projects / deliverable / linux.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
blame | history | raw | HEAD
Merge branch 'hwmon-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/jdelv...
[deliverable/linux.git] / net / ipv4 / tcp.c
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
268 #include <net/icmp.h>
269 #include <net/tcp.h>
270 #include <net/xfrm.h>
271 #include <net/ip.h>
272 #include <net/netdma.h>
273 #include <net/sock.h>
274
275 #include <asm/uaccess.h>
276 #include <asm/ioctls.h>
277
278 int sysctl_tcp_fin_timeout __read_mostly = TCP_FIN_TIMEOUT;
279
280 struct percpu_counter tcp_orphan_count;
281 EXPORT_SYMBOL_GPL(tcp_orphan_count);
282
283 int sysctl_tcp_mem[3] __read_mostly;
284 int sysctl_tcp_wmem[3] __read_mostly;
285 int sysctl_tcp_rmem[3] __read_mostly;
286
287 EXPORT_SYMBOL(sysctl_tcp_mem);
288 EXPORT_SYMBOL(sysctl_tcp_rmem);
289 EXPORT_SYMBOL(sysctl_tcp_wmem);
290
291 atomic_t tcp_memory_allocated; /* Current allocated memory. */
292 EXPORT_SYMBOL(tcp_memory_allocated);
293
294 /*
295 * Current number of TCP sockets.
296 */
297 struct percpu_counter tcp_sockets_allocated;
298 EXPORT_SYMBOL(tcp_sockets_allocated);
299
300 /*
301 * TCP splice context
302 */
303 struct tcp_splice_state {
304 struct pipe_inode_info *pipe;
305 size_t len;
306 unsigned int flags;
307 };
308
309 /*
310 * Pressure flag: try to collapse.
311 * Technical note: it is used by multiple contexts non atomically.
312 * All the __sk_mem_schedule() is of this nature: accounting
313 * is strict, actions are advisory and have some latency.
314 */
315 int tcp_memory_pressure __read_mostly;
316
317 EXPORT_SYMBOL(tcp_memory_pressure);
318
319 void tcp_enter_memory_pressure(struct sock *sk)
320 {
321 if (!tcp_memory_pressure) {
322 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMEMORYPRESSURES);
323 tcp_memory_pressure = 1;
324 }
325 }
326
327 EXPORT_SYMBOL(tcp_enter_memory_pressure);
328
329 /*
330 * Wait for a TCP event.
331 *
332 * Note that we don't need to lock the socket, as the upper poll layers
333 * take care of normal races (between the test and the event) and we don't
334 * go look at any of the socket buffers directly.
335 */
336 unsigned int tcp_poll(struct file *file, struct socket *sock, poll_table *wait)
337 {
338 unsigned int mask;
339 struct sock *sk = sock->sk;
340 struct tcp_sock *tp = tcp_sk(sk);
341
342 poll_wait(file, sk->sk_sleep, wait);
343 if (sk->sk_state == TCP_LISTEN)
344 return inet_csk_listen_poll(sk);
345
346 /* Socket is not locked. We are protected from async events
347 * by poll logic and correct handling of state changes
348 * made by other threads is impossible in any case.
349 */
350
351 mask = 0;
352 if (sk->sk_err)
353 mask = POLLERR;
354
355 /*
356 * POLLHUP is certainly not done right. But poll() doesn't
357 * have a notion of HUP in just one direction, and for a
358 * socket the read side is more interesting.
359 *
360 * Some poll() documentation says that POLLHUP is incompatible
361 * with the POLLOUT/POLLWR flags, so somebody should check this
362 * all. But careful, it tends to be safer to return too many
363 * bits than too few, and you can easily break real applications
364 * if you don't tell them that something has hung up!
365 *
366 * Check-me.
367 *
368 * Check number 1. POLLHUP is _UNMASKABLE_ event (see UNIX98 and
369 * our fs/select.c). It means that after we received EOF,
370 * poll always returns immediately, making impossible poll() on write()
371 * in state CLOSE_WAIT. One solution is evident --- to set POLLHUP
372 * if and only if shutdown has been made in both directions.
373 * Actually, it is interesting to look how Solaris and DUX
374 * solve this dilemma. I would prefer, if POLLHUP were maskable,
375 * then we could set it on SND_SHUTDOWN. BTW examples given
376 * in Stevens' books assume exactly this behaviour, it explains
377 * why POLLHUP is incompatible with POLLOUT. --ANK
378 *
379 * NOTE. Check for TCP_CLOSE is added. The goal is to prevent
380 * blocking on fresh not-connected or disconnected socket. --ANK
381 */
382 if (sk->sk_shutdown == SHUTDOWN_MASK || sk->sk_state == TCP_CLOSE)
383 mask |= POLLHUP;
384 if (sk->sk_shutdown & RCV_SHUTDOWN)
385 mask |= POLLIN | POLLRDNORM | POLLRDHUP;
386
387 /* Connected? */
388 if ((1 << sk->sk_state) & ~(TCPF_SYN_SENT | TCPF_SYN_RECV)) {
389 int target = sock_rcvlowat(sk, 0, INT_MAX);
390
391 if (tp->urg_seq == tp->copied_seq &&
392 !sock_flag(sk, SOCK_URGINLINE) &&
393 tp->urg_data)
394 target--;
395
396 /* Potential race condition. If read of tp below will
397 * escape above sk->sk_state, we can be illegally awaken
398 * in SYN_* states. */
399 if (tp->rcv_nxt - tp->copied_seq >= target)
400 mask |= POLLIN | POLLRDNORM;
401
402 if (!(sk->sk_shutdown & SEND_SHUTDOWN)) {
403 if (sk_stream_wspace(sk) >= sk_stream_min_wspace(sk)) {
404 mask |= POLLOUT | POLLWRNORM;
405 } else { /* send SIGIO later */
406 set_bit(SOCK_ASYNC_NOSPACE,
407 &sk->sk_socket->flags);
408 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
409
410 /* Race breaker. If space is freed after
411 * wspace test but before the flags are set,
412 * IO signal will be lost.
413 */
414 if (sk_stream_wspace(sk) >= sk_stream_min_wspace(sk))
415 mask |= POLLOUT | POLLWRNORM;
416 }
417 }
418
419 if (tp->urg_data & TCP_URG_VALID)
420 mask |= POLLPRI;
421 }
422 return mask;
423 }
424
425 int tcp_ioctl(struct sock *sk, int cmd, unsigned long arg)
426 {
427 struct tcp_sock *tp = tcp_sk(sk);
428 int answ;
429
430 switch (cmd) {
431 case SIOCINQ:
432 if (sk->sk_state == TCP_LISTEN)
433 return -EINVAL;
434
435 lock_sock(sk);
436 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
437 answ = 0;
438 else if (sock_flag(sk, SOCK_URGINLINE) ||
439 !tp->urg_data ||
440 before(tp->urg_seq, tp->copied_seq) ||
441 !before(tp->urg_seq, tp->rcv_nxt)) {
442 struct sk_buff *skb;
443
444 answ = tp->rcv_nxt - tp->copied_seq;
445
446 /* Subtract 1, if FIN is in queue. */
447 skb = skb_peek_tail(&sk->sk_receive_queue);
448 if (answ && skb)
449 answ -= tcp_hdr(skb)->fin;
450 } else
451 answ = tp->urg_seq - tp->copied_seq;
452 release_sock(sk);
453 break;
454 case SIOCATMARK:
455 answ = tp->urg_data && tp->urg_seq == tp->copied_seq;
456 break;
457 case SIOCOUTQ:
458 if (sk->sk_state == TCP_LISTEN)
459 return -EINVAL;
460
461 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
462 answ = 0;
463 else
464 answ = tp->write_seq - tp->snd_una;
465 break;
466 default:
467 return -ENOIOCTLCMD;
468 }
469
470 return put_user(answ, (int __user *)arg);
471 }
472
473 static inline void tcp_mark_push(struct tcp_sock *tp, struct sk_buff *skb)
474 {
475 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH;
476 tp->pushed_seq = tp->write_seq;
477 }
478
479 static inline int forced_push(struct tcp_sock *tp)
480 {
481 return after(tp->write_seq, tp->pushed_seq + (tp->max_window >> 1));
482 }
483
484 static inline void skb_entail(struct sock *sk, struct sk_buff *skb)
485 {
486 struct tcp_sock *tp = tcp_sk(sk);
487 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
488
489 skb->csum = 0;
490 tcb->seq = tcb->end_seq = tp->write_seq;
491 tcb->flags = TCPCB_FLAG_ACK;
492 tcb->sacked = 0;
493 skb_header_release(skb);
494 tcp_add_write_queue_tail(sk, skb);
495 sk->sk_wmem_queued += skb->truesize;
496 sk_mem_charge(sk, skb->truesize);
497 if (tp->nonagle & TCP_NAGLE_PUSH)
498 tp->nonagle &= ~TCP_NAGLE_PUSH;
499 }
500
501 static inline void tcp_mark_urg(struct tcp_sock *tp, int flags,
502 struct sk_buff *skb)
503 {
504 if (flags & MSG_OOB)
505 tp->snd_up = tp->write_seq;
506 }
507
508 static inline void tcp_push(struct sock *sk, int flags, int mss_now,
509 int nonagle)
510 {
511 struct tcp_sock *tp = tcp_sk(sk);
512
513 if (tcp_send_head(sk)) {
514 struct sk_buff *skb = tcp_write_queue_tail(sk);
515 if (!(flags & MSG_MORE) || forced_push(tp))
516 tcp_mark_push(tp, skb);
517 tcp_mark_urg(tp, flags, skb);
518 __tcp_push_pending_frames(sk, mss_now,
519 (flags & MSG_MORE) ? TCP_NAGLE_CORK : nonagle);
520 }
521 }
522
523 static int tcp_splice_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb,
524 unsigned int offset, size_t len)
525 {
526 struct tcp_splice_state *tss = rd_desc->arg.data;
527 int ret;
528
529 ret = skb_splice_bits(skb, offset, tss->pipe, min(rd_desc->count, len),
530 tss->flags);
531 if (ret > 0)
532 rd_desc->count -= ret;
533 return ret;
534 }
535
536 static int __tcp_splice_read(struct sock *sk, struct tcp_splice_state *tss)
537 {
538 /* Store TCP splice context information in read_descriptor_t. */
539 read_descriptor_t rd_desc = {
540 .arg.data = tss,
541 .count = tss->len,
542 };
543
544 return tcp_read_sock(sk, &rd_desc, tcp_splice_data_recv);
545 }
546
547 /**
548 * tcp_splice_read - splice data from TCP socket to a pipe
549 * @sock: socket to splice from
550 * @ppos: position (not valid)
551 * @pipe: pipe to splice to
552 * @len: number of bytes to splice
553 * @flags: splice modifier flags
554 *
555 * Description:
556 * Will read pages from given socket and fill them into a pipe.
557 *
558 **/
559 ssize_t tcp_splice_read(struct socket *sock, loff_t *ppos,
560 struct pipe_inode_info *pipe, size_t len,
561 unsigned int flags)
562 {
563 struct sock *sk = sock->sk;
564 struct tcp_splice_state tss = {
565 .pipe = pipe,
566 .len = len,
567 .flags = flags,
568 };
569 long timeo;
570 ssize_t spliced;
571 int ret;
572
573 /*
574 * We can't seek on a socket input
575 */
576 if (unlikely(*ppos))
577 return -ESPIPE;
578
579 ret = spliced = 0;
580
581 lock_sock(sk);
582
583 timeo = sock_rcvtimeo(sk, flags & SPLICE_F_NONBLOCK);
584 while (tss.len) {
585 ret = __tcp_splice_read(sk, &tss);
586 if (ret < 0)
587 break;
588 else if (!ret) {
589 if (spliced)
590 break;
591 if (sock_flag(sk, SOCK_DONE))
592 break;
593 if (sk->sk_err) {
594 ret = sock_error(sk);
595 break;
596 }
597 if (sk->sk_shutdown & RCV_SHUTDOWN)
598 break;
599 if (sk->sk_state == TCP_CLOSE) {
600 /*
601 * This occurs when user tries to read
602 * from never connected socket.
603 */
604 if (!sock_flag(sk, SOCK_DONE))
605 ret = -ENOTCONN;
606 break;
607 }
608 if (!timeo) {
609 ret = -EAGAIN;
610 break;
611 }
612 sk_wait_data(sk, &timeo);
613 if (signal_pending(current)) {
614 ret = sock_intr_errno(timeo);
615 break;
616 }
617 continue;
618 }
619 tss.len -= ret;
620 spliced += ret;
621
622 if (!timeo)
623 break;
624 release_sock(sk);
625 lock_sock(sk);
626
627 if (sk->sk_err || sk->sk_state == TCP_CLOSE ||
628 (sk->sk_shutdown & RCV_SHUTDOWN) ||
629 signal_pending(current))
630 break;
631 }
632
633 release_sock(sk);
634
635 if (spliced)
636 return spliced;
637
638 return ret;
639 }
640
641 struct sk_buff *sk_stream_alloc_skb(struct sock *sk, int size, gfp_t gfp)
642 {
643 struct sk_buff *skb;
644
645 /* The TCP header must be at least 32-bit aligned. */
646 size = ALIGN(size, 4);
647
648 skb = alloc_skb_fclone(size + sk->sk_prot->max_header, gfp);
649 if (skb) {
650 if (sk_wmem_schedule(sk, skb->truesize)) {
651 /*
652 * Make sure that we have exactly size bytes
653 * available to the caller, no more, no less.
654 */
655 skb_reserve(skb, skb_tailroom(skb) - size);
656 return skb;
657 }
658 __kfree_skb(skb);
659 } else {
660 sk->sk_prot->enter_memory_pressure(sk);
661 sk_stream_moderate_sndbuf(sk);
662 }
663 return NULL;
664 }
665
666 static unsigned int tcp_xmit_size_goal(struct sock *sk, u32 mss_now,
667 int large_allowed)
668 {
669 struct tcp_sock *tp = tcp_sk(sk);
670 u32 xmit_size_goal, old_size_goal;
671
672 xmit_size_goal = mss_now;
673
674 if (large_allowed && sk_can_gso(sk)) {
675 xmit_size_goal = ((sk->sk_gso_max_size - 1) -
676 inet_csk(sk)->icsk_af_ops->net_header_len -
677 inet_csk(sk)->icsk_ext_hdr_len -
678 tp->tcp_header_len);
679
680 xmit_size_goal = tcp_bound_to_half_wnd(tp, xmit_size_goal);
681
682 /* We try hard to avoid divides here */
683 old_size_goal = tp->xmit_size_goal_segs * mss_now;
684
685 if (likely(old_size_goal <= xmit_size_goal &&
686 old_size_goal + mss_now > xmit_size_goal)) {
687 xmit_size_goal = old_size_goal;
688 } else {
689 tp->xmit_size_goal_segs = xmit_size_goal / mss_now;
690 xmit_size_goal = tp->xmit_size_goal_segs * mss_now;
691 }
692 }
693
694 return max(xmit_size_goal, mss_now);
695 }
696
697 static int tcp_send_mss(struct sock *sk, int *size_goal, int flags)
698 {
699 int mss_now;
700
701 mss_now = tcp_current_mss(sk);
702 *size_goal = tcp_xmit_size_goal(sk, mss_now, !(flags & MSG_OOB));
703
704 return mss_now;
705 }
706
707 static ssize_t do_tcp_sendpages(struct sock *sk, struct page **pages, int poffset,
708 size_t psize, int flags)
709 {
710 struct tcp_sock *tp = tcp_sk(sk);
711 int mss_now, size_goal;
712 int err;
713 ssize_t copied;
714 long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
715
716 /* Wait for a connection to finish. */
717 if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
718 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
719 goto out_err;
720
721 clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
722
723 mss_now = tcp_send_mss(sk, &size_goal, flags);
724 copied = 0;
725
726 err = -EPIPE;
727 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
728 goto out_err;
729
730 while (psize > 0) {
731 struct sk_buff *skb = tcp_write_queue_tail(sk);
732 struct page *page = pages[poffset / PAGE_SIZE];
733 int copy, i, can_coalesce;
734 int offset = poffset % PAGE_SIZE;
735 int size = min_t(size_t, psize, PAGE_SIZE - offset);
736
737 if (!tcp_send_head(sk) || (copy = size_goal - skb->len) <= 0) {
738 new_segment:
739 if (!sk_stream_memory_free(sk))
740 goto wait_for_sndbuf;
741
742 skb = sk_stream_alloc_skb(sk, 0, sk->sk_allocation);
743 if (!skb)
744 goto wait_for_memory;
745
746 skb_entail(sk, skb);
747 copy = size_goal;
748 }
749
750 if (copy > size)
751 copy = size;
752
753 i = skb_shinfo(skb)->nr_frags;
754 can_coalesce = skb_can_coalesce(skb, i, page, offset);
755 if (!can_coalesce && i >= MAX_SKB_FRAGS) {
756 tcp_mark_push(tp, skb);
757 goto new_segment;
758 }
759 if (!sk_wmem_schedule(sk, copy))
760 goto wait_for_memory;
761
762 if (can_coalesce) {
763 skb_shinfo(skb)->frags[i - 1].size += copy;
764 } else {
765 get_page(page);
766 skb_fill_page_desc(skb, i, page, offset, copy);
767 }
768
769 skb->len += copy;
770 skb->data_len += copy;
771 skb->truesize += copy;
772 sk->sk_wmem_queued += copy;
773 sk_mem_charge(sk, copy);
774 skb->ip_summed = CHECKSUM_PARTIAL;
775 tp->write_seq += copy;
776 TCP_SKB_CB(skb)->end_seq += copy;
777 skb_shinfo(skb)->gso_segs = 0;
778
779 if (!copied)
780 TCP_SKB_CB(skb)->flags &= ~TCPCB_FLAG_PSH;
781
782 copied += copy;
783 poffset += copy;
784 if (!(psize -= copy))
785 goto out;
786
787 if (skb->len < size_goal || (flags & MSG_OOB))
788 continue;
789
790 if (forced_push(tp)) {
791 tcp_mark_push(tp, skb);
792 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
793 } else if (skb == tcp_send_head(sk))
794 tcp_push_one(sk, mss_now);
795 continue;
796
797 wait_for_sndbuf:
798 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
799 wait_for_memory:
800 if (copied)
801 tcp_push(sk, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH);
802
803 if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
804 goto do_error;
805
806 mss_now = tcp_send_mss(sk, &size_goal, flags);
807 }
808
809 out:
810 if (copied)
811 tcp_push(sk, flags, mss_now, tp->nonagle);
812 return copied;
813
814 do_error:
815 if (copied)
816 goto out;
817 out_err:
818 return sk_stream_error(sk, flags, err);
819 }
820
821 ssize_t tcp_sendpage(struct socket *sock, struct page *page, int offset,
822 size_t size, int flags)
823 {
824 ssize_t res;
825 struct sock *sk = sock->sk;
826
827 if (!(sk->sk_route_caps & NETIF_F_SG) ||
828 !(sk->sk_route_caps & NETIF_F_ALL_CSUM))
829 return sock_no_sendpage(sock, page, offset, size, flags);
830
831 lock_sock(sk);
832 TCP_CHECK_TIMER(sk);
833 res = do_tcp_sendpages(sk, &page, offset, size, flags);
834 TCP_CHECK_TIMER(sk);
835 release_sock(sk);
836 return res;
837 }
838
839 #define TCP_PAGE(sk) (sk->sk_sndmsg_page)
840 #define TCP_OFF(sk) (sk->sk_sndmsg_off)
841
842 static inline int select_size(struct sock *sk)
843 {
844 struct tcp_sock *tp = tcp_sk(sk);
845 int tmp = tp->mss_cache;
846
847 if (sk->sk_route_caps & NETIF_F_SG) {
848 if (sk_can_gso(sk))
849 tmp = 0;
850 else {
851 int pgbreak = SKB_MAX_HEAD(MAX_TCP_HEADER);
852
853 if (tmp >= pgbreak &&
854 tmp <= pgbreak + (MAX_SKB_FRAGS - 1) * PAGE_SIZE)
855 tmp = pgbreak;
856 }
857 }
858
859 return tmp;
860 }
861
862 int tcp_sendmsg(struct kiocb *iocb, struct socket *sock, struct msghdr *msg,
863 size_t size)
864 {
865 struct sock *sk = sock->sk;
866 struct iovec *iov;
867 struct tcp_sock *tp = tcp_sk(sk);
868 struct sk_buff *skb;
869 int iovlen, flags;
870 int mss_now, size_goal;
871 int err, copied;
872 long timeo;
873
874 lock_sock(sk);
875 TCP_CHECK_TIMER(sk);
876
877 flags = msg->msg_flags;
878 timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
879
880 /* Wait for a connection to finish. */
881 if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
882 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
883 goto out_err;
884
885 /* This should be in poll */
886 clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
887
888 mss_now = tcp_send_mss(sk, &size_goal, flags);
889
890 /* Ok commence sending. */
891 iovlen = msg->msg_iovlen;
892 iov = msg->msg_iov;
893 copied = 0;
894
895 err = -EPIPE;
896 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
897 goto out_err;
898
899 while (--iovlen >= 0) {
900 int seglen = iov->iov_len;
901 unsigned char __user *from = iov->iov_base;
902
903 iov++;
904
905 while (seglen > 0) {
906 int copy;
907
908 skb = tcp_write_queue_tail(sk);
909
910 if (!tcp_send_head(sk) ||
911 (copy = size_goal - skb->len) <= 0) {
912
913 new_segment:
914 /* Allocate new segment. If the interface is SG,
915 * allocate skb fitting to single page.
916 */
917 if (!sk_stream_memory_free(sk))
918 goto wait_for_sndbuf;
919
920 skb = sk_stream_alloc_skb(sk, select_size(sk),
921 sk->sk_allocation);
922 if (!skb)
923 goto wait_for_memory;
924
925 /*
926 * Check whether we can use HW checksum.
927 */
928 if (sk->sk_route_caps & NETIF_F_ALL_CSUM)
929 skb->ip_summed = CHECKSUM_PARTIAL;
930
931 skb_entail(sk, skb);
932 copy = size_goal;
933 }
934
935 /* Try to append data to the end of skb. */
936 if (copy > seglen)
937 copy = seglen;
938
939 /* Where to copy to? */
940 if (skb_tailroom(skb) > 0) {
941 /* We have some space in skb head. Superb! */
942 if (copy > skb_tailroom(skb))
943 copy = skb_tailroom(skb);
944 if ((err = skb_add_data(skb, from, copy)) != 0)
945 goto do_fault;
946 } else {
947 int merge = 0;
948 int i = skb_shinfo(skb)->nr_frags;
949 struct page *page = TCP_PAGE(sk);
950 int off = TCP_OFF(sk);
951
952 if (skb_can_coalesce(skb, i, page, off) &&
953 off != PAGE_SIZE) {
954 /* We can extend the last page
955 * fragment. */
956 merge = 1;
957 } else if (i == MAX_SKB_FRAGS ||
958 (!i &&
959 !(sk->sk_route_caps & NETIF_F_SG))) {
960 /* Need to add new fragment and cannot
961 * do this because interface is non-SG,
962 * or because all the page slots are
963 * busy. */
964 tcp_mark_push(tp, skb);
965 goto new_segment;
966 } else if (page) {
967 if (off == PAGE_SIZE) {
968 put_page(page);
969 TCP_PAGE(sk) = page = NULL;
970 off = 0;
971 }
972 } else
973 off = 0;
974
975 if (copy > PAGE_SIZE - off)
976 copy = PAGE_SIZE - off;
977
978 if (!sk_wmem_schedule(sk, copy))
979 goto wait_for_memory;
980
981 if (!page) {
982 /* Allocate new cache page. */
983 if (!(page = sk_stream_alloc_page(sk)))
984 goto wait_for_memory;
985 }
986
987 /* Time to copy data. We are close to
988 * the end! */
989 err = skb_copy_to_page(sk, from, skb, page,
990 off, copy);
991 if (err) {
992 /* If this page was new, give it to the
993 * socket so it does not get leaked.
994 */
995 if (!TCP_PAGE(sk)) {
996 TCP_PAGE(sk) = page;
997 TCP_OFF(sk) = 0;
998 }
999 goto do_error;
1000 }
1001
1002 /* Update the skb. */
1003 if (merge) {
1004 skb_shinfo(skb)->frags[i - 1].size +=
1005 copy;
1006 } else {
1007 skb_fill_page_desc(skb, i, page, off, copy);
1008 if (TCP_PAGE(sk)) {
1009 get_page(page);
1010 } else if (off + copy < PAGE_SIZE) {
1011 get_page(page);
1012 TCP_PAGE(sk) = page;
1013 }
1014 }
1015
1016 TCP_OFF(sk) = off + copy;
1017 }
1018
1019 if (!copied)
1020 TCP_SKB_CB(skb)->flags &= ~TCPCB_FLAG_PSH;
1021
1022 tp->write_seq += copy;
1023 TCP_SKB_CB(skb)->end_seq += copy;
1024 skb_shinfo(skb)->gso_segs = 0;
1025
1026 from += copy;
1027 copied += copy;
1028 if ((seglen -= copy) == 0 && iovlen == 0)
1029 goto out;
1030
1031 if (skb->len < size_goal || (flags & MSG_OOB))
1032 continue;
1033
1034 if (forced_push(tp)) {
1035 tcp_mark_push(tp, skb);
1036 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
1037 } else if (skb == tcp_send_head(sk))
1038 tcp_push_one(sk, mss_now);
1039 continue;
1040
1041 wait_for_sndbuf:
1042 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1043 wait_for_memory:
1044 if (copied)
1045 tcp_push(sk, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH);
1046
1047 if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
1048 goto do_error;
1049
1050 mss_now = tcp_send_mss(sk, &size_goal, flags);
1051 }
1052 }
1053
1054 out:
1055 if (copied)
1056 tcp_push(sk, flags, mss_now, tp->nonagle);
1057 TCP_CHECK_TIMER(sk);
1058 release_sock(sk);
1059 return copied;
1060
1061 do_fault:
1062 if (!skb->len) {
1063 tcp_unlink_write_queue(skb, sk);
1064 /* It is the one place in all of TCP, except connection
1065 * reset, where we can be unlinking the send_head.
1066 */
1067 tcp_check_send_head(sk, skb);
1068 sk_wmem_free_skb(sk, skb);
1069 }
1070
1071 do_error:
1072 if (copied)
1073 goto out;
1074 out_err:
1075 err = sk_stream_error(sk, flags, err);
1076 TCP_CHECK_TIMER(sk);
1077 release_sock(sk);
1078 return err;
1079 }
1080
1081 /*
1082 * Handle reading urgent data. BSD has very simple semantics for
1083 * this, no blocking and very strange errors 8)
1084 */
1085
1086 static int tcp_recv_urg(struct sock *sk, struct msghdr *msg, int len, int flags)
1087 {
1088 struct tcp_sock *tp = tcp_sk(sk);
1089
1090 /* No URG data to read. */
1091 if (sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data ||
1092 tp->urg_data == TCP_URG_READ)
1093 return -EINVAL; /* Yes this is right ! */
1094
1095 if (sk->sk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DONE))
1096 return -ENOTCONN;
1097
1098 if (tp->urg_data & TCP_URG_VALID) {
1099 int err = 0;
1100 char c = tp->urg_data;
1101
1102 if (!(flags & MSG_PEEK))
1103 tp->urg_data = TCP_URG_READ;
1104
1105 /* Read urgent data. */
1106 msg->msg_flags |= MSG_OOB;
1107
1108 if (len > 0) {
1109 if (!(flags & MSG_TRUNC))
1110 err = memcpy_toiovec(msg->msg_iov, &c, 1);
1111 len = 1;
1112 } else
1113 msg->msg_flags |= MSG_TRUNC;
1114
1115 return err ? -EFAULT : len;
1116 }
1117
1118 if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
1119 return 0;
1120
1121 /* Fixed the recv(..., MSG_OOB) behaviour. BSD docs and
1122 * the available implementations agree in this case:
1123 * this call should never block, independent of the
1124 * blocking state of the socket.
1125 * Mike <pall@rz.uni-karlsruhe.de>
1126 */
1127 return -EAGAIN;
1128 }
1129
1130 /* Clean up the receive buffer for full frames taken by the user,
1131 * then send an ACK if necessary. COPIED is the number of bytes
1132 * tcp_recvmsg has given to the user so far, it speeds up the
1133 * calculation of whether or not we must ACK for the sake of
1134 * a window update.
1135 */
1136 void tcp_cleanup_rbuf(struct sock *sk, int copied)
1137 {
1138 struct tcp_sock *tp = tcp_sk(sk);
1139 int time_to_ack = 0;
1140
1141 #if TCP_DEBUG
1142 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
1143
1144 WARN_ON(skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq));
1145 #endif
1146
1147 if (inet_csk_ack_scheduled(sk)) {
1148 const struct inet_connection_sock *icsk = inet_csk(sk);
1149 /* Delayed ACKs frequently hit locked sockets during bulk
1150 * receive. */
1151 if (icsk->icsk_ack.blocked ||
1152 /* Once-per-two-segments ACK was not sent by tcp_input.c */
1153 tp->rcv_nxt - tp->rcv_wup > icsk->icsk_ack.rcv_mss ||
1154 /*
1155 * If this read emptied read buffer, we send ACK, if
1156 * connection is not bidirectional, user drained
1157 * receive buffer and there was a small segment
1158 * in queue.
1159 */
1160 (copied > 0 &&
1161 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED2) ||
1162 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED) &&
1163 !icsk->icsk_ack.pingpong)) &&
1164 !atomic_read(&sk->sk_rmem_alloc)))
1165 time_to_ack = 1;
1166 }
1167
1168 /* We send an ACK if we can now advertise a non-zero window
1169 * which has been raised "significantly".
1170 *
1171 * Even if window raised up to infinity, do not send window open ACK
1172 * in states, where we will not receive more. It is useless.
1173 */
1174 if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) {
1175 __u32 rcv_window_now = tcp_receive_window(tp);
1176
1177 /* Optimize, __tcp_select_window() is not cheap. */
1178 if (2*rcv_window_now <= tp->window_clamp) {
1179 __u32 new_window = __tcp_select_window(sk);
1180
1181 /* Send ACK now, if this read freed lots of space
1182 * in our buffer. Certainly, new_window is new window.
1183 * We can advertise it now, if it is not less than current one.
1184 * "Lots" means "at least twice" here.
1185 */
1186 if (new_window && new_window >= 2 * rcv_window_now)
1187 time_to_ack = 1;
1188 }
1189 }
1190 if (time_to_ack)
1191 tcp_send_ack(sk);
1192 }
1193
1194 static void tcp_prequeue_process(struct sock *sk)
1195 {
1196 struct sk_buff *skb;
1197 struct tcp_sock *tp = tcp_sk(sk);
1198
1199 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPPREQUEUED);
1200
1201 /* RX process wants to run with disabled BHs, though it is not
1202 * necessary */
1203 local_bh_disable();
1204 while ((skb = __skb_dequeue(&tp->ucopy.prequeue)) != NULL)
1205 sk_backlog_rcv(sk, skb);
1206 local_bh_enable();
1207
1208 /* Clear memory counter. */
1209 tp->ucopy.memory = 0;
1210 }
1211
1212 static inline struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off)
1213 {
1214 struct sk_buff *skb;
1215 u32 offset;
1216
1217 skb_queue_walk(&sk->sk_receive_queue, skb) {
1218 offset = seq - TCP_SKB_CB(skb)->seq;
1219 if (tcp_hdr(skb)->syn)
1220 offset--;
1221 if (offset < skb->len || tcp_hdr(skb)->fin) {
1222 *off = offset;
1223 return skb;
1224 }
1225 }
1226 return NULL;
1227 }
1228
1229 /*
1230 * This routine provides an alternative to tcp_recvmsg() for routines
1231 * that would like to handle copying from skbuffs directly in 'sendfile'
1232 * fashion.
1233 * Note:
1234 * - It is assumed that the socket was locked by the caller.
1235 * - The routine does not block.
1236 * - At present, there is no support for reading OOB data
1237 * or for 'peeking' the socket using this routine
1238 * (although both would be easy to implement).
1239 */
1240 int tcp_read_sock(struct sock *sk, read_descriptor_t *desc,
1241 sk_read_actor_t recv_actor)
1242 {
1243 struct sk_buff *skb;
1244 struct tcp_sock *tp = tcp_sk(sk);
1245 u32 seq = tp->copied_seq;
1246 u32 offset;
1247 int copied = 0;
1248
1249 if (sk->sk_state == TCP_LISTEN)
1250 return -ENOTCONN;
1251 while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1252 if (offset < skb->len) {
1253 int used;
1254 size_t len;
1255
1256 len = skb->len - offset;
1257 /* Stop reading if we hit a patch of urgent data */
1258 if (tp->urg_data) {
1259 u32 urg_offset = tp->urg_seq - seq;
1260 if (urg_offset < len)
1261 len = urg_offset;
1262 if (!len)
1263 break;
1264 }
1265 used = recv_actor(desc, skb, offset, len);
1266 if (used < 0) {
1267 if (!copied)
1268 copied = used;
1269 break;
1270 } else if (used <= len) {
1271 seq += used;
1272 copied += used;
1273 offset += used;
1274 }
1275 /*
1276 * If recv_actor drops the lock (e.g. TCP splice
1277 * receive) the skb pointer might be invalid when
1278 * getting here: tcp_collapse might have deleted it
1279 * while aggregating skbs from the socket queue.
1280 */
1281 skb = tcp_recv_skb(sk, seq-1, &offset);
1282 if (!skb || (offset+1 != skb->len))
1283 break;
1284 }
1285 if (tcp_hdr(skb)->fin) {
1286 sk_eat_skb(sk, skb, 0);
1287 ++seq;
1288 break;
1289 }
1290 sk_eat_skb(sk, skb, 0);
1291 if (!desc->count)
1292 break;
1293 }
1294 tp->copied_seq = seq;
1295
1296 tcp_rcv_space_adjust(sk);
1297
1298 /* Clean up data we have read: This will do ACK frames. */
1299 if (copied > 0)
1300 tcp_cleanup_rbuf(sk, copied);
1301 return copied;
1302 }
1303
1304 /*
1305 * This routine copies from a sock struct into the user buffer.
1306 *
1307 * Technical note: in 2.3 we work on _locked_ socket, so that
1308 * tricks with *seq access order and skb->users are not required.
1309 * Probably, code can be easily improved even more.
1310 */
1311
1312 int tcp_recvmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg,
1313 size_t len, int nonblock, int flags, int *addr_len)
1314 {
1315 struct tcp_sock *tp = tcp_sk(sk);
1316 int copied = 0;
1317 u32 peek_seq;
1318 u32 *seq;
1319 unsigned long used;
1320 int err;
1321 int target; /* Read at least this many bytes */
1322 long timeo;
1323 struct task_struct *user_recv = NULL;
1324 int copied_early = 0;
1325 struct sk_buff *skb;
1326 u32 urg_hole = 0;
1327
1328 lock_sock(sk);
1329
1330 TCP_CHECK_TIMER(sk);
1331
1332 err = -ENOTCONN;
1333 if (sk->sk_state == TCP_LISTEN)
1334 goto out;
1335
1336 timeo = sock_rcvtimeo(sk, nonblock);
1337
1338 /* Urgent data needs to be handled specially. */
1339 if (flags & MSG_OOB)
1340 goto recv_urg;
1341
1342 seq = &tp->copied_seq;
1343 if (flags & MSG_PEEK) {
1344 peek_seq = tp->copied_seq;
1345 seq = &peek_seq;
1346 }
1347
1348 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
1349
1350 #ifdef CONFIG_NET_DMA
1351 tp->ucopy.dma_chan = NULL;
1352 preempt_disable();
1353 skb = skb_peek_tail(&sk->sk_receive_queue);
1354 {
1355 int available = 0;
1356
1357 if (skb)
1358 available = TCP_SKB_CB(skb)->seq + skb->len - (*seq);
1359 if ((available < target) &&
1360 (len > sysctl_tcp_dma_copybreak) && !(flags & MSG_PEEK) &&
1361 !sysctl_tcp_low_latency &&
1362 dma_find_channel(DMA_MEMCPY)) {
1363 preempt_enable_no_resched();
1364 tp->ucopy.pinned_list =
1365 dma_pin_iovec_pages(msg->msg_iov, len);
1366 } else {
1367 preempt_enable_no_resched();
1368 }
1369 }
1370 #endif
1371
1372 do {
1373 u32 offset;
1374
1375 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
1376 if (tp->urg_data && tp->urg_seq == *seq) {
1377 if (copied)
1378 break;
1379 if (signal_pending(current)) {
1380 copied = timeo ? sock_intr_errno(timeo) : -EAGAIN;
1381 break;
1382 }
1383 }
1384
1385 /* Next get a buffer. */
1386
1387 skb_queue_walk(&sk->sk_receive_queue, skb) {
1388 /* Now that we have two receive queues this
1389 * shouldn't happen.
1390 */
1391 if (before(*seq, TCP_SKB_CB(skb)->seq)) {
1392 printk(KERN_INFO "recvmsg bug: copied %X "
1393 "seq %X\n", *seq, TCP_SKB_CB(skb)->seq);
1394 break;
1395 }
1396 offset = *seq - TCP_SKB_CB(skb)->seq;
1397 if (tcp_hdr(skb)->syn)
1398 offset--;
1399 if (offset < skb->len)
1400 goto found_ok_skb;
1401 if (tcp_hdr(skb)->fin)
1402 goto found_fin_ok;
1403 WARN_ON(!(flags & MSG_PEEK));
1404 }
1405
1406 /* Well, if we have backlog, try to process it now yet. */
1407
1408 if (copied >= target && !sk->sk_backlog.tail)
1409 break;
1410
1411 if (copied) {
1412 if (sk->sk_err ||
1413 sk->sk_state == TCP_CLOSE ||
1414 (sk->sk_shutdown & RCV_SHUTDOWN) ||
1415 !timeo ||
1416 signal_pending(current))
1417 break;
1418 } else {
1419 if (sock_flag(sk, SOCK_DONE))
1420 break;
1421
1422 if (sk->sk_err) {
1423 copied = sock_error(sk);
1424 break;
1425 }
1426
1427 if (sk->sk_shutdown & RCV_SHUTDOWN)
1428 break;
1429
1430 if (sk->sk_state == TCP_CLOSE) {
1431 if (!sock_flag(sk, SOCK_DONE)) {
1432 /* This occurs when user tries to read
1433 * from never connected socket.
1434 */
1435 copied = -ENOTCONN;
1436 break;
1437 }
1438 break;
1439 }
1440
1441 if (!timeo) {
1442 copied = -EAGAIN;
1443 break;
1444 }
1445
1446 if (signal_pending(current)) {
1447 copied = sock_intr_errno(timeo);
1448 break;
1449 }
1450 }
1451
1452 tcp_cleanup_rbuf(sk, copied);
1453
1454 if (!sysctl_tcp_low_latency && tp->ucopy.task == user_recv) {
1455 /* Install new reader */
1456 if (!user_recv && !(flags & (MSG_TRUNC | MSG_PEEK))) {
1457 user_recv = current;
1458 tp->ucopy.task = user_recv;
1459 tp->ucopy.iov = msg->msg_iov;
1460 }
1461
1462 tp->ucopy.len = len;
1463
1464 WARN_ON(tp->copied_seq != tp->rcv_nxt &&
1465 !(flags & (MSG_PEEK | MSG_TRUNC)));
1466
1467 /* Ugly... If prequeue is not empty, we have to
1468 * process it before releasing socket, otherwise
1469 * order will be broken at second iteration.
1470 * More elegant solution is required!!!
1471 *
1472 * Look: we have the following (pseudo)queues:
1473 *
1474 * 1. packets in flight
1475 * 2. backlog
1476 * 3. prequeue
1477 * 4. receive_queue
1478 *
1479 * Each queue can be processed only if the next ones
1480 * are empty. At this point we have empty receive_queue.
1481 * But prequeue _can_ be not empty after 2nd iteration,
1482 * when we jumped to start of loop because backlog
1483 * processing added something to receive_queue.
1484 * We cannot release_sock(), because backlog contains
1485 * packets arrived _after_ prequeued ones.
1486 *
1487 * Shortly, algorithm is clear --- to process all
1488 * the queues in order. We could make it more directly,
1489 * requeueing packets from backlog to prequeue, if
1490 * is not empty. It is more elegant, but eats cycles,
1491 * unfortunately.
1492 */
1493 if (!skb_queue_empty(&tp->ucopy.prequeue))
1494 goto do_prequeue;
1495
1496 /* __ Set realtime policy in scheduler __ */
1497 }
1498
1499 if (copied >= target) {
1500 /* Do not sleep, just process backlog. */
1501 release_sock(sk);
1502 lock_sock(sk);
1503 } else
1504 sk_wait_data(sk, &timeo);
1505
1506 #ifdef CONFIG_NET_DMA
1507 tp->ucopy.wakeup = 0;
1508 #endif
1509
1510 if (user_recv) {
1511 int chunk;
1512
1513 /* __ Restore normal policy in scheduler __ */
1514
1515 if ((chunk = len - tp->ucopy.len) != 0) {
1516 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMBACKLOG, chunk);
1517 len -= chunk;
1518 copied += chunk;
1519 }
1520
1521 if (tp->rcv_nxt == tp->copied_seq &&
1522 !skb_queue_empty(&tp->ucopy.prequeue)) {
1523 do_prequeue:
1524 tcp_prequeue_process(sk);
1525
1526 if ((chunk = len - tp->ucopy.len) != 0) {
1527 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1528 len -= chunk;
1529 copied += chunk;
1530 }
1531 }
1532 }
1533 if ((flags & MSG_PEEK) &&
1534 (peek_seq - copied - urg_hole != tp->copied_seq)) {
1535 if (net_ratelimit())
1536 printk(KERN_DEBUG "TCP(%s:%d): Application bug, race in MSG_PEEK.\n",
1537 current->comm, task_pid_nr(current));
1538 peek_seq = tp->copied_seq;
1539 }
1540 continue;
1541
1542 found_ok_skb:
1543 /* Ok so how much can we use? */
1544 used = skb->len - offset;
1545 if (len < used)
1546 used = len;
1547
1548 /* Do we have urgent data here? */
1549 if (tp->urg_data) {
1550 u32 urg_offset = tp->urg_seq - *seq;
1551 if (urg_offset < used) {
1552 if (!urg_offset) {
1553 if (!sock_flag(sk, SOCK_URGINLINE)) {
1554 ++*seq;
1555 urg_hole++;
1556 offset++;
1557 used--;
1558 if (!used)
1559 goto skip_copy;
1560 }
1561 } else
1562 used = urg_offset;
1563 }
1564 }
1565
1566 if (!(flags & MSG_TRUNC)) {
1567 #ifdef CONFIG_NET_DMA
1568 if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list)
1569 tp->ucopy.dma_chan = dma_find_channel(DMA_MEMCPY);
1570
1571 if (tp->ucopy.dma_chan) {
1572 tp->ucopy.dma_cookie = dma_skb_copy_datagram_iovec(
1573 tp->ucopy.dma_chan, skb, offset,
1574 msg->msg_iov, used,
1575 tp->ucopy.pinned_list);
1576
1577 if (tp->ucopy.dma_cookie < 0) {
1578
1579 printk(KERN_ALERT "dma_cookie < 0\n");
1580
1581 /* Exception. Bailout! */
1582 if (!copied)
1583 copied = -EFAULT;
1584 break;
1585 }
1586 if ((offset + used) == skb->len)
1587 copied_early = 1;
1588
1589 } else
1590 #endif
1591 {
1592 err = skb_copy_datagram_iovec(skb, offset,
1593 msg->msg_iov, used);
1594 if (err) {
1595 /* Exception. Bailout! */
1596 if (!copied)
1597 copied = -EFAULT;
1598 break;
1599 }
1600 }
1601 }
1602
1603 *seq += used;
1604 copied += used;
1605 len -= used;
1606
1607 tcp_rcv_space_adjust(sk);
1608
1609 skip_copy:
1610 if (tp->urg_data && after(tp->copied_seq, tp->urg_seq)) {
1611 tp->urg_data = 0;
1612 tcp_fast_path_check(sk);
1613 }
1614 if (used + offset < skb->len)
1615 continue;
1616
1617 if (tcp_hdr(skb)->fin)
1618 goto found_fin_ok;
1619 if (!(flags & MSG_PEEK)) {
1620 sk_eat_skb(sk, skb, copied_early);
1621 copied_early = 0;
1622 }
1623 continue;
1624
1625 found_fin_ok:
1626 /* Process the FIN. */
1627 ++*seq;
1628 if (!(flags & MSG_PEEK)) {
1629 sk_eat_skb(sk, skb, copied_early);
1630 copied_early = 0;
1631 }
1632 break;
1633 } while (len > 0);
1634
1635 if (user_recv) {
1636 if (!skb_queue_empty(&tp->ucopy.prequeue)) {
1637 int chunk;
1638
1639 tp->ucopy.len = copied > 0 ? len : 0;
1640
1641 tcp_prequeue_process(sk);
1642
1643 if (copied > 0 && (chunk = len - tp->ucopy.len) != 0) {
1644 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1645 len -= chunk;
1646 copied += chunk;
1647 }
1648 }
1649
1650 tp->ucopy.task = NULL;
1651 tp->ucopy.len = 0;
1652 }
1653
1654 #ifdef CONFIG_NET_DMA
1655 if (tp->ucopy.dma_chan) {
1656 dma_cookie_t done, used;
1657
1658 dma_async_memcpy_issue_pending(tp->ucopy.dma_chan);
1659
1660 while (dma_async_memcpy_complete(tp->ucopy.dma_chan,
1661 tp->ucopy.dma_cookie, &done,
1662 &used) == DMA_IN_PROGRESS) {
1663 /* do partial cleanup of sk_async_wait_queue */
1664 while ((skb = skb_peek(&sk->sk_async_wait_queue)) &&
1665 (dma_async_is_complete(skb->dma_cookie, done,
1666 used) == DMA_SUCCESS)) {
1667 __skb_dequeue(&sk->sk_async_wait_queue);
1668 kfree_skb(skb);
1669 }
1670 }
1671
1672 /* Safe to free early-copied skbs now */
1673 __skb_queue_purge(&sk->sk_async_wait_queue);
1674 tp->ucopy.dma_chan = NULL;
1675 }
1676 if (tp->ucopy.pinned_list) {
1677 dma_unpin_iovec_pages(tp->ucopy.pinned_list);
1678 tp->ucopy.pinned_list = NULL;
1679 }
1680 #endif
1681
1682 /* According to UNIX98, msg_name/msg_namelen are ignored
1683 * on connected socket. I was just happy when found this 8) --ANK
1684 */
1685
1686 /* Clean up data we have read: This will do ACK frames. */
1687 tcp_cleanup_rbuf(sk, copied);
1688
1689 TCP_CHECK_TIMER(sk);
1690 release_sock(sk);
1691 return copied;
1692
1693 out:
1694 TCP_CHECK_TIMER(sk);
1695 release_sock(sk);
1696 return err;
1697
1698 recv_urg:
1699 err = tcp_recv_urg(sk, msg, len, flags);
1700 goto out;
1701 }
1702
1703 void tcp_set_state(struct sock *sk, int state)
1704 {
1705 int oldstate = sk->sk_state;
1706
1707 switch (state) {
1708 case TCP_ESTABLISHED:
1709 if (oldstate != TCP_ESTABLISHED)
1710 TCP_INC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
1711 break;
1712
1713 case TCP_CLOSE:
1714 if (oldstate == TCP_CLOSE_WAIT || oldstate == TCP_ESTABLISHED)
1715 TCP_INC_STATS(sock_net(sk), TCP_MIB_ESTABRESETS);
1716
1717 sk->sk_prot->unhash(sk);
1718 if (inet_csk(sk)->icsk_bind_hash &&
1719 !(sk->sk_userlocks & SOCK_BINDPORT_LOCK))
1720 inet_put_port(sk);
1721 /* fall through */
1722 default:
1723 if (oldstate == TCP_ESTABLISHED)
1724 TCP_DEC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
1725 }
1726
1727 /* Change state AFTER socket is unhashed to avoid closed
1728 * socket sitting in hash tables.
1729 */
1730 sk->sk_state = state;
1731
1732 #ifdef STATE_TRACE
1733 SOCK_DEBUG(sk, "TCP sk=%p, State %s -> %s\n", sk, statename[oldstate], statename[state]);
1734 #endif
1735 }
1736 EXPORT_SYMBOL_GPL(tcp_set_state);
1737
1738 /*
1739 * State processing on a close. This implements the state shift for
1740 * sending our FIN frame. Note that we only send a FIN for some
1741 * states. A shutdown() may have already sent the FIN, or we may be
1742 * closed.
1743 */
1744
1745 static const unsigned char new_state[16] = {
1746 /* current state: new state: action: */
1747 /* (Invalid) */ TCP_CLOSE,
1748 /* TCP_ESTABLISHED */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1749 /* TCP_SYN_SENT */ TCP_CLOSE,
1750 /* TCP_SYN_RECV */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1751 /* TCP_FIN_WAIT1 */ TCP_FIN_WAIT1,
1752 /* TCP_FIN_WAIT2 */ TCP_FIN_WAIT2,
1753 /* TCP_TIME_WAIT */ TCP_CLOSE,
1754 /* TCP_CLOSE */ TCP_CLOSE,
1755 /* TCP_CLOSE_WAIT */ TCP_LAST_ACK | TCP_ACTION_FIN,
1756 /* TCP_LAST_ACK */ TCP_LAST_ACK,
1757 /* TCP_LISTEN */ TCP_CLOSE,
1758 /* TCP_CLOSING */ TCP_CLOSING,
1759 };
1760
1761 static int tcp_close_state(struct sock *sk)
1762 {
1763 int next = (int)new_state[sk->sk_state];
1764 int ns = next & TCP_STATE_MASK;
1765
1766 tcp_set_state(sk, ns);
1767
1768 return next & TCP_ACTION_FIN;
1769 }
1770
1771 /*
1772 * Shutdown the sending side of a connection. Much like close except
1773 * that we don't receive shut down or sock_set_flag(sk, SOCK_DEAD).
1774 */
1775
1776 void tcp_shutdown(struct sock *sk, int how)
1777 {
1778 /* We need to grab some memory, and put together a FIN,
1779 * and then put it into the queue to be sent.
1780 * Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
1781 */
1782 if (!(how & SEND_SHUTDOWN))
1783 return;
1784
1785 /* If we've already sent a FIN, or it's a closed state, skip this. */
1786 if ((1 << sk->sk_state) &
1787 (TCPF_ESTABLISHED | TCPF_SYN_SENT |
1788 TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) {
1789 /* Clear out any half completed packets. FIN if needed. */
1790 if (tcp_close_state(sk))
1791 tcp_send_fin(sk);
1792 }
1793 }
1794
1795 void tcp_close(struct sock *sk, long timeout)
1796 {
1797 struct sk_buff *skb;
1798 int data_was_unread = 0;
1799 int state;
1800
1801 lock_sock(sk);
1802 sk->sk_shutdown = SHUTDOWN_MASK;
1803
1804 if (sk->sk_state == TCP_LISTEN) {
1805 tcp_set_state(sk, TCP_CLOSE);
1806
1807 /* Special case. */
1808 inet_csk_listen_stop(sk);
1809
1810 goto adjudge_to_death;
1811 }
1812
1813 /* We need to flush the recv. buffs. We do this only on the
1814 * descriptor close, not protocol-sourced closes, because the
1815 * reader process may not have drained the data yet!
1816 */
1817 while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) {
1818 u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq -
1819 tcp_hdr(skb)->fin;
1820 data_was_unread += len;
1821 __kfree_skb(skb);
1822 }
1823
1824 sk_mem_reclaim(sk);
1825
1826 /* As outlined in RFC 2525, section 2.17, we send a RST here because
1827 * data was lost. To witness the awful effects of the old behavior of
1828 * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk
1829 * GET in an FTP client, suspend the process, wait for the client to
1830 * advertise a zero window, then kill -9 the FTP client, wheee...
1831 * Note: timeout is always zero in such a case.
1832 */
1833 if (data_was_unread) {
1834 /* Unread data was tossed, zap the connection. */
1835 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONCLOSE);
1836 tcp_set_state(sk, TCP_CLOSE);
1837 tcp_send_active_reset(sk, GFP_KERNEL);
1838 } else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) {
1839 /* Check zero linger _after_ checking for unread data. */
1840 sk->sk_prot->disconnect(sk, 0);
1841 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
1842 } else if (tcp_close_state(sk)) {
1843 /* We FIN if the application ate all the data before
1844 * zapping the connection.
1845 */
1846
1847 /* RED-PEN. Formally speaking, we have broken TCP state
1848 * machine. State transitions:
1849 *
1850 * TCP_ESTABLISHED -> TCP_FIN_WAIT1
1851 * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible)
1852 * TCP_CLOSE_WAIT -> TCP_LAST_ACK
1853 *
1854 * are legal only when FIN has been sent (i.e. in window),
1855 * rather than queued out of window. Purists blame.
1856 *
1857 * F.e. "RFC state" is ESTABLISHED,
1858 * if Linux state is FIN-WAIT-1, but FIN is still not sent.
1859 *
1860 * The visible declinations are that sometimes
1861 * we enter time-wait state, when it is not required really
1862 * (harmless), do not send active resets, when they are
1863 * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
1864 * they look as CLOSING or LAST_ACK for Linux)
1865 * Probably, I missed some more holelets.
1866 * --ANK
1867 */
1868 tcp_send_fin(sk);
1869 }
1870
1871 sk_stream_wait_close(sk, timeout);
1872
1873 adjudge_to_death:
1874 state = sk->sk_state;
1875 sock_hold(sk);
1876 sock_orphan(sk);
1877
1878 /* It is the last release_sock in its life. It will remove backlog. */
1879 release_sock(sk);
1880
1881
1882 /* Now socket is owned by kernel and we acquire BH lock
1883 to finish close. No need to check for user refs.
1884 */
1885 local_bh_disable();
1886 bh_lock_sock(sk);
1887 WARN_ON(sock_owned_by_user(sk));
1888
1889 percpu_counter_inc(sk->sk_prot->orphan_count);
1890
1891 /* Have we already been destroyed by a softirq or backlog? */
1892 if (state != TCP_CLOSE && sk->sk_state == TCP_CLOSE)
1893 goto out;
1894
1895 /* This is a (useful) BSD violating of the RFC. There is a
1896 * problem with TCP as specified in that the other end could
1897 * keep a socket open forever with no application left this end.
1898 * We use a 3 minute timeout (about the same as BSD) then kill
1899 * our end. If they send after that then tough - BUT: long enough
1900 * that we won't make the old 4*rto = almost no time - whoops
1901 * reset mistake.
1902 *
1903 * Nope, it was not mistake. It is really desired behaviour
1904 * f.e. on http servers, when such sockets are useless, but
1905 * consume significant resources. Let's do it with special
1906 * linger2 option. --ANK
1907 */
1908
1909 if (sk->sk_state == TCP_FIN_WAIT2) {
1910 struct tcp_sock *tp = tcp_sk(sk);
1911 if (tp->linger2 < 0) {
1912 tcp_set_state(sk, TCP_CLOSE);
1913 tcp_send_active_reset(sk, GFP_ATOMIC);
1914 NET_INC_STATS_BH(sock_net(sk),
1915 LINUX_MIB_TCPABORTONLINGER);
1916 } else {
1917 const int tmo = tcp_fin_time(sk);
1918
1919 if (tmo > TCP_TIMEWAIT_LEN) {
1920 inet_csk_reset_keepalive_timer(sk,
1921 tmo - TCP_TIMEWAIT_LEN);
1922 } else {
1923 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
1924 goto out;
1925 }
1926 }
1927 }
1928 if (sk->sk_state != TCP_CLOSE) {
1929 int orphan_count = percpu_counter_read_positive(
1930 sk->sk_prot->orphan_count);
1931
1932 sk_mem_reclaim(sk);
1933 if (tcp_too_many_orphans(sk, orphan_count)) {
1934 if (net_ratelimit())
1935 printk(KERN_INFO "TCP: too many of orphaned "
1936 "sockets\n");
1937 tcp_set_state(sk, TCP_CLOSE);
1938 tcp_send_active_reset(sk, GFP_ATOMIC);
1939 NET_INC_STATS_BH(sock_net(sk),
1940 LINUX_MIB_TCPABORTONMEMORY);
1941 }
1942 }
1943
1944 if (sk->sk_state == TCP_CLOSE)
1945 inet_csk_destroy_sock(sk);
1946 /* Otherwise, socket is reprieved until protocol close. */
1947
1948 out:
1949 bh_unlock_sock(sk);
1950 local_bh_enable();
1951 sock_put(sk);
1952 }
1953
1954 /* These states need RST on ABORT according to RFC793 */
1955
1956 static inline int tcp_need_reset(int state)
1957 {
1958 return (1 << state) &
1959 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 |
1960 TCPF_FIN_WAIT2 | TCPF_SYN_RECV);
1961 }
1962
1963 int tcp_disconnect(struct sock *sk, int flags)
1964 {
1965 struct inet_sock *inet = inet_sk(sk);
1966 struct inet_connection_sock *icsk = inet_csk(sk);
1967 struct tcp_sock *tp = tcp_sk(sk);
1968 int err = 0;
1969 int old_state = sk->sk_state;
1970
1971 if (old_state != TCP_CLOSE)
1972 tcp_set_state(sk, TCP_CLOSE);
1973
1974 /* ABORT function of RFC793 */
1975 if (old_state == TCP_LISTEN) {
1976 inet_csk_listen_stop(sk);
1977 } else if (tcp_need_reset(old_state) ||
1978 (tp->snd_nxt != tp->write_seq &&
1979 (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK))) {
1980 /* The last check adjusts for discrepancy of Linux wrt. RFC
1981 * states
1982 */
1983 tcp_send_active_reset(sk, gfp_any());
1984 sk->sk_err = ECONNRESET;
1985 } else if (old_state == TCP_SYN_SENT)
1986 sk->sk_err = ECONNRESET;
1987
1988 tcp_clear_xmit_timers(sk);
1989 __skb_queue_purge(&sk->sk_receive_queue);
1990 tcp_write_queue_purge(sk);
1991 __skb_queue_purge(&tp->out_of_order_queue);
1992 #ifdef CONFIG_NET_DMA
1993 __skb_queue_purge(&sk->sk_async_wait_queue);
1994 #endif
1995
1996 inet->dport = 0;
1997
1998 if (!(sk->sk_userlocks & SOCK_BINDADDR_LOCK))
1999 inet_reset_saddr(sk);
2000
2001 sk->sk_shutdown = 0;
2002 sock_reset_flag(sk, SOCK_DONE);
2003 tp->srtt = 0;
2004 if ((tp->write_seq += tp->max_window + 2) == 0)
2005 tp->write_seq = 1;
2006 icsk->icsk_backoff = 0;
2007 tp->snd_cwnd = 2;
2008 icsk->icsk_probes_out = 0;
2009 tp->packets_out = 0;
2010 tp->snd_ssthresh = 0x7fffffff;
2011 tp->snd_cwnd_cnt = 0;
2012 tp->bytes_acked = 0;
2013 tcp_set_ca_state(sk, TCP_CA_Open);
2014 tcp_clear_retrans(tp);
2015 inet_csk_delack_init(sk);
2016 tcp_init_send_head(sk);
2017 memset(&tp->rx_opt, 0, sizeof(tp->rx_opt));
2018 __sk_dst_reset(sk);
2019
2020 WARN_ON(inet->num && !icsk->icsk_bind_hash);
2021
2022 sk->sk_error_report(sk);
2023 return err;
2024 }
2025
2026 /*
2027 * Socket option code for TCP.
2028 */
2029 static int do_tcp_setsockopt(struct sock *sk, int level,
2030 int optname, char __user *optval, int optlen)
2031 {
2032 struct tcp_sock *tp = tcp_sk(sk);
2033 struct inet_connection_sock *icsk = inet_csk(sk);
2034 int val;
2035 int err = 0;
2036
2037 /* This is a string value all the others are int's */
2038 if (optname == TCP_CONGESTION) {
2039 char name[TCP_CA_NAME_MAX];
2040
2041 if (optlen < 1)
2042 return -EINVAL;
2043
2044 val = strncpy_from_user(name, optval,
2045 min(TCP_CA_NAME_MAX-1, optlen));
2046 if (val < 0)
2047 return -EFAULT;
2048 name[val] = 0;
2049
2050 lock_sock(sk);
2051 err = tcp_set_congestion_control(sk, name);
2052 release_sock(sk);
2053 return err;
2054 }
2055
2056 if (optlen < sizeof(int))
2057 return -EINVAL;
2058
2059 if (get_user(val, (int __user *)optval))
2060 return -EFAULT;
2061
2062 lock_sock(sk);
2063
2064 switch (optname) {
2065 case TCP_MAXSEG:
2066 /* Values greater than interface MTU won't take effect. However
2067 * at the point when this call is done we typically don't yet
2068 * know which interface is going to be used */
2069 if (val < 8 || val > MAX_TCP_WINDOW) {
2070 err = -EINVAL;
2071 break;
2072 }
2073 tp->rx_opt.user_mss = val;
2074 break;
2075
2076 case TCP_NODELAY:
2077 if (val) {
2078 /* TCP_NODELAY is weaker than TCP_CORK, so that
2079 * this option on corked socket is remembered, but
2080 * it is not activated until cork is cleared.
2081 *
2082 * However, when TCP_NODELAY is set we make
2083 * an explicit push, which overrides even TCP_CORK
2084 * for currently queued segments.
2085 */
2086 tp->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH;
2087 tcp_push_pending_frames(sk);
2088 } else {
2089 tp->nonagle &= ~TCP_NAGLE_OFF;
2090 }
2091 break;
2092
2093 case TCP_CORK:
2094 /* When set indicates to always queue non-full frames.
2095 * Later the user clears this option and we transmit
2096 * any pending partial frames in the queue. This is
2097 * meant to be used alongside sendfile() to get properly
2098 * filled frames when the user (for example) must write
2099 * out headers with a write() call first and then use
2100 * sendfile to send out the data parts.
2101 *
2102 * TCP_CORK can be set together with TCP_NODELAY and it is
2103 * stronger than TCP_NODELAY.
2104 */
2105 if (val) {
2106 tp->nonagle |= TCP_NAGLE_CORK;
2107 } else {
2108 tp->nonagle &= ~TCP_NAGLE_CORK;
2109 if (tp->nonagle&TCP_NAGLE_OFF)
2110 tp->nonagle |= TCP_NAGLE_PUSH;
2111 tcp_push_pending_frames(sk);
2112 }
2113 break;
2114
2115 case TCP_KEEPIDLE:
2116 if (val < 1 || val > MAX_TCP_KEEPIDLE)
2117 err = -EINVAL;
2118 else {
2119 tp->keepalive_time = val * HZ;
2120 if (sock_flag(sk, SOCK_KEEPOPEN) &&
2121 !((1 << sk->sk_state) &
2122 (TCPF_CLOSE | TCPF_LISTEN))) {
2123 __u32 elapsed = tcp_time_stamp - tp->rcv_tstamp;
2124 if (tp->keepalive_time > elapsed)
2125 elapsed = tp->keepalive_time - elapsed;
2126 else
2127 elapsed = 0;
2128 inet_csk_reset_keepalive_timer(sk, elapsed);
2129 }
2130 }
2131 break;
2132 case TCP_KEEPINTVL:
2133 if (val < 1 || val > MAX_TCP_KEEPINTVL)
2134 err = -EINVAL;
2135 else
2136 tp->keepalive_intvl = val * HZ;
2137 break;
2138 case TCP_KEEPCNT:
2139 if (val < 1 || val > MAX_TCP_KEEPCNT)
2140 err = -EINVAL;
2141 else
2142 tp->keepalive_probes = val;
2143 break;
2144 case TCP_SYNCNT:
2145 if (val < 1 || val > MAX_TCP_SYNCNT)
2146 err = -EINVAL;
2147 else
2148 icsk->icsk_syn_retries = val;
2149 break;
2150
2151 case TCP_LINGER2:
2152 if (val < 0)
2153 tp->linger2 = -1;
2154 else if (val > sysctl_tcp_fin_timeout / HZ)
2155 tp->linger2 = 0;
2156 else
2157 tp->linger2 = val * HZ;
2158 break;
2159
2160 case TCP_DEFER_ACCEPT:
2161 icsk->icsk_accept_queue.rskq_defer_accept = 0;
2162 if (val > 0) {
2163 /* Translate value in seconds to number of
2164 * retransmits */
2165 while (icsk->icsk_accept_queue.rskq_defer_accept < 32 &&
2166 val > ((TCP_TIMEOUT_INIT / HZ) <<
2167 icsk->icsk_accept_queue.rskq_defer_accept))
2168 icsk->icsk_accept_queue.rskq_defer_accept++;
2169 icsk->icsk_accept_queue.rskq_defer_accept++;
2170 }
2171 break;
2172
2173 case TCP_WINDOW_CLAMP:
2174 if (!val) {
2175 if (sk->sk_state != TCP_CLOSE) {
2176 err = -EINVAL;
2177 break;
2178 }
2179 tp->window_clamp = 0;
2180 } else
2181 tp->window_clamp = val < SOCK_MIN_RCVBUF / 2 ?
2182 SOCK_MIN_RCVBUF / 2 : val;
2183 break;
2184
2185 case TCP_QUICKACK:
2186 if (!val) {
2187 icsk->icsk_ack.pingpong = 1;
2188 } else {
2189 icsk->icsk_ack.pingpong = 0;
2190 if ((1 << sk->sk_state) &
2191 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) &&
2192 inet_csk_ack_scheduled(sk)) {
2193 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
2194 tcp_cleanup_rbuf(sk, 1);
2195 if (!(val & 1))
2196 icsk->icsk_ack.pingpong = 1;
2197 }
2198 }
2199 break;
2200
2201 #ifdef CONFIG_TCP_MD5SIG
2202 case TCP_MD5SIG:
2203 /* Read the IP->Key mappings from userspace */
2204 err = tp->af_specific->md5_parse(sk, optval, optlen);
2205 break;
2206 #endif
2207
2208 default:
2209 err = -ENOPROTOOPT;
2210 break;
2211 }
2212
2213 release_sock(sk);
2214 return err;
2215 }
2216
2217 int tcp_setsockopt(struct sock *sk, int level, int optname, char __user *optval,
2218 int optlen)
2219 {
2220 struct inet_connection_sock *icsk = inet_csk(sk);
2221
2222 if (level != SOL_TCP)
2223 return icsk->icsk_af_ops->setsockopt(sk, level, optname,
2224 optval, optlen);
2225 return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2226 }
2227
2228 #ifdef CONFIG_COMPAT
2229 int compat_tcp_setsockopt(struct sock *sk, int level, int optname,
2230 char __user *optval, int optlen)
2231 {
2232 if (level != SOL_TCP)
2233 return inet_csk_compat_setsockopt(sk, level, optname,
2234 optval, optlen);
2235 return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2236 }
2237
2238 EXPORT_SYMBOL(compat_tcp_setsockopt);
2239 #endif
2240
2241 /* Return information about state of tcp endpoint in API format. */
2242 void tcp_get_info(struct sock *sk, struct tcp_info *info)
2243 {
2244 struct tcp_sock *tp = tcp_sk(sk);
2245 const struct inet_connection_sock *icsk = inet_csk(sk);
2246 u32 now = tcp_time_stamp;
2247
2248 memset(info, 0, sizeof(*info));
2249
2250 info->tcpi_state = sk->sk_state;
2251 info->tcpi_ca_state = icsk->icsk_ca_state;
2252 info->tcpi_retransmits = icsk->icsk_retransmits;
2253 info->tcpi_probes = icsk->icsk_probes_out;
2254 info->tcpi_backoff = icsk->icsk_backoff;
2255
2256 if (tp->rx_opt.tstamp_ok)
2257 info->tcpi_options |= TCPI_OPT_TIMESTAMPS;
2258 if (tcp_is_sack(tp))
2259 info->tcpi_options |= TCPI_OPT_SACK;
2260 if (tp->rx_opt.wscale_ok) {
2261 info->tcpi_options |= TCPI_OPT_WSCALE;
2262 info->tcpi_snd_wscale = tp->rx_opt.snd_wscale;
2263 info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale;
2264 }
2265
2266 if (tp->ecn_flags&TCP_ECN_OK)
2267 info->tcpi_options |= TCPI_OPT_ECN;
2268
2269 info->tcpi_rto = jiffies_to_usecs(icsk->icsk_rto);
2270 info->tcpi_ato = jiffies_to_usecs(icsk->icsk_ack.ato);
2271 info->tcpi_snd_mss = tp->mss_cache;
2272 info->tcpi_rcv_mss = icsk->icsk_ack.rcv_mss;
2273
2274 if (sk->sk_state == TCP_LISTEN) {
2275 info->tcpi_unacked = sk->sk_ack_backlog;
2276 info->tcpi_sacked = sk->sk_max_ack_backlog;
2277 } else {
2278 info->tcpi_unacked = tp->packets_out;
2279 info->tcpi_sacked = tp->sacked_out;
2280 }
2281 info->tcpi_lost = tp->lost_out;
2282 info->tcpi_retrans = tp->retrans_out;
2283 info->tcpi_fackets = tp->fackets_out;
2284
2285 info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime);
2286 info->tcpi_last_data_recv = jiffies_to_msecs(now - icsk->icsk_ack.lrcvtime);
2287 info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp);
2288
2289 info->tcpi_pmtu = icsk->icsk_pmtu_cookie;
2290 info->tcpi_rcv_ssthresh = tp->rcv_ssthresh;
2291 info->tcpi_rtt = jiffies_to_usecs(tp->srtt)>>3;
2292 info->tcpi_rttvar = jiffies_to_usecs(tp->mdev)>>2;
2293 info->tcpi_snd_ssthresh = tp->snd_ssthresh;
2294 info->tcpi_snd_cwnd = tp->snd_cwnd;
2295 info->tcpi_advmss = tp->advmss;
2296 info->tcpi_reordering = tp->reordering;
2297
2298 info->tcpi_rcv_rtt = jiffies_to_usecs(tp->rcv_rtt_est.rtt)>>3;
2299 info->tcpi_rcv_space = tp->rcvq_space.space;
2300
2301 info->tcpi_total_retrans = tp->total_retrans;
2302 }
2303
2304 EXPORT_SYMBOL_GPL(tcp_get_info);
2305
2306 static int do_tcp_getsockopt(struct sock *sk, int level,
2307 int optname, char __user *optval, int __user *optlen)
2308 {
2309 struct inet_connection_sock *icsk = inet_csk(sk);
2310 struct tcp_sock *tp = tcp_sk(sk);
2311 int val, len;
2312
2313 if (get_user(len, optlen))
2314 return -EFAULT;
2315
2316 len = min_t(unsigned int, len, sizeof(int));
2317
2318 if (len < 0)
2319 return -EINVAL;
2320
2321 switch (optname) {
2322 case TCP_MAXSEG:
2323 val = tp->mss_cache;
2324 if (!val && ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
2325 val = tp->rx_opt.user_mss;
2326 break;
2327 case TCP_NODELAY:
2328 val = !!(tp->nonagle&TCP_NAGLE_OFF);
2329 break;
2330 case TCP_CORK:
2331 val = !!(tp->nonagle&TCP_NAGLE_CORK);
2332 break;
2333 case TCP_KEEPIDLE:
2334 val = (tp->keepalive_time ? : sysctl_tcp_keepalive_time) / HZ;
2335 break;
2336 case TCP_KEEPINTVL:
2337 val = (tp->keepalive_intvl ? : sysctl_tcp_keepalive_intvl) / HZ;
2338 break;
2339 case TCP_KEEPCNT:
2340 val = tp->keepalive_probes ? : sysctl_tcp_keepalive_probes;
2341 break;
2342 case TCP_SYNCNT:
2343 val = icsk->icsk_syn_retries ? : sysctl_tcp_syn_retries;
2344 break;
2345 case TCP_LINGER2:
2346 val = tp->linger2;
2347 if (val >= 0)
2348 val = (val ? : sysctl_tcp_fin_timeout) / HZ;
2349 break;
2350 case TCP_DEFER_ACCEPT:
2351 val = !icsk->icsk_accept_queue.rskq_defer_accept ? 0 :
2352 ((TCP_TIMEOUT_INIT / HZ) << (icsk->icsk_accept_queue.rskq_defer_accept - 1));
2353 break;
2354 case TCP_WINDOW_CLAMP:
2355 val = tp->window_clamp;
2356 break;
2357 case TCP_INFO: {
2358 struct tcp_info info;
2359
2360 if (get_user(len, optlen))
2361 return -EFAULT;
2362
2363 tcp_get_info(sk, &info);
2364
2365 len = min_t(unsigned int, len, sizeof(info));
2366 if (put_user(len, optlen))
2367 return -EFAULT;
2368 if (copy_to_user(optval, &info, len))
2369 return -EFAULT;
2370 return 0;
2371 }
2372 case TCP_QUICKACK:
2373 val = !icsk->icsk_ack.pingpong;
2374 break;
2375
2376 case TCP_CONGESTION:
2377 if (get_user(len, optlen))
2378 return -EFAULT;
2379 len = min_t(unsigned int, len, TCP_CA_NAME_MAX);
2380 if (put_user(len, optlen))
2381 return -EFAULT;
2382 if (copy_to_user(optval, icsk->icsk_ca_ops->name, len))
2383 return -EFAULT;
2384 return 0;
2385 default:
2386 return -ENOPROTOOPT;
2387 }
2388
2389 if (put_user(len, optlen))
2390 return -EFAULT;
2391 if (copy_to_user(optval, &val, len))
2392 return -EFAULT;
2393 return 0;
2394 }
2395
2396 int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval,
2397 int __user *optlen)
2398 {
2399 struct inet_connection_sock *icsk = inet_csk(sk);
2400
2401 if (level != SOL_TCP)
2402 return icsk->icsk_af_ops->getsockopt(sk, level, optname,
2403 optval, optlen);
2404 return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2405 }
2406
2407 #ifdef CONFIG_COMPAT
2408 int compat_tcp_getsockopt(struct sock *sk, int level, int optname,
2409 char __user *optval, int __user *optlen)
2410 {
2411 if (level != SOL_TCP)
2412 return inet_csk_compat_getsockopt(sk, level, optname,
2413 optval, optlen);
2414 return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2415 }
2416
2417 EXPORT_SYMBOL(compat_tcp_getsockopt);
2418 #endif
2419
2420 struct sk_buff *tcp_tso_segment(struct sk_buff *skb, int features)
2421 {
2422 struct sk_buff *segs = ERR_PTR(-EINVAL);
2423 struct tcphdr *th;
2424 unsigned thlen;
2425 unsigned int seq;
2426 __be32 delta;
2427 unsigned int oldlen;
2428 unsigned int mss;
2429
2430 if (!pskb_may_pull(skb, sizeof(*th)))
2431 goto out;
2432
2433 th = tcp_hdr(skb);
2434 thlen = th->doff * 4;
2435 if (thlen < sizeof(*th))
2436 goto out;
2437
2438 if (!pskb_may_pull(skb, thlen))
2439 goto out;
2440
2441 oldlen = (u16)~skb->len;
2442 __skb_pull(skb, thlen);
2443
2444 mss = skb_shinfo(skb)->gso_size;
2445 if (unlikely(skb->len <= mss))
2446 goto out;
2447
2448 if (skb_gso_ok(skb, features | NETIF_F_GSO_ROBUST)) {
2449 /* Packet is from an untrusted source, reset gso_segs. */
2450 int type = skb_shinfo(skb)->gso_type;
2451
2452 if (unlikely(type &
2453 ~(SKB_GSO_TCPV4 |
2454 SKB_GSO_DODGY |
2455 SKB_GSO_TCP_ECN |
2456 SKB_GSO_TCPV6 |
2457 0) ||
2458 !(type & (SKB_GSO_TCPV4 | SKB_GSO_TCPV6))))
2459 goto out;
2460
2461 skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss);
2462
2463 segs = NULL;
2464 goto out;
2465 }
2466
2467 segs = skb_segment(skb, features);
2468 if (IS_ERR(segs))
2469 goto out;
2470
2471 delta = htonl(oldlen + (thlen + mss));
2472
2473 skb = segs;
2474 th = tcp_hdr(skb);
2475 seq = ntohl(th->seq);
2476
2477 do {
2478 th->fin = th->psh = 0;
2479
2480 th->check = ~csum_fold((__force __wsum)((__force u32)th->check +
2481 (__force u32)delta));
2482 if (skb->ip_summed != CHECKSUM_PARTIAL)
2483 th->check =
2484 csum_fold(csum_partial(skb_transport_header(skb),
2485 thlen, skb->csum));
2486
2487 seq += mss;
2488 skb = skb->next;
2489 th = tcp_hdr(skb);
2490
2491 th->seq = htonl(seq);
2492 th->cwr = 0;
2493 } while (skb->next);
2494
2495 delta = htonl(oldlen + (skb->tail - skb->transport_header) +
2496 skb->data_len);
2497 th->check = ~csum_fold((__force __wsum)((__force u32)th->check +
2498 (__force u32)delta));
2499 if (skb->ip_summed != CHECKSUM_PARTIAL)
2500 th->check = csum_fold(csum_partial(skb_transport_header(skb),
2501 thlen, skb->csum));
2502
2503 out:
2504 return segs;
2505 }
2506 EXPORT_SYMBOL(tcp_tso_segment);
2507
2508 struct sk_buff **tcp_gro_receive(struct sk_buff **head, struct sk_buff *skb)
2509 {
2510 struct sk_buff **pp = NULL;
2511 struct sk_buff *p;
2512 struct tcphdr *th;
2513 struct tcphdr *th2;
2514 unsigned int len;
2515 unsigned int thlen;
2516 unsigned int flags;
2517 unsigned int mss = 1;
2518 unsigned int hlen;
2519 unsigned int off;
2520 int flush = 1;
2521 int i;
2522
2523 off = skb_gro_offset(skb);
2524 hlen = off + sizeof(*th);
2525 th = skb_gro_header_fast(skb, off);
2526 if (skb_gro_header_hard(skb, hlen)) {
2527 th = skb_gro_header_slow(skb, hlen, off);
2528 if (unlikely(!th))
2529 goto out;
2530 }
2531
2532 thlen = th->doff * 4;
2533 if (thlen < sizeof(*th))
2534 goto out;
2535
2536 hlen = off + thlen;
2537 if (skb_gro_header_hard(skb, hlen)) {
2538 th = skb_gro_header_slow(skb, hlen, off);
2539 if (unlikely(!th))
2540 goto out;
2541 }
2542
2543 skb_gro_pull(skb, thlen);
2544
2545 len = skb_gro_len(skb);
2546 flags = tcp_flag_word(th);
2547
2548 for (; (p = *head); head = &p->next) {
2549 if (!NAPI_GRO_CB(p)->same_flow)
2550 continue;
2551
2552 th2 = tcp_hdr(p);
2553
2554 if (*(u32 *)&th->source ^ *(u32 *)&th2->source) {
2555 NAPI_GRO_CB(p)->same_flow = 0;
2556 continue;
2557 }
2558
2559 goto found;
2560 }
2561
2562 goto out_check_final;
2563
2564 found:
2565 flush = NAPI_GRO_CB(p)->flush;
2566 flush |= flags & TCP_FLAG_CWR;
2567 flush |= (flags ^ tcp_flag_word(th2)) &
2568 ~(TCP_FLAG_CWR | TCP_FLAG_FIN | TCP_FLAG_PSH);
2569 flush |= th->ack_seq ^ th2->ack_seq;
2570 for (i = sizeof(*th); i < thlen; i += 4)
2571 flush |= *(u32 *)((u8 *)th + i) ^
2572 *(u32 *)((u8 *)th2 + i);
2573
2574 mss = skb_shinfo(p)->gso_size;
2575
2576 flush |= (len - 1) >= mss;
2577 flush |= (ntohl(th2->seq) + skb_gro_len(p)) ^ ntohl(th->seq);
2578
2579 if (flush || skb_gro_receive(head, skb)) {
2580 mss = 1;
2581 goto out_check_final;
2582 }
2583
2584 p = *head;
2585 th2 = tcp_hdr(p);
2586 tcp_flag_word(th2) |= flags & (TCP_FLAG_FIN | TCP_FLAG_PSH);
2587
2588 out_check_final:
2589 flush = len < mss;
2590 flush |= flags & (TCP_FLAG_URG | TCP_FLAG_PSH | TCP_FLAG_RST |
2591 TCP_FLAG_SYN | TCP_FLAG_FIN);
2592
2593 if (p && (!NAPI_GRO_CB(skb)->same_flow || flush))
2594 pp = head;
2595
2596 out:
2597 NAPI_GRO_CB(skb)->flush |= flush;
2598
2599 return pp;
2600 }
2601 EXPORT_SYMBOL(tcp_gro_receive);
2602
2603 int tcp_gro_complete(struct sk_buff *skb)
2604 {
2605 struct tcphdr *th = tcp_hdr(skb);
2606
2607 skb->csum_start = skb_transport_header(skb) - skb->head;
2608 skb->csum_offset = offsetof(struct tcphdr, check);
2609 skb->ip_summed = CHECKSUM_PARTIAL;
2610
2611 skb_shinfo(skb)->gso_segs = NAPI_GRO_CB(skb)->count;
2612
2613 if (th->cwr)
2614 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
2615
2616 return 0;
2617 }
2618 EXPORT_SYMBOL(tcp_gro_complete);
2619
2620 #ifdef CONFIG_TCP_MD5SIG
2621 static unsigned long tcp_md5sig_users;
2622 static struct tcp_md5sig_pool **tcp_md5sig_pool;
2623 static DEFINE_SPINLOCK(tcp_md5sig_pool_lock);
2624
2625 static void __tcp_free_md5sig_pool(struct tcp_md5sig_pool **pool)
2626 {
2627 int cpu;
2628 for_each_possible_cpu(cpu) {
2629 struct tcp_md5sig_pool *p = *per_cpu_ptr(pool, cpu);
2630 if (p) {
2631 if (p->md5_desc.tfm)
2632 crypto_free_hash(p->md5_desc.tfm);
2633 kfree(p);
2634 p = NULL;
2635 }
2636 }
2637 free_percpu(pool);
2638 }
2639
2640 void tcp_free_md5sig_pool(void)
2641 {
2642 struct tcp_md5sig_pool **pool = NULL;
2643
2644 spin_lock_bh(&tcp_md5sig_pool_lock);
2645 if (--tcp_md5sig_users == 0) {
2646 pool = tcp_md5sig_pool;
2647 tcp_md5sig_pool = NULL;
2648 }
2649 spin_unlock_bh(&tcp_md5sig_pool_lock);
2650 if (pool)
2651 __tcp_free_md5sig_pool(pool);
2652 }
2653
2654 EXPORT_SYMBOL(tcp_free_md5sig_pool);
2655
2656 static struct tcp_md5sig_pool **__tcp_alloc_md5sig_pool(void)
2657 {
2658 int cpu;
2659 struct tcp_md5sig_pool **pool;
2660
2661 pool = alloc_percpu(struct tcp_md5sig_pool *);
2662 if (!pool)
2663 return NULL;
2664
2665 for_each_possible_cpu(cpu) {
2666 struct tcp_md5sig_pool *p;
2667 struct crypto_hash *hash;
2668
2669 p = kzalloc(sizeof(*p), GFP_KERNEL);
2670 if (!p)
2671 goto out_free;
2672 *per_cpu_ptr(pool, cpu) = p;
2673
2674 hash = crypto_alloc_hash("md5", 0, CRYPTO_ALG_ASYNC);
2675 if (!hash || IS_ERR(hash))
2676 goto out_free;
2677
2678 p->md5_desc.tfm = hash;
2679 }
2680 return pool;
2681 out_free:
2682 __tcp_free_md5sig_pool(pool);
2683 return NULL;
2684 }
2685
2686 struct tcp_md5sig_pool **tcp_alloc_md5sig_pool(void)
2687 {
2688 struct tcp_md5sig_pool **pool;
2689 int alloc = 0;
2690
2691 retry:
2692 spin_lock_bh(&tcp_md5sig_pool_lock);
2693 pool = tcp_md5sig_pool;
2694 if (tcp_md5sig_users++ == 0) {
2695 alloc = 1;
2696 spin_unlock_bh(&tcp_md5sig_pool_lock);
2697 } else if (!pool) {
2698 tcp_md5sig_users--;
2699 spin_unlock_bh(&tcp_md5sig_pool_lock);
2700 cpu_relax();
2701 goto retry;
2702 } else
2703 spin_unlock_bh(&tcp_md5sig_pool_lock);
2704
2705 if (alloc) {
2706 /* we cannot hold spinlock here because this may sleep. */
2707 struct tcp_md5sig_pool **p = __tcp_alloc_md5sig_pool();
2708 spin_lock_bh(&tcp_md5sig_pool_lock);
2709 if (!p) {
2710 tcp_md5sig_users--;
2711 spin_unlock_bh(&tcp_md5sig_pool_lock);
2712 return NULL;
2713 }
2714 pool = tcp_md5sig_pool;
2715 if (pool) {
2716 /* oops, it has already been assigned. */
2717 spin_unlock_bh(&tcp_md5sig_pool_lock);
2718 __tcp_free_md5sig_pool(p);
2719 } else {
2720 tcp_md5sig_pool = pool = p;
2721 spin_unlock_bh(&tcp_md5sig_pool_lock);
2722 }
2723 }
2724 return pool;
2725 }
2726
2727 EXPORT_SYMBOL(tcp_alloc_md5sig_pool);
2728
2729 struct tcp_md5sig_pool *__tcp_get_md5sig_pool(int cpu)
2730 {
2731 struct tcp_md5sig_pool **p;
2732 spin_lock_bh(&tcp_md5sig_pool_lock);
2733 p = tcp_md5sig_pool;
2734 if (p)
2735 tcp_md5sig_users++;
2736 spin_unlock_bh(&tcp_md5sig_pool_lock);
2737 return (p ? *per_cpu_ptr(p, cpu) : NULL);
2738 }
2739
2740 EXPORT_SYMBOL(__tcp_get_md5sig_pool);
2741
2742 void __tcp_put_md5sig_pool(void)
2743 {
2744 tcp_free_md5sig_pool();
2745 }
2746
2747 EXPORT_SYMBOL(__tcp_put_md5sig_pool);
2748
2749 int tcp_md5_hash_header(struct tcp_md5sig_pool *hp,
2750 struct tcphdr *th)
2751 {
2752 struct scatterlist sg;
2753 int err;
2754
2755 __sum16 old_checksum = th->check;
2756 th->check = 0;
2757 /* options aren't included in the hash */
2758 sg_init_one(&sg, th, sizeof(struct tcphdr));
2759 err = crypto_hash_update(&hp->md5_desc, &sg, sizeof(struct tcphdr));
2760 th->check = old_checksum;
2761 return err;
2762 }
2763
2764 EXPORT_SYMBOL(tcp_md5_hash_header);
2765
2766 int tcp_md5_hash_skb_data(struct tcp_md5sig_pool *hp,
2767 struct sk_buff *skb, unsigned header_len)
2768 {
2769 struct scatterlist sg;
2770 const struct tcphdr *tp = tcp_hdr(skb);
2771 struct hash_desc *desc = &hp->md5_desc;
2772 unsigned i;
2773 const unsigned head_data_len = skb_headlen(skb) > header_len ?
2774 skb_headlen(skb) - header_len : 0;
2775 const struct skb_shared_info *shi = skb_shinfo(skb);
2776
2777 sg_init_table(&sg, 1);
2778
2779 sg_set_buf(&sg, ((u8 *) tp) + header_len, head_data_len);
2780 if (crypto_hash_update(desc, &sg, head_data_len))
2781 return 1;
2782
2783 for (i = 0; i < shi->nr_frags; ++i) {
2784 const struct skb_frag_struct *f = &shi->frags[i];
2785 sg_set_page(&sg, f->page, f->size, f->page_offset);
2786 if (crypto_hash_update(desc, &sg, f->size))
2787 return 1;
2788 }
2789
2790 return 0;
2791 }
2792
2793 EXPORT_SYMBOL(tcp_md5_hash_skb_data);
2794
2795 int tcp_md5_hash_key(struct tcp_md5sig_pool *hp, struct tcp_md5sig_key *key)
2796 {
2797 struct scatterlist sg;
2798
2799 sg_init_one(&sg, key->key, key->keylen);
2800 return crypto_hash_update(&hp->md5_desc, &sg, key->keylen);
2801 }
2802
2803 EXPORT_SYMBOL(tcp_md5_hash_key);
2804
2805 #endif
2806
2807 void tcp_done(struct sock *sk)
2808 {
2809 if (sk->sk_state == TCP_SYN_SENT || sk->sk_state == TCP_SYN_RECV)
2810 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_ATTEMPTFAILS);
2811
2812 tcp_set_state(sk, TCP_CLOSE);
2813 tcp_clear_xmit_timers(sk);
2814
2815 sk->sk_shutdown = SHUTDOWN_MASK;
2816
2817 if (!sock_flag(sk, SOCK_DEAD))
2818 sk->sk_state_change(sk);
2819 else
2820 inet_csk_destroy_sock(sk);
2821 }
2822 EXPORT_SYMBOL_GPL(tcp_done);
2823
2824 extern struct tcp_congestion_ops tcp_reno;
2825
2826 static __initdata unsigned long thash_entries;
2827 static int __init set_thash_entries(char *str)
2828 {
2829 if (!str)
2830 return 0;
2831 thash_entries = simple_strtoul(str, &str, 0);
2832 return 1;
2833 }
2834 __setup("thash_entries=", set_thash_entries);
2835
2836 void __init tcp_init(void)
2837 {
2838 struct sk_buff *skb = NULL;
2839 unsigned long nr_pages, limit;
2840 int order, i, max_share;
2841
2842 BUILD_BUG_ON(sizeof(struct tcp_skb_cb) > sizeof(skb->cb));
2843
2844 percpu_counter_init(&tcp_sockets_allocated, 0);
2845 percpu_counter_init(&tcp_orphan_count, 0);
2846 tcp_hashinfo.bind_bucket_cachep =
2847 kmem_cache_create("tcp_bind_bucket",
2848 sizeof(struct inet_bind_bucket), 0,
2849 SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL);
2850
2851 /* Size and allocate the main established and bind bucket
2852 * hash tables.
2853 *
2854 * The methodology is similar to that of the buffer cache.
2855 */
2856 tcp_hashinfo.ehash =
2857 alloc_large_system_hash("TCP established",
2858 sizeof(struct inet_ehash_bucket),
2859 thash_entries,
2860 (num_physpages >= 128 * 1024) ?
2861 13 : 15,
2862 0,
2863 &tcp_hashinfo.ehash_size,
2864 NULL,
2865 thash_entries ? 0 : 512 * 1024);
2866 tcp_hashinfo.ehash_size = 1 << tcp_hashinfo.ehash_size;
2867 for (i = 0; i < tcp_hashinfo.ehash_size; i++) {
2868 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].chain, i);
2869 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].twchain, i);
2870 }
2871 if (inet_ehash_locks_alloc(&tcp_hashinfo))
2872 panic("TCP: failed to alloc ehash_locks");
2873 tcp_hashinfo.bhash =
2874 alloc_large_system_hash("TCP bind",
2875 sizeof(struct inet_bind_hashbucket),
2876 tcp_hashinfo.ehash_size,
2877 (num_physpages >= 128 * 1024) ?
2878 13 : 15,
2879 0,
2880 &tcp_hashinfo.bhash_size,
2881 NULL,
2882 64 * 1024);
2883 tcp_hashinfo.bhash_size = 1 << tcp_hashinfo.bhash_size;
2884 for (i = 0; i < tcp_hashinfo.bhash_size; i++) {
2885 spin_lock_init(&tcp_hashinfo.bhash[i].lock);
2886 INIT_HLIST_HEAD(&tcp_hashinfo.bhash[i].chain);
2887 }
2888
2889 /* Try to be a bit smarter and adjust defaults depending
2890 * on available memory.
2891 */
2892 for (order = 0; ((1 << order) << PAGE_SHIFT) <
2893 (tcp_hashinfo.bhash_size * sizeof(struct inet_bind_hashbucket));
2894 order++)
2895 ;
2896 if (order >= 4) {
2897 tcp_death_row.sysctl_max_tw_buckets = 180000;
2898 sysctl_tcp_max_orphans = 4096 << (order - 4);
2899 sysctl_max_syn_backlog = 1024;
2900 } else if (order < 3) {
2901 tcp_death_row.sysctl_max_tw_buckets >>= (3 - order);
2902 sysctl_tcp_max_orphans >>= (3 - order);
2903 sysctl_max_syn_backlog = 128;
2904 }
2905
2906 /* Set the pressure threshold to be a fraction of global memory that
2907 * is up to 1/2 at 256 MB, decreasing toward zero with the amount of
2908 * memory, with a floor of 128 pages.
2909 */
2910 nr_pages = totalram_pages - totalhigh_pages;
2911 limit = min(nr_pages, 1UL<<(28-PAGE_SHIFT)) >> (20-PAGE_SHIFT);
2912 limit = (limit * (nr_pages >> (20-PAGE_SHIFT))) >> (PAGE_SHIFT-11);
2913 limit = max(limit, 128UL);
2914 sysctl_tcp_mem[0] = limit / 4 * 3;
2915 sysctl_tcp_mem[1] = limit;
2916 sysctl_tcp_mem[2] = sysctl_tcp_mem[0] * 2;
2917
2918 /* Set per-socket limits to no more than 1/128 the pressure threshold */
2919 limit = ((unsigned long)sysctl_tcp_mem[1]) << (PAGE_SHIFT - 7);
2920 max_share = min(4UL*1024*1024, limit);
2921
2922 sysctl_tcp_wmem[0] = SK_MEM_QUANTUM;
2923 sysctl_tcp_wmem[1] = 16*1024;
2924 sysctl_tcp_wmem[2] = max(64*1024, max_share);
2925
2926 sysctl_tcp_rmem[0] = SK_MEM_QUANTUM;
2927 sysctl_tcp_rmem[1] = 87380;
2928 sysctl_tcp_rmem[2] = max(87380, max_share);
2929
2930 printk(KERN_INFO "TCP: Hash tables configured "
2931 "(established %d bind %d)\n",
2932 tcp_hashinfo.ehash_size, tcp_hashinfo.bhash_size);
2933
2934 tcp_register_congestion_control(&tcp_reno);
2935 }
2936
2937 EXPORT_SYMBOL(tcp_close);
2938 EXPORT_SYMBOL(tcp_disconnect);
2939 EXPORT_SYMBOL(tcp_getsockopt);
2940 EXPORT_SYMBOL(tcp_ioctl);
2941 EXPORT_SYMBOL(tcp_poll);
2942 EXPORT_SYMBOL(tcp_read_sock);
2943 EXPORT_SYMBOL(tcp_recvmsg);
2944 EXPORT_SYMBOL(tcp_sendmsg);
2945 EXPORT_SYMBOL(tcp_splice_read);
2946 EXPORT_SYMBOL(tcp_sendpage);
2947 EXPORT_SYMBOL(tcp_setsockopt);
2948 EXPORT_SYMBOL(tcp_shutdown);
Linux kernel - Deliverables
This page took 0.1862 seconds and 5 git commands to generate.