Loading...
1/*
2 * INET An implementation of the TCP/IP protocol suite for the LINUX
3 * operating system. INET is implemented using the BSD Socket
4 * interface as the means of communication with the user level.
5 *
6 * Implementation of the Transmission Control Protocol(TCP).
7 *
8 * Authors: Ross Biro
9 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
10 * Mark Evans, <evansmp@uhura.aston.ac.uk>
11 * Corey Minyard <wf-rch!minyard@relay.EU.net>
12 * Florian La Roche, <flla@stud.uni-sb.de>
13 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
14 * Linus Torvalds, <torvalds@cs.helsinki.fi>
15 * Alan Cox, <gw4pts@gw4pts.ampr.org>
16 * Matthew Dillon, <dillon@apollo.west.oic.com>
17 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
18 * Jorge Cwik, <jorge@laser.satlink.net>
19 *
20 * Fixes:
21 * Alan Cox : Numerous verify_area() calls
22 * Alan Cox : Set the ACK bit on a reset
23 * Alan Cox : Stopped it crashing if it closed while
24 * sk->inuse=1 and was trying to connect
25 * (tcp_err()).
26 * Alan Cox : All icmp error handling was broken
27 * pointers passed where wrong and the
28 * socket was looked up backwards. Nobody
29 * tested any icmp error code obviously.
30 * Alan Cox : tcp_err() now handled properly. It
31 * wakes people on errors. poll
32 * behaves and the icmp error race
33 * has gone by moving it into sock.c
34 * Alan Cox : tcp_send_reset() fixed to work for
35 * everything not just packets for
36 * unknown sockets.
37 * Alan Cox : tcp option processing.
38 * Alan Cox : Reset tweaked (still not 100%) [Had
39 * syn rule wrong]
40 * Herp Rosmanith : More reset fixes
41 * Alan Cox : No longer acks invalid rst frames.
42 * Acking any kind of RST is right out.
43 * Alan Cox : Sets an ignore me flag on an rst
44 * receive otherwise odd bits of prattle
45 * escape still
46 * Alan Cox : Fixed another acking RST frame bug.
47 * Should stop LAN workplace lockups.
48 * Alan Cox : Some tidyups using the new skb list
49 * facilities
50 * Alan Cox : sk->keepopen now seems to work
51 * Alan Cox : Pulls options out correctly on accepts
52 * Alan Cox : Fixed assorted sk->rqueue->next errors
53 * Alan Cox : PSH doesn't end a TCP read. Switched a
54 * bit to skb ops.
55 * Alan Cox : Tidied tcp_data to avoid a potential
56 * nasty.
57 * Alan Cox : Added some better commenting, as the
58 * tcp is hard to follow
59 * Alan Cox : Removed incorrect check for 20 * psh
60 * Michael O'Reilly : ack < copied bug fix.
61 * Johannes Stille : Misc tcp fixes (not all in yet).
62 * Alan Cox : FIN with no memory -> CRASH
63 * Alan Cox : Added socket option proto entries.
64 * Also added awareness of them to accept.
65 * Alan Cox : Added TCP options (SOL_TCP)
66 * Alan Cox : Switched wakeup calls to callbacks,
67 * so the kernel can layer network
68 * sockets.
69 * Alan Cox : Use ip_tos/ip_ttl settings.
70 * Alan Cox : Handle FIN (more) properly (we hope).
71 * Alan Cox : RST frames sent on unsynchronised
72 * state ack error.
73 * Alan Cox : Put in missing check for SYN bit.
74 * Alan Cox : Added tcp_select_window() aka NET2E
75 * window non shrink trick.
76 * Alan Cox : Added a couple of small NET2E timer
77 * fixes
78 * Charles Hedrick : TCP fixes
79 * Toomas Tamm : TCP window fixes
80 * Alan Cox : Small URG fix to rlogin ^C ack fight
81 * Charles Hedrick : Rewrote most of it to actually work
82 * Linus : Rewrote tcp_read() and URG handling
83 * completely
84 * Gerhard Koerting: Fixed some missing timer handling
85 * Matthew Dillon : Reworked TCP machine states as per RFC
86 * Gerhard Koerting: PC/TCP workarounds
87 * Adam Caldwell : Assorted timer/timing errors
88 * Matthew Dillon : Fixed another RST bug
89 * Alan Cox : Move to kernel side addressing changes.
90 * Alan Cox : Beginning work on TCP fastpathing
91 * (not yet usable)
92 * Arnt Gulbrandsen: Turbocharged tcp_check() routine.
93 * Alan Cox : TCP fast path debugging
94 * Alan Cox : Window clamping
95 * Michael Riepe : Bug in tcp_check()
96 * Matt Dillon : More TCP improvements and RST bug fixes
97 * Matt Dillon : Yet more small nasties remove from the
98 * TCP code (Be very nice to this man if
99 * tcp finally works 100%) 8)
100 * Alan Cox : BSD accept semantics.
101 * Alan Cox : Reset on closedown bug.
102 * Peter De Schrijver : ENOTCONN check missing in tcp_sendto().
103 * Michael Pall : Handle poll() after URG properly in
104 * all cases.
105 * Michael Pall : Undo the last fix in tcp_read_urg()
106 * (multi URG PUSH broke rlogin).
107 * Michael Pall : Fix the multi URG PUSH problem in
108 * tcp_readable(), poll() after URG
109 * works now.
110 * Michael Pall : recv(...,MSG_OOB) never blocks in the
111 * BSD api.
112 * Alan Cox : Changed the semantics of sk->socket to
113 * fix a race and a signal problem with
114 * accept() and async I/O.
115 * Alan Cox : Relaxed the rules on tcp_sendto().
116 * Yury Shevchuk : Really fixed accept() blocking problem.
117 * Craig I. Hagan : Allow for BSD compatible TIME_WAIT for
118 * clients/servers which listen in on
119 * fixed ports.
120 * Alan Cox : Cleaned the above up and shrank it to
121 * a sensible code size.
122 * Alan Cox : Self connect lockup fix.
123 * Alan Cox : No connect to multicast.
124 * Ross Biro : Close unaccepted children on master
125 * socket close.
126 * Alan Cox : Reset tracing code.
127 * Alan Cox : Spurious resets on shutdown.
128 * Alan Cox : Giant 15 minute/60 second timer error
129 * Alan Cox : Small whoops in polling before an
130 * accept.
131 * Alan Cox : Kept the state trace facility since
132 * it's handy for debugging.
133 * Alan Cox : More reset handler fixes.
134 * Alan Cox : Started rewriting the code based on
135 * the RFC's for other useful protocol
136 * references see: Comer, KA9Q NOS, and
137 * for a reference on the difference
138 * between specifications and how BSD
139 * works see the 4.4lite source.
140 * A.N.Kuznetsov : Don't time wait on completion of tidy
141 * close.
142 * Linus Torvalds : Fin/Shutdown & copied_seq changes.
143 * Linus Torvalds : Fixed BSD port reuse to work first syn
144 * Alan Cox : Reimplemented timers as per the RFC
145 * and using multiple timers for sanity.
146 * Alan Cox : Small bug fixes, and a lot of new
147 * comments.
148 * Alan Cox : Fixed dual reader crash by locking
149 * the buffers (much like datagram.c)
150 * Alan Cox : Fixed stuck sockets in probe. A probe
151 * now gets fed up of retrying without
152 * (even a no space) answer.
153 * Alan Cox : Extracted closing code better
154 * Alan Cox : Fixed the closing state machine to
155 * resemble the RFC.
156 * Alan Cox : More 'per spec' fixes.
157 * Jorge Cwik : Even faster checksumming.
158 * Alan Cox : tcp_data() doesn't ack illegal PSH
159 * only frames. At least one pc tcp stack
160 * generates them.
161 * Alan Cox : Cache last socket.
162 * Alan Cox : Per route irtt.
163 * Matt Day : poll()->select() match BSD precisely on error
164 * Alan Cox : New buffers
165 * Marc Tamsky : Various sk->prot->retransmits and
166 * sk->retransmits misupdating fixed.
167 * Fixed tcp_write_timeout: stuck close,
168 * and TCP syn retries gets used now.
169 * Mark Yarvis : In tcp_read_wakeup(), don't send an
170 * ack if state is TCP_CLOSED.
171 * Alan Cox : Look up device on a retransmit - routes may
172 * change. Doesn't yet cope with MSS shrink right
173 * but it's a start!
174 * Marc Tamsky : Closing in closing fixes.
175 * Mike Shaver : RFC1122 verifications.
176 * Alan Cox : rcv_saddr errors.
177 * Alan Cox : Block double connect().
178 * Alan Cox : Small hooks for enSKIP.
179 * Alexey Kuznetsov: Path MTU discovery.
180 * Alan Cox : Support soft errors.
181 * Alan Cox : Fix MTU discovery pathological case
182 * when the remote claims no mtu!
183 * Marc Tamsky : TCP_CLOSE fix.
184 * Colin (G3TNE) : Send a reset on syn ack replies in
185 * window but wrong (fixes NT lpd problems)
186 * Pedro Roque : Better TCP window handling, delayed ack.
187 * Joerg Reuter : No modification of locked buffers in
188 * tcp_do_retransmit()
189 * Eric Schenk : Changed receiver side silly window
190 * avoidance algorithm to BSD style
191 * algorithm. This doubles throughput
192 * against machines running Solaris,
193 * and seems to result in general
194 * improvement.
195 * Stefan Magdalinski : adjusted tcp_readable() to fix FIONREAD
196 * Willy Konynenberg : Transparent proxying support.
197 * Mike McLagan : Routing by source
198 * Keith Owens : Do proper merging with partial SKB's in
199 * tcp_do_sendmsg to avoid burstiness.
200 * Eric Schenk : Fix fast close down bug with
201 * shutdown() followed by close().
202 * Andi Kleen : Make poll agree with SIGIO
203 * Salvatore Sanfilippo : Support SO_LINGER with linger == 1 and
204 * lingertime == 0 (RFC 793 ABORT Call)
205 * Hirokazu Takahashi : Use copy_from_user() instead of
206 * csum_and_copy_from_user() if possible.
207 *
208 * This program is free software; you can redistribute it and/or
209 * modify it under the terms of the GNU General Public License
210 * as published by the Free Software Foundation; either version
211 * 2 of the License, or(at your option) any later version.
212 *
213 * Description of States:
214 *
215 * TCP_SYN_SENT sent a connection request, waiting for ack
216 *
217 * TCP_SYN_RECV received a connection request, sent ack,
218 * waiting for final ack in three-way handshake.
219 *
220 * TCP_ESTABLISHED connection established
221 *
222 * TCP_FIN_WAIT1 our side has shutdown, waiting to complete
223 * transmission of remaining buffered data
224 *
225 * TCP_FIN_WAIT2 all buffered data sent, waiting for remote
226 * to shutdown
227 *
228 * TCP_CLOSING both sides have shutdown but we still have
229 * data we have to finish sending
230 *
231 * TCP_TIME_WAIT timeout to catch resent junk before entering
232 * closed, can only be entered from FIN_WAIT2
233 * or CLOSING. Required because the other end
234 * may not have gotten our last ACK causing it
235 * to retransmit the data packet (which we ignore)
236 *
237 * TCP_CLOSE_WAIT remote side has shutdown and is waiting for
238 * us to finish writing our data and to shutdown
239 * (we have to close() to move on to LAST_ACK)
240 *
241 * TCP_LAST_ACK out side has shutdown after remote has
242 * shutdown. There may still be data in our
243 * buffer that we have to finish sending
244 *
245 * TCP_CLOSE socket is finished
246 */
247
248#define pr_fmt(fmt) "TCP: " fmt
249
250#include <linux/kernel.h>
251#include <linux/module.h>
252#include <linux/types.h>
253#include <linux/fcntl.h>
254#include <linux/poll.h>
255#include <linux/init.h>
256#include <linux/fs.h>
257#include <linux/skbuff.h>
258#include <linux/scatterlist.h>
259#include <linux/splice.h>
260#include <linux/net.h>
261#include <linux/socket.h>
262#include <linux/random.h>
263#include <linux/bootmem.h>
264#include <linux/highmem.h>
265#include <linux/swap.h>
266#include <linux/cache.h>
267#include <linux/err.h>
268#include <linux/crypto.h>
269#include <linux/time.h>
270#include <linux/slab.h>
271
272#include <net/icmp.h>
273#include <net/inet_common.h>
274#include <net/tcp.h>
275#include <net/xfrm.h>
276#include <net/ip.h>
277#include <net/netdma.h>
278#include <net/sock.h>
279
280#include <asm/uaccess.h>
281#include <asm/ioctls.h>
282#include <net/busy_poll.h>
283
284int sysctl_tcp_fin_timeout __read_mostly = TCP_FIN_TIMEOUT;
285
286int sysctl_tcp_min_tso_segs __read_mostly = 2;
287
288int sysctl_tcp_autocorking __read_mostly = 1;
289
290struct percpu_counter tcp_orphan_count;
291EXPORT_SYMBOL_GPL(tcp_orphan_count);
292
293long sysctl_tcp_mem[3] __read_mostly;
294int sysctl_tcp_wmem[3] __read_mostly;
295int sysctl_tcp_rmem[3] __read_mostly;
296
297EXPORT_SYMBOL(sysctl_tcp_mem);
298EXPORT_SYMBOL(sysctl_tcp_rmem);
299EXPORT_SYMBOL(sysctl_tcp_wmem);
300
301atomic_long_t tcp_memory_allocated; /* Current allocated memory. */
302EXPORT_SYMBOL(tcp_memory_allocated);
303
304/*
305 * Current number of TCP sockets.
306 */
307struct percpu_counter tcp_sockets_allocated;
308EXPORT_SYMBOL(tcp_sockets_allocated);
309
310/*
311 * TCP splice context
312 */
313struct tcp_splice_state {
314 struct pipe_inode_info *pipe;
315 size_t len;
316 unsigned int flags;
317};
318
319/*
320 * Pressure flag: try to collapse.
321 * Technical note: it is used by multiple contexts non atomically.
322 * All the __sk_mem_schedule() is of this nature: accounting
323 * is strict, actions are advisory and have some latency.
324 */
325int tcp_memory_pressure __read_mostly;
326EXPORT_SYMBOL(tcp_memory_pressure);
327
328void tcp_enter_memory_pressure(struct sock *sk)
329{
330 if (!tcp_memory_pressure) {
331 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMEMORYPRESSURES);
332 tcp_memory_pressure = 1;
333 }
334}
335EXPORT_SYMBOL(tcp_enter_memory_pressure);
336
337/* Convert seconds to retransmits based on initial and max timeout */
338static u8 secs_to_retrans(int seconds, int timeout, int rto_max)
339{
340 u8 res = 0;
341
342 if (seconds > 0) {
343 int period = timeout;
344
345 res = 1;
346 while (seconds > period && res < 255) {
347 res++;
348 timeout <<= 1;
349 if (timeout > rto_max)
350 timeout = rto_max;
351 period += timeout;
352 }
353 }
354 return res;
355}
356
357/* Convert retransmits to seconds based on initial and max timeout */
358static int retrans_to_secs(u8 retrans, int timeout, int rto_max)
359{
360 int period = 0;
361
362 if (retrans > 0) {
363 period = timeout;
364 while (--retrans) {
365 timeout <<= 1;
366 if (timeout > rto_max)
367 timeout = rto_max;
368 period += timeout;
369 }
370 }
371 return period;
372}
373
374/* Address-family independent initialization for a tcp_sock.
375 *
376 * NOTE: A lot of things set to zero explicitly by call to
377 * sk_alloc() so need not be done here.
378 */
379void tcp_init_sock(struct sock *sk)
380{
381 struct inet_connection_sock *icsk = inet_csk(sk);
382 struct tcp_sock *tp = tcp_sk(sk);
383
384 __skb_queue_head_init(&tp->out_of_order_queue);
385 tcp_init_xmit_timers(sk);
386 tcp_prequeue_init(tp);
387 INIT_LIST_HEAD(&tp->tsq_node);
388
389 icsk->icsk_rto = TCP_TIMEOUT_INIT;
390 tp->mdev_us = jiffies_to_usecs(TCP_TIMEOUT_INIT);
391
392 /* So many TCP implementations out there (incorrectly) count the
393 * initial SYN frame in their delayed-ACK and congestion control
394 * algorithms that we must have the following bandaid to talk
395 * efficiently to them. -DaveM
396 */
397 tp->snd_cwnd = TCP_INIT_CWND;
398
399 /* See draft-stevens-tcpca-spec-01 for discussion of the
400 * initialization of these values.
401 */
402 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
403 tp->snd_cwnd_clamp = ~0;
404 tp->mss_cache = TCP_MSS_DEFAULT;
405
406 tp->reordering = sysctl_tcp_reordering;
407 tcp_enable_early_retrans(tp);
408 icsk->icsk_ca_ops = &tcp_init_congestion_ops;
409
410 tp->tsoffset = 0;
411
412 sk->sk_state = TCP_CLOSE;
413
414 sk->sk_write_space = sk_stream_write_space;
415 sock_set_flag(sk, SOCK_USE_WRITE_QUEUE);
416
417 icsk->icsk_sync_mss = tcp_sync_mss;
418
419 sk->sk_sndbuf = sysctl_tcp_wmem[1];
420 sk->sk_rcvbuf = sysctl_tcp_rmem[1];
421
422 local_bh_disable();
423 sock_update_memcg(sk);
424 sk_sockets_allocated_inc(sk);
425 local_bh_enable();
426}
427EXPORT_SYMBOL(tcp_init_sock);
428
429/*
430 * Wait for a TCP event.
431 *
432 * Note that we don't need to lock the socket, as the upper poll layers
433 * take care of normal races (between the test and the event) and we don't
434 * go look at any of the socket buffers directly.
435 */
436unsigned int tcp_poll(struct file *file, struct socket *sock, poll_table *wait)
437{
438 unsigned int mask;
439 struct sock *sk = sock->sk;
440 const struct tcp_sock *tp = tcp_sk(sk);
441
442 sock_rps_record_flow(sk);
443
444 sock_poll_wait(file, sk_sleep(sk), wait);
445 if (sk->sk_state == TCP_LISTEN)
446 return inet_csk_listen_poll(sk);
447
448 /* Socket is not locked. We are protected from async events
449 * by poll logic and correct handling of state changes
450 * made by other threads is impossible in any case.
451 */
452
453 mask = 0;
454
455 /*
456 * POLLHUP is certainly not done right. But poll() doesn't
457 * have a notion of HUP in just one direction, and for a
458 * socket the read side is more interesting.
459 *
460 * Some poll() documentation says that POLLHUP is incompatible
461 * with the POLLOUT/POLLWR flags, so somebody should check this
462 * all. But careful, it tends to be safer to return too many
463 * bits than too few, and you can easily break real applications
464 * if you don't tell them that something has hung up!
465 *
466 * Check-me.
467 *
468 * Check number 1. POLLHUP is _UNMASKABLE_ event (see UNIX98 and
469 * our fs/select.c). It means that after we received EOF,
470 * poll always returns immediately, making impossible poll() on write()
471 * in state CLOSE_WAIT. One solution is evident --- to set POLLHUP
472 * if and only if shutdown has been made in both directions.
473 * Actually, it is interesting to look how Solaris and DUX
474 * solve this dilemma. I would prefer, if POLLHUP were maskable,
475 * then we could set it on SND_SHUTDOWN. BTW examples given
476 * in Stevens' books assume exactly this behaviour, it explains
477 * why POLLHUP is incompatible with POLLOUT. --ANK
478 *
479 * NOTE. Check for TCP_CLOSE is added. The goal is to prevent
480 * blocking on fresh not-connected or disconnected socket. --ANK
481 */
482 if (sk->sk_shutdown == SHUTDOWN_MASK || sk->sk_state == TCP_CLOSE)
483 mask |= POLLHUP;
484 if (sk->sk_shutdown & RCV_SHUTDOWN)
485 mask |= POLLIN | POLLRDNORM | POLLRDHUP;
486
487 /* Connected or passive Fast Open socket? */
488 if (sk->sk_state != TCP_SYN_SENT &&
489 (sk->sk_state != TCP_SYN_RECV || tp->fastopen_rsk != NULL)) {
490 int target = sock_rcvlowat(sk, 0, INT_MAX);
491
492 if (tp->urg_seq == tp->copied_seq &&
493 !sock_flag(sk, SOCK_URGINLINE) &&
494 tp->urg_data)
495 target++;
496
497 /* Potential race condition. If read of tp below will
498 * escape above sk->sk_state, we can be illegally awaken
499 * in SYN_* states. */
500 if (tp->rcv_nxt - tp->copied_seq >= target)
501 mask |= POLLIN | POLLRDNORM;
502
503 if (!(sk->sk_shutdown & SEND_SHUTDOWN)) {
504 if (sk_stream_is_writeable(sk)) {
505 mask |= POLLOUT | POLLWRNORM;
506 } else { /* send SIGIO later */
507 set_bit(SOCK_ASYNC_NOSPACE,
508 &sk->sk_socket->flags);
509 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
510
511 /* Race breaker. If space is freed after
512 * wspace test but before the flags are set,
513 * IO signal will be lost.
514 */
515 if (sk_stream_is_writeable(sk))
516 mask |= POLLOUT | POLLWRNORM;
517 }
518 } else
519 mask |= POLLOUT | POLLWRNORM;
520
521 if (tp->urg_data & TCP_URG_VALID)
522 mask |= POLLPRI;
523 }
524 /* This barrier is coupled with smp_wmb() in tcp_reset() */
525 smp_rmb();
526 if (sk->sk_err)
527 mask |= POLLERR;
528
529 return mask;
530}
531EXPORT_SYMBOL(tcp_poll);
532
533int tcp_ioctl(struct sock *sk, int cmd, unsigned long arg)
534{
535 struct tcp_sock *tp = tcp_sk(sk);
536 int answ;
537 bool slow;
538
539 switch (cmd) {
540 case SIOCINQ:
541 if (sk->sk_state == TCP_LISTEN)
542 return -EINVAL;
543
544 slow = lock_sock_fast(sk);
545 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
546 answ = 0;
547 else if (sock_flag(sk, SOCK_URGINLINE) ||
548 !tp->urg_data ||
549 before(tp->urg_seq, tp->copied_seq) ||
550 !before(tp->urg_seq, tp->rcv_nxt)) {
551
552 answ = tp->rcv_nxt - tp->copied_seq;
553
554 /* Subtract 1, if FIN was received */
555 if (answ && sock_flag(sk, SOCK_DONE))
556 answ--;
557 } else
558 answ = tp->urg_seq - tp->copied_seq;
559 unlock_sock_fast(sk, slow);
560 break;
561 case SIOCATMARK:
562 answ = tp->urg_data && tp->urg_seq == tp->copied_seq;
563 break;
564 case SIOCOUTQ:
565 if (sk->sk_state == TCP_LISTEN)
566 return -EINVAL;
567
568 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
569 answ = 0;
570 else
571 answ = tp->write_seq - tp->snd_una;
572 break;
573 case SIOCOUTQNSD:
574 if (sk->sk_state == TCP_LISTEN)
575 return -EINVAL;
576
577 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
578 answ = 0;
579 else
580 answ = tp->write_seq - tp->snd_nxt;
581 break;
582 default:
583 return -ENOIOCTLCMD;
584 }
585
586 return put_user(answ, (int __user *)arg);
587}
588EXPORT_SYMBOL(tcp_ioctl);
589
590static inline void tcp_mark_push(struct tcp_sock *tp, struct sk_buff *skb)
591{
592 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
593 tp->pushed_seq = tp->write_seq;
594}
595
596static inline bool forced_push(const struct tcp_sock *tp)
597{
598 return after(tp->write_seq, tp->pushed_seq + (tp->max_window >> 1));
599}
600
601static inline void skb_entail(struct sock *sk, struct sk_buff *skb)
602{
603 struct tcp_sock *tp = tcp_sk(sk);
604 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
605
606 skb->csum = 0;
607 tcb->seq = tcb->end_seq = tp->write_seq;
608 tcb->tcp_flags = TCPHDR_ACK;
609 tcb->sacked = 0;
610 skb_header_release(skb);
611 tcp_add_write_queue_tail(sk, skb);
612 sk->sk_wmem_queued += skb->truesize;
613 sk_mem_charge(sk, skb->truesize);
614 if (tp->nonagle & TCP_NAGLE_PUSH)
615 tp->nonagle &= ~TCP_NAGLE_PUSH;
616}
617
618static inline void tcp_mark_urg(struct tcp_sock *tp, int flags)
619{
620 if (flags & MSG_OOB)
621 tp->snd_up = tp->write_seq;
622}
623
624/* If a not yet filled skb is pushed, do not send it if
625 * we have data packets in Qdisc or NIC queues :
626 * Because TX completion will happen shortly, it gives a chance
627 * to coalesce future sendmsg() payload into this skb, without
628 * need for a timer, and with no latency trade off.
629 * As packets containing data payload have a bigger truesize
630 * than pure acks (dataless) packets, the last checks prevent
631 * autocorking if we only have an ACK in Qdisc/NIC queues,
632 * or if TX completion was delayed after we processed ACK packet.
633 */
634static bool tcp_should_autocork(struct sock *sk, struct sk_buff *skb,
635 int size_goal)
636{
637 return skb->len < size_goal &&
638 sysctl_tcp_autocorking &&
639 skb != tcp_write_queue_head(sk) &&
640 atomic_read(&sk->sk_wmem_alloc) > skb->truesize;
641}
642
643static void tcp_push(struct sock *sk, int flags, int mss_now,
644 int nonagle, int size_goal)
645{
646 struct tcp_sock *tp = tcp_sk(sk);
647 struct sk_buff *skb;
648
649 if (!tcp_send_head(sk))
650 return;
651
652 skb = tcp_write_queue_tail(sk);
653 if (!(flags & MSG_MORE) || forced_push(tp))
654 tcp_mark_push(tp, skb);
655
656 tcp_mark_urg(tp, flags);
657
658 if (tcp_should_autocork(sk, skb, size_goal)) {
659
660 /* avoid atomic op if TSQ_THROTTLED bit is already set */
661 if (!test_bit(TSQ_THROTTLED, &tp->tsq_flags)) {
662 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPAUTOCORKING);
663 set_bit(TSQ_THROTTLED, &tp->tsq_flags);
664 }
665 /* It is possible TX completion already happened
666 * before we set TSQ_THROTTLED.
667 */
668 if (atomic_read(&sk->sk_wmem_alloc) > skb->truesize)
669 return;
670 }
671
672 if (flags & MSG_MORE)
673 nonagle = TCP_NAGLE_CORK;
674
675 __tcp_push_pending_frames(sk, mss_now, nonagle);
676}
677
678static int tcp_splice_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb,
679 unsigned int offset, size_t len)
680{
681 struct tcp_splice_state *tss = rd_desc->arg.data;
682 int ret;
683
684 ret = skb_splice_bits(skb, offset, tss->pipe, min(rd_desc->count, len),
685 tss->flags);
686 if (ret > 0)
687 rd_desc->count -= ret;
688 return ret;
689}
690
691static int __tcp_splice_read(struct sock *sk, struct tcp_splice_state *tss)
692{
693 /* Store TCP splice context information in read_descriptor_t. */
694 read_descriptor_t rd_desc = {
695 .arg.data = tss,
696 .count = tss->len,
697 };
698
699 return tcp_read_sock(sk, &rd_desc, tcp_splice_data_recv);
700}
701
702/**
703 * tcp_splice_read - splice data from TCP socket to a pipe
704 * @sock: socket to splice from
705 * @ppos: position (not valid)
706 * @pipe: pipe to splice to
707 * @len: number of bytes to splice
708 * @flags: splice modifier flags
709 *
710 * Description:
711 * Will read pages from given socket and fill them into a pipe.
712 *
713 **/
714ssize_t tcp_splice_read(struct socket *sock, loff_t *ppos,
715 struct pipe_inode_info *pipe, size_t len,
716 unsigned int flags)
717{
718 struct sock *sk = sock->sk;
719 struct tcp_splice_state tss = {
720 .pipe = pipe,
721 .len = len,
722 .flags = flags,
723 };
724 long timeo;
725 ssize_t spliced;
726 int ret;
727
728 sock_rps_record_flow(sk);
729 /*
730 * We can't seek on a socket input
731 */
732 if (unlikely(*ppos))
733 return -ESPIPE;
734
735 ret = spliced = 0;
736
737 lock_sock(sk);
738
739 timeo = sock_rcvtimeo(sk, sock->file->f_flags & O_NONBLOCK);
740 while (tss.len) {
741 ret = __tcp_splice_read(sk, &tss);
742 if (ret < 0)
743 break;
744 else if (!ret) {
745 if (spliced)
746 break;
747 if (sock_flag(sk, SOCK_DONE))
748 break;
749 if (sk->sk_err) {
750 ret = sock_error(sk);
751 break;
752 }
753 if (sk->sk_shutdown & RCV_SHUTDOWN)
754 break;
755 if (sk->sk_state == TCP_CLOSE) {
756 /*
757 * This occurs when user tries to read
758 * from never connected socket.
759 */
760 if (!sock_flag(sk, SOCK_DONE))
761 ret = -ENOTCONN;
762 break;
763 }
764 if (!timeo) {
765 ret = -EAGAIN;
766 break;
767 }
768 sk_wait_data(sk, &timeo);
769 if (signal_pending(current)) {
770 ret = sock_intr_errno(timeo);
771 break;
772 }
773 continue;
774 }
775 tss.len -= ret;
776 spliced += ret;
777
778 if (!timeo)
779 break;
780 release_sock(sk);
781 lock_sock(sk);
782
783 if (sk->sk_err || sk->sk_state == TCP_CLOSE ||
784 (sk->sk_shutdown & RCV_SHUTDOWN) ||
785 signal_pending(current))
786 break;
787 }
788
789 release_sock(sk);
790
791 if (spliced)
792 return spliced;
793
794 return ret;
795}
796EXPORT_SYMBOL(tcp_splice_read);
797
798struct sk_buff *sk_stream_alloc_skb(struct sock *sk, int size, gfp_t gfp)
799{
800 struct sk_buff *skb;
801
802 /* The TCP header must be at least 32-bit aligned. */
803 size = ALIGN(size, 4);
804
805 skb = alloc_skb_fclone(size + sk->sk_prot->max_header, gfp);
806 if (skb) {
807 if (sk_wmem_schedule(sk, skb->truesize)) {
808 skb_reserve(skb, sk->sk_prot->max_header);
809 /*
810 * Make sure that we have exactly size bytes
811 * available to the caller, no more, no less.
812 */
813 skb->reserved_tailroom = skb->end - skb->tail - size;
814 return skb;
815 }
816 __kfree_skb(skb);
817 } else {
818 sk->sk_prot->enter_memory_pressure(sk);
819 sk_stream_moderate_sndbuf(sk);
820 }
821 return NULL;
822}
823
824static unsigned int tcp_xmit_size_goal(struct sock *sk, u32 mss_now,
825 int large_allowed)
826{
827 struct tcp_sock *tp = tcp_sk(sk);
828 u32 xmit_size_goal, old_size_goal;
829
830 xmit_size_goal = mss_now;
831
832 if (large_allowed && sk_can_gso(sk)) {
833 u32 gso_size, hlen;
834
835 /* Maybe we should/could use sk->sk_prot->max_header here ? */
836 hlen = inet_csk(sk)->icsk_af_ops->net_header_len +
837 inet_csk(sk)->icsk_ext_hdr_len +
838 tp->tcp_header_len;
839
840 /* Goal is to send at least one packet per ms,
841 * not one big TSO packet every 100 ms.
842 * This preserves ACK clocking and is consistent
843 * with tcp_tso_should_defer() heuristic.
844 */
845 gso_size = sk->sk_pacing_rate / (2 * MSEC_PER_SEC);
846 gso_size = max_t(u32, gso_size,
847 sysctl_tcp_min_tso_segs * mss_now);
848
849 xmit_size_goal = min_t(u32, gso_size,
850 sk->sk_gso_max_size - 1 - hlen);
851
852 xmit_size_goal = tcp_bound_to_half_wnd(tp, xmit_size_goal);
853
854 /* We try hard to avoid divides here */
855 old_size_goal = tp->xmit_size_goal_segs * mss_now;
856
857 if (likely(old_size_goal <= xmit_size_goal &&
858 old_size_goal + mss_now > xmit_size_goal)) {
859 xmit_size_goal = old_size_goal;
860 } else {
861 tp->xmit_size_goal_segs =
862 min_t(u16, xmit_size_goal / mss_now,
863 sk->sk_gso_max_segs);
864 xmit_size_goal = tp->xmit_size_goal_segs * mss_now;
865 }
866 }
867
868 return max(xmit_size_goal, mss_now);
869}
870
871static int tcp_send_mss(struct sock *sk, int *size_goal, int flags)
872{
873 int mss_now;
874
875 mss_now = tcp_current_mss(sk);
876 *size_goal = tcp_xmit_size_goal(sk, mss_now, !(flags & MSG_OOB));
877
878 return mss_now;
879}
880
881static ssize_t do_tcp_sendpages(struct sock *sk, struct page *page, int offset,
882 size_t size, int flags)
883{
884 struct tcp_sock *tp = tcp_sk(sk);
885 int mss_now, size_goal;
886 int err;
887 ssize_t copied;
888 long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
889
890 /* Wait for a connection to finish. One exception is TCP Fast Open
891 * (passive side) where data is allowed to be sent before a connection
892 * is fully established.
893 */
894 if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) &&
895 !tcp_passive_fastopen(sk)) {
896 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
897 goto out_err;
898 }
899
900 clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
901
902 mss_now = tcp_send_mss(sk, &size_goal, flags);
903 copied = 0;
904
905 err = -EPIPE;
906 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
907 goto out_err;
908
909 while (size > 0) {
910 struct sk_buff *skb = tcp_write_queue_tail(sk);
911 int copy, i;
912 bool can_coalesce;
913
914 if (!tcp_send_head(sk) || (copy = size_goal - skb->len) <= 0) {
915new_segment:
916 if (!sk_stream_memory_free(sk))
917 goto wait_for_sndbuf;
918
919 skb = sk_stream_alloc_skb(sk, 0, sk->sk_allocation);
920 if (!skb)
921 goto wait_for_memory;
922
923 skb_entail(sk, skb);
924 copy = size_goal;
925 }
926
927 if (copy > size)
928 copy = size;
929
930 i = skb_shinfo(skb)->nr_frags;
931 can_coalesce = skb_can_coalesce(skb, i, page, offset);
932 if (!can_coalesce && i >= MAX_SKB_FRAGS) {
933 tcp_mark_push(tp, skb);
934 goto new_segment;
935 }
936 if (!sk_wmem_schedule(sk, copy))
937 goto wait_for_memory;
938
939 if (can_coalesce) {
940 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
941 } else {
942 get_page(page);
943 skb_fill_page_desc(skb, i, page, offset, copy);
944 }
945 skb_shinfo(skb)->tx_flags |= SKBTX_SHARED_FRAG;
946
947 skb->len += copy;
948 skb->data_len += copy;
949 skb->truesize += copy;
950 sk->sk_wmem_queued += copy;
951 sk_mem_charge(sk, copy);
952 skb->ip_summed = CHECKSUM_PARTIAL;
953 tp->write_seq += copy;
954 TCP_SKB_CB(skb)->end_seq += copy;
955 skb_shinfo(skb)->gso_segs = 0;
956
957 if (!copied)
958 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
959
960 copied += copy;
961 offset += copy;
962 if (!(size -= copy))
963 goto out;
964
965 if (skb->len < size_goal || (flags & MSG_OOB))
966 continue;
967
968 if (forced_push(tp)) {
969 tcp_mark_push(tp, skb);
970 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
971 } else if (skb == tcp_send_head(sk))
972 tcp_push_one(sk, mss_now);
973 continue;
974
975wait_for_sndbuf:
976 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
977wait_for_memory:
978 tcp_push(sk, flags & ~MSG_MORE, mss_now,
979 TCP_NAGLE_PUSH, size_goal);
980
981 if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
982 goto do_error;
983
984 mss_now = tcp_send_mss(sk, &size_goal, flags);
985 }
986
987out:
988 if (copied && !(flags & MSG_SENDPAGE_NOTLAST))
989 tcp_push(sk, flags, mss_now, tp->nonagle, size_goal);
990 return copied;
991
992do_error:
993 if (copied)
994 goto out;
995out_err:
996 return sk_stream_error(sk, flags, err);
997}
998
999int tcp_sendpage(struct sock *sk, struct page *page, int offset,
1000 size_t size, int flags)
1001{
1002 ssize_t res;
1003
1004 if (!(sk->sk_route_caps & NETIF_F_SG) ||
1005 !(sk->sk_route_caps & NETIF_F_ALL_CSUM))
1006 return sock_no_sendpage(sk->sk_socket, page, offset, size,
1007 flags);
1008
1009 lock_sock(sk);
1010 res = do_tcp_sendpages(sk, page, offset, size, flags);
1011 release_sock(sk);
1012 return res;
1013}
1014EXPORT_SYMBOL(tcp_sendpage);
1015
1016static inline int select_size(const struct sock *sk, bool sg)
1017{
1018 const struct tcp_sock *tp = tcp_sk(sk);
1019 int tmp = tp->mss_cache;
1020
1021 if (sg) {
1022 if (sk_can_gso(sk)) {
1023 /* Small frames wont use a full page:
1024 * Payload will immediately follow tcp header.
1025 */
1026 tmp = SKB_WITH_OVERHEAD(2048 - MAX_TCP_HEADER);
1027 } else {
1028 int pgbreak = SKB_MAX_HEAD(MAX_TCP_HEADER);
1029
1030 if (tmp >= pgbreak &&
1031 tmp <= pgbreak + (MAX_SKB_FRAGS - 1) * PAGE_SIZE)
1032 tmp = pgbreak;
1033 }
1034 }
1035
1036 return tmp;
1037}
1038
1039void tcp_free_fastopen_req(struct tcp_sock *tp)
1040{
1041 if (tp->fastopen_req != NULL) {
1042 kfree(tp->fastopen_req);
1043 tp->fastopen_req = NULL;
1044 }
1045}
1046
1047static int tcp_sendmsg_fastopen(struct sock *sk, struct msghdr *msg,
1048 int *copied, size_t size)
1049{
1050 struct tcp_sock *tp = tcp_sk(sk);
1051 int err, flags;
1052
1053 if (!(sysctl_tcp_fastopen & TFO_CLIENT_ENABLE))
1054 return -EOPNOTSUPP;
1055 if (tp->fastopen_req != NULL)
1056 return -EALREADY; /* Another Fast Open is in progress */
1057
1058 tp->fastopen_req = kzalloc(sizeof(struct tcp_fastopen_request),
1059 sk->sk_allocation);
1060 if (unlikely(tp->fastopen_req == NULL))
1061 return -ENOBUFS;
1062 tp->fastopen_req->data = msg;
1063 tp->fastopen_req->size = size;
1064
1065 flags = (msg->msg_flags & MSG_DONTWAIT) ? O_NONBLOCK : 0;
1066 err = __inet_stream_connect(sk->sk_socket, msg->msg_name,
1067 msg->msg_namelen, flags);
1068 *copied = tp->fastopen_req->copied;
1069 tcp_free_fastopen_req(tp);
1070 return err;
1071}
1072
1073int tcp_sendmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg,
1074 size_t size)
1075{
1076 struct iovec *iov;
1077 struct tcp_sock *tp = tcp_sk(sk);
1078 struct sk_buff *skb;
1079 int iovlen, flags, err, copied = 0;
1080 int mss_now = 0, size_goal, copied_syn = 0, offset = 0;
1081 bool sg;
1082 long timeo;
1083
1084 lock_sock(sk);
1085
1086 flags = msg->msg_flags;
1087 if (flags & MSG_FASTOPEN) {
1088 err = tcp_sendmsg_fastopen(sk, msg, &copied_syn, size);
1089 if (err == -EINPROGRESS && copied_syn > 0)
1090 goto out;
1091 else if (err)
1092 goto out_err;
1093 offset = copied_syn;
1094 }
1095
1096 timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
1097
1098 /* Wait for a connection to finish. One exception is TCP Fast Open
1099 * (passive side) where data is allowed to be sent before a connection
1100 * is fully established.
1101 */
1102 if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) &&
1103 !tcp_passive_fastopen(sk)) {
1104 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
1105 goto do_error;
1106 }
1107
1108 if (unlikely(tp->repair)) {
1109 if (tp->repair_queue == TCP_RECV_QUEUE) {
1110 copied = tcp_send_rcvq(sk, msg, size);
1111 goto out;
1112 }
1113
1114 err = -EINVAL;
1115 if (tp->repair_queue == TCP_NO_QUEUE)
1116 goto out_err;
1117
1118 /* 'common' sending to sendq */
1119 }
1120
1121 /* This should be in poll */
1122 clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
1123
1124 mss_now = tcp_send_mss(sk, &size_goal, flags);
1125
1126 /* Ok commence sending. */
1127 iovlen = msg->msg_iovlen;
1128 iov = msg->msg_iov;
1129 copied = 0;
1130
1131 err = -EPIPE;
1132 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
1133 goto out_err;
1134
1135 sg = !!(sk->sk_route_caps & NETIF_F_SG);
1136
1137 while (--iovlen >= 0) {
1138 size_t seglen = iov->iov_len;
1139 unsigned char __user *from = iov->iov_base;
1140
1141 iov++;
1142 if (unlikely(offset > 0)) { /* Skip bytes copied in SYN */
1143 if (offset >= seglen) {
1144 offset -= seglen;
1145 continue;
1146 }
1147 seglen -= offset;
1148 from += offset;
1149 offset = 0;
1150 }
1151
1152 while (seglen > 0) {
1153 int copy = 0;
1154 int max = size_goal;
1155
1156 skb = tcp_write_queue_tail(sk);
1157 if (tcp_send_head(sk)) {
1158 if (skb->ip_summed == CHECKSUM_NONE)
1159 max = mss_now;
1160 copy = max - skb->len;
1161 }
1162
1163 if (copy <= 0) {
1164new_segment:
1165 /* Allocate new segment. If the interface is SG,
1166 * allocate skb fitting to single page.
1167 */
1168 if (!sk_stream_memory_free(sk))
1169 goto wait_for_sndbuf;
1170
1171 skb = sk_stream_alloc_skb(sk,
1172 select_size(sk, sg),
1173 sk->sk_allocation);
1174 if (!skb)
1175 goto wait_for_memory;
1176
1177 /*
1178 * All packets are restored as if they have
1179 * already been sent.
1180 */
1181 if (tp->repair)
1182 TCP_SKB_CB(skb)->when = tcp_time_stamp;
1183
1184 /*
1185 * Check whether we can use HW checksum.
1186 */
1187 if (sk->sk_route_caps & NETIF_F_ALL_CSUM)
1188 skb->ip_summed = CHECKSUM_PARTIAL;
1189
1190 skb_entail(sk, skb);
1191 copy = size_goal;
1192 max = size_goal;
1193 }
1194
1195 /* Try to append data to the end of skb. */
1196 if (copy > seglen)
1197 copy = seglen;
1198
1199 /* Where to copy to? */
1200 if (skb_availroom(skb) > 0) {
1201 /* We have some space in skb head. Superb! */
1202 copy = min_t(int, copy, skb_availroom(skb));
1203 err = skb_add_data_nocache(sk, skb, from, copy);
1204 if (err)
1205 goto do_fault;
1206 } else {
1207 bool merge = true;
1208 int i = skb_shinfo(skb)->nr_frags;
1209 struct page_frag *pfrag = sk_page_frag(sk);
1210
1211 if (!sk_page_frag_refill(sk, pfrag))
1212 goto wait_for_memory;
1213
1214 if (!skb_can_coalesce(skb, i, pfrag->page,
1215 pfrag->offset)) {
1216 if (i == MAX_SKB_FRAGS || !sg) {
1217 tcp_mark_push(tp, skb);
1218 goto new_segment;
1219 }
1220 merge = false;
1221 }
1222
1223 copy = min_t(int, copy, pfrag->size - pfrag->offset);
1224
1225 if (!sk_wmem_schedule(sk, copy))
1226 goto wait_for_memory;
1227
1228 err = skb_copy_to_page_nocache(sk, from, skb,
1229 pfrag->page,
1230 pfrag->offset,
1231 copy);
1232 if (err)
1233 goto do_error;
1234
1235 /* Update the skb. */
1236 if (merge) {
1237 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1238 } else {
1239 skb_fill_page_desc(skb, i, pfrag->page,
1240 pfrag->offset, copy);
1241 get_page(pfrag->page);
1242 }
1243 pfrag->offset += copy;
1244 }
1245
1246 if (!copied)
1247 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
1248
1249 tp->write_seq += copy;
1250 TCP_SKB_CB(skb)->end_seq += copy;
1251 skb_shinfo(skb)->gso_segs = 0;
1252
1253 from += copy;
1254 copied += copy;
1255 if ((seglen -= copy) == 0 && iovlen == 0)
1256 goto out;
1257
1258 if (skb->len < max || (flags & MSG_OOB) || unlikely(tp->repair))
1259 continue;
1260
1261 if (forced_push(tp)) {
1262 tcp_mark_push(tp, skb);
1263 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
1264 } else if (skb == tcp_send_head(sk))
1265 tcp_push_one(sk, mss_now);
1266 continue;
1267
1268wait_for_sndbuf:
1269 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1270wait_for_memory:
1271 if (copied)
1272 tcp_push(sk, flags & ~MSG_MORE, mss_now,
1273 TCP_NAGLE_PUSH, size_goal);
1274
1275 if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
1276 goto do_error;
1277
1278 mss_now = tcp_send_mss(sk, &size_goal, flags);
1279 }
1280 }
1281
1282out:
1283 if (copied)
1284 tcp_push(sk, flags, mss_now, tp->nonagle, size_goal);
1285 release_sock(sk);
1286 return copied + copied_syn;
1287
1288do_fault:
1289 if (!skb->len) {
1290 tcp_unlink_write_queue(skb, sk);
1291 /* It is the one place in all of TCP, except connection
1292 * reset, where we can be unlinking the send_head.
1293 */
1294 tcp_check_send_head(sk, skb);
1295 sk_wmem_free_skb(sk, skb);
1296 }
1297
1298do_error:
1299 if (copied + copied_syn)
1300 goto out;
1301out_err:
1302 err = sk_stream_error(sk, flags, err);
1303 release_sock(sk);
1304 return err;
1305}
1306EXPORT_SYMBOL(tcp_sendmsg);
1307
1308/*
1309 * Handle reading urgent data. BSD has very simple semantics for
1310 * this, no blocking and very strange errors 8)
1311 */
1312
1313static int tcp_recv_urg(struct sock *sk, struct msghdr *msg, int len, int flags)
1314{
1315 struct tcp_sock *tp = tcp_sk(sk);
1316
1317 /* No URG data to read. */
1318 if (sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data ||
1319 tp->urg_data == TCP_URG_READ)
1320 return -EINVAL; /* Yes this is right ! */
1321
1322 if (sk->sk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DONE))
1323 return -ENOTCONN;
1324
1325 if (tp->urg_data & TCP_URG_VALID) {
1326 int err = 0;
1327 char c = tp->urg_data;
1328
1329 if (!(flags & MSG_PEEK))
1330 tp->urg_data = TCP_URG_READ;
1331
1332 /* Read urgent data. */
1333 msg->msg_flags |= MSG_OOB;
1334
1335 if (len > 0) {
1336 if (!(flags & MSG_TRUNC))
1337 err = memcpy_toiovec(msg->msg_iov, &c, 1);
1338 len = 1;
1339 } else
1340 msg->msg_flags |= MSG_TRUNC;
1341
1342 return err ? -EFAULT : len;
1343 }
1344
1345 if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
1346 return 0;
1347
1348 /* Fixed the recv(..., MSG_OOB) behaviour. BSD docs and
1349 * the available implementations agree in this case:
1350 * this call should never block, independent of the
1351 * blocking state of the socket.
1352 * Mike <pall@rz.uni-karlsruhe.de>
1353 */
1354 return -EAGAIN;
1355}
1356
1357static int tcp_peek_sndq(struct sock *sk, struct msghdr *msg, int len)
1358{
1359 struct sk_buff *skb;
1360 int copied = 0, err = 0;
1361
1362 /* XXX -- need to support SO_PEEK_OFF */
1363
1364 skb_queue_walk(&sk->sk_write_queue, skb) {
1365 err = skb_copy_datagram_iovec(skb, 0, msg->msg_iov, skb->len);
1366 if (err)
1367 break;
1368
1369 copied += skb->len;
1370 }
1371
1372 return err ?: copied;
1373}
1374
1375/* Clean up the receive buffer for full frames taken by the user,
1376 * then send an ACK if necessary. COPIED is the number of bytes
1377 * tcp_recvmsg has given to the user so far, it speeds up the
1378 * calculation of whether or not we must ACK for the sake of
1379 * a window update.
1380 */
1381void tcp_cleanup_rbuf(struct sock *sk, int copied)
1382{
1383 struct tcp_sock *tp = tcp_sk(sk);
1384 bool time_to_ack = false;
1385
1386 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
1387
1388 WARN(skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq),
1389 "cleanup rbuf bug: copied %X seq %X rcvnxt %X\n",
1390 tp->copied_seq, TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt);
1391
1392 if (inet_csk_ack_scheduled(sk)) {
1393 const struct inet_connection_sock *icsk = inet_csk(sk);
1394 /* Delayed ACKs frequently hit locked sockets during bulk
1395 * receive. */
1396 if (icsk->icsk_ack.blocked ||
1397 /* Once-per-two-segments ACK was not sent by tcp_input.c */
1398 tp->rcv_nxt - tp->rcv_wup > icsk->icsk_ack.rcv_mss ||
1399 /*
1400 * If this read emptied read buffer, we send ACK, if
1401 * connection is not bidirectional, user drained
1402 * receive buffer and there was a small segment
1403 * in queue.
1404 */
1405 (copied > 0 &&
1406 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED2) ||
1407 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED) &&
1408 !icsk->icsk_ack.pingpong)) &&
1409 !atomic_read(&sk->sk_rmem_alloc)))
1410 time_to_ack = true;
1411 }
1412
1413 /* We send an ACK if we can now advertise a non-zero window
1414 * which has been raised "significantly".
1415 *
1416 * Even if window raised up to infinity, do not send window open ACK
1417 * in states, where we will not receive more. It is useless.
1418 */
1419 if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) {
1420 __u32 rcv_window_now = tcp_receive_window(tp);
1421
1422 /* Optimize, __tcp_select_window() is not cheap. */
1423 if (2*rcv_window_now <= tp->window_clamp) {
1424 __u32 new_window = __tcp_select_window(sk);
1425
1426 /* Send ACK now, if this read freed lots of space
1427 * in our buffer. Certainly, new_window is new window.
1428 * We can advertise it now, if it is not less than current one.
1429 * "Lots" means "at least twice" here.
1430 */
1431 if (new_window && new_window >= 2 * rcv_window_now)
1432 time_to_ack = true;
1433 }
1434 }
1435 if (time_to_ack)
1436 tcp_send_ack(sk);
1437}
1438
1439static void tcp_prequeue_process(struct sock *sk)
1440{
1441 struct sk_buff *skb;
1442 struct tcp_sock *tp = tcp_sk(sk);
1443
1444 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPPREQUEUED);
1445
1446 /* RX process wants to run with disabled BHs, though it is not
1447 * necessary */
1448 local_bh_disable();
1449 while ((skb = __skb_dequeue(&tp->ucopy.prequeue)) != NULL)
1450 sk_backlog_rcv(sk, skb);
1451 local_bh_enable();
1452
1453 /* Clear memory counter. */
1454 tp->ucopy.memory = 0;
1455}
1456
1457#ifdef CONFIG_NET_DMA
1458static void tcp_service_net_dma(struct sock *sk, bool wait)
1459{
1460 dma_cookie_t done, used;
1461 dma_cookie_t last_issued;
1462 struct tcp_sock *tp = tcp_sk(sk);
1463
1464 if (!tp->ucopy.dma_chan)
1465 return;
1466
1467 last_issued = tp->ucopy.dma_cookie;
1468 dma_async_issue_pending(tp->ucopy.dma_chan);
1469
1470 do {
1471 if (dma_async_is_tx_complete(tp->ucopy.dma_chan,
1472 last_issued, &done,
1473 &used) == DMA_COMPLETE) {
1474 /* Safe to free early-copied skbs now */
1475 __skb_queue_purge(&sk->sk_async_wait_queue);
1476 break;
1477 } else {
1478 struct sk_buff *skb;
1479 while ((skb = skb_peek(&sk->sk_async_wait_queue)) &&
1480 (dma_async_is_complete(skb->dma_cookie, done,
1481 used) == DMA_COMPLETE)) {
1482 __skb_dequeue(&sk->sk_async_wait_queue);
1483 kfree_skb(skb);
1484 }
1485 }
1486 } while (wait);
1487}
1488#endif
1489
1490static struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off)
1491{
1492 struct sk_buff *skb;
1493 u32 offset;
1494
1495 while ((skb = skb_peek(&sk->sk_receive_queue)) != NULL) {
1496 offset = seq - TCP_SKB_CB(skb)->seq;
1497 if (tcp_hdr(skb)->syn)
1498 offset--;
1499 if (offset < skb->len || tcp_hdr(skb)->fin) {
1500 *off = offset;
1501 return skb;
1502 }
1503 /* This looks weird, but this can happen if TCP collapsing
1504 * splitted a fat GRO packet, while we released socket lock
1505 * in skb_splice_bits()
1506 */
1507 sk_eat_skb(sk, skb, false);
1508 }
1509 return NULL;
1510}
1511
1512/*
1513 * This routine provides an alternative to tcp_recvmsg() for routines
1514 * that would like to handle copying from skbuffs directly in 'sendfile'
1515 * fashion.
1516 * Note:
1517 * - It is assumed that the socket was locked by the caller.
1518 * - The routine does not block.
1519 * - At present, there is no support for reading OOB data
1520 * or for 'peeking' the socket using this routine
1521 * (although both would be easy to implement).
1522 */
1523int tcp_read_sock(struct sock *sk, read_descriptor_t *desc,
1524 sk_read_actor_t recv_actor)
1525{
1526 struct sk_buff *skb;
1527 struct tcp_sock *tp = tcp_sk(sk);
1528 u32 seq = tp->copied_seq;
1529 u32 offset;
1530 int copied = 0;
1531
1532 if (sk->sk_state == TCP_LISTEN)
1533 return -ENOTCONN;
1534 while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1535 if (offset < skb->len) {
1536 int used;
1537 size_t len;
1538
1539 len = skb->len - offset;
1540 /* Stop reading if we hit a patch of urgent data */
1541 if (tp->urg_data) {
1542 u32 urg_offset = tp->urg_seq - seq;
1543 if (urg_offset < len)
1544 len = urg_offset;
1545 if (!len)
1546 break;
1547 }
1548 used = recv_actor(desc, skb, offset, len);
1549 if (used <= 0) {
1550 if (!copied)
1551 copied = used;
1552 break;
1553 } else if (used <= len) {
1554 seq += used;
1555 copied += used;
1556 offset += used;
1557 }
1558 /* If recv_actor drops the lock (e.g. TCP splice
1559 * receive) the skb pointer might be invalid when
1560 * getting here: tcp_collapse might have deleted it
1561 * while aggregating skbs from the socket queue.
1562 */
1563 skb = tcp_recv_skb(sk, seq - 1, &offset);
1564 if (!skb)
1565 break;
1566 /* TCP coalescing might have appended data to the skb.
1567 * Try to splice more frags
1568 */
1569 if (offset + 1 != skb->len)
1570 continue;
1571 }
1572 if (tcp_hdr(skb)->fin) {
1573 sk_eat_skb(sk, skb, false);
1574 ++seq;
1575 break;
1576 }
1577 sk_eat_skb(sk, skb, false);
1578 if (!desc->count)
1579 break;
1580 tp->copied_seq = seq;
1581 }
1582 tp->copied_seq = seq;
1583
1584 tcp_rcv_space_adjust(sk);
1585
1586 /* Clean up data we have read: This will do ACK frames. */
1587 if (copied > 0) {
1588 tcp_recv_skb(sk, seq, &offset);
1589 tcp_cleanup_rbuf(sk, copied);
1590 }
1591 return copied;
1592}
1593EXPORT_SYMBOL(tcp_read_sock);
1594
1595/*
1596 * This routine copies from a sock struct into the user buffer.
1597 *
1598 * Technical note: in 2.3 we work on _locked_ socket, so that
1599 * tricks with *seq access order and skb->users are not required.
1600 * Probably, code can be easily improved even more.
1601 */
1602
1603int tcp_recvmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg,
1604 size_t len, int nonblock, int flags, int *addr_len)
1605{
1606 struct tcp_sock *tp = tcp_sk(sk);
1607 int copied = 0;
1608 u32 peek_seq;
1609 u32 *seq;
1610 unsigned long used;
1611 int err;
1612 int target; /* Read at least this many bytes */
1613 long timeo;
1614 struct task_struct *user_recv = NULL;
1615 bool copied_early = false;
1616 struct sk_buff *skb;
1617 u32 urg_hole = 0;
1618
1619 if (sk_can_busy_loop(sk) && skb_queue_empty(&sk->sk_receive_queue) &&
1620 (sk->sk_state == TCP_ESTABLISHED))
1621 sk_busy_loop(sk, nonblock);
1622
1623 lock_sock(sk);
1624
1625 err = -ENOTCONN;
1626 if (sk->sk_state == TCP_LISTEN)
1627 goto out;
1628
1629 timeo = sock_rcvtimeo(sk, nonblock);
1630
1631 /* Urgent data needs to be handled specially. */
1632 if (flags & MSG_OOB)
1633 goto recv_urg;
1634
1635 if (unlikely(tp->repair)) {
1636 err = -EPERM;
1637 if (!(flags & MSG_PEEK))
1638 goto out;
1639
1640 if (tp->repair_queue == TCP_SEND_QUEUE)
1641 goto recv_sndq;
1642
1643 err = -EINVAL;
1644 if (tp->repair_queue == TCP_NO_QUEUE)
1645 goto out;
1646
1647 /* 'common' recv queue MSG_PEEK-ing */
1648 }
1649
1650 seq = &tp->copied_seq;
1651 if (flags & MSG_PEEK) {
1652 peek_seq = tp->copied_seq;
1653 seq = &peek_seq;
1654 }
1655
1656 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
1657
1658#ifdef CONFIG_NET_DMA
1659 tp->ucopy.dma_chan = NULL;
1660 preempt_disable();
1661 skb = skb_peek_tail(&sk->sk_receive_queue);
1662 {
1663 int available = 0;
1664
1665 if (skb)
1666 available = TCP_SKB_CB(skb)->seq + skb->len - (*seq);
1667 if ((available < target) &&
1668 (len > sysctl_tcp_dma_copybreak) && !(flags & MSG_PEEK) &&
1669 !sysctl_tcp_low_latency &&
1670 net_dma_find_channel()) {
1671 preempt_enable();
1672 tp->ucopy.pinned_list =
1673 dma_pin_iovec_pages(msg->msg_iov, len);
1674 } else {
1675 preempt_enable();
1676 }
1677 }
1678#endif
1679
1680 do {
1681 u32 offset;
1682
1683 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
1684 if (tp->urg_data && tp->urg_seq == *seq) {
1685 if (copied)
1686 break;
1687 if (signal_pending(current)) {
1688 copied = timeo ? sock_intr_errno(timeo) : -EAGAIN;
1689 break;
1690 }
1691 }
1692
1693 /* Next get a buffer. */
1694
1695 skb_queue_walk(&sk->sk_receive_queue, skb) {
1696 /* Now that we have two receive queues this
1697 * shouldn't happen.
1698 */
1699 if (WARN(before(*seq, TCP_SKB_CB(skb)->seq),
1700 "recvmsg bug: copied %X seq %X rcvnxt %X fl %X\n",
1701 *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt,
1702 flags))
1703 break;
1704
1705 offset = *seq - TCP_SKB_CB(skb)->seq;
1706 if (tcp_hdr(skb)->syn)
1707 offset--;
1708 if (offset < skb->len)
1709 goto found_ok_skb;
1710 if (tcp_hdr(skb)->fin)
1711 goto found_fin_ok;
1712 WARN(!(flags & MSG_PEEK),
1713 "recvmsg bug 2: copied %X seq %X rcvnxt %X fl %X\n",
1714 *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt, flags);
1715 }
1716
1717 /* Well, if we have backlog, try to process it now yet. */
1718
1719 if (copied >= target && !sk->sk_backlog.tail)
1720 break;
1721
1722 if (copied) {
1723 if (sk->sk_err ||
1724 sk->sk_state == TCP_CLOSE ||
1725 (sk->sk_shutdown & RCV_SHUTDOWN) ||
1726 !timeo ||
1727 signal_pending(current))
1728 break;
1729 } else {
1730 if (sock_flag(sk, SOCK_DONE))
1731 break;
1732
1733 if (sk->sk_err) {
1734 copied = sock_error(sk);
1735 break;
1736 }
1737
1738 if (sk->sk_shutdown & RCV_SHUTDOWN)
1739 break;
1740
1741 if (sk->sk_state == TCP_CLOSE) {
1742 if (!sock_flag(sk, SOCK_DONE)) {
1743 /* This occurs when user tries to read
1744 * from never connected socket.
1745 */
1746 copied = -ENOTCONN;
1747 break;
1748 }
1749 break;
1750 }
1751
1752 if (!timeo) {
1753 copied = -EAGAIN;
1754 break;
1755 }
1756
1757 if (signal_pending(current)) {
1758 copied = sock_intr_errno(timeo);
1759 break;
1760 }
1761 }
1762
1763 tcp_cleanup_rbuf(sk, copied);
1764
1765 if (!sysctl_tcp_low_latency && tp->ucopy.task == user_recv) {
1766 /* Install new reader */
1767 if (!user_recv && !(flags & (MSG_TRUNC | MSG_PEEK))) {
1768 user_recv = current;
1769 tp->ucopy.task = user_recv;
1770 tp->ucopy.iov = msg->msg_iov;
1771 }
1772
1773 tp->ucopy.len = len;
1774
1775 WARN_ON(tp->copied_seq != tp->rcv_nxt &&
1776 !(flags & (MSG_PEEK | MSG_TRUNC)));
1777
1778 /* Ugly... If prequeue is not empty, we have to
1779 * process it before releasing socket, otherwise
1780 * order will be broken at second iteration.
1781 * More elegant solution is required!!!
1782 *
1783 * Look: we have the following (pseudo)queues:
1784 *
1785 * 1. packets in flight
1786 * 2. backlog
1787 * 3. prequeue
1788 * 4. receive_queue
1789 *
1790 * Each queue can be processed only if the next ones
1791 * are empty. At this point we have empty receive_queue.
1792 * But prequeue _can_ be not empty after 2nd iteration,
1793 * when we jumped to start of loop because backlog
1794 * processing added something to receive_queue.
1795 * We cannot release_sock(), because backlog contains
1796 * packets arrived _after_ prequeued ones.
1797 *
1798 * Shortly, algorithm is clear --- to process all
1799 * the queues in order. We could make it more directly,
1800 * requeueing packets from backlog to prequeue, if
1801 * is not empty. It is more elegant, but eats cycles,
1802 * unfortunately.
1803 */
1804 if (!skb_queue_empty(&tp->ucopy.prequeue))
1805 goto do_prequeue;
1806
1807 /* __ Set realtime policy in scheduler __ */
1808 }
1809
1810#ifdef CONFIG_NET_DMA
1811 if (tp->ucopy.dma_chan) {
1812 if (tp->rcv_wnd == 0 &&
1813 !skb_queue_empty(&sk->sk_async_wait_queue)) {
1814 tcp_service_net_dma(sk, true);
1815 tcp_cleanup_rbuf(sk, copied);
1816 } else
1817 dma_async_issue_pending(tp->ucopy.dma_chan);
1818 }
1819#endif
1820 if (copied >= target) {
1821 /* Do not sleep, just process backlog. */
1822 release_sock(sk);
1823 lock_sock(sk);
1824 } else
1825 sk_wait_data(sk, &timeo);
1826
1827#ifdef CONFIG_NET_DMA
1828 tcp_service_net_dma(sk, false); /* Don't block */
1829 tp->ucopy.wakeup = 0;
1830#endif
1831
1832 if (user_recv) {
1833 int chunk;
1834
1835 /* __ Restore normal policy in scheduler __ */
1836
1837 if ((chunk = len - tp->ucopy.len) != 0) {
1838 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMBACKLOG, chunk);
1839 len -= chunk;
1840 copied += chunk;
1841 }
1842
1843 if (tp->rcv_nxt == tp->copied_seq &&
1844 !skb_queue_empty(&tp->ucopy.prequeue)) {
1845do_prequeue:
1846 tcp_prequeue_process(sk);
1847
1848 if ((chunk = len - tp->ucopy.len) != 0) {
1849 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1850 len -= chunk;
1851 copied += chunk;
1852 }
1853 }
1854 }
1855 if ((flags & MSG_PEEK) &&
1856 (peek_seq - copied - urg_hole != tp->copied_seq)) {
1857 net_dbg_ratelimited("TCP(%s:%d): Application bug, race in MSG_PEEK\n",
1858 current->comm,
1859 task_pid_nr(current));
1860 peek_seq = tp->copied_seq;
1861 }
1862 continue;
1863
1864 found_ok_skb:
1865 /* Ok so how much can we use? */
1866 used = skb->len - offset;
1867 if (len < used)
1868 used = len;
1869
1870 /* Do we have urgent data here? */
1871 if (tp->urg_data) {
1872 u32 urg_offset = tp->urg_seq - *seq;
1873 if (urg_offset < used) {
1874 if (!urg_offset) {
1875 if (!sock_flag(sk, SOCK_URGINLINE)) {
1876 ++*seq;
1877 urg_hole++;
1878 offset++;
1879 used--;
1880 if (!used)
1881 goto skip_copy;
1882 }
1883 } else
1884 used = urg_offset;
1885 }
1886 }
1887
1888 if (!(flags & MSG_TRUNC)) {
1889#ifdef CONFIG_NET_DMA
1890 if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list)
1891 tp->ucopy.dma_chan = net_dma_find_channel();
1892
1893 if (tp->ucopy.dma_chan) {
1894 tp->ucopy.dma_cookie = dma_skb_copy_datagram_iovec(
1895 tp->ucopy.dma_chan, skb, offset,
1896 msg->msg_iov, used,
1897 tp->ucopy.pinned_list);
1898
1899 if (tp->ucopy.dma_cookie < 0) {
1900
1901 pr_alert("%s: dma_cookie < 0\n",
1902 __func__);
1903
1904 /* Exception. Bailout! */
1905 if (!copied)
1906 copied = -EFAULT;
1907 break;
1908 }
1909
1910 dma_async_issue_pending(tp->ucopy.dma_chan);
1911
1912 if ((offset + used) == skb->len)
1913 copied_early = true;
1914
1915 } else
1916#endif
1917 {
1918 err = skb_copy_datagram_iovec(skb, offset,
1919 msg->msg_iov, used);
1920 if (err) {
1921 /* Exception. Bailout! */
1922 if (!copied)
1923 copied = -EFAULT;
1924 break;
1925 }
1926 }
1927 }
1928
1929 *seq += used;
1930 copied += used;
1931 len -= used;
1932
1933 tcp_rcv_space_adjust(sk);
1934
1935skip_copy:
1936 if (tp->urg_data && after(tp->copied_seq, tp->urg_seq)) {
1937 tp->urg_data = 0;
1938 tcp_fast_path_check(sk);
1939 }
1940 if (used + offset < skb->len)
1941 continue;
1942
1943 if (tcp_hdr(skb)->fin)
1944 goto found_fin_ok;
1945 if (!(flags & MSG_PEEK)) {
1946 sk_eat_skb(sk, skb, copied_early);
1947 copied_early = false;
1948 }
1949 continue;
1950
1951 found_fin_ok:
1952 /* Process the FIN. */
1953 ++*seq;
1954 if (!(flags & MSG_PEEK)) {
1955 sk_eat_skb(sk, skb, copied_early);
1956 copied_early = false;
1957 }
1958 break;
1959 } while (len > 0);
1960
1961 if (user_recv) {
1962 if (!skb_queue_empty(&tp->ucopy.prequeue)) {
1963 int chunk;
1964
1965 tp->ucopy.len = copied > 0 ? len : 0;
1966
1967 tcp_prequeue_process(sk);
1968
1969 if (copied > 0 && (chunk = len - tp->ucopy.len) != 0) {
1970 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1971 len -= chunk;
1972 copied += chunk;
1973 }
1974 }
1975
1976 tp->ucopy.task = NULL;
1977 tp->ucopy.len = 0;
1978 }
1979
1980#ifdef CONFIG_NET_DMA
1981 tcp_service_net_dma(sk, true); /* Wait for queue to drain */
1982 tp->ucopy.dma_chan = NULL;
1983
1984 if (tp->ucopy.pinned_list) {
1985 dma_unpin_iovec_pages(tp->ucopy.pinned_list);
1986 tp->ucopy.pinned_list = NULL;
1987 }
1988#endif
1989
1990 /* According to UNIX98, msg_name/msg_namelen are ignored
1991 * on connected socket. I was just happy when found this 8) --ANK
1992 */
1993
1994 /* Clean up data we have read: This will do ACK frames. */
1995 tcp_cleanup_rbuf(sk, copied);
1996
1997 release_sock(sk);
1998 return copied;
1999
2000out:
2001 release_sock(sk);
2002 return err;
2003
2004recv_urg:
2005 err = tcp_recv_urg(sk, msg, len, flags);
2006 goto out;
2007
2008recv_sndq:
2009 err = tcp_peek_sndq(sk, msg, len);
2010 goto out;
2011}
2012EXPORT_SYMBOL(tcp_recvmsg);
2013
2014void tcp_set_state(struct sock *sk, int state)
2015{
2016 int oldstate = sk->sk_state;
2017
2018 switch (state) {
2019 case TCP_ESTABLISHED:
2020 if (oldstate != TCP_ESTABLISHED)
2021 TCP_INC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
2022 break;
2023
2024 case TCP_CLOSE:
2025 if (oldstate == TCP_CLOSE_WAIT || oldstate == TCP_ESTABLISHED)
2026 TCP_INC_STATS(sock_net(sk), TCP_MIB_ESTABRESETS);
2027
2028 sk->sk_prot->unhash(sk);
2029 if (inet_csk(sk)->icsk_bind_hash &&
2030 !(sk->sk_userlocks & SOCK_BINDPORT_LOCK))
2031 inet_put_port(sk);
2032 /* fall through */
2033 default:
2034 if (oldstate == TCP_ESTABLISHED)
2035 TCP_DEC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
2036 }
2037
2038 /* Change state AFTER socket is unhashed to avoid closed
2039 * socket sitting in hash tables.
2040 */
2041 sk->sk_state = state;
2042
2043#ifdef STATE_TRACE
2044 SOCK_DEBUG(sk, "TCP sk=%p, State %s -> %s\n", sk, statename[oldstate], statename[state]);
2045#endif
2046}
2047EXPORT_SYMBOL_GPL(tcp_set_state);
2048
2049/*
2050 * State processing on a close. This implements the state shift for
2051 * sending our FIN frame. Note that we only send a FIN for some
2052 * states. A shutdown() may have already sent the FIN, or we may be
2053 * closed.
2054 */
2055
2056static const unsigned char new_state[16] = {
2057 /* current state: new state: action: */
2058 /* (Invalid) */ TCP_CLOSE,
2059 /* TCP_ESTABLISHED */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2060 /* TCP_SYN_SENT */ TCP_CLOSE,
2061 /* TCP_SYN_RECV */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2062 /* TCP_FIN_WAIT1 */ TCP_FIN_WAIT1,
2063 /* TCP_FIN_WAIT2 */ TCP_FIN_WAIT2,
2064 /* TCP_TIME_WAIT */ TCP_CLOSE,
2065 /* TCP_CLOSE */ TCP_CLOSE,
2066 /* TCP_CLOSE_WAIT */ TCP_LAST_ACK | TCP_ACTION_FIN,
2067 /* TCP_LAST_ACK */ TCP_LAST_ACK,
2068 /* TCP_LISTEN */ TCP_CLOSE,
2069 /* TCP_CLOSING */ TCP_CLOSING,
2070};
2071
2072static int tcp_close_state(struct sock *sk)
2073{
2074 int next = (int)new_state[sk->sk_state];
2075 int ns = next & TCP_STATE_MASK;
2076
2077 tcp_set_state(sk, ns);
2078
2079 return next & TCP_ACTION_FIN;
2080}
2081
2082/*
2083 * Shutdown the sending side of a connection. Much like close except
2084 * that we don't receive shut down or sock_set_flag(sk, SOCK_DEAD).
2085 */
2086
2087void tcp_shutdown(struct sock *sk, int how)
2088{
2089 /* We need to grab some memory, and put together a FIN,
2090 * and then put it into the queue to be sent.
2091 * Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
2092 */
2093 if (!(how & SEND_SHUTDOWN))
2094 return;
2095
2096 /* If we've already sent a FIN, or it's a closed state, skip this. */
2097 if ((1 << sk->sk_state) &
2098 (TCPF_ESTABLISHED | TCPF_SYN_SENT |
2099 TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) {
2100 /* Clear out any half completed packets. FIN if needed. */
2101 if (tcp_close_state(sk))
2102 tcp_send_fin(sk);
2103 }
2104}
2105EXPORT_SYMBOL(tcp_shutdown);
2106
2107bool tcp_check_oom(struct sock *sk, int shift)
2108{
2109 bool too_many_orphans, out_of_socket_memory;
2110
2111 too_many_orphans = tcp_too_many_orphans(sk, shift);
2112 out_of_socket_memory = tcp_out_of_memory(sk);
2113
2114 if (too_many_orphans)
2115 net_info_ratelimited("too many orphaned sockets\n");
2116 if (out_of_socket_memory)
2117 net_info_ratelimited("out of memory -- consider tuning tcp_mem\n");
2118 return too_many_orphans || out_of_socket_memory;
2119}
2120
2121void tcp_close(struct sock *sk, long timeout)
2122{
2123 struct sk_buff *skb;
2124 int data_was_unread = 0;
2125 int state;
2126
2127 lock_sock(sk);
2128 sk->sk_shutdown = SHUTDOWN_MASK;
2129
2130 if (sk->sk_state == TCP_LISTEN) {
2131 tcp_set_state(sk, TCP_CLOSE);
2132
2133 /* Special case. */
2134 inet_csk_listen_stop(sk);
2135
2136 goto adjudge_to_death;
2137 }
2138
2139 /* We need to flush the recv. buffs. We do this only on the
2140 * descriptor close, not protocol-sourced closes, because the
2141 * reader process may not have drained the data yet!
2142 */
2143 while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) {
2144 u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq -
2145 tcp_hdr(skb)->fin;
2146 data_was_unread += len;
2147 __kfree_skb(skb);
2148 }
2149
2150 sk_mem_reclaim(sk);
2151
2152 /* If socket has been already reset (e.g. in tcp_reset()) - kill it. */
2153 if (sk->sk_state == TCP_CLOSE)
2154 goto adjudge_to_death;
2155
2156 /* As outlined in RFC 2525, section 2.17, we send a RST here because
2157 * data was lost. To witness the awful effects of the old behavior of
2158 * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk
2159 * GET in an FTP client, suspend the process, wait for the client to
2160 * advertise a zero window, then kill -9 the FTP client, wheee...
2161 * Note: timeout is always zero in such a case.
2162 */
2163 if (unlikely(tcp_sk(sk)->repair)) {
2164 sk->sk_prot->disconnect(sk, 0);
2165 } else if (data_was_unread) {
2166 /* Unread data was tossed, zap the connection. */
2167 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONCLOSE);
2168 tcp_set_state(sk, TCP_CLOSE);
2169 tcp_send_active_reset(sk, sk->sk_allocation);
2170 } else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) {
2171 /* Check zero linger _after_ checking for unread data. */
2172 sk->sk_prot->disconnect(sk, 0);
2173 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
2174 } else if (tcp_close_state(sk)) {
2175 /* We FIN if the application ate all the data before
2176 * zapping the connection.
2177 */
2178
2179 /* RED-PEN. Formally speaking, we have broken TCP state
2180 * machine. State transitions:
2181 *
2182 * TCP_ESTABLISHED -> TCP_FIN_WAIT1
2183 * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible)
2184 * TCP_CLOSE_WAIT -> TCP_LAST_ACK
2185 *
2186 * are legal only when FIN has been sent (i.e. in window),
2187 * rather than queued out of window. Purists blame.
2188 *
2189 * F.e. "RFC state" is ESTABLISHED,
2190 * if Linux state is FIN-WAIT-1, but FIN is still not sent.
2191 *
2192 * The visible declinations are that sometimes
2193 * we enter time-wait state, when it is not required really
2194 * (harmless), do not send active resets, when they are
2195 * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
2196 * they look as CLOSING or LAST_ACK for Linux)
2197 * Probably, I missed some more holelets.
2198 * --ANK
2199 * XXX (TFO) - To start off we don't support SYN+ACK+FIN
2200 * in a single packet! (May consider it later but will
2201 * probably need API support or TCP_CORK SYN-ACK until
2202 * data is written and socket is closed.)
2203 */
2204 tcp_send_fin(sk);
2205 }
2206
2207 sk_stream_wait_close(sk, timeout);
2208
2209adjudge_to_death:
2210 state = sk->sk_state;
2211 sock_hold(sk);
2212 sock_orphan(sk);
2213
2214 /* It is the last release_sock in its life. It will remove backlog. */
2215 release_sock(sk);
2216
2217
2218 /* Now socket is owned by kernel and we acquire BH lock
2219 to finish close. No need to check for user refs.
2220 */
2221 local_bh_disable();
2222 bh_lock_sock(sk);
2223 WARN_ON(sock_owned_by_user(sk));
2224
2225 percpu_counter_inc(sk->sk_prot->orphan_count);
2226
2227 /* Have we already been destroyed by a softirq or backlog? */
2228 if (state != TCP_CLOSE && sk->sk_state == TCP_CLOSE)
2229 goto out;
2230
2231 /* This is a (useful) BSD violating of the RFC. There is a
2232 * problem with TCP as specified in that the other end could
2233 * keep a socket open forever with no application left this end.
2234 * We use a 1 minute timeout (about the same as BSD) then kill
2235 * our end. If they send after that then tough - BUT: long enough
2236 * that we won't make the old 4*rto = almost no time - whoops
2237 * reset mistake.
2238 *
2239 * Nope, it was not mistake. It is really desired behaviour
2240 * f.e. on http servers, when such sockets are useless, but
2241 * consume significant resources. Let's do it with special
2242 * linger2 option. --ANK
2243 */
2244
2245 if (sk->sk_state == TCP_FIN_WAIT2) {
2246 struct tcp_sock *tp = tcp_sk(sk);
2247 if (tp->linger2 < 0) {
2248 tcp_set_state(sk, TCP_CLOSE);
2249 tcp_send_active_reset(sk, GFP_ATOMIC);
2250 NET_INC_STATS_BH(sock_net(sk),
2251 LINUX_MIB_TCPABORTONLINGER);
2252 } else {
2253 const int tmo = tcp_fin_time(sk);
2254
2255 if (tmo > TCP_TIMEWAIT_LEN) {
2256 inet_csk_reset_keepalive_timer(sk,
2257 tmo - TCP_TIMEWAIT_LEN);
2258 } else {
2259 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
2260 goto out;
2261 }
2262 }
2263 }
2264 if (sk->sk_state != TCP_CLOSE) {
2265 sk_mem_reclaim(sk);
2266 if (tcp_check_oom(sk, 0)) {
2267 tcp_set_state(sk, TCP_CLOSE);
2268 tcp_send_active_reset(sk, GFP_ATOMIC);
2269 NET_INC_STATS_BH(sock_net(sk),
2270 LINUX_MIB_TCPABORTONMEMORY);
2271 }
2272 }
2273
2274 if (sk->sk_state == TCP_CLOSE) {
2275 struct request_sock *req = tcp_sk(sk)->fastopen_rsk;
2276 /* We could get here with a non-NULL req if the socket is
2277 * aborted (e.g., closed with unread data) before 3WHS
2278 * finishes.
2279 */
2280 if (req != NULL)
2281 reqsk_fastopen_remove(sk, req, false);
2282 inet_csk_destroy_sock(sk);
2283 }
2284 /* Otherwise, socket is reprieved until protocol close. */
2285
2286out:
2287 bh_unlock_sock(sk);
2288 local_bh_enable();
2289 sock_put(sk);
2290}
2291EXPORT_SYMBOL(tcp_close);
2292
2293/* These states need RST on ABORT according to RFC793 */
2294
2295static inline bool tcp_need_reset(int state)
2296{
2297 return (1 << state) &
2298 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 |
2299 TCPF_FIN_WAIT2 | TCPF_SYN_RECV);
2300}
2301
2302int tcp_disconnect(struct sock *sk, int flags)
2303{
2304 struct inet_sock *inet = inet_sk(sk);
2305 struct inet_connection_sock *icsk = inet_csk(sk);
2306 struct tcp_sock *tp = tcp_sk(sk);
2307 int err = 0;
2308 int old_state = sk->sk_state;
2309
2310 if (old_state != TCP_CLOSE)
2311 tcp_set_state(sk, TCP_CLOSE);
2312
2313 /* ABORT function of RFC793 */
2314 if (old_state == TCP_LISTEN) {
2315 inet_csk_listen_stop(sk);
2316 } else if (unlikely(tp->repair)) {
2317 sk->sk_err = ECONNABORTED;
2318 } else if (tcp_need_reset(old_state) ||
2319 (tp->snd_nxt != tp->write_seq &&
2320 (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK))) {
2321 /* The last check adjusts for discrepancy of Linux wrt. RFC
2322 * states
2323 */
2324 tcp_send_active_reset(sk, gfp_any());
2325 sk->sk_err = ECONNRESET;
2326 } else if (old_state == TCP_SYN_SENT)
2327 sk->sk_err = ECONNRESET;
2328
2329 tcp_clear_xmit_timers(sk);
2330 __skb_queue_purge(&sk->sk_receive_queue);
2331 tcp_write_queue_purge(sk);
2332 __skb_queue_purge(&tp->out_of_order_queue);
2333#ifdef CONFIG_NET_DMA
2334 __skb_queue_purge(&sk->sk_async_wait_queue);
2335#endif
2336
2337 inet->inet_dport = 0;
2338
2339 if (!(sk->sk_userlocks & SOCK_BINDADDR_LOCK))
2340 inet_reset_saddr(sk);
2341
2342 sk->sk_shutdown = 0;
2343 sock_reset_flag(sk, SOCK_DONE);
2344 tp->srtt_us = 0;
2345 if ((tp->write_seq += tp->max_window + 2) == 0)
2346 tp->write_seq = 1;
2347 icsk->icsk_backoff = 0;
2348 tp->snd_cwnd = 2;
2349 icsk->icsk_probes_out = 0;
2350 tp->packets_out = 0;
2351 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
2352 tp->snd_cwnd_cnt = 0;
2353 tp->window_clamp = 0;
2354 tcp_set_ca_state(sk, TCP_CA_Open);
2355 tcp_clear_retrans(tp);
2356 inet_csk_delack_init(sk);
2357 tcp_init_send_head(sk);
2358 memset(&tp->rx_opt, 0, sizeof(tp->rx_opt));
2359 __sk_dst_reset(sk);
2360
2361 WARN_ON(inet->inet_num && !icsk->icsk_bind_hash);
2362
2363 sk->sk_error_report(sk);
2364 return err;
2365}
2366EXPORT_SYMBOL(tcp_disconnect);
2367
2368void tcp_sock_destruct(struct sock *sk)
2369{
2370 inet_sock_destruct(sk);
2371
2372 kfree(inet_csk(sk)->icsk_accept_queue.fastopenq);
2373}
2374
2375static inline bool tcp_can_repair_sock(const struct sock *sk)
2376{
2377 return ns_capable(sock_net(sk)->user_ns, CAP_NET_ADMIN) &&
2378 ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_ESTABLISHED));
2379}
2380
2381static int tcp_repair_options_est(struct tcp_sock *tp,
2382 struct tcp_repair_opt __user *optbuf, unsigned int len)
2383{
2384 struct tcp_repair_opt opt;
2385
2386 while (len >= sizeof(opt)) {
2387 if (copy_from_user(&opt, optbuf, sizeof(opt)))
2388 return -EFAULT;
2389
2390 optbuf++;
2391 len -= sizeof(opt);
2392
2393 switch (opt.opt_code) {
2394 case TCPOPT_MSS:
2395 tp->rx_opt.mss_clamp = opt.opt_val;
2396 break;
2397 case TCPOPT_WINDOW:
2398 {
2399 u16 snd_wscale = opt.opt_val & 0xFFFF;
2400 u16 rcv_wscale = opt.opt_val >> 16;
2401
2402 if (snd_wscale > 14 || rcv_wscale > 14)
2403 return -EFBIG;
2404
2405 tp->rx_opt.snd_wscale = snd_wscale;
2406 tp->rx_opt.rcv_wscale = rcv_wscale;
2407 tp->rx_opt.wscale_ok = 1;
2408 }
2409 break;
2410 case TCPOPT_SACK_PERM:
2411 if (opt.opt_val != 0)
2412 return -EINVAL;
2413
2414 tp->rx_opt.sack_ok |= TCP_SACK_SEEN;
2415 if (sysctl_tcp_fack)
2416 tcp_enable_fack(tp);
2417 break;
2418 case TCPOPT_TIMESTAMP:
2419 if (opt.opt_val != 0)
2420 return -EINVAL;
2421
2422 tp->rx_opt.tstamp_ok = 1;
2423 break;
2424 }
2425 }
2426
2427 return 0;
2428}
2429
2430/*
2431 * Socket option code for TCP.
2432 */
2433static int do_tcp_setsockopt(struct sock *sk, int level,
2434 int optname, char __user *optval, unsigned int optlen)
2435{
2436 struct tcp_sock *tp = tcp_sk(sk);
2437 struct inet_connection_sock *icsk = inet_csk(sk);
2438 int val;
2439 int err = 0;
2440
2441 /* These are data/string values, all the others are ints */
2442 switch (optname) {
2443 case TCP_CONGESTION: {
2444 char name[TCP_CA_NAME_MAX];
2445
2446 if (optlen < 1)
2447 return -EINVAL;
2448
2449 val = strncpy_from_user(name, optval,
2450 min_t(long, TCP_CA_NAME_MAX-1, optlen));
2451 if (val < 0)
2452 return -EFAULT;
2453 name[val] = 0;
2454
2455 lock_sock(sk);
2456 err = tcp_set_congestion_control(sk, name);
2457 release_sock(sk);
2458 return err;
2459 }
2460 default:
2461 /* fallthru */
2462 break;
2463 }
2464
2465 if (optlen < sizeof(int))
2466 return -EINVAL;
2467
2468 if (get_user(val, (int __user *)optval))
2469 return -EFAULT;
2470
2471 lock_sock(sk);
2472
2473 switch (optname) {
2474 case TCP_MAXSEG:
2475 /* Values greater than interface MTU won't take effect. However
2476 * at the point when this call is done we typically don't yet
2477 * know which interface is going to be used */
2478 if (val < TCP_MIN_MSS || val > MAX_TCP_WINDOW) {
2479 err = -EINVAL;
2480 break;
2481 }
2482 tp->rx_opt.user_mss = val;
2483 break;
2484
2485 case TCP_NODELAY:
2486 if (val) {
2487 /* TCP_NODELAY is weaker than TCP_CORK, so that
2488 * this option on corked socket is remembered, but
2489 * it is not activated until cork is cleared.
2490 *
2491 * However, when TCP_NODELAY is set we make
2492 * an explicit push, which overrides even TCP_CORK
2493 * for currently queued segments.
2494 */
2495 tp->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH;
2496 tcp_push_pending_frames(sk);
2497 } else {
2498 tp->nonagle &= ~TCP_NAGLE_OFF;
2499 }
2500 break;
2501
2502 case TCP_THIN_LINEAR_TIMEOUTS:
2503 if (val < 0 || val > 1)
2504 err = -EINVAL;
2505 else
2506 tp->thin_lto = val;
2507 break;
2508
2509 case TCP_THIN_DUPACK:
2510 if (val < 0 || val > 1)
2511 err = -EINVAL;
2512 else {
2513 tp->thin_dupack = val;
2514 if (tp->thin_dupack)
2515 tcp_disable_early_retrans(tp);
2516 }
2517 break;
2518
2519 case TCP_REPAIR:
2520 if (!tcp_can_repair_sock(sk))
2521 err = -EPERM;
2522 else if (val == 1) {
2523 tp->repair = 1;
2524 sk->sk_reuse = SK_FORCE_REUSE;
2525 tp->repair_queue = TCP_NO_QUEUE;
2526 } else if (val == 0) {
2527 tp->repair = 0;
2528 sk->sk_reuse = SK_NO_REUSE;
2529 tcp_send_window_probe(sk);
2530 } else
2531 err = -EINVAL;
2532
2533 break;
2534
2535 case TCP_REPAIR_QUEUE:
2536 if (!tp->repair)
2537 err = -EPERM;
2538 else if (val < TCP_QUEUES_NR)
2539 tp->repair_queue = val;
2540 else
2541 err = -EINVAL;
2542 break;
2543
2544 case TCP_QUEUE_SEQ:
2545 if (sk->sk_state != TCP_CLOSE)
2546 err = -EPERM;
2547 else if (tp->repair_queue == TCP_SEND_QUEUE)
2548 tp->write_seq = val;
2549 else if (tp->repair_queue == TCP_RECV_QUEUE)
2550 tp->rcv_nxt = val;
2551 else
2552 err = -EINVAL;
2553 break;
2554
2555 case TCP_REPAIR_OPTIONS:
2556 if (!tp->repair)
2557 err = -EINVAL;
2558 else if (sk->sk_state == TCP_ESTABLISHED)
2559 err = tcp_repair_options_est(tp,
2560 (struct tcp_repair_opt __user *)optval,
2561 optlen);
2562 else
2563 err = -EPERM;
2564 break;
2565
2566 case TCP_CORK:
2567 /* When set indicates to always queue non-full frames.
2568 * Later the user clears this option and we transmit
2569 * any pending partial frames in the queue. This is
2570 * meant to be used alongside sendfile() to get properly
2571 * filled frames when the user (for example) must write
2572 * out headers with a write() call first and then use
2573 * sendfile to send out the data parts.
2574 *
2575 * TCP_CORK can be set together with TCP_NODELAY and it is
2576 * stronger than TCP_NODELAY.
2577 */
2578 if (val) {
2579 tp->nonagle |= TCP_NAGLE_CORK;
2580 } else {
2581 tp->nonagle &= ~TCP_NAGLE_CORK;
2582 if (tp->nonagle&TCP_NAGLE_OFF)
2583 tp->nonagle |= TCP_NAGLE_PUSH;
2584 tcp_push_pending_frames(sk);
2585 }
2586 break;
2587
2588 case TCP_KEEPIDLE:
2589 if (val < 1 || val > MAX_TCP_KEEPIDLE)
2590 err = -EINVAL;
2591 else {
2592 tp->keepalive_time = val * HZ;
2593 if (sock_flag(sk, SOCK_KEEPOPEN) &&
2594 !((1 << sk->sk_state) &
2595 (TCPF_CLOSE | TCPF_LISTEN))) {
2596 u32 elapsed = keepalive_time_elapsed(tp);
2597 if (tp->keepalive_time > elapsed)
2598 elapsed = tp->keepalive_time - elapsed;
2599 else
2600 elapsed = 0;
2601 inet_csk_reset_keepalive_timer(sk, elapsed);
2602 }
2603 }
2604 break;
2605 case TCP_KEEPINTVL:
2606 if (val < 1 || val > MAX_TCP_KEEPINTVL)
2607 err = -EINVAL;
2608 else
2609 tp->keepalive_intvl = val * HZ;
2610 break;
2611 case TCP_KEEPCNT:
2612 if (val < 1 || val > MAX_TCP_KEEPCNT)
2613 err = -EINVAL;
2614 else
2615 tp->keepalive_probes = val;
2616 break;
2617 case TCP_SYNCNT:
2618 if (val < 1 || val > MAX_TCP_SYNCNT)
2619 err = -EINVAL;
2620 else
2621 icsk->icsk_syn_retries = val;
2622 break;
2623
2624 case TCP_LINGER2:
2625 if (val < 0)
2626 tp->linger2 = -1;
2627 else if (val > sysctl_tcp_fin_timeout / HZ)
2628 tp->linger2 = 0;
2629 else
2630 tp->linger2 = val * HZ;
2631 break;
2632
2633 case TCP_DEFER_ACCEPT:
2634 /* Translate value in seconds to number of retransmits */
2635 icsk->icsk_accept_queue.rskq_defer_accept =
2636 secs_to_retrans(val, TCP_TIMEOUT_INIT / HZ,
2637 TCP_RTO_MAX / HZ);
2638 break;
2639
2640 case TCP_WINDOW_CLAMP:
2641 if (!val) {
2642 if (sk->sk_state != TCP_CLOSE) {
2643 err = -EINVAL;
2644 break;
2645 }
2646 tp->window_clamp = 0;
2647 } else
2648 tp->window_clamp = val < SOCK_MIN_RCVBUF / 2 ?
2649 SOCK_MIN_RCVBUF / 2 : val;
2650 break;
2651
2652 case TCP_QUICKACK:
2653 if (!val) {
2654 icsk->icsk_ack.pingpong = 1;
2655 } else {
2656 icsk->icsk_ack.pingpong = 0;
2657 if ((1 << sk->sk_state) &
2658 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) &&
2659 inet_csk_ack_scheduled(sk)) {
2660 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
2661 tcp_cleanup_rbuf(sk, 1);
2662 if (!(val & 1))
2663 icsk->icsk_ack.pingpong = 1;
2664 }
2665 }
2666 break;
2667
2668#ifdef CONFIG_TCP_MD5SIG
2669 case TCP_MD5SIG:
2670 /* Read the IP->Key mappings from userspace */
2671 err = tp->af_specific->md5_parse(sk, optval, optlen);
2672 break;
2673#endif
2674 case TCP_USER_TIMEOUT:
2675 /* Cap the max timeout in ms TCP will retry/retrans
2676 * before giving up and aborting (ETIMEDOUT) a connection.
2677 */
2678 if (val < 0)
2679 err = -EINVAL;
2680 else
2681 icsk->icsk_user_timeout = msecs_to_jiffies(val);
2682 break;
2683
2684 case TCP_FASTOPEN:
2685 if (val >= 0 && ((1 << sk->sk_state) & (TCPF_CLOSE |
2686 TCPF_LISTEN)))
2687 err = fastopen_init_queue(sk, val);
2688 else
2689 err = -EINVAL;
2690 break;
2691 case TCP_TIMESTAMP:
2692 if (!tp->repair)
2693 err = -EPERM;
2694 else
2695 tp->tsoffset = val - tcp_time_stamp;
2696 break;
2697 case TCP_NOTSENT_LOWAT:
2698 tp->notsent_lowat = val;
2699 sk->sk_write_space(sk);
2700 break;
2701 default:
2702 err = -ENOPROTOOPT;
2703 break;
2704 }
2705
2706 release_sock(sk);
2707 return err;
2708}
2709
2710int tcp_setsockopt(struct sock *sk, int level, int optname, char __user *optval,
2711 unsigned int optlen)
2712{
2713 const struct inet_connection_sock *icsk = inet_csk(sk);
2714
2715 if (level != SOL_TCP)
2716 return icsk->icsk_af_ops->setsockopt(sk, level, optname,
2717 optval, optlen);
2718 return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2719}
2720EXPORT_SYMBOL(tcp_setsockopt);
2721
2722#ifdef CONFIG_COMPAT
2723int compat_tcp_setsockopt(struct sock *sk, int level, int optname,
2724 char __user *optval, unsigned int optlen)
2725{
2726 if (level != SOL_TCP)
2727 return inet_csk_compat_setsockopt(sk, level, optname,
2728 optval, optlen);
2729 return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2730}
2731EXPORT_SYMBOL(compat_tcp_setsockopt);
2732#endif
2733
2734/* Return information about state of tcp endpoint in API format. */
2735void tcp_get_info(const struct sock *sk, struct tcp_info *info)
2736{
2737 const struct tcp_sock *tp = tcp_sk(sk);
2738 const struct inet_connection_sock *icsk = inet_csk(sk);
2739 u32 now = tcp_time_stamp;
2740
2741 memset(info, 0, sizeof(*info));
2742
2743 info->tcpi_state = sk->sk_state;
2744 info->tcpi_ca_state = icsk->icsk_ca_state;
2745 info->tcpi_retransmits = icsk->icsk_retransmits;
2746 info->tcpi_probes = icsk->icsk_probes_out;
2747 info->tcpi_backoff = icsk->icsk_backoff;
2748
2749 if (tp->rx_opt.tstamp_ok)
2750 info->tcpi_options |= TCPI_OPT_TIMESTAMPS;
2751 if (tcp_is_sack(tp))
2752 info->tcpi_options |= TCPI_OPT_SACK;
2753 if (tp->rx_opt.wscale_ok) {
2754 info->tcpi_options |= TCPI_OPT_WSCALE;
2755 info->tcpi_snd_wscale = tp->rx_opt.snd_wscale;
2756 info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale;
2757 }
2758
2759 if (tp->ecn_flags & TCP_ECN_OK)
2760 info->tcpi_options |= TCPI_OPT_ECN;
2761 if (tp->ecn_flags & TCP_ECN_SEEN)
2762 info->tcpi_options |= TCPI_OPT_ECN_SEEN;
2763 if (tp->syn_data_acked)
2764 info->tcpi_options |= TCPI_OPT_SYN_DATA;
2765
2766 info->tcpi_rto = jiffies_to_usecs(icsk->icsk_rto);
2767 info->tcpi_ato = jiffies_to_usecs(icsk->icsk_ack.ato);
2768 info->tcpi_snd_mss = tp->mss_cache;
2769 info->tcpi_rcv_mss = icsk->icsk_ack.rcv_mss;
2770
2771 if (sk->sk_state == TCP_LISTEN) {
2772 info->tcpi_unacked = sk->sk_ack_backlog;
2773 info->tcpi_sacked = sk->sk_max_ack_backlog;
2774 } else {
2775 info->tcpi_unacked = tp->packets_out;
2776 info->tcpi_sacked = tp->sacked_out;
2777 }
2778 info->tcpi_lost = tp->lost_out;
2779 info->tcpi_retrans = tp->retrans_out;
2780 info->tcpi_fackets = tp->fackets_out;
2781
2782 info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime);
2783 info->tcpi_last_data_recv = jiffies_to_msecs(now - icsk->icsk_ack.lrcvtime);
2784 info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp);
2785
2786 info->tcpi_pmtu = icsk->icsk_pmtu_cookie;
2787 info->tcpi_rcv_ssthresh = tp->rcv_ssthresh;
2788 info->tcpi_rtt = tp->srtt_us >> 3;
2789 info->tcpi_rttvar = tp->mdev_us >> 2;
2790 info->tcpi_snd_ssthresh = tp->snd_ssthresh;
2791 info->tcpi_snd_cwnd = tp->snd_cwnd;
2792 info->tcpi_advmss = tp->advmss;
2793 info->tcpi_reordering = tp->reordering;
2794
2795 info->tcpi_rcv_rtt = jiffies_to_usecs(tp->rcv_rtt_est.rtt)>>3;
2796 info->tcpi_rcv_space = tp->rcvq_space.space;
2797
2798 info->tcpi_total_retrans = tp->total_retrans;
2799
2800 info->tcpi_pacing_rate = sk->sk_pacing_rate != ~0U ?
2801 sk->sk_pacing_rate : ~0ULL;
2802 info->tcpi_max_pacing_rate = sk->sk_max_pacing_rate != ~0U ?
2803 sk->sk_max_pacing_rate : ~0ULL;
2804}
2805EXPORT_SYMBOL_GPL(tcp_get_info);
2806
2807static int do_tcp_getsockopt(struct sock *sk, int level,
2808 int optname, char __user *optval, int __user *optlen)
2809{
2810 struct inet_connection_sock *icsk = inet_csk(sk);
2811 struct tcp_sock *tp = tcp_sk(sk);
2812 int val, len;
2813
2814 if (get_user(len, optlen))
2815 return -EFAULT;
2816
2817 len = min_t(unsigned int, len, sizeof(int));
2818
2819 if (len < 0)
2820 return -EINVAL;
2821
2822 switch (optname) {
2823 case TCP_MAXSEG:
2824 val = tp->mss_cache;
2825 if (!val && ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
2826 val = tp->rx_opt.user_mss;
2827 if (tp->repair)
2828 val = tp->rx_opt.mss_clamp;
2829 break;
2830 case TCP_NODELAY:
2831 val = !!(tp->nonagle&TCP_NAGLE_OFF);
2832 break;
2833 case TCP_CORK:
2834 val = !!(tp->nonagle&TCP_NAGLE_CORK);
2835 break;
2836 case TCP_KEEPIDLE:
2837 val = keepalive_time_when(tp) / HZ;
2838 break;
2839 case TCP_KEEPINTVL:
2840 val = keepalive_intvl_when(tp) / HZ;
2841 break;
2842 case TCP_KEEPCNT:
2843 val = keepalive_probes(tp);
2844 break;
2845 case TCP_SYNCNT:
2846 val = icsk->icsk_syn_retries ? : sysctl_tcp_syn_retries;
2847 break;
2848 case TCP_LINGER2:
2849 val = tp->linger2;
2850 if (val >= 0)
2851 val = (val ? : sysctl_tcp_fin_timeout) / HZ;
2852 break;
2853 case TCP_DEFER_ACCEPT:
2854 val = retrans_to_secs(icsk->icsk_accept_queue.rskq_defer_accept,
2855 TCP_TIMEOUT_INIT / HZ, TCP_RTO_MAX / HZ);
2856 break;
2857 case TCP_WINDOW_CLAMP:
2858 val = tp->window_clamp;
2859 break;
2860 case TCP_INFO: {
2861 struct tcp_info info;
2862
2863 if (get_user(len, optlen))
2864 return -EFAULT;
2865
2866 tcp_get_info(sk, &info);
2867
2868 len = min_t(unsigned int, len, sizeof(info));
2869 if (put_user(len, optlen))
2870 return -EFAULT;
2871 if (copy_to_user(optval, &info, len))
2872 return -EFAULT;
2873 return 0;
2874 }
2875 case TCP_QUICKACK:
2876 val = !icsk->icsk_ack.pingpong;
2877 break;
2878
2879 case TCP_CONGESTION:
2880 if (get_user(len, optlen))
2881 return -EFAULT;
2882 len = min_t(unsigned int, len, TCP_CA_NAME_MAX);
2883 if (put_user(len, optlen))
2884 return -EFAULT;
2885 if (copy_to_user(optval, icsk->icsk_ca_ops->name, len))
2886 return -EFAULT;
2887 return 0;
2888
2889 case TCP_THIN_LINEAR_TIMEOUTS:
2890 val = tp->thin_lto;
2891 break;
2892 case TCP_THIN_DUPACK:
2893 val = tp->thin_dupack;
2894 break;
2895
2896 case TCP_REPAIR:
2897 val = tp->repair;
2898 break;
2899
2900 case TCP_REPAIR_QUEUE:
2901 if (tp->repair)
2902 val = tp->repair_queue;
2903 else
2904 return -EINVAL;
2905 break;
2906
2907 case TCP_QUEUE_SEQ:
2908 if (tp->repair_queue == TCP_SEND_QUEUE)
2909 val = tp->write_seq;
2910 else if (tp->repair_queue == TCP_RECV_QUEUE)
2911 val = tp->rcv_nxt;
2912 else
2913 return -EINVAL;
2914 break;
2915
2916 case TCP_USER_TIMEOUT:
2917 val = jiffies_to_msecs(icsk->icsk_user_timeout);
2918 break;
2919 case TCP_TIMESTAMP:
2920 val = tcp_time_stamp + tp->tsoffset;
2921 break;
2922 case TCP_NOTSENT_LOWAT:
2923 val = tp->notsent_lowat;
2924 break;
2925 default:
2926 return -ENOPROTOOPT;
2927 }
2928
2929 if (put_user(len, optlen))
2930 return -EFAULT;
2931 if (copy_to_user(optval, &val, len))
2932 return -EFAULT;
2933 return 0;
2934}
2935
2936int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval,
2937 int __user *optlen)
2938{
2939 struct inet_connection_sock *icsk = inet_csk(sk);
2940
2941 if (level != SOL_TCP)
2942 return icsk->icsk_af_ops->getsockopt(sk, level, optname,
2943 optval, optlen);
2944 return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2945}
2946EXPORT_SYMBOL(tcp_getsockopt);
2947
2948#ifdef CONFIG_COMPAT
2949int compat_tcp_getsockopt(struct sock *sk, int level, int optname,
2950 char __user *optval, int __user *optlen)
2951{
2952 if (level != SOL_TCP)
2953 return inet_csk_compat_getsockopt(sk, level, optname,
2954 optval, optlen);
2955 return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2956}
2957EXPORT_SYMBOL(compat_tcp_getsockopt);
2958#endif
2959
2960#ifdef CONFIG_TCP_MD5SIG
2961static struct tcp_md5sig_pool __percpu *tcp_md5sig_pool __read_mostly;
2962static DEFINE_MUTEX(tcp_md5sig_mutex);
2963
2964static void __tcp_free_md5sig_pool(struct tcp_md5sig_pool __percpu *pool)
2965{
2966 int cpu;
2967
2968 for_each_possible_cpu(cpu) {
2969 struct tcp_md5sig_pool *p = per_cpu_ptr(pool, cpu);
2970
2971 if (p->md5_desc.tfm)
2972 crypto_free_hash(p->md5_desc.tfm);
2973 }
2974 free_percpu(pool);
2975}
2976
2977static void __tcp_alloc_md5sig_pool(void)
2978{
2979 int cpu;
2980 struct tcp_md5sig_pool __percpu *pool;
2981
2982 pool = alloc_percpu(struct tcp_md5sig_pool);
2983 if (!pool)
2984 return;
2985
2986 for_each_possible_cpu(cpu) {
2987 struct crypto_hash *hash;
2988
2989 hash = crypto_alloc_hash("md5", 0, CRYPTO_ALG_ASYNC);
2990 if (IS_ERR_OR_NULL(hash))
2991 goto out_free;
2992
2993 per_cpu_ptr(pool, cpu)->md5_desc.tfm = hash;
2994 }
2995 /* before setting tcp_md5sig_pool, we must commit all writes
2996 * to memory. See ACCESS_ONCE() in tcp_get_md5sig_pool()
2997 */
2998 smp_wmb();
2999 tcp_md5sig_pool = pool;
3000 return;
3001out_free:
3002 __tcp_free_md5sig_pool(pool);
3003}
3004
3005bool tcp_alloc_md5sig_pool(void)
3006{
3007 if (unlikely(!tcp_md5sig_pool)) {
3008 mutex_lock(&tcp_md5sig_mutex);
3009
3010 if (!tcp_md5sig_pool)
3011 __tcp_alloc_md5sig_pool();
3012
3013 mutex_unlock(&tcp_md5sig_mutex);
3014 }
3015 return tcp_md5sig_pool != NULL;
3016}
3017EXPORT_SYMBOL(tcp_alloc_md5sig_pool);
3018
3019
3020/**
3021 * tcp_get_md5sig_pool - get md5sig_pool for this user
3022 *
3023 * We use percpu structure, so if we succeed, we exit with preemption
3024 * and BH disabled, to make sure another thread or softirq handling
3025 * wont try to get same context.
3026 */
3027struct tcp_md5sig_pool *tcp_get_md5sig_pool(void)
3028{
3029 struct tcp_md5sig_pool __percpu *p;
3030
3031 local_bh_disable();
3032 p = ACCESS_ONCE(tcp_md5sig_pool);
3033 if (p)
3034 return __this_cpu_ptr(p);
3035
3036 local_bh_enable();
3037 return NULL;
3038}
3039EXPORT_SYMBOL(tcp_get_md5sig_pool);
3040
3041int tcp_md5_hash_header(struct tcp_md5sig_pool *hp,
3042 const struct tcphdr *th)
3043{
3044 struct scatterlist sg;
3045 struct tcphdr hdr;
3046 int err;
3047
3048 /* We are not allowed to change tcphdr, make a local copy */
3049 memcpy(&hdr, th, sizeof(hdr));
3050 hdr.check = 0;
3051
3052 /* options aren't included in the hash */
3053 sg_init_one(&sg, &hdr, sizeof(hdr));
3054 err = crypto_hash_update(&hp->md5_desc, &sg, sizeof(hdr));
3055 return err;
3056}
3057EXPORT_SYMBOL(tcp_md5_hash_header);
3058
3059int tcp_md5_hash_skb_data(struct tcp_md5sig_pool *hp,
3060 const struct sk_buff *skb, unsigned int header_len)
3061{
3062 struct scatterlist sg;
3063 const struct tcphdr *tp = tcp_hdr(skb);
3064 struct hash_desc *desc = &hp->md5_desc;
3065 unsigned int i;
3066 const unsigned int head_data_len = skb_headlen(skb) > header_len ?
3067 skb_headlen(skb) - header_len : 0;
3068 const struct skb_shared_info *shi = skb_shinfo(skb);
3069 struct sk_buff *frag_iter;
3070
3071 sg_init_table(&sg, 1);
3072
3073 sg_set_buf(&sg, ((u8 *) tp) + header_len, head_data_len);
3074 if (crypto_hash_update(desc, &sg, head_data_len))
3075 return 1;
3076
3077 for (i = 0; i < shi->nr_frags; ++i) {
3078 const struct skb_frag_struct *f = &shi->frags[i];
3079 unsigned int offset = f->page_offset;
3080 struct page *page = skb_frag_page(f) + (offset >> PAGE_SHIFT);
3081
3082 sg_set_page(&sg, page, skb_frag_size(f),
3083 offset_in_page(offset));
3084 if (crypto_hash_update(desc, &sg, skb_frag_size(f)))
3085 return 1;
3086 }
3087
3088 skb_walk_frags(skb, frag_iter)
3089 if (tcp_md5_hash_skb_data(hp, frag_iter, 0))
3090 return 1;
3091
3092 return 0;
3093}
3094EXPORT_SYMBOL(tcp_md5_hash_skb_data);
3095
3096int tcp_md5_hash_key(struct tcp_md5sig_pool *hp, const struct tcp_md5sig_key *key)
3097{
3098 struct scatterlist sg;
3099
3100 sg_init_one(&sg, key->key, key->keylen);
3101 return crypto_hash_update(&hp->md5_desc, &sg, key->keylen);
3102}
3103EXPORT_SYMBOL(tcp_md5_hash_key);
3104
3105#endif
3106
3107void tcp_done(struct sock *sk)
3108{
3109 struct request_sock *req = tcp_sk(sk)->fastopen_rsk;
3110
3111 if (sk->sk_state == TCP_SYN_SENT || sk->sk_state == TCP_SYN_RECV)
3112 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_ATTEMPTFAILS);
3113
3114 tcp_set_state(sk, TCP_CLOSE);
3115 tcp_clear_xmit_timers(sk);
3116 if (req != NULL)
3117 reqsk_fastopen_remove(sk, req, false);
3118
3119 sk->sk_shutdown = SHUTDOWN_MASK;
3120
3121 if (!sock_flag(sk, SOCK_DEAD))
3122 sk->sk_state_change(sk);
3123 else
3124 inet_csk_destroy_sock(sk);
3125}
3126EXPORT_SYMBOL_GPL(tcp_done);
3127
3128extern struct tcp_congestion_ops tcp_reno;
3129
3130static __initdata unsigned long thash_entries;
3131static int __init set_thash_entries(char *str)
3132{
3133 ssize_t ret;
3134
3135 if (!str)
3136 return 0;
3137
3138 ret = kstrtoul(str, 0, &thash_entries);
3139 if (ret)
3140 return 0;
3141
3142 return 1;
3143}
3144__setup("thash_entries=", set_thash_entries);
3145
3146static void tcp_init_mem(void)
3147{
3148 unsigned long limit = nr_free_buffer_pages() / 8;
3149 limit = max(limit, 128UL);
3150 sysctl_tcp_mem[0] = limit / 4 * 3;
3151 sysctl_tcp_mem[1] = limit;
3152 sysctl_tcp_mem[2] = sysctl_tcp_mem[0] * 2;
3153}
3154
3155void __init tcp_init(void)
3156{
3157 struct sk_buff *skb = NULL;
3158 unsigned long limit;
3159 int max_rshare, max_wshare, cnt;
3160 unsigned int i;
3161
3162 BUILD_BUG_ON(sizeof(struct tcp_skb_cb) > sizeof(skb->cb));
3163
3164 percpu_counter_init(&tcp_sockets_allocated, 0);
3165 percpu_counter_init(&tcp_orphan_count, 0);
3166 tcp_hashinfo.bind_bucket_cachep =
3167 kmem_cache_create("tcp_bind_bucket",
3168 sizeof(struct inet_bind_bucket), 0,
3169 SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL);
3170
3171 /* Size and allocate the main established and bind bucket
3172 * hash tables.
3173 *
3174 * The methodology is similar to that of the buffer cache.
3175 */
3176 tcp_hashinfo.ehash =
3177 alloc_large_system_hash("TCP established",
3178 sizeof(struct inet_ehash_bucket),
3179 thash_entries,
3180 17, /* one slot per 128 KB of memory */
3181 0,
3182 NULL,
3183 &tcp_hashinfo.ehash_mask,
3184 0,
3185 thash_entries ? 0 : 512 * 1024);
3186 for (i = 0; i <= tcp_hashinfo.ehash_mask; i++)
3187 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].chain, i);
3188
3189 if (inet_ehash_locks_alloc(&tcp_hashinfo))
3190 panic("TCP: failed to alloc ehash_locks");
3191 tcp_hashinfo.bhash =
3192 alloc_large_system_hash("TCP bind",
3193 sizeof(struct inet_bind_hashbucket),
3194 tcp_hashinfo.ehash_mask + 1,
3195 17, /* one slot per 128 KB of memory */
3196 0,
3197 &tcp_hashinfo.bhash_size,
3198 NULL,
3199 0,
3200 64 * 1024);
3201 tcp_hashinfo.bhash_size = 1U << tcp_hashinfo.bhash_size;
3202 for (i = 0; i < tcp_hashinfo.bhash_size; i++) {
3203 spin_lock_init(&tcp_hashinfo.bhash[i].lock);
3204 INIT_HLIST_HEAD(&tcp_hashinfo.bhash[i].chain);
3205 }
3206
3207
3208 cnt = tcp_hashinfo.ehash_mask + 1;
3209
3210 tcp_death_row.sysctl_max_tw_buckets = cnt / 2;
3211 sysctl_tcp_max_orphans = cnt / 2;
3212 sysctl_max_syn_backlog = max(128, cnt / 256);
3213
3214 tcp_init_mem();
3215 /* Set per-socket limits to no more than 1/128 the pressure threshold */
3216 limit = nr_free_buffer_pages() << (PAGE_SHIFT - 7);
3217 max_wshare = min(4UL*1024*1024, limit);
3218 max_rshare = min(6UL*1024*1024, limit);
3219
3220 sysctl_tcp_wmem[0] = SK_MEM_QUANTUM;
3221 sysctl_tcp_wmem[1] = 16*1024;
3222 sysctl_tcp_wmem[2] = max(64*1024, max_wshare);
3223
3224 sysctl_tcp_rmem[0] = SK_MEM_QUANTUM;
3225 sysctl_tcp_rmem[1] = 87380;
3226 sysctl_tcp_rmem[2] = max(87380, max_rshare);
3227
3228 pr_info("Hash tables configured (established %u bind %u)\n",
3229 tcp_hashinfo.ehash_mask + 1, tcp_hashinfo.bhash_size);
3230
3231 tcp_metrics_init();
3232
3233 tcp_register_congestion_control(&tcp_reno);
3234
3235 tcp_tasklet_init();
3236}
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 <crypto/hash.h>
251#include <linux/kernel.h>
252#include <linux/module.h>
253#include <linux/types.h>
254#include <linux/fcntl.h>
255#include <linux/poll.h>
256#include <linux/inet_diag.h>
257#include <linux/init.h>
258#include <linux/fs.h>
259#include <linux/skbuff.h>
260#include <linux/scatterlist.h>
261#include <linux/splice.h>
262#include <linux/net.h>
263#include <linux/socket.h>
264#include <linux/random.h>
265#include <linux/bootmem.h>
266#include <linux/highmem.h>
267#include <linux/swap.h>
268#include <linux/cache.h>
269#include <linux/err.h>
270#include <linux/time.h>
271#include <linux/slab.h>
272#include <linux/errqueue.h>
273#include <linux/static_key.h>
274
275#include <net/icmp.h>
276#include <net/inet_common.h>
277#include <net/tcp.h>
278#include <net/xfrm.h>
279#include <net/ip.h>
280#include <net/sock.h>
281
282#include <linux/uaccess.h>
283#include <asm/ioctls.h>
284#include <net/busy_poll.h>
285
286struct percpu_counter tcp_orphan_count;
287EXPORT_SYMBOL_GPL(tcp_orphan_count);
288
289long sysctl_tcp_mem[3] __read_mostly;
290EXPORT_SYMBOL(sysctl_tcp_mem);
291
292atomic_long_t tcp_memory_allocated; /* Current allocated memory. */
293EXPORT_SYMBOL(tcp_memory_allocated);
294
295#if IS_ENABLED(CONFIG_SMC)
296DEFINE_STATIC_KEY_FALSE(tcp_have_smc);
297EXPORT_SYMBOL(tcp_have_smc);
298#endif
299
300/*
301 * Current number of TCP sockets.
302 */
303struct percpu_counter tcp_sockets_allocated;
304EXPORT_SYMBOL(tcp_sockets_allocated);
305
306/*
307 * TCP splice context
308 */
309struct tcp_splice_state {
310 struct pipe_inode_info *pipe;
311 size_t len;
312 unsigned int flags;
313};
314
315/*
316 * Pressure flag: try to collapse.
317 * Technical note: it is used by multiple contexts non atomically.
318 * All the __sk_mem_schedule() is of this nature: accounting
319 * is strict, actions are advisory and have some latency.
320 */
321unsigned long tcp_memory_pressure __read_mostly;
322EXPORT_SYMBOL_GPL(tcp_memory_pressure);
323
324void tcp_enter_memory_pressure(struct sock *sk)
325{
326 unsigned long val;
327
328 if (tcp_memory_pressure)
329 return;
330 val = jiffies;
331
332 if (!val)
333 val--;
334 if (!cmpxchg(&tcp_memory_pressure, 0, val))
335 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMEMORYPRESSURES);
336}
337EXPORT_SYMBOL_GPL(tcp_enter_memory_pressure);
338
339void tcp_leave_memory_pressure(struct sock *sk)
340{
341 unsigned long val;
342
343 if (!tcp_memory_pressure)
344 return;
345 val = xchg(&tcp_memory_pressure, 0);
346 if (val)
347 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPMEMORYPRESSURESCHRONO,
348 jiffies_to_msecs(jiffies - val));
349}
350EXPORT_SYMBOL_GPL(tcp_leave_memory_pressure);
351
352/* Convert seconds to retransmits based on initial and max timeout */
353static u8 secs_to_retrans(int seconds, int timeout, int rto_max)
354{
355 u8 res = 0;
356
357 if (seconds > 0) {
358 int period = timeout;
359
360 res = 1;
361 while (seconds > period && res < 255) {
362 res++;
363 timeout <<= 1;
364 if (timeout > rto_max)
365 timeout = rto_max;
366 period += timeout;
367 }
368 }
369 return res;
370}
371
372/* Convert retransmits to seconds based on initial and max timeout */
373static int retrans_to_secs(u8 retrans, int timeout, int rto_max)
374{
375 int period = 0;
376
377 if (retrans > 0) {
378 period = timeout;
379 while (--retrans) {
380 timeout <<= 1;
381 if (timeout > rto_max)
382 timeout = rto_max;
383 period += timeout;
384 }
385 }
386 return period;
387}
388
389static u64 tcp_compute_delivery_rate(const struct tcp_sock *tp)
390{
391 u32 rate = READ_ONCE(tp->rate_delivered);
392 u32 intv = READ_ONCE(tp->rate_interval_us);
393 u64 rate64 = 0;
394
395 if (rate && intv) {
396 rate64 = (u64)rate * tp->mss_cache * USEC_PER_SEC;
397 do_div(rate64, intv);
398 }
399 return rate64;
400}
401
402/* Address-family independent initialization for a tcp_sock.
403 *
404 * NOTE: A lot of things set to zero explicitly by call to
405 * sk_alloc() so need not be done here.
406 */
407void tcp_init_sock(struct sock *sk)
408{
409 struct inet_connection_sock *icsk = inet_csk(sk);
410 struct tcp_sock *tp = tcp_sk(sk);
411
412 tp->out_of_order_queue = RB_ROOT;
413 sk->tcp_rtx_queue = RB_ROOT;
414 tcp_init_xmit_timers(sk);
415 INIT_LIST_HEAD(&tp->tsq_node);
416 INIT_LIST_HEAD(&tp->tsorted_sent_queue);
417
418 icsk->icsk_rto = TCP_TIMEOUT_INIT;
419 tp->mdev_us = jiffies_to_usecs(TCP_TIMEOUT_INIT);
420 minmax_reset(&tp->rtt_min, tcp_jiffies32, ~0U);
421
422 /* So many TCP implementations out there (incorrectly) count the
423 * initial SYN frame in their delayed-ACK and congestion control
424 * algorithms that we must have the following bandaid to talk
425 * efficiently to them. -DaveM
426 */
427 tp->snd_cwnd = TCP_INIT_CWND;
428
429 /* There's a bubble in the pipe until at least the first ACK. */
430 tp->app_limited = ~0U;
431
432 /* See draft-stevens-tcpca-spec-01 for discussion of the
433 * initialization of these values.
434 */
435 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
436 tp->snd_cwnd_clamp = ~0;
437 tp->mss_cache = TCP_MSS_DEFAULT;
438
439 tp->reordering = sock_net(sk)->ipv4.sysctl_tcp_reordering;
440 tcp_assign_congestion_control(sk);
441
442 tp->tsoffset = 0;
443 tp->rack.reo_wnd_steps = 1;
444
445 sk->sk_state = TCP_CLOSE;
446
447 sk->sk_write_space = sk_stream_write_space;
448 sock_set_flag(sk, SOCK_USE_WRITE_QUEUE);
449
450 icsk->icsk_sync_mss = tcp_sync_mss;
451
452 sk->sk_sndbuf = sock_net(sk)->ipv4.sysctl_tcp_wmem[1];
453 sk->sk_rcvbuf = sock_net(sk)->ipv4.sysctl_tcp_rmem[1];
454
455 sk_sockets_allocated_inc(sk);
456 sk->sk_route_forced_caps = NETIF_F_GSO;
457}
458EXPORT_SYMBOL(tcp_init_sock);
459
460void tcp_init_transfer(struct sock *sk, int bpf_op)
461{
462 struct inet_connection_sock *icsk = inet_csk(sk);
463
464 tcp_mtup_init(sk);
465 icsk->icsk_af_ops->rebuild_header(sk);
466 tcp_init_metrics(sk);
467 tcp_call_bpf(sk, bpf_op, 0, NULL);
468 tcp_init_congestion_control(sk);
469 tcp_init_buffer_space(sk);
470}
471
472static void tcp_tx_timestamp(struct sock *sk, u16 tsflags)
473{
474 struct sk_buff *skb = tcp_write_queue_tail(sk);
475
476 if (tsflags && skb) {
477 struct skb_shared_info *shinfo = skb_shinfo(skb);
478 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
479
480 sock_tx_timestamp(sk, tsflags, &shinfo->tx_flags);
481 if (tsflags & SOF_TIMESTAMPING_TX_ACK)
482 tcb->txstamp_ack = 1;
483 if (tsflags & SOF_TIMESTAMPING_TX_RECORD_MASK)
484 shinfo->tskey = TCP_SKB_CB(skb)->seq + skb->len - 1;
485 }
486}
487
488static inline bool tcp_stream_is_readable(const struct tcp_sock *tp,
489 int target, struct sock *sk)
490{
491 return (tp->rcv_nxt - tp->copied_seq >= target) ||
492 (sk->sk_prot->stream_memory_read ?
493 sk->sk_prot->stream_memory_read(sk) : false);
494}
495
496/*
497 * Wait for a TCP event.
498 *
499 * Note that we don't need to lock the socket, as the upper poll layers
500 * take care of normal races (between the test and the event) and we don't
501 * go look at any of the socket buffers directly.
502 */
503__poll_t tcp_poll(struct file *file, struct socket *sock, poll_table *wait)
504{
505 __poll_t mask;
506 struct sock *sk = sock->sk;
507 const struct tcp_sock *tp = tcp_sk(sk);
508 int state;
509
510 sock_poll_wait(file, sk_sleep(sk), wait);
511
512 state = inet_sk_state_load(sk);
513 if (state == TCP_LISTEN)
514 return inet_csk_listen_poll(sk);
515
516 /* Socket is not locked. We are protected from async events
517 * by poll logic and correct handling of state changes
518 * made by other threads is impossible in any case.
519 */
520
521 mask = 0;
522
523 /*
524 * EPOLLHUP is certainly not done right. But poll() doesn't
525 * have a notion of HUP in just one direction, and for a
526 * socket the read side is more interesting.
527 *
528 * Some poll() documentation says that EPOLLHUP is incompatible
529 * with the EPOLLOUT/POLLWR flags, so somebody should check this
530 * all. But careful, it tends to be safer to return too many
531 * bits than too few, and you can easily break real applications
532 * if you don't tell them that something has hung up!
533 *
534 * Check-me.
535 *
536 * Check number 1. EPOLLHUP is _UNMASKABLE_ event (see UNIX98 and
537 * our fs/select.c). It means that after we received EOF,
538 * poll always returns immediately, making impossible poll() on write()
539 * in state CLOSE_WAIT. One solution is evident --- to set EPOLLHUP
540 * if and only if shutdown has been made in both directions.
541 * Actually, it is interesting to look how Solaris and DUX
542 * solve this dilemma. I would prefer, if EPOLLHUP were maskable,
543 * then we could set it on SND_SHUTDOWN. BTW examples given
544 * in Stevens' books assume exactly this behaviour, it explains
545 * why EPOLLHUP is incompatible with EPOLLOUT. --ANK
546 *
547 * NOTE. Check for TCP_CLOSE is added. The goal is to prevent
548 * blocking on fresh not-connected or disconnected socket. --ANK
549 */
550 if (sk->sk_shutdown == SHUTDOWN_MASK || state == TCP_CLOSE)
551 mask |= EPOLLHUP;
552 if (sk->sk_shutdown & RCV_SHUTDOWN)
553 mask |= EPOLLIN | EPOLLRDNORM | EPOLLRDHUP;
554
555 /* Connected or passive Fast Open socket? */
556 if (state != TCP_SYN_SENT &&
557 (state != TCP_SYN_RECV || tp->fastopen_rsk)) {
558 int target = sock_rcvlowat(sk, 0, INT_MAX);
559
560 if (tp->urg_seq == tp->copied_seq &&
561 !sock_flag(sk, SOCK_URGINLINE) &&
562 tp->urg_data)
563 target++;
564
565 if (tcp_stream_is_readable(tp, target, sk))
566 mask |= EPOLLIN | EPOLLRDNORM;
567
568 if (!(sk->sk_shutdown & SEND_SHUTDOWN)) {
569 if (sk_stream_is_writeable(sk)) {
570 mask |= EPOLLOUT | EPOLLWRNORM;
571 } else { /* send SIGIO later */
572 sk_set_bit(SOCKWQ_ASYNC_NOSPACE, sk);
573 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
574
575 /* Race breaker. If space is freed after
576 * wspace test but before the flags are set,
577 * IO signal will be lost. Memory barrier
578 * pairs with the input side.
579 */
580 smp_mb__after_atomic();
581 if (sk_stream_is_writeable(sk))
582 mask |= EPOLLOUT | EPOLLWRNORM;
583 }
584 } else
585 mask |= EPOLLOUT | EPOLLWRNORM;
586
587 if (tp->urg_data & TCP_URG_VALID)
588 mask |= EPOLLPRI;
589 } else if (state == TCP_SYN_SENT && inet_sk(sk)->defer_connect) {
590 /* Active TCP fastopen socket with defer_connect
591 * Return EPOLLOUT so application can call write()
592 * in order for kernel to generate SYN+data
593 */
594 mask |= EPOLLOUT | EPOLLWRNORM;
595 }
596 /* This barrier is coupled with smp_wmb() in tcp_reset() */
597 smp_rmb();
598 if (sk->sk_err || !skb_queue_empty(&sk->sk_error_queue))
599 mask |= EPOLLERR;
600
601 return mask;
602}
603EXPORT_SYMBOL(tcp_poll);
604
605int tcp_ioctl(struct sock *sk, int cmd, unsigned long arg)
606{
607 struct tcp_sock *tp = tcp_sk(sk);
608 int answ;
609 bool slow;
610
611 switch (cmd) {
612 case SIOCINQ:
613 if (sk->sk_state == TCP_LISTEN)
614 return -EINVAL;
615
616 slow = lock_sock_fast(sk);
617 answ = tcp_inq(sk);
618 unlock_sock_fast(sk, slow);
619 break;
620 case SIOCATMARK:
621 answ = tp->urg_data && tp->urg_seq == tp->copied_seq;
622 break;
623 case SIOCOUTQ:
624 if (sk->sk_state == TCP_LISTEN)
625 return -EINVAL;
626
627 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
628 answ = 0;
629 else
630 answ = tp->write_seq - tp->snd_una;
631 break;
632 case SIOCOUTQNSD:
633 if (sk->sk_state == TCP_LISTEN)
634 return -EINVAL;
635
636 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
637 answ = 0;
638 else
639 answ = tp->write_seq - tp->snd_nxt;
640 break;
641 default:
642 return -ENOIOCTLCMD;
643 }
644
645 return put_user(answ, (int __user *)arg);
646}
647EXPORT_SYMBOL(tcp_ioctl);
648
649static inline void tcp_mark_push(struct tcp_sock *tp, struct sk_buff *skb)
650{
651 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
652 tp->pushed_seq = tp->write_seq;
653}
654
655static inline bool forced_push(const struct tcp_sock *tp)
656{
657 return after(tp->write_seq, tp->pushed_seq + (tp->max_window >> 1));
658}
659
660static void skb_entail(struct sock *sk, struct sk_buff *skb)
661{
662 struct tcp_sock *tp = tcp_sk(sk);
663 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
664
665 skb->csum = 0;
666 tcb->seq = tcb->end_seq = tp->write_seq;
667 tcb->tcp_flags = TCPHDR_ACK;
668 tcb->sacked = 0;
669 __skb_header_release(skb);
670 tcp_add_write_queue_tail(sk, skb);
671 sk->sk_wmem_queued += skb->truesize;
672 sk_mem_charge(sk, skb->truesize);
673 if (tp->nonagle & TCP_NAGLE_PUSH)
674 tp->nonagle &= ~TCP_NAGLE_PUSH;
675
676 tcp_slow_start_after_idle_check(sk);
677}
678
679static inline void tcp_mark_urg(struct tcp_sock *tp, int flags)
680{
681 if (flags & MSG_OOB)
682 tp->snd_up = tp->write_seq;
683}
684
685/* If a not yet filled skb is pushed, do not send it if
686 * we have data packets in Qdisc or NIC queues :
687 * Because TX completion will happen shortly, it gives a chance
688 * to coalesce future sendmsg() payload into this skb, without
689 * need for a timer, and with no latency trade off.
690 * As packets containing data payload have a bigger truesize
691 * than pure acks (dataless) packets, the last checks prevent
692 * autocorking if we only have an ACK in Qdisc/NIC queues,
693 * or if TX completion was delayed after we processed ACK packet.
694 */
695static bool tcp_should_autocork(struct sock *sk, struct sk_buff *skb,
696 int size_goal)
697{
698 return skb->len < size_goal &&
699 sock_net(sk)->ipv4.sysctl_tcp_autocorking &&
700 !tcp_rtx_queue_empty(sk) &&
701 refcount_read(&sk->sk_wmem_alloc) > skb->truesize;
702}
703
704static void tcp_push(struct sock *sk, int flags, int mss_now,
705 int nonagle, int size_goal)
706{
707 struct tcp_sock *tp = tcp_sk(sk);
708 struct sk_buff *skb;
709
710 skb = tcp_write_queue_tail(sk);
711 if (!skb)
712 return;
713 if (!(flags & MSG_MORE) || forced_push(tp))
714 tcp_mark_push(tp, skb);
715
716 tcp_mark_urg(tp, flags);
717
718 if (tcp_should_autocork(sk, skb, size_goal)) {
719
720 /* avoid atomic op if TSQ_THROTTLED bit is already set */
721 if (!test_bit(TSQ_THROTTLED, &sk->sk_tsq_flags)) {
722 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPAUTOCORKING);
723 set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags);
724 }
725 /* It is possible TX completion already happened
726 * before we set TSQ_THROTTLED.
727 */
728 if (refcount_read(&sk->sk_wmem_alloc) > skb->truesize)
729 return;
730 }
731
732 if (flags & MSG_MORE)
733 nonagle = TCP_NAGLE_CORK;
734
735 __tcp_push_pending_frames(sk, mss_now, nonagle);
736}
737
738static int tcp_splice_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb,
739 unsigned int offset, size_t len)
740{
741 struct tcp_splice_state *tss = rd_desc->arg.data;
742 int ret;
743
744 ret = skb_splice_bits(skb, skb->sk, offset, tss->pipe,
745 min(rd_desc->count, len), tss->flags);
746 if (ret > 0)
747 rd_desc->count -= ret;
748 return ret;
749}
750
751static int __tcp_splice_read(struct sock *sk, struct tcp_splice_state *tss)
752{
753 /* Store TCP splice context information in read_descriptor_t. */
754 read_descriptor_t rd_desc = {
755 .arg.data = tss,
756 .count = tss->len,
757 };
758
759 return tcp_read_sock(sk, &rd_desc, tcp_splice_data_recv);
760}
761
762/**
763 * tcp_splice_read - splice data from TCP socket to a pipe
764 * @sock: socket to splice from
765 * @ppos: position (not valid)
766 * @pipe: pipe to splice to
767 * @len: number of bytes to splice
768 * @flags: splice modifier flags
769 *
770 * Description:
771 * Will read pages from given socket and fill them into a pipe.
772 *
773 **/
774ssize_t tcp_splice_read(struct socket *sock, loff_t *ppos,
775 struct pipe_inode_info *pipe, size_t len,
776 unsigned int flags)
777{
778 struct sock *sk = sock->sk;
779 struct tcp_splice_state tss = {
780 .pipe = pipe,
781 .len = len,
782 .flags = flags,
783 };
784 long timeo;
785 ssize_t spliced;
786 int ret;
787
788 sock_rps_record_flow(sk);
789 /*
790 * We can't seek on a socket input
791 */
792 if (unlikely(*ppos))
793 return -ESPIPE;
794
795 ret = spliced = 0;
796
797 lock_sock(sk);
798
799 timeo = sock_rcvtimeo(sk, sock->file->f_flags & O_NONBLOCK);
800 while (tss.len) {
801 ret = __tcp_splice_read(sk, &tss);
802 if (ret < 0)
803 break;
804 else if (!ret) {
805 if (spliced)
806 break;
807 if (sock_flag(sk, SOCK_DONE))
808 break;
809 if (sk->sk_err) {
810 ret = sock_error(sk);
811 break;
812 }
813 if (sk->sk_shutdown & RCV_SHUTDOWN)
814 break;
815 if (sk->sk_state == TCP_CLOSE) {
816 /*
817 * This occurs when user tries to read
818 * from never connected socket.
819 */
820 if (!sock_flag(sk, SOCK_DONE))
821 ret = -ENOTCONN;
822 break;
823 }
824 if (!timeo) {
825 ret = -EAGAIN;
826 break;
827 }
828 /* if __tcp_splice_read() got nothing while we have
829 * an skb in receive queue, we do not want to loop.
830 * This might happen with URG data.
831 */
832 if (!skb_queue_empty(&sk->sk_receive_queue))
833 break;
834 sk_wait_data(sk, &timeo, NULL);
835 if (signal_pending(current)) {
836 ret = sock_intr_errno(timeo);
837 break;
838 }
839 continue;
840 }
841 tss.len -= ret;
842 spliced += ret;
843
844 if (!timeo)
845 break;
846 release_sock(sk);
847 lock_sock(sk);
848
849 if (sk->sk_err || sk->sk_state == TCP_CLOSE ||
850 (sk->sk_shutdown & RCV_SHUTDOWN) ||
851 signal_pending(current))
852 break;
853 }
854
855 release_sock(sk);
856
857 if (spliced)
858 return spliced;
859
860 return ret;
861}
862EXPORT_SYMBOL(tcp_splice_read);
863
864struct sk_buff *sk_stream_alloc_skb(struct sock *sk, int size, gfp_t gfp,
865 bool force_schedule)
866{
867 struct sk_buff *skb;
868
869 /* The TCP header must be at least 32-bit aligned. */
870 size = ALIGN(size, 4);
871
872 if (unlikely(tcp_under_memory_pressure(sk)))
873 sk_mem_reclaim_partial(sk);
874
875 skb = alloc_skb_fclone(size + sk->sk_prot->max_header, gfp);
876 if (likely(skb)) {
877 bool mem_scheduled;
878
879 if (force_schedule) {
880 mem_scheduled = true;
881 sk_forced_mem_schedule(sk, skb->truesize);
882 } else {
883 mem_scheduled = sk_wmem_schedule(sk, skb->truesize);
884 }
885 if (likely(mem_scheduled)) {
886 skb_reserve(skb, sk->sk_prot->max_header);
887 /*
888 * Make sure that we have exactly size bytes
889 * available to the caller, no more, no less.
890 */
891 skb->reserved_tailroom = skb->end - skb->tail - size;
892 INIT_LIST_HEAD(&skb->tcp_tsorted_anchor);
893 return skb;
894 }
895 __kfree_skb(skb);
896 } else {
897 sk->sk_prot->enter_memory_pressure(sk);
898 sk_stream_moderate_sndbuf(sk);
899 }
900 return NULL;
901}
902
903static unsigned int tcp_xmit_size_goal(struct sock *sk, u32 mss_now,
904 int large_allowed)
905{
906 struct tcp_sock *tp = tcp_sk(sk);
907 u32 new_size_goal, size_goal;
908
909 if (!large_allowed)
910 return mss_now;
911
912 /* Note : tcp_tso_autosize() will eventually split this later */
913 new_size_goal = sk->sk_gso_max_size - 1 - MAX_TCP_HEADER;
914 new_size_goal = tcp_bound_to_half_wnd(tp, new_size_goal);
915
916 /* We try hard to avoid divides here */
917 size_goal = tp->gso_segs * mss_now;
918 if (unlikely(new_size_goal < size_goal ||
919 new_size_goal >= size_goal + mss_now)) {
920 tp->gso_segs = min_t(u16, new_size_goal / mss_now,
921 sk->sk_gso_max_segs);
922 size_goal = tp->gso_segs * mss_now;
923 }
924
925 return max(size_goal, mss_now);
926}
927
928static int tcp_send_mss(struct sock *sk, int *size_goal, int flags)
929{
930 int mss_now;
931
932 mss_now = tcp_current_mss(sk);
933 *size_goal = tcp_xmit_size_goal(sk, mss_now, !(flags & MSG_OOB));
934
935 return mss_now;
936}
937
938ssize_t do_tcp_sendpages(struct sock *sk, struct page *page, int offset,
939 size_t size, int flags)
940{
941 struct tcp_sock *tp = tcp_sk(sk);
942 int mss_now, size_goal;
943 int err;
944 ssize_t copied;
945 long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
946
947 /* Wait for a connection to finish. One exception is TCP Fast Open
948 * (passive side) where data is allowed to be sent before a connection
949 * is fully established.
950 */
951 if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) &&
952 !tcp_passive_fastopen(sk)) {
953 err = sk_stream_wait_connect(sk, &timeo);
954 if (err != 0)
955 goto out_err;
956 }
957
958 sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
959
960 mss_now = tcp_send_mss(sk, &size_goal, flags);
961 copied = 0;
962
963 err = -EPIPE;
964 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
965 goto out_err;
966
967 while (size > 0) {
968 struct sk_buff *skb = tcp_write_queue_tail(sk);
969 int copy, i;
970 bool can_coalesce;
971
972 if (!skb || (copy = size_goal - skb->len) <= 0 ||
973 !tcp_skb_can_collapse_to(skb)) {
974new_segment:
975 if (!sk_stream_memory_free(sk))
976 goto wait_for_sndbuf;
977
978 skb = sk_stream_alloc_skb(sk, 0, sk->sk_allocation,
979 tcp_rtx_and_write_queues_empty(sk));
980 if (!skb)
981 goto wait_for_memory;
982
983 skb_entail(sk, skb);
984 copy = size_goal;
985 }
986
987 if (copy > size)
988 copy = size;
989
990 i = skb_shinfo(skb)->nr_frags;
991 can_coalesce = skb_can_coalesce(skb, i, page, offset);
992 if (!can_coalesce && i >= sysctl_max_skb_frags) {
993 tcp_mark_push(tp, skb);
994 goto new_segment;
995 }
996 if (!sk_wmem_schedule(sk, copy))
997 goto wait_for_memory;
998
999 if (can_coalesce) {
1000 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1001 } else {
1002 get_page(page);
1003 skb_fill_page_desc(skb, i, page, offset, copy);
1004 }
1005
1006 if (!(flags & MSG_NO_SHARED_FRAGS))
1007 skb_shinfo(skb)->tx_flags |= SKBTX_SHARED_FRAG;
1008
1009 skb->len += copy;
1010 skb->data_len += copy;
1011 skb->truesize += copy;
1012 sk->sk_wmem_queued += copy;
1013 sk_mem_charge(sk, copy);
1014 skb->ip_summed = CHECKSUM_PARTIAL;
1015 tp->write_seq += copy;
1016 TCP_SKB_CB(skb)->end_seq += copy;
1017 tcp_skb_pcount_set(skb, 0);
1018
1019 if (!copied)
1020 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
1021
1022 copied += copy;
1023 offset += copy;
1024 size -= copy;
1025 if (!size)
1026 goto out;
1027
1028 if (skb->len < size_goal || (flags & MSG_OOB))
1029 continue;
1030
1031 if (forced_push(tp)) {
1032 tcp_mark_push(tp, skb);
1033 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
1034 } else if (skb == tcp_send_head(sk))
1035 tcp_push_one(sk, mss_now);
1036 continue;
1037
1038wait_for_sndbuf:
1039 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1040wait_for_memory:
1041 tcp_push(sk, flags & ~MSG_MORE, mss_now,
1042 TCP_NAGLE_PUSH, size_goal);
1043
1044 err = sk_stream_wait_memory(sk, &timeo);
1045 if (err != 0)
1046 goto do_error;
1047
1048 mss_now = tcp_send_mss(sk, &size_goal, flags);
1049 }
1050
1051out:
1052 if (copied) {
1053 tcp_tx_timestamp(sk, sk->sk_tsflags);
1054 if (!(flags & MSG_SENDPAGE_NOTLAST))
1055 tcp_push(sk, flags, mss_now, tp->nonagle, size_goal);
1056 }
1057 return copied;
1058
1059do_error:
1060 if (copied)
1061 goto out;
1062out_err:
1063 /* make sure we wake any epoll edge trigger waiter */
1064 if (unlikely(skb_queue_len(&sk->sk_write_queue) == 0 &&
1065 err == -EAGAIN)) {
1066 sk->sk_write_space(sk);
1067 tcp_chrono_stop(sk, TCP_CHRONO_SNDBUF_LIMITED);
1068 }
1069 return sk_stream_error(sk, flags, err);
1070}
1071EXPORT_SYMBOL_GPL(do_tcp_sendpages);
1072
1073int tcp_sendpage_locked(struct sock *sk, struct page *page, int offset,
1074 size_t size, int flags)
1075{
1076 if (!(sk->sk_route_caps & NETIF_F_SG))
1077 return sock_no_sendpage_locked(sk, page, offset, size, flags);
1078
1079 tcp_rate_check_app_limited(sk); /* is sending application-limited? */
1080
1081 return do_tcp_sendpages(sk, page, offset, size, flags);
1082}
1083EXPORT_SYMBOL_GPL(tcp_sendpage_locked);
1084
1085int tcp_sendpage(struct sock *sk, struct page *page, int offset,
1086 size_t size, int flags)
1087{
1088 int ret;
1089
1090 lock_sock(sk);
1091 ret = tcp_sendpage_locked(sk, page, offset, size, flags);
1092 release_sock(sk);
1093
1094 return ret;
1095}
1096EXPORT_SYMBOL(tcp_sendpage);
1097
1098/* Do not bother using a page frag for very small frames.
1099 * But use this heuristic only for the first skb in write queue.
1100 *
1101 * Having no payload in skb->head allows better SACK shifting
1102 * in tcp_shift_skb_data(), reducing sack/rack overhead, because
1103 * write queue has less skbs.
1104 * Each skb can hold up to MAX_SKB_FRAGS * 32Kbytes, or ~0.5 MB.
1105 * This also speeds up tso_fragment(), since it wont fallback
1106 * to tcp_fragment().
1107 */
1108static int linear_payload_sz(bool first_skb)
1109{
1110 if (first_skb)
1111 return SKB_WITH_OVERHEAD(2048 - MAX_TCP_HEADER);
1112 return 0;
1113}
1114
1115static int select_size(bool first_skb, bool zc)
1116{
1117 if (zc)
1118 return 0;
1119 return linear_payload_sz(first_skb);
1120}
1121
1122void tcp_free_fastopen_req(struct tcp_sock *tp)
1123{
1124 if (tp->fastopen_req) {
1125 kfree(tp->fastopen_req);
1126 tp->fastopen_req = NULL;
1127 }
1128}
1129
1130static int tcp_sendmsg_fastopen(struct sock *sk, struct msghdr *msg,
1131 int *copied, size_t size)
1132{
1133 struct tcp_sock *tp = tcp_sk(sk);
1134 struct inet_sock *inet = inet_sk(sk);
1135 struct sockaddr *uaddr = msg->msg_name;
1136 int err, flags;
1137
1138 if (!(sock_net(sk)->ipv4.sysctl_tcp_fastopen & TFO_CLIENT_ENABLE) ||
1139 (uaddr && msg->msg_namelen >= sizeof(uaddr->sa_family) &&
1140 uaddr->sa_family == AF_UNSPEC))
1141 return -EOPNOTSUPP;
1142 if (tp->fastopen_req)
1143 return -EALREADY; /* Another Fast Open is in progress */
1144
1145 tp->fastopen_req = kzalloc(sizeof(struct tcp_fastopen_request),
1146 sk->sk_allocation);
1147 if (unlikely(!tp->fastopen_req))
1148 return -ENOBUFS;
1149 tp->fastopen_req->data = msg;
1150 tp->fastopen_req->size = size;
1151
1152 if (inet->defer_connect) {
1153 err = tcp_connect(sk);
1154 /* Same failure procedure as in tcp_v4/6_connect */
1155 if (err) {
1156 tcp_set_state(sk, TCP_CLOSE);
1157 inet->inet_dport = 0;
1158 sk->sk_route_caps = 0;
1159 }
1160 }
1161 flags = (msg->msg_flags & MSG_DONTWAIT) ? O_NONBLOCK : 0;
1162 err = __inet_stream_connect(sk->sk_socket, uaddr,
1163 msg->msg_namelen, flags, 1);
1164 /* fastopen_req could already be freed in __inet_stream_connect
1165 * if the connection times out or gets rst
1166 */
1167 if (tp->fastopen_req) {
1168 *copied = tp->fastopen_req->copied;
1169 tcp_free_fastopen_req(tp);
1170 inet->defer_connect = 0;
1171 }
1172 return err;
1173}
1174
1175int tcp_sendmsg_locked(struct sock *sk, struct msghdr *msg, size_t size)
1176{
1177 struct tcp_sock *tp = tcp_sk(sk);
1178 struct ubuf_info *uarg = NULL;
1179 struct sk_buff *skb;
1180 struct sockcm_cookie sockc;
1181 int flags, err, copied = 0;
1182 int mss_now = 0, size_goal, copied_syn = 0;
1183 bool process_backlog = false;
1184 bool zc = false;
1185 long timeo;
1186
1187 flags = msg->msg_flags;
1188
1189 if (flags & MSG_ZEROCOPY && size) {
1190 if (sk->sk_state != TCP_ESTABLISHED) {
1191 err = -EINVAL;
1192 goto out_err;
1193 }
1194
1195 skb = tcp_write_queue_tail(sk);
1196 uarg = sock_zerocopy_realloc(sk, size, skb_zcopy(skb));
1197 if (!uarg) {
1198 err = -ENOBUFS;
1199 goto out_err;
1200 }
1201
1202 zc = sk->sk_route_caps & NETIF_F_SG;
1203 if (!zc)
1204 uarg->zerocopy = 0;
1205 }
1206
1207 if (unlikely(flags & MSG_FASTOPEN || inet_sk(sk)->defer_connect) &&
1208 !tp->repair) {
1209 err = tcp_sendmsg_fastopen(sk, msg, &copied_syn, size);
1210 if (err == -EINPROGRESS && copied_syn > 0)
1211 goto out;
1212 else if (err)
1213 goto out_err;
1214 }
1215
1216 timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
1217
1218 tcp_rate_check_app_limited(sk); /* is sending application-limited? */
1219
1220 /* Wait for a connection to finish. One exception is TCP Fast Open
1221 * (passive side) where data is allowed to be sent before a connection
1222 * is fully established.
1223 */
1224 if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) &&
1225 !tcp_passive_fastopen(sk)) {
1226 err = sk_stream_wait_connect(sk, &timeo);
1227 if (err != 0)
1228 goto do_error;
1229 }
1230
1231 if (unlikely(tp->repair)) {
1232 if (tp->repair_queue == TCP_RECV_QUEUE) {
1233 copied = tcp_send_rcvq(sk, msg, size);
1234 goto out_nopush;
1235 }
1236
1237 err = -EINVAL;
1238 if (tp->repair_queue == TCP_NO_QUEUE)
1239 goto out_err;
1240
1241 /* 'common' sending to sendq */
1242 }
1243
1244 sockc.tsflags = sk->sk_tsflags;
1245 if (msg->msg_controllen) {
1246 err = sock_cmsg_send(sk, msg, &sockc);
1247 if (unlikely(err)) {
1248 err = -EINVAL;
1249 goto out_err;
1250 }
1251 }
1252
1253 /* This should be in poll */
1254 sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
1255
1256 /* Ok commence sending. */
1257 copied = 0;
1258
1259restart:
1260 mss_now = tcp_send_mss(sk, &size_goal, flags);
1261
1262 err = -EPIPE;
1263 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
1264 goto do_error;
1265
1266 while (msg_data_left(msg)) {
1267 int copy = 0;
1268
1269 skb = tcp_write_queue_tail(sk);
1270 if (skb)
1271 copy = size_goal - skb->len;
1272
1273 if (copy <= 0 || !tcp_skb_can_collapse_to(skb)) {
1274 bool first_skb;
1275 int linear;
1276
1277new_segment:
1278 /* Allocate new segment. If the interface is SG,
1279 * allocate skb fitting to single page.
1280 */
1281 if (!sk_stream_memory_free(sk))
1282 goto wait_for_sndbuf;
1283
1284 if (process_backlog && sk_flush_backlog(sk)) {
1285 process_backlog = false;
1286 goto restart;
1287 }
1288 first_skb = tcp_rtx_and_write_queues_empty(sk);
1289 linear = select_size(first_skb, zc);
1290 skb = sk_stream_alloc_skb(sk, linear, sk->sk_allocation,
1291 first_skb);
1292 if (!skb)
1293 goto wait_for_memory;
1294
1295 process_backlog = true;
1296 skb->ip_summed = CHECKSUM_PARTIAL;
1297
1298 skb_entail(sk, skb);
1299 copy = size_goal;
1300
1301 /* All packets are restored as if they have
1302 * already been sent. skb_mstamp isn't set to
1303 * avoid wrong rtt estimation.
1304 */
1305 if (tp->repair)
1306 TCP_SKB_CB(skb)->sacked |= TCPCB_REPAIRED;
1307 }
1308
1309 /* Try to append data to the end of skb. */
1310 if (copy > msg_data_left(msg))
1311 copy = msg_data_left(msg);
1312
1313 /* Where to copy to? */
1314 if (skb_availroom(skb) > 0 && !zc) {
1315 /* We have some space in skb head. Superb! */
1316 copy = min_t(int, copy, skb_availroom(skb));
1317 err = skb_add_data_nocache(sk, skb, &msg->msg_iter, copy);
1318 if (err)
1319 goto do_fault;
1320 } else if (!zc) {
1321 bool merge = true;
1322 int i = skb_shinfo(skb)->nr_frags;
1323 struct page_frag *pfrag = sk_page_frag(sk);
1324
1325 if (!sk_page_frag_refill(sk, pfrag))
1326 goto wait_for_memory;
1327
1328 if (!skb_can_coalesce(skb, i, pfrag->page,
1329 pfrag->offset)) {
1330 if (i >= sysctl_max_skb_frags) {
1331 tcp_mark_push(tp, skb);
1332 goto new_segment;
1333 }
1334 merge = false;
1335 }
1336
1337 copy = min_t(int, copy, pfrag->size - pfrag->offset);
1338
1339 if (!sk_wmem_schedule(sk, copy))
1340 goto wait_for_memory;
1341
1342 err = skb_copy_to_page_nocache(sk, &msg->msg_iter, skb,
1343 pfrag->page,
1344 pfrag->offset,
1345 copy);
1346 if (err)
1347 goto do_error;
1348
1349 /* Update the skb. */
1350 if (merge) {
1351 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1352 } else {
1353 skb_fill_page_desc(skb, i, pfrag->page,
1354 pfrag->offset, copy);
1355 page_ref_inc(pfrag->page);
1356 }
1357 pfrag->offset += copy;
1358 } else {
1359 err = skb_zerocopy_iter_stream(sk, skb, msg, copy, uarg);
1360 if (err == -EMSGSIZE || err == -EEXIST) {
1361 tcp_mark_push(tp, skb);
1362 goto new_segment;
1363 }
1364 if (err < 0)
1365 goto do_error;
1366 copy = err;
1367 }
1368
1369 if (!copied)
1370 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
1371
1372 tp->write_seq += copy;
1373 TCP_SKB_CB(skb)->end_seq += copy;
1374 tcp_skb_pcount_set(skb, 0);
1375
1376 copied += copy;
1377 if (!msg_data_left(msg)) {
1378 if (unlikely(flags & MSG_EOR))
1379 TCP_SKB_CB(skb)->eor = 1;
1380 goto out;
1381 }
1382
1383 if (skb->len < size_goal || (flags & MSG_OOB) || unlikely(tp->repair))
1384 continue;
1385
1386 if (forced_push(tp)) {
1387 tcp_mark_push(tp, skb);
1388 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
1389 } else if (skb == tcp_send_head(sk))
1390 tcp_push_one(sk, mss_now);
1391 continue;
1392
1393wait_for_sndbuf:
1394 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1395wait_for_memory:
1396 if (copied)
1397 tcp_push(sk, flags & ~MSG_MORE, mss_now,
1398 TCP_NAGLE_PUSH, size_goal);
1399
1400 err = sk_stream_wait_memory(sk, &timeo);
1401 if (err != 0)
1402 goto do_error;
1403
1404 mss_now = tcp_send_mss(sk, &size_goal, flags);
1405 }
1406
1407out:
1408 if (copied) {
1409 tcp_tx_timestamp(sk, sockc.tsflags);
1410 tcp_push(sk, flags, mss_now, tp->nonagle, size_goal);
1411 }
1412out_nopush:
1413 sock_zerocopy_put(uarg);
1414 return copied + copied_syn;
1415
1416do_fault:
1417 if (!skb->len) {
1418 tcp_unlink_write_queue(skb, sk);
1419 /* It is the one place in all of TCP, except connection
1420 * reset, where we can be unlinking the send_head.
1421 */
1422 tcp_check_send_head(sk, skb);
1423 sk_wmem_free_skb(sk, skb);
1424 }
1425
1426do_error:
1427 if (copied + copied_syn)
1428 goto out;
1429out_err:
1430 sock_zerocopy_put_abort(uarg);
1431 err = sk_stream_error(sk, flags, err);
1432 /* make sure we wake any epoll edge trigger waiter */
1433 if (unlikely(skb_queue_len(&sk->sk_write_queue) == 0 &&
1434 err == -EAGAIN)) {
1435 sk->sk_write_space(sk);
1436 tcp_chrono_stop(sk, TCP_CHRONO_SNDBUF_LIMITED);
1437 }
1438 return err;
1439}
1440EXPORT_SYMBOL_GPL(tcp_sendmsg_locked);
1441
1442int tcp_sendmsg(struct sock *sk, struct msghdr *msg, size_t size)
1443{
1444 int ret;
1445
1446 lock_sock(sk);
1447 ret = tcp_sendmsg_locked(sk, msg, size);
1448 release_sock(sk);
1449
1450 return ret;
1451}
1452EXPORT_SYMBOL(tcp_sendmsg);
1453
1454/*
1455 * Handle reading urgent data. BSD has very simple semantics for
1456 * this, no blocking and very strange errors 8)
1457 */
1458
1459static int tcp_recv_urg(struct sock *sk, struct msghdr *msg, int len, int flags)
1460{
1461 struct tcp_sock *tp = tcp_sk(sk);
1462
1463 /* No URG data to read. */
1464 if (sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data ||
1465 tp->urg_data == TCP_URG_READ)
1466 return -EINVAL; /* Yes this is right ! */
1467
1468 if (sk->sk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DONE))
1469 return -ENOTCONN;
1470
1471 if (tp->urg_data & TCP_URG_VALID) {
1472 int err = 0;
1473 char c = tp->urg_data;
1474
1475 if (!(flags & MSG_PEEK))
1476 tp->urg_data = TCP_URG_READ;
1477
1478 /* Read urgent data. */
1479 msg->msg_flags |= MSG_OOB;
1480
1481 if (len > 0) {
1482 if (!(flags & MSG_TRUNC))
1483 err = memcpy_to_msg(msg, &c, 1);
1484 len = 1;
1485 } else
1486 msg->msg_flags |= MSG_TRUNC;
1487
1488 return err ? -EFAULT : len;
1489 }
1490
1491 if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
1492 return 0;
1493
1494 /* Fixed the recv(..., MSG_OOB) behaviour. BSD docs and
1495 * the available implementations agree in this case:
1496 * this call should never block, independent of the
1497 * blocking state of the socket.
1498 * Mike <pall@rz.uni-karlsruhe.de>
1499 */
1500 return -EAGAIN;
1501}
1502
1503static int tcp_peek_sndq(struct sock *sk, struct msghdr *msg, int len)
1504{
1505 struct sk_buff *skb;
1506 int copied = 0, err = 0;
1507
1508 /* XXX -- need to support SO_PEEK_OFF */
1509
1510 skb_rbtree_walk(skb, &sk->tcp_rtx_queue) {
1511 err = skb_copy_datagram_msg(skb, 0, msg, skb->len);
1512 if (err)
1513 return err;
1514 copied += skb->len;
1515 }
1516
1517 skb_queue_walk(&sk->sk_write_queue, skb) {
1518 err = skb_copy_datagram_msg(skb, 0, msg, skb->len);
1519 if (err)
1520 break;
1521
1522 copied += skb->len;
1523 }
1524
1525 return err ?: copied;
1526}
1527
1528/* Clean up the receive buffer for full frames taken by the user,
1529 * then send an ACK if necessary. COPIED is the number of bytes
1530 * tcp_recvmsg has given to the user so far, it speeds up the
1531 * calculation of whether or not we must ACK for the sake of
1532 * a window update.
1533 */
1534static void tcp_cleanup_rbuf(struct sock *sk, int copied)
1535{
1536 struct tcp_sock *tp = tcp_sk(sk);
1537 bool time_to_ack = false;
1538
1539 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
1540
1541 WARN(skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq),
1542 "cleanup rbuf bug: copied %X seq %X rcvnxt %X\n",
1543 tp->copied_seq, TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt);
1544
1545 if (inet_csk_ack_scheduled(sk)) {
1546 const struct inet_connection_sock *icsk = inet_csk(sk);
1547 /* Delayed ACKs frequently hit locked sockets during bulk
1548 * receive. */
1549 if (icsk->icsk_ack.blocked ||
1550 /* Once-per-two-segments ACK was not sent by tcp_input.c */
1551 tp->rcv_nxt - tp->rcv_wup > icsk->icsk_ack.rcv_mss ||
1552 /*
1553 * If this read emptied read buffer, we send ACK, if
1554 * connection is not bidirectional, user drained
1555 * receive buffer and there was a small segment
1556 * in queue.
1557 */
1558 (copied > 0 &&
1559 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED2) ||
1560 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED) &&
1561 !icsk->icsk_ack.pingpong)) &&
1562 !atomic_read(&sk->sk_rmem_alloc)))
1563 time_to_ack = true;
1564 }
1565
1566 /* We send an ACK if we can now advertise a non-zero window
1567 * which has been raised "significantly".
1568 *
1569 * Even if window raised up to infinity, do not send window open ACK
1570 * in states, where we will not receive more. It is useless.
1571 */
1572 if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) {
1573 __u32 rcv_window_now = tcp_receive_window(tp);
1574
1575 /* Optimize, __tcp_select_window() is not cheap. */
1576 if (2*rcv_window_now <= tp->window_clamp) {
1577 __u32 new_window = __tcp_select_window(sk);
1578
1579 /* Send ACK now, if this read freed lots of space
1580 * in our buffer. Certainly, new_window is new window.
1581 * We can advertise it now, if it is not less than current one.
1582 * "Lots" means "at least twice" here.
1583 */
1584 if (new_window && new_window >= 2 * rcv_window_now)
1585 time_to_ack = true;
1586 }
1587 }
1588 if (time_to_ack)
1589 tcp_send_ack(sk);
1590}
1591
1592static struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off)
1593{
1594 struct sk_buff *skb;
1595 u32 offset;
1596
1597 while ((skb = skb_peek(&sk->sk_receive_queue)) != NULL) {
1598 offset = seq - TCP_SKB_CB(skb)->seq;
1599 if (unlikely(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
1600 pr_err_once("%s: found a SYN, please report !\n", __func__);
1601 offset--;
1602 }
1603 if (offset < skb->len || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)) {
1604 *off = offset;
1605 return skb;
1606 }
1607 /* This looks weird, but this can happen if TCP collapsing
1608 * splitted a fat GRO packet, while we released socket lock
1609 * in skb_splice_bits()
1610 */
1611 sk_eat_skb(sk, skb);
1612 }
1613 return NULL;
1614}
1615
1616/*
1617 * This routine provides an alternative to tcp_recvmsg() for routines
1618 * that would like to handle copying from skbuffs directly in 'sendfile'
1619 * fashion.
1620 * Note:
1621 * - It is assumed that the socket was locked by the caller.
1622 * - The routine does not block.
1623 * - At present, there is no support for reading OOB data
1624 * or for 'peeking' the socket using this routine
1625 * (although both would be easy to implement).
1626 */
1627int tcp_read_sock(struct sock *sk, read_descriptor_t *desc,
1628 sk_read_actor_t recv_actor)
1629{
1630 struct sk_buff *skb;
1631 struct tcp_sock *tp = tcp_sk(sk);
1632 u32 seq = tp->copied_seq;
1633 u32 offset;
1634 int copied = 0;
1635
1636 if (sk->sk_state == TCP_LISTEN)
1637 return -ENOTCONN;
1638 while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1639 if (offset < skb->len) {
1640 int used;
1641 size_t len;
1642
1643 len = skb->len - offset;
1644 /* Stop reading if we hit a patch of urgent data */
1645 if (tp->urg_data) {
1646 u32 urg_offset = tp->urg_seq - seq;
1647 if (urg_offset < len)
1648 len = urg_offset;
1649 if (!len)
1650 break;
1651 }
1652 used = recv_actor(desc, skb, offset, len);
1653 if (used <= 0) {
1654 if (!copied)
1655 copied = used;
1656 break;
1657 } else if (used <= len) {
1658 seq += used;
1659 copied += used;
1660 offset += used;
1661 }
1662 /* If recv_actor drops the lock (e.g. TCP splice
1663 * receive) the skb pointer might be invalid when
1664 * getting here: tcp_collapse might have deleted it
1665 * while aggregating skbs from the socket queue.
1666 */
1667 skb = tcp_recv_skb(sk, seq - 1, &offset);
1668 if (!skb)
1669 break;
1670 /* TCP coalescing might have appended data to the skb.
1671 * Try to splice more frags
1672 */
1673 if (offset + 1 != skb->len)
1674 continue;
1675 }
1676 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) {
1677 sk_eat_skb(sk, skb);
1678 ++seq;
1679 break;
1680 }
1681 sk_eat_skb(sk, skb);
1682 if (!desc->count)
1683 break;
1684 tp->copied_seq = seq;
1685 }
1686 tp->copied_seq = seq;
1687
1688 tcp_rcv_space_adjust(sk);
1689
1690 /* Clean up data we have read: This will do ACK frames. */
1691 if (copied > 0) {
1692 tcp_recv_skb(sk, seq, &offset);
1693 tcp_cleanup_rbuf(sk, copied);
1694 }
1695 return copied;
1696}
1697EXPORT_SYMBOL(tcp_read_sock);
1698
1699int tcp_peek_len(struct socket *sock)
1700{
1701 return tcp_inq(sock->sk);
1702}
1703EXPORT_SYMBOL(tcp_peek_len);
1704
1705static void tcp_update_recv_tstamps(struct sk_buff *skb,
1706 struct scm_timestamping *tss)
1707{
1708 if (skb->tstamp)
1709 tss->ts[0] = ktime_to_timespec(skb->tstamp);
1710 else
1711 tss->ts[0] = (struct timespec) {0};
1712
1713 if (skb_hwtstamps(skb)->hwtstamp)
1714 tss->ts[2] = ktime_to_timespec(skb_hwtstamps(skb)->hwtstamp);
1715 else
1716 tss->ts[2] = (struct timespec) {0};
1717}
1718
1719/* Similar to __sock_recv_timestamp, but does not require an skb */
1720static void tcp_recv_timestamp(struct msghdr *msg, const struct sock *sk,
1721 struct scm_timestamping *tss)
1722{
1723 struct timeval tv;
1724 bool has_timestamping = false;
1725
1726 if (tss->ts[0].tv_sec || tss->ts[0].tv_nsec) {
1727 if (sock_flag(sk, SOCK_RCVTSTAMP)) {
1728 if (sock_flag(sk, SOCK_RCVTSTAMPNS)) {
1729 put_cmsg(msg, SOL_SOCKET, SCM_TIMESTAMPNS,
1730 sizeof(tss->ts[0]), &tss->ts[0]);
1731 } else {
1732 tv.tv_sec = tss->ts[0].tv_sec;
1733 tv.tv_usec = tss->ts[0].tv_nsec / 1000;
1734
1735 put_cmsg(msg, SOL_SOCKET, SCM_TIMESTAMP,
1736 sizeof(tv), &tv);
1737 }
1738 }
1739
1740 if (sk->sk_tsflags & SOF_TIMESTAMPING_SOFTWARE)
1741 has_timestamping = true;
1742 else
1743 tss->ts[0] = (struct timespec) {0};
1744 }
1745
1746 if (tss->ts[2].tv_sec || tss->ts[2].tv_nsec) {
1747 if (sk->sk_tsflags & SOF_TIMESTAMPING_RAW_HARDWARE)
1748 has_timestamping = true;
1749 else
1750 tss->ts[2] = (struct timespec) {0};
1751 }
1752
1753 if (has_timestamping) {
1754 tss->ts[1] = (struct timespec) {0};
1755 put_cmsg(msg, SOL_SOCKET, SCM_TIMESTAMPING,
1756 sizeof(*tss), tss);
1757 }
1758}
1759
1760/*
1761 * This routine copies from a sock struct into the user buffer.
1762 *
1763 * Technical note: in 2.3 we work on _locked_ socket, so that
1764 * tricks with *seq access order and skb->users are not required.
1765 * Probably, code can be easily improved even more.
1766 */
1767
1768int tcp_recvmsg(struct sock *sk, struct msghdr *msg, size_t len, int nonblock,
1769 int flags, int *addr_len)
1770{
1771 struct tcp_sock *tp = tcp_sk(sk);
1772 int copied = 0;
1773 u32 peek_seq;
1774 u32 *seq;
1775 unsigned long used;
1776 int err;
1777 int target; /* Read at least this many bytes */
1778 long timeo;
1779 struct sk_buff *skb, *last;
1780 u32 urg_hole = 0;
1781 struct scm_timestamping tss;
1782 bool has_tss = false;
1783
1784 if (unlikely(flags & MSG_ERRQUEUE))
1785 return inet_recv_error(sk, msg, len, addr_len);
1786
1787 if (sk_can_busy_loop(sk) && skb_queue_empty(&sk->sk_receive_queue) &&
1788 (sk->sk_state == TCP_ESTABLISHED))
1789 sk_busy_loop(sk, nonblock);
1790
1791 lock_sock(sk);
1792
1793 err = -ENOTCONN;
1794 if (sk->sk_state == TCP_LISTEN)
1795 goto out;
1796
1797 timeo = sock_rcvtimeo(sk, nonblock);
1798
1799 /* Urgent data needs to be handled specially. */
1800 if (flags & MSG_OOB)
1801 goto recv_urg;
1802
1803 if (unlikely(tp->repair)) {
1804 err = -EPERM;
1805 if (!(flags & MSG_PEEK))
1806 goto out;
1807
1808 if (tp->repair_queue == TCP_SEND_QUEUE)
1809 goto recv_sndq;
1810
1811 err = -EINVAL;
1812 if (tp->repair_queue == TCP_NO_QUEUE)
1813 goto out;
1814
1815 /* 'common' recv queue MSG_PEEK-ing */
1816 }
1817
1818 seq = &tp->copied_seq;
1819 if (flags & MSG_PEEK) {
1820 peek_seq = tp->copied_seq;
1821 seq = &peek_seq;
1822 }
1823
1824 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
1825
1826 do {
1827 u32 offset;
1828
1829 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
1830 if (tp->urg_data && tp->urg_seq == *seq) {
1831 if (copied)
1832 break;
1833 if (signal_pending(current)) {
1834 copied = timeo ? sock_intr_errno(timeo) : -EAGAIN;
1835 break;
1836 }
1837 }
1838
1839 /* Next get a buffer. */
1840
1841 last = skb_peek_tail(&sk->sk_receive_queue);
1842 skb_queue_walk(&sk->sk_receive_queue, skb) {
1843 last = skb;
1844 /* Now that we have two receive queues this
1845 * shouldn't happen.
1846 */
1847 if (WARN(before(*seq, TCP_SKB_CB(skb)->seq),
1848 "recvmsg bug: copied %X seq %X rcvnxt %X fl %X\n",
1849 *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt,
1850 flags))
1851 break;
1852
1853 offset = *seq - TCP_SKB_CB(skb)->seq;
1854 if (unlikely(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
1855 pr_err_once("%s: found a SYN, please report !\n", __func__);
1856 offset--;
1857 }
1858 if (offset < skb->len)
1859 goto found_ok_skb;
1860 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1861 goto found_fin_ok;
1862 WARN(!(flags & MSG_PEEK),
1863 "recvmsg bug 2: copied %X seq %X rcvnxt %X fl %X\n",
1864 *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt, flags);
1865 }
1866
1867 /* Well, if we have backlog, try to process it now yet. */
1868
1869 if (copied >= target && !sk->sk_backlog.tail)
1870 break;
1871
1872 if (copied) {
1873 if (sk->sk_err ||
1874 sk->sk_state == TCP_CLOSE ||
1875 (sk->sk_shutdown & RCV_SHUTDOWN) ||
1876 !timeo ||
1877 signal_pending(current))
1878 break;
1879 } else {
1880 if (sock_flag(sk, SOCK_DONE))
1881 break;
1882
1883 if (sk->sk_err) {
1884 copied = sock_error(sk);
1885 break;
1886 }
1887
1888 if (sk->sk_shutdown & RCV_SHUTDOWN)
1889 break;
1890
1891 if (sk->sk_state == TCP_CLOSE) {
1892 if (!sock_flag(sk, SOCK_DONE)) {
1893 /* This occurs when user tries to read
1894 * from never connected socket.
1895 */
1896 copied = -ENOTCONN;
1897 break;
1898 }
1899 break;
1900 }
1901
1902 if (!timeo) {
1903 copied = -EAGAIN;
1904 break;
1905 }
1906
1907 if (signal_pending(current)) {
1908 copied = sock_intr_errno(timeo);
1909 break;
1910 }
1911 }
1912
1913 tcp_cleanup_rbuf(sk, copied);
1914
1915 if (copied >= target) {
1916 /* Do not sleep, just process backlog. */
1917 release_sock(sk);
1918 lock_sock(sk);
1919 } else {
1920 sk_wait_data(sk, &timeo, last);
1921 }
1922
1923 if ((flags & MSG_PEEK) &&
1924 (peek_seq - copied - urg_hole != tp->copied_seq)) {
1925 net_dbg_ratelimited("TCP(%s:%d): Application bug, race in MSG_PEEK\n",
1926 current->comm,
1927 task_pid_nr(current));
1928 peek_seq = tp->copied_seq;
1929 }
1930 continue;
1931
1932 found_ok_skb:
1933 /* Ok so how much can we use? */
1934 used = skb->len - offset;
1935 if (len < used)
1936 used = len;
1937
1938 /* Do we have urgent data here? */
1939 if (tp->urg_data) {
1940 u32 urg_offset = tp->urg_seq - *seq;
1941 if (urg_offset < used) {
1942 if (!urg_offset) {
1943 if (!sock_flag(sk, SOCK_URGINLINE)) {
1944 ++*seq;
1945 urg_hole++;
1946 offset++;
1947 used--;
1948 if (!used)
1949 goto skip_copy;
1950 }
1951 } else
1952 used = urg_offset;
1953 }
1954 }
1955
1956 if (!(flags & MSG_TRUNC)) {
1957 err = skb_copy_datagram_msg(skb, offset, msg, used);
1958 if (err) {
1959 /* Exception. Bailout! */
1960 if (!copied)
1961 copied = -EFAULT;
1962 break;
1963 }
1964 }
1965
1966 *seq += used;
1967 copied += used;
1968 len -= used;
1969
1970 tcp_rcv_space_adjust(sk);
1971
1972skip_copy:
1973 if (tp->urg_data && after(tp->copied_seq, tp->urg_seq)) {
1974 tp->urg_data = 0;
1975 tcp_fast_path_check(sk);
1976 }
1977 if (used + offset < skb->len)
1978 continue;
1979
1980 if (TCP_SKB_CB(skb)->has_rxtstamp) {
1981 tcp_update_recv_tstamps(skb, &tss);
1982 has_tss = true;
1983 }
1984 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1985 goto found_fin_ok;
1986 if (!(flags & MSG_PEEK))
1987 sk_eat_skb(sk, skb);
1988 continue;
1989
1990 found_fin_ok:
1991 /* Process the FIN. */
1992 ++*seq;
1993 if (!(flags & MSG_PEEK))
1994 sk_eat_skb(sk, skb);
1995 break;
1996 } while (len > 0);
1997
1998 /* According to UNIX98, msg_name/msg_namelen are ignored
1999 * on connected socket. I was just happy when found this 8) --ANK
2000 */
2001
2002 if (has_tss)
2003 tcp_recv_timestamp(msg, sk, &tss);
2004
2005 /* Clean up data we have read: This will do ACK frames. */
2006 tcp_cleanup_rbuf(sk, copied);
2007
2008 release_sock(sk);
2009 return copied;
2010
2011out:
2012 release_sock(sk);
2013 return err;
2014
2015recv_urg:
2016 err = tcp_recv_urg(sk, msg, len, flags);
2017 goto out;
2018
2019recv_sndq:
2020 err = tcp_peek_sndq(sk, msg, len);
2021 goto out;
2022}
2023EXPORT_SYMBOL(tcp_recvmsg);
2024
2025void tcp_set_state(struct sock *sk, int state)
2026{
2027 int oldstate = sk->sk_state;
2028
2029 /* We defined a new enum for TCP states that are exported in BPF
2030 * so as not force the internal TCP states to be frozen. The
2031 * following checks will detect if an internal state value ever
2032 * differs from the BPF value. If this ever happens, then we will
2033 * need to remap the internal value to the BPF value before calling
2034 * tcp_call_bpf_2arg.
2035 */
2036 BUILD_BUG_ON((int)BPF_TCP_ESTABLISHED != (int)TCP_ESTABLISHED);
2037 BUILD_BUG_ON((int)BPF_TCP_SYN_SENT != (int)TCP_SYN_SENT);
2038 BUILD_BUG_ON((int)BPF_TCP_SYN_RECV != (int)TCP_SYN_RECV);
2039 BUILD_BUG_ON((int)BPF_TCP_FIN_WAIT1 != (int)TCP_FIN_WAIT1);
2040 BUILD_BUG_ON((int)BPF_TCP_FIN_WAIT2 != (int)TCP_FIN_WAIT2);
2041 BUILD_BUG_ON((int)BPF_TCP_TIME_WAIT != (int)TCP_TIME_WAIT);
2042 BUILD_BUG_ON((int)BPF_TCP_CLOSE != (int)TCP_CLOSE);
2043 BUILD_BUG_ON((int)BPF_TCP_CLOSE_WAIT != (int)TCP_CLOSE_WAIT);
2044 BUILD_BUG_ON((int)BPF_TCP_LAST_ACK != (int)TCP_LAST_ACK);
2045 BUILD_BUG_ON((int)BPF_TCP_LISTEN != (int)TCP_LISTEN);
2046 BUILD_BUG_ON((int)BPF_TCP_CLOSING != (int)TCP_CLOSING);
2047 BUILD_BUG_ON((int)BPF_TCP_NEW_SYN_RECV != (int)TCP_NEW_SYN_RECV);
2048 BUILD_BUG_ON((int)BPF_TCP_MAX_STATES != (int)TCP_MAX_STATES);
2049
2050 if (BPF_SOCK_OPS_TEST_FLAG(tcp_sk(sk), BPF_SOCK_OPS_STATE_CB_FLAG))
2051 tcp_call_bpf_2arg(sk, BPF_SOCK_OPS_STATE_CB, oldstate, state);
2052
2053 switch (state) {
2054 case TCP_ESTABLISHED:
2055 if (oldstate != TCP_ESTABLISHED)
2056 TCP_INC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
2057 break;
2058
2059 case TCP_CLOSE:
2060 if (oldstate == TCP_CLOSE_WAIT || oldstate == TCP_ESTABLISHED)
2061 TCP_INC_STATS(sock_net(sk), TCP_MIB_ESTABRESETS);
2062
2063 sk->sk_prot->unhash(sk);
2064 if (inet_csk(sk)->icsk_bind_hash &&
2065 !(sk->sk_userlocks & SOCK_BINDPORT_LOCK))
2066 inet_put_port(sk);
2067 /* fall through */
2068 default:
2069 if (oldstate == TCP_ESTABLISHED)
2070 TCP_DEC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
2071 }
2072
2073 /* Change state AFTER socket is unhashed to avoid closed
2074 * socket sitting in hash tables.
2075 */
2076 inet_sk_state_store(sk, state);
2077
2078#ifdef STATE_TRACE
2079 SOCK_DEBUG(sk, "TCP sk=%p, State %s -> %s\n", sk, statename[oldstate], statename[state]);
2080#endif
2081}
2082EXPORT_SYMBOL_GPL(tcp_set_state);
2083
2084/*
2085 * State processing on a close. This implements the state shift for
2086 * sending our FIN frame. Note that we only send a FIN for some
2087 * states. A shutdown() may have already sent the FIN, or we may be
2088 * closed.
2089 */
2090
2091static const unsigned char new_state[16] = {
2092 /* current state: new state: action: */
2093 [0 /* (Invalid) */] = TCP_CLOSE,
2094 [TCP_ESTABLISHED] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2095 [TCP_SYN_SENT] = TCP_CLOSE,
2096 [TCP_SYN_RECV] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2097 [TCP_FIN_WAIT1] = TCP_FIN_WAIT1,
2098 [TCP_FIN_WAIT2] = TCP_FIN_WAIT2,
2099 [TCP_TIME_WAIT] = TCP_CLOSE,
2100 [TCP_CLOSE] = TCP_CLOSE,
2101 [TCP_CLOSE_WAIT] = TCP_LAST_ACK | TCP_ACTION_FIN,
2102 [TCP_LAST_ACK] = TCP_LAST_ACK,
2103 [TCP_LISTEN] = TCP_CLOSE,
2104 [TCP_CLOSING] = TCP_CLOSING,
2105 [TCP_NEW_SYN_RECV] = TCP_CLOSE, /* should not happen ! */
2106};
2107
2108static int tcp_close_state(struct sock *sk)
2109{
2110 int next = (int)new_state[sk->sk_state];
2111 int ns = next & TCP_STATE_MASK;
2112
2113 tcp_set_state(sk, ns);
2114
2115 return next & TCP_ACTION_FIN;
2116}
2117
2118/*
2119 * Shutdown the sending side of a connection. Much like close except
2120 * that we don't receive shut down or sock_set_flag(sk, SOCK_DEAD).
2121 */
2122
2123void tcp_shutdown(struct sock *sk, int how)
2124{
2125 /* We need to grab some memory, and put together a FIN,
2126 * and then put it into the queue to be sent.
2127 * Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
2128 */
2129 if (!(how & SEND_SHUTDOWN))
2130 return;
2131
2132 /* If we've already sent a FIN, or it's a closed state, skip this. */
2133 if ((1 << sk->sk_state) &
2134 (TCPF_ESTABLISHED | TCPF_SYN_SENT |
2135 TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) {
2136 /* Clear out any half completed packets. FIN if needed. */
2137 if (tcp_close_state(sk))
2138 tcp_send_fin(sk);
2139 }
2140}
2141EXPORT_SYMBOL(tcp_shutdown);
2142
2143bool tcp_check_oom(struct sock *sk, int shift)
2144{
2145 bool too_many_orphans, out_of_socket_memory;
2146
2147 too_many_orphans = tcp_too_many_orphans(sk, shift);
2148 out_of_socket_memory = tcp_out_of_memory(sk);
2149
2150 if (too_many_orphans)
2151 net_info_ratelimited("too many orphaned sockets\n");
2152 if (out_of_socket_memory)
2153 net_info_ratelimited("out of memory -- consider tuning tcp_mem\n");
2154 return too_many_orphans || out_of_socket_memory;
2155}
2156
2157void tcp_close(struct sock *sk, long timeout)
2158{
2159 struct sk_buff *skb;
2160 int data_was_unread = 0;
2161 int state;
2162
2163 lock_sock(sk);
2164 sk->sk_shutdown = SHUTDOWN_MASK;
2165
2166 if (sk->sk_state == TCP_LISTEN) {
2167 tcp_set_state(sk, TCP_CLOSE);
2168
2169 /* Special case. */
2170 inet_csk_listen_stop(sk);
2171
2172 goto adjudge_to_death;
2173 }
2174
2175 /* We need to flush the recv. buffs. We do this only on the
2176 * descriptor close, not protocol-sourced closes, because the
2177 * reader process may not have drained the data yet!
2178 */
2179 while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) {
2180 u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq;
2181
2182 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
2183 len--;
2184 data_was_unread += len;
2185 __kfree_skb(skb);
2186 }
2187
2188 sk_mem_reclaim(sk);
2189
2190 /* If socket has been already reset (e.g. in tcp_reset()) - kill it. */
2191 if (sk->sk_state == TCP_CLOSE)
2192 goto adjudge_to_death;
2193
2194 /* As outlined in RFC 2525, section 2.17, we send a RST here because
2195 * data was lost. To witness the awful effects of the old behavior of
2196 * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk
2197 * GET in an FTP client, suspend the process, wait for the client to
2198 * advertise a zero window, then kill -9 the FTP client, wheee...
2199 * Note: timeout is always zero in such a case.
2200 */
2201 if (unlikely(tcp_sk(sk)->repair)) {
2202 sk->sk_prot->disconnect(sk, 0);
2203 } else if (data_was_unread) {
2204 /* Unread data was tossed, zap the connection. */
2205 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONCLOSE);
2206 tcp_set_state(sk, TCP_CLOSE);
2207 tcp_send_active_reset(sk, sk->sk_allocation);
2208 } else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) {
2209 /* Check zero linger _after_ checking for unread data. */
2210 sk->sk_prot->disconnect(sk, 0);
2211 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
2212 } else if (tcp_close_state(sk)) {
2213 /* We FIN if the application ate all the data before
2214 * zapping the connection.
2215 */
2216
2217 /* RED-PEN. Formally speaking, we have broken TCP state
2218 * machine. State transitions:
2219 *
2220 * TCP_ESTABLISHED -> TCP_FIN_WAIT1
2221 * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible)
2222 * TCP_CLOSE_WAIT -> TCP_LAST_ACK
2223 *
2224 * are legal only when FIN has been sent (i.e. in window),
2225 * rather than queued out of window. Purists blame.
2226 *
2227 * F.e. "RFC state" is ESTABLISHED,
2228 * if Linux state is FIN-WAIT-1, but FIN is still not sent.
2229 *
2230 * The visible declinations are that sometimes
2231 * we enter time-wait state, when it is not required really
2232 * (harmless), do not send active resets, when they are
2233 * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
2234 * they look as CLOSING or LAST_ACK for Linux)
2235 * Probably, I missed some more holelets.
2236 * --ANK
2237 * XXX (TFO) - To start off we don't support SYN+ACK+FIN
2238 * in a single packet! (May consider it later but will
2239 * probably need API support or TCP_CORK SYN-ACK until
2240 * data is written and socket is closed.)
2241 */
2242 tcp_send_fin(sk);
2243 }
2244
2245 sk_stream_wait_close(sk, timeout);
2246
2247adjudge_to_death:
2248 state = sk->sk_state;
2249 sock_hold(sk);
2250 sock_orphan(sk);
2251
2252 /* It is the last release_sock in its life. It will remove backlog. */
2253 release_sock(sk);
2254
2255
2256 /* Now socket is owned by kernel and we acquire BH lock
2257 * to finish close. No need to check for user refs.
2258 */
2259 local_bh_disable();
2260 bh_lock_sock(sk);
2261 WARN_ON(sock_owned_by_user(sk));
2262
2263 percpu_counter_inc(sk->sk_prot->orphan_count);
2264
2265 /* Have we already been destroyed by a softirq or backlog? */
2266 if (state != TCP_CLOSE && sk->sk_state == TCP_CLOSE)
2267 goto out;
2268
2269 /* This is a (useful) BSD violating of the RFC. There is a
2270 * problem with TCP as specified in that the other end could
2271 * keep a socket open forever with no application left this end.
2272 * We use a 1 minute timeout (about the same as BSD) then kill
2273 * our end. If they send after that then tough - BUT: long enough
2274 * that we won't make the old 4*rto = almost no time - whoops
2275 * reset mistake.
2276 *
2277 * Nope, it was not mistake. It is really desired behaviour
2278 * f.e. on http servers, when such sockets are useless, but
2279 * consume significant resources. Let's do it with special
2280 * linger2 option. --ANK
2281 */
2282
2283 if (sk->sk_state == TCP_FIN_WAIT2) {
2284 struct tcp_sock *tp = tcp_sk(sk);
2285 if (tp->linger2 < 0) {
2286 tcp_set_state(sk, TCP_CLOSE);
2287 tcp_send_active_reset(sk, GFP_ATOMIC);
2288 __NET_INC_STATS(sock_net(sk),
2289 LINUX_MIB_TCPABORTONLINGER);
2290 } else {
2291 const int tmo = tcp_fin_time(sk);
2292
2293 if (tmo > TCP_TIMEWAIT_LEN) {
2294 inet_csk_reset_keepalive_timer(sk,
2295 tmo - TCP_TIMEWAIT_LEN);
2296 } else {
2297 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
2298 goto out;
2299 }
2300 }
2301 }
2302 if (sk->sk_state != TCP_CLOSE) {
2303 sk_mem_reclaim(sk);
2304 if (tcp_check_oom(sk, 0)) {
2305 tcp_set_state(sk, TCP_CLOSE);
2306 tcp_send_active_reset(sk, GFP_ATOMIC);
2307 __NET_INC_STATS(sock_net(sk),
2308 LINUX_MIB_TCPABORTONMEMORY);
2309 } else if (!check_net(sock_net(sk))) {
2310 /* Not possible to send reset; just close */
2311 tcp_set_state(sk, TCP_CLOSE);
2312 }
2313 }
2314
2315 if (sk->sk_state == TCP_CLOSE) {
2316 struct request_sock *req = tcp_sk(sk)->fastopen_rsk;
2317 /* We could get here with a non-NULL req if the socket is
2318 * aborted (e.g., closed with unread data) before 3WHS
2319 * finishes.
2320 */
2321 if (req)
2322 reqsk_fastopen_remove(sk, req, false);
2323 inet_csk_destroy_sock(sk);
2324 }
2325 /* Otherwise, socket is reprieved until protocol close. */
2326
2327out:
2328 bh_unlock_sock(sk);
2329 local_bh_enable();
2330 sock_put(sk);
2331}
2332EXPORT_SYMBOL(tcp_close);
2333
2334/* These states need RST on ABORT according to RFC793 */
2335
2336static inline bool tcp_need_reset(int state)
2337{
2338 return (1 << state) &
2339 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 |
2340 TCPF_FIN_WAIT2 | TCPF_SYN_RECV);
2341}
2342
2343static void tcp_rtx_queue_purge(struct sock *sk)
2344{
2345 struct rb_node *p = rb_first(&sk->tcp_rtx_queue);
2346
2347 while (p) {
2348 struct sk_buff *skb = rb_to_skb(p);
2349
2350 p = rb_next(p);
2351 /* Since we are deleting whole queue, no need to
2352 * list_del(&skb->tcp_tsorted_anchor)
2353 */
2354 tcp_rtx_queue_unlink(skb, sk);
2355 sk_wmem_free_skb(sk, skb);
2356 }
2357}
2358
2359void tcp_write_queue_purge(struct sock *sk)
2360{
2361 struct sk_buff *skb;
2362
2363 tcp_chrono_stop(sk, TCP_CHRONO_BUSY);
2364 while ((skb = __skb_dequeue(&sk->sk_write_queue)) != NULL) {
2365 tcp_skb_tsorted_anchor_cleanup(skb);
2366 sk_wmem_free_skb(sk, skb);
2367 }
2368 tcp_rtx_queue_purge(sk);
2369 INIT_LIST_HEAD(&tcp_sk(sk)->tsorted_sent_queue);
2370 sk_mem_reclaim(sk);
2371 tcp_clear_all_retrans_hints(tcp_sk(sk));
2372 tcp_sk(sk)->packets_out = 0;
2373}
2374
2375int tcp_disconnect(struct sock *sk, int flags)
2376{
2377 struct inet_sock *inet = inet_sk(sk);
2378 struct inet_connection_sock *icsk = inet_csk(sk);
2379 struct tcp_sock *tp = tcp_sk(sk);
2380 int err = 0;
2381 int old_state = sk->sk_state;
2382
2383 if (old_state != TCP_CLOSE)
2384 tcp_set_state(sk, TCP_CLOSE);
2385
2386 /* ABORT function of RFC793 */
2387 if (old_state == TCP_LISTEN) {
2388 inet_csk_listen_stop(sk);
2389 } else if (unlikely(tp->repair)) {
2390 sk->sk_err = ECONNABORTED;
2391 } else if (tcp_need_reset(old_state) ||
2392 (tp->snd_nxt != tp->write_seq &&
2393 (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK))) {
2394 /* The last check adjusts for discrepancy of Linux wrt. RFC
2395 * states
2396 */
2397 tcp_send_active_reset(sk, gfp_any());
2398 sk->sk_err = ECONNRESET;
2399 } else if (old_state == TCP_SYN_SENT)
2400 sk->sk_err = ECONNRESET;
2401
2402 tcp_clear_xmit_timers(sk);
2403 __skb_queue_purge(&sk->sk_receive_queue);
2404 tcp_write_queue_purge(sk);
2405 tcp_fastopen_active_disable_ofo_check(sk);
2406 skb_rbtree_purge(&tp->out_of_order_queue);
2407
2408 inet->inet_dport = 0;
2409
2410 if (!(sk->sk_userlocks & SOCK_BINDADDR_LOCK))
2411 inet_reset_saddr(sk);
2412
2413 sk->sk_shutdown = 0;
2414 sock_reset_flag(sk, SOCK_DONE);
2415 tp->srtt_us = 0;
2416 tp->write_seq += tp->max_window + 2;
2417 if (tp->write_seq == 0)
2418 tp->write_seq = 1;
2419 icsk->icsk_backoff = 0;
2420 tp->snd_cwnd = 2;
2421 icsk->icsk_probes_out = 0;
2422 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
2423 tp->snd_cwnd_cnt = 0;
2424 tp->window_clamp = 0;
2425 tcp_set_ca_state(sk, TCP_CA_Open);
2426 tp->is_sack_reneg = 0;
2427 tcp_clear_retrans(tp);
2428 inet_csk_delack_init(sk);
2429 /* Initialize rcv_mss to TCP_MIN_MSS to avoid division by 0
2430 * issue in __tcp_select_window()
2431 */
2432 icsk->icsk_ack.rcv_mss = TCP_MIN_MSS;
2433 memset(&tp->rx_opt, 0, sizeof(tp->rx_opt));
2434 __sk_dst_reset(sk);
2435 dst_release(sk->sk_rx_dst);
2436 sk->sk_rx_dst = NULL;
2437 tcp_saved_syn_free(tp);
2438
2439 /* Clean up fastopen related fields */
2440 tcp_free_fastopen_req(tp);
2441 inet->defer_connect = 0;
2442
2443 WARN_ON(inet->inet_num && !icsk->icsk_bind_hash);
2444
2445 if (sk->sk_frag.page) {
2446 put_page(sk->sk_frag.page);
2447 sk->sk_frag.page = NULL;
2448 sk->sk_frag.offset = 0;
2449 }
2450
2451 sk->sk_error_report(sk);
2452 return err;
2453}
2454EXPORT_SYMBOL(tcp_disconnect);
2455
2456static inline bool tcp_can_repair_sock(const struct sock *sk)
2457{
2458 return ns_capable(sock_net(sk)->user_ns, CAP_NET_ADMIN) &&
2459 (sk->sk_state != TCP_LISTEN);
2460}
2461
2462static int tcp_repair_set_window(struct tcp_sock *tp, char __user *optbuf, int len)
2463{
2464 struct tcp_repair_window opt;
2465
2466 if (!tp->repair)
2467 return -EPERM;
2468
2469 if (len != sizeof(opt))
2470 return -EINVAL;
2471
2472 if (copy_from_user(&opt, optbuf, sizeof(opt)))
2473 return -EFAULT;
2474
2475 if (opt.max_window < opt.snd_wnd)
2476 return -EINVAL;
2477
2478 if (after(opt.snd_wl1, tp->rcv_nxt + opt.rcv_wnd))
2479 return -EINVAL;
2480
2481 if (after(opt.rcv_wup, tp->rcv_nxt))
2482 return -EINVAL;
2483
2484 tp->snd_wl1 = opt.snd_wl1;
2485 tp->snd_wnd = opt.snd_wnd;
2486 tp->max_window = opt.max_window;
2487
2488 tp->rcv_wnd = opt.rcv_wnd;
2489 tp->rcv_wup = opt.rcv_wup;
2490
2491 return 0;
2492}
2493
2494static int tcp_repair_options_est(struct sock *sk,
2495 struct tcp_repair_opt __user *optbuf, unsigned int len)
2496{
2497 struct tcp_sock *tp = tcp_sk(sk);
2498 struct tcp_repair_opt opt;
2499
2500 while (len >= sizeof(opt)) {
2501 if (copy_from_user(&opt, optbuf, sizeof(opt)))
2502 return -EFAULT;
2503
2504 optbuf++;
2505 len -= sizeof(opt);
2506
2507 switch (opt.opt_code) {
2508 case TCPOPT_MSS:
2509 tp->rx_opt.mss_clamp = opt.opt_val;
2510 tcp_mtup_init(sk);
2511 break;
2512 case TCPOPT_WINDOW:
2513 {
2514 u16 snd_wscale = opt.opt_val & 0xFFFF;
2515 u16 rcv_wscale = opt.opt_val >> 16;
2516
2517 if (snd_wscale > TCP_MAX_WSCALE || rcv_wscale > TCP_MAX_WSCALE)
2518 return -EFBIG;
2519
2520 tp->rx_opt.snd_wscale = snd_wscale;
2521 tp->rx_opt.rcv_wscale = rcv_wscale;
2522 tp->rx_opt.wscale_ok = 1;
2523 }
2524 break;
2525 case TCPOPT_SACK_PERM:
2526 if (opt.opt_val != 0)
2527 return -EINVAL;
2528
2529 tp->rx_opt.sack_ok |= TCP_SACK_SEEN;
2530 break;
2531 case TCPOPT_TIMESTAMP:
2532 if (opt.opt_val != 0)
2533 return -EINVAL;
2534
2535 tp->rx_opt.tstamp_ok = 1;
2536 break;
2537 }
2538 }
2539
2540 return 0;
2541}
2542
2543/*
2544 * Socket option code for TCP.
2545 */
2546static int do_tcp_setsockopt(struct sock *sk, int level,
2547 int optname, char __user *optval, unsigned int optlen)
2548{
2549 struct tcp_sock *tp = tcp_sk(sk);
2550 struct inet_connection_sock *icsk = inet_csk(sk);
2551 struct net *net = sock_net(sk);
2552 int val;
2553 int err = 0;
2554
2555 /* These are data/string values, all the others are ints */
2556 switch (optname) {
2557 case TCP_CONGESTION: {
2558 char name[TCP_CA_NAME_MAX];
2559
2560 if (optlen < 1)
2561 return -EINVAL;
2562
2563 val = strncpy_from_user(name, optval,
2564 min_t(long, TCP_CA_NAME_MAX-1, optlen));
2565 if (val < 0)
2566 return -EFAULT;
2567 name[val] = 0;
2568
2569 lock_sock(sk);
2570 err = tcp_set_congestion_control(sk, name, true, true);
2571 release_sock(sk);
2572 return err;
2573 }
2574 case TCP_ULP: {
2575 char name[TCP_ULP_NAME_MAX];
2576
2577 if (optlen < 1)
2578 return -EINVAL;
2579
2580 val = strncpy_from_user(name, optval,
2581 min_t(long, TCP_ULP_NAME_MAX - 1,
2582 optlen));
2583 if (val < 0)
2584 return -EFAULT;
2585 name[val] = 0;
2586
2587 lock_sock(sk);
2588 err = tcp_set_ulp(sk, name);
2589 release_sock(sk);
2590 return err;
2591 }
2592 case TCP_FASTOPEN_KEY: {
2593 __u8 key[TCP_FASTOPEN_KEY_LENGTH];
2594
2595 if (optlen != sizeof(key))
2596 return -EINVAL;
2597
2598 if (copy_from_user(key, optval, optlen))
2599 return -EFAULT;
2600
2601 return tcp_fastopen_reset_cipher(net, sk, key, sizeof(key));
2602 }
2603 default:
2604 /* fallthru */
2605 break;
2606 }
2607
2608 if (optlen < sizeof(int))
2609 return -EINVAL;
2610
2611 if (get_user(val, (int __user *)optval))
2612 return -EFAULT;
2613
2614 lock_sock(sk);
2615
2616 switch (optname) {
2617 case TCP_MAXSEG:
2618 /* Values greater than interface MTU won't take effect. However
2619 * at the point when this call is done we typically don't yet
2620 * know which interface is going to be used
2621 */
2622 if (val && (val < TCP_MIN_MSS || val > MAX_TCP_WINDOW)) {
2623 err = -EINVAL;
2624 break;
2625 }
2626 tp->rx_opt.user_mss = val;
2627 break;
2628
2629 case TCP_NODELAY:
2630 if (val) {
2631 /* TCP_NODELAY is weaker than TCP_CORK, so that
2632 * this option on corked socket is remembered, but
2633 * it is not activated until cork is cleared.
2634 *
2635 * However, when TCP_NODELAY is set we make
2636 * an explicit push, which overrides even TCP_CORK
2637 * for currently queued segments.
2638 */
2639 tp->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH;
2640 tcp_push_pending_frames(sk);
2641 } else {
2642 tp->nonagle &= ~TCP_NAGLE_OFF;
2643 }
2644 break;
2645
2646 case TCP_THIN_LINEAR_TIMEOUTS:
2647 if (val < 0 || val > 1)
2648 err = -EINVAL;
2649 else
2650 tp->thin_lto = val;
2651 break;
2652
2653 case TCP_THIN_DUPACK:
2654 if (val < 0 || val > 1)
2655 err = -EINVAL;
2656 break;
2657
2658 case TCP_REPAIR:
2659 if (!tcp_can_repair_sock(sk))
2660 err = -EPERM;
2661 else if (val == 1) {
2662 tp->repair = 1;
2663 sk->sk_reuse = SK_FORCE_REUSE;
2664 tp->repair_queue = TCP_NO_QUEUE;
2665 } else if (val == 0) {
2666 tp->repair = 0;
2667 sk->sk_reuse = SK_NO_REUSE;
2668 tcp_send_window_probe(sk);
2669 } else
2670 err = -EINVAL;
2671
2672 break;
2673
2674 case TCP_REPAIR_QUEUE:
2675 if (!tp->repair)
2676 err = -EPERM;
2677 else if ((unsigned int)val < TCP_QUEUES_NR)
2678 tp->repair_queue = val;
2679 else
2680 err = -EINVAL;
2681 break;
2682
2683 case TCP_QUEUE_SEQ:
2684 if (sk->sk_state != TCP_CLOSE)
2685 err = -EPERM;
2686 else if (tp->repair_queue == TCP_SEND_QUEUE)
2687 tp->write_seq = val;
2688 else if (tp->repair_queue == TCP_RECV_QUEUE)
2689 tp->rcv_nxt = val;
2690 else
2691 err = -EINVAL;
2692 break;
2693
2694 case TCP_REPAIR_OPTIONS:
2695 if (!tp->repair)
2696 err = -EINVAL;
2697 else if (sk->sk_state == TCP_ESTABLISHED)
2698 err = tcp_repair_options_est(sk,
2699 (struct tcp_repair_opt __user *)optval,
2700 optlen);
2701 else
2702 err = -EPERM;
2703 break;
2704
2705 case TCP_CORK:
2706 /* When set indicates to always queue non-full frames.
2707 * Later the user clears this option and we transmit
2708 * any pending partial frames in the queue. This is
2709 * meant to be used alongside sendfile() to get properly
2710 * filled frames when the user (for example) must write
2711 * out headers with a write() call first and then use
2712 * sendfile to send out the data parts.
2713 *
2714 * TCP_CORK can be set together with TCP_NODELAY and it is
2715 * stronger than TCP_NODELAY.
2716 */
2717 if (val) {
2718 tp->nonagle |= TCP_NAGLE_CORK;
2719 } else {
2720 tp->nonagle &= ~TCP_NAGLE_CORK;
2721 if (tp->nonagle&TCP_NAGLE_OFF)
2722 tp->nonagle |= TCP_NAGLE_PUSH;
2723 tcp_push_pending_frames(sk);
2724 }
2725 break;
2726
2727 case TCP_KEEPIDLE:
2728 if (val < 1 || val > MAX_TCP_KEEPIDLE)
2729 err = -EINVAL;
2730 else {
2731 tp->keepalive_time = val * HZ;
2732 if (sock_flag(sk, SOCK_KEEPOPEN) &&
2733 !((1 << sk->sk_state) &
2734 (TCPF_CLOSE | TCPF_LISTEN))) {
2735 u32 elapsed = keepalive_time_elapsed(tp);
2736 if (tp->keepalive_time > elapsed)
2737 elapsed = tp->keepalive_time - elapsed;
2738 else
2739 elapsed = 0;
2740 inet_csk_reset_keepalive_timer(sk, elapsed);
2741 }
2742 }
2743 break;
2744 case TCP_KEEPINTVL:
2745 if (val < 1 || val > MAX_TCP_KEEPINTVL)
2746 err = -EINVAL;
2747 else
2748 tp->keepalive_intvl = val * HZ;
2749 break;
2750 case TCP_KEEPCNT:
2751 if (val < 1 || val > MAX_TCP_KEEPCNT)
2752 err = -EINVAL;
2753 else
2754 tp->keepalive_probes = val;
2755 break;
2756 case TCP_SYNCNT:
2757 if (val < 1 || val > MAX_TCP_SYNCNT)
2758 err = -EINVAL;
2759 else
2760 icsk->icsk_syn_retries = val;
2761 break;
2762
2763 case TCP_SAVE_SYN:
2764 if (val < 0 || val > 1)
2765 err = -EINVAL;
2766 else
2767 tp->save_syn = val;
2768 break;
2769
2770 case TCP_LINGER2:
2771 if (val < 0)
2772 tp->linger2 = -1;
2773 else if (val > net->ipv4.sysctl_tcp_fin_timeout / HZ)
2774 tp->linger2 = 0;
2775 else
2776 tp->linger2 = val * HZ;
2777 break;
2778
2779 case TCP_DEFER_ACCEPT:
2780 /* Translate value in seconds to number of retransmits */
2781 icsk->icsk_accept_queue.rskq_defer_accept =
2782 secs_to_retrans(val, TCP_TIMEOUT_INIT / HZ,
2783 TCP_RTO_MAX / HZ);
2784 break;
2785
2786 case TCP_WINDOW_CLAMP:
2787 if (!val) {
2788 if (sk->sk_state != TCP_CLOSE) {
2789 err = -EINVAL;
2790 break;
2791 }
2792 tp->window_clamp = 0;
2793 } else
2794 tp->window_clamp = val < SOCK_MIN_RCVBUF / 2 ?
2795 SOCK_MIN_RCVBUF / 2 : val;
2796 break;
2797
2798 case TCP_QUICKACK:
2799 if (!val) {
2800 icsk->icsk_ack.pingpong = 1;
2801 } else {
2802 icsk->icsk_ack.pingpong = 0;
2803 if ((1 << sk->sk_state) &
2804 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) &&
2805 inet_csk_ack_scheduled(sk)) {
2806 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
2807 tcp_cleanup_rbuf(sk, 1);
2808 if (!(val & 1))
2809 icsk->icsk_ack.pingpong = 1;
2810 }
2811 }
2812 break;
2813
2814#ifdef CONFIG_TCP_MD5SIG
2815 case TCP_MD5SIG:
2816 case TCP_MD5SIG_EXT:
2817 if ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN))
2818 err = tp->af_specific->md5_parse(sk, optname, optval, optlen);
2819 else
2820 err = -EINVAL;
2821 break;
2822#endif
2823 case TCP_USER_TIMEOUT:
2824 /* Cap the max time in ms TCP will retry or probe the window
2825 * before giving up and aborting (ETIMEDOUT) a connection.
2826 */
2827 if (val < 0)
2828 err = -EINVAL;
2829 else
2830 icsk->icsk_user_timeout = msecs_to_jiffies(val);
2831 break;
2832
2833 case TCP_FASTOPEN:
2834 if (val >= 0 && ((1 << sk->sk_state) & (TCPF_CLOSE |
2835 TCPF_LISTEN))) {
2836 tcp_fastopen_init_key_once(net);
2837
2838 fastopen_queue_tune(sk, val);
2839 } else {
2840 err = -EINVAL;
2841 }
2842 break;
2843 case TCP_FASTOPEN_CONNECT:
2844 if (val > 1 || val < 0) {
2845 err = -EINVAL;
2846 } else if (net->ipv4.sysctl_tcp_fastopen & TFO_CLIENT_ENABLE) {
2847 if (sk->sk_state == TCP_CLOSE)
2848 tp->fastopen_connect = val;
2849 else
2850 err = -EINVAL;
2851 } else {
2852 err = -EOPNOTSUPP;
2853 }
2854 break;
2855 case TCP_FASTOPEN_NO_COOKIE:
2856 if (val > 1 || val < 0)
2857 err = -EINVAL;
2858 else if (!((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
2859 err = -EINVAL;
2860 else
2861 tp->fastopen_no_cookie = val;
2862 break;
2863 case TCP_TIMESTAMP:
2864 if (!tp->repair)
2865 err = -EPERM;
2866 else
2867 tp->tsoffset = val - tcp_time_stamp_raw();
2868 break;
2869 case TCP_REPAIR_WINDOW:
2870 err = tcp_repair_set_window(tp, optval, optlen);
2871 break;
2872 case TCP_NOTSENT_LOWAT:
2873 tp->notsent_lowat = val;
2874 sk->sk_write_space(sk);
2875 break;
2876 default:
2877 err = -ENOPROTOOPT;
2878 break;
2879 }
2880
2881 release_sock(sk);
2882 return err;
2883}
2884
2885int tcp_setsockopt(struct sock *sk, int level, int optname, char __user *optval,
2886 unsigned int optlen)
2887{
2888 const struct inet_connection_sock *icsk = inet_csk(sk);
2889
2890 if (level != SOL_TCP)
2891 return icsk->icsk_af_ops->setsockopt(sk, level, optname,
2892 optval, optlen);
2893 return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2894}
2895EXPORT_SYMBOL(tcp_setsockopt);
2896
2897#ifdef CONFIG_COMPAT
2898int compat_tcp_setsockopt(struct sock *sk, int level, int optname,
2899 char __user *optval, unsigned int optlen)
2900{
2901 if (level != SOL_TCP)
2902 return inet_csk_compat_setsockopt(sk, level, optname,
2903 optval, optlen);
2904 return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2905}
2906EXPORT_SYMBOL(compat_tcp_setsockopt);
2907#endif
2908
2909static void tcp_get_info_chrono_stats(const struct tcp_sock *tp,
2910 struct tcp_info *info)
2911{
2912 u64 stats[__TCP_CHRONO_MAX], total = 0;
2913 enum tcp_chrono i;
2914
2915 for (i = TCP_CHRONO_BUSY; i < __TCP_CHRONO_MAX; ++i) {
2916 stats[i] = tp->chrono_stat[i - 1];
2917 if (i == tp->chrono_type)
2918 stats[i] += tcp_jiffies32 - tp->chrono_start;
2919 stats[i] *= USEC_PER_SEC / HZ;
2920 total += stats[i];
2921 }
2922
2923 info->tcpi_busy_time = total;
2924 info->tcpi_rwnd_limited = stats[TCP_CHRONO_RWND_LIMITED];
2925 info->tcpi_sndbuf_limited = stats[TCP_CHRONO_SNDBUF_LIMITED];
2926}
2927
2928/* Return information about state of tcp endpoint in API format. */
2929void tcp_get_info(struct sock *sk, struct tcp_info *info)
2930{
2931 const struct tcp_sock *tp = tcp_sk(sk); /* iff sk_type == SOCK_STREAM */
2932 const struct inet_connection_sock *icsk = inet_csk(sk);
2933 u32 now;
2934 u64 rate64;
2935 bool slow;
2936 u32 rate;
2937
2938 memset(info, 0, sizeof(*info));
2939 if (sk->sk_type != SOCK_STREAM)
2940 return;
2941
2942 info->tcpi_state = inet_sk_state_load(sk);
2943
2944 /* Report meaningful fields for all TCP states, including listeners */
2945 rate = READ_ONCE(sk->sk_pacing_rate);
2946 rate64 = rate != ~0U ? rate : ~0ULL;
2947 info->tcpi_pacing_rate = rate64;
2948
2949 rate = READ_ONCE(sk->sk_max_pacing_rate);
2950 rate64 = rate != ~0U ? rate : ~0ULL;
2951 info->tcpi_max_pacing_rate = rate64;
2952
2953 info->tcpi_reordering = tp->reordering;
2954 info->tcpi_snd_cwnd = tp->snd_cwnd;
2955
2956 if (info->tcpi_state == TCP_LISTEN) {
2957 /* listeners aliased fields :
2958 * tcpi_unacked -> Number of children ready for accept()
2959 * tcpi_sacked -> max backlog
2960 */
2961 info->tcpi_unacked = sk->sk_ack_backlog;
2962 info->tcpi_sacked = sk->sk_max_ack_backlog;
2963 return;
2964 }
2965
2966 slow = lock_sock_fast(sk);
2967
2968 info->tcpi_ca_state = icsk->icsk_ca_state;
2969 info->tcpi_retransmits = icsk->icsk_retransmits;
2970 info->tcpi_probes = icsk->icsk_probes_out;
2971 info->tcpi_backoff = icsk->icsk_backoff;
2972
2973 if (tp->rx_opt.tstamp_ok)
2974 info->tcpi_options |= TCPI_OPT_TIMESTAMPS;
2975 if (tcp_is_sack(tp))
2976 info->tcpi_options |= TCPI_OPT_SACK;
2977 if (tp->rx_opt.wscale_ok) {
2978 info->tcpi_options |= TCPI_OPT_WSCALE;
2979 info->tcpi_snd_wscale = tp->rx_opt.snd_wscale;
2980 info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale;
2981 }
2982
2983 if (tp->ecn_flags & TCP_ECN_OK)
2984 info->tcpi_options |= TCPI_OPT_ECN;
2985 if (tp->ecn_flags & TCP_ECN_SEEN)
2986 info->tcpi_options |= TCPI_OPT_ECN_SEEN;
2987 if (tp->syn_data_acked)
2988 info->tcpi_options |= TCPI_OPT_SYN_DATA;
2989
2990 info->tcpi_rto = jiffies_to_usecs(icsk->icsk_rto);
2991 info->tcpi_ato = jiffies_to_usecs(icsk->icsk_ack.ato);
2992 info->tcpi_snd_mss = tp->mss_cache;
2993 info->tcpi_rcv_mss = icsk->icsk_ack.rcv_mss;
2994
2995 info->tcpi_unacked = tp->packets_out;
2996 info->tcpi_sacked = tp->sacked_out;
2997
2998 info->tcpi_lost = tp->lost_out;
2999 info->tcpi_retrans = tp->retrans_out;
3000
3001 now = tcp_jiffies32;
3002 info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime);
3003 info->tcpi_last_data_recv = jiffies_to_msecs(now - icsk->icsk_ack.lrcvtime);
3004 info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp);
3005
3006 info->tcpi_pmtu = icsk->icsk_pmtu_cookie;
3007 info->tcpi_rcv_ssthresh = tp->rcv_ssthresh;
3008 info->tcpi_rtt = tp->srtt_us >> 3;
3009 info->tcpi_rttvar = tp->mdev_us >> 2;
3010 info->tcpi_snd_ssthresh = tp->snd_ssthresh;
3011 info->tcpi_advmss = tp->advmss;
3012
3013 info->tcpi_rcv_rtt = tp->rcv_rtt_est.rtt_us >> 3;
3014 info->tcpi_rcv_space = tp->rcvq_space.space;
3015
3016 info->tcpi_total_retrans = tp->total_retrans;
3017
3018 info->tcpi_bytes_acked = tp->bytes_acked;
3019 info->tcpi_bytes_received = tp->bytes_received;
3020 info->tcpi_notsent_bytes = max_t(int, 0, tp->write_seq - tp->snd_nxt);
3021 tcp_get_info_chrono_stats(tp, info);
3022
3023 info->tcpi_segs_out = tp->segs_out;
3024 info->tcpi_segs_in = tp->segs_in;
3025
3026 info->tcpi_min_rtt = tcp_min_rtt(tp);
3027 info->tcpi_data_segs_in = tp->data_segs_in;
3028 info->tcpi_data_segs_out = tp->data_segs_out;
3029
3030 info->tcpi_delivery_rate_app_limited = tp->rate_app_limited ? 1 : 0;
3031 rate64 = tcp_compute_delivery_rate(tp);
3032 if (rate64)
3033 info->tcpi_delivery_rate = rate64;
3034 unlock_sock_fast(sk, slow);
3035}
3036EXPORT_SYMBOL_GPL(tcp_get_info);
3037
3038struct sk_buff *tcp_get_timestamping_opt_stats(const struct sock *sk)
3039{
3040 const struct tcp_sock *tp = tcp_sk(sk);
3041 struct sk_buff *stats;
3042 struct tcp_info info;
3043 u64 rate64;
3044 u32 rate;
3045
3046 stats = alloc_skb(7 * nla_total_size_64bit(sizeof(u64)) +
3047 5 * nla_total_size(sizeof(u32)) +
3048 3 * nla_total_size(sizeof(u8)), GFP_ATOMIC);
3049 if (!stats)
3050 return NULL;
3051
3052 tcp_get_info_chrono_stats(tp, &info);
3053 nla_put_u64_64bit(stats, TCP_NLA_BUSY,
3054 info.tcpi_busy_time, TCP_NLA_PAD);
3055 nla_put_u64_64bit(stats, TCP_NLA_RWND_LIMITED,
3056 info.tcpi_rwnd_limited, TCP_NLA_PAD);
3057 nla_put_u64_64bit(stats, TCP_NLA_SNDBUF_LIMITED,
3058 info.tcpi_sndbuf_limited, TCP_NLA_PAD);
3059 nla_put_u64_64bit(stats, TCP_NLA_DATA_SEGS_OUT,
3060 tp->data_segs_out, TCP_NLA_PAD);
3061 nla_put_u64_64bit(stats, TCP_NLA_TOTAL_RETRANS,
3062 tp->total_retrans, TCP_NLA_PAD);
3063
3064 rate = READ_ONCE(sk->sk_pacing_rate);
3065 rate64 = rate != ~0U ? rate : ~0ULL;
3066 nla_put_u64_64bit(stats, TCP_NLA_PACING_RATE, rate64, TCP_NLA_PAD);
3067
3068 rate64 = tcp_compute_delivery_rate(tp);
3069 nla_put_u64_64bit(stats, TCP_NLA_DELIVERY_RATE, rate64, TCP_NLA_PAD);
3070
3071 nla_put_u32(stats, TCP_NLA_SND_CWND, tp->snd_cwnd);
3072 nla_put_u32(stats, TCP_NLA_REORDERING, tp->reordering);
3073 nla_put_u32(stats, TCP_NLA_MIN_RTT, tcp_min_rtt(tp));
3074
3075 nla_put_u8(stats, TCP_NLA_RECUR_RETRANS, inet_csk(sk)->icsk_retransmits);
3076 nla_put_u8(stats, TCP_NLA_DELIVERY_RATE_APP_LMT, !!tp->rate_app_limited);
3077 nla_put_u32(stats, TCP_NLA_SND_SSTHRESH, tp->snd_ssthresh);
3078
3079 nla_put_u32(stats, TCP_NLA_SNDQ_SIZE, tp->write_seq - tp->snd_una);
3080 nla_put_u8(stats, TCP_NLA_CA_STATE, inet_csk(sk)->icsk_ca_state);
3081 return stats;
3082}
3083
3084static int do_tcp_getsockopt(struct sock *sk, int level,
3085 int optname, char __user *optval, int __user *optlen)
3086{
3087 struct inet_connection_sock *icsk = inet_csk(sk);
3088 struct tcp_sock *tp = tcp_sk(sk);
3089 struct net *net = sock_net(sk);
3090 int val, len;
3091
3092 if (get_user(len, optlen))
3093 return -EFAULT;
3094
3095 len = min_t(unsigned int, len, sizeof(int));
3096
3097 if (len < 0)
3098 return -EINVAL;
3099
3100 switch (optname) {
3101 case TCP_MAXSEG:
3102 val = tp->mss_cache;
3103 if (!val && ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
3104 val = tp->rx_opt.user_mss;
3105 if (tp->repair)
3106 val = tp->rx_opt.mss_clamp;
3107 break;
3108 case TCP_NODELAY:
3109 val = !!(tp->nonagle&TCP_NAGLE_OFF);
3110 break;
3111 case TCP_CORK:
3112 val = !!(tp->nonagle&TCP_NAGLE_CORK);
3113 break;
3114 case TCP_KEEPIDLE:
3115 val = keepalive_time_when(tp) / HZ;
3116 break;
3117 case TCP_KEEPINTVL:
3118 val = keepalive_intvl_when(tp) / HZ;
3119 break;
3120 case TCP_KEEPCNT:
3121 val = keepalive_probes(tp);
3122 break;
3123 case TCP_SYNCNT:
3124 val = icsk->icsk_syn_retries ? : net->ipv4.sysctl_tcp_syn_retries;
3125 break;
3126 case TCP_LINGER2:
3127 val = tp->linger2;
3128 if (val >= 0)
3129 val = (val ? : net->ipv4.sysctl_tcp_fin_timeout) / HZ;
3130 break;
3131 case TCP_DEFER_ACCEPT:
3132 val = retrans_to_secs(icsk->icsk_accept_queue.rskq_defer_accept,
3133 TCP_TIMEOUT_INIT / HZ, TCP_RTO_MAX / HZ);
3134 break;
3135 case TCP_WINDOW_CLAMP:
3136 val = tp->window_clamp;
3137 break;
3138 case TCP_INFO: {
3139 struct tcp_info info;
3140
3141 if (get_user(len, optlen))
3142 return -EFAULT;
3143
3144 tcp_get_info(sk, &info);
3145
3146 len = min_t(unsigned int, len, sizeof(info));
3147 if (put_user(len, optlen))
3148 return -EFAULT;
3149 if (copy_to_user(optval, &info, len))
3150 return -EFAULT;
3151 return 0;
3152 }
3153 case TCP_CC_INFO: {
3154 const struct tcp_congestion_ops *ca_ops;
3155 union tcp_cc_info info;
3156 size_t sz = 0;
3157 int attr;
3158
3159 if (get_user(len, optlen))
3160 return -EFAULT;
3161
3162 ca_ops = icsk->icsk_ca_ops;
3163 if (ca_ops && ca_ops->get_info)
3164 sz = ca_ops->get_info(sk, ~0U, &attr, &info);
3165
3166 len = min_t(unsigned int, len, sz);
3167 if (put_user(len, optlen))
3168 return -EFAULT;
3169 if (copy_to_user(optval, &info, len))
3170 return -EFAULT;
3171 return 0;
3172 }
3173 case TCP_QUICKACK:
3174 val = !icsk->icsk_ack.pingpong;
3175 break;
3176
3177 case TCP_CONGESTION:
3178 if (get_user(len, optlen))
3179 return -EFAULT;
3180 len = min_t(unsigned int, len, TCP_CA_NAME_MAX);
3181 if (put_user(len, optlen))
3182 return -EFAULT;
3183 if (copy_to_user(optval, icsk->icsk_ca_ops->name, len))
3184 return -EFAULT;
3185 return 0;
3186
3187 case TCP_ULP:
3188 if (get_user(len, optlen))
3189 return -EFAULT;
3190 len = min_t(unsigned int, len, TCP_ULP_NAME_MAX);
3191 if (!icsk->icsk_ulp_ops) {
3192 if (put_user(0, optlen))
3193 return -EFAULT;
3194 return 0;
3195 }
3196 if (put_user(len, optlen))
3197 return -EFAULT;
3198 if (copy_to_user(optval, icsk->icsk_ulp_ops->name, len))
3199 return -EFAULT;
3200 return 0;
3201
3202 case TCP_FASTOPEN_KEY: {
3203 __u8 key[TCP_FASTOPEN_KEY_LENGTH];
3204 struct tcp_fastopen_context *ctx;
3205
3206 if (get_user(len, optlen))
3207 return -EFAULT;
3208
3209 rcu_read_lock();
3210 ctx = rcu_dereference(icsk->icsk_accept_queue.fastopenq.ctx);
3211 if (ctx)
3212 memcpy(key, ctx->key, sizeof(key));
3213 else
3214 len = 0;
3215 rcu_read_unlock();
3216
3217 len = min_t(unsigned int, len, sizeof(key));
3218 if (put_user(len, optlen))
3219 return -EFAULT;
3220 if (copy_to_user(optval, key, len))
3221 return -EFAULT;
3222 return 0;
3223 }
3224 case TCP_THIN_LINEAR_TIMEOUTS:
3225 val = tp->thin_lto;
3226 break;
3227
3228 case TCP_THIN_DUPACK:
3229 val = 0;
3230 break;
3231
3232 case TCP_REPAIR:
3233 val = tp->repair;
3234 break;
3235
3236 case TCP_REPAIR_QUEUE:
3237 if (tp->repair)
3238 val = tp->repair_queue;
3239 else
3240 return -EINVAL;
3241 break;
3242
3243 case TCP_REPAIR_WINDOW: {
3244 struct tcp_repair_window opt;
3245
3246 if (get_user(len, optlen))
3247 return -EFAULT;
3248
3249 if (len != sizeof(opt))
3250 return -EINVAL;
3251
3252 if (!tp->repair)
3253 return -EPERM;
3254
3255 opt.snd_wl1 = tp->snd_wl1;
3256 opt.snd_wnd = tp->snd_wnd;
3257 opt.max_window = tp->max_window;
3258 opt.rcv_wnd = tp->rcv_wnd;
3259 opt.rcv_wup = tp->rcv_wup;
3260
3261 if (copy_to_user(optval, &opt, len))
3262 return -EFAULT;
3263 return 0;
3264 }
3265 case TCP_QUEUE_SEQ:
3266 if (tp->repair_queue == TCP_SEND_QUEUE)
3267 val = tp->write_seq;
3268 else if (tp->repair_queue == TCP_RECV_QUEUE)
3269 val = tp->rcv_nxt;
3270 else
3271 return -EINVAL;
3272 break;
3273
3274 case TCP_USER_TIMEOUT:
3275 val = jiffies_to_msecs(icsk->icsk_user_timeout);
3276 break;
3277
3278 case TCP_FASTOPEN:
3279 val = icsk->icsk_accept_queue.fastopenq.max_qlen;
3280 break;
3281
3282 case TCP_FASTOPEN_CONNECT:
3283 val = tp->fastopen_connect;
3284 break;
3285
3286 case TCP_FASTOPEN_NO_COOKIE:
3287 val = tp->fastopen_no_cookie;
3288 break;
3289
3290 case TCP_TIMESTAMP:
3291 val = tcp_time_stamp_raw() + tp->tsoffset;
3292 break;
3293 case TCP_NOTSENT_LOWAT:
3294 val = tp->notsent_lowat;
3295 break;
3296 case TCP_SAVE_SYN:
3297 val = tp->save_syn;
3298 break;
3299 case TCP_SAVED_SYN: {
3300 if (get_user(len, optlen))
3301 return -EFAULT;
3302
3303 lock_sock(sk);
3304 if (tp->saved_syn) {
3305 if (len < tp->saved_syn[0]) {
3306 if (put_user(tp->saved_syn[0], optlen)) {
3307 release_sock(sk);
3308 return -EFAULT;
3309 }
3310 release_sock(sk);
3311 return -EINVAL;
3312 }
3313 len = tp->saved_syn[0];
3314 if (put_user(len, optlen)) {
3315 release_sock(sk);
3316 return -EFAULT;
3317 }
3318 if (copy_to_user(optval, tp->saved_syn + 1, len)) {
3319 release_sock(sk);
3320 return -EFAULT;
3321 }
3322 tcp_saved_syn_free(tp);
3323 release_sock(sk);
3324 } else {
3325 release_sock(sk);
3326 len = 0;
3327 if (put_user(len, optlen))
3328 return -EFAULT;
3329 }
3330 return 0;
3331 }
3332 default:
3333 return -ENOPROTOOPT;
3334 }
3335
3336 if (put_user(len, optlen))
3337 return -EFAULT;
3338 if (copy_to_user(optval, &val, len))
3339 return -EFAULT;
3340 return 0;
3341}
3342
3343int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval,
3344 int __user *optlen)
3345{
3346 struct inet_connection_sock *icsk = inet_csk(sk);
3347
3348 if (level != SOL_TCP)
3349 return icsk->icsk_af_ops->getsockopt(sk, level, optname,
3350 optval, optlen);
3351 return do_tcp_getsockopt(sk, level, optname, optval, optlen);
3352}
3353EXPORT_SYMBOL(tcp_getsockopt);
3354
3355#ifdef CONFIG_COMPAT
3356int compat_tcp_getsockopt(struct sock *sk, int level, int optname,
3357 char __user *optval, int __user *optlen)
3358{
3359 if (level != SOL_TCP)
3360 return inet_csk_compat_getsockopt(sk, level, optname,
3361 optval, optlen);
3362 return do_tcp_getsockopt(sk, level, optname, optval, optlen);
3363}
3364EXPORT_SYMBOL(compat_tcp_getsockopt);
3365#endif
3366
3367#ifdef CONFIG_TCP_MD5SIG
3368static DEFINE_PER_CPU(struct tcp_md5sig_pool, tcp_md5sig_pool);
3369static DEFINE_MUTEX(tcp_md5sig_mutex);
3370static bool tcp_md5sig_pool_populated = false;
3371
3372static void __tcp_alloc_md5sig_pool(void)
3373{
3374 struct crypto_ahash *hash;
3375 int cpu;
3376
3377 hash = crypto_alloc_ahash("md5", 0, CRYPTO_ALG_ASYNC);
3378 if (IS_ERR(hash))
3379 return;
3380
3381 for_each_possible_cpu(cpu) {
3382 void *scratch = per_cpu(tcp_md5sig_pool, cpu).scratch;
3383 struct ahash_request *req;
3384
3385 if (!scratch) {
3386 scratch = kmalloc_node(sizeof(union tcp_md5sum_block) +
3387 sizeof(struct tcphdr),
3388 GFP_KERNEL,
3389 cpu_to_node(cpu));
3390 if (!scratch)
3391 return;
3392 per_cpu(tcp_md5sig_pool, cpu).scratch = scratch;
3393 }
3394 if (per_cpu(tcp_md5sig_pool, cpu).md5_req)
3395 continue;
3396
3397 req = ahash_request_alloc(hash, GFP_KERNEL);
3398 if (!req)
3399 return;
3400
3401 ahash_request_set_callback(req, 0, NULL, NULL);
3402
3403 per_cpu(tcp_md5sig_pool, cpu).md5_req = req;
3404 }
3405 /* before setting tcp_md5sig_pool_populated, we must commit all writes
3406 * to memory. See smp_rmb() in tcp_get_md5sig_pool()
3407 */
3408 smp_wmb();
3409 tcp_md5sig_pool_populated = true;
3410}
3411
3412bool tcp_alloc_md5sig_pool(void)
3413{
3414 if (unlikely(!tcp_md5sig_pool_populated)) {
3415 mutex_lock(&tcp_md5sig_mutex);
3416
3417 if (!tcp_md5sig_pool_populated)
3418 __tcp_alloc_md5sig_pool();
3419
3420 mutex_unlock(&tcp_md5sig_mutex);
3421 }
3422 return tcp_md5sig_pool_populated;
3423}
3424EXPORT_SYMBOL(tcp_alloc_md5sig_pool);
3425
3426
3427/**
3428 * tcp_get_md5sig_pool - get md5sig_pool for this user
3429 *
3430 * We use percpu structure, so if we succeed, we exit with preemption
3431 * and BH disabled, to make sure another thread or softirq handling
3432 * wont try to get same context.
3433 */
3434struct tcp_md5sig_pool *tcp_get_md5sig_pool(void)
3435{
3436 local_bh_disable();
3437
3438 if (tcp_md5sig_pool_populated) {
3439 /* coupled with smp_wmb() in __tcp_alloc_md5sig_pool() */
3440 smp_rmb();
3441 return this_cpu_ptr(&tcp_md5sig_pool);
3442 }
3443 local_bh_enable();
3444 return NULL;
3445}
3446EXPORT_SYMBOL(tcp_get_md5sig_pool);
3447
3448int tcp_md5_hash_skb_data(struct tcp_md5sig_pool *hp,
3449 const struct sk_buff *skb, unsigned int header_len)
3450{
3451 struct scatterlist sg;
3452 const struct tcphdr *tp = tcp_hdr(skb);
3453 struct ahash_request *req = hp->md5_req;
3454 unsigned int i;
3455 const unsigned int head_data_len = skb_headlen(skb) > header_len ?
3456 skb_headlen(skb) - header_len : 0;
3457 const struct skb_shared_info *shi = skb_shinfo(skb);
3458 struct sk_buff *frag_iter;
3459
3460 sg_init_table(&sg, 1);
3461
3462 sg_set_buf(&sg, ((u8 *) tp) + header_len, head_data_len);
3463 ahash_request_set_crypt(req, &sg, NULL, head_data_len);
3464 if (crypto_ahash_update(req))
3465 return 1;
3466
3467 for (i = 0; i < shi->nr_frags; ++i) {
3468 const struct skb_frag_struct *f = &shi->frags[i];
3469 unsigned int offset = f->page_offset;
3470 struct page *page = skb_frag_page(f) + (offset >> PAGE_SHIFT);
3471
3472 sg_set_page(&sg, page, skb_frag_size(f),
3473 offset_in_page(offset));
3474 ahash_request_set_crypt(req, &sg, NULL, skb_frag_size(f));
3475 if (crypto_ahash_update(req))
3476 return 1;
3477 }
3478
3479 skb_walk_frags(skb, frag_iter)
3480 if (tcp_md5_hash_skb_data(hp, frag_iter, 0))
3481 return 1;
3482
3483 return 0;
3484}
3485EXPORT_SYMBOL(tcp_md5_hash_skb_data);
3486
3487int tcp_md5_hash_key(struct tcp_md5sig_pool *hp, const struct tcp_md5sig_key *key)
3488{
3489 struct scatterlist sg;
3490
3491 sg_init_one(&sg, key->key, key->keylen);
3492 ahash_request_set_crypt(hp->md5_req, &sg, NULL, key->keylen);
3493 return crypto_ahash_update(hp->md5_req);
3494}
3495EXPORT_SYMBOL(tcp_md5_hash_key);
3496
3497#endif
3498
3499void tcp_done(struct sock *sk)
3500{
3501 struct request_sock *req = tcp_sk(sk)->fastopen_rsk;
3502
3503 if (sk->sk_state == TCP_SYN_SENT || sk->sk_state == TCP_SYN_RECV)
3504 TCP_INC_STATS(sock_net(sk), TCP_MIB_ATTEMPTFAILS);
3505
3506 tcp_set_state(sk, TCP_CLOSE);
3507 tcp_clear_xmit_timers(sk);
3508 if (req)
3509 reqsk_fastopen_remove(sk, req, false);
3510
3511 sk->sk_shutdown = SHUTDOWN_MASK;
3512
3513 if (!sock_flag(sk, SOCK_DEAD))
3514 sk->sk_state_change(sk);
3515 else
3516 inet_csk_destroy_sock(sk);
3517}
3518EXPORT_SYMBOL_GPL(tcp_done);
3519
3520int tcp_abort(struct sock *sk, int err)
3521{
3522 if (!sk_fullsock(sk)) {
3523 if (sk->sk_state == TCP_NEW_SYN_RECV) {
3524 struct request_sock *req = inet_reqsk(sk);
3525
3526 local_bh_disable();
3527 inet_csk_reqsk_queue_drop_and_put(req->rsk_listener,
3528 req);
3529 local_bh_enable();
3530 return 0;
3531 }
3532 return -EOPNOTSUPP;
3533 }
3534
3535 /* Don't race with userspace socket closes such as tcp_close. */
3536 lock_sock(sk);
3537
3538 if (sk->sk_state == TCP_LISTEN) {
3539 tcp_set_state(sk, TCP_CLOSE);
3540 inet_csk_listen_stop(sk);
3541 }
3542
3543 /* Don't race with BH socket closes such as inet_csk_listen_stop. */
3544 local_bh_disable();
3545 bh_lock_sock(sk);
3546
3547 if (!sock_flag(sk, SOCK_DEAD)) {
3548 sk->sk_err = err;
3549 /* This barrier is coupled with smp_rmb() in tcp_poll() */
3550 smp_wmb();
3551 sk->sk_error_report(sk);
3552 if (tcp_need_reset(sk->sk_state))
3553 tcp_send_active_reset(sk, GFP_ATOMIC);
3554 tcp_done(sk);
3555 }
3556
3557 bh_unlock_sock(sk);
3558 local_bh_enable();
3559 tcp_write_queue_purge(sk);
3560 release_sock(sk);
3561 return 0;
3562}
3563EXPORT_SYMBOL_GPL(tcp_abort);
3564
3565extern struct tcp_congestion_ops tcp_reno;
3566
3567static __initdata unsigned long thash_entries;
3568static int __init set_thash_entries(char *str)
3569{
3570 ssize_t ret;
3571
3572 if (!str)
3573 return 0;
3574
3575 ret = kstrtoul(str, 0, &thash_entries);
3576 if (ret)
3577 return 0;
3578
3579 return 1;
3580}
3581__setup("thash_entries=", set_thash_entries);
3582
3583static void __init tcp_init_mem(void)
3584{
3585 unsigned long limit = nr_free_buffer_pages() / 16;
3586
3587 limit = max(limit, 128UL);
3588 sysctl_tcp_mem[0] = limit / 4 * 3; /* 4.68 % */
3589 sysctl_tcp_mem[1] = limit; /* 6.25 % */
3590 sysctl_tcp_mem[2] = sysctl_tcp_mem[0] * 2; /* 9.37 % */
3591}
3592
3593void __init tcp_init(void)
3594{
3595 int max_rshare, max_wshare, cnt;
3596 unsigned long limit;
3597 unsigned int i;
3598
3599 BUILD_BUG_ON(sizeof(struct tcp_skb_cb) >
3600 FIELD_SIZEOF(struct sk_buff, cb));
3601
3602 percpu_counter_init(&tcp_sockets_allocated, 0, GFP_KERNEL);
3603 percpu_counter_init(&tcp_orphan_count, 0, GFP_KERNEL);
3604 inet_hashinfo_init(&tcp_hashinfo);
3605 inet_hashinfo2_init(&tcp_hashinfo, "tcp_listen_portaddr_hash",
3606 thash_entries, 21, /* one slot per 2 MB*/
3607 0, 64 * 1024);
3608 tcp_hashinfo.bind_bucket_cachep =
3609 kmem_cache_create("tcp_bind_bucket",
3610 sizeof(struct inet_bind_bucket), 0,
3611 SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL);
3612
3613 /* Size and allocate the main established and bind bucket
3614 * hash tables.
3615 *
3616 * The methodology is similar to that of the buffer cache.
3617 */
3618 tcp_hashinfo.ehash =
3619 alloc_large_system_hash("TCP established",
3620 sizeof(struct inet_ehash_bucket),
3621 thash_entries,
3622 17, /* one slot per 128 KB of memory */
3623 0,
3624 NULL,
3625 &tcp_hashinfo.ehash_mask,
3626 0,
3627 thash_entries ? 0 : 512 * 1024);
3628 for (i = 0; i <= tcp_hashinfo.ehash_mask; i++)
3629 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].chain, i);
3630
3631 if (inet_ehash_locks_alloc(&tcp_hashinfo))
3632 panic("TCP: failed to alloc ehash_locks");
3633 tcp_hashinfo.bhash =
3634 alloc_large_system_hash("TCP bind",
3635 sizeof(struct inet_bind_hashbucket),
3636 tcp_hashinfo.ehash_mask + 1,
3637 17, /* one slot per 128 KB of memory */
3638 0,
3639 &tcp_hashinfo.bhash_size,
3640 NULL,
3641 0,
3642 64 * 1024);
3643 tcp_hashinfo.bhash_size = 1U << tcp_hashinfo.bhash_size;
3644 for (i = 0; i < tcp_hashinfo.bhash_size; i++) {
3645 spin_lock_init(&tcp_hashinfo.bhash[i].lock);
3646 INIT_HLIST_HEAD(&tcp_hashinfo.bhash[i].chain);
3647 }
3648
3649
3650 cnt = tcp_hashinfo.ehash_mask + 1;
3651 sysctl_tcp_max_orphans = cnt / 2;
3652
3653 tcp_init_mem();
3654 /* Set per-socket limits to no more than 1/128 the pressure threshold */
3655 limit = nr_free_buffer_pages() << (PAGE_SHIFT - 7);
3656 max_wshare = min(4UL*1024*1024, limit);
3657 max_rshare = min(6UL*1024*1024, limit);
3658
3659 init_net.ipv4.sysctl_tcp_wmem[0] = SK_MEM_QUANTUM;
3660 init_net.ipv4.sysctl_tcp_wmem[1] = 16*1024;
3661 init_net.ipv4.sysctl_tcp_wmem[2] = max(64*1024, max_wshare);
3662
3663 init_net.ipv4.sysctl_tcp_rmem[0] = SK_MEM_QUANTUM;
3664 init_net.ipv4.sysctl_tcp_rmem[1] = 87380;
3665 init_net.ipv4.sysctl_tcp_rmem[2] = max(87380, max_rshare);
3666
3667 pr_info("Hash tables configured (established %u bind %u)\n",
3668 tcp_hashinfo.ehash_mask + 1, tcp_hashinfo.bhash_size);
3669
3670 tcp_v4_init();
3671 tcp_metrics_init();
3672 BUG_ON(tcp_register_congestion_control(&tcp_reno) != 0);
3673 tcp_tasklet_init();
3674}