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