Loading...
1// SPDX-License-Identifier: GPL-2.0-or-later
2/*
3 * INET An implementation of the TCP/IP protocol suite for the LINUX
4 * operating system. INET is implemented using the BSD Socket
5 * interface as the means of communication with the user level.
6 *
7 * Implementation of the Transmission Control Protocol(TCP).
8 *
9 * Authors: Ross Biro
10 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
11 * Mark Evans, <evansmp@uhura.aston.ac.uk>
12 * Corey Minyard <wf-rch!minyard@relay.EU.net>
13 * Florian La Roche, <flla@stud.uni-sb.de>
14 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
15 * Linus Torvalds, <torvalds@cs.helsinki.fi>
16 * Alan Cox, <gw4pts@gw4pts.ampr.org>
17 * Matthew Dillon, <dillon@apollo.west.oic.com>
18 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
19 * Jorge Cwik, <jorge@laser.satlink.net>
20 *
21 * Fixes:
22 * Alan Cox : Numerous verify_area() calls
23 * Alan Cox : Set the ACK bit on a reset
24 * Alan Cox : Stopped it crashing if it closed while
25 * sk->inuse=1 and was trying to connect
26 * (tcp_err()).
27 * Alan Cox : All icmp error handling was broken
28 * pointers passed where wrong and the
29 * socket was looked up backwards. Nobody
30 * tested any icmp error code obviously.
31 * Alan Cox : tcp_err() now handled properly. It
32 * wakes people on errors. poll
33 * behaves and the icmp error race
34 * has gone by moving it into sock.c
35 * Alan Cox : tcp_send_reset() fixed to work for
36 * everything not just packets for
37 * unknown sockets.
38 * Alan Cox : tcp option processing.
39 * Alan Cox : Reset tweaked (still not 100%) [Had
40 * syn rule wrong]
41 * Herp Rosmanith : More reset fixes
42 * Alan Cox : No longer acks invalid rst frames.
43 * Acking any kind of RST is right out.
44 * Alan Cox : Sets an ignore me flag on an rst
45 * receive otherwise odd bits of prattle
46 * escape still
47 * Alan Cox : Fixed another acking RST frame bug.
48 * Should stop LAN workplace lockups.
49 * Alan Cox : Some tidyups using the new skb list
50 * facilities
51 * Alan Cox : sk->keepopen now seems to work
52 * Alan Cox : Pulls options out correctly on accepts
53 * Alan Cox : Fixed assorted sk->rqueue->next errors
54 * Alan Cox : PSH doesn't end a TCP read. Switched a
55 * bit to skb ops.
56 * Alan Cox : Tidied tcp_data to avoid a potential
57 * nasty.
58 * Alan Cox : Added some better commenting, as the
59 * tcp is hard to follow
60 * Alan Cox : Removed incorrect check for 20 * psh
61 * Michael O'Reilly : ack < copied bug fix.
62 * Johannes Stille : Misc tcp fixes (not all in yet).
63 * Alan Cox : FIN with no memory -> CRASH
64 * Alan Cox : Added socket option proto entries.
65 * Also added awareness of them to accept.
66 * Alan Cox : Added TCP options (SOL_TCP)
67 * Alan Cox : Switched wakeup calls to callbacks,
68 * so the kernel can layer network
69 * sockets.
70 * Alan Cox : Use ip_tos/ip_ttl settings.
71 * Alan Cox : Handle FIN (more) properly (we hope).
72 * Alan Cox : RST frames sent on unsynchronised
73 * state ack error.
74 * Alan Cox : Put in missing check for SYN bit.
75 * Alan Cox : Added tcp_select_window() aka NET2E
76 * window non shrink trick.
77 * Alan Cox : Added a couple of small NET2E timer
78 * fixes
79 * Charles Hedrick : TCP fixes
80 * Toomas Tamm : TCP window fixes
81 * Alan Cox : Small URG fix to rlogin ^C ack fight
82 * Charles Hedrick : Rewrote most of it to actually work
83 * Linus : Rewrote tcp_read() and URG handling
84 * completely
85 * Gerhard Koerting: Fixed some missing timer handling
86 * Matthew Dillon : Reworked TCP machine states as per RFC
87 * Gerhard Koerting: PC/TCP workarounds
88 * Adam Caldwell : Assorted timer/timing errors
89 * Matthew Dillon : Fixed another RST bug
90 * Alan Cox : Move to kernel side addressing changes.
91 * Alan Cox : Beginning work on TCP fastpathing
92 * (not yet usable)
93 * Arnt Gulbrandsen: Turbocharged tcp_check() routine.
94 * Alan Cox : TCP fast path debugging
95 * Alan Cox : Window clamping
96 * Michael Riepe : Bug in tcp_check()
97 * Matt Dillon : More TCP improvements and RST bug fixes
98 * Matt Dillon : Yet more small nasties remove from the
99 * TCP code (Be very nice to this man if
100 * tcp finally works 100%) 8)
101 * Alan Cox : BSD accept semantics.
102 * Alan Cox : Reset on closedown bug.
103 * Peter De Schrijver : ENOTCONN check missing in tcp_sendto().
104 * Michael Pall : Handle poll() after URG properly in
105 * all cases.
106 * Michael Pall : Undo the last fix in tcp_read_urg()
107 * (multi URG PUSH broke rlogin).
108 * Michael Pall : Fix the multi URG PUSH problem in
109 * tcp_readable(), poll() after URG
110 * works now.
111 * Michael Pall : recv(...,MSG_OOB) never blocks in the
112 * BSD api.
113 * Alan Cox : Changed the semantics of sk->socket to
114 * fix a race and a signal problem with
115 * accept() and async I/O.
116 * Alan Cox : Relaxed the rules on tcp_sendto().
117 * Yury Shevchuk : Really fixed accept() blocking problem.
118 * Craig I. Hagan : Allow for BSD compatible TIME_WAIT for
119 * clients/servers which listen in on
120 * fixed ports.
121 * Alan Cox : Cleaned the above up and shrank it to
122 * a sensible code size.
123 * Alan Cox : Self connect lockup fix.
124 * Alan Cox : No connect to multicast.
125 * Ross Biro : Close unaccepted children on master
126 * socket close.
127 * Alan Cox : Reset tracing code.
128 * Alan Cox : Spurious resets on shutdown.
129 * Alan Cox : Giant 15 minute/60 second timer error
130 * Alan Cox : Small whoops in polling before an
131 * accept.
132 * Alan Cox : Kept the state trace facility since
133 * it's handy for debugging.
134 * Alan Cox : More reset handler fixes.
135 * Alan Cox : Started rewriting the code based on
136 * the RFC's for other useful protocol
137 * references see: Comer, KA9Q NOS, and
138 * for a reference on the difference
139 * between specifications and how BSD
140 * works see the 4.4lite source.
141 * A.N.Kuznetsov : Don't time wait on completion of tidy
142 * close.
143 * Linus Torvalds : Fin/Shutdown & copied_seq changes.
144 * Linus Torvalds : Fixed BSD port reuse to work first syn
145 * Alan Cox : Reimplemented timers as per the RFC
146 * and using multiple timers for sanity.
147 * Alan Cox : Small bug fixes, and a lot of new
148 * comments.
149 * Alan Cox : Fixed dual reader crash by locking
150 * the buffers (much like datagram.c)
151 * Alan Cox : Fixed stuck sockets in probe. A probe
152 * now gets fed up of retrying without
153 * (even a no space) answer.
154 * Alan Cox : Extracted closing code better
155 * Alan Cox : Fixed the closing state machine to
156 * resemble the RFC.
157 * Alan Cox : More 'per spec' fixes.
158 * Jorge Cwik : Even faster checksumming.
159 * Alan Cox : tcp_data() doesn't ack illegal PSH
160 * only frames. At least one pc tcp stack
161 * generates them.
162 * Alan Cox : Cache last socket.
163 * Alan Cox : Per route irtt.
164 * Matt Day : poll()->select() match BSD precisely on error
165 * Alan Cox : New buffers
166 * Marc Tamsky : Various sk->prot->retransmits and
167 * sk->retransmits misupdating fixed.
168 * Fixed tcp_write_timeout: stuck close,
169 * and TCP syn retries gets used now.
170 * Mark Yarvis : In tcp_read_wakeup(), don't send an
171 * ack if state is TCP_CLOSED.
172 * Alan Cox : Look up device on a retransmit - routes may
173 * change. Doesn't yet cope with MSS shrink right
174 * but it's a start!
175 * Marc Tamsky : Closing in closing fixes.
176 * Mike Shaver : RFC1122 verifications.
177 * Alan Cox : rcv_saddr errors.
178 * Alan Cox : Block double connect().
179 * Alan Cox : Small hooks for enSKIP.
180 * Alexey Kuznetsov: Path MTU discovery.
181 * Alan Cox : Support soft errors.
182 * Alan Cox : Fix MTU discovery pathological case
183 * when the remote claims no mtu!
184 * Marc Tamsky : TCP_CLOSE fix.
185 * Colin (G3TNE) : Send a reset on syn ack replies in
186 * window but wrong (fixes NT lpd problems)
187 * Pedro Roque : Better TCP window handling, delayed ack.
188 * Joerg Reuter : No modification of locked buffers in
189 * tcp_do_retransmit()
190 * Eric Schenk : Changed receiver side silly window
191 * avoidance algorithm to BSD style
192 * algorithm. This doubles throughput
193 * against machines running Solaris,
194 * and seems to result in general
195 * improvement.
196 * Stefan Magdalinski : adjusted tcp_readable() to fix FIONREAD
197 * Willy Konynenberg : Transparent proxying support.
198 * Mike McLagan : Routing by source
199 * Keith Owens : Do proper merging with partial SKB's in
200 * tcp_do_sendmsg to avoid burstiness.
201 * Eric Schenk : Fix fast close down bug with
202 * shutdown() followed by close().
203 * Andi Kleen : Make poll agree with SIGIO
204 * Salvatore Sanfilippo : Support SO_LINGER with linger == 1 and
205 * lingertime == 0 (RFC 793 ABORT Call)
206 * Hirokazu Takahashi : Use copy_from_user() instead of
207 * csum_and_copy_from_user() if possible.
208 *
209 * Description of States:
210 *
211 * TCP_SYN_SENT sent a connection request, waiting for ack
212 *
213 * TCP_SYN_RECV received a connection request, sent ack,
214 * waiting for final ack in three-way handshake.
215 *
216 * TCP_ESTABLISHED connection established
217 *
218 * TCP_FIN_WAIT1 our side has shutdown, waiting to complete
219 * transmission of remaining buffered data
220 *
221 * TCP_FIN_WAIT2 all buffered data sent, waiting for remote
222 * to shutdown
223 *
224 * TCP_CLOSING both sides have shutdown but we still have
225 * data we have to finish sending
226 *
227 * TCP_TIME_WAIT timeout to catch resent junk before entering
228 * closed, can only be entered from FIN_WAIT2
229 * or CLOSING. Required because the other end
230 * may not have gotten our last ACK causing it
231 * to retransmit the data packet (which we ignore)
232 *
233 * TCP_CLOSE_WAIT remote side has shutdown and is waiting for
234 * us to finish writing our data and to shutdown
235 * (we have to close() to move on to LAST_ACK)
236 *
237 * TCP_LAST_ACK out side has shutdown after remote has
238 * shutdown. There may still be data in our
239 * buffer that we have to finish sending
240 *
241 * TCP_CLOSE socket is finished
242 */
243
244#define pr_fmt(fmt) "TCP: " fmt
245
246#include <crypto/hash.h>
247#include <linux/kernel.h>
248#include <linux/module.h>
249#include <linux/types.h>
250#include <linux/fcntl.h>
251#include <linux/poll.h>
252#include <linux/inet_diag.h>
253#include <linux/init.h>
254#include <linux/fs.h>
255#include <linux/skbuff.h>
256#include <linux/scatterlist.h>
257#include <linux/splice.h>
258#include <linux/net.h>
259#include <linux/socket.h>
260#include <linux/random.h>
261#include <linux/memblock.h>
262#include <linux/highmem.h>
263#include <linux/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/proto_memory.h>
276#include <net/xfrm.h>
277#include <net/ip.h>
278#include <net/sock.h>
279#include <net/rstreason.h>
280
281#include <linux/uaccess.h>
282#include <asm/ioctls.h>
283#include <net/busy_poll.h>
284#include <net/hotdata.h>
285#include <trace/events/tcp.h>
286#include <net/rps.h>
287
288#include "../core/devmem.h"
289
290/* Track pending CMSGs. */
291enum {
292 TCP_CMSG_INQ = 1,
293 TCP_CMSG_TS = 2
294};
295
296DEFINE_PER_CPU(unsigned int, tcp_orphan_count);
297EXPORT_PER_CPU_SYMBOL_GPL(tcp_orphan_count);
298
299DEFINE_PER_CPU(u32, tcp_tw_isn);
300EXPORT_PER_CPU_SYMBOL_GPL(tcp_tw_isn);
301
302long sysctl_tcp_mem[3] __read_mostly;
303EXPORT_SYMBOL(sysctl_tcp_mem);
304
305atomic_long_t tcp_memory_allocated ____cacheline_aligned_in_smp; /* Current allocated memory. */
306EXPORT_SYMBOL(tcp_memory_allocated);
307DEFINE_PER_CPU(int, tcp_memory_per_cpu_fw_alloc);
308EXPORT_PER_CPU_SYMBOL_GPL(tcp_memory_per_cpu_fw_alloc);
309
310#if IS_ENABLED(CONFIG_SMC)
311DEFINE_STATIC_KEY_FALSE(tcp_have_smc);
312EXPORT_SYMBOL(tcp_have_smc);
313#endif
314
315/*
316 * Current number of TCP sockets.
317 */
318struct percpu_counter tcp_sockets_allocated ____cacheline_aligned_in_smp;
319EXPORT_SYMBOL(tcp_sockets_allocated);
320
321/*
322 * TCP splice context
323 */
324struct tcp_splice_state {
325 struct pipe_inode_info *pipe;
326 size_t len;
327 unsigned int flags;
328};
329
330/*
331 * Pressure flag: try to collapse.
332 * Technical note: it is used by multiple contexts non atomically.
333 * All the __sk_mem_schedule() is of this nature: accounting
334 * is strict, actions are advisory and have some latency.
335 */
336unsigned long tcp_memory_pressure __read_mostly;
337EXPORT_SYMBOL_GPL(tcp_memory_pressure);
338
339void tcp_enter_memory_pressure(struct sock *sk)
340{
341 unsigned long val;
342
343 if (READ_ONCE(tcp_memory_pressure))
344 return;
345 val = jiffies;
346
347 if (!val)
348 val--;
349 if (!cmpxchg(&tcp_memory_pressure, 0, val))
350 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMEMORYPRESSURES);
351}
352EXPORT_SYMBOL_GPL(tcp_enter_memory_pressure);
353
354void tcp_leave_memory_pressure(struct sock *sk)
355{
356 unsigned long val;
357
358 if (!READ_ONCE(tcp_memory_pressure))
359 return;
360 val = xchg(&tcp_memory_pressure, 0);
361 if (val)
362 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPMEMORYPRESSURESCHRONO,
363 jiffies_to_msecs(jiffies - val));
364}
365EXPORT_SYMBOL_GPL(tcp_leave_memory_pressure);
366
367/* Convert seconds to retransmits based on initial and max timeout */
368static u8 secs_to_retrans(int seconds, int timeout, int rto_max)
369{
370 u8 res = 0;
371
372 if (seconds > 0) {
373 int period = timeout;
374
375 res = 1;
376 while (seconds > period && res < 255) {
377 res++;
378 timeout <<= 1;
379 if (timeout > rto_max)
380 timeout = rto_max;
381 period += timeout;
382 }
383 }
384 return res;
385}
386
387/* Convert retransmits to seconds based on initial and max timeout */
388static int retrans_to_secs(u8 retrans, int timeout, int rto_max)
389{
390 int period = 0;
391
392 if (retrans > 0) {
393 period = timeout;
394 while (--retrans) {
395 timeout <<= 1;
396 if (timeout > rto_max)
397 timeout = rto_max;
398 period += timeout;
399 }
400 }
401 return period;
402}
403
404static u64 tcp_compute_delivery_rate(const struct tcp_sock *tp)
405{
406 u32 rate = READ_ONCE(tp->rate_delivered);
407 u32 intv = READ_ONCE(tp->rate_interval_us);
408 u64 rate64 = 0;
409
410 if (rate && intv) {
411 rate64 = (u64)rate * tp->mss_cache * USEC_PER_SEC;
412 do_div(rate64, intv);
413 }
414 return rate64;
415}
416
417/* Address-family independent initialization for a tcp_sock.
418 *
419 * NOTE: A lot of things set to zero explicitly by call to
420 * sk_alloc() so need not be done here.
421 */
422void tcp_init_sock(struct sock *sk)
423{
424 struct inet_connection_sock *icsk = inet_csk(sk);
425 struct tcp_sock *tp = tcp_sk(sk);
426 int rto_min_us;
427
428 tp->out_of_order_queue = RB_ROOT;
429 sk->tcp_rtx_queue = RB_ROOT;
430 tcp_init_xmit_timers(sk);
431 INIT_LIST_HEAD(&tp->tsq_node);
432 INIT_LIST_HEAD(&tp->tsorted_sent_queue);
433
434 icsk->icsk_rto = TCP_TIMEOUT_INIT;
435 rto_min_us = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_rto_min_us);
436 icsk->icsk_rto_min = usecs_to_jiffies(rto_min_us);
437 icsk->icsk_delack_max = TCP_DELACK_MAX;
438 tp->mdev_us = jiffies_to_usecs(TCP_TIMEOUT_INIT);
439 minmax_reset(&tp->rtt_min, tcp_jiffies32, ~0U);
440
441 /* So many TCP implementations out there (incorrectly) count the
442 * initial SYN frame in their delayed-ACK and congestion control
443 * algorithms that we must have the following bandaid to talk
444 * efficiently to them. -DaveM
445 */
446 tcp_snd_cwnd_set(tp, TCP_INIT_CWND);
447
448 /* There's a bubble in the pipe until at least the first ACK. */
449 tp->app_limited = ~0U;
450 tp->rate_app_limited = 1;
451
452 /* See draft-stevens-tcpca-spec-01 for discussion of the
453 * initialization of these values.
454 */
455 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
456 tp->snd_cwnd_clamp = ~0;
457 tp->mss_cache = TCP_MSS_DEFAULT;
458
459 tp->reordering = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_reordering);
460 tcp_assign_congestion_control(sk);
461
462 tp->tsoffset = 0;
463 tp->rack.reo_wnd_steps = 1;
464
465 sk->sk_write_space = sk_stream_write_space;
466 sock_set_flag(sk, SOCK_USE_WRITE_QUEUE);
467
468 icsk->icsk_sync_mss = tcp_sync_mss;
469
470 WRITE_ONCE(sk->sk_sndbuf, READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_wmem[1]));
471 WRITE_ONCE(sk->sk_rcvbuf, READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_rmem[1]));
472 tcp_scaling_ratio_init(sk);
473
474 set_bit(SOCK_SUPPORT_ZC, &sk->sk_socket->flags);
475 sk_sockets_allocated_inc(sk);
476 xa_init_flags(&sk->sk_user_frags, XA_FLAGS_ALLOC1);
477}
478EXPORT_SYMBOL(tcp_init_sock);
479
480static void tcp_tx_timestamp(struct sock *sk, struct sockcm_cookie *sockc)
481{
482 struct sk_buff *skb = tcp_write_queue_tail(sk);
483 u32 tsflags = sockc->tsflags;
484
485 if (tsflags && skb) {
486 struct skb_shared_info *shinfo = skb_shinfo(skb);
487 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
488
489 sock_tx_timestamp(sk, sockc, &shinfo->tx_flags);
490 if (tsflags & SOF_TIMESTAMPING_TX_ACK)
491 tcb->txstamp_ack = 1;
492 if (tsflags & SOF_TIMESTAMPING_TX_RECORD_MASK)
493 shinfo->tskey = TCP_SKB_CB(skb)->seq + skb->len - 1;
494 }
495}
496
497static bool tcp_stream_is_readable(struct sock *sk, int target)
498{
499 if (tcp_epollin_ready(sk, target))
500 return true;
501 return sk_is_readable(sk);
502}
503
504/*
505 * Wait for a TCP event.
506 *
507 * Note that we don't need to lock the socket, as the upper poll layers
508 * take care of normal races (between the test and the event) and we don't
509 * go look at any of the socket buffers directly.
510 */
511__poll_t tcp_poll(struct file *file, struct socket *sock, poll_table *wait)
512{
513 __poll_t mask;
514 struct sock *sk = sock->sk;
515 const struct tcp_sock *tp = tcp_sk(sk);
516 u8 shutdown;
517 int state;
518
519 sock_poll_wait(file, sock, wait);
520
521 state = inet_sk_state_load(sk);
522 if (state == TCP_LISTEN)
523 return inet_csk_listen_poll(sk);
524
525 /* Socket is not locked. We are protected from async events
526 * by poll logic and correct handling of state changes
527 * made by other threads is impossible in any case.
528 */
529
530 mask = 0;
531
532 /*
533 * EPOLLHUP is certainly not done right. But poll() doesn't
534 * have a notion of HUP in just one direction, and for a
535 * socket the read side is more interesting.
536 *
537 * Some poll() documentation says that EPOLLHUP is incompatible
538 * with the EPOLLOUT/POLLWR flags, so somebody should check this
539 * all. But careful, it tends to be safer to return too many
540 * bits than too few, and you can easily break real applications
541 * if you don't tell them that something has hung up!
542 *
543 * Check-me.
544 *
545 * Check number 1. EPOLLHUP is _UNMASKABLE_ event (see UNIX98 and
546 * our fs/select.c). It means that after we received EOF,
547 * poll always returns immediately, making impossible poll() on write()
548 * in state CLOSE_WAIT. One solution is evident --- to set EPOLLHUP
549 * if and only if shutdown has been made in both directions.
550 * Actually, it is interesting to look how Solaris and DUX
551 * solve this dilemma. I would prefer, if EPOLLHUP were maskable,
552 * then we could set it on SND_SHUTDOWN. BTW examples given
553 * in Stevens' books assume exactly this behaviour, it explains
554 * why EPOLLHUP is incompatible with EPOLLOUT. --ANK
555 *
556 * NOTE. Check for TCP_CLOSE is added. The goal is to prevent
557 * blocking on fresh not-connected or disconnected socket. --ANK
558 */
559 shutdown = READ_ONCE(sk->sk_shutdown);
560 if (shutdown == SHUTDOWN_MASK || state == TCP_CLOSE)
561 mask |= EPOLLHUP;
562 if (shutdown & RCV_SHUTDOWN)
563 mask |= EPOLLIN | EPOLLRDNORM | EPOLLRDHUP;
564
565 /* Connected or passive Fast Open socket? */
566 if (state != TCP_SYN_SENT &&
567 (state != TCP_SYN_RECV || rcu_access_pointer(tp->fastopen_rsk))) {
568 int target = sock_rcvlowat(sk, 0, INT_MAX);
569 u16 urg_data = READ_ONCE(tp->urg_data);
570
571 if (unlikely(urg_data) &&
572 READ_ONCE(tp->urg_seq) == READ_ONCE(tp->copied_seq) &&
573 !sock_flag(sk, SOCK_URGINLINE))
574 target++;
575
576 if (tcp_stream_is_readable(sk, target))
577 mask |= EPOLLIN | EPOLLRDNORM;
578
579 if (!(shutdown & SEND_SHUTDOWN)) {
580 if (__sk_stream_is_writeable(sk, 1)) {
581 mask |= EPOLLOUT | EPOLLWRNORM;
582 } else { /* send SIGIO later */
583 sk_set_bit(SOCKWQ_ASYNC_NOSPACE, sk);
584 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
585
586 /* Race breaker. If space is freed after
587 * wspace test but before the flags are set,
588 * IO signal will be lost. Memory barrier
589 * pairs with the input side.
590 */
591 smp_mb__after_atomic();
592 if (__sk_stream_is_writeable(sk, 1))
593 mask |= EPOLLOUT | EPOLLWRNORM;
594 }
595 } else
596 mask |= EPOLLOUT | EPOLLWRNORM;
597
598 if (urg_data & TCP_URG_VALID)
599 mask |= EPOLLPRI;
600 } else if (state == TCP_SYN_SENT &&
601 inet_test_bit(DEFER_CONNECT, sk)) {
602 /* Active TCP fastopen socket with defer_connect
603 * Return EPOLLOUT so application can call write()
604 * in order for kernel to generate SYN+data
605 */
606 mask |= EPOLLOUT | EPOLLWRNORM;
607 }
608 /* This barrier is coupled with smp_wmb() in tcp_done_with_error() */
609 smp_rmb();
610 if (READ_ONCE(sk->sk_err) ||
611 !skb_queue_empty_lockless(&sk->sk_error_queue))
612 mask |= EPOLLERR;
613
614 return mask;
615}
616EXPORT_SYMBOL(tcp_poll);
617
618int tcp_ioctl(struct sock *sk, int cmd, int *karg)
619{
620 struct tcp_sock *tp = tcp_sk(sk);
621 int answ;
622 bool slow;
623
624 switch (cmd) {
625 case SIOCINQ:
626 if (sk->sk_state == TCP_LISTEN)
627 return -EINVAL;
628
629 slow = lock_sock_fast(sk);
630 answ = tcp_inq(sk);
631 unlock_sock_fast(sk, slow);
632 break;
633 case SIOCATMARK:
634 answ = READ_ONCE(tp->urg_data) &&
635 READ_ONCE(tp->urg_seq) == READ_ONCE(tp->copied_seq);
636 break;
637 case SIOCOUTQ:
638 if (sk->sk_state == TCP_LISTEN)
639 return -EINVAL;
640
641 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
642 answ = 0;
643 else
644 answ = READ_ONCE(tp->write_seq) - tp->snd_una;
645 break;
646 case SIOCOUTQNSD:
647 if (sk->sk_state == TCP_LISTEN)
648 return -EINVAL;
649
650 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
651 answ = 0;
652 else
653 answ = READ_ONCE(tp->write_seq) -
654 READ_ONCE(tp->snd_nxt);
655 break;
656 default:
657 return -ENOIOCTLCMD;
658 }
659
660 *karg = answ;
661 return 0;
662}
663EXPORT_SYMBOL(tcp_ioctl);
664
665void tcp_mark_push(struct tcp_sock *tp, struct sk_buff *skb)
666{
667 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
668 tp->pushed_seq = tp->write_seq;
669}
670
671static inline bool forced_push(const struct tcp_sock *tp)
672{
673 return after(tp->write_seq, tp->pushed_seq + (tp->max_window >> 1));
674}
675
676void tcp_skb_entail(struct sock *sk, struct sk_buff *skb)
677{
678 struct tcp_sock *tp = tcp_sk(sk);
679 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
680
681 tcb->seq = tcb->end_seq = tp->write_seq;
682 tcb->tcp_flags = TCPHDR_ACK;
683 __skb_header_release(skb);
684 tcp_add_write_queue_tail(sk, skb);
685 sk_wmem_queued_add(sk, skb->truesize);
686 sk_mem_charge(sk, skb->truesize);
687 if (tp->nonagle & TCP_NAGLE_PUSH)
688 tp->nonagle &= ~TCP_NAGLE_PUSH;
689
690 tcp_slow_start_after_idle_check(sk);
691}
692
693static inline void tcp_mark_urg(struct tcp_sock *tp, int flags)
694{
695 if (flags & MSG_OOB)
696 tp->snd_up = tp->write_seq;
697}
698
699/* If a not yet filled skb is pushed, do not send it if
700 * we have data packets in Qdisc or NIC queues :
701 * Because TX completion will happen shortly, it gives a chance
702 * to coalesce future sendmsg() payload into this skb, without
703 * need for a timer, and with no latency trade off.
704 * As packets containing data payload have a bigger truesize
705 * than pure acks (dataless) packets, the last checks prevent
706 * autocorking if we only have an ACK in Qdisc/NIC queues,
707 * or if TX completion was delayed after we processed ACK packet.
708 */
709static bool tcp_should_autocork(struct sock *sk, struct sk_buff *skb,
710 int size_goal)
711{
712 return skb->len < size_goal &&
713 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_autocorking) &&
714 !tcp_rtx_queue_empty(sk) &&
715 refcount_read(&sk->sk_wmem_alloc) > skb->truesize &&
716 tcp_skb_can_collapse_to(skb);
717}
718
719void tcp_push(struct sock *sk, int flags, int mss_now,
720 int nonagle, int size_goal)
721{
722 struct tcp_sock *tp = tcp_sk(sk);
723 struct sk_buff *skb;
724
725 skb = tcp_write_queue_tail(sk);
726 if (!skb)
727 return;
728 if (!(flags & MSG_MORE) || forced_push(tp))
729 tcp_mark_push(tp, skb);
730
731 tcp_mark_urg(tp, flags);
732
733 if (tcp_should_autocork(sk, skb, size_goal)) {
734
735 /* avoid atomic op if TSQ_THROTTLED bit is already set */
736 if (!test_bit(TSQ_THROTTLED, &sk->sk_tsq_flags)) {
737 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPAUTOCORKING);
738 set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags);
739 smp_mb__after_atomic();
740 }
741 /* It is possible TX completion already happened
742 * before we set TSQ_THROTTLED.
743 */
744 if (refcount_read(&sk->sk_wmem_alloc) > skb->truesize)
745 return;
746 }
747
748 if (flags & MSG_MORE)
749 nonagle = TCP_NAGLE_CORK;
750
751 __tcp_push_pending_frames(sk, mss_now, nonagle);
752}
753
754static int tcp_splice_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb,
755 unsigned int offset, size_t len)
756{
757 struct tcp_splice_state *tss = rd_desc->arg.data;
758 int ret;
759
760 ret = skb_splice_bits(skb, skb->sk, offset, tss->pipe,
761 min(rd_desc->count, len), tss->flags);
762 if (ret > 0)
763 rd_desc->count -= ret;
764 return ret;
765}
766
767static int __tcp_splice_read(struct sock *sk, struct tcp_splice_state *tss)
768{
769 /* Store TCP splice context information in read_descriptor_t. */
770 read_descriptor_t rd_desc = {
771 .arg.data = tss,
772 .count = tss->len,
773 };
774
775 return tcp_read_sock(sk, &rd_desc, tcp_splice_data_recv);
776}
777
778/**
779 * tcp_splice_read - splice data from TCP socket to a pipe
780 * @sock: socket to splice from
781 * @ppos: position (not valid)
782 * @pipe: pipe to splice to
783 * @len: number of bytes to splice
784 * @flags: splice modifier flags
785 *
786 * Description:
787 * Will read pages from given socket and fill them into a pipe.
788 *
789 **/
790ssize_t tcp_splice_read(struct socket *sock, loff_t *ppos,
791 struct pipe_inode_info *pipe, size_t len,
792 unsigned int flags)
793{
794 struct sock *sk = sock->sk;
795 struct tcp_splice_state tss = {
796 .pipe = pipe,
797 .len = len,
798 .flags = flags,
799 };
800 long timeo;
801 ssize_t spliced;
802 int ret;
803
804 sock_rps_record_flow(sk);
805 /*
806 * We can't seek on a socket input
807 */
808 if (unlikely(*ppos))
809 return -ESPIPE;
810
811 ret = spliced = 0;
812
813 lock_sock(sk);
814
815 timeo = sock_rcvtimeo(sk, sock->file->f_flags & O_NONBLOCK);
816 while (tss.len) {
817 ret = __tcp_splice_read(sk, &tss);
818 if (ret < 0)
819 break;
820 else if (!ret) {
821 if (spliced)
822 break;
823 if (sock_flag(sk, SOCK_DONE))
824 break;
825 if (sk->sk_err) {
826 ret = sock_error(sk);
827 break;
828 }
829 if (sk->sk_shutdown & RCV_SHUTDOWN)
830 break;
831 if (sk->sk_state == TCP_CLOSE) {
832 /*
833 * This occurs when user tries to read
834 * from never connected socket.
835 */
836 ret = -ENOTCONN;
837 break;
838 }
839 if (!timeo) {
840 ret = -EAGAIN;
841 break;
842 }
843 /* if __tcp_splice_read() got nothing while we have
844 * an skb in receive queue, we do not want to loop.
845 * This might happen with URG data.
846 */
847 if (!skb_queue_empty(&sk->sk_receive_queue))
848 break;
849 ret = sk_wait_data(sk, &timeo, NULL);
850 if (ret < 0)
851 break;
852 if (signal_pending(current)) {
853 ret = sock_intr_errno(timeo);
854 break;
855 }
856 continue;
857 }
858 tss.len -= ret;
859 spliced += ret;
860
861 if (!tss.len || !timeo)
862 break;
863 release_sock(sk);
864 lock_sock(sk);
865
866 if (sk->sk_err || sk->sk_state == TCP_CLOSE ||
867 (sk->sk_shutdown & RCV_SHUTDOWN) ||
868 signal_pending(current))
869 break;
870 }
871
872 release_sock(sk);
873
874 if (spliced)
875 return spliced;
876
877 return ret;
878}
879EXPORT_SYMBOL(tcp_splice_read);
880
881struct sk_buff *tcp_stream_alloc_skb(struct sock *sk, gfp_t gfp,
882 bool force_schedule)
883{
884 struct sk_buff *skb;
885
886 skb = alloc_skb_fclone(MAX_TCP_HEADER, gfp);
887 if (likely(skb)) {
888 bool mem_scheduled;
889
890 skb->truesize = SKB_TRUESIZE(skb_end_offset(skb));
891 if (force_schedule) {
892 mem_scheduled = true;
893 sk_forced_mem_schedule(sk, skb->truesize);
894 } else {
895 mem_scheduled = sk_wmem_schedule(sk, skb->truesize);
896 }
897 if (likely(mem_scheduled)) {
898 skb_reserve(skb, MAX_TCP_HEADER);
899 skb->ip_summed = CHECKSUM_PARTIAL;
900 INIT_LIST_HEAD(&skb->tcp_tsorted_anchor);
901 return skb;
902 }
903 __kfree_skb(skb);
904 } else {
905 sk->sk_prot->enter_memory_pressure(sk);
906 sk_stream_moderate_sndbuf(sk);
907 }
908 return NULL;
909}
910
911static unsigned int tcp_xmit_size_goal(struct sock *sk, u32 mss_now,
912 int large_allowed)
913{
914 struct tcp_sock *tp = tcp_sk(sk);
915 u32 new_size_goal, size_goal;
916
917 if (!large_allowed)
918 return mss_now;
919
920 /* Note : tcp_tso_autosize() will eventually split this later */
921 new_size_goal = tcp_bound_to_half_wnd(tp, sk->sk_gso_max_size);
922
923 /* We try hard to avoid divides here */
924 size_goal = tp->gso_segs * mss_now;
925 if (unlikely(new_size_goal < size_goal ||
926 new_size_goal >= size_goal + mss_now)) {
927 tp->gso_segs = min_t(u16, new_size_goal / mss_now,
928 sk->sk_gso_max_segs);
929 size_goal = tp->gso_segs * mss_now;
930 }
931
932 return max(size_goal, mss_now);
933}
934
935int tcp_send_mss(struct sock *sk, int *size_goal, int flags)
936{
937 int mss_now;
938
939 mss_now = tcp_current_mss(sk);
940 *size_goal = tcp_xmit_size_goal(sk, mss_now, !(flags & MSG_OOB));
941
942 return mss_now;
943}
944
945/* In some cases, sendmsg() could have added an skb to the write queue,
946 * but failed adding payload on it. We need to remove it to consume less
947 * memory, but more importantly be able to generate EPOLLOUT for Edge Trigger
948 * epoll() users. Another reason is that tcp_write_xmit() does not like
949 * finding an empty skb in the write queue.
950 */
951void tcp_remove_empty_skb(struct sock *sk)
952{
953 struct sk_buff *skb = tcp_write_queue_tail(sk);
954
955 if (skb && TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq) {
956 tcp_unlink_write_queue(skb, sk);
957 if (tcp_write_queue_empty(sk))
958 tcp_chrono_stop(sk, TCP_CHRONO_BUSY);
959 tcp_wmem_free_skb(sk, skb);
960 }
961}
962
963/* skb changing from pure zc to mixed, must charge zc */
964static int tcp_downgrade_zcopy_pure(struct sock *sk, struct sk_buff *skb)
965{
966 if (unlikely(skb_zcopy_pure(skb))) {
967 u32 extra = skb->truesize -
968 SKB_TRUESIZE(skb_end_offset(skb));
969
970 if (!sk_wmem_schedule(sk, extra))
971 return -ENOMEM;
972
973 sk_mem_charge(sk, extra);
974 skb_shinfo(skb)->flags &= ~SKBFL_PURE_ZEROCOPY;
975 }
976 return 0;
977}
978
979
980int tcp_wmem_schedule(struct sock *sk, int copy)
981{
982 int left;
983
984 if (likely(sk_wmem_schedule(sk, copy)))
985 return copy;
986
987 /* We could be in trouble if we have nothing queued.
988 * Use whatever is left in sk->sk_forward_alloc and tcp_wmem[0]
989 * to guarantee some progress.
990 */
991 left = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_wmem[0]) - sk->sk_wmem_queued;
992 if (left > 0)
993 sk_forced_mem_schedule(sk, min(left, copy));
994 return min(copy, sk->sk_forward_alloc);
995}
996
997void tcp_free_fastopen_req(struct tcp_sock *tp)
998{
999 if (tp->fastopen_req) {
1000 kfree(tp->fastopen_req);
1001 tp->fastopen_req = NULL;
1002 }
1003}
1004
1005int tcp_sendmsg_fastopen(struct sock *sk, struct msghdr *msg, int *copied,
1006 size_t size, struct ubuf_info *uarg)
1007{
1008 struct tcp_sock *tp = tcp_sk(sk);
1009 struct inet_sock *inet = inet_sk(sk);
1010 struct sockaddr *uaddr = msg->msg_name;
1011 int err, flags;
1012
1013 if (!(READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_fastopen) &
1014 TFO_CLIENT_ENABLE) ||
1015 (uaddr && msg->msg_namelen >= sizeof(uaddr->sa_family) &&
1016 uaddr->sa_family == AF_UNSPEC))
1017 return -EOPNOTSUPP;
1018 if (tp->fastopen_req)
1019 return -EALREADY; /* Another Fast Open is in progress */
1020
1021 tp->fastopen_req = kzalloc(sizeof(struct tcp_fastopen_request),
1022 sk->sk_allocation);
1023 if (unlikely(!tp->fastopen_req))
1024 return -ENOBUFS;
1025 tp->fastopen_req->data = msg;
1026 tp->fastopen_req->size = size;
1027 tp->fastopen_req->uarg = uarg;
1028
1029 if (inet_test_bit(DEFER_CONNECT, sk)) {
1030 err = tcp_connect(sk);
1031 /* Same failure procedure as in tcp_v4/6_connect */
1032 if (err) {
1033 tcp_set_state(sk, TCP_CLOSE);
1034 inet->inet_dport = 0;
1035 sk->sk_route_caps = 0;
1036 }
1037 }
1038 flags = (msg->msg_flags & MSG_DONTWAIT) ? O_NONBLOCK : 0;
1039 err = __inet_stream_connect(sk->sk_socket, uaddr,
1040 msg->msg_namelen, flags, 1);
1041 /* fastopen_req could already be freed in __inet_stream_connect
1042 * if the connection times out or gets rst
1043 */
1044 if (tp->fastopen_req) {
1045 *copied = tp->fastopen_req->copied;
1046 tcp_free_fastopen_req(tp);
1047 inet_clear_bit(DEFER_CONNECT, sk);
1048 }
1049 return err;
1050}
1051
1052int tcp_sendmsg_locked(struct sock *sk, struct msghdr *msg, size_t size)
1053{
1054 struct tcp_sock *tp = tcp_sk(sk);
1055 struct ubuf_info *uarg = NULL;
1056 struct sk_buff *skb;
1057 struct sockcm_cookie sockc;
1058 int flags, err, copied = 0;
1059 int mss_now = 0, size_goal, copied_syn = 0;
1060 int process_backlog = 0;
1061 int zc = 0;
1062 long timeo;
1063
1064 flags = msg->msg_flags;
1065
1066 if ((flags & MSG_ZEROCOPY) && size) {
1067 if (msg->msg_ubuf) {
1068 uarg = msg->msg_ubuf;
1069 if (sk->sk_route_caps & NETIF_F_SG)
1070 zc = MSG_ZEROCOPY;
1071 } else if (sock_flag(sk, SOCK_ZEROCOPY)) {
1072 skb = tcp_write_queue_tail(sk);
1073 uarg = msg_zerocopy_realloc(sk, size, skb_zcopy(skb));
1074 if (!uarg) {
1075 err = -ENOBUFS;
1076 goto out_err;
1077 }
1078 if (sk->sk_route_caps & NETIF_F_SG)
1079 zc = MSG_ZEROCOPY;
1080 else
1081 uarg_to_msgzc(uarg)->zerocopy = 0;
1082 }
1083 } else if (unlikely(msg->msg_flags & MSG_SPLICE_PAGES) && size) {
1084 if (sk->sk_route_caps & NETIF_F_SG)
1085 zc = MSG_SPLICE_PAGES;
1086 }
1087
1088 if (unlikely(flags & MSG_FASTOPEN ||
1089 inet_test_bit(DEFER_CONNECT, sk)) &&
1090 !tp->repair) {
1091 err = tcp_sendmsg_fastopen(sk, msg, &copied_syn, size, uarg);
1092 if (err == -EINPROGRESS && copied_syn > 0)
1093 goto out;
1094 else if (err)
1095 goto out_err;
1096 }
1097
1098 timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
1099
1100 tcp_rate_check_app_limited(sk); /* is sending application-limited? */
1101
1102 /* Wait for a connection to finish. One exception is TCP Fast Open
1103 * (passive side) where data is allowed to be sent before a connection
1104 * is fully established.
1105 */
1106 if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) &&
1107 !tcp_passive_fastopen(sk)) {
1108 err = sk_stream_wait_connect(sk, &timeo);
1109 if (err != 0)
1110 goto do_error;
1111 }
1112
1113 if (unlikely(tp->repair)) {
1114 if (tp->repair_queue == TCP_RECV_QUEUE) {
1115 copied = tcp_send_rcvq(sk, msg, size);
1116 goto out_nopush;
1117 }
1118
1119 err = -EINVAL;
1120 if (tp->repair_queue == TCP_NO_QUEUE)
1121 goto out_err;
1122
1123 /* 'common' sending to sendq */
1124 }
1125
1126 sockcm_init(&sockc, sk);
1127 if (msg->msg_controllen) {
1128 err = sock_cmsg_send(sk, msg, &sockc);
1129 if (unlikely(err)) {
1130 err = -EINVAL;
1131 goto out_err;
1132 }
1133 }
1134
1135 /* This should be in poll */
1136 sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
1137
1138 /* Ok commence sending. */
1139 copied = 0;
1140
1141restart:
1142 mss_now = tcp_send_mss(sk, &size_goal, flags);
1143
1144 err = -EPIPE;
1145 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
1146 goto do_error;
1147
1148 while (msg_data_left(msg)) {
1149 ssize_t copy = 0;
1150
1151 skb = tcp_write_queue_tail(sk);
1152 if (skb)
1153 copy = size_goal - skb->len;
1154
1155 if (copy <= 0 || !tcp_skb_can_collapse_to(skb)) {
1156 bool first_skb;
1157
1158new_segment:
1159 if (!sk_stream_memory_free(sk))
1160 goto wait_for_space;
1161
1162 if (unlikely(process_backlog >= 16)) {
1163 process_backlog = 0;
1164 if (sk_flush_backlog(sk))
1165 goto restart;
1166 }
1167 first_skb = tcp_rtx_and_write_queues_empty(sk);
1168 skb = tcp_stream_alloc_skb(sk, sk->sk_allocation,
1169 first_skb);
1170 if (!skb)
1171 goto wait_for_space;
1172
1173 process_backlog++;
1174
1175#ifdef CONFIG_SKB_DECRYPTED
1176 skb->decrypted = !!(flags & MSG_SENDPAGE_DECRYPTED);
1177#endif
1178 tcp_skb_entail(sk, skb);
1179 copy = size_goal;
1180
1181 /* All packets are restored as if they have
1182 * already been sent. skb_mstamp_ns isn't set to
1183 * avoid wrong rtt estimation.
1184 */
1185 if (tp->repair)
1186 TCP_SKB_CB(skb)->sacked |= TCPCB_REPAIRED;
1187 }
1188
1189 /* Try to append data to the end of skb. */
1190 if (copy > msg_data_left(msg))
1191 copy = msg_data_left(msg);
1192
1193 if (zc == 0) {
1194 bool merge = true;
1195 int i = skb_shinfo(skb)->nr_frags;
1196 struct page_frag *pfrag = sk_page_frag(sk);
1197
1198 if (!sk_page_frag_refill(sk, pfrag))
1199 goto wait_for_space;
1200
1201 if (!skb_can_coalesce(skb, i, pfrag->page,
1202 pfrag->offset)) {
1203 if (i >= READ_ONCE(net_hotdata.sysctl_max_skb_frags)) {
1204 tcp_mark_push(tp, skb);
1205 goto new_segment;
1206 }
1207 merge = false;
1208 }
1209
1210 copy = min_t(int, copy, pfrag->size - pfrag->offset);
1211
1212 if (unlikely(skb_zcopy_pure(skb) || skb_zcopy_managed(skb))) {
1213 if (tcp_downgrade_zcopy_pure(sk, skb))
1214 goto wait_for_space;
1215 skb_zcopy_downgrade_managed(skb);
1216 }
1217
1218 copy = tcp_wmem_schedule(sk, copy);
1219 if (!copy)
1220 goto wait_for_space;
1221
1222 err = skb_copy_to_page_nocache(sk, &msg->msg_iter, skb,
1223 pfrag->page,
1224 pfrag->offset,
1225 copy);
1226 if (err)
1227 goto do_error;
1228
1229 /* Update the skb. */
1230 if (merge) {
1231 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1232 } else {
1233 skb_fill_page_desc(skb, i, pfrag->page,
1234 pfrag->offset, copy);
1235 page_ref_inc(pfrag->page);
1236 }
1237 pfrag->offset += copy;
1238 } else if (zc == MSG_ZEROCOPY) {
1239 /* First append to a fragless skb builds initial
1240 * pure zerocopy skb
1241 */
1242 if (!skb->len)
1243 skb_shinfo(skb)->flags |= SKBFL_PURE_ZEROCOPY;
1244
1245 if (!skb_zcopy_pure(skb)) {
1246 copy = tcp_wmem_schedule(sk, copy);
1247 if (!copy)
1248 goto wait_for_space;
1249 }
1250
1251 err = skb_zerocopy_iter_stream(sk, skb, msg, copy, uarg);
1252 if (err == -EMSGSIZE || err == -EEXIST) {
1253 tcp_mark_push(tp, skb);
1254 goto new_segment;
1255 }
1256 if (err < 0)
1257 goto do_error;
1258 copy = err;
1259 } else if (zc == MSG_SPLICE_PAGES) {
1260 /* Splice in data if we can; copy if we can't. */
1261 if (tcp_downgrade_zcopy_pure(sk, skb))
1262 goto wait_for_space;
1263 copy = tcp_wmem_schedule(sk, copy);
1264 if (!copy)
1265 goto wait_for_space;
1266
1267 err = skb_splice_from_iter(skb, &msg->msg_iter, copy,
1268 sk->sk_allocation);
1269 if (err < 0) {
1270 if (err == -EMSGSIZE) {
1271 tcp_mark_push(tp, skb);
1272 goto new_segment;
1273 }
1274 goto do_error;
1275 }
1276 copy = err;
1277
1278 if (!(flags & MSG_NO_SHARED_FRAGS))
1279 skb_shinfo(skb)->flags |= SKBFL_SHARED_FRAG;
1280
1281 sk_wmem_queued_add(sk, copy);
1282 sk_mem_charge(sk, copy);
1283 }
1284
1285 if (!copied)
1286 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
1287
1288 WRITE_ONCE(tp->write_seq, tp->write_seq + copy);
1289 TCP_SKB_CB(skb)->end_seq += copy;
1290 tcp_skb_pcount_set(skb, 0);
1291
1292 copied += copy;
1293 if (!msg_data_left(msg)) {
1294 if (unlikely(flags & MSG_EOR))
1295 TCP_SKB_CB(skb)->eor = 1;
1296 goto out;
1297 }
1298
1299 if (skb->len < size_goal || (flags & MSG_OOB) || unlikely(tp->repair))
1300 continue;
1301
1302 if (forced_push(tp)) {
1303 tcp_mark_push(tp, skb);
1304 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
1305 } else if (skb == tcp_send_head(sk))
1306 tcp_push_one(sk, mss_now);
1307 continue;
1308
1309wait_for_space:
1310 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1311 tcp_remove_empty_skb(sk);
1312 if (copied)
1313 tcp_push(sk, flags & ~MSG_MORE, mss_now,
1314 TCP_NAGLE_PUSH, size_goal);
1315
1316 err = sk_stream_wait_memory(sk, &timeo);
1317 if (err != 0)
1318 goto do_error;
1319
1320 mss_now = tcp_send_mss(sk, &size_goal, flags);
1321 }
1322
1323out:
1324 if (copied) {
1325 tcp_tx_timestamp(sk, &sockc);
1326 tcp_push(sk, flags, mss_now, tp->nonagle, size_goal);
1327 }
1328out_nopush:
1329 /* msg->msg_ubuf is pinned by the caller so we don't take extra refs */
1330 if (uarg && !msg->msg_ubuf)
1331 net_zcopy_put(uarg);
1332 return copied + copied_syn;
1333
1334do_error:
1335 tcp_remove_empty_skb(sk);
1336
1337 if (copied + copied_syn)
1338 goto out;
1339out_err:
1340 /* msg->msg_ubuf is pinned by the caller so we don't take extra refs */
1341 if (uarg && !msg->msg_ubuf)
1342 net_zcopy_put_abort(uarg, true);
1343 err = sk_stream_error(sk, flags, err);
1344 /* make sure we wake any epoll edge trigger waiter */
1345 if (unlikely(tcp_rtx_and_write_queues_empty(sk) && err == -EAGAIN)) {
1346 sk->sk_write_space(sk);
1347 tcp_chrono_stop(sk, TCP_CHRONO_SNDBUF_LIMITED);
1348 }
1349 return err;
1350}
1351EXPORT_SYMBOL_GPL(tcp_sendmsg_locked);
1352
1353int tcp_sendmsg(struct sock *sk, struct msghdr *msg, size_t size)
1354{
1355 int ret;
1356
1357 lock_sock(sk);
1358 ret = tcp_sendmsg_locked(sk, msg, size);
1359 release_sock(sk);
1360
1361 return ret;
1362}
1363EXPORT_SYMBOL(tcp_sendmsg);
1364
1365void tcp_splice_eof(struct socket *sock)
1366{
1367 struct sock *sk = sock->sk;
1368 struct tcp_sock *tp = tcp_sk(sk);
1369 int mss_now, size_goal;
1370
1371 if (!tcp_write_queue_tail(sk))
1372 return;
1373
1374 lock_sock(sk);
1375 mss_now = tcp_send_mss(sk, &size_goal, 0);
1376 tcp_push(sk, 0, mss_now, tp->nonagle, size_goal);
1377 release_sock(sk);
1378}
1379EXPORT_SYMBOL_GPL(tcp_splice_eof);
1380
1381/*
1382 * Handle reading urgent data. BSD has very simple semantics for
1383 * this, no blocking and very strange errors 8)
1384 */
1385
1386static int tcp_recv_urg(struct sock *sk, struct msghdr *msg, int len, int flags)
1387{
1388 struct tcp_sock *tp = tcp_sk(sk);
1389
1390 /* No URG data to read. */
1391 if (sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data ||
1392 tp->urg_data == TCP_URG_READ)
1393 return -EINVAL; /* Yes this is right ! */
1394
1395 if (sk->sk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DONE))
1396 return -ENOTCONN;
1397
1398 if (tp->urg_data & TCP_URG_VALID) {
1399 int err = 0;
1400 char c = tp->urg_data;
1401
1402 if (!(flags & MSG_PEEK))
1403 WRITE_ONCE(tp->urg_data, TCP_URG_READ);
1404
1405 /* Read urgent data. */
1406 msg->msg_flags |= MSG_OOB;
1407
1408 if (len > 0) {
1409 if (!(flags & MSG_TRUNC))
1410 err = memcpy_to_msg(msg, &c, 1);
1411 len = 1;
1412 } else
1413 msg->msg_flags |= MSG_TRUNC;
1414
1415 return err ? -EFAULT : len;
1416 }
1417
1418 if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
1419 return 0;
1420
1421 /* Fixed the recv(..., MSG_OOB) behaviour. BSD docs and
1422 * the available implementations agree in this case:
1423 * this call should never block, independent of the
1424 * blocking state of the socket.
1425 * Mike <pall@rz.uni-karlsruhe.de>
1426 */
1427 return -EAGAIN;
1428}
1429
1430static int tcp_peek_sndq(struct sock *sk, struct msghdr *msg, int len)
1431{
1432 struct sk_buff *skb;
1433 int copied = 0, err = 0;
1434
1435 skb_rbtree_walk(skb, &sk->tcp_rtx_queue) {
1436 err = skb_copy_datagram_msg(skb, 0, msg, skb->len);
1437 if (err)
1438 return err;
1439 copied += skb->len;
1440 }
1441
1442 skb_queue_walk(&sk->sk_write_queue, skb) {
1443 err = skb_copy_datagram_msg(skb, 0, msg, skb->len);
1444 if (err)
1445 break;
1446
1447 copied += skb->len;
1448 }
1449
1450 return err ?: copied;
1451}
1452
1453/* Clean up the receive buffer for full frames taken by the user,
1454 * then send an ACK if necessary. COPIED is the number of bytes
1455 * tcp_recvmsg has given to the user so far, it speeds up the
1456 * calculation of whether or not we must ACK for the sake of
1457 * a window update.
1458 */
1459void __tcp_cleanup_rbuf(struct sock *sk, int copied)
1460{
1461 struct tcp_sock *tp = tcp_sk(sk);
1462 bool time_to_ack = false;
1463
1464 if (inet_csk_ack_scheduled(sk)) {
1465 const struct inet_connection_sock *icsk = inet_csk(sk);
1466
1467 if (/* Once-per-two-segments ACK was not sent by tcp_input.c */
1468 tp->rcv_nxt - tp->rcv_wup > icsk->icsk_ack.rcv_mss ||
1469 /*
1470 * If this read emptied read buffer, we send ACK, if
1471 * connection is not bidirectional, user drained
1472 * receive buffer and there was a small segment
1473 * in queue.
1474 */
1475 (copied > 0 &&
1476 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED2) ||
1477 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED) &&
1478 !inet_csk_in_pingpong_mode(sk))) &&
1479 !atomic_read(&sk->sk_rmem_alloc)))
1480 time_to_ack = true;
1481 }
1482
1483 /* We send an ACK if we can now advertise a non-zero window
1484 * which has been raised "significantly".
1485 *
1486 * Even if window raised up to infinity, do not send window open ACK
1487 * in states, where we will not receive more. It is useless.
1488 */
1489 if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) {
1490 __u32 rcv_window_now = tcp_receive_window(tp);
1491
1492 /* Optimize, __tcp_select_window() is not cheap. */
1493 if (2*rcv_window_now <= tp->window_clamp) {
1494 __u32 new_window = __tcp_select_window(sk);
1495
1496 /* Send ACK now, if this read freed lots of space
1497 * in our buffer. Certainly, new_window is new window.
1498 * We can advertise it now, if it is not less than current one.
1499 * "Lots" means "at least twice" here.
1500 */
1501 if (new_window && new_window >= 2 * rcv_window_now)
1502 time_to_ack = true;
1503 }
1504 }
1505 if (time_to_ack)
1506 tcp_send_ack(sk);
1507}
1508
1509void tcp_cleanup_rbuf(struct sock *sk, int copied)
1510{
1511 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
1512 struct tcp_sock *tp = tcp_sk(sk);
1513
1514 WARN(skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq),
1515 "cleanup rbuf bug: copied %X seq %X rcvnxt %X\n",
1516 tp->copied_seq, TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt);
1517 __tcp_cleanup_rbuf(sk, copied);
1518}
1519
1520static void tcp_eat_recv_skb(struct sock *sk, struct sk_buff *skb)
1521{
1522 __skb_unlink(skb, &sk->sk_receive_queue);
1523 if (likely(skb->destructor == sock_rfree)) {
1524 sock_rfree(skb);
1525 skb->destructor = NULL;
1526 skb->sk = NULL;
1527 return skb_attempt_defer_free(skb);
1528 }
1529 __kfree_skb(skb);
1530}
1531
1532struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off)
1533{
1534 struct sk_buff *skb;
1535 u32 offset;
1536
1537 while ((skb = skb_peek(&sk->sk_receive_queue)) != NULL) {
1538 offset = seq - TCP_SKB_CB(skb)->seq;
1539 if (unlikely(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
1540 pr_err_once("%s: found a SYN, please report !\n", __func__);
1541 offset--;
1542 }
1543 if (offset < skb->len || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)) {
1544 *off = offset;
1545 return skb;
1546 }
1547 /* This looks weird, but this can happen if TCP collapsing
1548 * splitted a fat GRO packet, while we released socket lock
1549 * in skb_splice_bits()
1550 */
1551 tcp_eat_recv_skb(sk, skb);
1552 }
1553 return NULL;
1554}
1555EXPORT_SYMBOL(tcp_recv_skb);
1556
1557/*
1558 * This routine provides an alternative to tcp_recvmsg() for routines
1559 * that would like to handle copying from skbuffs directly in 'sendfile'
1560 * fashion.
1561 * Note:
1562 * - It is assumed that the socket was locked by the caller.
1563 * - The routine does not block.
1564 * - At present, there is no support for reading OOB data
1565 * or for 'peeking' the socket using this routine
1566 * (although both would be easy to implement).
1567 */
1568int tcp_read_sock(struct sock *sk, read_descriptor_t *desc,
1569 sk_read_actor_t recv_actor)
1570{
1571 struct sk_buff *skb;
1572 struct tcp_sock *tp = tcp_sk(sk);
1573 u32 seq = tp->copied_seq;
1574 u32 offset;
1575 int copied = 0;
1576
1577 if (sk->sk_state == TCP_LISTEN)
1578 return -ENOTCONN;
1579 while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1580 if (offset < skb->len) {
1581 int used;
1582 size_t len;
1583
1584 len = skb->len - offset;
1585 /* Stop reading if we hit a patch of urgent data */
1586 if (unlikely(tp->urg_data)) {
1587 u32 urg_offset = tp->urg_seq - seq;
1588 if (urg_offset < len)
1589 len = urg_offset;
1590 if (!len)
1591 break;
1592 }
1593 used = recv_actor(desc, skb, offset, len);
1594 if (used <= 0) {
1595 if (!copied)
1596 copied = used;
1597 break;
1598 }
1599 if (WARN_ON_ONCE(used > len))
1600 used = len;
1601 seq += used;
1602 copied += used;
1603 offset += used;
1604
1605 /* If recv_actor drops the lock (e.g. TCP splice
1606 * receive) the skb pointer might be invalid when
1607 * getting here: tcp_collapse might have deleted it
1608 * while aggregating skbs from the socket queue.
1609 */
1610 skb = tcp_recv_skb(sk, seq - 1, &offset);
1611 if (!skb)
1612 break;
1613 /* TCP coalescing might have appended data to the skb.
1614 * Try to splice more frags
1615 */
1616 if (offset + 1 != skb->len)
1617 continue;
1618 }
1619 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) {
1620 tcp_eat_recv_skb(sk, skb);
1621 ++seq;
1622 break;
1623 }
1624 tcp_eat_recv_skb(sk, skb);
1625 if (!desc->count)
1626 break;
1627 WRITE_ONCE(tp->copied_seq, seq);
1628 }
1629 WRITE_ONCE(tp->copied_seq, seq);
1630
1631 tcp_rcv_space_adjust(sk);
1632
1633 /* Clean up data we have read: This will do ACK frames. */
1634 if (copied > 0) {
1635 tcp_recv_skb(sk, seq, &offset);
1636 tcp_cleanup_rbuf(sk, copied);
1637 }
1638 return copied;
1639}
1640EXPORT_SYMBOL(tcp_read_sock);
1641
1642int tcp_read_skb(struct sock *sk, skb_read_actor_t recv_actor)
1643{
1644 struct sk_buff *skb;
1645 int copied = 0;
1646
1647 if (sk->sk_state == TCP_LISTEN)
1648 return -ENOTCONN;
1649
1650 while ((skb = skb_peek(&sk->sk_receive_queue)) != NULL) {
1651 u8 tcp_flags;
1652 int used;
1653
1654 __skb_unlink(skb, &sk->sk_receive_queue);
1655 WARN_ON_ONCE(!skb_set_owner_sk_safe(skb, sk));
1656 tcp_flags = TCP_SKB_CB(skb)->tcp_flags;
1657 used = recv_actor(sk, skb);
1658 if (used < 0) {
1659 if (!copied)
1660 copied = used;
1661 break;
1662 }
1663 copied += used;
1664
1665 if (tcp_flags & TCPHDR_FIN)
1666 break;
1667 }
1668 return copied;
1669}
1670EXPORT_SYMBOL(tcp_read_skb);
1671
1672void tcp_read_done(struct sock *sk, size_t len)
1673{
1674 struct tcp_sock *tp = tcp_sk(sk);
1675 u32 seq = tp->copied_seq;
1676 struct sk_buff *skb;
1677 size_t left;
1678 u32 offset;
1679
1680 if (sk->sk_state == TCP_LISTEN)
1681 return;
1682
1683 left = len;
1684 while (left && (skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1685 int used;
1686
1687 used = min_t(size_t, skb->len - offset, left);
1688 seq += used;
1689 left -= used;
1690
1691 if (skb->len > offset + used)
1692 break;
1693
1694 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) {
1695 tcp_eat_recv_skb(sk, skb);
1696 ++seq;
1697 break;
1698 }
1699 tcp_eat_recv_skb(sk, skb);
1700 }
1701 WRITE_ONCE(tp->copied_seq, seq);
1702
1703 tcp_rcv_space_adjust(sk);
1704
1705 /* Clean up data we have read: This will do ACK frames. */
1706 if (left != len)
1707 tcp_cleanup_rbuf(sk, len - left);
1708}
1709EXPORT_SYMBOL(tcp_read_done);
1710
1711int tcp_peek_len(struct socket *sock)
1712{
1713 return tcp_inq(sock->sk);
1714}
1715EXPORT_SYMBOL(tcp_peek_len);
1716
1717/* Make sure sk_rcvbuf is big enough to satisfy SO_RCVLOWAT hint */
1718int tcp_set_rcvlowat(struct sock *sk, int val)
1719{
1720 int space, cap;
1721
1722 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK)
1723 cap = sk->sk_rcvbuf >> 1;
1724 else
1725 cap = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_rmem[2]) >> 1;
1726 val = min(val, cap);
1727 WRITE_ONCE(sk->sk_rcvlowat, val ? : 1);
1728
1729 /* Check if we need to signal EPOLLIN right now */
1730 tcp_data_ready(sk);
1731
1732 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK)
1733 return 0;
1734
1735 space = tcp_space_from_win(sk, val);
1736 if (space > sk->sk_rcvbuf) {
1737 WRITE_ONCE(sk->sk_rcvbuf, space);
1738 WRITE_ONCE(tcp_sk(sk)->window_clamp, val);
1739 }
1740 return 0;
1741}
1742EXPORT_SYMBOL(tcp_set_rcvlowat);
1743
1744void tcp_update_recv_tstamps(struct sk_buff *skb,
1745 struct scm_timestamping_internal *tss)
1746{
1747 if (skb->tstamp)
1748 tss->ts[0] = ktime_to_timespec64(skb->tstamp);
1749 else
1750 tss->ts[0] = (struct timespec64) {0};
1751
1752 if (skb_hwtstamps(skb)->hwtstamp)
1753 tss->ts[2] = ktime_to_timespec64(skb_hwtstamps(skb)->hwtstamp);
1754 else
1755 tss->ts[2] = (struct timespec64) {0};
1756}
1757
1758#ifdef CONFIG_MMU
1759static const struct vm_operations_struct tcp_vm_ops = {
1760};
1761
1762int tcp_mmap(struct file *file, struct socket *sock,
1763 struct vm_area_struct *vma)
1764{
1765 if (vma->vm_flags & (VM_WRITE | VM_EXEC))
1766 return -EPERM;
1767 vm_flags_clear(vma, VM_MAYWRITE | VM_MAYEXEC);
1768
1769 /* Instruct vm_insert_page() to not mmap_read_lock(mm) */
1770 vm_flags_set(vma, VM_MIXEDMAP);
1771
1772 vma->vm_ops = &tcp_vm_ops;
1773 return 0;
1774}
1775EXPORT_SYMBOL(tcp_mmap);
1776
1777static skb_frag_t *skb_advance_to_frag(struct sk_buff *skb, u32 offset_skb,
1778 u32 *offset_frag)
1779{
1780 skb_frag_t *frag;
1781
1782 if (unlikely(offset_skb >= skb->len))
1783 return NULL;
1784
1785 offset_skb -= skb_headlen(skb);
1786 if ((int)offset_skb < 0 || skb_has_frag_list(skb))
1787 return NULL;
1788
1789 frag = skb_shinfo(skb)->frags;
1790 while (offset_skb) {
1791 if (skb_frag_size(frag) > offset_skb) {
1792 *offset_frag = offset_skb;
1793 return frag;
1794 }
1795 offset_skb -= skb_frag_size(frag);
1796 ++frag;
1797 }
1798 *offset_frag = 0;
1799 return frag;
1800}
1801
1802static bool can_map_frag(const skb_frag_t *frag)
1803{
1804 struct page *page;
1805
1806 if (skb_frag_size(frag) != PAGE_SIZE || skb_frag_off(frag))
1807 return false;
1808
1809 page = skb_frag_page(frag);
1810
1811 if (PageCompound(page) || page->mapping)
1812 return false;
1813
1814 return true;
1815}
1816
1817static int find_next_mappable_frag(const skb_frag_t *frag,
1818 int remaining_in_skb)
1819{
1820 int offset = 0;
1821
1822 if (likely(can_map_frag(frag)))
1823 return 0;
1824
1825 while (offset < remaining_in_skb && !can_map_frag(frag)) {
1826 offset += skb_frag_size(frag);
1827 ++frag;
1828 }
1829 return offset;
1830}
1831
1832static void tcp_zerocopy_set_hint_for_skb(struct sock *sk,
1833 struct tcp_zerocopy_receive *zc,
1834 struct sk_buff *skb, u32 offset)
1835{
1836 u32 frag_offset, partial_frag_remainder = 0;
1837 int mappable_offset;
1838 skb_frag_t *frag;
1839
1840 /* worst case: skip to next skb. try to improve on this case below */
1841 zc->recv_skip_hint = skb->len - offset;
1842
1843 /* Find the frag containing this offset (and how far into that frag) */
1844 frag = skb_advance_to_frag(skb, offset, &frag_offset);
1845 if (!frag)
1846 return;
1847
1848 if (frag_offset) {
1849 struct skb_shared_info *info = skb_shinfo(skb);
1850
1851 /* We read part of the last frag, must recvmsg() rest of skb. */
1852 if (frag == &info->frags[info->nr_frags - 1])
1853 return;
1854
1855 /* Else, we must at least read the remainder in this frag. */
1856 partial_frag_remainder = skb_frag_size(frag) - frag_offset;
1857 zc->recv_skip_hint -= partial_frag_remainder;
1858 ++frag;
1859 }
1860
1861 /* partial_frag_remainder: If part way through a frag, must read rest.
1862 * mappable_offset: Bytes till next mappable frag, *not* counting bytes
1863 * in partial_frag_remainder.
1864 */
1865 mappable_offset = find_next_mappable_frag(frag, zc->recv_skip_hint);
1866 zc->recv_skip_hint = mappable_offset + partial_frag_remainder;
1867}
1868
1869static int tcp_recvmsg_locked(struct sock *sk, struct msghdr *msg, size_t len,
1870 int flags, struct scm_timestamping_internal *tss,
1871 int *cmsg_flags);
1872static int receive_fallback_to_copy(struct sock *sk,
1873 struct tcp_zerocopy_receive *zc, int inq,
1874 struct scm_timestamping_internal *tss)
1875{
1876 unsigned long copy_address = (unsigned long)zc->copybuf_address;
1877 struct msghdr msg = {};
1878 int err;
1879
1880 zc->length = 0;
1881 zc->recv_skip_hint = 0;
1882
1883 if (copy_address != zc->copybuf_address)
1884 return -EINVAL;
1885
1886 err = import_ubuf(ITER_DEST, (void __user *)copy_address, inq,
1887 &msg.msg_iter);
1888 if (err)
1889 return err;
1890
1891 err = tcp_recvmsg_locked(sk, &msg, inq, MSG_DONTWAIT,
1892 tss, &zc->msg_flags);
1893 if (err < 0)
1894 return err;
1895
1896 zc->copybuf_len = err;
1897 if (likely(zc->copybuf_len)) {
1898 struct sk_buff *skb;
1899 u32 offset;
1900
1901 skb = tcp_recv_skb(sk, tcp_sk(sk)->copied_seq, &offset);
1902 if (skb)
1903 tcp_zerocopy_set_hint_for_skb(sk, zc, skb, offset);
1904 }
1905 return 0;
1906}
1907
1908static int tcp_copy_straggler_data(struct tcp_zerocopy_receive *zc,
1909 struct sk_buff *skb, u32 copylen,
1910 u32 *offset, u32 *seq)
1911{
1912 unsigned long copy_address = (unsigned long)zc->copybuf_address;
1913 struct msghdr msg = {};
1914 int err;
1915
1916 if (copy_address != zc->copybuf_address)
1917 return -EINVAL;
1918
1919 err = import_ubuf(ITER_DEST, (void __user *)copy_address, copylen,
1920 &msg.msg_iter);
1921 if (err)
1922 return err;
1923 err = skb_copy_datagram_msg(skb, *offset, &msg, copylen);
1924 if (err)
1925 return err;
1926 zc->recv_skip_hint -= copylen;
1927 *offset += copylen;
1928 *seq += copylen;
1929 return (__s32)copylen;
1930}
1931
1932static int tcp_zc_handle_leftover(struct tcp_zerocopy_receive *zc,
1933 struct sock *sk,
1934 struct sk_buff *skb,
1935 u32 *seq,
1936 s32 copybuf_len,
1937 struct scm_timestamping_internal *tss)
1938{
1939 u32 offset, copylen = min_t(u32, copybuf_len, zc->recv_skip_hint);
1940
1941 if (!copylen)
1942 return 0;
1943 /* skb is null if inq < PAGE_SIZE. */
1944 if (skb) {
1945 offset = *seq - TCP_SKB_CB(skb)->seq;
1946 } else {
1947 skb = tcp_recv_skb(sk, *seq, &offset);
1948 if (TCP_SKB_CB(skb)->has_rxtstamp) {
1949 tcp_update_recv_tstamps(skb, tss);
1950 zc->msg_flags |= TCP_CMSG_TS;
1951 }
1952 }
1953
1954 zc->copybuf_len = tcp_copy_straggler_data(zc, skb, copylen, &offset,
1955 seq);
1956 return zc->copybuf_len < 0 ? 0 : copylen;
1957}
1958
1959static int tcp_zerocopy_vm_insert_batch_error(struct vm_area_struct *vma,
1960 struct page **pending_pages,
1961 unsigned long pages_remaining,
1962 unsigned long *address,
1963 u32 *length,
1964 u32 *seq,
1965 struct tcp_zerocopy_receive *zc,
1966 u32 total_bytes_to_map,
1967 int err)
1968{
1969 /* At least one page did not map. Try zapping if we skipped earlier. */
1970 if (err == -EBUSY &&
1971 zc->flags & TCP_RECEIVE_ZEROCOPY_FLAG_TLB_CLEAN_HINT) {
1972 u32 maybe_zap_len;
1973
1974 maybe_zap_len = total_bytes_to_map - /* All bytes to map */
1975 *length + /* Mapped or pending */
1976 (pages_remaining * PAGE_SIZE); /* Failed map. */
1977 zap_page_range_single(vma, *address, maybe_zap_len, NULL);
1978 err = 0;
1979 }
1980
1981 if (!err) {
1982 unsigned long leftover_pages = pages_remaining;
1983 int bytes_mapped;
1984
1985 /* We called zap_page_range_single, try to reinsert. */
1986 err = vm_insert_pages(vma, *address,
1987 pending_pages,
1988 &pages_remaining);
1989 bytes_mapped = PAGE_SIZE * (leftover_pages - pages_remaining);
1990 *seq += bytes_mapped;
1991 *address += bytes_mapped;
1992 }
1993 if (err) {
1994 /* Either we were unable to zap, OR we zapped, retried an
1995 * insert, and still had an issue. Either ways, pages_remaining
1996 * is the number of pages we were unable to map, and we unroll
1997 * some state we speculatively touched before.
1998 */
1999 const int bytes_not_mapped = PAGE_SIZE * pages_remaining;
2000
2001 *length -= bytes_not_mapped;
2002 zc->recv_skip_hint += bytes_not_mapped;
2003 }
2004 return err;
2005}
2006
2007static int tcp_zerocopy_vm_insert_batch(struct vm_area_struct *vma,
2008 struct page **pages,
2009 unsigned int pages_to_map,
2010 unsigned long *address,
2011 u32 *length,
2012 u32 *seq,
2013 struct tcp_zerocopy_receive *zc,
2014 u32 total_bytes_to_map)
2015{
2016 unsigned long pages_remaining = pages_to_map;
2017 unsigned int pages_mapped;
2018 unsigned int bytes_mapped;
2019 int err;
2020
2021 err = vm_insert_pages(vma, *address, pages, &pages_remaining);
2022 pages_mapped = pages_to_map - (unsigned int)pages_remaining;
2023 bytes_mapped = PAGE_SIZE * pages_mapped;
2024 /* Even if vm_insert_pages fails, it may have partially succeeded in
2025 * mapping (some but not all of the pages).
2026 */
2027 *seq += bytes_mapped;
2028 *address += bytes_mapped;
2029
2030 if (likely(!err))
2031 return 0;
2032
2033 /* Error: maybe zap and retry + rollback state for failed inserts. */
2034 return tcp_zerocopy_vm_insert_batch_error(vma, pages + pages_mapped,
2035 pages_remaining, address, length, seq, zc, total_bytes_to_map,
2036 err);
2037}
2038
2039#define TCP_VALID_ZC_MSG_FLAGS (TCP_CMSG_TS)
2040static void tcp_zc_finalize_rx_tstamp(struct sock *sk,
2041 struct tcp_zerocopy_receive *zc,
2042 struct scm_timestamping_internal *tss)
2043{
2044 unsigned long msg_control_addr;
2045 struct msghdr cmsg_dummy;
2046
2047 msg_control_addr = (unsigned long)zc->msg_control;
2048 cmsg_dummy.msg_control_user = (void __user *)msg_control_addr;
2049 cmsg_dummy.msg_controllen =
2050 (__kernel_size_t)zc->msg_controllen;
2051 cmsg_dummy.msg_flags = in_compat_syscall()
2052 ? MSG_CMSG_COMPAT : 0;
2053 cmsg_dummy.msg_control_is_user = true;
2054 zc->msg_flags = 0;
2055 if (zc->msg_control == msg_control_addr &&
2056 zc->msg_controllen == cmsg_dummy.msg_controllen) {
2057 tcp_recv_timestamp(&cmsg_dummy, sk, tss);
2058 zc->msg_control = (__u64)
2059 ((uintptr_t)cmsg_dummy.msg_control_user);
2060 zc->msg_controllen =
2061 (__u64)cmsg_dummy.msg_controllen;
2062 zc->msg_flags = (__u32)cmsg_dummy.msg_flags;
2063 }
2064}
2065
2066static struct vm_area_struct *find_tcp_vma(struct mm_struct *mm,
2067 unsigned long address,
2068 bool *mmap_locked)
2069{
2070 struct vm_area_struct *vma = lock_vma_under_rcu(mm, address);
2071
2072 if (vma) {
2073 if (vma->vm_ops != &tcp_vm_ops) {
2074 vma_end_read(vma);
2075 return NULL;
2076 }
2077 *mmap_locked = false;
2078 return vma;
2079 }
2080
2081 mmap_read_lock(mm);
2082 vma = vma_lookup(mm, address);
2083 if (!vma || vma->vm_ops != &tcp_vm_ops) {
2084 mmap_read_unlock(mm);
2085 return NULL;
2086 }
2087 *mmap_locked = true;
2088 return vma;
2089}
2090
2091#define TCP_ZEROCOPY_PAGE_BATCH_SIZE 32
2092static int tcp_zerocopy_receive(struct sock *sk,
2093 struct tcp_zerocopy_receive *zc,
2094 struct scm_timestamping_internal *tss)
2095{
2096 u32 length = 0, offset, vma_len, avail_len, copylen = 0;
2097 unsigned long address = (unsigned long)zc->address;
2098 struct page *pages[TCP_ZEROCOPY_PAGE_BATCH_SIZE];
2099 s32 copybuf_len = zc->copybuf_len;
2100 struct tcp_sock *tp = tcp_sk(sk);
2101 const skb_frag_t *frags = NULL;
2102 unsigned int pages_to_map = 0;
2103 struct vm_area_struct *vma;
2104 struct sk_buff *skb = NULL;
2105 u32 seq = tp->copied_seq;
2106 u32 total_bytes_to_map;
2107 int inq = tcp_inq(sk);
2108 bool mmap_locked;
2109 int ret;
2110
2111 zc->copybuf_len = 0;
2112 zc->msg_flags = 0;
2113
2114 if (address & (PAGE_SIZE - 1) || address != zc->address)
2115 return -EINVAL;
2116
2117 if (sk->sk_state == TCP_LISTEN)
2118 return -ENOTCONN;
2119
2120 sock_rps_record_flow(sk);
2121
2122 if (inq && inq <= copybuf_len)
2123 return receive_fallback_to_copy(sk, zc, inq, tss);
2124
2125 if (inq < PAGE_SIZE) {
2126 zc->length = 0;
2127 zc->recv_skip_hint = inq;
2128 if (!inq && sock_flag(sk, SOCK_DONE))
2129 return -EIO;
2130 return 0;
2131 }
2132
2133 vma = find_tcp_vma(current->mm, address, &mmap_locked);
2134 if (!vma)
2135 return -EINVAL;
2136
2137 vma_len = min_t(unsigned long, zc->length, vma->vm_end - address);
2138 avail_len = min_t(u32, vma_len, inq);
2139 total_bytes_to_map = avail_len & ~(PAGE_SIZE - 1);
2140 if (total_bytes_to_map) {
2141 if (!(zc->flags & TCP_RECEIVE_ZEROCOPY_FLAG_TLB_CLEAN_HINT))
2142 zap_page_range_single(vma, address, total_bytes_to_map,
2143 NULL);
2144 zc->length = total_bytes_to_map;
2145 zc->recv_skip_hint = 0;
2146 } else {
2147 zc->length = avail_len;
2148 zc->recv_skip_hint = avail_len;
2149 }
2150 ret = 0;
2151 while (length + PAGE_SIZE <= zc->length) {
2152 int mappable_offset;
2153 struct page *page;
2154
2155 if (zc->recv_skip_hint < PAGE_SIZE) {
2156 u32 offset_frag;
2157
2158 if (skb) {
2159 if (zc->recv_skip_hint > 0)
2160 break;
2161 skb = skb->next;
2162 offset = seq - TCP_SKB_CB(skb)->seq;
2163 } else {
2164 skb = tcp_recv_skb(sk, seq, &offset);
2165 }
2166
2167 if (!skb_frags_readable(skb))
2168 break;
2169
2170 if (TCP_SKB_CB(skb)->has_rxtstamp) {
2171 tcp_update_recv_tstamps(skb, tss);
2172 zc->msg_flags |= TCP_CMSG_TS;
2173 }
2174 zc->recv_skip_hint = skb->len - offset;
2175 frags = skb_advance_to_frag(skb, offset, &offset_frag);
2176 if (!frags || offset_frag)
2177 break;
2178 }
2179
2180 mappable_offset = find_next_mappable_frag(frags,
2181 zc->recv_skip_hint);
2182 if (mappable_offset) {
2183 zc->recv_skip_hint = mappable_offset;
2184 break;
2185 }
2186 page = skb_frag_page(frags);
2187 if (WARN_ON_ONCE(!page))
2188 break;
2189
2190 prefetchw(page);
2191 pages[pages_to_map++] = page;
2192 length += PAGE_SIZE;
2193 zc->recv_skip_hint -= PAGE_SIZE;
2194 frags++;
2195 if (pages_to_map == TCP_ZEROCOPY_PAGE_BATCH_SIZE ||
2196 zc->recv_skip_hint < PAGE_SIZE) {
2197 /* Either full batch, or we're about to go to next skb
2198 * (and we cannot unroll failed ops across skbs).
2199 */
2200 ret = tcp_zerocopy_vm_insert_batch(vma, pages,
2201 pages_to_map,
2202 &address, &length,
2203 &seq, zc,
2204 total_bytes_to_map);
2205 if (ret)
2206 goto out;
2207 pages_to_map = 0;
2208 }
2209 }
2210 if (pages_to_map) {
2211 ret = tcp_zerocopy_vm_insert_batch(vma, pages, pages_to_map,
2212 &address, &length, &seq,
2213 zc, total_bytes_to_map);
2214 }
2215out:
2216 if (mmap_locked)
2217 mmap_read_unlock(current->mm);
2218 else
2219 vma_end_read(vma);
2220 /* Try to copy straggler data. */
2221 if (!ret)
2222 copylen = tcp_zc_handle_leftover(zc, sk, skb, &seq, copybuf_len, tss);
2223
2224 if (length + copylen) {
2225 WRITE_ONCE(tp->copied_seq, seq);
2226 tcp_rcv_space_adjust(sk);
2227
2228 /* Clean up data we have read: This will do ACK frames. */
2229 tcp_recv_skb(sk, seq, &offset);
2230 tcp_cleanup_rbuf(sk, length + copylen);
2231 ret = 0;
2232 if (length == zc->length)
2233 zc->recv_skip_hint = 0;
2234 } else {
2235 if (!zc->recv_skip_hint && sock_flag(sk, SOCK_DONE))
2236 ret = -EIO;
2237 }
2238 zc->length = length;
2239 return ret;
2240}
2241#endif
2242
2243/* Similar to __sock_recv_timestamp, but does not require an skb */
2244void tcp_recv_timestamp(struct msghdr *msg, const struct sock *sk,
2245 struct scm_timestamping_internal *tss)
2246{
2247 int new_tstamp = sock_flag(sk, SOCK_TSTAMP_NEW);
2248 u32 tsflags = READ_ONCE(sk->sk_tsflags);
2249 bool has_timestamping = false;
2250
2251 if (tss->ts[0].tv_sec || tss->ts[0].tv_nsec) {
2252 if (sock_flag(sk, SOCK_RCVTSTAMP)) {
2253 if (sock_flag(sk, SOCK_RCVTSTAMPNS)) {
2254 if (new_tstamp) {
2255 struct __kernel_timespec kts = {
2256 .tv_sec = tss->ts[0].tv_sec,
2257 .tv_nsec = tss->ts[0].tv_nsec,
2258 };
2259 put_cmsg(msg, SOL_SOCKET, SO_TIMESTAMPNS_NEW,
2260 sizeof(kts), &kts);
2261 } else {
2262 struct __kernel_old_timespec ts_old = {
2263 .tv_sec = tss->ts[0].tv_sec,
2264 .tv_nsec = tss->ts[0].tv_nsec,
2265 };
2266 put_cmsg(msg, SOL_SOCKET, SO_TIMESTAMPNS_OLD,
2267 sizeof(ts_old), &ts_old);
2268 }
2269 } else {
2270 if (new_tstamp) {
2271 struct __kernel_sock_timeval stv = {
2272 .tv_sec = tss->ts[0].tv_sec,
2273 .tv_usec = tss->ts[0].tv_nsec / 1000,
2274 };
2275 put_cmsg(msg, SOL_SOCKET, SO_TIMESTAMP_NEW,
2276 sizeof(stv), &stv);
2277 } else {
2278 struct __kernel_old_timeval tv = {
2279 .tv_sec = tss->ts[0].tv_sec,
2280 .tv_usec = tss->ts[0].tv_nsec / 1000,
2281 };
2282 put_cmsg(msg, SOL_SOCKET, SO_TIMESTAMP_OLD,
2283 sizeof(tv), &tv);
2284 }
2285 }
2286 }
2287
2288 if (tsflags & SOF_TIMESTAMPING_SOFTWARE &&
2289 (tsflags & SOF_TIMESTAMPING_RX_SOFTWARE ||
2290 !(tsflags & SOF_TIMESTAMPING_OPT_RX_FILTER)))
2291 has_timestamping = true;
2292 else
2293 tss->ts[0] = (struct timespec64) {0};
2294 }
2295
2296 if (tss->ts[2].tv_sec || tss->ts[2].tv_nsec) {
2297 if (tsflags & SOF_TIMESTAMPING_RAW_HARDWARE &&
2298 (tsflags & SOF_TIMESTAMPING_RX_HARDWARE ||
2299 !(tsflags & SOF_TIMESTAMPING_OPT_RX_FILTER)))
2300 has_timestamping = true;
2301 else
2302 tss->ts[2] = (struct timespec64) {0};
2303 }
2304
2305 if (has_timestamping) {
2306 tss->ts[1] = (struct timespec64) {0};
2307 if (sock_flag(sk, SOCK_TSTAMP_NEW))
2308 put_cmsg_scm_timestamping64(msg, tss);
2309 else
2310 put_cmsg_scm_timestamping(msg, tss);
2311 }
2312}
2313
2314static int tcp_inq_hint(struct sock *sk)
2315{
2316 const struct tcp_sock *tp = tcp_sk(sk);
2317 u32 copied_seq = READ_ONCE(tp->copied_seq);
2318 u32 rcv_nxt = READ_ONCE(tp->rcv_nxt);
2319 int inq;
2320
2321 inq = rcv_nxt - copied_seq;
2322 if (unlikely(inq < 0 || copied_seq != READ_ONCE(tp->copied_seq))) {
2323 lock_sock(sk);
2324 inq = tp->rcv_nxt - tp->copied_seq;
2325 release_sock(sk);
2326 }
2327 /* After receiving a FIN, tell the user-space to continue reading
2328 * by returning a non-zero inq.
2329 */
2330 if (inq == 0 && sock_flag(sk, SOCK_DONE))
2331 inq = 1;
2332 return inq;
2333}
2334
2335/* batch __xa_alloc() calls and reduce xa_lock()/xa_unlock() overhead. */
2336struct tcp_xa_pool {
2337 u8 max; /* max <= MAX_SKB_FRAGS */
2338 u8 idx; /* idx <= max */
2339 __u32 tokens[MAX_SKB_FRAGS];
2340 netmem_ref netmems[MAX_SKB_FRAGS];
2341};
2342
2343static void tcp_xa_pool_commit_locked(struct sock *sk, struct tcp_xa_pool *p)
2344{
2345 int i;
2346
2347 /* Commit part that has been copied to user space. */
2348 for (i = 0; i < p->idx; i++)
2349 __xa_cmpxchg(&sk->sk_user_frags, p->tokens[i], XA_ZERO_ENTRY,
2350 (__force void *)p->netmems[i], GFP_KERNEL);
2351 /* Rollback what has been pre-allocated and is no longer needed. */
2352 for (; i < p->max; i++)
2353 __xa_erase(&sk->sk_user_frags, p->tokens[i]);
2354
2355 p->max = 0;
2356 p->idx = 0;
2357}
2358
2359static void tcp_xa_pool_commit(struct sock *sk, struct tcp_xa_pool *p)
2360{
2361 if (!p->max)
2362 return;
2363
2364 xa_lock_bh(&sk->sk_user_frags);
2365
2366 tcp_xa_pool_commit_locked(sk, p);
2367
2368 xa_unlock_bh(&sk->sk_user_frags);
2369}
2370
2371static int tcp_xa_pool_refill(struct sock *sk, struct tcp_xa_pool *p,
2372 unsigned int max_frags)
2373{
2374 int err, k;
2375
2376 if (p->idx < p->max)
2377 return 0;
2378
2379 xa_lock_bh(&sk->sk_user_frags);
2380
2381 tcp_xa_pool_commit_locked(sk, p);
2382
2383 for (k = 0; k < max_frags; k++) {
2384 err = __xa_alloc(&sk->sk_user_frags, &p->tokens[k],
2385 XA_ZERO_ENTRY, xa_limit_31b, GFP_KERNEL);
2386 if (err)
2387 break;
2388 }
2389
2390 xa_unlock_bh(&sk->sk_user_frags);
2391
2392 p->max = k;
2393 p->idx = 0;
2394 return k ? 0 : err;
2395}
2396
2397/* On error, returns the -errno. On success, returns number of bytes sent to the
2398 * user. May not consume all of @remaining_len.
2399 */
2400static int tcp_recvmsg_dmabuf(struct sock *sk, const struct sk_buff *skb,
2401 unsigned int offset, struct msghdr *msg,
2402 int remaining_len)
2403{
2404 struct dmabuf_cmsg dmabuf_cmsg = { 0 };
2405 struct tcp_xa_pool tcp_xa_pool;
2406 unsigned int start;
2407 int i, copy, n;
2408 int sent = 0;
2409 int err = 0;
2410
2411 tcp_xa_pool.max = 0;
2412 tcp_xa_pool.idx = 0;
2413 do {
2414 start = skb_headlen(skb);
2415
2416 if (skb_frags_readable(skb)) {
2417 err = -ENODEV;
2418 goto out;
2419 }
2420
2421 /* Copy header. */
2422 copy = start - offset;
2423 if (copy > 0) {
2424 copy = min(copy, remaining_len);
2425
2426 n = copy_to_iter(skb->data + offset, copy,
2427 &msg->msg_iter);
2428 if (n != copy) {
2429 err = -EFAULT;
2430 goto out;
2431 }
2432
2433 offset += copy;
2434 remaining_len -= copy;
2435
2436 /* First a dmabuf_cmsg for # bytes copied to user
2437 * buffer.
2438 */
2439 memset(&dmabuf_cmsg, 0, sizeof(dmabuf_cmsg));
2440 dmabuf_cmsg.frag_size = copy;
2441 err = put_cmsg_notrunc(msg, SOL_SOCKET,
2442 SO_DEVMEM_LINEAR,
2443 sizeof(dmabuf_cmsg),
2444 &dmabuf_cmsg);
2445 if (err)
2446 goto out;
2447
2448 sent += copy;
2449
2450 if (remaining_len == 0)
2451 goto out;
2452 }
2453
2454 /* after that, send information of dmabuf pages through a
2455 * sequence of cmsg
2456 */
2457 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
2458 skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
2459 struct net_iov *niov;
2460 u64 frag_offset;
2461 int end;
2462
2463 /* !skb_frags_readable() should indicate that ALL the
2464 * frags in this skb are dmabuf net_iovs. We're checking
2465 * for that flag above, but also check individual frags
2466 * here. If the tcp stack is not setting
2467 * skb_frags_readable() correctly, we still don't want
2468 * to crash here.
2469 */
2470 if (!skb_frag_net_iov(frag)) {
2471 net_err_ratelimited("Found non-dmabuf skb with net_iov");
2472 err = -ENODEV;
2473 goto out;
2474 }
2475
2476 niov = skb_frag_net_iov(frag);
2477 end = start + skb_frag_size(frag);
2478 copy = end - offset;
2479
2480 if (copy > 0) {
2481 copy = min(copy, remaining_len);
2482
2483 frag_offset = net_iov_virtual_addr(niov) +
2484 skb_frag_off(frag) + offset -
2485 start;
2486 dmabuf_cmsg.frag_offset = frag_offset;
2487 dmabuf_cmsg.frag_size = copy;
2488 err = tcp_xa_pool_refill(sk, &tcp_xa_pool,
2489 skb_shinfo(skb)->nr_frags - i);
2490 if (err)
2491 goto out;
2492
2493 /* Will perform the exchange later */
2494 dmabuf_cmsg.frag_token = tcp_xa_pool.tokens[tcp_xa_pool.idx];
2495 dmabuf_cmsg.dmabuf_id = net_iov_binding_id(niov);
2496
2497 offset += copy;
2498 remaining_len -= copy;
2499
2500 err = put_cmsg_notrunc(msg, SOL_SOCKET,
2501 SO_DEVMEM_DMABUF,
2502 sizeof(dmabuf_cmsg),
2503 &dmabuf_cmsg);
2504 if (err)
2505 goto out;
2506
2507 atomic_long_inc(&niov->pp_ref_count);
2508 tcp_xa_pool.netmems[tcp_xa_pool.idx++] = skb_frag_netmem(frag);
2509
2510 sent += copy;
2511
2512 if (remaining_len == 0)
2513 goto out;
2514 }
2515 start = end;
2516 }
2517
2518 tcp_xa_pool_commit(sk, &tcp_xa_pool);
2519 if (!remaining_len)
2520 goto out;
2521
2522 /* if remaining_len is not satisfied yet, we need to go to the
2523 * next frag in the frag_list to satisfy remaining_len.
2524 */
2525 skb = skb_shinfo(skb)->frag_list ?: skb->next;
2526
2527 offset = offset - start;
2528 } while (skb);
2529
2530 if (remaining_len) {
2531 err = -EFAULT;
2532 goto out;
2533 }
2534
2535out:
2536 tcp_xa_pool_commit(sk, &tcp_xa_pool);
2537 if (!sent)
2538 sent = err;
2539
2540 return sent;
2541}
2542
2543/*
2544 * This routine copies from a sock struct into the user buffer.
2545 *
2546 * Technical note: in 2.3 we work on _locked_ socket, so that
2547 * tricks with *seq access order and skb->users are not required.
2548 * Probably, code can be easily improved even more.
2549 */
2550
2551static int tcp_recvmsg_locked(struct sock *sk, struct msghdr *msg, size_t len,
2552 int flags, struct scm_timestamping_internal *tss,
2553 int *cmsg_flags)
2554{
2555 struct tcp_sock *tp = tcp_sk(sk);
2556 int last_copied_dmabuf = -1; /* uninitialized */
2557 int copied = 0;
2558 u32 peek_seq;
2559 u32 *seq;
2560 unsigned long used;
2561 int err;
2562 int target; /* Read at least this many bytes */
2563 long timeo;
2564 struct sk_buff *skb, *last;
2565 u32 peek_offset = 0;
2566 u32 urg_hole = 0;
2567
2568 err = -ENOTCONN;
2569 if (sk->sk_state == TCP_LISTEN)
2570 goto out;
2571
2572 if (tp->recvmsg_inq) {
2573 *cmsg_flags = TCP_CMSG_INQ;
2574 msg->msg_get_inq = 1;
2575 }
2576 timeo = sock_rcvtimeo(sk, flags & MSG_DONTWAIT);
2577
2578 /* Urgent data needs to be handled specially. */
2579 if (flags & MSG_OOB)
2580 goto recv_urg;
2581
2582 if (unlikely(tp->repair)) {
2583 err = -EPERM;
2584 if (!(flags & MSG_PEEK))
2585 goto out;
2586
2587 if (tp->repair_queue == TCP_SEND_QUEUE)
2588 goto recv_sndq;
2589
2590 err = -EINVAL;
2591 if (tp->repair_queue == TCP_NO_QUEUE)
2592 goto out;
2593
2594 /* 'common' recv queue MSG_PEEK-ing */
2595 }
2596
2597 seq = &tp->copied_seq;
2598 if (flags & MSG_PEEK) {
2599 peek_offset = max(sk_peek_offset(sk, flags), 0);
2600 peek_seq = tp->copied_seq + peek_offset;
2601 seq = &peek_seq;
2602 }
2603
2604 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
2605
2606 do {
2607 u32 offset;
2608
2609 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
2610 if (unlikely(tp->urg_data) && tp->urg_seq == *seq) {
2611 if (copied)
2612 break;
2613 if (signal_pending(current)) {
2614 copied = timeo ? sock_intr_errno(timeo) : -EAGAIN;
2615 break;
2616 }
2617 }
2618
2619 /* Next get a buffer. */
2620
2621 last = skb_peek_tail(&sk->sk_receive_queue);
2622 skb_queue_walk(&sk->sk_receive_queue, skb) {
2623 last = skb;
2624 /* Now that we have two receive queues this
2625 * shouldn't happen.
2626 */
2627 if (WARN(before(*seq, TCP_SKB_CB(skb)->seq),
2628 "TCP recvmsg seq # bug: copied %X, seq %X, rcvnxt %X, fl %X\n",
2629 *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt,
2630 flags))
2631 break;
2632
2633 offset = *seq - TCP_SKB_CB(skb)->seq;
2634 if (unlikely(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
2635 pr_err_once("%s: found a SYN, please report !\n", __func__);
2636 offset--;
2637 }
2638 if (offset < skb->len)
2639 goto found_ok_skb;
2640 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
2641 goto found_fin_ok;
2642 WARN(!(flags & MSG_PEEK),
2643 "TCP recvmsg seq # bug 2: copied %X, seq %X, rcvnxt %X, fl %X\n",
2644 *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt, flags);
2645 }
2646
2647 /* Well, if we have backlog, try to process it now yet. */
2648
2649 if (copied >= target && !READ_ONCE(sk->sk_backlog.tail))
2650 break;
2651
2652 if (copied) {
2653 if (!timeo ||
2654 sk->sk_err ||
2655 sk->sk_state == TCP_CLOSE ||
2656 (sk->sk_shutdown & RCV_SHUTDOWN) ||
2657 signal_pending(current))
2658 break;
2659 } else {
2660 if (sock_flag(sk, SOCK_DONE))
2661 break;
2662
2663 if (sk->sk_err) {
2664 copied = sock_error(sk);
2665 break;
2666 }
2667
2668 if (sk->sk_shutdown & RCV_SHUTDOWN)
2669 break;
2670
2671 if (sk->sk_state == TCP_CLOSE) {
2672 /* This occurs when user tries to read
2673 * from never connected socket.
2674 */
2675 copied = -ENOTCONN;
2676 break;
2677 }
2678
2679 if (!timeo) {
2680 copied = -EAGAIN;
2681 break;
2682 }
2683
2684 if (signal_pending(current)) {
2685 copied = sock_intr_errno(timeo);
2686 break;
2687 }
2688 }
2689
2690 if (copied >= target) {
2691 /* Do not sleep, just process backlog. */
2692 __sk_flush_backlog(sk);
2693 } else {
2694 tcp_cleanup_rbuf(sk, copied);
2695 err = sk_wait_data(sk, &timeo, last);
2696 if (err < 0) {
2697 err = copied ? : err;
2698 goto out;
2699 }
2700 }
2701
2702 if ((flags & MSG_PEEK) &&
2703 (peek_seq - peek_offset - copied - urg_hole != tp->copied_seq)) {
2704 net_dbg_ratelimited("TCP(%s:%d): Application bug, race in MSG_PEEK\n",
2705 current->comm,
2706 task_pid_nr(current));
2707 peek_seq = tp->copied_seq + peek_offset;
2708 }
2709 continue;
2710
2711found_ok_skb:
2712 /* Ok so how much can we use? */
2713 used = skb->len - offset;
2714 if (len < used)
2715 used = len;
2716
2717 /* Do we have urgent data here? */
2718 if (unlikely(tp->urg_data)) {
2719 u32 urg_offset = tp->urg_seq - *seq;
2720 if (urg_offset < used) {
2721 if (!urg_offset) {
2722 if (!sock_flag(sk, SOCK_URGINLINE)) {
2723 WRITE_ONCE(*seq, *seq + 1);
2724 urg_hole++;
2725 offset++;
2726 used--;
2727 if (!used)
2728 goto skip_copy;
2729 }
2730 } else
2731 used = urg_offset;
2732 }
2733 }
2734
2735 if (!(flags & MSG_TRUNC)) {
2736 if (last_copied_dmabuf != -1 &&
2737 last_copied_dmabuf != !skb_frags_readable(skb))
2738 break;
2739
2740 if (skb_frags_readable(skb)) {
2741 err = skb_copy_datagram_msg(skb, offset, msg,
2742 used);
2743 if (err) {
2744 /* Exception. Bailout! */
2745 if (!copied)
2746 copied = -EFAULT;
2747 break;
2748 }
2749 } else {
2750 if (!(flags & MSG_SOCK_DEVMEM)) {
2751 /* dmabuf skbs can only be received
2752 * with the MSG_SOCK_DEVMEM flag.
2753 */
2754 if (!copied)
2755 copied = -EFAULT;
2756
2757 break;
2758 }
2759
2760 err = tcp_recvmsg_dmabuf(sk, skb, offset, msg,
2761 used);
2762 if (err <= 0) {
2763 if (!copied)
2764 copied = -EFAULT;
2765
2766 break;
2767 }
2768 used = err;
2769 }
2770 }
2771
2772 last_copied_dmabuf = !skb_frags_readable(skb);
2773
2774 WRITE_ONCE(*seq, *seq + used);
2775 copied += used;
2776 len -= used;
2777 if (flags & MSG_PEEK)
2778 sk_peek_offset_fwd(sk, used);
2779 else
2780 sk_peek_offset_bwd(sk, used);
2781 tcp_rcv_space_adjust(sk);
2782
2783skip_copy:
2784 if (unlikely(tp->urg_data) && after(tp->copied_seq, tp->urg_seq)) {
2785 WRITE_ONCE(tp->urg_data, 0);
2786 tcp_fast_path_check(sk);
2787 }
2788
2789 if (TCP_SKB_CB(skb)->has_rxtstamp) {
2790 tcp_update_recv_tstamps(skb, tss);
2791 *cmsg_flags |= TCP_CMSG_TS;
2792 }
2793
2794 if (used + offset < skb->len)
2795 continue;
2796
2797 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
2798 goto found_fin_ok;
2799 if (!(flags & MSG_PEEK))
2800 tcp_eat_recv_skb(sk, skb);
2801 continue;
2802
2803found_fin_ok:
2804 /* Process the FIN. */
2805 WRITE_ONCE(*seq, *seq + 1);
2806 if (!(flags & MSG_PEEK))
2807 tcp_eat_recv_skb(sk, skb);
2808 break;
2809 } while (len > 0);
2810
2811 /* According to UNIX98, msg_name/msg_namelen are ignored
2812 * on connected socket. I was just happy when found this 8) --ANK
2813 */
2814
2815 /* Clean up data we have read: This will do ACK frames. */
2816 tcp_cleanup_rbuf(sk, copied);
2817 return copied;
2818
2819out:
2820 return err;
2821
2822recv_urg:
2823 err = tcp_recv_urg(sk, msg, len, flags);
2824 goto out;
2825
2826recv_sndq:
2827 err = tcp_peek_sndq(sk, msg, len);
2828 goto out;
2829}
2830
2831int tcp_recvmsg(struct sock *sk, struct msghdr *msg, size_t len, int flags,
2832 int *addr_len)
2833{
2834 int cmsg_flags = 0, ret;
2835 struct scm_timestamping_internal tss;
2836
2837 if (unlikely(flags & MSG_ERRQUEUE))
2838 return inet_recv_error(sk, msg, len, addr_len);
2839
2840 if (sk_can_busy_loop(sk) &&
2841 skb_queue_empty_lockless(&sk->sk_receive_queue) &&
2842 sk->sk_state == TCP_ESTABLISHED)
2843 sk_busy_loop(sk, flags & MSG_DONTWAIT);
2844
2845 lock_sock(sk);
2846 ret = tcp_recvmsg_locked(sk, msg, len, flags, &tss, &cmsg_flags);
2847 release_sock(sk);
2848
2849 if ((cmsg_flags || msg->msg_get_inq) && ret >= 0) {
2850 if (cmsg_flags & TCP_CMSG_TS)
2851 tcp_recv_timestamp(msg, sk, &tss);
2852 if (msg->msg_get_inq) {
2853 msg->msg_inq = tcp_inq_hint(sk);
2854 if (cmsg_flags & TCP_CMSG_INQ)
2855 put_cmsg(msg, SOL_TCP, TCP_CM_INQ,
2856 sizeof(msg->msg_inq), &msg->msg_inq);
2857 }
2858 }
2859 return ret;
2860}
2861EXPORT_SYMBOL(tcp_recvmsg);
2862
2863void tcp_set_state(struct sock *sk, int state)
2864{
2865 int oldstate = sk->sk_state;
2866
2867 /* We defined a new enum for TCP states that are exported in BPF
2868 * so as not force the internal TCP states to be frozen. The
2869 * following checks will detect if an internal state value ever
2870 * differs from the BPF value. If this ever happens, then we will
2871 * need to remap the internal value to the BPF value before calling
2872 * tcp_call_bpf_2arg.
2873 */
2874 BUILD_BUG_ON((int)BPF_TCP_ESTABLISHED != (int)TCP_ESTABLISHED);
2875 BUILD_BUG_ON((int)BPF_TCP_SYN_SENT != (int)TCP_SYN_SENT);
2876 BUILD_BUG_ON((int)BPF_TCP_SYN_RECV != (int)TCP_SYN_RECV);
2877 BUILD_BUG_ON((int)BPF_TCP_FIN_WAIT1 != (int)TCP_FIN_WAIT1);
2878 BUILD_BUG_ON((int)BPF_TCP_FIN_WAIT2 != (int)TCP_FIN_WAIT2);
2879 BUILD_BUG_ON((int)BPF_TCP_TIME_WAIT != (int)TCP_TIME_WAIT);
2880 BUILD_BUG_ON((int)BPF_TCP_CLOSE != (int)TCP_CLOSE);
2881 BUILD_BUG_ON((int)BPF_TCP_CLOSE_WAIT != (int)TCP_CLOSE_WAIT);
2882 BUILD_BUG_ON((int)BPF_TCP_LAST_ACK != (int)TCP_LAST_ACK);
2883 BUILD_BUG_ON((int)BPF_TCP_LISTEN != (int)TCP_LISTEN);
2884 BUILD_BUG_ON((int)BPF_TCP_CLOSING != (int)TCP_CLOSING);
2885 BUILD_BUG_ON((int)BPF_TCP_NEW_SYN_RECV != (int)TCP_NEW_SYN_RECV);
2886 BUILD_BUG_ON((int)BPF_TCP_BOUND_INACTIVE != (int)TCP_BOUND_INACTIVE);
2887 BUILD_BUG_ON((int)BPF_TCP_MAX_STATES != (int)TCP_MAX_STATES);
2888
2889 /* bpf uapi header bpf.h defines an anonymous enum with values
2890 * BPF_TCP_* used by bpf programs. Currently gcc built vmlinux
2891 * is able to emit this enum in DWARF due to the above BUILD_BUG_ON.
2892 * But clang built vmlinux does not have this enum in DWARF
2893 * since clang removes the above code before generating IR/debuginfo.
2894 * Let us explicitly emit the type debuginfo to ensure the
2895 * above-mentioned anonymous enum in the vmlinux DWARF and hence BTF
2896 * regardless of which compiler is used.
2897 */
2898 BTF_TYPE_EMIT_ENUM(BPF_TCP_ESTABLISHED);
2899
2900 if (BPF_SOCK_OPS_TEST_FLAG(tcp_sk(sk), BPF_SOCK_OPS_STATE_CB_FLAG))
2901 tcp_call_bpf_2arg(sk, BPF_SOCK_OPS_STATE_CB, oldstate, state);
2902
2903 switch (state) {
2904 case TCP_ESTABLISHED:
2905 if (oldstate != TCP_ESTABLISHED)
2906 TCP_INC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
2907 break;
2908 case TCP_CLOSE_WAIT:
2909 if (oldstate == TCP_SYN_RECV)
2910 TCP_INC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
2911 break;
2912
2913 case TCP_CLOSE:
2914 if (oldstate == TCP_CLOSE_WAIT || oldstate == TCP_ESTABLISHED)
2915 TCP_INC_STATS(sock_net(sk), TCP_MIB_ESTABRESETS);
2916
2917 sk->sk_prot->unhash(sk);
2918 if (inet_csk(sk)->icsk_bind_hash &&
2919 !(sk->sk_userlocks & SOCK_BINDPORT_LOCK))
2920 inet_put_port(sk);
2921 fallthrough;
2922 default:
2923 if (oldstate == TCP_ESTABLISHED || oldstate == TCP_CLOSE_WAIT)
2924 TCP_DEC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
2925 }
2926
2927 /* Change state AFTER socket is unhashed to avoid closed
2928 * socket sitting in hash tables.
2929 */
2930 inet_sk_state_store(sk, state);
2931}
2932EXPORT_SYMBOL_GPL(tcp_set_state);
2933
2934/*
2935 * State processing on a close. This implements the state shift for
2936 * sending our FIN frame. Note that we only send a FIN for some
2937 * states. A shutdown() may have already sent the FIN, or we may be
2938 * closed.
2939 */
2940
2941static const unsigned char new_state[16] = {
2942 /* current state: new state: action: */
2943 [0 /* (Invalid) */] = TCP_CLOSE,
2944 [TCP_ESTABLISHED] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2945 [TCP_SYN_SENT] = TCP_CLOSE,
2946 [TCP_SYN_RECV] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2947 [TCP_FIN_WAIT1] = TCP_FIN_WAIT1,
2948 [TCP_FIN_WAIT2] = TCP_FIN_WAIT2,
2949 [TCP_TIME_WAIT] = TCP_CLOSE,
2950 [TCP_CLOSE] = TCP_CLOSE,
2951 [TCP_CLOSE_WAIT] = TCP_LAST_ACK | TCP_ACTION_FIN,
2952 [TCP_LAST_ACK] = TCP_LAST_ACK,
2953 [TCP_LISTEN] = TCP_CLOSE,
2954 [TCP_CLOSING] = TCP_CLOSING,
2955 [TCP_NEW_SYN_RECV] = TCP_CLOSE, /* should not happen ! */
2956};
2957
2958static int tcp_close_state(struct sock *sk)
2959{
2960 int next = (int)new_state[sk->sk_state];
2961 int ns = next & TCP_STATE_MASK;
2962
2963 tcp_set_state(sk, ns);
2964
2965 return next & TCP_ACTION_FIN;
2966}
2967
2968/*
2969 * Shutdown the sending side of a connection. Much like close except
2970 * that we don't receive shut down or sock_set_flag(sk, SOCK_DEAD).
2971 */
2972
2973void tcp_shutdown(struct sock *sk, int how)
2974{
2975 /* We need to grab some memory, and put together a FIN,
2976 * and then put it into the queue to be sent.
2977 * Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
2978 */
2979 if (!(how & SEND_SHUTDOWN))
2980 return;
2981
2982 /* If we've already sent a FIN, or it's a closed state, skip this. */
2983 if ((1 << sk->sk_state) &
2984 (TCPF_ESTABLISHED | TCPF_SYN_SENT |
2985 TCPF_CLOSE_WAIT)) {
2986 /* Clear out any half completed packets. FIN if needed. */
2987 if (tcp_close_state(sk))
2988 tcp_send_fin(sk);
2989 }
2990}
2991EXPORT_SYMBOL(tcp_shutdown);
2992
2993int tcp_orphan_count_sum(void)
2994{
2995 int i, total = 0;
2996
2997 for_each_possible_cpu(i)
2998 total += per_cpu(tcp_orphan_count, i);
2999
3000 return max(total, 0);
3001}
3002
3003static int tcp_orphan_cache;
3004static struct timer_list tcp_orphan_timer;
3005#define TCP_ORPHAN_TIMER_PERIOD msecs_to_jiffies(100)
3006
3007static void tcp_orphan_update(struct timer_list *unused)
3008{
3009 WRITE_ONCE(tcp_orphan_cache, tcp_orphan_count_sum());
3010 mod_timer(&tcp_orphan_timer, jiffies + TCP_ORPHAN_TIMER_PERIOD);
3011}
3012
3013static bool tcp_too_many_orphans(int shift)
3014{
3015 return READ_ONCE(tcp_orphan_cache) << shift >
3016 READ_ONCE(sysctl_tcp_max_orphans);
3017}
3018
3019static bool tcp_out_of_memory(const struct sock *sk)
3020{
3021 if (sk->sk_wmem_queued > SOCK_MIN_SNDBUF &&
3022 sk_memory_allocated(sk) > sk_prot_mem_limits(sk, 2))
3023 return true;
3024 return false;
3025}
3026
3027bool tcp_check_oom(const struct sock *sk, int shift)
3028{
3029 bool too_many_orphans, out_of_socket_memory;
3030
3031 too_many_orphans = tcp_too_many_orphans(shift);
3032 out_of_socket_memory = tcp_out_of_memory(sk);
3033
3034 if (too_many_orphans)
3035 net_info_ratelimited("too many orphaned sockets\n");
3036 if (out_of_socket_memory)
3037 net_info_ratelimited("out of memory -- consider tuning tcp_mem\n");
3038 return too_many_orphans || out_of_socket_memory;
3039}
3040
3041void __tcp_close(struct sock *sk, long timeout)
3042{
3043 struct sk_buff *skb;
3044 int data_was_unread = 0;
3045 int state;
3046
3047 WRITE_ONCE(sk->sk_shutdown, SHUTDOWN_MASK);
3048
3049 if (sk->sk_state == TCP_LISTEN) {
3050 tcp_set_state(sk, TCP_CLOSE);
3051
3052 /* Special case. */
3053 inet_csk_listen_stop(sk);
3054
3055 goto adjudge_to_death;
3056 }
3057
3058 /* We need to flush the recv. buffs. We do this only on the
3059 * descriptor close, not protocol-sourced closes, because the
3060 * reader process may not have drained the data yet!
3061 */
3062 while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) {
3063 u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq;
3064
3065 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
3066 len--;
3067 data_was_unread += len;
3068 __kfree_skb(skb);
3069 }
3070
3071 /* If socket has been already reset (e.g. in tcp_reset()) - kill it. */
3072 if (sk->sk_state == TCP_CLOSE)
3073 goto adjudge_to_death;
3074
3075 /* As outlined in RFC 2525, section 2.17, we send a RST here because
3076 * data was lost. To witness the awful effects of the old behavior of
3077 * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk
3078 * GET in an FTP client, suspend the process, wait for the client to
3079 * advertise a zero window, then kill -9 the FTP client, wheee...
3080 * Note: timeout is always zero in such a case.
3081 */
3082 if (unlikely(tcp_sk(sk)->repair)) {
3083 sk->sk_prot->disconnect(sk, 0);
3084 } else if (data_was_unread) {
3085 /* Unread data was tossed, zap the connection. */
3086 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONCLOSE);
3087 tcp_set_state(sk, TCP_CLOSE);
3088 tcp_send_active_reset(sk, sk->sk_allocation,
3089 SK_RST_REASON_TCP_ABORT_ON_CLOSE);
3090 } else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) {
3091 /* Check zero linger _after_ checking for unread data. */
3092 sk->sk_prot->disconnect(sk, 0);
3093 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
3094 } else if (tcp_close_state(sk)) {
3095 /* We FIN if the application ate all the data before
3096 * zapping the connection.
3097 */
3098
3099 /* RED-PEN. Formally speaking, we have broken TCP state
3100 * machine. State transitions:
3101 *
3102 * TCP_ESTABLISHED -> TCP_FIN_WAIT1
3103 * TCP_SYN_RECV -> TCP_FIN_WAIT1 (it is difficult)
3104 * TCP_CLOSE_WAIT -> TCP_LAST_ACK
3105 *
3106 * are legal only when FIN has been sent (i.e. in window),
3107 * rather than queued out of window. Purists blame.
3108 *
3109 * F.e. "RFC state" is ESTABLISHED,
3110 * if Linux state is FIN-WAIT-1, but FIN is still not sent.
3111 *
3112 * The visible declinations are that sometimes
3113 * we enter time-wait state, when it is not required really
3114 * (harmless), do not send active resets, when they are
3115 * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
3116 * they look as CLOSING or LAST_ACK for Linux)
3117 * Probably, I missed some more holelets.
3118 * --ANK
3119 * XXX (TFO) - To start off we don't support SYN+ACK+FIN
3120 * in a single packet! (May consider it later but will
3121 * probably need API support or TCP_CORK SYN-ACK until
3122 * data is written and socket is closed.)
3123 */
3124 tcp_send_fin(sk);
3125 }
3126
3127 sk_stream_wait_close(sk, timeout);
3128
3129adjudge_to_death:
3130 state = sk->sk_state;
3131 sock_hold(sk);
3132 sock_orphan(sk);
3133
3134 local_bh_disable();
3135 bh_lock_sock(sk);
3136 /* remove backlog if any, without releasing ownership. */
3137 __release_sock(sk);
3138
3139 this_cpu_inc(tcp_orphan_count);
3140
3141 /* Have we already been destroyed by a softirq or backlog? */
3142 if (state != TCP_CLOSE && sk->sk_state == TCP_CLOSE)
3143 goto out;
3144
3145 /* This is a (useful) BSD violating of the RFC. There is a
3146 * problem with TCP as specified in that the other end could
3147 * keep a socket open forever with no application left this end.
3148 * We use a 1 minute timeout (about the same as BSD) then kill
3149 * our end. If they send after that then tough - BUT: long enough
3150 * that we won't make the old 4*rto = almost no time - whoops
3151 * reset mistake.
3152 *
3153 * Nope, it was not mistake. It is really desired behaviour
3154 * f.e. on http servers, when such sockets are useless, but
3155 * consume significant resources. Let's do it with special
3156 * linger2 option. --ANK
3157 */
3158
3159 if (sk->sk_state == TCP_FIN_WAIT2) {
3160 struct tcp_sock *tp = tcp_sk(sk);
3161 if (READ_ONCE(tp->linger2) < 0) {
3162 tcp_set_state(sk, TCP_CLOSE);
3163 tcp_send_active_reset(sk, GFP_ATOMIC,
3164 SK_RST_REASON_TCP_ABORT_ON_LINGER);
3165 __NET_INC_STATS(sock_net(sk),
3166 LINUX_MIB_TCPABORTONLINGER);
3167 } else {
3168 const int tmo = tcp_fin_time(sk);
3169
3170 if (tmo > TCP_TIMEWAIT_LEN) {
3171 inet_csk_reset_keepalive_timer(sk,
3172 tmo - TCP_TIMEWAIT_LEN);
3173 } else {
3174 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
3175 goto out;
3176 }
3177 }
3178 }
3179 if (sk->sk_state != TCP_CLOSE) {
3180 if (tcp_check_oom(sk, 0)) {
3181 tcp_set_state(sk, TCP_CLOSE);
3182 tcp_send_active_reset(sk, GFP_ATOMIC,
3183 SK_RST_REASON_TCP_ABORT_ON_MEMORY);
3184 __NET_INC_STATS(sock_net(sk),
3185 LINUX_MIB_TCPABORTONMEMORY);
3186 } else if (!check_net(sock_net(sk))) {
3187 /* Not possible to send reset; just close */
3188 tcp_set_state(sk, TCP_CLOSE);
3189 }
3190 }
3191
3192 if (sk->sk_state == TCP_CLOSE) {
3193 struct request_sock *req;
3194
3195 req = rcu_dereference_protected(tcp_sk(sk)->fastopen_rsk,
3196 lockdep_sock_is_held(sk));
3197 /* We could get here with a non-NULL req if the socket is
3198 * aborted (e.g., closed with unread data) before 3WHS
3199 * finishes.
3200 */
3201 if (req)
3202 reqsk_fastopen_remove(sk, req, false);
3203 inet_csk_destroy_sock(sk);
3204 }
3205 /* Otherwise, socket is reprieved until protocol close. */
3206
3207out:
3208 bh_unlock_sock(sk);
3209 local_bh_enable();
3210}
3211
3212void tcp_close(struct sock *sk, long timeout)
3213{
3214 lock_sock(sk);
3215 __tcp_close(sk, timeout);
3216 release_sock(sk);
3217 if (!sk->sk_net_refcnt)
3218 inet_csk_clear_xmit_timers_sync(sk);
3219 sock_put(sk);
3220}
3221EXPORT_SYMBOL(tcp_close);
3222
3223/* These states need RST on ABORT according to RFC793 */
3224
3225static inline bool tcp_need_reset(int state)
3226{
3227 return (1 << state) &
3228 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 |
3229 TCPF_FIN_WAIT2 | TCPF_SYN_RECV);
3230}
3231
3232static void tcp_rtx_queue_purge(struct sock *sk)
3233{
3234 struct rb_node *p = rb_first(&sk->tcp_rtx_queue);
3235
3236 tcp_sk(sk)->highest_sack = NULL;
3237 while (p) {
3238 struct sk_buff *skb = rb_to_skb(p);
3239
3240 p = rb_next(p);
3241 /* Since we are deleting whole queue, no need to
3242 * list_del(&skb->tcp_tsorted_anchor)
3243 */
3244 tcp_rtx_queue_unlink(skb, sk);
3245 tcp_wmem_free_skb(sk, skb);
3246 }
3247}
3248
3249void tcp_write_queue_purge(struct sock *sk)
3250{
3251 struct sk_buff *skb;
3252
3253 tcp_chrono_stop(sk, TCP_CHRONO_BUSY);
3254 while ((skb = __skb_dequeue(&sk->sk_write_queue)) != NULL) {
3255 tcp_skb_tsorted_anchor_cleanup(skb);
3256 tcp_wmem_free_skb(sk, skb);
3257 }
3258 tcp_rtx_queue_purge(sk);
3259 INIT_LIST_HEAD(&tcp_sk(sk)->tsorted_sent_queue);
3260 tcp_clear_all_retrans_hints(tcp_sk(sk));
3261 tcp_sk(sk)->packets_out = 0;
3262 inet_csk(sk)->icsk_backoff = 0;
3263}
3264
3265int tcp_disconnect(struct sock *sk, int flags)
3266{
3267 struct inet_sock *inet = inet_sk(sk);
3268 struct inet_connection_sock *icsk = inet_csk(sk);
3269 struct tcp_sock *tp = tcp_sk(sk);
3270 int old_state = sk->sk_state;
3271 u32 seq;
3272
3273 if (old_state != TCP_CLOSE)
3274 tcp_set_state(sk, TCP_CLOSE);
3275
3276 /* ABORT function of RFC793 */
3277 if (old_state == TCP_LISTEN) {
3278 inet_csk_listen_stop(sk);
3279 } else if (unlikely(tp->repair)) {
3280 WRITE_ONCE(sk->sk_err, ECONNABORTED);
3281 } else if (tcp_need_reset(old_state)) {
3282 tcp_send_active_reset(sk, gfp_any(), SK_RST_REASON_TCP_STATE);
3283 WRITE_ONCE(sk->sk_err, ECONNRESET);
3284 } else if (tp->snd_nxt != tp->write_seq &&
3285 (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK)) {
3286 /* The last check adjusts for discrepancy of Linux wrt. RFC
3287 * states
3288 */
3289 tcp_send_active_reset(sk, gfp_any(),
3290 SK_RST_REASON_TCP_DISCONNECT_WITH_DATA);
3291 WRITE_ONCE(sk->sk_err, ECONNRESET);
3292 } else if (old_state == TCP_SYN_SENT)
3293 WRITE_ONCE(sk->sk_err, ECONNRESET);
3294
3295 tcp_clear_xmit_timers(sk);
3296 __skb_queue_purge(&sk->sk_receive_queue);
3297 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt);
3298 WRITE_ONCE(tp->urg_data, 0);
3299 sk_set_peek_off(sk, -1);
3300 tcp_write_queue_purge(sk);
3301 tcp_fastopen_active_disable_ofo_check(sk);
3302 skb_rbtree_purge(&tp->out_of_order_queue);
3303
3304 inet->inet_dport = 0;
3305
3306 inet_bhash2_reset_saddr(sk);
3307
3308 WRITE_ONCE(sk->sk_shutdown, 0);
3309 sock_reset_flag(sk, SOCK_DONE);
3310 tp->srtt_us = 0;
3311 tp->mdev_us = jiffies_to_usecs(TCP_TIMEOUT_INIT);
3312 tp->rcv_rtt_last_tsecr = 0;
3313
3314 seq = tp->write_seq + tp->max_window + 2;
3315 if (!seq)
3316 seq = 1;
3317 WRITE_ONCE(tp->write_seq, seq);
3318
3319 icsk->icsk_backoff = 0;
3320 icsk->icsk_probes_out = 0;
3321 icsk->icsk_probes_tstamp = 0;
3322 icsk->icsk_rto = TCP_TIMEOUT_INIT;
3323 icsk->icsk_rto_min = TCP_RTO_MIN;
3324 icsk->icsk_delack_max = TCP_DELACK_MAX;
3325 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
3326 tcp_snd_cwnd_set(tp, TCP_INIT_CWND);
3327 tp->snd_cwnd_cnt = 0;
3328 tp->is_cwnd_limited = 0;
3329 tp->max_packets_out = 0;
3330 tp->window_clamp = 0;
3331 tp->delivered = 0;
3332 tp->delivered_ce = 0;
3333 if (icsk->icsk_ca_initialized && icsk->icsk_ca_ops->release)
3334 icsk->icsk_ca_ops->release(sk);
3335 memset(icsk->icsk_ca_priv, 0, sizeof(icsk->icsk_ca_priv));
3336 icsk->icsk_ca_initialized = 0;
3337 tcp_set_ca_state(sk, TCP_CA_Open);
3338 tp->is_sack_reneg = 0;
3339 tcp_clear_retrans(tp);
3340 tp->total_retrans = 0;
3341 inet_csk_delack_init(sk);
3342 /* Initialize rcv_mss to TCP_MIN_MSS to avoid division by 0
3343 * issue in __tcp_select_window()
3344 */
3345 icsk->icsk_ack.rcv_mss = TCP_MIN_MSS;
3346 memset(&tp->rx_opt, 0, sizeof(tp->rx_opt));
3347 __sk_dst_reset(sk);
3348 dst_release(unrcu_pointer(xchg(&sk->sk_rx_dst, NULL)));
3349 tcp_saved_syn_free(tp);
3350 tp->compressed_ack = 0;
3351 tp->segs_in = 0;
3352 tp->segs_out = 0;
3353 tp->bytes_sent = 0;
3354 tp->bytes_acked = 0;
3355 tp->bytes_received = 0;
3356 tp->bytes_retrans = 0;
3357 tp->data_segs_in = 0;
3358 tp->data_segs_out = 0;
3359 tp->duplicate_sack[0].start_seq = 0;
3360 tp->duplicate_sack[0].end_seq = 0;
3361 tp->dsack_dups = 0;
3362 tp->reord_seen = 0;
3363 tp->retrans_out = 0;
3364 tp->sacked_out = 0;
3365 tp->tlp_high_seq = 0;
3366 tp->last_oow_ack_time = 0;
3367 tp->plb_rehash = 0;
3368 /* There's a bubble in the pipe until at least the first ACK. */
3369 tp->app_limited = ~0U;
3370 tp->rate_app_limited = 1;
3371 tp->rack.mstamp = 0;
3372 tp->rack.advanced = 0;
3373 tp->rack.reo_wnd_steps = 1;
3374 tp->rack.last_delivered = 0;
3375 tp->rack.reo_wnd_persist = 0;
3376 tp->rack.dsack_seen = 0;
3377 tp->syn_data_acked = 0;
3378 tp->rx_opt.saw_tstamp = 0;
3379 tp->rx_opt.dsack = 0;
3380 tp->rx_opt.num_sacks = 0;
3381 tp->rcv_ooopack = 0;
3382
3383
3384 /* Clean up fastopen related fields */
3385 tcp_free_fastopen_req(tp);
3386 inet_clear_bit(DEFER_CONNECT, sk);
3387 tp->fastopen_client_fail = 0;
3388
3389 WARN_ON(inet->inet_num && !icsk->icsk_bind_hash);
3390
3391 if (sk->sk_frag.page) {
3392 put_page(sk->sk_frag.page);
3393 sk->sk_frag.page = NULL;
3394 sk->sk_frag.offset = 0;
3395 }
3396 sk_error_report(sk);
3397 return 0;
3398}
3399EXPORT_SYMBOL(tcp_disconnect);
3400
3401static inline bool tcp_can_repair_sock(const struct sock *sk)
3402{
3403 return sockopt_ns_capable(sock_net(sk)->user_ns, CAP_NET_ADMIN) &&
3404 (sk->sk_state != TCP_LISTEN);
3405}
3406
3407static int tcp_repair_set_window(struct tcp_sock *tp, sockptr_t optbuf, int len)
3408{
3409 struct tcp_repair_window opt;
3410
3411 if (!tp->repair)
3412 return -EPERM;
3413
3414 if (len != sizeof(opt))
3415 return -EINVAL;
3416
3417 if (copy_from_sockptr(&opt, optbuf, sizeof(opt)))
3418 return -EFAULT;
3419
3420 if (opt.max_window < opt.snd_wnd)
3421 return -EINVAL;
3422
3423 if (after(opt.snd_wl1, tp->rcv_nxt + opt.rcv_wnd))
3424 return -EINVAL;
3425
3426 if (after(opt.rcv_wup, tp->rcv_nxt))
3427 return -EINVAL;
3428
3429 tp->snd_wl1 = opt.snd_wl1;
3430 tp->snd_wnd = opt.snd_wnd;
3431 tp->max_window = opt.max_window;
3432
3433 tp->rcv_wnd = opt.rcv_wnd;
3434 tp->rcv_wup = opt.rcv_wup;
3435
3436 return 0;
3437}
3438
3439static int tcp_repair_options_est(struct sock *sk, sockptr_t optbuf,
3440 unsigned int len)
3441{
3442 struct tcp_sock *tp = tcp_sk(sk);
3443 struct tcp_repair_opt opt;
3444 size_t offset = 0;
3445
3446 while (len >= sizeof(opt)) {
3447 if (copy_from_sockptr_offset(&opt, optbuf, offset, sizeof(opt)))
3448 return -EFAULT;
3449
3450 offset += sizeof(opt);
3451 len -= sizeof(opt);
3452
3453 switch (opt.opt_code) {
3454 case TCPOPT_MSS:
3455 tp->rx_opt.mss_clamp = opt.opt_val;
3456 tcp_mtup_init(sk);
3457 break;
3458 case TCPOPT_WINDOW:
3459 {
3460 u16 snd_wscale = opt.opt_val & 0xFFFF;
3461 u16 rcv_wscale = opt.opt_val >> 16;
3462
3463 if (snd_wscale > TCP_MAX_WSCALE || rcv_wscale > TCP_MAX_WSCALE)
3464 return -EFBIG;
3465
3466 tp->rx_opt.snd_wscale = snd_wscale;
3467 tp->rx_opt.rcv_wscale = rcv_wscale;
3468 tp->rx_opt.wscale_ok = 1;
3469 }
3470 break;
3471 case TCPOPT_SACK_PERM:
3472 if (opt.opt_val != 0)
3473 return -EINVAL;
3474
3475 tp->rx_opt.sack_ok |= TCP_SACK_SEEN;
3476 break;
3477 case TCPOPT_TIMESTAMP:
3478 if (opt.opt_val != 0)
3479 return -EINVAL;
3480
3481 tp->rx_opt.tstamp_ok = 1;
3482 break;
3483 }
3484 }
3485
3486 return 0;
3487}
3488
3489DEFINE_STATIC_KEY_FALSE(tcp_tx_delay_enabled);
3490EXPORT_SYMBOL(tcp_tx_delay_enabled);
3491
3492static void tcp_enable_tx_delay(void)
3493{
3494 if (!static_branch_unlikely(&tcp_tx_delay_enabled)) {
3495 static int __tcp_tx_delay_enabled = 0;
3496
3497 if (cmpxchg(&__tcp_tx_delay_enabled, 0, 1) == 0) {
3498 static_branch_enable(&tcp_tx_delay_enabled);
3499 pr_info("TCP_TX_DELAY enabled\n");
3500 }
3501 }
3502}
3503
3504/* When set indicates to always queue non-full frames. Later the user clears
3505 * this option and we transmit any pending partial frames in the queue. This is
3506 * meant to be used alongside sendfile() to get properly filled frames when the
3507 * user (for example) must write out headers with a write() call first and then
3508 * use sendfile to send out the data parts.
3509 *
3510 * TCP_CORK can be set together with TCP_NODELAY and it is stronger than
3511 * TCP_NODELAY.
3512 */
3513void __tcp_sock_set_cork(struct sock *sk, bool on)
3514{
3515 struct tcp_sock *tp = tcp_sk(sk);
3516
3517 if (on) {
3518 tp->nonagle |= TCP_NAGLE_CORK;
3519 } else {
3520 tp->nonagle &= ~TCP_NAGLE_CORK;
3521 if (tp->nonagle & TCP_NAGLE_OFF)
3522 tp->nonagle |= TCP_NAGLE_PUSH;
3523 tcp_push_pending_frames(sk);
3524 }
3525}
3526
3527void tcp_sock_set_cork(struct sock *sk, bool on)
3528{
3529 lock_sock(sk);
3530 __tcp_sock_set_cork(sk, on);
3531 release_sock(sk);
3532}
3533EXPORT_SYMBOL(tcp_sock_set_cork);
3534
3535/* TCP_NODELAY is weaker than TCP_CORK, so that this option on corked socket is
3536 * remembered, but it is not activated until cork is cleared.
3537 *
3538 * However, when TCP_NODELAY is set we make an explicit push, which overrides
3539 * even TCP_CORK for currently queued segments.
3540 */
3541void __tcp_sock_set_nodelay(struct sock *sk, bool on)
3542{
3543 if (on) {
3544 tcp_sk(sk)->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH;
3545 tcp_push_pending_frames(sk);
3546 } else {
3547 tcp_sk(sk)->nonagle &= ~TCP_NAGLE_OFF;
3548 }
3549}
3550
3551void tcp_sock_set_nodelay(struct sock *sk)
3552{
3553 lock_sock(sk);
3554 __tcp_sock_set_nodelay(sk, true);
3555 release_sock(sk);
3556}
3557EXPORT_SYMBOL(tcp_sock_set_nodelay);
3558
3559static void __tcp_sock_set_quickack(struct sock *sk, int val)
3560{
3561 if (!val) {
3562 inet_csk_enter_pingpong_mode(sk);
3563 return;
3564 }
3565
3566 inet_csk_exit_pingpong_mode(sk);
3567 if ((1 << sk->sk_state) & (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) &&
3568 inet_csk_ack_scheduled(sk)) {
3569 inet_csk(sk)->icsk_ack.pending |= ICSK_ACK_PUSHED;
3570 tcp_cleanup_rbuf(sk, 1);
3571 if (!(val & 1))
3572 inet_csk_enter_pingpong_mode(sk);
3573 }
3574}
3575
3576void tcp_sock_set_quickack(struct sock *sk, int val)
3577{
3578 lock_sock(sk);
3579 __tcp_sock_set_quickack(sk, val);
3580 release_sock(sk);
3581}
3582EXPORT_SYMBOL(tcp_sock_set_quickack);
3583
3584int tcp_sock_set_syncnt(struct sock *sk, int val)
3585{
3586 if (val < 1 || val > MAX_TCP_SYNCNT)
3587 return -EINVAL;
3588
3589 WRITE_ONCE(inet_csk(sk)->icsk_syn_retries, val);
3590 return 0;
3591}
3592EXPORT_SYMBOL(tcp_sock_set_syncnt);
3593
3594int tcp_sock_set_user_timeout(struct sock *sk, int val)
3595{
3596 /* Cap the max time in ms TCP will retry or probe the window
3597 * before giving up and aborting (ETIMEDOUT) a connection.
3598 */
3599 if (val < 0)
3600 return -EINVAL;
3601
3602 WRITE_ONCE(inet_csk(sk)->icsk_user_timeout, val);
3603 return 0;
3604}
3605EXPORT_SYMBOL(tcp_sock_set_user_timeout);
3606
3607int tcp_sock_set_keepidle_locked(struct sock *sk, int val)
3608{
3609 struct tcp_sock *tp = tcp_sk(sk);
3610
3611 if (val < 1 || val > MAX_TCP_KEEPIDLE)
3612 return -EINVAL;
3613
3614 /* Paired with WRITE_ONCE() in keepalive_time_when() */
3615 WRITE_ONCE(tp->keepalive_time, val * HZ);
3616 if (sock_flag(sk, SOCK_KEEPOPEN) &&
3617 !((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN))) {
3618 u32 elapsed = keepalive_time_elapsed(tp);
3619
3620 if (tp->keepalive_time > elapsed)
3621 elapsed = tp->keepalive_time - elapsed;
3622 else
3623 elapsed = 0;
3624 inet_csk_reset_keepalive_timer(sk, elapsed);
3625 }
3626
3627 return 0;
3628}
3629
3630int tcp_sock_set_keepidle(struct sock *sk, int val)
3631{
3632 int err;
3633
3634 lock_sock(sk);
3635 err = tcp_sock_set_keepidle_locked(sk, val);
3636 release_sock(sk);
3637 return err;
3638}
3639EXPORT_SYMBOL(tcp_sock_set_keepidle);
3640
3641int tcp_sock_set_keepintvl(struct sock *sk, int val)
3642{
3643 if (val < 1 || val > MAX_TCP_KEEPINTVL)
3644 return -EINVAL;
3645
3646 WRITE_ONCE(tcp_sk(sk)->keepalive_intvl, val * HZ);
3647 return 0;
3648}
3649EXPORT_SYMBOL(tcp_sock_set_keepintvl);
3650
3651int tcp_sock_set_keepcnt(struct sock *sk, int val)
3652{
3653 if (val < 1 || val > MAX_TCP_KEEPCNT)
3654 return -EINVAL;
3655
3656 /* Paired with READ_ONCE() in keepalive_probes() */
3657 WRITE_ONCE(tcp_sk(sk)->keepalive_probes, val);
3658 return 0;
3659}
3660EXPORT_SYMBOL(tcp_sock_set_keepcnt);
3661
3662int tcp_set_window_clamp(struct sock *sk, int val)
3663{
3664 struct tcp_sock *tp = tcp_sk(sk);
3665
3666 if (!val) {
3667 if (sk->sk_state != TCP_CLOSE)
3668 return -EINVAL;
3669 WRITE_ONCE(tp->window_clamp, 0);
3670 } else {
3671 u32 new_rcv_ssthresh, old_window_clamp = tp->window_clamp;
3672 u32 new_window_clamp = val < SOCK_MIN_RCVBUF / 2 ?
3673 SOCK_MIN_RCVBUF / 2 : val;
3674
3675 if (new_window_clamp == old_window_clamp)
3676 return 0;
3677
3678 WRITE_ONCE(tp->window_clamp, new_window_clamp);
3679 if (new_window_clamp < old_window_clamp) {
3680 /* need to apply the reserved mem provisioning only
3681 * when shrinking the window clamp
3682 */
3683 __tcp_adjust_rcv_ssthresh(sk, tp->window_clamp);
3684
3685 } else {
3686 new_rcv_ssthresh = min(tp->rcv_wnd, tp->window_clamp);
3687 tp->rcv_ssthresh = max(new_rcv_ssthresh,
3688 tp->rcv_ssthresh);
3689 }
3690 }
3691 return 0;
3692}
3693
3694/*
3695 * Socket option code for TCP.
3696 */
3697int do_tcp_setsockopt(struct sock *sk, int level, int optname,
3698 sockptr_t optval, unsigned int optlen)
3699{
3700 struct tcp_sock *tp = tcp_sk(sk);
3701 struct inet_connection_sock *icsk = inet_csk(sk);
3702 struct net *net = sock_net(sk);
3703 int val;
3704 int err = 0;
3705
3706 /* These are data/string values, all the others are ints */
3707 switch (optname) {
3708 case TCP_CONGESTION: {
3709 char name[TCP_CA_NAME_MAX];
3710
3711 if (optlen < 1)
3712 return -EINVAL;
3713
3714 val = strncpy_from_sockptr(name, optval,
3715 min_t(long, TCP_CA_NAME_MAX-1, optlen));
3716 if (val < 0)
3717 return -EFAULT;
3718 name[val] = 0;
3719
3720 sockopt_lock_sock(sk);
3721 err = tcp_set_congestion_control(sk, name, !has_current_bpf_ctx(),
3722 sockopt_ns_capable(sock_net(sk)->user_ns,
3723 CAP_NET_ADMIN));
3724 sockopt_release_sock(sk);
3725 return err;
3726 }
3727 case TCP_ULP: {
3728 char name[TCP_ULP_NAME_MAX];
3729
3730 if (optlen < 1)
3731 return -EINVAL;
3732
3733 val = strncpy_from_sockptr(name, optval,
3734 min_t(long, TCP_ULP_NAME_MAX - 1,
3735 optlen));
3736 if (val < 0)
3737 return -EFAULT;
3738 name[val] = 0;
3739
3740 sockopt_lock_sock(sk);
3741 err = tcp_set_ulp(sk, name);
3742 sockopt_release_sock(sk);
3743 return err;
3744 }
3745 case TCP_FASTOPEN_KEY: {
3746 __u8 key[TCP_FASTOPEN_KEY_BUF_LENGTH];
3747 __u8 *backup_key = NULL;
3748
3749 /* Allow a backup key as well to facilitate key rotation
3750 * First key is the active one.
3751 */
3752 if (optlen != TCP_FASTOPEN_KEY_LENGTH &&
3753 optlen != TCP_FASTOPEN_KEY_BUF_LENGTH)
3754 return -EINVAL;
3755
3756 if (copy_from_sockptr(key, optval, optlen))
3757 return -EFAULT;
3758
3759 if (optlen == TCP_FASTOPEN_KEY_BUF_LENGTH)
3760 backup_key = key + TCP_FASTOPEN_KEY_LENGTH;
3761
3762 return tcp_fastopen_reset_cipher(net, sk, key, backup_key);
3763 }
3764 default:
3765 /* fallthru */
3766 break;
3767 }
3768
3769 if (optlen < sizeof(int))
3770 return -EINVAL;
3771
3772 if (copy_from_sockptr(&val, optval, sizeof(val)))
3773 return -EFAULT;
3774
3775 /* Handle options that can be set without locking the socket. */
3776 switch (optname) {
3777 case TCP_SYNCNT:
3778 return tcp_sock_set_syncnt(sk, val);
3779 case TCP_USER_TIMEOUT:
3780 return tcp_sock_set_user_timeout(sk, val);
3781 case TCP_KEEPINTVL:
3782 return tcp_sock_set_keepintvl(sk, val);
3783 case TCP_KEEPCNT:
3784 return tcp_sock_set_keepcnt(sk, val);
3785 case TCP_LINGER2:
3786 if (val < 0)
3787 WRITE_ONCE(tp->linger2, -1);
3788 else if (val > TCP_FIN_TIMEOUT_MAX / HZ)
3789 WRITE_ONCE(tp->linger2, TCP_FIN_TIMEOUT_MAX);
3790 else
3791 WRITE_ONCE(tp->linger2, val * HZ);
3792 return 0;
3793 case TCP_DEFER_ACCEPT:
3794 /* Translate value in seconds to number of retransmits */
3795 WRITE_ONCE(icsk->icsk_accept_queue.rskq_defer_accept,
3796 secs_to_retrans(val, TCP_TIMEOUT_INIT / HZ,
3797 TCP_RTO_MAX / HZ));
3798 return 0;
3799 }
3800
3801 sockopt_lock_sock(sk);
3802
3803 switch (optname) {
3804 case TCP_MAXSEG:
3805 /* Values greater than interface MTU won't take effect. However
3806 * at the point when this call is done we typically don't yet
3807 * know which interface is going to be used
3808 */
3809 if (val && (val < TCP_MIN_MSS || val > MAX_TCP_WINDOW)) {
3810 err = -EINVAL;
3811 break;
3812 }
3813 tp->rx_opt.user_mss = val;
3814 break;
3815
3816 case TCP_NODELAY:
3817 __tcp_sock_set_nodelay(sk, val);
3818 break;
3819
3820 case TCP_THIN_LINEAR_TIMEOUTS:
3821 if (val < 0 || val > 1)
3822 err = -EINVAL;
3823 else
3824 tp->thin_lto = val;
3825 break;
3826
3827 case TCP_THIN_DUPACK:
3828 if (val < 0 || val > 1)
3829 err = -EINVAL;
3830 break;
3831
3832 case TCP_REPAIR:
3833 if (!tcp_can_repair_sock(sk))
3834 err = -EPERM;
3835 else if (val == TCP_REPAIR_ON) {
3836 tp->repair = 1;
3837 sk->sk_reuse = SK_FORCE_REUSE;
3838 tp->repair_queue = TCP_NO_QUEUE;
3839 } else if (val == TCP_REPAIR_OFF) {
3840 tp->repair = 0;
3841 sk->sk_reuse = SK_NO_REUSE;
3842 tcp_send_window_probe(sk);
3843 } else if (val == TCP_REPAIR_OFF_NO_WP) {
3844 tp->repair = 0;
3845 sk->sk_reuse = SK_NO_REUSE;
3846 } else
3847 err = -EINVAL;
3848
3849 break;
3850
3851 case TCP_REPAIR_QUEUE:
3852 if (!tp->repair)
3853 err = -EPERM;
3854 else if ((unsigned int)val < TCP_QUEUES_NR)
3855 tp->repair_queue = val;
3856 else
3857 err = -EINVAL;
3858 break;
3859
3860 case TCP_QUEUE_SEQ:
3861 if (sk->sk_state != TCP_CLOSE) {
3862 err = -EPERM;
3863 } else if (tp->repair_queue == TCP_SEND_QUEUE) {
3864 if (!tcp_rtx_queue_empty(sk))
3865 err = -EPERM;
3866 else
3867 WRITE_ONCE(tp->write_seq, val);
3868 } else if (tp->repair_queue == TCP_RECV_QUEUE) {
3869 if (tp->rcv_nxt != tp->copied_seq) {
3870 err = -EPERM;
3871 } else {
3872 WRITE_ONCE(tp->rcv_nxt, val);
3873 WRITE_ONCE(tp->copied_seq, val);
3874 }
3875 } else {
3876 err = -EINVAL;
3877 }
3878 break;
3879
3880 case TCP_REPAIR_OPTIONS:
3881 if (!tp->repair)
3882 err = -EINVAL;
3883 else if (sk->sk_state == TCP_ESTABLISHED && !tp->bytes_sent)
3884 err = tcp_repair_options_est(sk, optval, optlen);
3885 else
3886 err = -EPERM;
3887 break;
3888
3889 case TCP_CORK:
3890 __tcp_sock_set_cork(sk, val);
3891 break;
3892
3893 case TCP_KEEPIDLE:
3894 err = tcp_sock_set_keepidle_locked(sk, val);
3895 break;
3896 case TCP_SAVE_SYN:
3897 /* 0: disable, 1: enable, 2: start from ether_header */
3898 if (val < 0 || val > 2)
3899 err = -EINVAL;
3900 else
3901 tp->save_syn = val;
3902 break;
3903
3904 case TCP_WINDOW_CLAMP:
3905 err = tcp_set_window_clamp(sk, val);
3906 break;
3907
3908 case TCP_QUICKACK:
3909 __tcp_sock_set_quickack(sk, val);
3910 break;
3911
3912 case TCP_AO_REPAIR:
3913 if (!tcp_can_repair_sock(sk)) {
3914 err = -EPERM;
3915 break;
3916 }
3917 err = tcp_ao_set_repair(sk, optval, optlen);
3918 break;
3919#ifdef CONFIG_TCP_AO
3920 case TCP_AO_ADD_KEY:
3921 case TCP_AO_DEL_KEY:
3922 case TCP_AO_INFO: {
3923 /* If this is the first TCP-AO setsockopt() on the socket,
3924 * sk_state has to be LISTEN or CLOSE. Allow TCP_REPAIR
3925 * in any state.
3926 */
3927 if ((1 << sk->sk_state) & (TCPF_LISTEN | TCPF_CLOSE))
3928 goto ao_parse;
3929 if (rcu_dereference_protected(tcp_sk(sk)->ao_info,
3930 lockdep_sock_is_held(sk)))
3931 goto ao_parse;
3932 if (tp->repair)
3933 goto ao_parse;
3934 err = -EISCONN;
3935 break;
3936ao_parse:
3937 err = tp->af_specific->ao_parse(sk, optname, optval, optlen);
3938 break;
3939 }
3940#endif
3941#ifdef CONFIG_TCP_MD5SIG
3942 case TCP_MD5SIG:
3943 case TCP_MD5SIG_EXT:
3944 err = tp->af_specific->md5_parse(sk, optname, optval, optlen);
3945 break;
3946#endif
3947 case TCP_FASTOPEN:
3948 if (val >= 0 && ((1 << sk->sk_state) & (TCPF_CLOSE |
3949 TCPF_LISTEN))) {
3950 tcp_fastopen_init_key_once(net);
3951
3952 fastopen_queue_tune(sk, val);
3953 } else {
3954 err = -EINVAL;
3955 }
3956 break;
3957 case TCP_FASTOPEN_CONNECT:
3958 if (val > 1 || val < 0) {
3959 err = -EINVAL;
3960 } else if (READ_ONCE(net->ipv4.sysctl_tcp_fastopen) &
3961 TFO_CLIENT_ENABLE) {
3962 if (sk->sk_state == TCP_CLOSE)
3963 tp->fastopen_connect = val;
3964 else
3965 err = -EINVAL;
3966 } else {
3967 err = -EOPNOTSUPP;
3968 }
3969 break;
3970 case TCP_FASTOPEN_NO_COOKIE:
3971 if (val > 1 || val < 0)
3972 err = -EINVAL;
3973 else if (!((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
3974 err = -EINVAL;
3975 else
3976 tp->fastopen_no_cookie = val;
3977 break;
3978 case TCP_TIMESTAMP:
3979 if (!tp->repair) {
3980 err = -EPERM;
3981 break;
3982 }
3983 /* val is an opaque field,
3984 * and low order bit contains usec_ts enable bit.
3985 * Its a best effort, and we do not care if user makes an error.
3986 */
3987 tp->tcp_usec_ts = val & 1;
3988 WRITE_ONCE(tp->tsoffset, val - tcp_clock_ts(tp->tcp_usec_ts));
3989 break;
3990 case TCP_REPAIR_WINDOW:
3991 err = tcp_repair_set_window(tp, optval, optlen);
3992 break;
3993 case TCP_NOTSENT_LOWAT:
3994 WRITE_ONCE(tp->notsent_lowat, val);
3995 sk->sk_write_space(sk);
3996 break;
3997 case TCP_INQ:
3998 if (val > 1 || val < 0)
3999 err = -EINVAL;
4000 else
4001 tp->recvmsg_inq = val;
4002 break;
4003 case TCP_TX_DELAY:
4004 if (val)
4005 tcp_enable_tx_delay();
4006 WRITE_ONCE(tp->tcp_tx_delay, val);
4007 break;
4008 default:
4009 err = -ENOPROTOOPT;
4010 break;
4011 }
4012
4013 sockopt_release_sock(sk);
4014 return err;
4015}
4016
4017int tcp_setsockopt(struct sock *sk, int level, int optname, sockptr_t optval,
4018 unsigned int optlen)
4019{
4020 const struct inet_connection_sock *icsk = inet_csk(sk);
4021
4022 if (level != SOL_TCP)
4023 /* Paired with WRITE_ONCE() in do_ipv6_setsockopt() and tcp_v6_connect() */
4024 return READ_ONCE(icsk->icsk_af_ops)->setsockopt(sk, level, optname,
4025 optval, optlen);
4026 return do_tcp_setsockopt(sk, level, optname, optval, optlen);
4027}
4028EXPORT_SYMBOL(tcp_setsockopt);
4029
4030static void tcp_get_info_chrono_stats(const struct tcp_sock *tp,
4031 struct tcp_info *info)
4032{
4033 u64 stats[__TCP_CHRONO_MAX], total = 0;
4034 enum tcp_chrono i;
4035
4036 for (i = TCP_CHRONO_BUSY; i < __TCP_CHRONO_MAX; ++i) {
4037 stats[i] = tp->chrono_stat[i - 1];
4038 if (i == tp->chrono_type)
4039 stats[i] += tcp_jiffies32 - tp->chrono_start;
4040 stats[i] *= USEC_PER_SEC / HZ;
4041 total += stats[i];
4042 }
4043
4044 info->tcpi_busy_time = total;
4045 info->tcpi_rwnd_limited = stats[TCP_CHRONO_RWND_LIMITED];
4046 info->tcpi_sndbuf_limited = stats[TCP_CHRONO_SNDBUF_LIMITED];
4047}
4048
4049/* Return information about state of tcp endpoint in API format. */
4050void tcp_get_info(struct sock *sk, struct tcp_info *info)
4051{
4052 const struct tcp_sock *tp = tcp_sk(sk); /* iff sk_type == SOCK_STREAM */
4053 const struct inet_connection_sock *icsk = inet_csk(sk);
4054 unsigned long rate;
4055 u32 now;
4056 u64 rate64;
4057 bool slow;
4058
4059 memset(info, 0, sizeof(*info));
4060 if (sk->sk_type != SOCK_STREAM)
4061 return;
4062
4063 info->tcpi_state = inet_sk_state_load(sk);
4064
4065 /* Report meaningful fields for all TCP states, including listeners */
4066 rate = READ_ONCE(sk->sk_pacing_rate);
4067 rate64 = (rate != ~0UL) ? rate : ~0ULL;
4068 info->tcpi_pacing_rate = rate64;
4069
4070 rate = READ_ONCE(sk->sk_max_pacing_rate);
4071 rate64 = (rate != ~0UL) ? rate : ~0ULL;
4072 info->tcpi_max_pacing_rate = rate64;
4073
4074 info->tcpi_reordering = tp->reordering;
4075 info->tcpi_snd_cwnd = tcp_snd_cwnd(tp);
4076
4077 if (info->tcpi_state == TCP_LISTEN) {
4078 /* listeners aliased fields :
4079 * tcpi_unacked -> Number of children ready for accept()
4080 * tcpi_sacked -> max backlog
4081 */
4082 info->tcpi_unacked = READ_ONCE(sk->sk_ack_backlog);
4083 info->tcpi_sacked = READ_ONCE(sk->sk_max_ack_backlog);
4084 return;
4085 }
4086
4087 slow = lock_sock_fast(sk);
4088
4089 info->tcpi_ca_state = icsk->icsk_ca_state;
4090 info->tcpi_retransmits = icsk->icsk_retransmits;
4091 info->tcpi_probes = icsk->icsk_probes_out;
4092 info->tcpi_backoff = icsk->icsk_backoff;
4093
4094 if (tp->rx_opt.tstamp_ok)
4095 info->tcpi_options |= TCPI_OPT_TIMESTAMPS;
4096 if (tcp_is_sack(tp))
4097 info->tcpi_options |= TCPI_OPT_SACK;
4098 if (tp->rx_opt.wscale_ok) {
4099 info->tcpi_options |= TCPI_OPT_WSCALE;
4100 info->tcpi_snd_wscale = tp->rx_opt.snd_wscale;
4101 info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale;
4102 }
4103
4104 if (tp->ecn_flags & TCP_ECN_OK)
4105 info->tcpi_options |= TCPI_OPT_ECN;
4106 if (tp->ecn_flags & TCP_ECN_SEEN)
4107 info->tcpi_options |= TCPI_OPT_ECN_SEEN;
4108 if (tp->syn_data_acked)
4109 info->tcpi_options |= TCPI_OPT_SYN_DATA;
4110 if (tp->tcp_usec_ts)
4111 info->tcpi_options |= TCPI_OPT_USEC_TS;
4112
4113 info->tcpi_rto = jiffies_to_usecs(icsk->icsk_rto);
4114 info->tcpi_ato = jiffies_to_usecs(min_t(u32, icsk->icsk_ack.ato,
4115 tcp_delack_max(sk)));
4116 info->tcpi_snd_mss = tp->mss_cache;
4117 info->tcpi_rcv_mss = icsk->icsk_ack.rcv_mss;
4118
4119 info->tcpi_unacked = tp->packets_out;
4120 info->tcpi_sacked = tp->sacked_out;
4121
4122 info->tcpi_lost = tp->lost_out;
4123 info->tcpi_retrans = tp->retrans_out;
4124
4125 now = tcp_jiffies32;
4126 info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime);
4127 info->tcpi_last_data_recv = jiffies_to_msecs(now - icsk->icsk_ack.lrcvtime);
4128 info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp);
4129
4130 info->tcpi_pmtu = icsk->icsk_pmtu_cookie;
4131 info->tcpi_rcv_ssthresh = tp->rcv_ssthresh;
4132 info->tcpi_rtt = tp->srtt_us >> 3;
4133 info->tcpi_rttvar = tp->mdev_us >> 2;
4134 info->tcpi_snd_ssthresh = tp->snd_ssthresh;
4135 info->tcpi_advmss = tp->advmss;
4136
4137 info->tcpi_rcv_rtt = tp->rcv_rtt_est.rtt_us >> 3;
4138 info->tcpi_rcv_space = tp->rcvq_space.space;
4139
4140 info->tcpi_total_retrans = tp->total_retrans;
4141
4142 info->tcpi_bytes_acked = tp->bytes_acked;
4143 info->tcpi_bytes_received = tp->bytes_received;
4144 info->tcpi_notsent_bytes = max_t(int, 0, tp->write_seq - tp->snd_nxt);
4145 tcp_get_info_chrono_stats(tp, info);
4146
4147 info->tcpi_segs_out = tp->segs_out;
4148
4149 /* segs_in and data_segs_in can be updated from tcp_segs_in() from BH */
4150 info->tcpi_segs_in = READ_ONCE(tp->segs_in);
4151 info->tcpi_data_segs_in = READ_ONCE(tp->data_segs_in);
4152
4153 info->tcpi_min_rtt = tcp_min_rtt(tp);
4154 info->tcpi_data_segs_out = tp->data_segs_out;
4155
4156 info->tcpi_delivery_rate_app_limited = tp->rate_app_limited ? 1 : 0;
4157 rate64 = tcp_compute_delivery_rate(tp);
4158 if (rate64)
4159 info->tcpi_delivery_rate = rate64;
4160 info->tcpi_delivered = tp->delivered;
4161 info->tcpi_delivered_ce = tp->delivered_ce;
4162 info->tcpi_bytes_sent = tp->bytes_sent;
4163 info->tcpi_bytes_retrans = tp->bytes_retrans;
4164 info->tcpi_dsack_dups = tp->dsack_dups;
4165 info->tcpi_reord_seen = tp->reord_seen;
4166 info->tcpi_rcv_ooopack = tp->rcv_ooopack;
4167 info->tcpi_snd_wnd = tp->snd_wnd;
4168 info->tcpi_rcv_wnd = tp->rcv_wnd;
4169 info->tcpi_rehash = tp->plb_rehash + tp->timeout_rehash;
4170 info->tcpi_fastopen_client_fail = tp->fastopen_client_fail;
4171
4172 info->tcpi_total_rto = tp->total_rto;
4173 info->tcpi_total_rto_recoveries = tp->total_rto_recoveries;
4174 info->tcpi_total_rto_time = tp->total_rto_time;
4175 if (tp->rto_stamp)
4176 info->tcpi_total_rto_time += tcp_clock_ms() - tp->rto_stamp;
4177
4178 unlock_sock_fast(sk, slow);
4179}
4180EXPORT_SYMBOL_GPL(tcp_get_info);
4181
4182static size_t tcp_opt_stats_get_size(void)
4183{
4184 return
4185 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_BUSY */
4186 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_RWND_LIMITED */
4187 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_SNDBUF_LIMITED */
4188 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_DATA_SEGS_OUT */
4189 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_TOTAL_RETRANS */
4190 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_PACING_RATE */
4191 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_DELIVERY_RATE */
4192 nla_total_size(sizeof(u32)) + /* TCP_NLA_SND_CWND */
4193 nla_total_size(sizeof(u32)) + /* TCP_NLA_REORDERING */
4194 nla_total_size(sizeof(u32)) + /* TCP_NLA_MIN_RTT */
4195 nla_total_size(sizeof(u8)) + /* TCP_NLA_RECUR_RETRANS */
4196 nla_total_size(sizeof(u8)) + /* TCP_NLA_DELIVERY_RATE_APP_LMT */
4197 nla_total_size(sizeof(u32)) + /* TCP_NLA_SNDQ_SIZE */
4198 nla_total_size(sizeof(u8)) + /* TCP_NLA_CA_STATE */
4199 nla_total_size(sizeof(u32)) + /* TCP_NLA_SND_SSTHRESH */
4200 nla_total_size(sizeof(u32)) + /* TCP_NLA_DELIVERED */
4201 nla_total_size(sizeof(u32)) + /* TCP_NLA_DELIVERED_CE */
4202 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_BYTES_SENT */
4203 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_BYTES_RETRANS */
4204 nla_total_size(sizeof(u32)) + /* TCP_NLA_DSACK_DUPS */
4205 nla_total_size(sizeof(u32)) + /* TCP_NLA_REORD_SEEN */
4206 nla_total_size(sizeof(u32)) + /* TCP_NLA_SRTT */
4207 nla_total_size(sizeof(u16)) + /* TCP_NLA_TIMEOUT_REHASH */
4208 nla_total_size(sizeof(u32)) + /* TCP_NLA_BYTES_NOTSENT */
4209 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_EDT */
4210 nla_total_size(sizeof(u8)) + /* TCP_NLA_TTL */
4211 nla_total_size(sizeof(u32)) + /* TCP_NLA_REHASH */
4212 0;
4213}
4214
4215/* Returns TTL or hop limit of an incoming packet from skb. */
4216static u8 tcp_skb_ttl_or_hop_limit(const struct sk_buff *skb)
4217{
4218 if (skb->protocol == htons(ETH_P_IP))
4219 return ip_hdr(skb)->ttl;
4220 else if (skb->protocol == htons(ETH_P_IPV6))
4221 return ipv6_hdr(skb)->hop_limit;
4222 else
4223 return 0;
4224}
4225
4226struct sk_buff *tcp_get_timestamping_opt_stats(const struct sock *sk,
4227 const struct sk_buff *orig_skb,
4228 const struct sk_buff *ack_skb)
4229{
4230 const struct tcp_sock *tp = tcp_sk(sk);
4231 struct sk_buff *stats;
4232 struct tcp_info info;
4233 unsigned long rate;
4234 u64 rate64;
4235
4236 stats = alloc_skb(tcp_opt_stats_get_size(), GFP_ATOMIC);
4237 if (!stats)
4238 return NULL;
4239
4240 tcp_get_info_chrono_stats(tp, &info);
4241 nla_put_u64_64bit(stats, TCP_NLA_BUSY,
4242 info.tcpi_busy_time, TCP_NLA_PAD);
4243 nla_put_u64_64bit(stats, TCP_NLA_RWND_LIMITED,
4244 info.tcpi_rwnd_limited, TCP_NLA_PAD);
4245 nla_put_u64_64bit(stats, TCP_NLA_SNDBUF_LIMITED,
4246 info.tcpi_sndbuf_limited, TCP_NLA_PAD);
4247 nla_put_u64_64bit(stats, TCP_NLA_DATA_SEGS_OUT,
4248 tp->data_segs_out, TCP_NLA_PAD);
4249 nla_put_u64_64bit(stats, TCP_NLA_TOTAL_RETRANS,
4250 tp->total_retrans, TCP_NLA_PAD);
4251
4252 rate = READ_ONCE(sk->sk_pacing_rate);
4253 rate64 = (rate != ~0UL) ? rate : ~0ULL;
4254 nla_put_u64_64bit(stats, TCP_NLA_PACING_RATE, rate64, TCP_NLA_PAD);
4255
4256 rate64 = tcp_compute_delivery_rate(tp);
4257 nla_put_u64_64bit(stats, TCP_NLA_DELIVERY_RATE, rate64, TCP_NLA_PAD);
4258
4259 nla_put_u32(stats, TCP_NLA_SND_CWND, tcp_snd_cwnd(tp));
4260 nla_put_u32(stats, TCP_NLA_REORDERING, tp->reordering);
4261 nla_put_u32(stats, TCP_NLA_MIN_RTT, tcp_min_rtt(tp));
4262
4263 nla_put_u8(stats, TCP_NLA_RECUR_RETRANS, inet_csk(sk)->icsk_retransmits);
4264 nla_put_u8(stats, TCP_NLA_DELIVERY_RATE_APP_LMT, !!tp->rate_app_limited);
4265 nla_put_u32(stats, TCP_NLA_SND_SSTHRESH, tp->snd_ssthresh);
4266 nla_put_u32(stats, TCP_NLA_DELIVERED, tp->delivered);
4267 nla_put_u32(stats, TCP_NLA_DELIVERED_CE, tp->delivered_ce);
4268
4269 nla_put_u32(stats, TCP_NLA_SNDQ_SIZE, tp->write_seq - tp->snd_una);
4270 nla_put_u8(stats, TCP_NLA_CA_STATE, inet_csk(sk)->icsk_ca_state);
4271
4272 nla_put_u64_64bit(stats, TCP_NLA_BYTES_SENT, tp->bytes_sent,
4273 TCP_NLA_PAD);
4274 nla_put_u64_64bit(stats, TCP_NLA_BYTES_RETRANS, tp->bytes_retrans,
4275 TCP_NLA_PAD);
4276 nla_put_u32(stats, TCP_NLA_DSACK_DUPS, tp->dsack_dups);
4277 nla_put_u32(stats, TCP_NLA_REORD_SEEN, tp->reord_seen);
4278 nla_put_u32(stats, TCP_NLA_SRTT, tp->srtt_us >> 3);
4279 nla_put_u16(stats, TCP_NLA_TIMEOUT_REHASH, tp->timeout_rehash);
4280 nla_put_u32(stats, TCP_NLA_BYTES_NOTSENT,
4281 max_t(int, 0, tp->write_seq - tp->snd_nxt));
4282 nla_put_u64_64bit(stats, TCP_NLA_EDT, orig_skb->skb_mstamp_ns,
4283 TCP_NLA_PAD);
4284 if (ack_skb)
4285 nla_put_u8(stats, TCP_NLA_TTL,
4286 tcp_skb_ttl_or_hop_limit(ack_skb));
4287
4288 nla_put_u32(stats, TCP_NLA_REHASH, tp->plb_rehash + tp->timeout_rehash);
4289 return stats;
4290}
4291
4292int do_tcp_getsockopt(struct sock *sk, int level,
4293 int optname, sockptr_t optval, sockptr_t optlen)
4294{
4295 struct inet_connection_sock *icsk = inet_csk(sk);
4296 struct tcp_sock *tp = tcp_sk(sk);
4297 struct net *net = sock_net(sk);
4298 int val, len;
4299
4300 if (copy_from_sockptr(&len, optlen, sizeof(int)))
4301 return -EFAULT;
4302
4303 if (len < 0)
4304 return -EINVAL;
4305
4306 len = min_t(unsigned int, len, sizeof(int));
4307
4308 switch (optname) {
4309 case TCP_MAXSEG:
4310 val = tp->mss_cache;
4311 if (tp->rx_opt.user_mss &&
4312 ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
4313 val = tp->rx_opt.user_mss;
4314 if (tp->repair)
4315 val = tp->rx_opt.mss_clamp;
4316 break;
4317 case TCP_NODELAY:
4318 val = !!(tp->nonagle&TCP_NAGLE_OFF);
4319 break;
4320 case TCP_CORK:
4321 val = !!(tp->nonagle&TCP_NAGLE_CORK);
4322 break;
4323 case TCP_KEEPIDLE:
4324 val = keepalive_time_when(tp) / HZ;
4325 break;
4326 case TCP_KEEPINTVL:
4327 val = keepalive_intvl_when(tp) / HZ;
4328 break;
4329 case TCP_KEEPCNT:
4330 val = keepalive_probes(tp);
4331 break;
4332 case TCP_SYNCNT:
4333 val = READ_ONCE(icsk->icsk_syn_retries) ? :
4334 READ_ONCE(net->ipv4.sysctl_tcp_syn_retries);
4335 break;
4336 case TCP_LINGER2:
4337 val = READ_ONCE(tp->linger2);
4338 if (val >= 0)
4339 val = (val ? : READ_ONCE(net->ipv4.sysctl_tcp_fin_timeout)) / HZ;
4340 break;
4341 case TCP_DEFER_ACCEPT:
4342 val = READ_ONCE(icsk->icsk_accept_queue.rskq_defer_accept);
4343 val = retrans_to_secs(val, TCP_TIMEOUT_INIT / HZ,
4344 TCP_RTO_MAX / HZ);
4345 break;
4346 case TCP_WINDOW_CLAMP:
4347 val = READ_ONCE(tp->window_clamp);
4348 break;
4349 case TCP_INFO: {
4350 struct tcp_info info;
4351
4352 if (copy_from_sockptr(&len, optlen, sizeof(int)))
4353 return -EFAULT;
4354
4355 tcp_get_info(sk, &info);
4356
4357 len = min_t(unsigned int, len, sizeof(info));
4358 if (copy_to_sockptr(optlen, &len, sizeof(int)))
4359 return -EFAULT;
4360 if (copy_to_sockptr(optval, &info, len))
4361 return -EFAULT;
4362 return 0;
4363 }
4364 case TCP_CC_INFO: {
4365 const struct tcp_congestion_ops *ca_ops;
4366 union tcp_cc_info info;
4367 size_t sz = 0;
4368 int attr;
4369
4370 if (copy_from_sockptr(&len, optlen, sizeof(int)))
4371 return -EFAULT;
4372
4373 ca_ops = icsk->icsk_ca_ops;
4374 if (ca_ops && ca_ops->get_info)
4375 sz = ca_ops->get_info(sk, ~0U, &attr, &info);
4376
4377 len = min_t(unsigned int, len, sz);
4378 if (copy_to_sockptr(optlen, &len, sizeof(int)))
4379 return -EFAULT;
4380 if (copy_to_sockptr(optval, &info, len))
4381 return -EFAULT;
4382 return 0;
4383 }
4384 case TCP_QUICKACK:
4385 val = !inet_csk_in_pingpong_mode(sk);
4386 break;
4387
4388 case TCP_CONGESTION:
4389 if (copy_from_sockptr(&len, optlen, sizeof(int)))
4390 return -EFAULT;
4391 len = min_t(unsigned int, len, TCP_CA_NAME_MAX);
4392 if (copy_to_sockptr(optlen, &len, sizeof(int)))
4393 return -EFAULT;
4394 if (copy_to_sockptr(optval, icsk->icsk_ca_ops->name, len))
4395 return -EFAULT;
4396 return 0;
4397
4398 case TCP_ULP:
4399 if (copy_from_sockptr(&len, optlen, sizeof(int)))
4400 return -EFAULT;
4401 len = min_t(unsigned int, len, TCP_ULP_NAME_MAX);
4402 if (!icsk->icsk_ulp_ops) {
4403 len = 0;
4404 if (copy_to_sockptr(optlen, &len, sizeof(int)))
4405 return -EFAULT;
4406 return 0;
4407 }
4408 if (copy_to_sockptr(optlen, &len, sizeof(int)))
4409 return -EFAULT;
4410 if (copy_to_sockptr(optval, icsk->icsk_ulp_ops->name, len))
4411 return -EFAULT;
4412 return 0;
4413
4414 case TCP_FASTOPEN_KEY: {
4415 u64 key[TCP_FASTOPEN_KEY_BUF_LENGTH / sizeof(u64)];
4416 unsigned int key_len;
4417
4418 if (copy_from_sockptr(&len, optlen, sizeof(int)))
4419 return -EFAULT;
4420
4421 key_len = tcp_fastopen_get_cipher(net, icsk, key) *
4422 TCP_FASTOPEN_KEY_LENGTH;
4423 len = min_t(unsigned int, len, key_len);
4424 if (copy_to_sockptr(optlen, &len, sizeof(int)))
4425 return -EFAULT;
4426 if (copy_to_sockptr(optval, key, len))
4427 return -EFAULT;
4428 return 0;
4429 }
4430 case TCP_THIN_LINEAR_TIMEOUTS:
4431 val = tp->thin_lto;
4432 break;
4433
4434 case TCP_THIN_DUPACK:
4435 val = 0;
4436 break;
4437
4438 case TCP_REPAIR:
4439 val = tp->repair;
4440 break;
4441
4442 case TCP_REPAIR_QUEUE:
4443 if (tp->repair)
4444 val = tp->repair_queue;
4445 else
4446 return -EINVAL;
4447 break;
4448
4449 case TCP_REPAIR_WINDOW: {
4450 struct tcp_repair_window opt;
4451
4452 if (copy_from_sockptr(&len, optlen, sizeof(int)))
4453 return -EFAULT;
4454
4455 if (len != sizeof(opt))
4456 return -EINVAL;
4457
4458 if (!tp->repair)
4459 return -EPERM;
4460
4461 opt.snd_wl1 = tp->snd_wl1;
4462 opt.snd_wnd = tp->snd_wnd;
4463 opt.max_window = tp->max_window;
4464 opt.rcv_wnd = tp->rcv_wnd;
4465 opt.rcv_wup = tp->rcv_wup;
4466
4467 if (copy_to_sockptr(optval, &opt, len))
4468 return -EFAULT;
4469 return 0;
4470 }
4471 case TCP_QUEUE_SEQ:
4472 if (tp->repair_queue == TCP_SEND_QUEUE)
4473 val = tp->write_seq;
4474 else if (tp->repair_queue == TCP_RECV_QUEUE)
4475 val = tp->rcv_nxt;
4476 else
4477 return -EINVAL;
4478 break;
4479
4480 case TCP_USER_TIMEOUT:
4481 val = READ_ONCE(icsk->icsk_user_timeout);
4482 break;
4483
4484 case TCP_FASTOPEN:
4485 val = READ_ONCE(icsk->icsk_accept_queue.fastopenq.max_qlen);
4486 break;
4487
4488 case TCP_FASTOPEN_CONNECT:
4489 val = tp->fastopen_connect;
4490 break;
4491
4492 case TCP_FASTOPEN_NO_COOKIE:
4493 val = tp->fastopen_no_cookie;
4494 break;
4495
4496 case TCP_TX_DELAY:
4497 val = READ_ONCE(tp->tcp_tx_delay);
4498 break;
4499
4500 case TCP_TIMESTAMP:
4501 val = tcp_clock_ts(tp->tcp_usec_ts) + READ_ONCE(tp->tsoffset);
4502 if (tp->tcp_usec_ts)
4503 val |= 1;
4504 else
4505 val &= ~1;
4506 break;
4507 case TCP_NOTSENT_LOWAT:
4508 val = READ_ONCE(tp->notsent_lowat);
4509 break;
4510 case TCP_INQ:
4511 val = tp->recvmsg_inq;
4512 break;
4513 case TCP_SAVE_SYN:
4514 val = tp->save_syn;
4515 break;
4516 case TCP_SAVED_SYN: {
4517 if (copy_from_sockptr(&len, optlen, sizeof(int)))
4518 return -EFAULT;
4519
4520 sockopt_lock_sock(sk);
4521 if (tp->saved_syn) {
4522 if (len < tcp_saved_syn_len(tp->saved_syn)) {
4523 len = tcp_saved_syn_len(tp->saved_syn);
4524 if (copy_to_sockptr(optlen, &len, sizeof(int))) {
4525 sockopt_release_sock(sk);
4526 return -EFAULT;
4527 }
4528 sockopt_release_sock(sk);
4529 return -EINVAL;
4530 }
4531 len = tcp_saved_syn_len(tp->saved_syn);
4532 if (copy_to_sockptr(optlen, &len, sizeof(int))) {
4533 sockopt_release_sock(sk);
4534 return -EFAULT;
4535 }
4536 if (copy_to_sockptr(optval, tp->saved_syn->data, len)) {
4537 sockopt_release_sock(sk);
4538 return -EFAULT;
4539 }
4540 tcp_saved_syn_free(tp);
4541 sockopt_release_sock(sk);
4542 } else {
4543 sockopt_release_sock(sk);
4544 len = 0;
4545 if (copy_to_sockptr(optlen, &len, sizeof(int)))
4546 return -EFAULT;
4547 }
4548 return 0;
4549 }
4550#ifdef CONFIG_MMU
4551 case TCP_ZEROCOPY_RECEIVE: {
4552 struct scm_timestamping_internal tss;
4553 struct tcp_zerocopy_receive zc = {};
4554 int err;
4555
4556 if (copy_from_sockptr(&len, optlen, sizeof(int)))
4557 return -EFAULT;
4558 if (len < 0 ||
4559 len < offsetofend(struct tcp_zerocopy_receive, length))
4560 return -EINVAL;
4561 if (unlikely(len > sizeof(zc))) {
4562 err = check_zeroed_sockptr(optval, sizeof(zc),
4563 len - sizeof(zc));
4564 if (err < 1)
4565 return err == 0 ? -EINVAL : err;
4566 len = sizeof(zc);
4567 if (copy_to_sockptr(optlen, &len, sizeof(int)))
4568 return -EFAULT;
4569 }
4570 if (copy_from_sockptr(&zc, optval, len))
4571 return -EFAULT;
4572 if (zc.reserved)
4573 return -EINVAL;
4574 if (zc.msg_flags & ~(TCP_VALID_ZC_MSG_FLAGS))
4575 return -EINVAL;
4576 sockopt_lock_sock(sk);
4577 err = tcp_zerocopy_receive(sk, &zc, &tss);
4578 err = BPF_CGROUP_RUN_PROG_GETSOCKOPT_KERN(sk, level, optname,
4579 &zc, &len, err);
4580 sockopt_release_sock(sk);
4581 if (len >= offsetofend(struct tcp_zerocopy_receive, msg_flags))
4582 goto zerocopy_rcv_cmsg;
4583 switch (len) {
4584 case offsetofend(struct tcp_zerocopy_receive, msg_flags):
4585 goto zerocopy_rcv_cmsg;
4586 case offsetofend(struct tcp_zerocopy_receive, msg_controllen):
4587 case offsetofend(struct tcp_zerocopy_receive, msg_control):
4588 case offsetofend(struct tcp_zerocopy_receive, flags):
4589 case offsetofend(struct tcp_zerocopy_receive, copybuf_len):
4590 case offsetofend(struct tcp_zerocopy_receive, copybuf_address):
4591 case offsetofend(struct tcp_zerocopy_receive, err):
4592 goto zerocopy_rcv_sk_err;
4593 case offsetofend(struct tcp_zerocopy_receive, inq):
4594 goto zerocopy_rcv_inq;
4595 case offsetofend(struct tcp_zerocopy_receive, length):
4596 default:
4597 goto zerocopy_rcv_out;
4598 }
4599zerocopy_rcv_cmsg:
4600 if (zc.msg_flags & TCP_CMSG_TS)
4601 tcp_zc_finalize_rx_tstamp(sk, &zc, &tss);
4602 else
4603 zc.msg_flags = 0;
4604zerocopy_rcv_sk_err:
4605 if (!err)
4606 zc.err = sock_error(sk);
4607zerocopy_rcv_inq:
4608 zc.inq = tcp_inq_hint(sk);
4609zerocopy_rcv_out:
4610 if (!err && copy_to_sockptr(optval, &zc, len))
4611 err = -EFAULT;
4612 return err;
4613 }
4614#endif
4615 case TCP_AO_REPAIR:
4616 if (!tcp_can_repair_sock(sk))
4617 return -EPERM;
4618 return tcp_ao_get_repair(sk, optval, optlen);
4619 case TCP_AO_GET_KEYS:
4620 case TCP_AO_INFO: {
4621 int err;
4622
4623 sockopt_lock_sock(sk);
4624 if (optname == TCP_AO_GET_KEYS)
4625 err = tcp_ao_get_mkts(sk, optval, optlen);
4626 else
4627 err = tcp_ao_get_sock_info(sk, optval, optlen);
4628 sockopt_release_sock(sk);
4629
4630 return err;
4631 }
4632 case TCP_IS_MPTCP:
4633 val = 0;
4634 break;
4635 default:
4636 return -ENOPROTOOPT;
4637 }
4638
4639 if (copy_to_sockptr(optlen, &len, sizeof(int)))
4640 return -EFAULT;
4641 if (copy_to_sockptr(optval, &val, len))
4642 return -EFAULT;
4643 return 0;
4644}
4645
4646bool tcp_bpf_bypass_getsockopt(int level, int optname)
4647{
4648 /* TCP do_tcp_getsockopt has optimized getsockopt implementation
4649 * to avoid extra socket lock for TCP_ZEROCOPY_RECEIVE.
4650 */
4651 if (level == SOL_TCP && optname == TCP_ZEROCOPY_RECEIVE)
4652 return true;
4653
4654 return false;
4655}
4656EXPORT_SYMBOL(tcp_bpf_bypass_getsockopt);
4657
4658int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval,
4659 int __user *optlen)
4660{
4661 struct inet_connection_sock *icsk = inet_csk(sk);
4662
4663 if (level != SOL_TCP)
4664 /* Paired with WRITE_ONCE() in do_ipv6_setsockopt() and tcp_v6_connect() */
4665 return READ_ONCE(icsk->icsk_af_ops)->getsockopt(sk, level, optname,
4666 optval, optlen);
4667 return do_tcp_getsockopt(sk, level, optname, USER_SOCKPTR(optval),
4668 USER_SOCKPTR(optlen));
4669}
4670EXPORT_SYMBOL(tcp_getsockopt);
4671
4672#ifdef CONFIG_TCP_MD5SIG
4673int tcp_md5_sigpool_id = -1;
4674EXPORT_SYMBOL_GPL(tcp_md5_sigpool_id);
4675
4676int tcp_md5_alloc_sigpool(void)
4677{
4678 size_t scratch_size;
4679 int ret;
4680
4681 scratch_size = sizeof(union tcp_md5sum_block) + sizeof(struct tcphdr);
4682 ret = tcp_sigpool_alloc_ahash("md5", scratch_size);
4683 if (ret >= 0) {
4684 /* As long as any md5 sigpool was allocated, the return
4685 * id would stay the same. Re-write the id only for the case
4686 * when previously all MD5 keys were deleted and this call
4687 * allocates the first MD5 key, which may return a different
4688 * sigpool id than was used previously.
4689 */
4690 WRITE_ONCE(tcp_md5_sigpool_id, ret); /* Avoids the compiler potentially being smart here */
4691 return 0;
4692 }
4693 return ret;
4694}
4695
4696void tcp_md5_release_sigpool(void)
4697{
4698 tcp_sigpool_release(READ_ONCE(tcp_md5_sigpool_id));
4699}
4700
4701void tcp_md5_add_sigpool(void)
4702{
4703 tcp_sigpool_get(READ_ONCE(tcp_md5_sigpool_id));
4704}
4705
4706int tcp_md5_hash_key(struct tcp_sigpool *hp,
4707 const struct tcp_md5sig_key *key)
4708{
4709 u8 keylen = READ_ONCE(key->keylen); /* paired with WRITE_ONCE() in tcp_md5_do_add */
4710 struct scatterlist sg;
4711
4712 sg_init_one(&sg, key->key, keylen);
4713 ahash_request_set_crypt(hp->req, &sg, NULL, keylen);
4714
4715 /* We use data_race() because tcp_md5_do_add() might change
4716 * key->key under us
4717 */
4718 return data_race(crypto_ahash_update(hp->req));
4719}
4720EXPORT_SYMBOL(tcp_md5_hash_key);
4721
4722/* Called with rcu_read_lock() */
4723static enum skb_drop_reason
4724tcp_inbound_md5_hash(const struct sock *sk, const struct sk_buff *skb,
4725 const void *saddr, const void *daddr,
4726 int family, int l3index, const __u8 *hash_location)
4727{
4728 /* This gets called for each TCP segment that has TCP-MD5 option.
4729 * We have 3 drop cases:
4730 * o No MD5 hash and one expected.
4731 * o MD5 hash and we're not expecting one.
4732 * o MD5 hash and its wrong.
4733 */
4734 const struct tcp_sock *tp = tcp_sk(sk);
4735 struct tcp_md5sig_key *key;
4736 u8 newhash[16];
4737 int genhash;
4738
4739 key = tcp_md5_do_lookup(sk, l3index, saddr, family);
4740
4741 if (!key && hash_location) {
4742 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMD5UNEXPECTED);
4743 trace_tcp_hash_md5_unexpected(sk, skb);
4744 return SKB_DROP_REASON_TCP_MD5UNEXPECTED;
4745 }
4746
4747 /* Check the signature.
4748 * To support dual stack listeners, we need to handle
4749 * IPv4-mapped case.
4750 */
4751 if (family == AF_INET)
4752 genhash = tcp_v4_md5_hash_skb(newhash, key, NULL, skb);
4753 else
4754 genhash = tp->af_specific->calc_md5_hash(newhash, key,
4755 NULL, skb);
4756 if (genhash || memcmp(hash_location, newhash, 16) != 0) {
4757 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMD5FAILURE);
4758 trace_tcp_hash_md5_mismatch(sk, skb);
4759 return SKB_DROP_REASON_TCP_MD5FAILURE;
4760 }
4761 return SKB_NOT_DROPPED_YET;
4762}
4763#else
4764static inline enum skb_drop_reason
4765tcp_inbound_md5_hash(const struct sock *sk, const struct sk_buff *skb,
4766 const void *saddr, const void *daddr,
4767 int family, int l3index, const __u8 *hash_location)
4768{
4769 return SKB_NOT_DROPPED_YET;
4770}
4771
4772#endif
4773
4774/* Called with rcu_read_lock() */
4775enum skb_drop_reason
4776tcp_inbound_hash(struct sock *sk, const struct request_sock *req,
4777 const struct sk_buff *skb,
4778 const void *saddr, const void *daddr,
4779 int family, int dif, int sdif)
4780{
4781 const struct tcphdr *th = tcp_hdr(skb);
4782 const struct tcp_ao_hdr *aoh;
4783 const __u8 *md5_location;
4784 int l3index;
4785
4786 /* Invalid option or two times meet any of auth options */
4787 if (tcp_parse_auth_options(th, &md5_location, &aoh)) {
4788 trace_tcp_hash_bad_header(sk, skb);
4789 return SKB_DROP_REASON_TCP_AUTH_HDR;
4790 }
4791
4792 if (req) {
4793 if (tcp_rsk_used_ao(req) != !!aoh) {
4794 u8 keyid, rnext, maclen;
4795
4796 if (aoh) {
4797 keyid = aoh->keyid;
4798 rnext = aoh->rnext_keyid;
4799 maclen = tcp_ao_hdr_maclen(aoh);
4800 } else {
4801 keyid = rnext = maclen = 0;
4802 }
4803
4804 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPAOBAD);
4805 trace_tcp_ao_handshake_failure(sk, skb, keyid, rnext, maclen);
4806 return SKB_DROP_REASON_TCP_AOFAILURE;
4807 }
4808 }
4809
4810 /* sdif set, means packet ingressed via a device
4811 * in an L3 domain and dif is set to the l3mdev
4812 */
4813 l3index = sdif ? dif : 0;
4814
4815 /* Fast path: unsigned segments */
4816 if (likely(!md5_location && !aoh)) {
4817 /* Drop if there's TCP-MD5 or TCP-AO key with any rcvid/sndid
4818 * for the remote peer. On TCP-AO established connection
4819 * the last key is impossible to remove, so there's
4820 * always at least one current_key.
4821 */
4822 if (tcp_ao_required(sk, saddr, family, l3index, true)) {
4823 trace_tcp_hash_ao_required(sk, skb);
4824 return SKB_DROP_REASON_TCP_AONOTFOUND;
4825 }
4826 if (unlikely(tcp_md5_do_lookup(sk, l3index, saddr, family))) {
4827 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMD5NOTFOUND);
4828 trace_tcp_hash_md5_required(sk, skb);
4829 return SKB_DROP_REASON_TCP_MD5NOTFOUND;
4830 }
4831 return SKB_NOT_DROPPED_YET;
4832 }
4833
4834 if (aoh)
4835 return tcp_inbound_ao_hash(sk, skb, family, req, l3index, aoh);
4836
4837 return tcp_inbound_md5_hash(sk, skb, saddr, daddr, family,
4838 l3index, md5_location);
4839}
4840EXPORT_SYMBOL_GPL(tcp_inbound_hash);
4841
4842void tcp_done(struct sock *sk)
4843{
4844 struct request_sock *req;
4845
4846 /* We might be called with a new socket, after
4847 * inet_csk_prepare_forced_close() has been called
4848 * so we can not use lockdep_sock_is_held(sk)
4849 */
4850 req = rcu_dereference_protected(tcp_sk(sk)->fastopen_rsk, 1);
4851
4852 if (sk->sk_state == TCP_SYN_SENT || sk->sk_state == TCP_SYN_RECV)
4853 TCP_INC_STATS(sock_net(sk), TCP_MIB_ATTEMPTFAILS);
4854
4855 tcp_set_state(sk, TCP_CLOSE);
4856 tcp_clear_xmit_timers(sk);
4857 if (req)
4858 reqsk_fastopen_remove(sk, req, false);
4859
4860 WRITE_ONCE(sk->sk_shutdown, SHUTDOWN_MASK);
4861
4862 if (!sock_flag(sk, SOCK_DEAD))
4863 sk->sk_state_change(sk);
4864 else
4865 inet_csk_destroy_sock(sk);
4866}
4867EXPORT_SYMBOL_GPL(tcp_done);
4868
4869int tcp_abort(struct sock *sk, int err)
4870{
4871 int state = inet_sk_state_load(sk);
4872
4873 if (state == TCP_NEW_SYN_RECV) {
4874 struct request_sock *req = inet_reqsk(sk);
4875
4876 local_bh_disable();
4877 inet_csk_reqsk_queue_drop(req->rsk_listener, req);
4878 local_bh_enable();
4879 return 0;
4880 }
4881 if (state == TCP_TIME_WAIT) {
4882 struct inet_timewait_sock *tw = inet_twsk(sk);
4883
4884 refcount_inc(&tw->tw_refcnt);
4885 local_bh_disable();
4886 inet_twsk_deschedule_put(tw);
4887 local_bh_enable();
4888 return 0;
4889 }
4890
4891 /* BPF context ensures sock locking. */
4892 if (!has_current_bpf_ctx())
4893 /* Don't race with userspace socket closes such as tcp_close. */
4894 lock_sock(sk);
4895
4896 /* Avoid closing the same socket twice. */
4897 if (sk->sk_state == TCP_CLOSE) {
4898 if (!has_current_bpf_ctx())
4899 release_sock(sk);
4900 return -ENOENT;
4901 }
4902
4903 if (sk->sk_state == TCP_LISTEN) {
4904 tcp_set_state(sk, TCP_CLOSE);
4905 inet_csk_listen_stop(sk);
4906 }
4907
4908 /* Don't race with BH socket closes such as inet_csk_listen_stop. */
4909 local_bh_disable();
4910 bh_lock_sock(sk);
4911
4912 if (tcp_need_reset(sk->sk_state))
4913 tcp_send_active_reset(sk, GFP_ATOMIC,
4914 SK_RST_REASON_TCP_STATE);
4915 tcp_done_with_error(sk, err);
4916
4917 bh_unlock_sock(sk);
4918 local_bh_enable();
4919 if (!has_current_bpf_ctx())
4920 release_sock(sk);
4921 return 0;
4922}
4923EXPORT_SYMBOL_GPL(tcp_abort);
4924
4925extern struct tcp_congestion_ops tcp_reno;
4926
4927static __initdata unsigned long thash_entries;
4928static int __init set_thash_entries(char *str)
4929{
4930 ssize_t ret;
4931
4932 if (!str)
4933 return 0;
4934
4935 ret = kstrtoul(str, 0, &thash_entries);
4936 if (ret)
4937 return 0;
4938
4939 return 1;
4940}
4941__setup("thash_entries=", set_thash_entries);
4942
4943static void __init tcp_init_mem(void)
4944{
4945 unsigned long limit = nr_free_buffer_pages() / 16;
4946
4947 limit = max(limit, 128UL);
4948 sysctl_tcp_mem[0] = limit / 4 * 3; /* 4.68 % */
4949 sysctl_tcp_mem[1] = limit; /* 6.25 % */
4950 sysctl_tcp_mem[2] = sysctl_tcp_mem[0] * 2; /* 9.37 % */
4951}
4952
4953static void __init tcp_struct_check(void)
4954{
4955 /* TX read-mostly hotpath cache lines */
4956 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_tx, max_window);
4957 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_tx, rcv_ssthresh);
4958 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_tx, reordering);
4959 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_tx, notsent_lowat);
4960 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_tx, gso_segs);
4961 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_tx, lost_skb_hint);
4962 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_tx, retransmit_skb_hint);
4963 CACHELINE_ASSERT_GROUP_SIZE(struct tcp_sock, tcp_sock_read_tx, 40);
4964
4965 /* TXRX read-mostly hotpath cache lines */
4966 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_txrx, tsoffset);
4967 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_txrx, snd_wnd);
4968 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_txrx, mss_cache);
4969 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_txrx, snd_cwnd);
4970 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_txrx, prr_out);
4971 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_txrx, lost_out);
4972 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_txrx, sacked_out);
4973 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_txrx, scaling_ratio);
4974 CACHELINE_ASSERT_GROUP_SIZE(struct tcp_sock, tcp_sock_read_txrx, 32);
4975
4976 /* RX read-mostly hotpath cache lines */
4977 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_rx, copied_seq);
4978 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_rx, rcv_tstamp);
4979 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_rx, snd_wl1);
4980 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_rx, tlp_high_seq);
4981 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_rx, rttvar_us);
4982 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_rx, retrans_out);
4983 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_rx, advmss);
4984 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_rx, urg_data);
4985 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_rx, lost);
4986 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_rx, rtt_min);
4987 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_rx, out_of_order_queue);
4988 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_rx, snd_ssthresh);
4989 CACHELINE_ASSERT_GROUP_SIZE(struct tcp_sock, tcp_sock_read_rx, 69);
4990
4991 /* TX read-write hotpath cache lines */
4992 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, segs_out);
4993 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, data_segs_out);
4994 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, bytes_sent);
4995 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, snd_sml);
4996 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, chrono_start);
4997 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, chrono_stat);
4998 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, write_seq);
4999 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, pushed_seq);
5000 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, lsndtime);
5001 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, mdev_us);
5002 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, tcp_wstamp_ns);
5003 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, rtt_seq);
5004 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, tsorted_sent_queue);
5005 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, highest_sack);
5006 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, ecn_flags);
5007 CACHELINE_ASSERT_GROUP_SIZE(struct tcp_sock, tcp_sock_write_tx, 89);
5008
5009 /* TXRX read-write hotpath cache lines */
5010 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_txrx, pred_flags);
5011 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_txrx, tcp_clock_cache);
5012 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_txrx, tcp_mstamp);
5013 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_txrx, rcv_nxt);
5014 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_txrx, snd_nxt);
5015 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_txrx, snd_una);
5016 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_txrx, window_clamp);
5017 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_txrx, srtt_us);
5018 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_txrx, packets_out);
5019 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_txrx, snd_up);
5020 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_txrx, delivered);
5021 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_txrx, delivered_ce);
5022 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_txrx, app_limited);
5023 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_txrx, rcv_wnd);
5024 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_txrx, rx_opt);
5025
5026 /* 32bit arches with 8byte alignment on u64 fields might need padding
5027 * before tcp_clock_cache.
5028 */
5029 CACHELINE_ASSERT_GROUP_SIZE(struct tcp_sock, tcp_sock_write_txrx, 92 + 4);
5030
5031 /* RX read-write hotpath cache lines */
5032 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_rx, bytes_received);
5033 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_rx, segs_in);
5034 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_rx, data_segs_in);
5035 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_rx, rcv_wup);
5036 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_rx, max_packets_out);
5037 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_rx, cwnd_usage_seq);
5038 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_rx, rate_delivered);
5039 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_rx, rate_interval_us);
5040 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_rx, rcv_rtt_last_tsecr);
5041 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_rx, first_tx_mstamp);
5042 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_rx, delivered_mstamp);
5043 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_rx, bytes_acked);
5044 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_rx, rcv_rtt_est);
5045 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_rx, rcvq_space);
5046 CACHELINE_ASSERT_GROUP_SIZE(struct tcp_sock, tcp_sock_write_rx, 99);
5047}
5048
5049void __init tcp_init(void)
5050{
5051 int max_rshare, max_wshare, cnt;
5052 unsigned long limit;
5053 unsigned int i;
5054
5055 BUILD_BUG_ON(TCP_MIN_SND_MSS <= MAX_TCP_OPTION_SPACE);
5056 BUILD_BUG_ON(sizeof(struct tcp_skb_cb) >
5057 sizeof_field(struct sk_buff, cb));
5058
5059 tcp_struct_check();
5060
5061 percpu_counter_init(&tcp_sockets_allocated, 0, GFP_KERNEL);
5062
5063 timer_setup(&tcp_orphan_timer, tcp_orphan_update, TIMER_DEFERRABLE);
5064 mod_timer(&tcp_orphan_timer, jiffies + TCP_ORPHAN_TIMER_PERIOD);
5065
5066 inet_hashinfo2_init(&tcp_hashinfo, "tcp_listen_portaddr_hash",
5067 thash_entries, 21, /* one slot per 2 MB*/
5068 0, 64 * 1024);
5069 tcp_hashinfo.bind_bucket_cachep =
5070 kmem_cache_create("tcp_bind_bucket",
5071 sizeof(struct inet_bind_bucket), 0,
5072 SLAB_HWCACHE_ALIGN | SLAB_PANIC |
5073 SLAB_ACCOUNT,
5074 NULL);
5075 tcp_hashinfo.bind2_bucket_cachep =
5076 kmem_cache_create("tcp_bind2_bucket",
5077 sizeof(struct inet_bind2_bucket), 0,
5078 SLAB_HWCACHE_ALIGN | SLAB_PANIC |
5079 SLAB_ACCOUNT,
5080 NULL);
5081
5082 /* Size and allocate the main established and bind bucket
5083 * hash tables.
5084 *
5085 * The methodology is similar to that of the buffer cache.
5086 */
5087 tcp_hashinfo.ehash =
5088 alloc_large_system_hash("TCP established",
5089 sizeof(struct inet_ehash_bucket),
5090 thash_entries,
5091 17, /* one slot per 128 KB of memory */
5092 0,
5093 NULL,
5094 &tcp_hashinfo.ehash_mask,
5095 0,
5096 thash_entries ? 0 : 512 * 1024);
5097 for (i = 0; i <= tcp_hashinfo.ehash_mask; i++)
5098 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].chain, i);
5099
5100 if (inet_ehash_locks_alloc(&tcp_hashinfo))
5101 panic("TCP: failed to alloc ehash_locks");
5102 tcp_hashinfo.bhash =
5103 alloc_large_system_hash("TCP bind",
5104 2 * sizeof(struct inet_bind_hashbucket),
5105 tcp_hashinfo.ehash_mask + 1,
5106 17, /* one slot per 128 KB of memory */
5107 0,
5108 &tcp_hashinfo.bhash_size,
5109 NULL,
5110 0,
5111 64 * 1024);
5112 tcp_hashinfo.bhash_size = 1U << tcp_hashinfo.bhash_size;
5113 tcp_hashinfo.bhash2 = tcp_hashinfo.bhash + tcp_hashinfo.bhash_size;
5114 for (i = 0; i < tcp_hashinfo.bhash_size; i++) {
5115 spin_lock_init(&tcp_hashinfo.bhash[i].lock);
5116 INIT_HLIST_HEAD(&tcp_hashinfo.bhash[i].chain);
5117 spin_lock_init(&tcp_hashinfo.bhash2[i].lock);
5118 INIT_HLIST_HEAD(&tcp_hashinfo.bhash2[i].chain);
5119 }
5120
5121 tcp_hashinfo.pernet = false;
5122
5123 cnt = tcp_hashinfo.ehash_mask + 1;
5124 sysctl_tcp_max_orphans = cnt / 2;
5125
5126 tcp_init_mem();
5127 /* Set per-socket limits to no more than 1/128 the pressure threshold */
5128 limit = nr_free_buffer_pages() << (PAGE_SHIFT - 7);
5129 max_wshare = min(4UL*1024*1024, limit);
5130 max_rshare = min(6UL*1024*1024, limit);
5131
5132 init_net.ipv4.sysctl_tcp_wmem[0] = PAGE_SIZE;
5133 init_net.ipv4.sysctl_tcp_wmem[1] = 16*1024;
5134 init_net.ipv4.sysctl_tcp_wmem[2] = max(64*1024, max_wshare);
5135
5136 init_net.ipv4.sysctl_tcp_rmem[0] = PAGE_SIZE;
5137 init_net.ipv4.sysctl_tcp_rmem[1] = 131072;
5138 init_net.ipv4.sysctl_tcp_rmem[2] = max(131072, max_rshare);
5139
5140 pr_info("Hash tables configured (established %u bind %u)\n",
5141 tcp_hashinfo.ehash_mask + 1, tcp_hashinfo.bhash_size);
5142
5143 tcp_v4_init();
5144 tcp_metrics_init();
5145 BUG_ON(tcp_register_congestion_control(&tcp_reno) != 0);
5146 tcp_tasklet_init();
5147 mptcp_init();
5148}
1// SPDX-License-Identifier: GPL-2.0-or-later
2/*
3 * INET An implementation of the TCP/IP protocol suite for the LINUX
4 * operating system. INET is implemented using the BSD Socket
5 * interface as the means of communication with the user level.
6 *
7 * Implementation of the Transmission Control Protocol(TCP).
8 *
9 * Authors: Ross Biro
10 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
11 * Mark Evans, <evansmp@uhura.aston.ac.uk>
12 * Corey Minyard <wf-rch!minyard@relay.EU.net>
13 * Florian La Roche, <flla@stud.uni-sb.de>
14 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
15 * Linus Torvalds, <torvalds@cs.helsinki.fi>
16 * Alan Cox, <gw4pts@gw4pts.ampr.org>
17 * Matthew Dillon, <dillon@apollo.west.oic.com>
18 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
19 * Jorge Cwik, <jorge@laser.satlink.net>
20 *
21 * Fixes:
22 * Alan Cox : Numerous verify_area() calls
23 * Alan Cox : Set the ACK bit on a reset
24 * Alan Cox : Stopped it crashing if it closed while
25 * sk->inuse=1 and was trying to connect
26 * (tcp_err()).
27 * Alan Cox : All icmp error handling was broken
28 * pointers passed where wrong and the
29 * socket was looked up backwards. Nobody
30 * tested any icmp error code obviously.
31 * Alan Cox : tcp_err() now handled properly. It
32 * wakes people on errors. poll
33 * behaves and the icmp error race
34 * has gone by moving it into sock.c
35 * Alan Cox : tcp_send_reset() fixed to work for
36 * everything not just packets for
37 * unknown sockets.
38 * Alan Cox : tcp option processing.
39 * Alan Cox : Reset tweaked (still not 100%) [Had
40 * syn rule wrong]
41 * Herp Rosmanith : More reset fixes
42 * Alan Cox : No longer acks invalid rst frames.
43 * Acking any kind of RST is right out.
44 * Alan Cox : Sets an ignore me flag on an rst
45 * receive otherwise odd bits of prattle
46 * escape still
47 * Alan Cox : Fixed another acking RST frame bug.
48 * Should stop LAN workplace lockups.
49 * Alan Cox : Some tidyups using the new skb list
50 * facilities
51 * Alan Cox : sk->keepopen now seems to work
52 * Alan Cox : Pulls options out correctly on accepts
53 * Alan Cox : Fixed assorted sk->rqueue->next errors
54 * Alan Cox : PSH doesn't end a TCP read. Switched a
55 * bit to skb ops.
56 * Alan Cox : Tidied tcp_data to avoid a potential
57 * nasty.
58 * Alan Cox : Added some better commenting, as the
59 * tcp is hard to follow
60 * Alan Cox : Removed incorrect check for 20 * psh
61 * Michael O'Reilly : ack < copied bug fix.
62 * Johannes Stille : Misc tcp fixes (not all in yet).
63 * Alan Cox : FIN with no memory -> CRASH
64 * Alan Cox : Added socket option proto entries.
65 * Also added awareness of them to accept.
66 * Alan Cox : Added TCP options (SOL_TCP)
67 * Alan Cox : Switched wakeup calls to callbacks,
68 * so the kernel can layer network
69 * sockets.
70 * Alan Cox : Use ip_tos/ip_ttl settings.
71 * Alan Cox : Handle FIN (more) properly (we hope).
72 * Alan Cox : RST frames sent on unsynchronised
73 * state ack error.
74 * Alan Cox : Put in missing check for SYN bit.
75 * Alan Cox : Added tcp_select_window() aka NET2E
76 * window non shrink trick.
77 * Alan Cox : Added a couple of small NET2E timer
78 * fixes
79 * Charles Hedrick : TCP fixes
80 * Toomas Tamm : TCP window fixes
81 * Alan Cox : Small URG fix to rlogin ^C ack fight
82 * Charles Hedrick : Rewrote most of it to actually work
83 * Linus : Rewrote tcp_read() and URG handling
84 * completely
85 * Gerhard Koerting: Fixed some missing timer handling
86 * Matthew Dillon : Reworked TCP machine states as per RFC
87 * Gerhard Koerting: PC/TCP workarounds
88 * Adam Caldwell : Assorted timer/timing errors
89 * Matthew Dillon : Fixed another RST bug
90 * Alan Cox : Move to kernel side addressing changes.
91 * Alan Cox : Beginning work on TCP fastpathing
92 * (not yet usable)
93 * Arnt Gulbrandsen: Turbocharged tcp_check() routine.
94 * Alan Cox : TCP fast path debugging
95 * Alan Cox : Window clamping
96 * Michael Riepe : Bug in tcp_check()
97 * Matt Dillon : More TCP improvements and RST bug fixes
98 * Matt Dillon : Yet more small nasties remove from the
99 * TCP code (Be very nice to this man if
100 * tcp finally works 100%) 8)
101 * Alan Cox : BSD accept semantics.
102 * Alan Cox : Reset on closedown bug.
103 * Peter De Schrijver : ENOTCONN check missing in tcp_sendto().
104 * Michael Pall : Handle poll() after URG properly in
105 * all cases.
106 * Michael Pall : Undo the last fix in tcp_read_urg()
107 * (multi URG PUSH broke rlogin).
108 * Michael Pall : Fix the multi URG PUSH problem in
109 * tcp_readable(), poll() after URG
110 * works now.
111 * Michael Pall : recv(...,MSG_OOB) never blocks in the
112 * BSD api.
113 * Alan Cox : Changed the semantics of sk->socket to
114 * fix a race and a signal problem with
115 * accept() and async I/O.
116 * Alan Cox : Relaxed the rules on tcp_sendto().
117 * Yury Shevchuk : Really fixed accept() blocking problem.
118 * Craig I. Hagan : Allow for BSD compatible TIME_WAIT for
119 * clients/servers which listen in on
120 * fixed ports.
121 * Alan Cox : Cleaned the above up and shrank it to
122 * a sensible code size.
123 * Alan Cox : Self connect lockup fix.
124 * Alan Cox : No connect to multicast.
125 * Ross Biro : Close unaccepted children on master
126 * socket close.
127 * Alan Cox : Reset tracing code.
128 * Alan Cox : Spurious resets on shutdown.
129 * Alan Cox : Giant 15 minute/60 second timer error
130 * Alan Cox : Small whoops in polling before an
131 * accept.
132 * Alan Cox : Kept the state trace facility since
133 * it's handy for debugging.
134 * Alan Cox : More reset handler fixes.
135 * Alan Cox : Started rewriting the code based on
136 * the RFC's for other useful protocol
137 * references see: Comer, KA9Q NOS, and
138 * for a reference on the difference
139 * between specifications and how BSD
140 * works see the 4.4lite source.
141 * A.N.Kuznetsov : Don't time wait on completion of tidy
142 * close.
143 * Linus Torvalds : Fin/Shutdown & copied_seq changes.
144 * Linus Torvalds : Fixed BSD port reuse to work first syn
145 * Alan Cox : Reimplemented timers as per the RFC
146 * and using multiple timers for sanity.
147 * Alan Cox : Small bug fixes, and a lot of new
148 * comments.
149 * Alan Cox : Fixed dual reader crash by locking
150 * the buffers (much like datagram.c)
151 * Alan Cox : Fixed stuck sockets in probe. A probe
152 * now gets fed up of retrying without
153 * (even a no space) answer.
154 * Alan Cox : Extracted closing code better
155 * Alan Cox : Fixed the closing state machine to
156 * resemble the RFC.
157 * Alan Cox : More 'per spec' fixes.
158 * Jorge Cwik : Even faster checksumming.
159 * Alan Cox : tcp_data() doesn't ack illegal PSH
160 * only frames. At least one pc tcp stack
161 * generates them.
162 * Alan Cox : Cache last socket.
163 * Alan Cox : Per route irtt.
164 * Matt Day : poll()->select() match BSD precisely on error
165 * Alan Cox : New buffers
166 * Marc Tamsky : Various sk->prot->retransmits and
167 * sk->retransmits misupdating fixed.
168 * Fixed tcp_write_timeout: stuck close,
169 * and TCP syn retries gets used now.
170 * Mark Yarvis : In tcp_read_wakeup(), don't send an
171 * ack if state is TCP_CLOSED.
172 * Alan Cox : Look up device on a retransmit - routes may
173 * change. Doesn't yet cope with MSS shrink right
174 * but it's a start!
175 * Marc Tamsky : Closing in closing fixes.
176 * Mike Shaver : RFC1122 verifications.
177 * Alan Cox : rcv_saddr errors.
178 * Alan Cox : Block double connect().
179 * Alan Cox : Small hooks for enSKIP.
180 * Alexey Kuznetsov: Path MTU discovery.
181 * Alan Cox : Support soft errors.
182 * Alan Cox : Fix MTU discovery pathological case
183 * when the remote claims no mtu!
184 * Marc Tamsky : TCP_CLOSE fix.
185 * Colin (G3TNE) : Send a reset on syn ack replies in
186 * window but wrong (fixes NT lpd problems)
187 * Pedro Roque : Better TCP window handling, delayed ack.
188 * Joerg Reuter : No modification of locked buffers in
189 * tcp_do_retransmit()
190 * Eric Schenk : Changed receiver side silly window
191 * avoidance algorithm to BSD style
192 * algorithm. This doubles throughput
193 * against machines running Solaris,
194 * and seems to result in general
195 * improvement.
196 * Stefan Magdalinski : adjusted tcp_readable() to fix FIONREAD
197 * Willy Konynenberg : Transparent proxying support.
198 * Mike McLagan : Routing by source
199 * Keith Owens : Do proper merging with partial SKB's in
200 * tcp_do_sendmsg to avoid burstiness.
201 * Eric Schenk : Fix fast close down bug with
202 * shutdown() followed by close().
203 * Andi Kleen : Make poll agree with SIGIO
204 * Salvatore Sanfilippo : Support SO_LINGER with linger == 1 and
205 * lingertime == 0 (RFC 793 ABORT Call)
206 * Hirokazu Takahashi : Use copy_from_user() instead of
207 * csum_and_copy_from_user() if possible.
208 *
209 * Description of States:
210 *
211 * TCP_SYN_SENT sent a connection request, waiting for ack
212 *
213 * TCP_SYN_RECV received a connection request, sent ack,
214 * waiting for final ack in three-way handshake.
215 *
216 * TCP_ESTABLISHED connection established
217 *
218 * TCP_FIN_WAIT1 our side has shutdown, waiting to complete
219 * transmission of remaining buffered data
220 *
221 * TCP_FIN_WAIT2 all buffered data sent, waiting for remote
222 * to shutdown
223 *
224 * TCP_CLOSING both sides have shutdown but we still have
225 * data we have to finish sending
226 *
227 * TCP_TIME_WAIT timeout to catch resent junk before entering
228 * closed, can only be entered from FIN_WAIT2
229 * or CLOSING. Required because the other end
230 * may not have gotten our last ACK causing it
231 * to retransmit the data packet (which we ignore)
232 *
233 * TCP_CLOSE_WAIT remote side has shutdown and is waiting for
234 * us to finish writing our data and to shutdown
235 * (we have to close() to move on to LAST_ACK)
236 *
237 * TCP_LAST_ACK out side has shutdown after remote has
238 * shutdown. There may still be data in our
239 * buffer that we have to finish sending
240 *
241 * TCP_CLOSE socket is finished
242 */
243
244#define pr_fmt(fmt) "TCP: " fmt
245
246#include <crypto/hash.h>
247#include <linux/kernel.h>
248#include <linux/module.h>
249#include <linux/types.h>
250#include <linux/fcntl.h>
251#include <linux/poll.h>
252#include <linux/inet_diag.h>
253#include <linux/init.h>
254#include <linux/fs.h>
255#include <linux/skbuff.h>
256#include <linux/scatterlist.h>
257#include <linux/splice.h>
258#include <linux/net.h>
259#include <linux/socket.h>
260#include <linux/random.h>
261#include <linux/memblock.h>
262#include <linux/highmem.h>
263#include <linux/cache.h>
264#include <linux/err.h>
265#include <linux/time.h>
266#include <linux/slab.h>
267#include <linux/errqueue.h>
268#include <linux/static_key.h>
269#include <linux/btf.h>
270
271#include <net/icmp.h>
272#include <net/inet_common.h>
273#include <net/tcp.h>
274#include <net/mptcp.h>
275#include <net/xfrm.h>
276#include <net/ip.h>
277#include <net/sock.h>
278
279#include <linux/uaccess.h>
280#include <asm/ioctls.h>
281#include <net/busy_poll.h>
282
283/* Track pending CMSGs. */
284enum {
285 TCP_CMSG_INQ = 1,
286 TCP_CMSG_TS = 2
287};
288
289DEFINE_PER_CPU(unsigned int, tcp_orphan_count);
290EXPORT_PER_CPU_SYMBOL_GPL(tcp_orphan_count);
291
292long sysctl_tcp_mem[3] __read_mostly;
293EXPORT_SYMBOL(sysctl_tcp_mem);
294
295atomic_long_t tcp_memory_allocated ____cacheline_aligned_in_smp; /* Current allocated memory. */
296EXPORT_SYMBOL(tcp_memory_allocated);
297DEFINE_PER_CPU(int, tcp_memory_per_cpu_fw_alloc);
298EXPORT_PER_CPU_SYMBOL_GPL(tcp_memory_per_cpu_fw_alloc);
299
300#if IS_ENABLED(CONFIG_SMC)
301DEFINE_STATIC_KEY_FALSE(tcp_have_smc);
302EXPORT_SYMBOL(tcp_have_smc);
303#endif
304
305/*
306 * Current number of TCP sockets.
307 */
308struct percpu_counter tcp_sockets_allocated ____cacheline_aligned_in_smp;
309EXPORT_SYMBOL(tcp_sockets_allocated);
310
311/*
312 * TCP splice context
313 */
314struct tcp_splice_state {
315 struct pipe_inode_info *pipe;
316 size_t len;
317 unsigned int flags;
318};
319
320/*
321 * Pressure flag: try to collapse.
322 * Technical note: it is used by multiple contexts non atomically.
323 * All the __sk_mem_schedule() is of this nature: accounting
324 * is strict, actions are advisory and have some latency.
325 */
326unsigned long tcp_memory_pressure __read_mostly;
327EXPORT_SYMBOL_GPL(tcp_memory_pressure);
328
329void tcp_enter_memory_pressure(struct sock *sk)
330{
331 unsigned long val;
332
333 if (READ_ONCE(tcp_memory_pressure))
334 return;
335 val = jiffies;
336
337 if (!val)
338 val--;
339 if (!cmpxchg(&tcp_memory_pressure, 0, val))
340 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMEMORYPRESSURES);
341}
342EXPORT_SYMBOL_GPL(tcp_enter_memory_pressure);
343
344void tcp_leave_memory_pressure(struct sock *sk)
345{
346 unsigned long val;
347
348 if (!READ_ONCE(tcp_memory_pressure))
349 return;
350 val = xchg(&tcp_memory_pressure, 0);
351 if (val)
352 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPMEMORYPRESSURESCHRONO,
353 jiffies_to_msecs(jiffies - val));
354}
355EXPORT_SYMBOL_GPL(tcp_leave_memory_pressure);
356
357/* Convert seconds to retransmits based on initial and max timeout */
358static u8 secs_to_retrans(int seconds, int timeout, int rto_max)
359{
360 u8 res = 0;
361
362 if (seconds > 0) {
363 int period = timeout;
364
365 res = 1;
366 while (seconds > period && res < 255) {
367 res++;
368 timeout <<= 1;
369 if (timeout > rto_max)
370 timeout = rto_max;
371 period += timeout;
372 }
373 }
374 return res;
375}
376
377/* Convert retransmits to seconds based on initial and max timeout */
378static int retrans_to_secs(u8 retrans, int timeout, int rto_max)
379{
380 int period = 0;
381
382 if (retrans > 0) {
383 period = timeout;
384 while (--retrans) {
385 timeout <<= 1;
386 if (timeout > rto_max)
387 timeout = rto_max;
388 period += timeout;
389 }
390 }
391 return period;
392}
393
394static u64 tcp_compute_delivery_rate(const struct tcp_sock *tp)
395{
396 u32 rate = READ_ONCE(tp->rate_delivered);
397 u32 intv = READ_ONCE(tp->rate_interval_us);
398 u64 rate64 = 0;
399
400 if (rate && intv) {
401 rate64 = (u64)rate * tp->mss_cache * USEC_PER_SEC;
402 do_div(rate64, intv);
403 }
404 return rate64;
405}
406
407/* Address-family independent initialization for a tcp_sock.
408 *
409 * NOTE: A lot of things set to zero explicitly by call to
410 * sk_alloc() so need not be done here.
411 */
412void tcp_init_sock(struct sock *sk)
413{
414 struct inet_connection_sock *icsk = inet_csk(sk);
415 struct tcp_sock *tp = tcp_sk(sk);
416
417 tp->out_of_order_queue = RB_ROOT;
418 sk->tcp_rtx_queue = RB_ROOT;
419 tcp_init_xmit_timers(sk);
420 INIT_LIST_HEAD(&tp->tsq_node);
421 INIT_LIST_HEAD(&tp->tsorted_sent_queue);
422
423 icsk->icsk_rto = TCP_TIMEOUT_INIT;
424 icsk->icsk_rto_min = TCP_RTO_MIN;
425 icsk->icsk_delack_max = TCP_DELACK_MAX;
426 tp->mdev_us = jiffies_to_usecs(TCP_TIMEOUT_INIT);
427 minmax_reset(&tp->rtt_min, tcp_jiffies32, ~0U);
428
429 /* So many TCP implementations out there (incorrectly) count the
430 * initial SYN frame in their delayed-ACK and congestion control
431 * algorithms that we must have the following bandaid to talk
432 * efficiently to them. -DaveM
433 */
434 tcp_snd_cwnd_set(tp, TCP_INIT_CWND);
435
436 /* There's a bubble in the pipe until at least the first ACK. */
437 tp->app_limited = ~0U;
438 tp->rate_app_limited = 1;
439
440 /* See draft-stevens-tcpca-spec-01 for discussion of the
441 * initialization of these values.
442 */
443 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
444 tp->snd_cwnd_clamp = ~0;
445 tp->mss_cache = TCP_MSS_DEFAULT;
446
447 tp->reordering = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_reordering);
448 tcp_assign_congestion_control(sk);
449
450 tp->tsoffset = 0;
451 tp->rack.reo_wnd_steps = 1;
452
453 sk->sk_write_space = sk_stream_write_space;
454 sock_set_flag(sk, SOCK_USE_WRITE_QUEUE);
455
456 icsk->icsk_sync_mss = tcp_sync_mss;
457
458 WRITE_ONCE(sk->sk_sndbuf, READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_wmem[1]));
459 WRITE_ONCE(sk->sk_rcvbuf, READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_rmem[1]));
460 tcp_scaling_ratio_init(sk);
461
462 set_bit(SOCK_SUPPORT_ZC, &sk->sk_socket->flags);
463 sk_sockets_allocated_inc(sk);
464}
465EXPORT_SYMBOL(tcp_init_sock);
466
467static void tcp_tx_timestamp(struct sock *sk, u16 tsflags)
468{
469 struct sk_buff *skb = tcp_write_queue_tail(sk);
470
471 if (tsflags && skb) {
472 struct skb_shared_info *shinfo = skb_shinfo(skb);
473 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
474
475 sock_tx_timestamp(sk, tsflags, &shinfo->tx_flags);
476 if (tsflags & SOF_TIMESTAMPING_TX_ACK)
477 tcb->txstamp_ack = 1;
478 if (tsflags & SOF_TIMESTAMPING_TX_RECORD_MASK)
479 shinfo->tskey = TCP_SKB_CB(skb)->seq + skb->len - 1;
480 }
481}
482
483static bool tcp_stream_is_readable(struct sock *sk, int target)
484{
485 if (tcp_epollin_ready(sk, target))
486 return true;
487 return sk_is_readable(sk);
488}
489
490/*
491 * Wait for a TCP event.
492 *
493 * Note that we don't need to lock the socket, as the upper poll layers
494 * take care of normal races (between the test and the event) and we don't
495 * go look at any of the socket buffers directly.
496 */
497__poll_t tcp_poll(struct file *file, struct socket *sock, poll_table *wait)
498{
499 __poll_t mask;
500 struct sock *sk = sock->sk;
501 const struct tcp_sock *tp = tcp_sk(sk);
502 u8 shutdown;
503 int state;
504
505 sock_poll_wait(file, sock, wait);
506
507 state = inet_sk_state_load(sk);
508 if (state == TCP_LISTEN)
509 return inet_csk_listen_poll(sk);
510
511 /* Socket is not locked. We are protected from async events
512 * by poll logic and correct handling of state changes
513 * made by other threads is impossible in any case.
514 */
515
516 mask = 0;
517
518 /*
519 * EPOLLHUP is certainly not done right. But poll() doesn't
520 * have a notion of HUP in just one direction, and for a
521 * socket the read side is more interesting.
522 *
523 * Some poll() documentation says that EPOLLHUP is incompatible
524 * with the EPOLLOUT/POLLWR flags, so somebody should check this
525 * all. But careful, it tends to be safer to return too many
526 * bits than too few, and you can easily break real applications
527 * if you don't tell them that something has hung up!
528 *
529 * Check-me.
530 *
531 * Check number 1. EPOLLHUP is _UNMASKABLE_ event (see UNIX98 and
532 * our fs/select.c). It means that after we received EOF,
533 * poll always returns immediately, making impossible poll() on write()
534 * in state CLOSE_WAIT. One solution is evident --- to set EPOLLHUP
535 * if and only if shutdown has been made in both directions.
536 * Actually, it is interesting to look how Solaris and DUX
537 * solve this dilemma. I would prefer, if EPOLLHUP were maskable,
538 * then we could set it on SND_SHUTDOWN. BTW examples given
539 * in Stevens' books assume exactly this behaviour, it explains
540 * why EPOLLHUP is incompatible with EPOLLOUT. --ANK
541 *
542 * NOTE. Check for TCP_CLOSE is added. The goal is to prevent
543 * blocking on fresh not-connected or disconnected socket. --ANK
544 */
545 shutdown = READ_ONCE(sk->sk_shutdown);
546 if (shutdown == SHUTDOWN_MASK || state == TCP_CLOSE)
547 mask |= EPOLLHUP;
548 if (shutdown & RCV_SHUTDOWN)
549 mask |= EPOLLIN | EPOLLRDNORM | EPOLLRDHUP;
550
551 /* Connected or passive Fast Open socket? */
552 if (state != TCP_SYN_SENT &&
553 (state != TCP_SYN_RECV || rcu_access_pointer(tp->fastopen_rsk))) {
554 int target = sock_rcvlowat(sk, 0, INT_MAX);
555 u16 urg_data = READ_ONCE(tp->urg_data);
556
557 if (unlikely(urg_data) &&
558 READ_ONCE(tp->urg_seq) == READ_ONCE(tp->copied_seq) &&
559 !sock_flag(sk, SOCK_URGINLINE))
560 target++;
561
562 if (tcp_stream_is_readable(sk, target))
563 mask |= EPOLLIN | EPOLLRDNORM;
564
565 if (!(shutdown & SEND_SHUTDOWN)) {
566 if (__sk_stream_is_writeable(sk, 1)) {
567 mask |= EPOLLOUT | EPOLLWRNORM;
568 } else { /* send SIGIO later */
569 sk_set_bit(SOCKWQ_ASYNC_NOSPACE, sk);
570 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
571
572 /* Race breaker. If space is freed after
573 * wspace test but before the flags are set,
574 * IO signal will be lost. Memory barrier
575 * pairs with the input side.
576 */
577 smp_mb__after_atomic();
578 if (__sk_stream_is_writeable(sk, 1))
579 mask |= EPOLLOUT | EPOLLWRNORM;
580 }
581 } else
582 mask |= EPOLLOUT | EPOLLWRNORM;
583
584 if (urg_data & TCP_URG_VALID)
585 mask |= EPOLLPRI;
586 } else if (state == TCP_SYN_SENT &&
587 inet_test_bit(DEFER_CONNECT, sk)) {
588 /* Active TCP fastopen socket with defer_connect
589 * Return EPOLLOUT so application can call write()
590 * in order for kernel to generate SYN+data
591 */
592 mask |= EPOLLOUT | EPOLLWRNORM;
593 }
594 /* This barrier is coupled with smp_wmb() in tcp_reset() */
595 smp_rmb();
596 if (READ_ONCE(sk->sk_err) ||
597 !skb_queue_empty_lockless(&sk->sk_error_queue))
598 mask |= EPOLLERR;
599
600 return mask;
601}
602EXPORT_SYMBOL(tcp_poll);
603
604int tcp_ioctl(struct sock *sk, int cmd, int *karg)
605{
606 struct tcp_sock *tp = tcp_sk(sk);
607 int answ;
608 bool slow;
609
610 switch (cmd) {
611 case SIOCINQ:
612 if (sk->sk_state == TCP_LISTEN)
613 return -EINVAL;
614
615 slow = lock_sock_fast(sk);
616 answ = tcp_inq(sk);
617 unlock_sock_fast(sk, slow);
618 break;
619 case SIOCATMARK:
620 answ = READ_ONCE(tp->urg_data) &&
621 READ_ONCE(tp->urg_seq) == READ_ONCE(tp->copied_seq);
622 break;
623 case SIOCOUTQ:
624 if (sk->sk_state == TCP_LISTEN)
625 return -EINVAL;
626
627 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
628 answ = 0;
629 else
630 answ = READ_ONCE(tp->write_seq) - tp->snd_una;
631 break;
632 case SIOCOUTQNSD:
633 if (sk->sk_state == TCP_LISTEN)
634 return -EINVAL;
635
636 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
637 answ = 0;
638 else
639 answ = READ_ONCE(tp->write_seq) -
640 READ_ONCE(tp->snd_nxt);
641 break;
642 default:
643 return -ENOIOCTLCMD;
644 }
645
646 *karg = answ;
647 return 0;
648}
649EXPORT_SYMBOL(tcp_ioctl);
650
651void tcp_mark_push(struct tcp_sock *tp, struct sk_buff *skb)
652{
653 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
654 tp->pushed_seq = tp->write_seq;
655}
656
657static inline bool forced_push(const struct tcp_sock *tp)
658{
659 return after(tp->write_seq, tp->pushed_seq + (tp->max_window >> 1));
660}
661
662void tcp_skb_entail(struct sock *sk, struct sk_buff *skb)
663{
664 struct tcp_sock *tp = tcp_sk(sk);
665 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
666
667 tcb->seq = tcb->end_seq = tp->write_seq;
668 tcb->tcp_flags = TCPHDR_ACK;
669 __skb_header_release(skb);
670 tcp_add_write_queue_tail(sk, skb);
671 sk_wmem_queued_add(sk, skb->truesize);
672 sk_mem_charge(sk, skb->truesize);
673 if (tp->nonagle & TCP_NAGLE_PUSH)
674 tp->nonagle &= ~TCP_NAGLE_PUSH;
675
676 tcp_slow_start_after_idle_check(sk);
677}
678
679static inline void tcp_mark_urg(struct tcp_sock *tp, int flags)
680{
681 if (flags & MSG_OOB)
682 tp->snd_up = tp->write_seq;
683}
684
685/* If a not yet filled skb is pushed, do not send it if
686 * we have data packets in Qdisc or NIC queues :
687 * Because TX completion will happen shortly, it gives a chance
688 * to coalesce future sendmsg() payload into this skb, without
689 * need for a timer, and with no latency trade off.
690 * As packets containing data payload have a bigger truesize
691 * than pure acks (dataless) packets, the last checks prevent
692 * autocorking if we only have an ACK in Qdisc/NIC queues,
693 * or if TX completion was delayed after we processed ACK packet.
694 */
695static bool tcp_should_autocork(struct sock *sk, struct sk_buff *skb,
696 int size_goal)
697{
698 return skb->len < size_goal &&
699 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_autocorking) &&
700 !tcp_rtx_queue_empty(sk) &&
701 refcount_read(&sk->sk_wmem_alloc) > skb->truesize &&
702 tcp_skb_can_collapse_to(skb);
703}
704
705void tcp_push(struct sock *sk, int flags, int mss_now,
706 int nonagle, int size_goal)
707{
708 struct tcp_sock *tp = tcp_sk(sk);
709 struct sk_buff *skb;
710
711 skb = tcp_write_queue_tail(sk);
712 if (!skb)
713 return;
714 if (!(flags & MSG_MORE) || forced_push(tp))
715 tcp_mark_push(tp, skb);
716
717 tcp_mark_urg(tp, flags);
718
719 if (tcp_should_autocork(sk, skb, size_goal)) {
720
721 /* avoid atomic op if TSQ_THROTTLED bit is already set */
722 if (!test_bit(TSQ_THROTTLED, &sk->sk_tsq_flags)) {
723 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPAUTOCORKING);
724 set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags);
725 smp_mb__after_atomic();
726 }
727 /* It is possible TX completion already happened
728 * before we set TSQ_THROTTLED.
729 */
730 if (refcount_read(&sk->sk_wmem_alloc) > skb->truesize)
731 return;
732 }
733
734 if (flags & MSG_MORE)
735 nonagle = TCP_NAGLE_CORK;
736
737 __tcp_push_pending_frames(sk, mss_now, nonagle);
738}
739
740static int tcp_splice_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb,
741 unsigned int offset, size_t len)
742{
743 struct tcp_splice_state *tss = rd_desc->arg.data;
744 int ret;
745
746 ret = skb_splice_bits(skb, skb->sk, offset, tss->pipe,
747 min(rd_desc->count, len), tss->flags);
748 if (ret > 0)
749 rd_desc->count -= ret;
750 return ret;
751}
752
753static int __tcp_splice_read(struct sock *sk, struct tcp_splice_state *tss)
754{
755 /* Store TCP splice context information in read_descriptor_t. */
756 read_descriptor_t rd_desc = {
757 .arg.data = tss,
758 .count = tss->len,
759 };
760
761 return tcp_read_sock(sk, &rd_desc, tcp_splice_data_recv);
762}
763
764/**
765 * tcp_splice_read - splice data from TCP socket to a pipe
766 * @sock: socket to splice from
767 * @ppos: position (not valid)
768 * @pipe: pipe to splice to
769 * @len: number of bytes to splice
770 * @flags: splice modifier flags
771 *
772 * Description:
773 * Will read pages from given socket and fill them into a pipe.
774 *
775 **/
776ssize_t tcp_splice_read(struct socket *sock, loff_t *ppos,
777 struct pipe_inode_info *pipe, size_t len,
778 unsigned int flags)
779{
780 struct sock *sk = sock->sk;
781 struct tcp_splice_state tss = {
782 .pipe = pipe,
783 .len = len,
784 .flags = flags,
785 };
786 long timeo;
787 ssize_t spliced;
788 int ret;
789
790 sock_rps_record_flow(sk);
791 /*
792 * We can't seek on a socket input
793 */
794 if (unlikely(*ppos))
795 return -ESPIPE;
796
797 ret = spliced = 0;
798
799 lock_sock(sk);
800
801 timeo = sock_rcvtimeo(sk, sock->file->f_flags & O_NONBLOCK);
802 while (tss.len) {
803 ret = __tcp_splice_read(sk, &tss);
804 if (ret < 0)
805 break;
806 else if (!ret) {
807 if (spliced)
808 break;
809 if (sock_flag(sk, SOCK_DONE))
810 break;
811 if (sk->sk_err) {
812 ret = sock_error(sk);
813 break;
814 }
815 if (sk->sk_shutdown & RCV_SHUTDOWN)
816 break;
817 if (sk->sk_state == TCP_CLOSE) {
818 /*
819 * This occurs when user tries to read
820 * from never connected socket.
821 */
822 ret = -ENOTCONN;
823 break;
824 }
825 if (!timeo) {
826 ret = -EAGAIN;
827 break;
828 }
829 /* if __tcp_splice_read() got nothing while we have
830 * an skb in receive queue, we do not want to loop.
831 * This might happen with URG data.
832 */
833 if (!skb_queue_empty(&sk->sk_receive_queue))
834 break;
835 ret = sk_wait_data(sk, &timeo, NULL);
836 if (ret < 0)
837 break;
838 if (signal_pending(current)) {
839 ret = sock_intr_errno(timeo);
840 break;
841 }
842 continue;
843 }
844 tss.len -= ret;
845 spliced += ret;
846
847 if (!tss.len || !timeo)
848 break;
849 release_sock(sk);
850 lock_sock(sk);
851
852 if (sk->sk_err || sk->sk_state == TCP_CLOSE ||
853 (sk->sk_shutdown & RCV_SHUTDOWN) ||
854 signal_pending(current))
855 break;
856 }
857
858 release_sock(sk);
859
860 if (spliced)
861 return spliced;
862
863 return ret;
864}
865EXPORT_SYMBOL(tcp_splice_read);
866
867struct sk_buff *tcp_stream_alloc_skb(struct sock *sk, gfp_t gfp,
868 bool force_schedule)
869{
870 struct sk_buff *skb;
871
872 skb = alloc_skb_fclone(MAX_TCP_HEADER, gfp);
873 if (likely(skb)) {
874 bool mem_scheduled;
875
876 skb->truesize = SKB_TRUESIZE(skb_end_offset(skb));
877 if (force_schedule) {
878 mem_scheduled = true;
879 sk_forced_mem_schedule(sk, skb->truesize);
880 } else {
881 mem_scheduled = sk_wmem_schedule(sk, skb->truesize);
882 }
883 if (likely(mem_scheduled)) {
884 skb_reserve(skb, MAX_TCP_HEADER);
885 skb->ip_summed = CHECKSUM_PARTIAL;
886 INIT_LIST_HEAD(&skb->tcp_tsorted_anchor);
887 return skb;
888 }
889 __kfree_skb(skb);
890 } else {
891 sk->sk_prot->enter_memory_pressure(sk);
892 sk_stream_moderate_sndbuf(sk);
893 }
894 return NULL;
895}
896
897static unsigned int tcp_xmit_size_goal(struct sock *sk, u32 mss_now,
898 int large_allowed)
899{
900 struct tcp_sock *tp = tcp_sk(sk);
901 u32 new_size_goal, size_goal;
902
903 if (!large_allowed)
904 return mss_now;
905
906 /* Note : tcp_tso_autosize() will eventually split this later */
907 new_size_goal = tcp_bound_to_half_wnd(tp, sk->sk_gso_max_size);
908
909 /* We try hard to avoid divides here */
910 size_goal = tp->gso_segs * mss_now;
911 if (unlikely(new_size_goal < size_goal ||
912 new_size_goal >= size_goal + mss_now)) {
913 tp->gso_segs = min_t(u16, new_size_goal / mss_now,
914 sk->sk_gso_max_segs);
915 size_goal = tp->gso_segs * mss_now;
916 }
917
918 return max(size_goal, mss_now);
919}
920
921int tcp_send_mss(struct sock *sk, int *size_goal, int flags)
922{
923 int mss_now;
924
925 mss_now = tcp_current_mss(sk);
926 *size_goal = tcp_xmit_size_goal(sk, mss_now, !(flags & MSG_OOB));
927
928 return mss_now;
929}
930
931/* In some cases, sendmsg() could have added an skb to the write queue,
932 * but failed adding payload on it. We need to remove it to consume less
933 * memory, but more importantly be able to generate EPOLLOUT for Edge Trigger
934 * epoll() users. Another reason is that tcp_write_xmit() does not like
935 * finding an empty skb in the write queue.
936 */
937void tcp_remove_empty_skb(struct sock *sk)
938{
939 struct sk_buff *skb = tcp_write_queue_tail(sk);
940
941 if (skb && TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq) {
942 tcp_unlink_write_queue(skb, sk);
943 if (tcp_write_queue_empty(sk))
944 tcp_chrono_stop(sk, TCP_CHRONO_BUSY);
945 tcp_wmem_free_skb(sk, skb);
946 }
947}
948
949/* skb changing from pure zc to mixed, must charge zc */
950static int tcp_downgrade_zcopy_pure(struct sock *sk, struct sk_buff *skb)
951{
952 if (unlikely(skb_zcopy_pure(skb))) {
953 u32 extra = skb->truesize -
954 SKB_TRUESIZE(skb_end_offset(skb));
955
956 if (!sk_wmem_schedule(sk, extra))
957 return -ENOMEM;
958
959 sk_mem_charge(sk, extra);
960 skb_shinfo(skb)->flags &= ~SKBFL_PURE_ZEROCOPY;
961 }
962 return 0;
963}
964
965
966int tcp_wmem_schedule(struct sock *sk, int copy)
967{
968 int left;
969
970 if (likely(sk_wmem_schedule(sk, copy)))
971 return copy;
972
973 /* We could be in trouble if we have nothing queued.
974 * Use whatever is left in sk->sk_forward_alloc and tcp_wmem[0]
975 * to guarantee some progress.
976 */
977 left = sock_net(sk)->ipv4.sysctl_tcp_wmem[0] - sk->sk_wmem_queued;
978 if (left > 0)
979 sk_forced_mem_schedule(sk, min(left, copy));
980 return min(copy, sk->sk_forward_alloc);
981}
982
983void tcp_free_fastopen_req(struct tcp_sock *tp)
984{
985 if (tp->fastopen_req) {
986 kfree(tp->fastopen_req);
987 tp->fastopen_req = NULL;
988 }
989}
990
991int tcp_sendmsg_fastopen(struct sock *sk, struct msghdr *msg, int *copied,
992 size_t size, struct ubuf_info *uarg)
993{
994 struct tcp_sock *tp = tcp_sk(sk);
995 struct inet_sock *inet = inet_sk(sk);
996 struct sockaddr *uaddr = msg->msg_name;
997 int err, flags;
998
999 if (!(READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_fastopen) &
1000 TFO_CLIENT_ENABLE) ||
1001 (uaddr && msg->msg_namelen >= sizeof(uaddr->sa_family) &&
1002 uaddr->sa_family == AF_UNSPEC))
1003 return -EOPNOTSUPP;
1004 if (tp->fastopen_req)
1005 return -EALREADY; /* Another Fast Open is in progress */
1006
1007 tp->fastopen_req = kzalloc(sizeof(struct tcp_fastopen_request),
1008 sk->sk_allocation);
1009 if (unlikely(!tp->fastopen_req))
1010 return -ENOBUFS;
1011 tp->fastopen_req->data = msg;
1012 tp->fastopen_req->size = size;
1013 tp->fastopen_req->uarg = uarg;
1014
1015 if (inet_test_bit(DEFER_CONNECT, sk)) {
1016 err = tcp_connect(sk);
1017 /* Same failure procedure as in tcp_v4/6_connect */
1018 if (err) {
1019 tcp_set_state(sk, TCP_CLOSE);
1020 inet->inet_dport = 0;
1021 sk->sk_route_caps = 0;
1022 }
1023 }
1024 flags = (msg->msg_flags & MSG_DONTWAIT) ? O_NONBLOCK : 0;
1025 err = __inet_stream_connect(sk->sk_socket, uaddr,
1026 msg->msg_namelen, flags, 1);
1027 /* fastopen_req could already be freed in __inet_stream_connect
1028 * if the connection times out or gets rst
1029 */
1030 if (tp->fastopen_req) {
1031 *copied = tp->fastopen_req->copied;
1032 tcp_free_fastopen_req(tp);
1033 inet_clear_bit(DEFER_CONNECT, sk);
1034 }
1035 return err;
1036}
1037
1038int tcp_sendmsg_locked(struct sock *sk, struct msghdr *msg, size_t size)
1039{
1040 struct tcp_sock *tp = tcp_sk(sk);
1041 struct ubuf_info *uarg = NULL;
1042 struct sk_buff *skb;
1043 struct sockcm_cookie sockc;
1044 int flags, err, copied = 0;
1045 int mss_now = 0, size_goal, copied_syn = 0;
1046 int process_backlog = 0;
1047 int zc = 0;
1048 long timeo;
1049
1050 flags = msg->msg_flags;
1051
1052 if ((flags & MSG_ZEROCOPY) && size) {
1053 if (msg->msg_ubuf) {
1054 uarg = msg->msg_ubuf;
1055 if (sk->sk_route_caps & NETIF_F_SG)
1056 zc = MSG_ZEROCOPY;
1057 } else if (sock_flag(sk, SOCK_ZEROCOPY)) {
1058 skb = tcp_write_queue_tail(sk);
1059 uarg = msg_zerocopy_realloc(sk, size, skb_zcopy(skb));
1060 if (!uarg) {
1061 err = -ENOBUFS;
1062 goto out_err;
1063 }
1064 if (sk->sk_route_caps & NETIF_F_SG)
1065 zc = MSG_ZEROCOPY;
1066 else
1067 uarg_to_msgzc(uarg)->zerocopy = 0;
1068 }
1069 } else if (unlikely(msg->msg_flags & MSG_SPLICE_PAGES) && size) {
1070 if (sk->sk_route_caps & NETIF_F_SG)
1071 zc = MSG_SPLICE_PAGES;
1072 }
1073
1074 if (unlikely(flags & MSG_FASTOPEN ||
1075 inet_test_bit(DEFER_CONNECT, sk)) &&
1076 !tp->repair) {
1077 err = tcp_sendmsg_fastopen(sk, msg, &copied_syn, size, uarg);
1078 if (err == -EINPROGRESS && copied_syn > 0)
1079 goto out;
1080 else if (err)
1081 goto out_err;
1082 }
1083
1084 timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
1085
1086 tcp_rate_check_app_limited(sk); /* is sending application-limited? */
1087
1088 /* Wait for a connection to finish. One exception is TCP Fast Open
1089 * (passive side) where data is allowed to be sent before a connection
1090 * is fully established.
1091 */
1092 if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) &&
1093 !tcp_passive_fastopen(sk)) {
1094 err = sk_stream_wait_connect(sk, &timeo);
1095 if (err != 0)
1096 goto do_error;
1097 }
1098
1099 if (unlikely(tp->repair)) {
1100 if (tp->repair_queue == TCP_RECV_QUEUE) {
1101 copied = tcp_send_rcvq(sk, msg, size);
1102 goto out_nopush;
1103 }
1104
1105 err = -EINVAL;
1106 if (tp->repair_queue == TCP_NO_QUEUE)
1107 goto out_err;
1108
1109 /* 'common' sending to sendq */
1110 }
1111
1112 sockcm_init(&sockc, sk);
1113 if (msg->msg_controllen) {
1114 err = sock_cmsg_send(sk, msg, &sockc);
1115 if (unlikely(err)) {
1116 err = -EINVAL;
1117 goto out_err;
1118 }
1119 }
1120
1121 /* This should be in poll */
1122 sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
1123
1124 /* Ok commence sending. */
1125 copied = 0;
1126
1127restart:
1128 mss_now = tcp_send_mss(sk, &size_goal, flags);
1129
1130 err = -EPIPE;
1131 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
1132 goto do_error;
1133
1134 while (msg_data_left(msg)) {
1135 ssize_t copy = 0;
1136
1137 skb = tcp_write_queue_tail(sk);
1138 if (skb)
1139 copy = size_goal - skb->len;
1140
1141 if (copy <= 0 || !tcp_skb_can_collapse_to(skb)) {
1142 bool first_skb;
1143
1144new_segment:
1145 if (!sk_stream_memory_free(sk))
1146 goto wait_for_space;
1147
1148 if (unlikely(process_backlog >= 16)) {
1149 process_backlog = 0;
1150 if (sk_flush_backlog(sk))
1151 goto restart;
1152 }
1153 first_skb = tcp_rtx_and_write_queues_empty(sk);
1154 skb = tcp_stream_alloc_skb(sk, sk->sk_allocation,
1155 first_skb);
1156 if (!skb)
1157 goto wait_for_space;
1158
1159 process_backlog++;
1160
1161 tcp_skb_entail(sk, skb);
1162 copy = size_goal;
1163
1164 /* All packets are restored as if they have
1165 * already been sent. skb_mstamp_ns isn't set to
1166 * avoid wrong rtt estimation.
1167 */
1168 if (tp->repair)
1169 TCP_SKB_CB(skb)->sacked |= TCPCB_REPAIRED;
1170 }
1171
1172 /* Try to append data to the end of skb. */
1173 if (copy > msg_data_left(msg))
1174 copy = msg_data_left(msg);
1175
1176 if (zc == 0) {
1177 bool merge = true;
1178 int i = skb_shinfo(skb)->nr_frags;
1179 struct page_frag *pfrag = sk_page_frag(sk);
1180
1181 if (!sk_page_frag_refill(sk, pfrag))
1182 goto wait_for_space;
1183
1184 if (!skb_can_coalesce(skb, i, pfrag->page,
1185 pfrag->offset)) {
1186 if (i >= READ_ONCE(sysctl_max_skb_frags)) {
1187 tcp_mark_push(tp, skb);
1188 goto new_segment;
1189 }
1190 merge = false;
1191 }
1192
1193 copy = min_t(int, copy, pfrag->size - pfrag->offset);
1194
1195 if (unlikely(skb_zcopy_pure(skb) || skb_zcopy_managed(skb))) {
1196 if (tcp_downgrade_zcopy_pure(sk, skb))
1197 goto wait_for_space;
1198 skb_zcopy_downgrade_managed(skb);
1199 }
1200
1201 copy = tcp_wmem_schedule(sk, copy);
1202 if (!copy)
1203 goto wait_for_space;
1204
1205 err = skb_copy_to_page_nocache(sk, &msg->msg_iter, skb,
1206 pfrag->page,
1207 pfrag->offset,
1208 copy);
1209 if (err)
1210 goto do_error;
1211
1212 /* Update the skb. */
1213 if (merge) {
1214 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1215 } else {
1216 skb_fill_page_desc(skb, i, pfrag->page,
1217 pfrag->offset, copy);
1218 page_ref_inc(pfrag->page);
1219 }
1220 pfrag->offset += copy;
1221 } else if (zc == MSG_ZEROCOPY) {
1222 /* First append to a fragless skb builds initial
1223 * pure zerocopy skb
1224 */
1225 if (!skb->len)
1226 skb_shinfo(skb)->flags |= SKBFL_PURE_ZEROCOPY;
1227
1228 if (!skb_zcopy_pure(skb)) {
1229 copy = tcp_wmem_schedule(sk, copy);
1230 if (!copy)
1231 goto wait_for_space;
1232 }
1233
1234 err = skb_zerocopy_iter_stream(sk, skb, msg, copy, uarg);
1235 if (err == -EMSGSIZE || err == -EEXIST) {
1236 tcp_mark_push(tp, skb);
1237 goto new_segment;
1238 }
1239 if (err < 0)
1240 goto do_error;
1241 copy = err;
1242 } else if (zc == MSG_SPLICE_PAGES) {
1243 /* Splice in data if we can; copy if we can't. */
1244 if (tcp_downgrade_zcopy_pure(sk, skb))
1245 goto wait_for_space;
1246 copy = tcp_wmem_schedule(sk, copy);
1247 if (!copy)
1248 goto wait_for_space;
1249
1250 err = skb_splice_from_iter(skb, &msg->msg_iter, copy,
1251 sk->sk_allocation);
1252 if (err < 0) {
1253 if (err == -EMSGSIZE) {
1254 tcp_mark_push(tp, skb);
1255 goto new_segment;
1256 }
1257 goto do_error;
1258 }
1259 copy = err;
1260
1261 if (!(flags & MSG_NO_SHARED_FRAGS))
1262 skb_shinfo(skb)->flags |= SKBFL_SHARED_FRAG;
1263
1264 sk_wmem_queued_add(sk, copy);
1265 sk_mem_charge(sk, copy);
1266 }
1267
1268 if (!copied)
1269 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
1270
1271 WRITE_ONCE(tp->write_seq, tp->write_seq + copy);
1272 TCP_SKB_CB(skb)->end_seq += copy;
1273 tcp_skb_pcount_set(skb, 0);
1274
1275 copied += copy;
1276 if (!msg_data_left(msg)) {
1277 if (unlikely(flags & MSG_EOR))
1278 TCP_SKB_CB(skb)->eor = 1;
1279 goto out;
1280 }
1281
1282 if (skb->len < size_goal || (flags & MSG_OOB) || unlikely(tp->repair))
1283 continue;
1284
1285 if (forced_push(tp)) {
1286 tcp_mark_push(tp, skb);
1287 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
1288 } else if (skb == tcp_send_head(sk))
1289 tcp_push_one(sk, mss_now);
1290 continue;
1291
1292wait_for_space:
1293 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1294 tcp_remove_empty_skb(sk);
1295 if (copied)
1296 tcp_push(sk, flags & ~MSG_MORE, mss_now,
1297 TCP_NAGLE_PUSH, size_goal);
1298
1299 err = sk_stream_wait_memory(sk, &timeo);
1300 if (err != 0)
1301 goto do_error;
1302
1303 mss_now = tcp_send_mss(sk, &size_goal, flags);
1304 }
1305
1306out:
1307 if (copied) {
1308 tcp_tx_timestamp(sk, sockc.tsflags);
1309 tcp_push(sk, flags, mss_now, tp->nonagle, size_goal);
1310 }
1311out_nopush:
1312 /* msg->msg_ubuf is pinned by the caller so we don't take extra refs */
1313 if (uarg && !msg->msg_ubuf)
1314 net_zcopy_put(uarg);
1315 return copied + copied_syn;
1316
1317do_error:
1318 tcp_remove_empty_skb(sk);
1319
1320 if (copied + copied_syn)
1321 goto out;
1322out_err:
1323 /* msg->msg_ubuf is pinned by the caller so we don't take extra refs */
1324 if (uarg && !msg->msg_ubuf)
1325 net_zcopy_put_abort(uarg, true);
1326 err = sk_stream_error(sk, flags, err);
1327 /* make sure we wake any epoll edge trigger waiter */
1328 if (unlikely(tcp_rtx_and_write_queues_empty(sk) && err == -EAGAIN)) {
1329 sk->sk_write_space(sk);
1330 tcp_chrono_stop(sk, TCP_CHRONO_SNDBUF_LIMITED);
1331 }
1332 return err;
1333}
1334EXPORT_SYMBOL_GPL(tcp_sendmsg_locked);
1335
1336int tcp_sendmsg(struct sock *sk, struct msghdr *msg, size_t size)
1337{
1338 int ret;
1339
1340 lock_sock(sk);
1341 ret = tcp_sendmsg_locked(sk, msg, size);
1342 release_sock(sk);
1343
1344 return ret;
1345}
1346EXPORT_SYMBOL(tcp_sendmsg);
1347
1348void tcp_splice_eof(struct socket *sock)
1349{
1350 struct sock *sk = sock->sk;
1351 struct tcp_sock *tp = tcp_sk(sk);
1352 int mss_now, size_goal;
1353
1354 if (!tcp_write_queue_tail(sk))
1355 return;
1356
1357 lock_sock(sk);
1358 mss_now = tcp_send_mss(sk, &size_goal, 0);
1359 tcp_push(sk, 0, mss_now, tp->nonagle, size_goal);
1360 release_sock(sk);
1361}
1362EXPORT_SYMBOL_GPL(tcp_splice_eof);
1363
1364/*
1365 * Handle reading urgent data. BSD has very simple semantics for
1366 * this, no blocking and very strange errors 8)
1367 */
1368
1369static int tcp_recv_urg(struct sock *sk, struct msghdr *msg, int len, int flags)
1370{
1371 struct tcp_sock *tp = tcp_sk(sk);
1372
1373 /* No URG data to read. */
1374 if (sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data ||
1375 tp->urg_data == TCP_URG_READ)
1376 return -EINVAL; /* Yes this is right ! */
1377
1378 if (sk->sk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DONE))
1379 return -ENOTCONN;
1380
1381 if (tp->urg_data & TCP_URG_VALID) {
1382 int err = 0;
1383 char c = tp->urg_data;
1384
1385 if (!(flags & MSG_PEEK))
1386 WRITE_ONCE(tp->urg_data, TCP_URG_READ);
1387
1388 /* Read urgent data. */
1389 msg->msg_flags |= MSG_OOB;
1390
1391 if (len > 0) {
1392 if (!(flags & MSG_TRUNC))
1393 err = memcpy_to_msg(msg, &c, 1);
1394 len = 1;
1395 } else
1396 msg->msg_flags |= MSG_TRUNC;
1397
1398 return err ? -EFAULT : len;
1399 }
1400
1401 if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
1402 return 0;
1403
1404 /* Fixed the recv(..., MSG_OOB) behaviour. BSD docs and
1405 * the available implementations agree in this case:
1406 * this call should never block, independent of the
1407 * blocking state of the socket.
1408 * Mike <pall@rz.uni-karlsruhe.de>
1409 */
1410 return -EAGAIN;
1411}
1412
1413static int tcp_peek_sndq(struct sock *sk, struct msghdr *msg, int len)
1414{
1415 struct sk_buff *skb;
1416 int copied = 0, err = 0;
1417
1418 /* XXX -- need to support SO_PEEK_OFF */
1419
1420 skb_rbtree_walk(skb, &sk->tcp_rtx_queue) {
1421 err = skb_copy_datagram_msg(skb, 0, msg, skb->len);
1422 if (err)
1423 return err;
1424 copied += skb->len;
1425 }
1426
1427 skb_queue_walk(&sk->sk_write_queue, skb) {
1428 err = skb_copy_datagram_msg(skb, 0, msg, skb->len);
1429 if (err)
1430 break;
1431
1432 copied += skb->len;
1433 }
1434
1435 return err ?: copied;
1436}
1437
1438/* Clean up the receive buffer for full frames taken by the user,
1439 * then send an ACK if necessary. COPIED is the number of bytes
1440 * tcp_recvmsg has given to the user so far, it speeds up the
1441 * calculation of whether or not we must ACK for the sake of
1442 * a window update.
1443 */
1444void __tcp_cleanup_rbuf(struct sock *sk, int copied)
1445{
1446 struct tcp_sock *tp = tcp_sk(sk);
1447 bool time_to_ack = false;
1448
1449 if (inet_csk_ack_scheduled(sk)) {
1450 const struct inet_connection_sock *icsk = inet_csk(sk);
1451
1452 if (/* Once-per-two-segments ACK was not sent by tcp_input.c */
1453 tp->rcv_nxt - tp->rcv_wup > icsk->icsk_ack.rcv_mss ||
1454 /*
1455 * If this read emptied read buffer, we send ACK, if
1456 * connection is not bidirectional, user drained
1457 * receive buffer and there was a small segment
1458 * in queue.
1459 */
1460 (copied > 0 &&
1461 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED2) ||
1462 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED) &&
1463 !inet_csk_in_pingpong_mode(sk))) &&
1464 !atomic_read(&sk->sk_rmem_alloc)))
1465 time_to_ack = true;
1466 }
1467
1468 /* We send an ACK if we can now advertise a non-zero window
1469 * which has been raised "significantly".
1470 *
1471 * Even if window raised up to infinity, do not send window open ACK
1472 * in states, where we will not receive more. It is useless.
1473 */
1474 if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) {
1475 __u32 rcv_window_now = tcp_receive_window(tp);
1476
1477 /* Optimize, __tcp_select_window() is not cheap. */
1478 if (2*rcv_window_now <= tp->window_clamp) {
1479 __u32 new_window = __tcp_select_window(sk);
1480
1481 /* Send ACK now, if this read freed lots of space
1482 * in our buffer. Certainly, new_window is new window.
1483 * We can advertise it now, if it is not less than current one.
1484 * "Lots" means "at least twice" here.
1485 */
1486 if (new_window && new_window >= 2 * rcv_window_now)
1487 time_to_ack = true;
1488 }
1489 }
1490 if (time_to_ack)
1491 tcp_send_ack(sk);
1492}
1493
1494void tcp_cleanup_rbuf(struct sock *sk, int copied)
1495{
1496 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
1497 struct tcp_sock *tp = tcp_sk(sk);
1498
1499 WARN(skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq),
1500 "cleanup rbuf bug: copied %X seq %X rcvnxt %X\n",
1501 tp->copied_seq, TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt);
1502 __tcp_cleanup_rbuf(sk, copied);
1503}
1504
1505static void tcp_eat_recv_skb(struct sock *sk, struct sk_buff *skb)
1506{
1507 __skb_unlink(skb, &sk->sk_receive_queue);
1508 if (likely(skb->destructor == sock_rfree)) {
1509 sock_rfree(skb);
1510 skb->destructor = NULL;
1511 skb->sk = NULL;
1512 return skb_attempt_defer_free(skb);
1513 }
1514 __kfree_skb(skb);
1515}
1516
1517struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off)
1518{
1519 struct sk_buff *skb;
1520 u32 offset;
1521
1522 while ((skb = skb_peek(&sk->sk_receive_queue)) != NULL) {
1523 offset = seq - TCP_SKB_CB(skb)->seq;
1524 if (unlikely(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
1525 pr_err_once("%s: found a SYN, please report !\n", __func__);
1526 offset--;
1527 }
1528 if (offset < skb->len || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)) {
1529 *off = offset;
1530 return skb;
1531 }
1532 /* This looks weird, but this can happen if TCP collapsing
1533 * splitted a fat GRO packet, while we released socket lock
1534 * in skb_splice_bits()
1535 */
1536 tcp_eat_recv_skb(sk, skb);
1537 }
1538 return NULL;
1539}
1540EXPORT_SYMBOL(tcp_recv_skb);
1541
1542/*
1543 * This routine provides an alternative to tcp_recvmsg() for routines
1544 * that would like to handle copying from skbuffs directly in 'sendfile'
1545 * fashion.
1546 * Note:
1547 * - It is assumed that the socket was locked by the caller.
1548 * - The routine does not block.
1549 * - At present, there is no support for reading OOB data
1550 * or for 'peeking' the socket using this routine
1551 * (although both would be easy to implement).
1552 */
1553int tcp_read_sock(struct sock *sk, read_descriptor_t *desc,
1554 sk_read_actor_t recv_actor)
1555{
1556 struct sk_buff *skb;
1557 struct tcp_sock *tp = tcp_sk(sk);
1558 u32 seq = tp->copied_seq;
1559 u32 offset;
1560 int copied = 0;
1561
1562 if (sk->sk_state == TCP_LISTEN)
1563 return -ENOTCONN;
1564 while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1565 if (offset < skb->len) {
1566 int used;
1567 size_t len;
1568
1569 len = skb->len - offset;
1570 /* Stop reading if we hit a patch of urgent data */
1571 if (unlikely(tp->urg_data)) {
1572 u32 urg_offset = tp->urg_seq - seq;
1573 if (urg_offset < len)
1574 len = urg_offset;
1575 if (!len)
1576 break;
1577 }
1578 used = recv_actor(desc, skb, offset, len);
1579 if (used <= 0) {
1580 if (!copied)
1581 copied = used;
1582 break;
1583 }
1584 if (WARN_ON_ONCE(used > len))
1585 used = len;
1586 seq += used;
1587 copied += used;
1588 offset += used;
1589
1590 /* If recv_actor drops the lock (e.g. TCP splice
1591 * receive) the skb pointer might be invalid when
1592 * getting here: tcp_collapse might have deleted it
1593 * while aggregating skbs from the socket queue.
1594 */
1595 skb = tcp_recv_skb(sk, seq - 1, &offset);
1596 if (!skb)
1597 break;
1598 /* TCP coalescing might have appended data to the skb.
1599 * Try to splice more frags
1600 */
1601 if (offset + 1 != skb->len)
1602 continue;
1603 }
1604 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) {
1605 tcp_eat_recv_skb(sk, skb);
1606 ++seq;
1607 break;
1608 }
1609 tcp_eat_recv_skb(sk, skb);
1610 if (!desc->count)
1611 break;
1612 WRITE_ONCE(tp->copied_seq, seq);
1613 }
1614 WRITE_ONCE(tp->copied_seq, seq);
1615
1616 tcp_rcv_space_adjust(sk);
1617
1618 /* Clean up data we have read: This will do ACK frames. */
1619 if (copied > 0) {
1620 tcp_recv_skb(sk, seq, &offset);
1621 tcp_cleanup_rbuf(sk, copied);
1622 }
1623 return copied;
1624}
1625EXPORT_SYMBOL(tcp_read_sock);
1626
1627int tcp_read_skb(struct sock *sk, skb_read_actor_t recv_actor)
1628{
1629 struct sk_buff *skb;
1630 int copied = 0;
1631
1632 if (sk->sk_state == TCP_LISTEN)
1633 return -ENOTCONN;
1634
1635 while ((skb = skb_peek(&sk->sk_receive_queue)) != NULL) {
1636 u8 tcp_flags;
1637 int used;
1638
1639 __skb_unlink(skb, &sk->sk_receive_queue);
1640 WARN_ON_ONCE(!skb_set_owner_sk_safe(skb, sk));
1641 tcp_flags = TCP_SKB_CB(skb)->tcp_flags;
1642 used = recv_actor(sk, skb);
1643 if (used < 0) {
1644 if (!copied)
1645 copied = used;
1646 break;
1647 }
1648 copied += used;
1649
1650 if (tcp_flags & TCPHDR_FIN)
1651 break;
1652 }
1653 return copied;
1654}
1655EXPORT_SYMBOL(tcp_read_skb);
1656
1657void tcp_read_done(struct sock *sk, size_t len)
1658{
1659 struct tcp_sock *tp = tcp_sk(sk);
1660 u32 seq = tp->copied_seq;
1661 struct sk_buff *skb;
1662 size_t left;
1663 u32 offset;
1664
1665 if (sk->sk_state == TCP_LISTEN)
1666 return;
1667
1668 left = len;
1669 while (left && (skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1670 int used;
1671
1672 used = min_t(size_t, skb->len - offset, left);
1673 seq += used;
1674 left -= used;
1675
1676 if (skb->len > offset + used)
1677 break;
1678
1679 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) {
1680 tcp_eat_recv_skb(sk, skb);
1681 ++seq;
1682 break;
1683 }
1684 tcp_eat_recv_skb(sk, skb);
1685 }
1686 WRITE_ONCE(tp->copied_seq, seq);
1687
1688 tcp_rcv_space_adjust(sk);
1689
1690 /* Clean up data we have read: This will do ACK frames. */
1691 if (left != len)
1692 tcp_cleanup_rbuf(sk, len - left);
1693}
1694EXPORT_SYMBOL(tcp_read_done);
1695
1696int tcp_peek_len(struct socket *sock)
1697{
1698 return tcp_inq(sock->sk);
1699}
1700EXPORT_SYMBOL(tcp_peek_len);
1701
1702/* Make sure sk_rcvbuf is big enough to satisfy SO_RCVLOWAT hint */
1703int tcp_set_rcvlowat(struct sock *sk, int val)
1704{
1705 int space, cap;
1706
1707 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK)
1708 cap = sk->sk_rcvbuf >> 1;
1709 else
1710 cap = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_rmem[2]) >> 1;
1711 val = min(val, cap);
1712 WRITE_ONCE(sk->sk_rcvlowat, val ? : 1);
1713
1714 /* Check if we need to signal EPOLLIN right now */
1715 tcp_data_ready(sk);
1716
1717 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK)
1718 return 0;
1719
1720 space = tcp_space_from_win(sk, val);
1721 if (space > sk->sk_rcvbuf) {
1722 WRITE_ONCE(sk->sk_rcvbuf, space);
1723 tcp_sk(sk)->window_clamp = val;
1724 }
1725 return 0;
1726}
1727EXPORT_SYMBOL(tcp_set_rcvlowat);
1728
1729void tcp_update_recv_tstamps(struct sk_buff *skb,
1730 struct scm_timestamping_internal *tss)
1731{
1732 if (skb->tstamp)
1733 tss->ts[0] = ktime_to_timespec64(skb->tstamp);
1734 else
1735 tss->ts[0] = (struct timespec64) {0};
1736
1737 if (skb_hwtstamps(skb)->hwtstamp)
1738 tss->ts[2] = ktime_to_timespec64(skb_hwtstamps(skb)->hwtstamp);
1739 else
1740 tss->ts[2] = (struct timespec64) {0};
1741}
1742
1743#ifdef CONFIG_MMU
1744static const struct vm_operations_struct tcp_vm_ops = {
1745};
1746
1747int tcp_mmap(struct file *file, struct socket *sock,
1748 struct vm_area_struct *vma)
1749{
1750 if (vma->vm_flags & (VM_WRITE | VM_EXEC))
1751 return -EPERM;
1752 vm_flags_clear(vma, VM_MAYWRITE | VM_MAYEXEC);
1753
1754 /* Instruct vm_insert_page() to not mmap_read_lock(mm) */
1755 vm_flags_set(vma, VM_MIXEDMAP);
1756
1757 vma->vm_ops = &tcp_vm_ops;
1758 return 0;
1759}
1760EXPORT_SYMBOL(tcp_mmap);
1761
1762static skb_frag_t *skb_advance_to_frag(struct sk_buff *skb, u32 offset_skb,
1763 u32 *offset_frag)
1764{
1765 skb_frag_t *frag;
1766
1767 if (unlikely(offset_skb >= skb->len))
1768 return NULL;
1769
1770 offset_skb -= skb_headlen(skb);
1771 if ((int)offset_skb < 0 || skb_has_frag_list(skb))
1772 return NULL;
1773
1774 frag = skb_shinfo(skb)->frags;
1775 while (offset_skb) {
1776 if (skb_frag_size(frag) > offset_skb) {
1777 *offset_frag = offset_skb;
1778 return frag;
1779 }
1780 offset_skb -= skb_frag_size(frag);
1781 ++frag;
1782 }
1783 *offset_frag = 0;
1784 return frag;
1785}
1786
1787static bool can_map_frag(const skb_frag_t *frag)
1788{
1789 struct page *page;
1790
1791 if (skb_frag_size(frag) != PAGE_SIZE || skb_frag_off(frag))
1792 return false;
1793
1794 page = skb_frag_page(frag);
1795
1796 if (PageCompound(page) || page->mapping)
1797 return false;
1798
1799 return true;
1800}
1801
1802static int find_next_mappable_frag(const skb_frag_t *frag,
1803 int remaining_in_skb)
1804{
1805 int offset = 0;
1806
1807 if (likely(can_map_frag(frag)))
1808 return 0;
1809
1810 while (offset < remaining_in_skb && !can_map_frag(frag)) {
1811 offset += skb_frag_size(frag);
1812 ++frag;
1813 }
1814 return offset;
1815}
1816
1817static void tcp_zerocopy_set_hint_for_skb(struct sock *sk,
1818 struct tcp_zerocopy_receive *zc,
1819 struct sk_buff *skb, u32 offset)
1820{
1821 u32 frag_offset, partial_frag_remainder = 0;
1822 int mappable_offset;
1823 skb_frag_t *frag;
1824
1825 /* worst case: skip to next skb. try to improve on this case below */
1826 zc->recv_skip_hint = skb->len - offset;
1827
1828 /* Find the frag containing this offset (and how far into that frag) */
1829 frag = skb_advance_to_frag(skb, offset, &frag_offset);
1830 if (!frag)
1831 return;
1832
1833 if (frag_offset) {
1834 struct skb_shared_info *info = skb_shinfo(skb);
1835
1836 /* We read part of the last frag, must recvmsg() rest of skb. */
1837 if (frag == &info->frags[info->nr_frags - 1])
1838 return;
1839
1840 /* Else, we must at least read the remainder in this frag. */
1841 partial_frag_remainder = skb_frag_size(frag) - frag_offset;
1842 zc->recv_skip_hint -= partial_frag_remainder;
1843 ++frag;
1844 }
1845
1846 /* partial_frag_remainder: If part way through a frag, must read rest.
1847 * mappable_offset: Bytes till next mappable frag, *not* counting bytes
1848 * in partial_frag_remainder.
1849 */
1850 mappable_offset = find_next_mappable_frag(frag, zc->recv_skip_hint);
1851 zc->recv_skip_hint = mappable_offset + partial_frag_remainder;
1852}
1853
1854static int tcp_recvmsg_locked(struct sock *sk, struct msghdr *msg, size_t len,
1855 int flags, struct scm_timestamping_internal *tss,
1856 int *cmsg_flags);
1857static int receive_fallback_to_copy(struct sock *sk,
1858 struct tcp_zerocopy_receive *zc, int inq,
1859 struct scm_timestamping_internal *tss)
1860{
1861 unsigned long copy_address = (unsigned long)zc->copybuf_address;
1862 struct msghdr msg = {};
1863 int err;
1864
1865 zc->length = 0;
1866 zc->recv_skip_hint = 0;
1867
1868 if (copy_address != zc->copybuf_address)
1869 return -EINVAL;
1870
1871 err = import_ubuf(ITER_DEST, (void __user *)copy_address, inq,
1872 &msg.msg_iter);
1873 if (err)
1874 return err;
1875
1876 err = tcp_recvmsg_locked(sk, &msg, inq, MSG_DONTWAIT,
1877 tss, &zc->msg_flags);
1878 if (err < 0)
1879 return err;
1880
1881 zc->copybuf_len = err;
1882 if (likely(zc->copybuf_len)) {
1883 struct sk_buff *skb;
1884 u32 offset;
1885
1886 skb = tcp_recv_skb(sk, tcp_sk(sk)->copied_seq, &offset);
1887 if (skb)
1888 tcp_zerocopy_set_hint_for_skb(sk, zc, skb, offset);
1889 }
1890 return 0;
1891}
1892
1893static int tcp_copy_straggler_data(struct tcp_zerocopy_receive *zc,
1894 struct sk_buff *skb, u32 copylen,
1895 u32 *offset, u32 *seq)
1896{
1897 unsigned long copy_address = (unsigned long)zc->copybuf_address;
1898 struct msghdr msg = {};
1899 int err;
1900
1901 if (copy_address != zc->copybuf_address)
1902 return -EINVAL;
1903
1904 err = import_ubuf(ITER_DEST, (void __user *)copy_address, copylen,
1905 &msg.msg_iter);
1906 if (err)
1907 return err;
1908 err = skb_copy_datagram_msg(skb, *offset, &msg, copylen);
1909 if (err)
1910 return err;
1911 zc->recv_skip_hint -= copylen;
1912 *offset += copylen;
1913 *seq += copylen;
1914 return (__s32)copylen;
1915}
1916
1917static int tcp_zc_handle_leftover(struct tcp_zerocopy_receive *zc,
1918 struct sock *sk,
1919 struct sk_buff *skb,
1920 u32 *seq,
1921 s32 copybuf_len,
1922 struct scm_timestamping_internal *tss)
1923{
1924 u32 offset, copylen = min_t(u32, copybuf_len, zc->recv_skip_hint);
1925
1926 if (!copylen)
1927 return 0;
1928 /* skb is null if inq < PAGE_SIZE. */
1929 if (skb) {
1930 offset = *seq - TCP_SKB_CB(skb)->seq;
1931 } else {
1932 skb = tcp_recv_skb(sk, *seq, &offset);
1933 if (TCP_SKB_CB(skb)->has_rxtstamp) {
1934 tcp_update_recv_tstamps(skb, tss);
1935 zc->msg_flags |= TCP_CMSG_TS;
1936 }
1937 }
1938
1939 zc->copybuf_len = tcp_copy_straggler_data(zc, skb, copylen, &offset,
1940 seq);
1941 return zc->copybuf_len < 0 ? 0 : copylen;
1942}
1943
1944static int tcp_zerocopy_vm_insert_batch_error(struct vm_area_struct *vma,
1945 struct page **pending_pages,
1946 unsigned long pages_remaining,
1947 unsigned long *address,
1948 u32 *length,
1949 u32 *seq,
1950 struct tcp_zerocopy_receive *zc,
1951 u32 total_bytes_to_map,
1952 int err)
1953{
1954 /* At least one page did not map. Try zapping if we skipped earlier. */
1955 if (err == -EBUSY &&
1956 zc->flags & TCP_RECEIVE_ZEROCOPY_FLAG_TLB_CLEAN_HINT) {
1957 u32 maybe_zap_len;
1958
1959 maybe_zap_len = total_bytes_to_map - /* All bytes to map */
1960 *length + /* Mapped or pending */
1961 (pages_remaining * PAGE_SIZE); /* Failed map. */
1962 zap_page_range_single(vma, *address, maybe_zap_len, NULL);
1963 err = 0;
1964 }
1965
1966 if (!err) {
1967 unsigned long leftover_pages = pages_remaining;
1968 int bytes_mapped;
1969
1970 /* We called zap_page_range_single, try to reinsert. */
1971 err = vm_insert_pages(vma, *address,
1972 pending_pages,
1973 &pages_remaining);
1974 bytes_mapped = PAGE_SIZE * (leftover_pages - pages_remaining);
1975 *seq += bytes_mapped;
1976 *address += bytes_mapped;
1977 }
1978 if (err) {
1979 /* Either we were unable to zap, OR we zapped, retried an
1980 * insert, and still had an issue. Either ways, pages_remaining
1981 * is the number of pages we were unable to map, and we unroll
1982 * some state we speculatively touched before.
1983 */
1984 const int bytes_not_mapped = PAGE_SIZE * pages_remaining;
1985
1986 *length -= bytes_not_mapped;
1987 zc->recv_skip_hint += bytes_not_mapped;
1988 }
1989 return err;
1990}
1991
1992static int tcp_zerocopy_vm_insert_batch(struct vm_area_struct *vma,
1993 struct page **pages,
1994 unsigned int pages_to_map,
1995 unsigned long *address,
1996 u32 *length,
1997 u32 *seq,
1998 struct tcp_zerocopy_receive *zc,
1999 u32 total_bytes_to_map)
2000{
2001 unsigned long pages_remaining = pages_to_map;
2002 unsigned int pages_mapped;
2003 unsigned int bytes_mapped;
2004 int err;
2005
2006 err = vm_insert_pages(vma, *address, pages, &pages_remaining);
2007 pages_mapped = pages_to_map - (unsigned int)pages_remaining;
2008 bytes_mapped = PAGE_SIZE * pages_mapped;
2009 /* Even if vm_insert_pages fails, it may have partially succeeded in
2010 * mapping (some but not all of the pages).
2011 */
2012 *seq += bytes_mapped;
2013 *address += bytes_mapped;
2014
2015 if (likely(!err))
2016 return 0;
2017
2018 /* Error: maybe zap and retry + rollback state for failed inserts. */
2019 return tcp_zerocopy_vm_insert_batch_error(vma, pages + pages_mapped,
2020 pages_remaining, address, length, seq, zc, total_bytes_to_map,
2021 err);
2022}
2023
2024#define TCP_VALID_ZC_MSG_FLAGS (TCP_CMSG_TS)
2025static void tcp_zc_finalize_rx_tstamp(struct sock *sk,
2026 struct tcp_zerocopy_receive *zc,
2027 struct scm_timestamping_internal *tss)
2028{
2029 unsigned long msg_control_addr;
2030 struct msghdr cmsg_dummy;
2031
2032 msg_control_addr = (unsigned long)zc->msg_control;
2033 cmsg_dummy.msg_control_user = (void __user *)msg_control_addr;
2034 cmsg_dummy.msg_controllen =
2035 (__kernel_size_t)zc->msg_controllen;
2036 cmsg_dummy.msg_flags = in_compat_syscall()
2037 ? MSG_CMSG_COMPAT : 0;
2038 cmsg_dummy.msg_control_is_user = true;
2039 zc->msg_flags = 0;
2040 if (zc->msg_control == msg_control_addr &&
2041 zc->msg_controllen == cmsg_dummy.msg_controllen) {
2042 tcp_recv_timestamp(&cmsg_dummy, sk, tss);
2043 zc->msg_control = (__u64)
2044 ((uintptr_t)cmsg_dummy.msg_control_user);
2045 zc->msg_controllen =
2046 (__u64)cmsg_dummy.msg_controllen;
2047 zc->msg_flags = (__u32)cmsg_dummy.msg_flags;
2048 }
2049}
2050
2051static struct vm_area_struct *find_tcp_vma(struct mm_struct *mm,
2052 unsigned long address,
2053 bool *mmap_locked)
2054{
2055 struct vm_area_struct *vma = lock_vma_under_rcu(mm, address);
2056
2057 if (vma) {
2058 if (vma->vm_ops != &tcp_vm_ops) {
2059 vma_end_read(vma);
2060 return NULL;
2061 }
2062 *mmap_locked = false;
2063 return vma;
2064 }
2065
2066 mmap_read_lock(mm);
2067 vma = vma_lookup(mm, address);
2068 if (!vma || vma->vm_ops != &tcp_vm_ops) {
2069 mmap_read_unlock(mm);
2070 return NULL;
2071 }
2072 *mmap_locked = true;
2073 return vma;
2074}
2075
2076#define TCP_ZEROCOPY_PAGE_BATCH_SIZE 32
2077static int tcp_zerocopy_receive(struct sock *sk,
2078 struct tcp_zerocopy_receive *zc,
2079 struct scm_timestamping_internal *tss)
2080{
2081 u32 length = 0, offset, vma_len, avail_len, copylen = 0;
2082 unsigned long address = (unsigned long)zc->address;
2083 struct page *pages[TCP_ZEROCOPY_PAGE_BATCH_SIZE];
2084 s32 copybuf_len = zc->copybuf_len;
2085 struct tcp_sock *tp = tcp_sk(sk);
2086 const skb_frag_t *frags = NULL;
2087 unsigned int pages_to_map = 0;
2088 struct vm_area_struct *vma;
2089 struct sk_buff *skb = NULL;
2090 u32 seq = tp->copied_seq;
2091 u32 total_bytes_to_map;
2092 int inq = tcp_inq(sk);
2093 bool mmap_locked;
2094 int ret;
2095
2096 zc->copybuf_len = 0;
2097 zc->msg_flags = 0;
2098
2099 if (address & (PAGE_SIZE - 1) || address != zc->address)
2100 return -EINVAL;
2101
2102 if (sk->sk_state == TCP_LISTEN)
2103 return -ENOTCONN;
2104
2105 sock_rps_record_flow(sk);
2106
2107 if (inq && inq <= copybuf_len)
2108 return receive_fallback_to_copy(sk, zc, inq, tss);
2109
2110 if (inq < PAGE_SIZE) {
2111 zc->length = 0;
2112 zc->recv_skip_hint = inq;
2113 if (!inq && sock_flag(sk, SOCK_DONE))
2114 return -EIO;
2115 return 0;
2116 }
2117
2118 vma = find_tcp_vma(current->mm, address, &mmap_locked);
2119 if (!vma)
2120 return -EINVAL;
2121
2122 vma_len = min_t(unsigned long, zc->length, vma->vm_end - address);
2123 avail_len = min_t(u32, vma_len, inq);
2124 total_bytes_to_map = avail_len & ~(PAGE_SIZE - 1);
2125 if (total_bytes_to_map) {
2126 if (!(zc->flags & TCP_RECEIVE_ZEROCOPY_FLAG_TLB_CLEAN_HINT))
2127 zap_page_range_single(vma, address, total_bytes_to_map,
2128 NULL);
2129 zc->length = total_bytes_to_map;
2130 zc->recv_skip_hint = 0;
2131 } else {
2132 zc->length = avail_len;
2133 zc->recv_skip_hint = avail_len;
2134 }
2135 ret = 0;
2136 while (length + PAGE_SIZE <= zc->length) {
2137 int mappable_offset;
2138 struct page *page;
2139
2140 if (zc->recv_skip_hint < PAGE_SIZE) {
2141 u32 offset_frag;
2142
2143 if (skb) {
2144 if (zc->recv_skip_hint > 0)
2145 break;
2146 skb = skb->next;
2147 offset = seq - TCP_SKB_CB(skb)->seq;
2148 } else {
2149 skb = tcp_recv_skb(sk, seq, &offset);
2150 }
2151
2152 if (TCP_SKB_CB(skb)->has_rxtstamp) {
2153 tcp_update_recv_tstamps(skb, tss);
2154 zc->msg_flags |= TCP_CMSG_TS;
2155 }
2156 zc->recv_skip_hint = skb->len - offset;
2157 frags = skb_advance_to_frag(skb, offset, &offset_frag);
2158 if (!frags || offset_frag)
2159 break;
2160 }
2161
2162 mappable_offset = find_next_mappable_frag(frags,
2163 zc->recv_skip_hint);
2164 if (mappable_offset) {
2165 zc->recv_skip_hint = mappable_offset;
2166 break;
2167 }
2168 page = skb_frag_page(frags);
2169 prefetchw(page);
2170 pages[pages_to_map++] = page;
2171 length += PAGE_SIZE;
2172 zc->recv_skip_hint -= PAGE_SIZE;
2173 frags++;
2174 if (pages_to_map == TCP_ZEROCOPY_PAGE_BATCH_SIZE ||
2175 zc->recv_skip_hint < PAGE_SIZE) {
2176 /* Either full batch, or we're about to go to next skb
2177 * (and we cannot unroll failed ops across skbs).
2178 */
2179 ret = tcp_zerocopy_vm_insert_batch(vma, pages,
2180 pages_to_map,
2181 &address, &length,
2182 &seq, zc,
2183 total_bytes_to_map);
2184 if (ret)
2185 goto out;
2186 pages_to_map = 0;
2187 }
2188 }
2189 if (pages_to_map) {
2190 ret = tcp_zerocopy_vm_insert_batch(vma, pages, pages_to_map,
2191 &address, &length, &seq,
2192 zc, total_bytes_to_map);
2193 }
2194out:
2195 if (mmap_locked)
2196 mmap_read_unlock(current->mm);
2197 else
2198 vma_end_read(vma);
2199 /* Try to copy straggler data. */
2200 if (!ret)
2201 copylen = tcp_zc_handle_leftover(zc, sk, skb, &seq, copybuf_len, tss);
2202
2203 if (length + copylen) {
2204 WRITE_ONCE(tp->copied_seq, seq);
2205 tcp_rcv_space_adjust(sk);
2206
2207 /* Clean up data we have read: This will do ACK frames. */
2208 tcp_recv_skb(sk, seq, &offset);
2209 tcp_cleanup_rbuf(sk, length + copylen);
2210 ret = 0;
2211 if (length == zc->length)
2212 zc->recv_skip_hint = 0;
2213 } else {
2214 if (!zc->recv_skip_hint && sock_flag(sk, SOCK_DONE))
2215 ret = -EIO;
2216 }
2217 zc->length = length;
2218 return ret;
2219}
2220#endif
2221
2222/* Similar to __sock_recv_timestamp, but does not require an skb */
2223void tcp_recv_timestamp(struct msghdr *msg, const struct sock *sk,
2224 struct scm_timestamping_internal *tss)
2225{
2226 int new_tstamp = sock_flag(sk, SOCK_TSTAMP_NEW);
2227 bool has_timestamping = false;
2228
2229 if (tss->ts[0].tv_sec || tss->ts[0].tv_nsec) {
2230 if (sock_flag(sk, SOCK_RCVTSTAMP)) {
2231 if (sock_flag(sk, SOCK_RCVTSTAMPNS)) {
2232 if (new_tstamp) {
2233 struct __kernel_timespec kts = {
2234 .tv_sec = tss->ts[0].tv_sec,
2235 .tv_nsec = tss->ts[0].tv_nsec,
2236 };
2237 put_cmsg(msg, SOL_SOCKET, SO_TIMESTAMPNS_NEW,
2238 sizeof(kts), &kts);
2239 } else {
2240 struct __kernel_old_timespec ts_old = {
2241 .tv_sec = tss->ts[0].tv_sec,
2242 .tv_nsec = tss->ts[0].tv_nsec,
2243 };
2244 put_cmsg(msg, SOL_SOCKET, SO_TIMESTAMPNS_OLD,
2245 sizeof(ts_old), &ts_old);
2246 }
2247 } else {
2248 if (new_tstamp) {
2249 struct __kernel_sock_timeval stv = {
2250 .tv_sec = tss->ts[0].tv_sec,
2251 .tv_usec = tss->ts[0].tv_nsec / 1000,
2252 };
2253 put_cmsg(msg, SOL_SOCKET, SO_TIMESTAMP_NEW,
2254 sizeof(stv), &stv);
2255 } else {
2256 struct __kernel_old_timeval tv = {
2257 .tv_sec = tss->ts[0].tv_sec,
2258 .tv_usec = tss->ts[0].tv_nsec / 1000,
2259 };
2260 put_cmsg(msg, SOL_SOCKET, SO_TIMESTAMP_OLD,
2261 sizeof(tv), &tv);
2262 }
2263 }
2264 }
2265
2266 if (READ_ONCE(sk->sk_tsflags) & SOF_TIMESTAMPING_SOFTWARE)
2267 has_timestamping = true;
2268 else
2269 tss->ts[0] = (struct timespec64) {0};
2270 }
2271
2272 if (tss->ts[2].tv_sec || tss->ts[2].tv_nsec) {
2273 if (READ_ONCE(sk->sk_tsflags) & SOF_TIMESTAMPING_RAW_HARDWARE)
2274 has_timestamping = true;
2275 else
2276 tss->ts[2] = (struct timespec64) {0};
2277 }
2278
2279 if (has_timestamping) {
2280 tss->ts[1] = (struct timespec64) {0};
2281 if (sock_flag(sk, SOCK_TSTAMP_NEW))
2282 put_cmsg_scm_timestamping64(msg, tss);
2283 else
2284 put_cmsg_scm_timestamping(msg, tss);
2285 }
2286}
2287
2288static int tcp_inq_hint(struct sock *sk)
2289{
2290 const struct tcp_sock *tp = tcp_sk(sk);
2291 u32 copied_seq = READ_ONCE(tp->copied_seq);
2292 u32 rcv_nxt = READ_ONCE(tp->rcv_nxt);
2293 int inq;
2294
2295 inq = rcv_nxt - copied_seq;
2296 if (unlikely(inq < 0 || copied_seq != READ_ONCE(tp->copied_seq))) {
2297 lock_sock(sk);
2298 inq = tp->rcv_nxt - tp->copied_seq;
2299 release_sock(sk);
2300 }
2301 /* After receiving a FIN, tell the user-space to continue reading
2302 * by returning a non-zero inq.
2303 */
2304 if (inq == 0 && sock_flag(sk, SOCK_DONE))
2305 inq = 1;
2306 return inq;
2307}
2308
2309/*
2310 * This routine copies from a sock struct into the user buffer.
2311 *
2312 * Technical note: in 2.3 we work on _locked_ socket, so that
2313 * tricks with *seq access order and skb->users are not required.
2314 * Probably, code can be easily improved even more.
2315 */
2316
2317static int tcp_recvmsg_locked(struct sock *sk, struct msghdr *msg, size_t len,
2318 int flags, struct scm_timestamping_internal *tss,
2319 int *cmsg_flags)
2320{
2321 struct tcp_sock *tp = tcp_sk(sk);
2322 int copied = 0;
2323 u32 peek_seq;
2324 u32 *seq;
2325 unsigned long used;
2326 int err;
2327 int target; /* Read at least this many bytes */
2328 long timeo;
2329 struct sk_buff *skb, *last;
2330 u32 urg_hole = 0;
2331
2332 err = -ENOTCONN;
2333 if (sk->sk_state == TCP_LISTEN)
2334 goto out;
2335
2336 if (tp->recvmsg_inq) {
2337 *cmsg_flags = TCP_CMSG_INQ;
2338 msg->msg_get_inq = 1;
2339 }
2340 timeo = sock_rcvtimeo(sk, flags & MSG_DONTWAIT);
2341
2342 /* Urgent data needs to be handled specially. */
2343 if (flags & MSG_OOB)
2344 goto recv_urg;
2345
2346 if (unlikely(tp->repair)) {
2347 err = -EPERM;
2348 if (!(flags & MSG_PEEK))
2349 goto out;
2350
2351 if (tp->repair_queue == TCP_SEND_QUEUE)
2352 goto recv_sndq;
2353
2354 err = -EINVAL;
2355 if (tp->repair_queue == TCP_NO_QUEUE)
2356 goto out;
2357
2358 /* 'common' recv queue MSG_PEEK-ing */
2359 }
2360
2361 seq = &tp->copied_seq;
2362 if (flags & MSG_PEEK) {
2363 peek_seq = tp->copied_seq;
2364 seq = &peek_seq;
2365 }
2366
2367 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
2368
2369 do {
2370 u32 offset;
2371
2372 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
2373 if (unlikely(tp->urg_data) && tp->urg_seq == *seq) {
2374 if (copied)
2375 break;
2376 if (signal_pending(current)) {
2377 copied = timeo ? sock_intr_errno(timeo) : -EAGAIN;
2378 break;
2379 }
2380 }
2381
2382 /* Next get a buffer. */
2383
2384 last = skb_peek_tail(&sk->sk_receive_queue);
2385 skb_queue_walk(&sk->sk_receive_queue, skb) {
2386 last = skb;
2387 /* Now that we have two receive queues this
2388 * shouldn't happen.
2389 */
2390 if (WARN(before(*seq, TCP_SKB_CB(skb)->seq),
2391 "TCP recvmsg seq # bug: copied %X, seq %X, rcvnxt %X, fl %X\n",
2392 *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt,
2393 flags))
2394 break;
2395
2396 offset = *seq - TCP_SKB_CB(skb)->seq;
2397 if (unlikely(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
2398 pr_err_once("%s: found a SYN, please report !\n", __func__);
2399 offset--;
2400 }
2401 if (offset < skb->len)
2402 goto found_ok_skb;
2403 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
2404 goto found_fin_ok;
2405 WARN(!(flags & MSG_PEEK),
2406 "TCP recvmsg seq # bug 2: copied %X, seq %X, rcvnxt %X, fl %X\n",
2407 *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt, flags);
2408 }
2409
2410 /* Well, if we have backlog, try to process it now yet. */
2411
2412 if (copied >= target && !READ_ONCE(sk->sk_backlog.tail))
2413 break;
2414
2415 if (copied) {
2416 if (!timeo ||
2417 sk->sk_err ||
2418 sk->sk_state == TCP_CLOSE ||
2419 (sk->sk_shutdown & RCV_SHUTDOWN) ||
2420 signal_pending(current))
2421 break;
2422 } else {
2423 if (sock_flag(sk, SOCK_DONE))
2424 break;
2425
2426 if (sk->sk_err) {
2427 copied = sock_error(sk);
2428 break;
2429 }
2430
2431 if (sk->sk_shutdown & RCV_SHUTDOWN)
2432 break;
2433
2434 if (sk->sk_state == TCP_CLOSE) {
2435 /* This occurs when user tries to read
2436 * from never connected socket.
2437 */
2438 copied = -ENOTCONN;
2439 break;
2440 }
2441
2442 if (!timeo) {
2443 copied = -EAGAIN;
2444 break;
2445 }
2446
2447 if (signal_pending(current)) {
2448 copied = sock_intr_errno(timeo);
2449 break;
2450 }
2451 }
2452
2453 if (copied >= target) {
2454 /* Do not sleep, just process backlog. */
2455 __sk_flush_backlog(sk);
2456 } else {
2457 tcp_cleanup_rbuf(sk, copied);
2458 err = sk_wait_data(sk, &timeo, last);
2459 if (err < 0) {
2460 err = copied ? : err;
2461 goto out;
2462 }
2463 }
2464
2465 if ((flags & MSG_PEEK) &&
2466 (peek_seq - copied - urg_hole != tp->copied_seq)) {
2467 net_dbg_ratelimited("TCP(%s:%d): Application bug, race in MSG_PEEK\n",
2468 current->comm,
2469 task_pid_nr(current));
2470 peek_seq = tp->copied_seq;
2471 }
2472 continue;
2473
2474found_ok_skb:
2475 /* Ok so how much can we use? */
2476 used = skb->len - offset;
2477 if (len < used)
2478 used = len;
2479
2480 /* Do we have urgent data here? */
2481 if (unlikely(tp->urg_data)) {
2482 u32 urg_offset = tp->urg_seq - *seq;
2483 if (urg_offset < used) {
2484 if (!urg_offset) {
2485 if (!sock_flag(sk, SOCK_URGINLINE)) {
2486 WRITE_ONCE(*seq, *seq + 1);
2487 urg_hole++;
2488 offset++;
2489 used--;
2490 if (!used)
2491 goto skip_copy;
2492 }
2493 } else
2494 used = urg_offset;
2495 }
2496 }
2497
2498 if (!(flags & MSG_TRUNC)) {
2499 err = skb_copy_datagram_msg(skb, offset, msg, used);
2500 if (err) {
2501 /* Exception. Bailout! */
2502 if (!copied)
2503 copied = -EFAULT;
2504 break;
2505 }
2506 }
2507
2508 WRITE_ONCE(*seq, *seq + used);
2509 copied += used;
2510 len -= used;
2511
2512 tcp_rcv_space_adjust(sk);
2513
2514skip_copy:
2515 if (unlikely(tp->urg_data) && after(tp->copied_seq, tp->urg_seq)) {
2516 WRITE_ONCE(tp->urg_data, 0);
2517 tcp_fast_path_check(sk);
2518 }
2519
2520 if (TCP_SKB_CB(skb)->has_rxtstamp) {
2521 tcp_update_recv_tstamps(skb, tss);
2522 *cmsg_flags |= TCP_CMSG_TS;
2523 }
2524
2525 if (used + offset < skb->len)
2526 continue;
2527
2528 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
2529 goto found_fin_ok;
2530 if (!(flags & MSG_PEEK))
2531 tcp_eat_recv_skb(sk, skb);
2532 continue;
2533
2534found_fin_ok:
2535 /* Process the FIN. */
2536 WRITE_ONCE(*seq, *seq + 1);
2537 if (!(flags & MSG_PEEK))
2538 tcp_eat_recv_skb(sk, skb);
2539 break;
2540 } while (len > 0);
2541
2542 /* According to UNIX98, msg_name/msg_namelen are ignored
2543 * on connected socket. I was just happy when found this 8) --ANK
2544 */
2545
2546 /* Clean up data we have read: This will do ACK frames. */
2547 tcp_cleanup_rbuf(sk, copied);
2548 return copied;
2549
2550out:
2551 return err;
2552
2553recv_urg:
2554 err = tcp_recv_urg(sk, msg, len, flags);
2555 goto out;
2556
2557recv_sndq:
2558 err = tcp_peek_sndq(sk, msg, len);
2559 goto out;
2560}
2561
2562int tcp_recvmsg(struct sock *sk, struct msghdr *msg, size_t len, int flags,
2563 int *addr_len)
2564{
2565 int cmsg_flags = 0, ret;
2566 struct scm_timestamping_internal tss;
2567
2568 if (unlikely(flags & MSG_ERRQUEUE))
2569 return inet_recv_error(sk, msg, len, addr_len);
2570
2571 if (sk_can_busy_loop(sk) &&
2572 skb_queue_empty_lockless(&sk->sk_receive_queue) &&
2573 sk->sk_state == TCP_ESTABLISHED)
2574 sk_busy_loop(sk, flags & MSG_DONTWAIT);
2575
2576 lock_sock(sk);
2577 ret = tcp_recvmsg_locked(sk, msg, len, flags, &tss, &cmsg_flags);
2578 release_sock(sk);
2579
2580 if ((cmsg_flags || msg->msg_get_inq) && ret >= 0) {
2581 if (cmsg_flags & TCP_CMSG_TS)
2582 tcp_recv_timestamp(msg, sk, &tss);
2583 if (msg->msg_get_inq) {
2584 msg->msg_inq = tcp_inq_hint(sk);
2585 if (cmsg_flags & TCP_CMSG_INQ)
2586 put_cmsg(msg, SOL_TCP, TCP_CM_INQ,
2587 sizeof(msg->msg_inq), &msg->msg_inq);
2588 }
2589 }
2590 return ret;
2591}
2592EXPORT_SYMBOL(tcp_recvmsg);
2593
2594void tcp_set_state(struct sock *sk, int state)
2595{
2596 int oldstate = sk->sk_state;
2597
2598 /* We defined a new enum for TCP states that are exported in BPF
2599 * so as not force the internal TCP states to be frozen. The
2600 * following checks will detect if an internal state value ever
2601 * differs from the BPF value. If this ever happens, then we will
2602 * need to remap the internal value to the BPF value before calling
2603 * tcp_call_bpf_2arg.
2604 */
2605 BUILD_BUG_ON((int)BPF_TCP_ESTABLISHED != (int)TCP_ESTABLISHED);
2606 BUILD_BUG_ON((int)BPF_TCP_SYN_SENT != (int)TCP_SYN_SENT);
2607 BUILD_BUG_ON((int)BPF_TCP_SYN_RECV != (int)TCP_SYN_RECV);
2608 BUILD_BUG_ON((int)BPF_TCP_FIN_WAIT1 != (int)TCP_FIN_WAIT1);
2609 BUILD_BUG_ON((int)BPF_TCP_FIN_WAIT2 != (int)TCP_FIN_WAIT2);
2610 BUILD_BUG_ON((int)BPF_TCP_TIME_WAIT != (int)TCP_TIME_WAIT);
2611 BUILD_BUG_ON((int)BPF_TCP_CLOSE != (int)TCP_CLOSE);
2612 BUILD_BUG_ON((int)BPF_TCP_CLOSE_WAIT != (int)TCP_CLOSE_WAIT);
2613 BUILD_BUG_ON((int)BPF_TCP_LAST_ACK != (int)TCP_LAST_ACK);
2614 BUILD_BUG_ON((int)BPF_TCP_LISTEN != (int)TCP_LISTEN);
2615 BUILD_BUG_ON((int)BPF_TCP_CLOSING != (int)TCP_CLOSING);
2616 BUILD_BUG_ON((int)BPF_TCP_NEW_SYN_RECV != (int)TCP_NEW_SYN_RECV);
2617 BUILD_BUG_ON((int)BPF_TCP_BOUND_INACTIVE != (int)TCP_BOUND_INACTIVE);
2618 BUILD_BUG_ON((int)BPF_TCP_MAX_STATES != (int)TCP_MAX_STATES);
2619
2620 /* bpf uapi header bpf.h defines an anonymous enum with values
2621 * BPF_TCP_* used by bpf programs. Currently gcc built vmlinux
2622 * is able to emit this enum in DWARF due to the above BUILD_BUG_ON.
2623 * But clang built vmlinux does not have this enum in DWARF
2624 * since clang removes the above code before generating IR/debuginfo.
2625 * Let us explicitly emit the type debuginfo to ensure the
2626 * above-mentioned anonymous enum in the vmlinux DWARF and hence BTF
2627 * regardless of which compiler is used.
2628 */
2629 BTF_TYPE_EMIT_ENUM(BPF_TCP_ESTABLISHED);
2630
2631 if (BPF_SOCK_OPS_TEST_FLAG(tcp_sk(sk), BPF_SOCK_OPS_STATE_CB_FLAG))
2632 tcp_call_bpf_2arg(sk, BPF_SOCK_OPS_STATE_CB, oldstate, state);
2633
2634 switch (state) {
2635 case TCP_ESTABLISHED:
2636 if (oldstate != TCP_ESTABLISHED)
2637 TCP_INC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
2638 break;
2639
2640 case TCP_CLOSE:
2641 if (oldstate == TCP_CLOSE_WAIT || oldstate == TCP_ESTABLISHED)
2642 TCP_INC_STATS(sock_net(sk), TCP_MIB_ESTABRESETS);
2643
2644 sk->sk_prot->unhash(sk);
2645 if (inet_csk(sk)->icsk_bind_hash &&
2646 !(sk->sk_userlocks & SOCK_BINDPORT_LOCK))
2647 inet_put_port(sk);
2648 fallthrough;
2649 default:
2650 if (oldstate == TCP_ESTABLISHED)
2651 TCP_DEC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
2652 }
2653
2654 /* Change state AFTER socket is unhashed to avoid closed
2655 * socket sitting in hash tables.
2656 */
2657 inet_sk_state_store(sk, state);
2658}
2659EXPORT_SYMBOL_GPL(tcp_set_state);
2660
2661/*
2662 * State processing on a close. This implements the state shift for
2663 * sending our FIN frame. Note that we only send a FIN for some
2664 * states. A shutdown() may have already sent the FIN, or we may be
2665 * closed.
2666 */
2667
2668static const unsigned char new_state[16] = {
2669 /* current state: new state: action: */
2670 [0 /* (Invalid) */] = TCP_CLOSE,
2671 [TCP_ESTABLISHED] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2672 [TCP_SYN_SENT] = TCP_CLOSE,
2673 [TCP_SYN_RECV] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2674 [TCP_FIN_WAIT1] = TCP_FIN_WAIT1,
2675 [TCP_FIN_WAIT2] = TCP_FIN_WAIT2,
2676 [TCP_TIME_WAIT] = TCP_CLOSE,
2677 [TCP_CLOSE] = TCP_CLOSE,
2678 [TCP_CLOSE_WAIT] = TCP_LAST_ACK | TCP_ACTION_FIN,
2679 [TCP_LAST_ACK] = TCP_LAST_ACK,
2680 [TCP_LISTEN] = TCP_CLOSE,
2681 [TCP_CLOSING] = TCP_CLOSING,
2682 [TCP_NEW_SYN_RECV] = TCP_CLOSE, /* should not happen ! */
2683};
2684
2685static int tcp_close_state(struct sock *sk)
2686{
2687 int next = (int)new_state[sk->sk_state];
2688 int ns = next & TCP_STATE_MASK;
2689
2690 tcp_set_state(sk, ns);
2691
2692 return next & TCP_ACTION_FIN;
2693}
2694
2695/*
2696 * Shutdown the sending side of a connection. Much like close except
2697 * that we don't receive shut down or sock_set_flag(sk, SOCK_DEAD).
2698 */
2699
2700void tcp_shutdown(struct sock *sk, int how)
2701{
2702 /* We need to grab some memory, and put together a FIN,
2703 * and then put it into the queue to be sent.
2704 * Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
2705 */
2706 if (!(how & SEND_SHUTDOWN))
2707 return;
2708
2709 /* If we've already sent a FIN, or it's a closed state, skip this. */
2710 if ((1 << sk->sk_state) &
2711 (TCPF_ESTABLISHED | TCPF_SYN_SENT |
2712 TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) {
2713 /* Clear out any half completed packets. FIN if needed. */
2714 if (tcp_close_state(sk))
2715 tcp_send_fin(sk);
2716 }
2717}
2718EXPORT_SYMBOL(tcp_shutdown);
2719
2720int tcp_orphan_count_sum(void)
2721{
2722 int i, total = 0;
2723
2724 for_each_possible_cpu(i)
2725 total += per_cpu(tcp_orphan_count, i);
2726
2727 return max(total, 0);
2728}
2729
2730static int tcp_orphan_cache;
2731static struct timer_list tcp_orphan_timer;
2732#define TCP_ORPHAN_TIMER_PERIOD msecs_to_jiffies(100)
2733
2734static void tcp_orphan_update(struct timer_list *unused)
2735{
2736 WRITE_ONCE(tcp_orphan_cache, tcp_orphan_count_sum());
2737 mod_timer(&tcp_orphan_timer, jiffies + TCP_ORPHAN_TIMER_PERIOD);
2738}
2739
2740static bool tcp_too_many_orphans(int shift)
2741{
2742 return READ_ONCE(tcp_orphan_cache) << shift >
2743 READ_ONCE(sysctl_tcp_max_orphans);
2744}
2745
2746bool tcp_check_oom(struct sock *sk, int shift)
2747{
2748 bool too_many_orphans, out_of_socket_memory;
2749
2750 too_many_orphans = tcp_too_many_orphans(shift);
2751 out_of_socket_memory = tcp_out_of_memory(sk);
2752
2753 if (too_many_orphans)
2754 net_info_ratelimited("too many orphaned sockets\n");
2755 if (out_of_socket_memory)
2756 net_info_ratelimited("out of memory -- consider tuning tcp_mem\n");
2757 return too_many_orphans || out_of_socket_memory;
2758}
2759
2760void __tcp_close(struct sock *sk, long timeout)
2761{
2762 struct sk_buff *skb;
2763 int data_was_unread = 0;
2764 int state;
2765
2766 WRITE_ONCE(sk->sk_shutdown, SHUTDOWN_MASK);
2767
2768 if (sk->sk_state == TCP_LISTEN) {
2769 tcp_set_state(sk, TCP_CLOSE);
2770
2771 /* Special case. */
2772 inet_csk_listen_stop(sk);
2773
2774 goto adjudge_to_death;
2775 }
2776
2777 /* We need to flush the recv. buffs. We do this only on the
2778 * descriptor close, not protocol-sourced closes, because the
2779 * reader process may not have drained the data yet!
2780 */
2781 while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) {
2782 u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq;
2783
2784 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
2785 len--;
2786 data_was_unread += len;
2787 __kfree_skb(skb);
2788 }
2789
2790 /* If socket has been already reset (e.g. in tcp_reset()) - kill it. */
2791 if (sk->sk_state == TCP_CLOSE)
2792 goto adjudge_to_death;
2793
2794 /* As outlined in RFC 2525, section 2.17, we send a RST here because
2795 * data was lost. To witness the awful effects of the old behavior of
2796 * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk
2797 * GET in an FTP client, suspend the process, wait for the client to
2798 * advertise a zero window, then kill -9 the FTP client, wheee...
2799 * Note: timeout is always zero in such a case.
2800 */
2801 if (unlikely(tcp_sk(sk)->repair)) {
2802 sk->sk_prot->disconnect(sk, 0);
2803 } else if (data_was_unread) {
2804 /* Unread data was tossed, zap the connection. */
2805 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONCLOSE);
2806 tcp_set_state(sk, TCP_CLOSE);
2807 tcp_send_active_reset(sk, sk->sk_allocation);
2808 } else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) {
2809 /* Check zero linger _after_ checking for unread data. */
2810 sk->sk_prot->disconnect(sk, 0);
2811 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
2812 } else if (tcp_close_state(sk)) {
2813 /* We FIN if the application ate all the data before
2814 * zapping the connection.
2815 */
2816
2817 /* RED-PEN. Formally speaking, we have broken TCP state
2818 * machine. State transitions:
2819 *
2820 * TCP_ESTABLISHED -> TCP_FIN_WAIT1
2821 * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible)
2822 * TCP_CLOSE_WAIT -> TCP_LAST_ACK
2823 *
2824 * are legal only when FIN has been sent (i.e. in window),
2825 * rather than queued out of window. Purists blame.
2826 *
2827 * F.e. "RFC state" is ESTABLISHED,
2828 * if Linux state is FIN-WAIT-1, but FIN is still not sent.
2829 *
2830 * The visible declinations are that sometimes
2831 * we enter time-wait state, when it is not required really
2832 * (harmless), do not send active resets, when they are
2833 * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
2834 * they look as CLOSING or LAST_ACK for Linux)
2835 * Probably, I missed some more holelets.
2836 * --ANK
2837 * XXX (TFO) - To start off we don't support SYN+ACK+FIN
2838 * in a single packet! (May consider it later but will
2839 * probably need API support or TCP_CORK SYN-ACK until
2840 * data is written and socket is closed.)
2841 */
2842 tcp_send_fin(sk);
2843 }
2844
2845 sk_stream_wait_close(sk, timeout);
2846
2847adjudge_to_death:
2848 state = sk->sk_state;
2849 sock_hold(sk);
2850 sock_orphan(sk);
2851
2852 local_bh_disable();
2853 bh_lock_sock(sk);
2854 /* remove backlog if any, without releasing ownership. */
2855 __release_sock(sk);
2856
2857 this_cpu_inc(tcp_orphan_count);
2858
2859 /* Have we already been destroyed by a softirq or backlog? */
2860 if (state != TCP_CLOSE && sk->sk_state == TCP_CLOSE)
2861 goto out;
2862
2863 /* This is a (useful) BSD violating of the RFC. There is a
2864 * problem with TCP as specified in that the other end could
2865 * keep a socket open forever with no application left this end.
2866 * We use a 1 minute timeout (about the same as BSD) then kill
2867 * our end. If they send after that then tough - BUT: long enough
2868 * that we won't make the old 4*rto = almost no time - whoops
2869 * reset mistake.
2870 *
2871 * Nope, it was not mistake. It is really desired behaviour
2872 * f.e. on http servers, when such sockets are useless, but
2873 * consume significant resources. Let's do it with special
2874 * linger2 option. --ANK
2875 */
2876
2877 if (sk->sk_state == TCP_FIN_WAIT2) {
2878 struct tcp_sock *tp = tcp_sk(sk);
2879 if (READ_ONCE(tp->linger2) < 0) {
2880 tcp_set_state(sk, TCP_CLOSE);
2881 tcp_send_active_reset(sk, GFP_ATOMIC);
2882 __NET_INC_STATS(sock_net(sk),
2883 LINUX_MIB_TCPABORTONLINGER);
2884 } else {
2885 const int tmo = tcp_fin_time(sk);
2886
2887 if (tmo > TCP_TIMEWAIT_LEN) {
2888 inet_csk_reset_keepalive_timer(sk,
2889 tmo - TCP_TIMEWAIT_LEN);
2890 } else {
2891 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
2892 goto out;
2893 }
2894 }
2895 }
2896 if (sk->sk_state != TCP_CLOSE) {
2897 if (tcp_check_oom(sk, 0)) {
2898 tcp_set_state(sk, TCP_CLOSE);
2899 tcp_send_active_reset(sk, GFP_ATOMIC);
2900 __NET_INC_STATS(sock_net(sk),
2901 LINUX_MIB_TCPABORTONMEMORY);
2902 } else if (!check_net(sock_net(sk))) {
2903 /* Not possible to send reset; just close */
2904 tcp_set_state(sk, TCP_CLOSE);
2905 }
2906 }
2907
2908 if (sk->sk_state == TCP_CLOSE) {
2909 struct request_sock *req;
2910
2911 req = rcu_dereference_protected(tcp_sk(sk)->fastopen_rsk,
2912 lockdep_sock_is_held(sk));
2913 /* We could get here with a non-NULL req if the socket is
2914 * aborted (e.g., closed with unread data) before 3WHS
2915 * finishes.
2916 */
2917 if (req)
2918 reqsk_fastopen_remove(sk, req, false);
2919 inet_csk_destroy_sock(sk);
2920 }
2921 /* Otherwise, socket is reprieved until protocol close. */
2922
2923out:
2924 bh_unlock_sock(sk);
2925 local_bh_enable();
2926}
2927
2928void tcp_close(struct sock *sk, long timeout)
2929{
2930 lock_sock(sk);
2931 __tcp_close(sk, timeout);
2932 release_sock(sk);
2933 sock_put(sk);
2934}
2935EXPORT_SYMBOL(tcp_close);
2936
2937/* These states need RST on ABORT according to RFC793 */
2938
2939static inline bool tcp_need_reset(int state)
2940{
2941 return (1 << state) &
2942 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 |
2943 TCPF_FIN_WAIT2 | TCPF_SYN_RECV);
2944}
2945
2946static void tcp_rtx_queue_purge(struct sock *sk)
2947{
2948 struct rb_node *p = rb_first(&sk->tcp_rtx_queue);
2949
2950 tcp_sk(sk)->highest_sack = NULL;
2951 while (p) {
2952 struct sk_buff *skb = rb_to_skb(p);
2953
2954 p = rb_next(p);
2955 /* Since we are deleting whole queue, no need to
2956 * list_del(&skb->tcp_tsorted_anchor)
2957 */
2958 tcp_rtx_queue_unlink(skb, sk);
2959 tcp_wmem_free_skb(sk, skb);
2960 }
2961}
2962
2963void tcp_write_queue_purge(struct sock *sk)
2964{
2965 struct sk_buff *skb;
2966
2967 tcp_chrono_stop(sk, TCP_CHRONO_BUSY);
2968 while ((skb = __skb_dequeue(&sk->sk_write_queue)) != NULL) {
2969 tcp_skb_tsorted_anchor_cleanup(skb);
2970 tcp_wmem_free_skb(sk, skb);
2971 }
2972 tcp_rtx_queue_purge(sk);
2973 INIT_LIST_HEAD(&tcp_sk(sk)->tsorted_sent_queue);
2974 tcp_clear_all_retrans_hints(tcp_sk(sk));
2975 tcp_sk(sk)->packets_out = 0;
2976 inet_csk(sk)->icsk_backoff = 0;
2977}
2978
2979int tcp_disconnect(struct sock *sk, int flags)
2980{
2981 struct inet_sock *inet = inet_sk(sk);
2982 struct inet_connection_sock *icsk = inet_csk(sk);
2983 struct tcp_sock *tp = tcp_sk(sk);
2984 int old_state = sk->sk_state;
2985 u32 seq;
2986
2987 if (old_state != TCP_CLOSE)
2988 tcp_set_state(sk, TCP_CLOSE);
2989
2990 /* ABORT function of RFC793 */
2991 if (old_state == TCP_LISTEN) {
2992 inet_csk_listen_stop(sk);
2993 } else if (unlikely(tp->repair)) {
2994 WRITE_ONCE(sk->sk_err, ECONNABORTED);
2995 } else if (tcp_need_reset(old_state) ||
2996 (tp->snd_nxt != tp->write_seq &&
2997 (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK))) {
2998 /* The last check adjusts for discrepancy of Linux wrt. RFC
2999 * states
3000 */
3001 tcp_send_active_reset(sk, gfp_any());
3002 WRITE_ONCE(sk->sk_err, ECONNRESET);
3003 } else if (old_state == TCP_SYN_SENT)
3004 WRITE_ONCE(sk->sk_err, ECONNRESET);
3005
3006 tcp_clear_xmit_timers(sk);
3007 __skb_queue_purge(&sk->sk_receive_queue);
3008 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt);
3009 WRITE_ONCE(tp->urg_data, 0);
3010 tcp_write_queue_purge(sk);
3011 tcp_fastopen_active_disable_ofo_check(sk);
3012 skb_rbtree_purge(&tp->out_of_order_queue);
3013
3014 inet->inet_dport = 0;
3015
3016 inet_bhash2_reset_saddr(sk);
3017
3018 WRITE_ONCE(sk->sk_shutdown, 0);
3019 sock_reset_flag(sk, SOCK_DONE);
3020 tp->srtt_us = 0;
3021 tp->mdev_us = jiffies_to_usecs(TCP_TIMEOUT_INIT);
3022 tp->rcv_rtt_last_tsecr = 0;
3023
3024 seq = tp->write_seq + tp->max_window + 2;
3025 if (!seq)
3026 seq = 1;
3027 WRITE_ONCE(tp->write_seq, seq);
3028
3029 icsk->icsk_backoff = 0;
3030 icsk->icsk_probes_out = 0;
3031 icsk->icsk_probes_tstamp = 0;
3032 icsk->icsk_rto = TCP_TIMEOUT_INIT;
3033 icsk->icsk_rto_min = TCP_RTO_MIN;
3034 icsk->icsk_delack_max = TCP_DELACK_MAX;
3035 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
3036 tcp_snd_cwnd_set(tp, TCP_INIT_CWND);
3037 tp->snd_cwnd_cnt = 0;
3038 tp->is_cwnd_limited = 0;
3039 tp->max_packets_out = 0;
3040 tp->window_clamp = 0;
3041 tp->delivered = 0;
3042 tp->delivered_ce = 0;
3043 if (icsk->icsk_ca_ops->release)
3044 icsk->icsk_ca_ops->release(sk);
3045 memset(icsk->icsk_ca_priv, 0, sizeof(icsk->icsk_ca_priv));
3046 icsk->icsk_ca_initialized = 0;
3047 tcp_set_ca_state(sk, TCP_CA_Open);
3048 tp->is_sack_reneg = 0;
3049 tcp_clear_retrans(tp);
3050 tp->total_retrans = 0;
3051 inet_csk_delack_init(sk);
3052 /* Initialize rcv_mss to TCP_MIN_MSS to avoid division by 0
3053 * issue in __tcp_select_window()
3054 */
3055 icsk->icsk_ack.rcv_mss = TCP_MIN_MSS;
3056 memset(&tp->rx_opt, 0, sizeof(tp->rx_opt));
3057 __sk_dst_reset(sk);
3058 dst_release(xchg((__force struct dst_entry **)&sk->sk_rx_dst, NULL));
3059 tcp_saved_syn_free(tp);
3060 tp->compressed_ack = 0;
3061 tp->segs_in = 0;
3062 tp->segs_out = 0;
3063 tp->bytes_sent = 0;
3064 tp->bytes_acked = 0;
3065 tp->bytes_received = 0;
3066 tp->bytes_retrans = 0;
3067 tp->data_segs_in = 0;
3068 tp->data_segs_out = 0;
3069 tp->duplicate_sack[0].start_seq = 0;
3070 tp->duplicate_sack[0].end_seq = 0;
3071 tp->dsack_dups = 0;
3072 tp->reord_seen = 0;
3073 tp->retrans_out = 0;
3074 tp->sacked_out = 0;
3075 tp->tlp_high_seq = 0;
3076 tp->last_oow_ack_time = 0;
3077 tp->plb_rehash = 0;
3078 /* There's a bubble in the pipe until at least the first ACK. */
3079 tp->app_limited = ~0U;
3080 tp->rate_app_limited = 1;
3081 tp->rack.mstamp = 0;
3082 tp->rack.advanced = 0;
3083 tp->rack.reo_wnd_steps = 1;
3084 tp->rack.last_delivered = 0;
3085 tp->rack.reo_wnd_persist = 0;
3086 tp->rack.dsack_seen = 0;
3087 tp->syn_data_acked = 0;
3088 tp->rx_opt.saw_tstamp = 0;
3089 tp->rx_opt.dsack = 0;
3090 tp->rx_opt.num_sacks = 0;
3091 tp->rcv_ooopack = 0;
3092
3093
3094 /* Clean up fastopen related fields */
3095 tcp_free_fastopen_req(tp);
3096 inet_clear_bit(DEFER_CONNECT, sk);
3097 tp->fastopen_client_fail = 0;
3098
3099 WARN_ON(inet->inet_num && !icsk->icsk_bind_hash);
3100
3101 if (sk->sk_frag.page) {
3102 put_page(sk->sk_frag.page);
3103 sk->sk_frag.page = NULL;
3104 sk->sk_frag.offset = 0;
3105 }
3106 sk_error_report(sk);
3107 return 0;
3108}
3109EXPORT_SYMBOL(tcp_disconnect);
3110
3111static inline bool tcp_can_repair_sock(const struct sock *sk)
3112{
3113 return sockopt_ns_capable(sock_net(sk)->user_ns, CAP_NET_ADMIN) &&
3114 (sk->sk_state != TCP_LISTEN);
3115}
3116
3117static int tcp_repair_set_window(struct tcp_sock *tp, sockptr_t optbuf, int len)
3118{
3119 struct tcp_repair_window opt;
3120
3121 if (!tp->repair)
3122 return -EPERM;
3123
3124 if (len != sizeof(opt))
3125 return -EINVAL;
3126
3127 if (copy_from_sockptr(&opt, optbuf, sizeof(opt)))
3128 return -EFAULT;
3129
3130 if (opt.max_window < opt.snd_wnd)
3131 return -EINVAL;
3132
3133 if (after(opt.snd_wl1, tp->rcv_nxt + opt.rcv_wnd))
3134 return -EINVAL;
3135
3136 if (after(opt.rcv_wup, tp->rcv_nxt))
3137 return -EINVAL;
3138
3139 tp->snd_wl1 = opt.snd_wl1;
3140 tp->snd_wnd = opt.snd_wnd;
3141 tp->max_window = opt.max_window;
3142
3143 tp->rcv_wnd = opt.rcv_wnd;
3144 tp->rcv_wup = opt.rcv_wup;
3145
3146 return 0;
3147}
3148
3149static int tcp_repair_options_est(struct sock *sk, sockptr_t optbuf,
3150 unsigned int len)
3151{
3152 struct tcp_sock *tp = tcp_sk(sk);
3153 struct tcp_repair_opt opt;
3154 size_t offset = 0;
3155
3156 while (len >= sizeof(opt)) {
3157 if (copy_from_sockptr_offset(&opt, optbuf, offset, sizeof(opt)))
3158 return -EFAULT;
3159
3160 offset += sizeof(opt);
3161 len -= sizeof(opt);
3162
3163 switch (opt.opt_code) {
3164 case TCPOPT_MSS:
3165 tp->rx_opt.mss_clamp = opt.opt_val;
3166 tcp_mtup_init(sk);
3167 break;
3168 case TCPOPT_WINDOW:
3169 {
3170 u16 snd_wscale = opt.opt_val & 0xFFFF;
3171 u16 rcv_wscale = opt.opt_val >> 16;
3172
3173 if (snd_wscale > TCP_MAX_WSCALE || rcv_wscale > TCP_MAX_WSCALE)
3174 return -EFBIG;
3175
3176 tp->rx_opt.snd_wscale = snd_wscale;
3177 tp->rx_opt.rcv_wscale = rcv_wscale;
3178 tp->rx_opt.wscale_ok = 1;
3179 }
3180 break;
3181 case TCPOPT_SACK_PERM:
3182 if (opt.opt_val != 0)
3183 return -EINVAL;
3184
3185 tp->rx_opt.sack_ok |= TCP_SACK_SEEN;
3186 break;
3187 case TCPOPT_TIMESTAMP:
3188 if (opt.opt_val != 0)
3189 return -EINVAL;
3190
3191 tp->rx_opt.tstamp_ok = 1;
3192 break;
3193 }
3194 }
3195
3196 return 0;
3197}
3198
3199DEFINE_STATIC_KEY_FALSE(tcp_tx_delay_enabled);
3200EXPORT_SYMBOL(tcp_tx_delay_enabled);
3201
3202static void tcp_enable_tx_delay(void)
3203{
3204 if (!static_branch_unlikely(&tcp_tx_delay_enabled)) {
3205 static int __tcp_tx_delay_enabled = 0;
3206
3207 if (cmpxchg(&__tcp_tx_delay_enabled, 0, 1) == 0) {
3208 static_branch_enable(&tcp_tx_delay_enabled);
3209 pr_info("TCP_TX_DELAY enabled\n");
3210 }
3211 }
3212}
3213
3214/* When set indicates to always queue non-full frames. Later the user clears
3215 * this option and we transmit any pending partial frames in the queue. This is
3216 * meant to be used alongside sendfile() to get properly filled frames when the
3217 * user (for example) must write out headers with a write() call first and then
3218 * use sendfile to send out the data parts.
3219 *
3220 * TCP_CORK can be set together with TCP_NODELAY and it is stronger than
3221 * TCP_NODELAY.
3222 */
3223void __tcp_sock_set_cork(struct sock *sk, bool on)
3224{
3225 struct tcp_sock *tp = tcp_sk(sk);
3226
3227 if (on) {
3228 tp->nonagle |= TCP_NAGLE_CORK;
3229 } else {
3230 tp->nonagle &= ~TCP_NAGLE_CORK;
3231 if (tp->nonagle & TCP_NAGLE_OFF)
3232 tp->nonagle |= TCP_NAGLE_PUSH;
3233 tcp_push_pending_frames(sk);
3234 }
3235}
3236
3237void tcp_sock_set_cork(struct sock *sk, bool on)
3238{
3239 lock_sock(sk);
3240 __tcp_sock_set_cork(sk, on);
3241 release_sock(sk);
3242}
3243EXPORT_SYMBOL(tcp_sock_set_cork);
3244
3245/* TCP_NODELAY is weaker than TCP_CORK, so that this option on corked socket is
3246 * remembered, but it is not activated until cork is cleared.
3247 *
3248 * However, when TCP_NODELAY is set we make an explicit push, which overrides
3249 * even TCP_CORK for currently queued segments.
3250 */
3251void __tcp_sock_set_nodelay(struct sock *sk, bool on)
3252{
3253 if (on) {
3254 tcp_sk(sk)->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH;
3255 tcp_push_pending_frames(sk);
3256 } else {
3257 tcp_sk(sk)->nonagle &= ~TCP_NAGLE_OFF;
3258 }
3259}
3260
3261void tcp_sock_set_nodelay(struct sock *sk)
3262{
3263 lock_sock(sk);
3264 __tcp_sock_set_nodelay(sk, true);
3265 release_sock(sk);
3266}
3267EXPORT_SYMBOL(tcp_sock_set_nodelay);
3268
3269static void __tcp_sock_set_quickack(struct sock *sk, int val)
3270{
3271 if (!val) {
3272 inet_csk_enter_pingpong_mode(sk);
3273 return;
3274 }
3275
3276 inet_csk_exit_pingpong_mode(sk);
3277 if ((1 << sk->sk_state) & (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) &&
3278 inet_csk_ack_scheduled(sk)) {
3279 inet_csk(sk)->icsk_ack.pending |= ICSK_ACK_PUSHED;
3280 tcp_cleanup_rbuf(sk, 1);
3281 if (!(val & 1))
3282 inet_csk_enter_pingpong_mode(sk);
3283 }
3284}
3285
3286void tcp_sock_set_quickack(struct sock *sk, int val)
3287{
3288 lock_sock(sk);
3289 __tcp_sock_set_quickack(sk, val);
3290 release_sock(sk);
3291}
3292EXPORT_SYMBOL(tcp_sock_set_quickack);
3293
3294int tcp_sock_set_syncnt(struct sock *sk, int val)
3295{
3296 if (val < 1 || val > MAX_TCP_SYNCNT)
3297 return -EINVAL;
3298
3299 WRITE_ONCE(inet_csk(sk)->icsk_syn_retries, val);
3300 return 0;
3301}
3302EXPORT_SYMBOL(tcp_sock_set_syncnt);
3303
3304int tcp_sock_set_user_timeout(struct sock *sk, int val)
3305{
3306 /* Cap the max time in ms TCP will retry or probe the window
3307 * before giving up and aborting (ETIMEDOUT) a connection.
3308 */
3309 if (val < 0)
3310 return -EINVAL;
3311
3312 WRITE_ONCE(inet_csk(sk)->icsk_user_timeout, val);
3313 return 0;
3314}
3315EXPORT_SYMBOL(tcp_sock_set_user_timeout);
3316
3317int tcp_sock_set_keepidle_locked(struct sock *sk, int val)
3318{
3319 struct tcp_sock *tp = tcp_sk(sk);
3320
3321 if (val < 1 || val > MAX_TCP_KEEPIDLE)
3322 return -EINVAL;
3323
3324 /* Paired with WRITE_ONCE() in keepalive_time_when() */
3325 WRITE_ONCE(tp->keepalive_time, val * HZ);
3326 if (sock_flag(sk, SOCK_KEEPOPEN) &&
3327 !((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN))) {
3328 u32 elapsed = keepalive_time_elapsed(tp);
3329
3330 if (tp->keepalive_time > elapsed)
3331 elapsed = tp->keepalive_time - elapsed;
3332 else
3333 elapsed = 0;
3334 inet_csk_reset_keepalive_timer(sk, elapsed);
3335 }
3336
3337 return 0;
3338}
3339
3340int tcp_sock_set_keepidle(struct sock *sk, int val)
3341{
3342 int err;
3343
3344 lock_sock(sk);
3345 err = tcp_sock_set_keepidle_locked(sk, val);
3346 release_sock(sk);
3347 return err;
3348}
3349EXPORT_SYMBOL(tcp_sock_set_keepidle);
3350
3351int tcp_sock_set_keepintvl(struct sock *sk, int val)
3352{
3353 if (val < 1 || val > MAX_TCP_KEEPINTVL)
3354 return -EINVAL;
3355
3356 WRITE_ONCE(tcp_sk(sk)->keepalive_intvl, val * HZ);
3357 return 0;
3358}
3359EXPORT_SYMBOL(tcp_sock_set_keepintvl);
3360
3361int tcp_sock_set_keepcnt(struct sock *sk, int val)
3362{
3363 if (val < 1 || val > MAX_TCP_KEEPCNT)
3364 return -EINVAL;
3365
3366 /* Paired with READ_ONCE() in keepalive_probes() */
3367 WRITE_ONCE(tcp_sk(sk)->keepalive_probes, val);
3368 return 0;
3369}
3370EXPORT_SYMBOL(tcp_sock_set_keepcnt);
3371
3372int tcp_set_window_clamp(struct sock *sk, int val)
3373{
3374 struct tcp_sock *tp = tcp_sk(sk);
3375
3376 if (!val) {
3377 if (sk->sk_state != TCP_CLOSE)
3378 return -EINVAL;
3379 tp->window_clamp = 0;
3380 } else {
3381 u32 new_rcv_ssthresh, old_window_clamp = tp->window_clamp;
3382 u32 new_window_clamp = val < SOCK_MIN_RCVBUF / 2 ?
3383 SOCK_MIN_RCVBUF / 2 : val;
3384
3385 if (new_window_clamp == old_window_clamp)
3386 return 0;
3387
3388 tp->window_clamp = new_window_clamp;
3389 if (new_window_clamp < old_window_clamp) {
3390 /* need to apply the reserved mem provisioning only
3391 * when shrinking the window clamp
3392 */
3393 __tcp_adjust_rcv_ssthresh(sk, tp->window_clamp);
3394
3395 } else {
3396 new_rcv_ssthresh = min(tp->rcv_wnd, tp->window_clamp);
3397 tp->rcv_ssthresh = max(new_rcv_ssthresh,
3398 tp->rcv_ssthresh);
3399 }
3400 }
3401 return 0;
3402}
3403
3404/*
3405 * Socket option code for TCP.
3406 */
3407int do_tcp_setsockopt(struct sock *sk, int level, int optname,
3408 sockptr_t optval, unsigned int optlen)
3409{
3410 struct tcp_sock *tp = tcp_sk(sk);
3411 struct inet_connection_sock *icsk = inet_csk(sk);
3412 struct net *net = sock_net(sk);
3413 int val;
3414 int err = 0;
3415
3416 /* These are data/string values, all the others are ints */
3417 switch (optname) {
3418 case TCP_CONGESTION: {
3419 char name[TCP_CA_NAME_MAX];
3420
3421 if (optlen < 1)
3422 return -EINVAL;
3423
3424 val = strncpy_from_sockptr(name, optval,
3425 min_t(long, TCP_CA_NAME_MAX-1, optlen));
3426 if (val < 0)
3427 return -EFAULT;
3428 name[val] = 0;
3429
3430 sockopt_lock_sock(sk);
3431 err = tcp_set_congestion_control(sk, name, !has_current_bpf_ctx(),
3432 sockopt_ns_capable(sock_net(sk)->user_ns,
3433 CAP_NET_ADMIN));
3434 sockopt_release_sock(sk);
3435 return err;
3436 }
3437 case TCP_ULP: {
3438 char name[TCP_ULP_NAME_MAX];
3439
3440 if (optlen < 1)
3441 return -EINVAL;
3442
3443 val = strncpy_from_sockptr(name, optval,
3444 min_t(long, TCP_ULP_NAME_MAX - 1,
3445 optlen));
3446 if (val < 0)
3447 return -EFAULT;
3448 name[val] = 0;
3449
3450 sockopt_lock_sock(sk);
3451 err = tcp_set_ulp(sk, name);
3452 sockopt_release_sock(sk);
3453 return err;
3454 }
3455 case TCP_FASTOPEN_KEY: {
3456 __u8 key[TCP_FASTOPEN_KEY_BUF_LENGTH];
3457 __u8 *backup_key = NULL;
3458
3459 /* Allow a backup key as well to facilitate key rotation
3460 * First key is the active one.
3461 */
3462 if (optlen != TCP_FASTOPEN_KEY_LENGTH &&
3463 optlen != TCP_FASTOPEN_KEY_BUF_LENGTH)
3464 return -EINVAL;
3465
3466 if (copy_from_sockptr(key, optval, optlen))
3467 return -EFAULT;
3468
3469 if (optlen == TCP_FASTOPEN_KEY_BUF_LENGTH)
3470 backup_key = key + TCP_FASTOPEN_KEY_LENGTH;
3471
3472 return tcp_fastopen_reset_cipher(net, sk, key, backup_key);
3473 }
3474 default:
3475 /* fallthru */
3476 break;
3477 }
3478
3479 if (optlen < sizeof(int))
3480 return -EINVAL;
3481
3482 if (copy_from_sockptr(&val, optval, sizeof(val)))
3483 return -EFAULT;
3484
3485 /* Handle options that can be set without locking the socket. */
3486 switch (optname) {
3487 case TCP_SYNCNT:
3488 return tcp_sock_set_syncnt(sk, val);
3489 case TCP_USER_TIMEOUT:
3490 return tcp_sock_set_user_timeout(sk, val);
3491 case TCP_KEEPINTVL:
3492 return tcp_sock_set_keepintvl(sk, val);
3493 case TCP_KEEPCNT:
3494 return tcp_sock_set_keepcnt(sk, val);
3495 case TCP_LINGER2:
3496 if (val < 0)
3497 WRITE_ONCE(tp->linger2, -1);
3498 else if (val > TCP_FIN_TIMEOUT_MAX / HZ)
3499 WRITE_ONCE(tp->linger2, TCP_FIN_TIMEOUT_MAX);
3500 else
3501 WRITE_ONCE(tp->linger2, val * HZ);
3502 return 0;
3503 case TCP_DEFER_ACCEPT:
3504 /* Translate value in seconds to number of retransmits */
3505 WRITE_ONCE(icsk->icsk_accept_queue.rskq_defer_accept,
3506 secs_to_retrans(val, TCP_TIMEOUT_INIT / HZ,
3507 TCP_RTO_MAX / HZ));
3508 return 0;
3509 }
3510
3511 sockopt_lock_sock(sk);
3512
3513 switch (optname) {
3514 case TCP_MAXSEG:
3515 /* Values greater than interface MTU won't take effect. However
3516 * at the point when this call is done we typically don't yet
3517 * know which interface is going to be used
3518 */
3519 if (val && (val < TCP_MIN_MSS || val > MAX_TCP_WINDOW)) {
3520 err = -EINVAL;
3521 break;
3522 }
3523 tp->rx_opt.user_mss = val;
3524 break;
3525
3526 case TCP_NODELAY:
3527 __tcp_sock_set_nodelay(sk, val);
3528 break;
3529
3530 case TCP_THIN_LINEAR_TIMEOUTS:
3531 if (val < 0 || val > 1)
3532 err = -EINVAL;
3533 else
3534 tp->thin_lto = val;
3535 break;
3536
3537 case TCP_THIN_DUPACK:
3538 if (val < 0 || val > 1)
3539 err = -EINVAL;
3540 break;
3541
3542 case TCP_REPAIR:
3543 if (!tcp_can_repair_sock(sk))
3544 err = -EPERM;
3545 else if (val == TCP_REPAIR_ON) {
3546 tp->repair = 1;
3547 sk->sk_reuse = SK_FORCE_REUSE;
3548 tp->repair_queue = TCP_NO_QUEUE;
3549 } else if (val == TCP_REPAIR_OFF) {
3550 tp->repair = 0;
3551 sk->sk_reuse = SK_NO_REUSE;
3552 tcp_send_window_probe(sk);
3553 } else if (val == TCP_REPAIR_OFF_NO_WP) {
3554 tp->repair = 0;
3555 sk->sk_reuse = SK_NO_REUSE;
3556 } else
3557 err = -EINVAL;
3558
3559 break;
3560
3561 case TCP_REPAIR_QUEUE:
3562 if (!tp->repair)
3563 err = -EPERM;
3564 else if ((unsigned int)val < TCP_QUEUES_NR)
3565 tp->repair_queue = val;
3566 else
3567 err = -EINVAL;
3568 break;
3569
3570 case TCP_QUEUE_SEQ:
3571 if (sk->sk_state != TCP_CLOSE) {
3572 err = -EPERM;
3573 } else if (tp->repair_queue == TCP_SEND_QUEUE) {
3574 if (!tcp_rtx_queue_empty(sk))
3575 err = -EPERM;
3576 else
3577 WRITE_ONCE(tp->write_seq, val);
3578 } else if (tp->repair_queue == TCP_RECV_QUEUE) {
3579 if (tp->rcv_nxt != tp->copied_seq) {
3580 err = -EPERM;
3581 } else {
3582 WRITE_ONCE(tp->rcv_nxt, val);
3583 WRITE_ONCE(tp->copied_seq, val);
3584 }
3585 } else {
3586 err = -EINVAL;
3587 }
3588 break;
3589
3590 case TCP_REPAIR_OPTIONS:
3591 if (!tp->repair)
3592 err = -EINVAL;
3593 else if (sk->sk_state == TCP_ESTABLISHED && !tp->bytes_sent)
3594 err = tcp_repair_options_est(sk, optval, optlen);
3595 else
3596 err = -EPERM;
3597 break;
3598
3599 case TCP_CORK:
3600 __tcp_sock_set_cork(sk, val);
3601 break;
3602
3603 case TCP_KEEPIDLE:
3604 err = tcp_sock_set_keepidle_locked(sk, val);
3605 break;
3606 case TCP_SAVE_SYN:
3607 /* 0: disable, 1: enable, 2: start from ether_header */
3608 if (val < 0 || val > 2)
3609 err = -EINVAL;
3610 else
3611 tp->save_syn = val;
3612 break;
3613
3614 case TCP_WINDOW_CLAMP:
3615 err = tcp_set_window_clamp(sk, val);
3616 break;
3617
3618 case TCP_QUICKACK:
3619 __tcp_sock_set_quickack(sk, val);
3620 break;
3621
3622 case TCP_AO_REPAIR:
3623 if (!tcp_can_repair_sock(sk)) {
3624 err = -EPERM;
3625 break;
3626 }
3627 err = tcp_ao_set_repair(sk, optval, optlen);
3628 break;
3629#ifdef CONFIG_TCP_AO
3630 case TCP_AO_ADD_KEY:
3631 case TCP_AO_DEL_KEY:
3632 case TCP_AO_INFO: {
3633 /* If this is the first TCP-AO setsockopt() on the socket,
3634 * sk_state has to be LISTEN or CLOSE. Allow TCP_REPAIR
3635 * in any state.
3636 */
3637 if ((1 << sk->sk_state) & (TCPF_LISTEN | TCPF_CLOSE))
3638 goto ao_parse;
3639 if (rcu_dereference_protected(tcp_sk(sk)->ao_info,
3640 lockdep_sock_is_held(sk)))
3641 goto ao_parse;
3642 if (tp->repair)
3643 goto ao_parse;
3644 err = -EISCONN;
3645 break;
3646ao_parse:
3647 err = tp->af_specific->ao_parse(sk, optname, optval, optlen);
3648 break;
3649 }
3650#endif
3651#ifdef CONFIG_TCP_MD5SIG
3652 case TCP_MD5SIG:
3653 case TCP_MD5SIG_EXT:
3654 err = tp->af_specific->md5_parse(sk, optname, optval, optlen);
3655 break;
3656#endif
3657 case TCP_FASTOPEN:
3658 if (val >= 0 && ((1 << sk->sk_state) & (TCPF_CLOSE |
3659 TCPF_LISTEN))) {
3660 tcp_fastopen_init_key_once(net);
3661
3662 fastopen_queue_tune(sk, val);
3663 } else {
3664 err = -EINVAL;
3665 }
3666 break;
3667 case TCP_FASTOPEN_CONNECT:
3668 if (val > 1 || val < 0) {
3669 err = -EINVAL;
3670 } else if (READ_ONCE(net->ipv4.sysctl_tcp_fastopen) &
3671 TFO_CLIENT_ENABLE) {
3672 if (sk->sk_state == TCP_CLOSE)
3673 tp->fastopen_connect = val;
3674 else
3675 err = -EINVAL;
3676 } else {
3677 err = -EOPNOTSUPP;
3678 }
3679 break;
3680 case TCP_FASTOPEN_NO_COOKIE:
3681 if (val > 1 || val < 0)
3682 err = -EINVAL;
3683 else if (!((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
3684 err = -EINVAL;
3685 else
3686 tp->fastopen_no_cookie = val;
3687 break;
3688 case TCP_TIMESTAMP:
3689 if (!tp->repair) {
3690 err = -EPERM;
3691 break;
3692 }
3693 /* val is an opaque field,
3694 * and low order bit contains usec_ts enable bit.
3695 * Its a best effort, and we do not care if user makes an error.
3696 */
3697 tp->tcp_usec_ts = val & 1;
3698 WRITE_ONCE(tp->tsoffset, val - tcp_clock_ts(tp->tcp_usec_ts));
3699 break;
3700 case TCP_REPAIR_WINDOW:
3701 err = tcp_repair_set_window(tp, optval, optlen);
3702 break;
3703 case TCP_NOTSENT_LOWAT:
3704 WRITE_ONCE(tp->notsent_lowat, val);
3705 sk->sk_write_space(sk);
3706 break;
3707 case TCP_INQ:
3708 if (val > 1 || val < 0)
3709 err = -EINVAL;
3710 else
3711 tp->recvmsg_inq = val;
3712 break;
3713 case TCP_TX_DELAY:
3714 if (val)
3715 tcp_enable_tx_delay();
3716 WRITE_ONCE(tp->tcp_tx_delay, val);
3717 break;
3718 default:
3719 err = -ENOPROTOOPT;
3720 break;
3721 }
3722
3723 sockopt_release_sock(sk);
3724 return err;
3725}
3726
3727int tcp_setsockopt(struct sock *sk, int level, int optname, sockptr_t optval,
3728 unsigned int optlen)
3729{
3730 const struct inet_connection_sock *icsk = inet_csk(sk);
3731
3732 if (level != SOL_TCP)
3733 /* Paired with WRITE_ONCE() in do_ipv6_setsockopt() and tcp_v6_connect() */
3734 return READ_ONCE(icsk->icsk_af_ops)->setsockopt(sk, level, optname,
3735 optval, optlen);
3736 return do_tcp_setsockopt(sk, level, optname, optval, optlen);
3737}
3738EXPORT_SYMBOL(tcp_setsockopt);
3739
3740static void tcp_get_info_chrono_stats(const struct tcp_sock *tp,
3741 struct tcp_info *info)
3742{
3743 u64 stats[__TCP_CHRONO_MAX], total = 0;
3744 enum tcp_chrono i;
3745
3746 for (i = TCP_CHRONO_BUSY; i < __TCP_CHRONO_MAX; ++i) {
3747 stats[i] = tp->chrono_stat[i - 1];
3748 if (i == tp->chrono_type)
3749 stats[i] += tcp_jiffies32 - tp->chrono_start;
3750 stats[i] *= USEC_PER_SEC / HZ;
3751 total += stats[i];
3752 }
3753
3754 info->tcpi_busy_time = total;
3755 info->tcpi_rwnd_limited = stats[TCP_CHRONO_RWND_LIMITED];
3756 info->tcpi_sndbuf_limited = stats[TCP_CHRONO_SNDBUF_LIMITED];
3757}
3758
3759/* Return information about state of tcp endpoint in API format. */
3760void tcp_get_info(struct sock *sk, struct tcp_info *info)
3761{
3762 const struct tcp_sock *tp = tcp_sk(sk); /* iff sk_type == SOCK_STREAM */
3763 const struct inet_connection_sock *icsk = inet_csk(sk);
3764 unsigned long rate;
3765 u32 now;
3766 u64 rate64;
3767 bool slow;
3768
3769 memset(info, 0, sizeof(*info));
3770 if (sk->sk_type != SOCK_STREAM)
3771 return;
3772
3773 info->tcpi_state = inet_sk_state_load(sk);
3774
3775 /* Report meaningful fields for all TCP states, including listeners */
3776 rate = READ_ONCE(sk->sk_pacing_rate);
3777 rate64 = (rate != ~0UL) ? rate : ~0ULL;
3778 info->tcpi_pacing_rate = rate64;
3779
3780 rate = READ_ONCE(sk->sk_max_pacing_rate);
3781 rate64 = (rate != ~0UL) ? rate : ~0ULL;
3782 info->tcpi_max_pacing_rate = rate64;
3783
3784 info->tcpi_reordering = tp->reordering;
3785 info->tcpi_snd_cwnd = tcp_snd_cwnd(tp);
3786
3787 if (info->tcpi_state == TCP_LISTEN) {
3788 /* listeners aliased fields :
3789 * tcpi_unacked -> Number of children ready for accept()
3790 * tcpi_sacked -> max backlog
3791 */
3792 info->tcpi_unacked = READ_ONCE(sk->sk_ack_backlog);
3793 info->tcpi_sacked = READ_ONCE(sk->sk_max_ack_backlog);
3794 return;
3795 }
3796
3797 slow = lock_sock_fast(sk);
3798
3799 info->tcpi_ca_state = icsk->icsk_ca_state;
3800 info->tcpi_retransmits = icsk->icsk_retransmits;
3801 info->tcpi_probes = icsk->icsk_probes_out;
3802 info->tcpi_backoff = icsk->icsk_backoff;
3803
3804 if (tp->rx_opt.tstamp_ok)
3805 info->tcpi_options |= TCPI_OPT_TIMESTAMPS;
3806 if (tcp_is_sack(tp))
3807 info->tcpi_options |= TCPI_OPT_SACK;
3808 if (tp->rx_opt.wscale_ok) {
3809 info->tcpi_options |= TCPI_OPT_WSCALE;
3810 info->tcpi_snd_wscale = tp->rx_opt.snd_wscale;
3811 info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale;
3812 }
3813
3814 if (tp->ecn_flags & TCP_ECN_OK)
3815 info->tcpi_options |= TCPI_OPT_ECN;
3816 if (tp->ecn_flags & TCP_ECN_SEEN)
3817 info->tcpi_options |= TCPI_OPT_ECN_SEEN;
3818 if (tp->syn_data_acked)
3819 info->tcpi_options |= TCPI_OPT_SYN_DATA;
3820 if (tp->tcp_usec_ts)
3821 info->tcpi_options |= TCPI_OPT_USEC_TS;
3822
3823 info->tcpi_rto = jiffies_to_usecs(icsk->icsk_rto);
3824 info->tcpi_ato = jiffies_to_usecs(min_t(u32, icsk->icsk_ack.ato,
3825 tcp_delack_max(sk)));
3826 info->tcpi_snd_mss = tp->mss_cache;
3827 info->tcpi_rcv_mss = icsk->icsk_ack.rcv_mss;
3828
3829 info->tcpi_unacked = tp->packets_out;
3830 info->tcpi_sacked = tp->sacked_out;
3831
3832 info->tcpi_lost = tp->lost_out;
3833 info->tcpi_retrans = tp->retrans_out;
3834
3835 now = tcp_jiffies32;
3836 info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime);
3837 info->tcpi_last_data_recv = jiffies_to_msecs(now - icsk->icsk_ack.lrcvtime);
3838 info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp);
3839
3840 info->tcpi_pmtu = icsk->icsk_pmtu_cookie;
3841 info->tcpi_rcv_ssthresh = tp->rcv_ssthresh;
3842 info->tcpi_rtt = tp->srtt_us >> 3;
3843 info->tcpi_rttvar = tp->mdev_us >> 2;
3844 info->tcpi_snd_ssthresh = tp->snd_ssthresh;
3845 info->tcpi_advmss = tp->advmss;
3846
3847 info->tcpi_rcv_rtt = tp->rcv_rtt_est.rtt_us >> 3;
3848 info->tcpi_rcv_space = tp->rcvq_space.space;
3849
3850 info->tcpi_total_retrans = tp->total_retrans;
3851
3852 info->tcpi_bytes_acked = tp->bytes_acked;
3853 info->tcpi_bytes_received = tp->bytes_received;
3854 info->tcpi_notsent_bytes = max_t(int, 0, tp->write_seq - tp->snd_nxt);
3855 tcp_get_info_chrono_stats(tp, info);
3856
3857 info->tcpi_segs_out = tp->segs_out;
3858
3859 /* segs_in and data_segs_in can be updated from tcp_segs_in() from BH */
3860 info->tcpi_segs_in = READ_ONCE(tp->segs_in);
3861 info->tcpi_data_segs_in = READ_ONCE(tp->data_segs_in);
3862
3863 info->tcpi_min_rtt = tcp_min_rtt(tp);
3864 info->tcpi_data_segs_out = tp->data_segs_out;
3865
3866 info->tcpi_delivery_rate_app_limited = tp->rate_app_limited ? 1 : 0;
3867 rate64 = tcp_compute_delivery_rate(tp);
3868 if (rate64)
3869 info->tcpi_delivery_rate = rate64;
3870 info->tcpi_delivered = tp->delivered;
3871 info->tcpi_delivered_ce = tp->delivered_ce;
3872 info->tcpi_bytes_sent = tp->bytes_sent;
3873 info->tcpi_bytes_retrans = tp->bytes_retrans;
3874 info->tcpi_dsack_dups = tp->dsack_dups;
3875 info->tcpi_reord_seen = tp->reord_seen;
3876 info->tcpi_rcv_ooopack = tp->rcv_ooopack;
3877 info->tcpi_snd_wnd = tp->snd_wnd;
3878 info->tcpi_rcv_wnd = tp->rcv_wnd;
3879 info->tcpi_rehash = tp->plb_rehash + tp->timeout_rehash;
3880 info->tcpi_fastopen_client_fail = tp->fastopen_client_fail;
3881
3882 info->tcpi_total_rto = tp->total_rto;
3883 info->tcpi_total_rto_recoveries = tp->total_rto_recoveries;
3884 info->tcpi_total_rto_time = tp->total_rto_time;
3885 if (tp->rto_stamp)
3886 info->tcpi_total_rto_time += tcp_clock_ms() - tp->rto_stamp;
3887
3888 unlock_sock_fast(sk, slow);
3889}
3890EXPORT_SYMBOL_GPL(tcp_get_info);
3891
3892static size_t tcp_opt_stats_get_size(void)
3893{
3894 return
3895 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_BUSY */
3896 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_RWND_LIMITED */
3897 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_SNDBUF_LIMITED */
3898 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_DATA_SEGS_OUT */
3899 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_TOTAL_RETRANS */
3900 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_PACING_RATE */
3901 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_DELIVERY_RATE */
3902 nla_total_size(sizeof(u32)) + /* TCP_NLA_SND_CWND */
3903 nla_total_size(sizeof(u32)) + /* TCP_NLA_REORDERING */
3904 nla_total_size(sizeof(u32)) + /* TCP_NLA_MIN_RTT */
3905 nla_total_size(sizeof(u8)) + /* TCP_NLA_RECUR_RETRANS */
3906 nla_total_size(sizeof(u8)) + /* TCP_NLA_DELIVERY_RATE_APP_LMT */
3907 nla_total_size(sizeof(u32)) + /* TCP_NLA_SNDQ_SIZE */
3908 nla_total_size(sizeof(u8)) + /* TCP_NLA_CA_STATE */
3909 nla_total_size(sizeof(u32)) + /* TCP_NLA_SND_SSTHRESH */
3910 nla_total_size(sizeof(u32)) + /* TCP_NLA_DELIVERED */
3911 nla_total_size(sizeof(u32)) + /* TCP_NLA_DELIVERED_CE */
3912 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_BYTES_SENT */
3913 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_BYTES_RETRANS */
3914 nla_total_size(sizeof(u32)) + /* TCP_NLA_DSACK_DUPS */
3915 nla_total_size(sizeof(u32)) + /* TCP_NLA_REORD_SEEN */
3916 nla_total_size(sizeof(u32)) + /* TCP_NLA_SRTT */
3917 nla_total_size(sizeof(u16)) + /* TCP_NLA_TIMEOUT_REHASH */
3918 nla_total_size(sizeof(u32)) + /* TCP_NLA_BYTES_NOTSENT */
3919 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_EDT */
3920 nla_total_size(sizeof(u8)) + /* TCP_NLA_TTL */
3921 nla_total_size(sizeof(u32)) + /* TCP_NLA_REHASH */
3922 0;
3923}
3924
3925/* Returns TTL or hop limit of an incoming packet from skb. */
3926static u8 tcp_skb_ttl_or_hop_limit(const struct sk_buff *skb)
3927{
3928 if (skb->protocol == htons(ETH_P_IP))
3929 return ip_hdr(skb)->ttl;
3930 else if (skb->protocol == htons(ETH_P_IPV6))
3931 return ipv6_hdr(skb)->hop_limit;
3932 else
3933 return 0;
3934}
3935
3936struct sk_buff *tcp_get_timestamping_opt_stats(const struct sock *sk,
3937 const struct sk_buff *orig_skb,
3938 const struct sk_buff *ack_skb)
3939{
3940 const struct tcp_sock *tp = tcp_sk(sk);
3941 struct sk_buff *stats;
3942 struct tcp_info info;
3943 unsigned long rate;
3944 u64 rate64;
3945
3946 stats = alloc_skb(tcp_opt_stats_get_size(), GFP_ATOMIC);
3947 if (!stats)
3948 return NULL;
3949
3950 tcp_get_info_chrono_stats(tp, &info);
3951 nla_put_u64_64bit(stats, TCP_NLA_BUSY,
3952 info.tcpi_busy_time, TCP_NLA_PAD);
3953 nla_put_u64_64bit(stats, TCP_NLA_RWND_LIMITED,
3954 info.tcpi_rwnd_limited, TCP_NLA_PAD);
3955 nla_put_u64_64bit(stats, TCP_NLA_SNDBUF_LIMITED,
3956 info.tcpi_sndbuf_limited, TCP_NLA_PAD);
3957 nla_put_u64_64bit(stats, TCP_NLA_DATA_SEGS_OUT,
3958 tp->data_segs_out, TCP_NLA_PAD);
3959 nla_put_u64_64bit(stats, TCP_NLA_TOTAL_RETRANS,
3960 tp->total_retrans, TCP_NLA_PAD);
3961
3962 rate = READ_ONCE(sk->sk_pacing_rate);
3963 rate64 = (rate != ~0UL) ? rate : ~0ULL;
3964 nla_put_u64_64bit(stats, TCP_NLA_PACING_RATE, rate64, TCP_NLA_PAD);
3965
3966 rate64 = tcp_compute_delivery_rate(tp);
3967 nla_put_u64_64bit(stats, TCP_NLA_DELIVERY_RATE, rate64, TCP_NLA_PAD);
3968
3969 nla_put_u32(stats, TCP_NLA_SND_CWND, tcp_snd_cwnd(tp));
3970 nla_put_u32(stats, TCP_NLA_REORDERING, tp->reordering);
3971 nla_put_u32(stats, TCP_NLA_MIN_RTT, tcp_min_rtt(tp));
3972
3973 nla_put_u8(stats, TCP_NLA_RECUR_RETRANS, inet_csk(sk)->icsk_retransmits);
3974 nla_put_u8(stats, TCP_NLA_DELIVERY_RATE_APP_LMT, !!tp->rate_app_limited);
3975 nla_put_u32(stats, TCP_NLA_SND_SSTHRESH, tp->snd_ssthresh);
3976 nla_put_u32(stats, TCP_NLA_DELIVERED, tp->delivered);
3977 nla_put_u32(stats, TCP_NLA_DELIVERED_CE, tp->delivered_ce);
3978
3979 nla_put_u32(stats, TCP_NLA_SNDQ_SIZE, tp->write_seq - tp->snd_una);
3980 nla_put_u8(stats, TCP_NLA_CA_STATE, inet_csk(sk)->icsk_ca_state);
3981
3982 nla_put_u64_64bit(stats, TCP_NLA_BYTES_SENT, tp->bytes_sent,
3983 TCP_NLA_PAD);
3984 nla_put_u64_64bit(stats, TCP_NLA_BYTES_RETRANS, tp->bytes_retrans,
3985 TCP_NLA_PAD);
3986 nla_put_u32(stats, TCP_NLA_DSACK_DUPS, tp->dsack_dups);
3987 nla_put_u32(stats, TCP_NLA_REORD_SEEN, tp->reord_seen);
3988 nla_put_u32(stats, TCP_NLA_SRTT, tp->srtt_us >> 3);
3989 nla_put_u16(stats, TCP_NLA_TIMEOUT_REHASH, tp->timeout_rehash);
3990 nla_put_u32(stats, TCP_NLA_BYTES_NOTSENT,
3991 max_t(int, 0, tp->write_seq - tp->snd_nxt));
3992 nla_put_u64_64bit(stats, TCP_NLA_EDT, orig_skb->skb_mstamp_ns,
3993 TCP_NLA_PAD);
3994 if (ack_skb)
3995 nla_put_u8(stats, TCP_NLA_TTL,
3996 tcp_skb_ttl_or_hop_limit(ack_skb));
3997
3998 nla_put_u32(stats, TCP_NLA_REHASH, tp->plb_rehash + tp->timeout_rehash);
3999 return stats;
4000}
4001
4002int do_tcp_getsockopt(struct sock *sk, int level,
4003 int optname, sockptr_t optval, sockptr_t optlen)
4004{
4005 struct inet_connection_sock *icsk = inet_csk(sk);
4006 struct tcp_sock *tp = tcp_sk(sk);
4007 struct net *net = sock_net(sk);
4008 int val, len;
4009
4010 if (copy_from_sockptr(&len, optlen, sizeof(int)))
4011 return -EFAULT;
4012
4013 len = min_t(unsigned int, len, sizeof(int));
4014
4015 if (len < 0)
4016 return -EINVAL;
4017
4018 switch (optname) {
4019 case TCP_MAXSEG:
4020 val = tp->mss_cache;
4021 if (tp->rx_opt.user_mss &&
4022 ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
4023 val = tp->rx_opt.user_mss;
4024 if (tp->repair)
4025 val = tp->rx_opt.mss_clamp;
4026 break;
4027 case TCP_NODELAY:
4028 val = !!(tp->nonagle&TCP_NAGLE_OFF);
4029 break;
4030 case TCP_CORK:
4031 val = !!(tp->nonagle&TCP_NAGLE_CORK);
4032 break;
4033 case TCP_KEEPIDLE:
4034 val = keepalive_time_when(tp) / HZ;
4035 break;
4036 case TCP_KEEPINTVL:
4037 val = keepalive_intvl_when(tp) / HZ;
4038 break;
4039 case TCP_KEEPCNT:
4040 val = keepalive_probes(tp);
4041 break;
4042 case TCP_SYNCNT:
4043 val = READ_ONCE(icsk->icsk_syn_retries) ? :
4044 READ_ONCE(net->ipv4.sysctl_tcp_syn_retries);
4045 break;
4046 case TCP_LINGER2:
4047 val = READ_ONCE(tp->linger2);
4048 if (val >= 0)
4049 val = (val ? : READ_ONCE(net->ipv4.sysctl_tcp_fin_timeout)) / HZ;
4050 break;
4051 case TCP_DEFER_ACCEPT:
4052 val = READ_ONCE(icsk->icsk_accept_queue.rskq_defer_accept);
4053 val = retrans_to_secs(val, TCP_TIMEOUT_INIT / HZ,
4054 TCP_RTO_MAX / HZ);
4055 break;
4056 case TCP_WINDOW_CLAMP:
4057 val = tp->window_clamp;
4058 break;
4059 case TCP_INFO: {
4060 struct tcp_info info;
4061
4062 if (copy_from_sockptr(&len, optlen, sizeof(int)))
4063 return -EFAULT;
4064
4065 tcp_get_info(sk, &info);
4066
4067 len = min_t(unsigned int, len, sizeof(info));
4068 if (copy_to_sockptr(optlen, &len, sizeof(int)))
4069 return -EFAULT;
4070 if (copy_to_sockptr(optval, &info, len))
4071 return -EFAULT;
4072 return 0;
4073 }
4074 case TCP_CC_INFO: {
4075 const struct tcp_congestion_ops *ca_ops;
4076 union tcp_cc_info info;
4077 size_t sz = 0;
4078 int attr;
4079
4080 if (copy_from_sockptr(&len, optlen, sizeof(int)))
4081 return -EFAULT;
4082
4083 ca_ops = icsk->icsk_ca_ops;
4084 if (ca_ops && ca_ops->get_info)
4085 sz = ca_ops->get_info(sk, ~0U, &attr, &info);
4086
4087 len = min_t(unsigned int, len, sz);
4088 if (copy_to_sockptr(optlen, &len, sizeof(int)))
4089 return -EFAULT;
4090 if (copy_to_sockptr(optval, &info, len))
4091 return -EFAULT;
4092 return 0;
4093 }
4094 case TCP_QUICKACK:
4095 val = !inet_csk_in_pingpong_mode(sk);
4096 break;
4097
4098 case TCP_CONGESTION:
4099 if (copy_from_sockptr(&len, optlen, sizeof(int)))
4100 return -EFAULT;
4101 len = min_t(unsigned int, len, TCP_CA_NAME_MAX);
4102 if (copy_to_sockptr(optlen, &len, sizeof(int)))
4103 return -EFAULT;
4104 if (copy_to_sockptr(optval, icsk->icsk_ca_ops->name, len))
4105 return -EFAULT;
4106 return 0;
4107
4108 case TCP_ULP:
4109 if (copy_from_sockptr(&len, optlen, sizeof(int)))
4110 return -EFAULT;
4111 len = min_t(unsigned int, len, TCP_ULP_NAME_MAX);
4112 if (!icsk->icsk_ulp_ops) {
4113 len = 0;
4114 if (copy_to_sockptr(optlen, &len, sizeof(int)))
4115 return -EFAULT;
4116 return 0;
4117 }
4118 if (copy_to_sockptr(optlen, &len, sizeof(int)))
4119 return -EFAULT;
4120 if (copy_to_sockptr(optval, icsk->icsk_ulp_ops->name, len))
4121 return -EFAULT;
4122 return 0;
4123
4124 case TCP_FASTOPEN_KEY: {
4125 u64 key[TCP_FASTOPEN_KEY_BUF_LENGTH / sizeof(u64)];
4126 unsigned int key_len;
4127
4128 if (copy_from_sockptr(&len, optlen, sizeof(int)))
4129 return -EFAULT;
4130
4131 key_len = tcp_fastopen_get_cipher(net, icsk, key) *
4132 TCP_FASTOPEN_KEY_LENGTH;
4133 len = min_t(unsigned int, len, key_len);
4134 if (copy_to_sockptr(optlen, &len, sizeof(int)))
4135 return -EFAULT;
4136 if (copy_to_sockptr(optval, key, len))
4137 return -EFAULT;
4138 return 0;
4139 }
4140 case TCP_THIN_LINEAR_TIMEOUTS:
4141 val = tp->thin_lto;
4142 break;
4143
4144 case TCP_THIN_DUPACK:
4145 val = 0;
4146 break;
4147
4148 case TCP_REPAIR:
4149 val = tp->repair;
4150 break;
4151
4152 case TCP_REPAIR_QUEUE:
4153 if (tp->repair)
4154 val = tp->repair_queue;
4155 else
4156 return -EINVAL;
4157 break;
4158
4159 case TCP_REPAIR_WINDOW: {
4160 struct tcp_repair_window opt;
4161
4162 if (copy_from_sockptr(&len, optlen, sizeof(int)))
4163 return -EFAULT;
4164
4165 if (len != sizeof(opt))
4166 return -EINVAL;
4167
4168 if (!tp->repair)
4169 return -EPERM;
4170
4171 opt.snd_wl1 = tp->snd_wl1;
4172 opt.snd_wnd = tp->snd_wnd;
4173 opt.max_window = tp->max_window;
4174 opt.rcv_wnd = tp->rcv_wnd;
4175 opt.rcv_wup = tp->rcv_wup;
4176
4177 if (copy_to_sockptr(optval, &opt, len))
4178 return -EFAULT;
4179 return 0;
4180 }
4181 case TCP_QUEUE_SEQ:
4182 if (tp->repair_queue == TCP_SEND_QUEUE)
4183 val = tp->write_seq;
4184 else if (tp->repair_queue == TCP_RECV_QUEUE)
4185 val = tp->rcv_nxt;
4186 else
4187 return -EINVAL;
4188 break;
4189
4190 case TCP_USER_TIMEOUT:
4191 val = READ_ONCE(icsk->icsk_user_timeout);
4192 break;
4193
4194 case TCP_FASTOPEN:
4195 val = READ_ONCE(icsk->icsk_accept_queue.fastopenq.max_qlen);
4196 break;
4197
4198 case TCP_FASTOPEN_CONNECT:
4199 val = tp->fastopen_connect;
4200 break;
4201
4202 case TCP_FASTOPEN_NO_COOKIE:
4203 val = tp->fastopen_no_cookie;
4204 break;
4205
4206 case TCP_TX_DELAY:
4207 val = READ_ONCE(tp->tcp_tx_delay);
4208 break;
4209
4210 case TCP_TIMESTAMP:
4211 val = tcp_clock_ts(tp->tcp_usec_ts) + READ_ONCE(tp->tsoffset);
4212 if (tp->tcp_usec_ts)
4213 val |= 1;
4214 else
4215 val &= ~1;
4216 break;
4217 case TCP_NOTSENT_LOWAT:
4218 val = READ_ONCE(tp->notsent_lowat);
4219 break;
4220 case TCP_INQ:
4221 val = tp->recvmsg_inq;
4222 break;
4223 case TCP_SAVE_SYN:
4224 val = tp->save_syn;
4225 break;
4226 case TCP_SAVED_SYN: {
4227 if (copy_from_sockptr(&len, optlen, sizeof(int)))
4228 return -EFAULT;
4229
4230 sockopt_lock_sock(sk);
4231 if (tp->saved_syn) {
4232 if (len < tcp_saved_syn_len(tp->saved_syn)) {
4233 len = tcp_saved_syn_len(tp->saved_syn);
4234 if (copy_to_sockptr(optlen, &len, sizeof(int))) {
4235 sockopt_release_sock(sk);
4236 return -EFAULT;
4237 }
4238 sockopt_release_sock(sk);
4239 return -EINVAL;
4240 }
4241 len = tcp_saved_syn_len(tp->saved_syn);
4242 if (copy_to_sockptr(optlen, &len, sizeof(int))) {
4243 sockopt_release_sock(sk);
4244 return -EFAULT;
4245 }
4246 if (copy_to_sockptr(optval, tp->saved_syn->data, len)) {
4247 sockopt_release_sock(sk);
4248 return -EFAULT;
4249 }
4250 tcp_saved_syn_free(tp);
4251 sockopt_release_sock(sk);
4252 } else {
4253 sockopt_release_sock(sk);
4254 len = 0;
4255 if (copy_to_sockptr(optlen, &len, sizeof(int)))
4256 return -EFAULT;
4257 }
4258 return 0;
4259 }
4260#ifdef CONFIG_MMU
4261 case TCP_ZEROCOPY_RECEIVE: {
4262 struct scm_timestamping_internal tss;
4263 struct tcp_zerocopy_receive zc = {};
4264 int err;
4265
4266 if (copy_from_sockptr(&len, optlen, sizeof(int)))
4267 return -EFAULT;
4268 if (len < 0 ||
4269 len < offsetofend(struct tcp_zerocopy_receive, length))
4270 return -EINVAL;
4271 if (unlikely(len > sizeof(zc))) {
4272 err = check_zeroed_sockptr(optval, sizeof(zc),
4273 len - sizeof(zc));
4274 if (err < 1)
4275 return err == 0 ? -EINVAL : err;
4276 len = sizeof(zc);
4277 if (copy_to_sockptr(optlen, &len, sizeof(int)))
4278 return -EFAULT;
4279 }
4280 if (copy_from_sockptr(&zc, optval, len))
4281 return -EFAULT;
4282 if (zc.reserved)
4283 return -EINVAL;
4284 if (zc.msg_flags & ~(TCP_VALID_ZC_MSG_FLAGS))
4285 return -EINVAL;
4286 sockopt_lock_sock(sk);
4287 err = tcp_zerocopy_receive(sk, &zc, &tss);
4288 err = BPF_CGROUP_RUN_PROG_GETSOCKOPT_KERN(sk, level, optname,
4289 &zc, &len, err);
4290 sockopt_release_sock(sk);
4291 if (len >= offsetofend(struct tcp_zerocopy_receive, msg_flags))
4292 goto zerocopy_rcv_cmsg;
4293 switch (len) {
4294 case offsetofend(struct tcp_zerocopy_receive, msg_flags):
4295 goto zerocopy_rcv_cmsg;
4296 case offsetofend(struct tcp_zerocopy_receive, msg_controllen):
4297 case offsetofend(struct tcp_zerocopy_receive, msg_control):
4298 case offsetofend(struct tcp_zerocopy_receive, flags):
4299 case offsetofend(struct tcp_zerocopy_receive, copybuf_len):
4300 case offsetofend(struct tcp_zerocopy_receive, copybuf_address):
4301 case offsetofend(struct tcp_zerocopy_receive, err):
4302 goto zerocopy_rcv_sk_err;
4303 case offsetofend(struct tcp_zerocopy_receive, inq):
4304 goto zerocopy_rcv_inq;
4305 case offsetofend(struct tcp_zerocopy_receive, length):
4306 default:
4307 goto zerocopy_rcv_out;
4308 }
4309zerocopy_rcv_cmsg:
4310 if (zc.msg_flags & TCP_CMSG_TS)
4311 tcp_zc_finalize_rx_tstamp(sk, &zc, &tss);
4312 else
4313 zc.msg_flags = 0;
4314zerocopy_rcv_sk_err:
4315 if (!err)
4316 zc.err = sock_error(sk);
4317zerocopy_rcv_inq:
4318 zc.inq = tcp_inq_hint(sk);
4319zerocopy_rcv_out:
4320 if (!err && copy_to_sockptr(optval, &zc, len))
4321 err = -EFAULT;
4322 return err;
4323 }
4324#endif
4325 case TCP_AO_REPAIR:
4326 if (!tcp_can_repair_sock(sk))
4327 return -EPERM;
4328 return tcp_ao_get_repair(sk, optval, optlen);
4329 case TCP_AO_GET_KEYS:
4330 case TCP_AO_INFO: {
4331 int err;
4332
4333 sockopt_lock_sock(sk);
4334 if (optname == TCP_AO_GET_KEYS)
4335 err = tcp_ao_get_mkts(sk, optval, optlen);
4336 else
4337 err = tcp_ao_get_sock_info(sk, optval, optlen);
4338 sockopt_release_sock(sk);
4339
4340 return err;
4341 }
4342 default:
4343 return -ENOPROTOOPT;
4344 }
4345
4346 if (copy_to_sockptr(optlen, &len, sizeof(int)))
4347 return -EFAULT;
4348 if (copy_to_sockptr(optval, &val, len))
4349 return -EFAULT;
4350 return 0;
4351}
4352
4353bool tcp_bpf_bypass_getsockopt(int level, int optname)
4354{
4355 /* TCP do_tcp_getsockopt has optimized getsockopt implementation
4356 * to avoid extra socket lock for TCP_ZEROCOPY_RECEIVE.
4357 */
4358 if (level == SOL_TCP && optname == TCP_ZEROCOPY_RECEIVE)
4359 return true;
4360
4361 return false;
4362}
4363EXPORT_SYMBOL(tcp_bpf_bypass_getsockopt);
4364
4365int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval,
4366 int __user *optlen)
4367{
4368 struct inet_connection_sock *icsk = inet_csk(sk);
4369
4370 if (level != SOL_TCP)
4371 /* Paired with WRITE_ONCE() in do_ipv6_setsockopt() and tcp_v6_connect() */
4372 return READ_ONCE(icsk->icsk_af_ops)->getsockopt(sk, level, optname,
4373 optval, optlen);
4374 return do_tcp_getsockopt(sk, level, optname, USER_SOCKPTR(optval),
4375 USER_SOCKPTR(optlen));
4376}
4377EXPORT_SYMBOL(tcp_getsockopt);
4378
4379#ifdef CONFIG_TCP_MD5SIG
4380int tcp_md5_sigpool_id = -1;
4381EXPORT_SYMBOL_GPL(tcp_md5_sigpool_id);
4382
4383int tcp_md5_alloc_sigpool(void)
4384{
4385 size_t scratch_size;
4386 int ret;
4387
4388 scratch_size = sizeof(union tcp_md5sum_block) + sizeof(struct tcphdr);
4389 ret = tcp_sigpool_alloc_ahash("md5", scratch_size);
4390 if (ret >= 0) {
4391 /* As long as any md5 sigpool was allocated, the return
4392 * id would stay the same. Re-write the id only for the case
4393 * when previously all MD5 keys were deleted and this call
4394 * allocates the first MD5 key, which may return a different
4395 * sigpool id than was used previously.
4396 */
4397 WRITE_ONCE(tcp_md5_sigpool_id, ret); /* Avoids the compiler potentially being smart here */
4398 return 0;
4399 }
4400 return ret;
4401}
4402
4403void tcp_md5_release_sigpool(void)
4404{
4405 tcp_sigpool_release(READ_ONCE(tcp_md5_sigpool_id));
4406}
4407
4408void tcp_md5_add_sigpool(void)
4409{
4410 tcp_sigpool_get(READ_ONCE(tcp_md5_sigpool_id));
4411}
4412
4413int tcp_md5_hash_key(struct tcp_sigpool *hp,
4414 const struct tcp_md5sig_key *key)
4415{
4416 u8 keylen = READ_ONCE(key->keylen); /* paired with WRITE_ONCE() in tcp_md5_do_add */
4417 struct scatterlist sg;
4418
4419 sg_init_one(&sg, key->key, keylen);
4420 ahash_request_set_crypt(hp->req, &sg, NULL, keylen);
4421
4422 /* We use data_race() because tcp_md5_do_add() might change
4423 * key->key under us
4424 */
4425 return data_race(crypto_ahash_update(hp->req));
4426}
4427EXPORT_SYMBOL(tcp_md5_hash_key);
4428
4429/* Called with rcu_read_lock() */
4430enum skb_drop_reason
4431tcp_inbound_md5_hash(const struct sock *sk, const struct sk_buff *skb,
4432 const void *saddr, const void *daddr,
4433 int family, int l3index, const __u8 *hash_location)
4434{
4435 /* This gets called for each TCP segment that has TCP-MD5 option.
4436 * We have 3 drop cases:
4437 * o No MD5 hash and one expected.
4438 * o MD5 hash and we're not expecting one.
4439 * o MD5 hash and its wrong.
4440 */
4441 const struct tcp_sock *tp = tcp_sk(sk);
4442 struct tcp_md5sig_key *key;
4443 u8 newhash[16];
4444 int genhash;
4445
4446 key = tcp_md5_do_lookup(sk, l3index, saddr, family);
4447
4448 if (!key && hash_location) {
4449 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMD5UNEXPECTED);
4450 tcp_hash_fail("Unexpected MD5 Hash found", family, skb, "");
4451 return SKB_DROP_REASON_TCP_MD5UNEXPECTED;
4452 }
4453
4454 /* Check the signature.
4455 * To support dual stack listeners, we need to handle
4456 * IPv4-mapped case.
4457 */
4458 if (family == AF_INET)
4459 genhash = tcp_v4_md5_hash_skb(newhash, key, NULL, skb);
4460 else
4461 genhash = tp->af_specific->calc_md5_hash(newhash, key,
4462 NULL, skb);
4463 if (genhash || memcmp(hash_location, newhash, 16) != 0) {
4464 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMD5FAILURE);
4465 if (family == AF_INET) {
4466 tcp_hash_fail("MD5 Hash failed", AF_INET, skb, "%s L3 index %d",
4467 genhash ? "tcp_v4_calc_md5_hash failed"
4468 : "", l3index);
4469 } else {
4470 if (genhash) {
4471 tcp_hash_fail("MD5 Hash failed",
4472 AF_INET6, skb, "L3 index %d",
4473 l3index);
4474 } else {
4475 tcp_hash_fail("MD5 Hash mismatch",
4476 AF_INET6, skb, "L3 index %d",
4477 l3index);
4478 }
4479 }
4480 return SKB_DROP_REASON_TCP_MD5FAILURE;
4481 }
4482 return SKB_NOT_DROPPED_YET;
4483}
4484EXPORT_SYMBOL(tcp_inbound_md5_hash);
4485
4486#endif
4487
4488void tcp_done(struct sock *sk)
4489{
4490 struct request_sock *req;
4491
4492 /* We might be called with a new socket, after
4493 * inet_csk_prepare_forced_close() has been called
4494 * so we can not use lockdep_sock_is_held(sk)
4495 */
4496 req = rcu_dereference_protected(tcp_sk(sk)->fastopen_rsk, 1);
4497
4498 if (sk->sk_state == TCP_SYN_SENT || sk->sk_state == TCP_SYN_RECV)
4499 TCP_INC_STATS(sock_net(sk), TCP_MIB_ATTEMPTFAILS);
4500
4501 tcp_set_state(sk, TCP_CLOSE);
4502 tcp_clear_xmit_timers(sk);
4503 if (req)
4504 reqsk_fastopen_remove(sk, req, false);
4505
4506 WRITE_ONCE(sk->sk_shutdown, SHUTDOWN_MASK);
4507
4508 if (!sock_flag(sk, SOCK_DEAD))
4509 sk->sk_state_change(sk);
4510 else
4511 inet_csk_destroy_sock(sk);
4512}
4513EXPORT_SYMBOL_GPL(tcp_done);
4514
4515int tcp_abort(struct sock *sk, int err)
4516{
4517 int state = inet_sk_state_load(sk);
4518
4519 if (state == TCP_NEW_SYN_RECV) {
4520 struct request_sock *req = inet_reqsk(sk);
4521
4522 local_bh_disable();
4523 inet_csk_reqsk_queue_drop(req->rsk_listener, req);
4524 local_bh_enable();
4525 return 0;
4526 }
4527 if (state == TCP_TIME_WAIT) {
4528 struct inet_timewait_sock *tw = inet_twsk(sk);
4529
4530 refcount_inc(&tw->tw_refcnt);
4531 local_bh_disable();
4532 inet_twsk_deschedule_put(tw);
4533 local_bh_enable();
4534 return 0;
4535 }
4536
4537 /* BPF context ensures sock locking. */
4538 if (!has_current_bpf_ctx())
4539 /* Don't race with userspace socket closes such as tcp_close. */
4540 lock_sock(sk);
4541
4542 if (sk->sk_state == TCP_LISTEN) {
4543 tcp_set_state(sk, TCP_CLOSE);
4544 inet_csk_listen_stop(sk);
4545 }
4546
4547 /* Don't race with BH socket closes such as inet_csk_listen_stop. */
4548 local_bh_disable();
4549 bh_lock_sock(sk);
4550
4551 if (!sock_flag(sk, SOCK_DEAD)) {
4552 WRITE_ONCE(sk->sk_err, err);
4553 /* This barrier is coupled with smp_rmb() in tcp_poll() */
4554 smp_wmb();
4555 sk_error_report(sk);
4556 if (tcp_need_reset(sk->sk_state))
4557 tcp_send_active_reset(sk, GFP_ATOMIC);
4558 tcp_done(sk);
4559 }
4560
4561 bh_unlock_sock(sk);
4562 local_bh_enable();
4563 tcp_write_queue_purge(sk);
4564 if (!has_current_bpf_ctx())
4565 release_sock(sk);
4566 return 0;
4567}
4568EXPORT_SYMBOL_GPL(tcp_abort);
4569
4570extern struct tcp_congestion_ops tcp_reno;
4571
4572static __initdata unsigned long thash_entries;
4573static int __init set_thash_entries(char *str)
4574{
4575 ssize_t ret;
4576
4577 if (!str)
4578 return 0;
4579
4580 ret = kstrtoul(str, 0, &thash_entries);
4581 if (ret)
4582 return 0;
4583
4584 return 1;
4585}
4586__setup("thash_entries=", set_thash_entries);
4587
4588static void __init tcp_init_mem(void)
4589{
4590 unsigned long limit = nr_free_buffer_pages() / 16;
4591
4592 limit = max(limit, 128UL);
4593 sysctl_tcp_mem[0] = limit / 4 * 3; /* 4.68 % */
4594 sysctl_tcp_mem[1] = limit; /* 6.25 % */
4595 sysctl_tcp_mem[2] = sysctl_tcp_mem[0] * 2; /* 9.37 % */
4596}
4597
4598static void __init tcp_struct_check(void)
4599{
4600 /* TX read-mostly hotpath cache lines */
4601 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_tx, max_window);
4602 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_tx, rcv_ssthresh);
4603 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_tx, reordering);
4604 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_tx, notsent_lowat);
4605 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_tx, gso_segs);
4606 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_tx, lost_skb_hint);
4607 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_tx, retransmit_skb_hint);
4608 CACHELINE_ASSERT_GROUP_SIZE(struct tcp_sock, tcp_sock_read_tx, 40);
4609
4610 /* TXRX read-mostly hotpath cache lines */
4611 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_txrx, tsoffset);
4612 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_txrx, snd_wnd);
4613 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_txrx, mss_cache);
4614 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_txrx, snd_cwnd);
4615 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_txrx, prr_out);
4616 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_txrx, lost_out);
4617 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_txrx, sacked_out);
4618 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_txrx, scaling_ratio);
4619 CACHELINE_ASSERT_GROUP_SIZE(struct tcp_sock, tcp_sock_read_txrx, 32);
4620
4621 /* RX read-mostly hotpath cache lines */
4622 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_rx, copied_seq);
4623 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_rx, rcv_tstamp);
4624 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_rx, snd_wl1);
4625 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_rx, tlp_high_seq);
4626 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_rx, rttvar_us);
4627 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_rx, retrans_out);
4628 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_rx, advmss);
4629 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_rx, urg_data);
4630 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_rx, lost);
4631 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_rx, rtt_min);
4632 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_rx, out_of_order_queue);
4633 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_rx, snd_ssthresh);
4634 CACHELINE_ASSERT_GROUP_SIZE(struct tcp_sock, tcp_sock_read_rx, 69);
4635
4636 /* TX read-write hotpath cache lines */
4637 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, segs_out);
4638 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, data_segs_out);
4639 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, bytes_sent);
4640 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, snd_sml);
4641 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, chrono_start);
4642 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, chrono_stat);
4643 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, write_seq);
4644 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, pushed_seq);
4645 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, lsndtime);
4646 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, mdev_us);
4647 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, tcp_wstamp_ns);
4648 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, tcp_clock_cache);
4649 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, tcp_mstamp);
4650 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, rtt_seq);
4651 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, tsorted_sent_queue);
4652 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, highest_sack);
4653 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, ecn_flags);
4654 CACHELINE_ASSERT_GROUP_SIZE(struct tcp_sock, tcp_sock_write_tx, 113);
4655
4656 /* TXRX read-write hotpath cache lines */
4657 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_txrx, pred_flags);
4658 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_txrx, rcv_nxt);
4659 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_txrx, snd_nxt);
4660 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_txrx, snd_una);
4661 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_txrx, window_clamp);
4662 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_txrx, srtt_us);
4663 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_txrx, packets_out);
4664 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_txrx, snd_up);
4665 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_txrx, delivered);
4666 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_txrx, delivered_ce);
4667 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_txrx, app_limited);
4668 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_txrx, rcv_wnd);
4669 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_txrx, rx_opt);
4670 CACHELINE_ASSERT_GROUP_SIZE(struct tcp_sock, tcp_sock_write_txrx, 76);
4671
4672 /* RX read-write hotpath cache lines */
4673 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_rx, bytes_received);
4674 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_rx, segs_in);
4675 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_rx, data_segs_in);
4676 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_rx, rcv_wup);
4677 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_rx, max_packets_out);
4678 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_rx, cwnd_usage_seq);
4679 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_rx, rate_delivered);
4680 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_rx, rate_interval_us);
4681 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_rx, rcv_rtt_last_tsecr);
4682 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_rx, first_tx_mstamp);
4683 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_rx, delivered_mstamp);
4684 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_rx, bytes_acked);
4685 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_rx, rcv_rtt_est);
4686 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_rx, rcvq_space);
4687 CACHELINE_ASSERT_GROUP_SIZE(struct tcp_sock, tcp_sock_write_rx, 99);
4688}
4689
4690void __init tcp_init(void)
4691{
4692 int max_rshare, max_wshare, cnt;
4693 unsigned long limit;
4694 unsigned int i;
4695
4696 BUILD_BUG_ON(TCP_MIN_SND_MSS <= MAX_TCP_OPTION_SPACE);
4697 BUILD_BUG_ON(sizeof(struct tcp_skb_cb) >
4698 sizeof_field(struct sk_buff, cb));
4699
4700 tcp_struct_check();
4701
4702 percpu_counter_init(&tcp_sockets_allocated, 0, GFP_KERNEL);
4703
4704 timer_setup(&tcp_orphan_timer, tcp_orphan_update, TIMER_DEFERRABLE);
4705 mod_timer(&tcp_orphan_timer, jiffies + TCP_ORPHAN_TIMER_PERIOD);
4706
4707 inet_hashinfo2_init(&tcp_hashinfo, "tcp_listen_portaddr_hash",
4708 thash_entries, 21, /* one slot per 2 MB*/
4709 0, 64 * 1024);
4710 tcp_hashinfo.bind_bucket_cachep =
4711 kmem_cache_create("tcp_bind_bucket",
4712 sizeof(struct inet_bind_bucket), 0,
4713 SLAB_HWCACHE_ALIGN | SLAB_PANIC |
4714 SLAB_ACCOUNT,
4715 NULL);
4716 tcp_hashinfo.bind2_bucket_cachep =
4717 kmem_cache_create("tcp_bind2_bucket",
4718 sizeof(struct inet_bind2_bucket), 0,
4719 SLAB_HWCACHE_ALIGN | SLAB_PANIC |
4720 SLAB_ACCOUNT,
4721 NULL);
4722
4723 /* Size and allocate the main established and bind bucket
4724 * hash tables.
4725 *
4726 * The methodology is similar to that of the buffer cache.
4727 */
4728 tcp_hashinfo.ehash =
4729 alloc_large_system_hash("TCP established",
4730 sizeof(struct inet_ehash_bucket),
4731 thash_entries,
4732 17, /* one slot per 128 KB of memory */
4733 0,
4734 NULL,
4735 &tcp_hashinfo.ehash_mask,
4736 0,
4737 thash_entries ? 0 : 512 * 1024);
4738 for (i = 0; i <= tcp_hashinfo.ehash_mask; i++)
4739 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].chain, i);
4740
4741 if (inet_ehash_locks_alloc(&tcp_hashinfo))
4742 panic("TCP: failed to alloc ehash_locks");
4743 tcp_hashinfo.bhash =
4744 alloc_large_system_hash("TCP bind",
4745 2 * sizeof(struct inet_bind_hashbucket),
4746 tcp_hashinfo.ehash_mask + 1,
4747 17, /* one slot per 128 KB of memory */
4748 0,
4749 &tcp_hashinfo.bhash_size,
4750 NULL,
4751 0,
4752 64 * 1024);
4753 tcp_hashinfo.bhash_size = 1U << tcp_hashinfo.bhash_size;
4754 tcp_hashinfo.bhash2 = tcp_hashinfo.bhash + tcp_hashinfo.bhash_size;
4755 for (i = 0; i < tcp_hashinfo.bhash_size; i++) {
4756 spin_lock_init(&tcp_hashinfo.bhash[i].lock);
4757 INIT_HLIST_HEAD(&tcp_hashinfo.bhash[i].chain);
4758 spin_lock_init(&tcp_hashinfo.bhash2[i].lock);
4759 INIT_HLIST_HEAD(&tcp_hashinfo.bhash2[i].chain);
4760 }
4761
4762 tcp_hashinfo.pernet = false;
4763
4764 cnt = tcp_hashinfo.ehash_mask + 1;
4765 sysctl_tcp_max_orphans = cnt / 2;
4766
4767 tcp_init_mem();
4768 /* Set per-socket limits to no more than 1/128 the pressure threshold */
4769 limit = nr_free_buffer_pages() << (PAGE_SHIFT - 7);
4770 max_wshare = min(4UL*1024*1024, limit);
4771 max_rshare = min(6UL*1024*1024, limit);
4772
4773 init_net.ipv4.sysctl_tcp_wmem[0] = PAGE_SIZE;
4774 init_net.ipv4.sysctl_tcp_wmem[1] = 16*1024;
4775 init_net.ipv4.sysctl_tcp_wmem[2] = max(64*1024, max_wshare);
4776
4777 init_net.ipv4.sysctl_tcp_rmem[0] = PAGE_SIZE;
4778 init_net.ipv4.sysctl_tcp_rmem[1] = 131072;
4779 init_net.ipv4.sysctl_tcp_rmem[2] = max(131072, max_rshare);
4780
4781 pr_info("Hash tables configured (established %u bind %u)\n",
4782 tcp_hashinfo.ehash_mask + 1, tcp_hashinfo.bhash_size);
4783
4784 tcp_v4_init();
4785 tcp_metrics_init();
4786 BUG_ON(tcp_register_congestion_control(&tcp_reno) != 0);
4787 tcp_tasklet_init();
4788 mptcp_init();
4789}