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 <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}
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/tcp.h>
274#include <net/xfrm.h>
275#include <net/ip.h>
276#include <net/netdma.h>
277#include <net/sock.h>
278
279#include <asm/uaccess.h>
280#include <asm/ioctls.h>
281
282int sysctl_tcp_fin_timeout __read_mostly = TCP_FIN_TIMEOUT;
283
284struct percpu_counter tcp_orphan_count;
285EXPORT_SYMBOL_GPL(tcp_orphan_count);
286
287int sysctl_tcp_wmem[3] __read_mostly;
288int sysctl_tcp_rmem[3] __read_mostly;
289
290EXPORT_SYMBOL(sysctl_tcp_rmem);
291EXPORT_SYMBOL(sysctl_tcp_wmem);
292
293atomic_long_t tcp_memory_allocated; /* Current allocated memory. */
294EXPORT_SYMBOL(tcp_memory_allocated);
295
296/*
297 * Current number of TCP sockets.
298 */
299struct percpu_counter tcp_sockets_allocated;
300EXPORT_SYMBOL(tcp_sockets_allocated);
301
302/*
303 * TCP splice context
304 */
305struct tcp_splice_state {
306 struct pipe_inode_info *pipe;
307 size_t len;
308 unsigned int flags;
309};
310
311/*
312 * Pressure flag: try to collapse.
313 * Technical note: it is used by multiple contexts non atomically.
314 * All the __sk_mem_schedule() is of this nature: accounting
315 * is strict, actions are advisory and have some latency.
316 */
317int tcp_memory_pressure __read_mostly;
318EXPORT_SYMBOL(tcp_memory_pressure);
319
320void tcp_enter_memory_pressure(struct sock *sk)
321{
322 if (!tcp_memory_pressure) {
323 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMEMORYPRESSURES);
324 tcp_memory_pressure = 1;
325 }
326}
327EXPORT_SYMBOL(tcp_enter_memory_pressure);
328
329/* Convert seconds to retransmits based on initial and max timeout */
330static u8 secs_to_retrans(int seconds, int timeout, int rto_max)
331{
332 u8 res = 0;
333
334 if (seconds > 0) {
335 int period = timeout;
336
337 res = 1;
338 while (seconds > period && res < 255) {
339 res++;
340 timeout <<= 1;
341 if (timeout > rto_max)
342 timeout = rto_max;
343 period += timeout;
344 }
345 }
346 return res;
347}
348
349/* Convert retransmits to seconds based on initial and max timeout */
350static int retrans_to_secs(u8 retrans, int timeout, int rto_max)
351{
352 int period = 0;
353
354 if (retrans > 0) {
355 period = timeout;
356 while (--retrans) {
357 timeout <<= 1;
358 if (timeout > rto_max)
359 timeout = rto_max;
360 period += timeout;
361 }
362 }
363 return period;
364}
365
366/* Address-family independent initialization for a tcp_sock.
367 *
368 * NOTE: A lot of things set to zero explicitly by call to
369 * sk_alloc() so need not be done here.
370 */
371void tcp_init_sock(struct sock *sk)
372{
373 struct inet_connection_sock *icsk = inet_csk(sk);
374 struct tcp_sock *tp = tcp_sk(sk);
375
376 skb_queue_head_init(&tp->out_of_order_queue);
377 tcp_init_xmit_timers(sk);
378 tcp_prequeue_init(tp);
379
380 icsk->icsk_rto = TCP_TIMEOUT_INIT;
381 tp->mdev = TCP_TIMEOUT_INIT;
382
383 /* So many TCP implementations out there (incorrectly) count the
384 * initial SYN frame in their delayed-ACK and congestion control
385 * algorithms that we must have the following bandaid to talk
386 * efficiently to them. -DaveM
387 */
388 tp->snd_cwnd = TCP_INIT_CWND;
389
390 /* See draft-stevens-tcpca-spec-01 for discussion of the
391 * initialization of these values.
392 */
393 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
394 tp->snd_cwnd_clamp = ~0;
395 tp->mss_cache = TCP_MSS_DEFAULT;
396
397 tp->reordering = sysctl_tcp_reordering;
398 tcp_enable_early_retrans(tp);
399 icsk->icsk_ca_ops = &tcp_init_congestion_ops;
400
401 sk->sk_state = TCP_CLOSE;
402
403 sk->sk_write_space = sk_stream_write_space;
404 sock_set_flag(sk, SOCK_USE_WRITE_QUEUE);
405
406 icsk->icsk_sync_mss = tcp_sync_mss;
407
408 /* TCP Cookie Transactions */
409 if (sysctl_tcp_cookie_size > 0) {
410 /* Default, cookies without s_data_payload. */
411 tp->cookie_values =
412 kzalloc(sizeof(*tp->cookie_values),
413 sk->sk_allocation);
414 if (tp->cookie_values != NULL)
415 kref_init(&tp->cookie_values->kref);
416 }
417 /* Presumed zeroed, in order of appearance:
418 * cookie_in_always, cookie_out_never,
419 * s_data_constant, s_data_in, s_data_out
420 */
421 sk->sk_sndbuf = sysctl_tcp_wmem[1];
422 sk->sk_rcvbuf = sysctl_tcp_rmem[1];
423
424 local_bh_disable();
425 sock_update_memcg(sk);
426 sk_sockets_allocated_inc(sk);
427 local_bh_enable();
428}
429EXPORT_SYMBOL(tcp_init_sock);
430
431/*
432 * Wait for a TCP event.
433 *
434 * Note that we don't need to lock the socket, as the upper poll layers
435 * take care of normal races (between the test and the event) and we don't
436 * go look at any of the socket buffers directly.
437 */
438unsigned int tcp_poll(struct file *file, struct socket *sock, poll_table *wait)
439{
440 unsigned int mask;
441 struct sock *sk = sock->sk;
442 const struct tcp_sock *tp = tcp_sk(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? */
488 if ((1 << sk->sk_state) & ~(TCPF_SYN_SENT | TCPF_SYN_RECV)) {
489 int target = sock_rcvlowat(sk, 0, INT_MAX);
490
491 if (tp->urg_seq == tp->copied_seq &&
492 !sock_flag(sk, SOCK_URGINLINE) &&
493 tp->urg_data)
494 target++;
495
496 /* Potential race condition. If read of tp below will
497 * escape above sk->sk_state, we can be illegally awaken
498 * in SYN_* states. */
499 if (tp->rcv_nxt - tp->copied_seq >= target)
500 mask |= POLLIN | POLLRDNORM;
501
502 if (!(sk->sk_shutdown & SEND_SHUTDOWN)) {
503 if (sk_stream_wspace(sk) >= sk_stream_min_wspace(sk)) {
504 mask |= POLLOUT | POLLWRNORM;
505 } else { /* send SIGIO later */
506 set_bit(SOCK_ASYNC_NOSPACE,
507 &sk->sk_socket->flags);
508 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
509
510 /* Race breaker. If space is freed after
511 * wspace test but before the flags are set,
512 * IO signal will be lost.
513 */
514 if (sk_stream_wspace(sk) >= sk_stream_min_wspace(sk))
515 mask |= POLLOUT | POLLWRNORM;
516 }
517 } else
518 mask |= POLLOUT | POLLWRNORM;
519
520 if (tp->urg_data & TCP_URG_VALID)
521 mask |= POLLPRI;
522 }
523 /* This barrier is coupled with smp_wmb() in tcp_reset() */
524 smp_rmb();
525 if (sk->sk_err)
526 mask |= POLLERR;
527
528 return mask;
529}
530EXPORT_SYMBOL(tcp_poll);
531
532int tcp_ioctl(struct sock *sk, int cmd, unsigned long arg)
533{
534 struct tcp_sock *tp = tcp_sk(sk);
535 int answ;
536
537 switch (cmd) {
538 case SIOCINQ:
539 if (sk->sk_state == TCP_LISTEN)
540 return -EINVAL;
541
542 lock_sock(sk);
543 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
544 answ = 0;
545 else if (sock_flag(sk, SOCK_URGINLINE) ||
546 !tp->urg_data ||
547 before(tp->urg_seq, tp->copied_seq) ||
548 !before(tp->urg_seq, tp->rcv_nxt)) {
549 struct sk_buff *skb;
550
551 answ = tp->rcv_nxt - tp->copied_seq;
552
553 /* Subtract 1, if FIN is in queue. */
554 skb = skb_peek_tail(&sk->sk_receive_queue);
555 if (answ && skb)
556 answ -= tcp_hdr(skb)->fin;
557 } else
558 answ = tp->urg_seq - tp->copied_seq;
559 release_sock(sk);
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
624static inline void tcp_push(struct sock *sk, int flags, int mss_now,
625 int nonagle)
626{
627 if (tcp_send_head(sk)) {
628 struct tcp_sock *tp = tcp_sk(sk);
629
630 if (!(flags & MSG_MORE) || forced_push(tp))
631 tcp_mark_push(tp, tcp_write_queue_tail(sk));
632
633 tcp_mark_urg(tp, flags);
634 __tcp_push_pending_frames(sk, mss_now,
635 (flags & MSG_MORE) ? TCP_NAGLE_CORK : nonagle);
636 }
637}
638
639static int tcp_splice_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb,
640 unsigned int offset, size_t len)
641{
642 struct tcp_splice_state *tss = rd_desc->arg.data;
643 int ret;
644
645 ret = skb_splice_bits(skb, offset, tss->pipe, min(rd_desc->count, len),
646 tss->flags);
647 if (ret > 0)
648 rd_desc->count -= ret;
649 return ret;
650}
651
652static int __tcp_splice_read(struct sock *sk, struct tcp_splice_state *tss)
653{
654 /* Store TCP splice context information in read_descriptor_t. */
655 read_descriptor_t rd_desc = {
656 .arg.data = tss,
657 .count = tss->len,
658 };
659
660 return tcp_read_sock(sk, &rd_desc, tcp_splice_data_recv);
661}
662
663/**
664 * tcp_splice_read - splice data from TCP socket to a pipe
665 * @sock: socket to splice from
666 * @ppos: position (not valid)
667 * @pipe: pipe to splice to
668 * @len: number of bytes to splice
669 * @flags: splice modifier flags
670 *
671 * Description:
672 * Will read pages from given socket and fill them into a pipe.
673 *
674 **/
675ssize_t tcp_splice_read(struct socket *sock, loff_t *ppos,
676 struct pipe_inode_info *pipe, size_t len,
677 unsigned int flags)
678{
679 struct sock *sk = sock->sk;
680 struct tcp_splice_state tss = {
681 .pipe = pipe,
682 .len = len,
683 .flags = flags,
684 };
685 long timeo;
686 ssize_t spliced;
687 int ret;
688
689 sock_rps_record_flow(sk);
690 /*
691 * We can't seek on a socket input
692 */
693 if (unlikely(*ppos))
694 return -ESPIPE;
695
696 ret = spliced = 0;
697
698 lock_sock(sk);
699
700 timeo = sock_rcvtimeo(sk, sock->file->f_flags & O_NONBLOCK);
701 while (tss.len) {
702 ret = __tcp_splice_read(sk, &tss);
703 if (ret < 0)
704 break;
705 else if (!ret) {
706 if (spliced)
707 break;
708 if (sock_flag(sk, SOCK_DONE))
709 break;
710 if (sk->sk_err) {
711 ret = sock_error(sk);
712 break;
713 }
714 if (sk->sk_shutdown & RCV_SHUTDOWN)
715 break;
716 if (sk->sk_state == TCP_CLOSE) {
717 /*
718 * This occurs when user tries to read
719 * from never connected socket.
720 */
721 if (!sock_flag(sk, SOCK_DONE))
722 ret = -ENOTCONN;
723 break;
724 }
725 if (!timeo) {
726 ret = -EAGAIN;
727 break;
728 }
729 sk_wait_data(sk, &timeo);
730 if (signal_pending(current)) {
731 ret = sock_intr_errno(timeo);
732 break;
733 }
734 continue;
735 }
736 tss.len -= ret;
737 spliced += ret;
738
739 if (!timeo)
740 break;
741 release_sock(sk);
742 lock_sock(sk);
743
744 if (sk->sk_err || sk->sk_state == TCP_CLOSE ||
745 (sk->sk_shutdown & RCV_SHUTDOWN) ||
746 signal_pending(current))
747 break;
748 }
749
750 release_sock(sk);
751
752 if (spliced)
753 return spliced;
754
755 return ret;
756}
757EXPORT_SYMBOL(tcp_splice_read);
758
759struct sk_buff *sk_stream_alloc_skb(struct sock *sk, int size, gfp_t gfp)
760{
761 struct sk_buff *skb;
762
763 /* The TCP header must be at least 32-bit aligned. */
764 size = ALIGN(size, 4);
765
766 skb = alloc_skb_fclone(size + sk->sk_prot->max_header, gfp);
767 if (skb) {
768 if (sk_wmem_schedule(sk, skb->truesize)) {
769 skb_reserve(skb, sk->sk_prot->max_header);
770 /*
771 * Make sure that we have exactly size bytes
772 * available to the caller, no more, no less.
773 */
774 skb->avail_size = size;
775 return skb;
776 }
777 __kfree_skb(skb);
778 } else {
779 sk->sk_prot->enter_memory_pressure(sk);
780 sk_stream_moderate_sndbuf(sk);
781 }
782 return NULL;
783}
784
785static unsigned int tcp_xmit_size_goal(struct sock *sk, u32 mss_now,
786 int large_allowed)
787{
788 struct tcp_sock *tp = tcp_sk(sk);
789 u32 xmit_size_goal, old_size_goal;
790
791 xmit_size_goal = mss_now;
792
793 if (large_allowed && sk_can_gso(sk)) {
794 xmit_size_goal = ((sk->sk_gso_max_size - 1) -
795 inet_csk(sk)->icsk_af_ops->net_header_len -
796 inet_csk(sk)->icsk_ext_hdr_len -
797 tp->tcp_header_len);
798
799 xmit_size_goal = tcp_bound_to_half_wnd(tp, xmit_size_goal);
800
801 /* We try hard to avoid divides here */
802 old_size_goal = tp->xmit_size_goal_segs * mss_now;
803
804 if (likely(old_size_goal <= xmit_size_goal &&
805 old_size_goal + mss_now > xmit_size_goal)) {
806 xmit_size_goal = old_size_goal;
807 } else {
808 tp->xmit_size_goal_segs =
809 min_t(u16, xmit_size_goal / mss_now,
810 sk->sk_gso_max_segs);
811 xmit_size_goal = tp->xmit_size_goal_segs * mss_now;
812 }
813 }
814
815 return max(xmit_size_goal, mss_now);
816}
817
818static int tcp_send_mss(struct sock *sk, int *size_goal, int flags)
819{
820 int mss_now;
821
822 mss_now = tcp_current_mss(sk);
823 *size_goal = tcp_xmit_size_goal(sk, mss_now, !(flags & MSG_OOB));
824
825 return mss_now;
826}
827
828static ssize_t do_tcp_sendpages(struct sock *sk, struct page **pages, int poffset,
829 size_t psize, int flags)
830{
831 struct tcp_sock *tp = tcp_sk(sk);
832 int mss_now, size_goal;
833 int err;
834 ssize_t copied;
835 long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
836
837 /* Wait for a connection to finish. */
838 if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
839 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
840 goto out_err;
841
842 clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
843
844 mss_now = tcp_send_mss(sk, &size_goal, flags);
845 copied = 0;
846
847 err = -EPIPE;
848 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
849 goto out_err;
850
851 while (psize > 0) {
852 struct sk_buff *skb = tcp_write_queue_tail(sk);
853 struct page *page = pages[poffset / PAGE_SIZE];
854 int copy, i;
855 int offset = poffset % PAGE_SIZE;
856 int size = min_t(size_t, psize, PAGE_SIZE - offset);
857 bool can_coalesce;
858
859 if (!tcp_send_head(sk) || (copy = size_goal - skb->len) <= 0) {
860new_segment:
861 if (!sk_stream_memory_free(sk))
862 goto wait_for_sndbuf;
863
864 skb = sk_stream_alloc_skb(sk, 0, sk->sk_allocation);
865 if (!skb)
866 goto wait_for_memory;
867
868 skb_entail(sk, skb);
869 copy = size_goal;
870 }
871
872 if (copy > size)
873 copy = size;
874
875 i = skb_shinfo(skb)->nr_frags;
876 can_coalesce = skb_can_coalesce(skb, i, page, offset);
877 if (!can_coalesce && i >= MAX_SKB_FRAGS) {
878 tcp_mark_push(tp, skb);
879 goto new_segment;
880 }
881 if (!sk_wmem_schedule(sk, copy))
882 goto wait_for_memory;
883
884 if (can_coalesce) {
885 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
886 } else {
887 get_page(page);
888 skb_fill_page_desc(skb, i, page, offset, copy);
889 }
890
891 skb->len += copy;
892 skb->data_len += copy;
893 skb->truesize += copy;
894 sk->sk_wmem_queued += copy;
895 sk_mem_charge(sk, copy);
896 skb->ip_summed = CHECKSUM_PARTIAL;
897 tp->write_seq += copy;
898 TCP_SKB_CB(skb)->end_seq += copy;
899 skb_shinfo(skb)->gso_segs = 0;
900
901 if (!copied)
902 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
903
904 copied += copy;
905 poffset += copy;
906 if (!(psize -= copy))
907 goto out;
908
909 if (skb->len < size_goal || (flags & MSG_OOB))
910 continue;
911
912 if (forced_push(tp)) {
913 tcp_mark_push(tp, skb);
914 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
915 } else if (skb == tcp_send_head(sk))
916 tcp_push_one(sk, mss_now);
917 continue;
918
919wait_for_sndbuf:
920 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
921wait_for_memory:
922 tcp_push(sk, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH);
923
924 if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
925 goto do_error;
926
927 mss_now = tcp_send_mss(sk, &size_goal, flags);
928 }
929
930out:
931 if (copied && !(flags & MSG_SENDPAGE_NOTLAST))
932 tcp_push(sk, flags, mss_now, tp->nonagle);
933 return copied;
934
935do_error:
936 if (copied)
937 goto out;
938out_err:
939 return sk_stream_error(sk, flags, err);
940}
941
942int tcp_sendpage(struct sock *sk, struct page *page, int offset,
943 size_t size, int flags)
944{
945 ssize_t res;
946
947 if (!(sk->sk_route_caps & NETIF_F_SG) ||
948 !(sk->sk_route_caps & NETIF_F_ALL_CSUM))
949 return sock_no_sendpage(sk->sk_socket, page, offset, size,
950 flags);
951
952 lock_sock(sk);
953 res = do_tcp_sendpages(sk, &page, offset, size, flags);
954 release_sock(sk);
955 return res;
956}
957EXPORT_SYMBOL(tcp_sendpage);
958
959static inline int select_size(const struct sock *sk, bool sg)
960{
961 const struct tcp_sock *tp = tcp_sk(sk);
962 int tmp = tp->mss_cache;
963
964 if (sg) {
965 if (sk_can_gso(sk)) {
966 /* Small frames wont use a full page:
967 * Payload will immediately follow tcp header.
968 */
969 tmp = SKB_WITH_OVERHEAD(2048 - MAX_TCP_HEADER);
970 } else {
971 int pgbreak = SKB_MAX_HEAD(MAX_TCP_HEADER);
972
973 if (tmp >= pgbreak &&
974 tmp <= pgbreak + (MAX_SKB_FRAGS - 1) * PAGE_SIZE)
975 tmp = pgbreak;
976 }
977 }
978
979 return tmp;
980}
981
982int tcp_sendmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg,
983 size_t size)
984{
985 struct iovec *iov;
986 struct tcp_sock *tp = tcp_sk(sk);
987 struct sk_buff *skb;
988 int iovlen, flags, err, copied;
989 int mss_now = 0, size_goal;
990 bool sg;
991 long timeo;
992
993 lock_sock(sk);
994
995 flags = msg->msg_flags;
996 timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
997
998 /* Wait for a connection to finish. */
999 if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
1000 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
1001 goto out_err;
1002
1003 if (unlikely(tp->repair)) {
1004 if (tp->repair_queue == TCP_RECV_QUEUE) {
1005 copied = tcp_send_rcvq(sk, msg, size);
1006 goto out;
1007 }
1008
1009 err = -EINVAL;
1010 if (tp->repair_queue == TCP_NO_QUEUE)
1011 goto out_err;
1012
1013 /* 'common' sending to sendq */
1014 }
1015
1016 /* This should be in poll */
1017 clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
1018
1019 mss_now = tcp_send_mss(sk, &size_goal, flags);
1020
1021 /* Ok commence sending. */
1022 iovlen = msg->msg_iovlen;
1023 iov = msg->msg_iov;
1024 copied = 0;
1025
1026 err = -EPIPE;
1027 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
1028 goto out_err;
1029
1030 sg = !!(sk->sk_route_caps & NETIF_F_SG);
1031
1032 while (--iovlen >= 0) {
1033 size_t seglen = iov->iov_len;
1034 unsigned char __user *from = iov->iov_base;
1035
1036 iov++;
1037
1038 while (seglen > 0) {
1039 int copy = 0;
1040 int max = size_goal;
1041
1042 skb = tcp_write_queue_tail(sk);
1043 if (tcp_send_head(sk)) {
1044 if (skb->ip_summed == CHECKSUM_NONE)
1045 max = mss_now;
1046 copy = max - skb->len;
1047 }
1048
1049 if (copy <= 0) {
1050new_segment:
1051 /* Allocate new segment. If the interface is SG,
1052 * allocate skb fitting to single page.
1053 */
1054 if (!sk_stream_memory_free(sk))
1055 goto wait_for_sndbuf;
1056
1057 skb = sk_stream_alloc_skb(sk,
1058 select_size(sk, sg),
1059 sk->sk_allocation);
1060 if (!skb)
1061 goto wait_for_memory;
1062
1063 /*
1064 * Check whether we can use HW checksum.
1065 */
1066 if (sk->sk_route_caps & NETIF_F_ALL_CSUM)
1067 skb->ip_summed = CHECKSUM_PARTIAL;
1068
1069 skb_entail(sk, skb);
1070 copy = size_goal;
1071 max = size_goal;
1072 }
1073
1074 /* Try to append data to the end of skb. */
1075 if (copy > seglen)
1076 copy = seglen;
1077
1078 /* Where to copy to? */
1079 if (skb_availroom(skb) > 0) {
1080 /* We have some space in skb head. Superb! */
1081 copy = min_t(int, copy, skb_availroom(skb));
1082 err = skb_add_data_nocache(sk, skb, from, copy);
1083 if (err)
1084 goto do_fault;
1085 } else {
1086 bool merge = false;
1087 int i = skb_shinfo(skb)->nr_frags;
1088 struct page *page = sk->sk_sndmsg_page;
1089 int off;
1090
1091 if (page && page_count(page) == 1)
1092 sk->sk_sndmsg_off = 0;
1093
1094 off = sk->sk_sndmsg_off;
1095
1096 if (skb_can_coalesce(skb, i, page, off) &&
1097 off != PAGE_SIZE) {
1098 /* We can extend the last page
1099 * fragment. */
1100 merge = true;
1101 } else if (i == MAX_SKB_FRAGS || !sg) {
1102 /* Need to add new fragment and cannot
1103 * do this because interface is non-SG,
1104 * or because all the page slots are
1105 * busy. */
1106 tcp_mark_push(tp, skb);
1107 goto new_segment;
1108 } else if (page) {
1109 if (off == PAGE_SIZE) {
1110 put_page(page);
1111 sk->sk_sndmsg_page = page = NULL;
1112 off = 0;
1113 }
1114 } else
1115 off = 0;
1116
1117 if (copy > PAGE_SIZE - off)
1118 copy = PAGE_SIZE - off;
1119
1120 if (!sk_wmem_schedule(sk, copy))
1121 goto wait_for_memory;
1122
1123 if (!page) {
1124 /* Allocate new cache page. */
1125 if (!(page = sk_stream_alloc_page(sk)))
1126 goto wait_for_memory;
1127 }
1128
1129 /* Time to copy data. We are close to
1130 * the end! */
1131 err = skb_copy_to_page_nocache(sk, from, skb,
1132 page, off, copy);
1133 if (err) {
1134 /* If this page was new, give it to the
1135 * socket so it does not get leaked.
1136 */
1137 if (!sk->sk_sndmsg_page) {
1138 sk->sk_sndmsg_page = page;
1139 sk->sk_sndmsg_off = 0;
1140 }
1141 goto do_error;
1142 }
1143
1144 /* Update the skb. */
1145 if (merge) {
1146 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1147 } else {
1148 skb_fill_page_desc(skb, i, page, off, copy);
1149 if (sk->sk_sndmsg_page) {
1150 get_page(page);
1151 } else if (off + copy < PAGE_SIZE) {
1152 get_page(page);
1153 sk->sk_sndmsg_page = page;
1154 }
1155 }
1156
1157 sk->sk_sndmsg_off = off + copy;
1158 }
1159
1160 if (!copied)
1161 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
1162
1163 tp->write_seq += copy;
1164 TCP_SKB_CB(skb)->end_seq += copy;
1165 skb_shinfo(skb)->gso_segs = 0;
1166
1167 from += copy;
1168 copied += copy;
1169 if ((seglen -= copy) == 0 && iovlen == 0)
1170 goto out;
1171
1172 if (skb->len < max || (flags & MSG_OOB) || unlikely(tp->repair))
1173 continue;
1174
1175 if (forced_push(tp)) {
1176 tcp_mark_push(tp, skb);
1177 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
1178 } else if (skb == tcp_send_head(sk))
1179 tcp_push_one(sk, mss_now);
1180 continue;
1181
1182wait_for_sndbuf:
1183 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1184wait_for_memory:
1185 if (copied && likely(!tp->repair))
1186 tcp_push(sk, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH);
1187
1188 if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
1189 goto do_error;
1190
1191 mss_now = tcp_send_mss(sk, &size_goal, flags);
1192 }
1193 }
1194
1195out:
1196 if (copied && likely(!tp->repair))
1197 tcp_push(sk, flags, mss_now, tp->nonagle);
1198 release_sock(sk);
1199 return copied;
1200
1201do_fault:
1202 if (!skb->len) {
1203 tcp_unlink_write_queue(skb, sk);
1204 /* It is the one place in all of TCP, except connection
1205 * reset, where we can be unlinking the send_head.
1206 */
1207 tcp_check_send_head(sk, skb);
1208 sk_wmem_free_skb(sk, skb);
1209 }
1210
1211do_error:
1212 if (copied)
1213 goto out;
1214out_err:
1215 err = sk_stream_error(sk, flags, err);
1216 release_sock(sk);
1217 return err;
1218}
1219EXPORT_SYMBOL(tcp_sendmsg);
1220
1221/*
1222 * Handle reading urgent data. BSD has very simple semantics for
1223 * this, no blocking and very strange errors 8)
1224 */
1225
1226static int tcp_recv_urg(struct sock *sk, struct msghdr *msg, int len, int flags)
1227{
1228 struct tcp_sock *tp = tcp_sk(sk);
1229
1230 /* No URG data to read. */
1231 if (sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data ||
1232 tp->urg_data == TCP_URG_READ)
1233 return -EINVAL; /* Yes this is right ! */
1234
1235 if (sk->sk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DONE))
1236 return -ENOTCONN;
1237
1238 if (tp->urg_data & TCP_URG_VALID) {
1239 int err = 0;
1240 char c = tp->urg_data;
1241
1242 if (!(flags & MSG_PEEK))
1243 tp->urg_data = TCP_URG_READ;
1244
1245 /* Read urgent data. */
1246 msg->msg_flags |= MSG_OOB;
1247
1248 if (len > 0) {
1249 if (!(flags & MSG_TRUNC))
1250 err = memcpy_toiovec(msg->msg_iov, &c, 1);
1251 len = 1;
1252 } else
1253 msg->msg_flags |= MSG_TRUNC;
1254
1255 return err ? -EFAULT : len;
1256 }
1257
1258 if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
1259 return 0;
1260
1261 /* Fixed the recv(..., MSG_OOB) behaviour. BSD docs and
1262 * the available implementations agree in this case:
1263 * this call should never block, independent of the
1264 * blocking state of the socket.
1265 * Mike <pall@rz.uni-karlsruhe.de>
1266 */
1267 return -EAGAIN;
1268}
1269
1270static int tcp_peek_sndq(struct sock *sk, struct msghdr *msg, int len)
1271{
1272 struct sk_buff *skb;
1273 int copied = 0, err = 0;
1274
1275 /* XXX -- need to support SO_PEEK_OFF */
1276
1277 skb_queue_walk(&sk->sk_write_queue, skb) {
1278 err = skb_copy_datagram_iovec(skb, 0, msg->msg_iov, skb->len);
1279 if (err)
1280 break;
1281
1282 copied += skb->len;
1283 }
1284
1285 return err ?: copied;
1286}
1287
1288/* Clean up the receive buffer for full frames taken by the user,
1289 * then send an ACK if necessary. COPIED is the number of bytes
1290 * tcp_recvmsg has given to the user so far, it speeds up the
1291 * calculation of whether or not we must ACK for the sake of
1292 * a window update.
1293 */
1294void tcp_cleanup_rbuf(struct sock *sk, int copied)
1295{
1296 struct tcp_sock *tp = tcp_sk(sk);
1297 bool time_to_ack = false;
1298
1299 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
1300
1301 WARN(skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq),
1302 "cleanup rbuf bug: copied %X seq %X rcvnxt %X\n",
1303 tp->copied_seq, TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt);
1304
1305 if (inet_csk_ack_scheduled(sk)) {
1306 const struct inet_connection_sock *icsk = inet_csk(sk);
1307 /* Delayed ACKs frequently hit locked sockets during bulk
1308 * receive. */
1309 if (icsk->icsk_ack.blocked ||
1310 /* Once-per-two-segments ACK was not sent by tcp_input.c */
1311 tp->rcv_nxt - tp->rcv_wup > icsk->icsk_ack.rcv_mss ||
1312 /*
1313 * If this read emptied read buffer, we send ACK, if
1314 * connection is not bidirectional, user drained
1315 * receive buffer and there was a small segment
1316 * in queue.
1317 */
1318 (copied > 0 &&
1319 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED2) ||
1320 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED) &&
1321 !icsk->icsk_ack.pingpong)) &&
1322 !atomic_read(&sk->sk_rmem_alloc)))
1323 time_to_ack = true;
1324 }
1325
1326 /* We send an ACK if we can now advertise a non-zero window
1327 * which has been raised "significantly".
1328 *
1329 * Even if window raised up to infinity, do not send window open ACK
1330 * in states, where we will not receive more. It is useless.
1331 */
1332 if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) {
1333 __u32 rcv_window_now = tcp_receive_window(tp);
1334
1335 /* Optimize, __tcp_select_window() is not cheap. */
1336 if (2*rcv_window_now <= tp->window_clamp) {
1337 __u32 new_window = __tcp_select_window(sk);
1338
1339 /* Send ACK now, if this read freed lots of space
1340 * in our buffer. Certainly, new_window is new window.
1341 * We can advertise it now, if it is not less than current one.
1342 * "Lots" means "at least twice" here.
1343 */
1344 if (new_window && new_window >= 2 * rcv_window_now)
1345 time_to_ack = true;
1346 }
1347 }
1348 if (time_to_ack)
1349 tcp_send_ack(sk);
1350}
1351
1352static void tcp_prequeue_process(struct sock *sk)
1353{
1354 struct sk_buff *skb;
1355 struct tcp_sock *tp = tcp_sk(sk);
1356
1357 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPPREQUEUED);
1358
1359 /* RX process wants to run with disabled BHs, though it is not
1360 * necessary */
1361 local_bh_disable();
1362 while ((skb = __skb_dequeue(&tp->ucopy.prequeue)) != NULL)
1363 sk_backlog_rcv(sk, skb);
1364 local_bh_enable();
1365
1366 /* Clear memory counter. */
1367 tp->ucopy.memory = 0;
1368}
1369
1370#ifdef CONFIG_NET_DMA
1371static void tcp_service_net_dma(struct sock *sk, bool wait)
1372{
1373 dma_cookie_t done, used;
1374 dma_cookie_t last_issued;
1375 struct tcp_sock *tp = tcp_sk(sk);
1376
1377 if (!tp->ucopy.dma_chan)
1378 return;
1379
1380 last_issued = tp->ucopy.dma_cookie;
1381 dma_async_memcpy_issue_pending(tp->ucopy.dma_chan);
1382
1383 do {
1384 if (dma_async_memcpy_complete(tp->ucopy.dma_chan,
1385 last_issued, &done,
1386 &used) == DMA_SUCCESS) {
1387 /* Safe to free early-copied skbs now */
1388 __skb_queue_purge(&sk->sk_async_wait_queue);
1389 break;
1390 } else {
1391 struct sk_buff *skb;
1392 while ((skb = skb_peek(&sk->sk_async_wait_queue)) &&
1393 (dma_async_is_complete(skb->dma_cookie, done,
1394 used) == DMA_SUCCESS)) {
1395 __skb_dequeue(&sk->sk_async_wait_queue);
1396 kfree_skb(skb);
1397 }
1398 }
1399 } while (wait);
1400}
1401#endif
1402
1403static inline struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off)
1404{
1405 struct sk_buff *skb;
1406 u32 offset;
1407
1408 skb_queue_walk(&sk->sk_receive_queue, skb) {
1409 offset = seq - TCP_SKB_CB(skb)->seq;
1410 if (tcp_hdr(skb)->syn)
1411 offset--;
1412 if (offset < skb->len || tcp_hdr(skb)->fin) {
1413 *off = offset;
1414 return skb;
1415 }
1416 }
1417 return NULL;
1418}
1419
1420/*
1421 * This routine provides an alternative to tcp_recvmsg() for routines
1422 * that would like to handle copying from skbuffs directly in 'sendfile'
1423 * fashion.
1424 * Note:
1425 * - It is assumed that the socket was locked by the caller.
1426 * - The routine does not block.
1427 * - At present, there is no support for reading OOB data
1428 * or for 'peeking' the socket using this routine
1429 * (although both would be easy to implement).
1430 */
1431int tcp_read_sock(struct sock *sk, read_descriptor_t *desc,
1432 sk_read_actor_t recv_actor)
1433{
1434 struct sk_buff *skb;
1435 struct tcp_sock *tp = tcp_sk(sk);
1436 u32 seq = tp->copied_seq;
1437 u32 offset;
1438 int copied = 0;
1439
1440 if (sk->sk_state == TCP_LISTEN)
1441 return -ENOTCONN;
1442 while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1443 if (offset < skb->len) {
1444 int used;
1445 size_t len;
1446
1447 len = skb->len - offset;
1448 /* Stop reading if we hit a patch of urgent data */
1449 if (tp->urg_data) {
1450 u32 urg_offset = tp->urg_seq - seq;
1451 if (urg_offset < len)
1452 len = urg_offset;
1453 if (!len)
1454 break;
1455 }
1456 used = recv_actor(desc, skb, offset, len);
1457 if (used < 0) {
1458 if (!copied)
1459 copied = used;
1460 break;
1461 } else if (used <= len) {
1462 seq += used;
1463 copied += used;
1464 offset += used;
1465 }
1466 /*
1467 * If recv_actor drops the lock (e.g. TCP splice
1468 * receive) the skb pointer might be invalid when
1469 * getting here: tcp_collapse might have deleted it
1470 * while aggregating skbs from the socket queue.
1471 */
1472 skb = tcp_recv_skb(sk, seq-1, &offset);
1473 if (!skb || (offset+1 != skb->len))
1474 break;
1475 }
1476 if (tcp_hdr(skb)->fin) {
1477 sk_eat_skb(sk, skb, false);
1478 ++seq;
1479 break;
1480 }
1481 sk_eat_skb(sk, skb, false);
1482 if (!desc->count)
1483 break;
1484 tp->copied_seq = seq;
1485 }
1486 tp->copied_seq = seq;
1487
1488 tcp_rcv_space_adjust(sk);
1489
1490 /* Clean up data we have read: This will do ACK frames. */
1491 if (copied > 0)
1492 tcp_cleanup_rbuf(sk, copied);
1493 return copied;
1494}
1495EXPORT_SYMBOL(tcp_read_sock);
1496
1497/*
1498 * This routine copies from a sock struct into the user buffer.
1499 *
1500 * Technical note: in 2.3 we work on _locked_ socket, so that
1501 * tricks with *seq access order and skb->users are not required.
1502 * Probably, code can be easily improved even more.
1503 */
1504
1505int tcp_recvmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg,
1506 size_t len, int nonblock, int flags, int *addr_len)
1507{
1508 struct tcp_sock *tp = tcp_sk(sk);
1509 int copied = 0;
1510 u32 peek_seq;
1511 u32 *seq;
1512 unsigned long used;
1513 int err;
1514 int target; /* Read at least this many bytes */
1515 long timeo;
1516 struct task_struct *user_recv = NULL;
1517 bool copied_early = false;
1518 struct sk_buff *skb;
1519 u32 urg_hole = 0;
1520
1521 lock_sock(sk);
1522
1523 err = -ENOTCONN;
1524 if (sk->sk_state == TCP_LISTEN)
1525 goto out;
1526
1527 timeo = sock_rcvtimeo(sk, nonblock);
1528
1529 /* Urgent data needs to be handled specially. */
1530 if (flags & MSG_OOB)
1531 goto recv_urg;
1532
1533 if (unlikely(tp->repair)) {
1534 err = -EPERM;
1535 if (!(flags & MSG_PEEK))
1536 goto out;
1537
1538 if (tp->repair_queue == TCP_SEND_QUEUE)
1539 goto recv_sndq;
1540
1541 err = -EINVAL;
1542 if (tp->repair_queue == TCP_NO_QUEUE)
1543 goto out;
1544
1545 /* 'common' recv queue MSG_PEEK-ing */
1546 }
1547
1548 seq = &tp->copied_seq;
1549 if (flags & MSG_PEEK) {
1550 peek_seq = tp->copied_seq;
1551 seq = &peek_seq;
1552 }
1553
1554 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
1555
1556#ifdef CONFIG_NET_DMA
1557 tp->ucopy.dma_chan = NULL;
1558 preempt_disable();
1559 skb = skb_peek_tail(&sk->sk_receive_queue);
1560 {
1561 int available = 0;
1562
1563 if (skb)
1564 available = TCP_SKB_CB(skb)->seq + skb->len - (*seq);
1565 if ((available < target) &&
1566 (len > sysctl_tcp_dma_copybreak) && !(flags & MSG_PEEK) &&
1567 !sysctl_tcp_low_latency &&
1568 net_dma_find_channel()) {
1569 preempt_enable_no_resched();
1570 tp->ucopy.pinned_list =
1571 dma_pin_iovec_pages(msg->msg_iov, len);
1572 } else {
1573 preempt_enable_no_resched();
1574 }
1575 }
1576#endif
1577
1578 do {
1579 u32 offset;
1580
1581 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
1582 if (tp->urg_data && tp->urg_seq == *seq) {
1583 if (copied)
1584 break;
1585 if (signal_pending(current)) {
1586 copied = timeo ? sock_intr_errno(timeo) : -EAGAIN;
1587 break;
1588 }
1589 }
1590
1591 /* Next get a buffer. */
1592
1593 skb_queue_walk(&sk->sk_receive_queue, skb) {
1594 /* Now that we have two receive queues this
1595 * shouldn't happen.
1596 */
1597 if (WARN(before(*seq, TCP_SKB_CB(skb)->seq),
1598 "recvmsg bug: copied %X seq %X rcvnxt %X fl %X\n",
1599 *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt,
1600 flags))
1601 break;
1602
1603 offset = *seq - TCP_SKB_CB(skb)->seq;
1604 if (tcp_hdr(skb)->syn)
1605 offset--;
1606 if (offset < skb->len)
1607 goto found_ok_skb;
1608 if (tcp_hdr(skb)->fin)
1609 goto found_fin_ok;
1610 WARN(!(flags & MSG_PEEK),
1611 "recvmsg bug 2: copied %X seq %X rcvnxt %X fl %X\n",
1612 *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt, flags);
1613 }
1614
1615 /* Well, if we have backlog, try to process it now yet. */
1616
1617 if (copied >= target && !sk->sk_backlog.tail)
1618 break;
1619
1620 if (copied) {
1621 if (sk->sk_err ||
1622 sk->sk_state == TCP_CLOSE ||
1623 (sk->sk_shutdown & RCV_SHUTDOWN) ||
1624 !timeo ||
1625 signal_pending(current))
1626 break;
1627 } else {
1628 if (sock_flag(sk, SOCK_DONE))
1629 break;
1630
1631 if (sk->sk_err) {
1632 copied = sock_error(sk);
1633 break;
1634 }
1635
1636 if (sk->sk_shutdown & RCV_SHUTDOWN)
1637 break;
1638
1639 if (sk->sk_state == TCP_CLOSE) {
1640 if (!sock_flag(sk, SOCK_DONE)) {
1641 /* This occurs when user tries to read
1642 * from never connected socket.
1643 */
1644 copied = -ENOTCONN;
1645 break;
1646 }
1647 break;
1648 }
1649
1650 if (!timeo) {
1651 copied = -EAGAIN;
1652 break;
1653 }
1654
1655 if (signal_pending(current)) {
1656 copied = sock_intr_errno(timeo);
1657 break;
1658 }
1659 }
1660
1661 tcp_cleanup_rbuf(sk, copied);
1662
1663 if (!sysctl_tcp_low_latency && tp->ucopy.task == user_recv) {
1664 /* Install new reader */
1665 if (!user_recv && !(flags & (MSG_TRUNC | MSG_PEEK))) {
1666 user_recv = current;
1667 tp->ucopy.task = user_recv;
1668 tp->ucopy.iov = msg->msg_iov;
1669 }
1670
1671 tp->ucopy.len = len;
1672
1673 WARN_ON(tp->copied_seq != tp->rcv_nxt &&
1674 !(flags & (MSG_PEEK | MSG_TRUNC)));
1675
1676 /* Ugly... If prequeue is not empty, we have to
1677 * process it before releasing socket, otherwise
1678 * order will be broken at second iteration.
1679 * More elegant solution is required!!!
1680 *
1681 * Look: we have the following (pseudo)queues:
1682 *
1683 * 1. packets in flight
1684 * 2. backlog
1685 * 3. prequeue
1686 * 4. receive_queue
1687 *
1688 * Each queue can be processed only if the next ones
1689 * are empty. At this point we have empty receive_queue.
1690 * But prequeue _can_ be not empty after 2nd iteration,
1691 * when we jumped to start of loop because backlog
1692 * processing added something to receive_queue.
1693 * We cannot release_sock(), because backlog contains
1694 * packets arrived _after_ prequeued ones.
1695 *
1696 * Shortly, algorithm is clear --- to process all
1697 * the queues in order. We could make it more directly,
1698 * requeueing packets from backlog to prequeue, if
1699 * is not empty. It is more elegant, but eats cycles,
1700 * unfortunately.
1701 */
1702 if (!skb_queue_empty(&tp->ucopy.prequeue))
1703 goto do_prequeue;
1704
1705 /* __ Set realtime policy in scheduler __ */
1706 }
1707
1708#ifdef CONFIG_NET_DMA
1709 if (tp->ucopy.dma_chan)
1710 dma_async_memcpy_issue_pending(tp->ucopy.dma_chan);
1711#endif
1712 if (copied >= target) {
1713 /* Do not sleep, just process backlog. */
1714 release_sock(sk);
1715 lock_sock(sk);
1716 } else
1717 sk_wait_data(sk, &timeo);
1718
1719#ifdef CONFIG_NET_DMA
1720 tcp_service_net_dma(sk, false); /* Don't block */
1721 tp->ucopy.wakeup = 0;
1722#endif
1723
1724 if (user_recv) {
1725 int chunk;
1726
1727 /* __ Restore normal policy in scheduler __ */
1728
1729 if ((chunk = len - tp->ucopy.len) != 0) {
1730 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMBACKLOG, chunk);
1731 len -= chunk;
1732 copied += chunk;
1733 }
1734
1735 if (tp->rcv_nxt == tp->copied_seq &&
1736 !skb_queue_empty(&tp->ucopy.prequeue)) {
1737do_prequeue:
1738 tcp_prequeue_process(sk);
1739
1740 if ((chunk = len - tp->ucopy.len) != 0) {
1741 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1742 len -= chunk;
1743 copied += chunk;
1744 }
1745 }
1746 }
1747 if ((flags & MSG_PEEK) &&
1748 (peek_seq - copied - urg_hole != tp->copied_seq)) {
1749 net_dbg_ratelimited("TCP(%s:%d): Application bug, race in MSG_PEEK\n",
1750 current->comm,
1751 task_pid_nr(current));
1752 peek_seq = tp->copied_seq;
1753 }
1754 continue;
1755
1756 found_ok_skb:
1757 /* Ok so how much can we use? */
1758 used = skb->len - offset;
1759 if (len < used)
1760 used = len;
1761
1762 /* Do we have urgent data here? */
1763 if (tp->urg_data) {
1764 u32 urg_offset = tp->urg_seq - *seq;
1765 if (urg_offset < used) {
1766 if (!urg_offset) {
1767 if (!sock_flag(sk, SOCK_URGINLINE)) {
1768 ++*seq;
1769 urg_hole++;
1770 offset++;
1771 used--;
1772 if (!used)
1773 goto skip_copy;
1774 }
1775 } else
1776 used = urg_offset;
1777 }
1778 }
1779
1780 if (!(flags & MSG_TRUNC)) {
1781#ifdef CONFIG_NET_DMA
1782 if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list)
1783 tp->ucopy.dma_chan = net_dma_find_channel();
1784
1785 if (tp->ucopy.dma_chan) {
1786 tp->ucopy.dma_cookie = dma_skb_copy_datagram_iovec(
1787 tp->ucopy.dma_chan, skb, offset,
1788 msg->msg_iov, used,
1789 tp->ucopy.pinned_list);
1790
1791 if (tp->ucopy.dma_cookie < 0) {
1792
1793 pr_alert("%s: dma_cookie < 0\n",
1794 __func__);
1795
1796 /* Exception. Bailout! */
1797 if (!copied)
1798 copied = -EFAULT;
1799 break;
1800 }
1801
1802 dma_async_memcpy_issue_pending(tp->ucopy.dma_chan);
1803
1804 if ((offset + used) == skb->len)
1805 copied_early = true;
1806
1807 } else
1808#endif
1809 {
1810 err = skb_copy_datagram_iovec(skb, offset,
1811 msg->msg_iov, used);
1812 if (err) {
1813 /* Exception. Bailout! */
1814 if (!copied)
1815 copied = -EFAULT;
1816 break;
1817 }
1818 }
1819 }
1820
1821 *seq += used;
1822 copied += used;
1823 len -= used;
1824
1825 tcp_rcv_space_adjust(sk);
1826
1827skip_copy:
1828 if (tp->urg_data && after(tp->copied_seq, tp->urg_seq)) {
1829 tp->urg_data = 0;
1830 tcp_fast_path_check(sk);
1831 }
1832 if (used + offset < skb->len)
1833 continue;
1834
1835 if (tcp_hdr(skb)->fin)
1836 goto found_fin_ok;
1837 if (!(flags & MSG_PEEK)) {
1838 sk_eat_skb(sk, skb, copied_early);
1839 copied_early = false;
1840 }
1841 continue;
1842
1843 found_fin_ok:
1844 /* Process the FIN. */
1845 ++*seq;
1846 if (!(flags & MSG_PEEK)) {
1847 sk_eat_skb(sk, skb, copied_early);
1848 copied_early = false;
1849 }
1850 break;
1851 } while (len > 0);
1852
1853 if (user_recv) {
1854 if (!skb_queue_empty(&tp->ucopy.prequeue)) {
1855 int chunk;
1856
1857 tp->ucopy.len = copied > 0 ? len : 0;
1858
1859 tcp_prequeue_process(sk);
1860
1861 if (copied > 0 && (chunk = len - tp->ucopy.len) != 0) {
1862 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1863 len -= chunk;
1864 copied += chunk;
1865 }
1866 }
1867
1868 tp->ucopy.task = NULL;
1869 tp->ucopy.len = 0;
1870 }
1871
1872#ifdef CONFIG_NET_DMA
1873 tcp_service_net_dma(sk, true); /* Wait for queue to drain */
1874 tp->ucopy.dma_chan = NULL;
1875
1876 if (tp->ucopy.pinned_list) {
1877 dma_unpin_iovec_pages(tp->ucopy.pinned_list);
1878 tp->ucopy.pinned_list = NULL;
1879 }
1880#endif
1881
1882 /* According to UNIX98, msg_name/msg_namelen are ignored
1883 * on connected socket. I was just happy when found this 8) --ANK
1884 */
1885
1886 /* Clean up data we have read: This will do ACK frames. */
1887 tcp_cleanup_rbuf(sk, copied);
1888
1889 release_sock(sk);
1890 return copied;
1891
1892out:
1893 release_sock(sk);
1894 return err;
1895
1896recv_urg:
1897 err = tcp_recv_urg(sk, msg, len, flags);
1898 goto out;
1899
1900recv_sndq:
1901 err = tcp_peek_sndq(sk, msg, len);
1902 goto out;
1903}
1904EXPORT_SYMBOL(tcp_recvmsg);
1905
1906void tcp_set_state(struct sock *sk, int state)
1907{
1908 int oldstate = sk->sk_state;
1909
1910 switch (state) {
1911 case TCP_ESTABLISHED:
1912 if (oldstate != TCP_ESTABLISHED)
1913 TCP_INC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
1914 break;
1915
1916 case TCP_CLOSE:
1917 if (oldstate == TCP_CLOSE_WAIT || oldstate == TCP_ESTABLISHED)
1918 TCP_INC_STATS(sock_net(sk), TCP_MIB_ESTABRESETS);
1919
1920 sk->sk_prot->unhash(sk);
1921 if (inet_csk(sk)->icsk_bind_hash &&
1922 !(sk->sk_userlocks & SOCK_BINDPORT_LOCK))
1923 inet_put_port(sk);
1924 /* fall through */
1925 default:
1926 if (oldstate == TCP_ESTABLISHED)
1927 TCP_DEC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
1928 }
1929
1930 /* Change state AFTER socket is unhashed to avoid closed
1931 * socket sitting in hash tables.
1932 */
1933 sk->sk_state = state;
1934
1935#ifdef STATE_TRACE
1936 SOCK_DEBUG(sk, "TCP sk=%p, State %s -> %s\n", sk, statename[oldstate], statename[state]);
1937#endif
1938}
1939EXPORT_SYMBOL_GPL(tcp_set_state);
1940
1941/*
1942 * State processing on a close. This implements the state shift for
1943 * sending our FIN frame. Note that we only send a FIN for some
1944 * states. A shutdown() may have already sent the FIN, or we may be
1945 * closed.
1946 */
1947
1948static const unsigned char new_state[16] = {
1949 /* current state: new state: action: */
1950 /* (Invalid) */ TCP_CLOSE,
1951 /* TCP_ESTABLISHED */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1952 /* TCP_SYN_SENT */ TCP_CLOSE,
1953 /* TCP_SYN_RECV */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1954 /* TCP_FIN_WAIT1 */ TCP_FIN_WAIT1,
1955 /* TCP_FIN_WAIT2 */ TCP_FIN_WAIT2,
1956 /* TCP_TIME_WAIT */ TCP_CLOSE,
1957 /* TCP_CLOSE */ TCP_CLOSE,
1958 /* TCP_CLOSE_WAIT */ TCP_LAST_ACK | TCP_ACTION_FIN,
1959 /* TCP_LAST_ACK */ TCP_LAST_ACK,
1960 /* TCP_LISTEN */ TCP_CLOSE,
1961 /* TCP_CLOSING */ TCP_CLOSING,
1962};
1963
1964static int tcp_close_state(struct sock *sk)
1965{
1966 int next = (int)new_state[sk->sk_state];
1967 int ns = next & TCP_STATE_MASK;
1968
1969 tcp_set_state(sk, ns);
1970
1971 return next & TCP_ACTION_FIN;
1972}
1973
1974/*
1975 * Shutdown the sending side of a connection. Much like close except
1976 * that we don't receive shut down or sock_set_flag(sk, SOCK_DEAD).
1977 */
1978
1979void tcp_shutdown(struct sock *sk, int how)
1980{
1981 /* We need to grab some memory, and put together a FIN,
1982 * and then put it into the queue to be sent.
1983 * Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
1984 */
1985 if (!(how & SEND_SHUTDOWN))
1986 return;
1987
1988 /* If we've already sent a FIN, or it's a closed state, skip this. */
1989 if ((1 << sk->sk_state) &
1990 (TCPF_ESTABLISHED | TCPF_SYN_SENT |
1991 TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) {
1992 /* Clear out any half completed packets. FIN if needed. */
1993 if (tcp_close_state(sk))
1994 tcp_send_fin(sk);
1995 }
1996}
1997EXPORT_SYMBOL(tcp_shutdown);
1998
1999bool tcp_check_oom(struct sock *sk, int shift)
2000{
2001 bool too_many_orphans, out_of_socket_memory;
2002
2003 too_many_orphans = tcp_too_many_orphans(sk, shift);
2004 out_of_socket_memory = tcp_out_of_memory(sk);
2005
2006 if (too_many_orphans)
2007 net_info_ratelimited("too many orphaned sockets\n");
2008 if (out_of_socket_memory)
2009 net_info_ratelimited("out of memory -- consider tuning tcp_mem\n");
2010 return too_many_orphans || out_of_socket_memory;
2011}
2012
2013void tcp_close(struct sock *sk, long timeout)
2014{
2015 struct sk_buff *skb;
2016 int data_was_unread = 0;
2017 int state;
2018
2019 lock_sock(sk);
2020 sk->sk_shutdown = SHUTDOWN_MASK;
2021
2022 if (sk->sk_state == TCP_LISTEN) {
2023 tcp_set_state(sk, TCP_CLOSE);
2024
2025 /* Special case. */
2026 inet_csk_listen_stop(sk);
2027
2028 goto adjudge_to_death;
2029 }
2030
2031 /* We need to flush the recv. buffs. We do this only on the
2032 * descriptor close, not protocol-sourced closes, because the
2033 * reader process may not have drained the data yet!
2034 */
2035 while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) {
2036 u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq -
2037 tcp_hdr(skb)->fin;
2038 data_was_unread += len;
2039 __kfree_skb(skb);
2040 }
2041
2042 sk_mem_reclaim(sk);
2043
2044 /* If socket has been already reset (e.g. in tcp_reset()) - kill it. */
2045 if (sk->sk_state == TCP_CLOSE)
2046 goto adjudge_to_death;
2047
2048 /* As outlined in RFC 2525, section 2.17, we send a RST here because
2049 * data was lost. To witness the awful effects of the old behavior of
2050 * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk
2051 * GET in an FTP client, suspend the process, wait for the client to
2052 * advertise a zero window, then kill -9 the FTP client, wheee...
2053 * Note: timeout is always zero in such a case.
2054 */
2055 if (unlikely(tcp_sk(sk)->repair)) {
2056 sk->sk_prot->disconnect(sk, 0);
2057 } else if (data_was_unread) {
2058 /* Unread data was tossed, zap the connection. */
2059 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONCLOSE);
2060 tcp_set_state(sk, TCP_CLOSE);
2061 tcp_send_active_reset(sk, sk->sk_allocation);
2062 } else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) {
2063 /* Check zero linger _after_ checking for unread data. */
2064 sk->sk_prot->disconnect(sk, 0);
2065 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
2066 } else if (tcp_close_state(sk)) {
2067 /* We FIN if the application ate all the data before
2068 * zapping the connection.
2069 */
2070
2071 /* RED-PEN. Formally speaking, we have broken TCP state
2072 * machine. State transitions:
2073 *
2074 * TCP_ESTABLISHED -> TCP_FIN_WAIT1
2075 * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible)
2076 * TCP_CLOSE_WAIT -> TCP_LAST_ACK
2077 *
2078 * are legal only when FIN has been sent (i.e. in window),
2079 * rather than queued out of window. Purists blame.
2080 *
2081 * F.e. "RFC state" is ESTABLISHED,
2082 * if Linux state is FIN-WAIT-1, but FIN is still not sent.
2083 *
2084 * The visible declinations are that sometimes
2085 * we enter time-wait state, when it is not required really
2086 * (harmless), do not send active resets, when they are
2087 * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
2088 * they look as CLOSING or LAST_ACK for Linux)
2089 * Probably, I missed some more holelets.
2090 * --ANK
2091 */
2092 tcp_send_fin(sk);
2093 }
2094
2095 sk_stream_wait_close(sk, timeout);
2096
2097adjudge_to_death:
2098 state = sk->sk_state;
2099 sock_hold(sk);
2100 sock_orphan(sk);
2101
2102 /* It is the last release_sock in its life. It will remove backlog. */
2103 release_sock(sk);
2104
2105
2106 /* Now socket is owned by kernel and we acquire BH lock
2107 to finish close. No need to check for user refs.
2108 */
2109 local_bh_disable();
2110 bh_lock_sock(sk);
2111 WARN_ON(sock_owned_by_user(sk));
2112
2113 percpu_counter_inc(sk->sk_prot->orphan_count);
2114
2115 /* Have we already been destroyed by a softirq or backlog? */
2116 if (state != TCP_CLOSE && sk->sk_state == TCP_CLOSE)
2117 goto out;
2118
2119 /* This is a (useful) BSD violating of the RFC. There is a
2120 * problem with TCP as specified in that the other end could
2121 * keep a socket open forever with no application left this end.
2122 * We use a 3 minute timeout (about the same as BSD) then kill
2123 * our end. If they send after that then tough - BUT: long enough
2124 * that we won't make the old 4*rto = almost no time - whoops
2125 * reset mistake.
2126 *
2127 * Nope, it was not mistake. It is really desired behaviour
2128 * f.e. on http servers, when such sockets are useless, but
2129 * consume significant resources. Let's do it with special
2130 * linger2 option. --ANK
2131 */
2132
2133 if (sk->sk_state == TCP_FIN_WAIT2) {
2134 struct tcp_sock *tp = tcp_sk(sk);
2135 if (tp->linger2 < 0) {
2136 tcp_set_state(sk, TCP_CLOSE);
2137 tcp_send_active_reset(sk, GFP_ATOMIC);
2138 NET_INC_STATS_BH(sock_net(sk),
2139 LINUX_MIB_TCPABORTONLINGER);
2140 } else {
2141 const int tmo = tcp_fin_time(sk);
2142
2143 if (tmo > TCP_TIMEWAIT_LEN) {
2144 inet_csk_reset_keepalive_timer(sk,
2145 tmo - TCP_TIMEWAIT_LEN);
2146 } else {
2147 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
2148 goto out;
2149 }
2150 }
2151 }
2152 if (sk->sk_state != TCP_CLOSE) {
2153 sk_mem_reclaim(sk);
2154 if (tcp_check_oom(sk, 0)) {
2155 tcp_set_state(sk, TCP_CLOSE);
2156 tcp_send_active_reset(sk, GFP_ATOMIC);
2157 NET_INC_STATS_BH(sock_net(sk),
2158 LINUX_MIB_TCPABORTONMEMORY);
2159 }
2160 }
2161
2162 if (sk->sk_state == TCP_CLOSE)
2163 inet_csk_destroy_sock(sk);
2164 /* Otherwise, socket is reprieved until protocol close. */
2165
2166out:
2167 bh_unlock_sock(sk);
2168 local_bh_enable();
2169 sock_put(sk);
2170}
2171EXPORT_SYMBOL(tcp_close);
2172
2173/* These states need RST on ABORT according to RFC793 */
2174
2175static inline bool tcp_need_reset(int state)
2176{
2177 return (1 << state) &
2178 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 |
2179 TCPF_FIN_WAIT2 | TCPF_SYN_RECV);
2180}
2181
2182int tcp_disconnect(struct sock *sk, int flags)
2183{
2184 struct inet_sock *inet = inet_sk(sk);
2185 struct inet_connection_sock *icsk = inet_csk(sk);
2186 struct tcp_sock *tp = tcp_sk(sk);
2187 int err = 0;
2188 int old_state = sk->sk_state;
2189
2190 if (old_state != TCP_CLOSE)
2191 tcp_set_state(sk, TCP_CLOSE);
2192
2193 /* ABORT function of RFC793 */
2194 if (old_state == TCP_LISTEN) {
2195 inet_csk_listen_stop(sk);
2196 } else if (unlikely(tp->repair)) {
2197 sk->sk_err = ECONNABORTED;
2198 } else if (tcp_need_reset(old_state) ||
2199 (tp->snd_nxt != tp->write_seq &&
2200 (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK))) {
2201 /* The last check adjusts for discrepancy of Linux wrt. RFC
2202 * states
2203 */
2204 tcp_send_active_reset(sk, gfp_any());
2205 sk->sk_err = ECONNRESET;
2206 } else if (old_state == TCP_SYN_SENT)
2207 sk->sk_err = ECONNRESET;
2208
2209 tcp_clear_xmit_timers(sk);
2210 __skb_queue_purge(&sk->sk_receive_queue);
2211 tcp_write_queue_purge(sk);
2212 __skb_queue_purge(&tp->out_of_order_queue);
2213#ifdef CONFIG_NET_DMA
2214 __skb_queue_purge(&sk->sk_async_wait_queue);
2215#endif
2216
2217 inet->inet_dport = 0;
2218
2219 if (!(sk->sk_userlocks & SOCK_BINDADDR_LOCK))
2220 inet_reset_saddr(sk);
2221
2222 sk->sk_shutdown = 0;
2223 sock_reset_flag(sk, SOCK_DONE);
2224 tp->srtt = 0;
2225 if ((tp->write_seq += tp->max_window + 2) == 0)
2226 tp->write_seq = 1;
2227 icsk->icsk_backoff = 0;
2228 tp->snd_cwnd = 2;
2229 icsk->icsk_probes_out = 0;
2230 tp->packets_out = 0;
2231 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
2232 tp->snd_cwnd_cnt = 0;
2233 tp->bytes_acked = 0;
2234 tp->window_clamp = 0;
2235 tcp_set_ca_state(sk, TCP_CA_Open);
2236 tcp_clear_retrans(tp);
2237 inet_csk_delack_init(sk);
2238 tcp_init_send_head(sk);
2239 memset(&tp->rx_opt, 0, sizeof(tp->rx_opt));
2240 __sk_dst_reset(sk);
2241
2242 WARN_ON(inet->inet_num && !icsk->icsk_bind_hash);
2243
2244 sk->sk_error_report(sk);
2245 return err;
2246}
2247EXPORT_SYMBOL(tcp_disconnect);
2248
2249static inline bool tcp_can_repair_sock(const struct sock *sk)
2250{
2251 return capable(CAP_NET_ADMIN) &&
2252 ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_ESTABLISHED));
2253}
2254
2255static int tcp_repair_options_est(struct tcp_sock *tp,
2256 struct tcp_repair_opt __user *optbuf, unsigned int len)
2257{
2258 struct tcp_repair_opt opt;
2259
2260 while (len >= sizeof(opt)) {
2261 if (copy_from_user(&opt, optbuf, sizeof(opt)))
2262 return -EFAULT;
2263
2264 optbuf++;
2265 len -= sizeof(opt);
2266
2267 switch (opt.opt_code) {
2268 case TCPOPT_MSS:
2269 tp->rx_opt.mss_clamp = opt.opt_val;
2270 break;
2271 case TCPOPT_WINDOW:
2272 if (opt.opt_val > 14)
2273 return -EFBIG;
2274
2275 tp->rx_opt.snd_wscale = opt.opt_val;
2276 break;
2277 case TCPOPT_SACK_PERM:
2278 if (opt.opt_val != 0)
2279 return -EINVAL;
2280
2281 tp->rx_opt.sack_ok |= TCP_SACK_SEEN;
2282 if (sysctl_tcp_fack)
2283 tcp_enable_fack(tp);
2284 break;
2285 case TCPOPT_TIMESTAMP:
2286 if (opt.opt_val != 0)
2287 return -EINVAL;
2288
2289 tp->rx_opt.tstamp_ok = 1;
2290 break;
2291 }
2292 }
2293
2294 return 0;
2295}
2296
2297/*
2298 * Socket option code for TCP.
2299 */
2300static int do_tcp_setsockopt(struct sock *sk, int level,
2301 int optname, char __user *optval, unsigned int optlen)
2302{
2303 struct tcp_sock *tp = tcp_sk(sk);
2304 struct inet_connection_sock *icsk = inet_csk(sk);
2305 int val;
2306 int err = 0;
2307
2308 /* These are data/string values, all the others are ints */
2309 switch (optname) {
2310 case TCP_CONGESTION: {
2311 char name[TCP_CA_NAME_MAX];
2312
2313 if (optlen < 1)
2314 return -EINVAL;
2315
2316 val = strncpy_from_user(name, optval,
2317 min_t(long, TCP_CA_NAME_MAX-1, optlen));
2318 if (val < 0)
2319 return -EFAULT;
2320 name[val] = 0;
2321
2322 lock_sock(sk);
2323 err = tcp_set_congestion_control(sk, name);
2324 release_sock(sk);
2325 return err;
2326 }
2327 case TCP_COOKIE_TRANSACTIONS: {
2328 struct tcp_cookie_transactions ctd;
2329 struct tcp_cookie_values *cvp = NULL;
2330
2331 if (sizeof(ctd) > optlen)
2332 return -EINVAL;
2333 if (copy_from_user(&ctd, optval, sizeof(ctd)))
2334 return -EFAULT;
2335
2336 if (ctd.tcpct_used > sizeof(ctd.tcpct_value) ||
2337 ctd.tcpct_s_data_desired > TCP_MSS_DESIRED)
2338 return -EINVAL;
2339
2340 if (ctd.tcpct_cookie_desired == 0) {
2341 /* default to global value */
2342 } else if ((0x1 & ctd.tcpct_cookie_desired) ||
2343 ctd.tcpct_cookie_desired > TCP_COOKIE_MAX ||
2344 ctd.tcpct_cookie_desired < TCP_COOKIE_MIN) {
2345 return -EINVAL;
2346 }
2347
2348 if (TCP_COOKIE_OUT_NEVER & ctd.tcpct_flags) {
2349 /* Supercedes all other values */
2350 lock_sock(sk);
2351 if (tp->cookie_values != NULL) {
2352 kref_put(&tp->cookie_values->kref,
2353 tcp_cookie_values_release);
2354 tp->cookie_values = NULL;
2355 }
2356 tp->rx_opt.cookie_in_always = 0; /* false */
2357 tp->rx_opt.cookie_out_never = 1; /* true */
2358 release_sock(sk);
2359 return err;
2360 }
2361
2362 /* Allocate ancillary memory before locking.
2363 */
2364 if (ctd.tcpct_used > 0 ||
2365 (tp->cookie_values == NULL &&
2366 (sysctl_tcp_cookie_size > 0 ||
2367 ctd.tcpct_cookie_desired > 0 ||
2368 ctd.tcpct_s_data_desired > 0))) {
2369 cvp = kzalloc(sizeof(*cvp) + ctd.tcpct_used,
2370 GFP_KERNEL);
2371 if (cvp == NULL)
2372 return -ENOMEM;
2373
2374 kref_init(&cvp->kref);
2375 }
2376 lock_sock(sk);
2377 tp->rx_opt.cookie_in_always =
2378 (TCP_COOKIE_IN_ALWAYS & ctd.tcpct_flags);
2379 tp->rx_opt.cookie_out_never = 0; /* false */
2380
2381 if (tp->cookie_values != NULL) {
2382 if (cvp != NULL) {
2383 /* Changed values are recorded by a changed
2384 * pointer, ensuring the cookie will differ,
2385 * without separately hashing each value later.
2386 */
2387 kref_put(&tp->cookie_values->kref,
2388 tcp_cookie_values_release);
2389 } else {
2390 cvp = tp->cookie_values;
2391 }
2392 }
2393
2394 if (cvp != NULL) {
2395 cvp->cookie_desired = ctd.tcpct_cookie_desired;
2396
2397 if (ctd.tcpct_used > 0) {
2398 memcpy(cvp->s_data_payload, ctd.tcpct_value,
2399 ctd.tcpct_used);
2400 cvp->s_data_desired = ctd.tcpct_used;
2401 cvp->s_data_constant = 1; /* true */
2402 } else {
2403 /* No constant payload data. */
2404 cvp->s_data_desired = ctd.tcpct_s_data_desired;
2405 cvp->s_data_constant = 0; /* false */
2406 }
2407
2408 tp->cookie_values = cvp;
2409 }
2410 release_sock(sk);
2411 return err;
2412 }
2413 default:
2414 /* fallthru */
2415 break;
2416 }
2417
2418 if (optlen < sizeof(int))
2419 return -EINVAL;
2420
2421 if (get_user(val, (int __user *)optval))
2422 return -EFAULT;
2423
2424 lock_sock(sk);
2425
2426 switch (optname) {
2427 case TCP_MAXSEG:
2428 /* Values greater than interface MTU won't take effect. However
2429 * at the point when this call is done we typically don't yet
2430 * know which interface is going to be used */
2431 if (val < TCP_MIN_MSS || val > MAX_TCP_WINDOW) {
2432 err = -EINVAL;
2433 break;
2434 }
2435 tp->rx_opt.user_mss = val;
2436 break;
2437
2438 case TCP_NODELAY:
2439 if (val) {
2440 /* TCP_NODELAY is weaker than TCP_CORK, so that
2441 * this option on corked socket is remembered, but
2442 * it is not activated until cork is cleared.
2443 *
2444 * However, when TCP_NODELAY is set we make
2445 * an explicit push, which overrides even TCP_CORK
2446 * for currently queued segments.
2447 */
2448 tp->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH;
2449 tcp_push_pending_frames(sk);
2450 } else {
2451 tp->nonagle &= ~TCP_NAGLE_OFF;
2452 }
2453 break;
2454
2455 case TCP_THIN_LINEAR_TIMEOUTS:
2456 if (val < 0 || val > 1)
2457 err = -EINVAL;
2458 else
2459 tp->thin_lto = val;
2460 break;
2461
2462 case TCP_THIN_DUPACK:
2463 if (val < 0 || val > 1)
2464 err = -EINVAL;
2465 else
2466 tp->thin_dupack = val;
2467 if (tp->thin_dupack)
2468 tcp_disable_early_retrans(tp);
2469 break;
2470
2471 case TCP_REPAIR:
2472 if (!tcp_can_repair_sock(sk))
2473 err = -EPERM;
2474 else if (val == 1) {
2475 tp->repair = 1;
2476 sk->sk_reuse = SK_FORCE_REUSE;
2477 tp->repair_queue = TCP_NO_QUEUE;
2478 } else if (val == 0) {
2479 tp->repair = 0;
2480 sk->sk_reuse = SK_NO_REUSE;
2481 tcp_send_window_probe(sk);
2482 } else
2483 err = -EINVAL;
2484
2485 break;
2486
2487 case TCP_REPAIR_QUEUE:
2488 if (!tp->repair)
2489 err = -EPERM;
2490 else if (val < TCP_QUEUES_NR)
2491 tp->repair_queue = val;
2492 else
2493 err = -EINVAL;
2494 break;
2495
2496 case TCP_QUEUE_SEQ:
2497 if (sk->sk_state != TCP_CLOSE)
2498 err = -EPERM;
2499 else if (tp->repair_queue == TCP_SEND_QUEUE)
2500 tp->write_seq = val;
2501 else if (tp->repair_queue == TCP_RECV_QUEUE)
2502 tp->rcv_nxt = val;
2503 else
2504 err = -EINVAL;
2505 break;
2506
2507 case TCP_REPAIR_OPTIONS:
2508 if (!tp->repair)
2509 err = -EINVAL;
2510 else if (sk->sk_state == TCP_ESTABLISHED)
2511 err = tcp_repair_options_est(tp,
2512 (struct tcp_repair_opt __user *)optval,
2513 optlen);
2514 else
2515 err = -EPERM;
2516 break;
2517
2518 case TCP_CORK:
2519 /* When set indicates to always queue non-full frames.
2520 * Later the user clears this option and we transmit
2521 * any pending partial frames in the queue. This is
2522 * meant to be used alongside sendfile() to get properly
2523 * filled frames when the user (for example) must write
2524 * out headers with a write() call first and then use
2525 * sendfile to send out the data parts.
2526 *
2527 * TCP_CORK can be set together with TCP_NODELAY and it is
2528 * stronger than TCP_NODELAY.
2529 */
2530 if (val) {
2531 tp->nonagle |= TCP_NAGLE_CORK;
2532 } else {
2533 tp->nonagle &= ~TCP_NAGLE_CORK;
2534 if (tp->nonagle&TCP_NAGLE_OFF)
2535 tp->nonagle |= TCP_NAGLE_PUSH;
2536 tcp_push_pending_frames(sk);
2537 }
2538 break;
2539
2540 case TCP_KEEPIDLE:
2541 if (val < 1 || val > MAX_TCP_KEEPIDLE)
2542 err = -EINVAL;
2543 else {
2544 tp->keepalive_time = val * HZ;
2545 if (sock_flag(sk, SOCK_KEEPOPEN) &&
2546 !((1 << sk->sk_state) &
2547 (TCPF_CLOSE | TCPF_LISTEN))) {
2548 u32 elapsed = keepalive_time_elapsed(tp);
2549 if (tp->keepalive_time > elapsed)
2550 elapsed = tp->keepalive_time - elapsed;
2551 else
2552 elapsed = 0;
2553 inet_csk_reset_keepalive_timer(sk, elapsed);
2554 }
2555 }
2556 break;
2557 case TCP_KEEPINTVL:
2558 if (val < 1 || val > MAX_TCP_KEEPINTVL)
2559 err = -EINVAL;
2560 else
2561 tp->keepalive_intvl = val * HZ;
2562 break;
2563 case TCP_KEEPCNT:
2564 if (val < 1 || val > MAX_TCP_KEEPCNT)
2565 err = -EINVAL;
2566 else
2567 tp->keepalive_probes = val;
2568 break;
2569 case TCP_SYNCNT:
2570 if (val < 1 || val > MAX_TCP_SYNCNT)
2571 err = -EINVAL;
2572 else
2573 icsk->icsk_syn_retries = val;
2574 break;
2575
2576 case TCP_LINGER2:
2577 if (val < 0)
2578 tp->linger2 = -1;
2579 else if (val > sysctl_tcp_fin_timeout / HZ)
2580 tp->linger2 = 0;
2581 else
2582 tp->linger2 = val * HZ;
2583 break;
2584
2585 case TCP_DEFER_ACCEPT:
2586 /* Translate value in seconds to number of retransmits */
2587 icsk->icsk_accept_queue.rskq_defer_accept =
2588 secs_to_retrans(val, TCP_TIMEOUT_INIT / HZ,
2589 TCP_RTO_MAX / HZ);
2590 break;
2591
2592 case TCP_WINDOW_CLAMP:
2593 if (!val) {
2594 if (sk->sk_state != TCP_CLOSE) {
2595 err = -EINVAL;
2596 break;
2597 }
2598 tp->window_clamp = 0;
2599 } else
2600 tp->window_clamp = val < SOCK_MIN_RCVBUF / 2 ?
2601 SOCK_MIN_RCVBUF / 2 : val;
2602 break;
2603
2604 case TCP_QUICKACK:
2605 if (!val) {
2606 icsk->icsk_ack.pingpong = 1;
2607 } else {
2608 icsk->icsk_ack.pingpong = 0;
2609 if ((1 << sk->sk_state) &
2610 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) &&
2611 inet_csk_ack_scheduled(sk)) {
2612 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
2613 tcp_cleanup_rbuf(sk, 1);
2614 if (!(val & 1))
2615 icsk->icsk_ack.pingpong = 1;
2616 }
2617 }
2618 break;
2619
2620#ifdef CONFIG_TCP_MD5SIG
2621 case TCP_MD5SIG:
2622 /* Read the IP->Key mappings from userspace */
2623 err = tp->af_specific->md5_parse(sk, optval, optlen);
2624 break;
2625#endif
2626 case TCP_USER_TIMEOUT:
2627 /* Cap the max timeout in ms TCP will retry/retrans
2628 * before giving up and aborting (ETIMEDOUT) a connection.
2629 */
2630 if (val < 0)
2631 err = -EINVAL;
2632 else
2633 icsk->icsk_user_timeout = msecs_to_jiffies(val);
2634 break;
2635 default:
2636 err = -ENOPROTOOPT;
2637 break;
2638 }
2639
2640 release_sock(sk);
2641 return err;
2642}
2643
2644int tcp_setsockopt(struct sock *sk, int level, int optname, char __user *optval,
2645 unsigned int optlen)
2646{
2647 const struct inet_connection_sock *icsk = inet_csk(sk);
2648
2649 if (level != SOL_TCP)
2650 return icsk->icsk_af_ops->setsockopt(sk, level, optname,
2651 optval, optlen);
2652 return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2653}
2654EXPORT_SYMBOL(tcp_setsockopt);
2655
2656#ifdef CONFIG_COMPAT
2657int compat_tcp_setsockopt(struct sock *sk, int level, int optname,
2658 char __user *optval, unsigned int optlen)
2659{
2660 if (level != SOL_TCP)
2661 return inet_csk_compat_setsockopt(sk, level, optname,
2662 optval, optlen);
2663 return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2664}
2665EXPORT_SYMBOL(compat_tcp_setsockopt);
2666#endif
2667
2668/* Return information about state of tcp endpoint in API format. */
2669void tcp_get_info(const struct sock *sk, struct tcp_info *info)
2670{
2671 const struct tcp_sock *tp = tcp_sk(sk);
2672 const struct inet_connection_sock *icsk = inet_csk(sk);
2673 u32 now = tcp_time_stamp;
2674
2675 memset(info, 0, sizeof(*info));
2676
2677 info->tcpi_state = sk->sk_state;
2678 info->tcpi_ca_state = icsk->icsk_ca_state;
2679 info->tcpi_retransmits = icsk->icsk_retransmits;
2680 info->tcpi_probes = icsk->icsk_probes_out;
2681 info->tcpi_backoff = icsk->icsk_backoff;
2682
2683 if (tp->rx_opt.tstamp_ok)
2684 info->tcpi_options |= TCPI_OPT_TIMESTAMPS;
2685 if (tcp_is_sack(tp))
2686 info->tcpi_options |= TCPI_OPT_SACK;
2687 if (tp->rx_opt.wscale_ok) {
2688 info->tcpi_options |= TCPI_OPT_WSCALE;
2689 info->tcpi_snd_wscale = tp->rx_opt.snd_wscale;
2690 info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale;
2691 }
2692
2693 if (tp->ecn_flags & TCP_ECN_OK)
2694 info->tcpi_options |= TCPI_OPT_ECN;
2695 if (tp->ecn_flags & TCP_ECN_SEEN)
2696 info->tcpi_options |= TCPI_OPT_ECN_SEEN;
2697
2698 info->tcpi_rto = jiffies_to_usecs(icsk->icsk_rto);
2699 info->tcpi_ato = jiffies_to_usecs(icsk->icsk_ack.ato);
2700 info->tcpi_snd_mss = tp->mss_cache;
2701 info->tcpi_rcv_mss = icsk->icsk_ack.rcv_mss;
2702
2703 if (sk->sk_state == TCP_LISTEN) {
2704 info->tcpi_unacked = sk->sk_ack_backlog;
2705 info->tcpi_sacked = sk->sk_max_ack_backlog;
2706 } else {
2707 info->tcpi_unacked = tp->packets_out;
2708 info->tcpi_sacked = tp->sacked_out;
2709 }
2710 info->tcpi_lost = tp->lost_out;
2711 info->tcpi_retrans = tp->retrans_out;
2712 info->tcpi_fackets = tp->fackets_out;
2713
2714 info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime);
2715 info->tcpi_last_data_recv = jiffies_to_msecs(now - icsk->icsk_ack.lrcvtime);
2716 info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp);
2717
2718 info->tcpi_pmtu = icsk->icsk_pmtu_cookie;
2719 info->tcpi_rcv_ssthresh = tp->rcv_ssthresh;
2720 info->tcpi_rtt = jiffies_to_usecs(tp->srtt)>>3;
2721 info->tcpi_rttvar = jiffies_to_usecs(tp->mdev)>>2;
2722 info->tcpi_snd_ssthresh = tp->snd_ssthresh;
2723 info->tcpi_snd_cwnd = tp->snd_cwnd;
2724 info->tcpi_advmss = tp->advmss;
2725 info->tcpi_reordering = tp->reordering;
2726
2727 info->tcpi_rcv_rtt = jiffies_to_usecs(tp->rcv_rtt_est.rtt)>>3;
2728 info->tcpi_rcv_space = tp->rcvq_space.space;
2729
2730 info->tcpi_total_retrans = tp->total_retrans;
2731}
2732EXPORT_SYMBOL_GPL(tcp_get_info);
2733
2734static int do_tcp_getsockopt(struct sock *sk, int level,
2735 int optname, char __user *optval, int __user *optlen)
2736{
2737 struct inet_connection_sock *icsk = inet_csk(sk);
2738 struct tcp_sock *tp = tcp_sk(sk);
2739 int val, len;
2740
2741 if (get_user(len, optlen))
2742 return -EFAULT;
2743
2744 len = min_t(unsigned int, len, sizeof(int));
2745
2746 if (len < 0)
2747 return -EINVAL;
2748
2749 switch (optname) {
2750 case TCP_MAXSEG:
2751 val = tp->mss_cache;
2752 if (!val && ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
2753 val = tp->rx_opt.user_mss;
2754 if (tp->repair)
2755 val = tp->rx_opt.mss_clamp;
2756 break;
2757 case TCP_NODELAY:
2758 val = !!(tp->nonagle&TCP_NAGLE_OFF);
2759 break;
2760 case TCP_CORK:
2761 val = !!(tp->nonagle&TCP_NAGLE_CORK);
2762 break;
2763 case TCP_KEEPIDLE:
2764 val = keepalive_time_when(tp) / HZ;
2765 break;
2766 case TCP_KEEPINTVL:
2767 val = keepalive_intvl_when(tp) / HZ;
2768 break;
2769 case TCP_KEEPCNT:
2770 val = keepalive_probes(tp);
2771 break;
2772 case TCP_SYNCNT:
2773 val = icsk->icsk_syn_retries ? : sysctl_tcp_syn_retries;
2774 break;
2775 case TCP_LINGER2:
2776 val = tp->linger2;
2777 if (val >= 0)
2778 val = (val ? : sysctl_tcp_fin_timeout) / HZ;
2779 break;
2780 case TCP_DEFER_ACCEPT:
2781 val = retrans_to_secs(icsk->icsk_accept_queue.rskq_defer_accept,
2782 TCP_TIMEOUT_INIT / HZ, TCP_RTO_MAX / HZ);
2783 break;
2784 case TCP_WINDOW_CLAMP:
2785 val = tp->window_clamp;
2786 break;
2787 case TCP_INFO: {
2788 struct tcp_info info;
2789
2790 if (get_user(len, optlen))
2791 return -EFAULT;
2792
2793 tcp_get_info(sk, &info);
2794
2795 len = min_t(unsigned int, len, sizeof(info));
2796 if (put_user(len, optlen))
2797 return -EFAULT;
2798 if (copy_to_user(optval, &info, len))
2799 return -EFAULT;
2800 return 0;
2801 }
2802 case TCP_QUICKACK:
2803 val = !icsk->icsk_ack.pingpong;
2804 break;
2805
2806 case TCP_CONGESTION:
2807 if (get_user(len, optlen))
2808 return -EFAULT;
2809 len = min_t(unsigned int, len, TCP_CA_NAME_MAX);
2810 if (put_user(len, optlen))
2811 return -EFAULT;
2812 if (copy_to_user(optval, icsk->icsk_ca_ops->name, len))
2813 return -EFAULT;
2814 return 0;
2815
2816 case TCP_COOKIE_TRANSACTIONS: {
2817 struct tcp_cookie_transactions ctd;
2818 struct tcp_cookie_values *cvp = tp->cookie_values;
2819
2820 if (get_user(len, optlen))
2821 return -EFAULT;
2822 if (len < sizeof(ctd))
2823 return -EINVAL;
2824
2825 memset(&ctd, 0, sizeof(ctd));
2826 ctd.tcpct_flags = (tp->rx_opt.cookie_in_always ?
2827 TCP_COOKIE_IN_ALWAYS : 0)
2828 | (tp->rx_opt.cookie_out_never ?
2829 TCP_COOKIE_OUT_NEVER : 0);
2830
2831 if (cvp != NULL) {
2832 ctd.tcpct_flags |= (cvp->s_data_in ?
2833 TCP_S_DATA_IN : 0)
2834 | (cvp->s_data_out ?
2835 TCP_S_DATA_OUT : 0);
2836
2837 ctd.tcpct_cookie_desired = cvp->cookie_desired;
2838 ctd.tcpct_s_data_desired = cvp->s_data_desired;
2839
2840 memcpy(&ctd.tcpct_value[0], &cvp->cookie_pair[0],
2841 cvp->cookie_pair_size);
2842 ctd.tcpct_used = cvp->cookie_pair_size;
2843 }
2844
2845 if (put_user(sizeof(ctd), optlen))
2846 return -EFAULT;
2847 if (copy_to_user(optval, &ctd, sizeof(ctd)))
2848 return -EFAULT;
2849 return 0;
2850 }
2851 case TCP_THIN_LINEAR_TIMEOUTS:
2852 val = tp->thin_lto;
2853 break;
2854 case TCP_THIN_DUPACK:
2855 val = tp->thin_dupack;
2856 break;
2857
2858 case TCP_REPAIR:
2859 val = tp->repair;
2860 break;
2861
2862 case TCP_REPAIR_QUEUE:
2863 if (tp->repair)
2864 val = tp->repair_queue;
2865 else
2866 return -EINVAL;
2867 break;
2868
2869 case TCP_QUEUE_SEQ:
2870 if (tp->repair_queue == TCP_SEND_QUEUE)
2871 val = tp->write_seq;
2872 else if (tp->repair_queue == TCP_RECV_QUEUE)
2873 val = tp->rcv_nxt;
2874 else
2875 return -EINVAL;
2876 break;
2877
2878 case TCP_USER_TIMEOUT:
2879 val = jiffies_to_msecs(icsk->icsk_user_timeout);
2880 break;
2881 default:
2882 return -ENOPROTOOPT;
2883 }
2884
2885 if (put_user(len, optlen))
2886 return -EFAULT;
2887 if (copy_to_user(optval, &val, len))
2888 return -EFAULT;
2889 return 0;
2890}
2891
2892int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval,
2893 int __user *optlen)
2894{
2895 struct inet_connection_sock *icsk = inet_csk(sk);
2896
2897 if (level != SOL_TCP)
2898 return icsk->icsk_af_ops->getsockopt(sk, level, optname,
2899 optval, optlen);
2900 return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2901}
2902EXPORT_SYMBOL(tcp_getsockopt);
2903
2904#ifdef CONFIG_COMPAT
2905int compat_tcp_getsockopt(struct sock *sk, int level, int optname,
2906 char __user *optval, int __user *optlen)
2907{
2908 if (level != SOL_TCP)
2909 return inet_csk_compat_getsockopt(sk, level, optname,
2910 optval, optlen);
2911 return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2912}
2913EXPORT_SYMBOL(compat_tcp_getsockopt);
2914#endif
2915
2916struct sk_buff *tcp_tso_segment(struct sk_buff *skb,
2917 netdev_features_t features)
2918{
2919 struct sk_buff *segs = ERR_PTR(-EINVAL);
2920 struct tcphdr *th;
2921 unsigned int thlen;
2922 unsigned int seq;
2923 __be32 delta;
2924 unsigned int oldlen;
2925 unsigned int mss;
2926
2927 if (!pskb_may_pull(skb, sizeof(*th)))
2928 goto out;
2929
2930 th = tcp_hdr(skb);
2931 thlen = th->doff * 4;
2932 if (thlen < sizeof(*th))
2933 goto out;
2934
2935 if (!pskb_may_pull(skb, thlen))
2936 goto out;
2937
2938 oldlen = (u16)~skb->len;
2939 __skb_pull(skb, thlen);
2940
2941 mss = skb_shinfo(skb)->gso_size;
2942 if (unlikely(skb->len <= mss))
2943 goto out;
2944
2945 if (skb_gso_ok(skb, features | NETIF_F_GSO_ROBUST)) {
2946 /* Packet is from an untrusted source, reset gso_segs. */
2947 int type = skb_shinfo(skb)->gso_type;
2948
2949 if (unlikely(type &
2950 ~(SKB_GSO_TCPV4 |
2951 SKB_GSO_DODGY |
2952 SKB_GSO_TCP_ECN |
2953 SKB_GSO_TCPV6 |
2954 0) ||
2955 !(type & (SKB_GSO_TCPV4 | SKB_GSO_TCPV6))))
2956 goto out;
2957
2958 skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss);
2959
2960 segs = NULL;
2961 goto out;
2962 }
2963
2964 segs = skb_segment(skb, features);
2965 if (IS_ERR(segs))
2966 goto out;
2967
2968 delta = htonl(oldlen + (thlen + mss));
2969
2970 skb = segs;
2971 th = tcp_hdr(skb);
2972 seq = ntohl(th->seq);
2973
2974 do {
2975 th->fin = th->psh = 0;
2976
2977 th->check = ~csum_fold((__force __wsum)((__force u32)th->check +
2978 (__force u32)delta));
2979 if (skb->ip_summed != CHECKSUM_PARTIAL)
2980 th->check =
2981 csum_fold(csum_partial(skb_transport_header(skb),
2982 thlen, skb->csum));
2983
2984 seq += mss;
2985 skb = skb->next;
2986 th = tcp_hdr(skb);
2987
2988 th->seq = htonl(seq);
2989 th->cwr = 0;
2990 } while (skb->next);
2991
2992 delta = htonl(oldlen + (skb->tail - skb->transport_header) +
2993 skb->data_len);
2994 th->check = ~csum_fold((__force __wsum)((__force u32)th->check +
2995 (__force u32)delta));
2996 if (skb->ip_summed != CHECKSUM_PARTIAL)
2997 th->check = csum_fold(csum_partial(skb_transport_header(skb),
2998 thlen, skb->csum));
2999
3000out:
3001 return segs;
3002}
3003EXPORT_SYMBOL(tcp_tso_segment);
3004
3005struct sk_buff **tcp_gro_receive(struct sk_buff **head, struct sk_buff *skb)
3006{
3007 struct sk_buff **pp = NULL;
3008 struct sk_buff *p;
3009 struct tcphdr *th;
3010 struct tcphdr *th2;
3011 unsigned int len;
3012 unsigned int thlen;
3013 __be32 flags;
3014 unsigned int mss = 1;
3015 unsigned int hlen;
3016 unsigned int off;
3017 int flush = 1;
3018 int i;
3019
3020 off = skb_gro_offset(skb);
3021 hlen = off + sizeof(*th);
3022 th = skb_gro_header_fast(skb, off);
3023 if (skb_gro_header_hard(skb, hlen)) {
3024 th = skb_gro_header_slow(skb, hlen, off);
3025 if (unlikely(!th))
3026 goto out;
3027 }
3028
3029 thlen = th->doff * 4;
3030 if (thlen < sizeof(*th))
3031 goto out;
3032
3033 hlen = off + thlen;
3034 if (skb_gro_header_hard(skb, hlen)) {
3035 th = skb_gro_header_slow(skb, hlen, off);
3036 if (unlikely(!th))
3037 goto out;
3038 }
3039
3040 skb_gro_pull(skb, thlen);
3041
3042 len = skb_gro_len(skb);
3043 flags = tcp_flag_word(th);
3044
3045 for (; (p = *head); head = &p->next) {
3046 if (!NAPI_GRO_CB(p)->same_flow)
3047 continue;
3048
3049 th2 = tcp_hdr(p);
3050
3051 if (*(u32 *)&th->source ^ *(u32 *)&th2->source) {
3052 NAPI_GRO_CB(p)->same_flow = 0;
3053 continue;
3054 }
3055
3056 goto found;
3057 }
3058
3059 goto out_check_final;
3060
3061found:
3062 flush = NAPI_GRO_CB(p)->flush;
3063 flush |= (__force int)(flags & TCP_FLAG_CWR);
3064 flush |= (__force int)((flags ^ tcp_flag_word(th2)) &
3065 ~(TCP_FLAG_CWR | TCP_FLAG_FIN | TCP_FLAG_PSH));
3066 flush |= (__force int)(th->ack_seq ^ th2->ack_seq);
3067 for (i = sizeof(*th); i < thlen; i += 4)
3068 flush |= *(u32 *)((u8 *)th + i) ^
3069 *(u32 *)((u8 *)th2 + i);
3070
3071 mss = skb_shinfo(p)->gso_size;
3072
3073 flush |= (len - 1) >= mss;
3074 flush |= (ntohl(th2->seq) + skb_gro_len(p)) ^ ntohl(th->seq);
3075
3076 if (flush || skb_gro_receive(head, skb)) {
3077 mss = 1;
3078 goto out_check_final;
3079 }
3080
3081 p = *head;
3082 th2 = tcp_hdr(p);
3083 tcp_flag_word(th2) |= flags & (TCP_FLAG_FIN | TCP_FLAG_PSH);
3084
3085out_check_final:
3086 flush = len < mss;
3087 flush |= (__force int)(flags & (TCP_FLAG_URG | TCP_FLAG_PSH |
3088 TCP_FLAG_RST | TCP_FLAG_SYN |
3089 TCP_FLAG_FIN));
3090
3091 if (p && (!NAPI_GRO_CB(skb)->same_flow || flush))
3092 pp = head;
3093
3094out:
3095 NAPI_GRO_CB(skb)->flush |= flush;
3096
3097 return pp;
3098}
3099EXPORT_SYMBOL(tcp_gro_receive);
3100
3101int tcp_gro_complete(struct sk_buff *skb)
3102{
3103 struct tcphdr *th = tcp_hdr(skb);
3104
3105 skb->csum_start = skb_transport_header(skb) - skb->head;
3106 skb->csum_offset = offsetof(struct tcphdr, check);
3107 skb->ip_summed = CHECKSUM_PARTIAL;
3108
3109 skb_shinfo(skb)->gso_segs = NAPI_GRO_CB(skb)->count;
3110
3111 if (th->cwr)
3112 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
3113
3114 return 0;
3115}
3116EXPORT_SYMBOL(tcp_gro_complete);
3117
3118#ifdef CONFIG_TCP_MD5SIG
3119static unsigned long tcp_md5sig_users;
3120static struct tcp_md5sig_pool __percpu *tcp_md5sig_pool;
3121static DEFINE_SPINLOCK(tcp_md5sig_pool_lock);
3122
3123static void __tcp_free_md5sig_pool(struct tcp_md5sig_pool __percpu *pool)
3124{
3125 int cpu;
3126
3127 for_each_possible_cpu(cpu) {
3128 struct tcp_md5sig_pool *p = per_cpu_ptr(pool, cpu);
3129
3130 if (p->md5_desc.tfm)
3131 crypto_free_hash(p->md5_desc.tfm);
3132 }
3133 free_percpu(pool);
3134}
3135
3136void tcp_free_md5sig_pool(void)
3137{
3138 struct tcp_md5sig_pool __percpu *pool = NULL;
3139
3140 spin_lock_bh(&tcp_md5sig_pool_lock);
3141 if (--tcp_md5sig_users == 0) {
3142 pool = tcp_md5sig_pool;
3143 tcp_md5sig_pool = NULL;
3144 }
3145 spin_unlock_bh(&tcp_md5sig_pool_lock);
3146 if (pool)
3147 __tcp_free_md5sig_pool(pool);
3148}
3149EXPORT_SYMBOL(tcp_free_md5sig_pool);
3150
3151static struct tcp_md5sig_pool __percpu *
3152__tcp_alloc_md5sig_pool(struct sock *sk)
3153{
3154 int cpu;
3155 struct tcp_md5sig_pool __percpu *pool;
3156
3157 pool = alloc_percpu(struct tcp_md5sig_pool);
3158 if (!pool)
3159 return NULL;
3160
3161 for_each_possible_cpu(cpu) {
3162 struct crypto_hash *hash;
3163
3164 hash = crypto_alloc_hash("md5", 0, CRYPTO_ALG_ASYNC);
3165 if (!hash || IS_ERR(hash))
3166 goto out_free;
3167
3168 per_cpu_ptr(pool, cpu)->md5_desc.tfm = hash;
3169 }
3170 return pool;
3171out_free:
3172 __tcp_free_md5sig_pool(pool);
3173 return NULL;
3174}
3175
3176struct tcp_md5sig_pool __percpu *tcp_alloc_md5sig_pool(struct sock *sk)
3177{
3178 struct tcp_md5sig_pool __percpu *pool;
3179 bool alloc = false;
3180
3181retry:
3182 spin_lock_bh(&tcp_md5sig_pool_lock);
3183 pool = tcp_md5sig_pool;
3184 if (tcp_md5sig_users++ == 0) {
3185 alloc = true;
3186 spin_unlock_bh(&tcp_md5sig_pool_lock);
3187 } else if (!pool) {
3188 tcp_md5sig_users--;
3189 spin_unlock_bh(&tcp_md5sig_pool_lock);
3190 cpu_relax();
3191 goto retry;
3192 } else
3193 spin_unlock_bh(&tcp_md5sig_pool_lock);
3194
3195 if (alloc) {
3196 /* we cannot hold spinlock here because this may sleep. */
3197 struct tcp_md5sig_pool __percpu *p;
3198
3199 p = __tcp_alloc_md5sig_pool(sk);
3200 spin_lock_bh(&tcp_md5sig_pool_lock);
3201 if (!p) {
3202 tcp_md5sig_users--;
3203 spin_unlock_bh(&tcp_md5sig_pool_lock);
3204 return NULL;
3205 }
3206 pool = tcp_md5sig_pool;
3207 if (pool) {
3208 /* oops, it has already been assigned. */
3209 spin_unlock_bh(&tcp_md5sig_pool_lock);
3210 __tcp_free_md5sig_pool(p);
3211 } else {
3212 tcp_md5sig_pool = pool = p;
3213 spin_unlock_bh(&tcp_md5sig_pool_lock);
3214 }
3215 }
3216 return pool;
3217}
3218EXPORT_SYMBOL(tcp_alloc_md5sig_pool);
3219
3220
3221/**
3222 * tcp_get_md5sig_pool - get md5sig_pool for this user
3223 *
3224 * We use percpu structure, so if we succeed, we exit with preemption
3225 * and BH disabled, to make sure another thread or softirq handling
3226 * wont try to get same context.
3227 */
3228struct tcp_md5sig_pool *tcp_get_md5sig_pool(void)
3229{
3230 struct tcp_md5sig_pool __percpu *p;
3231
3232 local_bh_disable();
3233
3234 spin_lock(&tcp_md5sig_pool_lock);
3235 p = tcp_md5sig_pool;
3236 if (p)
3237 tcp_md5sig_users++;
3238 spin_unlock(&tcp_md5sig_pool_lock);
3239
3240 if (p)
3241 return this_cpu_ptr(p);
3242
3243 local_bh_enable();
3244 return NULL;
3245}
3246EXPORT_SYMBOL(tcp_get_md5sig_pool);
3247
3248void tcp_put_md5sig_pool(void)
3249{
3250 local_bh_enable();
3251 tcp_free_md5sig_pool();
3252}
3253EXPORT_SYMBOL(tcp_put_md5sig_pool);
3254
3255int tcp_md5_hash_header(struct tcp_md5sig_pool *hp,
3256 const struct tcphdr *th)
3257{
3258 struct scatterlist sg;
3259 struct tcphdr hdr;
3260 int err;
3261
3262 /* We are not allowed to change tcphdr, make a local copy */
3263 memcpy(&hdr, th, sizeof(hdr));
3264 hdr.check = 0;
3265
3266 /* options aren't included in the hash */
3267 sg_init_one(&sg, &hdr, sizeof(hdr));
3268 err = crypto_hash_update(&hp->md5_desc, &sg, sizeof(hdr));
3269 return err;
3270}
3271EXPORT_SYMBOL(tcp_md5_hash_header);
3272
3273int tcp_md5_hash_skb_data(struct tcp_md5sig_pool *hp,
3274 const struct sk_buff *skb, unsigned int header_len)
3275{
3276 struct scatterlist sg;
3277 const struct tcphdr *tp = tcp_hdr(skb);
3278 struct hash_desc *desc = &hp->md5_desc;
3279 unsigned int i;
3280 const unsigned int head_data_len = skb_headlen(skb) > header_len ?
3281 skb_headlen(skb) - header_len : 0;
3282 const struct skb_shared_info *shi = skb_shinfo(skb);
3283 struct sk_buff *frag_iter;
3284
3285 sg_init_table(&sg, 1);
3286
3287 sg_set_buf(&sg, ((u8 *) tp) + header_len, head_data_len);
3288 if (crypto_hash_update(desc, &sg, head_data_len))
3289 return 1;
3290
3291 for (i = 0; i < shi->nr_frags; ++i) {
3292 const struct skb_frag_struct *f = &shi->frags[i];
3293 struct page *page = skb_frag_page(f);
3294 sg_set_page(&sg, page, skb_frag_size(f), f->page_offset);
3295 if (crypto_hash_update(desc, &sg, skb_frag_size(f)))
3296 return 1;
3297 }
3298
3299 skb_walk_frags(skb, frag_iter)
3300 if (tcp_md5_hash_skb_data(hp, frag_iter, 0))
3301 return 1;
3302
3303 return 0;
3304}
3305EXPORT_SYMBOL(tcp_md5_hash_skb_data);
3306
3307int tcp_md5_hash_key(struct tcp_md5sig_pool *hp, const struct tcp_md5sig_key *key)
3308{
3309 struct scatterlist sg;
3310
3311 sg_init_one(&sg, key->key, key->keylen);
3312 return crypto_hash_update(&hp->md5_desc, &sg, key->keylen);
3313}
3314EXPORT_SYMBOL(tcp_md5_hash_key);
3315
3316#endif
3317
3318/**
3319 * Each Responder maintains up to two secret values concurrently for
3320 * efficient secret rollover. Each secret value has 4 states:
3321 *
3322 * Generating. (tcp_secret_generating != tcp_secret_primary)
3323 * Generates new Responder-Cookies, but not yet used for primary
3324 * verification. This is a short-term state, typically lasting only
3325 * one round trip time (RTT).
3326 *
3327 * Primary. (tcp_secret_generating == tcp_secret_primary)
3328 * Used both for generation and primary verification.
3329 *
3330 * Retiring. (tcp_secret_retiring != tcp_secret_secondary)
3331 * Used for verification, until the first failure that can be
3332 * verified by the newer Generating secret. At that time, this
3333 * cookie's state is changed to Secondary, and the Generating
3334 * cookie's state is changed to Primary. This is a short-term state,
3335 * typically lasting only one round trip time (RTT).
3336 *
3337 * Secondary. (tcp_secret_retiring == tcp_secret_secondary)
3338 * Used for secondary verification, after primary verification
3339 * failures. This state lasts no more than twice the Maximum Segment
3340 * Lifetime (2MSL). Then, the secret is discarded.
3341 */
3342struct tcp_cookie_secret {
3343 /* The secret is divided into two parts. The digest part is the
3344 * equivalent of previously hashing a secret and saving the state,
3345 * and serves as an initialization vector (IV). The message part
3346 * serves as the trailing secret.
3347 */
3348 u32 secrets[COOKIE_WORKSPACE_WORDS];
3349 unsigned long expires;
3350};
3351
3352#define TCP_SECRET_1MSL (HZ * TCP_PAWS_MSL)
3353#define TCP_SECRET_2MSL (HZ * TCP_PAWS_MSL * 2)
3354#define TCP_SECRET_LIFE (HZ * 600)
3355
3356static struct tcp_cookie_secret tcp_secret_one;
3357static struct tcp_cookie_secret tcp_secret_two;
3358
3359/* Essentially a circular list, without dynamic allocation. */
3360static struct tcp_cookie_secret *tcp_secret_generating;
3361static struct tcp_cookie_secret *tcp_secret_primary;
3362static struct tcp_cookie_secret *tcp_secret_retiring;
3363static struct tcp_cookie_secret *tcp_secret_secondary;
3364
3365static DEFINE_SPINLOCK(tcp_secret_locker);
3366
3367/* Select a pseudo-random word in the cookie workspace.
3368 */
3369static inline u32 tcp_cookie_work(const u32 *ws, const int n)
3370{
3371 return ws[COOKIE_DIGEST_WORDS + ((COOKIE_MESSAGE_WORDS-1) & ws[n])];
3372}
3373
3374/* Fill bakery[COOKIE_WORKSPACE_WORDS] with generator, updating as needed.
3375 * Called in softirq context.
3376 * Returns: 0 for success.
3377 */
3378int tcp_cookie_generator(u32 *bakery)
3379{
3380 unsigned long jiffy = jiffies;
3381
3382 if (unlikely(time_after_eq(jiffy, tcp_secret_generating->expires))) {
3383 spin_lock_bh(&tcp_secret_locker);
3384 if (!time_after_eq(jiffy, tcp_secret_generating->expires)) {
3385 /* refreshed by another */
3386 memcpy(bakery,
3387 &tcp_secret_generating->secrets[0],
3388 COOKIE_WORKSPACE_WORDS);
3389 } else {
3390 /* still needs refreshing */
3391 get_random_bytes(bakery, COOKIE_WORKSPACE_WORDS);
3392
3393 /* The first time, paranoia assumes that the
3394 * randomization function isn't as strong. But,
3395 * this secret initialization is delayed until
3396 * the last possible moment (packet arrival).
3397 * Although that time is observable, it is
3398 * unpredictably variable. Mash in the most
3399 * volatile clock bits available, and expire the
3400 * secret extra quickly.
3401 */
3402 if (unlikely(tcp_secret_primary->expires ==
3403 tcp_secret_secondary->expires)) {
3404 struct timespec tv;
3405
3406 getnstimeofday(&tv);
3407 bakery[COOKIE_DIGEST_WORDS+0] ^=
3408 (u32)tv.tv_nsec;
3409
3410 tcp_secret_secondary->expires = jiffy
3411 + TCP_SECRET_1MSL
3412 + (0x0f & tcp_cookie_work(bakery, 0));
3413 } else {
3414 tcp_secret_secondary->expires = jiffy
3415 + TCP_SECRET_LIFE
3416 + (0xff & tcp_cookie_work(bakery, 1));
3417 tcp_secret_primary->expires = jiffy
3418 + TCP_SECRET_2MSL
3419 + (0x1f & tcp_cookie_work(bakery, 2));
3420 }
3421 memcpy(&tcp_secret_secondary->secrets[0],
3422 bakery, COOKIE_WORKSPACE_WORDS);
3423
3424 rcu_assign_pointer(tcp_secret_generating,
3425 tcp_secret_secondary);
3426 rcu_assign_pointer(tcp_secret_retiring,
3427 tcp_secret_primary);
3428 /*
3429 * Neither call_rcu() nor synchronize_rcu() needed.
3430 * Retiring data is not freed. It is replaced after
3431 * further (locked) pointer updates, and a quiet time
3432 * (minimum 1MSL, maximum LIFE - 2MSL).
3433 */
3434 }
3435 spin_unlock_bh(&tcp_secret_locker);
3436 } else {
3437 rcu_read_lock_bh();
3438 memcpy(bakery,
3439 &rcu_dereference(tcp_secret_generating)->secrets[0],
3440 COOKIE_WORKSPACE_WORDS);
3441 rcu_read_unlock_bh();
3442 }
3443 return 0;
3444}
3445EXPORT_SYMBOL(tcp_cookie_generator);
3446
3447void tcp_done(struct sock *sk)
3448{
3449 if (sk->sk_state == TCP_SYN_SENT || sk->sk_state == TCP_SYN_RECV)
3450 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_ATTEMPTFAILS);
3451
3452 tcp_set_state(sk, TCP_CLOSE);
3453 tcp_clear_xmit_timers(sk);
3454
3455 sk->sk_shutdown = SHUTDOWN_MASK;
3456
3457 if (!sock_flag(sk, SOCK_DEAD))
3458 sk->sk_state_change(sk);
3459 else
3460 inet_csk_destroy_sock(sk);
3461}
3462EXPORT_SYMBOL_GPL(tcp_done);
3463
3464extern struct tcp_congestion_ops tcp_reno;
3465
3466static __initdata unsigned long thash_entries;
3467static int __init set_thash_entries(char *str)
3468{
3469 ssize_t ret;
3470
3471 if (!str)
3472 return 0;
3473
3474 ret = kstrtoul(str, 0, &thash_entries);
3475 if (ret)
3476 return 0;
3477
3478 return 1;
3479}
3480__setup("thash_entries=", set_thash_entries);
3481
3482void tcp_init_mem(struct net *net)
3483{
3484 unsigned long limit = nr_free_buffer_pages() / 8;
3485 limit = max(limit, 128UL);
3486 net->ipv4.sysctl_tcp_mem[0] = limit / 4 * 3;
3487 net->ipv4.sysctl_tcp_mem[1] = limit;
3488 net->ipv4.sysctl_tcp_mem[2] = net->ipv4.sysctl_tcp_mem[0] * 2;
3489}
3490
3491void __init tcp_init(void)
3492{
3493 struct sk_buff *skb = NULL;
3494 unsigned long limit;
3495 int max_rshare, max_wshare, cnt;
3496 unsigned int i;
3497 unsigned long jiffy = jiffies;
3498
3499 BUILD_BUG_ON(sizeof(struct tcp_skb_cb) > sizeof(skb->cb));
3500
3501 percpu_counter_init(&tcp_sockets_allocated, 0);
3502 percpu_counter_init(&tcp_orphan_count, 0);
3503 tcp_hashinfo.bind_bucket_cachep =
3504 kmem_cache_create("tcp_bind_bucket",
3505 sizeof(struct inet_bind_bucket), 0,
3506 SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL);
3507
3508 /* Size and allocate the main established and bind bucket
3509 * hash tables.
3510 *
3511 * The methodology is similar to that of the buffer cache.
3512 */
3513 tcp_hashinfo.ehash =
3514 alloc_large_system_hash("TCP established",
3515 sizeof(struct inet_ehash_bucket),
3516 thash_entries,
3517 (totalram_pages >= 128 * 1024) ?
3518 13 : 15,
3519 0,
3520 NULL,
3521 &tcp_hashinfo.ehash_mask,
3522 0,
3523 thash_entries ? 0 : 512 * 1024);
3524 for (i = 0; i <= tcp_hashinfo.ehash_mask; i++) {
3525 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].chain, i);
3526 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].twchain, i);
3527 }
3528 if (inet_ehash_locks_alloc(&tcp_hashinfo))
3529 panic("TCP: failed to alloc ehash_locks");
3530 tcp_hashinfo.bhash =
3531 alloc_large_system_hash("TCP bind",
3532 sizeof(struct inet_bind_hashbucket),
3533 tcp_hashinfo.ehash_mask + 1,
3534 (totalram_pages >= 128 * 1024) ?
3535 13 : 15,
3536 0,
3537 &tcp_hashinfo.bhash_size,
3538 NULL,
3539 0,
3540 64 * 1024);
3541 tcp_hashinfo.bhash_size = 1U << tcp_hashinfo.bhash_size;
3542 for (i = 0; i < tcp_hashinfo.bhash_size; i++) {
3543 spin_lock_init(&tcp_hashinfo.bhash[i].lock);
3544 INIT_HLIST_HEAD(&tcp_hashinfo.bhash[i].chain);
3545 }
3546
3547
3548 cnt = tcp_hashinfo.ehash_mask + 1;
3549
3550 tcp_death_row.sysctl_max_tw_buckets = cnt / 2;
3551 sysctl_tcp_max_orphans = cnt / 2;
3552 sysctl_max_syn_backlog = max(128, cnt / 256);
3553
3554 tcp_init_mem(&init_net);
3555 /* Set per-socket limits to no more than 1/128 the pressure threshold */
3556 limit = nr_free_buffer_pages() << (PAGE_SHIFT - 7);
3557 max_wshare = min(4UL*1024*1024, limit);
3558 max_rshare = min(6UL*1024*1024, limit);
3559
3560 sysctl_tcp_wmem[0] = SK_MEM_QUANTUM;
3561 sysctl_tcp_wmem[1] = 16*1024;
3562 sysctl_tcp_wmem[2] = max(64*1024, max_wshare);
3563
3564 sysctl_tcp_rmem[0] = SK_MEM_QUANTUM;
3565 sysctl_tcp_rmem[1] = 87380;
3566 sysctl_tcp_rmem[2] = max(87380, max_rshare);
3567
3568 pr_info("Hash tables configured (established %u bind %u)\n",
3569 tcp_hashinfo.ehash_mask + 1, tcp_hashinfo.bhash_size);
3570
3571 tcp_register_congestion_control(&tcp_reno);
3572
3573 memset(&tcp_secret_one.secrets[0], 0, sizeof(tcp_secret_one.secrets));
3574 memset(&tcp_secret_two.secrets[0], 0, sizeof(tcp_secret_two.secrets));
3575 tcp_secret_one.expires = jiffy; /* past due */
3576 tcp_secret_two.expires = jiffy; /* past due */
3577 tcp_secret_generating = &tcp_secret_one;
3578 tcp_secret_primary = &tcp_secret_one;
3579 tcp_secret_retiring = &tcp_secret_two;
3580 tcp_secret_secondary = &tcp_secret_two;
3581}