net: tcp: fix rx timestamp behavior for tcp_recvmsg
[linux-2.6-microblaze.git] / net / ipv4 / tcp.c
1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /*
3  * INET         An implementation of the TCP/IP protocol suite for the LINUX
4  *              operating system.  INET is implemented using the  BSD Socket
5  *              interface as the means of communication with the user level.
6  *
7  *              Implementation of the Transmission Control Protocol(TCP).
8  *
9  * Authors:     Ross Biro
10  *              Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
11  *              Mark Evans, <evansmp@uhura.aston.ac.uk>
12  *              Corey Minyard <wf-rch!minyard@relay.EU.net>
13  *              Florian La Roche, <flla@stud.uni-sb.de>
14  *              Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
15  *              Linus Torvalds, <torvalds@cs.helsinki.fi>
16  *              Alan Cox, <gw4pts@gw4pts.ampr.org>
17  *              Matthew Dillon, <dillon@apollo.west.oic.com>
18  *              Arnt Gulbrandsen, <agulbra@nvg.unit.no>
19  *              Jorge Cwik, <jorge@laser.satlink.net>
20  *
21  * Fixes:
22  *              Alan Cox        :       Numerous verify_area() calls
23  *              Alan Cox        :       Set the ACK bit on a reset
24  *              Alan Cox        :       Stopped it crashing if it closed while
25  *                                      sk->inuse=1 and was trying to connect
26  *                                      (tcp_err()).
27  *              Alan Cox        :       All icmp error handling was broken
28  *                                      pointers passed where wrong and the
29  *                                      socket was looked up backwards. Nobody
30  *                                      tested any icmp error code obviously.
31  *              Alan Cox        :       tcp_err() now handled properly. It
32  *                                      wakes people on errors. poll
33  *                                      behaves and the icmp error race
34  *                                      has gone by moving it into sock.c
35  *              Alan Cox        :       tcp_send_reset() fixed to work for
36  *                                      everything not just packets for
37  *                                      unknown sockets.
38  *              Alan Cox        :       tcp option processing.
39  *              Alan Cox        :       Reset tweaked (still not 100%) [Had
40  *                                      syn rule wrong]
41  *              Herp Rosmanith  :       More reset fixes
42  *              Alan Cox        :       No longer acks invalid rst frames.
43  *                                      Acking any kind of RST is right out.
44  *              Alan Cox        :       Sets an ignore me flag on an rst
45  *                                      receive otherwise odd bits of prattle
46  *                                      escape still
47  *              Alan Cox        :       Fixed another acking RST frame bug.
48  *                                      Should stop LAN workplace lockups.
49  *              Alan Cox        :       Some tidyups using the new skb list
50  *                                      facilities
51  *              Alan Cox        :       sk->keepopen now seems to work
52  *              Alan Cox        :       Pulls options out correctly on accepts
53  *              Alan Cox        :       Fixed assorted sk->rqueue->next errors
54  *              Alan Cox        :       PSH doesn't end a TCP read. Switched a
55  *                                      bit to skb ops.
56  *              Alan Cox        :       Tidied tcp_data to avoid a potential
57  *                                      nasty.
58  *              Alan Cox        :       Added some better commenting, as the
59  *                                      tcp is hard to follow
60  *              Alan Cox        :       Removed incorrect check for 20 * psh
61  *      Michael O'Reilly        :       ack < copied bug fix.
62  *      Johannes Stille         :       Misc tcp fixes (not all in yet).
63  *              Alan Cox        :       FIN with no memory -> CRASH
64  *              Alan Cox        :       Added socket option proto entries.
65  *                                      Also added awareness of them to accept.
66  *              Alan Cox        :       Added TCP options (SOL_TCP)
67  *              Alan Cox        :       Switched wakeup calls to callbacks,
68  *                                      so the kernel can layer network
69  *                                      sockets.
70  *              Alan Cox        :       Use ip_tos/ip_ttl settings.
71  *              Alan Cox        :       Handle FIN (more) properly (we hope).
72  *              Alan Cox        :       RST frames sent on unsynchronised
73  *                                      state ack error.
74  *              Alan Cox        :       Put in missing check for SYN bit.
75  *              Alan Cox        :       Added tcp_select_window() aka NET2E
76  *                                      window non shrink trick.
77  *              Alan Cox        :       Added a couple of small NET2E timer
78  *                                      fixes
79  *              Charles Hedrick :       TCP fixes
80  *              Toomas Tamm     :       TCP window fixes
81  *              Alan Cox        :       Small URG fix to rlogin ^C ack fight
82  *              Charles Hedrick :       Rewrote most of it to actually work
83  *              Linus           :       Rewrote tcp_read() and URG handling
84  *                                      completely
85  *              Gerhard Koerting:       Fixed some missing timer handling
86  *              Matthew Dillon  :       Reworked TCP machine states as per RFC
87  *              Gerhard Koerting:       PC/TCP workarounds
88  *              Adam Caldwell   :       Assorted timer/timing errors
89  *              Matthew Dillon  :       Fixed another RST bug
90  *              Alan Cox        :       Move to kernel side addressing changes.
91  *              Alan Cox        :       Beginning work on TCP fastpathing
92  *                                      (not yet usable)
93  *              Arnt Gulbrandsen:       Turbocharged tcp_check() routine.
94  *              Alan Cox        :       TCP fast path debugging
95  *              Alan Cox        :       Window clamping
96  *              Michael Riepe   :       Bug in tcp_check()
97  *              Matt Dillon     :       More TCP improvements and RST bug fixes
98  *              Matt Dillon     :       Yet more small nasties remove from the
99  *                                      TCP code (Be very nice to this man if
100  *                                      tcp finally works 100%) 8)
101  *              Alan Cox        :       BSD accept semantics.
102  *              Alan Cox        :       Reset on closedown bug.
103  *      Peter De Schrijver      :       ENOTCONN check missing in tcp_sendto().
104  *              Michael Pall    :       Handle poll() after URG properly in
105  *                                      all cases.
106  *              Michael Pall    :       Undo the last fix in tcp_read_urg()
107  *                                      (multi URG PUSH broke rlogin).
108  *              Michael Pall    :       Fix the multi URG PUSH problem in
109  *                                      tcp_readable(), poll() after URG
110  *                                      works now.
111  *              Michael Pall    :       recv(...,MSG_OOB) never blocks in the
112  *                                      BSD api.
113  *              Alan Cox        :       Changed the semantics of sk->socket to
114  *                                      fix a race and a signal problem with
115  *                                      accept() and async I/O.
116  *              Alan Cox        :       Relaxed the rules on tcp_sendto().
117  *              Yury Shevchuk   :       Really fixed accept() blocking problem.
118  *              Craig I. Hagan  :       Allow for BSD compatible TIME_WAIT for
119  *                                      clients/servers which listen in on
120  *                                      fixed ports.
121  *              Alan Cox        :       Cleaned the above up and shrank it to
122  *                                      a sensible code size.
123  *              Alan Cox        :       Self connect lockup fix.
124  *              Alan Cox        :       No connect to multicast.
125  *              Ross Biro       :       Close unaccepted children on master
126  *                                      socket close.
127  *              Alan Cox        :       Reset tracing code.
128  *              Alan Cox        :       Spurious resets on shutdown.
129  *              Alan Cox        :       Giant 15 minute/60 second timer error
130  *              Alan Cox        :       Small whoops in polling before an
131  *                                      accept.
132  *              Alan Cox        :       Kept the state trace facility since
133  *                                      it's handy for debugging.
134  *              Alan Cox        :       More reset handler fixes.
135  *              Alan Cox        :       Started rewriting the code based on
136  *                                      the RFC's for other useful protocol
137  *                                      references see: Comer, KA9Q NOS, and
138  *                                      for a reference on the difference
139  *                                      between specifications and how BSD
140  *                                      works see the 4.4lite source.
141  *              A.N.Kuznetsov   :       Don't time wait on completion of tidy
142  *                                      close.
143  *              Linus Torvalds  :       Fin/Shutdown & copied_seq changes.
144  *              Linus Torvalds  :       Fixed BSD port reuse to work first syn
145  *              Alan Cox        :       Reimplemented timers as per the RFC
146  *                                      and using multiple timers for sanity.
147  *              Alan Cox        :       Small bug fixes, and a lot of new
148  *                                      comments.
149  *              Alan Cox        :       Fixed dual reader crash by locking
150  *                                      the buffers (much like datagram.c)
151  *              Alan Cox        :       Fixed stuck sockets in probe. A probe
152  *                                      now gets fed up of retrying without
153  *                                      (even a no space) answer.
154  *              Alan Cox        :       Extracted closing code better
155  *              Alan Cox        :       Fixed the closing state machine to
156  *                                      resemble the RFC.
157  *              Alan Cox        :       More 'per spec' fixes.
158  *              Jorge Cwik      :       Even faster checksumming.
159  *              Alan Cox        :       tcp_data() doesn't ack illegal PSH
160  *                                      only frames. At least one pc tcp stack
161  *                                      generates them.
162  *              Alan Cox        :       Cache last socket.
163  *              Alan Cox        :       Per route irtt.
164  *              Matt Day        :       poll()->select() match BSD precisely on error
165  *              Alan Cox        :       New buffers
166  *              Marc Tamsky     :       Various sk->prot->retransmits and
167  *                                      sk->retransmits misupdating fixed.
168  *                                      Fixed tcp_write_timeout: stuck close,
169  *                                      and TCP syn retries gets used now.
170  *              Mark Yarvis     :       In tcp_read_wakeup(), don't send an
171  *                                      ack if state is TCP_CLOSED.
172  *              Alan Cox        :       Look up device on a retransmit - routes may
173  *                                      change. Doesn't yet cope with MSS shrink right
174  *                                      but it's a start!
175  *              Marc Tamsky     :       Closing in closing fixes.
176  *              Mike Shaver     :       RFC1122 verifications.
177  *              Alan Cox        :       rcv_saddr errors.
178  *              Alan Cox        :       Block double connect().
179  *              Alan Cox        :       Small hooks for enSKIP.
180  *              Alexey Kuznetsov:       Path MTU discovery.
181  *              Alan Cox        :       Support soft errors.
182  *              Alan Cox        :       Fix MTU discovery pathological case
183  *                                      when the remote claims no mtu!
184  *              Marc Tamsky     :       TCP_CLOSE fix.
185  *              Colin (G3TNE)   :       Send a reset on syn ack replies in
186  *                                      window but wrong (fixes NT lpd problems)
187  *              Pedro Roque     :       Better TCP window handling, delayed ack.
188  *              Joerg Reuter    :       No modification of locked buffers in
189  *                                      tcp_do_retransmit()
190  *              Eric Schenk     :       Changed receiver side silly window
191  *                                      avoidance algorithm to BSD style
192  *                                      algorithm. This doubles throughput
193  *                                      against machines running Solaris,
194  *                                      and seems to result in general
195  *                                      improvement.
196  *      Stefan Magdalinski      :       adjusted tcp_readable() to fix FIONREAD
197  *      Willy Konynenberg       :       Transparent proxying support.
198  *      Mike McLagan            :       Routing by source
199  *              Keith Owens     :       Do proper merging with partial SKB's in
200  *                                      tcp_do_sendmsg to avoid burstiness.
201  *              Eric Schenk     :       Fix fast close down bug with
202  *                                      shutdown() followed by close().
203  *              Andi Kleen      :       Make poll agree with SIGIO
204  *      Salvatore Sanfilippo    :       Support SO_LINGER with linger == 1 and
205  *                                      lingertime == 0 (RFC 793 ABORT Call)
206  *      Hirokazu Takahashi      :       Use copy_from_user() instead of
207  *                                      csum_and_copy_from_user() if possible.
208  *
209  * Description of States:
210  *
211  *      TCP_SYN_SENT            sent a connection request, waiting for ack
212  *
213  *      TCP_SYN_RECV            received a connection request, sent ack,
214  *                              waiting for final ack in three-way handshake.
215  *
216  *      TCP_ESTABLISHED         connection established
217  *
218  *      TCP_FIN_WAIT1           our side has shutdown, waiting to complete
219  *                              transmission of remaining buffered data
220  *
221  *      TCP_FIN_WAIT2           all buffered data sent, waiting for remote
222  *                              to shutdown
223  *
224  *      TCP_CLOSING             both sides have shutdown but we still have
225  *                              data we have to finish sending
226  *
227  *      TCP_TIME_WAIT           timeout to catch resent junk before entering
228  *                              closed, can only be entered from FIN_WAIT2
229  *                              or CLOSING.  Required because the other end
230  *                              may not have gotten our last ACK causing it
231  *                              to retransmit the data packet (which we ignore)
232  *
233  *      TCP_CLOSE_WAIT          remote side has shutdown and is waiting for
234  *                              us to finish writing our data and to shutdown
235  *                              (we have to close() to move on to LAST_ACK)
236  *
237  *      TCP_LAST_ACK            out side has shutdown after remote has
238  *                              shutdown.  There may still be data in our
239  *                              buffer that we have to finish sending
240  *
241  *      TCP_CLOSE               socket is finished
242  */
243
244 #define pr_fmt(fmt) "TCP: " fmt
245
246 #include <crypto/hash.h>
247 #include <linux/kernel.h>
248 #include <linux/module.h>
249 #include <linux/types.h>
250 #include <linux/fcntl.h>
251 #include <linux/poll.h>
252 #include <linux/inet_diag.h>
253 #include <linux/init.h>
254 #include <linux/fs.h>
255 #include <linux/skbuff.h>
256 #include <linux/scatterlist.h>
257 #include <linux/splice.h>
258 #include <linux/net.h>
259 #include <linux/socket.h>
260 #include <linux/random.h>
261 #include <linux/memblock.h>
262 #include <linux/highmem.h>
263 #include <linux/swap.h>
264 #include <linux/cache.h>
265 #include <linux/err.h>
266 #include <linux/time.h>
267 #include <linux/slab.h>
268 #include <linux/errqueue.h>
269 #include <linux/static_key.h>
270
271 #include <net/icmp.h>
272 #include <net/inet_common.h>
273 #include <net/tcp.h>
274 #include <net/mptcp.h>
275 #include <net/xfrm.h>
276 #include <net/ip.h>
277 #include <net/sock.h>
278
279 #include <linux/uaccess.h>
280 #include <asm/ioctls.h>
281 #include <net/busy_poll.h>
282
283 struct percpu_counter tcp_orphan_count;
284 EXPORT_SYMBOL_GPL(tcp_orphan_count);
285
286 long sysctl_tcp_mem[3] __read_mostly;
287 EXPORT_SYMBOL(sysctl_tcp_mem);
288
289 atomic_long_t tcp_memory_allocated;     /* Current allocated memory. */
290 EXPORT_SYMBOL(tcp_memory_allocated);
291
292 #if IS_ENABLED(CONFIG_SMC)
293 DEFINE_STATIC_KEY_FALSE(tcp_have_smc);
294 EXPORT_SYMBOL(tcp_have_smc);
295 #endif
296
297 /*
298  * Current number of TCP sockets.
299  */
300 struct percpu_counter tcp_sockets_allocated;
301 EXPORT_SYMBOL(tcp_sockets_allocated);
302
303 /*
304  * TCP splice context
305  */
306 struct tcp_splice_state {
307         struct pipe_inode_info *pipe;
308         size_t len;
309         unsigned int flags;
310 };
311
312 /*
313  * Pressure flag: try to collapse.
314  * Technical note: it is used by multiple contexts non atomically.
315  * All the __sk_mem_schedule() is of this nature: accounting
316  * is strict, actions are advisory and have some latency.
317  */
318 unsigned long tcp_memory_pressure __read_mostly;
319 EXPORT_SYMBOL_GPL(tcp_memory_pressure);
320
321 DEFINE_STATIC_KEY_FALSE(tcp_rx_skb_cache_key);
322 EXPORT_SYMBOL(tcp_rx_skb_cache_key);
323
324 DEFINE_STATIC_KEY_FALSE(tcp_tx_skb_cache_key);
325
326 void tcp_enter_memory_pressure(struct sock *sk)
327 {
328         unsigned long val;
329
330         if (READ_ONCE(tcp_memory_pressure))
331                 return;
332         val = jiffies;
333
334         if (!val)
335                 val--;
336         if (!cmpxchg(&tcp_memory_pressure, 0, val))
337                 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMEMORYPRESSURES);
338 }
339 EXPORT_SYMBOL_GPL(tcp_enter_memory_pressure);
340
341 void tcp_leave_memory_pressure(struct sock *sk)
342 {
343         unsigned long val;
344
345         if (!READ_ONCE(tcp_memory_pressure))
346                 return;
347         val = xchg(&tcp_memory_pressure, 0);
348         if (val)
349                 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPMEMORYPRESSURESCHRONO,
350                               jiffies_to_msecs(jiffies - val));
351 }
352 EXPORT_SYMBOL_GPL(tcp_leave_memory_pressure);
353
354 /* Convert seconds to retransmits based on initial and max timeout */
355 static u8 secs_to_retrans(int seconds, int timeout, int rto_max)
356 {
357         u8 res = 0;
358
359         if (seconds > 0) {
360                 int period = timeout;
361
362                 res = 1;
363                 while (seconds > period && res < 255) {
364                         res++;
365                         timeout <<= 1;
366                         if (timeout > rto_max)
367                                 timeout = rto_max;
368                         period += timeout;
369                 }
370         }
371         return res;
372 }
373
374 /* Convert retransmits to seconds based on initial and max timeout */
375 static int retrans_to_secs(u8 retrans, int timeout, int rto_max)
376 {
377         int period = 0;
378
379         if (retrans > 0) {
380                 period = timeout;
381                 while (--retrans) {
382                         timeout <<= 1;
383                         if (timeout > rto_max)
384                                 timeout = rto_max;
385                         period += timeout;
386                 }
387         }
388         return period;
389 }
390
391 static u64 tcp_compute_delivery_rate(const struct tcp_sock *tp)
392 {
393         u32 rate = READ_ONCE(tp->rate_delivered);
394         u32 intv = READ_ONCE(tp->rate_interval_us);
395         u64 rate64 = 0;
396
397         if (rate && intv) {
398                 rate64 = (u64)rate * tp->mss_cache * USEC_PER_SEC;
399                 do_div(rate64, intv);
400         }
401         return rate64;
402 }
403
404 /* Address-family independent initialization for a tcp_sock.
405  *
406  * NOTE: A lot of things set to zero explicitly by call to
407  *       sk_alloc() so need not be done here.
408  */
409 void tcp_init_sock(struct sock *sk)
410 {
411         struct inet_connection_sock *icsk = inet_csk(sk);
412         struct tcp_sock *tp = tcp_sk(sk);
413
414         tp->out_of_order_queue = RB_ROOT;
415         sk->tcp_rtx_queue = RB_ROOT;
416         tcp_init_xmit_timers(sk);
417         INIT_LIST_HEAD(&tp->tsq_node);
418         INIT_LIST_HEAD(&tp->tsorted_sent_queue);
419
420         icsk->icsk_rto = TCP_TIMEOUT_INIT;
421         tp->mdev_us = jiffies_to_usecs(TCP_TIMEOUT_INIT);
422         minmax_reset(&tp->rtt_min, tcp_jiffies32, ~0U);
423
424         /* So many TCP implementations out there (incorrectly) count the
425          * initial SYN frame in their delayed-ACK and congestion control
426          * algorithms that we must have the following bandaid to talk
427          * efficiently to them.  -DaveM
428          */
429         tp->snd_cwnd = TCP_INIT_CWND;
430
431         /* There's a bubble in the pipe until at least the first ACK. */
432         tp->app_limited = ~0U;
433
434         /* See draft-stevens-tcpca-spec-01 for discussion of the
435          * initialization of these values.
436          */
437         tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
438         tp->snd_cwnd_clamp = ~0;
439         tp->mss_cache = TCP_MSS_DEFAULT;
440
441         tp->reordering = sock_net(sk)->ipv4.sysctl_tcp_reordering;
442         tcp_assign_congestion_control(sk);
443
444         tp->tsoffset = 0;
445         tp->rack.reo_wnd_steps = 1;
446
447         sk->sk_write_space = sk_stream_write_space;
448         sock_set_flag(sk, SOCK_USE_WRITE_QUEUE);
449
450         icsk->icsk_sync_mss = tcp_sync_mss;
451
452         WRITE_ONCE(sk->sk_sndbuf, sock_net(sk)->ipv4.sysctl_tcp_wmem[1]);
453         WRITE_ONCE(sk->sk_rcvbuf, sock_net(sk)->ipv4.sysctl_tcp_rmem[1]);
454
455         sk_sockets_allocated_inc(sk);
456         sk->sk_route_forced_caps = NETIF_F_GSO;
457 }
458 EXPORT_SYMBOL(tcp_init_sock);
459
460 static void tcp_tx_timestamp(struct sock *sk, u16 tsflags)
461 {
462         struct sk_buff *skb = tcp_write_queue_tail(sk);
463
464         if (tsflags && skb) {
465                 struct skb_shared_info *shinfo = skb_shinfo(skb);
466                 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
467
468                 sock_tx_timestamp(sk, tsflags, &shinfo->tx_flags);
469                 if (tsflags & SOF_TIMESTAMPING_TX_ACK)
470                         tcb->txstamp_ack = 1;
471                 if (tsflags & SOF_TIMESTAMPING_TX_RECORD_MASK)
472                         shinfo->tskey = TCP_SKB_CB(skb)->seq + skb->len - 1;
473         }
474 }
475
476 static inline bool tcp_stream_is_readable(const struct tcp_sock *tp,
477                                           int target, struct sock *sk)
478 {
479         return (READ_ONCE(tp->rcv_nxt) - READ_ONCE(tp->copied_seq) >= target) ||
480                 (sk->sk_prot->stream_memory_read ?
481                 sk->sk_prot->stream_memory_read(sk) : false);
482 }
483
484 /*
485  *      Wait for a TCP event.
486  *
487  *      Note that we don't need to lock the socket, as the upper poll layers
488  *      take care of normal races (between the test and the event) and we don't
489  *      go look at any of the socket buffers directly.
490  */
491 __poll_t tcp_poll(struct file *file, struct socket *sock, poll_table *wait)
492 {
493         __poll_t mask;
494         struct sock *sk = sock->sk;
495         const struct tcp_sock *tp = tcp_sk(sk);
496         int state;
497
498         sock_poll_wait(file, sock, wait);
499
500         state = inet_sk_state_load(sk);
501         if (state == TCP_LISTEN)
502                 return inet_csk_listen_poll(sk);
503
504         /* Socket is not locked. We are protected from async events
505          * by poll logic and correct handling of state changes
506          * made by other threads is impossible in any case.
507          */
508
509         mask = 0;
510
511         /*
512          * EPOLLHUP is certainly not done right. But poll() doesn't
513          * have a notion of HUP in just one direction, and for a
514          * socket the read side is more interesting.
515          *
516          * Some poll() documentation says that EPOLLHUP is incompatible
517          * with the EPOLLOUT/POLLWR flags, so somebody should check this
518          * all. But careful, it tends to be safer to return too many
519          * bits than too few, and you can easily break real applications
520          * if you don't tell them that something has hung up!
521          *
522          * Check-me.
523          *
524          * Check number 1. EPOLLHUP is _UNMASKABLE_ event (see UNIX98 and
525          * our fs/select.c). It means that after we received EOF,
526          * poll always returns immediately, making impossible poll() on write()
527          * in state CLOSE_WAIT. One solution is evident --- to set EPOLLHUP
528          * if and only if shutdown has been made in both directions.
529          * Actually, it is interesting to look how Solaris and DUX
530          * solve this dilemma. I would prefer, if EPOLLHUP were maskable,
531          * then we could set it on SND_SHUTDOWN. BTW examples given
532          * in Stevens' books assume exactly this behaviour, it explains
533          * why EPOLLHUP is incompatible with EPOLLOUT.  --ANK
534          *
535          * NOTE. Check for TCP_CLOSE is added. The goal is to prevent
536          * blocking on fresh not-connected or disconnected socket. --ANK
537          */
538         if (sk->sk_shutdown == SHUTDOWN_MASK || state == TCP_CLOSE)
539                 mask |= EPOLLHUP;
540         if (sk->sk_shutdown & RCV_SHUTDOWN)
541                 mask |= EPOLLIN | EPOLLRDNORM | EPOLLRDHUP;
542
543         /* Connected or passive Fast Open socket? */
544         if (state != TCP_SYN_SENT &&
545             (state != TCP_SYN_RECV || rcu_access_pointer(tp->fastopen_rsk))) {
546                 int target = sock_rcvlowat(sk, 0, INT_MAX);
547
548                 if (READ_ONCE(tp->urg_seq) == READ_ONCE(tp->copied_seq) &&
549                     !sock_flag(sk, SOCK_URGINLINE) &&
550                     tp->urg_data)
551                         target++;
552
553                 if (tcp_stream_is_readable(tp, target, sk))
554                         mask |= EPOLLIN | EPOLLRDNORM;
555
556                 if (!(sk->sk_shutdown & SEND_SHUTDOWN)) {
557                         if (sk_stream_is_writeable(sk)) {
558                                 mask |= EPOLLOUT | EPOLLWRNORM;
559                         } else {  /* send SIGIO later */
560                                 sk_set_bit(SOCKWQ_ASYNC_NOSPACE, sk);
561                                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
562
563                                 /* Race breaker. If space is freed after
564                                  * wspace test but before the flags are set,
565                                  * IO signal will be lost. Memory barrier
566                                  * pairs with the input side.
567                                  */
568                                 smp_mb__after_atomic();
569                                 if (sk_stream_is_writeable(sk))
570                                         mask |= EPOLLOUT | EPOLLWRNORM;
571                         }
572                 } else
573                         mask |= EPOLLOUT | EPOLLWRNORM;
574
575                 if (tp->urg_data & TCP_URG_VALID)
576                         mask |= EPOLLPRI;
577         } else if (state == TCP_SYN_SENT && inet_sk(sk)->defer_connect) {
578                 /* Active TCP fastopen socket with defer_connect
579                  * Return EPOLLOUT so application can call write()
580                  * in order for kernel to generate SYN+data
581                  */
582                 mask |= EPOLLOUT | EPOLLWRNORM;
583         }
584         /* This barrier is coupled with smp_wmb() in tcp_reset() */
585         smp_rmb();
586         if (sk->sk_err || !skb_queue_empty_lockless(&sk->sk_error_queue))
587                 mask |= EPOLLERR;
588
589         return mask;
590 }
591 EXPORT_SYMBOL(tcp_poll);
592
593 int tcp_ioctl(struct sock *sk, int cmd, unsigned long arg)
594 {
595         struct tcp_sock *tp = tcp_sk(sk);
596         int answ;
597         bool slow;
598
599         switch (cmd) {
600         case SIOCINQ:
601                 if (sk->sk_state == TCP_LISTEN)
602                         return -EINVAL;
603
604                 slow = lock_sock_fast(sk);
605                 answ = tcp_inq(sk);
606                 unlock_sock_fast(sk, slow);
607                 break;
608         case SIOCATMARK:
609                 answ = tp->urg_data &&
610                        READ_ONCE(tp->urg_seq) == READ_ONCE(tp->copied_seq);
611                 break;
612         case SIOCOUTQ:
613                 if (sk->sk_state == TCP_LISTEN)
614                         return -EINVAL;
615
616                 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
617                         answ = 0;
618                 else
619                         answ = READ_ONCE(tp->write_seq) - tp->snd_una;
620                 break;
621         case SIOCOUTQNSD:
622                 if (sk->sk_state == TCP_LISTEN)
623                         return -EINVAL;
624
625                 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
626                         answ = 0;
627                 else
628                         answ = READ_ONCE(tp->write_seq) -
629                                READ_ONCE(tp->snd_nxt);
630                 break;
631         default:
632                 return -ENOIOCTLCMD;
633         }
634
635         return put_user(answ, (int __user *)arg);
636 }
637 EXPORT_SYMBOL(tcp_ioctl);
638
639 static inline void tcp_mark_push(struct tcp_sock *tp, struct sk_buff *skb)
640 {
641         TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
642         tp->pushed_seq = tp->write_seq;
643 }
644
645 static inline bool forced_push(const struct tcp_sock *tp)
646 {
647         return after(tp->write_seq, tp->pushed_seq + (tp->max_window >> 1));
648 }
649
650 static void skb_entail(struct sock *sk, struct sk_buff *skb)
651 {
652         struct tcp_sock *tp = tcp_sk(sk);
653         struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
654
655         skb->csum    = 0;
656         tcb->seq     = tcb->end_seq = tp->write_seq;
657         tcb->tcp_flags = TCPHDR_ACK;
658         tcb->sacked  = 0;
659         __skb_header_release(skb);
660         tcp_add_write_queue_tail(sk, skb);
661         sk_wmem_queued_add(sk, skb->truesize);
662         sk_mem_charge(sk, skb->truesize);
663         if (tp->nonagle & TCP_NAGLE_PUSH)
664                 tp->nonagle &= ~TCP_NAGLE_PUSH;
665
666         tcp_slow_start_after_idle_check(sk);
667 }
668
669 static inline void tcp_mark_urg(struct tcp_sock *tp, int flags)
670 {
671         if (flags & MSG_OOB)
672                 tp->snd_up = tp->write_seq;
673 }
674
675 /* If a not yet filled skb is pushed, do not send it if
676  * we have data packets in Qdisc or NIC queues :
677  * Because TX completion will happen shortly, it gives a chance
678  * to coalesce future sendmsg() payload into this skb, without
679  * need for a timer, and with no latency trade off.
680  * As packets containing data payload have a bigger truesize
681  * than pure acks (dataless) packets, the last checks prevent
682  * autocorking if we only have an ACK in Qdisc/NIC queues,
683  * or if TX completion was delayed after we processed ACK packet.
684  */
685 static bool tcp_should_autocork(struct sock *sk, struct sk_buff *skb,
686                                 int size_goal)
687 {
688         return skb->len < size_goal &&
689                sock_net(sk)->ipv4.sysctl_tcp_autocorking &&
690                !tcp_rtx_queue_empty(sk) &&
691                refcount_read(&sk->sk_wmem_alloc) > skb->truesize;
692 }
693
694 void tcp_push(struct sock *sk, int flags, int mss_now,
695               int nonagle, int size_goal)
696 {
697         struct tcp_sock *tp = tcp_sk(sk);
698         struct sk_buff *skb;
699
700         skb = tcp_write_queue_tail(sk);
701         if (!skb)
702                 return;
703         if (!(flags & MSG_MORE) || forced_push(tp))
704                 tcp_mark_push(tp, skb);
705
706         tcp_mark_urg(tp, flags);
707
708         if (tcp_should_autocork(sk, skb, size_goal)) {
709
710                 /* avoid atomic op if TSQ_THROTTLED bit is already set */
711                 if (!test_bit(TSQ_THROTTLED, &sk->sk_tsq_flags)) {
712                         NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPAUTOCORKING);
713                         set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags);
714                 }
715                 /* It is possible TX completion already happened
716                  * before we set TSQ_THROTTLED.
717                  */
718                 if (refcount_read(&sk->sk_wmem_alloc) > skb->truesize)
719                         return;
720         }
721
722         if (flags & MSG_MORE)
723                 nonagle = TCP_NAGLE_CORK;
724
725         __tcp_push_pending_frames(sk, mss_now, nonagle);
726 }
727
728 static int tcp_splice_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb,
729                                 unsigned int offset, size_t len)
730 {
731         struct tcp_splice_state *tss = rd_desc->arg.data;
732         int ret;
733
734         ret = skb_splice_bits(skb, skb->sk, offset, tss->pipe,
735                               min(rd_desc->count, len), tss->flags);
736         if (ret > 0)
737                 rd_desc->count -= ret;
738         return ret;
739 }
740
741 static int __tcp_splice_read(struct sock *sk, struct tcp_splice_state *tss)
742 {
743         /* Store TCP splice context information in read_descriptor_t. */
744         read_descriptor_t rd_desc = {
745                 .arg.data = tss,
746                 .count    = tss->len,
747         };
748
749         return tcp_read_sock(sk, &rd_desc, tcp_splice_data_recv);
750 }
751
752 /**
753  *  tcp_splice_read - splice data from TCP socket to a pipe
754  * @sock:       socket to splice from
755  * @ppos:       position (not valid)
756  * @pipe:       pipe to splice to
757  * @len:        number of bytes to splice
758  * @flags:      splice modifier flags
759  *
760  * Description:
761  *    Will read pages from given socket and fill them into a pipe.
762  *
763  **/
764 ssize_t tcp_splice_read(struct socket *sock, loff_t *ppos,
765                         struct pipe_inode_info *pipe, size_t len,
766                         unsigned int flags)
767 {
768         struct sock *sk = sock->sk;
769         struct tcp_splice_state tss = {
770                 .pipe = pipe,
771                 .len = len,
772                 .flags = flags,
773         };
774         long timeo;
775         ssize_t spliced;
776         int ret;
777
778         sock_rps_record_flow(sk);
779         /*
780          * We can't seek on a socket input
781          */
782         if (unlikely(*ppos))
783                 return -ESPIPE;
784
785         ret = spliced = 0;
786
787         lock_sock(sk);
788
789         timeo = sock_rcvtimeo(sk, sock->file->f_flags & O_NONBLOCK);
790         while (tss.len) {
791                 ret = __tcp_splice_read(sk, &tss);
792                 if (ret < 0)
793                         break;
794                 else if (!ret) {
795                         if (spliced)
796                                 break;
797                         if (sock_flag(sk, SOCK_DONE))
798                                 break;
799                         if (sk->sk_err) {
800                                 ret = sock_error(sk);
801                                 break;
802                         }
803                         if (sk->sk_shutdown & RCV_SHUTDOWN)
804                                 break;
805                         if (sk->sk_state == TCP_CLOSE) {
806                                 /*
807                                  * This occurs when user tries to read
808                                  * from never connected socket.
809                                  */
810                                 ret = -ENOTCONN;
811                                 break;
812                         }
813                         if (!timeo) {
814                                 ret = -EAGAIN;
815                                 break;
816                         }
817                         /* if __tcp_splice_read() got nothing while we have
818                          * an skb in receive queue, we do not want to loop.
819                          * This might happen with URG data.
820                          */
821                         if (!skb_queue_empty(&sk->sk_receive_queue))
822                                 break;
823                         sk_wait_data(sk, &timeo, NULL);
824                         if (signal_pending(current)) {
825                                 ret = sock_intr_errno(timeo);
826                                 break;
827                         }
828                         continue;
829                 }
830                 tss.len -= ret;
831                 spliced += ret;
832
833                 if (!timeo)
834                         break;
835                 release_sock(sk);
836                 lock_sock(sk);
837
838                 if (sk->sk_err || sk->sk_state == TCP_CLOSE ||
839                     (sk->sk_shutdown & RCV_SHUTDOWN) ||
840                     signal_pending(current))
841                         break;
842         }
843
844         release_sock(sk);
845
846         if (spliced)
847                 return spliced;
848
849         return ret;
850 }
851 EXPORT_SYMBOL(tcp_splice_read);
852
853 struct sk_buff *sk_stream_alloc_skb(struct sock *sk, int size, gfp_t gfp,
854                                     bool force_schedule)
855 {
856         struct sk_buff *skb;
857
858         if (likely(!size)) {
859                 skb = sk->sk_tx_skb_cache;
860                 if (skb) {
861                         skb->truesize = SKB_TRUESIZE(skb_end_offset(skb));
862                         sk->sk_tx_skb_cache = NULL;
863                         pskb_trim(skb, 0);
864                         INIT_LIST_HEAD(&skb->tcp_tsorted_anchor);
865                         skb_shinfo(skb)->tx_flags = 0;
866                         memset(TCP_SKB_CB(skb), 0, sizeof(struct tcp_skb_cb));
867                         return skb;
868                 }
869         }
870         /* The TCP header must be at least 32-bit aligned.  */
871         size = ALIGN(size, 4);
872
873         if (unlikely(tcp_under_memory_pressure(sk)))
874                 sk_mem_reclaim_partial(sk);
875
876         skb = alloc_skb_fclone(size + sk->sk_prot->max_header, gfp);
877         if (likely(skb)) {
878                 bool mem_scheduled;
879
880                 if (force_schedule) {
881                         mem_scheduled = true;
882                         sk_forced_mem_schedule(sk, skb->truesize);
883                 } else {
884                         mem_scheduled = sk_wmem_schedule(sk, skb->truesize);
885                 }
886                 if (likely(mem_scheduled)) {
887                         skb_reserve(skb, sk->sk_prot->max_header);
888                         /*
889                          * Make sure that we have exactly size bytes
890                          * available to the caller, no more, no less.
891                          */
892                         skb->reserved_tailroom = skb->end - skb->tail - size;
893                         INIT_LIST_HEAD(&skb->tcp_tsorted_anchor);
894                         return skb;
895                 }
896                 __kfree_skb(skb);
897         } else {
898                 sk->sk_prot->enter_memory_pressure(sk);
899                 sk_stream_moderate_sndbuf(sk);
900         }
901         return NULL;
902 }
903
904 static unsigned int tcp_xmit_size_goal(struct sock *sk, u32 mss_now,
905                                        int large_allowed)
906 {
907         struct tcp_sock *tp = tcp_sk(sk);
908         u32 new_size_goal, size_goal;
909
910         if (!large_allowed)
911                 return mss_now;
912
913         /* Note : tcp_tso_autosize() will eventually split this later */
914         new_size_goal = sk->sk_gso_max_size - 1 - MAX_TCP_HEADER;
915         new_size_goal = tcp_bound_to_half_wnd(tp, new_size_goal);
916
917         /* We try hard to avoid divides here */
918         size_goal = tp->gso_segs * mss_now;
919         if (unlikely(new_size_goal < size_goal ||
920                      new_size_goal >= size_goal + mss_now)) {
921                 tp->gso_segs = min_t(u16, new_size_goal / mss_now,
922                                      sk->sk_gso_max_segs);
923                 size_goal = tp->gso_segs * mss_now;
924         }
925
926         return max(size_goal, mss_now);
927 }
928
929 int tcp_send_mss(struct sock *sk, int *size_goal, int flags)
930 {
931         int mss_now;
932
933         mss_now = tcp_current_mss(sk);
934         *size_goal = tcp_xmit_size_goal(sk, mss_now, !(flags & MSG_OOB));
935
936         return mss_now;
937 }
938
939 /* In some cases, both sendpage() and sendmsg() could have added
940  * an skb to the write queue, but failed adding payload on it.
941  * We need to remove it to consume less memory, but more
942  * importantly be able to generate EPOLLOUT for Edge Trigger epoll()
943  * users.
944  */
945 static void tcp_remove_empty_skb(struct sock *sk, struct sk_buff *skb)
946 {
947         if (skb && !skb->len) {
948                 tcp_unlink_write_queue(skb, sk);
949                 if (tcp_write_queue_empty(sk))
950                         tcp_chrono_stop(sk, TCP_CHRONO_BUSY);
951                 sk_wmem_free_skb(sk, skb);
952         }
953 }
954
955 ssize_t do_tcp_sendpages(struct sock *sk, struct page *page, int offset,
956                          size_t size, int flags)
957 {
958         struct tcp_sock *tp = tcp_sk(sk);
959         int mss_now, size_goal;
960         int err;
961         ssize_t copied;
962         long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
963
964         if (IS_ENABLED(CONFIG_DEBUG_VM) &&
965             WARN_ONCE(PageSlab(page), "page must not be a Slab one"))
966                 return -EINVAL;
967
968         /* Wait for a connection to finish. One exception is TCP Fast Open
969          * (passive side) where data is allowed to be sent before a connection
970          * is fully established.
971          */
972         if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) &&
973             !tcp_passive_fastopen(sk)) {
974                 err = sk_stream_wait_connect(sk, &timeo);
975                 if (err != 0)
976                         goto out_err;
977         }
978
979         sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
980
981         mss_now = tcp_send_mss(sk, &size_goal, flags);
982         copied = 0;
983
984         err = -EPIPE;
985         if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
986                 goto out_err;
987
988         while (size > 0) {
989                 struct sk_buff *skb = tcp_write_queue_tail(sk);
990                 int copy, i;
991                 bool can_coalesce;
992
993                 if (!skb || (copy = size_goal - skb->len) <= 0 ||
994                     !tcp_skb_can_collapse_to(skb)) {
995 new_segment:
996                         if (!sk_stream_memory_free(sk))
997                                 goto wait_for_sndbuf;
998
999                         skb = sk_stream_alloc_skb(sk, 0, sk->sk_allocation,
1000                                         tcp_rtx_and_write_queues_empty(sk));
1001                         if (!skb)
1002                                 goto wait_for_memory;
1003
1004 #ifdef CONFIG_TLS_DEVICE
1005                         skb->decrypted = !!(flags & MSG_SENDPAGE_DECRYPTED);
1006 #endif
1007                         skb_entail(sk, skb);
1008                         copy = size_goal;
1009                 }
1010
1011                 if (copy > size)
1012                         copy = size;
1013
1014                 i = skb_shinfo(skb)->nr_frags;
1015                 can_coalesce = skb_can_coalesce(skb, i, page, offset);
1016                 if (!can_coalesce && i >= sysctl_max_skb_frags) {
1017                         tcp_mark_push(tp, skb);
1018                         goto new_segment;
1019                 }
1020                 if (!sk_wmem_schedule(sk, copy))
1021                         goto wait_for_memory;
1022
1023                 if (can_coalesce) {
1024                         skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1025                 } else {
1026                         get_page(page);
1027                         skb_fill_page_desc(skb, i, page, offset, copy);
1028                 }
1029
1030                 if (!(flags & MSG_NO_SHARED_FRAGS))
1031                         skb_shinfo(skb)->tx_flags |= SKBTX_SHARED_FRAG;
1032
1033                 skb->len += copy;
1034                 skb->data_len += copy;
1035                 skb->truesize += copy;
1036                 sk_wmem_queued_add(sk, copy);
1037                 sk_mem_charge(sk, copy);
1038                 skb->ip_summed = CHECKSUM_PARTIAL;
1039                 WRITE_ONCE(tp->write_seq, tp->write_seq + copy);
1040                 TCP_SKB_CB(skb)->end_seq += copy;
1041                 tcp_skb_pcount_set(skb, 0);
1042
1043                 if (!copied)
1044                         TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
1045
1046                 copied += copy;
1047                 offset += copy;
1048                 size -= copy;
1049                 if (!size)
1050                         goto out;
1051
1052                 if (skb->len < size_goal || (flags & MSG_OOB))
1053                         continue;
1054
1055                 if (forced_push(tp)) {
1056                         tcp_mark_push(tp, skb);
1057                         __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
1058                 } else if (skb == tcp_send_head(sk))
1059                         tcp_push_one(sk, mss_now);
1060                 continue;
1061
1062 wait_for_sndbuf:
1063                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1064 wait_for_memory:
1065                 tcp_push(sk, flags & ~MSG_MORE, mss_now,
1066                          TCP_NAGLE_PUSH, size_goal);
1067
1068                 err = sk_stream_wait_memory(sk, &timeo);
1069                 if (err != 0)
1070                         goto do_error;
1071
1072                 mss_now = tcp_send_mss(sk, &size_goal, flags);
1073         }
1074
1075 out:
1076         if (copied) {
1077                 tcp_tx_timestamp(sk, sk->sk_tsflags);
1078                 if (!(flags & MSG_SENDPAGE_NOTLAST))
1079                         tcp_push(sk, flags, mss_now, tp->nonagle, size_goal);
1080         }
1081         return copied;
1082
1083 do_error:
1084         tcp_remove_empty_skb(sk, tcp_write_queue_tail(sk));
1085         if (copied)
1086                 goto out;
1087 out_err:
1088         /* make sure we wake any epoll edge trigger waiter */
1089         if (unlikely(tcp_rtx_and_write_queues_empty(sk) && err == -EAGAIN)) {
1090                 sk->sk_write_space(sk);
1091                 tcp_chrono_stop(sk, TCP_CHRONO_SNDBUF_LIMITED);
1092         }
1093         return sk_stream_error(sk, flags, err);
1094 }
1095 EXPORT_SYMBOL_GPL(do_tcp_sendpages);
1096
1097 int tcp_sendpage_locked(struct sock *sk, struct page *page, int offset,
1098                         size_t size, int flags)
1099 {
1100         if (!(sk->sk_route_caps & NETIF_F_SG))
1101                 return sock_no_sendpage_locked(sk, page, offset, size, flags);
1102
1103         tcp_rate_check_app_limited(sk);  /* is sending application-limited? */
1104
1105         return do_tcp_sendpages(sk, page, offset, size, flags);
1106 }
1107 EXPORT_SYMBOL_GPL(tcp_sendpage_locked);
1108
1109 int tcp_sendpage(struct sock *sk, struct page *page, int offset,
1110                  size_t size, int flags)
1111 {
1112         int ret;
1113
1114         lock_sock(sk);
1115         ret = tcp_sendpage_locked(sk, page, offset, size, flags);
1116         release_sock(sk);
1117
1118         return ret;
1119 }
1120 EXPORT_SYMBOL(tcp_sendpage);
1121
1122 void tcp_free_fastopen_req(struct tcp_sock *tp)
1123 {
1124         if (tp->fastopen_req) {
1125                 kfree(tp->fastopen_req);
1126                 tp->fastopen_req = NULL;
1127         }
1128 }
1129
1130 static int tcp_sendmsg_fastopen(struct sock *sk, struct msghdr *msg,
1131                                 int *copied, size_t size,
1132                                 struct ubuf_info *uarg)
1133 {
1134         struct tcp_sock *tp = tcp_sk(sk);
1135         struct inet_sock *inet = inet_sk(sk);
1136         struct sockaddr *uaddr = msg->msg_name;
1137         int err, flags;
1138
1139         if (!(sock_net(sk)->ipv4.sysctl_tcp_fastopen & TFO_CLIENT_ENABLE) ||
1140             (uaddr && msg->msg_namelen >= sizeof(uaddr->sa_family) &&
1141              uaddr->sa_family == AF_UNSPEC))
1142                 return -EOPNOTSUPP;
1143         if (tp->fastopen_req)
1144                 return -EALREADY; /* Another Fast Open is in progress */
1145
1146         tp->fastopen_req = kzalloc(sizeof(struct tcp_fastopen_request),
1147                                    sk->sk_allocation);
1148         if (unlikely(!tp->fastopen_req))
1149                 return -ENOBUFS;
1150         tp->fastopen_req->data = msg;
1151         tp->fastopen_req->size = size;
1152         tp->fastopen_req->uarg = uarg;
1153
1154         if (inet->defer_connect) {
1155                 err = tcp_connect(sk);
1156                 /* Same failure procedure as in tcp_v4/6_connect */
1157                 if (err) {
1158                         tcp_set_state(sk, TCP_CLOSE);
1159                         inet->inet_dport = 0;
1160                         sk->sk_route_caps = 0;
1161                 }
1162         }
1163         flags = (msg->msg_flags & MSG_DONTWAIT) ? O_NONBLOCK : 0;
1164         err = __inet_stream_connect(sk->sk_socket, uaddr,
1165                                     msg->msg_namelen, flags, 1);
1166         /* fastopen_req could already be freed in __inet_stream_connect
1167          * if the connection times out or gets rst
1168          */
1169         if (tp->fastopen_req) {
1170                 *copied = tp->fastopen_req->copied;
1171                 tcp_free_fastopen_req(tp);
1172                 inet->defer_connect = 0;
1173         }
1174         return err;
1175 }
1176
1177 int tcp_sendmsg_locked(struct sock *sk, struct msghdr *msg, size_t size)
1178 {
1179         struct tcp_sock *tp = tcp_sk(sk);
1180         struct ubuf_info *uarg = NULL;
1181         struct sk_buff *skb;
1182         struct sockcm_cookie sockc;
1183         int flags, err, copied = 0;
1184         int mss_now = 0, size_goal, copied_syn = 0;
1185         int process_backlog = 0;
1186         bool zc = false;
1187         long timeo;
1188
1189         flags = msg->msg_flags;
1190
1191         if (flags & MSG_ZEROCOPY && size && sock_flag(sk, SOCK_ZEROCOPY)) {
1192                 skb = tcp_write_queue_tail(sk);
1193                 uarg = sock_zerocopy_realloc(sk, size, skb_zcopy(skb));
1194                 if (!uarg) {
1195                         err = -ENOBUFS;
1196                         goto out_err;
1197                 }
1198
1199                 zc = sk->sk_route_caps & NETIF_F_SG;
1200                 if (!zc)
1201                         uarg->zerocopy = 0;
1202         }
1203
1204         if (unlikely(flags & MSG_FASTOPEN || inet_sk(sk)->defer_connect) &&
1205             !tp->repair) {
1206                 err = tcp_sendmsg_fastopen(sk, msg, &copied_syn, size, uarg);
1207                 if (err == -EINPROGRESS && copied_syn > 0)
1208                         goto out;
1209                 else if (err)
1210                         goto out_err;
1211         }
1212
1213         timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
1214
1215         tcp_rate_check_app_limited(sk);  /* is sending application-limited? */
1216
1217         /* Wait for a connection to finish. One exception is TCP Fast Open
1218          * (passive side) where data is allowed to be sent before a connection
1219          * is fully established.
1220          */
1221         if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) &&
1222             !tcp_passive_fastopen(sk)) {
1223                 err = sk_stream_wait_connect(sk, &timeo);
1224                 if (err != 0)
1225                         goto do_error;
1226         }
1227
1228         if (unlikely(tp->repair)) {
1229                 if (tp->repair_queue == TCP_RECV_QUEUE) {
1230                         copied = tcp_send_rcvq(sk, msg, size);
1231                         goto out_nopush;
1232                 }
1233
1234                 err = -EINVAL;
1235                 if (tp->repair_queue == TCP_NO_QUEUE)
1236                         goto out_err;
1237
1238                 /* 'common' sending to sendq */
1239         }
1240
1241         sockcm_init(&sockc, sk);
1242         if (msg->msg_controllen) {
1243                 err = sock_cmsg_send(sk, msg, &sockc);
1244                 if (unlikely(err)) {
1245                         err = -EINVAL;
1246                         goto out_err;
1247                 }
1248         }
1249
1250         /* This should be in poll */
1251         sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
1252
1253         /* Ok commence sending. */
1254         copied = 0;
1255
1256 restart:
1257         mss_now = tcp_send_mss(sk, &size_goal, flags);
1258
1259         err = -EPIPE;
1260         if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
1261                 goto do_error;
1262
1263         while (msg_data_left(msg)) {
1264                 int copy = 0;
1265
1266                 skb = tcp_write_queue_tail(sk);
1267                 if (skb)
1268                         copy = size_goal - skb->len;
1269
1270                 if (copy <= 0 || !tcp_skb_can_collapse_to(skb)) {
1271                         bool first_skb;
1272
1273 new_segment:
1274                         if (!sk_stream_memory_free(sk))
1275                                 goto wait_for_sndbuf;
1276
1277                         if (unlikely(process_backlog >= 16)) {
1278                                 process_backlog = 0;
1279                                 if (sk_flush_backlog(sk))
1280                                         goto restart;
1281                         }
1282                         first_skb = tcp_rtx_and_write_queues_empty(sk);
1283                         skb = sk_stream_alloc_skb(sk, 0, sk->sk_allocation,
1284                                                   first_skb);
1285                         if (!skb)
1286                                 goto wait_for_memory;
1287
1288                         process_backlog++;
1289                         skb->ip_summed = CHECKSUM_PARTIAL;
1290
1291                         skb_entail(sk, skb);
1292                         copy = size_goal;
1293
1294                         /* All packets are restored as if they have
1295                          * already been sent. skb_mstamp_ns isn't set to
1296                          * avoid wrong rtt estimation.
1297                          */
1298                         if (tp->repair)
1299                                 TCP_SKB_CB(skb)->sacked |= TCPCB_REPAIRED;
1300                 }
1301
1302                 /* Try to append data to the end of skb. */
1303                 if (copy > msg_data_left(msg))
1304                         copy = msg_data_left(msg);
1305
1306                 /* Where to copy to? */
1307                 if (skb_availroom(skb) > 0 && !zc) {
1308                         /* We have some space in skb head. Superb! */
1309                         copy = min_t(int, copy, skb_availroom(skb));
1310                         err = skb_add_data_nocache(sk, skb, &msg->msg_iter, copy);
1311                         if (err)
1312                                 goto do_fault;
1313                 } else if (!zc) {
1314                         bool merge = true;
1315                         int i = skb_shinfo(skb)->nr_frags;
1316                         struct page_frag *pfrag = sk_page_frag(sk);
1317
1318                         if (!sk_page_frag_refill(sk, pfrag))
1319                                 goto wait_for_memory;
1320
1321                         if (!skb_can_coalesce(skb, i, pfrag->page,
1322                                               pfrag->offset)) {
1323                                 if (i >= sysctl_max_skb_frags) {
1324                                         tcp_mark_push(tp, skb);
1325                                         goto new_segment;
1326                                 }
1327                                 merge = false;
1328                         }
1329
1330                         copy = min_t(int, copy, pfrag->size - pfrag->offset);
1331
1332                         if (!sk_wmem_schedule(sk, copy))
1333                                 goto wait_for_memory;
1334
1335                         err = skb_copy_to_page_nocache(sk, &msg->msg_iter, skb,
1336                                                        pfrag->page,
1337                                                        pfrag->offset,
1338                                                        copy);
1339                         if (err)
1340                                 goto do_error;
1341
1342                         /* Update the skb. */
1343                         if (merge) {
1344                                 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1345                         } else {
1346                                 skb_fill_page_desc(skb, i, pfrag->page,
1347                                                    pfrag->offset, copy);
1348                                 page_ref_inc(pfrag->page);
1349                         }
1350                         pfrag->offset += copy;
1351                 } else {
1352                         err = skb_zerocopy_iter_stream(sk, skb, msg, copy, uarg);
1353                         if (err == -EMSGSIZE || err == -EEXIST) {
1354                                 tcp_mark_push(tp, skb);
1355                                 goto new_segment;
1356                         }
1357                         if (err < 0)
1358                                 goto do_error;
1359                         copy = err;
1360                 }
1361
1362                 if (!copied)
1363                         TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
1364
1365                 WRITE_ONCE(tp->write_seq, tp->write_seq + copy);
1366                 TCP_SKB_CB(skb)->end_seq += copy;
1367                 tcp_skb_pcount_set(skb, 0);
1368
1369                 copied += copy;
1370                 if (!msg_data_left(msg)) {
1371                         if (unlikely(flags & MSG_EOR))
1372                                 TCP_SKB_CB(skb)->eor = 1;
1373                         goto out;
1374                 }
1375
1376                 if (skb->len < size_goal || (flags & MSG_OOB) || unlikely(tp->repair))
1377                         continue;
1378
1379                 if (forced_push(tp)) {
1380                         tcp_mark_push(tp, skb);
1381                         __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
1382                 } else if (skb == tcp_send_head(sk))
1383                         tcp_push_one(sk, mss_now);
1384                 continue;
1385
1386 wait_for_sndbuf:
1387                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1388 wait_for_memory:
1389                 if (copied)
1390                         tcp_push(sk, flags & ~MSG_MORE, mss_now,
1391                                  TCP_NAGLE_PUSH, size_goal);
1392
1393                 err = sk_stream_wait_memory(sk, &timeo);
1394                 if (err != 0)
1395                         goto do_error;
1396
1397                 mss_now = tcp_send_mss(sk, &size_goal, flags);
1398         }
1399
1400 out:
1401         if (copied) {
1402                 tcp_tx_timestamp(sk, sockc.tsflags);
1403                 tcp_push(sk, flags, mss_now, tp->nonagle, size_goal);
1404         }
1405 out_nopush:
1406         sock_zerocopy_put(uarg);
1407         return copied + copied_syn;
1408
1409 do_error:
1410         skb = tcp_write_queue_tail(sk);
1411 do_fault:
1412         tcp_remove_empty_skb(sk, skb);
1413
1414         if (copied + copied_syn)
1415                 goto out;
1416 out_err:
1417         sock_zerocopy_put_abort(uarg, true);
1418         err = sk_stream_error(sk, flags, err);
1419         /* make sure we wake any epoll edge trigger waiter */
1420         if (unlikely(tcp_rtx_and_write_queues_empty(sk) && err == -EAGAIN)) {
1421                 sk->sk_write_space(sk);
1422                 tcp_chrono_stop(sk, TCP_CHRONO_SNDBUF_LIMITED);
1423         }
1424         return err;
1425 }
1426 EXPORT_SYMBOL_GPL(tcp_sendmsg_locked);
1427
1428 int tcp_sendmsg(struct sock *sk, struct msghdr *msg, size_t size)
1429 {
1430         int ret;
1431
1432         lock_sock(sk);
1433         ret = tcp_sendmsg_locked(sk, msg, size);
1434         release_sock(sk);
1435
1436         return ret;
1437 }
1438 EXPORT_SYMBOL(tcp_sendmsg);
1439
1440 /*
1441  *      Handle reading urgent data. BSD has very simple semantics for
1442  *      this, no blocking and very strange errors 8)
1443  */
1444
1445 static int tcp_recv_urg(struct sock *sk, struct msghdr *msg, int len, int flags)
1446 {
1447         struct tcp_sock *tp = tcp_sk(sk);
1448
1449         /* No URG data to read. */
1450         if (sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data ||
1451             tp->urg_data == TCP_URG_READ)
1452                 return -EINVAL; /* Yes this is right ! */
1453
1454         if (sk->sk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DONE))
1455                 return -ENOTCONN;
1456
1457         if (tp->urg_data & TCP_URG_VALID) {
1458                 int err = 0;
1459                 char c = tp->urg_data;
1460
1461                 if (!(flags & MSG_PEEK))
1462                         tp->urg_data = TCP_URG_READ;
1463
1464                 /* Read urgent data. */
1465                 msg->msg_flags |= MSG_OOB;
1466
1467                 if (len > 0) {
1468                         if (!(flags & MSG_TRUNC))
1469                                 err = memcpy_to_msg(msg, &c, 1);
1470                         len = 1;
1471                 } else
1472                         msg->msg_flags |= MSG_TRUNC;
1473
1474                 return err ? -EFAULT : len;
1475         }
1476
1477         if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
1478                 return 0;
1479
1480         /* Fixed the recv(..., MSG_OOB) behaviour.  BSD docs and
1481          * the available implementations agree in this case:
1482          * this call should never block, independent of the
1483          * blocking state of the socket.
1484          * Mike <pall@rz.uni-karlsruhe.de>
1485          */
1486         return -EAGAIN;
1487 }
1488
1489 static int tcp_peek_sndq(struct sock *sk, struct msghdr *msg, int len)
1490 {
1491         struct sk_buff *skb;
1492         int copied = 0, err = 0;
1493
1494         /* XXX -- need to support SO_PEEK_OFF */
1495
1496         skb_rbtree_walk(skb, &sk->tcp_rtx_queue) {
1497                 err = skb_copy_datagram_msg(skb, 0, msg, skb->len);
1498                 if (err)
1499                         return err;
1500                 copied += skb->len;
1501         }
1502
1503         skb_queue_walk(&sk->sk_write_queue, skb) {
1504                 err = skb_copy_datagram_msg(skb, 0, msg, skb->len);
1505                 if (err)
1506                         break;
1507
1508                 copied += skb->len;
1509         }
1510
1511         return err ?: copied;
1512 }
1513
1514 /* Clean up the receive buffer for full frames taken by the user,
1515  * then send an ACK if necessary.  COPIED is the number of bytes
1516  * tcp_recvmsg has given to the user so far, it speeds up the
1517  * calculation of whether or not we must ACK for the sake of
1518  * a window update.
1519  */
1520 static void tcp_cleanup_rbuf(struct sock *sk, int copied)
1521 {
1522         struct tcp_sock *tp = tcp_sk(sk);
1523         bool time_to_ack = false;
1524
1525         struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
1526
1527         WARN(skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq),
1528              "cleanup rbuf bug: copied %X seq %X rcvnxt %X\n",
1529              tp->copied_seq, TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt);
1530
1531         if (inet_csk_ack_scheduled(sk)) {
1532                 const struct inet_connection_sock *icsk = inet_csk(sk);
1533                    /* Delayed ACKs frequently hit locked sockets during bulk
1534                     * receive. */
1535                 if (icsk->icsk_ack.blocked ||
1536                     /* Once-per-two-segments ACK was not sent by tcp_input.c */
1537                     tp->rcv_nxt - tp->rcv_wup > icsk->icsk_ack.rcv_mss ||
1538                     /*
1539                      * If this read emptied read buffer, we send ACK, if
1540                      * connection is not bidirectional, user drained
1541                      * receive buffer and there was a small segment
1542                      * in queue.
1543                      */
1544                     (copied > 0 &&
1545                      ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED2) ||
1546                       ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED) &&
1547                        !inet_csk_in_pingpong_mode(sk))) &&
1548                       !atomic_read(&sk->sk_rmem_alloc)))
1549                         time_to_ack = true;
1550         }
1551
1552         /* We send an ACK if we can now advertise a non-zero window
1553          * which has been raised "significantly".
1554          *
1555          * Even if window raised up to infinity, do not send window open ACK
1556          * in states, where we will not receive more. It is useless.
1557          */
1558         if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) {
1559                 __u32 rcv_window_now = tcp_receive_window(tp);
1560
1561                 /* Optimize, __tcp_select_window() is not cheap. */
1562                 if (2*rcv_window_now <= tp->window_clamp) {
1563                         __u32 new_window = __tcp_select_window(sk);
1564
1565                         /* Send ACK now, if this read freed lots of space
1566                          * in our buffer. Certainly, new_window is new window.
1567                          * We can advertise it now, if it is not less than current one.
1568                          * "Lots" means "at least twice" here.
1569                          */
1570                         if (new_window && new_window >= 2 * rcv_window_now)
1571                                 time_to_ack = true;
1572                 }
1573         }
1574         if (time_to_ack)
1575                 tcp_send_ack(sk);
1576 }
1577
1578 static struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off)
1579 {
1580         struct sk_buff *skb;
1581         u32 offset;
1582
1583         while ((skb = skb_peek(&sk->sk_receive_queue)) != NULL) {
1584                 offset = seq - TCP_SKB_CB(skb)->seq;
1585                 if (unlikely(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
1586                         pr_err_once("%s: found a SYN, please report !\n", __func__);
1587                         offset--;
1588                 }
1589                 if (offset < skb->len || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)) {
1590                         *off = offset;
1591                         return skb;
1592                 }
1593                 /* This looks weird, but this can happen if TCP collapsing
1594                  * splitted a fat GRO packet, while we released socket lock
1595                  * in skb_splice_bits()
1596                  */
1597                 sk_eat_skb(sk, skb);
1598         }
1599         return NULL;
1600 }
1601
1602 /*
1603  * This routine provides an alternative to tcp_recvmsg() for routines
1604  * that would like to handle copying from skbuffs directly in 'sendfile'
1605  * fashion.
1606  * Note:
1607  *      - It is assumed that the socket was locked by the caller.
1608  *      - The routine does not block.
1609  *      - At present, there is no support for reading OOB data
1610  *        or for 'peeking' the socket using this routine
1611  *        (although both would be easy to implement).
1612  */
1613 int tcp_read_sock(struct sock *sk, read_descriptor_t *desc,
1614                   sk_read_actor_t recv_actor)
1615 {
1616         struct sk_buff *skb;
1617         struct tcp_sock *tp = tcp_sk(sk);
1618         u32 seq = tp->copied_seq;
1619         u32 offset;
1620         int copied = 0;
1621
1622         if (sk->sk_state == TCP_LISTEN)
1623                 return -ENOTCONN;
1624         while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1625                 if (offset < skb->len) {
1626                         int used;
1627                         size_t len;
1628
1629                         len = skb->len - offset;
1630                         /* Stop reading if we hit a patch of urgent data */
1631                         if (tp->urg_data) {
1632                                 u32 urg_offset = tp->urg_seq - seq;
1633                                 if (urg_offset < len)
1634                                         len = urg_offset;
1635                                 if (!len)
1636                                         break;
1637                         }
1638                         used = recv_actor(desc, skb, offset, len);
1639                         if (used <= 0) {
1640                                 if (!copied)
1641                                         copied = used;
1642                                 break;
1643                         } else if (used <= len) {
1644                                 seq += used;
1645                                 copied += used;
1646                                 offset += used;
1647                         }
1648                         /* If recv_actor drops the lock (e.g. TCP splice
1649                          * receive) the skb pointer might be invalid when
1650                          * getting here: tcp_collapse might have deleted it
1651                          * while aggregating skbs from the socket queue.
1652                          */
1653                         skb = tcp_recv_skb(sk, seq - 1, &offset);
1654                         if (!skb)
1655                                 break;
1656                         /* TCP coalescing might have appended data to the skb.
1657                          * Try to splice more frags
1658                          */
1659                         if (offset + 1 != skb->len)
1660                                 continue;
1661                 }
1662                 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) {
1663                         sk_eat_skb(sk, skb);
1664                         ++seq;
1665                         break;
1666                 }
1667                 sk_eat_skb(sk, skb);
1668                 if (!desc->count)
1669                         break;
1670                 WRITE_ONCE(tp->copied_seq, seq);
1671         }
1672         WRITE_ONCE(tp->copied_seq, seq);
1673
1674         tcp_rcv_space_adjust(sk);
1675
1676         /* Clean up data we have read: This will do ACK frames. */
1677         if (copied > 0) {
1678                 tcp_recv_skb(sk, seq, &offset);
1679                 tcp_cleanup_rbuf(sk, copied);
1680         }
1681         return copied;
1682 }
1683 EXPORT_SYMBOL(tcp_read_sock);
1684
1685 int tcp_peek_len(struct socket *sock)
1686 {
1687         return tcp_inq(sock->sk);
1688 }
1689 EXPORT_SYMBOL(tcp_peek_len);
1690
1691 /* Make sure sk_rcvbuf is big enough to satisfy SO_RCVLOWAT hint */
1692 int tcp_set_rcvlowat(struct sock *sk, int val)
1693 {
1694         int cap;
1695
1696         if (sk->sk_userlocks & SOCK_RCVBUF_LOCK)
1697                 cap = sk->sk_rcvbuf >> 1;
1698         else
1699                 cap = sock_net(sk)->ipv4.sysctl_tcp_rmem[2] >> 1;
1700         val = min(val, cap);
1701         WRITE_ONCE(sk->sk_rcvlowat, val ? : 1);
1702
1703         /* Check if we need to signal EPOLLIN right now */
1704         tcp_data_ready(sk);
1705
1706         if (sk->sk_userlocks & SOCK_RCVBUF_LOCK)
1707                 return 0;
1708
1709         val <<= 1;
1710         if (val > sk->sk_rcvbuf) {
1711                 WRITE_ONCE(sk->sk_rcvbuf, val);
1712                 tcp_sk(sk)->window_clamp = tcp_win_from_space(sk, val);
1713         }
1714         return 0;
1715 }
1716 EXPORT_SYMBOL(tcp_set_rcvlowat);
1717
1718 #ifdef CONFIG_MMU
1719 static const struct vm_operations_struct tcp_vm_ops = {
1720 };
1721
1722 int tcp_mmap(struct file *file, struct socket *sock,
1723              struct vm_area_struct *vma)
1724 {
1725         if (vma->vm_flags & (VM_WRITE | VM_EXEC))
1726                 return -EPERM;
1727         vma->vm_flags &= ~(VM_MAYWRITE | VM_MAYEXEC);
1728
1729         /* Instruct vm_insert_page() to not down_read(mmap_sem) */
1730         vma->vm_flags |= VM_MIXEDMAP;
1731
1732         vma->vm_ops = &tcp_vm_ops;
1733         return 0;
1734 }
1735 EXPORT_SYMBOL(tcp_mmap);
1736
1737 static int tcp_zerocopy_receive(struct sock *sk,
1738                                 struct tcp_zerocopy_receive *zc)
1739 {
1740         unsigned long address = (unsigned long)zc->address;
1741         u32 length = 0, seq, offset, zap_len;
1742         const skb_frag_t *frags = NULL;
1743         struct vm_area_struct *vma;
1744         struct sk_buff *skb = NULL;
1745         struct tcp_sock *tp;
1746         int inq;
1747         int ret;
1748
1749         if (address & (PAGE_SIZE - 1) || address != zc->address)
1750                 return -EINVAL;
1751
1752         if (sk->sk_state == TCP_LISTEN)
1753                 return -ENOTCONN;
1754
1755         sock_rps_record_flow(sk);
1756
1757         down_read(&current->mm->mmap_sem);
1758
1759         ret = -EINVAL;
1760         vma = find_vma(current->mm, address);
1761         if (!vma || vma->vm_start > address || vma->vm_ops != &tcp_vm_ops)
1762                 goto out;
1763         zc->length = min_t(unsigned long, zc->length, vma->vm_end - address);
1764
1765         tp = tcp_sk(sk);
1766         seq = tp->copied_seq;
1767         inq = tcp_inq(sk);
1768         zc->length = min_t(u32, zc->length, inq);
1769         zap_len = zc->length & ~(PAGE_SIZE - 1);
1770         if (zap_len) {
1771                 zap_page_range(vma, address, zap_len);
1772                 zc->recv_skip_hint = 0;
1773         } else {
1774                 zc->recv_skip_hint = zc->length;
1775         }
1776         ret = 0;
1777         while (length + PAGE_SIZE <= zc->length) {
1778                 if (zc->recv_skip_hint < PAGE_SIZE) {
1779                         if (skb) {
1780                                 if (zc->recv_skip_hint > 0)
1781                                         break;
1782                                 skb = skb->next;
1783                                 offset = seq - TCP_SKB_CB(skb)->seq;
1784                         } else {
1785                                 skb = tcp_recv_skb(sk, seq, &offset);
1786                         }
1787
1788                         zc->recv_skip_hint = skb->len - offset;
1789                         offset -= skb_headlen(skb);
1790                         if ((int)offset < 0 || skb_has_frag_list(skb))
1791                                 break;
1792                         frags = skb_shinfo(skb)->frags;
1793                         while (offset) {
1794                                 if (skb_frag_size(frags) > offset)
1795                                         goto out;
1796                                 offset -= skb_frag_size(frags);
1797                                 frags++;
1798                         }
1799                 }
1800                 if (skb_frag_size(frags) != PAGE_SIZE || skb_frag_off(frags)) {
1801                         int remaining = zc->recv_skip_hint;
1802
1803                         while (remaining && (skb_frag_size(frags) != PAGE_SIZE ||
1804                                              skb_frag_off(frags))) {
1805                                 remaining -= skb_frag_size(frags);
1806                                 frags++;
1807                         }
1808                         zc->recv_skip_hint -= remaining;
1809                         break;
1810                 }
1811                 ret = vm_insert_page(vma, address + length,
1812                                      skb_frag_page(frags));
1813                 if (ret)
1814                         break;
1815                 length += PAGE_SIZE;
1816                 seq += PAGE_SIZE;
1817                 zc->recv_skip_hint -= PAGE_SIZE;
1818                 frags++;
1819         }
1820 out:
1821         up_read(&current->mm->mmap_sem);
1822         if (length) {
1823                 WRITE_ONCE(tp->copied_seq, seq);
1824                 tcp_rcv_space_adjust(sk);
1825
1826                 /* Clean up data we have read: This will do ACK frames. */
1827                 tcp_recv_skb(sk, seq, &offset);
1828                 tcp_cleanup_rbuf(sk, length);
1829                 ret = 0;
1830                 if (length == zc->length)
1831                         zc->recv_skip_hint = 0;
1832         } else {
1833                 if (!zc->recv_skip_hint && sock_flag(sk, SOCK_DONE))
1834                         ret = -EIO;
1835         }
1836         zc->length = length;
1837         return ret;
1838 }
1839 #endif
1840
1841 static void tcp_update_recv_tstamps(struct sk_buff *skb,
1842                                     struct scm_timestamping_internal *tss)
1843 {
1844         if (skb->tstamp)
1845                 tss->ts[0] = ktime_to_timespec64(skb->tstamp);
1846         else
1847                 tss->ts[0] = (struct timespec64) {0};
1848
1849         if (skb_hwtstamps(skb)->hwtstamp)
1850                 tss->ts[2] = ktime_to_timespec64(skb_hwtstamps(skb)->hwtstamp);
1851         else
1852                 tss->ts[2] = (struct timespec64) {0};
1853 }
1854
1855 /* Similar to __sock_recv_timestamp, but does not require an skb */
1856 static void tcp_recv_timestamp(struct msghdr *msg, const struct sock *sk,
1857                                struct scm_timestamping_internal *tss)
1858 {
1859         int new_tstamp = sock_flag(sk, SOCK_TSTAMP_NEW);
1860         bool has_timestamping = false;
1861
1862         if (tss->ts[0].tv_sec || tss->ts[0].tv_nsec) {
1863                 if (sock_flag(sk, SOCK_RCVTSTAMP)) {
1864                         if (sock_flag(sk, SOCK_RCVTSTAMPNS)) {
1865                                 if (new_tstamp) {
1866                                         struct __kernel_timespec kts = {
1867                                                 .tv_sec = tss->ts[0].tv_sec,
1868                                                 .tv_nsec = tss->ts[0].tv_nsec,
1869                                         };
1870                                         put_cmsg(msg, SOL_SOCKET, SO_TIMESTAMPNS_NEW,
1871                                                  sizeof(kts), &kts);
1872                                 } else {
1873                                         struct __kernel_old_timespec ts_old = {
1874                                                 .tv_sec = tss->ts[0].tv_sec,
1875                                                 .tv_nsec = tss->ts[0].tv_nsec,
1876                                         };
1877                                         put_cmsg(msg, SOL_SOCKET, SO_TIMESTAMPNS_OLD,
1878                                                  sizeof(ts_old), &ts_old);
1879                                 }
1880                         } else {
1881                                 if (new_tstamp) {
1882                                         struct __kernel_sock_timeval stv = {
1883                                                 .tv_sec = tss->ts[0].tv_sec,
1884                                                 .tv_usec = tss->ts[0].tv_nsec / 1000,
1885                                         };
1886                                         put_cmsg(msg, SOL_SOCKET, SO_TIMESTAMP_NEW,
1887                                                  sizeof(stv), &stv);
1888                                 } else {
1889                                         struct __kernel_old_timeval tv = {
1890                                                 .tv_sec = tss->ts[0].tv_sec,
1891                                                 .tv_usec = tss->ts[0].tv_nsec / 1000,
1892                                         };
1893                                         put_cmsg(msg, SOL_SOCKET, SO_TIMESTAMP_OLD,
1894                                                  sizeof(tv), &tv);
1895                                 }
1896                         }
1897                 }
1898
1899                 if (sk->sk_tsflags & SOF_TIMESTAMPING_SOFTWARE)
1900                         has_timestamping = true;
1901                 else
1902                         tss->ts[0] = (struct timespec64) {0};
1903         }
1904
1905         if (tss->ts[2].tv_sec || tss->ts[2].tv_nsec) {
1906                 if (sk->sk_tsflags & SOF_TIMESTAMPING_RAW_HARDWARE)
1907                         has_timestamping = true;
1908                 else
1909                         tss->ts[2] = (struct timespec64) {0};
1910         }
1911
1912         if (has_timestamping) {
1913                 tss->ts[1] = (struct timespec64) {0};
1914                 if (sock_flag(sk, SOCK_TSTAMP_NEW))
1915                         put_cmsg_scm_timestamping64(msg, tss);
1916                 else
1917                         put_cmsg_scm_timestamping(msg, tss);
1918         }
1919 }
1920
1921 static int tcp_inq_hint(struct sock *sk)
1922 {
1923         const struct tcp_sock *tp = tcp_sk(sk);
1924         u32 copied_seq = READ_ONCE(tp->copied_seq);
1925         u32 rcv_nxt = READ_ONCE(tp->rcv_nxt);
1926         int inq;
1927
1928         inq = rcv_nxt - copied_seq;
1929         if (unlikely(inq < 0 || copied_seq != READ_ONCE(tp->copied_seq))) {
1930                 lock_sock(sk);
1931                 inq = tp->rcv_nxt - tp->copied_seq;
1932                 release_sock(sk);
1933         }
1934         /* After receiving a FIN, tell the user-space to continue reading
1935          * by returning a non-zero inq.
1936          */
1937         if (inq == 0 && sock_flag(sk, SOCK_DONE))
1938                 inq = 1;
1939         return inq;
1940 }
1941
1942 /*
1943  *      This routine copies from a sock struct into the user buffer.
1944  *
1945  *      Technical note: in 2.3 we work on _locked_ socket, so that
1946  *      tricks with *seq access order and skb->users are not required.
1947  *      Probably, code can be easily improved even more.
1948  */
1949
1950 int tcp_recvmsg(struct sock *sk, struct msghdr *msg, size_t len, int nonblock,
1951                 int flags, int *addr_len)
1952 {
1953         struct tcp_sock *tp = tcp_sk(sk);
1954         int copied = 0;
1955         u32 peek_seq;
1956         u32 *seq;
1957         unsigned long used;
1958         int err, inq;
1959         int target;             /* Read at least this many bytes */
1960         long timeo;
1961         struct sk_buff *skb, *last;
1962         u32 urg_hole = 0;
1963         struct scm_timestamping_internal tss;
1964         int cmsg_flags;
1965
1966         if (unlikely(flags & MSG_ERRQUEUE))
1967                 return inet_recv_error(sk, msg, len, addr_len);
1968
1969         if (sk_can_busy_loop(sk) && skb_queue_empty_lockless(&sk->sk_receive_queue) &&
1970             (sk->sk_state == TCP_ESTABLISHED))
1971                 sk_busy_loop(sk, nonblock);
1972
1973         lock_sock(sk);
1974
1975         err = -ENOTCONN;
1976         if (sk->sk_state == TCP_LISTEN)
1977                 goto out;
1978
1979         cmsg_flags = tp->recvmsg_inq ? 1 : 0;
1980         timeo = sock_rcvtimeo(sk, nonblock);
1981
1982         /* Urgent data needs to be handled specially. */
1983         if (flags & MSG_OOB)
1984                 goto recv_urg;
1985
1986         if (unlikely(tp->repair)) {
1987                 err = -EPERM;
1988                 if (!(flags & MSG_PEEK))
1989                         goto out;
1990
1991                 if (tp->repair_queue == TCP_SEND_QUEUE)
1992                         goto recv_sndq;
1993
1994                 err = -EINVAL;
1995                 if (tp->repair_queue == TCP_NO_QUEUE)
1996                         goto out;
1997
1998                 /* 'common' recv queue MSG_PEEK-ing */
1999         }
2000
2001         seq = &tp->copied_seq;
2002         if (flags & MSG_PEEK) {
2003                 peek_seq = tp->copied_seq;
2004                 seq = &peek_seq;
2005         }
2006
2007         target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
2008
2009         do {
2010                 u32 offset;
2011
2012                 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
2013                 if (tp->urg_data && tp->urg_seq == *seq) {
2014                         if (copied)
2015                                 break;
2016                         if (signal_pending(current)) {
2017                                 copied = timeo ? sock_intr_errno(timeo) : -EAGAIN;
2018                                 break;
2019                         }
2020                 }
2021
2022                 /* Next get a buffer. */
2023
2024                 last = skb_peek_tail(&sk->sk_receive_queue);
2025                 skb_queue_walk(&sk->sk_receive_queue, skb) {
2026                         last = skb;
2027                         /* Now that we have two receive queues this
2028                          * shouldn't happen.
2029                          */
2030                         if (WARN(before(*seq, TCP_SKB_CB(skb)->seq),
2031                                  "TCP recvmsg seq # bug: copied %X, seq %X, rcvnxt %X, fl %X\n",
2032                                  *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt,
2033                                  flags))
2034                                 break;
2035
2036                         offset = *seq - TCP_SKB_CB(skb)->seq;
2037                         if (unlikely(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
2038                                 pr_err_once("%s: found a SYN, please report !\n", __func__);
2039                                 offset--;
2040                         }
2041                         if (offset < skb->len)
2042                                 goto found_ok_skb;
2043                         if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
2044                                 goto found_fin_ok;
2045                         WARN(!(flags & MSG_PEEK),
2046                              "TCP recvmsg seq # bug 2: copied %X, seq %X, rcvnxt %X, fl %X\n",
2047                              *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt, flags);
2048                 }
2049
2050                 /* Well, if we have backlog, try to process it now yet. */
2051
2052                 if (copied >= target && !READ_ONCE(sk->sk_backlog.tail))
2053                         break;
2054
2055                 if (copied) {
2056                         if (sk->sk_err ||
2057                             sk->sk_state == TCP_CLOSE ||
2058                             (sk->sk_shutdown & RCV_SHUTDOWN) ||
2059                             !timeo ||
2060                             signal_pending(current))
2061                                 break;
2062                 } else {
2063                         if (sock_flag(sk, SOCK_DONE))
2064                                 break;
2065
2066                         if (sk->sk_err) {
2067                                 copied = sock_error(sk);
2068                                 break;
2069                         }
2070
2071                         if (sk->sk_shutdown & RCV_SHUTDOWN)
2072                                 break;
2073
2074                         if (sk->sk_state == TCP_CLOSE) {
2075                                 /* This occurs when user tries to read
2076                                  * from never connected socket.
2077                                  */
2078                                 copied = -ENOTCONN;
2079                                 break;
2080                         }
2081
2082                         if (!timeo) {
2083                                 copied = -EAGAIN;
2084                                 break;
2085                         }
2086
2087                         if (signal_pending(current)) {
2088                                 copied = sock_intr_errno(timeo);
2089                                 break;
2090                         }
2091                 }
2092
2093                 tcp_cleanup_rbuf(sk, copied);
2094
2095                 if (copied >= target) {
2096                         /* Do not sleep, just process backlog. */
2097                         release_sock(sk);
2098                         lock_sock(sk);
2099                 } else {
2100                         sk_wait_data(sk, &timeo, last);
2101                 }
2102
2103                 if ((flags & MSG_PEEK) &&
2104                     (peek_seq - copied - urg_hole != tp->copied_seq)) {
2105                         net_dbg_ratelimited("TCP(%s:%d): Application bug, race in MSG_PEEK\n",
2106                                             current->comm,
2107                                             task_pid_nr(current));
2108                         peek_seq = tp->copied_seq;
2109                 }
2110                 continue;
2111
2112 found_ok_skb:
2113                 /* Ok so how much can we use? */
2114                 used = skb->len - offset;
2115                 if (len < used)
2116                         used = len;
2117
2118                 /* Do we have urgent data here? */
2119                 if (tp->urg_data) {
2120                         u32 urg_offset = tp->urg_seq - *seq;
2121                         if (urg_offset < used) {
2122                                 if (!urg_offset) {
2123                                         if (!sock_flag(sk, SOCK_URGINLINE)) {
2124                                                 WRITE_ONCE(*seq, *seq + 1);
2125                                                 urg_hole++;
2126                                                 offset++;
2127                                                 used--;
2128                                                 if (!used)
2129                                                         goto skip_copy;
2130                                         }
2131                                 } else
2132                                         used = urg_offset;
2133                         }
2134                 }
2135
2136                 if (!(flags & MSG_TRUNC)) {
2137                         err = skb_copy_datagram_msg(skb, offset, msg, used);
2138                         if (err) {
2139                                 /* Exception. Bailout! */
2140                                 if (!copied)
2141                                         copied = -EFAULT;
2142                                 break;
2143                         }
2144                 }
2145
2146                 WRITE_ONCE(*seq, *seq + used);
2147                 copied += used;
2148                 len -= used;
2149
2150                 tcp_rcv_space_adjust(sk);
2151
2152 skip_copy:
2153                 if (tp->urg_data && after(tp->copied_seq, tp->urg_seq)) {
2154                         tp->urg_data = 0;
2155                         tcp_fast_path_check(sk);
2156                 }
2157
2158                 if (TCP_SKB_CB(skb)->has_rxtstamp) {
2159                         tcp_update_recv_tstamps(skb, &tss);
2160                         cmsg_flags |= 2;
2161                 }
2162
2163                 if (used + offset < skb->len)
2164                         continue;
2165
2166                 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
2167                         goto found_fin_ok;
2168                 if (!(flags & MSG_PEEK))
2169                         sk_eat_skb(sk, skb);
2170                 continue;
2171
2172 found_fin_ok:
2173                 /* Process the FIN. */
2174                 WRITE_ONCE(*seq, *seq + 1);
2175                 if (!(flags & MSG_PEEK))
2176                         sk_eat_skb(sk, skb);
2177                 break;
2178         } while (len > 0);
2179
2180         /* According to UNIX98, msg_name/msg_namelen are ignored
2181          * on connected socket. I was just happy when found this 8) --ANK
2182          */
2183
2184         /* Clean up data we have read: This will do ACK frames. */
2185         tcp_cleanup_rbuf(sk, copied);
2186
2187         release_sock(sk);
2188
2189         if (cmsg_flags) {
2190                 if (cmsg_flags & 2)
2191                         tcp_recv_timestamp(msg, sk, &tss);
2192                 if (cmsg_flags & 1) {
2193                         inq = tcp_inq_hint(sk);
2194                         put_cmsg(msg, SOL_TCP, TCP_CM_INQ, sizeof(inq), &inq);
2195                 }
2196         }
2197
2198         return copied;
2199
2200 out:
2201         release_sock(sk);
2202         return err;
2203
2204 recv_urg:
2205         err = tcp_recv_urg(sk, msg, len, flags);
2206         goto out;
2207
2208 recv_sndq:
2209         err = tcp_peek_sndq(sk, msg, len);
2210         goto out;
2211 }
2212 EXPORT_SYMBOL(tcp_recvmsg);
2213
2214 void tcp_set_state(struct sock *sk, int state)
2215 {
2216         int oldstate = sk->sk_state;
2217
2218         /* We defined a new enum for TCP states that are exported in BPF
2219          * so as not force the internal TCP states to be frozen. The
2220          * following checks will detect if an internal state value ever
2221          * differs from the BPF value. If this ever happens, then we will
2222          * need to remap the internal value to the BPF value before calling
2223          * tcp_call_bpf_2arg.
2224          */
2225         BUILD_BUG_ON((int)BPF_TCP_ESTABLISHED != (int)TCP_ESTABLISHED);
2226         BUILD_BUG_ON((int)BPF_TCP_SYN_SENT != (int)TCP_SYN_SENT);
2227         BUILD_BUG_ON((int)BPF_TCP_SYN_RECV != (int)TCP_SYN_RECV);
2228         BUILD_BUG_ON((int)BPF_TCP_FIN_WAIT1 != (int)TCP_FIN_WAIT1);
2229         BUILD_BUG_ON((int)BPF_TCP_FIN_WAIT2 != (int)TCP_FIN_WAIT2);
2230         BUILD_BUG_ON((int)BPF_TCP_TIME_WAIT != (int)TCP_TIME_WAIT);
2231         BUILD_BUG_ON((int)BPF_TCP_CLOSE != (int)TCP_CLOSE);
2232         BUILD_BUG_ON((int)BPF_TCP_CLOSE_WAIT != (int)TCP_CLOSE_WAIT);
2233         BUILD_BUG_ON((int)BPF_TCP_LAST_ACK != (int)TCP_LAST_ACK);
2234         BUILD_BUG_ON((int)BPF_TCP_LISTEN != (int)TCP_LISTEN);
2235         BUILD_BUG_ON((int)BPF_TCP_CLOSING != (int)TCP_CLOSING);
2236         BUILD_BUG_ON((int)BPF_TCP_NEW_SYN_RECV != (int)TCP_NEW_SYN_RECV);
2237         BUILD_BUG_ON((int)BPF_TCP_MAX_STATES != (int)TCP_MAX_STATES);
2238
2239         if (BPF_SOCK_OPS_TEST_FLAG(tcp_sk(sk), BPF_SOCK_OPS_STATE_CB_FLAG))
2240                 tcp_call_bpf_2arg(sk, BPF_SOCK_OPS_STATE_CB, oldstate, state);
2241
2242         switch (state) {
2243         case TCP_ESTABLISHED:
2244                 if (oldstate != TCP_ESTABLISHED)
2245                         TCP_INC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
2246                 break;
2247
2248         case TCP_CLOSE:
2249                 if (oldstate == TCP_CLOSE_WAIT || oldstate == TCP_ESTABLISHED)
2250                         TCP_INC_STATS(sock_net(sk), TCP_MIB_ESTABRESETS);
2251
2252                 sk->sk_prot->unhash(sk);
2253                 if (inet_csk(sk)->icsk_bind_hash &&
2254                     !(sk->sk_userlocks & SOCK_BINDPORT_LOCK))
2255                         inet_put_port(sk);
2256                 fallthrough;
2257         default:
2258                 if (oldstate == TCP_ESTABLISHED)
2259                         TCP_DEC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
2260         }
2261
2262         /* Change state AFTER socket is unhashed to avoid closed
2263          * socket sitting in hash tables.
2264          */
2265         inet_sk_state_store(sk, state);
2266 }
2267 EXPORT_SYMBOL_GPL(tcp_set_state);
2268
2269 /*
2270  *      State processing on a close. This implements the state shift for
2271  *      sending our FIN frame. Note that we only send a FIN for some
2272  *      states. A shutdown() may have already sent the FIN, or we may be
2273  *      closed.
2274  */
2275
2276 static const unsigned char new_state[16] = {
2277   /* current state:        new state:      action:      */
2278   [0 /* (Invalid) */]   = TCP_CLOSE,
2279   [TCP_ESTABLISHED]     = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2280   [TCP_SYN_SENT]        = TCP_CLOSE,
2281   [TCP_SYN_RECV]        = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2282   [TCP_FIN_WAIT1]       = TCP_FIN_WAIT1,
2283   [TCP_FIN_WAIT2]       = TCP_FIN_WAIT2,
2284   [TCP_TIME_WAIT]       = TCP_CLOSE,
2285   [TCP_CLOSE]           = TCP_CLOSE,
2286   [TCP_CLOSE_WAIT]      = TCP_LAST_ACK  | TCP_ACTION_FIN,
2287   [TCP_LAST_ACK]        = TCP_LAST_ACK,
2288   [TCP_LISTEN]          = TCP_CLOSE,
2289   [TCP_CLOSING]         = TCP_CLOSING,
2290   [TCP_NEW_SYN_RECV]    = TCP_CLOSE,    /* should not happen ! */
2291 };
2292
2293 static int tcp_close_state(struct sock *sk)
2294 {
2295         int next = (int)new_state[sk->sk_state];
2296         int ns = next & TCP_STATE_MASK;
2297
2298         tcp_set_state(sk, ns);
2299
2300         return next & TCP_ACTION_FIN;
2301 }
2302
2303 /*
2304  *      Shutdown the sending side of a connection. Much like close except
2305  *      that we don't receive shut down or sock_set_flag(sk, SOCK_DEAD).
2306  */
2307
2308 void tcp_shutdown(struct sock *sk, int how)
2309 {
2310         /*      We need to grab some memory, and put together a FIN,
2311          *      and then put it into the queue to be sent.
2312          *              Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
2313          */
2314         if (!(how & SEND_SHUTDOWN))
2315                 return;
2316
2317         /* If we've already sent a FIN, or it's a closed state, skip this. */
2318         if ((1 << sk->sk_state) &
2319             (TCPF_ESTABLISHED | TCPF_SYN_SENT |
2320              TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) {
2321                 /* Clear out any half completed packets.  FIN if needed. */
2322                 if (tcp_close_state(sk))
2323                         tcp_send_fin(sk);
2324         }
2325 }
2326 EXPORT_SYMBOL(tcp_shutdown);
2327
2328 bool tcp_check_oom(struct sock *sk, int shift)
2329 {
2330         bool too_many_orphans, out_of_socket_memory;
2331
2332         too_many_orphans = tcp_too_many_orphans(sk, shift);
2333         out_of_socket_memory = tcp_out_of_memory(sk);
2334
2335         if (too_many_orphans)
2336                 net_info_ratelimited("too many orphaned sockets\n");
2337         if (out_of_socket_memory)
2338                 net_info_ratelimited("out of memory -- consider tuning tcp_mem\n");
2339         return too_many_orphans || out_of_socket_memory;
2340 }
2341
2342 void tcp_close(struct sock *sk, long timeout)
2343 {
2344         struct sk_buff *skb;
2345         int data_was_unread = 0;
2346         int state;
2347
2348         lock_sock(sk);
2349         sk->sk_shutdown = SHUTDOWN_MASK;
2350
2351         if (sk->sk_state == TCP_LISTEN) {
2352                 tcp_set_state(sk, TCP_CLOSE);
2353
2354                 /* Special case. */
2355                 inet_csk_listen_stop(sk);
2356
2357                 goto adjudge_to_death;
2358         }
2359
2360         /*  We need to flush the recv. buffs.  We do this only on the
2361          *  descriptor close, not protocol-sourced closes, because the
2362          *  reader process may not have drained the data yet!
2363          */
2364         while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) {
2365                 u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq;
2366
2367                 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
2368                         len--;
2369                 data_was_unread += len;
2370                 __kfree_skb(skb);
2371         }
2372
2373         sk_mem_reclaim(sk);
2374
2375         /* If socket has been already reset (e.g. in tcp_reset()) - kill it. */
2376         if (sk->sk_state == TCP_CLOSE)
2377                 goto adjudge_to_death;
2378
2379         /* As outlined in RFC 2525, section 2.17, we send a RST here because
2380          * data was lost. To witness the awful effects of the old behavior of
2381          * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk
2382          * GET in an FTP client, suspend the process, wait for the client to
2383          * advertise a zero window, then kill -9 the FTP client, wheee...
2384          * Note: timeout is always zero in such a case.
2385          */
2386         if (unlikely(tcp_sk(sk)->repair)) {
2387                 sk->sk_prot->disconnect(sk, 0);
2388         } else if (data_was_unread) {
2389                 /* Unread data was tossed, zap the connection. */
2390                 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONCLOSE);
2391                 tcp_set_state(sk, TCP_CLOSE);
2392                 tcp_send_active_reset(sk, sk->sk_allocation);
2393         } else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) {
2394                 /* Check zero linger _after_ checking for unread data. */
2395                 sk->sk_prot->disconnect(sk, 0);
2396                 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
2397         } else if (tcp_close_state(sk)) {
2398                 /* We FIN if the application ate all the data before
2399                  * zapping the connection.
2400                  */
2401
2402                 /* RED-PEN. Formally speaking, we have broken TCP state
2403                  * machine. State transitions:
2404                  *
2405                  * TCP_ESTABLISHED -> TCP_FIN_WAIT1
2406                  * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible)
2407                  * TCP_CLOSE_WAIT -> TCP_LAST_ACK
2408                  *
2409                  * are legal only when FIN has been sent (i.e. in window),
2410                  * rather than queued out of window. Purists blame.
2411                  *
2412                  * F.e. "RFC state" is ESTABLISHED,
2413                  * if Linux state is FIN-WAIT-1, but FIN is still not sent.
2414                  *
2415                  * The visible declinations are that sometimes
2416                  * we enter time-wait state, when it is not required really
2417                  * (harmless), do not send active resets, when they are
2418                  * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
2419                  * they look as CLOSING or LAST_ACK for Linux)
2420                  * Probably, I missed some more holelets.
2421                  *                                              --ANK
2422                  * XXX (TFO) - To start off we don't support SYN+ACK+FIN
2423                  * in a single packet! (May consider it later but will
2424                  * probably need API support or TCP_CORK SYN-ACK until
2425                  * data is written and socket is closed.)
2426                  */
2427                 tcp_send_fin(sk);
2428         }
2429
2430         sk_stream_wait_close(sk, timeout);
2431
2432 adjudge_to_death:
2433         state = sk->sk_state;
2434         sock_hold(sk);
2435         sock_orphan(sk);
2436
2437         local_bh_disable();
2438         bh_lock_sock(sk);
2439         /* remove backlog if any, without releasing ownership. */
2440         __release_sock(sk);
2441
2442         percpu_counter_inc(sk->sk_prot->orphan_count);
2443
2444         /* Have we already been destroyed by a softirq or backlog? */
2445         if (state != TCP_CLOSE && sk->sk_state == TCP_CLOSE)
2446                 goto out;
2447
2448         /*      This is a (useful) BSD violating of the RFC. There is a
2449          *      problem with TCP as specified in that the other end could
2450          *      keep a socket open forever with no application left this end.
2451          *      We use a 1 minute timeout (about the same as BSD) then kill
2452          *      our end. If they send after that then tough - BUT: long enough
2453          *      that we won't make the old 4*rto = almost no time - whoops
2454          *      reset mistake.
2455          *
2456          *      Nope, it was not mistake. It is really desired behaviour
2457          *      f.e. on http servers, when such sockets are useless, but
2458          *      consume significant resources. Let's do it with special
2459          *      linger2 option.                                 --ANK
2460          */
2461
2462         if (sk->sk_state == TCP_FIN_WAIT2) {
2463                 struct tcp_sock *tp = tcp_sk(sk);
2464                 if (tp->linger2 < 0) {
2465                         tcp_set_state(sk, TCP_CLOSE);
2466                         tcp_send_active_reset(sk, GFP_ATOMIC);
2467                         __NET_INC_STATS(sock_net(sk),
2468                                         LINUX_MIB_TCPABORTONLINGER);
2469                 } else {
2470                         const int tmo = tcp_fin_time(sk);
2471
2472                         if (tmo > TCP_TIMEWAIT_LEN) {
2473                                 inet_csk_reset_keepalive_timer(sk,
2474                                                 tmo - TCP_TIMEWAIT_LEN);
2475                         } else {
2476                                 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
2477                                 goto out;
2478                         }
2479                 }
2480         }
2481         if (sk->sk_state != TCP_CLOSE) {
2482                 sk_mem_reclaim(sk);
2483                 if (tcp_check_oom(sk, 0)) {
2484                         tcp_set_state(sk, TCP_CLOSE);
2485                         tcp_send_active_reset(sk, GFP_ATOMIC);
2486                         __NET_INC_STATS(sock_net(sk),
2487                                         LINUX_MIB_TCPABORTONMEMORY);
2488                 } else if (!check_net(sock_net(sk))) {
2489                         /* Not possible to send reset; just close */
2490                         tcp_set_state(sk, TCP_CLOSE);
2491                 }
2492         }
2493
2494         if (sk->sk_state == TCP_CLOSE) {
2495                 struct request_sock *req;
2496
2497                 req = rcu_dereference_protected(tcp_sk(sk)->fastopen_rsk,
2498                                                 lockdep_sock_is_held(sk));
2499                 /* We could get here with a non-NULL req if the socket is
2500                  * aborted (e.g., closed with unread data) before 3WHS
2501                  * finishes.
2502                  */
2503                 if (req)
2504                         reqsk_fastopen_remove(sk, req, false);
2505                 inet_csk_destroy_sock(sk);
2506         }
2507         /* Otherwise, socket is reprieved until protocol close. */
2508
2509 out:
2510         bh_unlock_sock(sk);
2511         local_bh_enable();
2512         release_sock(sk);
2513         sock_put(sk);
2514 }
2515 EXPORT_SYMBOL(tcp_close);
2516
2517 /* These states need RST on ABORT according to RFC793 */
2518
2519 static inline bool tcp_need_reset(int state)
2520 {
2521         return (1 << state) &
2522                (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 |
2523                 TCPF_FIN_WAIT2 | TCPF_SYN_RECV);
2524 }
2525
2526 static void tcp_rtx_queue_purge(struct sock *sk)
2527 {
2528         struct rb_node *p = rb_first(&sk->tcp_rtx_queue);
2529
2530         tcp_sk(sk)->highest_sack = NULL;
2531         while (p) {
2532                 struct sk_buff *skb = rb_to_skb(p);
2533
2534                 p = rb_next(p);
2535                 /* Since we are deleting whole queue, no need to
2536                  * list_del(&skb->tcp_tsorted_anchor)
2537                  */
2538                 tcp_rtx_queue_unlink(skb, sk);
2539                 sk_wmem_free_skb(sk, skb);
2540         }
2541 }
2542
2543 void tcp_write_queue_purge(struct sock *sk)
2544 {
2545         struct sk_buff *skb;
2546
2547         tcp_chrono_stop(sk, TCP_CHRONO_BUSY);
2548         while ((skb = __skb_dequeue(&sk->sk_write_queue)) != NULL) {
2549                 tcp_skb_tsorted_anchor_cleanup(skb);
2550                 sk_wmem_free_skb(sk, skb);
2551         }
2552         tcp_rtx_queue_purge(sk);
2553         skb = sk->sk_tx_skb_cache;
2554         if (skb) {
2555                 __kfree_skb(skb);
2556                 sk->sk_tx_skb_cache = NULL;
2557         }
2558         INIT_LIST_HEAD(&tcp_sk(sk)->tsorted_sent_queue);
2559         sk_mem_reclaim(sk);
2560         tcp_clear_all_retrans_hints(tcp_sk(sk));
2561         tcp_sk(sk)->packets_out = 0;
2562         inet_csk(sk)->icsk_backoff = 0;
2563 }
2564
2565 int tcp_disconnect(struct sock *sk, int flags)
2566 {
2567         struct inet_sock *inet = inet_sk(sk);
2568         struct inet_connection_sock *icsk = inet_csk(sk);
2569         struct tcp_sock *tp = tcp_sk(sk);
2570         int old_state = sk->sk_state;
2571         u32 seq;
2572
2573         if (old_state != TCP_CLOSE)
2574                 tcp_set_state(sk, TCP_CLOSE);
2575
2576         /* ABORT function of RFC793 */
2577         if (old_state == TCP_LISTEN) {
2578                 inet_csk_listen_stop(sk);
2579         } else if (unlikely(tp->repair)) {
2580                 sk->sk_err = ECONNABORTED;
2581         } else if (tcp_need_reset(old_state) ||
2582                    (tp->snd_nxt != tp->write_seq &&
2583                     (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK))) {
2584                 /* The last check adjusts for discrepancy of Linux wrt. RFC
2585                  * states
2586                  */
2587                 tcp_send_active_reset(sk, gfp_any());
2588                 sk->sk_err = ECONNRESET;
2589         } else if (old_state == TCP_SYN_SENT)
2590                 sk->sk_err = ECONNRESET;
2591
2592         tcp_clear_xmit_timers(sk);
2593         __skb_queue_purge(&sk->sk_receive_queue);
2594         if (sk->sk_rx_skb_cache) {
2595                 __kfree_skb(sk->sk_rx_skb_cache);
2596                 sk->sk_rx_skb_cache = NULL;
2597         }
2598         WRITE_ONCE(tp->copied_seq, tp->rcv_nxt);
2599         tp->urg_data = 0;
2600         tcp_write_queue_purge(sk);
2601         tcp_fastopen_active_disable_ofo_check(sk);
2602         skb_rbtree_purge(&tp->out_of_order_queue);
2603
2604         inet->inet_dport = 0;
2605
2606         if (!(sk->sk_userlocks & SOCK_BINDADDR_LOCK))
2607                 inet_reset_saddr(sk);
2608
2609         sk->sk_shutdown = 0;
2610         sock_reset_flag(sk, SOCK_DONE);
2611         tp->srtt_us = 0;
2612         tp->mdev_us = jiffies_to_usecs(TCP_TIMEOUT_INIT);
2613         tp->rcv_rtt_last_tsecr = 0;
2614
2615         seq = tp->write_seq + tp->max_window + 2;
2616         if (!seq)
2617                 seq = 1;
2618         WRITE_ONCE(tp->write_seq, seq);
2619
2620         icsk->icsk_backoff = 0;
2621         icsk->icsk_probes_out = 0;
2622         icsk->icsk_rto = TCP_TIMEOUT_INIT;
2623         tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
2624         tp->snd_cwnd = TCP_INIT_CWND;
2625         tp->snd_cwnd_cnt = 0;
2626         tp->window_clamp = 0;
2627         tp->delivered = 0;
2628         tp->delivered_ce = 0;
2629         tcp_set_ca_state(sk, TCP_CA_Open);
2630         tp->is_sack_reneg = 0;
2631         tcp_clear_retrans(tp);
2632         tp->total_retrans = 0;
2633         inet_csk_delack_init(sk);
2634         /* Initialize rcv_mss to TCP_MIN_MSS to avoid division by 0
2635          * issue in __tcp_select_window()
2636          */
2637         icsk->icsk_ack.rcv_mss = TCP_MIN_MSS;
2638         memset(&tp->rx_opt, 0, sizeof(tp->rx_opt));
2639         __sk_dst_reset(sk);
2640         dst_release(sk->sk_rx_dst);
2641         sk->sk_rx_dst = NULL;
2642         tcp_saved_syn_free(tp);
2643         tp->compressed_ack = 0;
2644         tp->segs_in = 0;
2645         tp->segs_out = 0;
2646         tp->bytes_sent = 0;
2647         tp->bytes_acked = 0;
2648         tp->bytes_received = 0;
2649         tp->bytes_retrans = 0;
2650         tp->data_segs_in = 0;
2651         tp->data_segs_out = 0;
2652         tp->duplicate_sack[0].start_seq = 0;
2653         tp->duplicate_sack[0].end_seq = 0;
2654         tp->dsack_dups = 0;
2655         tp->reord_seen = 0;
2656         tp->retrans_out = 0;
2657         tp->sacked_out = 0;
2658         tp->tlp_high_seq = 0;
2659         tp->last_oow_ack_time = 0;
2660         /* There's a bubble in the pipe until at least the first ACK. */
2661         tp->app_limited = ~0U;
2662         tp->rack.mstamp = 0;
2663         tp->rack.advanced = 0;
2664         tp->rack.reo_wnd_steps = 1;
2665         tp->rack.last_delivered = 0;
2666         tp->rack.reo_wnd_persist = 0;
2667         tp->rack.dsack_seen = 0;
2668         tp->syn_data_acked = 0;
2669         tp->rx_opt.saw_tstamp = 0;
2670         tp->rx_opt.dsack = 0;
2671         tp->rx_opt.num_sacks = 0;
2672         tp->rcv_ooopack = 0;
2673
2674
2675         /* Clean up fastopen related fields */
2676         tcp_free_fastopen_req(tp);
2677         inet->defer_connect = 0;
2678         tp->fastopen_client_fail = 0;
2679
2680         WARN_ON(inet->inet_num && !icsk->icsk_bind_hash);
2681
2682         if (sk->sk_frag.page) {
2683                 put_page(sk->sk_frag.page);
2684                 sk->sk_frag.page = NULL;
2685                 sk->sk_frag.offset = 0;
2686         }
2687
2688         sk->sk_error_report(sk);
2689         return 0;
2690 }
2691 EXPORT_SYMBOL(tcp_disconnect);
2692
2693 static inline bool tcp_can_repair_sock(const struct sock *sk)
2694 {
2695         return ns_capable(sock_net(sk)->user_ns, CAP_NET_ADMIN) &&
2696                 (sk->sk_state != TCP_LISTEN);
2697 }
2698
2699 static int tcp_repair_set_window(struct tcp_sock *tp, char __user *optbuf, int len)
2700 {
2701         struct tcp_repair_window opt;
2702
2703         if (!tp->repair)
2704                 return -EPERM;
2705
2706         if (len != sizeof(opt))
2707                 return -EINVAL;
2708
2709         if (copy_from_user(&opt, optbuf, sizeof(opt)))
2710                 return -EFAULT;
2711
2712         if (opt.max_window < opt.snd_wnd)
2713                 return -EINVAL;
2714
2715         if (after(opt.snd_wl1, tp->rcv_nxt + opt.rcv_wnd))
2716                 return -EINVAL;
2717
2718         if (after(opt.rcv_wup, tp->rcv_nxt))
2719                 return -EINVAL;
2720
2721         tp->snd_wl1     = opt.snd_wl1;
2722         tp->snd_wnd     = opt.snd_wnd;
2723         tp->max_window  = opt.max_window;
2724
2725         tp->rcv_wnd     = opt.rcv_wnd;
2726         tp->rcv_wup     = opt.rcv_wup;
2727
2728         return 0;
2729 }
2730
2731 static int tcp_repair_options_est(struct sock *sk,
2732                 struct tcp_repair_opt __user *optbuf, unsigned int len)
2733 {
2734         struct tcp_sock *tp = tcp_sk(sk);
2735         struct tcp_repair_opt opt;
2736
2737         while (len >= sizeof(opt)) {
2738                 if (copy_from_user(&opt, optbuf, sizeof(opt)))
2739                         return -EFAULT;
2740
2741                 optbuf++;
2742                 len -= sizeof(opt);
2743
2744                 switch (opt.opt_code) {
2745                 case TCPOPT_MSS:
2746                         tp->rx_opt.mss_clamp = opt.opt_val;
2747                         tcp_mtup_init(sk);
2748                         break;
2749                 case TCPOPT_WINDOW:
2750                         {
2751                                 u16 snd_wscale = opt.opt_val & 0xFFFF;
2752                                 u16 rcv_wscale = opt.opt_val >> 16;
2753
2754                                 if (snd_wscale > TCP_MAX_WSCALE || rcv_wscale > TCP_MAX_WSCALE)
2755                                         return -EFBIG;
2756
2757                                 tp->rx_opt.snd_wscale = snd_wscale;
2758                                 tp->rx_opt.rcv_wscale = rcv_wscale;
2759                                 tp->rx_opt.wscale_ok = 1;
2760                         }
2761                         break;
2762                 case TCPOPT_SACK_PERM:
2763                         if (opt.opt_val != 0)
2764                                 return -EINVAL;
2765
2766                         tp->rx_opt.sack_ok |= TCP_SACK_SEEN;
2767                         break;
2768                 case TCPOPT_TIMESTAMP:
2769                         if (opt.opt_val != 0)
2770                                 return -EINVAL;
2771
2772                         tp->rx_opt.tstamp_ok = 1;
2773                         break;
2774                 }
2775         }
2776
2777         return 0;
2778 }
2779
2780 DEFINE_STATIC_KEY_FALSE(tcp_tx_delay_enabled);
2781 EXPORT_SYMBOL(tcp_tx_delay_enabled);
2782
2783 static void tcp_enable_tx_delay(void)
2784 {
2785         if (!static_branch_unlikely(&tcp_tx_delay_enabled)) {
2786                 static int __tcp_tx_delay_enabled = 0;
2787
2788                 if (cmpxchg(&__tcp_tx_delay_enabled, 0, 1) == 0) {
2789                         static_branch_enable(&tcp_tx_delay_enabled);
2790                         pr_info("TCP_TX_DELAY enabled\n");
2791                 }
2792         }
2793 }
2794
2795 /*
2796  *      Socket option code for TCP.
2797  */
2798 static int do_tcp_setsockopt(struct sock *sk, int level,
2799                 int optname, char __user *optval, unsigned int optlen)
2800 {
2801         struct tcp_sock *tp = tcp_sk(sk);
2802         struct inet_connection_sock *icsk = inet_csk(sk);
2803         struct net *net = sock_net(sk);
2804         int val;
2805         int err = 0;
2806
2807         /* These are data/string values, all the others are ints */
2808         switch (optname) {
2809         case TCP_CONGESTION: {
2810                 char name[TCP_CA_NAME_MAX];
2811
2812                 if (optlen < 1)
2813                         return -EINVAL;
2814
2815                 val = strncpy_from_user(name, optval,
2816                                         min_t(long, TCP_CA_NAME_MAX-1, optlen));
2817                 if (val < 0)
2818                         return -EFAULT;
2819                 name[val] = 0;
2820
2821                 lock_sock(sk);
2822                 err = tcp_set_congestion_control(sk, name, true, true,
2823                                                  ns_capable(sock_net(sk)->user_ns,
2824                                                             CAP_NET_ADMIN));
2825                 release_sock(sk);
2826                 return err;
2827         }
2828         case TCP_ULP: {
2829                 char name[TCP_ULP_NAME_MAX];
2830
2831                 if (optlen < 1)
2832                         return -EINVAL;
2833
2834                 val = strncpy_from_user(name, optval,
2835                                         min_t(long, TCP_ULP_NAME_MAX - 1,
2836                                               optlen));
2837                 if (val < 0)
2838                         return -EFAULT;
2839                 name[val] = 0;
2840
2841                 lock_sock(sk);
2842                 err = tcp_set_ulp(sk, name);
2843                 release_sock(sk);
2844                 return err;
2845         }
2846         case TCP_FASTOPEN_KEY: {
2847                 __u8 key[TCP_FASTOPEN_KEY_BUF_LENGTH];
2848                 __u8 *backup_key = NULL;
2849
2850                 /* Allow a backup key as well to facilitate key rotation
2851                  * First key is the active one.
2852                  */
2853                 if (optlen != TCP_FASTOPEN_KEY_LENGTH &&
2854                     optlen != TCP_FASTOPEN_KEY_BUF_LENGTH)
2855                         return -EINVAL;
2856
2857                 if (copy_from_user(key, optval, optlen))
2858                         return -EFAULT;
2859
2860                 if (optlen == TCP_FASTOPEN_KEY_BUF_LENGTH)
2861                         backup_key = key + TCP_FASTOPEN_KEY_LENGTH;
2862
2863                 return tcp_fastopen_reset_cipher(net, sk, key, backup_key);
2864         }
2865         default:
2866                 /* fallthru */
2867                 break;
2868         }
2869
2870         if (optlen < sizeof(int))
2871                 return -EINVAL;
2872
2873         if (get_user(val, (int __user *)optval))
2874                 return -EFAULT;
2875
2876         lock_sock(sk);
2877
2878         switch (optname) {
2879         case TCP_MAXSEG:
2880                 /* Values greater than interface MTU won't take effect. However
2881                  * at the point when this call is done we typically don't yet
2882                  * know which interface is going to be used
2883                  */
2884                 if (val && (val < TCP_MIN_MSS || val > MAX_TCP_WINDOW)) {
2885                         err = -EINVAL;
2886                         break;
2887                 }
2888                 tp->rx_opt.user_mss = val;
2889                 break;
2890
2891         case TCP_NODELAY:
2892                 if (val) {
2893                         /* TCP_NODELAY is weaker than TCP_CORK, so that
2894                          * this option on corked socket is remembered, but
2895                          * it is not activated until cork is cleared.
2896                          *
2897                          * However, when TCP_NODELAY is set we make
2898                          * an explicit push, which overrides even TCP_CORK
2899                          * for currently queued segments.
2900                          */
2901                         tp->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH;
2902                         tcp_push_pending_frames(sk);
2903                 } else {
2904                         tp->nonagle &= ~TCP_NAGLE_OFF;
2905                 }
2906                 break;
2907
2908         case TCP_THIN_LINEAR_TIMEOUTS:
2909                 if (val < 0 || val > 1)
2910                         err = -EINVAL;
2911                 else
2912                         tp->thin_lto = val;
2913                 break;
2914
2915         case TCP_THIN_DUPACK:
2916                 if (val < 0 || val > 1)
2917                         err = -EINVAL;
2918                 break;
2919
2920         case TCP_REPAIR:
2921                 if (!tcp_can_repair_sock(sk))
2922                         err = -EPERM;
2923                 else if (val == TCP_REPAIR_ON) {
2924                         tp->repair = 1;
2925                         sk->sk_reuse = SK_FORCE_REUSE;
2926                         tp->repair_queue = TCP_NO_QUEUE;
2927                 } else if (val == TCP_REPAIR_OFF) {
2928                         tp->repair = 0;
2929                         sk->sk_reuse = SK_NO_REUSE;
2930                         tcp_send_window_probe(sk);
2931                 } else if (val == TCP_REPAIR_OFF_NO_WP) {
2932                         tp->repair = 0;
2933                         sk->sk_reuse = SK_NO_REUSE;
2934                 } else
2935                         err = -EINVAL;
2936
2937                 break;
2938
2939         case TCP_REPAIR_QUEUE:
2940                 if (!tp->repair)
2941                         err = -EPERM;
2942                 else if ((unsigned int)val < TCP_QUEUES_NR)
2943                         tp->repair_queue = val;
2944                 else
2945                         err = -EINVAL;
2946                 break;
2947
2948         case TCP_QUEUE_SEQ:
2949                 if (sk->sk_state != TCP_CLOSE)
2950                         err = -EPERM;
2951                 else if (tp->repair_queue == TCP_SEND_QUEUE)
2952                         WRITE_ONCE(tp->write_seq, val);
2953                 else if (tp->repair_queue == TCP_RECV_QUEUE) {
2954                         WRITE_ONCE(tp->rcv_nxt, val);
2955                         WRITE_ONCE(tp->copied_seq, val);
2956                 }
2957                 else
2958                         err = -EINVAL;
2959                 break;
2960
2961         case TCP_REPAIR_OPTIONS:
2962                 if (!tp->repair)
2963                         err = -EINVAL;
2964                 else if (sk->sk_state == TCP_ESTABLISHED)
2965                         err = tcp_repair_options_est(sk,
2966                                         (struct tcp_repair_opt __user *)optval,
2967                                         optlen);
2968                 else
2969                         err = -EPERM;
2970                 break;
2971
2972         case TCP_CORK:
2973                 /* When set indicates to always queue non-full frames.
2974                  * Later the user clears this option and we transmit
2975                  * any pending partial frames in the queue.  This is
2976                  * meant to be used alongside sendfile() to get properly
2977                  * filled frames when the user (for example) must write
2978                  * out headers with a write() call first and then use
2979                  * sendfile to send out the data parts.
2980                  *
2981                  * TCP_CORK can be set together with TCP_NODELAY and it is
2982                  * stronger than TCP_NODELAY.
2983                  */
2984                 if (val) {
2985                         tp->nonagle |= TCP_NAGLE_CORK;
2986                 } else {
2987                         tp->nonagle &= ~TCP_NAGLE_CORK;
2988                         if (tp->nonagle&TCP_NAGLE_OFF)
2989                                 tp->nonagle |= TCP_NAGLE_PUSH;
2990                         tcp_push_pending_frames(sk);
2991                 }
2992                 break;
2993
2994         case TCP_KEEPIDLE:
2995                 if (val < 1 || val > MAX_TCP_KEEPIDLE)
2996                         err = -EINVAL;
2997                 else {
2998                         tp->keepalive_time = val * HZ;
2999                         if (sock_flag(sk, SOCK_KEEPOPEN) &&
3000                             !((1 << sk->sk_state) &
3001                               (TCPF_CLOSE | TCPF_LISTEN))) {
3002                                 u32 elapsed = keepalive_time_elapsed(tp);
3003                                 if (tp->keepalive_time > elapsed)
3004                                         elapsed = tp->keepalive_time - elapsed;
3005                                 else
3006                                         elapsed = 0;
3007                                 inet_csk_reset_keepalive_timer(sk, elapsed);
3008                         }
3009                 }
3010                 break;
3011         case TCP_KEEPINTVL:
3012                 if (val < 1 || val > MAX_TCP_KEEPINTVL)
3013                         err = -EINVAL;
3014                 else
3015                         tp->keepalive_intvl = val * HZ;
3016                 break;
3017         case TCP_KEEPCNT:
3018                 if (val < 1 || val > MAX_TCP_KEEPCNT)
3019                         err = -EINVAL;
3020                 else
3021                         tp->keepalive_probes = val;
3022                 break;
3023         case TCP_SYNCNT:
3024                 if (val < 1 || val > MAX_TCP_SYNCNT)
3025                         err = -EINVAL;
3026                 else
3027                         icsk->icsk_syn_retries = val;
3028                 break;
3029
3030         case TCP_SAVE_SYN:
3031                 if (val < 0 || val > 1)
3032                         err = -EINVAL;
3033                 else
3034                         tp->save_syn = val;
3035                 break;
3036
3037         case TCP_LINGER2:
3038                 if (val < 0)
3039                         tp->linger2 = -1;
3040                 else if (val > net->ipv4.sysctl_tcp_fin_timeout / HZ)
3041                         tp->linger2 = 0;
3042                 else
3043                         tp->linger2 = val * HZ;
3044                 break;
3045
3046         case TCP_DEFER_ACCEPT:
3047                 /* Translate value in seconds to number of retransmits */
3048                 icsk->icsk_accept_queue.rskq_defer_accept =
3049                         secs_to_retrans(val, TCP_TIMEOUT_INIT / HZ,
3050                                         TCP_RTO_MAX / HZ);
3051                 break;
3052
3053         case TCP_WINDOW_CLAMP:
3054                 if (!val) {
3055                         if (sk->sk_state != TCP_CLOSE) {
3056                                 err = -EINVAL;
3057                                 break;
3058                         }
3059                         tp->window_clamp = 0;
3060                 } else
3061                         tp->window_clamp = val < SOCK_MIN_RCVBUF / 2 ?
3062                                                 SOCK_MIN_RCVBUF / 2 : val;
3063                 break;
3064
3065         case TCP_QUICKACK:
3066                 if (!val) {
3067                         inet_csk_enter_pingpong_mode(sk);
3068                 } else {
3069                         inet_csk_exit_pingpong_mode(sk);
3070                         if ((1 << sk->sk_state) &
3071                             (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) &&
3072                             inet_csk_ack_scheduled(sk)) {
3073                                 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
3074                                 tcp_cleanup_rbuf(sk, 1);
3075                                 if (!(val & 1))
3076                                         inet_csk_enter_pingpong_mode(sk);
3077                         }
3078                 }
3079                 break;
3080
3081 #ifdef CONFIG_TCP_MD5SIG
3082         case TCP_MD5SIG:
3083         case TCP_MD5SIG_EXT:
3084                 if ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN))
3085                         err = tp->af_specific->md5_parse(sk, optname, optval, optlen);
3086                 else
3087                         err = -EINVAL;
3088                 break;
3089 #endif
3090         case TCP_USER_TIMEOUT:
3091                 /* Cap the max time in ms TCP will retry or probe the window
3092                  * before giving up and aborting (ETIMEDOUT) a connection.
3093                  */
3094                 if (val < 0)
3095                         err = -EINVAL;
3096                 else
3097                         icsk->icsk_user_timeout = val;
3098                 break;
3099
3100         case TCP_FASTOPEN:
3101                 if (val >= 0 && ((1 << sk->sk_state) & (TCPF_CLOSE |
3102                     TCPF_LISTEN))) {
3103                         tcp_fastopen_init_key_once(net);
3104
3105                         fastopen_queue_tune(sk, val);
3106                 } else {
3107                         err = -EINVAL;
3108                 }
3109                 break;
3110         case TCP_FASTOPEN_CONNECT:
3111                 if (val > 1 || val < 0) {
3112                         err = -EINVAL;
3113                 } else if (net->ipv4.sysctl_tcp_fastopen & TFO_CLIENT_ENABLE) {
3114                         if (sk->sk_state == TCP_CLOSE)
3115                                 tp->fastopen_connect = val;
3116                         else
3117                                 err = -EINVAL;
3118                 } else {
3119                         err = -EOPNOTSUPP;
3120                 }
3121                 break;
3122         case TCP_FASTOPEN_NO_COOKIE:
3123                 if (val > 1 || val < 0)
3124                         err = -EINVAL;
3125                 else if (!((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
3126                         err = -EINVAL;
3127                 else
3128                         tp->fastopen_no_cookie = val;
3129                 break;
3130         case TCP_TIMESTAMP:
3131                 if (!tp->repair)
3132                         err = -EPERM;
3133                 else
3134                         tp->tsoffset = val - tcp_time_stamp_raw();
3135                 break;
3136         case TCP_REPAIR_WINDOW:
3137                 err = tcp_repair_set_window(tp, optval, optlen);
3138                 break;
3139         case TCP_NOTSENT_LOWAT:
3140                 tp->notsent_lowat = val;
3141                 sk->sk_write_space(sk);
3142                 break;
3143         case TCP_INQ:
3144                 if (val > 1 || val < 0)
3145                         err = -EINVAL;
3146                 else
3147                         tp->recvmsg_inq = val;
3148                 break;
3149         case TCP_TX_DELAY:
3150                 if (val)
3151                         tcp_enable_tx_delay();
3152                 tp->tcp_tx_delay = val;
3153                 break;
3154         default:
3155                 err = -ENOPROTOOPT;
3156                 break;
3157         }
3158
3159         release_sock(sk);
3160         return err;
3161 }
3162
3163 int tcp_setsockopt(struct sock *sk, int level, int optname, char __user *optval,
3164                    unsigned int optlen)
3165 {
3166         const struct inet_connection_sock *icsk = inet_csk(sk);
3167
3168         if (level != SOL_TCP)
3169                 return icsk->icsk_af_ops->setsockopt(sk, level, optname,
3170                                                      optval, optlen);
3171         return do_tcp_setsockopt(sk, level, optname, optval, optlen);
3172 }
3173 EXPORT_SYMBOL(tcp_setsockopt);
3174
3175 #ifdef CONFIG_COMPAT
3176 int compat_tcp_setsockopt(struct sock *sk, int level, int optname,
3177                           char __user *optval, unsigned int optlen)
3178 {
3179         if (level != SOL_TCP)
3180                 return inet_csk_compat_setsockopt(sk, level, optname,
3181                                                   optval, optlen);
3182         return do_tcp_setsockopt(sk, level, optname, optval, optlen);
3183 }
3184 EXPORT_SYMBOL(compat_tcp_setsockopt);
3185 #endif
3186
3187 static void tcp_get_info_chrono_stats(const struct tcp_sock *tp,
3188                                       struct tcp_info *info)
3189 {
3190         u64 stats[__TCP_CHRONO_MAX], total = 0;
3191         enum tcp_chrono i;
3192
3193         for (i = TCP_CHRONO_BUSY; i < __TCP_CHRONO_MAX; ++i) {
3194                 stats[i] = tp->chrono_stat[i - 1];
3195                 if (i == tp->chrono_type)
3196                         stats[i] += tcp_jiffies32 - tp->chrono_start;
3197                 stats[i] *= USEC_PER_SEC / HZ;
3198                 total += stats[i];
3199         }
3200
3201         info->tcpi_busy_time = total;
3202         info->tcpi_rwnd_limited = stats[TCP_CHRONO_RWND_LIMITED];
3203         info->tcpi_sndbuf_limited = stats[TCP_CHRONO_SNDBUF_LIMITED];
3204 }
3205
3206 /* Return information about state of tcp endpoint in API format. */
3207 void tcp_get_info(struct sock *sk, struct tcp_info *info)
3208 {
3209         const struct tcp_sock *tp = tcp_sk(sk); /* iff sk_type == SOCK_STREAM */
3210         const struct inet_connection_sock *icsk = inet_csk(sk);
3211         unsigned long rate;
3212         u32 now;
3213         u64 rate64;
3214         bool slow;
3215
3216         memset(info, 0, sizeof(*info));
3217         if (sk->sk_type != SOCK_STREAM)
3218                 return;
3219
3220         info->tcpi_state = inet_sk_state_load(sk);
3221
3222         /* Report meaningful fields for all TCP states, including listeners */
3223         rate = READ_ONCE(sk->sk_pacing_rate);
3224         rate64 = (rate != ~0UL) ? rate : ~0ULL;
3225         info->tcpi_pacing_rate = rate64;
3226
3227         rate = READ_ONCE(sk->sk_max_pacing_rate);
3228         rate64 = (rate != ~0UL) ? rate : ~0ULL;
3229         info->tcpi_max_pacing_rate = rate64;
3230
3231         info->tcpi_reordering = tp->reordering;
3232         info->tcpi_snd_cwnd = tp->snd_cwnd;
3233
3234         if (info->tcpi_state == TCP_LISTEN) {
3235                 /* listeners aliased fields :
3236                  * tcpi_unacked -> Number of children ready for accept()
3237                  * tcpi_sacked  -> max backlog
3238                  */
3239                 info->tcpi_unacked = READ_ONCE(sk->sk_ack_backlog);
3240                 info->tcpi_sacked = READ_ONCE(sk->sk_max_ack_backlog);
3241                 return;
3242         }
3243
3244         slow = lock_sock_fast(sk);
3245
3246         info->tcpi_ca_state = icsk->icsk_ca_state;
3247         info->tcpi_retransmits = icsk->icsk_retransmits;
3248         info->tcpi_probes = icsk->icsk_probes_out;
3249         info->tcpi_backoff = icsk->icsk_backoff;
3250
3251         if (tp->rx_opt.tstamp_ok)
3252                 info->tcpi_options |= TCPI_OPT_TIMESTAMPS;
3253         if (tcp_is_sack(tp))
3254                 info->tcpi_options |= TCPI_OPT_SACK;
3255         if (tp->rx_opt.wscale_ok) {
3256                 info->tcpi_options |= TCPI_OPT_WSCALE;
3257                 info->tcpi_snd_wscale = tp->rx_opt.snd_wscale;
3258                 info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale;
3259         }
3260
3261         if (tp->ecn_flags & TCP_ECN_OK)
3262                 info->tcpi_options |= TCPI_OPT_ECN;
3263         if (tp->ecn_flags & TCP_ECN_SEEN)
3264                 info->tcpi_options |= TCPI_OPT_ECN_SEEN;
3265         if (tp->syn_data_acked)
3266                 info->tcpi_options |= TCPI_OPT_SYN_DATA;
3267
3268         info->tcpi_rto = jiffies_to_usecs(icsk->icsk_rto);
3269         info->tcpi_ato = jiffies_to_usecs(icsk->icsk_ack.ato);
3270         info->tcpi_snd_mss = tp->mss_cache;
3271         info->tcpi_rcv_mss = icsk->icsk_ack.rcv_mss;
3272
3273         info->tcpi_unacked = tp->packets_out;
3274         info->tcpi_sacked = tp->sacked_out;
3275
3276         info->tcpi_lost = tp->lost_out;
3277         info->tcpi_retrans = tp->retrans_out;
3278
3279         now = tcp_jiffies32;
3280         info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime);
3281         info->tcpi_last_data_recv = jiffies_to_msecs(now - icsk->icsk_ack.lrcvtime);
3282         info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp);
3283
3284         info->tcpi_pmtu = icsk->icsk_pmtu_cookie;
3285         info->tcpi_rcv_ssthresh = tp->rcv_ssthresh;
3286         info->tcpi_rtt = tp->srtt_us >> 3;
3287         info->tcpi_rttvar = tp->mdev_us >> 2;
3288         info->tcpi_snd_ssthresh = tp->snd_ssthresh;
3289         info->tcpi_advmss = tp->advmss;
3290
3291         info->tcpi_rcv_rtt = tp->rcv_rtt_est.rtt_us >> 3;
3292         info->tcpi_rcv_space = tp->rcvq_space.space;
3293
3294         info->tcpi_total_retrans = tp->total_retrans;
3295
3296         info->tcpi_bytes_acked = tp->bytes_acked;
3297         info->tcpi_bytes_received = tp->bytes_received;
3298         info->tcpi_notsent_bytes = max_t(int, 0, tp->write_seq - tp->snd_nxt);
3299         tcp_get_info_chrono_stats(tp, info);
3300
3301         info->tcpi_segs_out = tp->segs_out;
3302         info->tcpi_segs_in = tp->segs_in;
3303
3304         info->tcpi_min_rtt = tcp_min_rtt(tp);
3305         info->tcpi_data_segs_in = tp->data_segs_in;
3306         info->tcpi_data_segs_out = tp->data_segs_out;
3307
3308         info->tcpi_delivery_rate_app_limited = tp->rate_app_limited ? 1 : 0;
3309         rate64 = tcp_compute_delivery_rate(tp);
3310         if (rate64)
3311                 info->tcpi_delivery_rate = rate64;
3312         info->tcpi_delivered = tp->delivered;
3313         info->tcpi_delivered_ce = tp->delivered_ce;
3314         info->tcpi_bytes_sent = tp->bytes_sent;
3315         info->tcpi_bytes_retrans = tp->bytes_retrans;
3316         info->tcpi_dsack_dups = tp->dsack_dups;
3317         info->tcpi_reord_seen = tp->reord_seen;
3318         info->tcpi_rcv_ooopack = tp->rcv_ooopack;
3319         info->tcpi_snd_wnd = tp->snd_wnd;
3320         info->tcpi_fastopen_client_fail = tp->fastopen_client_fail;
3321         unlock_sock_fast(sk, slow);
3322 }
3323 EXPORT_SYMBOL_GPL(tcp_get_info);
3324
3325 static size_t tcp_opt_stats_get_size(void)
3326 {
3327         return
3328                 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_BUSY */
3329                 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_RWND_LIMITED */
3330                 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_SNDBUF_LIMITED */
3331                 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_DATA_SEGS_OUT */
3332                 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_TOTAL_RETRANS */
3333                 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_PACING_RATE */
3334                 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_DELIVERY_RATE */
3335                 nla_total_size(sizeof(u32)) + /* TCP_NLA_SND_CWND */
3336                 nla_total_size(sizeof(u32)) + /* TCP_NLA_REORDERING */
3337                 nla_total_size(sizeof(u32)) + /* TCP_NLA_MIN_RTT */
3338                 nla_total_size(sizeof(u8)) + /* TCP_NLA_RECUR_RETRANS */
3339                 nla_total_size(sizeof(u8)) + /* TCP_NLA_DELIVERY_RATE_APP_LMT */
3340                 nla_total_size(sizeof(u32)) + /* TCP_NLA_SNDQ_SIZE */
3341                 nla_total_size(sizeof(u8)) + /* TCP_NLA_CA_STATE */
3342                 nla_total_size(sizeof(u32)) + /* TCP_NLA_SND_SSTHRESH */
3343                 nla_total_size(sizeof(u32)) + /* TCP_NLA_DELIVERED */
3344                 nla_total_size(sizeof(u32)) + /* TCP_NLA_DELIVERED_CE */
3345                 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_BYTES_SENT */
3346                 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_BYTES_RETRANS */
3347                 nla_total_size(sizeof(u32)) + /* TCP_NLA_DSACK_DUPS */
3348                 nla_total_size(sizeof(u32)) + /* TCP_NLA_REORD_SEEN */
3349                 nla_total_size(sizeof(u32)) + /* TCP_NLA_SRTT */
3350                 nla_total_size(sizeof(u16)) + /* TCP_NLA_TIMEOUT_REHASH */
3351                 nla_total_size(sizeof(u32)) + /* TCP_NLA_BYTES_NOTSENT */
3352                 0;
3353 }
3354
3355 struct sk_buff *tcp_get_timestamping_opt_stats(const struct sock *sk)
3356 {
3357         const struct tcp_sock *tp = tcp_sk(sk);
3358         struct sk_buff *stats;
3359         struct tcp_info info;
3360         unsigned long rate;
3361         u64 rate64;
3362
3363         stats = alloc_skb(tcp_opt_stats_get_size(), GFP_ATOMIC);
3364         if (!stats)
3365                 return NULL;
3366
3367         tcp_get_info_chrono_stats(tp, &info);
3368         nla_put_u64_64bit(stats, TCP_NLA_BUSY,
3369                           info.tcpi_busy_time, TCP_NLA_PAD);
3370         nla_put_u64_64bit(stats, TCP_NLA_RWND_LIMITED,
3371                           info.tcpi_rwnd_limited, TCP_NLA_PAD);
3372         nla_put_u64_64bit(stats, TCP_NLA_SNDBUF_LIMITED,
3373                           info.tcpi_sndbuf_limited, TCP_NLA_PAD);
3374         nla_put_u64_64bit(stats, TCP_NLA_DATA_SEGS_OUT,
3375                           tp->data_segs_out, TCP_NLA_PAD);
3376         nla_put_u64_64bit(stats, TCP_NLA_TOTAL_RETRANS,
3377                           tp->total_retrans, TCP_NLA_PAD);
3378
3379         rate = READ_ONCE(sk->sk_pacing_rate);
3380         rate64 = (rate != ~0UL) ? rate : ~0ULL;
3381         nla_put_u64_64bit(stats, TCP_NLA_PACING_RATE, rate64, TCP_NLA_PAD);
3382
3383         rate64 = tcp_compute_delivery_rate(tp);
3384         nla_put_u64_64bit(stats, TCP_NLA_DELIVERY_RATE, rate64, TCP_NLA_PAD);
3385
3386         nla_put_u32(stats, TCP_NLA_SND_CWND, tp->snd_cwnd);
3387         nla_put_u32(stats, TCP_NLA_REORDERING, tp->reordering);
3388         nla_put_u32(stats, TCP_NLA_MIN_RTT, tcp_min_rtt(tp));
3389
3390         nla_put_u8(stats, TCP_NLA_RECUR_RETRANS, inet_csk(sk)->icsk_retransmits);
3391         nla_put_u8(stats, TCP_NLA_DELIVERY_RATE_APP_LMT, !!tp->rate_app_limited);
3392         nla_put_u32(stats, TCP_NLA_SND_SSTHRESH, tp->snd_ssthresh);
3393         nla_put_u32(stats, TCP_NLA_DELIVERED, tp->delivered);
3394         nla_put_u32(stats, TCP_NLA_DELIVERED_CE, tp->delivered_ce);
3395
3396         nla_put_u32(stats, TCP_NLA_SNDQ_SIZE, tp->write_seq - tp->snd_una);
3397         nla_put_u8(stats, TCP_NLA_CA_STATE, inet_csk(sk)->icsk_ca_state);
3398
3399         nla_put_u64_64bit(stats, TCP_NLA_BYTES_SENT, tp->bytes_sent,
3400                           TCP_NLA_PAD);
3401         nla_put_u64_64bit(stats, TCP_NLA_BYTES_RETRANS, tp->bytes_retrans,
3402                           TCP_NLA_PAD);
3403         nla_put_u32(stats, TCP_NLA_DSACK_DUPS, tp->dsack_dups);
3404         nla_put_u32(stats, TCP_NLA_REORD_SEEN, tp->reord_seen);
3405         nla_put_u32(stats, TCP_NLA_SRTT, tp->srtt_us >> 3);
3406         nla_put_u16(stats, TCP_NLA_TIMEOUT_REHASH, tp->timeout_rehash);
3407         nla_put_u32(stats, TCP_NLA_BYTES_NOTSENT,
3408                     max_t(int, 0, tp->write_seq - tp->snd_nxt));
3409
3410         return stats;
3411 }
3412
3413 static int do_tcp_getsockopt(struct sock *sk, int level,
3414                 int optname, char __user *optval, int __user *optlen)
3415 {
3416         struct inet_connection_sock *icsk = inet_csk(sk);
3417         struct tcp_sock *tp = tcp_sk(sk);
3418         struct net *net = sock_net(sk);
3419         int val, len;
3420
3421         if (get_user(len, optlen))
3422                 return -EFAULT;
3423
3424         len = min_t(unsigned int, len, sizeof(int));
3425
3426         if (len < 0)
3427                 return -EINVAL;
3428
3429         switch (optname) {
3430         case TCP_MAXSEG:
3431                 val = tp->mss_cache;
3432                 if (!val && ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
3433                         val = tp->rx_opt.user_mss;
3434                 if (tp->repair)
3435                         val = tp->rx_opt.mss_clamp;
3436                 break;
3437         case TCP_NODELAY:
3438                 val = !!(tp->nonagle&TCP_NAGLE_OFF);
3439                 break;
3440         case TCP_CORK:
3441                 val = !!(tp->nonagle&TCP_NAGLE_CORK);
3442                 break;
3443         case TCP_KEEPIDLE:
3444                 val = keepalive_time_when(tp) / HZ;
3445                 break;
3446         case TCP_KEEPINTVL:
3447                 val = keepalive_intvl_when(tp) / HZ;
3448                 break;
3449         case TCP_KEEPCNT:
3450                 val = keepalive_probes(tp);
3451                 break;
3452         case TCP_SYNCNT:
3453                 val = icsk->icsk_syn_retries ? : net->ipv4.sysctl_tcp_syn_retries;
3454                 break;
3455         case TCP_LINGER2:
3456                 val = tp->linger2;
3457                 if (val >= 0)
3458                         val = (val ? : net->ipv4.sysctl_tcp_fin_timeout) / HZ;
3459                 break;
3460         case TCP_DEFER_ACCEPT:
3461                 val = retrans_to_secs(icsk->icsk_accept_queue.rskq_defer_accept,
3462                                       TCP_TIMEOUT_INIT / HZ, TCP_RTO_MAX / HZ);
3463                 break;
3464         case TCP_WINDOW_CLAMP:
3465                 val = tp->window_clamp;
3466                 break;
3467         case TCP_INFO: {
3468                 struct tcp_info info;
3469
3470                 if (get_user(len, optlen))
3471                         return -EFAULT;
3472
3473                 tcp_get_info(sk, &info);
3474
3475                 len = min_t(unsigned int, len, sizeof(info));
3476                 if (put_user(len, optlen))
3477                         return -EFAULT;
3478                 if (copy_to_user(optval, &info, len))
3479                         return -EFAULT;
3480                 return 0;
3481         }
3482         case TCP_CC_INFO: {
3483                 const struct tcp_congestion_ops *ca_ops;
3484                 union tcp_cc_info info;
3485                 size_t sz = 0;
3486                 int attr;
3487
3488                 if (get_user(len, optlen))
3489                         return -EFAULT;
3490
3491                 ca_ops = icsk->icsk_ca_ops;
3492                 if (ca_ops && ca_ops->get_info)
3493                         sz = ca_ops->get_info(sk, ~0U, &attr, &info);
3494
3495                 len = min_t(unsigned int, len, sz);
3496                 if (put_user(len, optlen))
3497                         return -EFAULT;
3498                 if (copy_to_user(optval, &info, len))
3499                         return -EFAULT;
3500                 return 0;
3501         }
3502         case TCP_QUICKACK:
3503                 val = !inet_csk_in_pingpong_mode(sk);
3504                 break;
3505
3506         case TCP_CONGESTION:
3507                 if (get_user(len, optlen))
3508                         return -EFAULT;
3509                 len = min_t(unsigned int, len, TCP_CA_NAME_MAX);
3510                 if (put_user(len, optlen))
3511                         return -EFAULT;
3512                 if (copy_to_user(optval, icsk->icsk_ca_ops->name, len))
3513                         return -EFAULT;
3514                 return 0;
3515
3516         case TCP_ULP:
3517                 if (get_user(len, optlen))
3518                         return -EFAULT;
3519                 len = min_t(unsigned int, len, TCP_ULP_NAME_MAX);
3520                 if (!icsk->icsk_ulp_ops) {
3521                         if (put_user(0, optlen))
3522                                 return -EFAULT;
3523                         return 0;
3524                 }
3525                 if (put_user(len, optlen))
3526                         return -EFAULT;
3527                 if (copy_to_user(optval, icsk->icsk_ulp_ops->name, len))
3528                         return -EFAULT;
3529                 return 0;
3530
3531         case TCP_FASTOPEN_KEY: {
3532                 __u8 key[TCP_FASTOPEN_KEY_BUF_LENGTH];
3533                 struct tcp_fastopen_context *ctx;
3534                 unsigned int key_len = 0;
3535
3536                 if (get_user(len, optlen))
3537                         return -EFAULT;
3538
3539                 rcu_read_lock();
3540                 ctx = rcu_dereference(icsk->icsk_accept_queue.fastopenq.ctx);
3541                 if (ctx) {
3542                         key_len = tcp_fastopen_context_len(ctx) *
3543                                         TCP_FASTOPEN_KEY_LENGTH;
3544                         memcpy(&key[0], &ctx->key[0], key_len);
3545                 }
3546                 rcu_read_unlock();
3547
3548                 len = min_t(unsigned int, len, key_len);
3549                 if (put_user(len, optlen))
3550                         return -EFAULT;
3551                 if (copy_to_user(optval, key, len))
3552                         return -EFAULT;
3553                 return 0;
3554         }
3555         case TCP_THIN_LINEAR_TIMEOUTS:
3556                 val = tp->thin_lto;
3557                 break;
3558
3559         case TCP_THIN_DUPACK:
3560                 val = 0;
3561                 break;
3562
3563         case TCP_REPAIR:
3564                 val = tp->repair;
3565                 break;
3566
3567         case TCP_REPAIR_QUEUE:
3568                 if (tp->repair)
3569                         val = tp->repair_queue;
3570                 else
3571                         return -EINVAL;
3572                 break;
3573
3574         case TCP_REPAIR_WINDOW: {
3575                 struct tcp_repair_window opt;
3576
3577                 if (get_user(len, optlen))
3578                         return -EFAULT;
3579
3580                 if (len != sizeof(opt))
3581                         return -EINVAL;
3582
3583                 if (!tp->repair)
3584                         return -EPERM;
3585
3586                 opt.snd_wl1     = tp->snd_wl1;
3587                 opt.snd_wnd     = tp->snd_wnd;
3588                 opt.max_window  = tp->max_window;
3589                 opt.rcv_wnd     = tp->rcv_wnd;
3590                 opt.rcv_wup     = tp->rcv_wup;
3591
3592                 if (copy_to_user(optval, &opt, len))
3593                         return -EFAULT;
3594                 return 0;
3595         }
3596         case TCP_QUEUE_SEQ:
3597                 if (tp->repair_queue == TCP_SEND_QUEUE)
3598                         val = tp->write_seq;
3599                 else if (tp->repair_queue == TCP_RECV_QUEUE)
3600                         val = tp->rcv_nxt;
3601                 else
3602                         return -EINVAL;
3603                 break;
3604
3605         case TCP_USER_TIMEOUT:
3606                 val = icsk->icsk_user_timeout;
3607                 break;
3608
3609         case TCP_FASTOPEN:
3610                 val = icsk->icsk_accept_queue.fastopenq.max_qlen;
3611                 break;
3612
3613         case TCP_FASTOPEN_CONNECT:
3614                 val = tp->fastopen_connect;
3615                 break;
3616
3617         case TCP_FASTOPEN_NO_COOKIE:
3618                 val = tp->fastopen_no_cookie;
3619                 break;
3620
3621         case TCP_TX_DELAY:
3622                 val = tp->tcp_tx_delay;
3623                 break;
3624
3625         case TCP_TIMESTAMP:
3626                 val = tcp_time_stamp_raw() + tp->tsoffset;
3627                 break;
3628         case TCP_NOTSENT_LOWAT:
3629                 val = tp->notsent_lowat;
3630                 break;
3631         case TCP_INQ:
3632                 val = tp->recvmsg_inq;
3633                 break;
3634         case TCP_SAVE_SYN:
3635                 val = tp->save_syn;
3636                 break;
3637         case TCP_SAVED_SYN: {
3638                 if (get_user(len, optlen))
3639                         return -EFAULT;
3640
3641                 lock_sock(sk);
3642                 if (tp->saved_syn) {
3643                         if (len < tp->saved_syn[0]) {
3644                                 if (put_user(tp->saved_syn[0], optlen)) {
3645                                         release_sock(sk);
3646                                         return -EFAULT;
3647                                 }
3648                                 release_sock(sk);
3649                                 return -EINVAL;
3650                         }
3651                         len = tp->saved_syn[0];
3652                         if (put_user(len, optlen)) {
3653                                 release_sock(sk);
3654                                 return -EFAULT;
3655                         }
3656                         if (copy_to_user(optval, tp->saved_syn + 1, len)) {
3657                                 release_sock(sk);
3658                                 return -EFAULT;
3659                         }
3660                         tcp_saved_syn_free(tp);
3661                         release_sock(sk);
3662                 } else {
3663                         release_sock(sk);
3664                         len = 0;
3665                         if (put_user(len, optlen))
3666                                 return -EFAULT;
3667                 }
3668                 return 0;
3669         }
3670 #ifdef CONFIG_MMU
3671         case TCP_ZEROCOPY_RECEIVE: {
3672                 struct tcp_zerocopy_receive zc;
3673                 int err;
3674
3675                 if (get_user(len, optlen))
3676                         return -EFAULT;
3677                 if (len < offsetofend(struct tcp_zerocopy_receive, length))
3678                         return -EINVAL;
3679                 if (len > sizeof(zc)) {
3680                         len = sizeof(zc);
3681                         if (put_user(len, optlen))
3682                                 return -EFAULT;
3683                 }
3684                 if (copy_from_user(&zc, optval, len))
3685                         return -EFAULT;
3686                 lock_sock(sk);
3687                 err = tcp_zerocopy_receive(sk, &zc);
3688                 release_sock(sk);
3689                 if (len == sizeof(zc))
3690                         goto zerocopy_rcv_sk_err;
3691                 switch (len) {
3692                 case offsetofend(struct tcp_zerocopy_receive, err):
3693                         goto zerocopy_rcv_sk_err;
3694                 case offsetofend(struct tcp_zerocopy_receive, inq):
3695                         goto zerocopy_rcv_inq;
3696                 case offsetofend(struct tcp_zerocopy_receive, length):
3697                 default:
3698                         goto zerocopy_rcv_out;
3699                 }
3700 zerocopy_rcv_sk_err:
3701                 if (!err)
3702                         zc.err = sock_error(sk);
3703 zerocopy_rcv_inq:
3704                 zc.inq = tcp_inq_hint(sk);
3705 zerocopy_rcv_out:
3706                 if (!err && copy_to_user(optval, &zc, len))
3707                         err = -EFAULT;
3708                 return err;
3709         }
3710 #endif
3711         default:
3712                 return -ENOPROTOOPT;
3713         }
3714
3715         if (put_user(len, optlen))
3716                 return -EFAULT;
3717         if (copy_to_user(optval, &val, len))
3718                 return -EFAULT;
3719         return 0;
3720 }
3721
3722 int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval,
3723                    int __user *optlen)
3724 {
3725         struct inet_connection_sock *icsk = inet_csk(sk);
3726
3727         if (level != SOL_TCP)
3728                 return icsk->icsk_af_ops->getsockopt(sk, level, optname,
3729                                                      optval, optlen);
3730         return do_tcp_getsockopt(sk, level, optname, optval, optlen);
3731 }
3732 EXPORT_SYMBOL(tcp_getsockopt);
3733
3734 #ifdef CONFIG_COMPAT
3735 int compat_tcp_getsockopt(struct sock *sk, int level, int optname,
3736                           char __user *optval, int __user *optlen)
3737 {
3738         if (level != SOL_TCP)
3739                 return inet_csk_compat_getsockopt(sk, level, optname,
3740                                                   optval, optlen);
3741         return do_tcp_getsockopt(sk, level, optname, optval, optlen);
3742 }
3743 EXPORT_SYMBOL(compat_tcp_getsockopt);
3744 #endif
3745
3746 #ifdef CONFIG_TCP_MD5SIG
3747 static DEFINE_PER_CPU(struct tcp_md5sig_pool, tcp_md5sig_pool);
3748 static DEFINE_MUTEX(tcp_md5sig_mutex);
3749 static bool tcp_md5sig_pool_populated = false;
3750
3751 static void __tcp_alloc_md5sig_pool(void)
3752 {
3753         struct crypto_ahash *hash;
3754         int cpu;
3755
3756         hash = crypto_alloc_ahash("md5", 0, CRYPTO_ALG_ASYNC);
3757         if (IS_ERR(hash))
3758                 return;
3759
3760         for_each_possible_cpu(cpu) {
3761                 void *scratch = per_cpu(tcp_md5sig_pool, cpu).scratch;
3762                 struct ahash_request *req;
3763
3764                 if (!scratch) {
3765                         scratch = kmalloc_node(sizeof(union tcp_md5sum_block) +
3766                                                sizeof(struct tcphdr),
3767                                                GFP_KERNEL,
3768                                                cpu_to_node(cpu));
3769                         if (!scratch)
3770                                 return;
3771                         per_cpu(tcp_md5sig_pool, cpu).scratch = scratch;
3772                 }
3773                 if (per_cpu(tcp_md5sig_pool, cpu).md5_req)
3774                         continue;
3775
3776                 req = ahash_request_alloc(hash, GFP_KERNEL);
3777                 if (!req)
3778                         return;
3779
3780                 ahash_request_set_callback(req, 0, NULL, NULL);
3781
3782                 per_cpu(tcp_md5sig_pool, cpu).md5_req = req;
3783         }
3784         /* before setting tcp_md5sig_pool_populated, we must commit all writes
3785          * to memory. See smp_rmb() in tcp_get_md5sig_pool()
3786          */
3787         smp_wmb();
3788         tcp_md5sig_pool_populated = true;
3789 }
3790
3791 bool tcp_alloc_md5sig_pool(void)
3792 {
3793         if (unlikely(!tcp_md5sig_pool_populated)) {
3794                 mutex_lock(&tcp_md5sig_mutex);
3795
3796                 if (!tcp_md5sig_pool_populated) {
3797                         __tcp_alloc_md5sig_pool();
3798                         if (tcp_md5sig_pool_populated)
3799                                 static_branch_inc(&tcp_md5_needed);
3800                 }
3801
3802                 mutex_unlock(&tcp_md5sig_mutex);
3803         }
3804         return tcp_md5sig_pool_populated;
3805 }
3806 EXPORT_SYMBOL(tcp_alloc_md5sig_pool);
3807
3808
3809 /**
3810  *      tcp_get_md5sig_pool - get md5sig_pool for this user
3811  *
3812  *      We use percpu structure, so if we succeed, we exit with preemption
3813  *      and BH disabled, to make sure another thread or softirq handling
3814  *      wont try to get same context.
3815  */
3816 struct tcp_md5sig_pool *tcp_get_md5sig_pool(void)
3817 {
3818         local_bh_disable();
3819
3820         if (tcp_md5sig_pool_populated) {
3821                 /* coupled with smp_wmb() in __tcp_alloc_md5sig_pool() */
3822                 smp_rmb();
3823                 return this_cpu_ptr(&tcp_md5sig_pool);
3824         }
3825         local_bh_enable();
3826         return NULL;
3827 }
3828 EXPORT_SYMBOL(tcp_get_md5sig_pool);
3829
3830 int tcp_md5_hash_skb_data(struct tcp_md5sig_pool *hp,
3831                           const struct sk_buff *skb, unsigned int header_len)
3832 {
3833         struct scatterlist sg;
3834         const struct tcphdr *tp = tcp_hdr(skb);
3835         struct ahash_request *req = hp->md5_req;
3836         unsigned int i;
3837         const unsigned int head_data_len = skb_headlen(skb) > header_len ?
3838                                            skb_headlen(skb) - header_len : 0;
3839         const struct skb_shared_info *shi = skb_shinfo(skb);
3840         struct sk_buff *frag_iter;
3841
3842         sg_init_table(&sg, 1);
3843
3844         sg_set_buf(&sg, ((u8 *) tp) + header_len, head_data_len);
3845         ahash_request_set_crypt(req, &sg, NULL, head_data_len);
3846         if (crypto_ahash_update(req))
3847                 return 1;
3848
3849         for (i = 0; i < shi->nr_frags; ++i) {
3850                 const skb_frag_t *f = &shi->frags[i];
3851                 unsigned int offset = skb_frag_off(f);
3852                 struct page *page = skb_frag_page(f) + (offset >> PAGE_SHIFT);
3853
3854                 sg_set_page(&sg, page, skb_frag_size(f),
3855                             offset_in_page(offset));
3856                 ahash_request_set_crypt(req, &sg, NULL, skb_frag_size(f));
3857                 if (crypto_ahash_update(req))
3858                         return 1;
3859         }
3860
3861         skb_walk_frags(skb, frag_iter)
3862                 if (tcp_md5_hash_skb_data(hp, frag_iter, 0))
3863                         return 1;
3864
3865         return 0;
3866 }
3867 EXPORT_SYMBOL(tcp_md5_hash_skb_data);
3868
3869 int tcp_md5_hash_key(struct tcp_md5sig_pool *hp, const struct tcp_md5sig_key *key)
3870 {
3871         struct scatterlist sg;
3872
3873         sg_init_one(&sg, key->key, key->keylen);
3874         ahash_request_set_crypt(hp->md5_req, &sg, NULL, key->keylen);
3875         return crypto_ahash_update(hp->md5_req);
3876 }
3877 EXPORT_SYMBOL(tcp_md5_hash_key);
3878
3879 #endif
3880
3881 void tcp_done(struct sock *sk)
3882 {
3883         struct request_sock *req;
3884
3885         /* We might be called with a new socket, after
3886          * inet_csk_prepare_forced_close() has been called
3887          * so we can not use lockdep_sock_is_held(sk)
3888          */
3889         req = rcu_dereference_protected(tcp_sk(sk)->fastopen_rsk, 1);
3890
3891         if (sk->sk_state == TCP_SYN_SENT || sk->sk_state == TCP_SYN_RECV)
3892                 TCP_INC_STATS(sock_net(sk), TCP_MIB_ATTEMPTFAILS);
3893
3894         tcp_set_state(sk, TCP_CLOSE);
3895         tcp_clear_xmit_timers(sk);
3896         if (req)
3897                 reqsk_fastopen_remove(sk, req, false);
3898
3899         sk->sk_shutdown = SHUTDOWN_MASK;
3900
3901         if (!sock_flag(sk, SOCK_DEAD))
3902                 sk->sk_state_change(sk);
3903         else
3904                 inet_csk_destroy_sock(sk);
3905 }
3906 EXPORT_SYMBOL_GPL(tcp_done);
3907
3908 int tcp_abort(struct sock *sk, int err)
3909 {
3910         if (!sk_fullsock(sk)) {
3911                 if (sk->sk_state == TCP_NEW_SYN_RECV) {
3912                         struct request_sock *req = inet_reqsk(sk);
3913
3914                         local_bh_disable();
3915                         inet_csk_reqsk_queue_drop(req->rsk_listener, req);
3916                         local_bh_enable();
3917                         return 0;
3918                 }
3919                 return -EOPNOTSUPP;
3920         }
3921
3922         /* Don't race with userspace socket closes such as tcp_close. */
3923         lock_sock(sk);
3924
3925         if (sk->sk_state == TCP_LISTEN) {
3926                 tcp_set_state(sk, TCP_CLOSE);
3927                 inet_csk_listen_stop(sk);
3928         }
3929
3930         /* Don't race with BH socket closes such as inet_csk_listen_stop. */
3931         local_bh_disable();
3932         bh_lock_sock(sk);
3933
3934         if (!sock_flag(sk, SOCK_DEAD)) {
3935                 sk->sk_err = err;
3936                 /* This barrier is coupled with smp_rmb() in tcp_poll() */
3937                 smp_wmb();
3938                 sk->sk_error_report(sk);
3939                 if (tcp_need_reset(sk->sk_state))
3940                         tcp_send_active_reset(sk, GFP_ATOMIC);
3941                 tcp_done(sk);
3942         }
3943
3944         bh_unlock_sock(sk);
3945         local_bh_enable();
3946         tcp_write_queue_purge(sk);
3947         release_sock(sk);
3948         return 0;
3949 }
3950 EXPORT_SYMBOL_GPL(tcp_abort);
3951
3952 extern struct tcp_congestion_ops tcp_reno;
3953
3954 static __initdata unsigned long thash_entries;
3955 static int __init set_thash_entries(char *str)
3956 {
3957         ssize_t ret;
3958
3959         if (!str)
3960                 return 0;
3961
3962         ret = kstrtoul(str, 0, &thash_entries);
3963         if (ret)
3964                 return 0;
3965
3966         return 1;
3967 }
3968 __setup("thash_entries=", set_thash_entries);
3969
3970 static void __init tcp_init_mem(void)
3971 {
3972         unsigned long limit = nr_free_buffer_pages() / 16;
3973
3974         limit = max(limit, 128UL);
3975         sysctl_tcp_mem[0] = limit / 4 * 3;              /* 4.68 % */
3976         sysctl_tcp_mem[1] = limit;                      /* 6.25 % */
3977         sysctl_tcp_mem[2] = sysctl_tcp_mem[0] * 2;      /* 9.37 % */
3978 }
3979
3980 void __init tcp_init(void)
3981 {
3982         int max_rshare, max_wshare, cnt;
3983         unsigned long limit;
3984         unsigned int i;
3985
3986         BUILD_BUG_ON(TCP_MIN_SND_MSS <= MAX_TCP_OPTION_SPACE);
3987         BUILD_BUG_ON(sizeof(struct tcp_skb_cb) >
3988                      sizeof_field(struct sk_buff, cb));
3989
3990         percpu_counter_init(&tcp_sockets_allocated, 0, GFP_KERNEL);
3991         percpu_counter_init(&tcp_orphan_count, 0, GFP_KERNEL);
3992         inet_hashinfo_init(&tcp_hashinfo);
3993         inet_hashinfo2_init(&tcp_hashinfo, "tcp_listen_portaddr_hash",
3994                             thash_entries, 21,  /* one slot per 2 MB*/
3995                             0, 64 * 1024);
3996         tcp_hashinfo.bind_bucket_cachep =
3997                 kmem_cache_create("tcp_bind_bucket",
3998                                   sizeof(struct inet_bind_bucket), 0,
3999                                   SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL);
4000
4001         /* Size and allocate the main established and bind bucket
4002          * hash tables.
4003          *
4004          * The methodology is similar to that of the buffer cache.
4005          */
4006         tcp_hashinfo.ehash =
4007                 alloc_large_system_hash("TCP established",
4008                                         sizeof(struct inet_ehash_bucket),
4009                                         thash_entries,
4010                                         17, /* one slot per 128 KB of memory */
4011                                         0,
4012                                         NULL,
4013                                         &tcp_hashinfo.ehash_mask,
4014                                         0,
4015                                         thash_entries ? 0 : 512 * 1024);
4016         for (i = 0; i <= tcp_hashinfo.ehash_mask; i++)
4017                 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].chain, i);
4018
4019         if (inet_ehash_locks_alloc(&tcp_hashinfo))
4020                 panic("TCP: failed to alloc ehash_locks");
4021         tcp_hashinfo.bhash =
4022                 alloc_large_system_hash("TCP bind",
4023                                         sizeof(struct inet_bind_hashbucket),
4024                                         tcp_hashinfo.ehash_mask + 1,
4025                                         17, /* one slot per 128 KB of memory */
4026                                         0,
4027                                         &tcp_hashinfo.bhash_size,
4028                                         NULL,
4029                                         0,
4030                                         64 * 1024);
4031         tcp_hashinfo.bhash_size = 1U << tcp_hashinfo.bhash_size;
4032         for (i = 0; i < tcp_hashinfo.bhash_size; i++) {
4033                 spin_lock_init(&tcp_hashinfo.bhash[i].lock);
4034                 INIT_HLIST_HEAD(&tcp_hashinfo.bhash[i].chain);
4035         }
4036
4037
4038         cnt = tcp_hashinfo.ehash_mask + 1;
4039         sysctl_tcp_max_orphans = cnt / 2;
4040
4041         tcp_init_mem();
4042         /* Set per-socket limits to no more than 1/128 the pressure threshold */
4043         limit = nr_free_buffer_pages() << (PAGE_SHIFT - 7);
4044         max_wshare = min(4UL*1024*1024, limit);
4045         max_rshare = min(6UL*1024*1024, limit);
4046
4047         init_net.ipv4.sysctl_tcp_wmem[0] = SK_MEM_QUANTUM;
4048         init_net.ipv4.sysctl_tcp_wmem[1] = 16*1024;
4049         init_net.ipv4.sysctl_tcp_wmem[2] = max(64*1024, max_wshare);
4050
4051         init_net.ipv4.sysctl_tcp_rmem[0] = SK_MEM_QUANTUM;
4052         init_net.ipv4.sysctl_tcp_rmem[1] = 131072;
4053         init_net.ipv4.sysctl_tcp_rmem[2] = max(131072, max_rshare);
4054
4055         pr_info("Hash tables configured (established %u bind %u)\n",
4056                 tcp_hashinfo.ehash_mask + 1, tcp_hashinfo.bhash_size);
4057
4058         tcp_v4_init();
4059         tcp_metrics_init();
4060         BUG_ON(tcp_register_congestion_control(&tcp_reno) != 0);
4061         tcp_tasklet_init();
4062         mptcp_init();
4063 }