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// SPDX-License-Identifier: GPL-2.0-or-later
2/*
3 * INET An implementation of the TCP/IP protocol suite for the LINUX
4 * operating system. INET is implemented using the BSD Socket
5 * interface as the means of communication with the user level.
6 *
7 * Implementation of the Transmission Control Protocol(TCP).
8 *
9 * Authors: Ross Biro
10 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
11 * Mark Evans, <evansmp@uhura.aston.ac.uk>
12 * Corey Minyard <wf-rch!minyard@relay.EU.net>
13 * Florian La Roche, <flla@stud.uni-sb.de>
14 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
15 * Linus Torvalds, <torvalds@cs.helsinki.fi>
16 * Alan Cox, <gw4pts@gw4pts.ampr.org>
17 * Matthew Dillon, <dillon@apollo.west.oic.com>
18 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
19 * Jorge Cwik, <jorge@laser.satlink.net>
20 *
21 * Fixes:
22 * Alan Cox : Numerous verify_area() calls
23 * Alan Cox : Set the ACK bit on a reset
24 * Alan Cox : Stopped it crashing if it closed while
25 * sk->inuse=1 and was trying to connect
26 * (tcp_err()).
27 * Alan Cox : All icmp error handling was broken
28 * pointers passed where wrong and the
29 * socket was looked up backwards. Nobody
30 * tested any icmp error code obviously.
31 * Alan Cox : tcp_err() now handled properly. It
32 * wakes people on errors. poll
33 * behaves and the icmp error race
34 * has gone by moving it into sock.c
35 * Alan Cox : tcp_send_reset() fixed to work for
36 * everything not just packets for
37 * unknown sockets.
38 * Alan Cox : tcp option processing.
39 * Alan Cox : Reset tweaked (still not 100%) [Had
40 * syn rule wrong]
41 * Herp Rosmanith : More reset fixes
42 * Alan Cox : No longer acks invalid rst frames.
43 * Acking any kind of RST is right out.
44 * Alan Cox : Sets an ignore me flag on an rst
45 * receive otherwise odd bits of prattle
46 * escape still
47 * Alan Cox : Fixed another acking RST frame bug.
48 * Should stop LAN workplace lockups.
49 * Alan Cox : Some tidyups using the new skb list
50 * facilities
51 * Alan Cox : sk->keepopen now seems to work
52 * Alan Cox : Pulls options out correctly on accepts
53 * Alan Cox : Fixed assorted sk->rqueue->next errors
54 * Alan Cox : PSH doesn't end a TCP read. Switched a
55 * bit to skb ops.
56 * Alan Cox : Tidied tcp_data to avoid a potential
57 * nasty.
58 * Alan Cox : Added some better commenting, as the
59 * tcp is hard to follow
60 * Alan Cox : Removed incorrect check for 20 * psh
61 * Michael O'Reilly : ack < copied bug fix.
62 * Johannes Stille : Misc tcp fixes (not all in yet).
63 * Alan Cox : FIN with no memory -> CRASH
64 * Alan Cox : Added socket option proto entries.
65 * Also added awareness of them to accept.
66 * Alan Cox : Added TCP options (SOL_TCP)
67 * Alan Cox : Switched wakeup calls to callbacks,
68 * so the kernel can layer network
69 * sockets.
70 * Alan Cox : Use ip_tos/ip_ttl settings.
71 * Alan Cox : Handle FIN (more) properly (we hope).
72 * Alan Cox : RST frames sent on unsynchronised
73 * state ack error.
74 * Alan Cox : Put in missing check for SYN bit.
75 * Alan Cox : Added tcp_select_window() aka NET2E
76 * window non shrink trick.
77 * Alan Cox : Added a couple of small NET2E timer
78 * fixes
79 * Charles Hedrick : TCP fixes
80 * Toomas Tamm : TCP window fixes
81 * Alan Cox : Small URG fix to rlogin ^C ack fight
82 * Charles Hedrick : Rewrote most of it to actually work
83 * Linus : Rewrote tcp_read() and URG handling
84 * completely
85 * Gerhard Koerting: Fixed some missing timer handling
86 * Matthew Dillon : Reworked TCP machine states as per RFC
87 * Gerhard Koerting: PC/TCP workarounds
88 * Adam Caldwell : Assorted timer/timing errors
89 * Matthew Dillon : Fixed another RST bug
90 * Alan Cox : Move to kernel side addressing changes.
91 * Alan Cox : Beginning work on TCP fastpathing
92 * (not yet usable)
93 * Arnt Gulbrandsen: Turbocharged tcp_check() routine.
94 * Alan Cox : TCP fast path debugging
95 * Alan Cox : Window clamping
96 * Michael Riepe : Bug in tcp_check()
97 * Matt Dillon : More TCP improvements and RST bug fixes
98 * Matt Dillon : Yet more small nasties remove from the
99 * TCP code (Be very nice to this man if
100 * tcp finally works 100%) 8)
101 * Alan Cox : BSD accept semantics.
102 * Alan Cox : Reset on closedown bug.
103 * Peter De Schrijver : ENOTCONN check missing in tcp_sendto().
104 * Michael Pall : Handle poll() after URG properly in
105 * all cases.
106 * Michael Pall : Undo the last fix in tcp_read_urg()
107 * (multi URG PUSH broke rlogin).
108 * Michael Pall : Fix the multi URG PUSH problem in
109 * tcp_readable(), poll() after URG
110 * works now.
111 * Michael Pall : recv(...,MSG_OOB) never blocks in the
112 * BSD api.
113 * Alan Cox : Changed the semantics of sk->socket to
114 * fix a race and a signal problem with
115 * accept() and async I/O.
116 * Alan Cox : Relaxed the rules on tcp_sendto().
117 * Yury Shevchuk : Really fixed accept() blocking problem.
118 * Craig I. Hagan : Allow for BSD compatible TIME_WAIT for
119 * clients/servers which listen in on
120 * fixed ports.
121 * Alan Cox : Cleaned the above up and shrank it to
122 * a sensible code size.
123 * Alan Cox : Self connect lockup fix.
124 * Alan Cox : No connect to multicast.
125 * Ross Biro : Close unaccepted children on master
126 * socket close.
127 * Alan Cox : Reset tracing code.
128 * Alan Cox : Spurious resets on shutdown.
129 * Alan Cox : Giant 15 minute/60 second timer error
130 * Alan Cox : Small whoops in polling before an
131 * accept.
132 * Alan Cox : Kept the state trace facility since
133 * it's handy for debugging.
134 * Alan Cox : More reset handler fixes.
135 * Alan Cox : Started rewriting the code based on
136 * the RFC's for other useful protocol
137 * references see: Comer, KA9Q NOS, and
138 * for a reference on the difference
139 * between specifications and how BSD
140 * works see the 4.4lite source.
141 * A.N.Kuznetsov : Don't time wait on completion of tidy
142 * close.
143 * Linus Torvalds : Fin/Shutdown & copied_seq changes.
144 * Linus Torvalds : Fixed BSD port reuse to work first syn
145 * Alan Cox : Reimplemented timers as per the RFC
146 * and using multiple timers for sanity.
147 * Alan Cox : Small bug fixes, and a lot of new
148 * comments.
149 * Alan Cox : Fixed dual reader crash by locking
150 * the buffers (much like datagram.c)
151 * Alan Cox : Fixed stuck sockets in probe. A probe
152 * now gets fed up of retrying without
153 * (even a no space) answer.
154 * Alan Cox : Extracted closing code better
155 * Alan Cox : Fixed the closing state machine to
156 * resemble the RFC.
157 * Alan Cox : More 'per spec' fixes.
158 * Jorge Cwik : Even faster checksumming.
159 * Alan Cox : tcp_data() doesn't ack illegal PSH
160 * only frames. At least one pc tcp stack
161 * generates them.
162 * Alan Cox : Cache last socket.
163 * Alan Cox : Per route irtt.
164 * Matt Day : poll()->select() match BSD precisely on error
165 * Alan Cox : New buffers
166 * Marc Tamsky : Various sk->prot->retransmits and
167 * sk->retransmits misupdating fixed.
168 * Fixed tcp_write_timeout: stuck close,
169 * and TCP syn retries gets used now.
170 * Mark Yarvis : In tcp_read_wakeup(), don't send an
171 * ack if state is TCP_CLOSED.
172 * Alan Cox : Look up device on a retransmit - routes may
173 * change. Doesn't yet cope with MSS shrink right
174 * but it's a start!
175 * Marc Tamsky : Closing in closing fixes.
176 * Mike Shaver : RFC1122 verifications.
177 * Alan Cox : rcv_saddr errors.
178 * Alan Cox : Block double connect().
179 * Alan Cox : Small hooks for enSKIP.
180 * Alexey Kuznetsov: Path MTU discovery.
181 * Alan Cox : Support soft errors.
182 * Alan Cox : Fix MTU discovery pathological case
183 * when the remote claims no mtu!
184 * Marc Tamsky : TCP_CLOSE fix.
185 * Colin (G3TNE) : Send a reset on syn ack replies in
186 * window but wrong (fixes NT lpd problems)
187 * Pedro Roque : Better TCP window handling, delayed ack.
188 * Joerg Reuter : No modification of locked buffers in
189 * tcp_do_retransmit()
190 * Eric Schenk : Changed receiver side silly window
191 * avoidance algorithm to BSD style
192 * algorithm. This doubles throughput
193 * against machines running Solaris,
194 * and seems to result in general
195 * improvement.
196 * Stefan Magdalinski : adjusted tcp_readable() to fix FIONREAD
197 * Willy Konynenberg : Transparent proxying support.
198 * Mike McLagan : Routing by source
199 * Keith Owens : Do proper merging with partial SKB's in
200 * tcp_do_sendmsg to avoid burstiness.
201 * Eric Schenk : Fix fast close down bug with
202 * shutdown() followed by close().
203 * Andi Kleen : Make poll agree with SIGIO
204 * Salvatore Sanfilippo : Support SO_LINGER with linger == 1 and
205 * lingertime == 0 (RFC 793 ABORT Call)
206 * Hirokazu Takahashi : Use copy_from_user() instead of
207 * csum_and_copy_from_user() if possible.
208 *
209 * Description of States:
210 *
211 * TCP_SYN_SENT sent a connection request, waiting for ack
212 *
213 * TCP_SYN_RECV received a connection request, sent ack,
214 * waiting for final ack in three-way handshake.
215 *
216 * TCP_ESTABLISHED connection established
217 *
218 * TCP_FIN_WAIT1 our side has shutdown, waiting to complete
219 * transmission of remaining buffered data
220 *
221 * TCP_FIN_WAIT2 all buffered data sent, waiting for remote
222 * to shutdown
223 *
224 * TCP_CLOSING both sides have shutdown but we still have
225 * data we have to finish sending
226 *
227 * TCP_TIME_WAIT timeout to catch resent junk before entering
228 * closed, can only be entered from FIN_WAIT2
229 * or CLOSING. Required because the other end
230 * may not have gotten our last ACK causing it
231 * to retransmit the data packet (which we ignore)
232 *
233 * TCP_CLOSE_WAIT remote side has shutdown and is waiting for
234 * us to finish writing our data and to shutdown
235 * (we have to close() to move on to LAST_ACK)
236 *
237 * TCP_LAST_ACK out side has shutdown after remote has
238 * shutdown. There may still be data in our
239 * buffer that we have to finish sending
240 *
241 * TCP_CLOSE socket is finished
242 */
243
244#define pr_fmt(fmt) "TCP: " fmt
245
246#include <crypto/hash.h>
247#include <linux/kernel.h>
248#include <linux/module.h>
249#include <linux/types.h>
250#include <linux/fcntl.h>
251#include <linux/poll.h>
252#include <linux/inet_diag.h>
253#include <linux/init.h>
254#include <linux/fs.h>
255#include <linux/skbuff.h>
256#include <linux/scatterlist.h>
257#include <linux/splice.h>
258#include <linux/net.h>
259#include <linux/socket.h>
260#include <linux/random.h>
261#include <linux/memblock.h>
262#include <linux/highmem.h>
263#include <linux/swap.h>
264#include <linux/cache.h>
265#include <linux/err.h>
266#include <linux/time.h>
267#include <linux/slab.h>
268#include <linux/errqueue.h>
269#include <linux/static_key.h>
270
271#include <net/icmp.h>
272#include <net/inet_common.h>
273#include <net/tcp.h>
274#include <net/mptcp.h>
275#include <net/xfrm.h>
276#include <net/ip.h>
277#include <net/sock.h>
278
279#include <linux/uaccess.h>
280#include <asm/ioctls.h>
281#include <net/busy_poll.h>
282
283struct percpu_counter tcp_orphan_count;
284EXPORT_SYMBOL_GPL(tcp_orphan_count);
285
286long sysctl_tcp_mem[3] __read_mostly;
287EXPORT_SYMBOL(sysctl_tcp_mem);
288
289atomic_long_t tcp_memory_allocated; /* Current allocated memory. */
290EXPORT_SYMBOL(tcp_memory_allocated);
291
292#if IS_ENABLED(CONFIG_SMC)
293DEFINE_STATIC_KEY_FALSE(tcp_have_smc);
294EXPORT_SYMBOL(tcp_have_smc);
295#endif
296
297/*
298 * Current number of TCP sockets.
299 */
300struct percpu_counter tcp_sockets_allocated;
301EXPORT_SYMBOL(tcp_sockets_allocated);
302
303/*
304 * TCP splice context
305 */
306struct tcp_splice_state {
307 struct pipe_inode_info *pipe;
308 size_t len;
309 unsigned int flags;
310};
311
312/*
313 * Pressure flag: try to collapse.
314 * Technical note: it is used by multiple contexts non atomically.
315 * All the __sk_mem_schedule() is of this nature: accounting
316 * is strict, actions are advisory and have some latency.
317 */
318unsigned long tcp_memory_pressure __read_mostly;
319EXPORT_SYMBOL_GPL(tcp_memory_pressure);
320
321DEFINE_STATIC_KEY_FALSE(tcp_rx_skb_cache_key);
322EXPORT_SYMBOL(tcp_rx_skb_cache_key);
323
324DEFINE_STATIC_KEY_FALSE(tcp_tx_skb_cache_key);
325
326void tcp_enter_memory_pressure(struct sock *sk)
327{
328 unsigned long val;
329
330 if (READ_ONCE(tcp_memory_pressure))
331 return;
332 val = jiffies;
333
334 if (!val)
335 val--;
336 if (!cmpxchg(&tcp_memory_pressure, 0, val))
337 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMEMORYPRESSURES);
338}
339EXPORT_SYMBOL_GPL(tcp_enter_memory_pressure);
340
341void tcp_leave_memory_pressure(struct sock *sk)
342{
343 unsigned long val;
344
345 if (!READ_ONCE(tcp_memory_pressure))
346 return;
347 val = xchg(&tcp_memory_pressure, 0);
348 if (val)
349 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPMEMORYPRESSURESCHRONO,
350 jiffies_to_msecs(jiffies - val));
351}
352EXPORT_SYMBOL_GPL(tcp_leave_memory_pressure);
353
354/* Convert seconds to retransmits based on initial and max timeout */
355static u8 secs_to_retrans(int seconds, int timeout, int rto_max)
356{
357 u8 res = 0;
358
359 if (seconds > 0) {
360 int period = timeout;
361
362 res = 1;
363 while (seconds > period && res < 255) {
364 res++;
365 timeout <<= 1;
366 if (timeout > rto_max)
367 timeout = rto_max;
368 period += timeout;
369 }
370 }
371 return res;
372}
373
374/* Convert retransmits to seconds based on initial and max timeout */
375static int retrans_to_secs(u8 retrans, int timeout, int rto_max)
376{
377 int period = 0;
378
379 if (retrans > 0) {
380 period = timeout;
381 while (--retrans) {
382 timeout <<= 1;
383 if (timeout > rto_max)
384 timeout = rto_max;
385 period += timeout;
386 }
387 }
388 return period;
389}
390
391static u64 tcp_compute_delivery_rate(const struct tcp_sock *tp)
392{
393 u32 rate = READ_ONCE(tp->rate_delivered);
394 u32 intv = READ_ONCE(tp->rate_interval_us);
395 u64 rate64 = 0;
396
397 if (rate && intv) {
398 rate64 = (u64)rate * tp->mss_cache * USEC_PER_SEC;
399 do_div(rate64, intv);
400 }
401 return rate64;
402}
403
404/* Address-family independent initialization for a tcp_sock.
405 *
406 * NOTE: A lot of things set to zero explicitly by call to
407 * sk_alloc() so need not be done here.
408 */
409void tcp_init_sock(struct sock *sk)
410{
411 struct inet_connection_sock *icsk = inet_csk(sk);
412 struct tcp_sock *tp = tcp_sk(sk);
413
414 tp->out_of_order_queue = RB_ROOT;
415 sk->tcp_rtx_queue = RB_ROOT;
416 tcp_init_xmit_timers(sk);
417 INIT_LIST_HEAD(&tp->tsq_node);
418 INIT_LIST_HEAD(&tp->tsorted_sent_queue);
419
420 icsk->icsk_rto = TCP_TIMEOUT_INIT;
421 tp->mdev_us = jiffies_to_usecs(TCP_TIMEOUT_INIT);
422 minmax_reset(&tp->rtt_min, tcp_jiffies32, ~0U);
423
424 /* So many TCP implementations out there (incorrectly) count the
425 * initial SYN frame in their delayed-ACK and congestion control
426 * algorithms that we must have the following bandaid to talk
427 * efficiently to them. -DaveM
428 */
429 tp->snd_cwnd = TCP_INIT_CWND;
430
431 /* There's a bubble in the pipe until at least the first ACK. */
432 tp->app_limited = ~0U;
433
434 /* See draft-stevens-tcpca-spec-01 for discussion of the
435 * initialization of these values.
436 */
437 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
438 tp->snd_cwnd_clamp = ~0;
439 tp->mss_cache = TCP_MSS_DEFAULT;
440
441 tp->reordering = sock_net(sk)->ipv4.sysctl_tcp_reordering;
442 tcp_assign_congestion_control(sk);
443
444 tp->tsoffset = 0;
445 tp->rack.reo_wnd_steps = 1;
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 WRITE_ONCE(sk->sk_sndbuf, sock_net(sk)->ipv4.sysctl_tcp_wmem[1]);
453 WRITE_ONCE(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
460static void tcp_tx_timestamp(struct sock *sk, u16 tsflags)
461{
462 struct sk_buff *skb = tcp_write_queue_tail(sk);
463
464 if (tsflags && skb) {
465 struct skb_shared_info *shinfo = skb_shinfo(skb);
466 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
467
468 sock_tx_timestamp(sk, tsflags, &shinfo->tx_flags);
469 if (tsflags & SOF_TIMESTAMPING_TX_ACK)
470 tcb->txstamp_ack = 1;
471 if (tsflags & SOF_TIMESTAMPING_TX_RECORD_MASK)
472 shinfo->tskey = TCP_SKB_CB(skb)->seq + skb->len - 1;
473 }
474}
475
476static inline bool tcp_stream_is_readable(const struct tcp_sock *tp,
477 int target, struct sock *sk)
478{
479 int avail = READ_ONCE(tp->rcv_nxt) - READ_ONCE(tp->copied_seq);
480
481 if (avail > 0) {
482 if (avail >= target)
483 return true;
484 if (tcp_rmem_pressure(sk))
485 return true;
486 }
487 if (sk->sk_prot->stream_memory_read)
488 return sk->sk_prot->stream_memory_read(sk);
489 return false;
490}
491
492/*
493 * Wait for a TCP event.
494 *
495 * Note that we don't need to lock the socket, as the upper poll layers
496 * take care of normal races (between the test and the event) and we don't
497 * go look at any of the socket buffers directly.
498 */
499__poll_t tcp_poll(struct file *file, struct socket *sock, poll_table *wait)
500{
501 __poll_t mask;
502 struct sock *sk = sock->sk;
503 const struct tcp_sock *tp = tcp_sk(sk);
504 int state;
505
506 sock_poll_wait(file, sock, wait);
507
508 state = inet_sk_state_load(sk);
509 if (state == TCP_LISTEN)
510 return inet_csk_listen_poll(sk);
511
512 /* Socket is not locked. We are protected from async events
513 * by poll logic and correct handling of state changes
514 * made by other threads is impossible in any case.
515 */
516
517 mask = 0;
518
519 /*
520 * EPOLLHUP is certainly not done right. But poll() doesn't
521 * have a notion of HUP in just one direction, and for a
522 * socket the read side is more interesting.
523 *
524 * Some poll() documentation says that EPOLLHUP is incompatible
525 * with the EPOLLOUT/POLLWR flags, so somebody should check this
526 * all. But careful, it tends to be safer to return too many
527 * bits than too few, and you can easily break real applications
528 * if you don't tell them that something has hung up!
529 *
530 * Check-me.
531 *
532 * Check number 1. EPOLLHUP is _UNMASKABLE_ event (see UNIX98 and
533 * our fs/select.c). It means that after we received EOF,
534 * poll always returns immediately, making impossible poll() on write()
535 * in state CLOSE_WAIT. One solution is evident --- to set EPOLLHUP
536 * if and only if shutdown has been made in both directions.
537 * Actually, it is interesting to look how Solaris and DUX
538 * solve this dilemma. I would prefer, if EPOLLHUP were maskable,
539 * then we could set it on SND_SHUTDOWN. BTW examples given
540 * in Stevens' books assume exactly this behaviour, it explains
541 * why EPOLLHUP is incompatible with EPOLLOUT. --ANK
542 *
543 * NOTE. Check for TCP_CLOSE is added. The goal is to prevent
544 * blocking on fresh not-connected or disconnected socket. --ANK
545 */
546 if (sk->sk_shutdown == SHUTDOWN_MASK || state == TCP_CLOSE)
547 mask |= EPOLLHUP;
548 if (sk->sk_shutdown & RCV_SHUTDOWN)
549 mask |= EPOLLIN | EPOLLRDNORM | EPOLLRDHUP;
550
551 /* Connected or passive Fast Open socket? */
552 if (state != TCP_SYN_SENT &&
553 (state != TCP_SYN_RECV || rcu_access_pointer(tp->fastopen_rsk))) {
554 int target = sock_rcvlowat(sk, 0, INT_MAX);
555
556 if (READ_ONCE(tp->urg_seq) == READ_ONCE(tp->copied_seq) &&
557 !sock_flag(sk, SOCK_URGINLINE) &&
558 tp->urg_data)
559 target++;
560
561 if (tcp_stream_is_readable(tp, target, sk))
562 mask |= EPOLLIN | EPOLLRDNORM;
563
564 if (!(sk->sk_shutdown & SEND_SHUTDOWN)) {
565 if (sk_stream_is_writeable(sk)) {
566 mask |= EPOLLOUT | EPOLLWRNORM;
567 } else { /* send SIGIO later */
568 sk_set_bit(SOCKWQ_ASYNC_NOSPACE, sk);
569 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
570
571 /* Race breaker. If space is freed after
572 * wspace test but before the flags are set,
573 * IO signal will be lost. Memory barrier
574 * pairs with the input side.
575 */
576 smp_mb__after_atomic();
577 if (sk_stream_is_writeable(sk))
578 mask |= EPOLLOUT | EPOLLWRNORM;
579 }
580 } else
581 mask |= EPOLLOUT | EPOLLWRNORM;
582
583 if (tp->urg_data & TCP_URG_VALID)
584 mask |= EPOLLPRI;
585 } else if (state == TCP_SYN_SENT && inet_sk(sk)->defer_connect) {
586 /* Active TCP fastopen socket with defer_connect
587 * Return EPOLLOUT so application can call write()
588 * in order for kernel to generate SYN+data
589 */
590 mask |= EPOLLOUT | EPOLLWRNORM;
591 }
592 /* This barrier is coupled with smp_wmb() in tcp_reset() */
593 smp_rmb();
594 if (sk->sk_err || !skb_queue_empty_lockless(&sk->sk_error_queue))
595 mask |= EPOLLERR;
596
597 return mask;
598}
599EXPORT_SYMBOL(tcp_poll);
600
601int tcp_ioctl(struct sock *sk, int cmd, unsigned long arg)
602{
603 struct tcp_sock *tp = tcp_sk(sk);
604 int answ;
605 bool slow;
606
607 switch (cmd) {
608 case SIOCINQ:
609 if (sk->sk_state == TCP_LISTEN)
610 return -EINVAL;
611
612 slow = lock_sock_fast(sk);
613 answ = tcp_inq(sk);
614 unlock_sock_fast(sk, slow);
615 break;
616 case SIOCATMARK:
617 answ = tp->urg_data &&
618 READ_ONCE(tp->urg_seq) == READ_ONCE(tp->copied_seq);
619 break;
620 case SIOCOUTQ:
621 if (sk->sk_state == TCP_LISTEN)
622 return -EINVAL;
623
624 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
625 answ = 0;
626 else
627 answ = READ_ONCE(tp->write_seq) - tp->snd_una;
628 break;
629 case SIOCOUTQNSD:
630 if (sk->sk_state == TCP_LISTEN)
631 return -EINVAL;
632
633 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
634 answ = 0;
635 else
636 answ = READ_ONCE(tp->write_seq) -
637 READ_ONCE(tp->snd_nxt);
638 break;
639 default:
640 return -ENOIOCTLCMD;
641 }
642
643 return put_user(answ, (int __user *)arg);
644}
645EXPORT_SYMBOL(tcp_ioctl);
646
647static inline void tcp_mark_push(struct tcp_sock *tp, struct sk_buff *skb)
648{
649 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
650 tp->pushed_seq = tp->write_seq;
651}
652
653static inline bool forced_push(const struct tcp_sock *tp)
654{
655 return after(tp->write_seq, tp->pushed_seq + (tp->max_window >> 1));
656}
657
658static void skb_entail(struct sock *sk, struct sk_buff *skb)
659{
660 struct tcp_sock *tp = tcp_sk(sk);
661 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
662
663 skb->csum = 0;
664 tcb->seq = tcb->end_seq = tp->write_seq;
665 tcb->tcp_flags = TCPHDR_ACK;
666 tcb->sacked = 0;
667 __skb_header_release(skb);
668 tcp_add_write_queue_tail(sk, skb);
669 sk_wmem_queued_add(sk, skb->truesize);
670 sk_mem_charge(sk, skb->truesize);
671 if (tp->nonagle & TCP_NAGLE_PUSH)
672 tp->nonagle &= ~TCP_NAGLE_PUSH;
673
674 tcp_slow_start_after_idle_check(sk);
675}
676
677static inline void tcp_mark_urg(struct tcp_sock *tp, int flags)
678{
679 if (flags & MSG_OOB)
680 tp->snd_up = tp->write_seq;
681}
682
683/* If a not yet filled skb is pushed, do not send it if
684 * we have data packets in Qdisc or NIC queues :
685 * Because TX completion will happen shortly, it gives a chance
686 * to coalesce future sendmsg() payload into this skb, without
687 * need for a timer, and with no latency trade off.
688 * As packets containing data payload have a bigger truesize
689 * than pure acks (dataless) packets, the last checks prevent
690 * autocorking if we only have an ACK in Qdisc/NIC queues,
691 * or if TX completion was delayed after we processed ACK packet.
692 */
693static bool tcp_should_autocork(struct sock *sk, struct sk_buff *skb,
694 int size_goal)
695{
696 return skb->len < size_goal &&
697 sock_net(sk)->ipv4.sysctl_tcp_autocorking &&
698 !tcp_rtx_queue_empty(sk) &&
699 refcount_read(&sk->sk_wmem_alloc) > skb->truesize;
700}
701
702void tcp_push(struct sock *sk, int flags, int mss_now,
703 int nonagle, int size_goal)
704{
705 struct tcp_sock *tp = tcp_sk(sk);
706 struct sk_buff *skb;
707
708 skb = tcp_write_queue_tail(sk);
709 if (!skb)
710 return;
711 if (!(flags & MSG_MORE) || forced_push(tp))
712 tcp_mark_push(tp, skb);
713
714 tcp_mark_urg(tp, flags);
715
716 if (tcp_should_autocork(sk, skb, size_goal)) {
717
718 /* avoid atomic op if TSQ_THROTTLED bit is already set */
719 if (!test_bit(TSQ_THROTTLED, &sk->sk_tsq_flags)) {
720 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPAUTOCORKING);
721 set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags);
722 }
723 /* It is possible TX completion already happened
724 * before we set TSQ_THROTTLED.
725 */
726 if (refcount_read(&sk->sk_wmem_alloc) > skb->truesize)
727 return;
728 }
729
730 if (flags & MSG_MORE)
731 nonagle = TCP_NAGLE_CORK;
732
733 __tcp_push_pending_frames(sk, mss_now, nonagle);
734}
735
736static int tcp_splice_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb,
737 unsigned int offset, size_t len)
738{
739 struct tcp_splice_state *tss = rd_desc->arg.data;
740 int ret;
741
742 ret = skb_splice_bits(skb, skb->sk, offset, tss->pipe,
743 min(rd_desc->count, len), tss->flags);
744 if (ret > 0)
745 rd_desc->count -= ret;
746 return ret;
747}
748
749static int __tcp_splice_read(struct sock *sk, struct tcp_splice_state *tss)
750{
751 /* Store TCP splice context information in read_descriptor_t. */
752 read_descriptor_t rd_desc = {
753 .arg.data = tss,
754 .count = tss->len,
755 };
756
757 return tcp_read_sock(sk, &rd_desc, tcp_splice_data_recv);
758}
759
760/**
761 * tcp_splice_read - splice data from TCP socket to a pipe
762 * @sock: socket to splice from
763 * @ppos: position (not valid)
764 * @pipe: pipe to splice to
765 * @len: number of bytes to splice
766 * @flags: splice modifier flags
767 *
768 * Description:
769 * Will read pages from given socket and fill them into a pipe.
770 *
771 **/
772ssize_t tcp_splice_read(struct socket *sock, loff_t *ppos,
773 struct pipe_inode_info *pipe, size_t len,
774 unsigned int flags)
775{
776 struct sock *sk = sock->sk;
777 struct tcp_splice_state tss = {
778 .pipe = pipe,
779 .len = len,
780 .flags = flags,
781 };
782 long timeo;
783 ssize_t spliced;
784 int ret;
785
786 sock_rps_record_flow(sk);
787 /*
788 * We can't seek on a socket input
789 */
790 if (unlikely(*ppos))
791 return -ESPIPE;
792
793 ret = spliced = 0;
794
795 lock_sock(sk);
796
797 timeo = sock_rcvtimeo(sk, sock->file->f_flags & O_NONBLOCK);
798 while (tss.len) {
799 ret = __tcp_splice_read(sk, &tss);
800 if (ret < 0)
801 break;
802 else if (!ret) {
803 if (spliced)
804 break;
805 if (sock_flag(sk, SOCK_DONE))
806 break;
807 if (sk->sk_err) {
808 ret = sock_error(sk);
809 break;
810 }
811 if (sk->sk_shutdown & RCV_SHUTDOWN)
812 break;
813 if (sk->sk_state == TCP_CLOSE) {
814 /*
815 * This occurs when user tries to read
816 * from never connected socket.
817 */
818 ret = -ENOTCONN;
819 break;
820 }
821 if (!timeo) {
822 ret = -EAGAIN;
823 break;
824 }
825 /* if __tcp_splice_read() got nothing while we have
826 * an skb in receive queue, we do not want to loop.
827 * This might happen with URG data.
828 */
829 if (!skb_queue_empty(&sk->sk_receive_queue))
830 break;
831 sk_wait_data(sk, &timeo, NULL);
832 if (signal_pending(current)) {
833 ret = sock_intr_errno(timeo);
834 break;
835 }
836 continue;
837 }
838 tss.len -= ret;
839 spliced += ret;
840
841 if (!timeo)
842 break;
843 release_sock(sk);
844 lock_sock(sk);
845
846 if (sk->sk_err || sk->sk_state == TCP_CLOSE ||
847 (sk->sk_shutdown & RCV_SHUTDOWN) ||
848 signal_pending(current))
849 break;
850 }
851
852 release_sock(sk);
853
854 if (spliced)
855 return spliced;
856
857 return ret;
858}
859EXPORT_SYMBOL(tcp_splice_read);
860
861struct sk_buff *sk_stream_alloc_skb(struct sock *sk, int size, gfp_t gfp,
862 bool force_schedule)
863{
864 struct sk_buff *skb;
865
866 if (likely(!size)) {
867 skb = sk->sk_tx_skb_cache;
868 if (skb) {
869 skb->truesize = SKB_TRUESIZE(skb_end_offset(skb));
870 sk->sk_tx_skb_cache = NULL;
871 pskb_trim(skb, 0);
872 INIT_LIST_HEAD(&skb->tcp_tsorted_anchor);
873 skb_shinfo(skb)->tx_flags = 0;
874 memset(TCP_SKB_CB(skb), 0, sizeof(struct tcp_skb_cb));
875 return skb;
876 }
877 }
878 /* The TCP header must be at least 32-bit aligned. */
879 size = ALIGN(size, 4);
880
881 if (unlikely(tcp_under_memory_pressure(sk)))
882 sk_mem_reclaim_partial(sk);
883
884 skb = alloc_skb_fclone(size + sk->sk_prot->max_header, gfp);
885 if (likely(skb)) {
886 bool mem_scheduled;
887
888 if (force_schedule) {
889 mem_scheduled = true;
890 sk_forced_mem_schedule(sk, skb->truesize);
891 } else {
892 mem_scheduled = sk_wmem_schedule(sk, skb->truesize);
893 }
894 if (likely(mem_scheduled)) {
895 skb_reserve(skb, sk->sk_prot->max_header);
896 /*
897 * Make sure that we have exactly size bytes
898 * available to the caller, no more, no less.
899 */
900 skb->reserved_tailroom = skb->end - skb->tail - size;
901 INIT_LIST_HEAD(&skb->tcp_tsorted_anchor);
902 return skb;
903 }
904 __kfree_skb(skb);
905 } else {
906 sk->sk_prot->enter_memory_pressure(sk);
907 sk_stream_moderate_sndbuf(sk);
908 }
909 return NULL;
910}
911
912static unsigned int tcp_xmit_size_goal(struct sock *sk, u32 mss_now,
913 int large_allowed)
914{
915 struct tcp_sock *tp = tcp_sk(sk);
916 u32 new_size_goal, size_goal;
917
918 if (!large_allowed)
919 return mss_now;
920
921 /* Note : tcp_tso_autosize() will eventually split this later */
922 new_size_goal = sk->sk_gso_max_size - 1 - MAX_TCP_HEADER;
923 new_size_goal = tcp_bound_to_half_wnd(tp, new_size_goal);
924
925 /* We try hard to avoid divides here */
926 size_goal = tp->gso_segs * mss_now;
927 if (unlikely(new_size_goal < size_goal ||
928 new_size_goal >= size_goal + mss_now)) {
929 tp->gso_segs = min_t(u16, new_size_goal / mss_now,
930 sk->sk_gso_max_segs);
931 size_goal = tp->gso_segs * mss_now;
932 }
933
934 return max(size_goal, mss_now);
935}
936
937int tcp_send_mss(struct sock *sk, int *size_goal, int flags)
938{
939 int mss_now;
940
941 mss_now = tcp_current_mss(sk);
942 *size_goal = tcp_xmit_size_goal(sk, mss_now, !(flags & MSG_OOB));
943
944 return mss_now;
945}
946
947/* In some cases, both sendpage() and sendmsg() could have added
948 * an skb to the write queue, but failed adding payload on it.
949 * We need to remove it to consume less memory, but more
950 * importantly be able to generate EPOLLOUT for Edge Trigger epoll()
951 * users.
952 */
953static void tcp_remove_empty_skb(struct sock *sk, struct sk_buff *skb)
954{
955 if (skb && !skb->len) {
956 tcp_unlink_write_queue(skb, sk);
957 if (tcp_write_queue_empty(sk))
958 tcp_chrono_stop(sk, TCP_CHRONO_BUSY);
959 sk_wmem_free_skb(sk, skb);
960 }
961}
962
963ssize_t do_tcp_sendpages(struct sock *sk, struct page *page, int offset,
964 size_t size, int flags)
965{
966 struct tcp_sock *tp = tcp_sk(sk);
967 int mss_now, size_goal;
968 int err;
969 ssize_t copied;
970 long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
971
972 if (IS_ENABLED(CONFIG_DEBUG_VM) &&
973 WARN_ONCE(!sendpage_ok(page),
974 "page must not be a Slab one and have page_count > 0"))
975 return -EINVAL;
976
977 /* Wait for a connection to finish. One exception is TCP Fast Open
978 * (passive side) where data is allowed to be sent before a connection
979 * is fully established.
980 */
981 if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) &&
982 !tcp_passive_fastopen(sk)) {
983 err = sk_stream_wait_connect(sk, &timeo);
984 if (err != 0)
985 goto out_err;
986 }
987
988 sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
989
990 mss_now = tcp_send_mss(sk, &size_goal, flags);
991 copied = 0;
992
993 err = -EPIPE;
994 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
995 goto out_err;
996
997 while (size > 0) {
998 struct sk_buff *skb = tcp_write_queue_tail(sk);
999 int copy, i;
1000 bool can_coalesce;
1001
1002 if (!skb || (copy = size_goal - skb->len) <= 0 ||
1003 !tcp_skb_can_collapse_to(skb)) {
1004new_segment:
1005 if (!sk_stream_memory_free(sk))
1006 goto wait_for_sndbuf;
1007
1008 skb = sk_stream_alloc_skb(sk, 0, sk->sk_allocation,
1009 tcp_rtx_and_write_queues_empty(sk));
1010 if (!skb)
1011 goto wait_for_memory;
1012
1013#ifdef CONFIG_TLS_DEVICE
1014 skb->decrypted = !!(flags & MSG_SENDPAGE_DECRYPTED);
1015#endif
1016 skb_entail(sk, skb);
1017 copy = size_goal;
1018 }
1019
1020 if (copy > size)
1021 copy = size;
1022
1023 i = skb_shinfo(skb)->nr_frags;
1024 can_coalesce = skb_can_coalesce(skb, i, page, offset);
1025 if (!can_coalesce && i >= sysctl_max_skb_frags) {
1026 tcp_mark_push(tp, skb);
1027 goto new_segment;
1028 }
1029 if (!sk_wmem_schedule(sk, copy))
1030 goto wait_for_memory;
1031
1032 if (can_coalesce) {
1033 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1034 } else {
1035 get_page(page);
1036 skb_fill_page_desc(skb, i, page, offset, copy);
1037 }
1038
1039 if (!(flags & MSG_NO_SHARED_FRAGS))
1040 skb_shinfo(skb)->tx_flags |= SKBTX_SHARED_FRAG;
1041
1042 skb->len += copy;
1043 skb->data_len += copy;
1044 skb->truesize += copy;
1045 sk_wmem_queued_add(sk, copy);
1046 sk_mem_charge(sk, copy);
1047 skb->ip_summed = CHECKSUM_PARTIAL;
1048 WRITE_ONCE(tp->write_seq, tp->write_seq + copy);
1049 TCP_SKB_CB(skb)->end_seq += copy;
1050 tcp_skb_pcount_set(skb, 0);
1051
1052 if (!copied)
1053 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
1054
1055 copied += copy;
1056 offset += copy;
1057 size -= copy;
1058 if (!size)
1059 goto out;
1060
1061 if (skb->len < size_goal || (flags & MSG_OOB))
1062 continue;
1063
1064 if (forced_push(tp)) {
1065 tcp_mark_push(tp, skb);
1066 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
1067 } else if (skb == tcp_send_head(sk))
1068 tcp_push_one(sk, mss_now);
1069 continue;
1070
1071wait_for_sndbuf:
1072 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1073wait_for_memory:
1074 tcp_push(sk, flags & ~MSG_MORE, mss_now,
1075 TCP_NAGLE_PUSH, size_goal);
1076
1077 err = sk_stream_wait_memory(sk, &timeo);
1078 if (err != 0)
1079 goto do_error;
1080
1081 mss_now = tcp_send_mss(sk, &size_goal, flags);
1082 }
1083
1084out:
1085 if (copied) {
1086 tcp_tx_timestamp(sk, sk->sk_tsflags);
1087 if (!(flags & MSG_SENDPAGE_NOTLAST))
1088 tcp_push(sk, flags, mss_now, tp->nonagle, size_goal);
1089 }
1090 return copied;
1091
1092do_error:
1093 tcp_remove_empty_skb(sk, tcp_write_queue_tail(sk));
1094 if (copied)
1095 goto out;
1096out_err:
1097 /* make sure we wake any epoll edge trigger waiter */
1098 if (unlikely(tcp_rtx_and_write_queues_empty(sk) && err == -EAGAIN)) {
1099 sk->sk_write_space(sk);
1100 tcp_chrono_stop(sk, TCP_CHRONO_SNDBUF_LIMITED);
1101 }
1102 return sk_stream_error(sk, flags, err);
1103}
1104EXPORT_SYMBOL_GPL(do_tcp_sendpages);
1105
1106int tcp_sendpage_locked(struct sock *sk, struct page *page, int offset,
1107 size_t size, int flags)
1108{
1109 if (!(sk->sk_route_caps & NETIF_F_SG))
1110 return sock_no_sendpage_locked(sk, page, offset, size, flags);
1111
1112 tcp_rate_check_app_limited(sk); /* is sending application-limited? */
1113
1114 return do_tcp_sendpages(sk, page, offset, size, flags);
1115}
1116EXPORT_SYMBOL_GPL(tcp_sendpage_locked);
1117
1118int tcp_sendpage(struct sock *sk, struct page *page, int offset,
1119 size_t size, int flags)
1120{
1121 int ret;
1122
1123 lock_sock(sk);
1124 ret = tcp_sendpage_locked(sk, page, offset, size, flags);
1125 release_sock(sk);
1126
1127 return ret;
1128}
1129EXPORT_SYMBOL(tcp_sendpage);
1130
1131void tcp_free_fastopen_req(struct tcp_sock *tp)
1132{
1133 if (tp->fastopen_req) {
1134 kfree(tp->fastopen_req);
1135 tp->fastopen_req = NULL;
1136 }
1137}
1138
1139static int tcp_sendmsg_fastopen(struct sock *sk, struct msghdr *msg,
1140 int *copied, size_t size,
1141 struct ubuf_info *uarg)
1142{
1143 struct tcp_sock *tp = tcp_sk(sk);
1144 struct inet_sock *inet = inet_sk(sk);
1145 struct sockaddr *uaddr = msg->msg_name;
1146 int err, flags;
1147
1148 if (!(sock_net(sk)->ipv4.sysctl_tcp_fastopen & TFO_CLIENT_ENABLE) ||
1149 (uaddr && msg->msg_namelen >= sizeof(uaddr->sa_family) &&
1150 uaddr->sa_family == AF_UNSPEC))
1151 return -EOPNOTSUPP;
1152 if (tp->fastopen_req)
1153 return -EALREADY; /* Another Fast Open is in progress */
1154
1155 tp->fastopen_req = kzalloc(sizeof(struct tcp_fastopen_request),
1156 sk->sk_allocation);
1157 if (unlikely(!tp->fastopen_req))
1158 return -ENOBUFS;
1159 tp->fastopen_req->data = msg;
1160 tp->fastopen_req->size = size;
1161 tp->fastopen_req->uarg = uarg;
1162
1163 if (inet->defer_connect) {
1164 err = tcp_connect(sk);
1165 /* Same failure procedure as in tcp_v4/6_connect */
1166 if (err) {
1167 tcp_set_state(sk, TCP_CLOSE);
1168 inet->inet_dport = 0;
1169 sk->sk_route_caps = 0;
1170 }
1171 }
1172 flags = (msg->msg_flags & MSG_DONTWAIT) ? O_NONBLOCK : 0;
1173 err = __inet_stream_connect(sk->sk_socket, uaddr,
1174 msg->msg_namelen, flags, 1);
1175 /* fastopen_req could already be freed in __inet_stream_connect
1176 * if the connection times out or gets rst
1177 */
1178 if (tp->fastopen_req) {
1179 *copied = tp->fastopen_req->copied;
1180 tcp_free_fastopen_req(tp);
1181 inet->defer_connect = 0;
1182 }
1183 return err;
1184}
1185
1186int tcp_sendmsg_locked(struct sock *sk, struct msghdr *msg, size_t size)
1187{
1188 struct tcp_sock *tp = tcp_sk(sk);
1189 struct ubuf_info *uarg = NULL;
1190 struct sk_buff *skb;
1191 struct sockcm_cookie sockc;
1192 int flags, err, copied = 0;
1193 int mss_now = 0, size_goal, copied_syn = 0;
1194 int process_backlog = 0;
1195 bool zc = false;
1196 long timeo;
1197
1198 flags = msg->msg_flags;
1199
1200 if (flags & MSG_ZEROCOPY && size && sock_flag(sk, SOCK_ZEROCOPY)) {
1201 skb = tcp_write_queue_tail(sk);
1202 uarg = sock_zerocopy_realloc(sk, size, skb_zcopy(skb));
1203 if (!uarg) {
1204 err = -ENOBUFS;
1205 goto out_err;
1206 }
1207
1208 zc = sk->sk_route_caps & NETIF_F_SG;
1209 if (!zc)
1210 uarg->zerocopy = 0;
1211 }
1212
1213 if (unlikely(flags & MSG_FASTOPEN || inet_sk(sk)->defer_connect) &&
1214 !tp->repair) {
1215 err = tcp_sendmsg_fastopen(sk, msg, &copied_syn, size, uarg);
1216 if (err == -EINPROGRESS && copied_syn > 0)
1217 goto out;
1218 else if (err)
1219 goto out_err;
1220 }
1221
1222 timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
1223
1224 tcp_rate_check_app_limited(sk); /* is sending application-limited? */
1225
1226 /* Wait for a connection to finish. One exception is TCP Fast Open
1227 * (passive side) where data is allowed to be sent before a connection
1228 * is fully established.
1229 */
1230 if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) &&
1231 !tcp_passive_fastopen(sk)) {
1232 err = sk_stream_wait_connect(sk, &timeo);
1233 if (err != 0)
1234 goto do_error;
1235 }
1236
1237 if (unlikely(tp->repair)) {
1238 if (tp->repair_queue == TCP_RECV_QUEUE) {
1239 copied = tcp_send_rcvq(sk, msg, size);
1240 goto out_nopush;
1241 }
1242
1243 err = -EINVAL;
1244 if (tp->repair_queue == TCP_NO_QUEUE)
1245 goto out_err;
1246
1247 /* 'common' sending to sendq */
1248 }
1249
1250 sockcm_init(&sockc, sk);
1251 if (msg->msg_controllen) {
1252 err = sock_cmsg_send(sk, msg, &sockc);
1253 if (unlikely(err)) {
1254 err = -EINVAL;
1255 goto out_err;
1256 }
1257 }
1258
1259 /* This should be in poll */
1260 sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
1261
1262 /* Ok commence sending. */
1263 copied = 0;
1264
1265restart:
1266 mss_now = tcp_send_mss(sk, &size_goal, flags);
1267
1268 err = -EPIPE;
1269 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
1270 goto do_error;
1271
1272 while (msg_data_left(msg)) {
1273 int copy = 0;
1274
1275 skb = tcp_write_queue_tail(sk);
1276 if (skb)
1277 copy = size_goal - skb->len;
1278
1279 if (copy <= 0 || !tcp_skb_can_collapse_to(skb)) {
1280 bool first_skb;
1281
1282new_segment:
1283 if (!sk_stream_memory_free(sk))
1284 goto wait_for_sndbuf;
1285
1286 if (unlikely(process_backlog >= 16)) {
1287 process_backlog = 0;
1288 if (sk_flush_backlog(sk))
1289 goto restart;
1290 }
1291 first_skb = tcp_rtx_and_write_queues_empty(sk);
1292 skb = sk_stream_alloc_skb(sk, 0, sk->sk_allocation,
1293 first_skb);
1294 if (!skb)
1295 goto wait_for_memory;
1296
1297 process_backlog++;
1298 skb->ip_summed = CHECKSUM_PARTIAL;
1299
1300 skb_entail(sk, skb);
1301 copy = size_goal;
1302
1303 /* All packets are restored as if they have
1304 * already been sent. skb_mstamp_ns isn't set to
1305 * avoid wrong rtt estimation.
1306 */
1307 if (tp->repair)
1308 TCP_SKB_CB(skb)->sacked |= TCPCB_REPAIRED;
1309 }
1310
1311 /* Try to append data to the end of skb. */
1312 if (copy > msg_data_left(msg))
1313 copy = msg_data_left(msg);
1314
1315 /* Where to copy to? */
1316 if (skb_availroom(skb) > 0 && !zc) {
1317 /* We have some space in skb head. Superb! */
1318 copy = min_t(int, copy, skb_availroom(skb));
1319 err = skb_add_data_nocache(sk, skb, &msg->msg_iter, copy);
1320 if (err)
1321 goto do_fault;
1322 } else if (!zc) {
1323 bool merge = true;
1324 int i = skb_shinfo(skb)->nr_frags;
1325 struct page_frag *pfrag = sk_page_frag(sk);
1326
1327 if (!sk_page_frag_refill(sk, pfrag))
1328 goto wait_for_memory;
1329
1330 if (!skb_can_coalesce(skb, i, pfrag->page,
1331 pfrag->offset)) {
1332 if (i >= sysctl_max_skb_frags) {
1333 tcp_mark_push(tp, skb);
1334 goto new_segment;
1335 }
1336 merge = false;
1337 }
1338
1339 copy = min_t(int, copy, pfrag->size - pfrag->offset);
1340
1341 if (!sk_wmem_schedule(sk, copy))
1342 goto wait_for_memory;
1343
1344 err = skb_copy_to_page_nocache(sk, &msg->msg_iter, skb,
1345 pfrag->page,
1346 pfrag->offset,
1347 copy);
1348 if (err)
1349 goto do_error;
1350
1351 /* Update the skb. */
1352 if (merge) {
1353 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1354 } else {
1355 skb_fill_page_desc(skb, i, pfrag->page,
1356 pfrag->offset, copy);
1357 page_ref_inc(pfrag->page);
1358 }
1359 pfrag->offset += copy;
1360 } else {
1361 err = skb_zerocopy_iter_stream(sk, skb, msg, copy, uarg);
1362 if (err == -EMSGSIZE || err == -EEXIST) {
1363 tcp_mark_push(tp, skb);
1364 goto new_segment;
1365 }
1366 if (err < 0)
1367 goto do_error;
1368 copy = err;
1369 }
1370
1371 if (!copied)
1372 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
1373
1374 WRITE_ONCE(tp->write_seq, tp->write_seq + copy);
1375 TCP_SKB_CB(skb)->end_seq += copy;
1376 tcp_skb_pcount_set(skb, 0);
1377
1378 copied += copy;
1379 if (!msg_data_left(msg)) {
1380 if (unlikely(flags & MSG_EOR))
1381 TCP_SKB_CB(skb)->eor = 1;
1382 goto out;
1383 }
1384
1385 if (skb->len < size_goal || (flags & MSG_OOB) || unlikely(tp->repair))
1386 continue;
1387
1388 if (forced_push(tp)) {
1389 tcp_mark_push(tp, skb);
1390 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
1391 } else if (skb == tcp_send_head(sk))
1392 tcp_push_one(sk, mss_now);
1393 continue;
1394
1395wait_for_sndbuf:
1396 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1397wait_for_memory:
1398 if (copied)
1399 tcp_push(sk, flags & ~MSG_MORE, mss_now,
1400 TCP_NAGLE_PUSH, size_goal);
1401
1402 err = sk_stream_wait_memory(sk, &timeo);
1403 if (err != 0)
1404 goto do_error;
1405
1406 mss_now = tcp_send_mss(sk, &size_goal, flags);
1407 }
1408
1409out:
1410 if (copied) {
1411 tcp_tx_timestamp(sk, sockc.tsflags);
1412 tcp_push(sk, flags, mss_now, tp->nonagle, size_goal);
1413 }
1414out_nopush:
1415 sock_zerocopy_put(uarg);
1416 return copied + copied_syn;
1417
1418do_error:
1419 skb = tcp_write_queue_tail(sk);
1420do_fault:
1421 tcp_remove_empty_skb(sk, skb);
1422
1423 if (copied + copied_syn)
1424 goto out;
1425out_err:
1426 sock_zerocopy_put_abort(uarg, true);
1427 err = sk_stream_error(sk, flags, err);
1428 /* make sure we wake any epoll edge trigger waiter */
1429 if (unlikely(tcp_rtx_and_write_queues_empty(sk) && err == -EAGAIN)) {
1430 sk->sk_write_space(sk);
1431 tcp_chrono_stop(sk, TCP_CHRONO_SNDBUF_LIMITED);
1432 }
1433 return err;
1434}
1435EXPORT_SYMBOL_GPL(tcp_sendmsg_locked);
1436
1437int tcp_sendmsg(struct sock *sk, struct msghdr *msg, size_t size)
1438{
1439 int ret;
1440
1441 lock_sock(sk);
1442 ret = tcp_sendmsg_locked(sk, msg, size);
1443 release_sock(sk);
1444
1445 return ret;
1446}
1447EXPORT_SYMBOL(tcp_sendmsg);
1448
1449/*
1450 * Handle reading urgent data. BSD has very simple semantics for
1451 * this, no blocking and very strange errors 8)
1452 */
1453
1454static int tcp_recv_urg(struct sock *sk, struct msghdr *msg, int len, int flags)
1455{
1456 struct tcp_sock *tp = tcp_sk(sk);
1457
1458 /* No URG data to read. */
1459 if (sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data ||
1460 tp->urg_data == TCP_URG_READ)
1461 return -EINVAL; /* Yes this is right ! */
1462
1463 if (sk->sk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DONE))
1464 return -ENOTCONN;
1465
1466 if (tp->urg_data & TCP_URG_VALID) {
1467 int err = 0;
1468 char c = tp->urg_data;
1469
1470 if (!(flags & MSG_PEEK))
1471 tp->urg_data = TCP_URG_READ;
1472
1473 /* Read urgent data. */
1474 msg->msg_flags |= MSG_OOB;
1475
1476 if (len > 0) {
1477 if (!(flags & MSG_TRUNC))
1478 err = memcpy_to_msg(msg, &c, 1);
1479 len = 1;
1480 } else
1481 msg->msg_flags |= MSG_TRUNC;
1482
1483 return err ? -EFAULT : len;
1484 }
1485
1486 if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
1487 return 0;
1488
1489 /* Fixed the recv(..., MSG_OOB) behaviour. BSD docs and
1490 * the available implementations agree in this case:
1491 * this call should never block, independent of the
1492 * blocking state of the socket.
1493 * Mike <pall@rz.uni-karlsruhe.de>
1494 */
1495 return -EAGAIN;
1496}
1497
1498static int tcp_peek_sndq(struct sock *sk, struct msghdr *msg, int len)
1499{
1500 struct sk_buff *skb;
1501 int copied = 0, err = 0;
1502
1503 /* XXX -- need to support SO_PEEK_OFF */
1504
1505 skb_rbtree_walk(skb, &sk->tcp_rtx_queue) {
1506 err = skb_copy_datagram_msg(skb, 0, msg, skb->len);
1507 if (err)
1508 return err;
1509 copied += skb->len;
1510 }
1511
1512 skb_queue_walk(&sk->sk_write_queue, skb) {
1513 err = skb_copy_datagram_msg(skb, 0, msg, skb->len);
1514 if (err)
1515 break;
1516
1517 copied += skb->len;
1518 }
1519
1520 return err ?: copied;
1521}
1522
1523/* Clean up the receive buffer for full frames taken by the user,
1524 * then send an ACK if necessary. COPIED is the number of bytes
1525 * tcp_recvmsg has given to the user so far, it speeds up the
1526 * calculation of whether or not we must ACK for the sake of
1527 * a window update.
1528 */
1529static void tcp_cleanup_rbuf(struct sock *sk, int copied)
1530{
1531 struct tcp_sock *tp = tcp_sk(sk);
1532 bool time_to_ack = false;
1533
1534 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
1535
1536 WARN(skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq),
1537 "cleanup rbuf bug: copied %X seq %X rcvnxt %X\n",
1538 tp->copied_seq, TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt);
1539
1540 if (inet_csk_ack_scheduled(sk)) {
1541 const struct inet_connection_sock *icsk = inet_csk(sk);
1542 /* Delayed ACKs frequently hit locked sockets during bulk
1543 * receive. */
1544 if (icsk->icsk_ack.blocked ||
1545 /* Once-per-two-segments ACK was not sent by tcp_input.c */
1546 tp->rcv_nxt - tp->rcv_wup > icsk->icsk_ack.rcv_mss ||
1547 /*
1548 * If this read emptied read buffer, we send ACK, if
1549 * connection is not bidirectional, user drained
1550 * receive buffer and there was a small segment
1551 * in queue.
1552 */
1553 (copied > 0 &&
1554 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED2) ||
1555 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED) &&
1556 !inet_csk_in_pingpong_mode(sk))) &&
1557 !atomic_read(&sk->sk_rmem_alloc)))
1558 time_to_ack = true;
1559 }
1560
1561 /* We send an ACK if we can now advertise a non-zero window
1562 * which has been raised "significantly".
1563 *
1564 * Even if window raised up to infinity, do not send window open ACK
1565 * in states, where we will not receive more. It is useless.
1566 */
1567 if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) {
1568 __u32 rcv_window_now = tcp_receive_window(tp);
1569
1570 /* Optimize, __tcp_select_window() is not cheap. */
1571 if (2*rcv_window_now <= tp->window_clamp) {
1572 __u32 new_window = __tcp_select_window(sk);
1573
1574 /* Send ACK now, if this read freed lots of space
1575 * in our buffer. Certainly, new_window is new window.
1576 * We can advertise it now, if it is not less than current one.
1577 * "Lots" means "at least twice" here.
1578 */
1579 if (new_window && new_window >= 2 * rcv_window_now)
1580 time_to_ack = true;
1581 }
1582 }
1583 if (time_to_ack)
1584 tcp_send_ack(sk);
1585}
1586
1587static struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off)
1588{
1589 struct sk_buff *skb;
1590 u32 offset;
1591
1592 while ((skb = skb_peek(&sk->sk_receive_queue)) != NULL) {
1593 offset = seq - TCP_SKB_CB(skb)->seq;
1594 if (unlikely(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
1595 pr_err_once("%s: found a SYN, please report !\n", __func__);
1596 offset--;
1597 }
1598 if (offset < skb->len || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)) {
1599 *off = offset;
1600 return skb;
1601 }
1602 /* This looks weird, but this can happen if TCP collapsing
1603 * splitted a fat GRO packet, while we released socket lock
1604 * in skb_splice_bits()
1605 */
1606 sk_eat_skb(sk, skb);
1607 }
1608 return NULL;
1609}
1610
1611/*
1612 * This routine provides an alternative to tcp_recvmsg() for routines
1613 * that would like to handle copying from skbuffs directly in 'sendfile'
1614 * fashion.
1615 * Note:
1616 * - It is assumed that the socket was locked by the caller.
1617 * - The routine does not block.
1618 * - At present, there is no support for reading OOB data
1619 * or for 'peeking' the socket using this routine
1620 * (although both would be easy to implement).
1621 */
1622int tcp_read_sock(struct sock *sk, read_descriptor_t *desc,
1623 sk_read_actor_t recv_actor)
1624{
1625 struct sk_buff *skb;
1626 struct tcp_sock *tp = tcp_sk(sk);
1627 u32 seq = tp->copied_seq;
1628 u32 offset;
1629 int copied = 0;
1630
1631 if (sk->sk_state == TCP_LISTEN)
1632 return -ENOTCONN;
1633 while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1634 if (offset < skb->len) {
1635 int used;
1636 size_t len;
1637
1638 len = skb->len - offset;
1639 /* Stop reading if we hit a patch of urgent data */
1640 if (tp->urg_data) {
1641 u32 urg_offset = tp->urg_seq - seq;
1642 if (urg_offset < len)
1643 len = urg_offset;
1644 if (!len)
1645 break;
1646 }
1647 used = recv_actor(desc, skb, offset, len);
1648 if (used <= 0) {
1649 if (!copied)
1650 copied = used;
1651 break;
1652 } else if (used <= len) {
1653 seq += used;
1654 copied += used;
1655 offset += used;
1656 }
1657 /* If recv_actor drops the lock (e.g. TCP splice
1658 * receive) the skb pointer might be invalid when
1659 * getting here: tcp_collapse might have deleted it
1660 * while aggregating skbs from the socket queue.
1661 */
1662 skb = tcp_recv_skb(sk, seq - 1, &offset);
1663 if (!skb)
1664 break;
1665 /* TCP coalescing might have appended data to the skb.
1666 * Try to splice more frags
1667 */
1668 if (offset + 1 != skb->len)
1669 continue;
1670 }
1671 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) {
1672 sk_eat_skb(sk, skb);
1673 ++seq;
1674 break;
1675 }
1676 sk_eat_skb(sk, skb);
1677 if (!desc->count)
1678 break;
1679 WRITE_ONCE(tp->copied_seq, seq);
1680 }
1681 WRITE_ONCE(tp->copied_seq, seq);
1682
1683 tcp_rcv_space_adjust(sk);
1684
1685 /* Clean up data we have read: This will do ACK frames. */
1686 if (copied > 0) {
1687 tcp_recv_skb(sk, seq, &offset);
1688 tcp_cleanup_rbuf(sk, copied);
1689 }
1690 return copied;
1691}
1692EXPORT_SYMBOL(tcp_read_sock);
1693
1694int tcp_peek_len(struct socket *sock)
1695{
1696 return tcp_inq(sock->sk);
1697}
1698EXPORT_SYMBOL(tcp_peek_len);
1699
1700/* Make sure sk_rcvbuf is big enough to satisfy SO_RCVLOWAT hint */
1701int tcp_set_rcvlowat(struct sock *sk, int val)
1702{
1703 int cap;
1704
1705 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK)
1706 cap = sk->sk_rcvbuf >> 1;
1707 else
1708 cap = sock_net(sk)->ipv4.sysctl_tcp_rmem[2] >> 1;
1709 val = min(val, cap);
1710 WRITE_ONCE(sk->sk_rcvlowat, val ? : 1);
1711
1712 /* Check if we need to signal EPOLLIN right now */
1713 tcp_data_ready(sk);
1714
1715 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK)
1716 return 0;
1717
1718 val <<= 1;
1719 if (val > sk->sk_rcvbuf) {
1720 WRITE_ONCE(sk->sk_rcvbuf, val);
1721 tcp_sk(sk)->window_clamp = tcp_win_from_space(sk, val);
1722 }
1723 return 0;
1724}
1725EXPORT_SYMBOL(tcp_set_rcvlowat);
1726
1727#ifdef CONFIG_MMU
1728static const struct vm_operations_struct tcp_vm_ops = {
1729};
1730
1731int tcp_mmap(struct file *file, struct socket *sock,
1732 struct vm_area_struct *vma)
1733{
1734 if (vma->vm_flags & (VM_WRITE | VM_EXEC))
1735 return -EPERM;
1736 vma->vm_flags &= ~(VM_MAYWRITE | VM_MAYEXEC);
1737
1738 /* Instruct vm_insert_page() to not mmap_read_lock(mm) */
1739 vma->vm_flags |= VM_MIXEDMAP;
1740
1741 vma->vm_ops = &tcp_vm_ops;
1742 return 0;
1743}
1744EXPORT_SYMBOL(tcp_mmap);
1745
1746static int tcp_zerocopy_vm_insert_batch(struct vm_area_struct *vma,
1747 struct page **pages,
1748 unsigned long pages_to_map,
1749 unsigned long *insert_addr,
1750 u32 *length_with_pending,
1751 u32 *seq,
1752 struct tcp_zerocopy_receive *zc)
1753{
1754 unsigned long pages_remaining = pages_to_map;
1755 int bytes_mapped;
1756 int ret;
1757
1758 ret = vm_insert_pages(vma, *insert_addr, pages, &pages_remaining);
1759 bytes_mapped = PAGE_SIZE * (pages_to_map - pages_remaining);
1760 /* Even if vm_insert_pages fails, it may have partially succeeded in
1761 * mapping (some but not all of the pages).
1762 */
1763 *seq += bytes_mapped;
1764 *insert_addr += bytes_mapped;
1765 if (ret) {
1766 /* But if vm_insert_pages did fail, we have to unroll some state
1767 * we speculatively touched before.
1768 */
1769 const int bytes_not_mapped = PAGE_SIZE * pages_remaining;
1770 *length_with_pending -= bytes_not_mapped;
1771 zc->recv_skip_hint += bytes_not_mapped;
1772 }
1773 return ret;
1774}
1775
1776static int tcp_zerocopy_receive(struct sock *sk,
1777 struct tcp_zerocopy_receive *zc)
1778{
1779 unsigned long address = (unsigned long)zc->address;
1780 u32 length = 0, seq, offset, zap_len;
1781 #define PAGE_BATCH_SIZE 8
1782 struct page *pages[PAGE_BATCH_SIZE];
1783 const skb_frag_t *frags = NULL;
1784 struct vm_area_struct *vma;
1785 struct sk_buff *skb = NULL;
1786 unsigned long pg_idx = 0;
1787 unsigned long curr_addr;
1788 struct tcp_sock *tp;
1789 int inq;
1790 int ret;
1791
1792 if (address & (PAGE_SIZE - 1) || address != zc->address)
1793 return -EINVAL;
1794
1795 if (sk->sk_state == TCP_LISTEN)
1796 return -ENOTCONN;
1797
1798 sock_rps_record_flow(sk);
1799
1800 tp = tcp_sk(sk);
1801
1802 mmap_read_lock(current->mm);
1803
1804 vma = find_vma(current->mm, address);
1805 if (!vma || vma->vm_start > address || vma->vm_ops != &tcp_vm_ops) {
1806 mmap_read_unlock(current->mm);
1807 return -EINVAL;
1808 }
1809 zc->length = min_t(unsigned long, zc->length, vma->vm_end - address);
1810
1811 seq = tp->copied_seq;
1812 inq = tcp_inq(sk);
1813 zc->length = min_t(u32, zc->length, inq);
1814 zap_len = zc->length & ~(PAGE_SIZE - 1);
1815 if (zap_len) {
1816 zap_page_range(vma, address, zap_len);
1817 zc->recv_skip_hint = 0;
1818 } else {
1819 zc->recv_skip_hint = zc->length;
1820 }
1821 ret = 0;
1822 curr_addr = address;
1823 while (length + PAGE_SIZE <= zc->length) {
1824 if (zc->recv_skip_hint < PAGE_SIZE) {
1825 /* If we're here, finish the current batch. */
1826 if (pg_idx) {
1827 ret = tcp_zerocopy_vm_insert_batch(vma, pages,
1828 pg_idx,
1829 &curr_addr,
1830 &length,
1831 &seq, zc);
1832 if (ret)
1833 goto out;
1834 pg_idx = 0;
1835 }
1836 if (skb) {
1837 if (zc->recv_skip_hint > 0)
1838 break;
1839 skb = skb->next;
1840 offset = seq - TCP_SKB_CB(skb)->seq;
1841 } else {
1842 skb = tcp_recv_skb(sk, seq, &offset);
1843 }
1844 zc->recv_skip_hint = skb->len - offset;
1845 offset -= skb_headlen(skb);
1846 if ((int)offset < 0 || skb_has_frag_list(skb))
1847 break;
1848 frags = skb_shinfo(skb)->frags;
1849 while (offset) {
1850 if (skb_frag_size(frags) > offset)
1851 goto out;
1852 offset -= skb_frag_size(frags);
1853 frags++;
1854 }
1855 }
1856 if (skb_frag_size(frags) != PAGE_SIZE || skb_frag_off(frags)) {
1857 int remaining = zc->recv_skip_hint;
1858
1859 while (remaining && (skb_frag_size(frags) != PAGE_SIZE ||
1860 skb_frag_off(frags))) {
1861 remaining -= skb_frag_size(frags);
1862 frags++;
1863 }
1864 zc->recv_skip_hint -= remaining;
1865 break;
1866 }
1867 pages[pg_idx] = skb_frag_page(frags);
1868 pg_idx++;
1869 length += PAGE_SIZE;
1870 zc->recv_skip_hint -= PAGE_SIZE;
1871 frags++;
1872 if (pg_idx == PAGE_BATCH_SIZE) {
1873 ret = tcp_zerocopy_vm_insert_batch(vma, pages, pg_idx,
1874 &curr_addr, &length,
1875 &seq, zc);
1876 if (ret)
1877 goto out;
1878 pg_idx = 0;
1879 }
1880 }
1881 if (pg_idx) {
1882 ret = tcp_zerocopy_vm_insert_batch(vma, pages, pg_idx,
1883 &curr_addr, &length, &seq,
1884 zc);
1885 }
1886out:
1887 mmap_read_unlock(current->mm);
1888 if (length) {
1889 WRITE_ONCE(tp->copied_seq, seq);
1890 tcp_rcv_space_adjust(sk);
1891
1892 /* Clean up data we have read: This will do ACK frames. */
1893 tcp_recv_skb(sk, seq, &offset);
1894 tcp_cleanup_rbuf(sk, length);
1895 ret = 0;
1896 if (length == zc->length)
1897 zc->recv_skip_hint = 0;
1898 } else {
1899 if (!zc->recv_skip_hint && sock_flag(sk, SOCK_DONE))
1900 ret = -EIO;
1901 }
1902 zc->length = length;
1903 return ret;
1904}
1905#endif
1906
1907static void tcp_update_recv_tstamps(struct sk_buff *skb,
1908 struct scm_timestamping_internal *tss)
1909{
1910 if (skb->tstamp)
1911 tss->ts[0] = ktime_to_timespec64(skb->tstamp);
1912 else
1913 tss->ts[0] = (struct timespec64) {0};
1914
1915 if (skb_hwtstamps(skb)->hwtstamp)
1916 tss->ts[2] = ktime_to_timespec64(skb_hwtstamps(skb)->hwtstamp);
1917 else
1918 tss->ts[2] = (struct timespec64) {0};
1919}
1920
1921/* Similar to __sock_recv_timestamp, but does not require an skb */
1922static void tcp_recv_timestamp(struct msghdr *msg, const struct sock *sk,
1923 struct scm_timestamping_internal *tss)
1924{
1925 int new_tstamp = sock_flag(sk, SOCK_TSTAMP_NEW);
1926 bool has_timestamping = false;
1927
1928 if (tss->ts[0].tv_sec || tss->ts[0].tv_nsec) {
1929 if (sock_flag(sk, SOCK_RCVTSTAMP)) {
1930 if (sock_flag(sk, SOCK_RCVTSTAMPNS)) {
1931 if (new_tstamp) {
1932 struct __kernel_timespec kts = {
1933 .tv_sec = tss->ts[0].tv_sec,
1934 .tv_nsec = tss->ts[0].tv_nsec,
1935 };
1936 put_cmsg(msg, SOL_SOCKET, SO_TIMESTAMPNS_NEW,
1937 sizeof(kts), &kts);
1938 } else {
1939 struct __kernel_old_timespec ts_old = {
1940 .tv_sec = tss->ts[0].tv_sec,
1941 .tv_nsec = tss->ts[0].tv_nsec,
1942 };
1943 put_cmsg(msg, SOL_SOCKET, SO_TIMESTAMPNS_OLD,
1944 sizeof(ts_old), &ts_old);
1945 }
1946 } else {
1947 if (new_tstamp) {
1948 struct __kernel_sock_timeval stv = {
1949 .tv_sec = tss->ts[0].tv_sec,
1950 .tv_usec = tss->ts[0].tv_nsec / 1000,
1951 };
1952 put_cmsg(msg, SOL_SOCKET, SO_TIMESTAMP_NEW,
1953 sizeof(stv), &stv);
1954 } else {
1955 struct __kernel_old_timeval tv = {
1956 .tv_sec = tss->ts[0].tv_sec,
1957 .tv_usec = tss->ts[0].tv_nsec / 1000,
1958 };
1959 put_cmsg(msg, SOL_SOCKET, SO_TIMESTAMP_OLD,
1960 sizeof(tv), &tv);
1961 }
1962 }
1963 }
1964
1965 if (sk->sk_tsflags & SOF_TIMESTAMPING_SOFTWARE)
1966 has_timestamping = true;
1967 else
1968 tss->ts[0] = (struct timespec64) {0};
1969 }
1970
1971 if (tss->ts[2].tv_sec || tss->ts[2].tv_nsec) {
1972 if (sk->sk_tsflags & SOF_TIMESTAMPING_RAW_HARDWARE)
1973 has_timestamping = true;
1974 else
1975 tss->ts[2] = (struct timespec64) {0};
1976 }
1977
1978 if (has_timestamping) {
1979 tss->ts[1] = (struct timespec64) {0};
1980 if (sock_flag(sk, SOCK_TSTAMP_NEW))
1981 put_cmsg_scm_timestamping64(msg, tss);
1982 else
1983 put_cmsg_scm_timestamping(msg, tss);
1984 }
1985}
1986
1987static int tcp_inq_hint(struct sock *sk)
1988{
1989 const struct tcp_sock *tp = tcp_sk(sk);
1990 u32 copied_seq = READ_ONCE(tp->copied_seq);
1991 u32 rcv_nxt = READ_ONCE(tp->rcv_nxt);
1992 int inq;
1993
1994 inq = rcv_nxt - copied_seq;
1995 if (unlikely(inq < 0 || copied_seq != READ_ONCE(tp->copied_seq))) {
1996 lock_sock(sk);
1997 inq = tp->rcv_nxt - tp->copied_seq;
1998 release_sock(sk);
1999 }
2000 /* After receiving a FIN, tell the user-space to continue reading
2001 * by returning a non-zero inq.
2002 */
2003 if (inq == 0 && sock_flag(sk, SOCK_DONE))
2004 inq = 1;
2005 return inq;
2006}
2007
2008/*
2009 * This routine copies from a sock struct into the user buffer.
2010 *
2011 * Technical note: in 2.3 we work on _locked_ socket, so that
2012 * tricks with *seq access order and skb->users are not required.
2013 * Probably, code can be easily improved even more.
2014 */
2015
2016int tcp_recvmsg(struct sock *sk, struct msghdr *msg, size_t len, int nonblock,
2017 int flags, int *addr_len)
2018{
2019 struct tcp_sock *tp = tcp_sk(sk);
2020 int copied = 0;
2021 u32 peek_seq;
2022 u32 *seq;
2023 unsigned long used;
2024 int err, inq;
2025 int target; /* Read at least this many bytes */
2026 long timeo;
2027 struct sk_buff *skb, *last;
2028 u32 urg_hole = 0;
2029 struct scm_timestamping_internal tss;
2030 int cmsg_flags;
2031
2032 if (unlikely(flags & MSG_ERRQUEUE))
2033 return inet_recv_error(sk, msg, len, addr_len);
2034
2035 if (sk_can_busy_loop(sk) && skb_queue_empty_lockless(&sk->sk_receive_queue) &&
2036 (sk->sk_state == TCP_ESTABLISHED))
2037 sk_busy_loop(sk, nonblock);
2038
2039 lock_sock(sk);
2040
2041 err = -ENOTCONN;
2042 if (sk->sk_state == TCP_LISTEN)
2043 goto out;
2044
2045 cmsg_flags = tp->recvmsg_inq ? 1 : 0;
2046 timeo = sock_rcvtimeo(sk, nonblock);
2047
2048 /* Urgent data needs to be handled specially. */
2049 if (flags & MSG_OOB)
2050 goto recv_urg;
2051
2052 if (unlikely(tp->repair)) {
2053 err = -EPERM;
2054 if (!(flags & MSG_PEEK))
2055 goto out;
2056
2057 if (tp->repair_queue == TCP_SEND_QUEUE)
2058 goto recv_sndq;
2059
2060 err = -EINVAL;
2061 if (tp->repair_queue == TCP_NO_QUEUE)
2062 goto out;
2063
2064 /* 'common' recv queue MSG_PEEK-ing */
2065 }
2066
2067 seq = &tp->copied_seq;
2068 if (flags & MSG_PEEK) {
2069 peek_seq = tp->copied_seq;
2070 seq = &peek_seq;
2071 }
2072
2073 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
2074
2075 do {
2076 u32 offset;
2077
2078 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
2079 if (tp->urg_data && tp->urg_seq == *seq) {
2080 if (copied)
2081 break;
2082 if (signal_pending(current)) {
2083 copied = timeo ? sock_intr_errno(timeo) : -EAGAIN;
2084 break;
2085 }
2086 }
2087
2088 /* Next get a buffer. */
2089
2090 last = skb_peek_tail(&sk->sk_receive_queue);
2091 skb_queue_walk(&sk->sk_receive_queue, skb) {
2092 last = skb;
2093 /* Now that we have two receive queues this
2094 * shouldn't happen.
2095 */
2096 if (WARN(before(*seq, TCP_SKB_CB(skb)->seq),
2097 "TCP recvmsg seq # bug: copied %X, seq %X, rcvnxt %X, fl %X\n",
2098 *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt,
2099 flags))
2100 break;
2101
2102 offset = *seq - TCP_SKB_CB(skb)->seq;
2103 if (unlikely(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
2104 pr_err_once("%s: found a SYN, please report !\n", __func__);
2105 offset--;
2106 }
2107 if (offset < skb->len)
2108 goto found_ok_skb;
2109 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
2110 goto found_fin_ok;
2111 WARN(!(flags & MSG_PEEK),
2112 "TCP recvmsg seq # bug 2: copied %X, seq %X, rcvnxt %X, fl %X\n",
2113 *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt, flags);
2114 }
2115
2116 /* Well, if we have backlog, try to process it now yet. */
2117
2118 if (copied >= target && !READ_ONCE(sk->sk_backlog.tail))
2119 break;
2120
2121 if (copied) {
2122 if (sk->sk_err ||
2123 sk->sk_state == TCP_CLOSE ||
2124 (sk->sk_shutdown & RCV_SHUTDOWN) ||
2125 !timeo ||
2126 signal_pending(current))
2127 break;
2128 } else {
2129 if (sock_flag(sk, SOCK_DONE))
2130 break;
2131
2132 if (sk->sk_err) {
2133 copied = sock_error(sk);
2134 break;
2135 }
2136
2137 if (sk->sk_shutdown & RCV_SHUTDOWN)
2138 break;
2139
2140 if (sk->sk_state == TCP_CLOSE) {
2141 /* This occurs when user tries to read
2142 * from never connected socket.
2143 */
2144 copied = -ENOTCONN;
2145 break;
2146 }
2147
2148 if (!timeo) {
2149 copied = -EAGAIN;
2150 break;
2151 }
2152
2153 if (signal_pending(current)) {
2154 copied = sock_intr_errno(timeo);
2155 break;
2156 }
2157 }
2158
2159 tcp_cleanup_rbuf(sk, copied);
2160
2161 if (copied >= target) {
2162 /* Do not sleep, just process backlog. */
2163 release_sock(sk);
2164 lock_sock(sk);
2165 } else {
2166 sk_wait_data(sk, &timeo, last);
2167 }
2168
2169 if ((flags & MSG_PEEK) &&
2170 (peek_seq - copied - urg_hole != tp->copied_seq)) {
2171 net_dbg_ratelimited("TCP(%s:%d): Application bug, race in MSG_PEEK\n",
2172 current->comm,
2173 task_pid_nr(current));
2174 peek_seq = tp->copied_seq;
2175 }
2176 continue;
2177
2178found_ok_skb:
2179 /* Ok so how much can we use? */
2180 used = skb->len - offset;
2181 if (len < used)
2182 used = len;
2183
2184 /* Do we have urgent data here? */
2185 if (tp->urg_data) {
2186 u32 urg_offset = tp->urg_seq - *seq;
2187 if (urg_offset < used) {
2188 if (!urg_offset) {
2189 if (!sock_flag(sk, SOCK_URGINLINE)) {
2190 WRITE_ONCE(*seq, *seq + 1);
2191 urg_hole++;
2192 offset++;
2193 used--;
2194 if (!used)
2195 goto skip_copy;
2196 }
2197 } else
2198 used = urg_offset;
2199 }
2200 }
2201
2202 if (!(flags & MSG_TRUNC)) {
2203 err = skb_copy_datagram_msg(skb, offset, msg, used);
2204 if (err) {
2205 /* Exception. Bailout! */
2206 if (!copied)
2207 copied = -EFAULT;
2208 break;
2209 }
2210 }
2211
2212 WRITE_ONCE(*seq, *seq + used);
2213 copied += used;
2214 len -= used;
2215
2216 tcp_rcv_space_adjust(sk);
2217
2218skip_copy:
2219 if (tp->urg_data && after(tp->copied_seq, tp->urg_seq)) {
2220 tp->urg_data = 0;
2221 tcp_fast_path_check(sk);
2222 }
2223
2224 if (TCP_SKB_CB(skb)->has_rxtstamp) {
2225 tcp_update_recv_tstamps(skb, &tss);
2226 cmsg_flags |= 2;
2227 }
2228
2229 if (used + offset < skb->len)
2230 continue;
2231
2232 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
2233 goto found_fin_ok;
2234 if (!(flags & MSG_PEEK))
2235 sk_eat_skb(sk, skb);
2236 continue;
2237
2238found_fin_ok:
2239 /* Process the FIN. */
2240 WRITE_ONCE(*seq, *seq + 1);
2241 if (!(flags & MSG_PEEK))
2242 sk_eat_skb(sk, skb);
2243 break;
2244 } while (len > 0);
2245
2246 /* According to UNIX98, msg_name/msg_namelen are ignored
2247 * on connected socket. I was just happy when found this 8) --ANK
2248 */
2249
2250 /* Clean up data we have read: This will do ACK frames. */
2251 tcp_cleanup_rbuf(sk, copied);
2252
2253 release_sock(sk);
2254
2255 if (cmsg_flags) {
2256 if (cmsg_flags & 2)
2257 tcp_recv_timestamp(msg, sk, &tss);
2258 if (cmsg_flags & 1) {
2259 inq = tcp_inq_hint(sk);
2260 put_cmsg(msg, SOL_TCP, TCP_CM_INQ, sizeof(inq), &inq);
2261 }
2262 }
2263
2264 return copied;
2265
2266out:
2267 release_sock(sk);
2268 return err;
2269
2270recv_urg:
2271 err = tcp_recv_urg(sk, msg, len, flags);
2272 goto out;
2273
2274recv_sndq:
2275 err = tcp_peek_sndq(sk, msg, len);
2276 goto out;
2277}
2278EXPORT_SYMBOL(tcp_recvmsg);
2279
2280void tcp_set_state(struct sock *sk, int state)
2281{
2282 int oldstate = sk->sk_state;
2283
2284 /* We defined a new enum for TCP states that are exported in BPF
2285 * so as not force the internal TCP states to be frozen. The
2286 * following checks will detect if an internal state value ever
2287 * differs from the BPF value. If this ever happens, then we will
2288 * need to remap the internal value to the BPF value before calling
2289 * tcp_call_bpf_2arg.
2290 */
2291 BUILD_BUG_ON((int)BPF_TCP_ESTABLISHED != (int)TCP_ESTABLISHED);
2292 BUILD_BUG_ON((int)BPF_TCP_SYN_SENT != (int)TCP_SYN_SENT);
2293 BUILD_BUG_ON((int)BPF_TCP_SYN_RECV != (int)TCP_SYN_RECV);
2294 BUILD_BUG_ON((int)BPF_TCP_FIN_WAIT1 != (int)TCP_FIN_WAIT1);
2295 BUILD_BUG_ON((int)BPF_TCP_FIN_WAIT2 != (int)TCP_FIN_WAIT2);
2296 BUILD_BUG_ON((int)BPF_TCP_TIME_WAIT != (int)TCP_TIME_WAIT);
2297 BUILD_BUG_ON((int)BPF_TCP_CLOSE != (int)TCP_CLOSE);
2298 BUILD_BUG_ON((int)BPF_TCP_CLOSE_WAIT != (int)TCP_CLOSE_WAIT);
2299 BUILD_BUG_ON((int)BPF_TCP_LAST_ACK != (int)TCP_LAST_ACK);
2300 BUILD_BUG_ON((int)BPF_TCP_LISTEN != (int)TCP_LISTEN);
2301 BUILD_BUG_ON((int)BPF_TCP_CLOSING != (int)TCP_CLOSING);
2302 BUILD_BUG_ON((int)BPF_TCP_NEW_SYN_RECV != (int)TCP_NEW_SYN_RECV);
2303 BUILD_BUG_ON((int)BPF_TCP_MAX_STATES != (int)TCP_MAX_STATES);
2304
2305 if (BPF_SOCK_OPS_TEST_FLAG(tcp_sk(sk), BPF_SOCK_OPS_STATE_CB_FLAG))
2306 tcp_call_bpf_2arg(sk, BPF_SOCK_OPS_STATE_CB, oldstate, state);
2307
2308 switch (state) {
2309 case TCP_ESTABLISHED:
2310 if (oldstate != TCP_ESTABLISHED)
2311 TCP_INC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
2312 break;
2313
2314 case TCP_CLOSE:
2315 if (oldstate == TCP_CLOSE_WAIT || oldstate == TCP_ESTABLISHED)
2316 TCP_INC_STATS(sock_net(sk), TCP_MIB_ESTABRESETS);
2317
2318 sk->sk_prot->unhash(sk);
2319 if (inet_csk(sk)->icsk_bind_hash &&
2320 !(sk->sk_userlocks & SOCK_BINDPORT_LOCK))
2321 inet_put_port(sk);
2322 fallthrough;
2323 default:
2324 if (oldstate == TCP_ESTABLISHED)
2325 TCP_DEC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
2326 }
2327
2328 /* Change state AFTER socket is unhashed to avoid closed
2329 * socket sitting in hash tables.
2330 */
2331 inet_sk_state_store(sk, state);
2332}
2333EXPORT_SYMBOL_GPL(tcp_set_state);
2334
2335/*
2336 * State processing on a close. This implements the state shift for
2337 * sending our FIN frame. Note that we only send a FIN for some
2338 * states. A shutdown() may have already sent the FIN, or we may be
2339 * closed.
2340 */
2341
2342static const unsigned char new_state[16] = {
2343 /* current state: new state: action: */
2344 [0 /* (Invalid) */] = TCP_CLOSE,
2345 [TCP_ESTABLISHED] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2346 [TCP_SYN_SENT] = TCP_CLOSE,
2347 [TCP_SYN_RECV] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2348 [TCP_FIN_WAIT1] = TCP_FIN_WAIT1,
2349 [TCP_FIN_WAIT2] = TCP_FIN_WAIT2,
2350 [TCP_TIME_WAIT] = TCP_CLOSE,
2351 [TCP_CLOSE] = TCP_CLOSE,
2352 [TCP_CLOSE_WAIT] = TCP_LAST_ACK | TCP_ACTION_FIN,
2353 [TCP_LAST_ACK] = TCP_LAST_ACK,
2354 [TCP_LISTEN] = TCP_CLOSE,
2355 [TCP_CLOSING] = TCP_CLOSING,
2356 [TCP_NEW_SYN_RECV] = TCP_CLOSE, /* should not happen ! */
2357};
2358
2359static int tcp_close_state(struct sock *sk)
2360{
2361 int next = (int)new_state[sk->sk_state];
2362 int ns = next & TCP_STATE_MASK;
2363
2364 tcp_set_state(sk, ns);
2365
2366 return next & TCP_ACTION_FIN;
2367}
2368
2369/*
2370 * Shutdown the sending side of a connection. Much like close except
2371 * that we don't receive shut down or sock_set_flag(sk, SOCK_DEAD).
2372 */
2373
2374void tcp_shutdown(struct sock *sk, int how)
2375{
2376 /* We need to grab some memory, and put together a FIN,
2377 * and then put it into the queue to be sent.
2378 * Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
2379 */
2380 if (!(how & SEND_SHUTDOWN))
2381 return;
2382
2383 /* If we've already sent a FIN, or it's a closed state, skip this. */
2384 if ((1 << sk->sk_state) &
2385 (TCPF_ESTABLISHED | TCPF_SYN_SENT |
2386 TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) {
2387 /* Clear out any half completed packets. FIN if needed. */
2388 if (tcp_close_state(sk))
2389 tcp_send_fin(sk);
2390 }
2391}
2392EXPORT_SYMBOL(tcp_shutdown);
2393
2394bool tcp_check_oom(struct sock *sk, int shift)
2395{
2396 bool too_many_orphans, out_of_socket_memory;
2397
2398 too_many_orphans = tcp_too_many_orphans(sk, shift);
2399 out_of_socket_memory = tcp_out_of_memory(sk);
2400
2401 if (too_many_orphans)
2402 net_info_ratelimited("too many orphaned sockets\n");
2403 if (out_of_socket_memory)
2404 net_info_ratelimited("out of memory -- consider tuning tcp_mem\n");
2405 return too_many_orphans || out_of_socket_memory;
2406}
2407
2408void tcp_close(struct sock *sk, long timeout)
2409{
2410 struct sk_buff *skb;
2411 int data_was_unread = 0;
2412 int state;
2413
2414 lock_sock(sk);
2415 sk->sk_shutdown = SHUTDOWN_MASK;
2416
2417 if (sk->sk_state == TCP_LISTEN) {
2418 tcp_set_state(sk, TCP_CLOSE);
2419
2420 /* Special case. */
2421 inet_csk_listen_stop(sk);
2422
2423 goto adjudge_to_death;
2424 }
2425
2426 /* We need to flush the recv. buffs. We do this only on the
2427 * descriptor close, not protocol-sourced closes, because the
2428 * reader process may not have drained the data yet!
2429 */
2430 while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) {
2431 u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq;
2432
2433 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
2434 len--;
2435 data_was_unread += len;
2436 __kfree_skb(skb);
2437 }
2438
2439 sk_mem_reclaim(sk);
2440
2441 /* If socket has been already reset (e.g. in tcp_reset()) - kill it. */
2442 if (sk->sk_state == TCP_CLOSE)
2443 goto adjudge_to_death;
2444
2445 /* As outlined in RFC 2525, section 2.17, we send a RST here because
2446 * data was lost. To witness the awful effects of the old behavior of
2447 * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk
2448 * GET in an FTP client, suspend the process, wait for the client to
2449 * advertise a zero window, then kill -9 the FTP client, wheee...
2450 * Note: timeout is always zero in such a case.
2451 */
2452 if (unlikely(tcp_sk(sk)->repair)) {
2453 sk->sk_prot->disconnect(sk, 0);
2454 } else if (data_was_unread) {
2455 /* Unread data was tossed, zap the connection. */
2456 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONCLOSE);
2457 tcp_set_state(sk, TCP_CLOSE);
2458 tcp_send_active_reset(sk, sk->sk_allocation);
2459 } else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) {
2460 /* Check zero linger _after_ checking for unread data. */
2461 sk->sk_prot->disconnect(sk, 0);
2462 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
2463 } else if (tcp_close_state(sk)) {
2464 /* We FIN if the application ate all the data before
2465 * zapping the connection.
2466 */
2467
2468 /* RED-PEN. Formally speaking, we have broken TCP state
2469 * machine. State transitions:
2470 *
2471 * TCP_ESTABLISHED -> TCP_FIN_WAIT1
2472 * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible)
2473 * TCP_CLOSE_WAIT -> TCP_LAST_ACK
2474 *
2475 * are legal only when FIN has been sent (i.e. in window),
2476 * rather than queued out of window. Purists blame.
2477 *
2478 * F.e. "RFC state" is ESTABLISHED,
2479 * if Linux state is FIN-WAIT-1, but FIN is still not sent.
2480 *
2481 * The visible declinations are that sometimes
2482 * we enter time-wait state, when it is not required really
2483 * (harmless), do not send active resets, when they are
2484 * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
2485 * they look as CLOSING or LAST_ACK for Linux)
2486 * Probably, I missed some more holelets.
2487 * --ANK
2488 * XXX (TFO) - To start off we don't support SYN+ACK+FIN
2489 * in a single packet! (May consider it later but will
2490 * probably need API support or TCP_CORK SYN-ACK until
2491 * data is written and socket is closed.)
2492 */
2493 tcp_send_fin(sk);
2494 }
2495
2496 sk_stream_wait_close(sk, timeout);
2497
2498adjudge_to_death:
2499 state = sk->sk_state;
2500 sock_hold(sk);
2501 sock_orphan(sk);
2502
2503 local_bh_disable();
2504 bh_lock_sock(sk);
2505 /* remove backlog if any, without releasing ownership. */
2506 __release_sock(sk);
2507
2508 percpu_counter_inc(sk->sk_prot->orphan_count);
2509
2510 /* Have we already been destroyed by a softirq or backlog? */
2511 if (state != TCP_CLOSE && sk->sk_state == TCP_CLOSE)
2512 goto out;
2513
2514 /* This is a (useful) BSD violating of the RFC. There is a
2515 * problem with TCP as specified in that the other end could
2516 * keep a socket open forever with no application left this end.
2517 * We use a 1 minute timeout (about the same as BSD) then kill
2518 * our end. If they send after that then tough - BUT: long enough
2519 * that we won't make the old 4*rto = almost no time - whoops
2520 * reset mistake.
2521 *
2522 * Nope, it was not mistake. It is really desired behaviour
2523 * f.e. on http servers, when such sockets are useless, but
2524 * consume significant resources. Let's do it with special
2525 * linger2 option. --ANK
2526 */
2527
2528 if (sk->sk_state == TCP_FIN_WAIT2) {
2529 struct tcp_sock *tp = tcp_sk(sk);
2530 if (tp->linger2 < 0) {
2531 tcp_set_state(sk, TCP_CLOSE);
2532 tcp_send_active_reset(sk, GFP_ATOMIC);
2533 __NET_INC_STATS(sock_net(sk),
2534 LINUX_MIB_TCPABORTONLINGER);
2535 } else {
2536 const int tmo = tcp_fin_time(sk);
2537
2538 if (tmo > TCP_TIMEWAIT_LEN) {
2539 inet_csk_reset_keepalive_timer(sk,
2540 tmo - TCP_TIMEWAIT_LEN);
2541 } else {
2542 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
2543 goto out;
2544 }
2545 }
2546 }
2547 if (sk->sk_state != TCP_CLOSE) {
2548 sk_mem_reclaim(sk);
2549 if (tcp_check_oom(sk, 0)) {
2550 tcp_set_state(sk, TCP_CLOSE);
2551 tcp_send_active_reset(sk, GFP_ATOMIC);
2552 __NET_INC_STATS(sock_net(sk),
2553 LINUX_MIB_TCPABORTONMEMORY);
2554 } else if (!check_net(sock_net(sk))) {
2555 /* Not possible to send reset; just close */
2556 tcp_set_state(sk, TCP_CLOSE);
2557 }
2558 }
2559
2560 if (sk->sk_state == TCP_CLOSE) {
2561 struct request_sock *req;
2562
2563 req = rcu_dereference_protected(tcp_sk(sk)->fastopen_rsk,
2564 lockdep_sock_is_held(sk));
2565 /* We could get here with a non-NULL req if the socket is
2566 * aborted (e.g., closed with unread data) before 3WHS
2567 * finishes.
2568 */
2569 if (req)
2570 reqsk_fastopen_remove(sk, req, false);
2571 inet_csk_destroy_sock(sk);
2572 }
2573 /* Otherwise, socket is reprieved until protocol close. */
2574
2575out:
2576 bh_unlock_sock(sk);
2577 local_bh_enable();
2578 release_sock(sk);
2579 sock_put(sk);
2580}
2581EXPORT_SYMBOL(tcp_close);
2582
2583/* These states need RST on ABORT according to RFC793 */
2584
2585static inline bool tcp_need_reset(int state)
2586{
2587 return (1 << state) &
2588 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 |
2589 TCPF_FIN_WAIT2 | TCPF_SYN_RECV);
2590}
2591
2592static void tcp_rtx_queue_purge(struct sock *sk)
2593{
2594 struct rb_node *p = rb_first(&sk->tcp_rtx_queue);
2595
2596 tcp_sk(sk)->highest_sack = NULL;
2597 while (p) {
2598 struct sk_buff *skb = rb_to_skb(p);
2599
2600 p = rb_next(p);
2601 /* Since we are deleting whole queue, no need to
2602 * list_del(&skb->tcp_tsorted_anchor)
2603 */
2604 tcp_rtx_queue_unlink(skb, sk);
2605 sk_wmem_free_skb(sk, skb);
2606 }
2607}
2608
2609void tcp_write_queue_purge(struct sock *sk)
2610{
2611 struct sk_buff *skb;
2612
2613 tcp_chrono_stop(sk, TCP_CHRONO_BUSY);
2614 while ((skb = __skb_dequeue(&sk->sk_write_queue)) != NULL) {
2615 tcp_skb_tsorted_anchor_cleanup(skb);
2616 sk_wmem_free_skb(sk, skb);
2617 }
2618 tcp_rtx_queue_purge(sk);
2619 skb = sk->sk_tx_skb_cache;
2620 if (skb) {
2621 __kfree_skb(skb);
2622 sk->sk_tx_skb_cache = NULL;
2623 }
2624 INIT_LIST_HEAD(&tcp_sk(sk)->tsorted_sent_queue);
2625 sk_mem_reclaim(sk);
2626 tcp_clear_all_retrans_hints(tcp_sk(sk));
2627 tcp_sk(sk)->packets_out = 0;
2628 inet_csk(sk)->icsk_backoff = 0;
2629}
2630
2631int tcp_disconnect(struct sock *sk, int flags)
2632{
2633 struct inet_sock *inet = inet_sk(sk);
2634 struct inet_connection_sock *icsk = inet_csk(sk);
2635 struct tcp_sock *tp = tcp_sk(sk);
2636 int old_state = sk->sk_state;
2637 u32 seq;
2638
2639 if (old_state != TCP_CLOSE)
2640 tcp_set_state(sk, TCP_CLOSE);
2641
2642 /* ABORT function of RFC793 */
2643 if (old_state == TCP_LISTEN) {
2644 inet_csk_listen_stop(sk);
2645 } else if (unlikely(tp->repair)) {
2646 sk->sk_err = ECONNABORTED;
2647 } else if (tcp_need_reset(old_state) ||
2648 (tp->snd_nxt != tp->write_seq &&
2649 (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK))) {
2650 /* The last check adjusts for discrepancy of Linux wrt. RFC
2651 * states
2652 */
2653 tcp_send_active_reset(sk, gfp_any());
2654 sk->sk_err = ECONNRESET;
2655 } else if (old_state == TCP_SYN_SENT)
2656 sk->sk_err = ECONNRESET;
2657
2658 tcp_clear_xmit_timers(sk);
2659 __skb_queue_purge(&sk->sk_receive_queue);
2660 if (sk->sk_rx_skb_cache) {
2661 __kfree_skb(sk->sk_rx_skb_cache);
2662 sk->sk_rx_skb_cache = NULL;
2663 }
2664 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt);
2665 tp->urg_data = 0;
2666 tcp_write_queue_purge(sk);
2667 tcp_fastopen_active_disable_ofo_check(sk);
2668 skb_rbtree_purge(&tp->out_of_order_queue);
2669
2670 inet->inet_dport = 0;
2671
2672 if (!(sk->sk_userlocks & SOCK_BINDADDR_LOCK))
2673 inet_reset_saddr(sk);
2674
2675 sk->sk_shutdown = 0;
2676 sock_reset_flag(sk, SOCK_DONE);
2677 tp->srtt_us = 0;
2678 tp->mdev_us = jiffies_to_usecs(TCP_TIMEOUT_INIT);
2679 tp->rcv_rtt_last_tsecr = 0;
2680
2681 seq = tp->write_seq + tp->max_window + 2;
2682 if (!seq)
2683 seq = 1;
2684 WRITE_ONCE(tp->write_seq, seq);
2685
2686 icsk->icsk_backoff = 0;
2687 icsk->icsk_probes_out = 0;
2688 icsk->icsk_rto = TCP_TIMEOUT_INIT;
2689 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
2690 tp->snd_cwnd = TCP_INIT_CWND;
2691 tp->snd_cwnd_cnt = 0;
2692 tp->window_clamp = 0;
2693 tp->delivered = 0;
2694 tp->delivered_ce = 0;
2695 if (icsk->icsk_ca_ops->release)
2696 icsk->icsk_ca_ops->release(sk);
2697 memset(icsk->icsk_ca_priv, 0, sizeof(icsk->icsk_ca_priv));
2698 tcp_set_ca_state(sk, TCP_CA_Open);
2699 tp->is_sack_reneg = 0;
2700 tcp_clear_retrans(tp);
2701 tp->total_retrans = 0;
2702 inet_csk_delack_init(sk);
2703 /* Initialize rcv_mss to TCP_MIN_MSS to avoid division by 0
2704 * issue in __tcp_select_window()
2705 */
2706 icsk->icsk_ack.rcv_mss = TCP_MIN_MSS;
2707 memset(&tp->rx_opt, 0, sizeof(tp->rx_opt));
2708 __sk_dst_reset(sk);
2709 dst_release(sk->sk_rx_dst);
2710 sk->sk_rx_dst = NULL;
2711 tcp_saved_syn_free(tp);
2712 tp->compressed_ack = 0;
2713 tp->segs_in = 0;
2714 tp->segs_out = 0;
2715 tp->bytes_sent = 0;
2716 tp->bytes_acked = 0;
2717 tp->bytes_received = 0;
2718 tp->bytes_retrans = 0;
2719 tp->data_segs_in = 0;
2720 tp->data_segs_out = 0;
2721 tp->duplicate_sack[0].start_seq = 0;
2722 tp->duplicate_sack[0].end_seq = 0;
2723 tp->dsack_dups = 0;
2724 tp->reord_seen = 0;
2725 tp->retrans_out = 0;
2726 tp->sacked_out = 0;
2727 tp->tlp_high_seq = 0;
2728 tp->last_oow_ack_time = 0;
2729 /* There's a bubble in the pipe until at least the first ACK. */
2730 tp->app_limited = ~0U;
2731 tp->rack.mstamp = 0;
2732 tp->rack.advanced = 0;
2733 tp->rack.reo_wnd_steps = 1;
2734 tp->rack.last_delivered = 0;
2735 tp->rack.reo_wnd_persist = 0;
2736 tp->rack.dsack_seen = 0;
2737 tp->syn_data_acked = 0;
2738 tp->rx_opt.saw_tstamp = 0;
2739 tp->rx_opt.dsack = 0;
2740 tp->rx_opt.num_sacks = 0;
2741 tp->rcv_ooopack = 0;
2742
2743
2744 /* Clean up fastopen related fields */
2745 tcp_free_fastopen_req(tp);
2746 inet->defer_connect = 0;
2747 tp->fastopen_client_fail = 0;
2748
2749 WARN_ON(inet->inet_num && !icsk->icsk_bind_hash);
2750
2751 if (sk->sk_frag.page) {
2752 put_page(sk->sk_frag.page);
2753 sk->sk_frag.page = NULL;
2754 sk->sk_frag.offset = 0;
2755 }
2756
2757 sk->sk_error_report(sk);
2758 return 0;
2759}
2760EXPORT_SYMBOL(tcp_disconnect);
2761
2762static inline bool tcp_can_repair_sock(const struct sock *sk)
2763{
2764 return ns_capable(sock_net(sk)->user_ns, CAP_NET_ADMIN) &&
2765 (sk->sk_state != TCP_LISTEN);
2766}
2767
2768static int tcp_repair_set_window(struct tcp_sock *tp, sockptr_t optbuf, int len)
2769{
2770 struct tcp_repair_window opt;
2771
2772 if (!tp->repair)
2773 return -EPERM;
2774
2775 if (len != sizeof(opt))
2776 return -EINVAL;
2777
2778 if (copy_from_sockptr(&opt, optbuf, sizeof(opt)))
2779 return -EFAULT;
2780
2781 if (opt.max_window < opt.snd_wnd)
2782 return -EINVAL;
2783
2784 if (after(opt.snd_wl1, tp->rcv_nxt + opt.rcv_wnd))
2785 return -EINVAL;
2786
2787 if (after(opt.rcv_wup, tp->rcv_nxt))
2788 return -EINVAL;
2789
2790 tp->snd_wl1 = opt.snd_wl1;
2791 tp->snd_wnd = opt.snd_wnd;
2792 tp->max_window = opt.max_window;
2793
2794 tp->rcv_wnd = opt.rcv_wnd;
2795 tp->rcv_wup = opt.rcv_wup;
2796
2797 return 0;
2798}
2799
2800static int tcp_repair_options_est(struct sock *sk, sockptr_t optbuf,
2801 unsigned int len)
2802{
2803 struct tcp_sock *tp = tcp_sk(sk);
2804 struct tcp_repair_opt opt;
2805 size_t offset = 0;
2806
2807 while (len >= sizeof(opt)) {
2808 if (copy_from_sockptr_offset(&opt, optbuf, offset, sizeof(opt)))
2809 return -EFAULT;
2810
2811 offset += sizeof(opt);
2812 len -= sizeof(opt);
2813
2814 switch (opt.opt_code) {
2815 case TCPOPT_MSS:
2816 tp->rx_opt.mss_clamp = opt.opt_val;
2817 tcp_mtup_init(sk);
2818 break;
2819 case TCPOPT_WINDOW:
2820 {
2821 u16 snd_wscale = opt.opt_val & 0xFFFF;
2822 u16 rcv_wscale = opt.opt_val >> 16;
2823
2824 if (snd_wscale > TCP_MAX_WSCALE || rcv_wscale > TCP_MAX_WSCALE)
2825 return -EFBIG;
2826
2827 tp->rx_opt.snd_wscale = snd_wscale;
2828 tp->rx_opt.rcv_wscale = rcv_wscale;
2829 tp->rx_opt.wscale_ok = 1;
2830 }
2831 break;
2832 case TCPOPT_SACK_PERM:
2833 if (opt.opt_val != 0)
2834 return -EINVAL;
2835
2836 tp->rx_opt.sack_ok |= TCP_SACK_SEEN;
2837 break;
2838 case TCPOPT_TIMESTAMP:
2839 if (opt.opt_val != 0)
2840 return -EINVAL;
2841
2842 tp->rx_opt.tstamp_ok = 1;
2843 break;
2844 }
2845 }
2846
2847 return 0;
2848}
2849
2850DEFINE_STATIC_KEY_FALSE(tcp_tx_delay_enabled);
2851EXPORT_SYMBOL(tcp_tx_delay_enabled);
2852
2853static void tcp_enable_tx_delay(void)
2854{
2855 if (!static_branch_unlikely(&tcp_tx_delay_enabled)) {
2856 static int __tcp_tx_delay_enabled = 0;
2857
2858 if (cmpxchg(&__tcp_tx_delay_enabled, 0, 1) == 0) {
2859 static_branch_enable(&tcp_tx_delay_enabled);
2860 pr_info("TCP_TX_DELAY enabled\n");
2861 }
2862 }
2863}
2864
2865/* When set indicates to always queue non-full frames. Later the user clears
2866 * this option and we transmit any pending partial frames in the queue. This is
2867 * meant to be used alongside sendfile() to get properly filled frames when the
2868 * user (for example) must write out headers with a write() call first and then
2869 * use sendfile to send out the data parts.
2870 *
2871 * TCP_CORK can be set together with TCP_NODELAY and it is stronger than
2872 * TCP_NODELAY.
2873 */
2874static void __tcp_sock_set_cork(struct sock *sk, bool on)
2875{
2876 struct tcp_sock *tp = tcp_sk(sk);
2877
2878 if (on) {
2879 tp->nonagle |= TCP_NAGLE_CORK;
2880 } else {
2881 tp->nonagle &= ~TCP_NAGLE_CORK;
2882 if (tp->nonagle & TCP_NAGLE_OFF)
2883 tp->nonagle |= TCP_NAGLE_PUSH;
2884 tcp_push_pending_frames(sk);
2885 }
2886}
2887
2888void tcp_sock_set_cork(struct sock *sk, bool on)
2889{
2890 lock_sock(sk);
2891 __tcp_sock_set_cork(sk, on);
2892 release_sock(sk);
2893}
2894EXPORT_SYMBOL(tcp_sock_set_cork);
2895
2896/* TCP_NODELAY is weaker than TCP_CORK, so that this option on corked socket is
2897 * remembered, but it is not activated until cork is cleared.
2898 *
2899 * However, when TCP_NODELAY is set we make an explicit push, which overrides
2900 * even TCP_CORK for currently queued segments.
2901 */
2902static void __tcp_sock_set_nodelay(struct sock *sk, bool on)
2903{
2904 if (on) {
2905 tcp_sk(sk)->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH;
2906 tcp_push_pending_frames(sk);
2907 } else {
2908 tcp_sk(sk)->nonagle &= ~TCP_NAGLE_OFF;
2909 }
2910}
2911
2912void tcp_sock_set_nodelay(struct sock *sk)
2913{
2914 lock_sock(sk);
2915 __tcp_sock_set_nodelay(sk, true);
2916 release_sock(sk);
2917}
2918EXPORT_SYMBOL(tcp_sock_set_nodelay);
2919
2920static void __tcp_sock_set_quickack(struct sock *sk, int val)
2921{
2922 if (!val) {
2923 inet_csk_enter_pingpong_mode(sk);
2924 return;
2925 }
2926
2927 inet_csk_exit_pingpong_mode(sk);
2928 if ((1 << sk->sk_state) & (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) &&
2929 inet_csk_ack_scheduled(sk)) {
2930 inet_csk(sk)->icsk_ack.pending |= ICSK_ACK_PUSHED;
2931 tcp_cleanup_rbuf(sk, 1);
2932 if (!(val & 1))
2933 inet_csk_enter_pingpong_mode(sk);
2934 }
2935}
2936
2937void tcp_sock_set_quickack(struct sock *sk, int val)
2938{
2939 lock_sock(sk);
2940 __tcp_sock_set_quickack(sk, val);
2941 release_sock(sk);
2942}
2943EXPORT_SYMBOL(tcp_sock_set_quickack);
2944
2945int tcp_sock_set_syncnt(struct sock *sk, int val)
2946{
2947 if (val < 1 || val > MAX_TCP_SYNCNT)
2948 return -EINVAL;
2949
2950 lock_sock(sk);
2951 inet_csk(sk)->icsk_syn_retries = val;
2952 release_sock(sk);
2953 return 0;
2954}
2955EXPORT_SYMBOL(tcp_sock_set_syncnt);
2956
2957void tcp_sock_set_user_timeout(struct sock *sk, u32 val)
2958{
2959 lock_sock(sk);
2960 inet_csk(sk)->icsk_user_timeout = val;
2961 release_sock(sk);
2962}
2963EXPORT_SYMBOL(tcp_sock_set_user_timeout);
2964
2965int tcp_sock_set_keepidle_locked(struct sock *sk, int val)
2966{
2967 struct tcp_sock *tp = tcp_sk(sk);
2968
2969 if (val < 1 || val > MAX_TCP_KEEPIDLE)
2970 return -EINVAL;
2971
2972 tp->keepalive_time = val * HZ;
2973 if (sock_flag(sk, SOCK_KEEPOPEN) &&
2974 !((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN))) {
2975 u32 elapsed = keepalive_time_elapsed(tp);
2976
2977 if (tp->keepalive_time > elapsed)
2978 elapsed = tp->keepalive_time - elapsed;
2979 else
2980 elapsed = 0;
2981 inet_csk_reset_keepalive_timer(sk, elapsed);
2982 }
2983
2984 return 0;
2985}
2986
2987int tcp_sock_set_keepidle(struct sock *sk, int val)
2988{
2989 int err;
2990
2991 lock_sock(sk);
2992 err = tcp_sock_set_keepidle_locked(sk, val);
2993 release_sock(sk);
2994 return err;
2995}
2996EXPORT_SYMBOL(tcp_sock_set_keepidle);
2997
2998int tcp_sock_set_keepintvl(struct sock *sk, int val)
2999{
3000 if (val < 1 || val > MAX_TCP_KEEPINTVL)
3001 return -EINVAL;
3002
3003 lock_sock(sk);
3004 tcp_sk(sk)->keepalive_intvl = val * HZ;
3005 release_sock(sk);
3006 return 0;
3007}
3008EXPORT_SYMBOL(tcp_sock_set_keepintvl);
3009
3010int tcp_sock_set_keepcnt(struct sock *sk, int val)
3011{
3012 if (val < 1 || val > MAX_TCP_KEEPCNT)
3013 return -EINVAL;
3014
3015 lock_sock(sk);
3016 tcp_sk(sk)->keepalive_probes = val;
3017 release_sock(sk);
3018 return 0;
3019}
3020EXPORT_SYMBOL(tcp_sock_set_keepcnt);
3021
3022/*
3023 * Socket option code for TCP.
3024 */
3025static int do_tcp_setsockopt(struct sock *sk, int level, int optname,
3026 sockptr_t optval, unsigned int optlen)
3027{
3028 struct tcp_sock *tp = tcp_sk(sk);
3029 struct inet_connection_sock *icsk = inet_csk(sk);
3030 struct net *net = sock_net(sk);
3031 int val;
3032 int err = 0;
3033
3034 /* These are data/string values, all the others are ints */
3035 switch (optname) {
3036 case TCP_CONGESTION: {
3037 char name[TCP_CA_NAME_MAX];
3038
3039 if (optlen < 1)
3040 return -EINVAL;
3041
3042 val = strncpy_from_sockptr(name, optval,
3043 min_t(long, TCP_CA_NAME_MAX-1, optlen));
3044 if (val < 0)
3045 return -EFAULT;
3046 name[val] = 0;
3047
3048 lock_sock(sk);
3049 err = tcp_set_congestion_control(sk, name, true, true,
3050 ns_capable(sock_net(sk)->user_ns,
3051 CAP_NET_ADMIN));
3052 release_sock(sk);
3053 return err;
3054 }
3055 case TCP_ULP: {
3056 char name[TCP_ULP_NAME_MAX];
3057
3058 if (optlen < 1)
3059 return -EINVAL;
3060
3061 val = strncpy_from_sockptr(name, optval,
3062 min_t(long, TCP_ULP_NAME_MAX - 1,
3063 optlen));
3064 if (val < 0)
3065 return -EFAULT;
3066 name[val] = 0;
3067
3068 lock_sock(sk);
3069 err = tcp_set_ulp(sk, name);
3070 release_sock(sk);
3071 return err;
3072 }
3073 case TCP_FASTOPEN_KEY: {
3074 __u8 key[TCP_FASTOPEN_KEY_BUF_LENGTH];
3075 __u8 *backup_key = NULL;
3076
3077 /* Allow a backup key as well to facilitate key rotation
3078 * First key is the active one.
3079 */
3080 if (optlen != TCP_FASTOPEN_KEY_LENGTH &&
3081 optlen != TCP_FASTOPEN_KEY_BUF_LENGTH)
3082 return -EINVAL;
3083
3084 if (copy_from_sockptr(key, optval, optlen))
3085 return -EFAULT;
3086
3087 if (optlen == TCP_FASTOPEN_KEY_BUF_LENGTH)
3088 backup_key = key + TCP_FASTOPEN_KEY_LENGTH;
3089
3090 return tcp_fastopen_reset_cipher(net, sk, key, backup_key);
3091 }
3092 default:
3093 /* fallthru */
3094 break;
3095 }
3096
3097 if (optlen < sizeof(int))
3098 return -EINVAL;
3099
3100 if (copy_from_sockptr(&val, optval, sizeof(val)))
3101 return -EFAULT;
3102
3103 lock_sock(sk);
3104
3105 switch (optname) {
3106 case TCP_MAXSEG:
3107 /* Values greater than interface MTU won't take effect. However
3108 * at the point when this call is done we typically don't yet
3109 * know which interface is going to be used
3110 */
3111 if (val && (val < TCP_MIN_MSS || val > MAX_TCP_WINDOW)) {
3112 err = -EINVAL;
3113 break;
3114 }
3115 tp->rx_opt.user_mss = val;
3116 break;
3117
3118 case TCP_NODELAY:
3119 __tcp_sock_set_nodelay(sk, val);
3120 break;
3121
3122 case TCP_THIN_LINEAR_TIMEOUTS:
3123 if (val < 0 || val > 1)
3124 err = -EINVAL;
3125 else
3126 tp->thin_lto = val;
3127 break;
3128
3129 case TCP_THIN_DUPACK:
3130 if (val < 0 || val > 1)
3131 err = -EINVAL;
3132 break;
3133
3134 case TCP_REPAIR:
3135 if (!tcp_can_repair_sock(sk))
3136 err = -EPERM;
3137 else if (val == TCP_REPAIR_ON) {
3138 tp->repair = 1;
3139 sk->sk_reuse = SK_FORCE_REUSE;
3140 tp->repair_queue = TCP_NO_QUEUE;
3141 } else if (val == TCP_REPAIR_OFF) {
3142 tp->repair = 0;
3143 sk->sk_reuse = SK_NO_REUSE;
3144 tcp_send_window_probe(sk);
3145 } else if (val == TCP_REPAIR_OFF_NO_WP) {
3146 tp->repair = 0;
3147 sk->sk_reuse = SK_NO_REUSE;
3148 } else
3149 err = -EINVAL;
3150
3151 break;
3152
3153 case TCP_REPAIR_QUEUE:
3154 if (!tp->repair)
3155 err = -EPERM;
3156 else if ((unsigned int)val < TCP_QUEUES_NR)
3157 tp->repair_queue = val;
3158 else
3159 err = -EINVAL;
3160 break;
3161
3162 case TCP_QUEUE_SEQ:
3163 if (sk->sk_state != TCP_CLOSE)
3164 err = -EPERM;
3165 else if (tp->repair_queue == TCP_SEND_QUEUE)
3166 WRITE_ONCE(tp->write_seq, val);
3167 else if (tp->repair_queue == TCP_RECV_QUEUE) {
3168 WRITE_ONCE(tp->rcv_nxt, val);
3169 WRITE_ONCE(tp->copied_seq, val);
3170 }
3171 else
3172 err = -EINVAL;
3173 break;
3174
3175 case TCP_REPAIR_OPTIONS:
3176 if (!tp->repair)
3177 err = -EINVAL;
3178 else if (sk->sk_state == TCP_ESTABLISHED)
3179 err = tcp_repair_options_est(sk, optval, optlen);
3180 else
3181 err = -EPERM;
3182 break;
3183
3184 case TCP_CORK:
3185 __tcp_sock_set_cork(sk, val);
3186 break;
3187
3188 case TCP_KEEPIDLE:
3189 err = tcp_sock_set_keepidle_locked(sk, val);
3190 break;
3191 case TCP_KEEPINTVL:
3192 if (val < 1 || val > MAX_TCP_KEEPINTVL)
3193 err = -EINVAL;
3194 else
3195 tp->keepalive_intvl = val * HZ;
3196 break;
3197 case TCP_KEEPCNT:
3198 if (val < 1 || val > MAX_TCP_KEEPCNT)
3199 err = -EINVAL;
3200 else
3201 tp->keepalive_probes = val;
3202 break;
3203 case TCP_SYNCNT:
3204 if (val < 1 || val > MAX_TCP_SYNCNT)
3205 err = -EINVAL;
3206 else
3207 icsk->icsk_syn_retries = val;
3208 break;
3209
3210 case TCP_SAVE_SYN:
3211 if (val < 0 || val > 1)
3212 err = -EINVAL;
3213 else
3214 tp->save_syn = val;
3215 break;
3216
3217 case TCP_LINGER2:
3218 if (val < 0)
3219 tp->linger2 = -1;
3220 else if (val > TCP_FIN_TIMEOUT_MAX / HZ)
3221 tp->linger2 = TCP_FIN_TIMEOUT_MAX;
3222 else
3223 tp->linger2 = val * HZ;
3224 break;
3225
3226 case TCP_DEFER_ACCEPT:
3227 /* Translate value in seconds to number of retransmits */
3228 icsk->icsk_accept_queue.rskq_defer_accept =
3229 secs_to_retrans(val, TCP_TIMEOUT_INIT / HZ,
3230 TCP_RTO_MAX / HZ);
3231 break;
3232
3233 case TCP_WINDOW_CLAMP:
3234 if (!val) {
3235 if (sk->sk_state != TCP_CLOSE) {
3236 err = -EINVAL;
3237 break;
3238 }
3239 tp->window_clamp = 0;
3240 } else
3241 tp->window_clamp = val < SOCK_MIN_RCVBUF / 2 ?
3242 SOCK_MIN_RCVBUF / 2 : val;
3243 break;
3244
3245 case TCP_QUICKACK:
3246 __tcp_sock_set_quickack(sk, val);
3247 break;
3248
3249#ifdef CONFIG_TCP_MD5SIG
3250 case TCP_MD5SIG:
3251 case TCP_MD5SIG_EXT:
3252 err = tp->af_specific->md5_parse(sk, optname, optval, optlen);
3253 break;
3254#endif
3255 case TCP_USER_TIMEOUT:
3256 /* Cap the max time in ms TCP will retry or probe the window
3257 * before giving up and aborting (ETIMEDOUT) a connection.
3258 */
3259 if (val < 0)
3260 err = -EINVAL;
3261 else
3262 icsk->icsk_user_timeout = val;
3263 break;
3264
3265 case TCP_FASTOPEN:
3266 if (val >= 0 && ((1 << sk->sk_state) & (TCPF_CLOSE |
3267 TCPF_LISTEN))) {
3268 tcp_fastopen_init_key_once(net);
3269
3270 fastopen_queue_tune(sk, val);
3271 } else {
3272 err = -EINVAL;
3273 }
3274 break;
3275 case TCP_FASTOPEN_CONNECT:
3276 if (val > 1 || val < 0) {
3277 err = -EINVAL;
3278 } else if (net->ipv4.sysctl_tcp_fastopen & TFO_CLIENT_ENABLE) {
3279 if (sk->sk_state == TCP_CLOSE)
3280 tp->fastopen_connect = val;
3281 else
3282 err = -EINVAL;
3283 } else {
3284 err = -EOPNOTSUPP;
3285 }
3286 break;
3287 case TCP_FASTOPEN_NO_COOKIE:
3288 if (val > 1 || val < 0)
3289 err = -EINVAL;
3290 else if (!((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
3291 err = -EINVAL;
3292 else
3293 tp->fastopen_no_cookie = val;
3294 break;
3295 case TCP_TIMESTAMP:
3296 if (!tp->repair)
3297 err = -EPERM;
3298 else
3299 tp->tsoffset = val - tcp_time_stamp_raw();
3300 break;
3301 case TCP_REPAIR_WINDOW:
3302 err = tcp_repair_set_window(tp, optval, optlen);
3303 break;
3304 case TCP_NOTSENT_LOWAT:
3305 tp->notsent_lowat = val;
3306 sk->sk_write_space(sk);
3307 break;
3308 case TCP_INQ:
3309 if (val > 1 || val < 0)
3310 err = -EINVAL;
3311 else
3312 tp->recvmsg_inq = val;
3313 break;
3314 case TCP_TX_DELAY:
3315 if (val)
3316 tcp_enable_tx_delay();
3317 tp->tcp_tx_delay = val;
3318 break;
3319 default:
3320 err = -ENOPROTOOPT;
3321 break;
3322 }
3323
3324 release_sock(sk);
3325 return err;
3326}
3327
3328int tcp_setsockopt(struct sock *sk, int level, int optname, sockptr_t optval,
3329 unsigned int optlen)
3330{
3331 const struct inet_connection_sock *icsk = inet_csk(sk);
3332
3333 if (level != SOL_TCP)
3334 return icsk->icsk_af_ops->setsockopt(sk, level, optname,
3335 optval, optlen);
3336 return do_tcp_setsockopt(sk, level, optname, optval, optlen);
3337}
3338EXPORT_SYMBOL(tcp_setsockopt);
3339
3340static void tcp_get_info_chrono_stats(const struct tcp_sock *tp,
3341 struct tcp_info *info)
3342{
3343 u64 stats[__TCP_CHRONO_MAX], total = 0;
3344 enum tcp_chrono i;
3345
3346 for (i = TCP_CHRONO_BUSY; i < __TCP_CHRONO_MAX; ++i) {
3347 stats[i] = tp->chrono_stat[i - 1];
3348 if (i == tp->chrono_type)
3349 stats[i] += tcp_jiffies32 - tp->chrono_start;
3350 stats[i] *= USEC_PER_SEC / HZ;
3351 total += stats[i];
3352 }
3353
3354 info->tcpi_busy_time = total;
3355 info->tcpi_rwnd_limited = stats[TCP_CHRONO_RWND_LIMITED];
3356 info->tcpi_sndbuf_limited = stats[TCP_CHRONO_SNDBUF_LIMITED];
3357}
3358
3359/* Return information about state of tcp endpoint in API format. */
3360void tcp_get_info(struct sock *sk, struct tcp_info *info)
3361{
3362 const struct tcp_sock *tp = tcp_sk(sk); /* iff sk_type == SOCK_STREAM */
3363 const struct inet_connection_sock *icsk = inet_csk(sk);
3364 unsigned long rate;
3365 u32 now;
3366 u64 rate64;
3367 bool slow;
3368
3369 memset(info, 0, sizeof(*info));
3370 if (sk->sk_type != SOCK_STREAM)
3371 return;
3372
3373 info->tcpi_state = inet_sk_state_load(sk);
3374
3375 /* Report meaningful fields for all TCP states, including listeners */
3376 rate = READ_ONCE(sk->sk_pacing_rate);
3377 rate64 = (rate != ~0UL) ? rate : ~0ULL;
3378 info->tcpi_pacing_rate = rate64;
3379
3380 rate = READ_ONCE(sk->sk_max_pacing_rate);
3381 rate64 = (rate != ~0UL) ? rate : ~0ULL;
3382 info->tcpi_max_pacing_rate = rate64;
3383
3384 info->tcpi_reordering = tp->reordering;
3385 info->tcpi_snd_cwnd = tp->snd_cwnd;
3386
3387 if (info->tcpi_state == TCP_LISTEN) {
3388 /* listeners aliased fields :
3389 * tcpi_unacked -> Number of children ready for accept()
3390 * tcpi_sacked -> max backlog
3391 */
3392 info->tcpi_unacked = READ_ONCE(sk->sk_ack_backlog);
3393 info->tcpi_sacked = READ_ONCE(sk->sk_max_ack_backlog);
3394 return;
3395 }
3396
3397 slow = lock_sock_fast(sk);
3398
3399 info->tcpi_ca_state = icsk->icsk_ca_state;
3400 info->tcpi_retransmits = icsk->icsk_retransmits;
3401 info->tcpi_probes = icsk->icsk_probes_out;
3402 info->tcpi_backoff = icsk->icsk_backoff;
3403
3404 if (tp->rx_opt.tstamp_ok)
3405 info->tcpi_options |= TCPI_OPT_TIMESTAMPS;
3406 if (tcp_is_sack(tp))
3407 info->tcpi_options |= TCPI_OPT_SACK;
3408 if (tp->rx_opt.wscale_ok) {
3409 info->tcpi_options |= TCPI_OPT_WSCALE;
3410 info->tcpi_snd_wscale = tp->rx_opt.snd_wscale;
3411 info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale;
3412 }
3413
3414 if (tp->ecn_flags & TCP_ECN_OK)
3415 info->tcpi_options |= TCPI_OPT_ECN;
3416 if (tp->ecn_flags & TCP_ECN_SEEN)
3417 info->tcpi_options |= TCPI_OPT_ECN_SEEN;
3418 if (tp->syn_data_acked)
3419 info->tcpi_options |= TCPI_OPT_SYN_DATA;
3420
3421 info->tcpi_rto = jiffies_to_usecs(icsk->icsk_rto);
3422 info->tcpi_ato = jiffies_to_usecs(icsk->icsk_ack.ato);
3423 info->tcpi_snd_mss = tp->mss_cache;
3424 info->tcpi_rcv_mss = icsk->icsk_ack.rcv_mss;
3425
3426 info->tcpi_unacked = tp->packets_out;
3427 info->tcpi_sacked = tp->sacked_out;
3428
3429 info->tcpi_lost = tp->lost_out;
3430 info->tcpi_retrans = tp->retrans_out;
3431
3432 now = tcp_jiffies32;
3433 info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime);
3434 info->tcpi_last_data_recv = jiffies_to_msecs(now - icsk->icsk_ack.lrcvtime);
3435 info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp);
3436
3437 info->tcpi_pmtu = icsk->icsk_pmtu_cookie;
3438 info->tcpi_rcv_ssthresh = tp->rcv_ssthresh;
3439 info->tcpi_rtt = tp->srtt_us >> 3;
3440 info->tcpi_rttvar = tp->mdev_us >> 2;
3441 info->tcpi_snd_ssthresh = tp->snd_ssthresh;
3442 info->tcpi_advmss = tp->advmss;
3443
3444 info->tcpi_rcv_rtt = tp->rcv_rtt_est.rtt_us >> 3;
3445 info->tcpi_rcv_space = tp->rcvq_space.space;
3446
3447 info->tcpi_total_retrans = tp->total_retrans;
3448
3449 info->tcpi_bytes_acked = tp->bytes_acked;
3450 info->tcpi_bytes_received = tp->bytes_received;
3451 info->tcpi_notsent_bytes = max_t(int, 0, tp->write_seq - tp->snd_nxt);
3452 tcp_get_info_chrono_stats(tp, info);
3453
3454 info->tcpi_segs_out = tp->segs_out;
3455 info->tcpi_segs_in = tp->segs_in;
3456
3457 info->tcpi_min_rtt = tcp_min_rtt(tp);
3458 info->tcpi_data_segs_in = tp->data_segs_in;
3459 info->tcpi_data_segs_out = tp->data_segs_out;
3460
3461 info->tcpi_delivery_rate_app_limited = tp->rate_app_limited ? 1 : 0;
3462 rate64 = tcp_compute_delivery_rate(tp);
3463 if (rate64)
3464 info->tcpi_delivery_rate = rate64;
3465 info->tcpi_delivered = tp->delivered;
3466 info->tcpi_delivered_ce = tp->delivered_ce;
3467 info->tcpi_bytes_sent = tp->bytes_sent;
3468 info->tcpi_bytes_retrans = tp->bytes_retrans;
3469 info->tcpi_dsack_dups = tp->dsack_dups;
3470 info->tcpi_reord_seen = tp->reord_seen;
3471 info->tcpi_rcv_ooopack = tp->rcv_ooopack;
3472 info->tcpi_snd_wnd = tp->snd_wnd;
3473 info->tcpi_fastopen_client_fail = tp->fastopen_client_fail;
3474 unlock_sock_fast(sk, slow);
3475}
3476EXPORT_SYMBOL_GPL(tcp_get_info);
3477
3478static size_t tcp_opt_stats_get_size(void)
3479{
3480 return
3481 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_BUSY */
3482 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_RWND_LIMITED */
3483 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_SNDBUF_LIMITED */
3484 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_DATA_SEGS_OUT */
3485 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_TOTAL_RETRANS */
3486 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_PACING_RATE */
3487 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_DELIVERY_RATE */
3488 nla_total_size(sizeof(u32)) + /* TCP_NLA_SND_CWND */
3489 nla_total_size(sizeof(u32)) + /* TCP_NLA_REORDERING */
3490 nla_total_size(sizeof(u32)) + /* TCP_NLA_MIN_RTT */
3491 nla_total_size(sizeof(u8)) + /* TCP_NLA_RECUR_RETRANS */
3492 nla_total_size(sizeof(u8)) + /* TCP_NLA_DELIVERY_RATE_APP_LMT */
3493 nla_total_size(sizeof(u32)) + /* TCP_NLA_SNDQ_SIZE */
3494 nla_total_size(sizeof(u8)) + /* TCP_NLA_CA_STATE */
3495 nla_total_size(sizeof(u32)) + /* TCP_NLA_SND_SSTHRESH */
3496 nla_total_size(sizeof(u32)) + /* TCP_NLA_DELIVERED */
3497 nla_total_size(sizeof(u32)) + /* TCP_NLA_DELIVERED_CE */
3498 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_BYTES_SENT */
3499 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_BYTES_RETRANS */
3500 nla_total_size(sizeof(u32)) + /* TCP_NLA_DSACK_DUPS */
3501 nla_total_size(sizeof(u32)) + /* TCP_NLA_REORD_SEEN */
3502 nla_total_size(sizeof(u32)) + /* TCP_NLA_SRTT */
3503 nla_total_size(sizeof(u16)) + /* TCP_NLA_TIMEOUT_REHASH */
3504 nla_total_size(sizeof(u32)) + /* TCP_NLA_BYTES_NOTSENT */
3505 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_EDT */
3506 0;
3507}
3508
3509struct sk_buff *tcp_get_timestamping_opt_stats(const struct sock *sk,
3510 const struct sk_buff *orig_skb)
3511{
3512 const struct tcp_sock *tp = tcp_sk(sk);
3513 struct sk_buff *stats;
3514 struct tcp_info info;
3515 unsigned long rate;
3516 u64 rate64;
3517
3518 stats = alloc_skb(tcp_opt_stats_get_size(), GFP_ATOMIC);
3519 if (!stats)
3520 return NULL;
3521
3522 tcp_get_info_chrono_stats(tp, &info);
3523 nla_put_u64_64bit(stats, TCP_NLA_BUSY,
3524 info.tcpi_busy_time, TCP_NLA_PAD);
3525 nla_put_u64_64bit(stats, TCP_NLA_RWND_LIMITED,
3526 info.tcpi_rwnd_limited, TCP_NLA_PAD);
3527 nla_put_u64_64bit(stats, TCP_NLA_SNDBUF_LIMITED,
3528 info.tcpi_sndbuf_limited, TCP_NLA_PAD);
3529 nla_put_u64_64bit(stats, TCP_NLA_DATA_SEGS_OUT,
3530 tp->data_segs_out, TCP_NLA_PAD);
3531 nla_put_u64_64bit(stats, TCP_NLA_TOTAL_RETRANS,
3532 tp->total_retrans, TCP_NLA_PAD);
3533
3534 rate = READ_ONCE(sk->sk_pacing_rate);
3535 rate64 = (rate != ~0UL) ? rate : ~0ULL;
3536 nla_put_u64_64bit(stats, TCP_NLA_PACING_RATE, rate64, TCP_NLA_PAD);
3537
3538 rate64 = tcp_compute_delivery_rate(tp);
3539 nla_put_u64_64bit(stats, TCP_NLA_DELIVERY_RATE, rate64, TCP_NLA_PAD);
3540
3541 nla_put_u32(stats, TCP_NLA_SND_CWND, tp->snd_cwnd);
3542 nla_put_u32(stats, TCP_NLA_REORDERING, tp->reordering);
3543 nla_put_u32(stats, TCP_NLA_MIN_RTT, tcp_min_rtt(tp));
3544
3545 nla_put_u8(stats, TCP_NLA_RECUR_RETRANS, inet_csk(sk)->icsk_retransmits);
3546 nla_put_u8(stats, TCP_NLA_DELIVERY_RATE_APP_LMT, !!tp->rate_app_limited);
3547 nla_put_u32(stats, TCP_NLA_SND_SSTHRESH, tp->snd_ssthresh);
3548 nla_put_u32(stats, TCP_NLA_DELIVERED, tp->delivered);
3549 nla_put_u32(stats, TCP_NLA_DELIVERED_CE, tp->delivered_ce);
3550
3551 nla_put_u32(stats, TCP_NLA_SNDQ_SIZE, tp->write_seq - tp->snd_una);
3552 nla_put_u8(stats, TCP_NLA_CA_STATE, inet_csk(sk)->icsk_ca_state);
3553
3554 nla_put_u64_64bit(stats, TCP_NLA_BYTES_SENT, tp->bytes_sent,
3555 TCP_NLA_PAD);
3556 nla_put_u64_64bit(stats, TCP_NLA_BYTES_RETRANS, tp->bytes_retrans,
3557 TCP_NLA_PAD);
3558 nla_put_u32(stats, TCP_NLA_DSACK_DUPS, tp->dsack_dups);
3559 nla_put_u32(stats, TCP_NLA_REORD_SEEN, tp->reord_seen);
3560 nla_put_u32(stats, TCP_NLA_SRTT, tp->srtt_us >> 3);
3561 nla_put_u16(stats, TCP_NLA_TIMEOUT_REHASH, tp->timeout_rehash);
3562 nla_put_u32(stats, TCP_NLA_BYTES_NOTSENT,
3563 max_t(int, 0, tp->write_seq - tp->snd_nxt));
3564 nla_put_u64_64bit(stats, TCP_NLA_EDT, orig_skb->skb_mstamp_ns,
3565 TCP_NLA_PAD);
3566
3567 return stats;
3568}
3569
3570static int do_tcp_getsockopt(struct sock *sk, int level,
3571 int optname, char __user *optval, int __user *optlen)
3572{
3573 struct inet_connection_sock *icsk = inet_csk(sk);
3574 struct tcp_sock *tp = tcp_sk(sk);
3575 struct net *net = sock_net(sk);
3576 int val, len;
3577
3578 if (get_user(len, optlen))
3579 return -EFAULT;
3580
3581 len = min_t(unsigned int, len, sizeof(int));
3582
3583 if (len < 0)
3584 return -EINVAL;
3585
3586 switch (optname) {
3587 case TCP_MAXSEG:
3588 val = tp->mss_cache;
3589 if (!val && ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
3590 val = tp->rx_opt.user_mss;
3591 if (tp->repair)
3592 val = tp->rx_opt.mss_clamp;
3593 break;
3594 case TCP_NODELAY:
3595 val = !!(tp->nonagle&TCP_NAGLE_OFF);
3596 break;
3597 case TCP_CORK:
3598 val = !!(tp->nonagle&TCP_NAGLE_CORK);
3599 break;
3600 case TCP_KEEPIDLE:
3601 val = keepalive_time_when(tp) / HZ;
3602 break;
3603 case TCP_KEEPINTVL:
3604 val = keepalive_intvl_when(tp) / HZ;
3605 break;
3606 case TCP_KEEPCNT:
3607 val = keepalive_probes(tp);
3608 break;
3609 case TCP_SYNCNT:
3610 val = icsk->icsk_syn_retries ? : net->ipv4.sysctl_tcp_syn_retries;
3611 break;
3612 case TCP_LINGER2:
3613 val = tp->linger2;
3614 if (val >= 0)
3615 val = (val ? : net->ipv4.sysctl_tcp_fin_timeout) / HZ;
3616 break;
3617 case TCP_DEFER_ACCEPT:
3618 val = retrans_to_secs(icsk->icsk_accept_queue.rskq_defer_accept,
3619 TCP_TIMEOUT_INIT / HZ, TCP_RTO_MAX / HZ);
3620 break;
3621 case TCP_WINDOW_CLAMP:
3622 val = tp->window_clamp;
3623 break;
3624 case TCP_INFO: {
3625 struct tcp_info info;
3626
3627 if (get_user(len, optlen))
3628 return -EFAULT;
3629
3630 tcp_get_info(sk, &info);
3631
3632 len = min_t(unsigned int, len, sizeof(info));
3633 if (put_user(len, optlen))
3634 return -EFAULT;
3635 if (copy_to_user(optval, &info, len))
3636 return -EFAULT;
3637 return 0;
3638 }
3639 case TCP_CC_INFO: {
3640 const struct tcp_congestion_ops *ca_ops;
3641 union tcp_cc_info info;
3642 size_t sz = 0;
3643 int attr;
3644
3645 if (get_user(len, optlen))
3646 return -EFAULT;
3647
3648 ca_ops = icsk->icsk_ca_ops;
3649 if (ca_ops && ca_ops->get_info)
3650 sz = ca_ops->get_info(sk, ~0U, &attr, &info);
3651
3652 len = min_t(unsigned int, len, sz);
3653 if (put_user(len, optlen))
3654 return -EFAULT;
3655 if (copy_to_user(optval, &info, len))
3656 return -EFAULT;
3657 return 0;
3658 }
3659 case TCP_QUICKACK:
3660 val = !inet_csk_in_pingpong_mode(sk);
3661 break;
3662
3663 case TCP_CONGESTION:
3664 if (get_user(len, optlen))
3665 return -EFAULT;
3666 len = min_t(unsigned int, len, TCP_CA_NAME_MAX);
3667 if (put_user(len, optlen))
3668 return -EFAULT;
3669 if (copy_to_user(optval, icsk->icsk_ca_ops->name, len))
3670 return -EFAULT;
3671 return 0;
3672
3673 case TCP_ULP:
3674 if (get_user(len, optlen))
3675 return -EFAULT;
3676 len = min_t(unsigned int, len, TCP_ULP_NAME_MAX);
3677 if (!icsk->icsk_ulp_ops) {
3678 if (put_user(0, optlen))
3679 return -EFAULT;
3680 return 0;
3681 }
3682 if (put_user(len, optlen))
3683 return -EFAULT;
3684 if (copy_to_user(optval, icsk->icsk_ulp_ops->name, len))
3685 return -EFAULT;
3686 return 0;
3687
3688 case TCP_FASTOPEN_KEY: {
3689 u64 key[TCP_FASTOPEN_KEY_BUF_LENGTH / sizeof(u64)];
3690 unsigned int key_len;
3691
3692 if (get_user(len, optlen))
3693 return -EFAULT;
3694
3695 key_len = tcp_fastopen_get_cipher(net, icsk, key) *
3696 TCP_FASTOPEN_KEY_LENGTH;
3697 len = min_t(unsigned int, len, key_len);
3698 if (put_user(len, optlen))
3699 return -EFAULT;
3700 if (copy_to_user(optval, key, len))
3701 return -EFAULT;
3702 return 0;
3703 }
3704 case TCP_THIN_LINEAR_TIMEOUTS:
3705 val = tp->thin_lto;
3706 break;
3707
3708 case TCP_THIN_DUPACK:
3709 val = 0;
3710 break;
3711
3712 case TCP_REPAIR:
3713 val = tp->repair;
3714 break;
3715
3716 case TCP_REPAIR_QUEUE:
3717 if (tp->repair)
3718 val = tp->repair_queue;
3719 else
3720 return -EINVAL;
3721 break;
3722
3723 case TCP_REPAIR_WINDOW: {
3724 struct tcp_repair_window opt;
3725
3726 if (get_user(len, optlen))
3727 return -EFAULT;
3728
3729 if (len != sizeof(opt))
3730 return -EINVAL;
3731
3732 if (!tp->repair)
3733 return -EPERM;
3734
3735 opt.snd_wl1 = tp->snd_wl1;
3736 opt.snd_wnd = tp->snd_wnd;
3737 opt.max_window = tp->max_window;
3738 opt.rcv_wnd = tp->rcv_wnd;
3739 opt.rcv_wup = tp->rcv_wup;
3740
3741 if (copy_to_user(optval, &opt, len))
3742 return -EFAULT;
3743 return 0;
3744 }
3745 case TCP_QUEUE_SEQ:
3746 if (tp->repair_queue == TCP_SEND_QUEUE)
3747 val = tp->write_seq;
3748 else if (tp->repair_queue == TCP_RECV_QUEUE)
3749 val = tp->rcv_nxt;
3750 else
3751 return -EINVAL;
3752 break;
3753
3754 case TCP_USER_TIMEOUT:
3755 val = icsk->icsk_user_timeout;
3756 break;
3757
3758 case TCP_FASTOPEN:
3759 val = icsk->icsk_accept_queue.fastopenq.max_qlen;
3760 break;
3761
3762 case TCP_FASTOPEN_CONNECT:
3763 val = tp->fastopen_connect;
3764 break;
3765
3766 case TCP_FASTOPEN_NO_COOKIE:
3767 val = tp->fastopen_no_cookie;
3768 break;
3769
3770 case TCP_TX_DELAY:
3771 val = tp->tcp_tx_delay;
3772 break;
3773
3774 case TCP_TIMESTAMP:
3775 val = tcp_time_stamp_raw() + tp->tsoffset;
3776 break;
3777 case TCP_NOTSENT_LOWAT:
3778 val = tp->notsent_lowat;
3779 break;
3780 case TCP_INQ:
3781 val = tp->recvmsg_inq;
3782 break;
3783 case TCP_SAVE_SYN:
3784 val = tp->save_syn;
3785 break;
3786 case TCP_SAVED_SYN: {
3787 if (get_user(len, optlen))
3788 return -EFAULT;
3789
3790 lock_sock(sk);
3791 if (tp->saved_syn) {
3792 if (len < tp->saved_syn[0]) {
3793 if (put_user(tp->saved_syn[0], optlen)) {
3794 release_sock(sk);
3795 return -EFAULT;
3796 }
3797 release_sock(sk);
3798 return -EINVAL;
3799 }
3800 len = tp->saved_syn[0];
3801 if (put_user(len, optlen)) {
3802 release_sock(sk);
3803 return -EFAULT;
3804 }
3805 if (copy_to_user(optval, tp->saved_syn + 1, len)) {
3806 release_sock(sk);
3807 return -EFAULT;
3808 }
3809 tcp_saved_syn_free(tp);
3810 release_sock(sk);
3811 } else {
3812 release_sock(sk);
3813 len = 0;
3814 if (put_user(len, optlen))
3815 return -EFAULT;
3816 }
3817 return 0;
3818 }
3819#ifdef CONFIG_MMU
3820 case TCP_ZEROCOPY_RECEIVE: {
3821 struct tcp_zerocopy_receive zc;
3822 int err;
3823
3824 if (get_user(len, optlen))
3825 return -EFAULT;
3826 if (len < offsetofend(struct tcp_zerocopy_receive, length))
3827 return -EINVAL;
3828 if (len > sizeof(zc)) {
3829 len = sizeof(zc);
3830 if (put_user(len, optlen))
3831 return -EFAULT;
3832 }
3833 if (copy_from_user(&zc, optval, len))
3834 return -EFAULT;
3835 lock_sock(sk);
3836 err = tcp_zerocopy_receive(sk, &zc);
3837 release_sock(sk);
3838 if (len == sizeof(zc))
3839 goto zerocopy_rcv_sk_err;
3840 switch (len) {
3841 case offsetofend(struct tcp_zerocopy_receive, err):
3842 goto zerocopy_rcv_sk_err;
3843 case offsetofend(struct tcp_zerocopy_receive, inq):
3844 goto zerocopy_rcv_inq;
3845 case offsetofend(struct tcp_zerocopy_receive, length):
3846 default:
3847 goto zerocopy_rcv_out;
3848 }
3849zerocopy_rcv_sk_err:
3850 if (!err)
3851 zc.err = sock_error(sk);
3852zerocopy_rcv_inq:
3853 zc.inq = tcp_inq_hint(sk);
3854zerocopy_rcv_out:
3855 if (!err && copy_to_user(optval, &zc, len))
3856 err = -EFAULT;
3857 return err;
3858 }
3859#endif
3860 default:
3861 return -ENOPROTOOPT;
3862 }
3863
3864 if (put_user(len, optlen))
3865 return -EFAULT;
3866 if (copy_to_user(optval, &val, len))
3867 return -EFAULT;
3868 return 0;
3869}
3870
3871int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval,
3872 int __user *optlen)
3873{
3874 struct inet_connection_sock *icsk = inet_csk(sk);
3875
3876 if (level != SOL_TCP)
3877 return icsk->icsk_af_ops->getsockopt(sk, level, optname,
3878 optval, optlen);
3879 return do_tcp_getsockopt(sk, level, optname, optval, optlen);
3880}
3881EXPORT_SYMBOL(tcp_getsockopt);
3882
3883#ifdef CONFIG_TCP_MD5SIG
3884static DEFINE_PER_CPU(struct tcp_md5sig_pool, tcp_md5sig_pool);
3885static DEFINE_MUTEX(tcp_md5sig_mutex);
3886static bool tcp_md5sig_pool_populated = false;
3887
3888static void __tcp_alloc_md5sig_pool(void)
3889{
3890 struct crypto_ahash *hash;
3891 int cpu;
3892
3893 hash = crypto_alloc_ahash("md5", 0, CRYPTO_ALG_ASYNC);
3894 if (IS_ERR(hash))
3895 return;
3896
3897 for_each_possible_cpu(cpu) {
3898 void *scratch = per_cpu(tcp_md5sig_pool, cpu).scratch;
3899 struct ahash_request *req;
3900
3901 if (!scratch) {
3902 scratch = kmalloc_node(sizeof(union tcp_md5sum_block) +
3903 sizeof(struct tcphdr),
3904 GFP_KERNEL,
3905 cpu_to_node(cpu));
3906 if (!scratch)
3907 return;
3908 per_cpu(tcp_md5sig_pool, cpu).scratch = scratch;
3909 }
3910 if (per_cpu(tcp_md5sig_pool, cpu).md5_req)
3911 continue;
3912
3913 req = ahash_request_alloc(hash, GFP_KERNEL);
3914 if (!req)
3915 return;
3916
3917 ahash_request_set_callback(req, 0, NULL, NULL);
3918
3919 per_cpu(tcp_md5sig_pool, cpu).md5_req = req;
3920 }
3921 /* before setting tcp_md5sig_pool_populated, we must commit all writes
3922 * to memory. See smp_rmb() in tcp_get_md5sig_pool()
3923 */
3924 smp_wmb();
3925 tcp_md5sig_pool_populated = true;
3926}
3927
3928bool tcp_alloc_md5sig_pool(void)
3929{
3930 if (unlikely(!tcp_md5sig_pool_populated)) {
3931 mutex_lock(&tcp_md5sig_mutex);
3932
3933 if (!tcp_md5sig_pool_populated) {
3934 __tcp_alloc_md5sig_pool();
3935 if (tcp_md5sig_pool_populated)
3936 static_branch_inc(&tcp_md5_needed);
3937 }
3938
3939 mutex_unlock(&tcp_md5sig_mutex);
3940 }
3941 return tcp_md5sig_pool_populated;
3942}
3943EXPORT_SYMBOL(tcp_alloc_md5sig_pool);
3944
3945
3946/**
3947 * tcp_get_md5sig_pool - get md5sig_pool for this user
3948 *
3949 * We use percpu structure, so if we succeed, we exit with preemption
3950 * and BH disabled, to make sure another thread or softirq handling
3951 * wont try to get same context.
3952 */
3953struct tcp_md5sig_pool *tcp_get_md5sig_pool(void)
3954{
3955 local_bh_disable();
3956
3957 if (tcp_md5sig_pool_populated) {
3958 /* coupled with smp_wmb() in __tcp_alloc_md5sig_pool() */
3959 smp_rmb();
3960 return this_cpu_ptr(&tcp_md5sig_pool);
3961 }
3962 local_bh_enable();
3963 return NULL;
3964}
3965EXPORT_SYMBOL(tcp_get_md5sig_pool);
3966
3967int tcp_md5_hash_skb_data(struct tcp_md5sig_pool *hp,
3968 const struct sk_buff *skb, unsigned int header_len)
3969{
3970 struct scatterlist sg;
3971 const struct tcphdr *tp = tcp_hdr(skb);
3972 struct ahash_request *req = hp->md5_req;
3973 unsigned int i;
3974 const unsigned int head_data_len = skb_headlen(skb) > header_len ?
3975 skb_headlen(skb) - header_len : 0;
3976 const struct skb_shared_info *shi = skb_shinfo(skb);
3977 struct sk_buff *frag_iter;
3978
3979 sg_init_table(&sg, 1);
3980
3981 sg_set_buf(&sg, ((u8 *) tp) + header_len, head_data_len);
3982 ahash_request_set_crypt(req, &sg, NULL, head_data_len);
3983 if (crypto_ahash_update(req))
3984 return 1;
3985
3986 for (i = 0; i < shi->nr_frags; ++i) {
3987 const skb_frag_t *f = &shi->frags[i];
3988 unsigned int offset = skb_frag_off(f);
3989 struct page *page = skb_frag_page(f) + (offset >> PAGE_SHIFT);
3990
3991 sg_set_page(&sg, page, skb_frag_size(f),
3992 offset_in_page(offset));
3993 ahash_request_set_crypt(req, &sg, NULL, skb_frag_size(f));
3994 if (crypto_ahash_update(req))
3995 return 1;
3996 }
3997
3998 skb_walk_frags(skb, frag_iter)
3999 if (tcp_md5_hash_skb_data(hp, frag_iter, 0))
4000 return 1;
4001
4002 return 0;
4003}
4004EXPORT_SYMBOL(tcp_md5_hash_skb_data);
4005
4006int tcp_md5_hash_key(struct tcp_md5sig_pool *hp, const struct tcp_md5sig_key *key)
4007{
4008 u8 keylen = READ_ONCE(key->keylen); /* paired with WRITE_ONCE() in tcp_md5_do_add */
4009 struct scatterlist sg;
4010
4011 sg_init_one(&sg, key->key, keylen);
4012 ahash_request_set_crypt(hp->md5_req, &sg, NULL, keylen);
4013
4014 /* We use data_race() because tcp_md5_do_add() might change key->key under us */
4015 return data_race(crypto_ahash_update(hp->md5_req));
4016}
4017EXPORT_SYMBOL(tcp_md5_hash_key);
4018
4019#endif
4020
4021void tcp_done(struct sock *sk)
4022{
4023 struct request_sock *req;
4024
4025 /* We might be called with a new socket, after
4026 * inet_csk_prepare_forced_close() has been called
4027 * so we can not use lockdep_sock_is_held(sk)
4028 */
4029 req = rcu_dereference_protected(tcp_sk(sk)->fastopen_rsk, 1);
4030
4031 if (sk->sk_state == TCP_SYN_SENT || sk->sk_state == TCP_SYN_RECV)
4032 TCP_INC_STATS(sock_net(sk), TCP_MIB_ATTEMPTFAILS);
4033
4034 tcp_set_state(sk, TCP_CLOSE);
4035 tcp_clear_xmit_timers(sk);
4036 if (req)
4037 reqsk_fastopen_remove(sk, req, false);
4038
4039 sk->sk_shutdown = SHUTDOWN_MASK;
4040
4041 if (!sock_flag(sk, SOCK_DEAD))
4042 sk->sk_state_change(sk);
4043 else
4044 inet_csk_destroy_sock(sk);
4045}
4046EXPORT_SYMBOL_GPL(tcp_done);
4047
4048int tcp_abort(struct sock *sk, int err)
4049{
4050 if (!sk_fullsock(sk)) {
4051 if (sk->sk_state == TCP_NEW_SYN_RECV) {
4052 struct request_sock *req = inet_reqsk(sk);
4053
4054 local_bh_disable();
4055 inet_csk_reqsk_queue_drop(req->rsk_listener, req);
4056 local_bh_enable();
4057 return 0;
4058 }
4059 return -EOPNOTSUPP;
4060 }
4061
4062 /* Don't race with userspace socket closes such as tcp_close. */
4063 lock_sock(sk);
4064
4065 if (sk->sk_state == TCP_LISTEN) {
4066 tcp_set_state(sk, TCP_CLOSE);
4067 inet_csk_listen_stop(sk);
4068 }
4069
4070 /* Don't race with BH socket closes such as inet_csk_listen_stop. */
4071 local_bh_disable();
4072 bh_lock_sock(sk);
4073
4074 if (!sock_flag(sk, SOCK_DEAD)) {
4075 sk->sk_err = err;
4076 /* This barrier is coupled with smp_rmb() in tcp_poll() */
4077 smp_wmb();
4078 sk->sk_error_report(sk);
4079 if (tcp_need_reset(sk->sk_state))
4080 tcp_send_active_reset(sk, GFP_ATOMIC);
4081 tcp_done(sk);
4082 }
4083
4084 bh_unlock_sock(sk);
4085 local_bh_enable();
4086 tcp_write_queue_purge(sk);
4087 release_sock(sk);
4088 return 0;
4089}
4090EXPORT_SYMBOL_GPL(tcp_abort);
4091
4092extern struct tcp_congestion_ops tcp_reno;
4093
4094static __initdata unsigned long thash_entries;
4095static int __init set_thash_entries(char *str)
4096{
4097 ssize_t ret;
4098
4099 if (!str)
4100 return 0;
4101
4102 ret = kstrtoul(str, 0, &thash_entries);
4103 if (ret)
4104 return 0;
4105
4106 return 1;
4107}
4108__setup("thash_entries=", set_thash_entries);
4109
4110static void __init tcp_init_mem(void)
4111{
4112 unsigned long limit = nr_free_buffer_pages() / 16;
4113
4114 limit = max(limit, 128UL);
4115 sysctl_tcp_mem[0] = limit / 4 * 3; /* 4.68 % */
4116 sysctl_tcp_mem[1] = limit; /* 6.25 % */
4117 sysctl_tcp_mem[2] = sysctl_tcp_mem[0] * 2; /* 9.37 % */
4118}
4119
4120void __init tcp_init(void)
4121{
4122 int max_rshare, max_wshare, cnt;
4123 unsigned long limit;
4124 unsigned int i;
4125
4126 BUILD_BUG_ON(TCP_MIN_SND_MSS <= MAX_TCP_OPTION_SPACE);
4127 BUILD_BUG_ON(sizeof(struct tcp_skb_cb) >
4128 sizeof_field(struct sk_buff, cb));
4129
4130 percpu_counter_init(&tcp_sockets_allocated, 0, GFP_KERNEL);
4131 percpu_counter_init(&tcp_orphan_count, 0, GFP_KERNEL);
4132 inet_hashinfo_init(&tcp_hashinfo);
4133 inet_hashinfo2_init(&tcp_hashinfo, "tcp_listen_portaddr_hash",
4134 thash_entries, 21, /* one slot per 2 MB*/
4135 0, 64 * 1024);
4136 tcp_hashinfo.bind_bucket_cachep =
4137 kmem_cache_create("tcp_bind_bucket",
4138 sizeof(struct inet_bind_bucket), 0,
4139 SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL);
4140
4141 /* Size and allocate the main established and bind bucket
4142 * hash tables.
4143 *
4144 * The methodology is similar to that of the buffer cache.
4145 */
4146 tcp_hashinfo.ehash =
4147 alloc_large_system_hash("TCP established",
4148 sizeof(struct inet_ehash_bucket),
4149 thash_entries,
4150 17, /* one slot per 128 KB of memory */
4151 0,
4152 NULL,
4153 &tcp_hashinfo.ehash_mask,
4154 0,
4155 thash_entries ? 0 : 512 * 1024);
4156 for (i = 0; i <= tcp_hashinfo.ehash_mask; i++)
4157 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].chain, i);
4158
4159 if (inet_ehash_locks_alloc(&tcp_hashinfo))
4160 panic("TCP: failed to alloc ehash_locks");
4161 tcp_hashinfo.bhash =
4162 alloc_large_system_hash("TCP bind",
4163 sizeof(struct inet_bind_hashbucket),
4164 tcp_hashinfo.ehash_mask + 1,
4165 17, /* one slot per 128 KB of memory */
4166 0,
4167 &tcp_hashinfo.bhash_size,
4168 NULL,
4169 0,
4170 64 * 1024);
4171 tcp_hashinfo.bhash_size = 1U << tcp_hashinfo.bhash_size;
4172 for (i = 0; i < tcp_hashinfo.bhash_size; i++) {
4173 spin_lock_init(&tcp_hashinfo.bhash[i].lock);
4174 INIT_HLIST_HEAD(&tcp_hashinfo.bhash[i].chain);
4175 }
4176
4177
4178 cnt = tcp_hashinfo.ehash_mask + 1;
4179 sysctl_tcp_max_orphans = cnt / 2;
4180
4181 tcp_init_mem();
4182 /* Set per-socket limits to no more than 1/128 the pressure threshold */
4183 limit = nr_free_buffer_pages() << (PAGE_SHIFT - 7);
4184 max_wshare = min(4UL*1024*1024, limit);
4185 max_rshare = min(6UL*1024*1024, limit);
4186
4187 init_net.ipv4.sysctl_tcp_wmem[0] = SK_MEM_QUANTUM;
4188 init_net.ipv4.sysctl_tcp_wmem[1] = 16*1024;
4189 init_net.ipv4.sysctl_tcp_wmem[2] = max(64*1024, max_wshare);
4190
4191 init_net.ipv4.sysctl_tcp_rmem[0] = SK_MEM_QUANTUM;
4192 init_net.ipv4.sysctl_tcp_rmem[1] = 131072;
4193 init_net.ipv4.sysctl_tcp_rmem[2] = max(131072, max_rshare);
4194
4195 pr_info("Hash tables configured (established %u bind %u)\n",
4196 tcp_hashinfo.ehash_mask + 1, tcp_hashinfo.bhash_size);
4197
4198 tcp_v4_init();
4199 tcp_metrics_init();
4200 BUG_ON(tcp_register_congestion_control(&tcp_reno) != 0);
4201 tcp_tasklet_init();
4202 mptcp_init();
4203}