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 #include <linux/kernel.h>
249 #include <linux/module.h>
250 #include <linux/types.h>
251 #include <linux/fcntl.h>
252 #include <linux/poll.h>
253 #include <linux/init.h>
254 #include <linux/fs.h>
255 #include <linux/skbuff.h>
256 #include <linux/scatterlist.h>
257 #include <linux/splice.h>
258 #include <linux/net.h>
259 #include <linux/socket.h>
260 #include <linux/random.h>
261 #include <linux/bootmem.h>
262 #include <linux/highmem.h>
263 #include <linux/swap.h>
264 #include <linux/cache.h>
265 #include <linux/err.h>
266 #include <linux/crypto.h>
267 #include <linux/time.h>
268 #include <linux/slab.h>
269 #include <linux/uid_stat.h>
271 #include <net/icmp.h>
273 #include <net/xfrm.h>
275 #include <net/netdma.h>
276 #include <net/sock.h>
278 #include <asm/uaccess.h>
279 #include <asm/ioctls.h>
281 int sysctl_tcp_fin_timeout __read_mostly = TCP_FIN_TIMEOUT;
283 struct percpu_counter tcp_orphan_count;
284 EXPORT_SYMBOL_GPL(tcp_orphan_count);
286 int sysctl_tcp_mem[3] __read_mostly;
287 int sysctl_tcp_wmem[3] __read_mostly;
288 int sysctl_tcp_rmem[3] __read_mostly;
290 EXPORT_SYMBOL(sysctl_tcp_mem);
291 EXPORT_SYMBOL(sysctl_tcp_rmem);
292 EXPORT_SYMBOL(sysctl_tcp_wmem);
294 atomic_t tcp_memory_allocated; /* Current allocated memory. */
295 EXPORT_SYMBOL(tcp_memory_allocated);
298 * Current number of TCP sockets.
300 struct percpu_counter tcp_sockets_allocated;
301 EXPORT_SYMBOL(tcp_sockets_allocated);
306 struct tcp_splice_state {
307 struct pipe_inode_info *pipe;
313 * Pressure flag: try to collapse.
314 * Technical note: it is used by multiple contexts non atomically.
315 * All the __sk_mem_schedule() is of this nature: accounting
316 * is strict, actions are advisory and have some latency.
318 int tcp_memory_pressure __read_mostly;
319 EXPORT_SYMBOL(tcp_memory_pressure);
321 void tcp_enter_memory_pressure(struct sock *sk)
323 if (!tcp_memory_pressure) {
324 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMEMORYPRESSURES);
325 tcp_memory_pressure = 1;
328 EXPORT_SYMBOL(tcp_enter_memory_pressure);
330 /* Convert seconds to retransmits based on initial and max timeout */
331 static u8 secs_to_retrans(int seconds, int timeout, int rto_max)
336 int period = timeout;
339 while (seconds > period && res < 255) {
342 if (timeout > rto_max)
350 /* Convert retransmits to seconds based on initial and max timeout */
351 static int retrans_to_secs(u8 retrans, int timeout, int rto_max)
359 if (timeout > rto_max)
368 * Wait for a TCP event.
370 * Note that we don't need to lock the socket, as the upper poll layers
371 * take care of normal races (between the test and the event) and we don't
372 * go look at any of the socket buffers directly.
374 unsigned int tcp_poll(struct file *file, struct socket *sock, poll_table *wait)
377 struct sock *sk = sock->sk;
378 struct tcp_sock *tp = tcp_sk(sk);
380 sock_poll_wait(file, sk_sleep(sk), wait);
381 if (sk->sk_state == TCP_LISTEN)
382 return inet_csk_listen_poll(sk);
384 /* Socket is not locked. We are protected from async events
385 * by poll logic and correct handling of state changes
386 * made by other threads is impossible in any case.
392 * POLLHUP is certainly not done right. But poll() doesn't
393 * have a notion of HUP in just one direction, and for a
394 * socket the read side is more interesting.
396 * Some poll() documentation says that POLLHUP is incompatible
397 * with the POLLOUT/POLLWR flags, so somebody should check this
398 * all. But careful, it tends to be safer to return too many
399 * bits than too few, and you can easily break real applications
400 * if you don't tell them that something has hung up!
404 * Check number 1. POLLHUP is _UNMASKABLE_ event (see UNIX98 and
405 * our fs/select.c). It means that after we received EOF,
406 * poll always returns immediately, making impossible poll() on write()
407 * in state CLOSE_WAIT. One solution is evident --- to set POLLHUP
408 * if and only if shutdown has been made in both directions.
409 * Actually, it is interesting to look how Solaris and DUX
410 * solve this dilemma. I would prefer, if POLLHUP were maskable,
411 * then we could set it on SND_SHUTDOWN. BTW examples given
412 * in Stevens' books assume exactly this behaviour, it explains
413 * why POLLHUP is incompatible with POLLOUT. --ANK
415 * NOTE. Check for TCP_CLOSE is added. The goal is to prevent
416 * blocking on fresh not-connected or disconnected socket. --ANK
418 if (sk->sk_shutdown == SHUTDOWN_MASK || sk->sk_state == TCP_CLOSE)
420 if (sk->sk_shutdown & RCV_SHUTDOWN)
421 mask |= POLLIN | POLLRDNORM | POLLRDHUP;
424 if ((1 << sk->sk_state) & ~(TCPF_SYN_SENT | TCPF_SYN_RECV)) {
425 int target = sock_rcvlowat(sk, 0, INT_MAX);
427 if (tp->urg_seq == tp->copied_seq &&
428 !sock_flag(sk, SOCK_URGINLINE) &&
432 /* Potential race condition. If read of tp below will
433 * escape above sk->sk_state, we can be illegally awaken
434 * in SYN_* states. */
435 if (tp->rcv_nxt - tp->copied_seq >= target)
436 mask |= POLLIN | POLLRDNORM;
438 if (!(sk->sk_shutdown & SEND_SHUTDOWN)) {
439 if (sk_stream_wspace(sk) >= sk_stream_min_wspace(sk)) {
440 mask |= POLLOUT | POLLWRNORM;
441 } else { /* send SIGIO later */
442 set_bit(SOCK_ASYNC_NOSPACE,
443 &sk->sk_socket->flags);
444 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
446 /* Race breaker. If space is freed after
447 * wspace test but before the flags are set,
448 * IO signal will be lost.
450 if (sk_stream_wspace(sk) >= sk_stream_min_wspace(sk))
451 mask |= POLLOUT | POLLWRNORM;
454 mask |= POLLOUT | POLLWRNORM;
456 if (tp->urg_data & TCP_URG_VALID)
459 /* This barrier is coupled with smp_wmb() in tcp_reset() */
466 EXPORT_SYMBOL(tcp_poll);
468 int tcp_ioctl(struct sock *sk, int cmd, unsigned long arg)
470 struct tcp_sock *tp = tcp_sk(sk);
475 if (sk->sk_state == TCP_LISTEN)
479 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
481 else if (sock_flag(sk, SOCK_URGINLINE) ||
483 before(tp->urg_seq, tp->copied_seq) ||
484 !before(tp->urg_seq, tp->rcv_nxt)) {
487 answ = tp->rcv_nxt - tp->copied_seq;
489 /* Subtract 1, if FIN is in queue. */
490 skb = skb_peek_tail(&sk->sk_receive_queue);
492 answ -= tcp_hdr(skb)->fin;
494 answ = tp->urg_seq - tp->copied_seq;
498 answ = tp->urg_data && tp->urg_seq == tp->copied_seq;
501 if (sk->sk_state == TCP_LISTEN)
504 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
507 answ = tp->write_seq - tp->snd_una;
513 return put_user(answ, (int __user *)arg);
515 EXPORT_SYMBOL(tcp_ioctl);
517 static inline void tcp_mark_push(struct tcp_sock *tp, struct sk_buff *skb)
519 TCP_SKB_CB(skb)->flags |= TCPHDR_PSH;
520 tp->pushed_seq = tp->write_seq;
523 static inline int forced_push(struct tcp_sock *tp)
525 return after(tp->write_seq, tp->pushed_seq + (tp->max_window >> 1));
528 static inline void skb_entail(struct sock *sk, struct sk_buff *skb)
530 struct tcp_sock *tp = tcp_sk(sk);
531 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
534 tcb->seq = tcb->end_seq = tp->write_seq;
535 tcb->flags = TCPHDR_ACK;
537 skb_header_release(skb);
538 tcp_add_write_queue_tail(sk, skb);
539 sk->sk_wmem_queued += skb->truesize;
540 sk_mem_charge(sk, skb->truesize);
541 if (tp->nonagle & TCP_NAGLE_PUSH)
542 tp->nonagle &= ~TCP_NAGLE_PUSH;
545 static inline void tcp_mark_urg(struct tcp_sock *tp, int flags)
548 tp->snd_up = tp->write_seq;
551 static inline void tcp_push(struct sock *sk, int flags, int mss_now,
554 if (tcp_send_head(sk)) {
555 struct tcp_sock *tp = tcp_sk(sk);
557 if (!(flags & MSG_MORE) || forced_push(tp))
558 tcp_mark_push(tp, tcp_write_queue_tail(sk));
560 tcp_mark_urg(tp, flags);
561 __tcp_push_pending_frames(sk, mss_now,
562 (flags & MSG_MORE) ? TCP_NAGLE_CORK : nonagle);
566 static int tcp_splice_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb,
567 unsigned int offset, size_t len)
569 struct tcp_splice_state *tss = rd_desc->arg.data;
572 ret = skb_splice_bits(skb, offset, tss->pipe, min(rd_desc->count, len),
575 rd_desc->count -= ret;
579 static int __tcp_splice_read(struct sock *sk, struct tcp_splice_state *tss)
581 /* Store TCP splice context information in read_descriptor_t. */
582 read_descriptor_t rd_desc = {
587 return tcp_read_sock(sk, &rd_desc, tcp_splice_data_recv);
591 * tcp_splice_read - splice data from TCP socket to a pipe
592 * @sock: socket to splice from
593 * @ppos: position (not valid)
594 * @pipe: pipe to splice to
595 * @len: number of bytes to splice
596 * @flags: splice modifier flags
599 * Will read pages from given socket and fill them into a pipe.
602 ssize_t tcp_splice_read(struct socket *sock, loff_t *ppos,
603 struct pipe_inode_info *pipe, size_t len,
606 struct sock *sk = sock->sk;
607 struct tcp_splice_state tss = {
616 sock_rps_record_flow(sk);
618 * We can't seek on a socket input
627 timeo = sock_rcvtimeo(sk, sock->file->f_flags & O_NONBLOCK);
629 ret = __tcp_splice_read(sk, &tss);
635 if (sock_flag(sk, SOCK_DONE))
638 ret = sock_error(sk);
641 if (sk->sk_shutdown & RCV_SHUTDOWN)
643 if (sk->sk_state == TCP_CLOSE) {
645 * This occurs when user tries to read
646 * from never connected socket.
648 if (!sock_flag(sk, SOCK_DONE))
656 sk_wait_data(sk, &timeo);
657 if (signal_pending(current)) {
658 ret = sock_intr_errno(timeo);
671 if (sk->sk_err || sk->sk_state == TCP_CLOSE ||
672 (sk->sk_shutdown & RCV_SHUTDOWN) ||
673 signal_pending(current))
684 EXPORT_SYMBOL(tcp_splice_read);
686 struct sk_buff *sk_stream_alloc_skb(struct sock *sk, int size, gfp_t gfp)
690 /* The TCP header must be at least 32-bit aligned. */
691 size = ALIGN(size, 4);
693 skb = alloc_skb_fclone(size + sk->sk_prot->max_header, gfp);
695 if (sk_wmem_schedule(sk, skb->truesize)) {
697 * Make sure that we have exactly size bytes
698 * available to the caller, no more, no less.
700 skb_reserve(skb, skb_tailroom(skb) - size);
705 sk->sk_prot->enter_memory_pressure(sk);
706 sk_stream_moderate_sndbuf(sk);
711 static unsigned int tcp_xmit_size_goal(struct sock *sk, u32 mss_now,
714 struct tcp_sock *tp = tcp_sk(sk);
715 u32 xmit_size_goal, old_size_goal;
717 xmit_size_goal = mss_now;
719 if (large_allowed && sk_can_gso(sk)) {
720 xmit_size_goal = ((sk->sk_gso_max_size - 1) -
721 inet_csk(sk)->icsk_af_ops->net_header_len -
722 inet_csk(sk)->icsk_ext_hdr_len -
725 xmit_size_goal = tcp_bound_to_half_wnd(tp, xmit_size_goal);
727 /* We try hard to avoid divides here */
728 old_size_goal = tp->xmit_size_goal_segs * mss_now;
730 if (likely(old_size_goal <= xmit_size_goal &&
731 old_size_goal + mss_now > xmit_size_goal)) {
732 xmit_size_goal = old_size_goal;
734 tp->xmit_size_goal_segs = xmit_size_goal / mss_now;
735 xmit_size_goal = tp->xmit_size_goal_segs * mss_now;
739 return max(xmit_size_goal, mss_now);
742 static int tcp_send_mss(struct sock *sk, int *size_goal, int flags)
746 mss_now = tcp_current_mss(sk);
747 *size_goal = tcp_xmit_size_goal(sk, mss_now, !(flags & MSG_OOB));
752 static ssize_t do_tcp_sendpages(struct sock *sk, struct page **pages, int poffset,
753 size_t psize, int flags)
755 struct tcp_sock *tp = tcp_sk(sk);
756 int mss_now, size_goal;
759 long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
761 /* Wait for a connection to finish. */
762 if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
763 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
766 clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
768 mss_now = tcp_send_mss(sk, &size_goal, flags);
772 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
776 struct sk_buff *skb = tcp_write_queue_tail(sk);
777 struct page *page = pages[poffset / PAGE_SIZE];
778 int copy, i, can_coalesce;
779 int offset = poffset % PAGE_SIZE;
780 int size = min_t(size_t, psize, PAGE_SIZE - offset);
782 if (!tcp_send_head(sk) || (copy = size_goal - skb->len) <= 0) {
784 if (!sk_stream_memory_free(sk))
785 goto wait_for_sndbuf;
787 skb = sk_stream_alloc_skb(sk, 0, sk->sk_allocation);
789 goto wait_for_memory;
798 i = skb_shinfo(skb)->nr_frags;
799 can_coalesce = skb_can_coalesce(skb, i, page, offset);
800 if (!can_coalesce && i >= MAX_SKB_FRAGS) {
801 tcp_mark_push(tp, skb);
804 if (!sk_wmem_schedule(sk, copy))
805 goto wait_for_memory;
808 skb_shinfo(skb)->frags[i - 1].size += copy;
811 skb_fill_page_desc(skb, i, page, offset, copy);
815 skb->data_len += copy;
816 skb->truesize += copy;
817 sk->sk_wmem_queued += copy;
818 sk_mem_charge(sk, copy);
819 skb->ip_summed = CHECKSUM_PARTIAL;
820 tp->write_seq += copy;
821 TCP_SKB_CB(skb)->end_seq += copy;
822 skb_shinfo(skb)->gso_segs = 0;
825 TCP_SKB_CB(skb)->flags &= ~TCPHDR_PSH;
829 if (!(psize -= copy))
832 if (skb->len < size_goal || (flags & MSG_OOB))
835 if (forced_push(tp)) {
836 tcp_mark_push(tp, skb);
837 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
838 } else if (skb == tcp_send_head(sk))
839 tcp_push_one(sk, mss_now);
843 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
846 tcp_push(sk, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH);
848 if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
851 mss_now = tcp_send_mss(sk, &size_goal, flags);
856 tcp_push(sk, flags, mss_now, tp->nonagle);
863 return sk_stream_error(sk, flags, err);
866 int tcp_sendpage(struct sock *sk, struct page *page, int offset,
867 size_t size, int flags)
871 if (!(sk->sk_route_caps & NETIF_F_SG) ||
872 !(sk->sk_route_caps & NETIF_F_ALL_CSUM))
873 return sock_no_sendpage(sk->sk_socket, page, offset, size,
878 res = do_tcp_sendpages(sk, &page, offset, size, flags);
883 EXPORT_SYMBOL(tcp_sendpage);
885 #define TCP_PAGE(sk) (sk->sk_sndmsg_page)
886 #define TCP_OFF(sk) (sk->sk_sndmsg_off)
888 static inline int select_size(struct sock *sk, int sg)
890 struct tcp_sock *tp = tcp_sk(sk);
891 int tmp = tp->mss_cache;
897 int pgbreak = SKB_MAX_HEAD(MAX_TCP_HEADER);
899 if (tmp >= pgbreak &&
900 tmp <= pgbreak + (MAX_SKB_FRAGS - 1) * PAGE_SIZE)
908 int tcp_sendmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg,
912 struct tcp_sock *tp = tcp_sk(sk);
915 int mss_now, size_goal;
922 flags = msg->msg_flags;
923 timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
925 /* Wait for a connection to finish. */
926 if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
927 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
930 /* This should be in poll */
931 clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
933 mss_now = tcp_send_mss(sk, &size_goal, flags);
935 /* Ok commence sending. */
936 iovlen = msg->msg_iovlen;
941 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
944 sg = sk->sk_route_caps & NETIF_F_SG;
946 while (--iovlen >= 0) {
947 size_t seglen = iov->iov_len;
948 unsigned char __user *from = iov->iov_base;
956 skb = tcp_write_queue_tail(sk);
957 if (tcp_send_head(sk)) {
958 if (skb->ip_summed == CHECKSUM_NONE)
960 copy = max - skb->len;
965 /* Allocate new segment. If the interface is SG,
966 * allocate skb fitting to single page.
968 if (!sk_stream_memory_free(sk))
969 goto wait_for_sndbuf;
971 skb = sk_stream_alloc_skb(sk,
975 goto wait_for_memory;
978 * Check whether we can use HW checksum.
980 if (sk->sk_route_caps & NETIF_F_ALL_CSUM)
981 skb->ip_summed = CHECKSUM_PARTIAL;
988 /* Try to append data to the end of skb. */
992 /* Where to copy to? */
993 if (skb_tailroom(skb) > 0) {
994 /* We have some space in skb head. Superb! */
995 if (copy > skb_tailroom(skb))
996 copy = skb_tailroom(skb);
997 if ((err = skb_add_data(skb, from, copy)) != 0)
1001 int i = skb_shinfo(skb)->nr_frags;
1002 struct page *page = TCP_PAGE(sk);
1003 int off = TCP_OFF(sk);
1005 if (skb_can_coalesce(skb, i, page, off) &&
1007 /* We can extend the last page
1010 } else if (i == MAX_SKB_FRAGS || !sg) {
1011 /* Need to add new fragment and cannot
1012 * do this because interface is non-SG,
1013 * or because all the page slots are
1015 tcp_mark_push(tp, skb);
1018 if (off == PAGE_SIZE) {
1020 TCP_PAGE(sk) = page = NULL;
1026 if (copy > PAGE_SIZE - off)
1027 copy = PAGE_SIZE - off;
1029 if (!sk_wmem_schedule(sk, copy))
1030 goto wait_for_memory;
1033 /* Allocate new cache page. */
1034 if (!(page = sk_stream_alloc_page(sk)))
1035 goto wait_for_memory;
1038 /* Time to copy data. We are close to
1040 err = skb_copy_to_page(sk, from, skb, page,
1043 /* If this page was new, give it to the
1044 * socket so it does not get leaked.
1046 if (!TCP_PAGE(sk)) {
1047 TCP_PAGE(sk) = page;
1053 /* Update the skb. */
1055 skb_shinfo(skb)->frags[i - 1].size +=
1058 skb_fill_page_desc(skb, i, page, off, copy);
1061 } else if (off + copy < PAGE_SIZE) {
1063 TCP_PAGE(sk) = page;
1067 TCP_OFF(sk) = off + copy;
1071 TCP_SKB_CB(skb)->flags &= ~TCPHDR_PSH;
1073 tp->write_seq += copy;
1074 TCP_SKB_CB(skb)->end_seq += copy;
1075 skb_shinfo(skb)->gso_segs = 0;
1079 if ((seglen -= copy) == 0 && iovlen == 0)
1082 if (skb->len < max || (flags & MSG_OOB))
1085 if (forced_push(tp)) {
1086 tcp_mark_push(tp, skb);
1087 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
1088 } else if (skb == tcp_send_head(sk))
1089 tcp_push_one(sk, mss_now);
1093 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1096 tcp_push(sk, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH);
1098 if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
1101 mss_now = tcp_send_mss(sk, &size_goal, flags);
1107 tcp_push(sk, flags, mss_now, tp->nonagle);
1108 TCP_CHECK_TIMER(sk);
1112 uid_stat_tcp_snd(current_uid(), copied);
1117 tcp_unlink_write_queue(skb, sk);
1118 /* It is the one place in all of TCP, except connection
1119 * reset, where we can be unlinking the send_head.
1121 tcp_check_send_head(sk, skb);
1122 sk_wmem_free_skb(sk, skb);
1129 err = sk_stream_error(sk, flags, err);
1130 TCP_CHECK_TIMER(sk);
1134 EXPORT_SYMBOL(tcp_sendmsg);
1137 * Handle reading urgent data. BSD has very simple semantics for
1138 * this, no blocking and very strange errors 8)
1141 static int tcp_recv_urg(struct sock *sk, struct msghdr *msg, int len, int flags)
1143 struct tcp_sock *tp = tcp_sk(sk);
1145 /* No URG data to read. */
1146 if (sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data ||
1147 tp->urg_data == TCP_URG_READ)
1148 return -EINVAL; /* Yes this is right ! */
1150 if (sk->sk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DONE))
1153 if (tp->urg_data & TCP_URG_VALID) {
1155 char c = tp->urg_data;
1157 if (!(flags & MSG_PEEK))
1158 tp->urg_data = TCP_URG_READ;
1160 /* Read urgent data. */
1161 msg->msg_flags |= MSG_OOB;
1164 if (!(flags & MSG_TRUNC))
1165 err = memcpy_toiovec(msg->msg_iov, &c, 1);
1168 msg->msg_flags |= MSG_TRUNC;
1170 return err ? -EFAULT : len;
1173 if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
1176 /* Fixed the recv(..., MSG_OOB) behaviour. BSD docs and
1177 * the available implementations agree in this case:
1178 * this call should never block, independent of the
1179 * blocking state of the socket.
1180 * Mike <pall@rz.uni-karlsruhe.de>
1185 /* Clean up the receive buffer for full frames taken by the user,
1186 * then send an ACK if necessary. COPIED is the number of bytes
1187 * tcp_recvmsg has given to the user so far, it speeds up the
1188 * calculation of whether or not we must ACK for the sake of
1191 void tcp_cleanup_rbuf(struct sock *sk, int copied)
1193 struct tcp_sock *tp = tcp_sk(sk);
1194 int time_to_ack = 0;
1197 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
1199 WARN(skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq),
1200 KERN_INFO "cleanup rbuf bug: copied %X seq %X rcvnxt %X\n",
1201 tp->copied_seq, TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt);
1204 if (inet_csk_ack_scheduled(sk)) {
1205 const struct inet_connection_sock *icsk = inet_csk(sk);
1206 /* Delayed ACKs frequently hit locked sockets during bulk
1208 if (icsk->icsk_ack.blocked ||
1209 /* Once-per-two-segments ACK was not sent by tcp_input.c */
1210 tp->rcv_nxt - tp->rcv_wup > icsk->icsk_ack.rcv_mss ||
1212 * If this read emptied read buffer, we send ACK, if
1213 * connection is not bidirectional, user drained
1214 * receive buffer and there was a small segment
1218 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED2) ||
1219 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED) &&
1220 !icsk->icsk_ack.pingpong)) &&
1221 !atomic_read(&sk->sk_rmem_alloc)))
1225 /* We send an ACK if we can now advertise a non-zero window
1226 * which has been raised "significantly".
1228 * Even if window raised up to infinity, do not send window open ACK
1229 * in states, where we will not receive more. It is useless.
1231 if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) {
1232 __u32 rcv_window_now = tcp_receive_window(tp);
1234 /* Optimize, __tcp_select_window() is not cheap. */
1235 if (2*rcv_window_now <= tp->window_clamp) {
1236 __u32 new_window = __tcp_select_window(sk);
1238 /* Send ACK now, if this read freed lots of space
1239 * in our buffer. Certainly, new_window is new window.
1240 * We can advertise it now, if it is not less than current one.
1241 * "Lots" means "at least twice" here.
1243 if (new_window && new_window >= 2 * rcv_window_now)
1251 static void tcp_prequeue_process(struct sock *sk)
1253 struct sk_buff *skb;
1254 struct tcp_sock *tp = tcp_sk(sk);
1256 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPPREQUEUED);
1258 /* RX process wants to run with disabled BHs, though it is not
1261 while ((skb = __skb_dequeue(&tp->ucopy.prequeue)) != NULL)
1262 sk_backlog_rcv(sk, skb);
1265 /* Clear memory counter. */
1266 tp->ucopy.memory = 0;
1269 #ifdef CONFIG_NET_DMA
1270 static void tcp_service_net_dma(struct sock *sk, bool wait)
1272 dma_cookie_t done, used;
1273 dma_cookie_t last_issued;
1274 struct tcp_sock *tp = tcp_sk(sk);
1276 if (!tp->ucopy.dma_chan)
1279 last_issued = tp->ucopy.dma_cookie;
1280 dma_async_memcpy_issue_pending(tp->ucopy.dma_chan);
1283 if (dma_async_memcpy_complete(tp->ucopy.dma_chan,
1285 &used) == DMA_SUCCESS) {
1286 /* Safe to free early-copied skbs now */
1287 __skb_queue_purge(&sk->sk_async_wait_queue);
1290 struct sk_buff *skb;
1291 while ((skb = skb_peek(&sk->sk_async_wait_queue)) &&
1292 (dma_async_is_complete(skb->dma_cookie, done,
1293 used) == DMA_SUCCESS)) {
1294 __skb_dequeue(&sk->sk_async_wait_queue);
1302 static inline struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off)
1304 struct sk_buff *skb;
1307 skb_queue_walk(&sk->sk_receive_queue, skb) {
1308 offset = seq - TCP_SKB_CB(skb)->seq;
1309 if (tcp_hdr(skb)->syn)
1311 if (offset < skb->len || tcp_hdr(skb)->fin) {
1320 * This routine provides an alternative to tcp_recvmsg() for routines
1321 * that would like to handle copying from skbuffs directly in 'sendfile'
1324 * - It is assumed that the socket was locked by the caller.
1325 * - The routine does not block.
1326 * - At present, there is no support for reading OOB data
1327 * or for 'peeking' the socket using this routine
1328 * (although both would be easy to implement).
1330 int tcp_read_sock(struct sock *sk, read_descriptor_t *desc,
1331 sk_read_actor_t recv_actor)
1333 struct sk_buff *skb;
1334 struct tcp_sock *tp = tcp_sk(sk);
1335 u32 seq = tp->copied_seq;
1339 if (sk->sk_state == TCP_LISTEN)
1341 while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1342 if (offset < skb->len) {
1346 len = skb->len - offset;
1347 /* Stop reading if we hit a patch of urgent data */
1349 u32 urg_offset = tp->urg_seq - seq;
1350 if (urg_offset < len)
1355 used = recv_actor(desc, skb, offset, len);
1360 } else if (used <= len) {
1366 * If recv_actor drops the lock (e.g. TCP splice
1367 * receive) the skb pointer might be invalid when
1368 * getting here: tcp_collapse might have deleted it
1369 * while aggregating skbs from the socket queue.
1371 skb = tcp_recv_skb(sk, seq-1, &offset);
1372 if (!skb || (offset+1 != skb->len))
1375 if (tcp_hdr(skb)->fin) {
1376 sk_eat_skb(sk, skb, 0);
1380 sk_eat_skb(sk, skb, 0);
1383 tp->copied_seq = seq;
1385 tp->copied_seq = seq;
1387 tcp_rcv_space_adjust(sk);
1389 /* Clean up data we have read: This will do ACK frames. */
1391 tcp_cleanup_rbuf(sk, copied);
1392 uid_stat_tcp_rcv(current_uid(), copied);
1397 EXPORT_SYMBOL(tcp_read_sock);
1400 * This routine copies from a sock struct into the user buffer.
1402 * Technical note: in 2.3 we work on _locked_ socket, so that
1403 * tricks with *seq access order and skb->users are not required.
1404 * Probably, code can be easily improved even more.
1407 int tcp_recvmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg,
1408 size_t len, int nonblock, int flags, int *addr_len)
1410 struct tcp_sock *tp = tcp_sk(sk);
1416 int target; /* Read at least this many bytes */
1418 struct task_struct *user_recv = NULL;
1419 int copied_early = 0;
1420 struct sk_buff *skb;
1425 TCP_CHECK_TIMER(sk);
1428 if (sk->sk_state == TCP_LISTEN)
1431 timeo = sock_rcvtimeo(sk, nonblock);
1433 /* Urgent data needs to be handled specially. */
1434 if (flags & MSG_OOB)
1437 seq = &tp->copied_seq;
1438 if (flags & MSG_PEEK) {
1439 peek_seq = tp->copied_seq;
1443 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
1445 #ifdef CONFIG_NET_DMA
1446 tp->ucopy.dma_chan = NULL;
1448 skb = skb_peek_tail(&sk->sk_receive_queue);
1453 available = TCP_SKB_CB(skb)->seq + skb->len - (*seq);
1454 if ((available < target) &&
1455 (len > sysctl_tcp_dma_copybreak) && !(flags & MSG_PEEK) &&
1456 !sysctl_tcp_low_latency &&
1457 dma_find_channel(DMA_MEMCPY)) {
1458 preempt_enable_no_resched();
1459 tp->ucopy.pinned_list =
1460 dma_pin_iovec_pages(msg->msg_iov, len);
1462 preempt_enable_no_resched();
1470 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
1471 if (tp->urg_data && tp->urg_seq == *seq) {
1474 if (signal_pending(current)) {
1475 copied = timeo ? sock_intr_errno(timeo) : -EAGAIN;
1480 /* Next get a buffer. */
1482 skb_queue_walk(&sk->sk_receive_queue, skb) {
1483 /* Now that we have two receive queues this
1486 if (WARN(before(*seq, TCP_SKB_CB(skb)->seq),
1487 KERN_INFO "recvmsg bug: copied %X "
1488 "seq %X rcvnxt %X fl %X\n", *seq,
1489 TCP_SKB_CB(skb)->seq, tp->rcv_nxt,
1493 offset = *seq - TCP_SKB_CB(skb)->seq;
1494 if (tcp_hdr(skb)->syn)
1496 if (offset < skb->len)
1498 if (tcp_hdr(skb)->fin)
1500 WARN(!(flags & MSG_PEEK), KERN_INFO "recvmsg bug 2: "
1501 "copied %X seq %X rcvnxt %X fl %X\n",
1502 *seq, TCP_SKB_CB(skb)->seq,
1503 tp->rcv_nxt, flags);
1506 /* Well, if we have backlog, try to process it now yet. */
1508 if (copied >= target && !sk->sk_backlog.tail)
1513 sk->sk_state == TCP_CLOSE ||
1514 (sk->sk_shutdown & RCV_SHUTDOWN) ||
1516 signal_pending(current))
1519 if (sock_flag(sk, SOCK_DONE))
1523 copied = sock_error(sk);
1527 if (sk->sk_shutdown & RCV_SHUTDOWN)
1530 if (sk->sk_state == TCP_CLOSE) {
1531 if (!sock_flag(sk, SOCK_DONE)) {
1532 /* This occurs when user tries to read
1533 * from never connected socket.
1546 if (signal_pending(current)) {
1547 copied = sock_intr_errno(timeo);
1552 tcp_cleanup_rbuf(sk, copied);
1554 if (!sysctl_tcp_low_latency && tp->ucopy.task == user_recv) {
1555 /* Install new reader */
1556 if (!user_recv && !(flags & (MSG_TRUNC | MSG_PEEK))) {
1557 user_recv = current;
1558 tp->ucopy.task = user_recv;
1559 tp->ucopy.iov = msg->msg_iov;
1562 tp->ucopy.len = len;
1564 WARN_ON(tp->copied_seq != tp->rcv_nxt &&
1565 !(flags & (MSG_PEEK | MSG_TRUNC)));
1567 /* Ugly... If prequeue is not empty, we have to
1568 * process it before releasing socket, otherwise
1569 * order will be broken at second iteration.
1570 * More elegant solution is required!!!
1572 * Look: we have the following (pseudo)queues:
1574 * 1. packets in flight
1579 * Each queue can be processed only if the next ones
1580 * are empty. At this point we have empty receive_queue.
1581 * But prequeue _can_ be not empty after 2nd iteration,
1582 * when we jumped to start of loop because backlog
1583 * processing added something to receive_queue.
1584 * We cannot release_sock(), because backlog contains
1585 * packets arrived _after_ prequeued ones.
1587 * Shortly, algorithm is clear --- to process all
1588 * the queues in order. We could make it more directly,
1589 * requeueing packets from backlog to prequeue, if
1590 * is not empty. It is more elegant, but eats cycles,
1593 if (!skb_queue_empty(&tp->ucopy.prequeue))
1596 /* __ Set realtime policy in scheduler __ */
1599 #ifdef CONFIG_NET_DMA
1600 if (tp->ucopy.dma_chan)
1601 dma_async_memcpy_issue_pending(tp->ucopy.dma_chan);
1603 if (copied >= target) {
1604 /* Do not sleep, just process backlog. */
1608 sk_wait_data(sk, &timeo);
1610 #ifdef CONFIG_NET_DMA
1611 tcp_service_net_dma(sk, false); /* Don't block */
1612 tp->ucopy.wakeup = 0;
1618 /* __ Restore normal policy in scheduler __ */
1620 if ((chunk = len - tp->ucopy.len) != 0) {
1621 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMBACKLOG, chunk);
1626 if (tp->rcv_nxt == tp->copied_seq &&
1627 !skb_queue_empty(&tp->ucopy.prequeue)) {
1629 tcp_prequeue_process(sk);
1631 if ((chunk = len - tp->ucopy.len) != 0) {
1632 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1638 if ((flags & MSG_PEEK) &&
1639 (peek_seq - copied - urg_hole != tp->copied_seq)) {
1640 if (net_ratelimit())
1641 printk(KERN_DEBUG "TCP(%s:%d): Application bug, race in MSG_PEEK.\n",
1642 current->comm, task_pid_nr(current));
1643 peek_seq = tp->copied_seq;
1648 /* Ok so how much can we use? */
1649 used = skb->len - offset;
1653 /* Do we have urgent data here? */
1655 u32 urg_offset = tp->urg_seq - *seq;
1656 if (urg_offset < used) {
1658 if (!sock_flag(sk, SOCK_URGINLINE)) {
1671 if (!(flags & MSG_TRUNC)) {
1672 #ifdef CONFIG_NET_DMA
1673 if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list)
1674 tp->ucopy.dma_chan = dma_find_channel(DMA_MEMCPY);
1676 if (tp->ucopy.dma_chan) {
1677 tp->ucopy.dma_cookie = dma_skb_copy_datagram_iovec(
1678 tp->ucopy.dma_chan, skb, offset,
1680 tp->ucopy.pinned_list);
1682 if (tp->ucopy.dma_cookie < 0) {
1684 printk(KERN_ALERT "dma_cookie < 0\n");
1686 /* Exception. Bailout! */
1692 dma_async_memcpy_issue_pending(tp->ucopy.dma_chan);
1694 if ((offset + used) == skb->len)
1700 err = skb_copy_datagram_iovec(skb, offset,
1701 msg->msg_iov, used);
1703 /* Exception. Bailout! */
1715 tcp_rcv_space_adjust(sk);
1718 if (tp->urg_data && after(tp->copied_seq, tp->urg_seq)) {
1720 tcp_fast_path_check(sk);
1722 if (used + offset < skb->len)
1725 if (tcp_hdr(skb)->fin)
1727 if (!(flags & MSG_PEEK)) {
1728 sk_eat_skb(sk, skb, copied_early);
1734 /* Process the FIN. */
1736 if (!(flags & MSG_PEEK)) {
1737 sk_eat_skb(sk, skb, copied_early);
1744 if (!skb_queue_empty(&tp->ucopy.prequeue)) {
1747 tp->ucopy.len = copied > 0 ? len : 0;
1749 tcp_prequeue_process(sk);
1751 if (copied > 0 && (chunk = len - tp->ucopy.len) != 0) {
1752 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1758 tp->ucopy.task = NULL;
1762 #ifdef CONFIG_NET_DMA
1763 tcp_service_net_dma(sk, true); /* Wait for queue to drain */
1764 tp->ucopy.dma_chan = NULL;
1766 if (tp->ucopy.pinned_list) {
1767 dma_unpin_iovec_pages(tp->ucopy.pinned_list);
1768 tp->ucopy.pinned_list = NULL;
1772 /* According to UNIX98, msg_name/msg_namelen are ignored
1773 * on connected socket. I was just happy when found this 8) --ANK
1776 /* Clean up data we have read: This will do ACK frames. */
1777 tcp_cleanup_rbuf(sk, copied);
1779 TCP_CHECK_TIMER(sk);
1783 uid_stat_tcp_rcv(current_uid(), copied);
1787 TCP_CHECK_TIMER(sk);
1792 err = tcp_recv_urg(sk, msg, len, flags);
1794 uid_stat_tcp_rcv(current_uid(), err);
1797 EXPORT_SYMBOL(tcp_recvmsg);
1799 void tcp_set_state(struct sock *sk, int state)
1801 int oldstate = sk->sk_state;
1804 case TCP_ESTABLISHED:
1805 if (oldstate != TCP_ESTABLISHED)
1806 TCP_INC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
1810 if (oldstate == TCP_CLOSE_WAIT || oldstate == TCP_ESTABLISHED)
1811 TCP_INC_STATS(sock_net(sk), TCP_MIB_ESTABRESETS);
1813 sk->sk_prot->unhash(sk);
1814 if (inet_csk(sk)->icsk_bind_hash &&
1815 !(sk->sk_userlocks & SOCK_BINDPORT_LOCK))
1819 if (oldstate == TCP_ESTABLISHED)
1820 TCP_DEC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
1823 /* Change state AFTER socket is unhashed to avoid closed
1824 * socket sitting in hash tables.
1826 sk->sk_state = state;
1829 SOCK_DEBUG(sk, "TCP sk=%p, State %s -> %s\n", sk, statename[oldstate], statename[state]);
1832 EXPORT_SYMBOL_GPL(tcp_set_state);
1835 * State processing on a close. This implements the state shift for
1836 * sending our FIN frame. Note that we only send a FIN for some
1837 * states. A shutdown() may have already sent the FIN, or we may be
1841 static const unsigned char new_state[16] = {
1842 /* current state: new state: action: */
1843 /* (Invalid) */ TCP_CLOSE,
1844 /* TCP_ESTABLISHED */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1845 /* TCP_SYN_SENT */ TCP_CLOSE,
1846 /* TCP_SYN_RECV */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1847 /* TCP_FIN_WAIT1 */ TCP_FIN_WAIT1,
1848 /* TCP_FIN_WAIT2 */ TCP_FIN_WAIT2,
1849 /* TCP_TIME_WAIT */ TCP_CLOSE,
1850 /* TCP_CLOSE */ TCP_CLOSE,
1851 /* TCP_CLOSE_WAIT */ TCP_LAST_ACK | TCP_ACTION_FIN,
1852 /* TCP_LAST_ACK */ TCP_LAST_ACK,
1853 /* TCP_LISTEN */ TCP_CLOSE,
1854 /* TCP_CLOSING */ TCP_CLOSING,
1857 static int tcp_close_state(struct sock *sk)
1859 int next = (int)new_state[sk->sk_state];
1860 int ns = next & TCP_STATE_MASK;
1862 tcp_set_state(sk, ns);
1864 return next & TCP_ACTION_FIN;
1868 * Shutdown the sending side of a connection. Much like close except
1869 * that we don't receive shut down or sock_set_flag(sk, SOCK_DEAD).
1872 void tcp_shutdown(struct sock *sk, int how)
1874 /* We need to grab some memory, and put together a FIN,
1875 * and then put it into the queue to be sent.
1876 * Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
1878 if (!(how & SEND_SHUTDOWN))
1881 /* If we've already sent a FIN, or it's a closed state, skip this. */
1882 if ((1 << sk->sk_state) &
1883 (TCPF_ESTABLISHED | TCPF_SYN_SENT |
1884 TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) {
1885 /* Clear out any half completed packets. FIN if needed. */
1886 if (tcp_close_state(sk))
1890 EXPORT_SYMBOL(tcp_shutdown);
1892 void tcp_close(struct sock *sk, long timeout)
1894 struct sk_buff *skb;
1895 int data_was_unread = 0;
1899 sk->sk_shutdown = SHUTDOWN_MASK;
1901 if (sk->sk_state == TCP_LISTEN) {
1902 tcp_set_state(sk, TCP_CLOSE);
1905 inet_csk_listen_stop(sk);
1907 goto adjudge_to_death;
1910 /* We need to flush the recv. buffs. We do this only on the
1911 * descriptor close, not protocol-sourced closes, because the
1912 * reader process may not have drained the data yet!
1914 while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) {
1915 u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq -
1917 data_was_unread += len;
1923 /* If socket has been already reset (e.g. in tcp_reset()) - kill it. */
1924 if (sk->sk_state == TCP_CLOSE)
1925 goto adjudge_to_death;
1927 /* As outlined in RFC 2525, section 2.17, we send a RST here because
1928 * data was lost. To witness the awful effects of the old behavior of
1929 * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk
1930 * GET in an FTP client, suspend the process, wait for the client to
1931 * advertise a zero window, then kill -9 the FTP client, wheee...
1932 * Note: timeout is always zero in such a case.
1934 if (data_was_unread) {
1935 /* Unread data was tossed, zap the connection. */
1936 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONCLOSE);
1937 tcp_set_state(sk, TCP_CLOSE);
1938 tcp_send_active_reset(sk, sk->sk_allocation);
1939 } else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) {
1940 /* Check zero linger _after_ checking for unread data. */
1941 sk->sk_prot->disconnect(sk, 0);
1942 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
1943 } else if (tcp_close_state(sk)) {
1944 /* We FIN if the application ate all the data before
1945 * zapping the connection.
1948 /* RED-PEN. Formally speaking, we have broken TCP state
1949 * machine. State transitions:
1951 * TCP_ESTABLISHED -> TCP_FIN_WAIT1
1952 * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible)
1953 * TCP_CLOSE_WAIT -> TCP_LAST_ACK
1955 * are legal only when FIN has been sent (i.e. in window),
1956 * rather than queued out of window. Purists blame.
1958 * F.e. "RFC state" is ESTABLISHED,
1959 * if Linux state is FIN-WAIT-1, but FIN is still not sent.
1961 * The visible declinations are that sometimes
1962 * we enter time-wait state, when it is not required really
1963 * (harmless), do not send active resets, when they are
1964 * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
1965 * they look as CLOSING or LAST_ACK for Linux)
1966 * Probably, I missed some more holelets.
1972 sk_stream_wait_close(sk, timeout);
1975 state = sk->sk_state;
1979 /* It is the last release_sock in its life. It will remove backlog. */
1983 /* Now socket is owned by kernel and we acquire BH lock
1984 to finish close. No need to check for user refs.
1988 WARN_ON(sock_owned_by_user(sk));
1990 percpu_counter_inc(sk->sk_prot->orphan_count);
1992 /* Have we already been destroyed by a softirq or backlog? */
1993 if (state != TCP_CLOSE && sk->sk_state == TCP_CLOSE)
1996 /* This is a (useful) BSD violating of the RFC. There is a
1997 * problem with TCP as specified in that the other end could
1998 * keep a socket open forever with no application left this end.
1999 * We use a 3 minute timeout (about the same as BSD) then kill
2000 * our end. If they send after that then tough - BUT: long enough
2001 * that we won't make the old 4*rto = almost no time - whoops
2004 * Nope, it was not mistake. It is really desired behaviour
2005 * f.e. on http servers, when such sockets are useless, but
2006 * consume significant resources. Let's do it with special
2007 * linger2 option. --ANK
2010 if (sk->sk_state == TCP_FIN_WAIT2) {
2011 struct tcp_sock *tp = tcp_sk(sk);
2012 if (tp->linger2 < 0) {
2013 tcp_set_state(sk, TCP_CLOSE);
2014 tcp_send_active_reset(sk, GFP_ATOMIC);
2015 NET_INC_STATS_BH(sock_net(sk),
2016 LINUX_MIB_TCPABORTONLINGER);
2018 const int tmo = tcp_fin_time(sk);
2020 if (tmo > TCP_TIMEWAIT_LEN) {
2021 inet_csk_reset_keepalive_timer(sk,
2022 tmo - TCP_TIMEWAIT_LEN);
2024 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
2029 if (sk->sk_state != TCP_CLOSE) {
2031 if (tcp_too_many_orphans(sk, 0)) {
2032 if (net_ratelimit())
2033 printk(KERN_INFO "TCP: too many of orphaned "
2035 tcp_set_state(sk, TCP_CLOSE);
2036 tcp_send_active_reset(sk, GFP_ATOMIC);
2037 NET_INC_STATS_BH(sock_net(sk),
2038 LINUX_MIB_TCPABORTONMEMORY);
2042 if (sk->sk_state == TCP_CLOSE)
2043 inet_csk_destroy_sock(sk);
2044 /* Otherwise, socket is reprieved until protocol close. */
2051 EXPORT_SYMBOL(tcp_close);
2053 /* These states need RST on ABORT according to RFC793 */
2055 static inline int tcp_need_reset(int state)
2057 return (1 << state) &
2058 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 |
2059 TCPF_FIN_WAIT2 | TCPF_SYN_RECV);
2062 int tcp_disconnect(struct sock *sk, int flags)
2064 struct inet_sock *inet = inet_sk(sk);
2065 struct inet_connection_sock *icsk = inet_csk(sk);
2066 struct tcp_sock *tp = tcp_sk(sk);
2068 int old_state = sk->sk_state;
2070 if (old_state != TCP_CLOSE)
2071 tcp_set_state(sk, TCP_CLOSE);
2073 /* ABORT function of RFC793 */
2074 if (old_state == TCP_LISTEN) {
2075 inet_csk_listen_stop(sk);
2076 } else if (tcp_need_reset(old_state) ||
2077 (tp->snd_nxt != tp->write_seq &&
2078 (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK))) {
2079 /* The last check adjusts for discrepancy of Linux wrt. RFC
2082 tcp_send_active_reset(sk, gfp_any());
2083 sk->sk_err = ECONNRESET;
2084 } else if (old_state == TCP_SYN_SENT)
2085 sk->sk_err = ECONNRESET;
2087 tcp_clear_xmit_timers(sk);
2088 __skb_queue_purge(&sk->sk_receive_queue);
2089 tcp_write_queue_purge(sk);
2090 __skb_queue_purge(&tp->out_of_order_queue);
2091 #ifdef CONFIG_NET_DMA
2092 __skb_queue_purge(&sk->sk_async_wait_queue);
2095 inet->inet_dport = 0;
2097 if (!(sk->sk_userlocks & SOCK_BINDADDR_LOCK))
2098 inet_reset_saddr(sk);
2100 sk->sk_shutdown = 0;
2101 sock_reset_flag(sk, SOCK_DONE);
2103 if ((tp->write_seq += tp->max_window + 2) == 0)
2105 icsk->icsk_backoff = 0;
2107 icsk->icsk_probes_out = 0;
2108 tp->packets_out = 0;
2109 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
2110 tp->snd_cwnd_cnt = 0;
2111 tp->bytes_acked = 0;
2112 tp->window_clamp = 0;
2113 tcp_set_ca_state(sk, TCP_CA_Open);
2114 tcp_clear_retrans(tp);
2115 inet_csk_delack_init(sk);
2116 tcp_init_send_head(sk);
2117 memset(&tp->rx_opt, 0, sizeof(tp->rx_opt));
2120 WARN_ON(inet->inet_num && !icsk->icsk_bind_hash);
2122 sk->sk_error_report(sk);
2125 EXPORT_SYMBOL(tcp_disconnect);
2128 * Socket option code for TCP.
2130 static int do_tcp_setsockopt(struct sock *sk, int level,
2131 int optname, char __user *optval, unsigned int optlen)
2133 struct tcp_sock *tp = tcp_sk(sk);
2134 struct inet_connection_sock *icsk = inet_csk(sk);
2138 /* These are data/string values, all the others are ints */
2140 case TCP_CONGESTION: {
2141 char name[TCP_CA_NAME_MAX];
2146 val = strncpy_from_user(name, optval,
2147 min_t(long, TCP_CA_NAME_MAX-1, optlen));
2153 err = tcp_set_congestion_control(sk, name);
2157 case TCP_COOKIE_TRANSACTIONS: {
2158 struct tcp_cookie_transactions ctd;
2159 struct tcp_cookie_values *cvp = NULL;
2161 if (sizeof(ctd) > optlen)
2163 if (copy_from_user(&ctd, optval, sizeof(ctd)))
2166 if (ctd.tcpct_used > sizeof(ctd.tcpct_value) ||
2167 ctd.tcpct_s_data_desired > TCP_MSS_DESIRED)
2170 if (ctd.tcpct_cookie_desired == 0) {
2171 /* default to global value */
2172 } else if ((0x1 & ctd.tcpct_cookie_desired) ||
2173 ctd.tcpct_cookie_desired > TCP_COOKIE_MAX ||
2174 ctd.tcpct_cookie_desired < TCP_COOKIE_MIN) {
2178 if (TCP_COOKIE_OUT_NEVER & ctd.tcpct_flags) {
2179 /* Supercedes all other values */
2181 if (tp->cookie_values != NULL) {
2182 kref_put(&tp->cookie_values->kref,
2183 tcp_cookie_values_release);
2184 tp->cookie_values = NULL;
2186 tp->rx_opt.cookie_in_always = 0; /* false */
2187 tp->rx_opt.cookie_out_never = 1; /* true */
2192 /* Allocate ancillary memory before locking.
2194 if (ctd.tcpct_used > 0 ||
2195 (tp->cookie_values == NULL &&
2196 (sysctl_tcp_cookie_size > 0 ||
2197 ctd.tcpct_cookie_desired > 0 ||
2198 ctd.tcpct_s_data_desired > 0))) {
2199 cvp = kzalloc(sizeof(*cvp) + ctd.tcpct_used,
2204 kref_init(&cvp->kref);
2207 tp->rx_opt.cookie_in_always =
2208 (TCP_COOKIE_IN_ALWAYS & ctd.tcpct_flags);
2209 tp->rx_opt.cookie_out_never = 0; /* false */
2211 if (tp->cookie_values != NULL) {
2213 /* Changed values are recorded by a changed
2214 * pointer, ensuring the cookie will differ,
2215 * without separately hashing each value later.
2217 kref_put(&tp->cookie_values->kref,
2218 tcp_cookie_values_release);
2220 cvp = tp->cookie_values;
2225 cvp->cookie_desired = ctd.tcpct_cookie_desired;
2227 if (ctd.tcpct_used > 0) {
2228 memcpy(cvp->s_data_payload, ctd.tcpct_value,
2230 cvp->s_data_desired = ctd.tcpct_used;
2231 cvp->s_data_constant = 1; /* true */
2233 /* No constant payload data. */
2234 cvp->s_data_desired = ctd.tcpct_s_data_desired;
2235 cvp->s_data_constant = 0; /* false */
2238 tp->cookie_values = cvp;
2248 if (optlen < sizeof(int))
2251 if (get_user(val, (int __user *)optval))
2258 /* Values greater than interface MTU won't take effect. However
2259 * at the point when this call is done we typically don't yet
2260 * know which interface is going to be used */
2261 if (val < 8 || val > MAX_TCP_WINDOW) {
2265 tp->rx_opt.user_mss = val;
2270 /* TCP_NODELAY is weaker than TCP_CORK, so that
2271 * this option on corked socket is remembered, but
2272 * it is not activated until cork is cleared.
2274 * However, when TCP_NODELAY is set we make
2275 * an explicit push, which overrides even TCP_CORK
2276 * for currently queued segments.
2278 tp->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH;
2279 tcp_push_pending_frames(sk);
2281 tp->nonagle &= ~TCP_NAGLE_OFF;
2285 case TCP_THIN_LINEAR_TIMEOUTS:
2286 if (val < 0 || val > 1)
2292 case TCP_THIN_DUPACK:
2293 if (val < 0 || val > 1)
2296 tp->thin_dupack = val;
2300 /* When set indicates to always queue non-full frames.
2301 * Later the user clears this option and we transmit
2302 * any pending partial frames in the queue. This is
2303 * meant to be used alongside sendfile() to get properly
2304 * filled frames when the user (for example) must write
2305 * out headers with a write() call first and then use
2306 * sendfile to send out the data parts.
2308 * TCP_CORK can be set together with TCP_NODELAY and it is
2309 * stronger than TCP_NODELAY.
2312 tp->nonagle |= TCP_NAGLE_CORK;
2314 tp->nonagle &= ~TCP_NAGLE_CORK;
2315 if (tp->nonagle&TCP_NAGLE_OFF)
2316 tp->nonagle |= TCP_NAGLE_PUSH;
2317 tcp_push_pending_frames(sk);
2322 if (val < 1 || val > MAX_TCP_KEEPIDLE)
2325 tp->keepalive_time = val * HZ;
2326 if (sock_flag(sk, SOCK_KEEPOPEN) &&
2327 !((1 << sk->sk_state) &
2328 (TCPF_CLOSE | TCPF_LISTEN))) {
2329 u32 elapsed = keepalive_time_elapsed(tp);
2330 if (tp->keepalive_time > elapsed)
2331 elapsed = tp->keepalive_time - elapsed;
2334 inet_csk_reset_keepalive_timer(sk, elapsed);
2339 if (val < 1 || val > MAX_TCP_KEEPINTVL)
2342 tp->keepalive_intvl = val * HZ;
2345 if (val < 1 || val > MAX_TCP_KEEPCNT)
2348 tp->keepalive_probes = val;
2351 if (val < 1 || val > MAX_TCP_SYNCNT)
2354 icsk->icsk_syn_retries = val;
2360 else if (val > sysctl_tcp_fin_timeout / HZ)
2363 tp->linger2 = val * HZ;
2366 case TCP_DEFER_ACCEPT:
2367 /* Translate value in seconds to number of retransmits */
2368 icsk->icsk_accept_queue.rskq_defer_accept =
2369 secs_to_retrans(val, TCP_TIMEOUT_INIT / HZ,
2373 case TCP_WINDOW_CLAMP:
2375 if (sk->sk_state != TCP_CLOSE) {
2379 tp->window_clamp = 0;
2381 tp->window_clamp = val < SOCK_MIN_RCVBUF / 2 ?
2382 SOCK_MIN_RCVBUF / 2 : val;
2387 icsk->icsk_ack.pingpong = 1;
2389 icsk->icsk_ack.pingpong = 0;
2390 if ((1 << sk->sk_state) &
2391 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) &&
2392 inet_csk_ack_scheduled(sk)) {
2393 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
2394 tcp_cleanup_rbuf(sk, 1);
2396 icsk->icsk_ack.pingpong = 1;
2401 #ifdef CONFIG_TCP_MD5SIG
2403 /* Read the IP->Key mappings from userspace */
2404 err = tp->af_specific->md5_parse(sk, optval, optlen);
2417 int tcp_setsockopt(struct sock *sk, int level, int optname, char __user *optval,
2418 unsigned int optlen)
2420 struct inet_connection_sock *icsk = inet_csk(sk);
2422 if (level != SOL_TCP)
2423 return icsk->icsk_af_ops->setsockopt(sk, level, optname,
2425 return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2427 EXPORT_SYMBOL(tcp_setsockopt);
2429 #ifdef CONFIG_COMPAT
2430 int compat_tcp_setsockopt(struct sock *sk, int level, int optname,
2431 char __user *optval, unsigned int optlen)
2433 if (level != SOL_TCP)
2434 return inet_csk_compat_setsockopt(sk, level, optname,
2436 return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2438 EXPORT_SYMBOL(compat_tcp_setsockopt);
2441 /* Return information about state of tcp endpoint in API format. */
2442 void tcp_get_info(struct sock *sk, struct tcp_info *info)
2444 struct tcp_sock *tp = tcp_sk(sk);
2445 const struct inet_connection_sock *icsk = inet_csk(sk);
2446 u32 now = tcp_time_stamp;
2448 memset(info, 0, sizeof(*info));
2450 info->tcpi_state = sk->sk_state;
2451 info->tcpi_ca_state = icsk->icsk_ca_state;
2452 info->tcpi_retransmits = icsk->icsk_retransmits;
2453 info->tcpi_probes = icsk->icsk_probes_out;
2454 info->tcpi_backoff = icsk->icsk_backoff;
2456 if (tp->rx_opt.tstamp_ok)
2457 info->tcpi_options |= TCPI_OPT_TIMESTAMPS;
2458 if (tcp_is_sack(tp))
2459 info->tcpi_options |= TCPI_OPT_SACK;
2460 if (tp->rx_opt.wscale_ok) {
2461 info->tcpi_options |= TCPI_OPT_WSCALE;
2462 info->tcpi_snd_wscale = tp->rx_opt.snd_wscale;
2463 info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale;
2466 if (tp->ecn_flags&TCP_ECN_OK)
2467 info->tcpi_options |= TCPI_OPT_ECN;
2469 info->tcpi_rto = jiffies_to_usecs(icsk->icsk_rto);
2470 info->tcpi_ato = jiffies_to_usecs(icsk->icsk_ack.ato);
2471 info->tcpi_snd_mss = tp->mss_cache;
2472 info->tcpi_rcv_mss = icsk->icsk_ack.rcv_mss;
2474 if (sk->sk_state == TCP_LISTEN) {
2475 info->tcpi_unacked = sk->sk_ack_backlog;
2476 info->tcpi_sacked = sk->sk_max_ack_backlog;
2478 info->tcpi_unacked = tp->packets_out;
2479 info->tcpi_sacked = tp->sacked_out;
2481 info->tcpi_lost = tp->lost_out;
2482 info->tcpi_retrans = tp->retrans_out;
2483 info->tcpi_fackets = tp->fackets_out;
2485 info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime);
2486 info->tcpi_last_data_recv = jiffies_to_msecs(now - icsk->icsk_ack.lrcvtime);
2487 info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp);
2489 info->tcpi_pmtu = icsk->icsk_pmtu_cookie;
2490 info->tcpi_rcv_ssthresh = tp->rcv_ssthresh;
2491 info->tcpi_rtt = jiffies_to_usecs(tp->srtt)>>3;
2492 info->tcpi_rttvar = jiffies_to_usecs(tp->mdev)>>2;
2493 info->tcpi_snd_ssthresh = tp->snd_ssthresh;
2494 info->tcpi_snd_cwnd = tp->snd_cwnd;
2495 info->tcpi_advmss = tp->advmss;
2496 info->tcpi_reordering = tp->reordering;
2498 info->tcpi_rcv_rtt = jiffies_to_usecs(tp->rcv_rtt_est.rtt)>>3;
2499 info->tcpi_rcv_space = tp->rcvq_space.space;
2501 info->tcpi_total_retrans = tp->total_retrans;
2503 EXPORT_SYMBOL_GPL(tcp_get_info);
2505 static int do_tcp_getsockopt(struct sock *sk, int level,
2506 int optname, char __user *optval, int __user *optlen)
2508 struct inet_connection_sock *icsk = inet_csk(sk);
2509 struct tcp_sock *tp = tcp_sk(sk);
2512 if (get_user(len, optlen))
2515 len = min_t(unsigned int, len, sizeof(int));
2522 val = tp->mss_cache;
2523 if (!val && ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
2524 val = tp->rx_opt.user_mss;
2527 val = !!(tp->nonagle&TCP_NAGLE_OFF);
2530 val = !!(tp->nonagle&TCP_NAGLE_CORK);
2533 val = keepalive_time_when(tp) / HZ;
2536 val = keepalive_intvl_when(tp) / HZ;
2539 val = keepalive_probes(tp);
2542 val = icsk->icsk_syn_retries ? : sysctl_tcp_syn_retries;
2547 val = (val ? : sysctl_tcp_fin_timeout) / HZ;
2549 case TCP_DEFER_ACCEPT:
2550 val = retrans_to_secs(icsk->icsk_accept_queue.rskq_defer_accept,
2551 TCP_TIMEOUT_INIT / HZ, TCP_RTO_MAX / HZ);
2553 case TCP_WINDOW_CLAMP:
2554 val = tp->window_clamp;
2557 struct tcp_info info;
2559 if (get_user(len, optlen))
2562 tcp_get_info(sk, &info);
2564 len = min_t(unsigned int, len, sizeof(info));
2565 if (put_user(len, optlen))
2567 if (copy_to_user(optval, &info, len))
2572 val = !icsk->icsk_ack.pingpong;
2575 case TCP_CONGESTION:
2576 if (get_user(len, optlen))
2578 len = min_t(unsigned int, len, TCP_CA_NAME_MAX);
2579 if (put_user(len, optlen))
2581 if (copy_to_user(optval, icsk->icsk_ca_ops->name, len))
2585 case TCP_COOKIE_TRANSACTIONS: {
2586 struct tcp_cookie_transactions ctd;
2587 struct tcp_cookie_values *cvp = tp->cookie_values;
2589 if (get_user(len, optlen))
2591 if (len < sizeof(ctd))
2594 memset(&ctd, 0, sizeof(ctd));
2595 ctd.tcpct_flags = (tp->rx_opt.cookie_in_always ?
2596 TCP_COOKIE_IN_ALWAYS : 0)
2597 | (tp->rx_opt.cookie_out_never ?
2598 TCP_COOKIE_OUT_NEVER : 0);
2601 ctd.tcpct_flags |= (cvp->s_data_in ?
2603 | (cvp->s_data_out ?
2604 TCP_S_DATA_OUT : 0);
2606 ctd.tcpct_cookie_desired = cvp->cookie_desired;
2607 ctd.tcpct_s_data_desired = cvp->s_data_desired;
2609 memcpy(&ctd.tcpct_value[0], &cvp->cookie_pair[0],
2610 cvp->cookie_pair_size);
2611 ctd.tcpct_used = cvp->cookie_pair_size;
2614 if (put_user(sizeof(ctd), optlen))
2616 if (copy_to_user(optval, &ctd, sizeof(ctd)))
2620 case TCP_THIN_LINEAR_TIMEOUTS:
2623 case TCP_THIN_DUPACK:
2624 val = tp->thin_dupack;
2627 return -ENOPROTOOPT;
2630 if (put_user(len, optlen))
2632 if (copy_to_user(optval, &val, len))
2637 int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval,
2640 struct inet_connection_sock *icsk = inet_csk(sk);
2642 if (level != SOL_TCP)
2643 return icsk->icsk_af_ops->getsockopt(sk, level, optname,
2645 return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2647 EXPORT_SYMBOL(tcp_getsockopt);
2649 #ifdef CONFIG_COMPAT
2650 int compat_tcp_getsockopt(struct sock *sk, int level, int optname,
2651 char __user *optval, int __user *optlen)
2653 if (level != SOL_TCP)
2654 return inet_csk_compat_getsockopt(sk, level, optname,
2656 return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2658 EXPORT_SYMBOL(compat_tcp_getsockopt);
2661 struct sk_buff *tcp_tso_segment(struct sk_buff *skb, int features)
2663 struct sk_buff *segs = ERR_PTR(-EINVAL);
2668 unsigned int oldlen;
2671 if (!pskb_may_pull(skb, sizeof(*th)))
2675 thlen = th->doff * 4;
2676 if (thlen < sizeof(*th))
2679 if (!pskb_may_pull(skb, thlen))
2682 oldlen = (u16)~skb->len;
2683 __skb_pull(skb, thlen);
2685 mss = skb_shinfo(skb)->gso_size;
2686 if (unlikely(skb->len <= mss))
2689 if (skb_gso_ok(skb, features | NETIF_F_GSO_ROBUST)) {
2690 /* Packet is from an untrusted source, reset gso_segs. */
2691 int type = skb_shinfo(skb)->gso_type;
2699 !(type & (SKB_GSO_TCPV4 | SKB_GSO_TCPV6))))
2702 skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss);
2708 segs = skb_segment(skb, features);
2712 delta = htonl(oldlen + (thlen + mss));
2716 seq = ntohl(th->seq);
2719 th->fin = th->psh = 0;
2721 th->check = ~csum_fold((__force __wsum)((__force u32)th->check +
2722 (__force u32)delta));
2723 if (skb->ip_summed != CHECKSUM_PARTIAL)
2725 csum_fold(csum_partial(skb_transport_header(skb),
2732 th->seq = htonl(seq);
2734 } while (skb->next);
2736 delta = htonl(oldlen + (skb->tail - skb->transport_header) +
2738 th->check = ~csum_fold((__force __wsum)((__force u32)th->check +
2739 (__force u32)delta));
2740 if (skb->ip_summed != CHECKSUM_PARTIAL)
2741 th->check = csum_fold(csum_partial(skb_transport_header(skb),
2747 EXPORT_SYMBOL(tcp_tso_segment);
2749 struct sk_buff **tcp_gro_receive(struct sk_buff **head, struct sk_buff *skb)
2751 struct sk_buff **pp = NULL;
2758 unsigned int mss = 1;
2764 off = skb_gro_offset(skb);
2765 hlen = off + sizeof(*th);
2766 th = skb_gro_header_fast(skb, off);
2767 if (skb_gro_header_hard(skb, hlen)) {
2768 th = skb_gro_header_slow(skb, hlen, off);
2773 thlen = th->doff * 4;
2774 if (thlen < sizeof(*th))
2778 if (skb_gro_header_hard(skb, hlen)) {
2779 th = skb_gro_header_slow(skb, hlen, off);
2784 skb_gro_pull(skb, thlen);
2786 len = skb_gro_len(skb);
2787 flags = tcp_flag_word(th);
2789 for (; (p = *head); head = &p->next) {
2790 if (!NAPI_GRO_CB(p)->same_flow)
2795 if (*(u32 *)&th->source ^ *(u32 *)&th2->source) {
2796 NAPI_GRO_CB(p)->same_flow = 0;
2803 goto out_check_final;
2806 flush = NAPI_GRO_CB(p)->flush;
2807 flush |= (__force int)(flags & TCP_FLAG_CWR);
2808 flush |= (__force int)((flags ^ tcp_flag_word(th2)) &
2809 ~(TCP_FLAG_CWR | TCP_FLAG_FIN | TCP_FLAG_PSH));
2810 flush |= (__force int)(th->ack_seq ^ th2->ack_seq);
2811 for (i = sizeof(*th); i < thlen; i += 4)
2812 flush |= *(u32 *)((u8 *)th + i) ^
2813 *(u32 *)((u8 *)th2 + i);
2815 mss = skb_shinfo(p)->gso_size;
2817 flush |= (len - 1) >= mss;
2818 flush |= (ntohl(th2->seq) + skb_gro_len(p)) ^ ntohl(th->seq);
2820 if (flush || skb_gro_receive(head, skb)) {
2822 goto out_check_final;
2827 tcp_flag_word(th2) |= flags & (TCP_FLAG_FIN | TCP_FLAG_PSH);
2831 flush |= (__force int)(flags & (TCP_FLAG_URG | TCP_FLAG_PSH |
2832 TCP_FLAG_RST | TCP_FLAG_SYN |
2835 if (p && (!NAPI_GRO_CB(skb)->same_flow || flush))
2839 NAPI_GRO_CB(skb)->flush |= flush;
2843 EXPORT_SYMBOL(tcp_gro_receive);
2845 int tcp_gro_complete(struct sk_buff *skb)
2847 struct tcphdr *th = tcp_hdr(skb);
2849 skb->csum_start = skb_transport_header(skb) - skb->head;
2850 skb->csum_offset = offsetof(struct tcphdr, check);
2851 skb->ip_summed = CHECKSUM_PARTIAL;
2853 skb_shinfo(skb)->gso_segs = NAPI_GRO_CB(skb)->count;
2856 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
2860 EXPORT_SYMBOL(tcp_gro_complete);
2862 #ifdef CONFIG_TCP_MD5SIG
2863 static unsigned long tcp_md5sig_users;
2864 static struct tcp_md5sig_pool * __percpu *tcp_md5sig_pool;
2865 static DEFINE_SPINLOCK(tcp_md5sig_pool_lock);
2867 static void __tcp_free_md5sig_pool(struct tcp_md5sig_pool * __percpu *pool)
2870 for_each_possible_cpu(cpu) {
2871 struct tcp_md5sig_pool *p = *per_cpu_ptr(pool, cpu);
2873 if (p->md5_desc.tfm)
2874 crypto_free_hash(p->md5_desc.tfm);
2881 void tcp_free_md5sig_pool(void)
2883 struct tcp_md5sig_pool * __percpu *pool = NULL;
2885 spin_lock_bh(&tcp_md5sig_pool_lock);
2886 if (--tcp_md5sig_users == 0) {
2887 pool = tcp_md5sig_pool;
2888 tcp_md5sig_pool = NULL;
2890 spin_unlock_bh(&tcp_md5sig_pool_lock);
2892 __tcp_free_md5sig_pool(pool);
2894 EXPORT_SYMBOL(tcp_free_md5sig_pool);
2896 static struct tcp_md5sig_pool * __percpu *
2897 __tcp_alloc_md5sig_pool(struct sock *sk)
2900 struct tcp_md5sig_pool * __percpu *pool;
2902 pool = alloc_percpu(struct tcp_md5sig_pool *);
2906 for_each_possible_cpu(cpu) {
2907 struct tcp_md5sig_pool *p;
2908 struct crypto_hash *hash;
2910 p = kzalloc(sizeof(*p), sk->sk_allocation);
2913 *per_cpu_ptr(pool, cpu) = p;
2915 hash = crypto_alloc_hash("md5", 0, CRYPTO_ALG_ASYNC);
2916 if (!hash || IS_ERR(hash))
2919 p->md5_desc.tfm = hash;
2923 __tcp_free_md5sig_pool(pool);
2927 struct tcp_md5sig_pool * __percpu *tcp_alloc_md5sig_pool(struct sock *sk)
2929 struct tcp_md5sig_pool * __percpu *pool;
2933 spin_lock_bh(&tcp_md5sig_pool_lock);
2934 pool = tcp_md5sig_pool;
2935 if (tcp_md5sig_users++ == 0) {
2937 spin_unlock_bh(&tcp_md5sig_pool_lock);
2940 spin_unlock_bh(&tcp_md5sig_pool_lock);
2944 spin_unlock_bh(&tcp_md5sig_pool_lock);
2947 /* we cannot hold spinlock here because this may sleep. */
2948 struct tcp_md5sig_pool * __percpu *p;
2950 p = __tcp_alloc_md5sig_pool(sk);
2951 spin_lock_bh(&tcp_md5sig_pool_lock);
2954 spin_unlock_bh(&tcp_md5sig_pool_lock);
2957 pool = tcp_md5sig_pool;
2959 /* oops, it has already been assigned. */
2960 spin_unlock_bh(&tcp_md5sig_pool_lock);
2961 __tcp_free_md5sig_pool(p);
2963 tcp_md5sig_pool = pool = p;
2964 spin_unlock_bh(&tcp_md5sig_pool_lock);
2969 EXPORT_SYMBOL(tcp_alloc_md5sig_pool);
2973 * tcp_get_md5sig_pool - get md5sig_pool for this user
2975 * We use percpu structure, so if we succeed, we exit with preemption
2976 * and BH disabled, to make sure another thread or softirq handling
2977 * wont try to get same context.
2979 struct tcp_md5sig_pool *tcp_get_md5sig_pool(void)
2981 struct tcp_md5sig_pool * __percpu *p;
2985 spin_lock(&tcp_md5sig_pool_lock);
2986 p = tcp_md5sig_pool;
2989 spin_unlock(&tcp_md5sig_pool_lock);
2992 return *this_cpu_ptr(p);
2997 EXPORT_SYMBOL(tcp_get_md5sig_pool);
2999 void tcp_put_md5sig_pool(void)
3002 tcp_free_md5sig_pool();
3004 EXPORT_SYMBOL(tcp_put_md5sig_pool);
3006 int tcp_md5_hash_header(struct tcp_md5sig_pool *hp,
3009 struct scatterlist sg;
3012 __sum16 old_checksum = th->check;
3014 /* options aren't included in the hash */
3015 sg_init_one(&sg, th, sizeof(struct tcphdr));
3016 err = crypto_hash_update(&hp->md5_desc, &sg, sizeof(struct tcphdr));
3017 th->check = old_checksum;
3020 EXPORT_SYMBOL(tcp_md5_hash_header);
3022 int tcp_md5_hash_skb_data(struct tcp_md5sig_pool *hp,
3023 struct sk_buff *skb, unsigned header_len)
3025 struct scatterlist sg;
3026 const struct tcphdr *tp = tcp_hdr(skb);
3027 struct hash_desc *desc = &hp->md5_desc;
3029 const unsigned head_data_len = skb_headlen(skb) > header_len ?
3030 skb_headlen(skb) - header_len : 0;
3031 const struct skb_shared_info *shi = skb_shinfo(skb);
3032 struct sk_buff *frag_iter;
3034 sg_init_table(&sg, 1);
3036 sg_set_buf(&sg, ((u8 *) tp) + header_len, head_data_len);
3037 if (crypto_hash_update(desc, &sg, head_data_len))
3040 for (i = 0; i < shi->nr_frags; ++i) {
3041 const struct skb_frag_struct *f = &shi->frags[i];
3042 sg_set_page(&sg, f->page, f->size, f->page_offset);
3043 if (crypto_hash_update(desc, &sg, f->size))
3047 skb_walk_frags(skb, frag_iter)
3048 if (tcp_md5_hash_skb_data(hp, frag_iter, 0))
3053 EXPORT_SYMBOL(tcp_md5_hash_skb_data);
3055 int tcp_md5_hash_key(struct tcp_md5sig_pool *hp, struct tcp_md5sig_key *key)
3057 struct scatterlist sg;
3059 sg_init_one(&sg, key->key, key->keylen);
3060 return crypto_hash_update(&hp->md5_desc, &sg, key->keylen);
3062 EXPORT_SYMBOL(tcp_md5_hash_key);
3067 * Each Responder maintains up to two secret values concurrently for
3068 * efficient secret rollover. Each secret value has 4 states:
3070 * Generating. (tcp_secret_generating != tcp_secret_primary)
3071 * Generates new Responder-Cookies, but not yet used for primary
3072 * verification. This is a short-term state, typically lasting only
3073 * one round trip time (RTT).
3075 * Primary. (tcp_secret_generating == tcp_secret_primary)
3076 * Used both for generation and primary verification.
3078 * Retiring. (tcp_secret_retiring != tcp_secret_secondary)
3079 * Used for verification, until the first failure that can be
3080 * verified by the newer Generating secret. At that time, this
3081 * cookie's state is changed to Secondary, and the Generating
3082 * cookie's state is changed to Primary. This is a short-term state,
3083 * typically lasting only one round trip time (RTT).
3085 * Secondary. (tcp_secret_retiring == tcp_secret_secondary)
3086 * Used for secondary verification, after primary verification
3087 * failures. This state lasts no more than twice the Maximum Segment
3088 * Lifetime (2MSL). Then, the secret is discarded.
3090 struct tcp_cookie_secret {
3091 /* The secret is divided into two parts. The digest part is the
3092 * equivalent of previously hashing a secret and saving the state,
3093 * and serves as an initialization vector (IV). The message part
3094 * serves as the trailing secret.
3096 u32 secrets[COOKIE_WORKSPACE_WORDS];
3097 unsigned long expires;
3100 #define TCP_SECRET_1MSL (HZ * TCP_PAWS_MSL)
3101 #define TCP_SECRET_2MSL (HZ * TCP_PAWS_MSL * 2)
3102 #define TCP_SECRET_LIFE (HZ * 600)
3104 static struct tcp_cookie_secret tcp_secret_one;
3105 static struct tcp_cookie_secret tcp_secret_two;
3107 /* Essentially a circular list, without dynamic allocation. */
3108 static struct tcp_cookie_secret *tcp_secret_generating;
3109 static struct tcp_cookie_secret *tcp_secret_primary;
3110 static struct tcp_cookie_secret *tcp_secret_retiring;
3111 static struct tcp_cookie_secret *tcp_secret_secondary;
3113 static DEFINE_SPINLOCK(tcp_secret_locker);
3115 /* Select a pseudo-random word in the cookie workspace.
3117 static inline u32 tcp_cookie_work(const u32 *ws, const int n)
3119 return ws[COOKIE_DIGEST_WORDS + ((COOKIE_MESSAGE_WORDS-1) & ws[n])];
3122 /* Fill bakery[COOKIE_WORKSPACE_WORDS] with generator, updating as needed.
3123 * Called in softirq context.
3124 * Returns: 0 for success.
3126 int tcp_cookie_generator(u32 *bakery)
3128 unsigned long jiffy = jiffies;
3130 if (unlikely(time_after_eq(jiffy, tcp_secret_generating->expires))) {
3131 spin_lock_bh(&tcp_secret_locker);
3132 if (!time_after_eq(jiffy, tcp_secret_generating->expires)) {
3133 /* refreshed by another */
3135 &tcp_secret_generating->secrets[0],
3136 COOKIE_WORKSPACE_WORDS);
3138 /* still needs refreshing */
3139 get_random_bytes(bakery, COOKIE_WORKSPACE_WORDS);
3141 /* The first time, paranoia assumes that the
3142 * randomization function isn't as strong. But,
3143 * this secret initialization is delayed until
3144 * the last possible moment (packet arrival).
3145 * Although that time is observable, it is
3146 * unpredictably variable. Mash in the most
3147 * volatile clock bits available, and expire the
3148 * secret extra quickly.
3150 if (unlikely(tcp_secret_primary->expires ==
3151 tcp_secret_secondary->expires)) {
3154 getnstimeofday(&tv);
3155 bakery[COOKIE_DIGEST_WORDS+0] ^=
3158 tcp_secret_secondary->expires = jiffy
3160 + (0x0f & tcp_cookie_work(bakery, 0));
3162 tcp_secret_secondary->expires = jiffy
3164 + (0xff & tcp_cookie_work(bakery, 1));
3165 tcp_secret_primary->expires = jiffy
3167 + (0x1f & tcp_cookie_work(bakery, 2));
3169 memcpy(&tcp_secret_secondary->secrets[0],
3170 bakery, COOKIE_WORKSPACE_WORDS);
3172 rcu_assign_pointer(tcp_secret_generating,
3173 tcp_secret_secondary);
3174 rcu_assign_pointer(tcp_secret_retiring,
3175 tcp_secret_primary);
3177 * Neither call_rcu() nor synchronize_rcu() needed.
3178 * Retiring data is not freed. It is replaced after
3179 * further (locked) pointer updates, and a quiet time
3180 * (minimum 1MSL, maximum LIFE - 2MSL).
3183 spin_unlock_bh(&tcp_secret_locker);
3187 &rcu_dereference(tcp_secret_generating)->secrets[0],
3188 COOKIE_WORKSPACE_WORDS);
3189 rcu_read_unlock_bh();
3193 EXPORT_SYMBOL(tcp_cookie_generator);
3195 void tcp_done(struct sock *sk)
3197 if (sk->sk_state == TCP_SYN_SENT || sk->sk_state == TCP_SYN_RECV)
3198 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_ATTEMPTFAILS);
3200 tcp_set_state(sk, TCP_CLOSE);
3201 tcp_clear_xmit_timers(sk);
3203 sk->sk_shutdown = SHUTDOWN_MASK;
3205 if (!sock_flag(sk, SOCK_DEAD))
3206 sk->sk_state_change(sk);
3208 inet_csk_destroy_sock(sk);
3210 EXPORT_SYMBOL_GPL(tcp_done);
3212 extern struct tcp_congestion_ops tcp_reno;
3214 static __initdata unsigned long thash_entries;
3215 static int __init set_thash_entries(char *str)
3219 thash_entries = simple_strtoul(str, &str, 0);
3222 __setup("thash_entries=", set_thash_entries);
3224 void __init tcp_init(void)
3226 struct sk_buff *skb = NULL;
3227 unsigned long nr_pages, limit;
3228 int i, max_share, cnt;
3229 unsigned long jiffy = jiffies;
3231 BUILD_BUG_ON(sizeof(struct tcp_skb_cb) > sizeof(skb->cb));
3233 percpu_counter_init(&tcp_sockets_allocated, 0);
3234 percpu_counter_init(&tcp_orphan_count, 0);
3235 tcp_hashinfo.bind_bucket_cachep =
3236 kmem_cache_create("tcp_bind_bucket",
3237 sizeof(struct inet_bind_bucket), 0,
3238 SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL);
3240 /* Size and allocate the main established and bind bucket
3243 * The methodology is similar to that of the buffer cache.
3245 tcp_hashinfo.ehash =
3246 alloc_large_system_hash("TCP established",
3247 sizeof(struct inet_ehash_bucket),
3249 (totalram_pages >= 128 * 1024) ?
3253 &tcp_hashinfo.ehash_mask,
3254 thash_entries ? 0 : 512 * 1024);
3255 for (i = 0; i <= tcp_hashinfo.ehash_mask; i++) {
3256 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].chain, i);
3257 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].twchain, i);
3259 if (inet_ehash_locks_alloc(&tcp_hashinfo))
3260 panic("TCP: failed to alloc ehash_locks");
3261 tcp_hashinfo.bhash =
3262 alloc_large_system_hash("TCP bind",
3263 sizeof(struct inet_bind_hashbucket),
3264 tcp_hashinfo.ehash_mask + 1,
3265 (totalram_pages >= 128 * 1024) ?
3268 &tcp_hashinfo.bhash_size,
3271 tcp_hashinfo.bhash_size = 1 << tcp_hashinfo.bhash_size;
3272 for (i = 0; i < tcp_hashinfo.bhash_size; i++) {
3273 spin_lock_init(&tcp_hashinfo.bhash[i].lock);
3274 INIT_HLIST_HEAD(&tcp_hashinfo.bhash[i].chain);
3278 cnt = tcp_hashinfo.ehash_mask + 1;
3280 tcp_death_row.sysctl_max_tw_buckets = cnt / 2;
3281 sysctl_tcp_max_orphans = cnt / 2;
3282 sysctl_max_syn_backlog = max(128, cnt / 256);
3284 /* Set the pressure threshold to be a fraction of global memory that
3285 * is up to 1/2 at 256 MB, decreasing toward zero with the amount of
3286 * memory, with a floor of 128 pages.
3288 nr_pages = totalram_pages - totalhigh_pages;
3289 limit = min(nr_pages, 1UL<<(28-PAGE_SHIFT)) >> (20-PAGE_SHIFT);
3290 limit = (limit * (nr_pages >> (20-PAGE_SHIFT))) >> (PAGE_SHIFT-11);
3291 limit = max(limit, 128UL);
3292 sysctl_tcp_mem[0] = limit / 4 * 3;
3293 sysctl_tcp_mem[1] = limit;
3294 sysctl_tcp_mem[2] = sysctl_tcp_mem[0] * 2;
3296 /* Set per-socket limits to no more than 1/128 the pressure threshold */
3297 limit = ((unsigned long)sysctl_tcp_mem[1]) << (PAGE_SHIFT - 7);
3298 max_share = min(4UL*1024*1024, limit);
3300 sysctl_tcp_wmem[0] = SK_MEM_QUANTUM;
3301 sysctl_tcp_wmem[1] = 16*1024;
3302 sysctl_tcp_wmem[2] = max(64*1024, max_share);
3304 sysctl_tcp_rmem[0] = SK_MEM_QUANTUM;
3305 sysctl_tcp_rmem[1] = 87380;
3306 sysctl_tcp_rmem[2] = max(87380, max_share);
3308 printk(KERN_INFO "TCP: Hash tables configured "
3309 "(established %u bind %u)\n",
3310 tcp_hashinfo.ehash_mask + 1, tcp_hashinfo.bhash_size);
3312 tcp_register_congestion_control(&tcp_reno);
3314 memset(&tcp_secret_one.secrets[0], 0, sizeof(tcp_secret_one.secrets));
3315 memset(&tcp_secret_two.secrets[0], 0, sizeof(tcp_secret_two.secrets));
3316 tcp_secret_one.expires = jiffy; /* past due */
3317 tcp_secret_two.expires = jiffy; /* past due */
3318 tcp_secret_generating = &tcp_secret_one;
3319 tcp_secret_primary = &tcp_secret_one;
3320 tcp_secret_retiring = &tcp_secret_two;
3321 tcp_secret_secondary = &tcp_secret_two;