xref: /linux/net/ipv4/tcp.c (revision 1e15510b71c99c6e49134d756df91069f7d18141)
1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /*
3  * INET		An implementation of the TCP/IP protocol suite for the LINUX
4  *		operating system.  INET is implemented using the  BSD Socket
5  *		interface as the means of communication with the user level.
6  *
7  *		Implementation of the Transmission Control Protocol(TCP).
8  *
9  * Authors:	Ross Biro
10  *		Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
11  *		Mark Evans, <evansmp@uhura.aston.ac.uk>
12  *		Corey Minyard <wf-rch!minyard@relay.EU.net>
13  *		Florian La Roche, <flla@stud.uni-sb.de>
14  *		Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
15  *		Linus Torvalds, <torvalds@cs.helsinki.fi>
16  *		Alan Cox, <gw4pts@gw4pts.ampr.org>
17  *		Matthew Dillon, <dillon@apollo.west.oic.com>
18  *		Arnt Gulbrandsen, <agulbra@nvg.unit.no>
19  *		Jorge Cwik, <jorge@laser.satlink.net>
20  *
21  * Fixes:
22  *		Alan Cox	:	Numerous verify_area() calls
23  *		Alan Cox	:	Set the ACK bit on a reset
24  *		Alan Cox	:	Stopped it crashing if it closed while
25  *					sk->inuse=1 and was trying to connect
26  *					(tcp_err()).
27  *		Alan Cox	:	All icmp error handling was broken
28  *					pointers passed where wrong and the
29  *					socket was looked up backwards. Nobody
30  *					tested any icmp error code obviously.
31  *		Alan Cox	:	tcp_err() now handled properly. It
32  *					wakes people on errors. poll
33  *					behaves and the icmp error race
34  *					has gone by moving it into sock.c
35  *		Alan Cox	:	tcp_send_reset() fixed to work for
36  *					everything not just packets for
37  *					unknown sockets.
38  *		Alan Cox	:	tcp option processing.
39  *		Alan Cox	:	Reset tweaked (still not 100%) [Had
40  *					syn rule wrong]
41  *		Herp Rosmanith  :	More reset fixes
42  *		Alan Cox	:	No longer acks invalid rst frames.
43  *					Acking any kind of RST is right out.
44  *		Alan Cox	:	Sets an ignore me flag on an rst
45  *					receive otherwise odd bits of prattle
46  *					escape still
47  *		Alan Cox	:	Fixed another acking RST frame bug.
48  *					Should stop LAN workplace lockups.
49  *		Alan Cox	: 	Some tidyups using the new skb list
50  *					facilities
51  *		Alan Cox	:	sk->keepopen now seems to work
52  *		Alan Cox	:	Pulls options out correctly on accepts
53  *		Alan Cox	:	Fixed assorted sk->rqueue->next errors
54  *		Alan Cox	:	PSH doesn't end a TCP read. Switched a
55  *					bit to skb ops.
56  *		Alan Cox	:	Tidied tcp_data to avoid a potential
57  *					nasty.
58  *		Alan Cox	:	Added some better commenting, as the
59  *					tcp is hard to follow
60  *		Alan Cox	:	Removed incorrect check for 20 * psh
61  *	Michael O'Reilly	:	ack < copied bug fix.
62  *	Johannes Stille		:	Misc tcp fixes (not all in yet).
63  *		Alan Cox	:	FIN with no memory -> CRASH
64  *		Alan Cox	:	Added socket option proto entries.
65  *					Also added awareness of them to accept.
66  *		Alan Cox	:	Added TCP options (SOL_TCP)
67  *		Alan Cox	:	Switched wakeup calls to callbacks,
68  *					so the kernel can layer network
69  *					sockets.
70  *		Alan Cox	:	Use ip_tos/ip_ttl settings.
71  *		Alan Cox	:	Handle FIN (more) properly (we hope).
72  *		Alan Cox	:	RST frames sent on unsynchronised
73  *					state ack error.
74  *		Alan Cox	:	Put in missing check for SYN bit.
75  *		Alan Cox	:	Added tcp_select_window() aka NET2E
76  *					window non shrink trick.
77  *		Alan Cox	:	Added a couple of small NET2E timer
78  *					fixes
79  *		Charles Hedrick :	TCP fixes
80  *		Toomas Tamm	:	TCP window fixes
81  *		Alan Cox	:	Small URG fix to rlogin ^C ack fight
82  *		Charles Hedrick	:	Rewrote most of it to actually work
83  *		Linus		:	Rewrote tcp_read() and URG handling
84  *					completely
85  *		Gerhard Koerting:	Fixed some missing timer handling
86  *		Matthew Dillon  :	Reworked TCP machine states as per RFC
87  *		Gerhard Koerting:	PC/TCP workarounds
88  *		Adam Caldwell	:	Assorted timer/timing errors
89  *		Matthew Dillon	:	Fixed another RST bug
90  *		Alan Cox	:	Move to kernel side addressing changes.
91  *		Alan Cox	:	Beginning work on TCP fastpathing
92  *					(not yet usable)
93  *		Arnt Gulbrandsen:	Turbocharged tcp_check() routine.
94  *		Alan Cox	:	TCP fast path debugging
95  *		Alan Cox	:	Window clamping
96  *		Michael Riepe	:	Bug in tcp_check()
97  *		Matt Dillon	:	More TCP improvements and RST bug fixes
98  *		Matt Dillon	:	Yet more small nasties remove from the
99  *					TCP code (Be very nice to this man if
100  *					tcp finally works 100%) 8)
101  *		Alan Cox	:	BSD accept semantics.
102  *		Alan Cox	:	Reset on closedown bug.
103  *	Peter De Schrijver	:	ENOTCONN check missing in tcp_sendto().
104  *		Michael Pall	:	Handle poll() after URG properly in
105  *					all cases.
106  *		Michael Pall	:	Undo the last fix in tcp_read_urg()
107  *					(multi URG PUSH broke rlogin).
108  *		Michael Pall	:	Fix the multi URG PUSH problem in
109  *					tcp_readable(), poll() after URG
110  *					works now.
111  *		Michael Pall	:	recv(...,MSG_OOB) never blocks in the
112  *					BSD api.
113  *		Alan Cox	:	Changed the semantics of sk->socket to
114  *					fix a race and a signal problem with
115  *					accept() and async I/O.
116  *		Alan Cox	:	Relaxed the rules on tcp_sendto().
117  *		Yury Shevchuk	:	Really fixed accept() blocking problem.
118  *		Craig I. Hagan  :	Allow for BSD compatible TIME_WAIT for
119  *					clients/servers which listen in on
120  *					fixed ports.
121  *		Alan Cox	:	Cleaned the above up and shrank it to
122  *					a sensible code size.
123  *		Alan Cox	:	Self connect lockup fix.
124  *		Alan Cox	:	No connect to multicast.
125  *		Ross Biro	:	Close unaccepted children on master
126  *					socket close.
127  *		Alan Cox	:	Reset tracing code.
128  *		Alan Cox	:	Spurious resets on shutdown.
129  *		Alan Cox	:	Giant 15 minute/60 second timer error
130  *		Alan Cox	:	Small whoops in polling before an
131  *					accept.
132  *		Alan Cox	:	Kept the state trace facility since
133  *					it's handy for debugging.
134  *		Alan Cox	:	More reset handler fixes.
135  *		Alan Cox	:	Started rewriting the code based on
136  *					the RFC's for other useful protocol
137  *					references see: Comer, KA9Q NOS, and
138  *					for a reference on the difference
139  *					between specifications and how BSD
140  *					works see the 4.4lite source.
141  *		A.N.Kuznetsov	:	Don't time wait on completion of tidy
142  *					close.
143  *		Linus Torvalds	:	Fin/Shutdown & copied_seq changes.
144  *		Linus Torvalds	:	Fixed BSD port reuse to work first syn
145  *		Alan Cox	:	Reimplemented timers as per the RFC
146  *					and using multiple timers for sanity.
147  *		Alan Cox	:	Small bug fixes, and a lot of new
148  *					comments.
149  *		Alan Cox	:	Fixed dual reader crash by locking
150  *					the buffers (much like datagram.c)
151  *		Alan Cox	:	Fixed stuck sockets in probe. A probe
152  *					now gets fed up of retrying without
153  *					(even a no space) answer.
154  *		Alan Cox	:	Extracted closing code better
155  *		Alan Cox	:	Fixed the closing state machine to
156  *					resemble the RFC.
157  *		Alan Cox	:	More 'per spec' fixes.
158  *		Jorge Cwik	:	Even faster checksumming.
159  *		Alan Cox	:	tcp_data() doesn't ack illegal PSH
160  *					only frames. At least one pc tcp stack
161  *					generates them.
162  *		Alan Cox	:	Cache last socket.
163  *		Alan Cox	:	Per route irtt.
164  *		Matt Day	:	poll()->select() match BSD precisely on error
165  *		Alan Cox	:	New buffers
166  *		Marc Tamsky	:	Various sk->prot->retransmits and
167  *					sk->retransmits misupdating fixed.
168  *					Fixed tcp_write_timeout: stuck close,
169  *					and TCP syn retries gets used now.
170  *		Mark Yarvis	:	In tcp_read_wakeup(), don't send an
171  *					ack if state is TCP_CLOSED.
172  *		Alan Cox	:	Look up device on a retransmit - routes may
173  *					change. Doesn't yet cope with MSS shrink right
174  *					but it's a start!
175  *		Marc Tamsky	:	Closing in closing fixes.
176  *		Mike Shaver	:	RFC1122 verifications.
177  *		Alan Cox	:	rcv_saddr errors.
178  *		Alan Cox	:	Block double connect().
179  *		Alan Cox	:	Small hooks for enSKIP.
180  *		Alexey Kuznetsov:	Path MTU discovery.
181  *		Alan Cox	:	Support soft errors.
182  *		Alan Cox	:	Fix MTU discovery pathological case
183  *					when the remote claims no mtu!
184  *		Marc Tamsky	:	TCP_CLOSE fix.
185  *		Colin (G3TNE)	:	Send a reset on syn ack replies in
186  *					window but wrong (fixes NT lpd problems)
187  *		Pedro Roque	:	Better TCP window handling, delayed ack.
188  *		Joerg Reuter	:	No modification of locked buffers in
189  *					tcp_do_retransmit()
190  *		Eric Schenk	:	Changed receiver side silly window
191  *					avoidance algorithm to BSD style
192  *					algorithm. This doubles throughput
193  *					against machines running Solaris,
194  *					and seems to result in general
195  *					improvement.
196  *	Stefan Magdalinski	:	adjusted tcp_readable() to fix FIONREAD
197  *	Willy Konynenberg	:	Transparent proxying support.
198  *	Mike McLagan		:	Routing by source
199  *		Keith Owens	:	Do proper merging with partial SKB's in
200  *					tcp_do_sendmsg to avoid burstiness.
201  *		Eric Schenk	:	Fix fast close down bug with
202  *					shutdown() followed by close().
203  *		Andi Kleen 	:	Make poll agree with SIGIO
204  *	Salvatore Sanfilippo	:	Support SO_LINGER with linger == 1 and
205  *					lingertime == 0 (RFC 793 ABORT Call)
206  *	Hirokazu Takahashi	:	Use copy_from_user() instead of
207  *					csum_and_copy_from_user() if possible.
208  *
209  * Description of States:
210  *
211  *	TCP_SYN_SENT		sent a connection request, waiting for ack
212  *
213  *	TCP_SYN_RECV		received a connection request, sent ack,
214  *				waiting for final ack in three-way handshake.
215  *
216  *	TCP_ESTABLISHED		connection established
217  *
218  *	TCP_FIN_WAIT1		our side has shutdown, waiting to complete
219  *				transmission of remaining buffered data
220  *
221  *	TCP_FIN_WAIT2		all buffered data sent, waiting for remote
222  *				to shutdown
223  *
224  *	TCP_CLOSING		both sides have shutdown but we still have
225  *				data we have to finish sending
226  *
227  *	TCP_TIME_WAIT		timeout to catch resent junk before entering
228  *				closed, can only be entered from FIN_WAIT2
229  *				or CLOSING.  Required because the other end
230  *				may not have gotten our last ACK causing it
231  *				to retransmit the data packet (which we ignore)
232  *
233  *	TCP_CLOSE_WAIT		remote side has shutdown and is waiting for
234  *				us to finish writing our data and to shutdown
235  *				(we have to close() to move on to LAST_ACK)
236  *
237  *	TCP_LAST_ACK		out side has shutdown after remote has
238  *				shutdown.  There may still be data in our
239  *				buffer that we have to finish sending
240  *
241  *	TCP_CLOSE		socket is finished
242  */
243 
244 #define pr_fmt(fmt) "TCP: " fmt
245 
246 #include <crypto/hash.h>
247 #include <linux/kernel.h>
248 #include <linux/module.h>
249 #include <linux/types.h>
250 #include <linux/fcntl.h>
251 #include <linux/poll.h>
252 #include <linux/inet_diag.h>
253 #include <linux/init.h>
254 #include <linux/fs.h>
255 #include <linux/skbuff.h>
256 #include <linux/scatterlist.h>
257 #include <linux/splice.h>
258 #include <linux/net.h>
259 #include <linux/socket.h>
260 #include <linux/random.h>
261 #include <linux/memblock.h>
262 #include <linux/highmem.h>
263 #include <linux/cache.h>
264 #include <linux/err.h>
265 #include <linux/time.h>
266 #include <linux/slab.h>
267 #include <linux/errqueue.h>
268 #include <linux/static_key.h>
269 #include <linux/btf.h>
270 
271 #include <net/icmp.h>
272 #include <net/inet_common.h>
273 #include <net/tcp.h>
274 #include <net/mptcp.h>
275 #include <net/proto_memory.h>
276 #include <net/xfrm.h>
277 #include <net/ip.h>
278 #include <net/sock.h>
279 #include <net/rstreason.h>
280 
281 #include <linux/uaccess.h>
282 #include <asm/ioctls.h>
283 #include <net/busy_poll.h>
284 #include <net/hotdata.h>
285 #include <trace/events/tcp.h>
286 #include <net/rps.h>
287 
288 #include "../core/devmem.h"
289 
290 /* Track pending CMSGs. */
291 enum {
292 	TCP_CMSG_INQ = 1,
293 	TCP_CMSG_TS = 2
294 };
295 
296 DEFINE_PER_CPU(unsigned int, tcp_orphan_count);
297 EXPORT_PER_CPU_SYMBOL_GPL(tcp_orphan_count);
298 
299 DEFINE_PER_CPU(u32, tcp_tw_isn);
300 EXPORT_PER_CPU_SYMBOL_GPL(tcp_tw_isn);
301 
302 long sysctl_tcp_mem[3] __read_mostly;
303 EXPORT_SYMBOL(sysctl_tcp_mem);
304 
305 atomic_long_t tcp_memory_allocated ____cacheline_aligned_in_smp;	/* Current allocated memory. */
306 EXPORT_SYMBOL(tcp_memory_allocated);
307 DEFINE_PER_CPU(int, tcp_memory_per_cpu_fw_alloc);
308 EXPORT_PER_CPU_SYMBOL_GPL(tcp_memory_per_cpu_fw_alloc);
309 
310 #if IS_ENABLED(CONFIG_SMC)
311 DEFINE_STATIC_KEY_FALSE(tcp_have_smc);
312 EXPORT_SYMBOL(tcp_have_smc);
313 #endif
314 
315 /*
316  * Current number of TCP sockets.
317  */
318 struct percpu_counter tcp_sockets_allocated ____cacheline_aligned_in_smp;
319 EXPORT_SYMBOL(tcp_sockets_allocated);
320 
321 /*
322  * TCP splice context
323  */
324 struct tcp_splice_state {
325 	struct pipe_inode_info *pipe;
326 	size_t len;
327 	unsigned int flags;
328 };
329 
330 /*
331  * Pressure flag: try to collapse.
332  * Technical note: it is used by multiple contexts non atomically.
333  * All the __sk_mem_schedule() is of this nature: accounting
334  * is strict, actions are advisory and have some latency.
335  */
336 unsigned long tcp_memory_pressure __read_mostly;
337 EXPORT_SYMBOL_GPL(tcp_memory_pressure);
338 
tcp_enter_memory_pressure(struct sock * sk)339 void tcp_enter_memory_pressure(struct sock *sk)
340 {
341 	unsigned long val;
342 
343 	if (READ_ONCE(tcp_memory_pressure))
344 		return;
345 	val = jiffies;
346 
347 	if (!val)
348 		val--;
349 	if (!cmpxchg(&tcp_memory_pressure, 0, val))
350 		NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMEMORYPRESSURES);
351 }
352 EXPORT_SYMBOL_GPL(tcp_enter_memory_pressure);
353 
tcp_leave_memory_pressure(struct sock * sk)354 void tcp_leave_memory_pressure(struct sock *sk)
355 {
356 	unsigned long val;
357 
358 	if (!READ_ONCE(tcp_memory_pressure))
359 		return;
360 	val = xchg(&tcp_memory_pressure, 0);
361 	if (val)
362 		NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPMEMORYPRESSURESCHRONO,
363 			      jiffies_to_msecs(jiffies - val));
364 }
365 EXPORT_SYMBOL_GPL(tcp_leave_memory_pressure);
366 
367 /* Convert seconds to retransmits based on initial and max timeout */
secs_to_retrans(int seconds,int timeout,int rto_max)368 static u8 secs_to_retrans(int seconds, int timeout, int rto_max)
369 {
370 	u8 res = 0;
371 
372 	if (seconds > 0) {
373 		int period = timeout;
374 
375 		res = 1;
376 		while (seconds > period && res < 255) {
377 			res++;
378 			timeout <<= 1;
379 			if (timeout > rto_max)
380 				timeout = rto_max;
381 			period += timeout;
382 		}
383 	}
384 	return res;
385 }
386 
387 /* Convert retransmits to seconds based on initial and max timeout */
retrans_to_secs(u8 retrans,int timeout,int rto_max)388 static int retrans_to_secs(u8 retrans, int timeout, int rto_max)
389 {
390 	int period = 0;
391 
392 	if (retrans > 0) {
393 		period = timeout;
394 		while (--retrans) {
395 			timeout <<= 1;
396 			if (timeout > rto_max)
397 				timeout = rto_max;
398 			period += timeout;
399 		}
400 	}
401 	return period;
402 }
403 
tcp_compute_delivery_rate(const struct tcp_sock * tp)404 static u64 tcp_compute_delivery_rate(const struct tcp_sock *tp)
405 {
406 	u32 rate = READ_ONCE(tp->rate_delivered);
407 	u32 intv = READ_ONCE(tp->rate_interval_us);
408 	u64 rate64 = 0;
409 
410 	if (rate && intv) {
411 		rate64 = (u64)rate * tp->mss_cache * USEC_PER_SEC;
412 		do_div(rate64, intv);
413 	}
414 	return rate64;
415 }
416 
417 /* Address-family independent initialization for a tcp_sock.
418  *
419  * NOTE: A lot of things set to zero explicitly by call to
420  *       sk_alloc() so need not be done here.
421  */
tcp_init_sock(struct sock * sk)422 void tcp_init_sock(struct sock *sk)
423 {
424 	struct inet_connection_sock *icsk = inet_csk(sk);
425 	struct tcp_sock *tp = tcp_sk(sk);
426 	int rto_min_us;
427 
428 	tp->out_of_order_queue = RB_ROOT;
429 	sk->tcp_rtx_queue = RB_ROOT;
430 	tcp_init_xmit_timers(sk);
431 	INIT_LIST_HEAD(&tp->tsq_node);
432 	INIT_LIST_HEAD(&tp->tsorted_sent_queue);
433 
434 	icsk->icsk_rto = TCP_TIMEOUT_INIT;
435 	rto_min_us = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_rto_min_us);
436 	icsk->icsk_rto_min = usecs_to_jiffies(rto_min_us);
437 	icsk->icsk_delack_max = TCP_DELACK_MAX;
438 	tp->mdev_us = jiffies_to_usecs(TCP_TIMEOUT_INIT);
439 	minmax_reset(&tp->rtt_min, tcp_jiffies32, ~0U);
440 
441 	/* So many TCP implementations out there (incorrectly) count the
442 	 * initial SYN frame in their delayed-ACK and congestion control
443 	 * algorithms that we must have the following bandaid to talk
444 	 * efficiently to them.  -DaveM
445 	 */
446 	tcp_snd_cwnd_set(tp, TCP_INIT_CWND);
447 
448 	/* There's a bubble in the pipe until at least the first ACK. */
449 	tp->app_limited = ~0U;
450 	tp->rate_app_limited = 1;
451 
452 	/* See draft-stevens-tcpca-spec-01 for discussion of the
453 	 * initialization of these values.
454 	 */
455 	tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
456 	tp->snd_cwnd_clamp = ~0;
457 	tp->mss_cache = TCP_MSS_DEFAULT;
458 
459 	tp->reordering = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_reordering);
460 	tcp_assign_congestion_control(sk);
461 
462 	tp->tsoffset = 0;
463 	tp->rack.reo_wnd_steps = 1;
464 
465 	sk->sk_write_space = sk_stream_write_space;
466 	sock_set_flag(sk, SOCK_USE_WRITE_QUEUE);
467 
468 	icsk->icsk_sync_mss = tcp_sync_mss;
469 
470 	WRITE_ONCE(sk->sk_sndbuf, READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_wmem[1]));
471 	WRITE_ONCE(sk->sk_rcvbuf, READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_rmem[1]));
472 	tcp_scaling_ratio_init(sk);
473 
474 	set_bit(SOCK_SUPPORT_ZC, &sk->sk_socket->flags);
475 	sk_sockets_allocated_inc(sk);
476 	xa_init_flags(&sk->sk_user_frags, XA_FLAGS_ALLOC1);
477 }
478 EXPORT_SYMBOL(tcp_init_sock);
479 
tcp_tx_timestamp(struct sock * sk,struct sockcm_cookie * sockc)480 static void tcp_tx_timestamp(struct sock *sk, struct sockcm_cookie *sockc)
481 {
482 	struct sk_buff *skb = tcp_write_queue_tail(sk);
483 	u32 tsflags = sockc->tsflags;
484 
485 	if (tsflags && skb) {
486 		struct skb_shared_info *shinfo = skb_shinfo(skb);
487 		struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
488 
489 		sock_tx_timestamp(sk, sockc, &shinfo->tx_flags);
490 		if (tsflags & SOF_TIMESTAMPING_TX_ACK)
491 			tcb->txstamp_ack = 1;
492 		if (tsflags & SOF_TIMESTAMPING_TX_RECORD_MASK)
493 			shinfo->tskey = TCP_SKB_CB(skb)->seq + skb->len - 1;
494 	}
495 }
496 
tcp_stream_is_readable(struct sock * sk,int target)497 static bool tcp_stream_is_readable(struct sock *sk, int target)
498 {
499 	if (tcp_epollin_ready(sk, target))
500 		return true;
501 	return sk_is_readable(sk);
502 }
503 
504 /*
505  *	Wait for a TCP event.
506  *
507  *	Note that we don't need to lock the socket, as the upper poll layers
508  *	take care of normal races (between the test and the event) and we don't
509  *	go look at any of the socket buffers directly.
510  */
tcp_poll(struct file * file,struct socket * sock,poll_table * wait)511 __poll_t tcp_poll(struct file *file, struct socket *sock, poll_table *wait)
512 {
513 	__poll_t mask;
514 	struct sock *sk = sock->sk;
515 	const struct tcp_sock *tp = tcp_sk(sk);
516 	u8 shutdown;
517 	int state;
518 
519 	sock_poll_wait(file, sock, wait);
520 
521 	state = inet_sk_state_load(sk);
522 	if (state == TCP_LISTEN)
523 		return inet_csk_listen_poll(sk);
524 
525 	/* Socket is not locked. We are protected from async events
526 	 * by poll logic and correct handling of state changes
527 	 * made by other threads is impossible in any case.
528 	 */
529 
530 	mask = 0;
531 
532 	/*
533 	 * EPOLLHUP is certainly not done right. But poll() doesn't
534 	 * have a notion of HUP in just one direction, and for a
535 	 * socket the read side is more interesting.
536 	 *
537 	 * Some poll() documentation says that EPOLLHUP is incompatible
538 	 * with the EPOLLOUT/POLLWR flags, so somebody should check this
539 	 * all. But careful, it tends to be safer to return too many
540 	 * bits than too few, and you can easily break real applications
541 	 * if you don't tell them that something has hung up!
542 	 *
543 	 * Check-me.
544 	 *
545 	 * Check number 1. EPOLLHUP is _UNMASKABLE_ event (see UNIX98 and
546 	 * our fs/select.c). It means that after we received EOF,
547 	 * poll always returns immediately, making impossible poll() on write()
548 	 * in state CLOSE_WAIT. One solution is evident --- to set EPOLLHUP
549 	 * if and only if shutdown has been made in both directions.
550 	 * Actually, it is interesting to look how Solaris and DUX
551 	 * solve this dilemma. I would prefer, if EPOLLHUP were maskable,
552 	 * then we could set it on SND_SHUTDOWN. BTW examples given
553 	 * in Stevens' books assume exactly this behaviour, it explains
554 	 * why EPOLLHUP is incompatible with EPOLLOUT.	--ANK
555 	 *
556 	 * NOTE. Check for TCP_CLOSE is added. The goal is to prevent
557 	 * blocking on fresh not-connected or disconnected socket. --ANK
558 	 */
559 	shutdown = READ_ONCE(sk->sk_shutdown);
560 	if (shutdown == SHUTDOWN_MASK || state == TCP_CLOSE)
561 		mask |= EPOLLHUP;
562 	if (shutdown & RCV_SHUTDOWN)
563 		mask |= EPOLLIN | EPOLLRDNORM | EPOLLRDHUP;
564 
565 	/* Connected or passive Fast Open socket? */
566 	if (state != TCP_SYN_SENT &&
567 	    (state != TCP_SYN_RECV || rcu_access_pointer(tp->fastopen_rsk))) {
568 		int target = sock_rcvlowat(sk, 0, INT_MAX);
569 		u16 urg_data = READ_ONCE(tp->urg_data);
570 
571 		if (unlikely(urg_data) &&
572 		    READ_ONCE(tp->urg_seq) == READ_ONCE(tp->copied_seq) &&
573 		    !sock_flag(sk, SOCK_URGINLINE))
574 			target++;
575 
576 		if (tcp_stream_is_readable(sk, target))
577 			mask |= EPOLLIN | EPOLLRDNORM;
578 
579 		if (!(shutdown & SEND_SHUTDOWN)) {
580 			if (__sk_stream_is_writeable(sk, 1)) {
581 				mask |= EPOLLOUT | EPOLLWRNORM;
582 			} else {  /* send SIGIO later */
583 				sk_set_bit(SOCKWQ_ASYNC_NOSPACE, sk);
584 				set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
585 
586 				/* Race breaker. If space is freed after
587 				 * wspace test but before the flags are set,
588 				 * IO signal will be lost. Memory barrier
589 				 * pairs with the input side.
590 				 */
591 				smp_mb__after_atomic();
592 				if (__sk_stream_is_writeable(sk, 1))
593 					mask |= EPOLLOUT | EPOLLWRNORM;
594 			}
595 		} else
596 			mask |= EPOLLOUT | EPOLLWRNORM;
597 
598 		if (urg_data & TCP_URG_VALID)
599 			mask |= EPOLLPRI;
600 	} else if (state == TCP_SYN_SENT &&
601 		   inet_test_bit(DEFER_CONNECT, sk)) {
602 		/* Active TCP fastopen socket with defer_connect
603 		 * Return EPOLLOUT so application can call write()
604 		 * in order for kernel to generate SYN+data
605 		 */
606 		mask |= EPOLLOUT | EPOLLWRNORM;
607 	}
608 	/* This barrier is coupled with smp_wmb() in tcp_done_with_error() */
609 	smp_rmb();
610 	if (READ_ONCE(sk->sk_err) ||
611 	    !skb_queue_empty_lockless(&sk->sk_error_queue))
612 		mask |= EPOLLERR;
613 
614 	return mask;
615 }
616 EXPORT_SYMBOL(tcp_poll);
617 
tcp_ioctl(struct sock * sk,int cmd,int * karg)618 int tcp_ioctl(struct sock *sk, int cmd, int *karg)
619 {
620 	struct tcp_sock *tp = tcp_sk(sk);
621 	int answ;
622 	bool slow;
623 
624 	switch (cmd) {
625 	case SIOCINQ:
626 		if (sk->sk_state == TCP_LISTEN)
627 			return -EINVAL;
628 
629 		slow = lock_sock_fast(sk);
630 		answ = tcp_inq(sk);
631 		unlock_sock_fast(sk, slow);
632 		break;
633 	case SIOCATMARK:
634 		answ = READ_ONCE(tp->urg_data) &&
635 		       READ_ONCE(tp->urg_seq) == READ_ONCE(tp->copied_seq);
636 		break;
637 	case SIOCOUTQ:
638 		if (sk->sk_state == TCP_LISTEN)
639 			return -EINVAL;
640 
641 		if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
642 			answ = 0;
643 		else
644 			answ = READ_ONCE(tp->write_seq) - tp->snd_una;
645 		break;
646 	case SIOCOUTQNSD:
647 		if (sk->sk_state == TCP_LISTEN)
648 			return -EINVAL;
649 
650 		if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
651 			answ = 0;
652 		else
653 			answ = READ_ONCE(tp->write_seq) -
654 			       READ_ONCE(tp->snd_nxt);
655 		break;
656 	default:
657 		return -ENOIOCTLCMD;
658 	}
659 
660 	*karg = answ;
661 	return 0;
662 }
663 EXPORT_SYMBOL(tcp_ioctl);
664 
tcp_mark_push(struct tcp_sock * tp,struct sk_buff * skb)665 void tcp_mark_push(struct tcp_sock *tp, struct sk_buff *skb)
666 {
667 	TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
668 	tp->pushed_seq = tp->write_seq;
669 }
670 
forced_push(const struct tcp_sock * tp)671 static inline bool forced_push(const struct tcp_sock *tp)
672 {
673 	return after(tp->write_seq, tp->pushed_seq + (tp->max_window >> 1));
674 }
675 
tcp_skb_entail(struct sock * sk,struct sk_buff * skb)676 void tcp_skb_entail(struct sock *sk, struct sk_buff *skb)
677 {
678 	struct tcp_sock *tp = tcp_sk(sk);
679 	struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
680 
681 	tcb->seq     = tcb->end_seq = tp->write_seq;
682 	tcb->tcp_flags = TCPHDR_ACK;
683 	__skb_header_release(skb);
684 	tcp_add_write_queue_tail(sk, skb);
685 	sk_wmem_queued_add(sk, skb->truesize);
686 	sk_mem_charge(sk, skb->truesize);
687 	if (tp->nonagle & TCP_NAGLE_PUSH)
688 		tp->nonagle &= ~TCP_NAGLE_PUSH;
689 
690 	tcp_slow_start_after_idle_check(sk);
691 }
692 
tcp_mark_urg(struct tcp_sock * tp,int flags)693 static inline void tcp_mark_urg(struct tcp_sock *tp, int flags)
694 {
695 	if (flags & MSG_OOB)
696 		tp->snd_up = tp->write_seq;
697 }
698 
699 /* If a not yet filled skb is pushed, do not send it if
700  * we have data packets in Qdisc or NIC queues :
701  * Because TX completion will happen shortly, it gives a chance
702  * to coalesce future sendmsg() payload into this skb, without
703  * need for a timer, and with no latency trade off.
704  * As packets containing data payload have a bigger truesize
705  * than pure acks (dataless) packets, the last checks prevent
706  * autocorking if we only have an ACK in Qdisc/NIC queues,
707  * or if TX completion was delayed after we processed ACK packet.
708  */
tcp_should_autocork(struct sock * sk,struct sk_buff * skb,int size_goal)709 static bool tcp_should_autocork(struct sock *sk, struct sk_buff *skb,
710 				int size_goal)
711 {
712 	return skb->len < size_goal &&
713 	       READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_autocorking) &&
714 	       !tcp_rtx_queue_empty(sk) &&
715 	       refcount_read(&sk->sk_wmem_alloc) > skb->truesize &&
716 	       tcp_skb_can_collapse_to(skb);
717 }
718 
tcp_push(struct sock * sk,int flags,int mss_now,int nonagle,int size_goal)719 void tcp_push(struct sock *sk, int flags, int mss_now,
720 	      int nonagle, int size_goal)
721 {
722 	struct tcp_sock *tp = tcp_sk(sk);
723 	struct sk_buff *skb;
724 
725 	skb = tcp_write_queue_tail(sk);
726 	if (!skb)
727 		return;
728 	if (!(flags & MSG_MORE) || forced_push(tp))
729 		tcp_mark_push(tp, skb);
730 
731 	tcp_mark_urg(tp, flags);
732 
733 	if (tcp_should_autocork(sk, skb, size_goal)) {
734 
735 		/* avoid atomic op if TSQ_THROTTLED bit is already set */
736 		if (!test_bit(TSQ_THROTTLED, &sk->sk_tsq_flags)) {
737 			NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPAUTOCORKING);
738 			set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags);
739 			smp_mb__after_atomic();
740 		}
741 		/* It is possible TX completion already happened
742 		 * before we set TSQ_THROTTLED.
743 		 */
744 		if (refcount_read(&sk->sk_wmem_alloc) > skb->truesize)
745 			return;
746 	}
747 
748 	if (flags & MSG_MORE)
749 		nonagle = TCP_NAGLE_CORK;
750 
751 	__tcp_push_pending_frames(sk, mss_now, nonagle);
752 }
753 
tcp_splice_data_recv(read_descriptor_t * rd_desc,struct sk_buff * skb,unsigned int offset,size_t len)754 static int tcp_splice_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb,
755 				unsigned int offset, size_t len)
756 {
757 	struct tcp_splice_state *tss = rd_desc->arg.data;
758 	int ret;
759 
760 	ret = skb_splice_bits(skb, skb->sk, offset, tss->pipe,
761 			      min(rd_desc->count, len), tss->flags);
762 	if (ret > 0)
763 		rd_desc->count -= ret;
764 	return ret;
765 }
766 
__tcp_splice_read(struct sock * sk,struct tcp_splice_state * tss)767 static int __tcp_splice_read(struct sock *sk, struct tcp_splice_state *tss)
768 {
769 	/* Store TCP splice context information in read_descriptor_t. */
770 	read_descriptor_t rd_desc = {
771 		.arg.data = tss,
772 		.count	  = tss->len,
773 	};
774 
775 	return tcp_read_sock(sk, &rd_desc, tcp_splice_data_recv);
776 }
777 
778 /**
779  *  tcp_splice_read - splice data from TCP socket to a pipe
780  * @sock:	socket to splice from
781  * @ppos:	position (not valid)
782  * @pipe:	pipe to splice to
783  * @len:	number of bytes to splice
784  * @flags:	splice modifier flags
785  *
786  * Description:
787  *    Will read pages from given socket and fill them into a pipe.
788  *
789  **/
tcp_splice_read(struct socket * sock,loff_t * ppos,struct pipe_inode_info * pipe,size_t len,unsigned int flags)790 ssize_t tcp_splice_read(struct socket *sock, loff_t *ppos,
791 			struct pipe_inode_info *pipe, size_t len,
792 			unsigned int flags)
793 {
794 	struct sock *sk = sock->sk;
795 	struct tcp_splice_state tss = {
796 		.pipe = pipe,
797 		.len = len,
798 		.flags = flags,
799 	};
800 	long timeo;
801 	ssize_t spliced;
802 	int ret;
803 
804 	sock_rps_record_flow(sk);
805 	/*
806 	 * We can't seek on a socket input
807 	 */
808 	if (unlikely(*ppos))
809 		return -ESPIPE;
810 
811 	ret = spliced = 0;
812 
813 	lock_sock(sk);
814 
815 	timeo = sock_rcvtimeo(sk, sock->file->f_flags & O_NONBLOCK);
816 	while (tss.len) {
817 		ret = __tcp_splice_read(sk, &tss);
818 		if (ret < 0)
819 			break;
820 		else if (!ret) {
821 			if (spliced)
822 				break;
823 			if (sock_flag(sk, SOCK_DONE))
824 				break;
825 			if (sk->sk_err) {
826 				ret = sock_error(sk);
827 				break;
828 			}
829 			if (sk->sk_shutdown & RCV_SHUTDOWN)
830 				break;
831 			if (sk->sk_state == TCP_CLOSE) {
832 				/*
833 				 * This occurs when user tries to read
834 				 * from never connected socket.
835 				 */
836 				ret = -ENOTCONN;
837 				break;
838 			}
839 			if (!timeo) {
840 				ret = -EAGAIN;
841 				break;
842 			}
843 			/* if __tcp_splice_read() got nothing while we have
844 			 * an skb in receive queue, we do not want to loop.
845 			 * This might happen with URG data.
846 			 */
847 			if (!skb_queue_empty(&sk->sk_receive_queue))
848 				break;
849 			ret = sk_wait_data(sk, &timeo, NULL);
850 			if (ret < 0)
851 				break;
852 			if (signal_pending(current)) {
853 				ret = sock_intr_errno(timeo);
854 				break;
855 			}
856 			continue;
857 		}
858 		tss.len -= ret;
859 		spliced += ret;
860 
861 		if (!tss.len || !timeo)
862 			break;
863 		release_sock(sk);
864 		lock_sock(sk);
865 
866 		if (sk->sk_err || sk->sk_state == TCP_CLOSE ||
867 		    (sk->sk_shutdown & RCV_SHUTDOWN) ||
868 		    signal_pending(current))
869 			break;
870 	}
871 
872 	release_sock(sk);
873 
874 	if (spliced)
875 		return spliced;
876 
877 	return ret;
878 }
879 EXPORT_SYMBOL(tcp_splice_read);
880 
tcp_stream_alloc_skb(struct sock * sk,gfp_t gfp,bool force_schedule)881 struct sk_buff *tcp_stream_alloc_skb(struct sock *sk, gfp_t gfp,
882 				     bool force_schedule)
883 {
884 	struct sk_buff *skb;
885 
886 	skb = alloc_skb_fclone(MAX_TCP_HEADER, gfp);
887 	if (likely(skb)) {
888 		bool mem_scheduled;
889 
890 		skb->truesize = SKB_TRUESIZE(skb_end_offset(skb));
891 		if (force_schedule) {
892 			mem_scheduled = true;
893 			sk_forced_mem_schedule(sk, skb->truesize);
894 		} else {
895 			mem_scheduled = sk_wmem_schedule(sk, skb->truesize);
896 		}
897 		if (likely(mem_scheduled)) {
898 			skb_reserve(skb, MAX_TCP_HEADER);
899 			skb->ip_summed = CHECKSUM_PARTIAL;
900 			INIT_LIST_HEAD(&skb->tcp_tsorted_anchor);
901 			return skb;
902 		}
903 		__kfree_skb(skb);
904 	} else {
905 		sk->sk_prot->enter_memory_pressure(sk);
906 		sk_stream_moderate_sndbuf(sk);
907 	}
908 	return NULL;
909 }
910 
tcp_xmit_size_goal(struct sock * sk,u32 mss_now,int large_allowed)911 static unsigned int tcp_xmit_size_goal(struct sock *sk, u32 mss_now,
912 				       int large_allowed)
913 {
914 	struct tcp_sock *tp = tcp_sk(sk);
915 	u32 new_size_goal, size_goal;
916 
917 	if (!large_allowed)
918 		return mss_now;
919 
920 	/* Note : tcp_tso_autosize() will eventually split this later */
921 	new_size_goal = tcp_bound_to_half_wnd(tp, sk->sk_gso_max_size);
922 
923 	/* We try hard to avoid divides here */
924 	size_goal = tp->gso_segs * mss_now;
925 	if (unlikely(new_size_goal < size_goal ||
926 		     new_size_goal >= size_goal + mss_now)) {
927 		tp->gso_segs = min_t(u16, new_size_goal / mss_now,
928 				     sk->sk_gso_max_segs);
929 		size_goal = tp->gso_segs * mss_now;
930 	}
931 
932 	return max(size_goal, mss_now);
933 }
934 
tcp_send_mss(struct sock * sk,int * size_goal,int flags)935 int tcp_send_mss(struct sock *sk, int *size_goal, int flags)
936 {
937 	int mss_now;
938 
939 	mss_now = tcp_current_mss(sk);
940 	*size_goal = tcp_xmit_size_goal(sk, mss_now, !(flags & MSG_OOB));
941 
942 	return mss_now;
943 }
944 
945 /* In some cases, sendmsg() could have added an skb to the write queue,
946  * but failed adding payload on it. We need to remove it to consume less
947  * memory, but more importantly be able to generate EPOLLOUT for Edge Trigger
948  * epoll() users. Another reason is that tcp_write_xmit() does not like
949  * finding an empty skb in the write queue.
950  */
tcp_remove_empty_skb(struct sock * sk)951 void tcp_remove_empty_skb(struct sock *sk)
952 {
953 	struct sk_buff *skb = tcp_write_queue_tail(sk);
954 
955 	if (skb && TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq) {
956 		tcp_unlink_write_queue(skb, sk);
957 		if (tcp_write_queue_empty(sk))
958 			tcp_chrono_stop(sk, TCP_CHRONO_BUSY);
959 		tcp_wmem_free_skb(sk, skb);
960 	}
961 }
962 
963 /* skb changing from pure zc to mixed, must charge zc */
tcp_downgrade_zcopy_pure(struct sock * sk,struct sk_buff * skb)964 static int tcp_downgrade_zcopy_pure(struct sock *sk, struct sk_buff *skb)
965 {
966 	if (unlikely(skb_zcopy_pure(skb))) {
967 		u32 extra = skb->truesize -
968 			    SKB_TRUESIZE(skb_end_offset(skb));
969 
970 		if (!sk_wmem_schedule(sk, extra))
971 			return -ENOMEM;
972 
973 		sk_mem_charge(sk, extra);
974 		skb_shinfo(skb)->flags &= ~SKBFL_PURE_ZEROCOPY;
975 	}
976 	return 0;
977 }
978 
979 
tcp_wmem_schedule(struct sock * sk,int copy)980 int tcp_wmem_schedule(struct sock *sk, int copy)
981 {
982 	int left;
983 
984 	if (likely(sk_wmem_schedule(sk, copy)))
985 		return copy;
986 
987 	/* We could be in trouble if we have nothing queued.
988 	 * Use whatever is left in sk->sk_forward_alloc and tcp_wmem[0]
989 	 * to guarantee some progress.
990 	 */
991 	left = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_wmem[0]) - sk->sk_wmem_queued;
992 	if (left > 0)
993 		sk_forced_mem_schedule(sk, min(left, copy));
994 	return min(copy, sk->sk_forward_alloc);
995 }
996 
tcp_free_fastopen_req(struct tcp_sock * tp)997 void tcp_free_fastopen_req(struct tcp_sock *tp)
998 {
999 	if (tp->fastopen_req) {
1000 		kfree(tp->fastopen_req);
1001 		tp->fastopen_req = NULL;
1002 	}
1003 }
1004 
tcp_sendmsg_fastopen(struct sock * sk,struct msghdr * msg,int * copied,size_t size,struct ubuf_info * uarg)1005 int tcp_sendmsg_fastopen(struct sock *sk, struct msghdr *msg, int *copied,
1006 			 size_t size, struct ubuf_info *uarg)
1007 {
1008 	struct tcp_sock *tp = tcp_sk(sk);
1009 	struct inet_sock *inet = inet_sk(sk);
1010 	struct sockaddr *uaddr = msg->msg_name;
1011 	int err, flags;
1012 
1013 	if (!(READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_fastopen) &
1014 	      TFO_CLIENT_ENABLE) ||
1015 	    (uaddr && msg->msg_namelen >= sizeof(uaddr->sa_family) &&
1016 	     uaddr->sa_family == AF_UNSPEC))
1017 		return -EOPNOTSUPP;
1018 	if (tp->fastopen_req)
1019 		return -EALREADY; /* Another Fast Open is in progress */
1020 
1021 	tp->fastopen_req = kzalloc(sizeof(struct tcp_fastopen_request),
1022 				   sk->sk_allocation);
1023 	if (unlikely(!tp->fastopen_req))
1024 		return -ENOBUFS;
1025 	tp->fastopen_req->data = msg;
1026 	tp->fastopen_req->size = size;
1027 	tp->fastopen_req->uarg = uarg;
1028 
1029 	if (inet_test_bit(DEFER_CONNECT, sk)) {
1030 		err = tcp_connect(sk);
1031 		/* Same failure procedure as in tcp_v4/6_connect */
1032 		if (err) {
1033 			tcp_set_state(sk, TCP_CLOSE);
1034 			inet->inet_dport = 0;
1035 			sk->sk_route_caps = 0;
1036 		}
1037 	}
1038 	flags = (msg->msg_flags & MSG_DONTWAIT) ? O_NONBLOCK : 0;
1039 	err = __inet_stream_connect(sk->sk_socket, uaddr,
1040 				    msg->msg_namelen, flags, 1);
1041 	/* fastopen_req could already be freed in __inet_stream_connect
1042 	 * if the connection times out or gets rst
1043 	 */
1044 	if (tp->fastopen_req) {
1045 		*copied = tp->fastopen_req->copied;
1046 		tcp_free_fastopen_req(tp);
1047 		inet_clear_bit(DEFER_CONNECT, sk);
1048 	}
1049 	return err;
1050 }
1051 
tcp_sendmsg_locked(struct sock * sk,struct msghdr * msg,size_t size)1052 int tcp_sendmsg_locked(struct sock *sk, struct msghdr *msg, size_t size)
1053 {
1054 	struct tcp_sock *tp = tcp_sk(sk);
1055 	struct ubuf_info *uarg = NULL;
1056 	struct sk_buff *skb;
1057 	struct sockcm_cookie sockc;
1058 	int flags, err, copied = 0;
1059 	int mss_now = 0, size_goal, copied_syn = 0;
1060 	int process_backlog = 0;
1061 	int zc = 0;
1062 	long timeo;
1063 
1064 	flags = msg->msg_flags;
1065 
1066 	if ((flags & MSG_ZEROCOPY) && size) {
1067 		if (msg->msg_ubuf) {
1068 			uarg = msg->msg_ubuf;
1069 			if (sk->sk_route_caps & NETIF_F_SG)
1070 				zc = MSG_ZEROCOPY;
1071 		} else if (sock_flag(sk, SOCK_ZEROCOPY)) {
1072 			skb = tcp_write_queue_tail(sk);
1073 			uarg = msg_zerocopy_realloc(sk, size, skb_zcopy(skb));
1074 			if (!uarg) {
1075 				err = -ENOBUFS;
1076 				goto out_err;
1077 			}
1078 			if (sk->sk_route_caps & NETIF_F_SG)
1079 				zc = MSG_ZEROCOPY;
1080 			else
1081 				uarg_to_msgzc(uarg)->zerocopy = 0;
1082 		}
1083 	} else if (unlikely(msg->msg_flags & MSG_SPLICE_PAGES) && size) {
1084 		if (sk->sk_route_caps & NETIF_F_SG)
1085 			zc = MSG_SPLICE_PAGES;
1086 	}
1087 
1088 	if (unlikely(flags & MSG_FASTOPEN ||
1089 		     inet_test_bit(DEFER_CONNECT, sk)) &&
1090 	    !tp->repair) {
1091 		err = tcp_sendmsg_fastopen(sk, msg, &copied_syn, size, uarg);
1092 		if (err == -EINPROGRESS && copied_syn > 0)
1093 			goto out;
1094 		else if (err)
1095 			goto out_err;
1096 	}
1097 
1098 	timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
1099 
1100 	tcp_rate_check_app_limited(sk);  /* is sending application-limited? */
1101 
1102 	/* Wait for a connection to finish. One exception is TCP Fast Open
1103 	 * (passive side) where data is allowed to be sent before a connection
1104 	 * is fully established.
1105 	 */
1106 	if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) &&
1107 	    !tcp_passive_fastopen(sk)) {
1108 		err = sk_stream_wait_connect(sk, &timeo);
1109 		if (err != 0)
1110 			goto do_error;
1111 	}
1112 
1113 	if (unlikely(tp->repair)) {
1114 		if (tp->repair_queue == TCP_RECV_QUEUE) {
1115 			copied = tcp_send_rcvq(sk, msg, size);
1116 			goto out_nopush;
1117 		}
1118 
1119 		err = -EINVAL;
1120 		if (tp->repair_queue == TCP_NO_QUEUE)
1121 			goto out_err;
1122 
1123 		/* 'common' sending to sendq */
1124 	}
1125 
1126 	sockcm_init(&sockc, sk);
1127 	if (msg->msg_controllen) {
1128 		err = sock_cmsg_send(sk, msg, &sockc);
1129 		if (unlikely(err)) {
1130 			err = -EINVAL;
1131 			goto out_err;
1132 		}
1133 	}
1134 
1135 	/* This should be in poll */
1136 	sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
1137 
1138 	/* Ok commence sending. */
1139 	copied = 0;
1140 
1141 restart:
1142 	mss_now = tcp_send_mss(sk, &size_goal, flags);
1143 
1144 	err = -EPIPE;
1145 	if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
1146 		goto do_error;
1147 
1148 	while (msg_data_left(msg)) {
1149 		ssize_t copy = 0;
1150 
1151 		skb = tcp_write_queue_tail(sk);
1152 		if (skb)
1153 			copy = size_goal - skb->len;
1154 
1155 		if (copy <= 0 || !tcp_skb_can_collapse_to(skb)) {
1156 			bool first_skb;
1157 
1158 new_segment:
1159 			if (!sk_stream_memory_free(sk))
1160 				goto wait_for_space;
1161 
1162 			if (unlikely(process_backlog >= 16)) {
1163 				process_backlog = 0;
1164 				if (sk_flush_backlog(sk))
1165 					goto restart;
1166 			}
1167 			first_skb = tcp_rtx_and_write_queues_empty(sk);
1168 			skb = tcp_stream_alloc_skb(sk, sk->sk_allocation,
1169 						   first_skb);
1170 			if (!skb)
1171 				goto wait_for_space;
1172 
1173 			process_backlog++;
1174 
1175 #ifdef CONFIG_SKB_DECRYPTED
1176 			skb->decrypted = !!(flags & MSG_SENDPAGE_DECRYPTED);
1177 #endif
1178 			tcp_skb_entail(sk, skb);
1179 			copy = size_goal;
1180 
1181 			/* All packets are restored as if they have
1182 			 * already been sent. skb_mstamp_ns isn't set to
1183 			 * avoid wrong rtt estimation.
1184 			 */
1185 			if (tp->repair)
1186 				TCP_SKB_CB(skb)->sacked |= TCPCB_REPAIRED;
1187 		}
1188 
1189 		/* Try to append data to the end of skb. */
1190 		if (copy > msg_data_left(msg))
1191 			copy = msg_data_left(msg);
1192 
1193 		if (zc == 0) {
1194 			bool merge = true;
1195 			int i = skb_shinfo(skb)->nr_frags;
1196 			struct page_frag *pfrag = sk_page_frag(sk);
1197 
1198 			if (!sk_page_frag_refill(sk, pfrag))
1199 				goto wait_for_space;
1200 
1201 			if (!skb_can_coalesce(skb, i, pfrag->page,
1202 					      pfrag->offset)) {
1203 				if (i >= READ_ONCE(net_hotdata.sysctl_max_skb_frags)) {
1204 					tcp_mark_push(tp, skb);
1205 					goto new_segment;
1206 				}
1207 				merge = false;
1208 			}
1209 
1210 			copy = min_t(int, copy, pfrag->size - pfrag->offset);
1211 
1212 			if (unlikely(skb_zcopy_pure(skb) || skb_zcopy_managed(skb))) {
1213 				if (tcp_downgrade_zcopy_pure(sk, skb))
1214 					goto wait_for_space;
1215 				skb_zcopy_downgrade_managed(skb);
1216 			}
1217 
1218 			copy = tcp_wmem_schedule(sk, copy);
1219 			if (!copy)
1220 				goto wait_for_space;
1221 
1222 			err = skb_copy_to_page_nocache(sk, &msg->msg_iter, skb,
1223 						       pfrag->page,
1224 						       pfrag->offset,
1225 						       copy);
1226 			if (err)
1227 				goto do_error;
1228 
1229 			/* Update the skb. */
1230 			if (merge) {
1231 				skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1232 			} else {
1233 				skb_fill_page_desc(skb, i, pfrag->page,
1234 						   pfrag->offset, copy);
1235 				page_ref_inc(pfrag->page);
1236 			}
1237 			pfrag->offset += copy;
1238 		} else if (zc == MSG_ZEROCOPY)  {
1239 			/* First append to a fragless skb builds initial
1240 			 * pure zerocopy skb
1241 			 */
1242 			if (!skb->len)
1243 				skb_shinfo(skb)->flags |= SKBFL_PURE_ZEROCOPY;
1244 
1245 			if (!skb_zcopy_pure(skb)) {
1246 				copy = tcp_wmem_schedule(sk, copy);
1247 				if (!copy)
1248 					goto wait_for_space;
1249 			}
1250 
1251 			err = skb_zerocopy_iter_stream(sk, skb, msg, copy, uarg);
1252 			if (err == -EMSGSIZE || err == -EEXIST) {
1253 				tcp_mark_push(tp, skb);
1254 				goto new_segment;
1255 			}
1256 			if (err < 0)
1257 				goto do_error;
1258 			copy = err;
1259 		} else if (zc == MSG_SPLICE_PAGES) {
1260 			/* Splice in data if we can; copy if we can't. */
1261 			if (tcp_downgrade_zcopy_pure(sk, skb))
1262 				goto wait_for_space;
1263 			copy = tcp_wmem_schedule(sk, copy);
1264 			if (!copy)
1265 				goto wait_for_space;
1266 
1267 			err = skb_splice_from_iter(skb, &msg->msg_iter, copy,
1268 						   sk->sk_allocation);
1269 			if (err < 0) {
1270 				if (err == -EMSGSIZE) {
1271 					tcp_mark_push(tp, skb);
1272 					goto new_segment;
1273 				}
1274 				goto do_error;
1275 			}
1276 			copy = err;
1277 
1278 			if (!(flags & MSG_NO_SHARED_FRAGS))
1279 				skb_shinfo(skb)->flags |= SKBFL_SHARED_FRAG;
1280 
1281 			sk_wmem_queued_add(sk, copy);
1282 			sk_mem_charge(sk, copy);
1283 		}
1284 
1285 		if (!copied)
1286 			TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
1287 
1288 		WRITE_ONCE(tp->write_seq, tp->write_seq + copy);
1289 		TCP_SKB_CB(skb)->end_seq += copy;
1290 		tcp_skb_pcount_set(skb, 0);
1291 
1292 		copied += copy;
1293 		if (!msg_data_left(msg)) {
1294 			if (unlikely(flags & MSG_EOR))
1295 				TCP_SKB_CB(skb)->eor = 1;
1296 			goto out;
1297 		}
1298 
1299 		if (skb->len < size_goal || (flags & MSG_OOB) || unlikely(tp->repair))
1300 			continue;
1301 
1302 		if (forced_push(tp)) {
1303 			tcp_mark_push(tp, skb);
1304 			__tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
1305 		} else if (skb == tcp_send_head(sk))
1306 			tcp_push_one(sk, mss_now);
1307 		continue;
1308 
1309 wait_for_space:
1310 		set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1311 		tcp_remove_empty_skb(sk);
1312 		if (copied)
1313 			tcp_push(sk, flags & ~MSG_MORE, mss_now,
1314 				 TCP_NAGLE_PUSH, size_goal);
1315 
1316 		err = sk_stream_wait_memory(sk, &timeo);
1317 		if (err != 0)
1318 			goto do_error;
1319 
1320 		mss_now = tcp_send_mss(sk, &size_goal, flags);
1321 	}
1322 
1323 out:
1324 	if (copied) {
1325 		tcp_tx_timestamp(sk, &sockc);
1326 		tcp_push(sk, flags, mss_now, tp->nonagle, size_goal);
1327 	}
1328 out_nopush:
1329 	/* msg->msg_ubuf is pinned by the caller so we don't take extra refs */
1330 	if (uarg && !msg->msg_ubuf)
1331 		net_zcopy_put(uarg);
1332 	return copied + copied_syn;
1333 
1334 do_error:
1335 	tcp_remove_empty_skb(sk);
1336 
1337 	if (copied + copied_syn)
1338 		goto out;
1339 out_err:
1340 	/* msg->msg_ubuf is pinned by the caller so we don't take extra refs */
1341 	if (uarg && !msg->msg_ubuf)
1342 		net_zcopy_put_abort(uarg, true);
1343 	err = sk_stream_error(sk, flags, err);
1344 	/* make sure we wake any epoll edge trigger waiter */
1345 	if (unlikely(tcp_rtx_and_write_queues_empty(sk) && err == -EAGAIN)) {
1346 		sk->sk_write_space(sk);
1347 		tcp_chrono_stop(sk, TCP_CHRONO_SNDBUF_LIMITED);
1348 	}
1349 	return err;
1350 }
1351 EXPORT_SYMBOL_GPL(tcp_sendmsg_locked);
1352 
tcp_sendmsg(struct sock * sk,struct msghdr * msg,size_t size)1353 int tcp_sendmsg(struct sock *sk, struct msghdr *msg, size_t size)
1354 {
1355 	int ret;
1356 
1357 	lock_sock(sk);
1358 	ret = tcp_sendmsg_locked(sk, msg, size);
1359 	release_sock(sk);
1360 
1361 	return ret;
1362 }
1363 EXPORT_SYMBOL(tcp_sendmsg);
1364 
tcp_splice_eof(struct socket * sock)1365 void tcp_splice_eof(struct socket *sock)
1366 {
1367 	struct sock *sk = sock->sk;
1368 	struct tcp_sock *tp = tcp_sk(sk);
1369 	int mss_now, size_goal;
1370 
1371 	if (!tcp_write_queue_tail(sk))
1372 		return;
1373 
1374 	lock_sock(sk);
1375 	mss_now = tcp_send_mss(sk, &size_goal, 0);
1376 	tcp_push(sk, 0, mss_now, tp->nonagle, size_goal);
1377 	release_sock(sk);
1378 }
1379 EXPORT_SYMBOL_GPL(tcp_splice_eof);
1380 
1381 /*
1382  *	Handle reading urgent data. BSD has very simple semantics for
1383  *	this, no blocking and very strange errors 8)
1384  */
1385 
tcp_recv_urg(struct sock * sk,struct msghdr * msg,int len,int flags)1386 static int tcp_recv_urg(struct sock *sk, struct msghdr *msg, int len, int flags)
1387 {
1388 	struct tcp_sock *tp = tcp_sk(sk);
1389 
1390 	/* No URG data to read. */
1391 	if (sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data ||
1392 	    tp->urg_data == TCP_URG_READ)
1393 		return -EINVAL;	/* Yes this is right ! */
1394 
1395 	if (sk->sk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DONE))
1396 		return -ENOTCONN;
1397 
1398 	if (tp->urg_data & TCP_URG_VALID) {
1399 		int err = 0;
1400 		char c = tp->urg_data;
1401 
1402 		if (!(flags & MSG_PEEK))
1403 			WRITE_ONCE(tp->urg_data, TCP_URG_READ);
1404 
1405 		/* Read urgent data. */
1406 		msg->msg_flags |= MSG_OOB;
1407 
1408 		if (len > 0) {
1409 			if (!(flags & MSG_TRUNC))
1410 				err = memcpy_to_msg(msg, &c, 1);
1411 			len = 1;
1412 		} else
1413 			msg->msg_flags |= MSG_TRUNC;
1414 
1415 		return err ? -EFAULT : len;
1416 	}
1417 
1418 	if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
1419 		return 0;
1420 
1421 	/* Fixed the recv(..., MSG_OOB) behaviour.  BSD docs and
1422 	 * the available implementations agree in this case:
1423 	 * this call should never block, independent of the
1424 	 * blocking state of the socket.
1425 	 * Mike <pall@rz.uni-karlsruhe.de>
1426 	 */
1427 	return -EAGAIN;
1428 }
1429 
tcp_peek_sndq(struct sock * sk,struct msghdr * msg,int len)1430 static int tcp_peek_sndq(struct sock *sk, struct msghdr *msg, int len)
1431 {
1432 	struct sk_buff *skb;
1433 	int copied = 0, err = 0;
1434 
1435 	skb_rbtree_walk(skb, &sk->tcp_rtx_queue) {
1436 		err = skb_copy_datagram_msg(skb, 0, msg, skb->len);
1437 		if (err)
1438 			return err;
1439 		copied += skb->len;
1440 	}
1441 
1442 	skb_queue_walk(&sk->sk_write_queue, skb) {
1443 		err = skb_copy_datagram_msg(skb, 0, msg, skb->len);
1444 		if (err)
1445 			break;
1446 
1447 		copied += skb->len;
1448 	}
1449 
1450 	return err ?: copied;
1451 }
1452 
1453 /* Clean up the receive buffer for full frames taken by the user,
1454  * then send an ACK if necessary.  COPIED is the number of bytes
1455  * tcp_recvmsg has given to the user so far, it speeds up the
1456  * calculation of whether or not we must ACK for the sake of
1457  * a window update.
1458  */
__tcp_cleanup_rbuf(struct sock * sk,int copied)1459 void __tcp_cleanup_rbuf(struct sock *sk, int copied)
1460 {
1461 	struct tcp_sock *tp = tcp_sk(sk);
1462 	bool time_to_ack = false;
1463 
1464 	if (inet_csk_ack_scheduled(sk)) {
1465 		const struct inet_connection_sock *icsk = inet_csk(sk);
1466 
1467 		if (/* Once-per-two-segments ACK was not sent by tcp_input.c */
1468 		    tp->rcv_nxt - tp->rcv_wup > icsk->icsk_ack.rcv_mss ||
1469 		    /*
1470 		     * If this read emptied read buffer, we send ACK, if
1471 		     * connection is not bidirectional, user drained
1472 		     * receive buffer and there was a small segment
1473 		     * in queue.
1474 		     */
1475 		    (copied > 0 &&
1476 		     ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED2) ||
1477 		      ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED) &&
1478 		       !inet_csk_in_pingpong_mode(sk))) &&
1479 		      !atomic_read(&sk->sk_rmem_alloc)))
1480 			time_to_ack = true;
1481 	}
1482 
1483 	/* We send an ACK if we can now advertise a non-zero window
1484 	 * which has been raised "significantly".
1485 	 *
1486 	 * Even if window raised up to infinity, do not send window open ACK
1487 	 * in states, where we will not receive more. It is useless.
1488 	 */
1489 	if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) {
1490 		__u32 rcv_window_now = tcp_receive_window(tp);
1491 
1492 		/* Optimize, __tcp_select_window() is not cheap. */
1493 		if (2*rcv_window_now <= tp->window_clamp) {
1494 			__u32 new_window = __tcp_select_window(sk);
1495 
1496 			/* Send ACK now, if this read freed lots of space
1497 			 * in our buffer. Certainly, new_window is new window.
1498 			 * We can advertise it now, if it is not less than current one.
1499 			 * "Lots" means "at least twice" here.
1500 			 */
1501 			if (new_window && new_window >= 2 * rcv_window_now)
1502 				time_to_ack = true;
1503 		}
1504 	}
1505 	if (time_to_ack)
1506 		tcp_send_ack(sk);
1507 }
1508 
tcp_cleanup_rbuf(struct sock * sk,int copied)1509 void tcp_cleanup_rbuf(struct sock *sk, int copied)
1510 {
1511 	struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
1512 	struct tcp_sock *tp = tcp_sk(sk);
1513 
1514 	WARN(skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq),
1515 	     "cleanup rbuf bug: copied %X seq %X rcvnxt %X\n",
1516 	     tp->copied_seq, TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt);
1517 	__tcp_cleanup_rbuf(sk, copied);
1518 }
1519 
tcp_eat_recv_skb(struct sock * sk,struct sk_buff * skb)1520 static void tcp_eat_recv_skb(struct sock *sk, struct sk_buff *skb)
1521 {
1522 	__skb_unlink(skb, &sk->sk_receive_queue);
1523 	if (likely(skb->destructor == sock_rfree)) {
1524 		sock_rfree(skb);
1525 		skb->destructor = NULL;
1526 		skb->sk = NULL;
1527 		return skb_attempt_defer_free(skb);
1528 	}
1529 	__kfree_skb(skb);
1530 }
1531 
tcp_recv_skb(struct sock * sk,u32 seq,u32 * off)1532 struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off)
1533 {
1534 	struct sk_buff *skb;
1535 	u32 offset;
1536 
1537 	while ((skb = skb_peek(&sk->sk_receive_queue)) != NULL) {
1538 		offset = seq - TCP_SKB_CB(skb)->seq;
1539 		if (unlikely(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
1540 			pr_err_once("%s: found a SYN, please report !\n", __func__);
1541 			offset--;
1542 		}
1543 		if (offset < skb->len || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)) {
1544 			*off = offset;
1545 			return skb;
1546 		}
1547 		/* This looks weird, but this can happen if TCP collapsing
1548 		 * splitted a fat GRO packet, while we released socket lock
1549 		 * in skb_splice_bits()
1550 		 */
1551 		tcp_eat_recv_skb(sk, skb);
1552 	}
1553 	return NULL;
1554 }
1555 EXPORT_SYMBOL(tcp_recv_skb);
1556 
1557 /*
1558  * This routine provides an alternative to tcp_recvmsg() for routines
1559  * that would like to handle copying from skbuffs directly in 'sendfile'
1560  * fashion.
1561  * Note:
1562  *	- It is assumed that the socket was locked by the caller.
1563  *	- The routine does not block.
1564  *	- At present, there is no support for reading OOB data
1565  *	  or for 'peeking' the socket using this routine
1566  *	  (although both would be easy to implement).
1567  */
__tcp_read_sock(struct sock * sk,read_descriptor_t * desc,sk_read_actor_t recv_actor,bool noack,u32 * copied_seq)1568 static int __tcp_read_sock(struct sock *sk, read_descriptor_t *desc,
1569 			   sk_read_actor_t recv_actor, bool noack,
1570 			   u32 *copied_seq)
1571 {
1572 	struct sk_buff *skb;
1573 	struct tcp_sock *tp = tcp_sk(sk);
1574 	u32 seq = *copied_seq;
1575 	u32 offset;
1576 	int copied = 0;
1577 
1578 	if (sk->sk_state == TCP_LISTEN)
1579 		return -ENOTCONN;
1580 	while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1581 		if (offset < skb->len) {
1582 			int used;
1583 			size_t len;
1584 
1585 			len = skb->len - offset;
1586 			/* Stop reading if we hit a patch of urgent data */
1587 			if (unlikely(tp->urg_data)) {
1588 				u32 urg_offset = tp->urg_seq - seq;
1589 				if (urg_offset < len)
1590 					len = urg_offset;
1591 				if (!len)
1592 					break;
1593 			}
1594 			used = recv_actor(desc, skb, offset, len);
1595 			if (used <= 0) {
1596 				if (!copied)
1597 					copied = used;
1598 				break;
1599 			}
1600 			if (WARN_ON_ONCE(used > len))
1601 				used = len;
1602 			seq += used;
1603 			copied += used;
1604 			offset += used;
1605 
1606 			/* If recv_actor drops the lock (e.g. TCP splice
1607 			 * receive) the skb pointer might be invalid when
1608 			 * getting here: tcp_collapse might have deleted it
1609 			 * while aggregating skbs from the socket queue.
1610 			 */
1611 			skb = tcp_recv_skb(sk, seq - 1, &offset);
1612 			if (!skb)
1613 				break;
1614 			/* TCP coalescing might have appended data to the skb.
1615 			 * Try to splice more frags
1616 			 */
1617 			if (offset + 1 != skb->len)
1618 				continue;
1619 		}
1620 		if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) {
1621 			tcp_eat_recv_skb(sk, skb);
1622 			++seq;
1623 			break;
1624 		}
1625 		tcp_eat_recv_skb(sk, skb);
1626 		if (!desc->count)
1627 			break;
1628 		WRITE_ONCE(*copied_seq, seq);
1629 	}
1630 	WRITE_ONCE(*copied_seq, seq);
1631 
1632 	if (noack)
1633 		goto out;
1634 
1635 	tcp_rcv_space_adjust(sk);
1636 
1637 	/* Clean up data we have read: This will do ACK frames. */
1638 	if (copied > 0) {
1639 		tcp_recv_skb(sk, seq, &offset);
1640 		tcp_cleanup_rbuf(sk, copied);
1641 	}
1642 out:
1643 	return copied;
1644 }
1645 
tcp_read_sock(struct sock * sk,read_descriptor_t * desc,sk_read_actor_t recv_actor)1646 int tcp_read_sock(struct sock *sk, read_descriptor_t *desc,
1647 		  sk_read_actor_t recv_actor)
1648 {
1649 	return __tcp_read_sock(sk, desc, recv_actor, false,
1650 			       &tcp_sk(sk)->copied_seq);
1651 }
1652 EXPORT_SYMBOL(tcp_read_sock);
1653 
tcp_read_sock_noack(struct sock * sk,read_descriptor_t * desc,sk_read_actor_t recv_actor,bool noack,u32 * copied_seq)1654 int tcp_read_sock_noack(struct sock *sk, read_descriptor_t *desc,
1655 			sk_read_actor_t recv_actor, bool noack,
1656 			u32 *copied_seq)
1657 {
1658 	return __tcp_read_sock(sk, desc, recv_actor, noack, copied_seq);
1659 }
1660 
tcp_read_skb(struct sock * sk,skb_read_actor_t recv_actor)1661 int tcp_read_skb(struct sock *sk, skb_read_actor_t recv_actor)
1662 {
1663 	struct sk_buff *skb;
1664 	int copied = 0;
1665 
1666 	if (sk->sk_state == TCP_LISTEN)
1667 		return -ENOTCONN;
1668 
1669 	while ((skb = skb_peek(&sk->sk_receive_queue)) != NULL) {
1670 		u8 tcp_flags;
1671 		int used;
1672 
1673 		__skb_unlink(skb, &sk->sk_receive_queue);
1674 		WARN_ON_ONCE(!skb_set_owner_sk_safe(skb, sk));
1675 		tcp_flags = TCP_SKB_CB(skb)->tcp_flags;
1676 		used = recv_actor(sk, skb);
1677 		if (used < 0) {
1678 			if (!copied)
1679 				copied = used;
1680 			break;
1681 		}
1682 		copied += used;
1683 
1684 		if (tcp_flags & TCPHDR_FIN)
1685 			break;
1686 	}
1687 	return copied;
1688 }
1689 EXPORT_SYMBOL(tcp_read_skb);
1690 
tcp_read_done(struct sock * sk,size_t len)1691 void tcp_read_done(struct sock *sk, size_t len)
1692 {
1693 	struct tcp_sock *tp = tcp_sk(sk);
1694 	u32 seq = tp->copied_seq;
1695 	struct sk_buff *skb;
1696 	size_t left;
1697 	u32 offset;
1698 
1699 	if (sk->sk_state == TCP_LISTEN)
1700 		return;
1701 
1702 	left = len;
1703 	while (left && (skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1704 		int used;
1705 
1706 		used = min_t(size_t, skb->len - offset, left);
1707 		seq += used;
1708 		left -= used;
1709 
1710 		if (skb->len > offset + used)
1711 			break;
1712 
1713 		if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) {
1714 			tcp_eat_recv_skb(sk, skb);
1715 			++seq;
1716 			break;
1717 		}
1718 		tcp_eat_recv_skb(sk, skb);
1719 	}
1720 	WRITE_ONCE(tp->copied_seq, seq);
1721 
1722 	tcp_rcv_space_adjust(sk);
1723 
1724 	/* Clean up data we have read: This will do ACK frames. */
1725 	if (left != len)
1726 		tcp_cleanup_rbuf(sk, len - left);
1727 }
1728 EXPORT_SYMBOL(tcp_read_done);
1729 
tcp_peek_len(struct socket * sock)1730 int tcp_peek_len(struct socket *sock)
1731 {
1732 	return tcp_inq(sock->sk);
1733 }
1734 EXPORT_SYMBOL(tcp_peek_len);
1735 
1736 /* Make sure sk_rcvbuf is big enough to satisfy SO_RCVLOWAT hint */
tcp_set_rcvlowat(struct sock * sk,int val)1737 int tcp_set_rcvlowat(struct sock *sk, int val)
1738 {
1739 	int space, cap;
1740 
1741 	if (sk->sk_userlocks & SOCK_RCVBUF_LOCK)
1742 		cap = sk->sk_rcvbuf >> 1;
1743 	else
1744 		cap = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_rmem[2]) >> 1;
1745 	val = min(val, cap);
1746 	WRITE_ONCE(sk->sk_rcvlowat, val ? : 1);
1747 
1748 	/* Check if we need to signal EPOLLIN right now */
1749 	tcp_data_ready(sk);
1750 
1751 	if (sk->sk_userlocks & SOCK_RCVBUF_LOCK)
1752 		return 0;
1753 
1754 	space = tcp_space_from_win(sk, val);
1755 	if (space > sk->sk_rcvbuf) {
1756 		WRITE_ONCE(sk->sk_rcvbuf, space);
1757 		WRITE_ONCE(tcp_sk(sk)->window_clamp, val);
1758 	}
1759 	return 0;
1760 }
1761 EXPORT_SYMBOL(tcp_set_rcvlowat);
1762 
tcp_update_recv_tstamps(struct sk_buff * skb,struct scm_timestamping_internal * tss)1763 void tcp_update_recv_tstamps(struct sk_buff *skb,
1764 			     struct scm_timestamping_internal *tss)
1765 {
1766 	if (skb->tstamp)
1767 		tss->ts[0] = ktime_to_timespec64(skb->tstamp);
1768 	else
1769 		tss->ts[0] = (struct timespec64) {0};
1770 
1771 	if (skb_hwtstamps(skb)->hwtstamp)
1772 		tss->ts[2] = ktime_to_timespec64(skb_hwtstamps(skb)->hwtstamp);
1773 	else
1774 		tss->ts[2] = (struct timespec64) {0};
1775 }
1776 
1777 #ifdef CONFIG_MMU
1778 static const struct vm_operations_struct tcp_vm_ops = {
1779 };
1780 
tcp_mmap(struct file * file,struct socket * sock,struct vm_area_struct * vma)1781 int tcp_mmap(struct file *file, struct socket *sock,
1782 	     struct vm_area_struct *vma)
1783 {
1784 	if (vma->vm_flags & (VM_WRITE | VM_EXEC))
1785 		return -EPERM;
1786 	vm_flags_clear(vma, VM_MAYWRITE | VM_MAYEXEC);
1787 
1788 	/* Instruct vm_insert_page() to not mmap_read_lock(mm) */
1789 	vm_flags_set(vma, VM_MIXEDMAP);
1790 
1791 	vma->vm_ops = &tcp_vm_ops;
1792 	return 0;
1793 }
1794 EXPORT_SYMBOL(tcp_mmap);
1795 
skb_advance_to_frag(struct sk_buff * skb,u32 offset_skb,u32 * offset_frag)1796 static skb_frag_t *skb_advance_to_frag(struct sk_buff *skb, u32 offset_skb,
1797 				       u32 *offset_frag)
1798 {
1799 	skb_frag_t *frag;
1800 
1801 	if (unlikely(offset_skb >= skb->len))
1802 		return NULL;
1803 
1804 	offset_skb -= skb_headlen(skb);
1805 	if ((int)offset_skb < 0 || skb_has_frag_list(skb))
1806 		return NULL;
1807 
1808 	frag = skb_shinfo(skb)->frags;
1809 	while (offset_skb) {
1810 		if (skb_frag_size(frag) > offset_skb) {
1811 			*offset_frag = offset_skb;
1812 			return frag;
1813 		}
1814 		offset_skb -= skb_frag_size(frag);
1815 		++frag;
1816 	}
1817 	*offset_frag = 0;
1818 	return frag;
1819 }
1820 
can_map_frag(const skb_frag_t * frag)1821 static bool can_map_frag(const skb_frag_t *frag)
1822 {
1823 	struct page *page;
1824 
1825 	if (skb_frag_size(frag) != PAGE_SIZE || skb_frag_off(frag))
1826 		return false;
1827 
1828 	page = skb_frag_page(frag);
1829 
1830 	if (PageCompound(page) || page->mapping)
1831 		return false;
1832 
1833 	return true;
1834 }
1835 
find_next_mappable_frag(const skb_frag_t * frag,int remaining_in_skb)1836 static int find_next_mappable_frag(const skb_frag_t *frag,
1837 				   int remaining_in_skb)
1838 {
1839 	int offset = 0;
1840 
1841 	if (likely(can_map_frag(frag)))
1842 		return 0;
1843 
1844 	while (offset < remaining_in_skb && !can_map_frag(frag)) {
1845 		offset += skb_frag_size(frag);
1846 		++frag;
1847 	}
1848 	return offset;
1849 }
1850 
tcp_zerocopy_set_hint_for_skb(struct sock * sk,struct tcp_zerocopy_receive * zc,struct sk_buff * skb,u32 offset)1851 static void tcp_zerocopy_set_hint_for_skb(struct sock *sk,
1852 					  struct tcp_zerocopy_receive *zc,
1853 					  struct sk_buff *skb, u32 offset)
1854 {
1855 	u32 frag_offset, partial_frag_remainder = 0;
1856 	int mappable_offset;
1857 	skb_frag_t *frag;
1858 
1859 	/* worst case: skip to next skb. try to improve on this case below */
1860 	zc->recv_skip_hint = skb->len - offset;
1861 
1862 	/* Find the frag containing this offset (and how far into that frag) */
1863 	frag = skb_advance_to_frag(skb, offset, &frag_offset);
1864 	if (!frag)
1865 		return;
1866 
1867 	if (frag_offset) {
1868 		struct skb_shared_info *info = skb_shinfo(skb);
1869 
1870 		/* We read part of the last frag, must recvmsg() rest of skb. */
1871 		if (frag == &info->frags[info->nr_frags - 1])
1872 			return;
1873 
1874 		/* Else, we must at least read the remainder in this frag. */
1875 		partial_frag_remainder = skb_frag_size(frag) - frag_offset;
1876 		zc->recv_skip_hint -= partial_frag_remainder;
1877 		++frag;
1878 	}
1879 
1880 	/* partial_frag_remainder: If part way through a frag, must read rest.
1881 	 * mappable_offset: Bytes till next mappable frag, *not* counting bytes
1882 	 * in partial_frag_remainder.
1883 	 */
1884 	mappable_offset = find_next_mappable_frag(frag, zc->recv_skip_hint);
1885 	zc->recv_skip_hint = mappable_offset + partial_frag_remainder;
1886 }
1887 
1888 static int tcp_recvmsg_locked(struct sock *sk, struct msghdr *msg, size_t len,
1889 			      int flags, struct scm_timestamping_internal *tss,
1890 			      int *cmsg_flags);
receive_fallback_to_copy(struct sock * sk,struct tcp_zerocopy_receive * zc,int inq,struct scm_timestamping_internal * tss)1891 static int receive_fallback_to_copy(struct sock *sk,
1892 				    struct tcp_zerocopy_receive *zc, int inq,
1893 				    struct scm_timestamping_internal *tss)
1894 {
1895 	unsigned long copy_address = (unsigned long)zc->copybuf_address;
1896 	struct msghdr msg = {};
1897 	int err;
1898 
1899 	zc->length = 0;
1900 	zc->recv_skip_hint = 0;
1901 
1902 	if (copy_address != zc->copybuf_address)
1903 		return -EINVAL;
1904 
1905 	err = import_ubuf(ITER_DEST, (void __user *)copy_address, inq,
1906 			  &msg.msg_iter);
1907 	if (err)
1908 		return err;
1909 
1910 	err = tcp_recvmsg_locked(sk, &msg, inq, MSG_DONTWAIT,
1911 				 tss, &zc->msg_flags);
1912 	if (err < 0)
1913 		return err;
1914 
1915 	zc->copybuf_len = err;
1916 	if (likely(zc->copybuf_len)) {
1917 		struct sk_buff *skb;
1918 		u32 offset;
1919 
1920 		skb = tcp_recv_skb(sk, tcp_sk(sk)->copied_seq, &offset);
1921 		if (skb)
1922 			tcp_zerocopy_set_hint_for_skb(sk, zc, skb, offset);
1923 	}
1924 	return 0;
1925 }
1926 
tcp_copy_straggler_data(struct tcp_zerocopy_receive * zc,struct sk_buff * skb,u32 copylen,u32 * offset,u32 * seq)1927 static int tcp_copy_straggler_data(struct tcp_zerocopy_receive *zc,
1928 				   struct sk_buff *skb, u32 copylen,
1929 				   u32 *offset, u32 *seq)
1930 {
1931 	unsigned long copy_address = (unsigned long)zc->copybuf_address;
1932 	struct msghdr msg = {};
1933 	int err;
1934 
1935 	if (copy_address != zc->copybuf_address)
1936 		return -EINVAL;
1937 
1938 	err = import_ubuf(ITER_DEST, (void __user *)copy_address, copylen,
1939 			  &msg.msg_iter);
1940 	if (err)
1941 		return err;
1942 	err = skb_copy_datagram_msg(skb, *offset, &msg, copylen);
1943 	if (err)
1944 		return err;
1945 	zc->recv_skip_hint -= copylen;
1946 	*offset += copylen;
1947 	*seq += copylen;
1948 	return (__s32)copylen;
1949 }
1950 
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)1951 static int tcp_zc_handle_leftover(struct tcp_zerocopy_receive *zc,
1952 				  struct sock *sk,
1953 				  struct sk_buff *skb,
1954 				  u32 *seq,
1955 				  s32 copybuf_len,
1956 				  struct scm_timestamping_internal *tss)
1957 {
1958 	u32 offset, copylen = min_t(u32, copybuf_len, zc->recv_skip_hint);
1959 
1960 	if (!copylen)
1961 		return 0;
1962 	/* skb is null if inq < PAGE_SIZE. */
1963 	if (skb) {
1964 		offset = *seq - TCP_SKB_CB(skb)->seq;
1965 	} else {
1966 		skb = tcp_recv_skb(sk, *seq, &offset);
1967 		if (TCP_SKB_CB(skb)->has_rxtstamp) {
1968 			tcp_update_recv_tstamps(skb, tss);
1969 			zc->msg_flags |= TCP_CMSG_TS;
1970 		}
1971 	}
1972 
1973 	zc->copybuf_len = tcp_copy_straggler_data(zc, skb, copylen, &offset,
1974 						  seq);
1975 	return zc->copybuf_len < 0 ? 0 : copylen;
1976 }
1977 
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)1978 static int tcp_zerocopy_vm_insert_batch_error(struct vm_area_struct *vma,
1979 					      struct page **pending_pages,
1980 					      unsigned long pages_remaining,
1981 					      unsigned long *address,
1982 					      u32 *length,
1983 					      u32 *seq,
1984 					      struct tcp_zerocopy_receive *zc,
1985 					      u32 total_bytes_to_map,
1986 					      int err)
1987 {
1988 	/* At least one page did not map. Try zapping if we skipped earlier. */
1989 	if (err == -EBUSY &&
1990 	    zc->flags & TCP_RECEIVE_ZEROCOPY_FLAG_TLB_CLEAN_HINT) {
1991 		u32 maybe_zap_len;
1992 
1993 		maybe_zap_len = total_bytes_to_map -  /* All bytes to map */
1994 				*length + /* Mapped or pending */
1995 				(pages_remaining * PAGE_SIZE); /* Failed map. */
1996 		zap_page_range_single(vma, *address, maybe_zap_len, NULL);
1997 		err = 0;
1998 	}
1999 
2000 	if (!err) {
2001 		unsigned long leftover_pages = pages_remaining;
2002 		int bytes_mapped;
2003 
2004 		/* We called zap_page_range_single, try to reinsert. */
2005 		err = vm_insert_pages(vma, *address,
2006 				      pending_pages,
2007 				      &pages_remaining);
2008 		bytes_mapped = PAGE_SIZE * (leftover_pages - pages_remaining);
2009 		*seq += bytes_mapped;
2010 		*address += bytes_mapped;
2011 	}
2012 	if (err) {
2013 		/* Either we were unable to zap, OR we zapped, retried an
2014 		 * insert, and still had an issue. Either ways, pages_remaining
2015 		 * is the number of pages we were unable to map, and we unroll
2016 		 * some state we speculatively touched before.
2017 		 */
2018 		const int bytes_not_mapped = PAGE_SIZE * pages_remaining;
2019 
2020 		*length -= bytes_not_mapped;
2021 		zc->recv_skip_hint += bytes_not_mapped;
2022 	}
2023 	return err;
2024 }
2025 
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)2026 static int tcp_zerocopy_vm_insert_batch(struct vm_area_struct *vma,
2027 					struct page **pages,
2028 					unsigned int pages_to_map,
2029 					unsigned long *address,
2030 					u32 *length,
2031 					u32 *seq,
2032 					struct tcp_zerocopy_receive *zc,
2033 					u32 total_bytes_to_map)
2034 {
2035 	unsigned long pages_remaining = pages_to_map;
2036 	unsigned int pages_mapped;
2037 	unsigned int bytes_mapped;
2038 	int err;
2039 
2040 	err = vm_insert_pages(vma, *address, pages, &pages_remaining);
2041 	pages_mapped = pages_to_map - (unsigned int)pages_remaining;
2042 	bytes_mapped = PAGE_SIZE * pages_mapped;
2043 	/* Even if vm_insert_pages fails, it may have partially succeeded in
2044 	 * mapping (some but not all of the pages).
2045 	 */
2046 	*seq += bytes_mapped;
2047 	*address += bytes_mapped;
2048 
2049 	if (likely(!err))
2050 		return 0;
2051 
2052 	/* Error: maybe zap and retry + rollback state for failed inserts. */
2053 	return tcp_zerocopy_vm_insert_batch_error(vma, pages + pages_mapped,
2054 		pages_remaining, address, length, seq, zc, total_bytes_to_map,
2055 		err);
2056 }
2057 
2058 #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)2059 static void tcp_zc_finalize_rx_tstamp(struct sock *sk,
2060 				      struct tcp_zerocopy_receive *zc,
2061 				      struct scm_timestamping_internal *tss)
2062 {
2063 	unsigned long msg_control_addr;
2064 	struct msghdr cmsg_dummy;
2065 
2066 	msg_control_addr = (unsigned long)zc->msg_control;
2067 	cmsg_dummy.msg_control_user = (void __user *)msg_control_addr;
2068 	cmsg_dummy.msg_controllen =
2069 		(__kernel_size_t)zc->msg_controllen;
2070 	cmsg_dummy.msg_flags = in_compat_syscall()
2071 		? MSG_CMSG_COMPAT : 0;
2072 	cmsg_dummy.msg_control_is_user = true;
2073 	zc->msg_flags = 0;
2074 	if (zc->msg_control == msg_control_addr &&
2075 	    zc->msg_controllen == cmsg_dummy.msg_controllen) {
2076 		tcp_recv_timestamp(&cmsg_dummy, sk, tss);
2077 		zc->msg_control = (__u64)
2078 			((uintptr_t)cmsg_dummy.msg_control_user);
2079 		zc->msg_controllen =
2080 			(__u64)cmsg_dummy.msg_controllen;
2081 		zc->msg_flags = (__u32)cmsg_dummy.msg_flags;
2082 	}
2083 }
2084 
find_tcp_vma(struct mm_struct * mm,unsigned long address,bool * mmap_locked)2085 static struct vm_area_struct *find_tcp_vma(struct mm_struct *mm,
2086 					   unsigned long address,
2087 					   bool *mmap_locked)
2088 {
2089 	struct vm_area_struct *vma = lock_vma_under_rcu(mm, address);
2090 
2091 	if (vma) {
2092 		if (vma->vm_ops != &tcp_vm_ops) {
2093 			vma_end_read(vma);
2094 			return NULL;
2095 		}
2096 		*mmap_locked = false;
2097 		return vma;
2098 	}
2099 
2100 	mmap_read_lock(mm);
2101 	vma = vma_lookup(mm, address);
2102 	if (!vma || vma->vm_ops != &tcp_vm_ops) {
2103 		mmap_read_unlock(mm);
2104 		return NULL;
2105 	}
2106 	*mmap_locked = true;
2107 	return vma;
2108 }
2109 
2110 #define TCP_ZEROCOPY_PAGE_BATCH_SIZE 32
tcp_zerocopy_receive(struct sock * sk,struct tcp_zerocopy_receive * zc,struct scm_timestamping_internal * tss)2111 static int tcp_zerocopy_receive(struct sock *sk,
2112 				struct tcp_zerocopy_receive *zc,
2113 				struct scm_timestamping_internal *tss)
2114 {
2115 	u32 length = 0, offset, vma_len, avail_len, copylen = 0;
2116 	unsigned long address = (unsigned long)zc->address;
2117 	struct page *pages[TCP_ZEROCOPY_PAGE_BATCH_SIZE];
2118 	s32 copybuf_len = zc->copybuf_len;
2119 	struct tcp_sock *tp = tcp_sk(sk);
2120 	const skb_frag_t *frags = NULL;
2121 	unsigned int pages_to_map = 0;
2122 	struct vm_area_struct *vma;
2123 	struct sk_buff *skb = NULL;
2124 	u32 seq = tp->copied_seq;
2125 	u32 total_bytes_to_map;
2126 	int inq = tcp_inq(sk);
2127 	bool mmap_locked;
2128 	int ret;
2129 
2130 	zc->copybuf_len = 0;
2131 	zc->msg_flags = 0;
2132 
2133 	if (address & (PAGE_SIZE - 1) || address != zc->address)
2134 		return -EINVAL;
2135 
2136 	if (sk->sk_state == TCP_LISTEN)
2137 		return -ENOTCONN;
2138 
2139 	sock_rps_record_flow(sk);
2140 
2141 	if (inq && inq <= copybuf_len)
2142 		return receive_fallback_to_copy(sk, zc, inq, tss);
2143 
2144 	if (inq < PAGE_SIZE) {
2145 		zc->length = 0;
2146 		zc->recv_skip_hint = inq;
2147 		if (!inq && sock_flag(sk, SOCK_DONE))
2148 			return -EIO;
2149 		return 0;
2150 	}
2151 
2152 	vma = find_tcp_vma(current->mm, address, &mmap_locked);
2153 	if (!vma)
2154 		return -EINVAL;
2155 
2156 	vma_len = min_t(unsigned long, zc->length, vma->vm_end - address);
2157 	avail_len = min_t(u32, vma_len, inq);
2158 	total_bytes_to_map = avail_len & ~(PAGE_SIZE - 1);
2159 	if (total_bytes_to_map) {
2160 		if (!(zc->flags & TCP_RECEIVE_ZEROCOPY_FLAG_TLB_CLEAN_HINT))
2161 			zap_page_range_single(vma, address, total_bytes_to_map,
2162 					      NULL);
2163 		zc->length = total_bytes_to_map;
2164 		zc->recv_skip_hint = 0;
2165 	} else {
2166 		zc->length = avail_len;
2167 		zc->recv_skip_hint = avail_len;
2168 	}
2169 	ret = 0;
2170 	while (length + PAGE_SIZE <= zc->length) {
2171 		int mappable_offset;
2172 		struct page *page;
2173 
2174 		if (zc->recv_skip_hint < PAGE_SIZE) {
2175 			u32 offset_frag;
2176 
2177 			if (skb) {
2178 				if (zc->recv_skip_hint > 0)
2179 					break;
2180 				skb = skb->next;
2181 				offset = seq - TCP_SKB_CB(skb)->seq;
2182 			} else {
2183 				skb = tcp_recv_skb(sk, seq, &offset);
2184 			}
2185 
2186 			if (!skb_frags_readable(skb))
2187 				break;
2188 
2189 			if (TCP_SKB_CB(skb)->has_rxtstamp) {
2190 				tcp_update_recv_tstamps(skb, tss);
2191 				zc->msg_flags |= TCP_CMSG_TS;
2192 			}
2193 			zc->recv_skip_hint = skb->len - offset;
2194 			frags = skb_advance_to_frag(skb, offset, &offset_frag);
2195 			if (!frags || offset_frag)
2196 				break;
2197 		}
2198 
2199 		mappable_offset = find_next_mappable_frag(frags,
2200 							  zc->recv_skip_hint);
2201 		if (mappable_offset) {
2202 			zc->recv_skip_hint = mappable_offset;
2203 			break;
2204 		}
2205 		page = skb_frag_page(frags);
2206 		if (WARN_ON_ONCE(!page))
2207 			break;
2208 
2209 		prefetchw(page);
2210 		pages[pages_to_map++] = page;
2211 		length += PAGE_SIZE;
2212 		zc->recv_skip_hint -= PAGE_SIZE;
2213 		frags++;
2214 		if (pages_to_map == TCP_ZEROCOPY_PAGE_BATCH_SIZE ||
2215 		    zc->recv_skip_hint < PAGE_SIZE) {
2216 			/* Either full batch, or we're about to go to next skb
2217 			 * (and we cannot unroll failed ops across skbs).
2218 			 */
2219 			ret = tcp_zerocopy_vm_insert_batch(vma, pages,
2220 							   pages_to_map,
2221 							   &address, &length,
2222 							   &seq, zc,
2223 							   total_bytes_to_map);
2224 			if (ret)
2225 				goto out;
2226 			pages_to_map = 0;
2227 		}
2228 	}
2229 	if (pages_to_map) {
2230 		ret = tcp_zerocopy_vm_insert_batch(vma, pages, pages_to_map,
2231 						   &address, &length, &seq,
2232 						   zc, total_bytes_to_map);
2233 	}
2234 out:
2235 	if (mmap_locked)
2236 		mmap_read_unlock(current->mm);
2237 	else
2238 		vma_end_read(vma);
2239 	/* Try to copy straggler data. */
2240 	if (!ret)
2241 		copylen = tcp_zc_handle_leftover(zc, sk, skb, &seq, copybuf_len, tss);
2242 
2243 	if (length + copylen) {
2244 		WRITE_ONCE(tp->copied_seq, seq);
2245 		tcp_rcv_space_adjust(sk);
2246 
2247 		/* Clean up data we have read: This will do ACK frames. */
2248 		tcp_recv_skb(sk, seq, &offset);
2249 		tcp_cleanup_rbuf(sk, length + copylen);
2250 		ret = 0;
2251 		if (length == zc->length)
2252 			zc->recv_skip_hint = 0;
2253 	} else {
2254 		if (!zc->recv_skip_hint && sock_flag(sk, SOCK_DONE))
2255 			ret = -EIO;
2256 	}
2257 	zc->length = length;
2258 	return ret;
2259 }
2260 #endif
2261 
2262 /* 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)2263 void tcp_recv_timestamp(struct msghdr *msg, const struct sock *sk,
2264 			struct scm_timestamping_internal *tss)
2265 {
2266 	int new_tstamp = sock_flag(sk, SOCK_TSTAMP_NEW);
2267 	u32 tsflags = READ_ONCE(sk->sk_tsflags);
2268 	bool has_timestamping = false;
2269 
2270 	if (tss->ts[0].tv_sec || tss->ts[0].tv_nsec) {
2271 		if (sock_flag(sk, SOCK_RCVTSTAMP)) {
2272 			if (sock_flag(sk, SOCK_RCVTSTAMPNS)) {
2273 				if (new_tstamp) {
2274 					struct __kernel_timespec kts = {
2275 						.tv_sec = tss->ts[0].tv_sec,
2276 						.tv_nsec = tss->ts[0].tv_nsec,
2277 					};
2278 					put_cmsg(msg, SOL_SOCKET, SO_TIMESTAMPNS_NEW,
2279 						 sizeof(kts), &kts);
2280 				} else {
2281 					struct __kernel_old_timespec ts_old = {
2282 						.tv_sec = tss->ts[0].tv_sec,
2283 						.tv_nsec = tss->ts[0].tv_nsec,
2284 					};
2285 					put_cmsg(msg, SOL_SOCKET, SO_TIMESTAMPNS_OLD,
2286 						 sizeof(ts_old), &ts_old);
2287 				}
2288 			} else {
2289 				if (new_tstamp) {
2290 					struct __kernel_sock_timeval stv = {
2291 						.tv_sec = tss->ts[0].tv_sec,
2292 						.tv_usec = tss->ts[0].tv_nsec / 1000,
2293 					};
2294 					put_cmsg(msg, SOL_SOCKET, SO_TIMESTAMP_NEW,
2295 						 sizeof(stv), &stv);
2296 				} else {
2297 					struct __kernel_old_timeval tv = {
2298 						.tv_sec = tss->ts[0].tv_sec,
2299 						.tv_usec = tss->ts[0].tv_nsec / 1000,
2300 					};
2301 					put_cmsg(msg, SOL_SOCKET, SO_TIMESTAMP_OLD,
2302 						 sizeof(tv), &tv);
2303 				}
2304 			}
2305 		}
2306 
2307 		if (tsflags & SOF_TIMESTAMPING_SOFTWARE &&
2308 		    (tsflags & SOF_TIMESTAMPING_RX_SOFTWARE ||
2309 		     !(tsflags & SOF_TIMESTAMPING_OPT_RX_FILTER)))
2310 			has_timestamping = true;
2311 		else
2312 			tss->ts[0] = (struct timespec64) {0};
2313 	}
2314 
2315 	if (tss->ts[2].tv_sec || tss->ts[2].tv_nsec) {
2316 		if (tsflags & SOF_TIMESTAMPING_RAW_HARDWARE &&
2317 		    (tsflags & SOF_TIMESTAMPING_RX_HARDWARE ||
2318 		     !(tsflags & SOF_TIMESTAMPING_OPT_RX_FILTER)))
2319 			has_timestamping = true;
2320 		else
2321 			tss->ts[2] = (struct timespec64) {0};
2322 	}
2323 
2324 	if (has_timestamping) {
2325 		tss->ts[1] = (struct timespec64) {0};
2326 		if (sock_flag(sk, SOCK_TSTAMP_NEW))
2327 			put_cmsg_scm_timestamping64(msg, tss);
2328 		else
2329 			put_cmsg_scm_timestamping(msg, tss);
2330 	}
2331 }
2332 
tcp_inq_hint(struct sock * sk)2333 static int tcp_inq_hint(struct sock *sk)
2334 {
2335 	const struct tcp_sock *tp = tcp_sk(sk);
2336 	u32 copied_seq = READ_ONCE(tp->copied_seq);
2337 	u32 rcv_nxt = READ_ONCE(tp->rcv_nxt);
2338 	int inq;
2339 
2340 	inq = rcv_nxt - copied_seq;
2341 	if (unlikely(inq < 0 || copied_seq != READ_ONCE(tp->copied_seq))) {
2342 		lock_sock(sk);
2343 		inq = tp->rcv_nxt - tp->copied_seq;
2344 		release_sock(sk);
2345 	}
2346 	/* After receiving a FIN, tell the user-space to continue reading
2347 	 * by returning a non-zero inq.
2348 	 */
2349 	if (inq == 0 && sock_flag(sk, SOCK_DONE))
2350 		inq = 1;
2351 	return inq;
2352 }
2353 
2354 /* batch __xa_alloc() calls and reduce xa_lock()/xa_unlock() overhead. */
2355 struct tcp_xa_pool {
2356 	u8		max; /* max <= MAX_SKB_FRAGS */
2357 	u8		idx; /* idx <= max */
2358 	__u32		tokens[MAX_SKB_FRAGS];
2359 	netmem_ref	netmems[MAX_SKB_FRAGS];
2360 };
2361 
tcp_xa_pool_commit_locked(struct sock * sk,struct tcp_xa_pool * p)2362 static void tcp_xa_pool_commit_locked(struct sock *sk, struct tcp_xa_pool *p)
2363 {
2364 	int i;
2365 
2366 	/* Commit part that has been copied to user space. */
2367 	for (i = 0; i < p->idx; i++)
2368 		__xa_cmpxchg(&sk->sk_user_frags, p->tokens[i], XA_ZERO_ENTRY,
2369 			     (__force void *)p->netmems[i], GFP_KERNEL);
2370 	/* Rollback what has been pre-allocated and is no longer needed. */
2371 	for (; i < p->max; i++)
2372 		__xa_erase(&sk->sk_user_frags, p->tokens[i]);
2373 
2374 	p->max = 0;
2375 	p->idx = 0;
2376 }
2377 
tcp_xa_pool_commit(struct sock * sk,struct tcp_xa_pool * p)2378 static void tcp_xa_pool_commit(struct sock *sk, struct tcp_xa_pool *p)
2379 {
2380 	if (!p->max)
2381 		return;
2382 
2383 	xa_lock_bh(&sk->sk_user_frags);
2384 
2385 	tcp_xa_pool_commit_locked(sk, p);
2386 
2387 	xa_unlock_bh(&sk->sk_user_frags);
2388 }
2389 
tcp_xa_pool_refill(struct sock * sk,struct tcp_xa_pool * p,unsigned int max_frags)2390 static int tcp_xa_pool_refill(struct sock *sk, struct tcp_xa_pool *p,
2391 			      unsigned int max_frags)
2392 {
2393 	int err, k;
2394 
2395 	if (p->idx < p->max)
2396 		return 0;
2397 
2398 	xa_lock_bh(&sk->sk_user_frags);
2399 
2400 	tcp_xa_pool_commit_locked(sk, p);
2401 
2402 	for (k = 0; k < max_frags; k++) {
2403 		err = __xa_alloc(&sk->sk_user_frags, &p->tokens[k],
2404 				 XA_ZERO_ENTRY, xa_limit_31b, GFP_KERNEL);
2405 		if (err)
2406 			break;
2407 	}
2408 
2409 	xa_unlock_bh(&sk->sk_user_frags);
2410 
2411 	p->max = k;
2412 	p->idx = 0;
2413 	return k ? 0 : err;
2414 }
2415 
2416 /* On error, returns the -errno. On success, returns number of bytes sent to the
2417  * user. May not consume all of @remaining_len.
2418  */
tcp_recvmsg_dmabuf(struct sock * sk,const struct sk_buff * skb,unsigned int offset,struct msghdr * msg,int remaining_len)2419 static int tcp_recvmsg_dmabuf(struct sock *sk, const struct sk_buff *skb,
2420 			      unsigned int offset, struct msghdr *msg,
2421 			      int remaining_len)
2422 {
2423 	struct dmabuf_cmsg dmabuf_cmsg = { 0 };
2424 	struct tcp_xa_pool tcp_xa_pool;
2425 	unsigned int start;
2426 	int i, copy, n;
2427 	int sent = 0;
2428 	int err = 0;
2429 
2430 	tcp_xa_pool.max = 0;
2431 	tcp_xa_pool.idx = 0;
2432 	do {
2433 		start = skb_headlen(skb);
2434 
2435 		if (skb_frags_readable(skb)) {
2436 			err = -ENODEV;
2437 			goto out;
2438 		}
2439 
2440 		/* Copy header. */
2441 		copy = start - offset;
2442 		if (copy > 0) {
2443 			copy = min(copy, remaining_len);
2444 
2445 			n = copy_to_iter(skb->data + offset, copy,
2446 					 &msg->msg_iter);
2447 			if (n != copy) {
2448 				err = -EFAULT;
2449 				goto out;
2450 			}
2451 
2452 			offset += copy;
2453 			remaining_len -= copy;
2454 
2455 			/* First a dmabuf_cmsg for # bytes copied to user
2456 			 * buffer.
2457 			 */
2458 			memset(&dmabuf_cmsg, 0, sizeof(dmabuf_cmsg));
2459 			dmabuf_cmsg.frag_size = copy;
2460 			err = put_cmsg_notrunc(msg, SOL_SOCKET,
2461 					       SO_DEVMEM_LINEAR,
2462 					       sizeof(dmabuf_cmsg),
2463 					       &dmabuf_cmsg);
2464 			if (err)
2465 				goto out;
2466 
2467 			sent += copy;
2468 
2469 			if (remaining_len == 0)
2470 				goto out;
2471 		}
2472 
2473 		/* after that, send information of dmabuf pages through a
2474 		 * sequence of cmsg
2475 		 */
2476 		for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
2477 			skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
2478 			struct net_iov *niov;
2479 			u64 frag_offset;
2480 			int end;
2481 
2482 			/* !skb_frags_readable() should indicate that ALL the
2483 			 * frags in this skb are dmabuf net_iovs. We're checking
2484 			 * for that flag above, but also check individual frags
2485 			 * here. If the tcp stack is not setting
2486 			 * skb_frags_readable() correctly, we still don't want
2487 			 * to crash here.
2488 			 */
2489 			if (!skb_frag_net_iov(frag)) {
2490 				net_err_ratelimited("Found non-dmabuf skb with net_iov");
2491 				err = -ENODEV;
2492 				goto out;
2493 			}
2494 
2495 			niov = skb_frag_net_iov(frag);
2496 			end = start + skb_frag_size(frag);
2497 			copy = end - offset;
2498 
2499 			if (copy > 0) {
2500 				copy = min(copy, remaining_len);
2501 
2502 				frag_offset = net_iov_virtual_addr(niov) +
2503 					      skb_frag_off(frag) + offset -
2504 					      start;
2505 				dmabuf_cmsg.frag_offset = frag_offset;
2506 				dmabuf_cmsg.frag_size = copy;
2507 				err = tcp_xa_pool_refill(sk, &tcp_xa_pool,
2508 							 skb_shinfo(skb)->nr_frags - i);
2509 				if (err)
2510 					goto out;
2511 
2512 				/* Will perform the exchange later */
2513 				dmabuf_cmsg.frag_token = tcp_xa_pool.tokens[tcp_xa_pool.idx];
2514 				dmabuf_cmsg.dmabuf_id = net_iov_binding_id(niov);
2515 
2516 				offset += copy;
2517 				remaining_len -= copy;
2518 
2519 				err = put_cmsg_notrunc(msg, SOL_SOCKET,
2520 						       SO_DEVMEM_DMABUF,
2521 						       sizeof(dmabuf_cmsg),
2522 						       &dmabuf_cmsg);
2523 				if (err)
2524 					goto out;
2525 
2526 				atomic_long_inc(&niov->pp_ref_count);
2527 				tcp_xa_pool.netmems[tcp_xa_pool.idx++] = skb_frag_netmem(frag);
2528 
2529 				sent += copy;
2530 
2531 				if (remaining_len == 0)
2532 					goto out;
2533 			}
2534 			start = end;
2535 		}
2536 
2537 		tcp_xa_pool_commit(sk, &tcp_xa_pool);
2538 		if (!remaining_len)
2539 			goto out;
2540 
2541 		/* if remaining_len is not satisfied yet, we need to go to the
2542 		 * next frag in the frag_list to satisfy remaining_len.
2543 		 */
2544 		skb = skb_shinfo(skb)->frag_list ?: skb->next;
2545 
2546 		offset = offset - start;
2547 	} while (skb);
2548 
2549 	if (remaining_len) {
2550 		err = -EFAULT;
2551 		goto out;
2552 	}
2553 
2554 out:
2555 	tcp_xa_pool_commit(sk, &tcp_xa_pool);
2556 	if (!sent)
2557 		sent = err;
2558 
2559 	return sent;
2560 }
2561 
2562 /*
2563  *	This routine copies from a sock struct into the user buffer.
2564  *
2565  *	Technical note: in 2.3 we work on _locked_ socket, so that
2566  *	tricks with *seq access order and skb->users are not required.
2567  *	Probably, code can be easily improved even more.
2568  */
2569 
tcp_recvmsg_locked(struct sock * sk,struct msghdr * msg,size_t len,int flags,struct scm_timestamping_internal * tss,int * cmsg_flags)2570 static int tcp_recvmsg_locked(struct sock *sk, struct msghdr *msg, size_t len,
2571 			      int flags, struct scm_timestamping_internal *tss,
2572 			      int *cmsg_flags)
2573 {
2574 	struct tcp_sock *tp = tcp_sk(sk);
2575 	int last_copied_dmabuf = -1; /* uninitialized */
2576 	int copied = 0;
2577 	u32 peek_seq;
2578 	u32 *seq;
2579 	unsigned long used;
2580 	int err;
2581 	int target;		/* Read at least this many bytes */
2582 	long timeo;
2583 	struct sk_buff *skb, *last;
2584 	u32 peek_offset = 0;
2585 	u32 urg_hole = 0;
2586 
2587 	err = -ENOTCONN;
2588 	if (sk->sk_state == TCP_LISTEN)
2589 		goto out;
2590 
2591 	if (tp->recvmsg_inq) {
2592 		*cmsg_flags = TCP_CMSG_INQ;
2593 		msg->msg_get_inq = 1;
2594 	}
2595 	timeo = sock_rcvtimeo(sk, flags & MSG_DONTWAIT);
2596 
2597 	/* Urgent data needs to be handled specially. */
2598 	if (flags & MSG_OOB)
2599 		goto recv_urg;
2600 
2601 	if (unlikely(tp->repair)) {
2602 		err = -EPERM;
2603 		if (!(flags & MSG_PEEK))
2604 			goto out;
2605 
2606 		if (tp->repair_queue == TCP_SEND_QUEUE)
2607 			goto recv_sndq;
2608 
2609 		err = -EINVAL;
2610 		if (tp->repair_queue == TCP_NO_QUEUE)
2611 			goto out;
2612 
2613 		/* 'common' recv queue MSG_PEEK-ing */
2614 	}
2615 
2616 	seq = &tp->copied_seq;
2617 	if (flags & MSG_PEEK) {
2618 		peek_offset = max(sk_peek_offset(sk, flags), 0);
2619 		peek_seq = tp->copied_seq + peek_offset;
2620 		seq = &peek_seq;
2621 	}
2622 
2623 	target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
2624 
2625 	do {
2626 		u32 offset;
2627 
2628 		/* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
2629 		if (unlikely(tp->urg_data) && tp->urg_seq == *seq) {
2630 			if (copied)
2631 				break;
2632 			if (signal_pending(current)) {
2633 				copied = timeo ? sock_intr_errno(timeo) : -EAGAIN;
2634 				break;
2635 			}
2636 		}
2637 
2638 		/* Next get a buffer. */
2639 
2640 		last = skb_peek_tail(&sk->sk_receive_queue);
2641 		skb_queue_walk(&sk->sk_receive_queue, skb) {
2642 			last = skb;
2643 			/* Now that we have two receive queues this
2644 			 * shouldn't happen.
2645 			 */
2646 			if (WARN(before(*seq, TCP_SKB_CB(skb)->seq),
2647 				 "TCP recvmsg seq # bug: copied %X, seq %X, rcvnxt %X, fl %X\n",
2648 				 *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt,
2649 				 flags))
2650 				break;
2651 
2652 			offset = *seq - TCP_SKB_CB(skb)->seq;
2653 			if (unlikely(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
2654 				pr_err_once("%s: found a SYN, please report !\n", __func__);
2655 				offset--;
2656 			}
2657 			if (offset < skb->len)
2658 				goto found_ok_skb;
2659 			if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
2660 				goto found_fin_ok;
2661 			WARN(!(flags & MSG_PEEK),
2662 			     "TCP recvmsg seq # bug 2: copied %X, seq %X, rcvnxt %X, fl %X\n",
2663 			     *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt, flags);
2664 		}
2665 
2666 		/* Well, if we have backlog, try to process it now yet. */
2667 
2668 		if (copied >= target && !READ_ONCE(sk->sk_backlog.tail))
2669 			break;
2670 
2671 		if (copied) {
2672 			if (!timeo ||
2673 			    sk->sk_err ||
2674 			    sk->sk_state == TCP_CLOSE ||
2675 			    (sk->sk_shutdown & RCV_SHUTDOWN) ||
2676 			    signal_pending(current))
2677 				break;
2678 		} else {
2679 			if (sock_flag(sk, SOCK_DONE))
2680 				break;
2681 
2682 			if (sk->sk_err) {
2683 				copied = sock_error(sk);
2684 				break;
2685 			}
2686 
2687 			if (sk->sk_shutdown & RCV_SHUTDOWN)
2688 				break;
2689 
2690 			if (sk->sk_state == TCP_CLOSE) {
2691 				/* This occurs when user tries to read
2692 				 * from never connected socket.
2693 				 */
2694 				copied = -ENOTCONN;
2695 				break;
2696 			}
2697 
2698 			if (!timeo) {
2699 				copied = -EAGAIN;
2700 				break;
2701 			}
2702 
2703 			if (signal_pending(current)) {
2704 				copied = sock_intr_errno(timeo);
2705 				break;
2706 			}
2707 		}
2708 
2709 		if (copied >= target) {
2710 			/* Do not sleep, just process backlog. */
2711 			__sk_flush_backlog(sk);
2712 		} else {
2713 			tcp_cleanup_rbuf(sk, copied);
2714 			err = sk_wait_data(sk, &timeo, last);
2715 			if (err < 0) {
2716 				err = copied ? : err;
2717 				goto out;
2718 			}
2719 		}
2720 
2721 		if ((flags & MSG_PEEK) &&
2722 		    (peek_seq - peek_offset - copied - urg_hole != tp->copied_seq)) {
2723 			net_dbg_ratelimited("TCP(%s:%d): Application bug, race in MSG_PEEK\n",
2724 					    current->comm,
2725 					    task_pid_nr(current));
2726 			peek_seq = tp->copied_seq + peek_offset;
2727 		}
2728 		continue;
2729 
2730 found_ok_skb:
2731 		/* Ok so how much can we use? */
2732 		used = skb->len - offset;
2733 		if (len < used)
2734 			used = len;
2735 
2736 		/* Do we have urgent data here? */
2737 		if (unlikely(tp->urg_data)) {
2738 			u32 urg_offset = tp->urg_seq - *seq;
2739 			if (urg_offset < used) {
2740 				if (!urg_offset) {
2741 					if (!sock_flag(sk, SOCK_URGINLINE)) {
2742 						WRITE_ONCE(*seq, *seq + 1);
2743 						urg_hole++;
2744 						offset++;
2745 						used--;
2746 						if (!used)
2747 							goto skip_copy;
2748 					}
2749 				} else
2750 					used = urg_offset;
2751 			}
2752 		}
2753 
2754 		if (!(flags & MSG_TRUNC)) {
2755 			if (last_copied_dmabuf != -1 &&
2756 			    last_copied_dmabuf != !skb_frags_readable(skb))
2757 				break;
2758 
2759 			if (skb_frags_readable(skb)) {
2760 				err = skb_copy_datagram_msg(skb, offset, msg,
2761 							    used);
2762 				if (err) {
2763 					/* Exception. Bailout! */
2764 					if (!copied)
2765 						copied = -EFAULT;
2766 					break;
2767 				}
2768 			} else {
2769 				if (!(flags & MSG_SOCK_DEVMEM)) {
2770 					/* dmabuf skbs can only be received
2771 					 * with the MSG_SOCK_DEVMEM flag.
2772 					 */
2773 					if (!copied)
2774 						copied = -EFAULT;
2775 
2776 					break;
2777 				}
2778 
2779 				err = tcp_recvmsg_dmabuf(sk, skb, offset, msg,
2780 							 used);
2781 				if (err <= 0) {
2782 					if (!copied)
2783 						copied = -EFAULT;
2784 
2785 					break;
2786 				}
2787 				used = err;
2788 			}
2789 		}
2790 
2791 		last_copied_dmabuf = !skb_frags_readable(skb);
2792 
2793 		WRITE_ONCE(*seq, *seq + used);
2794 		copied += used;
2795 		len -= used;
2796 		if (flags & MSG_PEEK)
2797 			sk_peek_offset_fwd(sk, used);
2798 		else
2799 			sk_peek_offset_bwd(sk, used);
2800 		tcp_rcv_space_adjust(sk);
2801 
2802 skip_copy:
2803 		if (unlikely(tp->urg_data) && after(tp->copied_seq, tp->urg_seq)) {
2804 			WRITE_ONCE(tp->urg_data, 0);
2805 			tcp_fast_path_check(sk);
2806 		}
2807 
2808 		if (TCP_SKB_CB(skb)->has_rxtstamp) {
2809 			tcp_update_recv_tstamps(skb, tss);
2810 			*cmsg_flags |= TCP_CMSG_TS;
2811 		}
2812 
2813 		if (used + offset < skb->len)
2814 			continue;
2815 
2816 		if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
2817 			goto found_fin_ok;
2818 		if (!(flags & MSG_PEEK))
2819 			tcp_eat_recv_skb(sk, skb);
2820 		continue;
2821 
2822 found_fin_ok:
2823 		/* Process the FIN. */
2824 		WRITE_ONCE(*seq, *seq + 1);
2825 		if (!(flags & MSG_PEEK))
2826 			tcp_eat_recv_skb(sk, skb);
2827 		break;
2828 	} while (len > 0);
2829 
2830 	/* According to UNIX98, msg_name/msg_namelen are ignored
2831 	 * on connected socket. I was just happy when found this 8) --ANK
2832 	 */
2833 
2834 	/* Clean up data we have read: This will do ACK frames. */
2835 	tcp_cleanup_rbuf(sk, copied);
2836 	return copied;
2837 
2838 out:
2839 	return err;
2840 
2841 recv_urg:
2842 	err = tcp_recv_urg(sk, msg, len, flags);
2843 	goto out;
2844 
2845 recv_sndq:
2846 	err = tcp_peek_sndq(sk, msg, len);
2847 	goto out;
2848 }
2849 
tcp_recvmsg(struct sock * sk,struct msghdr * msg,size_t len,int flags,int * addr_len)2850 int tcp_recvmsg(struct sock *sk, struct msghdr *msg, size_t len, int flags,
2851 		int *addr_len)
2852 {
2853 	int cmsg_flags = 0, ret;
2854 	struct scm_timestamping_internal tss;
2855 
2856 	if (unlikely(flags & MSG_ERRQUEUE))
2857 		return inet_recv_error(sk, msg, len, addr_len);
2858 
2859 	if (sk_can_busy_loop(sk) &&
2860 	    skb_queue_empty_lockless(&sk->sk_receive_queue) &&
2861 	    sk->sk_state == TCP_ESTABLISHED)
2862 		sk_busy_loop(sk, flags & MSG_DONTWAIT);
2863 
2864 	lock_sock(sk);
2865 	ret = tcp_recvmsg_locked(sk, msg, len, flags, &tss, &cmsg_flags);
2866 	release_sock(sk);
2867 
2868 	if ((cmsg_flags || msg->msg_get_inq) && ret >= 0) {
2869 		if (cmsg_flags & TCP_CMSG_TS)
2870 			tcp_recv_timestamp(msg, sk, &tss);
2871 		if (msg->msg_get_inq) {
2872 			msg->msg_inq = tcp_inq_hint(sk);
2873 			if (cmsg_flags & TCP_CMSG_INQ)
2874 				put_cmsg(msg, SOL_TCP, TCP_CM_INQ,
2875 					 sizeof(msg->msg_inq), &msg->msg_inq);
2876 		}
2877 	}
2878 	return ret;
2879 }
2880 EXPORT_SYMBOL(tcp_recvmsg);
2881 
tcp_set_state(struct sock * sk,int state)2882 void tcp_set_state(struct sock *sk, int state)
2883 {
2884 	int oldstate = sk->sk_state;
2885 
2886 	/* We defined a new enum for TCP states that are exported in BPF
2887 	 * so as not force the internal TCP states to be frozen. The
2888 	 * following checks will detect if an internal state value ever
2889 	 * differs from the BPF value. If this ever happens, then we will
2890 	 * need to remap the internal value to the BPF value before calling
2891 	 * tcp_call_bpf_2arg.
2892 	 */
2893 	BUILD_BUG_ON((int)BPF_TCP_ESTABLISHED != (int)TCP_ESTABLISHED);
2894 	BUILD_BUG_ON((int)BPF_TCP_SYN_SENT != (int)TCP_SYN_SENT);
2895 	BUILD_BUG_ON((int)BPF_TCP_SYN_RECV != (int)TCP_SYN_RECV);
2896 	BUILD_BUG_ON((int)BPF_TCP_FIN_WAIT1 != (int)TCP_FIN_WAIT1);
2897 	BUILD_BUG_ON((int)BPF_TCP_FIN_WAIT2 != (int)TCP_FIN_WAIT2);
2898 	BUILD_BUG_ON((int)BPF_TCP_TIME_WAIT != (int)TCP_TIME_WAIT);
2899 	BUILD_BUG_ON((int)BPF_TCP_CLOSE != (int)TCP_CLOSE);
2900 	BUILD_BUG_ON((int)BPF_TCP_CLOSE_WAIT != (int)TCP_CLOSE_WAIT);
2901 	BUILD_BUG_ON((int)BPF_TCP_LAST_ACK != (int)TCP_LAST_ACK);
2902 	BUILD_BUG_ON((int)BPF_TCP_LISTEN != (int)TCP_LISTEN);
2903 	BUILD_BUG_ON((int)BPF_TCP_CLOSING != (int)TCP_CLOSING);
2904 	BUILD_BUG_ON((int)BPF_TCP_NEW_SYN_RECV != (int)TCP_NEW_SYN_RECV);
2905 	BUILD_BUG_ON((int)BPF_TCP_BOUND_INACTIVE != (int)TCP_BOUND_INACTIVE);
2906 	BUILD_BUG_ON((int)BPF_TCP_MAX_STATES != (int)TCP_MAX_STATES);
2907 
2908 	/* bpf uapi header bpf.h defines an anonymous enum with values
2909 	 * BPF_TCP_* used by bpf programs. Currently gcc built vmlinux
2910 	 * is able to emit this enum in DWARF due to the above BUILD_BUG_ON.
2911 	 * But clang built vmlinux does not have this enum in DWARF
2912 	 * since clang removes the above code before generating IR/debuginfo.
2913 	 * Let us explicitly emit the type debuginfo to ensure the
2914 	 * above-mentioned anonymous enum in the vmlinux DWARF and hence BTF
2915 	 * regardless of which compiler is used.
2916 	 */
2917 	BTF_TYPE_EMIT_ENUM(BPF_TCP_ESTABLISHED);
2918 
2919 	if (BPF_SOCK_OPS_TEST_FLAG(tcp_sk(sk), BPF_SOCK_OPS_STATE_CB_FLAG))
2920 		tcp_call_bpf_2arg(sk, BPF_SOCK_OPS_STATE_CB, oldstate, state);
2921 
2922 	switch (state) {
2923 	case TCP_ESTABLISHED:
2924 		if (oldstate != TCP_ESTABLISHED)
2925 			TCP_INC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
2926 		break;
2927 	case TCP_CLOSE_WAIT:
2928 		if (oldstate == TCP_SYN_RECV)
2929 			TCP_INC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
2930 		break;
2931 
2932 	case TCP_CLOSE:
2933 		if (oldstate == TCP_CLOSE_WAIT || oldstate == TCP_ESTABLISHED)
2934 			TCP_INC_STATS(sock_net(sk), TCP_MIB_ESTABRESETS);
2935 
2936 		sk->sk_prot->unhash(sk);
2937 		if (inet_csk(sk)->icsk_bind_hash &&
2938 		    !(sk->sk_userlocks & SOCK_BINDPORT_LOCK))
2939 			inet_put_port(sk);
2940 		fallthrough;
2941 	default:
2942 		if (oldstate == TCP_ESTABLISHED || oldstate == TCP_CLOSE_WAIT)
2943 			TCP_DEC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
2944 	}
2945 
2946 	/* Change state AFTER socket is unhashed to avoid closed
2947 	 * socket sitting in hash tables.
2948 	 */
2949 	inet_sk_state_store(sk, state);
2950 }
2951 EXPORT_SYMBOL_GPL(tcp_set_state);
2952 
2953 /*
2954  *	State processing on a close. This implements the state shift for
2955  *	sending our FIN frame. Note that we only send a FIN for some
2956  *	states. A shutdown() may have already sent the FIN, or we may be
2957  *	closed.
2958  */
2959 
2960 static const unsigned char new_state[16] = {
2961   /* current state:        new state:      action:	*/
2962   [0 /* (Invalid) */]	= TCP_CLOSE,
2963   [TCP_ESTABLISHED]	= TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2964   [TCP_SYN_SENT]	= TCP_CLOSE,
2965   [TCP_SYN_RECV]	= TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2966   [TCP_FIN_WAIT1]	= TCP_FIN_WAIT1,
2967   [TCP_FIN_WAIT2]	= TCP_FIN_WAIT2,
2968   [TCP_TIME_WAIT]	= TCP_CLOSE,
2969   [TCP_CLOSE]		= TCP_CLOSE,
2970   [TCP_CLOSE_WAIT]	= TCP_LAST_ACK  | TCP_ACTION_FIN,
2971   [TCP_LAST_ACK]	= TCP_LAST_ACK,
2972   [TCP_LISTEN]		= TCP_CLOSE,
2973   [TCP_CLOSING]		= TCP_CLOSING,
2974   [TCP_NEW_SYN_RECV]	= TCP_CLOSE,	/* should not happen ! */
2975 };
2976 
tcp_close_state(struct sock * sk)2977 static int tcp_close_state(struct sock *sk)
2978 {
2979 	int next = (int)new_state[sk->sk_state];
2980 	int ns = next & TCP_STATE_MASK;
2981 
2982 	tcp_set_state(sk, ns);
2983 
2984 	return next & TCP_ACTION_FIN;
2985 }
2986 
2987 /*
2988  *	Shutdown the sending side of a connection. Much like close except
2989  *	that we don't receive shut down or sock_set_flag(sk, SOCK_DEAD).
2990  */
2991 
tcp_shutdown(struct sock * sk,int how)2992 void tcp_shutdown(struct sock *sk, int how)
2993 {
2994 	/*	We need to grab some memory, and put together a FIN,
2995 	 *	and then put it into the queue to be sent.
2996 	 *		Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
2997 	 */
2998 	if (!(how & SEND_SHUTDOWN))
2999 		return;
3000 
3001 	/* If we've already sent a FIN, or it's a closed state, skip this. */
3002 	if ((1 << sk->sk_state) &
3003 	    (TCPF_ESTABLISHED | TCPF_SYN_SENT |
3004 	     TCPF_CLOSE_WAIT)) {
3005 		/* Clear out any half completed packets.  FIN if needed. */
3006 		if (tcp_close_state(sk))
3007 			tcp_send_fin(sk);
3008 	}
3009 }
3010 EXPORT_SYMBOL(tcp_shutdown);
3011 
tcp_orphan_count_sum(void)3012 int tcp_orphan_count_sum(void)
3013 {
3014 	int i, total = 0;
3015 
3016 	for_each_possible_cpu(i)
3017 		total += per_cpu(tcp_orphan_count, i);
3018 
3019 	return max(total, 0);
3020 }
3021 
3022 static int tcp_orphan_cache;
3023 static struct timer_list tcp_orphan_timer;
3024 #define TCP_ORPHAN_TIMER_PERIOD msecs_to_jiffies(100)
3025 
tcp_orphan_update(struct timer_list * unused)3026 static void tcp_orphan_update(struct timer_list *unused)
3027 {
3028 	WRITE_ONCE(tcp_orphan_cache, tcp_orphan_count_sum());
3029 	mod_timer(&tcp_orphan_timer, jiffies + TCP_ORPHAN_TIMER_PERIOD);
3030 }
3031 
tcp_too_many_orphans(int shift)3032 static bool tcp_too_many_orphans(int shift)
3033 {
3034 	return READ_ONCE(tcp_orphan_cache) << shift >
3035 		READ_ONCE(sysctl_tcp_max_orphans);
3036 }
3037 
tcp_out_of_memory(const struct sock * sk)3038 static bool tcp_out_of_memory(const struct sock *sk)
3039 {
3040 	if (sk->sk_wmem_queued > SOCK_MIN_SNDBUF &&
3041 	    sk_memory_allocated(sk) > sk_prot_mem_limits(sk, 2))
3042 		return true;
3043 	return false;
3044 }
3045 
tcp_check_oom(const struct sock * sk,int shift)3046 bool tcp_check_oom(const struct sock *sk, int shift)
3047 {
3048 	bool too_many_orphans, out_of_socket_memory;
3049 
3050 	too_many_orphans = tcp_too_many_orphans(shift);
3051 	out_of_socket_memory = tcp_out_of_memory(sk);
3052 
3053 	if (too_many_orphans)
3054 		net_info_ratelimited("too many orphaned sockets\n");
3055 	if (out_of_socket_memory)
3056 		net_info_ratelimited("out of memory -- consider tuning tcp_mem\n");
3057 	return too_many_orphans || out_of_socket_memory;
3058 }
3059 
__tcp_close(struct sock * sk,long timeout)3060 void __tcp_close(struct sock *sk, long timeout)
3061 {
3062 	struct sk_buff *skb;
3063 	int data_was_unread = 0;
3064 	int state;
3065 
3066 	WRITE_ONCE(sk->sk_shutdown, SHUTDOWN_MASK);
3067 
3068 	if (sk->sk_state == TCP_LISTEN) {
3069 		tcp_set_state(sk, TCP_CLOSE);
3070 
3071 		/* Special case. */
3072 		inet_csk_listen_stop(sk);
3073 
3074 		goto adjudge_to_death;
3075 	}
3076 
3077 	/*  We need to flush the recv. buffs.  We do this only on the
3078 	 *  descriptor close, not protocol-sourced closes, because the
3079 	 *  reader process may not have drained the data yet!
3080 	 */
3081 	while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) {
3082 		u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq;
3083 
3084 		if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
3085 			len--;
3086 		data_was_unread += len;
3087 		__kfree_skb(skb);
3088 	}
3089 
3090 	/* If socket has been already reset (e.g. in tcp_reset()) - kill it. */
3091 	if (sk->sk_state == TCP_CLOSE)
3092 		goto adjudge_to_death;
3093 
3094 	/* As outlined in RFC 2525, section 2.17, we send a RST here because
3095 	 * data was lost. To witness the awful effects of the old behavior of
3096 	 * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk
3097 	 * GET in an FTP client, suspend the process, wait for the client to
3098 	 * advertise a zero window, then kill -9 the FTP client, wheee...
3099 	 * Note: timeout is always zero in such a case.
3100 	 */
3101 	if (unlikely(tcp_sk(sk)->repair)) {
3102 		sk->sk_prot->disconnect(sk, 0);
3103 	} else if (data_was_unread) {
3104 		/* Unread data was tossed, zap the connection. */
3105 		NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONCLOSE);
3106 		tcp_set_state(sk, TCP_CLOSE);
3107 		tcp_send_active_reset(sk, sk->sk_allocation,
3108 				      SK_RST_REASON_TCP_ABORT_ON_CLOSE);
3109 	} else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) {
3110 		/* Check zero linger _after_ checking for unread data. */
3111 		sk->sk_prot->disconnect(sk, 0);
3112 		NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
3113 	} else if (tcp_close_state(sk)) {
3114 		/* We FIN if the application ate all the data before
3115 		 * zapping the connection.
3116 		 */
3117 
3118 		/* RED-PEN. Formally speaking, we have broken TCP state
3119 		 * machine. State transitions:
3120 		 *
3121 		 * TCP_ESTABLISHED -> TCP_FIN_WAIT1
3122 		 * TCP_SYN_RECV	-> TCP_FIN_WAIT1 (it is difficult)
3123 		 * TCP_CLOSE_WAIT -> TCP_LAST_ACK
3124 		 *
3125 		 * are legal only when FIN has been sent (i.e. in window),
3126 		 * rather than queued out of window. Purists blame.
3127 		 *
3128 		 * F.e. "RFC state" is ESTABLISHED,
3129 		 * if Linux state is FIN-WAIT-1, but FIN is still not sent.
3130 		 *
3131 		 * The visible declinations are that sometimes
3132 		 * we enter time-wait state, when it is not required really
3133 		 * (harmless), do not send active resets, when they are
3134 		 * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
3135 		 * they look as CLOSING or LAST_ACK for Linux)
3136 		 * Probably, I missed some more holelets.
3137 		 * 						--ANK
3138 		 * XXX (TFO) - To start off we don't support SYN+ACK+FIN
3139 		 * in a single packet! (May consider it later but will
3140 		 * probably need API support or TCP_CORK SYN-ACK until
3141 		 * data is written and socket is closed.)
3142 		 */
3143 		tcp_send_fin(sk);
3144 	}
3145 
3146 	sk_stream_wait_close(sk, timeout);
3147 
3148 adjudge_to_death:
3149 	state = sk->sk_state;
3150 	sock_hold(sk);
3151 	sock_orphan(sk);
3152 
3153 	local_bh_disable();
3154 	bh_lock_sock(sk);
3155 	/* remove backlog if any, without releasing ownership. */
3156 	__release_sock(sk);
3157 
3158 	this_cpu_inc(tcp_orphan_count);
3159 
3160 	/* Have we already been destroyed by a softirq or backlog? */
3161 	if (state != TCP_CLOSE && sk->sk_state == TCP_CLOSE)
3162 		goto out;
3163 
3164 	/*	This is a (useful) BSD violating of the RFC. There is a
3165 	 *	problem with TCP as specified in that the other end could
3166 	 *	keep a socket open forever with no application left this end.
3167 	 *	We use a 1 minute timeout (about the same as BSD) then kill
3168 	 *	our end. If they send after that then tough - BUT: long enough
3169 	 *	that we won't make the old 4*rto = almost no time - whoops
3170 	 *	reset mistake.
3171 	 *
3172 	 *	Nope, it was not mistake. It is really desired behaviour
3173 	 *	f.e. on http servers, when such sockets are useless, but
3174 	 *	consume significant resources. Let's do it with special
3175 	 *	linger2	option.					--ANK
3176 	 */
3177 
3178 	if (sk->sk_state == TCP_FIN_WAIT2) {
3179 		struct tcp_sock *tp = tcp_sk(sk);
3180 		if (READ_ONCE(tp->linger2) < 0) {
3181 			tcp_set_state(sk, TCP_CLOSE);
3182 			tcp_send_active_reset(sk, GFP_ATOMIC,
3183 					      SK_RST_REASON_TCP_ABORT_ON_LINGER);
3184 			__NET_INC_STATS(sock_net(sk),
3185 					LINUX_MIB_TCPABORTONLINGER);
3186 		} else {
3187 			const int tmo = tcp_fin_time(sk);
3188 
3189 			if (tmo > TCP_TIMEWAIT_LEN) {
3190 				inet_csk_reset_keepalive_timer(sk,
3191 						tmo - TCP_TIMEWAIT_LEN);
3192 			} else {
3193 				tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
3194 				goto out;
3195 			}
3196 		}
3197 	}
3198 	if (sk->sk_state != TCP_CLOSE) {
3199 		if (tcp_check_oom(sk, 0)) {
3200 			tcp_set_state(sk, TCP_CLOSE);
3201 			tcp_send_active_reset(sk, GFP_ATOMIC,
3202 					      SK_RST_REASON_TCP_ABORT_ON_MEMORY);
3203 			__NET_INC_STATS(sock_net(sk),
3204 					LINUX_MIB_TCPABORTONMEMORY);
3205 		} else if (!check_net(sock_net(sk))) {
3206 			/* Not possible to send reset; just close */
3207 			tcp_set_state(sk, TCP_CLOSE);
3208 		}
3209 	}
3210 
3211 	if (sk->sk_state == TCP_CLOSE) {
3212 		struct request_sock *req;
3213 
3214 		req = rcu_dereference_protected(tcp_sk(sk)->fastopen_rsk,
3215 						lockdep_sock_is_held(sk));
3216 		/* We could get here with a non-NULL req if the socket is
3217 		 * aborted (e.g., closed with unread data) before 3WHS
3218 		 * finishes.
3219 		 */
3220 		if (req)
3221 			reqsk_fastopen_remove(sk, req, false);
3222 		inet_csk_destroy_sock(sk);
3223 	}
3224 	/* Otherwise, socket is reprieved until protocol close. */
3225 
3226 out:
3227 	bh_unlock_sock(sk);
3228 	local_bh_enable();
3229 }
3230 
tcp_close(struct sock * sk,long timeout)3231 void tcp_close(struct sock *sk, long timeout)
3232 {
3233 	lock_sock(sk);
3234 	__tcp_close(sk, timeout);
3235 	release_sock(sk);
3236 	if (!sk->sk_net_refcnt)
3237 		inet_csk_clear_xmit_timers_sync(sk);
3238 	sock_put(sk);
3239 }
3240 EXPORT_SYMBOL(tcp_close);
3241 
3242 /* These states need RST on ABORT according to RFC793 */
3243 
tcp_need_reset(int state)3244 static inline bool tcp_need_reset(int state)
3245 {
3246 	return (1 << state) &
3247 	       (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 |
3248 		TCPF_FIN_WAIT2 | TCPF_SYN_RECV);
3249 }
3250 
tcp_rtx_queue_purge(struct sock * sk)3251 static void tcp_rtx_queue_purge(struct sock *sk)
3252 {
3253 	struct rb_node *p = rb_first(&sk->tcp_rtx_queue);
3254 
3255 	tcp_sk(sk)->highest_sack = NULL;
3256 	while (p) {
3257 		struct sk_buff *skb = rb_to_skb(p);
3258 
3259 		p = rb_next(p);
3260 		/* Since we are deleting whole queue, no need to
3261 		 * list_del(&skb->tcp_tsorted_anchor)
3262 		 */
3263 		tcp_rtx_queue_unlink(skb, sk);
3264 		tcp_wmem_free_skb(sk, skb);
3265 	}
3266 }
3267 
tcp_write_queue_purge(struct sock * sk)3268 void tcp_write_queue_purge(struct sock *sk)
3269 {
3270 	struct sk_buff *skb;
3271 
3272 	tcp_chrono_stop(sk, TCP_CHRONO_BUSY);
3273 	while ((skb = __skb_dequeue(&sk->sk_write_queue)) != NULL) {
3274 		tcp_skb_tsorted_anchor_cleanup(skb);
3275 		tcp_wmem_free_skb(sk, skb);
3276 	}
3277 	tcp_rtx_queue_purge(sk);
3278 	INIT_LIST_HEAD(&tcp_sk(sk)->tsorted_sent_queue);
3279 	tcp_clear_all_retrans_hints(tcp_sk(sk));
3280 	tcp_sk(sk)->packets_out = 0;
3281 	inet_csk(sk)->icsk_backoff = 0;
3282 }
3283 
tcp_disconnect(struct sock * sk,int flags)3284 int tcp_disconnect(struct sock *sk, int flags)
3285 {
3286 	struct inet_sock *inet = inet_sk(sk);
3287 	struct inet_connection_sock *icsk = inet_csk(sk);
3288 	struct tcp_sock *tp = tcp_sk(sk);
3289 	int old_state = sk->sk_state;
3290 	u32 seq;
3291 
3292 	if (old_state != TCP_CLOSE)
3293 		tcp_set_state(sk, TCP_CLOSE);
3294 
3295 	/* ABORT function of RFC793 */
3296 	if (old_state == TCP_LISTEN) {
3297 		inet_csk_listen_stop(sk);
3298 	} else if (unlikely(tp->repair)) {
3299 		WRITE_ONCE(sk->sk_err, ECONNABORTED);
3300 	} else if (tcp_need_reset(old_state)) {
3301 		tcp_send_active_reset(sk, gfp_any(), SK_RST_REASON_TCP_STATE);
3302 		WRITE_ONCE(sk->sk_err, ECONNRESET);
3303 	} else if (tp->snd_nxt != tp->write_seq &&
3304 		   (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK)) {
3305 		/* The last check adjusts for discrepancy of Linux wrt. RFC
3306 		 * states
3307 		 */
3308 		tcp_send_active_reset(sk, gfp_any(),
3309 				      SK_RST_REASON_TCP_DISCONNECT_WITH_DATA);
3310 		WRITE_ONCE(sk->sk_err, ECONNRESET);
3311 	} else if (old_state == TCP_SYN_SENT)
3312 		WRITE_ONCE(sk->sk_err, ECONNRESET);
3313 
3314 	tcp_clear_xmit_timers(sk);
3315 	__skb_queue_purge(&sk->sk_receive_queue);
3316 	WRITE_ONCE(tp->copied_seq, tp->rcv_nxt);
3317 	WRITE_ONCE(tp->urg_data, 0);
3318 	sk_set_peek_off(sk, -1);
3319 	tcp_write_queue_purge(sk);
3320 	tcp_fastopen_active_disable_ofo_check(sk);
3321 	skb_rbtree_purge(&tp->out_of_order_queue);
3322 
3323 	inet->inet_dport = 0;
3324 
3325 	inet_bhash2_reset_saddr(sk);
3326 
3327 	WRITE_ONCE(sk->sk_shutdown, 0);
3328 	sock_reset_flag(sk, SOCK_DONE);
3329 	tp->srtt_us = 0;
3330 	tp->mdev_us = jiffies_to_usecs(TCP_TIMEOUT_INIT);
3331 	tp->rcv_rtt_last_tsecr = 0;
3332 
3333 	seq = tp->write_seq + tp->max_window + 2;
3334 	if (!seq)
3335 		seq = 1;
3336 	WRITE_ONCE(tp->write_seq, seq);
3337 
3338 	icsk->icsk_backoff = 0;
3339 	icsk->icsk_probes_out = 0;
3340 	icsk->icsk_probes_tstamp = 0;
3341 	icsk->icsk_rto = TCP_TIMEOUT_INIT;
3342 	icsk->icsk_rto_min = TCP_RTO_MIN;
3343 	icsk->icsk_delack_max = TCP_DELACK_MAX;
3344 	tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
3345 	tcp_snd_cwnd_set(tp, TCP_INIT_CWND);
3346 	tp->snd_cwnd_cnt = 0;
3347 	tp->is_cwnd_limited = 0;
3348 	tp->max_packets_out = 0;
3349 	tp->window_clamp = 0;
3350 	tp->delivered = 0;
3351 	tp->delivered_ce = 0;
3352 	if (icsk->icsk_ca_initialized && icsk->icsk_ca_ops->release)
3353 		icsk->icsk_ca_ops->release(sk);
3354 	memset(icsk->icsk_ca_priv, 0, sizeof(icsk->icsk_ca_priv));
3355 	icsk->icsk_ca_initialized = 0;
3356 	tcp_set_ca_state(sk, TCP_CA_Open);
3357 	tp->is_sack_reneg = 0;
3358 	tcp_clear_retrans(tp);
3359 	tp->total_retrans = 0;
3360 	inet_csk_delack_init(sk);
3361 	/* Initialize rcv_mss to TCP_MIN_MSS to avoid division by 0
3362 	 * issue in __tcp_select_window()
3363 	 */
3364 	icsk->icsk_ack.rcv_mss = TCP_MIN_MSS;
3365 	memset(&tp->rx_opt, 0, sizeof(tp->rx_opt));
3366 	__sk_dst_reset(sk);
3367 	dst_release(unrcu_pointer(xchg(&sk->sk_rx_dst, NULL)));
3368 	tcp_saved_syn_free(tp);
3369 	tp->compressed_ack = 0;
3370 	tp->segs_in = 0;
3371 	tp->segs_out = 0;
3372 	tp->bytes_sent = 0;
3373 	tp->bytes_acked = 0;
3374 	tp->bytes_received = 0;
3375 	tp->bytes_retrans = 0;
3376 	tp->data_segs_in = 0;
3377 	tp->data_segs_out = 0;
3378 	tp->duplicate_sack[0].start_seq = 0;
3379 	tp->duplicate_sack[0].end_seq = 0;
3380 	tp->dsack_dups = 0;
3381 	tp->reord_seen = 0;
3382 	tp->retrans_out = 0;
3383 	tp->sacked_out = 0;
3384 	tp->tlp_high_seq = 0;
3385 	tp->last_oow_ack_time = 0;
3386 	tp->plb_rehash = 0;
3387 	/* There's a bubble in the pipe until at least the first ACK. */
3388 	tp->app_limited = ~0U;
3389 	tp->rate_app_limited = 1;
3390 	tp->rack.mstamp = 0;
3391 	tp->rack.advanced = 0;
3392 	tp->rack.reo_wnd_steps = 1;
3393 	tp->rack.last_delivered = 0;
3394 	tp->rack.reo_wnd_persist = 0;
3395 	tp->rack.dsack_seen = 0;
3396 	tp->syn_data_acked = 0;
3397 	tp->rx_opt.saw_tstamp = 0;
3398 	tp->rx_opt.dsack = 0;
3399 	tp->rx_opt.num_sacks = 0;
3400 	tp->rcv_ooopack = 0;
3401 
3402 
3403 	/* Clean up fastopen related fields */
3404 	tcp_free_fastopen_req(tp);
3405 	inet_clear_bit(DEFER_CONNECT, sk);
3406 	tp->fastopen_client_fail = 0;
3407 
3408 	WARN_ON(inet->inet_num && !icsk->icsk_bind_hash);
3409 
3410 	if (sk->sk_frag.page) {
3411 		put_page(sk->sk_frag.page);
3412 		sk->sk_frag.page = NULL;
3413 		sk->sk_frag.offset = 0;
3414 	}
3415 	sk_error_report(sk);
3416 	return 0;
3417 }
3418 EXPORT_SYMBOL(tcp_disconnect);
3419 
tcp_can_repair_sock(const struct sock * sk)3420 static inline bool tcp_can_repair_sock(const struct sock *sk)
3421 {
3422 	return sockopt_ns_capable(sock_net(sk)->user_ns, CAP_NET_ADMIN) &&
3423 		(sk->sk_state != TCP_LISTEN);
3424 }
3425 
tcp_repair_set_window(struct tcp_sock * tp,sockptr_t optbuf,int len)3426 static int tcp_repair_set_window(struct tcp_sock *tp, sockptr_t optbuf, int len)
3427 {
3428 	struct tcp_repair_window opt;
3429 
3430 	if (!tp->repair)
3431 		return -EPERM;
3432 
3433 	if (len != sizeof(opt))
3434 		return -EINVAL;
3435 
3436 	if (copy_from_sockptr(&opt, optbuf, sizeof(opt)))
3437 		return -EFAULT;
3438 
3439 	if (opt.max_window < opt.snd_wnd)
3440 		return -EINVAL;
3441 
3442 	if (after(opt.snd_wl1, tp->rcv_nxt + opt.rcv_wnd))
3443 		return -EINVAL;
3444 
3445 	if (after(opt.rcv_wup, tp->rcv_nxt))
3446 		return -EINVAL;
3447 
3448 	tp->snd_wl1	= opt.snd_wl1;
3449 	tp->snd_wnd	= opt.snd_wnd;
3450 	tp->max_window	= opt.max_window;
3451 
3452 	tp->rcv_wnd	= opt.rcv_wnd;
3453 	tp->rcv_wup	= opt.rcv_wup;
3454 
3455 	return 0;
3456 }
3457 
tcp_repair_options_est(struct sock * sk,sockptr_t optbuf,unsigned int len)3458 static int tcp_repair_options_est(struct sock *sk, sockptr_t optbuf,
3459 		unsigned int len)
3460 {
3461 	struct tcp_sock *tp = tcp_sk(sk);
3462 	struct tcp_repair_opt opt;
3463 	size_t offset = 0;
3464 
3465 	while (len >= sizeof(opt)) {
3466 		if (copy_from_sockptr_offset(&opt, optbuf, offset, sizeof(opt)))
3467 			return -EFAULT;
3468 
3469 		offset += sizeof(opt);
3470 		len -= sizeof(opt);
3471 
3472 		switch (opt.opt_code) {
3473 		case TCPOPT_MSS:
3474 			tp->rx_opt.mss_clamp = opt.opt_val;
3475 			tcp_mtup_init(sk);
3476 			break;
3477 		case TCPOPT_WINDOW:
3478 			{
3479 				u16 snd_wscale = opt.opt_val & 0xFFFF;
3480 				u16 rcv_wscale = opt.opt_val >> 16;
3481 
3482 				if (snd_wscale > TCP_MAX_WSCALE || rcv_wscale > TCP_MAX_WSCALE)
3483 					return -EFBIG;
3484 
3485 				tp->rx_opt.snd_wscale = snd_wscale;
3486 				tp->rx_opt.rcv_wscale = rcv_wscale;
3487 				tp->rx_opt.wscale_ok = 1;
3488 			}
3489 			break;
3490 		case TCPOPT_SACK_PERM:
3491 			if (opt.opt_val != 0)
3492 				return -EINVAL;
3493 
3494 			tp->rx_opt.sack_ok |= TCP_SACK_SEEN;
3495 			break;
3496 		case TCPOPT_TIMESTAMP:
3497 			if (opt.opt_val != 0)
3498 				return -EINVAL;
3499 
3500 			tp->rx_opt.tstamp_ok = 1;
3501 			break;
3502 		}
3503 	}
3504 
3505 	return 0;
3506 }
3507 
3508 DEFINE_STATIC_KEY_FALSE(tcp_tx_delay_enabled);
3509 EXPORT_SYMBOL(tcp_tx_delay_enabled);
3510 
tcp_enable_tx_delay(void)3511 static void tcp_enable_tx_delay(void)
3512 {
3513 	if (!static_branch_unlikely(&tcp_tx_delay_enabled)) {
3514 		static int __tcp_tx_delay_enabled = 0;
3515 
3516 		if (cmpxchg(&__tcp_tx_delay_enabled, 0, 1) == 0) {
3517 			static_branch_enable(&tcp_tx_delay_enabled);
3518 			pr_info("TCP_TX_DELAY enabled\n");
3519 		}
3520 	}
3521 }
3522 
3523 /* When set indicates to always queue non-full frames.  Later the user clears
3524  * this option and we transmit any pending partial frames in the queue.  This is
3525  * meant to be used alongside sendfile() to get properly filled frames when the
3526  * user (for example) must write out headers with a write() call first and then
3527  * use sendfile to send out the data parts.
3528  *
3529  * TCP_CORK can be set together with TCP_NODELAY and it is stronger than
3530  * TCP_NODELAY.
3531  */
__tcp_sock_set_cork(struct sock * sk,bool on)3532 void __tcp_sock_set_cork(struct sock *sk, bool on)
3533 {
3534 	struct tcp_sock *tp = tcp_sk(sk);
3535 
3536 	if (on) {
3537 		tp->nonagle |= TCP_NAGLE_CORK;
3538 	} else {
3539 		tp->nonagle &= ~TCP_NAGLE_CORK;
3540 		if (tp->nonagle & TCP_NAGLE_OFF)
3541 			tp->nonagle |= TCP_NAGLE_PUSH;
3542 		tcp_push_pending_frames(sk);
3543 	}
3544 }
3545 
tcp_sock_set_cork(struct sock * sk,bool on)3546 void tcp_sock_set_cork(struct sock *sk, bool on)
3547 {
3548 	lock_sock(sk);
3549 	__tcp_sock_set_cork(sk, on);
3550 	release_sock(sk);
3551 }
3552 EXPORT_SYMBOL(tcp_sock_set_cork);
3553 
3554 /* TCP_NODELAY is weaker than TCP_CORK, so that this option on corked socket is
3555  * remembered, but it is not activated until cork is cleared.
3556  *
3557  * However, when TCP_NODELAY is set we make an explicit push, which overrides
3558  * even TCP_CORK for currently queued segments.
3559  */
__tcp_sock_set_nodelay(struct sock * sk,bool on)3560 void __tcp_sock_set_nodelay(struct sock *sk, bool on)
3561 {
3562 	if (on) {
3563 		tcp_sk(sk)->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH;
3564 		tcp_push_pending_frames(sk);
3565 	} else {
3566 		tcp_sk(sk)->nonagle &= ~TCP_NAGLE_OFF;
3567 	}
3568 }
3569 
tcp_sock_set_nodelay(struct sock * sk)3570 void tcp_sock_set_nodelay(struct sock *sk)
3571 {
3572 	lock_sock(sk);
3573 	__tcp_sock_set_nodelay(sk, true);
3574 	release_sock(sk);
3575 }
3576 EXPORT_SYMBOL(tcp_sock_set_nodelay);
3577 
__tcp_sock_set_quickack(struct sock * sk,int val)3578 static void __tcp_sock_set_quickack(struct sock *sk, int val)
3579 {
3580 	if (!val) {
3581 		inet_csk_enter_pingpong_mode(sk);
3582 		return;
3583 	}
3584 
3585 	inet_csk_exit_pingpong_mode(sk);
3586 	if ((1 << sk->sk_state) & (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) &&
3587 	    inet_csk_ack_scheduled(sk)) {
3588 		inet_csk(sk)->icsk_ack.pending |= ICSK_ACK_PUSHED;
3589 		tcp_cleanup_rbuf(sk, 1);
3590 		if (!(val & 1))
3591 			inet_csk_enter_pingpong_mode(sk);
3592 	}
3593 }
3594 
tcp_sock_set_quickack(struct sock * sk,int val)3595 void tcp_sock_set_quickack(struct sock *sk, int val)
3596 {
3597 	lock_sock(sk);
3598 	__tcp_sock_set_quickack(sk, val);
3599 	release_sock(sk);
3600 }
3601 EXPORT_SYMBOL(tcp_sock_set_quickack);
3602 
tcp_sock_set_syncnt(struct sock * sk,int val)3603 int tcp_sock_set_syncnt(struct sock *sk, int val)
3604 {
3605 	if (val < 1 || val > MAX_TCP_SYNCNT)
3606 		return -EINVAL;
3607 
3608 	WRITE_ONCE(inet_csk(sk)->icsk_syn_retries, val);
3609 	return 0;
3610 }
3611 EXPORT_SYMBOL(tcp_sock_set_syncnt);
3612 
tcp_sock_set_user_timeout(struct sock * sk,int val)3613 int tcp_sock_set_user_timeout(struct sock *sk, int val)
3614 {
3615 	/* Cap the max time in ms TCP will retry or probe the window
3616 	 * before giving up and aborting (ETIMEDOUT) a connection.
3617 	 */
3618 	if (val < 0)
3619 		return -EINVAL;
3620 
3621 	WRITE_ONCE(inet_csk(sk)->icsk_user_timeout, val);
3622 	return 0;
3623 }
3624 EXPORT_SYMBOL(tcp_sock_set_user_timeout);
3625 
tcp_sock_set_keepidle_locked(struct sock * sk,int val)3626 int tcp_sock_set_keepidle_locked(struct sock *sk, int val)
3627 {
3628 	struct tcp_sock *tp = tcp_sk(sk);
3629 
3630 	if (val < 1 || val > MAX_TCP_KEEPIDLE)
3631 		return -EINVAL;
3632 
3633 	/* Paired with WRITE_ONCE() in keepalive_time_when() */
3634 	WRITE_ONCE(tp->keepalive_time, val * HZ);
3635 	if (sock_flag(sk, SOCK_KEEPOPEN) &&
3636 	    !((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN))) {
3637 		u32 elapsed = keepalive_time_elapsed(tp);
3638 
3639 		if (tp->keepalive_time > elapsed)
3640 			elapsed = tp->keepalive_time - elapsed;
3641 		else
3642 			elapsed = 0;
3643 		inet_csk_reset_keepalive_timer(sk, elapsed);
3644 	}
3645 
3646 	return 0;
3647 }
3648 
tcp_sock_set_keepidle(struct sock * sk,int val)3649 int tcp_sock_set_keepidle(struct sock *sk, int val)
3650 {
3651 	int err;
3652 
3653 	lock_sock(sk);
3654 	err = tcp_sock_set_keepidle_locked(sk, val);
3655 	release_sock(sk);
3656 	return err;
3657 }
3658 EXPORT_SYMBOL(tcp_sock_set_keepidle);
3659 
tcp_sock_set_keepintvl(struct sock * sk,int val)3660 int tcp_sock_set_keepintvl(struct sock *sk, int val)
3661 {
3662 	if (val < 1 || val > MAX_TCP_KEEPINTVL)
3663 		return -EINVAL;
3664 
3665 	WRITE_ONCE(tcp_sk(sk)->keepalive_intvl, val * HZ);
3666 	return 0;
3667 }
3668 EXPORT_SYMBOL(tcp_sock_set_keepintvl);
3669 
tcp_sock_set_keepcnt(struct sock * sk,int val)3670 int tcp_sock_set_keepcnt(struct sock *sk, int val)
3671 {
3672 	if (val < 1 || val > MAX_TCP_KEEPCNT)
3673 		return -EINVAL;
3674 
3675 	/* Paired with READ_ONCE() in keepalive_probes() */
3676 	WRITE_ONCE(tcp_sk(sk)->keepalive_probes, val);
3677 	return 0;
3678 }
3679 EXPORT_SYMBOL(tcp_sock_set_keepcnt);
3680 
tcp_set_window_clamp(struct sock * sk,int val)3681 int tcp_set_window_clamp(struct sock *sk, int val)
3682 {
3683 	struct tcp_sock *tp = tcp_sk(sk);
3684 
3685 	if (!val) {
3686 		if (sk->sk_state != TCP_CLOSE)
3687 			return -EINVAL;
3688 		WRITE_ONCE(tp->window_clamp, 0);
3689 	} else {
3690 		u32 new_rcv_ssthresh, old_window_clamp = tp->window_clamp;
3691 		u32 new_window_clamp = val < SOCK_MIN_RCVBUF / 2 ?
3692 						SOCK_MIN_RCVBUF / 2 : val;
3693 
3694 		if (new_window_clamp == old_window_clamp)
3695 			return 0;
3696 
3697 		WRITE_ONCE(tp->window_clamp, new_window_clamp);
3698 		if (new_window_clamp < old_window_clamp) {
3699 			/* need to apply the reserved mem provisioning only
3700 			 * when shrinking the window clamp
3701 			 */
3702 			__tcp_adjust_rcv_ssthresh(sk, tp->window_clamp);
3703 
3704 		} else {
3705 			new_rcv_ssthresh = min(tp->rcv_wnd, tp->window_clamp);
3706 			tp->rcv_ssthresh = max(new_rcv_ssthresh,
3707 					       tp->rcv_ssthresh);
3708 		}
3709 	}
3710 	return 0;
3711 }
3712 
3713 /*
3714  *	Socket option code for TCP.
3715  */
do_tcp_setsockopt(struct sock * sk,int level,int optname,sockptr_t optval,unsigned int optlen)3716 int do_tcp_setsockopt(struct sock *sk, int level, int optname,
3717 		      sockptr_t optval, unsigned int optlen)
3718 {
3719 	struct tcp_sock *tp = tcp_sk(sk);
3720 	struct inet_connection_sock *icsk = inet_csk(sk);
3721 	struct net *net = sock_net(sk);
3722 	int val;
3723 	int err = 0;
3724 
3725 	/* These are data/string values, all the others are ints */
3726 	switch (optname) {
3727 	case TCP_CONGESTION: {
3728 		char name[TCP_CA_NAME_MAX];
3729 
3730 		if (optlen < 1)
3731 			return -EINVAL;
3732 
3733 		val = strncpy_from_sockptr(name, optval,
3734 					min_t(long, TCP_CA_NAME_MAX-1, optlen));
3735 		if (val < 0)
3736 			return -EFAULT;
3737 		name[val] = 0;
3738 
3739 		sockopt_lock_sock(sk);
3740 		err = tcp_set_congestion_control(sk, name, !has_current_bpf_ctx(),
3741 						 sockopt_ns_capable(sock_net(sk)->user_ns,
3742 								    CAP_NET_ADMIN));
3743 		sockopt_release_sock(sk);
3744 		return err;
3745 	}
3746 	case TCP_ULP: {
3747 		char name[TCP_ULP_NAME_MAX];
3748 
3749 		if (optlen < 1)
3750 			return -EINVAL;
3751 
3752 		val = strncpy_from_sockptr(name, optval,
3753 					min_t(long, TCP_ULP_NAME_MAX - 1,
3754 					      optlen));
3755 		if (val < 0)
3756 			return -EFAULT;
3757 		name[val] = 0;
3758 
3759 		sockopt_lock_sock(sk);
3760 		err = tcp_set_ulp(sk, name);
3761 		sockopt_release_sock(sk);
3762 		return err;
3763 	}
3764 	case TCP_FASTOPEN_KEY: {
3765 		__u8 key[TCP_FASTOPEN_KEY_BUF_LENGTH];
3766 		__u8 *backup_key = NULL;
3767 
3768 		/* Allow a backup key as well to facilitate key rotation
3769 		 * First key is the active one.
3770 		 */
3771 		if (optlen != TCP_FASTOPEN_KEY_LENGTH &&
3772 		    optlen != TCP_FASTOPEN_KEY_BUF_LENGTH)
3773 			return -EINVAL;
3774 
3775 		if (copy_from_sockptr(key, optval, optlen))
3776 			return -EFAULT;
3777 
3778 		if (optlen == TCP_FASTOPEN_KEY_BUF_LENGTH)
3779 			backup_key = key + TCP_FASTOPEN_KEY_LENGTH;
3780 
3781 		return tcp_fastopen_reset_cipher(net, sk, key, backup_key);
3782 	}
3783 	default:
3784 		/* fallthru */
3785 		break;
3786 	}
3787 
3788 	if (optlen < sizeof(int))
3789 		return -EINVAL;
3790 
3791 	if (copy_from_sockptr(&val, optval, sizeof(val)))
3792 		return -EFAULT;
3793 
3794 	/* Handle options that can be set without locking the socket. */
3795 	switch (optname) {
3796 	case TCP_SYNCNT:
3797 		return tcp_sock_set_syncnt(sk, val);
3798 	case TCP_USER_TIMEOUT:
3799 		return tcp_sock_set_user_timeout(sk, val);
3800 	case TCP_KEEPINTVL:
3801 		return tcp_sock_set_keepintvl(sk, val);
3802 	case TCP_KEEPCNT:
3803 		return tcp_sock_set_keepcnt(sk, val);
3804 	case TCP_LINGER2:
3805 		if (val < 0)
3806 			WRITE_ONCE(tp->linger2, -1);
3807 		else if (val > TCP_FIN_TIMEOUT_MAX / HZ)
3808 			WRITE_ONCE(tp->linger2, TCP_FIN_TIMEOUT_MAX);
3809 		else
3810 			WRITE_ONCE(tp->linger2, val * HZ);
3811 		return 0;
3812 	case TCP_DEFER_ACCEPT:
3813 		/* Translate value in seconds to number of retransmits */
3814 		WRITE_ONCE(icsk->icsk_accept_queue.rskq_defer_accept,
3815 			   secs_to_retrans(val, TCP_TIMEOUT_INIT / HZ,
3816 					   TCP_RTO_MAX / HZ));
3817 		return 0;
3818 	}
3819 
3820 	sockopt_lock_sock(sk);
3821 
3822 	switch (optname) {
3823 	case TCP_MAXSEG:
3824 		/* Values greater than interface MTU won't take effect. However
3825 		 * at the point when this call is done we typically don't yet
3826 		 * know which interface is going to be used
3827 		 */
3828 		if (val && (val < TCP_MIN_MSS || val > MAX_TCP_WINDOW)) {
3829 			err = -EINVAL;
3830 			break;
3831 		}
3832 		tp->rx_opt.user_mss = val;
3833 		break;
3834 
3835 	case TCP_NODELAY:
3836 		__tcp_sock_set_nodelay(sk, val);
3837 		break;
3838 
3839 	case TCP_THIN_LINEAR_TIMEOUTS:
3840 		if (val < 0 || val > 1)
3841 			err = -EINVAL;
3842 		else
3843 			tp->thin_lto = val;
3844 		break;
3845 
3846 	case TCP_THIN_DUPACK:
3847 		if (val < 0 || val > 1)
3848 			err = -EINVAL;
3849 		break;
3850 
3851 	case TCP_REPAIR:
3852 		if (!tcp_can_repair_sock(sk))
3853 			err = -EPERM;
3854 		else if (val == TCP_REPAIR_ON) {
3855 			tp->repair = 1;
3856 			sk->sk_reuse = SK_FORCE_REUSE;
3857 			tp->repair_queue = TCP_NO_QUEUE;
3858 		} else if (val == TCP_REPAIR_OFF) {
3859 			tp->repair = 0;
3860 			sk->sk_reuse = SK_NO_REUSE;
3861 			tcp_send_window_probe(sk);
3862 		} else if (val == TCP_REPAIR_OFF_NO_WP) {
3863 			tp->repair = 0;
3864 			sk->sk_reuse = SK_NO_REUSE;
3865 		} else
3866 			err = -EINVAL;
3867 
3868 		break;
3869 
3870 	case TCP_REPAIR_QUEUE:
3871 		if (!tp->repair)
3872 			err = -EPERM;
3873 		else if ((unsigned int)val < TCP_QUEUES_NR)
3874 			tp->repair_queue = val;
3875 		else
3876 			err = -EINVAL;
3877 		break;
3878 
3879 	case TCP_QUEUE_SEQ:
3880 		if (sk->sk_state != TCP_CLOSE) {
3881 			err = -EPERM;
3882 		} else if (tp->repair_queue == TCP_SEND_QUEUE) {
3883 			if (!tcp_rtx_queue_empty(sk))
3884 				err = -EPERM;
3885 			else
3886 				WRITE_ONCE(tp->write_seq, val);
3887 		} else if (tp->repair_queue == TCP_RECV_QUEUE) {
3888 			if (tp->rcv_nxt != tp->copied_seq) {
3889 				err = -EPERM;
3890 			} else {
3891 				WRITE_ONCE(tp->rcv_nxt, val);
3892 				WRITE_ONCE(tp->copied_seq, val);
3893 			}
3894 		} else {
3895 			err = -EINVAL;
3896 		}
3897 		break;
3898 
3899 	case TCP_REPAIR_OPTIONS:
3900 		if (!tp->repair)
3901 			err = -EINVAL;
3902 		else if (sk->sk_state == TCP_ESTABLISHED && !tp->bytes_sent)
3903 			err = tcp_repair_options_est(sk, optval, optlen);
3904 		else
3905 			err = -EPERM;
3906 		break;
3907 
3908 	case TCP_CORK:
3909 		__tcp_sock_set_cork(sk, val);
3910 		break;
3911 
3912 	case TCP_KEEPIDLE:
3913 		err = tcp_sock_set_keepidle_locked(sk, val);
3914 		break;
3915 	case TCP_SAVE_SYN:
3916 		/* 0: disable, 1: enable, 2: start from ether_header */
3917 		if (val < 0 || val > 2)
3918 			err = -EINVAL;
3919 		else
3920 			tp->save_syn = val;
3921 		break;
3922 
3923 	case TCP_WINDOW_CLAMP:
3924 		err = tcp_set_window_clamp(sk, val);
3925 		break;
3926 
3927 	case TCP_QUICKACK:
3928 		__tcp_sock_set_quickack(sk, val);
3929 		break;
3930 
3931 	case TCP_AO_REPAIR:
3932 		if (!tcp_can_repair_sock(sk)) {
3933 			err = -EPERM;
3934 			break;
3935 		}
3936 		err = tcp_ao_set_repair(sk, optval, optlen);
3937 		break;
3938 #ifdef CONFIG_TCP_AO
3939 	case TCP_AO_ADD_KEY:
3940 	case TCP_AO_DEL_KEY:
3941 	case TCP_AO_INFO: {
3942 		/* If this is the first TCP-AO setsockopt() on the socket,
3943 		 * sk_state has to be LISTEN or CLOSE. Allow TCP_REPAIR
3944 		 * in any state.
3945 		 */
3946 		if ((1 << sk->sk_state) & (TCPF_LISTEN | TCPF_CLOSE))
3947 			goto ao_parse;
3948 		if (rcu_dereference_protected(tcp_sk(sk)->ao_info,
3949 					      lockdep_sock_is_held(sk)))
3950 			goto ao_parse;
3951 		if (tp->repair)
3952 			goto ao_parse;
3953 		err = -EISCONN;
3954 		break;
3955 ao_parse:
3956 		err = tp->af_specific->ao_parse(sk, optname, optval, optlen);
3957 		break;
3958 	}
3959 #endif
3960 #ifdef CONFIG_TCP_MD5SIG
3961 	case TCP_MD5SIG:
3962 	case TCP_MD5SIG_EXT:
3963 		err = tp->af_specific->md5_parse(sk, optname, optval, optlen);
3964 		break;
3965 #endif
3966 	case TCP_FASTOPEN:
3967 		if (val >= 0 && ((1 << sk->sk_state) & (TCPF_CLOSE |
3968 		    TCPF_LISTEN))) {
3969 			tcp_fastopen_init_key_once(net);
3970 
3971 			fastopen_queue_tune(sk, val);
3972 		} else {
3973 			err = -EINVAL;
3974 		}
3975 		break;
3976 	case TCP_FASTOPEN_CONNECT:
3977 		if (val > 1 || val < 0) {
3978 			err = -EINVAL;
3979 		} else if (READ_ONCE(net->ipv4.sysctl_tcp_fastopen) &
3980 			   TFO_CLIENT_ENABLE) {
3981 			if (sk->sk_state == TCP_CLOSE)
3982 				tp->fastopen_connect = val;
3983 			else
3984 				err = -EINVAL;
3985 		} else {
3986 			err = -EOPNOTSUPP;
3987 		}
3988 		break;
3989 	case TCP_FASTOPEN_NO_COOKIE:
3990 		if (val > 1 || val < 0)
3991 			err = -EINVAL;
3992 		else if (!((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
3993 			err = -EINVAL;
3994 		else
3995 			tp->fastopen_no_cookie = val;
3996 		break;
3997 	case TCP_TIMESTAMP:
3998 		if (!tp->repair) {
3999 			err = -EPERM;
4000 			break;
4001 		}
4002 		/* val is an opaque field,
4003 		 * and low order bit contains usec_ts enable bit.
4004 		 * Its a best effort, and we do not care if user makes an error.
4005 		 */
4006 		tp->tcp_usec_ts = val & 1;
4007 		WRITE_ONCE(tp->tsoffset, val - tcp_clock_ts(tp->tcp_usec_ts));
4008 		break;
4009 	case TCP_REPAIR_WINDOW:
4010 		err = tcp_repair_set_window(tp, optval, optlen);
4011 		break;
4012 	case TCP_NOTSENT_LOWAT:
4013 		WRITE_ONCE(tp->notsent_lowat, val);
4014 		sk->sk_write_space(sk);
4015 		break;
4016 	case TCP_INQ:
4017 		if (val > 1 || val < 0)
4018 			err = -EINVAL;
4019 		else
4020 			tp->recvmsg_inq = val;
4021 		break;
4022 	case TCP_TX_DELAY:
4023 		if (val)
4024 			tcp_enable_tx_delay();
4025 		WRITE_ONCE(tp->tcp_tx_delay, val);
4026 		break;
4027 	default:
4028 		err = -ENOPROTOOPT;
4029 		break;
4030 	}
4031 
4032 	sockopt_release_sock(sk);
4033 	return err;
4034 }
4035 
tcp_setsockopt(struct sock * sk,int level,int optname,sockptr_t optval,unsigned int optlen)4036 int tcp_setsockopt(struct sock *sk, int level, int optname, sockptr_t optval,
4037 		   unsigned int optlen)
4038 {
4039 	const struct inet_connection_sock *icsk = inet_csk(sk);
4040 
4041 	if (level != SOL_TCP)
4042 		/* Paired with WRITE_ONCE() in do_ipv6_setsockopt() and tcp_v6_connect() */
4043 		return READ_ONCE(icsk->icsk_af_ops)->setsockopt(sk, level, optname,
4044 								optval, optlen);
4045 	return do_tcp_setsockopt(sk, level, optname, optval, optlen);
4046 }
4047 EXPORT_SYMBOL(tcp_setsockopt);
4048 
tcp_get_info_chrono_stats(const struct tcp_sock * tp,struct tcp_info * info)4049 static void tcp_get_info_chrono_stats(const struct tcp_sock *tp,
4050 				      struct tcp_info *info)
4051 {
4052 	u64 stats[__TCP_CHRONO_MAX], total = 0;
4053 	enum tcp_chrono i;
4054 
4055 	for (i = TCP_CHRONO_BUSY; i < __TCP_CHRONO_MAX; ++i) {
4056 		stats[i] = tp->chrono_stat[i - 1];
4057 		if (i == tp->chrono_type)
4058 			stats[i] += tcp_jiffies32 - tp->chrono_start;
4059 		stats[i] *= USEC_PER_SEC / HZ;
4060 		total += stats[i];
4061 	}
4062 
4063 	info->tcpi_busy_time = total;
4064 	info->tcpi_rwnd_limited = stats[TCP_CHRONO_RWND_LIMITED];
4065 	info->tcpi_sndbuf_limited = stats[TCP_CHRONO_SNDBUF_LIMITED];
4066 }
4067 
4068 /* Return information about state of tcp endpoint in API format. */
tcp_get_info(struct sock * sk,struct tcp_info * info)4069 void tcp_get_info(struct sock *sk, struct tcp_info *info)
4070 {
4071 	const struct tcp_sock *tp = tcp_sk(sk); /* iff sk_type == SOCK_STREAM */
4072 	const struct inet_connection_sock *icsk = inet_csk(sk);
4073 	unsigned long rate;
4074 	u32 now;
4075 	u64 rate64;
4076 	bool slow;
4077 
4078 	memset(info, 0, sizeof(*info));
4079 	if (sk->sk_type != SOCK_STREAM)
4080 		return;
4081 
4082 	info->tcpi_state = inet_sk_state_load(sk);
4083 
4084 	/* Report meaningful fields for all TCP states, including listeners */
4085 	rate = READ_ONCE(sk->sk_pacing_rate);
4086 	rate64 = (rate != ~0UL) ? rate : ~0ULL;
4087 	info->tcpi_pacing_rate = rate64;
4088 
4089 	rate = READ_ONCE(sk->sk_max_pacing_rate);
4090 	rate64 = (rate != ~0UL) ? rate : ~0ULL;
4091 	info->tcpi_max_pacing_rate = rate64;
4092 
4093 	info->tcpi_reordering = tp->reordering;
4094 	info->tcpi_snd_cwnd = tcp_snd_cwnd(tp);
4095 
4096 	if (info->tcpi_state == TCP_LISTEN) {
4097 		/* listeners aliased fields :
4098 		 * tcpi_unacked -> Number of children ready for accept()
4099 		 * tcpi_sacked  -> max backlog
4100 		 */
4101 		info->tcpi_unacked = READ_ONCE(sk->sk_ack_backlog);
4102 		info->tcpi_sacked = READ_ONCE(sk->sk_max_ack_backlog);
4103 		return;
4104 	}
4105 
4106 	slow = lock_sock_fast(sk);
4107 
4108 	info->tcpi_ca_state = icsk->icsk_ca_state;
4109 	info->tcpi_retransmits = icsk->icsk_retransmits;
4110 	info->tcpi_probes = icsk->icsk_probes_out;
4111 	info->tcpi_backoff = icsk->icsk_backoff;
4112 
4113 	if (tp->rx_opt.tstamp_ok)
4114 		info->tcpi_options |= TCPI_OPT_TIMESTAMPS;
4115 	if (tcp_is_sack(tp))
4116 		info->tcpi_options |= TCPI_OPT_SACK;
4117 	if (tp->rx_opt.wscale_ok) {
4118 		info->tcpi_options |= TCPI_OPT_WSCALE;
4119 		info->tcpi_snd_wscale = tp->rx_opt.snd_wscale;
4120 		info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale;
4121 	}
4122 
4123 	if (tp->ecn_flags & TCP_ECN_OK)
4124 		info->tcpi_options |= TCPI_OPT_ECN;
4125 	if (tp->ecn_flags & TCP_ECN_SEEN)
4126 		info->tcpi_options |= TCPI_OPT_ECN_SEEN;
4127 	if (tp->syn_data_acked)
4128 		info->tcpi_options |= TCPI_OPT_SYN_DATA;
4129 	if (tp->tcp_usec_ts)
4130 		info->tcpi_options |= TCPI_OPT_USEC_TS;
4131 
4132 	info->tcpi_rto = jiffies_to_usecs(icsk->icsk_rto);
4133 	info->tcpi_ato = jiffies_to_usecs(min_t(u32, icsk->icsk_ack.ato,
4134 						tcp_delack_max(sk)));
4135 	info->tcpi_snd_mss = tp->mss_cache;
4136 	info->tcpi_rcv_mss = icsk->icsk_ack.rcv_mss;
4137 
4138 	info->tcpi_unacked = tp->packets_out;
4139 	info->tcpi_sacked = tp->sacked_out;
4140 
4141 	info->tcpi_lost = tp->lost_out;
4142 	info->tcpi_retrans = tp->retrans_out;
4143 
4144 	now = tcp_jiffies32;
4145 	info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime);
4146 	info->tcpi_last_data_recv = jiffies_to_msecs(now - icsk->icsk_ack.lrcvtime);
4147 	info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp);
4148 
4149 	info->tcpi_pmtu = icsk->icsk_pmtu_cookie;
4150 	info->tcpi_rcv_ssthresh = tp->rcv_ssthresh;
4151 	info->tcpi_rtt = tp->srtt_us >> 3;
4152 	info->tcpi_rttvar = tp->mdev_us >> 2;
4153 	info->tcpi_snd_ssthresh = tp->snd_ssthresh;
4154 	info->tcpi_advmss = tp->advmss;
4155 
4156 	info->tcpi_rcv_rtt = tp->rcv_rtt_est.rtt_us >> 3;
4157 	info->tcpi_rcv_space = tp->rcvq_space.space;
4158 
4159 	info->tcpi_total_retrans = tp->total_retrans;
4160 
4161 	info->tcpi_bytes_acked = tp->bytes_acked;
4162 	info->tcpi_bytes_received = tp->bytes_received;
4163 	info->tcpi_notsent_bytes = max_t(int, 0, tp->write_seq - tp->snd_nxt);
4164 	tcp_get_info_chrono_stats(tp, info);
4165 
4166 	info->tcpi_segs_out = tp->segs_out;
4167 
4168 	/* segs_in and data_segs_in can be updated from tcp_segs_in() from BH */
4169 	info->tcpi_segs_in = READ_ONCE(tp->segs_in);
4170 	info->tcpi_data_segs_in = READ_ONCE(tp->data_segs_in);
4171 
4172 	info->tcpi_min_rtt = tcp_min_rtt(tp);
4173 	info->tcpi_data_segs_out = tp->data_segs_out;
4174 
4175 	info->tcpi_delivery_rate_app_limited = tp->rate_app_limited ? 1 : 0;
4176 	rate64 = tcp_compute_delivery_rate(tp);
4177 	if (rate64)
4178 		info->tcpi_delivery_rate = rate64;
4179 	info->tcpi_delivered = tp->delivered;
4180 	info->tcpi_delivered_ce = tp->delivered_ce;
4181 	info->tcpi_bytes_sent = tp->bytes_sent;
4182 	info->tcpi_bytes_retrans = tp->bytes_retrans;
4183 	info->tcpi_dsack_dups = tp->dsack_dups;
4184 	info->tcpi_reord_seen = tp->reord_seen;
4185 	info->tcpi_rcv_ooopack = tp->rcv_ooopack;
4186 	info->tcpi_snd_wnd = tp->snd_wnd;
4187 	info->tcpi_rcv_wnd = tp->rcv_wnd;
4188 	info->tcpi_rehash = tp->plb_rehash + tp->timeout_rehash;
4189 	info->tcpi_fastopen_client_fail = tp->fastopen_client_fail;
4190 
4191 	info->tcpi_total_rto = tp->total_rto;
4192 	info->tcpi_total_rto_recoveries = tp->total_rto_recoveries;
4193 	info->tcpi_total_rto_time = tp->total_rto_time;
4194 	if (tp->rto_stamp)
4195 		info->tcpi_total_rto_time += tcp_clock_ms() - tp->rto_stamp;
4196 
4197 	unlock_sock_fast(sk, slow);
4198 }
4199 EXPORT_SYMBOL_GPL(tcp_get_info);
4200 
tcp_opt_stats_get_size(void)4201 static size_t tcp_opt_stats_get_size(void)
4202 {
4203 	return
4204 		nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_BUSY */
4205 		nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_RWND_LIMITED */
4206 		nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_SNDBUF_LIMITED */
4207 		nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_DATA_SEGS_OUT */
4208 		nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_TOTAL_RETRANS */
4209 		nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_PACING_RATE */
4210 		nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_DELIVERY_RATE */
4211 		nla_total_size(sizeof(u32)) + /* TCP_NLA_SND_CWND */
4212 		nla_total_size(sizeof(u32)) + /* TCP_NLA_REORDERING */
4213 		nla_total_size(sizeof(u32)) + /* TCP_NLA_MIN_RTT */
4214 		nla_total_size(sizeof(u8)) + /* TCP_NLA_RECUR_RETRANS */
4215 		nla_total_size(sizeof(u8)) + /* TCP_NLA_DELIVERY_RATE_APP_LMT */
4216 		nla_total_size(sizeof(u32)) + /* TCP_NLA_SNDQ_SIZE */
4217 		nla_total_size(sizeof(u8)) + /* TCP_NLA_CA_STATE */
4218 		nla_total_size(sizeof(u32)) + /* TCP_NLA_SND_SSTHRESH */
4219 		nla_total_size(sizeof(u32)) + /* TCP_NLA_DELIVERED */
4220 		nla_total_size(sizeof(u32)) + /* TCP_NLA_DELIVERED_CE */
4221 		nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_BYTES_SENT */
4222 		nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_BYTES_RETRANS */
4223 		nla_total_size(sizeof(u32)) + /* TCP_NLA_DSACK_DUPS */
4224 		nla_total_size(sizeof(u32)) + /* TCP_NLA_REORD_SEEN */
4225 		nla_total_size(sizeof(u32)) + /* TCP_NLA_SRTT */
4226 		nla_total_size(sizeof(u16)) + /* TCP_NLA_TIMEOUT_REHASH */
4227 		nla_total_size(sizeof(u32)) + /* TCP_NLA_BYTES_NOTSENT */
4228 		nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_EDT */
4229 		nla_total_size(sizeof(u8)) + /* TCP_NLA_TTL */
4230 		nla_total_size(sizeof(u32)) + /* TCP_NLA_REHASH */
4231 		0;
4232 }
4233 
4234 /* Returns TTL or hop limit of an incoming packet from skb. */
tcp_skb_ttl_or_hop_limit(const struct sk_buff * skb)4235 static u8 tcp_skb_ttl_or_hop_limit(const struct sk_buff *skb)
4236 {
4237 	if (skb->protocol == htons(ETH_P_IP))
4238 		return ip_hdr(skb)->ttl;
4239 	else if (skb->protocol == htons(ETH_P_IPV6))
4240 		return ipv6_hdr(skb)->hop_limit;
4241 	else
4242 		return 0;
4243 }
4244 
tcp_get_timestamping_opt_stats(const struct sock * sk,const struct sk_buff * orig_skb,const struct sk_buff * ack_skb)4245 struct sk_buff *tcp_get_timestamping_opt_stats(const struct sock *sk,
4246 					       const struct sk_buff *orig_skb,
4247 					       const struct sk_buff *ack_skb)
4248 {
4249 	const struct tcp_sock *tp = tcp_sk(sk);
4250 	struct sk_buff *stats;
4251 	struct tcp_info info;
4252 	unsigned long rate;
4253 	u64 rate64;
4254 
4255 	stats = alloc_skb(tcp_opt_stats_get_size(), GFP_ATOMIC);
4256 	if (!stats)
4257 		return NULL;
4258 
4259 	tcp_get_info_chrono_stats(tp, &info);
4260 	nla_put_u64_64bit(stats, TCP_NLA_BUSY,
4261 			  info.tcpi_busy_time, TCP_NLA_PAD);
4262 	nla_put_u64_64bit(stats, TCP_NLA_RWND_LIMITED,
4263 			  info.tcpi_rwnd_limited, TCP_NLA_PAD);
4264 	nla_put_u64_64bit(stats, TCP_NLA_SNDBUF_LIMITED,
4265 			  info.tcpi_sndbuf_limited, TCP_NLA_PAD);
4266 	nla_put_u64_64bit(stats, TCP_NLA_DATA_SEGS_OUT,
4267 			  tp->data_segs_out, TCP_NLA_PAD);
4268 	nla_put_u64_64bit(stats, TCP_NLA_TOTAL_RETRANS,
4269 			  tp->total_retrans, TCP_NLA_PAD);
4270 
4271 	rate = READ_ONCE(sk->sk_pacing_rate);
4272 	rate64 = (rate != ~0UL) ? rate : ~0ULL;
4273 	nla_put_u64_64bit(stats, TCP_NLA_PACING_RATE, rate64, TCP_NLA_PAD);
4274 
4275 	rate64 = tcp_compute_delivery_rate(tp);
4276 	nla_put_u64_64bit(stats, TCP_NLA_DELIVERY_RATE, rate64, TCP_NLA_PAD);
4277 
4278 	nla_put_u32(stats, TCP_NLA_SND_CWND, tcp_snd_cwnd(tp));
4279 	nla_put_u32(stats, TCP_NLA_REORDERING, tp->reordering);
4280 	nla_put_u32(stats, TCP_NLA_MIN_RTT, tcp_min_rtt(tp));
4281 
4282 	nla_put_u8(stats, TCP_NLA_RECUR_RETRANS, inet_csk(sk)->icsk_retransmits);
4283 	nla_put_u8(stats, TCP_NLA_DELIVERY_RATE_APP_LMT, !!tp->rate_app_limited);
4284 	nla_put_u32(stats, TCP_NLA_SND_SSTHRESH, tp->snd_ssthresh);
4285 	nla_put_u32(stats, TCP_NLA_DELIVERED, tp->delivered);
4286 	nla_put_u32(stats, TCP_NLA_DELIVERED_CE, tp->delivered_ce);
4287 
4288 	nla_put_u32(stats, TCP_NLA_SNDQ_SIZE, tp->write_seq - tp->snd_una);
4289 	nla_put_u8(stats, TCP_NLA_CA_STATE, inet_csk(sk)->icsk_ca_state);
4290 
4291 	nla_put_u64_64bit(stats, TCP_NLA_BYTES_SENT, tp->bytes_sent,
4292 			  TCP_NLA_PAD);
4293 	nla_put_u64_64bit(stats, TCP_NLA_BYTES_RETRANS, tp->bytes_retrans,
4294 			  TCP_NLA_PAD);
4295 	nla_put_u32(stats, TCP_NLA_DSACK_DUPS, tp->dsack_dups);
4296 	nla_put_u32(stats, TCP_NLA_REORD_SEEN, tp->reord_seen);
4297 	nla_put_u32(stats, TCP_NLA_SRTT, tp->srtt_us >> 3);
4298 	nla_put_u16(stats, TCP_NLA_TIMEOUT_REHASH, tp->timeout_rehash);
4299 	nla_put_u32(stats, TCP_NLA_BYTES_NOTSENT,
4300 		    max_t(int, 0, tp->write_seq - tp->snd_nxt));
4301 	nla_put_u64_64bit(stats, TCP_NLA_EDT, orig_skb->skb_mstamp_ns,
4302 			  TCP_NLA_PAD);
4303 	if (ack_skb)
4304 		nla_put_u8(stats, TCP_NLA_TTL,
4305 			   tcp_skb_ttl_or_hop_limit(ack_skb));
4306 
4307 	nla_put_u32(stats, TCP_NLA_REHASH, tp->plb_rehash + tp->timeout_rehash);
4308 	return stats;
4309 }
4310 
do_tcp_getsockopt(struct sock * sk,int level,int optname,sockptr_t optval,sockptr_t optlen)4311 int do_tcp_getsockopt(struct sock *sk, int level,
4312 		      int optname, sockptr_t optval, sockptr_t optlen)
4313 {
4314 	struct inet_connection_sock *icsk = inet_csk(sk);
4315 	struct tcp_sock *tp = tcp_sk(sk);
4316 	struct net *net = sock_net(sk);
4317 	int val, len;
4318 
4319 	if (copy_from_sockptr(&len, optlen, sizeof(int)))
4320 		return -EFAULT;
4321 
4322 	if (len < 0)
4323 		return -EINVAL;
4324 
4325 	len = min_t(unsigned int, len, sizeof(int));
4326 
4327 	switch (optname) {
4328 	case TCP_MAXSEG:
4329 		val = tp->mss_cache;
4330 		if (tp->rx_opt.user_mss &&
4331 		    ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
4332 			val = tp->rx_opt.user_mss;
4333 		if (tp->repair)
4334 			val = tp->rx_opt.mss_clamp;
4335 		break;
4336 	case TCP_NODELAY:
4337 		val = !!(tp->nonagle&TCP_NAGLE_OFF);
4338 		break;
4339 	case TCP_CORK:
4340 		val = !!(tp->nonagle&TCP_NAGLE_CORK);
4341 		break;
4342 	case TCP_KEEPIDLE:
4343 		val = keepalive_time_when(tp) / HZ;
4344 		break;
4345 	case TCP_KEEPINTVL:
4346 		val = keepalive_intvl_when(tp) / HZ;
4347 		break;
4348 	case TCP_KEEPCNT:
4349 		val = keepalive_probes(tp);
4350 		break;
4351 	case TCP_SYNCNT:
4352 		val = READ_ONCE(icsk->icsk_syn_retries) ? :
4353 			READ_ONCE(net->ipv4.sysctl_tcp_syn_retries);
4354 		break;
4355 	case TCP_LINGER2:
4356 		val = READ_ONCE(tp->linger2);
4357 		if (val >= 0)
4358 			val = (val ? : READ_ONCE(net->ipv4.sysctl_tcp_fin_timeout)) / HZ;
4359 		break;
4360 	case TCP_DEFER_ACCEPT:
4361 		val = READ_ONCE(icsk->icsk_accept_queue.rskq_defer_accept);
4362 		val = retrans_to_secs(val, TCP_TIMEOUT_INIT / HZ,
4363 				      TCP_RTO_MAX / HZ);
4364 		break;
4365 	case TCP_WINDOW_CLAMP:
4366 		val = READ_ONCE(tp->window_clamp);
4367 		break;
4368 	case TCP_INFO: {
4369 		struct tcp_info info;
4370 
4371 		if (copy_from_sockptr(&len, optlen, sizeof(int)))
4372 			return -EFAULT;
4373 
4374 		tcp_get_info(sk, &info);
4375 
4376 		len = min_t(unsigned int, len, sizeof(info));
4377 		if (copy_to_sockptr(optlen, &len, sizeof(int)))
4378 			return -EFAULT;
4379 		if (copy_to_sockptr(optval, &info, len))
4380 			return -EFAULT;
4381 		return 0;
4382 	}
4383 	case TCP_CC_INFO: {
4384 		const struct tcp_congestion_ops *ca_ops;
4385 		union tcp_cc_info info;
4386 		size_t sz = 0;
4387 		int attr;
4388 
4389 		if (copy_from_sockptr(&len, optlen, sizeof(int)))
4390 			return -EFAULT;
4391 
4392 		ca_ops = icsk->icsk_ca_ops;
4393 		if (ca_ops && ca_ops->get_info)
4394 			sz = ca_ops->get_info(sk, ~0U, &attr, &info);
4395 
4396 		len = min_t(unsigned int, len, sz);
4397 		if (copy_to_sockptr(optlen, &len, sizeof(int)))
4398 			return -EFAULT;
4399 		if (copy_to_sockptr(optval, &info, len))
4400 			return -EFAULT;
4401 		return 0;
4402 	}
4403 	case TCP_QUICKACK:
4404 		val = !inet_csk_in_pingpong_mode(sk);
4405 		break;
4406 
4407 	case TCP_CONGESTION:
4408 		if (copy_from_sockptr(&len, optlen, sizeof(int)))
4409 			return -EFAULT;
4410 		len = min_t(unsigned int, len, TCP_CA_NAME_MAX);
4411 		if (copy_to_sockptr(optlen, &len, sizeof(int)))
4412 			return -EFAULT;
4413 		if (copy_to_sockptr(optval, icsk->icsk_ca_ops->name, len))
4414 			return -EFAULT;
4415 		return 0;
4416 
4417 	case TCP_ULP:
4418 		if (copy_from_sockptr(&len, optlen, sizeof(int)))
4419 			return -EFAULT;
4420 		len = min_t(unsigned int, len, TCP_ULP_NAME_MAX);
4421 		if (!icsk->icsk_ulp_ops) {
4422 			len = 0;
4423 			if (copy_to_sockptr(optlen, &len, sizeof(int)))
4424 				return -EFAULT;
4425 			return 0;
4426 		}
4427 		if (copy_to_sockptr(optlen, &len, sizeof(int)))
4428 			return -EFAULT;
4429 		if (copy_to_sockptr(optval, icsk->icsk_ulp_ops->name, len))
4430 			return -EFAULT;
4431 		return 0;
4432 
4433 	case TCP_FASTOPEN_KEY: {
4434 		u64 key[TCP_FASTOPEN_KEY_BUF_LENGTH / sizeof(u64)];
4435 		unsigned int key_len;
4436 
4437 		if (copy_from_sockptr(&len, optlen, sizeof(int)))
4438 			return -EFAULT;
4439 
4440 		key_len = tcp_fastopen_get_cipher(net, icsk, key) *
4441 				TCP_FASTOPEN_KEY_LENGTH;
4442 		len = min_t(unsigned int, len, key_len);
4443 		if (copy_to_sockptr(optlen, &len, sizeof(int)))
4444 			return -EFAULT;
4445 		if (copy_to_sockptr(optval, key, len))
4446 			return -EFAULT;
4447 		return 0;
4448 	}
4449 	case TCP_THIN_LINEAR_TIMEOUTS:
4450 		val = tp->thin_lto;
4451 		break;
4452 
4453 	case TCP_THIN_DUPACK:
4454 		val = 0;
4455 		break;
4456 
4457 	case TCP_REPAIR:
4458 		val = tp->repair;
4459 		break;
4460 
4461 	case TCP_REPAIR_QUEUE:
4462 		if (tp->repair)
4463 			val = tp->repair_queue;
4464 		else
4465 			return -EINVAL;
4466 		break;
4467 
4468 	case TCP_REPAIR_WINDOW: {
4469 		struct tcp_repair_window opt;
4470 
4471 		if (copy_from_sockptr(&len, optlen, sizeof(int)))
4472 			return -EFAULT;
4473 
4474 		if (len != sizeof(opt))
4475 			return -EINVAL;
4476 
4477 		if (!tp->repair)
4478 			return -EPERM;
4479 
4480 		opt.snd_wl1	= tp->snd_wl1;
4481 		opt.snd_wnd	= tp->snd_wnd;
4482 		opt.max_window	= tp->max_window;
4483 		opt.rcv_wnd	= tp->rcv_wnd;
4484 		opt.rcv_wup	= tp->rcv_wup;
4485 
4486 		if (copy_to_sockptr(optval, &opt, len))
4487 			return -EFAULT;
4488 		return 0;
4489 	}
4490 	case TCP_QUEUE_SEQ:
4491 		if (tp->repair_queue == TCP_SEND_QUEUE)
4492 			val = tp->write_seq;
4493 		else if (tp->repair_queue == TCP_RECV_QUEUE)
4494 			val = tp->rcv_nxt;
4495 		else
4496 			return -EINVAL;
4497 		break;
4498 
4499 	case TCP_USER_TIMEOUT:
4500 		val = READ_ONCE(icsk->icsk_user_timeout);
4501 		break;
4502 
4503 	case TCP_FASTOPEN:
4504 		val = READ_ONCE(icsk->icsk_accept_queue.fastopenq.max_qlen);
4505 		break;
4506 
4507 	case TCP_FASTOPEN_CONNECT:
4508 		val = tp->fastopen_connect;
4509 		break;
4510 
4511 	case TCP_FASTOPEN_NO_COOKIE:
4512 		val = tp->fastopen_no_cookie;
4513 		break;
4514 
4515 	case TCP_TX_DELAY:
4516 		val = READ_ONCE(tp->tcp_tx_delay);
4517 		break;
4518 
4519 	case TCP_TIMESTAMP:
4520 		val = tcp_clock_ts(tp->tcp_usec_ts) + READ_ONCE(tp->tsoffset);
4521 		if (tp->tcp_usec_ts)
4522 			val |= 1;
4523 		else
4524 			val &= ~1;
4525 		break;
4526 	case TCP_NOTSENT_LOWAT:
4527 		val = READ_ONCE(tp->notsent_lowat);
4528 		break;
4529 	case TCP_INQ:
4530 		val = tp->recvmsg_inq;
4531 		break;
4532 	case TCP_SAVE_SYN:
4533 		val = tp->save_syn;
4534 		break;
4535 	case TCP_SAVED_SYN: {
4536 		if (copy_from_sockptr(&len, optlen, sizeof(int)))
4537 			return -EFAULT;
4538 
4539 		sockopt_lock_sock(sk);
4540 		if (tp->saved_syn) {
4541 			if (len < tcp_saved_syn_len(tp->saved_syn)) {
4542 				len = tcp_saved_syn_len(tp->saved_syn);
4543 				if (copy_to_sockptr(optlen, &len, sizeof(int))) {
4544 					sockopt_release_sock(sk);
4545 					return -EFAULT;
4546 				}
4547 				sockopt_release_sock(sk);
4548 				return -EINVAL;
4549 			}
4550 			len = tcp_saved_syn_len(tp->saved_syn);
4551 			if (copy_to_sockptr(optlen, &len, sizeof(int))) {
4552 				sockopt_release_sock(sk);
4553 				return -EFAULT;
4554 			}
4555 			if (copy_to_sockptr(optval, tp->saved_syn->data, len)) {
4556 				sockopt_release_sock(sk);
4557 				return -EFAULT;
4558 			}
4559 			tcp_saved_syn_free(tp);
4560 			sockopt_release_sock(sk);
4561 		} else {
4562 			sockopt_release_sock(sk);
4563 			len = 0;
4564 			if (copy_to_sockptr(optlen, &len, sizeof(int)))
4565 				return -EFAULT;
4566 		}
4567 		return 0;
4568 	}
4569 #ifdef CONFIG_MMU
4570 	case TCP_ZEROCOPY_RECEIVE: {
4571 		struct scm_timestamping_internal tss;
4572 		struct tcp_zerocopy_receive zc = {};
4573 		int err;
4574 
4575 		if (copy_from_sockptr(&len, optlen, sizeof(int)))
4576 			return -EFAULT;
4577 		if (len < 0 ||
4578 		    len < offsetofend(struct tcp_zerocopy_receive, length))
4579 			return -EINVAL;
4580 		if (unlikely(len > sizeof(zc))) {
4581 			err = check_zeroed_sockptr(optval, sizeof(zc),
4582 						   len - sizeof(zc));
4583 			if (err < 1)
4584 				return err == 0 ? -EINVAL : err;
4585 			len = sizeof(zc);
4586 			if (copy_to_sockptr(optlen, &len, sizeof(int)))
4587 				return -EFAULT;
4588 		}
4589 		if (copy_from_sockptr(&zc, optval, len))
4590 			return -EFAULT;
4591 		if (zc.reserved)
4592 			return -EINVAL;
4593 		if (zc.msg_flags &  ~(TCP_VALID_ZC_MSG_FLAGS))
4594 			return -EINVAL;
4595 		sockopt_lock_sock(sk);
4596 		err = tcp_zerocopy_receive(sk, &zc, &tss);
4597 		err = BPF_CGROUP_RUN_PROG_GETSOCKOPT_KERN(sk, level, optname,
4598 							  &zc, &len, err);
4599 		sockopt_release_sock(sk);
4600 		if (len >= offsetofend(struct tcp_zerocopy_receive, msg_flags))
4601 			goto zerocopy_rcv_cmsg;
4602 		switch (len) {
4603 		case offsetofend(struct tcp_zerocopy_receive, msg_flags):
4604 			goto zerocopy_rcv_cmsg;
4605 		case offsetofend(struct tcp_zerocopy_receive, msg_controllen):
4606 		case offsetofend(struct tcp_zerocopy_receive, msg_control):
4607 		case offsetofend(struct tcp_zerocopy_receive, flags):
4608 		case offsetofend(struct tcp_zerocopy_receive, copybuf_len):
4609 		case offsetofend(struct tcp_zerocopy_receive, copybuf_address):
4610 		case offsetofend(struct tcp_zerocopy_receive, err):
4611 			goto zerocopy_rcv_sk_err;
4612 		case offsetofend(struct tcp_zerocopy_receive, inq):
4613 			goto zerocopy_rcv_inq;
4614 		case offsetofend(struct tcp_zerocopy_receive, length):
4615 		default:
4616 			goto zerocopy_rcv_out;
4617 		}
4618 zerocopy_rcv_cmsg:
4619 		if (zc.msg_flags & TCP_CMSG_TS)
4620 			tcp_zc_finalize_rx_tstamp(sk, &zc, &tss);
4621 		else
4622 			zc.msg_flags = 0;
4623 zerocopy_rcv_sk_err:
4624 		if (!err)
4625 			zc.err = sock_error(sk);
4626 zerocopy_rcv_inq:
4627 		zc.inq = tcp_inq_hint(sk);
4628 zerocopy_rcv_out:
4629 		if (!err && copy_to_sockptr(optval, &zc, len))
4630 			err = -EFAULT;
4631 		return err;
4632 	}
4633 #endif
4634 	case TCP_AO_REPAIR:
4635 		if (!tcp_can_repair_sock(sk))
4636 			return -EPERM;
4637 		return tcp_ao_get_repair(sk, optval, optlen);
4638 	case TCP_AO_GET_KEYS:
4639 	case TCP_AO_INFO: {
4640 		int err;
4641 
4642 		sockopt_lock_sock(sk);
4643 		if (optname == TCP_AO_GET_KEYS)
4644 			err = tcp_ao_get_mkts(sk, optval, optlen);
4645 		else
4646 			err = tcp_ao_get_sock_info(sk, optval, optlen);
4647 		sockopt_release_sock(sk);
4648 
4649 		return err;
4650 	}
4651 	case TCP_IS_MPTCP:
4652 		val = 0;
4653 		break;
4654 	default:
4655 		return -ENOPROTOOPT;
4656 	}
4657 
4658 	if (copy_to_sockptr(optlen, &len, sizeof(int)))
4659 		return -EFAULT;
4660 	if (copy_to_sockptr(optval, &val, len))
4661 		return -EFAULT;
4662 	return 0;
4663 }
4664 
tcp_bpf_bypass_getsockopt(int level,int optname)4665 bool tcp_bpf_bypass_getsockopt(int level, int optname)
4666 {
4667 	/* TCP do_tcp_getsockopt has optimized getsockopt implementation
4668 	 * to avoid extra socket lock for TCP_ZEROCOPY_RECEIVE.
4669 	 */
4670 	if (level == SOL_TCP && optname == TCP_ZEROCOPY_RECEIVE)
4671 		return true;
4672 
4673 	return false;
4674 }
4675 EXPORT_SYMBOL(tcp_bpf_bypass_getsockopt);
4676 
tcp_getsockopt(struct sock * sk,int level,int optname,char __user * optval,int __user * optlen)4677 int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval,
4678 		   int __user *optlen)
4679 {
4680 	struct inet_connection_sock *icsk = inet_csk(sk);
4681 
4682 	if (level != SOL_TCP)
4683 		/* Paired with WRITE_ONCE() in do_ipv6_setsockopt() and tcp_v6_connect() */
4684 		return READ_ONCE(icsk->icsk_af_ops)->getsockopt(sk, level, optname,
4685 								optval, optlen);
4686 	return do_tcp_getsockopt(sk, level, optname, USER_SOCKPTR(optval),
4687 				 USER_SOCKPTR(optlen));
4688 }
4689 EXPORT_SYMBOL(tcp_getsockopt);
4690 
4691 #ifdef CONFIG_TCP_MD5SIG
4692 int tcp_md5_sigpool_id = -1;
4693 EXPORT_SYMBOL_GPL(tcp_md5_sigpool_id);
4694 
tcp_md5_alloc_sigpool(void)4695 int tcp_md5_alloc_sigpool(void)
4696 {
4697 	size_t scratch_size;
4698 	int ret;
4699 
4700 	scratch_size = sizeof(union tcp_md5sum_block) + sizeof(struct tcphdr);
4701 	ret = tcp_sigpool_alloc_ahash("md5", scratch_size);
4702 	if (ret >= 0) {
4703 		/* As long as any md5 sigpool was allocated, the return
4704 		 * id would stay the same. Re-write the id only for the case
4705 		 * when previously all MD5 keys were deleted and this call
4706 		 * allocates the first MD5 key, which may return a different
4707 		 * sigpool id than was used previously.
4708 		 */
4709 		WRITE_ONCE(tcp_md5_sigpool_id, ret); /* Avoids the compiler potentially being smart here */
4710 		return 0;
4711 	}
4712 	return ret;
4713 }
4714 
tcp_md5_release_sigpool(void)4715 void tcp_md5_release_sigpool(void)
4716 {
4717 	tcp_sigpool_release(READ_ONCE(tcp_md5_sigpool_id));
4718 }
4719 
tcp_md5_add_sigpool(void)4720 void tcp_md5_add_sigpool(void)
4721 {
4722 	tcp_sigpool_get(READ_ONCE(tcp_md5_sigpool_id));
4723 }
4724 
tcp_md5_hash_key(struct tcp_sigpool * hp,const struct tcp_md5sig_key * key)4725 int tcp_md5_hash_key(struct tcp_sigpool *hp,
4726 		     const struct tcp_md5sig_key *key)
4727 {
4728 	u8 keylen = READ_ONCE(key->keylen); /* paired with WRITE_ONCE() in tcp_md5_do_add */
4729 	struct scatterlist sg;
4730 
4731 	sg_init_one(&sg, key->key, keylen);
4732 	ahash_request_set_crypt(hp->req, &sg, NULL, keylen);
4733 
4734 	/* We use data_race() because tcp_md5_do_add() might change
4735 	 * key->key under us
4736 	 */
4737 	return data_race(crypto_ahash_update(hp->req));
4738 }
4739 EXPORT_SYMBOL(tcp_md5_hash_key);
4740 
4741 /* Called with rcu_read_lock() */
4742 static 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 l3index,const __u8 * hash_location)4743 tcp_inbound_md5_hash(const struct sock *sk, const struct sk_buff *skb,
4744 		     const void *saddr, const void *daddr,
4745 		     int family, int l3index, const __u8 *hash_location)
4746 {
4747 	/* This gets called for each TCP segment that has TCP-MD5 option.
4748 	 * We have 3 drop cases:
4749 	 * o No MD5 hash and one expected.
4750 	 * o MD5 hash and we're not expecting one.
4751 	 * o MD5 hash and its wrong.
4752 	 */
4753 	const struct tcp_sock *tp = tcp_sk(sk);
4754 	struct tcp_md5sig_key *key;
4755 	u8 newhash[16];
4756 	int genhash;
4757 
4758 	key = tcp_md5_do_lookup(sk, l3index, saddr, family);
4759 
4760 	if (!key && hash_location) {
4761 		NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMD5UNEXPECTED);
4762 		trace_tcp_hash_md5_unexpected(sk, skb);
4763 		return SKB_DROP_REASON_TCP_MD5UNEXPECTED;
4764 	}
4765 
4766 	/* Check the signature.
4767 	 * To support dual stack listeners, we need to handle
4768 	 * IPv4-mapped case.
4769 	 */
4770 	if (family == AF_INET)
4771 		genhash = tcp_v4_md5_hash_skb(newhash, key, NULL, skb);
4772 	else
4773 		genhash = tp->af_specific->calc_md5_hash(newhash, key,
4774 							 NULL, skb);
4775 	if (genhash || memcmp(hash_location, newhash, 16) != 0) {
4776 		NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMD5FAILURE);
4777 		trace_tcp_hash_md5_mismatch(sk, skb);
4778 		return SKB_DROP_REASON_TCP_MD5FAILURE;
4779 	}
4780 	return SKB_NOT_DROPPED_YET;
4781 }
4782 #else
4783 static inline 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 l3index,const __u8 * hash_location)4784 tcp_inbound_md5_hash(const struct sock *sk, const struct sk_buff *skb,
4785 		     const void *saddr, const void *daddr,
4786 		     int family, int l3index, const __u8 *hash_location)
4787 {
4788 	return SKB_NOT_DROPPED_YET;
4789 }
4790 
4791 #endif
4792 
4793 /* Called with rcu_read_lock() */
4794 enum skb_drop_reason
tcp_inbound_hash(struct sock * sk,const struct request_sock * req,const struct sk_buff * skb,const void * saddr,const void * daddr,int family,int dif,int sdif)4795 tcp_inbound_hash(struct sock *sk, const struct request_sock *req,
4796 		 const struct sk_buff *skb,
4797 		 const void *saddr, const void *daddr,
4798 		 int family, int dif, int sdif)
4799 {
4800 	const struct tcphdr *th = tcp_hdr(skb);
4801 	const struct tcp_ao_hdr *aoh;
4802 	const __u8 *md5_location;
4803 	int l3index;
4804 
4805 	/* Invalid option or two times meet any of auth options */
4806 	if (tcp_parse_auth_options(th, &md5_location, &aoh)) {
4807 		trace_tcp_hash_bad_header(sk, skb);
4808 		return SKB_DROP_REASON_TCP_AUTH_HDR;
4809 	}
4810 
4811 	if (req) {
4812 		if (tcp_rsk_used_ao(req) != !!aoh) {
4813 			u8 keyid, rnext, maclen;
4814 
4815 			if (aoh) {
4816 				keyid = aoh->keyid;
4817 				rnext = aoh->rnext_keyid;
4818 				maclen = tcp_ao_hdr_maclen(aoh);
4819 			} else {
4820 				keyid = rnext = maclen = 0;
4821 			}
4822 
4823 			NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPAOBAD);
4824 			trace_tcp_ao_handshake_failure(sk, skb, keyid, rnext, maclen);
4825 			return SKB_DROP_REASON_TCP_AOFAILURE;
4826 		}
4827 	}
4828 
4829 	/* sdif set, means packet ingressed via a device
4830 	 * in an L3 domain and dif is set to the l3mdev
4831 	 */
4832 	l3index = sdif ? dif : 0;
4833 
4834 	/* Fast path: unsigned segments */
4835 	if (likely(!md5_location && !aoh)) {
4836 		/* Drop if there's TCP-MD5 or TCP-AO key with any rcvid/sndid
4837 		 * for the remote peer. On TCP-AO established connection
4838 		 * the last key is impossible to remove, so there's
4839 		 * always at least one current_key.
4840 		 */
4841 		if (tcp_ao_required(sk, saddr, family, l3index, true)) {
4842 			trace_tcp_hash_ao_required(sk, skb);
4843 			return SKB_DROP_REASON_TCP_AONOTFOUND;
4844 		}
4845 		if (unlikely(tcp_md5_do_lookup(sk, l3index, saddr, family))) {
4846 			NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMD5NOTFOUND);
4847 			trace_tcp_hash_md5_required(sk, skb);
4848 			return SKB_DROP_REASON_TCP_MD5NOTFOUND;
4849 		}
4850 		return SKB_NOT_DROPPED_YET;
4851 	}
4852 
4853 	if (aoh)
4854 		return tcp_inbound_ao_hash(sk, skb, family, req, l3index, aoh);
4855 
4856 	return tcp_inbound_md5_hash(sk, skb, saddr, daddr, family,
4857 				    l3index, md5_location);
4858 }
4859 EXPORT_SYMBOL_GPL(tcp_inbound_hash);
4860 
tcp_done(struct sock * sk)4861 void tcp_done(struct sock *sk)
4862 {
4863 	struct request_sock *req;
4864 
4865 	/* We might be called with a new socket, after
4866 	 * inet_csk_prepare_forced_close() has been called
4867 	 * so we can not use lockdep_sock_is_held(sk)
4868 	 */
4869 	req = rcu_dereference_protected(tcp_sk(sk)->fastopen_rsk, 1);
4870 
4871 	if (sk->sk_state == TCP_SYN_SENT || sk->sk_state == TCP_SYN_RECV)
4872 		TCP_INC_STATS(sock_net(sk), TCP_MIB_ATTEMPTFAILS);
4873 
4874 	tcp_set_state(sk, TCP_CLOSE);
4875 	tcp_clear_xmit_timers(sk);
4876 	if (req)
4877 		reqsk_fastopen_remove(sk, req, false);
4878 
4879 	WRITE_ONCE(sk->sk_shutdown, SHUTDOWN_MASK);
4880 
4881 	if (!sock_flag(sk, SOCK_DEAD))
4882 		sk->sk_state_change(sk);
4883 	else
4884 		inet_csk_destroy_sock(sk);
4885 }
4886 EXPORT_SYMBOL_GPL(tcp_done);
4887 
tcp_abort(struct sock * sk,int err)4888 int tcp_abort(struct sock *sk, int err)
4889 {
4890 	int state = inet_sk_state_load(sk);
4891 
4892 	if (state == TCP_NEW_SYN_RECV) {
4893 		struct request_sock *req = inet_reqsk(sk);
4894 
4895 		local_bh_disable();
4896 		inet_csk_reqsk_queue_drop(req->rsk_listener, req);
4897 		local_bh_enable();
4898 		return 0;
4899 	}
4900 	if (state == TCP_TIME_WAIT) {
4901 		struct inet_timewait_sock *tw = inet_twsk(sk);
4902 
4903 		refcount_inc(&tw->tw_refcnt);
4904 		local_bh_disable();
4905 		inet_twsk_deschedule_put(tw);
4906 		local_bh_enable();
4907 		return 0;
4908 	}
4909 
4910 	/* BPF context ensures sock locking. */
4911 	if (!has_current_bpf_ctx())
4912 		/* Don't race with userspace socket closes such as tcp_close. */
4913 		lock_sock(sk);
4914 
4915 	/* Avoid closing the same socket twice. */
4916 	if (sk->sk_state == TCP_CLOSE) {
4917 		if (!has_current_bpf_ctx())
4918 			release_sock(sk);
4919 		return -ENOENT;
4920 	}
4921 
4922 	if (sk->sk_state == TCP_LISTEN) {
4923 		tcp_set_state(sk, TCP_CLOSE);
4924 		inet_csk_listen_stop(sk);
4925 	}
4926 
4927 	/* Don't race with BH socket closes such as inet_csk_listen_stop. */
4928 	local_bh_disable();
4929 	bh_lock_sock(sk);
4930 
4931 	if (tcp_need_reset(sk->sk_state))
4932 		tcp_send_active_reset(sk, GFP_ATOMIC,
4933 				      SK_RST_REASON_TCP_STATE);
4934 	tcp_done_with_error(sk, err);
4935 
4936 	bh_unlock_sock(sk);
4937 	local_bh_enable();
4938 	if (!has_current_bpf_ctx())
4939 		release_sock(sk);
4940 	return 0;
4941 }
4942 EXPORT_SYMBOL_GPL(tcp_abort);
4943 
4944 extern struct tcp_congestion_ops tcp_reno;
4945 
4946 static __initdata unsigned long thash_entries;
set_thash_entries(char * str)4947 static int __init set_thash_entries(char *str)
4948 {
4949 	ssize_t ret;
4950 
4951 	if (!str)
4952 		return 0;
4953 
4954 	ret = kstrtoul(str, 0, &thash_entries);
4955 	if (ret)
4956 		return 0;
4957 
4958 	return 1;
4959 }
4960 __setup("thash_entries=", set_thash_entries);
4961 
tcp_init_mem(void)4962 static void __init tcp_init_mem(void)
4963 {
4964 	unsigned long limit = nr_free_buffer_pages() / 16;
4965 
4966 	limit = max(limit, 128UL);
4967 	sysctl_tcp_mem[0] = limit / 4 * 3;		/* 4.68 % */
4968 	sysctl_tcp_mem[1] = limit;			/* 6.25 % */
4969 	sysctl_tcp_mem[2] = sysctl_tcp_mem[0] * 2;	/* 9.37 % */
4970 }
4971 
tcp_struct_check(void)4972 static void __init tcp_struct_check(void)
4973 {
4974 	/* TX read-mostly hotpath cache lines */
4975 	CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_tx, max_window);
4976 	CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_tx, rcv_ssthresh);
4977 	CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_tx, reordering);
4978 	CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_tx, notsent_lowat);
4979 	CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_tx, gso_segs);
4980 	CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_tx, lost_skb_hint);
4981 	CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_tx, retransmit_skb_hint);
4982 	CACHELINE_ASSERT_GROUP_SIZE(struct tcp_sock, tcp_sock_read_tx, 40);
4983 
4984 	/* TXRX read-mostly hotpath cache lines */
4985 	CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_txrx, tsoffset);
4986 	CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_txrx, snd_wnd);
4987 	CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_txrx, mss_cache);
4988 	CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_txrx, snd_cwnd);
4989 	CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_txrx, prr_out);
4990 	CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_txrx, lost_out);
4991 	CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_txrx, sacked_out);
4992 	CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_txrx, scaling_ratio);
4993 	CACHELINE_ASSERT_GROUP_SIZE(struct tcp_sock, tcp_sock_read_txrx, 32);
4994 
4995 	/* RX read-mostly hotpath cache lines */
4996 	CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_rx, copied_seq);
4997 	CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_rx, rcv_tstamp);
4998 	CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_rx, snd_wl1);
4999 	CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_rx, tlp_high_seq);
5000 	CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_rx, rttvar_us);
5001 	CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_rx, retrans_out);
5002 	CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_rx, advmss);
5003 	CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_rx, urg_data);
5004 	CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_rx, lost);
5005 	CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_rx, rtt_min);
5006 	CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_rx, out_of_order_queue);
5007 	CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_rx, snd_ssthresh);
5008 	CACHELINE_ASSERT_GROUP_SIZE(struct tcp_sock, tcp_sock_read_rx, 69);
5009 
5010 	/* TX read-write hotpath cache lines */
5011 	CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, segs_out);
5012 	CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, data_segs_out);
5013 	CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, bytes_sent);
5014 	CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, snd_sml);
5015 	CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, chrono_start);
5016 	CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, chrono_stat);
5017 	CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, write_seq);
5018 	CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, pushed_seq);
5019 	CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, lsndtime);
5020 	CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, mdev_us);
5021 	CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, tcp_wstamp_ns);
5022 	CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, rtt_seq);
5023 	CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, tsorted_sent_queue);
5024 	CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, highest_sack);
5025 	CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, ecn_flags);
5026 	CACHELINE_ASSERT_GROUP_SIZE(struct tcp_sock, tcp_sock_write_tx, 89);
5027 
5028 	/* TXRX read-write hotpath cache lines */
5029 	CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_txrx, pred_flags);
5030 	CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_txrx, tcp_clock_cache);
5031 	CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_txrx, tcp_mstamp);
5032 	CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_txrx, rcv_nxt);
5033 	CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_txrx, snd_nxt);
5034 	CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_txrx, snd_una);
5035 	CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_txrx, window_clamp);
5036 	CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_txrx, srtt_us);
5037 	CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_txrx, packets_out);
5038 	CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_txrx, snd_up);
5039 	CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_txrx, delivered);
5040 	CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_txrx, delivered_ce);
5041 	CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_txrx, app_limited);
5042 	CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_txrx, rcv_wnd);
5043 	CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_txrx, rx_opt);
5044 
5045 	/* 32bit arches with 8byte alignment on u64 fields might need padding
5046 	 * before tcp_clock_cache.
5047 	 */
5048 	CACHELINE_ASSERT_GROUP_SIZE(struct tcp_sock, tcp_sock_write_txrx, 92 + 4);
5049 
5050 	/* RX read-write hotpath cache lines */
5051 	CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_rx, bytes_received);
5052 	CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_rx, segs_in);
5053 	CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_rx, data_segs_in);
5054 	CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_rx, rcv_wup);
5055 	CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_rx, max_packets_out);
5056 	CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_rx, cwnd_usage_seq);
5057 	CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_rx, rate_delivered);
5058 	CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_rx, rate_interval_us);
5059 	CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_rx, rcv_rtt_last_tsecr);
5060 	CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_rx, first_tx_mstamp);
5061 	CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_rx, delivered_mstamp);
5062 	CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_rx, bytes_acked);
5063 	CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_rx, rcv_rtt_est);
5064 	CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_rx, rcvq_space);
5065 	CACHELINE_ASSERT_GROUP_SIZE(struct tcp_sock, tcp_sock_write_rx, 99);
5066 }
5067 
tcp_init(void)5068 void __init tcp_init(void)
5069 {
5070 	int max_rshare, max_wshare, cnt;
5071 	unsigned long limit;
5072 	unsigned int i;
5073 
5074 	BUILD_BUG_ON(TCP_MIN_SND_MSS <= MAX_TCP_OPTION_SPACE);
5075 	BUILD_BUG_ON(sizeof(struct tcp_skb_cb) >
5076 		     sizeof_field(struct sk_buff, cb));
5077 
5078 	tcp_struct_check();
5079 
5080 	percpu_counter_init(&tcp_sockets_allocated, 0, GFP_KERNEL);
5081 
5082 	timer_setup(&tcp_orphan_timer, tcp_orphan_update, TIMER_DEFERRABLE);
5083 	mod_timer(&tcp_orphan_timer, jiffies + TCP_ORPHAN_TIMER_PERIOD);
5084 
5085 	inet_hashinfo2_init(&tcp_hashinfo, "tcp_listen_portaddr_hash",
5086 			    thash_entries, 21,  /* one slot per 2 MB*/
5087 			    0, 64 * 1024);
5088 	tcp_hashinfo.bind_bucket_cachep =
5089 		kmem_cache_create("tcp_bind_bucket",
5090 				  sizeof(struct inet_bind_bucket), 0,
5091 				  SLAB_HWCACHE_ALIGN | SLAB_PANIC |
5092 				  SLAB_ACCOUNT,
5093 				  NULL);
5094 	tcp_hashinfo.bind2_bucket_cachep =
5095 		kmem_cache_create("tcp_bind2_bucket",
5096 				  sizeof(struct inet_bind2_bucket), 0,
5097 				  SLAB_HWCACHE_ALIGN | SLAB_PANIC |
5098 				  SLAB_ACCOUNT,
5099 				  NULL);
5100 
5101 	/* Size and allocate the main established and bind bucket
5102 	 * hash tables.
5103 	 *
5104 	 * The methodology is similar to that of the buffer cache.
5105 	 */
5106 	tcp_hashinfo.ehash =
5107 		alloc_large_system_hash("TCP established",
5108 					sizeof(struct inet_ehash_bucket),
5109 					thash_entries,
5110 					17, /* one slot per 128 KB of memory */
5111 					0,
5112 					NULL,
5113 					&tcp_hashinfo.ehash_mask,
5114 					0,
5115 					thash_entries ? 0 : 512 * 1024);
5116 	for (i = 0; i <= tcp_hashinfo.ehash_mask; i++)
5117 		INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].chain, i);
5118 
5119 	if (inet_ehash_locks_alloc(&tcp_hashinfo))
5120 		panic("TCP: failed to alloc ehash_locks");
5121 	tcp_hashinfo.bhash =
5122 		alloc_large_system_hash("TCP bind",
5123 					2 * sizeof(struct inet_bind_hashbucket),
5124 					tcp_hashinfo.ehash_mask + 1,
5125 					17, /* one slot per 128 KB of memory */
5126 					0,
5127 					&tcp_hashinfo.bhash_size,
5128 					NULL,
5129 					0,
5130 					64 * 1024);
5131 	tcp_hashinfo.bhash_size = 1U << tcp_hashinfo.bhash_size;
5132 	tcp_hashinfo.bhash2 = tcp_hashinfo.bhash + tcp_hashinfo.bhash_size;
5133 	for (i = 0; i < tcp_hashinfo.bhash_size; i++) {
5134 		spin_lock_init(&tcp_hashinfo.bhash[i].lock);
5135 		INIT_HLIST_HEAD(&tcp_hashinfo.bhash[i].chain);
5136 		spin_lock_init(&tcp_hashinfo.bhash2[i].lock);
5137 		INIT_HLIST_HEAD(&tcp_hashinfo.bhash2[i].chain);
5138 	}
5139 
5140 	tcp_hashinfo.pernet = false;
5141 
5142 	cnt = tcp_hashinfo.ehash_mask + 1;
5143 	sysctl_tcp_max_orphans = cnt / 2;
5144 
5145 	tcp_init_mem();
5146 	/* Set per-socket limits to no more than 1/128 the pressure threshold */
5147 	limit = nr_free_buffer_pages() << (PAGE_SHIFT - 7);
5148 	max_wshare = min(4UL*1024*1024, limit);
5149 	max_rshare = min(6UL*1024*1024, limit);
5150 
5151 	init_net.ipv4.sysctl_tcp_wmem[0] = PAGE_SIZE;
5152 	init_net.ipv4.sysctl_tcp_wmem[1] = 16*1024;
5153 	init_net.ipv4.sysctl_tcp_wmem[2] = max(64*1024, max_wshare);
5154 
5155 	init_net.ipv4.sysctl_tcp_rmem[0] = PAGE_SIZE;
5156 	init_net.ipv4.sysctl_tcp_rmem[1] = 131072;
5157 	init_net.ipv4.sysctl_tcp_rmem[2] = max(131072, max_rshare);
5158 
5159 	pr_info("Hash tables configured (established %u bind %u)\n",
5160 		tcp_hashinfo.ehash_mask + 1, tcp_hashinfo.bhash_size);
5161 
5162 	tcp_v4_init();
5163 	tcp_metrics_init();
5164 	BUG_ON(tcp_register_congestion_control(&tcp_reno) != 0);
5165 	tcp_tasklet_init();
5166 	mptcp_init();
5167 }
5168