1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /*
3 * INET An implementation of the TCP/IP protocol suite for the LINUX
4 * operating system. INET is implemented using the BSD Socket
5 * interface as the means of communication with the user level.
6 *
7 * Implementation of the Transmission Control Protocol(TCP).
8 *
9 * Authors: Ross Biro
10 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
11 * Mark Evans, <evansmp@uhura.aston.ac.uk>
12 * Corey Minyard <wf-rch!minyard@relay.EU.net>
13 * Florian La Roche, <flla@stud.uni-sb.de>
14 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
15 * Linus Torvalds, <torvalds@cs.helsinki.fi>
16 * Alan Cox, <gw4pts@gw4pts.ampr.org>
17 * Matthew Dillon, <dillon@apollo.west.oic.com>
18 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
19 * Jorge Cwik, <jorge@laser.satlink.net>
20 *
21 * Fixes:
22 * Alan Cox : Numerous verify_area() calls
23 * Alan Cox : Set the ACK bit on a reset
24 * Alan Cox : Stopped it crashing if it closed while
25 * sk->inuse=1 and was trying to connect
26 * (tcp_err()).
27 * Alan Cox : All icmp error handling was broken
28 * pointers passed where wrong and the
29 * socket was looked up backwards. Nobody
30 * tested any icmp error code obviously.
31 * Alan Cox : tcp_err() now handled properly. It
32 * wakes people on errors. poll
33 * behaves and the icmp error race
34 * has gone by moving it into sock.c
35 * Alan Cox : tcp_send_reset() fixed to work for
36 * everything not just packets for
37 * unknown sockets.
38 * Alan Cox : tcp option processing.
39 * Alan Cox : Reset tweaked (still not 100%) [Had
40 * syn rule wrong]
41 * Herp Rosmanith : More reset fixes
42 * Alan Cox : No longer acks invalid rst frames.
43 * Acking any kind of RST is right out.
44 * Alan Cox : Sets an ignore me flag on an rst
45 * receive otherwise odd bits of prattle
46 * escape still
47 * Alan Cox : Fixed another acking RST frame bug.
48 * Should stop LAN workplace lockups.
49 * Alan Cox : Some tidyups using the new skb list
50 * facilities
51 * Alan Cox : sk->keepopen now seems to work
52 * Alan Cox : Pulls options out correctly on accepts
53 * Alan Cox : Fixed assorted sk->rqueue->next errors
54 * Alan Cox : PSH doesn't end a TCP read. Switched a
55 * bit to skb ops.
56 * Alan Cox : Tidied tcp_data to avoid a potential
57 * nasty.
58 * Alan Cox : Added some better commenting, as the
59 * tcp is hard to follow
60 * Alan Cox : Removed incorrect check for 20 * psh
61 * Michael O'Reilly : ack < copied bug fix.
62 * Johannes Stille : Misc tcp fixes (not all in yet).
63 * Alan Cox : FIN with no memory -> CRASH
64 * Alan Cox : Added socket option proto entries.
65 * Also added awareness of them to accept.
66 * Alan Cox : Added TCP options (SOL_TCP)
67 * Alan Cox : Switched wakeup calls to callbacks,
68 * so the kernel can layer network
69 * sockets.
70 * Alan Cox : Use ip_tos/ip_ttl settings.
71 * Alan Cox : Handle FIN (more) properly (we hope).
72 * Alan Cox : RST frames sent on unsynchronised
73 * state ack error.
74 * Alan Cox : Put in missing check for SYN bit.
75 * Alan Cox : Added tcp_select_window() aka NET2E
76 * window non shrink trick.
77 * Alan Cox : Added a couple of small NET2E timer
78 * fixes
79 * Charles Hedrick : TCP fixes
80 * Toomas Tamm : TCP window fixes
81 * Alan Cox : Small URG fix to rlogin ^C ack fight
82 * Charles Hedrick : Rewrote most of it to actually work
83 * Linus : Rewrote tcp_read() and URG handling
84 * completely
85 * Gerhard Koerting: Fixed some missing timer handling
86 * Matthew Dillon : Reworked TCP machine states as per RFC
87 * Gerhard Koerting: PC/TCP workarounds
88 * Adam Caldwell : Assorted timer/timing errors
89 * Matthew Dillon : Fixed another RST bug
90 * Alan Cox : Move to kernel side addressing changes.
91 * Alan Cox : Beginning work on TCP fastpathing
92 * (not yet usable)
93 * Arnt Gulbrandsen: Turbocharged tcp_check() routine.
94 * Alan Cox : TCP fast path debugging
95 * Alan Cox : Window clamping
96 * Michael Riepe : Bug in tcp_check()
97 * Matt Dillon : More TCP improvements and RST bug fixes
98 * Matt Dillon : Yet more small nasties remove from the
99 * TCP code (Be very nice to this man if
100 * tcp finally works 100%) 8)
101 * Alan Cox : BSD accept semantics.
102 * Alan Cox : Reset on closedown bug.
103 * Peter De Schrijver : ENOTCONN check missing in tcp_sendto().
104 * Michael Pall : Handle poll() after URG properly in
105 * all cases.
106 * Michael Pall : Undo the last fix in tcp_read_urg()
107 * (multi URG PUSH broke rlogin).
108 * Michael Pall : Fix the multi URG PUSH problem in
109 * tcp_readable(), poll() after URG
110 * works now.
111 * Michael Pall : recv(...,MSG_OOB) never blocks in the
112 * BSD api.
113 * Alan Cox : Changed the semantics of sk->socket to
114 * fix a race and a signal problem with
115 * accept() and async I/O.
116 * Alan Cox : Relaxed the rules on tcp_sendto().
117 * Yury Shevchuk : Really fixed accept() blocking problem.
118 * Craig I. Hagan : Allow for BSD compatible TIME_WAIT for
119 * clients/servers which listen in on
120 * fixed ports.
121 * Alan Cox : Cleaned the above up and shrank it to
122 * a sensible code size.
123 * Alan Cox : Self connect lockup fix.
124 * Alan Cox : No connect to multicast.
125 * Ross Biro : Close unaccepted children on master
126 * socket close.
127 * Alan Cox : Reset tracing code.
128 * Alan Cox : Spurious resets on shutdown.
129 * Alan Cox : Giant 15 minute/60 second timer error
130 * Alan Cox : Small whoops in polling before an
131 * accept.
132 * Alan Cox : Kept the state trace facility since
133 * it's handy for debugging.
134 * Alan Cox : More reset handler fixes.
135 * Alan Cox : Started rewriting the code based on
136 * the RFC's for other useful protocol
137 * references see: Comer, KA9Q NOS, and
138 * for a reference on the difference
139 * between specifications and how BSD
140 * works see the 4.4lite source.
141 * A.N.Kuznetsov : Don't time wait on completion of tidy
142 * close.
143 * Linus Torvalds : Fin/Shutdown & copied_seq changes.
144 * Linus Torvalds : Fixed BSD port reuse to work first syn
145 * Alan Cox : Reimplemented timers as per the RFC
146 * and using multiple timers for sanity.
147 * Alan Cox : Small bug fixes, and a lot of new
148 * comments.
149 * Alan Cox : Fixed dual reader crash by locking
150 * the buffers (much like datagram.c)
151 * Alan Cox : Fixed stuck sockets in probe. A probe
152 * now gets fed up of retrying without
153 * (even a no space) answer.
154 * Alan Cox : Extracted closing code better
155 * Alan Cox : Fixed the closing state machine to
156 * resemble the RFC.
157 * Alan Cox : More 'per spec' fixes.
158 * Jorge Cwik : Even faster checksumming.
159 * Alan Cox : tcp_data() doesn't ack illegal PSH
160 * only frames. At least one pc tcp stack
161 * generates them.
162 * Alan Cox : Cache last socket.
163 * Alan Cox : Per route irtt.
164 * Matt Day : poll()->select() match BSD precisely on error
165 * Alan Cox : New buffers
166 * Marc Tamsky : Various sk->prot->retransmits and
167 * sk->retransmits misupdating fixed.
168 * Fixed tcp_write_timeout: stuck close,
169 * and TCP syn retries gets used now.
170 * Mark Yarvis : In tcp_read_wakeup(), don't send an
171 * ack if state is TCP_CLOSED.
172 * Alan Cox : Look up device on a retransmit - routes may
173 * change. Doesn't yet cope with MSS shrink right
174 * but it's a start!
175 * Marc Tamsky : Closing in closing fixes.
176 * Mike Shaver : RFC1122 verifications.
177 * Alan Cox : rcv_saddr errors.
178 * Alan Cox : Block double connect().
179 * Alan Cox : Small hooks for enSKIP.
180 * Alexey Kuznetsov: Path MTU discovery.
181 * Alan Cox : Support soft errors.
182 * Alan Cox : Fix MTU discovery pathological case
183 * when the remote claims no mtu!
184 * Marc Tamsky : TCP_CLOSE fix.
185 * Colin (G3TNE) : Send a reset on syn ack replies in
186 * window but wrong (fixes NT lpd problems)
187 * Pedro Roque : Better TCP window handling, delayed ack.
188 * Joerg Reuter : No modification of locked buffers in
189 * tcp_do_retransmit()
190 * Eric Schenk : Changed receiver side silly window
191 * avoidance algorithm to BSD style
192 * algorithm. This doubles throughput
193 * against machines running Solaris,
194 * and seems to result in general
195 * improvement.
196 * Stefan Magdalinski : adjusted tcp_readable() to fix FIONREAD
197 * Willy Konynenberg : Transparent proxying support.
198 * Mike McLagan : Routing by source
199 * Keith Owens : Do proper merging with partial SKB's in
200 * tcp_do_sendmsg to avoid burstiness.
201 * Eric Schenk : Fix fast close down bug with
202 * shutdown() followed by close().
203 * Andi Kleen : Make poll agree with SIGIO
204 * Salvatore Sanfilippo : Support SO_LINGER with linger == 1 and
205 * lingertime == 0 (RFC 793 ABORT Call)
206 * Hirokazu Takahashi : Use copy_from_user() instead of
207 * csum_and_copy_from_user() if possible.
208 *
209 * Description of States:
210 *
211 * TCP_SYN_SENT sent a connection request, waiting for ack
212 *
213 * TCP_SYN_RECV received a connection request, sent ack,
214 * waiting for final ack in three-way handshake.
215 *
216 * TCP_ESTABLISHED connection established
217 *
218 * TCP_FIN_WAIT1 our side has shutdown, waiting to complete
219 * transmission of remaining buffered data
220 *
221 * TCP_FIN_WAIT2 all buffered data sent, waiting for remote
222 * to shutdown
223 *
224 * TCP_CLOSING both sides have shutdown but we still have
225 * data we have to finish sending
226 *
227 * TCP_TIME_WAIT timeout to catch resent junk before entering
228 * closed, can only be entered from FIN_WAIT2
229 * or CLOSING. Required because the other end
230 * may not have gotten our last ACK causing it
231 * to retransmit the data packet (which we ignore)
232 *
233 * TCP_CLOSE_WAIT remote side has shutdown and is waiting for
234 * us to finish writing our data and to shutdown
235 * (we have to close() to move on to LAST_ACK)
236 *
237 * TCP_LAST_ACK out side has shutdown after remote has
238 * shutdown. There may still be data in our
239 * buffer that we have to finish sending
240 *
241 * TCP_CLOSE socket is finished
242 */
243
244 #define pr_fmt(fmt) "TCP: " fmt
245
246 #include <crypto/hash.h>
247 #include <linux/kernel.h>
248 #include <linux/module.h>
249 #include <linux/types.h>
250 #include <linux/fcntl.h>
251 #include <linux/poll.h>
252 #include <linux/inet_diag.h>
253 #include <linux/init.h>
254 #include <linux/fs.h>
255 #include <linux/skbuff.h>
256 #include <linux/scatterlist.h>
257 #include <linux/splice.h>
258 #include <linux/net.h>
259 #include <linux/socket.h>
260 #include <linux/random.h>
261 #include <linux/memblock.h>
262 #include <linux/highmem.h>
263 #include <linux/swap.h>
264 #include <linux/cache.h>
265 #include <linux/err.h>
266 #include <linux/time.h>
267 #include <linux/slab.h>
268 #include <linux/errqueue.h>
269 #include <linux/static_key.h>
270 #include <linux/btf.h>
271
272 #include <net/icmp.h>
273 #include <net/inet_common.h>
274 #include <net/tcp.h>
275 #include <net/mptcp.h>
276 #include <net/xfrm.h>
277 #include <net/ip.h>
278 #include <net/sock.h>
279
280 #include <linux/uaccess.h>
281 #include <asm/ioctls.h>
282 #include <net/busy_poll.h>
283
284 /* Track pending CMSGs. */
285 enum {
286 TCP_CMSG_INQ = 1,
287 TCP_CMSG_TS = 2
288 };
289
290 struct percpu_counter tcp_orphan_count;
291 EXPORT_SYMBOL_GPL(tcp_orphan_count);
292
293 long sysctl_tcp_mem[3] __read_mostly;
294 EXPORT_SYMBOL(sysctl_tcp_mem);
295
296 atomic_long_t tcp_memory_allocated; /* Current allocated memory. */
297 EXPORT_SYMBOL(tcp_memory_allocated);
298
299 #if IS_ENABLED(CONFIG_SMC)
300 DEFINE_STATIC_KEY_FALSE(tcp_have_smc);
301 EXPORT_SYMBOL(tcp_have_smc);
302 #endif
303
304 /*
305 * Current number of TCP sockets.
306 */
307 struct percpu_counter tcp_sockets_allocated;
308 EXPORT_SYMBOL(tcp_sockets_allocated);
309
310 /*
311 * TCP splice context
312 */
313 struct tcp_splice_state {
314 struct pipe_inode_info *pipe;
315 size_t len;
316 unsigned int flags;
317 };
318
319 /*
320 * Pressure flag: try to collapse.
321 * Technical note: it is used by multiple contexts non atomically.
322 * All the __sk_mem_schedule() is of this nature: accounting
323 * is strict, actions are advisory and have some latency.
324 */
325 unsigned long tcp_memory_pressure __read_mostly;
326 EXPORT_SYMBOL_GPL(tcp_memory_pressure);
327
328 DEFINE_STATIC_KEY_FALSE(tcp_rx_skb_cache_key);
329 EXPORT_SYMBOL(tcp_rx_skb_cache_key);
330
331 DEFINE_STATIC_KEY_FALSE(tcp_tx_skb_cache_key);
332
tcp_enter_memory_pressure(struct sock * sk)333 void tcp_enter_memory_pressure(struct sock *sk)
334 {
335 unsigned long val;
336
337 if (READ_ONCE(tcp_memory_pressure))
338 return;
339 val = jiffies;
340
341 if (!val)
342 val--;
343 if (!cmpxchg(&tcp_memory_pressure, 0, val))
344 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMEMORYPRESSURES);
345 }
346 EXPORT_SYMBOL_GPL(tcp_enter_memory_pressure);
347
tcp_leave_memory_pressure(struct sock * sk)348 void tcp_leave_memory_pressure(struct sock *sk)
349 {
350 unsigned long val;
351
352 if (!READ_ONCE(tcp_memory_pressure))
353 return;
354 val = xchg(&tcp_memory_pressure, 0);
355 if (val)
356 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPMEMORYPRESSURESCHRONO,
357 jiffies_to_msecs(jiffies - val));
358 }
359 EXPORT_SYMBOL_GPL(tcp_leave_memory_pressure);
360
361 /* Convert seconds to retransmits based on initial and max timeout */
secs_to_retrans(int seconds,int timeout,int rto_max)362 static u8 secs_to_retrans(int seconds, int timeout, int rto_max)
363 {
364 u8 res = 0;
365
366 if (seconds > 0) {
367 int period = timeout;
368
369 res = 1;
370 while (seconds > period && res < 255) {
371 res++;
372 timeout <<= 1;
373 if (timeout > rto_max)
374 timeout = rto_max;
375 period += timeout;
376 }
377 }
378 return res;
379 }
380
381 /* Convert retransmits to seconds based on initial and max timeout */
retrans_to_secs(u8 retrans,int timeout,int rto_max)382 static int retrans_to_secs(u8 retrans, int timeout, int rto_max)
383 {
384 int period = 0;
385
386 if (retrans > 0) {
387 period = timeout;
388 while (--retrans) {
389 timeout <<= 1;
390 if (timeout > rto_max)
391 timeout = rto_max;
392 period += timeout;
393 }
394 }
395 return period;
396 }
397
tcp_compute_delivery_rate(const struct tcp_sock * tp)398 static u64 tcp_compute_delivery_rate(const struct tcp_sock *tp)
399 {
400 u32 rate = READ_ONCE(tp->rate_delivered);
401 u32 intv = READ_ONCE(tp->rate_interval_us);
402 u64 rate64 = 0;
403
404 if (rate && intv) {
405 rate64 = (u64)rate * tp->mss_cache * USEC_PER_SEC;
406 do_div(rate64, intv);
407 }
408 return rate64;
409 }
410
411 /* Address-family independent initialization for a tcp_sock.
412 *
413 * NOTE: A lot of things set to zero explicitly by call to
414 * sk_alloc() so need not be done here.
415 */
tcp_init_sock(struct sock * sk)416 void tcp_init_sock(struct sock *sk)
417 {
418 struct inet_connection_sock *icsk = inet_csk(sk);
419 struct tcp_sock *tp = tcp_sk(sk);
420
421 tp->out_of_order_queue = RB_ROOT;
422 sk->tcp_rtx_queue = RB_ROOT;
423 tcp_init_xmit_timers(sk);
424 INIT_LIST_HEAD(&tp->tsq_node);
425 INIT_LIST_HEAD(&tp->tsorted_sent_queue);
426
427 icsk->icsk_rto = TCP_TIMEOUT_INIT;
428 icsk->icsk_rto_min = TCP_RTO_MIN;
429 icsk->icsk_delack_max = TCP_DELACK_MAX;
430 tp->mdev_us = jiffies_to_usecs(TCP_TIMEOUT_INIT);
431 minmax_reset(&tp->rtt_min, tcp_jiffies32, ~0U);
432
433 /* So many TCP implementations out there (incorrectly) count the
434 * initial SYN frame in their delayed-ACK and congestion control
435 * algorithms that we must have the following bandaid to talk
436 * efficiently to them. -DaveM
437 */
438 tp->snd_cwnd = TCP_INIT_CWND;
439
440 /* There's a bubble in the pipe until at least the first ACK. */
441 tp->app_limited = ~0U;
442
443 /* See draft-stevens-tcpca-spec-01 for discussion of the
444 * initialization of these values.
445 */
446 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
447 tp->snd_cwnd_clamp = ~0;
448 tp->mss_cache = TCP_MSS_DEFAULT;
449
450 tp->reordering = sock_net(sk)->ipv4.sysctl_tcp_reordering;
451 tcp_assign_congestion_control(sk);
452
453 tp->tsoffset = 0;
454 tp->rack.reo_wnd_steps = 1;
455
456 sk->sk_write_space = sk_stream_write_space;
457 sock_set_flag(sk, SOCK_USE_WRITE_QUEUE);
458
459 icsk->icsk_sync_mss = tcp_sync_mss;
460
461 WRITE_ONCE(sk->sk_sndbuf, sock_net(sk)->ipv4.sysctl_tcp_wmem[1]);
462 WRITE_ONCE(sk->sk_rcvbuf, sock_net(sk)->ipv4.sysctl_tcp_rmem[1]);
463
464 sk_sockets_allocated_inc(sk);
465 sk->sk_route_forced_caps = NETIF_F_GSO;
466 }
467 EXPORT_SYMBOL(tcp_init_sock);
468
tcp_tx_timestamp(struct sock * sk,u16 tsflags)469 static void tcp_tx_timestamp(struct sock *sk, u16 tsflags)
470 {
471 struct sk_buff *skb = tcp_write_queue_tail(sk);
472
473 if (tsflags && skb) {
474 struct skb_shared_info *shinfo = skb_shinfo(skb);
475 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
476
477 sock_tx_timestamp(sk, tsflags, &shinfo->tx_flags);
478 if (tsflags & SOF_TIMESTAMPING_TX_ACK)
479 tcb->txstamp_ack = 1;
480 if (tsflags & SOF_TIMESTAMPING_TX_RECORD_MASK)
481 shinfo->tskey = TCP_SKB_CB(skb)->seq + skb->len - 1;
482 }
483 }
484
tcp_stream_is_readable(struct sock * sk,int target)485 static bool tcp_stream_is_readable(struct sock *sk, int target)
486 {
487 if (tcp_epollin_ready(sk, target))
488 return true;
489
490 if (sk->sk_prot->stream_memory_read)
491 return sk->sk_prot->stream_memory_read(sk);
492 return false;
493 }
494
495 /*
496 * Wait for a TCP event.
497 *
498 * Note that we don't need to lock the socket, as the upper poll layers
499 * take care of normal races (between the test and the event) and we don't
500 * go look at any of the socket buffers directly.
501 */
tcp_poll(struct file * file,struct socket * sock,poll_table * wait)502 __poll_t tcp_poll(struct file *file, struct socket *sock, poll_table *wait)
503 {
504 __poll_t mask;
505 struct sock *sk = sock->sk;
506 const struct tcp_sock *tp = tcp_sk(sk);
507 int state;
508
509 sock_poll_wait(file, sock, wait);
510
511 state = inet_sk_state_load(sk);
512 if (state == TCP_LISTEN)
513 return inet_csk_listen_poll(sk);
514
515 /* Socket is not locked. We are protected from async events
516 * by poll logic and correct handling of state changes
517 * made by other threads is impossible in any case.
518 */
519
520 mask = 0;
521
522 /*
523 * EPOLLHUP is certainly not done right. But poll() doesn't
524 * have a notion of HUP in just one direction, and for a
525 * socket the read side is more interesting.
526 *
527 * Some poll() documentation says that EPOLLHUP is incompatible
528 * with the EPOLLOUT/POLLWR flags, so somebody should check this
529 * all. But careful, it tends to be safer to return too many
530 * bits than too few, and you can easily break real applications
531 * if you don't tell them that something has hung up!
532 *
533 * Check-me.
534 *
535 * Check number 1. EPOLLHUP is _UNMASKABLE_ event (see UNIX98 and
536 * our fs/select.c). It means that after we received EOF,
537 * poll always returns immediately, making impossible poll() on write()
538 * in state CLOSE_WAIT. One solution is evident --- to set EPOLLHUP
539 * if and only if shutdown has been made in both directions.
540 * Actually, it is interesting to look how Solaris and DUX
541 * solve this dilemma. I would prefer, if EPOLLHUP were maskable,
542 * then we could set it on SND_SHUTDOWN. BTW examples given
543 * in Stevens' books assume exactly this behaviour, it explains
544 * why EPOLLHUP is incompatible with EPOLLOUT. --ANK
545 *
546 * NOTE. Check for TCP_CLOSE is added. The goal is to prevent
547 * blocking on fresh not-connected or disconnected socket. --ANK
548 */
549 if (sk->sk_shutdown == SHUTDOWN_MASK || state == TCP_CLOSE)
550 mask |= EPOLLHUP;
551 if (sk->sk_shutdown & RCV_SHUTDOWN)
552 mask |= EPOLLIN | EPOLLRDNORM | EPOLLRDHUP;
553
554 /* Connected or passive Fast Open socket? */
555 if (state != TCP_SYN_SENT &&
556 (state != TCP_SYN_RECV || rcu_access_pointer(tp->fastopen_rsk))) {
557 int target = sock_rcvlowat(sk, 0, INT_MAX);
558
559 if (READ_ONCE(tp->urg_seq) == READ_ONCE(tp->copied_seq) &&
560 !sock_flag(sk, SOCK_URGINLINE) &&
561 tp->urg_data)
562 target++;
563
564 if (tcp_stream_is_readable(sk, target))
565 mask |= EPOLLIN | EPOLLRDNORM;
566
567 if (!(sk->sk_shutdown & SEND_SHUTDOWN)) {
568 if (__sk_stream_is_writeable(sk, 1)) {
569 mask |= EPOLLOUT | EPOLLWRNORM;
570 } else { /* send SIGIO later */
571 sk_set_bit(SOCKWQ_ASYNC_NOSPACE, sk);
572 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
573
574 /* Race breaker. If space is freed after
575 * wspace test but before the flags are set,
576 * IO signal will be lost. Memory barrier
577 * pairs with the input side.
578 */
579 smp_mb__after_atomic();
580 if (__sk_stream_is_writeable(sk, 1))
581 mask |= EPOLLOUT | EPOLLWRNORM;
582 }
583 } else
584 mask |= EPOLLOUT | EPOLLWRNORM;
585
586 if (tp->urg_data & TCP_URG_VALID)
587 mask |= EPOLLPRI;
588 } else if (state == TCP_SYN_SENT && inet_sk(sk)->defer_connect) {
589 /* Active TCP fastopen socket with defer_connect
590 * Return EPOLLOUT so application can call write()
591 * in order for kernel to generate SYN+data
592 */
593 mask |= EPOLLOUT | EPOLLWRNORM;
594 }
595 /* This barrier is coupled with smp_wmb() in tcp_reset() */
596 smp_rmb();
597 if (sk->sk_err || !skb_queue_empty_lockless(&sk->sk_error_queue))
598 mask |= EPOLLERR;
599
600 return mask;
601 }
602 EXPORT_SYMBOL(tcp_poll);
603
tcp_ioctl(struct sock * sk,int cmd,unsigned long arg)604 int tcp_ioctl(struct sock *sk, int cmd, unsigned long arg)
605 {
606 struct tcp_sock *tp = tcp_sk(sk);
607 int answ;
608 bool slow;
609
610 switch (cmd) {
611 case SIOCINQ:
612 if (sk->sk_state == TCP_LISTEN)
613 return -EINVAL;
614
615 slow = lock_sock_fast(sk);
616 answ = tcp_inq(sk);
617 unlock_sock_fast(sk, slow);
618 break;
619 case SIOCATMARK:
620 answ = tp->urg_data &&
621 READ_ONCE(tp->urg_seq) == READ_ONCE(tp->copied_seq);
622 break;
623 case SIOCOUTQ:
624 if (sk->sk_state == TCP_LISTEN)
625 return -EINVAL;
626
627 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
628 answ = 0;
629 else
630 answ = READ_ONCE(tp->write_seq) - tp->snd_una;
631 break;
632 case SIOCOUTQNSD:
633 if (sk->sk_state == TCP_LISTEN)
634 return -EINVAL;
635
636 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
637 answ = 0;
638 else
639 answ = READ_ONCE(tp->write_seq) -
640 READ_ONCE(tp->snd_nxt);
641 break;
642 default:
643 return -ENOIOCTLCMD;
644 }
645
646 return put_user(answ, (int __user *)arg);
647 }
648 EXPORT_SYMBOL(tcp_ioctl);
649
tcp_mark_push(struct tcp_sock * tp,struct sk_buff * skb)650 static inline void tcp_mark_push(struct tcp_sock *tp, struct sk_buff *skb)
651 {
652 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
653 tp->pushed_seq = tp->write_seq;
654 }
655
forced_push(const struct tcp_sock * tp)656 static inline bool forced_push(const struct tcp_sock *tp)
657 {
658 return after(tp->write_seq, tp->pushed_seq + (tp->max_window >> 1));
659 }
660
skb_entail(struct sock * sk,struct sk_buff * skb)661 static void skb_entail(struct sock *sk, struct sk_buff *skb)
662 {
663 struct tcp_sock *tp = tcp_sk(sk);
664 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
665
666 skb->csum = 0;
667 tcb->seq = tcb->end_seq = tp->write_seq;
668 tcb->tcp_flags = TCPHDR_ACK;
669 tcb->sacked = 0;
670 __skb_header_release(skb);
671 tcp_add_write_queue_tail(sk, skb);
672 sk_wmem_queued_add(sk, skb->truesize);
673 sk_mem_charge(sk, skb->truesize);
674 if (tp->nonagle & TCP_NAGLE_PUSH)
675 tp->nonagle &= ~TCP_NAGLE_PUSH;
676
677 tcp_slow_start_after_idle_check(sk);
678 }
679
tcp_mark_urg(struct tcp_sock * tp,int flags)680 static inline void tcp_mark_urg(struct tcp_sock *tp, int flags)
681 {
682 if (flags & MSG_OOB)
683 tp->snd_up = tp->write_seq;
684 }
685
686 /* If a not yet filled skb is pushed, do not send it if
687 * we have data packets in Qdisc or NIC queues :
688 * Because TX completion will happen shortly, it gives a chance
689 * to coalesce future sendmsg() payload into this skb, without
690 * need for a timer, and with no latency trade off.
691 * As packets containing data payload have a bigger truesize
692 * than pure acks (dataless) packets, the last checks prevent
693 * autocorking if we only have an ACK in Qdisc/NIC queues,
694 * or if TX completion was delayed after we processed ACK packet.
695 */
tcp_should_autocork(struct sock * sk,struct sk_buff * skb,int size_goal)696 static bool tcp_should_autocork(struct sock *sk, struct sk_buff *skb,
697 int size_goal)
698 {
699 return skb->len < size_goal &&
700 sock_net(sk)->ipv4.sysctl_tcp_autocorking &&
701 !tcp_rtx_queue_empty(sk) &&
702 refcount_read(&sk->sk_wmem_alloc) > skb->truesize;
703 }
704
tcp_push(struct sock * sk,int flags,int mss_now,int nonagle,int size_goal)705 void tcp_push(struct sock *sk, int flags, int mss_now,
706 int nonagle, int size_goal)
707 {
708 struct tcp_sock *tp = tcp_sk(sk);
709 struct sk_buff *skb;
710
711 skb = tcp_write_queue_tail(sk);
712 if (!skb)
713 return;
714 if (!(flags & MSG_MORE) || forced_push(tp))
715 tcp_mark_push(tp, skb);
716
717 tcp_mark_urg(tp, flags);
718
719 if (tcp_should_autocork(sk, skb, size_goal)) {
720
721 /* avoid atomic op if TSQ_THROTTLED bit is already set */
722 if (!test_bit(TSQ_THROTTLED, &sk->sk_tsq_flags)) {
723 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPAUTOCORKING);
724 set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags);
725 }
726 /* It is possible TX completion already happened
727 * before we set TSQ_THROTTLED.
728 */
729 if (refcount_read(&sk->sk_wmem_alloc) > skb->truesize)
730 return;
731 }
732
733 if (flags & MSG_MORE)
734 nonagle = TCP_NAGLE_CORK;
735
736 __tcp_push_pending_frames(sk, mss_now, nonagle);
737 }
738
tcp_splice_data_recv(read_descriptor_t * rd_desc,struct sk_buff * skb,unsigned int offset,size_t len)739 static int tcp_splice_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb,
740 unsigned int offset, size_t len)
741 {
742 struct tcp_splice_state *tss = rd_desc->arg.data;
743 int ret;
744
745 ret = skb_splice_bits(skb, skb->sk, offset, tss->pipe,
746 min(rd_desc->count, len), tss->flags);
747 if (ret > 0)
748 rd_desc->count -= ret;
749 return ret;
750 }
751
__tcp_splice_read(struct sock * sk,struct tcp_splice_state * tss)752 static int __tcp_splice_read(struct sock *sk, struct tcp_splice_state *tss)
753 {
754 /* Store TCP splice context information in read_descriptor_t. */
755 read_descriptor_t rd_desc = {
756 .arg.data = tss,
757 .count = tss->len,
758 };
759
760 return tcp_read_sock(sk, &rd_desc, tcp_splice_data_recv);
761 }
762
763 /**
764 * tcp_splice_read - splice data from TCP socket to a pipe
765 * @sock: socket to splice from
766 * @ppos: position (not valid)
767 * @pipe: pipe to splice to
768 * @len: number of bytes to splice
769 * @flags: splice modifier flags
770 *
771 * Description:
772 * Will read pages from given socket and fill them into a pipe.
773 *
774 **/
tcp_splice_read(struct socket * sock,loff_t * ppos,struct pipe_inode_info * pipe,size_t len,unsigned int flags)775 ssize_t tcp_splice_read(struct socket *sock, loff_t *ppos,
776 struct pipe_inode_info *pipe, size_t len,
777 unsigned int flags)
778 {
779 struct sock *sk = sock->sk;
780 struct tcp_splice_state tss = {
781 .pipe = pipe,
782 .len = len,
783 .flags = flags,
784 };
785 long timeo;
786 ssize_t spliced;
787 int ret;
788
789 sock_rps_record_flow(sk);
790 /*
791 * We can't seek on a socket input
792 */
793 if (unlikely(*ppos))
794 return -ESPIPE;
795
796 ret = spliced = 0;
797
798 lock_sock(sk);
799
800 timeo = sock_rcvtimeo(sk, sock->file->f_flags & O_NONBLOCK);
801 while (tss.len) {
802 ret = __tcp_splice_read(sk, &tss);
803 if (ret < 0)
804 break;
805 else if (!ret) {
806 if (spliced)
807 break;
808 if (sock_flag(sk, SOCK_DONE))
809 break;
810 if (sk->sk_err) {
811 ret = sock_error(sk);
812 break;
813 }
814 if (sk->sk_shutdown & RCV_SHUTDOWN)
815 break;
816 if (sk->sk_state == TCP_CLOSE) {
817 /*
818 * This occurs when user tries to read
819 * from never connected socket.
820 */
821 ret = -ENOTCONN;
822 break;
823 }
824 if (!timeo) {
825 ret = -EAGAIN;
826 break;
827 }
828 /* if __tcp_splice_read() got nothing while we have
829 * an skb in receive queue, we do not want to loop.
830 * This might happen with URG data.
831 */
832 if (!skb_queue_empty(&sk->sk_receive_queue))
833 break;
834 sk_wait_data(sk, &timeo, NULL);
835 if (signal_pending(current)) {
836 ret = sock_intr_errno(timeo);
837 break;
838 }
839 continue;
840 }
841 tss.len -= ret;
842 spliced += ret;
843
844 if (!timeo)
845 break;
846 release_sock(sk);
847 lock_sock(sk);
848
849 if (sk->sk_err || sk->sk_state == TCP_CLOSE ||
850 (sk->sk_shutdown & RCV_SHUTDOWN) ||
851 signal_pending(current))
852 break;
853 }
854
855 release_sock(sk);
856
857 if (spliced)
858 return spliced;
859
860 return ret;
861 }
862 EXPORT_SYMBOL(tcp_splice_read);
863
sk_stream_alloc_skb(struct sock * sk,int size,gfp_t gfp,bool force_schedule)864 struct sk_buff *sk_stream_alloc_skb(struct sock *sk, int size, gfp_t gfp,
865 bool force_schedule)
866 {
867 struct sk_buff *skb;
868
869 if (likely(!size)) {
870 skb = sk->sk_tx_skb_cache;
871 if (skb) {
872 skb->truesize = SKB_TRUESIZE(skb_end_offset(skb));
873 sk->sk_tx_skb_cache = NULL;
874 pskb_trim(skb, 0);
875 INIT_LIST_HEAD(&skb->tcp_tsorted_anchor);
876 skb_shinfo(skb)->tx_flags = 0;
877 memset(TCP_SKB_CB(skb), 0, sizeof(struct tcp_skb_cb));
878 return skb;
879 }
880 }
881 /* The TCP header must be at least 32-bit aligned. */
882 size = ALIGN(size, 4);
883
884 if (unlikely(tcp_under_memory_pressure(sk)))
885 sk_mem_reclaim_partial(sk);
886
887 skb = alloc_skb_fclone(size + sk->sk_prot->max_header, gfp);
888 if (likely(skb)) {
889 bool mem_scheduled;
890
891 if (force_schedule) {
892 mem_scheduled = true;
893 sk_forced_mem_schedule(sk, skb->truesize);
894 } else {
895 mem_scheduled = sk_wmem_schedule(sk, skb->truesize);
896 }
897 if (likely(mem_scheduled)) {
898 skb_reserve(skb, sk->sk_prot->max_header);
899 /*
900 * Make sure that we have exactly size bytes
901 * available to the caller, no more, no less.
902 */
903 skb->reserved_tailroom = skb->end - skb->tail - size;
904 INIT_LIST_HEAD(&skb->tcp_tsorted_anchor);
905 return skb;
906 }
907 __kfree_skb(skb);
908 } else {
909 sk->sk_prot->enter_memory_pressure(sk);
910 sk_stream_moderate_sndbuf(sk);
911 }
912 return NULL;
913 }
914
tcp_xmit_size_goal(struct sock * sk,u32 mss_now,int large_allowed)915 static unsigned int tcp_xmit_size_goal(struct sock *sk, u32 mss_now,
916 int large_allowed)
917 {
918 struct tcp_sock *tp = tcp_sk(sk);
919 u32 new_size_goal, size_goal;
920
921 if (!large_allowed)
922 return mss_now;
923
924 /* Note : tcp_tso_autosize() will eventually split this later */
925 new_size_goal = sk->sk_gso_max_size - 1 - MAX_TCP_HEADER;
926 new_size_goal = tcp_bound_to_half_wnd(tp, new_size_goal);
927
928 /* We try hard to avoid divides here */
929 size_goal = tp->gso_segs * mss_now;
930 if (unlikely(new_size_goal < size_goal ||
931 new_size_goal >= size_goal + mss_now)) {
932 tp->gso_segs = min_t(u16, new_size_goal / mss_now,
933 sk->sk_gso_max_segs);
934 size_goal = tp->gso_segs * mss_now;
935 }
936
937 return max(size_goal, mss_now);
938 }
939
tcp_send_mss(struct sock * sk,int * size_goal,int flags)940 int tcp_send_mss(struct sock *sk, int *size_goal, int flags)
941 {
942 int mss_now;
943
944 mss_now = tcp_current_mss(sk);
945 *size_goal = tcp_xmit_size_goal(sk, mss_now, !(flags & MSG_OOB));
946
947 return mss_now;
948 }
949
950 /* In some cases, both sendpage() and sendmsg() could have added
951 * an skb to the write queue, but failed adding payload on it.
952 * We need to remove it to consume less memory, but more
953 * importantly be able to generate EPOLLOUT for Edge Trigger epoll()
954 * users.
955 */
tcp_remove_empty_skb(struct sock * sk,struct sk_buff * skb)956 void tcp_remove_empty_skb(struct sock *sk, struct sk_buff *skb)
957 {
958 if (skb && !skb->len) {
959 tcp_unlink_write_queue(skb, sk);
960 if (tcp_write_queue_empty(sk))
961 tcp_chrono_stop(sk, TCP_CHRONO_BUSY);
962 sk_wmem_free_skb(sk, skb);
963 }
964 }
965
tcp_build_frag(struct sock * sk,int size_goal,int flags,struct page * page,int offset,size_t * size)966 struct sk_buff *tcp_build_frag(struct sock *sk, int size_goal, int flags,
967 struct page *page, int offset, size_t *size)
968 {
969 struct sk_buff *skb = tcp_write_queue_tail(sk);
970 struct tcp_sock *tp = tcp_sk(sk);
971 bool can_coalesce;
972 int copy, i;
973
974 if (!skb || (copy = size_goal - skb->len) <= 0 ||
975 !tcp_skb_can_collapse_to(skb)) {
976 new_segment:
977 if (!sk_stream_memory_free(sk))
978 return NULL;
979
980 skb = sk_stream_alloc_skb(sk, 0, sk->sk_allocation,
981 tcp_rtx_and_write_queues_empty(sk));
982 if (!skb)
983 return NULL;
984
985 #ifdef CONFIG_TLS_DEVICE
986 skb->decrypted = !!(flags & MSG_SENDPAGE_DECRYPTED);
987 #endif
988 skb_entail(sk, skb);
989 copy = size_goal;
990 }
991
992 if (copy > *size)
993 copy = *size;
994
995 i = skb_shinfo(skb)->nr_frags;
996 can_coalesce = skb_can_coalesce(skb, i, page, offset);
997 if (!can_coalesce && i >= sysctl_max_skb_frags) {
998 tcp_mark_push(tp, skb);
999 goto new_segment;
1000 }
1001 if (!sk_wmem_schedule(sk, copy))
1002 return NULL;
1003
1004 if (can_coalesce) {
1005 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1006 } else {
1007 get_page(page);
1008 skb_fill_page_desc(skb, i, page, offset, copy);
1009 }
1010
1011 if (!(flags & MSG_NO_SHARED_FRAGS))
1012 skb_shinfo(skb)->flags |= SKBFL_SHARED_FRAG;
1013
1014 skb->len += copy;
1015 skb->data_len += copy;
1016 skb->truesize += copy;
1017 sk_wmem_queued_add(sk, copy);
1018 sk_mem_charge(sk, copy);
1019 skb->ip_summed = CHECKSUM_PARTIAL;
1020 WRITE_ONCE(tp->write_seq, tp->write_seq + copy);
1021 TCP_SKB_CB(skb)->end_seq += copy;
1022 tcp_skb_pcount_set(skb, 0);
1023
1024 *size = copy;
1025 return skb;
1026 }
1027
do_tcp_sendpages(struct sock * sk,struct page * page,int offset,size_t size,int flags)1028 ssize_t do_tcp_sendpages(struct sock *sk, struct page *page, int offset,
1029 size_t size, int flags)
1030 {
1031 struct tcp_sock *tp = tcp_sk(sk);
1032 int mss_now, size_goal;
1033 int err;
1034 ssize_t copied;
1035 long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
1036
1037 if (IS_ENABLED(CONFIG_DEBUG_VM) &&
1038 WARN_ONCE(!sendpage_ok(page),
1039 "page must not be a Slab one and have page_count > 0"))
1040 return -EINVAL;
1041
1042 /* Wait for a connection to finish. One exception is TCP Fast Open
1043 * (passive side) where data is allowed to be sent before a connection
1044 * is fully established.
1045 */
1046 if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) &&
1047 !tcp_passive_fastopen(sk)) {
1048 err = sk_stream_wait_connect(sk, &timeo);
1049 if (err != 0)
1050 goto out_err;
1051 }
1052
1053 sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
1054
1055 mss_now = tcp_send_mss(sk, &size_goal, flags);
1056 copied = 0;
1057
1058 err = -EPIPE;
1059 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
1060 goto out_err;
1061
1062 while (size > 0) {
1063 struct sk_buff *skb;
1064 size_t copy = size;
1065
1066 skb = tcp_build_frag(sk, size_goal, flags, page, offset, ©);
1067 if (!skb)
1068 goto wait_for_space;
1069
1070 if (!copied)
1071 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
1072
1073 copied += copy;
1074 offset += copy;
1075 size -= copy;
1076 if (!size)
1077 goto out;
1078
1079 if (skb->len < size_goal || (flags & MSG_OOB))
1080 continue;
1081
1082 if (forced_push(tp)) {
1083 tcp_mark_push(tp, skb);
1084 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
1085 } else if (skb == tcp_send_head(sk))
1086 tcp_push_one(sk, mss_now);
1087 continue;
1088
1089 wait_for_space:
1090 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1091 tcp_push(sk, flags & ~MSG_MORE, mss_now,
1092 TCP_NAGLE_PUSH, size_goal);
1093
1094 err = sk_stream_wait_memory(sk, &timeo);
1095 if (err != 0)
1096 goto do_error;
1097
1098 mss_now = tcp_send_mss(sk, &size_goal, flags);
1099 }
1100
1101 out:
1102 if (copied) {
1103 tcp_tx_timestamp(sk, sk->sk_tsflags);
1104 if (!(flags & MSG_SENDPAGE_NOTLAST))
1105 tcp_push(sk, flags, mss_now, tp->nonagle, size_goal);
1106 }
1107 return copied;
1108
1109 do_error:
1110 tcp_remove_empty_skb(sk, tcp_write_queue_tail(sk));
1111 if (copied)
1112 goto out;
1113 out_err:
1114 /* make sure we wake any epoll edge trigger waiter */
1115 if (unlikely(tcp_rtx_and_write_queues_empty(sk) && err == -EAGAIN)) {
1116 sk->sk_write_space(sk);
1117 tcp_chrono_stop(sk, TCP_CHRONO_SNDBUF_LIMITED);
1118 }
1119 return sk_stream_error(sk, flags, err);
1120 }
1121 EXPORT_SYMBOL_GPL(do_tcp_sendpages);
1122
tcp_sendpage_locked(struct sock * sk,struct page * page,int offset,size_t size,int flags)1123 int tcp_sendpage_locked(struct sock *sk, struct page *page, int offset,
1124 size_t size, int flags)
1125 {
1126 if (!(sk->sk_route_caps & NETIF_F_SG))
1127 return sock_no_sendpage_locked(sk, page, offset, size, flags);
1128
1129 tcp_rate_check_app_limited(sk); /* is sending application-limited? */
1130
1131 return do_tcp_sendpages(sk, page, offset, size, flags);
1132 }
1133 EXPORT_SYMBOL_GPL(tcp_sendpage_locked);
1134
tcp_sendpage(struct sock * sk,struct page * page,int offset,size_t size,int flags)1135 int tcp_sendpage(struct sock *sk, struct page *page, int offset,
1136 size_t size, int flags)
1137 {
1138 int ret;
1139
1140 lock_sock(sk);
1141 ret = tcp_sendpage_locked(sk, page, offset, size, flags);
1142 release_sock(sk);
1143
1144 return ret;
1145 }
1146 EXPORT_SYMBOL(tcp_sendpage);
1147
tcp_free_fastopen_req(struct tcp_sock * tp)1148 void tcp_free_fastopen_req(struct tcp_sock *tp)
1149 {
1150 if (tp->fastopen_req) {
1151 kfree(tp->fastopen_req);
1152 tp->fastopen_req = NULL;
1153 }
1154 }
1155
tcp_sendmsg_fastopen(struct sock * sk,struct msghdr * msg,int * copied,size_t size,struct ubuf_info * uarg)1156 static int tcp_sendmsg_fastopen(struct sock *sk, struct msghdr *msg,
1157 int *copied, size_t size,
1158 struct ubuf_info *uarg)
1159 {
1160 struct tcp_sock *tp = tcp_sk(sk);
1161 struct inet_sock *inet = inet_sk(sk);
1162 struct sockaddr *uaddr = msg->msg_name;
1163 int err, flags;
1164
1165 if (!(sock_net(sk)->ipv4.sysctl_tcp_fastopen & TFO_CLIENT_ENABLE) ||
1166 (uaddr && msg->msg_namelen >= sizeof(uaddr->sa_family) &&
1167 uaddr->sa_family == AF_UNSPEC))
1168 return -EOPNOTSUPP;
1169 if (tp->fastopen_req)
1170 return -EALREADY; /* Another Fast Open is in progress */
1171
1172 tp->fastopen_req = kzalloc(sizeof(struct tcp_fastopen_request),
1173 sk->sk_allocation);
1174 if (unlikely(!tp->fastopen_req))
1175 return -ENOBUFS;
1176 tp->fastopen_req->data = msg;
1177 tp->fastopen_req->size = size;
1178 tp->fastopen_req->uarg = uarg;
1179
1180 if (inet->defer_connect) {
1181 err = tcp_connect(sk);
1182 /* Same failure procedure as in tcp_v4/6_connect */
1183 if (err) {
1184 tcp_set_state(sk, TCP_CLOSE);
1185 inet->inet_dport = 0;
1186 sk->sk_route_caps = 0;
1187 }
1188 }
1189 flags = (msg->msg_flags & MSG_DONTWAIT) ? O_NONBLOCK : 0;
1190 err = __inet_stream_connect(sk->sk_socket, uaddr,
1191 msg->msg_namelen, flags, 1);
1192 /* fastopen_req could already be freed in __inet_stream_connect
1193 * if the connection times out or gets rst
1194 */
1195 if (tp->fastopen_req) {
1196 *copied = tp->fastopen_req->copied;
1197 tcp_free_fastopen_req(tp);
1198 inet->defer_connect = 0;
1199 }
1200 return err;
1201 }
1202
tcp_sendmsg_locked(struct sock * sk,struct msghdr * msg,size_t size)1203 int tcp_sendmsg_locked(struct sock *sk, struct msghdr *msg, size_t size)
1204 {
1205 struct tcp_sock *tp = tcp_sk(sk);
1206 struct ubuf_info *uarg = NULL;
1207 struct sk_buff *skb;
1208 struct sockcm_cookie sockc;
1209 int flags, err, copied = 0;
1210 int mss_now = 0, size_goal, copied_syn = 0;
1211 int process_backlog = 0;
1212 bool zc = false;
1213 long timeo;
1214
1215 flags = msg->msg_flags;
1216
1217 if (flags & MSG_ZEROCOPY && size && sock_flag(sk, SOCK_ZEROCOPY)) {
1218 skb = tcp_write_queue_tail(sk);
1219 uarg = msg_zerocopy_realloc(sk, size, skb_zcopy(skb));
1220 if (!uarg) {
1221 err = -ENOBUFS;
1222 goto out_err;
1223 }
1224
1225 zc = sk->sk_route_caps & NETIF_F_SG;
1226 if (!zc)
1227 uarg->zerocopy = 0;
1228 }
1229
1230 if (unlikely(flags & MSG_FASTOPEN || inet_sk(sk)->defer_connect) &&
1231 !tp->repair) {
1232 err = tcp_sendmsg_fastopen(sk, msg, &copied_syn, size, uarg);
1233 if (err == -EINPROGRESS && copied_syn > 0)
1234 goto out;
1235 else if (err)
1236 goto out_err;
1237 }
1238
1239 timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
1240
1241 tcp_rate_check_app_limited(sk); /* is sending application-limited? */
1242
1243 /* Wait for a connection to finish. One exception is TCP Fast Open
1244 * (passive side) where data is allowed to be sent before a connection
1245 * is fully established.
1246 */
1247 if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) &&
1248 !tcp_passive_fastopen(sk)) {
1249 err = sk_stream_wait_connect(sk, &timeo);
1250 if (err != 0)
1251 goto do_error;
1252 }
1253
1254 if (unlikely(tp->repair)) {
1255 if (tp->repair_queue == TCP_RECV_QUEUE) {
1256 copied = tcp_send_rcvq(sk, msg, size);
1257 goto out_nopush;
1258 }
1259
1260 err = -EINVAL;
1261 if (tp->repair_queue == TCP_NO_QUEUE)
1262 goto out_err;
1263
1264 /* 'common' sending to sendq */
1265 }
1266
1267 sockcm_init(&sockc, sk);
1268 if (msg->msg_controllen) {
1269 err = sock_cmsg_send(sk, msg, &sockc);
1270 if (unlikely(err)) {
1271 err = -EINVAL;
1272 goto out_err;
1273 }
1274 }
1275
1276 /* This should be in poll */
1277 sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
1278
1279 /* Ok commence sending. */
1280 copied = 0;
1281
1282 restart:
1283 mss_now = tcp_send_mss(sk, &size_goal, flags);
1284
1285 err = -EPIPE;
1286 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
1287 goto do_error;
1288
1289 while (msg_data_left(msg)) {
1290 int copy = 0;
1291
1292 skb = tcp_write_queue_tail(sk);
1293 if (skb)
1294 copy = size_goal - skb->len;
1295
1296 if (copy <= 0 || !tcp_skb_can_collapse_to(skb)) {
1297 bool first_skb;
1298
1299 new_segment:
1300 if (!sk_stream_memory_free(sk))
1301 goto wait_for_space;
1302
1303 if (unlikely(process_backlog >= 16)) {
1304 process_backlog = 0;
1305 if (sk_flush_backlog(sk))
1306 goto restart;
1307 }
1308 first_skb = tcp_rtx_and_write_queues_empty(sk);
1309 skb = sk_stream_alloc_skb(sk, 0, sk->sk_allocation,
1310 first_skb);
1311 if (!skb)
1312 goto wait_for_space;
1313
1314 process_backlog++;
1315 skb->ip_summed = CHECKSUM_PARTIAL;
1316
1317 skb_entail(sk, skb);
1318 copy = size_goal;
1319
1320 /* All packets are restored as if they have
1321 * already been sent. skb_mstamp_ns isn't set to
1322 * avoid wrong rtt estimation.
1323 */
1324 if (tp->repair)
1325 TCP_SKB_CB(skb)->sacked |= TCPCB_REPAIRED;
1326 }
1327
1328 /* Try to append data to the end of skb. */
1329 if (copy > msg_data_left(msg))
1330 copy = msg_data_left(msg);
1331
1332 /* Where to copy to? */
1333 if (skb_availroom(skb) > 0 && !zc) {
1334 /* We have some space in skb head. Superb! */
1335 copy = min_t(int, copy, skb_availroom(skb));
1336 err = skb_add_data_nocache(sk, skb, &msg->msg_iter, copy);
1337 if (err)
1338 goto do_fault;
1339 } else if (!zc) {
1340 bool merge = true;
1341 int i = skb_shinfo(skb)->nr_frags;
1342 struct page_frag *pfrag = sk_page_frag(sk);
1343
1344 if (!sk_page_frag_refill(sk, pfrag))
1345 goto wait_for_space;
1346
1347 if (!skb_can_coalesce(skb, i, pfrag->page,
1348 pfrag->offset)) {
1349 if (i >= sysctl_max_skb_frags) {
1350 tcp_mark_push(tp, skb);
1351 goto new_segment;
1352 }
1353 merge = false;
1354 }
1355
1356 copy = min_t(int, copy, pfrag->size - pfrag->offset);
1357
1358 if (!sk_wmem_schedule(sk, copy))
1359 goto wait_for_space;
1360
1361 err = skb_copy_to_page_nocache(sk, &msg->msg_iter, skb,
1362 pfrag->page,
1363 pfrag->offset,
1364 copy);
1365 if (err)
1366 goto do_error;
1367
1368 /* Update the skb. */
1369 if (merge) {
1370 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1371 } else {
1372 skb_fill_page_desc(skb, i, pfrag->page,
1373 pfrag->offset, copy);
1374 page_ref_inc(pfrag->page);
1375 }
1376 pfrag->offset += copy;
1377 } else {
1378 err = skb_zerocopy_iter_stream(sk, skb, msg, copy, uarg);
1379 if (err == -EMSGSIZE || err == -EEXIST) {
1380 tcp_mark_push(tp, skb);
1381 goto new_segment;
1382 }
1383 if (err < 0)
1384 goto do_error;
1385 copy = err;
1386 }
1387
1388 if (!copied)
1389 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
1390
1391 WRITE_ONCE(tp->write_seq, tp->write_seq + copy);
1392 TCP_SKB_CB(skb)->end_seq += copy;
1393 tcp_skb_pcount_set(skb, 0);
1394
1395 copied += copy;
1396 if (!msg_data_left(msg)) {
1397 if (unlikely(flags & MSG_EOR))
1398 TCP_SKB_CB(skb)->eor = 1;
1399 goto out;
1400 }
1401
1402 if (skb->len < size_goal || (flags & MSG_OOB) || unlikely(tp->repair))
1403 continue;
1404
1405 if (forced_push(tp)) {
1406 tcp_mark_push(tp, skb);
1407 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
1408 } else if (skb == tcp_send_head(sk))
1409 tcp_push_one(sk, mss_now);
1410 continue;
1411
1412 wait_for_space:
1413 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1414 if (copied)
1415 tcp_push(sk, flags & ~MSG_MORE, mss_now,
1416 TCP_NAGLE_PUSH, size_goal);
1417
1418 err = sk_stream_wait_memory(sk, &timeo);
1419 if (err != 0)
1420 goto do_error;
1421
1422 mss_now = tcp_send_mss(sk, &size_goal, flags);
1423 }
1424
1425 out:
1426 if (copied) {
1427 tcp_tx_timestamp(sk, sockc.tsflags);
1428 tcp_push(sk, flags, mss_now, tp->nonagle, size_goal);
1429 }
1430 out_nopush:
1431 net_zcopy_put(uarg);
1432 return copied + copied_syn;
1433
1434 do_error:
1435 skb = tcp_write_queue_tail(sk);
1436 do_fault:
1437 tcp_remove_empty_skb(sk, skb);
1438
1439 if (copied + copied_syn)
1440 goto out;
1441 out_err:
1442 net_zcopy_put_abort(uarg, true);
1443 err = sk_stream_error(sk, flags, err);
1444 /* make sure we wake any epoll edge trigger waiter */
1445 if (unlikely(tcp_rtx_and_write_queues_empty(sk) && err == -EAGAIN)) {
1446 sk->sk_write_space(sk);
1447 tcp_chrono_stop(sk, TCP_CHRONO_SNDBUF_LIMITED);
1448 }
1449 return err;
1450 }
1451 EXPORT_SYMBOL_GPL(tcp_sendmsg_locked);
1452
tcp_sendmsg(struct sock * sk,struct msghdr * msg,size_t size)1453 int tcp_sendmsg(struct sock *sk, struct msghdr *msg, size_t size)
1454 {
1455 int ret;
1456
1457 lock_sock(sk);
1458 ret = tcp_sendmsg_locked(sk, msg, size);
1459 release_sock(sk);
1460
1461 return ret;
1462 }
1463 EXPORT_SYMBOL(tcp_sendmsg);
1464
1465 /*
1466 * Handle reading urgent data. BSD has very simple semantics for
1467 * this, no blocking and very strange errors 8)
1468 */
1469
tcp_recv_urg(struct sock * sk,struct msghdr * msg,int len,int flags)1470 static int tcp_recv_urg(struct sock *sk, struct msghdr *msg, int len, int flags)
1471 {
1472 struct tcp_sock *tp = tcp_sk(sk);
1473
1474 /* No URG data to read. */
1475 if (sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data ||
1476 tp->urg_data == TCP_URG_READ)
1477 return -EINVAL; /* Yes this is right ! */
1478
1479 if (sk->sk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DONE))
1480 return -ENOTCONN;
1481
1482 if (tp->urg_data & TCP_URG_VALID) {
1483 int err = 0;
1484 char c = tp->urg_data;
1485
1486 if (!(flags & MSG_PEEK))
1487 tp->urg_data = TCP_URG_READ;
1488
1489 /* Read urgent data. */
1490 msg->msg_flags |= MSG_OOB;
1491
1492 if (len > 0) {
1493 if (!(flags & MSG_TRUNC))
1494 err = memcpy_to_msg(msg, &c, 1);
1495 len = 1;
1496 } else
1497 msg->msg_flags |= MSG_TRUNC;
1498
1499 return err ? -EFAULT : len;
1500 }
1501
1502 if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
1503 return 0;
1504
1505 /* Fixed the recv(..., MSG_OOB) behaviour. BSD docs and
1506 * the available implementations agree in this case:
1507 * this call should never block, independent of the
1508 * blocking state of the socket.
1509 * Mike <pall@rz.uni-karlsruhe.de>
1510 */
1511 return -EAGAIN;
1512 }
1513
tcp_peek_sndq(struct sock * sk,struct msghdr * msg,int len)1514 static int tcp_peek_sndq(struct sock *sk, struct msghdr *msg, int len)
1515 {
1516 struct sk_buff *skb;
1517 int copied = 0, err = 0;
1518
1519 /* XXX -- need to support SO_PEEK_OFF */
1520
1521 skb_rbtree_walk(skb, &sk->tcp_rtx_queue) {
1522 err = skb_copy_datagram_msg(skb, 0, msg, skb->len);
1523 if (err)
1524 return err;
1525 copied += skb->len;
1526 }
1527
1528 skb_queue_walk(&sk->sk_write_queue, skb) {
1529 err = skb_copy_datagram_msg(skb, 0, msg, skb->len);
1530 if (err)
1531 break;
1532
1533 copied += skb->len;
1534 }
1535
1536 return err ?: copied;
1537 }
1538
1539 /* Clean up the receive buffer for full frames taken by the user,
1540 * then send an ACK if necessary. COPIED is the number of bytes
1541 * tcp_recvmsg has given to the user so far, it speeds up the
1542 * calculation of whether or not we must ACK for the sake of
1543 * a window update.
1544 */
tcp_cleanup_rbuf(struct sock * sk,int copied)1545 void tcp_cleanup_rbuf(struct sock *sk, int copied)
1546 {
1547 struct tcp_sock *tp = tcp_sk(sk);
1548 bool time_to_ack = false;
1549
1550 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
1551
1552 WARN(skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq),
1553 "cleanup rbuf bug: copied %X seq %X rcvnxt %X\n",
1554 tp->copied_seq, TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt);
1555
1556 if (inet_csk_ack_scheduled(sk)) {
1557 const struct inet_connection_sock *icsk = inet_csk(sk);
1558
1559 if (/* Once-per-two-segments ACK was not sent by tcp_input.c */
1560 tp->rcv_nxt - tp->rcv_wup > icsk->icsk_ack.rcv_mss ||
1561 /*
1562 * If this read emptied read buffer, we send ACK, if
1563 * connection is not bidirectional, user drained
1564 * receive buffer and there was a small segment
1565 * in queue.
1566 */
1567 (copied > 0 &&
1568 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED2) ||
1569 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED) &&
1570 !inet_csk_in_pingpong_mode(sk))) &&
1571 !atomic_read(&sk->sk_rmem_alloc)))
1572 time_to_ack = true;
1573 }
1574
1575 /* We send an ACK if we can now advertise a non-zero window
1576 * which has been raised "significantly".
1577 *
1578 * Even if window raised up to infinity, do not send window open ACK
1579 * in states, where we will not receive more. It is useless.
1580 */
1581 if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) {
1582 __u32 rcv_window_now = tcp_receive_window(tp);
1583
1584 /* Optimize, __tcp_select_window() is not cheap. */
1585 if (2*rcv_window_now <= tp->window_clamp) {
1586 __u32 new_window = __tcp_select_window(sk);
1587
1588 /* Send ACK now, if this read freed lots of space
1589 * in our buffer. Certainly, new_window is new window.
1590 * We can advertise it now, if it is not less than current one.
1591 * "Lots" means "at least twice" here.
1592 */
1593 if (new_window && new_window >= 2 * rcv_window_now)
1594 time_to_ack = true;
1595 }
1596 }
1597 if (time_to_ack)
1598 tcp_send_ack(sk);
1599 }
1600
tcp_recv_skb(struct sock * sk,u32 seq,u32 * off)1601 static struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off)
1602 {
1603 struct sk_buff *skb;
1604 u32 offset;
1605
1606 while ((skb = skb_peek(&sk->sk_receive_queue)) != NULL) {
1607 offset = seq - TCP_SKB_CB(skb)->seq;
1608 if (unlikely(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
1609 pr_err_once("%s: found a SYN, please report !\n", __func__);
1610 offset--;
1611 }
1612 if (offset < skb->len || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)) {
1613 *off = offset;
1614 return skb;
1615 }
1616 /* This looks weird, but this can happen if TCP collapsing
1617 * splitted a fat GRO packet, while we released socket lock
1618 * in skb_splice_bits()
1619 */
1620 sk_eat_skb(sk, skb);
1621 }
1622 return NULL;
1623 }
1624
1625 /*
1626 * This routine provides an alternative to tcp_recvmsg() for routines
1627 * that would like to handle copying from skbuffs directly in 'sendfile'
1628 * fashion.
1629 * Note:
1630 * - It is assumed that the socket was locked by the caller.
1631 * - The routine does not block.
1632 * - At present, there is no support for reading OOB data
1633 * or for 'peeking' the socket using this routine
1634 * (although both would be easy to implement).
1635 */
tcp_read_sock(struct sock * sk,read_descriptor_t * desc,sk_read_actor_t recv_actor)1636 int tcp_read_sock(struct sock *sk, read_descriptor_t *desc,
1637 sk_read_actor_t recv_actor)
1638 {
1639 struct sk_buff *skb;
1640 struct tcp_sock *tp = tcp_sk(sk);
1641 u32 seq = tp->copied_seq;
1642 u32 offset;
1643 int copied = 0;
1644
1645 if (sk->sk_state == TCP_LISTEN)
1646 return -ENOTCONN;
1647 while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1648 if (offset < skb->len) {
1649 int used;
1650 size_t len;
1651
1652 len = skb->len - offset;
1653 /* Stop reading if we hit a patch of urgent data */
1654 if (tp->urg_data) {
1655 u32 urg_offset = tp->urg_seq - seq;
1656 if (urg_offset < len)
1657 len = urg_offset;
1658 if (!len)
1659 break;
1660 }
1661 used = recv_actor(desc, skb, offset, len);
1662 if (used <= 0) {
1663 if (!copied)
1664 copied = used;
1665 break;
1666 } else if (used <= len) {
1667 seq += used;
1668 copied += used;
1669 offset += used;
1670 }
1671 /* If recv_actor drops the lock (e.g. TCP splice
1672 * receive) the skb pointer might be invalid when
1673 * getting here: tcp_collapse might have deleted it
1674 * while aggregating skbs from the socket queue.
1675 */
1676 skb = tcp_recv_skb(sk, seq - 1, &offset);
1677 if (!skb)
1678 break;
1679 /* TCP coalescing might have appended data to the skb.
1680 * Try to splice more frags
1681 */
1682 if (offset + 1 != skb->len)
1683 continue;
1684 }
1685 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) {
1686 sk_eat_skb(sk, skb);
1687 ++seq;
1688 break;
1689 }
1690 sk_eat_skb(sk, skb);
1691 if (!desc->count)
1692 break;
1693 WRITE_ONCE(tp->copied_seq, seq);
1694 }
1695 WRITE_ONCE(tp->copied_seq, seq);
1696
1697 tcp_rcv_space_adjust(sk);
1698
1699 /* Clean up data we have read: This will do ACK frames. */
1700 if (copied > 0) {
1701 tcp_recv_skb(sk, seq, &offset);
1702 tcp_cleanup_rbuf(sk, copied);
1703 }
1704 return copied;
1705 }
1706 EXPORT_SYMBOL(tcp_read_sock);
1707
tcp_peek_len(struct socket * sock)1708 int tcp_peek_len(struct socket *sock)
1709 {
1710 return tcp_inq(sock->sk);
1711 }
1712 EXPORT_SYMBOL(tcp_peek_len);
1713
1714 /* Make sure sk_rcvbuf is big enough to satisfy SO_RCVLOWAT hint */
tcp_set_rcvlowat(struct sock * sk,int val)1715 int tcp_set_rcvlowat(struct sock *sk, int val)
1716 {
1717 int cap;
1718
1719 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK)
1720 cap = sk->sk_rcvbuf >> 1;
1721 else
1722 cap = sock_net(sk)->ipv4.sysctl_tcp_rmem[2] >> 1;
1723 val = min(val, cap);
1724 WRITE_ONCE(sk->sk_rcvlowat, val ? : 1);
1725
1726 /* Check if we need to signal EPOLLIN right now */
1727 tcp_data_ready(sk);
1728
1729 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK)
1730 return 0;
1731
1732 val <<= 1;
1733 if (val > sk->sk_rcvbuf) {
1734 WRITE_ONCE(sk->sk_rcvbuf, val);
1735 tcp_sk(sk)->window_clamp = tcp_win_from_space(sk, val);
1736 }
1737 return 0;
1738 }
1739 EXPORT_SYMBOL(tcp_set_rcvlowat);
1740
tcp_update_recv_tstamps(struct sk_buff * skb,struct scm_timestamping_internal * tss)1741 static void tcp_update_recv_tstamps(struct sk_buff *skb,
1742 struct scm_timestamping_internal *tss)
1743 {
1744 if (skb->tstamp)
1745 tss->ts[0] = ktime_to_timespec64(skb->tstamp);
1746 else
1747 tss->ts[0] = (struct timespec64) {0};
1748
1749 if (skb_hwtstamps(skb)->hwtstamp)
1750 tss->ts[2] = ktime_to_timespec64(skb_hwtstamps(skb)->hwtstamp);
1751 else
1752 tss->ts[2] = (struct timespec64) {0};
1753 }
1754
1755 #ifdef CONFIG_MMU
1756 static const struct vm_operations_struct tcp_vm_ops = {
1757 };
1758
tcp_mmap(struct file * file,struct socket * sock,struct vm_area_struct * vma)1759 int tcp_mmap(struct file *file, struct socket *sock,
1760 struct vm_area_struct *vma)
1761 {
1762 if (vma->vm_flags & (VM_WRITE | VM_EXEC))
1763 return -EPERM;
1764 vma->vm_flags &= ~(VM_MAYWRITE | VM_MAYEXEC);
1765
1766 /* Instruct vm_insert_page() to not mmap_read_lock(mm) */
1767 vma->vm_flags |= VM_MIXEDMAP;
1768
1769 vma->vm_ops = &tcp_vm_ops;
1770 return 0;
1771 }
1772 EXPORT_SYMBOL(tcp_mmap);
1773
skb_advance_to_frag(struct sk_buff * skb,u32 offset_skb,u32 * offset_frag)1774 static skb_frag_t *skb_advance_to_frag(struct sk_buff *skb, u32 offset_skb,
1775 u32 *offset_frag)
1776 {
1777 skb_frag_t *frag;
1778
1779 offset_skb -= skb_headlen(skb);
1780 if ((int)offset_skb < 0 || skb_has_frag_list(skb))
1781 return NULL;
1782
1783 frag = skb_shinfo(skb)->frags;
1784 while (offset_skb) {
1785 if (skb_frag_size(frag) > offset_skb) {
1786 *offset_frag = offset_skb;
1787 return frag;
1788 }
1789 offset_skb -= skb_frag_size(frag);
1790 ++frag;
1791 }
1792 *offset_frag = 0;
1793 return frag;
1794 }
1795
can_map_frag(const skb_frag_t * frag)1796 static bool can_map_frag(const skb_frag_t *frag)
1797 {
1798 return skb_frag_size(frag) == PAGE_SIZE && !skb_frag_off(frag);
1799 }
1800
find_next_mappable_frag(const skb_frag_t * frag,int remaining_in_skb)1801 static int find_next_mappable_frag(const skb_frag_t *frag,
1802 int remaining_in_skb)
1803 {
1804 int offset = 0;
1805
1806 if (likely(can_map_frag(frag)))
1807 return 0;
1808
1809 while (offset < remaining_in_skb && !can_map_frag(frag)) {
1810 offset += skb_frag_size(frag);
1811 ++frag;
1812 }
1813 return offset;
1814 }
1815
tcp_zerocopy_set_hint_for_skb(struct sock * sk,struct tcp_zerocopy_receive * zc,struct sk_buff * skb,u32 offset)1816 static void tcp_zerocopy_set_hint_for_skb(struct sock *sk,
1817 struct tcp_zerocopy_receive *zc,
1818 struct sk_buff *skb, u32 offset)
1819 {
1820 u32 frag_offset, partial_frag_remainder = 0;
1821 int mappable_offset;
1822 skb_frag_t *frag;
1823
1824 /* worst case: skip to next skb. try to improve on this case below */
1825 zc->recv_skip_hint = skb->len - offset;
1826
1827 /* Find the frag containing this offset (and how far into that frag) */
1828 frag = skb_advance_to_frag(skb, offset, &frag_offset);
1829 if (!frag)
1830 return;
1831
1832 if (frag_offset) {
1833 struct skb_shared_info *info = skb_shinfo(skb);
1834
1835 /* We read part of the last frag, must recvmsg() rest of skb. */
1836 if (frag == &info->frags[info->nr_frags - 1])
1837 return;
1838
1839 /* Else, we must at least read the remainder in this frag. */
1840 partial_frag_remainder = skb_frag_size(frag) - frag_offset;
1841 zc->recv_skip_hint -= partial_frag_remainder;
1842 ++frag;
1843 }
1844
1845 /* partial_frag_remainder: If part way through a frag, must read rest.
1846 * mappable_offset: Bytes till next mappable frag, *not* counting bytes
1847 * in partial_frag_remainder.
1848 */
1849 mappable_offset = find_next_mappable_frag(frag, zc->recv_skip_hint);
1850 zc->recv_skip_hint = mappable_offset + partial_frag_remainder;
1851 }
1852
1853 static int tcp_recvmsg_locked(struct sock *sk, struct msghdr *msg, size_t len,
1854 int nonblock, int flags,
1855 struct scm_timestamping_internal *tss,
1856 int *cmsg_flags);
receive_fallback_to_copy(struct sock * sk,struct tcp_zerocopy_receive * zc,int inq,struct scm_timestamping_internal * tss)1857 static int receive_fallback_to_copy(struct sock *sk,
1858 struct tcp_zerocopy_receive *zc, int inq,
1859 struct scm_timestamping_internal *tss)
1860 {
1861 unsigned long copy_address = (unsigned long)zc->copybuf_address;
1862 struct msghdr msg = {};
1863 struct iovec iov;
1864 int err;
1865
1866 zc->length = 0;
1867 zc->recv_skip_hint = 0;
1868
1869 if (copy_address != zc->copybuf_address)
1870 return -EINVAL;
1871
1872 err = import_single_range(READ, (void __user *)copy_address,
1873 inq, &iov, &msg.msg_iter);
1874 if (err)
1875 return err;
1876
1877 err = tcp_recvmsg_locked(sk, &msg, inq, /*nonblock=*/1, /*flags=*/0,
1878 tss, &zc->msg_flags);
1879 if (err < 0)
1880 return err;
1881
1882 zc->copybuf_len = err;
1883 if (likely(zc->copybuf_len)) {
1884 struct sk_buff *skb;
1885 u32 offset;
1886
1887 skb = tcp_recv_skb(sk, tcp_sk(sk)->copied_seq, &offset);
1888 if (skb)
1889 tcp_zerocopy_set_hint_for_skb(sk, zc, skb, offset);
1890 }
1891 return 0;
1892 }
1893
tcp_copy_straggler_data(struct tcp_zerocopy_receive * zc,struct sk_buff * skb,u32 copylen,u32 * offset,u32 * seq)1894 static int tcp_copy_straggler_data(struct tcp_zerocopy_receive *zc,
1895 struct sk_buff *skb, u32 copylen,
1896 u32 *offset, u32 *seq)
1897 {
1898 unsigned long copy_address = (unsigned long)zc->copybuf_address;
1899 struct msghdr msg = {};
1900 struct iovec iov;
1901 int err;
1902
1903 if (copy_address != zc->copybuf_address)
1904 return -EINVAL;
1905
1906 err = import_single_range(READ, (void __user *)copy_address,
1907 copylen, &iov, &msg.msg_iter);
1908 if (err)
1909 return err;
1910 err = skb_copy_datagram_msg(skb, *offset, &msg, copylen);
1911 if (err)
1912 return err;
1913 zc->recv_skip_hint -= copylen;
1914 *offset += copylen;
1915 *seq += copylen;
1916 return (__s32)copylen;
1917 }
1918
tcp_zc_handle_leftover(struct tcp_zerocopy_receive * zc,struct sock * sk,struct sk_buff * skb,u32 * seq,s32 copybuf_len,struct scm_timestamping_internal * tss)1919 static int tcp_zc_handle_leftover(struct tcp_zerocopy_receive *zc,
1920 struct sock *sk,
1921 struct sk_buff *skb,
1922 u32 *seq,
1923 s32 copybuf_len,
1924 struct scm_timestamping_internal *tss)
1925 {
1926 u32 offset, copylen = min_t(u32, copybuf_len, zc->recv_skip_hint);
1927
1928 if (!copylen)
1929 return 0;
1930 /* skb is null if inq < PAGE_SIZE. */
1931 if (skb) {
1932 offset = *seq - TCP_SKB_CB(skb)->seq;
1933 } else {
1934 skb = tcp_recv_skb(sk, *seq, &offset);
1935 if (TCP_SKB_CB(skb)->has_rxtstamp) {
1936 tcp_update_recv_tstamps(skb, tss);
1937 zc->msg_flags |= TCP_CMSG_TS;
1938 }
1939 }
1940
1941 zc->copybuf_len = tcp_copy_straggler_data(zc, skb, copylen, &offset,
1942 seq);
1943 return zc->copybuf_len < 0 ? 0 : copylen;
1944 }
1945
tcp_zerocopy_vm_insert_batch_error(struct vm_area_struct * vma,struct page ** pending_pages,unsigned long pages_remaining,unsigned long * address,u32 * length,u32 * seq,struct tcp_zerocopy_receive * zc,u32 total_bytes_to_map,int err)1946 static int tcp_zerocopy_vm_insert_batch_error(struct vm_area_struct *vma,
1947 struct page **pending_pages,
1948 unsigned long pages_remaining,
1949 unsigned long *address,
1950 u32 *length,
1951 u32 *seq,
1952 struct tcp_zerocopy_receive *zc,
1953 u32 total_bytes_to_map,
1954 int err)
1955 {
1956 /* At least one page did not map. Try zapping if we skipped earlier. */
1957 if (err == -EBUSY &&
1958 zc->flags & TCP_RECEIVE_ZEROCOPY_FLAG_TLB_CLEAN_HINT) {
1959 u32 maybe_zap_len;
1960
1961 maybe_zap_len = total_bytes_to_map - /* All bytes to map */
1962 *length + /* Mapped or pending */
1963 (pages_remaining * PAGE_SIZE); /* Failed map. */
1964 zap_page_range(vma, *address, maybe_zap_len);
1965 err = 0;
1966 }
1967
1968 if (!err) {
1969 unsigned long leftover_pages = pages_remaining;
1970 int bytes_mapped;
1971
1972 /* We called zap_page_range, try to reinsert. */
1973 err = vm_insert_pages(vma, *address,
1974 pending_pages,
1975 &pages_remaining);
1976 bytes_mapped = PAGE_SIZE * (leftover_pages - pages_remaining);
1977 *seq += bytes_mapped;
1978 *address += bytes_mapped;
1979 }
1980 if (err) {
1981 /* Either we were unable to zap, OR we zapped, retried an
1982 * insert, and still had an issue. Either ways, pages_remaining
1983 * is the number of pages we were unable to map, and we unroll
1984 * some state we speculatively touched before.
1985 */
1986 const int bytes_not_mapped = PAGE_SIZE * pages_remaining;
1987
1988 *length -= bytes_not_mapped;
1989 zc->recv_skip_hint += bytes_not_mapped;
1990 }
1991 return err;
1992 }
1993
tcp_zerocopy_vm_insert_batch(struct vm_area_struct * vma,struct page ** pages,unsigned int pages_to_map,unsigned long * address,u32 * length,u32 * seq,struct tcp_zerocopy_receive * zc,u32 total_bytes_to_map)1994 static int tcp_zerocopy_vm_insert_batch(struct vm_area_struct *vma,
1995 struct page **pages,
1996 unsigned int pages_to_map,
1997 unsigned long *address,
1998 u32 *length,
1999 u32 *seq,
2000 struct tcp_zerocopy_receive *zc,
2001 u32 total_bytes_to_map)
2002 {
2003 unsigned long pages_remaining = pages_to_map;
2004 unsigned int pages_mapped;
2005 unsigned int bytes_mapped;
2006 int err;
2007
2008 err = vm_insert_pages(vma, *address, pages, &pages_remaining);
2009 pages_mapped = pages_to_map - (unsigned int)pages_remaining;
2010 bytes_mapped = PAGE_SIZE * pages_mapped;
2011 /* Even if vm_insert_pages fails, it may have partially succeeded in
2012 * mapping (some but not all of the pages).
2013 */
2014 *seq += bytes_mapped;
2015 *address += bytes_mapped;
2016
2017 if (likely(!err))
2018 return 0;
2019
2020 /* Error: maybe zap and retry + rollback state for failed inserts. */
2021 return tcp_zerocopy_vm_insert_batch_error(vma, pages + pages_mapped,
2022 pages_remaining, address, length, seq, zc, total_bytes_to_map,
2023 err);
2024 }
2025
2026 #define TCP_VALID_ZC_MSG_FLAGS (TCP_CMSG_TS)
2027 static void tcp_recv_timestamp(struct msghdr *msg, const struct sock *sk,
2028 struct scm_timestamping_internal *tss);
tcp_zc_finalize_rx_tstamp(struct sock * sk,struct tcp_zerocopy_receive * zc,struct scm_timestamping_internal * tss)2029 static void tcp_zc_finalize_rx_tstamp(struct sock *sk,
2030 struct tcp_zerocopy_receive *zc,
2031 struct scm_timestamping_internal *tss)
2032 {
2033 unsigned long msg_control_addr;
2034 struct msghdr cmsg_dummy;
2035
2036 msg_control_addr = (unsigned long)zc->msg_control;
2037 cmsg_dummy.msg_control = (void *)msg_control_addr;
2038 cmsg_dummy.msg_controllen =
2039 (__kernel_size_t)zc->msg_controllen;
2040 cmsg_dummy.msg_flags = in_compat_syscall()
2041 ? MSG_CMSG_COMPAT : 0;
2042 cmsg_dummy.msg_control_is_user = true;
2043 zc->msg_flags = 0;
2044 if (zc->msg_control == msg_control_addr &&
2045 zc->msg_controllen == cmsg_dummy.msg_controllen) {
2046 tcp_recv_timestamp(&cmsg_dummy, sk, tss);
2047 zc->msg_control = (__u64)
2048 ((uintptr_t)cmsg_dummy.msg_control);
2049 zc->msg_controllen =
2050 (__u64)cmsg_dummy.msg_controllen;
2051 zc->msg_flags = (__u32)cmsg_dummy.msg_flags;
2052 }
2053 }
2054
2055 #define TCP_ZEROCOPY_PAGE_BATCH_SIZE 32
tcp_zerocopy_receive(struct sock * sk,struct tcp_zerocopy_receive * zc,struct scm_timestamping_internal * tss)2056 static int tcp_zerocopy_receive(struct sock *sk,
2057 struct tcp_zerocopy_receive *zc,
2058 struct scm_timestamping_internal *tss)
2059 {
2060 u32 length = 0, offset, vma_len, avail_len, copylen = 0;
2061 unsigned long address = (unsigned long)zc->address;
2062 struct page *pages[TCP_ZEROCOPY_PAGE_BATCH_SIZE];
2063 s32 copybuf_len = zc->copybuf_len;
2064 struct tcp_sock *tp = tcp_sk(sk);
2065 const skb_frag_t *frags = NULL;
2066 unsigned int pages_to_map = 0;
2067 struct vm_area_struct *vma;
2068 struct sk_buff *skb = NULL;
2069 u32 seq = tp->copied_seq;
2070 u32 total_bytes_to_map;
2071 int inq = tcp_inq(sk);
2072 int ret;
2073
2074 zc->copybuf_len = 0;
2075 zc->msg_flags = 0;
2076
2077 if (address & (PAGE_SIZE - 1) || address != zc->address)
2078 return -EINVAL;
2079
2080 if (sk->sk_state == TCP_LISTEN)
2081 return -ENOTCONN;
2082
2083 sock_rps_record_flow(sk);
2084
2085 if (inq && inq <= copybuf_len)
2086 return receive_fallback_to_copy(sk, zc, inq, tss);
2087
2088 if (inq < PAGE_SIZE) {
2089 zc->length = 0;
2090 zc->recv_skip_hint = inq;
2091 if (!inq && sock_flag(sk, SOCK_DONE))
2092 return -EIO;
2093 return 0;
2094 }
2095
2096 mmap_read_lock(current->mm);
2097
2098 vma = find_vma(current->mm, address);
2099 if (!vma || vma->vm_start > address || vma->vm_ops != &tcp_vm_ops) {
2100 mmap_read_unlock(current->mm);
2101 return -EINVAL;
2102 }
2103 vma_len = min_t(unsigned long, zc->length, vma->vm_end - address);
2104 avail_len = min_t(u32, vma_len, inq);
2105 total_bytes_to_map = avail_len & ~(PAGE_SIZE - 1);
2106 if (total_bytes_to_map) {
2107 if (!(zc->flags & TCP_RECEIVE_ZEROCOPY_FLAG_TLB_CLEAN_HINT))
2108 zap_page_range(vma, address, total_bytes_to_map);
2109 zc->length = total_bytes_to_map;
2110 zc->recv_skip_hint = 0;
2111 } else {
2112 zc->length = avail_len;
2113 zc->recv_skip_hint = avail_len;
2114 }
2115 ret = 0;
2116 while (length + PAGE_SIZE <= zc->length) {
2117 int mappable_offset;
2118 struct page *page;
2119
2120 if (zc->recv_skip_hint < PAGE_SIZE) {
2121 u32 offset_frag;
2122
2123 if (skb) {
2124 if (zc->recv_skip_hint > 0)
2125 break;
2126 skb = skb->next;
2127 offset = seq - TCP_SKB_CB(skb)->seq;
2128 } else {
2129 skb = tcp_recv_skb(sk, seq, &offset);
2130 }
2131
2132 if (TCP_SKB_CB(skb)->has_rxtstamp) {
2133 tcp_update_recv_tstamps(skb, tss);
2134 zc->msg_flags |= TCP_CMSG_TS;
2135 }
2136 zc->recv_skip_hint = skb->len - offset;
2137 frags = skb_advance_to_frag(skb, offset, &offset_frag);
2138 if (!frags || offset_frag)
2139 break;
2140 }
2141
2142 mappable_offset = find_next_mappable_frag(frags,
2143 zc->recv_skip_hint);
2144 if (mappable_offset) {
2145 zc->recv_skip_hint = mappable_offset;
2146 break;
2147 }
2148 page = skb_frag_page(frags);
2149 prefetchw(page);
2150 pages[pages_to_map++] = page;
2151 length += PAGE_SIZE;
2152 zc->recv_skip_hint -= PAGE_SIZE;
2153 frags++;
2154 if (pages_to_map == TCP_ZEROCOPY_PAGE_BATCH_SIZE ||
2155 zc->recv_skip_hint < PAGE_SIZE) {
2156 /* Either full batch, or we're about to go to next skb
2157 * (and we cannot unroll failed ops across skbs).
2158 */
2159 ret = tcp_zerocopy_vm_insert_batch(vma, pages,
2160 pages_to_map,
2161 &address, &length,
2162 &seq, zc,
2163 total_bytes_to_map);
2164 if (ret)
2165 goto out;
2166 pages_to_map = 0;
2167 }
2168 }
2169 if (pages_to_map) {
2170 ret = tcp_zerocopy_vm_insert_batch(vma, pages, pages_to_map,
2171 &address, &length, &seq,
2172 zc, total_bytes_to_map);
2173 }
2174 out:
2175 mmap_read_unlock(current->mm);
2176 /* Try to copy straggler data. */
2177 if (!ret)
2178 copylen = tcp_zc_handle_leftover(zc, sk, skb, &seq, copybuf_len, tss);
2179
2180 if (length + copylen) {
2181 WRITE_ONCE(tp->copied_seq, seq);
2182 tcp_rcv_space_adjust(sk);
2183
2184 /* Clean up data we have read: This will do ACK frames. */
2185 tcp_recv_skb(sk, seq, &offset);
2186 tcp_cleanup_rbuf(sk, length + copylen);
2187 ret = 0;
2188 if (length == zc->length)
2189 zc->recv_skip_hint = 0;
2190 } else {
2191 if (!zc->recv_skip_hint && sock_flag(sk, SOCK_DONE))
2192 ret = -EIO;
2193 }
2194 zc->length = length;
2195 return ret;
2196 }
2197 #endif
2198
2199 /* Similar to __sock_recv_timestamp, but does not require an skb */
tcp_recv_timestamp(struct msghdr * msg,const struct sock * sk,struct scm_timestamping_internal * tss)2200 static void tcp_recv_timestamp(struct msghdr *msg, const struct sock *sk,
2201 struct scm_timestamping_internal *tss)
2202 {
2203 int new_tstamp = sock_flag(sk, SOCK_TSTAMP_NEW);
2204 bool has_timestamping = false;
2205
2206 if (tss->ts[0].tv_sec || tss->ts[0].tv_nsec) {
2207 if (sock_flag(sk, SOCK_RCVTSTAMP)) {
2208 if (sock_flag(sk, SOCK_RCVTSTAMPNS)) {
2209 if (new_tstamp) {
2210 struct __kernel_timespec kts = {
2211 .tv_sec = tss->ts[0].tv_sec,
2212 .tv_nsec = tss->ts[0].tv_nsec,
2213 };
2214 put_cmsg(msg, SOL_SOCKET, SO_TIMESTAMPNS_NEW,
2215 sizeof(kts), &kts);
2216 } else {
2217 struct __kernel_old_timespec ts_old = {
2218 .tv_sec = tss->ts[0].tv_sec,
2219 .tv_nsec = tss->ts[0].tv_nsec,
2220 };
2221 put_cmsg(msg, SOL_SOCKET, SO_TIMESTAMPNS_OLD,
2222 sizeof(ts_old), &ts_old);
2223 }
2224 } else {
2225 if (new_tstamp) {
2226 struct __kernel_sock_timeval stv = {
2227 .tv_sec = tss->ts[0].tv_sec,
2228 .tv_usec = tss->ts[0].tv_nsec / 1000,
2229 };
2230 put_cmsg(msg, SOL_SOCKET, SO_TIMESTAMP_NEW,
2231 sizeof(stv), &stv);
2232 } else {
2233 struct __kernel_old_timeval tv = {
2234 .tv_sec = tss->ts[0].tv_sec,
2235 .tv_usec = tss->ts[0].tv_nsec / 1000,
2236 };
2237 put_cmsg(msg, SOL_SOCKET, SO_TIMESTAMP_OLD,
2238 sizeof(tv), &tv);
2239 }
2240 }
2241 }
2242
2243 if (sk->sk_tsflags & SOF_TIMESTAMPING_SOFTWARE)
2244 has_timestamping = true;
2245 else
2246 tss->ts[0] = (struct timespec64) {0};
2247 }
2248
2249 if (tss->ts[2].tv_sec || tss->ts[2].tv_nsec) {
2250 if (sk->sk_tsflags & SOF_TIMESTAMPING_RAW_HARDWARE)
2251 has_timestamping = true;
2252 else
2253 tss->ts[2] = (struct timespec64) {0};
2254 }
2255
2256 if (has_timestamping) {
2257 tss->ts[1] = (struct timespec64) {0};
2258 if (sock_flag(sk, SOCK_TSTAMP_NEW))
2259 put_cmsg_scm_timestamping64(msg, tss);
2260 else
2261 put_cmsg_scm_timestamping(msg, tss);
2262 }
2263 }
2264
tcp_inq_hint(struct sock * sk)2265 static int tcp_inq_hint(struct sock *sk)
2266 {
2267 const struct tcp_sock *tp = tcp_sk(sk);
2268 u32 copied_seq = READ_ONCE(tp->copied_seq);
2269 u32 rcv_nxt = READ_ONCE(tp->rcv_nxt);
2270 int inq;
2271
2272 inq = rcv_nxt - copied_seq;
2273 if (unlikely(inq < 0 || copied_seq != READ_ONCE(tp->copied_seq))) {
2274 lock_sock(sk);
2275 inq = tp->rcv_nxt - tp->copied_seq;
2276 release_sock(sk);
2277 }
2278 /* After receiving a FIN, tell the user-space to continue reading
2279 * by returning a non-zero inq.
2280 */
2281 if (inq == 0 && sock_flag(sk, SOCK_DONE))
2282 inq = 1;
2283 return inq;
2284 }
2285
2286 /*
2287 * This routine copies from a sock struct into the user buffer.
2288 *
2289 * Technical note: in 2.3 we work on _locked_ socket, so that
2290 * tricks with *seq access order and skb->users are not required.
2291 * Probably, code can be easily improved even more.
2292 */
2293
tcp_recvmsg_locked(struct sock * sk,struct msghdr * msg,size_t len,int nonblock,int flags,struct scm_timestamping_internal * tss,int * cmsg_flags)2294 static int tcp_recvmsg_locked(struct sock *sk, struct msghdr *msg, size_t len,
2295 int nonblock, int flags,
2296 struct scm_timestamping_internal *tss,
2297 int *cmsg_flags)
2298 {
2299 struct tcp_sock *tp = tcp_sk(sk);
2300 int copied = 0;
2301 u32 peek_seq;
2302 u32 *seq;
2303 unsigned long used;
2304 int err;
2305 int target; /* Read at least this many bytes */
2306 long timeo;
2307 struct sk_buff *skb, *last;
2308 u32 urg_hole = 0;
2309
2310 err = -ENOTCONN;
2311 if (sk->sk_state == TCP_LISTEN)
2312 goto out;
2313
2314 if (tp->recvmsg_inq)
2315 *cmsg_flags = TCP_CMSG_INQ;
2316 timeo = sock_rcvtimeo(sk, nonblock);
2317
2318 /* Urgent data needs to be handled specially. */
2319 if (flags & MSG_OOB)
2320 goto recv_urg;
2321
2322 if (unlikely(tp->repair)) {
2323 err = -EPERM;
2324 if (!(flags & MSG_PEEK))
2325 goto out;
2326
2327 if (tp->repair_queue == TCP_SEND_QUEUE)
2328 goto recv_sndq;
2329
2330 err = -EINVAL;
2331 if (tp->repair_queue == TCP_NO_QUEUE)
2332 goto out;
2333
2334 /* 'common' recv queue MSG_PEEK-ing */
2335 }
2336
2337 seq = &tp->copied_seq;
2338 if (flags & MSG_PEEK) {
2339 peek_seq = tp->copied_seq;
2340 seq = &peek_seq;
2341 }
2342
2343 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
2344
2345 do {
2346 u32 offset;
2347
2348 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
2349 if (tp->urg_data && tp->urg_seq == *seq) {
2350 if (copied)
2351 break;
2352 if (signal_pending(current)) {
2353 copied = timeo ? sock_intr_errno(timeo) : -EAGAIN;
2354 break;
2355 }
2356 }
2357
2358 /* Next get a buffer. */
2359
2360 last = skb_peek_tail(&sk->sk_receive_queue);
2361 skb_queue_walk(&sk->sk_receive_queue, skb) {
2362 last = skb;
2363 /* Now that we have two receive queues this
2364 * shouldn't happen.
2365 */
2366 if (WARN(before(*seq, TCP_SKB_CB(skb)->seq),
2367 "TCP recvmsg seq # bug: copied %X, seq %X, rcvnxt %X, fl %X\n",
2368 *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt,
2369 flags))
2370 break;
2371
2372 offset = *seq - TCP_SKB_CB(skb)->seq;
2373 if (unlikely(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
2374 pr_err_once("%s: found a SYN, please report !\n", __func__);
2375 offset--;
2376 }
2377 if (offset < skb->len)
2378 goto found_ok_skb;
2379 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
2380 goto found_fin_ok;
2381 WARN(!(flags & MSG_PEEK),
2382 "TCP recvmsg seq # bug 2: copied %X, seq %X, rcvnxt %X, fl %X\n",
2383 *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt, flags);
2384 }
2385
2386 /* Well, if we have backlog, try to process it now yet. */
2387
2388 if (copied >= target && !READ_ONCE(sk->sk_backlog.tail))
2389 break;
2390
2391 if (copied) {
2392 if (sk->sk_err ||
2393 sk->sk_state == TCP_CLOSE ||
2394 (sk->sk_shutdown & RCV_SHUTDOWN) ||
2395 !timeo ||
2396 signal_pending(current))
2397 break;
2398 } else {
2399 if (sock_flag(sk, SOCK_DONE))
2400 break;
2401
2402 if (sk->sk_err) {
2403 copied = sock_error(sk);
2404 break;
2405 }
2406
2407 if (sk->sk_shutdown & RCV_SHUTDOWN)
2408 break;
2409
2410 if (sk->sk_state == TCP_CLOSE) {
2411 /* This occurs when user tries to read
2412 * from never connected socket.
2413 */
2414 copied = -ENOTCONN;
2415 break;
2416 }
2417
2418 if (!timeo) {
2419 copied = -EAGAIN;
2420 break;
2421 }
2422
2423 if (signal_pending(current)) {
2424 copied = sock_intr_errno(timeo);
2425 break;
2426 }
2427 }
2428
2429 tcp_cleanup_rbuf(sk, copied);
2430
2431 if (copied >= target) {
2432 /* Do not sleep, just process backlog. */
2433 release_sock(sk);
2434 lock_sock(sk);
2435 } else {
2436 sk_wait_data(sk, &timeo, last);
2437 }
2438
2439 if ((flags & MSG_PEEK) &&
2440 (peek_seq - copied - urg_hole != tp->copied_seq)) {
2441 net_dbg_ratelimited("TCP(%s:%d): Application bug, race in MSG_PEEK\n",
2442 current->comm,
2443 task_pid_nr(current));
2444 peek_seq = tp->copied_seq;
2445 }
2446 continue;
2447
2448 found_ok_skb:
2449 /* Ok so how much can we use? */
2450 used = skb->len - offset;
2451 if (len < used)
2452 used = len;
2453
2454 /* Do we have urgent data here? */
2455 if (tp->urg_data) {
2456 u32 urg_offset = tp->urg_seq - *seq;
2457 if (urg_offset < used) {
2458 if (!urg_offset) {
2459 if (!sock_flag(sk, SOCK_URGINLINE)) {
2460 WRITE_ONCE(*seq, *seq + 1);
2461 urg_hole++;
2462 offset++;
2463 used--;
2464 if (!used)
2465 goto skip_copy;
2466 }
2467 } else
2468 used = urg_offset;
2469 }
2470 }
2471
2472 if (!(flags & MSG_TRUNC)) {
2473 err = skb_copy_datagram_msg(skb, offset, msg, used);
2474 if (err) {
2475 /* Exception. Bailout! */
2476 if (!copied)
2477 copied = -EFAULT;
2478 break;
2479 }
2480 }
2481
2482 WRITE_ONCE(*seq, *seq + used);
2483 copied += used;
2484 len -= used;
2485
2486 tcp_rcv_space_adjust(sk);
2487
2488 skip_copy:
2489 if (tp->urg_data && after(tp->copied_seq, tp->urg_seq)) {
2490 tp->urg_data = 0;
2491 tcp_fast_path_check(sk);
2492 }
2493
2494 if (TCP_SKB_CB(skb)->has_rxtstamp) {
2495 tcp_update_recv_tstamps(skb, tss);
2496 *cmsg_flags |= TCP_CMSG_TS;
2497 }
2498
2499 if (used + offset < skb->len)
2500 continue;
2501
2502 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
2503 goto found_fin_ok;
2504 if (!(flags & MSG_PEEK))
2505 sk_eat_skb(sk, skb);
2506 continue;
2507
2508 found_fin_ok:
2509 /* Process the FIN. */
2510 WRITE_ONCE(*seq, *seq + 1);
2511 if (!(flags & MSG_PEEK))
2512 sk_eat_skb(sk, skb);
2513 break;
2514 } while (len > 0);
2515
2516 /* According to UNIX98, msg_name/msg_namelen are ignored
2517 * on connected socket. I was just happy when found this 8) --ANK
2518 */
2519
2520 /* Clean up data we have read: This will do ACK frames. */
2521 tcp_cleanup_rbuf(sk, copied);
2522 return copied;
2523
2524 out:
2525 return err;
2526
2527 recv_urg:
2528 err = tcp_recv_urg(sk, msg, len, flags);
2529 goto out;
2530
2531 recv_sndq:
2532 err = tcp_peek_sndq(sk, msg, len);
2533 goto out;
2534 }
2535
tcp_recvmsg(struct sock * sk,struct msghdr * msg,size_t len,int nonblock,int flags,int * addr_len)2536 int tcp_recvmsg(struct sock *sk, struct msghdr *msg, size_t len, int nonblock,
2537 int flags, int *addr_len)
2538 {
2539 int cmsg_flags = 0, ret, inq;
2540 struct scm_timestamping_internal tss;
2541
2542 if (unlikely(flags & MSG_ERRQUEUE))
2543 return inet_recv_error(sk, msg, len, addr_len);
2544
2545 if (sk_can_busy_loop(sk) &&
2546 skb_queue_empty_lockless(&sk->sk_receive_queue) &&
2547 sk->sk_state == TCP_ESTABLISHED)
2548 sk_busy_loop(sk, nonblock);
2549
2550 lock_sock(sk);
2551 ret = tcp_recvmsg_locked(sk, msg, len, nonblock, flags, &tss,
2552 &cmsg_flags);
2553 release_sock(sk);
2554
2555 if (cmsg_flags && ret >= 0) {
2556 if (cmsg_flags & TCP_CMSG_TS)
2557 tcp_recv_timestamp(msg, sk, &tss);
2558 if (cmsg_flags & TCP_CMSG_INQ) {
2559 inq = tcp_inq_hint(sk);
2560 put_cmsg(msg, SOL_TCP, TCP_CM_INQ, sizeof(inq), &inq);
2561 }
2562 }
2563 return ret;
2564 }
2565 EXPORT_SYMBOL(tcp_recvmsg);
2566
tcp_set_state(struct sock * sk,int state)2567 void tcp_set_state(struct sock *sk, int state)
2568 {
2569 int oldstate = sk->sk_state;
2570
2571 /* We defined a new enum for TCP states that are exported in BPF
2572 * so as not force the internal TCP states to be frozen. The
2573 * following checks will detect if an internal state value ever
2574 * differs from the BPF value. If this ever happens, then we will
2575 * need to remap the internal value to the BPF value before calling
2576 * tcp_call_bpf_2arg.
2577 */
2578 BUILD_BUG_ON((int)BPF_TCP_ESTABLISHED != (int)TCP_ESTABLISHED);
2579 BUILD_BUG_ON((int)BPF_TCP_SYN_SENT != (int)TCP_SYN_SENT);
2580 BUILD_BUG_ON((int)BPF_TCP_SYN_RECV != (int)TCP_SYN_RECV);
2581 BUILD_BUG_ON((int)BPF_TCP_FIN_WAIT1 != (int)TCP_FIN_WAIT1);
2582 BUILD_BUG_ON((int)BPF_TCP_FIN_WAIT2 != (int)TCP_FIN_WAIT2);
2583 BUILD_BUG_ON((int)BPF_TCP_TIME_WAIT != (int)TCP_TIME_WAIT);
2584 BUILD_BUG_ON((int)BPF_TCP_CLOSE != (int)TCP_CLOSE);
2585 BUILD_BUG_ON((int)BPF_TCP_CLOSE_WAIT != (int)TCP_CLOSE_WAIT);
2586 BUILD_BUG_ON((int)BPF_TCP_LAST_ACK != (int)TCP_LAST_ACK);
2587 BUILD_BUG_ON((int)BPF_TCP_LISTEN != (int)TCP_LISTEN);
2588 BUILD_BUG_ON((int)BPF_TCP_CLOSING != (int)TCP_CLOSING);
2589 BUILD_BUG_ON((int)BPF_TCP_NEW_SYN_RECV != (int)TCP_NEW_SYN_RECV);
2590 BUILD_BUG_ON((int)BPF_TCP_MAX_STATES != (int)TCP_MAX_STATES);
2591
2592 /* bpf uapi header bpf.h defines an anonymous enum with values
2593 * BPF_TCP_* used by bpf programs. Currently gcc built vmlinux
2594 * is able to emit this enum in DWARF due to the above BUILD_BUG_ON.
2595 * But clang built vmlinux does not have this enum in DWARF
2596 * since clang removes the above code before generating IR/debuginfo.
2597 * Let us explicitly emit the type debuginfo to ensure the
2598 * above-mentioned anonymous enum in the vmlinux DWARF and hence BTF
2599 * regardless of which compiler is used.
2600 */
2601 BTF_TYPE_EMIT_ENUM(BPF_TCP_ESTABLISHED);
2602
2603 if (BPF_SOCK_OPS_TEST_FLAG(tcp_sk(sk), BPF_SOCK_OPS_STATE_CB_FLAG))
2604 tcp_call_bpf_2arg(sk, BPF_SOCK_OPS_STATE_CB, oldstate, state);
2605
2606 switch (state) {
2607 case TCP_ESTABLISHED:
2608 if (oldstate != TCP_ESTABLISHED)
2609 TCP_INC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
2610 break;
2611
2612 case TCP_CLOSE:
2613 if (oldstate == TCP_CLOSE_WAIT || oldstate == TCP_ESTABLISHED)
2614 TCP_INC_STATS(sock_net(sk), TCP_MIB_ESTABRESETS);
2615
2616 sk->sk_prot->unhash(sk);
2617 if (inet_csk(sk)->icsk_bind_hash &&
2618 !(sk->sk_userlocks & SOCK_BINDPORT_LOCK))
2619 inet_put_port(sk);
2620 fallthrough;
2621 default:
2622 if (oldstate == TCP_ESTABLISHED)
2623 TCP_DEC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
2624 }
2625
2626 /* Change state AFTER socket is unhashed to avoid closed
2627 * socket sitting in hash tables.
2628 */
2629 inet_sk_state_store(sk, state);
2630 }
2631 EXPORT_SYMBOL_GPL(tcp_set_state);
2632
2633 /*
2634 * State processing on a close. This implements the state shift for
2635 * sending our FIN frame. Note that we only send a FIN for some
2636 * states. A shutdown() may have already sent the FIN, or we may be
2637 * closed.
2638 */
2639
2640 static const unsigned char new_state[16] = {
2641 /* current state: new state: action: */
2642 [0 /* (Invalid) */] = TCP_CLOSE,
2643 [TCP_ESTABLISHED] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2644 [TCP_SYN_SENT] = TCP_CLOSE,
2645 [TCP_SYN_RECV] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2646 [TCP_FIN_WAIT1] = TCP_FIN_WAIT1,
2647 [TCP_FIN_WAIT2] = TCP_FIN_WAIT2,
2648 [TCP_TIME_WAIT] = TCP_CLOSE,
2649 [TCP_CLOSE] = TCP_CLOSE,
2650 [TCP_CLOSE_WAIT] = TCP_LAST_ACK | TCP_ACTION_FIN,
2651 [TCP_LAST_ACK] = TCP_LAST_ACK,
2652 [TCP_LISTEN] = TCP_CLOSE,
2653 [TCP_CLOSING] = TCP_CLOSING,
2654 [TCP_NEW_SYN_RECV] = TCP_CLOSE, /* should not happen ! */
2655 };
2656
tcp_close_state(struct sock * sk)2657 static int tcp_close_state(struct sock *sk)
2658 {
2659 int next = (int)new_state[sk->sk_state];
2660 int ns = next & TCP_STATE_MASK;
2661
2662 tcp_set_state(sk, ns);
2663
2664 return next & TCP_ACTION_FIN;
2665 }
2666
2667 /*
2668 * Shutdown the sending side of a connection. Much like close except
2669 * that we don't receive shut down or sock_set_flag(sk, SOCK_DEAD).
2670 */
2671
tcp_shutdown(struct sock * sk,int how)2672 void tcp_shutdown(struct sock *sk, int how)
2673 {
2674 /* We need to grab some memory, and put together a FIN,
2675 * and then put it into the queue to be sent.
2676 * Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
2677 */
2678 if (!(how & SEND_SHUTDOWN))
2679 return;
2680
2681 /* If we've already sent a FIN, or it's a closed state, skip this. */
2682 if ((1 << sk->sk_state) &
2683 (TCPF_ESTABLISHED | TCPF_SYN_SENT |
2684 TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) {
2685 /* Clear out any half completed packets. FIN if needed. */
2686 if (tcp_close_state(sk))
2687 tcp_send_fin(sk);
2688 }
2689 }
2690 EXPORT_SYMBOL(tcp_shutdown);
2691
tcp_check_oom(struct sock * sk,int shift)2692 bool tcp_check_oom(struct sock *sk, int shift)
2693 {
2694 bool too_many_orphans, out_of_socket_memory;
2695
2696 too_many_orphans = tcp_too_many_orphans(sk, shift);
2697 out_of_socket_memory = tcp_out_of_memory(sk);
2698
2699 if (too_many_orphans)
2700 net_info_ratelimited("too many orphaned sockets\n");
2701 if (out_of_socket_memory)
2702 net_info_ratelimited("out of memory -- consider tuning tcp_mem\n");
2703 return too_many_orphans || out_of_socket_memory;
2704 }
2705
__tcp_close(struct sock * sk,long timeout)2706 void __tcp_close(struct sock *sk, long timeout)
2707 {
2708 struct sk_buff *skb;
2709 int data_was_unread = 0;
2710 int state;
2711
2712 sk->sk_shutdown = SHUTDOWN_MASK;
2713
2714 if (sk->sk_state == TCP_LISTEN) {
2715 tcp_set_state(sk, TCP_CLOSE);
2716
2717 /* Special case. */
2718 inet_csk_listen_stop(sk);
2719
2720 goto adjudge_to_death;
2721 }
2722
2723 /* We need to flush the recv. buffs. We do this only on the
2724 * descriptor close, not protocol-sourced closes, because the
2725 * reader process may not have drained the data yet!
2726 */
2727 while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) {
2728 u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq;
2729
2730 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
2731 len--;
2732 data_was_unread += len;
2733 __kfree_skb(skb);
2734 }
2735
2736 sk_mem_reclaim(sk);
2737
2738 /* If socket has been already reset (e.g. in tcp_reset()) - kill it. */
2739 if (sk->sk_state == TCP_CLOSE)
2740 goto adjudge_to_death;
2741
2742 /* As outlined in RFC 2525, section 2.17, we send a RST here because
2743 * data was lost. To witness the awful effects of the old behavior of
2744 * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk
2745 * GET in an FTP client, suspend the process, wait for the client to
2746 * advertise a zero window, then kill -9 the FTP client, wheee...
2747 * Note: timeout is always zero in such a case.
2748 */
2749 if (unlikely(tcp_sk(sk)->repair)) {
2750 sk->sk_prot->disconnect(sk, 0);
2751 } else if (data_was_unread) {
2752 /* Unread data was tossed, zap the connection. */
2753 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONCLOSE);
2754 tcp_set_state(sk, TCP_CLOSE);
2755 tcp_send_active_reset(sk, sk->sk_allocation);
2756 } else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) {
2757 /* Check zero linger _after_ checking for unread data. */
2758 sk->sk_prot->disconnect(sk, 0);
2759 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
2760 } else if (tcp_close_state(sk)) {
2761 /* We FIN if the application ate all the data before
2762 * zapping the connection.
2763 */
2764
2765 /* RED-PEN. Formally speaking, we have broken TCP state
2766 * machine. State transitions:
2767 *
2768 * TCP_ESTABLISHED -> TCP_FIN_WAIT1
2769 * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible)
2770 * TCP_CLOSE_WAIT -> TCP_LAST_ACK
2771 *
2772 * are legal only when FIN has been sent (i.e. in window),
2773 * rather than queued out of window. Purists blame.
2774 *
2775 * F.e. "RFC state" is ESTABLISHED,
2776 * if Linux state is FIN-WAIT-1, but FIN is still not sent.
2777 *
2778 * The visible declinations are that sometimes
2779 * we enter time-wait state, when it is not required really
2780 * (harmless), do not send active resets, when they are
2781 * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
2782 * they look as CLOSING or LAST_ACK for Linux)
2783 * Probably, I missed some more holelets.
2784 * --ANK
2785 * XXX (TFO) - To start off we don't support SYN+ACK+FIN
2786 * in a single packet! (May consider it later but will
2787 * probably need API support or TCP_CORK SYN-ACK until
2788 * data is written and socket is closed.)
2789 */
2790 tcp_send_fin(sk);
2791 }
2792
2793 sk_stream_wait_close(sk, timeout);
2794
2795 adjudge_to_death:
2796 state = sk->sk_state;
2797 sock_hold(sk);
2798 sock_orphan(sk);
2799
2800 local_bh_disable();
2801 bh_lock_sock(sk);
2802 /* remove backlog if any, without releasing ownership. */
2803 __release_sock(sk);
2804
2805 percpu_counter_inc(sk->sk_prot->orphan_count);
2806
2807 /* Have we already been destroyed by a softirq or backlog? */
2808 if (state != TCP_CLOSE && sk->sk_state == TCP_CLOSE)
2809 goto out;
2810
2811 /* This is a (useful) BSD violating of the RFC. There is a
2812 * problem with TCP as specified in that the other end could
2813 * keep a socket open forever with no application left this end.
2814 * We use a 1 minute timeout (about the same as BSD) then kill
2815 * our end. If they send after that then tough - BUT: long enough
2816 * that we won't make the old 4*rto = almost no time - whoops
2817 * reset mistake.
2818 *
2819 * Nope, it was not mistake. It is really desired behaviour
2820 * f.e. on http servers, when such sockets are useless, but
2821 * consume significant resources. Let's do it with special
2822 * linger2 option. --ANK
2823 */
2824
2825 if (sk->sk_state == TCP_FIN_WAIT2) {
2826 struct tcp_sock *tp = tcp_sk(sk);
2827 if (tp->linger2 < 0) {
2828 tcp_set_state(sk, TCP_CLOSE);
2829 tcp_send_active_reset(sk, GFP_ATOMIC);
2830 __NET_INC_STATS(sock_net(sk),
2831 LINUX_MIB_TCPABORTONLINGER);
2832 } else {
2833 const int tmo = tcp_fin_time(sk);
2834
2835 if (tmo > TCP_TIMEWAIT_LEN) {
2836 inet_csk_reset_keepalive_timer(sk,
2837 tmo - TCP_TIMEWAIT_LEN);
2838 } else {
2839 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
2840 goto out;
2841 }
2842 }
2843 }
2844 if (sk->sk_state != TCP_CLOSE) {
2845 sk_mem_reclaim(sk);
2846 if (tcp_check_oom(sk, 0)) {
2847 tcp_set_state(sk, TCP_CLOSE);
2848 tcp_send_active_reset(sk, GFP_ATOMIC);
2849 __NET_INC_STATS(sock_net(sk),
2850 LINUX_MIB_TCPABORTONMEMORY);
2851 } else if (!check_net(sock_net(sk))) {
2852 /* Not possible to send reset; just close */
2853 tcp_set_state(sk, TCP_CLOSE);
2854 }
2855 }
2856
2857 if (sk->sk_state == TCP_CLOSE) {
2858 struct request_sock *req;
2859
2860 req = rcu_dereference_protected(tcp_sk(sk)->fastopen_rsk,
2861 lockdep_sock_is_held(sk));
2862 /* We could get here with a non-NULL req if the socket is
2863 * aborted (e.g., closed with unread data) before 3WHS
2864 * finishes.
2865 */
2866 if (req)
2867 reqsk_fastopen_remove(sk, req, false);
2868 inet_csk_destroy_sock(sk);
2869 }
2870 /* Otherwise, socket is reprieved until protocol close. */
2871
2872 out:
2873 bh_unlock_sock(sk);
2874 local_bh_enable();
2875 }
2876
tcp_close(struct sock * sk,long timeout)2877 void tcp_close(struct sock *sk, long timeout)
2878 {
2879 lock_sock(sk);
2880 __tcp_close(sk, timeout);
2881 release_sock(sk);
2882 sock_put(sk);
2883 }
2884 EXPORT_SYMBOL(tcp_close);
2885
2886 /* These states need RST on ABORT according to RFC793 */
2887
tcp_need_reset(int state)2888 static inline bool tcp_need_reset(int state)
2889 {
2890 return (1 << state) &
2891 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 |
2892 TCPF_FIN_WAIT2 | TCPF_SYN_RECV);
2893 }
2894
tcp_rtx_queue_purge(struct sock * sk)2895 static void tcp_rtx_queue_purge(struct sock *sk)
2896 {
2897 struct rb_node *p = rb_first(&sk->tcp_rtx_queue);
2898
2899 tcp_sk(sk)->highest_sack = NULL;
2900 while (p) {
2901 struct sk_buff *skb = rb_to_skb(p);
2902
2903 p = rb_next(p);
2904 /* Since we are deleting whole queue, no need to
2905 * list_del(&skb->tcp_tsorted_anchor)
2906 */
2907 tcp_rtx_queue_unlink(skb, sk);
2908 sk_wmem_free_skb(sk, skb);
2909 }
2910 }
2911
tcp_write_queue_purge(struct sock * sk)2912 void tcp_write_queue_purge(struct sock *sk)
2913 {
2914 struct sk_buff *skb;
2915
2916 tcp_chrono_stop(sk, TCP_CHRONO_BUSY);
2917 while ((skb = __skb_dequeue(&sk->sk_write_queue)) != NULL) {
2918 tcp_skb_tsorted_anchor_cleanup(skb);
2919 sk_wmem_free_skb(sk, skb);
2920 }
2921 tcp_rtx_queue_purge(sk);
2922 skb = sk->sk_tx_skb_cache;
2923 if (skb) {
2924 __kfree_skb(skb);
2925 sk->sk_tx_skb_cache = NULL;
2926 }
2927 INIT_LIST_HEAD(&tcp_sk(sk)->tsorted_sent_queue);
2928 sk_mem_reclaim(sk);
2929 tcp_clear_all_retrans_hints(tcp_sk(sk));
2930 tcp_sk(sk)->packets_out = 0;
2931 inet_csk(sk)->icsk_backoff = 0;
2932 }
2933
tcp_disconnect(struct sock * sk,int flags)2934 int tcp_disconnect(struct sock *sk, int flags)
2935 {
2936 struct inet_sock *inet = inet_sk(sk);
2937 struct inet_connection_sock *icsk = inet_csk(sk);
2938 struct tcp_sock *tp = tcp_sk(sk);
2939 int old_state = sk->sk_state;
2940 u32 seq;
2941
2942 if (old_state != TCP_CLOSE)
2943 tcp_set_state(sk, TCP_CLOSE);
2944
2945 /* ABORT function of RFC793 */
2946 if (old_state == TCP_LISTEN) {
2947 inet_csk_listen_stop(sk);
2948 } else if (unlikely(tp->repair)) {
2949 sk->sk_err = ECONNABORTED;
2950 } else if (tcp_need_reset(old_state) ||
2951 (tp->snd_nxt != tp->write_seq &&
2952 (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK))) {
2953 /* The last check adjusts for discrepancy of Linux wrt. RFC
2954 * states
2955 */
2956 tcp_send_active_reset(sk, gfp_any());
2957 sk->sk_err = ECONNRESET;
2958 } else if (old_state == TCP_SYN_SENT)
2959 sk->sk_err = ECONNRESET;
2960
2961 tcp_clear_xmit_timers(sk);
2962 __skb_queue_purge(&sk->sk_receive_queue);
2963 if (sk->sk_rx_skb_cache) {
2964 __kfree_skb(sk->sk_rx_skb_cache);
2965 sk->sk_rx_skb_cache = NULL;
2966 }
2967 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt);
2968 tp->urg_data = 0;
2969 tcp_write_queue_purge(sk);
2970 tcp_fastopen_active_disable_ofo_check(sk);
2971 skb_rbtree_purge(&tp->out_of_order_queue);
2972
2973 inet->inet_dport = 0;
2974
2975 if (!(sk->sk_userlocks & SOCK_BINDADDR_LOCK))
2976 inet_reset_saddr(sk);
2977
2978 sk->sk_shutdown = 0;
2979 sock_reset_flag(sk, SOCK_DONE);
2980 tp->srtt_us = 0;
2981 tp->mdev_us = jiffies_to_usecs(TCP_TIMEOUT_INIT);
2982 tp->rcv_rtt_last_tsecr = 0;
2983
2984 seq = tp->write_seq + tp->max_window + 2;
2985 if (!seq)
2986 seq = 1;
2987 WRITE_ONCE(tp->write_seq, seq);
2988
2989 icsk->icsk_backoff = 0;
2990 icsk->icsk_probes_out = 0;
2991 icsk->icsk_probes_tstamp = 0;
2992 icsk->icsk_rto = TCP_TIMEOUT_INIT;
2993 icsk->icsk_rto_min = TCP_RTO_MIN;
2994 icsk->icsk_delack_max = TCP_DELACK_MAX;
2995 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
2996 tp->snd_cwnd = TCP_INIT_CWND;
2997 tp->snd_cwnd_cnt = 0;
2998 tp->window_clamp = 0;
2999 tp->delivered = 0;
3000 tp->delivered_ce = 0;
3001 if (icsk->icsk_ca_ops->release)
3002 icsk->icsk_ca_ops->release(sk);
3003 memset(icsk->icsk_ca_priv, 0, sizeof(icsk->icsk_ca_priv));
3004 icsk->icsk_ca_initialized = 0;
3005 tcp_set_ca_state(sk, TCP_CA_Open);
3006 tp->is_sack_reneg = 0;
3007 tcp_clear_retrans(tp);
3008 tp->total_retrans = 0;
3009 inet_csk_delack_init(sk);
3010 /* Initialize rcv_mss to TCP_MIN_MSS to avoid division by 0
3011 * issue in __tcp_select_window()
3012 */
3013 icsk->icsk_ack.rcv_mss = TCP_MIN_MSS;
3014 memset(&tp->rx_opt, 0, sizeof(tp->rx_opt));
3015 __sk_dst_reset(sk);
3016 dst_release(sk->sk_rx_dst);
3017 sk->sk_rx_dst = NULL;
3018 tcp_saved_syn_free(tp);
3019 tp->compressed_ack = 0;
3020 tp->segs_in = 0;
3021 tp->segs_out = 0;
3022 tp->bytes_sent = 0;
3023 tp->bytes_acked = 0;
3024 tp->bytes_received = 0;
3025 tp->bytes_retrans = 0;
3026 tp->data_segs_in = 0;
3027 tp->data_segs_out = 0;
3028 tp->duplicate_sack[0].start_seq = 0;
3029 tp->duplicate_sack[0].end_seq = 0;
3030 tp->dsack_dups = 0;
3031 tp->reord_seen = 0;
3032 tp->retrans_out = 0;
3033 tp->sacked_out = 0;
3034 tp->tlp_high_seq = 0;
3035 tp->last_oow_ack_time = 0;
3036 /* There's a bubble in the pipe until at least the first ACK. */
3037 tp->app_limited = ~0U;
3038 tp->rack.mstamp = 0;
3039 tp->rack.advanced = 0;
3040 tp->rack.reo_wnd_steps = 1;
3041 tp->rack.last_delivered = 0;
3042 tp->rack.reo_wnd_persist = 0;
3043 tp->rack.dsack_seen = 0;
3044 tp->syn_data_acked = 0;
3045 tp->rx_opt.saw_tstamp = 0;
3046 tp->rx_opt.dsack = 0;
3047 tp->rx_opt.num_sacks = 0;
3048 tp->rcv_ooopack = 0;
3049
3050
3051 /* Clean up fastopen related fields */
3052 tcp_free_fastopen_req(tp);
3053 inet->defer_connect = 0;
3054 tp->fastopen_client_fail = 0;
3055
3056 WARN_ON(inet->inet_num && !icsk->icsk_bind_hash);
3057
3058 if (sk->sk_frag.page) {
3059 put_page(sk->sk_frag.page);
3060 sk->sk_frag.page = NULL;
3061 sk->sk_frag.offset = 0;
3062 }
3063
3064 sk->sk_error_report(sk);
3065 return 0;
3066 }
3067 EXPORT_SYMBOL(tcp_disconnect);
3068
tcp_can_repair_sock(const struct sock * sk)3069 static inline bool tcp_can_repair_sock(const struct sock *sk)
3070 {
3071 return ns_capable(sock_net(sk)->user_ns, CAP_NET_ADMIN) &&
3072 (sk->sk_state != TCP_LISTEN);
3073 }
3074
tcp_repair_set_window(struct tcp_sock * tp,sockptr_t optbuf,int len)3075 static int tcp_repair_set_window(struct tcp_sock *tp, sockptr_t optbuf, int len)
3076 {
3077 struct tcp_repair_window opt;
3078
3079 if (!tp->repair)
3080 return -EPERM;
3081
3082 if (len != sizeof(opt))
3083 return -EINVAL;
3084
3085 if (copy_from_sockptr(&opt, optbuf, sizeof(opt)))
3086 return -EFAULT;
3087
3088 if (opt.max_window < opt.snd_wnd)
3089 return -EINVAL;
3090
3091 if (after(opt.snd_wl1, tp->rcv_nxt + opt.rcv_wnd))
3092 return -EINVAL;
3093
3094 if (after(opt.rcv_wup, tp->rcv_nxt))
3095 return -EINVAL;
3096
3097 tp->snd_wl1 = opt.snd_wl1;
3098 tp->snd_wnd = opt.snd_wnd;
3099 tp->max_window = opt.max_window;
3100
3101 tp->rcv_wnd = opt.rcv_wnd;
3102 tp->rcv_wup = opt.rcv_wup;
3103
3104 return 0;
3105 }
3106
tcp_repair_options_est(struct sock * sk,sockptr_t optbuf,unsigned int len)3107 static int tcp_repair_options_est(struct sock *sk, sockptr_t optbuf,
3108 unsigned int len)
3109 {
3110 struct tcp_sock *tp = tcp_sk(sk);
3111 struct tcp_repair_opt opt;
3112 size_t offset = 0;
3113
3114 while (len >= sizeof(opt)) {
3115 if (copy_from_sockptr_offset(&opt, optbuf, offset, sizeof(opt)))
3116 return -EFAULT;
3117
3118 offset += sizeof(opt);
3119 len -= sizeof(opt);
3120
3121 switch (opt.opt_code) {
3122 case TCPOPT_MSS:
3123 tp->rx_opt.mss_clamp = opt.opt_val;
3124 tcp_mtup_init(sk);
3125 break;
3126 case TCPOPT_WINDOW:
3127 {
3128 u16 snd_wscale = opt.opt_val & 0xFFFF;
3129 u16 rcv_wscale = opt.opt_val >> 16;
3130
3131 if (snd_wscale > TCP_MAX_WSCALE || rcv_wscale > TCP_MAX_WSCALE)
3132 return -EFBIG;
3133
3134 tp->rx_opt.snd_wscale = snd_wscale;
3135 tp->rx_opt.rcv_wscale = rcv_wscale;
3136 tp->rx_opt.wscale_ok = 1;
3137 }
3138 break;
3139 case TCPOPT_SACK_PERM:
3140 if (opt.opt_val != 0)
3141 return -EINVAL;
3142
3143 tp->rx_opt.sack_ok |= TCP_SACK_SEEN;
3144 break;
3145 case TCPOPT_TIMESTAMP:
3146 if (opt.opt_val != 0)
3147 return -EINVAL;
3148
3149 tp->rx_opt.tstamp_ok = 1;
3150 break;
3151 }
3152 }
3153
3154 return 0;
3155 }
3156
3157 DEFINE_STATIC_KEY_FALSE(tcp_tx_delay_enabled);
3158 EXPORT_SYMBOL(tcp_tx_delay_enabled);
3159
tcp_enable_tx_delay(void)3160 static void tcp_enable_tx_delay(void)
3161 {
3162 if (!static_branch_unlikely(&tcp_tx_delay_enabled)) {
3163 static int __tcp_tx_delay_enabled = 0;
3164
3165 if (cmpxchg(&__tcp_tx_delay_enabled, 0, 1) == 0) {
3166 static_branch_enable(&tcp_tx_delay_enabled);
3167 pr_info("TCP_TX_DELAY enabled\n");
3168 }
3169 }
3170 }
3171
3172 /* When set indicates to always queue non-full frames. Later the user clears
3173 * this option and we transmit any pending partial frames in the queue. This is
3174 * meant to be used alongside sendfile() to get properly filled frames when the
3175 * user (for example) must write out headers with a write() call first and then
3176 * use sendfile to send out the data parts.
3177 *
3178 * TCP_CORK can be set together with TCP_NODELAY and it is stronger than
3179 * TCP_NODELAY.
3180 */
__tcp_sock_set_cork(struct sock * sk,bool on)3181 static void __tcp_sock_set_cork(struct sock *sk, bool on)
3182 {
3183 struct tcp_sock *tp = tcp_sk(sk);
3184
3185 if (on) {
3186 tp->nonagle |= TCP_NAGLE_CORK;
3187 } else {
3188 tp->nonagle &= ~TCP_NAGLE_CORK;
3189 if (tp->nonagle & TCP_NAGLE_OFF)
3190 tp->nonagle |= TCP_NAGLE_PUSH;
3191 tcp_push_pending_frames(sk);
3192 }
3193 }
3194
tcp_sock_set_cork(struct sock * sk,bool on)3195 void tcp_sock_set_cork(struct sock *sk, bool on)
3196 {
3197 lock_sock(sk);
3198 __tcp_sock_set_cork(sk, on);
3199 release_sock(sk);
3200 }
3201 EXPORT_SYMBOL(tcp_sock_set_cork);
3202
3203 /* TCP_NODELAY is weaker than TCP_CORK, so that this option on corked socket is
3204 * remembered, but it is not activated until cork is cleared.
3205 *
3206 * However, when TCP_NODELAY is set we make an explicit push, which overrides
3207 * even TCP_CORK for currently queued segments.
3208 */
__tcp_sock_set_nodelay(struct sock * sk,bool on)3209 static void __tcp_sock_set_nodelay(struct sock *sk, bool on)
3210 {
3211 if (on) {
3212 tcp_sk(sk)->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH;
3213 tcp_push_pending_frames(sk);
3214 } else {
3215 tcp_sk(sk)->nonagle &= ~TCP_NAGLE_OFF;
3216 }
3217 }
3218
tcp_sock_set_nodelay(struct sock * sk)3219 void tcp_sock_set_nodelay(struct sock *sk)
3220 {
3221 lock_sock(sk);
3222 __tcp_sock_set_nodelay(sk, true);
3223 release_sock(sk);
3224 }
3225 EXPORT_SYMBOL(tcp_sock_set_nodelay);
3226
__tcp_sock_set_quickack(struct sock * sk,int val)3227 static void __tcp_sock_set_quickack(struct sock *sk, int val)
3228 {
3229 if (!val) {
3230 inet_csk_enter_pingpong_mode(sk);
3231 return;
3232 }
3233
3234 inet_csk_exit_pingpong_mode(sk);
3235 if ((1 << sk->sk_state) & (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) &&
3236 inet_csk_ack_scheduled(sk)) {
3237 inet_csk(sk)->icsk_ack.pending |= ICSK_ACK_PUSHED;
3238 tcp_cleanup_rbuf(sk, 1);
3239 if (!(val & 1))
3240 inet_csk_enter_pingpong_mode(sk);
3241 }
3242 }
3243
tcp_sock_set_quickack(struct sock * sk,int val)3244 void tcp_sock_set_quickack(struct sock *sk, int val)
3245 {
3246 lock_sock(sk);
3247 __tcp_sock_set_quickack(sk, val);
3248 release_sock(sk);
3249 }
3250 EXPORT_SYMBOL(tcp_sock_set_quickack);
3251
tcp_sock_set_syncnt(struct sock * sk,int val)3252 int tcp_sock_set_syncnt(struct sock *sk, int val)
3253 {
3254 if (val < 1 || val > MAX_TCP_SYNCNT)
3255 return -EINVAL;
3256
3257 lock_sock(sk);
3258 inet_csk(sk)->icsk_syn_retries = val;
3259 release_sock(sk);
3260 return 0;
3261 }
3262 EXPORT_SYMBOL(tcp_sock_set_syncnt);
3263
tcp_sock_set_user_timeout(struct sock * sk,u32 val)3264 void tcp_sock_set_user_timeout(struct sock *sk, u32 val)
3265 {
3266 lock_sock(sk);
3267 inet_csk(sk)->icsk_user_timeout = val;
3268 release_sock(sk);
3269 }
3270 EXPORT_SYMBOL(tcp_sock_set_user_timeout);
3271
tcp_sock_set_keepidle_locked(struct sock * sk,int val)3272 int tcp_sock_set_keepidle_locked(struct sock *sk, int val)
3273 {
3274 struct tcp_sock *tp = tcp_sk(sk);
3275
3276 if (val < 1 || val > MAX_TCP_KEEPIDLE)
3277 return -EINVAL;
3278
3279 tp->keepalive_time = val * HZ;
3280 if (sock_flag(sk, SOCK_KEEPOPEN) &&
3281 !((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN))) {
3282 u32 elapsed = keepalive_time_elapsed(tp);
3283
3284 if (tp->keepalive_time > elapsed)
3285 elapsed = tp->keepalive_time - elapsed;
3286 else
3287 elapsed = 0;
3288 inet_csk_reset_keepalive_timer(sk, elapsed);
3289 }
3290
3291 return 0;
3292 }
3293
tcp_sock_set_keepidle(struct sock * sk,int val)3294 int tcp_sock_set_keepidle(struct sock *sk, int val)
3295 {
3296 int err;
3297
3298 lock_sock(sk);
3299 err = tcp_sock_set_keepidle_locked(sk, val);
3300 release_sock(sk);
3301 return err;
3302 }
3303 EXPORT_SYMBOL(tcp_sock_set_keepidle);
3304
tcp_sock_set_keepintvl(struct sock * sk,int val)3305 int tcp_sock_set_keepintvl(struct sock *sk, int val)
3306 {
3307 if (val < 1 || val > MAX_TCP_KEEPINTVL)
3308 return -EINVAL;
3309
3310 lock_sock(sk);
3311 tcp_sk(sk)->keepalive_intvl = val * HZ;
3312 release_sock(sk);
3313 return 0;
3314 }
3315 EXPORT_SYMBOL(tcp_sock_set_keepintvl);
3316
tcp_sock_set_keepcnt(struct sock * sk,int val)3317 int tcp_sock_set_keepcnt(struct sock *sk, int val)
3318 {
3319 if (val < 1 || val > MAX_TCP_KEEPCNT)
3320 return -EINVAL;
3321
3322 lock_sock(sk);
3323 tcp_sk(sk)->keepalive_probes = val;
3324 release_sock(sk);
3325 return 0;
3326 }
3327 EXPORT_SYMBOL(tcp_sock_set_keepcnt);
3328
tcp_set_window_clamp(struct sock * sk,int val)3329 int tcp_set_window_clamp(struct sock *sk, int val)
3330 {
3331 struct tcp_sock *tp = tcp_sk(sk);
3332
3333 if (!val) {
3334 if (sk->sk_state != TCP_CLOSE)
3335 return -EINVAL;
3336 tp->window_clamp = 0;
3337 } else {
3338 tp->window_clamp = val < SOCK_MIN_RCVBUF / 2 ?
3339 SOCK_MIN_RCVBUF / 2 : val;
3340 }
3341 return 0;
3342 }
3343
3344 /*
3345 * Socket option code for TCP.
3346 */
do_tcp_setsockopt(struct sock * sk,int level,int optname,sockptr_t optval,unsigned int optlen)3347 static int do_tcp_setsockopt(struct sock *sk, int level, int optname,
3348 sockptr_t optval, unsigned int optlen)
3349 {
3350 struct tcp_sock *tp = tcp_sk(sk);
3351 struct inet_connection_sock *icsk = inet_csk(sk);
3352 struct net *net = sock_net(sk);
3353 int val;
3354 int err = 0;
3355
3356 /* These are data/string values, all the others are ints */
3357 switch (optname) {
3358 case TCP_CONGESTION: {
3359 char name[TCP_CA_NAME_MAX];
3360
3361 if (optlen < 1)
3362 return -EINVAL;
3363
3364 val = strncpy_from_sockptr(name, optval,
3365 min_t(long, TCP_CA_NAME_MAX-1, optlen));
3366 if (val < 0)
3367 return -EFAULT;
3368 name[val] = 0;
3369
3370 lock_sock(sk);
3371 err = tcp_set_congestion_control(sk, name, true,
3372 ns_capable(sock_net(sk)->user_ns,
3373 CAP_NET_ADMIN));
3374 release_sock(sk);
3375 return err;
3376 }
3377 case TCP_ULP: {
3378 char name[TCP_ULP_NAME_MAX];
3379
3380 if (optlen < 1)
3381 return -EINVAL;
3382
3383 val = strncpy_from_sockptr(name, optval,
3384 min_t(long, TCP_ULP_NAME_MAX - 1,
3385 optlen));
3386 if (val < 0)
3387 return -EFAULT;
3388 name[val] = 0;
3389
3390 lock_sock(sk);
3391 err = tcp_set_ulp(sk, name);
3392 release_sock(sk);
3393 return err;
3394 }
3395 case TCP_FASTOPEN_KEY: {
3396 __u8 key[TCP_FASTOPEN_KEY_BUF_LENGTH];
3397 __u8 *backup_key = NULL;
3398
3399 /* Allow a backup key as well to facilitate key rotation
3400 * First key is the active one.
3401 */
3402 if (optlen != TCP_FASTOPEN_KEY_LENGTH &&
3403 optlen != TCP_FASTOPEN_KEY_BUF_LENGTH)
3404 return -EINVAL;
3405
3406 if (copy_from_sockptr(key, optval, optlen))
3407 return -EFAULT;
3408
3409 if (optlen == TCP_FASTOPEN_KEY_BUF_LENGTH)
3410 backup_key = key + TCP_FASTOPEN_KEY_LENGTH;
3411
3412 return tcp_fastopen_reset_cipher(net, sk, key, backup_key);
3413 }
3414 default:
3415 /* fallthru */
3416 break;
3417 }
3418
3419 if (optlen < sizeof(int))
3420 return -EINVAL;
3421
3422 if (copy_from_sockptr(&val, optval, sizeof(val)))
3423 return -EFAULT;
3424
3425 lock_sock(sk);
3426
3427 switch (optname) {
3428 case TCP_MAXSEG:
3429 /* Values greater than interface MTU won't take effect. However
3430 * at the point when this call is done we typically don't yet
3431 * know which interface is going to be used
3432 */
3433 if (val && (val < TCP_MIN_MSS || val > MAX_TCP_WINDOW)) {
3434 err = -EINVAL;
3435 break;
3436 }
3437 tp->rx_opt.user_mss = val;
3438 break;
3439
3440 case TCP_NODELAY:
3441 __tcp_sock_set_nodelay(sk, val);
3442 break;
3443
3444 case TCP_THIN_LINEAR_TIMEOUTS:
3445 if (val < 0 || val > 1)
3446 err = -EINVAL;
3447 else
3448 tp->thin_lto = val;
3449 break;
3450
3451 case TCP_THIN_DUPACK:
3452 if (val < 0 || val > 1)
3453 err = -EINVAL;
3454 break;
3455
3456 case TCP_REPAIR:
3457 if (!tcp_can_repair_sock(sk))
3458 err = -EPERM;
3459 else if (val == TCP_REPAIR_ON) {
3460 tp->repair = 1;
3461 sk->sk_reuse = SK_FORCE_REUSE;
3462 tp->repair_queue = TCP_NO_QUEUE;
3463 } else if (val == TCP_REPAIR_OFF) {
3464 tp->repair = 0;
3465 sk->sk_reuse = SK_NO_REUSE;
3466 tcp_send_window_probe(sk);
3467 } else if (val == TCP_REPAIR_OFF_NO_WP) {
3468 tp->repair = 0;
3469 sk->sk_reuse = SK_NO_REUSE;
3470 } else
3471 err = -EINVAL;
3472
3473 break;
3474
3475 case TCP_REPAIR_QUEUE:
3476 if (!tp->repair)
3477 err = -EPERM;
3478 else if ((unsigned int)val < TCP_QUEUES_NR)
3479 tp->repair_queue = val;
3480 else
3481 err = -EINVAL;
3482 break;
3483
3484 case TCP_QUEUE_SEQ:
3485 if (sk->sk_state != TCP_CLOSE) {
3486 err = -EPERM;
3487 } else if (tp->repair_queue == TCP_SEND_QUEUE) {
3488 if (!tcp_rtx_queue_empty(sk))
3489 err = -EPERM;
3490 else
3491 WRITE_ONCE(tp->write_seq, val);
3492 } else if (tp->repair_queue == TCP_RECV_QUEUE) {
3493 if (tp->rcv_nxt != tp->copied_seq) {
3494 err = -EPERM;
3495 } else {
3496 WRITE_ONCE(tp->rcv_nxt, val);
3497 WRITE_ONCE(tp->copied_seq, val);
3498 }
3499 } else {
3500 err = -EINVAL;
3501 }
3502 break;
3503
3504 case TCP_REPAIR_OPTIONS:
3505 if (!tp->repair)
3506 err = -EINVAL;
3507 else if (sk->sk_state == TCP_ESTABLISHED)
3508 err = tcp_repair_options_est(sk, optval, optlen);
3509 else
3510 err = -EPERM;
3511 break;
3512
3513 case TCP_CORK:
3514 __tcp_sock_set_cork(sk, val);
3515 break;
3516
3517 case TCP_KEEPIDLE:
3518 err = tcp_sock_set_keepidle_locked(sk, val);
3519 break;
3520 case TCP_KEEPINTVL:
3521 if (val < 1 || val > MAX_TCP_KEEPINTVL)
3522 err = -EINVAL;
3523 else
3524 tp->keepalive_intvl = val * HZ;
3525 break;
3526 case TCP_KEEPCNT:
3527 if (val < 1 || val > MAX_TCP_KEEPCNT)
3528 err = -EINVAL;
3529 else
3530 tp->keepalive_probes = val;
3531 break;
3532 case TCP_SYNCNT:
3533 if (val < 1 || val > MAX_TCP_SYNCNT)
3534 err = -EINVAL;
3535 else
3536 icsk->icsk_syn_retries = val;
3537 break;
3538
3539 case TCP_SAVE_SYN:
3540 /* 0: disable, 1: enable, 2: start from ether_header */
3541 if (val < 0 || val > 2)
3542 err = -EINVAL;
3543 else
3544 tp->save_syn = val;
3545 break;
3546
3547 case TCP_LINGER2:
3548 if (val < 0)
3549 tp->linger2 = -1;
3550 else if (val > TCP_FIN_TIMEOUT_MAX / HZ)
3551 tp->linger2 = TCP_FIN_TIMEOUT_MAX;
3552 else
3553 tp->linger2 = val * HZ;
3554 break;
3555
3556 case TCP_DEFER_ACCEPT:
3557 /* Translate value in seconds to number of retransmits */
3558 icsk->icsk_accept_queue.rskq_defer_accept =
3559 secs_to_retrans(val, TCP_TIMEOUT_INIT / HZ,
3560 TCP_RTO_MAX / HZ);
3561 break;
3562
3563 case TCP_WINDOW_CLAMP:
3564 err = tcp_set_window_clamp(sk, val);
3565 break;
3566
3567 case TCP_QUICKACK:
3568 __tcp_sock_set_quickack(sk, val);
3569 break;
3570
3571 #ifdef CONFIG_TCP_MD5SIG
3572 case TCP_MD5SIG:
3573 case TCP_MD5SIG_EXT:
3574 err = tp->af_specific->md5_parse(sk, optname, optval, optlen);
3575 break;
3576 #endif
3577 case TCP_USER_TIMEOUT:
3578 /* Cap the max time in ms TCP will retry or probe the window
3579 * before giving up and aborting (ETIMEDOUT) a connection.
3580 */
3581 if (val < 0)
3582 err = -EINVAL;
3583 else
3584 icsk->icsk_user_timeout = val;
3585 break;
3586
3587 case TCP_FASTOPEN:
3588 if (val >= 0 && ((1 << sk->sk_state) & (TCPF_CLOSE |
3589 TCPF_LISTEN))) {
3590 tcp_fastopen_init_key_once(net);
3591
3592 fastopen_queue_tune(sk, val);
3593 } else {
3594 err = -EINVAL;
3595 }
3596 break;
3597 case TCP_FASTOPEN_CONNECT:
3598 if (val > 1 || val < 0) {
3599 err = -EINVAL;
3600 } else if (net->ipv4.sysctl_tcp_fastopen & TFO_CLIENT_ENABLE) {
3601 if (sk->sk_state == TCP_CLOSE)
3602 tp->fastopen_connect = val;
3603 else
3604 err = -EINVAL;
3605 } else {
3606 err = -EOPNOTSUPP;
3607 }
3608 break;
3609 case TCP_FASTOPEN_NO_COOKIE:
3610 if (val > 1 || val < 0)
3611 err = -EINVAL;
3612 else if (!((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
3613 err = -EINVAL;
3614 else
3615 tp->fastopen_no_cookie = val;
3616 break;
3617 case TCP_TIMESTAMP:
3618 if (!tp->repair)
3619 err = -EPERM;
3620 else
3621 tp->tsoffset = val - tcp_time_stamp_raw();
3622 break;
3623 case TCP_REPAIR_WINDOW:
3624 err = tcp_repair_set_window(tp, optval, optlen);
3625 break;
3626 case TCP_NOTSENT_LOWAT:
3627 tp->notsent_lowat = val;
3628 sk->sk_write_space(sk);
3629 break;
3630 case TCP_INQ:
3631 if (val > 1 || val < 0)
3632 err = -EINVAL;
3633 else
3634 tp->recvmsg_inq = val;
3635 break;
3636 case TCP_TX_DELAY:
3637 if (val)
3638 tcp_enable_tx_delay();
3639 tp->tcp_tx_delay = val;
3640 break;
3641 default:
3642 err = -ENOPROTOOPT;
3643 break;
3644 }
3645
3646 release_sock(sk);
3647 return err;
3648 }
3649
tcp_setsockopt(struct sock * sk,int level,int optname,sockptr_t optval,unsigned int optlen)3650 int tcp_setsockopt(struct sock *sk, int level, int optname, sockptr_t optval,
3651 unsigned int optlen)
3652 {
3653 const struct inet_connection_sock *icsk = inet_csk(sk);
3654
3655 if (level != SOL_TCP)
3656 return icsk->icsk_af_ops->setsockopt(sk, level, optname,
3657 optval, optlen);
3658 return do_tcp_setsockopt(sk, level, optname, optval, optlen);
3659 }
3660 EXPORT_SYMBOL(tcp_setsockopt);
3661
tcp_get_info_chrono_stats(const struct tcp_sock * tp,struct tcp_info * info)3662 static void tcp_get_info_chrono_stats(const struct tcp_sock *tp,
3663 struct tcp_info *info)
3664 {
3665 u64 stats[__TCP_CHRONO_MAX], total = 0;
3666 enum tcp_chrono i;
3667
3668 for (i = TCP_CHRONO_BUSY; i < __TCP_CHRONO_MAX; ++i) {
3669 stats[i] = tp->chrono_stat[i - 1];
3670 if (i == tp->chrono_type)
3671 stats[i] += tcp_jiffies32 - tp->chrono_start;
3672 stats[i] *= USEC_PER_SEC / HZ;
3673 total += stats[i];
3674 }
3675
3676 info->tcpi_busy_time = total;
3677 info->tcpi_rwnd_limited = stats[TCP_CHRONO_RWND_LIMITED];
3678 info->tcpi_sndbuf_limited = stats[TCP_CHRONO_SNDBUF_LIMITED];
3679 }
3680
3681 /* Return information about state of tcp endpoint in API format. */
tcp_get_info(struct sock * sk,struct tcp_info * info)3682 void tcp_get_info(struct sock *sk, struct tcp_info *info)
3683 {
3684 const struct tcp_sock *tp = tcp_sk(sk); /* iff sk_type == SOCK_STREAM */
3685 const struct inet_connection_sock *icsk = inet_csk(sk);
3686 unsigned long rate;
3687 u32 now;
3688 u64 rate64;
3689 bool slow;
3690
3691 memset(info, 0, sizeof(*info));
3692 if (sk->sk_type != SOCK_STREAM)
3693 return;
3694
3695 info->tcpi_state = inet_sk_state_load(sk);
3696
3697 /* Report meaningful fields for all TCP states, including listeners */
3698 rate = READ_ONCE(sk->sk_pacing_rate);
3699 rate64 = (rate != ~0UL) ? rate : ~0ULL;
3700 info->tcpi_pacing_rate = rate64;
3701
3702 rate = READ_ONCE(sk->sk_max_pacing_rate);
3703 rate64 = (rate != ~0UL) ? rate : ~0ULL;
3704 info->tcpi_max_pacing_rate = rate64;
3705
3706 info->tcpi_reordering = tp->reordering;
3707 info->tcpi_snd_cwnd = tp->snd_cwnd;
3708
3709 if (info->tcpi_state == TCP_LISTEN) {
3710 /* listeners aliased fields :
3711 * tcpi_unacked -> Number of children ready for accept()
3712 * tcpi_sacked -> max backlog
3713 */
3714 info->tcpi_unacked = READ_ONCE(sk->sk_ack_backlog);
3715 info->tcpi_sacked = READ_ONCE(sk->sk_max_ack_backlog);
3716 return;
3717 }
3718
3719 slow = lock_sock_fast(sk);
3720
3721 info->tcpi_ca_state = icsk->icsk_ca_state;
3722 info->tcpi_retransmits = icsk->icsk_retransmits;
3723 info->tcpi_probes = icsk->icsk_probes_out;
3724 info->tcpi_backoff = icsk->icsk_backoff;
3725
3726 if (tp->rx_opt.tstamp_ok)
3727 info->tcpi_options |= TCPI_OPT_TIMESTAMPS;
3728 if (tcp_is_sack(tp))
3729 info->tcpi_options |= TCPI_OPT_SACK;
3730 if (tp->rx_opt.wscale_ok) {
3731 info->tcpi_options |= TCPI_OPT_WSCALE;
3732 info->tcpi_snd_wscale = tp->rx_opt.snd_wscale;
3733 info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale;
3734 }
3735
3736 if (tp->ecn_flags & TCP_ECN_OK)
3737 info->tcpi_options |= TCPI_OPT_ECN;
3738 if (tp->ecn_flags & TCP_ECN_SEEN)
3739 info->tcpi_options |= TCPI_OPT_ECN_SEEN;
3740 if (tp->syn_data_acked)
3741 info->tcpi_options |= TCPI_OPT_SYN_DATA;
3742
3743 info->tcpi_rto = jiffies_to_usecs(icsk->icsk_rto);
3744 info->tcpi_ato = jiffies_to_usecs(icsk->icsk_ack.ato);
3745 info->tcpi_snd_mss = tp->mss_cache;
3746 info->tcpi_rcv_mss = icsk->icsk_ack.rcv_mss;
3747
3748 info->tcpi_unacked = tp->packets_out;
3749 info->tcpi_sacked = tp->sacked_out;
3750
3751 info->tcpi_lost = tp->lost_out;
3752 info->tcpi_retrans = tp->retrans_out;
3753
3754 now = tcp_jiffies32;
3755 info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime);
3756 info->tcpi_last_data_recv = jiffies_to_msecs(now - icsk->icsk_ack.lrcvtime);
3757 info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp);
3758
3759 info->tcpi_pmtu = icsk->icsk_pmtu_cookie;
3760 info->tcpi_rcv_ssthresh = tp->rcv_ssthresh;
3761 info->tcpi_rtt = tp->srtt_us >> 3;
3762 info->tcpi_rttvar = tp->mdev_us >> 2;
3763 info->tcpi_snd_ssthresh = tp->snd_ssthresh;
3764 info->tcpi_advmss = tp->advmss;
3765
3766 info->tcpi_rcv_rtt = tp->rcv_rtt_est.rtt_us >> 3;
3767 info->tcpi_rcv_space = tp->rcvq_space.space;
3768
3769 info->tcpi_total_retrans = tp->total_retrans;
3770
3771 info->tcpi_bytes_acked = tp->bytes_acked;
3772 info->tcpi_bytes_received = tp->bytes_received;
3773 info->tcpi_notsent_bytes = max_t(int, 0, tp->write_seq - tp->snd_nxt);
3774 tcp_get_info_chrono_stats(tp, info);
3775
3776 info->tcpi_segs_out = tp->segs_out;
3777 info->tcpi_segs_in = tp->segs_in;
3778
3779 info->tcpi_min_rtt = tcp_min_rtt(tp);
3780 info->tcpi_data_segs_in = tp->data_segs_in;
3781 info->tcpi_data_segs_out = tp->data_segs_out;
3782
3783 info->tcpi_delivery_rate_app_limited = tp->rate_app_limited ? 1 : 0;
3784 rate64 = tcp_compute_delivery_rate(tp);
3785 if (rate64)
3786 info->tcpi_delivery_rate = rate64;
3787 info->tcpi_delivered = tp->delivered;
3788 info->tcpi_delivered_ce = tp->delivered_ce;
3789 info->tcpi_bytes_sent = tp->bytes_sent;
3790 info->tcpi_bytes_retrans = tp->bytes_retrans;
3791 info->tcpi_dsack_dups = tp->dsack_dups;
3792 info->tcpi_reord_seen = tp->reord_seen;
3793 info->tcpi_rcv_ooopack = tp->rcv_ooopack;
3794 info->tcpi_snd_wnd = tp->snd_wnd;
3795 info->tcpi_fastopen_client_fail = tp->fastopen_client_fail;
3796 unlock_sock_fast(sk, slow);
3797 }
3798 EXPORT_SYMBOL_GPL(tcp_get_info);
3799
tcp_opt_stats_get_size(void)3800 static size_t tcp_opt_stats_get_size(void)
3801 {
3802 return
3803 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_BUSY */
3804 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_RWND_LIMITED */
3805 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_SNDBUF_LIMITED */
3806 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_DATA_SEGS_OUT */
3807 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_TOTAL_RETRANS */
3808 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_PACING_RATE */
3809 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_DELIVERY_RATE */
3810 nla_total_size(sizeof(u32)) + /* TCP_NLA_SND_CWND */
3811 nla_total_size(sizeof(u32)) + /* TCP_NLA_REORDERING */
3812 nla_total_size(sizeof(u32)) + /* TCP_NLA_MIN_RTT */
3813 nla_total_size(sizeof(u8)) + /* TCP_NLA_RECUR_RETRANS */
3814 nla_total_size(sizeof(u8)) + /* TCP_NLA_DELIVERY_RATE_APP_LMT */
3815 nla_total_size(sizeof(u32)) + /* TCP_NLA_SNDQ_SIZE */
3816 nla_total_size(sizeof(u8)) + /* TCP_NLA_CA_STATE */
3817 nla_total_size(sizeof(u32)) + /* TCP_NLA_SND_SSTHRESH */
3818 nla_total_size(sizeof(u32)) + /* TCP_NLA_DELIVERED */
3819 nla_total_size(sizeof(u32)) + /* TCP_NLA_DELIVERED_CE */
3820 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_BYTES_SENT */
3821 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_BYTES_RETRANS */
3822 nla_total_size(sizeof(u32)) + /* TCP_NLA_DSACK_DUPS */
3823 nla_total_size(sizeof(u32)) + /* TCP_NLA_REORD_SEEN */
3824 nla_total_size(sizeof(u32)) + /* TCP_NLA_SRTT */
3825 nla_total_size(sizeof(u16)) + /* TCP_NLA_TIMEOUT_REHASH */
3826 nla_total_size(sizeof(u32)) + /* TCP_NLA_BYTES_NOTSENT */
3827 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_EDT */
3828 nla_total_size(sizeof(u8)) + /* TCP_NLA_TTL */
3829 0;
3830 }
3831
3832 /* Returns TTL or hop limit of an incoming packet from skb. */
tcp_skb_ttl_or_hop_limit(const struct sk_buff * skb)3833 static u8 tcp_skb_ttl_or_hop_limit(const struct sk_buff *skb)
3834 {
3835 if (skb->protocol == htons(ETH_P_IP))
3836 return ip_hdr(skb)->ttl;
3837 else if (skb->protocol == htons(ETH_P_IPV6))
3838 return ipv6_hdr(skb)->hop_limit;
3839 else
3840 return 0;
3841 }
3842
tcp_get_timestamping_opt_stats(const struct sock * sk,const struct sk_buff * orig_skb,const struct sk_buff * ack_skb)3843 struct sk_buff *tcp_get_timestamping_opt_stats(const struct sock *sk,
3844 const struct sk_buff *orig_skb,
3845 const struct sk_buff *ack_skb)
3846 {
3847 const struct tcp_sock *tp = tcp_sk(sk);
3848 struct sk_buff *stats;
3849 struct tcp_info info;
3850 unsigned long rate;
3851 u64 rate64;
3852
3853 stats = alloc_skb(tcp_opt_stats_get_size(), GFP_ATOMIC);
3854 if (!stats)
3855 return NULL;
3856
3857 tcp_get_info_chrono_stats(tp, &info);
3858 nla_put_u64_64bit(stats, TCP_NLA_BUSY,
3859 info.tcpi_busy_time, TCP_NLA_PAD);
3860 nla_put_u64_64bit(stats, TCP_NLA_RWND_LIMITED,
3861 info.tcpi_rwnd_limited, TCP_NLA_PAD);
3862 nla_put_u64_64bit(stats, TCP_NLA_SNDBUF_LIMITED,
3863 info.tcpi_sndbuf_limited, TCP_NLA_PAD);
3864 nla_put_u64_64bit(stats, TCP_NLA_DATA_SEGS_OUT,
3865 tp->data_segs_out, TCP_NLA_PAD);
3866 nla_put_u64_64bit(stats, TCP_NLA_TOTAL_RETRANS,
3867 tp->total_retrans, TCP_NLA_PAD);
3868
3869 rate = READ_ONCE(sk->sk_pacing_rate);
3870 rate64 = (rate != ~0UL) ? rate : ~0ULL;
3871 nla_put_u64_64bit(stats, TCP_NLA_PACING_RATE, rate64, TCP_NLA_PAD);
3872
3873 rate64 = tcp_compute_delivery_rate(tp);
3874 nla_put_u64_64bit(stats, TCP_NLA_DELIVERY_RATE, rate64, TCP_NLA_PAD);
3875
3876 nla_put_u32(stats, TCP_NLA_SND_CWND, tp->snd_cwnd);
3877 nla_put_u32(stats, TCP_NLA_REORDERING, tp->reordering);
3878 nla_put_u32(stats, TCP_NLA_MIN_RTT, tcp_min_rtt(tp));
3879
3880 nla_put_u8(stats, TCP_NLA_RECUR_RETRANS, inet_csk(sk)->icsk_retransmits);
3881 nla_put_u8(stats, TCP_NLA_DELIVERY_RATE_APP_LMT, !!tp->rate_app_limited);
3882 nla_put_u32(stats, TCP_NLA_SND_SSTHRESH, tp->snd_ssthresh);
3883 nla_put_u32(stats, TCP_NLA_DELIVERED, tp->delivered);
3884 nla_put_u32(stats, TCP_NLA_DELIVERED_CE, tp->delivered_ce);
3885
3886 nla_put_u32(stats, TCP_NLA_SNDQ_SIZE, tp->write_seq - tp->snd_una);
3887 nla_put_u8(stats, TCP_NLA_CA_STATE, inet_csk(sk)->icsk_ca_state);
3888
3889 nla_put_u64_64bit(stats, TCP_NLA_BYTES_SENT, tp->bytes_sent,
3890 TCP_NLA_PAD);
3891 nla_put_u64_64bit(stats, TCP_NLA_BYTES_RETRANS, tp->bytes_retrans,
3892 TCP_NLA_PAD);
3893 nla_put_u32(stats, TCP_NLA_DSACK_DUPS, tp->dsack_dups);
3894 nla_put_u32(stats, TCP_NLA_REORD_SEEN, tp->reord_seen);
3895 nla_put_u32(stats, TCP_NLA_SRTT, tp->srtt_us >> 3);
3896 nla_put_u16(stats, TCP_NLA_TIMEOUT_REHASH, tp->timeout_rehash);
3897 nla_put_u32(stats, TCP_NLA_BYTES_NOTSENT,
3898 max_t(int, 0, tp->write_seq - tp->snd_nxt));
3899 nla_put_u64_64bit(stats, TCP_NLA_EDT, orig_skb->skb_mstamp_ns,
3900 TCP_NLA_PAD);
3901 if (ack_skb)
3902 nla_put_u8(stats, TCP_NLA_TTL,
3903 tcp_skb_ttl_or_hop_limit(ack_skb));
3904
3905 return stats;
3906 }
3907
do_tcp_getsockopt(struct sock * sk,int level,int optname,char __user * optval,int __user * optlen)3908 static int do_tcp_getsockopt(struct sock *sk, int level,
3909 int optname, char __user *optval, int __user *optlen)
3910 {
3911 struct inet_connection_sock *icsk = inet_csk(sk);
3912 struct tcp_sock *tp = tcp_sk(sk);
3913 struct net *net = sock_net(sk);
3914 int val, len;
3915
3916 if (get_user(len, optlen))
3917 return -EFAULT;
3918
3919 len = min_t(unsigned int, len, sizeof(int));
3920
3921 if (len < 0)
3922 return -EINVAL;
3923
3924 switch (optname) {
3925 case TCP_MAXSEG:
3926 val = tp->mss_cache;
3927 if (!val && ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
3928 val = tp->rx_opt.user_mss;
3929 if (tp->repair)
3930 val = tp->rx_opt.mss_clamp;
3931 break;
3932 case TCP_NODELAY:
3933 val = !!(tp->nonagle&TCP_NAGLE_OFF);
3934 break;
3935 case TCP_CORK:
3936 val = !!(tp->nonagle&TCP_NAGLE_CORK);
3937 break;
3938 case TCP_KEEPIDLE:
3939 val = keepalive_time_when(tp) / HZ;
3940 break;
3941 case TCP_KEEPINTVL:
3942 val = keepalive_intvl_when(tp) / HZ;
3943 break;
3944 case TCP_KEEPCNT:
3945 val = keepalive_probes(tp);
3946 break;
3947 case TCP_SYNCNT:
3948 val = icsk->icsk_syn_retries ? : net->ipv4.sysctl_tcp_syn_retries;
3949 break;
3950 case TCP_LINGER2:
3951 val = tp->linger2;
3952 if (val >= 0)
3953 val = (val ? : net->ipv4.sysctl_tcp_fin_timeout) / HZ;
3954 break;
3955 case TCP_DEFER_ACCEPT:
3956 val = retrans_to_secs(icsk->icsk_accept_queue.rskq_defer_accept,
3957 TCP_TIMEOUT_INIT / HZ, TCP_RTO_MAX / HZ);
3958 break;
3959 case TCP_WINDOW_CLAMP:
3960 val = tp->window_clamp;
3961 break;
3962 case TCP_INFO: {
3963 struct tcp_info info;
3964
3965 if (get_user(len, optlen))
3966 return -EFAULT;
3967
3968 tcp_get_info(sk, &info);
3969
3970 len = min_t(unsigned int, len, sizeof(info));
3971 if (put_user(len, optlen))
3972 return -EFAULT;
3973 if (copy_to_user(optval, &info, len))
3974 return -EFAULT;
3975 return 0;
3976 }
3977 case TCP_CC_INFO: {
3978 const struct tcp_congestion_ops *ca_ops;
3979 union tcp_cc_info info;
3980 size_t sz = 0;
3981 int attr;
3982
3983 if (get_user(len, optlen))
3984 return -EFAULT;
3985
3986 ca_ops = icsk->icsk_ca_ops;
3987 if (ca_ops && ca_ops->get_info)
3988 sz = ca_ops->get_info(sk, ~0U, &attr, &info);
3989
3990 len = min_t(unsigned int, len, sz);
3991 if (put_user(len, optlen))
3992 return -EFAULT;
3993 if (copy_to_user(optval, &info, len))
3994 return -EFAULT;
3995 return 0;
3996 }
3997 case TCP_QUICKACK:
3998 val = !inet_csk_in_pingpong_mode(sk);
3999 break;
4000
4001 case TCP_CONGESTION:
4002 if (get_user(len, optlen))
4003 return -EFAULT;
4004 len = min_t(unsigned int, len, TCP_CA_NAME_MAX);
4005 if (put_user(len, optlen))
4006 return -EFAULT;
4007 if (copy_to_user(optval, icsk->icsk_ca_ops->name, len))
4008 return -EFAULT;
4009 return 0;
4010
4011 case TCP_ULP:
4012 if (get_user(len, optlen))
4013 return -EFAULT;
4014 len = min_t(unsigned int, len, TCP_ULP_NAME_MAX);
4015 if (!icsk->icsk_ulp_ops) {
4016 if (put_user(0, optlen))
4017 return -EFAULT;
4018 return 0;
4019 }
4020 if (put_user(len, optlen))
4021 return -EFAULT;
4022 if (copy_to_user(optval, icsk->icsk_ulp_ops->name, len))
4023 return -EFAULT;
4024 return 0;
4025
4026 case TCP_FASTOPEN_KEY: {
4027 u64 key[TCP_FASTOPEN_KEY_BUF_LENGTH / sizeof(u64)];
4028 unsigned int key_len;
4029
4030 if (get_user(len, optlen))
4031 return -EFAULT;
4032
4033 key_len = tcp_fastopen_get_cipher(net, icsk, key) *
4034 TCP_FASTOPEN_KEY_LENGTH;
4035 len = min_t(unsigned int, len, key_len);
4036 if (put_user(len, optlen))
4037 return -EFAULT;
4038 if (copy_to_user(optval, key, len))
4039 return -EFAULT;
4040 return 0;
4041 }
4042 case TCP_THIN_LINEAR_TIMEOUTS:
4043 val = tp->thin_lto;
4044 break;
4045
4046 case TCP_THIN_DUPACK:
4047 val = 0;
4048 break;
4049
4050 case TCP_REPAIR:
4051 val = tp->repair;
4052 break;
4053
4054 case TCP_REPAIR_QUEUE:
4055 if (tp->repair)
4056 val = tp->repair_queue;
4057 else
4058 return -EINVAL;
4059 break;
4060
4061 case TCP_REPAIR_WINDOW: {
4062 struct tcp_repair_window opt;
4063
4064 if (get_user(len, optlen))
4065 return -EFAULT;
4066
4067 if (len != sizeof(opt))
4068 return -EINVAL;
4069
4070 if (!tp->repair)
4071 return -EPERM;
4072
4073 opt.snd_wl1 = tp->snd_wl1;
4074 opt.snd_wnd = tp->snd_wnd;
4075 opt.max_window = tp->max_window;
4076 opt.rcv_wnd = tp->rcv_wnd;
4077 opt.rcv_wup = tp->rcv_wup;
4078
4079 if (copy_to_user(optval, &opt, len))
4080 return -EFAULT;
4081 return 0;
4082 }
4083 case TCP_QUEUE_SEQ:
4084 if (tp->repair_queue == TCP_SEND_QUEUE)
4085 val = tp->write_seq;
4086 else if (tp->repair_queue == TCP_RECV_QUEUE)
4087 val = tp->rcv_nxt;
4088 else
4089 return -EINVAL;
4090 break;
4091
4092 case TCP_USER_TIMEOUT:
4093 val = icsk->icsk_user_timeout;
4094 break;
4095
4096 case TCP_FASTOPEN:
4097 val = icsk->icsk_accept_queue.fastopenq.max_qlen;
4098 break;
4099
4100 case TCP_FASTOPEN_CONNECT:
4101 val = tp->fastopen_connect;
4102 break;
4103
4104 case TCP_FASTOPEN_NO_COOKIE:
4105 val = tp->fastopen_no_cookie;
4106 break;
4107
4108 case TCP_TX_DELAY:
4109 val = tp->tcp_tx_delay;
4110 break;
4111
4112 case TCP_TIMESTAMP:
4113 val = tcp_time_stamp_raw() + tp->tsoffset;
4114 break;
4115 case TCP_NOTSENT_LOWAT:
4116 val = tp->notsent_lowat;
4117 break;
4118 case TCP_INQ:
4119 val = tp->recvmsg_inq;
4120 break;
4121 case TCP_SAVE_SYN:
4122 val = tp->save_syn;
4123 break;
4124 case TCP_SAVED_SYN: {
4125 if (get_user(len, optlen))
4126 return -EFAULT;
4127
4128 lock_sock(sk);
4129 if (tp->saved_syn) {
4130 if (len < tcp_saved_syn_len(tp->saved_syn)) {
4131 if (put_user(tcp_saved_syn_len(tp->saved_syn),
4132 optlen)) {
4133 release_sock(sk);
4134 return -EFAULT;
4135 }
4136 release_sock(sk);
4137 return -EINVAL;
4138 }
4139 len = tcp_saved_syn_len(tp->saved_syn);
4140 if (put_user(len, optlen)) {
4141 release_sock(sk);
4142 return -EFAULT;
4143 }
4144 if (copy_to_user(optval, tp->saved_syn->data, len)) {
4145 release_sock(sk);
4146 return -EFAULT;
4147 }
4148 tcp_saved_syn_free(tp);
4149 release_sock(sk);
4150 } else {
4151 release_sock(sk);
4152 len = 0;
4153 if (put_user(len, optlen))
4154 return -EFAULT;
4155 }
4156 return 0;
4157 }
4158 #ifdef CONFIG_MMU
4159 case TCP_ZEROCOPY_RECEIVE: {
4160 struct scm_timestamping_internal tss;
4161 struct tcp_zerocopy_receive zc = {};
4162 int err;
4163
4164 if (get_user(len, optlen))
4165 return -EFAULT;
4166 if (len < 0 ||
4167 len < offsetofend(struct tcp_zerocopy_receive, length))
4168 return -EINVAL;
4169 if (unlikely(len > sizeof(zc))) {
4170 err = check_zeroed_user(optval + sizeof(zc),
4171 len - sizeof(zc));
4172 if (err < 1)
4173 return err == 0 ? -EINVAL : err;
4174 len = sizeof(zc);
4175 if (put_user(len, optlen))
4176 return -EFAULT;
4177 }
4178 if (copy_from_user(&zc, optval, len))
4179 return -EFAULT;
4180 if (zc.reserved)
4181 return -EINVAL;
4182 if (zc.msg_flags & ~(TCP_VALID_ZC_MSG_FLAGS))
4183 return -EINVAL;
4184 lock_sock(sk);
4185 err = tcp_zerocopy_receive(sk, &zc, &tss);
4186 err = BPF_CGROUP_RUN_PROG_GETSOCKOPT_KERN(sk, level, optname,
4187 &zc, &len, err);
4188 release_sock(sk);
4189 if (len >= offsetofend(struct tcp_zerocopy_receive, msg_flags))
4190 goto zerocopy_rcv_cmsg;
4191 switch (len) {
4192 case offsetofend(struct tcp_zerocopy_receive, msg_flags):
4193 goto zerocopy_rcv_cmsg;
4194 case offsetofend(struct tcp_zerocopy_receive, msg_controllen):
4195 case offsetofend(struct tcp_zerocopy_receive, msg_control):
4196 case offsetofend(struct tcp_zerocopy_receive, flags):
4197 case offsetofend(struct tcp_zerocopy_receive, copybuf_len):
4198 case offsetofend(struct tcp_zerocopy_receive, copybuf_address):
4199 case offsetofend(struct tcp_zerocopy_receive, err):
4200 goto zerocopy_rcv_sk_err;
4201 case offsetofend(struct tcp_zerocopy_receive, inq):
4202 goto zerocopy_rcv_inq;
4203 case offsetofend(struct tcp_zerocopy_receive, length):
4204 default:
4205 goto zerocopy_rcv_out;
4206 }
4207 zerocopy_rcv_cmsg:
4208 if (zc.msg_flags & TCP_CMSG_TS)
4209 tcp_zc_finalize_rx_tstamp(sk, &zc, &tss);
4210 else
4211 zc.msg_flags = 0;
4212 zerocopy_rcv_sk_err:
4213 if (!err)
4214 zc.err = sock_error(sk);
4215 zerocopy_rcv_inq:
4216 zc.inq = tcp_inq_hint(sk);
4217 zerocopy_rcv_out:
4218 if (!err && copy_to_user(optval, &zc, len))
4219 err = -EFAULT;
4220 return err;
4221 }
4222 #endif
4223 default:
4224 return -ENOPROTOOPT;
4225 }
4226
4227 if (put_user(len, optlen))
4228 return -EFAULT;
4229 if (copy_to_user(optval, &val, len))
4230 return -EFAULT;
4231 return 0;
4232 }
4233
tcp_bpf_bypass_getsockopt(int level,int optname)4234 bool tcp_bpf_bypass_getsockopt(int level, int optname)
4235 {
4236 /* TCP do_tcp_getsockopt has optimized getsockopt implementation
4237 * to avoid extra socket lock for TCP_ZEROCOPY_RECEIVE.
4238 */
4239 if (level == SOL_TCP && optname == TCP_ZEROCOPY_RECEIVE)
4240 return true;
4241
4242 return false;
4243 }
4244 EXPORT_SYMBOL(tcp_bpf_bypass_getsockopt);
4245
tcp_getsockopt(struct sock * sk,int level,int optname,char __user * optval,int __user * optlen)4246 int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval,
4247 int __user *optlen)
4248 {
4249 struct inet_connection_sock *icsk = inet_csk(sk);
4250
4251 if (level != SOL_TCP)
4252 return icsk->icsk_af_ops->getsockopt(sk, level, optname,
4253 optval, optlen);
4254 return do_tcp_getsockopt(sk, level, optname, optval, optlen);
4255 }
4256 EXPORT_SYMBOL(tcp_getsockopt);
4257
4258 #ifdef CONFIG_TCP_MD5SIG
4259 static DEFINE_PER_CPU(struct tcp_md5sig_pool, tcp_md5sig_pool);
4260 static DEFINE_MUTEX(tcp_md5sig_mutex);
4261 static bool tcp_md5sig_pool_populated = false;
4262
__tcp_alloc_md5sig_pool(void)4263 static void __tcp_alloc_md5sig_pool(void)
4264 {
4265 struct crypto_ahash *hash;
4266 int cpu;
4267
4268 hash = crypto_alloc_ahash("md5", 0, CRYPTO_ALG_ASYNC);
4269 if (IS_ERR(hash))
4270 return;
4271
4272 for_each_possible_cpu(cpu) {
4273 void *scratch = per_cpu(tcp_md5sig_pool, cpu).scratch;
4274 struct ahash_request *req;
4275
4276 if (!scratch) {
4277 scratch = kmalloc_node(sizeof(union tcp_md5sum_block) +
4278 sizeof(struct tcphdr),
4279 GFP_KERNEL,
4280 cpu_to_node(cpu));
4281 if (!scratch)
4282 return;
4283 per_cpu(tcp_md5sig_pool, cpu).scratch = scratch;
4284 }
4285 if (per_cpu(tcp_md5sig_pool, cpu).md5_req)
4286 continue;
4287
4288 req = ahash_request_alloc(hash, GFP_KERNEL);
4289 if (!req)
4290 return;
4291
4292 ahash_request_set_callback(req, 0, NULL, NULL);
4293
4294 per_cpu(tcp_md5sig_pool, cpu).md5_req = req;
4295 }
4296 /* before setting tcp_md5sig_pool_populated, we must commit all writes
4297 * to memory. See smp_rmb() in tcp_get_md5sig_pool()
4298 */
4299 smp_wmb();
4300 tcp_md5sig_pool_populated = true;
4301 }
4302
tcp_alloc_md5sig_pool(void)4303 bool tcp_alloc_md5sig_pool(void)
4304 {
4305 if (unlikely(!tcp_md5sig_pool_populated)) {
4306 mutex_lock(&tcp_md5sig_mutex);
4307
4308 if (!tcp_md5sig_pool_populated) {
4309 __tcp_alloc_md5sig_pool();
4310 if (tcp_md5sig_pool_populated)
4311 static_branch_inc(&tcp_md5_needed);
4312 }
4313
4314 mutex_unlock(&tcp_md5sig_mutex);
4315 }
4316 return tcp_md5sig_pool_populated;
4317 }
4318 EXPORT_SYMBOL(tcp_alloc_md5sig_pool);
4319
4320
4321 /**
4322 * tcp_get_md5sig_pool - get md5sig_pool for this user
4323 *
4324 * We use percpu structure, so if we succeed, we exit with preemption
4325 * and BH disabled, to make sure another thread or softirq handling
4326 * wont try to get same context.
4327 */
tcp_get_md5sig_pool(void)4328 struct tcp_md5sig_pool *tcp_get_md5sig_pool(void)
4329 {
4330 local_bh_disable();
4331
4332 if (tcp_md5sig_pool_populated) {
4333 /* coupled with smp_wmb() in __tcp_alloc_md5sig_pool() */
4334 smp_rmb();
4335 return this_cpu_ptr(&tcp_md5sig_pool);
4336 }
4337 local_bh_enable();
4338 return NULL;
4339 }
4340 EXPORT_SYMBOL(tcp_get_md5sig_pool);
4341
tcp_md5_hash_skb_data(struct tcp_md5sig_pool * hp,const struct sk_buff * skb,unsigned int header_len)4342 int tcp_md5_hash_skb_data(struct tcp_md5sig_pool *hp,
4343 const struct sk_buff *skb, unsigned int header_len)
4344 {
4345 struct scatterlist sg;
4346 const struct tcphdr *tp = tcp_hdr(skb);
4347 struct ahash_request *req = hp->md5_req;
4348 unsigned int i;
4349 const unsigned int head_data_len = skb_headlen(skb) > header_len ?
4350 skb_headlen(skb) - header_len : 0;
4351 const struct skb_shared_info *shi = skb_shinfo(skb);
4352 struct sk_buff *frag_iter;
4353
4354 sg_init_table(&sg, 1);
4355
4356 sg_set_buf(&sg, ((u8 *) tp) + header_len, head_data_len);
4357 ahash_request_set_crypt(req, &sg, NULL, head_data_len);
4358 if (crypto_ahash_update(req))
4359 return 1;
4360
4361 for (i = 0; i < shi->nr_frags; ++i) {
4362 const skb_frag_t *f = &shi->frags[i];
4363 unsigned int offset = skb_frag_off(f);
4364 struct page *page = skb_frag_page(f) + (offset >> PAGE_SHIFT);
4365
4366 sg_set_page(&sg, page, skb_frag_size(f),
4367 offset_in_page(offset));
4368 ahash_request_set_crypt(req, &sg, NULL, skb_frag_size(f));
4369 if (crypto_ahash_update(req))
4370 return 1;
4371 }
4372
4373 skb_walk_frags(skb, frag_iter)
4374 if (tcp_md5_hash_skb_data(hp, frag_iter, 0))
4375 return 1;
4376
4377 return 0;
4378 }
4379 EXPORT_SYMBOL(tcp_md5_hash_skb_data);
4380
tcp_md5_hash_key(struct tcp_md5sig_pool * hp,const struct tcp_md5sig_key * key)4381 int tcp_md5_hash_key(struct tcp_md5sig_pool *hp, const struct tcp_md5sig_key *key)
4382 {
4383 u8 keylen = READ_ONCE(key->keylen); /* paired with WRITE_ONCE() in tcp_md5_do_add */
4384 struct scatterlist sg;
4385
4386 sg_init_one(&sg, key->key, keylen);
4387 ahash_request_set_crypt(hp->md5_req, &sg, NULL, keylen);
4388
4389 /* We use data_race() because tcp_md5_do_add() might change key->key under us */
4390 return data_race(crypto_ahash_update(hp->md5_req));
4391 }
4392 EXPORT_SYMBOL(tcp_md5_hash_key);
4393
4394 #endif
4395
tcp_done(struct sock * sk)4396 void tcp_done(struct sock *sk)
4397 {
4398 struct request_sock *req;
4399
4400 /* We might be called with a new socket, after
4401 * inet_csk_prepare_forced_close() has been called
4402 * so we can not use lockdep_sock_is_held(sk)
4403 */
4404 req = rcu_dereference_protected(tcp_sk(sk)->fastopen_rsk, 1);
4405
4406 if (sk->sk_state == TCP_SYN_SENT || sk->sk_state == TCP_SYN_RECV)
4407 TCP_INC_STATS(sock_net(sk), TCP_MIB_ATTEMPTFAILS);
4408
4409 tcp_set_state(sk, TCP_CLOSE);
4410 tcp_clear_xmit_timers(sk);
4411 if (req)
4412 reqsk_fastopen_remove(sk, req, false);
4413
4414 sk->sk_shutdown = SHUTDOWN_MASK;
4415
4416 if (!sock_flag(sk, SOCK_DEAD))
4417 sk->sk_state_change(sk);
4418 else
4419 inet_csk_destroy_sock(sk);
4420 }
4421 EXPORT_SYMBOL_GPL(tcp_done);
4422
tcp_abort(struct sock * sk,int err)4423 int tcp_abort(struct sock *sk, int err)
4424 {
4425 if (!sk_fullsock(sk)) {
4426 if (sk->sk_state == TCP_NEW_SYN_RECV) {
4427 struct request_sock *req = inet_reqsk(sk);
4428
4429 local_bh_disable();
4430 inet_csk_reqsk_queue_drop(req->rsk_listener, req);
4431 local_bh_enable();
4432 return 0;
4433 }
4434 return -EOPNOTSUPP;
4435 }
4436
4437 /* Don't race with userspace socket closes such as tcp_close. */
4438 lock_sock(sk);
4439
4440 if (sk->sk_state == TCP_LISTEN) {
4441 tcp_set_state(sk, TCP_CLOSE);
4442 inet_csk_listen_stop(sk);
4443 }
4444
4445 /* Don't race with BH socket closes such as inet_csk_listen_stop. */
4446 local_bh_disable();
4447 bh_lock_sock(sk);
4448
4449 if (!sock_flag(sk, SOCK_DEAD)) {
4450 sk->sk_err = err;
4451 /* This barrier is coupled with smp_rmb() in tcp_poll() */
4452 smp_wmb();
4453 sk->sk_error_report(sk);
4454 if (tcp_need_reset(sk->sk_state))
4455 tcp_send_active_reset(sk, GFP_ATOMIC);
4456 tcp_done(sk);
4457 }
4458
4459 bh_unlock_sock(sk);
4460 local_bh_enable();
4461 tcp_write_queue_purge(sk);
4462 release_sock(sk);
4463 return 0;
4464 }
4465 EXPORT_SYMBOL_GPL(tcp_abort);
4466
4467 extern struct tcp_congestion_ops tcp_reno;
4468
4469 static __initdata unsigned long thash_entries;
set_thash_entries(char * str)4470 static int __init set_thash_entries(char *str)
4471 {
4472 ssize_t ret;
4473
4474 if (!str)
4475 return 0;
4476
4477 ret = kstrtoul(str, 0, &thash_entries);
4478 if (ret)
4479 return 0;
4480
4481 return 1;
4482 }
4483 __setup("thash_entries=", set_thash_entries);
4484
tcp_init_mem(void)4485 static void __init tcp_init_mem(void)
4486 {
4487 unsigned long limit = nr_free_buffer_pages() / 16;
4488
4489 limit = max(limit, 128UL);
4490 sysctl_tcp_mem[0] = limit / 4 * 3; /* 4.68 % */
4491 sysctl_tcp_mem[1] = limit; /* 6.25 % */
4492 sysctl_tcp_mem[2] = sysctl_tcp_mem[0] * 2; /* 9.37 % */
4493 }
4494
tcp_init(void)4495 void __init tcp_init(void)
4496 {
4497 int max_rshare, max_wshare, cnt;
4498 unsigned long limit;
4499 unsigned int i;
4500
4501 BUILD_BUG_ON(TCP_MIN_SND_MSS <= MAX_TCP_OPTION_SPACE);
4502 BUILD_BUG_ON(sizeof(struct tcp_skb_cb) >
4503 sizeof_field(struct sk_buff, cb));
4504
4505 percpu_counter_init(&tcp_sockets_allocated, 0, GFP_KERNEL);
4506 percpu_counter_init(&tcp_orphan_count, 0, GFP_KERNEL);
4507 inet_hashinfo_init(&tcp_hashinfo);
4508 inet_hashinfo2_init(&tcp_hashinfo, "tcp_listen_portaddr_hash",
4509 thash_entries, 21, /* one slot per 2 MB*/
4510 0, 64 * 1024);
4511 tcp_hashinfo.bind_bucket_cachep =
4512 kmem_cache_create("tcp_bind_bucket",
4513 sizeof(struct inet_bind_bucket), 0,
4514 SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL);
4515
4516 /* Size and allocate the main established and bind bucket
4517 * hash tables.
4518 *
4519 * The methodology is similar to that of the buffer cache.
4520 */
4521 tcp_hashinfo.ehash =
4522 alloc_large_system_hash("TCP established",
4523 sizeof(struct inet_ehash_bucket),
4524 thash_entries,
4525 17, /* one slot per 128 KB of memory */
4526 0,
4527 NULL,
4528 &tcp_hashinfo.ehash_mask,
4529 0,
4530 thash_entries ? 0 : 512 * 1024);
4531 for (i = 0; i <= tcp_hashinfo.ehash_mask; i++)
4532 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].chain, i);
4533
4534 if (inet_ehash_locks_alloc(&tcp_hashinfo))
4535 panic("TCP: failed to alloc ehash_locks");
4536 tcp_hashinfo.bhash =
4537 alloc_large_system_hash("TCP bind",
4538 sizeof(struct inet_bind_hashbucket),
4539 tcp_hashinfo.ehash_mask + 1,
4540 17, /* one slot per 128 KB of memory */
4541 0,
4542 &tcp_hashinfo.bhash_size,
4543 NULL,
4544 0,
4545 64 * 1024);
4546 tcp_hashinfo.bhash_size = 1U << tcp_hashinfo.bhash_size;
4547 for (i = 0; i < tcp_hashinfo.bhash_size; i++) {
4548 spin_lock_init(&tcp_hashinfo.bhash[i].lock);
4549 INIT_HLIST_HEAD(&tcp_hashinfo.bhash[i].chain);
4550 }
4551
4552
4553 cnt = tcp_hashinfo.ehash_mask + 1;
4554 sysctl_tcp_max_orphans = cnt / 2;
4555
4556 tcp_init_mem();
4557 /* Set per-socket limits to no more than 1/128 the pressure threshold */
4558 limit = nr_free_buffer_pages() << (PAGE_SHIFT - 7);
4559 max_wshare = min(4UL*1024*1024, limit);
4560 max_rshare = min(6UL*1024*1024, limit);
4561
4562 init_net.ipv4.sysctl_tcp_wmem[0] = SK_MEM_QUANTUM;
4563 init_net.ipv4.sysctl_tcp_wmem[1] = 16*1024;
4564 init_net.ipv4.sysctl_tcp_wmem[2] = max(64*1024, max_wshare);
4565
4566 init_net.ipv4.sysctl_tcp_rmem[0] = SK_MEM_QUANTUM;
4567 init_net.ipv4.sysctl_tcp_rmem[1] = 131072;
4568 init_net.ipv4.sysctl_tcp_rmem[2] = max(131072, max_rshare);
4569
4570 pr_info("Hash tables configured (established %u bind %u)\n",
4571 tcp_hashinfo.ehash_mask + 1, tcp_hashinfo.bhash_size);
4572
4573 tcp_v4_init();
4574 tcp_metrics_init();
4575 BUG_ON(tcp_register_congestion_control(&tcp_reno) != 0);
4576 tcp_tasklet_init();
4577 mptcp_init();
4578 }
4579