Merge tag 'fixes-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/arm...
[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/inet_diag.h>
256 #include <linux/init.h>
257 #include <linux/fs.h>
258 #include <linux/skbuff.h>
259 #include <linux/scatterlist.h>
260 #include <linux/splice.h>
261 #include <linux/net.h>
262 #include <linux/socket.h>
263 #include <linux/random.h>
264 #include <linux/bootmem.h>
265 #include <linux/highmem.h>
266 #include <linux/swap.h>
267 #include <linux/cache.h>
268 #include <linux/err.h>
269 #include <linux/crypto.h>
270 #include <linux/time.h>
271 #include <linux/slab.h>
272
273 #include <net/icmp.h>
274 #include <net/inet_common.h>
275 #include <net/tcp.h>
276 #include <net/xfrm.h>
277 #include <net/ip.h>
278 #include <net/sock.h>
279
280 #include <asm/uaccess.h>
281 #include <asm/ioctls.h>
282 #include <net/busy_poll.h>
283
284 int sysctl_tcp_fin_timeout __read_mostly = TCP_FIN_TIMEOUT;
285
286 int sysctl_tcp_min_tso_segs __read_mostly = 2;
287
288 int sysctl_tcp_autocorking __read_mostly = 1;
289
290 struct percpu_counter tcp_orphan_count;
291 EXPORT_SYMBOL_GPL(tcp_orphan_count);
292
293 long sysctl_tcp_mem[3] __read_mostly;
294 int sysctl_tcp_wmem[3] __read_mostly;
295 int sysctl_tcp_rmem[3] __read_mostly;
296
297 EXPORT_SYMBOL(sysctl_tcp_mem);
298 EXPORT_SYMBOL(sysctl_tcp_rmem);
299 EXPORT_SYMBOL(sysctl_tcp_wmem);
300
301 atomic_long_t tcp_memory_allocated; /* Current allocated memory. */
302 EXPORT_SYMBOL(tcp_memory_allocated);
303
304 /*
305 * Current number of TCP sockets.
306 */
307 struct percpu_counter tcp_sockets_allocated;
308 EXPORT_SYMBOL(tcp_sockets_allocated);
309
310 /*
311 * TCP splice context
312 */
313 struct tcp_splice_state {
314 struct pipe_inode_info *pipe;
315 size_t len;
316 unsigned int flags;
317 };
318
319 /*
320 * Pressure flag: try to collapse.
321 * Technical note: it is used by multiple contexts non atomically.
322 * All the __sk_mem_schedule() is of this nature: accounting
323 * is strict, actions are advisory and have some latency.
324 */
325 int tcp_memory_pressure __read_mostly;
326 EXPORT_SYMBOL(tcp_memory_pressure);
327
328 void tcp_enter_memory_pressure(struct sock *sk)
329 {
330 if (!tcp_memory_pressure) {
331 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMEMORYPRESSURES);
332 tcp_memory_pressure = 1;
333 }
334 }
335 EXPORT_SYMBOL(tcp_enter_memory_pressure);
336
337 /* Convert seconds to retransmits based on initial and max timeout */
338 static u8 secs_to_retrans(int seconds, int timeout, int rto_max)
339 {
340 u8 res = 0;
341
342 if (seconds > 0) {
343 int period = timeout;
344
345 res = 1;
346 while (seconds > period && res < 255) {
347 res++;
348 timeout <<= 1;
349 if (timeout > rto_max)
350 timeout = rto_max;
351 period += timeout;
352 }
353 }
354 return res;
355 }
356
357 /* Convert retransmits to seconds based on initial and max timeout */
358 static int retrans_to_secs(u8 retrans, int timeout, int rto_max)
359 {
360 int period = 0;
361
362 if (retrans > 0) {
363 period = timeout;
364 while (--retrans) {
365 timeout <<= 1;
366 if (timeout > rto_max)
367 timeout = rto_max;
368 period += timeout;
369 }
370 }
371 return period;
372 }
373
374 /* Address-family independent initialization for a tcp_sock.
375 *
376 * NOTE: A lot of things set to zero explicitly by call to
377 * sk_alloc() so need not be done here.
378 */
379 void tcp_init_sock(struct sock *sk)
380 {
381 struct inet_connection_sock *icsk = inet_csk(sk);
382 struct tcp_sock *tp = tcp_sk(sk);
383
384 __skb_queue_head_init(&tp->out_of_order_queue);
385 tcp_init_xmit_timers(sk);
386 tcp_prequeue_init(tp);
387 INIT_LIST_HEAD(&tp->tsq_node);
388
389 icsk->icsk_rto = TCP_TIMEOUT_INIT;
390 tp->mdev_us = jiffies_to_usecs(TCP_TIMEOUT_INIT);
391 tp->rtt_min[0].rtt = ~0U;
392
393 /* So many TCP implementations out there (incorrectly) count the
394 * initial SYN frame in their delayed-ACK and congestion control
395 * algorithms that we must have the following bandaid to talk
396 * efficiently to them. -DaveM
397 */
398 tp->snd_cwnd = TCP_INIT_CWND;
399
400 /* See draft-stevens-tcpca-spec-01 for discussion of the
401 * initialization of these values.
402 */
403 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
404 tp->snd_cwnd_clamp = ~0;
405 tp->mss_cache = TCP_MSS_DEFAULT;
406 u64_stats_init(&tp->syncp);
407
408 tp->reordering = sysctl_tcp_reordering;
409 tcp_enable_early_retrans(tp);
410 tcp_assign_congestion_control(sk);
411
412 tp->tsoffset = 0;
413
414 sk->sk_state = TCP_CLOSE;
415
416 sk->sk_write_space = sk_stream_write_space;
417 sock_set_flag(sk, SOCK_USE_WRITE_QUEUE);
418
419 icsk->icsk_sync_mss = tcp_sync_mss;
420
421 sk->sk_sndbuf = sysctl_tcp_wmem[1];
422 sk->sk_rcvbuf = sysctl_tcp_rmem[1];
423
424 local_bh_disable();
425 sock_update_memcg(sk);
426 sk_sockets_allocated_inc(sk);
427 local_bh_enable();
428 }
429 EXPORT_SYMBOL(tcp_init_sock);
430
431 static void tcp_tx_timestamp(struct sock *sk, struct sk_buff *skb)
432 {
433 if (sk->sk_tsflags) {
434 struct skb_shared_info *shinfo = skb_shinfo(skb);
435
436 sock_tx_timestamp(sk, &shinfo->tx_flags);
437 if (shinfo->tx_flags & SKBTX_ANY_TSTAMP)
438 shinfo->tskey = TCP_SKB_CB(skb)->seq + skb->len - 1;
439 }
440 }
441
442 /*
443 * Wait for a TCP event.
444 *
445 * Note that we don't need to lock the socket, as the upper poll layers
446 * take care of normal races (between the test and the event) and we don't
447 * go look at any of the socket buffers directly.
448 */
449 unsigned int tcp_poll(struct file *file, struct socket *sock, poll_table *wait)
450 {
451 unsigned int mask;
452 struct sock *sk = sock->sk;
453 const struct tcp_sock *tp = tcp_sk(sk);
454 int state;
455
456 sock_rps_record_flow(sk);
457
458 sock_poll_wait(file, sk_sleep(sk), wait);
459
460 state = sk_state_load(sk);
461 if (state == TCP_LISTEN)
462 return inet_csk_listen_poll(sk);
463
464 /* Socket is not locked. We are protected from async events
465 * by poll logic and correct handling of state changes
466 * made by other threads is impossible in any case.
467 */
468
469 mask = 0;
470
471 /*
472 * POLLHUP is certainly not done right. But poll() doesn't
473 * have a notion of HUP in just one direction, and for a
474 * socket the read side is more interesting.
475 *
476 * Some poll() documentation says that POLLHUP is incompatible
477 * with the POLLOUT/POLLWR flags, so somebody should check this
478 * all. But careful, it tends to be safer to return too many
479 * bits than too few, and you can easily break real applications
480 * if you don't tell them that something has hung up!
481 *
482 * Check-me.
483 *
484 * Check number 1. POLLHUP is _UNMASKABLE_ event (see UNIX98 and
485 * our fs/select.c). It means that after we received EOF,
486 * poll always returns immediately, making impossible poll() on write()
487 * in state CLOSE_WAIT. One solution is evident --- to set POLLHUP
488 * if and only if shutdown has been made in both directions.
489 * Actually, it is interesting to look how Solaris and DUX
490 * solve this dilemma. I would prefer, if POLLHUP were maskable,
491 * then we could set it on SND_SHUTDOWN. BTW examples given
492 * in Stevens' books assume exactly this behaviour, it explains
493 * why POLLHUP is incompatible with POLLOUT. --ANK
494 *
495 * NOTE. Check for TCP_CLOSE is added. The goal is to prevent
496 * blocking on fresh not-connected or disconnected socket. --ANK
497 */
498 if (sk->sk_shutdown == SHUTDOWN_MASK || state == TCP_CLOSE)
499 mask |= POLLHUP;
500 if (sk->sk_shutdown & RCV_SHUTDOWN)
501 mask |= POLLIN | POLLRDNORM | POLLRDHUP;
502
503 /* Connected or passive Fast Open socket? */
504 if (state != TCP_SYN_SENT &&
505 (state != TCP_SYN_RECV || tp->fastopen_rsk)) {
506 int target = sock_rcvlowat(sk, 0, INT_MAX);
507
508 if (tp->urg_seq == tp->copied_seq &&
509 !sock_flag(sk, SOCK_URGINLINE) &&
510 tp->urg_data)
511 target++;
512
513 if (tp->rcv_nxt - tp->copied_seq >= target)
514 mask |= POLLIN | POLLRDNORM;
515
516 if (!(sk->sk_shutdown & SEND_SHUTDOWN)) {
517 if (sk_stream_is_writeable(sk)) {
518 mask |= POLLOUT | POLLWRNORM;
519 } else { /* send SIGIO later */
520 sk_set_bit(SOCKWQ_ASYNC_NOSPACE, sk);
521 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
522
523 /* Race breaker. If space is freed after
524 * wspace test but before the flags are set,
525 * IO signal will be lost. Memory barrier
526 * pairs with the input side.
527 */
528 smp_mb__after_atomic();
529 if (sk_stream_is_writeable(sk))
530 mask |= POLLOUT | POLLWRNORM;
531 }
532 } else
533 mask |= POLLOUT | POLLWRNORM;
534
535 if (tp->urg_data & TCP_URG_VALID)
536 mask |= POLLPRI;
537 }
538 /* This barrier is coupled with smp_wmb() in tcp_reset() */
539 smp_rmb();
540 if (sk->sk_err || !skb_queue_empty(&sk->sk_error_queue))
541 mask |= POLLERR;
542
543 return mask;
544 }
545 EXPORT_SYMBOL(tcp_poll);
546
547 int tcp_ioctl(struct sock *sk, int cmd, unsigned long arg)
548 {
549 struct tcp_sock *tp = tcp_sk(sk);
550 int answ;
551 bool slow;
552
553 switch (cmd) {
554 case SIOCINQ:
555 if (sk->sk_state == TCP_LISTEN)
556 return -EINVAL;
557
558 slow = lock_sock_fast(sk);
559 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
560 answ = 0;
561 else if (sock_flag(sk, SOCK_URGINLINE) ||
562 !tp->urg_data ||
563 before(tp->urg_seq, tp->copied_seq) ||
564 !before(tp->urg_seq, tp->rcv_nxt)) {
565
566 answ = tp->rcv_nxt - tp->copied_seq;
567
568 /* Subtract 1, if FIN was received */
569 if (answ && sock_flag(sk, SOCK_DONE))
570 answ--;
571 } else
572 answ = tp->urg_seq - tp->copied_seq;
573 unlock_sock_fast(sk, slow);
574 break;
575 case SIOCATMARK:
576 answ = tp->urg_data && tp->urg_seq == tp->copied_seq;
577 break;
578 case SIOCOUTQ:
579 if (sk->sk_state == TCP_LISTEN)
580 return -EINVAL;
581
582 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
583 answ = 0;
584 else
585 answ = tp->write_seq - tp->snd_una;
586 break;
587 case SIOCOUTQNSD:
588 if (sk->sk_state == TCP_LISTEN)
589 return -EINVAL;
590
591 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
592 answ = 0;
593 else
594 answ = tp->write_seq - tp->snd_nxt;
595 break;
596 default:
597 return -ENOIOCTLCMD;
598 }
599
600 return put_user(answ, (int __user *)arg);
601 }
602 EXPORT_SYMBOL(tcp_ioctl);
603
604 static inline void tcp_mark_push(struct tcp_sock *tp, struct sk_buff *skb)
605 {
606 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
607 tp->pushed_seq = tp->write_seq;
608 }
609
610 static inline bool forced_push(const struct tcp_sock *tp)
611 {
612 return after(tp->write_seq, tp->pushed_seq + (tp->max_window >> 1));
613 }
614
615 static void skb_entail(struct sock *sk, struct sk_buff *skb)
616 {
617 struct tcp_sock *tp = tcp_sk(sk);
618 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
619
620 skb->csum = 0;
621 tcb->seq = tcb->end_seq = tp->write_seq;
622 tcb->tcp_flags = TCPHDR_ACK;
623 tcb->sacked = 0;
624 __skb_header_release(skb);
625 tcp_add_write_queue_tail(sk, skb);
626 sk->sk_wmem_queued += skb->truesize;
627 sk_mem_charge(sk, skb->truesize);
628 if (tp->nonagle & TCP_NAGLE_PUSH)
629 tp->nonagle &= ~TCP_NAGLE_PUSH;
630
631 tcp_slow_start_after_idle_check(sk);
632 }
633
634 static inline void tcp_mark_urg(struct tcp_sock *tp, int flags)
635 {
636 if (flags & MSG_OOB)
637 tp->snd_up = tp->write_seq;
638 }
639
640 /* If a not yet filled skb is pushed, do not send it if
641 * we have data packets in Qdisc or NIC queues :
642 * Because TX completion will happen shortly, it gives a chance
643 * to coalesce future sendmsg() payload into this skb, without
644 * need for a timer, and with no latency trade off.
645 * As packets containing data payload have a bigger truesize
646 * than pure acks (dataless) packets, the last checks prevent
647 * autocorking if we only have an ACK in Qdisc/NIC queues,
648 * or if TX completion was delayed after we processed ACK packet.
649 */
650 static bool tcp_should_autocork(struct sock *sk, struct sk_buff *skb,
651 int size_goal)
652 {
653 return skb->len < size_goal &&
654 sysctl_tcp_autocorking &&
655 skb != tcp_write_queue_head(sk) &&
656 atomic_read(&sk->sk_wmem_alloc) > skb->truesize;
657 }
658
659 static void tcp_push(struct sock *sk, int flags, int mss_now,
660 int nonagle, int size_goal)
661 {
662 struct tcp_sock *tp = tcp_sk(sk);
663 struct sk_buff *skb;
664
665 if (!tcp_send_head(sk))
666 return;
667
668 skb = tcp_write_queue_tail(sk);
669 if (!(flags & MSG_MORE) || forced_push(tp))
670 tcp_mark_push(tp, skb);
671
672 tcp_mark_urg(tp, flags);
673
674 if (tcp_should_autocork(sk, skb, size_goal)) {
675
676 /* avoid atomic op if TSQ_THROTTLED bit is already set */
677 if (!test_bit(TSQ_THROTTLED, &tp->tsq_flags)) {
678 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPAUTOCORKING);
679 set_bit(TSQ_THROTTLED, &tp->tsq_flags);
680 }
681 /* It is possible TX completion already happened
682 * before we set TSQ_THROTTLED.
683 */
684 if (atomic_read(&sk->sk_wmem_alloc) > skb->truesize)
685 return;
686 }
687
688 if (flags & MSG_MORE)
689 nonagle = TCP_NAGLE_CORK;
690
691 __tcp_push_pending_frames(sk, mss_now, nonagle);
692 }
693
694 static int tcp_splice_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb,
695 unsigned int offset, size_t len)
696 {
697 struct tcp_splice_state *tss = rd_desc->arg.data;
698 int ret;
699
700 ret = skb_splice_bits(skb, skb->sk, offset, tss->pipe,
701 min(rd_desc->count, len), tss->flags,
702 skb_socket_splice);
703 if (ret > 0)
704 rd_desc->count -= ret;
705 return ret;
706 }
707
708 static int __tcp_splice_read(struct sock *sk, struct tcp_splice_state *tss)
709 {
710 /* Store TCP splice context information in read_descriptor_t. */
711 read_descriptor_t rd_desc = {
712 .arg.data = tss,
713 .count = tss->len,
714 };
715
716 return tcp_read_sock(sk, &rd_desc, tcp_splice_data_recv);
717 }
718
719 /**
720 * tcp_splice_read - splice data from TCP socket to a pipe
721 * @sock: socket to splice from
722 * @ppos: position (not valid)
723 * @pipe: pipe to splice to
724 * @len: number of bytes to splice
725 * @flags: splice modifier flags
726 *
727 * Description:
728 * Will read pages from given socket and fill them into a pipe.
729 *
730 **/
731 ssize_t tcp_splice_read(struct socket *sock, loff_t *ppos,
732 struct pipe_inode_info *pipe, size_t len,
733 unsigned int flags)
734 {
735 struct sock *sk = sock->sk;
736 struct tcp_splice_state tss = {
737 .pipe = pipe,
738 .len = len,
739 .flags = flags,
740 };
741 long timeo;
742 ssize_t spliced;
743 int ret;
744
745 sock_rps_record_flow(sk);
746 /*
747 * We can't seek on a socket input
748 */
749 if (unlikely(*ppos))
750 return -ESPIPE;
751
752 ret = spliced = 0;
753
754 lock_sock(sk);
755
756 timeo = sock_rcvtimeo(sk, sock->file->f_flags & O_NONBLOCK);
757 while (tss.len) {
758 ret = __tcp_splice_read(sk, &tss);
759 if (ret < 0)
760 break;
761 else if (!ret) {
762 if (spliced)
763 break;
764 if (sock_flag(sk, SOCK_DONE))
765 break;
766 if (sk->sk_err) {
767 ret = sock_error(sk);
768 break;
769 }
770 if (sk->sk_shutdown & RCV_SHUTDOWN)
771 break;
772 if (sk->sk_state == TCP_CLOSE) {
773 /*
774 * This occurs when user tries to read
775 * from never connected socket.
776 */
777 if (!sock_flag(sk, SOCK_DONE))
778 ret = -ENOTCONN;
779 break;
780 }
781 if (!timeo) {
782 ret = -EAGAIN;
783 break;
784 }
785 sk_wait_data(sk, &timeo, NULL);
786 if (signal_pending(current)) {
787 ret = sock_intr_errno(timeo);
788 break;
789 }
790 continue;
791 }
792 tss.len -= ret;
793 spliced += ret;
794
795 if (!timeo)
796 break;
797 release_sock(sk);
798 lock_sock(sk);
799
800 if (sk->sk_err || sk->sk_state == TCP_CLOSE ||
801 (sk->sk_shutdown & RCV_SHUTDOWN) ||
802 signal_pending(current))
803 break;
804 }
805
806 release_sock(sk);
807
808 if (spliced)
809 return spliced;
810
811 return ret;
812 }
813 EXPORT_SYMBOL(tcp_splice_read);
814
815 struct sk_buff *sk_stream_alloc_skb(struct sock *sk, int size, gfp_t gfp,
816 bool force_schedule)
817 {
818 struct sk_buff *skb;
819
820 /* The TCP header must be at least 32-bit aligned. */
821 size = ALIGN(size, 4);
822
823 if (unlikely(tcp_under_memory_pressure(sk)))
824 sk_mem_reclaim_partial(sk);
825
826 skb = alloc_skb_fclone(size + sk->sk_prot->max_header, gfp);
827 if (likely(skb)) {
828 bool mem_scheduled;
829
830 if (force_schedule) {
831 mem_scheduled = true;
832 sk_forced_mem_schedule(sk, skb->truesize);
833 } else {
834 mem_scheduled = sk_wmem_schedule(sk, skb->truesize);
835 }
836 if (likely(mem_scheduled)) {
837 skb_reserve(skb, sk->sk_prot->max_header);
838 /*
839 * Make sure that we have exactly size bytes
840 * available to the caller, no more, no less.
841 */
842 skb->reserved_tailroom = skb->end - skb->tail - size;
843 return skb;
844 }
845 __kfree_skb(skb);
846 } else {
847 sk->sk_prot->enter_memory_pressure(sk);
848 sk_stream_moderate_sndbuf(sk);
849 }
850 return NULL;
851 }
852
853 static unsigned int tcp_xmit_size_goal(struct sock *sk, u32 mss_now,
854 int large_allowed)
855 {
856 struct tcp_sock *tp = tcp_sk(sk);
857 u32 new_size_goal, size_goal;
858
859 if (!large_allowed || !sk_can_gso(sk))
860 return mss_now;
861
862 /* Note : tcp_tso_autosize() will eventually split this later */
863 new_size_goal = sk->sk_gso_max_size - 1 - MAX_TCP_HEADER;
864 new_size_goal = tcp_bound_to_half_wnd(tp, new_size_goal);
865
866 /* We try hard to avoid divides here */
867 size_goal = tp->gso_segs * mss_now;
868 if (unlikely(new_size_goal < size_goal ||
869 new_size_goal >= size_goal + mss_now)) {
870 tp->gso_segs = min_t(u16, new_size_goal / mss_now,
871 sk->sk_gso_max_segs);
872 size_goal = tp->gso_segs * mss_now;
873 }
874
875 return max(size_goal, mss_now);
876 }
877
878 static int tcp_send_mss(struct sock *sk, int *size_goal, int flags)
879 {
880 int mss_now;
881
882 mss_now = tcp_current_mss(sk);
883 *size_goal = tcp_xmit_size_goal(sk, mss_now, !(flags & MSG_OOB));
884
885 return mss_now;
886 }
887
888 static ssize_t do_tcp_sendpages(struct sock *sk, struct page *page, int offset,
889 size_t size, int flags)
890 {
891 struct tcp_sock *tp = tcp_sk(sk);
892 int mss_now, size_goal;
893 int err;
894 ssize_t copied;
895 long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
896
897 /* Wait for a connection to finish. One exception is TCP Fast Open
898 * (passive side) where data is allowed to be sent before a connection
899 * is fully established.
900 */
901 if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) &&
902 !tcp_passive_fastopen(sk)) {
903 err = sk_stream_wait_connect(sk, &timeo);
904 if (err != 0)
905 goto out_err;
906 }
907
908 sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
909
910 mss_now = tcp_send_mss(sk, &size_goal, flags);
911 copied = 0;
912
913 err = -EPIPE;
914 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
915 goto out_err;
916
917 while (size > 0) {
918 struct sk_buff *skb = tcp_write_queue_tail(sk);
919 int copy, i;
920 bool can_coalesce;
921
922 if (!tcp_send_head(sk) || (copy = size_goal - skb->len) <= 0) {
923 new_segment:
924 if (!sk_stream_memory_free(sk))
925 goto wait_for_sndbuf;
926
927 skb = sk_stream_alloc_skb(sk, 0, sk->sk_allocation,
928 skb_queue_empty(&sk->sk_write_queue));
929 if (!skb)
930 goto wait_for_memory;
931
932 skb_entail(sk, skb);
933 copy = size_goal;
934 }
935
936 if (copy > size)
937 copy = size;
938
939 i = skb_shinfo(skb)->nr_frags;
940 can_coalesce = skb_can_coalesce(skb, i, page, offset);
941 if (!can_coalesce && i >= MAX_SKB_FRAGS) {
942 tcp_mark_push(tp, skb);
943 goto new_segment;
944 }
945 if (!sk_wmem_schedule(sk, copy))
946 goto wait_for_memory;
947
948 if (can_coalesce) {
949 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
950 } else {
951 get_page(page);
952 skb_fill_page_desc(skb, i, page, offset, copy);
953 }
954 skb_shinfo(skb)->tx_flags |= SKBTX_SHARED_FRAG;
955
956 skb->len += copy;
957 skb->data_len += copy;
958 skb->truesize += copy;
959 sk->sk_wmem_queued += copy;
960 sk_mem_charge(sk, copy);
961 skb->ip_summed = CHECKSUM_PARTIAL;
962 tp->write_seq += copy;
963 TCP_SKB_CB(skb)->end_seq += copy;
964 tcp_skb_pcount_set(skb, 0);
965
966 if (!copied)
967 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
968
969 copied += copy;
970 offset += copy;
971 size -= copy;
972 if (!size) {
973 tcp_tx_timestamp(sk, skb);
974 goto out;
975 }
976
977 if (skb->len < size_goal || (flags & MSG_OOB))
978 continue;
979
980 if (forced_push(tp)) {
981 tcp_mark_push(tp, skb);
982 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
983 } else if (skb == tcp_send_head(sk))
984 tcp_push_one(sk, mss_now);
985 continue;
986
987 wait_for_sndbuf:
988 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
989 wait_for_memory:
990 tcp_push(sk, flags & ~MSG_MORE, mss_now,
991 TCP_NAGLE_PUSH, size_goal);
992
993 err = sk_stream_wait_memory(sk, &timeo);
994 if (err != 0)
995 goto do_error;
996
997 mss_now = tcp_send_mss(sk, &size_goal, flags);
998 }
999
1000 out:
1001 if (copied && !(flags & MSG_SENDPAGE_NOTLAST))
1002 tcp_push(sk, flags, mss_now, tp->nonagle, size_goal);
1003 return copied;
1004
1005 do_error:
1006 if (copied)
1007 goto out;
1008 out_err:
1009 /* make sure we wake any epoll edge trigger waiter */
1010 if (unlikely(skb_queue_len(&sk->sk_write_queue) == 0 && err == -EAGAIN))
1011 sk->sk_write_space(sk);
1012 return sk_stream_error(sk, flags, err);
1013 }
1014
1015 int tcp_sendpage(struct sock *sk, struct page *page, int offset,
1016 size_t size, int flags)
1017 {
1018 ssize_t res;
1019
1020 if (!(sk->sk_route_caps & NETIF_F_SG) ||
1021 !(sk->sk_route_caps & NETIF_F_ALL_CSUM))
1022 return sock_no_sendpage(sk->sk_socket, page, offset, size,
1023 flags);
1024
1025 lock_sock(sk);
1026 res = do_tcp_sendpages(sk, page, offset, size, flags);
1027 release_sock(sk);
1028 return res;
1029 }
1030 EXPORT_SYMBOL(tcp_sendpage);
1031
1032 static inline int select_size(const struct sock *sk, bool sg)
1033 {
1034 const struct tcp_sock *tp = tcp_sk(sk);
1035 int tmp = tp->mss_cache;
1036
1037 if (sg) {
1038 if (sk_can_gso(sk)) {
1039 /* Small frames wont use a full page:
1040 * Payload will immediately follow tcp header.
1041 */
1042 tmp = SKB_WITH_OVERHEAD(2048 - MAX_TCP_HEADER);
1043 } else {
1044 int pgbreak = SKB_MAX_HEAD(MAX_TCP_HEADER);
1045
1046 if (tmp >= pgbreak &&
1047 tmp <= pgbreak + (MAX_SKB_FRAGS - 1) * PAGE_SIZE)
1048 tmp = pgbreak;
1049 }
1050 }
1051
1052 return tmp;
1053 }
1054
1055 void tcp_free_fastopen_req(struct tcp_sock *tp)
1056 {
1057 if (tp->fastopen_req) {
1058 kfree(tp->fastopen_req);
1059 tp->fastopen_req = NULL;
1060 }
1061 }
1062
1063 static int tcp_sendmsg_fastopen(struct sock *sk, struct msghdr *msg,
1064 int *copied, size_t size)
1065 {
1066 struct tcp_sock *tp = tcp_sk(sk);
1067 int err, flags;
1068
1069 if (!(sysctl_tcp_fastopen & TFO_CLIENT_ENABLE))
1070 return -EOPNOTSUPP;
1071 if (tp->fastopen_req)
1072 return -EALREADY; /* Another Fast Open is in progress */
1073
1074 tp->fastopen_req = kzalloc(sizeof(struct tcp_fastopen_request),
1075 sk->sk_allocation);
1076 if (unlikely(!tp->fastopen_req))
1077 return -ENOBUFS;
1078 tp->fastopen_req->data = msg;
1079 tp->fastopen_req->size = size;
1080
1081 flags = (msg->msg_flags & MSG_DONTWAIT) ? O_NONBLOCK : 0;
1082 err = __inet_stream_connect(sk->sk_socket, msg->msg_name,
1083 msg->msg_namelen, flags);
1084 *copied = tp->fastopen_req->copied;
1085 tcp_free_fastopen_req(tp);
1086 return err;
1087 }
1088
1089 int tcp_sendmsg(struct sock *sk, struct msghdr *msg, size_t size)
1090 {
1091 struct tcp_sock *tp = tcp_sk(sk);
1092 struct sk_buff *skb;
1093 int flags, err, copied = 0;
1094 int mss_now = 0, size_goal, copied_syn = 0;
1095 bool sg;
1096 long timeo;
1097
1098 lock_sock(sk);
1099
1100 flags = msg->msg_flags;
1101 if (flags & MSG_FASTOPEN) {
1102 err = tcp_sendmsg_fastopen(sk, msg, &copied_syn, size);
1103 if (err == -EINPROGRESS && copied_syn > 0)
1104 goto out;
1105 else if (err)
1106 goto out_err;
1107 }
1108
1109 timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
1110
1111 /* Wait for a connection to finish. One exception is TCP Fast Open
1112 * (passive side) where data is allowed to be sent before a connection
1113 * is fully established.
1114 */
1115 if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) &&
1116 !tcp_passive_fastopen(sk)) {
1117 err = sk_stream_wait_connect(sk, &timeo);
1118 if (err != 0)
1119 goto do_error;
1120 }
1121
1122 if (unlikely(tp->repair)) {
1123 if (tp->repair_queue == TCP_RECV_QUEUE) {
1124 copied = tcp_send_rcvq(sk, msg, size);
1125 goto out_nopush;
1126 }
1127
1128 err = -EINVAL;
1129 if (tp->repair_queue == TCP_NO_QUEUE)
1130 goto out_err;
1131
1132 /* 'common' sending to sendq */
1133 }
1134
1135 /* This should be in poll */
1136 sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
1137
1138 mss_now = tcp_send_mss(sk, &size_goal, flags);
1139
1140 /* Ok commence sending. */
1141 copied = 0;
1142
1143 err = -EPIPE;
1144 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
1145 goto out_err;
1146
1147 sg = !!(sk->sk_route_caps & NETIF_F_SG);
1148
1149 while (msg_data_left(msg)) {
1150 int copy = 0;
1151 int max = size_goal;
1152
1153 skb = tcp_write_queue_tail(sk);
1154 if (tcp_send_head(sk)) {
1155 if (skb->ip_summed == CHECKSUM_NONE)
1156 max = mss_now;
1157 copy = max - skb->len;
1158 }
1159
1160 if (copy <= 0) {
1161 new_segment:
1162 /* Allocate new segment. If the interface is SG,
1163 * allocate skb fitting to single page.
1164 */
1165 if (!sk_stream_memory_free(sk))
1166 goto wait_for_sndbuf;
1167
1168 skb = sk_stream_alloc_skb(sk,
1169 select_size(sk, sg),
1170 sk->sk_allocation,
1171 skb_queue_empty(&sk->sk_write_queue));
1172 if (!skb)
1173 goto wait_for_memory;
1174
1175 /*
1176 * Check whether we can use HW checksum.
1177 */
1178 if (sk->sk_route_caps & NETIF_F_ALL_CSUM)
1179 skb->ip_summed = CHECKSUM_PARTIAL;
1180
1181 skb_entail(sk, skb);
1182 copy = size_goal;
1183 max = size_goal;
1184
1185 /* All packets are restored as if they have
1186 * already been sent. skb_mstamp isn't set to
1187 * avoid wrong rtt estimation.
1188 */
1189 if (tp->repair)
1190 TCP_SKB_CB(skb)->sacked |= TCPCB_REPAIRED;
1191 }
1192
1193 /* Try to append data to the end of skb. */
1194 if (copy > msg_data_left(msg))
1195 copy = msg_data_left(msg);
1196
1197 /* Where to copy to? */
1198 if (skb_availroom(skb) > 0) {
1199 /* We have some space in skb head. Superb! */
1200 copy = min_t(int, copy, skb_availroom(skb));
1201 err = skb_add_data_nocache(sk, skb, &msg->msg_iter, copy);
1202 if (err)
1203 goto do_fault;
1204 } else {
1205 bool merge = true;
1206 int i = skb_shinfo(skb)->nr_frags;
1207 struct page_frag *pfrag = sk_page_frag(sk);
1208
1209 if (!sk_page_frag_refill(sk, pfrag))
1210 goto wait_for_memory;
1211
1212 if (!skb_can_coalesce(skb, i, pfrag->page,
1213 pfrag->offset)) {
1214 if (i == MAX_SKB_FRAGS || !sg) {
1215 tcp_mark_push(tp, skb);
1216 goto new_segment;
1217 }
1218 merge = false;
1219 }
1220
1221 copy = min_t(int, copy, pfrag->size - pfrag->offset);
1222
1223 if (!sk_wmem_schedule(sk, copy))
1224 goto wait_for_memory;
1225
1226 err = skb_copy_to_page_nocache(sk, &msg->msg_iter, skb,
1227 pfrag->page,
1228 pfrag->offset,
1229 copy);
1230 if (err)
1231 goto do_error;
1232
1233 /* Update the skb. */
1234 if (merge) {
1235 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1236 } else {
1237 skb_fill_page_desc(skb, i, pfrag->page,
1238 pfrag->offset, copy);
1239 get_page(pfrag->page);
1240 }
1241 pfrag->offset += copy;
1242 }
1243
1244 if (!copied)
1245 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
1246
1247 tp->write_seq += copy;
1248 TCP_SKB_CB(skb)->end_seq += copy;
1249 tcp_skb_pcount_set(skb, 0);
1250
1251 copied += copy;
1252 if (!msg_data_left(msg)) {
1253 tcp_tx_timestamp(sk, skb);
1254 goto out;
1255 }
1256
1257 if (skb->len < max || (flags & MSG_OOB) || unlikely(tp->repair))
1258 continue;
1259
1260 if (forced_push(tp)) {
1261 tcp_mark_push(tp, skb);
1262 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
1263 } else if (skb == tcp_send_head(sk))
1264 tcp_push_one(sk, mss_now);
1265 continue;
1266
1267 wait_for_sndbuf:
1268 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1269 wait_for_memory:
1270 if (copied)
1271 tcp_push(sk, flags & ~MSG_MORE, mss_now,
1272 TCP_NAGLE_PUSH, size_goal);
1273
1274 err = sk_stream_wait_memory(sk, &timeo);
1275 if (err != 0)
1276 goto do_error;
1277
1278 mss_now = tcp_send_mss(sk, &size_goal, flags);
1279 }
1280
1281 out:
1282 if (copied)
1283 tcp_push(sk, flags, mss_now, tp->nonagle, size_goal);
1284 out_nopush:
1285 release_sock(sk);
1286 return copied + copied_syn;
1287
1288 do_fault:
1289 if (!skb->len) {
1290 tcp_unlink_write_queue(skb, sk);
1291 /* It is the one place in all of TCP, except connection
1292 * reset, where we can be unlinking the send_head.
1293 */
1294 tcp_check_send_head(sk, skb);
1295 sk_wmem_free_skb(sk, skb);
1296 }
1297
1298 do_error:
1299 if (copied + copied_syn)
1300 goto out;
1301 out_err:
1302 err = sk_stream_error(sk, flags, err);
1303 /* make sure we wake any epoll edge trigger waiter */
1304 if (unlikely(skb_queue_len(&sk->sk_write_queue) == 0 && err == -EAGAIN))
1305 sk->sk_write_space(sk);
1306 release_sock(sk);
1307 return err;
1308 }
1309 EXPORT_SYMBOL(tcp_sendmsg);
1310
1311 /*
1312 * Handle reading urgent data. BSD has very simple semantics for
1313 * this, no blocking and very strange errors 8)
1314 */
1315
1316 static int tcp_recv_urg(struct sock *sk, struct msghdr *msg, int len, int flags)
1317 {
1318 struct tcp_sock *tp = tcp_sk(sk);
1319
1320 /* No URG data to read. */
1321 if (sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data ||
1322 tp->urg_data == TCP_URG_READ)
1323 return -EINVAL; /* Yes this is right ! */
1324
1325 if (sk->sk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DONE))
1326 return -ENOTCONN;
1327
1328 if (tp->urg_data & TCP_URG_VALID) {
1329 int err = 0;
1330 char c = tp->urg_data;
1331
1332 if (!(flags & MSG_PEEK))
1333 tp->urg_data = TCP_URG_READ;
1334
1335 /* Read urgent data. */
1336 msg->msg_flags |= MSG_OOB;
1337
1338 if (len > 0) {
1339 if (!(flags & MSG_TRUNC))
1340 err = memcpy_to_msg(msg, &c, 1);
1341 len = 1;
1342 } else
1343 msg->msg_flags |= MSG_TRUNC;
1344
1345 return err ? -EFAULT : len;
1346 }
1347
1348 if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
1349 return 0;
1350
1351 /* Fixed the recv(..., MSG_OOB) behaviour. BSD docs and
1352 * the available implementations agree in this case:
1353 * this call should never block, independent of the
1354 * blocking state of the socket.
1355 * Mike <pall@rz.uni-karlsruhe.de>
1356 */
1357 return -EAGAIN;
1358 }
1359
1360 static int tcp_peek_sndq(struct sock *sk, struct msghdr *msg, int len)
1361 {
1362 struct sk_buff *skb;
1363 int copied = 0, err = 0;
1364
1365 /* XXX -- need to support SO_PEEK_OFF */
1366
1367 skb_queue_walk(&sk->sk_write_queue, skb) {
1368 err = skb_copy_datagram_msg(skb, 0, msg, skb->len);
1369 if (err)
1370 break;
1371
1372 copied += skb->len;
1373 }
1374
1375 return err ?: copied;
1376 }
1377
1378 /* Clean up the receive buffer for full frames taken by the user,
1379 * then send an ACK if necessary. COPIED is the number of bytes
1380 * tcp_recvmsg has given to the user so far, it speeds up the
1381 * calculation of whether or not we must ACK for the sake of
1382 * a window update.
1383 */
1384 static void tcp_cleanup_rbuf(struct sock *sk, int copied)
1385 {
1386 struct tcp_sock *tp = tcp_sk(sk);
1387 bool time_to_ack = false;
1388
1389 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
1390
1391 WARN(skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq),
1392 "cleanup rbuf bug: copied %X seq %X rcvnxt %X\n",
1393 tp->copied_seq, TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt);
1394
1395 if (inet_csk_ack_scheduled(sk)) {
1396 const struct inet_connection_sock *icsk = inet_csk(sk);
1397 /* Delayed ACKs frequently hit locked sockets during bulk
1398 * receive. */
1399 if (icsk->icsk_ack.blocked ||
1400 /* Once-per-two-segments ACK was not sent by tcp_input.c */
1401 tp->rcv_nxt - tp->rcv_wup > icsk->icsk_ack.rcv_mss ||
1402 /*
1403 * If this read emptied read buffer, we send ACK, if
1404 * connection is not bidirectional, user drained
1405 * receive buffer and there was a small segment
1406 * in queue.
1407 */
1408 (copied > 0 &&
1409 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED2) ||
1410 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED) &&
1411 !icsk->icsk_ack.pingpong)) &&
1412 !atomic_read(&sk->sk_rmem_alloc)))
1413 time_to_ack = true;
1414 }
1415
1416 /* We send an ACK if we can now advertise a non-zero window
1417 * which has been raised "significantly".
1418 *
1419 * Even if window raised up to infinity, do not send window open ACK
1420 * in states, where we will not receive more. It is useless.
1421 */
1422 if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) {
1423 __u32 rcv_window_now = tcp_receive_window(tp);
1424
1425 /* Optimize, __tcp_select_window() is not cheap. */
1426 if (2*rcv_window_now <= tp->window_clamp) {
1427 __u32 new_window = __tcp_select_window(sk);
1428
1429 /* Send ACK now, if this read freed lots of space
1430 * in our buffer. Certainly, new_window is new window.
1431 * We can advertise it now, if it is not less than current one.
1432 * "Lots" means "at least twice" here.
1433 */
1434 if (new_window && new_window >= 2 * rcv_window_now)
1435 time_to_ack = true;
1436 }
1437 }
1438 if (time_to_ack)
1439 tcp_send_ack(sk);
1440 }
1441
1442 static void tcp_prequeue_process(struct sock *sk)
1443 {
1444 struct sk_buff *skb;
1445 struct tcp_sock *tp = tcp_sk(sk);
1446
1447 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPPREQUEUED);
1448
1449 /* RX process wants to run with disabled BHs, though it is not
1450 * necessary */
1451 local_bh_disable();
1452 while ((skb = __skb_dequeue(&tp->ucopy.prequeue)) != NULL)
1453 sk_backlog_rcv(sk, skb);
1454 local_bh_enable();
1455
1456 /* Clear memory counter. */
1457 tp->ucopy.memory = 0;
1458 }
1459
1460 static struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off)
1461 {
1462 struct sk_buff *skb;
1463 u32 offset;
1464
1465 while ((skb = skb_peek(&sk->sk_receive_queue)) != NULL) {
1466 offset = seq - TCP_SKB_CB(skb)->seq;
1467 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
1468 offset--;
1469 if (offset < skb->len || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)) {
1470 *off = offset;
1471 return skb;
1472 }
1473 /* This looks weird, but this can happen if TCP collapsing
1474 * splitted a fat GRO packet, while we released socket lock
1475 * in skb_splice_bits()
1476 */
1477 sk_eat_skb(sk, skb);
1478 }
1479 return NULL;
1480 }
1481
1482 /*
1483 * This routine provides an alternative to tcp_recvmsg() for routines
1484 * that would like to handle copying from skbuffs directly in 'sendfile'
1485 * fashion.
1486 * Note:
1487 * - It is assumed that the socket was locked by the caller.
1488 * - The routine does not block.
1489 * - At present, there is no support for reading OOB data
1490 * or for 'peeking' the socket using this routine
1491 * (although both would be easy to implement).
1492 */
1493 int tcp_read_sock(struct sock *sk, read_descriptor_t *desc,
1494 sk_read_actor_t recv_actor)
1495 {
1496 struct sk_buff *skb;
1497 struct tcp_sock *tp = tcp_sk(sk);
1498 u32 seq = tp->copied_seq;
1499 u32 offset;
1500 int copied = 0;
1501
1502 if (sk->sk_state == TCP_LISTEN)
1503 return -ENOTCONN;
1504 while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1505 if (offset < skb->len) {
1506 int used;
1507 size_t len;
1508
1509 len = skb->len - offset;
1510 /* Stop reading if we hit a patch of urgent data */
1511 if (tp->urg_data) {
1512 u32 urg_offset = tp->urg_seq - seq;
1513 if (urg_offset < len)
1514 len = urg_offset;
1515 if (!len)
1516 break;
1517 }
1518 used = recv_actor(desc, skb, offset, len);
1519 if (used <= 0) {
1520 if (!copied)
1521 copied = used;
1522 break;
1523 } else if (used <= len) {
1524 seq += used;
1525 copied += used;
1526 offset += used;
1527 }
1528 /* If recv_actor drops the lock (e.g. TCP splice
1529 * receive) the skb pointer might be invalid when
1530 * getting here: tcp_collapse might have deleted it
1531 * while aggregating skbs from the socket queue.
1532 */
1533 skb = tcp_recv_skb(sk, seq - 1, &offset);
1534 if (!skb)
1535 break;
1536 /* TCP coalescing might have appended data to the skb.
1537 * Try to splice more frags
1538 */
1539 if (offset + 1 != skb->len)
1540 continue;
1541 }
1542 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) {
1543 sk_eat_skb(sk, skb);
1544 ++seq;
1545 break;
1546 }
1547 sk_eat_skb(sk, skb);
1548 if (!desc->count)
1549 break;
1550 tp->copied_seq = seq;
1551 }
1552 tp->copied_seq = seq;
1553
1554 tcp_rcv_space_adjust(sk);
1555
1556 /* Clean up data we have read: This will do ACK frames. */
1557 if (copied > 0) {
1558 tcp_recv_skb(sk, seq, &offset);
1559 tcp_cleanup_rbuf(sk, copied);
1560 }
1561 return copied;
1562 }
1563 EXPORT_SYMBOL(tcp_read_sock);
1564
1565 /*
1566 * This routine copies from a sock struct into the user buffer.
1567 *
1568 * Technical note: in 2.3 we work on _locked_ socket, so that
1569 * tricks with *seq access order and skb->users are not required.
1570 * Probably, code can be easily improved even more.
1571 */
1572
1573 int tcp_recvmsg(struct sock *sk, struct msghdr *msg, size_t len, int nonblock,
1574 int flags, int *addr_len)
1575 {
1576 struct tcp_sock *tp = tcp_sk(sk);
1577 int copied = 0;
1578 u32 peek_seq;
1579 u32 *seq;
1580 unsigned long used;
1581 int err;
1582 int target; /* Read at least this many bytes */
1583 long timeo;
1584 struct task_struct *user_recv = NULL;
1585 struct sk_buff *skb, *last;
1586 u32 urg_hole = 0;
1587
1588 if (unlikely(flags & MSG_ERRQUEUE))
1589 return inet_recv_error(sk, msg, len, addr_len);
1590
1591 if (sk_can_busy_loop(sk) && skb_queue_empty(&sk->sk_receive_queue) &&
1592 (sk->sk_state == TCP_ESTABLISHED))
1593 sk_busy_loop(sk, nonblock);
1594
1595 lock_sock(sk);
1596
1597 err = -ENOTCONN;
1598 if (sk->sk_state == TCP_LISTEN)
1599 goto out;
1600
1601 timeo = sock_rcvtimeo(sk, nonblock);
1602
1603 /* Urgent data needs to be handled specially. */
1604 if (flags & MSG_OOB)
1605 goto recv_urg;
1606
1607 if (unlikely(tp->repair)) {
1608 err = -EPERM;
1609 if (!(flags & MSG_PEEK))
1610 goto out;
1611
1612 if (tp->repair_queue == TCP_SEND_QUEUE)
1613 goto recv_sndq;
1614
1615 err = -EINVAL;
1616 if (tp->repair_queue == TCP_NO_QUEUE)
1617 goto out;
1618
1619 /* 'common' recv queue MSG_PEEK-ing */
1620 }
1621
1622 seq = &tp->copied_seq;
1623 if (flags & MSG_PEEK) {
1624 peek_seq = tp->copied_seq;
1625 seq = &peek_seq;
1626 }
1627
1628 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
1629
1630 do {
1631 u32 offset;
1632
1633 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
1634 if (tp->urg_data && tp->urg_seq == *seq) {
1635 if (copied)
1636 break;
1637 if (signal_pending(current)) {
1638 copied = timeo ? sock_intr_errno(timeo) : -EAGAIN;
1639 break;
1640 }
1641 }
1642
1643 /* Next get a buffer. */
1644
1645 last = skb_peek_tail(&sk->sk_receive_queue);
1646 skb_queue_walk(&sk->sk_receive_queue, skb) {
1647 last = 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_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
1659 offset--;
1660 if (offset < skb->len)
1661 goto found_ok_skb;
1662 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_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.msg = msg;
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 if (copied >= target) {
1763 /* Do not sleep, just process backlog. */
1764 release_sock(sk);
1765 lock_sock(sk);
1766 } else {
1767 sk_wait_data(sk, &timeo, last);
1768 }
1769
1770 if (user_recv) {
1771 int chunk;
1772
1773 /* __ Restore normal policy in scheduler __ */
1774
1775 chunk = len - tp->ucopy.len;
1776 if (chunk != 0) {
1777 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMBACKLOG, chunk);
1778 len -= chunk;
1779 copied += chunk;
1780 }
1781
1782 if (tp->rcv_nxt == tp->copied_seq &&
1783 !skb_queue_empty(&tp->ucopy.prequeue)) {
1784 do_prequeue:
1785 tcp_prequeue_process(sk);
1786
1787 chunk = len - tp->ucopy.len;
1788 if (chunk != 0) {
1789 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1790 len -= chunk;
1791 copied += chunk;
1792 }
1793 }
1794 }
1795 if ((flags & MSG_PEEK) &&
1796 (peek_seq - copied - urg_hole != tp->copied_seq)) {
1797 net_dbg_ratelimited("TCP(%s:%d): Application bug, race in MSG_PEEK\n",
1798 current->comm,
1799 task_pid_nr(current));
1800 peek_seq = tp->copied_seq;
1801 }
1802 continue;
1803
1804 found_ok_skb:
1805 /* Ok so how much can we use? */
1806 used = skb->len - offset;
1807 if (len < used)
1808 used = len;
1809
1810 /* Do we have urgent data here? */
1811 if (tp->urg_data) {
1812 u32 urg_offset = tp->urg_seq - *seq;
1813 if (urg_offset < used) {
1814 if (!urg_offset) {
1815 if (!sock_flag(sk, SOCK_URGINLINE)) {
1816 ++*seq;
1817 urg_hole++;
1818 offset++;
1819 used--;
1820 if (!used)
1821 goto skip_copy;
1822 }
1823 } else
1824 used = urg_offset;
1825 }
1826 }
1827
1828 if (!(flags & MSG_TRUNC)) {
1829 err = skb_copy_datagram_msg(skb, offset, msg, used);
1830 if (err) {
1831 /* Exception. Bailout! */
1832 if (!copied)
1833 copied = -EFAULT;
1834 break;
1835 }
1836 }
1837
1838 *seq += used;
1839 copied += used;
1840 len -= used;
1841
1842 tcp_rcv_space_adjust(sk);
1843
1844 skip_copy:
1845 if (tp->urg_data && after(tp->copied_seq, tp->urg_seq)) {
1846 tp->urg_data = 0;
1847 tcp_fast_path_check(sk);
1848 }
1849 if (used + offset < skb->len)
1850 continue;
1851
1852 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1853 goto found_fin_ok;
1854 if (!(flags & MSG_PEEK))
1855 sk_eat_skb(sk, skb);
1856 continue;
1857
1858 found_fin_ok:
1859 /* Process the FIN. */
1860 ++*seq;
1861 if (!(flags & MSG_PEEK))
1862 sk_eat_skb(sk, skb);
1863 break;
1864 } while (len > 0);
1865
1866 if (user_recv) {
1867 if (!skb_queue_empty(&tp->ucopy.prequeue)) {
1868 int chunk;
1869
1870 tp->ucopy.len = copied > 0 ? len : 0;
1871
1872 tcp_prequeue_process(sk);
1873
1874 if (copied > 0 && (chunk = len - tp->ucopy.len) != 0) {
1875 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1876 len -= chunk;
1877 copied += chunk;
1878 }
1879 }
1880
1881 tp->ucopy.task = NULL;
1882 tp->ucopy.len = 0;
1883 }
1884
1885 /* According to UNIX98, msg_name/msg_namelen are ignored
1886 * on connected socket. I was just happy when found this 8) --ANK
1887 */
1888
1889 /* Clean up data we have read: This will do ACK frames. */
1890 tcp_cleanup_rbuf(sk, copied);
1891
1892 release_sock(sk);
1893 return copied;
1894
1895 out:
1896 release_sock(sk);
1897 return err;
1898
1899 recv_urg:
1900 err = tcp_recv_urg(sk, msg, len, flags);
1901 goto out;
1902
1903 recv_sndq:
1904 err = tcp_peek_sndq(sk, msg, len);
1905 goto out;
1906 }
1907 EXPORT_SYMBOL(tcp_recvmsg);
1908
1909 void tcp_set_state(struct sock *sk, int state)
1910 {
1911 int oldstate = sk->sk_state;
1912
1913 switch (state) {
1914 case TCP_ESTABLISHED:
1915 if (oldstate != TCP_ESTABLISHED)
1916 TCP_INC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
1917 break;
1918
1919 case TCP_CLOSE:
1920 if (oldstate == TCP_CLOSE_WAIT || oldstate == TCP_ESTABLISHED)
1921 TCP_INC_STATS(sock_net(sk), TCP_MIB_ESTABRESETS);
1922
1923 sk->sk_prot->unhash(sk);
1924 if (inet_csk(sk)->icsk_bind_hash &&
1925 !(sk->sk_userlocks & SOCK_BINDPORT_LOCK))
1926 inet_put_port(sk);
1927 /* fall through */
1928 default:
1929 if (oldstate == TCP_ESTABLISHED)
1930 TCP_DEC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
1931 }
1932
1933 /* Change state AFTER socket is unhashed to avoid closed
1934 * socket sitting in hash tables.
1935 */
1936 sk_state_store(sk, state);
1937
1938 #ifdef STATE_TRACE
1939 SOCK_DEBUG(sk, "TCP sk=%p, State %s -> %s\n", sk, statename[oldstate], statename[state]);
1940 #endif
1941 }
1942 EXPORT_SYMBOL_GPL(tcp_set_state);
1943
1944 /*
1945 * State processing on a close. This implements the state shift for
1946 * sending our FIN frame. Note that we only send a FIN for some
1947 * states. A shutdown() may have already sent the FIN, or we may be
1948 * closed.
1949 */
1950
1951 static const unsigned char new_state[16] = {
1952 /* current state: new state: action: */
1953 [0 /* (Invalid) */] = TCP_CLOSE,
1954 [TCP_ESTABLISHED] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1955 [TCP_SYN_SENT] = TCP_CLOSE,
1956 [TCP_SYN_RECV] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1957 [TCP_FIN_WAIT1] = TCP_FIN_WAIT1,
1958 [TCP_FIN_WAIT2] = TCP_FIN_WAIT2,
1959 [TCP_TIME_WAIT] = TCP_CLOSE,
1960 [TCP_CLOSE] = TCP_CLOSE,
1961 [TCP_CLOSE_WAIT] = TCP_LAST_ACK | TCP_ACTION_FIN,
1962 [TCP_LAST_ACK] = TCP_LAST_ACK,
1963 [TCP_LISTEN] = TCP_CLOSE,
1964 [TCP_CLOSING] = TCP_CLOSING,
1965 [TCP_NEW_SYN_RECV] = TCP_CLOSE, /* should not happen ! */
1966 };
1967
1968 static int tcp_close_state(struct sock *sk)
1969 {
1970 int next = (int)new_state[sk->sk_state];
1971 int ns = next & TCP_STATE_MASK;
1972
1973 tcp_set_state(sk, ns);
1974
1975 return next & TCP_ACTION_FIN;
1976 }
1977
1978 /*
1979 * Shutdown the sending side of a connection. Much like close except
1980 * that we don't receive shut down or sock_set_flag(sk, SOCK_DEAD).
1981 */
1982
1983 void tcp_shutdown(struct sock *sk, int how)
1984 {
1985 /* We need to grab some memory, and put together a FIN,
1986 * and then put it into the queue to be sent.
1987 * Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
1988 */
1989 if (!(how & SEND_SHUTDOWN))
1990 return;
1991
1992 /* If we've already sent a FIN, or it's a closed state, skip this. */
1993 if ((1 << sk->sk_state) &
1994 (TCPF_ESTABLISHED | TCPF_SYN_SENT |
1995 TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) {
1996 /* Clear out any half completed packets. FIN if needed. */
1997 if (tcp_close_state(sk))
1998 tcp_send_fin(sk);
1999 }
2000 }
2001 EXPORT_SYMBOL(tcp_shutdown);
2002
2003 bool tcp_check_oom(struct sock *sk, int shift)
2004 {
2005 bool too_many_orphans, out_of_socket_memory;
2006
2007 too_many_orphans = tcp_too_many_orphans(sk, shift);
2008 out_of_socket_memory = tcp_out_of_memory(sk);
2009
2010 if (too_many_orphans)
2011 net_info_ratelimited("too many orphaned sockets\n");
2012 if (out_of_socket_memory)
2013 net_info_ratelimited("out of memory -- consider tuning tcp_mem\n");
2014 return too_many_orphans || out_of_socket_memory;
2015 }
2016
2017 void tcp_close(struct sock *sk, long timeout)
2018 {
2019 struct sk_buff *skb;
2020 int data_was_unread = 0;
2021 int state;
2022
2023 lock_sock(sk);
2024 sk->sk_shutdown = SHUTDOWN_MASK;
2025
2026 if (sk->sk_state == TCP_LISTEN) {
2027 tcp_set_state(sk, TCP_CLOSE);
2028
2029 /* Special case. */
2030 inet_csk_listen_stop(sk);
2031
2032 goto adjudge_to_death;
2033 }
2034
2035 /* We need to flush the recv. buffs. We do this only on the
2036 * descriptor close, not protocol-sourced closes, because the
2037 * reader process may not have drained the data yet!
2038 */
2039 while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) {
2040 u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq;
2041
2042 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
2043 len--;
2044 data_was_unread += len;
2045 __kfree_skb(skb);
2046 }
2047
2048 sk_mem_reclaim(sk);
2049
2050 /* If socket has been already reset (e.g. in tcp_reset()) - kill it. */
2051 if (sk->sk_state == TCP_CLOSE)
2052 goto adjudge_to_death;
2053
2054 /* As outlined in RFC 2525, section 2.17, we send a RST here because
2055 * data was lost. To witness the awful effects of the old behavior of
2056 * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk
2057 * GET in an FTP client, suspend the process, wait for the client to
2058 * advertise a zero window, then kill -9 the FTP client, wheee...
2059 * Note: timeout is always zero in such a case.
2060 */
2061 if (unlikely(tcp_sk(sk)->repair)) {
2062 sk->sk_prot->disconnect(sk, 0);
2063 } else if (data_was_unread) {
2064 /* Unread data was tossed, zap the connection. */
2065 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONCLOSE);
2066 tcp_set_state(sk, TCP_CLOSE);
2067 tcp_send_active_reset(sk, sk->sk_allocation);
2068 } else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) {
2069 /* Check zero linger _after_ checking for unread data. */
2070 sk->sk_prot->disconnect(sk, 0);
2071 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
2072 } else if (tcp_close_state(sk)) {
2073 /* We FIN if the application ate all the data before
2074 * zapping the connection.
2075 */
2076
2077 /* RED-PEN. Formally speaking, we have broken TCP state
2078 * machine. State transitions:
2079 *
2080 * TCP_ESTABLISHED -> TCP_FIN_WAIT1
2081 * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible)
2082 * TCP_CLOSE_WAIT -> TCP_LAST_ACK
2083 *
2084 * are legal only when FIN has been sent (i.e. in window),
2085 * rather than queued out of window. Purists blame.
2086 *
2087 * F.e. "RFC state" is ESTABLISHED,
2088 * if Linux state is FIN-WAIT-1, but FIN is still not sent.
2089 *
2090 * The visible declinations are that sometimes
2091 * we enter time-wait state, when it is not required really
2092 * (harmless), do not send active resets, when they are
2093 * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
2094 * they look as CLOSING or LAST_ACK for Linux)
2095 * Probably, I missed some more holelets.
2096 * --ANK
2097 * XXX (TFO) - To start off we don't support SYN+ACK+FIN
2098 * in a single packet! (May consider it later but will
2099 * probably need API support or TCP_CORK SYN-ACK until
2100 * data is written and socket is closed.)
2101 */
2102 tcp_send_fin(sk);
2103 }
2104
2105 sk_stream_wait_close(sk, timeout);
2106
2107 adjudge_to_death:
2108 state = sk->sk_state;
2109 sock_hold(sk);
2110 sock_orphan(sk);
2111
2112 /* It is the last release_sock in its life. It will remove backlog. */
2113 release_sock(sk);
2114
2115
2116 /* Now socket is owned by kernel and we acquire BH lock
2117 to finish close. No need to check for user refs.
2118 */
2119 local_bh_disable();
2120 bh_lock_sock(sk);
2121 WARN_ON(sock_owned_by_user(sk));
2122
2123 percpu_counter_inc(sk->sk_prot->orphan_count);
2124
2125 /* Have we already been destroyed by a softirq or backlog? */
2126 if (state != TCP_CLOSE && sk->sk_state == TCP_CLOSE)
2127 goto out;
2128
2129 /* This is a (useful) BSD violating of the RFC. There is a
2130 * problem with TCP as specified in that the other end could
2131 * keep a socket open forever with no application left this end.
2132 * We use a 1 minute timeout (about the same as BSD) then kill
2133 * our end. If they send after that then tough - BUT: long enough
2134 * that we won't make the old 4*rto = almost no time - whoops
2135 * reset mistake.
2136 *
2137 * Nope, it was not mistake. It is really desired behaviour
2138 * f.e. on http servers, when such sockets are useless, but
2139 * consume significant resources. Let's do it with special
2140 * linger2 option. --ANK
2141 */
2142
2143 if (sk->sk_state == TCP_FIN_WAIT2) {
2144 struct tcp_sock *tp = tcp_sk(sk);
2145 if (tp->linger2 < 0) {
2146 tcp_set_state(sk, TCP_CLOSE);
2147 tcp_send_active_reset(sk, GFP_ATOMIC);
2148 NET_INC_STATS_BH(sock_net(sk),
2149 LINUX_MIB_TCPABORTONLINGER);
2150 } else {
2151 const int tmo = tcp_fin_time(sk);
2152
2153 if (tmo > TCP_TIMEWAIT_LEN) {
2154 inet_csk_reset_keepalive_timer(sk,
2155 tmo - TCP_TIMEWAIT_LEN);
2156 } else {
2157 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
2158 goto out;
2159 }
2160 }
2161 }
2162 if (sk->sk_state != TCP_CLOSE) {
2163 sk_mem_reclaim(sk);
2164 if (tcp_check_oom(sk, 0)) {
2165 tcp_set_state(sk, TCP_CLOSE);
2166 tcp_send_active_reset(sk, GFP_ATOMIC);
2167 NET_INC_STATS_BH(sock_net(sk),
2168 LINUX_MIB_TCPABORTONMEMORY);
2169 }
2170 }
2171
2172 if (sk->sk_state == TCP_CLOSE) {
2173 struct request_sock *req = tcp_sk(sk)->fastopen_rsk;
2174 /* We could get here with a non-NULL req if the socket is
2175 * aborted (e.g., closed with unread data) before 3WHS
2176 * finishes.
2177 */
2178 if (req)
2179 reqsk_fastopen_remove(sk, req, false);
2180 inet_csk_destroy_sock(sk);
2181 }
2182 /* Otherwise, socket is reprieved until protocol close. */
2183
2184 out:
2185 bh_unlock_sock(sk);
2186 local_bh_enable();
2187 sock_put(sk);
2188 }
2189 EXPORT_SYMBOL(tcp_close);
2190
2191 /* These states need RST on ABORT according to RFC793 */
2192
2193 static inline bool tcp_need_reset(int state)
2194 {
2195 return (1 << state) &
2196 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 |
2197 TCPF_FIN_WAIT2 | TCPF_SYN_RECV);
2198 }
2199
2200 int tcp_disconnect(struct sock *sk, int flags)
2201 {
2202 struct inet_sock *inet = inet_sk(sk);
2203 struct inet_connection_sock *icsk = inet_csk(sk);
2204 struct tcp_sock *tp = tcp_sk(sk);
2205 int err = 0;
2206 int old_state = sk->sk_state;
2207
2208 if (old_state != TCP_CLOSE)
2209 tcp_set_state(sk, TCP_CLOSE);
2210
2211 /* ABORT function of RFC793 */
2212 if (old_state == TCP_LISTEN) {
2213 inet_csk_listen_stop(sk);
2214 } else if (unlikely(tp->repair)) {
2215 sk->sk_err = ECONNABORTED;
2216 } else if (tcp_need_reset(old_state) ||
2217 (tp->snd_nxt != tp->write_seq &&
2218 (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK))) {
2219 /* The last check adjusts for discrepancy of Linux wrt. RFC
2220 * states
2221 */
2222 tcp_send_active_reset(sk, gfp_any());
2223 sk->sk_err = ECONNRESET;
2224 } else if (old_state == TCP_SYN_SENT)
2225 sk->sk_err = ECONNRESET;
2226
2227 tcp_clear_xmit_timers(sk);
2228 __skb_queue_purge(&sk->sk_receive_queue);
2229 tcp_write_queue_purge(sk);
2230 __skb_queue_purge(&tp->out_of_order_queue);
2231
2232 inet->inet_dport = 0;
2233
2234 if (!(sk->sk_userlocks & SOCK_BINDADDR_LOCK))
2235 inet_reset_saddr(sk);
2236
2237 sk->sk_shutdown = 0;
2238 sock_reset_flag(sk, SOCK_DONE);
2239 tp->srtt_us = 0;
2240 tp->write_seq += tp->max_window + 2;
2241 if (tp->write_seq == 0)
2242 tp->write_seq = 1;
2243 icsk->icsk_backoff = 0;
2244 tp->snd_cwnd = 2;
2245 icsk->icsk_probes_out = 0;
2246 tp->packets_out = 0;
2247 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
2248 tp->snd_cwnd_cnt = 0;
2249 tp->window_clamp = 0;
2250 tcp_set_ca_state(sk, TCP_CA_Open);
2251 tcp_clear_retrans(tp);
2252 inet_csk_delack_init(sk);
2253 tcp_init_send_head(sk);
2254 memset(&tp->rx_opt, 0, sizeof(tp->rx_opt));
2255 __sk_dst_reset(sk);
2256
2257 WARN_ON(inet->inet_num && !icsk->icsk_bind_hash);
2258
2259 sk->sk_error_report(sk);
2260 return err;
2261 }
2262 EXPORT_SYMBOL(tcp_disconnect);
2263
2264 static inline bool tcp_can_repair_sock(const struct sock *sk)
2265 {
2266 return ns_capable(sock_net(sk)->user_ns, CAP_NET_ADMIN) &&
2267 ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_ESTABLISHED));
2268 }
2269
2270 static int tcp_repair_options_est(struct tcp_sock *tp,
2271 struct tcp_repair_opt __user *optbuf, unsigned int len)
2272 {
2273 struct tcp_repair_opt opt;
2274
2275 while (len >= sizeof(opt)) {
2276 if (copy_from_user(&opt, optbuf, sizeof(opt)))
2277 return -EFAULT;
2278
2279 optbuf++;
2280 len -= sizeof(opt);
2281
2282 switch (opt.opt_code) {
2283 case TCPOPT_MSS:
2284 tp->rx_opt.mss_clamp = opt.opt_val;
2285 break;
2286 case TCPOPT_WINDOW:
2287 {
2288 u16 snd_wscale = opt.opt_val & 0xFFFF;
2289 u16 rcv_wscale = opt.opt_val >> 16;
2290
2291 if (snd_wscale > 14 || rcv_wscale > 14)
2292 return -EFBIG;
2293
2294 tp->rx_opt.snd_wscale = snd_wscale;
2295 tp->rx_opt.rcv_wscale = rcv_wscale;
2296 tp->rx_opt.wscale_ok = 1;
2297 }
2298 break;
2299 case TCPOPT_SACK_PERM:
2300 if (opt.opt_val != 0)
2301 return -EINVAL;
2302
2303 tp->rx_opt.sack_ok |= TCP_SACK_SEEN;
2304 if (sysctl_tcp_fack)
2305 tcp_enable_fack(tp);
2306 break;
2307 case TCPOPT_TIMESTAMP:
2308 if (opt.opt_val != 0)
2309 return -EINVAL;
2310
2311 tp->rx_opt.tstamp_ok = 1;
2312 break;
2313 }
2314 }
2315
2316 return 0;
2317 }
2318
2319 /*
2320 * Socket option code for TCP.
2321 */
2322 static int do_tcp_setsockopt(struct sock *sk, int level,
2323 int optname, char __user *optval, unsigned int optlen)
2324 {
2325 struct tcp_sock *tp = tcp_sk(sk);
2326 struct inet_connection_sock *icsk = inet_csk(sk);
2327 int val;
2328 int err = 0;
2329
2330 /* These are data/string values, all the others are ints */
2331 switch (optname) {
2332 case TCP_CONGESTION: {
2333 char name[TCP_CA_NAME_MAX];
2334
2335 if (optlen < 1)
2336 return -EINVAL;
2337
2338 val = strncpy_from_user(name, optval,
2339 min_t(long, TCP_CA_NAME_MAX-1, optlen));
2340 if (val < 0)
2341 return -EFAULT;
2342 name[val] = 0;
2343
2344 lock_sock(sk);
2345 err = tcp_set_congestion_control(sk, name);
2346 release_sock(sk);
2347 return err;
2348 }
2349 default:
2350 /* fallthru */
2351 break;
2352 }
2353
2354 if (optlen < sizeof(int))
2355 return -EINVAL;
2356
2357 if (get_user(val, (int __user *)optval))
2358 return -EFAULT;
2359
2360 lock_sock(sk);
2361
2362 switch (optname) {
2363 case TCP_MAXSEG:
2364 /* Values greater than interface MTU won't take effect. However
2365 * at the point when this call is done we typically don't yet
2366 * know which interface is going to be used */
2367 if (val < TCP_MIN_MSS || val > MAX_TCP_WINDOW) {
2368 err = -EINVAL;
2369 break;
2370 }
2371 tp->rx_opt.user_mss = val;
2372 break;
2373
2374 case TCP_NODELAY:
2375 if (val) {
2376 /* TCP_NODELAY is weaker than TCP_CORK, so that
2377 * this option on corked socket is remembered, but
2378 * it is not activated until cork is cleared.
2379 *
2380 * However, when TCP_NODELAY is set we make
2381 * an explicit push, which overrides even TCP_CORK
2382 * for currently queued segments.
2383 */
2384 tp->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH;
2385 tcp_push_pending_frames(sk);
2386 } else {
2387 tp->nonagle &= ~TCP_NAGLE_OFF;
2388 }
2389 break;
2390
2391 case TCP_THIN_LINEAR_TIMEOUTS:
2392 if (val < 0 || val > 1)
2393 err = -EINVAL;
2394 else
2395 tp->thin_lto = val;
2396 break;
2397
2398 case TCP_THIN_DUPACK:
2399 if (val < 0 || val > 1)
2400 err = -EINVAL;
2401 else {
2402 tp->thin_dupack = val;
2403 if (tp->thin_dupack)
2404 tcp_disable_early_retrans(tp);
2405 }
2406 break;
2407
2408 case TCP_REPAIR:
2409 if (!tcp_can_repair_sock(sk))
2410 err = -EPERM;
2411 else if (val == 1) {
2412 tp->repair = 1;
2413 sk->sk_reuse = SK_FORCE_REUSE;
2414 tp->repair_queue = TCP_NO_QUEUE;
2415 } else if (val == 0) {
2416 tp->repair = 0;
2417 sk->sk_reuse = SK_NO_REUSE;
2418 tcp_send_window_probe(sk);
2419 } else
2420 err = -EINVAL;
2421
2422 break;
2423
2424 case TCP_REPAIR_QUEUE:
2425 if (!tp->repair)
2426 err = -EPERM;
2427 else if (val < TCP_QUEUES_NR)
2428 tp->repair_queue = val;
2429 else
2430 err = -EINVAL;
2431 break;
2432
2433 case TCP_QUEUE_SEQ:
2434 if (sk->sk_state != TCP_CLOSE)
2435 err = -EPERM;
2436 else if (tp->repair_queue == TCP_SEND_QUEUE)
2437 tp->write_seq = val;
2438 else if (tp->repair_queue == TCP_RECV_QUEUE)
2439 tp->rcv_nxt = val;
2440 else
2441 err = -EINVAL;
2442 break;
2443
2444 case TCP_REPAIR_OPTIONS:
2445 if (!tp->repair)
2446 err = -EINVAL;
2447 else if (sk->sk_state == TCP_ESTABLISHED)
2448 err = tcp_repair_options_est(tp,
2449 (struct tcp_repair_opt __user *)optval,
2450 optlen);
2451 else
2452 err = -EPERM;
2453 break;
2454
2455 case TCP_CORK:
2456 /* When set indicates to always queue non-full frames.
2457 * Later the user clears this option and we transmit
2458 * any pending partial frames in the queue. This is
2459 * meant to be used alongside sendfile() to get properly
2460 * filled frames when the user (for example) must write
2461 * out headers with a write() call first and then use
2462 * sendfile to send out the data parts.
2463 *
2464 * TCP_CORK can be set together with TCP_NODELAY and it is
2465 * stronger than TCP_NODELAY.
2466 */
2467 if (val) {
2468 tp->nonagle |= TCP_NAGLE_CORK;
2469 } else {
2470 tp->nonagle &= ~TCP_NAGLE_CORK;
2471 if (tp->nonagle&TCP_NAGLE_OFF)
2472 tp->nonagle |= TCP_NAGLE_PUSH;
2473 tcp_push_pending_frames(sk);
2474 }
2475 break;
2476
2477 case TCP_KEEPIDLE:
2478 if (val < 1 || val > MAX_TCP_KEEPIDLE)
2479 err = -EINVAL;
2480 else {
2481 tp->keepalive_time = val * HZ;
2482 if (sock_flag(sk, SOCK_KEEPOPEN) &&
2483 !((1 << sk->sk_state) &
2484 (TCPF_CLOSE | TCPF_LISTEN))) {
2485 u32 elapsed = keepalive_time_elapsed(tp);
2486 if (tp->keepalive_time > elapsed)
2487 elapsed = tp->keepalive_time - elapsed;
2488 else
2489 elapsed = 0;
2490 inet_csk_reset_keepalive_timer(sk, elapsed);
2491 }
2492 }
2493 break;
2494 case TCP_KEEPINTVL:
2495 if (val < 1 || val > MAX_TCP_KEEPINTVL)
2496 err = -EINVAL;
2497 else
2498 tp->keepalive_intvl = val * HZ;
2499 break;
2500 case TCP_KEEPCNT:
2501 if (val < 1 || val > MAX_TCP_KEEPCNT)
2502 err = -EINVAL;
2503 else
2504 tp->keepalive_probes = val;
2505 break;
2506 case TCP_SYNCNT:
2507 if (val < 1 || val > MAX_TCP_SYNCNT)
2508 err = -EINVAL;
2509 else
2510 icsk->icsk_syn_retries = val;
2511 break;
2512
2513 case TCP_SAVE_SYN:
2514 if (val < 0 || val > 1)
2515 err = -EINVAL;
2516 else
2517 tp->save_syn = val;
2518 break;
2519
2520 case TCP_LINGER2:
2521 if (val < 0)
2522 tp->linger2 = -1;
2523 else if (val > sysctl_tcp_fin_timeout / HZ)
2524 tp->linger2 = 0;
2525 else
2526 tp->linger2 = val * HZ;
2527 break;
2528
2529 case TCP_DEFER_ACCEPT:
2530 /* Translate value in seconds to number of retransmits */
2531 icsk->icsk_accept_queue.rskq_defer_accept =
2532 secs_to_retrans(val, TCP_TIMEOUT_INIT / HZ,
2533 TCP_RTO_MAX / HZ);
2534 break;
2535
2536 case TCP_WINDOW_CLAMP:
2537 if (!val) {
2538 if (sk->sk_state != TCP_CLOSE) {
2539 err = -EINVAL;
2540 break;
2541 }
2542 tp->window_clamp = 0;
2543 } else
2544 tp->window_clamp = val < SOCK_MIN_RCVBUF / 2 ?
2545 SOCK_MIN_RCVBUF / 2 : val;
2546 break;
2547
2548 case TCP_QUICKACK:
2549 if (!val) {
2550 icsk->icsk_ack.pingpong = 1;
2551 } else {
2552 icsk->icsk_ack.pingpong = 0;
2553 if ((1 << sk->sk_state) &
2554 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) &&
2555 inet_csk_ack_scheduled(sk)) {
2556 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
2557 tcp_cleanup_rbuf(sk, 1);
2558 if (!(val & 1))
2559 icsk->icsk_ack.pingpong = 1;
2560 }
2561 }
2562 break;
2563
2564 #ifdef CONFIG_TCP_MD5SIG
2565 case TCP_MD5SIG:
2566 /* Read the IP->Key mappings from userspace */
2567 err = tp->af_specific->md5_parse(sk, optval, optlen);
2568 break;
2569 #endif
2570 case TCP_USER_TIMEOUT:
2571 /* Cap the max time in ms TCP will retry or probe the window
2572 * before giving up and aborting (ETIMEDOUT) a connection.
2573 */
2574 if (val < 0)
2575 err = -EINVAL;
2576 else
2577 icsk->icsk_user_timeout = msecs_to_jiffies(val);
2578 break;
2579
2580 case TCP_FASTOPEN:
2581 if (val >= 0 && ((1 << sk->sk_state) & (TCPF_CLOSE |
2582 TCPF_LISTEN))) {
2583 tcp_fastopen_init_key_once(true);
2584
2585 fastopen_queue_tune(sk, val);
2586 } else {
2587 err = -EINVAL;
2588 }
2589 break;
2590 case TCP_TIMESTAMP:
2591 if (!tp->repair)
2592 err = -EPERM;
2593 else
2594 tp->tsoffset = val - tcp_time_stamp;
2595 break;
2596 case TCP_NOTSENT_LOWAT:
2597 tp->notsent_lowat = val;
2598 sk->sk_write_space(sk);
2599 break;
2600 default:
2601 err = -ENOPROTOOPT;
2602 break;
2603 }
2604
2605 release_sock(sk);
2606 return err;
2607 }
2608
2609 int tcp_setsockopt(struct sock *sk, int level, int optname, char __user *optval,
2610 unsigned int optlen)
2611 {
2612 const struct inet_connection_sock *icsk = inet_csk(sk);
2613
2614 if (level != SOL_TCP)
2615 return icsk->icsk_af_ops->setsockopt(sk, level, optname,
2616 optval, optlen);
2617 return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2618 }
2619 EXPORT_SYMBOL(tcp_setsockopt);
2620
2621 #ifdef CONFIG_COMPAT
2622 int compat_tcp_setsockopt(struct sock *sk, int level, int optname,
2623 char __user *optval, unsigned int optlen)
2624 {
2625 if (level != SOL_TCP)
2626 return inet_csk_compat_setsockopt(sk, level, optname,
2627 optval, optlen);
2628 return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2629 }
2630 EXPORT_SYMBOL(compat_tcp_setsockopt);
2631 #endif
2632
2633 /* Return information about state of tcp endpoint in API format. */
2634 void tcp_get_info(struct sock *sk, struct tcp_info *info)
2635 {
2636 const struct tcp_sock *tp = tcp_sk(sk); /* iff sk_type == SOCK_STREAM */
2637 const struct inet_connection_sock *icsk = inet_csk(sk);
2638 u32 now = tcp_time_stamp;
2639 unsigned int start;
2640 u32 rate;
2641
2642 memset(info, 0, sizeof(*info));
2643 if (sk->sk_type != SOCK_STREAM)
2644 return;
2645
2646 info->tcpi_state = sk_state_load(sk);
2647
2648 info->tcpi_ca_state = icsk->icsk_ca_state;
2649 info->tcpi_retransmits = icsk->icsk_retransmits;
2650 info->tcpi_probes = icsk->icsk_probes_out;
2651 info->tcpi_backoff = icsk->icsk_backoff;
2652
2653 if (tp->rx_opt.tstamp_ok)
2654 info->tcpi_options |= TCPI_OPT_TIMESTAMPS;
2655 if (tcp_is_sack(tp))
2656 info->tcpi_options |= TCPI_OPT_SACK;
2657 if (tp->rx_opt.wscale_ok) {
2658 info->tcpi_options |= TCPI_OPT_WSCALE;
2659 info->tcpi_snd_wscale = tp->rx_opt.snd_wscale;
2660 info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale;
2661 }
2662
2663 if (tp->ecn_flags & TCP_ECN_OK)
2664 info->tcpi_options |= TCPI_OPT_ECN;
2665 if (tp->ecn_flags & TCP_ECN_SEEN)
2666 info->tcpi_options |= TCPI_OPT_ECN_SEEN;
2667 if (tp->syn_data_acked)
2668 info->tcpi_options |= TCPI_OPT_SYN_DATA;
2669
2670 info->tcpi_rto = jiffies_to_usecs(icsk->icsk_rto);
2671 info->tcpi_ato = jiffies_to_usecs(icsk->icsk_ack.ato);
2672 info->tcpi_snd_mss = tp->mss_cache;
2673 info->tcpi_rcv_mss = icsk->icsk_ack.rcv_mss;
2674
2675 if (info->tcpi_state == TCP_LISTEN) {
2676 info->tcpi_unacked = sk->sk_ack_backlog;
2677 info->tcpi_sacked = sk->sk_max_ack_backlog;
2678 } else {
2679 info->tcpi_unacked = tp->packets_out;
2680 info->tcpi_sacked = tp->sacked_out;
2681 }
2682 info->tcpi_lost = tp->lost_out;
2683 info->tcpi_retrans = tp->retrans_out;
2684 info->tcpi_fackets = tp->fackets_out;
2685
2686 info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime);
2687 info->tcpi_last_data_recv = jiffies_to_msecs(now - icsk->icsk_ack.lrcvtime);
2688 info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp);
2689
2690 info->tcpi_pmtu = icsk->icsk_pmtu_cookie;
2691 info->tcpi_rcv_ssthresh = tp->rcv_ssthresh;
2692 info->tcpi_rtt = tp->srtt_us >> 3;
2693 info->tcpi_rttvar = tp->mdev_us >> 2;
2694 info->tcpi_snd_ssthresh = tp->snd_ssthresh;
2695 info->tcpi_snd_cwnd = tp->snd_cwnd;
2696 info->tcpi_advmss = tp->advmss;
2697 info->tcpi_reordering = tp->reordering;
2698
2699 info->tcpi_rcv_rtt = jiffies_to_usecs(tp->rcv_rtt_est.rtt)>>3;
2700 info->tcpi_rcv_space = tp->rcvq_space.space;
2701
2702 info->tcpi_total_retrans = tp->total_retrans;
2703
2704 rate = READ_ONCE(sk->sk_pacing_rate);
2705 info->tcpi_pacing_rate = rate != ~0U ? rate : ~0ULL;
2706
2707 rate = READ_ONCE(sk->sk_max_pacing_rate);
2708 info->tcpi_max_pacing_rate = rate != ~0U ? rate : ~0ULL;
2709
2710 do {
2711 start = u64_stats_fetch_begin_irq(&tp->syncp);
2712 info->tcpi_bytes_acked = tp->bytes_acked;
2713 info->tcpi_bytes_received = tp->bytes_received;
2714 } while (u64_stats_fetch_retry_irq(&tp->syncp, start));
2715 info->tcpi_segs_out = tp->segs_out;
2716 info->tcpi_segs_in = tp->segs_in;
2717 }
2718 EXPORT_SYMBOL_GPL(tcp_get_info);
2719
2720 static int do_tcp_getsockopt(struct sock *sk, int level,
2721 int optname, char __user *optval, int __user *optlen)
2722 {
2723 struct inet_connection_sock *icsk = inet_csk(sk);
2724 struct tcp_sock *tp = tcp_sk(sk);
2725 int val, len;
2726
2727 if (get_user(len, optlen))
2728 return -EFAULT;
2729
2730 len = min_t(unsigned int, len, sizeof(int));
2731
2732 if (len < 0)
2733 return -EINVAL;
2734
2735 switch (optname) {
2736 case TCP_MAXSEG:
2737 val = tp->mss_cache;
2738 if (!val && ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
2739 val = tp->rx_opt.user_mss;
2740 if (tp->repair)
2741 val = tp->rx_opt.mss_clamp;
2742 break;
2743 case TCP_NODELAY:
2744 val = !!(tp->nonagle&TCP_NAGLE_OFF);
2745 break;
2746 case TCP_CORK:
2747 val = !!(tp->nonagle&TCP_NAGLE_CORK);
2748 break;
2749 case TCP_KEEPIDLE:
2750 val = keepalive_time_when(tp) / HZ;
2751 break;
2752 case TCP_KEEPINTVL:
2753 val = keepalive_intvl_when(tp) / HZ;
2754 break;
2755 case TCP_KEEPCNT:
2756 val = keepalive_probes(tp);
2757 break;
2758 case TCP_SYNCNT:
2759 val = icsk->icsk_syn_retries ? : sysctl_tcp_syn_retries;
2760 break;
2761 case TCP_LINGER2:
2762 val = tp->linger2;
2763 if (val >= 0)
2764 val = (val ? : sysctl_tcp_fin_timeout) / HZ;
2765 break;
2766 case TCP_DEFER_ACCEPT:
2767 val = retrans_to_secs(icsk->icsk_accept_queue.rskq_defer_accept,
2768 TCP_TIMEOUT_INIT / HZ, TCP_RTO_MAX / HZ);
2769 break;
2770 case TCP_WINDOW_CLAMP:
2771 val = tp->window_clamp;
2772 break;
2773 case TCP_INFO: {
2774 struct tcp_info info;
2775
2776 if (get_user(len, optlen))
2777 return -EFAULT;
2778
2779 tcp_get_info(sk, &info);
2780
2781 len = min_t(unsigned int, len, sizeof(info));
2782 if (put_user(len, optlen))
2783 return -EFAULT;
2784 if (copy_to_user(optval, &info, len))
2785 return -EFAULT;
2786 return 0;
2787 }
2788 case TCP_CC_INFO: {
2789 const struct tcp_congestion_ops *ca_ops;
2790 union tcp_cc_info info;
2791 size_t sz = 0;
2792 int attr;
2793
2794 if (get_user(len, optlen))
2795 return -EFAULT;
2796
2797 ca_ops = icsk->icsk_ca_ops;
2798 if (ca_ops && ca_ops->get_info)
2799 sz = ca_ops->get_info(sk, ~0U, &attr, &info);
2800
2801 len = min_t(unsigned int, len, sz);
2802 if (put_user(len, optlen))
2803 return -EFAULT;
2804 if (copy_to_user(optval, &info, len))
2805 return -EFAULT;
2806 return 0;
2807 }
2808 case TCP_QUICKACK:
2809 val = !icsk->icsk_ack.pingpong;
2810 break;
2811
2812 case TCP_CONGESTION:
2813 if (get_user(len, optlen))
2814 return -EFAULT;
2815 len = min_t(unsigned int, len, TCP_CA_NAME_MAX);
2816 if (put_user(len, optlen))
2817 return -EFAULT;
2818 if (copy_to_user(optval, icsk->icsk_ca_ops->name, len))
2819 return -EFAULT;
2820 return 0;
2821
2822 case TCP_THIN_LINEAR_TIMEOUTS:
2823 val = tp->thin_lto;
2824 break;
2825 case TCP_THIN_DUPACK:
2826 val = tp->thin_dupack;
2827 break;
2828
2829 case TCP_REPAIR:
2830 val = tp->repair;
2831 break;
2832
2833 case TCP_REPAIR_QUEUE:
2834 if (tp->repair)
2835 val = tp->repair_queue;
2836 else
2837 return -EINVAL;
2838 break;
2839
2840 case TCP_QUEUE_SEQ:
2841 if (tp->repair_queue == TCP_SEND_QUEUE)
2842 val = tp->write_seq;
2843 else if (tp->repair_queue == TCP_RECV_QUEUE)
2844 val = tp->rcv_nxt;
2845 else
2846 return -EINVAL;
2847 break;
2848
2849 case TCP_USER_TIMEOUT:
2850 val = jiffies_to_msecs(icsk->icsk_user_timeout);
2851 break;
2852
2853 case TCP_FASTOPEN:
2854 val = icsk->icsk_accept_queue.fastopenq.max_qlen;
2855 break;
2856
2857 case TCP_TIMESTAMP:
2858 val = tcp_time_stamp + tp->tsoffset;
2859 break;
2860 case TCP_NOTSENT_LOWAT:
2861 val = tp->notsent_lowat;
2862 break;
2863 case TCP_SAVE_SYN:
2864 val = tp->save_syn;
2865 break;
2866 case TCP_SAVED_SYN: {
2867 if (get_user(len, optlen))
2868 return -EFAULT;
2869
2870 lock_sock(sk);
2871 if (tp->saved_syn) {
2872 if (len < tp->saved_syn[0]) {
2873 if (put_user(tp->saved_syn[0], optlen)) {
2874 release_sock(sk);
2875 return -EFAULT;
2876 }
2877 release_sock(sk);
2878 return -EINVAL;
2879 }
2880 len = tp->saved_syn[0];
2881 if (put_user(len, optlen)) {
2882 release_sock(sk);
2883 return -EFAULT;
2884 }
2885 if (copy_to_user(optval, tp->saved_syn + 1, len)) {
2886 release_sock(sk);
2887 return -EFAULT;
2888 }
2889 tcp_saved_syn_free(tp);
2890 release_sock(sk);
2891 } else {
2892 release_sock(sk);
2893 len = 0;
2894 if (put_user(len, optlen))
2895 return -EFAULT;
2896 }
2897 return 0;
2898 }
2899 default:
2900 return -ENOPROTOOPT;
2901 }
2902
2903 if (put_user(len, optlen))
2904 return -EFAULT;
2905 if (copy_to_user(optval, &val, len))
2906 return -EFAULT;
2907 return 0;
2908 }
2909
2910 int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval,
2911 int __user *optlen)
2912 {
2913 struct inet_connection_sock *icsk = inet_csk(sk);
2914
2915 if (level != SOL_TCP)
2916 return icsk->icsk_af_ops->getsockopt(sk, level, optname,
2917 optval, optlen);
2918 return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2919 }
2920 EXPORT_SYMBOL(tcp_getsockopt);
2921
2922 #ifdef CONFIG_COMPAT
2923 int compat_tcp_getsockopt(struct sock *sk, int level, int optname,
2924 char __user *optval, int __user *optlen)
2925 {
2926 if (level != SOL_TCP)
2927 return inet_csk_compat_getsockopt(sk, level, optname,
2928 optval, optlen);
2929 return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2930 }
2931 EXPORT_SYMBOL(compat_tcp_getsockopt);
2932 #endif
2933
2934 #ifdef CONFIG_TCP_MD5SIG
2935 static DEFINE_PER_CPU(struct tcp_md5sig_pool, tcp_md5sig_pool);
2936 static DEFINE_MUTEX(tcp_md5sig_mutex);
2937 static bool tcp_md5sig_pool_populated = false;
2938
2939 static void __tcp_alloc_md5sig_pool(void)
2940 {
2941 int cpu;
2942
2943 for_each_possible_cpu(cpu) {
2944 if (!per_cpu(tcp_md5sig_pool, cpu).md5_desc.tfm) {
2945 struct crypto_hash *hash;
2946
2947 hash = crypto_alloc_hash("md5", 0, CRYPTO_ALG_ASYNC);
2948 if (IS_ERR_OR_NULL(hash))
2949 return;
2950 per_cpu(tcp_md5sig_pool, cpu).md5_desc.tfm = hash;
2951 }
2952 }
2953 /* before setting tcp_md5sig_pool_populated, we must commit all writes
2954 * to memory. See smp_rmb() in tcp_get_md5sig_pool()
2955 */
2956 smp_wmb();
2957 tcp_md5sig_pool_populated = true;
2958 }
2959
2960 bool tcp_alloc_md5sig_pool(void)
2961 {
2962 if (unlikely(!tcp_md5sig_pool_populated)) {
2963 mutex_lock(&tcp_md5sig_mutex);
2964
2965 if (!tcp_md5sig_pool_populated)
2966 __tcp_alloc_md5sig_pool();
2967
2968 mutex_unlock(&tcp_md5sig_mutex);
2969 }
2970 return tcp_md5sig_pool_populated;
2971 }
2972 EXPORT_SYMBOL(tcp_alloc_md5sig_pool);
2973
2974
2975 /**
2976 * tcp_get_md5sig_pool - get md5sig_pool for this user
2977 *
2978 * We use percpu structure, so if we succeed, we exit with preemption
2979 * and BH disabled, to make sure another thread or softirq handling
2980 * wont try to get same context.
2981 */
2982 struct tcp_md5sig_pool *tcp_get_md5sig_pool(void)
2983 {
2984 local_bh_disable();
2985
2986 if (tcp_md5sig_pool_populated) {
2987 /* coupled with smp_wmb() in __tcp_alloc_md5sig_pool() */
2988 smp_rmb();
2989 return this_cpu_ptr(&tcp_md5sig_pool);
2990 }
2991 local_bh_enable();
2992 return NULL;
2993 }
2994 EXPORT_SYMBOL(tcp_get_md5sig_pool);
2995
2996 int tcp_md5_hash_header(struct tcp_md5sig_pool *hp,
2997 const struct tcphdr *th)
2998 {
2999 struct scatterlist sg;
3000 struct tcphdr hdr;
3001 int err;
3002
3003 /* We are not allowed to change tcphdr, make a local copy */
3004 memcpy(&hdr, th, sizeof(hdr));
3005 hdr.check = 0;
3006
3007 /* options aren't included in the hash */
3008 sg_init_one(&sg, &hdr, sizeof(hdr));
3009 err = crypto_hash_update(&hp->md5_desc, &sg, sizeof(hdr));
3010 return err;
3011 }
3012 EXPORT_SYMBOL(tcp_md5_hash_header);
3013
3014 int tcp_md5_hash_skb_data(struct tcp_md5sig_pool *hp,
3015 const struct sk_buff *skb, unsigned int header_len)
3016 {
3017 struct scatterlist sg;
3018 const struct tcphdr *tp = tcp_hdr(skb);
3019 struct hash_desc *desc = &hp->md5_desc;
3020 unsigned int i;
3021 const unsigned int head_data_len = skb_headlen(skb) > header_len ?
3022 skb_headlen(skb) - header_len : 0;
3023 const struct skb_shared_info *shi = skb_shinfo(skb);
3024 struct sk_buff *frag_iter;
3025
3026 sg_init_table(&sg, 1);
3027
3028 sg_set_buf(&sg, ((u8 *) tp) + header_len, head_data_len);
3029 if (crypto_hash_update(desc, &sg, head_data_len))
3030 return 1;
3031
3032 for (i = 0; i < shi->nr_frags; ++i) {
3033 const struct skb_frag_struct *f = &shi->frags[i];
3034 unsigned int offset = f->page_offset;
3035 struct page *page = skb_frag_page(f) + (offset >> PAGE_SHIFT);
3036
3037 sg_set_page(&sg, page, skb_frag_size(f),
3038 offset_in_page(offset));
3039 if (crypto_hash_update(desc, &sg, skb_frag_size(f)))
3040 return 1;
3041 }
3042
3043 skb_walk_frags(skb, frag_iter)
3044 if (tcp_md5_hash_skb_data(hp, frag_iter, 0))
3045 return 1;
3046
3047 return 0;
3048 }
3049 EXPORT_SYMBOL(tcp_md5_hash_skb_data);
3050
3051 int tcp_md5_hash_key(struct tcp_md5sig_pool *hp, const struct tcp_md5sig_key *key)
3052 {
3053 struct scatterlist sg;
3054
3055 sg_init_one(&sg, key->key, key->keylen);
3056 return crypto_hash_update(&hp->md5_desc, &sg, key->keylen);
3057 }
3058 EXPORT_SYMBOL(tcp_md5_hash_key);
3059
3060 #endif
3061
3062 void tcp_done(struct sock *sk)
3063 {
3064 struct request_sock *req = tcp_sk(sk)->fastopen_rsk;
3065
3066 if (sk->sk_state == TCP_SYN_SENT || sk->sk_state == TCP_SYN_RECV)
3067 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_ATTEMPTFAILS);
3068
3069 tcp_set_state(sk, TCP_CLOSE);
3070 tcp_clear_xmit_timers(sk);
3071 if (req)
3072 reqsk_fastopen_remove(sk, req, false);
3073
3074 sk->sk_shutdown = SHUTDOWN_MASK;
3075
3076 if (!sock_flag(sk, SOCK_DEAD))
3077 sk->sk_state_change(sk);
3078 else
3079 inet_csk_destroy_sock(sk);
3080 }
3081 EXPORT_SYMBOL_GPL(tcp_done);
3082
3083 extern struct tcp_congestion_ops tcp_reno;
3084
3085 static __initdata unsigned long thash_entries;
3086 static int __init set_thash_entries(char *str)
3087 {
3088 ssize_t ret;
3089
3090 if (!str)
3091 return 0;
3092
3093 ret = kstrtoul(str, 0, &thash_entries);
3094 if (ret)
3095 return 0;
3096
3097 return 1;
3098 }
3099 __setup("thash_entries=", set_thash_entries);
3100
3101 static void __init tcp_init_mem(void)
3102 {
3103 unsigned long limit = nr_free_buffer_pages() / 16;
3104
3105 limit = max(limit, 128UL);
3106 sysctl_tcp_mem[0] = limit / 4 * 3; /* 4.68 % */
3107 sysctl_tcp_mem[1] = limit; /* 6.25 % */
3108 sysctl_tcp_mem[2] = sysctl_tcp_mem[0] * 2; /* 9.37 % */
3109 }
3110
3111 void __init tcp_init(void)
3112 {
3113 unsigned long limit;
3114 int max_rshare, max_wshare, cnt;
3115 unsigned int i;
3116
3117 sock_skb_cb_check_size(sizeof(struct tcp_skb_cb));
3118
3119 percpu_counter_init(&tcp_sockets_allocated, 0, GFP_KERNEL);
3120 percpu_counter_init(&tcp_orphan_count, 0, GFP_KERNEL);
3121 tcp_hashinfo.bind_bucket_cachep =
3122 kmem_cache_create("tcp_bind_bucket",
3123 sizeof(struct inet_bind_bucket), 0,
3124 SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL);
3125
3126 /* Size and allocate the main established and bind bucket
3127 * hash tables.
3128 *
3129 * The methodology is similar to that of the buffer cache.
3130 */
3131 tcp_hashinfo.ehash =
3132 alloc_large_system_hash("TCP established",
3133 sizeof(struct inet_ehash_bucket),
3134 thash_entries,
3135 17, /* one slot per 128 KB of memory */
3136 0,
3137 NULL,
3138 &tcp_hashinfo.ehash_mask,
3139 0,
3140 thash_entries ? 0 : 512 * 1024);
3141 for (i = 0; i <= tcp_hashinfo.ehash_mask; i++)
3142 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].chain, i);
3143
3144 if (inet_ehash_locks_alloc(&tcp_hashinfo))
3145 panic("TCP: failed to alloc ehash_locks");
3146 tcp_hashinfo.bhash =
3147 alloc_large_system_hash("TCP bind",
3148 sizeof(struct inet_bind_hashbucket),
3149 tcp_hashinfo.ehash_mask + 1,
3150 17, /* one slot per 128 KB of memory */
3151 0,
3152 &tcp_hashinfo.bhash_size,
3153 NULL,
3154 0,
3155 64 * 1024);
3156 tcp_hashinfo.bhash_size = 1U << tcp_hashinfo.bhash_size;
3157 for (i = 0; i < tcp_hashinfo.bhash_size; i++) {
3158 spin_lock_init(&tcp_hashinfo.bhash[i].lock);
3159 INIT_HLIST_HEAD(&tcp_hashinfo.bhash[i].chain);
3160 }
3161
3162
3163 cnt = tcp_hashinfo.ehash_mask + 1;
3164
3165 tcp_death_row.sysctl_max_tw_buckets = cnt / 2;
3166 sysctl_tcp_max_orphans = cnt / 2;
3167 sysctl_max_syn_backlog = max(128, cnt / 256);
3168
3169 tcp_init_mem();
3170 /* Set per-socket limits to no more than 1/128 the pressure threshold */
3171 limit = nr_free_buffer_pages() << (PAGE_SHIFT - 7);
3172 max_wshare = min(4UL*1024*1024, limit);
3173 max_rshare = min(6UL*1024*1024, limit);
3174
3175 sysctl_tcp_wmem[0] = SK_MEM_QUANTUM;
3176 sysctl_tcp_wmem[1] = 16*1024;
3177 sysctl_tcp_wmem[2] = max(64*1024, max_wshare);
3178
3179 sysctl_tcp_rmem[0] = SK_MEM_QUANTUM;
3180 sysctl_tcp_rmem[1] = 87380;
3181 sysctl_tcp_rmem[2] = max(87380, max_rshare);
3182
3183 pr_info("Hash tables configured (established %u bind %u)\n",
3184 tcp_hashinfo.ehash_mask + 1, tcp_hashinfo.bhash_size);
3185
3186 tcp_metrics_init();
3187 BUG_ON(tcp_register_congestion_control(&tcp_reno) != 0);
3188 tcp_tasklet_init();
3189 }
This page took 0.099134 seconds and 5 git commands to generate.