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