN(skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq),
1425 "cleanup rbuf bug: copied %X seq %X rcvnxt %X\n",
1426 tp->copied_seq, TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt);
1428 if (inet_csk_ack_scheduled(sk)) {
1429 const struct inet_connection_sock *icsk = inet_csk(sk);
1430 /* Delayed ACKs frequently hit locked sockets during bulk
1432 if (icsk->icsk_ack.blocked ||
1433 /* Once-per-two-segments ACK was not sent by tcp_input.c */
1434 tp->rcv_nxt - tp->rcv_wup > icsk->icsk_ack.rcv_mss ||
1436 * If this read emptied read buffer, we send ACK, if
1437 * connection is not bidirectional, user drained
1438 * receive buffer and there was a small segment
1442 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED2) ||
1443 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED) &&
1444 !icsk->icsk_ack.pingpong)) &&
1445 !atomic_read(&sk->sk_rmem_alloc)))
1449 /* We send an ACK if we can now advertise a non-zero window
1450 * which has been raised "significantly".
1452 * Even if window raised up to infinity, do not send window open ACK
1453 * in states, where we will not receive more. It is useless.
1455 if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) {
1456 __u32 rcv_window_now = tcp_receive_window(tp);
1458 /* Optimize, __tcp_select_window() is not cheap. */
1459 if (2*rcv_window_now <= tp->window_clamp) {
1460 __u32 new_window = __tcp_select_window(sk);
1462 /* Send ACK now, if this read freed lots of space
1463 * in our buffer. Certainly, new_window is new window.
1464 * We can advertise it now, if it is not less than current one.
1465 * "Lots" means "at least twice" here.
1467 if (new_window && new_window >= 2 * rcv_window_now)
1475 static void tcp_prequeue_process(struct sock *sk)
1477 struct sk_buff *skb;
1478 struct tcp_sock *tp = tcp_sk(sk);
1480 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPPREQUEUED);
1482 while ((skb = __skb_dequeue(&tp->ucopy.prequeue)) != NULL)
1483 sk_backlog_rcv(sk, skb);
1485 /* Clear memory counter. */
1486 tp->ucopy.memory = 0;
1489 static struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off)
1491 struct sk_buff *skb;
1494 while ((skb = skb_peek(&sk->sk_receive_queue)) != NULL) {
1495 offset = seq - TCP_SKB_CB(skb)->seq;
1496 if (unlikely(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
1497 pr_err_once("%s: found a SYN, please report !\n", __func__);
1500 if (offset < skb->len || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)) {
1504 /* This looks weird, but this can happen if TCP collapsing
1505 * splitted a fat GRO packet, while we released socket lock
1506 * in skb_splice_bits()
1508 sk_eat_skb(sk, skb);
1514 * This routine provides an alternative to tcp_recvmsg() for routines
1515 * that would like to handle copying from skbuffs directly in 'sendfile'
1518 * - It is assumed that the socket was locked by the caller.
1519 * - The routine does not block.
1520 * - At present, there is no support for reading OOB data
1521 * or for 'peeking' the socket using this routine
1522 * (although both would be easy to implement).
1524 int tcp_read_sock(struct sock *sk, read_descriptor_t *desc,
1525 sk_read_actor_t recv_actor)
1527 struct sk_buff *skb;
1528 struct tcp_sock *tp = tcp_sk(sk);
1529 u32 seq = tp->copied_seq;
1533 if (sk->sk_state == TCP_LISTEN)
1535 while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1536 if (offset < skb->len) {
1540 len = skb->len - offset;
1541 /* Stop reading if we hit a patch of urgent data */
1543 u32 urg_offset = tp->urg_seq - seq;
1544 if (urg_offset < len)
1549 used = recv_actor(desc, skb, offset, len);
1554 } else if (used <= len) {
1559 /* If recv_actor drops the lock (e.g. TCP splice
1560 * receive) the skb pointer might be invalid when
1561 * getting here: tcp_collapse might have deleted it
1562 * while aggregating skbs from the socket queue.
1564 skb = tcp_recv_skb(sk, seq - 1, &offset);
1567 /* TCP coalescing might have appended data to the skb.
1568 * Try to splice more frags
1570 if (offset + 1 != skb->len)
1573 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) {
1574 sk_eat_skb(sk, skb);
1578 sk_eat_skb(sk, skb);
1581 tp->copied_seq = seq;
1583 tp->copied_seq = seq;
1585 tcp_rcv_space_adjust(sk);
1587 /* Clean up data we have read: This will do ACK frames. */
1589 tcp_recv_skb(sk, seq, &offset);
1590 tcp_cleanup_rbuf(sk, copied);
1594 EXPORT_SYMBOL(tcp_read_sock);
1596 int tcp_peek_len(struct socket *sock)
1598 return tcp_inq(sock->sk);
1600 EXPORT_SYMBOL(tcp_peek_len);
1603 * This routine copies from a sock struct into the user buffer.
1605 * Technical note: in 2.3 we work on _locked_ socket, so that
1606 * tricks with *seq access order and skb->users are not required.
1607 * Probably, code can be easily improved even more.
1610 int tcp_recvmsg(struct sock *sk, struct msghdr *msg, size_t len, int nonblock,
1611 int flags, int *addr_len)
1613 struct tcp_sock *tp = tcp_sk(sk);
1619 int target; /* Read at least this many bytes */
1621 struct task_struct *user_recv = NULL;
1622 struct sk_buff *skb, *last;
1625 if (unlikely(flags & MSG_ERRQUEUE))
1626 return inet_recv_error(sk, msg, len, addr_len);
1628 if (sk_can_busy_loop(sk) && skb_queue_empty(&sk->sk_receive_queue) &&
1629 (sk->sk_state == TCP_ESTABLISHED))
1630 sk_busy_loop(sk, nonblock);
1635 if (sk->sk_state == TCP_LISTEN)
1638 timeo = sock_rcvtimeo(sk, nonblock);
1640 /* Urgent data needs to be handled specially. */
1641 if (flags & MSG_OOB)
1644 if (unlikely(tp->repair)) {
1646 if (!(flags & MSG_PEEK))
1649 if (tp->repair_queue == TCP_SEND_QUEUE)
1653 if (tp->repair_queue == TCP_NO_QUEUE)
1656 /* 'common' recv queue MSG_PEEK-ing */
1659 seq = &tp->copied_seq;
1660 if (flags & MSG_PEEK) {
1661 peek_seq = tp->copied_seq;
1665 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
1670 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
1671 if (tp->urg_data && tp->urg_seq == *seq) {
1674 if (signal_pending(current)) {
1675 copied = timeo ? sock_intr_errno(timeo) : -EAGAIN;
1680 /* Next get a buffer. */
1682 last = skb_peek_tail(&sk->sk_receive_queue);
1683 skb_queue_walk(&sk->sk_receive_queue, skb) {
1685 /* Now that we have two receive queues this
1688 if (WARN(before(*seq, TCP_SKB_CB(skb)->seq),
1689 "recvmsg bug: copied %X seq %X rcvnxt %X fl %X\n",
1690 *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt,
1694 offset = *seq - TCP_SKB_CB(skb)->seq;
1695 if (unlikely(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
1696 pr_err_once("%s: found a SYN, please report !\n", __func__);
1699 if (offset < skb->len)
1701 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1703 WARN(!(flags & MSG_PEEK),
1704 "recvmsg bug 2: copied %X seq %X rcvnxt %X fl %X\n",
1705 *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt, flags);
1708 /* Well, if we have backlog, try to process it now yet. */
1710 if (copied >= target && !sk->sk_backlog.tail)
1715 sk->sk_state == TCP_CLOSE ||
1716 (sk->sk_shutdown & RCV_SHUTDOWN) ||
1718 signal_pending(current))
1721 if (sock_flag(sk, SOCK_DONE))
1725 copied = sock_error(sk);
1729 if (sk->sk_shutdown & RCV_SHUTDOWN)
1732 if (sk->sk_state == TCP_CLOSE) {
1733 if (!sock_flag(sk, SOCK_DONE)) {
1734 /* This occurs when user tries to read
1735 * from never connected socket.
1748 if (signal_pending(current)) {
1749 copied = sock_intr_errno(timeo);
1754 tcp_cleanup_rbuf(sk, copied);
1756 if (!sysctl_tcp_low_latency && tp->ucopy.task == user_recv) {
1757 /* Install new reader */
1758 if (!user_recv && !(flags & (MSG_TRUNC | MSG_PEEK))) {
1759 user_recv = current;
1760 tp->ucopy.task = user_recv;
1761 tp->ucopy.msg = msg;
1764 tp->ucopy.len = len;
1766 WARN_ON(tp->copied_seq != tp->rcv_nxt &&
1767 !(flags & (MSG_PEEK | MSG_TRUNC)));
1769 /* Ugly... If prequeue is not empty, we have to
1770 * process it before releasing socket, otherwise
1771 * order will be broken at second iteration.
1772 * More elegant solution is required!!!
1774 * Look: we have the following (pseudo)queues:
1776 * 1. packets in flight
1781 * Each queue can be processed only if the next ones
1782 * are empty. At this point we have empty receive_queue.
1783 * But prequeue _can_ be not empty after 2nd iteration,
1784 * when we jumped to start of loop because backlog
1785 * processing added something to receive_queue.
1786 * We cannot release_sock(), because backlog contains
1787 * packets arrived _after_ prequeued ones.
1789 * Shortly, algorithm is clear --- to process all
1790 * the queues in order. We could make it more directly,
1791 * requeueing packets from backlog to prequeue, if
1792 * is not empty. It is more elegant, but eats cycles,
1795 if (!skb_queue_empty(&tp->ucopy.prequeue))
1798 /* __ Set realtime policy in scheduler __ */
1801 if (copied >= target) {
1802 /* Do not sleep, just process backlog. */
1806 sk_wait_data(sk, &timeo, last);
1812 /* __ Restore normal policy in scheduler __ */
1814 chunk = len - tp->ucopy.len;
1816 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMBACKLOG, chunk);
1821 if (tp->rcv_nxt == tp->copied_seq &&
1822 !skb_queue_empty(&tp->ucopy.prequeue)) {
1824 tcp_prequeue_process(sk);
1826 chunk = len - tp->ucopy.len;
1828 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1834 if ((flags & MSG_PEEK) &&
1835 (peek_seq - copied - urg_hole != tp->copied_seq)) {
1836 net_dbg_ratelimited("TCP(%s:%d): Application bug, race in MSG_PEEK\n",
1838 task_pid_nr(current));
1839 peek_seq = tp->copied_seq;
1844 /* Ok so how much can we use? */
1845 used = skb->len - offset;
1849 /* Do we have urgent data here? */
1851 u32 urg_offset = tp->urg_seq - *seq;
1852 if (urg_offset < used) {
1854 if (!sock_flag(sk, SOCK_URGINLINE)) {
1867 if (!(flags & MSG_TRUNC)) {
1868 err = skb_copy_datagram_msg(skb, offset, msg, used);
1870 /* Exception. Bailout! */
1881 tcp_rcv_space_adjust(sk);
1884 if (tp->urg_data && after(tp->copied_seq, tp->urg_seq)) {
1886 tcp_fast_path_check(sk);
1888 if (used + offset < skb->len)
1891 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1893 if (!(flags & MSG_PEEK))
1894 sk_eat_skb(sk, skb);
1898 /* Process the FIN. */
1900 if (!(flags & MSG_PEEK))
1901 sk_eat_skb(sk, skb);
1906 if (!skb_queue_empty(&tp->ucopy.prequeue)) {
1909 tp->ucopy.len = copied > 0 ? len : 0;
1911 tcp_prequeue_process(sk);
1913 if (copied > 0 && (chunk = len - tp->ucopy.len) != 0) {
1914 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1920 tp->ucopy.task = NULL;
1924 /* According to UNIX98, msg_name/msg_namelen are ignored
1925 * on connected socket. I was just happy when found this 8) --ANK
1928 /* Clean up data we have read: This will do ACK frames. */
1929 tcp_cleanup_rbuf(sk, copied);
1939 err = tcp_recv_urg(sk, msg, len, flags);
1943 err = tcp_peek_sndq(sk, msg, len);
1946 EXPORT_SYMBOL(tcp_recvmsg);
1948 void tcp_set_state(struct sock *sk, int state)
1950 int oldstate = sk->sk_state;
1953 case TCP_ESTABLISHED:
1954 if (oldstate != TCP_ESTABLISHED)
1955 TCP_INC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
1959 if (oldstate == TCP_CLOSE_WAIT || oldstate == TCP_ESTABLISHED)
1960 TCP_INC_STATS(sock_net(sk), TCP_MIB_ESTABRESETS);
1962 sk->sk_prot->unhash(sk);
1963 if (inet_csk(sk)->icsk_bind_hash &&
1964 !(sk->sk_userlocks & SOCK_BINDPORT_LOCK))
1968 if (oldstate == TCP_ESTABLISHED)
1969 TCP_DEC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
1972 /* Change state AFTER socket is unhashed to avoid closed
1973 * socket sitting in hash tables.
1975 sk_state_store(sk, state);
1978 SOCK_DEBUG(sk, "TCP sk=%p, State %s -> %s\n", sk, statename[oldstate], statename[state]);
1981 EXPORT_SYMBOL_GPL(tcp_set_state);
1984 * State processing on a close. This implements the state shift for
1985 * sending our FIN frame. Note that we only send a FIN for some
1986 * states. A shutdown() may have already sent the FIN, or we may be
1990 static const unsigned char new_state[16] = {
1991 /* current state: new state: action: */
1992 [0 /* (Invalid) */] = TCP_CLOSE,
1993 [TCP_ESTABLISHED] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1994 [TCP_SYN_SENT] = TCP_CLOSE,
1995 [TCP_SYN_RECV] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1996 [TCP_FIN_WAIT1] = TCP_FIN_WAIT1,
1997 [TCP_FIN_WAIT2] = TCP_FIN_WAIT2,
1998 [TCP_TIME_WAIT] = TCP_CLOSE,
1999 [TCP_CLOSE] = TCP_CLOSE,
2000 [TCP_CLOSE_WAIT] = TCP_LAST_ACK | TCP_ACTION_FIN,
2001 [TCP_LAST_ACK] = TCP_LAST_ACK,
2002 [TCP_LISTEN] = TCP_CLOSE,
2003 [TCP_CLOSING] = TCP_CLOSING,
2004 [TCP_NEW_SYN_RECV] = TCP_CLOSE, /* should not happen ! */
2007 static int tcp_close_state(struct sock *sk)
2009 int next = (int)new_state[sk->sk_state];
2010 int ns = next & TCP_STATE_MASK;
2012 tcp_set_state(sk, ns);
2014 return next & TCP_ACTION_FIN;
2018 * Shutdown the sending side of a connection. Much like close except
2019 * that we don't receive shut down or sock_set_flag(sk, SOCK_DEAD).
2022 void tcp_shutdown(struct sock *sk, int how)
2024 /* We need to grab some memory, and put together a FIN,
2025 * and then put it into the queue to be sent.
2026 * Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
2028 if (!(how & SEND_SHUTDOWN))
2031 /* If we've already sent a FIN, or it's a closed state, skip this. */
2032 if ((1 << sk->sk_state) &
2033 (TCPF_ESTABLISHED | TCPF_SYN_SENT |
2034 TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) {
2035 /* Clear out any half completed packets. FIN if needed. */
2036 if (tcp_close_state(sk))
2040 EXPORT_SYMBOL(tcp_shutdown);
2042 bool tcp_check_oom(struct sock *sk, int shift)
2044 bool too_many_orphans, out_of_socket_memory;
2046 too_many_orphans = tcp_too_many_orphans(sk, shift);
2047 out_of_socket_memory = tcp_out_of_memory(sk);
2049 if (too_many_orphans)
2050 net_info_ratelimited("too many orphaned sockets\n");
2051 if (out_of_socket_memory)
2052 net_info_ratelimited("out of memory -- consider tuning tcp_mem\n");
2053 return too_many_orphans || out_of_socket_memory;
2056 void tcp_close(struct sock *sk, long timeout)
2058 struct sk_buff *skb;
2059 int data_was_unread = 0;
2063 sk->sk_shutdown = SHUTDOWN_MASK;
2065 if (sk->sk_state == TCP_LISTEN) {
2066 tcp_set_state(sk, TCP_CLOSE);
2069 inet_csk_listen_stop(sk);
2071 goto adjudge_to_death;
2074 /* We need to flush the recv. buffs. We do this only on the
2075 * descriptor close, not protocol-sourced closes, because the
2076 * reader process may not have drained the data yet!
2078 while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) {
2079 u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq;
2081 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
2083 data_was_unread += len;
2089 /* If socket has been already reset (e.g. in tcp_reset()) - kill it. */
2090 if (sk->sk_state == TCP_CLOSE)
2091 goto adjudge_to_death;
2093 /* As outlined in RFC 2525, section 2.17, we send a RST here because
2094 * data was lost. To witness the awful effects of the old behavior of
2095 * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk
2096 * GET in an FTP client, suspend the process, wait for the client to
2097 * advertise a zero window, then kill -9 the FTP client, wheee...
2098 * Note: timeout is always zero in such a case.
2100 if (unlikely(tcp_sk(sk)->repair)) {
2101 sk->sk_prot->disconnect(sk, 0);
2102 } else if (data_was_unread) {
2103 /* Unread data was tossed, zap the connection. */
2104 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONCLOSE);
2105 tcp_set_state(sk, TCP_CLOSE);
2106 tcp_send_active_reset(sk, sk->sk_allocation);
2107 } else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) {
2108 /* Check zero linger _after_ checking for unread data. */
2109 sk->sk_prot->disconnect(sk, 0);
2110 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
2111 } else if (tcp_close_state(sk)) {
2112 /* We FIN if the application ate all the data before
2113 * zapping the connection.
2116 /* RED-PEN. Formally speaking, we have broken TCP state
2117 * machine. State transitions:
2119 * TCP_ESTABLISHED -> TCP_FIN_WAIT1
2120 * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible)
2121 * TCP_CLOSE_WAIT -> TCP_LAST_ACK
2123 * are legal only when FIN has been sent (i.e. in window),
2124 * rather than queued out of window. Purists blame.
2126 * F.e. "RFC state" is ESTABLISHED,
2127 * if Linux state is FIN-WAIT-1, but FIN is still not sent.
2129 * The visible declinations are that sometimes
2130 * we enter time-wait state, when it is not required really
2131 * (harmless), do not send active resets, when they are
2132 * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
2133 * they look as CLOSING or LAST_ACK for Linux)
2134 * Probably, I missed some more holelets.
2136 * XXX (TFO) - To start off we don't support SYN+ACK+FIN
2137 * in a single packet! (May consider it later but will
2138 * probably need API support or TCP_CORK SYN-ACK until
2139 * data is written and socket is closed.)
2144 sk_stream_wait_close(sk, timeout);
2147 state = sk->sk_state;
2151 /* It is the last release_sock in its life. It will remove backlog. */
2155 /* Now socket is owned by kernel and we acquire BH lock
2156 to finish close. No need to check for user refs.
2160 WARN_ON(sock_owned_by_user(sk));
2162 percpu_counter_inc(sk->sk_prot->orphan_count);
2164 /* Have we already been destroyed by a softirq or backlog? */
2165 if (state != TCP_CLOSE && sk->sk_state == TCP_CLOSE)
2168 /* This is a (useful) BSD violating of the RFC. There is a
2169 * problem with TCP as specified in that the other end could
2170 * keep a socket open forever with no application left this end.
2171 * We use a 1 minute timeout (about the same as BSD) then kill
2172 * our end. If they send after that then tough - BUT: long enough
2173 * that we won't make the old 4*rto = almost no time - whoops
2176 * Nope, it was not mistake. It is really desired behaviour
2177 * f.e. on http servers, when such sockets are useless, but
2178 * consume significant resources. Let's do it with special
2179 * linger2 option. --ANK
2182 if (sk->sk_state == TCP_FIN_WAIT2) {
2183 struct tcp_sock *tp = tcp_sk(sk);
2184 if (tp->linger2 < 0) {
2185 tcp_set_state(sk, TCP_CLOSE);
2186 tcp_send_active_reset(sk, GFP_ATOMIC);
2187 __NET_INC_STATS(sock_net(sk),
2188 LINUX_MIB_TCPABORTONLINGER);
2190 const int tmo = tcp_fin_time(sk);
2192 if (tmo > TCP_TIMEWAIT_LEN) {
2193 inet_csk_reset_keepalive_timer(sk,
2194 tmo - TCP_TIMEWAIT_LEN);
2196 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
2201 if (sk->sk_state != TCP_CLOSE) {
2203 if (tcp_check_oom(sk, 0)) {
2204 tcp_set_state(sk, TCP_CLOSE);
2205 tcp_send_active_reset(sk, GFP_ATOMIC);
2206 __NET_INC_STATS(sock_net(sk),
2207 LINUX_MIB_TCPABORTONMEMORY);
2211 if (sk->sk_state == TCP_CLOSE) {
2212 struct request_sock *req = tcp_sk(sk)->fastopen_rsk;
2213 /* We could get here with a non-NULL req if the socket is
2214 * aborted (e.g., closed with unread data) before 3WHS
2218 reqsk_fastopen_remove(sk, req, false);
2219 inet_csk_destroy_sock(sk);
2221 /* Otherwise, socket is reprieved until protocol close. */
2228 EXPORT_SYMBOL(tcp_close);
2230 /* These states need RST on ABORT according to RFC793 */
2232 static inline bool tcp_need_reset(int state)
2234 return (1 << state) &
2235 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 |
2236 TCPF_FIN_WAIT2 | TCPF_SYN_RECV);
2239 int tcp_disconnect(struct sock *sk, int flags)
2241 struct inet_sock *inet = inet_sk(sk);
2242 struct inet_connection_sock *icsk = inet_csk(sk);
2243 struct tcp_sock *tp = tcp_sk(sk);
2245 int old_state = sk->sk_state;
2247 if (old_state != TCP_CLOSE)
2248 tcp_set_state(sk, TCP_CLOSE);
2250 /* ABORT function of RFC793 */
2251 if (old_state == TCP_LISTEN) {
2252 inet_csk_listen_stop(sk);
2253 } else if (unlikely(tp->repair)) {
2254 sk->sk_err = ECONNABORTED;
2255 } else if (tcp_need_reset(old_state) ||
2256 (tp->snd_nxt != tp->write_seq &&
2257 (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK))) {
2258 /* The last check adjusts for discrepancy of Linux wrt. RFC
2261 tcp_send_active_reset(sk, gfp_any());
2262 sk->sk_err = ECONNRESET;
2263 } else if (old_state == TCP_SYN_SENT)
2264 sk->sk_err = ECONNRESET;
2266 tcp_clear_xmit_timers(sk);
2267 __skb_queue_purge(&sk->sk_receive_queue);
2268 tcp_write_queue_purge(sk);
2269 skb_rbtree_purge(&tp->out_of_order_queue);
2271 inet->inet_dport = 0;
2273 if (!(sk->sk_userlocks & SOCK_BINDADDR_LOCK))
2274 inet_reset_saddr(sk);
2276 sk->sk_shutdown = 0;
2277 sock_reset_flag(sk, SOCK_DONE);
2279 tp->write_seq += tp->max_window + 2;
2280 if (tp->write_seq == 0)
2282 icsk->icsk_backoff = 0;
2284 icsk->icsk_probes_out = 0;
2285 tp->packets_out = 0;
2286 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
2287 tp->snd_cwnd_cnt = 0;
2288 tp->window_clamp = 0;
2289 tcp_set_ca_state(sk, TCP_CA_Open);
2290 tcp_clear_retrans(tp);
2291 inet_csk_delack_init(sk);
2292 tcp_init_send_head(sk);
2293 memset(&tp->rx_opt, 0, sizeof(tp->rx_opt));
2296 WARN_ON(inet->inet_num && !icsk->icsk_bind_hash);
2298 sk->sk_error_report(sk);
2301 EXPORT_SYMBOL(tcp_disconnect);
2303 static inline bool tcp_can_repair_sock(const struct sock *sk)
2305 return ns_capable(sock_net(sk)->user_ns, CAP_NET_ADMIN) &&
2306 ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_ESTABLISHED));
2309 static int tcp_repair_set_window(struct tcp_sock *tp, char __user *optbuf, int len)
2311 struct tcp_repair_window opt;
2316 if (len != sizeof(opt))
2319 if (copy_from_user(&opt, optbuf, sizeof(opt)))
2322 if (opt.max_window < opt.snd_wnd)
2325 if (after(opt.snd_wl1, tp->rcv_nxt + opt.rcv_wnd))
2328 if (after(opt.rcv_wup, tp->rcv_nxt))
2331 tp->snd_wl1 = opt.snd_wl1;
2332 tp->snd_wnd = opt.snd_wnd;
2333 tp->max_window = opt.max_window;
2335 tp->rcv_wnd = opt.rcv_wnd;
2336 tp->rcv_wup = opt.rcv_wup;
2341 static int tcp_repair_options_est(struct tcp_sock *tp,
2342 struct tcp_repair_opt __user *optbuf, unsigned int len)
2344 struct tcp_repair_opt opt;
2346 while (len >= sizeof(opt)) {
2347 if (copy_from_user(&opt, optbuf, sizeof(opt)))
2353 switch (opt.opt_code) {
2355 tp->rx_opt.mss_clamp = opt.opt_val;
2359 u16 snd_wscale = opt.opt_val & 0xFFFF;
2360 u16 rcv_wscale = opt.opt_val >> 16;
2362 if (snd_wscale > 14 || rcv_wscale > 14)
2365 tp->rx_opt.snd_wscale = snd_wscale;
2366 tp->rx_opt.rcv_wscale = rcv_wscale;
2367 tp->rx_opt.wscale_ok = 1;
2370 case TCPOPT_SACK_PERM:
2371 if (opt.opt_val != 0)
2374 tp->rx_opt.sack_ok |= TCP_SACK_SEEN;
2375 if (sysctl_tcp_fack)
2376 tcp_enable_fack(tp);
2378 case TCPOPT_TIMESTAMP:
2379 if (opt.opt_val != 0)
2382 tp->rx_opt.tstamp_ok = 1;
2391 * Socket option code for TCP.
2393 static int do_tcp_setsockopt(struct sock *sk, int level,
2394 int optname, char __user *optval, unsigned int optlen)
2396 struct tcp_sock *tp = tcp_sk(sk);
2397 struct inet_connection_sock *icsk = inet_csk(sk);
2398 struct net *net = sock_net(sk);
2402 /* These are data/string values, all the others are ints */
2404 case TCP_CONGESTION: {
2405 char name[TCP_CA_NAME_MAX];
2410 val = strncpy_from_user(name, optval,
2411 min_t(long, TCP_CA_NAME_MAX-1, optlen));
2417 err = tcp_set_congestion_control(sk, name);
2426 if (optlen < sizeof(int))
2429 if (get_user(val, (int __user *)optval))
2436 /* Values greater than interface MTU won't take effect. However
2437 * at the point when this call is done we typically don't yet
2438 * know which interface is going to be used */
2439 if (val < TCP_MIN_MSS || val > MAX_TCP_WINDOW) {
2443 tp->rx_opt.user_mss = val;
2448 /* TCP_NODELAY is weaker than TCP_CORK, so that
2449 * this option on corked socket is remembered, but
2450 * it is not activated until cork is cleared.
2452 * However, when TCP_NODELAY is set we make
2453 * an explicit push, which overrides even TCP_CORK
2454 * for currently queued segments.
2456 tp->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH;
2457 tcp_push_pending_frames(sk);
2459 tp->nonagle &= ~TCP_NAGLE_OFF;
2463 case TCP_THIN_LINEAR_TIMEOUTS:
2464 if (val < 0 || val > 1)
2470 case TCP_THIN_DUPACK:
2471 if (val < 0 || val > 1)
2474 tp->thin_dupack = val;
2475 if (tp->thin_dupack)
2476 tcp_disable_early_retrans(tp);
2481 if (!tcp_can_repair_sock(sk))
2483 else if (val == 1) {
2485 sk->sk_reuse = SK_FORCE_REUSE;
2486 tp->repair_queue = TCP_NO_QUEUE;
2487 } else if (val == 0) {
2489 sk->sk_reuse = SK_NO_REUSE;
2490 tcp_send_window_probe(sk);
2496 case TCP_REPAIR_QUEUE:
2499 else if (val < TCP_QUEUES_NR)
2500 tp->repair_queue = val;
2506 if (sk->sk_state != TCP_CLOSE)
2508 else if (tp->repair_queue == TCP_SEND_QUEUE)
2509 tp->write_seq = val;
2510 else if (tp->repair_queue == TCP_RECV_QUEUE)
2516 case TCP_REPAIR_OPTIONS:
2519 else if (sk->sk_state == TCP_ESTABLISHED)
2520 err = tcp_repair_options_est(tp,
2521 (struct tcp_repair_opt __user *)optval,
2528 /* When set indicates to always queue non-full frames.
2529 * Later the user clears this option and we transmit
2530 * any pending partial frames in the queue. This is
2531 * meant to be used alongside sendfile() to get properly
2532 * filled frames when the user (for example) must write
2533 * out headers with a write() call first and then use
2534 * sendfile to send out the data parts.
2536 * TCP_CORK can be set together with TCP_NODELAY and it is
2537 * stronger than TCP_NODELAY.
2540 tp->nonagle |= TCP_NAGLE_CORK;
2542 tp->nonagle &= ~TCP_NAGLE_CORK;
2543 if (tp->nonagle&TCP_NAGLE_OFF)
2544 tp->nonagle |= TCP_NAGLE_PUSH;
2545 tcp_push_pending_frames(sk);
2550 if (val < 1 || val > MAX_TCP_KEEPIDLE)
2553 tp->keepalive_time = val * HZ;
2554 if (sock_flag(sk, SOCK_KEEPOPEN) &&
2555 !((1 << sk->sk_state) &
2556 (TCPF_CLOSE | TCPF_LISTEN))) {
2557 u32 elapsed = keepalive_time_elapsed(tp);
2558 if (tp->keepalive_time > elapsed)
2559 elapsed = tp->keepalive_time - elapsed;
2562 inet_csk_reset_keepalive_timer(sk, elapsed);
2567 if (val < 1 || val > MAX_TCP_KEEPINTVL)
2570 tp->keepalive_intvl = val * HZ;
2573 if (val < 1 || val > MAX_TCP_KEEPCNT)
2576 tp->keepalive_probes = val;
2579 if (val < 1 || val > MAX_TCP_SYNCNT)
2582 icsk->icsk_syn_retries = val;
2586 if (val < 0 || val > 1)
2595 else if (val > net->ipv4.sysctl_tcp_fin_timeout / HZ)
2598 tp->linger2 = val * HZ;
2601 case TCP_DEFER_ACCEPT:
2602 /* Translate value in seconds to number of retransmits */
2603 icsk->icsk_accept_queue.rskq_defer_accept =
2604 secs_to_retrans(val, TCP_TIMEOUT_INIT / HZ,
2608 case TCP_WINDOW_CLAMP:
2610 if (sk->sk_state != TCP_CLOSE) {
2614 tp->window_clamp = 0;
2616 tp->window_clamp = val < SOCK_MIN_RCVBUF / 2 ?
2617 SOCK_MIN_RCVBUF / 2 : val;
2622 icsk->icsk_ack.pingpong = 1;
2624 icsk->icsk_ack.pingpong = 0;
2625 if ((1 << sk->sk_state) &
2626 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) &&
2627 inet_csk_ack_scheduled(sk)) {
2628 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
2629 tcp_cleanup_rbuf(sk, 1);
2631 icsk->icsk_ack.pingpong = 1;
2636 #ifdef CONFIG_TCP_MD5SIG
2638 /* Read the IP->Key mappings from userspace */
2639 err = tp->af_specific->md5_parse(sk, optval, optlen);
2642 case TCP_USER_TIMEOUT:
2643 /* Cap the max time in ms TCP will retry or probe the window
2644 * before giving up and aborting (ETIMEDOUT) a connection.
2649 icsk->icsk_user_timeout = msecs_to_jiffies(val);
2653 if (val >= 0 && ((1 << sk->sk_state) & (TCPF_CLOSE |
2655 tcp_fastopen_init_key_once(true);
2657 fastopen_queue_tune(sk, val);
2666 tp->tsoffset = val - tcp_time_stamp;
2668 case TCP_REPAIR_WINDOW:
2669 err = tcp_repair_set_window(tp, optval, optlen);
2671 case TCP_NOTSENT_LOWAT:
2672 tp->notsent_lowat = val;
2673 sk->sk_write_space(sk);
2684 int tcp_setsockopt(struct sock *sk, int level, int optname, char __user *optval,
2685 unsigned int optlen)
2687 const struct inet_connection_sock *icsk = inet_csk(sk);
2689 if (level != SOL_TCP)
2690 return icsk->icsk_af_ops->setsockopt(sk, level, optname,
2692 return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2694 EXPORT_SYMBOL(tcp_setsockopt);
2696 #ifdef CONFIG_COMPAT
2697 int compat_tcp_setsockopt(struct sock *sk, int level, int optname,
2698 char __user *optval, unsigned int optlen)
2700 if (level != SOL_TCP)
2701 return inet_csk_compat_setsockopt(sk, level, optname,
2703 return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2705 EXPORT_SYMBOL(compat_tcp_setsockopt);
2708 /* Return information about state of tcp endpoint in API format. */
2709 void tcp_get_info(struct sock *sk, struct tcp_info *info)
2711 const struct tcp_sock *tp = tcp_sk(sk); /* iff sk_type == SOCK_STREAM */
2712 const struct inet_connection_sock *icsk = inet_csk(sk);
2713 u32 now = tcp_time_stamp, intv;
2719 memset(info, 0, sizeof(*info));
2720 if (sk->sk_type != SOCK_STREAM)
2723 info->tcpi_state = sk_state_load(sk);
2725 info->tcpi_ca_state = icsk->icsk_ca_state;
2726 info->tcpi_retransmits = icsk->icsk_retransmits;
2727 info->tcpi_probes = icsk->icsk_probes_out;
2728 info->tcpi_backoff = icsk->icsk_backoff;
2730 if (tp->rx_opt.tstamp_ok)
2731 info->tcpi_options |= TCPI_OPT_TIMESTAMPS;
2732 if (tcp_is_sack(tp))
2733 info->tcpi_options |= TCPI_OPT_SACK;
2734 if (tp->rx_opt.wscale_ok) {
2735 info->tcpi_options |= TCPI_OPT_WSCALE;
2736 info->tcpi_snd_wscale = tp->rx_opt.snd_wscale;
2737 info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale;
2740 if (tp->ecn_flags & TCP_ECN_OK)
2741 info->tcpi_options |= TCPI_OPT_ECN;
2742 if (tp->ecn_flags & TCP_ECN_SEEN)
2743 info->tcpi_options |= TCPI_OPT_ECN_SEEN;
2744 if (tp->syn_data_acked)
2745 info->tcpi_options |= TCPI_OPT_SYN_DATA;
2747 info->tcpi_rto = jiffies_to_usecs(icsk->icsk_rto);
2748 info->tcpi_ato = jiffies_to_usecs(icsk->icsk_ack.ato);
2749 info->tcpi_snd_mss = tp->mss_cache;
2750 info->tcpi_rcv_mss = icsk->icsk_ack.rcv_mss;
2752 if (info->tcpi_state == TCP_LISTEN) {
2753 info->tcpi_unacked = sk->sk_ack_backlog;
2754 info->tcpi_sacked = sk->sk_max_ack_backlog;
2756 info->tcpi_unacked = tp->packets_out;
2757 info->tcpi_sacked = tp->sacked_out;
2759 info->tcpi_lost = tp->lost_out;
2760 info->tcpi_retrans = tp->retrans_out;
2761 info->tcpi_fackets = tp->fackets_out;
2763 info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime);
2764 info->tcpi_last_data_recv = jiffies_to_msecs(now - icsk->icsk_ack.lrcvtime);
2765 info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp);
2767 info->tcpi_pmtu = icsk->icsk_pmtu_cookie;
2768 info->tcpi_rcv_ssthresh = tp->rcv_ssthresh;
2769 info->tcpi_rtt = tp->srtt_us >> 3;
2770 info->tcpi_rttvar = tp->mdev_us >> 2;
2771 info->tcpi_snd_ssthresh = tp->snd_ssthresh;
2772 info->tcpi_snd_cwnd = tp->snd_cwnd;
2773 info->tcpi_advmss = tp->advmss;
2774 info->tcpi_reordering = tp->reordering;
2776 info->tcpi_rcv_rtt = jiffies_to_usecs(tp->rcv_rtt_est.rtt)>>3;
2777 info->tcpi_rcv_space = tp->rcvq_space.space;
2779 info->tcpi_total_retrans = tp->total_retrans;
2781 rate = READ_ONCE(sk->sk_pacing_rate);
2782 rate64 = rate != ~0U ? rate : ~0ULL;
2783 put_unaligned(rate64, &info->tcpi_pacing_rate);
2785 rate = READ_ONCE(sk->sk_max_pacing_rate);
2786 rate64 = rate != ~0U ? rate : ~0ULL;
2787 put_unaligned(rate64, &info->tcpi_max_pacing_rate);
2790 start = u64_stats_fetch_begin_irq(&tp->syncp);
2791 put_unaligned(tp->bytes_acked, &info->tcpi_bytes_acked);
2792 put_unaligned(tp->bytes_received, &info->tcpi_bytes_received);
2793 } while (u64_stats_fetch_retry_irq(&tp->syncp, start));
2794 info->tcpi_segs_out = tp->segs_out;
2795 info->tcpi_segs_in = tp->segs_in;
2797 notsent_bytes = READ_ONCE(tp->write_seq) - READ_ONCE(tp->snd_nxt);
2798 info->tcpi_notsent_bytes = max(0, notsent_bytes);
2800 info->tcpi_min_rtt = tcp_min_rtt(tp);
2801 info->tcpi_data_segs_in = tp->data_segs_in;
2802 info->tcpi_data_segs_out = tp->data_segs_out;
2804 info->tcpi_delivery_rate_app_limited = tp->rate_app_limited ? 1 : 0;
2805 rate = READ_ONCE(tp->rate_delivered);
2806 intv = READ_ONCE(tp->rate_interval_us);
2808 rate64 = (u64)rate * tp->mss_cache * USEC_PER_SEC;
2809 do_div(rate64, intv);
2810 put_unaligned(rate64, &info->tcpi_delivery_rate);
2813 EXPORT_SYMBOL_GPL(tcp_get_info);
2815 static int do_tcp_getsockopt(struct sock *sk, int level,
2816 int optname, char __user *optval, int __user *optlen)
2818 struct inet_connection_sock *icsk = inet_csk(sk);
2819 struct tcp_sock *tp = tcp_sk(sk);
2820 struct net *net = sock_net(sk);
2823 if (get_user(len, optlen))
2826 len = min_t(unsigned int, len, sizeof(int));
2833 val = tp->mss_cache;
2834 if (!val && ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
2835 val = tp->rx_opt.user_mss;
2837 val = tp->rx_opt.mss_clamp;
2840 val = !!(tp->nonagle&TCP_NAGLE_OFF);
2843 val = !!(tp->nonagle&TCP_NAGLE_CORK);
2846 val = keepalive_time_when(tp) / HZ;
2849 val = keepalive_intvl_when(tp) / HZ;
2852 val = keepalive_probes(tp);
2855 val = icsk->icsk_syn_retries ? : net->ipv4.sysctl_tcp_syn_retries;
2860 val = (val ? : net->ipv4.sysctl_tcp_fin_timeout) / HZ;
2862 case TCP_DEFER_ACCEPT:
2863 val = retrans_to_secs(icsk->icsk_accept_queue.rskq_defer_accept,
2864 TCP_TIMEOUT_INIT / HZ, TCP_RTO_MAX / HZ);
2866 case TCP_WINDOW_CLAMP:
2867 val = tp->window_clamp;
2870 struct tcp_info info;
2872 if (get_user(len, optlen))
2875 tcp_get_info(sk, &info);
2877 len = min_t(unsigned int, len, sizeof(info));
2878 if (put_user(len, optlen))
2880 if (copy_to_user(optval, &info, len))
2885 const struct tcp_congestion_ops *ca_ops;
2886 union tcp_cc_info info;
2890 if (get_user(len, optlen))
2893 ca_ops = icsk->icsk_ca_ops;
2894 if (ca_ops && ca_ops->get_info)
2895 sz = ca_ops->get_info(sk, ~0U, &attr, &info);
2897 len = min_t(unsigned int, len, sz);
2898 if (put_user(len, optlen))
2900 if (copy_to_user(optval, &info, len))
2905 val = !icsk->icsk_ack.pingpong;
2908 case TCP_CONGESTION:
2909 if (get_user(len, optlen))
2911 len = min_t(unsigned int, len, TCP_CA_NAME_MAX);
2912 if (put_user(len, optlen))
2914 if (copy_to_user(optval, icsk->icsk_ca_ops->name, len))
2918 case TCP_THIN_LINEAR_TIMEOUTS:
2921 case TCP_THIN_DUPACK:
2922 val = tp->thin_dupack;
2929 case TCP_REPAIR_QUEUE:
2931 val = tp->repair_queue;
2936 case TCP_REPAIR_WINDOW: {
2937 struct tcp_repair_window opt;
2939 if (get_user(len, optlen))
2942 if (len != sizeof(opt))
2948 opt.snd_wl1 = tp->snd_wl1;
2949 opt.snd_wnd = tp->snd_wnd;
2950 opt.max_window = tp->max_window;
2951 opt.rcv_wnd = tp->rcv_wnd;
2952 opt.rcv_wup = tp->rcv_wup;
2954 if (copy_to_user(optval, &opt, len))
2959 if (tp->repair_queue == TCP_SEND_QUEUE)
2960 val = tp->write_seq;
2961 else if (tp->repair_queue == TCP_RECV_QUEUE)
2967 case TCP_USER_TIMEOUT:
2968 val = jiffies_to_msecs(icsk->icsk_user_timeout);
2972 val = icsk->icsk_accept_queue.fastopenq.max_qlen;
2976 val = tcp_time_stamp + tp->tsoffset;
2978 case TCP_NOTSENT_LOWAT:
2979 val = tp->notsent_lowat;
2984 case TCP_SAVED_SYN: {
2985 if (get_user(len, optlen))
2989 if (tp->saved_syn) {
2990 if (len < tp->saved_syn[0]) {
2991 if (put_user(tp->saved_syn[0], optlen)) {
2998 len = tp->saved_syn[0];
2999 if (put_user(len, optlen)) {
3003 if (copy_to_user(optval, tp->saved_syn + 1, len)) {
3007 tcp_saved_syn_free(tp);
3012 if (put_user(len, optlen))
3018 return -ENOPROTOOPT;
3021 if (put_user(len, optlen))
3023 if (copy_to_user(optval, &val, len))
3028 int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval,
3031 struct inet_connection_sock *icsk = inet_csk(sk);
3033 if (level != SOL_TCP)
3034 return icsk->icsk_af_ops->getsockopt(sk, level, optname,
3036 return do_tcp_getsockopt(sk, level, optname, optval, optlen);
3038 EXPORT_SYMBOL(tcp_getsockopt);
3040 #ifdef CONFIG_COMPAT
3041 int compat_tcp_getsockopt(struct sock *sk, int level, int optname,
3042 char __user *optval, int __user *optlen)
3044 if (level != SOL_TCP)
3045 return inet_csk_compat_getsockopt(sk, level, optname,
3047 return do_tcp_getsockopt(sk, level, optname, optval, optlen);
3049 EXPORT_SYMBOL(compat_tcp_getsockopt);
3052 #ifdef CONFIG_TCP_MD5SIG
3053 static DEFINE_PER_CPU(struct tcp_md5sig_pool, tcp_md5sig_pool);
3054 static DEFINE_MUTEX(tcp_md5sig_mutex);
3055 static bool tcp_md5sig_pool_populated = false;
3057 static void __tcp_alloc_md5sig_pool(void)
3059 struct crypto_ahash *hash;
3062 hash = crypto_alloc_ahash("md5", 0, CRYPTO_ALG_ASYNC);
3066 for_each_possible_cpu(cpu) {
3067 void *scratch = per_cpu(tcp_md5sig_pool, cpu).scratch;
3068 struct ahash_request *req;
3071 scratch = kmalloc_node(sizeof(union tcp_md5sum_block) +
3072 sizeof(struct tcphdr),
3077 per_cpu(tcp_md5sig_pool, cpu).scratch = scratch;
3079 if (per_cpu(tcp_md5sig_pool, cpu).md5_req)
3082 req = ahash_request_alloc(hash, GFP_KERNEL);
3086 ahash_request_set_callback(req, 0, NULL, NULL);
3088 per_cpu(tcp_md5sig_pool, cpu).md5_req = req;
3090 /* before setting tcp_md5sig_pool_populated, we must commit all writes
3091 * to memory. See smp_rmb() in tcp_get_md5sig_pool()
3094 tcp_md5sig_pool_populated = true;
3097 bool tcp_alloc_md5sig_pool(void)
3099 if (unlikely(!tcp_md5sig_pool_populated)) {
3100 mutex_lock(&tcp_md5sig_mutex);
3102 if (!tcp_md5sig_pool_populated)
3103 __tcp_alloc_md5sig_pool();
3105 mutex_unlock(&tcp_md5sig_mutex);
3107 return tcp_md5sig_pool_populated;
3109 EXPORT_SYMBOL(tcp_alloc_md5sig_pool);
3113 * tcp_get_md5sig_pool - get md5sig_pool for this user
3115 * We use percpu structure, so if we succeed, we exit with preemption
3116 * and BH disabled, to make sure another thread or softirq handling
3117 * wont try to get same context.
3119 struct tcp_md5sig_pool *tcp_get_md5sig_pool(void)
3123 if (tcp_md5sig_pool_populated) {
3124 /* coupled with smp_wmb() in __tcp_alloc_md5sig_pool() */
3126 return this_cpu_ptr(&tcp_md5sig_pool);
3131 EXPORT_SYMBOL(tcp_get_md5sig_pool);
3133 int tcp_md5_hash_skb_data(struct tcp_md5sig_pool *hp,
3134 const struct sk_buff *skb, unsigned int header_len)
3136 struct scatterlist sg;
3137 const struct tcphdr *tp = tcp_hdr(skb);
3138 struct ahash_request *req = hp->md5_req;
3140 const unsigned int head_data_len = skb_headlen(skb) > header_len ?
3141 skb_headlen(skb) - header_len : 0;
3142 const struct skb_shared_info *shi = skb_shinfo(skb);
3143 struct sk_buff *frag_iter;
3145 sg_init_table(&sg, 1);
3147 sg_set_buf(&sg, ((u8 *) tp) + header_len, head_data_len);
3148 ahash_request_set_crypt(req, &sg, NULL, head_data_len);
3149 if (crypto_ahash_update(req))
3152 for (i = 0; i < shi->nr_frags; ++i) {
3153 const struct skb_frag_struct *f = &shi->frags[i];
3154 unsigned int offset = f->page_offset;
3155 struct page *page = skb_frag_page(f) + (offset >> PAGE_SHIFT);
3157 sg_set_page(&sg, page, skb_frag_size(f),
3158 offset_in_page(offset));
3159 ahash_request_set_crypt(req, &sg, NULL, skb_frag_size(f));
3160 if (crypto_ahash_update(req))
3164 skb_walk_frags(skb, frag_iter)
3165 if (tcp_md5_hash_skb_data(hp, frag_iter, 0))
3170 EXPORT_SYMBOL(tcp_md5_hash_skb_data);
3172 int tcp_md5_hash_key(struct tcp_md5sig_pool *hp, const struct tcp_md5sig_key *key)
3174 struct scatterlist sg;
3176 sg_init_one(&sg, key->key, key->keylen);
3177 ahash_request_set_crypt(hp->md5_req, &sg, NULL, key->keylen);
3178 return crypto_ahash_update(hp->md5_req);
3180 EXPORT_SYMBOL(tcp_md5_hash_key);
3184 void tcp_done(struct sock *sk)
3186 struct request_sock *req = tcp_sk(sk)->fastopen_rsk;
3188 if (sk->sk_state == TCP_SYN_SENT || sk->sk_state == TCP_SYN_RECV)
3189 TCP_INC_STATS(sock_net(sk), TCP_MIB_ATTEMPTFAILS);
3191 tcp_set_state(sk, TCP_CLOSE);
3192 tcp_clear_xmit_timers(sk);
3194 reqsk_fastopen_remove(sk, req, false);
3196 sk->sk_shutdown = SHUTDOWN_MASK;
3198 if (!sock_flag(sk, SOCK_DEAD))
3199 sk->sk_state_change(sk);
3201 inet_csk_destroy_sock(sk);
3203 EXPORT_SYMBOL_GPL(tcp_done);
3205 int tcp_abort(struct sock *sk, int err)
3207 if (!sk_fullsock(sk)) {
3208 if (sk->sk_state == TCP_NEW_SYN_RECV) {
3209 struct request_sock *req = inet_reqsk(sk);
3212 inet_csk_reqsk_queue_drop_and_put(req->rsk_listener,
3220 /* Don't race with userspace socket closes such as tcp_close. */
3223 if (sk->sk_state == TCP_LISTEN) {
3224 tcp_set_state(sk, TCP_CLOSE);
3225 inet_csk_listen_stop(sk);
3228 /* Don't race with BH socket closes such as inet_csk_listen_stop. */
3232 if (!sock_flag(sk, SOCK_DEAD)) {
3234 /* This barrier is coupled with smp_rmb() in tcp_poll() */
3236 sk->sk_error_report(sk);
3237 if (tcp_need_reset(sk->sk_state))
3238 tcp_send_active_reset(sk, GFP_ATOMIC);
3247 EXPORT_SYMBOL_GPL(tcp_abort);
3249 extern struct tcp_congestion_ops tcp_reno;
3251 static __initdata unsigned long thash_entries;
3252 static int __init set_thash_entries(char *str)
3259 ret = kstrtoul(str, 0, &thash_entries);
3265 __setup("thash_entries=", set_thash_entries);
3267 static void __init tcp_init_mem(void)
3269 unsigned long limit = nr_free_buffer_pages() / 16;
3271 limit = max(limit, 128UL);
3272 sysctl_tcp_mem[0] = limit / 4 * 3; /* 4.68 % */
3273 sysctl_tcp_mem[1] = limit; /* 6.25 % */
3274 sysctl_tcp_mem[2] = sysctl_tcp_mem[0] * 2; /* 9.37 % */
3277 void __init tcp_init(void)
3279 int max_rshare, max_wshare, cnt;
3280 unsigned long limit;
3283 BUILD_BUG_ON(sizeof(struct tcp_skb_cb) >
3284 FIELD_SIZEOF(struct sk_buff, cb));
3286 percpu_counter_init(&tcp_sockets_allocated, 0, GFP_KERNEL);
3287 percpu_counter_init(&tcp_orphan_count, 0, GFP_KERNEL);
3288 tcp_hashinfo.bind_bucket_cachep =
3289 kmem_cache_create("tcp_bind_bucket",
3290 sizeof(struct inet_bind_bucket), 0,
3291 SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL);
3293 /* Size and allocate the main established and bind bucket
3296 * The methodology is similar to that of the buffer cache.
3298 tcp_hashinfo.ehash =
3299 alloc_large_system_hash("TCP established",
3300 sizeof(struct inet_ehash_bucket),
3302 17, /* one slot per 128 KB of memory */
3305 &tcp_hashinfo.ehash_mask,
3307 thash_entries ? 0 : 512 * 1024);
3308 for (i = 0; i <= tcp_hashinfo.ehash_mask; i++)
3309 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].chain, i);
3311 if (inet_ehash_locks_alloc(&tcp_hashinfo))
3312 panic("TCP: failed to alloc ehash_locks");
3313 tcp_hashinfo.bhash =
3314 alloc_large_system_hash("TCP bind",
3315 sizeof(struct inet_bind_hashbucket),
3316 tcp_hashinfo.ehash_mask + 1,
3317 17, /* one slot per 128 KB of memory */
3319 &tcp_hashinfo.bhash_size,
3323 tcp_hashinfo.bhash_size = 1U << tcp_hashinfo.bhash_size;
3324 for (i = 0; i < tcp_hashinfo.bhash_size; i++) {
3325 spin_lock_init(&tcp_hashinfo.bhash[i].lock);
3326 INIT_HLIST_HEAD(&tcp_hashinfo.bhash[i].chain);
3330 cnt = tcp_hashinfo.ehash_mask + 1;
3332 tcp_death_row.sysctl_max_tw_buckets = cnt / 2;
3333 sysctl_tcp_max_orphans = cnt / 2;
3334 sysctl_max_syn_backlog = max(128, cnt / 256);
3337 /* Set per-socket limits to no more than 1/128 the pressure threshold */
3338 limit = nr_free_buffer_pages() << (PAGE_SHIFT - 7);
3339 max_wshare = min(4UL*1024*1024, limit);
3340 max_rshare = min(6UL*1024*1024, limit);
3342 sysctl_tcp_wmem[0] = SK_MEM_QUANTUM;
3343 sysctl_tcp_wmem[1] = 16*1024;
3344 sysctl_tcp_wmem[2] = max(64*1024, max_wshare);
3346 sysctl_tcp_rmem[0] = SK_MEM_QUANTUM;
3347 sysctl_tcp_rmem[1] = 87380;
3348 sysctl_tcp_rmem[2] = max(87380, max_rshare);
3350 pr_info("Hash tables configured (established %u bind %u)\n",
3351 tcp_hashinfo.ehash_mask + 1, tcp_hashinfo.bhash_size);
3354 BUG_ON(tcp_register_congestion_control(&tcp_reno) != 0);