1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /*
3  * INET		An implementation of the TCP/IP protocol suite for the LINUX
4  *		operating system.  INET is implemented using the  BSD Socket
5  *		interface as the means of communication with the user level.
6  *
7  *		Implementation of the Transmission Control Protocol(TCP).
8  *
9  * Authors:	Ross Biro
10  *		Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
11  *		Mark Evans, <evansmp@uhura.aston.ac.uk>
12  *		Corey Minyard <wf-rch!minyard@relay.EU.net>
13  *		Florian La Roche, <flla@stud.uni-sb.de>
14  *		Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
15  *		Linus Torvalds, <torvalds@cs.helsinki.fi>
16  *		Alan Cox, <gw4pts@gw4pts.ampr.org>
17  *		Matthew Dillon, <dillon@apollo.west.oic.com>
18  *		Arnt Gulbrandsen, <agulbra@nvg.unit.no>
19  *		Jorge Cwik, <jorge@laser.satlink.net>
20  *
21  * Fixes:
22  *		Alan Cox	:	Numerous verify_area() calls
23  *		Alan Cox	:	Set the ACK bit on a reset
24  *		Alan Cox	:	Stopped it crashing if it closed while
25  *					sk->inuse=1 and was trying to connect
26  *					(tcp_err()).
27  *		Alan Cox	:	All icmp error handling was broken
28  *					pointers passed where wrong and the
29  *					socket was looked up backwards. Nobody
30  *					tested any icmp error code obviously.
31  *		Alan Cox	:	tcp_err() now handled properly. It
32  *					wakes people on errors. poll
33  *					behaves and the icmp error race
34  *					has gone by moving it into sock.c
35  *		Alan Cox	:	tcp_send_reset() fixed to work for
36  *					everything not just packets for
37  *					unknown sockets.
38  *		Alan Cox	:	tcp option processing.
39  *		Alan Cox	:	Reset tweaked (still not 100%) [Had
40  *					syn rule wrong]
41  *		Herp Rosmanith  :	More reset fixes
42  *		Alan Cox	:	No longer acks invalid rst frames.
43  *					Acking any kind of RST is right out.
44  *		Alan Cox	:	Sets an ignore me flag on an rst
45  *					receive otherwise odd bits of prattle
46  *					escape still
47  *		Alan Cox	:	Fixed another acking RST frame bug.
48  *					Should stop LAN workplace lockups.
49  *		Alan Cox	: 	Some tidyups using the new skb list
50  *					facilities
51  *		Alan Cox	:	sk->keepopen now seems to work
52  *		Alan Cox	:	Pulls options out correctly on accepts
53  *		Alan Cox	:	Fixed assorted sk->rqueue->next errors
54  *		Alan Cox	:	PSH doesn't end a TCP read. Switched a
55  *					bit to skb ops.
56  *		Alan Cox	:	Tidied tcp_data to avoid a potential
57  *					nasty.
58  *		Alan Cox	:	Added some better commenting, as the
59  *					tcp is hard to follow
60  *		Alan Cox	:	Removed incorrect check for 20 * psh
61  *	Michael O'Reilly	:	ack < copied bug fix.
62  *	Johannes Stille		:	Misc tcp fixes (not all in yet).
63  *		Alan Cox	:	FIN with no memory -> CRASH
64  *		Alan Cox	:	Added socket option proto entries.
65  *					Also added awareness of them to accept.
66  *		Alan Cox	:	Added TCP options (SOL_TCP)
67  *		Alan Cox	:	Switched wakeup calls to callbacks,
68  *					so the kernel can layer network
69  *					sockets.
70  *		Alan Cox	:	Use ip_tos/ip_ttl settings.
71  *		Alan Cox	:	Handle FIN (more) properly (we hope).
72  *		Alan Cox	:	RST frames sent on unsynchronised
73  *					state ack error.
74  *		Alan Cox	:	Put in missing check for SYN bit.
75  *		Alan Cox	:	Added tcp_select_window() aka NET2E
76  *					window non shrink trick.
77  *		Alan Cox	:	Added a couple of small NET2E timer
78  *					fixes
79  *		Charles Hedrick :	TCP fixes
80  *		Toomas Tamm	:	TCP window fixes
81  *		Alan Cox	:	Small URG fix to rlogin ^C ack fight
82  *		Charles Hedrick	:	Rewrote most of it to actually work
83  *		Linus		:	Rewrote tcp_read() and URG handling
84  *					completely
85  *		Gerhard Koerting:	Fixed some missing timer handling
86  *		Matthew Dillon  :	Reworked TCP machine states as per RFC
87  *		Gerhard Koerting:	PC/TCP workarounds
88  *		Adam Caldwell	:	Assorted timer/timing errors
89  *		Matthew Dillon	:	Fixed another RST bug
90  *		Alan Cox	:	Move to kernel side addressing changes.
91  *		Alan Cox	:	Beginning work on TCP fastpathing
92  *					(not yet usable)
93  *		Arnt Gulbrandsen:	Turbocharged tcp_check() routine.
94  *		Alan Cox	:	TCP fast path debugging
95  *		Alan Cox	:	Window clamping
96  *		Michael Riepe	:	Bug in tcp_check()
97  *		Matt Dillon	:	More TCP improvements and RST bug fixes
98  *		Matt Dillon	:	Yet more small nasties remove from the
99  *					TCP code (Be very nice to this man if
100  *					tcp finally works 100%) 8)
101  *		Alan Cox	:	BSD accept semantics.
102  *		Alan Cox	:	Reset on closedown bug.
103  *	Peter De Schrijver	:	ENOTCONN check missing in tcp_sendto().
104  *		Michael Pall	:	Handle poll() after URG properly in
105  *					all cases.
106  *		Michael Pall	:	Undo the last fix in tcp_read_urg()
107  *					(multi URG PUSH broke rlogin).
108  *		Michael Pall	:	Fix the multi URG PUSH problem in
109  *					tcp_readable(), poll() after URG
110  *					works now.
111  *		Michael Pall	:	recv(...,MSG_OOB) never blocks in the
112  *					BSD api.
113  *		Alan Cox	:	Changed the semantics of sk->socket to
114  *					fix a race and a signal problem with
115  *					accept() and async I/O.
116  *		Alan Cox	:	Relaxed the rules on tcp_sendto().
117  *		Yury Shevchuk	:	Really fixed accept() blocking problem.
118  *		Craig I. Hagan  :	Allow for BSD compatible TIME_WAIT for
119  *					clients/servers which listen in on
120  *					fixed ports.
121  *		Alan Cox	:	Cleaned the above up and shrank it to
122  *					a sensible code size.
123  *		Alan Cox	:	Self connect lockup fix.
124  *		Alan Cox	:	No connect to multicast.
125  *		Ross Biro	:	Close unaccepted children on master
126  *					socket close.
127  *		Alan Cox	:	Reset tracing code.
128  *		Alan Cox	:	Spurious resets on shutdown.
129  *		Alan Cox	:	Giant 15 minute/60 second timer error
130  *		Alan Cox	:	Small whoops in polling before an
131  *					accept.
132  *		Alan Cox	:	Kept the state trace facility since
133  *					it's handy for debugging.
134  *		Alan Cox	:	More reset handler fixes.
135  *		Alan Cox	:	Started rewriting the code based on
136  *					the RFC's for other useful protocol
137  *					references see: Comer, KA9Q NOS, and
138  *					for a reference on the difference
139  *					between specifications and how BSD
140  *					works see the 4.4lite source.
141  *		A.N.Kuznetsov	:	Don't time wait on completion of tidy
142  *					close.
143  *		Linus Torvalds	:	Fin/Shutdown & copied_seq changes.
144  *		Linus Torvalds	:	Fixed BSD port reuse to work first syn
145  *		Alan Cox	:	Reimplemented timers as per the RFC
146  *					and using multiple timers for sanity.
147  *		Alan Cox	:	Small bug fixes, and a lot of new
148  *					comments.
149  *		Alan Cox	:	Fixed dual reader crash by locking
150  *					the buffers (much like datagram.c)
151  *		Alan Cox	:	Fixed stuck sockets in probe. A probe
152  *					now gets fed up of retrying without
153  *					(even a no space) answer.
154  *		Alan Cox	:	Extracted closing code better
155  *		Alan Cox	:	Fixed the closing state machine to
156  *					resemble the RFC.
157  *		Alan Cox	:	More 'per spec' fixes.
158  *		Jorge Cwik	:	Even faster checksumming.
159  *		Alan Cox	:	tcp_data() doesn't ack illegal PSH
160  *					only frames. At least one pc tcp stack
161  *					generates them.
162  *		Alan Cox	:	Cache last socket.
163  *		Alan Cox	:	Per route irtt.
164  *		Matt Day	:	poll()->select() match BSD precisely on error
165  *		Alan Cox	:	New buffers
166  *		Marc Tamsky	:	Various sk->prot->retransmits and
167  *					sk->retransmits misupdating fixed.
168  *					Fixed tcp_write_timeout: stuck close,
169  *					and TCP syn retries gets used now.
170  *		Mark Yarvis	:	In tcp_read_wakeup(), don't send an
171  *					ack if state is TCP_CLOSED.
172  *		Alan Cox	:	Look up device on a retransmit - routes may
173  *					change. Doesn't yet cope with MSS shrink right
174  *					but it's a start!
175  *		Marc Tamsky	:	Closing in closing fixes.
176  *		Mike Shaver	:	RFC1122 verifications.
177  *		Alan Cox	:	rcv_saddr errors.
178  *		Alan Cox	:	Block double connect().
179  *		Alan Cox	:	Small hooks for enSKIP.
180  *		Alexey Kuznetsov:	Path MTU discovery.
181  *		Alan Cox	:	Support soft errors.
182  *		Alan Cox	:	Fix MTU discovery pathological case
183  *					when the remote claims no mtu!
184  *		Marc Tamsky	:	TCP_CLOSE fix.
185  *		Colin (G3TNE)	:	Send a reset on syn ack replies in
186  *					window but wrong (fixes NT lpd problems)
187  *		Pedro Roque	:	Better TCP window handling, delayed ack.
188  *		Joerg Reuter	:	No modification of locked buffers in
189  *					tcp_do_retransmit()
190  *		Eric Schenk	:	Changed receiver side silly window
191  *					avoidance algorithm to BSD style
192  *					algorithm. This doubles throughput
193  *					against machines running Solaris,
194  *					and seems to result in general
195  *					improvement.
196  *	Stefan Magdalinski	:	adjusted tcp_readable() to fix FIONREAD
197  *	Willy Konynenberg	:	Transparent proxying support.
198  *	Mike McLagan		:	Routing by source
199  *		Keith Owens	:	Do proper merging with partial SKB's in
200  *					tcp_do_sendmsg to avoid burstiness.
201  *		Eric Schenk	:	Fix fast close down bug with
202  *					shutdown() followed by close().
203  *		Andi Kleen 	:	Make poll agree with SIGIO
204  *	Salvatore Sanfilippo	:	Support SO_LINGER with linger == 1 and
205  *					lingertime == 0 (RFC 793 ABORT Call)
206  *	Hirokazu Takahashi	:	Use copy_from_user() instead of
207  *					csum_and_copy_from_user() if possible.
208  *
209  * Description of States:
210  *
211  *	TCP_SYN_SENT		sent a connection request, waiting for ack
212  *
213  *	TCP_SYN_RECV		received a connection request, sent ack,
214  *				waiting for final ack in three-way handshake.
215  *
216  *	TCP_ESTABLISHED		connection established
217  *
218  *	TCP_FIN_WAIT1		our side has shutdown, waiting to complete
219  *				transmission of remaining buffered data
220  *
221  *	TCP_FIN_WAIT2		all buffered data sent, waiting for remote
222  *				to shutdown
223  *
224  *	TCP_CLOSING		both sides have shutdown but we still have
225  *				data we have to finish sending
226  *
227  *	TCP_TIME_WAIT		timeout to catch resent junk before entering
228  *				closed, can only be entered from FIN_WAIT2
229  *				or CLOSING.  Required because the other end
230  *				may not have gotten our last ACK causing it
231  *				to retransmit the data packet (which we ignore)
232  *
233  *	TCP_CLOSE_WAIT		remote side has shutdown and is waiting for
234  *				us to finish writing our data and to shutdown
235  *				(we have to close() to move on to LAST_ACK)
236  *
237  *	TCP_LAST_ACK		out side has shutdown after remote has
238  *				shutdown.  There may still be data in our
239  *				buffer that we have to finish sending
240  *
241  *	TCP_CLOSE		socket is finished
242  */
243 
244 #define pr_fmt(fmt) "TCP: " fmt
245 
246 #include <crypto/hash.h>
247 #include <linux/kernel.h>
248 #include <linux/module.h>
249 #include <linux/types.h>
250 #include <linux/fcntl.h>
251 #include <linux/poll.h>
252 #include <linux/inet_diag.h>
253 #include <linux/init.h>
254 #include <linux/fs.h>
255 #include <linux/skbuff.h>
256 #include <linux/scatterlist.h>
257 #include <linux/splice.h>
258 #include <linux/net.h>
259 #include <linux/socket.h>
260 #include <linux/random.h>
261 #include <linux/memblock.h>
262 #include <linux/highmem.h>
263 #include <linux/cache.h>
264 #include <linux/err.h>
265 #include <linux/time.h>
266 #include <linux/slab.h>
267 #include <linux/errqueue.h>
268 #include <linux/static_key.h>
269 #include <linux/btf.h>
270 
271 #include <net/icmp.h>
272 #include <net/inet_common.h>
273 #include <net/tcp.h>
274 #include <net/mptcp.h>
275 #include <net/xfrm.h>
276 #include <net/ip.h>
277 #include <net/sock.h>
278 
279 #include <linux/uaccess.h>
280 #include <asm/ioctls.h>
281 #include <net/busy_poll.h>
282 
283 /* Track pending CMSGs. */
284 enum {
285 	TCP_CMSG_INQ = 1,
286 	TCP_CMSG_TS = 2
287 };
288 
289 DEFINE_PER_CPU(unsigned int, tcp_orphan_count);
290 EXPORT_PER_CPU_SYMBOL_GPL(tcp_orphan_count);
291 
292 long sysctl_tcp_mem[3] __read_mostly;
293 EXPORT_SYMBOL(sysctl_tcp_mem);
294 
295 atomic_long_t tcp_memory_allocated ____cacheline_aligned_in_smp;	/* Current allocated memory. */
296 EXPORT_SYMBOL(tcp_memory_allocated);
297 DEFINE_PER_CPU(int, tcp_memory_per_cpu_fw_alloc);
298 EXPORT_PER_CPU_SYMBOL_GPL(tcp_memory_per_cpu_fw_alloc);
299 
300 #if IS_ENABLED(CONFIG_SMC)
301 DEFINE_STATIC_KEY_FALSE(tcp_have_smc);
302 EXPORT_SYMBOL(tcp_have_smc);
303 #endif
304 
305 /*
306  * Current number of TCP sockets.
307  */
308 struct percpu_counter tcp_sockets_allocated ____cacheline_aligned_in_smp;
309 EXPORT_SYMBOL(tcp_sockets_allocated);
310 
311 /*
312  * TCP splice context
313  */
314 struct tcp_splice_state {
315 	struct pipe_inode_info *pipe;
316 	size_t len;
317 	unsigned int flags;
318 };
319 
320 /*
321  * Pressure flag: try to collapse.
322  * Technical note: it is used by multiple contexts non atomically.
323  * All the __sk_mem_schedule() is of this nature: accounting
324  * is strict, actions are advisory and have some latency.
325  */
326 unsigned long tcp_memory_pressure __read_mostly;
327 EXPORT_SYMBOL_GPL(tcp_memory_pressure);
328 
tcp_enter_memory_pressure(struct sock * sk)329 void tcp_enter_memory_pressure(struct sock *sk)
330 {
331 	unsigned long val;
332 
333 	if (READ_ONCE(tcp_memory_pressure))
334 		return;
335 	val = jiffies;
336 
337 	if (!val)
338 		val--;
339 	if (!cmpxchg(&tcp_memory_pressure, 0, val))
340 		NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMEMORYPRESSURES);
341 }
342 EXPORT_SYMBOL_GPL(tcp_enter_memory_pressure);
343 
tcp_leave_memory_pressure(struct sock * sk)344 void tcp_leave_memory_pressure(struct sock *sk)
345 {
346 	unsigned long val;
347 
348 	if (!READ_ONCE(tcp_memory_pressure))
349 		return;
350 	val = xchg(&tcp_memory_pressure, 0);
351 	if (val)
352 		NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPMEMORYPRESSURESCHRONO,
353 			      jiffies_to_msecs(jiffies - val));
354 }
355 EXPORT_SYMBOL_GPL(tcp_leave_memory_pressure);
356 
357 /* Convert seconds to retransmits based on initial and max timeout */
secs_to_retrans(int seconds,int timeout,int rto_max)358 static u8 secs_to_retrans(int seconds, int timeout, int rto_max)
359 {
360 	u8 res = 0;
361 
362 	if (seconds > 0) {
363 		int period = timeout;
364 
365 		res = 1;
366 		while (seconds > period && res < 255) {
367 			res++;
368 			timeout <<= 1;
369 			if (timeout > rto_max)
370 				timeout = rto_max;
371 			period += timeout;
372 		}
373 	}
374 	return res;
375 }
376 
377 /* Convert retransmits to seconds based on initial and max timeout */
retrans_to_secs(u8 retrans,int timeout,int rto_max)378 static int retrans_to_secs(u8 retrans, int timeout, int rto_max)
379 {
380 	int period = 0;
381 
382 	if (retrans > 0) {
383 		period = timeout;
384 		while (--retrans) {
385 			timeout <<= 1;
386 			if (timeout > rto_max)
387 				timeout = rto_max;
388 			period += timeout;
389 		}
390 	}
391 	return period;
392 }
393 
tcp_compute_delivery_rate(const struct tcp_sock * tp)394 static u64 tcp_compute_delivery_rate(const struct tcp_sock *tp)
395 {
396 	u32 rate = READ_ONCE(tp->rate_delivered);
397 	u32 intv = READ_ONCE(tp->rate_interval_us);
398 	u64 rate64 = 0;
399 
400 	if (rate && intv) {
401 		rate64 = (u64)rate * tp->mss_cache * USEC_PER_SEC;
402 		do_div(rate64, intv);
403 	}
404 	return rate64;
405 }
406 
407 /* Address-family independent initialization for a tcp_sock.
408  *
409  * NOTE: A lot of things set to zero explicitly by call to
410  *       sk_alloc() so need not be done here.
411  */
tcp_init_sock(struct sock * sk)412 void tcp_init_sock(struct sock *sk)
413 {
414 	struct inet_connection_sock *icsk = inet_csk(sk);
415 	struct tcp_sock *tp = tcp_sk(sk);
416 
417 	tp->out_of_order_queue = RB_ROOT;
418 	sk->tcp_rtx_queue = RB_ROOT;
419 	tcp_init_xmit_timers(sk);
420 	INIT_LIST_HEAD(&tp->tsq_node);
421 	INIT_LIST_HEAD(&tp->tsorted_sent_queue);
422 
423 	icsk->icsk_rto = TCP_TIMEOUT_INIT;
424 	icsk->icsk_rto_min = TCP_RTO_MIN;
425 	icsk->icsk_delack_max = TCP_DELACK_MAX;
426 	tp->mdev_us = jiffies_to_usecs(TCP_TIMEOUT_INIT);
427 	minmax_reset(&tp->rtt_min, tcp_jiffies32, ~0U);
428 
429 	/* So many TCP implementations out there (incorrectly) count the
430 	 * initial SYN frame in their delayed-ACK and congestion control
431 	 * algorithms that we must have the following bandaid to talk
432 	 * efficiently to them.  -DaveM
433 	 */
434 	tcp_snd_cwnd_set(tp, TCP_INIT_CWND);
435 
436 	/* There's a bubble in the pipe until at least the first ACK. */
437 	tp->app_limited = ~0U;
438 	tp->rate_app_limited = 1;
439 
440 	/* See draft-stevens-tcpca-spec-01 for discussion of the
441 	 * initialization of these values.
442 	 */
443 	tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
444 	tp->snd_cwnd_clamp = ~0;
445 	tp->mss_cache = TCP_MSS_DEFAULT;
446 
447 	tp->reordering = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_reordering);
448 	tcp_assign_congestion_control(sk);
449 
450 	tp->tsoffset = 0;
451 	tp->rack.reo_wnd_steps = 1;
452 
453 	sk->sk_write_space = sk_stream_write_space;
454 	sock_set_flag(sk, SOCK_USE_WRITE_QUEUE);
455 
456 	icsk->icsk_sync_mss = tcp_sync_mss;
457 
458 	WRITE_ONCE(sk->sk_sndbuf, READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_wmem[1]));
459 	WRITE_ONCE(sk->sk_rcvbuf, READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_rmem[1]));
460 
461 	set_bit(SOCK_SUPPORT_ZC, &sk->sk_socket->flags);
462 	sk_sockets_allocated_inc(sk);
463 }
464 EXPORT_SYMBOL(tcp_init_sock);
465 
tcp_tx_timestamp(struct sock * sk,u16 tsflags)466 static void tcp_tx_timestamp(struct sock *sk, u16 tsflags)
467 {
468 	struct sk_buff *skb = tcp_write_queue_tail(sk);
469 
470 	if (tsflags && skb) {
471 		struct skb_shared_info *shinfo = skb_shinfo(skb);
472 		struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
473 
474 		sock_tx_timestamp(sk, tsflags, &shinfo->tx_flags);
475 		if (tsflags & SOF_TIMESTAMPING_TX_ACK)
476 			tcb->txstamp_ack = 1;
477 		if (tsflags & SOF_TIMESTAMPING_TX_RECORD_MASK)
478 			shinfo->tskey = TCP_SKB_CB(skb)->seq + skb->len - 1;
479 	}
480 }
481 
tcp_stream_is_readable(struct sock * sk,int target)482 static bool tcp_stream_is_readable(struct sock *sk, int target)
483 {
484 	if (tcp_epollin_ready(sk, target))
485 		return true;
486 	return sk_is_readable(sk);
487 }
488 
489 /*
490  *	Wait for a TCP event.
491  *
492  *	Note that we don't need to lock the socket, as the upper poll layers
493  *	take care of normal races (between the test and the event) and we don't
494  *	go look at any of the socket buffers directly.
495  */
tcp_poll(struct file * file,struct socket * sock,poll_table * wait)496 __poll_t tcp_poll(struct file *file, struct socket *sock, poll_table *wait)
497 {
498 	__poll_t mask;
499 	struct sock *sk = sock->sk;
500 	const struct tcp_sock *tp = tcp_sk(sk);
501 	int state;
502 
503 	sock_poll_wait(file, sock, wait);
504 
505 	state = inet_sk_state_load(sk);
506 	if (state == TCP_LISTEN)
507 		return inet_csk_listen_poll(sk);
508 
509 	/* Socket is not locked. We are protected from async events
510 	 * by poll logic and correct handling of state changes
511 	 * made by other threads is impossible in any case.
512 	 */
513 
514 	mask = 0;
515 
516 	/*
517 	 * EPOLLHUP is certainly not done right. But poll() doesn't
518 	 * have a notion of HUP in just one direction, and for a
519 	 * socket the read side is more interesting.
520 	 *
521 	 * Some poll() documentation says that EPOLLHUP is incompatible
522 	 * with the EPOLLOUT/POLLWR flags, so somebody should check this
523 	 * all. But careful, it tends to be safer to return too many
524 	 * bits than too few, and you can easily break real applications
525 	 * if you don't tell them that something has hung up!
526 	 *
527 	 * Check-me.
528 	 *
529 	 * Check number 1. EPOLLHUP is _UNMASKABLE_ event (see UNIX98 and
530 	 * our fs/select.c). It means that after we received EOF,
531 	 * poll always returns immediately, making impossible poll() on write()
532 	 * in state CLOSE_WAIT. One solution is evident --- to set EPOLLHUP
533 	 * if and only if shutdown has been made in both directions.
534 	 * Actually, it is interesting to look how Solaris and DUX
535 	 * solve this dilemma. I would prefer, if EPOLLHUP were maskable,
536 	 * then we could set it on SND_SHUTDOWN. BTW examples given
537 	 * in Stevens' books assume exactly this behaviour, it explains
538 	 * why EPOLLHUP is incompatible with EPOLLOUT.	--ANK
539 	 *
540 	 * NOTE. Check for TCP_CLOSE is added. The goal is to prevent
541 	 * blocking on fresh not-connected or disconnected socket. --ANK
542 	 */
543 	if (sk->sk_shutdown == SHUTDOWN_MASK || state == TCP_CLOSE)
544 		mask |= EPOLLHUP;
545 	if (sk->sk_shutdown & RCV_SHUTDOWN)
546 		mask |= EPOLLIN | EPOLLRDNORM | EPOLLRDHUP;
547 
548 	/* Connected or passive Fast Open socket? */
549 	if (state != TCP_SYN_SENT &&
550 	    (state != TCP_SYN_RECV || rcu_access_pointer(tp->fastopen_rsk))) {
551 		int target = sock_rcvlowat(sk, 0, INT_MAX);
552 		u16 urg_data = READ_ONCE(tp->urg_data);
553 
554 		if (unlikely(urg_data) &&
555 		    READ_ONCE(tp->urg_seq) == READ_ONCE(tp->copied_seq) &&
556 		    !sock_flag(sk, SOCK_URGINLINE))
557 			target++;
558 
559 		if (tcp_stream_is_readable(sk, target))
560 			mask |= EPOLLIN | EPOLLRDNORM;
561 
562 		if (!(sk->sk_shutdown & SEND_SHUTDOWN)) {
563 			if (__sk_stream_is_writeable(sk, 1)) {
564 				mask |= EPOLLOUT | EPOLLWRNORM;
565 			} else {  /* send SIGIO later */
566 				sk_set_bit(SOCKWQ_ASYNC_NOSPACE, sk);
567 				set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
568 
569 				/* Race breaker. If space is freed after
570 				 * wspace test but before the flags are set,
571 				 * IO signal will be lost. Memory barrier
572 				 * pairs with the input side.
573 				 */
574 				smp_mb__after_atomic();
575 				if (__sk_stream_is_writeable(sk, 1))
576 					mask |= EPOLLOUT | EPOLLWRNORM;
577 			}
578 		} else
579 			mask |= EPOLLOUT | EPOLLWRNORM;
580 
581 		if (urg_data & TCP_URG_VALID)
582 			mask |= EPOLLPRI;
583 	} else if (state == TCP_SYN_SENT && inet_sk(sk)->defer_connect) {
584 		/* Active TCP fastopen socket with defer_connect
585 		 * Return EPOLLOUT so application can call write()
586 		 * in order for kernel to generate SYN+data
587 		 */
588 		mask |= EPOLLOUT | EPOLLWRNORM;
589 	}
590 	/* This barrier is coupled with smp_wmb() in tcp_reset() */
591 	smp_rmb();
592 	if (sk->sk_err || !skb_queue_empty_lockless(&sk->sk_error_queue))
593 		mask |= EPOLLERR;
594 
595 	return mask;
596 }
597 EXPORT_SYMBOL(tcp_poll);
598 
tcp_ioctl(struct sock * sk,int cmd,unsigned long arg)599 int tcp_ioctl(struct sock *sk, int cmd, unsigned long arg)
600 {
601 	struct tcp_sock *tp = tcp_sk(sk);
602 	int answ;
603 	bool slow;
604 
605 	switch (cmd) {
606 	case SIOCINQ:
607 		if (sk->sk_state == TCP_LISTEN)
608 			return -EINVAL;
609 
610 		slow = lock_sock_fast(sk);
611 		answ = tcp_inq(sk);
612 		unlock_sock_fast(sk, slow);
613 		break;
614 	case SIOCATMARK:
615 		answ = READ_ONCE(tp->urg_data) &&
616 		       READ_ONCE(tp->urg_seq) == READ_ONCE(tp->copied_seq);
617 		break;
618 	case SIOCOUTQ:
619 		if (sk->sk_state == TCP_LISTEN)
620 			return -EINVAL;
621 
622 		if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
623 			answ = 0;
624 		else
625 			answ = READ_ONCE(tp->write_seq) - tp->snd_una;
626 		break;
627 	case SIOCOUTQNSD:
628 		if (sk->sk_state == TCP_LISTEN)
629 			return -EINVAL;
630 
631 		if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
632 			answ = 0;
633 		else
634 			answ = READ_ONCE(tp->write_seq) -
635 			       READ_ONCE(tp->snd_nxt);
636 		break;
637 	default:
638 		return -ENOIOCTLCMD;
639 	}
640 
641 	return put_user(answ, (int __user *)arg);
642 }
643 EXPORT_SYMBOL(tcp_ioctl);
644 
tcp_mark_push(struct tcp_sock * tp,struct sk_buff * skb)645 void tcp_mark_push(struct tcp_sock *tp, struct sk_buff *skb)
646 {
647 	TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
648 	tp->pushed_seq = tp->write_seq;
649 }
650 
forced_push(const struct tcp_sock * tp)651 static inline bool forced_push(const struct tcp_sock *tp)
652 {
653 	return after(tp->write_seq, tp->pushed_seq + (tp->max_window >> 1));
654 }
655 
tcp_skb_entail(struct sock * sk,struct sk_buff * skb)656 void tcp_skb_entail(struct sock *sk, struct sk_buff *skb)
657 {
658 	struct tcp_sock *tp = tcp_sk(sk);
659 	struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
660 
661 	tcb->seq     = tcb->end_seq = tp->write_seq;
662 	tcb->tcp_flags = TCPHDR_ACK;
663 	__skb_header_release(skb);
664 	tcp_add_write_queue_tail(sk, skb);
665 	sk_wmem_queued_add(sk, skb->truesize);
666 	sk_mem_charge(sk, skb->truesize);
667 	if (tp->nonagle & TCP_NAGLE_PUSH)
668 		tp->nonagle &= ~TCP_NAGLE_PUSH;
669 
670 	tcp_slow_start_after_idle_check(sk);
671 }
672 
tcp_mark_urg(struct tcp_sock * tp,int flags)673 static inline void tcp_mark_urg(struct tcp_sock *tp, int flags)
674 {
675 	if (flags & MSG_OOB)
676 		tp->snd_up = tp->write_seq;
677 }
678 
679 /* If a not yet filled skb is pushed, do not send it if
680  * we have data packets in Qdisc or NIC queues :
681  * Because TX completion will happen shortly, it gives a chance
682  * to coalesce future sendmsg() payload into this skb, without
683  * need for a timer, and with no latency trade off.
684  * As packets containing data payload have a bigger truesize
685  * than pure acks (dataless) packets, the last checks prevent
686  * autocorking if we only have an ACK in Qdisc/NIC queues,
687  * or if TX completion was delayed after we processed ACK packet.
688  */
tcp_should_autocork(struct sock * sk,struct sk_buff * skb,int size_goal)689 static bool tcp_should_autocork(struct sock *sk, struct sk_buff *skb,
690 				int size_goal)
691 {
692 	return skb->len < size_goal &&
693 	       READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_autocorking) &&
694 	       !tcp_rtx_queue_empty(sk) &&
695 	       refcount_read(&sk->sk_wmem_alloc) > skb->truesize &&
696 	       tcp_skb_can_collapse_to(skb);
697 }
698 
tcp_push(struct sock * sk,int flags,int mss_now,int nonagle,int size_goal)699 void tcp_push(struct sock *sk, int flags, int mss_now,
700 	      int nonagle, int size_goal)
701 {
702 	struct tcp_sock *tp = tcp_sk(sk);
703 	struct sk_buff *skb;
704 
705 	skb = tcp_write_queue_tail(sk);
706 	if (!skb)
707 		return;
708 	if (!(flags & MSG_MORE) || forced_push(tp))
709 		tcp_mark_push(tp, skb);
710 
711 	tcp_mark_urg(tp, flags);
712 
713 	if (tcp_should_autocork(sk, skb, size_goal)) {
714 
715 		/* avoid atomic op if TSQ_THROTTLED bit is already set */
716 		if (!test_bit(TSQ_THROTTLED, &sk->sk_tsq_flags)) {
717 			NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPAUTOCORKING);
718 			set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags);
719 		}
720 		/* It is possible TX completion already happened
721 		 * before we set TSQ_THROTTLED.
722 		 */
723 		if (refcount_read(&sk->sk_wmem_alloc) > skb->truesize)
724 			return;
725 	}
726 
727 	if (flags & MSG_MORE)
728 		nonagle = TCP_NAGLE_CORK;
729 
730 	__tcp_push_pending_frames(sk, mss_now, nonagle);
731 }
732 
tcp_splice_data_recv(read_descriptor_t * rd_desc,struct sk_buff * skb,unsigned int offset,size_t len)733 static int tcp_splice_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb,
734 				unsigned int offset, size_t len)
735 {
736 	struct tcp_splice_state *tss = rd_desc->arg.data;
737 	int ret;
738 
739 	ret = skb_splice_bits(skb, skb->sk, offset, tss->pipe,
740 			      min(rd_desc->count, len), tss->flags);
741 	if (ret > 0)
742 		rd_desc->count -= ret;
743 	return ret;
744 }
745 
__tcp_splice_read(struct sock * sk,struct tcp_splice_state * tss)746 static int __tcp_splice_read(struct sock *sk, struct tcp_splice_state *tss)
747 {
748 	/* Store TCP splice context information in read_descriptor_t. */
749 	read_descriptor_t rd_desc = {
750 		.arg.data = tss,
751 		.count	  = tss->len,
752 	};
753 
754 	return tcp_read_sock(sk, &rd_desc, tcp_splice_data_recv);
755 }
756 
757 /**
758  *  tcp_splice_read - splice data from TCP socket to a pipe
759  * @sock:	socket to splice from
760  * @ppos:	position (not valid)
761  * @pipe:	pipe to splice to
762  * @len:	number of bytes to splice
763  * @flags:	splice modifier flags
764  *
765  * Description:
766  *    Will read pages from given socket and fill them into a pipe.
767  *
768  **/
tcp_splice_read(struct socket * sock,loff_t * ppos,struct pipe_inode_info * pipe,size_t len,unsigned int flags)769 ssize_t tcp_splice_read(struct socket *sock, loff_t *ppos,
770 			struct pipe_inode_info *pipe, size_t len,
771 			unsigned int flags)
772 {
773 	struct sock *sk = sock->sk;
774 	struct tcp_splice_state tss = {
775 		.pipe = pipe,
776 		.len = len,
777 		.flags = flags,
778 	};
779 	long timeo;
780 	ssize_t spliced;
781 	int ret;
782 
783 	sock_rps_record_flow(sk);
784 	/*
785 	 * We can't seek on a socket input
786 	 */
787 	if (unlikely(*ppos))
788 		return -ESPIPE;
789 
790 	ret = spliced = 0;
791 
792 	lock_sock(sk);
793 
794 	timeo = sock_rcvtimeo(sk, sock->file->f_flags & O_NONBLOCK);
795 	while (tss.len) {
796 		ret = __tcp_splice_read(sk, &tss);
797 		if (ret < 0)
798 			break;
799 		else if (!ret) {
800 			if (spliced)
801 				break;
802 			if (sock_flag(sk, SOCK_DONE))
803 				break;
804 			if (sk->sk_err) {
805 				ret = sock_error(sk);
806 				break;
807 			}
808 			if (sk->sk_shutdown & RCV_SHUTDOWN)
809 				break;
810 			if (sk->sk_state == TCP_CLOSE) {
811 				/*
812 				 * This occurs when user tries to read
813 				 * from never connected socket.
814 				 */
815 				ret = -ENOTCONN;
816 				break;
817 			}
818 			if (!timeo) {
819 				ret = -EAGAIN;
820 				break;
821 			}
822 			/* if __tcp_splice_read() got nothing while we have
823 			 * an skb in receive queue, we do not want to loop.
824 			 * This might happen with URG data.
825 			 */
826 			if (!skb_queue_empty(&sk->sk_receive_queue))
827 				break;
828 			sk_wait_data(sk, &timeo, NULL);
829 			if (signal_pending(current)) {
830 				ret = sock_intr_errno(timeo);
831 				break;
832 			}
833 			continue;
834 		}
835 		tss.len -= ret;
836 		spliced += ret;
837 
838 		if (!timeo)
839 			break;
840 		release_sock(sk);
841 		lock_sock(sk);
842 
843 		if (sk->sk_err || sk->sk_state == TCP_CLOSE ||
844 		    (sk->sk_shutdown & RCV_SHUTDOWN) ||
845 		    signal_pending(current))
846 			break;
847 	}
848 
849 	release_sock(sk);
850 
851 	if (spliced)
852 		return spliced;
853 
854 	return ret;
855 }
856 EXPORT_SYMBOL(tcp_splice_read);
857 
tcp_stream_alloc_skb(struct sock * sk,int size,gfp_t gfp,bool force_schedule)858 struct sk_buff *tcp_stream_alloc_skb(struct sock *sk, int size, gfp_t gfp,
859 				     bool force_schedule)
860 {
861 	struct sk_buff *skb;
862 
863 	skb = alloc_skb_fclone(size + MAX_TCP_HEADER, gfp);
864 	if (likely(skb)) {
865 		bool mem_scheduled;
866 
867 		skb->truesize = SKB_TRUESIZE(skb_end_offset(skb));
868 		if (force_schedule) {
869 			mem_scheduled = true;
870 			sk_forced_mem_schedule(sk, skb->truesize);
871 		} else {
872 			mem_scheduled = sk_wmem_schedule(sk, skb->truesize);
873 		}
874 		if (likely(mem_scheduled)) {
875 			skb_reserve(skb, MAX_TCP_HEADER);
876 			skb->ip_summed = CHECKSUM_PARTIAL;
877 			INIT_LIST_HEAD(&skb->tcp_tsorted_anchor);
878 			return skb;
879 		}
880 		__kfree_skb(skb);
881 	} else {
882 		sk->sk_prot->enter_memory_pressure(sk);
883 		sk_stream_moderate_sndbuf(sk);
884 	}
885 	return NULL;
886 }
887 
tcp_xmit_size_goal(struct sock * sk,u32 mss_now,int large_allowed)888 static unsigned int tcp_xmit_size_goal(struct sock *sk, u32 mss_now,
889 				       int large_allowed)
890 {
891 	struct tcp_sock *tp = tcp_sk(sk);
892 	u32 new_size_goal, size_goal;
893 
894 	if (!large_allowed)
895 		return mss_now;
896 
897 	/* Note : tcp_tso_autosize() will eventually split this later */
898 	new_size_goal = tcp_bound_to_half_wnd(tp, sk->sk_gso_max_size);
899 
900 	/* We try hard to avoid divides here */
901 	size_goal = tp->gso_segs * mss_now;
902 	if (unlikely(new_size_goal < size_goal ||
903 		     new_size_goal >= size_goal + mss_now)) {
904 		tp->gso_segs = min_t(u16, new_size_goal / mss_now,
905 				     sk->sk_gso_max_segs);
906 		size_goal = tp->gso_segs * mss_now;
907 	}
908 
909 	return max(size_goal, mss_now);
910 }
911 
tcp_send_mss(struct sock * sk,int * size_goal,int flags)912 int tcp_send_mss(struct sock *sk, int *size_goal, int flags)
913 {
914 	int mss_now;
915 
916 	mss_now = tcp_current_mss(sk);
917 	*size_goal = tcp_xmit_size_goal(sk, mss_now, !(flags & MSG_OOB));
918 
919 	return mss_now;
920 }
921 
922 /* In some cases, both sendpage() and sendmsg() could have added
923  * an skb to the write queue, but failed adding payload on it.
924  * We need to remove it to consume less memory, but more
925  * importantly be able to generate EPOLLOUT for Edge Trigger epoll()
926  * users.
927  */
tcp_remove_empty_skb(struct sock * sk)928 void tcp_remove_empty_skb(struct sock *sk)
929 {
930 	struct sk_buff *skb = tcp_write_queue_tail(sk);
931 
932 	if (skb && TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq) {
933 		tcp_unlink_write_queue(skb, sk);
934 		if (tcp_write_queue_empty(sk))
935 			tcp_chrono_stop(sk, TCP_CHRONO_BUSY);
936 		tcp_wmem_free_skb(sk, skb);
937 	}
938 }
939 
940 /* skb changing from pure zc to mixed, must charge zc */
tcp_downgrade_zcopy_pure(struct sock * sk,struct sk_buff * skb)941 static int tcp_downgrade_zcopy_pure(struct sock *sk, struct sk_buff *skb)
942 {
943 	if (unlikely(skb_zcopy_pure(skb))) {
944 		u32 extra = skb->truesize -
945 			    SKB_TRUESIZE(skb_end_offset(skb));
946 
947 		if (!sk_wmem_schedule(sk, extra))
948 			return -ENOMEM;
949 
950 		sk_mem_charge(sk, extra);
951 		skb_shinfo(skb)->flags &= ~SKBFL_PURE_ZEROCOPY;
952 	}
953 	return 0;
954 }
955 
956 
tcp_wmem_schedule(struct sock * sk,int copy)957 static int tcp_wmem_schedule(struct sock *sk, int copy)
958 {
959 	int left;
960 
961 	if (likely(sk_wmem_schedule(sk, copy)))
962 		return copy;
963 
964 	/* We could be in trouble if we have nothing queued.
965 	 * Use whatever is left in sk->sk_forward_alloc and tcp_wmem[0]
966 	 * to guarantee some progress.
967 	 */
968 	left = sock_net(sk)->ipv4.sysctl_tcp_wmem[0] - sk->sk_wmem_queued;
969 	if (left > 0)
970 		sk_forced_mem_schedule(sk, min(left, copy));
971 	return min(copy, sk->sk_forward_alloc);
972 }
973 
tcp_build_frag(struct sock * sk,int size_goal,int flags,struct page * page,int offset,size_t * size)974 static struct sk_buff *tcp_build_frag(struct sock *sk, int size_goal, int flags,
975 				      struct page *page, int offset, size_t *size)
976 {
977 	struct sk_buff *skb = tcp_write_queue_tail(sk);
978 	struct tcp_sock *tp = tcp_sk(sk);
979 	bool can_coalesce;
980 	int copy, i;
981 
982 	if (!skb || (copy = size_goal - skb->len) <= 0 ||
983 	    !tcp_skb_can_collapse_to(skb)) {
984 new_segment:
985 		if (!sk_stream_memory_free(sk))
986 			return NULL;
987 
988 		skb = tcp_stream_alloc_skb(sk, 0, sk->sk_allocation,
989 					   tcp_rtx_and_write_queues_empty(sk));
990 		if (!skb)
991 			return NULL;
992 
993 #ifdef CONFIG_TLS_DEVICE
994 		skb->decrypted = !!(flags & MSG_SENDPAGE_DECRYPTED);
995 #endif
996 		tcp_skb_entail(sk, skb);
997 		copy = size_goal;
998 	}
999 
1000 	if (copy > *size)
1001 		copy = *size;
1002 
1003 	i = skb_shinfo(skb)->nr_frags;
1004 	can_coalesce = skb_can_coalesce(skb, i, page, offset);
1005 	if (!can_coalesce && i >= READ_ONCE(sysctl_max_skb_frags)) {
1006 		tcp_mark_push(tp, skb);
1007 		goto new_segment;
1008 	}
1009 	if (tcp_downgrade_zcopy_pure(sk, skb))
1010 		return NULL;
1011 
1012 	copy = tcp_wmem_schedule(sk, copy);
1013 	if (!copy)
1014 		return NULL;
1015 
1016 	if (can_coalesce) {
1017 		skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1018 	} else {
1019 		get_page(page);
1020 		skb_fill_page_desc_noacc(skb, i, page, offset, copy);
1021 	}
1022 
1023 	if (!(flags & MSG_NO_SHARED_FRAGS))
1024 		skb_shinfo(skb)->flags |= SKBFL_SHARED_FRAG;
1025 
1026 	skb->len += copy;
1027 	skb->data_len += copy;
1028 	skb->truesize += copy;
1029 	sk_wmem_queued_add(sk, copy);
1030 	sk_mem_charge(sk, copy);
1031 	WRITE_ONCE(tp->write_seq, tp->write_seq + copy);
1032 	TCP_SKB_CB(skb)->end_seq += copy;
1033 	tcp_skb_pcount_set(skb, 0);
1034 
1035 	*size = copy;
1036 	return skb;
1037 }
1038 
do_tcp_sendpages(struct sock * sk,struct page * page,int offset,size_t size,int flags)1039 ssize_t do_tcp_sendpages(struct sock *sk, struct page *page, int offset,
1040 			 size_t size, int flags)
1041 {
1042 	struct tcp_sock *tp = tcp_sk(sk);
1043 	int mss_now, size_goal;
1044 	int err;
1045 	ssize_t copied;
1046 	long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
1047 
1048 	if (IS_ENABLED(CONFIG_DEBUG_VM) &&
1049 	    WARN_ONCE(!sendpage_ok(page),
1050 		      "page must not be a Slab one and have page_count > 0"))
1051 		return -EINVAL;
1052 
1053 	/* Wait for a connection to finish. One exception is TCP Fast Open
1054 	 * (passive side) where data is allowed to be sent before a connection
1055 	 * is fully established.
1056 	 */
1057 	if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) &&
1058 	    !tcp_passive_fastopen(sk)) {
1059 		err = sk_stream_wait_connect(sk, &timeo);
1060 		if (err != 0)
1061 			goto out_err;
1062 	}
1063 
1064 	sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
1065 
1066 	mss_now = tcp_send_mss(sk, &size_goal, flags);
1067 	copied = 0;
1068 
1069 	err = -EPIPE;
1070 	if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
1071 		goto out_err;
1072 
1073 	while (size > 0) {
1074 		struct sk_buff *skb;
1075 		size_t copy = size;
1076 
1077 		skb = tcp_build_frag(sk, size_goal, flags, page, offset, &copy);
1078 		if (!skb)
1079 			goto wait_for_space;
1080 
1081 		if (!copied)
1082 			TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
1083 
1084 		copied += copy;
1085 		offset += copy;
1086 		size -= copy;
1087 		if (!size)
1088 			goto out;
1089 
1090 		if (skb->len < size_goal || (flags & MSG_OOB))
1091 			continue;
1092 
1093 		if (forced_push(tp)) {
1094 			tcp_mark_push(tp, skb);
1095 			__tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
1096 		} else if (skb == tcp_send_head(sk))
1097 			tcp_push_one(sk, mss_now);
1098 		continue;
1099 
1100 wait_for_space:
1101 		set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1102 		tcp_push(sk, flags & ~MSG_MORE, mss_now,
1103 			 TCP_NAGLE_PUSH, size_goal);
1104 
1105 		err = sk_stream_wait_memory(sk, &timeo);
1106 		if (err != 0)
1107 			goto do_error;
1108 
1109 		mss_now = tcp_send_mss(sk, &size_goal, flags);
1110 	}
1111 
1112 out:
1113 	if (copied) {
1114 		tcp_tx_timestamp(sk, sk->sk_tsflags);
1115 		if (!(flags & MSG_SENDPAGE_NOTLAST))
1116 			tcp_push(sk, flags, mss_now, tp->nonagle, size_goal);
1117 	}
1118 	return copied;
1119 
1120 do_error:
1121 	tcp_remove_empty_skb(sk);
1122 	if (copied)
1123 		goto out;
1124 out_err:
1125 	/* make sure we wake any epoll edge trigger waiter */
1126 	if (unlikely(tcp_rtx_and_write_queues_empty(sk) && err == -EAGAIN)) {
1127 		sk->sk_write_space(sk);
1128 		tcp_chrono_stop(sk, TCP_CHRONO_SNDBUF_LIMITED);
1129 	}
1130 	return sk_stream_error(sk, flags, err);
1131 }
1132 EXPORT_SYMBOL_GPL(do_tcp_sendpages);
1133 
tcp_sendpage_locked(struct sock * sk,struct page * page,int offset,size_t size,int flags)1134 int tcp_sendpage_locked(struct sock *sk, struct page *page, int offset,
1135 			size_t size, int flags)
1136 {
1137 	if (!(sk->sk_route_caps & NETIF_F_SG))
1138 		return sock_no_sendpage_locked(sk, page, offset, size, flags);
1139 
1140 	tcp_rate_check_app_limited(sk);  /* is sending application-limited? */
1141 
1142 	return do_tcp_sendpages(sk, page, offset, size, flags);
1143 }
1144 EXPORT_SYMBOL_GPL(tcp_sendpage_locked);
1145 
tcp_sendpage(struct sock * sk,struct page * page,int offset,size_t size,int flags)1146 int tcp_sendpage(struct sock *sk, struct page *page, int offset,
1147 		 size_t size, int flags)
1148 {
1149 	int ret;
1150 
1151 	lock_sock(sk);
1152 	ret = tcp_sendpage_locked(sk, page, offset, size, flags);
1153 	release_sock(sk);
1154 
1155 	return ret;
1156 }
1157 EXPORT_SYMBOL(tcp_sendpage);
1158 
tcp_free_fastopen_req(struct tcp_sock * tp)1159 void tcp_free_fastopen_req(struct tcp_sock *tp)
1160 {
1161 	if (tp->fastopen_req) {
1162 		kfree(tp->fastopen_req);
1163 		tp->fastopen_req = NULL;
1164 	}
1165 }
1166 
tcp_sendmsg_fastopen(struct sock * sk,struct msghdr * msg,int * copied,size_t size,struct ubuf_info * uarg)1167 int tcp_sendmsg_fastopen(struct sock *sk, struct msghdr *msg, int *copied,
1168 			 size_t size, struct ubuf_info *uarg)
1169 {
1170 	struct tcp_sock *tp = tcp_sk(sk);
1171 	struct inet_sock *inet = inet_sk(sk);
1172 	struct sockaddr *uaddr = msg->msg_name;
1173 	int err, flags;
1174 
1175 	if (!(READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_fastopen) &
1176 	      TFO_CLIENT_ENABLE) ||
1177 	    (uaddr && msg->msg_namelen >= sizeof(uaddr->sa_family) &&
1178 	     uaddr->sa_family == AF_UNSPEC))
1179 		return -EOPNOTSUPP;
1180 	if (tp->fastopen_req)
1181 		return -EALREADY; /* Another Fast Open is in progress */
1182 
1183 	tp->fastopen_req = kzalloc(sizeof(struct tcp_fastopen_request),
1184 				   sk->sk_allocation);
1185 	if (unlikely(!tp->fastopen_req))
1186 		return -ENOBUFS;
1187 	tp->fastopen_req->data = msg;
1188 	tp->fastopen_req->size = size;
1189 	tp->fastopen_req->uarg = uarg;
1190 
1191 	if (inet->defer_connect) {
1192 		err = tcp_connect(sk);
1193 		/* Same failure procedure as in tcp_v4/6_connect */
1194 		if (err) {
1195 			tcp_set_state(sk, TCP_CLOSE);
1196 			inet->inet_dport = 0;
1197 			sk->sk_route_caps = 0;
1198 		}
1199 	}
1200 	flags = (msg->msg_flags & MSG_DONTWAIT) ? O_NONBLOCK : 0;
1201 	err = __inet_stream_connect(sk->sk_socket, uaddr,
1202 				    msg->msg_namelen, flags, 1);
1203 	/* fastopen_req could already be freed in __inet_stream_connect
1204 	 * if the connection times out or gets rst
1205 	 */
1206 	if (tp->fastopen_req) {
1207 		*copied = tp->fastopen_req->copied;
1208 		tcp_free_fastopen_req(tp);
1209 		inet->defer_connect = 0;
1210 	}
1211 	return err;
1212 }
1213 
tcp_sendmsg_locked(struct sock * sk,struct msghdr * msg,size_t size)1214 int tcp_sendmsg_locked(struct sock *sk, struct msghdr *msg, size_t size)
1215 {
1216 	struct tcp_sock *tp = tcp_sk(sk);
1217 	struct ubuf_info *uarg = NULL;
1218 	struct sk_buff *skb;
1219 	struct sockcm_cookie sockc;
1220 	int flags, err, copied = 0;
1221 	int mss_now = 0, size_goal, copied_syn = 0;
1222 	int process_backlog = 0;
1223 	bool zc = false;
1224 	long timeo;
1225 
1226 	flags = msg->msg_flags;
1227 
1228 	if ((flags & MSG_ZEROCOPY) && size) {
1229 		skb = tcp_write_queue_tail(sk);
1230 
1231 		if (msg->msg_ubuf) {
1232 			uarg = msg->msg_ubuf;
1233 			net_zcopy_get(uarg);
1234 			zc = sk->sk_route_caps & NETIF_F_SG;
1235 		} else if (sock_flag(sk, SOCK_ZEROCOPY)) {
1236 			uarg = msg_zerocopy_realloc(sk, size, skb_zcopy(skb));
1237 			if (!uarg) {
1238 				err = -ENOBUFS;
1239 				goto out_err;
1240 			}
1241 			zc = sk->sk_route_caps & NETIF_F_SG;
1242 			if (!zc)
1243 				uarg_to_msgzc(uarg)->zerocopy = 0;
1244 		}
1245 	}
1246 
1247 	if (unlikely(flags & MSG_FASTOPEN || inet_sk(sk)->defer_connect) &&
1248 	    !tp->repair) {
1249 		err = tcp_sendmsg_fastopen(sk, msg, &copied_syn, size, uarg);
1250 		if (err == -EINPROGRESS && copied_syn > 0)
1251 			goto out;
1252 		else if (err)
1253 			goto out_err;
1254 	}
1255 
1256 	timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
1257 
1258 	tcp_rate_check_app_limited(sk);  /* is sending application-limited? */
1259 
1260 	/* Wait for a connection to finish. One exception is TCP Fast Open
1261 	 * (passive side) where data is allowed to be sent before a connection
1262 	 * is fully established.
1263 	 */
1264 	if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) &&
1265 	    !tcp_passive_fastopen(sk)) {
1266 		err = sk_stream_wait_connect(sk, &timeo);
1267 		if (err != 0)
1268 			goto do_error;
1269 	}
1270 
1271 	if (unlikely(tp->repair)) {
1272 		if (tp->repair_queue == TCP_RECV_QUEUE) {
1273 			copied = tcp_send_rcvq(sk, msg, size);
1274 			goto out_nopush;
1275 		}
1276 
1277 		err = -EINVAL;
1278 		if (tp->repair_queue == TCP_NO_QUEUE)
1279 			goto out_err;
1280 
1281 		/* 'common' sending to sendq */
1282 	}
1283 
1284 	sockcm_init(&sockc, sk);
1285 	if (msg->msg_controllen) {
1286 		err = sock_cmsg_send(sk, msg, &sockc);
1287 		if (unlikely(err)) {
1288 			err = -EINVAL;
1289 			goto out_err;
1290 		}
1291 	}
1292 
1293 	/* This should be in poll */
1294 	sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
1295 
1296 	/* Ok commence sending. */
1297 	copied = 0;
1298 
1299 restart:
1300 	mss_now = tcp_send_mss(sk, &size_goal, flags);
1301 
1302 	err = -EPIPE;
1303 	if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
1304 		goto do_error;
1305 
1306 	while (msg_data_left(msg)) {
1307 		int copy = 0;
1308 
1309 		skb = tcp_write_queue_tail(sk);
1310 		if (skb)
1311 			copy = size_goal - skb->len;
1312 
1313 		if (copy <= 0 || !tcp_skb_can_collapse_to(skb)) {
1314 			bool first_skb;
1315 
1316 new_segment:
1317 			if (!sk_stream_memory_free(sk))
1318 				goto wait_for_space;
1319 
1320 			if (unlikely(process_backlog >= 16)) {
1321 				process_backlog = 0;
1322 				if (sk_flush_backlog(sk))
1323 					goto restart;
1324 			}
1325 			first_skb = tcp_rtx_and_write_queues_empty(sk);
1326 			skb = tcp_stream_alloc_skb(sk, 0, sk->sk_allocation,
1327 						   first_skb);
1328 			if (!skb)
1329 				goto wait_for_space;
1330 
1331 			process_backlog++;
1332 
1333 			tcp_skb_entail(sk, skb);
1334 			copy = size_goal;
1335 
1336 			/* All packets are restored as if they have
1337 			 * already been sent. skb_mstamp_ns isn't set to
1338 			 * avoid wrong rtt estimation.
1339 			 */
1340 			if (tp->repair)
1341 				TCP_SKB_CB(skb)->sacked |= TCPCB_REPAIRED;
1342 		}
1343 
1344 		/* Try to append data to the end of skb. */
1345 		if (copy > msg_data_left(msg))
1346 			copy = msg_data_left(msg);
1347 
1348 		if (!zc) {
1349 			bool merge = true;
1350 			int i = skb_shinfo(skb)->nr_frags;
1351 			struct page_frag *pfrag = sk_page_frag(sk);
1352 
1353 			if (!sk_page_frag_refill(sk, pfrag))
1354 				goto wait_for_space;
1355 
1356 			if (!skb_can_coalesce(skb, i, pfrag->page,
1357 					      pfrag->offset)) {
1358 				if (i >= READ_ONCE(sysctl_max_skb_frags)) {
1359 					tcp_mark_push(tp, skb);
1360 					goto new_segment;
1361 				}
1362 				merge = false;
1363 			}
1364 
1365 			copy = min_t(int, copy, pfrag->size - pfrag->offset);
1366 
1367 			if (unlikely(skb_zcopy_pure(skb) || skb_zcopy_managed(skb))) {
1368 				if (tcp_downgrade_zcopy_pure(sk, skb))
1369 					goto wait_for_space;
1370 				skb_zcopy_downgrade_managed(skb);
1371 			}
1372 
1373 			copy = tcp_wmem_schedule(sk, copy);
1374 			if (!copy)
1375 				goto wait_for_space;
1376 
1377 			err = skb_copy_to_page_nocache(sk, &msg->msg_iter, skb,
1378 						       pfrag->page,
1379 						       pfrag->offset,
1380 						       copy);
1381 			if (err)
1382 				goto do_error;
1383 
1384 			/* Update the skb. */
1385 			if (merge) {
1386 				skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1387 			} else {
1388 				skb_fill_page_desc(skb, i, pfrag->page,
1389 						   pfrag->offset, copy);
1390 				page_ref_inc(pfrag->page);
1391 			}
1392 			pfrag->offset += copy;
1393 		} else {
1394 			/* First append to a fragless skb builds initial
1395 			 * pure zerocopy skb
1396 			 */
1397 			if (!skb->len)
1398 				skb_shinfo(skb)->flags |= SKBFL_PURE_ZEROCOPY;
1399 
1400 			if (!skb_zcopy_pure(skb)) {
1401 				copy = tcp_wmem_schedule(sk, copy);
1402 				if (!copy)
1403 					goto wait_for_space;
1404 			}
1405 
1406 			err = skb_zerocopy_iter_stream(sk, skb, msg, copy, uarg);
1407 			if (err == -EMSGSIZE || err == -EEXIST) {
1408 				tcp_mark_push(tp, skb);
1409 				goto new_segment;
1410 			}
1411 			if (err < 0)
1412 				goto do_error;
1413 			copy = err;
1414 		}
1415 
1416 		if (!copied)
1417 			TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
1418 
1419 		WRITE_ONCE(tp->write_seq, tp->write_seq + copy);
1420 		TCP_SKB_CB(skb)->end_seq += copy;
1421 		tcp_skb_pcount_set(skb, 0);
1422 
1423 		copied += copy;
1424 		if (!msg_data_left(msg)) {
1425 			if (unlikely(flags & MSG_EOR))
1426 				TCP_SKB_CB(skb)->eor = 1;
1427 			goto out;
1428 		}
1429 
1430 		if (skb->len < size_goal || (flags & MSG_OOB) || unlikely(tp->repair))
1431 			continue;
1432 
1433 		if (forced_push(tp)) {
1434 			tcp_mark_push(tp, skb);
1435 			__tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
1436 		} else if (skb == tcp_send_head(sk))
1437 			tcp_push_one(sk, mss_now);
1438 		continue;
1439 
1440 wait_for_space:
1441 		set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1442 		if (copied)
1443 			tcp_push(sk, flags & ~MSG_MORE, mss_now,
1444 				 TCP_NAGLE_PUSH, size_goal);
1445 
1446 		err = sk_stream_wait_memory(sk, &timeo);
1447 		if (err != 0)
1448 			goto do_error;
1449 
1450 		mss_now = tcp_send_mss(sk, &size_goal, flags);
1451 	}
1452 
1453 out:
1454 	if (copied) {
1455 		tcp_tx_timestamp(sk, sockc.tsflags);
1456 		tcp_push(sk, flags, mss_now, tp->nonagle, size_goal);
1457 	}
1458 out_nopush:
1459 	net_zcopy_put(uarg);
1460 	return copied + copied_syn;
1461 
1462 do_error:
1463 	tcp_remove_empty_skb(sk);
1464 
1465 	if (copied + copied_syn)
1466 		goto out;
1467 out_err:
1468 	net_zcopy_put_abort(uarg, true);
1469 	err = sk_stream_error(sk, flags, err);
1470 	/* make sure we wake any epoll edge trigger waiter */
1471 	if (unlikely(tcp_rtx_and_write_queues_empty(sk) && err == -EAGAIN)) {
1472 		sk->sk_write_space(sk);
1473 		tcp_chrono_stop(sk, TCP_CHRONO_SNDBUF_LIMITED);
1474 	}
1475 	return err;
1476 }
1477 EXPORT_SYMBOL_GPL(tcp_sendmsg_locked);
1478 
tcp_sendmsg(struct sock * sk,struct msghdr * msg,size_t size)1479 int tcp_sendmsg(struct sock *sk, struct msghdr *msg, size_t size)
1480 {
1481 	int ret;
1482 
1483 	lock_sock(sk);
1484 	ret = tcp_sendmsg_locked(sk, msg, size);
1485 	release_sock(sk);
1486 
1487 	return ret;
1488 }
1489 EXPORT_SYMBOL(tcp_sendmsg);
1490 
1491 /*
1492  *	Handle reading urgent data. BSD has very simple semantics for
1493  *	this, no blocking and very strange errors 8)
1494  */
1495 
tcp_recv_urg(struct sock * sk,struct msghdr * msg,int len,int flags)1496 static int tcp_recv_urg(struct sock *sk, struct msghdr *msg, int len, int flags)
1497 {
1498 	struct tcp_sock *tp = tcp_sk(sk);
1499 
1500 	/* No URG data to read. */
1501 	if (sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data ||
1502 	    tp->urg_data == TCP_URG_READ)
1503 		return -EINVAL;	/* Yes this is right ! */
1504 
1505 	if (sk->sk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DONE))
1506 		return -ENOTCONN;
1507 
1508 	if (tp->urg_data & TCP_URG_VALID) {
1509 		int err = 0;
1510 		char c = tp->urg_data;
1511 
1512 		if (!(flags & MSG_PEEK))
1513 			WRITE_ONCE(tp->urg_data, TCP_URG_READ);
1514 
1515 		/* Read urgent data. */
1516 		msg->msg_flags |= MSG_OOB;
1517 
1518 		if (len > 0) {
1519 			if (!(flags & MSG_TRUNC))
1520 				err = memcpy_to_msg(msg, &c, 1);
1521 			len = 1;
1522 		} else
1523 			msg->msg_flags |= MSG_TRUNC;
1524 
1525 		return err ? -EFAULT : len;
1526 	}
1527 
1528 	if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
1529 		return 0;
1530 
1531 	/* Fixed the recv(..., MSG_OOB) behaviour.  BSD docs and
1532 	 * the available implementations agree in this case:
1533 	 * this call should never block, independent of the
1534 	 * blocking state of the socket.
1535 	 * Mike <pall@rz.uni-karlsruhe.de>
1536 	 */
1537 	return -EAGAIN;
1538 }
1539 
tcp_peek_sndq(struct sock * sk,struct msghdr * msg,int len)1540 static int tcp_peek_sndq(struct sock *sk, struct msghdr *msg, int len)
1541 {
1542 	struct sk_buff *skb;
1543 	int copied = 0, err = 0;
1544 
1545 	/* XXX -- need to support SO_PEEK_OFF */
1546 
1547 	skb_rbtree_walk(skb, &sk->tcp_rtx_queue) {
1548 		err = skb_copy_datagram_msg(skb, 0, msg, skb->len);
1549 		if (err)
1550 			return err;
1551 		copied += skb->len;
1552 	}
1553 
1554 	skb_queue_walk(&sk->sk_write_queue, skb) {
1555 		err = skb_copy_datagram_msg(skb, 0, msg, skb->len);
1556 		if (err)
1557 			break;
1558 
1559 		copied += skb->len;
1560 	}
1561 
1562 	return err ?: copied;
1563 }
1564 
1565 /* Clean up the receive buffer for full frames taken by the user,
1566  * then send an ACK if necessary.  COPIED is the number of bytes
1567  * tcp_recvmsg has given to the user so far, it speeds up the
1568  * calculation of whether or not we must ACK for the sake of
1569  * a window update.
1570  */
__tcp_cleanup_rbuf(struct sock * sk,int copied)1571 static void __tcp_cleanup_rbuf(struct sock *sk, int copied)
1572 {
1573 	struct tcp_sock *tp = tcp_sk(sk);
1574 	bool time_to_ack = false;
1575 
1576 	if (inet_csk_ack_scheduled(sk)) {
1577 		const struct inet_connection_sock *icsk = inet_csk(sk);
1578 
1579 		if (/* Once-per-two-segments ACK was not sent by tcp_input.c */
1580 		    tp->rcv_nxt - tp->rcv_wup > icsk->icsk_ack.rcv_mss ||
1581 		    /*
1582 		     * If this read emptied read buffer, we send ACK, if
1583 		     * connection is not bidirectional, user drained
1584 		     * receive buffer and there was a small segment
1585 		     * in queue.
1586 		     */
1587 		    (copied > 0 &&
1588 		     ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED2) ||
1589 		      ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED) &&
1590 		       !inet_csk_in_pingpong_mode(sk))) &&
1591 		      !atomic_read(&sk->sk_rmem_alloc)))
1592 			time_to_ack = true;
1593 	}
1594 
1595 	/* We send an ACK if we can now advertise a non-zero window
1596 	 * which has been raised "significantly".
1597 	 *
1598 	 * Even if window raised up to infinity, do not send window open ACK
1599 	 * in states, where we will not receive more. It is useless.
1600 	 */
1601 	if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) {
1602 		__u32 rcv_window_now = tcp_receive_window(tp);
1603 
1604 		/* Optimize, __tcp_select_window() is not cheap. */
1605 		if (2*rcv_window_now <= tp->window_clamp) {
1606 			__u32 new_window = __tcp_select_window(sk);
1607 
1608 			/* Send ACK now, if this read freed lots of space
1609 			 * in our buffer. Certainly, new_window is new window.
1610 			 * We can advertise it now, if it is not less than current one.
1611 			 * "Lots" means "at least twice" here.
1612 			 */
1613 			if (new_window && new_window >= 2 * rcv_window_now)
1614 				time_to_ack = true;
1615 		}
1616 	}
1617 	if (time_to_ack)
1618 		tcp_send_ack(sk);
1619 }
1620 
tcp_cleanup_rbuf(struct sock * sk,int copied)1621 void tcp_cleanup_rbuf(struct sock *sk, int copied)
1622 {
1623 	struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
1624 	struct tcp_sock *tp = tcp_sk(sk);
1625 
1626 	WARN(skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq),
1627 	     "cleanup rbuf bug: copied %X seq %X rcvnxt %X\n",
1628 	     tp->copied_seq, TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt);
1629 	__tcp_cleanup_rbuf(sk, copied);
1630 }
1631 
tcp_eat_recv_skb(struct sock * sk,struct sk_buff * skb)1632 static void tcp_eat_recv_skb(struct sock *sk, struct sk_buff *skb)
1633 {
1634 	__skb_unlink(skb, &sk->sk_receive_queue);
1635 	if (likely(skb->destructor == sock_rfree)) {
1636 		sock_rfree(skb);
1637 		skb->destructor = NULL;
1638 		skb->sk = NULL;
1639 		return skb_attempt_defer_free(skb);
1640 	}
1641 	__kfree_skb(skb);
1642 }
1643 
tcp_recv_skb(struct sock * sk,u32 seq,u32 * off)1644 struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off)
1645 {
1646 	struct sk_buff *skb;
1647 	u32 offset;
1648 
1649 	while ((skb = skb_peek(&sk->sk_receive_queue)) != NULL) {
1650 		offset = seq - TCP_SKB_CB(skb)->seq;
1651 		if (unlikely(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
1652 			pr_err_once("%s: found a SYN, please report !\n", __func__);
1653 			offset--;
1654 		}
1655 		if (offset < skb->len || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)) {
1656 			*off = offset;
1657 			return skb;
1658 		}
1659 		/* This looks weird, but this can happen if TCP collapsing
1660 		 * splitted a fat GRO packet, while we released socket lock
1661 		 * in skb_splice_bits()
1662 		 */
1663 		tcp_eat_recv_skb(sk, skb);
1664 	}
1665 	return NULL;
1666 }
1667 EXPORT_SYMBOL(tcp_recv_skb);
1668 
1669 /*
1670  * This routine provides an alternative to tcp_recvmsg() for routines
1671  * that would like to handle copying from skbuffs directly in 'sendfile'
1672  * fashion.
1673  * Note:
1674  *	- It is assumed that the socket was locked by the caller.
1675  *	- The routine does not block.
1676  *	- At present, there is no support for reading OOB data
1677  *	  or for 'peeking' the socket using this routine
1678  *	  (although both would be easy to implement).
1679  */
tcp_read_sock(struct sock * sk,read_descriptor_t * desc,sk_read_actor_t recv_actor)1680 int tcp_read_sock(struct sock *sk, read_descriptor_t *desc,
1681 		  sk_read_actor_t recv_actor)
1682 {
1683 	struct sk_buff *skb;
1684 	struct tcp_sock *tp = tcp_sk(sk);
1685 	u32 seq = tp->copied_seq;
1686 	u32 offset;
1687 	int copied = 0;
1688 
1689 	if (sk->sk_state == TCP_LISTEN)
1690 		return -ENOTCONN;
1691 	while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1692 		if (offset < skb->len) {
1693 			int used;
1694 			size_t len;
1695 
1696 			len = skb->len - offset;
1697 			/* Stop reading if we hit a patch of urgent data */
1698 			if (unlikely(tp->urg_data)) {
1699 				u32 urg_offset = tp->urg_seq - seq;
1700 				if (urg_offset < len)
1701 					len = urg_offset;
1702 				if (!len)
1703 					break;
1704 			}
1705 			used = recv_actor(desc, skb, offset, len);
1706 			if (used <= 0) {
1707 				if (!copied)
1708 					copied = used;
1709 				break;
1710 			}
1711 			if (WARN_ON_ONCE(used > len))
1712 				used = len;
1713 			seq += used;
1714 			copied += used;
1715 			offset += used;
1716 
1717 			/* If recv_actor drops the lock (e.g. TCP splice
1718 			 * receive) the skb pointer might be invalid when
1719 			 * getting here: tcp_collapse might have deleted it
1720 			 * while aggregating skbs from the socket queue.
1721 			 */
1722 			skb = tcp_recv_skb(sk, seq - 1, &offset);
1723 			if (!skb)
1724 				break;
1725 			/* TCP coalescing might have appended data to the skb.
1726 			 * Try to splice more frags
1727 			 */
1728 			if (offset + 1 != skb->len)
1729 				continue;
1730 		}
1731 		if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) {
1732 			tcp_eat_recv_skb(sk, skb);
1733 			++seq;
1734 			break;
1735 		}
1736 		tcp_eat_recv_skb(sk, skb);
1737 		if (!desc->count)
1738 			break;
1739 		WRITE_ONCE(tp->copied_seq, seq);
1740 	}
1741 	WRITE_ONCE(tp->copied_seq, seq);
1742 
1743 	tcp_rcv_space_adjust(sk);
1744 
1745 	/* Clean up data we have read: This will do ACK frames. */
1746 	if (copied > 0) {
1747 		tcp_recv_skb(sk, seq, &offset);
1748 		tcp_cleanup_rbuf(sk, copied);
1749 	}
1750 	return copied;
1751 }
1752 EXPORT_SYMBOL(tcp_read_sock);
1753 
tcp_read_skb(struct sock * sk,skb_read_actor_t recv_actor)1754 int tcp_read_skb(struct sock *sk, skb_read_actor_t recv_actor)
1755 {
1756 	struct tcp_sock *tp = tcp_sk(sk);
1757 	u32 seq = tp->copied_seq;
1758 	struct sk_buff *skb;
1759 	int copied = 0;
1760 	u32 offset;
1761 
1762 	if (sk->sk_state == TCP_LISTEN)
1763 		return -ENOTCONN;
1764 
1765 	while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1766 		u8 tcp_flags;
1767 		int used;
1768 
1769 		__skb_unlink(skb, &sk->sk_receive_queue);
1770 		WARN_ON_ONCE(!skb_set_owner_sk_safe(skb, sk));
1771 		tcp_flags = TCP_SKB_CB(skb)->tcp_flags;
1772 		used = recv_actor(sk, skb);
1773 		consume_skb(skb);
1774 		if (used < 0) {
1775 			if (!copied)
1776 				copied = used;
1777 			break;
1778 		}
1779 		seq += used;
1780 		copied += used;
1781 
1782 		if (tcp_flags & TCPHDR_FIN) {
1783 			++seq;
1784 			break;
1785 		}
1786 	}
1787 	WRITE_ONCE(tp->copied_seq, seq);
1788 
1789 	tcp_rcv_space_adjust(sk);
1790 
1791 	/* Clean up data we have read: This will do ACK frames. */
1792 	if (copied > 0)
1793 		__tcp_cleanup_rbuf(sk, copied);
1794 
1795 	return copied;
1796 }
1797 EXPORT_SYMBOL(tcp_read_skb);
1798 
tcp_read_done(struct sock * sk,size_t len)1799 void tcp_read_done(struct sock *sk, size_t len)
1800 {
1801 	struct tcp_sock *tp = tcp_sk(sk);
1802 	u32 seq = tp->copied_seq;
1803 	struct sk_buff *skb;
1804 	size_t left;
1805 	u32 offset;
1806 
1807 	if (sk->sk_state == TCP_LISTEN)
1808 		return;
1809 
1810 	left = len;
1811 	while (left && (skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1812 		int used;
1813 
1814 		used = min_t(size_t, skb->len - offset, left);
1815 		seq += used;
1816 		left -= used;
1817 
1818 		if (skb->len > offset + used)
1819 			break;
1820 
1821 		if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) {
1822 			tcp_eat_recv_skb(sk, skb);
1823 			++seq;
1824 			break;
1825 		}
1826 		tcp_eat_recv_skb(sk, skb);
1827 	}
1828 	WRITE_ONCE(tp->copied_seq, seq);
1829 
1830 	tcp_rcv_space_adjust(sk);
1831 
1832 	/* Clean up data we have read: This will do ACK frames. */
1833 	if (left != len)
1834 		tcp_cleanup_rbuf(sk, len - left);
1835 }
1836 EXPORT_SYMBOL(tcp_read_done);
1837 
tcp_peek_len(struct socket * sock)1838 int tcp_peek_len(struct socket *sock)
1839 {
1840 	return tcp_inq(sock->sk);
1841 }
1842 EXPORT_SYMBOL(tcp_peek_len);
1843 
1844 /* Make sure sk_rcvbuf is big enough to satisfy SO_RCVLOWAT hint */
tcp_set_rcvlowat(struct sock * sk,int val)1845 int tcp_set_rcvlowat(struct sock *sk, int val)
1846 {
1847 	int cap;
1848 
1849 	if (sk->sk_userlocks & SOCK_RCVBUF_LOCK)
1850 		cap = sk->sk_rcvbuf >> 1;
1851 	else
1852 		cap = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_rmem[2]) >> 1;
1853 	val = min(val, cap);
1854 	WRITE_ONCE(sk->sk_rcvlowat, val ? : 1);
1855 
1856 	/* Check if we need to signal EPOLLIN right now */
1857 	tcp_data_ready(sk);
1858 
1859 	if (sk->sk_userlocks & SOCK_RCVBUF_LOCK)
1860 		return 0;
1861 
1862 	val <<= 1;
1863 	if (val > sk->sk_rcvbuf) {
1864 		WRITE_ONCE(sk->sk_rcvbuf, val);
1865 		tcp_sk(sk)->window_clamp = tcp_win_from_space(sk, val);
1866 	}
1867 	return 0;
1868 }
1869 EXPORT_SYMBOL(tcp_set_rcvlowat);
1870 
tcp_update_recv_tstamps(struct sk_buff * skb,struct scm_timestamping_internal * tss)1871 void tcp_update_recv_tstamps(struct sk_buff *skb,
1872 			     struct scm_timestamping_internal *tss)
1873 {
1874 	if (skb->tstamp)
1875 		tss->ts[0] = ktime_to_timespec64(skb->tstamp);
1876 	else
1877 		tss->ts[0] = (struct timespec64) {0};
1878 
1879 	if (skb_hwtstamps(skb)->hwtstamp)
1880 		tss->ts[2] = ktime_to_timespec64(skb_hwtstamps(skb)->hwtstamp);
1881 	else
1882 		tss->ts[2] = (struct timespec64) {0};
1883 }
1884 
1885 #ifdef CONFIG_MMU
1886 static const struct vm_operations_struct tcp_vm_ops = {
1887 };
1888 
tcp_mmap(struct file * file,struct socket * sock,struct vm_area_struct * vma)1889 int tcp_mmap(struct file *file, struct socket *sock,
1890 	     struct vm_area_struct *vma)
1891 {
1892 	if (vma->vm_flags & (VM_WRITE | VM_EXEC))
1893 		return -EPERM;
1894 	vma->vm_flags &= ~(VM_MAYWRITE | VM_MAYEXEC);
1895 
1896 	/* Instruct vm_insert_page() to not mmap_read_lock(mm) */
1897 	vma->vm_flags |= VM_MIXEDMAP;
1898 
1899 	vma->vm_ops = &tcp_vm_ops;
1900 	return 0;
1901 }
1902 EXPORT_SYMBOL(tcp_mmap);
1903 
skb_advance_to_frag(struct sk_buff * skb,u32 offset_skb,u32 * offset_frag)1904 static skb_frag_t *skb_advance_to_frag(struct sk_buff *skb, u32 offset_skb,
1905 				       u32 *offset_frag)
1906 {
1907 	skb_frag_t *frag;
1908 
1909 	if (unlikely(offset_skb >= skb->len))
1910 		return NULL;
1911 
1912 	offset_skb -= skb_headlen(skb);
1913 	if ((int)offset_skb < 0 || skb_has_frag_list(skb))
1914 		return NULL;
1915 
1916 	frag = skb_shinfo(skb)->frags;
1917 	while (offset_skb) {
1918 		if (skb_frag_size(frag) > offset_skb) {
1919 			*offset_frag = offset_skb;
1920 			return frag;
1921 		}
1922 		offset_skb -= skb_frag_size(frag);
1923 		++frag;
1924 	}
1925 	*offset_frag = 0;
1926 	return frag;
1927 }
1928 
can_map_frag(const skb_frag_t * frag)1929 static bool can_map_frag(const skb_frag_t *frag)
1930 {
1931 	return skb_frag_size(frag) == PAGE_SIZE && !skb_frag_off(frag);
1932 }
1933 
find_next_mappable_frag(const skb_frag_t * frag,int remaining_in_skb)1934 static int find_next_mappable_frag(const skb_frag_t *frag,
1935 				   int remaining_in_skb)
1936 {
1937 	int offset = 0;
1938 
1939 	if (likely(can_map_frag(frag)))
1940 		return 0;
1941 
1942 	while (offset < remaining_in_skb && !can_map_frag(frag)) {
1943 		offset += skb_frag_size(frag);
1944 		++frag;
1945 	}
1946 	return offset;
1947 }
1948 
tcp_zerocopy_set_hint_for_skb(struct sock * sk,struct tcp_zerocopy_receive * zc,struct sk_buff * skb,u32 offset)1949 static void tcp_zerocopy_set_hint_for_skb(struct sock *sk,
1950 					  struct tcp_zerocopy_receive *zc,
1951 					  struct sk_buff *skb, u32 offset)
1952 {
1953 	u32 frag_offset, partial_frag_remainder = 0;
1954 	int mappable_offset;
1955 	skb_frag_t *frag;
1956 
1957 	/* worst case: skip to next skb. try to improve on this case below */
1958 	zc->recv_skip_hint = skb->len - offset;
1959 
1960 	/* Find the frag containing this offset (and how far into that frag) */
1961 	frag = skb_advance_to_frag(skb, offset, &frag_offset);
1962 	if (!frag)
1963 		return;
1964 
1965 	if (frag_offset) {
1966 		struct skb_shared_info *info = skb_shinfo(skb);
1967 
1968 		/* We read part of the last frag, must recvmsg() rest of skb. */
1969 		if (frag == &info->frags[info->nr_frags - 1])
1970 			return;
1971 
1972 		/* Else, we must at least read the remainder in this frag. */
1973 		partial_frag_remainder = skb_frag_size(frag) - frag_offset;
1974 		zc->recv_skip_hint -= partial_frag_remainder;
1975 		++frag;
1976 	}
1977 
1978 	/* partial_frag_remainder: If part way through a frag, must read rest.
1979 	 * mappable_offset: Bytes till next mappable frag, *not* counting bytes
1980 	 * in partial_frag_remainder.
1981 	 */
1982 	mappable_offset = find_next_mappable_frag(frag, zc->recv_skip_hint);
1983 	zc->recv_skip_hint = mappable_offset + partial_frag_remainder;
1984 }
1985 
1986 static int tcp_recvmsg_locked(struct sock *sk, struct msghdr *msg, size_t len,
1987 			      int flags, struct scm_timestamping_internal *tss,
1988 			      int *cmsg_flags);
receive_fallback_to_copy(struct sock * sk,struct tcp_zerocopy_receive * zc,int inq,struct scm_timestamping_internal * tss)1989 static int receive_fallback_to_copy(struct sock *sk,
1990 				    struct tcp_zerocopy_receive *zc, int inq,
1991 				    struct scm_timestamping_internal *tss)
1992 {
1993 	unsigned long copy_address = (unsigned long)zc->copybuf_address;
1994 	struct msghdr msg = {};
1995 	struct iovec iov;
1996 	int err;
1997 
1998 	zc->length = 0;
1999 	zc->recv_skip_hint = 0;
2000 
2001 	if (copy_address != zc->copybuf_address)
2002 		return -EINVAL;
2003 
2004 	err = import_single_range(READ, (void __user *)copy_address,
2005 				  inq, &iov, &msg.msg_iter);
2006 	if (err)
2007 		return err;
2008 
2009 	err = tcp_recvmsg_locked(sk, &msg, inq, MSG_DONTWAIT,
2010 				 tss, &zc->msg_flags);
2011 	if (err < 0)
2012 		return err;
2013 
2014 	zc->copybuf_len = err;
2015 	if (likely(zc->copybuf_len)) {
2016 		struct sk_buff *skb;
2017 		u32 offset;
2018 
2019 		skb = tcp_recv_skb(sk, tcp_sk(sk)->copied_seq, &offset);
2020 		if (skb)
2021 			tcp_zerocopy_set_hint_for_skb(sk, zc, skb, offset);
2022 	}
2023 	return 0;
2024 }
2025 
tcp_copy_straggler_data(struct tcp_zerocopy_receive * zc,struct sk_buff * skb,u32 copylen,u32 * offset,u32 * seq)2026 static int tcp_copy_straggler_data(struct tcp_zerocopy_receive *zc,
2027 				   struct sk_buff *skb, u32 copylen,
2028 				   u32 *offset, u32 *seq)
2029 {
2030 	unsigned long copy_address = (unsigned long)zc->copybuf_address;
2031 	struct msghdr msg = {};
2032 	struct iovec iov;
2033 	int err;
2034 
2035 	if (copy_address != zc->copybuf_address)
2036 		return -EINVAL;
2037 
2038 	err = import_single_range(READ, (void __user *)copy_address,
2039 				  copylen, &iov, &msg.msg_iter);
2040 	if (err)
2041 		return err;
2042 	err = skb_copy_datagram_msg(skb, *offset, &msg, copylen);
2043 	if (err)
2044 		return err;
2045 	zc->recv_skip_hint -= copylen;
2046 	*offset += copylen;
2047 	*seq += copylen;
2048 	return (__s32)copylen;
2049 }
2050 
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)2051 static int tcp_zc_handle_leftover(struct tcp_zerocopy_receive *zc,
2052 				  struct sock *sk,
2053 				  struct sk_buff *skb,
2054 				  u32 *seq,
2055 				  s32 copybuf_len,
2056 				  struct scm_timestamping_internal *tss)
2057 {
2058 	u32 offset, copylen = min_t(u32, copybuf_len, zc->recv_skip_hint);
2059 
2060 	if (!copylen)
2061 		return 0;
2062 	/* skb is null if inq < PAGE_SIZE. */
2063 	if (skb) {
2064 		offset = *seq - TCP_SKB_CB(skb)->seq;
2065 	} else {
2066 		skb = tcp_recv_skb(sk, *seq, &offset);
2067 		if (TCP_SKB_CB(skb)->has_rxtstamp) {
2068 			tcp_update_recv_tstamps(skb, tss);
2069 			zc->msg_flags |= TCP_CMSG_TS;
2070 		}
2071 	}
2072 
2073 	zc->copybuf_len = tcp_copy_straggler_data(zc, skb, copylen, &offset,
2074 						  seq);
2075 	return zc->copybuf_len < 0 ? 0 : copylen;
2076 }
2077 
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)2078 static int tcp_zerocopy_vm_insert_batch_error(struct vm_area_struct *vma,
2079 					      struct page **pending_pages,
2080 					      unsigned long pages_remaining,
2081 					      unsigned long *address,
2082 					      u32 *length,
2083 					      u32 *seq,
2084 					      struct tcp_zerocopy_receive *zc,
2085 					      u32 total_bytes_to_map,
2086 					      int err)
2087 {
2088 	/* At least one page did not map. Try zapping if we skipped earlier. */
2089 	if (err == -EBUSY &&
2090 	    zc->flags & TCP_RECEIVE_ZEROCOPY_FLAG_TLB_CLEAN_HINT) {
2091 		u32 maybe_zap_len;
2092 
2093 		maybe_zap_len = total_bytes_to_map -  /* All bytes to map */
2094 				*length + /* Mapped or pending */
2095 				(pages_remaining * PAGE_SIZE); /* Failed map. */
2096 		zap_page_range(vma, *address, maybe_zap_len);
2097 		err = 0;
2098 	}
2099 
2100 	if (!err) {
2101 		unsigned long leftover_pages = pages_remaining;
2102 		int bytes_mapped;
2103 
2104 		/* We called zap_page_range, try to reinsert. */
2105 		err = vm_insert_pages(vma, *address,
2106 				      pending_pages,
2107 				      &pages_remaining);
2108 		bytes_mapped = PAGE_SIZE * (leftover_pages - pages_remaining);
2109 		*seq += bytes_mapped;
2110 		*address += bytes_mapped;
2111 	}
2112 	if (err) {
2113 		/* Either we were unable to zap, OR we zapped, retried an
2114 		 * insert, and still had an issue. Either ways, pages_remaining
2115 		 * is the number of pages we were unable to map, and we unroll
2116 		 * some state we speculatively touched before.
2117 		 */
2118 		const int bytes_not_mapped = PAGE_SIZE * pages_remaining;
2119 
2120 		*length -= bytes_not_mapped;
2121 		zc->recv_skip_hint += bytes_not_mapped;
2122 	}
2123 	return err;
2124 }
2125 
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)2126 static int tcp_zerocopy_vm_insert_batch(struct vm_area_struct *vma,
2127 					struct page **pages,
2128 					unsigned int pages_to_map,
2129 					unsigned long *address,
2130 					u32 *length,
2131 					u32 *seq,
2132 					struct tcp_zerocopy_receive *zc,
2133 					u32 total_bytes_to_map)
2134 {
2135 	unsigned long pages_remaining = pages_to_map;
2136 	unsigned int pages_mapped;
2137 	unsigned int bytes_mapped;
2138 	int err;
2139 
2140 	err = vm_insert_pages(vma, *address, pages, &pages_remaining);
2141 	pages_mapped = pages_to_map - (unsigned int)pages_remaining;
2142 	bytes_mapped = PAGE_SIZE * pages_mapped;
2143 	/* Even if vm_insert_pages fails, it may have partially succeeded in
2144 	 * mapping (some but not all of the pages).
2145 	 */
2146 	*seq += bytes_mapped;
2147 	*address += bytes_mapped;
2148 
2149 	if (likely(!err))
2150 		return 0;
2151 
2152 	/* Error: maybe zap and retry + rollback state for failed inserts. */
2153 	return tcp_zerocopy_vm_insert_batch_error(vma, pages + pages_mapped,
2154 		pages_remaining, address, length, seq, zc, total_bytes_to_map,
2155 		err);
2156 }
2157 
2158 #define TCP_VALID_ZC_MSG_FLAGS   (TCP_CMSG_TS)
tcp_zc_finalize_rx_tstamp(struct sock * sk,struct tcp_zerocopy_receive * zc,struct scm_timestamping_internal * tss)2159 static void tcp_zc_finalize_rx_tstamp(struct sock *sk,
2160 				      struct tcp_zerocopy_receive *zc,
2161 				      struct scm_timestamping_internal *tss)
2162 {
2163 	unsigned long msg_control_addr;
2164 	struct msghdr cmsg_dummy;
2165 
2166 	msg_control_addr = (unsigned long)zc->msg_control;
2167 	cmsg_dummy.msg_control = (void *)msg_control_addr;
2168 	cmsg_dummy.msg_controllen =
2169 		(__kernel_size_t)zc->msg_controllen;
2170 	cmsg_dummy.msg_flags = in_compat_syscall()
2171 		? MSG_CMSG_COMPAT : 0;
2172 	cmsg_dummy.msg_control_is_user = true;
2173 	zc->msg_flags = 0;
2174 	if (zc->msg_control == msg_control_addr &&
2175 	    zc->msg_controllen == cmsg_dummy.msg_controllen) {
2176 		tcp_recv_timestamp(&cmsg_dummy, sk, tss);
2177 		zc->msg_control = (__u64)
2178 			((uintptr_t)cmsg_dummy.msg_control);
2179 		zc->msg_controllen =
2180 			(__u64)cmsg_dummy.msg_controllen;
2181 		zc->msg_flags = (__u32)cmsg_dummy.msg_flags;
2182 	}
2183 }
2184 
2185 #define TCP_ZEROCOPY_PAGE_BATCH_SIZE 32
tcp_zerocopy_receive(struct sock * sk,struct tcp_zerocopy_receive * zc,struct scm_timestamping_internal * tss)2186 static int tcp_zerocopy_receive(struct sock *sk,
2187 				struct tcp_zerocopy_receive *zc,
2188 				struct scm_timestamping_internal *tss)
2189 {
2190 	u32 length = 0, offset, vma_len, avail_len, copylen = 0;
2191 	unsigned long address = (unsigned long)zc->address;
2192 	struct page *pages[TCP_ZEROCOPY_PAGE_BATCH_SIZE];
2193 	s32 copybuf_len = zc->copybuf_len;
2194 	struct tcp_sock *tp = tcp_sk(sk);
2195 	const skb_frag_t *frags = NULL;
2196 	unsigned int pages_to_map = 0;
2197 	struct vm_area_struct *vma;
2198 	struct sk_buff *skb = NULL;
2199 	u32 seq = tp->copied_seq;
2200 	u32 total_bytes_to_map;
2201 	int inq = tcp_inq(sk);
2202 	int ret;
2203 
2204 	zc->copybuf_len = 0;
2205 	zc->msg_flags = 0;
2206 
2207 	if (address & (PAGE_SIZE - 1) || address != zc->address)
2208 		return -EINVAL;
2209 
2210 	if (sk->sk_state == TCP_LISTEN)
2211 		return -ENOTCONN;
2212 
2213 	sock_rps_record_flow(sk);
2214 
2215 	if (inq && inq <= copybuf_len)
2216 		return receive_fallback_to_copy(sk, zc, inq, tss);
2217 
2218 	if (inq < PAGE_SIZE) {
2219 		zc->length = 0;
2220 		zc->recv_skip_hint = inq;
2221 		if (!inq && sock_flag(sk, SOCK_DONE))
2222 			return -EIO;
2223 		return 0;
2224 	}
2225 
2226 	mmap_read_lock(current->mm);
2227 
2228 	vma = vma_lookup(current->mm, address);
2229 	if (!vma || vma->vm_ops != &tcp_vm_ops) {
2230 		mmap_read_unlock(current->mm);
2231 		return -EINVAL;
2232 	}
2233 	vma_len = min_t(unsigned long, zc->length, vma->vm_end - address);
2234 	avail_len = min_t(u32, vma_len, inq);
2235 	total_bytes_to_map = avail_len & ~(PAGE_SIZE - 1);
2236 	if (total_bytes_to_map) {
2237 		if (!(zc->flags & TCP_RECEIVE_ZEROCOPY_FLAG_TLB_CLEAN_HINT))
2238 			zap_page_range(vma, address, total_bytes_to_map);
2239 		zc->length = total_bytes_to_map;
2240 		zc->recv_skip_hint = 0;
2241 	} else {
2242 		zc->length = avail_len;
2243 		zc->recv_skip_hint = avail_len;
2244 	}
2245 	ret = 0;
2246 	while (length + PAGE_SIZE <= zc->length) {
2247 		int mappable_offset;
2248 		struct page *page;
2249 
2250 		if (zc->recv_skip_hint < PAGE_SIZE) {
2251 			u32 offset_frag;
2252 
2253 			if (skb) {
2254 				if (zc->recv_skip_hint > 0)
2255 					break;
2256 				skb = skb->next;
2257 				offset = seq - TCP_SKB_CB(skb)->seq;
2258 			} else {
2259 				skb = tcp_recv_skb(sk, seq, &offset);
2260 			}
2261 
2262 			if (TCP_SKB_CB(skb)->has_rxtstamp) {
2263 				tcp_update_recv_tstamps(skb, tss);
2264 				zc->msg_flags |= TCP_CMSG_TS;
2265 			}
2266 			zc->recv_skip_hint = skb->len - offset;
2267 			frags = skb_advance_to_frag(skb, offset, &offset_frag);
2268 			if (!frags || offset_frag)
2269 				break;
2270 		}
2271 
2272 		mappable_offset = find_next_mappable_frag(frags,
2273 							  zc->recv_skip_hint);
2274 		if (mappable_offset) {
2275 			zc->recv_skip_hint = mappable_offset;
2276 			break;
2277 		}
2278 		page = skb_frag_page(frags);
2279 		prefetchw(page);
2280 		pages[pages_to_map++] = page;
2281 		length += PAGE_SIZE;
2282 		zc->recv_skip_hint -= PAGE_SIZE;
2283 		frags++;
2284 		if (pages_to_map == TCP_ZEROCOPY_PAGE_BATCH_SIZE ||
2285 		    zc->recv_skip_hint < PAGE_SIZE) {
2286 			/* Either full batch, or we're about to go to next skb
2287 			 * (and we cannot unroll failed ops across skbs).
2288 			 */
2289 			ret = tcp_zerocopy_vm_insert_batch(vma, pages,
2290 							   pages_to_map,
2291 							   &address, &length,
2292 							   &seq, zc,
2293 							   total_bytes_to_map);
2294 			if (ret)
2295 				goto out;
2296 			pages_to_map = 0;
2297 		}
2298 	}
2299 	if (pages_to_map) {
2300 		ret = tcp_zerocopy_vm_insert_batch(vma, pages, pages_to_map,
2301 						   &address, &length, &seq,
2302 						   zc, total_bytes_to_map);
2303 	}
2304 out:
2305 	mmap_read_unlock(current->mm);
2306 	/* Try to copy straggler data. */
2307 	if (!ret)
2308 		copylen = tcp_zc_handle_leftover(zc, sk, skb, &seq, copybuf_len, tss);
2309 
2310 	if (length + copylen) {
2311 		WRITE_ONCE(tp->copied_seq, seq);
2312 		tcp_rcv_space_adjust(sk);
2313 
2314 		/* Clean up data we have read: This will do ACK frames. */
2315 		tcp_recv_skb(sk, seq, &offset);
2316 		tcp_cleanup_rbuf(sk, length + copylen);
2317 		ret = 0;
2318 		if (length == zc->length)
2319 			zc->recv_skip_hint = 0;
2320 	} else {
2321 		if (!zc->recv_skip_hint && sock_flag(sk, SOCK_DONE))
2322 			ret = -EIO;
2323 	}
2324 	zc->length = length;
2325 	return ret;
2326 }
2327 #endif
2328 
2329 /* Similar to __sock_recv_timestamp, but does not require an skb */
tcp_recv_timestamp(struct msghdr * msg,const struct sock * sk,struct scm_timestamping_internal * tss)2330 void tcp_recv_timestamp(struct msghdr *msg, const struct sock *sk,
2331 			struct scm_timestamping_internal *tss)
2332 {
2333 	int new_tstamp = sock_flag(sk, SOCK_TSTAMP_NEW);
2334 	bool has_timestamping = false;
2335 
2336 	if (tss->ts[0].tv_sec || tss->ts[0].tv_nsec) {
2337 		if (sock_flag(sk, SOCK_RCVTSTAMP)) {
2338 			if (sock_flag(sk, SOCK_RCVTSTAMPNS)) {
2339 				if (new_tstamp) {
2340 					struct __kernel_timespec kts = {
2341 						.tv_sec = tss->ts[0].tv_sec,
2342 						.tv_nsec = tss->ts[0].tv_nsec,
2343 					};
2344 					put_cmsg(msg, SOL_SOCKET, SO_TIMESTAMPNS_NEW,
2345 						 sizeof(kts), &kts);
2346 				} else {
2347 					struct __kernel_old_timespec ts_old = {
2348 						.tv_sec = tss->ts[0].tv_sec,
2349 						.tv_nsec = tss->ts[0].tv_nsec,
2350 					};
2351 					put_cmsg(msg, SOL_SOCKET, SO_TIMESTAMPNS_OLD,
2352 						 sizeof(ts_old), &ts_old);
2353 				}
2354 			} else {
2355 				if (new_tstamp) {
2356 					struct __kernel_sock_timeval stv = {
2357 						.tv_sec = tss->ts[0].tv_sec,
2358 						.tv_usec = tss->ts[0].tv_nsec / 1000,
2359 					};
2360 					put_cmsg(msg, SOL_SOCKET, SO_TIMESTAMP_NEW,
2361 						 sizeof(stv), &stv);
2362 				} else {
2363 					struct __kernel_old_timeval tv = {
2364 						.tv_sec = tss->ts[0].tv_sec,
2365 						.tv_usec = tss->ts[0].tv_nsec / 1000,
2366 					};
2367 					put_cmsg(msg, SOL_SOCKET, SO_TIMESTAMP_OLD,
2368 						 sizeof(tv), &tv);
2369 				}
2370 			}
2371 		}
2372 
2373 		if (sk->sk_tsflags & SOF_TIMESTAMPING_SOFTWARE)
2374 			has_timestamping = true;
2375 		else
2376 			tss->ts[0] = (struct timespec64) {0};
2377 	}
2378 
2379 	if (tss->ts[2].tv_sec || tss->ts[2].tv_nsec) {
2380 		if (sk->sk_tsflags & SOF_TIMESTAMPING_RAW_HARDWARE)
2381 			has_timestamping = true;
2382 		else
2383 			tss->ts[2] = (struct timespec64) {0};
2384 	}
2385 
2386 	if (has_timestamping) {
2387 		tss->ts[1] = (struct timespec64) {0};
2388 		if (sock_flag(sk, SOCK_TSTAMP_NEW))
2389 			put_cmsg_scm_timestamping64(msg, tss);
2390 		else
2391 			put_cmsg_scm_timestamping(msg, tss);
2392 	}
2393 }
2394 
tcp_inq_hint(struct sock * sk)2395 static int tcp_inq_hint(struct sock *sk)
2396 {
2397 	const struct tcp_sock *tp = tcp_sk(sk);
2398 	u32 copied_seq = READ_ONCE(tp->copied_seq);
2399 	u32 rcv_nxt = READ_ONCE(tp->rcv_nxt);
2400 	int inq;
2401 
2402 	inq = rcv_nxt - copied_seq;
2403 	if (unlikely(inq < 0 || copied_seq != READ_ONCE(tp->copied_seq))) {
2404 		lock_sock(sk);
2405 		inq = tp->rcv_nxt - tp->copied_seq;
2406 		release_sock(sk);
2407 	}
2408 	/* After receiving a FIN, tell the user-space to continue reading
2409 	 * by returning a non-zero inq.
2410 	 */
2411 	if (inq == 0 && sock_flag(sk, SOCK_DONE))
2412 		inq = 1;
2413 	return inq;
2414 }
2415 
2416 /*
2417  *	This routine copies from a sock struct into the user buffer.
2418  *
2419  *	Technical note: in 2.3 we work on _locked_ socket, so that
2420  *	tricks with *seq access order and skb->users are not required.
2421  *	Probably, code can be easily improved even more.
2422  */
2423 
tcp_recvmsg_locked(struct sock * sk,struct msghdr * msg,size_t len,int flags,struct scm_timestamping_internal * tss,int * cmsg_flags)2424 static int tcp_recvmsg_locked(struct sock *sk, struct msghdr *msg, size_t len,
2425 			      int flags, struct scm_timestamping_internal *tss,
2426 			      int *cmsg_flags)
2427 {
2428 	struct tcp_sock *tp = tcp_sk(sk);
2429 	int copied = 0;
2430 	u32 peek_seq;
2431 	u32 *seq;
2432 	unsigned long used;
2433 	int err;
2434 	int target;		/* Read at least this many bytes */
2435 	long timeo;
2436 	struct sk_buff *skb, *last;
2437 	u32 urg_hole = 0;
2438 
2439 	err = -ENOTCONN;
2440 	if (sk->sk_state == TCP_LISTEN)
2441 		goto out;
2442 
2443 	if (tp->recvmsg_inq) {
2444 		*cmsg_flags = TCP_CMSG_INQ;
2445 		msg->msg_get_inq = 1;
2446 	}
2447 	timeo = sock_rcvtimeo(sk, flags & MSG_DONTWAIT);
2448 
2449 	/* Urgent data needs to be handled specially. */
2450 	if (flags & MSG_OOB)
2451 		goto recv_urg;
2452 
2453 	if (unlikely(tp->repair)) {
2454 		err = -EPERM;
2455 		if (!(flags & MSG_PEEK))
2456 			goto out;
2457 
2458 		if (tp->repair_queue == TCP_SEND_QUEUE)
2459 			goto recv_sndq;
2460 
2461 		err = -EINVAL;
2462 		if (tp->repair_queue == TCP_NO_QUEUE)
2463 			goto out;
2464 
2465 		/* 'common' recv queue MSG_PEEK-ing */
2466 	}
2467 
2468 	seq = &tp->copied_seq;
2469 	if (flags & MSG_PEEK) {
2470 		peek_seq = tp->copied_seq;
2471 		seq = &peek_seq;
2472 	}
2473 
2474 	target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
2475 
2476 	do {
2477 		u32 offset;
2478 
2479 		/* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
2480 		if (unlikely(tp->urg_data) && tp->urg_seq == *seq) {
2481 			if (copied)
2482 				break;
2483 			if (signal_pending(current)) {
2484 				copied = timeo ? sock_intr_errno(timeo) : -EAGAIN;
2485 				break;
2486 			}
2487 		}
2488 
2489 		/* Next get a buffer. */
2490 
2491 		last = skb_peek_tail(&sk->sk_receive_queue);
2492 		skb_queue_walk(&sk->sk_receive_queue, skb) {
2493 			last = skb;
2494 			/* Now that we have two receive queues this
2495 			 * shouldn't happen.
2496 			 */
2497 			if (WARN(before(*seq, TCP_SKB_CB(skb)->seq),
2498 				 "TCP recvmsg seq # bug: copied %X, seq %X, rcvnxt %X, fl %X\n",
2499 				 *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt,
2500 				 flags))
2501 				break;
2502 
2503 			offset = *seq - TCP_SKB_CB(skb)->seq;
2504 			if (unlikely(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
2505 				pr_err_once("%s: found a SYN, please report !\n", __func__);
2506 				offset--;
2507 			}
2508 			if (offset < skb->len)
2509 				goto found_ok_skb;
2510 			if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
2511 				goto found_fin_ok;
2512 			WARN(!(flags & MSG_PEEK),
2513 			     "TCP recvmsg seq # bug 2: copied %X, seq %X, rcvnxt %X, fl %X\n",
2514 			     *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt, flags);
2515 		}
2516 
2517 		/* Well, if we have backlog, try to process it now yet. */
2518 
2519 		if (copied >= target && !READ_ONCE(sk->sk_backlog.tail))
2520 			break;
2521 
2522 		if (copied) {
2523 			if (!timeo ||
2524 			    sk->sk_err ||
2525 			    sk->sk_state == TCP_CLOSE ||
2526 			    (sk->sk_shutdown & RCV_SHUTDOWN) ||
2527 			    signal_pending(current))
2528 				break;
2529 		} else {
2530 			if (sock_flag(sk, SOCK_DONE))
2531 				break;
2532 
2533 			if (sk->sk_err) {
2534 				copied = sock_error(sk);
2535 				break;
2536 			}
2537 
2538 			if (sk->sk_shutdown & RCV_SHUTDOWN)
2539 				break;
2540 
2541 			if (sk->sk_state == TCP_CLOSE) {
2542 				/* This occurs when user tries to read
2543 				 * from never connected socket.
2544 				 */
2545 				copied = -ENOTCONN;
2546 				break;
2547 			}
2548 
2549 			if (!timeo) {
2550 				copied = -EAGAIN;
2551 				break;
2552 			}
2553 
2554 			if (signal_pending(current)) {
2555 				copied = sock_intr_errno(timeo);
2556 				break;
2557 			}
2558 		}
2559 
2560 		if (copied >= target) {
2561 			/* Do not sleep, just process backlog. */
2562 			__sk_flush_backlog(sk);
2563 		} else {
2564 			tcp_cleanup_rbuf(sk, copied);
2565 			sk_wait_data(sk, &timeo, last);
2566 		}
2567 
2568 		if ((flags & MSG_PEEK) &&
2569 		    (peek_seq - copied - urg_hole != tp->copied_seq)) {
2570 			net_dbg_ratelimited("TCP(%s:%d): Application bug, race in MSG_PEEK\n",
2571 					    current->comm,
2572 					    task_pid_nr(current));
2573 			peek_seq = tp->copied_seq;
2574 		}
2575 		continue;
2576 
2577 found_ok_skb:
2578 		/* Ok so how much can we use? */
2579 		used = skb->len - offset;
2580 		if (len < used)
2581 			used = len;
2582 
2583 		/* Do we have urgent data here? */
2584 		if (unlikely(tp->urg_data)) {
2585 			u32 urg_offset = tp->urg_seq - *seq;
2586 			if (urg_offset < used) {
2587 				if (!urg_offset) {
2588 					if (!sock_flag(sk, SOCK_URGINLINE)) {
2589 						WRITE_ONCE(*seq, *seq + 1);
2590 						urg_hole++;
2591 						offset++;
2592 						used--;
2593 						if (!used)
2594 							goto skip_copy;
2595 					}
2596 				} else
2597 					used = urg_offset;
2598 			}
2599 		}
2600 
2601 		if (!(flags & MSG_TRUNC)) {
2602 			err = skb_copy_datagram_msg(skb, offset, msg, used);
2603 			if (err) {
2604 				/* Exception. Bailout! */
2605 				if (!copied)
2606 					copied = -EFAULT;
2607 				break;
2608 			}
2609 		}
2610 
2611 		WRITE_ONCE(*seq, *seq + used);
2612 		copied += used;
2613 		len -= used;
2614 
2615 		tcp_rcv_space_adjust(sk);
2616 
2617 skip_copy:
2618 		if (unlikely(tp->urg_data) && after(tp->copied_seq, tp->urg_seq)) {
2619 			WRITE_ONCE(tp->urg_data, 0);
2620 			tcp_fast_path_check(sk);
2621 		}
2622 
2623 		if (TCP_SKB_CB(skb)->has_rxtstamp) {
2624 			tcp_update_recv_tstamps(skb, tss);
2625 			*cmsg_flags |= TCP_CMSG_TS;
2626 		}
2627 
2628 		if (used + offset < skb->len)
2629 			continue;
2630 
2631 		if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
2632 			goto found_fin_ok;
2633 		if (!(flags & MSG_PEEK))
2634 			tcp_eat_recv_skb(sk, skb);
2635 		continue;
2636 
2637 found_fin_ok:
2638 		/* Process the FIN. */
2639 		WRITE_ONCE(*seq, *seq + 1);
2640 		if (!(flags & MSG_PEEK))
2641 			tcp_eat_recv_skb(sk, skb);
2642 		break;
2643 	} while (len > 0);
2644 
2645 	/* According to UNIX98, msg_name/msg_namelen are ignored
2646 	 * on connected socket. I was just happy when found this 8) --ANK
2647 	 */
2648 
2649 	/* Clean up data we have read: This will do ACK frames. */
2650 	tcp_cleanup_rbuf(sk, copied);
2651 	return copied;
2652 
2653 out:
2654 	return err;
2655 
2656 recv_urg:
2657 	err = tcp_recv_urg(sk, msg, len, flags);
2658 	goto out;
2659 
2660 recv_sndq:
2661 	err = tcp_peek_sndq(sk, msg, len);
2662 	goto out;
2663 }
2664 
tcp_recvmsg(struct sock * sk,struct msghdr * msg,size_t len,int flags,int * addr_len)2665 int tcp_recvmsg(struct sock *sk, struct msghdr *msg, size_t len, int flags,
2666 		int *addr_len)
2667 {
2668 	int cmsg_flags = 0, ret;
2669 	struct scm_timestamping_internal tss;
2670 
2671 	if (unlikely(flags & MSG_ERRQUEUE))
2672 		return inet_recv_error(sk, msg, len, addr_len);
2673 
2674 	if (sk_can_busy_loop(sk) &&
2675 	    skb_queue_empty_lockless(&sk->sk_receive_queue) &&
2676 	    sk->sk_state == TCP_ESTABLISHED)
2677 		sk_busy_loop(sk, flags & MSG_DONTWAIT);
2678 
2679 	lock_sock(sk);
2680 	ret = tcp_recvmsg_locked(sk, msg, len, flags, &tss, &cmsg_flags);
2681 	release_sock(sk);
2682 
2683 	if ((cmsg_flags || msg->msg_get_inq) && ret >= 0) {
2684 		if (cmsg_flags & TCP_CMSG_TS)
2685 			tcp_recv_timestamp(msg, sk, &tss);
2686 		if (msg->msg_get_inq) {
2687 			msg->msg_inq = tcp_inq_hint(sk);
2688 			if (cmsg_flags & TCP_CMSG_INQ)
2689 				put_cmsg(msg, SOL_TCP, TCP_CM_INQ,
2690 					 sizeof(msg->msg_inq), &msg->msg_inq);
2691 		}
2692 	}
2693 	return ret;
2694 }
2695 EXPORT_SYMBOL(tcp_recvmsg);
2696 
tcp_set_state(struct sock * sk,int state)2697 void tcp_set_state(struct sock *sk, int state)
2698 {
2699 	int oldstate = sk->sk_state;
2700 
2701 	/* We defined a new enum for TCP states that are exported in BPF
2702 	 * so as not force the internal TCP states to be frozen. The
2703 	 * following checks will detect if an internal state value ever
2704 	 * differs from the BPF value. If this ever happens, then we will
2705 	 * need to remap the internal value to the BPF value before calling
2706 	 * tcp_call_bpf_2arg.
2707 	 */
2708 	BUILD_BUG_ON((int)BPF_TCP_ESTABLISHED != (int)TCP_ESTABLISHED);
2709 	BUILD_BUG_ON((int)BPF_TCP_SYN_SENT != (int)TCP_SYN_SENT);
2710 	BUILD_BUG_ON((int)BPF_TCP_SYN_RECV != (int)TCP_SYN_RECV);
2711 	BUILD_BUG_ON((int)BPF_TCP_FIN_WAIT1 != (int)TCP_FIN_WAIT1);
2712 	BUILD_BUG_ON((int)BPF_TCP_FIN_WAIT2 != (int)TCP_FIN_WAIT2);
2713 	BUILD_BUG_ON((int)BPF_TCP_TIME_WAIT != (int)TCP_TIME_WAIT);
2714 	BUILD_BUG_ON((int)BPF_TCP_CLOSE != (int)TCP_CLOSE);
2715 	BUILD_BUG_ON((int)BPF_TCP_CLOSE_WAIT != (int)TCP_CLOSE_WAIT);
2716 	BUILD_BUG_ON((int)BPF_TCP_LAST_ACK != (int)TCP_LAST_ACK);
2717 	BUILD_BUG_ON((int)BPF_TCP_LISTEN != (int)TCP_LISTEN);
2718 	BUILD_BUG_ON((int)BPF_TCP_CLOSING != (int)TCP_CLOSING);
2719 	BUILD_BUG_ON((int)BPF_TCP_NEW_SYN_RECV != (int)TCP_NEW_SYN_RECV);
2720 	BUILD_BUG_ON((int)BPF_TCP_MAX_STATES != (int)TCP_MAX_STATES);
2721 
2722 	/* bpf uapi header bpf.h defines an anonymous enum with values
2723 	 * BPF_TCP_* used by bpf programs. Currently gcc built vmlinux
2724 	 * is able to emit this enum in DWARF due to the above BUILD_BUG_ON.
2725 	 * But clang built vmlinux does not have this enum in DWARF
2726 	 * since clang removes the above code before generating IR/debuginfo.
2727 	 * Let us explicitly emit the type debuginfo to ensure the
2728 	 * above-mentioned anonymous enum in the vmlinux DWARF and hence BTF
2729 	 * regardless of which compiler is used.
2730 	 */
2731 	BTF_TYPE_EMIT_ENUM(BPF_TCP_ESTABLISHED);
2732 
2733 	if (BPF_SOCK_OPS_TEST_FLAG(tcp_sk(sk), BPF_SOCK_OPS_STATE_CB_FLAG))
2734 		tcp_call_bpf_2arg(sk, BPF_SOCK_OPS_STATE_CB, oldstate, state);
2735 
2736 	switch (state) {
2737 	case TCP_ESTABLISHED:
2738 		if (oldstate != TCP_ESTABLISHED)
2739 			TCP_INC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
2740 		break;
2741 
2742 	case TCP_CLOSE:
2743 		if (oldstate == TCP_CLOSE_WAIT || oldstate == TCP_ESTABLISHED)
2744 			TCP_INC_STATS(sock_net(sk), TCP_MIB_ESTABRESETS);
2745 
2746 		sk->sk_prot->unhash(sk);
2747 		if (inet_csk(sk)->icsk_bind_hash &&
2748 		    !(sk->sk_userlocks & SOCK_BINDPORT_LOCK))
2749 			inet_put_port(sk);
2750 		fallthrough;
2751 	default:
2752 		if (oldstate == TCP_ESTABLISHED)
2753 			TCP_DEC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
2754 	}
2755 
2756 	/* Change state AFTER socket is unhashed to avoid closed
2757 	 * socket sitting in hash tables.
2758 	 */
2759 	inet_sk_state_store(sk, state);
2760 }
2761 EXPORT_SYMBOL_GPL(tcp_set_state);
2762 
2763 /*
2764  *	State processing on a close. This implements the state shift for
2765  *	sending our FIN frame. Note that we only send a FIN for some
2766  *	states. A shutdown() may have already sent the FIN, or we may be
2767  *	closed.
2768  */
2769 
2770 static const unsigned char new_state[16] = {
2771   /* current state:        new state:      action:	*/
2772   [0 /* (Invalid) */]	= TCP_CLOSE,
2773   [TCP_ESTABLISHED]	= TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2774   [TCP_SYN_SENT]	= TCP_CLOSE,
2775   [TCP_SYN_RECV]	= TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2776   [TCP_FIN_WAIT1]	= TCP_FIN_WAIT1,
2777   [TCP_FIN_WAIT2]	= TCP_FIN_WAIT2,
2778   [TCP_TIME_WAIT]	= TCP_CLOSE,
2779   [TCP_CLOSE]		= TCP_CLOSE,
2780   [TCP_CLOSE_WAIT]	= TCP_LAST_ACK  | TCP_ACTION_FIN,
2781   [TCP_LAST_ACK]	= TCP_LAST_ACK,
2782   [TCP_LISTEN]		= TCP_CLOSE,
2783   [TCP_CLOSING]		= TCP_CLOSING,
2784   [TCP_NEW_SYN_RECV]	= TCP_CLOSE,	/* should not happen ! */
2785 };
2786 
tcp_close_state(struct sock * sk)2787 static int tcp_close_state(struct sock *sk)
2788 {
2789 	int next = (int)new_state[sk->sk_state];
2790 	int ns = next & TCP_STATE_MASK;
2791 
2792 	tcp_set_state(sk, ns);
2793 
2794 	return next & TCP_ACTION_FIN;
2795 }
2796 
2797 /*
2798  *	Shutdown the sending side of a connection. Much like close except
2799  *	that we don't receive shut down or sock_set_flag(sk, SOCK_DEAD).
2800  */
2801 
tcp_shutdown(struct sock * sk,int how)2802 void tcp_shutdown(struct sock *sk, int how)
2803 {
2804 	/*	We need to grab some memory, and put together a FIN,
2805 	 *	and then put it into the queue to be sent.
2806 	 *		Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
2807 	 */
2808 	if (!(how & SEND_SHUTDOWN))
2809 		return;
2810 
2811 	/* If we've already sent a FIN, or it's a closed state, skip this. */
2812 	if ((1 << sk->sk_state) &
2813 	    (TCPF_ESTABLISHED | TCPF_SYN_SENT |
2814 	     TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) {
2815 		/* Clear out any half completed packets.  FIN if needed. */
2816 		if (tcp_close_state(sk))
2817 			tcp_send_fin(sk);
2818 	}
2819 }
2820 EXPORT_SYMBOL(tcp_shutdown);
2821 
tcp_orphan_count_sum(void)2822 int tcp_orphan_count_sum(void)
2823 {
2824 	int i, total = 0;
2825 
2826 	for_each_possible_cpu(i)
2827 		total += per_cpu(tcp_orphan_count, i);
2828 
2829 	return max(total, 0);
2830 }
2831 
2832 static int tcp_orphan_cache;
2833 static struct timer_list tcp_orphan_timer;
2834 #define TCP_ORPHAN_TIMER_PERIOD msecs_to_jiffies(100)
2835 
tcp_orphan_update(struct timer_list * unused)2836 static void tcp_orphan_update(struct timer_list *unused)
2837 {
2838 	WRITE_ONCE(tcp_orphan_cache, tcp_orphan_count_sum());
2839 	mod_timer(&tcp_orphan_timer, jiffies + TCP_ORPHAN_TIMER_PERIOD);
2840 }
2841 
tcp_too_many_orphans(int shift)2842 static bool tcp_too_many_orphans(int shift)
2843 {
2844 	return READ_ONCE(tcp_orphan_cache) << shift >
2845 		READ_ONCE(sysctl_tcp_max_orphans);
2846 }
2847 
tcp_check_oom(struct sock * sk,int shift)2848 bool tcp_check_oom(struct sock *sk, int shift)
2849 {
2850 	bool too_many_orphans, out_of_socket_memory;
2851 
2852 	too_many_orphans = tcp_too_many_orphans(shift);
2853 	out_of_socket_memory = tcp_out_of_memory(sk);
2854 
2855 	if (too_many_orphans)
2856 		net_info_ratelimited("too many orphaned sockets\n");
2857 	if (out_of_socket_memory)
2858 		net_info_ratelimited("out of memory -- consider tuning tcp_mem\n");
2859 	return too_many_orphans || out_of_socket_memory;
2860 }
2861 
__tcp_close(struct sock * sk,long timeout)2862 void __tcp_close(struct sock *sk, long timeout)
2863 {
2864 	struct sk_buff *skb;
2865 	int data_was_unread = 0;
2866 	int state;
2867 
2868 	sk->sk_shutdown = SHUTDOWN_MASK;
2869 
2870 	if (sk->sk_state == TCP_LISTEN) {
2871 		tcp_set_state(sk, TCP_CLOSE);
2872 
2873 		/* Special case. */
2874 		inet_csk_listen_stop(sk);
2875 
2876 		goto adjudge_to_death;
2877 	}
2878 
2879 	/*  We need to flush the recv. buffs.  We do this only on the
2880 	 *  descriptor close, not protocol-sourced closes, because the
2881 	 *  reader process may not have drained the data yet!
2882 	 */
2883 	while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) {
2884 		u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq;
2885 
2886 		if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
2887 			len--;
2888 		data_was_unread += len;
2889 		__kfree_skb(skb);
2890 	}
2891 
2892 	/* If socket has been already reset (e.g. in tcp_reset()) - kill it. */
2893 	if (sk->sk_state == TCP_CLOSE)
2894 		goto adjudge_to_death;
2895 
2896 	/* As outlined in RFC 2525, section 2.17, we send a RST here because
2897 	 * data was lost. To witness the awful effects of the old behavior of
2898 	 * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk
2899 	 * GET in an FTP client, suspend the process, wait for the client to
2900 	 * advertise a zero window, then kill -9 the FTP client, wheee...
2901 	 * Note: timeout is always zero in such a case.
2902 	 */
2903 	if (unlikely(tcp_sk(sk)->repair)) {
2904 		sk->sk_prot->disconnect(sk, 0);
2905 	} else if (data_was_unread) {
2906 		/* Unread data was tossed, zap the connection. */
2907 		NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONCLOSE);
2908 		tcp_set_state(sk, TCP_CLOSE);
2909 		tcp_send_active_reset(sk, sk->sk_allocation);
2910 	} else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) {
2911 		/* Check zero linger _after_ checking for unread data. */
2912 		sk->sk_prot->disconnect(sk, 0);
2913 		NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
2914 	} else if (tcp_close_state(sk)) {
2915 		/* We FIN if the application ate all the data before
2916 		 * zapping the connection.
2917 		 */
2918 
2919 		/* RED-PEN. Formally speaking, we have broken TCP state
2920 		 * machine. State transitions:
2921 		 *
2922 		 * TCP_ESTABLISHED -> TCP_FIN_WAIT1
2923 		 * TCP_SYN_RECV	-> TCP_FIN_WAIT1 (forget it, it's impossible)
2924 		 * TCP_CLOSE_WAIT -> TCP_LAST_ACK
2925 		 *
2926 		 * are legal only when FIN has been sent (i.e. in window),
2927 		 * rather than queued out of window. Purists blame.
2928 		 *
2929 		 * F.e. "RFC state" is ESTABLISHED,
2930 		 * if Linux state is FIN-WAIT-1, but FIN is still not sent.
2931 		 *
2932 		 * The visible declinations are that sometimes
2933 		 * we enter time-wait state, when it is not required really
2934 		 * (harmless), do not send active resets, when they are
2935 		 * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
2936 		 * they look as CLOSING or LAST_ACK for Linux)
2937 		 * Probably, I missed some more holelets.
2938 		 * 						--ANK
2939 		 * XXX (TFO) - To start off we don't support SYN+ACK+FIN
2940 		 * in a single packet! (May consider it later but will
2941 		 * probably need API support or TCP_CORK SYN-ACK until
2942 		 * data is written and socket is closed.)
2943 		 */
2944 		tcp_send_fin(sk);
2945 	}
2946 
2947 	sk_stream_wait_close(sk, timeout);
2948 
2949 adjudge_to_death:
2950 	state = sk->sk_state;
2951 	sock_hold(sk);
2952 	sock_orphan(sk);
2953 
2954 	local_bh_disable();
2955 	bh_lock_sock(sk);
2956 	/* remove backlog if any, without releasing ownership. */
2957 	__release_sock(sk);
2958 
2959 	this_cpu_inc(tcp_orphan_count);
2960 
2961 	/* Have we already been destroyed by a softirq or backlog? */
2962 	if (state != TCP_CLOSE && sk->sk_state == TCP_CLOSE)
2963 		goto out;
2964 
2965 	/*	This is a (useful) BSD violating of the RFC. There is a
2966 	 *	problem with TCP as specified in that the other end could
2967 	 *	keep a socket open forever with no application left this end.
2968 	 *	We use a 1 minute timeout (about the same as BSD) then kill
2969 	 *	our end. If they send after that then tough - BUT: long enough
2970 	 *	that we won't make the old 4*rto = almost no time - whoops
2971 	 *	reset mistake.
2972 	 *
2973 	 *	Nope, it was not mistake. It is really desired behaviour
2974 	 *	f.e. on http servers, when such sockets are useless, but
2975 	 *	consume significant resources. Let's do it with special
2976 	 *	linger2	option.					--ANK
2977 	 */
2978 
2979 	if (sk->sk_state == TCP_FIN_WAIT2) {
2980 		struct tcp_sock *tp = tcp_sk(sk);
2981 		if (tp->linger2 < 0) {
2982 			tcp_set_state(sk, TCP_CLOSE);
2983 			tcp_send_active_reset(sk, GFP_ATOMIC);
2984 			__NET_INC_STATS(sock_net(sk),
2985 					LINUX_MIB_TCPABORTONLINGER);
2986 		} else {
2987 			const int tmo = tcp_fin_time(sk);
2988 
2989 			if (tmo > TCP_TIMEWAIT_LEN) {
2990 				inet_csk_reset_keepalive_timer(sk,
2991 						tmo - TCP_TIMEWAIT_LEN);
2992 			} else {
2993 				tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
2994 				goto out;
2995 			}
2996 		}
2997 	}
2998 	if (sk->sk_state != TCP_CLOSE) {
2999 		if (tcp_check_oom(sk, 0)) {
3000 			tcp_set_state(sk, TCP_CLOSE);
3001 			tcp_send_active_reset(sk, GFP_ATOMIC);
3002 			__NET_INC_STATS(sock_net(sk),
3003 					LINUX_MIB_TCPABORTONMEMORY);
3004 		} else if (!check_net(sock_net(sk))) {
3005 			/* Not possible to send reset; just close */
3006 			tcp_set_state(sk, TCP_CLOSE);
3007 		}
3008 	}
3009 
3010 	if (sk->sk_state == TCP_CLOSE) {
3011 		struct request_sock *req;
3012 
3013 		req = rcu_dereference_protected(tcp_sk(sk)->fastopen_rsk,
3014 						lockdep_sock_is_held(sk));
3015 		/* We could get here with a non-NULL req if the socket is
3016 		 * aborted (e.g., closed with unread data) before 3WHS
3017 		 * finishes.
3018 		 */
3019 		if (req)
3020 			reqsk_fastopen_remove(sk, req, false);
3021 		inet_csk_destroy_sock(sk);
3022 	}
3023 	/* Otherwise, socket is reprieved until protocol close. */
3024 
3025 out:
3026 	bh_unlock_sock(sk);
3027 	local_bh_enable();
3028 }
3029 
tcp_close(struct sock * sk,long timeout)3030 void tcp_close(struct sock *sk, long timeout)
3031 {
3032 	lock_sock(sk);
3033 	__tcp_close(sk, timeout);
3034 	release_sock(sk);
3035 	sock_put(sk);
3036 }
3037 EXPORT_SYMBOL(tcp_close);
3038 
3039 /* These states need RST on ABORT according to RFC793 */
3040 
tcp_need_reset(int state)3041 static inline bool tcp_need_reset(int state)
3042 {
3043 	return (1 << state) &
3044 	       (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 |
3045 		TCPF_FIN_WAIT2 | TCPF_SYN_RECV);
3046 }
3047 
tcp_rtx_queue_purge(struct sock * sk)3048 static void tcp_rtx_queue_purge(struct sock *sk)
3049 {
3050 	struct rb_node *p = rb_first(&sk->tcp_rtx_queue);
3051 
3052 	tcp_sk(sk)->highest_sack = NULL;
3053 	while (p) {
3054 		struct sk_buff *skb = rb_to_skb(p);
3055 
3056 		p = rb_next(p);
3057 		/* Since we are deleting whole queue, no need to
3058 		 * list_del(&skb->tcp_tsorted_anchor)
3059 		 */
3060 		tcp_rtx_queue_unlink(skb, sk);
3061 		tcp_wmem_free_skb(sk, skb);
3062 	}
3063 }
3064 
tcp_write_queue_purge(struct sock * sk)3065 void tcp_write_queue_purge(struct sock *sk)
3066 {
3067 	struct sk_buff *skb;
3068 
3069 	tcp_chrono_stop(sk, TCP_CHRONO_BUSY);
3070 	while ((skb = __skb_dequeue(&sk->sk_write_queue)) != NULL) {
3071 		tcp_skb_tsorted_anchor_cleanup(skb);
3072 		tcp_wmem_free_skb(sk, skb);
3073 	}
3074 	tcp_rtx_queue_purge(sk);
3075 	INIT_LIST_HEAD(&tcp_sk(sk)->tsorted_sent_queue);
3076 	tcp_clear_all_retrans_hints(tcp_sk(sk));
3077 	tcp_sk(sk)->packets_out = 0;
3078 	inet_csk(sk)->icsk_backoff = 0;
3079 }
3080 
tcp_disconnect(struct sock * sk,int flags)3081 int tcp_disconnect(struct sock *sk, int flags)
3082 {
3083 	struct inet_sock *inet = inet_sk(sk);
3084 	struct inet_connection_sock *icsk = inet_csk(sk);
3085 	struct tcp_sock *tp = tcp_sk(sk);
3086 	int old_state = sk->sk_state;
3087 	u32 seq;
3088 
3089 	if (old_state != TCP_CLOSE)
3090 		tcp_set_state(sk, TCP_CLOSE);
3091 
3092 	/* ABORT function of RFC793 */
3093 	if (old_state == TCP_LISTEN) {
3094 		inet_csk_listen_stop(sk);
3095 	} else if (unlikely(tp->repair)) {
3096 		sk->sk_err = ECONNABORTED;
3097 	} else if (tcp_need_reset(old_state) ||
3098 		   (tp->snd_nxt != tp->write_seq &&
3099 		    (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK))) {
3100 		/* The last check adjusts for discrepancy of Linux wrt. RFC
3101 		 * states
3102 		 */
3103 		tcp_send_active_reset(sk, gfp_any());
3104 		sk->sk_err = ECONNRESET;
3105 	} else if (old_state == TCP_SYN_SENT)
3106 		sk->sk_err = ECONNRESET;
3107 
3108 	tcp_clear_xmit_timers(sk);
3109 	__skb_queue_purge(&sk->sk_receive_queue);
3110 	WRITE_ONCE(tp->copied_seq, tp->rcv_nxt);
3111 	WRITE_ONCE(tp->urg_data, 0);
3112 	tcp_write_queue_purge(sk);
3113 	tcp_fastopen_active_disable_ofo_check(sk);
3114 	skb_rbtree_purge(&tp->out_of_order_queue);
3115 
3116 	inet->inet_dport = 0;
3117 
3118 	inet_bhash2_reset_saddr(sk);
3119 
3120 	sk->sk_shutdown = 0;
3121 	sock_reset_flag(sk, SOCK_DONE);
3122 	tp->srtt_us = 0;
3123 	tp->mdev_us = jiffies_to_usecs(TCP_TIMEOUT_INIT);
3124 	tp->rcv_rtt_last_tsecr = 0;
3125 
3126 	seq = tp->write_seq + tp->max_window + 2;
3127 	if (!seq)
3128 		seq = 1;
3129 	WRITE_ONCE(tp->write_seq, seq);
3130 
3131 	icsk->icsk_backoff = 0;
3132 	icsk->icsk_probes_out = 0;
3133 	icsk->icsk_probes_tstamp = 0;
3134 	icsk->icsk_rto = TCP_TIMEOUT_INIT;
3135 	icsk->icsk_rto_min = TCP_RTO_MIN;
3136 	icsk->icsk_delack_max = TCP_DELACK_MAX;
3137 	tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
3138 	tcp_snd_cwnd_set(tp, TCP_INIT_CWND);
3139 	tp->snd_cwnd_cnt = 0;
3140 	tp->is_cwnd_limited = 0;
3141 	tp->max_packets_out = 0;
3142 	tp->window_clamp = 0;
3143 	tp->delivered = 0;
3144 	tp->delivered_ce = 0;
3145 	if (icsk->icsk_ca_ops->release)
3146 		icsk->icsk_ca_ops->release(sk);
3147 	memset(icsk->icsk_ca_priv, 0, sizeof(icsk->icsk_ca_priv));
3148 	icsk->icsk_ca_initialized = 0;
3149 	tcp_set_ca_state(sk, TCP_CA_Open);
3150 	tp->is_sack_reneg = 0;
3151 	tcp_clear_retrans(tp);
3152 	tp->total_retrans = 0;
3153 	inet_csk_delack_init(sk);
3154 	/* Initialize rcv_mss to TCP_MIN_MSS to avoid division by 0
3155 	 * issue in __tcp_select_window()
3156 	 */
3157 	icsk->icsk_ack.rcv_mss = TCP_MIN_MSS;
3158 	memset(&tp->rx_opt, 0, sizeof(tp->rx_opt));
3159 	__sk_dst_reset(sk);
3160 	dst_release(xchg((__force struct dst_entry **)&sk->sk_rx_dst, NULL));
3161 	tcp_saved_syn_free(tp);
3162 	tp->compressed_ack = 0;
3163 	tp->segs_in = 0;
3164 	tp->segs_out = 0;
3165 	tp->bytes_sent = 0;
3166 	tp->bytes_acked = 0;
3167 	tp->bytes_received = 0;
3168 	tp->bytes_retrans = 0;
3169 	tp->data_segs_in = 0;
3170 	tp->data_segs_out = 0;
3171 	tp->duplicate_sack[0].start_seq = 0;
3172 	tp->duplicate_sack[0].end_seq = 0;
3173 	tp->dsack_dups = 0;
3174 	tp->reord_seen = 0;
3175 	tp->retrans_out = 0;
3176 	tp->sacked_out = 0;
3177 	tp->tlp_high_seq = 0;
3178 	tp->last_oow_ack_time = 0;
3179 	/* There's a bubble in the pipe until at least the first ACK. */
3180 	tp->app_limited = ~0U;
3181 	tp->rate_app_limited = 1;
3182 	tp->rack.mstamp = 0;
3183 	tp->rack.advanced = 0;
3184 	tp->rack.reo_wnd_steps = 1;
3185 	tp->rack.last_delivered = 0;
3186 	tp->rack.reo_wnd_persist = 0;
3187 	tp->rack.dsack_seen = 0;
3188 	tp->syn_data_acked = 0;
3189 	tp->rx_opt.saw_tstamp = 0;
3190 	tp->rx_opt.dsack = 0;
3191 	tp->rx_opt.num_sacks = 0;
3192 	tp->rcv_ooopack = 0;
3193 
3194 
3195 	/* Clean up fastopen related fields */
3196 	tcp_free_fastopen_req(tp);
3197 	inet->defer_connect = 0;
3198 	tp->fastopen_client_fail = 0;
3199 
3200 	WARN_ON(inet->inet_num && !icsk->icsk_bind_hash);
3201 
3202 	if (sk->sk_frag.page) {
3203 		put_page(sk->sk_frag.page);
3204 		sk->sk_frag.page = NULL;
3205 		sk->sk_frag.offset = 0;
3206 	}
3207 	sk_error_report(sk);
3208 	return 0;
3209 }
3210 EXPORT_SYMBOL(tcp_disconnect);
3211 
tcp_can_repair_sock(const struct sock * sk)3212 static inline bool tcp_can_repair_sock(const struct sock *sk)
3213 {
3214 	return sockopt_ns_capable(sock_net(sk)->user_ns, CAP_NET_ADMIN) &&
3215 		(sk->sk_state != TCP_LISTEN);
3216 }
3217 
tcp_repair_set_window(struct tcp_sock * tp,sockptr_t optbuf,int len)3218 static int tcp_repair_set_window(struct tcp_sock *tp, sockptr_t optbuf, int len)
3219 {
3220 	struct tcp_repair_window opt;
3221 
3222 	if (!tp->repair)
3223 		return -EPERM;
3224 
3225 	if (len != sizeof(opt))
3226 		return -EINVAL;
3227 
3228 	if (copy_from_sockptr(&opt, optbuf, sizeof(opt)))
3229 		return -EFAULT;
3230 
3231 	if (opt.max_window < opt.snd_wnd)
3232 		return -EINVAL;
3233 
3234 	if (after(opt.snd_wl1, tp->rcv_nxt + opt.rcv_wnd))
3235 		return -EINVAL;
3236 
3237 	if (after(opt.rcv_wup, tp->rcv_nxt))
3238 		return -EINVAL;
3239 
3240 	tp->snd_wl1	= opt.snd_wl1;
3241 	tp->snd_wnd	= opt.snd_wnd;
3242 	tp->max_window	= opt.max_window;
3243 
3244 	tp->rcv_wnd	= opt.rcv_wnd;
3245 	tp->rcv_wup	= opt.rcv_wup;
3246 
3247 	return 0;
3248 }
3249 
tcp_repair_options_est(struct sock * sk,sockptr_t optbuf,unsigned int len)3250 static int tcp_repair_options_est(struct sock *sk, sockptr_t optbuf,
3251 		unsigned int len)
3252 {
3253 	struct tcp_sock *tp = tcp_sk(sk);
3254 	struct tcp_repair_opt opt;
3255 	size_t offset = 0;
3256 
3257 	while (len >= sizeof(opt)) {
3258 		if (copy_from_sockptr_offset(&opt, optbuf, offset, sizeof(opt)))
3259 			return -EFAULT;
3260 
3261 		offset += sizeof(opt);
3262 		len -= sizeof(opt);
3263 
3264 		switch (opt.opt_code) {
3265 		case TCPOPT_MSS:
3266 			tp->rx_opt.mss_clamp = opt.opt_val;
3267 			tcp_mtup_init(sk);
3268 			break;
3269 		case TCPOPT_WINDOW:
3270 			{
3271 				u16 snd_wscale = opt.opt_val & 0xFFFF;
3272 				u16 rcv_wscale = opt.opt_val >> 16;
3273 
3274 				if (snd_wscale > TCP_MAX_WSCALE || rcv_wscale > TCP_MAX_WSCALE)
3275 					return -EFBIG;
3276 
3277 				tp->rx_opt.snd_wscale = snd_wscale;
3278 				tp->rx_opt.rcv_wscale = rcv_wscale;
3279 				tp->rx_opt.wscale_ok = 1;
3280 			}
3281 			break;
3282 		case TCPOPT_SACK_PERM:
3283 			if (opt.opt_val != 0)
3284 				return -EINVAL;
3285 
3286 			tp->rx_opt.sack_ok |= TCP_SACK_SEEN;
3287 			break;
3288 		case TCPOPT_TIMESTAMP:
3289 			if (opt.opt_val != 0)
3290 				return -EINVAL;
3291 
3292 			tp->rx_opt.tstamp_ok = 1;
3293 			break;
3294 		}
3295 	}
3296 
3297 	return 0;
3298 }
3299 
3300 DEFINE_STATIC_KEY_FALSE(tcp_tx_delay_enabled);
3301 EXPORT_SYMBOL(tcp_tx_delay_enabled);
3302 
tcp_enable_tx_delay(void)3303 static void tcp_enable_tx_delay(void)
3304 {
3305 	if (!static_branch_unlikely(&tcp_tx_delay_enabled)) {
3306 		static int __tcp_tx_delay_enabled = 0;
3307 
3308 		if (cmpxchg(&__tcp_tx_delay_enabled, 0, 1) == 0) {
3309 			static_branch_enable(&tcp_tx_delay_enabled);
3310 			pr_info("TCP_TX_DELAY enabled\n");
3311 		}
3312 	}
3313 }
3314 
3315 /* When set indicates to always queue non-full frames.  Later the user clears
3316  * this option and we transmit any pending partial frames in the queue.  This is
3317  * meant to be used alongside sendfile() to get properly filled frames when the
3318  * user (for example) must write out headers with a write() call first and then
3319  * use sendfile to send out the data parts.
3320  *
3321  * TCP_CORK can be set together with TCP_NODELAY and it is stronger than
3322  * TCP_NODELAY.
3323  */
__tcp_sock_set_cork(struct sock * sk,bool on)3324 void __tcp_sock_set_cork(struct sock *sk, bool on)
3325 {
3326 	struct tcp_sock *tp = tcp_sk(sk);
3327 
3328 	if (on) {
3329 		tp->nonagle |= TCP_NAGLE_CORK;
3330 	} else {
3331 		tp->nonagle &= ~TCP_NAGLE_CORK;
3332 		if (tp->nonagle & TCP_NAGLE_OFF)
3333 			tp->nonagle |= TCP_NAGLE_PUSH;
3334 		tcp_push_pending_frames(sk);
3335 	}
3336 }
3337 
tcp_sock_set_cork(struct sock * sk,bool on)3338 void tcp_sock_set_cork(struct sock *sk, bool on)
3339 {
3340 	lock_sock(sk);
3341 	__tcp_sock_set_cork(sk, on);
3342 	release_sock(sk);
3343 }
3344 EXPORT_SYMBOL(tcp_sock_set_cork);
3345 
3346 /* TCP_NODELAY is weaker than TCP_CORK, so that this option on corked socket is
3347  * remembered, but it is not activated until cork is cleared.
3348  *
3349  * However, when TCP_NODELAY is set we make an explicit push, which overrides
3350  * even TCP_CORK for currently queued segments.
3351  */
__tcp_sock_set_nodelay(struct sock * sk,bool on)3352 void __tcp_sock_set_nodelay(struct sock *sk, bool on)
3353 {
3354 	if (on) {
3355 		tcp_sk(sk)->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH;
3356 		tcp_push_pending_frames(sk);
3357 	} else {
3358 		tcp_sk(sk)->nonagle &= ~TCP_NAGLE_OFF;
3359 	}
3360 }
3361 
tcp_sock_set_nodelay(struct sock * sk)3362 void tcp_sock_set_nodelay(struct sock *sk)
3363 {
3364 	lock_sock(sk);
3365 	__tcp_sock_set_nodelay(sk, true);
3366 	release_sock(sk);
3367 }
3368 EXPORT_SYMBOL(tcp_sock_set_nodelay);
3369 
__tcp_sock_set_quickack(struct sock * sk,int val)3370 static void __tcp_sock_set_quickack(struct sock *sk, int val)
3371 {
3372 	if (!val) {
3373 		inet_csk_enter_pingpong_mode(sk);
3374 		return;
3375 	}
3376 
3377 	inet_csk_exit_pingpong_mode(sk);
3378 	if ((1 << sk->sk_state) & (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) &&
3379 	    inet_csk_ack_scheduled(sk)) {
3380 		inet_csk(sk)->icsk_ack.pending |= ICSK_ACK_PUSHED;
3381 		tcp_cleanup_rbuf(sk, 1);
3382 		if (!(val & 1))
3383 			inet_csk_enter_pingpong_mode(sk);
3384 	}
3385 }
3386 
tcp_sock_set_quickack(struct sock * sk,int val)3387 void tcp_sock_set_quickack(struct sock *sk, int val)
3388 {
3389 	lock_sock(sk);
3390 	__tcp_sock_set_quickack(sk, val);
3391 	release_sock(sk);
3392 }
3393 EXPORT_SYMBOL(tcp_sock_set_quickack);
3394 
tcp_sock_set_syncnt(struct sock * sk,int val)3395 int tcp_sock_set_syncnt(struct sock *sk, int val)
3396 {
3397 	if (val < 1 || val > MAX_TCP_SYNCNT)
3398 		return -EINVAL;
3399 
3400 	lock_sock(sk);
3401 	inet_csk(sk)->icsk_syn_retries = val;
3402 	release_sock(sk);
3403 	return 0;
3404 }
3405 EXPORT_SYMBOL(tcp_sock_set_syncnt);
3406 
tcp_sock_set_user_timeout(struct sock * sk,u32 val)3407 void tcp_sock_set_user_timeout(struct sock *sk, u32 val)
3408 {
3409 	lock_sock(sk);
3410 	inet_csk(sk)->icsk_user_timeout = val;
3411 	release_sock(sk);
3412 }
3413 EXPORT_SYMBOL(tcp_sock_set_user_timeout);
3414 
tcp_sock_set_keepidle_locked(struct sock * sk,int val)3415 int tcp_sock_set_keepidle_locked(struct sock *sk, int val)
3416 {
3417 	struct tcp_sock *tp = tcp_sk(sk);
3418 
3419 	if (val < 1 || val > MAX_TCP_KEEPIDLE)
3420 		return -EINVAL;
3421 
3422 	tp->keepalive_time = val * HZ;
3423 	if (sock_flag(sk, SOCK_KEEPOPEN) &&
3424 	    !((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN))) {
3425 		u32 elapsed = keepalive_time_elapsed(tp);
3426 
3427 		if (tp->keepalive_time > elapsed)
3428 			elapsed = tp->keepalive_time - elapsed;
3429 		else
3430 			elapsed = 0;
3431 		inet_csk_reset_keepalive_timer(sk, elapsed);
3432 	}
3433 
3434 	return 0;
3435 }
3436 
tcp_sock_set_keepidle(struct sock * sk,int val)3437 int tcp_sock_set_keepidle(struct sock *sk, int val)
3438 {
3439 	int err;
3440 
3441 	lock_sock(sk);
3442 	err = tcp_sock_set_keepidle_locked(sk, val);
3443 	release_sock(sk);
3444 	return err;
3445 }
3446 EXPORT_SYMBOL(tcp_sock_set_keepidle);
3447 
tcp_sock_set_keepintvl(struct sock * sk,int val)3448 int tcp_sock_set_keepintvl(struct sock *sk, int val)
3449 {
3450 	if (val < 1 || val > MAX_TCP_KEEPINTVL)
3451 		return -EINVAL;
3452 
3453 	lock_sock(sk);
3454 	tcp_sk(sk)->keepalive_intvl = val * HZ;
3455 	release_sock(sk);
3456 	return 0;
3457 }
3458 EXPORT_SYMBOL(tcp_sock_set_keepintvl);
3459 
tcp_sock_set_keepcnt(struct sock * sk,int val)3460 int tcp_sock_set_keepcnt(struct sock *sk, int val)
3461 {
3462 	if (val < 1 || val > MAX_TCP_KEEPCNT)
3463 		return -EINVAL;
3464 
3465 	lock_sock(sk);
3466 	tcp_sk(sk)->keepalive_probes = val;
3467 	release_sock(sk);
3468 	return 0;
3469 }
3470 EXPORT_SYMBOL(tcp_sock_set_keepcnt);
3471 
tcp_set_window_clamp(struct sock * sk,int val)3472 int tcp_set_window_clamp(struct sock *sk, int val)
3473 {
3474 	struct tcp_sock *tp = tcp_sk(sk);
3475 
3476 	if (!val) {
3477 		if (sk->sk_state != TCP_CLOSE)
3478 			return -EINVAL;
3479 		tp->window_clamp = 0;
3480 	} else {
3481 		tp->window_clamp = val < SOCK_MIN_RCVBUF / 2 ?
3482 			SOCK_MIN_RCVBUF / 2 : val;
3483 		tp->rcv_ssthresh = min(tp->rcv_wnd, tp->window_clamp);
3484 	}
3485 	return 0;
3486 }
3487 
3488 /*
3489  *	Socket option code for TCP.
3490  */
do_tcp_setsockopt(struct sock * sk,int level,int optname,sockptr_t optval,unsigned int optlen)3491 int do_tcp_setsockopt(struct sock *sk, int level, int optname,
3492 		      sockptr_t optval, unsigned int optlen)
3493 {
3494 	struct tcp_sock *tp = tcp_sk(sk);
3495 	struct inet_connection_sock *icsk = inet_csk(sk);
3496 	struct net *net = sock_net(sk);
3497 	int val;
3498 	int err = 0;
3499 
3500 	/* These are data/string values, all the others are ints */
3501 	switch (optname) {
3502 	case TCP_CONGESTION: {
3503 		char name[TCP_CA_NAME_MAX];
3504 
3505 		if (optlen < 1)
3506 			return -EINVAL;
3507 
3508 		val = strncpy_from_sockptr(name, optval,
3509 					min_t(long, TCP_CA_NAME_MAX-1, optlen));
3510 		if (val < 0)
3511 			return -EFAULT;
3512 		name[val] = 0;
3513 
3514 		sockopt_lock_sock(sk);
3515 		err = tcp_set_congestion_control(sk, name, !has_current_bpf_ctx(),
3516 						 sockopt_ns_capable(sock_net(sk)->user_ns,
3517 								    CAP_NET_ADMIN));
3518 		sockopt_release_sock(sk);
3519 		return err;
3520 	}
3521 	case TCP_ULP: {
3522 		char name[TCP_ULP_NAME_MAX];
3523 
3524 		if (optlen < 1)
3525 			return -EINVAL;
3526 
3527 		val = strncpy_from_sockptr(name, optval,
3528 					min_t(long, TCP_ULP_NAME_MAX - 1,
3529 					      optlen));
3530 		if (val < 0)
3531 			return -EFAULT;
3532 		name[val] = 0;
3533 
3534 		sockopt_lock_sock(sk);
3535 		err = tcp_set_ulp(sk, name);
3536 		sockopt_release_sock(sk);
3537 		return err;
3538 	}
3539 	case TCP_FASTOPEN_KEY: {
3540 		__u8 key[TCP_FASTOPEN_KEY_BUF_LENGTH];
3541 		__u8 *backup_key = NULL;
3542 
3543 		/* Allow a backup key as well to facilitate key rotation
3544 		 * First key is the active one.
3545 		 */
3546 		if (optlen != TCP_FASTOPEN_KEY_LENGTH &&
3547 		    optlen != TCP_FASTOPEN_KEY_BUF_LENGTH)
3548 			return -EINVAL;
3549 
3550 		if (copy_from_sockptr(key, optval, optlen))
3551 			return -EFAULT;
3552 
3553 		if (optlen == TCP_FASTOPEN_KEY_BUF_LENGTH)
3554 			backup_key = key + TCP_FASTOPEN_KEY_LENGTH;
3555 
3556 		return tcp_fastopen_reset_cipher(net, sk, key, backup_key);
3557 	}
3558 	default:
3559 		/* fallthru */
3560 		break;
3561 	}
3562 
3563 	if (optlen < sizeof(int))
3564 		return -EINVAL;
3565 
3566 	if (copy_from_sockptr(&val, optval, sizeof(val)))
3567 		return -EFAULT;
3568 
3569 	sockopt_lock_sock(sk);
3570 
3571 	switch (optname) {
3572 	case TCP_MAXSEG:
3573 		/* Values greater than interface MTU won't take effect. However
3574 		 * at the point when this call is done we typically don't yet
3575 		 * know which interface is going to be used
3576 		 */
3577 		if (val && (val < TCP_MIN_MSS || val > MAX_TCP_WINDOW)) {
3578 			err = -EINVAL;
3579 			break;
3580 		}
3581 		tp->rx_opt.user_mss = val;
3582 		break;
3583 
3584 	case TCP_NODELAY:
3585 		__tcp_sock_set_nodelay(sk, val);
3586 		break;
3587 
3588 	case TCP_THIN_LINEAR_TIMEOUTS:
3589 		if (val < 0 || val > 1)
3590 			err = -EINVAL;
3591 		else
3592 			tp->thin_lto = val;
3593 		break;
3594 
3595 	case TCP_THIN_DUPACK:
3596 		if (val < 0 || val > 1)
3597 			err = -EINVAL;
3598 		break;
3599 
3600 	case TCP_REPAIR:
3601 		if (!tcp_can_repair_sock(sk))
3602 			err = -EPERM;
3603 		else if (val == TCP_REPAIR_ON) {
3604 			tp->repair = 1;
3605 			sk->sk_reuse = SK_FORCE_REUSE;
3606 			tp->repair_queue = TCP_NO_QUEUE;
3607 		} else if (val == TCP_REPAIR_OFF) {
3608 			tp->repair = 0;
3609 			sk->sk_reuse = SK_NO_REUSE;
3610 			tcp_send_window_probe(sk);
3611 		} else if (val == TCP_REPAIR_OFF_NO_WP) {
3612 			tp->repair = 0;
3613 			sk->sk_reuse = SK_NO_REUSE;
3614 		} else
3615 			err = -EINVAL;
3616 
3617 		break;
3618 
3619 	case TCP_REPAIR_QUEUE:
3620 		if (!tp->repair)
3621 			err = -EPERM;
3622 		else if ((unsigned int)val < TCP_QUEUES_NR)
3623 			tp->repair_queue = val;
3624 		else
3625 			err = -EINVAL;
3626 		break;
3627 
3628 	case TCP_QUEUE_SEQ:
3629 		if (sk->sk_state != TCP_CLOSE) {
3630 			err = -EPERM;
3631 		} else if (tp->repair_queue == TCP_SEND_QUEUE) {
3632 			if (!tcp_rtx_queue_empty(sk))
3633 				err = -EPERM;
3634 			else
3635 				WRITE_ONCE(tp->write_seq, val);
3636 		} else if (tp->repair_queue == TCP_RECV_QUEUE) {
3637 			if (tp->rcv_nxt != tp->copied_seq) {
3638 				err = -EPERM;
3639 			} else {
3640 				WRITE_ONCE(tp->rcv_nxt, val);
3641 				WRITE_ONCE(tp->copied_seq, val);
3642 			}
3643 		} else {
3644 			err = -EINVAL;
3645 		}
3646 		break;
3647 
3648 	case TCP_REPAIR_OPTIONS:
3649 		if (!tp->repair)
3650 			err = -EINVAL;
3651 		else if (sk->sk_state == TCP_ESTABLISHED && !tp->bytes_sent)
3652 			err = tcp_repair_options_est(sk, optval, optlen);
3653 		else
3654 			err = -EPERM;
3655 		break;
3656 
3657 	case TCP_CORK:
3658 		__tcp_sock_set_cork(sk, val);
3659 		break;
3660 
3661 	case TCP_KEEPIDLE:
3662 		err = tcp_sock_set_keepidle_locked(sk, val);
3663 		break;
3664 	case TCP_KEEPINTVL:
3665 		if (val < 1 || val > MAX_TCP_KEEPINTVL)
3666 			err = -EINVAL;
3667 		else
3668 			tp->keepalive_intvl = val * HZ;
3669 		break;
3670 	case TCP_KEEPCNT:
3671 		if (val < 1 || val > MAX_TCP_KEEPCNT)
3672 			err = -EINVAL;
3673 		else
3674 			tp->keepalive_probes = val;
3675 		break;
3676 	case TCP_SYNCNT:
3677 		if (val < 1 || val > MAX_TCP_SYNCNT)
3678 			err = -EINVAL;
3679 		else
3680 			icsk->icsk_syn_retries = val;
3681 		break;
3682 
3683 	case TCP_SAVE_SYN:
3684 		/* 0: disable, 1: enable, 2: start from ether_header */
3685 		if (val < 0 || val > 2)
3686 			err = -EINVAL;
3687 		else
3688 			tp->save_syn = val;
3689 		break;
3690 
3691 	case TCP_LINGER2:
3692 		if (val < 0)
3693 			tp->linger2 = -1;
3694 		else if (val > TCP_FIN_TIMEOUT_MAX / HZ)
3695 			tp->linger2 = TCP_FIN_TIMEOUT_MAX;
3696 		else
3697 			tp->linger2 = val * HZ;
3698 		break;
3699 
3700 	case TCP_DEFER_ACCEPT:
3701 		/* Translate value in seconds to number of retransmits */
3702 		icsk->icsk_accept_queue.rskq_defer_accept =
3703 			secs_to_retrans(val, TCP_TIMEOUT_INIT / HZ,
3704 					TCP_RTO_MAX / HZ);
3705 		break;
3706 
3707 	case TCP_WINDOW_CLAMP:
3708 		err = tcp_set_window_clamp(sk, val);
3709 		break;
3710 
3711 	case TCP_QUICKACK:
3712 		__tcp_sock_set_quickack(sk, val);
3713 		break;
3714 
3715 #ifdef CONFIG_TCP_MD5SIG
3716 	case TCP_MD5SIG:
3717 	case TCP_MD5SIG_EXT:
3718 		err = tp->af_specific->md5_parse(sk, optname, optval, optlen);
3719 		break;
3720 #endif
3721 	case TCP_USER_TIMEOUT:
3722 		/* Cap the max time in ms TCP will retry or probe the window
3723 		 * before giving up and aborting (ETIMEDOUT) a connection.
3724 		 */
3725 		if (val < 0)
3726 			err = -EINVAL;
3727 		else
3728 			icsk->icsk_user_timeout = val;
3729 		break;
3730 
3731 	case TCP_FASTOPEN:
3732 		if (val >= 0 && ((1 << sk->sk_state) & (TCPF_CLOSE |
3733 		    TCPF_LISTEN))) {
3734 			tcp_fastopen_init_key_once(net);
3735 
3736 			fastopen_queue_tune(sk, val);
3737 		} else {
3738 			err = -EINVAL;
3739 		}
3740 		break;
3741 	case TCP_FASTOPEN_CONNECT:
3742 		if (val > 1 || val < 0) {
3743 			err = -EINVAL;
3744 		} else if (READ_ONCE(net->ipv4.sysctl_tcp_fastopen) &
3745 			   TFO_CLIENT_ENABLE) {
3746 			if (sk->sk_state == TCP_CLOSE)
3747 				tp->fastopen_connect = val;
3748 			else
3749 				err = -EINVAL;
3750 		} else {
3751 			err = -EOPNOTSUPP;
3752 		}
3753 		break;
3754 	case TCP_FASTOPEN_NO_COOKIE:
3755 		if (val > 1 || val < 0)
3756 			err = -EINVAL;
3757 		else if (!((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
3758 			err = -EINVAL;
3759 		else
3760 			tp->fastopen_no_cookie = val;
3761 		break;
3762 	case TCP_TIMESTAMP:
3763 		if (!tp->repair)
3764 			err = -EPERM;
3765 		else
3766 			tp->tsoffset = val - tcp_time_stamp_raw();
3767 		break;
3768 	case TCP_REPAIR_WINDOW:
3769 		err = tcp_repair_set_window(tp, optval, optlen);
3770 		break;
3771 	case TCP_NOTSENT_LOWAT:
3772 		tp->notsent_lowat = val;
3773 		sk->sk_write_space(sk);
3774 		break;
3775 	case TCP_INQ:
3776 		if (val > 1 || val < 0)
3777 			err = -EINVAL;
3778 		else
3779 			tp->recvmsg_inq = val;
3780 		break;
3781 	case TCP_TX_DELAY:
3782 		if (val)
3783 			tcp_enable_tx_delay();
3784 		tp->tcp_tx_delay = val;
3785 		break;
3786 	default:
3787 		err = -ENOPROTOOPT;
3788 		break;
3789 	}
3790 
3791 	sockopt_release_sock(sk);
3792 	return err;
3793 }
3794 
tcp_setsockopt(struct sock * sk,int level,int optname,sockptr_t optval,unsigned int optlen)3795 int tcp_setsockopt(struct sock *sk, int level, int optname, sockptr_t optval,
3796 		   unsigned int optlen)
3797 {
3798 	const struct inet_connection_sock *icsk = inet_csk(sk);
3799 
3800 	if (level != SOL_TCP)
3801 		/* Paired with WRITE_ONCE() in do_ipv6_setsockopt() and tcp_v6_connect() */
3802 		return READ_ONCE(icsk->icsk_af_ops)->setsockopt(sk, level, optname,
3803 								optval, optlen);
3804 	return do_tcp_setsockopt(sk, level, optname, optval, optlen);
3805 }
3806 EXPORT_SYMBOL(tcp_setsockopt);
3807 
tcp_get_info_chrono_stats(const struct tcp_sock * tp,struct tcp_info * info)3808 static void tcp_get_info_chrono_stats(const struct tcp_sock *tp,
3809 				      struct tcp_info *info)
3810 {
3811 	u64 stats[__TCP_CHRONO_MAX], total = 0;
3812 	enum tcp_chrono i;
3813 
3814 	for (i = TCP_CHRONO_BUSY; i < __TCP_CHRONO_MAX; ++i) {
3815 		stats[i] = tp->chrono_stat[i - 1];
3816 		if (i == tp->chrono_type)
3817 			stats[i] += tcp_jiffies32 - tp->chrono_start;
3818 		stats[i] *= USEC_PER_SEC / HZ;
3819 		total += stats[i];
3820 	}
3821 
3822 	info->tcpi_busy_time = total;
3823 	info->tcpi_rwnd_limited = stats[TCP_CHRONO_RWND_LIMITED];
3824 	info->tcpi_sndbuf_limited = stats[TCP_CHRONO_SNDBUF_LIMITED];
3825 }
3826 
3827 /* Return information about state of tcp endpoint in API format. */
tcp_get_info(struct sock * sk,struct tcp_info * info)3828 void tcp_get_info(struct sock *sk, struct tcp_info *info)
3829 {
3830 	const struct tcp_sock *tp = tcp_sk(sk); /* iff sk_type == SOCK_STREAM */
3831 	const struct inet_connection_sock *icsk = inet_csk(sk);
3832 	unsigned long rate;
3833 	u32 now;
3834 	u64 rate64;
3835 	bool slow;
3836 
3837 	memset(info, 0, sizeof(*info));
3838 	if (sk->sk_type != SOCK_STREAM)
3839 		return;
3840 
3841 	info->tcpi_state = inet_sk_state_load(sk);
3842 
3843 	/* Report meaningful fields for all TCP states, including listeners */
3844 	rate = READ_ONCE(sk->sk_pacing_rate);
3845 	rate64 = (rate != ~0UL) ? rate : ~0ULL;
3846 	info->tcpi_pacing_rate = rate64;
3847 
3848 	rate = READ_ONCE(sk->sk_max_pacing_rate);
3849 	rate64 = (rate != ~0UL) ? rate : ~0ULL;
3850 	info->tcpi_max_pacing_rate = rate64;
3851 
3852 	info->tcpi_reordering = tp->reordering;
3853 	info->tcpi_snd_cwnd = tcp_snd_cwnd(tp);
3854 
3855 	if (info->tcpi_state == TCP_LISTEN) {
3856 		/* listeners aliased fields :
3857 		 * tcpi_unacked -> Number of children ready for accept()
3858 		 * tcpi_sacked  -> max backlog
3859 		 */
3860 		info->tcpi_unacked = READ_ONCE(sk->sk_ack_backlog);
3861 		info->tcpi_sacked = READ_ONCE(sk->sk_max_ack_backlog);
3862 		return;
3863 	}
3864 
3865 	slow = lock_sock_fast(sk);
3866 
3867 	info->tcpi_ca_state = icsk->icsk_ca_state;
3868 	info->tcpi_retransmits = icsk->icsk_retransmits;
3869 	info->tcpi_probes = icsk->icsk_probes_out;
3870 	info->tcpi_backoff = icsk->icsk_backoff;
3871 
3872 	if (tp->rx_opt.tstamp_ok)
3873 		info->tcpi_options |= TCPI_OPT_TIMESTAMPS;
3874 	if (tcp_is_sack(tp))
3875 		info->tcpi_options |= TCPI_OPT_SACK;
3876 	if (tp->rx_opt.wscale_ok) {
3877 		info->tcpi_options |= TCPI_OPT_WSCALE;
3878 		info->tcpi_snd_wscale = tp->rx_opt.snd_wscale;
3879 		info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale;
3880 	}
3881 
3882 	if (tp->ecn_flags & TCP_ECN_OK)
3883 		info->tcpi_options |= TCPI_OPT_ECN;
3884 	if (tp->ecn_flags & TCP_ECN_SEEN)
3885 		info->tcpi_options |= TCPI_OPT_ECN_SEEN;
3886 	if (tp->syn_data_acked)
3887 		info->tcpi_options |= TCPI_OPT_SYN_DATA;
3888 
3889 	info->tcpi_rto = jiffies_to_usecs(icsk->icsk_rto);
3890 	info->tcpi_ato = jiffies_to_usecs(icsk->icsk_ack.ato);
3891 	info->tcpi_snd_mss = tp->mss_cache;
3892 	info->tcpi_rcv_mss = icsk->icsk_ack.rcv_mss;
3893 
3894 	info->tcpi_unacked = tp->packets_out;
3895 	info->tcpi_sacked = tp->sacked_out;
3896 
3897 	info->tcpi_lost = tp->lost_out;
3898 	info->tcpi_retrans = tp->retrans_out;
3899 
3900 	now = tcp_jiffies32;
3901 	info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime);
3902 	info->tcpi_last_data_recv = jiffies_to_msecs(now - icsk->icsk_ack.lrcvtime);
3903 	info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp);
3904 
3905 	info->tcpi_pmtu = icsk->icsk_pmtu_cookie;
3906 	info->tcpi_rcv_ssthresh = tp->rcv_ssthresh;
3907 	info->tcpi_rtt = tp->srtt_us >> 3;
3908 	info->tcpi_rttvar = tp->mdev_us >> 2;
3909 	info->tcpi_snd_ssthresh = tp->snd_ssthresh;
3910 	info->tcpi_advmss = tp->advmss;
3911 
3912 	info->tcpi_rcv_rtt = tp->rcv_rtt_est.rtt_us >> 3;
3913 	info->tcpi_rcv_space = tp->rcvq_space.space;
3914 
3915 	info->tcpi_total_retrans = tp->total_retrans;
3916 
3917 	info->tcpi_bytes_acked = tp->bytes_acked;
3918 	info->tcpi_bytes_received = tp->bytes_received;
3919 	info->tcpi_notsent_bytes = max_t(int, 0, tp->write_seq - tp->snd_nxt);
3920 	tcp_get_info_chrono_stats(tp, info);
3921 
3922 	info->tcpi_segs_out = tp->segs_out;
3923 
3924 	/* segs_in and data_segs_in can be updated from tcp_segs_in() from BH */
3925 	info->tcpi_segs_in = READ_ONCE(tp->segs_in);
3926 	info->tcpi_data_segs_in = READ_ONCE(tp->data_segs_in);
3927 
3928 	info->tcpi_min_rtt = tcp_min_rtt(tp);
3929 	info->tcpi_data_segs_out = tp->data_segs_out;
3930 
3931 	info->tcpi_delivery_rate_app_limited = tp->rate_app_limited ? 1 : 0;
3932 	rate64 = tcp_compute_delivery_rate(tp);
3933 	if (rate64)
3934 		info->tcpi_delivery_rate = rate64;
3935 	info->tcpi_delivered = tp->delivered;
3936 	info->tcpi_delivered_ce = tp->delivered_ce;
3937 	info->tcpi_bytes_sent = tp->bytes_sent;
3938 	info->tcpi_bytes_retrans = tp->bytes_retrans;
3939 	info->tcpi_dsack_dups = tp->dsack_dups;
3940 	info->tcpi_reord_seen = tp->reord_seen;
3941 	info->tcpi_rcv_ooopack = tp->rcv_ooopack;
3942 	info->tcpi_snd_wnd = tp->snd_wnd;
3943 	info->tcpi_fastopen_client_fail = tp->fastopen_client_fail;
3944 	unlock_sock_fast(sk, slow);
3945 }
3946 EXPORT_SYMBOL_GPL(tcp_get_info);
3947 
tcp_opt_stats_get_size(void)3948 static size_t tcp_opt_stats_get_size(void)
3949 {
3950 	return
3951 		nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_BUSY */
3952 		nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_RWND_LIMITED */
3953 		nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_SNDBUF_LIMITED */
3954 		nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_DATA_SEGS_OUT */
3955 		nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_TOTAL_RETRANS */
3956 		nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_PACING_RATE */
3957 		nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_DELIVERY_RATE */
3958 		nla_total_size(sizeof(u32)) + /* TCP_NLA_SND_CWND */
3959 		nla_total_size(sizeof(u32)) + /* TCP_NLA_REORDERING */
3960 		nla_total_size(sizeof(u32)) + /* TCP_NLA_MIN_RTT */
3961 		nla_total_size(sizeof(u8)) + /* TCP_NLA_RECUR_RETRANS */
3962 		nla_total_size(sizeof(u8)) + /* TCP_NLA_DELIVERY_RATE_APP_LMT */
3963 		nla_total_size(sizeof(u32)) + /* TCP_NLA_SNDQ_SIZE */
3964 		nla_total_size(sizeof(u8)) + /* TCP_NLA_CA_STATE */
3965 		nla_total_size(sizeof(u32)) + /* TCP_NLA_SND_SSTHRESH */
3966 		nla_total_size(sizeof(u32)) + /* TCP_NLA_DELIVERED */
3967 		nla_total_size(sizeof(u32)) + /* TCP_NLA_DELIVERED_CE */
3968 		nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_BYTES_SENT */
3969 		nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_BYTES_RETRANS */
3970 		nla_total_size(sizeof(u32)) + /* TCP_NLA_DSACK_DUPS */
3971 		nla_total_size(sizeof(u32)) + /* TCP_NLA_REORD_SEEN */
3972 		nla_total_size(sizeof(u32)) + /* TCP_NLA_SRTT */
3973 		nla_total_size(sizeof(u16)) + /* TCP_NLA_TIMEOUT_REHASH */
3974 		nla_total_size(sizeof(u32)) + /* TCP_NLA_BYTES_NOTSENT */
3975 		nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_EDT */
3976 		nla_total_size(sizeof(u8)) + /* TCP_NLA_TTL */
3977 		0;
3978 }
3979 
3980 /* Returns TTL or hop limit of an incoming packet from skb. */
tcp_skb_ttl_or_hop_limit(const struct sk_buff * skb)3981 static u8 tcp_skb_ttl_or_hop_limit(const struct sk_buff *skb)
3982 {
3983 	if (skb->protocol == htons(ETH_P_IP))
3984 		return ip_hdr(skb)->ttl;
3985 	else if (skb->protocol == htons(ETH_P_IPV6))
3986 		return ipv6_hdr(skb)->hop_limit;
3987 	else
3988 		return 0;
3989 }
3990 
tcp_get_timestamping_opt_stats(const struct sock * sk,const struct sk_buff * orig_skb,const struct sk_buff * ack_skb)3991 struct sk_buff *tcp_get_timestamping_opt_stats(const struct sock *sk,
3992 					       const struct sk_buff *orig_skb,
3993 					       const struct sk_buff *ack_skb)
3994 {
3995 	const struct tcp_sock *tp = tcp_sk(sk);
3996 	struct sk_buff *stats;
3997 	struct tcp_info info;
3998 	unsigned long rate;
3999 	u64 rate64;
4000 
4001 	stats = alloc_skb(tcp_opt_stats_get_size(), GFP_ATOMIC);
4002 	if (!stats)
4003 		return NULL;
4004 
4005 	tcp_get_info_chrono_stats(tp, &info);
4006 	nla_put_u64_64bit(stats, TCP_NLA_BUSY,
4007 			  info.tcpi_busy_time, TCP_NLA_PAD);
4008 	nla_put_u64_64bit(stats, TCP_NLA_RWND_LIMITED,
4009 			  info.tcpi_rwnd_limited, TCP_NLA_PAD);
4010 	nla_put_u64_64bit(stats, TCP_NLA_SNDBUF_LIMITED,
4011 			  info.tcpi_sndbuf_limited, TCP_NLA_PAD);
4012 	nla_put_u64_64bit(stats, TCP_NLA_DATA_SEGS_OUT,
4013 			  tp->data_segs_out, TCP_NLA_PAD);
4014 	nla_put_u64_64bit(stats, TCP_NLA_TOTAL_RETRANS,
4015 			  tp->total_retrans, TCP_NLA_PAD);
4016 
4017 	rate = READ_ONCE(sk->sk_pacing_rate);
4018 	rate64 = (rate != ~0UL) ? rate : ~0ULL;
4019 	nla_put_u64_64bit(stats, TCP_NLA_PACING_RATE, rate64, TCP_NLA_PAD);
4020 
4021 	rate64 = tcp_compute_delivery_rate(tp);
4022 	nla_put_u64_64bit(stats, TCP_NLA_DELIVERY_RATE, rate64, TCP_NLA_PAD);
4023 
4024 	nla_put_u32(stats, TCP_NLA_SND_CWND, tcp_snd_cwnd(tp));
4025 	nla_put_u32(stats, TCP_NLA_REORDERING, tp->reordering);
4026 	nla_put_u32(stats, TCP_NLA_MIN_RTT, tcp_min_rtt(tp));
4027 
4028 	nla_put_u8(stats, TCP_NLA_RECUR_RETRANS, inet_csk(sk)->icsk_retransmits);
4029 	nla_put_u8(stats, TCP_NLA_DELIVERY_RATE_APP_LMT, !!tp->rate_app_limited);
4030 	nla_put_u32(stats, TCP_NLA_SND_SSTHRESH, tp->snd_ssthresh);
4031 	nla_put_u32(stats, TCP_NLA_DELIVERED, tp->delivered);
4032 	nla_put_u32(stats, TCP_NLA_DELIVERED_CE, tp->delivered_ce);
4033 
4034 	nla_put_u32(stats, TCP_NLA_SNDQ_SIZE, tp->write_seq - tp->snd_una);
4035 	nla_put_u8(stats, TCP_NLA_CA_STATE, inet_csk(sk)->icsk_ca_state);
4036 
4037 	nla_put_u64_64bit(stats, TCP_NLA_BYTES_SENT, tp->bytes_sent,
4038 			  TCP_NLA_PAD);
4039 	nla_put_u64_64bit(stats, TCP_NLA_BYTES_RETRANS, tp->bytes_retrans,
4040 			  TCP_NLA_PAD);
4041 	nla_put_u32(stats, TCP_NLA_DSACK_DUPS, tp->dsack_dups);
4042 	nla_put_u32(stats, TCP_NLA_REORD_SEEN, tp->reord_seen);
4043 	nla_put_u32(stats, TCP_NLA_SRTT, tp->srtt_us >> 3);
4044 	nla_put_u16(stats, TCP_NLA_TIMEOUT_REHASH, tp->timeout_rehash);
4045 	nla_put_u32(stats, TCP_NLA_BYTES_NOTSENT,
4046 		    max_t(int, 0, tp->write_seq - tp->snd_nxt));
4047 	nla_put_u64_64bit(stats, TCP_NLA_EDT, orig_skb->skb_mstamp_ns,
4048 			  TCP_NLA_PAD);
4049 	if (ack_skb)
4050 		nla_put_u8(stats, TCP_NLA_TTL,
4051 			   tcp_skb_ttl_or_hop_limit(ack_skb));
4052 
4053 	return stats;
4054 }
4055 
do_tcp_getsockopt(struct sock * sk,int level,int optname,sockptr_t optval,sockptr_t optlen)4056 int do_tcp_getsockopt(struct sock *sk, int level,
4057 		      int optname, sockptr_t optval, sockptr_t optlen)
4058 {
4059 	struct inet_connection_sock *icsk = inet_csk(sk);
4060 	struct tcp_sock *tp = tcp_sk(sk);
4061 	struct net *net = sock_net(sk);
4062 	int val, len;
4063 
4064 	if (copy_from_sockptr(&len, optlen, sizeof(int)))
4065 		return -EFAULT;
4066 
4067 	len = min_t(unsigned int, len, sizeof(int));
4068 
4069 	if (len < 0)
4070 		return -EINVAL;
4071 
4072 	switch (optname) {
4073 	case TCP_MAXSEG:
4074 		val = tp->mss_cache;
4075 		if (!val && ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
4076 			val = tp->rx_opt.user_mss;
4077 		if (tp->repair)
4078 			val = tp->rx_opt.mss_clamp;
4079 		break;
4080 	case TCP_NODELAY:
4081 		val = !!(tp->nonagle&TCP_NAGLE_OFF);
4082 		break;
4083 	case TCP_CORK:
4084 		val = !!(tp->nonagle&TCP_NAGLE_CORK);
4085 		break;
4086 	case TCP_KEEPIDLE:
4087 		val = keepalive_time_when(tp) / HZ;
4088 		break;
4089 	case TCP_KEEPINTVL:
4090 		val = keepalive_intvl_when(tp) / HZ;
4091 		break;
4092 	case TCP_KEEPCNT:
4093 		val = keepalive_probes(tp);
4094 		break;
4095 	case TCP_SYNCNT:
4096 		val = icsk->icsk_syn_retries ? :
4097 			READ_ONCE(net->ipv4.sysctl_tcp_syn_retries);
4098 		break;
4099 	case TCP_LINGER2:
4100 		val = tp->linger2;
4101 		if (val >= 0)
4102 			val = (val ? : READ_ONCE(net->ipv4.sysctl_tcp_fin_timeout)) / HZ;
4103 		break;
4104 	case TCP_DEFER_ACCEPT:
4105 		val = retrans_to_secs(icsk->icsk_accept_queue.rskq_defer_accept,
4106 				      TCP_TIMEOUT_INIT / HZ, TCP_RTO_MAX / HZ);
4107 		break;
4108 	case TCP_WINDOW_CLAMP:
4109 		val = tp->window_clamp;
4110 		break;
4111 	case TCP_INFO: {
4112 		struct tcp_info info;
4113 
4114 		if (copy_from_sockptr(&len, optlen, sizeof(int)))
4115 			return -EFAULT;
4116 
4117 		tcp_get_info(sk, &info);
4118 
4119 		len = min_t(unsigned int, len, sizeof(info));
4120 		if (copy_to_sockptr(optlen, &len, sizeof(int)))
4121 			return -EFAULT;
4122 		if (copy_to_sockptr(optval, &info, len))
4123 			return -EFAULT;
4124 		return 0;
4125 	}
4126 	case TCP_CC_INFO: {
4127 		const struct tcp_congestion_ops *ca_ops;
4128 		union tcp_cc_info info;
4129 		size_t sz = 0;
4130 		int attr;
4131 
4132 		if (copy_from_sockptr(&len, optlen, sizeof(int)))
4133 			return -EFAULT;
4134 
4135 		ca_ops = icsk->icsk_ca_ops;
4136 		if (ca_ops && ca_ops->get_info)
4137 			sz = ca_ops->get_info(sk, ~0U, &attr, &info);
4138 
4139 		len = min_t(unsigned int, len, sz);
4140 		if (copy_to_sockptr(optlen, &len, sizeof(int)))
4141 			return -EFAULT;
4142 		if (copy_to_sockptr(optval, &info, len))
4143 			return -EFAULT;
4144 		return 0;
4145 	}
4146 	case TCP_QUICKACK:
4147 		val = !inet_csk_in_pingpong_mode(sk);
4148 		break;
4149 
4150 	case TCP_CONGESTION:
4151 		if (copy_from_sockptr(&len, optlen, sizeof(int)))
4152 			return -EFAULT;
4153 		len = min_t(unsigned int, len, TCP_CA_NAME_MAX);
4154 		if (copy_to_sockptr(optlen, &len, sizeof(int)))
4155 			return -EFAULT;
4156 		if (copy_to_sockptr(optval, icsk->icsk_ca_ops->name, len))
4157 			return -EFAULT;
4158 		return 0;
4159 
4160 	case TCP_ULP:
4161 		if (copy_from_sockptr(&len, optlen, sizeof(int)))
4162 			return -EFAULT;
4163 		len = min_t(unsigned int, len, TCP_ULP_NAME_MAX);
4164 		if (!icsk->icsk_ulp_ops) {
4165 			len = 0;
4166 			if (copy_to_sockptr(optlen, &len, sizeof(int)))
4167 				return -EFAULT;
4168 			return 0;
4169 		}
4170 		if (copy_to_sockptr(optlen, &len, sizeof(int)))
4171 			return -EFAULT;
4172 		if (copy_to_sockptr(optval, icsk->icsk_ulp_ops->name, len))
4173 			return -EFAULT;
4174 		return 0;
4175 
4176 	case TCP_FASTOPEN_KEY: {
4177 		u64 key[TCP_FASTOPEN_KEY_BUF_LENGTH / sizeof(u64)];
4178 		unsigned int key_len;
4179 
4180 		if (copy_from_sockptr(&len, optlen, sizeof(int)))
4181 			return -EFAULT;
4182 
4183 		key_len = tcp_fastopen_get_cipher(net, icsk, key) *
4184 				TCP_FASTOPEN_KEY_LENGTH;
4185 		len = min_t(unsigned int, len, key_len);
4186 		if (copy_to_sockptr(optlen, &len, sizeof(int)))
4187 			return -EFAULT;
4188 		if (copy_to_sockptr(optval, key, len))
4189 			return -EFAULT;
4190 		return 0;
4191 	}
4192 	case TCP_THIN_LINEAR_TIMEOUTS:
4193 		val = tp->thin_lto;
4194 		break;
4195 
4196 	case TCP_THIN_DUPACK:
4197 		val = 0;
4198 		break;
4199 
4200 	case TCP_REPAIR:
4201 		val = tp->repair;
4202 		break;
4203 
4204 	case TCP_REPAIR_QUEUE:
4205 		if (tp->repair)
4206 			val = tp->repair_queue;
4207 		else
4208 			return -EINVAL;
4209 		break;
4210 
4211 	case TCP_REPAIR_WINDOW: {
4212 		struct tcp_repair_window opt;
4213 
4214 		if (copy_from_sockptr(&len, optlen, sizeof(int)))
4215 			return -EFAULT;
4216 
4217 		if (len != sizeof(opt))
4218 			return -EINVAL;
4219 
4220 		if (!tp->repair)
4221 			return -EPERM;
4222 
4223 		opt.snd_wl1	= tp->snd_wl1;
4224 		opt.snd_wnd	= tp->snd_wnd;
4225 		opt.max_window	= tp->max_window;
4226 		opt.rcv_wnd	= tp->rcv_wnd;
4227 		opt.rcv_wup	= tp->rcv_wup;
4228 
4229 		if (copy_to_sockptr(optval, &opt, len))
4230 			return -EFAULT;
4231 		return 0;
4232 	}
4233 	case TCP_QUEUE_SEQ:
4234 		if (tp->repair_queue == TCP_SEND_QUEUE)
4235 			val = tp->write_seq;
4236 		else if (tp->repair_queue == TCP_RECV_QUEUE)
4237 			val = tp->rcv_nxt;
4238 		else
4239 			return -EINVAL;
4240 		break;
4241 
4242 	case TCP_USER_TIMEOUT:
4243 		val = icsk->icsk_user_timeout;
4244 		break;
4245 
4246 	case TCP_FASTOPEN:
4247 		val = icsk->icsk_accept_queue.fastopenq.max_qlen;
4248 		break;
4249 
4250 	case TCP_FASTOPEN_CONNECT:
4251 		val = tp->fastopen_connect;
4252 		break;
4253 
4254 	case TCP_FASTOPEN_NO_COOKIE:
4255 		val = tp->fastopen_no_cookie;
4256 		break;
4257 
4258 	case TCP_TX_DELAY:
4259 		val = tp->tcp_tx_delay;
4260 		break;
4261 
4262 	case TCP_TIMESTAMP:
4263 		val = tcp_time_stamp_raw() + tp->tsoffset;
4264 		break;
4265 	case TCP_NOTSENT_LOWAT:
4266 		val = tp->notsent_lowat;
4267 		break;
4268 	case TCP_INQ:
4269 		val = tp->recvmsg_inq;
4270 		break;
4271 	case TCP_SAVE_SYN:
4272 		val = tp->save_syn;
4273 		break;
4274 	case TCP_SAVED_SYN: {
4275 		if (copy_from_sockptr(&len, optlen, sizeof(int)))
4276 			return -EFAULT;
4277 
4278 		sockopt_lock_sock(sk);
4279 		if (tp->saved_syn) {
4280 			if (len < tcp_saved_syn_len(tp->saved_syn)) {
4281 				len = tcp_saved_syn_len(tp->saved_syn);
4282 				if (copy_to_sockptr(optlen, &len, sizeof(int))) {
4283 					sockopt_release_sock(sk);
4284 					return -EFAULT;
4285 				}
4286 				sockopt_release_sock(sk);
4287 				return -EINVAL;
4288 			}
4289 			len = tcp_saved_syn_len(tp->saved_syn);
4290 			if (copy_to_sockptr(optlen, &len, sizeof(int))) {
4291 				sockopt_release_sock(sk);
4292 				return -EFAULT;
4293 			}
4294 			if (copy_to_sockptr(optval, tp->saved_syn->data, len)) {
4295 				sockopt_release_sock(sk);
4296 				return -EFAULT;
4297 			}
4298 			tcp_saved_syn_free(tp);
4299 			sockopt_release_sock(sk);
4300 		} else {
4301 			sockopt_release_sock(sk);
4302 			len = 0;
4303 			if (copy_to_sockptr(optlen, &len, sizeof(int)))
4304 				return -EFAULT;
4305 		}
4306 		return 0;
4307 	}
4308 #ifdef CONFIG_MMU
4309 	case TCP_ZEROCOPY_RECEIVE: {
4310 		struct scm_timestamping_internal tss;
4311 		struct tcp_zerocopy_receive zc = {};
4312 		int err;
4313 
4314 		if (copy_from_sockptr(&len, optlen, sizeof(int)))
4315 			return -EFAULT;
4316 		if (len < 0 ||
4317 		    len < offsetofend(struct tcp_zerocopy_receive, length))
4318 			return -EINVAL;
4319 		if (unlikely(len > sizeof(zc))) {
4320 			err = check_zeroed_sockptr(optval, sizeof(zc),
4321 						   len - sizeof(zc));
4322 			if (err < 1)
4323 				return err == 0 ? -EINVAL : err;
4324 			len = sizeof(zc);
4325 			if (copy_to_sockptr(optlen, &len, sizeof(int)))
4326 				return -EFAULT;
4327 		}
4328 		if (copy_from_sockptr(&zc, optval, len))
4329 			return -EFAULT;
4330 		if (zc.reserved)
4331 			return -EINVAL;
4332 		if (zc.msg_flags &  ~(TCP_VALID_ZC_MSG_FLAGS))
4333 			return -EINVAL;
4334 		sockopt_lock_sock(sk);
4335 		err = tcp_zerocopy_receive(sk, &zc, &tss);
4336 		err = BPF_CGROUP_RUN_PROG_GETSOCKOPT_KERN(sk, level, optname,
4337 							  &zc, &len, err);
4338 		sockopt_release_sock(sk);
4339 		if (len >= offsetofend(struct tcp_zerocopy_receive, msg_flags))
4340 			goto zerocopy_rcv_cmsg;
4341 		switch (len) {
4342 		case offsetofend(struct tcp_zerocopy_receive, msg_flags):
4343 			goto zerocopy_rcv_cmsg;
4344 		case offsetofend(struct tcp_zerocopy_receive, msg_controllen):
4345 		case offsetofend(struct tcp_zerocopy_receive, msg_control):
4346 		case offsetofend(struct tcp_zerocopy_receive, flags):
4347 		case offsetofend(struct tcp_zerocopy_receive, copybuf_len):
4348 		case offsetofend(struct tcp_zerocopy_receive, copybuf_address):
4349 		case offsetofend(struct tcp_zerocopy_receive, err):
4350 			goto zerocopy_rcv_sk_err;
4351 		case offsetofend(struct tcp_zerocopy_receive, inq):
4352 			goto zerocopy_rcv_inq;
4353 		case offsetofend(struct tcp_zerocopy_receive, length):
4354 		default:
4355 			goto zerocopy_rcv_out;
4356 		}
4357 zerocopy_rcv_cmsg:
4358 		if (zc.msg_flags & TCP_CMSG_TS)
4359 			tcp_zc_finalize_rx_tstamp(sk, &zc, &tss);
4360 		else
4361 			zc.msg_flags = 0;
4362 zerocopy_rcv_sk_err:
4363 		if (!err)
4364 			zc.err = sock_error(sk);
4365 zerocopy_rcv_inq:
4366 		zc.inq = tcp_inq_hint(sk);
4367 zerocopy_rcv_out:
4368 		if (!err && copy_to_sockptr(optval, &zc, len))
4369 			err = -EFAULT;
4370 		return err;
4371 	}
4372 #endif
4373 	default:
4374 		return -ENOPROTOOPT;
4375 	}
4376 
4377 	if (copy_to_sockptr(optlen, &len, sizeof(int)))
4378 		return -EFAULT;
4379 	if (copy_to_sockptr(optval, &val, len))
4380 		return -EFAULT;
4381 	return 0;
4382 }
4383 
tcp_bpf_bypass_getsockopt(int level,int optname)4384 bool tcp_bpf_bypass_getsockopt(int level, int optname)
4385 {
4386 	/* TCP do_tcp_getsockopt has optimized getsockopt implementation
4387 	 * to avoid extra socket lock for TCP_ZEROCOPY_RECEIVE.
4388 	 */
4389 	if (level == SOL_TCP && optname == TCP_ZEROCOPY_RECEIVE)
4390 		return true;
4391 
4392 	return false;
4393 }
4394 EXPORT_SYMBOL(tcp_bpf_bypass_getsockopt);
4395 
tcp_getsockopt(struct sock * sk,int level,int optname,char __user * optval,int __user * optlen)4396 int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval,
4397 		   int __user *optlen)
4398 {
4399 	struct inet_connection_sock *icsk = inet_csk(sk);
4400 
4401 	if (level != SOL_TCP)
4402 		/* Paired with WRITE_ONCE() in do_ipv6_setsockopt() and tcp_v6_connect() */
4403 		return READ_ONCE(icsk->icsk_af_ops)->getsockopt(sk, level, optname,
4404 								optval, optlen);
4405 	return do_tcp_getsockopt(sk, level, optname, USER_SOCKPTR(optval),
4406 				 USER_SOCKPTR(optlen));
4407 }
4408 EXPORT_SYMBOL(tcp_getsockopt);
4409 
4410 #ifdef CONFIG_TCP_MD5SIG
4411 static DEFINE_PER_CPU(struct tcp_md5sig_pool, tcp_md5sig_pool);
4412 static DEFINE_MUTEX(tcp_md5sig_mutex);
4413 static bool tcp_md5sig_pool_populated = false;
4414 
__tcp_alloc_md5sig_pool(void)4415 static void __tcp_alloc_md5sig_pool(void)
4416 {
4417 	struct crypto_ahash *hash;
4418 	int cpu;
4419 
4420 	hash = crypto_alloc_ahash("md5", 0, CRYPTO_ALG_ASYNC);
4421 	if (IS_ERR(hash))
4422 		return;
4423 
4424 	for_each_possible_cpu(cpu) {
4425 		void *scratch = per_cpu(tcp_md5sig_pool, cpu).scratch;
4426 		struct ahash_request *req;
4427 
4428 		if (!scratch) {
4429 			scratch = kmalloc_node(sizeof(union tcp_md5sum_block) +
4430 					       sizeof(struct tcphdr),
4431 					       GFP_KERNEL,
4432 					       cpu_to_node(cpu));
4433 			if (!scratch)
4434 				return;
4435 			per_cpu(tcp_md5sig_pool, cpu).scratch = scratch;
4436 		}
4437 		if (per_cpu(tcp_md5sig_pool, cpu).md5_req)
4438 			continue;
4439 
4440 		req = ahash_request_alloc(hash, GFP_KERNEL);
4441 		if (!req)
4442 			return;
4443 
4444 		ahash_request_set_callback(req, 0, NULL, NULL);
4445 
4446 		per_cpu(tcp_md5sig_pool, cpu).md5_req = req;
4447 	}
4448 	/* before setting tcp_md5sig_pool_populated, we must commit all writes
4449 	 * to memory. See smp_rmb() in tcp_get_md5sig_pool()
4450 	 */
4451 	smp_wmb();
4452 	/* Paired with READ_ONCE() from tcp_alloc_md5sig_pool()
4453 	 * and tcp_get_md5sig_pool().
4454 	*/
4455 	WRITE_ONCE(tcp_md5sig_pool_populated, true);
4456 }
4457 
tcp_alloc_md5sig_pool(void)4458 bool tcp_alloc_md5sig_pool(void)
4459 {
4460 	/* Paired with WRITE_ONCE() from __tcp_alloc_md5sig_pool() */
4461 	if (unlikely(!READ_ONCE(tcp_md5sig_pool_populated))) {
4462 		mutex_lock(&tcp_md5sig_mutex);
4463 
4464 		if (!tcp_md5sig_pool_populated) {
4465 			__tcp_alloc_md5sig_pool();
4466 			if (tcp_md5sig_pool_populated)
4467 				static_branch_inc(&tcp_md5_needed);
4468 		}
4469 
4470 		mutex_unlock(&tcp_md5sig_mutex);
4471 	}
4472 	/* Paired with WRITE_ONCE() from __tcp_alloc_md5sig_pool() */
4473 	return READ_ONCE(tcp_md5sig_pool_populated);
4474 }
4475 EXPORT_SYMBOL(tcp_alloc_md5sig_pool);
4476 
4477 
4478 /**
4479  *	tcp_get_md5sig_pool - get md5sig_pool for this user
4480  *
4481  *	We use percpu structure, so if we succeed, we exit with preemption
4482  *	and BH disabled, to make sure another thread or softirq handling
4483  *	wont try to get same context.
4484  */
tcp_get_md5sig_pool(void)4485 struct tcp_md5sig_pool *tcp_get_md5sig_pool(void)
4486 {
4487 	local_bh_disable();
4488 
4489 	/* Paired with WRITE_ONCE() from __tcp_alloc_md5sig_pool() */
4490 	if (READ_ONCE(tcp_md5sig_pool_populated)) {
4491 		/* coupled with smp_wmb() in __tcp_alloc_md5sig_pool() */
4492 		smp_rmb();
4493 		return this_cpu_ptr(&tcp_md5sig_pool);
4494 	}
4495 	local_bh_enable();
4496 	return NULL;
4497 }
4498 EXPORT_SYMBOL(tcp_get_md5sig_pool);
4499 
tcp_md5_hash_skb_data(struct tcp_md5sig_pool * hp,const struct sk_buff * skb,unsigned int header_len)4500 int tcp_md5_hash_skb_data(struct tcp_md5sig_pool *hp,
4501 			  const struct sk_buff *skb, unsigned int header_len)
4502 {
4503 	struct scatterlist sg;
4504 	const struct tcphdr *tp = tcp_hdr(skb);
4505 	struct ahash_request *req = hp->md5_req;
4506 	unsigned int i;
4507 	const unsigned int head_data_len = skb_headlen(skb) > header_len ?
4508 					   skb_headlen(skb) - header_len : 0;
4509 	const struct skb_shared_info *shi = skb_shinfo(skb);
4510 	struct sk_buff *frag_iter;
4511 
4512 	sg_init_table(&sg, 1);
4513 
4514 	sg_set_buf(&sg, ((u8 *) tp) + header_len, head_data_len);
4515 	ahash_request_set_crypt(req, &sg, NULL, head_data_len);
4516 	if (crypto_ahash_update(req))
4517 		return 1;
4518 
4519 	for (i = 0; i < shi->nr_frags; ++i) {
4520 		const skb_frag_t *f = &shi->frags[i];
4521 		unsigned int offset = skb_frag_off(f);
4522 		struct page *page = skb_frag_page(f) + (offset >> PAGE_SHIFT);
4523 
4524 		sg_set_page(&sg, page, skb_frag_size(f),
4525 			    offset_in_page(offset));
4526 		ahash_request_set_crypt(req, &sg, NULL, skb_frag_size(f));
4527 		if (crypto_ahash_update(req))
4528 			return 1;
4529 	}
4530 
4531 	skb_walk_frags(skb, frag_iter)
4532 		if (tcp_md5_hash_skb_data(hp, frag_iter, 0))
4533 			return 1;
4534 
4535 	return 0;
4536 }
4537 EXPORT_SYMBOL(tcp_md5_hash_skb_data);
4538 
tcp_md5_hash_key(struct tcp_md5sig_pool * hp,const struct tcp_md5sig_key * key)4539 int tcp_md5_hash_key(struct tcp_md5sig_pool *hp, const struct tcp_md5sig_key *key)
4540 {
4541 	u8 keylen = READ_ONCE(key->keylen); /* paired with WRITE_ONCE() in tcp_md5_do_add */
4542 	struct scatterlist sg;
4543 
4544 	sg_init_one(&sg, key->key, keylen);
4545 	ahash_request_set_crypt(hp->md5_req, &sg, NULL, keylen);
4546 
4547 	/* We use data_race() because tcp_md5_do_add() might change key->key under us */
4548 	return data_race(crypto_ahash_update(hp->md5_req));
4549 }
4550 EXPORT_SYMBOL(tcp_md5_hash_key);
4551 
4552 /* Called with rcu_read_lock() */
4553 enum skb_drop_reason
tcp_inbound_md5_hash(const struct sock * sk,const struct sk_buff * skb,const void * saddr,const void * daddr,int family,int dif,int sdif)4554 tcp_inbound_md5_hash(const struct sock *sk, const struct sk_buff *skb,
4555 		     const void *saddr, const void *daddr,
4556 		     int family, int dif, int sdif)
4557 {
4558 	/*
4559 	 * This gets called for each TCP segment that arrives
4560 	 * so we want to be efficient.
4561 	 * We have 3 drop cases:
4562 	 * o No MD5 hash and one expected.
4563 	 * o MD5 hash and we're not expecting one.
4564 	 * o MD5 hash and its wrong.
4565 	 */
4566 	const __u8 *hash_location = NULL;
4567 	struct tcp_md5sig_key *hash_expected;
4568 	const struct tcphdr *th = tcp_hdr(skb);
4569 	struct tcp_sock *tp = tcp_sk(sk);
4570 	int genhash, l3index;
4571 	u8 newhash[16];
4572 
4573 	/* sdif set, means packet ingressed via a device
4574 	 * in an L3 domain and dif is set to the l3mdev
4575 	 */
4576 	l3index = sdif ? dif : 0;
4577 
4578 	hash_expected = tcp_md5_do_lookup(sk, l3index, saddr, family);
4579 	hash_location = tcp_parse_md5sig_option(th);
4580 
4581 	/* We've parsed the options - do we have a hash? */
4582 	if (!hash_expected && !hash_location)
4583 		return SKB_NOT_DROPPED_YET;
4584 
4585 	if (hash_expected && !hash_location) {
4586 		NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMD5NOTFOUND);
4587 		return SKB_DROP_REASON_TCP_MD5NOTFOUND;
4588 	}
4589 
4590 	if (!hash_expected && hash_location) {
4591 		NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMD5UNEXPECTED);
4592 		return SKB_DROP_REASON_TCP_MD5UNEXPECTED;
4593 	}
4594 
4595 	/* Check the signature.
4596 	 * To support dual stack listeners, we need to handle
4597 	 * IPv4-mapped case.
4598 	 */
4599 	if (family == AF_INET)
4600 		genhash = tcp_v4_md5_hash_skb(newhash,
4601 					      hash_expected,
4602 					      NULL, skb);
4603 	else
4604 		genhash = tp->af_specific->calc_md5_hash(newhash,
4605 							 hash_expected,
4606 							 NULL, skb);
4607 
4608 	if (genhash || memcmp(hash_location, newhash, 16) != 0) {
4609 		NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMD5FAILURE);
4610 		if (family == AF_INET) {
4611 			net_info_ratelimited("MD5 Hash failed for (%pI4, %d)->(%pI4, %d)%s L3 index %d\n",
4612 					saddr, ntohs(th->source),
4613 					daddr, ntohs(th->dest),
4614 					genhash ? " tcp_v4_calc_md5_hash failed"
4615 					: "", l3index);
4616 		} else {
4617 			net_info_ratelimited("MD5 Hash %s for [%pI6c]:%u->[%pI6c]:%u L3 index %d\n",
4618 					genhash ? "failed" : "mismatch",
4619 					saddr, ntohs(th->source),
4620 					daddr, ntohs(th->dest), l3index);
4621 		}
4622 		return SKB_DROP_REASON_TCP_MD5FAILURE;
4623 	}
4624 	return SKB_NOT_DROPPED_YET;
4625 }
4626 EXPORT_SYMBOL(tcp_inbound_md5_hash);
4627 
4628 #endif
4629 
tcp_done(struct sock * sk)4630 void tcp_done(struct sock *sk)
4631 {
4632 	struct request_sock *req;
4633 
4634 	/* We might be called with a new socket, after
4635 	 * inet_csk_prepare_forced_close() has been called
4636 	 * so we can not use lockdep_sock_is_held(sk)
4637 	 */
4638 	req = rcu_dereference_protected(tcp_sk(sk)->fastopen_rsk, 1);
4639 
4640 	if (sk->sk_state == TCP_SYN_SENT || sk->sk_state == TCP_SYN_RECV)
4641 		TCP_INC_STATS(sock_net(sk), TCP_MIB_ATTEMPTFAILS);
4642 
4643 	tcp_set_state(sk, TCP_CLOSE);
4644 	tcp_clear_xmit_timers(sk);
4645 	if (req)
4646 		reqsk_fastopen_remove(sk, req, false);
4647 
4648 	sk->sk_shutdown = SHUTDOWN_MASK;
4649 
4650 	if (!sock_flag(sk, SOCK_DEAD))
4651 		sk->sk_state_change(sk);
4652 	else
4653 		inet_csk_destroy_sock(sk);
4654 }
4655 EXPORT_SYMBOL_GPL(tcp_done);
4656 
tcp_abort(struct sock * sk,int err)4657 int tcp_abort(struct sock *sk, int err)
4658 {
4659 	int state = inet_sk_state_load(sk);
4660 
4661 	if (state == TCP_NEW_SYN_RECV) {
4662 		struct request_sock *req = inet_reqsk(sk);
4663 
4664 		local_bh_disable();
4665 		inet_csk_reqsk_queue_drop(req->rsk_listener, req);
4666 		local_bh_enable();
4667 		return 0;
4668 	}
4669 	if (state == TCP_TIME_WAIT) {
4670 		struct inet_timewait_sock *tw = inet_twsk(sk);
4671 
4672 		refcount_inc(&tw->tw_refcnt);
4673 		local_bh_disable();
4674 		inet_twsk_deschedule_put(tw);
4675 		local_bh_enable();
4676 		return 0;
4677 	}
4678 
4679 	/* Don't race with userspace socket closes such as tcp_close. */
4680 	lock_sock(sk);
4681 
4682 	if (sk->sk_state == TCP_LISTEN) {
4683 		tcp_set_state(sk, TCP_CLOSE);
4684 		inet_csk_listen_stop(sk);
4685 	}
4686 
4687 	/* Don't race with BH socket closes such as inet_csk_listen_stop. */
4688 	local_bh_disable();
4689 	bh_lock_sock(sk);
4690 
4691 	if (!sock_flag(sk, SOCK_DEAD)) {
4692 		sk->sk_err = err;
4693 		/* This barrier is coupled with smp_rmb() in tcp_poll() */
4694 		smp_wmb();
4695 		sk_error_report(sk);
4696 		if (tcp_need_reset(sk->sk_state))
4697 			tcp_send_active_reset(sk, GFP_ATOMIC);
4698 		tcp_done(sk);
4699 	}
4700 
4701 	bh_unlock_sock(sk);
4702 	local_bh_enable();
4703 	tcp_write_queue_purge(sk);
4704 	release_sock(sk);
4705 	return 0;
4706 }
4707 EXPORT_SYMBOL_GPL(tcp_abort);
4708 
4709 extern struct tcp_congestion_ops tcp_reno;
4710 
4711 static __initdata unsigned long thash_entries;
set_thash_entries(char * str)4712 static int __init set_thash_entries(char *str)
4713 {
4714 	ssize_t ret;
4715 
4716 	if (!str)
4717 		return 0;
4718 
4719 	ret = kstrtoul(str, 0, &thash_entries);
4720 	if (ret)
4721 		return 0;
4722 
4723 	return 1;
4724 }
4725 __setup("thash_entries=", set_thash_entries);
4726 
tcp_init_mem(void)4727 static void __init tcp_init_mem(void)
4728 {
4729 	unsigned long limit = nr_free_buffer_pages() / 16;
4730 
4731 	limit = max(limit, 128UL);
4732 	sysctl_tcp_mem[0] = limit / 4 * 3;		/* 4.68 % */
4733 	sysctl_tcp_mem[1] = limit;			/* 6.25 % */
4734 	sysctl_tcp_mem[2] = sysctl_tcp_mem[0] * 2;	/* 9.37 % */
4735 }
4736 
tcp_init(void)4737 void __init tcp_init(void)
4738 {
4739 	int max_rshare, max_wshare, cnt;
4740 	unsigned long limit;
4741 	unsigned int i;
4742 
4743 	BUILD_BUG_ON(TCP_MIN_SND_MSS <= MAX_TCP_OPTION_SPACE);
4744 	BUILD_BUG_ON(sizeof(struct tcp_skb_cb) >
4745 		     sizeof_field(struct sk_buff, cb));
4746 
4747 	percpu_counter_init(&tcp_sockets_allocated, 0, GFP_KERNEL);
4748 
4749 	timer_setup(&tcp_orphan_timer, tcp_orphan_update, TIMER_DEFERRABLE);
4750 	mod_timer(&tcp_orphan_timer, jiffies + TCP_ORPHAN_TIMER_PERIOD);
4751 
4752 	inet_hashinfo2_init(&tcp_hashinfo, "tcp_listen_portaddr_hash",
4753 			    thash_entries, 21,  /* one slot per 2 MB*/
4754 			    0, 64 * 1024);
4755 	tcp_hashinfo.bind_bucket_cachep =
4756 		kmem_cache_create("tcp_bind_bucket",
4757 				  sizeof(struct inet_bind_bucket), 0,
4758 				  SLAB_HWCACHE_ALIGN | SLAB_PANIC |
4759 				  SLAB_ACCOUNT,
4760 				  NULL);
4761 	tcp_hashinfo.bind2_bucket_cachep =
4762 		kmem_cache_create("tcp_bind2_bucket",
4763 				  sizeof(struct inet_bind2_bucket), 0,
4764 				  SLAB_HWCACHE_ALIGN | SLAB_PANIC |
4765 				  SLAB_ACCOUNT,
4766 				  NULL);
4767 
4768 	/* Size and allocate the main established and bind bucket
4769 	 * hash tables.
4770 	 *
4771 	 * The methodology is similar to that of the buffer cache.
4772 	 */
4773 	tcp_hashinfo.ehash =
4774 		alloc_large_system_hash("TCP established",
4775 					sizeof(struct inet_ehash_bucket),
4776 					thash_entries,
4777 					17, /* one slot per 128 KB of memory */
4778 					0,
4779 					NULL,
4780 					&tcp_hashinfo.ehash_mask,
4781 					0,
4782 					thash_entries ? 0 : 512 * 1024);
4783 	for (i = 0; i <= tcp_hashinfo.ehash_mask; i++)
4784 		INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].chain, i);
4785 
4786 	if (inet_ehash_locks_alloc(&tcp_hashinfo))
4787 		panic("TCP: failed to alloc ehash_locks");
4788 	tcp_hashinfo.bhash =
4789 		alloc_large_system_hash("TCP bind",
4790 					2 * sizeof(struct inet_bind_hashbucket),
4791 					tcp_hashinfo.ehash_mask + 1,
4792 					17, /* one slot per 128 KB of memory */
4793 					0,
4794 					&tcp_hashinfo.bhash_size,
4795 					NULL,
4796 					0,
4797 					64 * 1024);
4798 	tcp_hashinfo.bhash_size = 1U << tcp_hashinfo.bhash_size;
4799 	tcp_hashinfo.bhash2 = tcp_hashinfo.bhash + tcp_hashinfo.bhash_size;
4800 	for (i = 0; i < tcp_hashinfo.bhash_size; i++) {
4801 		spin_lock_init(&tcp_hashinfo.bhash[i].lock);
4802 		INIT_HLIST_HEAD(&tcp_hashinfo.bhash[i].chain);
4803 		spin_lock_init(&tcp_hashinfo.bhash2[i].lock);
4804 		INIT_HLIST_HEAD(&tcp_hashinfo.bhash2[i].chain);
4805 	}
4806 
4807 	tcp_hashinfo.pernet = false;
4808 
4809 	cnt = tcp_hashinfo.ehash_mask + 1;
4810 	sysctl_tcp_max_orphans = cnt / 2;
4811 
4812 	tcp_init_mem();
4813 	/* Set per-socket limits to no more than 1/128 the pressure threshold */
4814 	limit = nr_free_buffer_pages() << (PAGE_SHIFT - 7);
4815 	max_wshare = min(4UL*1024*1024, limit);
4816 	max_rshare = min(6UL*1024*1024, limit);
4817 
4818 	init_net.ipv4.sysctl_tcp_wmem[0] = PAGE_SIZE;
4819 	init_net.ipv4.sysctl_tcp_wmem[1] = 16*1024;
4820 	init_net.ipv4.sysctl_tcp_wmem[2] = max(64*1024, max_wshare);
4821 
4822 	init_net.ipv4.sysctl_tcp_rmem[0] = PAGE_SIZE;
4823 	init_net.ipv4.sysctl_tcp_rmem[1] = 131072;
4824 	init_net.ipv4.sysctl_tcp_rmem[2] = max(131072, max_rshare);
4825 
4826 	pr_info("Hash tables configured (established %u bind %u)\n",
4827 		tcp_hashinfo.ehash_mask + 1, tcp_hashinfo.bhash_size);
4828 
4829 	tcp_v4_init();
4830 	tcp_metrics_init();
4831 	BUG_ON(tcp_register_congestion_control(&tcp_reno) != 0);
4832 	tcp_tasklet_init();
4833 	mptcp_init();
4834 }
4835