Loading...
1/*
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.
5 *
6 * Implementation of the Transmission Control Protocol(TCP).
7 *
8 * Authors: Ross Biro
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>
19 *
20 * Fixes:
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
25 * (tcp_err()).
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
36 * unknown sockets.
37 * Alan Cox : tcp option processing.
38 * Alan Cox : Reset tweaked (still not 100%) [Had
39 * syn rule wrong]
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
45 * escape still
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
49 * facilities
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
54 * bit to skb ops.
55 * Alan Cox : Tidied tcp_data to avoid a potential
56 * nasty.
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
68 * sockets.
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
72 * state ack error.
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
77 * fixes
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
83 * completely
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
91 * (not yet usable)
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
104 * all cases.
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
109 * works now.
110 * Michael Pall : recv(...,MSG_OOB) never blocks in the
111 * BSD api.
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
119 * fixed ports.
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
125 * socket close.
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
130 * accept.
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
141 * close.
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
147 * comments.
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
155 * resemble the RFC.
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
160 * generates them.
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
173 * but it's a start!
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
194 * improvement.
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.
207 *
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.
212 *
213 * Description of States:
214 *
215 * TCP_SYN_SENT sent a connection request, waiting for ack
216 *
217 * TCP_SYN_RECV received a connection request, sent ack,
218 * waiting for final ack in three-way handshake.
219 *
220 * TCP_ESTABLISHED connection established
221 *
222 * TCP_FIN_WAIT1 our side has shutdown, waiting to complete
223 * transmission of remaining buffered data
224 *
225 * TCP_FIN_WAIT2 all buffered data sent, waiting for remote
226 * to shutdown
227 *
228 * TCP_CLOSING both sides have shutdown but we still have
229 * data we have to finish sending
230 *
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)
236 *
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)
240 *
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
244 *
245 * TCP_CLOSE socket is finished
246 */
247
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
270#include <net/icmp.h>
271#include <net/tcp.h>
272#include <net/xfrm.h>
273#include <net/ip.h>
274#include <net/netdma.h>
275#include <net/sock.h>
276
277#include <asm/uaccess.h>
278#include <asm/ioctls.h>
279
280int sysctl_tcp_fin_timeout __read_mostly = TCP_FIN_TIMEOUT;
281
282struct percpu_counter tcp_orphan_count;
283EXPORT_SYMBOL_GPL(tcp_orphan_count);
284
285long sysctl_tcp_mem[3] __read_mostly;
286int sysctl_tcp_wmem[3] __read_mostly;
287int sysctl_tcp_rmem[3] __read_mostly;
288
289EXPORT_SYMBOL(sysctl_tcp_mem);
290EXPORT_SYMBOL(sysctl_tcp_rmem);
291EXPORT_SYMBOL(sysctl_tcp_wmem);
292
293atomic_long_t tcp_memory_allocated; /* Current allocated memory. */
294EXPORT_SYMBOL(tcp_memory_allocated);
295
296/*
297 * Current number of TCP sockets.
298 */
299struct percpu_counter tcp_sockets_allocated;
300EXPORT_SYMBOL(tcp_sockets_allocated);
301
302/*
303 * TCP splice context
304 */
305struct tcp_splice_state {
306 struct pipe_inode_info *pipe;
307 size_t len;
308 unsigned int flags;
309};
310
311/*
312 * Pressure flag: try to collapse.
313 * Technical note: it is used by multiple contexts non atomically.
314 * All the __sk_mem_schedule() is of this nature: accounting
315 * is strict, actions are advisory and have some latency.
316 */
317int tcp_memory_pressure __read_mostly;
318EXPORT_SYMBOL(tcp_memory_pressure);
319
320void tcp_enter_memory_pressure(struct sock *sk)
321{
322 if (!tcp_memory_pressure) {
323 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMEMORYPRESSURES);
324 tcp_memory_pressure = 1;
325 }
326}
327EXPORT_SYMBOL(tcp_enter_memory_pressure);
328
329/* Convert seconds to retransmits based on initial and max timeout */
330static u8 secs_to_retrans(int seconds, int timeout, int rto_max)
331{
332 u8 res = 0;
333
334 if (seconds > 0) {
335 int period = timeout;
336
337 res = 1;
338 while (seconds > period && res < 255) {
339 res++;
340 timeout <<= 1;
341 if (timeout > rto_max)
342 timeout = rto_max;
343 period += timeout;
344 }
345 }
346 return res;
347}
348
349/* Convert retransmits to seconds based on initial and max timeout */
350static int retrans_to_secs(u8 retrans, int timeout, int rto_max)
351{
352 int period = 0;
353
354 if (retrans > 0) {
355 period = timeout;
356 while (--retrans) {
357 timeout <<= 1;
358 if (timeout > rto_max)
359 timeout = rto_max;
360 period += timeout;
361 }
362 }
363 return period;
364}
365
366/*
367 * Wait for a TCP event.
368 *
369 * Note that we don't need to lock the socket, as the upper poll layers
370 * take care of normal races (between the test and the event) and we don't
371 * go look at any of the socket buffers directly.
372 */
373unsigned int tcp_poll(struct file *file, struct socket *sock, poll_table *wait)
374{
375 unsigned int mask;
376 struct sock *sk = sock->sk;
377 struct tcp_sock *tp = tcp_sk(sk);
378
379 sock_poll_wait(file, sk_sleep(sk), wait);
380 if (sk->sk_state == TCP_LISTEN)
381 return inet_csk_listen_poll(sk);
382
383 /* Socket is not locked. We are protected from async events
384 * by poll logic and correct handling of state changes
385 * made by other threads is impossible in any case.
386 */
387
388 mask = 0;
389
390 /*
391 * POLLHUP is certainly not done right. But poll() doesn't
392 * have a notion of HUP in just one direction, and for a
393 * socket the read side is more interesting.
394 *
395 * Some poll() documentation says that POLLHUP is incompatible
396 * with the POLLOUT/POLLWR flags, so somebody should check this
397 * all. But careful, it tends to be safer to return too many
398 * bits than too few, and you can easily break real applications
399 * if you don't tell them that something has hung up!
400 *
401 * Check-me.
402 *
403 * Check number 1. POLLHUP is _UNMASKABLE_ event (see UNIX98 and
404 * our fs/select.c). It means that after we received EOF,
405 * poll always returns immediately, making impossible poll() on write()
406 * in state CLOSE_WAIT. One solution is evident --- to set POLLHUP
407 * if and only if shutdown has been made in both directions.
408 * Actually, it is interesting to look how Solaris and DUX
409 * solve this dilemma. I would prefer, if POLLHUP were maskable,
410 * then we could set it on SND_SHUTDOWN. BTW examples given
411 * in Stevens' books assume exactly this behaviour, it explains
412 * why POLLHUP is incompatible with POLLOUT. --ANK
413 *
414 * NOTE. Check for TCP_CLOSE is added. The goal is to prevent
415 * blocking on fresh not-connected or disconnected socket. --ANK
416 */
417 if (sk->sk_shutdown == SHUTDOWN_MASK || sk->sk_state == TCP_CLOSE)
418 mask |= POLLHUP;
419 if (sk->sk_shutdown & RCV_SHUTDOWN)
420 mask |= POLLIN | POLLRDNORM | POLLRDHUP;
421
422 /* Connected? */
423 if ((1 << sk->sk_state) & ~(TCPF_SYN_SENT | TCPF_SYN_RECV)) {
424 int target = sock_rcvlowat(sk, 0, INT_MAX);
425
426 if (tp->urg_seq == tp->copied_seq &&
427 !sock_flag(sk, SOCK_URGINLINE) &&
428 tp->urg_data)
429 target++;
430
431 /* Potential race condition. If read of tp below will
432 * escape above sk->sk_state, we can be illegally awaken
433 * in SYN_* states. */
434 if (tp->rcv_nxt - tp->copied_seq >= target)
435 mask |= POLLIN | POLLRDNORM;
436
437 if (!(sk->sk_shutdown & SEND_SHUTDOWN)) {
438 if (sk_stream_wspace(sk) >= sk_stream_min_wspace(sk)) {
439 mask |= POLLOUT | POLLWRNORM;
440 } else { /* send SIGIO later */
441 set_bit(SOCK_ASYNC_NOSPACE,
442 &sk->sk_socket->flags);
443 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
444
445 /* Race breaker. If space is freed after
446 * wspace test but before the flags are set,
447 * IO signal will be lost.
448 */
449 if (sk_stream_wspace(sk) >= sk_stream_min_wspace(sk))
450 mask |= POLLOUT | POLLWRNORM;
451 }
452 } else
453 mask |= POLLOUT | POLLWRNORM;
454
455 if (tp->urg_data & TCP_URG_VALID)
456 mask |= POLLPRI;
457 }
458 /* This barrier is coupled with smp_wmb() in tcp_reset() */
459 smp_rmb();
460 if (sk->sk_err)
461 mask |= POLLERR;
462
463 return mask;
464}
465EXPORT_SYMBOL(tcp_poll);
466
467int tcp_ioctl(struct sock *sk, int cmd, unsigned long arg)
468{
469 struct tcp_sock *tp = tcp_sk(sk);
470 int answ;
471
472 switch (cmd) {
473 case SIOCINQ:
474 if (sk->sk_state == TCP_LISTEN)
475 return -EINVAL;
476
477 lock_sock(sk);
478 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
479 answ = 0;
480 else if (sock_flag(sk, SOCK_URGINLINE) ||
481 !tp->urg_data ||
482 before(tp->urg_seq, tp->copied_seq) ||
483 !before(tp->urg_seq, tp->rcv_nxt)) {
484 struct sk_buff *skb;
485
486 answ = tp->rcv_nxt - tp->copied_seq;
487
488 /* Subtract 1, if FIN is in queue. */
489 skb = skb_peek_tail(&sk->sk_receive_queue);
490 if (answ && skb)
491 answ -= tcp_hdr(skb)->fin;
492 } else
493 answ = tp->urg_seq - tp->copied_seq;
494 release_sock(sk);
495 break;
496 case SIOCATMARK:
497 answ = tp->urg_data && tp->urg_seq == tp->copied_seq;
498 break;
499 case SIOCOUTQ:
500 if (sk->sk_state == TCP_LISTEN)
501 return -EINVAL;
502
503 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
504 answ = 0;
505 else
506 answ = tp->write_seq - tp->snd_una;
507 break;
508 case SIOCOUTQNSD:
509 if (sk->sk_state == TCP_LISTEN)
510 return -EINVAL;
511
512 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
513 answ = 0;
514 else
515 answ = tp->write_seq - tp->snd_nxt;
516 break;
517 default:
518 return -ENOIOCTLCMD;
519 }
520
521 return put_user(answ, (int __user *)arg);
522}
523EXPORT_SYMBOL(tcp_ioctl);
524
525static inline void tcp_mark_push(struct tcp_sock *tp, struct sk_buff *skb)
526{
527 TCP_SKB_CB(skb)->flags |= TCPHDR_PSH;
528 tp->pushed_seq = tp->write_seq;
529}
530
531static inline int forced_push(struct tcp_sock *tp)
532{
533 return after(tp->write_seq, tp->pushed_seq + (tp->max_window >> 1));
534}
535
536static inline void skb_entail(struct sock *sk, struct sk_buff *skb)
537{
538 struct tcp_sock *tp = tcp_sk(sk);
539 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
540
541 skb->csum = 0;
542 tcb->seq = tcb->end_seq = tp->write_seq;
543 tcb->flags = TCPHDR_ACK;
544 tcb->sacked = 0;
545 skb_header_release(skb);
546 tcp_add_write_queue_tail(sk, skb);
547 sk->sk_wmem_queued += skb->truesize;
548 sk_mem_charge(sk, skb->truesize);
549 if (tp->nonagle & TCP_NAGLE_PUSH)
550 tp->nonagle &= ~TCP_NAGLE_PUSH;
551}
552
553static inline void tcp_mark_urg(struct tcp_sock *tp, int flags)
554{
555 if (flags & MSG_OOB)
556 tp->snd_up = tp->write_seq;
557}
558
559static inline void tcp_push(struct sock *sk, int flags, int mss_now,
560 int nonagle)
561{
562 if (tcp_send_head(sk)) {
563 struct tcp_sock *tp = tcp_sk(sk);
564
565 if (!(flags & MSG_MORE) || forced_push(tp))
566 tcp_mark_push(tp, tcp_write_queue_tail(sk));
567
568 tcp_mark_urg(tp, flags);
569 __tcp_push_pending_frames(sk, mss_now,
570 (flags & MSG_MORE) ? TCP_NAGLE_CORK : nonagle);
571 }
572}
573
574static int tcp_splice_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb,
575 unsigned int offset, size_t len)
576{
577 struct tcp_splice_state *tss = rd_desc->arg.data;
578 int ret;
579
580 ret = skb_splice_bits(skb, offset, tss->pipe, min(rd_desc->count, len),
581 tss->flags);
582 if (ret > 0)
583 rd_desc->count -= ret;
584 return ret;
585}
586
587static int __tcp_splice_read(struct sock *sk, struct tcp_splice_state *tss)
588{
589 /* Store TCP splice context information in read_descriptor_t. */
590 read_descriptor_t rd_desc = {
591 .arg.data = tss,
592 .count = tss->len,
593 };
594
595 return tcp_read_sock(sk, &rd_desc, tcp_splice_data_recv);
596}
597
598/**
599 * tcp_splice_read - splice data from TCP socket to a pipe
600 * @sock: socket to splice from
601 * @ppos: position (not valid)
602 * @pipe: pipe to splice to
603 * @len: number of bytes to splice
604 * @flags: splice modifier flags
605 *
606 * Description:
607 * Will read pages from given socket and fill them into a pipe.
608 *
609 **/
610ssize_t tcp_splice_read(struct socket *sock, loff_t *ppos,
611 struct pipe_inode_info *pipe, size_t len,
612 unsigned int flags)
613{
614 struct sock *sk = sock->sk;
615 struct tcp_splice_state tss = {
616 .pipe = pipe,
617 .len = len,
618 .flags = flags,
619 };
620 long timeo;
621 ssize_t spliced;
622 int ret;
623
624 sock_rps_record_flow(sk);
625 /*
626 * We can't seek on a socket input
627 */
628 if (unlikely(*ppos))
629 return -ESPIPE;
630
631 ret = spliced = 0;
632
633 lock_sock(sk);
634
635 timeo = sock_rcvtimeo(sk, sock->file->f_flags & O_NONBLOCK);
636 while (tss.len) {
637 ret = __tcp_splice_read(sk, &tss);
638 if (ret < 0)
639 break;
640 else if (!ret) {
641 if (spliced)
642 break;
643 if (sock_flag(sk, SOCK_DONE))
644 break;
645 if (sk->sk_err) {
646 ret = sock_error(sk);
647 break;
648 }
649 if (sk->sk_shutdown & RCV_SHUTDOWN)
650 break;
651 if (sk->sk_state == TCP_CLOSE) {
652 /*
653 * This occurs when user tries to read
654 * from never connected socket.
655 */
656 if (!sock_flag(sk, SOCK_DONE))
657 ret = -ENOTCONN;
658 break;
659 }
660 if (!timeo) {
661 ret = -EAGAIN;
662 break;
663 }
664 sk_wait_data(sk, &timeo);
665 if (signal_pending(current)) {
666 ret = sock_intr_errno(timeo);
667 break;
668 }
669 continue;
670 }
671 tss.len -= ret;
672 spliced += ret;
673
674 if (!timeo)
675 break;
676 release_sock(sk);
677 lock_sock(sk);
678
679 if (sk->sk_err || sk->sk_state == TCP_CLOSE ||
680 (sk->sk_shutdown & RCV_SHUTDOWN) ||
681 signal_pending(current))
682 break;
683 }
684
685 release_sock(sk);
686
687 if (spliced)
688 return spliced;
689
690 return ret;
691}
692EXPORT_SYMBOL(tcp_splice_read);
693
694struct sk_buff *sk_stream_alloc_skb(struct sock *sk, int size, gfp_t gfp)
695{
696 struct sk_buff *skb;
697
698 /* The TCP header must be at least 32-bit aligned. */
699 size = ALIGN(size, 4);
700
701 skb = alloc_skb_fclone(size + sk->sk_prot->max_header, gfp);
702 if (skb) {
703 if (sk_wmem_schedule(sk, skb->truesize)) {
704 /*
705 * Make sure that we have exactly size bytes
706 * available to the caller, no more, no less.
707 */
708 skb_reserve(skb, skb_tailroom(skb) - size);
709 return skb;
710 }
711 __kfree_skb(skb);
712 } else {
713 sk->sk_prot->enter_memory_pressure(sk);
714 sk_stream_moderate_sndbuf(sk);
715 }
716 return NULL;
717}
718
719static unsigned int tcp_xmit_size_goal(struct sock *sk, u32 mss_now,
720 int large_allowed)
721{
722 struct tcp_sock *tp = tcp_sk(sk);
723 u32 xmit_size_goal, old_size_goal;
724
725 xmit_size_goal = mss_now;
726
727 if (large_allowed && sk_can_gso(sk)) {
728 xmit_size_goal = ((sk->sk_gso_max_size - 1) -
729 inet_csk(sk)->icsk_af_ops->net_header_len -
730 inet_csk(sk)->icsk_ext_hdr_len -
731 tp->tcp_header_len);
732
733 xmit_size_goal = tcp_bound_to_half_wnd(tp, xmit_size_goal);
734
735 /* We try hard to avoid divides here */
736 old_size_goal = tp->xmit_size_goal_segs * mss_now;
737
738 if (likely(old_size_goal <= xmit_size_goal &&
739 old_size_goal + mss_now > xmit_size_goal)) {
740 xmit_size_goal = old_size_goal;
741 } else {
742 tp->xmit_size_goal_segs = xmit_size_goal / mss_now;
743 xmit_size_goal = tp->xmit_size_goal_segs * mss_now;
744 }
745 }
746
747 return max(xmit_size_goal, mss_now);
748}
749
750static int tcp_send_mss(struct sock *sk, int *size_goal, int flags)
751{
752 int mss_now;
753
754 mss_now = tcp_current_mss(sk);
755 *size_goal = tcp_xmit_size_goal(sk, mss_now, !(flags & MSG_OOB));
756
757 return mss_now;
758}
759
760static ssize_t do_tcp_sendpages(struct sock *sk, struct page **pages, int poffset,
761 size_t psize, int flags)
762{
763 struct tcp_sock *tp = tcp_sk(sk);
764 int mss_now, size_goal;
765 int err;
766 ssize_t copied;
767 long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
768
769 /* Wait for a connection to finish. */
770 if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
771 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
772 goto out_err;
773
774 clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
775
776 mss_now = tcp_send_mss(sk, &size_goal, flags);
777 copied = 0;
778
779 err = -EPIPE;
780 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
781 goto out_err;
782
783 while (psize > 0) {
784 struct sk_buff *skb = tcp_write_queue_tail(sk);
785 struct page *page = pages[poffset / PAGE_SIZE];
786 int copy, i, can_coalesce;
787 int offset = poffset % PAGE_SIZE;
788 int size = min_t(size_t, psize, PAGE_SIZE - offset);
789
790 if (!tcp_send_head(sk) || (copy = size_goal - skb->len) <= 0) {
791new_segment:
792 if (!sk_stream_memory_free(sk))
793 goto wait_for_sndbuf;
794
795 skb = sk_stream_alloc_skb(sk, 0, sk->sk_allocation);
796 if (!skb)
797 goto wait_for_memory;
798
799 skb_entail(sk, skb);
800 copy = size_goal;
801 }
802
803 if (copy > size)
804 copy = size;
805
806 i = skb_shinfo(skb)->nr_frags;
807 can_coalesce = skb_can_coalesce(skb, i, page, offset);
808 if (!can_coalesce && i >= MAX_SKB_FRAGS) {
809 tcp_mark_push(tp, skb);
810 goto new_segment;
811 }
812 if (!sk_wmem_schedule(sk, copy))
813 goto wait_for_memory;
814
815 if (can_coalesce) {
816 skb_shinfo(skb)->frags[i - 1].size += copy;
817 } else {
818 get_page(page);
819 skb_fill_page_desc(skb, i, page, offset, copy);
820 }
821
822 skb->len += copy;
823 skb->data_len += copy;
824 skb->truesize += copy;
825 sk->sk_wmem_queued += copy;
826 sk_mem_charge(sk, copy);
827 skb->ip_summed = CHECKSUM_PARTIAL;
828 tp->write_seq += copy;
829 TCP_SKB_CB(skb)->end_seq += copy;
830 skb_shinfo(skb)->gso_segs = 0;
831
832 if (!copied)
833 TCP_SKB_CB(skb)->flags &= ~TCPHDR_PSH;
834
835 copied += copy;
836 poffset += copy;
837 if (!(psize -= copy))
838 goto out;
839
840 if (skb->len < size_goal || (flags & MSG_OOB))
841 continue;
842
843 if (forced_push(tp)) {
844 tcp_mark_push(tp, skb);
845 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
846 } else if (skb == tcp_send_head(sk))
847 tcp_push_one(sk, mss_now);
848 continue;
849
850wait_for_sndbuf:
851 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
852wait_for_memory:
853 if (copied)
854 tcp_push(sk, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH);
855
856 if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
857 goto do_error;
858
859 mss_now = tcp_send_mss(sk, &size_goal, flags);
860 }
861
862out:
863 if (copied)
864 tcp_push(sk, flags, mss_now, tp->nonagle);
865 return copied;
866
867do_error:
868 if (copied)
869 goto out;
870out_err:
871 return sk_stream_error(sk, flags, err);
872}
873
874int tcp_sendpage(struct sock *sk, struct page *page, int offset,
875 size_t size, int flags)
876{
877 ssize_t res;
878
879 if (!(sk->sk_route_caps & NETIF_F_SG) ||
880 !(sk->sk_route_caps & NETIF_F_ALL_CSUM))
881 return sock_no_sendpage(sk->sk_socket, page, offset, size,
882 flags);
883
884 lock_sock(sk);
885 res = do_tcp_sendpages(sk, &page, offset, size, flags);
886 release_sock(sk);
887 return res;
888}
889EXPORT_SYMBOL(tcp_sendpage);
890
891#define TCP_PAGE(sk) (sk->sk_sndmsg_page)
892#define TCP_OFF(sk) (sk->sk_sndmsg_off)
893
894static inline int select_size(struct sock *sk, int sg)
895{
896 struct tcp_sock *tp = tcp_sk(sk);
897 int tmp = tp->mss_cache;
898
899 if (sg) {
900 if (sk_can_gso(sk))
901 tmp = 0;
902 else {
903 int pgbreak = SKB_MAX_HEAD(MAX_TCP_HEADER);
904
905 if (tmp >= pgbreak &&
906 tmp <= pgbreak + (MAX_SKB_FRAGS - 1) * PAGE_SIZE)
907 tmp = pgbreak;
908 }
909 }
910
911 return tmp;
912}
913
914int tcp_sendmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg,
915 size_t size)
916{
917 struct iovec *iov;
918 struct tcp_sock *tp = tcp_sk(sk);
919 struct sk_buff *skb;
920 int iovlen, flags;
921 int mss_now, size_goal;
922 int sg, err, copied;
923 long timeo;
924
925 lock_sock(sk);
926
927 flags = msg->msg_flags;
928 timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
929
930 /* Wait for a connection to finish. */
931 if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
932 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
933 goto out_err;
934
935 /* This should be in poll */
936 clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
937
938 mss_now = tcp_send_mss(sk, &size_goal, flags);
939
940 /* Ok commence sending. */
941 iovlen = msg->msg_iovlen;
942 iov = msg->msg_iov;
943 copied = 0;
944
945 err = -EPIPE;
946 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
947 goto out_err;
948
949 sg = sk->sk_route_caps & NETIF_F_SG;
950
951 while (--iovlen >= 0) {
952 size_t seglen = iov->iov_len;
953 unsigned char __user *from = iov->iov_base;
954
955 iov++;
956
957 while (seglen > 0) {
958 int copy = 0;
959 int max = size_goal;
960
961 skb = tcp_write_queue_tail(sk);
962 if (tcp_send_head(sk)) {
963 if (skb->ip_summed == CHECKSUM_NONE)
964 max = mss_now;
965 copy = max - skb->len;
966 }
967
968 if (copy <= 0) {
969new_segment:
970 /* Allocate new segment. If the interface is SG,
971 * allocate skb fitting to single page.
972 */
973 if (!sk_stream_memory_free(sk))
974 goto wait_for_sndbuf;
975
976 skb = sk_stream_alloc_skb(sk,
977 select_size(sk, sg),
978 sk->sk_allocation);
979 if (!skb)
980 goto wait_for_memory;
981
982 /*
983 * Check whether we can use HW checksum.
984 */
985 if (sk->sk_route_caps & NETIF_F_ALL_CSUM)
986 skb->ip_summed = CHECKSUM_PARTIAL;
987
988 skb_entail(sk, skb);
989 copy = size_goal;
990 max = size_goal;
991 }
992
993 /* Try to append data to the end of skb. */
994 if (copy > seglen)
995 copy = seglen;
996
997 /* Where to copy to? */
998 if (skb_tailroom(skb) > 0) {
999 /* We have some space in skb head. Superb! */
1000 if (copy > skb_tailroom(skb))
1001 copy = skb_tailroom(skb);
1002 err = skb_add_data_nocache(sk, skb, from, copy);
1003 if (err)
1004 goto do_fault;
1005 } else {
1006 int merge = 0;
1007 int i = skb_shinfo(skb)->nr_frags;
1008 struct page *page = TCP_PAGE(sk);
1009 int off = TCP_OFF(sk);
1010
1011 if (skb_can_coalesce(skb, i, page, off) &&
1012 off != PAGE_SIZE) {
1013 /* We can extend the last page
1014 * fragment. */
1015 merge = 1;
1016 } else if (i == MAX_SKB_FRAGS || !sg) {
1017 /* Need to add new fragment and cannot
1018 * do this because interface is non-SG,
1019 * or because all the page slots are
1020 * busy. */
1021 tcp_mark_push(tp, skb);
1022 goto new_segment;
1023 } else if (page) {
1024 if (off == PAGE_SIZE) {
1025 put_page(page);
1026 TCP_PAGE(sk) = page = NULL;
1027 off = 0;
1028 }
1029 } else
1030 off = 0;
1031
1032 if (copy > PAGE_SIZE - off)
1033 copy = PAGE_SIZE - off;
1034
1035 if (!sk_wmem_schedule(sk, copy))
1036 goto wait_for_memory;
1037
1038 if (!page) {
1039 /* Allocate new cache page. */
1040 if (!(page = sk_stream_alloc_page(sk)))
1041 goto wait_for_memory;
1042 }
1043
1044 /* Time to copy data. We are close to
1045 * the end! */
1046 err = skb_copy_to_page_nocache(sk, from, skb,
1047 page, off, copy);
1048 if (err) {
1049 /* If this page was new, give it to the
1050 * socket so it does not get leaked.
1051 */
1052 if (!TCP_PAGE(sk)) {
1053 TCP_PAGE(sk) = page;
1054 TCP_OFF(sk) = 0;
1055 }
1056 goto do_error;
1057 }
1058
1059 /* Update the skb. */
1060 if (merge) {
1061 skb_shinfo(skb)->frags[i - 1].size +=
1062 copy;
1063 } else {
1064 skb_fill_page_desc(skb, i, page, off, copy);
1065 if (TCP_PAGE(sk)) {
1066 get_page(page);
1067 } else if (off + copy < PAGE_SIZE) {
1068 get_page(page);
1069 TCP_PAGE(sk) = page;
1070 }
1071 }
1072
1073 TCP_OFF(sk) = off + copy;
1074 }
1075
1076 if (!copied)
1077 TCP_SKB_CB(skb)->flags &= ~TCPHDR_PSH;
1078
1079 tp->write_seq += copy;
1080 TCP_SKB_CB(skb)->end_seq += copy;
1081 skb_shinfo(skb)->gso_segs = 0;
1082
1083 from += copy;
1084 copied += copy;
1085 if ((seglen -= copy) == 0 && iovlen == 0)
1086 goto out;
1087
1088 if (skb->len < max || (flags & MSG_OOB))
1089 continue;
1090
1091 if (forced_push(tp)) {
1092 tcp_mark_push(tp, skb);
1093 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
1094 } else if (skb == tcp_send_head(sk))
1095 tcp_push_one(sk, mss_now);
1096 continue;
1097
1098wait_for_sndbuf:
1099 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1100wait_for_memory:
1101 if (copied)
1102 tcp_push(sk, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH);
1103
1104 if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
1105 goto do_error;
1106
1107 mss_now = tcp_send_mss(sk, &size_goal, flags);
1108 }
1109 }
1110
1111out:
1112 if (copied)
1113 tcp_push(sk, flags, mss_now, tp->nonagle);
1114 release_sock(sk);
1115 return copied;
1116
1117do_fault:
1118 if (!skb->len) {
1119 tcp_unlink_write_queue(skb, sk);
1120 /* It is the one place in all of TCP, except connection
1121 * reset, where we can be unlinking the send_head.
1122 */
1123 tcp_check_send_head(sk, skb);
1124 sk_wmem_free_skb(sk, skb);
1125 }
1126
1127do_error:
1128 if (copied)
1129 goto out;
1130out_err:
1131 err = sk_stream_error(sk, flags, err);
1132 release_sock(sk);
1133 return err;
1134}
1135EXPORT_SYMBOL(tcp_sendmsg);
1136
1137/*
1138 * Handle reading urgent data. BSD has very simple semantics for
1139 * this, no blocking and very strange errors 8)
1140 */
1141
1142static int tcp_recv_urg(struct sock *sk, struct msghdr *msg, int len, int flags)
1143{
1144 struct tcp_sock *tp = tcp_sk(sk);
1145
1146 /* No URG data to read. */
1147 if (sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data ||
1148 tp->urg_data == TCP_URG_READ)
1149 return -EINVAL; /* Yes this is right ! */
1150
1151 if (sk->sk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DONE))
1152 return -ENOTCONN;
1153
1154 if (tp->urg_data & TCP_URG_VALID) {
1155 int err = 0;
1156 char c = tp->urg_data;
1157
1158 if (!(flags & MSG_PEEK))
1159 tp->urg_data = TCP_URG_READ;
1160
1161 /* Read urgent data. */
1162 msg->msg_flags |= MSG_OOB;
1163
1164 if (len > 0) {
1165 if (!(flags & MSG_TRUNC))
1166 err = memcpy_toiovec(msg->msg_iov, &c, 1);
1167 len = 1;
1168 } else
1169 msg->msg_flags |= MSG_TRUNC;
1170
1171 return err ? -EFAULT : len;
1172 }
1173
1174 if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
1175 return 0;
1176
1177 /* Fixed the recv(..., MSG_OOB) behaviour. BSD docs and
1178 * the available implementations agree in this case:
1179 * this call should never block, independent of the
1180 * blocking state of the socket.
1181 * Mike <pall@rz.uni-karlsruhe.de>
1182 */
1183 return -EAGAIN;
1184}
1185
1186/* Clean up the receive buffer for full frames taken by the user,
1187 * then send an ACK if necessary. COPIED is the number of bytes
1188 * tcp_recvmsg has given to the user so far, it speeds up the
1189 * calculation of whether or not we must ACK for the sake of
1190 * a window update.
1191 */
1192void tcp_cleanup_rbuf(struct sock *sk, int copied)
1193{
1194 struct tcp_sock *tp = tcp_sk(sk);
1195 int time_to_ack = 0;
1196
1197#if TCP_DEBUG
1198 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
1199
1200 WARN(skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq),
1201 "cleanup rbuf bug: copied %X seq %X rcvnxt %X\n",
1202 tp->copied_seq, TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt);
1203#endif
1204
1205 if (inet_csk_ack_scheduled(sk)) {
1206 const struct inet_connection_sock *icsk = inet_csk(sk);
1207 /* Delayed ACKs frequently hit locked sockets during bulk
1208 * receive. */
1209 if (icsk->icsk_ack.blocked ||
1210 /* Once-per-two-segments ACK was not sent by tcp_input.c */
1211 tp->rcv_nxt - tp->rcv_wup > icsk->icsk_ack.rcv_mss ||
1212 /*
1213 * If this read emptied read buffer, we send ACK, if
1214 * connection is not bidirectional, user drained
1215 * receive buffer and there was a small segment
1216 * in queue.
1217 */
1218 (copied > 0 &&
1219 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED2) ||
1220 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED) &&
1221 !icsk->icsk_ack.pingpong)) &&
1222 !atomic_read(&sk->sk_rmem_alloc)))
1223 time_to_ack = 1;
1224 }
1225
1226 /* We send an ACK if we can now advertise a non-zero window
1227 * which has been raised "significantly".
1228 *
1229 * Even if window raised up to infinity, do not send window open ACK
1230 * in states, where we will not receive more. It is useless.
1231 */
1232 if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) {
1233 __u32 rcv_window_now = tcp_receive_window(tp);
1234
1235 /* Optimize, __tcp_select_window() is not cheap. */
1236 if (2*rcv_window_now <= tp->window_clamp) {
1237 __u32 new_window = __tcp_select_window(sk);
1238
1239 /* Send ACK now, if this read freed lots of space
1240 * in our buffer. Certainly, new_window is new window.
1241 * We can advertise it now, if it is not less than current one.
1242 * "Lots" means "at least twice" here.
1243 */
1244 if (new_window && new_window >= 2 * rcv_window_now)
1245 time_to_ack = 1;
1246 }
1247 }
1248 if (time_to_ack)
1249 tcp_send_ack(sk);
1250}
1251
1252static void tcp_prequeue_process(struct sock *sk)
1253{
1254 struct sk_buff *skb;
1255 struct tcp_sock *tp = tcp_sk(sk);
1256
1257 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPPREQUEUED);
1258
1259 /* RX process wants to run with disabled BHs, though it is not
1260 * necessary */
1261 local_bh_disable();
1262 while ((skb = __skb_dequeue(&tp->ucopy.prequeue)) != NULL)
1263 sk_backlog_rcv(sk, skb);
1264 local_bh_enable();
1265
1266 /* Clear memory counter. */
1267 tp->ucopy.memory = 0;
1268}
1269
1270#ifdef CONFIG_NET_DMA
1271static void tcp_service_net_dma(struct sock *sk, bool wait)
1272{
1273 dma_cookie_t done, used;
1274 dma_cookie_t last_issued;
1275 struct tcp_sock *tp = tcp_sk(sk);
1276
1277 if (!tp->ucopy.dma_chan)
1278 return;
1279
1280 last_issued = tp->ucopy.dma_cookie;
1281 dma_async_memcpy_issue_pending(tp->ucopy.dma_chan);
1282
1283 do {
1284 if (dma_async_memcpy_complete(tp->ucopy.dma_chan,
1285 last_issued, &done,
1286 &used) == DMA_SUCCESS) {
1287 /* Safe to free early-copied skbs now */
1288 __skb_queue_purge(&sk->sk_async_wait_queue);
1289 break;
1290 } else {
1291 struct sk_buff *skb;
1292 while ((skb = skb_peek(&sk->sk_async_wait_queue)) &&
1293 (dma_async_is_complete(skb->dma_cookie, done,
1294 used) == DMA_SUCCESS)) {
1295 __skb_dequeue(&sk->sk_async_wait_queue);
1296 kfree_skb(skb);
1297 }
1298 }
1299 } while (wait);
1300}
1301#endif
1302
1303static inline struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off)
1304{
1305 struct sk_buff *skb;
1306 u32 offset;
1307
1308 skb_queue_walk(&sk->sk_receive_queue, skb) {
1309 offset = seq - TCP_SKB_CB(skb)->seq;
1310 if (tcp_hdr(skb)->syn)
1311 offset--;
1312 if (offset < skb->len || tcp_hdr(skb)->fin) {
1313 *off = offset;
1314 return skb;
1315 }
1316 }
1317 return NULL;
1318}
1319
1320/*
1321 * This routine provides an alternative to tcp_recvmsg() for routines
1322 * that would like to handle copying from skbuffs directly in 'sendfile'
1323 * fashion.
1324 * Note:
1325 * - It is assumed that the socket was locked by the caller.
1326 * - The routine does not block.
1327 * - At present, there is no support for reading OOB data
1328 * or for 'peeking' the socket using this routine
1329 * (although both would be easy to implement).
1330 */
1331int tcp_read_sock(struct sock *sk, read_descriptor_t *desc,
1332 sk_read_actor_t recv_actor)
1333{
1334 struct sk_buff *skb;
1335 struct tcp_sock *tp = tcp_sk(sk);
1336 u32 seq = tp->copied_seq;
1337 u32 offset;
1338 int copied = 0;
1339
1340 if (sk->sk_state == TCP_LISTEN)
1341 return -ENOTCONN;
1342 while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1343 if (offset < skb->len) {
1344 int used;
1345 size_t len;
1346
1347 len = skb->len - offset;
1348 /* Stop reading if we hit a patch of urgent data */
1349 if (tp->urg_data) {
1350 u32 urg_offset = tp->urg_seq - seq;
1351 if (urg_offset < len)
1352 len = urg_offset;
1353 if (!len)
1354 break;
1355 }
1356 used = recv_actor(desc, skb, offset, len);
1357 if (used < 0) {
1358 if (!copied)
1359 copied = used;
1360 break;
1361 } else if (used <= len) {
1362 seq += used;
1363 copied += used;
1364 offset += used;
1365 }
1366 /*
1367 * If recv_actor drops the lock (e.g. TCP splice
1368 * receive) the skb pointer might be invalid when
1369 * getting here: tcp_collapse might have deleted it
1370 * while aggregating skbs from the socket queue.
1371 */
1372 skb = tcp_recv_skb(sk, seq-1, &offset);
1373 if (!skb || (offset+1 != skb->len))
1374 break;
1375 }
1376 if (tcp_hdr(skb)->fin) {
1377 sk_eat_skb(sk, skb, 0);
1378 ++seq;
1379 break;
1380 }
1381 sk_eat_skb(sk, skb, 0);
1382 if (!desc->count)
1383 break;
1384 tp->copied_seq = seq;
1385 }
1386 tp->copied_seq = seq;
1387
1388 tcp_rcv_space_adjust(sk);
1389
1390 /* Clean up data we have read: This will do ACK frames. */
1391 if (copied > 0)
1392 tcp_cleanup_rbuf(sk, copied);
1393 return copied;
1394}
1395EXPORT_SYMBOL(tcp_read_sock);
1396
1397/*
1398 * This routine copies from a sock struct into the user buffer.
1399 *
1400 * Technical note: in 2.3 we work on _locked_ socket, so that
1401 * tricks with *seq access order and skb->users are not required.
1402 * Probably, code can be easily improved even more.
1403 */
1404
1405int tcp_recvmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg,
1406 size_t len, int nonblock, int flags, int *addr_len)
1407{
1408 struct tcp_sock *tp = tcp_sk(sk);
1409 int copied = 0;
1410 u32 peek_seq;
1411 u32 *seq;
1412 unsigned long used;
1413 int err;
1414 int target; /* Read at least this many bytes */
1415 long timeo;
1416 struct task_struct *user_recv = NULL;
1417 int copied_early = 0;
1418 struct sk_buff *skb;
1419 u32 urg_hole = 0;
1420
1421 lock_sock(sk);
1422
1423 err = -ENOTCONN;
1424 if (sk->sk_state == TCP_LISTEN)
1425 goto out;
1426
1427 timeo = sock_rcvtimeo(sk, nonblock);
1428
1429 /* Urgent data needs to be handled specially. */
1430 if (flags & MSG_OOB)
1431 goto recv_urg;
1432
1433 seq = &tp->copied_seq;
1434 if (flags & MSG_PEEK) {
1435 peek_seq = tp->copied_seq;
1436 seq = &peek_seq;
1437 }
1438
1439 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
1440
1441#ifdef CONFIG_NET_DMA
1442 tp->ucopy.dma_chan = NULL;
1443 preempt_disable();
1444 skb = skb_peek_tail(&sk->sk_receive_queue);
1445 {
1446 int available = 0;
1447
1448 if (skb)
1449 available = TCP_SKB_CB(skb)->seq + skb->len - (*seq);
1450 if ((available < target) &&
1451 (len > sysctl_tcp_dma_copybreak) && !(flags & MSG_PEEK) &&
1452 !sysctl_tcp_low_latency &&
1453 dma_find_channel(DMA_MEMCPY)) {
1454 preempt_enable_no_resched();
1455 tp->ucopy.pinned_list =
1456 dma_pin_iovec_pages(msg->msg_iov, len);
1457 } else {
1458 preempt_enable_no_resched();
1459 }
1460 }
1461#endif
1462
1463 do {
1464 u32 offset;
1465
1466 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
1467 if (tp->urg_data && tp->urg_seq == *seq) {
1468 if (copied)
1469 break;
1470 if (signal_pending(current)) {
1471 copied = timeo ? sock_intr_errno(timeo) : -EAGAIN;
1472 break;
1473 }
1474 }
1475
1476 /* Next get a buffer. */
1477
1478 skb_queue_walk(&sk->sk_receive_queue, skb) {
1479 /* Now that we have two receive queues this
1480 * shouldn't happen.
1481 */
1482 if (WARN(before(*seq, TCP_SKB_CB(skb)->seq),
1483 "recvmsg bug: copied %X seq %X rcvnxt %X fl %X\n",
1484 *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt,
1485 flags))
1486 break;
1487
1488 offset = *seq - TCP_SKB_CB(skb)->seq;
1489 if (tcp_hdr(skb)->syn)
1490 offset--;
1491 if (offset < skb->len)
1492 goto found_ok_skb;
1493 if (tcp_hdr(skb)->fin)
1494 goto found_fin_ok;
1495 WARN(!(flags & MSG_PEEK),
1496 "recvmsg bug 2: copied %X seq %X rcvnxt %X fl %X\n",
1497 *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt, flags);
1498 }
1499
1500 /* Well, if we have backlog, try to process it now yet. */
1501
1502 if (copied >= target && !sk->sk_backlog.tail)
1503 break;
1504
1505 if (copied) {
1506 if (sk->sk_err ||
1507 sk->sk_state == TCP_CLOSE ||
1508 (sk->sk_shutdown & RCV_SHUTDOWN) ||
1509 !timeo ||
1510 signal_pending(current))
1511 break;
1512 } else {
1513 if (sock_flag(sk, SOCK_DONE))
1514 break;
1515
1516 if (sk->sk_err) {
1517 copied = sock_error(sk);
1518 break;
1519 }
1520
1521 if (sk->sk_shutdown & RCV_SHUTDOWN)
1522 break;
1523
1524 if (sk->sk_state == TCP_CLOSE) {
1525 if (!sock_flag(sk, SOCK_DONE)) {
1526 /* This occurs when user tries to read
1527 * from never connected socket.
1528 */
1529 copied = -ENOTCONN;
1530 break;
1531 }
1532 break;
1533 }
1534
1535 if (!timeo) {
1536 copied = -EAGAIN;
1537 break;
1538 }
1539
1540 if (signal_pending(current)) {
1541 copied = sock_intr_errno(timeo);
1542 break;
1543 }
1544 }
1545
1546 tcp_cleanup_rbuf(sk, copied);
1547
1548 if (!sysctl_tcp_low_latency && tp->ucopy.task == user_recv) {
1549 /* Install new reader */
1550 if (!user_recv && !(flags & (MSG_TRUNC | MSG_PEEK))) {
1551 user_recv = current;
1552 tp->ucopy.task = user_recv;
1553 tp->ucopy.iov = msg->msg_iov;
1554 }
1555
1556 tp->ucopy.len = len;
1557
1558 WARN_ON(tp->copied_seq != tp->rcv_nxt &&
1559 !(flags & (MSG_PEEK | MSG_TRUNC)));
1560
1561 /* Ugly... If prequeue is not empty, we have to
1562 * process it before releasing socket, otherwise
1563 * order will be broken at second iteration.
1564 * More elegant solution is required!!!
1565 *
1566 * Look: we have the following (pseudo)queues:
1567 *
1568 * 1. packets in flight
1569 * 2. backlog
1570 * 3. prequeue
1571 * 4. receive_queue
1572 *
1573 * Each queue can be processed only if the next ones
1574 * are empty. At this point we have empty receive_queue.
1575 * But prequeue _can_ be not empty after 2nd iteration,
1576 * when we jumped to start of loop because backlog
1577 * processing added something to receive_queue.
1578 * We cannot release_sock(), because backlog contains
1579 * packets arrived _after_ prequeued ones.
1580 *
1581 * Shortly, algorithm is clear --- to process all
1582 * the queues in order. We could make it more directly,
1583 * requeueing packets from backlog to prequeue, if
1584 * is not empty. It is more elegant, but eats cycles,
1585 * unfortunately.
1586 */
1587 if (!skb_queue_empty(&tp->ucopy.prequeue))
1588 goto do_prequeue;
1589
1590 /* __ Set realtime policy in scheduler __ */
1591 }
1592
1593#ifdef CONFIG_NET_DMA
1594 if (tp->ucopy.dma_chan)
1595 dma_async_memcpy_issue_pending(tp->ucopy.dma_chan);
1596#endif
1597 if (copied >= target) {
1598 /* Do not sleep, just process backlog. */
1599 release_sock(sk);
1600 lock_sock(sk);
1601 } else
1602 sk_wait_data(sk, &timeo);
1603
1604#ifdef CONFIG_NET_DMA
1605 tcp_service_net_dma(sk, false); /* Don't block */
1606 tp->ucopy.wakeup = 0;
1607#endif
1608
1609 if (user_recv) {
1610 int chunk;
1611
1612 /* __ Restore normal policy in scheduler __ */
1613
1614 if ((chunk = len - tp->ucopy.len) != 0) {
1615 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMBACKLOG, chunk);
1616 len -= chunk;
1617 copied += chunk;
1618 }
1619
1620 if (tp->rcv_nxt == tp->copied_seq &&
1621 !skb_queue_empty(&tp->ucopy.prequeue)) {
1622do_prequeue:
1623 tcp_prequeue_process(sk);
1624
1625 if ((chunk = len - tp->ucopy.len) != 0) {
1626 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1627 len -= chunk;
1628 copied += chunk;
1629 }
1630 }
1631 }
1632 if ((flags & MSG_PEEK) &&
1633 (peek_seq - copied - urg_hole != tp->copied_seq)) {
1634 if (net_ratelimit())
1635 printk(KERN_DEBUG "TCP(%s:%d): Application bug, race in MSG_PEEK.\n",
1636 current->comm, task_pid_nr(current));
1637 peek_seq = tp->copied_seq;
1638 }
1639 continue;
1640
1641 found_ok_skb:
1642 /* Ok so how much can we use? */
1643 used = skb->len - offset;
1644 if (len < used)
1645 used = len;
1646
1647 /* Do we have urgent data here? */
1648 if (tp->urg_data) {
1649 u32 urg_offset = tp->urg_seq - *seq;
1650 if (urg_offset < used) {
1651 if (!urg_offset) {
1652 if (!sock_flag(sk, SOCK_URGINLINE)) {
1653 ++*seq;
1654 urg_hole++;
1655 offset++;
1656 used--;
1657 if (!used)
1658 goto skip_copy;
1659 }
1660 } else
1661 used = urg_offset;
1662 }
1663 }
1664
1665 if (!(flags & MSG_TRUNC)) {
1666#ifdef CONFIG_NET_DMA
1667 if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list)
1668 tp->ucopy.dma_chan = dma_find_channel(DMA_MEMCPY);
1669
1670 if (tp->ucopy.dma_chan) {
1671 tp->ucopy.dma_cookie = dma_skb_copy_datagram_iovec(
1672 tp->ucopy.dma_chan, skb, offset,
1673 msg->msg_iov, used,
1674 tp->ucopy.pinned_list);
1675
1676 if (tp->ucopy.dma_cookie < 0) {
1677
1678 printk(KERN_ALERT "dma_cookie < 0\n");
1679
1680 /* Exception. Bailout! */
1681 if (!copied)
1682 copied = -EFAULT;
1683 break;
1684 }
1685
1686 dma_async_memcpy_issue_pending(tp->ucopy.dma_chan);
1687
1688 if ((offset + used) == skb->len)
1689 copied_early = 1;
1690
1691 } else
1692#endif
1693 {
1694 err = skb_copy_datagram_iovec(skb, offset,
1695 msg->msg_iov, used);
1696 if (err) {
1697 /* Exception. Bailout! */
1698 if (!copied)
1699 copied = -EFAULT;
1700 break;
1701 }
1702 }
1703 }
1704
1705 *seq += used;
1706 copied += used;
1707 len -= used;
1708
1709 tcp_rcv_space_adjust(sk);
1710
1711skip_copy:
1712 if (tp->urg_data && after(tp->copied_seq, tp->urg_seq)) {
1713 tp->urg_data = 0;
1714 tcp_fast_path_check(sk);
1715 }
1716 if (used + offset < skb->len)
1717 continue;
1718
1719 if (tcp_hdr(skb)->fin)
1720 goto found_fin_ok;
1721 if (!(flags & MSG_PEEK)) {
1722 sk_eat_skb(sk, skb, copied_early);
1723 copied_early = 0;
1724 }
1725 continue;
1726
1727 found_fin_ok:
1728 /* Process the FIN. */
1729 ++*seq;
1730 if (!(flags & MSG_PEEK)) {
1731 sk_eat_skb(sk, skb, copied_early);
1732 copied_early = 0;
1733 }
1734 break;
1735 } while (len > 0);
1736
1737 if (user_recv) {
1738 if (!skb_queue_empty(&tp->ucopy.prequeue)) {
1739 int chunk;
1740
1741 tp->ucopy.len = copied > 0 ? len : 0;
1742
1743 tcp_prequeue_process(sk);
1744
1745 if (copied > 0 && (chunk = len - tp->ucopy.len) != 0) {
1746 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1747 len -= chunk;
1748 copied += chunk;
1749 }
1750 }
1751
1752 tp->ucopy.task = NULL;
1753 tp->ucopy.len = 0;
1754 }
1755
1756#ifdef CONFIG_NET_DMA
1757 tcp_service_net_dma(sk, true); /* Wait for queue to drain */
1758 tp->ucopy.dma_chan = NULL;
1759
1760 if (tp->ucopy.pinned_list) {
1761 dma_unpin_iovec_pages(tp->ucopy.pinned_list);
1762 tp->ucopy.pinned_list = NULL;
1763 }
1764#endif
1765
1766 /* According to UNIX98, msg_name/msg_namelen are ignored
1767 * on connected socket. I was just happy when found this 8) --ANK
1768 */
1769
1770 /* Clean up data we have read: This will do ACK frames. */
1771 tcp_cleanup_rbuf(sk, copied);
1772
1773 release_sock(sk);
1774 return copied;
1775
1776out:
1777 release_sock(sk);
1778 return err;
1779
1780recv_urg:
1781 err = tcp_recv_urg(sk, msg, len, flags);
1782 goto out;
1783}
1784EXPORT_SYMBOL(tcp_recvmsg);
1785
1786void tcp_set_state(struct sock *sk, int state)
1787{
1788 int oldstate = sk->sk_state;
1789
1790 switch (state) {
1791 case TCP_ESTABLISHED:
1792 if (oldstate != TCP_ESTABLISHED)
1793 TCP_INC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
1794 break;
1795
1796 case TCP_CLOSE:
1797 if (oldstate == TCP_CLOSE_WAIT || oldstate == TCP_ESTABLISHED)
1798 TCP_INC_STATS(sock_net(sk), TCP_MIB_ESTABRESETS);
1799
1800 sk->sk_prot->unhash(sk);
1801 if (inet_csk(sk)->icsk_bind_hash &&
1802 !(sk->sk_userlocks & SOCK_BINDPORT_LOCK))
1803 inet_put_port(sk);
1804 /* fall through */
1805 default:
1806 if (oldstate == TCP_ESTABLISHED)
1807 TCP_DEC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
1808 }
1809
1810 /* Change state AFTER socket is unhashed to avoid closed
1811 * socket sitting in hash tables.
1812 */
1813 sk->sk_state = state;
1814
1815#ifdef STATE_TRACE
1816 SOCK_DEBUG(sk, "TCP sk=%p, State %s -> %s\n", sk, statename[oldstate], statename[state]);
1817#endif
1818}
1819EXPORT_SYMBOL_GPL(tcp_set_state);
1820
1821/*
1822 * State processing on a close. This implements the state shift for
1823 * sending our FIN frame. Note that we only send a FIN for some
1824 * states. A shutdown() may have already sent the FIN, or we may be
1825 * closed.
1826 */
1827
1828static const unsigned char new_state[16] = {
1829 /* current state: new state: action: */
1830 /* (Invalid) */ TCP_CLOSE,
1831 /* TCP_ESTABLISHED */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1832 /* TCP_SYN_SENT */ TCP_CLOSE,
1833 /* TCP_SYN_RECV */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1834 /* TCP_FIN_WAIT1 */ TCP_FIN_WAIT1,
1835 /* TCP_FIN_WAIT2 */ TCP_FIN_WAIT2,
1836 /* TCP_TIME_WAIT */ TCP_CLOSE,
1837 /* TCP_CLOSE */ TCP_CLOSE,
1838 /* TCP_CLOSE_WAIT */ TCP_LAST_ACK | TCP_ACTION_FIN,
1839 /* TCP_LAST_ACK */ TCP_LAST_ACK,
1840 /* TCP_LISTEN */ TCP_CLOSE,
1841 /* TCP_CLOSING */ TCP_CLOSING,
1842};
1843
1844static int tcp_close_state(struct sock *sk)
1845{
1846 int next = (int)new_state[sk->sk_state];
1847 int ns = next & TCP_STATE_MASK;
1848
1849 tcp_set_state(sk, ns);
1850
1851 return next & TCP_ACTION_FIN;
1852}
1853
1854/*
1855 * Shutdown the sending side of a connection. Much like close except
1856 * that we don't receive shut down or sock_set_flag(sk, SOCK_DEAD).
1857 */
1858
1859void tcp_shutdown(struct sock *sk, int how)
1860{
1861 /* We need to grab some memory, and put together a FIN,
1862 * and then put it into the queue to be sent.
1863 * Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
1864 */
1865 if (!(how & SEND_SHUTDOWN))
1866 return;
1867
1868 /* If we've already sent a FIN, or it's a closed state, skip this. */
1869 if ((1 << sk->sk_state) &
1870 (TCPF_ESTABLISHED | TCPF_SYN_SENT |
1871 TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) {
1872 /* Clear out any half completed packets. FIN if needed. */
1873 if (tcp_close_state(sk))
1874 tcp_send_fin(sk);
1875 }
1876}
1877EXPORT_SYMBOL(tcp_shutdown);
1878
1879void tcp_close(struct sock *sk, long timeout)
1880{
1881 struct sk_buff *skb;
1882 int data_was_unread = 0;
1883 int state;
1884
1885 lock_sock(sk);
1886 sk->sk_shutdown = SHUTDOWN_MASK;
1887
1888 if (sk->sk_state == TCP_LISTEN) {
1889 tcp_set_state(sk, TCP_CLOSE);
1890
1891 /* Special case. */
1892 inet_csk_listen_stop(sk);
1893
1894 goto adjudge_to_death;
1895 }
1896
1897 /* We need to flush the recv. buffs. We do this only on the
1898 * descriptor close, not protocol-sourced closes, because the
1899 * reader process may not have drained the data yet!
1900 */
1901 while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) {
1902 u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq -
1903 tcp_hdr(skb)->fin;
1904 data_was_unread += len;
1905 __kfree_skb(skb);
1906 }
1907
1908 sk_mem_reclaim(sk);
1909
1910 /* If socket has been already reset (e.g. in tcp_reset()) - kill it. */
1911 if (sk->sk_state == TCP_CLOSE)
1912 goto adjudge_to_death;
1913
1914 /* As outlined in RFC 2525, section 2.17, we send a RST here because
1915 * data was lost. To witness the awful effects of the old behavior of
1916 * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk
1917 * GET in an FTP client, suspend the process, wait for the client to
1918 * advertise a zero window, then kill -9 the FTP client, wheee...
1919 * Note: timeout is always zero in such a case.
1920 */
1921 if (data_was_unread) {
1922 /* Unread data was tossed, zap the connection. */
1923 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONCLOSE);
1924 tcp_set_state(sk, TCP_CLOSE);
1925 tcp_send_active_reset(sk, sk->sk_allocation);
1926 } else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) {
1927 /* Check zero linger _after_ checking for unread data. */
1928 sk->sk_prot->disconnect(sk, 0);
1929 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
1930 } else if (tcp_close_state(sk)) {
1931 /* We FIN if the application ate all the data before
1932 * zapping the connection.
1933 */
1934
1935 /* RED-PEN. Formally speaking, we have broken TCP state
1936 * machine. State transitions:
1937 *
1938 * TCP_ESTABLISHED -> TCP_FIN_WAIT1
1939 * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible)
1940 * TCP_CLOSE_WAIT -> TCP_LAST_ACK
1941 *
1942 * are legal only when FIN has been sent (i.e. in window),
1943 * rather than queued out of window. Purists blame.
1944 *
1945 * F.e. "RFC state" is ESTABLISHED,
1946 * if Linux state is FIN-WAIT-1, but FIN is still not sent.
1947 *
1948 * The visible declinations are that sometimes
1949 * we enter time-wait state, when it is not required really
1950 * (harmless), do not send active resets, when they are
1951 * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
1952 * they look as CLOSING or LAST_ACK for Linux)
1953 * Probably, I missed some more holelets.
1954 * --ANK
1955 */
1956 tcp_send_fin(sk);
1957 }
1958
1959 sk_stream_wait_close(sk, timeout);
1960
1961adjudge_to_death:
1962 state = sk->sk_state;
1963 sock_hold(sk);
1964 sock_orphan(sk);
1965
1966 /* It is the last release_sock in its life. It will remove backlog. */
1967 release_sock(sk);
1968
1969
1970 /* Now socket is owned by kernel and we acquire BH lock
1971 to finish close. No need to check for user refs.
1972 */
1973 local_bh_disable();
1974 bh_lock_sock(sk);
1975 WARN_ON(sock_owned_by_user(sk));
1976
1977 percpu_counter_inc(sk->sk_prot->orphan_count);
1978
1979 /* Have we already been destroyed by a softirq or backlog? */
1980 if (state != TCP_CLOSE && sk->sk_state == TCP_CLOSE)
1981 goto out;
1982
1983 /* This is a (useful) BSD violating of the RFC. There is a
1984 * problem with TCP as specified in that the other end could
1985 * keep a socket open forever with no application left this end.
1986 * We use a 3 minute timeout (about the same as BSD) then kill
1987 * our end. If they send after that then tough - BUT: long enough
1988 * that we won't make the old 4*rto = almost no time - whoops
1989 * reset mistake.
1990 *
1991 * Nope, it was not mistake. It is really desired behaviour
1992 * f.e. on http servers, when such sockets are useless, but
1993 * consume significant resources. Let's do it with special
1994 * linger2 option. --ANK
1995 */
1996
1997 if (sk->sk_state == TCP_FIN_WAIT2) {
1998 struct tcp_sock *tp = tcp_sk(sk);
1999 if (tp->linger2 < 0) {
2000 tcp_set_state(sk, TCP_CLOSE);
2001 tcp_send_active_reset(sk, GFP_ATOMIC);
2002 NET_INC_STATS_BH(sock_net(sk),
2003 LINUX_MIB_TCPABORTONLINGER);
2004 } else {
2005 const int tmo = tcp_fin_time(sk);
2006
2007 if (tmo > TCP_TIMEWAIT_LEN) {
2008 inet_csk_reset_keepalive_timer(sk,
2009 tmo - TCP_TIMEWAIT_LEN);
2010 } else {
2011 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
2012 goto out;
2013 }
2014 }
2015 }
2016 if (sk->sk_state != TCP_CLOSE) {
2017 sk_mem_reclaim(sk);
2018 if (tcp_too_many_orphans(sk, 0)) {
2019 if (net_ratelimit())
2020 printk(KERN_INFO "TCP: too many of orphaned "
2021 "sockets\n");
2022 tcp_set_state(sk, TCP_CLOSE);
2023 tcp_send_active_reset(sk, GFP_ATOMIC);
2024 NET_INC_STATS_BH(sock_net(sk),
2025 LINUX_MIB_TCPABORTONMEMORY);
2026 }
2027 }
2028
2029 if (sk->sk_state == TCP_CLOSE)
2030 inet_csk_destroy_sock(sk);
2031 /* Otherwise, socket is reprieved until protocol close. */
2032
2033out:
2034 bh_unlock_sock(sk);
2035 local_bh_enable();
2036 sock_put(sk);
2037}
2038EXPORT_SYMBOL(tcp_close);
2039
2040/* These states need RST on ABORT according to RFC793 */
2041
2042static inline int tcp_need_reset(int state)
2043{
2044 return (1 << state) &
2045 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 |
2046 TCPF_FIN_WAIT2 | TCPF_SYN_RECV);
2047}
2048
2049int tcp_disconnect(struct sock *sk, int flags)
2050{
2051 struct inet_sock *inet = inet_sk(sk);
2052 struct inet_connection_sock *icsk = inet_csk(sk);
2053 struct tcp_sock *tp = tcp_sk(sk);
2054 int err = 0;
2055 int old_state = sk->sk_state;
2056
2057 if (old_state != TCP_CLOSE)
2058 tcp_set_state(sk, TCP_CLOSE);
2059
2060 /* ABORT function of RFC793 */
2061 if (old_state == TCP_LISTEN) {
2062 inet_csk_listen_stop(sk);
2063 } else if (tcp_need_reset(old_state) ||
2064 (tp->snd_nxt != tp->write_seq &&
2065 (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK))) {
2066 /* The last check adjusts for discrepancy of Linux wrt. RFC
2067 * states
2068 */
2069 tcp_send_active_reset(sk, gfp_any());
2070 sk->sk_err = ECONNRESET;
2071 } else if (old_state == TCP_SYN_SENT)
2072 sk->sk_err = ECONNRESET;
2073
2074 tcp_clear_xmit_timers(sk);
2075 __skb_queue_purge(&sk->sk_receive_queue);
2076 tcp_write_queue_purge(sk);
2077 __skb_queue_purge(&tp->out_of_order_queue);
2078#ifdef CONFIG_NET_DMA
2079 __skb_queue_purge(&sk->sk_async_wait_queue);
2080#endif
2081
2082 inet->inet_dport = 0;
2083
2084 if (!(sk->sk_userlocks & SOCK_BINDADDR_LOCK))
2085 inet_reset_saddr(sk);
2086
2087 sk->sk_shutdown = 0;
2088 sock_reset_flag(sk, SOCK_DONE);
2089 tp->srtt = 0;
2090 if ((tp->write_seq += tp->max_window + 2) == 0)
2091 tp->write_seq = 1;
2092 icsk->icsk_backoff = 0;
2093 tp->snd_cwnd = 2;
2094 icsk->icsk_probes_out = 0;
2095 tp->packets_out = 0;
2096 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
2097 tp->snd_cwnd_cnt = 0;
2098 tp->bytes_acked = 0;
2099 tp->window_clamp = 0;
2100 tcp_set_ca_state(sk, TCP_CA_Open);
2101 tcp_clear_retrans(tp);
2102 inet_csk_delack_init(sk);
2103 tcp_init_send_head(sk);
2104 memset(&tp->rx_opt, 0, sizeof(tp->rx_opt));
2105 __sk_dst_reset(sk);
2106
2107 WARN_ON(inet->inet_num && !icsk->icsk_bind_hash);
2108
2109 sk->sk_error_report(sk);
2110 return err;
2111}
2112EXPORT_SYMBOL(tcp_disconnect);
2113
2114/*
2115 * Socket option code for TCP.
2116 */
2117static int do_tcp_setsockopt(struct sock *sk, int level,
2118 int optname, char __user *optval, unsigned int optlen)
2119{
2120 struct tcp_sock *tp = tcp_sk(sk);
2121 struct inet_connection_sock *icsk = inet_csk(sk);
2122 int val;
2123 int err = 0;
2124
2125 /* These are data/string values, all the others are ints */
2126 switch (optname) {
2127 case TCP_CONGESTION: {
2128 char name[TCP_CA_NAME_MAX];
2129
2130 if (optlen < 1)
2131 return -EINVAL;
2132
2133 val = strncpy_from_user(name, optval,
2134 min_t(long, TCP_CA_NAME_MAX-1, optlen));
2135 if (val < 0)
2136 return -EFAULT;
2137 name[val] = 0;
2138
2139 lock_sock(sk);
2140 err = tcp_set_congestion_control(sk, name);
2141 release_sock(sk);
2142 return err;
2143 }
2144 case TCP_COOKIE_TRANSACTIONS: {
2145 struct tcp_cookie_transactions ctd;
2146 struct tcp_cookie_values *cvp = NULL;
2147
2148 if (sizeof(ctd) > optlen)
2149 return -EINVAL;
2150 if (copy_from_user(&ctd, optval, sizeof(ctd)))
2151 return -EFAULT;
2152
2153 if (ctd.tcpct_used > sizeof(ctd.tcpct_value) ||
2154 ctd.tcpct_s_data_desired > TCP_MSS_DESIRED)
2155 return -EINVAL;
2156
2157 if (ctd.tcpct_cookie_desired == 0) {
2158 /* default to global value */
2159 } else if ((0x1 & ctd.tcpct_cookie_desired) ||
2160 ctd.tcpct_cookie_desired > TCP_COOKIE_MAX ||
2161 ctd.tcpct_cookie_desired < TCP_COOKIE_MIN) {
2162 return -EINVAL;
2163 }
2164
2165 if (TCP_COOKIE_OUT_NEVER & ctd.tcpct_flags) {
2166 /* Supercedes all other values */
2167 lock_sock(sk);
2168 if (tp->cookie_values != NULL) {
2169 kref_put(&tp->cookie_values->kref,
2170 tcp_cookie_values_release);
2171 tp->cookie_values = NULL;
2172 }
2173 tp->rx_opt.cookie_in_always = 0; /* false */
2174 tp->rx_opt.cookie_out_never = 1; /* true */
2175 release_sock(sk);
2176 return err;
2177 }
2178
2179 /* Allocate ancillary memory before locking.
2180 */
2181 if (ctd.tcpct_used > 0 ||
2182 (tp->cookie_values == NULL &&
2183 (sysctl_tcp_cookie_size > 0 ||
2184 ctd.tcpct_cookie_desired > 0 ||
2185 ctd.tcpct_s_data_desired > 0))) {
2186 cvp = kzalloc(sizeof(*cvp) + ctd.tcpct_used,
2187 GFP_KERNEL);
2188 if (cvp == NULL)
2189 return -ENOMEM;
2190
2191 kref_init(&cvp->kref);
2192 }
2193 lock_sock(sk);
2194 tp->rx_opt.cookie_in_always =
2195 (TCP_COOKIE_IN_ALWAYS & ctd.tcpct_flags);
2196 tp->rx_opt.cookie_out_never = 0; /* false */
2197
2198 if (tp->cookie_values != NULL) {
2199 if (cvp != NULL) {
2200 /* Changed values are recorded by a changed
2201 * pointer, ensuring the cookie will differ,
2202 * without separately hashing each value later.
2203 */
2204 kref_put(&tp->cookie_values->kref,
2205 tcp_cookie_values_release);
2206 } else {
2207 cvp = tp->cookie_values;
2208 }
2209 }
2210
2211 if (cvp != NULL) {
2212 cvp->cookie_desired = ctd.tcpct_cookie_desired;
2213
2214 if (ctd.tcpct_used > 0) {
2215 memcpy(cvp->s_data_payload, ctd.tcpct_value,
2216 ctd.tcpct_used);
2217 cvp->s_data_desired = ctd.tcpct_used;
2218 cvp->s_data_constant = 1; /* true */
2219 } else {
2220 /* No constant payload data. */
2221 cvp->s_data_desired = ctd.tcpct_s_data_desired;
2222 cvp->s_data_constant = 0; /* false */
2223 }
2224
2225 tp->cookie_values = cvp;
2226 }
2227 release_sock(sk);
2228 return err;
2229 }
2230 default:
2231 /* fallthru */
2232 break;
2233 }
2234
2235 if (optlen < sizeof(int))
2236 return -EINVAL;
2237
2238 if (get_user(val, (int __user *)optval))
2239 return -EFAULT;
2240
2241 lock_sock(sk);
2242
2243 switch (optname) {
2244 case TCP_MAXSEG:
2245 /* Values greater than interface MTU won't take effect. However
2246 * at the point when this call is done we typically don't yet
2247 * know which interface is going to be used */
2248 if (val < TCP_MIN_MSS || val > MAX_TCP_WINDOW) {
2249 err = -EINVAL;
2250 break;
2251 }
2252 tp->rx_opt.user_mss = val;
2253 break;
2254
2255 case TCP_NODELAY:
2256 if (val) {
2257 /* TCP_NODELAY is weaker than TCP_CORK, so that
2258 * this option on corked socket is remembered, but
2259 * it is not activated until cork is cleared.
2260 *
2261 * However, when TCP_NODELAY is set we make
2262 * an explicit push, which overrides even TCP_CORK
2263 * for currently queued segments.
2264 */
2265 tp->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH;
2266 tcp_push_pending_frames(sk);
2267 } else {
2268 tp->nonagle &= ~TCP_NAGLE_OFF;
2269 }
2270 break;
2271
2272 case TCP_THIN_LINEAR_TIMEOUTS:
2273 if (val < 0 || val > 1)
2274 err = -EINVAL;
2275 else
2276 tp->thin_lto = val;
2277 break;
2278
2279 case TCP_THIN_DUPACK:
2280 if (val < 0 || val > 1)
2281 err = -EINVAL;
2282 else
2283 tp->thin_dupack = val;
2284 break;
2285
2286 case TCP_CORK:
2287 /* When set indicates to always queue non-full frames.
2288 * Later the user clears this option and we transmit
2289 * any pending partial frames in the queue. This is
2290 * meant to be used alongside sendfile() to get properly
2291 * filled frames when the user (for example) must write
2292 * out headers with a write() call first and then use
2293 * sendfile to send out the data parts.
2294 *
2295 * TCP_CORK can be set together with TCP_NODELAY and it is
2296 * stronger than TCP_NODELAY.
2297 */
2298 if (val) {
2299 tp->nonagle |= TCP_NAGLE_CORK;
2300 } else {
2301 tp->nonagle &= ~TCP_NAGLE_CORK;
2302 if (tp->nonagle&TCP_NAGLE_OFF)
2303 tp->nonagle |= TCP_NAGLE_PUSH;
2304 tcp_push_pending_frames(sk);
2305 }
2306 break;
2307
2308 case TCP_KEEPIDLE:
2309 if (val < 1 || val > MAX_TCP_KEEPIDLE)
2310 err = -EINVAL;
2311 else {
2312 tp->keepalive_time = val * HZ;
2313 if (sock_flag(sk, SOCK_KEEPOPEN) &&
2314 !((1 << sk->sk_state) &
2315 (TCPF_CLOSE | TCPF_LISTEN))) {
2316 u32 elapsed = keepalive_time_elapsed(tp);
2317 if (tp->keepalive_time > elapsed)
2318 elapsed = tp->keepalive_time - elapsed;
2319 else
2320 elapsed = 0;
2321 inet_csk_reset_keepalive_timer(sk, elapsed);
2322 }
2323 }
2324 break;
2325 case TCP_KEEPINTVL:
2326 if (val < 1 || val > MAX_TCP_KEEPINTVL)
2327 err = -EINVAL;
2328 else
2329 tp->keepalive_intvl = val * HZ;
2330 break;
2331 case TCP_KEEPCNT:
2332 if (val < 1 || val > MAX_TCP_KEEPCNT)
2333 err = -EINVAL;
2334 else
2335 tp->keepalive_probes = val;
2336 break;
2337 case TCP_SYNCNT:
2338 if (val < 1 || val > MAX_TCP_SYNCNT)
2339 err = -EINVAL;
2340 else
2341 icsk->icsk_syn_retries = val;
2342 break;
2343
2344 case TCP_LINGER2:
2345 if (val < 0)
2346 tp->linger2 = -1;
2347 else if (val > sysctl_tcp_fin_timeout / HZ)
2348 tp->linger2 = 0;
2349 else
2350 tp->linger2 = val * HZ;
2351 break;
2352
2353 case TCP_DEFER_ACCEPT:
2354 /* Translate value in seconds to number of retransmits */
2355 icsk->icsk_accept_queue.rskq_defer_accept =
2356 secs_to_retrans(val, TCP_TIMEOUT_INIT / HZ,
2357 TCP_RTO_MAX / HZ);
2358 break;
2359
2360 case TCP_WINDOW_CLAMP:
2361 if (!val) {
2362 if (sk->sk_state != TCP_CLOSE) {
2363 err = -EINVAL;
2364 break;
2365 }
2366 tp->window_clamp = 0;
2367 } else
2368 tp->window_clamp = val < SOCK_MIN_RCVBUF / 2 ?
2369 SOCK_MIN_RCVBUF / 2 : val;
2370 break;
2371
2372 case TCP_QUICKACK:
2373 if (!val) {
2374 icsk->icsk_ack.pingpong = 1;
2375 } else {
2376 icsk->icsk_ack.pingpong = 0;
2377 if ((1 << sk->sk_state) &
2378 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) &&
2379 inet_csk_ack_scheduled(sk)) {
2380 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
2381 tcp_cleanup_rbuf(sk, 1);
2382 if (!(val & 1))
2383 icsk->icsk_ack.pingpong = 1;
2384 }
2385 }
2386 break;
2387
2388#ifdef CONFIG_TCP_MD5SIG
2389 case TCP_MD5SIG:
2390 /* Read the IP->Key mappings from userspace */
2391 err = tp->af_specific->md5_parse(sk, optval, optlen);
2392 break;
2393#endif
2394 case TCP_USER_TIMEOUT:
2395 /* Cap the max timeout in ms TCP will retry/retrans
2396 * before giving up and aborting (ETIMEDOUT) a connection.
2397 */
2398 icsk->icsk_user_timeout = msecs_to_jiffies(val);
2399 break;
2400 default:
2401 err = -ENOPROTOOPT;
2402 break;
2403 }
2404
2405 release_sock(sk);
2406 return err;
2407}
2408
2409int tcp_setsockopt(struct sock *sk, int level, int optname, char __user *optval,
2410 unsigned int optlen)
2411{
2412 struct inet_connection_sock *icsk = inet_csk(sk);
2413
2414 if (level != SOL_TCP)
2415 return icsk->icsk_af_ops->setsockopt(sk, level, optname,
2416 optval, optlen);
2417 return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2418}
2419EXPORT_SYMBOL(tcp_setsockopt);
2420
2421#ifdef CONFIG_COMPAT
2422int compat_tcp_setsockopt(struct sock *sk, int level, int optname,
2423 char __user *optval, unsigned int optlen)
2424{
2425 if (level != SOL_TCP)
2426 return inet_csk_compat_setsockopt(sk, level, optname,
2427 optval, optlen);
2428 return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2429}
2430EXPORT_SYMBOL(compat_tcp_setsockopt);
2431#endif
2432
2433/* Return information about state of tcp endpoint in API format. */
2434void tcp_get_info(struct sock *sk, struct tcp_info *info)
2435{
2436 struct tcp_sock *tp = tcp_sk(sk);
2437 const struct inet_connection_sock *icsk = inet_csk(sk);
2438 u32 now = tcp_time_stamp;
2439
2440 memset(info, 0, sizeof(*info));
2441
2442 info->tcpi_state = sk->sk_state;
2443 info->tcpi_ca_state = icsk->icsk_ca_state;
2444 info->tcpi_retransmits = icsk->icsk_retransmits;
2445 info->tcpi_probes = icsk->icsk_probes_out;
2446 info->tcpi_backoff = icsk->icsk_backoff;
2447
2448 if (tp->rx_opt.tstamp_ok)
2449 info->tcpi_options |= TCPI_OPT_TIMESTAMPS;
2450 if (tcp_is_sack(tp))
2451 info->tcpi_options |= TCPI_OPT_SACK;
2452 if (tp->rx_opt.wscale_ok) {
2453 info->tcpi_options |= TCPI_OPT_WSCALE;
2454 info->tcpi_snd_wscale = tp->rx_opt.snd_wscale;
2455 info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale;
2456 }
2457
2458 if (tp->ecn_flags&TCP_ECN_OK)
2459 info->tcpi_options |= TCPI_OPT_ECN;
2460
2461 info->tcpi_rto = jiffies_to_usecs(icsk->icsk_rto);
2462 info->tcpi_ato = jiffies_to_usecs(icsk->icsk_ack.ato);
2463 info->tcpi_snd_mss = tp->mss_cache;
2464 info->tcpi_rcv_mss = icsk->icsk_ack.rcv_mss;
2465
2466 if (sk->sk_state == TCP_LISTEN) {
2467 info->tcpi_unacked = sk->sk_ack_backlog;
2468 info->tcpi_sacked = sk->sk_max_ack_backlog;
2469 } else {
2470 info->tcpi_unacked = tp->packets_out;
2471 info->tcpi_sacked = tp->sacked_out;
2472 }
2473 info->tcpi_lost = tp->lost_out;
2474 info->tcpi_retrans = tp->retrans_out;
2475 info->tcpi_fackets = tp->fackets_out;
2476
2477 info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime);
2478 info->tcpi_last_data_recv = jiffies_to_msecs(now - icsk->icsk_ack.lrcvtime);
2479 info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp);
2480
2481 info->tcpi_pmtu = icsk->icsk_pmtu_cookie;
2482 info->tcpi_rcv_ssthresh = tp->rcv_ssthresh;
2483 info->tcpi_rtt = jiffies_to_usecs(tp->srtt)>>3;
2484 info->tcpi_rttvar = jiffies_to_usecs(tp->mdev)>>2;
2485 info->tcpi_snd_ssthresh = tp->snd_ssthresh;
2486 info->tcpi_snd_cwnd = tp->snd_cwnd;
2487 info->tcpi_advmss = tp->advmss;
2488 info->tcpi_reordering = tp->reordering;
2489
2490 info->tcpi_rcv_rtt = jiffies_to_usecs(tp->rcv_rtt_est.rtt)>>3;
2491 info->tcpi_rcv_space = tp->rcvq_space.space;
2492
2493 info->tcpi_total_retrans = tp->total_retrans;
2494}
2495EXPORT_SYMBOL_GPL(tcp_get_info);
2496
2497static int do_tcp_getsockopt(struct sock *sk, int level,
2498 int optname, char __user *optval, int __user *optlen)
2499{
2500 struct inet_connection_sock *icsk = inet_csk(sk);
2501 struct tcp_sock *tp = tcp_sk(sk);
2502 int val, len;
2503
2504 if (get_user(len, optlen))
2505 return -EFAULT;
2506
2507 len = min_t(unsigned int, len, sizeof(int));
2508
2509 if (len < 0)
2510 return -EINVAL;
2511
2512 switch (optname) {
2513 case TCP_MAXSEG:
2514 val = tp->mss_cache;
2515 if (!val && ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
2516 val = tp->rx_opt.user_mss;
2517 break;
2518 case TCP_NODELAY:
2519 val = !!(tp->nonagle&TCP_NAGLE_OFF);
2520 break;
2521 case TCP_CORK:
2522 val = !!(tp->nonagle&TCP_NAGLE_CORK);
2523 break;
2524 case TCP_KEEPIDLE:
2525 val = keepalive_time_when(tp) / HZ;
2526 break;
2527 case TCP_KEEPINTVL:
2528 val = keepalive_intvl_when(tp) / HZ;
2529 break;
2530 case TCP_KEEPCNT:
2531 val = keepalive_probes(tp);
2532 break;
2533 case TCP_SYNCNT:
2534 val = icsk->icsk_syn_retries ? : sysctl_tcp_syn_retries;
2535 break;
2536 case TCP_LINGER2:
2537 val = tp->linger2;
2538 if (val >= 0)
2539 val = (val ? : sysctl_tcp_fin_timeout) / HZ;
2540 break;
2541 case TCP_DEFER_ACCEPT:
2542 val = retrans_to_secs(icsk->icsk_accept_queue.rskq_defer_accept,
2543 TCP_TIMEOUT_INIT / HZ, TCP_RTO_MAX / HZ);
2544 break;
2545 case TCP_WINDOW_CLAMP:
2546 val = tp->window_clamp;
2547 break;
2548 case TCP_INFO: {
2549 struct tcp_info info;
2550
2551 if (get_user(len, optlen))
2552 return -EFAULT;
2553
2554 tcp_get_info(sk, &info);
2555
2556 len = min_t(unsigned int, len, sizeof(info));
2557 if (put_user(len, optlen))
2558 return -EFAULT;
2559 if (copy_to_user(optval, &info, len))
2560 return -EFAULT;
2561 return 0;
2562 }
2563 case TCP_QUICKACK:
2564 val = !icsk->icsk_ack.pingpong;
2565 break;
2566
2567 case TCP_CONGESTION:
2568 if (get_user(len, optlen))
2569 return -EFAULT;
2570 len = min_t(unsigned int, len, TCP_CA_NAME_MAX);
2571 if (put_user(len, optlen))
2572 return -EFAULT;
2573 if (copy_to_user(optval, icsk->icsk_ca_ops->name, len))
2574 return -EFAULT;
2575 return 0;
2576
2577 case TCP_COOKIE_TRANSACTIONS: {
2578 struct tcp_cookie_transactions ctd;
2579 struct tcp_cookie_values *cvp = tp->cookie_values;
2580
2581 if (get_user(len, optlen))
2582 return -EFAULT;
2583 if (len < sizeof(ctd))
2584 return -EINVAL;
2585
2586 memset(&ctd, 0, sizeof(ctd));
2587 ctd.tcpct_flags = (tp->rx_opt.cookie_in_always ?
2588 TCP_COOKIE_IN_ALWAYS : 0)
2589 | (tp->rx_opt.cookie_out_never ?
2590 TCP_COOKIE_OUT_NEVER : 0);
2591
2592 if (cvp != NULL) {
2593 ctd.tcpct_flags |= (cvp->s_data_in ?
2594 TCP_S_DATA_IN : 0)
2595 | (cvp->s_data_out ?
2596 TCP_S_DATA_OUT : 0);
2597
2598 ctd.tcpct_cookie_desired = cvp->cookie_desired;
2599 ctd.tcpct_s_data_desired = cvp->s_data_desired;
2600
2601 memcpy(&ctd.tcpct_value[0], &cvp->cookie_pair[0],
2602 cvp->cookie_pair_size);
2603 ctd.tcpct_used = cvp->cookie_pair_size;
2604 }
2605
2606 if (put_user(sizeof(ctd), optlen))
2607 return -EFAULT;
2608 if (copy_to_user(optval, &ctd, sizeof(ctd)))
2609 return -EFAULT;
2610 return 0;
2611 }
2612 case TCP_THIN_LINEAR_TIMEOUTS:
2613 val = tp->thin_lto;
2614 break;
2615 case TCP_THIN_DUPACK:
2616 val = tp->thin_dupack;
2617 break;
2618
2619 case TCP_USER_TIMEOUT:
2620 val = jiffies_to_msecs(icsk->icsk_user_timeout);
2621 break;
2622 default:
2623 return -ENOPROTOOPT;
2624 }
2625
2626 if (put_user(len, optlen))
2627 return -EFAULT;
2628 if (copy_to_user(optval, &val, len))
2629 return -EFAULT;
2630 return 0;
2631}
2632
2633int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval,
2634 int __user *optlen)
2635{
2636 struct inet_connection_sock *icsk = inet_csk(sk);
2637
2638 if (level != SOL_TCP)
2639 return icsk->icsk_af_ops->getsockopt(sk, level, optname,
2640 optval, optlen);
2641 return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2642}
2643EXPORT_SYMBOL(tcp_getsockopt);
2644
2645#ifdef CONFIG_COMPAT
2646int compat_tcp_getsockopt(struct sock *sk, int level, int optname,
2647 char __user *optval, int __user *optlen)
2648{
2649 if (level != SOL_TCP)
2650 return inet_csk_compat_getsockopt(sk, level, optname,
2651 optval, optlen);
2652 return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2653}
2654EXPORT_SYMBOL(compat_tcp_getsockopt);
2655#endif
2656
2657struct sk_buff *tcp_tso_segment(struct sk_buff *skb, u32 features)
2658{
2659 struct sk_buff *segs = ERR_PTR(-EINVAL);
2660 struct tcphdr *th;
2661 unsigned thlen;
2662 unsigned int seq;
2663 __be32 delta;
2664 unsigned int oldlen;
2665 unsigned int mss;
2666
2667 if (!pskb_may_pull(skb, sizeof(*th)))
2668 goto out;
2669
2670 th = tcp_hdr(skb);
2671 thlen = th->doff * 4;
2672 if (thlen < sizeof(*th))
2673 goto out;
2674
2675 if (!pskb_may_pull(skb, thlen))
2676 goto out;
2677
2678 oldlen = (u16)~skb->len;
2679 __skb_pull(skb, thlen);
2680
2681 mss = skb_shinfo(skb)->gso_size;
2682 if (unlikely(skb->len <= mss))
2683 goto out;
2684
2685 if (skb_gso_ok(skb, features | NETIF_F_GSO_ROBUST)) {
2686 /* Packet is from an untrusted source, reset gso_segs. */
2687 int type = skb_shinfo(skb)->gso_type;
2688
2689 if (unlikely(type &
2690 ~(SKB_GSO_TCPV4 |
2691 SKB_GSO_DODGY |
2692 SKB_GSO_TCP_ECN |
2693 SKB_GSO_TCPV6 |
2694 0) ||
2695 !(type & (SKB_GSO_TCPV4 | SKB_GSO_TCPV6))))
2696 goto out;
2697
2698 skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss);
2699
2700 segs = NULL;
2701 goto out;
2702 }
2703
2704 segs = skb_segment(skb, features);
2705 if (IS_ERR(segs))
2706 goto out;
2707
2708 delta = htonl(oldlen + (thlen + mss));
2709
2710 skb = segs;
2711 th = tcp_hdr(skb);
2712 seq = ntohl(th->seq);
2713
2714 do {
2715 th->fin = th->psh = 0;
2716
2717 th->check = ~csum_fold((__force __wsum)((__force u32)th->check +
2718 (__force u32)delta));
2719 if (skb->ip_summed != CHECKSUM_PARTIAL)
2720 th->check =
2721 csum_fold(csum_partial(skb_transport_header(skb),
2722 thlen, skb->csum));
2723
2724 seq += mss;
2725 skb = skb->next;
2726 th = tcp_hdr(skb);
2727
2728 th->seq = htonl(seq);
2729 th->cwr = 0;
2730 } while (skb->next);
2731
2732 delta = htonl(oldlen + (skb->tail - skb->transport_header) +
2733 skb->data_len);
2734 th->check = ~csum_fold((__force __wsum)((__force u32)th->check +
2735 (__force u32)delta));
2736 if (skb->ip_summed != CHECKSUM_PARTIAL)
2737 th->check = csum_fold(csum_partial(skb_transport_header(skb),
2738 thlen, skb->csum));
2739
2740out:
2741 return segs;
2742}
2743EXPORT_SYMBOL(tcp_tso_segment);
2744
2745struct sk_buff **tcp_gro_receive(struct sk_buff **head, struct sk_buff *skb)
2746{
2747 struct sk_buff **pp = NULL;
2748 struct sk_buff *p;
2749 struct tcphdr *th;
2750 struct tcphdr *th2;
2751 unsigned int len;
2752 unsigned int thlen;
2753 __be32 flags;
2754 unsigned int mss = 1;
2755 unsigned int hlen;
2756 unsigned int off;
2757 int flush = 1;
2758 int i;
2759
2760 off = skb_gro_offset(skb);
2761 hlen = off + sizeof(*th);
2762 th = skb_gro_header_fast(skb, off);
2763 if (skb_gro_header_hard(skb, hlen)) {
2764 th = skb_gro_header_slow(skb, hlen, off);
2765 if (unlikely(!th))
2766 goto out;
2767 }
2768
2769 thlen = th->doff * 4;
2770 if (thlen < sizeof(*th))
2771 goto out;
2772
2773 hlen = off + thlen;
2774 if (skb_gro_header_hard(skb, hlen)) {
2775 th = skb_gro_header_slow(skb, hlen, off);
2776 if (unlikely(!th))
2777 goto out;
2778 }
2779
2780 skb_gro_pull(skb, thlen);
2781
2782 len = skb_gro_len(skb);
2783 flags = tcp_flag_word(th);
2784
2785 for (; (p = *head); head = &p->next) {
2786 if (!NAPI_GRO_CB(p)->same_flow)
2787 continue;
2788
2789 th2 = tcp_hdr(p);
2790
2791 if (*(u32 *)&th->source ^ *(u32 *)&th2->source) {
2792 NAPI_GRO_CB(p)->same_flow = 0;
2793 continue;
2794 }
2795
2796 goto found;
2797 }
2798
2799 goto out_check_final;
2800
2801found:
2802 flush = NAPI_GRO_CB(p)->flush;
2803 flush |= (__force int)(flags & TCP_FLAG_CWR);
2804 flush |= (__force int)((flags ^ tcp_flag_word(th2)) &
2805 ~(TCP_FLAG_CWR | TCP_FLAG_FIN | TCP_FLAG_PSH));
2806 flush |= (__force int)(th->ack_seq ^ th2->ack_seq);
2807 for (i = sizeof(*th); i < thlen; i += 4)
2808 flush |= *(u32 *)((u8 *)th + i) ^
2809 *(u32 *)((u8 *)th2 + i);
2810
2811 mss = skb_shinfo(p)->gso_size;
2812
2813 flush |= (len - 1) >= mss;
2814 flush |= (ntohl(th2->seq) + skb_gro_len(p)) ^ ntohl(th->seq);
2815
2816 if (flush || skb_gro_receive(head, skb)) {
2817 mss = 1;
2818 goto out_check_final;
2819 }
2820
2821 p = *head;
2822 th2 = tcp_hdr(p);
2823 tcp_flag_word(th2) |= flags & (TCP_FLAG_FIN | TCP_FLAG_PSH);
2824
2825out_check_final:
2826 flush = len < mss;
2827 flush |= (__force int)(flags & (TCP_FLAG_URG | TCP_FLAG_PSH |
2828 TCP_FLAG_RST | TCP_FLAG_SYN |
2829 TCP_FLAG_FIN));
2830
2831 if (p && (!NAPI_GRO_CB(skb)->same_flow || flush))
2832 pp = head;
2833
2834out:
2835 NAPI_GRO_CB(skb)->flush |= flush;
2836
2837 return pp;
2838}
2839EXPORT_SYMBOL(tcp_gro_receive);
2840
2841int tcp_gro_complete(struct sk_buff *skb)
2842{
2843 struct tcphdr *th = tcp_hdr(skb);
2844
2845 skb->csum_start = skb_transport_header(skb) - skb->head;
2846 skb->csum_offset = offsetof(struct tcphdr, check);
2847 skb->ip_summed = CHECKSUM_PARTIAL;
2848
2849 skb_shinfo(skb)->gso_segs = NAPI_GRO_CB(skb)->count;
2850
2851 if (th->cwr)
2852 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
2853
2854 return 0;
2855}
2856EXPORT_SYMBOL(tcp_gro_complete);
2857
2858#ifdef CONFIG_TCP_MD5SIG
2859static unsigned long tcp_md5sig_users;
2860static struct tcp_md5sig_pool * __percpu *tcp_md5sig_pool;
2861static DEFINE_SPINLOCK(tcp_md5sig_pool_lock);
2862
2863static void __tcp_free_md5sig_pool(struct tcp_md5sig_pool * __percpu *pool)
2864{
2865 int cpu;
2866 for_each_possible_cpu(cpu) {
2867 struct tcp_md5sig_pool *p = *per_cpu_ptr(pool, cpu);
2868 if (p) {
2869 if (p->md5_desc.tfm)
2870 crypto_free_hash(p->md5_desc.tfm);
2871 kfree(p);
2872 }
2873 }
2874 free_percpu(pool);
2875}
2876
2877void tcp_free_md5sig_pool(void)
2878{
2879 struct tcp_md5sig_pool * __percpu *pool = NULL;
2880
2881 spin_lock_bh(&tcp_md5sig_pool_lock);
2882 if (--tcp_md5sig_users == 0) {
2883 pool = tcp_md5sig_pool;
2884 tcp_md5sig_pool = NULL;
2885 }
2886 spin_unlock_bh(&tcp_md5sig_pool_lock);
2887 if (pool)
2888 __tcp_free_md5sig_pool(pool);
2889}
2890EXPORT_SYMBOL(tcp_free_md5sig_pool);
2891
2892static struct tcp_md5sig_pool * __percpu *
2893__tcp_alloc_md5sig_pool(struct sock *sk)
2894{
2895 int cpu;
2896 struct tcp_md5sig_pool * __percpu *pool;
2897
2898 pool = alloc_percpu(struct tcp_md5sig_pool *);
2899 if (!pool)
2900 return NULL;
2901
2902 for_each_possible_cpu(cpu) {
2903 struct tcp_md5sig_pool *p;
2904 struct crypto_hash *hash;
2905
2906 p = kzalloc(sizeof(*p), sk->sk_allocation);
2907 if (!p)
2908 goto out_free;
2909 *per_cpu_ptr(pool, cpu) = p;
2910
2911 hash = crypto_alloc_hash("md5", 0, CRYPTO_ALG_ASYNC);
2912 if (!hash || IS_ERR(hash))
2913 goto out_free;
2914
2915 p->md5_desc.tfm = hash;
2916 }
2917 return pool;
2918out_free:
2919 __tcp_free_md5sig_pool(pool);
2920 return NULL;
2921}
2922
2923struct tcp_md5sig_pool * __percpu *tcp_alloc_md5sig_pool(struct sock *sk)
2924{
2925 struct tcp_md5sig_pool * __percpu *pool;
2926 int alloc = 0;
2927
2928retry:
2929 spin_lock_bh(&tcp_md5sig_pool_lock);
2930 pool = tcp_md5sig_pool;
2931 if (tcp_md5sig_users++ == 0) {
2932 alloc = 1;
2933 spin_unlock_bh(&tcp_md5sig_pool_lock);
2934 } else if (!pool) {
2935 tcp_md5sig_users--;
2936 spin_unlock_bh(&tcp_md5sig_pool_lock);
2937 cpu_relax();
2938 goto retry;
2939 } else
2940 spin_unlock_bh(&tcp_md5sig_pool_lock);
2941
2942 if (alloc) {
2943 /* we cannot hold spinlock here because this may sleep. */
2944 struct tcp_md5sig_pool * __percpu *p;
2945
2946 p = __tcp_alloc_md5sig_pool(sk);
2947 spin_lock_bh(&tcp_md5sig_pool_lock);
2948 if (!p) {
2949 tcp_md5sig_users--;
2950 spin_unlock_bh(&tcp_md5sig_pool_lock);
2951 return NULL;
2952 }
2953 pool = tcp_md5sig_pool;
2954 if (pool) {
2955 /* oops, it has already been assigned. */
2956 spin_unlock_bh(&tcp_md5sig_pool_lock);
2957 __tcp_free_md5sig_pool(p);
2958 } else {
2959 tcp_md5sig_pool = pool = p;
2960 spin_unlock_bh(&tcp_md5sig_pool_lock);
2961 }
2962 }
2963 return pool;
2964}
2965EXPORT_SYMBOL(tcp_alloc_md5sig_pool);
2966
2967
2968/**
2969 * tcp_get_md5sig_pool - get md5sig_pool for this user
2970 *
2971 * We use percpu structure, so if we succeed, we exit with preemption
2972 * and BH disabled, to make sure another thread or softirq handling
2973 * wont try to get same context.
2974 */
2975struct tcp_md5sig_pool *tcp_get_md5sig_pool(void)
2976{
2977 struct tcp_md5sig_pool * __percpu *p;
2978
2979 local_bh_disable();
2980
2981 spin_lock(&tcp_md5sig_pool_lock);
2982 p = tcp_md5sig_pool;
2983 if (p)
2984 tcp_md5sig_users++;
2985 spin_unlock(&tcp_md5sig_pool_lock);
2986
2987 if (p)
2988 return *this_cpu_ptr(p);
2989
2990 local_bh_enable();
2991 return NULL;
2992}
2993EXPORT_SYMBOL(tcp_get_md5sig_pool);
2994
2995void tcp_put_md5sig_pool(void)
2996{
2997 local_bh_enable();
2998 tcp_free_md5sig_pool();
2999}
3000EXPORT_SYMBOL(tcp_put_md5sig_pool);
3001
3002int tcp_md5_hash_header(struct tcp_md5sig_pool *hp,
3003 struct tcphdr *th)
3004{
3005 struct scatterlist sg;
3006 int err;
3007
3008 __sum16 old_checksum = th->check;
3009 th->check = 0;
3010 /* options aren't included in the hash */
3011 sg_init_one(&sg, th, sizeof(struct tcphdr));
3012 err = crypto_hash_update(&hp->md5_desc, &sg, sizeof(struct tcphdr));
3013 th->check = old_checksum;
3014 return err;
3015}
3016EXPORT_SYMBOL(tcp_md5_hash_header);
3017
3018int tcp_md5_hash_skb_data(struct tcp_md5sig_pool *hp,
3019 struct sk_buff *skb, unsigned header_len)
3020{
3021 struct scatterlist sg;
3022 const struct tcphdr *tp = tcp_hdr(skb);
3023 struct hash_desc *desc = &hp->md5_desc;
3024 unsigned i;
3025 const unsigned head_data_len = skb_headlen(skb) > header_len ?
3026 skb_headlen(skb) - header_len : 0;
3027 const struct skb_shared_info *shi = skb_shinfo(skb);
3028 struct sk_buff *frag_iter;
3029
3030 sg_init_table(&sg, 1);
3031
3032 sg_set_buf(&sg, ((u8 *) tp) + header_len, head_data_len);
3033 if (crypto_hash_update(desc, &sg, head_data_len))
3034 return 1;
3035
3036 for (i = 0; i < shi->nr_frags; ++i) {
3037 const struct skb_frag_struct *f = &shi->frags[i];
3038 sg_set_page(&sg, f->page, f->size, f->page_offset);
3039 if (crypto_hash_update(desc, &sg, f->size))
3040 return 1;
3041 }
3042
3043 skb_walk_frags(skb, frag_iter)
3044 if (tcp_md5_hash_skb_data(hp, frag_iter, 0))
3045 return 1;
3046
3047 return 0;
3048}
3049EXPORT_SYMBOL(tcp_md5_hash_skb_data);
3050
3051int tcp_md5_hash_key(struct tcp_md5sig_pool *hp, struct tcp_md5sig_key *key)
3052{
3053 struct scatterlist sg;
3054
3055 sg_init_one(&sg, key->key, key->keylen);
3056 return crypto_hash_update(&hp->md5_desc, &sg, key->keylen);
3057}
3058EXPORT_SYMBOL(tcp_md5_hash_key);
3059
3060#endif
3061
3062/**
3063 * Each Responder maintains up to two secret values concurrently for
3064 * efficient secret rollover. Each secret value has 4 states:
3065 *
3066 * Generating. (tcp_secret_generating != tcp_secret_primary)
3067 * Generates new Responder-Cookies, but not yet used for primary
3068 * verification. This is a short-term state, typically lasting only
3069 * one round trip time (RTT).
3070 *
3071 * Primary. (tcp_secret_generating == tcp_secret_primary)
3072 * Used both for generation and primary verification.
3073 *
3074 * Retiring. (tcp_secret_retiring != tcp_secret_secondary)
3075 * Used for verification, until the first failure that can be
3076 * verified by the newer Generating secret. At that time, this
3077 * cookie's state is changed to Secondary, and the Generating
3078 * cookie's state is changed to Primary. This is a short-term state,
3079 * typically lasting only one round trip time (RTT).
3080 *
3081 * Secondary. (tcp_secret_retiring == tcp_secret_secondary)
3082 * Used for secondary verification, after primary verification
3083 * failures. This state lasts no more than twice the Maximum Segment
3084 * Lifetime (2MSL). Then, the secret is discarded.
3085 */
3086struct tcp_cookie_secret {
3087 /* The secret is divided into two parts. The digest part is the
3088 * equivalent of previously hashing a secret and saving the state,
3089 * and serves as an initialization vector (IV). The message part
3090 * serves as the trailing secret.
3091 */
3092 u32 secrets[COOKIE_WORKSPACE_WORDS];
3093 unsigned long expires;
3094};
3095
3096#define TCP_SECRET_1MSL (HZ * TCP_PAWS_MSL)
3097#define TCP_SECRET_2MSL (HZ * TCP_PAWS_MSL * 2)
3098#define TCP_SECRET_LIFE (HZ * 600)
3099
3100static struct tcp_cookie_secret tcp_secret_one;
3101static struct tcp_cookie_secret tcp_secret_two;
3102
3103/* Essentially a circular list, without dynamic allocation. */
3104static struct tcp_cookie_secret *tcp_secret_generating;
3105static struct tcp_cookie_secret *tcp_secret_primary;
3106static struct tcp_cookie_secret *tcp_secret_retiring;
3107static struct tcp_cookie_secret *tcp_secret_secondary;
3108
3109static DEFINE_SPINLOCK(tcp_secret_locker);
3110
3111/* Select a pseudo-random word in the cookie workspace.
3112 */
3113static inline u32 tcp_cookie_work(const u32 *ws, const int n)
3114{
3115 return ws[COOKIE_DIGEST_WORDS + ((COOKIE_MESSAGE_WORDS-1) & ws[n])];
3116}
3117
3118/* Fill bakery[COOKIE_WORKSPACE_WORDS] with generator, updating as needed.
3119 * Called in softirq context.
3120 * Returns: 0 for success.
3121 */
3122int tcp_cookie_generator(u32 *bakery)
3123{
3124 unsigned long jiffy = jiffies;
3125
3126 if (unlikely(time_after_eq(jiffy, tcp_secret_generating->expires))) {
3127 spin_lock_bh(&tcp_secret_locker);
3128 if (!time_after_eq(jiffy, tcp_secret_generating->expires)) {
3129 /* refreshed by another */
3130 memcpy(bakery,
3131 &tcp_secret_generating->secrets[0],
3132 COOKIE_WORKSPACE_WORDS);
3133 } else {
3134 /* still needs refreshing */
3135 get_random_bytes(bakery, COOKIE_WORKSPACE_WORDS);
3136
3137 /* The first time, paranoia assumes that the
3138 * randomization function isn't as strong. But,
3139 * this secret initialization is delayed until
3140 * the last possible moment (packet arrival).
3141 * Although that time is observable, it is
3142 * unpredictably variable. Mash in the most
3143 * volatile clock bits available, and expire the
3144 * secret extra quickly.
3145 */
3146 if (unlikely(tcp_secret_primary->expires ==
3147 tcp_secret_secondary->expires)) {
3148 struct timespec tv;
3149
3150 getnstimeofday(&tv);
3151 bakery[COOKIE_DIGEST_WORDS+0] ^=
3152 (u32)tv.tv_nsec;
3153
3154 tcp_secret_secondary->expires = jiffy
3155 + TCP_SECRET_1MSL
3156 + (0x0f & tcp_cookie_work(bakery, 0));
3157 } else {
3158 tcp_secret_secondary->expires = jiffy
3159 + TCP_SECRET_LIFE
3160 + (0xff & tcp_cookie_work(bakery, 1));
3161 tcp_secret_primary->expires = jiffy
3162 + TCP_SECRET_2MSL
3163 + (0x1f & tcp_cookie_work(bakery, 2));
3164 }
3165 memcpy(&tcp_secret_secondary->secrets[0],
3166 bakery, COOKIE_WORKSPACE_WORDS);
3167
3168 rcu_assign_pointer(tcp_secret_generating,
3169 tcp_secret_secondary);
3170 rcu_assign_pointer(tcp_secret_retiring,
3171 tcp_secret_primary);
3172 /*
3173 * Neither call_rcu() nor synchronize_rcu() needed.
3174 * Retiring data is not freed. It is replaced after
3175 * further (locked) pointer updates, and a quiet time
3176 * (minimum 1MSL, maximum LIFE - 2MSL).
3177 */
3178 }
3179 spin_unlock_bh(&tcp_secret_locker);
3180 } else {
3181 rcu_read_lock_bh();
3182 memcpy(bakery,
3183 &rcu_dereference(tcp_secret_generating)->secrets[0],
3184 COOKIE_WORKSPACE_WORDS);
3185 rcu_read_unlock_bh();
3186 }
3187 return 0;
3188}
3189EXPORT_SYMBOL(tcp_cookie_generator);
3190
3191void tcp_done(struct sock *sk)
3192{
3193 if (sk->sk_state == TCP_SYN_SENT || sk->sk_state == TCP_SYN_RECV)
3194 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_ATTEMPTFAILS);
3195
3196 tcp_set_state(sk, TCP_CLOSE);
3197 tcp_clear_xmit_timers(sk);
3198
3199 sk->sk_shutdown = SHUTDOWN_MASK;
3200
3201 if (!sock_flag(sk, SOCK_DEAD))
3202 sk->sk_state_change(sk);
3203 else
3204 inet_csk_destroy_sock(sk);
3205}
3206EXPORT_SYMBOL_GPL(tcp_done);
3207
3208extern struct tcp_congestion_ops tcp_reno;
3209
3210static __initdata unsigned long thash_entries;
3211static int __init set_thash_entries(char *str)
3212{
3213 if (!str)
3214 return 0;
3215 thash_entries = simple_strtoul(str, &str, 0);
3216 return 1;
3217}
3218__setup("thash_entries=", set_thash_entries);
3219
3220void __init tcp_init(void)
3221{
3222 struct sk_buff *skb = NULL;
3223 unsigned long limit;
3224 int i, max_share, cnt;
3225 unsigned long jiffy = jiffies;
3226
3227 BUILD_BUG_ON(sizeof(struct tcp_skb_cb) > sizeof(skb->cb));
3228
3229 percpu_counter_init(&tcp_sockets_allocated, 0);
3230 percpu_counter_init(&tcp_orphan_count, 0);
3231 tcp_hashinfo.bind_bucket_cachep =
3232 kmem_cache_create("tcp_bind_bucket",
3233 sizeof(struct inet_bind_bucket), 0,
3234 SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL);
3235
3236 /* Size and allocate the main established and bind bucket
3237 * hash tables.
3238 *
3239 * The methodology is similar to that of the buffer cache.
3240 */
3241 tcp_hashinfo.ehash =
3242 alloc_large_system_hash("TCP established",
3243 sizeof(struct inet_ehash_bucket),
3244 thash_entries,
3245 (totalram_pages >= 128 * 1024) ?
3246 13 : 15,
3247 0,
3248 NULL,
3249 &tcp_hashinfo.ehash_mask,
3250 thash_entries ? 0 : 512 * 1024);
3251 for (i = 0; i <= tcp_hashinfo.ehash_mask; i++) {
3252 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].chain, i);
3253 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].twchain, i);
3254 }
3255 if (inet_ehash_locks_alloc(&tcp_hashinfo))
3256 panic("TCP: failed to alloc ehash_locks");
3257 tcp_hashinfo.bhash =
3258 alloc_large_system_hash("TCP bind",
3259 sizeof(struct inet_bind_hashbucket),
3260 tcp_hashinfo.ehash_mask + 1,
3261 (totalram_pages >= 128 * 1024) ?
3262 13 : 15,
3263 0,
3264 &tcp_hashinfo.bhash_size,
3265 NULL,
3266 64 * 1024);
3267 tcp_hashinfo.bhash_size = 1 << tcp_hashinfo.bhash_size;
3268 for (i = 0; i < tcp_hashinfo.bhash_size; i++) {
3269 spin_lock_init(&tcp_hashinfo.bhash[i].lock);
3270 INIT_HLIST_HEAD(&tcp_hashinfo.bhash[i].chain);
3271 }
3272
3273
3274 cnt = tcp_hashinfo.ehash_mask + 1;
3275
3276 tcp_death_row.sysctl_max_tw_buckets = cnt / 2;
3277 sysctl_tcp_max_orphans = cnt / 2;
3278 sysctl_max_syn_backlog = max(128, cnt / 256);
3279
3280 limit = nr_free_buffer_pages() / 8;
3281 limit = max(limit, 128UL);
3282 sysctl_tcp_mem[0] = limit / 4 * 3;
3283 sysctl_tcp_mem[1] = limit;
3284 sysctl_tcp_mem[2] = sysctl_tcp_mem[0] * 2;
3285
3286 /* Set per-socket limits to no more than 1/128 the pressure threshold */
3287 limit = ((unsigned long)sysctl_tcp_mem[1]) << (PAGE_SHIFT - 7);
3288 max_share = min(4UL*1024*1024, limit);
3289
3290 sysctl_tcp_wmem[0] = SK_MEM_QUANTUM;
3291 sysctl_tcp_wmem[1] = 16*1024;
3292 sysctl_tcp_wmem[2] = max(64*1024, max_share);
3293
3294 sysctl_tcp_rmem[0] = SK_MEM_QUANTUM;
3295 sysctl_tcp_rmem[1] = 87380;
3296 sysctl_tcp_rmem[2] = max(87380, max_share);
3297
3298 printk(KERN_INFO "TCP: Hash tables configured "
3299 "(established %u bind %u)\n",
3300 tcp_hashinfo.ehash_mask + 1, tcp_hashinfo.bhash_size);
3301
3302 tcp_register_congestion_control(&tcp_reno);
3303
3304 memset(&tcp_secret_one.secrets[0], 0, sizeof(tcp_secret_one.secrets));
3305 memset(&tcp_secret_two.secrets[0], 0, sizeof(tcp_secret_two.secrets));
3306 tcp_secret_one.expires = jiffy; /* past due */
3307 tcp_secret_two.expires = jiffy; /* past due */
3308 tcp_secret_generating = &tcp_secret_one;
3309 tcp_secret_primary = &tcp_secret_one;
3310 tcp_secret_retiring = &tcp_secret_two;
3311 tcp_secret_secondary = &tcp_secret_two;
3312}
1/*
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.
5 *
6 * Implementation of the Transmission Control Protocol(TCP).
7 *
8 * Authors: Ross Biro
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>
19 *
20 * Fixes:
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
25 * (tcp_err()).
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
36 * unknown sockets.
37 * Alan Cox : tcp option processing.
38 * Alan Cox : Reset tweaked (still not 100%) [Had
39 * syn rule wrong]
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
45 * escape still
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
49 * facilities
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
54 * bit to skb ops.
55 * Alan Cox : Tidied tcp_data to avoid a potential
56 * nasty.
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
68 * sockets.
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
72 * state ack error.
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
77 * fixes
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
83 * completely
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
91 * (not yet usable)
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
104 * all cases.
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
109 * works now.
110 * Michael Pall : recv(...,MSG_OOB) never blocks in the
111 * BSD api.
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
119 * fixed ports.
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
125 * socket close.
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
130 * accept.
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
141 * close.
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
147 * comments.
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
155 * resemble the RFC.
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
160 * generates them.
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
173 * but it's a start!
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
194 * improvement.
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.
207 *
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.
212 *
213 * Description of States:
214 *
215 * TCP_SYN_SENT sent a connection request, waiting for ack
216 *
217 * TCP_SYN_RECV received a connection request, sent ack,
218 * waiting for final ack in three-way handshake.
219 *
220 * TCP_ESTABLISHED connection established
221 *
222 * TCP_FIN_WAIT1 our side has shutdown, waiting to complete
223 * transmission of remaining buffered data
224 *
225 * TCP_FIN_WAIT2 all buffered data sent, waiting for remote
226 * to shutdown
227 *
228 * TCP_CLOSING both sides have shutdown but we still have
229 * data we have to finish sending
230 *
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)
236 *
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)
240 *
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
244 *
245 * TCP_CLOSE socket is finished
246 */
247
248#define pr_fmt(fmt) "TCP: " fmt
249
250#include <linux/kernel.h>
251#include <linux/module.h>
252#include <linux/types.h>
253#include <linux/fcntl.h>
254#include <linux/poll.h>
255#include <linux/init.h>
256#include <linux/fs.h>
257#include <linux/skbuff.h>
258#include <linux/scatterlist.h>
259#include <linux/splice.h>
260#include <linux/net.h>
261#include <linux/socket.h>
262#include <linux/random.h>
263#include <linux/bootmem.h>
264#include <linux/highmem.h>
265#include <linux/swap.h>
266#include <linux/cache.h>
267#include <linux/err.h>
268#include <linux/crypto.h>
269#include <linux/time.h>
270#include <linux/slab.h>
271
272#include <net/icmp.h>
273#include <net/inet_common.h>
274#include <net/tcp.h>
275#include <net/xfrm.h>
276#include <net/ip.h>
277#include <net/netdma.h>
278#include <net/sock.h>
279
280#include <asm/uaccess.h>
281#include <asm/ioctls.h>
282#include <net/busy_poll.h>
283
284int sysctl_tcp_fin_timeout __read_mostly = TCP_FIN_TIMEOUT;
285
286int sysctl_tcp_min_tso_segs __read_mostly = 2;
287
288int sysctl_tcp_autocorking __read_mostly = 1;
289
290struct percpu_counter tcp_orphan_count;
291EXPORT_SYMBOL_GPL(tcp_orphan_count);
292
293long sysctl_tcp_mem[3] __read_mostly;
294int sysctl_tcp_wmem[3] __read_mostly;
295int sysctl_tcp_rmem[3] __read_mostly;
296
297EXPORT_SYMBOL(sysctl_tcp_mem);
298EXPORT_SYMBOL(sysctl_tcp_rmem);
299EXPORT_SYMBOL(sysctl_tcp_wmem);
300
301atomic_long_t tcp_memory_allocated; /* Current allocated memory. */
302EXPORT_SYMBOL(tcp_memory_allocated);
303
304/*
305 * Current number of TCP sockets.
306 */
307struct percpu_counter tcp_sockets_allocated;
308EXPORT_SYMBOL(tcp_sockets_allocated);
309
310/*
311 * TCP splice context
312 */
313struct tcp_splice_state {
314 struct pipe_inode_info *pipe;
315 size_t len;
316 unsigned int flags;
317};
318
319/*
320 * Pressure flag: try to collapse.
321 * Technical note: it is used by multiple contexts non atomically.
322 * All the __sk_mem_schedule() is of this nature: accounting
323 * is strict, actions are advisory and have some latency.
324 */
325int tcp_memory_pressure __read_mostly;
326EXPORT_SYMBOL(tcp_memory_pressure);
327
328void tcp_enter_memory_pressure(struct sock *sk)
329{
330 if (!tcp_memory_pressure) {
331 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMEMORYPRESSURES);
332 tcp_memory_pressure = 1;
333 }
334}
335EXPORT_SYMBOL(tcp_enter_memory_pressure);
336
337/* Convert seconds to retransmits based on initial and max timeout */
338static u8 secs_to_retrans(int seconds, int timeout, int rto_max)
339{
340 u8 res = 0;
341
342 if (seconds > 0) {
343 int period = timeout;
344
345 res = 1;
346 while (seconds > period && res < 255) {
347 res++;
348 timeout <<= 1;
349 if (timeout > rto_max)
350 timeout = rto_max;
351 period += timeout;
352 }
353 }
354 return res;
355}
356
357/* Convert retransmits to seconds based on initial and max timeout */
358static int retrans_to_secs(u8 retrans, int timeout, int rto_max)
359{
360 int period = 0;
361
362 if (retrans > 0) {
363 period = timeout;
364 while (--retrans) {
365 timeout <<= 1;
366 if (timeout > rto_max)
367 timeout = rto_max;
368 period += timeout;
369 }
370 }
371 return period;
372}
373
374/* Address-family independent initialization for a tcp_sock.
375 *
376 * NOTE: A lot of things set to zero explicitly by call to
377 * sk_alloc() so need not be done here.
378 */
379void tcp_init_sock(struct sock *sk)
380{
381 struct inet_connection_sock *icsk = inet_csk(sk);
382 struct tcp_sock *tp = tcp_sk(sk);
383
384 __skb_queue_head_init(&tp->out_of_order_queue);
385 tcp_init_xmit_timers(sk);
386 tcp_prequeue_init(tp);
387 INIT_LIST_HEAD(&tp->tsq_node);
388
389 icsk->icsk_rto = TCP_TIMEOUT_INIT;
390 tp->mdev_us = jiffies_to_usecs(TCP_TIMEOUT_INIT);
391
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
396 */
397 tp->snd_cwnd = TCP_INIT_CWND;
398
399 /* See draft-stevens-tcpca-spec-01 for discussion of the
400 * initialization of these values.
401 */
402 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
403 tp->snd_cwnd_clamp = ~0;
404 tp->mss_cache = TCP_MSS_DEFAULT;
405
406 tp->reordering = sysctl_tcp_reordering;
407 tcp_enable_early_retrans(tp);
408 icsk->icsk_ca_ops = &tcp_init_congestion_ops;
409
410 tp->tsoffset = 0;
411
412 sk->sk_state = TCP_CLOSE;
413
414 sk->sk_write_space = sk_stream_write_space;
415 sock_set_flag(sk, SOCK_USE_WRITE_QUEUE);
416
417 icsk->icsk_sync_mss = tcp_sync_mss;
418
419 sk->sk_sndbuf = sysctl_tcp_wmem[1];
420 sk->sk_rcvbuf = sysctl_tcp_rmem[1];
421
422 local_bh_disable();
423 sock_update_memcg(sk);
424 sk_sockets_allocated_inc(sk);
425 local_bh_enable();
426}
427EXPORT_SYMBOL(tcp_init_sock);
428
429/*
430 * Wait for a TCP event.
431 *
432 * Note that we don't need to lock the socket, as the upper poll layers
433 * take care of normal races (between the test and the event) and we don't
434 * go look at any of the socket buffers directly.
435 */
436unsigned int tcp_poll(struct file *file, struct socket *sock, poll_table *wait)
437{
438 unsigned int mask;
439 struct sock *sk = sock->sk;
440 const struct tcp_sock *tp = tcp_sk(sk);
441
442 sock_rps_record_flow(sk);
443
444 sock_poll_wait(file, sk_sleep(sk), wait);
445 if (sk->sk_state == TCP_LISTEN)
446 return inet_csk_listen_poll(sk);
447
448 /* Socket is not locked. We are protected from async events
449 * by poll logic and correct handling of state changes
450 * made by other threads is impossible in any case.
451 */
452
453 mask = 0;
454
455 /*
456 * POLLHUP is certainly not done right. But poll() doesn't
457 * have a notion of HUP in just one direction, and for a
458 * socket the read side is more interesting.
459 *
460 * Some poll() documentation says that POLLHUP is incompatible
461 * with the POLLOUT/POLLWR flags, so somebody should check this
462 * all. But careful, it tends to be safer to return too many
463 * bits than too few, and you can easily break real applications
464 * if you don't tell them that something has hung up!
465 *
466 * Check-me.
467 *
468 * Check number 1. POLLHUP is _UNMASKABLE_ event (see UNIX98 and
469 * our fs/select.c). It means that after we received EOF,
470 * poll always returns immediately, making impossible poll() on write()
471 * in state CLOSE_WAIT. One solution is evident --- to set POLLHUP
472 * if and only if shutdown has been made in both directions.
473 * Actually, it is interesting to look how Solaris and DUX
474 * solve this dilemma. I would prefer, if POLLHUP were maskable,
475 * then we could set it on SND_SHUTDOWN. BTW examples given
476 * in Stevens' books assume exactly this behaviour, it explains
477 * why POLLHUP is incompatible with POLLOUT. --ANK
478 *
479 * NOTE. Check for TCP_CLOSE is added. The goal is to prevent
480 * blocking on fresh not-connected or disconnected socket. --ANK
481 */
482 if (sk->sk_shutdown == SHUTDOWN_MASK || sk->sk_state == TCP_CLOSE)
483 mask |= POLLHUP;
484 if (sk->sk_shutdown & RCV_SHUTDOWN)
485 mask |= POLLIN | POLLRDNORM | POLLRDHUP;
486
487 /* Connected or passive Fast Open socket? */
488 if (sk->sk_state != TCP_SYN_SENT &&
489 (sk->sk_state != TCP_SYN_RECV || tp->fastopen_rsk != NULL)) {
490 int target = sock_rcvlowat(sk, 0, INT_MAX);
491
492 if (tp->urg_seq == tp->copied_seq &&
493 !sock_flag(sk, SOCK_URGINLINE) &&
494 tp->urg_data)
495 target++;
496
497 /* Potential race condition. If read of tp below will
498 * escape above sk->sk_state, we can be illegally awaken
499 * in SYN_* states. */
500 if (tp->rcv_nxt - tp->copied_seq >= target)
501 mask |= POLLIN | POLLRDNORM;
502
503 if (!(sk->sk_shutdown & SEND_SHUTDOWN)) {
504 if (sk_stream_is_writeable(sk)) {
505 mask |= POLLOUT | POLLWRNORM;
506 } else { /* send SIGIO later */
507 set_bit(SOCK_ASYNC_NOSPACE,
508 &sk->sk_socket->flags);
509 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
510
511 /* Race breaker. If space is freed after
512 * wspace test but before the flags are set,
513 * IO signal will be lost.
514 */
515 if (sk_stream_is_writeable(sk))
516 mask |= POLLOUT | POLLWRNORM;
517 }
518 } else
519 mask |= POLLOUT | POLLWRNORM;
520
521 if (tp->urg_data & TCP_URG_VALID)
522 mask |= POLLPRI;
523 }
524 /* This barrier is coupled with smp_wmb() in tcp_reset() */
525 smp_rmb();
526 if (sk->sk_err)
527 mask |= POLLERR;
528
529 return mask;
530}
531EXPORT_SYMBOL(tcp_poll);
532
533int tcp_ioctl(struct sock *sk, int cmd, unsigned long arg)
534{
535 struct tcp_sock *tp = tcp_sk(sk);
536 int answ;
537 bool slow;
538
539 switch (cmd) {
540 case SIOCINQ:
541 if (sk->sk_state == TCP_LISTEN)
542 return -EINVAL;
543
544 slow = lock_sock_fast(sk);
545 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
546 answ = 0;
547 else if (sock_flag(sk, SOCK_URGINLINE) ||
548 !tp->urg_data ||
549 before(tp->urg_seq, tp->copied_seq) ||
550 !before(tp->urg_seq, tp->rcv_nxt)) {
551
552 answ = tp->rcv_nxt - tp->copied_seq;
553
554 /* Subtract 1, if FIN was received */
555 if (answ && sock_flag(sk, SOCK_DONE))
556 answ--;
557 } else
558 answ = tp->urg_seq - tp->copied_seq;
559 unlock_sock_fast(sk, slow);
560 break;
561 case SIOCATMARK:
562 answ = tp->urg_data && tp->urg_seq == tp->copied_seq;
563 break;
564 case SIOCOUTQ:
565 if (sk->sk_state == TCP_LISTEN)
566 return -EINVAL;
567
568 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
569 answ = 0;
570 else
571 answ = tp->write_seq - tp->snd_una;
572 break;
573 case SIOCOUTQNSD:
574 if (sk->sk_state == TCP_LISTEN)
575 return -EINVAL;
576
577 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
578 answ = 0;
579 else
580 answ = tp->write_seq - tp->snd_nxt;
581 break;
582 default:
583 return -ENOIOCTLCMD;
584 }
585
586 return put_user(answ, (int __user *)arg);
587}
588EXPORT_SYMBOL(tcp_ioctl);
589
590static inline void tcp_mark_push(struct tcp_sock *tp, struct sk_buff *skb)
591{
592 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
593 tp->pushed_seq = tp->write_seq;
594}
595
596static inline bool forced_push(const struct tcp_sock *tp)
597{
598 return after(tp->write_seq, tp->pushed_seq + (tp->max_window >> 1));
599}
600
601static inline void skb_entail(struct sock *sk, struct sk_buff *skb)
602{
603 struct tcp_sock *tp = tcp_sk(sk);
604 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
605
606 skb->csum = 0;
607 tcb->seq = tcb->end_seq = tp->write_seq;
608 tcb->tcp_flags = TCPHDR_ACK;
609 tcb->sacked = 0;
610 skb_header_release(skb);
611 tcp_add_write_queue_tail(sk, skb);
612 sk->sk_wmem_queued += skb->truesize;
613 sk_mem_charge(sk, skb->truesize);
614 if (tp->nonagle & TCP_NAGLE_PUSH)
615 tp->nonagle &= ~TCP_NAGLE_PUSH;
616}
617
618static inline void tcp_mark_urg(struct tcp_sock *tp, int flags)
619{
620 if (flags & MSG_OOB)
621 tp->snd_up = tp->write_seq;
622}
623
624/* If a not yet filled skb is pushed, do not send it if
625 * we have data packets in Qdisc or NIC queues :
626 * Because TX completion will happen shortly, it gives a chance
627 * to coalesce future sendmsg() payload into this skb, without
628 * need for a timer, and with no latency trade off.
629 * As packets containing data payload have a bigger truesize
630 * than pure acks (dataless) packets, the last checks prevent
631 * autocorking if we only have an ACK in Qdisc/NIC queues,
632 * or if TX completion was delayed after we processed ACK packet.
633 */
634static bool tcp_should_autocork(struct sock *sk, struct sk_buff *skb,
635 int size_goal)
636{
637 return skb->len < size_goal &&
638 sysctl_tcp_autocorking &&
639 skb != tcp_write_queue_head(sk) &&
640 atomic_read(&sk->sk_wmem_alloc) > skb->truesize;
641}
642
643static void tcp_push(struct sock *sk, int flags, int mss_now,
644 int nonagle, int size_goal)
645{
646 struct tcp_sock *tp = tcp_sk(sk);
647 struct sk_buff *skb;
648
649 if (!tcp_send_head(sk))
650 return;
651
652 skb = tcp_write_queue_tail(sk);
653 if (!(flags & MSG_MORE) || forced_push(tp))
654 tcp_mark_push(tp, skb);
655
656 tcp_mark_urg(tp, flags);
657
658 if (tcp_should_autocork(sk, skb, size_goal)) {
659
660 /* avoid atomic op if TSQ_THROTTLED bit is already set */
661 if (!test_bit(TSQ_THROTTLED, &tp->tsq_flags)) {
662 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPAUTOCORKING);
663 set_bit(TSQ_THROTTLED, &tp->tsq_flags);
664 }
665 /* It is possible TX completion already happened
666 * before we set TSQ_THROTTLED.
667 */
668 if (atomic_read(&sk->sk_wmem_alloc) > skb->truesize)
669 return;
670 }
671
672 if (flags & MSG_MORE)
673 nonagle = TCP_NAGLE_CORK;
674
675 __tcp_push_pending_frames(sk, mss_now, nonagle);
676}
677
678static int tcp_splice_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb,
679 unsigned int offset, size_t len)
680{
681 struct tcp_splice_state *tss = rd_desc->arg.data;
682 int ret;
683
684 ret = skb_splice_bits(skb, offset, tss->pipe, min(rd_desc->count, len),
685 tss->flags);
686 if (ret > 0)
687 rd_desc->count -= ret;
688 return ret;
689}
690
691static int __tcp_splice_read(struct sock *sk, struct tcp_splice_state *tss)
692{
693 /* Store TCP splice context information in read_descriptor_t. */
694 read_descriptor_t rd_desc = {
695 .arg.data = tss,
696 .count = tss->len,
697 };
698
699 return tcp_read_sock(sk, &rd_desc, tcp_splice_data_recv);
700}
701
702/**
703 * tcp_splice_read - splice data from TCP socket to a pipe
704 * @sock: socket to splice from
705 * @ppos: position (not valid)
706 * @pipe: pipe to splice to
707 * @len: number of bytes to splice
708 * @flags: splice modifier flags
709 *
710 * Description:
711 * Will read pages from given socket and fill them into a pipe.
712 *
713 **/
714ssize_t tcp_splice_read(struct socket *sock, loff_t *ppos,
715 struct pipe_inode_info *pipe, size_t len,
716 unsigned int flags)
717{
718 struct sock *sk = sock->sk;
719 struct tcp_splice_state tss = {
720 .pipe = pipe,
721 .len = len,
722 .flags = flags,
723 };
724 long timeo;
725 ssize_t spliced;
726 int ret;
727
728 sock_rps_record_flow(sk);
729 /*
730 * We can't seek on a socket input
731 */
732 if (unlikely(*ppos))
733 return -ESPIPE;
734
735 ret = spliced = 0;
736
737 lock_sock(sk);
738
739 timeo = sock_rcvtimeo(sk, sock->file->f_flags & O_NONBLOCK);
740 while (tss.len) {
741 ret = __tcp_splice_read(sk, &tss);
742 if (ret < 0)
743 break;
744 else if (!ret) {
745 if (spliced)
746 break;
747 if (sock_flag(sk, SOCK_DONE))
748 break;
749 if (sk->sk_err) {
750 ret = sock_error(sk);
751 break;
752 }
753 if (sk->sk_shutdown & RCV_SHUTDOWN)
754 break;
755 if (sk->sk_state == TCP_CLOSE) {
756 /*
757 * This occurs when user tries to read
758 * from never connected socket.
759 */
760 if (!sock_flag(sk, SOCK_DONE))
761 ret = -ENOTCONN;
762 break;
763 }
764 if (!timeo) {
765 ret = -EAGAIN;
766 break;
767 }
768 sk_wait_data(sk, &timeo);
769 if (signal_pending(current)) {
770 ret = sock_intr_errno(timeo);
771 break;
772 }
773 continue;
774 }
775 tss.len -= ret;
776 spliced += ret;
777
778 if (!timeo)
779 break;
780 release_sock(sk);
781 lock_sock(sk);
782
783 if (sk->sk_err || sk->sk_state == TCP_CLOSE ||
784 (sk->sk_shutdown & RCV_SHUTDOWN) ||
785 signal_pending(current))
786 break;
787 }
788
789 release_sock(sk);
790
791 if (spliced)
792 return spliced;
793
794 return ret;
795}
796EXPORT_SYMBOL(tcp_splice_read);
797
798struct sk_buff *sk_stream_alloc_skb(struct sock *sk, int size, gfp_t gfp)
799{
800 struct sk_buff *skb;
801
802 /* The TCP header must be at least 32-bit aligned. */
803 size = ALIGN(size, 4);
804
805 skb = alloc_skb_fclone(size + sk->sk_prot->max_header, gfp);
806 if (skb) {
807 if (sk_wmem_schedule(sk, skb->truesize)) {
808 skb_reserve(skb, sk->sk_prot->max_header);
809 /*
810 * Make sure that we have exactly size bytes
811 * available to the caller, no more, no less.
812 */
813 skb->reserved_tailroom = skb->end - skb->tail - size;
814 return skb;
815 }
816 __kfree_skb(skb);
817 } else {
818 sk->sk_prot->enter_memory_pressure(sk);
819 sk_stream_moderate_sndbuf(sk);
820 }
821 return NULL;
822}
823
824static unsigned int tcp_xmit_size_goal(struct sock *sk, u32 mss_now,
825 int large_allowed)
826{
827 struct tcp_sock *tp = tcp_sk(sk);
828 u32 xmit_size_goal, old_size_goal;
829
830 xmit_size_goal = mss_now;
831
832 if (large_allowed && sk_can_gso(sk)) {
833 u32 gso_size, hlen;
834
835 /* Maybe we should/could use sk->sk_prot->max_header here ? */
836 hlen = inet_csk(sk)->icsk_af_ops->net_header_len +
837 inet_csk(sk)->icsk_ext_hdr_len +
838 tp->tcp_header_len;
839
840 /* Goal is to send at least one packet per ms,
841 * not one big TSO packet every 100 ms.
842 * This preserves ACK clocking and is consistent
843 * with tcp_tso_should_defer() heuristic.
844 */
845 gso_size = sk->sk_pacing_rate / (2 * MSEC_PER_SEC);
846 gso_size = max_t(u32, gso_size,
847 sysctl_tcp_min_tso_segs * mss_now);
848
849 xmit_size_goal = min_t(u32, gso_size,
850 sk->sk_gso_max_size - 1 - hlen);
851
852 xmit_size_goal = tcp_bound_to_half_wnd(tp, xmit_size_goal);
853
854 /* We try hard to avoid divides here */
855 old_size_goal = tp->xmit_size_goal_segs * mss_now;
856
857 if (likely(old_size_goal <= xmit_size_goal &&
858 old_size_goal + mss_now > xmit_size_goal)) {
859 xmit_size_goal = old_size_goal;
860 } else {
861 tp->xmit_size_goal_segs =
862 min_t(u16, xmit_size_goal / mss_now,
863 sk->sk_gso_max_segs);
864 xmit_size_goal = tp->xmit_size_goal_segs * mss_now;
865 }
866 }
867
868 return max(xmit_size_goal, mss_now);
869}
870
871static int tcp_send_mss(struct sock *sk, int *size_goal, int flags)
872{
873 int mss_now;
874
875 mss_now = tcp_current_mss(sk);
876 *size_goal = tcp_xmit_size_goal(sk, mss_now, !(flags & MSG_OOB));
877
878 return mss_now;
879}
880
881static ssize_t do_tcp_sendpages(struct sock *sk, struct page *page, int offset,
882 size_t size, int flags)
883{
884 struct tcp_sock *tp = tcp_sk(sk);
885 int mss_now, size_goal;
886 int err;
887 ssize_t copied;
888 long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
889
890 /* Wait for a connection to finish. One exception is TCP Fast Open
891 * (passive side) where data is allowed to be sent before a connection
892 * is fully established.
893 */
894 if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) &&
895 !tcp_passive_fastopen(sk)) {
896 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
897 goto out_err;
898 }
899
900 clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
901
902 mss_now = tcp_send_mss(sk, &size_goal, flags);
903 copied = 0;
904
905 err = -EPIPE;
906 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
907 goto out_err;
908
909 while (size > 0) {
910 struct sk_buff *skb = tcp_write_queue_tail(sk);
911 int copy, i;
912 bool can_coalesce;
913
914 if (!tcp_send_head(sk) || (copy = size_goal - skb->len) <= 0) {
915new_segment:
916 if (!sk_stream_memory_free(sk))
917 goto wait_for_sndbuf;
918
919 skb = sk_stream_alloc_skb(sk, 0, sk->sk_allocation);
920 if (!skb)
921 goto wait_for_memory;
922
923 skb_entail(sk, skb);
924 copy = size_goal;
925 }
926
927 if (copy > size)
928 copy = size;
929
930 i = skb_shinfo(skb)->nr_frags;
931 can_coalesce = skb_can_coalesce(skb, i, page, offset);
932 if (!can_coalesce && i >= MAX_SKB_FRAGS) {
933 tcp_mark_push(tp, skb);
934 goto new_segment;
935 }
936 if (!sk_wmem_schedule(sk, copy))
937 goto wait_for_memory;
938
939 if (can_coalesce) {
940 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
941 } else {
942 get_page(page);
943 skb_fill_page_desc(skb, i, page, offset, copy);
944 }
945 skb_shinfo(skb)->tx_flags |= SKBTX_SHARED_FRAG;
946
947 skb->len += copy;
948 skb->data_len += copy;
949 skb->truesize += copy;
950 sk->sk_wmem_queued += copy;
951 sk_mem_charge(sk, copy);
952 skb->ip_summed = CHECKSUM_PARTIAL;
953 tp->write_seq += copy;
954 TCP_SKB_CB(skb)->end_seq += copy;
955 skb_shinfo(skb)->gso_segs = 0;
956
957 if (!copied)
958 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
959
960 copied += copy;
961 offset += copy;
962 if (!(size -= copy))
963 goto out;
964
965 if (skb->len < size_goal || (flags & MSG_OOB))
966 continue;
967
968 if (forced_push(tp)) {
969 tcp_mark_push(tp, skb);
970 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
971 } else if (skb == tcp_send_head(sk))
972 tcp_push_one(sk, mss_now);
973 continue;
974
975wait_for_sndbuf:
976 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
977wait_for_memory:
978 tcp_push(sk, flags & ~MSG_MORE, mss_now,
979 TCP_NAGLE_PUSH, size_goal);
980
981 if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
982 goto do_error;
983
984 mss_now = tcp_send_mss(sk, &size_goal, flags);
985 }
986
987out:
988 if (copied && !(flags & MSG_SENDPAGE_NOTLAST))
989 tcp_push(sk, flags, mss_now, tp->nonagle, size_goal);
990 return copied;
991
992do_error:
993 if (copied)
994 goto out;
995out_err:
996 return sk_stream_error(sk, flags, err);
997}
998
999int tcp_sendpage(struct sock *sk, struct page *page, int offset,
1000 size_t size, int flags)
1001{
1002 ssize_t res;
1003
1004 if (!(sk->sk_route_caps & NETIF_F_SG) ||
1005 !(sk->sk_route_caps & NETIF_F_ALL_CSUM))
1006 return sock_no_sendpage(sk->sk_socket, page, offset, size,
1007 flags);
1008
1009 lock_sock(sk);
1010 res = do_tcp_sendpages(sk, page, offset, size, flags);
1011 release_sock(sk);
1012 return res;
1013}
1014EXPORT_SYMBOL(tcp_sendpage);
1015
1016static inline int select_size(const struct sock *sk, bool sg)
1017{
1018 const struct tcp_sock *tp = tcp_sk(sk);
1019 int tmp = tp->mss_cache;
1020
1021 if (sg) {
1022 if (sk_can_gso(sk)) {
1023 /* Small frames wont use a full page:
1024 * Payload will immediately follow tcp header.
1025 */
1026 tmp = SKB_WITH_OVERHEAD(2048 - MAX_TCP_HEADER);
1027 } else {
1028 int pgbreak = SKB_MAX_HEAD(MAX_TCP_HEADER);
1029
1030 if (tmp >= pgbreak &&
1031 tmp <= pgbreak + (MAX_SKB_FRAGS - 1) * PAGE_SIZE)
1032 tmp = pgbreak;
1033 }
1034 }
1035
1036 return tmp;
1037}
1038
1039void tcp_free_fastopen_req(struct tcp_sock *tp)
1040{
1041 if (tp->fastopen_req != NULL) {
1042 kfree(tp->fastopen_req);
1043 tp->fastopen_req = NULL;
1044 }
1045}
1046
1047static int tcp_sendmsg_fastopen(struct sock *sk, struct msghdr *msg,
1048 int *copied, size_t size)
1049{
1050 struct tcp_sock *tp = tcp_sk(sk);
1051 int err, flags;
1052
1053 if (!(sysctl_tcp_fastopen & TFO_CLIENT_ENABLE))
1054 return -EOPNOTSUPP;
1055 if (tp->fastopen_req != NULL)
1056 return -EALREADY; /* Another Fast Open is in progress */
1057
1058 tp->fastopen_req = kzalloc(sizeof(struct tcp_fastopen_request),
1059 sk->sk_allocation);
1060 if (unlikely(tp->fastopen_req == NULL))
1061 return -ENOBUFS;
1062 tp->fastopen_req->data = msg;
1063 tp->fastopen_req->size = size;
1064
1065 flags = (msg->msg_flags & MSG_DONTWAIT) ? O_NONBLOCK : 0;
1066 err = __inet_stream_connect(sk->sk_socket, msg->msg_name,
1067 msg->msg_namelen, flags);
1068 *copied = tp->fastopen_req->copied;
1069 tcp_free_fastopen_req(tp);
1070 return err;
1071}
1072
1073int tcp_sendmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg,
1074 size_t size)
1075{
1076 struct iovec *iov;
1077 struct tcp_sock *tp = tcp_sk(sk);
1078 struct sk_buff *skb;
1079 int iovlen, flags, err, copied = 0;
1080 int mss_now = 0, size_goal, copied_syn = 0, offset = 0;
1081 bool sg;
1082 long timeo;
1083
1084 lock_sock(sk);
1085
1086 flags = msg->msg_flags;
1087 if (flags & MSG_FASTOPEN) {
1088 err = tcp_sendmsg_fastopen(sk, msg, &copied_syn, size);
1089 if (err == -EINPROGRESS && copied_syn > 0)
1090 goto out;
1091 else if (err)
1092 goto out_err;
1093 offset = copied_syn;
1094 }
1095
1096 timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
1097
1098 /* Wait for a connection to finish. One exception is TCP Fast Open
1099 * (passive side) where data is allowed to be sent before a connection
1100 * is fully established.
1101 */
1102 if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) &&
1103 !tcp_passive_fastopen(sk)) {
1104 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
1105 goto do_error;
1106 }
1107
1108 if (unlikely(tp->repair)) {
1109 if (tp->repair_queue == TCP_RECV_QUEUE) {
1110 copied = tcp_send_rcvq(sk, msg, size);
1111 goto out;
1112 }
1113
1114 err = -EINVAL;
1115 if (tp->repair_queue == TCP_NO_QUEUE)
1116 goto out_err;
1117
1118 /* 'common' sending to sendq */
1119 }
1120
1121 /* This should be in poll */
1122 clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
1123
1124 mss_now = tcp_send_mss(sk, &size_goal, flags);
1125
1126 /* Ok commence sending. */
1127 iovlen = msg->msg_iovlen;
1128 iov = msg->msg_iov;
1129 copied = 0;
1130
1131 err = -EPIPE;
1132 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
1133 goto out_err;
1134
1135 sg = !!(sk->sk_route_caps & NETIF_F_SG);
1136
1137 while (--iovlen >= 0) {
1138 size_t seglen = iov->iov_len;
1139 unsigned char __user *from = iov->iov_base;
1140
1141 iov++;
1142 if (unlikely(offset > 0)) { /* Skip bytes copied in SYN */
1143 if (offset >= seglen) {
1144 offset -= seglen;
1145 continue;
1146 }
1147 seglen -= offset;
1148 from += offset;
1149 offset = 0;
1150 }
1151
1152 while (seglen > 0) {
1153 int copy = 0;
1154 int max = size_goal;
1155
1156 skb = tcp_write_queue_tail(sk);
1157 if (tcp_send_head(sk)) {
1158 if (skb->ip_summed == CHECKSUM_NONE)
1159 max = mss_now;
1160 copy = max - skb->len;
1161 }
1162
1163 if (copy <= 0) {
1164new_segment:
1165 /* Allocate new segment. If the interface is SG,
1166 * allocate skb fitting to single page.
1167 */
1168 if (!sk_stream_memory_free(sk))
1169 goto wait_for_sndbuf;
1170
1171 skb = sk_stream_alloc_skb(sk,
1172 select_size(sk, sg),
1173 sk->sk_allocation);
1174 if (!skb)
1175 goto wait_for_memory;
1176
1177 /*
1178 * All packets are restored as if they have
1179 * already been sent.
1180 */
1181 if (tp->repair)
1182 TCP_SKB_CB(skb)->when = tcp_time_stamp;
1183
1184 /*
1185 * Check whether we can use HW checksum.
1186 */
1187 if (sk->sk_route_caps & NETIF_F_ALL_CSUM)
1188 skb->ip_summed = CHECKSUM_PARTIAL;
1189
1190 skb_entail(sk, skb);
1191 copy = size_goal;
1192 max = size_goal;
1193 }
1194
1195 /* Try to append data to the end of skb. */
1196 if (copy > seglen)
1197 copy = seglen;
1198
1199 /* Where to copy to? */
1200 if (skb_availroom(skb) > 0) {
1201 /* We have some space in skb head. Superb! */
1202 copy = min_t(int, copy, skb_availroom(skb));
1203 err = skb_add_data_nocache(sk, skb, from, copy);
1204 if (err)
1205 goto do_fault;
1206 } else {
1207 bool merge = true;
1208 int i = skb_shinfo(skb)->nr_frags;
1209 struct page_frag *pfrag = sk_page_frag(sk);
1210
1211 if (!sk_page_frag_refill(sk, pfrag))
1212 goto wait_for_memory;
1213
1214 if (!skb_can_coalesce(skb, i, pfrag->page,
1215 pfrag->offset)) {
1216 if (i == MAX_SKB_FRAGS || !sg) {
1217 tcp_mark_push(tp, skb);
1218 goto new_segment;
1219 }
1220 merge = false;
1221 }
1222
1223 copy = min_t(int, copy, pfrag->size - pfrag->offset);
1224
1225 if (!sk_wmem_schedule(sk, copy))
1226 goto wait_for_memory;
1227
1228 err = skb_copy_to_page_nocache(sk, from, skb,
1229 pfrag->page,
1230 pfrag->offset,
1231 copy);
1232 if (err)
1233 goto do_error;
1234
1235 /* Update the skb. */
1236 if (merge) {
1237 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1238 } else {
1239 skb_fill_page_desc(skb, i, pfrag->page,
1240 pfrag->offset, copy);
1241 get_page(pfrag->page);
1242 }
1243 pfrag->offset += copy;
1244 }
1245
1246 if (!copied)
1247 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
1248
1249 tp->write_seq += copy;
1250 TCP_SKB_CB(skb)->end_seq += copy;
1251 skb_shinfo(skb)->gso_segs = 0;
1252
1253 from += copy;
1254 copied += copy;
1255 if ((seglen -= copy) == 0 && iovlen == 0)
1256 goto out;
1257
1258 if (skb->len < max || (flags & MSG_OOB) || unlikely(tp->repair))
1259 continue;
1260
1261 if (forced_push(tp)) {
1262 tcp_mark_push(tp, skb);
1263 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
1264 } else if (skb == tcp_send_head(sk))
1265 tcp_push_one(sk, mss_now);
1266 continue;
1267
1268wait_for_sndbuf:
1269 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1270wait_for_memory:
1271 if (copied)
1272 tcp_push(sk, flags & ~MSG_MORE, mss_now,
1273 TCP_NAGLE_PUSH, size_goal);
1274
1275 if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
1276 goto do_error;
1277
1278 mss_now = tcp_send_mss(sk, &size_goal, flags);
1279 }
1280 }
1281
1282out:
1283 if (copied)
1284 tcp_push(sk, flags, mss_now, tp->nonagle, size_goal);
1285 release_sock(sk);
1286 return copied + copied_syn;
1287
1288do_fault:
1289 if (!skb->len) {
1290 tcp_unlink_write_queue(skb, sk);
1291 /* It is the one place in all of TCP, except connection
1292 * reset, where we can be unlinking the send_head.
1293 */
1294 tcp_check_send_head(sk, skb);
1295 sk_wmem_free_skb(sk, skb);
1296 }
1297
1298do_error:
1299 if (copied + copied_syn)
1300 goto out;
1301out_err:
1302 err = sk_stream_error(sk, flags, err);
1303 release_sock(sk);
1304 return err;
1305}
1306EXPORT_SYMBOL(tcp_sendmsg);
1307
1308/*
1309 * Handle reading urgent data. BSD has very simple semantics for
1310 * this, no blocking and very strange errors 8)
1311 */
1312
1313static int tcp_recv_urg(struct sock *sk, struct msghdr *msg, int len, int flags)
1314{
1315 struct tcp_sock *tp = tcp_sk(sk);
1316
1317 /* No URG data to read. */
1318 if (sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data ||
1319 tp->urg_data == TCP_URG_READ)
1320 return -EINVAL; /* Yes this is right ! */
1321
1322 if (sk->sk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DONE))
1323 return -ENOTCONN;
1324
1325 if (tp->urg_data & TCP_URG_VALID) {
1326 int err = 0;
1327 char c = tp->urg_data;
1328
1329 if (!(flags & MSG_PEEK))
1330 tp->urg_data = TCP_URG_READ;
1331
1332 /* Read urgent data. */
1333 msg->msg_flags |= MSG_OOB;
1334
1335 if (len > 0) {
1336 if (!(flags & MSG_TRUNC))
1337 err = memcpy_toiovec(msg->msg_iov, &c, 1);
1338 len = 1;
1339 } else
1340 msg->msg_flags |= MSG_TRUNC;
1341
1342 return err ? -EFAULT : len;
1343 }
1344
1345 if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
1346 return 0;
1347
1348 /* Fixed the recv(..., MSG_OOB) behaviour. BSD docs and
1349 * the available implementations agree in this case:
1350 * this call should never block, independent of the
1351 * blocking state of the socket.
1352 * Mike <pall@rz.uni-karlsruhe.de>
1353 */
1354 return -EAGAIN;
1355}
1356
1357static int tcp_peek_sndq(struct sock *sk, struct msghdr *msg, int len)
1358{
1359 struct sk_buff *skb;
1360 int copied = 0, err = 0;
1361
1362 /* XXX -- need to support SO_PEEK_OFF */
1363
1364 skb_queue_walk(&sk->sk_write_queue, skb) {
1365 err = skb_copy_datagram_iovec(skb, 0, msg->msg_iov, skb->len);
1366 if (err)
1367 break;
1368
1369 copied += skb->len;
1370 }
1371
1372 return err ?: copied;
1373}
1374
1375/* Clean up the receive buffer for full frames taken by the user,
1376 * then send an ACK if necessary. COPIED is the number of bytes
1377 * tcp_recvmsg has given to the user so far, it speeds up the
1378 * calculation of whether or not we must ACK for the sake of
1379 * a window update.
1380 */
1381void tcp_cleanup_rbuf(struct sock *sk, int copied)
1382{
1383 struct tcp_sock *tp = tcp_sk(sk);
1384 bool time_to_ack = false;
1385
1386 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
1387
1388 WARN(skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq),
1389 "cleanup rbuf bug: copied %X seq %X rcvnxt %X\n",
1390 tp->copied_seq, TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt);
1391
1392 if (inet_csk_ack_scheduled(sk)) {
1393 const struct inet_connection_sock *icsk = inet_csk(sk);
1394 /* Delayed ACKs frequently hit locked sockets during bulk
1395 * receive. */
1396 if (icsk->icsk_ack.blocked ||
1397 /* Once-per-two-segments ACK was not sent by tcp_input.c */
1398 tp->rcv_nxt - tp->rcv_wup > icsk->icsk_ack.rcv_mss ||
1399 /*
1400 * If this read emptied read buffer, we send ACK, if
1401 * connection is not bidirectional, user drained
1402 * receive buffer and there was a small segment
1403 * in queue.
1404 */
1405 (copied > 0 &&
1406 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED2) ||
1407 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED) &&
1408 !icsk->icsk_ack.pingpong)) &&
1409 !atomic_read(&sk->sk_rmem_alloc)))
1410 time_to_ack = true;
1411 }
1412
1413 /* We send an ACK if we can now advertise a non-zero window
1414 * which has been raised "significantly".
1415 *
1416 * Even if window raised up to infinity, do not send window open ACK
1417 * in states, where we will not receive more. It is useless.
1418 */
1419 if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) {
1420 __u32 rcv_window_now = tcp_receive_window(tp);
1421
1422 /* Optimize, __tcp_select_window() is not cheap. */
1423 if (2*rcv_window_now <= tp->window_clamp) {
1424 __u32 new_window = __tcp_select_window(sk);
1425
1426 /* Send ACK now, if this read freed lots of space
1427 * in our buffer. Certainly, new_window is new window.
1428 * We can advertise it now, if it is not less than current one.
1429 * "Lots" means "at least twice" here.
1430 */
1431 if (new_window && new_window >= 2 * rcv_window_now)
1432 time_to_ack = true;
1433 }
1434 }
1435 if (time_to_ack)
1436 tcp_send_ack(sk);
1437}
1438
1439static void tcp_prequeue_process(struct sock *sk)
1440{
1441 struct sk_buff *skb;
1442 struct tcp_sock *tp = tcp_sk(sk);
1443
1444 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPPREQUEUED);
1445
1446 /* RX process wants to run with disabled BHs, though it is not
1447 * necessary */
1448 local_bh_disable();
1449 while ((skb = __skb_dequeue(&tp->ucopy.prequeue)) != NULL)
1450 sk_backlog_rcv(sk, skb);
1451 local_bh_enable();
1452
1453 /* Clear memory counter. */
1454 tp->ucopy.memory = 0;
1455}
1456
1457#ifdef CONFIG_NET_DMA
1458static void tcp_service_net_dma(struct sock *sk, bool wait)
1459{
1460 dma_cookie_t done, used;
1461 dma_cookie_t last_issued;
1462 struct tcp_sock *tp = tcp_sk(sk);
1463
1464 if (!tp->ucopy.dma_chan)
1465 return;
1466
1467 last_issued = tp->ucopy.dma_cookie;
1468 dma_async_issue_pending(tp->ucopy.dma_chan);
1469
1470 do {
1471 if (dma_async_is_tx_complete(tp->ucopy.dma_chan,
1472 last_issued, &done,
1473 &used) == DMA_COMPLETE) {
1474 /* Safe to free early-copied skbs now */
1475 __skb_queue_purge(&sk->sk_async_wait_queue);
1476 break;
1477 } else {
1478 struct sk_buff *skb;
1479 while ((skb = skb_peek(&sk->sk_async_wait_queue)) &&
1480 (dma_async_is_complete(skb->dma_cookie, done,
1481 used) == DMA_COMPLETE)) {
1482 __skb_dequeue(&sk->sk_async_wait_queue);
1483 kfree_skb(skb);
1484 }
1485 }
1486 } while (wait);
1487}
1488#endif
1489
1490static struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off)
1491{
1492 struct sk_buff *skb;
1493 u32 offset;
1494
1495 while ((skb = skb_peek(&sk->sk_receive_queue)) != NULL) {
1496 offset = seq - TCP_SKB_CB(skb)->seq;
1497 if (tcp_hdr(skb)->syn)
1498 offset--;
1499 if (offset < skb->len || tcp_hdr(skb)->fin) {
1500 *off = offset;
1501 return skb;
1502 }
1503 /* This looks weird, but this can happen if TCP collapsing
1504 * splitted a fat GRO packet, while we released socket lock
1505 * in skb_splice_bits()
1506 */
1507 sk_eat_skb(sk, skb, false);
1508 }
1509 return NULL;
1510}
1511
1512/*
1513 * This routine provides an alternative to tcp_recvmsg() for routines
1514 * that would like to handle copying from skbuffs directly in 'sendfile'
1515 * fashion.
1516 * Note:
1517 * - It is assumed that the socket was locked by the caller.
1518 * - The routine does not block.
1519 * - At present, there is no support for reading OOB data
1520 * or for 'peeking' the socket using this routine
1521 * (although both would be easy to implement).
1522 */
1523int tcp_read_sock(struct sock *sk, read_descriptor_t *desc,
1524 sk_read_actor_t recv_actor)
1525{
1526 struct sk_buff *skb;
1527 struct tcp_sock *tp = tcp_sk(sk);
1528 u32 seq = tp->copied_seq;
1529 u32 offset;
1530 int copied = 0;
1531
1532 if (sk->sk_state == TCP_LISTEN)
1533 return -ENOTCONN;
1534 while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1535 if (offset < skb->len) {
1536 int used;
1537 size_t len;
1538
1539 len = skb->len - offset;
1540 /* Stop reading if we hit a patch of urgent data */
1541 if (tp->urg_data) {
1542 u32 urg_offset = tp->urg_seq - seq;
1543 if (urg_offset < len)
1544 len = urg_offset;
1545 if (!len)
1546 break;
1547 }
1548 used = recv_actor(desc, skb, offset, len);
1549 if (used <= 0) {
1550 if (!copied)
1551 copied = used;
1552 break;
1553 } else if (used <= len) {
1554 seq += used;
1555 copied += used;
1556 offset += used;
1557 }
1558 /* If recv_actor drops the lock (e.g. TCP splice
1559 * receive) the skb pointer might be invalid when
1560 * getting here: tcp_collapse might have deleted it
1561 * while aggregating skbs from the socket queue.
1562 */
1563 skb = tcp_recv_skb(sk, seq - 1, &offset);
1564 if (!skb)
1565 break;
1566 /* TCP coalescing might have appended data to the skb.
1567 * Try to splice more frags
1568 */
1569 if (offset + 1 != skb->len)
1570 continue;
1571 }
1572 if (tcp_hdr(skb)->fin) {
1573 sk_eat_skb(sk, skb, false);
1574 ++seq;
1575 break;
1576 }
1577 sk_eat_skb(sk, skb, false);
1578 if (!desc->count)
1579 break;
1580 tp->copied_seq = seq;
1581 }
1582 tp->copied_seq = seq;
1583
1584 tcp_rcv_space_adjust(sk);
1585
1586 /* Clean up data we have read: This will do ACK frames. */
1587 if (copied > 0) {
1588 tcp_recv_skb(sk, seq, &offset);
1589 tcp_cleanup_rbuf(sk, copied);
1590 }
1591 return copied;
1592}
1593EXPORT_SYMBOL(tcp_read_sock);
1594
1595/*
1596 * This routine copies from a sock struct into the user buffer.
1597 *
1598 * Technical note: in 2.3 we work on _locked_ socket, so that
1599 * tricks with *seq access order and skb->users are not required.
1600 * Probably, code can be easily improved even more.
1601 */
1602
1603int tcp_recvmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg,
1604 size_t len, int nonblock, int flags, int *addr_len)
1605{
1606 struct tcp_sock *tp = tcp_sk(sk);
1607 int copied = 0;
1608 u32 peek_seq;
1609 u32 *seq;
1610 unsigned long used;
1611 int err;
1612 int target; /* Read at least this many bytes */
1613 long timeo;
1614 struct task_struct *user_recv = NULL;
1615 bool copied_early = false;
1616 struct sk_buff *skb;
1617 u32 urg_hole = 0;
1618
1619 if (sk_can_busy_loop(sk) && skb_queue_empty(&sk->sk_receive_queue) &&
1620 (sk->sk_state == TCP_ESTABLISHED))
1621 sk_busy_loop(sk, nonblock);
1622
1623 lock_sock(sk);
1624
1625 err = -ENOTCONN;
1626 if (sk->sk_state == TCP_LISTEN)
1627 goto out;
1628
1629 timeo = sock_rcvtimeo(sk, nonblock);
1630
1631 /* Urgent data needs to be handled specially. */
1632 if (flags & MSG_OOB)
1633 goto recv_urg;
1634
1635 if (unlikely(tp->repair)) {
1636 err = -EPERM;
1637 if (!(flags & MSG_PEEK))
1638 goto out;
1639
1640 if (tp->repair_queue == TCP_SEND_QUEUE)
1641 goto recv_sndq;
1642
1643 err = -EINVAL;
1644 if (tp->repair_queue == TCP_NO_QUEUE)
1645 goto out;
1646
1647 /* 'common' recv queue MSG_PEEK-ing */
1648 }
1649
1650 seq = &tp->copied_seq;
1651 if (flags & MSG_PEEK) {
1652 peek_seq = tp->copied_seq;
1653 seq = &peek_seq;
1654 }
1655
1656 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
1657
1658#ifdef CONFIG_NET_DMA
1659 tp->ucopy.dma_chan = NULL;
1660 preempt_disable();
1661 skb = skb_peek_tail(&sk->sk_receive_queue);
1662 {
1663 int available = 0;
1664
1665 if (skb)
1666 available = TCP_SKB_CB(skb)->seq + skb->len - (*seq);
1667 if ((available < target) &&
1668 (len > sysctl_tcp_dma_copybreak) && !(flags & MSG_PEEK) &&
1669 !sysctl_tcp_low_latency &&
1670 net_dma_find_channel()) {
1671 preempt_enable();
1672 tp->ucopy.pinned_list =
1673 dma_pin_iovec_pages(msg->msg_iov, len);
1674 } else {
1675 preempt_enable();
1676 }
1677 }
1678#endif
1679
1680 do {
1681 u32 offset;
1682
1683 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
1684 if (tp->urg_data && tp->urg_seq == *seq) {
1685 if (copied)
1686 break;
1687 if (signal_pending(current)) {
1688 copied = timeo ? sock_intr_errno(timeo) : -EAGAIN;
1689 break;
1690 }
1691 }
1692
1693 /* Next get a buffer. */
1694
1695 skb_queue_walk(&sk->sk_receive_queue, skb) {
1696 /* Now that we have two receive queues this
1697 * shouldn't happen.
1698 */
1699 if (WARN(before(*seq, TCP_SKB_CB(skb)->seq),
1700 "recvmsg bug: copied %X seq %X rcvnxt %X fl %X\n",
1701 *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt,
1702 flags))
1703 break;
1704
1705 offset = *seq - TCP_SKB_CB(skb)->seq;
1706 if (tcp_hdr(skb)->syn)
1707 offset--;
1708 if (offset < skb->len)
1709 goto found_ok_skb;
1710 if (tcp_hdr(skb)->fin)
1711 goto found_fin_ok;
1712 WARN(!(flags & MSG_PEEK),
1713 "recvmsg bug 2: copied %X seq %X rcvnxt %X fl %X\n",
1714 *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt, flags);
1715 }
1716
1717 /* Well, if we have backlog, try to process it now yet. */
1718
1719 if (copied >= target && !sk->sk_backlog.tail)
1720 break;
1721
1722 if (copied) {
1723 if (sk->sk_err ||
1724 sk->sk_state == TCP_CLOSE ||
1725 (sk->sk_shutdown & RCV_SHUTDOWN) ||
1726 !timeo ||
1727 signal_pending(current))
1728 break;
1729 } else {
1730 if (sock_flag(sk, SOCK_DONE))
1731 break;
1732
1733 if (sk->sk_err) {
1734 copied = sock_error(sk);
1735 break;
1736 }
1737
1738 if (sk->sk_shutdown & RCV_SHUTDOWN)
1739 break;
1740
1741 if (sk->sk_state == TCP_CLOSE) {
1742 if (!sock_flag(sk, SOCK_DONE)) {
1743 /* This occurs when user tries to read
1744 * from never connected socket.
1745 */
1746 copied = -ENOTCONN;
1747 break;
1748 }
1749 break;
1750 }
1751
1752 if (!timeo) {
1753 copied = -EAGAIN;
1754 break;
1755 }
1756
1757 if (signal_pending(current)) {
1758 copied = sock_intr_errno(timeo);
1759 break;
1760 }
1761 }
1762
1763 tcp_cleanup_rbuf(sk, copied);
1764
1765 if (!sysctl_tcp_low_latency && tp->ucopy.task == user_recv) {
1766 /* Install new reader */
1767 if (!user_recv && !(flags & (MSG_TRUNC | MSG_PEEK))) {
1768 user_recv = current;
1769 tp->ucopy.task = user_recv;
1770 tp->ucopy.iov = msg->msg_iov;
1771 }
1772
1773 tp->ucopy.len = len;
1774
1775 WARN_ON(tp->copied_seq != tp->rcv_nxt &&
1776 !(flags & (MSG_PEEK | MSG_TRUNC)));
1777
1778 /* Ugly... If prequeue is not empty, we have to
1779 * process it before releasing socket, otherwise
1780 * order will be broken at second iteration.
1781 * More elegant solution is required!!!
1782 *
1783 * Look: we have the following (pseudo)queues:
1784 *
1785 * 1. packets in flight
1786 * 2. backlog
1787 * 3. prequeue
1788 * 4. receive_queue
1789 *
1790 * Each queue can be processed only if the next ones
1791 * are empty. At this point we have empty receive_queue.
1792 * But prequeue _can_ be not empty after 2nd iteration,
1793 * when we jumped to start of loop because backlog
1794 * processing added something to receive_queue.
1795 * We cannot release_sock(), because backlog contains
1796 * packets arrived _after_ prequeued ones.
1797 *
1798 * Shortly, algorithm is clear --- to process all
1799 * the queues in order. We could make it more directly,
1800 * requeueing packets from backlog to prequeue, if
1801 * is not empty. It is more elegant, but eats cycles,
1802 * unfortunately.
1803 */
1804 if (!skb_queue_empty(&tp->ucopy.prequeue))
1805 goto do_prequeue;
1806
1807 /* __ Set realtime policy in scheduler __ */
1808 }
1809
1810#ifdef CONFIG_NET_DMA
1811 if (tp->ucopy.dma_chan) {
1812 if (tp->rcv_wnd == 0 &&
1813 !skb_queue_empty(&sk->sk_async_wait_queue)) {
1814 tcp_service_net_dma(sk, true);
1815 tcp_cleanup_rbuf(sk, copied);
1816 } else
1817 dma_async_issue_pending(tp->ucopy.dma_chan);
1818 }
1819#endif
1820 if (copied >= target) {
1821 /* Do not sleep, just process backlog. */
1822 release_sock(sk);
1823 lock_sock(sk);
1824 } else
1825 sk_wait_data(sk, &timeo);
1826
1827#ifdef CONFIG_NET_DMA
1828 tcp_service_net_dma(sk, false); /* Don't block */
1829 tp->ucopy.wakeup = 0;
1830#endif
1831
1832 if (user_recv) {
1833 int chunk;
1834
1835 /* __ Restore normal policy in scheduler __ */
1836
1837 if ((chunk = len - tp->ucopy.len) != 0) {
1838 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMBACKLOG, chunk);
1839 len -= chunk;
1840 copied += chunk;
1841 }
1842
1843 if (tp->rcv_nxt == tp->copied_seq &&
1844 !skb_queue_empty(&tp->ucopy.prequeue)) {
1845do_prequeue:
1846 tcp_prequeue_process(sk);
1847
1848 if ((chunk = len - tp->ucopy.len) != 0) {
1849 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1850 len -= chunk;
1851 copied += chunk;
1852 }
1853 }
1854 }
1855 if ((flags & MSG_PEEK) &&
1856 (peek_seq - copied - urg_hole != tp->copied_seq)) {
1857 net_dbg_ratelimited("TCP(%s:%d): Application bug, race in MSG_PEEK\n",
1858 current->comm,
1859 task_pid_nr(current));
1860 peek_seq = tp->copied_seq;
1861 }
1862 continue;
1863
1864 found_ok_skb:
1865 /* Ok so how much can we use? */
1866 used = skb->len - offset;
1867 if (len < used)
1868 used = len;
1869
1870 /* Do we have urgent data here? */
1871 if (tp->urg_data) {
1872 u32 urg_offset = tp->urg_seq - *seq;
1873 if (urg_offset < used) {
1874 if (!urg_offset) {
1875 if (!sock_flag(sk, SOCK_URGINLINE)) {
1876 ++*seq;
1877 urg_hole++;
1878 offset++;
1879 used--;
1880 if (!used)
1881 goto skip_copy;
1882 }
1883 } else
1884 used = urg_offset;
1885 }
1886 }
1887
1888 if (!(flags & MSG_TRUNC)) {
1889#ifdef CONFIG_NET_DMA
1890 if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list)
1891 tp->ucopy.dma_chan = net_dma_find_channel();
1892
1893 if (tp->ucopy.dma_chan) {
1894 tp->ucopy.dma_cookie = dma_skb_copy_datagram_iovec(
1895 tp->ucopy.dma_chan, skb, offset,
1896 msg->msg_iov, used,
1897 tp->ucopy.pinned_list);
1898
1899 if (tp->ucopy.dma_cookie < 0) {
1900
1901 pr_alert("%s: dma_cookie < 0\n",
1902 __func__);
1903
1904 /* Exception. Bailout! */
1905 if (!copied)
1906 copied = -EFAULT;
1907 break;
1908 }
1909
1910 dma_async_issue_pending(tp->ucopy.dma_chan);
1911
1912 if ((offset + used) == skb->len)
1913 copied_early = true;
1914
1915 } else
1916#endif
1917 {
1918 err = skb_copy_datagram_iovec(skb, offset,
1919 msg->msg_iov, used);
1920 if (err) {
1921 /* Exception. Bailout! */
1922 if (!copied)
1923 copied = -EFAULT;
1924 break;
1925 }
1926 }
1927 }
1928
1929 *seq += used;
1930 copied += used;
1931 len -= used;
1932
1933 tcp_rcv_space_adjust(sk);
1934
1935skip_copy:
1936 if (tp->urg_data && after(tp->copied_seq, tp->urg_seq)) {
1937 tp->urg_data = 0;
1938 tcp_fast_path_check(sk);
1939 }
1940 if (used + offset < skb->len)
1941 continue;
1942
1943 if (tcp_hdr(skb)->fin)
1944 goto found_fin_ok;
1945 if (!(flags & MSG_PEEK)) {
1946 sk_eat_skb(sk, skb, copied_early);
1947 copied_early = false;
1948 }
1949 continue;
1950
1951 found_fin_ok:
1952 /* Process the FIN. */
1953 ++*seq;
1954 if (!(flags & MSG_PEEK)) {
1955 sk_eat_skb(sk, skb, copied_early);
1956 copied_early = false;
1957 }
1958 break;
1959 } while (len > 0);
1960
1961 if (user_recv) {
1962 if (!skb_queue_empty(&tp->ucopy.prequeue)) {
1963 int chunk;
1964
1965 tp->ucopy.len = copied > 0 ? len : 0;
1966
1967 tcp_prequeue_process(sk);
1968
1969 if (copied > 0 && (chunk = len - tp->ucopy.len) != 0) {
1970 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1971 len -= chunk;
1972 copied += chunk;
1973 }
1974 }
1975
1976 tp->ucopy.task = NULL;
1977 tp->ucopy.len = 0;
1978 }
1979
1980#ifdef CONFIG_NET_DMA
1981 tcp_service_net_dma(sk, true); /* Wait for queue to drain */
1982 tp->ucopy.dma_chan = NULL;
1983
1984 if (tp->ucopy.pinned_list) {
1985 dma_unpin_iovec_pages(tp->ucopy.pinned_list);
1986 tp->ucopy.pinned_list = NULL;
1987 }
1988#endif
1989
1990 /* According to UNIX98, msg_name/msg_namelen are ignored
1991 * on connected socket. I was just happy when found this 8) --ANK
1992 */
1993
1994 /* Clean up data we have read: This will do ACK frames. */
1995 tcp_cleanup_rbuf(sk, copied);
1996
1997 release_sock(sk);
1998 return copied;
1999
2000out:
2001 release_sock(sk);
2002 return err;
2003
2004recv_urg:
2005 err = tcp_recv_urg(sk, msg, len, flags);
2006 goto out;
2007
2008recv_sndq:
2009 err = tcp_peek_sndq(sk, msg, len);
2010 goto out;
2011}
2012EXPORT_SYMBOL(tcp_recvmsg);
2013
2014void tcp_set_state(struct sock *sk, int state)
2015{
2016 int oldstate = sk->sk_state;
2017
2018 switch (state) {
2019 case TCP_ESTABLISHED:
2020 if (oldstate != TCP_ESTABLISHED)
2021 TCP_INC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
2022 break;
2023
2024 case TCP_CLOSE:
2025 if (oldstate == TCP_CLOSE_WAIT || oldstate == TCP_ESTABLISHED)
2026 TCP_INC_STATS(sock_net(sk), TCP_MIB_ESTABRESETS);
2027
2028 sk->sk_prot->unhash(sk);
2029 if (inet_csk(sk)->icsk_bind_hash &&
2030 !(sk->sk_userlocks & SOCK_BINDPORT_LOCK))
2031 inet_put_port(sk);
2032 /* fall through */
2033 default:
2034 if (oldstate == TCP_ESTABLISHED)
2035 TCP_DEC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
2036 }
2037
2038 /* Change state AFTER socket is unhashed to avoid closed
2039 * socket sitting in hash tables.
2040 */
2041 sk->sk_state = state;
2042
2043#ifdef STATE_TRACE
2044 SOCK_DEBUG(sk, "TCP sk=%p, State %s -> %s\n", sk, statename[oldstate], statename[state]);
2045#endif
2046}
2047EXPORT_SYMBOL_GPL(tcp_set_state);
2048
2049/*
2050 * State processing on a close. This implements the state shift for
2051 * sending our FIN frame. Note that we only send a FIN for some
2052 * states. A shutdown() may have already sent the FIN, or we may be
2053 * closed.
2054 */
2055
2056static const unsigned char new_state[16] = {
2057 /* current state: new state: action: */
2058 /* (Invalid) */ TCP_CLOSE,
2059 /* TCP_ESTABLISHED */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2060 /* TCP_SYN_SENT */ TCP_CLOSE,
2061 /* TCP_SYN_RECV */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2062 /* TCP_FIN_WAIT1 */ TCP_FIN_WAIT1,
2063 /* TCP_FIN_WAIT2 */ TCP_FIN_WAIT2,
2064 /* TCP_TIME_WAIT */ TCP_CLOSE,
2065 /* TCP_CLOSE */ TCP_CLOSE,
2066 /* TCP_CLOSE_WAIT */ TCP_LAST_ACK | TCP_ACTION_FIN,
2067 /* TCP_LAST_ACK */ TCP_LAST_ACK,
2068 /* TCP_LISTEN */ TCP_CLOSE,
2069 /* TCP_CLOSING */ TCP_CLOSING,
2070};
2071
2072static int tcp_close_state(struct sock *sk)
2073{
2074 int next = (int)new_state[sk->sk_state];
2075 int ns = next & TCP_STATE_MASK;
2076
2077 tcp_set_state(sk, ns);
2078
2079 return next & TCP_ACTION_FIN;
2080}
2081
2082/*
2083 * Shutdown the sending side of a connection. Much like close except
2084 * that we don't receive shut down or sock_set_flag(sk, SOCK_DEAD).
2085 */
2086
2087void tcp_shutdown(struct sock *sk, int how)
2088{
2089 /* We need to grab some memory, and put together a FIN,
2090 * and then put it into the queue to be sent.
2091 * Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
2092 */
2093 if (!(how & SEND_SHUTDOWN))
2094 return;
2095
2096 /* If we've already sent a FIN, or it's a closed state, skip this. */
2097 if ((1 << sk->sk_state) &
2098 (TCPF_ESTABLISHED | TCPF_SYN_SENT |
2099 TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) {
2100 /* Clear out any half completed packets. FIN if needed. */
2101 if (tcp_close_state(sk))
2102 tcp_send_fin(sk);
2103 }
2104}
2105EXPORT_SYMBOL(tcp_shutdown);
2106
2107bool tcp_check_oom(struct sock *sk, int shift)
2108{
2109 bool too_many_orphans, out_of_socket_memory;
2110
2111 too_many_orphans = tcp_too_many_orphans(sk, shift);
2112 out_of_socket_memory = tcp_out_of_memory(sk);
2113
2114 if (too_many_orphans)
2115 net_info_ratelimited("too many orphaned sockets\n");
2116 if (out_of_socket_memory)
2117 net_info_ratelimited("out of memory -- consider tuning tcp_mem\n");
2118 return too_many_orphans || out_of_socket_memory;
2119}
2120
2121void tcp_close(struct sock *sk, long timeout)
2122{
2123 struct sk_buff *skb;
2124 int data_was_unread = 0;
2125 int state;
2126
2127 lock_sock(sk);
2128 sk->sk_shutdown = SHUTDOWN_MASK;
2129
2130 if (sk->sk_state == TCP_LISTEN) {
2131 tcp_set_state(sk, TCP_CLOSE);
2132
2133 /* Special case. */
2134 inet_csk_listen_stop(sk);
2135
2136 goto adjudge_to_death;
2137 }
2138
2139 /* We need to flush the recv. buffs. We do this only on the
2140 * descriptor close, not protocol-sourced closes, because the
2141 * reader process may not have drained the data yet!
2142 */
2143 while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) {
2144 u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq -
2145 tcp_hdr(skb)->fin;
2146 data_was_unread += len;
2147 __kfree_skb(skb);
2148 }
2149
2150 sk_mem_reclaim(sk);
2151
2152 /* If socket has been already reset (e.g. in tcp_reset()) - kill it. */
2153 if (sk->sk_state == TCP_CLOSE)
2154 goto adjudge_to_death;
2155
2156 /* As outlined in RFC 2525, section 2.17, we send a RST here because
2157 * data was lost. To witness the awful effects of the old behavior of
2158 * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk
2159 * GET in an FTP client, suspend the process, wait for the client to
2160 * advertise a zero window, then kill -9 the FTP client, wheee...
2161 * Note: timeout is always zero in such a case.
2162 */
2163 if (unlikely(tcp_sk(sk)->repair)) {
2164 sk->sk_prot->disconnect(sk, 0);
2165 } else if (data_was_unread) {
2166 /* Unread data was tossed, zap the connection. */
2167 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONCLOSE);
2168 tcp_set_state(sk, TCP_CLOSE);
2169 tcp_send_active_reset(sk, sk->sk_allocation);
2170 } else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) {
2171 /* Check zero linger _after_ checking for unread data. */
2172 sk->sk_prot->disconnect(sk, 0);
2173 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
2174 } else if (tcp_close_state(sk)) {
2175 /* We FIN if the application ate all the data before
2176 * zapping the connection.
2177 */
2178
2179 /* RED-PEN. Formally speaking, we have broken TCP state
2180 * machine. State transitions:
2181 *
2182 * TCP_ESTABLISHED -> TCP_FIN_WAIT1
2183 * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible)
2184 * TCP_CLOSE_WAIT -> TCP_LAST_ACK
2185 *
2186 * are legal only when FIN has been sent (i.e. in window),
2187 * rather than queued out of window. Purists blame.
2188 *
2189 * F.e. "RFC state" is ESTABLISHED,
2190 * if Linux state is FIN-WAIT-1, but FIN is still not sent.
2191 *
2192 * The visible declinations are that sometimes
2193 * we enter time-wait state, when it is not required really
2194 * (harmless), do not send active resets, when they are
2195 * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
2196 * they look as CLOSING or LAST_ACK for Linux)
2197 * Probably, I missed some more holelets.
2198 * --ANK
2199 * XXX (TFO) - To start off we don't support SYN+ACK+FIN
2200 * in a single packet! (May consider it later but will
2201 * probably need API support or TCP_CORK SYN-ACK until
2202 * data is written and socket is closed.)
2203 */
2204 tcp_send_fin(sk);
2205 }
2206
2207 sk_stream_wait_close(sk, timeout);
2208
2209adjudge_to_death:
2210 state = sk->sk_state;
2211 sock_hold(sk);
2212 sock_orphan(sk);
2213
2214 /* It is the last release_sock in its life. It will remove backlog. */
2215 release_sock(sk);
2216
2217
2218 /* Now socket is owned by kernel and we acquire BH lock
2219 to finish close. No need to check for user refs.
2220 */
2221 local_bh_disable();
2222 bh_lock_sock(sk);
2223 WARN_ON(sock_owned_by_user(sk));
2224
2225 percpu_counter_inc(sk->sk_prot->orphan_count);
2226
2227 /* Have we already been destroyed by a softirq or backlog? */
2228 if (state != TCP_CLOSE && sk->sk_state == TCP_CLOSE)
2229 goto out;
2230
2231 /* This is a (useful) BSD violating of the RFC. There is a
2232 * problem with TCP as specified in that the other end could
2233 * keep a socket open forever with no application left this end.
2234 * We use a 1 minute timeout (about the same as BSD) then kill
2235 * our end. If they send after that then tough - BUT: long enough
2236 * that we won't make the old 4*rto = almost no time - whoops
2237 * reset mistake.
2238 *
2239 * Nope, it was not mistake. It is really desired behaviour
2240 * f.e. on http servers, when such sockets are useless, but
2241 * consume significant resources. Let's do it with special
2242 * linger2 option. --ANK
2243 */
2244
2245 if (sk->sk_state == TCP_FIN_WAIT2) {
2246 struct tcp_sock *tp = tcp_sk(sk);
2247 if (tp->linger2 < 0) {
2248 tcp_set_state(sk, TCP_CLOSE);
2249 tcp_send_active_reset(sk, GFP_ATOMIC);
2250 NET_INC_STATS_BH(sock_net(sk),
2251 LINUX_MIB_TCPABORTONLINGER);
2252 } else {
2253 const int tmo = tcp_fin_time(sk);
2254
2255 if (tmo > TCP_TIMEWAIT_LEN) {
2256 inet_csk_reset_keepalive_timer(sk,
2257 tmo - TCP_TIMEWAIT_LEN);
2258 } else {
2259 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
2260 goto out;
2261 }
2262 }
2263 }
2264 if (sk->sk_state != TCP_CLOSE) {
2265 sk_mem_reclaim(sk);
2266 if (tcp_check_oom(sk, 0)) {
2267 tcp_set_state(sk, TCP_CLOSE);
2268 tcp_send_active_reset(sk, GFP_ATOMIC);
2269 NET_INC_STATS_BH(sock_net(sk),
2270 LINUX_MIB_TCPABORTONMEMORY);
2271 }
2272 }
2273
2274 if (sk->sk_state == TCP_CLOSE) {
2275 struct request_sock *req = tcp_sk(sk)->fastopen_rsk;
2276 /* We could get here with a non-NULL req if the socket is
2277 * aborted (e.g., closed with unread data) before 3WHS
2278 * finishes.
2279 */
2280 if (req != NULL)
2281 reqsk_fastopen_remove(sk, req, false);
2282 inet_csk_destroy_sock(sk);
2283 }
2284 /* Otherwise, socket is reprieved until protocol close. */
2285
2286out:
2287 bh_unlock_sock(sk);
2288 local_bh_enable();
2289 sock_put(sk);
2290}
2291EXPORT_SYMBOL(tcp_close);
2292
2293/* These states need RST on ABORT according to RFC793 */
2294
2295static inline bool tcp_need_reset(int state)
2296{
2297 return (1 << state) &
2298 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 |
2299 TCPF_FIN_WAIT2 | TCPF_SYN_RECV);
2300}
2301
2302int tcp_disconnect(struct sock *sk, int flags)
2303{
2304 struct inet_sock *inet = inet_sk(sk);
2305 struct inet_connection_sock *icsk = inet_csk(sk);
2306 struct tcp_sock *tp = tcp_sk(sk);
2307 int err = 0;
2308 int old_state = sk->sk_state;
2309
2310 if (old_state != TCP_CLOSE)
2311 tcp_set_state(sk, TCP_CLOSE);
2312
2313 /* ABORT function of RFC793 */
2314 if (old_state == TCP_LISTEN) {
2315 inet_csk_listen_stop(sk);
2316 } else if (unlikely(tp->repair)) {
2317 sk->sk_err = ECONNABORTED;
2318 } else if (tcp_need_reset(old_state) ||
2319 (tp->snd_nxt != tp->write_seq &&
2320 (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK))) {
2321 /* The last check adjusts for discrepancy of Linux wrt. RFC
2322 * states
2323 */
2324 tcp_send_active_reset(sk, gfp_any());
2325 sk->sk_err = ECONNRESET;
2326 } else if (old_state == TCP_SYN_SENT)
2327 sk->sk_err = ECONNRESET;
2328
2329 tcp_clear_xmit_timers(sk);
2330 __skb_queue_purge(&sk->sk_receive_queue);
2331 tcp_write_queue_purge(sk);
2332 __skb_queue_purge(&tp->out_of_order_queue);
2333#ifdef CONFIG_NET_DMA
2334 __skb_queue_purge(&sk->sk_async_wait_queue);
2335#endif
2336
2337 inet->inet_dport = 0;
2338
2339 if (!(sk->sk_userlocks & SOCK_BINDADDR_LOCK))
2340 inet_reset_saddr(sk);
2341
2342 sk->sk_shutdown = 0;
2343 sock_reset_flag(sk, SOCK_DONE);
2344 tp->srtt_us = 0;
2345 if ((tp->write_seq += tp->max_window + 2) == 0)
2346 tp->write_seq = 1;
2347 icsk->icsk_backoff = 0;
2348 tp->snd_cwnd = 2;
2349 icsk->icsk_probes_out = 0;
2350 tp->packets_out = 0;
2351 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
2352 tp->snd_cwnd_cnt = 0;
2353 tp->window_clamp = 0;
2354 tcp_set_ca_state(sk, TCP_CA_Open);
2355 tcp_clear_retrans(tp);
2356 inet_csk_delack_init(sk);
2357 tcp_init_send_head(sk);
2358 memset(&tp->rx_opt, 0, sizeof(tp->rx_opt));
2359 __sk_dst_reset(sk);
2360
2361 WARN_ON(inet->inet_num && !icsk->icsk_bind_hash);
2362
2363 sk->sk_error_report(sk);
2364 return err;
2365}
2366EXPORT_SYMBOL(tcp_disconnect);
2367
2368void tcp_sock_destruct(struct sock *sk)
2369{
2370 inet_sock_destruct(sk);
2371
2372 kfree(inet_csk(sk)->icsk_accept_queue.fastopenq);
2373}
2374
2375static inline bool tcp_can_repair_sock(const struct sock *sk)
2376{
2377 return ns_capable(sock_net(sk)->user_ns, CAP_NET_ADMIN) &&
2378 ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_ESTABLISHED));
2379}
2380
2381static int tcp_repair_options_est(struct tcp_sock *tp,
2382 struct tcp_repair_opt __user *optbuf, unsigned int len)
2383{
2384 struct tcp_repair_opt opt;
2385
2386 while (len >= sizeof(opt)) {
2387 if (copy_from_user(&opt, optbuf, sizeof(opt)))
2388 return -EFAULT;
2389
2390 optbuf++;
2391 len -= sizeof(opt);
2392
2393 switch (opt.opt_code) {
2394 case TCPOPT_MSS:
2395 tp->rx_opt.mss_clamp = opt.opt_val;
2396 break;
2397 case TCPOPT_WINDOW:
2398 {
2399 u16 snd_wscale = opt.opt_val & 0xFFFF;
2400 u16 rcv_wscale = opt.opt_val >> 16;
2401
2402 if (snd_wscale > 14 || rcv_wscale > 14)
2403 return -EFBIG;
2404
2405 tp->rx_opt.snd_wscale = snd_wscale;
2406 tp->rx_opt.rcv_wscale = rcv_wscale;
2407 tp->rx_opt.wscale_ok = 1;
2408 }
2409 break;
2410 case TCPOPT_SACK_PERM:
2411 if (opt.opt_val != 0)
2412 return -EINVAL;
2413
2414 tp->rx_opt.sack_ok |= TCP_SACK_SEEN;
2415 if (sysctl_tcp_fack)
2416 tcp_enable_fack(tp);
2417 break;
2418 case TCPOPT_TIMESTAMP:
2419 if (opt.opt_val != 0)
2420 return -EINVAL;
2421
2422 tp->rx_opt.tstamp_ok = 1;
2423 break;
2424 }
2425 }
2426
2427 return 0;
2428}
2429
2430/*
2431 * Socket option code for TCP.
2432 */
2433static int do_tcp_setsockopt(struct sock *sk, int level,
2434 int optname, char __user *optval, unsigned int optlen)
2435{
2436 struct tcp_sock *tp = tcp_sk(sk);
2437 struct inet_connection_sock *icsk = inet_csk(sk);
2438 int val;
2439 int err = 0;
2440
2441 /* These are data/string values, all the others are ints */
2442 switch (optname) {
2443 case TCP_CONGESTION: {
2444 char name[TCP_CA_NAME_MAX];
2445
2446 if (optlen < 1)
2447 return -EINVAL;
2448
2449 val = strncpy_from_user(name, optval,
2450 min_t(long, TCP_CA_NAME_MAX-1, optlen));
2451 if (val < 0)
2452 return -EFAULT;
2453 name[val] = 0;
2454
2455 lock_sock(sk);
2456 err = tcp_set_congestion_control(sk, name);
2457 release_sock(sk);
2458 return err;
2459 }
2460 default:
2461 /* fallthru */
2462 break;
2463 }
2464
2465 if (optlen < sizeof(int))
2466 return -EINVAL;
2467
2468 if (get_user(val, (int __user *)optval))
2469 return -EFAULT;
2470
2471 lock_sock(sk);
2472
2473 switch (optname) {
2474 case TCP_MAXSEG:
2475 /* Values greater than interface MTU won't take effect. However
2476 * at the point when this call is done we typically don't yet
2477 * know which interface is going to be used */
2478 if (val < TCP_MIN_MSS || val > MAX_TCP_WINDOW) {
2479 err = -EINVAL;
2480 break;
2481 }
2482 tp->rx_opt.user_mss = val;
2483 break;
2484
2485 case TCP_NODELAY:
2486 if (val) {
2487 /* TCP_NODELAY is weaker than TCP_CORK, so that
2488 * this option on corked socket is remembered, but
2489 * it is not activated until cork is cleared.
2490 *
2491 * However, when TCP_NODELAY is set we make
2492 * an explicit push, which overrides even TCP_CORK
2493 * for currently queued segments.
2494 */
2495 tp->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH;
2496 tcp_push_pending_frames(sk);
2497 } else {
2498 tp->nonagle &= ~TCP_NAGLE_OFF;
2499 }
2500 break;
2501
2502 case TCP_THIN_LINEAR_TIMEOUTS:
2503 if (val < 0 || val > 1)
2504 err = -EINVAL;
2505 else
2506 tp->thin_lto = val;
2507 break;
2508
2509 case TCP_THIN_DUPACK:
2510 if (val < 0 || val > 1)
2511 err = -EINVAL;
2512 else {
2513 tp->thin_dupack = val;
2514 if (tp->thin_dupack)
2515 tcp_disable_early_retrans(tp);
2516 }
2517 break;
2518
2519 case TCP_REPAIR:
2520 if (!tcp_can_repair_sock(sk))
2521 err = -EPERM;
2522 else if (val == 1) {
2523 tp->repair = 1;
2524 sk->sk_reuse = SK_FORCE_REUSE;
2525 tp->repair_queue = TCP_NO_QUEUE;
2526 } else if (val == 0) {
2527 tp->repair = 0;
2528 sk->sk_reuse = SK_NO_REUSE;
2529 tcp_send_window_probe(sk);
2530 } else
2531 err = -EINVAL;
2532
2533 break;
2534
2535 case TCP_REPAIR_QUEUE:
2536 if (!tp->repair)
2537 err = -EPERM;
2538 else if (val < TCP_QUEUES_NR)
2539 tp->repair_queue = val;
2540 else
2541 err = -EINVAL;
2542 break;
2543
2544 case TCP_QUEUE_SEQ:
2545 if (sk->sk_state != TCP_CLOSE)
2546 err = -EPERM;
2547 else if (tp->repair_queue == TCP_SEND_QUEUE)
2548 tp->write_seq = val;
2549 else if (tp->repair_queue == TCP_RECV_QUEUE)
2550 tp->rcv_nxt = val;
2551 else
2552 err = -EINVAL;
2553 break;
2554
2555 case TCP_REPAIR_OPTIONS:
2556 if (!tp->repair)
2557 err = -EINVAL;
2558 else if (sk->sk_state == TCP_ESTABLISHED)
2559 err = tcp_repair_options_est(tp,
2560 (struct tcp_repair_opt __user *)optval,
2561 optlen);
2562 else
2563 err = -EPERM;
2564 break;
2565
2566 case TCP_CORK:
2567 /* When set indicates to always queue non-full frames.
2568 * Later the user clears this option and we transmit
2569 * any pending partial frames in the queue. This is
2570 * meant to be used alongside sendfile() to get properly
2571 * filled frames when the user (for example) must write
2572 * out headers with a write() call first and then use
2573 * sendfile to send out the data parts.
2574 *
2575 * TCP_CORK can be set together with TCP_NODELAY and it is
2576 * stronger than TCP_NODELAY.
2577 */
2578 if (val) {
2579 tp->nonagle |= TCP_NAGLE_CORK;
2580 } else {
2581 tp->nonagle &= ~TCP_NAGLE_CORK;
2582 if (tp->nonagle&TCP_NAGLE_OFF)
2583 tp->nonagle |= TCP_NAGLE_PUSH;
2584 tcp_push_pending_frames(sk);
2585 }
2586 break;
2587
2588 case TCP_KEEPIDLE:
2589 if (val < 1 || val > MAX_TCP_KEEPIDLE)
2590 err = -EINVAL;
2591 else {
2592 tp->keepalive_time = val * HZ;
2593 if (sock_flag(sk, SOCK_KEEPOPEN) &&
2594 !((1 << sk->sk_state) &
2595 (TCPF_CLOSE | TCPF_LISTEN))) {
2596 u32 elapsed = keepalive_time_elapsed(tp);
2597 if (tp->keepalive_time > elapsed)
2598 elapsed = tp->keepalive_time - elapsed;
2599 else
2600 elapsed = 0;
2601 inet_csk_reset_keepalive_timer(sk, elapsed);
2602 }
2603 }
2604 break;
2605 case TCP_KEEPINTVL:
2606 if (val < 1 || val > MAX_TCP_KEEPINTVL)
2607 err = -EINVAL;
2608 else
2609 tp->keepalive_intvl = val * HZ;
2610 break;
2611 case TCP_KEEPCNT:
2612 if (val < 1 || val > MAX_TCP_KEEPCNT)
2613 err = -EINVAL;
2614 else
2615 tp->keepalive_probes = val;
2616 break;
2617 case TCP_SYNCNT:
2618 if (val < 1 || val > MAX_TCP_SYNCNT)
2619 err = -EINVAL;
2620 else
2621 icsk->icsk_syn_retries = val;
2622 break;
2623
2624 case TCP_LINGER2:
2625 if (val < 0)
2626 tp->linger2 = -1;
2627 else if (val > sysctl_tcp_fin_timeout / HZ)
2628 tp->linger2 = 0;
2629 else
2630 tp->linger2 = val * HZ;
2631 break;
2632
2633 case TCP_DEFER_ACCEPT:
2634 /* Translate value in seconds to number of retransmits */
2635 icsk->icsk_accept_queue.rskq_defer_accept =
2636 secs_to_retrans(val, TCP_TIMEOUT_INIT / HZ,
2637 TCP_RTO_MAX / HZ);
2638 break;
2639
2640 case TCP_WINDOW_CLAMP:
2641 if (!val) {
2642 if (sk->sk_state != TCP_CLOSE) {
2643 err = -EINVAL;
2644 break;
2645 }
2646 tp->window_clamp = 0;
2647 } else
2648 tp->window_clamp = val < SOCK_MIN_RCVBUF / 2 ?
2649 SOCK_MIN_RCVBUF / 2 : val;
2650 break;
2651
2652 case TCP_QUICKACK:
2653 if (!val) {
2654 icsk->icsk_ack.pingpong = 1;
2655 } else {
2656 icsk->icsk_ack.pingpong = 0;
2657 if ((1 << sk->sk_state) &
2658 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) &&
2659 inet_csk_ack_scheduled(sk)) {
2660 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
2661 tcp_cleanup_rbuf(sk, 1);
2662 if (!(val & 1))
2663 icsk->icsk_ack.pingpong = 1;
2664 }
2665 }
2666 break;
2667
2668#ifdef CONFIG_TCP_MD5SIG
2669 case TCP_MD5SIG:
2670 /* Read the IP->Key mappings from userspace */
2671 err = tp->af_specific->md5_parse(sk, optval, optlen);
2672 break;
2673#endif
2674 case TCP_USER_TIMEOUT:
2675 /* Cap the max timeout in ms TCP will retry/retrans
2676 * before giving up and aborting (ETIMEDOUT) a connection.
2677 */
2678 if (val < 0)
2679 err = -EINVAL;
2680 else
2681 icsk->icsk_user_timeout = msecs_to_jiffies(val);
2682 break;
2683
2684 case TCP_FASTOPEN:
2685 if (val >= 0 && ((1 << sk->sk_state) & (TCPF_CLOSE |
2686 TCPF_LISTEN)))
2687 err = fastopen_init_queue(sk, val);
2688 else
2689 err = -EINVAL;
2690 break;
2691 case TCP_TIMESTAMP:
2692 if (!tp->repair)
2693 err = -EPERM;
2694 else
2695 tp->tsoffset = val - tcp_time_stamp;
2696 break;
2697 case TCP_NOTSENT_LOWAT:
2698 tp->notsent_lowat = val;
2699 sk->sk_write_space(sk);
2700 break;
2701 default:
2702 err = -ENOPROTOOPT;
2703 break;
2704 }
2705
2706 release_sock(sk);
2707 return err;
2708}
2709
2710int tcp_setsockopt(struct sock *sk, int level, int optname, char __user *optval,
2711 unsigned int optlen)
2712{
2713 const struct inet_connection_sock *icsk = inet_csk(sk);
2714
2715 if (level != SOL_TCP)
2716 return icsk->icsk_af_ops->setsockopt(sk, level, optname,
2717 optval, optlen);
2718 return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2719}
2720EXPORT_SYMBOL(tcp_setsockopt);
2721
2722#ifdef CONFIG_COMPAT
2723int compat_tcp_setsockopt(struct sock *sk, int level, int optname,
2724 char __user *optval, unsigned int optlen)
2725{
2726 if (level != SOL_TCP)
2727 return inet_csk_compat_setsockopt(sk, level, optname,
2728 optval, optlen);
2729 return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2730}
2731EXPORT_SYMBOL(compat_tcp_setsockopt);
2732#endif
2733
2734/* Return information about state of tcp endpoint in API format. */
2735void tcp_get_info(const struct sock *sk, struct tcp_info *info)
2736{
2737 const struct tcp_sock *tp = tcp_sk(sk);
2738 const struct inet_connection_sock *icsk = inet_csk(sk);
2739 u32 now = tcp_time_stamp;
2740
2741 memset(info, 0, sizeof(*info));
2742
2743 info->tcpi_state = sk->sk_state;
2744 info->tcpi_ca_state = icsk->icsk_ca_state;
2745 info->tcpi_retransmits = icsk->icsk_retransmits;
2746 info->tcpi_probes = icsk->icsk_probes_out;
2747 info->tcpi_backoff = icsk->icsk_backoff;
2748
2749 if (tp->rx_opt.tstamp_ok)
2750 info->tcpi_options |= TCPI_OPT_TIMESTAMPS;
2751 if (tcp_is_sack(tp))
2752 info->tcpi_options |= TCPI_OPT_SACK;
2753 if (tp->rx_opt.wscale_ok) {
2754 info->tcpi_options |= TCPI_OPT_WSCALE;
2755 info->tcpi_snd_wscale = tp->rx_opt.snd_wscale;
2756 info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale;
2757 }
2758
2759 if (tp->ecn_flags & TCP_ECN_OK)
2760 info->tcpi_options |= TCPI_OPT_ECN;
2761 if (tp->ecn_flags & TCP_ECN_SEEN)
2762 info->tcpi_options |= TCPI_OPT_ECN_SEEN;
2763 if (tp->syn_data_acked)
2764 info->tcpi_options |= TCPI_OPT_SYN_DATA;
2765
2766 info->tcpi_rto = jiffies_to_usecs(icsk->icsk_rto);
2767 info->tcpi_ato = jiffies_to_usecs(icsk->icsk_ack.ato);
2768 info->tcpi_snd_mss = tp->mss_cache;
2769 info->tcpi_rcv_mss = icsk->icsk_ack.rcv_mss;
2770
2771 if (sk->sk_state == TCP_LISTEN) {
2772 info->tcpi_unacked = sk->sk_ack_backlog;
2773 info->tcpi_sacked = sk->sk_max_ack_backlog;
2774 } else {
2775 info->tcpi_unacked = tp->packets_out;
2776 info->tcpi_sacked = tp->sacked_out;
2777 }
2778 info->tcpi_lost = tp->lost_out;
2779 info->tcpi_retrans = tp->retrans_out;
2780 info->tcpi_fackets = tp->fackets_out;
2781
2782 info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime);
2783 info->tcpi_last_data_recv = jiffies_to_msecs(now - icsk->icsk_ack.lrcvtime);
2784 info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp);
2785
2786 info->tcpi_pmtu = icsk->icsk_pmtu_cookie;
2787 info->tcpi_rcv_ssthresh = tp->rcv_ssthresh;
2788 info->tcpi_rtt = tp->srtt_us >> 3;
2789 info->tcpi_rttvar = tp->mdev_us >> 2;
2790 info->tcpi_snd_ssthresh = tp->snd_ssthresh;
2791 info->tcpi_snd_cwnd = tp->snd_cwnd;
2792 info->tcpi_advmss = tp->advmss;
2793 info->tcpi_reordering = tp->reordering;
2794
2795 info->tcpi_rcv_rtt = jiffies_to_usecs(tp->rcv_rtt_est.rtt)>>3;
2796 info->tcpi_rcv_space = tp->rcvq_space.space;
2797
2798 info->tcpi_total_retrans = tp->total_retrans;
2799
2800 info->tcpi_pacing_rate = sk->sk_pacing_rate != ~0U ?
2801 sk->sk_pacing_rate : ~0ULL;
2802 info->tcpi_max_pacing_rate = sk->sk_max_pacing_rate != ~0U ?
2803 sk->sk_max_pacing_rate : ~0ULL;
2804}
2805EXPORT_SYMBOL_GPL(tcp_get_info);
2806
2807static int do_tcp_getsockopt(struct sock *sk, int level,
2808 int optname, char __user *optval, int __user *optlen)
2809{
2810 struct inet_connection_sock *icsk = inet_csk(sk);
2811 struct tcp_sock *tp = tcp_sk(sk);
2812 int val, len;
2813
2814 if (get_user(len, optlen))
2815 return -EFAULT;
2816
2817 len = min_t(unsigned int, len, sizeof(int));
2818
2819 if (len < 0)
2820 return -EINVAL;
2821
2822 switch (optname) {
2823 case TCP_MAXSEG:
2824 val = tp->mss_cache;
2825 if (!val && ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
2826 val = tp->rx_opt.user_mss;
2827 if (tp->repair)
2828 val = tp->rx_opt.mss_clamp;
2829 break;
2830 case TCP_NODELAY:
2831 val = !!(tp->nonagle&TCP_NAGLE_OFF);
2832 break;
2833 case TCP_CORK:
2834 val = !!(tp->nonagle&TCP_NAGLE_CORK);
2835 break;
2836 case TCP_KEEPIDLE:
2837 val = keepalive_time_when(tp) / HZ;
2838 break;
2839 case TCP_KEEPINTVL:
2840 val = keepalive_intvl_when(tp) / HZ;
2841 break;
2842 case TCP_KEEPCNT:
2843 val = keepalive_probes(tp);
2844 break;
2845 case TCP_SYNCNT:
2846 val = icsk->icsk_syn_retries ? : sysctl_tcp_syn_retries;
2847 break;
2848 case TCP_LINGER2:
2849 val = tp->linger2;
2850 if (val >= 0)
2851 val = (val ? : sysctl_tcp_fin_timeout) / HZ;
2852 break;
2853 case TCP_DEFER_ACCEPT:
2854 val = retrans_to_secs(icsk->icsk_accept_queue.rskq_defer_accept,
2855 TCP_TIMEOUT_INIT / HZ, TCP_RTO_MAX / HZ);
2856 break;
2857 case TCP_WINDOW_CLAMP:
2858 val = tp->window_clamp;
2859 break;
2860 case TCP_INFO: {
2861 struct tcp_info info;
2862
2863 if (get_user(len, optlen))
2864 return -EFAULT;
2865
2866 tcp_get_info(sk, &info);
2867
2868 len = min_t(unsigned int, len, sizeof(info));
2869 if (put_user(len, optlen))
2870 return -EFAULT;
2871 if (copy_to_user(optval, &info, len))
2872 return -EFAULT;
2873 return 0;
2874 }
2875 case TCP_QUICKACK:
2876 val = !icsk->icsk_ack.pingpong;
2877 break;
2878
2879 case TCP_CONGESTION:
2880 if (get_user(len, optlen))
2881 return -EFAULT;
2882 len = min_t(unsigned int, len, TCP_CA_NAME_MAX);
2883 if (put_user(len, optlen))
2884 return -EFAULT;
2885 if (copy_to_user(optval, icsk->icsk_ca_ops->name, len))
2886 return -EFAULT;
2887 return 0;
2888
2889 case TCP_THIN_LINEAR_TIMEOUTS:
2890 val = tp->thin_lto;
2891 break;
2892 case TCP_THIN_DUPACK:
2893 val = tp->thin_dupack;
2894 break;
2895
2896 case TCP_REPAIR:
2897 val = tp->repair;
2898 break;
2899
2900 case TCP_REPAIR_QUEUE:
2901 if (tp->repair)
2902 val = tp->repair_queue;
2903 else
2904 return -EINVAL;
2905 break;
2906
2907 case TCP_QUEUE_SEQ:
2908 if (tp->repair_queue == TCP_SEND_QUEUE)
2909 val = tp->write_seq;
2910 else if (tp->repair_queue == TCP_RECV_QUEUE)
2911 val = tp->rcv_nxt;
2912 else
2913 return -EINVAL;
2914 break;
2915
2916 case TCP_USER_TIMEOUT:
2917 val = jiffies_to_msecs(icsk->icsk_user_timeout);
2918 break;
2919 case TCP_TIMESTAMP:
2920 val = tcp_time_stamp + tp->tsoffset;
2921 break;
2922 case TCP_NOTSENT_LOWAT:
2923 val = tp->notsent_lowat;
2924 break;
2925 default:
2926 return -ENOPROTOOPT;
2927 }
2928
2929 if (put_user(len, optlen))
2930 return -EFAULT;
2931 if (copy_to_user(optval, &val, len))
2932 return -EFAULT;
2933 return 0;
2934}
2935
2936int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval,
2937 int __user *optlen)
2938{
2939 struct inet_connection_sock *icsk = inet_csk(sk);
2940
2941 if (level != SOL_TCP)
2942 return icsk->icsk_af_ops->getsockopt(sk, level, optname,
2943 optval, optlen);
2944 return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2945}
2946EXPORT_SYMBOL(tcp_getsockopt);
2947
2948#ifdef CONFIG_COMPAT
2949int compat_tcp_getsockopt(struct sock *sk, int level, int optname,
2950 char __user *optval, int __user *optlen)
2951{
2952 if (level != SOL_TCP)
2953 return inet_csk_compat_getsockopt(sk, level, optname,
2954 optval, optlen);
2955 return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2956}
2957EXPORT_SYMBOL(compat_tcp_getsockopt);
2958#endif
2959
2960#ifdef CONFIG_TCP_MD5SIG
2961static struct tcp_md5sig_pool __percpu *tcp_md5sig_pool __read_mostly;
2962static DEFINE_MUTEX(tcp_md5sig_mutex);
2963
2964static void __tcp_free_md5sig_pool(struct tcp_md5sig_pool __percpu *pool)
2965{
2966 int cpu;
2967
2968 for_each_possible_cpu(cpu) {
2969 struct tcp_md5sig_pool *p = per_cpu_ptr(pool, cpu);
2970
2971 if (p->md5_desc.tfm)
2972 crypto_free_hash(p->md5_desc.tfm);
2973 }
2974 free_percpu(pool);
2975}
2976
2977static void __tcp_alloc_md5sig_pool(void)
2978{
2979 int cpu;
2980 struct tcp_md5sig_pool __percpu *pool;
2981
2982 pool = alloc_percpu(struct tcp_md5sig_pool);
2983 if (!pool)
2984 return;
2985
2986 for_each_possible_cpu(cpu) {
2987 struct crypto_hash *hash;
2988
2989 hash = crypto_alloc_hash("md5", 0, CRYPTO_ALG_ASYNC);
2990 if (IS_ERR_OR_NULL(hash))
2991 goto out_free;
2992
2993 per_cpu_ptr(pool, cpu)->md5_desc.tfm = hash;
2994 }
2995 /* before setting tcp_md5sig_pool, we must commit all writes
2996 * to memory. See ACCESS_ONCE() in tcp_get_md5sig_pool()
2997 */
2998 smp_wmb();
2999 tcp_md5sig_pool = pool;
3000 return;
3001out_free:
3002 __tcp_free_md5sig_pool(pool);
3003}
3004
3005bool tcp_alloc_md5sig_pool(void)
3006{
3007 if (unlikely(!tcp_md5sig_pool)) {
3008 mutex_lock(&tcp_md5sig_mutex);
3009
3010 if (!tcp_md5sig_pool)
3011 __tcp_alloc_md5sig_pool();
3012
3013 mutex_unlock(&tcp_md5sig_mutex);
3014 }
3015 return tcp_md5sig_pool != NULL;
3016}
3017EXPORT_SYMBOL(tcp_alloc_md5sig_pool);
3018
3019
3020/**
3021 * tcp_get_md5sig_pool - get md5sig_pool for this user
3022 *
3023 * We use percpu structure, so if we succeed, we exit with preemption
3024 * and BH disabled, to make sure another thread or softirq handling
3025 * wont try to get same context.
3026 */
3027struct tcp_md5sig_pool *tcp_get_md5sig_pool(void)
3028{
3029 struct tcp_md5sig_pool __percpu *p;
3030
3031 local_bh_disable();
3032 p = ACCESS_ONCE(tcp_md5sig_pool);
3033 if (p)
3034 return __this_cpu_ptr(p);
3035
3036 local_bh_enable();
3037 return NULL;
3038}
3039EXPORT_SYMBOL(tcp_get_md5sig_pool);
3040
3041int tcp_md5_hash_header(struct tcp_md5sig_pool *hp,
3042 const struct tcphdr *th)
3043{
3044 struct scatterlist sg;
3045 struct tcphdr hdr;
3046 int err;
3047
3048 /* We are not allowed to change tcphdr, make a local copy */
3049 memcpy(&hdr, th, sizeof(hdr));
3050 hdr.check = 0;
3051
3052 /* options aren't included in the hash */
3053 sg_init_one(&sg, &hdr, sizeof(hdr));
3054 err = crypto_hash_update(&hp->md5_desc, &sg, sizeof(hdr));
3055 return err;
3056}
3057EXPORT_SYMBOL(tcp_md5_hash_header);
3058
3059int tcp_md5_hash_skb_data(struct tcp_md5sig_pool *hp,
3060 const struct sk_buff *skb, unsigned int header_len)
3061{
3062 struct scatterlist sg;
3063 const struct tcphdr *tp = tcp_hdr(skb);
3064 struct hash_desc *desc = &hp->md5_desc;
3065 unsigned int i;
3066 const unsigned int head_data_len = skb_headlen(skb) > header_len ?
3067 skb_headlen(skb) - header_len : 0;
3068 const struct skb_shared_info *shi = skb_shinfo(skb);
3069 struct sk_buff *frag_iter;
3070
3071 sg_init_table(&sg, 1);
3072
3073 sg_set_buf(&sg, ((u8 *) tp) + header_len, head_data_len);
3074 if (crypto_hash_update(desc, &sg, head_data_len))
3075 return 1;
3076
3077 for (i = 0; i < shi->nr_frags; ++i) {
3078 const struct skb_frag_struct *f = &shi->frags[i];
3079 unsigned int offset = f->page_offset;
3080 struct page *page = skb_frag_page(f) + (offset >> PAGE_SHIFT);
3081
3082 sg_set_page(&sg, page, skb_frag_size(f),
3083 offset_in_page(offset));
3084 if (crypto_hash_update(desc, &sg, skb_frag_size(f)))
3085 return 1;
3086 }
3087
3088 skb_walk_frags(skb, frag_iter)
3089 if (tcp_md5_hash_skb_data(hp, frag_iter, 0))
3090 return 1;
3091
3092 return 0;
3093}
3094EXPORT_SYMBOL(tcp_md5_hash_skb_data);
3095
3096int tcp_md5_hash_key(struct tcp_md5sig_pool *hp, const struct tcp_md5sig_key *key)
3097{
3098 struct scatterlist sg;
3099
3100 sg_init_one(&sg, key->key, key->keylen);
3101 return crypto_hash_update(&hp->md5_desc, &sg, key->keylen);
3102}
3103EXPORT_SYMBOL(tcp_md5_hash_key);
3104
3105#endif
3106
3107void tcp_done(struct sock *sk)
3108{
3109 struct request_sock *req = tcp_sk(sk)->fastopen_rsk;
3110
3111 if (sk->sk_state == TCP_SYN_SENT || sk->sk_state == TCP_SYN_RECV)
3112 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_ATTEMPTFAILS);
3113
3114 tcp_set_state(sk, TCP_CLOSE);
3115 tcp_clear_xmit_timers(sk);
3116 if (req != NULL)
3117 reqsk_fastopen_remove(sk, req, false);
3118
3119 sk->sk_shutdown = SHUTDOWN_MASK;
3120
3121 if (!sock_flag(sk, SOCK_DEAD))
3122 sk->sk_state_change(sk);
3123 else
3124 inet_csk_destroy_sock(sk);
3125}
3126EXPORT_SYMBOL_GPL(tcp_done);
3127
3128extern struct tcp_congestion_ops tcp_reno;
3129
3130static __initdata unsigned long thash_entries;
3131static int __init set_thash_entries(char *str)
3132{
3133 ssize_t ret;
3134
3135 if (!str)
3136 return 0;
3137
3138 ret = kstrtoul(str, 0, &thash_entries);
3139 if (ret)
3140 return 0;
3141
3142 return 1;
3143}
3144__setup("thash_entries=", set_thash_entries);
3145
3146static void tcp_init_mem(void)
3147{
3148 unsigned long limit = nr_free_buffer_pages() / 8;
3149 limit = max(limit, 128UL);
3150 sysctl_tcp_mem[0] = limit / 4 * 3;
3151 sysctl_tcp_mem[1] = limit;
3152 sysctl_tcp_mem[2] = sysctl_tcp_mem[0] * 2;
3153}
3154
3155void __init tcp_init(void)
3156{
3157 struct sk_buff *skb = NULL;
3158 unsigned long limit;
3159 int max_rshare, max_wshare, cnt;
3160 unsigned int i;
3161
3162 BUILD_BUG_ON(sizeof(struct tcp_skb_cb) > sizeof(skb->cb));
3163
3164 percpu_counter_init(&tcp_sockets_allocated, 0);
3165 percpu_counter_init(&tcp_orphan_count, 0);
3166 tcp_hashinfo.bind_bucket_cachep =
3167 kmem_cache_create("tcp_bind_bucket",
3168 sizeof(struct inet_bind_bucket), 0,
3169 SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL);
3170
3171 /* Size and allocate the main established and bind bucket
3172 * hash tables.
3173 *
3174 * The methodology is similar to that of the buffer cache.
3175 */
3176 tcp_hashinfo.ehash =
3177 alloc_large_system_hash("TCP established",
3178 sizeof(struct inet_ehash_bucket),
3179 thash_entries,
3180 17, /* one slot per 128 KB of memory */
3181 0,
3182 NULL,
3183 &tcp_hashinfo.ehash_mask,
3184 0,
3185 thash_entries ? 0 : 512 * 1024);
3186 for (i = 0; i <= tcp_hashinfo.ehash_mask; i++)
3187 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].chain, i);
3188
3189 if (inet_ehash_locks_alloc(&tcp_hashinfo))
3190 panic("TCP: failed to alloc ehash_locks");
3191 tcp_hashinfo.bhash =
3192 alloc_large_system_hash("TCP bind",
3193 sizeof(struct inet_bind_hashbucket),
3194 tcp_hashinfo.ehash_mask + 1,
3195 17, /* one slot per 128 KB of memory */
3196 0,
3197 &tcp_hashinfo.bhash_size,
3198 NULL,
3199 0,
3200 64 * 1024);
3201 tcp_hashinfo.bhash_size = 1U << tcp_hashinfo.bhash_size;
3202 for (i = 0; i < tcp_hashinfo.bhash_size; i++) {
3203 spin_lock_init(&tcp_hashinfo.bhash[i].lock);
3204 INIT_HLIST_HEAD(&tcp_hashinfo.bhash[i].chain);
3205 }
3206
3207
3208 cnt = tcp_hashinfo.ehash_mask + 1;
3209
3210 tcp_death_row.sysctl_max_tw_buckets = cnt / 2;
3211 sysctl_tcp_max_orphans = cnt / 2;
3212 sysctl_max_syn_backlog = max(128, cnt / 256);
3213
3214 tcp_init_mem();
3215 /* Set per-socket limits to no more than 1/128 the pressure threshold */
3216 limit = nr_free_buffer_pages() << (PAGE_SHIFT - 7);
3217 max_wshare = min(4UL*1024*1024, limit);
3218 max_rshare = min(6UL*1024*1024, limit);
3219
3220 sysctl_tcp_wmem[0] = SK_MEM_QUANTUM;
3221 sysctl_tcp_wmem[1] = 16*1024;
3222 sysctl_tcp_wmem[2] = max(64*1024, max_wshare);
3223
3224 sysctl_tcp_rmem[0] = SK_MEM_QUANTUM;
3225 sysctl_tcp_rmem[1] = 87380;
3226 sysctl_tcp_rmem[2] = max(87380, max_rshare);
3227
3228 pr_info("Hash tables configured (established %u bind %u)\n",
3229 tcp_hashinfo.ehash_mask + 1, tcp_hashinfo.bhash_size);
3230
3231 tcp_metrics_init();
3232
3233 tcp_register_congestion_control(&tcp_reno);
3234
3235 tcp_tasklet_init();
3236}