projects / deliverable / linux.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
blame | history | raw | HEAD
ALSA: hda_intel: add position_fix quirk for Asus K53E
[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? */
490 if ((1 << sk->sk_state) & ~(TCPF_SYN_SENT | TCPF_SYN_RECV)) {
491 int target = sock_rcvlowat(sk, 0, INT_MAX);
492
493 if (tp->urg_seq == tp->copied_seq &&
494 !sock_flag(sk, SOCK_URGINLINE) &&
495 tp->urg_data)
496 target++;
497
498 /* Potential race condition. If read of tp below will
499 * escape above sk->sk_state, we can be illegally awaken
500 * in SYN_* states. */
501 if (tp->rcv_nxt - tp->copied_seq >= target)
502 mask |= POLLIN | POLLRDNORM;
503
504 if (!(sk->sk_shutdown & SEND_SHUTDOWN)) {
505 if (sk_stream_wspace(sk) >= sk_stream_min_wspace(sk)) {
506 mask |= POLLOUT | POLLWRNORM;
507 } else { /* send SIGIO later */
508 set_bit(SOCK_ASYNC_NOSPACE,
509 &sk->sk_socket->flags);
510 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
511
512 /* Race breaker. If space is freed after
513 * wspace test but before the flags are set,
514 * IO signal will be lost.
515 */
516 if (sk_stream_wspace(sk) >= sk_stream_min_wspace(sk))
517 mask |= POLLOUT | POLLWRNORM;
518 }
519 } else
520 mask |= POLLOUT | POLLWRNORM;
521
522 if (tp->urg_data & TCP_URG_VALID)
523 mask |= POLLPRI;
524 }
525 /* This barrier is coupled with smp_wmb() in tcp_reset() */
526 smp_rmb();
527 if (sk->sk_err)
528 mask |= POLLERR;
529
530 return mask;
531 }
532 EXPORT_SYMBOL(tcp_poll);
533
534 int tcp_ioctl(struct sock *sk, int cmd, unsigned long arg)
535 {
536 struct tcp_sock *tp = tcp_sk(sk);
537 int answ;
538
539 switch (cmd) {
540 case SIOCINQ:
541 if (sk->sk_state == TCP_LISTEN)
542 return -EINVAL;
543
544 lock_sock(sk);
545 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
546 answ = 0;
547 else if (sock_flag(sk, SOCK_URGINLINE) ||
548 !tp->urg_data ||
549 before(tp->urg_seq, tp->copied_seq) ||
550 !before(tp->urg_seq, tp->rcv_nxt)) {
551 struct sk_buff *skb;
552
553 answ = tp->rcv_nxt - tp->copied_seq;
554
555 /* Subtract 1, if FIN is in queue. */
556 skb = skb_peek_tail(&sk->sk_receive_queue);
557 if (answ && skb)
558 answ -= tcp_hdr(skb)->fin;
559 } else
560 answ = tp->urg_seq - tp->copied_seq;
561 release_sock(sk);
562 break;
563 case SIOCATMARK:
564 answ = tp->urg_data && tp->urg_seq == tp->copied_seq;
565 break;
566 case SIOCOUTQ:
567 if (sk->sk_state == TCP_LISTEN)
568 return -EINVAL;
569
570 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
571 answ = 0;
572 else
573 answ = tp->write_seq - tp->snd_una;
574 break;
575 case SIOCOUTQNSD:
576 if (sk->sk_state == TCP_LISTEN)
577 return -EINVAL;
578
579 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
580 answ = 0;
581 else
582 answ = tp->write_seq - tp->snd_nxt;
583 break;
584 default:
585 return -ENOIOCTLCMD;
586 }
587
588 return put_user(answ, (int __user *)arg);
589 }
590 EXPORT_SYMBOL(tcp_ioctl);
591
592 static inline void tcp_mark_push(struct tcp_sock *tp, struct sk_buff *skb)
593 {
594 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
595 tp->pushed_seq = tp->write_seq;
596 }
597
598 static inline bool forced_push(const struct tcp_sock *tp)
599 {
600 return after(tp->write_seq, tp->pushed_seq + (tp->max_window >> 1));
601 }
602
603 static inline void skb_entail(struct sock *sk, struct sk_buff *skb)
604 {
605 struct tcp_sock *tp = tcp_sk(sk);
606 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
607
608 skb->csum = 0;
609 tcb->seq = tcb->end_seq = tp->write_seq;
610 tcb->tcp_flags = TCPHDR_ACK;
611 tcb->sacked = 0;
612 skb_header_release(skb);
613 tcp_add_write_queue_tail(sk, skb);
614 sk->sk_wmem_queued += skb->truesize;
615 sk_mem_charge(sk, skb->truesize);
616 if (tp->nonagle & TCP_NAGLE_PUSH)
617 tp->nonagle &= ~TCP_NAGLE_PUSH;
618 }
619
620 static inline void tcp_mark_urg(struct tcp_sock *tp, int flags)
621 {
622 if (flags & MSG_OOB)
623 tp->snd_up = tp->write_seq;
624 }
625
626 static inline void tcp_push(struct sock *sk, int flags, int mss_now,
627 int nonagle)
628 {
629 if (tcp_send_head(sk)) {
630 struct tcp_sock *tp = tcp_sk(sk);
631
632 if (!(flags & MSG_MORE) || forced_push(tp))
633 tcp_mark_push(tp, tcp_write_queue_tail(sk));
634
635 tcp_mark_urg(tp, flags);
636 __tcp_push_pending_frames(sk, mss_now,
637 (flags & MSG_MORE) ? TCP_NAGLE_CORK : nonagle);
638 }
639 }
640
641 static int tcp_splice_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb,
642 unsigned int offset, size_t len)
643 {
644 struct tcp_splice_state *tss = rd_desc->arg.data;
645 int ret;
646
647 ret = skb_splice_bits(skb, offset, tss->pipe, min(rd_desc->count, len),
648 tss->flags);
649 if (ret > 0)
650 rd_desc->count -= ret;
651 return ret;
652 }
653
654 static int __tcp_splice_read(struct sock *sk, struct tcp_splice_state *tss)
655 {
656 /* Store TCP splice context information in read_descriptor_t. */
657 read_descriptor_t rd_desc = {
658 .arg.data = tss,
659 .count = tss->len,
660 };
661
662 return tcp_read_sock(sk, &rd_desc, tcp_splice_data_recv);
663 }
664
665 /**
666 * tcp_splice_read - splice data from TCP socket to a pipe
667 * @sock: socket to splice from
668 * @ppos: position (not valid)
669 * @pipe: pipe to splice to
670 * @len: number of bytes to splice
671 * @flags: splice modifier flags
672 *
673 * Description:
674 * Will read pages from given socket and fill them into a pipe.
675 *
676 **/
677 ssize_t tcp_splice_read(struct socket *sock, loff_t *ppos,
678 struct pipe_inode_info *pipe, size_t len,
679 unsigned int flags)
680 {
681 struct sock *sk = sock->sk;
682 struct tcp_splice_state tss = {
683 .pipe = pipe,
684 .len = len,
685 .flags = flags,
686 };
687 long timeo;
688 ssize_t spliced;
689 int ret;
690
691 sock_rps_record_flow(sk);
692 /*
693 * We can't seek on a socket input
694 */
695 if (unlikely(*ppos))
696 return -ESPIPE;
697
698 ret = spliced = 0;
699
700 lock_sock(sk);
701
702 timeo = sock_rcvtimeo(sk, sock->file->f_flags & O_NONBLOCK);
703 while (tss.len) {
704 ret = __tcp_splice_read(sk, &tss);
705 if (ret < 0)
706 break;
707 else if (!ret) {
708 if (spliced)
709 break;
710 if (sock_flag(sk, SOCK_DONE))
711 break;
712 if (sk->sk_err) {
713 ret = sock_error(sk);
714 break;
715 }
716 if (sk->sk_shutdown & RCV_SHUTDOWN)
717 break;
718 if (sk->sk_state == TCP_CLOSE) {
719 /*
720 * This occurs when user tries to read
721 * from never connected socket.
722 */
723 if (!sock_flag(sk, SOCK_DONE))
724 ret = -ENOTCONN;
725 break;
726 }
727 if (!timeo) {
728 ret = -EAGAIN;
729 break;
730 }
731 sk_wait_data(sk, &timeo);
732 if (signal_pending(current)) {
733 ret = sock_intr_errno(timeo);
734 break;
735 }
736 continue;
737 }
738 tss.len -= ret;
739 spliced += ret;
740
741 if (!timeo)
742 break;
743 release_sock(sk);
744 lock_sock(sk);
745
746 if (sk->sk_err || sk->sk_state == TCP_CLOSE ||
747 (sk->sk_shutdown & RCV_SHUTDOWN) ||
748 signal_pending(current))
749 break;
750 }
751
752 release_sock(sk);
753
754 if (spliced)
755 return spliced;
756
757 return ret;
758 }
759 EXPORT_SYMBOL(tcp_splice_read);
760
761 struct sk_buff *sk_stream_alloc_skb(struct sock *sk, int size, gfp_t gfp)
762 {
763 struct sk_buff *skb;
764
765 /* The TCP header must be at least 32-bit aligned. */
766 size = ALIGN(size, 4);
767
768 skb = alloc_skb_fclone(size + sk->sk_prot->max_header, gfp);
769 if (skb) {
770 if (sk_wmem_schedule(sk, skb->truesize)) {
771 skb_reserve(skb, sk->sk_prot->max_header);
772 /*
773 * Make sure that we have exactly size bytes
774 * available to the caller, no more, no less.
775 */
776 skb->avail_size = size;
777 return skb;
778 }
779 __kfree_skb(skb);
780 } else {
781 sk->sk_prot->enter_memory_pressure(sk);
782 sk_stream_moderate_sndbuf(sk);
783 }
784 return NULL;
785 }
786
787 static unsigned int tcp_xmit_size_goal(struct sock *sk, u32 mss_now,
788 int large_allowed)
789 {
790 struct tcp_sock *tp = tcp_sk(sk);
791 u32 xmit_size_goal, old_size_goal;
792
793 xmit_size_goal = mss_now;
794
795 if (large_allowed && sk_can_gso(sk)) {
796 xmit_size_goal = ((sk->sk_gso_max_size - 1) -
797 inet_csk(sk)->icsk_af_ops->net_header_len -
798 inet_csk(sk)->icsk_ext_hdr_len -
799 tp->tcp_header_len);
800
801 /* TSQ : try to have two TSO segments in flight */
802 xmit_size_goal = min_t(u32, xmit_size_goal,
803 sysctl_tcp_limit_output_bytes >> 1);
804
805 xmit_size_goal = tcp_bound_to_half_wnd(tp, xmit_size_goal);
806
807 /* We try hard to avoid divides here */
808 old_size_goal = tp->xmit_size_goal_segs * mss_now;
809
810 if (likely(old_size_goal <= xmit_size_goal &&
811 old_size_goal + mss_now > xmit_size_goal)) {
812 xmit_size_goal = old_size_goal;
813 } else {
814 tp->xmit_size_goal_segs = xmit_size_goal / mss_now;
815 xmit_size_goal = tp->xmit_size_goal_segs * mss_now;
816 }
817 }
818
819 return max(xmit_size_goal, mss_now);
820 }
821
822 static int tcp_send_mss(struct sock *sk, int *size_goal, int flags)
823 {
824 int mss_now;
825
826 mss_now = tcp_current_mss(sk);
827 *size_goal = tcp_xmit_size_goal(sk, mss_now, !(flags & MSG_OOB));
828
829 return mss_now;
830 }
831
832 static ssize_t do_tcp_sendpages(struct sock *sk, struct page **pages, int poffset,
833 size_t psize, int flags)
834 {
835 struct tcp_sock *tp = tcp_sk(sk);
836 int mss_now, size_goal;
837 int err;
838 ssize_t copied;
839 long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
840
841 /* Wait for a connection to finish. */
842 if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
843 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
844 goto out_err;
845
846 clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
847
848 mss_now = tcp_send_mss(sk, &size_goal, flags);
849 copied = 0;
850
851 err = -EPIPE;
852 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
853 goto out_err;
854
855 while (psize > 0) {
856 struct sk_buff *skb = tcp_write_queue_tail(sk);
857 struct page *page = pages[poffset / PAGE_SIZE];
858 int copy, i;
859 int offset = poffset % PAGE_SIZE;
860 int size = min_t(size_t, psize, PAGE_SIZE - offset);
861 bool can_coalesce;
862
863 if (!tcp_send_head(sk) || (copy = size_goal - skb->len) <= 0) {
864 new_segment:
865 if (!sk_stream_memory_free(sk))
866 goto wait_for_sndbuf;
867
868 skb = sk_stream_alloc_skb(sk, 0, sk->sk_allocation);
869 if (!skb)
870 goto wait_for_memory;
871
872 skb_entail(sk, skb);
873 copy = size_goal;
874 }
875
876 if (copy > size)
877 copy = size;
878
879 i = skb_shinfo(skb)->nr_frags;
880 can_coalesce = skb_can_coalesce(skb, i, page, offset);
881 if (!can_coalesce && i >= MAX_SKB_FRAGS) {
882 tcp_mark_push(tp, skb);
883 goto new_segment;
884 }
885 if (!sk_wmem_schedule(sk, copy))
886 goto wait_for_memory;
887
888 if (can_coalesce) {
889 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
890 } else {
891 get_page(page);
892 skb_fill_page_desc(skb, i, page, offset, copy);
893 }
894
895 skb->len += copy;
896 skb->data_len += copy;
897 skb->truesize += copy;
898 sk->sk_wmem_queued += copy;
899 sk_mem_charge(sk, copy);
900 skb->ip_summed = CHECKSUM_PARTIAL;
901 tp->write_seq += copy;
902 TCP_SKB_CB(skb)->end_seq += copy;
903 skb_shinfo(skb)->gso_segs = 0;
904
905 if (!copied)
906 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
907
908 copied += copy;
909 poffset += copy;
910 if (!(psize -= copy))
911 goto out;
912
913 if (skb->len < size_goal || (flags & MSG_OOB))
914 continue;
915
916 if (forced_push(tp)) {
917 tcp_mark_push(tp, skb);
918 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
919 } else if (skb == tcp_send_head(sk))
920 tcp_push_one(sk, mss_now);
921 continue;
922
923 wait_for_sndbuf:
924 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
925 wait_for_memory:
926 tcp_push(sk, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH);
927
928 if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
929 goto do_error;
930
931 mss_now = tcp_send_mss(sk, &size_goal, flags);
932 }
933
934 out:
935 if (copied && !(flags & MSG_SENDPAGE_NOTLAST))
936 tcp_push(sk, flags, mss_now, tp->nonagle);
937 return copied;
938
939 do_error:
940 if (copied)
941 goto out;
942 out_err:
943 return sk_stream_error(sk, flags, err);
944 }
945
946 int tcp_sendpage(struct sock *sk, struct page *page, int offset,
947 size_t size, int flags)
948 {
949 ssize_t res;
950
951 if (!(sk->sk_route_caps & NETIF_F_SG) ||
952 !(sk->sk_route_caps & NETIF_F_ALL_CSUM))
953 return sock_no_sendpage(sk->sk_socket, page, offset, size,
954 flags);
955
956 lock_sock(sk);
957 res = do_tcp_sendpages(sk, &page, offset, size, flags);
958 release_sock(sk);
959 return res;
960 }
961 EXPORT_SYMBOL(tcp_sendpage);
962
963 static inline int select_size(const struct sock *sk, bool sg)
964 {
965 const struct tcp_sock *tp = tcp_sk(sk);
966 int tmp = tp->mss_cache;
967
968 if (sg) {
969 if (sk_can_gso(sk)) {
970 /* Small frames wont use a full page:
971 * Payload will immediately follow tcp header.
972 */
973 tmp = SKB_WITH_OVERHEAD(2048 - MAX_TCP_HEADER);
974 } else {
975 int pgbreak = SKB_MAX_HEAD(MAX_TCP_HEADER);
976
977 if (tmp >= pgbreak &&
978 tmp <= pgbreak + (MAX_SKB_FRAGS - 1) * PAGE_SIZE)
979 tmp = pgbreak;
980 }
981 }
982
983 return tmp;
984 }
985
986 void tcp_free_fastopen_req(struct tcp_sock *tp)
987 {
988 if (tp->fastopen_req != NULL) {
989 kfree(tp->fastopen_req);
990 tp->fastopen_req = NULL;
991 }
992 }
993
994 static int tcp_sendmsg_fastopen(struct sock *sk, struct msghdr *msg, int *size)
995 {
996 struct tcp_sock *tp = tcp_sk(sk);
997 int err, flags;
998
999 if (!(sysctl_tcp_fastopen & TFO_CLIENT_ENABLE))
1000 return -EOPNOTSUPP;
1001 if (tp->fastopen_req != NULL)
1002 return -EALREADY; /* Another Fast Open is in progress */
1003
1004 tp->fastopen_req = kzalloc(sizeof(struct tcp_fastopen_request),
1005 sk->sk_allocation);
1006 if (unlikely(tp->fastopen_req == NULL))
1007 return -ENOBUFS;
1008 tp->fastopen_req->data = msg;
1009
1010 flags = (msg->msg_flags & MSG_DONTWAIT) ? O_NONBLOCK : 0;
1011 err = __inet_stream_connect(sk->sk_socket, msg->msg_name,
1012 msg->msg_namelen, flags);
1013 *size = tp->fastopen_req->copied;
1014 tcp_free_fastopen_req(tp);
1015 return err;
1016 }
1017
1018 int tcp_sendmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg,
1019 size_t size)
1020 {
1021 struct iovec *iov;
1022 struct tcp_sock *tp = tcp_sk(sk);
1023 struct sk_buff *skb;
1024 int iovlen, flags, err, copied = 0;
1025 int mss_now = 0, size_goal, copied_syn = 0, offset = 0;
1026 bool sg;
1027 long timeo;
1028
1029 lock_sock(sk);
1030
1031 flags = msg->msg_flags;
1032 if (flags & MSG_FASTOPEN) {
1033 err = tcp_sendmsg_fastopen(sk, msg, &copied_syn);
1034 if (err == -EINPROGRESS && copied_syn > 0)
1035 goto out;
1036 else if (err)
1037 goto out_err;
1038 offset = copied_syn;
1039 }
1040
1041 timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
1042
1043 /* Wait for a connection to finish. */
1044 if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
1045 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
1046 goto do_error;
1047
1048 if (unlikely(tp->repair)) {
1049 if (tp->repair_queue == TCP_RECV_QUEUE) {
1050 copied = tcp_send_rcvq(sk, msg, size);
1051 goto out;
1052 }
1053
1054 err = -EINVAL;
1055 if (tp->repair_queue == TCP_NO_QUEUE)
1056 goto out_err;
1057
1058 /* 'common' sending to sendq */
1059 }
1060
1061 /* This should be in poll */
1062 clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
1063
1064 mss_now = tcp_send_mss(sk, &size_goal, flags);
1065
1066 /* Ok commence sending. */
1067 iovlen = msg->msg_iovlen;
1068 iov = msg->msg_iov;
1069 copied = 0;
1070
1071 err = -EPIPE;
1072 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
1073 goto out_err;
1074
1075 sg = !!(sk->sk_route_caps & NETIF_F_SG);
1076
1077 while (--iovlen >= 0) {
1078 size_t seglen = iov->iov_len;
1079 unsigned char __user *from = iov->iov_base;
1080
1081 iov++;
1082 if (unlikely(offset > 0)) { /* Skip bytes copied in SYN */
1083 if (offset >= seglen) {
1084 offset -= seglen;
1085 continue;
1086 }
1087 seglen -= offset;
1088 from += offset;
1089 offset = 0;
1090 }
1091
1092 while (seglen > 0) {
1093 int copy = 0;
1094 int max = size_goal;
1095
1096 skb = tcp_write_queue_tail(sk);
1097 if (tcp_send_head(sk)) {
1098 if (skb->ip_summed == CHECKSUM_NONE)
1099 max = mss_now;
1100 copy = max - skb->len;
1101 }
1102
1103 if (copy <= 0) {
1104 new_segment:
1105 /* Allocate new segment. If the interface is SG,
1106 * allocate skb fitting to single page.
1107 */
1108 if (!sk_stream_memory_free(sk))
1109 goto wait_for_sndbuf;
1110
1111 skb = sk_stream_alloc_skb(sk,
1112 select_size(sk, sg),
1113 sk->sk_allocation);
1114 if (!skb)
1115 goto wait_for_memory;
1116
1117 /*
1118 * Check whether we can use HW checksum.
1119 */
1120 if (sk->sk_route_caps & NETIF_F_ALL_CSUM)
1121 skb->ip_summed = CHECKSUM_PARTIAL;
1122
1123 skb_entail(sk, skb);
1124 copy = size_goal;
1125 max = size_goal;
1126 }
1127
1128 /* Try to append data to the end of skb. */
1129 if (copy > seglen)
1130 copy = seglen;
1131
1132 /* Where to copy to? */
1133 if (skb_availroom(skb) > 0) {
1134 /* We have some space in skb head. Superb! */
1135 copy = min_t(int, copy, skb_availroom(skb));
1136 err = skb_add_data_nocache(sk, skb, from, copy);
1137 if (err)
1138 goto do_fault;
1139 } else {
1140 bool merge = false;
1141 int i = skb_shinfo(skb)->nr_frags;
1142 struct page *page = sk->sk_sndmsg_page;
1143 int off;
1144
1145 if (page && page_count(page) == 1)
1146 sk->sk_sndmsg_off = 0;
1147
1148 off = sk->sk_sndmsg_off;
1149
1150 if (skb_can_coalesce(skb, i, page, off) &&
1151 off != PAGE_SIZE) {
1152 /* We can extend the last page
1153 * fragment. */
1154 merge = true;
1155 } else if (i == MAX_SKB_FRAGS || !sg) {
1156 /* Need to add new fragment and cannot
1157 * do this because interface is non-SG,
1158 * or because all the page slots are
1159 * busy. */
1160 tcp_mark_push(tp, skb);
1161 goto new_segment;
1162 } else if (page) {
1163 if (off == PAGE_SIZE) {
1164 put_page(page);
1165 sk->sk_sndmsg_page = page = NULL;
1166 off = 0;
1167 }
1168 } else
1169 off = 0;
1170
1171 if (copy > PAGE_SIZE - off)
1172 copy = PAGE_SIZE - off;
1173
1174 if (!sk_wmem_schedule(sk, copy))
1175 goto wait_for_memory;
1176
1177 if (!page) {
1178 /* Allocate new cache page. */
1179 if (!(page = sk_stream_alloc_page(sk)))
1180 goto wait_for_memory;
1181 }
1182
1183 /* Time to copy data. We are close to
1184 * the end! */
1185 err = skb_copy_to_page_nocache(sk, from, skb,
1186 page, off, copy);
1187 if (err) {
1188 /* If this page was new, give it to the
1189 * socket so it does not get leaked.
1190 */
1191 if (!sk->sk_sndmsg_page) {
1192 sk->sk_sndmsg_page = page;
1193 sk->sk_sndmsg_off = 0;
1194 }
1195 goto do_error;
1196 }
1197
1198 /* Update the skb. */
1199 if (merge) {
1200 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1201 } else {
1202 skb_fill_page_desc(skb, i, page, off, copy);
1203 if (sk->sk_sndmsg_page) {
1204 get_page(page);
1205 } else if (off + copy < PAGE_SIZE) {
1206 get_page(page);
1207 sk->sk_sndmsg_page = page;
1208 }
1209 }
1210
1211 sk->sk_sndmsg_off = off + copy;
1212 }
1213
1214 if (!copied)
1215 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
1216
1217 tp->write_seq += copy;
1218 TCP_SKB_CB(skb)->end_seq += copy;
1219 skb_shinfo(skb)->gso_segs = 0;
1220
1221 from += copy;
1222 copied += copy;
1223 if ((seglen -= copy) == 0 && iovlen == 0)
1224 goto out;
1225
1226 if (skb->len < max || (flags & MSG_OOB) || unlikely(tp->repair))
1227 continue;
1228
1229 if (forced_push(tp)) {
1230 tcp_mark_push(tp, skb);
1231 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
1232 } else if (skb == tcp_send_head(sk))
1233 tcp_push_one(sk, mss_now);
1234 continue;
1235
1236 wait_for_sndbuf:
1237 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1238 wait_for_memory:
1239 if (copied && likely(!tp->repair))
1240 tcp_push(sk, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH);
1241
1242 if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
1243 goto do_error;
1244
1245 mss_now = tcp_send_mss(sk, &size_goal, flags);
1246 }
1247 }
1248
1249 out:
1250 if (copied && likely(!tp->repair))
1251 tcp_push(sk, flags, mss_now, tp->nonagle);
1252 release_sock(sk);
1253 return copied + copied_syn;
1254
1255 do_fault:
1256 if (!skb->len) {
1257 tcp_unlink_write_queue(skb, sk);
1258 /* It is the one place in all of TCP, except connection
1259 * reset, where we can be unlinking the send_head.
1260 */
1261 tcp_check_send_head(sk, skb);
1262 sk_wmem_free_skb(sk, skb);
1263 }
1264
1265 do_error:
1266 if (copied + copied_syn)
1267 goto out;
1268 out_err:
1269 err = sk_stream_error(sk, flags, err);
1270 release_sock(sk);
1271 return err;
1272 }
1273 EXPORT_SYMBOL(tcp_sendmsg);
1274
1275 /*
1276 * Handle reading urgent data. BSD has very simple semantics for
1277 * this, no blocking and very strange errors 8)
1278 */
1279
1280 static int tcp_recv_urg(struct sock *sk, struct msghdr *msg, int len, int flags)
1281 {
1282 struct tcp_sock *tp = tcp_sk(sk);
1283
1284 /* No URG data to read. */
1285 if (sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data ||
1286 tp->urg_data == TCP_URG_READ)
1287 return -EINVAL; /* Yes this is right ! */
1288
1289 if (sk->sk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DONE))
1290 return -ENOTCONN;
1291
1292 if (tp->urg_data & TCP_URG_VALID) {
1293 int err = 0;
1294 char c = tp->urg_data;
1295
1296 if (!(flags & MSG_PEEK))
1297 tp->urg_data = TCP_URG_READ;
1298
1299 /* Read urgent data. */
1300 msg->msg_flags |= MSG_OOB;
1301
1302 if (len > 0) {
1303 if (!(flags & MSG_TRUNC))
1304 err = memcpy_toiovec(msg->msg_iov, &c, 1);
1305 len = 1;
1306 } else
1307 msg->msg_flags |= MSG_TRUNC;
1308
1309 return err ? -EFAULT : len;
1310 }
1311
1312 if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
1313 return 0;
1314
1315 /* Fixed the recv(..., MSG_OOB) behaviour. BSD docs and
1316 * the available implementations agree in this case:
1317 * this call should never block, independent of the
1318 * blocking state of the socket.
1319 * Mike <pall@rz.uni-karlsruhe.de>
1320 */
1321 return -EAGAIN;
1322 }
1323
1324 static int tcp_peek_sndq(struct sock *sk, struct msghdr *msg, int len)
1325 {
1326 struct sk_buff *skb;
1327 int copied = 0, err = 0;
1328
1329 /* XXX -- need to support SO_PEEK_OFF */
1330
1331 skb_queue_walk(&sk->sk_write_queue, skb) {
1332 err = skb_copy_datagram_iovec(skb, 0, msg->msg_iov, skb->len);
1333 if (err)
1334 break;
1335
1336 copied += skb->len;
1337 }
1338
1339 return err ?: copied;
1340 }
1341
1342 /* Clean up the receive buffer for full frames taken by the user,
1343 * then send an ACK if necessary. COPIED is the number of bytes
1344 * tcp_recvmsg has given to the user so far, it speeds up the
1345 * calculation of whether or not we must ACK for the sake of
1346 * a window update.
1347 */
1348 void tcp_cleanup_rbuf(struct sock *sk, int copied)
1349 {
1350 struct tcp_sock *tp = tcp_sk(sk);
1351 bool time_to_ack = false;
1352
1353 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
1354
1355 WARN(skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq),
1356 "cleanup rbuf bug: copied %X seq %X rcvnxt %X\n",
1357 tp->copied_seq, TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt);
1358
1359 if (inet_csk_ack_scheduled(sk)) {
1360 const struct inet_connection_sock *icsk = inet_csk(sk);
1361 /* Delayed ACKs frequently hit locked sockets during bulk
1362 * receive. */
1363 if (icsk->icsk_ack.blocked ||
1364 /* Once-per-two-segments ACK was not sent by tcp_input.c */
1365 tp->rcv_nxt - tp->rcv_wup > icsk->icsk_ack.rcv_mss ||
1366 /*
1367 * If this read emptied read buffer, we send ACK, if
1368 * connection is not bidirectional, user drained
1369 * receive buffer and there was a small segment
1370 * in queue.
1371 */
1372 (copied > 0 &&
1373 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED2) ||
1374 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED) &&
1375 !icsk->icsk_ack.pingpong)) &&
1376 !atomic_read(&sk->sk_rmem_alloc)))
1377 time_to_ack = true;
1378 }
1379
1380 /* We send an ACK if we can now advertise a non-zero window
1381 * which has been raised "significantly".
1382 *
1383 * Even if window raised up to infinity, do not send window open ACK
1384 * in states, where we will not receive more. It is useless.
1385 */
1386 if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) {
1387 __u32 rcv_window_now = tcp_receive_window(tp);
1388
1389 /* Optimize, __tcp_select_window() is not cheap. */
1390 if (2*rcv_window_now <= tp->window_clamp) {
1391 __u32 new_window = __tcp_select_window(sk);
1392
1393 /* Send ACK now, if this read freed lots of space
1394 * in our buffer. Certainly, new_window is new window.
1395 * We can advertise it now, if it is not less than current one.
1396 * "Lots" means "at least twice" here.
1397 */
1398 if (new_window && new_window >= 2 * rcv_window_now)
1399 time_to_ack = true;
1400 }
1401 }
1402 if (time_to_ack)
1403 tcp_send_ack(sk);
1404 }
1405
1406 static void tcp_prequeue_process(struct sock *sk)
1407 {
1408 struct sk_buff *skb;
1409 struct tcp_sock *tp = tcp_sk(sk);
1410
1411 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPPREQUEUED);
1412
1413 /* RX process wants to run with disabled BHs, though it is not
1414 * necessary */
1415 local_bh_disable();
1416 while ((skb = __skb_dequeue(&tp->ucopy.prequeue)) != NULL)
1417 sk_backlog_rcv(sk, skb);
1418 local_bh_enable();
1419
1420 /* Clear memory counter. */
1421 tp->ucopy.memory = 0;
1422 }
1423
1424 #ifdef CONFIG_NET_DMA
1425 static void tcp_service_net_dma(struct sock *sk, bool wait)
1426 {
1427 dma_cookie_t done, used;
1428 dma_cookie_t last_issued;
1429 struct tcp_sock *tp = tcp_sk(sk);
1430
1431 if (!tp->ucopy.dma_chan)
1432 return;
1433
1434 last_issued = tp->ucopy.dma_cookie;
1435 dma_async_memcpy_issue_pending(tp->ucopy.dma_chan);
1436
1437 do {
1438 if (dma_async_memcpy_complete(tp->ucopy.dma_chan,
1439 last_issued, &done,
1440 &used) == DMA_SUCCESS) {
1441 /* Safe to free early-copied skbs now */
1442 __skb_queue_purge(&sk->sk_async_wait_queue);
1443 break;
1444 } else {
1445 struct sk_buff *skb;
1446 while ((skb = skb_peek(&sk->sk_async_wait_queue)) &&
1447 (dma_async_is_complete(skb->dma_cookie, done,
1448 used) == DMA_SUCCESS)) {
1449 __skb_dequeue(&sk->sk_async_wait_queue);
1450 kfree_skb(skb);
1451 }
1452 }
1453 } while (wait);
1454 }
1455 #endif
1456
1457 static inline struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off)
1458 {
1459 struct sk_buff *skb;
1460 u32 offset;
1461
1462 skb_queue_walk(&sk->sk_receive_queue, skb) {
1463 offset = seq - TCP_SKB_CB(skb)->seq;
1464 if (tcp_hdr(skb)->syn)
1465 offset--;
1466 if (offset < skb->len || tcp_hdr(skb)->fin) {
1467 *off = offset;
1468 return skb;
1469 }
1470 }
1471 return NULL;
1472 }
1473
1474 /*
1475 * This routine provides an alternative to tcp_recvmsg() for routines
1476 * that would like to handle copying from skbuffs directly in 'sendfile'
1477 * fashion.
1478 * Note:
1479 * - It is assumed that the socket was locked by the caller.
1480 * - The routine does not block.
1481 * - At present, there is no support for reading OOB data
1482 * or for 'peeking' the socket using this routine
1483 * (although both would be easy to implement).
1484 */
1485 int tcp_read_sock(struct sock *sk, read_descriptor_t *desc,
1486 sk_read_actor_t recv_actor)
1487 {
1488 struct sk_buff *skb;
1489 struct tcp_sock *tp = tcp_sk(sk);
1490 u32 seq = tp->copied_seq;
1491 u32 offset;
1492 int copied = 0;
1493
1494 if (sk->sk_state == TCP_LISTEN)
1495 return -ENOTCONN;
1496 while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1497 if (offset < skb->len) {
1498 int used;
1499 size_t len;
1500
1501 len = skb->len - offset;
1502 /* Stop reading if we hit a patch of urgent data */
1503 if (tp->urg_data) {
1504 u32 urg_offset = tp->urg_seq - seq;
1505 if (urg_offset < len)
1506 len = urg_offset;
1507 if (!len)
1508 break;
1509 }
1510 used = recv_actor(desc, skb, offset, len);
1511 if (used < 0) {
1512 if (!copied)
1513 copied = used;
1514 break;
1515 } else if (used <= len) {
1516 seq += used;
1517 copied += used;
1518 offset += used;
1519 }
1520 /*
1521 * If recv_actor drops the lock (e.g. TCP splice
1522 * receive) the skb pointer might be invalid when
1523 * getting here: tcp_collapse might have deleted it
1524 * while aggregating skbs from the socket queue.
1525 */
1526 skb = tcp_recv_skb(sk, seq-1, &offset);
1527 if (!skb || (offset+1 != skb->len))
1528 break;
1529 }
1530 if (tcp_hdr(skb)->fin) {
1531 sk_eat_skb(sk, skb, false);
1532 ++seq;
1533 break;
1534 }
1535 sk_eat_skb(sk, skb, false);
1536 if (!desc->count)
1537 break;
1538 tp->copied_seq = seq;
1539 }
1540 tp->copied_seq = seq;
1541
1542 tcp_rcv_space_adjust(sk);
1543
1544 /* Clean up data we have read: This will do ACK frames. */
1545 if (copied > 0)
1546 tcp_cleanup_rbuf(sk, copied);
1547 return copied;
1548 }
1549 EXPORT_SYMBOL(tcp_read_sock);
1550
1551 /*
1552 * This routine copies from a sock struct into the user buffer.
1553 *
1554 * Technical note: in 2.3 we work on _locked_ socket, so that
1555 * tricks with *seq access order and skb->users are not required.
1556 * Probably, code can be easily improved even more.
1557 */
1558
1559 int tcp_recvmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg,
1560 size_t len, int nonblock, int flags, int *addr_len)
1561 {
1562 struct tcp_sock *tp = tcp_sk(sk);
1563 int copied = 0;
1564 u32 peek_seq;
1565 u32 *seq;
1566 unsigned long used;
1567 int err;
1568 int target; /* Read at least this many bytes */
1569 long timeo;
1570 struct task_struct *user_recv = NULL;
1571 bool copied_early = false;
1572 struct sk_buff *skb;
1573 u32 urg_hole = 0;
1574
1575 lock_sock(sk);
1576
1577 err = -ENOTCONN;
1578 if (sk->sk_state == TCP_LISTEN)
1579 goto out;
1580
1581 timeo = sock_rcvtimeo(sk, nonblock);
1582
1583 /* Urgent data needs to be handled specially. */
1584 if (flags & MSG_OOB)
1585 goto recv_urg;
1586
1587 if (unlikely(tp->repair)) {
1588 err = -EPERM;
1589 if (!(flags & MSG_PEEK))
1590 goto out;
1591
1592 if (tp->repair_queue == TCP_SEND_QUEUE)
1593 goto recv_sndq;
1594
1595 err = -EINVAL;
1596 if (tp->repair_queue == TCP_NO_QUEUE)
1597 goto out;
1598
1599 /* 'common' recv queue MSG_PEEK-ing */
1600 }
1601
1602 seq = &tp->copied_seq;
1603 if (flags & MSG_PEEK) {
1604 peek_seq = tp->copied_seq;
1605 seq = &peek_seq;
1606 }
1607
1608 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
1609
1610 #ifdef CONFIG_NET_DMA
1611 tp->ucopy.dma_chan = NULL;
1612 preempt_disable();
1613 skb = skb_peek_tail(&sk->sk_receive_queue);
1614 {
1615 int available = 0;
1616
1617 if (skb)
1618 available = TCP_SKB_CB(skb)->seq + skb->len - (*seq);
1619 if ((available < target) &&
1620 (len > sysctl_tcp_dma_copybreak) && !(flags & MSG_PEEK) &&
1621 !sysctl_tcp_low_latency &&
1622 net_dma_find_channel()) {
1623 preempt_enable_no_resched();
1624 tp->ucopy.pinned_list =
1625 dma_pin_iovec_pages(msg->msg_iov, len);
1626 } else {
1627 preempt_enable_no_resched();
1628 }
1629 }
1630 #endif
1631
1632 do {
1633 u32 offset;
1634
1635 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
1636 if (tp->urg_data && tp->urg_seq == *seq) {
1637 if (copied)
1638 break;
1639 if (signal_pending(current)) {
1640 copied = timeo ? sock_intr_errno(timeo) : -EAGAIN;
1641 break;
1642 }
1643 }
1644
1645 /* Next get a buffer. */
1646
1647 skb_queue_walk(&sk->sk_receive_queue, skb) {
1648 /* Now that we have two receive queues this
1649 * shouldn't happen.
1650 */
1651 if (WARN(before(*seq, TCP_SKB_CB(skb)->seq),
1652 "recvmsg bug: copied %X seq %X rcvnxt %X fl %X\n",
1653 *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt,
1654 flags))
1655 break;
1656
1657 offset = *seq - TCP_SKB_CB(skb)->seq;
1658 if (tcp_hdr(skb)->syn)
1659 offset--;
1660 if (offset < skb->len)
1661 goto found_ok_skb;
1662 if (tcp_hdr(skb)->fin)
1663 goto found_fin_ok;
1664 WARN(!(flags & MSG_PEEK),
1665 "recvmsg bug 2: copied %X seq %X rcvnxt %X fl %X\n",
1666 *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt, flags);
1667 }
1668
1669 /* Well, if we have backlog, try to process it now yet. */
1670
1671 if (copied >= target && !sk->sk_backlog.tail)
1672 break;
1673
1674 if (copied) {
1675 if (sk->sk_err ||
1676 sk->sk_state == TCP_CLOSE ||
1677 (sk->sk_shutdown & RCV_SHUTDOWN) ||
1678 !timeo ||
1679 signal_pending(current))
1680 break;
1681 } else {
1682 if (sock_flag(sk, SOCK_DONE))
1683 break;
1684
1685 if (sk->sk_err) {
1686 copied = sock_error(sk);
1687 break;
1688 }
1689
1690 if (sk->sk_shutdown & RCV_SHUTDOWN)
1691 break;
1692
1693 if (sk->sk_state == TCP_CLOSE) {
1694 if (!sock_flag(sk, SOCK_DONE)) {
1695 /* This occurs when user tries to read
1696 * from never connected socket.
1697 */
1698 copied = -ENOTCONN;
1699 break;
1700 }
1701 break;
1702 }
1703
1704 if (!timeo) {
1705 copied = -EAGAIN;
1706 break;
1707 }
1708
1709 if (signal_pending(current)) {
1710 copied = sock_intr_errno(timeo);
1711 break;
1712 }
1713 }
1714
1715 tcp_cleanup_rbuf(sk, copied);
1716
1717 if (!sysctl_tcp_low_latency && tp->ucopy.task == user_recv) {
1718 /* Install new reader */
1719 if (!user_recv && !(flags & (MSG_TRUNC | MSG_PEEK))) {
1720 user_recv = current;
1721 tp->ucopy.task = user_recv;
1722 tp->ucopy.iov = msg->msg_iov;
1723 }
1724
1725 tp->ucopy.len = len;
1726
1727 WARN_ON(tp->copied_seq != tp->rcv_nxt &&
1728 !(flags & (MSG_PEEK | MSG_TRUNC)));
1729
1730 /* Ugly... If prequeue is not empty, we have to
1731 * process it before releasing socket, otherwise
1732 * order will be broken at second iteration.
1733 * More elegant solution is required!!!
1734 *
1735 * Look: we have the following (pseudo)queues:
1736 *
1737 * 1. packets in flight
1738 * 2. backlog
1739 * 3. prequeue
1740 * 4. receive_queue
1741 *
1742 * Each queue can be processed only if the next ones
1743 * are empty. At this point we have empty receive_queue.
1744 * But prequeue _can_ be not empty after 2nd iteration,
1745 * when we jumped to start of loop because backlog
1746 * processing added something to receive_queue.
1747 * We cannot release_sock(), because backlog contains
1748 * packets arrived _after_ prequeued ones.
1749 *
1750 * Shortly, algorithm is clear --- to process all
1751 * the queues in order. We could make it more directly,
1752 * requeueing packets from backlog to prequeue, if
1753 * is not empty. It is more elegant, but eats cycles,
1754 * unfortunately.
1755 */
1756 if (!skb_queue_empty(&tp->ucopy.prequeue))
1757 goto do_prequeue;
1758
1759 /* __ Set realtime policy in scheduler __ */
1760 }
1761
1762 #ifdef CONFIG_NET_DMA
1763 if (tp->ucopy.dma_chan)
1764 dma_async_memcpy_issue_pending(tp->ucopy.dma_chan);
1765 #endif
1766 if (copied >= target) {
1767 /* Do not sleep, just process backlog. */
1768 release_sock(sk);
1769 lock_sock(sk);
1770 } else
1771 sk_wait_data(sk, &timeo);
1772
1773 #ifdef CONFIG_NET_DMA
1774 tcp_service_net_dma(sk, false); /* Don't block */
1775 tp->ucopy.wakeup = 0;
1776 #endif
1777
1778 if (user_recv) {
1779 int chunk;
1780
1781 /* __ Restore normal policy in scheduler __ */
1782
1783 if ((chunk = len - tp->ucopy.len) != 0) {
1784 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMBACKLOG, chunk);
1785 len -= chunk;
1786 copied += chunk;
1787 }
1788
1789 if (tp->rcv_nxt == tp->copied_seq &&
1790 !skb_queue_empty(&tp->ucopy.prequeue)) {
1791 do_prequeue:
1792 tcp_prequeue_process(sk);
1793
1794 if ((chunk = len - tp->ucopy.len) != 0) {
1795 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1796 len -= chunk;
1797 copied += chunk;
1798 }
1799 }
1800 }
1801 if ((flags & MSG_PEEK) &&
1802 (peek_seq - copied - urg_hole != tp->copied_seq)) {
1803 net_dbg_ratelimited("TCP(%s:%d): Application bug, race in MSG_PEEK\n",
1804 current->comm,
1805 task_pid_nr(current));
1806 peek_seq = tp->copied_seq;
1807 }
1808 continue;
1809
1810 found_ok_skb:
1811 /* Ok so how much can we use? */
1812 used = skb->len - offset;
1813 if (len < used)
1814 used = len;
1815
1816 /* Do we have urgent data here? */
1817 if (tp->urg_data) {
1818 u32 urg_offset = tp->urg_seq - *seq;
1819 if (urg_offset < used) {
1820 if (!urg_offset) {
1821 if (!sock_flag(sk, SOCK_URGINLINE)) {
1822 ++*seq;
1823 urg_hole++;
1824 offset++;
1825 used--;
1826 if (!used)
1827 goto skip_copy;
1828 }
1829 } else
1830 used = urg_offset;
1831 }
1832 }
1833
1834 if (!(flags & MSG_TRUNC)) {
1835 #ifdef CONFIG_NET_DMA
1836 if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list)
1837 tp->ucopy.dma_chan = net_dma_find_channel();
1838
1839 if (tp->ucopy.dma_chan) {
1840 tp->ucopy.dma_cookie = dma_skb_copy_datagram_iovec(
1841 tp->ucopy.dma_chan, skb, offset,
1842 msg->msg_iov, used,
1843 tp->ucopy.pinned_list);
1844
1845 if (tp->ucopy.dma_cookie < 0) {
1846
1847 pr_alert("%s: dma_cookie < 0\n",
1848 __func__);
1849
1850 /* Exception. Bailout! */
1851 if (!copied)
1852 copied = -EFAULT;
1853 break;
1854 }
1855
1856 dma_async_memcpy_issue_pending(tp->ucopy.dma_chan);
1857
1858 if ((offset + used) == skb->len)
1859 copied_early = true;
1860
1861 } else
1862 #endif
1863 {
1864 err = skb_copy_datagram_iovec(skb, offset,
1865 msg->msg_iov, used);
1866 if (err) {
1867 /* Exception. Bailout! */
1868 if (!copied)
1869 copied = -EFAULT;
1870 break;
1871 }
1872 }
1873 }
1874
1875 *seq += used;
1876 copied += used;
1877 len -= used;
1878
1879 tcp_rcv_space_adjust(sk);
1880
1881 skip_copy:
1882 if (tp->urg_data && after(tp->copied_seq, tp->urg_seq)) {
1883 tp->urg_data = 0;
1884 tcp_fast_path_check(sk);
1885 }
1886 if (used + offset < skb->len)
1887 continue;
1888
1889 if (tcp_hdr(skb)->fin)
1890 goto found_fin_ok;
1891 if (!(flags & MSG_PEEK)) {
1892 sk_eat_skb(sk, skb, copied_early);
1893 copied_early = false;
1894 }
1895 continue;
1896
1897 found_fin_ok:
1898 /* Process the FIN. */
1899 ++*seq;
1900 if (!(flags & MSG_PEEK)) {
1901 sk_eat_skb(sk, skb, copied_early);
1902 copied_early = false;
1903 }
1904 break;
1905 } while (len > 0);
1906
1907 if (user_recv) {
1908 if (!skb_queue_empty(&tp->ucopy.prequeue)) {
1909 int chunk;
1910
1911 tp->ucopy.len = copied > 0 ? len : 0;
1912
1913 tcp_prequeue_process(sk);
1914
1915 if (copied > 0 && (chunk = len - tp->ucopy.len) != 0) {
1916 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1917 len -= chunk;
1918 copied += chunk;
1919 }
1920 }
1921
1922 tp->ucopy.task = NULL;
1923 tp->ucopy.len = 0;
1924 }
1925
1926 #ifdef CONFIG_NET_DMA
1927 tcp_service_net_dma(sk, true); /* Wait for queue to drain */
1928 tp->ucopy.dma_chan = NULL;
1929
1930 if (tp->ucopy.pinned_list) {
1931 dma_unpin_iovec_pages(tp->ucopy.pinned_list);
1932 tp->ucopy.pinned_list = NULL;
1933 }
1934 #endif
1935
1936 /* According to UNIX98, msg_name/msg_namelen are ignored
1937 * on connected socket. I was just happy when found this 8) --ANK
1938 */
1939
1940 /* Clean up data we have read: This will do ACK frames. */
1941 tcp_cleanup_rbuf(sk, copied);
1942
1943 release_sock(sk);
1944 return copied;
1945
1946 out:
1947 release_sock(sk);
1948 return err;
1949
1950 recv_urg:
1951 err = tcp_recv_urg(sk, msg, len, flags);
1952 goto out;
1953
1954 recv_sndq:
1955 err = tcp_peek_sndq(sk, msg, len);
1956 goto out;
1957 }
1958 EXPORT_SYMBOL(tcp_recvmsg);
1959
1960 void tcp_set_state(struct sock *sk, int state)
1961 {
1962 int oldstate = sk->sk_state;
1963
1964 switch (state) {
1965 case TCP_ESTABLISHED:
1966 if (oldstate != TCP_ESTABLISHED)
1967 TCP_INC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
1968 break;
1969
1970 case TCP_CLOSE:
1971 if (oldstate == TCP_CLOSE_WAIT || oldstate == TCP_ESTABLISHED)
1972 TCP_INC_STATS(sock_net(sk), TCP_MIB_ESTABRESETS);
1973
1974 sk->sk_prot->unhash(sk);
1975 if (inet_csk(sk)->icsk_bind_hash &&
1976 !(sk->sk_userlocks & SOCK_BINDPORT_LOCK))
1977 inet_put_port(sk);
1978 /* fall through */
1979 default:
1980 if (oldstate == TCP_ESTABLISHED)
1981 TCP_DEC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
1982 }
1983
1984 /* Change state AFTER socket is unhashed to avoid closed
1985 * socket sitting in hash tables.
1986 */
1987 sk->sk_state = state;
1988
1989 #ifdef STATE_TRACE
1990 SOCK_DEBUG(sk, "TCP sk=%p, State %s -> %s\n", sk, statename[oldstate], statename[state]);
1991 #endif
1992 }
1993 EXPORT_SYMBOL_GPL(tcp_set_state);
1994
1995 /*
1996 * State processing on a close. This implements the state shift for
1997 * sending our FIN frame. Note that we only send a FIN for some
1998 * states. A shutdown() may have already sent the FIN, or we may be
1999 * closed.
2000 */
2001
2002 static const unsigned char new_state[16] = {
2003 /* current state: new state: action: */
2004 /* (Invalid) */ TCP_CLOSE,
2005 /* TCP_ESTABLISHED */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2006 /* TCP_SYN_SENT */ TCP_CLOSE,
2007 /* TCP_SYN_RECV */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2008 /* TCP_FIN_WAIT1 */ TCP_FIN_WAIT1,
2009 /* TCP_FIN_WAIT2 */ TCP_FIN_WAIT2,
2010 /* TCP_TIME_WAIT */ TCP_CLOSE,
2011 /* TCP_CLOSE */ TCP_CLOSE,
2012 /* TCP_CLOSE_WAIT */ TCP_LAST_ACK | TCP_ACTION_FIN,
2013 /* TCP_LAST_ACK */ TCP_LAST_ACK,
2014 /* TCP_LISTEN */ TCP_CLOSE,
2015 /* TCP_CLOSING */ TCP_CLOSING,
2016 };
2017
2018 static int tcp_close_state(struct sock *sk)
2019 {
2020 int next = (int)new_state[sk->sk_state];
2021 int ns = next & TCP_STATE_MASK;
2022
2023 tcp_set_state(sk, ns);
2024
2025 return next & TCP_ACTION_FIN;
2026 }
2027
2028 /*
2029 * Shutdown the sending side of a connection. Much like close except
2030 * that we don't receive shut down or sock_set_flag(sk, SOCK_DEAD).
2031 */
2032
2033 void tcp_shutdown(struct sock *sk, int how)
2034 {
2035 /* We need to grab some memory, and put together a FIN,
2036 * and then put it into the queue to be sent.
2037 * Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
2038 */
2039 if (!(how & SEND_SHUTDOWN))
2040 return;
2041
2042 /* If we've already sent a FIN, or it's a closed state, skip this. */
2043 if ((1 << sk->sk_state) &
2044 (TCPF_ESTABLISHED | TCPF_SYN_SENT |
2045 TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) {
2046 /* Clear out any half completed packets. FIN if needed. */
2047 if (tcp_close_state(sk))
2048 tcp_send_fin(sk);
2049 }
2050 }
2051 EXPORT_SYMBOL(tcp_shutdown);
2052
2053 bool tcp_check_oom(struct sock *sk, int shift)
2054 {
2055 bool too_many_orphans, out_of_socket_memory;
2056
2057 too_many_orphans = tcp_too_many_orphans(sk, shift);
2058 out_of_socket_memory = tcp_out_of_memory(sk);
2059
2060 if (too_many_orphans)
2061 net_info_ratelimited("too many orphaned sockets\n");
2062 if (out_of_socket_memory)
2063 net_info_ratelimited("out of memory -- consider tuning tcp_mem\n");
2064 return too_many_orphans || out_of_socket_memory;
2065 }
2066
2067 void tcp_close(struct sock *sk, long timeout)
2068 {
2069 struct sk_buff *skb;
2070 int data_was_unread = 0;
2071 int state;
2072
2073 lock_sock(sk);
2074 sk->sk_shutdown = SHUTDOWN_MASK;
2075
2076 if (sk->sk_state == TCP_LISTEN) {
2077 tcp_set_state(sk, TCP_CLOSE);
2078
2079 /* Special case. */
2080 inet_csk_listen_stop(sk);
2081
2082 goto adjudge_to_death;
2083 }
2084
2085 /* We need to flush the recv. buffs. We do this only on the
2086 * descriptor close, not protocol-sourced closes, because the
2087 * reader process may not have drained the data yet!
2088 */
2089 while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) {
2090 u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq -
2091 tcp_hdr(skb)->fin;
2092 data_was_unread += len;
2093 __kfree_skb(skb);
2094 }
2095
2096 sk_mem_reclaim(sk);
2097
2098 /* If socket has been already reset (e.g. in tcp_reset()) - kill it. */
2099 if (sk->sk_state == TCP_CLOSE)
2100 goto adjudge_to_death;
2101
2102 /* As outlined in RFC 2525, section 2.17, we send a RST here because
2103 * data was lost. To witness the awful effects of the old behavior of
2104 * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk
2105 * GET in an FTP client, suspend the process, wait for the client to
2106 * advertise a zero window, then kill -9 the FTP client, wheee...
2107 * Note: timeout is always zero in such a case.
2108 */
2109 if (unlikely(tcp_sk(sk)->repair)) {
2110 sk->sk_prot->disconnect(sk, 0);
2111 } else if (data_was_unread) {
2112 /* Unread data was tossed, zap the connection. */
2113 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONCLOSE);
2114 tcp_set_state(sk, TCP_CLOSE);
2115 tcp_send_active_reset(sk, sk->sk_allocation);
2116 } else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) {
2117 /* Check zero linger _after_ checking for unread data. */
2118 sk->sk_prot->disconnect(sk, 0);
2119 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
2120 } else if (tcp_close_state(sk)) {
2121 /* We FIN if the application ate all the data before
2122 * zapping the connection.
2123 */
2124
2125 /* RED-PEN. Formally speaking, we have broken TCP state
2126 * machine. State transitions:
2127 *
2128 * TCP_ESTABLISHED -> TCP_FIN_WAIT1
2129 * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible)
2130 * TCP_CLOSE_WAIT -> TCP_LAST_ACK
2131 *
2132 * are legal only when FIN has been sent (i.e. in window),
2133 * rather than queued out of window. Purists blame.
2134 *
2135 * F.e. "RFC state" is ESTABLISHED,
2136 * if Linux state is FIN-WAIT-1, but FIN is still not sent.
2137 *
2138 * The visible declinations are that sometimes
2139 * we enter time-wait state, when it is not required really
2140 * (harmless), do not send active resets, when they are
2141 * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
2142 * they look as CLOSING or LAST_ACK for Linux)
2143 * Probably, I missed some more holelets.
2144 * --ANK
2145 */
2146 tcp_send_fin(sk);
2147 }
2148
2149 sk_stream_wait_close(sk, timeout);
2150
2151 adjudge_to_death:
2152 state = sk->sk_state;
2153 sock_hold(sk);
2154 sock_orphan(sk);
2155
2156 /* It is the last release_sock in its life. It will remove backlog. */
2157 release_sock(sk);
2158
2159
2160 /* Now socket is owned by kernel and we acquire BH lock
2161 to finish close. No need to check for user refs.
2162 */
2163 local_bh_disable();
2164 bh_lock_sock(sk);
2165 WARN_ON(sock_owned_by_user(sk));
2166
2167 percpu_counter_inc(sk->sk_prot->orphan_count);
2168
2169 /* Have we already been destroyed by a softirq or backlog? */
2170 if (state != TCP_CLOSE && sk->sk_state == TCP_CLOSE)
2171 goto out;
2172
2173 /* This is a (useful) BSD violating of the RFC. There is a
2174 * problem with TCP as specified in that the other end could
2175 * keep a socket open forever with no application left this end.
2176 * We use a 3 minute timeout (about the same as BSD) then kill
2177 * our end. If they send after that then tough - BUT: long enough
2178 * that we won't make the old 4*rto = almost no time - whoops
2179 * reset mistake.
2180 *
2181 * Nope, it was not mistake. It is really desired behaviour
2182 * f.e. on http servers, when such sockets are useless, but
2183 * consume significant resources. Let's do it with special
2184 * linger2 option. --ANK
2185 */
2186
2187 if (sk->sk_state == TCP_FIN_WAIT2) {
2188 struct tcp_sock *tp = tcp_sk(sk);
2189 if (tp->linger2 < 0) {
2190 tcp_set_state(sk, TCP_CLOSE);
2191 tcp_send_active_reset(sk, GFP_ATOMIC);
2192 NET_INC_STATS_BH(sock_net(sk),
2193 LINUX_MIB_TCPABORTONLINGER);
2194 } else {
2195 const int tmo = tcp_fin_time(sk);
2196
2197 if (tmo > TCP_TIMEWAIT_LEN) {
2198 inet_csk_reset_keepalive_timer(sk,
2199 tmo - TCP_TIMEWAIT_LEN);
2200 } else {
2201 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
2202 goto out;
2203 }
2204 }
2205 }
2206 if (sk->sk_state != TCP_CLOSE) {
2207 sk_mem_reclaim(sk);
2208 if (tcp_check_oom(sk, 0)) {
2209 tcp_set_state(sk, TCP_CLOSE);
2210 tcp_send_active_reset(sk, GFP_ATOMIC);
2211 NET_INC_STATS_BH(sock_net(sk),
2212 LINUX_MIB_TCPABORTONMEMORY);
2213 }
2214 }
2215
2216 if (sk->sk_state == TCP_CLOSE)
2217 inet_csk_destroy_sock(sk);
2218 /* Otherwise, socket is reprieved until protocol close. */
2219
2220 out:
2221 bh_unlock_sock(sk);
2222 local_bh_enable();
2223 sock_put(sk);
2224 }
2225 EXPORT_SYMBOL(tcp_close);
2226
2227 /* These states need RST on ABORT according to RFC793 */
2228
2229 static inline bool tcp_need_reset(int state)
2230 {
2231 return (1 << state) &
2232 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 |
2233 TCPF_FIN_WAIT2 | TCPF_SYN_RECV);
2234 }
2235
2236 int tcp_disconnect(struct sock *sk, int flags)
2237 {
2238 struct inet_sock *inet = inet_sk(sk);
2239 struct inet_connection_sock *icsk = inet_csk(sk);
2240 struct tcp_sock *tp = tcp_sk(sk);
2241 int err = 0;
2242 int old_state = sk->sk_state;
2243
2244 if (old_state != TCP_CLOSE)
2245 tcp_set_state(sk, TCP_CLOSE);
2246
2247 /* ABORT function of RFC793 */
2248 if (old_state == TCP_LISTEN) {
2249 inet_csk_listen_stop(sk);
2250 } else if (unlikely(tp->repair)) {
2251 sk->sk_err = ECONNABORTED;
2252 } else if (tcp_need_reset(old_state) ||
2253 (tp->snd_nxt != tp->write_seq &&
2254 (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK))) {
2255 /* The last check adjusts for discrepancy of Linux wrt. RFC
2256 * states
2257 */
2258 tcp_send_active_reset(sk, gfp_any());
2259 sk->sk_err = ECONNRESET;
2260 } else if (old_state == TCP_SYN_SENT)
2261 sk->sk_err = ECONNRESET;
2262
2263 tcp_clear_xmit_timers(sk);
2264 __skb_queue_purge(&sk->sk_receive_queue);
2265 tcp_write_queue_purge(sk);
2266 __skb_queue_purge(&tp->out_of_order_queue);
2267 #ifdef CONFIG_NET_DMA
2268 __skb_queue_purge(&sk->sk_async_wait_queue);
2269 #endif
2270
2271 inet->inet_dport = 0;
2272
2273 if (!(sk->sk_userlocks & SOCK_BINDADDR_LOCK))
2274 inet_reset_saddr(sk);
2275
2276 sk->sk_shutdown = 0;
2277 sock_reset_flag(sk, SOCK_DONE);
2278 tp->srtt = 0;
2279 if ((tp->write_seq += tp->max_window + 2) == 0)
2280 tp->write_seq = 1;
2281 icsk->icsk_backoff = 0;
2282 tp->snd_cwnd = 2;
2283 icsk->icsk_probes_out = 0;
2284 tp->packets_out = 0;
2285 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
2286 tp->snd_cwnd_cnt = 0;
2287 tp->bytes_acked = 0;
2288 tp->window_clamp = 0;
2289 tcp_set_ca_state(sk, TCP_CA_Open);
2290 tcp_clear_retrans(tp);
2291 inet_csk_delack_init(sk);
2292 tcp_init_send_head(sk);
2293 memset(&tp->rx_opt, 0, sizeof(tp->rx_opt));
2294 __sk_dst_reset(sk);
2295
2296 WARN_ON(inet->inet_num && !icsk->icsk_bind_hash);
2297
2298 sk->sk_error_report(sk);
2299 return err;
2300 }
2301 EXPORT_SYMBOL(tcp_disconnect);
2302
2303 static inline bool tcp_can_repair_sock(const struct sock *sk)
2304 {
2305 return capable(CAP_NET_ADMIN) &&
2306 ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_ESTABLISHED));
2307 }
2308
2309 static int tcp_repair_options_est(struct tcp_sock *tp,
2310 struct tcp_repair_opt __user *optbuf, unsigned int len)
2311 {
2312 struct tcp_repair_opt opt;
2313
2314 while (len >= sizeof(opt)) {
2315 if (copy_from_user(&opt, optbuf, sizeof(opt)))
2316 return -EFAULT;
2317
2318 optbuf++;
2319 len -= sizeof(opt);
2320
2321 switch (opt.opt_code) {
2322 case TCPOPT_MSS:
2323 tp->rx_opt.mss_clamp = opt.opt_val;
2324 break;
2325 case TCPOPT_WINDOW:
2326 if (opt.opt_val > 14)
2327 return -EFBIG;
2328
2329 tp->rx_opt.snd_wscale = opt.opt_val;
2330 break;
2331 case TCPOPT_SACK_PERM:
2332 if (opt.opt_val != 0)
2333 return -EINVAL;
2334
2335 tp->rx_opt.sack_ok |= TCP_SACK_SEEN;
2336 if (sysctl_tcp_fack)
2337 tcp_enable_fack(tp);
2338 break;
2339 case TCPOPT_TIMESTAMP:
2340 if (opt.opt_val != 0)
2341 return -EINVAL;
2342
2343 tp->rx_opt.tstamp_ok = 1;
2344 break;
2345 }
2346 }
2347
2348 return 0;
2349 }
2350
2351 /*
2352 * Socket option code for TCP.
2353 */
2354 static int do_tcp_setsockopt(struct sock *sk, int level,
2355 int optname, char __user *optval, unsigned int optlen)
2356 {
2357 struct tcp_sock *tp = tcp_sk(sk);
2358 struct inet_connection_sock *icsk = inet_csk(sk);
2359 int val;
2360 int err = 0;
2361
2362 /* These are data/string values, all the others are ints */
2363 switch (optname) {
2364 case TCP_CONGESTION: {
2365 char name[TCP_CA_NAME_MAX];
2366
2367 if (optlen < 1)
2368 return -EINVAL;
2369
2370 val = strncpy_from_user(name, optval,
2371 min_t(long, TCP_CA_NAME_MAX-1, optlen));
2372 if (val < 0)
2373 return -EFAULT;
2374 name[val] = 0;
2375
2376 lock_sock(sk);
2377 err = tcp_set_congestion_control(sk, name);
2378 release_sock(sk);
2379 return err;
2380 }
2381 case TCP_COOKIE_TRANSACTIONS: {
2382 struct tcp_cookie_transactions ctd;
2383 struct tcp_cookie_values *cvp = NULL;
2384
2385 if (sizeof(ctd) > optlen)
2386 return -EINVAL;
2387 if (copy_from_user(&ctd, optval, sizeof(ctd)))
2388 return -EFAULT;
2389
2390 if (ctd.tcpct_used > sizeof(ctd.tcpct_value) ||
2391 ctd.tcpct_s_data_desired > TCP_MSS_DESIRED)
2392 return -EINVAL;
2393
2394 if (ctd.tcpct_cookie_desired == 0) {
2395 /* default to global value */
2396 } else if ((0x1 & ctd.tcpct_cookie_desired) ||
2397 ctd.tcpct_cookie_desired > TCP_COOKIE_MAX ||
2398 ctd.tcpct_cookie_desired < TCP_COOKIE_MIN) {
2399 return -EINVAL;
2400 }
2401
2402 if (TCP_COOKIE_OUT_NEVER & ctd.tcpct_flags) {
2403 /* Supercedes all other values */
2404 lock_sock(sk);
2405 if (tp->cookie_values != NULL) {
2406 kref_put(&tp->cookie_values->kref,
2407 tcp_cookie_values_release);
2408 tp->cookie_values = NULL;
2409 }
2410 tp->rx_opt.cookie_in_always = 0; /* false */
2411 tp->rx_opt.cookie_out_never = 1; /* true */
2412 release_sock(sk);
2413 return err;
2414 }
2415
2416 /* Allocate ancillary memory before locking.
2417 */
2418 if (ctd.tcpct_used > 0 ||
2419 (tp->cookie_values == NULL &&
2420 (sysctl_tcp_cookie_size > 0 ||
2421 ctd.tcpct_cookie_desired > 0 ||
2422 ctd.tcpct_s_data_desired > 0))) {
2423 cvp = kzalloc(sizeof(*cvp) + ctd.tcpct_used,
2424 GFP_KERNEL);
2425 if (cvp == NULL)
2426 return -ENOMEM;
2427
2428 kref_init(&cvp->kref);
2429 }
2430 lock_sock(sk);
2431 tp->rx_opt.cookie_in_always =
2432 (TCP_COOKIE_IN_ALWAYS & ctd.tcpct_flags);
2433 tp->rx_opt.cookie_out_never = 0; /* false */
2434
2435 if (tp->cookie_values != NULL) {
2436 if (cvp != NULL) {
2437 /* Changed values are recorded by a changed
2438 * pointer, ensuring the cookie will differ,
2439 * without separately hashing each value later.
2440 */
2441 kref_put(&tp->cookie_values->kref,
2442 tcp_cookie_values_release);
2443 } else {
2444 cvp = tp->cookie_values;
2445 }
2446 }
2447
2448 if (cvp != NULL) {
2449 cvp->cookie_desired = ctd.tcpct_cookie_desired;
2450
2451 if (ctd.tcpct_used > 0) {
2452 memcpy(cvp->s_data_payload, ctd.tcpct_value,
2453 ctd.tcpct_used);
2454 cvp->s_data_desired = ctd.tcpct_used;
2455 cvp->s_data_constant = 1; /* true */
2456 } else {
2457 /* No constant payload data. */
2458 cvp->s_data_desired = ctd.tcpct_s_data_desired;
2459 cvp->s_data_constant = 0; /* false */
2460 }
2461
2462 tp->cookie_values = cvp;
2463 }
2464 release_sock(sk);
2465 return err;
2466 }
2467 default:
2468 /* fallthru */
2469 break;
2470 }
2471
2472 if (optlen < sizeof(int))
2473 return -EINVAL;
2474
2475 if (get_user(val, (int __user *)optval))
2476 return -EFAULT;
2477
2478 lock_sock(sk);
2479
2480 switch (optname) {
2481 case TCP_MAXSEG:
2482 /* Values greater than interface MTU won't take effect. However
2483 * at the point when this call is done we typically don't yet
2484 * know which interface is going to be used */
2485 if (val < TCP_MIN_MSS || val > MAX_TCP_WINDOW) {
2486 err = -EINVAL;
2487 break;
2488 }
2489 tp->rx_opt.user_mss = val;
2490 break;
2491
2492 case TCP_NODELAY:
2493 if (val) {
2494 /* TCP_NODELAY is weaker than TCP_CORK, so that
2495 * this option on corked socket is remembered, but
2496 * it is not activated until cork is cleared.
2497 *
2498 * However, when TCP_NODELAY is set we make
2499 * an explicit push, which overrides even TCP_CORK
2500 * for currently queued segments.
2501 */
2502 tp->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH;
2503 tcp_push_pending_frames(sk);
2504 } else {
2505 tp->nonagle &= ~TCP_NAGLE_OFF;
2506 }
2507 break;
2508
2509 case TCP_THIN_LINEAR_TIMEOUTS:
2510 if (val < 0 || val > 1)
2511 err = -EINVAL;
2512 else
2513 tp->thin_lto = val;
2514 break;
2515
2516 case TCP_THIN_DUPACK:
2517 if (val < 0 || val > 1)
2518 err = -EINVAL;
2519 else
2520 tp->thin_dupack = val;
2521 if (tp->thin_dupack)
2522 tcp_disable_early_retrans(tp);
2523 break;
2524
2525 case TCP_REPAIR:
2526 if (!tcp_can_repair_sock(sk))
2527 err = -EPERM;
2528 else if (val == 1) {
2529 tp->repair = 1;
2530 sk->sk_reuse = SK_FORCE_REUSE;
2531 tp->repair_queue = TCP_NO_QUEUE;
2532 } else if (val == 0) {
2533 tp->repair = 0;
2534 sk->sk_reuse = SK_NO_REUSE;
2535 tcp_send_window_probe(sk);
2536 } else
2537 err = -EINVAL;
2538
2539 break;
2540
2541 case TCP_REPAIR_QUEUE:
2542 if (!tp->repair)
2543 err = -EPERM;
2544 else if (val < TCP_QUEUES_NR)
2545 tp->repair_queue = val;
2546 else
2547 err = -EINVAL;
2548 break;
2549
2550 case TCP_QUEUE_SEQ:
2551 if (sk->sk_state != TCP_CLOSE)
2552 err = -EPERM;
2553 else if (tp->repair_queue == TCP_SEND_QUEUE)
2554 tp->write_seq = val;
2555 else if (tp->repair_queue == TCP_RECV_QUEUE)
2556 tp->rcv_nxt = val;
2557 else
2558 err = -EINVAL;
2559 break;
2560
2561 case TCP_REPAIR_OPTIONS:
2562 if (!tp->repair)
2563 err = -EINVAL;
2564 else if (sk->sk_state == TCP_ESTABLISHED)
2565 err = tcp_repair_options_est(tp,
2566 (struct tcp_repair_opt __user *)optval,
2567 optlen);
2568 else
2569 err = -EPERM;
2570 break;
2571
2572 case TCP_CORK:
2573 /* When set indicates to always queue non-full frames.
2574 * Later the user clears this option and we transmit
2575 * any pending partial frames in the queue. This is
2576 * meant to be used alongside sendfile() to get properly
2577 * filled frames when the user (for example) must write
2578 * out headers with a write() call first and then use
2579 * sendfile to send out the data parts.
2580 *
2581 * TCP_CORK can be set together with TCP_NODELAY and it is
2582 * stronger than TCP_NODELAY.
2583 */
2584 if (val) {
2585 tp->nonagle |= TCP_NAGLE_CORK;
2586 } else {
2587 tp->nonagle &= ~TCP_NAGLE_CORK;
2588 if (tp->nonagle&TCP_NAGLE_OFF)
2589 tp->nonagle |= TCP_NAGLE_PUSH;
2590 tcp_push_pending_frames(sk);
2591 }
2592 break;
2593
2594 case TCP_KEEPIDLE:
2595 if (val < 1 || val > MAX_TCP_KEEPIDLE)
2596 err = -EINVAL;
2597 else {
2598 tp->keepalive_time = val * HZ;
2599 if (sock_flag(sk, SOCK_KEEPOPEN) &&
2600 !((1 << sk->sk_state) &
2601 (TCPF_CLOSE | TCPF_LISTEN))) {
2602 u32 elapsed = keepalive_time_elapsed(tp);
2603 if (tp->keepalive_time > elapsed)
2604 elapsed = tp->keepalive_time - elapsed;
2605 else
2606 elapsed = 0;
2607 inet_csk_reset_keepalive_timer(sk, elapsed);
2608 }
2609 }
2610 break;
2611 case TCP_KEEPINTVL:
2612 if (val < 1 || val > MAX_TCP_KEEPINTVL)
2613 err = -EINVAL;
2614 else
2615 tp->keepalive_intvl = val * HZ;
2616 break;
2617 case TCP_KEEPCNT:
2618 if (val < 1 || val > MAX_TCP_KEEPCNT)
2619 err = -EINVAL;
2620 else
2621 tp->keepalive_probes = val;
2622 break;
2623 case TCP_SYNCNT:
2624 if (val < 1 || val > MAX_TCP_SYNCNT)
2625 err = -EINVAL;
2626 else
2627 icsk->icsk_syn_retries = val;
2628 break;
2629
2630 case TCP_LINGER2:
2631 if (val < 0)
2632 tp->linger2 = -1;
2633 else if (val > sysctl_tcp_fin_timeout / HZ)
2634 tp->linger2 = 0;
2635 else
2636 tp->linger2 = val * HZ;
2637 break;
2638
2639 case TCP_DEFER_ACCEPT:
2640 /* Translate value in seconds to number of retransmits */
2641 icsk->icsk_accept_queue.rskq_defer_accept =
2642 secs_to_retrans(val, TCP_TIMEOUT_INIT / HZ,
2643 TCP_RTO_MAX / HZ);
2644 break;
2645
2646 case TCP_WINDOW_CLAMP:
2647 if (!val) {
2648 if (sk->sk_state != TCP_CLOSE) {
2649 err = -EINVAL;
2650 break;
2651 }
2652 tp->window_clamp = 0;
2653 } else
2654 tp->window_clamp = val < SOCK_MIN_RCVBUF / 2 ?
2655 SOCK_MIN_RCVBUF / 2 : val;
2656 break;
2657
2658 case TCP_QUICKACK:
2659 if (!val) {
2660 icsk->icsk_ack.pingpong = 1;
2661 } else {
2662 icsk->icsk_ack.pingpong = 0;
2663 if ((1 << sk->sk_state) &
2664 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) &&
2665 inet_csk_ack_scheduled(sk)) {
2666 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
2667 tcp_cleanup_rbuf(sk, 1);
2668 if (!(val & 1))
2669 icsk->icsk_ack.pingpong = 1;
2670 }
2671 }
2672 break;
2673
2674 #ifdef CONFIG_TCP_MD5SIG
2675 case TCP_MD5SIG:
2676 /* Read the IP->Key mappings from userspace */
2677 err = tp->af_specific->md5_parse(sk, optval, optlen);
2678 break;
2679 #endif
2680 case TCP_USER_TIMEOUT:
2681 /* Cap the max timeout in ms TCP will retry/retrans
2682 * before giving up and aborting (ETIMEDOUT) a connection.
2683 */
2684 if (val < 0)
2685 err = -EINVAL;
2686 else
2687 icsk->icsk_user_timeout = msecs_to_jiffies(val);
2688 break;
2689 default:
2690 err = -ENOPROTOOPT;
2691 break;
2692 }
2693
2694 release_sock(sk);
2695 return err;
2696 }
2697
2698 int tcp_setsockopt(struct sock *sk, int level, int optname, char __user *optval,
2699 unsigned int optlen)
2700 {
2701 const struct inet_connection_sock *icsk = inet_csk(sk);
2702
2703 if (level != SOL_TCP)
2704 return icsk->icsk_af_ops->setsockopt(sk, level, optname,
2705 optval, optlen);
2706 return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2707 }
2708 EXPORT_SYMBOL(tcp_setsockopt);
2709
2710 #ifdef CONFIG_COMPAT
2711 int compat_tcp_setsockopt(struct sock *sk, int level, int optname,
2712 char __user *optval, unsigned int optlen)
2713 {
2714 if (level != SOL_TCP)
2715 return inet_csk_compat_setsockopt(sk, level, optname,
2716 optval, optlen);
2717 return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2718 }
2719 EXPORT_SYMBOL(compat_tcp_setsockopt);
2720 #endif
2721
2722 /* Return information about state of tcp endpoint in API format. */
2723 void tcp_get_info(const struct sock *sk, struct tcp_info *info)
2724 {
2725 const struct tcp_sock *tp = tcp_sk(sk);
2726 const struct inet_connection_sock *icsk = inet_csk(sk);
2727 u32 now = tcp_time_stamp;
2728
2729 memset(info, 0, sizeof(*info));
2730
2731 info->tcpi_state = sk->sk_state;
2732 info->tcpi_ca_state = icsk->icsk_ca_state;
2733 info->tcpi_retransmits = icsk->icsk_retransmits;
2734 info->tcpi_probes = icsk->icsk_probes_out;
2735 info->tcpi_backoff = icsk->icsk_backoff;
2736
2737 if (tp->rx_opt.tstamp_ok)
2738 info->tcpi_options |= TCPI_OPT_TIMESTAMPS;
2739 if (tcp_is_sack(tp))
2740 info->tcpi_options |= TCPI_OPT_SACK;
2741 if (tp->rx_opt.wscale_ok) {
2742 info->tcpi_options |= TCPI_OPT_WSCALE;
2743 info->tcpi_snd_wscale = tp->rx_opt.snd_wscale;
2744 info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale;
2745 }
2746
2747 if (tp->ecn_flags & TCP_ECN_OK)
2748 info->tcpi_options |= TCPI_OPT_ECN;
2749 if (tp->ecn_flags & TCP_ECN_SEEN)
2750 info->tcpi_options |= TCPI_OPT_ECN_SEEN;
2751
2752 info->tcpi_rto = jiffies_to_usecs(icsk->icsk_rto);
2753 info->tcpi_ato = jiffies_to_usecs(icsk->icsk_ack.ato);
2754 info->tcpi_snd_mss = tp->mss_cache;
2755 info->tcpi_rcv_mss = icsk->icsk_ack.rcv_mss;
2756
2757 if (sk->sk_state == TCP_LISTEN) {
2758 info->tcpi_unacked = sk->sk_ack_backlog;
2759 info->tcpi_sacked = sk->sk_max_ack_backlog;
2760 } else {
2761 info->tcpi_unacked = tp->packets_out;
2762 info->tcpi_sacked = tp->sacked_out;
2763 }
2764 info->tcpi_lost = tp->lost_out;
2765 info->tcpi_retrans = tp->retrans_out;
2766 info->tcpi_fackets = tp->fackets_out;
2767
2768 info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime);
2769 info->tcpi_last_data_recv = jiffies_to_msecs(now - icsk->icsk_ack.lrcvtime);
2770 info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp);
2771
2772 info->tcpi_pmtu = icsk->icsk_pmtu_cookie;
2773 info->tcpi_rcv_ssthresh = tp->rcv_ssthresh;
2774 info->tcpi_rtt = jiffies_to_usecs(tp->srtt)>>3;
2775 info->tcpi_rttvar = jiffies_to_usecs(tp->mdev)>>2;
2776 info->tcpi_snd_ssthresh = tp->snd_ssthresh;
2777 info->tcpi_snd_cwnd = tp->snd_cwnd;
2778 info->tcpi_advmss = tp->advmss;
2779 info->tcpi_reordering = tp->reordering;
2780
2781 info->tcpi_rcv_rtt = jiffies_to_usecs(tp->rcv_rtt_est.rtt)>>3;
2782 info->tcpi_rcv_space = tp->rcvq_space.space;
2783
2784 info->tcpi_total_retrans = tp->total_retrans;
2785 }
2786 EXPORT_SYMBOL_GPL(tcp_get_info);
2787
2788 static int do_tcp_getsockopt(struct sock *sk, int level,
2789 int optname, char __user *optval, int __user *optlen)
2790 {
2791 struct inet_connection_sock *icsk = inet_csk(sk);
2792 struct tcp_sock *tp = tcp_sk(sk);
2793 int val, len;
2794
2795 if (get_user(len, optlen))
2796 return -EFAULT;
2797
2798 len = min_t(unsigned int, len, sizeof(int));
2799
2800 if (len < 0)
2801 return -EINVAL;
2802
2803 switch (optname) {
2804 case TCP_MAXSEG:
2805 val = tp->mss_cache;
2806 if (!val && ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
2807 val = tp->rx_opt.user_mss;
2808 if (tp->repair)
2809 val = tp->rx_opt.mss_clamp;
2810 break;
2811 case TCP_NODELAY:
2812 val = !!(tp->nonagle&TCP_NAGLE_OFF);
2813 break;
2814 case TCP_CORK:
2815 val = !!(tp->nonagle&TCP_NAGLE_CORK);
2816 break;
2817 case TCP_KEEPIDLE:
2818 val = keepalive_time_when(tp) / HZ;
2819 break;
2820 case TCP_KEEPINTVL:
2821 val = keepalive_intvl_when(tp) / HZ;
2822 break;
2823 case TCP_KEEPCNT:
2824 val = keepalive_probes(tp);
2825 break;
2826 case TCP_SYNCNT:
2827 val = icsk->icsk_syn_retries ? : sysctl_tcp_syn_retries;
2828 break;
2829 case TCP_LINGER2:
2830 val = tp->linger2;
2831 if (val >= 0)
2832 val = (val ? : sysctl_tcp_fin_timeout) / HZ;
2833 break;
2834 case TCP_DEFER_ACCEPT:
2835 val = retrans_to_secs(icsk->icsk_accept_queue.rskq_defer_accept,
2836 TCP_TIMEOUT_INIT / HZ, TCP_RTO_MAX / HZ);
2837 break;
2838 case TCP_WINDOW_CLAMP:
2839 val = tp->window_clamp;
2840 break;
2841 case TCP_INFO: {
2842 struct tcp_info info;
2843
2844 if (get_user(len, optlen))
2845 return -EFAULT;
2846
2847 tcp_get_info(sk, &info);
2848
2849 len = min_t(unsigned int, len, sizeof(info));
2850 if (put_user(len, optlen))
2851 return -EFAULT;
2852 if (copy_to_user(optval, &info, len))
2853 return -EFAULT;
2854 return 0;
2855 }
2856 case TCP_QUICKACK:
2857 val = !icsk->icsk_ack.pingpong;
2858 break;
2859
2860 case TCP_CONGESTION:
2861 if (get_user(len, optlen))
2862 return -EFAULT;
2863 len = min_t(unsigned int, len, TCP_CA_NAME_MAX);
2864 if (put_user(len, optlen))
2865 return -EFAULT;
2866 if (copy_to_user(optval, icsk->icsk_ca_ops->name, len))
2867 return -EFAULT;
2868 return 0;
2869
2870 case TCP_COOKIE_TRANSACTIONS: {
2871 struct tcp_cookie_transactions ctd;
2872 struct tcp_cookie_values *cvp = tp->cookie_values;
2873
2874 if (get_user(len, optlen))
2875 return -EFAULT;
2876 if (len < sizeof(ctd))
2877 return -EINVAL;
2878
2879 memset(&ctd, 0, sizeof(ctd));
2880 ctd.tcpct_flags = (tp->rx_opt.cookie_in_always ?
2881 TCP_COOKIE_IN_ALWAYS : 0)
2882 | (tp->rx_opt.cookie_out_never ?
2883 TCP_COOKIE_OUT_NEVER : 0);
2884
2885 if (cvp != NULL) {
2886 ctd.tcpct_flags |= (cvp->s_data_in ?
2887 TCP_S_DATA_IN : 0)
2888 | (cvp->s_data_out ?
2889 TCP_S_DATA_OUT : 0);
2890
2891 ctd.tcpct_cookie_desired = cvp->cookie_desired;
2892 ctd.tcpct_s_data_desired = cvp->s_data_desired;
2893
2894 memcpy(&ctd.tcpct_value[0], &cvp->cookie_pair[0],
2895 cvp->cookie_pair_size);
2896 ctd.tcpct_used = cvp->cookie_pair_size;
2897 }
2898
2899 if (put_user(sizeof(ctd), optlen))
2900 return -EFAULT;
2901 if (copy_to_user(optval, &ctd, sizeof(ctd)))
2902 return -EFAULT;
2903 return 0;
2904 }
2905 case TCP_THIN_LINEAR_TIMEOUTS:
2906 val = tp->thin_lto;
2907 break;
2908 case TCP_THIN_DUPACK:
2909 val = tp->thin_dupack;
2910 break;
2911
2912 case TCP_REPAIR:
2913 val = tp->repair;
2914 break;
2915
2916 case TCP_REPAIR_QUEUE:
2917 if (tp->repair)
2918 val = tp->repair_queue;
2919 else
2920 return -EINVAL;
2921 break;
2922
2923 case TCP_QUEUE_SEQ:
2924 if (tp->repair_queue == TCP_SEND_QUEUE)
2925 val = tp->write_seq;
2926 else if (tp->repair_queue == TCP_RECV_QUEUE)
2927 val = tp->rcv_nxt;
2928 else
2929 return -EINVAL;
2930 break;
2931
2932 case TCP_USER_TIMEOUT:
2933 val = jiffies_to_msecs(icsk->icsk_user_timeout);
2934 break;
2935 default:
2936 return -ENOPROTOOPT;
2937 }
2938
2939 if (put_user(len, optlen))
2940 return -EFAULT;
2941 if (copy_to_user(optval, &val, len))
2942 return -EFAULT;
2943 return 0;
2944 }
2945
2946 int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval,
2947 int __user *optlen)
2948 {
2949 struct inet_connection_sock *icsk = inet_csk(sk);
2950
2951 if (level != SOL_TCP)
2952 return icsk->icsk_af_ops->getsockopt(sk, level, optname,
2953 optval, optlen);
2954 return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2955 }
2956 EXPORT_SYMBOL(tcp_getsockopt);
2957
2958 #ifdef CONFIG_COMPAT
2959 int compat_tcp_getsockopt(struct sock *sk, int level, int optname,
2960 char __user *optval, int __user *optlen)
2961 {
2962 if (level != SOL_TCP)
2963 return inet_csk_compat_getsockopt(sk, level, optname,
2964 optval, optlen);
2965 return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2966 }
2967 EXPORT_SYMBOL(compat_tcp_getsockopt);
2968 #endif
2969
2970 struct sk_buff *tcp_tso_segment(struct sk_buff *skb,
2971 netdev_features_t features)
2972 {
2973 struct sk_buff *segs = ERR_PTR(-EINVAL);
2974 struct tcphdr *th;
2975 unsigned int thlen;
2976 unsigned int seq;
2977 __be32 delta;
2978 unsigned int oldlen;
2979 unsigned int mss;
2980
2981 if (!pskb_may_pull(skb, sizeof(*th)))
2982 goto out;
2983
2984 th = tcp_hdr(skb);
2985 thlen = th->doff * 4;
2986 if (thlen < sizeof(*th))
2987 goto out;
2988
2989 if (!pskb_may_pull(skb, thlen))
2990 goto out;
2991
2992 oldlen = (u16)~skb->len;
2993 __skb_pull(skb, thlen);
2994
2995 mss = skb_shinfo(skb)->gso_size;
2996 if (unlikely(skb->len <= mss))
2997 goto out;
2998
2999 if (skb_gso_ok(skb, features | NETIF_F_GSO_ROBUST)) {
3000 /* Packet is from an untrusted source, reset gso_segs. */
3001 int type = skb_shinfo(skb)->gso_type;
3002
3003 if (unlikely(type &
3004 ~(SKB_GSO_TCPV4 |
3005 SKB_GSO_DODGY |
3006 SKB_GSO_TCP_ECN |
3007 SKB_GSO_TCPV6 |
3008 0) ||
3009 !(type & (SKB_GSO_TCPV4 | SKB_GSO_TCPV6))))
3010 goto out;
3011
3012 skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss);
3013
3014 segs = NULL;
3015 goto out;
3016 }
3017
3018 segs = skb_segment(skb, features);
3019 if (IS_ERR(segs))
3020 goto out;
3021
3022 delta = htonl(oldlen + (thlen + mss));
3023
3024 skb = segs;
3025 th = tcp_hdr(skb);
3026 seq = ntohl(th->seq);
3027
3028 do {
3029 th->fin = th->psh = 0;
3030
3031 th->check = ~csum_fold((__force __wsum)((__force u32)th->check +
3032 (__force u32)delta));
3033 if (skb->ip_summed != CHECKSUM_PARTIAL)
3034 th->check =
3035 csum_fold(csum_partial(skb_transport_header(skb),
3036 thlen, skb->csum));
3037
3038 seq += mss;
3039 skb = skb->next;
3040 th = tcp_hdr(skb);
3041
3042 th->seq = htonl(seq);
3043 th->cwr = 0;
3044 } while (skb->next);
3045
3046 delta = htonl(oldlen + (skb->tail - skb->transport_header) +
3047 skb->data_len);
3048 th->check = ~csum_fold((__force __wsum)((__force u32)th->check +
3049 (__force u32)delta));
3050 if (skb->ip_summed != CHECKSUM_PARTIAL)
3051 th->check = csum_fold(csum_partial(skb_transport_header(skb),
3052 thlen, skb->csum));
3053
3054 out:
3055 return segs;
3056 }
3057 EXPORT_SYMBOL(tcp_tso_segment);
3058
3059 struct sk_buff **tcp_gro_receive(struct sk_buff **head, struct sk_buff *skb)
3060 {
3061 struct sk_buff **pp = NULL;
3062 struct sk_buff *p;
3063 struct tcphdr *th;
3064 struct tcphdr *th2;
3065 unsigned int len;
3066 unsigned int thlen;
3067 __be32 flags;
3068 unsigned int mss = 1;
3069 unsigned int hlen;
3070 unsigned int off;
3071 int flush = 1;
3072 int i;
3073
3074 off = skb_gro_offset(skb);
3075 hlen = off + sizeof(*th);
3076 th = skb_gro_header_fast(skb, off);
3077 if (skb_gro_header_hard(skb, hlen)) {
3078 th = skb_gro_header_slow(skb, hlen, off);
3079 if (unlikely(!th))
3080 goto out;
3081 }
3082
3083 thlen = th->doff * 4;
3084 if (thlen < sizeof(*th))
3085 goto out;
3086
3087 hlen = off + thlen;
3088 if (skb_gro_header_hard(skb, hlen)) {
3089 th = skb_gro_header_slow(skb, hlen, off);
3090 if (unlikely(!th))
3091 goto out;
3092 }
3093
3094 skb_gro_pull(skb, thlen);
3095
3096 len = skb_gro_len(skb);
3097 flags = tcp_flag_word(th);
3098
3099 for (; (p = *head); head = &p->next) {
3100 if (!NAPI_GRO_CB(p)->same_flow)
3101 continue;
3102
3103 th2 = tcp_hdr(p);
3104
3105 if (*(u32 *)&th->source ^ *(u32 *)&th2->source) {
3106 NAPI_GRO_CB(p)->same_flow = 0;
3107 continue;
3108 }
3109
3110 goto found;
3111 }
3112
3113 goto out_check_final;
3114
3115 found:
3116 flush = NAPI_GRO_CB(p)->flush;
3117 flush |= (__force int)(flags & TCP_FLAG_CWR);
3118 flush |= (__force int)((flags ^ tcp_flag_word(th2)) &
3119 ~(TCP_FLAG_CWR | TCP_FLAG_FIN | TCP_FLAG_PSH));
3120 flush |= (__force int)(th->ack_seq ^ th2->ack_seq);
3121 for (i = sizeof(*th); i < thlen; i += 4)
3122 flush |= *(u32 *)((u8 *)th + i) ^
3123 *(u32 *)((u8 *)th2 + i);
3124
3125 mss = skb_shinfo(p)->gso_size;
3126
3127 flush |= (len - 1) >= mss;
3128 flush |= (ntohl(th2->seq) + skb_gro_len(p)) ^ ntohl(th->seq);
3129
3130 if (flush || skb_gro_receive(head, skb)) {
3131 mss = 1;
3132 goto out_check_final;
3133 }
3134
3135 p = *head;
3136 th2 = tcp_hdr(p);
3137 tcp_flag_word(th2) |= flags & (TCP_FLAG_FIN | TCP_FLAG_PSH);
3138
3139 out_check_final:
3140 flush = len < mss;
3141 flush |= (__force int)(flags & (TCP_FLAG_URG | TCP_FLAG_PSH |
3142 TCP_FLAG_RST | TCP_FLAG_SYN |
3143 TCP_FLAG_FIN));
3144
3145 if (p && (!NAPI_GRO_CB(skb)->same_flow || flush))
3146 pp = head;
3147
3148 out:
3149 NAPI_GRO_CB(skb)->flush |= flush;
3150
3151 return pp;
3152 }
3153 EXPORT_SYMBOL(tcp_gro_receive);
3154
3155 int tcp_gro_complete(struct sk_buff *skb)
3156 {
3157 struct tcphdr *th = tcp_hdr(skb);
3158
3159 skb->csum_start = skb_transport_header(skb) - skb->head;
3160 skb->csum_offset = offsetof(struct tcphdr, check);
3161 skb->ip_summed = CHECKSUM_PARTIAL;
3162
3163 skb_shinfo(skb)->gso_segs = NAPI_GRO_CB(skb)->count;
3164
3165 if (th->cwr)
3166 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
3167
3168 return 0;
3169 }
3170 EXPORT_SYMBOL(tcp_gro_complete);
3171
3172 #ifdef CONFIG_TCP_MD5SIG
3173 static unsigned long tcp_md5sig_users;
3174 static struct tcp_md5sig_pool __percpu *tcp_md5sig_pool;
3175 static DEFINE_SPINLOCK(tcp_md5sig_pool_lock);
3176
3177 static void __tcp_free_md5sig_pool(struct tcp_md5sig_pool __percpu *pool)
3178 {
3179 int cpu;
3180
3181 for_each_possible_cpu(cpu) {
3182 struct tcp_md5sig_pool *p = per_cpu_ptr(pool, cpu);
3183
3184 if (p->md5_desc.tfm)
3185 crypto_free_hash(p->md5_desc.tfm);
3186 }
3187 free_percpu(pool);
3188 }
3189
3190 void tcp_free_md5sig_pool(void)
3191 {
3192 struct tcp_md5sig_pool __percpu *pool = NULL;
3193
3194 spin_lock_bh(&tcp_md5sig_pool_lock);
3195 if (--tcp_md5sig_users == 0) {
3196 pool = tcp_md5sig_pool;
3197 tcp_md5sig_pool = NULL;
3198 }
3199 spin_unlock_bh(&tcp_md5sig_pool_lock);
3200 if (pool)
3201 __tcp_free_md5sig_pool(pool);
3202 }
3203 EXPORT_SYMBOL(tcp_free_md5sig_pool);
3204
3205 static struct tcp_md5sig_pool __percpu *
3206 __tcp_alloc_md5sig_pool(struct sock *sk)
3207 {
3208 int cpu;
3209 struct tcp_md5sig_pool __percpu *pool;
3210
3211 pool = alloc_percpu(struct tcp_md5sig_pool);
3212 if (!pool)
3213 return NULL;
3214
3215 for_each_possible_cpu(cpu) {
3216 struct crypto_hash *hash;
3217
3218 hash = crypto_alloc_hash("md5", 0, CRYPTO_ALG_ASYNC);
3219 if (!hash || IS_ERR(hash))
3220 goto out_free;
3221
3222 per_cpu_ptr(pool, cpu)->md5_desc.tfm = hash;
3223 }
3224 return pool;
3225 out_free:
3226 __tcp_free_md5sig_pool(pool);
3227 return NULL;
3228 }
3229
3230 struct tcp_md5sig_pool __percpu *tcp_alloc_md5sig_pool(struct sock *sk)
3231 {
3232 struct tcp_md5sig_pool __percpu *pool;
3233 bool alloc = false;
3234
3235 retry:
3236 spin_lock_bh(&tcp_md5sig_pool_lock);
3237 pool = tcp_md5sig_pool;
3238 if (tcp_md5sig_users++ == 0) {
3239 alloc = true;
3240 spin_unlock_bh(&tcp_md5sig_pool_lock);
3241 } else if (!pool) {
3242 tcp_md5sig_users--;
3243 spin_unlock_bh(&tcp_md5sig_pool_lock);
3244 cpu_relax();
3245 goto retry;
3246 } else
3247 spin_unlock_bh(&tcp_md5sig_pool_lock);
3248
3249 if (alloc) {
3250 /* we cannot hold spinlock here because this may sleep. */
3251 struct tcp_md5sig_pool __percpu *p;
3252
3253 p = __tcp_alloc_md5sig_pool(sk);
3254 spin_lock_bh(&tcp_md5sig_pool_lock);
3255 if (!p) {
3256 tcp_md5sig_users--;
3257 spin_unlock_bh(&tcp_md5sig_pool_lock);
3258 return NULL;
3259 }
3260 pool = tcp_md5sig_pool;
3261 if (pool) {
3262 /* oops, it has already been assigned. */
3263 spin_unlock_bh(&tcp_md5sig_pool_lock);
3264 __tcp_free_md5sig_pool(p);
3265 } else {
3266 tcp_md5sig_pool = pool = p;
3267 spin_unlock_bh(&tcp_md5sig_pool_lock);
3268 }
3269 }
3270 return pool;
3271 }
3272 EXPORT_SYMBOL(tcp_alloc_md5sig_pool);
3273
3274
3275 /**
3276 * tcp_get_md5sig_pool - get md5sig_pool for this user
3277 *
3278 * We use percpu structure, so if we succeed, we exit with preemption
3279 * and BH disabled, to make sure another thread or softirq handling
3280 * wont try to get same context.
3281 */
3282 struct tcp_md5sig_pool *tcp_get_md5sig_pool(void)
3283 {
3284 struct tcp_md5sig_pool __percpu *p;
3285
3286 local_bh_disable();
3287
3288 spin_lock(&tcp_md5sig_pool_lock);
3289 p = tcp_md5sig_pool;
3290 if (p)
3291 tcp_md5sig_users++;
3292 spin_unlock(&tcp_md5sig_pool_lock);
3293
3294 if (p)
3295 return this_cpu_ptr(p);
3296
3297 local_bh_enable();
3298 return NULL;
3299 }
3300 EXPORT_SYMBOL(tcp_get_md5sig_pool);
3301
3302 void tcp_put_md5sig_pool(void)
3303 {
3304 local_bh_enable();
3305 tcp_free_md5sig_pool();
3306 }
3307 EXPORT_SYMBOL(tcp_put_md5sig_pool);
3308
3309 int tcp_md5_hash_header(struct tcp_md5sig_pool *hp,
3310 const struct tcphdr *th)
3311 {
3312 struct scatterlist sg;
3313 struct tcphdr hdr;
3314 int err;
3315
3316 /* We are not allowed to change tcphdr, make a local copy */
3317 memcpy(&hdr, th, sizeof(hdr));
3318 hdr.check = 0;
3319
3320 /* options aren't included in the hash */
3321 sg_init_one(&sg, &hdr, sizeof(hdr));
3322 err = crypto_hash_update(&hp->md5_desc, &sg, sizeof(hdr));
3323 return err;
3324 }
3325 EXPORT_SYMBOL(tcp_md5_hash_header);
3326
3327 int tcp_md5_hash_skb_data(struct tcp_md5sig_pool *hp,
3328 const struct sk_buff *skb, unsigned int header_len)
3329 {
3330 struct scatterlist sg;
3331 const struct tcphdr *tp = tcp_hdr(skb);
3332 struct hash_desc *desc = &hp->md5_desc;
3333 unsigned int i;
3334 const unsigned int head_data_len = skb_headlen(skb) > header_len ?
3335 skb_headlen(skb) - header_len : 0;
3336 const struct skb_shared_info *shi = skb_shinfo(skb);
3337 struct sk_buff *frag_iter;
3338
3339 sg_init_table(&sg, 1);
3340
3341 sg_set_buf(&sg, ((u8 *) tp) + header_len, head_data_len);
3342 if (crypto_hash_update(desc, &sg, head_data_len))
3343 return 1;
3344
3345 for (i = 0; i < shi->nr_frags; ++i) {
3346 const struct skb_frag_struct *f = &shi->frags[i];
3347 struct page *page = skb_frag_page(f);
3348 sg_set_page(&sg, page, skb_frag_size(f), f->page_offset);
3349 if (crypto_hash_update(desc, &sg, skb_frag_size(f)))
3350 return 1;
3351 }
3352
3353 skb_walk_frags(skb, frag_iter)
3354 if (tcp_md5_hash_skb_data(hp, frag_iter, 0))
3355 return 1;
3356
3357 return 0;
3358 }
3359 EXPORT_SYMBOL(tcp_md5_hash_skb_data);
3360
3361 int tcp_md5_hash_key(struct tcp_md5sig_pool *hp, const struct tcp_md5sig_key *key)
3362 {
3363 struct scatterlist sg;
3364
3365 sg_init_one(&sg, key->key, key->keylen);
3366 return crypto_hash_update(&hp->md5_desc, &sg, key->keylen);
3367 }
3368 EXPORT_SYMBOL(tcp_md5_hash_key);
3369
3370 #endif
3371
3372 /* Each Responder maintains up to two secret values concurrently for
3373 * efficient secret rollover. Each secret value has 4 states:
3374 *
3375 * Generating. (tcp_secret_generating != tcp_secret_primary)
3376 * Generates new Responder-Cookies, but not yet used for primary
3377 * verification. This is a short-term state, typically lasting only
3378 * one round trip time (RTT).
3379 *
3380 * Primary. (tcp_secret_generating == tcp_secret_primary)
3381 * Used both for generation and primary verification.
3382 *
3383 * Retiring. (tcp_secret_retiring != tcp_secret_secondary)
3384 * Used for verification, until the first failure that can be
3385 * verified by the newer Generating secret. At that time, this
3386 * cookie's state is changed to Secondary, and the Generating
3387 * cookie's state is changed to Primary. This is a short-term state,
3388 * typically lasting only one round trip time (RTT).
3389 *
3390 * Secondary. (tcp_secret_retiring == tcp_secret_secondary)
3391 * Used for secondary verification, after primary verification
3392 * failures. This state lasts no more than twice the Maximum Segment
3393 * Lifetime (2MSL). Then, the secret is discarded.
3394 */
3395 struct tcp_cookie_secret {
3396 /* The secret is divided into two parts. The digest part is the
3397 * equivalent of previously hashing a secret and saving the state,
3398 * and serves as an initialization vector (IV). The message part
3399 * serves as the trailing secret.
3400 */
3401 u32 secrets[COOKIE_WORKSPACE_WORDS];
3402 unsigned long expires;
3403 };
3404
3405 #define TCP_SECRET_1MSL (HZ * TCP_PAWS_MSL)
3406 #define TCP_SECRET_2MSL (HZ * TCP_PAWS_MSL * 2)
3407 #define TCP_SECRET_LIFE (HZ * 600)
3408
3409 static struct tcp_cookie_secret tcp_secret_one;
3410 static struct tcp_cookie_secret tcp_secret_two;
3411
3412 /* Essentially a circular list, without dynamic allocation. */
3413 static struct tcp_cookie_secret *tcp_secret_generating;
3414 static struct tcp_cookie_secret *tcp_secret_primary;
3415 static struct tcp_cookie_secret *tcp_secret_retiring;
3416 static struct tcp_cookie_secret *tcp_secret_secondary;
3417
3418 static DEFINE_SPINLOCK(tcp_secret_locker);
3419
3420 /* Select a pseudo-random word in the cookie workspace.
3421 */
3422 static inline u32 tcp_cookie_work(const u32 *ws, const int n)
3423 {
3424 return ws[COOKIE_DIGEST_WORDS + ((COOKIE_MESSAGE_WORDS-1) & ws[n])];
3425 }
3426
3427 /* Fill bakery[COOKIE_WORKSPACE_WORDS] with generator, updating as needed.
3428 * Called in softirq context.
3429 * Returns: 0 for success.
3430 */
3431 int tcp_cookie_generator(u32 *bakery)
3432 {
3433 unsigned long jiffy = jiffies;
3434
3435 if (unlikely(time_after_eq(jiffy, tcp_secret_generating->expires))) {
3436 spin_lock_bh(&tcp_secret_locker);
3437 if (!time_after_eq(jiffy, tcp_secret_generating->expires)) {
3438 /* refreshed by another */
3439 memcpy(bakery,
3440 &tcp_secret_generating->secrets[0],
3441 COOKIE_WORKSPACE_WORDS);
3442 } else {
3443 /* still needs refreshing */
3444 get_random_bytes(bakery, COOKIE_WORKSPACE_WORDS);
3445
3446 /* The first time, paranoia assumes that the
3447 * randomization function isn't as strong. But,
3448 * this secret initialization is delayed until
3449 * the last possible moment (packet arrival).
3450 * Although that time is observable, it is
3451 * unpredictably variable. Mash in the most
3452 * volatile clock bits available, and expire the
3453 * secret extra quickly.
3454 */
3455 if (unlikely(tcp_secret_primary->expires ==
3456 tcp_secret_secondary->expires)) {
3457 struct timespec tv;
3458
3459 getnstimeofday(&tv);
3460 bakery[COOKIE_DIGEST_WORDS+0] ^=
3461 (u32)tv.tv_nsec;
3462
3463 tcp_secret_secondary->expires = jiffy
3464 + TCP_SECRET_1MSL
3465 + (0x0f & tcp_cookie_work(bakery, 0));
3466 } else {
3467 tcp_secret_secondary->expires = jiffy
3468 + TCP_SECRET_LIFE
3469 + (0xff & tcp_cookie_work(bakery, 1));
3470 tcp_secret_primary->expires = jiffy
3471 + TCP_SECRET_2MSL
3472 + (0x1f & tcp_cookie_work(bakery, 2));
3473 }
3474 memcpy(&tcp_secret_secondary->secrets[0],
3475 bakery, COOKIE_WORKSPACE_WORDS);
3476
3477 rcu_assign_pointer(tcp_secret_generating,
3478 tcp_secret_secondary);
3479 rcu_assign_pointer(tcp_secret_retiring,
3480 tcp_secret_primary);
3481 /*
3482 * Neither call_rcu() nor synchronize_rcu() needed.
3483 * Retiring data is not freed. It is replaced after
3484 * further (locked) pointer updates, and a quiet time
3485 * (minimum 1MSL, maximum LIFE - 2MSL).
3486 */
3487 }
3488 spin_unlock_bh(&tcp_secret_locker);
3489 } else {
3490 rcu_read_lock_bh();
3491 memcpy(bakery,
3492 &rcu_dereference(tcp_secret_generating)->secrets[0],
3493 COOKIE_WORKSPACE_WORDS);
3494 rcu_read_unlock_bh();
3495 }
3496 return 0;
3497 }
3498 EXPORT_SYMBOL(tcp_cookie_generator);
3499
3500 void tcp_done(struct sock *sk)
3501 {
3502 if (sk->sk_state == TCP_SYN_SENT || sk->sk_state == TCP_SYN_RECV)
3503 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_ATTEMPTFAILS);
3504
3505 tcp_set_state(sk, TCP_CLOSE);
3506 tcp_clear_xmit_timers(sk);
3507
3508 sk->sk_shutdown = SHUTDOWN_MASK;
3509
3510 if (!sock_flag(sk, SOCK_DEAD))
3511 sk->sk_state_change(sk);
3512 else
3513 inet_csk_destroy_sock(sk);
3514 }
3515 EXPORT_SYMBOL_GPL(tcp_done);
3516
3517 extern struct tcp_congestion_ops tcp_reno;
3518
3519 static __initdata unsigned long thash_entries;
3520 static int __init set_thash_entries(char *str)
3521 {
3522 ssize_t ret;
3523
3524 if (!str)
3525 return 0;
3526
3527 ret = kstrtoul(str, 0, &thash_entries);
3528 if (ret)
3529 return 0;
3530
3531 return 1;
3532 }
3533 __setup("thash_entries=", set_thash_entries);
3534
3535 void tcp_init_mem(struct net *net)
3536 {
3537 unsigned long limit = nr_free_buffer_pages() / 8;
3538 limit = max(limit, 128UL);
3539 net->ipv4.sysctl_tcp_mem[0] = limit / 4 * 3;
3540 net->ipv4.sysctl_tcp_mem[1] = limit;
3541 net->ipv4.sysctl_tcp_mem[2] = net->ipv4.sysctl_tcp_mem[0] * 2;
3542 }
3543
3544 void __init tcp_init(void)
3545 {
3546 struct sk_buff *skb = NULL;
3547 unsigned long limit;
3548 int max_rshare, max_wshare, cnt;
3549 unsigned int i;
3550 unsigned long jiffy = jiffies;
3551
3552 BUILD_BUG_ON(sizeof(struct tcp_skb_cb) > sizeof(skb->cb));
3553
3554 percpu_counter_init(&tcp_sockets_allocated, 0);
3555 percpu_counter_init(&tcp_orphan_count, 0);
3556 tcp_hashinfo.bind_bucket_cachep =
3557 kmem_cache_create("tcp_bind_bucket",
3558 sizeof(struct inet_bind_bucket), 0,
3559 SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL);
3560
3561 /* Size and allocate the main established and bind bucket
3562 * hash tables.
3563 *
3564 * The methodology is similar to that of the buffer cache.
3565 */
3566 tcp_hashinfo.ehash =
3567 alloc_large_system_hash("TCP established",
3568 sizeof(struct inet_ehash_bucket),
3569 thash_entries,
3570 (totalram_pages >= 128 * 1024) ?
3571 13 : 15,
3572 0,
3573 NULL,
3574 &tcp_hashinfo.ehash_mask,
3575 0,
3576 thash_entries ? 0 : 512 * 1024);
3577 for (i = 0; i <= tcp_hashinfo.ehash_mask; i++) {
3578 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].chain, i);
3579 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].twchain, i);
3580 }
3581 if (inet_ehash_locks_alloc(&tcp_hashinfo))
3582 panic("TCP: failed to alloc ehash_locks");
3583 tcp_hashinfo.bhash =
3584 alloc_large_system_hash("TCP bind",
3585 sizeof(struct inet_bind_hashbucket),
3586 tcp_hashinfo.ehash_mask + 1,
3587 (totalram_pages >= 128 * 1024) ?
3588 13 : 15,
3589 0,
3590 &tcp_hashinfo.bhash_size,
3591 NULL,
3592 0,
3593 64 * 1024);
3594 tcp_hashinfo.bhash_size = 1U << tcp_hashinfo.bhash_size;
3595 for (i = 0; i < tcp_hashinfo.bhash_size; i++) {
3596 spin_lock_init(&tcp_hashinfo.bhash[i].lock);
3597 INIT_HLIST_HEAD(&tcp_hashinfo.bhash[i].chain);
3598 }
3599
3600
3601 cnt = tcp_hashinfo.ehash_mask + 1;
3602
3603 tcp_death_row.sysctl_max_tw_buckets = cnt / 2;
3604 sysctl_tcp_max_orphans = cnt / 2;
3605 sysctl_max_syn_backlog = max(128, cnt / 256);
3606
3607 tcp_init_mem(&init_net);
3608 /* Set per-socket limits to no more than 1/128 the pressure threshold */
3609 limit = nr_free_buffer_pages() << (PAGE_SHIFT - 7);
3610 max_wshare = min(4UL*1024*1024, limit);
3611 max_rshare = min(6UL*1024*1024, limit);
3612
3613 sysctl_tcp_wmem[0] = SK_MEM_QUANTUM;
3614 sysctl_tcp_wmem[1] = 16*1024;
3615 sysctl_tcp_wmem[2] = max(64*1024, max_wshare);
3616
3617 sysctl_tcp_rmem[0] = SK_MEM_QUANTUM;
3618 sysctl_tcp_rmem[1] = 87380;
3619 sysctl_tcp_rmem[2] = max(87380, max_rshare);
3620
3621 pr_info("Hash tables configured (established %u bind %u)\n",
3622 tcp_hashinfo.ehash_mask + 1, tcp_hashinfo.bhash_size);
3623
3624 tcp_metrics_init();
3625
3626 tcp_register_congestion_control(&tcp_reno);
3627
3628 memset(&tcp_secret_one.secrets[0], 0, sizeof(tcp_secret_one.secrets));
3629 memset(&tcp_secret_two.secrets[0], 0, sizeof(tcp_secret_two.secrets));
3630 tcp_secret_one.expires = jiffy; /* past due */
3631 tcp_secret_two.expires = jiffy; /* past due */
3632 tcp_secret_generating = &tcp_secret_one;
3633 tcp_secret_primary = &tcp_secret_one;
3634 tcp_secret_retiring = &tcp_secret_two;
3635 tcp_secret_secondary = &tcp_secret_two;
3636 tcp_tasklet_init();
3637 }
Linux kernel - Deliverables
This page took 0.127805 seconds and 5 git commands to generate.