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
461 set_bit(SOCK_SUPPORT_ZC, &sk->sk_socket->flags);
462 sk_sockets_allocated_inc(sk);
463}
464EXPORT_SYMBOL(tcp_init_sock);
465
466static void tcp_tx_timestamp(struct sock *sk, u16 tsflags)
467{
468 struct sk_buff *skb = tcp_write_queue_tail(sk);
469
470 if (tsflags && skb) {
471 struct skb_shared_info *shinfo = skb_shinfo(skb);
472 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
473
474 sock_tx_timestamp(sk, tsflags, &shinfo->tx_flags);
475 if (tsflags & SOF_TIMESTAMPING_TX_ACK)
476 tcb->txstamp_ack = 1;
477 if (tsflags & SOF_TIMESTAMPING_TX_RECORD_MASK)
478 shinfo->tskey = TCP_SKB_CB(skb)->seq + skb->len - 1;
479 }
480}
481
482static bool tcp_stream_is_readable(struct sock *sk, int target)
483{
484 if (tcp_epollin_ready(sk, target))
485 return true;
486 return sk_is_readable(sk);
487}
488
489/*
490 * Wait for a TCP event.
491 *
492 * Note that we don't need to lock the socket, as the upper poll layers
493 * take care of normal races (between the test and the event) and we don't
494 * go look at any of the socket buffers directly.
495 */
496__poll_t tcp_poll(struct file *file, struct socket *sock, poll_table *wait)
497{
498 __poll_t mask;
499 struct sock *sk = sock->sk;
500 const struct tcp_sock *tp = tcp_sk(sk);
501 int state;
502
503 sock_poll_wait(file, sock, wait);
504
505 state = inet_sk_state_load(sk);
506 if (state == TCP_LISTEN)
507 return inet_csk_listen_poll(sk);
508
509 /* Socket is not locked. We are protected from async events
510 * by poll logic and correct handling of state changes
511 * made by other threads is impossible in any case.
512 */
513
514 mask = 0;
515
516 /*
517 * EPOLLHUP is certainly not done right. But poll() doesn't
518 * have a notion of HUP in just one direction, and for a
519 * socket the read side is more interesting.
520 *
521 * Some poll() documentation says that EPOLLHUP is incompatible
522 * with the EPOLLOUT/POLLWR flags, so somebody should check this
523 * all. But careful, it tends to be safer to return too many
524 * bits than too few, and you can easily break real applications
525 * if you don't tell them that something has hung up!
526 *
527 * Check-me.
528 *
529 * Check number 1. EPOLLHUP is _UNMASKABLE_ event (see UNIX98 and
530 * our fs/select.c). It means that after we received EOF,
531 * poll always returns immediately, making impossible poll() on write()
532 * in state CLOSE_WAIT. One solution is evident --- to set EPOLLHUP
533 * if and only if shutdown has been made in both directions.
534 * Actually, it is interesting to look how Solaris and DUX
535 * solve this dilemma. I would prefer, if EPOLLHUP were maskable,
536 * then we could set it on SND_SHUTDOWN. BTW examples given
537 * in Stevens' books assume exactly this behaviour, it explains
538 * why EPOLLHUP is incompatible with EPOLLOUT. --ANK
539 *
540 * NOTE. Check for TCP_CLOSE is added. The goal is to prevent
541 * blocking on fresh not-connected or disconnected socket. --ANK
542 */
543 if (sk->sk_shutdown == SHUTDOWN_MASK || state == TCP_CLOSE)
544 mask |= EPOLLHUP;
545 if (sk->sk_shutdown & RCV_SHUTDOWN)
546 mask |= EPOLLIN | EPOLLRDNORM | EPOLLRDHUP;
547
548 /* Connected or passive Fast Open socket? */
549 if (state != TCP_SYN_SENT &&
550 (state != TCP_SYN_RECV || rcu_access_pointer(tp->fastopen_rsk))) {
551 int target = sock_rcvlowat(sk, 0, INT_MAX);
552 u16 urg_data = READ_ONCE(tp->urg_data);
553
554 if (unlikely(urg_data) &&
555 READ_ONCE(tp->urg_seq) == READ_ONCE(tp->copied_seq) &&
556 !sock_flag(sk, SOCK_URGINLINE))
557 target++;
558
559 if (tcp_stream_is_readable(sk, target))
560 mask |= EPOLLIN | EPOLLRDNORM;
561
562 if (!(sk->sk_shutdown & SEND_SHUTDOWN)) {
563 if (__sk_stream_is_writeable(sk, 1)) {
564 mask |= EPOLLOUT | EPOLLWRNORM;
565 } else { /* send SIGIO later */
566 sk_set_bit(SOCKWQ_ASYNC_NOSPACE, sk);
567 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
568
569 /* Race breaker. If space is freed after
570 * wspace test but before the flags are set,
571 * IO signal will be lost. Memory barrier
572 * pairs with the input side.
573 */
574 smp_mb__after_atomic();
575 if (__sk_stream_is_writeable(sk, 1))
576 mask |= EPOLLOUT | EPOLLWRNORM;
577 }
578 } else
579 mask |= EPOLLOUT | EPOLLWRNORM;
580
581 if (urg_data & TCP_URG_VALID)
582 mask |= EPOLLPRI;
583 } else if (state == TCP_SYN_SENT && inet_sk(sk)->defer_connect) {
584 /* Active TCP fastopen socket with defer_connect
585 * Return EPOLLOUT so application can call write()
586 * in order for kernel to generate SYN+data
587 */
588 mask |= EPOLLOUT | EPOLLWRNORM;
589 }
590 /* This barrier is coupled with smp_wmb() in tcp_reset() */
591 smp_rmb();
592 if (sk->sk_err || !skb_queue_empty_lockless(&sk->sk_error_queue))
593 mask |= EPOLLERR;
594
595 return mask;
596}
597EXPORT_SYMBOL(tcp_poll);
598
599int tcp_ioctl(struct sock *sk, int cmd, unsigned long arg)
600{
601 struct tcp_sock *tp = tcp_sk(sk);
602 int answ;
603 bool slow;
604
605 switch (cmd) {
606 case SIOCINQ:
607 if (sk->sk_state == TCP_LISTEN)
608 return -EINVAL;
609
610 slow = lock_sock_fast(sk);
611 answ = tcp_inq(sk);
612 unlock_sock_fast(sk, slow);
613 break;
614 case SIOCATMARK:
615 answ = READ_ONCE(tp->urg_data) &&
616 READ_ONCE(tp->urg_seq) == READ_ONCE(tp->copied_seq);
617 break;
618 case SIOCOUTQ:
619 if (sk->sk_state == TCP_LISTEN)
620 return -EINVAL;
621
622 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
623 answ = 0;
624 else
625 answ = READ_ONCE(tp->write_seq) - tp->snd_una;
626 break;
627 case SIOCOUTQNSD:
628 if (sk->sk_state == TCP_LISTEN)
629 return -EINVAL;
630
631 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
632 answ = 0;
633 else
634 answ = READ_ONCE(tp->write_seq) -
635 READ_ONCE(tp->snd_nxt);
636 break;
637 default:
638 return -ENOIOCTLCMD;
639 }
640
641 return put_user(answ, (int __user *)arg);
642}
643EXPORT_SYMBOL(tcp_ioctl);
644
645void tcp_mark_push(struct tcp_sock *tp, struct sk_buff *skb)
646{
647 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
648 tp->pushed_seq = tp->write_seq;
649}
650
651static inline bool forced_push(const struct tcp_sock *tp)
652{
653 return after(tp->write_seq, tp->pushed_seq + (tp->max_window >> 1));
654}
655
656void tcp_skb_entail(struct sock *sk, struct sk_buff *skb)
657{
658 struct tcp_sock *tp = tcp_sk(sk);
659 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
660
661 tcb->seq = tcb->end_seq = tp->write_seq;
662 tcb->tcp_flags = TCPHDR_ACK;
663 __skb_header_release(skb);
664 tcp_add_write_queue_tail(sk, skb);
665 sk_wmem_queued_add(sk, skb->truesize);
666 sk_mem_charge(sk, skb->truesize);
667 if (tp->nonagle & TCP_NAGLE_PUSH)
668 tp->nonagle &= ~TCP_NAGLE_PUSH;
669
670 tcp_slow_start_after_idle_check(sk);
671}
672
673static inline void tcp_mark_urg(struct tcp_sock *tp, int flags)
674{
675 if (flags & MSG_OOB)
676 tp->snd_up = tp->write_seq;
677}
678
679/* If a not yet filled skb is pushed, do not send it if
680 * we have data packets in Qdisc or NIC queues :
681 * Because TX completion will happen shortly, it gives a chance
682 * to coalesce future sendmsg() payload into this skb, without
683 * need for a timer, and with no latency trade off.
684 * As packets containing data payload have a bigger truesize
685 * than pure acks (dataless) packets, the last checks prevent
686 * autocorking if we only have an ACK in Qdisc/NIC queues,
687 * or if TX completion was delayed after we processed ACK packet.
688 */
689static bool tcp_should_autocork(struct sock *sk, struct sk_buff *skb,
690 int size_goal)
691{
692 return skb->len < size_goal &&
693 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_autocorking) &&
694 !tcp_rtx_queue_empty(sk) &&
695 refcount_read(&sk->sk_wmem_alloc) > skb->truesize &&
696 tcp_skb_can_collapse_to(skb);
697}
698
699void tcp_push(struct sock *sk, int flags, int mss_now,
700 int nonagle, int size_goal)
701{
702 struct tcp_sock *tp = tcp_sk(sk);
703 struct sk_buff *skb;
704
705 skb = tcp_write_queue_tail(sk);
706 if (!skb)
707 return;
708 if (!(flags & MSG_MORE) || forced_push(tp))
709 tcp_mark_push(tp, skb);
710
711 tcp_mark_urg(tp, flags);
712
713 if (tcp_should_autocork(sk, skb, size_goal)) {
714
715 /* avoid atomic op if TSQ_THROTTLED bit is already set */
716 if (!test_bit(TSQ_THROTTLED, &sk->sk_tsq_flags)) {
717 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPAUTOCORKING);
718 set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags);
719 }
720 /* It is possible TX completion already happened
721 * before we set TSQ_THROTTLED.
722 */
723 if (refcount_read(&sk->sk_wmem_alloc) > skb->truesize)
724 return;
725 }
726
727 if (flags & MSG_MORE)
728 nonagle = TCP_NAGLE_CORK;
729
730 __tcp_push_pending_frames(sk, mss_now, nonagle);
731}
732
733static int tcp_splice_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb,
734 unsigned int offset, size_t len)
735{
736 struct tcp_splice_state *tss = rd_desc->arg.data;
737 int ret;
738
739 ret = skb_splice_bits(skb, skb->sk, offset, tss->pipe,
740 min(rd_desc->count, len), tss->flags);
741 if (ret > 0)
742 rd_desc->count -= ret;
743 return ret;
744}
745
746static int __tcp_splice_read(struct sock *sk, struct tcp_splice_state *tss)
747{
748 /* Store TCP splice context information in read_descriptor_t. */
749 read_descriptor_t rd_desc = {
750 .arg.data = tss,
751 .count = tss->len,
752 };
753
754 return tcp_read_sock(sk, &rd_desc, tcp_splice_data_recv);
755}
756
757/**
758 * tcp_splice_read - splice data from TCP socket to a pipe
759 * @sock: socket to splice from
760 * @ppos: position (not valid)
761 * @pipe: pipe to splice to
762 * @len: number of bytes to splice
763 * @flags: splice modifier flags
764 *
765 * Description:
766 * Will read pages from given socket and fill them into a pipe.
767 *
768 **/
769ssize_t tcp_splice_read(struct socket *sock, loff_t *ppos,
770 struct pipe_inode_info *pipe, size_t len,
771 unsigned int flags)
772{
773 struct sock *sk = sock->sk;
774 struct tcp_splice_state tss = {
775 .pipe = pipe,
776 .len = len,
777 .flags = flags,
778 };
779 long timeo;
780 ssize_t spliced;
781 int ret;
782
783 sock_rps_record_flow(sk);
784 /*
785 * We can't seek on a socket input
786 */
787 if (unlikely(*ppos))
788 return -ESPIPE;
789
790 ret = spliced = 0;
791
792 lock_sock(sk);
793
794 timeo = sock_rcvtimeo(sk, sock->file->f_flags & O_NONBLOCK);
795 while (tss.len) {
796 ret = __tcp_splice_read(sk, &tss);
797 if (ret < 0)
798 break;
799 else if (!ret) {
800 if (spliced)
801 break;
802 if (sock_flag(sk, SOCK_DONE))
803 break;
804 if (sk->sk_err) {
805 ret = sock_error(sk);
806 break;
807 }
808 if (sk->sk_shutdown & RCV_SHUTDOWN)
809 break;
810 if (sk->sk_state == TCP_CLOSE) {
811 /*
812 * This occurs when user tries to read
813 * from never connected socket.
814 */
815 ret = -ENOTCONN;
816 break;
817 }
818 if (!timeo) {
819 ret = -EAGAIN;
820 break;
821 }
822 /* if __tcp_splice_read() got nothing while we have
823 * an skb in receive queue, we do not want to loop.
824 * This might happen with URG data.
825 */
826 if (!skb_queue_empty(&sk->sk_receive_queue))
827 break;
828 sk_wait_data(sk, &timeo, NULL);
829 if (signal_pending(current)) {
830 ret = sock_intr_errno(timeo);
831 break;
832 }
833 continue;
834 }
835 tss.len -= ret;
836 spliced += ret;
837
838 if (!timeo)
839 break;
840 release_sock(sk);
841 lock_sock(sk);
842
843 if (sk->sk_err || sk->sk_state == TCP_CLOSE ||
844 (sk->sk_shutdown & RCV_SHUTDOWN) ||
845 signal_pending(current))
846 break;
847 }
848
849 release_sock(sk);
850
851 if (spliced)
852 return spliced;
853
854 return ret;
855}
856EXPORT_SYMBOL(tcp_splice_read);
857
858struct sk_buff *tcp_stream_alloc_skb(struct sock *sk, int size, gfp_t gfp,
859 bool force_schedule)
860{
861 struct sk_buff *skb;
862
863 skb = alloc_skb_fclone(size + MAX_TCP_HEADER, gfp);
864 if (likely(skb)) {
865 bool mem_scheduled;
866
867 skb->truesize = SKB_TRUESIZE(skb_end_offset(skb));
868 if (force_schedule) {
869 mem_scheduled = true;
870 sk_forced_mem_schedule(sk, skb->truesize);
871 } else {
872 mem_scheduled = sk_wmem_schedule(sk, skb->truesize);
873 }
874 if (likely(mem_scheduled)) {
875 skb_reserve(skb, MAX_TCP_HEADER);
876 skb->ip_summed = CHECKSUM_PARTIAL;
877 INIT_LIST_HEAD(&skb->tcp_tsorted_anchor);
878 return skb;
879 }
880 __kfree_skb(skb);
881 } else {
882 sk->sk_prot->enter_memory_pressure(sk);
883 sk_stream_moderate_sndbuf(sk);
884 }
885 return NULL;
886}
887
888static unsigned int tcp_xmit_size_goal(struct sock *sk, u32 mss_now,
889 int large_allowed)
890{
891 struct tcp_sock *tp = tcp_sk(sk);
892 u32 new_size_goal, size_goal;
893
894 if (!large_allowed)
895 return mss_now;
896
897 /* Note : tcp_tso_autosize() will eventually split this later */
898 new_size_goal = tcp_bound_to_half_wnd(tp, sk->sk_gso_max_size);
899
900 /* We try hard to avoid divides here */
901 size_goal = tp->gso_segs * mss_now;
902 if (unlikely(new_size_goal < size_goal ||
903 new_size_goal >= size_goal + mss_now)) {
904 tp->gso_segs = min_t(u16, new_size_goal / mss_now,
905 sk->sk_gso_max_segs);
906 size_goal = tp->gso_segs * mss_now;
907 }
908
909 return max(size_goal, mss_now);
910}
911
912int tcp_send_mss(struct sock *sk, int *size_goal, int flags)
913{
914 int mss_now;
915
916 mss_now = tcp_current_mss(sk);
917 *size_goal = tcp_xmit_size_goal(sk, mss_now, !(flags & MSG_OOB));
918
919 return mss_now;
920}
921
922/* In some cases, both sendpage() and sendmsg() could have added
923 * an skb to the write queue, but failed adding payload on it.
924 * We need to remove it to consume less memory, but more
925 * importantly be able to generate EPOLLOUT for Edge Trigger epoll()
926 * users.
927 */
928void tcp_remove_empty_skb(struct sock *sk)
929{
930 struct sk_buff *skb = tcp_write_queue_tail(sk);
931
932 if (skb && TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq) {
933 tcp_unlink_write_queue(skb, sk);
934 if (tcp_write_queue_empty(sk))
935 tcp_chrono_stop(sk, TCP_CHRONO_BUSY);
936 tcp_wmem_free_skb(sk, skb);
937 }
938}
939
940/* skb changing from pure zc to mixed, must charge zc */
941static int tcp_downgrade_zcopy_pure(struct sock *sk, struct sk_buff *skb)
942{
943 if (unlikely(skb_zcopy_pure(skb))) {
944 u32 extra = skb->truesize -
945 SKB_TRUESIZE(skb_end_offset(skb));
946
947 if (!sk_wmem_schedule(sk, extra))
948 return -ENOMEM;
949
950 sk_mem_charge(sk, extra);
951 skb_shinfo(skb)->flags &= ~SKBFL_PURE_ZEROCOPY;
952 }
953 return 0;
954}
955
956
957static int tcp_wmem_schedule(struct sock *sk, int copy)
958{
959 int left;
960
961 if (likely(sk_wmem_schedule(sk, copy)))
962 return copy;
963
964 /* We could be in trouble if we have nothing queued.
965 * Use whatever is left in sk->sk_forward_alloc and tcp_wmem[0]
966 * to guarantee some progress.
967 */
968 left = sock_net(sk)->ipv4.sysctl_tcp_wmem[0] - sk->sk_wmem_queued;
969 if (left > 0)
970 sk_forced_mem_schedule(sk, min(left, copy));
971 return min(copy, sk->sk_forward_alloc);
972}
973
974static struct sk_buff *tcp_build_frag(struct sock *sk, int size_goal, int flags,
975 struct page *page, int offset, size_t *size)
976{
977 struct sk_buff *skb = tcp_write_queue_tail(sk);
978 struct tcp_sock *tp = tcp_sk(sk);
979 bool can_coalesce;
980 int copy, i;
981
982 if (!skb || (copy = size_goal - skb->len) <= 0 ||
983 !tcp_skb_can_collapse_to(skb)) {
984new_segment:
985 if (!sk_stream_memory_free(sk))
986 return NULL;
987
988 skb = tcp_stream_alloc_skb(sk, 0, sk->sk_allocation,
989 tcp_rtx_and_write_queues_empty(sk));
990 if (!skb)
991 return NULL;
992
993#ifdef CONFIG_TLS_DEVICE
994 skb->decrypted = !!(flags & MSG_SENDPAGE_DECRYPTED);
995#endif
996 tcp_skb_entail(sk, skb);
997 copy = size_goal;
998 }
999
1000 if (copy > *size)
1001 copy = *size;
1002
1003 i = skb_shinfo(skb)->nr_frags;
1004 can_coalesce = skb_can_coalesce(skb, i, page, offset);
1005 if (!can_coalesce && i >= READ_ONCE(sysctl_max_skb_frags)) {
1006 tcp_mark_push(tp, skb);
1007 goto new_segment;
1008 }
1009 if (tcp_downgrade_zcopy_pure(sk, skb))
1010 return NULL;
1011
1012 copy = tcp_wmem_schedule(sk, copy);
1013 if (!copy)
1014 return NULL;
1015
1016 if (can_coalesce) {
1017 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1018 } else {
1019 get_page(page);
1020 skb_fill_page_desc_noacc(skb, i, page, offset, copy);
1021 }
1022
1023 if (!(flags & MSG_NO_SHARED_FRAGS))
1024 skb_shinfo(skb)->flags |= SKBFL_SHARED_FRAG;
1025
1026 skb->len += copy;
1027 skb->data_len += copy;
1028 skb->truesize += copy;
1029 sk_wmem_queued_add(sk, copy);
1030 sk_mem_charge(sk, copy);
1031 WRITE_ONCE(tp->write_seq, tp->write_seq + copy);
1032 TCP_SKB_CB(skb)->end_seq += copy;
1033 tcp_skb_pcount_set(skb, 0);
1034
1035 *size = copy;
1036 return skb;
1037}
1038
1039ssize_t do_tcp_sendpages(struct sock *sk, struct page *page, int offset,
1040 size_t size, int flags)
1041{
1042 struct tcp_sock *tp = tcp_sk(sk);
1043 int mss_now, size_goal;
1044 int err;
1045 ssize_t copied;
1046 long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
1047
1048 if (IS_ENABLED(CONFIG_DEBUG_VM) &&
1049 WARN_ONCE(!sendpage_ok(page),
1050 "page must not be a Slab one and have page_count > 0"))
1051 return -EINVAL;
1052
1053 /* Wait for a connection to finish. One exception is TCP Fast Open
1054 * (passive side) where data is allowed to be sent before a connection
1055 * is fully established.
1056 */
1057 if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) &&
1058 !tcp_passive_fastopen(sk)) {
1059 err = sk_stream_wait_connect(sk, &timeo);
1060 if (err != 0)
1061 goto out_err;
1062 }
1063
1064 sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
1065
1066 mss_now = tcp_send_mss(sk, &size_goal, flags);
1067 copied = 0;
1068
1069 err = -EPIPE;
1070 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
1071 goto out_err;
1072
1073 while (size > 0) {
1074 struct sk_buff *skb;
1075 size_t copy = size;
1076
1077 skb = tcp_build_frag(sk, size_goal, flags, page, offset, ©);
1078 if (!skb)
1079 goto wait_for_space;
1080
1081 if (!copied)
1082 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
1083
1084 copied += copy;
1085 offset += copy;
1086 size -= copy;
1087 if (!size)
1088 goto out;
1089
1090 if (skb->len < size_goal || (flags & MSG_OOB))
1091 continue;
1092
1093 if (forced_push(tp)) {
1094 tcp_mark_push(tp, skb);
1095 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
1096 } else if (skb == tcp_send_head(sk))
1097 tcp_push_one(sk, mss_now);
1098 continue;
1099
1100wait_for_space:
1101 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1102 tcp_push(sk, flags & ~MSG_MORE, mss_now,
1103 TCP_NAGLE_PUSH, size_goal);
1104
1105 err = sk_stream_wait_memory(sk, &timeo);
1106 if (err != 0)
1107 goto do_error;
1108
1109 mss_now = tcp_send_mss(sk, &size_goal, flags);
1110 }
1111
1112out:
1113 if (copied) {
1114 tcp_tx_timestamp(sk, sk->sk_tsflags);
1115 if (!(flags & MSG_SENDPAGE_NOTLAST))
1116 tcp_push(sk, flags, mss_now, tp->nonagle, size_goal);
1117 }
1118 return copied;
1119
1120do_error:
1121 tcp_remove_empty_skb(sk);
1122 if (copied)
1123 goto out;
1124out_err:
1125 /* make sure we wake any epoll edge trigger waiter */
1126 if (unlikely(tcp_rtx_and_write_queues_empty(sk) && err == -EAGAIN)) {
1127 sk->sk_write_space(sk);
1128 tcp_chrono_stop(sk, TCP_CHRONO_SNDBUF_LIMITED);
1129 }
1130 return sk_stream_error(sk, flags, err);
1131}
1132EXPORT_SYMBOL_GPL(do_tcp_sendpages);
1133
1134int tcp_sendpage_locked(struct sock *sk, struct page *page, int offset,
1135 size_t size, int flags)
1136{
1137 if (!(sk->sk_route_caps & NETIF_F_SG))
1138 return sock_no_sendpage_locked(sk, page, offset, size, flags);
1139
1140 tcp_rate_check_app_limited(sk); /* is sending application-limited? */
1141
1142 return do_tcp_sendpages(sk, page, offset, size, flags);
1143}
1144EXPORT_SYMBOL_GPL(tcp_sendpage_locked);
1145
1146int tcp_sendpage(struct sock *sk, struct page *page, int offset,
1147 size_t size, int flags)
1148{
1149 int ret;
1150
1151 lock_sock(sk);
1152 ret = tcp_sendpage_locked(sk, page, offset, size, flags);
1153 release_sock(sk);
1154
1155 return ret;
1156}
1157EXPORT_SYMBOL(tcp_sendpage);
1158
1159void tcp_free_fastopen_req(struct tcp_sock *tp)
1160{
1161 if (tp->fastopen_req) {
1162 kfree(tp->fastopen_req);
1163 tp->fastopen_req = NULL;
1164 }
1165}
1166
1167int tcp_sendmsg_fastopen(struct sock *sk, struct msghdr *msg, int *copied,
1168 size_t size, struct ubuf_info *uarg)
1169{
1170 struct tcp_sock *tp = tcp_sk(sk);
1171 struct inet_sock *inet = inet_sk(sk);
1172 struct sockaddr *uaddr = msg->msg_name;
1173 int err, flags;
1174
1175 if (!(READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_fastopen) &
1176 TFO_CLIENT_ENABLE) ||
1177 (uaddr && msg->msg_namelen >= sizeof(uaddr->sa_family) &&
1178 uaddr->sa_family == AF_UNSPEC))
1179 return -EOPNOTSUPP;
1180 if (tp->fastopen_req)
1181 return -EALREADY; /* Another Fast Open is in progress */
1182
1183 tp->fastopen_req = kzalloc(sizeof(struct tcp_fastopen_request),
1184 sk->sk_allocation);
1185 if (unlikely(!tp->fastopen_req))
1186 return -ENOBUFS;
1187 tp->fastopen_req->data = msg;
1188 tp->fastopen_req->size = size;
1189 tp->fastopen_req->uarg = uarg;
1190
1191 if (inet->defer_connect) {
1192 err = tcp_connect(sk);
1193 /* Same failure procedure as in tcp_v4/6_connect */
1194 if (err) {
1195 tcp_set_state(sk, TCP_CLOSE);
1196 inet->inet_dport = 0;
1197 sk->sk_route_caps = 0;
1198 }
1199 }
1200 flags = (msg->msg_flags & MSG_DONTWAIT) ? O_NONBLOCK : 0;
1201 err = __inet_stream_connect(sk->sk_socket, uaddr,
1202 msg->msg_namelen, flags, 1);
1203 /* fastopen_req could already be freed in __inet_stream_connect
1204 * if the connection times out or gets rst
1205 */
1206 if (tp->fastopen_req) {
1207 *copied = tp->fastopen_req->copied;
1208 tcp_free_fastopen_req(tp);
1209 inet->defer_connect = 0;
1210 }
1211 return err;
1212}
1213
1214int tcp_sendmsg_locked(struct sock *sk, struct msghdr *msg, size_t size)
1215{
1216 struct tcp_sock *tp = tcp_sk(sk);
1217 struct ubuf_info *uarg = NULL;
1218 struct sk_buff *skb;
1219 struct sockcm_cookie sockc;
1220 int flags, err, copied = 0;
1221 int mss_now = 0, size_goal, copied_syn = 0;
1222 int process_backlog = 0;
1223 bool zc = false;
1224 long timeo;
1225
1226 flags = msg->msg_flags;
1227
1228 if ((flags & MSG_ZEROCOPY) && size) {
1229 skb = tcp_write_queue_tail(sk);
1230
1231 if (msg->msg_ubuf) {
1232 uarg = msg->msg_ubuf;
1233 net_zcopy_get(uarg);
1234 zc = sk->sk_route_caps & NETIF_F_SG;
1235 } else if (sock_flag(sk, SOCK_ZEROCOPY)) {
1236 uarg = msg_zerocopy_realloc(sk, size, skb_zcopy(skb));
1237 if (!uarg) {
1238 err = -ENOBUFS;
1239 goto out_err;
1240 }
1241 zc = sk->sk_route_caps & NETIF_F_SG;
1242 if (!zc)
1243 uarg_to_msgzc(uarg)->zerocopy = 0;
1244 }
1245 }
1246
1247 if (unlikely(flags & MSG_FASTOPEN || inet_sk(sk)->defer_connect) &&
1248 !tp->repair) {
1249 err = tcp_sendmsg_fastopen(sk, msg, &copied_syn, size, uarg);
1250 if (err == -EINPROGRESS && copied_syn > 0)
1251 goto out;
1252 else if (err)
1253 goto out_err;
1254 }
1255
1256 timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
1257
1258 tcp_rate_check_app_limited(sk); /* is sending application-limited? */
1259
1260 /* Wait for a connection to finish. One exception is TCP Fast Open
1261 * (passive side) where data is allowed to be sent before a connection
1262 * is fully established.
1263 */
1264 if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) &&
1265 !tcp_passive_fastopen(sk)) {
1266 err = sk_stream_wait_connect(sk, &timeo);
1267 if (err != 0)
1268 goto do_error;
1269 }
1270
1271 if (unlikely(tp->repair)) {
1272 if (tp->repair_queue == TCP_RECV_QUEUE) {
1273 copied = tcp_send_rcvq(sk, msg, size);
1274 goto out_nopush;
1275 }
1276
1277 err = -EINVAL;
1278 if (tp->repair_queue == TCP_NO_QUEUE)
1279 goto out_err;
1280
1281 /* 'common' sending to sendq */
1282 }
1283
1284 sockcm_init(&sockc, sk);
1285 if (msg->msg_controllen) {
1286 err = sock_cmsg_send(sk, msg, &sockc);
1287 if (unlikely(err)) {
1288 err = -EINVAL;
1289 goto out_err;
1290 }
1291 }
1292
1293 /* This should be in poll */
1294 sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
1295
1296 /* Ok commence sending. */
1297 copied = 0;
1298
1299restart:
1300 mss_now = tcp_send_mss(sk, &size_goal, flags);
1301
1302 err = -EPIPE;
1303 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
1304 goto do_error;
1305
1306 while (msg_data_left(msg)) {
1307 int copy = 0;
1308
1309 skb = tcp_write_queue_tail(sk);
1310 if (skb)
1311 copy = size_goal - skb->len;
1312
1313 if (copy <= 0 || !tcp_skb_can_collapse_to(skb)) {
1314 bool first_skb;
1315
1316new_segment:
1317 if (!sk_stream_memory_free(sk))
1318 goto wait_for_space;
1319
1320 if (unlikely(process_backlog >= 16)) {
1321 process_backlog = 0;
1322 if (sk_flush_backlog(sk))
1323 goto restart;
1324 }
1325 first_skb = tcp_rtx_and_write_queues_empty(sk);
1326 skb = tcp_stream_alloc_skb(sk, 0, sk->sk_allocation,
1327 first_skb);
1328 if (!skb)
1329 goto wait_for_space;
1330
1331 process_backlog++;
1332
1333 tcp_skb_entail(sk, skb);
1334 copy = size_goal;
1335
1336 /* All packets are restored as if they have
1337 * already been sent. skb_mstamp_ns isn't set to
1338 * avoid wrong rtt estimation.
1339 */
1340 if (tp->repair)
1341 TCP_SKB_CB(skb)->sacked |= TCPCB_REPAIRED;
1342 }
1343
1344 /* Try to append data to the end of skb. */
1345 if (copy > msg_data_left(msg))
1346 copy = msg_data_left(msg);
1347
1348 if (!zc) {
1349 bool merge = true;
1350 int i = skb_shinfo(skb)->nr_frags;
1351 struct page_frag *pfrag = sk_page_frag(sk);
1352
1353 if (!sk_page_frag_refill(sk, pfrag))
1354 goto wait_for_space;
1355
1356 if (!skb_can_coalesce(skb, i, pfrag->page,
1357 pfrag->offset)) {
1358 if (i >= READ_ONCE(sysctl_max_skb_frags)) {
1359 tcp_mark_push(tp, skb);
1360 goto new_segment;
1361 }
1362 merge = false;
1363 }
1364
1365 copy = min_t(int, copy, pfrag->size - pfrag->offset);
1366
1367 if (unlikely(skb_zcopy_pure(skb) || skb_zcopy_managed(skb))) {
1368 if (tcp_downgrade_zcopy_pure(sk, skb))
1369 goto wait_for_space;
1370 skb_zcopy_downgrade_managed(skb);
1371 }
1372
1373 copy = tcp_wmem_schedule(sk, copy);
1374 if (!copy)
1375 goto wait_for_space;
1376
1377 err = skb_copy_to_page_nocache(sk, &msg->msg_iter, skb,
1378 pfrag->page,
1379 pfrag->offset,
1380 copy);
1381 if (err)
1382 goto do_error;
1383
1384 /* Update the skb. */
1385 if (merge) {
1386 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1387 } else {
1388 skb_fill_page_desc(skb, i, pfrag->page,
1389 pfrag->offset, copy);
1390 page_ref_inc(pfrag->page);
1391 }
1392 pfrag->offset += copy;
1393 } else {
1394 /* First append to a fragless skb builds initial
1395 * pure zerocopy skb
1396 */
1397 if (!skb->len)
1398 skb_shinfo(skb)->flags |= SKBFL_PURE_ZEROCOPY;
1399
1400 if (!skb_zcopy_pure(skb)) {
1401 copy = tcp_wmem_schedule(sk, copy);
1402 if (!copy)
1403 goto wait_for_space;
1404 }
1405
1406 err = skb_zerocopy_iter_stream(sk, skb, msg, copy, uarg);
1407 if (err == -EMSGSIZE || err == -EEXIST) {
1408 tcp_mark_push(tp, skb);
1409 goto new_segment;
1410 }
1411 if (err < 0)
1412 goto do_error;
1413 copy = err;
1414 }
1415
1416 if (!copied)
1417 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
1418
1419 WRITE_ONCE(tp->write_seq, tp->write_seq + copy);
1420 TCP_SKB_CB(skb)->end_seq += copy;
1421 tcp_skb_pcount_set(skb, 0);
1422
1423 copied += copy;
1424 if (!msg_data_left(msg)) {
1425 if (unlikely(flags & MSG_EOR))
1426 TCP_SKB_CB(skb)->eor = 1;
1427 goto out;
1428 }
1429
1430 if (skb->len < size_goal || (flags & MSG_OOB) || unlikely(tp->repair))
1431 continue;
1432
1433 if (forced_push(tp)) {
1434 tcp_mark_push(tp, skb);
1435 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
1436 } else if (skb == tcp_send_head(sk))
1437 tcp_push_one(sk, mss_now);
1438 continue;
1439
1440wait_for_space:
1441 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1442 if (copied)
1443 tcp_push(sk, flags & ~MSG_MORE, mss_now,
1444 TCP_NAGLE_PUSH, size_goal);
1445
1446 err = sk_stream_wait_memory(sk, &timeo);
1447 if (err != 0)
1448 goto do_error;
1449
1450 mss_now = tcp_send_mss(sk, &size_goal, flags);
1451 }
1452
1453out:
1454 if (copied) {
1455 tcp_tx_timestamp(sk, sockc.tsflags);
1456 tcp_push(sk, flags, mss_now, tp->nonagle, size_goal);
1457 }
1458out_nopush:
1459 net_zcopy_put(uarg);
1460 return copied + copied_syn;
1461
1462do_error:
1463 tcp_remove_empty_skb(sk);
1464
1465 if (copied + copied_syn)
1466 goto out;
1467out_err:
1468 net_zcopy_put_abort(uarg, true);
1469 err = sk_stream_error(sk, flags, err);
1470 /* make sure we wake any epoll edge trigger waiter */
1471 if (unlikely(tcp_rtx_and_write_queues_empty(sk) && err == -EAGAIN)) {
1472 sk->sk_write_space(sk);
1473 tcp_chrono_stop(sk, TCP_CHRONO_SNDBUF_LIMITED);
1474 }
1475 return err;
1476}
1477EXPORT_SYMBOL_GPL(tcp_sendmsg_locked);
1478
1479int tcp_sendmsg(struct sock *sk, struct msghdr *msg, size_t size)
1480{
1481 int ret;
1482
1483 lock_sock(sk);
1484 ret = tcp_sendmsg_locked(sk, msg, size);
1485 release_sock(sk);
1486
1487 return ret;
1488}
1489EXPORT_SYMBOL(tcp_sendmsg);
1490
1491/*
1492 * Handle reading urgent data. BSD has very simple semantics for
1493 * this, no blocking and very strange errors 8)
1494 */
1495
1496static int tcp_recv_urg(struct sock *sk, struct msghdr *msg, int len, int flags)
1497{
1498 struct tcp_sock *tp = tcp_sk(sk);
1499
1500 /* No URG data to read. */
1501 if (sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data ||
1502 tp->urg_data == TCP_URG_READ)
1503 return -EINVAL; /* Yes this is right ! */
1504
1505 if (sk->sk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DONE))
1506 return -ENOTCONN;
1507
1508 if (tp->urg_data & TCP_URG_VALID) {
1509 int err = 0;
1510 char c = tp->urg_data;
1511
1512 if (!(flags & MSG_PEEK))
1513 WRITE_ONCE(tp->urg_data, TCP_URG_READ);
1514
1515 /* Read urgent data. */
1516 msg->msg_flags |= MSG_OOB;
1517
1518 if (len > 0) {
1519 if (!(flags & MSG_TRUNC))
1520 err = memcpy_to_msg(msg, &c, 1);
1521 len = 1;
1522 } else
1523 msg->msg_flags |= MSG_TRUNC;
1524
1525 return err ? -EFAULT : len;
1526 }
1527
1528 if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
1529 return 0;
1530
1531 /* Fixed the recv(..., MSG_OOB) behaviour. BSD docs and
1532 * the available implementations agree in this case:
1533 * this call should never block, independent of the
1534 * blocking state of the socket.
1535 * Mike <pall@rz.uni-karlsruhe.de>
1536 */
1537 return -EAGAIN;
1538}
1539
1540static int tcp_peek_sndq(struct sock *sk, struct msghdr *msg, int len)
1541{
1542 struct sk_buff *skb;
1543 int copied = 0, err = 0;
1544
1545 /* XXX -- need to support SO_PEEK_OFF */
1546
1547 skb_rbtree_walk(skb, &sk->tcp_rtx_queue) {
1548 err = skb_copy_datagram_msg(skb, 0, msg, skb->len);
1549 if (err)
1550 return err;
1551 copied += skb->len;
1552 }
1553
1554 skb_queue_walk(&sk->sk_write_queue, skb) {
1555 err = skb_copy_datagram_msg(skb, 0, msg, skb->len);
1556 if (err)
1557 break;
1558
1559 copied += skb->len;
1560 }
1561
1562 return err ?: copied;
1563}
1564
1565/* Clean up the receive buffer for full frames taken by the user,
1566 * then send an ACK if necessary. COPIED is the number of bytes
1567 * tcp_recvmsg has given to the user so far, it speeds up the
1568 * calculation of whether or not we must ACK for the sake of
1569 * a window update.
1570 */
1571static void __tcp_cleanup_rbuf(struct sock *sk, int copied)
1572{
1573 struct tcp_sock *tp = tcp_sk(sk);
1574 bool time_to_ack = false;
1575
1576 if (inet_csk_ack_scheduled(sk)) {
1577 const struct inet_connection_sock *icsk = inet_csk(sk);
1578
1579 if (/* Once-per-two-segments ACK was not sent by tcp_input.c */
1580 tp->rcv_nxt - tp->rcv_wup > icsk->icsk_ack.rcv_mss ||
1581 /*
1582 * If this read emptied read buffer, we send ACK, if
1583 * connection is not bidirectional, user drained
1584 * receive buffer and there was a small segment
1585 * in queue.
1586 */
1587 (copied > 0 &&
1588 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED2) ||
1589 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED) &&
1590 !inet_csk_in_pingpong_mode(sk))) &&
1591 !atomic_read(&sk->sk_rmem_alloc)))
1592 time_to_ack = true;
1593 }
1594
1595 /* We send an ACK if we can now advertise a non-zero window
1596 * which has been raised "significantly".
1597 *
1598 * Even if window raised up to infinity, do not send window open ACK
1599 * in states, where we will not receive more. It is useless.
1600 */
1601 if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) {
1602 __u32 rcv_window_now = tcp_receive_window(tp);
1603
1604 /* Optimize, __tcp_select_window() is not cheap. */
1605 if (2*rcv_window_now <= tp->window_clamp) {
1606 __u32 new_window = __tcp_select_window(sk);
1607
1608 /* Send ACK now, if this read freed lots of space
1609 * in our buffer. Certainly, new_window is new window.
1610 * We can advertise it now, if it is not less than current one.
1611 * "Lots" means "at least twice" here.
1612 */
1613 if (new_window && new_window >= 2 * rcv_window_now)
1614 time_to_ack = true;
1615 }
1616 }
1617 if (time_to_ack)
1618 tcp_send_ack(sk);
1619}
1620
1621void tcp_cleanup_rbuf(struct sock *sk, int copied)
1622{
1623 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
1624 struct tcp_sock *tp = tcp_sk(sk);
1625
1626 WARN(skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq),
1627 "cleanup rbuf bug: copied %X seq %X rcvnxt %X\n",
1628 tp->copied_seq, TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt);
1629 __tcp_cleanup_rbuf(sk, copied);
1630}
1631
1632static void tcp_eat_recv_skb(struct sock *sk, struct sk_buff *skb)
1633{
1634 __skb_unlink(skb, &sk->sk_receive_queue);
1635 if (likely(skb->destructor == sock_rfree)) {
1636 sock_rfree(skb);
1637 skb->destructor = NULL;
1638 skb->sk = NULL;
1639 return skb_attempt_defer_free(skb);
1640 }
1641 __kfree_skb(skb);
1642}
1643
1644struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off)
1645{
1646 struct sk_buff *skb;
1647 u32 offset;
1648
1649 while ((skb = skb_peek(&sk->sk_receive_queue)) != NULL) {
1650 offset = seq - TCP_SKB_CB(skb)->seq;
1651 if (unlikely(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
1652 pr_err_once("%s: found a SYN, please report !\n", __func__);
1653 offset--;
1654 }
1655 if (offset < skb->len || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)) {
1656 *off = offset;
1657 return skb;
1658 }
1659 /* This looks weird, but this can happen if TCP collapsing
1660 * splitted a fat GRO packet, while we released socket lock
1661 * in skb_splice_bits()
1662 */
1663 tcp_eat_recv_skb(sk, skb);
1664 }
1665 return NULL;
1666}
1667EXPORT_SYMBOL(tcp_recv_skb);
1668
1669/*
1670 * This routine provides an alternative to tcp_recvmsg() for routines
1671 * that would like to handle copying from skbuffs directly in 'sendfile'
1672 * fashion.
1673 * Note:
1674 * - It is assumed that the socket was locked by the caller.
1675 * - The routine does not block.
1676 * - At present, there is no support for reading OOB data
1677 * or for 'peeking' the socket using this routine
1678 * (although both would be easy to implement).
1679 */
1680int tcp_read_sock(struct sock *sk, read_descriptor_t *desc,
1681 sk_read_actor_t recv_actor)
1682{
1683 struct sk_buff *skb;
1684 struct tcp_sock *tp = tcp_sk(sk);
1685 u32 seq = tp->copied_seq;
1686 u32 offset;
1687 int copied = 0;
1688
1689 if (sk->sk_state == TCP_LISTEN)
1690 return -ENOTCONN;
1691 while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1692 if (offset < skb->len) {
1693 int used;
1694 size_t len;
1695
1696 len = skb->len - offset;
1697 /* Stop reading if we hit a patch of urgent data */
1698 if (unlikely(tp->urg_data)) {
1699 u32 urg_offset = tp->urg_seq - seq;
1700 if (urg_offset < len)
1701 len = urg_offset;
1702 if (!len)
1703 break;
1704 }
1705 used = recv_actor(desc, skb, offset, len);
1706 if (used <= 0) {
1707 if (!copied)
1708 copied = used;
1709 break;
1710 }
1711 if (WARN_ON_ONCE(used > len))
1712 used = len;
1713 seq += used;
1714 copied += used;
1715 offset += used;
1716
1717 /* If recv_actor drops the lock (e.g. TCP splice
1718 * receive) the skb pointer might be invalid when
1719 * getting here: tcp_collapse might have deleted it
1720 * while aggregating skbs from the socket queue.
1721 */
1722 skb = tcp_recv_skb(sk, seq - 1, &offset);
1723 if (!skb)
1724 break;
1725 /* TCP coalescing might have appended data to the skb.
1726 * Try to splice more frags
1727 */
1728 if (offset + 1 != skb->len)
1729 continue;
1730 }
1731 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) {
1732 tcp_eat_recv_skb(sk, skb);
1733 ++seq;
1734 break;
1735 }
1736 tcp_eat_recv_skb(sk, skb);
1737 if (!desc->count)
1738 break;
1739 WRITE_ONCE(tp->copied_seq, seq);
1740 }
1741 WRITE_ONCE(tp->copied_seq, seq);
1742
1743 tcp_rcv_space_adjust(sk);
1744
1745 /* Clean up data we have read: This will do ACK frames. */
1746 if (copied > 0) {
1747 tcp_recv_skb(sk, seq, &offset);
1748 tcp_cleanup_rbuf(sk, copied);
1749 }
1750 return copied;
1751}
1752EXPORT_SYMBOL(tcp_read_sock);
1753
1754int tcp_read_skb(struct sock *sk, skb_read_actor_t recv_actor)
1755{
1756 struct tcp_sock *tp = tcp_sk(sk);
1757 u32 seq = tp->copied_seq;
1758 struct sk_buff *skb;
1759 int copied = 0;
1760 u32 offset;
1761
1762 if (sk->sk_state == TCP_LISTEN)
1763 return -ENOTCONN;
1764
1765 while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1766 u8 tcp_flags;
1767 int used;
1768
1769 __skb_unlink(skb, &sk->sk_receive_queue);
1770 WARN_ON_ONCE(!skb_set_owner_sk_safe(skb, sk));
1771 tcp_flags = TCP_SKB_CB(skb)->tcp_flags;
1772 used = recv_actor(sk, skb);
1773 consume_skb(skb);
1774 if (used < 0) {
1775 if (!copied)
1776 copied = used;
1777 break;
1778 }
1779 seq += used;
1780 copied += used;
1781
1782 if (tcp_flags & TCPHDR_FIN) {
1783 ++seq;
1784 break;
1785 }
1786 }
1787 WRITE_ONCE(tp->copied_seq, seq);
1788
1789 tcp_rcv_space_adjust(sk);
1790
1791 /* Clean up data we have read: This will do ACK frames. */
1792 if (copied > 0)
1793 __tcp_cleanup_rbuf(sk, copied);
1794
1795 return copied;
1796}
1797EXPORT_SYMBOL(tcp_read_skb);
1798
1799void tcp_read_done(struct sock *sk, size_t len)
1800{
1801 struct tcp_sock *tp = tcp_sk(sk);
1802 u32 seq = tp->copied_seq;
1803 struct sk_buff *skb;
1804 size_t left;
1805 u32 offset;
1806
1807 if (sk->sk_state == TCP_LISTEN)
1808 return;
1809
1810 left = len;
1811 while (left && (skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1812 int used;
1813
1814 used = min_t(size_t, skb->len - offset, left);
1815 seq += used;
1816 left -= used;
1817
1818 if (skb->len > offset + used)
1819 break;
1820
1821 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) {
1822 tcp_eat_recv_skb(sk, skb);
1823 ++seq;
1824 break;
1825 }
1826 tcp_eat_recv_skb(sk, skb);
1827 }
1828 WRITE_ONCE(tp->copied_seq, seq);
1829
1830 tcp_rcv_space_adjust(sk);
1831
1832 /* Clean up data we have read: This will do ACK frames. */
1833 if (left != len)
1834 tcp_cleanup_rbuf(sk, len - left);
1835}
1836EXPORT_SYMBOL(tcp_read_done);
1837
1838int tcp_peek_len(struct socket *sock)
1839{
1840 return tcp_inq(sock->sk);
1841}
1842EXPORT_SYMBOL(tcp_peek_len);
1843
1844/* Make sure sk_rcvbuf is big enough to satisfy SO_RCVLOWAT hint */
1845int tcp_set_rcvlowat(struct sock *sk, int val)
1846{
1847 int cap;
1848
1849 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK)
1850 cap = sk->sk_rcvbuf >> 1;
1851 else
1852 cap = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_rmem[2]) >> 1;
1853 val = min(val, cap);
1854 WRITE_ONCE(sk->sk_rcvlowat, val ? : 1);
1855
1856 /* Check if we need to signal EPOLLIN right now */
1857 tcp_data_ready(sk);
1858
1859 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK)
1860 return 0;
1861
1862 val <<= 1;
1863 if (val > sk->sk_rcvbuf) {
1864 WRITE_ONCE(sk->sk_rcvbuf, val);
1865 tcp_sk(sk)->window_clamp = tcp_win_from_space(sk, val);
1866 }
1867 return 0;
1868}
1869EXPORT_SYMBOL(tcp_set_rcvlowat);
1870
1871void tcp_update_recv_tstamps(struct sk_buff *skb,
1872 struct scm_timestamping_internal *tss)
1873{
1874 if (skb->tstamp)
1875 tss->ts[0] = ktime_to_timespec64(skb->tstamp);
1876 else
1877 tss->ts[0] = (struct timespec64) {0};
1878
1879 if (skb_hwtstamps(skb)->hwtstamp)
1880 tss->ts[2] = ktime_to_timespec64(skb_hwtstamps(skb)->hwtstamp);
1881 else
1882 tss->ts[2] = (struct timespec64) {0};
1883}
1884
1885#ifdef CONFIG_MMU
1886static const struct vm_operations_struct tcp_vm_ops = {
1887};
1888
1889int tcp_mmap(struct file *file, struct socket *sock,
1890 struct vm_area_struct *vma)
1891{
1892 if (vma->vm_flags & (VM_WRITE | VM_EXEC))
1893 return -EPERM;
1894 vma->vm_flags &= ~(VM_MAYWRITE | VM_MAYEXEC);
1895
1896 /* Instruct vm_insert_page() to not mmap_read_lock(mm) */
1897 vma->vm_flags |= VM_MIXEDMAP;
1898
1899 vma->vm_ops = &tcp_vm_ops;
1900 return 0;
1901}
1902EXPORT_SYMBOL(tcp_mmap);
1903
1904static skb_frag_t *skb_advance_to_frag(struct sk_buff *skb, u32 offset_skb,
1905 u32 *offset_frag)
1906{
1907 skb_frag_t *frag;
1908
1909 if (unlikely(offset_skb >= skb->len))
1910 return NULL;
1911
1912 offset_skb -= skb_headlen(skb);
1913 if ((int)offset_skb < 0 || skb_has_frag_list(skb))
1914 return NULL;
1915
1916 frag = skb_shinfo(skb)->frags;
1917 while (offset_skb) {
1918 if (skb_frag_size(frag) > offset_skb) {
1919 *offset_frag = offset_skb;
1920 return frag;
1921 }
1922 offset_skb -= skb_frag_size(frag);
1923 ++frag;
1924 }
1925 *offset_frag = 0;
1926 return frag;
1927}
1928
1929static bool can_map_frag(const skb_frag_t *frag)
1930{
1931 return skb_frag_size(frag) == PAGE_SIZE && !skb_frag_off(frag);
1932}
1933
1934static int find_next_mappable_frag(const skb_frag_t *frag,
1935 int remaining_in_skb)
1936{
1937 int offset = 0;
1938
1939 if (likely(can_map_frag(frag)))
1940 return 0;
1941
1942 while (offset < remaining_in_skb && !can_map_frag(frag)) {
1943 offset += skb_frag_size(frag);
1944 ++frag;
1945 }
1946 return offset;
1947}
1948
1949static void tcp_zerocopy_set_hint_for_skb(struct sock *sk,
1950 struct tcp_zerocopy_receive *zc,
1951 struct sk_buff *skb, u32 offset)
1952{
1953 u32 frag_offset, partial_frag_remainder = 0;
1954 int mappable_offset;
1955 skb_frag_t *frag;
1956
1957 /* worst case: skip to next skb. try to improve on this case below */
1958 zc->recv_skip_hint = skb->len - offset;
1959
1960 /* Find the frag containing this offset (and how far into that frag) */
1961 frag = skb_advance_to_frag(skb, offset, &frag_offset);
1962 if (!frag)
1963 return;
1964
1965 if (frag_offset) {
1966 struct skb_shared_info *info = skb_shinfo(skb);
1967
1968 /* We read part of the last frag, must recvmsg() rest of skb. */
1969 if (frag == &info->frags[info->nr_frags - 1])
1970 return;
1971
1972 /* Else, we must at least read the remainder in this frag. */
1973 partial_frag_remainder = skb_frag_size(frag) - frag_offset;
1974 zc->recv_skip_hint -= partial_frag_remainder;
1975 ++frag;
1976 }
1977
1978 /* partial_frag_remainder: If part way through a frag, must read rest.
1979 * mappable_offset: Bytes till next mappable frag, *not* counting bytes
1980 * in partial_frag_remainder.
1981 */
1982 mappable_offset = find_next_mappable_frag(frag, zc->recv_skip_hint);
1983 zc->recv_skip_hint = mappable_offset + partial_frag_remainder;
1984}
1985
1986static int tcp_recvmsg_locked(struct sock *sk, struct msghdr *msg, size_t len,
1987 int flags, struct scm_timestamping_internal *tss,
1988 int *cmsg_flags);
1989static int receive_fallback_to_copy(struct sock *sk,
1990 struct tcp_zerocopy_receive *zc, int inq,
1991 struct scm_timestamping_internal *tss)
1992{
1993 unsigned long copy_address = (unsigned long)zc->copybuf_address;
1994 struct msghdr msg = {};
1995 struct iovec iov;
1996 int err;
1997
1998 zc->length = 0;
1999 zc->recv_skip_hint = 0;
2000
2001 if (copy_address != zc->copybuf_address)
2002 return -EINVAL;
2003
2004 err = import_single_range(ITER_DEST, (void __user *)copy_address,
2005 inq, &iov, &msg.msg_iter);
2006 if (err)
2007 return err;
2008
2009 err = tcp_recvmsg_locked(sk, &msg, inq, MSG_DONTWAIT,
2010 tss, &zc->msg_flags);
2011 if (err < 0)
2012 return err;
2013
2014 zc->copybuf_len = err;
2015 if (likely(zc->copybuf_len)) {
2016 struct sk_buff *skb;
2017 u32 offset;
2018
2019 skb = tcp_recv_skb(sk, tcp_sk(sk)->copied_seq, &offset);
2020 if (skb)
2021 tcp_zerocopy_set_hint_for_skb(sk, zc, skb, offset);
2022 }
2023 return 0;
2024}
2025
2026static int tcp_copy_straggler_data(struct tcp_zerocopy_receive *zc,
2027 struct sk_buff *skb, u32 copylen,
2028 u32 *offset, u32 *seq)
2029{
2030 unsigned long copy_address = (unsigned long)zc->copybuf_address;
2031 struct msghdr msg = {};
2032 struct iovec iov;
2033 int err;
2034
2035 if (copy_address != zc->copybuf_address)
2036 return -EINVAL;
2037
2038 err = import_single_range(ITER_DEST, (void __user *)copy_address,
2039 copylen, &iov, &msg.msg_iter);
2040 if (err)
2041 return err;
2042 err = skb_copy_datagram_msg(skb, *offset, &msg, copylen);
2043 if (err)
2044 return err;
2045 zc->recv_skip_hint -= copylen;
2046 *offset += copylen;
2047 *seq += copylen;
2048 return (__s32)copylen;
2049}
2050
2051static int tcp_zc_handle_leftover(struct tcp_zerocopy_receive *zc,
2052 struct sock *sk,
2053 struct sk_buff *skb,
2054 u32 *seq,
2055 s32 copybuf_len,
2056 struct scm_timestamping_internal *tss)
2057{
2058 u32 offset, copylen = min_t(u32, copybuf_len, zc->recv_skip_hint);
2059
2060 if (!copylen)
2061 return 0;
2062 /* skb is null if inq < PAGE_SIZE. */
2063 if (skb) {
2064 offset = *seq - TCP_SKB_CB(skb)->seq;
2065 } else {
2066 skb = tcp_recv_skb(sk, *seq, &offset);
2067 if (TCP_SKB_CB(skb)->has_rxtstamp) {
2068 tcp_update_recv_tstamps(skb, tss);
2069 zc->msg_flags |= TCP_CMSG_TS;
2070 }
2071 }
2072
2073 zc->copybuf_len = tcp_copy_straggler_data(zc, skb, copylen, &offset,
2074 seq);
2075 return zc->copybuf_len < 0 ? 0 : copylen;
2076}
2077
2078static int tcp_zerocopy_vm_insert_batch_error(struct vm_area_struct *vma,
2079 struct page **pending_pages,
2080 unsigned long pages_remaining,
2081 unsigned long *address,
2082 u32 *length,
2083 u32 *seq,
2084 struct tcp_zerocopy_receive *zc,
2085 u32 total_bytes_to_map,
2086 int err)
2087{
2088 /* At least one page did not map. Try zapping if we skipped earlier. */
2089 if (err == -EBUSY &&
2090 zc->flags & TCP_RECEIVE_ZEROCOPY_FLAG_TLB_CLEAN_HINT) {
2091 u32 maybe_zap_len;
2092
2093 maybe_zap_len = total_bytes_to_map - /* All bytes to map */
2094 *length + /* Mapped or pending */
2095 (pages_remaining * PAGE_SIZE); /* Failed map. */
2096 zap_page_range(vma, *address, maybe_zap_len);
2097 err = 0;
2098 }
2099
2100 if (!err) {
2101 unsigned long leftover_pages = pages_remaining;
2102 int bytes_mapped;
2103
2104 /* We called zap_page_range, try to reinsert. */
2105 err = vm_insert_pages(vma, *address,
2106 pending_pages,
2107 &pages_remaining);
2108 bytes_mapped = PAGE_SIZE * (leftover_pages - pages_remaining);
2109 *seq += bytes_mapped;
2110 *address += bytes_mapped;
2111 }
2112 if (err) {
2113 /* Either we were unable to zap, OR we zapped, retried an
2114 * insert, and still had an issue. Either ways, pages_remaining
2115 * is the number of pages we were unable to map, and we unroll
2116 * some state we speculatively touched before.
2117 */
2118 const int bytes_not_mapped = PAGE_SIZE * pages_remaining;
2119
2120 *length -= bytes_not_mapped;
2121 zc->recv_skip_hint += bytes_not_mapped;
2122 }
2123 return err;
2124}
2125
2126static int tcp_zerocopy_vm_insert_batch(struct vm_area_struct *vma,
2127 struct page **pages,
2128 unsigned int pages_to_map,
2129 unsigned long *address,
2130 u32 *length,
2131 u32 *seq,
2132 struct tcp_zerocopy_receive *zc,
2133 u32 total_bytes_to_map)
2134{
2135 unsigned long pages_remaining = pages_to_map;
2136 unsigned int pages_mapped;
2137 unsigned int bytes_mapped;
2138 int err;
2139
2140 err = vm_insert_pages(vma, *address, pages, &pages_remaining);
2141 pages_mapped = pages_to_map - (unsigned int)pages_remaining;
2142 bytes_mapped = PAGE_SIZE * pages_mapped;
2143 /* Even if vm_insert_pages fails, it may have partially succeeded in
2144 * mapping (some but not all of the pages).
2145 */
2146 *seq += bytes_mapped;
2147 *address += bytes_mapped;
2148
2149 if (likely(!err))
2150 return 0;
2151
2152 /* Error: maybe zap and retry + rollback state for failed inserts. */
2153 return tcp_zerocopy_vm_insert_batch_error(vma, pages + pages_mapped,
2154 pages_remaining, address, length, seq, zc, total_bytes_to_map,
2155 err);
2156}
2157
2158#define TCP_VALID_ZC_MSG_FLAGS (TCP_CMSG_TS)
2159static void tcp_zc_finalize_rx_tstamp(struct sock *sk,
2160 struct tcp_zerocopy_receive *zc,
2161 struct scm_timestamping_internal *tss)
2162{
2163 unsigned long msg_control_addr;
2164 struct msghdr cmsg_dummy;
2165
2166 msg_control_addr = (unsigned long)zc->msg_control;
2167 cmsg_dummy.msg_control = (void *)msg_control_addr;
2168 cmsg_dummy.msg_controllen =
2169 (__kernel_size_t)zc->msg_controllen;
2170 cmsg_dummy.msg_flags = in_compat_syscall()
2171 ? MSG_CMSG_COMPAT : 0;
2172 cmsg_dummy.msg_control_is_user = true;
2173 zc->msg_flags = 0;
2174 if (zc->msg_control == msg_control_addr &&
2175 zc->msg_controllen == cmsg_dummy.msg_controllen) {
2176 tcp_recv_timestamp(&cmsg_dummy, sk, tss);
2177 zc->msg_control = (__u64)
2178 ((uintptr_t)cmsg_dummy.msg_control);
2179 zc->msg_controllen =
2180 (__u64)cmsg_dummy.msg_controllen;
2181 zc->msg_flags = (__u32)cmsg_dummy.msg_flags;
2182 }
2183}
2184
2185#define TCP_ZEROCOPY_PAGE_BATCH_SIZE 32
2186static int tcp_zerocopy_receive(struct sock *sk,
2187 struct tcp_zerocopy_receive *zc,
2188 struct scm_timestamping_internal *tss)
2189{
2190 u32 length = 0, offset, vma_len, avail_len, copylen = 0;
2191 unsigned long address = (unsigned long)zc->address;
2192 struct page *pages[TCP_ZEROCOPY_PAGE_BATCH_SIZE];
2193 s32 copybuf_len = zc->copybuf_len;
2194 struct tcp_sock *tp = tcp_sk(sk);
2195 const skb_frag_t *frags = NULL;
2196 unsigned int pages_to_map = 0;
2197 struct vm_area_struct *vma;
2198 struct sk_buff *skb = NULL;
2199 u32 seq = tp->copied_seq;
2200 u32 total_bytes_to_map;
2201 int inq = tcp_inq(sk);
2202 int ret;
2203
2204 zc->copybuf_len = 0;
2205 zc->msg_flags = 0;
2206
2207 if (address & (PAGE_SIZE - 1) || address != zc->address)
2208 return -EINVAL;
2209
2210 if (sk->sk_state == TCP_LISTEN)
2211 return -ENOTCONN;
2212
2213 sock_rps_record_flow(sk);
2214
2215 if (inq && inq <= copybuf_len)
2216 return receive_fallback_to_copy(sk, zc, inq, tss);
2217
2218 if (inq < PAGE_SIZE) {
2219 zc->length = 0;
2220 zc->recv_skip_hint = inq;
2221 if (!inq && sock_flag(sk, SOCK_DONE))
2222 return -EIO;
2223 return 0;
2224 }
2225
2226 mmap_read_lock(current->mm);
2227
2228 vma = vma_lookup(current->mm, address);
2229 if (!vma || vma->vm_ops != &tcp_vm_ops) {
2230 mmap_read_unlock(current->mm);
2231 return -EINVAL;
2232 }
2233 vma_len = min_t(unsigned long, zc->length, vma->vm_end - address);
2234 avail_len = min_t(u32, vma_len, inq);
2235 total_bytes_to_map = avail_len & ~(PAGE_SIZE - 1);
2236 if (total_bytes_to_map) {
2237 if (!(zc->flags & TCP_RECEIVE_ZEROCOPY_FLAG_TLB_CLEAN_HINT))
2238 zap_page_range(vma, address, total_bytes_to_map);
2239 zc->length = total_bytes_to_map;
2240 zc->recv_skip_hint = 0;
2241 } else {
2242 zc->length = avail_len;
2243 zc->recv_skip_hint = avail_len;
2244 }
2245 ret = 0;
2246 while (length + PAGE_SIZE <= zc->length) {
2247 int mappable_offset;
2248 struct page *page;
2249
2250 if (zc->recv_skip_hint < PAGE_SIZE) {
2251 u32 offset_frag;
2252
2253 if (skb) {
2254 if (zc->recv_skip_hint > 0)
2255 break;
2256 skb = skb->next;
2257 offset = seq - TCP_SKB_CB(skb)->seq;
2258 } else {
2259 skb = tcp_recv_skb(sk, seq, &offset);
2260 }
2261
2262 if (TCP_SKB_CB(skb)->has_rxtstamp) {
2263 tcp_update_recv_tstamps(skb, tss);
2264 zc->msg_flags |= TCP_CMSG_TS;
2265 }
2266 zc->recv_skip_hint = skb->len - offset;
2267 frags = skb_advance_to_frag(skb, offset, &offset_frag);
2268 if (!frags || offset_frag)
2269 break;
2270 }
2271
2272 mappable_offset = find_next_mappable_frag(frags,
2273 zc->recv_skip_hint);
2274 if (mappable_offset) {
2275 zc->recv_skip_hint = mappable_offset;
2276 break;
2277 }
2278 page = skb_frag_page(frags);
2279 prefetchw(page);
2280 pages[pages_to_map++] = page;
2281 length += PAGE_SIZE;
2282 zc->recv_skip_hint -= PAGE_SIZE;
2283 frags++;
2284 if (pages_to_map == TCP_ZEROCOPY_PAGE_BATCH_SIZE ||
2285 zc->recv_skip_hint < PAGE_SIZE) {
2286 /* Either full batch, or we're about to go to next skb
2287 * (and we cannot unroll failed ops across skbs).
2288 */
2289 ret = tcp_zerocopy_vm_insert_batch(vma, pages,
2290 pages_to_map,
2291 &address, &length,
2292 &seq, zc,
2293 total_bytes_to_map);
2294 if (ret)
2295 goto out;
2296 pages_to_map = 0;
2297 }
2298 }
2299 if (pages_to_map) {
2300 ret = tcp_zerocopy_vm_insert_batch(vma, pages, pages_to_map,
2301 &address, &length, &seq,
2302 zc, total_bytes_to_map);
2303 }
2304out:
2305 mmap_read_unlock(current->mm);
2306 /* Try to copy straggler data. */
2307 if (!ret)
2308 copylen = tcp_zc_handle_leftover(zc, sk, skb, &seq, copybuf_len, tss);
2309
2310 if (length + copylen) {
2311 WRITE_ONCE(tp->copied_seq, seq);
2312 tcp_rcv_space_adjust(sk);
2313
2314 /* Clean up data we have read: This will do ACK frames. */
2315 tcp_recv_skb(sk, seq, &offset);
2316 tcp_cleanup_rbuf(sk, length + copylen);
2317 ret = 0;
2318 if (length == zc->length)
2319 zc->recv_skip_hint = 0;
2320 } else {
2321 if (!zc->recv_skip_hint && sock_flag(sk, SOCK_DONE))
2322 ret = -EIO;
2323 }
2324 zc->length = length;
2325 return ret;
2326}
2327#endif
2328
2329/* Similar to __sock_recv_timestamp, but does not require an skb */
2330void tcp_recv_timestamp(struct msghdr *msg, const struct sock *sk,
2331 struct scm_timestamping_internal *tss)
2332{
2333 int new_tstamp = sock_flag(sk, SOCK_TSTAMP_NEW);
2334 bool has_timestamping = false;
2335
2336 if (tss->ts[0].tv_sec || tss->ts[0].tv_nsec) {
2337 if (sock_flag(sk, SOCK_RCVTSTAMP)) {
2338 if (sock_flag(sk, SOCK_RCVTSTAMPNS)) {
2339 if (new_tstamp) {
2340 struct __kernel_timespec kts = {
2341 .tv_sec = tss->ts[0].tv_sec,
2342 .tv_nsec = tss->ts[0].tv_nsec,
2343 };
2344 put_cmsg(msg, SOL_SOCKET, SO_TIMESTAMPNS_NEW,
2345 sizeof(kts), &kts);
2346 } else {
2347 struct __kernel_old_timespec ts_old = {
2348 .tv_sec = tss->ts[0].tv_sec,
2349 .tv_nsec = tss->ts[0].tv_nsec,
2350 };
2351 put_cmsg(msg, SOL_SOCKET, SO_TIMESTAMPNS_OLD,
2352 sizeof(ts_old), &ts_old);
2353 }
2354 } else {
2355 if (new_tstamp) {
2356 struct __kernel_sock_timeval stv = {
2357 .tv_sec = tss->ts[0].tv_sec,
2358 .tv_usec = tss->ts[0].tv_nsec / 1000,
2359 };
2360 put_cmsg(msg, SOL_SOCKET, SO_TIMESTAMP_NEW,
2361 sizeof(stv), &stv);
2362 } else {
2363 struct __kernel_old_timeval tv = {
2364 .tv_sec = tss->ts[0].tv_sec,
2365 .tv_usec = tss->ts[0].tv_nsec / 1000,
2366 };
2367 put_cmsg(msg, SOL_SOCKET, SO_TIMESTAMP_OLD,
2368 sizeof(tv), &tv);
2369 }
2370 }
2371 }
2372
2373 if (sk->sk_tsflags & SOF_TIMESTAMPING_SOFTWARE)
2374 has_timestamping = true;
2375 else
2376 tss->ts[0] = (struct timespec64) {0};
2377 }
2378
2379 if (tss->ts[2].tv_sec || tss->ts[2].tv_nsec) {
2380 if (sk->sk_tsflags & SOF_TIMESTAMPING_RAW_HARDWARE)
2381 has_timestamping = true;
2382 else
2383 tss->ts[2] = (struct timespec64) {0};
2384 }
2385
2386 if (has_timestamping) {
2387 tss->ts[1] = (struct timespec64) {0};
2388 if (sock_flag(sk, SOCK_TSTAMP_NEW))
2389 put_cmsg_scm_timestamping64(msg, tss);
2390 else
2391 put_cmsg_scm_timestamping(msg, tss);
2392 }
2393}
2394
2395static int tcp_inq_hint(struct sock *sk)
2396{
2397 const struct tcp_sock *tp = tcp_sk(sk);
2398 u32 copied_seq = READ_ONCE(tp->copied_seq);
2399 u32 rcv_nxt = READ_ONCE(tp->rcv_nxt);
2400 int inq;
2401
2402 inq = rcv_nxt - copied_seq;
2403 if (unlikely(inq < 0 || copied_seq != READ_ONCE(tp->copied_seq))) {
2404 lock_sock(sk);
2405 inq = tp->rcv_nxt - tp->copied_seq;
2406 release_sock(sk);
2407 }
2408 /* After receiving a FIN, tell the user-space to continue reading
2409 * by returning a non-zero inq.
2410 */
2411 if (inq == 0 && sock_flag(sk, SOCK_DONE))
2412 inq = 1;
2413 return inq;
2414}
2415
2416/*
2417 * This routine copies from a sock struct into the user buffer.
2418 *
2419 * Technical note: in 2.3 we work on _locked_ socket, so that
2420 * tricks with *seq access order and skb->users are not required.
2421 * Probably, code can be easily improved even more.
2422 */
2423
2424static int tcp_recvmsg_locked(struct sock *sk, struct msghdr *msg, size_t len,
2425 int flags, struct scm_timestamping_internal *tss,
2426 int *cmsg_flags)
2427{
2428 struct tcp_sock *tp = tcp_sk(sk);
2429 int copied = 0;
2430 u32 peek_seq;
2431 u32 *seq;
2432 unsigned long used;
2433 int err;
2434 int target; /* Read at least this many bytes */
2435 long timeo;
2436 struct sk_buff *skb, *last;
2437 u32 urg_hole = 0;
2438
2439 err = -ENOTCONN;
2440 if (sk->sk_state == TCP_LISTEN)
2441 goto out;
2442
2443 if (tp->recvmsg_inq) {
2444 *cmsg_flags = TCP_CMSG_INQ;
2445 msg->msg_get_inq = 1;
2446 }
2447 timeo = sock_rcvtimeo(sk, flags & MSG_DONTWAIT);
2448
2449 /* Urgent data needs to be handled specially. */
2450 if (flags & MSG_OOB)
2451 goto recv_urg;
2452
2453 if (unlikely(tp->repair)) {
2454 err = -EPERM;
2455 if (!(flags & MSG_PEEK))
2456 goto out;
2457
2458 if (tp->repair_queue == TCP_SEND_QUEUE)
2459 goto recv_sndq;
2460
2461 err = -EINVAL;
2462 if (tp->repair_queue == TCP_NO_QUEUE)
2463 goto out;
2464
2465 /* 'common' recv queue MSG_PEEK-ing */
2466 }
2467
2468 seq = &tp->copied_seq;
2469 if (flags & MSG_PEEK) {
2470 peek_seq = tp->copied_seq;
2471 seq = &peek_seq;
2472 }
2473
2474 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
2475
2476 do {
2477 u32 offset;
2478
2479 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
2480 if (unlikely(tp->urg_data) && tp->urg_seq == *seq) {
2481 if (copied)
2482 break;
2483 if (signal_pending(current)) {
2484 copied = timeo ? sock_intr_errno(timeo) : -EAGAIN;
2485 break;
2486 }
2487 }
2488
2489 /* Next get a buffer. */
2490
2491 last = skb_peek_tail(&sk->sk_receive_queue);
2492 skb_queue_walk(&sk->sk_receive_queue, skb) {
2493 last = skb;
2494 /* Now that we have two receive queues this
2495 * shouldn't happen.
2496 */
2497 if (WARN(before(*seq, TCP_SKB_CB(skb)->seq),
2498 "TCP recvmsg seq # bug: copied %X, seq %X, rcvnxt %X, fl %X\n",
2499 *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt,
2500 flags))
2501 break;
2502
2503 offset = *seq - TCP_SKB_CB(skb)->seq;
2504 if (unlikely(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
2505 pr_err_once("%s: found a SYN, please report !\n", __func__);
2506 offset--;
2507 }
2508 if (offset < skb->len)
2509 goto found_ok_skb;
2510 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
2511 goto found_fin_ok;
2512 WARN(!(flags & MSG_PEEK),
2513 "TCP recvmsg seq # bug 2: copied %X, seq %X, rcvnxt %X, fl %X\n",
2514 *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt, flags);
2515 }
2516
2517 /* Well, if we have backlog, try to process it now yet. */
2518
2519 if (copied >= target && !READ_ONCE(sk->sk_backlog.tail))
2520 break;
2521
2522 if (copied) {
2523 if (!timeo ||
2524 sk->sk_err ||
2525 sk->sk_state == TCP_CLOSE ||
2526 (sk->sk_shutdown & RCV_SHUTDOWN) ||
2527 signal_pending(current))
2528 break;
2529 } else {
2530 if (sock_flag(sk, SOCK_DONE))
2531 break;
2532
2533 if (sk->sk_err) {
2534 copied = sock_error(sk);
2535 break;
2536 }
2537
2538 if (sk->sk_shutdown & RCV_SHUTDOWN)
2539 break;
2540
2541 if (sk->sk_state == TCP_CLOSE) {
2542 /* This occurs when user tries to read
2543 * from never connected socket.
2544 */
2545 copied = -ENOTCONN;
2546 break;
2547 }
2548
2549 if (!timeo) {
2550 copied = -EAGAIN;
2551 break;
2552 }
2553
2554 if (signal_pending(current)) {
2555 copied = sock_intr_errno(timeo);
2556 break;
2557 }
2558 }
2559
2560 if (copied >= target) {
2561 /* Do not sleep, just process backlog. */
2562 __sk_flush_backlog(sk);
2563 } else {
2564 tcp_cleanup_rbuf(sk, copied);
2565 sk_wait_data(sk, &timeo, last);
2566 }
2567
2568 if ((flags & MSG_PEEK) &&
2569 (peek_seq - copied - urg_hole != tp->copied_seq)) {
2570 net_dbg_ratelimited("TCP(%s:%d): Application bug, race in MSG_PEEK\n",
2571 current->comm,
2572 task_pid_nr(current));
2573 peek_seq = tp->copied_seq;
2574 }
2575 continue;
2576
2577found_ok_skb:
2578 /* Ok so how much can we use? */
2579 used = skb->len - offset;
2580 if (len < used)
2581 used = len;
2582
2583 /* Do we have urgent data here? */
2584 if (unlikely(tp->urg_data)) {
2585 u32 urg_offset = tp->urg_seq - *seq;
2586 if (urg_offset < used) {
2587 if (!urg_offset) {
2588 if (!sock_flag(sk, SOCK_URGINLINE)) {
2589 WRITE_ONCE(*seq, *seq + 1);
2590 urg_hole++;
2591 offset++;
2592 used--;
2593 if (!used)
2594 goto skip_copy;
2595 }
2596 } else
2597 used = urg_offset;
2598 }
2599 }
2600
2601 if (!(flags & MSG_TRUNC)) {
2602 err = skb_copy_datagram_msg(skb, offset, msg, used);
2603 if (err) {
2604 /* Exception. Bailout! */
2605 if (!copied)
2606 copied = -EFAULT;
2607 break;
2608 }
2609 }
2610
2611 WRITE_ONCE(*seq, *seq + used);
2612 copied += used;
2613 len -= used;
2614
2615 tcp_rcv_space_adjust(sk);
2616
2617skip_copy:
2618 if (unlikely(tp->urg_data) && after(tp->copied_seq, tp->urg_seq)) {
2619 WRITE_ONCE(tp->urg_data, 0);
2620 tcp_fast_path_check(sk);
2621 }
2622
2623 if (TCP_SKB_CB(skb)->has_rxtstamp) {
2624 tcp_update_recv_tstamps(skb, tss);
2625 *cmsg_flags |= TCP_CMSG_TS;
2626 }
2627
2628 if (used + offset < skb->len)
2629 continue;
2630
2631 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
2632 goto found_fin_ok;
2633 if (!(flags & MSG_PEEK))
2634 tcp_eat_recv_skb(sk, skb);
2635 continue;
2636
2637found_fin_ok:
2638 /* Process the FIN. */
2639 WRITE_ONCE(*seq, *seq + 1);
2640 if (!(flags & MSG_PEEK))
2641 tcp_eat_recv_skb(sk, skb);
2642 break;
2643 } while (len > 0);
2644
2645 /* According to UNIX98, msg_name/msg_namelen are ignored
2646 * on connected socket. I was just happy when found this 8) --ANK
2647 */
2648
2649 /* Clean up data we have read: This will do ACK frames. */
2650 tcp_cleanup_rbuf(sk, copied);
2651 return copied;
2652
2653out:
2654 return err;
2655
2656recv_urg:
2657 err = tcp_recv_urg(sk, msg, len, flags);
2658 goto out;
2659
2660recv_sndq:
2661 err = tcp_peek_sndq(sk, msg, len);
2662 goto out;
2663}
2664
2665int tcp_recvmsg(struct sock *sk, struct msghdr *msg, size_t len, int flags,
2666 int *addr_len)
2667{
2668 int cmsg_flags = 0, ret;
2669 struct scm_timestamping_internal tss;
2670
2671 if (unlikely(flags & MSG_ERRQUEUE))
2672 return inet_recv_error(sk, msg, len, addr_len);
2673
2674 if (sk_can_busy_loop(sk) &&
2675 skb_queue_empty_lockless(&sk->sk_receive_queue) &&
2676 sk->sk_state == TCP_ESTABLISHED)
2677 sk_busy_loop(sk, flags & MSG_DONTWAIT);
2678
2679 lock_sock(sk);
2680 ret = tcp_recvmsg_locked(sk, msg, len, flags, &tss, &cmsg_flags);
2681 release_sock(sk);
2682
2683 if ((cmsg_flags || msg->msg_get_inq) && ret >= 0) {
2684 if (cmsg_flags & TCP_CMSG_TS)
2685 tcp_recv_timestamp(msg, sk, &tss);
2686 if (msg->msg_get_inq) {
2687 msg->msg_inq = tcp_inq_hint(sk);
2688 if (cmsg_flags & TCP_CMSG_INQ)
2689 put_cmsg(msg, SOL_TCP, TCP_CM_INQ,
2690 sizeof(msg->msg_inq), &msg->msg_inq);
2691 }
2692 }
2693 return ret;
2694}
2695EXPORT_SYMBOL(tcp_recvmsg);
2696
2697void tcp_set_state(struct sock *sk, int state)
2698{
2699 int oldstate = sk->sk_state;
2700
2701 /* We defined a new enum for TCP states that are exported in BPF
2702 * so as not force the internal TCP states to be frozen. The
2703 * following checks will detect if an internal state value ever
2704 * differs from the BPF value. If this ever happens, then we will
2705 * need to remap the internal value to the BPF value before calling
2706 * tcp_call_bpf_2arg.
2707 */
2708 BUILD_BUG_ON((int)BPF_TCP_ESTABLISHED != (int)TCP_ESTABLISHED);
2709 BUILD_BUG_ON((int)BPF_TCP_SYN_SENT != (int)TCP_SYN_SENT);
2710 BUILD_BUG_ON((int)BPF_TCP_SYN_RECV != (int)TCP_SYN_RECV);
2711 BUILD_BUG_ON((int)BPF_TCP_FIN_WAIT1 != (int)TCP_FIN_WAIT1);
2712 BUILD_BUG_ON((int)BPF_TCP_FIN_WAIT2 != (int)TCP_FIN_WAIT2);
2713 BUILD_BUG_ON((int)BPF_TCP_TIME_WAIT != (int)TCP_TIME_WAIT);
2714 BUILD_BUG_ON((int)BPF_TCP_CLOSE != (int)TCP_CLOSE);
2715 BUILD_BUG_ON((int)BPF_TCP_CLOSE_WAIT != (int)TCP_CLOSE_WAIT);
2716 BUILD_BUG_ON((int)BPF_TCP_LAST_ACK != (int)TCP_LAST_ACK);
2717 BUILD_BUG_ON((int)BPF_TCP_LISTEN != (int)TCP_LISTEN);
2718 BUILD_BUG_ON((int)BPF_TCP_CLOSING != (int)TCP_CLOSING);
2719 BUILD_BUG_ON((int)BPF_TCP_NEW_SYN_RECV != (int)TCP_NEW_SYN_RECV);
2720 BUILD_BUG_ON((int)BPF_TCP_MAX_STATES != (int)TCP_MAX_STATES);
2721
2722 /* bpf uapi header bpf.h defines an anonymous enum with values
2723 * BPF_TCP_* used by bpf programs. Currently gcc built vmlinux
2724 * is able to emit this enum in DWARF due to the above BUILD_BUG_ON.
2725 * But clang built vmlinux does not have this enum in DWARF
2726 * since clang removes the above code before generating IR/debuginfo.
2727 * Let us explicitly emit the type debuginfo to ensure the
2728 * above-mentioned anonymous enum in the vmlinux DWARF and hence BTF
2729 * regardless of which compiler is used.
2730 */
2731 BTF_TYPE_EMIT_ENUM(BPF_TCP_ESTABLISHED);
2732
2733 if (BPF_SOCK_OPS_TEST_FLAG(tcp_sk(sk), BPF_SOCK_OPS_STATE_CB_FLAG))
2734 tcp_call_bpf_2arg(sk, BPF_SOCK_OPS_STATE_CB, oldstate, state);
2735
2736 switch (state) {
2737 case TCP_ESTABLISHED:
2738 if (oldstate != TCP_ESTABLISHED)
2739 TCP_INC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
2740 break;
2741
2742 case TCP_CLOSE:
2743 if (oldstate == TCP_CLOSE_WAIT || oldstate == TCP_ESTABLISHED)
2744 TCP_INC_STATS(sock_net(sk), TCP_MIB_ESTABRESETS);
2745
2746 sk->sk_prot->unhash(sk);
2747 if (inet_csk(sk)->icsk_bind_hash &&
2748 !(sk->sk_userlocks & SOCK_BINDPORT_LOCK))
2749 inet_put_port(sk);
2750 fallthrough;
2751 default:
2752 if (oldstate == TCP_ESTABLISHED)
2753 TCP_DEC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
2754 }
2755
2756 /* Change state AFTER socket is unhashed to avoid closed
2757 * socket sitting in hash tables.
2758 */
2759 inet_sk_state_store(sk, state);
2760}
2761EXPORT_SYMBOL_GPL(tcp_set_state);
2762
2763/*
2764 * State processing on a close. This implements the state shift for
2765 * sending our FIN frame. Note that we only send a FIN for some
2766 * states. A shutdown() may have already sent the FIN, or we may be
2767 * closed.
2768 */
2769
2770static const unsigned char new_state[16] = {
2771 /* current state: new state: action: */
2772 [0 /* (Invalid) */] = TCP_CLOSE,
2773 [TCP_ESTABLISHED] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2774 [TCP_SYN_SENT] = TCP_CLOSE,
2775 [TCP_SYN_RECV] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2776 [TCP_FIN_WAIT1] = TCP_FIN_WAIT1,
2777 [TCP_FIN_WAIT2] = TCP_FIN_WAIT2,
2778 [TCP_TIME_WAIT] = TCP_CLOSE,
2779 [TCP_CLOSE] = TCP_CLOSE,
2780 [TCP_CLOSE_WAIT] = TCP_LAST_ACK | TCP_ACTION_FIN,
2781 [TCP_LAST_ACK] = TCP_LAST_ACK,
2782 [TCP_LISTEN] = TCP_CLOSE,
2783 [TCP_CLOSING] = TCP_CLOSING,
2784 [TCP_NEW_SYN_RECV] = TCP_CLOSE, /* should not happen ! */
2785};
2786
2787static int tcp_close_state(struct sock *sk)
2788{
2789 int next = (int)new_state[sk->sk_state];
2790 int ns = next & TCP_STATE_MASK;
2791
2792 tcp_set_state(sk, ns);
2793
2794 return next & TCP_ACTION_FIN;
2795}
2796
2797/*
2798 * Shutdown the sending side of a connection. Much like close except
2799 * that we don't receive shut down or sock_set_flag(sk, SOCK_DEAD).
2800 */
2801
2802void tcp_shutdown(struct sock *sk, int how)
2803{
2804 /* We need to grab some memory, and put together a FIN,
2805 * and then put it into the queue to be sent.
2806 * Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
2807 */
2808 if (!(how & SEND_SHUTDOWN))
2809 return;
2810
2811 /* If we've already sent a FIN, or it's a closed state, skip this. */
2812 if ((1 << sk->sk_state) &
2813 (TCPF_ESTABLISHED | TCPF_SYN_SENT |
2814 TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) {
2815 /* Clear out any half completed packets. FIN if needed. */
2816 if (tcp_close_state(sk))
2817 tcp_send_fin(sk);
2818 }
2819}
2820EXPORT_SYMBOL(tcp_shutdown);
2821
2822int tcp_orphan_count_sum(void)
2823{
2824 int i, total = 0;
2825
2826 for_each_possible_cpu(i)
2827 total += per_cpu(tcp_orphan_count, i);
2828
2829 return max(total, 0);
2830}
2831
2832static int tcp_orphan_cache;
2833static struct timer_list tcp_orphan_timer;
2834#define TCP_ORPHAN_TIMER_PERIOD msecs_to_jiffies(100)
2835
2836static void tcp_orphan_update(struct timer_list *unused)
2837{
2838 WRITE_ONCE(tcp_orphan_cache, tcp_orphan_count_sum());
2839 mod_timer(&tcp_orphan_timer, jiffies + TCP_ORPHAN_TIMER_PERIOD);
2840}
2841
2842static bool tcp_too_many_orphans(int shift)
2843{
2844 return READ_ONCE(tcp_orphan_cache) << shift >
2845 READ_ONCE(sysctl_tcp_max_orphans);
2846}
2847
2848bool tcp_check_oom(struct sock *sk, int shift)
2849{
2850 bool too_many_orphans, out_of_socket_memory;
2851
2852 too_many_orphans = tcp_too_many_orphans(shift);
2853 out_of_socket_memory = tcp_out_of_memory(sk);
2854
2855 if (too_many_orphans)
2856 net_info_ratelimited("too many orphaned sockets\n");
2857 if (out_of_socket_memory)
2858 net_info_ratelimited("out of memory -- consider tuning tcp_mem\n");
2859 return too_many_orphans || out_of_socket_memory;
2860}
2861
2862void __tcp_close(struct sock *sk, long timeout)
2863{
2864 struct sk_buff *skb;
2865 int data_was_unread = 0;
2866 int state;
2867
2868 sk->sk_shutdown = SHUTDOWN_MASK;
2869
2870 if (sk->sk_state == TCP_LISTEN) {
2871 tcp_set_state(sk, TCP_CLOSE);
2872
2873 /* Special case. */
2874 inet_csk_listen_stop(sk);
2875
2876 goto adjudge_to_death;
2877 }
2878
2879 /* We need to flush the recv. buffs. We do this only on the
2880 * descriptor close, not protocol-sourced closes, because the
2881 * reader process may not have drained the data yet!
2882 */
2883 while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) {
2884 u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq;
2885
2886 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
2887 len--;
2888 data_was_unread += len;
2889 __kfree_skb(skb);
2890 }
2891
2892 /* If socket has been already reset (e.g. in tcp_reset()) - kill it. */
2893 if (sk->sk_state == TCP_CLOSE)
2894 goto adjudge_to_death;
2895
2896 /* As outlined in RFC 2525, section 2.17, we send a RST here because
2897 * data was lost. To witness the awful effects of the old behavior of
2898 * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk
2899 * GET in an FTP client, suspend the process, wait for the client to
2900 * advertise a zero window, then kill -9 the FTP client, wheee...
2901 * Note: timeout is always zero in such a case.
2902 */
2903 if (unlikely(tcp_sk(sk)->repair)) {
2904 sk->sk_prot->disconnect(sk, 0);
2905 } else if (data_was_unread) {
2906 /* Unread data was tossed, zap the connection. */
2907 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONCLOSE);
2908 tcp_set_state(sk, TCP_CLOSE);
2909 tcp_send_active_reset(sk, sk->sk_allocation);
2910 } else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) {
2911 /* Check zero linger _after_ checking for unread data. */
2912 sk->sk_prot->disconnect(sk, 0);
2913 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
2914 } else if (tcp_close_state(sk)) {
2915 /* We FIN if the application ate all the data before
2916 * zapping the connection.
2917 */
2918
2919 /* RED-PEN. Formally speaking, we have broken TCP state
2920 * machine. State transitions:
2921 *
2922 * TCP_ESTABLISHED -> TCP_FIN_WAIT1
2923 * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible)
2924 * TCP_CLOSE_WAIT -> TCP_LAST_ACK
2925 *
2926 * are legal only when FIN has been sent (i.e. in window),
2927 * rather than queued out of window. Purists blame.
2928 *
2929 * F.e. "RFC state" is ESTABLISHED,
2930 * if Linux state is FIN-WAIT-1, but FIN is still not sent.
2931 *
2932 * The visible declinations are that sometimes
2933 * we enter time-wait state, when it is not required really
2934 * (harmless), do not send active resets, when they are
2935 * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
2936 * they look as CLOSING or LAST_ACK for Linux)
2937 * Probably, I missed some more holelets.
2938 * --ANK
2939 * XXX (TFO) - To start off we don't support SYN+ACK+FIN
2940 * in a single packet! (May consider it later but will
2941 * probably need API support or TCP_CORK SYN-ACK until
2942 * data is written and socket is closed.)
2943 */
2944 tcp_send_fin(sk);
2945 }
2946
2947 sk_stream_wait_close(sk, timeout);
2948
2949adjudge_to_death:
2950 state = sk->sk_state;
2951 sock_hold(sk);
2952 sock_orphan(sk);
2953
2954 local_bh_disable();
2955 bh_lock_sock(sk);
2956 /* remove backlog if any, without releasing ownership. */
2957 __release_sock(sk);
2958
2959 this_cpu_inc(tcp_orphan_count);
2960
2961 /* Have we already been destroyed by a softirq or backlog? */
2962 if (state != TCP_CLOSE && sk->sk_state == TCP_CLOSE)
2963 goto out;
2964
2965 /* This is a (useful) BSD violating of the RFC. There is a
2966 * problem with TCP as specified in that the other end could
2967 * keep a socket open forever with no application left this end.
2968 * We use a 1 minute timeout (about the same as BSD) then kill
2969 * our end. If they send after that then tough - BUT: long enough
2970 * that we won't make the old 4*rto = almost no time - whoops
2971 * reset mistake.
2972 *
2973 * Nope, it was not mistake. It is really desired behaviour
2974 * f.e. on http servers, when such sockets are useless, but
2975 * consume significant resources. Let's do it with special
2976 * linger2 option. --ANK
2977 */
2978
2979 if (sk->sk_state == TCP_FIN_WAIT2) {
2980 struct tcp_sock *tp = tcp_sk(sk);
2981 if (tp->linger2 < 0) {
2982 tcp_set_state(sk, TCP_CLOSE);
2983 tcp_send_active_reset(sk, GFP_ATOMIC);
2984 __NET_INC_STATS(sock_net(sk),
2985 LINUX_MIB_TCPABORTONLINGER);
2986 } else {
2987 const int tmo = tcp_fin_time(sk);
2988
2989 if (tmo > TCP_TIMEWAIT_LEN) {
2990 inet_csk_reset_keepalive_timer(sk,
2991 tmo - TCP_TIMEWAIT_LEN);
2992 } else {
2993 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
2994 goto out;
2995 }
2996 }
2997 }
2998 if (sk->sk_state != TCP_CLOSE) {
2999 if (tcp_check_oom(sk, 0)) {
3000 tcp_set_state(sk, TCP_CLOSE);
3001 tcp_send_active_reset(sk, GFP_ATOMIC);
3002 __NET_INC_STATS(sock_net(sk),
3003 LINUX_MIB_TCPABORTONMEMORY);
3004 } else if (!check_net(sock_net(sk))) {
3005 /* Not possible to send reset; just close */
3006 tcp_set_state(sk, TCP_CLOSE);
3007 }
3008 }
3009
3010 if (sk->sk_state == TCP_CLOSE) {
3011 struct request_sock *req;
3012
3013 req = rcu_dereference_protected(tcp_sk(sk)->fastopen_rsk,
3014 lockdep_sock_is_held(sk));
3015 /* We could get here with a non-NULL req if the socket is
3016 * aborted (e.g., closed with unread data) before 3WHS
3017 * finishes.
3018 */
3019 if (req)
3020 reqsk_fastopen_remove(sk, req, false);
3021 inet_csk_destroy_sock(sk);
3022 }
3023 /* Otherwise, socket is reprieved until protocol close. */
3024
3025out:
3026 bh_unlock_sock(sk);
3027 local_bh_enable();
3028}
3029
3030void tcp_close(struct sock *sk, long timeout)
3031{
3032 lock_sock(sk);
3033 __tcp_close(sk, timeout);
3034 release_sock(sk);
3035 sock_put(sk);
3036}
3037EXPORT_SYMBOL(tcp_close);
3038
3039/* These states need RST on ABORT according to RFC793 */
3040
3041static inline bool tcp_need_reset(int state)
3042{
3043 return (1 << state) &
3044 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 |
3045 TCPF_FIN_WAIT2 | TCPF_SYN_RECV);
3046}
3047
3048static void tcp_rtx_queue_purge(struct sock *sk)
3049{
3050 struct rb_node *p = rb_first(&sk->tcp_rtx_queue);
3051
3052 tcp_sk(sk)->highest_sack = NULL;
3053 while (p) {
3054 struct sk_buff *skb = rb_to_skb(p);
3055
3056 p = rb_next(p);
3057 /* Since we are deleting whole queue, no need to
3058 * list_del(&skb->tcp_tsorted_anchor)
3059 */
3060 tcp_rtx_queue_unlink(skb, sk);
3061 tcp_wmem_free_skb(sk, skb);
3062 }
3063}
3064
3065void tcp_write_queue_purge(struct sock *sk)
3066{
3067 struct sk_buff *skb;
3068
3069 tcp_chrono_stop(sk, TCP_CHRONO_BUSY);
3070 while ((skb = __skb_dequeue(&sk->sk_write_queue)) != NULL) {
3071 tcp_skb_tsorted_anchor_cleanup(skb);
3072 tcp_wmem_free_skb(sk, skb);
3073 }
3074 tcp_rtx_queue_purge(sk);
3075 INIT_LIST_HEAD(&tcp_sk(sk)->tsorted_sent_queue);
3076 tcp_clear_all_retrans_hints(tcp_sk(sk));
3077 tcp_sk(sk)->packets_out = 0;
3078 inet_csk(sk)->icsk_backoff = 0;
3079}
3080
3081int tcp_disconnect(struct sock *sk, int flags)
3082{
3083 struct inet_sock *inet = inet_sk(sk);
3084 struct inet_connection_sock *icsk = inet_csk(sk);
3085 struct tcp_sock *tp = tcp_sk(sk);
3086 int old_state = sk->sk_state;
3087 u32 seq;
3088
3089 if (old_state != TCP_CLOSE)
3090 tcp_set_state(sk, TCP_CLOSE);
3091
3092 /* ABORT function of RFC793 */
3093 if (old_state == TCP_LISTEN) {
3094 inet_csk_listen_stop(sk);
3095 } else if (unlikely(tp->repair)) {
3096 sk->sk_err = ECONNABORTED;
3097 } else if (tcp_need_reset(old_state) ||
3098 (tp->snd_nxt != tp->write_seq &&
3099 (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK))) {
3100 /* The last check adjusts for discrepancy of Linux wrt. RFC
3101 * states
3102 */
3103 tcp_send_active_reset(sk, gfp_any());
3104 sk->sk_err = ECONNRESET;
3105 } else if (old_state == TCP_SYN_SENT)
3106 sk->sk_err = ECONNRESET;
3107
3108 tcp_clear_xmit_timers(sk);
3109 __skb_queue_purge(&sk->sk_receive_queue);
3110 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt);
3111 WRITE_ONCE(tp->urg_data, 0);
3112 tcp_write_queue_purge(sk);
3113 tcp_fastopen_active_disable_ofo_check(sk);
3114 skb_rbtree_purge(&tp->out_of_order_queue);
3115
3116 inet->inet_dport = 0;
3117
3118 inet_bhash2_reset_saddr(sk);
3119
3120 sk->sk_shutdown = 0;
3121 sock_reset_flag(sk, SOCK_DONE);
3122 tp->srtt_us = 0;
3123 tp->mdev_us = jiffies_to_usecs(TCP_TIMEOUT_INIT);
3124 tp->rcv_rtt_last_tsecr = 0;
3125
3126 seq = tp->write_seq + tp->max_window + 2;
3127 if (!seq)
3128 seq = 1;
3129 WRITE_ONCE(tp->write_seq, seq);
3130
3131 icsk->icsk_backoff = 0;
3132 icsk->icsk_probes_out = 0;
3133 icsk->icsk_probes_tstamp = 0;
3134 icsk->icsk_rto = TCP_TIMEOUT_INIT;
3135 icsk->icsk_rto_min = TCP_RTO_MIN;
3136 icsk->icsk_delack_max = TCP_DELACK_MAX;
3137 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
3138 tcp_snd_cwnd_set(tp, TCP_INIT_CWND);
3139 tp->snd_cwnd_cnt = 0;
3140 tp->is_cwnd_limited = 0;
3141 tp->max_packets_out = 0;
3142 tp->window_clamp = 0;
3143 tp->delivered = 0;
3144 tp->delivered_ce = 0;
3145 if (icsk->icsk_ca_ops->release)
3146 icsk->icsk_ca_ops->release(sk);
3147 memset(icsk->icsk_ca_priv, 0, sizeof(icsk->icsk_ca_priv));
3148 icsk->icsk_ca_initialized = 0;
3149 tcp_set_ca_state(sk, TCP_CA_Open);
3150 tp->is_sack_reneg = 0;
3151 tcp_clear_retrans(tp);
3152 tp->total_retrans = 0;
3153 inet_csk_delack_init(sk);
3154 /* Initialize rcv_mss to TCP_MIN_MSS to avoid division by 0
3155 * issue in __tcp_select_window()
3156 */
3157 icsk->icsk_ack.rcv_mss = TCP_MIN_MSS;
3158 memset(&tp->rx_opt, 0, sizeof(tp->rx_opt));
3159 __sk_dst_reset(sk);
3160 dst_release(xchg((__force struct dst_entry **)&sk->sk_rx_dst, NULL));
3161 tcp_saved_syn_free(tp);
3162 tp->compressed_ack = 0;
3163 tp->segs_in = 0;
3164 tp->segs_out = 0;
3165 tp->bytes_sent = 0;
3166 tp->bytes_acked = 0;
3167 tp->bytes_received = 0;
3168 tp->bytes_retrans = 0;
3169 tp->data_segs_in = 0;
3170 tp->data_segs_out = 0;
3171 tp->duplicate_sack[0].start_seq = 0;
3172 tp->duplicate_sack[0].end_seq = 0;
3173 tp->dsack_dups = 0;
3174 tp->reord_seen = 0;
3175 tp->retrans_out = 0;
3176 tp->sacked_out = 0;
3177 tp->tlp_high_seq = 0;
3178 tp->last_oow_ack_time = 0;
3179 tp->plb_rehash = 0;
3180 /* There's a bubble in the pipe until at least the first ACK. */
3181 tp->app_limited = ~0U;
3182 tp->rate_app_limited = 1;
3183 tp->rack.mstamp = 0;
3184 tp->rack.advanced = 0;
3185 tp->rack.reo_wnd_steps = 1;
3186 tp->rack.last_delivered = 0;
3187 tp->rack.reo_wnd_persist = 0;
3188 tp->rack.dsack_seen = 0;
3189 tp->syn_data_acked = 0;
3190 tp->rx_opt.saw_tstamp = 0;
3191 tp->rx_opt.dsack = 0;
3192 tp->rx_opt.num_sacks = 0;
3193 tp->rcv_ooopack = 0;
3194
3195
3196 /* Clean up fastopen related fields */
3197 tcp_free_fastopen_req(tp);
3198 inet->defer_connect = 0;
3199 tp->fastopen_client_fail = 0;
3200
3201 WARN_ON(inet->inet_num && !icsk->icsk_bind_hash);
3202
3203 if (sk->sk_frag.page) {
3204 put_page(sk->sk_frag.page);
3205 sk->sk_frag.page = NULL;
3206 sk->sk_frag.offset = 0;
3207 }
3208 sk_error_report(sk);
3209 return 0;
3210}
3211EXPORT_SYMBOL(tcp_disconnect);
3212
3213static inline bool tcp_can_repair_sock(const struct sock *sk)
3214{
3215 return sockopt_ns_capable(sock_net(sk)->user_ns, CAP_NET_ADMIN) &&
3216 (sk->sk_state != TCP_LISTEN);
3217}
3218
3219static int tcp_repair_set_window(struct tcp_sock *tp, sockptr_t optbuf, int len)
3220{
3221 struct tcp_repair_window opt;
3222
3223 if (!tp->repair)
3224 return -EPERM;
3225
3226 if (len != sizeof(opt))
3227 return -EINVAL;
3228
3229 if (copy_from_sockptr(&opt, optbuf, sizeof(opt)))
3230 return -EFAULT;
3231
3232 if (opt.max_window < opt.snd_wnd)
3233 return -EINVAL;
3234
3235 if (after(opt.snd_wl1, tp->rcv_nxt + opt.rcv_wnd))
3236 return -EINVAL;
3237
3238 if (after(opt.rcv_wup, tp->rcv_nxt))
3239 return -EINVAL;
3240
3241 tp->snd_wl1 = opt.snd_wl1;
3242 tp->snd_wnd = opt.snd_wnd;
3243 tp->max_window = opt.max_window;
3244
3245 tp->rcv_wnd = opt.rcv_wnd;
3246 tp->rcv_wup = opt.rcv_wup;
3247
3248 return 0;
3249}
3250
3251static int tcp_repair_options_est(struct sock *sk, sockptr_t optbuf,
3252 unsigned int len)
3253{
3254 struct tcp_sock *tp = tcp_sk(sk);
3255 struct tcp_repair_opt opt;
3256 size_t offset = 0;
3257
3258 while (len >= sizeof(opt)) {
3259 if (copy_from_sockptr_offset(&opt, optbuf, offset, sizeof(opt)))
3260 return -EFAULT;
3261
3262 offset += sizeof(opt);
3263 len -= sizeof(opt);
3264
3265 switch (opt.opt_code) {
3266 case TCPOPT_MSS:
3267 tp->rx_opt.mss_clamp = opt.opt_val;
3268 tcp_mtup_init(sk);
3269 break;
3270 case TCPOPT_WINDOW:
3271 {
3272 u16 snd_wscale = opt.opt_val & 0xFFFF;
3273 u16 rcv_wscale = opt.opt_val >> 16;
3274
3275 if (snd_wscale > TCP_MAX_WSCALE || rcv_wscale > TCP_MAX_WSCALE)
3276 return -EFBIG;
3277
3278 tp->rx_opt.snd_wscale = snd_wscale;
3279 tp->rx_opt.rcv_wscale = rcv_wscale;
3280 tp->rx_opt.wscale_ok = 1;
3281 }
3282 break;
3283 case TCPOPT_SACK_PERM:
3284 if (opt.opt_val != 0)
3285 return -EINVAL;
3286
3287 tp->rx_opt.sack_ok |= TCP_SACK_SEEN;
3288 break;
3289 case TCPOPT_TIMESTAMP:
3290 if (opt.opt_val != 0)
3291 return -EINVAL;
3292
3293 tp->rx_opt.tstamp_ok = 1;
3294 break;
3295 }
3296 }
3297
3298 return 0;
3299}
3300
3301DEFINE_STATIC_KEY_FALSE(tcp_tx_delay_enabled);
3302EXPORT_SYMBOL(tcp_tx_delay_enabled);
3303
3304static void tcp_enable_tx_delay(void)
3305{
3306 if (!static_branch_unlikely(&tcp_tx_delay_enabled)) {
3307 static int __tcp_tx_delay_enabled = 0;
3308
3309 if (cmpxchg(&__tcp_tx_delay_enabled, 0, 1) == 0) {
3310 static_branch_enable(&tcp_tx_delay_enabled);
3311 pr_info("TCP_TX_DELAY enabled\n");
3312 }
3313 }
3314}
3315
3316/* When set indicates to always queue non-full frames. Later the user clears
3317 * this option and we transmit any pending partial frames in the queue. This is
3318 * meant to be used alongside sendfile() to get properly filled frames when the
3319 * user (for example) must write out headers with a write() call first and then
3320 * use sendfile to send out the data parts.
3321 *
3322 * TCP_CORK can be set together with TCP_NODELAY and it is stronger than
3323 * TCP_NODELAY.
3324 */
3325void __tcp_sock_set_cork(struct sock *sk, bool on)
3326{
3327 struct tcp_sock *tp = tcp_sk(sk);
3328
3329 if (on) {
3330 tp->nonagle |= TCP_NAGLE_CORK;
3331 } else {
3332 tp->nonagle &= ~TCP_NAGLE_CORK;
3333 if (tp->nonagle & TCP_NAGLE_OFF)
3334 tp->nonagle |= TCP_NAGLE_PUSH;
3335 tcp_push_pending_frames(sk);
3336 }
3337}
3338
3339void tcp_sock_set_cork(struct sock *sk, bool on)
3340{
3341 lock_sock(sk);
3342 __tcp_sock_set_cork(sk, on);
3343 release_sock(sk);
3344}
3345EXPORT_SYMBOL(tcp_sock_set_cork);
3346
3347/* TCP_NODELAY is weaker than TCP_CORK, so that this option on corked socket is
3348 * remembered, but it is not activated until cork is cleared.
3349 *
3350 * However, when TCP_NODELAY is set we make an explicit push, which overrides
3351 * even TCP_CORK for currently queued segments.
3352 */
3353void __tcp_sock_set_nodelay(struct sock *sk, bool on)
3354{
3355 if (on) {
3356 tcp_sk(sk)->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH;
3357 tcp_push_pending_frames(sk);
3358 } else {
3359 tcp_sk(sk)->nonagle &= ~TCP_NAGLE_OFF;
3360 }
3361}
3362
3363void tcp_sock_set_nodelay(struct sock *sk)
3364{
3365 lock_sock(sk);
3366 __tcp_sock_set_nodelay(sk, true);
3367 release_sock(sk);
3368}
3369EXPORT_SYMBOL(tcp_sock_set_nodelay);
3370
3371static void __tcp_sock_set_quickack(struct sock *sk, int val)
3372{
3373 if (!val) {
3374 inet_csk_enter_pingpong_mode(sk);
3375 return;
3376 }
3377
3378 inet_csk_exit_pingpong_mode(sk);
3379 if ((1 << sk->sk_state) & (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) &&
3380 inet_csk_ack_scheduled(sk)) {
3381 inet_csk(sk)->icsk_ack.pending |= ICSK_ACK_PUSHED;
3382 tcp_cleanup_rbuf(sk, 1);
3383 if (!(val & 1))
3384 inet_csk_enter_pingpong_mode(sk);
3385 }
3386}
3387
3388void tcp_sock_set_quickack(struct sock *sk, int val)
3389{
3390 lock_sock(sk);
3391 __tcp_sock_set_quickack(sk, val);
3392 release_sock(sk);
3393}
3394EXPORT_SYMBOL(tcp_sock_set_quickack);
3395
3396int tcp_sock_set_syncnt(struct sock *sk, int val)
3397{
3398 if (val < 1 || val > MAX_TCP_SYNCNT)
3399 return -EINVAL;
3400
3401 lock_sock(sk);
3402 inet_csk(sk)->icsk_syn_retries = val;
3403 release_sock(sk);
3404 return 0;
3405}
3406EXPORT_SYMBOL(tcp_sock_set_syncnt);
3407
3408void tcp_sock_set_user_timeout(struct sock *sk, u32 val)
3409{
3410 lock_sock(sk);
3411 inet_csk(sk)->icsk_user_timeout = val;
3412 release_sock(sk);
3413}
3414EXPORT_SYMBOL(tcp_sock_set_user_timeout);
3415
3416int tcp_sock_set_keepidle_locked(struct sock *sk, int val)
3417{
3418 struct tcp_sock *tp = tcp_sk(sk);
3419
3420 if (val < 1 || val > MAX_TCP_KEEPIDLE)
3421 return -EINVAL;
3422
3423 tp->keepalive_time = val * HZ;
3424 if (sock_flag(sk, SOCK_KEEPOPEN) &&
3425 !((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN))) {
3426 u32 elapsed = keepalive_time_elapsed(tp);
3427
3428 if (tp->keepalive_time > elapsed)
3429 elapsed = tp->keepalive_time - elapsed;
3430 else
3431 elapsed = 0;
3432 inet_csk_reset_keepalive_timer(sk, elapsed);
3433 }
3434
3435 return 0;
3436}
3437
3438int tcp_sock_set_keepidle(struct sock *sk, int val)
3439{
3440 int err;
3441
3442 lock_sock(sk);
3443 err = tcp_sock_set_keepidle_locked(sk, val);
3444 release_sock(sk);
3445 return err;
3446}
3447EXPORT_SYMBOL(tcp_sock_set_keepidle);
3448
3449int tcp_sock_set_keepintvl(struct sock *sk, int val)
3450{
3451 if (val < 1 || val > MAX_TCP_KEEPINTVL)
3452 return -EINVAL;
3453
3454 lock_sock(sk);
3455 tcp_sk(sk)->keepalive_intvl = val * HZ;
3456 release_sock(sk);
3457 return 0;
3458}
3459EXPORT_SYMBOL(tcp_sock_set_keepintvl);
3460
3461int tcp_sock_set_keepcnt(struct sock *sk, int val)
3462{
3463 if (val < 1 || val > MAX_TCP_KEEPCNT)
3464 return -EINVAL;
3465
3466 lock_sock(sk);
3467 tcp_sk(sk)->keepalive_probes = val;
3468 release_sock(sk);
3469 return 0;
3470}
3471EXPORT_SYMBOL(tcp_sock_set_keepcnt);
3472
3473int tcp_set_window_clamp(struct sock *sk, int val)
3474{
3475 struct tcp_sock *tp = tcp_sk(sk);
3476
3477 if (!val) {
3478 if (sk->sk_state != TCP_CLOSE)
3479 return -EINVAL;
3480 tp->window_clamp = 0;
3481 } else {
3482 tp->window_clamp = val < SOCK_MIN_RCVBUF / 2 ?
3483 SOCK_MIN_RCVBUF / 2 : val;
3484 tp->rcv_ssthresh = min(tp->rcv_wnd, tp->window_clamp);
3485 }
3486 return 0;
3487}
3488
3489/*
3490 * Socket option code for TCP.
3491 */
3492int do_tcp_setsockopt(struct sock *sk, int level, int optname,
3493 sockptr_t optval, unsigned int optlen)
3494{
3495 struct tcp_sock *tp = tcp_sk(sk);
3496 struct inet_connection_sock *icsk = inet_csk(sk);
3497 struct net *net = sock_net(sk);
3498 int val;
3499 int err = 0;
3500
3501 /* These are data/string values, all the others are ints */
3502 switch (optname) {
3503 case TCP_CONGESTION: {
3504 char name[TCP_CA_NAME_MAX];
3505
3506 if (optlen < 1)
3507 return -EINVAL;
3508
3509 val = strncpy_from_sockptr(name, optval,
3510 min_t(long, TCP_CA_NAME_MAX-1, optlen));
3511 if (val < 0)
3512 return -EFAULT;
3513 name[val] = 0;
3514
3515 sockopt_lock_sock(sk);
3516 err = tcp_set_congestion_control(sk, name, !has_current_bpf_ctx(),
3517 sockopt_ns_capable(sock_net(sk)->user_ns,
3518 CAP_NET_ADMIN));
3519 sockopt_release_sock(sk);
3520 return err;
3521 }
3522 case TCP_ULP: {
3523 char name[TCP_ULP_NAME_MAX];
3524
3525 if (optlen < 1)
3526 return -EINVAL;
3527
3528 val = strncpy_from_sockptr(name, optval,
3529 min_t(long, TCP_ULP_NAME_MAX - 1,
3530 optlen));
3531 if (val < 0)
3532 return -EFAULT;
3533 name[val] = 0;
3534
3535 sockopt_lock_sock(sk);
3536 err = tcp_set_ulp(sk, name);
3537 sockopt_release_sock(sk);
3538 return err;
3539 }
3540 case TCP_FASTOPEN_KEY: {
3541 __u8 key[TCP_FASTOPEN_KEY_BUF_LENGTH];
3542 __u8 *backup_key = NULL;
3543
3544 /* Allow a backup key as well to facilitate key rotation
3545 * First key is the active one.
3546 */
3547 if (optlen != TCP_FASTOPEN_KEY_LENGTH &&
3548 optlen != TCP_FASTOPEN_KEY_BUF_LENGTH)
3549 return -EINVAL;
3550
3551 if (copy_from_sockptr(key, optval, optlen))
3552 return -EFAULT;
3553
3554 if (optlen == TCP_FASTOPEN_KEY_BUF_LENGTH)
3555 backup_key = key + TCP_FASTOPEN_KEY_LENGTH;
3556
3557 return tcp_fastopen_reset_cipher(net, sk, key, backup_key);
3558 }
3559 default:
3560 /* fallthru */
3561 break;
3562 }
3563
3564 if (optlen < sizeof(int))
3565 return -EINVAL;
3566
3567 if (copy_from_sockptr(&val, optval, sizeof(val)))
3568 return -EFAULT;
3569
3570 sockopt_lock_sock(sk);
3571
3572 switch (optname) {
3573 case TCP_MAXSEG:
3574 /* Values greater than interface MTU won't take effect. However
3575 * at the point when this call is done we typically don't yet
3576 * know which interface is going to be used
3577 */
3578 if (val && (val < TCP_MIN_MSS || val > MAX_TCP_WINDOW)) {
3579 err = -EINVAL;
3580 break;
3581 }
3582 tp->rx_opt.user_mss = val;
3583 break;
3584
3585 case TCP_NODELAY:
3586 __tcp_sock_set_nodelay(sk, val);
3587 break;
3588
3589 case TCP_THIN_LINEAR_TIMEOUTS:
3590 if (val < 0 || val > 1)
3591 err = -EINVAL;
3592 else
3593 tp->thin_lto = val;
3594 break;
3595
3596 case TCP_THIN_DUPACK:
3597 if (val < 0 || val > 1)
3598 err = -EINVAL;
3599 break;
3600
3601 case TCP_REPAIR:
3602 if (!tcp_can_repair_sock(sk))
3603 err = -EPERM;
3604 else if (val == TCP_REPAIR_ON) {
3605 tp->repair = 1;
3606 sk->sk_reuse = SK_FORCE_REUSE;
3607 tp->repair_queue = TCP_NO_QUEUE;
3608 } else if (val == TCP_REPAIR_OFF) {
3609 tp->repair = 0;
3610 sk->sk_reuse = SK_NO_REUSE;
3611 tcp_send_window_probe(sk);
3612 } else if (val == TCP_REPAIR_OFF_NO_WP) {
3613 tp->repair = 0;
3614 sk->sk_reuse = SK_NO_REUSE;
3615 } else
3616 err = -EINVAL;
3617
3618 break;
3619
3620 case TCP_REPAIR_QUEUE:
3621 if (!tp->repair)
3622 err = -EPERM;
3623 else if ((unsigned int)val < TCP_QUEUES_NR)
3624 tp->repair_queue = val;
3625 else
3626 err = -EINVAL;
3627 break;
3628
3629 case TCP_QUEUE_SEQ:
3630 if (sk->sk_state != TCP_CLOSE) {
3631 err = -EPERM;
3632 } else if (tp->repair_queue == TCP_SEND_QUEUE) {
3633 if (!tcp_rtx_queue_empty(sk))
3634 err = -EPERM;
3635 else
3636 WRITE_ONCE(tp->write_seq, val);
3637 } else if (tp->repair_queue == TCP_RECV_QUEUE) {
3638 if (tp->rcv_nxt != tp->copied_seq) {
3639 err = -EPERM;
3640 } else {
3641 WRITE_ONCE(tp->rcv_nxt, val);
3642 WRITE_ONCE(tp->copied_seq, val);
3643 }
3644 } else {
3645 err = -EINVAL;
3646 }
3647 break;
3648
3649 case TCP_REPAIR_OPTIONS:
3650 if (!tp->repair)
3651 err = -EINVAL;
3652 else if (sk->sk_state == TCP_ESTABLISHED && !tp->bytes_sent)
3653 err = tcp_repair_options_est(sk, optval, optlen);
3654 else
3655 err = -EPERM;
3656 break;
3657
3658 case TCP_CORK:
3659 __tcp_sock_set_cork(sk, val);
3660 break;
3661
3662 case TCP_KEEPIDLE:
3663 err = tcp_sock_set_keepidle_locked(sk, val);
3664 break;
3665 case TCP_KEEPINTVL:
3666 if (val < 1 || val > MAX_TCP_KEEPINTVL)
3667 err = -EINVAL;
3668 else
3669 tp->keepalive_intvl = val * HZ;
3670 break;
3671 case TCP_KEEPCNT:
3672 if (val < 1 || val > MAX_TCP_KEEPCNT)
3673 err = -EINVAL;
3674 else
3675 tp->keepalive_probes = val;
3676 break;
3677 case TCP_SYNCNT:
3678 if (val < 1 || val > MAX_TCP_SYNCNT)
3679 err = -EINVAL;
3680 else
3681 icsk->icsk_syn_retries = val;
3682 break;
3683
3684 case TCP_SAVE_SYN:
3685 /* 0: disable, 1: enable, 2: start from ether_header */
3686 if (val < 0 || val > 2)
3687 err = -EINVAL;
3688 else
3689 tp->save_syn = val;
3690 break;
3691
3692 case TCP_LINGER2:
3693 if (val < 0)
3694 tp->linger2 = -1;
3695 else if (val > TCP_FIN_TIMEOUT_MAX / HZ)
3696 tp->linger2 = TCP_FIN_TIMEOUT_MAX;
3697 else
3698 tp->linger2 = val * HZ;
3699 break;
3700
3701 case TCP_DEFER_ACCEPT:
3702 /* Translate value in seconds to number of retransmits */
3703 icsk->icsk_accept_queue.rskq_defer_accept =
3704 secs_to_retrans(val, TCP_TIMEOUT_INIT / HZ,
3705 TCP_RTO_MAX / HZ);
3706 break;
3707
3708 case TCP_WINDOW_CLAMP:
3709 err = tcp_set_window_clamp(sk, val);
3710 break;
3711
3712 case TCP_QUICKACK:
3713 __tcp_sock_set_quickack(sk, val);
3714 break;
3715
3716#ifdef CONFIG_TCP_MD5SIG
3717 case TCP_MD5SIG:
3718 case TCP_MD5SIG_EXT:
3719 err = tp->af_specific->md5_parse(sk, optname, optval, optlen);
3720 break;
3721#endif
3722 case TCP_USER_TIMEOUT:
3723 /* Cap the max time in ms TCP will retry or probe the window
3724 * before giving up and aborting (ETIMEDOUT) a connection.
3725 */
3726 if (val < 0)
3727 err = -EINVAL;
3728 else
3729 icsk->icsk_user_timeout = val;
3730 break;
3731
3732 case TCP_FASTOPEN:
3733 if (val >= 0 && ((1 << sk->sk_state) & (TCPF_CLOSE |
3734 TCPF_LISTEN))) {
3735 tcp_fastopen_init_key_once(net);
3736
3737 fastopen_queue_tune(sk, val);
3738 } else {
3739 err = -EINVAL;
3740 }
3741 break;
3742 case TCP_FASTOPEN_CONNECT:
3743 if (val > 1 || val < 0) {
3744 err = -EINVAL;
3745 } else if (READ_ONCE(net->ipv4.sysctl_tcp_fastopen) &
3746 TFO_CLIENT_ENABLE) {
3747 if (sk->sk_state == TCP_CLOSE)
3748 tp->fastopen_connect = val;
3749 else
3750 err = -EINVAL;
3751 } else {
3752 err = -EOPNOTSUPP;
3753 }
3754 break;
3755 case TCP_FASTOPEN_NO_COOKIE:
3756 if (val > 1 || val < 0)
3757 err = -EINVAL;
3758 else if (!((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
3759 err = -EINVAL;
3760 else
3761 tp->fastopen_no_cookie = val;
3762 break;
3763 case TCP_TIMESTAMP:
3764 if (!tp->repair)
3765 err = -EPERM;
3766 else
3767 tp->tsoffset = val - tcp_time_stamp_raw();
3768 break;
3769 case TCP_REPAIR_WINDOW:
3770 err = tcp_repair_set_window(tp, optval, optlen);
3771 break;
3772 case TCP_NOTSENT_LOWAT:
3773 tp->notsent_lowat = val;
3774 sk->sk_write_space(sk);
3775 break;
3776 case TCP_INQ:
3777 if (val > 1 || val < 0)
3778 err = -EINVAL;
3779 else
3780 tp->recvmsg_inq = val;
3781 break;
3782 case TCP_TX_DELAY:
3783 if (val)
3784 tcp_enable_tx_delay();
3785 tp->tcp_tx_delay = val;
3786 break;
3787 default:
3788 err = -ENOPROTOOPT;
3789 break;
3790 }
3791
3792 sockopt_release_sock(sk);
3793 return err;
3794}
3795
3796int tcp_setsockopt(struct sock *sk, int level, int optname, sockptr_t optval,
3797 unsigned int optlen)
3798{
3799 const struct inet_connection_sock *icsk = inet_csk(sk);
3800
3801 if (level != SOL_TCP)
3802 /* Paired with WRITE_ONCE() in do_ipv6_setsockopt() and tcp_v6_connect() */
3803 return READ_ONCE(icsk->icsk_af_ops)->setsockopt(sk, level, optname,
3804 optval, optlen);
3805 return do_tcp_setsockopt(sk, level, optname, optval, optlen);
3806}
3807EXPORT_SYMBOL(tcp_setsockopt);
3808
3809static void tcp_get_info_chrono_stats(const struct tcp_sock *tp,
3810 struct tcp_info *info)
3811{
3812 u64 stats[__TCP_CHRONO_MAX], total = 0;
3813 enum tcp_chrono i;
3814
3815 for (i = TCP_CHRONO_BUSY; i < __TCP_CHRONO_MAX; ++i) {
3816 stats[i] = tp->chrono_stat[i - 1];
3817 if (i == tp->chrono_type)
3818 stats[i] += tcp_jiffies32 - tp->chrono_start;
3819 stats[i] *= USEC_PER_SEC / HZ;
3820 total += stats[i];
3821 }
3822
3823 info->tcpi_busy_time = total;
3824 info->tcpi_rwnd_limited = stats[TCP_CHRONO_RWND_LIMITED];
3825 info->tcpi_sndbuf_limited = stats[TCP_CHRONO_SNDBUF_LIMITED];
3826}
3827
3828/* Return information about state of tcp endpoint in API format. */
3829void tcp_get_info(struct sock *sk, struct tcp_info *info)
3830{
3831 const struct tcp_sock *tp = tcp_sk(sk); /* iff sk_type == SOCK_STREAM */
3832 const struct inet_connection_sock *icsk = inet_csk(sk);
3833 unsigned long rate;
3834 u32 now;
3835 u64 rate64;
3836 bool slow;
3837
3838 memset(info, 0, sizeof(*info));
3839 if (sk->sk_type != SOCK_STREAM)
3840 return;
3841
3842 info->tcpi_state = inet_sk_state_load(sk);
3843
3844 /* Report meaningful fields for all TCP states, including listeners */
3845 rate = READ_ONCE(sk->sk_pacing_rate);
3846 rate64 = (rate != ~0UL) ? rate : ~0ULL;
3847 info->tcpi_pacing_rate = rate64;
3848
3849 rate = READ_ONCE(sk->sk_max_pacing_rate);
3850 rate64 = (rate != ~0UL) ? rate : ~0ULL;
3851 info->tcpi_max_pacing_rate = rate64;
3852
3853 info->tcpi_reordering = tp->reordering;
3854 info->tcpi_snd_cwnd = tcp_snd_cwnd(tp);
3855
3856 if (info->tcpi_state == TCP_LISTEN) {
3857 /* listeners aliased fields :
3858 * tcpi_unacked -> Number of children ready for accept()
3859 * tcpi_sacked -> max backlog
3860 */
3861 info->tcpi_unacked = READ_ONCE(sk->sk_ack_backlog);
3862 info->tcpi_sacked = READ_ONCE(sk->sk_max_ack_backlog);
3863 return;
3864 }
3865
3866 slow = lock_sock_fast(sk);
3867
3868 info->tcpi_ca_state = icsk->icsk_ca_state;
3869 info->tcpi_retransmits = icsk->icsk_retransmits;
3870 info->tcpi_probes = icsk->icsk_probes_out;
3871 info->tcpi_backoff = icsk->icsk_backoff;
3872
3873 if (tp->rx_opt.tstamp_ok)
3874 info->tcpi_options |= TCPI_OPT_TIMESTAMPS;
3875 if (tcp_is_sack(tp))
3876 info->tcpi_options |= TCPI_OPT_SACK;
3877 if (tp->rx_opt.wscale_ok) {
3878 info->tcpi_options |= TCPI_OPT_WSCALE;
3879 info->tcpi_snd_wscale = tp->rx_opt.snd_wscale;
3880 info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale;
3881 }
3882
3883 if (tp->ecn_flags & TCP_ECN_OK)
3884 info->tcpi_options |= TCPI_OPT_ECN;
3885 if (tp->ecn_flags & TCP_ECN_SEEN)
3886 info->tcpi_options |= TCPI_OPT_ECN_SEEN;
3887 if (tp->syn_data_acked)
3888 info->tcpi_options |= TCPI_OPT_SYN_DATA;
3889
3890 info->tcpi_rto = jiffies_to_usecs(icsk->icsk_rto);
3891 info->tcpi_ato = jiffies_to_usecs(icsk->icsk_ack.ato);
3892 info->tcpi_snd_mss = tp->mss_cache;
3893 info->tcpi_rcv_mss = icsk->icsk_ack.rcv_mss;
3894
3895 info->tcpi_unacked = tp->packets_out;
3896 info->tcpi_sacked = tp->sacked_out;
3897
3898 info->tcpi_lost = tp->lost_out;
3899 info->tcpi_retrans = tp->retrans_out;
3900
3901 now = tcp_jiffies32;
3902 info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime);
3903 info->tcpi_last_data_recv = jiffies_to_msecs(now - icsk->icsk_ack.lrcvtime);
3904 info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp);
3905
3906 info->tcpi_pmtu = icsk->icsk_pmtu_cookie;
3907 info->tcpi_rcv_ssthresh = tp->rcv_ssthresh;
3908 info->tcpi_rtt = tp->srtt_us >> 3;
3909 info->tcpi_rttvar = tp->mdev_us >> 2;
3910 info->tcpi_snd_ssthresh = tp->snd_ssthresh;
3911 info->tcpi_advmss = tp->advmss;
3912
3913 info->tcpi_rcv_rtt = tp->rcv_rtt_est.rtt_us >> 3;
3914 info->tcpi_rcv_space = tp->rcvq_space.space;
3915
3916 info->tcpi_total_retrans = tp->total_retrans;
3917
3918 info->tcpi_bytes_acked = tp->bytes_acked;
3919 info->tcpi_bytes_received = tp->bytes_received;
3920 info->tcpi_notsent_bytes = max_t(int, 0, tp->write_seq - tp->snd_nxt);
3921 tcp_get_info_chrono_stats(tp, info);
3922
3923 info->tcpi_segs_out = tp->segs_out;
3924
3925 /* segs_in and data_segs_in can be updated from tcp_segs_in() from BH */
3926 info->tcpi_segs_in = READ_ONCE(tp->segs_in);
3927 info->tcpi_data_segs_in = READ_ONCE(tp->data_segs_in);
3928
3929 info->tcpi_min_rtt = tcp_min_rtt(tp);
3930 info->tcpi_data_segs_out = tp->data_segs_out;
3931
3932 info->tcpi_delivery_rate_app_limited = tp->rate_app_limited ? 1 : 0;
3933 rate64 = tcp_compute_delivery_rate(tp);
3934 if (rate64)
3935 info->tcpi_delivery_rate = rate64;
3936 info->tcpi_delivered = tp->delivered;
3937 info->tcpi_delivered_ce = tp->delivered_ce;
3938 info->tcpi_bytes_sent = tp->bytes_sent;
3939 info->tcpi_bytes_retrans = tp->bytes_retrans;
3940 info->tcpi_dsack_dups = tp->dsack_dups;
3941 info->tcpi_reord_seen = tp->reord_seen;
3942 info->tcpi_rcv_ooopack = tp->rcv_ooopack;
3943 info->tcpi_snd_wnd = tp->snd_wnd;
3944 info->tcpi_rcv_wnd = tp->rcv_wnd;
3945 info->tcpi_rehash = tp->plb_rehash + tp->timeout_rehash;
3946 info->tcpi_fastopen_client_fail = tp->fastopen_client_fail;
3947 unlock_sock_fast(sk, slow);
3948}
3949EXPORT_SYMBOL_GPL(tcp_get_info);
3950
3951static size_t tcp_opt_stats_get_size(void)
3952{
3953 return
3954 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_BUSY */
3955 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_RWND_LIMITED */
3956 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_SNDBUF_LIMITED */
3957 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_DATA_SEGS_OUT */
3958 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_TOTAL_RETRANS */
3959 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_PACING_RATE */
3960 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_DELIVERY_RATE */
3961 nla_total_size(sizeof(u32)) + /* TCP_NLA_SND_CWND */
3962 nla_total_size(sizeof(u32)) + /* TCP_NLA_REORDERING */
3963 nla_total_size(sizeof(u32)) + /* TCP_NLA_MIN_RTT */
3964 nla_total_size(sizeof(u8)) + /* TCP_NLA_RECUR_RETRANS */
3965 nla_total_size(sizeof(u8)) + /* TCP_NLA_DELIVERY_RATE_APP_LMT */
3966 nla_total_size(sizeof(u32)) + /* TCP_NLA_SNDQ_SIZE */
3967 nla_total_size(sizeof(u8)) + /* TCP_NLA_CA_STATE */
3968 nla_total_size(sizeof(u32)) + /* TCP_NLA_SND_SSTHRESH */
3969 nla_total_size(sizeof(u32)) + /* TCP_NLA_DELIVERED */
3970 nla_total_size(sizeof(u32)) + /* TCP_NLA_DELIVERED_CE */
3971 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_BYTES_SENT */
3972 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_BYTES_RETRANS */
3973 nla_total_size(sizeof(u32)) + /* TCP_NLA_DSACK_DUPS */
3974 nla_total_size(sizeof(u32)) + /* TCP_NLA_REORD_SEEN */
3975 nla_total_size(sizeof(u32)) + /* TCP_NLA_SRTT */
3976 nla_total_size(sizeof(u16)) + /* TCP_NLA_TIMEOUT_REHASH */
3977 nla_total_size(sizeof(u32)) + /* TCP_NLA_BYTES_NOTSENT */
3978 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_EDT */
3979 nla_total_size(sizeof(u8)) + /* TCP_NLA_TTL */
3980 nla_total_size(sizeof(u32)) + /* TCP_NLA_REHASH */
3981 0;
3982}
3983
3984/* Returns TTL or hop limit of an incoming packet from skb. */
3985static u8 tcp_skb_ttl_or_hop_limit(const struct sk_buff *skb)
3986{
3987 if (skb->protocol == htons(ETH_P_IP))
3988 return ip_hdr(skb)->ttl;
3989 else if (skb->protocol == htons(ETH_P_IPV6))
3990 return ipv6_hdr(skb)->hop_limit;
3991 else
3992 return 0;
3993}
3994
3995struct sk_buff *tcp_get_timestamping_opt_stats(const struct sock *sk,
3996 const struct sk_buff *orig_skb,
3997 const struct sk_buff *ack_skb)
3998{
3999 const struct tcp_sock *tp = tcp_sk(sk);
4000 struct sk_buff *stats;
4001 struct tcp_info info;
4002 unsigned long rate;
4003 u64 rate64;
4004
4005 stats = alloc_skb(tcp_opt_stats_get_size(), GFP_ATOMIC);
4006 if (!stats)
4007 return NULL;
4008
4009 tcp_get_info_chrono_stats(tp, &info);
4010 nla_put_u64_64bit(stats, TCP_NLA_BUSY,
4011 info.tcpi_busy_time, TCP_NLA_PAD);
4012 nla_put_u64_64bit(stats, TCP_NLA_RWND_LIMITED,
4013 info.tcpi_rwnd_limited, TCP_NLA_PAD);
4014 nla_put_u64_64bit(stats, TCP_NLA_SNDBUF_LIMITED,
4015 info.tcpi_sndbuf_limited, TCP_NLA_PAD);
4016 nla_put_u64_64bit(stats, TCP_NLA_DATA_SEGS_OUT,
4017 tp->data_segs_out, TCP_NLA_PAD);
4018 nla_put_u64_64bit(stats, TCP_NLA_TOTAL_RETRANS,
4019 tp->total_retrans, TCP_NLA_PAD);
4020
4021 rate = READ_ONCE(sk->sk_pacing_rate);
4022 rate64 = (rate != ~0UL) ? rate : ~0ULL;
4023 nla_put_u64_64bit(stats, TCP_NLA_PACING_RATE, rate64, TCP_NLA_PAD);
4024
4025 rate64 = tcp_compute_delivery_rate(tp);
4026 nla_put_u64_64bit(stats, TCP_NLA_DELIVERY_RATE, rate64, TCP_NLA_PAD);
4027
4028 nla_put_u32(stats, TCP_NLA_SND_CWND, tcp_snd_cwnd(tp));
4029 nla_put_u32(stats, TCP_NLA_REORDERING, tp->reordering);
4030 nla_put_u32(stats, TCP_NLA_MIN_RTT, tcp_min_rtt(tp));
4031
4032 nla_put_u8(stats, TCP_NLA_RECUR_RETRANS, inet_csk(sk)->icsk_retransmits);
4033 nla_put_u8(stats, TCP_NLA_DELIVERY_RATE_APP_LMT, !!tp->rate_app_limited);
4034 nla_put_u32(stats, TCP_NLA_SND_SSTHRESH, tp->snd_ssthresh);
4035 nla_put_u32(stats, TCP_NLA_DELIVERED, tp->delivered);
4036 nla_put_u32(stats, TCP_NLA_DELIVERED_CE, tp->delivered_ce);
4037
4038 nla_put_u32(stats, TCP_NLA_SNDQ_SIZE, tp->write_seq - tp->snd_una);
4039 nla_put_u8(stats, TCP_NLA_CA_STATE, inet_csk(sk)->icsk_ca_state);
4040
4041 nla_put_u64_64bit(stats, TCP_NLA_BYTES_SENT, tp->bytes_sent,
4042 TCP_NLA_PAD);
4043 nla_put_u64_64bit(stats, TCP_NLA_BYTES_RETRANS, tp->bytes_retrans,
4044 TCP_NLA_PAD);
4045 nla_put_u32(stats, TCP_NLA_DSACK_DUPS, tp->dsack_dups);
4046 nla_put_u32(stats, TCP_NLA_REORD_SEEN, tp->reord_seen);
4047 nla_put_u32(stats, TCP_NLA_SRTT, tp->srtt_us >> 3);
4048 nla_put_u16(stats, TCP_NLA_TIMEOUT_REHASH, tp->timeout_rehash);
4049 nla_put_u32(stats, TCP_NLA_BYTES_NOTSENT,
4050 max_t(int, 0, tp->write_seq - tp->snd_nxt));
4051 nla_put_u64_64bit(stats, TCP_NLA_EDT, orig_skb->skb_mstamp_ns,
4052 TCP_NLA_PAD);
4053 if (ack_skb)
4054 nla_put_u8(stats, TCP_NLA_TTL,
4055 tcp_skb_ttl_or_hop_limit(ack_skb));
4056
4057 nla_put_u32(stats, TCP_NLA_REHASH, tp->plb_rehash + tp->timeout_rehash);
4058 return stats;
4059}
4060
4061int do_tcp_getsockopt(struct sock *sk, int level,
4062 int optname, sockptr_t optval, sockptr_t optlen)
4063{
4064 struct inet_connection_sock *icsk = inet_csk(sk);
4065 struct tcp_sock *tp = tcp_sk(sk);
4066 struct net *net = sock_net(sk);
4067 int val, len;
4068
4069 if (copy_from_sockptr(&len, optlen, sizeof(int)))
4070 return -EFAULT;
4071
4072 len = min_t(unsigned int, len, sizeof(int));
4073
4074 if (len < 0)
4075 return -EINVAL;
4076
4077 switch (optname) {
4078 case TCP_MAXSEG:
4079 val = tp->mss_cache;
4080 if (!val && ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
4081 val = tp->rx_opt.user_mss;
4082 if (tp->repair)
4083 val = tp->rx_opt.mss_clamp;
4084 break;
4085 case TCP_NODELAY:
4086 val = !!(tp->nonagle&TCP_NAGLE_OFF);
4087 break;
4088 case TCP_CORK:
4089 val = !!(tp->nonagle&TCP_NAGLE_CORK);
4090 break;
4091 case TCP_KEEPIDLE:
4092 val = keepalive_time_when(tp) / HZ;
4093 break;
4094 case TCP_KEEPINTVL:
4095 val = keepalive_intvl_when(tp) / HZ;
4096 break;
4097 case TCP_KEEPCNT:
4098 val = keepalive_probes(tp);
4099 break;
4100 case TCP_SYNCNT:
4101 val = icsk->icsk_syn_retries ? :
4102 READ_ONCE(net->ipv4.sysctl_tcp_syn_retries);
4103 break;
4104 case TCP_LINGER2:
4105 val = tp->linger2;
4106 if (val >= 0)
4107 val = (val ? : READ_ONCE(net->ipv4.sysctl_tcp_fin_timeout)) / HZ;
4108 break;
4109 case TCP_DEFER_ACCEPT:
4110 val = retrans_to_secs(icsk->icsk_accept_queue.rskq_defer_accept,
4111 TCP_TIMEOUT_INIT / HZ, TCP_RTO_MAX / HZ);
4112 break;
4113 case TCP_WINDOW_CLAMP:
4114 val = tp->window_clamp;
4115 break;
4116 case TCP_INFO: {
4117 struct tcp_info info;
4118
4119 if (copy_from_sockptr(&len, optlen, sizeof(int)))
4120 return -EFAULT;
4121
4122 tcp_get_info(sk, &info);
4123
4124 len = min_t(unsigned int, len, sizeof(info));
4125 if (copy_to_sockptr(optlen, &len, sizeof(int)))
4126 return -EFAULT;
4127 if (copy_to_sockptr(optval, &info, len))
4128 return -EFAULT;
4129 return 0;
4130 }
4131 case TCP_CC_INFO: {
4132 const struct tcp_congestion_ops *ca_ops;
4133 union tcp_cc_info info;
4134 size_t sz = 0;
4135 int attr;
4136
4137 if (copy_from_sockptr(&len, optlen, sizeof(int)))
4138 return -EFAULT;
4139
4140 ca_ops = icsk->icsk_ca_ops;
4141 if (ca_ops && ca_ops->get_info)
4142 sz = ca_ops->get_info(sk, ~0U, &attr, &info);
4143
4144 len = min_t(unsigned int, len, sz);
4145 if (copy_to_sockptr(optlen, &len, sizeof(int)))
4146 return -EFAULT;
4147 if (copy_to_sockptr(optval, &info, len))
4148 return -EFAULT;
4149 return 0;
4150 }
4151 case TCP_QUICKACK:
4152 val = !inet_csk_in_pingpong_mode(sk);
4153 break;
4154
4155 case TCP_CONGESTION:
4156 if (copy_from_sockptr(&len, optlen, sizeof(int)))
4157 return -EFAULT;
4158 len = min_t(unsigned int, len, TCP_CA_NAME_MAX);
4159 if (copy_to_sockptr(optlen, &len, sizeof(int)))
4160 return -EFAULT;
4161 if (copy_to_sockptr(optval, icsk->icsk_ca_ops->name, len))
4162 return -EFAULT;
4163 return 0;
4164
4165 case TCP_ULP:
4166 if (copy_from_sockptr(&len, optlen, sizeof(int)))
4167 return -EFAULT;
4168 len = min_t(unsigned int, len, TCP_ULP_NAME_MAX);
4169 if (!icsk->icsk_ulp_ops) {
4170 len = 0;
4171 if (copy_to_sockptr(optlen, &len, sizeof(int)))
4172 return -EFAULT;
4173 return 0;
4174 }
4175 if (copy_to_sockptr(optlen, &len, sizeof(int)))
4176 return -EFAULT;
4177 if (copy_to_sockptr(optval, icsk->icsk_ulp_ops->name, len))
4178 return -EFAULT;
4179 return 0;
4180
4181 case TCP_FASTOPEN_KEY: {
4182 u64 key[TCP_FASTOPEN_KEY_BUF_LENGTH / sizeof(u64)];
4183 unsigned int key_len;
4184
4185 if (copy_from_sockptr(&len, optlen, sizeof(int)))
4186 return -EFAULT;
4187
4188 key_len = tcp_fastopen_get_cipher(net, icsk, key) *
4189 TCP_FASTOPEN_KEY_LENGTH;
4190 len = min_t(unsigned int, len, key_len);
4191 if (copy_to_sockptr(optlen, &len, sizeof(int)))
4192 return -EFAULT;
4193 if (copy_to_sockptr(optval, key, len))
4194 return -EFAULT;
4195 return 0;
4196 }
4197 case TCP_THIN_LINEAR_TIMEOUTS:
4198 val = tp->thin_lto;
4199 break;
4200
4201 case TCP_THIN_DUPACK:
4202 val = 0;
4203 break;
4204
4205 case TCP_REPAIR:
4206 val = tp->repair;
4207 break;
4208
4209 case TCP_REPAIR_QUEUE:
4210 if (tp->repair)
4211 val = tp->repair_queue;
4212 else
4213 return -EINVAL;
4214 break;
4215
4216 case TCP_REPAIR_WINDOW: {
4217 struct tcp_repair_window opt;
4218
4219 if (copy_from_sockptr(&len, optlen, sizeof(int)))
4220 return -EFAULT;
4221
4222 if (len != sizeof(opt))
4223 return -EINVAL;
4224
4225 if (!tp->repair)
4226 return -EPERM;
4227
4228 opt.snd_wl1 = tp->snd_wl1;
4229 opt.snd_wnd = tp->snd_wnd;
4230 opt.max_window = tp->max_window;
4231 opt.rcv_wnd = tp->rcv_wnd;
4232 opt.rcv_wup = tp->rcv_wup;
4233
4234 if (copy_to_sockptr(optval, &opt, len))
4235 return -EFAULT;
4236 return 0;
4237 }
4238 case TCP_QUEUE_SEQ:
4239 if (tp->repair_queue == TCP_SEND_QUEUE)
4240 val = tp->write_seq;
4241 else if (tp->repair_queue == TCP_RECV_QUEUE)
4242 val = tp->rcv_nxt;
4243 else
4244 return -EINVAL;
4245 break;
4246
4247 case TCP_USER_TIMEOUT:
4248 val = icsk->icsk_user_timeout;
4249 break;
4250
4251 case TCP_FASTOPEN:
4252 val = icsk->icsk_accept_queue.fastopenq.max_qlen;
4253 break;
4254
4255 case TCP_FASTOPEN_CONNECT:
4256 val = tp->fastopen_connect;
4257 break;
4258
4259 case TCP_FASTOPEN_NO_COOKIE:
4260 val = tp->fastopen_no_cookie;
4261 break;
4262
4263 case TCP_TX_DELAY:
4264 val = tp->tcp_tx_delay;
4265 break;
4266
4267 case TCP_TIMESTAMP:
4268 val = tcp_time_stamp_raw() + tp->tsoffset;
4269 break;
4270 case TCP_NOTSENT_LOWAT:
4271 val = tp->notsent_lowat;
4272 break;
4273 case TCP_INQ:
4274 val = tp->recvmsg_inq;
4275 break;
4276 case TCP_SAVE_SYN:
4277 val = tp->save_syn;
4278 break;
4279 case TCP_SAVED_SYN: {
4280 if (copy_from_sockptr(&len, optlen, sizeof(int)))
4281 return -EFAULT;
4282
4283 sockopt_lock_sock(sk);
4284 if (tp->saved_syn) {
4285 if (len < tcp_saved_syn_len(tp->saved_syn)) {
4286 len = tcp_saved_syn_len(tp->saved_syn);
4287 if (copy_to_sockptr(optlen, &len, sizeof(int))) {
4288 sockopt_release_sock(sk);
4289 return -EFAULT;
4290 }
4291 sockopt_release_sock(sk);
4292 return -EINVAL;
4293 }
4294 len = tcp_saved_syn_len(tp->saved_syn);
4295 if (copy_to_sockptr(optlen, &len, sizeof(int))) {
4296 sockopt_release_sock(sk);
4297 return -EFAULT;
4298 }
4299 if (copy_to_sockptr(optval, tp->saved_syn->data, len)) {
4300 sockopt_release_sock(sk);
4301 return -EFAULT;
4302 }
4303 tcp_saved_syn_free(tp);
4304 sockopt_release_sock(sk);
4305 } else {
4306 sockopt_release_sock(sk);
4307 len = 0;
4308 if (copy_to_sockptr(optlen, &len, sizeof(int)))
4309 return -EFAULT;
4310 }
4311 return 0;
4312 }
4313#ifdef CONFIG_MMU
4314 case TCP_ZEROCOPY_RECEIVE: {
4315 struct scm_timestamping_internal tss;
4316 struct tcp_zerocopy_receive zc = {};
4317 int err;
4318
4319 if (copy_from_sockptr(&len, optlen, sizeof(int)))
4320 return -EFAULT;
4321 if (len < 0 ||
4322 len < offsetofend(struct tcp_zerocopy_receive, length))
4323 return -EINVAL;
4324 if (unlikely(len > sizeof(zc))) {
4325 err = check_zeroed_sockptr(optval, sizeof(zc),
4326 len - sizeof(zc));
4327 if (err < 1)
4328 return err == 0 ? -EINVAL : err;
4329 len = sizeof(zc);
4330 if (copy_to_sockptr(optlen, &len, sizeof(int)))
4331 return -EFAULT;
4332 }
4333 if (copy_from_sockptr(&zc, optval, len))
4334 return -EFAULT;
4335 if (zc.reserved)
4336 return -EINVAL;
4337 if (zc.msg_flags & ~(TCP_VALID_ZC_MSG_FLAGS))
4338 return -EINVAL;
4339 sockopt_lock_sock(sk);
4340 err = tcp_zerocopy_receive(sk, &zc, &tss);
4341 err = BPF_CGROUP_RUN_PROG_GETSOCKOPT_KERN(sk, level, optname,
4342 &zc, &len, err);
4343 sockopt_release_sock(sk);
4344 if (len >= offsetofend(struct tcp_zerocopy_receive, msg_flags))
4345 goto zerocopy_rcv_cmsg;
4346 switch (len) {
4347 case offsetofend(struct tcp_zerocopy_receive, msg_flags):
4348 goto zerocopy_rcv_cmsg;
4349 case offsetofend(struct tcp_zerocopy_receive, msg_controllen):
4350 case offsetofend(struct tcp_zerocopy_receive, msg_control):
4351 case offsetofend(struct tcp_zerocopy_receive, flags):
4352 case offsetofend(struct tcp_zerocopy_receive, copybuf_len):
4353 case offsetofend(struct tcp_zerocopy_receive, copybuf_address):
4354 case offsetofend(struct tcp_zerocopy_receive, err):
4355 goto zerocopy_rcv_sk_err;
4356 case offsetofend(struct tcp_zerocopy_receive, inq):
4357 goto zerocopy_rcv_inq;
4358 case offsetofend(struct tcp_zerocopy_receive, length):
4359 default:
4360 goto zerocopy_rcv_out;
4361 }
4362zerocopy_rcv_cmsg:
4363 if (zc.msg_flags & TCP_CMSG_TS)
4364 tcp_zc_finalize_rx_tstamp(sk, &zc, &tss);
4365 else
4366 zc.msg_flags = 0;
4367zerocopy_rcv_sk_err:
4368 if (!err)
4369 zc.err = sock_error(sk);
4370zerocopy_rcv_inq:
4371 zc.inq = tcp_inq_hint(sk);
4372zerocopy_rcv_out:
4373 if (!err && copy_to_sockptr(optval, &zc, len))
4374 err = -EFAULT;
4375 return err;
4376 }
4377#endif
4378 default:
4379 return -ENOPROTOOPT;
4380 }
4381
4382 if (copy_to_sockptr(optlen, &len, sizeof(int)))
4383 return -EFAULT;
4384 if (copy_to_sockptr(optval, &val, len))
4385 return -EFAULT;
4386 return 0;
4387}
4388
4389bool tcp_bpf_bypass_getsockopt(int level, int optname)
4390{
4391 /* TCP do_tcp_getsockopt has optimized getsockopt implementation
4392 * to avoid extra socket lock for TCP_ZEROCOPY_RECEIVE.
4393 */
4394 if (level == SOL_TCP && optname == TCP_ZEROCOPY_RECEIVE)
4395 return true;
4396
4397 return false;
4398}
4399EXPORT_SYMBOL(tcp_bpf_bypass_getsockopt);
4400
4401int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval,
4402 int __user *optlen)
4403{
4404 struct inet_connection_sock *icsk = inet_csk(sk);
4405
4406 if (level != SOL_TCP)
4407 /* Paired with WRITE_ONCE() in do_ipv6_setsockopt() and tcp_v6_connect() */
4408 return READ_ONCE(icsk->icsk_af_ops)->getsockopt(sk, level, optname,
4409 optval, optlen);
4410 return do_tcp_getsockopt(sk, level, optname, USER_SOCKPTR(optval),
4411 USER_SOCKPTR(optlen));
4412}
4413EXPORT_SYMBOL(tcp_getsockopt);
4414
4415#ifdef CONFIG_TCP_MD5SIG
4416static DEFINE_PER_CPU(struct tcp_md5sig_pool, tcp_md5sig_pool);
4417static DEFINE_MUTEX(tcp_md5sig_mutex);
4418static bool tcp_md5sig_pool_populated = false;
4419
4420static void __tcp_alloc_md5sig_pool(void)
4421{
4422 struct crypto_ahash *hash;
4423 int cpu;
4424
4425 hash = crypto_alloc_ahash("md5", 0, CRYPTO_ALG_ASYNC);
4426 if (IS_ERR(hash))
4427 return;
4428
4429 for_each_possible_cpu(cpu) {
4430 void *scratch = per_cpu(tcp_md5sig_pool, cpu).scratch;
4431 struct ahash_request *req;
4432
4433 if (!scratch) {
4434 scratch = kmalloc_node(sizeof(union tcp_md5sum_block) +
4435 sizeof(struct tcphdr),
4436 GFP_KERNEL,
4437 cpu_to_node(cpu));
4438 if (!scratch)
4439 return;
4440 per_cpu(tcp_md5sig_pool, cpu).scratch = scratch;
4441 }
4442 if (per_cpu(tcp_md5sig_pool, cpu).md5_req)
4443 continue;
4444
4445 req = ahash_request_alloc(hash, GFP_KERNEL);
4446 if (!req)
4447 return;
4448
4449 ahash_request_set_callback(req, 0, NULL, NULL);
4450
4451 per_cpu(tcp_md5sig_pool, cpu).md5_req = req;
4452 }
4453 /* before setting tcp_md5sig_pool_populated, we must commit all writes
4454 * to memory. See smp_rmb() in tcp_get_md5sig_pool()
4455 */
4456 smp_wmb();
4457 /* Paired with READ_ONCE() from tcp_alloc_md5sig_pool()
4458 * and tcp_get_md5sig_pool().
4459 */
4460 WRITE_ONCE(tcp_md5sig_pool_populated, true);
4461}
4462
4463bool tcp_alloc_md5sig_pool(void)
4464{
4465 /* Paired with WRITE_ONCE() from __tcp_alloc_md5sig_pool() */
4466 if (unlikely(!READ_ONCE(tcp_md5sig_pool_populated))) {
4467 mutex_lock(&tcp_md5sig_mutex);
4468
4469 if (!tcp_md5sig_pool_populated)
4470 __tcp_alloc_md5sig_pool();
4471
4472 mutex_unlock(&tcp_md5sig_mutex);
4473 }
4474 /* Paired with WRITE_ONCE() from __tcp_alloc_md5sig_pool() */
4475 return READ_ONCE(tcp_md5sig_pool_populated);
4476}
4477EXPORT_SYMBOL(tcp_alloc_md5sig_pool);
4478
4479
4480/**
4481 * tcp_get_md5sig_pool - get md5sig_pool for this user
4482 *
4483 * We use percpu structure, so if we succeed, we exit with preemption
4484 * and BH disabled, to make sure another thread or softirq handling
4485 * wont try to get same context.
4486 */
4487struct tcp_md5sig_pool *tcp_get_md5sig_pool(void)
4488{
4489 local_bh_disable();
4490
4491 /* Paired with WRITE_ONCE() from __tcp_alloc_md5sig_pool() */
4492 if (READ_ONCE(tcp_md5sig_pool_populated)) {
4493 /* coupled with smp_wmb() in __tcp_alloc_md5sig_pool() */
4494 smp_rmb();
4495 return this_cpu_ptr(&tcp_md5sig_pool);
4496 }
4497 local_bh_enable();
4498 return NULL;
4499}
4500EXPORT_SYMBOL(tcp_get_md5sig_pool);
4501
4502int tcp_md5_hash_skb_data(struct tcp_md5sig_pool *hp,
4503 const struct sk_buff *skb, unsigned int header_len)
4504{
4505 struct scatterlist sg;
4506 const struct tcphdr *tp = tcp_hdr(skb);
4507 struct ahash_request *req = hp->md5_req;
4508 unsigned int i;
4509 const unsigned int head_data_len = skb_headlen(skb) > header_len ?
4510 skb_headlen(skb) - header_len : 0;
4511 const struct skb_shared_info *shi = skb_shinfo(skb);
4512 struct sk_buff *frag_iter;
4513
4514 sg_init_table(&sg, 1);
4515
4516 sg_set_buf(&sg, ((u8 *) tp) + header_len, head_data_len);
4517 ahash_request_set_crypt(req, &sg, NULL, head_data_len);
4518 if (crypto_ahash_update(req))
4519 return 1;
4520
4521 for (i = 0; i < shi->nr_frags; ++i) {
4522 const skb_frag_t *f = &shi->frags[i];
4523 unsigned int offset = skb_frag_off(f);
4524 struct page *page = skb_frag_page(f) + (offset >> PAGE_SHIFT);
4525
4526 sg_set_page(&sg, page, skb_frag_size(f),
4527 offset_in_page(offset));
4528 ahash_request_set_crypt(req, &sg, NULL, skb_frag_size(f));
4529 if (crypto_ahash_update(req))
4530 return 1;
4531 }
4532
4533 skb_walk_frags(skb, frag_iter)
4534 if (tcp_md5_hash_skb_data(hp, frag_iter, 0))
4535 return 1;
4536
4537 return 0;
4538}
4539EXPORT_SYMBOL(tcp_md5_hash_skb_data);
4540
4541int tcp_md5_hash_key(struct tcp_md5sig_pool *hp, const struct tcp_md5sig_key *key)
4542{
4543 u8 keylen = READ_ONCE(key->keylen); /* paired with WRITE_ONCE() in tcp_md5_do_add */
4544 struct scatterlist sg;
4545
4546 sg_init_one(&sg, key->key, keylen);
4547 ahash_request_set_crypt(hp->md5_req, &sg, NULL, keylen);
4548
4549 /* We use data_race() because tcp_md5_do_add() might change key->key under us */
4550 return data_race(crypto_ahash_update(hp->md5_req));
4551}
4552EXPORT_SYMBOL(tcp_md5_hash_key);
4553
4554/* Called with rcu_read_lock() */
4555enum skb_drop_reason
4556tcp_inbound_md5_hash(const struct sock *sk, const struct sk_buff *skb,
4557 const void *saddr, const void *daddr,
4558 int family, int dif, int sdif)
4559{
4560 /*
4561 * This gets called for each TCP segment that arrives
4562 * so we want to be efficient.
4563 * We have 3 drop cases:
4564 * o No MD5 hash and one expected.
4565 * o MD5 hash and we're not expecting one.
4566 * o MD5 hash and its wrong.
4567 */
4568 const __u8 *hash_location = NULL;
4569 struct tcp_md5sig_key *hash_expected;
4570 const struct tcphdr *th = tcp_hdr(skb);
4571 struct tcp_sock *tp = tcp_sk(sk);
4572 int genhash, l3index;
4573 u8 newhash[16];
4574
4575 /* sdif set, means packet ingressed via a device
4576 * in an L3 domain and dif is set to the l3mdev
4577 */
4578 l3index = sdif ? dif : 0;
4579
4580 hash_expected = tcp_md5_do_lookup(sk, l3index, saddr, family);
4581 hash_location = tcp_parse_md5sig_option(th);
4582
4583 /* We've parsed the options - do we have a hash? */
4584 if (!hash_expected && !hash_location)
4585 return SKB_NOT_DROPPED_YET;
4586
4587 if (hash_expected && !hash_location) {
4588 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMD5NOTFOUND);
4589 return SKB_DROP_REASON_TCP_MD5NOTFOUND;
4590 }
4591
4592 if (!hash_expected && hash_location) {
4593 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMD5UNEXPECTED);
4594 return SKB_DROP_REASON_TCP_MD5UNEXPECTED;
4595 }
4596
4597 /* Check the signature.
4598 * To support dual stack listeners, we need to handle
4599 * IPv4-mapped case.
4600 */
4601 if (family == AF_INET)
4602 genhash = tcp_v4_md5_hash_skb(newhash,
4603 hash_expected,
4604 NULL, skb);
4605 else
4606 genhash = tp->af_specific->calc_md5_hash(newhash,
4607 hash_expected,
4608 NULL, skb);
4609
4610 if (genhash || memcmp(hash_location, newhash, 16) != 0) {
4611 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMD5FAILURE);
4612 if (family == AF_INET) {
4613 net_info_ratelimited("MD5 Hash failed for (%pI4, %d)->(%pI4, %d)%s L3 index %d\n",
4614 saddr, ntohs(th->source),
4615 daddr, ntohs(th->dest),
4616 genhash ? " tcp_v4_calc_md5_hash failed"
4617 : "", l3index);
4618 } else {
4619 net_info_ratelimited("MD5 Hash %s for [%pI6c]:%u->[%pI6c]:%u L3 index %d\n",
4620 genhash ? "failed" : "mismatch",
4621 saddr, ntohs(th->source),
4622 daddr, ntohs(th->dest), l3index);
4623 }
4624 return SKB_DROP_REASON_TCP_MD5FAILURE;
4625 }
4626 return SKB_NOT_DROPPED_YET;
4627}
4628EXPORT_SYMBOL(tcp_inbound_md5_hash);
4629
4630#endif
4631
4632void tcp_done(struct sock *sk)
4633{
4634 struct request_sock *req;
4635
4636 /* We might be called with a new socket, after
4637 * inet_csk_prepare_forced_close() has been called
4638 * so we can not use lockdep_sock_is_held(sk)
4639 */
4640 req = rcu_dereference_protected(tcp_sk(sk)->fastopen_rsk, 1);
4641
4642 if (sk->sk_state == TCP_SYN_SENT || sk->sk_state == TCP_SYN_RECV)
4643 TCP_INC_STATS(sock_net(sk), TCP_MIB_ATTEMPTFAILS);
4644
4645 tcp_set_state(sk, TCP_CLOSE);
4646 tcp_clear_xmit_timers(sk);
4647 if (req)
4648 reqsk_fastopen_remove(sk, req, false);
4649
4650 sk->sk_shutdown = SHUTDOWN_MASK;
4651
4652 if (!sock_flag(sk, SOCK_DEAD))
4653 sk->sk_state_change(sk);
4654 else
4655 inet_csk_destroy_sock(sk);
4656}
4657EXPORT_SYMBOL_GPL(tcp_done);
4658
4659int tcp_abort(struct sock *sk, int err)
4660{
4661 int state = inet_sk_state_load(sk);
4662
4663 if (state == TCP_NEW_SYN_RECV) {
4664 struct request_sock *req = inet_reqsk(sk);
4665
4666 local_bh_disable();
4667 inet_csk_reqsk_queue_drop(req->rsk_listener, req);
4668 local_bh_enable();
4669 return 0;
4670 }
4671 if (state == TCP_TIME_WAIT) {
4672 struct inet_timewait_sock *tw = inet_twsk(sk);
4673
4674 refcount_inc(&tw->tw_refcnt);
4675 local_bh_disable();
4676 inet_twsk_deschedule_put(tw);
4677 local_bh_enable();
4678 return 0;
4679 }
4680
4681 /* Don't race with userspace socket closes such as tcp_close. */
4682 lock_sock(sk);
4683
4684 if (sk->sk_state == TCP_LISTEN) {
4685 tcp_set_state(sk, TCP_CLOSE);
4686 inet_csk_listen_stop(sk);
4687 }
4688
4689 /* Don't race with BH socket closes such as inet_csk_listen_stop. */
4690 local_bh_disable();
4691 bh_lock_sock(sk);
4692
4693 if (!sock_flag(sk, SOCK_DEAD)) {
4694 sk->sk_err = err;
4695 /* This barrier is coupled with smp_rmb() in tcp_poll() */
4696 smp_wmb();
4697 sk_error_report(sk);
4698 if (tcp_need_reset(sk->sk_state))
4699 tcp_send_active_reset(sk, GFP_ATOMIC);
4700 tcp_done(sk);
4701 }
4702
4703 bh_unlock_sock(sk);
4704 local_bh_enable();
4705 tcp_write_queue_purge(sk);
4706 release_sock(sk);
4707 return 0;
4708}
4709EXPORT_SYMBOL_GPL(tcp_abort);
4710
4711extern struct tcp_congestion_ops tcp_reno;
4712
4713static __initdata unsigned long thash_entries;
4714static int __init set_thash_entries(char *str)
4715{
4716 ssize_t ret;
4717
4718 if (!str)
4719 return 0;
4720
4721 ret = kstrtoul(str, 0, &thash_entries);
4722 if (ret)
4723 return 0;
4724
4725 return 1;
4726}
4727__setup("thash_entries=", set_thash_entries);
4728
4729static void __init tcp_init_mem(void)
4730{
4731 unsigned long limit = nr_free_buffer_pages() / 16;
4732
4733 limit = max(limit, 128UL);
4734 sysctl_tcp_mem[0] = limit / 4 * 3; /* 4.68 % */
4735 sysctl_tcp_mem[1] = limit; /* 6.25 % */
4736 sysctl_tcp_mem[2] = sysctl_tcp_mem[0] * 2; /* 9.37 % */
4737}
4738
4739void __init tcp_init(void)
4740{
4741 int max_rshare, max_wshare, cnt;
4742 unsigned long limit;
4743 unsigned int i;
4744
4745 BUILD_BUG_ON(TCP_MIN_SND_MSS <= MAX_TCP_OPTION_SPACE);
4746 BUILD_BUG_ON(sizeof(struct tcp_skb_cb) >
4747 sizeof_field(struct sk_buff, cb));
4748
4749 percpu_counter_init(&tcp_sockets_allocated, 0, GFP_KERNEL);
4750
4751 timer_setup(&tcp_orphan_timer, tcp_orphan_update, TIMER_DEFERRABLE);
4752 mod_timer(&tcp_orphan_timer, jiffies + TCP_ORPHAN_TIMER_PERIOD);
4753
4754 inet_hashinfo2_init(&tcp_hashinfo, "tcp_listen_portaddr_hash",
4755 thash_entries, 21, /* one slot per 2 MB*/
4756 0, 64 * 1024);
4757 tcp_hashinfo.bind_bucket_cachep =
4758 kmem_cache_create("tcp_bind_bucket",
4759 sizeof(struct inet_bind_bucket), 0,
4760 SLAB_HWCACHE_ALIGN | SLAB_PANIC |
4761 SLAB_ACCOUNT,
4762 NULL);
4763 tcp_hashinfo.bind2_bucket_cachep =
4764 kmem_cache_create("tcp_bind2_bucket",
4765 sizeof(struct inet_bind2_bucket), 0,
4766 SLAB_HWCACHE_ALIGN | SLAB_PANIC |
4767 SLAB_ACCOUNT,
4768 NULL);
4769
4770 /* Size and allocate the main established and bind bucket
4771 * hash tables.
4772 *
4773 * The methodology is similar to that of the buffer cache.
4774 */
4775 tcp_hashinfo.ehash =
4776 alloc_large_system_hash("TCP established",
4777 sizeof(struct inet_ehash_bucket),
4778 thash_entries,
4779 17, /* one slot per 128 KB of memory */
4780 0,
4781 NULL,
4782 &tcp_hashinfo.ehash_mask,
4783 0,
4784 thash_entries ? 0 : 512 * 1024);
4785 for (i = 0; i <= tcp_hashinfo.ehash_mask; i++)
4786 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].chain, i);
4787
4788 if (inet_ehash_locks_alloc(&tcp_hashinfo))
4789 panic("TCP: failed to alloc ehash_locks");
4790 tcp_hashinfo.bhash =
4791 alloc_large_system_hash("TCP bind",
4792 2 * sizeof(struct inet_bind_hashbucket),
4793 tcp_hashinfo.ehash_mask + 1,
4794 17, /* one slot per 128 KB of memory */
4795 0,
4796 &tcp_hashinfo.bhash_size,
4797 NULL,
4798 0,
4799 64 * 1024);
4800 tcp_hashinfo.bhash_size = 1U << tcp_hashinfo.bhash_size;
4801 tcp_hashinfo.bhash2 = tcp_hashinfo.bhash + tcp_hashinfo.bhash_size;
4802 for (i = 0; i < tcp_hashinfo.bhash_size; i++) {
4803 spin_lock_init(&tcp_hashinfo.bhash[i].lock);
4804 INIT_HLIST_HEAD(&tcp_hashinfo.bhash[i].chain);
4805 spin_lock_init(&tcp_hashinfo.bhash2[i].lock);
4806 INIT_HLIST_HEAD(&tcp_hashinfo.bhash2[i].chain);
4807 }
4808
4809 tcp_hashinfo.pernet = false;
4810
4811 cnt = tcp_hashinfo.ehash_mask + 1;
4812 sysctl_tcp_max_orphans = cnt / 2;
4813
4814 tcp_init_mem();
4815 /* Set per-socket limits to no more than 1/128 the pressure threshold */
4816 limit = nr_free_buffer_pages() << (PAGE_SHIFT - 7);
4817 max_wshare = min(4UL*1024*1024, limit);
4818 max_rshare = min(6UL*1024*1024, limit);
4819
4820 init_net.ipv4.sysctl_tcp_wmem[0] = PAGE_SIZE;
4821 init_net.ipv4.sysctl_tcp_wmem[1] = 16*1024;
4822 init_net.ipv4.sysctl_tcp_wmem[2] = max(64*1024, max_wshare);
4823
4824 init_net.ipv4.sysctl_tcp_rmem[0] = PAGE_SIZE;
4825 init_net.ipv4.sysctl_tcp_rmem[1] = 131072;
4826 init_net.ipv4.sysctl_tcp_rmem[2] = max(131072, max_rshare);
4827
4828 pr_info("Hash tables configured (established %u bind %u)\n",
4829 tcp_hashinfo.ehash_mask + 1, tcp_hashinfo.bhash_size);
4830
4831 tcp_v4_init();
4832 tcp_metrics_init();
4833 BUG_ON(tcp_register_congestion_control(&tcp_reno) != 0);
4834 tcp_tasklet_init();
4835 mptcp_init();
4836}