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