projects / deliverable / linux.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
blame | history | raw | HEAD
Merge branch 'for_3.8-rc1' into v4l_for_linus
[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 #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
283 int sysctl_tcp_fin_timeout __read_mostly = TCP_FIN_TIMEOUT;
284
285 struct percpu_counter tcp_orphan_count;
286 EXPORT_SYMBOL_GPL(tcp_orphan_count);
287
288 int sysctl_tcp_wmem[3] __read_mostly;
289 int sysctl_tcp_rmem[3] __read_mostly;
290
291 EXPORT_SYMBOL(sysctl_tcp_rmem);
292 EXPORT_SYMBOL(sysctl_tcp_wmem);
293
294 atomic_long_t tcp_memory_allocated; /* Current allocated memory. */
295 EXPORT_SYMBOL(tcp_memory_allocated);
296
297 /*
298 * Current number of TCP sockets.
299 */
300 struct percpu_counter tcp_sockets_allocated;
301 EXPORT_SYMBOL(tcp_sockets_allocated);
302
303 /*
304 * TCP splice context
305 */
306 struct tcp_splice_state {
307 struct pipe_inode_info *pipe;
308 size_t len;
309 unsigned int flags;
310 };
311
312 /*
313 * Pressure flag: try to collapse.
314 * Technical note: it is used by multiple contexts non atomically.
315 * All the __sk_mem_schedule() is of this nature: accounting
316 * is strict, actions are advisory and have some latency.
317 */
318 int tcp_memory_pressure __read_mostly;
319 EXPORT_SYMBOL(tcp_memory_pressure);
320
321 void tcp_enter_memory_pressure(struct sock *sk)
322 {
323 if (!tcp_memory_pressure) {
324 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMEMORYPRESSURES);
325 tcp_memory_pressure = 1;
326 }
327 }
328 EXPORT_SYMBOL(tcp_enter_memory_pressure);
329
330 /* Convert seconds to retransmits based on initial and max timeout */
331 static u8 secs_to_retrans(int seconds, int timeout, int rto_max)
332 {
333 u8 res = 0;
334
335 if (seconds > 0) {
336 int period = timeout;
337
338 res = 1;
339 while (seconds > period && res < 255) {
340 res++;
341 timeout <<= 1;
342 if (timeout > rto_max)
343 timeout = rto_max;
344 period += timeout;
345 }
346 }
347 return res;
348 }
349
350 /* Convert retransmits to seconds based on initial and max timeout */
351 static int retrans_to_secs(u8 retrans, int timeout, int rto_max)
352 {
353 int period = 0;
354
355 if (retrans > 0) {
356 period = timeout;
357 while (--retrans) {
358 timeout <<= 1;
359 if (timeout > rto_max)
360 timeout = rto_max;
361 period += timeout;
362 }
363 }
364 return period;
365 }
366
367 /* Address-family independent initialization for a tcp_sock.
368 *
369 * NOTE: A lot of things set to zero explicitly by call to
370 * sk_alloc() so need not be done here.
371 */
372 void tcp_init_sock(struct sock *sk)
373 {
374 struct inet_connection_sock *icsk = inet_csk(sk);
375 struct tcp_sock *tp = tcp_sk(sk);
376
377 skb_queue_head_init(&tp->out_of_order_queue);
378 tcp_init_xmit_timers(sk);
379 tcp_prequeue_init(tp);
380 INIT_LIST_HEAD(&tp->tsq_node);
381
382 icsk->icsk_rto = TCP_TIMEOUT_INIT;
383 tp->mdev = TCP_TIMEOUT_INIT;
384
385 /* So many TCP implementations out there (incorrectly) count the
386 * initial SYN frame in their delayed-ACK and congestion control
387 * algorithms that we must have the following bandaid to talk
388 * efficiently to them. -DaveM
389 */
390 tp->snd_cwnd = TCP_INIT_CWND;
391
392 /* See draft-stevens-tcpca-spec-01 for discussion of the
393 * initialization of these values.
394 */
395 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
396 tp->snd_cwnd_clamp = ~0;
397 tp->mss_cache = TCP_MSS_DEFAULT;
398
399 tp->reordering = sysctl_tcp_reordering;
400 tcp_enable_early_retrans(tp);
401 icsk->icsk_ca_ops = &tcp_init_congestion_ops;
402
403 sk->sk_state = TCP_CLOSE;
404
405 sk->sk_write_space = sk_stream_write_space;
406 sock_set_flag(sk, SOCK_USE_WRITE_QUEUE);
407
408 icsk->icsk_sync_mss = tcp_sync_mss;
409
410 /* TCP Cookie Transactions */
411 if (sysctl_tcp_cookie_size > 0) {
412 /* Default, cookies without s_data_payload. */
413 tp->cookie_values =
414 kzalloc(sizeof(*tp->cookie_values),
415 sk->sk_allocation);
416 if (tp->cookie_values != NULL)
417 kref_init(&tp->cookie_values->kref);
418 }
419 /* Presumed zeroed, in order of appearance:
420 * cookie_in_always, cookie_out_never,
421 * s_data_constant, s_data_in, s_data_out
422 */
423 sk->sk_sndbuf = sysctl_tcp_wmem[1];
424 sk->sk_rcvbuf = sysctl_tcp_rmem[1];
425
426 local_bh_disable();
427 sock_update_memcg(sk);
428 sk_sockets_allocated_inc(sk);
429 local_bh_enable();
430 }
431 EXPORT_SYMBOL(tcp_init_sock);
432
433 /*
434 * Wait for a TCP event.
435 *
436 * Note that we don't need to lock the socket, as the upper poll layers
437 * take care of normal races (between the test and the event) and we don't
438 * go look at any of the socket buffers directly.
439 */
440 unsigned int tcp_poll(struct file *file, struct socket *sock, poll_table *wait)
441 {
442 unsigned int mask;
443 struct sock *sk = sock->sk;
444 const struct tcp_sock *tp = tcp_sk(sk);
445
446 sock_poll_wait(file, sk_sleep(sk), wait);
447 if (sk->sk_state == TCP_LISTEN)
448 return inet_csk_listen_poll(sk);
449
450 /* Socket is not locked. We are protected from async events
451 * by poll logic and correct handling of state changes
452 * made by other threads is impossible in any case.
453 */
454
455 mask = 0;
456
457 /*
458 * POLLHUP is certainly not done right. But poll() doesn't
459 * have a notion of HUP in just one direction, and for a
460 * socket the read side is more interesting.
461 *
462 * Some poll() documentation says that POLLHUP is incompatible
463 * with the POLLOUT/POLLWR flags, so somebody should check this
464 * all. But careful, it tends to be safer to return too many
465 * bits than too few, and you can easily break real applications
466 * if you don't tell them that something has hung up!
467 *
468 * Check-me.
469 *
470 * Check number 1. POLLHUP is _UNMASKABLE_ event (see UNIX98 and
471 * our fs/select.c). It means that after we received EOF,
472 * poll always returns immediately, making impossible poll() on write()
473 * in state CLOSE_WAIT. One solution is evident --- to set POLLHUP
474 * if and only if shutdown has been made in both directions.
475 * Actually, it is interesting to look how Solaris and DUX
476 * solve this dilemma. I would prefer, if POLLHUP were maskable,
477 * then we could set it on SND_SHUTDOWN. BTW examples given
478 * in Stevens' books assume exactly this behaviour, it explains
479 * why POLLHUP is incompatible with POLLOUT. --ANK
480 *
481 * NOTE. Check for TCP_CLOSE is added. The goal is to prevent
482 * blocking on fresh not-connected or disconnected socket. --ANK
483 */
484 if (sk->sk_shutdown == SHUTDOWN_MASK || sk->sk_state == TCP_CLOSE)
485 mask |= POLLHUP;
486 if (sk->sk_shutdown & RCV_SHUTDOWN)
487 mask |= POLLIN | POLLRDNORM | POLLRDHUP;
488
489 /* Connected or passive Fast Open socket? */
490 if (sk->sk_state != TCP_SYN_SENT &&
491 (sk->sk_state != TCP_SYN_RECV || tp->fastopen_rsk != NULL)) {
492 int target = sock_rcvlowat(sk, 0, INT_MAX);
493
494 if (tp->urg_seq == tp->copied_seq &&
495 !sock_flag(sk, SOCK_URGINLINE) &&
496 tp->urg_data)
497 target++;
498
499 /* Potential race condition. If read of tp below will
500 * escape above sk->sk_state, we can be illegally awaken
501 * in SYN_* states. */
502 if (tp->rcv_nxt - tp->copied_seq >= target)
503 mask |= POLLIN | POLLRDNORM;
504
505 if (!(sk->sk_shutdown & SEND_SHUTDOWN)) {
506 if (sk_stream_wspace(sk) >= sk_stream_min_wspace(sk)) {
507 mask |= POLLOUT | POLLWRNORM;
508 } else { /* send SIGIO later */
509 set_bit(SOCK_ASYNC_NOSPACE,
510 &sk->sk_socket->flags);
511 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
512
513 /* Race breaker. If space is freed after
514 * wspace test but before the flags are set,
515 * IO signal will be lost.
516 */
517 if (sk_stream_wspace(sk) >= sk_stream_min_wspace(sk))
518 mask |= POLLOUT | POLLWRNORM;
519 }
520 } else
521 mask |= POLLOUT | POLLWRNORM;
522
523 if (tp->urg_data & TCP_URG_VALID)
524 mask |= POLLPRI;
525 }
526 /* This barrier is coupled with smp_wmb() in tcp_reset() */
527 smp_rmb();
528 if (sk->sk_err)
529 mask |= POLLERR;
530
531 return mask;
532 }
533 EXPORT_SYMBOL(tcp_poll);
534
535 int tcp_ioctl(struct sock *sk, int cmd, unsigned long arg)
536 {
537 struct tcp_sock *tp = tcp_sk(sk);
538 int answ;
539
540 switch (cmd) {
541 case SIOCINQ:
542 if (sk->sk_state == TCP_LISTEN)
543 return -EINVAL;
544
545 lock_sock(sk);
546 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
547 answ = 0;
548 else if (sock_flag(sk, SOCK_URGINLINE) ||
549 !tp->urg_data ||
550 before(tp->urg_seq, tp->copied_seq) ||
551 !before(tp->urg_seq, tp->rcv_nxt)) {
552
553 answ = tp->rcv_nxt - tp->copied_seq;
554
555 /* Subtract 1, if FIN was received */
556 if (answ && sock_flag(sk, SOCK_DONE))
557 answ--;
558 } else
559 answ = tp->urg_seq - tp->copied_seq;
560 release_sock(sk);
561 break;
562 case SIOCATMARK:
563 answ = tp->urg_data && tp->urg_seq == tp->copied_seq;
564 break;
565 case SIOCOUTQ:
566 if (sk->sk_state == TCP_LISTEN)
567 return -EINVAL;
568
569 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
570 answ = 0;
571 else
572 answ = tp->write_seq - tp->snd_una;
573 break;
574 case SIOCOUTQNSD:
575 if (sk->sk_state == TCP_LISTEN)
576 return -EINVAL;
577
578 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
579 answ = 0;
580 else
581 answ = tp->write_seq - tp->snd_nxt;
582 break;
583 default:
584 return -ENOIOCTLCMD;
585 }
586
587 return put_user(answ, (int __user *)arg);
588 }
589 EXPORT_SYMBOL(tcp_ioctl);
590
591 static inline void tcp_mark_push(struct tcp_sock *tp, struct sk_buff *skb)
592 {
593 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
594 tp->pushed_seq = tp->write_seq;
595 }
596
597 static inline bool forced_push(const struct tcp_sock *tp)
598 {
599 return after(tp->write_seq, tp->pushed_seq + (tp->max_window >> 1));
600 }
601
602 static inline void skb_entail(struct sock *sk, struct sk_buff *skb)
603 {
604 struct tcp_sock *tp = tcp_sk(sk);
605 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
606
607 skb->csum = 0;
608 tcb->seq = tcb->end_seq = tp->write_seq;
609 tcb->tcp_flags = TCPHDR_ACK;
610 tcb->sacked = 0;
611 skb_header_release(skb);
612 tcp_add_write_queue_tail(sk, skb);
613 sk->sk_wmem_queued += skb->truesize;
614 sk_mem_charge(sk, skb->truesize);
615 if (tp->nonagle & TCP_NAGLE_PUSH)
616 tp->nonagle &= ~TCP_NAGLE_PUSH;
617 }
618
619 static inline void tcp_mark_urg(struct tcp_sock *tp, int flags)
620 {
621 if (flags & MSG_OOB)
622 tp->snd_up = tp->write_seq;
623 }
624
625 static inline void tcp_push(struct sock *sk, int flags, int mss_now,
626 int nonagle)
627 {
628 if (tcp_send_head(sk)) {
629 struct tcp_sock *tp = tcp_sk(sk);
630
631 if (!(flags & MSG_MORE) || forced_push(tp))
632 tcp_mark_push(tp, tcp_write_queue_tail(sk));
633
634 tcp_mark_urg(tp, flags);
635 __tcp_push_pending_frames(sk, mss_now,
636 (flags & MSG_MORE) ? TCP_NAGLE_CORK : nonagle);
637 }
638 }
639
640 static int tcp_splice_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb,
641 unsigned int offset, size_t len)
642 {
643 struct tcp_splice_state *tss = rd_desc->arg.data;
644 int ret;
645
646 ret = skb_splice_bits(skb, offset, tss->pipe, min(rd_desc->count, len),
647 tss->flags);
648 if (ret > 0)
649 rd_desc->count -= ret;
650 return ret;
651 }
652
653 static int __tcp_splice_read(struct sock *sk, struct tcp_splice_state *tss)
654 {
655 /* Store TCP splice context information in read_descriptor_t. */
656 read_descriptor_t rd_desc = {
657 .arg.data = tss,
658 .count = tss->len,
659 };
660
661 return tcp_read_sock(sk, &rd_desc, tcp_splice_data_recv);
662 }
663
664 /**
665 * tcp_splice_read - splice data from TCP socket to a pipe
666 * @sock: socket to splice from
667 * @ppos: position (not valid)
668 * @pipe: pipe to splice to
669 * @len: number of bytes to splice
670 * @flags: splice modifier flags
671 *
672 * Description:
673 * Will read pages from given socket and fill them into a pipe.
674 *
675 **/
676 ssize_t tcp_splice_read(struct socket *sock, loff_t *ppos,
677 struct pipe_inode_info *pipe, size_t len,
678 unsigned int flags)
679 {
680 struct sock *sk = sock->sk;
681 struct tcp_splice_state tss = {
682 .pipe = pipe,
683 .len = len,
684 .flags = flags,
685 };
686 long timeo;
687 ssize_t spliced;
688 int ret;
689
690 sock_rps_record_flow(sk);
691 /*
692 * We can't seek on a socket input
693 */
694 if (unlikely(*ppos))
695 return -ESPIPE;
696
697 ret = spliced = 0;
698
699 lock_sock(sk);
700
701 timeo = sock_rcvtimeo(sk, sock->file->f_flags & O_NONBLOCK);
702 while (tss.len) {
703 ret = __tcp_splice_read(sk, &tss);
704 if (ret < 0)
705 break;
706 else if (!ret) {
707 if (spliced)
708 break;
709 if (sock_flag(sk, SOCK_DONE))
710 break;
711 if (sk->sk_err) {
712 ret = sock_error(sk);
713 break;
714 }
715 if (sk->sk_shutdown & RCV_SHUTDOWN)
716 break;
717 if (sk->sk_state == TCP_CLOSE) {
718 /*
719 * This occurs when user tries to read
720 * from never connected socket.
721 */
722 if (!sock_flag(sk, SOCK_DONE))
723 ret = -ENOTCONN;
724 break;
725 }
726 if (!timeo) {
727 ret = -EAGAIN;
728 break;
729 }
730 sk_wait_data(sk, &timeo);
731 if (signal_pending(current)) {
732 ret = sock_intr_errno(timeo);
733 break;
734 }
735 continue;
736 }
737 tss.len -= ret;
738 spliced += ret;
739
740 if (!timeo)
741 break;
742 release_sock(sk);
743 lock_sock(sk);
744
745 if (sk->sk_err || sk->sk_state == TCP_CLOSE ||
746 (sk->sk_shutdown & RCV_SHUTDOWN) ||
747 signal_pending(current))
748 break;
749 }
750
751 release_sock(sk);
752
753 if (spliced)
754 return spliced;
755
756 return ret;
757 }
758 EXPORT_SYMBOL(tcp_splice_read);
759
760 struct sk_buff *sk_stream_alloc_skb(struct sock *sk, int size, gfp_t gfp)
761 {
762 struct sk_buff *skb;
763
764 /* The TCP header must be at least 32-bit aligned. */
765 size = ALIGN(size, 4);
766
767 skb = alloc_skb_fclone(size + sk->sk_prot->max_header, gfp);
768 if (skb) {
769 if (sk_wmem_schedule(sk, skb->truesize)) {
770 skb_reserve(skb, sk->sk_prot->max_header);
771 /*
772 * Make sure that we have exactly size bytes
773 * available to the caller, no more, no less.
774 */
775 skb->avail_size = size;
776 return skb;
777 }
778 __kfree_skb(skb);
779 } else {
780 sk->sk_prot->enter_memory_pressure(sk);
781 sk_stream_moderate_sndbuf(sk);
782 }
783 return NULL;
784 }
785
786 static unsigned int tcp_xmit_size_goal(struct sock *sk, u32 mss_now,
787 int large_allowed)
788 {
789 struct tcp_sock *tp = tcp_sk(sk);
790 u32 xmit_size_goal, old_size_goal;
791
792 xmit_size_goal = mss_now;
793
794 if (large_allowed && sk_can_gso(sk)) {
795 xmit_size_goal = ((sk->sk_gso_max_size - 1) -
796 inet_csk(sk)->icsk_af_ops->net_header_len -
797 inet_csk(sk)->icsk_ext_hdr_len -
798 tp->tcp_header_len);
799
800 /* TSQ : try to have two TSO segments in flight */
801 xmit_size_goal = min_t(u32, xmit_size_goal,
802 sysctl_tcp_limit_output_bytes >> 1);
803
804 xmit_size_goal = tcp_bound_to_half_wnd(tp, xmit_size_goal);
805
806 /* We try hard to avoid divides here */
807 old_size_goal = tp->xmit_size_goal_segs * mss_now;
808
809 if (likely(old_size_goal <= xmit_size_goal &&
810 old_size_goal + mss_now > xmit_size_goal)) {
811 xmit_size_goal = old_size_goal;
812 } else {
813 tp->xmit_size_goal_segs =
814 min_t(u16, xmit_size_goal / mss_now,
815 sk->sk_gso_max_segs);
816 xmit_size_goal = tp->xmit_size_goal_segs * mss_now;
817 }
818 }
819
820 return max(xmit_size_goal, mss_now);
821 }
822
823 static int tcp_send_mss(struct sock *sk, int *size_goal, int flags)
824 {
825 int mss_now;
826
827 mss_now = tcp_current_mss(sk);
828 *size_goal = tcp_xmit_size_goal(sk, mss_now, !(flags & MSG_OOB));
829
830 return mss_now;
831 }
832
833 static ssize_t do_tcp_sendpages(struct sock *sk, struct page *page, int offset,
834 size_t size, int flags)
835 {
836 struct tcp_sock *tp = tcp_sk(sk);
837 int mss_now, size_goal;
838 int err;
839 ssize_t copied;
840 long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
841
842 /* Wait for a connection to finish. One exception is TCP Fast Open
843 * (passive side) where data is allowed to be sent before a connection
844 * is fully established.
845 */
846 if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) &&
847 !tcp_passive_fastopen(sk)) {
848 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
849 goto out_err;
850 }
851
852 clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
853
854 mss_now = tcp_send_mss(sk, &size_goal, flags);
855 copied = 0;
856
857 err = -EPIPE;
858 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
859 goto out_err;
860
861 while (size > 0) {
862 struct sk_buff *skb = tcp_write_queue_tail(sk);
863 int copy, i;
864 bool can_coalesce;
865
866 if (!tcp_send_head(sk) || (copy = size_goal - skb->len) <= 0) {
867 new_segment:
868 if (!sk_stream_memory_free(sk))
869 goto wait_for_sndbuf;
870
871 skb = sk_stream_alloc_skb(sk, 0, sk->sk_allocation);
872 if (!skb)
873 goto wait_for_memory;
874
875 skb_entail(sk, skb);
876 copy = size_goal;
877 }
878
879 if (copy > size)
880 copy = size;
881
882 i = skb_shinfo(skb)->nr_frags;
883 can_coalesce = skb_can_coalesce(skb, i, page, offset);
884 if (!can_coalesce && i >= MAX_SKB_FRAGS) {
885 tcp_mark_push(tp, skb);
886 goto new_segment;
887 }
888 if (!sk_wmem_schedule(sk, copy))
889 goto wait_for_memory;
890
891 if (can_coalesce) {
892 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
893 } else {
894 get_page(page);
895 skb_fill_page_desc(skb, i, page, offset, copy);
896 }
897
898 skb->len += copy;
899 skb->data_len += copy;
900 skb->truesize += copy;
901 sk->sk_wmem_queued += copy;
902 sk_mem_charge(sk, copy);
903 skb->ip_summed = CHECKSUM_PARTIAL;
904 tp->write_seq += copy;
905 TCP_SKB_CB(skb)->end_seq += copy;
906 skb_shinfo(skb)->gso_segs = 0;
907
908 if (!copied)
909 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
910
911 copied += copy;
912 offset += copy;
913 if (!(size -= copy))
914 goto out;
915
916 if (skb->len < size_goal || (flags & MSG_OOB))
917 continue;
918
919 if (forced_push(tp)) {
920 tcp_mark_push(tp, skb);
921 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
922 } else if (skb == tcp_send_head(sk))
923 tcp_push_one(sk, mss_now);
924 continue;
925
926 wait_for_sndbuf:
927 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
928 wait_for_memory:
929 tcp_push(sk, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH);
930
931 if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
932 goto do_error;
933
934 mss_now = tcp_send_mss(sk, &size_goal, flags);
935 }
936
937 out:
938 if (copied && !(flags & MSG_SENDPAGE_NOTLAST))
939 tcp_push(sk, flags, mss_now, tp->nonagle);
940 return copied;
941
942 do_error:
943 if (copied)
944 goto out;
945 out_err:
946 return sk_stream_error(sk, flags, err);
947 }
948
949 int tcp_sendpage(struct sock *sk, struct page *page, int offset,
950 size_t size, int flags)
951 {
952 ssize_t res;
953
954 if (!(sk->sk_route_caps & NETIF_F_SG) ||
955 !(sk->sk_route_caps & NETIF_F_ALL_CSUM))
956 return sock_no_sendpage(sk->sk_socket, page, offset, size,
957 flags);
958
959 lock_sock(sk);
960 res = do_tcp_sendpages(sk, page, offset, size, flags);
961 release_sock(sk);
962 return res;
963 }
964 EXPORT_SYMBOL(tcp_sendpage);
965
966 static inline int select_size(const struct sock *sk, bool sg)
967 {
968 const struct tcp_sock *tp = tcp_sk(sk);
969 int tmp = tp->mss_cache;
970
971 if (sg) {
972 if (sk_can_gso(sk)) {
973 /* Small frames wont use a full page:
974 * Payload will immediately follow tcp header.
975 */
976 tmp = SKB_WITH_OVERHEAD(2048 - MAX_TCP_HEADER);
977 } else {
978 int pgbreak = SKB_MAX_HEAD(MAX_TCP_HEADER);
979
980 if (tmp >= pgbreak &&
981 tmp <= pgbreak + (MAX_SKB_FRAGS - 1) * PAGE_SIZE)
982 tmp = pgbreak;
983 }
984 }
985
986 return tmp;
987 }
988
989 void tcp_free_fastopen_req(struct tcp_sock *tp)
990 {
991 if (tp->fastopen_req != NULL) {
992 kfree(tp->fastopen_req);
993 tp->fastopen_req = NULL;
994 }
995 }
996
997 static int tcp_sendmsg_fastopen(struct sock *sk, struct msghdr *msg, int *size)
998 {
999 struct tcp_sock *tp = tcp_sk(sk);
1000 int err, flags;
1001
1002 if (!(sysctl_tcp_fastopen & TFO_CLIENT_ENABLE))
1003 return -EOPNOTSUPP;
1004 if (tp->fastopen_req != NULL)
1005 return -EALREADY; /* Another Fast Open is in progress */
1006
1007 tp->fastopen_req = kzalloc(sizeof(struct tcp_fastopen_request),
1008 sk->sk_allocation);
1009 if (unlikely(tp->fastopen_req == NULL))
1010 return -ENOBUFS;
1011 tp->fastopen_req->data = msg;
1012
1013 flags = (msg->msg_flags & MSG_DONTWAIT) ? O_NONBLOCK : 0;
1014 err = __inet_stream_connect(sk->sk_socket, msg->msg_name,
1015 msg->msg_namelen, flags);
1016 *size = tp->fastopen_req->copied;
1017 tcp_free_fastopen_req(tp);
1018 return err;
1019 }
1020
1021 int tcp_sendmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg,
1022 size_t size)
1023 {
1024 struct iovec *iov;
1025 struct tcp_sock *tp = tcp_sk(sk);
1026 struct sk_buff *skb;
1027 int iovlen, flags, err, copied = 0;
1028 int mss_now = 0, size_goal, copied_syn = 0, offset = 0;
1029 bool sg;
1030 long timeo;
1031
1032 lock_sock(sk);
1033
1034 flags = msg->msg_flags;
1035 if (flags & MSG_FASTOPEN) {
1036 err = tcp_sendmsg_fastopen(sk, msg, &copied_syn);
1037 if (err == -EINPROGRESS && copied_syn > 0)
1038 goto out;
1039 else if (err)
1040 goto out_err;
1041 offset = copied_syn;
1042 }
1043
1044 timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
1045
1046 /* Wait for a connection to finish. One exception is TCP Fast Open
1047 * (passive side) where data is allowed to be sent before a connection
1048 * is fully established.
1049 */
1050 if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) &&
1051 !tcp_passive_fastopen(sk)) {
1052 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
1053 goto do_error;
1054 }
1055
1056 if (unlikely(tp->repair)) {
1057 if (tp->repair_queue == TCP_RECV_QUEUE) {
1058 copied = tcp_send_rcvq(sk, msg, size);
1059 goto out;
1060 }
1061
1062 err = -EINVAL;
1063 if (tp->repair_queue == TCP_NO_QUEUE)
1064 goto out_err;
1065
1066 /* 'common' sending to sendq */
1067 }
1068
1069 /* This should be in poll */
1070 clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
1071
1072 mss_now = tcp_send_mss(sk, &size_goal, flags);
1073
1074 /* Ok commence sending. */
1075 iovlen = msg->msg_iovlen;
1076 iov = msg->msg_iov;
1077 copied = 0;
1078
1079 err = -EPIPE;
1080 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
1081 goto out_err;
1082
1083 sg = !!(sk->sk_route_caps & NETIF_F_SG);
1084
1085 while (--iovlen >= 0) {
1086 size_t seglen = iov->iov_len;
1087 unsigned char __user *from = iov->iov_base;
1088
1089 iov++;
1090 if (unlikely(offset > 0)) { /* Skip bytes copied in SYN */
1091 if (offset >= seglen) {
1092 offset -= seglen;
1093 continue;
1094 }
1095 seglen -= offset;
1096 from += offset;
1097 offset = 0;
1098 }
1099
1100 while (seglen > 0) {
1101 int copy = 0;
1102 int max = size_goal;
1103
1104 skb = tcp_write_queue_tail(sk);
1105 if (tcp_send_head(sk)) {
1106 if (skb->ip_summed == CHECKSUM_NONE)
1107 max = mss_now;
1108 copy = max - skb->len;
1109 }
1110
1111 if (copy <= 0) {
1112 new_segment:
1113 /* Allocate new segment. If the interface is SG,
1114 * allocate skb fitting to single page.
1115 */
1116 if (!sk_stream_memory_free(sk))
1117 goto wait_for_sndbuf;
1118
1119 skb = sk_stream_alloc_skb(sk,
1120 select_size(sk, sg),
1121 sk->sk_allocation);
1122 if (!skb)
1123 goto wait_for_memory;
1124
1125 /*
1126 * Check whether we can use HW checksum.
1127 */
1128 if (sk->sk_route_caps & NETIF_F_ALL_CSUM)
1129 skb->ip_summed = CHECKSUM_PARTIAL;
1130
1131 skb_entail(sk, skb);
1132 copy = size_goal;
1133 max = size_goal;
1134 }
1135
1136 /* Try to append data to the end of skb. */
1137 if (copy > seglen)
1138 copy = seglen;
1139
1140 /* Where to copy to? */
1141 if (skb_availroom(skb) > 0) {
1142 /* We have some space in skb head. Superb! */
1143 copy = min_t(int, copy, skb_availroom(skb));
1144 err = skb_add_data_nocache(sk, skb, from, copy);
1145 if (err)
1146 goto do_fault;
1147 } else {
1148 bool merge = true;
1149 int i = skb_shinfo(skb)->nr_frags;
1150 struct page_frag *pfrag = sk_page_frag(sk);
1151
1152 if (!sk_page_frag_refill(sk, pfrag))
1153 goto wait_for_memory;
1154
1155 if (!skb_can_coalesce(skb, i, pfrag->page,
1156 pfrag->offset)) {
1157 if (i == MAX_SKB_FRAGS || !sg) {
1158 tcp_mark_push(tp, skb);
1159 goto new_segment;
1160 }
1161 merge = false;
1162 }
1163
1164 copy = min_t(int, copy, pfrag->size - pfrag->offset);
1165
1166 if (!sk_wmem_schedule(sk, copy))
1167 goto wait_for_memory;
1168
1169 err = skb_copy_to_page_nocache(sk, from, skb,
1170 pfrag->page,
1171 pfrag->offset,
1172 copy);
1173 if (err)
1174 goto do_error;
1175
1176 /* Update the skb. */
1177 if (merge) {
1178 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1179 } else {
1180 skb_fill_page_desc(skb, i, pfrag->page,
1181 pfrag->offset, copy);
1182 get_page(pfrag->page);
1183 }
1184 pfrag->offset += copy;
1185 }
1186
1187 if (!copied)
1188 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
1189
1190 tp->write_seq += copy;
1191 TCP_SKB_CB(skb)->end_seq += copy;
1192 skb_shinfo(skb)->gso_segs = 0;
1193
1194 from += copy;
1195 copied += copy;
1196 if ((seglen -= copy) == 0 && iovlen == 0)
1197 goto out;
1198
1199 if (skb->len < max || (flags & MSG_OOB) || unlikely(tp->repair))
1200 continue;
1201
1202 if (forced_push(tp)) {
1203 tcp_mark_push(tp, skb);
1204 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
1205 } else if (skb == tcp_send_head(sk))
1206 tcp_push_one(sk, mss_now);
1207 continue;
1208
1209 wait_for_sndbuf:
1210 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1211 wait_for_memory:
1212 if (copied)
1213 tcp_push(sk, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH);
1214
1215 if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
1216 goto do_error;
1217
1218 mss_now = tcp_send_mss(sk, &size_goal, flags);
1219 }
1220 }
1221
1222 out:
1223 if (copied)
1224 tcp_push(sk, flags, mss_now, tp->nonagle);
1225 release_sock(sk);
1226 return copied + copied_syn;
1227
1228 do_fault:
1229 if (!skb->len) {
1230 tcp_unlink_write_queue(skb, sk);
1231 /* It is the one place in all of TCP, except connection
1232 * reset, where we can be unlinking the send_head.
1233 */
1234 tcp_check_send_head(sk, skb);
1235 sk_wmem_free_skb(sk, skb);
1236 }
1237
1238 do_error:
1239 if (copied + copied_syn)
1240 goto out;
1241 out_err:
1242 err = sk_stream_error(sk, flags, err);
1243 release_sock(sk);
1244 return err;
1245 }
1246 EXPORT_SYMBOL(tcp_sendmsg);
1247
1248 /*
1249 * Handle reading urgent data. BSD has very simple semantics for
1250 * this, no blocking and very strange errors 8)
1251 */
1252
1253 static int tcp_recv_urg(struct sock *sk, struct msghdr *msg, int len, int flags)
1254 {
1255 struct tcp_sock *tp = tcp_sk(sk);
1256
1257 /* No URG data to read. */
1258 if (sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data ||
1259 tp->urg_data == TCP_URG_READ)
1260 return -EINVAL; /* Yes this is right ! */
1261
1262 if (sk->sk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DONE))
1263 return -ENOTCONN;
1264
1265 if (tp->urg_data & TCP_URG_VALID) {
1266 int err = 0;
1267 char c = tp->urg_data;
1268
1269 if (!(flags & MSG_PEEK))
1270 tp->urg_data = TCP_URG_READ;
1271
1272 /* Read urgent data. */
1273 msg->msg_flags |= MSG_OOB;
1274
1275 if (len > 0) {
1276 if (!(flags & MSG_TRUNC))
1277 err = memcpy_toiovec(msg->msg_iov, &c, 1);
1278 len = 1;
1279 } else
1280 msg->msg_flags |= MSG_TRUNC;
1281
1282 return err ? -EFAULT : len;
1283 }
1284
1285 if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
1286 return 0;
1287
1288 /* Fixed the recv(..., MSG_OOB) behaviour. BSD docs and
1289 * the available implementations agree in this case:
1290 * this call should never block, independent of the
1291 * blocking state of the socket.
1292 * Mike <pall@rz.uni-karlsruhe.de>
1293 */
1294 return -EAGAIN;
1295 }
1296
1297 static int tcp_peek_sndq(struct sock *sk, struct msghdr *msg, int len)
1298 {
1299 struct sk_buff *skb;
1300 int copied = 0, err = 0;
1301
1302 /* XXX -- need to support SO_PEEK_OFF */
1303
1304 skb_queue_walk(&sk->sk_write_queue, skb) {
1305 err = skb_copy_datagram_iovec(skb, 0, msg->msg_iov, skb->len);
1306 if (err)
1307 break;
1308
1309 copied += skb->len;
1310 }
1311
1312 return err ?: copied;
1313 }
1314
1315 /* Clean up the receive buffer for full frames taken by the user,
1316 * then send an ACK if necessary. COPIED is the number of bytes
1317 * tcp_recvmsg has given to the user so far, it speeds up the
1318 * calculation of whether or not we must ACK for the sake of
1319 * a window update.
1320 */
1321 void tcp_cleanup_rbuf(struct sock *sk, int copied)
1322 {
1323 struct tcp_sock *tp = tcp_sk(sk);
1324 bool time_to_ack = false;
1325
1326 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
1327
1328 WARN(skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq),
1329 "cleanup rbuf bug: copied %X seq %X rcvnxt %X\n",
1330 tp->copied_seq, TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt);
1331
1332 if (inet_csk_ack_scheduled(sk)) {
1333 const struct inet_connection_sock *icsk = inet_csk(sk);
1334 /* Delayed ACKs frequently hit locked sockets during bulk
1335 * receive. */
1336 if (icsk->icsk_ack.blocked ||
1337 /* Once-per-two-segments ACK was not sent by tcp_input.c */
1338 tp->rcv_nxt - tp->rcv_wup > icsk->icsk_ack.rcv_mss ||
1339 /*
1340 * If this read emptied read buffer, we send ACK, if
1341 * connection is not bidirectional, user drained
1342 * receive buffer and there was a small segment
1343 * in queue.
1344 */
1345 (copied > 0 &&
1346 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED2) ||
1347 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED) &&
1348 !icsk->icsk_ack.pingpong)) &&
1349 !atomic_read(&sk->sk_rmem_alloc)))
1350 time_to_ack = true;
1351 }
1352
1353 /* We send an ACK if we can now advertise a non-zero window
1354 * which has been raised "significantly".
1355 *
1356 * Even if window raised up to infinity, do not send window open ACK
1357 * in states, where we will not receive more. It is useless.
1358 */
1359 if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) {
1360 __u32 rcv_window_now = tcp_receive_window(tp);
1361
1362 /* Optimize, __tcp_select_window() is not cheap. */
1363 if (2*rcv_window_now <= tp->window_clamp) {
1364 __u32 new_window = __tcp_select_window(sk);
1365
1366 /* Send ACK now, if this read freed lots of space
1367 * in our buffer. Certainly, new_window is new window.
1368 * We can advertise it now, if it is not less than current one.
1369 * "Lots" means "at least twice" here.
1370 */
1371 if (new_window && new_window >= 2 * rcv_window_now)
1372 time_to_ack = true;
1373 }
1374 }
1375 if (time_to_ack)
1376 tcp_send_ack(sk);
1377 }
1378
1379 static void tcp_prequeue_process(struct sock *sk)
1380 {
1381 struct sk_buff *skb;
1382 struct tcp_sock *tp = tcp_sk(sk);
1383
1384 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPPREQUEUED);
1385
1386 /* RX process wants to run with disabled BHs, though it is not
1387 * necessary */
1388 local_bh_disable();
1389 while ((skb = __skb_dequeue(&tp->ucopy.prequeue)) != NULL)
1390 sk_backlog_rcv(sk, skb);
1391 local_bh_enable();
1392
1393 /* Clear memory counter. */
1394 tp->ucopy.memory = 0;
1395 }
1396
1397 #ifdef CONFIG_NET_DMA
1398 static void tcp_service_net_dma(struct sock *sk, bool wait)
1399 {
1400 dma_cookie_t done, used;
1401 dma_cookie_t last_issued;
1402 struct tcp_sock *tp = tcp_sk(sk);
1403
1404 if (!tp->ucopy.dma_chan)
1405 return;
1406
1407 last_issued = tp->ucopy.dma_cookie;
1408 dma_async_memcpy_issue_pending(tp->ucopy.dma_chan);
1409
1410 do {
1411 if (dma_async_memcpy_complete(tp->ucopy.dma_chan,
1412 last_issued, &done,
1413 &used) == DMA_SUCCESS) {
1414 /* Safe to free early-copied skbs now */
1415 __skb_queue_purge(&sk->sk_async_wait_queue);
1416 break;
1417 } else {
1418 struct sk_buff *skb;
1419 while ((skb = skb_peek(&sk->sk_async_wait_queue)) &&
1420 (dma_async_is_complete(skb->dma_cookie, done,
1421 used) == DMA_SUCCESS)) {
1422 __skb_dequeue(&sk->sk_async_wait_queue);
1423 kfree_skb(skb);
1424 }
1425 }
1426 } while (wait);
1427 }
1428 #endif
1429
1430 static inline struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off)
1431 {
1432 struct sk_buff *skb;
1433 u32 offset;
1434
1435 skb_queue_walk(&sk->sk_receive_queue, skb) {
1436 offset = seq - TCP_SKB_CB(skb)->seq;
1437 if (tcp_hdr(skb)->syn)
1438 offset--;
1439 if (offset < skb->len || tcp_hdr(skb)->fin) {
1440 *off = offset;
1441 return skb;
1442 }
1443 }
1444 return NULL;
1445 }
1446
1447 /*
1448 * This routine provides an alternative to tcp_recvmsg() for routines
1449 * that would like to handle copying from skbuffs directly in 'sendfile'
1450 * fashion.
1451 * Note:
1452 * - It is assumed that the socket was locked by the caller.
1453 * - The routine does not block.
1454 * - At present, there is no support for reading OOB data
1455 * or for 'peeking' the socket using this routine
1456 * (although both would be easy to implement).
1457 */
1458 int tcp_read_sock(struct sock *sk, read_descriptor_t *desc,
1459 sk_read_actor_t recv_actor)
1460 {
1461 struct sk_buff *skb;
1462 struct tcp_sock *tp = tcp_sk(sk);
1463 u32 seq = tp->copied_seq;
1464 u32 offset;
1465 int copied = 0;
1466
1467 if (sk->sk_state == TCP_LISTEN)
1468 return -ENOTCONN;
1469 while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1470 if (offset < skb->len) {
1471 int used;
1472 size_t len;
1473
1474 len = skb->len - offset;
1475 /* Stop reading if we hit a patch of urgent data */
1476 if (tp->urg_data) {
1477 u32 urg_offset = tp->urg_seq - seq;
1478 if (urg_offset < len)
1479 len = urg_offset;
1480 if (!len)
1481 break;
1482 }
1483 used = recv_actor(desc, skb, offset, len);
1484 if (used < 0) {
1485 if (!copied)
1486 copied = used;
1487 break;
1488 } else if (used <= len) {
1489 seq += used;
1490 copied += used;
1491 offset += used;
1492 }
1493 /*
1494 * If recv_actor drops the lock (e.g. TCP splice
1495 * receive) the skb pointer might be invalid when
1496 * getting here: tcp_collapse might have deleted it
1497 * while aggregating skbs from the socket queue.
1498 */
1499 skb = tcp_recv_skb(sk, seq-1, &offset);
1500 if (!skb || (offset+1 != skb->len))
1501 break;
1502 }
1503 if (tcp_hdr(skb)->fin) {
1504 sk_eat_skb(sk, skb, false);
1505 ++seq;
1506 break;
1507 }
1508 sk_eat_skb(sk, skb, false);
1509 if (!desc->count)
1510 break;
1511 tp->copied_seq = seq;
1512 }
1513 tp->copied_seq = seq;
1514
1515 tcp_rcv_space_adjust(sk);
1516
1517 /* Clean up data we have read: This will do ACK frames. */
1518 if (copied > 0)
1519 tcp_cleanup_rbuf(sk, copied);
1520 return copied;
1521 }
1522 EXPORT_SYMBOL(tcp_read_sock);
1523
1524 /*
1525 * This routine copies from a sock struct into the user buffer.
1526 *
1527 * Technical note: in 2.3 we work on _locked_ socket, so that
1528 * tricks with *seq access order and skb->users are not required.
1529 * Probably, code can be easily improved even more.
1530 */
1531
1532 int tcp_recvmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg,
1533 size_t len, int nonblock, int flags, int *addr_len)
1534 {
1535 struct tcp_sock *tp = tcp_sk(sk);
1536 int copied = 0;
1537 u32 peek_seq;
1538 u32 *seq;
1539 unsigned long used;
1540 int err;
1541 int target; /* Read at least this many bytes */
1542 long timeo;
1543 struct task_struct *user_recv = NULL;
1544 bool copied_early = false;
1545 struct sk_buff *skb;
1546 u32 urg_hole = 0;
1547
1548 lock_sock(sk);
1549
1550 err = -ENOTCONN;
1551 if (sk->sk_state == TCP_LISTEN)
1552 goto out;
1553
1554 timeo = sock_rcvtimeo(sk, nonblock);
1555
1556 /* Urgent data needs to be handled specially. */
1557 if (flags & MSG_OOB)
1558 goto recv_urg;
1559
1560 if (unlikely(tp->repair)) {
1561 err = -EPERM;
1562 if (!(flags & MSG_PEEK))
1563 goto out;
1564
1565 if (tp->repair_queue == TCP_SEND_QUEUE)
1566 goto recv_sndq;
1567
1568 err = -EINVAL;
1569 if (tp->repair_queue == TCP_NO_QUEUE)
1570 goto out;
1571
1572 /* 'common' recv queue MSG_PEEK-ing */
1573 }
1574
1575 seq = &tp->copied_seq;
1576 if (flags & MSG_PEEK) {
1577 peek_seq = tp->copied_seq;
1578 seq = &peek_seq;
1579 }
1580
1581 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
1582
1583 #ifdef CONFIG_NET_DMA
1584 tp->ucopy.dma_chan = NULL;
1585 preempt_disable();
1586 skb = skb_peek_tail(&sk->sk_receive_queue);
1587 {
1588 int available = 0;
1589
1590 if (skb)
1591 available = TCP_SKB_CB(skb)->seq + skb->len - (*seq);
1592 if ((available < target) &&
1593 (len > sysctl_tcp_dma_copybreak) && !(flags & MSG_PEEK) &&
1594 !sysctl_tcp_low_latency &&
1595 net_dma_find_channel()) {
1596 preempt_enable_no_resched();
1597 tp->ucopy.pinned_list =
1598 dma_pin_iovec_pages(msg->msg_iov, len);
1599 } else {
1600 preempt_enable_no_resched();
1601 }
1602 }
1603 #endif
1604
1605 do {
1606 u32 offset;
1607
1608 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
1609 if (tp->urg_data && tp->urg_seq == *seq) {
1610 if (copied)
1611 break;
1612 if (signal_pending(current)) {
1613 copied = timeo ? sock_intr_errno(timeo) : -EAGAIN;
1614 break;
1615 }
1616 }
1617
1618 /* Next get a buffer. */
1619
1620 skb_queue_walk(&sk->sk_receive_queue, skb) {
1621 /* Now that we have two receive queues this
1622 * shouldn't happen.
1623 */
1624 if (WARN(before(*seq, TCP_SKB_CB(skb)->seq),
1625 "recvmsg bug: copied %X seq %X rcvnxt %X fl %X\n",
1626 *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt,
1627 flags))
1628 break;
1629
1630 offset = *seq - TCP_SKB_CB(skb)->seq;
1631 if (tcp_hdr(skb)->syn)
1632 offset--;
1633 if (offset < skb->len)
1634 goto found_ok_skb;
1635 if (tcp_hdr(skb)->fin)
1636 goto found_fin_ok;
1637 WARN(!(flags & MSG_PEEK),
1638 "recvmsg bug 2: copied %X seq %X rcvnxt %X fl %X\n",
1639 *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt, flags);
1640 }
1641
1642 /* Well, if we have backlog, try to process it now yet. */
1643
1644 if (copied >= target && !sk->sk_backlog.tail)
1645 break;
1646
1647 if (copied) {
1648 if (sk->sk_err ||
1649 sk->sk_state == TCP_CLOSE ||
1650 (sk->sk_shutdown & RCV_SHUTDOWN) ||
1651 !timeo ||
1652 signal_pending(current))
1653 break;
1654 } else {
1655 if (sock_flag(sk, SOCK_DONE))
1656 break;
1657
1658 if (sk->sk_err) {
1659 copied = sock_error(sk);
1660 break;
1661 }
1662
1663 if (sk->sk_shutdown & RCV_SHUTDOWN)
1664 break;
1665
1666 if (sk->sk_state == TCP_CLOSE) {
1667 if (!sock_flag(sk, SOCK_DONE)) {
1668 /* This occurs when user tries to read
1669 * from never connected socket.
1670 */
1671 copied = -ENOTCONN;
1672 break;
1673 }
1674 break;
1675 }
1676
1677 if (!timeo) {
1678 copied = -EAGAIN;
1679 break;
1680 }
1681
1682 if (signal_pending(current)) {
1683 copied = sock_intr_errno(timeo);
1684 break;
1685 }
1686 }
1687
1688 tcp_cleanup_rbuf(sk, copied);
1689
1690 if (!sysctl_tcp_low_latency && tp->ucopy.task == user_recv) {
1691 /* Install new reader */
1692 if (!user_recv && !(flags & (MSG_TRUNC | MSG_PEEK))) {
1693 user_recv = current;
1694 tp->ucopy.task = user_recv;
1695 tp->ucopy.iov = msg->msg_iov;
1696 }
1697
1698 tp->ucopy.len = len;
1699
1700 WARN_ON(tp->copied_seq != tp->rcv_nxt &&
1701 !(flags & (MSG_PEEK | MSG_TRUNC)));
1702
1703 /* Ugly... If prequeue is not empty, we have to
1704 * process it before releasing socket, otherwise
1705 * order will be broken at second iteration.
1706 * More elegant solution is required!!!
1707 *
1708 * Look: we have the following (pseudo)queues:
1709 *
1710 * 1. packets in flight
1711 * 2. backlog
1712 * 3. prequeue
1713 * 4. receive_queue
1714 *
1715 * Each queue can be processed only if the next ones
1716 * are empty. At this point we have empty receive_queue.
1717 * But prequeue _can_ be not empty after 2nd iteration,
1718 * when we jumped to start of loop because backlog
1719 * processing added something to receive_queue.
1720 * We cannot release_sock(), because backlog contains
1721 * packets arrived _after_ prequeued ones.
1722 *
1723 * Shortly, algorithm is clear --- to process all
1724 * the queues in order. We could make it more directly,
1725 * requeueing packets from backlog to prequeue, if
1726 * is not empty. It is more elegant, but eats cycles,
1727 * unfortunately.
1728 */
1729 if (!skb_queue_empty(&tp->ucopy.prequeue))
1730 goto do_prequeue;
1731
1732 /* __ Set realtime policy in scheduler __ */
1733 }
1734
1735 #ifdef CONFIG_NET_DMA
1736 if (tp->ucopy.dma_chan) {
1737 if (tp->rcv_wnd == 0 &&
1738 !skb_queue_empty(&sk->sk_async_wait_queue)) {
1739 tcp_service_net_dma(sk, true);
1740 tcp_cleanup_rbuf(sk, copied);
1741 } else
1742 dma_async_memcpy_issue_pending(tp->ucopy.dma_chan);
1743 }
1744 #endif
1745 if (copied >= target) {
1746 /* Do not sleep, just process backlog. */
1747 release_sock(sk);
1748 lock_sock(sk);
1749 } else
1750 sk_wait_data(sk, &timeo);
1751
1752 #ifdef CONFIG_NET_DMA
1753 tcp_service_net_dma(sk, false); /* Don't block */
1754 tp->ucopy.wakeup = 0;
1755 #endif
1756
1757 if (user_recv) {
1758 int chunk;
1759
1760 /* __ Restore normal policy in scheduler __ */
1761
1762 if ((chunk = len - tp->ucopy.len) != 0) {
1763 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMBACKLOG, chunk);
1764 len -= chunk;
1765 copied += chunk;
1766 }
1767
1768 if (tp->rcv_nxt == tp->copied_seq &&
1769 !skb_queue_empty(&tp->ucopy.prequeue)) {
1770 do_prequeue:
1771 tcp_prequeue_process(sk);
1772
1773 if ((chunk = len - tp->ucopy.len) != 0) {
1774 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1775 len -= chunk;
1776 copied += chunk;
1777 }
1778 }
1779 }
1780 if ((flags & MSG_PEEK) &&
1781 (peek_seq - copied - urg_hole != tp->copied_seq)) {
1782 net_dbg_ratelimited("TCP(%s:%d): Application bug, race in MSG_PEEK\n",
1783 current->comm,
1784 task_pid_nr(current));
1785 peek_seq = tp->copied_seq;
1786 }
1787 continue;
1788
1789 found_ok_skb:
1790 /* Ok so how much can we use? */
1791 used = skb->len - offset;
1792 if (len < used)
1793 used = len;
1794
1795 /* Do we have urgent data here? */
1796 if (tp->urg_data) {
1797 u32 urg_offset = tp->urg_seq - *seq;
1798 if (urg_offset < used) {
1799 if (!urg_offset) {
1800 if (!sock_flag(sk, SOCK_URGINLINE)) {
1801 ++*seq;
1802 urg_hole++;
1803 offset++;
1804 used--;
1805 if (!used)
1806 goto skip_copy;
1807 }
1808 } else
1809 used = urg_offset;
1810 }
1811 }
1812
1813 if (!(flags & MSG_TRUNC)) {
1814 #ifdef CONFIG_NET_DMA
1815 if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list)
1816 tp->ucopy.dma_chan = net_dma_find_channel();
1817
1818 if (tp->ucopy.dma_chan) {
1819 tp->ucopy.dma_cookie = dma_skb_copy_datagram_iovec(
1820 tp->ucopy.dma_chan, skb, offset,
1821 msg->msg_iov, used,
1822 tp->ucopy.pinned_list);
1823
1824 if (tp->ucopy.dma_cookie < 0) {
1825
1826 pr_alert("%s: dma_cookie < 0\n",
1827 __func__);
1828
1829 /* Exception. Bailout! */
1830 if (!copied)
1831 copied = -EFAULT;
1832 break;
1833 }
1834
1835 dma_async_memcpy_issue_pending(tp->ucopy.dma_chan);
1836
1837 if ((offset + used) == skb->len)
1838 copied_early = true;
1839
1840 } else
1841 #endif
1842 {
1843 err = skb_copy_datagram_iovec(skb, offset,
1844 msg->msg_iov, used);
1845 if (err) {
1846 /* Exception. Bailout! */
1847 if (!copied)
1848 copied = -EFAULT;
1849 break;
1850 }
1851 }
1852 }
1853
1854 *seq += used;
1855 copied += used;
1856 len -= used;
1857
1858 tcp_rcv_space_adjust(sk);
1859
1860 skip_copy:
1861 if (tp->urg_data && after(tp->copied_seq, tp->urg_seq)) {
1862 tp->urg_data = 0;
1863 tcp_fast_path_check(sk);
1864 }
1865 if (used + offset < skb->len)
1866 continue;
1867
1868 if (tcp_hdr(skb)->fin)
1869 goto found_fin_ok;
1870 if (!(flags & MSG_PEEK)) {
1871 sk_eat_skb(sk, skb, copied_early);
1872 copied_early = false;
1873 }
1874 continue;
1875
1876 found_fin_ok:
1877 /* Process the FIN. */
1878 ++*seq;
1879 if (!(flags & MSG_PEEK)) {
1880 sk_eat_skb(sk, skb, copied_early);
1881 copied_early = false;
1882 }
1883 break;
1884 } while (len > 0);
1885
1886 if (user_recv) {
1887 if (!skb_queue_empty(&tp->ucopy.prequeue)) {
1888 int chunk;
1889
1890 tp->ucopy.len = copied > 0 ? len : 0;
1891
1892 tcp_prequeue_process(sk);
1893
1894 if (copied > 0 && (chunk = len - tp->ucopy.len) != 0) {
1895 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1896 len -= chunk;
1897 copied += chunk;
1898 }
1899 }
1900
1901 tp->ucopy.task = NULL;
1902 tp->ucopy.len = 0;
1903 }
1904
1905 #ifdef CONFIG_NET_DMA
1906 tcp_service_net_dma(sk, true); /* Wait for queue to drain */
1907 tp->ucopy.dma_chan = NULL;
1908
1909 if (tp->ucopy.pinned_list) {
1910 dma_unpin_iovec_pages(tp->ucopy.pinned_list);
1911 tp->ucopy.pinned_list = NULL;
1912 }
1913 #endif
1914
1915 /* According to UNIX98, msg_name/msg_namelen are ignored
1916 * on connected socket. I was just happy when found this 8) --ANK
1917 */
1918
1919 /* Clean up data we have read: This will do ACK frames. */
1920 tcp_cleanup_rbuf(sk, copied);
1921
1922 release_sock(sk);
1923 return copied;
1924
1925 out:
1926 release_sock(sk);
1927 return err;
1928
1929 recv_urg:
1930 err = tcp_recv_urg(sk, msg, len, flags);
1931 goto out;
1932
1933 recv_sndq:
1934 err = tcp_peek_sndq(sk, msg, len);
1935 goto out;
1936 }
1937 EXPORT_SYMBOL(tcp_recvmsg);
1938
1939 void tcp_set_state(struct sock *sk, int state)
1940 {
1941 int oldstate = sk->sk_state;
1942
1943 switch (state) {
1944 case TCP_ESTABLISHED:
1945 if (oldstate != TCP_ESTABLISHED)
1946 TCP_INC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
1947 break;
1948
1949 case TCP_CLOSE:
1950 if (oldstate == TCP_CLOSE_WAIT || oldstate == TCP_ESTABLISHED)
1951 TCP_INC_STATS(sock_net(sk), TCP_MIB_ESTABRESETS);
1952
1953 sk->sk_prot->unhash(sk);
1954 if (inet_csk(sk)->icsk_bind_hash &&
1955 !(sk->sk_userlocks & SOCK_BINDPORT_LOCK))
1956 inet_put_port(sk);
1957 /* fall through */
1958 default:
1959 if (oldstate == TCP_ESTABLISHED)
1960 TCP_DEC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
1961 }
1962
1963 /* Change state AFTER socket is unhashed to avoid closed
1964 * socket sitting in hash tables.
1965 */
1966 sk->sk_state = state;
1967
1968 #ifdef STATE_TRACE
1969 SOCK_DEBUG(sk, "TCP sk=%p, State %s -> %s\n", sk, statename[oldstate], statename[state]);
1970 #endif
1971 }
1972 EXPORT_SYMBOL_GPL(tcp_set_state);
1973
1974 /*
1975 * State processing on a close. This implements the state shift for
1976 * sending our FIN frame. Note that we only send a FIN for some
1977 * states. A shutdown() may have already sent the FIN, or we may be
1978 * closed.
1979 */
1980
1981 static const unsigned char new_state[16] = {
1982 /* current state: new state: action: */
1983 /* (Invalid) */ TCP_CLOSE,
1984 /* TCP_ESTABLISHED */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1985 /* TCP_SYN_SENT */ TCP_CLOSE,
1986 /* TCP_SYN_RECV */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1987 /* TCP_FIN_WAIT1 */ TCP_FIN_WAIT1,
1988 /* TCP_FIN_WAIT2 */ TCP_FIN_WAIT2,
1989 /* TCP_TIME_WAIT */ TCP_CLOSE,
1990 /* TCP_CLOSE */ TCP_CLOSE,
1991 /* TCP_CLOSE_WAIT */ TCP_LAST_ACK | TCP_ACTION_FIN,
1992 /* TCP_LAST_ACK */ TCP_LAST_ACK,
1993 /* TCP_LISTEN */ TCP_CLOSE,
1994 /* TCP_CLOSING */ TCP_CLOSING,
1995 };
1996
1997 static int tcp_close_state(struct sock *sk)
1998 {
1999 int next = (int)new_state[sk->sk_state];
2000 int ns = next & TCP_STATE_MASK;
2001
2002 tcp_set_state(sk, ns);
2003
2004 return next & TCP_ACTION_FIN;
2005 }
2006
2007 /*
2008 * Shutdown the sending side of a connection. Much like close except
2009 * that we don't receive shut down or sock_set_flag(sk, SOCK_DEAD).
2010 */
2011
2012 void tcp_shutdown(struct sock *sk, int how)
2013 {
2014 /* We need to grab some memory, and put together a FIN,
2015 * and then put it into the queue to be sent.
2016 * Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
2017 */
2018 if (!(how & SEND_SHUTDOWN))
2019 return;
2020
2021 /* If we've already sent a FIN, or it's a closed state, skip this. */
2022 if ((1 << sk->sk_state) &
2023 (TCPF_ESTABLISHED | TCPF_SYN_SENT |
2024 TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) {
2025 /* Clear out any half completed packets. FIN if needed. */
2026 if (tcp_close_state(sk))
2027 tcp_send_fin(sk);
2028 }
2029 }
2030 EXPORT_SYMBOL(tcp_shutdown);
2031
2032 bool tcp_check_oom(struct sock *sk, int shift)
2033 {
2034 bool too_many_orphans, out_of_socket_memory;
2035
2036 too_many_orphans = tcp_too_many_orphans(sk, shift);
2037 out_of_socket_memory = tcp_out_of_memory(sk);
2038
2039 if (too_many_orphans)
2040 net_info_ratelimited("too many orphaned sockets\n");
2041 if (out_of_socket_memory)
2042 net_info_ratelimited("out of memory -- consider tuning tcp_mem\n");
2043 return too_many_orphans || out_of_socket_memory;
2044 }
2045
2046 void tcp_close(struct sock *sk, long timeout)
2047 {
2048 struct sk_buff *skb;
2049 int data_was_unread = 0;
2050 int state;
2051
2052 lock_sock(sk);
2053 sk->sk_shutdown = SHUTDOWN_MASK;
2054
2055 if (sk->sk_state == TCP_LISTEN) {
2056 tcp_set_state(sk, TCP_CLOSE);
2057
2058 /* Special case. */
2059 inet_csk_listen_stop(sk);
2060
2061 goto adjudge_to_death;
2062 }
2063
2064 /* We need to flush the recv. buffs. We do this only on the
2065 * descriptor close, not protocol-sourced closes, because the
2066 * reader process may not have drained the data yet!
2067 */
2068 while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) {
2069 u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq -
2070 tcp_hdr(skb)->fin;
2071 data_was_unread += len;
2072 __kfree_skb(skb);
2073 }
2074
2075 sk_mem_reclaim(sk);
2076
2077 /* If socket has been already reset (e.g. in tcp_reset()) - kill it. */
2078 if (sk->sk_state == TCP_CLOSE)
2079 goto adjudge_to_death;
2080
2081 /* As outlined in RFC 2525, section 2.17, we send a RST here because
2082 * data was lost. To witness the awful effects of the old behavior of
2083 * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk
2084 * GET in an FTP client, suspend the process, wait for the client to
2085 * advertise a zero window, then kill -9 the FTP client, wheee...
2086 * Note: timeout is always zero in such a case.
2087 */
2088 if (unlikely(tcp_sk(sk)->repair)) {
2089 sk->sk_prot->disconnect(sk, 0);
2090 } else if (data_was_unread) {
2091 /* Unread data was tossed, zap the connection. */
2092 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONCLOSE);
2093 tcp_set_state(sk, TCP_CLOSE);
2094 tcp_send_active_reset(sk, sk->sk_allocation);
2095 } else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) {
2096 /* Check zero linger _after_ checking for unread data. */
2097 sk->sk_prot->disconnect(sk, 0);
2098 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
2099 } else if (tcp_close_state(sk)) {
2100 /* We FIN if the application ate all the data before
2101 * zapping the connection.
2102 */
2103
2104 /* RED-PEN. Formally speaking, we have broken TCP state
2105 * machine. State transitions:
2106 *
2107 * TCP_ESTABLISHED -> TCP_FIN_WAIT1
2108 * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible)
2109 * TCP_CLOSE_WAIT -> TCP_LAST_ACK
2110 *
2111 * are legal only when FIN has been sent (i.e. in window),
2112 * rather than queued out of window. Purists blame.
2113 *
2114 * F.e. "RFC state" is ESTABLISHED,
2115 * if Linux state is FIN-WAIT-1, but FIN is still not sent.
2116 *
2117 * The visible declinations are that sometimes
2118 * we enter time-wait state, when it is not required really
2119 * (harmless), do not send active resets, when they are
2120 * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
2121 * they look as CLOSING or LAST_ACK for Linux)
2122 * Probably, I missed some more holelets.
2123 * --ANK
2124 * XXX (TFO) - To start off we don't support SYN+ACK+FIN
2125 * in a single packet! (May consider it later but will
2126 * probably need API support or TCP_CORK SYN-ACK until
2127 * data is written and socket is closed.)
2128 */
2129 tcp_send_fin(sk);
2130 }
2131
2132 sk_stream_wait_close(sk, timeout);
2133
2134 adjudge_to_death:
2135 state = sk->sk_state;
2136 sock_hold(sk);
2137 sock_orphan(sk);
2138
2139 /* It is the last release_sock in its life. It will remove backlog. */
2140 release_sock(sk);
2141
2142
2143 /* Now socket is owned by kernel and we acquire BH lock
2144 to finish close. No need to check for user refs.
2145 */
2146 local_bh_disable();
2147 bh_lock_sock(sk);
2148 WARN_ON(sock_owned_by_user(sk));
2149
2150 percpu_counter_inc(sk->sk_prot->orphan_count);
2151
2152 /* Have we already been destroyed by a softirq or backlog? */
2153 if (state != TCP_CLOSE && sk->sk_state == TCP_CLOSE)
2154 goto out;
2155
2156 /* This is a (useful) BSD violating of the RFC. There is a
2157 * problem with TCP as specified in that the other end could
2158 * keep a socket open forever with no application left this end.
2159 * We use a 3 minute timeout (about the same as BSD) then kill
2160 * our end. If they send after that then tough - BUT: long enough
2161 * that we won't make the old 4*rto = almost no time - whoops
2162 * reset mistake.
2163 *
2164 * Nope, it was not mistake. It is really desired behaviour
2165 * f.e. on http servers, when such sockets are useless, but
2166 * consume significant resources. Let's do it with special
2167 * linger2 option. --ANK
2168 */
2169
2170 if (sk->sk_state == TCP_FIN_WAIT2) {
2171 struct tcp_sock *tp = tcp_sk(sk);
2172 if (tp->linger2 < 0) {
2173 tcp_set_state(sk, TCP_CLOSE);
2174 tcp_send_active_reset(sk, GFP_ATOMIC);
2175 NET_INC_STATS_BH(sock_net(sk),
2176 LINUX_MIB_TCPABORTONLINGER);
2177 } else {
2178 const int tmo = tcp_fin_time(sk);
2179
2180 if (tmo > TCP_TIMEWAIT_LEN) {
2181 inet_csk_reset_keepalive_timer(sk,
2182 tmo - TCP_TIMEWAIT_LEN);
2183 } else {
2184 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
2185 goto out;
2186 }
2187 }
2188 }
2189 if (sk->sk_state != TCP_CLOSE) {
2190 sk_mem_reclaim(sk);
2191 if (tcp_check_oom(sk, 0)) {
2192 tcp_set_state(sk, TCP_CLOSE);
2193 tcp_send_active_reset(sk, GFP_ATOMIC);
2194 NET_INC_STATS_BH(sock_net(sk),
2195 LINUX_MIB_TCPABORTONMEMORY);
2196 }
2197 }
2198
2199 if (sk->sk_state == TCP_CLOSE) {
2200 struct request_sock *req = tcp_sk(sk)->fastopen_rsk;
2201 /* We could get here with a non-NULL req if the socket is
2202 * aborted (e.g., closed with unread data) before 3WHS
2203 * finishes.
2204 */
2205 if (req != NULL)
2206 reqsk_fastopen_remove(sk, req, false);
2207 inet_csk_destroy_sock(sk);
2208 }
2209 /* Otherwise, socket is reprieved until protocol close. */
2210
2211 out:
2212 bh_unlock_sock(sk);
2213 local_bh_enable();
2214 sock_put(sk);
2215 }
2216 EXPORT_SYMBOL(tcp_close);
2217
2218 /* These states need RST on ABORT according to RFC793 */
2219
2220 static inline bool tcp_need_reset(int state)
2221 {
2222 return (1 << state) &
2223 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 |
2224 TCPF_FIN_WAIT2 | TCPF_SYN_RECV);
2225 }
2226
2227 int tcp_disconnect(struct sock *sk, int flags)
2228 {
2229 struct inet_sock *inet = inet_sk(sk);
2230 struct inet_connection_sock *icsk = inet_csk(sk);
2231 struct tcp_sock *tp = tcp_sk(sk);
2232 int err = 0;
2233 int old_state = sk->sk_state;
2234
2235 if (old_state != TCP_CLOSE)
2236 tcp_set_state(sk, TCP_CLOSE);
2237
2238 /* ABORT function of RFC793 */
2239 if (old_state == TCP_LISTEN) {
2240 inet_csk_listen_stop(sk);
2241 } else if (unlikely(tp->repair)) {
2242 sk->sk_err = ECONNABORTED;
2243 } else if (tcp_need_reset(old_state) ||
2244 (tp->snd_nxt != tp->write_seq &&
2245 (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK))) {
2246 /* The last check adjusts for discrepancy of Linux wrt. RFC
2247 * states
2248 */
2249 tcp_send_active_reset(sk, gfp_any());
2250 sk->sk_err = ECONNRESET;
2251 } else if (old_state == TCP_SYN_SENT)
2252 sk->sk_err = ECONNRESET;
2253
2254 tcp_clear_xmit_timers(sk);
2255 __skb_queue_purge(&sk->sk_receive_queue);
2256 tcp_write_queue_purge(sk);
2257 __skb_queue_purge(&tp->out_of_order_queue);
2258 #ifdef CONFIG_NET_DMA
2259 __skb_queue_purge(&sk->sk_async_wait_queue);
2260 #endif
2261
2262 inet->inet_dport = 0;
2263
2264 if (!(sk->sk_userlocks & SOCK_BINDADDR_LOCK))
2265 inet_reset_saddr(sk);
2266
2267 sk->sk_shutdown = 0;
2268 sock_reset_flag(sk, SOCK_DONE);
2269 tp->srtt = 0;
2270 if ((tp->write_seq += tp->max_window + 2) == 0)
2271 tp->write_seq = 1;
2272 icsk->icsk_backoff = 0;
2273 tp->snd_cwnd = 2;
2274 icsk->icsk_probes_out = 0;
2275 tp->packets_out = 0;
2276 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
2277 tp->snd_cwnd_cnt = 0;
2278 tp->bytes_acked = 0;
2279 tp->window_clamp = 0;
2280 tcp_set_ca_state(sk, TCP_CA_Open);
2281 tcp_clear_retrans(tp);
2282 inet_csk_delack_init(sk);
2283 tcp_init_send_head(sk);
2284 memset(&tp->rx_opt, 0, sizeof(tp->rx_opt));
2285 __sk_dst_reset(sk);
2286
2287 WARN_ON(inet->inet_num && !icsk->icsk_bind_hash);
2288
2289 sk->sk_error_report(sk);
2290 return err;
2291 }
2292 EXPORT_SYMBOL(tcp_disconnect);
2293
2294 void tcp_sock_destruct(struct sock *sk)
2295 {
2296 inet_sock_destruct(sk);
2297
2298 kfree(inet_csk(sk)->icsk_accept_queue.fastopenq);
2299 }
2300
2301 static inline bool tcp_can_repair_sock(const struct sock *sk)
2302 {
2303 return capable(CAP_NET_ADMIN) &&
2304 ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_ESTABLISHED));
2305 }
2306
2307 static int tcp_repair_options_est(struct tcp_sock *tp,
2308 struct tcp_repair_opt __user *optbuf, unsigned int len)
2309 {
2310 struct tcp_repair_opt opt;
2311
2312 while (len >= sizeof(opt)) {
2313 if (copy_from_user(&opt, optbuf, sizeof(opt)))
2314 return -EFAULT;
2315
2316 optbuf++;
2317 len -= sizeof(opt);
2318
2319 switch (opt.opt_code) {
2320 case TCPOPT_MSS:
2321 tp->rx_opt.mss_clamp = opt.opt_val;
2322 break;
2323 case TCPOPT_WINDOW:
2324 {
2325 u16 snd_wscale = opt.opt_val & 0xFFFF;
2326 u16 rcv_wscale = opt.opt_val >> 16;
2327
2328 if (snd_wscale > 14 || rcv_wscale > 14)
2329 return -EFBIG;
2330
2331 tp->rx_opt.snd_wscale = snd_wscale;
2332 tp->rx_opt.rcv_wscale = rcv_wscale;
2333 tp->rx_opt.wscale_ok = 1;
2334 }
2335 break;
2336 case TCPOPT_SACK_PERM:
2337 if (opt.opt_val != 0)
2338 return -EINVAL;
2339
2340 tp->rx_opt.sack_ok |= TCP_SACK_SEEN;
2341 if (sysctl_tcp_fack)
2342 tcp_enable_fack(tp);
2343 break;
2344 case TCPOPT_TIMESTAMP:
2345 if (opt.opt_val != 0)
2346 return -EINVAL;
2347
2348 tp->rx_opt.tstamp_ok = 1;
2349 break;
2350 }
2351 }
2352
2353 return 0;
2354 }
2355
2356 /*
2357 * Socket option code for TCP.
2358 */
2359 static int do_tcp_setsockopt(struct sock *sk, int level,
2360 int optname, char __user *optval, unsigned int optlen)
2361 {
2362 struct tcp_sock *tp = tcp_sk(sk);
2363 struct inet_connection_sock *icsk = inet_csk(sk);
2364 int val;
2365 int err = 0;
2366
2367 /* These are data/string values, all the others are ints */
2368 switch (optname) {
2369 case TCP_CONGESTION: {
2370 char name[TCP_CA_NAME_MAX];
2371
2372 if (optlen < 1)
2373 return -EINVAL;
2374
2375 val = strncpy_from_user(name, optval,
2376 min_t(long, TCP_CA_NAME_MAX-1, optlen));
2377 if (val < 0)
2378 return -EFAULT;
2379 name[val] = 0;
2380
2381 lock_sock(sk);
2382 err = tcp_set_congestion_control(sk, name);
2383 release_sock(sk);
2384 return err;
2385 }
2386 case TCP_COOKIE_TRANSACTIONS: {
2387 struct tcp_cookie_transactions ctd;
2388 struct tcp_cookie_values *cvp = NULL;
2389
2390 if (sizeof(ctd) > optlen)
2391 return -EINVAL;
2392 if (copy_from_user(&ctd, optval, sizeof(ctd)))
2393 return -EFAULT;
2394
2395 if (ctd.tcpct_used > sizeof(ctd.tcpct_value) ||
2396 ctd.tcpct_s_data_desired > TCP_MSS_DESIRED)
2397 return -EINVAL;
2398
2399 if (ctd.tcpct_cookie_desired == 0) {
2400 /* default to global value */
2401 } else if ((0x1 & ctd.tcpct_cookie_desired) ||
2402 ctd.tcpct_cookie_desired > TCP_COOKIE_MAX ||
2403 ctd.tcpct_cookie_desired < TCP_COOKIE_MIN) {
2404 return -EINVAL;
2405 }
2406
2407 if (TCP_COOKIE_OUT_NEVER & ctd.tcpct_flags) {
2408 /* Supercedes all other values */
2409 lock_sock(sk);
2410 if (tp->cookie_values != NULL) {
2411 kref_put(&tp->cookie_values->kref,
2412 tcp_cookie_values_release);
2413 tp->cookie_values = NULL;
2414 }
2415 tp->rx_opt.cookie_in_always = 0; /* false */
2416 tp->rx_opt.cookie_out_never = 1; /* true */
2417 release_sock(sk);
2418 return err;
2419 }
2420
2421 /* Allocate ancillary memory before locking.
2422 */
2423 if (ctd.tcpct_used > 0 ||
2424 (tp->cookie_values == NULL &&
2425 (sysctl_tcp_cookie_size > 0 ||
2426 ctd.tcpct_cookie_desired > 0 ||
2427 ctd.tcpct_s_data_desired > 0))) {
2428 cvp = kzalloc(sizeof(*cvp) + ctd.tcpct_used,
2429 GFP_KERNEL);
2430 if (cvp == NULL)
2431 return -ENOMEM;
2432
2433 kref_init(&cvp->kref);
2434 }
2435 lock_sock(sk);
2436 tp->rx_opt.cookie_in_always =
2437 (TCP_COOKIE_IN_ALWAYS & ctd.tcpct_flags);
2438 tp->rx_opt.cookie_out_never = 0; /* false */
2439
2440 if (tp->cookie_values != NULL) {
2441 if (cvp != NULL) {
2442 /* Changed values are recorded by a changed
2443 * pointer, ensuring the cookie will differ,
2444 * without separately hashing each value later.
2445 */
2446 kref_put(&tp->cookie_values->kref,
2447 tcp_cookie_values_release);
2448 } else {
2449 cvp = tp->cookie_values;
2450 }
2451 }
2452
2453 if (cvp != NULL) {
2454 cvp->cookie_desired = ctd.tcpct_cookie_desired;
2455
2456 if (ctd.tcpct_used > 0) {
2457 memcpy(cvp->s_data_payload, ctd.tcpct_value,
2458 ctd.tcpct_used);
2459 cvp->s_data_desired = ctd.tcpct_used;
2460 cvp->s_data_constant = 1; /* true */
2461 } else {
2462 /* No constant payload data. */
2463 cvp->s_data_desired = ctd.tcpct_s_data_desired;
2464 cvp->s_data_constant = 0; /* false */
2465 }
2466
2467 tp->cookie_values = cvp;
2468 }
2469 release_sock(sk);
2470 return err;
2471 }
2472 default:
2473 /* fallthru */
2474 break;
2475 }
2476
2477 if (optlen < sizeof(int))
2478 return -EINVAL;
2479
2480 if (get_user(val, (int __user *)optval))
2481 return -EFAULT;
2482
2483 lock_sock(sk);
2484
2485 switch (optname) {
2486 case TCP_MAXSEG:
2487 /* Values greater than interface MTU won't take effect. However
2488 * at the point when this call is done we typically don't yet
2489 * know which interface is going to be used */
2490 if (val < TCP_MIN_MSS || val > MAX_TCP_WINDOW) {
2491 err = -EINVAL;
2492 break;
2493 }
2494 tp->rx_opt.user_mss = val;
2495 break;
2496
2497 case TCP_NODELAY:
2498 if (val) {
2499 /* TCP_NODELAY is weaker than TCP_CORK, so that
2500 * this option on corked socket is remembered, but
2501 * it is not activated until cork is cleared.
2502 *
2503 * However, when TCP_NODELAY is set we make
2504 * an explicit push, which overrides even TCP_CORK
2505 * for currently queued segments.
2506 */
2507 tp->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH;
2508 tcp_push_pending_frames(sk);
2509 } else {
2510 tp->nonagle &= ~TCP_NAGLE_OFF;
2511 }
2512 break;
2513
2514 case TCP_THIN_LINEAR_TIMEOUTS:
2515 if (val < 0 || val > 1)
2516 err = -EINVAL;
2517 else
2518 tp->thin_lto = val;
2519 break;
2520
2521 case TCP_THIN_DUPACK:
2522 if (val < 0 || val > 1)
2523 err = -EINVAL;
2524 else
2525 tp->thin_dupack = val;
2526 if (tp->thin_dupack)
2527 tcp_disable_early_retrans(tp);
2528 break;
2529
2530 case TCP_REPAIR:
2531 if (!tcp_can_repair_sock(sk))
2532 err = -EPERM;
2533 else if (val == 1) {
2534 tp->repair = 1;
2535 sk->sk_reuse = SK_FORCE_REUSE;
2536 tp->repair_queue = TCP_NO_QUEUE;
2537 } else if (val == 0) {
2538 tp->repair = 0;
2539 sk->sk_reuse = SK_NO_REUSE;
2540 tcp_send_window_probe(sk);
2541 } else
2542 err = -EINVAL;
2543
2544 break;
2545
2546 case TCP_REPAIR_QUEUE:
2547 if (!tp->repair)
2548 err = -EPERM;
2549 else if (val < TCP_QUEUES_NR)
2550 tp->repair_queue = val;
2551 else
2552 err = -EINVAL;
2553 break;
2554
2555 case TCP_QUEUE_SEQ:
2556 if (sk->sk_state != TCP_CLOSE)
2557 err = -EPERM;
2558 else if (tp->repair_queue == TCP_SEND_QUEUE)
2559 tp->write_seq = val;
2560 else if (tp->repair_queue == TCP_RECV_QUEUE)
2561 tp->rcv_nxt = val;
2562 else
2563 err = -EINVAL;
2564 break;
2565
2566 case TCP_REPAIR_OPTIONS:
2567 if (!tp->repair)
2568 err = -EINVAL;
2569 else if (sk->sk_state == TCP_ESTABLISHED)
2570 err = tcp_repair_options_est(tp,
2571 (struct tcp_repair_opt __user *)optval,
2572 optlen);
2573 else
2574 err = -EPERM;
2575 break;
2576
2577 case TCP_CORK:
2578 /* When set indicates to always queue non-full frames.
2579 * Later the user clears this option and we transmit
2580 * any pending partial frames in the queue. This is
2581 * meant to be used alongside sendfile() to get properly
2582 * filled frames when the user (for example) must write
2583 * out headers with a write() call first and then use
2584 * sendfile to send out the data parts.
2585 *
2586 * TCP_CORK can be set together with TCP_NODELAY and it is
2587 * stronger than TCP_NODELAY.
2588 */
2589 if (val) {
2590 tp->nonagle |= TCP_NAGLE_CORK;
2591 } else {
2592 tp->nonagle &= ~TCP_NAGLE_CORK;
2593 if (tp->nonagle&TCP_NAGLE_OFF)
2594 tp->nonagle |= TCP_NAGLE_PUSH;
2595 tcp_push_pending_frames(sk);
2596 }
2597 break;
2598
2599 case TCP_KEEPIDLE:
2600 if (val < 1 || val > MAX_TCP_KEEPIDLE)
2601 err = -EINVAL;
2602 else {
2603 tp->keepalive_time = val * HZ;
2604 if (sock_flag(sk, SOCK_KEEPOPEN) &&
2605 !((1 << sk->sk_state) &
2606 (TCPF_CLOSE | TCPF_LISTEN))) {
2607 u32 elapsed = keepalive_time_elapsed(tp);
2608 if (tp->keepalive_time > elapsed)
2609 elapsed = tp->keepalive_time - elapsed;
2610 else
2611 elapsed = 0;
2612 inet_csk_reset_keepalive_timer(sk, elapsed);
2613 }
2614 }
2615 break;
2616 case TCP_KEEPINTVL:
2617 if (val < 1 || val > MAX_TCP_KEEPINTVL)
2618 err = -EINVAL;
2619 else
2620 tp->keepalive_intvl = val * HZ;
2621 break;
2622 case TCP_KEEPCNT:
2623 if (val < 1 || val > MAX_TCP_KEEPCNT)
2624 err = -EINVAL;
2625 else
2626 tp->keepalive_probes = val;
2627 break;
2628 case TCP_SYNCNT:
2629 if (val < 1 || val > MAX_TCP_SYNCNT)
2630 err = -EINVAL;
2631 else
2632 icsk->icsk_syn_retries = val;
2633 break;
2634
2635 case TCP_LINGER2:
2636 if (val < 0)
2637 tp->linger2 = -1;
2638 else if (val > sysctl_tcp_fin_timeout / HZ)
2639 tp->linger2 = 0;
2640 else
2641 tp->linger2 = val * HZ;
2642 break;
2643
2644 case TCP_DEFER_ACCEPT:
2645 /* Translate value in seconds to number of retransmits */
2646 icsk->icsk_accept_queue.rskq_defer_accept =
2647 secs_to_retrans(val, TCP_TIMEOUT_INIT / HZ,
2648 TCP_RTO_MAX / HZ);
2649 break;
2650
2651 case TCP_WINDOW_CLAMP:
2652 if (!val) {
2653 if (sk->sk_state != TCP_CLOSE) {
2654 err = -EINVAL;
2655 break;
2656 }
2657 tp->window_clamp = 0;
2658 } else
2659 tp->window_clamp = val < SOCK_MIN_RCVBUF / 2 ?
2660 SOCK_MIN_RCVBUF / 2 : val;
2661 break;
2662
2663 case TCP_QUICKACK:
2664 if (!val) {
2665 icsk->icsk_ack.pingpong = 1;
2666 } else {
2667 icsk->icsk_ack.pingpong = 0;
2668 if ((1 << sk->sk_state) &
2669 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) &&
2670 inet_csk_ack_scheduled(sk)) {
2671 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
2672 tcp_cleanup_rbuf(sk, 1);
2673 if (!(val & 1))
2674 icsk->icsk_ack.pingpong = 1;
2675 }
2676 }
2677 break;
2678
2679 #ifdef CONFIG_TCP_MD5SIG
2680 case TCP_MD5SIG:
2681 /* Read the IP->Key mappings from userspace */
2682 err = tp->af_specific->md5_parse(sk, optval, optlen);
2683 break;
2684 #endif
2685 case TCP_USER_TIMEOUT:
2686 /* Cap the max timeout in ms TCP will retry/retrans
2687 * before giving up and aborting (ETIMEDOUT) a connection.
2688 */
2689 if (val < 0)
2690 err = -EINVAL;
2691 else
2692 icsk->icsk_user_timeout = msecs_to_jiffies(val);
2693 break;
2694
2695 case TCP_FASTOPEN:
2696 if (val >= 0 && ((1 << sk->sk_state) & (TCPF_CLOSE |
2697 TCPF_LISTEN)))
2698 err = fastopen_init_queue(sk, val);
2699 else
2700 err = -EINVAL;
2701 break;
2702 default:
2703 err = -ENOPROTOOPT;
2704 break;
2705 }
2706
2707 release_sock(sk);
2708 return err;
2709 }
2710
2711 int tcp_setsockopt(struct sock *sk, int level, int optname, char __user *optval,
2712 unsigned int optlen)
2713 {
2714 const struct inet_connection_sock *icsk = inet_csk(sk);
2715
2716 if (level != SOL_TCP)
2717 return icsk->icsk_af_ops->setsockopt(sk, level, optname,
2718 optval, optlen);
2719 return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2720 }
2721 EXPORT_SYMBOL(tcp_setsockopt);
2722
2723 #ifdef CONFIG_COMPAT
2724 int compat_tcp_setsockopt(struct sock *sk, int level, int optname,
2725 char __user *optval, unsigned int optlen)
2726 {
2727 if (level != SOL_TCP)
2728 return inet_csk_compat_setsockopt(sk, level, optname,
2729 optval, optlen);
2730 return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2731 }
2732 EXPORT_SYMBOL(compat_tcp_setsockopt);
2733 #endif
2734
2735 /* Return information about state of tcp endpoint in API format. */
2736 void tcp_get_info(const struct sock *sk, struct tcp_info *info)
2737 {
2738 const struct tcp_sock *tp = tcp_sk(sk);
2739 const struct inet_connection_sock *icsk = inet_csk(sk);
2740 u32 now = tcp_time_stamp;
2741
2742 memset(info, 0, sizeof(*info));
2743
2744 info->tcpi_state = sk->sk_state;
2745 info->tcpi_ca_state = icsk->icsk_ca_state;
2746 info->tcpi_retransmits = icsk->icsk_retransmits;
2747 info->tcpi_probes = icsk->icsk_probes_out;
2748 info->tcpi_backoff = icsk->icsk_backoff;
2749
2750 if (tp->rx_opt.tstamp_ok)
2751 info->tcpi_options |= TCPI_OPT_TIMESTAMPS;
2752 if (tcp_is_sack(tp))
2753 info->tcpi_options |= TCPI_OPT_SACK;
2754 if (tp->rx_opt.wscale_ok) {
2755 info->tcpi_options |= TCPI_OPT_WSCALE;
2756 info->tcpi_snd_wscale = tp->rx_opt.snd_wscale;
2757 info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale;
2758 }
2759
2760 if (tp->ecn_flags & TCP_ECN_OK)
2761 info->tcpi_options |= TCPI_OPT_ECN;
2762 if (tp->ecn_flags & TCP_ECN_SEEN)
2763 info->tcpi_options |= TCPI_OPT_ECN_SEEN;
2764 if (tp->syn_data_acked)
2765 info->tcpi_options |= TCPI_OPT_SYN_DATA;
2766
2767 info->tcpi_rto = jiffies_to_usecs(icsk->icsk_rto);
2768 info->tcpi_ato = jiffies_to_usecs(icsk->icsk_ack.ato);
2769 info->tcpi_snd_mss = tp->mss_cache;
2770 info->tcpi_rcv_mss = icsk->icsk_ack.rcv_mss;
2771
2772 if (sk->sk_state == TCP_LISTEN) {
2773 info->tcpi_unacked = sk->sk_ack_backlog;
2774 info->tcpi_sacked = sk->sk_max_ack_backlog;
2775 } else {
2776 info->tcpi_unacked = tp->packets_out;
2777 info->tcpi_sacked = tp->sacked_out;
2778 }
2779 info->tcpi_lost = tp->lost_out;
2780 info->tcpi_retrans = tp->retrans_out;
2781 info->tcpi_fackets = tp->fackets_out;
2782
2783 info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime);
2784 info->tcpi_last_data_recv = jiffies_to_msecs(now - icsk->icsk_ack.lrcvtime);
2785 info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp);
2786
2787 info->tcpi_pmtu = icsk->icsk_pmtu_cookie;
2788 info->tcpi_rcv_ssthresh = tp->rcv_ssthresh;
2789 info->tcpi_rtt = jiffies_to_usecs(tp->srtt)>>3;
2790 info->tcpi_rttvar = jiffies_to_usecs(tp->mdev)>>2;
2791 info->tcpi_snd_ssthresh = tp->snd_ssthresh;
2792 info->tcpi_snd_cwnd = tp->snd_cwnd;
2793 info->tcpi_advmss = tp->advmss;
2794 info->tcpi_reordering = tp->reordering;
2795
2796 info->tcpi_rcv_rtt = jiffies_to_usecs(tp->rcv_rtt_est.rtt)>>3;
2797 info->tcpi_rcv_space = tp->rcvq_space.space;
2798
2799 info->tcpi_total_retrans = tp->total_retrans;
2800 }
2801 EXPORT_SYMBOL_GPL(tcp_get_info);
2802
2803 static int do_tcp_getsockopt(struct sock *sk, int level,
2804 int optname, char __user *optval, int __user *optlen)
2805 {
2806 struct inet_connection_sock *icsk = inet_csk(sk);
2807 struct tcp_sock *tp = tcp_sk(sk);
2808 int val, len;
2809
2810 if (get_user(len, optlen))
2811 return -EFAULT;
2812
2813 len = min_t(unsigned int, len, sizeof(int));
2814
2815 if (len < 0)
2816 return -EINVAL;
2817
2818 switch (optname) {
2819 case TCP_MAXSEG:
2820 val = tp->mss_cache;
2821 if (!val && ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
2822 val = tp->rx_opt.user_mss;
2823 if (tp->repair)
2824 val = tp->rx_opt.mss_clamp;
2825 break;
2826 case TCP_NODELAY:
2827 val = !!(tp->nonagle&TCP_NAGLE_OFF);
2828 break;
2829 case TCP_CORK:
2830 val = !!(tp->nonagle&TCP_NAGLE_CORK);
2831 break;
2832 case TCP_KEEPIDLE:
2833 val = keepalive_time_when(tp) / HZ;
2834 break;
2835 case TCP_KEEPINTVL:
2836 val = keepalive_intvl_when(tp) / HZ;
2837 break;
2838 case TCP_KEEPCNT:
2839 val = keepalive_probes(tp);
2840 break;
2841 case TCP_SYNCNT:
2842 val = icsk->icsk_syn_retries ? : sysctl_tcp_syn_retries;
2843 break;
2844 case TCP_LINGER2:
2845 val = tp->linger2;
2846 if (val >= 0)
2847 val = (val ? : sysctl_tcp_fin_timeout) / HZ;
2848 break;
2849 case TCP_DEFER_ACCEPT:
2850 val = retrans_to_secs(icsk->icsk_accept_queue.rskq_defer_accept,
2851 TCP_TIMEOUT_INIT / HZ, TCP_RTO_MAX / HZ);
2852 break;
2853 case TCP_WINDOW_CLAMP:
2854 val = tp->window_clamp;
2855 break;
2856 case TCP_INFO: {
2857 struct tcp_info info;
2858
2859 if (get_user(len, optlen))
2860 return -EFAULT;
2861
2862 tcp_get_info(sk, &info);
2863
2864 len = min_t(unsigned int, len, sizeof(info));
2865 if (put_user(len, optlen))
2866 return -EFAULT;
2867 if (copy_to_user(optval, &info, len))
2868 return -EFAULT;
2869 return 0;
2870 }
2871 case TCP_QUICKACK:
2872 val = !icsk->icsk_ack.pingpong;
2873 break;
2874
2875 case TCP_CONGESTION:
2876 if (get_user(len, optlen))
2877 return -EFAULT;
2878 len = min_t(unsigned int, len, TCP_CA_NAME_MAX);
2879 if (put_user(len, optlen))
2880 return -EFAULT;
2881 if (copy_to_user(optval, icsk->icsk_ca_ops->name, len))
2882 return -EFAULT;
2883 return 0;
2884
2885 case TCP_COOKIE_TRANSACTIONS: {
2886 struct tcp_cookie_transactions ctd;
2887 struct tcp_cookie_values *cvp = tp->cookie_values;
2888
2889 if (get_user(len, optlen))
2890 return -EFAULT;
2891 if (len < sizeof(ctd))
2892 return -EINVAL;
2893
2894 memset(&ctd, 0, sizeof(ctd));
2895 ctd.tcpct_flags = (tp->rx_opt.cookie_in_always ?
2896 TCP_COOKIE_IN_ALWAYS : 0)
2897 | (tp->rx_opt.cookie_out_never ?
2898 TCP_COOKIE_OUT_NEVER : 0);
2899
2900 if (cvp != NULL) {
2901 ctd.tcpct_flags |= (cvp->s_data_in ?
2902 TCP_S_DATA_IN : 0)
2903 | (cvp->s_data_out ?
2904 TCP_S_DATA_OUT : 0);
2905
2906 ctd.tcpct_cookie_desired = cvp->cookie_desired;
2907 ctd.tcpct_s_data_desired = cvp->s_data_desired;
2908
2909 memcpy(&ctd.tcpct_value[0], &cvp->cookie_pair[0],
2910 cvp->cookie_pair_size);
2911 ctd.tcpct_used = cvp->cookie_pair_size;
2912 }
2913
2914 if (put_user(sizeof(ctd), optlen))
2915 return -EFAULT;
2916 if (copy_to_user(optval, &ctd, sizeof(ctd)))
2917 return -EFAULT;
2918 return 0;
2919 }
2920 case TCP_THIN_LINEAR_TIMEOUTS:
2921 val = tp->thin_lto;
2922 break;
2923 case TCP_THIN_DUPACK:
2924 val = tp->thin_dupack;
2925 break;
2926
2927 case TCP_REPAIR:
2928 val = tp->repair;
2929 break;
2930
2931 case TCP_REPAIR_QUEUE:
2932 if (tp->repair)
2933 val = tp->repair_queue;
2934 else
2935 return -EINVAL;
2936 break;
2937
2938 case TCP_QUEUE_SEQ:
2939 if (tp->repair_queue == TCP_SEND_QUEUE)
2940 val = tp->write_seq;
2941 else if (tp->repair_queue == TCP_RECV_QUEUE)
2942 val = tp->rcv_nxt;
2943 else
2944 return -EINVAL;
2945 break;
2946
2947 case TCP_USER_TIMEOUT:
2948 val = jiffies_to_msecs(icsk->icsk_user_timeout);
2949 break;
2950 default:
2951 return -ENOPROTOOPT;
2952 }
2953
2954 if (put_user(len, optlen))
2955 return -EFAULT;
2956 if (copy_to_user(optval, &val, len))
2957 return -EFAULT;
2958 return 0;
2959 }
2960
2961 int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval,
2962 int __user *optlen)
2963 {
2964 struct inet_connection_sock *icsk = inet_csk(sk);
2965
2966 if (level != SOL_TCP)
2967 return icsk->icsk_af_ops->getsockopt(sk, level, optname,
2968 optval, optlen);
2969 return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2970 }
2971 EXPORT_SYMBOL(tcp_getsockopt);
2972
2973 #ifdef CONFIG_COMPAT
2974 int compat_tcp_getsockopt(struct sock *sk, int level, int optname,
2975 char __user *optval, int __user *optlen)
2976 {
2977 if (level != SOL_TCP)
2978 return inet_csk_compat_getsockopt(sk, level, optname,
2979 optval, optlen);
2980 return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2981 }
2982 EXPORT_SYMBOL(compat_tcp_getsockopt);
2983 #endif
2984
2985 struct sk_buff *tcp_tso_segment(struct sk_buff *skb,
2986 netdev_features_t features)
2987 {
2988 struct sk_buff *segs = ERR_PTR(-EINVAL);
2989 struct tcphdr *th;
2990 unsigned int thlen;
2991 unsigned int seq;
2992 __be32 delta;
2993 unsigned int oldlen;
2994 unsigned int mss;
2995
2996 if (!pskb_may_pull(skb, sizeof(*th)))
2997 goto out;
2998
2999 th = tcp_hdr(skb);
3000 thlen = th->doff * 4;
3001 if (thlen < sizeof(*th))
3002 goto out;
3003
3004 if (!pskb_may_pull(skb, thlen))
3005 goto out;
3006
3007 oldlen = (u16)~skb->len;
3008 __skb_pull(skb, thlen);
3009
3010 mss = skb_shinfo(skb)->gso_size;
3011 if (unlikely(skb->len <= mss))
3012 goto out;
3013
3014 if (skb_gso_ok(skb, features | NETIF_F_GSO_ROBUST)) {
3015 /* Packet is from an untrusted source, reset gso_segs. */
3016 int type = skb_shinfo(skb)->gso_type;
3017
3018 if (unlikely(type &
3019 ~(SKB_GSO_TCPV4 |
3020 SKB_GSO_DODGY |
3021 SKB_GSO_TCP_ECN |
3022 SKB_GSO_TCPV6 |
3023 0) ||
3024 !(type & (SKB_GSO_TCPV4 | SKB_GSO_TCPV6))))
3025 goto out;
3026
3027 skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss);
3028
3029 segs = NULL;
3030 goto out;
3031 }
3032
3033 segs = skb_segment(skb, features);
3034 if (IS_ERR(segs))
3035 goto out;
3036
3037 delta = htonl(oldlen + (thlen + mss));
3038
3039 skb = segs;
3040 th = tcp_hdr(skb);
3041 seq = ntohl(th->seq);
3042
3043 do {
3044 th->fin = th->psh = 0;
3045
3046 th->check = ~csum_fold((__force __wsum)((__force u32)th->check +
3047 (__force u32)delta));
3048 if (skb->ip_summed != CHECKSUM_PARTIAL)
3049 th->check =
3050 csum_fold(csum_partial(skb_transport_header(skb),
3051 thlen, skb->csum));
3052
3053 seq += mss;
3054 skb = skb->next;
3055 th = tcp_hdr(skb);
3056
3057 th->seq = htonl(seq);
3058 th->cwr = 0;
3059 } while (skb->next);
3060
3061 delta = htonl(oldlen + (skb->tail - skb->transport_header) +
3062 skb->data_len);
3063 th->check = ~csum_fold((__force __wsum)((__force u32)th->check +
3064 (__force u32)delta));
3065 if (skb->ip_summed != CHECKSUM_PARTIAL)
3066 th->check = csum_fold(csum_partial(skb_transport_header(skb),
3067 thlen, skb->csum));
3068
3069 out:
3070 return segs;
3071 }
3072 EXPORT_SYMBOL(tcp_tso_segment);
3073
3074 struct sk_buff **tcp_gro_receive(struct sk_buff **head, struct sk_buff *skb)
3075 {
3076 struct sk_buff **pp = NULL;
3077 struct sk_buff *p;
3078 struct tcphdr *th;
3079 struct tcphdr *th2;
3080 unsigned int len;
3081 unsigned int thlen;
3082 __be32 flags;
3083 unsigned int mss = 1;
3084 unsigned int hlen;
3085 unsigned int off;
3086 int flush = 1;
3087 int i;
3088
3089 off = skb_gro_offset(skb);
3090 hlen = off + sizeof(*th);
3091 th = skb_gro_header_fast(skb, off);
3092 if (skb_gro_header_hard(skb, hlen)) {
3093 th = skb_gro_header_slow(skb, hlen, off);
3094 if (unlikely(!th))
3095 goto out;
3096 }
3097
3098 thlen = th->doff * 4;
3099 if (thlen < sizeof(*th))
3100 goto out;
3101
3102 hlen = off + thlen;
3103 if (skb_gro_header_hard(skb, hlen)) {
3104 th = skb_gro_header_slow(skb, hlen, off);
3105 if (unlikely(!th))
3106 goto out;
3107 }
3108
3109 skb_gro_pull(skb, thlen);
3110
3111 len = skb_gro_len(skb);
3112 flags = tcp_flag_word(th);
3113
3114 for (; (p = *head); head = &p->next) {
3115 if (!NAPI_GRO_CB(p)->same_flow)
3116 continue;
3117
3118 th2 = tcp_hdr(p);
3119
3120 if (*(u32 *)&th->source ^ *(u32 *)&th2->source) {
3121 NAPI_GRO_CB(p)->same_flow = 0;
3122 continue;
3123 }
3124
3125 goto found;
3126 }
3127
3128 goto out_check_final;
3129
3130 found:
3131 flush = NAPI_GRO_CB(p)->flush;
3132 flush |= (__force int)(flags & TCP_FLAG_CWR);
3133 flush |= (__force int)((flags ^ tcp_flag_word(th2)) &
3134 ~(TCP_FLAG_CWR | TCP_FLAG_FIN | TCP_FLAG_PSH));
3135 flush |= (__force int)(th->ack_seq ^ th2->ack_seq);
3136 for (i = sizeof(*th); i < thlen; i += 4)
3137 flush |= *(u32 *)((u8 *)th + i) ^
3138 *(u32 *)((u8 *)th2 + i);
3139
3140 mss = skb_shinfo(p)->gso_size;
3141
3142 flush |= (len - 1) >= mss;
3143 flush |= (ntohl(th2->seq) + skb_gro_len(p)) ^ ntohl(th->seq);
3144
3145 if (flush || skb_gro_receive(head, skb)) {
3146 mss = 1;
3147 goto out_check_final;
3148 }
3149
3150 p = *head;
3151 th2 = tcp_hdr(p);
3152 tcp_flag_word(th2) |= flags & (TCP_FLAG_FIN | TCP_FLAG_PSH);
3153
3154 out_check_final:
3155 flush = len < mss;
3156 flush |= (__force int)(flags & (TCP_FLAG_URG | TCP_FLAG_PSH |
3157 TCP_FLAG_RST | TCP_FLAG_SYN |
3158 TCP_FLAG_FIN));
3159
3160 if (p && (!NAPI_GRO_CB(skb)->same_flow || flush))
3161 pp = head;
3162
3163 out:
3164 NAPI_GRO_CB(skb)->flush |= flush;
3165
3166 return pp;
3167 }
3168 EXPORT_SYMBOL(tcp_gro_receive);
3169
3170 int tcp_gro_complete(struct sk_buff *skb)
3171 {
3172 struct tcphdr *th = tcp_hdr(skb);
3173
3174 skb->csum_start = skb_transport_header(skb) - skb->head;
3175 skb->csum_offset = offsetof(struct tcphdr, check);
3176 skb->ip_summed = CHECKSUM_PARTIAL;
3177
3178 skb_shinfo(skb)->gso_segs = NAPI_GRO_CB(skb)->count;
3179
3180 if (th->cwr)
3181 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
3182
3183 return 0;
3184 }
3185 EXPORT_SYMBOL(tcp_gro_complete);
3186
3187 #ifdef CONFIG_TCP_MD5SIG
3188 static unsigned long tcp_md5sig_users;
3189 static struct tcp_md5sig_pool __percpu *tcp_md5sig_pool;
3190 static DEFINE_SPINLOCK(tcp_md5sig_pool_lock);
3191
3192 static void __tcp_free_md5sig_pool(struct tcp_md5sig_pool __percpu *pool)
3193 {
3194 int cpu;
3195
3196 for_each_possible_cpu(cpu) {
3197 struct tcp_md5sig_pool *p = per_cpu_ptr(pool, cpu);
3198
3199 if (p->md5_desc.tfm)
3200 crypto_free_hash(p->md5_desc.tfm);
3201 }
3202 free_percpu(pool);
3203 }
3204
3205 void tcp_free_md5sig_pool(void)
3206 {
3207 struct tcp_md5sig_pool __percpu *pool = NULL;
3208
3209 spin_lock_bh(&tcp_md5sig_pool_lock);
3210 if (--tcp_md5sig_users == 0) {
3211 pool = tcp_md5sig_pool;
3212 tcp_md5sig_pool = NULL;
3213 }
3214 spin_unlock_bh(&tcp_md5sig_pool_lock);
3215 if (pool)
3216 __tcp_free_md5sig_pool(pool);
3217 }
3218 EXPORT_SYMBOL(tcp_free_md5sig_pool);
3219
3220 static struct tcp_md5sig_pool __percpu *
3221 __tcp_alloc_md5sig_pool(struct sock *sk)
3222 {
3223 int cpu;
3224 struct tcp_md5sig_pool __percpu *pool;
3225
3226 pool = alloc_percpu(struct tcp_md5sig_pool);
3227 if (!pool)
3228 return NULL;
3229
3230 for_each_possible_cpu(cpu) {
3231 struct crypto_hash *hash;
3232
3233 hash = crypto_alloc_hash("md5", 0, CRYPTO_ALG_ASYNC);
3234 if (!hash || IS_ERR(hash))
3235 goto out_free;
3236
3237 per_cpu_ptr(pool, cpu)->md5_desc.tfm = hash;
3238 }
3239 return pool;
3240 out_free:
3241 __tcp_free_md5sig_pool(pool);
3242 return NULL;
3243 }
3244
3245 struct tcp_md5sig_pool __percpu *tcp_alloc_md5sig_pool(struct sock *sk)
3246 {
3247 struct tcp_md5sig_pool __percpu *pool;
3248 bool alloc = false;
3249
3250 retry:
3251 spin_lock_bh(&tcp_md5sig_pool_lock);
3252 pool = tcp_md5sig_pool;
3253 if (tcp_md5sig_users++ == 0) {
3254 alloc = true;
3255 spin_unlock_bh(&tcp_md5sig_pool_lock);
3256 } else if (!pool) {
3257 tcp_md5sig_users--;
3258 spin_unlock_bh(&tcp_md5sig_pool_lock);
3259 cpu_relax();
3260 goto retry;
3261 } else
3262 spin_unlock_bh(&tcp_md5sig_pool_lock);
3263
3264 if (alloc) {
3265 /* we cannot hold spinlock here because this may sleep. */
3266 struct tcp_md5sig_pool __percpu *p;
3267
3268 p = __tcp_alloc_md5sig_pool(sk);
3269 spin_lock_bh(&tcp_md5sig_pool_lock);
3270 if (!p) {
3271 tcp_md5sig_users--;
3272 spin_unlock_bh(&tcp_md5sig_pool_lock);
3273 return NULL;
3274 }
3275 pool = tcp_md5sig_pool;
3276 if (pool) {
3277 /* oops, it has already been assigned. */
3278 spin_unlock_bh(&tcp_md5sig_pool_lock);
3279 __tcp_free_md5sig_pool(p);
3280 } else {
3281 tcp_md5sig_pool = pool = p;
3282 spin_unlock_bh(&tcp_md5sig_pool_lock);
3283 }
3284 }
3285 return pool;
3286 }
3287 EXPORT_SYMBOL(tcp_alloc_md5sig_pool);
3288
3289
3290 /**
3291 * tcp_get_md5sig_pool - get md5sig_pool for this user
3292 *
3293 * We use percpu structure, so if we succeed, we exit with preemption
3294 * and BH disabled, to make sure another thread or softirq handling
3295 * wont try to get same context.
3296 */
3297 struct tcp_md5sig_pool *tcp_get_md5sig_pool(void)
3298 {
3299 struct tcp_md5sig_pool __percpu *p;
3300
3301 local_bh_disable();
3302
3303 spin_lock(&tcp_md5sig_pool_lock);
3304 p = tcp_md5sig_pool;
3305 if (p)
3306 tcp_md5sig_users++;
3307 spin_unlock(&tcp_md5sig_pool_lock);
3308
3309 if (p)
3310 return this_cpu_ptr(p);
3311
3312 local_bh_enable();
3313 return NULL;
3314 }
3315 EXPORT_SYMBOL(tcp_get_md5sig_pool);
3316
3317 void tcp_put_md5sig_pool(void)
3318 {
3319 local_bh_enable();
3320 tcp_free_md5sig_pool();
3321 }
3322 EXPORT_SYMBOL(tcp_put_md5sig_pool);
3323
3324 int tcp_md5_hash_header(struct tcp_md5sig_pool *hp,
3325 const struct tcphdr *th)
3326 {
3327 struct scatterlist sg;
3328 struct tcphdr hdr;
3329 int err;
3330
3331 /* We are not allowed to change tcphdr, make a local copy */
3332 memcpy(&hdr, th, sizeof(hdr));
3333 hdr.check = 0;
3334
3335 /* options aren't included in the hash */
3336 sg_init_one(&sg, &hdr, sizeof(hdr));
3337 err = crypto_hash_update(&hp->md5_desc, &sg, sizeof(hdr));
3338 return err;
3339 }
3340 EXPORT_SYMBOL(tcp_md5_hash_header);
3341
3342 int tcp_md5_hash_skb_data(struct tcp_md5sig_pool *hp,
3343 const struct sk_buff *skb, unsigned int header_len)
3344 {
3345 struct scatterlist sg;
3346 const struct tcphdr *tp = tcp_hdr(skb);
3347 struct hash_desc *desc = &hp->md5_desc;
3348 unsigned int i;
3349 const unsigned int head_data_len = skb_headlen(skb) > header_len ?
3350 skb_headlen(skb) - header_len : 0;
3351 const struct skb_shared_info *shi = skb_shinfo(skb);
3352 struct sk_buff *frag_iter;
3353
3354 sg_init_table(&sg, 1);
3355
3356 sg_set_buf(&sg, ((u8 *) tp) + header_len, head_data_len);
3357 if (crypto_hash_update(desc, &sg, head_data_len))
3358 return 1;
3359
3360 for (i = 0; i < shi->nr_frags; ++i) {
3361 const struct skb_frag_struct *f = &shi->frags[i];
3362 struct page *page = skb_frag_page(f);
3363 sg_set_page(&sg, page, skb_frag_size(f), f->page_offset);
3364 if (crypto_hash_update(desc, &sg, skb_frag_size(f)))
3365 return 1;
3366 }
3367
3368 skb_walk_frags(skb, frag_iter)
3369 if (tcp_md5_hash_skb_data(hp, frag_iter, 0))
3370 return 1;
3371
3372 return 0;
3373 }
3374 EXPORT_SYMBOL(tcp_md5_hash_skb_data);
3375
3376 int tcp_md5_hash_key(struct tcp_md5sig_pool *hp, const struct tcp_md5sig_key *key)
3377 {
3378 struct scatterlist sg;
3379
3380 sg_init_one(&sg, key->key, key->keylen);
3381 return crypto_hash_update(&hp->md5_desc, &sg, key->keylen);
3382 }
3383 EXPORT_SYMBOL(tcp_md5_hash_key);
3384
3385 #endif
3386
3387 /* Each Responder maintains up to two secret values concurrently for
3388 * efficient secret rollover. Each secret value has 4 states:
3389 *
3390 * Generating. (tcp_secret_generating != tcp_secret_primary)
3391 * Generates new Responder-Cookies, but not yet used for primary
3392 * verification. This is a short-term state, typically lasting only
3393 * one round trip time (RTT).
3394 *
3395 * Primary. (tcp_secret_generating == tcp_secret_primary)
3396 * Used both for generation and primary verification.
3397 *
3398 * Retiring. (tcp_secret_retiring != tcp_secret_secondary)
3399 * Used for verification, until the first failure that can be
3400 * verified by the newer Generating secret. At that time, this
3401 * cookie's state is changed to Secondary, and the Generating
3402 * cookie's state is changed to Primary. This is a short-term state,
3403 * typically lasting only one round trip time (RTT).
3404 *
3405 * Secondary. (tcp_secret_retiring == tcp_secret_secondary)
3406 * Used for secondary verification, after primary verification
3407 * failures. This state lasts no more than twice the Maximum Segment
3408 * Lifetime (2MSL). Then, the secret is discarded.
3409 */
3410 struct tcp_cookie_secret {
3411 /* The secret is divided into two parts. The digest part is the
3412 * equivalent of previously hashing a secret and saving the state,
3413 * and serves as an initialization vector (IV). The message part
3414 * serves as the trailing secret.
3415 */
3416 u32 secrets[COOKIE_WORKSPACE_WORDS];
3417 unsigned long expires;
3418 };
3419
3420 #define TCP_SECRET_1MSL (HZ * TCP_PAWS_MSL)
3421 #define TCP_SECRET_2MSL (HZ * TCP_PAWS_MSL * 2)
3422 #define TCP_SECRET_LIFE (HZ * 600)
3423
3424 static struct tcp_cookie_secret tcp_secret_one;
3425 static struct tcp_cookie_secret tcp_secret_two;
3426
3427 /* Essentially a circular list, without dynamic allocation. */
3428 static struct tcp_cookie_secret *tcp_secret_generating;
3429 static struct tcp_cookie_secret *tcp_secret_primary;
3430 static struct tcp_cookie_secret *tcp_secret_retiring;
3431 static struct tcp_cookie_secret *tcp_secret_secondary;
3432
3433 static DEFINE_SPINLOCK(tcp_secret_locker);
3434
3435 /* Select a pseudo-random word in the cookie workspace.
3436 */
3437 static inline u32 tcp_cookie_work(const u32 *ws, const int n)
3438 {
3439 return ws[COOKIE_DIGEST_WORDS + ((COOKIE_MESSAGE_WORDS-1) & ws[n])];
3440 }
3441
3442 /* Fill bakery[COOKIE_WORKSPACE_WORDS] with generator, updating as needed.
3443 * Called in softirq context.
3444 * Returns: 0 for success.
3445 */
3446 int tcp_cookie_generator(u32 *bakery)
3447 {
3448 unsigned long jiffy = jiffies;
3449
3450 if (unlikely(time_after_eq(jiffy, tcp_secret_generating->expires))) {
3451 spin_lock_bh(&tcp_secret_locker);
3452 if (!time_after_eq(jiffy, tcp_secret_generating->expires)) {
3453 /* refreshed by another */
3454 memcpy(bakery,
3455 &tcp_secret_generating->secrets[0],
3456 COOKIE_WORKSPACE_WORDS);
3457 } else {
3458 /* still needs refreshing */
3459 get_random_bytes(bakery, COOKIE_WORKSPACE_WORDS);
3460
3461 /* The first time, paranoia assumes that the
3462 * randomization function isn't as strong. But,
3463 * this secret initialization is delayed until
3464 * the last possible moment (packet arrival).
3465 * Although that time is observable, it is
3466 * unpredictably variable. Mash in the most
3467 * volatile clock bits available, and expire the
3468 * secret extra quickly.
3469 */
3470 if (unlikely(tcp_secret_primary->expires ==
3471 tcp_secret_secondary->expires)) {
3472 struct timespec tv;
3473
3474 getnstimeofday(&tv);
3475 bakery[COOKIE_DIGEST_WORDS+0] ^=
3476 (u32)tv.tv_nsec;
3477
3478 tcp_secret_secondary->expires = jiffy
3479 + TCP_SECRET_1MSL
3480 + (0x0f & tcp_cookie_work(bakery, 0));
3481 } else {
3482 tcp_secret_secondary->expires = jiffy
3483 + TCP_SECRET_LIFE
3484 + (0xff & tcp_cookie_work(bakery, 1));
3485 tcp_secret_primary->expires = jiffy
3486 + TCP_SECRET_2MSL
3487 + (0x1f & tcp_cookie_work(bakery, 2));
3488 }
3489 memcpy(&tcp_secret_secondary->secrets[0],
3490 bakery, COOKIE_WORKSPACE_WORDS);
3491
3492 rcu_assign_pointer(tcp_secret_generating,
3493 tcp_secret_secondary);
3494 rcu_assign_pointer(tcp_secret_retiring,
3495 tcp_secret_primary);
3496 /*
3497 * Neither call_rcu() nor synchronize_rcu() needed.
3498 * Retiring data is not freed. It is replaced after
3499 * further (locked) pointer updates, and a quiet time
3500 * (minimum 1MSL, maximum LIFE - 2MSL).
3501 */
3502 }
3503 spin_unlock_bh(&tcp_secret_locker);
3504 } else {
3505 rcu_read_lock_bh();
3506 memcpy(bakery,
3507 &rcu_dereference(tcp_secret_generating)->secrets[0],
3508 COOKIE_WORKSPACE_WORDS);
3509 rcu_read_unlock_bh();
3510 }
3511 return 0;
3512 }
3513 EXPORT_SYMBOL(tcp_cookie_generator);
3514
3515 void tcp_done(struct sock *sk)
3516 {
3517 struct request_sock *req = tcp_sk(sk)->fastopen_rsk;
3518
3519 if (sk->sk_state == TCP_SYN_SENT || sk->sk_state == TCP_SYN_RECV)
3520 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_ATTEMPTFAILS);
3521
3522 tcp_set_state(sk, TCP_CLOSE);
3523 tcp_clear_xmit_timers(sk);
3524 if (req != NULL)
3525 reqsk_fastopen_remove(sk, req, false);
3526
3527 sk->sk_shutdown = SHUTDOWN_MASK;
3528
3529 if (!sock_flag(sk, SOCK_DEAD))
3530 sk->sk_state_change(sk);
3531 else
3532 inet_csk_destroy_sock(sk);
3533 }
3534 EXPORT_SYMBOL_GPL(tcp_done);
3535
3536 extern struct tcp_congestion_ops tcp_reno;
3537
3538 static __initdata unsigned long thash_entries;
3539 static int __init set_thash_entries(char *str)
3540 {
3541 ssize_t ret;
3542
3543 if (!str)
3544 return 0;
3545
3546 ret = kstrtoul(str, 0, &thash_entries);
3547 if (ret)
3548 return 0;
3549
3550 return 1;
3551 }
3552 __setup("thash_entries=", set_thash_entries);
3553
3554 void tcp_init_mem(struct net *net)
3555 {
3556 unsigned long limit = nr_free_buffer_pages() / 8;
3557 limit = max(limit, 128UL);
3558 net->ipv4.sysctl_tcp_mem[0] = limit / 4 * 3;
3559 net->ipv4.sysctl_tcp_mem[1] = limit;
3560 net->ipv4.sysctl_tcp_mem[2] = net->ipv4.sysctl_tcp_mem[0] * 2;
3561 }
3562
3563 void __init tcp_init(void)
3564 {
3565 struct sk_buff *skb = NULL;
3566 unsigned long limit;
3567 int max_rshare, max_wshare, cnt;
3568 unsigned int i;
3569 unsigned long jiffy = jiffies;
3570
3571 BUILD_BUG_ON(sizeof(struct tcp_skb_cb) > sizeof(skb->cb));
3572
3573 percpu_counter_init(&tcp_sockets_allocated, 0);
3574 percpu_counter_init(&tcp_orphan_count, 0);
3575 tcp_hashinfo.bind_bucket_cachep =
3576 kmem_cache_create("tcp_bind_bucket",
3577 sizeof(struct inet_bind_bucket), 0,
3578 SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL);
3579
3580 /* Size and allocate the main established and bind bucket
3581 * hash tables.
3582 *
3583 * The methodology is similar to that of the buffer cache.
3584 */
3585 tcp_hashinfo.ehash =
3586 alloc_large_system_hash("TCP established",
3587 sizeof(struct inet_ehash_bucket),
3588 thash_entries,
3589 (totalram_pages >= 128 * 1024) ?
3590 13 : 15,
3591 0,
3592 NULL,
3593 &tcp_hashinfo.ehash_mask,
3594 0,
3595 thash_entries ? 0 : 512 * 1024);
3596 for (i = 0; i <= tcp_hashinfo.ehash_mask; i++) {
3597 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].chain, i);
3598 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].twchain, i);
3599 }
3600 if (inet_ehash_locks_alloc(&tcp_hashinfo))
3601 panic("TCP: failed to alloc ehash_locks");
3602 tcp_hashinfo.bhash =
3603 alloc_large_system_hash("TCP bind",
3604 sizeof(struct inet_bind_hashbucket),
3605 tcp_hashinfo.ehash_mask + 1,
3606 (totalram_pages >= 128 * 1024) ?
3607 13 : 15,
3608 0,
3609 &tcp_hashinfo.bhash_size,
3610 NULL,
3611 0,
3612 64 * 1024);
3613 tcp_hashinfo.bhash_size = 1U << tcp_hashinfo.bhash_size;
3614 for (i = 0; i < tcp_hashinfo.bhash_size; i++) {
3615 spin_lock_init(&tcp_hashinfo.bhash[i].lock);
3616 INIT_HLIST_HEAD(&tcp_hashinfo.bhash[i].chain);
3617 }
3618
3619
3620 cnt = tcp_hashinfo.ehash_mask + 1;
3621
3622 tcp_death_row.sysctl_max_tw_buckets = cnt / 2;
3623 sysctl_tcp_max_orphans = cnt / 2;
3624 sysctl_max_syn_backlog = max(128, cnt / 256);
3625
3626 tcp_init_mem(&init_net);
3627 /* Set per-socket limits to no more than 1/128 the pressure threshold */
3628 limit = nr_free_buffer_pages() << (PAGE_SHIFT - 7);
3629 max_wshare = min(4UL*1024*1024, limit);
3630 max_rshare = min(6UL*1024*1024, limit);
3631
3632 sysctl_tcp_wmem[0] = SK_MEM_QUANTUM;
3633 sysctl_tcp_wmem[1] = 16*1024;
3634 sysctl_tcp_wmem[2] = max(64*1024, max_wshare);
3635
3636 sysctl_tcp_rmem[0] = SK_MEM_QUANTUM;
3637 sysctl_tcp_rmem[1] = 87380;
3638 sysctl_tcp_rmem[2] = max(87380, max_rshare);
3639
3640 pr_info("Hash tables configured (established %u bind %u)\n",
3641 tcp_hashinfo.ehash_mask + 1, tcp_hashinfo.bhash_size);
3642
3643 tcp_metrics_init();
3644
3645 tcp_register_congestion_control(&tcp_reno);
3646
3647 memset(&tcp_secret_one.secrets[0], 0, sizeof(tcp_secret_one.secrets));
3648 memset(&tcp_secret_two.secrets[0], 0, sizeof(tcp_secret_two.secrets));
3649 tcp_secret_one.expires = jiffy; /* past due */
3650 tcp_secret_two.expires = jiffy; /* past due */
3651 tcp_secret_generating = &tcp_secret_one;
3652 tcp_secret_primary = &tcp_secret_one;
3653 tcp_secret_retiring = &tcp_secret_two;
3654 tcp_secret_secondary = &tcp_secret_two;
3655 tcp_tasklet_init();
3656 }
Linux kernel - Deliverables
This page took 0.150127 seconds and 5 git commands to generate.