Loading...
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}
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/cache.h>
264#include <linux/err.h>
265#include <linux/time.h>
266#include <linux/slab.h>
267#include <linux/errqueue.h>
268#include <linux/static_key.h>
269#include <linux/btf.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
283/* Track pending CMSGs. */
284enum {
285 TCP_CMSG_INQ = 1,
286 TCP_CMSG_TS = 2
287};
288
289DEFINE_PER_CPU(unsigned int, tcp_orphan_count);
290EXPORT_PER_CPU_SYMBOL_GPL(tcp_orphan_count);
291
292long sysctl_tcp_mem[3] __read_mostly;
293EXPORT_SYMBOL(sysctl_tcp_mem);
294
295atomic_long_t tcp_memory_allocated ____cacheline_aligned_in_smp; /* Current allocated memory. */
296EXPORT_SYMBOL(tcp_memory_allocated);
297DEFINE_PER_CPU(int, tcp_memory_per_cpu_fw_alloc);
298EXPORT_PER_CPU_SYMBOL_GPL(tcp_memory_per_cpu_fw_alloc);
299
300#if IS_ENABLED(CONFIG_SMC)
301DEFINE_STATIC_KEY_FALSE(tcp_have_smc);
302EXPORT_SYMBOL(tcp_have_smc);
303#endif
304
305/*
306 * Current number of TCP sockets.
307 */
308struct percpu_counter tcp_sockets_allocated ____cacheline_aligned_in_smp;
309EXPORT_SYMBOL(tcp_sockets_allocated);
310
311/*
312 * TCP splice context
313 */
314struct tcp_splice_state {
315 struct pipe_inode_info *pipe;
316 size_t len;
317 unsigned int flags;
318};
319
320/*
321 * Pressure flag: try to collapse.
322 * Technical note: it is used by multiple contexts non atomically.
323 * All the __sk_mem_schedule() is of this nature: accounting
324 * is strict, actions are advisory and have some latency.
325 */
326unsigned long tcp_memory_pressure __read_mostly;
327EXPORT_SYMBOL_GPL(tcp_memory_pressure);
328
329void tcp_enter_memory_pressure(struct sock *sk)
330{
331 unsigned long val;
332
333 if (READ_ONCE(tcp_memory_pressure))
334 return;
335 val = jiffies;
336
337 if (!val)
338 val--;
339 if (!cmpxchg(&tcp_memory_pressure, 0, val))
340 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMEMORYPRESSURES);
341}
342EXPORT_SYMBOL_GPL(tcp_enter_memory_pressure);
343
344void tcp_leave_memory_pressure(struct sock *sk)
345{
346 unsigned long val;
347
348 if (!READ_ONCE(tcp_memory_pressure))
349 return;
350 val = xchg(&tcp_memory_pressure, 0);
351 if (val)
352 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPMEMORYPRESSURESCHRONO,
353 jiffies_to_msecs(jiffies - val));
354}
355EXPORT_SYMBOL_GPL(tcp_leave_memory_pressure);
356
357/* Convert seconds to retransmits based on initial and max timeout */
358static u8 secs_to_retrans(int seconds, int timeout, int rto_max)
359{
360 u8 res = 0;
361
362 if (seconds > 0) {
363 int period = timeout;
364
365 res = 1;
366 while (seconds > period && res < 255) {
367 res++;
368 timeout <<= 1;
369 if (timeout > rto_max)
370 timeout = rto_max;
371 period += timeout;
372 }
373 }
374 return res;
375}
376
377/* Convert retransmits to seconds based on initial and max timeout */
378static int retrans_to_secs(u8 retrans, int timeout, int rto_max)
379{
380 int period = 0;
381
382 if (retrans > 0) {
383 period = timeout;
384 while (--retrans) {
385 timeout <<= 1;
386 if (timeout > rto_max)
387 timeout = rto_max;
388 period += timeout;
389 }
390 }
391 return period;
392}
393
394static u64 tcp_compute_delivery_rate(const struct tcp_sock *tp)
395{
396 u32 rate = READ_ONCE(tp->rate_delivered);
397 u32 intv = READ_ONCE(tp->rate_interval_us);
398 u64 rate64 = 0;
399
400 if (rate && intv) {
401 rate64 = (u64)rate * tp->mss_cache * USEC_PER_SEC;
402 do_div(rate64, intv);
403 }
404 return rate64;
405}
406
407/* Address-family independent initialization for a tcp_sock.
408 *
409 * NOTE: A lot of things set to zero explicitly by call to
410 * sk_alloc() so need not be done here.
411 */
412void tcp_init_sock(struct sock *sk)
413{
414 struct inet_connection_sock *icsk = inet_csk(sk);
415 struct tcp_sock *tp = tcp_sk(sk);
416
417 tp->out_of_order_queue = RB_ROOT;
418 sk->tcp_rtx_queue = RB_ROOT;
419 tcp_init_xmit_timers(sk);
420 INIT_LIST_HEAD(&tp->tsq_node);
421 INIT_LIST_HEAD(&tp->tsorted_sent_queue);
422
423 icsk->icsk_rto = TCP_TIMEOUT_INIT;
424 icsk->icsk_rto_min = TCP_RTO_MIN;
425 icsk->icsk_delack_max = TCP_DELACK_MAX;
426 tp->mdev_us = jiffies_to_usecs(TCP_TIMEOUT_INIT);
427 minmax_reset(&tp->rtt_min, tcp_jiffies32, ~0U);
428
429 /* So many TCP implementations out there (incorrectly) count the
430 * initial SYN frame in their delayed-ACK and congestion control
431 * algorithms that we must have the following bandaid to talk
432 * efficiently to them. -DaveM
433 */
434 tcp_snd_cwnd_set(tp, TCP_INIT_CWND);
435
436 /* There's a bubble in the pipe until at least the first ACK. */
437 tp->app_limited = ~0U;
438 tp->rate_app_limited = 1;
439
440 /* See draft-stevens-tcpca-spec-01 for discussion of the
441 * initialization of these values.
442 */
443 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
444 tp->snd_cwnd_clamp = ~0;
445 tp->mss_cache = TCP_MSS_DEFAULT;
446
447 tp->reordering = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_reordering);
448 tcp_assign_congestion_control(sk);
449
450 tp->tsoffset = 0;
451 tp->rack.reo_wnd_steps = 1;
452
453 sk->sk_write_space = sk_stream_write_space;
454 sock_set_flag(sk, SOCK_USE_WRITE_QUEUE);
455
456 icsk->icsk_sync_mss = tcp_sync_mss;
457
458 WRITE_ONCE(sk->sk_sndbuf, READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_wmem[1]));
459 WRITE_ONCE(sk->sk_rcvbuf, READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_rmem[1]));
460 tcp_scaling_ratio_init(sk);
461
462 set_bit(SOCK_SUPPORT_ZC, &sk->sk_socket->flags);
463 sk_sockets_allocated_inc(sk);
464}
465EXPORT_SYMBOL(tcp_init_sock);
466
467static void tcp_tx_timestamp(struct sock *sk, u16 tsflags)
468{
469 struct sk_buff *skb = tcp_write_queue_tail(sk);
470
471 if (tsflags && skb) {
472 struct skb_shared_info *shinfo = skb_shinfo(skb);
473 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
474
475 sock_tx_timestamp(sk, tsflags, &shinfo->tx_flags);
476 if (tsflags & SOF_TIMESTAMPING_TX_ACK)
477 tcb->txstamp_ack = 1;
478 if (tsflags & SOF_TIMESTAMPING_TX_RECORD_MASK)
479 shinfo->tskey = TCP_SKB_CB(skb)->seq + skb->len - 1;
480 }
481}
482
483static bool tcp_stream_is_readable(struct sock *sk, int target)
484{
485 if (tcp_epollin_ready(sk, target))
486 return true;
487 return sk_is_readable(sk);
488}
489
490/*
491 * Wait for a TCP event.
492 *
493 * Note that we don't need to lock the socket, as the upper poll layers
494 * take care of normal races (between the test and the event) and we don't
495 * go look at any of the socket buffers directly.
496 */
497__poll_t tcp_poll(struct file *file, struct socket *sock, poll_table *wait)
498{
499 __poll_t mask;
500 struct sock *sk = sock->sk;
501 const struct tcp_sock *tp = tcp_sk(sk);
502 u8 shutdown;
503 int state;
504
505 sock_poll_wait(file, sock, wait);
506
507 state = inet_sk_state_load(sk);
508 if (state == TCP_LISTEN)
509 return inet_csk_listen_poll(sk);
510
511 /* Socket is not locked. We are protected from async events
512 * by poll logic and correct handling of state changes
513 * made by other threads is impossible in any case.
514 */
515
516 mask = 0;
517
518 /*
519 * EPOLLHUP is certainly not done right. But poll() doesn't
520 * have a notion of HUP in just one direction, and for a
521 * socket the read side is more interesting.
522 *
523 * Some poll() documentation says that EPOLLHUP is incompatible
524 * with the EPOLLOUT/POLLWR flags, so somebody should check this
525 * all. But careful, it tends to be safer to return too many
526 * bits than too few, and you can easily break real applications
527 * if you don't tell them that something has hung up!
528 *
529 * Check-me.
530 *
531 * Check number 1. EPOLLHUP is _UNMASKABLE_ event (see UNIX98 and
532 * our fs/select.c). It means that after we received EOF,
533 * poll always returns immediately, making impossible poll() on write()
534 * in state CLOSE_WAIT. One solution is evident --- to set EPOLLHUP
535 * if and only if shutdown has been made in both directions.
536 * Actually, it is interesting to look how Solaris and DUX
537 * solve this dilemma. I would prefer, if EPOLLHUP were maskable,
538 * then we could set it on SND_SHUTDOWN. BTW examples given
539 * in Stevens' books assume exactly this behaviour, it explains
540 * why EPOLLHUP is incompatible with EPOLLOUT. --ANK
541 *
542 * NOTE. Check for TCP_CLOSE is added. The goal is to prevent
543 * blocking on fresh not-connected or disconnected socket. --ANK
544 */
545 shutdown = READ_ONCE(sk->sk_shutdown);
546 if (shutdown == SHUTDOWN_MASK || state == TCP_CLOSE)
547 mask |= EPOLLHUP;
548 if (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 u16 urg_data = READ_ONCE(tp->urg_data);
556
557 if (unlikely(urg_data) &&
558 READ_ONCE(tp->urg_seq) == READ_ONCE(tp->copied_seq) &&
559 !sock_flag(sk, SOCK_URGINLINE))
560 target++;
561
562 if (tcp_stream_is_readable(sk, target))
563 mask |= EPOLLIN | EPOLLRDNORM;
564
565 if (!(shutdown & SEND_SHUTDOWN)) {
566 if (__sk_stream_is_writeable(sk, 1)) {
567 mask |= EPOLLOUT | EPOLLWRNORM;
568 } else { /* send SIGIO later */
569 sk_set_bit(SOCKWQ_ASYNC_NOSPACE, sk);
570 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
571
572 /* Race breaker. If space is freed after
573 * wspace test but before the flags are set,
574 * IO signal will be lost. Memory barrier
575 * pairs with the input side.
576 */
577 smp_mb__after_atomic();
578 if (__sk_stream_is_writeable(sk, 1))
579 mask |= EPOLLOUT | EPOLLWRNORM;
580 }
581 } else
582 mask |= EPOLLOUT | EPOLLWRNORM;
583
584 if (urg_data & TCP_URG_VALID)
585 mask |= EPOLLPRI;
586 } else if (state == TCP_SYN_SENT &&
587 inet_test_bit(DEFER_CONNECT, sk)) {
588 /* Active TCP fastopen socket with defer_connect
589 * Return EPOLLOUT so application can call write()
590 * in order for kernel to generate SYN+data
591 */
592 mask |= EPOLLOUT | EPOLLWRNORM;
593 }
594 /* This barrier is coupled with smp_wmb() in tcp_reset() */
595 smp_rmb();
596 if (READ_ONCE(sk->sk_err) ||
597 !skb_queue_empty_lockless(&sk->sk_error_queue))
598 mask |= EPOLLERR;
599
600 return mask;
601}
602EXPORT_SYMBOL(tcp_poll);
603
604int tcp_ioctl(struct sock *sk, int cmd, int *karg)
605{
606 struct tcp_sock *tp = tcp_sk(sk);
607 int answ;
608 bool slow;
609
610 switch (cmd) {
611 case SIOCINQ:
612 if (sk->sk_state == TCP_LISTEN)
613 return -EINVAL;
614
615 slow = lock_sock_fast(sk);
616 answ = tcp_inq(sk);
617 unlock_sock_fast(sk, slow);
618 break;
619 case SIOCATMARK:
620 answ = READ_ONCE(tp->urg_data) &&
621 READ_ONCE(tp->urg_seq) == READ_ONCE(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 = READ_ONCE(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 = READ_ONCE(tp->write_seq) -
640 READ_ONCE(tp->snd_nxt);
641 break;
642 default:
643 return -ENOIOCTLCMD;
644 }
645
646 *karg = answ;
647 return 0;
648}
649EXPORT_SYMBOL(tcp_ioctl);
650
651void tcp_mark_push(struct tcp_sock *tp, struct sk_buff *skb)
652{
653 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
654 tp->pushed_seq = tp->write_seq;
655}
656
657static inline bool forced_push(const struct tcp_sock *tp)
658{
659 return after(tp->write_seq, tp->pushed_seq + (tp->max_window >> 1));
660}
661
662void tcp_skb_entail(struct sock *sk, struct sk_buff *skb)
663{
664 struct tcp_sock *tp = tcp_sk(sk);
665 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
666
667 tcb->seq = tcb->end_seq = tp->write_seq;
668 tcb->tcp_flags = TCPHDR_ACK;
669 __skb_header_release(skb);
670 tcp_add_write_queue_tail(sk, skb);
671 sk_wmem_queued_add(sk, 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 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_autocorking) &&
700 !tcp_rtx_queue_empty(sk) &&
701 refcount_read(&sk->sk_wmem_alloc) > skb->truesize &&
702 tcp_skb_can_collapse_to(skb);
703}
704
705void tcp_push(struct sock *sk, int flags, int mss_now,
706 int nonagle, int size_goal)
707{
708 struct tcp_sock *tp = tcp_sk(sk);
709 struct sk_buff *skb;
710
711 skb = tcp_write_queue_tail(sk);
712 if (!skb)
713 return;
714 if (!(flags & MSG_MORE) || forced_push(tp))
715 tcp_mark_push(tp, skb);
716
717 tcp_mark_urg(tp, flags);
718
719 if (tcp_should_autocork(sk, skb, size_goal)) {
720
721 /* avoid atomic op if TSQ_THROTTLED bit is already set */
722 if (!test_bit(TSQ_THROTTLED, &sk->sk_tsq_flags)) {
723 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPAUTOCORKING);
724 set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags);
725 smp_mb__after_atomic();
726 }
727 /* It is possible TX completion already happened
728 * before we set TSQ_THROTTLED.
729 */
730 if (refcount_read(&sk->sk_wmem_alloc) > skb->truesize)
731 return;
732 }
733
734 if (flags & MSG_MORE)
735 nonagle = TCP_NAGLE_CORK;
736
737 __tcp_push_pending_frames(sk, mss_now, nonagle);
738}
739
740static int tcp_splice_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb,
741 unsigned int offset, size_t len)
742{
743 struct tcp_splice_state *tss = rd_desc->arg.data;
744 int ret;
745
746 ret = skb_splice_bits(skb, skb->sk, offset, tss->pipe,
747 min(rd_desc->count, len), tss->flags);
748 if (ret > 0)
749 rd_desc->count -= ret;
750 return ret;
751}
752
753static int __tcp_splice_read(struct sock *sk, struct tcp_splice_state *tss)
754{
755 /* Store TCP splice context information in read_descriptor_t. */
756 read_descriptor_t rd_desc = {
757 .arg.data = tss,
758 .count = tss->len,
759 };
760
761 return tcp_read_sock(sk, &rd_desc, tcp_splice_data_recv);
762}
763
764/**
765 * tcp_splice_read - splice data from TCP socket to a pipe
766 * @sock: socket to splice from
767 * @ppos: position (not valid)
768 * @pipe: pipe to splice to
769 * @len: number of bytes to splice
770 * @flags: splice modifier flags
771 *
772 * Description:
773 * Will read pages from given socket and fill them into a pipe.
774 *
775 **/
776ssize_t tcp_splice_read(struct socket *sock, loff_t *ppos,
777 struct pipe_inode_info *pipe, size_t len,
778 unsigned int flags)
779{
780 struct sock *sk = sock->sk;
781 struct tcp_splice_state tss = {
782 .pipe = pipe,
783 .len = len,
784 .flags = flags,
785 };
786 long timeo;
787 ssize_t spliced;
788 int ret;
789
790 sock_rps_record_flow(sk);
791 /*
792 * We can't seek on a socket input
793 */
794 if (unlikely(*ppos))
795 return -ESPIPE;
796
797 ret = spliced = 0;
798
799 lock_sock(sk);
800
801 timeo = sock_rcvtimeo(sk, sock->file->f_flags & O_NONBLOCK);
802 while (tss.len) {
803 ret = __tcp_splice_read(sk, &tss);
804 if (ret < 0)
805 break;
806 else if (!ret) {
807 if (spliced)
808 break;
809 if (sock_flag(sk, SOCK_DONE))
810 break;
811 if (sk->sk_err) {
812 ret = sock_error(sk);
813 break;
814 }
815 if (sk->sk_shutdown & RCV_SHUTDOWN)
816 break;
817 if (sk->sk_state == TCP_CLOSE) {
818 /*
819 * This occurs when user tries to read
820 * from never connected socket.
821 */
822 ret = -ENOTCONN;
823 break;
824 }
825 if (!timeo) {
826 ret = -EAGAIN;
827 break;
828 }
829 /* if __tcp_splice_read() got nothing while we have
830 * an skb in receive queue, we do not want to loop.
831 * This might happen with URG data.
832 */
833 if (!skb_queue_empty(&sk->sk_receive_queue))
834 break;
835 ret = sk_wait_data(sk, &timeo, NULL);
836 if (ret < 0)
837 break;
838 if (signal_pending(current)) {
839 ret = sock_intr_errno(timeo);
840 break;
841 }
842 continue;
843 }
844 tss.len -= ret;
845 spliced += ret;
846
847 if (!tss.len || !timeo)
848 break;
849 release_sock(sk);
850 lock_sock(sk);
851
852 if (sk->sk_err || sk->sk_state == TCP_CLOSE ||
853 (sk->sk_shutdown & RCV_SHUTDOWN) ||
854 signal_pending(current))
855 break;
856 }
857
858 release_sock(sk);
859
860 if (spliced)
861 return spliced;
862
863 return ret;
864}
865EXPORT_SYMBOL(tcp_splice_read);
866
867struct sk_buff *tcp_stream_alloc_skb(struct sock *sk, gfp_t gfp,
868 bool force_schedule)
869{
870 struct sk_buff *skb;
871
872 skb = alloc_skb_fclone(MAX_TCP_HEADER, gfp);
873 if (likely(skb)) {
874 bool mem_scheduled;
875
876 skb->truesize = SKB_TRUESIZE(skb_end_offset(skb));
877 if (force_schedule) {
878 mem_scheduled = true;
879 sk_forced_mem_schedule(sk, skb->truesize);
880 } else {
881 mem_scheduled = sk_wmem_schedule(sk, skb->truesize);
882 }
883 if (likely(mem_scheduled)) {
884 skb_reserve(skb, MAX_TCP_HEADER);
885 skb->ip_summed = CHECKSUM_PARTIAL;
886 INIT_LIST_HEAD(&skb->tcp_tsorted_anchor);
887 return skb;
888 }
889 __kfree_skb(skb);
890 } else {
891 sk->sk_prot->enter_memory_pressure(sk);
892 sk_stream_moderate_sndbuf(sk);
893 }
894 return NULL;
895}
896
897static unsigned int tcp_xmit_size_goal(struct sock *sk, u32 mss_now,
898 int large_allowed)
899{
900 struct tcp_sock *tp = tcp_sk(sk);
901 u32 new_size_goal, size_goal;
902
903 if (!large_allowed)
904 return mss_now;
905
906 /* Note : tcp_tso_autosize() will eventually split this later */
907 new_size_goal = tcp_bound_to_half_wnd(tp, sk->sk_gso_max_size);
908
909 /* We try hard to avoid divides here */
910 size_goal = tp->gso_segs * mss_now;
911 if (unlikely(new_size_goal < size_goal ||
912 new_size_goal >= size_goal + mss_now)) {
913 tp->gso_segs = min_t(u16, new_size_goal / mss_now,
914 sk->sk_gso_max_segs);
915 size_goal = tp->gso_segs * mss_now;
916 }
917
918 return max(size_goal, mss_now);
919}
920
921int tcp_send_mss(struct sock *sk, int *size_goal, int flags)
922{
923 int mss_now;
924
925 mss_now = tcp_current_mss(sk);
926 *size_goal = tcp_xmit_size_goal(sk, mss_now, !(flags & MSG_OOB));
927
928 return mss_now;
929}
930
931/* In some cases, sendmsg() could have added an skb to the write queue,
932 * but failed adding payload on it. We need to remove it to consume less
933 * memory, but more importantly be able to generate EPOLLOUT for Edge Trigger
934 * epoll() users. Another reason is that tcp_write_xmit() does not like
935 * finding an empty skb in the write queue.
936 */
937void tcp_remove_empty_skb(struct sock *sk)
938{
939 struct sk_buff *skb = tcp_write_queue_tail(sk);
940
941 if (skb && TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq) {
942 tcp_unlink_write_queue(skb, sk);
943 if (tcp_write_queue_empty(sk))
944 tcp_chrono_stop(sk, TCP_CHRONO_BUSY);
945 tcp_wmem_free_skb(sk, skb);
946 }
947}
948
949/* skb changing from pure zc to mixed, must charge zc */
950static int tcp_downgrade_zcopy_pure(struct sock *sk, struct sk_buff *skb)
951{
952 if (unlikely(skb_zcopy_pure(skb))) {
953 u32 extra = skb->truesize -
954 SKB_TRUESIZE(skb_end_offset(skb));
955
956 if (!sk_wmem_schedule(sk, extra))
957 return -ENOMEM;
958
959 sk_mem_charge(sk, extra);
960 skb_shinfo(skb)->flags &= ~SKBFL_PURE_ZEROCOPY;
961 }
962 return 0;
963}
964
965
966int tcp_wmem_schedule(struct sock *sk, int copy)
967{
968 int left;
969
970 if (likely(sk_wmem_schedule(sk, copy)))
971 return copy;
972
973 /* We could be in trouble if we have nothing queued.
974 * Use whatever is left in sk->sk_forward_alloc and tcp_wmem[0]
975 * to guarantee some progress.
976 */
977 left = sock_net(sk)->ipv4.sysctl_tcp_wmem[0] - sk->sk_wmem_queued;
978 if (left > 0)
979 sk_forced_mem_schedule(sk, min(left, copy));
980 return min(copy, sk->sk_forward_alloc);
981}
982
983void tcp_free_fastopen_req(struct tcp_sock *tp)
984{
985 if (tp->fastopen_req) {
986 kfree(tp->fastopen_req);
987 tp->fastopen_req = NULL;
988 }
989}
990
991int tcp_sendmsg_fastopen(struct sock *sk, struct msghdr *msg, int *copied,
992 size_t size, struct ubuf_info *uarg)
993{
994 struct tcp_sock *tp = tcp_sk(sk);
995 struct inet_sock *inet = inet_sk(sk);
996 struct sockaddr *uaddr = msg->msg_name;
997 int err, flags;
998
999 if (!(READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_fastopen) &
1000 TFO_CLIENT_ENABLE) ||
1001 (uaddr && msg->msg_namelen >= sizeof(uaddr->sa_family) &&
1002 uaddr->sa_family == AF_UNSPEC))
1003 return -EOPNOTSUPP;
1004 if (tp->fastopen_req)
1005 return -EALREADY; /* Another Fast Open is in progress */
1006
1007 tp->fastopen_req = kzalloc(sizeof(struct tcp_fastopen_request),
1008 sk->sk_allocation);
1009 if (unlikely(!tp->fastopen_req))
1010 return -ENOBUFS;
1011 tp->fastopen_req->data = msg;
1012 tp->fastopen_req->size = size;
1013 tp->fastopen_req->uarg = uarg;
1014
1015 if (inet_test_bit(DEFER_CONNECT, sk)) {
1016 err = tcp_connect(sk);
1017 /* Same failure procedure as in tcp_v4/6_connect */
1018 if (err) {
1019 tcp_set_state(sk, TCP_CLOSE);
1020 inet->inet_dport = 0;
1021 sk->sk_route_caps = 0;
1022 }
1023 }
1024 flags = (msg->msg_flags & MSG_DONTWAIT) ? O_NONBLOCK : 0;
1025 err = __inet_stream_connect(sk->sk_socket, uaddr,
1026 msg->msg_namelen, flags, 1);
1027 /* fastopen_req could already be freed in __inet_stream_connect
1028 * if the connection times out or gets rst
1029 */
1030 if (tp->fastopen_req) {
1031 *copied = tp->fastopen_req->copied;
1032 tcp_free_fastopen_req(tp);
1033 inet_clear_bit(DEFER_CONNECT, sk);
1034 }
1035 return err;
1036}
1037
1038int tcp_sendmsg_locked(struct sock *sk, struct msghdr *msg, size_t size)
1039{
1040 struct tcp_sock *tp = tcp_sk(sk);
1041 struct ubuf_info *uarg = NULL;
1042 struct sk_buff *skb;
1043 struct sockcm_cookie sockc;
1044 int flags, err, copied = 0;
1045 int mss_now = 0, size_goal, copied_syn = 0;
1046 int process_backlog = 0;
1047 int zc = 0;
1048 long timeo;
1049
1050 flags = msg->msg_flags;
1051
1052 if ((flags & MSG_ZEROCOPY) && size) {
1053 if (msg->msg_ubuf) {
1054 uarg = msg->msg_ubuf;
1055 if (sk->sk_route_caps & NETIF_F_SG)
1056 zc = MSG_ZEROCOPY;
1057 } else if (sock_flag(sk, SOCK_ZEROCOPY)) {
1058 skb = tcp_write_queue_tail(sk);
1059 uarg = msg_zerocopy_realloc(sk, size, skb_zcopy(skb));
1060 if (!uarg) {
1061 err = -ENOBUFS;
1062 goto out_err;
1063 }
1064 if (sk->sk_route_caps & NETIF_F_SG)
1065 zc = MSG_ZEROCOPY;
1066 else
1067 uarg_to_msgzc(uarg)->zerocopy = 0;
1068 }
1069 } else if (unlikely(msg->msg_flags & MSG_SPLICE_PAGES) && size) {
1070 if (sk->sk_route_caps & NETIF_F_SG)
1071 zc = MSG_SPLICE_PAGES;
1072 }
1073
1074 if (unlikely(flags & MSG_FASTOPEN ||
1075 inet_test_bit(DEFER_CONNECT, sk)) &&
1076 !tp->repair) {
1077 err = tcp_sendmsg_fastopen(sk, msg, &copied_syn, size, uarg);
1078 if (err == -EINPROGRESS && copied_syn > 0)
1079 goto out;
1080 else if (err)
1081 goto out_err;
1082 }
1083
1084 timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
1085
1086 tcp_rate_check_app_limited(sk); /* is sending application-limited? */
1087
1088 /* Wait for a connection to finish. One exception is TCP Fast Open
1089 * (passive side) where data is allowed to be sent before a connection
1090 * is fully established.
1091 */
1092 if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) &&
1093 !tcp_passive_fastopen(sk)) {
1094 err = sk_stream_wait_connect(sk, &timeo);
1095 if (err != 0)
1096 goto do_error;
1097 }
1098
1099 if (unlikely(tp->repair)) {
1100 if (tp->repair_queue == TCP_RECV_QUEUE) {
1101 copied = tcp_send_rcvq(sk, msg, size);
1102 goto out_nopush;
1103 }
1104
1105 err = -EINVAL;
1106 if (tp->repair_queue == TCP_NO_QUEUE)
1107 goto out_err;
1108
1109 /* 'common' sending to sendq */
1110 }
1111
1112 sockcm_init(&sockc, sk);
1113 if (msg->msg_controllen) {
1114 err = sock_cmsg_send(sk, msg, &sockc);
1115 if (unlikely(err)) {
1116 err = -EINVAL;
1117 goto out_err;
1118 }
1119 }
1120
1121 /* This should be in poll */
1122 sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
1123
1124 /* Ok commence sending. */
1125 copied = 0;
1126
1127restart:
1128 mss_now = tcp_send_mss(sk, &size_goal, flags);
1129
1130 err = -EPIPE;
1131 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
1132 goto do_error;
1133
1134 while (msg_data_left(msg)) {
1135 ssize_t copy = 0;
1136
1137 skb = tcp_write_queue_tail(sk);
1138 if (skb)
1139 copy = size_goal - skb->len;
1140
1141 if (copy <= 0 || !tcp_skb_can_collapse_to(skb)) {
1142 bool first_skb;
1143
1144new_segment:
1145 if (!sk_stream_memory_free(sk))
1146 goto wait_for_space;
1147
1148 if (unlikely(process_backlog >= 16)) {
1149 process_backlog = 0;
1150 if (sk_flush_backlog(sk))
1151 goto restart;
1152 }
1153 first_skb = tcp_rtx_and_write_queues_empty(sk);
1154 skb = tcp_stream_alloc_skb(sk, sk->sk_allocation,
1155 first_skb);
1156 if (!skb)
1157 goto wait_for_space;
1158
1159 process_backlog++;
1160
1161 tcp_skb_entail(sk, skb);
1162 copy = size_goal;
1163
1164 /* All packets are restored as if they have
1165 * already been sent. skb_mstamp_ns isn't set to
1166 * avoid wrong rtt estimation.
1167 */
1168 if (tp->repair)
1169 TCP_SKB_CB(skb)->sacked |= TCPCB_REPAIRED;
1170 }
1171
1172 /* Try to append data to the end of skb. */
1173 if (copy > msg_data_left(msg))
1174 copy = msg_data_left(msg);
1175
1176 if (zc == 0) {
1177 bool merge = true;
1178 int i = skb_shinfo(skb)->nr_frags;
1179 struct page_frag *pfrag = sk_page_frag(sk);
1180
1181 if (!sk_page_frag_refill(sk, pfrag))
1182 goto wait_for_space;
1183
1184 if (!skb_can_coalesce(skb, i, pfrag->page,
1185 pfrag->offset)) {
1186 if (i >= READ_ONCE(sysctl_max_skb_frags)) {
1187 tcp_mark_push(tp, skb);
1188 goto new_segment;
1189 }
1190 merge = false;
1191 }
1192
1193 copy = min_t(int, copy, pfrag->size - pfrag->offset);
1194
1195 if (unlikely(skb_zcopy_pure(skb) || skb_zcopy_managed(skb))) {
1196 if (tcp_downgrade_zcopy_pure(sk, skb))
1197 goto wait_for_space;
1198 skb_zcopy_downgrade_managed(skb);
1199 }
1200
1201 copy = tcp_wmem_schedule(sk, copy);
1202 if (!copy)
1203 goto wait_for_space;
1204
1205 err = skb_copy_to_page_nocache(sk, &msg->msg_iter, skb,
1206 pfrag->page,
1207 pfrag->offset,
1208 copy);
1209 if (err)
1210 goto do_error;
1211
1212 /* Update the skb. */
1213 if (merge) {
1214 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1215 } else {
1216 skb_fill_page_desc(skb, i, pfrag->page,
1217 pfrag->offset, copy);
1218 page_ref_inc(pfrag->page);
1219 }
1220 pfrag->offset += copy;
1221 } else if (zc == MSG_ZEROCOPY) {
1222 /* First append to a fragless skb builds initial
1223 * pure zerocopy skb
1224 */
1225 if (!skb->len)
1226 skb_shinfo(skb)->flags |= SKBFL_PURE_ZEROCOPY;
1227
1228 if (!skb_zcopy_pure(skb)) {
1229 copy = tcp_wmem_schedule(sk, copy);
1230 if (!copy)
1231 goto wait_for_space;
1232 }
1233
1234 err = skb_zerocopy_iter_stream(sk, skb, msg, copy, uarg);
1235 if (err == -EMSGSIZE || err == -EEXIST) {
1236 tcp_mark_push(tp, skb);
1237 goto new_segment;
1238 }
1239 if (err < 0)
1240 goto do_error;
1241 copy = err;
1242 } else if (zc == MSG_SPLICE_PAGES) {
1243 /* Splice in data if we can; copy if we can't. */
1244 if (tcp_downgrade_zcopy_pure(sk, skb))
1245 goto wait_for_space;
1246 copy = tcp_wmem_schedule(sk, copy);
1247 if (!copy)
1248 goto wait_for_space;
1249
1250 err = skb_splice_from_iter(skb, &msg->msg_iter, copy,
1251 sk->sk_allocation);
1252 if (err < 0) {
1253 if (err == -EMSGSIZE) {
1254 tcp_mark_push(tp, skb);
1255 goto new_segment;
1256 }
1257 goto do_error;
1258 }
1259 copy = err;
1260
1261 if (!(flags & MSG_NO_SHARED_FRAGS))
1262 skb_shinfo(skb)->flags |= SKBFL_SHARED_FRAG;
1263
1264 sk_wmem_queued_add(sk, copy);
1265 sk_mem_charge(sk, copy);
1266 }
1267
1268 if (!copied)
1269 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
1270
1271 WRITE_ONCE(tp->write_seq, tp->write_seq + copy);
1272 TCP_SKB_CB(skb)->end_seq += copy;
1273 tcp_skb_pcount_set(skb, 0);
1274
1275 copied += copy;
1276 if (!msg_data_left(msg)) {
1277 if (unlikely(flags & MSG_EOR))
1278 TCP_SKB_CB(skb)->eor = 1;
1279 goto out;
1280 }
1281
1282 if (skb->len < size_goal || (flags & MSG_OOB) || unlikely(tp->repair))
1283 continue;
1284
1285 if (forced_push(tp)) {
1286 tcp_mark_push(tp, skb);
1287 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
1288 } else if (skb == tcp_send_head(sk))
1289 tcp_push_one(sk, mss_now);
1290 continue;
1291
1292wait_for_space:
1293 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1294 tcp_remove_empty_skb(sk);
1295 if (copied)
1296 tcp_push(sk, flags & ~MSG_MORE, mss_now,
1297 TCP_NAGLE_PUSH, size_goal);
1298
1299 err = sk_stream_wait_memory(sk, &timeo);
1300 if (err != 0)
1301 goto do_error;
1302
1303 mss_now = tcp_send_mss(sk, &size_goal, flags);
1304 }
1305
1306out:
1307 if (copied) {
1308 tcp_tx_timestamp(sk, sockc.tsflags);
1309 tcp_push(sk, flags, mss_now, tp->nonagle, size_goal);
1310 }
1311out_nopush:
1312 /* msg->msg_ubuf is pinned by the caller so we don't take extra refs */
1313 if (uarg && !msg->msg_ubuf)
1314 net_zcopy_put(uarg);
1315 return copied + copied_syn;
1316
1317do_error:
1318 tcp_remove_empty_skb(sk);
1319
1320 if (copied + copied_syn)
1321 goto out;
1322out_err:
1323 /* msg->msg_ubuf is pinned by the caller so we don't take extra refs */
1324 if (uarg && !msg->msg_ubuf)
1325 net_zcopy_put_abort(uarg, true);
1326 err = sk_stream_error(sk, flags, err);
1327 /* make sure we wake any epoll edge trigger waiter */
1328 if (unlikely(tcp_rtx_and_write_queues_empty(sk) && err == -EAGAIN)) {
1329 sk->sk_write_space(sk);
1330 tcp_chrono_stop(sk, TCP_CHRONO_SNDBUF_LIMITED);
1331 }
1332 return err;
1333}
1334EXPORT_SYMBOL_GPL(tcp_sendmsg_locked);
1335
1336int tcp_sendmsg(struct sock *sk, struct msghdr *msg, size_t size)
1337{
1338 int ret;
1339
1340 lock_sock(sk);
1341 ret = tcp_sendmsg_locked(sk, msg, size);
1342 release_sock(sk);
1343
1344 return ret;
1345}
1346EXPORT_SYMBOL(tcp_sendmsg);
1347
1348void tcp_splice_eof(struct socket *sock)
1349{
1350 struct sock *sk = sock->sk;
1351 struct tcp_sock *tp = tcp_sk(sk);
1352 int mss_now, size_goal;
1353
1354 if (!tcp_write_queue_tail(sk))
1355 return;
1356
1357 lock_sock(sk);
1358 mss_now = tcp_send_mss(sk, &size_goal, 0);
1359 tcp_push(sk, 0, mss_now, tp->nonagle, size_goal);
1360 release_sock(sk);
1361}
1362EXPORT_SYMBOL_GPL(tcp_splice_eof);
1363
1364/*
1365 * Handle reading urgent data. BSD has very simple semantics for
1366 * this, no blocking and very strange errors 8)
1367 */
1368
1369static int tcp_recv_urg(struct sock *sk, struct msghdr *msg, int len, int flags)
1370{
1371 struct tcp_sock *tp = tcp_sk(sk);
1372
1373 /* No URG data to read. */
1374 if (sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data ||
1375 tp->urg_data == TCP_URG_READ)
1376 return -EINVAL; /* Yes this is right ! */
1377
1378 if (sk->sk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DONE))
1379 return -ENOTCONN;
1380
1381 if (tp->urg_data & TCP_URG_VALID) {
1382 int err = 0;
1383 char c = tp->urg_data;
1384
1385 if (!(flags & MSG_PEEK))
1386 WRITE_ONCE(tp->urg_data, TCP_URG_READ);
1387
1388 /* Read urgent data. */
1389 msg->msg_flags |= MSG_OOB;
1390
1391 if (len > 0) {
1392 if (!(flags & MSG_TRUNC))
1393 err = memcpy_to_msg(msg, &c, 1);
1394 len = 1;
1395 } else
1396 msg->msg_flags |= MSG_TRUNC;
1397
1398 return err ? -EFAULT : len;
1399 }
1400
1401 if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
1402 return 0;
1403
1404 /* Fixed the recv(..., MSG_OOB) behaviour. BSD docs and
1405 * the available implementations agree in this case:
1406 * this call should never block, independent of the
1407 * blocking state of the socket.
1408 * Mike <pall@rz.uni-karlsruhe.de>
1409 */
1410 return -EAGAIN;
1411}
1412
1413static int tcp_peek_sndq(struct sock *sk, struct msghdr *msg, int len)
1414{
1415 struct sk_buff *skb;
1416 int copied = 0, err = 0;
1417
1418 /* XXX -- need to support SO_PEEK_OFF */
1419
1420 skb_rbtree_walk(skb, &sk->tcp_rtx_queue) {
1421 err = skb_copy_datagram_msg(skb, 0, msg, skb->len);
1422 if (err)
1423 return err;
1424 copied += skb->len;
1425 }
1426
1427 skb_queue_walk(&sk->sk_write_queue, skb) {
1428 err = skb_copy_datagram_msg(skb, 0, msg, skb->len);
1429 if (err)
1430 break;
1431
1432 copied += skb->len;
1433 }
1434
1435 return err ?: copied;
1436}
1437
1438/* Clean up the receive buffer for full frames taken by the user,
1439 * then send an ACK if necessary. COPIED is the number of bytes
1440 * tcp_recvmsg has given to the user so far, it speeds up the
1441 * calculation of whether or not we must ACK for the sake of
1442 * a window update.
1443 */
1444void __tcp_cleanup_rbuf(struct sock *sk, int copied)
1445{
1446 struct tcp_sock *tp = tcp_sk(sk);
1447 bool time_to_ack = false;
1448
1449 if (inet_csk_ack_scheduled(sk)) {
1450 const struct inet_connection_sock *icsk = inet_csk(sk);
1451
1452 if (/* Once-per-two-segments ACK was not sent by tcp_input.c */
1453 tp->rcv_nxt - tp->rcv_wup > icsk->icsk_ack.rcv_mss ||
1454 /*
1455 * If this read emptied read buffer, we send ACK, if
1456 * connection is not bidirectional, user drained
1457 * receive buffer and there was a small segment
1458 * in queue.
1459 */
1460 (copied > 0 &&
1461 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED2) ||
1462 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED) &&
1463 !inet_csk_in_pingpong_mode(sk))) &&
1464 !atomic_read(&sk->sk_rmem_alloc)))
1465 time_to_ack = true;
1466 }
1467
1468 /* We send an ACK if we can now advertise a non-zero window
1469 * which has been raised "significantly".
1470 *
1471 * Even if window raised up to infinity, do not send window open ACK
1472 * in states, where we will not receive more. It is useless.
1473 */
1474 if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) {
1475 __u32 rcv_window_now = tcp_receive_window(tp);
1476
1477 /* Optimize, __tcp_select_window() is not cheap. */
1478 if (2*rcv_window_now <= tp->window_clamp) {
1479 __u32 new_window = __tcp_select_window(sk);
1480
1481 /* Send ACK now, if this read freed lots of space
1482 * in our buffer. Certainly, new_window is new window.
1483 * We can advertise it now, if it is not less than current one.
1484 * "Lots" means "at least twice" here.
1485 */
1486 if (new_window && new_window >= 2 * rcv_window_now)
1487 time_to_ack = true;
1488 }
1489 }
1490 if (time_to_ack)
1491 tcp_send_ack(sk);
1492}
1493
1494void tcp_cleanup_rbuf(struct sock *sk, int copied)
1495{
1496 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
1497 struct tcp_sock *tp = tcp_sk(sk);
1498
1499 WARN(skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq),
1500 "cleanup rbuf bug: copied %X seq %X rcvnxt %X\n",
1501 tp->copied_seq, TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt);
1502 __tcp_cleanup_rbuf(sk, copied);
1503}
1504
1505static void tcp_eat_recv_skb(struct sock *sk, struct sk_buff *skb)
1506{
1507 __skb_unlink(skb, &sk->sk_receive_queue);
1508 if (likely(skb->destructor == sock_rfree)) {
1509 sock_rfree(skb);
1510 skb->destructor = NULL;
1511 skb->sk = NULL;
1512 return skb_attempt_defer_free(skb);
1513 }
1514 __kfree_skb(skb);
1515}
1516
1517struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off)
1518{
1519 struct sk_buff *skb;
1520 u32 offset;
1521
1522 while ((skb = skb_peek(&sk->sk_receive_queue)) != NULL) {
1523 offset = seq - TCP_SKB_CB(skb)->seq;
1524 if (unlikely(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
1525 pr_err_once("%s: found a SYN, please report !\n", __func__);
1526 offset--;
1527 }
1528 if (offset < skb->len || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)) {
1529 *off = offset;
1530 return skb;
1531 }
1532 /* This looks weird, but this can happen if TCP collapsing
1533 * splitted a fat GRO packet, while we released socket lock
1534 * in skb_splice_bits()
1535 */
1536 tcp_eat_recv_skb(sk, skb);
1537 }
1538 return NULL;
1539}
1540EXPORT_SYMBOL(tcp_recv_skb);
1541
1542/*
1543 * This routine provides an alternative to tcp_recvmsg() for routines
1544 * that would like to handle copying from skbuffs directly in 'sendfile'
1545 * fashion.
1546 * Note:
1547 * - It is assumed that the socket was locked by the caller.
1548 * - The routine does not block.
1549 * - At present, there is no support for reading OOB data
1550 * or for 'peeking' the socket using this routine
1551 * (although both would be easy to implement).
1552 */
1553int tcp_read_sock(struct sock *sk, read_descriptor_t *desc,
1554 sk_read_actor_t recv_actor)
1555{
1556 struct sk_buff *skb;
1557 struct tcp_sock *tp = tcp_sk(sk);
1558 u32 seq = tp->copied_seq;
1559 u32 offset;
1560 int copied = 0;
1561
1562 if (sk->sk_state == TCP_LISTEN)
1563 return -ENOTCONN;
1564 while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1565 if (offset < skb->len) {
1566 int used;
1567 size_t len;
1568
1569 len = skb->len - offset;
1570 /* Stop reading if we hit a patch of urgent data */
1571 if (unlikely(tp->urg_data)) {
1572 u32 urg_offset = tp->urg_seq - seq;
1573 if (urg_offset < len)
1574 len = urg_offset;
1575 if (!len)
1576 break;
1577 }
1578 used = recv_actor(desc, skb, offset, len);
1579 if (used <= 0) {
1580 if (!copied)
1581 copied = used;
1582 break;
1583 }
1584 if (WARN_ON_ONCE(used > len))
1585 used = len;
1586 seq += used;
1587 copied += used;
1588 offset += used;
1589
1590 /* If recv_actor drops the lock (e.g. TCP splice
1591 * receive) the skb pointer might be invalid when
1592 * getting here: tcp_collapse might have deleted it
1593 * while aggregating skbs from the socket queue.
1594 */
1595 skb = tcp_recv_skb(sk, seq - 1, &offset);
1596 if (!skb)
1597 break;
1598 /* TCP coalescing might have appended data to the skb.
1599 * Try to splice more frags
1600 */
1601 if (offset + 1 != skb->len)
1602 continue;
1603 }
1604 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) {
1605 tcp_eat_recv_skb(sk, skb);
1606 ++seq;
1607 break;
1608 }
1609 tcp_eat_recv_skb(sk, skb);
1610 if (!desc->count)
1611 break;
1612 WRITE_ONCE(tp->copied_seq, seq);
1613 }
1614 WRITE_ONCE(tp->copied_seq, seq);
1615
1616 tcp_rcv_space_adjust(sk);
1617
1618 /* Clean up data we have read: This will do ACK frames. */
1619 if (copied > 0) {
1620 tcp_recv_skb(sk, seq, &offset);
1621 tcp_cleanup_rbuf(sk, copied);
1622 }
1623 return copied;
1624}
1625EXPORT_SYMBOL(tcp_read_sock);
1626
1627int tcp_read_skb(struct sock *sk, skb_read_actor_t recv_actor)
1628{
1629 struct sk_buff *skb;
1630 int copied = 0;
1631
1632 if (sk->sk_state == TCP_LISTEN)
1633 return -ENOTCONN;
1634
1635 while ((skb = skb_peek(&sk->sk_receive_queue)) != NULL) {
1636 u8 tcp_flags;
1637 int used;
1638
1639 __skb_unlink(skb, &sk->sk_receive_queue);
1640 WARN_ON_ONCE(!skb_set_owner_sk_safe(skb, sk));
1641 tcp_flags = TCP_SKB_CB(skb)->tcp_flags;
1642 used = recv_actor(sk, skb);
1643 if (used < 0) {
1644 if (!copied)
1645 copied = used;
1646 break;
1647 }
1648 copied += used;
1649
1650 if (tcp_flags & TCPHDR_FIN)
1651 break;
1652 }
1653 return copied;
1654}
1655EXPORT_SYMBOL(tcp_read_skb);
1656
1657void tcp_read_done(struct sock *sk, size_t len)
1658{
1659 struct tcp_sock *tp = tcp_sk(sk);
1660 u32 seq = tp->copied_seq;
1661 struct sk_buff *skb;
1662 size_t left;
1663 u32 offset;
1664
1665 if (sk->sk_state == TCP_LISTEN)
1666 return;
1667
1668 left = len;
1669 while (left && (skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1670 int used;
1671
1672 used = min_t(size_t, skb->len - offset, left);
1673 seq += used;
1674 left -= used;
1675
1676 if (skb->len > offset + used)
1677 break;
1678
1679 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) {
1680 tcp_eat_recv_skb(sk, skb);
1681 ++seq;
1682 break;
1683 }
1684 tcp_eat_recv_skb(sk, skb);
1685 }
1686 WRITE_ONCE(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 (left != len)
1692 tcp_cleanup_rbuf(sk, len - left);
1693}
1694EXPORT_SYMBOL(tcp_read_done);
1695
1696int tcp_peek_len(struct socket *sock)
1697{
1698 return tcp_inq(sock->sk);
1699}
1700EXPORT_SYMBOL(tcp_peek_len);
1701
1702/* Make sure sk_rcvbuf is big enough to satisfy SO_RCVLOWAT hint */
1703int tcp_set_rcvlowat(struct sock *sk, int val)
1704{
1705 int space, cap;
1706
1707 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK)
1708 cap = sk->sk_rcvbuf >> 1;
1709 else
1710 cap = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_rmem[2]) >> 1;
1711 val = min(val, cap);
1712 WRITE_ONCE(sk->sk_rcvlowat, val ? : 1);
1713
1714 /* Check if we need to signal EPOLLIN right now */
1715 tcp_data_ready(sk);
1716
1717 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK)
1718 return 0;
1719
1720 space = tcp_space_from_win(sk, val);
1721 if (space > sk->sk_rcvbuf) {
1722 WRITE_ONCE(sk->sk_rcvbuf, space);
1723 tcp_sk(sk)->window_clamp = val;
1724 }
1725 return 0;
1726}
1727EXPORT_SYMBOL(tcp_set_rcvlowat);
1728
1729void tcp_update_recv_tstamps(struct sk_buff *skb,
1730 struct scm_timestamping_internal *tss)
1731{
1732 if (skb->tstamp)
1733 tss->ts[0] = ktime_to_timespec64(skb->tstamp);
1734 else
1735 tss->ts[0] = (struct timespec64) {0};
1736
1737 if (skb_hwtstamps(skb)->hwtstamp)
1738 tss->ts[2] = ktime_to_timespec64(skb_hwtstamps(skb)->hwtstamp);
1739 else
1740 tss->ts[2] = (struct timespec64) {0};
1741}
1742
1743#ifdef CONFIG_MMU
1744static const struct vm_operations_struct tcp_vm_ops = {
1745};
1746
1747int tcp_mmap(struct file *file, struct socket *sock,
1748 struct vm_area_struct *vma)
1749{
1750 if (vma->vm_flags & (VM_WRITE | VM_EXEC))
1751 return -EPERM;
1752 vm_flags_clear(vma, VM_MAYWRITE | VM_MAYEXEC);
1753
1754 /* Instruct vm_insert_page() to not mmap_read_lock(mm) */
1755 vm_flags_set(vma, VM_MIXEDMAP);
1756
1757 vma->vm_ops = &tcp_vm_ops;
1758 return 0;
1759}
1760EXPORT_SYMBOL(tcp_mmap);
1761
1762static skb_frag_t *skb_advance_to_frag(struct sk_buff *skb, u32 offset_skb,
1763 u32 *offset_frag)
1764{
1765 skb_frag_t *frag;
1766
1767 if (unlikely(offset_skb >= skb->len))
1768 return NULL;
1769
1770 offset_skb -= skb_headlen(skb);
1771 if ((int)offset_skb < 0 || skb_has_frag_list(skb))
1772 return NULL;
1773
1774 frag = skb_shinfo(skb)->frags;
1775 while (offset_skb) {
1776 if (skb_frag_size(frag) > offset_skb) {
1777 *offset_frag = offset_skb;
1778 return frag;
1779 }
1780 offset_skb -= skb_frag_size(frag);
1781 ++frag;
1782 }
1783 *offset_frag = 0;
1784 return frag;
1785}
1786
1787static bool can_map_frag(const skb_frag_t *frag)
1788{
1789 struct page *page;
1790
1791 if (skb_frag_size(frag) != PAGE_SIZE || skb_frag_off(frag))
1792 return false;
1793
1794 page = skb_frag_page(frag);
1795
1796 if (PageCompound(page) || page->mapping)
1797 return false;
1798
1799 return true;
1800}
1801
1802static int find_next_mappable_frag(const skb_frag_t *frag,
1803 int remaining_in_skb)
1804{
1805 int offset = 0;
1806
1807 if (likely(can_map_frag(frag)))
1808 return 0;
1809
1810 while (offset < remaining_in_skb && !can_map_frag(frag)) {
1811 offset += skb_frag_size(frag);
1812 ++frag;
1813 }
1814 return offset;
1815}
1816
1817static void tcp_zerocopy_set_hint_for_skb(struct sock *sk,
1818 struct tcp_zerocopy_receive *zc,
1819 struct sk_buff *skb, u32 offset)
1820{
1821 u32 frag_offset, partial_frag_remainder = 0;
1822 int mappable_offset;
1823 skb_frag_t *frag;
1824
1825 /* worst case: skip to next skb. try to improve on this case below */
1826 zc->recv_skip_hint = skb->len - offset;
1827
1828 /* Find the frag containing this offset (and how far into that frag) */
1829 frag = skb_advance_to_frag(skb, offset, &frag_offset);
1830 if (!frag)
1831 return;
1832
1833 if (frag_offset) {
1834 struct skb_shared_info *info = skb_shinfo(skb);
1835
1836 /* We read part of the last frag, must recvmsg() rest of skb. */
1837 if (frag == &info->frags[info->nr_frags - 1])
1838 return;
1839
1840 /* Else, we must at least read the remainder in this frag. */
1841 partial_frag_remainder = skb_frag_size(frag) - frag_offset;
1842 zc->recv_skip_hint -= partial_frag_remainder;
1843 ++frag;
1844 }
1845
1846 /* partial_frag_remainder: If part way through a frag, must read rest.
1847 * mappable_offset: Bytes till next mappable frag, *not* counting bytes
1848 * in partial_frag_remainder.
1849 */
1850 mappable_offset = find_next_mappable_frag(frag, zc->recv_skip_hint);
1851 zc->recv_skip_hint = mappable_offset + partial_frag_remainder;
1852}
1853
1854static int tcp_recvmsg_locked(struct sock *sk, struct msghdr *msg, size_t len,
1855 int flags, struct scm_timestamping_internal *tss,
1856 int *cmsg_flags);
1857static int receive_fallback_to_copy(struct sock *sk,
1858 struct tcp_zerocopy_receive *zc, int inq,
1859 struct scm_timestamping_internal *tss)
1860{
1861 unsigned long copy_address = (unsigned long)zc->copybuf_address;
1862 struct msghdr msg = {};
1863 int err;
1864
1865 zc->length = 0;
1866 zc->recv_skip_hint = 0;
1867
1868 if (copy_address != zc->copybuf_address)
1869 return -EINVAL;
1870
1871 err = import_ubuf(ITER_DEST, (void __user *)copy_address, inq,
1872 &msg.msg_iter);
1873 if (err)
1874 return err;
1875
1876 err = tcp_recvmsg_locked(sk, &msg, inq, MSG_DONTWAIT,
1877 tss, &zc->msg_flags);
1878 if (err < 0)
1879 return err;
1880
1881 zc->copybuf_len = err;
1882 if (likely(zc->copybuf_len)) {
1883 struct sk_buff *skb;
1884 u32 offset;
1885
1886 skb = tcp_recv_skb(sk, tcp_sk(sk)->copied_seq, &offset);
1887 if (skb)
1888 tcp_zerocopy_set_hint_for_skb(sk, zc, skb, offset);
1889 }
1890 return 0;
1891}
1892
1893static int tcp_copy_straggler_data(struct tcp_zerocopy_receive *zc,
1894 struct sk_buff *skb, u32 copylen,
1895 u32 *offset, u32 *seq)
1896{
1897 unsigned long copy_address = (unsigned long)zc->copybuf_address;
1898 struct msghdr msg = {};
1899 int err;
1900
1901 if (copy_address != zc->copybuf_address)
1902 return -EINVAL;
1903
1904 err = import_ubuf(ITER_DEST, (void __user *)copy_address, copylen,
1905 &msg.msg_iter);
1906 if (err)
1907 return err;
1908 err = skb_copy_datagram_msg(skb, *offset, &msg, copylen);
1909 if (err)
1910 return err;
1911 zc->recv_skip_hint -= copylen;
1912 *offset += copylen;
1913 *seq += copylen;
1914 return (__s32)copylen;
1915}
1916
1917static int tcp_zc_handle_leftover(struct tcp_zerocopy_receive *zc,
1918 struct sock *sk,
1919 struct sk_buff *skb,
1920 u32 *seq,
1921 s32 copybuf_len,
1922 struct scm_timestamping_internal *tss)
1923{
1924 u32 offset, copylen = min_t(u32, copybuf_len, zc->recv_skip_hint);
1925
1926 if (!copylen)
1927 return 0;
1928 /* skb is null if inq < PAGE_SIZE. */
1929 if (skb) {
1930 offset = *seq - TCP_SKB_CB(skb)->seq;
1931 } else {
1932 skb = tcp_recv_skb(sk, *seq, &offset);
1933 if (TCP_SKB_CB(skb)->has_rxtstamp) {
1934 tcp_update_recv_tstamps(skb, tss);
1935 zc->msg_flags |= TCP_CMSG_TS;
1936 }
1937 }
1938
1939 zc->copybuf_len = tcp_copy_straggler_data(zc, skb, copylen, &offset,
1940 seq);
1941 return zc->copybuf_len < 0 ? 0 : copylen;
1942}
1943
1944static int tcp_zerocopy_vm_insert_batch_error(struct vm_area_struct *vma,
1945 struct page **pending_pages,
1946 unsigned long pages_remaining,
1947 unsigned long *address,
1948 u32 *length,
1949 u32 *seq,
1950 struct tcp_zerocopy_receive *zc,
1951 u32 total_bytes_to_map,
1952 int err)
1953{
1954 /* At least one page did not map. Try zapping if we skipped earlier. */
1955 if (err == -EBUSY &&
1956 zc->flags & TCP_RECEIVE_ZEROCOPY_FLAG_TLB_CLEAN_HINT) {
1957 u32 maybe_zap_len;
1958
1959 maybe_zap_len = total_bytes_to_map - /* All bytes to map */
1960 *length + /* Mapped or pending */
1961 (pages_remaining * PAGE_SIZE); /* Failed map. */
1962 zap_page_range_single(vma, *address, maybe_zap_len, NULL);
1963 err = 0;
1964 }
1965
1966 if (!err) {
1967 unsigned long leftover_pages = pages_remaining;
1968 int bytes_mapped;
1969
1970 /* We called zap_page_range_single, try to reinsert. */
1971 err = vm_insert_pages(vma, *address,
1972 pending_pages,
1973 &pages_remaining);
1974 bytes_mapped = PAGE_SIZE * (leftover_pages - pages_remaining);
1975 *seq += bytes_mapped;
1976 *address += bytes_mapped;
1977 }
1978 if (err) {
1979 /* Either we were unable to zap, OR we zapped, retried an
1980 * insert, and still had an issue. Either ways, pages_remaining
1981 * is the number of pages we were unable to map, and we unroll
1982 * some state we speculatively touched before.
1983 */
1984 const int bytes_not_mapped = PAGE_SIZE * pages_remaining;
1985
1986 *length -= bytes_not_mapped;
1987 zc->recv_skip_hint += bytes_not_mapped;
1988 }
1989 return err;
1990}
1991
1992static int tcp_zerocopy_vm_insert_batch(struct vm_area_struct *vma,
1993 struct page **pages,
1994 unsigned int pages_to_map,
1995 unsigned long *address,
1996 u32 *length,
1997 u32 *seq,
1998 struct tcp_zerocopy_receive *zc,
1999 u32 total_bytes_to_map)
2000{
2001 unsigned long pages_remaining = pages_to_map;
2002 unsigned int pages_mapped;
2003 unsigned int bytes_mapped;
2004 int err;
2005
2006 err = vm_insert_pages(vma, *address, pages, &pages_remaining);
2007 pages_mapped = pages_to_map - (unsigned int)pages_remaining;
2008 bytes_mapped = PAGE_SIZE * pages_mapped;
2009 /* Even if vm_insert_pages fails, it may have partially succeeded in
2010 * mapping (some but not all of the pages).
2011 */
2012 *seq += bytes_mapped;
2013 *address += bytes_mapped;
2014
2015 if (likely(!err))
2016 return 0;
2017
2018 /* Error: maybe zap and retry + rollback state for failed inserts. */
2019 return tcp_zerocopy_vm_insert_batch_error(vma, pages + pages_mapped,
2020 pages_remaining, address, length, seq, zc, total_bytes_to_map,
2021 err);
2022}
2023
2024#define TCP_VALID_ZC_MSG_FLAGS (TCP_CMSG_TS)
2025static void tcp_zc_finalize_rx_tstamp(struct sock *sk,
2026 struct tcp_zerocopy_receive *zc,
2027 struct scm_timestamping_internal *tss)
2028{
2029 unsigned long msg_control_addr;
2030 struct msghdr cmsg_dummy;
2031
2032 msg_control_addr = (unsigned long)zc->msg_control;
2033 cmsg_dummy.msg_control_user = (void __user *)msg_control_addr;
2034 cmsg_dummy.msg_controllen =
2035 (__kernel_size_t)zc->msg_controllen;
2036 cmsg_dummy.msg_flags = in_compat_syscall()
2037 ? MSG_CMSG_COMPAT : 0;
2038 cmsg_dummy.msg_control_is_user = true;
2039 zc->msg_flags = 0;
2040 if (zc->msg_control == msg_control_addr &&
2041 zc->msg_controllen == cmsg_dummy.msg_controllen) {
2042 tcp_recv_timestamp(&cmsg_dummy, sk, tss);
2043 zc->msg_control = (__u64)
2044 ((uintptr_t)cmsg_dummy.msg_control_user);
2045 zc->msg_controllen =
2046 (__u64)cmsg_dummy.msg_controllen;
2047 zc->msg_flags = (__u32)cmsg_dummy.msg_flags;
2048 }
2049}
2050
2051static struct vm_area_struct *find_tcp_vma(struct mm_struct *mm,
2052 unsigned long address,
2053 bool *mmap_locked)
2054{
2055 struct vm_area_struct *vma = lock_vma_under_rcu(mm, address);
2056
2057 if (vma) {
2058 if (vma->vm_ops != &tcp_vm_ops) {
2059 vma_end_read(vma);
2060 return NULL;
2061 }
2062 *mmap_locked = false;
2063 return vma;
2064 }
2065
2066 mmap_read_lock(mm);
2067 vma = vma_lookup(mm, address);
2068 if (!vma || vma->vm_ops != &tcp_vm_ops) {
2069 mmap_read_unlock(mm);
2070 return NULL;
2071 }
2072 *mmap_locked = true;
2073 return vma;
2074}
2075
2076#define TCP_ZEROCOPY_PAGE_BATCH_SIZE 32
2077static int tcp_zerocopy_receive(struct sock *sk,
2078 struct tcp_zerocopy_receive *zc,
2079 struct scm_timestamping_internal *tss)
2080{
2081 u32 length = 0, offset, vma_len, avail_len, copylen = 0;
2082 unsigned long address = (unsigned long)zc->address;
2083 struct page *pages[TCP_ZEROCOPY_PAGE_BATCH_SIZE];
2084 s32 copybuf_len = zc->copybuf_len;
2085 struct tcp_sock *tp = tcp_sk(sk);
2086 const skb_frag_t *frags = NULL;
2087 unsigned int pages_to_map = 0;
2088 struct vm_area_struct *vma;
2089 struct sk_buff *skb = NULL;
2090 u32 seq = tp->copied_seq;
2091 u32 total_bytes_to_map;
2092 int inq = tcp_inq(sk);
2093 bool mmap_locked;
2094 int ret;
2095
2096 zc->copybuf_len = 0;
2097 zc->msg_flags = 0;
2098
2099 if (address & (PAGE_SIZE - 1) || address != zc->address)
2100 return -EINVAL;
2101
2102 if (sk->sk_state == TCP_LISTEN)
2103 return -ENOTCONN;
2104
2105 sock_rps_record_flow(sk);
2106
2107 if (inq && inq <= copybuf_len)
2108 return receive_fallback_to_copy(sk, zc, inq, tss);
2109
2110 if (inq < PAGE_SIZE) {
2111 zc->length = 0;
2112 zc->recv_skip_hint = inq;
2113 if (!inq && sock_flag(sk, SOCK_DONE))
2114 return -EIO;
2115 return 0;
2116 }
2117
2118 vma = find_tcp_vma(current->mm, address, &mmap_locked);
2119 if (!vma)
2120 return -EINVAL;
2121
2122 vma_len = min_t(unsigned long, zc->length, vma->vm_end - address);
2123 avail_len = min_t(u32, vma_len, inq);
2124 total_bytes_to_map = avail_len & ~(PAGE_SIZE - 1);
2125 if (total_bytes_to_map) {
2126 if (!(zc->flags & TCP_RECEIVE_ZEROCOPY_FLAG_TLB_CLEAN_HINT))
2127 zap_page_range_single(vma, address, total_bytes_to_map,
2128 NULL);
2129 zc->length = total_bytes_to_map;
2130 zc->recv_skip_hint = 0;
2131 } else {
2132 zc->length = avail_len;
2133 zc->recv_skip_hint = avail_len;
2134 }
2135 ret = 0;
2136 while (length + PAGE_SIZE <= zc->length) {
2137 int mappable_offset;
2138 struct page *page;
2139
2140 if (zc->recv_skip_hint < PAGE_SIZE) {
2141 u32 offset_frag;
2142
2143 if (skb) {
2144 if (zc->recv_skip_hint > 0)
2145 break;
2146 skb = skb->next;
2147 offset = seq - TCP_SKB_CB(skb)->seq;
2148 } else {
2149 skb = tcp_recv_skb(sk, seq, &offset);
2150 }
2151
2152 if (TCP_SKB_CB(skb)->has_rxtstamp) {
2153 tcp_update_recv_tstamps(skb, tss);
2154 zc->msg_flags |= TCP_CMSG_TS;
2155 }
2156 zc->recv_skip_hint = skb->len - offset;
2157 frags = skb_advance_to_frag(skb, offset, &offset_frag);
2158 if (!frags || offset_frag)
2159 break;
2160 }
2161
2162 mappable_offset = find_next_mappable_frag(frags,
2163 zc->recv_skip_hint);
2164 if (mappable_offset) {
2165 zc->recv_skip_hint = mappable_offset;
2166 break;
2167 }
2168 page = skb_frag_page(frags);
2169 prefetchw(page);
2170 pages[pages_to_map++] = page;
2171 length += PAGE_SIZE;
2172 zc->recv_skip_hint -= PAGE_SIZE;
2173 frags++;
2174 if (pages_to_map == TCP_ZEROCOPY_PAGE_BATCH_SIZE ||
2175 zc->recv_skip_hint < PAGE_SIZE) {
2176 /* Either full batch, or we're about to go to next skb
2177 * (and we cannot unroll failed ops across skbs).
2178 */
2179 ret = tcp_zerocopy_vm_insert_batch(vma, pages,
2180 pages_to_map,
2181 &address, &length,
2182 &seq, zc,
2183 total_bytes_to_map);
2184 if (ret)
2185 goto out;
2186 pages_to_map = 0;
2187 }
2188 }
2189 if (pages_to_map) {
2190 ret = tcp_zerocopy_vm_insert_batch(vma, pages, pages_to_map,
2191 &address, &length, &seq,
2192 zc, total_bytes_to_map);
2193 }
2194out:
2195 if (mmap_locked)
2196 mmap_read_unlock(current->mm);
2197 else
2198 vma_end_read(vma);
2199 /* Try to copy straggler data. */
2200 if (!ret)
2201 copylen = tcp_zc_handle_leftover(zc, sk, skb, &seq, copybuf_len, tss);
2202
2203 if (length + copylen) {
2204 WRITE_ONCE(tp->copied_seq, seq);
2205 tcp_rcv_space_adjust(sk);
2206
2207 /* Clean up data we have read: This will do ACK frames. */
2208 tcp_recv_skb(sk, seq, &offset);
2209 tcp_cleanup_rbuf(sk, length + copylen);
2210 ret = 0;
2211 if (length == zc->length)
2212 zc->recv_skip_hint = 0;
2213 } else {
2214 if (!zc->recv_skip_hint && sock_flag(sk, SOCK_DONE))
2215 ret = -EIO;
2216 }
2217 zc->length = length;
2218 return ret;
2219}
2220#endif
2221
2222/* Similar to __sock_recv_timestamp, but does not require an skb */
2223void tcp_recv_timestamp(struct msghdr *msg, const struct sock *sk,
2224 struct scm_timestamping_internal *tss)
2225{
2226 int new_tstamp = sock_flag(sk, SOCK_TSTAMP_NEW);
2227 bool has_timestamping = false;
2228
2229 if (tss->ts[0].tv_sec || tss->ts[0].tv_nsec) {
2230 if (sock_flag(sk, SOCK_RCVTSTAMP)) {
2231 if (sock_flag(sk, SOCK_RCVTSTAMPNS)) {
2232 if (new_tstamp) {
2233 struct __kernel_timespec kts = {
2234 .tv_sec = tss->ts[0].tv_sec,
2235 .tv_nsec = tss->ts[0].tv_nsec,
2236 };
2237 put_cmsg(msg, SOL_SOCKET, SO_TIMESTAMPNS_NEW,
2238 sizeof(kts), &kts);
2239 } else {
2240 struct __kernel_old_timespec ts_old = {
2241 .tv_sec = tss->ts[0].tv_sec,
2242 .tv_nsec = tss->ts[0].tv_nsec,
2243 };
2244 put_cmsg(msg, SOL_SOCKET, SO_TIMESTAMPNS_OLD,
2245 sizeof(ts_old), &ts_old);
2246 }
2247 } else {
2248 if (new_tstamp) {
2249 struct __kernel_sock_timeval stv = {
2250 .tv_sec = tss->ts[0].tv_sec,
2251 .tv_usec = tss->ts[0].tv_nsec / 1000,
2252 };
2253 put_cmsg(msg, SOL_SOCKET, SO_TIMESTAMP_NEW,
2254 sizeof(stv), &stv);
2255 } else {
2256 struct __kernel_old_timeval tv = {
2257 .tv_sec = tss->ts[0].tv_sec,
2258 .tv_usec = tss->ts[0].tv_nsec / 1000,
2259 };
2260 put_cmsg(msg, SOL_SOCKET, SO_TIMESTAMP_OLD,
2261 sizeof(tv), &tv);
2262 }
2263 }
2264 }
2265
2266 if (READ_ONCE(sk->sk_tsflags) & SOF_TIMESTAMPING_SOFTWARE)
2267 has_timestamping = true;
2268 else
2269 tss->ts[0] = (struct timespec64) {0};
2270 }
2271
2272 if (tss->ts[2].tv_sec || tss->ts[2].tv_nsec) {
2273 if (READ_ONCE(sk->sk_tsflags) & SOF_TIMESTAMPING_RAW_HARDWARE)
2274 has_timestamping = true;
2275 else
2276 tss->ts[2] = (struct timespec64) {0};
2277 }
2278
2279 if (has_timestamping) {
2280 tss->ts[1] = (struct timespec64) {0};
2281 if (sock_flag(sk, SOCK_TSTAMP_NEW))
2282 put_cmsg_scm_timestamping64(msg, tss);
2283 else
2284 put_cmsg_scm_timestamping(msg, tss);
2285 }
2286}
2287
2288static int tcp_inq_hint(struct sock *sk)
2289{
2290 const struct tcp_sock *tp = tcp_sk(sk);
2291 u32 copied_seq = READ_ONCE(tp->copied_seq);
2292 u32 rcv_nxt = READ_ONCE(tp->rcv_nxt);
2293 int inq;
2294
2295 inq = rcv_nxt - copied_seq;
2296 if (unlikely(inq < 0 || copied_seq != READ_ONCE(tp->copied_seq))) {
2297 lock_sock(sk);
2298 inq = tp->rcv_nxt - tp->copied_seq;
2299 release_sock(sk);
2300 }
2301 /* After receiving a FIN, tell the user-space to continue reading
2302 * by returning a non-zero inq.
2303 */
2304 if (inq == 0 && sock_flag(sk, SOCK_DONE))
2305 inq = 1;
2306 return inq;
2307}
2308
2309/*
2310 * This routine copies from a sock struct into the user buffer.
2311 *
2312 * Technical note: in 2.3 we work on _locked_ socket, so that
2313 * tricks with *seq access order and skb->users are not required.
2314 * Probably, code can be easily improved even more.
2315 */
2316
2317static int tcp_recvmsg_locked(struct sock *sk, struct msghdr *msg, size_t len,
2318 int flags, struct scm_timestamping_internal *tss,
2319 int *cmsg_flags)
2320{
2321 struct tcp_sock *tp = tcp_sk(sk);
2322 int copied = 0;
2323 u32 peek_seq;
2324 u32 *seq;
2325 unsigned long used;
2326 int err;
2327 int target; /* Read at least this many bytes */
2328 long timeo;
2329 struct sk_buff *skb, *last;
2330 u32 urg_hole = 0;
2331
2332 err = -ENOTCONN;
2333 if (sk->sk_state == TCP_LISTEN)
2334 goto out;
2335
2336 if (tp->recvmsg_inq) {
2337 *cmsg_flags = TCP_CMSG_INQ;
2338 msg->msg_get_inq = 1;
2339 }
2340 timeo = sock_rcvtimeo(sk, flags & MSG_DONTWAIT);
2341
2342 /* Urgent data needs to be handled specially. */
2343 if (flags & MSG_OOB)
2344 goto recv_urg;
2345
2346 if (unlikely(tp->repair)) {
2347 err = -EPERM;
2348 if (!(flags & MSG_PEEK))
2349 goto out;
2350
2351 if (tp->repair_queue == TCP_SEND_QUEUE)
2352 goto recv_sndq;
2353
2354 err = -EINVAL;
2355 if (tp->repair_queue == TCP_NO_QUEUE)
2356 goto out;
2357
2358 /* 'common' recv queue MSG_PEEK-ing */
2359 }
2360
2361 seq = &tp->copied_seq;
2362 if (flags & MSG_PEEK) {
2363 peek_seq = tp->copied_seq;
2364 seq = &peek_seq;
2365 }
2366
2367 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
2368
2369 do {
2370 u32 offset;
2371
2372 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
2373 if (unlikely(tp->urg_data) && tp->urg_seq == *seq) {
2374 if (copied)
2375 break;
2376 if (signal_pending(current)) {
2377 copied = timeo ? sock_intr_errno(timeo) : -EAGAIN;
2378 break;
2379 }
2380 }
2381
2382 /* Next get a buffer. */
2383
2384 last = skb_peek_tail(&sk->sk_receive_queue);
2385 skb_queue_walk(&sk->sk_receive_queue, skb) {
2386 last = skb;
2387 /* Now that we have two receive queues this
2388 * shouldn't happen.
2389 */
2390 if (WARN(before(*seq, TCP_SKB_CB(skb)->seq),
2391 "TCP recvmsg seq # bug: copied %X, seq %X, rcvnxt %X, fl %X\n",
2392 *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt,
2393 flags))
2394 break;
2395
2396 offset = *seq - TCP_SKB_CB(skb)->seq;
2397 if (unlikely(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
2398 pr_err_once("%s: found a SYN, please report !\n", __func__);
2399 offset--;
2400 }
2401 if (offset < skb->len)
2402 goto found_ok_skb;
2403 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
2404 goto found_fin_ok;
2405 WARN(!(flags & MSG_PEEK),
2406 "TCP recvmsg seq # bug 2: copied %X, seq %X, rcvnxt %X, fl %X\n",
2407 *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt, flags);
2408 }
2409
2410 /* Well, if we have backlog, try to process it now yet. */
2411
2412 if (copied >= target && !READ_ONCE(sk->sk_backlog.tail))
2413 break;
2414
2415 if (copied) {
2416 if (!timeo ||
2417 sk->sk_err ||
2418 sk->sk_state == TCP_CLOSE ||
2419 (sk->sk_shutdown & RCV_SHUTDOWN) ||
2420 signal_pending(current))
2421 break;
2422 } else {
2423 if (sock_flag(sk, SOCK_DONE))
2424 break;
2425
2426 if (sk->sk_err) {
2427 copied = sock_error(sk);
2428 break;
2429 }
2430
2431 if (sk->sk_shutdown & RCV_SHUTDOWN)
2432 break;
2433
2434 if (sk->sk_state == TCP_CLOSE) {
2435 /* This occurs when user tries to read
2436 * from never connected socket.
2437 */
2438 copied = -ENOTCONN;
2439 break;
2440 }
2441
2442 if (!timeo) {
2443 copied = -EAGAIN;
2444 break;
2445 }
2446
2447 if (signal_pending(current)) {
2448 copied = sock_intr_errno(timeo);
2449 break;
2450 }
2451 }
2452
2453 if (copied >= target) {
2454 /* Do not sleep, just process backlog. */
2455 __sk_flush_backlog(sk);
2456 } else {
2457 tcp_cleanup_rbuf(sk, copied);
2458 err = sk_wait_data(sk, &timeo, last);
2459 if (err < 0) {
2460 err = copied ? : err;
2461 goto out;
2462 }
2463 }
2464
2465 if ((flags & MSG_PEEK) &&
2466 (peek_seq - copied - urg_hole != tp->copied_seq)) {
2467 net_dbg_ratelimited("TCP(%s:%d): Application bug, race in MSG_PEEK\n",
2468 current->comm,
2469 task_pid_nr(current));
2470 peek_seq = tp->copied_seq;
2471 }
2472 continue;
2473
2474found_ok_skb:
2475 /* Ok so how much can we use? */
2476 used = skb->len - offset;
2477 if (len < used)
2478 used = len;
2479
2480 /* Do we have urgent data here? */
2481 if (unlikely(tp->urg_data)) {
2482 u32 urg_offset = tp->urg_seq - *seq;
2483 if (urg_offset < used) {
2484 if (!urg_offset) {
2485 if (!sock_flag(sk, SOCK_URGINLINE)) {
2486 WRITE_ONCE(*seq, *seq + 1);
2487 urg_hole++;
2488 offset++;
2489 used--;
2490 if (!used)
2491 goto skip_copy;
2492 }
2493 } else
2494 used = urg_offset;
2495 }
2496 }
2497
2498 if (!(flags & MSG_TRUNC)) {
2499 err = skb_copy_datagram_msg(skb, offset, msg, used);
2500 if (err) {
2501 /* Exception. Bailout! */
2502 if (!copied)
2503 copied = -EFAULT;
2504 break;
2505 }
2506 }
2507
2508 WRITE_ONCE(*seq, *seq + used);
2509 copied += used;
2510 len -= used;
2511
2512 tcp_rcv_space_adjust(sk);
2513
2514skip_copy:
2515 if (unlikely(tp->urg_data) && after(tp->copied_seq, tp->urg_seq)) {
2516 WRITE_ONCE(tp->urg_data, 0);
2517 tcp_fast_path_check(sk);
2518 }
2519
2520 if (TCP_SKB_CB(skb)->has_rxtstamp) {
2521 tcp_update_recv_tstamps(skb, tss);
2522 *cmsg_flags |= TCP_CMSG_TS;
2523 }
2524
2525 if (used + offset < skb->len)
2526 continue;
2527
2528 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
2529 goto found_fin_ok;
2530 if (!(flags & MSG_PEEK))
2531 tcp_eat_recv_skb(sk, skb);
2532 continue;
2533
2534found_fin_ok:
2535 /* Process the FIN. */
2536 WRITE_ONCE(*seq, *seq + 1);
2537 if (!(flags & MSG_PEEK))
2538 tcp_eat_recv_skb(sk, skb);
2539 break;
2540 } while (len > 0);
2541
2542 /* According to UNIX98, msg_name/msg_namelen are ignored
2543 * on connected socket. I was just happy when found this 8) --ANK
2544 */
2545
2546 /* Clean up data we have read: This will do ACK frames. */
2547 tcp_cleanup_rbuf(sk, copied);
2548 return copied;
2549
2550out:
2551 return err;
2552
2553recv_urg:
2554 err = tcp_recv_urg(sk, msg, len, flags);
2555 goto out;
2556
2557recv_sndq:
2558 err = tcp_peek_sndq(sk, msg, len);
2559 goto out;
2560}
2561
2562int tcp_recvmsg(struct sock *sk, struct msghdr *msg, size_t len, int flags,
2563 int *addr_len)
2564{
2565 int cmsg_flags = 0, ret;
2566 struct scm_timestamping_internal tss;
2567
2568 if (unlikely(flags & MSG_ERRQUEUE))
2569 return inet_recv_error(sk, msg, len, addr_len);
2570
2571 if (sk_can_busy_loop(sk) &&
2572 skb_queue_empty_lockless(&sk->sk_receive_queue) &&
2573 sk->sk_state == TCP_ESTABLISHED)
2574 sk_busy_loop(sk, flags & MSG_DONTWAIT);
2575
2576 lock_sock(sk);
2577 ret = tcp_recvmsg_locked(sk, msg, len, flags, &tss, &cmsg_flags);
2578 release_sock(sk);
2579
2580 if ((cmsg_flags || msg->msg_get_inq) && ret >= 0) {
2581 if (cmsg_flags & TCP_CMSG_TS)
2582 tcp_recv_timestamp(msg, sk, &tss);
2583 if (msg->msg_get_inq) {
2584 msg->msg_inq = tcp_inq_hint(sk);
2585 if (cmsg_flags & TCP_CMSG_INQ)
2586 put_cmsg(msg, SOL_TCP, TCP_CM_INQ,
2587 sizeof(msg->msg_inq), &msg->msg_inq);
2588 }
2589 }
2590 return ret;
2591}
2592EXPORT_SYMBOL(tcp_recvmsg);
2593
2594void tcp_set_state(struct sock *sk, int state)
2595{
2596 int oldstate = sk->sk_state;
2597
2598 /* We defined a new enum for TCP states that are exported in BPF
2599 * so as not force the internal TCP states to be frozen. The
2600 * following checks will detect if an internal state value ever
2601 * differs from the BPF value. If this ever happens, then we will
2602 * need to remap the internal value to the BPF value before calling
2603 * tcp_call_bpf_2arg.
2604 */
2605 BUILD_BUG_ON((int)BPF_TCP_ESTABLISHED != (int)TCP_ESTABLISHED);
2606 BUILD_BUG_ON((int)BPF_TCP_SYN_SENT != (int)TCP_SYN_SENT);
2607 BUILD_BUG_ON((int)BPF_TCP_SYN_RECV != (int)TCP_SYN_RECV);
2608 BUILD_BUG_ON((int)BPF_TCP_FIN_WAIT1 != (int)TCP_FIN_WAIT1);
2609 BUILD_BUG_ON((int)BPF_TCP_FIN_WAIT2 != (int)TCP_FIN_WAIT2);
2610 BUILD_BUG_ON((int)BPF_TCP_TIME_WAIT != (int)TCP_TIME_WAIT);
2611 BUILD_BUG_ON((int)BPF_TCP_CLOSE != (int)TCP_CLOSE);
2612 BUILD_BUG_ON((int)BPF_TCP_CLOSE_WAIT != (int)TCP_CLOSE_WAIT);
2613 BUILD_BUG_ON((int)BPF_TCP_LAST_ACK != (int)TCP_LAST_ACK);
2614 BUILD_BUG_ON((int)BPF_TCP_LISTEN != (int)TCP_LISTEN);
2615 BUILD_BUG_ON((int)BPF_TCP_CLOSING != (int)TCP_CLOSING);
2616 BUILD_BUG_ON((int)BPF_TCP_NEW_SYN_RECV != (int)TCP_NEW_SYN_RECV);
2617 BUILD_BUG_ON((int)BPF_TCP_BOUND_INACTIVE != (int)TCP_BOUND_INACTIVE);
2618 BUILD_BUG_ON((int)BPF_TCP_MAX_STATES != (int)TCP_MAX_STATES);
2619
2620 /* bpf uapi header bpf.h defines an anonymous enum with values
2621 * BPF_TCP_* used by bpf programs. Currently gcc built vmlinux
2622 * is able to emit this enum in DWARF due to the above BUILD_BUG_ON.
2623 * But clang built vmlinux does not have this enum in DWARF
2624 * since clang removes the above code before generating IR/debuginfo.
2625 * Let us explicitly emit the type debuginfo to ensure the
2626 * above-mentioned anonymous enum in the vmlinux DWARF and hence BTF
2627 * regardless of which compiler is used.
2628 */
2629 BTF_TYPE_EMIT_ENUM(BPF_TCP_ESTABLISHED);
2630
2631 if (BPF_SOCK_OPS_TEST_FLAG(tcp_sk(sk), BPF_SOCK_OPS_STATE_CB_FLAG))
2632 tcp_call_bpf_2arg(sk, BPF_SOCK_OPS_STATE_CB, oldstate, state);
2633
2634 switch (state) {
2635 case TCP_ESTABLISHED:
2636 if (oldstate != TCP_ESTABLISHED)
2637 TCP_INC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
2638 break;
2639
2640 case TCP_CLOSE:
2641 if (oldstate == TCP_CLOSE_WAIT || oldstate == TCP_ESTABLISHED)
2642 TCP_INC_STATS(sock_net(sk), TCP_MIB_ESTABRESETS);
2643
2644 sk->sk_prot->unhash(sk);
2645 if (inet_csk(sk)->icsk_bind_hash &&
2646 !(sk->sk_userlocks & SOCK_BINDPORT_LOCK))
2647 inet_put_port(sk);
2648 fallthrough;
2649 default:
2650 if (oldstate == TCP_ESTABLISHED)
2651 TCP_DEC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
2652 }
2653
2654 /* Change state AFTER socket is unhashed to avoid closed
2655 * socket sitting in hash tables.
2656 */
2657 inet_sk_state_store(sk, state);
2658}
2659EXPORT_SYMBOL_GPL(tcp_set_state);
2660
2661/*
2662 * State processing on a close. This implements the state shift for
2663 * sending our FIN frame. Note that we only send a FIN for some
2664 * states. A shutdown() may have already sent the FIN, or we may be
2665 * closed.
2666 */
2667
2668static const unsigned char new_state[16] = {
2669 /* current state: new state: action: */
2670 [0 /* (Invalid) */] = TCP_CLOSE,
2671 [TCP_ESTABLISHED] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2672 [TCP_SYN_SENT] = TCP_CLOSE,
2673 [TCP_SYN_RECV] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2674 [TCP_FIN_WAIT1] = TCP_FIN_WAIT1,
2675 [TCP_FIN_WAIT2] = TCP_FIN_WAIT2,
2676 [TCP_TIME_WAIT] = TCP_CLOSE,
2677 [TCP_CLOSE] = TCP_CLOSE,
2678 [TCP_CLOSE_WAIT] = TCP_LAST_ACK | TCP_ACTION_FIN,
2679 [TCP_LAST_ACK] = TCP_LAST_ACK,
2680 [TCP_LISTEN] = TCP_CLOSE,
2681 [TCP_CLOSING] = TCP_CLOSING,
2682 [TCP_NEW_SYN_RECV] = TCP_CLOSE, /* should not happen ! */
2683};
2684
2685static int tcp_close_state(struct sock *sk)
2686{
2687 int next = (int)new_state[sk->sk_state];
2688 int ns = next & TCP_STATE_MASK;
2689
2690 tcp_set_state(sk, ns);
2691
2692 return next & TCP_ACTION_FIN;
2693}
2694
2695/*
2696 * Shutdown the sending side of a connection. Much like close except
2697 * that we don't receive shut down or sock_set_flag(sk, SOCK_DEAD).
2698 */
2699
2700void tcp_shutdown(struct sock *sk, int how)
2701{
2702 /* We need to grab some memory, and put together a FIN,
2703 * and then put it into the queue to be sent.
2704 * Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
2705 */
2706 if (!(how & SEND_SHUTDOWN))
2707 return;
2708
2709 /* If we've already sent a FIN, or it's a closed state, skip this. */
2710 if ((1 << sk->sk_state) &
2711 (TCPF_ESTABLISHED | TCPF_SYN_SENT |
2712 TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) {
2713 /* Clear out any half completed packets. FIN if needed. */
2714 if (tcp_close_state(sk))
2715 tcp_send_fin(sk);
2716 }
2717}
2718EXPORT_SYMBOL(tcp_shutdown);
2719
2720int tcp_orphan_count_sum(void)
2721{
2722 int i, total = 0;
2723
2724 for_each_possible_cpu(i)
2725 total += per_cpu(tcp_orphan_count, i);
2726
2727 return max(total, 0);
2728}
2729
2730static int tcp_orphan_cache;
2731static struct timer_list tcp_orphan_timer;
2732#define TCP_ORPHAN_TIMER_PERIOD msecs_to_jiffies(100)
2733
2734static void tcp_orphan_update(struct timer_list *unused)
2735{
2736 WRITE_ONCE(tcp_orphan_cache, tcp_orphan_count_sum());
2737 mod_timer(&tcp_orphan_timer, jiffies + TCP_ORPHAN_TIMER_PERIOD);
2738}
2739
2740static bool tcp_too_many_orphans(int shift)
2741{
2742 return READ_ONCE(tcp_orphan_cache) << shift >
2743 READ_ONCE(sysctl_tcp_max_orphans);
2744}
2745
2746bool tcp_check_oom(struct sock *sk, int shift)
2747{
2748 bool too_many_orphans, out_of_socket_memory;
2749
2750 too_many_orphans = tcp_too_many_orphans(shift);
2751 out_of_socket_memory = tcp_out_of_memory(sk);
2752
2753 if (too_many_orphans)
2754 net_info_ratelimited("too many orphaned sockets\n");
2755 if (out_of_socket_memory)
2756 net_info_ratelimited("out of memory -- consider tuning tcp_mem\n");
2757 return too_many_orphans || out_of_socket_memory;
2758}
2759
2760void __tcp_close(struct sock *sk, long timeout)
2761{
2762 struct sk_buff *skb;
2763 int data_was_unread = 0;
2764 int state;
2765
2766 WRITE_ONCE(sk->sk_shutdown, SHUTDOWN_MASK);
2767
2768 if (sk->sk_state == TCP_LISTEN) {
2769 tcp_set_state(sk, TCP_CLOSE);
2770
2771 /* Special case. */
2772 inet_csk_listen_stop(sk);
2773
2774 goto adjudge_to_death;
2775 }
2776
2777 /* We need to flush the recv. buffs. We do this only on the
2778 * descriptor close, not protocol-sourced closes, because the
2779 * reader process may not have drained the data yet!
2780 */
2781 while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) {
2782 u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq;
2783
2784 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
2785 len--;
2786 data_was_unread += len;
2787 __kfree_skb(skb);
2788 }
2789
2790 /* If socket has been already reset (e.g. in tcp_reset()) - kill it. */
2791 if (sk->sk_state == TCP_CLOSE)
2792 goto adjudge_to_death;
2793
2794 /* As outlined in RFC 2525, section 2.17, we send a RST here because
2795 * data was lost. To witness the awful effects of the old behavior of
2796 * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk
2797 * GET in an FTP client, suspend the process, wait for the client to
2798 * advertise a zero window, then kill -9 the FTP client, wheee...
2799 * Note: timeout is always zero in such a case.
2800 */
2801 if (unlikely(tcp_sk(sk)->repair)) {
2802 sk->sk_prot->disconnect(sk, 0);
2803 } else if (data_was_unread) {
2804 /* Unread data was tossed, zap the connection. */
2805 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONCLOSE);
2806 tcp_set_state(sk, TCP_CLOSE);
2807 tcp_send_active_reset(sk, sk->sk_allocation);
2808 } else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) {
2809 /* Check zero linger _after_ checking for unread data. */
2810 sk->sk_prot->disconnect(sk, 0);
2811 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
2812 } else if (tcp_close_state(sk)) {
2813 /* We FIN if the application ate all the data before
2814 * zapping the connection.
2815 */
2816
2817 /* RED-PEN. Formally speaking, we have broken TCP state
2818 * machine. State transitions:
2819 *
2820 * TCP_ESTABLISHED -> TCP_FIN_WAIT1
2821 * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible)
2822 * TCP_CLOSE_WAIT -> TCP_LAST_ACK
2823 *
2824 * are legal only when FIN has been sent (i.e. in window),
2825 * rather than queued out of window. Purists blame.
2826 *
2827 * F.e. "RFC state" is ESTABLISHED,
2828 * if Linux state is FIN-WAIT-1, but FIN is still not sent.
2829 *
2830 * The visible declinations are that sometimes
2831 * we enter time-wait state, when it is not required really
2832 * (harmless), do not send active resets, when they are
2833 * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
2834 * they look as CLOSING or LAST_ACK for Linux)
2835 * Probably, I missed some more holelets.
2836 * --ANK
2837 * XXX (TFO) - To start off we don't support SYN+ACK+FIN
2838 * in a single packet! (May consider it later but will
2839 * probably need API support or TCP_CORK SYN-ACK until
2840 * data is written and socket is closed.)
2841 */
2842 tcp_send_fin(sk);
2843 }
2844
2845 sk_stream_wait_close(sk, timeout);
2846
2847adjudge_to_death:
2848 state = sk->sk_state;
2849 sock_hold(sk);
2850 sock_orphan(sk);
2851
2852 local_bh_disable();
2853 bh_lock_sock(sk);
2854 /* remove backlog if any, without releasing ownership. */
2855 __release_sock(sk);
2856
2857 this_cpu_inc(tcp_orphan_count);
2858
2859 /* Have we already been destroyed by a softirq or backlog? */
2860 if (state != TCP_CLOSE && sk->sk_state == TCP_CLOSE)
2861 goto out;
2862
2863 /* This is a (useful) BSD violating of the RFC. There is a
2864 * problem with TCP as specified in that the other end could
2865 * keep a socket open forever with no application left this end.
2866 * We use a 1 minute timeout (about the same as BSD) then kill
2867 * our end. If they send after that then tough - BUT: long enough
2868 * that we won't make the old 4*rto = almost no time - whoops
2869 * reset mistake.
2870 *
2871 * Nope, it was not mistake. It is really desired behaviour
2872 * f.e. on http servers, when such sockets are useless, but
2873 * consume significant resources. Let's do it with special
2874 * linger2 option. --ANK
2875 */
2876
2877 if (sk->sk_state == TCP_FIN_WAIT2) {
2878 struct tcp_sock *tp = tcp_sk(sk);
2879 if (READ_ONCE(tp->linger2) < 0) {
2880 tcp_set_state(sk, TCP_CLOSE);
2881 tcp_send_active_reset(sk, GFP_ATOMIC);
2882 __NET_INC_STATS(sock_net(sk),
2883 LINUX_MIB_TCPABORTONLINGER);
2884 } else {
2885 const int tmo = tcp_fin_time(sk);
2886
2887 if (tmo > TCP_TIMEWAIT_LEN) {
2888 inet_csk_reset_keepalive_timer(sk,
2889 tmo - TCP_TIMEWAIT_LEN);
2890 } else {
2891 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
2892 goto out;
2893 }
2894 }
2895 }
2896 if (sk->sk_state != TCP_CLOSE) {
2897 if (tcp_check_oom(sk, 0)) {
2898 tcp_set_state(sk, TCP_CLOSE);
2899 tcp_send_active_reset(sk, GFP_ATOMIC);
2900 __NET_INC_STATS(sock_net(sk),
2901 LINUX_MIB_TCPABORTONMEMORY);
2902 } else if (!check_net(sock_net(sk))) {
2903 /* Not possible to send reset; just close */
2904 tcp_set_state(sk, TCP_CLOSE);
2905 }
2906 }
2907
2908 if (sk->sk_state == TCP_CLOSE) {
2909 struct request_sock *req;
2910
2911 req = rcu_dereference_protected(tcp_sk(sk)->fastopen_rsk,
2912 lockdep_sock_is_held(sk));
2913 /* We could get here with a non-NULL req if the socket is
2914 * aborted (e.g., closed with unread data) before 3WHS
2915 * finishes.
2916 */
2917 if (req)
2918 reqsk_fastopen_remove(sk, req, false);
2919 inet_csk_destroy_sock(sk);
2920 }
2921 /* Otherwise, socket is reprieved until protocol close. */
2922
2923out:
2924 bh_unlock_sock(sk);
2925 local_bh_enable();
2926}
2927
2928void tcp_close(struct sock *sk, long timeout)
2929{
2930 lock_sock(sk);
2931 __tcp_close(sk, timeout);
2932 release_sock(sk);
2933 sock_put(sk);
2934}
2935EXPORT_SYMBOL(tcp_close);
2936
2937/* These states need RST on ABORT according to RFC793 */
2938
2939static inline bool tcp_need_reset(int state)
2940{
2941 return (1 << state) &
2942 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 |
2943 TCPF_FIN_WAIT2 | TCPF_SYN_RECV);
2944}
2945
2946static void tcp_rtx_queue_purge(struct sock *sk)
2947{
2948 struct rb_node *p = rb_first(&sk->tcp_rtx_queue);
2949
2950 tcp_sk(sk)->highest_sack = NULL;
2951 while (p) {
2952 struct sk_buff *skb = rb_to_skb(p);
2953
2954 p = rb_next(p);
2955 /* Since we are deleting whole queue, no need to
2956 * list_del(&skb->tcp_tsorted_anchor)
2957 */
2958 tcp_rtx_queue_unlink(skb, sk);
2959 tcp_wmem_free_skb(sk, skb);
2960 }
2961}
2962
2963void tcp_write_queue_purge(struct sock *sk)
2964{
2965 struct sk_buff *skb;
2966
2967 tcp_chrono_stop(sk, TCP_CHRONO_BUSY);
2968 while ((skb = __skb_dequeue(&sk->sk_write_queue)) != NULL) {
2969 tcp_skb_tsorted_anchor_cleanup(skb);
2970 tcp_wmem_free_skb(sk, skb);
2971 }
2972 tcp_rtx_queue_purge(sk);
2973 INIT_LIST_HEAD(&tcp_sk(sk)->tsorted_sent_queue);
2974 tcp_clear_all_retrans_hints(tcp_sk(sk));
2975 tcp_sk(sk)->packets_out = 0;
2976 inet_csk(sk)->icsk_backoff = 0;
2977}
2978
2979int tcp_disconnect(struct sock *sk, int flags)
2980{
2981 struct inet_sock *inet = inet_sk(sk);
2982 struct inet_connection_sock *icsk = inet_csk(sk);
2983 struct tcp_sock *tp = tcp_sk(sk);
2984 int old_state = sk->sk_state;
2985 u32 seq;
2986
2987 if (old_state != TCP_CLOSE)
2988 tcp_set_state(sk, TCP_CLOSE);
2989
2990 /* ABORT function of RFC793 */
2991 if (old_state == TCP_LISTEN) {
2992 inet_csk_listen_stop(sk);
2993 } else if (unlikely(tp->repair)) {
2994 WRITE_ONCE(sk->sk_err, ECONNABORTED);
2995 } else if (tcp_need_reset(old_state) ||
2996 (tp->snd_nxt != tp->write_seq &&
2997 (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK))) {
2998 /* The last check adjusts for discrepancy of Linux wrt. RFC
2999 * states
3000 */
3001 tcp_send_active_reset(sk, gfp_any());
3002 WRITE_ONCE(sk->sk_err, ECONNRESET);
3003 } else if (old_state == TCP_SYN_SENT)
3004 WRITE_ONCE(sk->sk_err, ECONNRESET);
3005
3006 tcp_clear_xmit_timers(sk);
3007 __skb_queue_purge(&sk->sk_receive_queue);
3008 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt);
3009 WRITE_ONCE(tp->urg_data, 0);
3010 tcp_write_queue_purge(sk);
3011 tcp_fastopen_active_disable_ofo_check(sk);
3012 skb_rbtree_purge(&tp->out_of_order_queue);
3013
3014 inet->inet_dport = 0;
3015
3016 inet_bhash2_reset_saddr(sk);
3017
3018 WRITE_ONCE(sk->sk_shutdown, 0);
3019 sock_reset_flag(sk, SOCK_DONE);
3020 tp->srtt_us = 0;
3021 tp->mdev_us = jiffies_to_usecs(TCP_TIMEOUT_INIT);
3022 tp->rcv_rtt_last_tsecr = 0;
3023
3024 seq = tp->write_seq + tp->max_window + 2;
3025 if (!seq)
3026 seq = 1;
3027 WRITE_ONCE(tp->write_seq, seq);
3028
3029 icsk->icsk_backoff = 0;
3030 icsk->icsk_probes_out = 0;
3031 icsk->icsk_probes_tstamp = 0;
3032 icsk->icsk_rto = TCP_TIMEOUT_INIT;
3033 icsk->icsk_rto_min = TCP_RTO_MIN;
3034 icsk->icsk_delack_max = TCP_DELACK_MAX;
3035 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
3036 tcp_snd_cwnd_set(tp, TCP_INIT_CWND);
3037 tp->snd_cwnd_cnt = 0;
3038 tp->is_cwnd_limited = 0;
3039 tp->max_packets_out = 0;
3040 tp->window_clamp = 0;
3041 tp->delivered = 0;
3042 tp->delivered_ce = 0;
3043 if (icsk->icsk_ca_ops->release)
3044 icsk->icsk_ca_ops->release(sk);
3045 memset(icsk->icsk_ca_priv, 0, sizeof(icsk->icsk_ca_priv));
3046 icsk->icsk_ca_initialized = 0;
3047 tcp_set_ca_state(sk, TCP_CA_Open);
3048 tp->is_sack_reneg = 0;
3049 tcp_clear_retrans(tp);
3050 tp->total_retrans = 0;
3051 inet_csk_delack_init(sk);
3052 /* Initialize rcv_mss to TCP_MIN_MSS to avoid division by 0
3053 * issue in __tcp_select_window()
3054 */
3055 icsk->icsk_ack.rcv_mss = TCP_MIN_MSS;
3056 memset(&tp->rx_opt, 0, sizeof(tp->rx_opt));
3057 __sk_dst_reset(sk);
3058 dst_release(xchg((__force struct dst_entry **)&sk->sk_rx_dst, NULL));
3059 tcp_saved_syn_free(tp);
3060 tp->compressed_ack = 0;
3061 tp->segs_in = 0;
3062 tp->segs_out = 0;
3063 tp->bytes_sent = 0;
3064 tp->bytes_acked = 0;
3065 tp->bytes_received = 0;
3066 tp->bytes_retrans = 0;
3067 tp->data_segs_in = 0;
3068 tp->data_segs_out = 0;
3069 tp->duplicate_sack[0].start_seq = 0;
3070 tp->duplicate_sack[0].end_seq = 0;
3071 tp->dsack_dups = 0;
3072 tp->reord_seen = 0;
3073 tp->retrans_out = 0;
3074 tp->sacked_out = 0;
3075 tp->tlp_high_seq = 0;
3076 tp->last_oow_ack_time = 0;
3077 tp->plb_rehash = 0;
3078 /* There's a bubble in the pipe until at least the first ACK. */
3079 tp->app_limited = ~0U;
3080 tp->rate_app_limited = 1;
3081 tp->rack.mstamp = 0;
3082 tp->rack.advanced = 0;
3083 tp->rack.reo_wnd_steps = 1;
3084 tp->rack.last_delivered = 0;
3085 tp->rack.reo_wnd_persist = 0;
3086 tp->rack.dsack_seen = 0;
3087 tp->syn_data_acked = 0;
3088 tp->rx_opt.saw_tstamp = 0;
3089 tp->rx_opt.dsack = 0;
3090 tp->rx_opt.num_sacks = 0;
3091 tp->rcv_ooopack = 0;
3092
3093
3094 /* Clean up fastopen related fields */
3095 tcp_free_fastopen_req(tp);
3096 inet_clear_bit(DEFER_CONNECT, sk);
3097 tp->fastopen_client_fail = 0;
3098
3099 WARN_ON(inet->inet_num && !icsk->icsk_bind_hash);
3100
3101 if (sk->sk_frag.page) {
3102 put_page(sk->sk_frag.page);
3103 sk->sk_frag.page = NULL;
3104 sk->sk_frag.offset = 0;
3105 }
3106 sk_error_report(sk);
3107 return 0;
3108}
3109EXPORT_SYMBOL(tcp_disconnect);
3110
3111static inline bool tcp_can_repair_sock(const struct sock *sk)
3112{
3113 return sockopt_ns_capable(sock_net(sk)->user_ns, CAP_NET_ADMIN) &&
3114 (sk->sk_state != TCP_LISTEN);
3115}
3116
3117static int tcp_repair_set_window(struct tcp_sock *tp, sockptr_t optbuf, int len)
3118{
3119 struct tcp_repair_window opt;
3120
3121 if (!tp->repair)
3122 return -EPERM;
3123
3124 if (len != sizeof(opt))
3125 return -EINVAL;
3126
3127 if (copy_from_sockptr(&opt, optbuf, sizeof(opt)))
3128 return -EFAULT;
3129
3130 if (opt.max_window < opt.snd_wnd)
3131 return -EINVAL;
3132
3133 if (after(opt.snd_wl1, tp->rcv_nxt + opt.rcv_wnd))
3134 return -EINVAL;
3135
3136 if (after(opt.rcv_wup, tp->rcv_nxt))
3137 return -EINVAL;
3138
3139 tp->snd_wl1 = opt.snd_wl1;
3140 tp->snd_wnd = opt.snd_wnd;
3141 tp->max_window = opt.max_window;
3142
3143 tp->rcv_wnd = opt.rcv_wnd;
3144 tp->rcv_wup = opt.rcv_wup;
3145
3146 return 0;
3147}
3148
3149static int tcp_repair_options_est(struct sock *sk, sockptr_t optbuf,
3150 unsigned int len)
3151{
3152 struct tcp_sock *tp = tcp_sk(sk);
3153 struct tcp_repair_opt opt;
3154 size_t offset = 0;
3155
3156 while (len >= sizeof(opt)) {
3157 if (copy_from_sockptr_offset(&opt, optbuf, offset, sizeof(opt)))
3158 return -EFAULT;
3159
3160 offset += sizeof(opt);
3161 len -= sizeof(opt);
3162
3163 switch (opt.opt_code) {
3164 case TCPOPT_MSS:
3165 tp->rx_opt.mss_clamp = opt.opt_val;
3166 tcp_mtup_init(sk);
3167 break;
3168 case TCPOPT_WINDOW:
3169 {
3170 u16 snd_wscale = opt.opt_val & 0xFFFF;
3171 u16 rcv_wscale = opt.opt_val >> 16;
3172
3173 if (snd_wscale > TCP_MAX_WSCALE || rcv_wscale > TCP_MAX_WSCALE)
3174 return -EFBIG;
3175
3176 tp->rx_opt.snd_wscale = snd_wscale;
3177 tp->rx_opt.rcv_wscale = rcv_wscale;
3178 tp->rx_opt.wscale_ok = 1;
3179 }
3180 break;
3181 case TCPOPT_SACK_PERM:
3182 if (opt.opt_val != 0)
3183 return -EINVAL;
3184
3185 tp->rx_opt.sack_ok |= TCP_SACK_SEEN;
3186 break;
3187 case TCPOPT_TIMESTAMP:
3188 if (opt.opt_val != 0)
3189 return -EINVAL;
3190
3191 tp->rx_opt.tstamp_ok = 1;
3192 break;
3193 }
3194 }
3195
3196 return 0;
3197}
3198
3199DEFINE_STATIC_KEY_FALSE(tcp_tx_delay_enabled);
3200EXPORT_SYMBOL(tcp_tx_delay_enabled);
3201
3202static void tcp_enable_tx_delay(void)
3203{
3204 if (!static_branch_unlikely(&tcp_tx_delay_enabled)) {
3205 static int __tcp_tx_delay_enabled = 0;
3206
3207 if (cmpxchg(&__tcp_tx_delay_enabled, 0, 1) == 0) {
3208 static_branch_enable(&tcp_tx_delay_enabled);
3209 pr_info("TCP_TX_DELAY enabled\n");
3210 }
3211 }
3212}
3213
3214/* When set indicates to always queue non-full frames. Later the user clears
3215 * this option and we transmit any pending partial frames in the queue. This is
3216 * meant to be used alongside sendfile() to get properly filled frames when the
3217 * user (for example) must write out headers with a write() call first and then
3218 * use sendfile to send out the data parts.
3219 *
3220 * TCP_CORK can be set together with TCP_NODELAY and it is stronger than
3221 * TCP_NODELAY.
3222 */
3223void __tcp_sock_set_cork(struct sock *sk, bool on)
3224{
3225 struct tcp_sock *tp = tcp_sk(sk);
3226
3227 if (on) {
3228 tp->nonagle |= TCP_NAGLE_CORK;
3229 } else {
3230 tp->nonagle &= ~TCP_NAGLE_CORK;
3231 if (tp->nonagle & TCP_NAGLE_OFF)
3232 tp->nonagle |= TCP_NAGLE_PUSH;
3233 tcp_push_pending_frames(sk);
3234 }
3235}
3236
3237void tcp_sock_set_cork(struct sock *sk, bool on)
3238{
3239 lock_sock(sk);
3240 __tcp_sock_set_cork(sk, on);
3241 release_sock(sk);
3242}
3243EXPORT_SYMBOL(tcp_sock_set_cork);
3244
3245/* TCP_NODELAY is weaker than TCP_CORK, so that this option on corked socket is
3246 * remembered, but it is not activated until cork is cleared.
3247 *
3248 * However, when TCP_NODELAY is set we make an explicit push, which overrides
3249 * even TCP_CORK for currently queued segments.
3250 */
3251void __tcp_sock_set_nodelay(struct sock *sk, bool on)
3252{
3253 if (on) {
3254 tcp_sk(sk)->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH;
3255 tcp_push_pending_frames(sk);
3256 } else {
3257 tcp_sk(sk)->nonagle &= ~TCP_NAGLE_OFF;
3258 }
3259}
3260
3261void tcp_sock_set_nodelay(struct sock *sk)
3262{
3263 lock_sock(sk);
3264 __tcp_sock_set_nodelay(sk, true);
3265 release_sock(sk);
3266}
3267EXPORT_SYMBOL(tcp_sock_set_nodelay);
3268
3269static void __tcp_sock_set_quickack(struct sock *sk, int val)
3270{
3271 if (!val) {
3272 inet_csk_enter_pingpong_mode(sk);
3273 return;
3274 }
3275
3276 inet_csk_exit_pingpong_mode(sk);
3277 if ((1 << sk->sk_state) & (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) &&
3278 inet_csk_ack_scheduled(sk)) {
3279 inet_csk(sk)->icsk_ack.pending |= ICSK_ACK_PUSHED;
3280 tcp_cleanup_rbuf(sk, 1);
3281 if (!(val & 1))
3282 inet_csk_enter_pingpong_mode(sk);
3283 }
3284}
3285
3286void tcp_sock_set_quickack(struct sock *sk, int val)
3287{
3288 lock_sock(sk);
3289 __tcp_sock_set_quickack(sk, val);
3290 release_sock(sk);
3291}
3292EXPORT_SYMBOL(tcp_sock_set_quickack);
3293
3294int tcp_sock_set_syncnt(struct sock *sk, int val)
3295{
3296 if (val < 1 || val > MAX_TCP_SYNCNT)
3297 return -EINVAL;
3298
3299 WRITE_ONCE(inet_csk(sk)->icsk_syn_retries, val);
3300 return 0;
3301}
3302EXPORT_SYMBOL(tcp_sock_set_syncnt);
3303
3304int tcp_sock_set_user_timeout(struct sock *sk, int val)
3305{
3306 /* Cap the max time in ms TCP will retry or probe the window
3307 * before giving up and aborting (ETIMEDOUT) a connection.
3308 */
3309 if (val < 0)
3310 return -EINVAL;
3311
3312 WRITE_ONCE(inet_csk(sk)->icsk_user_timeout, val);
3313 return 0;
3314}
3315EXPORT_SYMBOL(tcp_sock_set_user_timeout);
3316
3317int tcp_sock_set_keepidle_locked(struct sock *sk, int val)
3318{
3319 struct tcp_sock *tp = tcp_sk(sk);
3320
3321 if (val < 1 || val > MAX_TCP_KEEPIDLE)
3322 return -EINVAL;
3323
3324 /* Paired with WRITE_ONCE() in keepalive_time_when() */
3325 WRITE_ONCE(tp->keepalive_time, val * HZ);
3326 if (sock_flag(sk, SOCK_KEEPOPEN) &&
3327 !((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN))) {
3328 u32 elapsed = keepalive_time_elapsed(tp);
3329
3330 if (tp->keepalive_time > elapsed)
3331 elapsed = tp->keepalive_time - elapsed;
3332 else
3333 elapsed = 0;
3334 inet_csk_reset_keepalive_timer(sk, elapsed);
3335 }
3336
3337 return 0;
3338}
3339
3340int tcp_sock_set_keepidle(struct sock *sk, int val)
3341{
3342 int err;
3343
3344 lock_sock(sk);
3345 err = tcp_sock_set_keepidle_locked(sk, val);
3346 release_sock(sk);
3347 return err;
3348}
3349EXPORT_SYMBOL(tcp_sock_set_keepidle);
3350
3351int tcp_sock_set_keepintvl(struct sock *sk, int val)
3352{
3353 if (val < 1 || val > MAX_TCP_KEEPINTVL)
3354 return -EINVAL;
3355
3356 WRITE_ONCE(tcp_sk(sk)->keepalive_intvl, val * HZ);
3357 return 0;
3358}
3359EXPORT_SYMBOL(tcp_sock_set_keepintvl);
3360
3361int tcp_sock_set_keepcnt(struct sock *sk, int val)
3362{
3363 if (val < 1 || val > MAX_TCP_KEEPCNT)
3364 return -EINVAL;
3365
3366 /* Paired with READ_ONCE() in keepalive_probes() */
3367 WRITE_ONCE(tcp_sk(sk)->keepalive_probes, val);
3368 return 0;
3369}
3370EXPORT_SYMBOL(tcp_sock_set_keepcnt);
3371
3372int tcp_set_window_clamp(struct sock *sk, int val)
3373{
3374 struct tcp_sock *tp = tcp_sk(sk);
3375
3376 if (!val) {
3377 if (sk->sk_state != TCP_CLOSE)
3378 return -EINVAL;
3379 tp->window_clamp = 0;
3380 } else {
3381 u32 new_rcv_ssthresh, old_window_clamp = tp->window_clamp;
3382 u32 new_window_clamp = val < SOCK_MIN_RCVBUF / 2 ?
3383 SOCK_MIN_RCVBUF / 2 : val;
3384
3385 if (new_window_clamp == old_window_clamp)
3386 return 0;
3387
3388 tp->window_clamp = new_window_clamp;
3389 if (new_window_clamp < old_window_clamp) {
3390 /* need to apply the reserved mem provisioning only
3391 * when shrinking the window clamp
3392 */
3393 __tcp_adjust_rcv_ssthresh(sk, tp->window_clamp);
3394
3395 } else {
3396 new_rcv_ssthresh = min(tp->rcv_wnd, tp->window_clamp);
3397 tp->rcv_ssthresh = max(new_rcv_ssthresh,
3398 tp->rcv_ssthresh);
3399 }
3400 }
3401 return 0;
3402}
3403
3404/*
3405 * Socket option code for TCP.
3406 */
3407int do_tcp_setsockopt(struct sock *sk, int level, int optname,
3408 sockptr_t optval, unsigned int optlen)
3409{
3410 struct tcp_sock *tp = tcp_sk(sk);
3411 struct inet_connection_sock *icsk = inet_csk(sk);
3412 struct net *net = sock_net(sk);
3413 int val;
3414 int err = 0;
3415
3416 /* These are data/string values, all the others are ints */
3417 switch (optname) {
3418 case TCP_CONGESTION: {
3419 char name[TCP_CA_NAME_MAX];
3420
3421 if (optlen < 1)
3422 return -EINVAL;
3423
3424 val = strncpy_from_sockptr(name, optval,
3425 min_t(long, TCP_CA_NAME_MAX-1, optlen));
3426 if (val < 0)
3427 return -EFAULT;
3428 name[val] = 0;
3429
3430 sockopt_lock_sock(sk);
3431 err = tcp_set_congestion_control(sk, name, !has_current_bpf_ctx(),
3432 sockopt_ns_capable(sock_net(sk)->user_ns,
3433 CAP_NET_ADMIN));
3434 sockopt_release_sock(sk);
3435 return err;
3436 }
3437 case TCP_ULP: {
3438 char name[TCP_ULP_NAME_MAX];
3439
3440 if (optlen < 1)
3441 return -EINVAL;
3442
3443 val = strncpy_from_sockptr(name, optval,
3444 min_t(long, TCP_ULP_NAME_MAX - 1,
3445 optlen));
3446 if (val < 0)
3447 return -EFAULT;
3448 name[val] = 0;
3449
3450 sockopt_lock_sock(sk);
3451 err = tcp_set_ulp(sk, name);
3452 sockopt_release_sock(sk);
3453 return err;
3454 }
3455 case TCP_FASTOPEN_KEY: {
3456 __u8 key[TCP_FASTOPEN_KEY_BUF_LENGTH];
3457 __u8 *backup_key = NULL;
3458
3459 /* Allow a backup key as well to facilitate key rotation
3460 * First key is the active one.
3461 */
3462 if (optlen != TCP_FASTOPEN_KEY_LENGTH &&
3463 optlen != TCP_FASTOPEN_KEY_BUF_LENGTH)
3464 return -EINVAL;
3465
3466 if (copy_from_sockptr(key, optval, optlen))
3467 return -EFAULT;
3468
3469 if (optlen == TCP_FASTOPEN_KEY_BUF_LENGTH)
3470 backup_key = key + TCP_FASTOPEN_KEY_LENGTH;
3471
3472 return tcp_fastopen_reset_cipher(net, sk, key, backup_key);
3473 }
3474 default:
3475 /* fallthru */
3476 break;
3477 }
3478
3479 if (optlen < sizeof(int))
3480 return -EINVAL;
3481
3482 if (copy_from_sockptr(&val, optval, sizeof(val)))
3483 return -EFAULT;
3484
3485 /* Handle options that can be set without locking the socket. */
3486 switch (optname) {
3487 case TCP_SYNCNT:
3488 return tcp_sock_set_syncnt(sk, val);
3489 case TCP_USER_TIMEOUT:
3490 return tcp_sock_set_user_timeout(sk, val);
3491 case TCP_KEEPINTVL:
3492 return tcp_sock_set_keepintvl(sk, val);
3493 case TCP_KEEPCNT:
3494 return tcp_sock_set_keepcnt(sk, val);
3495 case TCP_LINGER2:
3496 if (val < 0)
3497 WRITE_ONCE(tp->linger2, -1);
3498 else if (val > TCP_FIN_TIMEOUT_MAX / HZ)
3499 WRITE_ONCE(tp->linger2, TCP_FIN_TIMEOUT_MAX);
3500 else
3501 WRITE_ONCE(tp->linger2, val * HZ);
3502 return 0;
3503 case TCP_DEFER_ACCEPT:
3504 /* Translate value in seconds to number of retransmits */
3505 WRITE_ONCE(icsk->icsk_accept_queue.rskq_defer_accept,
3506 secs_to_retrans(val, TCP_TIMEOUT_INIT / HZ,
3507 TCP_RTO_MAX / HZ));
3508 return 0;
3509 }
3510
3511 sockopt_lock_sock(sk);
3512
3513 switch (optname) {
3514 case TCP_MAXSEG:
3515 /* Values greater than interface MTU won't take effect. However
3516 * at the point when this call is done we typically don't yet
3517 * know which interface is going to be used
3518 */
3519 if (val && (val < TCP_MIN_MSS || val > MAX_TCP_WINDOW)) {
3520 err = -EINVAL;
3521 break;
3522 }
3523 tp->rx_opt.user_mss = val;
3524 break;
3525
3526 case TCP_NODELAY:
3527 __tcp_sock_set_nodelay(sk, val);
3528 break;
3529
3530 case TCP_THIN_LINEAR_TIMEOUTS:
3531 if (val < 0 || val > 1)
3532 err = -EINVAL;
3533 else
3534 tp->thin_lto = val;
3535 break;
3536
3537 case TCP_THIN_DUPACK:
3538 if (val < 0 || val > 1)
3539 err = -EINVAL;
3540 break;
3541
3542 case TCP_REPAIR:
3543 if (!tcp_can_repair_sock(sk))
3544 err = -EPERM;
3545 else if (val == TCP_REPAIR_ON) {
3546 tp->repair = 1;
3547 sk->sk_reuse = SK_FORCE_REUSE;
3548 tp->repair_queue = TCP_NO_QUEUE;
3549 } else if (val == TCP_REPAIR_OFF) {
3550 tp->repair = 0;
3551 sk->sk_reuse = SK_NO_REUSE;
3552 tcp_send_window_probe(sk);
3553 } else if (val == TCP_REPAIR_OFF_NO_WP) {
3554 tp->repair = 0;
3555 sk->sk_reuse = SK_NO_REUSE;
3556 } else
3557 err = -EINVAL;
3558
3559 break;
3560
3561 case TCP_REPAIR_QUEUE:
3562 if (!tp->repair)
3563 err = -EPERM;
3564 else if ((unsigned int)val < TCP_QUEUES_NR)
3565 tp->repair_queue = val;
3566 else
3567 err = -EINVAL;
3568 break;
3569
3570 case TCP_QUEUE_SEQ:
3571 if (sk->sk_state != TCP_CLOSE) {
3572 err = -EPERM;
3573 } else if (tp->repair_queue == TCP_SEND_QUEUE) {
3574 if (!tcp_rtx_queue_empty(sk))
3575 err = -EPERM;
3576 else
3577 WRITE_ONCE(tp->write_seq, val);
3578 } else if (tp->repair_queue == TCP_RECV_QUEUE) {
3579 if (tp->rcv_nxt != tp->copied_seq) {
3580 err = -EPERM;
3581 } else {
3582 WRITE_ONCE(tp->rcv_nxt, val);
3583 WRITE_ONCE(tp->copied_seq, val);
3584 }
3585 } else {
3586 err = -EINVAL;
3587 }
3588 break;
3589
3590 case TCP_REPAIR_OPTIONS:
3591 if (!tp->repair)
3592 err = -EINVAL;
3593 else if (sk->sk_state == TCP_ESTABLISHED && !tp->bytes_sent)
3594 err = tcp_repair_options_est(sk, optval, optlen);
3595 else
3596 err = -EPERM;
3597 break;
3598
3599 case TCP_CORK:
3600 __tcp_sock_set_cork(sk, val);
3601 break;
3602
3603 case TCP_KEEPIDLE:
3604 err = tcp_sock_set_keepidle_locked(sk, val);
3605 break;
3606 case TCP_SAVE_SYN:
3607 /* 0: disable, 1: enable, 2: start from ether_header */
3608 if (val < 0 || val > 2)
3609 err = -EINVAL;
3610 else
3611 tp->save_syn = val;
3612 break;
3613
3614 case TCP_WINDOW_CLAMP:
3615 err = tcp_set_window_clamp(sk, val);
3616 break;
3617
3618 case TCP_QUICKACK:
3619 __tcp_sock_set_quickack(sk, val);
3620 break;
3621
3622 case TCP_AO_REPAIR:
3623 if (!tcp_can_repair_sock(sk)) {
3624 err = -EPERM;
3625 break;
3626 }
3627 err = tcp_ao_set_repair(sk, optval, optlen);
3628 break;
3629#ifdef CONFIG_TCP_AO
3630 case TCP_AO_ADD_KEY:
3631 case TCP_AO_DEL_KEY:
3632 case TCP_AO_INFO: {
3633 /* If this is the first TCP-AO setsockopt() on the socket,
3634 * sk_state has to be LISTEN or CLOSE. Allow TCP_REPAIR
3635 * in any state.
3636 */
3637 if ((1 << sk->sk_state) & (TCPF_LISTEN | TCPF_CLOSE))
3638 goto ao_parse;
3639 if (rcu_dereference_protected(tcp_sk(sk)->ao_info,
3640 lockdep_sock_is_held(sk)))
3641 goto ao_parse;
3642 if (tp->repair)
3643 goto ao_parse;
3644 err = -EISCONN;
3645 break;
3646ao_parse:
3647 err = tp->af_specific->ao_parse(sk, optname, optval, optlen);
3648 break;
3649 }
3650#endif
3651#ifdef CONFIG_TCP_MD5SIG
3652 case TCP_MD5SIG:
3653 case TCP_MD5SIG_EXT:
3654 err = tp->af_specific->md5_parse(sk, optname, optval, optlen);
3655 break;
3656#endif
3657 case TCP_FASTOPEN:
3658 if (val >= 0 && ((1 << sk->sk_state) & (TCPF_CLOSE |
3659 TCPF_LISTEN))) {
3660 tcp_fastopen_init_key_once(net);
3661
3662 fastopen_queue_tune(sk, val);
3663 } else {
3664 err = -EINVAL;
3665 }
3666 break;
3667 case TCP_FASTOPEN_CONNECT:
3668 if (val > 1 || val < 0) {
3669 err = -EINVAL;
3670 } else if (READ_ONCE(net->ipv4.sysctl_tcp_fastopen) &
3671 TFO_CLIENT_ENABLE) {
3672 if (sk->sk_state == TCP_CLOSE)
3673 tp->fastopen_connect = val;
3674 else
3675 err = -EINVAL;
3676 } else {
3677 err = -EOPNOTSUPP;
3678 }
3679 break;
3680 case TCP_FASTOPEN_NO_COOKIE:
3681 if (val > 1 || val < 0)
3682 err = -EINVAL;
3683 else if (!((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
3684 err = -EINVAL;
3685 else
3686 tp->fastopen_no_cookie = val;
3687 break;
3688 case TCP_TIMESTAMP:
3689 if (!tp->repair) {
3690 err = -EPERM;
3691 break;
3692 }
3693 /* val is an opaque field,
3694 * and low order bit contains usec_ts enable bit.
3695 * Its a best effort, and we do not care if user makes an error.
3696 */
3697 tp->tcp_usec_ts = val & 1;
3698 WRITE_ONCE(tp->tsoffset, val - tcp_clock_ts(tp->tcp_usec_ts));
3699 break;
3700 case TCP_REPAIR_WINDOW:
3701 err = tcp_repair_set_window(tp, optval, optlen);
3702 break;
3703 case TCP_NOTSENT_LOWAT:
3704 WRITE_ONCE(tp->notsent_lowat, val);
3705 sk->sk_write_space(sk);
3706 break;
3707 case TCP_INQ:
3708 if (val > 1 || val < 0)
3709 err = -EINVAL;
3710 else
3711 tp->recvmsg_inq = val;
3712 break;
3713 case TCP_TX_DELAY:
3714 if (val)
3715 tcp_enable_tx_delay();
3716 WRITE_ONCE(tp->tcp_tx_delay, val);
3717 break;
3718 default:
3719 err = -ENOPROTOOPT;
3720 break;
3721 }
3722
3723 sockopt_release_sock(sk);
3724 return err;
3725}
3726
3727int tcp_setsockopt(struct sock *sk, int level, int optname, sockptr_t optval,
3728 unsigned int optlen)
3729{
3730 const struct inet_connection_sock *icsk = inet_csk(sk);
3731
3732 if (level != SOL_TCP)
3733 /* Paired with WRITE_ONCE() in do_ipv6_setsockopt() and tcp_v6_connect() */
3734 return READ_ONCE(icsk->icsk_af_ops)->setsockopt(sk, level, optname,
3735 optval, optlen);
3736 return do_tcp_setsockopt(sk, level, optname, optval, optlen);
3737}
3738EXPORT_SYMBOL(tcp_setsockopt);
3739
3740static void tcp_get_info_chrono_stats(const struct tcp_sock *tp,
3741 struct tcp_info *info)
3742{
3743 u64 stats[__TCP_CHRONO_MAX], total = 0;
3744 enum tcp_chrono i;
3745
3746 for (i = TCP_CHRONO_BUSY; i < __TCP_CHRONO_MAX; ++i) {
3747 stats[i] = tp->chrono_stat[i - 1];
3748 if (i == tp->chrono_type)
3749 stats[i] += tcp_jiffies32 - tp->chrono_start;
3750 stats[i] *= USEC_PER_SEC / HZ;
3751 total += stats[i];
3752 }
3753
3754 info->tcpi_busy_time = total;
3755 info->tcpi_rwnd_limited = stats[TCP_CHRONO_RWND_LIMITED];
3756 info->tcpi_sndbuf_limited = stats[TCP_CHRONO_SNDBUF_LIMITED];
3757}
3758
3759/* Return information about state of tcp endpoint in API format. */
3760void tcp_get_info(struct sock *sk, struct tcp_info *info)
3761{
3762 const struct tcp_sock *tp = tcp_sk(sk); /* iff sk_type == SOCK_STREAM */
3763 const struct inet_connection_sock *icsk = inet_csk(sk);
3764 unsigned long rate;
3765 u32 now;
3766 u64 rate64;
3767 bool slow;
3768
3769 memset(info, 0, sizeof(*info));
3770 if (sk->sk_type != SOCK_STREAM)
3771 return;
3772
3773 info->tcpi_state = inet_sk_state_load(sk);
3774
3775 /* Report meaningful fields for all TCP states, including listeners */
3776 rate = READ_ONCE(sk->sk_pacing_rate);
3777 rate64 = (rate != ~0UL) ? rate : ~0ULL;
3778 info->tcpi_pacing_rate = rate64;
3779
3780 rate = READ_ONCE(sk->sk_max_pacing_rate);
3781 rate64 = (rate != ~0UL) ? rate : ~0ULL;
3782 info->tcpi_max_pacing_rate = rate64;
3783
3784 info->tcpi_reordering = tp->reordering;
3785 info->tcpi_snd_cwnd = tcp_snd_cwnd(tp);
3786
3787 if (info->tcpi_state == TCP_LISTEN) {
3788 /* listeners aliased fields :
3789 * tcpi_unacked -> Number of children ready for accept()
3790 * tcpi_sacked -> max backlog
3791 */
3792 info->tcpi_unacked = READ_ONCE(sk->sk_ack_backlog);
3793 info->tcpi_sacked = READ_ONCE(sk->sk_max_ack_backlog);
3794 return;
3795 }
3796
3797 slow = lock_sock_fast(sk);
3798
3799 info->tcpi_ca_state = icsk->icsk_ca_state;
3800 info->tcpi_retransmits = icsk->icsk_retransmits;
3801 info->tcpi_probes = icsk->icsk_probes_out;
3802 info->tcpi_backoff = icsk->icsk_backoff;
3803
3804 if (tp->rx_opt.tstamp_ok)
3805 info->tcpi_options |= TCPI_OPT_TIMESTAMPS;
3806 if (tcp_is_sack(tp))
3807 info->tcpi_options |= TCPI_OPT_SACK;
3808 if (tp->rx_opt.wscale_ok) {
3809 info->tcpi_options |= TCPI_OPT_WSCALE;
3810 info->tcpi_snd_wscale = tp->rx_opt.snd_wscale;
3811 info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale;
3812 }
3813
3814 if (tp->ecn_flags & TCP_ECN_OK)
3815 info->tcpi_options |= TCPI_OPT_ECN;
3816 if (tp->ecn_flags & TCP_ECN_SEEN)
3817 info->tcpi_options |= TCPI_OPT_ECN_SEEN;
3818 if (tp->syn_data_acked)
3819 info->tcpi_options |= TCPI_OPT_SYN_DATA;
3820 if (tp->tcp_usec_ts)
3821 info->tcpi_options |= TCPI_OPT_USEC_TS;
3822
3823 info->tcpi_rto = jiffies_to_usecs(icsk->icsk_rto);
3824 info->tcpi_ato = jiffies_to_usecs(min_t(u32, icsk->icsk_ack.ato,
3825 tcp_delack_max(sk)));
3826 info->tcpi_snd_mss = tp->mss_cache;
3827 info->tcpi_rcv_mss = icsk->icsk_ack.rcv_mss;
3828
3829 info->tcpi_unacked = tp->packets_out;
3830 info->tcpi_sacked = tp->sacked_out;
3831
3832 info->tcpi_lost = tp->lost_out;
3833 info->tcpi_retrans = tp->retrans_out;
3834
3835 now = tcp_jiffies32;
3836 info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime);
3837 info->tcpi_last_data_recv = jiffies_to_msecs(now - icsk->icsk_ack.lrcvtime);
3838 info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp);
3839
3840 info->tcpi_pmtu = icsk->icsk_pmtu_cookie;
3841 info->tcpi_rcv_ssthresh = tp->rcv_ssthresh;
3842 info->tcpi_rtt = tp->srtt_us >> 3;
3843 info->tcpi_rttvar = tp->mdev_us >> 2;
3844 info->tcpi_snd_ssthresh = tp->snd_ssthresh;
3845 info->tcpi_advmss = tp->advmss;
3846
3847 info->tcpi_rcv_rtt = tp->rcv_rtt_est.rtt_us >> 3;
3848 info->tcpi_rcv_space = tp->rcvq_space.space;
3849
3850 info->tcpi_total_retrans = tp->total_retrans;
3851
3852 info->tcpi_bytes_acked = tp->bytes_acked;
3853 info->tcpi_bytes_received = tp->bytes_received;
3854 info->tcpi_notsent_bytes = max_t(int, 0, tp->write_seq - tp->snd_nxt);
3855 tcp_get_info_chrono_stats(tp, info);
3856
3857 info->tcpi_segs_out = tp->segs_out;
3858
3859 /* segs_in and data_segs_in can be updated from tcp_segs_in() from BH */
3860 info->tcpi_segs_in = READ_ONCE(tp->segs_in);
3861 info->tcpi_data_segs_in = READ_ONCE(tp->data_segs_in);
3862
3863 info->tcpi_min_rtt = tcp_min_rtt(tp);
3864 info->tcpi_data_segs_out = tp->data_segs_out;
3865
3866 info->tcpi_delivery_rate_app_limited = tp->rate_app_limited ? 1 : 0;
3867 rate64 = tcp_compute_delivery_rate(tp);
3868 if (rate64)
3869 info->tcpi_delivery_rate = rate64;
3870 info->tcpi_delivered = tp->delivered;
3871 info->tcpi_delivered_ce = tp->delivered_ce;
3872 info->tcpi_bytes_sent = tp->bytes_sent;
3873 info->tcpi_bytes_retrans = tp->bytes_retrans;
3874 info->tcpi_dsack_dups = tp->dsack_dups;
3875 info->tcpi_reord_seen = tp->reord_seen;
3876 info->tcpi_rcv_ooopack = tp->rcv_ooopack;
3877 info->tcpi_snd_wnd = tp->snd_wnd;
3878 info->tcpi_rcv_wnd = tp->rcv_wnd;
3879 info->tcpi_rehash = tp->plb_rehash + tp->timeout_rehash;
3880 info->tcpi_fastopen_client_fail = tp->fastopen_client_fail;
3881
3882 info->tcpi_total_rto = tp->total_rto;
3883 info->tcpi_total_rto_recoveries = tp->total_rto_recoveries;
3884 info->tcpi_total_rto_time = tp->total_rto_time;
3885 if (tp->rto_stamp)
3886 info->tcpi_total_rto_time += tcp_clock_ms() - tp->rto_stamp;
3887
3888 unlock_sock_fast(sk, slow);
3889}
3890EXPORT_SYMBOL_GPL(tcp_get_info);
3891
3892static size_t tcp_opt_stats_get_size(void)
3893{
3894 return
3895 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_BUSY */
3896 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_RWND_LIMITED */
3897 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_SNDBUF_LIMITED */
3898 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_DATA_SEGS_OUT */
3899 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_TOTAL_RETRANS */
3900 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_PACING_RATE */
3901 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_DELIVERY_RATE */
3902 nla_total_size(sizeof(u32)) + /* TCP_NLA_SND_CWND */
3903 nla_total_size(sizeof(u32)) + /* TCP_NLA_REORDERING */
3904 nla_total_size(sizeof(u32)) + /* TCP_NLA_MIN_RTT */
3905 nla_total_size(sizeof(u8)) + /* TCP_NLA_RECUR_RETRANS */
3906 nla_total_size(sizeof(u8)) + /* TCP_NLA_DELIVERY_RATE_APP_LMT */
3907 nla_total_size(sizeof(u32)) + /* TCP_NLA_SNDQ_SIZE */
3908 nla_total_size(sizeof(u8)) + /* TCP_NLA_CA_STATE */
3909 nla_total_size(sizeof(u32)) + /* TCP_NLA_SND_SSTHRESH */
3910 nla_total_size(sizeof(u32)) + /* TCP_NLA_DELIVERED */
3911 nla_total_size(sizeof(u32)) + /* TCP_NLA_DELIVERED_CE */
3912 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_BYTES_SENT */
3913 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_BYTES_RETRANS */
3914 nla_total_size(sizeof(u32)) + /* TCP_NLA_DSACK_DUPS */
3915 nla_total_size(sizeof(u32)) + /* TCP_NLA_REORD_SEEN */
3916 nla_total_size(sizeof(u32)) + /* TCP_NLA_SRTT */
3917 nla_total_size(sizeof(u16)) + /* TCP_NLA_TIMEOUT_REHASH */
3918 nla_total_size(sizeof(u32)) + /* TCP_NLA_BYTES_NOTSENT */
3919 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_EDT */
3920 nla_total_size(sizeof(u8)) + /* TCP_NLA_TTL */
3921 nla_total_size(sizeof(u32)) + /* TCP_NLA_REHASH */
3922 0;
3923}
3924
3925/* Returns TTL or hop limit of an incoming packet from skb. */
3926static u8 tcp_skb_ttl_or_hop_limit(const struct sk_buff *skb)
3927{
3928 if (skb->protocol == htons(ETH_P_IP))
3929 return ip_hdr(skb)->ttl;
3930 else if (skb->protocol == htons(ETH_P_IPV6))
3931 return ipv6_hdr(skb)->hop_limit;
3932 else
3933 return 0;
3934}
3935
3936struct sk_buff *tcp_get_timestamping_opt_stats(const struct sock *sk,
3937 const struct sk_buff *orig_skb,
3938 const struct sk_buff *ack_skb)
3939{
3940 const struct tcp_sock *tp = tcp_sk(sk);
3941 struct sk_buff *stats;
3942 struct tcp_info info;
3943 unsigned long rate;
3944 u64 rate64;
3945
3946 stats = alloc_skb(tcp_opt_stats_get_size(), GFP_ATOMIC);
3947 if (!stats)
3948 return NULL;
3949
3950 tcp_get_info_chrono_stats(tp, &info);
3951 nla_put_u64_64bit(stats, TCP_NLA_BUSY,
3952 info.tcpi_busy_time, TCP_NLA_PAD);
3953 nla_put_u64_64bit(stats, TCP_NLA_RWND_LIMITED,
3954 info.tcpi_rwnd_limited, TCP_NLA_PAD);
3955 nla_put_u64_64bit(stats, TCP_NLA_SNDBUF_LIMITED,
3956 info.tcpi_sndbuf_limited, TCP_NLA_PAD);
3957 nla_put_u64_64bit(stats, TCP_NLA_DATA_SEGS_OUT,
3958 tp->data_segs_out, TCP_NLA_PAD);
3959 nla_put_u64_64bit(stats, TCP_NLA_TOTAL_RETRANS,
3960 tp->total_retrans, TCP_NLA_PAD);
3961
3962 rate = READ_ONCE(sk->sk_pacing_rate);
3963 rate64 = (rate != ~0UL) ? rate : ~0ULL;
3964 nla_put_u64_64bit(stats, TCP_NLA_PACING_RATE, rate64, TCP_NLA_PAD);
3965
3966 rate64 = tcp_compute_delivery_rate(tp);
3967 nla_put_u64_64bit(stats, TCP_NLA_DELIVERY_RATE, rate64, TCP_NLA_PAD);
3968
3969 nla_put_u32(stats, TCP_NLA_SND_CWND, tcp_snd_cwnd(tp));
3970 nla_put_u32(stats, TCP_NLA_REORDERING, tp->reordering);
3971 nla_put_u32(stats, TCP_NLA_MIN_RTT, tcp_min_rtt(tp));
3972
3973 nla_put_u8(stats, TCP_NLA_RECUR_RETRANS, inet_csk(sk)->icsk_retransmits);
3974 nla_put_u8(stats, TCP_NLA_DELIVERY_RATE_APP_LMT, !!tp->rate_app_limited);
3975 nla_put_u32(stats, TCP_NLA_SND_SSTHRESH, tp->snd_ssthresh);
3976 nla_put_u32(stats, TCP_NLA_DELIVERED, tp->delivered);
3977 nla_put_u32(stats, TCP_NLA_DELIVERED_CE, tp->delivered_ce);
3978
3979 nla_put_u32(stats, TCP_NLA_SNDQ_SIZE, tp->write_seq - tp->snd_una);
3980 nla_put_u8(stats, TCP_NLA_CA_STATE, inet_csk(sk)->icsk_ca_state);
3981
3982 nla_put_u64_64bit(stats, TCP_NLA_BYTES_SENT, tp->bytes_sent,
3983 TCP_NLA_PAD);
3984 nla_put_u64_64bit(stats, TCP_NLA_BYTES_RETRANS, tp->bytes_retrans,
3985 TCP_NLA_PAD);
3986 nla_put_u32(stats, TCP_NLA_DSACK_DUPS, tp->dsack_dups);
3987 nla_put_u32(stats, TCP_NLA_REORD_SEEN, tp->reord_seen);
3988 nla_put_u32(stats, TCP_NLA_SRTT, tp->srtt_us >> 3);
3989 nla_put_u16(stats, TCP_NLA_TIMEOUT_REHASH, tp->timeout_rehash);
3990 nla_put_u32(stats, TCP_NLA_BYTES_NOTSENT,
3991 max_t(int, 0, tp->write_seq - tp->snd_nxt));
3992 nla_put_u64_64bit(stats, TCP_NLA_EDT, orig_skb->skb_mstamp_ns,
3993 TCP_NLA_PAD);
3994 if (ack_skb)
3995 nla_put_u8(stats, TCP_NLA_TTL,
3996 tcp_skb_ttl_or_hop_limit(ack_skb));
3997
3998 nla_put_u32(stats, TCP_NLA_REHASH, tp->plb_rehash + tp->timeout_rehash);
3999 return stats;
4000}
4001
4002int do_tcp_getsockopt(struct sock *sk, int level,
4003 int optname, sockptr_t optval, sockptr_t optlen)
4004{
4005 struct inet_connection_sock *icsk = inet_csk(sk);
4006 struct tcp_sock *tp = tcp_sk(sk);
4007 struct net *net = sock_net(sk);
4008 int val, len;
4009
4010 if (copy_from_sockptr(&len, optlen, sizeof(int)))
4011 return -EFAULT;
4012
4013 len = min_t(unsigned int, len, sizeof(int));
4014
4015 if (len < 0)
4016 return -EINVAL;
4017
4018 switch (optname) {
4019 case TCP_MAXSEG:
4020 val = tp->mss_cache;
4021 if (tp->rx_opt.user_mss &&
4022 ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
4023 val = tp->rx_opt.user_mss;
4024 if (tp->repair)
4025 val = tp->rx_opt.mss_clamp;
4026 break;
4027 case TCP_NODELAY:
4028 val = !!(tp->nonagle&TCP_NAGLE_OFF);
4029 break;
4030 case TCP_CORK:
4031 val = !!(tp->nonagle&TCP_NAGLE_CORK);
4032 break;
4033 case TCP_KEEPIDLE:
4034 val = keepalive_time_when(tp) / HZ;
4035 break;
4036 case TCP_KEEPINTVL:
4037 val = keepalive_intvl_when(tp) / HZ;
4038 break;
4039 case TCP_KEEPCNT:
4040 val = keepalive_probes(tp);
4041 break;
4042 case TCP_SYNCNT:
4043 val = READ_ONCE(icsk->icsk_syn_retries) ? :
4044 READ_ONCE(net->ipv4.sysctl_tcp_syn_retries);
4045 break;
4046 case TCP_LINGER2:
4047 val = READ_ONCE(tp->linger2);
4048 if (val >= 0)
4049 val = (val ? : READ_ONCE(net->ipv4.sysctl_tcp_fin_timeout)) / HZ;
4050 break;
4051 case TCP_DEFER_ACCEPT:
4052 val = READ_ONCE(icsk->icsk_accept_queue.rskq_defer_accept);
4053 val = retrans_to_secs(val, TCP_TIMEOUT_INIT / HZ,
4054 TCP_RTO_MAX / HZ);
4055 break;
4056 case TCP_WINDOW_CLAMP:
4057 val = tp->window_clamp;
4058 break;
4059 case TCP_INFO: {
4060 struct tcp_info info;
4061
4062 if (copy_from_sockptr(&len, optlen, sizeof(int)))
4063 return -EFAULT;
4064
4065 tcp_get_info(sk, &info);
4066
4067 len = min_t(unsigned int, len, sizeof(info));
4068 if (copy_to_sockptr(optlen, &len, sizeof(int)))
4069 return -EFAULT;
4070 if (copy_to_sockptr(optval, &info, len))
4071 return -EFAULT;
4072 return 0;
4073 }
4074 case TCP_CC_INFO: {
4075 const struct tcp_congestion_ops *ca_ops;
4076 union tcp_cc_info info;
4077 size_t sz = 0;
4078 int attr;
4079
4080 if (copy_from_sockptr(&len, optlen, sizeof(int)))
4081 return -EFAULT;
4082
4083 ca_ops = icsk->icsk_ca_ops;
4084 if (ca_ops && ca_ops->get_info)
4085 sz = ca_ops->get_info(sk, ~0U, &attr, &info);
4086
4087 len = min_t(unsigned int, len, sz);
4088 if (copy_to_sockptr(optlen, &len, sizeof(int)))
4089 return -EFAULT;
4090 if (copy_to_sockptr(optval, &info, len))
4091 return -EFAULT;
4092 return 0;
4093 }
4094 case TCP_QUICKACK:
4095 val = !inet_csk_in_pingpong_mode(sk);
4096 break;
4097
4098 case TCP_CONGESTION:
4099 if (copy_from_sockptr(&len, optlen, sizeof(int)))
4100 return -EFAULT;
4101 len = min_t(unsigned int, len, TCP_CA_NAME_MAX);
4102 if (copy_to_sockptr(optlen, &len, sizeof(int)))
4103 return -EFAULT;
4104 if (copy_to_sockptr(optval, icsk->icsk_ca_ops->name, len))
4105 return -EFAULT;
4106 return 0;
4107
4108 case TCP_ULP:
4109 if (copy_from_sockptr(&len, optlen, sizeof(int)))
4110 return -EFAULT;
4111 len = min_t(unsigned int, len, TCP_ULP_NAME_MAX);
4112 if (!icsk->icsk_ulp_ops) {
4113 len = 0;
4114 if (copy_to_sockptr(optlen, &len, sizeof(int)))
4115 return -EFAULT;
4116 return 0;
4117 }
4118 if (copy_to_sockptr(optlen, &len, sizeof(int)))
4119 return -EFAULT;
4120 if (copy_to_sockptr(optval, icsk->icsk_ulp_ops->name, len))
4121 return -EFAULT;
4122 return 0;
4123
4124 case TCP_FASTOPEN_KEY: {
4125 u64 key[TCP_FASTOPEN_KEY_BUF_LENGTH / sizeof(u64)];
4126 unsigned int key_len;
4127
4128 if (copy_from_sockptr(&len, optlen, sizeof(int)))
4129 return -EFAULT;
4130
4131 key_len = tcp_fastopen_get_cipher(net, icsk, key) *
4132 TCP_FASTOPEN_KEY_LENGTH;
4133 len = min_t(unsigned int, len, key_len);
4134 if (copy_to_sockptr(optlen, &len, sizeof(int)))
4135 return -EFAULT;
4136 if (copy_to_sockptr(optval, key, len))
4137 return -EFAULT;
4138 return 0;
4139 }
4140 case TCP_THIN_LINEAR_TIMEOUTS:
4141 val = tp->thin_lto;
4142 break;
4143
4144 case TCP_THIN_DUPACK:
4145 val = 0;
4146 break;
4147
4148 case TCP_REPAIR:
4149 val = tp->repair;
4150 break;
4151
4152 case TCP_REPAIR_QUEUE:
4153 if (tp->repair)
4154 val = tp->repair_queue;
4155 else
4156 return -EINVAL;
4157 break;
4158
4159 case TCP_REPAIR_WINDOW: {
4160 struct tcp_repair_window opt;
4161
4162 if (copy_from_sockptr(&len, optlen, sizeof(int)))
4163 return -EFAULT;
4164
4165 if (len != sizeof(opt))
4166 return -EINVAL;
4167
4168 if (!tp->repair)
4169 return -EPERM;
4170
4171 opt.snd_wl1 = tp->snd_wl1;
4172 opt.snd_wnd = tp->snd_wnd;
4173 opt.max_window = tp->max_window;
4174 opt.rcv_wnd = tp->rcv_wnd;
4175 opt.rcv_wup = tp->rcv_wup;
4176
4177 if (copy_to_sockptr(optval, &opt, len))
4178 return -EFAULT;
4179 return 0;
4180 }
4181 case TCP_QUEUE_SEQ:
4182 if (tp->repair_queue == TCP_SEND_QUEUE)
4183 val = tp->write_seq;
4184 else if (tp->repair_queue == TCP_RECV_QUEUE)
4185 val = tp->rcv_nxt;
4186 else
4187 return -EINVAL;
4188 break;
4189
4190 case TCP_USER_TIMEOUT:
4191 val = READ_ONCE(icsk->icsk_user_timeout);
4192 break;
4193
4194 case TCP_FASTOPEN:
4195 val = READ_ONCE(icsk->icsk_accept_queue.fastopenq.max_qlen);
4196 break;
4197
4198 case TCP_FASTOPEN_CONNECT:
4199 val = tp->fastopen_connect;
4200 break;
4201
4202 case TCP_FASTOPEN_NO_COOKIE:
4203 val = tp->fastopen_no_cookie;
4204 break;
4205
4206 case TCP_TX_DELAY:
4207 val = READ_ONCE(tp->tcp_tx_delay);
4208 break;
4209
4210 case TCP_TIMESTAMP:
4211 val = tcp_clock_ts(tp->tcp_usec_ts) + READ_ONCE(tp->tsoffset);
4212 if (tp->tcp_usec_ts)
4213 val |= 1;
4214 else
4215 val &= ~1;
4216 break;
4217 case TCP_NOTSENT_LOWAT:
4218 val = READ_ONCE(tp->notsent_lowat);
4219 break;
4220 case TCP_INQ:
4221 val = tp->recvmsg_inq;
4222 break;
4223 case TCP_SAVE_SYN:
4224 val = tp->save_syn;
4225 break;
4226 case TCP_SAVED_SYN: {
4227 if (copy_from_sockptr(&len, optlen, sizeof(int)))
4228 return -EFAULT;
4229
4230 sockopt_lock_sock(sk);
4231 if (tp->saved_syn) {
4232 if (len < tcp_saved_syn_len(tp->saved_syn)) {
4233 len = tcp_saved_syn_len(tp->saved_syn);
4234 if (copy_to_sockptr(optlen, &len, sizeof(int))) {
4235 sockopt_release_sock(sk);
4236 return -EFAULT;
4237 }
4238 sockopt_release_sock(sk);
4239 return -EINVAL;
4240 }
4241 len = tcp_saved_syn_len(tp->saved_syn);
4242 if (copy_to_sockptr(optlen, &len, sizeof(int))) {
4243 sockopt_release_sock(sk);
4244 return -EFAULT;
4245 }
4246 if (copy_to_sockptr(optval, tp->saved_syn->data, len)) {
4247 sockopt_release_sock(sk);
4248 return -EFAULT;
4249 }
4250 tcp_saved_syn_free(tp);
4251 sockopt_release_sock(sk);
4252 } else {
4253 sockopt_release_sock(sk);
4254 len = 0;
4255 if (copy_to_sockptr(optlen, &len, sizeof(int)))
4256 return -EFAULT;
4257 }
4258 return 0;
4259 }
4260#ifdef CONFIG_MMU
4261 case TCP_ZEROCOPY_RECEIVE: {
4262 struct scm_timestamping_internal tss;
4263 struct tcp_zerocopy_receive zc = {};
4264 int err;
4265
4266 if (copy_from_sockptr(&len, optlen, sizeof(int)))
4267 return -EFAULT;
4268 if (len < 0 ||
4269 len < offsetofend(struct tcp_zerocopy_receive, length))
4270 return -EINVAL;
4271 if (unlikely(len > sizeof(zc))) {
4272 err = check_zeroed_sockptr(optval, sizeof(zc),
4273 len - sizeof(zc));
4274 if (err < 1)
4275 return err == 0 ? -EINVAL : err;
4276 len = sizeof(zc);
4277 if (copy_to_sockptr(optlen, &len, sizeof(int)))
4278 return -EFAULT;
4279 }
4280 if (copy_from_sockptr(&zc, optval, len))
4281 return -EFAULT;
4282 if (zc.reserved)
4283 return -EINVAL;
4284 if (zc.msg_flags & ~(TCP_VALID_ZC_MSG_FLAGS))
4285 return -EINVAL;
4286 sockopt_lock_sock(sk);
4287 err = tcp_zerocopy_receive(sk, &zc, &tss);
4288 err = BPF_CGROUP_RUN_PROG_GETSOCKOPT_KERN(sk, level, optname,
4289 &zc, &len, err);
4290 sockopt_release_sock(sk);
4291 if (len >= offsetofend(struct tcp_zerocopy_receive, msg_flags))
4292 goto zerocopy_rcv_cmsg;
4293 switch (len) {
4294 case offsetofend(struct tcp_zerocopy_receive, msg_flags):
4295 goto zerocopy_rcv_cmsg;
4296 case offsetofend(struct tcp_zerocopy_receive, msg_controllen):
4297 case offsetofend(struct tcp_zerocopy_receive, msg_control):
4298 case offsetofend(struct tcp_zerocopy_receive, flags):
4299 case offsetofend(struct tcp_zerocopy_receive, copybuf_len):
4300 case offsetofend(struct tcp_zerocopy_receive, copybuf_address):
4301 case offsetofend(struct tcp_zerocopy_receive, err):
4302 goto zerocopy_rcv_sk_err;
4303 case offsetofend(struct tcp_zerocopy_receive, inq):
4304 goto zerocopy_rcv_inq;
4305 case offsetofend(struct tcp_zerocopy_receive, length):
4306 default:
4307 goto zerocopy_rcv_out;
4308 }
4309zerocopy_rcv_cmsg:
4310 if (zc.msg_flags & TCP_CMSG_TS)
4311 tcp_zc_finalize_rx_tstamp(sk, &zc, &tss);
4312 else
4313 zc.msg_flags = 0;
4314zerocopy_rcv_sk_err:
4315 if (!err)
4316 zc.err = sock_error(sk);
4317zerocopy_rcv_inq:
4318 zc.inq = tcp_inq_hint(sk);
4319zerocopy_rcv_out:
4320 if (!err && copy_to_sockptr(optval, &zc, len))
4321 err = -EFAULT;
4322 return err;
4323 }
4324#endif
4325 case TCP_AO_REPAIR:
4326 if (!tcp_can_repair_sock(sk))
4327 return -EPERM;
4328 return tcp_ao_get_repair(sk, optval, optlen);
4329 case TCP_AO_GET_KEYS:
4330 case TCP_AO_INFO: {
4331 int err;
4332
4333 sockopt_lock_sock(sk);
4334 if (optname == TCP_AO_GET_KEYS)
4335 err = tcp_ao_get_mkts(sk, optval, optlen);
4336 else
4337 err = tcp_ao_get_sock_info(sk, optval, optlen);
4338 sockopt_release_sock(sk);
4339
4340 return err;
4341 }
4342 default:
4343 return -ENOPROTOOPT;
4344 }
4345
4346 if (copy_to_sockptr(optlen, &len, sizeof(int)))
4347 return -EFAULT;
4348 if (copy_to_sockptr(optval, &val, len))
4349 return -EFAULT;
4350 return 0;
4351}
4352
4353bool tcp_bpf_bypass_getsockopt(int level, int optname)
4354{
4355 /* TCP do_tcp_getsockopt has optimized getsockopt implementation
4356 * to avoid extra socket lock for TCP_ZEROCOPY_RECEIVE.
4357 */
4358 if (level == SOL_TCP && optname == TCP_ZEROCOPY_RECEIVE)
4359 return true;
4360
4361 return false;
4362}
4363EXPORT_SYMBOL(tcp_bpf_bypass_getsockopt);
4364
4365int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval,
4366 int __user *optlen)
4367{
4368 struct inet_connection_sock *icsk = inet_csk(sk);
4369
4370 if (level != SOL_TCP)
4371 /* Paired with WRITE_ONCE() in do_ipv6_setsockopt() and tcp_v6_connect() */
4372 return READ_ONCE(icsk->icsk_af_ops)->getsockopt(sk, level, optname,
4373 optval, optlen);
4374 return do_tcp_getsockopt(sk, level, optname, USER_SOCKPTR(optval),
4375 USER_SOCKPTR(optlen));
4376}
4377EXPORT_SYMBOL(tcp_getsockopt);
4378
4379#ifdef CONFIG_TCP_MD5SIG
4380int tcp_md5_sigpool_id = -1;
4381EXPORT_SYMBOL_GPL(tcp_md5_sigpool_id);
4382
4383int tcp_md5_alloc_sigpool(void)
4384{
4385 size_t scratch_size;
4386 int ret;
4387
4388 scratch_size = sizeof(union tcp_md5sum_block) + sizeof(struct tcphdr);
4389 ret = tcp_sigpool_alloc_ahash("md5", scratch_size);
4390 if (ret >= 0) {
4391 /* As long as any md5 sigpool was allocated, the return
4392 * id would stay the same. Re-write the id only for the case
4393 * when previously all MD5 keys were deleted and this call
4394 * allocates the first MD5 key, which may return a different
4395 * sigpool id than was used previously.
4396 */
4397 WRITE_ONCE(tcp_md5_sigpool_id, ret); /* Avoids the compiler potentially being smart here */
4398 return 0;
4399 }
4400 return ret;
4401}
4402
4403void tcp_md5_release_sigpool(void)
4404{
4405 tcp_sigpool_release(READ_ONCE(tcp_md5_sigpool_id));
4406}
4407
4408void tcp_md5_add_sigpool(void)
4409{
4410 tcp_sigpool_get(READ_ONCE(tcp_md5_sigpool_id));
4411}
4412
4413int tcp_md5_hash_key(struct tcp_sigpool *hp,
4414 const struct tcp_md5sig_key *key)
4415{
4416 u8 keylen = READ_ONCE(key->keylen); /* paired with WRITE_ONCE() in tcp_md5_do_add */
4417 struct scatterlist sg;
4418
4419 sg_init_one(&sg, key->key, keylen);
4420 ahash_request_set_crypt(hp->req, &sg, NULL, keylen);
4421
4422 /* We use data_race() because tcp_md5_do_add() might change
4423 * key->key under us
4424 */
4425 return data_race(crypto_ahash_update(hp->req));
4426}
4427EXPORT_SYMBOL(tcp_md5_hash_key);
4428
4429/* Called with rcu_read_lock() */
4430enum skb_drop_reason
4431tcp_inbound_md5_hash(const struct sock *sk, const struct sk_buff *skb,
4432 const void *saddr, const void *daddr,
4433 int family, int l3index, const __u8 *hash_location)
4434{
4435 /* This gets called for each TCP segment that has TCP-MD5 option.
4436 * We have 3 drop cases:
4437 * o No MD5 hash and one expected.
4438 * o MD5 hash and we're not expecting one.
4439 * o MD5 hash and its wrong.
4440 */
4441 const struct tcp_sock *tp = tcp_sk(sk);
4442 struct tcp_md5sig_key *key;
4443 u8 newhash[16];
4444 int genhash;
4445
4446 key = tcp_md5_do_lookup(sk, l3index, saddr, family);
4447
4448 if (!key && hash_location) {
4449 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMD5UNEXPECTED);
4450 tcp_hash_fail("Unexpected MD5 Hash found", family, skb, "");
4451 return SKB_DROP_REASON_TCP_MD5UNEXPECTED;
4452 }
4453
4454 /* Check the signature.
4455 * To support dual stack listeners, we need to handle
4456 * IPv4-mapped case.
4457 */
4458 if (family == AF_INET)
4459 genhash = tcp_v4_md5_hash_skb(newhash, key, NULL, skb);
4460 else
4461 genhash = tp->af_specific->calc_md5_hash(newhash, key,
4462 NULL, skb);
4463 if (genhash || memcmp(hash_location, newhash, 16) != 0) {
4464 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMD5FAILURE);
4465 if (family == AF_INET) {
4466 tcp_hash_fail("MD5 Hash failed", AF_INET, skb, "%s L3 index %d",
4467 genhash ? "tcp_v4_calc_md5_hash failed"
4468 : "", l3index);
4469 } else {
4470 if (genhash) {
4471 tcp_hash_fail("MD5 Hash failed",
4472 AF_INET6, skb, "L3 index %d",
4473 l3index);
4474 } else {
4475 tcp_hash_fail("MD5 Hash mismatch",
4476 AF_INET6, skb, "L3 index %d",
4477 l3index);
4478 }
4479 }
4480 return SKB_DROP_REASON_TCP_MD5FAILURE;
4481 }
4482 return SKB_NOT_DROPPED_YET;
4483}
4484EXPORT_SYMBOL(tcp_inbound_md5_hash);
4485
4486#endif
4487
4488void tcp_done(struct sock *sk)
4489{
4490 struct request_sock *req;
4491
4492 /* We might be called with a new socket, after
4493 * inet_csk_prepare_forced_close() has been called
4494 * so we can not use lockdep_sock_is_held(sk)
4495 */
4496 req = rcu_dereference_protected(tcp_sk(sk)->fastopen_rsk, 1);
4497
4498 if (sk->sk_state == TCP_SYN_SENT || sk->sk_state == TCP_SYN_RECV)
4499 TCP_INC_STATS(sock_net(sk), TCP_MIB_ATTEMPTFAILS);
4500
4501 tcp_set_state(sk, TCP_CLOSE);
4502 tcp_clear_xmit_timers(sk);
4503 if (req)
4504 reqsk_fastopen_remove(sk, req, false);
4505
4506 WRITE_ONCE(sk->sk_shutdown, SHUTDOWN_MASK);
4507
4508 if (!sock_flag(sk, SOCK_DEAD))
4509 sk->sk_state_change(sk);
4510 else
4511 inet_csk_destroy_sock(sk);
4512}
4513EXPORT_SYMBOL_GPL(tcp_done);
4514
4515int tcp_abort(struct sock *sk, int err)
4516{
4517 int state = inet_sk_state_load(sk);
4518
4519 if (state == TCP_NEW_SYN_RECV) {
4520 struct request_sock *req = inet_reqsk(sk);
4521
4522 local_bh_disable();
4523 inet_csk_reqsk_queue_drop(req->rsk_listener, req);
4524 local_bh_enable();
4525 return 0;
4526 }
4527 if (state == TCP_TIME_WAIT) {
4528 struct inet_timewait_sock *tw = inet_twsk(sk);
4529
4530 refcount_inc(&tw->tw_refcnt);
4531 local_bh_disable();
4532 inet_twsk_deschedule_put(tw);
4533 local_bh_enable();
4534 return 0;
4535 }
4536
4537 /* BPF context ensures sock locking. */
4538 if (!has_current_bpf_ctx())
4539 /* Don't race with userspace socket closes such as tcp_close. */
4540 lock_sock(sk);
4541
4542 if (sk->sk_state == TCP_LISTEN) {
4543 tcp_set_state(sk, TCP_CLOSE);
4544 inet_csk_listen_stop(sk);
4545 }
4546
4547 /* Don't race with BH socket closes such as inet_csk_listen_stop. */
4548 local_bh_disable();
4549 bh_lock_sock(sk);
4550
4551 if (!sock_flag(sk, SOCK_DEAD)) {
4552 WRITE_ONCE(sk->sk_err, err);
4553 /* This barrier is coupled with smp_rmb() in tcp_poll() */
4554 smp_wmb();
4555 sk_error_report(sk);
4556 if (tcp_need_reset(sk->sk_state))
4557 tcp_send_active_reset(sk, GFP_ATOMIC);
4558 tcp_done(sk);
4559 }
4560
4561 bh_unlock_sock(sk);
4562 local_bh_enable();
4563 tcp_write_queue_purge(sk);
4564 if (!has_current_bpf_ctx())
4565 release_sock(sk);
4566 return 0;
4567}
4568EXPORT_SYMBOL_GPL(tcp_abort);
4569
4570extern struct tcp_congestion_ops tcp_reno;
4571
4572static __initdata unsigned long thash_entries;
4573static int __init set_thash_entries(char *str)
4574{
4575 ssize_t ret;
4576
4577 if (!str)
4578 return 0;
4579
4580 ret = kstrtoul(str, 0, &thash_entries);
4581 if (ret)
4582 return 0;
4583
4584 return 1;
4585}
4586__setup("thash_entries=", set_thash_entries);
4587
4588static void __init tcp_init_mem(void)
4589{
4590 unsigned long limit = nr_free_buffer_pages() / 16;
4591
4592 limit = max(limit, 128UL);
4593 sysctl_tcp_mem[0] = limit / 4 * 3; /* 4.68 % */
4594 sysctl_tcp_mem[1] = limit; /* 6.25 % */
4595 sysctl_tcp_mem[2] = sysctl_tcp_mem[0] * 2; /* 9.37 % */
4596}
4597
4598static void __init tcp_struct_check(void)
4599{
4600 /* TX read-mostly hotpath cache lines */
4601 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_tx, max_window);
4602 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_tx, rcv_ssthresh);
4603 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_tx, reordering);
4604 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_tx, notsent_lowat);
4605 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_tx, gso_segs);
4606 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_tx, lost_skb_hint);
4607 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_tx, retransmit_skb_hint);
4608 CACHELINE_ASSERT_GROUP_SIZE(struct tcp_sock, tcp_sock_read_tx, 40);
4609
4610 /* TXRX read-mostly hotpath cache lines */
4611 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_txrx, tsoffset);
4612 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_txrx, snd_wnd);
4613 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_txrx, mss_cache);
4614 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_txrx, snd_cwnd);
4615 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_txrx, prr_out);
4616 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_txrx, lost_out);
4617 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_txrx, sacked_out);
4618 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_txrx, scaling_ratio);
4619 CACHELINE_ASSERT_GROUP_SIZE(struct tcp_sock, tcp_sock_read_txrx, 32);
4620
4621 /* RX read-mostly hotpath cache lines */
4622 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_rx, copied_seq);
4623 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_rx, rcv_tstamp);
4624 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_rx, snd_wl1);
4625 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_rx, tlp_high_seq);
4626 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_rx, rttvar_us);
4627 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_rx, retrans_out);
4628 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_rx, advmss);
4629 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_rx, urg_data);
4630 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_rx, lost);
4631 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_rx, rtt_min);
4632 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_rx, out_of_order_queue);
4633 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_rx, snd_ssthresh);
4634 CACHELINE_ASSERT_GROUP_SIZE(struct tcp_sock, tcp_sock_read_rx, 69);
4635
4636 /* TX read-write hotpath cache lines */
4637 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, segs_out);
4638 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, data_segs_out);
4639 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, bytes_sent);
4640 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, snd_sml);
4641 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, chrono_start);
4642 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, chrono_stat);
4643 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, write_seq);
4644 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, pushed_seq);
4645 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, lsndtime);
4646 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, mdev_us);
4647 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, tcp_wstamp_ns);
4648 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, tcp_clock_cache);
4649 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, tcp_mstamp);
4650 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, rtt_seq);
4651 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, tsorted_sent_queue);
4652 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, highest_sack);
4653 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, ecn_flags);
4654 CACHELINE_ASSERT_GROUP_SIZE(struct tcp_sock, tcp_sock_write_tx, 113);
4655
4656 /* TXRX read-write hotpath cache lines */
4657 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_txrx, pred_flags);
4658 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_txrx, rcv_nxt);
4659 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_txrx, snd_nxt);
4660 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_txrx, snd_una);
4661 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_txrx, window_clamp);
4662 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_txrx, srtt_us);
4663 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_txrx, packets_out);
4664 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_txrx, snd_up);
4665 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_txrx, delivered);
4666 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_txrx, delivered_ce);
4667 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_txrx, app_limited);
4668 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_txrx, rcv_wnd);
4669 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_txrx, rx_opt);
4670 CACHELINE_ASSERT_GROUP_SIZE(struct tcp_sock, tcp_sock_write_txrx, 76);
4671
4672 /* RX read-write hotpath cache lines */
4673 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_rx, bytes_received);
4674 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_rx, segs_in);
4675 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_rx, data_segs_in);
4676 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_rx, rcv_wup);
4677 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_rx, max_packets_out);
4678 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_rx, cwnd_usage_seq);
4679 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_rx, rate_delivered);
4680 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_rx, rate_interval_us);
4681 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_rx, rcv_rtt_last_tsecr);
4682 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_rx, first_tx_mstamp);
4683 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_rx, delivered_mstamp);
4684 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_rx, bytes_acked);
4685 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_rx, rcv_rtt_est);
4686 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_rx, rcvq_space);
4687 CACHELINE_ASSERT_GROUP_SIZE(struct tcp_sock, tcp_sock_write_rx, 99);
4688}
4689
4690void __init tcp_init(void)
4691{
4692 int max_rshare, max_wshare, cnt;
4693 unsigned long limit;
4694 unsigned int i;
4695
4696 BUILD_BUG_ON(TCP_MIN_SND_MSS <= MAX_TCP_OPTION_SPACE);
4697 BUILD_BUG_ON(sizeof(struct tcp_skb_cb) >
4698 sizeof_field(struct sk_buff, cb));
4699
4700 tcp_struct_check();
4701
4702 percpu_counter_init(&tcp_sockets_allocated, 0, GFP_KERNEL);
4703
4704 timer_setup(&tcp_orphan_timer, tcp_orphan_update, TIMER_DEFERRABLE);
4705 mod_timer(&tcp_orphan_timer, jiffies + TCP_ORPHAN_TIMER_PERIOD);
4706
4707 inet_hashinfo2_init(&tcp_hashinfo, "tcp_listen_portaddr_hash",
4708 thash_entries, 21, /* one slot per 2 MB*/
4709 0, 64 * 1024);
4710 tcp_hashinfo.bind_bucket_cachep =
4711 kmem_cache_create("tcp_bind_bucket",
4712 sizeof(struct inet_bind_bucket), 0,
4713 SLAB_HWCACHE_ALIGN | SLAB_PANIC |
4714 SLAB_ACCOUNT,
4715 NULL);
4716 tcp_hashinfo.bind2_bucket_cachep =
4717 kmem_cache_create("tcp_bind2_bucket",
4718 sizeof(struct inet_bind2_bucket), 0,
4719 SLAB_HWCACHE_ALIGN | SLAB_PANIC |
4720 SLAB_ACCOUNT,
4721 NULL);
4722
4723 /* Size and allocate the main established and bind bucket
4724 * hash tables.
4725 *
4726 * The methodology is similar to that of the buffer cache.
4727 */
4728 tcp_hashinfo.ehash =
4729 alloc_large_system_hash("TCP established",
4730 sizeof(struct inet_ehash_bucket),
4731 thash_entries,
4732 17, /* one slot per 128 KB of memory */
4733 0,
4734 NULL,
4735 &tcp_hashinfo.ehash_mask,
4736 0,
4737 thash_entries ? 0 : 512 * 1024);
4738 for (i = 0; i <= tcp_hashinfo.ehash_mask; i++)
4739 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].chain, i);
4740
4741 if (inet_ehash_locks_alloc(&tcp_hashinfo))
4742 panic("TCP: failed to alloc ehash_locks");
4743 tcp_hashinfo.bhash =
4744 alloc_large_system_hash("TCP bind",
4745 2 * sizeof(struct inet_bind_hashbucket),
4746 tcp_hashinfo.ehash_mask + 1,
4747 17, /* one slot per 128 KB of memory */
4748 0,
4749 &tcp_hashinfo.bhash_size,
4750 NULL,
4751 0,
4752 64 * 1024);
4753 tcp_hashinfo.bhash_size = 1U << tcp_hashinfo.bhash_size;
4754 tcp_hashinfo.bhash2 = tcp_hashinfo.bhash + tcp_hashinfo.bhash_size;
4755 for (i = 0; i < tcp_hashinfo.bhash_size; i++) {
4756 spin_lock_init(&tcp_hashinfo.bhash[i].lock);
4757 INIT_HLIST_HEAD(&tcp_hashinfo.bhash[i].chain);
4758 spin_lock_init(&tcp_hashinfo.bhash2[i].lock);
4759 INIT_HLIST_HEAD(&tcp_hashinfo.bhash2[i].chain);
4760 }
4761
4762 tcp_hashinfo.pernet = false;
4763
4764 cnt = tcp_hashinfo.ehash_mask + 1;
4765 sysctl_tcp_max_orphans = cnt / 2;
4766
4767 tcp_init_mem();
4768 /* Set per-socket limits to no more than 1/128 the pressure threshold */
4769 limit = nr_free_buffer_pages() << (PAGE_SHIFT - 7);
4770 max_wshare = min(4UL*1024*1024, limit);
4771 max_rshare = min(6UL*1024*1024, limit);
4772
4773 init_net.ipv4.sysctl_tcp_wmem[0] = PAGE_SIZE;
4774 init_net.ipv4.sysctl_tcp_wmem[1] = 16*1024;
4775 init_net.ipv4.sysctl_tcp_wmem[2] = max(64*1024, max_wshare);
4776
4777 init_net.ipv4.sysctl_tcp_rmem[0] = PAGE_SIZE;
4778 init_net.ipv4.sysctl_tcp_rmem[1] = 131072;
4779 init_net.ipv4.sysctl_tcp_rmem[2] = max(131072, max_rshare);
4780
4781 pr_info("Hash tables configured (established %u bind %u)\n",
4782 tcp_hashinfo.ehash_mask + 1, tcp_hashinfo.bhash_size);
4783
4784 tcp_v4_init();
4785 tcp_metrics_init();
4786 BUG_ON(tcp_register_congestion_control(&tcp_reno) != 0);
4787 tcp_tasklet_init();
4788 mptcp_init();
4789}