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/inet_ecn.h>
274#include <net/tcp.h>
275#include <net/tcp_ecn.h>
276#include <net/mptcp.h>
277#include <net/proto_memory.h>
278#include <net/xfrm.h>
279#include <net/ip.h>
280#include <net/psp.h>
281#include <net/sock.h>
282#include <net/rstreason.h>
283
284#include <linux/uaccess.h>
285#include <asm/ioctls.h>
286#include <net/busy_poll.h>
287#include <net/hotdata.h>
288#include <trace/events/tcp.h>
289#include <net/rps.h>
290
291#include "../core/devmem.h"
292
293/* Track pending CMSGs. */
294enum {
295 TCP_CMSG_INQ = 1,
296 TCP_CMSG_TS = 2
297};
298
299DEFINE_PER_CPU(unsigned int, tcp_orphan_count);
300EXPORT_PER_CPU_SYMBOL_GPL(tcp_orphan_count);
301
302DEFINE_PER_CPU(u32, tcp_tw_isn);
303EXPORT_PER_CPU_SYMBOL_GPL(tcp_tw_isn);
304
305long sysctl_tcp_mem[3] __read_mostly;
306EXPORT_IPV6_MOD(sysctl_tcp_mem);
307
308DEFINE_PER_CPU(int, tcp_memory_per_cpu_fw_alloc);
309EXPORT_PER_CPU_SYMBOL_GPL(tcp_memory_per_cpu_fw_alloc);
310
311#if IS_ENABLED(CONFIG_SMC)
312DEFINE_STATIC_KEY_FALSE(tcp_have_smc);
313EXPORT_SYMBOL(tcp_have_smc);
314#endif
315
316/*
317 * Current number of TCP sockets.
318 */
319struct percpu_counter tcp_sockets_allocated ____cacheline_aligned_in_smp;
320EXPORT_IPV6_MOD(tcp_sockets_allocated);
321
322/*
323 * TCP splice context
324 */
325struct tcp_splice_state {
326 struct pipe_inode_info *pipe;
327 size_t len;
328 unsigned int flags;
329};
330
331/*
332 * Pressure flag: try to collapse.
333 * Technical note: it is used by multiple contexts non atomically.
334 * All the __sk_mem_schedule() is of this nature: accounting
335 * is strict, actions are advisory and have some latency.
336 */
337unsigned long tcp_memory_pressure __read_mostly;
338EXPORT_SYMBOL_GPL(tcp_memory_pressure);
339
340void tcp_enter_memory_pressure(struct sock *sk)
341{
342 unsigned long val;
343
344 if (READ_ONCE(tcp_memory_pressure))
345 return;
346 val = jiffies;
347
348 if (!val)
349 val--;
350 if (!cmpxchg(&tcp_memory_pressure, 0, val))
351 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMEMORYPRESSURES);
352}
353EXPORT_IPV6_MOD_GPL(tcp_enter_memory_pressure);
354
355void tcp_leave_memory_pressure(struct sock *sk)
356{
357 unsigned long val;
358
359 if (!READ_ONCE(tcp_memory_pressure))
360 return;
361 val = xchg(&tcp_memory_pressure, 0);
362 if (val)
363 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPMEMORYPRESSURESCHRONO,
364 jiffies_to_msecs(jiffies - val));
365}
366EXPORT_IPV6_MOD_GPL(tcp_leave_memory_pressure);
367
368/* Convert seconds to retransmits based on initial and max timeout */
369static u8 secs_to_retrans(int seconds, int timeout, int rto_max)
370{
371 u8 res = 0;
372
373 if (seconds > 0) {
374 int period = timeout;
375
376 res = 1;
377 while (seconds > period && res < 255) {
378 res++;
379 timeout <<= 1;
380 if (timeout > rto_max)
381 timeout = rto_max;
382 period += timeout;
383 }
384 }
385 return res;
386}
387
388/* Convert retransmits to seconds based on initial and max timeout */
389static int retrans_to_secs(u8 retrans, int timeout, int rto_max)
390{
391 int period = 0;
392
393 if (retrans > 0) {
394 period = timeout;
395 while (--retrans) {
396 timeout <<= 1;
397 if (timeout > rto_max)
398 timeout = rto_max;
399 period += timeout;
400 }
401 }
402 return period;
403}
404
405static u64 tcp_compute_delivery_rate(const struct tcp_sock *tp)
406{
407 u32 rate = READ_ONCE(tp->rate_delivered);
408 u32 intv = READ_ONCE(tp->rate_interval_us);
409 u64 rate64 = 0;
410
411 if (rate && intv) {
412 rate64 = (u64)rate * tp->mss_cache * USEC_PER_SEC;
413 do_div(rate64, intv);
414 }
415 return rate64;
416}
417
418#ifdef CONFIG_TCP_MD5SIG
419void tcp_md5_destruct_sock(struct sock *sk)
420{
421 struct tcp_sock *tp = tcp_sk(sk);
422
423 if (tp->md5sig_info) {
424
425 tcp_clear_md5_list(sk);
426 kfree(rcu_replace_pointer(tp->md5sig_info, NULL, 1));
427 static_branch_slow_dec_deferred(&tcp_md5_needed);
428 tcp_md5_release_sigpool();
429 }
430}
431EXPORT_IPV6_MOD_GPL(tcp_md5_destruct_sock);
432#endif
433
434/* Address-family independent initialization for a tcp_sock.
435 *
436 * NOTE: A lot of things set to zero explicitly by call to
437 * sk_alloc() so need not be done here.
438 */
439void tcp_init_sock(struct sock *sk)
440{
441 struct inet_connection_sock *icsk = inet_csk(sk);
442 struct tcp_sock *tp = tcp_sk(sk);
443 int rto_min_us, rto_max_ms;
444
445 tp->out_of_order_queue = RB_ROOT;
446 sk->tcp_rtx_queue = RB_ROOT;
447 tcp_init_xmit_timers(sk);
448 INIT_LIST_HEAD(list: &tp->tsq_node);
449 INIT_LIST_HEAD(list: &tp->tsorted_sent_queue);
450
451 icsk->icsk_rto = TCP_TIMEOUT_INIT;
452
453 rto_max_ms = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_rto_max_ms);
454 icsk->icsk_rto_max = msecs_to_jiffies(m: rto_max_ms);
455
456 rto_min_us = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_rto_min_us);
457 icsk->icsk_rto_min = usecs_to_jiffies(u: rto_min_us);
458 icsk->icsk_delack_max = TCP_DELACK_MAX;
459 tp->mdev_us = jiffies_to_usecs(TCP_TIMEOUT_INIT);
460 minmax_reset(m: &tp->rtt_min, tcp_jiffies32, meas: ~0U);
461
462 /* So many TCP implementations out there (incorrectly) count the
463 * initial SYN frame in their delayed-ACK and congestion control
464 * algorithms that we must have the following bandaid to talk
465 * efficiently to them. -DaveM
466 */
467 tcp_snd_cwnd_set(tp, TCP_INIT_CWND);
468
469 /* There's a bubble in the pipe until at least the first ACK. */
470 tp->app_limited = ~0U;
471 tp->rate_app_limited = 1;
472
473 /* See draft-stevens-tcpca-spec-01 for discussion of the
474 * initialization of these values.
475 */
476 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
477 tp->snd_cwnd_clamp = ~0;
478 tp->mss_cache = TCP_MSS_DEFAULT;
479
480 tp->reordering = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_reordering);
481 tcp_assign_congestion_control(sk);
482
483 tp->tsoffset = 0;
484 tp->rack.reo_wnd_steps = 1;
485
486 sk->sk_write_space = sk_stream_write_space;
487 sock_set_flag(sk, flag: SOCK_USE_WRITE_QUEUE);
488
489 icsk->icsk_sync_mss = tcp_sync_mss;
490
491 WRITE_ONCE(sk->sk_sndbuf, READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_wmem[1]));
492 WRITE_ONCE(sk->sk_rcvbuf, READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_rmem[1]));
493 tcp_scaling_ratio_init(sk);
494
495 set_bit(nr: SOCK_SUPPORT_ZC, addr: &sk->sk_socket->flags);
496 sk_sockets_allocated_inc(sk);
497 xa_init_flags(xa: &sk->sk_user_frags, XA_FLAGS_ALLOC1);
498}
499EXPORT_IPV6_MOD(tcp_init_sock);
500
501static void tcp_tx_timestamp(struct sock *sk, struct sockcm_cookie *sockc)
502{
503 struct sk_buff *skb = tcp_write_queue_tail(sk);
504 u32 tsflags = sockc->tsflags;
505
506 if (tsflags && skb) {
507 struct skb_shared_info *shinfo = skb_shinfo(skb);
508 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
509
510 sock_tx_timestamp(sk, sockc, tx_flags: &shinfo->tx_flags);
511 if (tsflags & SOF_TIMESTAMPING_TX_ACK)
512 tcb->txstamp_ack |= TSTAMP_ACK_SK;
513 if (tsflags & SOF_TIMESTAMPING_TX_RECORD_MASK)
514 shinfo->tskey = TCP_SKB_CB(skb)->seq + skb->len - 1;
515 }
516
517 if (cgroup_bpf_enabled(CGROUP_SOCK_OPS) &&
518 SK_BPF_CB_FLAG_TEST(sk, SK_BPF_CB_TX_TIMESTAMPING) && skb)
519 bpf_skops_tx_timestamping(sk, skb, op: BPF_SOCK_OPS_TSTAMP_SENDMSG_CB);
520}
521
522static bool tcp_stream_is_readable(struct sock *sk, int target)
523{
524 if (tcp_epollin_ready(sk, target))
525 return true;
526 return sk_is_readable(sk);
527}
528
529/*
530 * Wait for a TCP event.
531 *
532 * Note that we don't need to lock the socket, as the upper poll layers
533 * take care of normal races (between the test and the event) and we don't
534 * go look at any of the socket buffers directly.
535 */
536__poll_t tcp_poll(struct file *file, struct socket *sock, poll_table *wait)
537{
538 __poll_t mask;
539 struct sock *sk = sock->sk;
540 const struct tcp_sock *tp = tcp_sk(sk);
541 u8 shutdown;
542 int state;
543
544 sock_poll_wait(filp: file, sock, p: wait);
545
546 state = inet_sk_state_load(sk);
547 if (state == TCP_LISTEN)
548 return inet_csk_listen_poll(sk);
549
550 /* Socket is not locked. We are protected from async events
551 * by poll logic and correct handling of state changes
552 * made by other threads is impossible in any case.
553 */
554
555 mask = 0;
556
557 /*
558 * EPOLLHUP is certainly not done right. But poll() doesn't
559 * have a notion of HUP in just one direction, and for a
560 * socket the read side is more interesting.
561 *
562 * Some poll() documentation says that EPOLLHUP is incompatible
563 * with the EPOLLOUT/POLLWR flags, so somebody should check this
564 * all. But careful, it tends to be safer to return too many
565 * bits than too few, and you can easily break real applications
566 * if you don't tell them that something has hung up!
567 *
568 * Check-me.
569 *
570 * Check number 1. EPOLLHUP is _UNMASKABLE_ event (see UNIX98 and
571 * our fs/select.c). It means that after we received EOF,
572 * poll always returns immediately, making impossible poll() on write()
573 * in state CLOSE_WAIT. One solution is evident --- to set EPOLLHUP
574 * if and only if shutdown has been made in both directions.
575 * Actually, it is interesting to look how Solaris and DUX
576 * solve this dilemma. I would prefer, if EPOLLHUP were maskable,
577 * then we could set it on SND_SHUTDOWN. BTW examples given
578 * in Stevens' books assume exactly this behaviour, it explains
579 * why EPOLLHUP is incompatible with EPOLLOUT. --ANK
580 *
581 * NOTE. Check for TCP_CLOSE is added. The goal is to prevent
582 * blocking on fresh not-connected or disconnected socket. --ANK
583 */
584 shutdown = READ_ONCE(sk->sk_shutdown);
585 if (shutdown == SHUTDOWN_MASK || state == TCP_CLOSE)
586 mask |= EPOLLHUP;
587 if (shutdown & RCV_SHUTDOWN)
588 mask |= EPOLLIN | EPOLLRDNORM | EPOLLRDHUP;
589
590 /* Connected or passive Fast Open socket? */
591 if (state != TCP_SYN_SENT &&
592 (state != TCP_SYN_RECV || rcu_access_pointer(tp->fastopen_rsk))) {
593 int target = sock_rcvlowat(sk, waitall: 0, INT_MAX);
594 u16 urg_data = READ_ONCE(tp->urg_data);
595
596 if (unlikely(urg_data) &&
597 READ_ONCE(tp->urg_seq) == READ_ONCE(tp->copied_seq) &&
598 !sock_flag(sk, flag: SOCK_URGINLINE))
599 target++;
600
601 if (tcp_stream_is_readable(sk, target))
602 mask |= EPOLLIN | EPOLLRDNORM;
603
604 if (!(shutdown & SEND_SHUTDOWN)) {
605 if (__sk_stream_is_writeable(sk, wake: 1)) {
606 mask |= EPOLLOUT | EPOLLWRNORM;
607 } else { /* send SIGIO later */
608 sk_set_bit(nr: SOCKWQ_ASYNC_NOSPACE, sk);
609 set_bit(nr: SOCK_NOSPACE, addr: &sk->sk_socket->flags);
610
611 /* Race breaker. If space is freed after
612 * wspace test but before the flags are set,
613 * IO signal will be lost. Memory barrier
614 * pairs with the input side.
615 */
616 smp_mb__after_atomic();
617 if (__sk_stream_is_writeable(sk, wake: 1))
618 mask |= EPOLLOUT | EPOLLWRNORM;
619 }
620 } else
621 mask |= EPOLLOUT | EPOLLWRNORM;
622
623 if (urg_data & TCP_URG_VALID)
624 mask |= EPOLLPRI;
625 } else if (state == TCP_SYN_SENT &&
626 inet_test_bit(DEFER_CONNECT, sk)) {
627 /* Active TCP fastopen socket with defer_connect
628 * Return EPOLLOUT so application can call write()
629 * in order for kernel to generate SYN+data
630 */
631 mask |= EPOLLOUT | EPOLLWRNORM;
632 }
633 /* This barrier is coupled with smp_wmb() in tcp_done_with_error() */
634 smp_rmb();
635 if (READ_ONCE(sk->sk_err) ||
636 !skb_queue_empty_lockless(list: &sk->sk_error_queue))
637 mask |= EPOLLERR;
638
639 return mask;
640}
641EXPORT_SYMBOL(tcp_poll);
642
643int tcp_ioctl(struct sock *sk, int cmd, int *karg)
644{
645 struct tcp_sock *tp = tcp_sk(sk);
646 int answ;
647 bool slow;
648
649 switch (cmd) {
650 case SIOCINQ:
651 if (sk->sk_state == TCP_LISTEN)
652 return -EINVAL;
653
654 slow = lock_sock_fast(sk);
655 answ = tcp_inq(sk);
656 unlock_sock_fast(sk, slow);
657 break;
658 case SIOCATMARK:
659 answ = READ_ONCE(tp->urg_data) &&
660 READ_ONCE(tp->urg_seq) == READ_ONCE(tp->copied_seq);
661 break;
662 case SIOCOUTQ:
663 if (sk->sk_state == TCP_LISTEN)
664 return -EINVAL;
665
666 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
667 answ = 0;
668 else
669 answ = READ_ONCE(tp->write_seq) - tp->snd_una;
670 break;
671 case SIOCOUTQNSD:
672 if (sk->sk_state == TCP_LISTEN)
673 return -EINVAL;
674
675 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
676 answ = 0;
677 else
678 answ = READ_ONCE(tp->write_seq) -
679 READ_ONCE(tp->snd_nxt);
680 break;
681 default:
682 return -ENOIOCTLCMD;
683 }
684
685 *karg = answ;
686 return 0;
687}
688EXPORT_IPV6_MOD(tcp_ioctl);
689
690void tcp_mark_push(struct tcp_sock *tp, struct sk_buff *skb)
691{
692 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
693 tp->pushed_seq = tp->write_seq;
694}
695
696static inline bool forced_push(const struct tcp_sock *tp)
697{
698 return after(tp->write_seq, tp->pushed_seq + (tp->max_window >> 1));
699}
700
701void tcp_skb_entail(struct sock *sk, struct sk_buff *skb)
702{
703 struct tcp_sock *tp = tcp_sk(sk);
704 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
705
706 tcb->seq = tcb->end_seq = tp->write_seq;
707 tcb->tcp_flags = TCPHDR_ACK;
708 __skb_header_release(skb);
709 psp_enqueue_set_decrypted(sk, skb);
710 tcp_add_write_queue_tail(sk, skb);
711 sk_wmem_queued_add(sk, val: skb->truesize);
712 sk_mem_charge(sk, size: skb->truesize);
713 if (tp->nonagle & TCP_NAGLE_PUSH)
714 tp->nonagle &= ~TCP_NAGLE_PUSH;
715
716 tcp_slow_start_after_idle_check(sk);
717}
718
719static inline void tcp_mark_urg(struct tcp_sock *tp, int flags)
720{
721 if (flags & MSG_OOB)
722 tp->snd_up = tp->write_seq;
723}
724
725/* If a not yet filled skb is pushed, do not send it if
726 * we have data packets in Qdisc or NIC queues :
727 * Because TX completion will happen shortly, it gives a chance
728 * to coalesce future sendmsg() payload into this skb, without
729 * need for a timer, and with no latency trade off.
730 * As packets containing data payload have a bigger truesize
731 * than pure acks (dataless) packets, the last checks prevent
732 * autocorking if we only have an ACK in Qdisc/NIC queues,
733 * or if TX completion was delayed after we processed ACK packet.
734 */
735static bool tcp_should_autocork(struct sock *sk, struct sk_buff *skb,
736 int size_goal)
737{
738 return skb->len < size_goal &&
739 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_autocorking) &&
740 !tcp_rtx_queue_empty(sk) &&
741 refcount_read(r: &sk->sk_wmem_alloc) > skb->truesize &&
742 tcp_skb_can_collapse_to(skb);
743}
744
745void tcp_push(struct sock *sk, int flags, int mss_now,
746 int nonagle, int size_goal)
747{
748 struct tcp_sock *tp = tcp_sk(sk);
749 struct sk_buff *skb;
750
751 skb = tcp_write_queue_tail(sk);
752 if (!skb)
753 return;
754 if (!(flags & MSG_MORE) || forced_push(tp))
755 tcp_mark_push(tp, skb);
756
757 tcp_mark_urg(tp, flags);
758
759 if (tcp_should_autocork(sk, skb, size_goal)) {
760
761 /* avoid atomic op if TSQ_THROTTLED bit is already set */
762 if (!test_bit(TSQ_THROTTLED, &sk->sk_tsq_flags)) {
763 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPAUTOCORKING);
764 set_bit(nr: TSQ_THROTTLED, addr: &sk->sk_tsq_flags);
765 smp_mb__after_atomic();
766 }
767 /* It is possible TX completion already happened
768 * before we set TSQ_THROTTLED.
769 */
770 if (refcount_read(r: &sk->sk_wmem_alloc) > skb->truesize)
771 return;
772 }
773
774 if (flags & MSG_MORE)
775 nonagle = TCP_NAGLE_CORK;
776
777 __tcp_push_pending_frames(sk, cur_mss: mss_now, nonagle);
778}
779
780static int tcp_splice_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb,
781 unsigned int offset, size_t len)
782{
783 struct tcp_splice_state *tss = rd_desc->arg.data;
784 int ret;
785
786 ret = skb_splice_bits(skb, sk: skb->sk, offset, pipe: tss->pipe,
787 min(rd_desc->count, len), flags: tss->flags);
788 if (ret > 0)
789 rd_desc->count -= ret;
790 return ret;
791}
792
793static int __tcp_splice_read(struct sock *sk, struct tcp_splice_state *tss)
794{
795 /* Store TCP splice context information in read_descriptor_t. */
796 read_descriptor_t rd_desc = {
797 .arg.data = tss,
798 .count = tss->len,
799 };
800
801 return tcp_read_sock(sk, desc: &rd_desc, recv_actor: tcp_splice_data_recv);
802}
803
804/**
805 * tcp_splice_read - splice data from TCP socket to a pipe
806 * @sock: socket to splice from
807 * @ppos: position (not valid)
808 * @pipe: pipe to splice to
809 * @len: number of bytes to splice
810 * @flags: splice modifier flags
811 *
812 * Description:
813 * Will read pages from given socket and fill them into a pipe.
814 *
815 **/
816ssize_t tcp_splice_read(struct socket *sock, loff_t *ppos,
817 struct pipe_inode_info *pipe, size_t len,
818 unsigned int flags)
819{
820 struct sock *sk = sock->sk;
821 struct tcp_splice_state tss = {
822 .pipe = pipe,
823 .len = len,
824 .flags = flags,
825 };
826 long timeo;
827 ssize_t spliced;
828 int ret;
829
830 sock_rps_record_flow(sk);
831 /*
832 * We can't seek on a socket input
833 */
834 if (unlikely(*ppos))
835 return -ESPIPE;
836
837 ret = spliced = 0;
838
839 lock_sock(sk);
840
841 timeo = sock_rcvtimeo(sk, noblock: sock->file->f_flags & O_NONBLOCK);
842 while (tss.len) {
843 ret = __tcp_splice_read(sk, tss: &tss);
844 if (ret < 0)
845 break;
846 else if (!ret) {
847 if (spliced)
848 break;
849 if (sock_flag(sk, flag: SOCK_DONE))
850 break;
851 if (sk->sk_err) {
852 ret = sock_error(sk);
853 break;
854 }
855 if (sk->sk_shutdown & RCV_SHUTDOWN)
856 break;
857 if (sk->sk_state == TCP_CLOSE) {
858 /*
859 * This occurs when user tries to read
860 * from never connected socket.
861 */
862 ret = -ENOTCONN;
863 break;
864 }
865 if (!timeo) {
866 ret = -EAGAIN;
867 break;
868 }
869 /* if __tcp_splice_read() got nothing while we have
870 * an skb in receive queue, we do not want to loop.
871 * This might happen with URG data.
872 */
873 if (!skb_queue_empty(list: &sk->sk_receive_queue))
874 break;
875 ret = sk_wait_data(sk, timeo: &timeo, NULL);
876 if (ret < 0)
877 break;
878 if (signal_pending(current)) {
879 ret = sock_intr_errno(timeo);
880 break;
881 }
882 continue;
883 }
884 tss.len -= ret;
885 spliced += ret;
886
887 if (!tss.len || !timeo)
888 break;
889 release_sock(sk);
890 lock_sock(sk);
891
892 if (sk->sk_err || sk->sk_state == TCP_CLOSE ||
893 (sk->sk_shutdown & RCV_SHUTDOWN) ||
894 signal_pending(current))
895 break;
896 }
897
898 release_sock(sk);
899
900 if (spliced)
901 return spliced;
902
903 return ret;
904}
905EXPORT_IPV6_MOD(tcp_splice_read);
906
907struct sk_buff *tcp_stream_alloc_skb(struct sock *sk, gfp_t gfp,
908 bool force_schedule)
909{
910 struct sk_buff *skb;
911
912 skb = alloc_skb_fclone(MAX_TCP_HEADER, priority: gfp);
913 if (likely(skb)) {
914 bool mem_scheduled;
915
916 skb->truesize = SKB_TRUESIZE(skb_end_offset(skb));
917 if (force_schedule) {
918 mem_scheduled = true;
919 sk_forced_mem_schedule(sk, size: skb->truesize);
920 } else {
921 mem_scheduled = sk_wmem_schedule(sk, size: skb->truesize);
922 }
923 if (likely(mem_scheduled)) {
924 skb_reserve(skb, MAX_TCP_HEADER);
925 skb->ip_summed = CHECKSUM_PARTIAL;
926 INIT_LIST_HEAD(list: &skb->tcp_tsorted_anchor);
927 return skb;
928 }
929 __kfree_skb(skb);
930 } else {
931 sk->sk_prot->enter_memory_pressure(sk);
932 sk_stream_moderate_sndbuf(sk);
933 }
934 return NULL;
935}
936
937static unsigned int tcp_xmit_size_goal(struct sock *sk, u32 mss_now,
938 int large_allowed)
939{
940 struct tcp_sock *tp = tcp_sk(sk);
941 u32 new_size_goal, size_goal;
942
943 if (!large_allowed)
944 return mss_now;
945
946 /* Note : tcp_tso_autosize() will eventually split this later */
947 new_size_goal = tcp_bound_to_half_wnd(tp, pktsize: sk->sk_gso_max_size);
948
949 /* We try hard to avoid divides here */
950 size_goal = tp->gso_segs * mss_now;
951 if (unlikely(new_size_goal < size_goal ||
952 new_size_goal >= size_goal + mss_now)) {
953 tp->gso_segs = min_t(u16, new_size_goal / mss_now,
954 sk->sk_gso_max_segs);
955 size_goal = tp->gso_segs * mss_now;
956 }
957
958 return max(size_goal, mss_now);
959}
960
961int tcp_send_mss(struct sock *sk, int *size_goal, int flags)
962{
963 int mss_now;
964
965 mss_now = tcp_current_mss(sk);
966 *size_goal = tcp_xmit_size_goal(sk, mss_now, large_allowed: !(flags & MSG_OOB));
967
968 return mss_now;
969}
970
971/* In some cases, sendmsg() could have added an skb to the write queue,
972 * but failed adding payload on it. We need to remove it to consume less
973 * memory, but more importantly be able to generate EPOLLOUT for Edge Trigger
974 * epoll() users. Another reason is that tcp_write_xmit() does not like
975 * finding an empty skb in the write queue.
976 */
977void tcp_remove_empty_skb(struct sock *sk)
978{
979 struct sk_buff *skb = tcp_write_queue_tail(sk);
980
981 if (skb && TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq) {
982 tcp_unlink_write_queue(skb, sk);
983 if (tcp_write_queue_empty(sk))
984 tcp_chrono_stop(sk, type: TCP_CHRONO_BUSY);
985 tcp_wmem_free_skb(sk, skb);
986 }
987}
988
989/* skb changing from pure zc to mixed, must charge zc */
990static int tcp_downgrade_zcopy_pure(struct sock *sk, struct sk_buff *skb)
991{
992 if (unlikely(skb_zcopy_pure(skb))) {
993 u32 extra = skb->truesize -
994 SKB_TRUESIZE(skb_end_offset(skb));
995
996 if (!sk_wmem_schedule(sk, size: extra))
997 return -ENOMEM;
998
999 sk_mem_charge(sk, size: extra);
1000 skb_shinfo(skb)->flags &= ~SKBFL_PURE_ZEROCOPY;
1001 }
1002 return 0;
1003}
1004
1005
1006int tcp_wmem_schedule(struct sock *sk, int copy)
1007{
1008 int left;
1009
1010 if (likely(sk_wmem_schedule(sk, copy)))
1011 return copy;
1012
1013 /* We could be in trouble if we have nothing queued.
1014 * Use whatever is left in sk->sk_forward_alloc and tcp_wmem[0]
1015 * to guarantee some progress.
1016 */
1017 left = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_wmem[0]) - sk->sk_wmem_queued;
1018 if (left > 0)
1019 sk_forced_mem_schedule(sk, min(left, copy));
1020 return min(copy, sk->sk_forward_alloc);
1021}
1022
1023void tcp_free_fastopen_req(struct tcp_sock *tp)
1024{
1025 if (tp->fastopen_req) {
1026 kfree(objp: tp->fastopen_req);
1027 tp->fastopen_req = NULL;
1028 }
1029}
1030
1031int tcp_sendmsg_fastopen(struct sock *sk, struct msghdr *msg, int *copied,
1032 size_t size, struct ubuf_info *uarg)
1033{
1034 struct tcp_sock *tp = tcp_sk(sk);
1035 struct inet_sock *inet = inet_sk(sk);
1036 struct sockaddr *uaddr = msg->msg_name;
1037 int err, flags;
1038
1039 if (!(READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_fastopen) &
1040 TFO_CLIENT_ENABLE) ||
1041 (uaddr && msg->msg_namelen >= sizeof(uaddr->sa_family) &&
1042 uaddr->sa_family == AF_UNSPEC))
1043 return -EOPNOTSUPP;
1044 if (tp->fastopen_req)
1045 return -EALREADY; /* Another Fast Open is in progress */
1046
1047 tp->fastopen_req = kzalloc(sizeof(struct tcp_fastopen_request),
1048 sk->sk_allocation);
1049 if (unlikely(!tp->fastopen_req))
1050 return -ENOBUFS;
1051 tp->fastopen_req->data = msg;
1052 tp->fastopen_req->size = size;
1053 tp->fastopen_req->uarg = uarg;
1054
1055 if (inet_test_bit(DEFER_CONNECT, sk)) {
1056 err = tcp_connect(sk);
1057 /* Same failure procedure as in tcp_v4/6_connect */
1058 if (err) {
1059 tcp_set_state(sk, state: TCP_CLOSE);
1060 inet->inet_dport = 0;
1061 sk->sk_route_caps = 0;
1062 }
1063 }
1064 flags = (msg->msg_flags & MSG_DONTWAIT) ? O_NONBLOCK : 0;
1065 err = __inet_stream_connect(sock: sk->sk_socket, uaddr,
1066 addr_len: msg->msg_namelen, flags, is_sendmsg: 1);
1067 /* fastopen_req could already be freed in __inet_stream_connect
1068 * if the connection times out or gets rst
1069 */
1070 if (tp->fastopen_req) {
1071 *copied = tp->fastopen_req->copied;
1072 tcp_free_fastopen_req(tp);
1073 inet_clear_bit(DEFER_CONNECT, sk);
1074 }
1075 return err;
1076}
1077
1078int tcp_sendmsg_locked(struct sock *sk, struct msghdr *msg, size_t size)
1079{
1080 struct net_devmem_dmabuf_binding *binding = NULL;
1081 struct tcp_sock *tp = tcp_sk(sk);
1082 struct ubuf_info *uarg = NULL;
1083 struct sk_buff *skb;
1084 struct sockcm_cookie sockc;
1085 int flags, err, copied = 0;
1086 int mss_now = 0, size_goal, copied_syn = 0;
1087 int process_backlog = 0;
1088 int sockc_err = 0;
1089 int zc = 0;
1090 long timeo;
1091
1092 flags = msg->msg_flags;
1093
1094 sockc = (struct sockcm_cookie){ .tsflags = READ_ONCE(sk->sk_tsflags) };
1095 if (msg->msg_controllen) {
1096 sockc_err = sock_cmsg_send(sk, msg, sockc: &sockc);
1097 /* Don't return error until MSG_FASTOPEN has been processed;
1098 * that may succeed even if the cmsg is invalid.
1099 */
1100 }
1101
1102 if ((flags & MSG_ZEROCOPY) && size) {
1103 if (msg->msg_ubuf) {
1104 uarg = msg->msg_ubuf;
1105 if (sk->sk_route_caps & NETIF_F_SG)
1106 zc = MSG_ZEROCOPY;
1107 } else if (sock_flag(sk, flag: SOCK_ZEROCOPY)) {
1108 skb = tcp_write_queue_tail(sk);
1109 uarg = msg_zerocopy_realloc(sk, size, uarg: skb_zcopy(skb),
1110 devmem: !sockc_err && sockc.dmabuf_id);
1111 if (!uarg) {
1112 err = -ENOBUFS;
1113 goto out_err;
1114 }
1115 if (sk->sk_route_caps & NETIF_F_SG)
1116 zc = MSG_ZEROCOPY;
1117 else
1118 uarg_to_msgzc(uarg)->zerocopy = 0;
1119
1120 if (!sockc_err && sockc.dmabuf_id) {
1121 binding = net_devmem_get_binding(sk, dmabuf_id: sockc.dmabuf_id);
1122 if (IS_ERR(ptr: binding)) {
1123 err = PTR_ERR(ptr: binding);
1124 binding = NULL;
1125 goto out_err;
1126 }
1127 }
1128 }
1129 } else if (unlikely(msg->msg_flags & MSG_SPLICE_PAGES) && size) {
1130 if (sk->sk_route_caps & NETIF_F_SG)
1131 zc = MSG_SPLICE_PAGES;
1132 }
1133
1134 if (!sockc_err && sockc.dmabuf_id &&
1135 (!(flags & MSG_ZEROCOPY) || !sock_flag(sk, flag: SOCK_ZEROCOPY))) {
1136 err = -EINVAL;
1137 goto out_err;
1138 }
1139
1140 if (unlikely(flags & MSG_FASTOPEN ||
1141 inet_test_bit(DEFER_CONNECT, sk)) &&
1142 !tp->repair) {
1143 err = tcp_sendmsg_fastopen(sk, msg, copied: &copied_syn, size, uarg);
1144 if (err == -EINPROGRESS && copied_syn > 0)
1145 goto out;
1146 else if (err)
1147 goto out_err;
1148 }
1149
1150 timeo = sock_sndtimeo(sk, noblock: flags & MSG_DONTWAIT);
1151
1152 tcp_rate_check_app_limited(sk); /* is sending application-limited? */
1153
1154 /* Wait for a connection to finish. One exception is TCP Fast Open
1155 * (passive side) where data is allowed to be sent before a connection
1156 * is fully established.
1157 */
1158 if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) &&
1159 !tcp_passive_fastopen(sk)) {
1160 err = sk_stream_wait_connect(sk, timeo_p: &timeo);
1161 if (err != 0)
1162 goto do_error;
1163 }
1164
1165 if (unlikely(tp->repair)) {
1166 if (tp->repair_queue == TCP_RECV_QUEUE) {
1167 copied = tcp_send_rcvq(sk, msg, size);
1168 goto out_nopush;
1169 }
1170
1171 err = -EINVAL;
1172 if (tp->repair_queue == TCP_NO_QUEUE)
1173 goto out_err;
1174
1175 /* 'common' sending to sendq */
1176 }
1177
1178 if (sockc_err) {
1179 err = sockc_err;
1180 goto out_err;
1181 }
1182
1183 /* This should be in poll */
1184 sk_clear_bit(nr: SOCKWQ_ASYNC_NOSPACE, sk);
1185
1186 /* Ok commence sending. */
1187 copied = 0;
1188
1189restart:
1190 mss_now = tcp_send_mss(sk, size_goal: &size_goal, flags);
1191
1192 err = -EPIPE;
1193 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
1194 goto do_error;
1195
1196 while (msg_data_left(msg)) {
1197 int copy = 0;
1198
1199 skb = tcp_write_queue_tail(sk);
1200 if (skb)
1201 copy = size_goal - skb->len;
1202
1203 trace_tcp_sendmsg_locked(sk, msg, skb, size_goal);
1204
1205 if (copy <= 0 || !tcp_skb_can_collapse_to(skb)) {
1206 bool first_skb;
1207
1208new_segment:
1209 if (!sk_stream_memory_free(sk))
1210 goto wait_for_space;
1211
1212 if (unlikely(process_backlog >= 16)) {
1213 process_backlog = 0;
1214 if (sk_flush_backlog(sk))
1215 goto restart;
1216 }
1217 first_skb = tcp_rtx_and_write_queues_empty(sk);
1218 skb = tcp_stream_alloc_skb(sk, gfp: sk->sk_allocation,
1219 force_schedule: first_skb);
1220 if (!skb)
1221 goto wait_for_space;
1222
1223 process_backlog++;
1224
1225#ifdef CONFIG_SKB_DECRYPTED
1226 skb->decrypted = !!(flags & MSG_SENDPAGE_DECRYPTED);
1227#endif
1228 tcp_skb_entail(sk, skb);
1229 copy = size_goal;
1230
1231 /* All packets are restored as if they have
1232 * already been sent. skb_mstamp_ns isn't set to
1233 * avoid wrong rtt estimation.
1234 */
1235 if (tp->repair)
1236 TCP_SKB_CB(skb)->sacked |= TCPCB_REPAIRED;
1237 }
1238
1239 /* Try to append data to the end of skb. */
1240 if (copy > msg_data_left(msg))
1241 copy = msg_data_left(msg);
1242
1243 if (zc == 0) {
1244 bool merge = true;
1245 int i = skb_shinfo(skb)->nr_frags;
1246 struct page_frag *pfrag = sk_page_frag(sk);
1247
1248 if (!sk_page_frag_refill(sk, pfrag))
1249 goto wait_for_space;
1250
1251 if (!skb_can_coalesce(skb, i, page: pfrag->page,
1252 off: pfrag->offset)) {
1253 if (i >= READ_ONCE(net_hotdata.sysctl_max_skb_frags)) {
1254 tcp_mark_push(tp, skb);
1255 goto new_segment;
1256 }
1257 merge = false;
1258 }
1259
1260 copy = min_t(int, copy, pfrag->size - pfrag->offset);
1261
1262 if (unlikely(skb_zcopy_pure(skb) || skb_zcopy_managed(skb))) {
1263 if (tcp_downgrade_zcopy_pure(sk, skb))
1264 goto wait_for_space;
1265 skb_zcopy_downgrade_managed(skb);
1266 }
1267
1268 copy = tcp_wmem_schedule(sk, copy);
1269 if (!copy)
1270 goto wait_for_space;
1271
1272 err = skb_copy_to_page_nocache(sk, from: &msg->msg_iter, skb,
1273 page: pfrag->page,
1274 off: pfrag->offset,
1275 copy);
1276 if (err)
1277 goto do_error;
1278
1279 /* Update the skb. */
1280 if (merge) {
1281 skb_frag_size_add(frag: &skb_shinfo(skb)->frags[i - 1], delta: copy);
1282 } else {
1283 skb_fill_page_desc(skb, i, page: pfrag->page,
1284 off: pfrag->offset, size: copy);
1285 page_ref_inc(page: pfrag->page);
1286 }
1287 pfrag->offset += copy;
1288 } else if (zc == MSG_ZEROCOPY) {
1289 /* First append to a fragless skb builds initial
1290 * pure zerocopy skb
1291 */
1292 if (!skb->len)
1293 skb_shinfo(skb)->flags |= SKBFL_PURE_ZEROCOPY;
1294
1295 if (!skb_zcopy_pure(skb)) {
1296 copy = tcp_wmem_schedule(sk, copy);
1297 if (!copy)
1298 goto wait_for_space;
1299 }
1300
1301 err = skb_zerocopy_iter_stream(sk, skb, msg, len: copy, uarg,
1302 binding);
1303 if (err == -EMSGSIZE || err == -EEXIST) {
1304 tcp_mark_push(tp, skb);
1305 goto new_segment;
1306 }
1307 if (err < 0)
1308 goto do_error;
1309 copy = err;
1310 } else if (zc == MSG_SPLICE_PAGES) {
1311 /* Splice in data if we can; copy if we can't. */
1312 if (tcp_downgrade_zcopy_pure(sk, skb))
1313 goto wait_for_space;
1314 copy = tcp_wmem_schedule(sk, copy);
1315 if (!copy)
1316 goto wait_for_space;
1317
1318 err = skb_splice_from_iter(skb, iter: &msg->msg_iter, maxsize: copy);
1319 if (err < 0) {
1320 if (err == -EMSGSIZE) {
1321 tcp_mark_push(tp, skb);
1322 goto new_segment;
1323 }
1324 goto do_error;
1325 }
1326 copy = err;
1327
1328 if (!(flags & MSG_NO_SHARED_FRAGS))
1329 skb_shinfo(skb)->flags |= SKBFL_SHARED_FRAG;
1330
1331 sk_wmem_queued_add(sk, val: copy);
1332 sk_mem_charge(sk, size: copy);
1333 }
1334
1335 if (!copied)
1336 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
1337
1338 WRITE_ONCE(tp->write_seq, tp->write_seq + copy);
1339 TCP_SKB_CB(skb)->end_seq += copy;
1340 tcp_skb_pcount_set(skb, segs: 0);
1341
1342 copied += copy;
1343 if (!msg_data_left(msg)) {
1344 if (unlikely(flags & MSG_EOR))
1345 TCP_SKB_CB(skb)->eor = 1;
1346 goto out;
1347 }
1348
1349 if (skb->len < size_goal || (flags & MSG_OOB) || unlikely(tp->repair))
1350 continue;
1351
1352 if (forced_push(tp)) {
1353 tcp_mark_push(tp, skb);
1354 __tcp_push_pending_frames(sk, cur_mss: mss_now, TCP_NAGLE_PUSH);
1355 } else if (skb == tcp_send_head(sk))
1356 tcp_push_one(sk, mss_now);
1357 continue;
1358
1359wait_for_space:
1360 set_bit(nr: SOCK_NOSPACE, addr: &sk->sk_socket->flags);
1361 tcp_remove_empty_skb(sk);
1362 if (copied)
1363 tcp_push(sk, flags: flags & ~MSG_MORE, mss_now,
1364 TCP_NAGLE_PUSH, size_goal);
1365
1366 err = sk_stream_wait_memory(sk, timeo_p: &timeo);
1367 if (err != 0)
1368 goto do_error;
1369
1370 mss_now = tcp_send_mss(sk, size_goal: &size_goal, flags);
1371 }
1372
1373out:
1374 if (copied) {
1375 tcp_tx_timestamp(sk, sockc: &sockc);
1376 tcp_push(sk, flags, mss_now, nonagle: tp->nonagle, size_goal);
1377 }
1378out_nopush:
1379 /* msg->msg_ubuf is pinned by the caller so we don't take extra refs */
1380 if (uarg && !msg->msg_ubuf)
1381 net_zcopy_put(uarg);
1382 if (binding)
1383 net_devmem_dmabuf_binding_put(binding);
1384 return copied + copied_syn;
1385
1386do_error:
1387 tcp_remove_empty_skb(sk);
1388
1389 if (copied + copied_syn)
1390 goto out;
1391out_err:
1392 /* msg->msg_ubuf is pinned by the caller so we don't take extra refs */
1393 if (uarg && !msg->msg_ubuf)
1394 net_zcopy_put_abort(uarg, have_uref: true);
1395 err = sk_stream_error(sk, flags, err);
1396 /* make sure we wake any epoll edge trigger waiter */
1397 if (unlikely(tcp_rtx_and_write_queues_empty(sk) && err == -EAGAIN)) {
1398 sk->sk_write_space(sk);
1399 tcp_chrono_stop(sk, type: TCP_CHRONO_SNDBUF_LIMITED);
1400 }
1401 if (binding)
1402 net_devmem_dmabuf_binding_put(binding);
1403
1404 return err;
1405}
1406EXPORT_SYMBOL_GPL(tcp_sendmsg_locked);
1407
1408int tcp_sendmsg(struct sock *sk, struct msghdr *msg, size_t size)
1409{
1410 int ret;
1411
1412 lock_sock(sk);
1413 ret = tcp_sendmsg_locked(sk, msg, size);
1414 release_sock(sk);
1415
1416 return ret;
1417}
1418EXPORT_SYMBOL(tcp_sendmsg);
1419
1420void tcp_splice_eof(struct socket *sock)
1421{
1422 struct sock *sk = sock->sk;
1423 struct tcp_sock *tp = tcp_sk(sk);
1424 int mss_now, size_goal;
1425
1426 if (!tcp_write_queue_tail(sk))
1427 return;
1428
1429 lock_sock(sk);
1430 mss_now = tcp_send_mss(sk, size_goal: &size_goal, flags: 0);
1431 tcp_push(sk, flags: 0, mss_now, nonagle: tp->nonagle, size_goal);
1432 release_sock(sk);
1433}
1434EXPORT_IPV6_MOD_GPL(tcp_splice_eof);
1435
1436/*
1437 * Handle reading urgent data. BSD has very simple semantics for
1438 * this, no blocking and very strange errors 8)
1439 */
1440
1441static int tcp_recv_urg(struct sock *sk, struct msghdr *msg, int len, int flags)
1442{
1443 struct tcp_sock *tp = tcp_sk(sk);
1444
1445 /* No URG data to read. */
1446 if (sock_flag(sk, flag: SOCK_URGINLINE) || !tp->urg_data ||
1447 tp->urg_data == TCP_URG_READ)
1448 return -EINVAL; /* Yes this is right ! */
1449
1450 if (sk->sk_state == TCP_CLOSE && !sock_flag(sk, flag: SOCK_DONE))
1451 return -ENOTCONN;
1452
1453 if (tp->urg_data & TCP_URG_VALID) {
1454 int err = 0;
1455 char c = tp->urg_data;
1456
1457 if (!(flags & MSG_PEEK))
1458 WRITE_ONCE(tp->urg_data, TCP_URG_READ);
1459
1460 /* Read urgent data. */
1461 msg->msg_flags |= MSG_OOB;
1462
1463 if (len > 0) {
1464 if (!(flags & MSG_TRUNC))
1465 err = memcpy_to_msg(msg, data: &c, len: 1);
1466 len = 1;
1467 } else
1468 msg->msg_flags |= MSG_TRUNC;
1469
1470 return err ? -EFAULT : len;
1471 }
1472
1473 if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
1474 return 0;
1475
1476 /* Fixed the recv(..., MSG_OOB) behaviour. BSD docs and
1477 * the available implementations agree in this case:
1478 * this call should never block, independent of the
1479 * blocking state of the socket.
1480 * Mike <pall@rz.uni-karlsruhe.de>
1481 */
1482 return -EAGAIN;
1483}
1484
1485static int tcp_peek_sndq(struct sock *sk, struct msghdr *msg, int len)
1486{
1487 struct sk_buff *skb;
1488 int copied = 0, err = 0;
1489
1490 skb_rbtree_walk(skb, &sk->tcp_rtx_queue) {
1491 err = skb_copy_datagram_msg(from: skb, offset: 0, msg, size: skb->len);
1492 if (err)
1493 return err;
1494 copied += skb->len;
1495 }
1496
1497 skb_queue_walk(&sk->sk_write_queue, skb) {
1498 err = skb_copy_datagram_msg(from: skb, offset: 0, msg, size: skb->len);
1499 if (err)
1500 break;
1501
1502 copied += skb->len;
1503 }
1504
1505 return err ?: copied;
1506}
1507
1508/* Clean up the receive buffer for full frames taken by the user,
1509 * then send an ACK if necessary. COPIED is the number of bytes
1510 * tcp_recvmsg has given to the user so far, it speeds up the
1511 * calculation of whether or not we must ACK for the sake of
1512 * a window update.
1513 */
1514void __tcp_cleanup_rbuf(struct sock *sk, int copied)
1515{
1516 struct tcp_sock *tp = tcp_sk(sk);
1517 bool time_to_ack = false;
1518
1519 if (inet_csk_ack_scheduled(sk)) {
1520 const struct inet_connection_sock *icsk = inet_csk(sk);
1521
1522 if (/* Once-per-two-segments ACK was not sent by tcp_input.c */
1523 tp->rcv_nxt - tp->rcv_wup > icsk->icsk_ack.rcv_mss ||
1524 /*
1525 * If this read emptied read buffer, we send ACK, if
1526 * connection is not bidirectional, user drained
1527 * receive buffer and there was a small segment
1528 * in queue.
1529 */
1530 (copied > 0 &&
1531 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED2) ||
1532 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED) &&
1533 !inet_csk_in_pingpong_mode(sk))) &&
1534 !atomic_read(v: &sk->sk_rmem_alloc)))
1535 time_to_ack = true;
1536 }
1537
1538 /* We send an ACK if we can now advertise a non-zero window
1539 * which has been raised "significantly".
1540 *
1541 * Even if window raised up to infinity, do not send window open ACK
1542 * in states, where we will not receive more. It is useless.
1543 */
1544 if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) {
1545 __u32 rcv_window_now = tcp_receive_window(tp);
1546
1547 /* Optimize, __tcp_select_window() is not cheap. */
1548 if (2*rcv_window_now <= tp->window_clamp) {
1549 __u32 new_window = __tcp_select_window(sk);
1550
1551 /* Send ACK now, if this read freed lots of space
1552 * in our buffer. Certainly, new_window is new window.
1553 * We can advertise it now, if it is not less than current one.
1554 * "Lots" means "at least twice" here.
1555 */
1556 if (new_window && new_window >= 2 * rcv_window_now)
1557 time_to_ack = true;
1558 }
1559 }
1560 if (time_to_ack)
1561 tcp_send_ack(sk);
1562}
1563
1564void tcp_cleanup_rbuf(struct sock *sk, int copied)
1565{
1566 struct sk_buff *skb = skb_peek(list_: &sk->sk_receive_queue);
1567 struct tcp_sock *tp = tcp_sk(sk);
1568
1569 WARN(skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq),
1570 "cleanup rbuf bug: copied %X seq %X rcvnxt %X\n",
1571 tp->copied_seq, TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt);
1572 __tcp_cleanup_rbuf(sk, copied);
1573}
1574
1575static void tcp_eat_recv_skb(struct sock *sk, struct sk_buff *skb)
1576{
1577 __skb_unlink(skb, list: &sk->sk_receive_queue);
1578 if (likely(skb->destructor == sock_rfree)) {
1579 sock_rfree(skb);
1580 skb->destructor = NULL;
1581 skb->sk = NULL;
1582 return skb_attempt_defer_free(skb);
1583 }
1584 __kfree_skb(skb);
1585}
1586
1587struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off)
1588{
1589 struct sk_buff *skb;
1590 u32 offset;
1591
1592 while ((skb = skb_peek(list_: &sk->sk_receive_queue)) != NULL) {
1593 offset = seq - TCP_SKB_CB(skb)->seq;
1594 if (unlikely(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
1595 pr_err_once("%s: found a SYN, please report !\n", __func__);
1596 offset--;
1597 }
1598 if (offset < skb->len || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)) {
1599 *off = offset;
1600 return skb;
1601 }
1602 /* This looks weird, but this can happen if TCP collapsing
1603 * splitted a fat GRO packet, while we released socket lock
1604 * in skb_splice_bits()
1605 */
1606 tcp_eat_recv_skb(sk, skb);
1607 }
1608 return NULL;
1609}
1610EXPORT_SYMBOL(tcp_recv_skb);
1611
1612/*
1613 * This routine provides an alternative to tcp_recvmsg() for routines
1614 * that would like to handle copying from skbuffs directly in 'sendfile'
1615 * fashion.
1616 * Note:
1617 * - It is assumed that the socket was locked by the caller.
1618 * - The routine does not block.
1619 * - At present, there is no support for reading OOB data
1620 * or for 'peeking' the socket using this routine
1621 * (although both would be easy to implement).
1622 */
1623static int __tcp_read_sock(struct sock *sk, read_descriptor_t *desc,
1624 sk_read_actor_t recv_actor, bool noack,
1625 u32 *copied_seq)
1626{
1627 struct sk_buff *skb;
1628 struct tcp_sock *tp = tcp_sk(sk);
1629 u32 seq = *copied_seq;
1630 u32 offset;
1631 int copied = 0;
1632
1633 if (sk->sk_state == TCP_LISTEN)
1634 return -ENOTCONN;
1635 while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1636 if (offset < skb->len) {
1637 int used;
1638 size_t len;
1639
1640 len = skb->len - offset;
1641 /* Stop reading if we hit a patch of urgent data */
1642 if (unlikely(tp->urg_data)) {
1643 u32 urg_offset = tp->urg_seq - seq;
1644 if (urg_offset < len)
1645 len = urg_offset;
1646 if (!len)
1647 break;
1648 }
1649 used = recv_actor(desc, skb, offset, len);
1650 if (used <= 0) {
1651 if (!copied)
1652 copied = used;
1653 break;
1654 }
1655 if (WARN_ON_ONCE(used > len))
1656 used = len;
1657 seq += used;
1658 copied += used;
1659 offset += used;
1660
1661 /* If recv_actor drops the lock (e.g. TCP splice
1662 * receive) the skb pointer might be invalid when
1663 * getting here: tcp_collapse might have deleted it
1664 * while aggregating skbs from the socket queue.
1665 */
1666 skb = tcp_recv_skb(sk, seq - 1, &offset);
1667 if (!skb)
1668 break;
1669 /* TCP coalescing might have appended data to the skb.
1670 * Try to splice more frags
1671 */
1672 if (offset + 1 != skb->len)
1673 continue;
1674 }
1675 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) {
1676 tcp_eat_recv_skb(sk, skb);
1677 ++seq;
1678 break;
1679 }
1680 tcp_eat_recv_skb(sk, skb);
1681 if (!desc->count)
1682 break;
1683 WRITE_ONCE(*copied_seq, seq);
1684 }
1685 WRITE_ONCE(*copied_seq, seq);
1686
1687 if (noack)
1688 goto out;
1689
1690 tcp_rcv_space_adjust(sk);
1691
1692 /* Clean up data we have read: This will do ACK frames. */
1693 if (copied > 0) {
1694 tcp_recv_skb(sk, seq, &offset);
1695 tcp_cleanup_rbuf(sk, copied);
1696 }
1697out:
1698 return copied;
1699}
1700
1701int tcp_read_sock(struct sock *sk, read_descriptor_t *desc,
1702 sk_read_actor_t recv_actor)
1703{
1704 return __tcp_read_sock(sk, desc, recv_actor, noack: false,
1705 copied_seq: &tcp_sk(sk)->copied_seq);
1706}
1707EXPORT_SYMBOL(tcp_read_sock);
1708
1709int tcp_read_sock_noack(struct sock *sk, read_descriptor_t *desc,
1710 sk_read_actor_t recv_actor, bool noack,
1711 u32 *copied_seq)
1712{
1713 return __tcp_read_sock(sk, desc, recv_actor, noack, copied_seq);
1714}
1715
1716int tcp_read_skb(struct sock *sk, skb_read_actor_t recv_actor)
1717{
1718 struct sk_buff *skb;
1719 int copied = 0;
1720
1721 if (sk->sk_state == TCP_LISTEN)
1722 return -ENOTCONN;
1723
1724 while ((skb = skb_peek(list_: &sk->sk_receive_queue)) != NULL) {
1725 u8 tcp_flags;
1726 int used;
1727
1728 __skb_unlink(skb, list: &sk->sk_receive_queue);
1729 WARN_ON_ONCE(!skb_set_owner_sk_safe(skb, sk));
1730 tcp_flags = TCP_SKB_CB(skb)->tcp_flags;
1731 used = recv_actor(sk, skb);
1732 if (used < 0) {
1733 if (!copied)
1734 copied = used;
1735 break;
1736 }
1737 copied += used;
1738
1739 if (tcp_flags & TCPHDR_FIN)
1740 break;
1741 }
1742 return copied;
1743}
1744EXPORT_IPV6_MOD(tcp_read_skb);
1745
1746void tcp_read_done(struct sock *sk, size_t len)
1747{
1748 struct tcp_sock *tp = tcp_sk(sk);
1749 u32 seq = tp->copied_seq;
1750 struct sk_buff *skb;
1751 size_t left;
1752 u32 offset;
1753
1754 if (sk->sk_state == TCP_LISTEN)
1755 return;
1756
1757 left = len;
1758 while (left && (skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1759 int used;
1760
1761 used = min_t(size_t, skb->len - offset, left);
1762 seq += used;
1763 left -= used;
1764
1765 if (skb->len > offset + used)
1766 break;
1767
1768 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) {
1769 tcp_eat_recv_skb(sk, skb);
1770 ++seq;
1771 break;
1772 }
1773 tcp_eat_recv_skb(sk, skb);
1774 }
1775 WRITE_ONCE(tp->copied_seq, seq);
1776
1777 tcp_rcv_space_adjust(sk);
1778
1779 /* Clean up data we have read: This will do ACK frames. */
1780 if (left != len)
1781 tcp_cleanup_rbuf(sk, copied: len - left);
1782}
1783EXPORT_SYMBOL(tcp_read_done);
1784
1785int tcp_peek_len(struct socket *sock)
1786{
1787 return tcp_inq(sk: sock->sk);
1788}
1789EXPORT_IPV6_MOD(tcp_peek_len);
1790
1791/* Make sure sk_rcvbuf is big enough to satisfy SO_RCVLOWAT hint */
1792int tcp_set_rcvlowat(struct sock *sk, int val)
1793{
1794 struct tcp_sock *tp = tcp_sk(sk);
1795 int space, cap;
1796
1797 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK)
1798 cap = sk->sk_rcvbuf >> 1;
1799 else
1800 cap = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_rmem[2]) >> 1;
1801 val = min(val, cap);
1802 WRITE_ONCE(sk->sk_rcvlowat, val ? : 1);
1803
1804 /* Check if we need to signal EPOLLIN right now */
1805 tcp_data_ready(sk);
1806
1807 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK)
1808 return 0;
1809
1810 space = tcp_space_from_win(sk, win: val);
1811 if (space > sk->sk_rcvbuf) {
1812 WRITE_ONCE(sk->sk_rcvbuf, space);
1813
1814 if (tp->window_clamp && tp->window_clamp < val)
1815 WRITE_ONCE(tp->window_clamp, val);
1816 }
1817 return 0;
1818}
1819EXPORT_IPV6_MOD(tcp_set_rcvlowat);
1820
1821void tcp_update_recv_tstamps(struct sk_buff *skb,
1822 struct scm_timestamping_internal *tss)
1823{
1824 if (skb->tstamp)
1825 tss->ts[0] = ktime_to_timespec64(skb->tstamp);
1826 else
1827 tss->ts[0] = (struct timespec64) {0};
1828
1829 if (skb_hwtstamps(skb)->hwtstamp)
1830 tss->ts[2] = ktime_to_timespec64(skb_hwtstamps(skb)->hwtstamp);
1831 else
1832 tss->ts[2] = (struct timespec64) {0};
1833}
1834
1835#ifdef CONFIG_MMU
1836static const struct vm_operations_struct tcp_vm_ops = {
1837};
1838
1839int tcp_mmap(struct file *file, struct socket *sock,
1840 struct vm_area_struct *vma)
1841{
1842 if (vma->vm_flags & (VM_WRITE | VM_EXEC))
1843 return -EPERM;
1844 vm_flags_clear(vma, VM_MAYWRITE | VM_MAYEXEC);
1845
1846 /* Instruct vm_insert_page() to not mmap_read_lock(mm) */
1847 vm_flags_set(vma, VM_MIXEDMAP);
1848
1849 vma->vm_ops = &tcp_vm_ops;
1850 return 0;
1851}
1852EXPORT_IPV6_MOD(tcp_mmap);
1853
1854static skb_frag_t *skb_advance_to_frag(struct sk_buff *skb, u32 offset_skb,
1855 u32 *offset_frag)
1856{
1857 skb_frag_t *frag;
1858
1859 if (unlikely(offset_skb >= skb->len))
1860 return NULL;
1861
1862 offset_skb -= skb_headlen(skb);
1863 if ((int)offset_skb < 0 || skb_has_frag_list(skb))
1864 return NULL;
1865
1866 frag = skb_shinfo(skb)->frags;
1867 while (offset_skb) {
1868 if (skb_frag_size(frag) > offset_skb) {
1869 *offset_frag = offset_skb;
1870 return frag;
1871 }
1872 offset_skb -= skb_frag_size(frag);
1873 ++frag;
1874 }
1875 *offset_frag = 0;
1876 return frag;
1877}
1878
1879static bool can_map_frag(const skb_frag_t *frag)
1880{
1881 struct page *page;
1882
1883 if (skb_frag_size(frag) != PAGE_SIZE || skb_frag_off(frag))
1884 return false;
1885
1886 page = skb_frag_page(frag);
1887
1888 if (PageCompound(page) || page->mapping)
1889 return false;
1890
1891 return true;
1892}
1893
1894static int find_next_mappable_frag(const skb_frag_t *frag,
1895 int remaining_in_skb)
1896{
1897 int offset = 0;
1898
1899 if (likely(can_map_frag(frag)))
1900 return 0;
1901
1902 while (offset < remaining_in_skb && !can_map_frag(frag)) {
1903 offset += skb_frag_size(frag);
1904 ++frag;
1905 }
1906 return offset;
1907}
1908
1909static void tcp_zerocopy_set_hint_for_skb(struct sock *sk,
1910 struct tcp_zerocopy_receive *zc,
1911 struct sk_buff *skb, u32 offset)
1912{
1913 u32 frag_offset, partial_frag_remainder = 0;
1914 int mappable_offset;
1915 skb_frag_t *frag;
1916
1917 /* worst case: skip to next skb. try to improve on this case below */
1918 zc->recv_skip_hint = skb->len - offset;
1919
1920 /* Find the frag containing this offset (and how far into that frag) */
1921 frag = skb_advance_to_frag(skb, offset_skb: offset, offset_frag: &frag_offset);
1922 if (!frag)
1923 return;
1924
1925 if (frag_offset) {
1926 struct skb_shared_info *info = skb_shinfo(skb);
1927
1928 /* We read part of the last frag, must recvmsg() rest of skb. */
1929 if (frag == &info->frags[info->nr_frags - 1])
1930 return;
1931
1932 /* Else, we must at least read the remainder in this frag. */
1933 partial_frag_remainder = skb_frag_size(frag) - frag_offset;
1934 zc->recv_skip_hint -= partial_frag_remainder;
1935 ++frag;
1936 }
1937
1938 /* partial_frag_remainder: If part way through a frag, must read rest.
1939 * mappable_offset: Bytes till next mappable frag, *not* counting bytes
1940 * in partial_frag_remainder.
1941 */
1942 mappable_offset = find_next_mappable_frag(frag, remaining_in_skb: zc->recv_skip_hint);
1943 zc->recv_skip_hint = mappable_offset + partial_frag_remainder;
1944}
1945
1946static int tcp_recvmsg_locked(struct sock *sk, struct msghdr *msg, size_t len,
1947 int flags, struct scm_timestamping_internal *tss,
1948 int *cmsg_flags);
1949static int receive_fallback_to_copy(struct sock *sk,
1950 struct tcp_zerocopy_receive *zc, int inq,
1951 struct scm_timestamping_internal *tss)
1952{
1953 unsigned long copy_address = (unsigned long)zc->copybuf_address;
1954 struct msghdr msg = {};
1955 int err;
1956
1957 zc->length = 0;
1958 zc->recv_skip_hint = 0;
1959
1960 if (copy_address != zc->copybuf_address)
1961 return -EINVAL;
1962
1963 err = import_ubuf(ITER_DEST, buf: (void __user *)copy_address, len: inq,
1964 i: &msg.msg_iter);
1965 if (err)
1966 return err;
1967
1968 err = tcp_recvmsg_locked(sk, msg: &msg, len: inq, MSG_DONTWAIT,
1969 tss, cmsg_flags: &zc->msg_flags);
1970 if (err < 0)
1971 return err;
1972
1973 zc->copybuf_len = err;
1974 if (likely(zc->copybuf_len)) {
1975 struct sk_buff *skb;
1976 u32 offset;
1977
1978 skb = tcp_recv_skb(sk, tcp_sk(sk)->copied_seq, &offset);
1979 if (skb)
1980 tcp_zerocopy_set_hint_for_skb(sk, zc, skb, offset);
1981 }
1982 return 0;
1983}
1984
1985static int tcp_copy_straggler_data(struct tcp_zerocopy_receive *zc,
1986 struct sk_buff *skb, u32 copylen,
1987 u32 *offset, u32 *seq)
1988{
1989 unsigned long copy_address = (unsigned long)zc->copybuf_address;
1990 struct msghdr msg = {};
1991 int err;
1992
1993 if (copy_address != zc->copybuf_address)
1994 return -EINVAL;
1995
1996 err = import_ubuf(ITER_DEST, buf: (void __user *)copy_address, len: copylen,
1997 i: &msg.msg_iter);
1998 if (err)
1999 return err;
2000 err = skb_copy_datagram_msg(from: skb, offset: *offset, msg: &msg, size: copylen);
2001 if (err)
2002 return err;
2003 zc->recv_skip_hint -= copylen;
2004 *offset += copylen;
2005 *seq += copylen;
2006 return (__s32)copylen;
2007}
2008
2009static int tcp_zc_handle_leftover(struct tcp_zerocopy_receive *zc,
2010 struct sock *sk,
2011 struct sk_buff *skb,
2012 u32 *seq,
2013 s32 copybuf_len,
2014 struct scm_timestamping_internal *tss)
2015{
2016 u32 offset, copylen = min_t(u32, copybuf_len, zc->recv_skip_hint);
2017
2018 if (!copylen)
2019 return 0;
2020 /* skb is null if inq < PAGE_SIZE. */
2021 if (skb) {
2022 offset = *seq - TCP_SKB_CB(skb)->seq;
2023 } else {
2024 skb = tcp_recv_skb(sk, *seq, &offset);
2025 if (TCP_SKB_CB(skb)->has_rxtstamp) {
2026 tcp_update_recv_tstamps(skb, tss);
2027 zc->msg_flags |= TCP_CMSG_TS;
2028 }
2029 }
2030
2031 zc->copybuf_len = tcp_copy_straggler_data(zc, skb, copylen, offset: &offset,
2032 seq);
2033 return zc->copybuf_len < 0 ? 0 : copylen;
2034}
2035
2036static int tcp_zerocopy_vm_insert_batch_error(struct vm_area_struct *vma,
2037 struct page **pending_pages,
2038 unsigned long pages_remaining,
2039 unsigned long *address,
2040 u32 *length,
2041 u32 *seq,
2042 struct tcp_zerocopy_receive *zc,
2043 u32 total_bytes_to_map,
2044 int err)
2045{
2046 /* At least one page did not map. Try zapping if we skipped earlier. */
2047 if (err == -EBUSY &&
2048 zc->flags & TCP_RECEIVE_ZEROCOPY_FLAG_TLB_CLEAN_HINT) {
2049 u32 maybe_zap_len;
2050
2051 maybe_zap_len = total_bytes_to_map - /* All bytes to map */
2052 *length + /* Mapped or pending */
2053 (pages_remaining * PAGE_SIZE); /* Failed map. */
2054 zap_page_range_single(vma, address: *address, size: maybe_zap_len, NULL);
2055 err = 0;
2056 }
2057
2058 if (!err) {
2059 unsigned long leftover_pages = pages_remaining;
2060 int bytes_mapped;
2061
2062 /* We called zap_page_range_single, try to reinsert. */
2063 err = vm_insert_pages(vma, addr: *address,
2064 pages: pending_pages,
2065 num: &pages_remaining);
2066 bytes_mapped = PAGE_SIZE * (leftover_pages - pages_remaining);
2067 *seq += bytes_mapped;
2068 *address += bytes_mapped;
2069 }
2070 if (err) {
2071 /* Either we were unable to zap, OR we zapped, retried an
2072 * insert, and still had an issue. Either ways, pages_remaining
2073 * is the number of pages we were unable to map, and we unroll
2074 * some state we speculatively touched before.
2075 */
2076 const int bytes_not_mapped = PAGE_SIZE * pages_remaining;
2077
2078 *length -= bytes_not_mapped;
2079 zc->recv_skip_hint += bytes_not_mapped;
2080 }
2081 return err;
2082}
2083
2084static int tcp_zerocopy_vm_insert_batch(struct vm_area_struct *vma,
2085 struct page **pages,
2086 unsigned int pages_to_map,
2087 unsigned long *address,
2088 u32 *length,
2089 u32 *seq,
2090 struct tcp_zerocopy_receive *zc,
2091 u32 total_bytes_to_map)
2092{
2093 unsigned long pages_remaining = pages_to_map;
2094 unsigned int pages_mapped;
2095 unsigned int bytes_mapped;
2096 int err;
2097
2098 err = vm_insert_pages(vma, addr: *address, pages, num: &pages_remaining);
2099 pages_mapped = pages_to_map - (unsigned int)pages_remaining;
2100 bytes_mapped = PAGE_SIZE * pages_mapped;
2101 /* Even if vm_insert_pages fails, it may have partially succeeded in
2102 * mapping (some but not all of the pages).
2103 */
2104 *seq += bytes_mapped;
2105 *address += bytes_mapped;
2106
2107 if (likely(!err))
2108 return 0;
2109
2110 /* Error: maybe zap and retry + rollback state for failed inserts. */
2111 return tcp_zerocopy_vm_insert_batch_error(vma, pending_pages: pages + pages_mapped,
2112 pages_remaining, address, length, seq, zc, total_bytes_to_map,
2113 err);
2114}
2115
2116#define TCP_VALID_ZC_MSG_FLAGS (TCP_CMSG_TS)
2117static void tcp_zc_finalize_rx_tstamp(struct sock *sk,
2118 struct tcp_zerocopy_receive *zc,
2119 struct scm_timestamping_internal *tss)
2120{
2121 unsigned long msg_control_addr;
2122 struct msghdr cmsg_dummy;
2123
2124 msg_control_addr = (unsigned long)zc->msg_control;
2125 cmsg_dummy.msg_control_user = (void __user *)msg_control_addr;
2126 cmsg_dummy.msg_controllen =
2127 (__kernel_size_t)zc->msg_controllen;
2128 cmsg_dummy.msg_flags = in_compat_syscall()
2129 ? MSG_CMSG_COMPAT : 0;
2130 cmsg_dummy.msg_control_is_user = true;
2131 zc->msg_flags = 0;
2132 if (zc->msg_control == msg_control_addr &&
2133 zc->msg_controllen == cmsg_dummy.msg_controllen) {
2134 tcp_recv_timestamp(msg: &cmsg_dummy, sk, tss);
2135 zc->msg_control = (__u64)
2136 ((uintptr_t)cmsg_dummy.msg_control_user);
2137 zc->msg_controllen =
2138 (__u64)cmsg_dummy.msg_controllen;
2139 zc->msg_flags = (__u32)cmsg_dummy.msg_flags;
2140 }
2141}
2142
2143static struct vm_area_struct *find_tcp_vma(struct mm_struct *mm,
2144 unsigned long address,
2145 bool *mmap_locked)
2146{
2147 struct vm_area_struct *vma = lock_vma_under_rcu(mm, address);
2148
2149 if (vma) {
2150 if (vma->vm_ops != &tcp_vm_ops) {
2151 vma_end_read(vma);
2152 return NULL;
2153 }
2154 *mmap_locked = false;
2155 return vma;
2156 }
2157
2158 mmap_read_lock(mm);
2159 vma = vma_lookup(mm, addr: address);
2160 if (!vma || vma->vm_ops != &tcp_vm_ops) {
2161 mmap_read_unlock(mm);
2162 return NULL;
2163 }
2164 *mmap_locked = true;
2165 return vma;
2166}
2167
2168#define TCP_ZEROCOPY_PAGE_BATCH_SIZE 32
2169static int tcp_zerocopy_receive(struct sock *sk,
2170 struct tcp_zerocopy_receive *zc,
2171 struct scm_timestamping_internal *tss)
2172{
2173 u32 length = 0, offset, vma_len, avail_len, copylen = 0;
2174 unsigned long address = (unsigned long)zc->address;
2175 struct page *pages[TCP_ZEROCOPY_PAGE_BATCH_SIZE];
2176 s32 copybuf_len = zc->copybuf_len;
2177 struct tcp_sock *tp = tcp_sk(sk);
2178 const skb_frag_t *frags = NULL;
2179 unsigned int pages_to_map = 0;
2180 struct vm_area_struct *vma;
2181 struct sk_buff *skb = NULL;
2182 u32 seq = tp->copied_seq;
2183 u32 total_bytes_to_map;
2184 int inq = tcp_inq(sk);
2185 bool mmap_locked;
2186 int ret;
2187
2188 zc->copybuf_len = 0;
2189 zc->msg_flags = 0;
2190
2191 if (address & (PAGE_SIZE - 1) || address != zc->address)
2192 return -EINVAL;
2193
2194 if (sk->sk_state == TCP_LISTEN)
2195 return -ENOTCONN;
2196
2197 sock_rps_record_flow(sk);
2198
2199 if (inq && inq <= copybuf_len)
2200 return receive_fallback_to_copy(sk, zc, inq, tss);
2201
2202 if (inq < PAGE_SIZE) {
2203 zc->length = 0;
2204 zc->recv_skip_hint = inq;
2205 if (!inq && sock_flag(sk, flag: SOCK_DONE))
2206 return -EIO;
2207 return 0;
2208 }
2209
2210 vma = find_tcp_vma(current->mm, address, mmap_locked: &mmap_locked);
2211 if (!vma)
2212 return -EINVAL;
2213
2214 vma_len = min_t(unsigned long, zc->length, vma->vm_end - address);
2215 avail_len = min_t(u32, vma_len, inq);
2216 total_bytes_to_map = avail_len & ~(PAGE_SIZE - 1);
2217 if (total_bytes_to_map) {
2218 if (!(zc->flags & TCP_RECEIVE_ZEROCOPY_FLAG_TLB_CLEAN_HINT))
2219 zap_page_range_single(vma, address, size: total_bytes_to_map,
2220 NULL);
2221 zc->length = total_bytes_to_map;
2222 zc->recv_skip_hint = 0;
2223 } else {
2224 zc->length = avail_len;
2225 zc->recv_skip_hint = avail_len;
2226 }
2227 ret = 0;
2228 while (length + PAGE_SIZE <= zc->length) {
2229 int mappable_offset;
2230 struct page *page;
2231
2232 if (zc->recv_skip_hint < PAGE_SIZE) {
2233 u32 offset_frag;
2234
2235 if (skb) {
2236 if (zc->recv_skip_hint > 0)
2237 break;
2238 skb = skb->next;
2239 offset = seq - TCP_SKB_CB(skb)->seq;
2240 } else {
2241 skb = tcp_recv_skb(sk, seq, &offset);
2242 }
2243
2244 if (!skb_frags_readable(skb))
2245 break;
2246
2247 if (TCP_SKB_CB(skb)->has_rxtstamp) {
2248 tcp_update_recv_tstamps(skb, tss);
2249 zc->msg_flags |= TCP_CMSG_TS;
2250 }
2251 zc->recv_skip_hint = skb->len - offset;
2252 frags = skb_advance_to_frag(skb, offset_skb: offset, offset_frag: &offset_frag);
2253 if (!frags || offset_frag)
2254 break;
2255 }
2256
2257 mappable_offset = find_next_mappable_frag(frag: frags,
2258 remaining_in_skb: zc->recv_skip_hint);
2259 if (mappable_offset) {
2260 zc->recv_skip_hint = mappable_offset;
2261 break;
2262 }
2263 page = skb_frag_page(frag: frags);
2264 if (WARN_ON_ONCE(!page))
2265 break;
2266
2267 prefetchw(x: page);
2268 pages[pages_to_map++] = page;
2269 length += PAGE_SIZE;
2270 zc->recv_skip_hint -= PAGE_SIZE;
2271 frags++;
2272 if (pages_to_map == TCP_ZEROCOPY_PAGE_BATCH_SIZE ||
2273 zc->recv_skip_hint < PAGE_SIZE) {
2274 /* Either full batch, or we're about to go to next skb
2275 * (and we cannot unroll failed ops across skbs).
2276 */
2277 ret = tcp_zerocopy_vm_insert_batch(vma, pages,
2278 pages_to_map,
2279 address: &address, length: &length,
2280 seq: &seq, zc,
2281 total_bytes_to_map);
2282 if (ret)
2283 goto out;
2284 pages_to_map = 0;
2285 }
2286 }
2287 if (pages_to_map) {
2288 ret = tcp_zerocopy_vm_insert_batch(vma, pages, pages_to_map,
2289 address: &address, length: &length, seq: &seq,
2290 zc, total_bytes_to_map);
2291 }
2292out:
2293 if (mmap_locked)
2294 mmap_read_unlock(current->mm);
2295 else
2296 vma_end_read(vma);
2297 /* Try to copy straggler data. */
2298 if (!ret)
2299 copylen = tcp_zc_handle_leftover(zc, sk, skb, seq: &seq, copybuf_len, tss);
2300
2301 if (length + copylen) {
2302 WRITE_ONCE(tp->copied_seq, seq);
2303 tcp_rcv_space_adjust(sk);
2304
2305 /* Clean up data we have read: This will do ACK frames. */
2306 tcp_recv_skb(sk, seq, &offset);
2307 tcp_cleanup_rbuf(sk, copied: length + copylen);
2308 ret = 0;
2309 if (length == zc->length)
2310 zc->recv_skip_hint = 0;
2311 } else {
2312 if (!zc->recv_skip_hint && sock_flag(sk, flag: SOCK_DONE))
2313 ret = -EIO;
2314 }
2315 zc->length = length;
2316 return ret;
2317}
2318#endif
2319
2320/* Similar to __sock_recv_timestamp, but does not require an skb */
2321void tcp_recv_timestamp(struct msghdr *msg, const struct sock *sk,
2322 struct scm_timestamping_internal *tss)
2323{
2324 int new_tstamp = sock_flag(sk, flag: SOCK_TSTAMP_NEW);
2325 u32 tsflags = READ_ONCE(sk->sk_tsflags);
2326 bool has_timestamping = false;
2327
2328 if (tss->ts[0].tv_sec || tss->ts[0].tv_nsec) {
2329 if (sock_flag(sk, flag: SOCK_RCVTSTAMP)) {
2330 if (sock_flag(sk, flag: SOCK_RCVTSTAMPNS)) {
2331 if (new_tstamp) {
2332 struct __kernel_timespec kts = {
2333 .tv_sec = tss->ts[0].tv_sec,
2334 .tv_nsec = tss->ts[0].tv_nsec,
2335 };
2336 put_cmsg(msg, SOL_SOCKET, SO_TIMESTAMPNS_NEW,
2337 len: sizeof(kts), data: &kts);
2338 } else {
2339 struct __kernel_old_timespec ts_old = {
2340 .tv_sec = tss->ts[0].tv_sec,
2341 .tv_nsec = tss->ts[0].tv_nsec,
2342 };
2343 put_cmsg(msg, SOL_SOCKET, SO_TIMESTAMPNS_OLD,
2344 len: sizeof(ts_old), data: &ts_old);
2345 }
2346 } else {
2347 if (new_tstamp) {
2348 struct __kernel_sock_timeval stv = {
2349 .tv_sec = tss->ts[0].tv_sec,
2350 .tv_usec = tss->ts[0].tv_nsec / 1000,
2351 };
2352 put_cmsg(msg, SOL_SOCKET, SO_TIMESTAMP_NEW,
2353 len: sizeof(stv), data: &stv);
2354 } else {
2355 struct __kernel_old_timeval tv = {
2356 .tv_sec = tss->ts[0].tv_sec,
2357 .tv_usec = tss->ts[0].tv_nsec / 1000,
2358 };
2359 put_cmsg(msg, SOL_SOCKET, SO_TIMESTAMP_OLD,
2360 len: sizeof(tv), data: &tv);
2361 }
2362 }
2363 }
2364
2365 if (tsflags & SOF_TIMESTAMPING_SOFTWARE &&
2366 (tsflags & SOF_TIMESTAMPING_RX_SOFTWARE ||
2367 !(tsflags & SOF_TIMESTAMPING_OPT_RX_FILTER)))
2368 has_timestamping = true;
2369 else
2370 tss->ts[0] = (struct timespec64) {0};
2371 }
2372
2373 if (tss->ts[2].tv_sec || tss->ts[2].tv_nsec) {
2374 if (tsflags & SOF_TIMESTAMPING_RAW_HARDWARE &&
2375 (tsflags & SOF_TIMESTAMPING_RX_HARDWARE ||
2376 !(tsflags & SOF_TIMESTAMPING_OPT_RX_FILTER)))
2377 has_timestamping = true;
2378 else
2379 tss->ts[2] = (struct timespec64) {0};
2380 }
2381
2382 if (has_timestamping) {
2383 tss->ts[1] = (struct timespec64) {0};
2384 if (sock_flag(sk, SOCK_TSTAMP_NEW))
2385 put_cmsg_scm_timestamping64(msg, tss);
2386 else
2387 put_cmsg_scm_timestamping(msg, tss);
2388 }
2389}
2390
2391static int tcp_inq_hint(struct sock *sk)
2392{
2393 const struct tcp_sock *tp = tcp_sk(sk);
2394 u32 copied_seq = READ_ONCE(tp->copied_seq);
2395 u32 rcv_nxt = READ_ONCE(tp->rcv_nxt);
2396 int inq;
2397
2398 inq = rcv_nxt - copied_seq;
2399 if (unlikely(inq < 0 || copied_seq != READ_ONCE(tp->copied_seq))) {
2400 lock_sock(sk);
2401 inq = tp->rcv_nxt - tp->copied_seq;
2402 release_sock(sk);
2403 }
2404 /* After receiving a FIN, tell the user-space to continue reading
2405 * by returning a non-zero inq.
2406 */
2407 if (inq == 0 && sock_flag(sk, flag: SOCK_DONE))
2408 inq = 1;
2409 return inq;
2410}
2411
2412/* batch __xa_alloc() calls and reduce xa_lock()/xa_unlock() overhead. */
2413struct tcp_xa_pool {
2414 u8 max; /* max <= MAX_SKB_FRAGS */
2415 u8 idx; /* idx <= max */
2416 __u32 tokens[MAX_SKB_FRAGS];
2417 netmem_ref netmems[MAX_SKB_FRAGS];
2418};
2419
2420static void tcp_xa_pool_commit_locked(struct sock *sk, struct tcp_xa_pool *p)
2421{
2422 int i;
2423
2424 /* Commit part that has been copied to user space. */
2425 for (i = 0; i < p->idx; i++)
2426 __xa_cmpxchg(&sk->sk_user_frags, index: p->tokens[i], XA_ZERO_ENTRY,
2427 entry: (__force void *)p->netmems[i], GFP_KERNEL);
2428 /* Rollback what has been pre-allocated and is no longer needed. */
2429 for (; i < p->max; i++)
2430 __xa_erase(&sk->sk_user_frags, index: p->tokens[i]);
2431
2432 p->max = 0;
2433 p->idx = 0;
2434}
2435
2436static void tcp_xa_pool_commit(struct sock *sk, struct tcp_xa_pool *p)
2437{
2438 if (!p->max)
2439 return;
2440
2441 xa_lock_bh(&sk->sk_user_frags);
2442
2443 tcp_xa_pool_commit_locked(sk, p);
2444
2445 xa_unlock_bh(&sk->sk_user_frags);
2446}
2447
2448static int tcp_xa_pool_refill(struct sock *sk, struct tcp_xa_pool *p,
2449 unsigned int max_frags)
2450{
2451 int err, k;
2452
2453 if (p->idx < p->max)
2454 return 0;
2455
2456 xa_lock_bh(&sk->sk_user_frags);
2457
2458 tcp_xa_pool_commit_locked(sk, p);
2459
2460 for (k = 0; k < max_frags; k++) {
2461 err = __xa_alloc(&sk->sk_user_frags, id: &p->tokens[k],
2462 XA_ZERO_ENTRY, xa_limit_31b, GFP_KERNEL);
2463 if (err)
2464 break;
2465 }
2466
2467 xa_unlock_bh(&sk->sk_user_frags);
2468
2469 p->max = k;
2470 p->idx = 0;
2471 return k ? 0 : err;
2472}
2473
2474/* On error, returns the -errno. On success, returns number of bytes sent to the
2475 * user. May not consume all of @remaining_len.
2476 */
2477static int tcp_recvmsg_dmabuf(struct sock *sk, const struct sk_buff *skb,
2478 unsigned int offset, struct msghdr *msg,
2479 int remaining_len)
2480{
2481 struct dmabuf_cmsg dmabuf_cmsg = { 0 };
2482 struct tcp_xa_pool tcp_xa_pool;
2483 unsigned int start;
2484 int i, copy, n;
2485 int sent = 0;
2486 int err = 0;
2487
2488 tcp_xa_pool.max = 0;
2489 tcp_xa_pool.idx = 0;
2490 do {
2491 start = skb_headlen(skb);
2492
2493 if (skb_frags_readable(skb)) {
2494 err = -ENODEV;
2495 goto out;
2496 }
2497
2498 /* Copy header. */
2499 copy = start - offset;
2500 if (copy > 0) {
2501 copy = min(copy, remaining_len);
2502
2503 n = copy_to_iter(addr: skb->data + offset, bytes: copy,
2504 i: &msg->msg_iter);
2505 if (n != copy) {
2506 err = -EFAULT;
2507 goto out;
2508 }
2509
2510 offset += copy;
2511 remaining_len -= copy;
2512
2513 /* First a dmabuf_cmsg for # bytes copied to user
2514 * buffer.
2515 */
2516 memset(s: &dmabuf_cmsg, c: 0, n: sizeof(dmabuf_cmsg));
2517 dmabuf_cmsg.frag_size = copy;
2518 err = put_cmsg_notrunc(msg, SOL_SOCKET,
2519 SO_DEVMEM_LINEAR,
2520 len: sizeof(dmabuf_cmsg),
2521 data: &dmabuf_cmsg);
2522 if (err)
2523 goto out;
2524
2525 sent += copy;
2526
2527 if (remaining_len == 0)
2528 goto out;
2529 }
2530
2531 /* after that, send information of dmabuf pages through a
2532 * sequence of cmsg
2533 */
2534 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
2535 skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
2536 struct net_iov *niov;
2537 u64 frag_offset;
2538 int end;
2539
2540 /* !skb_frags_readable() should indicate that ALL the
2541 * frags in this skb are dmabuf net_iovs. We're checking
2542 * for that flag above, but also check individual frags
2543 * here. If the tcp stack is not setting
2544 * skb_frags_readable() correctly, we still don't want
2545 * to crash here.
2546 */
2547 if (!skb_frag_net_iov(frag)) {
2548 net_err_ratelimited("Found non-dmabuf skb with net_iov");
2549 err = -ENODEV;
2550 goto out;
2551 }
2552
2553 niov = skb_frag_net_iov(frag);
2554 if (!net_is_devmem_iov(niov)) {
2555 err = -ENODEV;
2556 goto out;
2557 }
2558
2559 end = start + skb_frag_size(frag);
2560 copy = end - offset;
2561
2562 if (copy > 0) {
2563 copy = min(copy, remaining_len);
2564
2565 frag_offset = net_iov_virtual_addr(niov) +
2566 skb_frag_off(frag) + offset -
2567 start;
2568 dmabuf_cmsg.frag_offset = frag_offset;
2569 dmabuf_cmsg.frag_size = copy;
2570 err = tcp_xa_pool_refill(sk, p: &tcp_xa_pool,
2571 skb_shinfo(skb)->nr_frags - i);
2572 if (err)
2573 goto out;
2574
2575 /* Will perform the exchange later */
2576 dmabuf_cmsg.frag_token = tcp_xa_pool.tokens[tcp_xa_pool.idx];
2577 dmabuf_cmsg.dmabuf_id = net_devmem_iov_binding_id(niov);
2578
2579 offset += copy;
2580 remaining_len -= copy;
2581
2582 err = put_cmsg_notrunc(msg, SOL_SOCKET,
2583 SO_DEVMEM_DMABUF,
2584 len: sizeof(dmabuf_cmsg),
2585 data: &dmabuf_cmsg);
2586 if (err)
2587 goto out;
2588
2589 atomic_long_inc(v: &niov->pp_ref_count);
2590 tcp_xa_pool.netmems[tcp_xa_pool.idx++] = skb_frag_netmem(frag);
2591
2592 sent += copy;
2593
2594 if (remaining_len == 0)
2595 goto out;
2596 }
2597 start = end;
2598 }
2599
2600 tcp_xa_pool_commit(sk, p: &tcp_xa_pool);
2601 if (!remaining_len)
2602 goto out;
2603
2604 /* if remaining_len is not satisfied yet, we need to go to the
2605 * next frag in the frag_list to satisfy remaining_len.
2606 */
2607 skb = skb_shinfo(skb)->frag_list ?: skb->next;
2608
2609 offset = offset - start;
2610 } while (skb);
2611
2612 if (remaining_len) {
2613 err = -EFAULT;
2614 goto out;
2615 }
2616
2617out:
2618 tcp_xa_pool_commit(sk, p: &tcp_xa_pool);
2619 if (!sent)
2620 sent = err;
2621
2622 return sent;
2623}
2624
2625/*
2626 * This routine copies from a sock struct into the user buffer.
2627 *
2628 * Technical note: in 2.3 we work on _locked_ socket, so that
2629 * tricks with *seq access order and skb->users are not required.
2630 * Probably, code can be easily improved even more.
2631 */
2632
2633static int tcp_recvmsg_locked(struct sock *sk, struct msghdr *msg, size_t len,
2634 int flags, struct scm_timestamping_internal *tss,
2635 int *cmsg_flags)
2636{
2637 struct tcp_sock *tp = tcp_sk(sk);
2638 int last_copied_dmabuf = -1; /* uninitialized */
2639 int copied = 0;
2640 u32 peek_seq;
2641 u32 *seq;
2642 unsigned long used;
2643 int err;
2644 int target; /* Read at least this many bytes */
2645 long timeo;
2646 struct sk_buff *skb, *last;
2647 u32 peek_offset = 0;
2648 u32 urg_hole = 0;
2649
2650 err = -ENOTCONN;
2651 if (sk->sk_state == TCP_LISTEN)
2652 goto out;
2653
2654 if (tp->recvmsg_inq) {
2655 *cmsg_flags = TCP_CMSG_INQ;
2656 msg->msg_get_inq = 1;
2657 }
2658 timeo = sock_rcvtimeo(sk, noblock: flags & MSG_DONTWAIT);
2659
2660 /* Urgent data needs to be handled specially. */
2661 if (flags & MSG_OOB)
2662 goto recv_urg;
2663
2664 if (unlikely(tp->repair)) {
2665 err = -EPERM;
2666 if (!(flags & MSG_PEEK))
2667 goto out;
2668
2669 if (tp->repair_queue == TCP_SEND_QUEUE)
2670 goto recv_sndq;
2671
2672 err = -EINVAL;
2673 if (tp->repair_queue == TCP_NO_QUEUE)
2674 goto out;
2675
2676 /* 'common' recv queue MSG_PEEK-ing */
2677 }
2678
2679 seq = &tp->copied_seq;
2680 if (flags & MSG_PEEK) {
2681 peek_offset = max(sk_peek_offset(sk, flags), 0);
2682 peek_seq = tp->copied_seq + peek_offset;
2683 seq = &peek_seq;
2684 }
2685
2686 target = sock_rcvlowat(sk, waitall: flags & MSG_WAITALL, len);
2687
2688 do {
2689 u32 offset;
2690
2691 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
2692 if (unlikely(tp->urg_data) && tp->urg_seq == *seq) {
2693 if (copied)
2694 break;
2695 if (signal_pending(current)) {
2696 copied = timeo ? sock_intr_errno(timeo) : -EAGAIN;
2697 break;
2698 }
2699 }
2700
2701 /* Next get a buffer. */
2702
2703 last = skb_peek_tail(list_: &sk->sk_receive_queue);
2704 skb_queue_walk(&sk->sk_receive_queue, skb) {
2705 last = skb;
2706 /* Now that we have two receive queues this
2707 * shouldn't happen.
2708 */
2709 if (WARN(before(*seq, TCP_SKB_CB(skb)->seq),
2710 "TCP recvmsg seq # bug: copied %X, seq %X, rcvnxt %X, fl %X\n",
2711 *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt,
2712 flags))
2713 break;
2714
2715 offset = *seq - TCP_SKB_CB(skb)->seq;
2716 if (unlikely(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
2717 pr_err_once("%s: found a SYN, please report !\n", __func__);
2718 offset--;
2719 }
2720 if (offset < skb->len)
2721 goto found_ok_skb;
2722 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
2723 goto found_fin_ok;
2724 WARN(!(flags & MSG_PEEK),
2725 "TCP recvmsg seq # bug 2: copied %X, seq %X, rcvnxt %X, fl %X\n",
2726 *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt, flags);
2727 }
2728
2729 /* Well, if we have backlog, try to process it now yet. */
2730
2731 if (copied >= target && !READ_ONCE(sk->sk_backlog.tail))
2732 break;
2733
2734 if (copied) {
2735 if (!timeo ||
2736 sk->sk_err ||
2737 sk->sk_state == TCP_CLOSE ||
2738 (sk->sk_shutdown & RCV_SHUTDOWN) ||
2739 signal_pending(current))
2740 break;
2741 } else {
2742 if (sock_flag(sk, flag: SOCK_DONE))
2743 break;
2744
2745 if (sk->sk_err) {
2746 copied = sock_error(sk);
2747 break;
2748 }
2749
2750 if (sk->sk_shutdown & RCV_SHUTDOWN)
2751 break;
2752
2753 if (sk->sk_state == TCP_CLOSE) {
2754 /* This occurs when user tries to read
2755 * from never connected socket.
2756 */
2757 copied = -ENOTCONN;
2758 break;
2759 }
2760
2761 if (!timeo) {
2762 copied = -EAGAIN;
2763 break;
2764 }
2765
2766 if (signal_pending(current)) {
2767 copied = sock_intr_errno(timeo);
2768 break;
2769 }
2770 }
2771
2772 if (copied >= target) {
2773 /* Do not sleep, just process backlog. */
2774 __sk_flush_backlog(sk);
2775 } else {
2776 tcp_cleanup_rbuf(sk, copied);
2777 err = sk_wait_data(sk, timeo: &timeo, skb: last);
2778 if (err < 0) {
2779 err = copied ? : err;
2780 goto out;
2781 }
2782 }
2783
2784 if ((flags & MSG_PEEK) &&
2785 (peek_seq - peek_offset - copied - urg_hole != tp->copied_seq)) {
2786 net_dbg_ratelimited("TCP(%s:%d): Application bug, race in MSG_PEEK\n",
2787 current->comm,
2788 task_pid_nr(current));
2789 peek_seq = tp->copied_seq + peek_offset;
2790 }
2791 continue;
2792
2793found_ok_skb:
2794 /* Ok so how much can we use? */
2795 used = skb->len - offset;
2796 if (len < used)
2797 used = len;
2798
2799 /* Do we have urgent data here? */
2800 if (unlikely(tp->urg_data)) {
2801 u32 urg_offset = tp->urg_seq - *seq;
2802 if (urg_offset < used) {
2803 if (!urg_offset) {
2804 if (!sock_flag(sk, flag: SOCK_URGINLINE)) {
2805 WRITE_ONCE(*seq, *seq + 1);
2806 urg_hole++;
2807 offset++;
2808 used--;
2809 if (!used)
2810 goto skip_copy;
2811 }
2812 } else
2813 used = urg_offset;
2814 }
2815 }
2816
2817 if (!(flags & MSG_TRUNC)) {
2818 if (last_copied_dmabuf != -1 &&
2819 last_copied_dmabuf != !skb_frags_readable(skb))
2820 break;
2821
2822 if (skb_frags_readable(skb)) {
2823 err = skb_copy_datagram_msg(from: skb, offset, msg,
2824 size: used);
2825 if (err) {
2826 /* Exception. Bailout! */
2827 if (!copied)
2828 copied = -EFAULT;
2829 break;
2830 }
2831 } else {
2832 if (!(flags & MSG_SOCK_DEVMEM)) {
2833 /* dmabuf skbs can only be received
2834 * with the MSG_SOCK_DEVMEM flag.
2835 */
2836 if (!copied)
2837 copied = -EFAULT;
2838
2839 break;
2840 }
2841
2842 err = tcp_recvmsg_dmabuf(sk, skb, offset, msg,
2843 remaining_len: used);
2844 if (err < 0) {
2845 if (!copied)
2846 copied = err;
2847
2848 break;
2849 }
2850 used = err;
2851 }
2852 }
2853
2854 last_copied_dmabuf = !skb_frags_readable(skb);
2855
2856 WRITE_ONCE(*seq, *seq + used);
2857 copied += used;
2858 len -= used;
2859 if (flags & MSG_PEEK)
2860 sk_peek_offset_fwd(sk, val: used);
2861 else
2862 sk_peek_offset_bwd(sk, val: used);
2863 tcp_rcv_space_adjust(sk);
2864
2865skip_copy:
2866 if (unlikely(tp->urg_data) && after(tp->copied_seq, tp->urg_seq)) {
2867 WRITE_ONCE(tp->urg_data, 0);
2868 tcp_fast_path_check(sk);
2869 }
2870
2871 if (TCP_SKB_CB(skb)->has_rxtstamp) {
2872 tcp_update_recv_tstamps(skb, tss);
2873 *cmsg_flags |= TCP_CMSG_TS;
2874 }
2875
2876 if (used + offset < skb->len)
2877 continue;
2878
2879 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
2880 goto found_fin_ok;
2881 if (!(flags & MSG_PEEK))
2882 tcp_eat_recv_skb(sk, skb);
2883 continue;
2884
2885found_fin_ok:
2886 /* Process the FIN. */
2887 WRITE_ONCE(*seq, *seq + 1);
2888 if (!(flags & MSG_PEEK))
2889 tcp_eat_recv_skb(sk, skb);
2890 break;
2891 } while (len > 0);
2892
2893 /* According to UNIX98, msg_name/msg_namelen are ignored
2894 * on connected socket. I was just happy when found this 8) --ANK
2895 */
2896
2897 /* Clean up data we have read: This will do ACK frames. */
2898 tcp_cleanup_rbuf(sk, copied);
2899 return copied;
2900
2901out:
2902 return err;
2903
2904recv_urg:
2905 err = tcp_recv_urg(sk, msg, len, flags);
2906 goto out;
2907
2908recv_sndq:
2909 err = tcp_peek_sndq(sk, msg, len);
2910 goto out;
2911}
2912
2913int tcp_recvmsg(struct sock *sk, struct msghdr *msg, size_t len, int flags,
2914 int *addr_len)
2915{
2916 int cmsg_flags = 0, ret;
2917 struct scm_timestamping_internal tss;
2918
2919 if (unlikely(flags & MSG_ERRQUEUE))
2920 return inet_recv_error(sk, msg, len, addr_len);
2921
2922 if (sk_can_busy_loop(sk) &&
2923 skb_queue_empty_lockless(list: &sk->sk_receive_queue) &&
2924 sk->sk_state == TCP_ESTABLISHED)
2925 sk_busy_loop(sk, nonblock: flags & MSG_DONTWAIT);
2926
2927 lock_sock(sk);
2928 ret = tcp_recvmsg_locked(sk, msg, len, flags, tss: &tss, cmsg_flags: &cmsg_flags);
2929 release_sock(sk);
2930
2931 if ((cmsg_flags || msg->msg_get_inq) && ret >= 0) {
2932 if (cmsg_flags & TCP_CMSG_TS)
2933 tcp_recv_timestamp(msg, sk, tss: &tss);
2934 if (msg->msg_get_inq) {
2935 msg->msg_inq = tcp_inq_hint(sk);
2936 if (cmsg_flags & TCP_CMSG_INQ)
2937 put_cmsg(msg, SOL_TCP, TCP_CM_INQ,
2938 len: sizeof(msg->msg_inq), data: &msg->msg_inq);
2939 }
2940 }
2941 return ret;
2942}
2943EXPORT_IPV6_MOD(tcp_recvmsg);
2944
2945void tcp_set_state(struct sock *sk, int state)
2946{
2947 int oldstate = sk->sk_state;
2948
2949 /* We defined a new enum for TCP states that are exported in BPF
2950 * so as not force the internal TCP states to be frozen. The
2951 * following checks will detect if an internal state value ever
2952 * differs from the BPF value. If this ever happens, then we will
2953 * need to remap the internal value to the BPF value before calling
2954 * tcp_call_bpf_2arg.
2955 */
2956 BUILD_BUG_ON((int)BPF_TCP_ESTABLISHED != (int)TCP_ESTABLISHED);
2957 BUILD_BUG_ON((int)BPF_TCP_SYN_SENT != (int)TCP_SYN_SENT);
2958 BUILD_BUG_ON((int)BPF_TCP_SYN_RECV != (int)TCP_SYN_RECV);
2959 BUILD_BUG_ON((int)BPF_TCP_FIN_WAIT1 != (int)TCP_FIN_WAIT1);
2960 BUILD_BUG_ON((int)BPF_TCP_FIN_WAIT2 != (int)TCP_FIN_WAIT2);
2961 BUILD_BUG_ON((int)BPF_TCP_TIME_WAIT != (int)TCP_TIME_WAIT);
2962 BUILD_BUG_ON((int)BPF_TCP_CLOSE != (int)TCP_CLOSE);
2963 BUILD_BUG_ON((int)BPF_TCP_CLOSE_WAIT != (int)TCP_CLOSE_WAIT);
2964 BUILD_BUG_ON((int)BPF_TCP_LAST_ACK != (int)TCP_LAST_ACK);
2965 BUILD_BUG_ON((int)BPF_TCP_LISTEN != (int)TCP_LISTEN);
2966 BUILD_BUG_ON((int)BPF_TCP_CLOSING != (int)TCP_CLOSING);
2967 BUILD_BUG_ON((int)BPF_TCP_NEW_SYN_RECV != (int)TCP_NEW_SYN_RECV);
2968 BUILD_BUG_ON((int)BPF_TCP_BOUND_INACTIVE != (int)TCP_BOUND_INACTIVE);
2969 BUILD_BUG_ON((int)BPF_TCP_MAX_STATES != (int)TCP_MAX_STATES);
2970
2971 /* bpf uapi header bpf.h defines an anonymous enum with values
2972 * BPF_TCP_* used by bpf programs. Currently gcc built vmlinux
2973 * is able to emit this enum in DWARF due to the above BUILD_BUG_ON.
2974 * But clang built vmlinux does not have this enum in DWARF
2975 * since clang removes the above code before generating IR/debuginfo.
2976 * Let us explicitly emit the type debuginfo to ensure the
2977 * above-mentioned anonymous enum in the vmlinux DWARF and hence BTF
2978 * regardless of which compiler is used.
2979 */
2980 BTF_TYPE_EMIT_ENUM(BPF_TCP_ESTABLISHED);
2981
2982 if (BPF_SOCK_OPS_TEST_FLAG(tcp_sk(sk), BPF_SOCK_OPS_STATE_CB_FLAG))
2983 tcp_call_bpf_2arg(sk, op: BPF_SOCK_OPS_STATE_CB, arg1: oldstate, arg2: state);
2984
2985 switch (state) {
2986 case TCP_ESTABLISHED:
2987 if (oldstate != TCP_ESTABLISHED)
2988 TCP_INC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
2989 break;
2990 case TCP_CLOSE_WAIT:
2991 if (oldstate == TCP_SYN_RECV)
2992 TCP_INC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
2993 break;
2994
2995 case TCP_CLOSE:
2996 if (oldstate == TCP_CLOSE_WAIT || oldstate == TCP_ESTABLISHED)
2997 TCP_INC_STATS(sock_net(sk), TCP_MIB_ESTABRESETS);
2998
2999 sk->sk_prot->unhash(sk);
3000 if (inet_csk(sk)->icsk_bind_hash &&
3001 !(sk->sk_userlocks & SOCK_BINDPORT_LOCK))
3002 inet_put_port(sk);
3003 fallthrough;
3004 default:
3005 if (oldstate == TCP_ESTABLISHED || oldstate == TCP_CLOSE_WAIT)
3006 TCP_DEC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
3007 }
3008
3009 /* Change state AFTER socket is unhashed to avoid closed
3010 * socket sitting in hash tables.
3011 */
3012 inet_sk_state_store(sk, newstate: state);
3013}
3014EXPORT_SYMBOL_GPL(tcp_set_state);
3015
3016/*
3017 * State processing on a close. This implements the state shift for
3018 * sending our FIN frame. Note that we only send a FIN for some
3019 * states. A shutdown() may have already sent the FIN, or we may be
3020 * closed.
3021 */
3022
3023static const unsigned char new_state[16] = {
3024 /* current state: new state: action: */
3025 [0 /* (Invalid) */] = TCP_CLOSE,
3026 [TCP_ESTABLISHED] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
3027 [TCP_SYN_SENT] = TCP_CLOSE,
3028 [TCP_SYN_RECV] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
3029 [TCP_FIN_WAIT1] = TCP_FIN_WAIT1,
3030 [TCP_FIN_WAIT2] = TCP_FIN_WAIT2,
3031 [TCP_TIME_WAIT] = TCP_CLOSE,
3032 [TCP_CLOSE] = TCP_CLOSE,
3033 [TCP_CLOSE_WAIT] = TCP_LAST_ACK | TCP_ACTION_FIN,
3034 [TCP_LAST_ACK] = TCP_LAST_ACK,
3035 [TCP_LISTEN] = TCP_CLOSE,
3036 [TCP_CLOSING] = TCP_CLOSING,
3037 [TCP_NEW_SYN_RECV] = TCP_CLOSE, /* should not happen ! */
3038};
3039
3040static int tcp_close_state(struct sock *sk)
3041{
3042 int next = (int)new_state[sk->sk_state];
3043 int ns = next & TCP_STATE_MASK;
3044
3045 tcp_set_state(sk, ns);
3046
3047 return next & TCP_ACTION_FIN;
3048}
3049
3050/*
3051 * Shutdown the sending side of a connection. Much like close except
3052 * that we don't receive shut down or sock_set_flag(sk, SOCK_DEAD).
3053 */
3054
3055void tcp_shutdown(struct sock *sk, int how)
3056{
3057 /* We need to grab some memory, and put together a FIN,
3058 * and then put it into the queue to be sent.
3059 * Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
3060 */
3061 if (!(how & SEND_SHUTDOWN))
3062 return;
3063
3064 /* If we've already sent a FIN, or it's a closed state, skip this. */
3065 if ((1 << sk->sk_state) &
3066 (TCPF_ESTABLISHED | TCPF_SYN_SENT |
3067 TCPF_CLOSE_WAIT)) {
3068 /* Clear out any half completed packets. FIN if needed. */
3069 if (tcp_close_state(sk))
3070 tcp_send_fin(sk);
3071 }
3072}
3073EXPORT_IPV6_MOD(tcp_shutdown);
3074
3075int tcp_orphan_count_sum(void)
3076{
3077 int i, total = 0;
3078
3079 for_each_possible_cpu(i)
3080 total += per_cpu(tcp_orphan_count, i);
3081
3082 return max(total, 0);
3083}
3084
3085static int tcp_orphan_cache;
3086static struct timer_list tcp_orphan_timer;
3087#define TCP_ORPHAN_TIMER_PERIOD msecs_to_jiffies(100)
3088
3089static void tcp_orphan_update(struct timer_list *unused)
3090{
3091 WRITE_ONCE(tcp_orphan_cache, tcp_orphan_count_sum());
3092 mod_timer(timer: &tcp_orphan_timer, expires: jiffies + TCP_ORPHAN_TIMER_PERIOD);
3093}
3094
3095static bool tcp_too_many_orphans(int shift)
3096{
3097 return READ_ONCE(tcp_orphan_cache) << shift >
3098 READ_ONCE(sysctl_tcp_max_orphans);
3099}
3100
3101static bool tcp_out_of_memory(const struct sock *sk)
3102{
3103 if (sk->sk_wmem_queued > SOCK_MIN_SNDBUF &&
3104 sk_memory_allocated(sk) > sk_prot_mem_limits(sk, index: 2))
3105 return true;
3106 return false;
3107}
3108
3109bool tcp_check_oom(const struct sock *sk, int shift)
3110{
3111 bool too_many_orphans, out_of_socket_memory;
3112
3113 too_many_orphans = tcp_too_many_orphans(shift);
3114 out_of_socket_memory = tcp_out_of_memory(sk);
3115
3116 if (too_many_orphans)
3117 net_info_ratelimited("too many orphaned sockets\n");
3118 if (out_of_socket_memory)
3119 net_info_ratelimited("out of memory -- consider tuning tcp_mem\n");
3120 return too_many_orphans || out_of_socket_memory;
3121}
3122
3123void __tcp_close(struct sock *sk, long timeout)
3124{
3125 bool data_was_unread = false;
3126 struct sk_buff *skb;
3127 int state;
3128
3129 WRITE_ONCE(sk->sk_shutdown, SHUTDOWN_MASK);
3130
3131 if (sk->sk_state == TCP_LISTEN) {
3132 tcp_set_state(sk, TCP_CLOSE);
3133
3134 /* Special case. */
3135 inet_csk_listen_stop(sk);
3136
3137 goto adjudge_to_death;
3138 }
3139
3140 /* We need to flush the recv. buffs. We do this only on the
3141 * descriptor close, not protocol-sourced closes, because the
3142 * reader process may not have drained the data yet!
3143 */
3144 while ((skb = skb_peek(list_: &sk->sk_receive_queue)) != NULL) {
3145 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
3146
3147 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
3148 end_seq--;
3149 if (after(end_seq, tcp_sk(sk)->copied_seq))
3150 data_was_unread = true;
3151 tcp_eat_recv_skb(sk, skb);
3152 }
3153
3154 /* If socket has been already reset (e.g. in tcp_reset()) - kill it. */
3155 if (sk->sk_state == TCP_CLOSE)
3156 goto adjudge_to_death;
3157
3158 /* As outlined in RFC 2525, section 2.17, we send a RST here because
3159 * data was lost. To witness the awful effects of the old behavior of
3160 * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk
3161 * GET in an FTP client, suspend the process, wait for the client to
3162 * advertise a zero window, then kill -9 the FTP client, wheee...
3163 * Note: timeout is always zero in such a case.
3164 */
3165 if (unlikely(tcp_sk(sk)->repair)) {
3166 sk->sk_prot->disconnect(sk, 0);
3167 } else if (data_was_unread) {
3168 /* Unread data was tossed, zap the connection. */
3169 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONCLOSE);
3170 tcp_set_state(sk, TCP_CLOSE);
3171 tcp_send_active_reset(sk, priority: sk->sk_allocation,
3172 reason: SK_RST_REASON_TCP_ABORT_ON_CLOSE);
3173 } else if (sock_flag(sk, flag: SOCK_LINGER) && !sk->sk_lingertime) {
3174 /* Check zero linger _after_ checking for unread data. */
3175 sk->sk_prot->disconnect(sk, 0);
3176 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
3177 } else if (tcp_close_state(sk)) {
3178 /* We FIN if the application ate all the data before
3179 * zapping the connection.
3180 */
3181
3182 /* RED-PEN. Formally speaking, we have broken TCP state
3183 * machine. State transitions:
3184 *
3185 * TCP_ESTABLISHED -> TCP_FIN_WAIT1
3186 * TCP_SYN_RECV -> TCP_FIN_WAIT1 (it is difficult)
3187 * TCP_CLOSE_WAIT -> TCP_LAST_ACK
3188 *
3189 * are legal only when FIN has been sent (i.e. in window),
3190 * rather than queued out of window. Purists blame.
3191 *
3192 * F.e. "RFC state" is ESTABLISHED,
3193 * if Linux state is FIN-WAIT-1, but FIN is still not sent.
3194 *
3195 * The visible declinations are that sometimes
3196 * we enter time-wait state, when it is not required really
3197 * (harmless), do not send active resets, when they are
3198 * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
3199 * they look as CLOSING or LAST_ACK for Linux)
3200 * Probably, I missed some more holelets.
3201 * --ANK
3202 * XXX (TFO) - To start off we don't support SYN+ACK+FIN
3203 * in a single packet! (May consider it later but will
3204 * probably need API support or TCP_CORK SYN-ACK until
3205 * data is written and socket is closed.)
3206 */
3207 tcp_send_fin(sk);
3208 }
3209
3210 sk_stream_wait_close(sk, timeo_p: timeout);
3211
3212adjudge_to_death:
3213 state = sk->sk_state;
3214 sock_hold(sk);
3215 sock_orphan(sk);
3216
3217 local_bh_disable();
3218 bh_lock_sock(sk);
3219 /* remove backlog if any, without releasing ownership. */
3220 __release_sock(sk);
3221
3222 tcp_orphan_count_inc();
3223
3224 /* Have we already been destroyed by a softirq or backlog? */
3225 if (state != TCP_CLOSE && sk->sk_state == TCP_CLOSE)
3226 goto out;
3227
3228 /* This is a (useful) BSD violating of the RFC. There is a
3229 * problem with TCP as specified in that the other end could
3230 * keep a socket open forever with no application left this end.
3231 * We use a 1 minute timeout (about the same as BSD) then kill
3232 * our end. If they send after that then tough - BUT: long enough
3233 * that we won't make the old 4*rto = almost no time - whoops
3234 * reset mistake.
3235 *
3236 * Nope, it was not mistake. It is really desired behaviour
3237 * f.e. on http servers, when such sockets are useless, but
3238 * consume significant resources. Let's do it with special
3239 * linger2 option. --ANK
3240 */
3241
3242 if (sk->sk_state == TCP_FIN_WAIT2) {
3243 struct tcp_sock *tp = tcp_sk(sk);
3244 if (READ_ONCE(tp->linger2) < 0) {
3245 tcp_set_state(sk, TCP_CLOSE);
3246 tcp_send_active_reset(sk, GFP_ATOMIC,
3247 reason: SK_RST_REASON_TCP_ABORT_ON_LINGER);
3248 __NET_INC_STATS(sock_net(sk),
3249 LINUX_MIB_TCPABORTONLINGER);
3250 } else {
3251 const int tmo = tcp_fin_time(sk);
3252
3253 if (tmo > TCP_TIMEWAIT_LEN) {
3254 tcp_reset_keepalive_timer(sk,
3255 timeout: tmo - TCP_TIMEWAIT_LEN);
3256 } else {
3257 tcp_time_wait(sk, state: TCP_FIN_WAIT2, timeo: tmo);
3258 goto out;
3259 }
3260 }
3261 }
3262 if (sk->sk_state != TCP_CLOSE) {
3263 if (tcp_check_oom(sk, shift: 0)) {
3264 tcp_set_state(sk, TCP_CLOSE);
3265 tcp_send_active_reset(sk, GFP_ATOMIC,
3266 reason: SK_RST_REASON_TCP_ABORT_ON_MEMORY);
3267 __NET_INC_STATS(sock_net(sk),
3268 LINUX_MIB_TCPABORTONMEMORY);
3269 } else if (!check_net(net: sock_net(sk))) {
3270 /* Not possible to send reset; just close */
3271 tcp_set_state(sk, TCP_CLOSE);
3272 }
3273 }
3274
3275 if (sk->sk_state == TCP_CLOSE) {
3276 struct request_sock *req;
3277
3278 req = rcu_dereference_protected(tcp_sk(sk)->fastopen_rsk,
3279 lockdep_sock_is_held(sk));
3280 /* We could get here with a non-NULL req if the socket is
3281 * aborted (e.g., closed with unread data) before 3WHS
3282 * finishes.
3283 */
3284 if (req)
3285 reqsk_fastopen_remove(sk, req, reset: false);
3286 inet_csk_destroy_sock(sk);
3287 }
3288 /* Otherwise, socket is reprieved until protocol close. */
3289
3290out:
3291 bh_unlock_sock(sk);
3292 local_bh_enable();
3293}
3294
3295void tcp_close(struct sock *sk, long timeout)
3296{
3297 lock_sock(sk);
3298 __tcp_close(sk, timeout);
3299 release_sock(sk);
3300 if (!sk->sk_net_refcnt)
3301 inet_csk_clear_xmit_timers_sync(sk);
3302 sock_put(sk);
3303}
3304EXPORT_SYMBOL(tcp_close);
3305
3306/* These states need RST on ABORT according to RFC793 */
3307
3308static inline bool tcp_need_reset(int state)
3309{
3310 return (1 << state) &
3311 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 |
3312 TCPF_FIN_WAIT2 | TCPF_SYN_RECV);
3313}
3314
3315static void tcp_rtx_queue_purge(struct sock *sk)
3316{
3317 struct rb_node *p = rb_first(&sk->tcp_rtx_queue);
3318
3319 tcp_sk(sk)->highest_sack = NULL;
3320 while (p) {
3321 struct sk_buff *skb = rb_to_skb(p);
3322
3323 p = rb_next(p);
3324 /* Since we are deleting whole queue, no need to
3325 * list_del(&skb->tcp_tsorted_anchor)
3326 */
3327 tcp_rtx_queue_unlink(skb, sk);
3328 tcp_wmem_free_skb(sk, skb);
3329 }
3330}
3331
3332void tcp_write_queue_purge(struct sock *sk)
3333{
3334 struct sk_buff *skb;
3335
3336 tcp_chrono_stop(sk, type: TCP_CHRONO_BUSY);
3337 while ((skb = __skb_dequeue(list: &sk->sk_write_queue)) != NULL) {
3338 tcp_skb_tsorted_anchor_cleanup(skb);
3339 tcp_wmem_free_skb(sk, skb);
3340 }
3341 tcp_rtx_queue_purge(sk);
3342 INIT_LIST_HEAD(list: &tcp_sk(sk)->tsorted_sent_queue);
3343 tcp_clear_all_retrans_hints(tcp_sk(sk));
3344 tcp_sk(sk)->packets_out = 0;
3345 inet_csk(sk)->icsk_backoff = 0;
3346}
3347
3348int tcp_disconnect(struct sock *sk, int flags)
3349{
3350 struct inet_sock *inet = inet_sk(sk);
3351 struct inet_connection_sock *icsk = inet_csk(sk);
3352 struct tcp_sock *tp = tcp_sk(sk);
3353 int old_state = sk->sk_state;
3354 struct request_sock *req;
3355 u32 seq;
3356
3357 if (old_state != TCP_CLOSE)
3358 tcp_set_state(sk, TCP_CLOSE);
3359
3360 /* ABORT function of RFC793 */
3361 if (old_state == TCP_LISTEN) {
3362 inet_csk_listen_stop(sk);
3363 } else if (unlikely(tp->repair)) {
3364 WRITE_ONCE(sk->sk_err, ECONNABORTED);
3365 } else if (tcp_need_reset(state: old_state)) {
3366 tcp_send_active_reset(sk, priority: gfp_any(), reason: SK_RST_REASON_TCP_STATE);
3367 WRITE_ONCE(sk->sk_err, ECONNRESET);
3368 } else if (tp->snd_nxt != tp->write_seq &&
3369 (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK)) {
3370 /* The last check adjusts for discrepancy of Linux wrt. RFC
3371 * states
3372 */
3373 tcp_send_active_reset(sk, priority: gfp_any(),
3374 reason: SK_RST_REASON_TCP_DISCONNECT_WITH_DATA);
3375 WRITE_ONCE(sk->sk_err, ECONNRESET);
3376 } else if (old_state == TCP_SYN_SENT)
3377 WRITE_ONCE(sk->sk_err, ECONNRESET);
3378
3379 tcp_clear_xmit_timers(sk);
3380 __skb_queue_purge(list: &sk->sk_receive_queue);
3381 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt);
3382 WRITE_ONCE(tp->urg_data, 0);
3383 sk_set_peek_off(sk, val: -1);
3384 tcp_write_queue_purge(sk);
3385 tcp_fastopen_active_disable_ofo_check(sk);
3386 skb_rbtree_purge(root: &tp->out_of_order_queue);
3387
3388 inet->inet_dport = 0;
3389
3390 inet_bhash2_reset_saddr(sk);
3391
3392 WRITE_ONCE(sk->sk_shutdown, 0);
3393 sock_reset_flag(sk, flag: SOCK_DONE);
3394 tp->srtt_us = 0;
3395 tp->mdev_us = jiffies_to_usecs(TCP_TIMEOUT_INIT);
3396 tp->rcv_rtt_last_tsecr = 0;
3397
3398 seq = tp->write_seq + tp->max_window + 2;
3399 if (!seq)
3400 seq = 1;
3401 WRITE_ONCE(tp->write_seq, seq);
3402
3403 icsk->icsk_backoff = 0;
3404 WRITE_ONCE(icsk->icsk_probes_out, 0);
3405 icsk->icsk_probes_tstamp = 0;
3406 icsk->icsk_rto = TCP_TIMEOUT_INIT;
3407 WRITE_ONCE(icsk->icsk_rto_min, TCP_RTO_MIN);
3408 WRITE_ONCE(icsk->icsk_delack_max, TCP_DELACK_MAX);
3409 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
3410 tcp_snd_cwnd_set(tp, TCP_INIT_CWND);
3411 tp->snd_cwnd_cnt = 0;
3412 tp->is_cwnd_limited = 0;
3413 tp->max_packets_out = 0;
3414 tp->window_clamp = 0;
3415 tp->delivered = 0;
3416 tp->delivered_ce = 0;
3417 tp->accecn_fail_mode = 0;
3418 tp->saw_accecn_opt = TCP_ACCECN_OPT_NOT_SEEN;
3419 tcp_accecn_init_counters(tp);
3420 tp->prev_ecnfield = 0;
3421 tp->accecn_opt_tstamp = 0;
3422 if (icsk->icsk_ca_initialized && icsk->icsk_ca_ops->release)
3423 icsk->icsk_ca_ops->release(sk);
3424 memset(s: icsk->icsk_ca_priv, c: 0, n: sizeof(icsk->icsk_ca_priv));
3425 icsk->icsk_ca_initialized = 0;
3426 tcp_set_ca_state(sk, ca_state: TCP_CA_Open);
3427 tp->is_sack_reneg = 0;
3428 tcp_clear_retrans(tp);
3429 tp->total_retrans = 0;
3430 inet_csk_delack_init(sk);
3431 /* Initialize rcv_mss to TCP_MIN_MSS to avoid division by 0
3432 * issue in __tcp_select_window()
3433 */
3434 icsk->icsk_ack.rcv_mss = TCP_MIN_MSS;
3435 memset(s: &tp->rx_opt, c: 0, n: sizeof(tp->rx_opt));
3436 __sk_dst_reset(sk);
3437 dst_release(unrcu_pointer(xchg(&sk->sk_rx_dst, NULL)));
3438 tcp_saved_syn_free(tp);
3439 tp->compressed_ack = 0;
3440 tp->segs_in = 0;
3441 tp->segs_out = 0;
3442 tp->bytes_sent = 0;
3443 tp->bytes_acked = 0;
3444 tp->bytes_received = 0;
3445 tp->bytes_retrans = 0;
3446 tp->data_segs_in = 0;
3447 tp->data_segs_out = 0;
3448 tp->duplicate_sack[0].start_seq = 0;
3449 tp->duplicate_sack[0].end_seq = 0;
3450 tp->dsack_dups = 0;
3451 tp->reord_seen = 0;
3452 tp->retrans_out = 0;
3453 tp->sacked_out = 0;
3454 tp->tlp_high_seq = 0;
3455 tp->last_oow_ack_time = 0;
3456 tp->plb_rehash = 0;
3457 /* There's a bubble in the pipe until at least the first ACK. */
3458 tp->app_limited = ~0U;
3459 tp->rate_app_limited = 1;
3460 tp->rack.mstamp = 0;
3461 tp->rack.advanced = 0;
3462 tp->rack.reo_wnd_steps = 1;
3463 tp->rack.last_delivered = 0;
3464 tp->rack.reo_wnd_persist = 0;
3465 tp->rack.dsack_seen = 0;
3466 tp->syn_data_acked = 0;
3467 tp->syn_fastopen_child = 0;
3468 tp->rx_opt.saw_tstamp = 0;
3469 tp->rx_opt.dsack = 0;
3470 tp->rx_opt.num_sacks = 0;
3471 tp->rcv_ooopack = 0;
3472
3473
3474 /* Clean up fastopen related fields */
3475 req = rcu_dereference_protected(tp->fastopen_rsk,
3476 lockdep_sock_is_held(sk));
3477 if (req)
3478 reqsk_fastopen_remove(sk, req, reset: false);
3479 tcp_free_fastopen_req(tp);
3480 inet_clear_bit(DEFER_CONNECT, sk);
3481 tp->fastopen_client_fail = 0;
3482
3483 WARN_ON(inet->inet_num && !icsk->icsk_bind_hash);
3484
3485 if (sk->sk_frag.page) {
3486 put_page(page: sk->sk_frag.page);
3487 sk->sk_frag.page = NULL;
3488 sk->sk_frag.offset = 0;
3489 }
3490 sk_error_report(sk);
3491 return 0;
3492}
3493EXPORT_SYMBOL(tcp_disconnect);
3494
3495static inline bool tcp_can_repair_sock(const struct sock *sk)
3496{
3497 return sockopt_ns_capable(ns: sock_net(sk)->user_ns, CAP_NET_ADMIN) &&
3498 (sk->sk_state != TCP_LISTEN);
3499}
3500
3501static int tcp_repair_set_window(struct tcp_sock *tp, sockptr_t optbuf, int len)
3502{
3503 struct tcp_repair_window opt;
3504
3505 if (!tp->repair)
3506 return -EPERM;
3507
3508 if (len != sizeof(opt))
3509 return -EINVAL;
3510
3511 if (copy_from_sockptr(dst: &opt, src: optbuf, size: sizeof(opt)))
3512 return -EFAULT;
3513
3514 if (opt.max_window < opt.snd_wnd)
3515 return -EINVAL;
3516
3517 if (after(opt.snd_wl1, tp->rcv_nxt + opt.rcv_wnd))
3518 return -EINVAL;
3519
3520 if (after(opt.rcv_wup, tp->rcv_nxt))
3521 return -EINVAL;
3522
3523 tp->snd_wl1 = opt.snd_wl1;
3524 tp->snd_wnd = opt.snd_wnd;
3525 tp->max_window = opt.max_window;
3526
3527 tp->rcv_wnd = opt.rcv_wnd;
3528 tp->rcv_wup = opt.rcv_wup;
3529
3530 return 0;
3531}
3532
3533static int tcp_repair_options_est(struct sock *sk, sockptr_t optbuf,
3534 unsigned int len)
3535{
3536 struct tcp_sock *tp = tcp_sk(sk);
3537 struct tcp_repair_opt opt;
3538 size_t offset = 0;
3539
3540 while (len >= sizeof(opt)) {
3541 if (copy_from_sockptr_offset(dst: &opt, src: optbuf, offset, size: sizeof(opt)))
3542 return -EFAULT;
3543
3544 offset += sizeof(opt);
3545 len -= sizeof(opt);
3546
3547 switch (opt.opt_code) {
3548 case TCPOPT_MSS:
3549 tp->rx_opt.mss_clamp = opt.opt_val;
3550 tcp_mtup_init(sk);
3551 break;
3552 case TCPOPT_WINDOW:
3553 {
3554 u16 snd_wscale = opt.opt_val & 0xFFFF;
3555 u16 rcv_wscale = opt.opt_val >> 16;
3556
3557 if (snd_wscale > TCP_MAX_WSCALE || rcv_wscale > TCP_MAX_WSCALE)
3558 return -EFBIG;
3559
3560 tp->rx_opt.snd_wscale = snd_wscale;
3561 tp->rx_opt.rcv_wscale = rcv_wscale;
3562 tp->rx_opt.wscale_ok = 1;
3563 }
3564 break;
3565 case TCPOPT_SACK_PERM:
3566 if (opt.opt_val != 0)
3567 return -EINVAL;
3568
3569 tp->rx_opt.sack_ok |= TCP_SACK_SEEN;
3570 break;
3571 case TCPOPT_TIMESTAMP:
3572 if (opt.opt_val != 0)
3573 return -EINVAL;
3574
3575 tp->rx_opt.tstamp_ok = 1;
3576 break;
3577 }
3578 }
3579
3580 return 0;
3581}
3582
3583DEFINE_STATIC_KEY_FALSE(tcp_tx_delay_enabled);
3584EXPORT_IPV6_MOD(tcp_tx_delay_enabled);
3585
3586static void tcp_enable_tx_delay(void)
3587{
3588 if (!static_branch_unlikely(&tcp_tx_delay_enabled)) {
3589 static int __tcp_tx_delay_enabled = 0;
3590
3591 if (cmpxchg(&__tcp_tx_delay_enabled, 0, 1) == 0) {
3592 static_branch_enable(&tcp_tx_delay_enabled);
3593 pr_info("TCP_TX_DELAY enabled\n");
3594 }
3595 }
3596}
3597
3598/* When set indicates to always queue non-full frames. Later the user clears
3599 * this option and we transmit any pending partial frames in the queue. This is
3600 * meant to be used alongside sendfile() to get properly filled frames when the
3601 * user (for example) must write out headers with a write() call first and then
3602 * use sendfile to send out the data parts.
3603 *
3604 * TCP_CORK can be set together with TCP_NODELAY and it is stronger than
3605 * TCP_NODELAY.
3606 */
3607void __tcp_sock_set_cork(struct sock *sk, bool on)
3608{
3609 struct tcp_sock *tp = tcp_sk(sk);
3610
3611 if (on) {
3612 tp->nonagle |= TCP_NAGLE_CORK;
3613 } else {
3614 tp->nonagle &= ~TCP_NAGLE_CORK;
3615 if (tp->nonagle & TCP_NAGLE_OFF)
3616 tp->nonagle |= TCP_NAGLE_PUSH;
3617 tcp_push_pending_frames(sk);
3618 }
3619}
3620
3621void tcp_sock_set_cork(struct sock *sk, bool on)
3622{
3623 lock_sock(sk);
3624 __tcp_sock_set_cork(sk, on);
3625 release_sock(sk);
3626}
3627EXPORT_SYMBOL(tcp_sock_set_cork);
3628
3629/* TCP_NODELAY is weaker than TCP_CORK, so that this option on corked socket is
3630 * remembered, but it is not activated until cork is cleared.
3631 *
3632 * However, when TCP_NODELAY is set we make an explicit push, which overrides
3633 * even TCP_CORK for currently queued segments.
3634 */
3635void __tcp_sock_set_nodelay(struct sock *sk, bool on)
3636{
3637 if (on) {
3638 tcp_sk(sk)->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH;
3639 tcp_push_pending_frames(sk);
3640 } else {
3641 tcp_sk(sk)->nonagle &= ~TCP_NAGLE_OFF;
3642 }
3643}
3644
3645void tcp_sock_set_nodelay(struct sock *sk)
3646{
3647 lock_sock(sk);
3648 __tcp_sock_set_nodelay(sk, on: true);
3649 release_sock(sk);
3650}
3651EXPORT_SYMBOL(tcp_sock_set_nodelay);
3652
3653static void __tcp_sock_set_quickack(struct sock *sk, int val)
3654{
3655 if (!val) {
3656 inet_csk_enter_pingpong_mode(sk);
3657 return;
3658 }
3659
3660 inet_csk_exit_pingpong_mode(sk);
3661 if ((1 << sk->sk_state) & (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) &&
3662 inet_csk_ack_scheduled(sk)) {
3663 inet_csk(sk)->icsk_ack.pending |= ICSK_ACK_PUSHED;
3664 tcp_cleanup_rbuf(sk, copied: 1);
3665 if (!(val & 1))
3666 inet_csk_enter_pingpong_mode(sk);
3667 }
3668}
3669
3670void tcp_sock_set_quickack(struct sock *sk, int val)
3671{
3672 lock_sock(sk);
3673 __tcp_sock_set_quickack(sk, val);
3674 release_sock(sk);
3675}
3676EXPORT_SYMBOL(tcp_sock_set_quickack);
3677
3678int tcp_sock_set_syncnt(struct sock *sk, int val)
3679{
3680 if (val < 1 || val > MAX_TCP_SYNCNT)
3681 return -EINVAL;
3682
3683 WRITE_ONCE(inet_csk(sk)->icsk_syn_retries, val);
3684 return 0;
3685}
3686EXPORT_SYMBOL(tcp_sock_set_syncnt);
3687
3688int tcp_sock_set_user_timeout(struct sock *sk, int val)
3689{
3690 /* Cap the max time in ms TCP will retry or probe the window
3691 * before giving up and aborting (ETIMEDOUT) a connection.
3692 */
3693 if (val < 0)
3694 return -EINVAL;
3695
3696 WRITE_ONCE(inet_csk(sk)->icsk_user_timeout, val);
3697 return 0;
3698}
3699EXPORT_SYMBOL(tcp_sock_set_user_timeout);
3700
3701int tcp_sock_set_keepidle_locked(struct sock *sk, int val)
3702{
3703 struct tcp_sock *tp = tcp_sk(sk);
3704
3705 if (val < 1 || val > MAX_TCP_KEEPIDLE)
3706 return -EINVAL;
3707
3708 /* Paired with WRITE_ONCE() in keepalive_time_when() */
3709 WRITE_ONCE(tp->keepalive_time, val * HZ);
3710 if (sock_flag(sk, flag: SOCK_KEEPOPEN) &&
3711 !((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN))) {
3712 u32 elapsed = keepalive_time_elapsed(tp);
3713
3714 if (tp->keepalive_time > elapsed)
3715 elapsed = tp->keepalive_time - elapsed;
3716 else
3717 elapsed = 0;
3718 tcp_reset_keepalive_timer(sk, timeout: elapsed);
3719 }
3720
3721 return 0;
3722}
3723
3724int tcp_sock_set_keepidle(struct sock *sk, int val)
3725{
3726 int err;
3727
3728 lock_sock(sk);
3729 err = tcp_sock_set_keepidle_locked(sk, val);
3730 release_sock(sk);
3731 return err;
3732}
3733EXPORT_SYMBOL(tcp_sock_set_keepidle);
3734
3735int tcp_sock_set_keepintvl(struct sock *sk, int val)
3736{
3737 if (val < 1 || val > MAX_TCP_KEEPINTVL)
3738 return -EINVAL;
3739
3740 WRITE_ONCE(tcp_sk(sk)->keepalive_intvl, val * HZ);
3741 return 0;
3742}
3743EXPORT_SYMBOL(tcp_sock_set_keepintvl);
3744
3745int tcp_sock_set_keepcnt(struct sock *sk, int val)
3746{
3747 if (val < 1 || val > MAX_TCP_KEEPCNT)
3748 return -EINVAL;
3749
3750 /* Paired with READ_ONCE() in keepalive_probes() */
3751 WRITE_ONCE(tcp_sk(sk)->keepalive_probes, val);
3752 return 0;
3753}
3754EXPORT_SYMBOL(tcp_sock_set_keepcnt);
3755
3756int tcp_set_window_clamp(struct sock *sk, int val)
3757{
3758 u32 old_window_clamp, new_window_clamp, new_rcv_ssthresh;
3759 struct tcp_sock *tp = tcp_sk(sk);
3760
3761 if (!val) {
3762 if (sk->sk_state != TCP_CLOSE)
3763 return -EINVAL;
3764 WRITE_ONCE(tp->window_clamp, 0);
3765 return 0;
3766 }
3767
3768 old_window_clamp = tp->window_clamp;
3769 new_window_clamp = max_t(int, SOCK_MIN_RCVBUF / 2, val);
3770
3771 if (new_window_clamp == old_window_clamp)
3772 return 0;
3773
3774 WRITE_ONCE(tp->window_clamp, new_window_clamp);
3775
3776 /* Need to apply the reserved mem provisioning only
3777 * when shrinking the window clamp.
3778 */
3779 if (new_window_clamp < old_window_clamp) {
3780 __tcp_adjust_rcv_ssthresh(sk, new_ssthresh: new_window_clamp);
3781 } else {
3782 new_rcv_ssthresh = min(tp->rcv_wnd, new_window_clamp);
3783 tp->rcv_ssthresh = max(new_rcv_ssthresh, tp->rcv_ssthresh);
3784 }
3785 return 0;
3786}
3787
3788int tcp_sock_set_maxseg(struct sock *sk, int val)
3789{
3790 /* Values greater than interface MTU won't take effect. However
3791 * at the point when this call is done we typically don't yet
3792 * know which interface is going to be used
3793 */
3794 if (val && (val < TCP_MIN_MSS || val > MAX_TCP_WINDOW))
3795 return -EINVAL;
3796
3797 WRITE_ONCE(tcp_sk(sk)->rx_opt.user_mss, val);
3798 return 0;
3799}
3800
3801/*
3802 * Socket option code for TCP.
3803 */
3804int do_tcp_setsockopt(struct sock *sk, int level, int optname,
3805 sockptr_t optval, unsigned int optlen)
3806{
3807 struct tcp_sock *tp = tcp_sk(sk);
3808 struct inet_connection_sock *icsk = inet_csk(sk);
3809 struct net *net = sock_net(sk);
3810 int val;
3811 int err = 0;
3812
3813 /* These are data/string values, all the others are ints */
3814 switch (optname) {
3815 case TCP_CONGESTION: {
3816 char name[TCP_CA_NAME_MAX];
3817
3818 if (optlen < 1)
3819 return -EINVAL;
3820
3821 val = strncpy_from_sockptr(dst: name, src: optval,
3822 min_t(long, TCP_CA_NAME_MAX-1, optlen));
3823 if (val < 0)
3824 return -EFAULT;
3825 name[val] = 0;
3826
3827 sockopt_lock_sock(sk);
3828 err = tcp_set_congestion_control(sk, name, load: !has_current_bpf_ctx(),
3829 cap_net_admin: sockopt_ns_capable(ns: sock_net(sk)->user_ns,
3830 CAP_NET_ADMIN));
3831 sockopt_release_sock(sk);
3832 return err;
3833 }
3834 case TCP_ULP: {
3835 char name[TCP_ULP_NAME_MAX];
3836
3837 if (optlen < 1)
3838 return -EINVAL;
3839
3840 val = strncpy_from_sockptr(dst: name, src: optval,
3841 min_t(long, TCP_ULP_NAME_MAX - 1,
3842 optlen));
3843 if (val < 0)
3844 return -EFAULT;
3845 name[val] = 0;
3846
3847 sockopt_lock_sock(sk);
3848 err = tcp_set_ulp(sk, name);
3849 sockopt_release_sock(sk);
3850 return err;
3851 }
3852 case TCP_FASTOPEN_KEY: {
3853 __u8 key[TCP_FASTOPEN_KEY_BUF_LENGTH];
3854 __u8 *backup_key = NULL;
3855
3856 /* Allow a backup key as well to facilitate key rotation
3857 * First key is the active one.
3858 */
3859 if (optlen != TCP_FASTOPEN_KEY_LENGTH &&
3860 optlen != TCP_FASTOPEN_KEY_BUF_LENGTH)
3861 return -EINVAL;
3862
3863 if (copy_from_sockptr(dst: key, src: optval, size: optlen))
3864 return -EFAULT;
3865
3866 if (optlen == TCP_FASTOPEN_KEY_BUF_LENGTH)
3867 backup_key = key + TCP_FASTOPEN_KEY_LENGTH;
3868
3869 return tcp_fastopen_reset_cipher(net, sk, primary_key: key, backup_key);
3870 }
3871 default:
3872 /* fallthru */
3873 break;
3874 }
3875
3876 if (optlen < sizeof(int))
3877 return -EINVAL;
3878
3879 if (copy_from_sockptr(dst: &val, src: optval, size: sizeof(val)))
3880 return -EFAULT;
3881
3882 /* Handle options that can be set without locking the socket. */
3883 switch (optname) {
3884 case TCP_SYNCNT:
3885 return tcp_sock_set_syncnt(sk, val);
3886 case TCP_USER_TIMEOUT:
3887 return tcp_sock_set_user_timeout(sk, val);
3888 case TCP_KEEPINTVL:
3889 return tcp_sock_set_keepintvl(sk, val);
3890 case TCP_KEEPCNT:
3891 return tcp_sock_set_keepcnt(sk, val);
3892 case TCP_LINGER2:
3893 if (val < 0)
3894 WRITE_ONCE(tp->linger2, -1);
3895 else if (val > TCP_FIN_TIMEOUT_MAX / HZ)
3896 WRITE_ONCE(tp->linger2, TCP_FIN_TIMEOUT_MAX);
3897 else
3898 WRITE_ONCE(tp->linger2, val * HZ);
3899 return 0;
3900 case TCP_DEFER_ACCEPT:
3901 /* Translate value in seconds to number of retransmits */
3902 WRITE_ONCE(icsk->icsk_accept_queue.rskq_defer_accept,
3903 secs_to_retrans(val, TCP_TIMEOUT_INIT / HZ,
3904 TCP_RTO_MAX / HZ));
3905 return 0;
3906 case TCP_RTO_MAX_MS:
3907 if (val < MSEC_PER_SEC || val > TCP_RTO_MAX_SEC * MSEC_PER_SEC)
3908 return -EINVAL;
3909 WRITE_ONCE(inet_csk(sk)->icsk_rto_max, msecs_to_jiffies(val));
3910 return 0;
3911 case TCP_RTO_MIN_US: {
3912 int rto_min = usecs_to_jiffies(u: val);
3913
3914 if (rto_min > TCP_RTO_MIN || rto_min < TCP_TIMEOUT_MIN)
3915 return -EINVAL;
3916 WRITE_ONCE(inet_csk(sk)->icsk_rto_min, rto_min);
3917 return 0;
3918 }
3919 case TCP_DELACK_MAX_US: {
3920 int delack_max = usecs_to_jiffies(u: val);
3921
3922 if (delack_max > TCP_DELACK_MAX || delack_max < TCP_TIMEOUT_MIN)
3923 return -EINVAL;
3924 WRITE_ONCE(inet_csk(sk)->icsk_delack_max, delack_max);
3925 return 0;
3926 }
3927 case TCP_MAXSEG:
3928 return tcp_sock_set_maxseg(sk, val);
3929 }
3930
3931 sockopt_lock_sock(sk);
3932
3933 switch (optname) {
3934 case TCP_NODELAY:
3935 __tcp_sock_set_nodelay(sk, on: val);
3936 break;
3937
3938 case TCP_THIN_LINEAR_TIMEOUTS:
3939 if (val < 0 || val > 1)
3940 err = -EINVAL;
3941 else
3942 tp->thin_lto = val;
3943 break;
3944
3945 case TCP_THIN_DUPACK:
3946 if (val < 0 || val > 1)
3947 err = -EINVAL;
3948 break;
3949
3950 case TCP_REPAIR:
3951 if (!tcp_can_repair_sock(sk))
3952 err = -EPERM;
3953 else if (val == TCP_REPAIR_ON) {
3954 tp->repair = 1;
3955 sk->sk_reuse = SK_FORCE_REUSE;
3956 tp->repair_queue = TCP_NO_QUEUE;
3957 } else if (val == TCP_REPAIR_OFF) {
3958 tp->repair = 0;
3959 sk->sk_reuse = SK_NO_REUSE;
3960 tcp_send_window_probe(sk);
3961 } else if (val == TCP_REPAIR_OFF_NO_WP) {
3962 tp->repair = 0;
3963 sk->sk_reuse = SK_NO_REUSE;
3964 } else
3965 err = -EINVAL;
3966
3967 break;
3968
3969 case TCP_REPAIR_QUEUE:
3970 if (!tp->repair)
3971 err = -EPERM;
3972 else if ((unsigned int)val < TCP_QUEUES_NR)
3973 tp->repair_queue = val;
3974 else
3975 err = -EINVAL;
3976 break;
3977
3978 case TCP_QUEUE_SEQ:
3979 if (sk->sk_state != TCP_CLOSE) {
3980 err = -EPERM;
3981 } else if (tp->repair_queue == TCP_SEND_QUEUE) {
3982 if (!tcp_rtx_queue_empty(sk))
3983 err = -EPERM;
3984 else
3985 WRITE_ONCE(tp->write_seq, val);
3986 } else if (tp->repair_queue == TCP_RECV_QUEUE) {
3987 if (tp->rcv_nxt != tp->copied_seq) {
3988 err = -EPERM;
3989 } else {
3990 WRITE_ONCE(tp->rcv_nxt, val);
3991 WRITE_ONCE(tp->copied_seq, val);
3992 }
3993 } else {
3994 err = -EINVAL;
3995 }
3996 break;
3997
3998 case TCP_REPAIR_OPTIONS:
3999 if (!tp->repair)
4000 err = -EINVAL;
4001 else if (sk->sk_state == TCP_ESTABLISHED && !tp->bytes_sent)
4002 err = tcp_repair_options_est(sk, optbuf: optval, len: optlen);
4003 else
4004 err = -EPERM;
4005 break;
4006
4007 case TCP_CORK:
4008 __tcp_sock_set_cork(sk, on: val);
4009 break;
4010
4011 case TCP_KEEPIDLE:
4012 err = tcp_sock_set_keepidle_locked(sk, val);
4013 break;
4014 case TCP_SAVE_SYN:
4015 /* 0: disable, 1: enable, 2: start from ether_header */
4016 if (val < 0 || val > 2)
4017 err = -EINVAL;
4018 else
4019 tp->save_syn = val;
4020 break;
4021
4022 case TCP_WINDOW_CLAMP:
4023 err = tcp_set_window_clamp(sk, val);
4024 break;
4025
4026 case TCP_QUICKACK:
4027 __tcp_sock_set_quickack(sk, val);
4028 break;
4029
4030 case TCP_AO_REPAIR:
4031 if (!tcp_can_repair_sock(sk)) {
4032 err = -EPERM;
4033 break;
4034 }
4035 err = tcp_ao_set_repair(sk, optval, optlen);
4036 break;
4037#ifdef CONFIG_TCP_AO
4038 case TCP_AO_ADD_KEY:
4039 case TCP_AO_DEL_KEY:
4040 case TCP_AO_INFO: {
4041 /* If this is the first TCP-AO setsockopt() on the socket,
4042 * sk_state has to be LISTEN or CLOSE. Allow TCP_REPAIR
4043 * in any state.
4044 */
4045 if ((1 << sk->sk_state) & (TCPF_LISTEN | TCPF_CLOSE))
4046 goto ao_parse;
4047 if (rcu_dereference_protected(tcp_sk(sk)->ao_info,
4048 lockdep_sock_is_held(sk)))
4049 goto ao_parse;
4050 if (tp->repair)
4051 goto ao_parse;
4052 err = -EISCONN;
4053 break;
4054ao_parse:
4055 err = tp->af_specific->ao_parse(sk, optname, optval, optlen);
4056 break;
4057 }
4058#endif
4059#ifdef CONFIG_TCP_MD5SIG
4060 case TCP_MD5SIG:
4061 case TCP_MD5SIG_EXT:
4062 err = tp->af_specific->md5_parse(sk, optname, optval, optlen);
4063 break;
4064#endif
4065 case TCP_FASTOPEN:
4066 if (val >= 0 && ((1 << sk->sk_state) & (TCPF_CLOSE |
4067 TCPF_LISTEN))) {
4068 tcp_fastopen_init_key_once(net);
4069
4070 fastopen_queue_tune(sk, backlog: val);
4071 } else {
4072 err = -EINVAL;
4073 }
4074 break;
4075 case TCP_FASTOPEN_CONNECT:
4076 if (val > 1 || val < 0) {
4077 err = -EINVAL;
4078 } else if (READ_ONCE(net->ipv4.sysctl_tcp_fastopen) &
4079 TFO_CLIENT_ENABLE) {
4080 if (sk->sk_state == TCP_CLOSE)
4081 tp->fastopen_connect = val;
4082 else
4083 err = -EINVAL;
4084 } else {
4085 err = -EOPNOTSUPP;
4086 }
4087 break;
4088 case TCP_FASTOPEN_NO_COOKIE:
4089 if (val > 1 || val < 0)
4090 err = -EINVAL;
4091 else if (!((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
4092 err = -EINVAL;
4093 else
4094 tp->fastopen_no_cookie = val;
4095 break;
4096 case TCP_TIMESTAMP:
4097 if (!tp->repair) {
4098 err = -EPERM;
4099 break;
4100 }
4101 /* val is an opaque field,
4102 * and low order bit contains usec_ts enable bit.
4103 * Its a best effort, and we do not care if user makes an error.
4104 */
4105 tp->tcp_usec_ts = val & 1;
4106 WRITE_ONCE(tp->tsoffset, val - tcp_clock_ts(tp->tcp_usec_ts));
4107 break;
4108 case TCP_REPAIR_WINDOW:
4109 err = tcp_repair_set_window(tp, optbuf: optval, len: optlen);
4110 break;
4111 case TCP_NOTSENT_LOWAT:
4112 WRITE_ONCE(tp->notsent_lowat, val);
4113 sk->sk_write_space(sk);
4114 break;
4115 case TCP_INQ:
4116 if (val > 1 || val < 0)
4117 err = -EINVAL;
4118 else
4119 tp->recvmsg_inq = val;
4120 break;
4121 case TCP_TX_DELAY:
4122 if (val)
4123 tcp_enable_tx_delay();
4124 WRITE_ONCE(tp->tcp_tx_delay, val);
4125 break;
4126 default:
4127 err = -ENOPROTOOPT;
4128 break;
4129 }
4130
4131 sockopt_release_sock(sk);
4132 return err;
4133}
4134
4135int tcp_setsockopt(struct sock *sk, int level, int optname, sockptr_t optval,
4136 unsigned int optlen)
4137{
4138 const struct inet_connection_sock *icsk = inet_csk(sk);
4139
4140 if (level != SOL_TCP)
4141 /* Paired with WRITE_ONCE() in do_ipv6_setsockopt() and tcp_v6_connect() */
4142 return READ_ONCE(icsk->icsk_af_ops)->setsockopt(sk, level, optname,
4143 optval, optlen);
4144 return do_tcp_setsockopt(sk, level, optname, optval, optlen);
4145}
4146EXPORT_IPV6_MOD(tcp_setsockopt);
4147
4148static void tcp_get_info_chrono_stats(const struct tcp_sock *tp,
4149 struct tcp_info *info)
4150{
4151 u64 stats[__TCP_CHRONO_MAX], total = 0;
4152 enum tcp_chrono i;
4153
4154 for (i = TCP_CHRONO_BUSY; i < __TCP_CHRONO_MAX; ++i) {
4155 stats[i] = tp->chrono_stat[i - 1];
4156 if (i == tp->chrono_type)
4157 stats[i] += tcp_jiffies32 - tp->chrono_start;
4158 stats[i] *= USEC_PER_SEC / HZ;
4159 total += stats[i];
4160 }
4161
4162 info->tcpi_busy_time = total;
4163 info->tcpi_rwnd_limited = stats[TCP_CHRONO_RWND_LIMITED];
4164 info->tcpi_sndbuf_limited = stats[TCP_CHRONO_SNDBUF_LIMITED];
4165}
4166
4167/* Return information about state of tcp endpoint in API format. */
4168void tcp_get_info(struct sock *sk, struct tcp_info *info)
4169{
4170 const struct tcp_sock *tp = tcp_sk(sk); /* iff sk_type == SOCK_STREAM */
4171 const struct inet_connection_sock *icsk = inet_csk(sk);
4172 const u8 ect1_idx = INET_ECN_ECT_1 - 1;
4173 const u8 ect0_idx = INET_ECN_ECT_0 - 1;
4174 const u8 ce_idx = INET_ECN_CE - 1;
4175 unsigned long rate;
4176 u32 now;
4177 u64 rate64;
4178 bool slow;
4179
4180 memset(s: info, c: 0, n: sizeof(*info));
4181 if (sk->sk_type != SOCK_STREAM)
4182 return;
4183
4184 info->tcpi_state = inet_sk_state_load(sk);
4185
4186 /* Report meaningful fields for all TCP states, including listeners */
4187 rate = READ_ONCE(sk->sk_pacing_rate);
4188 rate64 = (rate != ~0UL) ? rate : ~0ULL;
4189 info->tcpi_pacing_rate = rate64;
4190
4191 rate = READ_ONCE(sk->sk_max_pacing_rate);
4192 rate64 = (rate != ~0UL) ? rate : ~0ULL;
4193 info->tcpi_max_pacing_rate = rate64;
4194
4195 info->tcpi_reordering = tp->reordering;
4196 info->tcpi_snd_cwnd = tcp_snd_cwnd(tp);
4197
4198 if (info->tcpi_state == TCP_LISTEN) {
4199 /* listeners aliased fields :
4200 * tcpi_unacked -> Number of children ready for accept()
4201 * tcpi_sacked -> max backlog
4202 */
4203 info->tcpi_unacked = READ_ONCE(sk->sk_ack_backlog);
4204 info->tcpi_sacked = READ_ONCE(sk->sk_max_ack_backlog);
4205 return;
4206 }
4207
4208 slow = lock_sock_fast(sk);
4209
4210 info->tcpi_ca_state = icsk->icsk_ca_state;
4211 info->tcpi_retransmits = icsk->icsk_retransmits;
4212 info->tcpi_probes = icsk->icsk_probes_out;
4213 info->tcpi_backoff = icsk->icsk_backoff;
4214
4215 if (tp->rx_opt.tstamp_ok)
4216 info->tcpi_options |= TCPI_OPT_TIMESTAMPS;
4217 if (tcp_is_sack(tp))
4218 info->tcpi_options |= TCPI_OPT_SACK;
4219 if (tp->rx_opt.wscale_ok) {
4220 info->tcpi_options |= TCPI_OPT_WSCALE;
4221 info->tcpi_snd_wscale = tp->rx_opt.snd_wscale;
4222 info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale;
4223 }
4224
4225 if (tcp_ecn_mode_any(tp))
4226 info->tcpi_options |= TCPI_OPT_ECN;
4227 if (tp->ecn_flags & TCP_ECN_SEEN)
4228 info->tcpi_options |= TCPI_OPT_ECN_SEEN;
4229 if (tp->syn_data_acked)
4230 info->tcpi_options |= TCPI_OPT_SYN_DATA;
4231 if (tp->tcp_usec_ts)
4232 info->tcpi_options |= TCPI_OPT_USEC_TS;
4233 if (tp->syn_fastopen_child)
4234 info->tcpi_options |= TCPI_OPT_TFO_CHILD;
4235
4236 info->tcpi_rto = jiffies_to_usecs(j: icsk->icsk_rto);
4237 info->tcpi_ato = jiffies_to_usecs(min_t(u32, icsk->icsk_ack.ato,
4238 tcp_delack_max(sk)));
4239 info->tcpi_snd_mss = tp->mss_cache;
4240 info->tcpi_rcv_mss = icsk->icsk_ack.rcv_mss;
4241
4242 info->tcpi_unacked = tp->packets_out;
4243 info->tcpi_sacked = tp->sacked_out;
4244
4245 info->tcpi_lost = tp->lost_out;
4246 info->tcpi_retrans = tp->retrans_out;
4247
4248 now = tcp_jiffies32;
4249 info->tcpi_last_data_sent = jiffies_to_msecs(j: now - tp->lsndtime);
4250 info->tcpi_last_data_recv = jiffies_to_msecs(j: now - icsk->icsk_ack.lrcvtime);
4251 info->tcpi_last_ack_recv = jiffies_to_msecs(j: now - tp->rcv_tstamp);
4252
4253 info->tcpi_pmtu = icsk->icsk_pmtu_cookie;
4254 info->tcpi_rcv_ssthresh = tp->rcv_ssthresh;
4255 info->tcpi_rtt = tp->srtt_us >> 3;
4256 info->tcpi_rttvar = tp->mdev_us >> 2;
4257 info->tcpi_snd_ssthresh = tp->snd_ssthresh;
4258 info->tcpi_advmss = tp->advmss;
4259
4260 info->tcpi_rcv_rtt = tp->rcv_rtt_est.rtt_us >> 3;
4261 info->tcpi_rcv_space = tp->rcvq_space.space;
4262
4263 info->tcpi_total_retrans = tp->total_retrans;
4264
4265 info->tcpi_bytes_acked = tp->bytes_acked;
4266 info->tcpi_bytes_received = tp->bytes_received;
4267 info->tcpi_notsent_bytes = max_t(int, 0, tp->write_seq - tp->snd_nxt);
4268 tcp_get_info_chrono_stats(tp, info);
4269
4270 info->tcpi_segs_out = tp->segs_out;
4271
4272 /* segs_in and data_segs_in can be updated from tcp_segs_in() from BH */
4273 info->tcpi_segs_in = READ_ONCE(tp->segs_in);
4274 info->tcpi_data_segs_in = READ_ONCE(tp->data_segs_in);
4275
4276 info->tcpi_min_rtt = tcp_min_rtt(tp);
4277 info->tcpi_data_segs_out = tp->data_segs_out;
4278
4279 info->tcpi_delivery_rate_app_limited = tp->rate_app_limited ? 1 : 0;
4280 rate64 = tcp_compute_delivery_rate(tp);
4281 if (rate64)
4282 info->tcpi_delivery_rate = rate64;
4283 info->tcpi_delivered = tp->delivered;
4284 info->tcpi_delivered_ce = tp->delivered_ce;
4285 info->tcpi_bytes_sent = tp->bytes_sent;
4286 info->tcpi_bytes_retrans = tp->bytes_retrans;
4287 info->tcpi_dsack_dups = tp->dsack_dups;
4288 info->tcpi_reord_seen = tp->reord_seen;
4289 info->tcpi_rcv_ooopack = tp->rcv_ooopack;
4290 info->tcpi_snd_wnd = tp->snd_wnd;
4291 info->tcpi_rcv_wnd = tp->rcv_wnd;
4292 info->tcpi_rehash = tp->plb_rehash + tp->timeout_rehash;
4293 info->tcpi_fastopen_client_fail = tp->fastopen_client_fail;
4294
4295 info->tcpi_total_rto = tp->total_rto;
4296 info->tcpi_total_rto_recoveries = tp->total_rto_recoveries;
4297 info->tcpi_total_rto_time = tp->total_rto_time;
4298 if (tp->rto_stamp)
4299 info->tcpi_total_rto_time += tcp_clock_ms() - tp->rto_stamp;
4300
4301 info->tcpi_accecn_fail_mode = tp->accecn_fail_mode;
4302 info->tcpi_accecn_opt_seen = tp->saw_accecn_opt;
4303 info->tcpi_received_ce = tp->received_ce;
4304 info->tcpi_delivered_e1_bytes = tp->delivered_ecn_bytes[ect1_idx];
4305 info->tcpi_delivered_e0_bytes = tp->delivered_ecn_bytes[ect0_idx];
4306 info->tcpi_delivered_ce_bytes = tp->delivered_ecn_bytes[ce_idx];
4307 info->tcpi_received_e1_bytes = tp->received_ecn_bytes[ect1_idx];
4308 info->tcpi_received_e0_bytes = tp->received_ecn_bytes[ect0_idx];
4309 info->tcpi_received_ce_bytes = tp->received_ecn_bytes[ce_idx];
4310
4311 unlock_sock_fast(sk, slow);
4312}
4313EXPORT_SYMBOL_GPL(tcp_get_info);
4314
4315static size_t tcp_opt_stats_get_size(void)
4316{
4317 return
4318 nla_total_size_64bit(payload: sizeof(u64)) + /* TCP_NLA_BUSY */
4319 nla_total_size_64bit(payload: sizeof(u64)) + /* TCP_NLA_RWND_LIMITED */
4320 nla_total_size_64bit(payload: sizeof(u64)) + /* TCP_NLA_SNDBUF_LIMITED */
4321 nla_total_size_64bit(payload: sizeof(u64)) + /* TCP_NLA_DATA_SEGS_OUT */
4322 nla_total_size_64bit(payload: sizeof(u64)) + /* TCP_NLA_TOTAL_RETRANS */
4323 nla_total_size_64bit(payload: sizeof(u64)) + /* TCP_NLA_PACING_RATE */
4324 nla_total_size_64bit(payload: sizeof(u64)) + /* TCP_NLA_DELIVERY_RATE */
4325 nla_total_size(payload: sizeof(u32)) + /* TCP_NLA_SND_CWND */
4326 nla_total_size(payload: sizeof(u32)) + /* TCP_NLA_REORDERING */
4327 nla_total_size(payload: sizeof(u32)) + /* TCP_NLA_MIN_RTT */
4328 nla_total_size(payload: sizeof(u8)) + /* TCP_NLA_RECUR_RETRANS */
4329 nla_total_size(payload: sizeof(u8)) + /* TCP_NLA_DELIVERY_RATE_APP_LMT */
4330 nla_total_size(payload: sizeof(u32)) + /* TCP_NLA_SNDQ_SIZE */
4331 nla_total_size(payload: sizeof(u8)) + /* TCP_NLA_CA_STATE */
4332 nla_total_size(payload: sizeof(u32)) + /* TCP_NLA_SND_SSTHRESH */
4333 nla_total_size(payload: sizeof(u32)) + /* TCP_NLA_DELIVERED */
4334 nla_total_size(payload: sizeof(u32)) + /* TCP_NLA_DELIVERED_CE */
4335 nla_total_size_64bit(payload: sizeof(u64)) + /* TCP_NLA_BYTES_SENT */
4336 nla_total_size_64bit(payload: sizeof(u64)) + /* TCP_NLA_BYTES_RETRANS */
4337 nla_total_size(payload: sizeof(u32)) + /* TCP_NLA_DSACK_DUPS */
4338 nla_total_size(payload: sizeof(u32)) + /* TCP_NLA_REORD_SEEN */
4339 nla_total_size(payload: sizeof(u32)) + /* TCP_NLA_SRTT */
4340 nla_total_size(payload: sizeof(u16)) + /* TCP_NLA_TIMEOUT_REHASH */
4341 nla_total_size(payload: sizeof(u32)) + /* TCP_NLA_BYTES_NOTSENT */
4342 nla_total_size_64bit(payload: sizeof(u64)) + /* TCP_NLA_EDT */
4343 nla_total_size(payload: sizeof(u8)) + /* TCP_NLA_TTL */
4344 nla_total_size(payload: sizeof(u32)) + /* TCP_NLA_REHASH */
4345 0;
4346}
4347
4348/* Returns TTL or hop limit of an incoming packet from skb. */
4349static u8 tcp_skb_ttl_or_hop_limit(const struct sk_buff *skb)
4350{
4351 if (skb->protocol == htons(ETH_P_IP))
4352 return ip_hdr(skb)->ttl;
4353 else if (skb->protocol == htons(ETH_P_IPV6))
4354 return ipv6_hdr(skb)->hop_limit;
4355 else
4356 return 0;
4357}
4358
4359struct sk_buff *tcp_get_timestamping_opt_stats(const struct sock *sk,
4360 const struct sk_buff *orig_skb,
4361 const struct sk_buff *ack_skb)
4362{
4363 const struct tcp_sock *tp = tcp_sk(sk);
4364 struct sk_buff *stats;
4365 struct tcp_info info;
4366 unsigned long rate;
4367 u64 rate64;
4368
4369 stats = alloc_skb(size: tcp_opt_stats_get_size(), GFP_ATOMIC);
4370 if (!stats)
4371 return NULL;
4372
4373 tcp_get_info_chrono_stats(tp, info: &info);
4374 nla_put_u64_64bit(skb: stats, attrtype: TCP_NLA_BUSY,
4375 value: info.tcpi_busy_time, padattr: TCP_NLA_PAD);
4376 nla_put_u64_64bit(skb: stats, attrtype: TCP_NLA_RWND_LIMITED,
4377 value: info.tcpi_rwnd_limited, padattr: TCP_NLA_PAD);
4378 nla_put_u64_64bit(skb: stats, attrtype: TCP_NLA_SNDBUF_LIMITED,
4379 value: info.tcpi_sndbuf_limited, padattr: TCP_NLA_PAD);
4380 nla_put_u64_64bit(skb: stats, attrtype: TCP_NLA_DATA_SEGS_OUT,
4381 value: tp->data_segs_out, padattr: TCP_NLA_PAD);
4382 nla_put_u64_64bit(skb: stats, attrtype: TCP_NLA_TOTAL_RETRANS,
4383 value: tp->total_retrans, padattr: TCP_NLA_PAD);
4384
4385 rate = READ_ONCE(sk->sk_pacing_rate);
4386 rate64 = (rate != ~0UL) ? rate : ~0ULL;
4387 nla_put_u64_64bit(skb: stats, attrtype: TCP_NLA_PACING_RATE, value: rate64, padattr: TCP_NLA_PAD);
4388
4389 rate64 = tcp_compute_delivery_rate(tp);
4390 nla_put_u64_64bit(skb: stats, attrtype: TCP_NLA_DELIVERY_RATE, value: rate64, padattr: TCP_NLA_PAD);
4391
4392 nla_put_u32(skb: stats, attrtype: TCP_NLA_SND_CWND, value: tcp_snd_cwnd(tp));
4393 nla_put_u32(skb: stats, attrtype: TCP_NLA_REORDERING, value: tp->reordering);
4394 nla_put_u32(skb: stats, attrtype: TCP_NLA_MIN_RTT, value: tcp_min_rtt(tp));
4395
4396 nla_put_u8(skb: stats, attrtype: TCP_NLA_RECUR_RETRANS,
4397 READ_ONCE(inet_csk(sk)->icsk_retransmits));
4398 nla_put_u8(skb: stats, attrtype: TCP_NLA_DELIVERY_RATE_APP_LMT, value: !!tp->rate_app_limited);
4399 nla_put_u32(skb: stats, attrtype: TCP_NLA_SND_SSTHRESH, value: tp->snd_ssthresh);
4400 nla_put_u32(skb: stats, attrtype: TCP_NLA_DELIVERED, value: tp->delivered);
4401 nla_put_u32(skb: stats, attrtype: TCP_NLA_DELIVERED_CE, value: tp->delivered_ce);
4402
4403 nla_put_u32(skb: stats, attrtype: TCP_NLA_SNDQ_SIZE, value: tp->write_seq - tp->snd_una);
4404 nla_put_u8(skb: stats, attrtype: TCP_NLA_CA_STATE, inet_csk(sk)->icsk_ca_state);
4405
4406 nla_put_u64_64bit(skb: stats, attrtype: TCP_NLA_BYTES_SENT, value: tp->bytes_sent,
4407 padattr: TCP_NLA_PAD);
4408 nla_put_u64_64bit(skb: stats, attrtype: TCP_NLA_BYTES_RETRANS, value: tp->bytes_retrans,
4409 padattr: TCP_NLA_PAD);
4410 nla_put_u32(skb: stats, attrtype: TCP_NLA_DSACK_DUPS, value: tp->dsack_dups);
4411 nla_put_u32(skb: stats, attrtype: TCP_NLA_REORD_SEEN, value: tp->reord_seen);
4412 nla_put_u32(skb: stats, attrtype: TCP_NLA_SRTT, value: tp->srtt_us >> 3);
4413 nla_put_u16(skb: stats, attrtype: TCP_NLA_TIMEOUT_REHASH, value: tp->timeout_rehash);
4414 nla_put_u32(skb: stats, attrtype: TCP_NLA_BYTES_NOTSENT,
4415 max_t(int, 0, tp->write_seq - tp->snd_nxt));
4416 nla_put_u64_64bit(skb: stats, attrtype: TCP_NLA_EDT, value: orig_skb->skb_mstamp_ns,
4417 padattr: TCP_NLA_PAD);
4418 if (ack_skb)
4419 nla_put_u8(skb: stats, attrtype: TCP_NLA_TTL,
4420 value: tcp_skb_ttl_or_hop_limit(skb: ack_skb));
4421
4422 nla_put_u32(skb: stats, attrtype: TCP_NLA_REHASH, value: tp->plb_rehash + tp->timeout_rehash);
4423 return stats;
4424}
4425
4426int do_tcp_getsockopt(struct sock *sk, int level,
4427 int optname, sockptr_t optval, sockptr_t optlen)
4428{
4429 struct inet_connection_sock *icsk = inet_csk(sk);
4430 struct tcp_sock *tp = tcp_sk(sk);
4431 struct net *net = sock_net(sk);
4432 int user_mss;
4433 int val, len;
4434
4435 if (copy_from_sockptr(dst: &len, src: optlen, size: sizeof(int)))
4436 return -EFAULT;
4437
4438 if (len < 0)
4439 return -EINVAL;
4440
4441 len = min_t(unsigned int, len, sizeof(int));
4442
4443 switch (optname) {
4444 case TCP_MAXSEG:
4445 val = tp->mss_cache;
4446 user_mss = READ_ONCE(tp->rx_opt.user_mss);
4447 if (user_mss &&
4448 ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
4449 val = user_mss;
4450 if (tp->repair)
4451 val = tp->rx_opt.mss_clamp;
4452 break;
4453 case TCP_NODELAY:
4454 val = !!(tp->nonagle&TCP_NAGLE_OFF);
4455 break;
4456 case TCP_CORK:
4457 val = !!(tp->nonagle&TCP_NAGLE_CORK);
4458 break;
4459 case TCP_KEEPIDLE:
4460 val = keepalive_time_when(tp) / HZ;
4461 break;
4462 case TCP_KEEPINTVL:
4463 val = keepalive_intvl_when(tp) / HZ;
4464 break;
4465 case TCP_KEEPCNT:
4466 val = keepalive_probes(tp);
4467 break;
4468 case TCP_SYNCNT:
4469 val = READ_ONCE(icsk->icsk_syn_retries) ? :
4470 READ_ONCE(net->ipv4.sysctl_tcp_syn_retries);
4471 break;
4472 case TCP_LINGER2:
4473 val = READ_ONCE(tp->linger2);
4474 if (val >= 0)
4475 val = (val ? : READ_ONCE(net->ipv4.sysctl_tcp_fin_timeout)) / HZ;
4476 break;
4477 case TCP_DEFER_ACCEPT:
4478 val = READ_ONCE(icsk->icsk_accept_queue.rskq_defer_accept);
4479 val = retrans_to_secs(retrans: val, TCP_TIMEOUT_INIT / HZ,
4480 TCP_RTO_MAX / HZ);
4481 break;
4482 case TCP_WINDOW_CLAMP:
4483 val = READ_ONCE(tp->window_clamp);
4484 break;
4485 case TCP_INFO: {
4486 struct tcp_info info;
4487
4488 if (copy_from_sockptr(dst: &len, src: optlen, size: sizeof(int)))
4489 return -EFAULT;
4490
4491 tcp_get_info(sk, &info);
4492
4493 len = min_t(unsigned int, len, sizeof(info));
4494 if (copy_to_sockptr(dst: optlen, src: &len, size: sizeof(int)))
4495 return -EFAULT;
4496 if (copy_to_sockptr(dst: optval, src: &info, size: len))
4497 return -EFAULT;
4498 return 0;
4499 }
4500 case TCP_CC_INFO: {
4501 const struct tcp_congestion_ops *ca_ops;
4502 union tcp_cc_info info;
4503 size_t sz = 0;
4504 int attr;
4505
4506 if (copy_from_sockptr(dst: &len, src: optlen, size: sizeof(int)))
4507 return -EFAULT;
4508
4509 ca_ops = icsk->icsk_ca_ops;
4510 if (ca_ops && ca_ops->get_info)
4511 sz = ca_ops->get_info(sk, ~0U, &attr, &info);
4512
4513 len = min_t(unsigned int, len, sz);
4514 if (copy_to_sockptr(dst: optlen, src: &len, size: sizeof(int)))
4515 return -EFAULT;
4516 if (copy_to_sockptr(dst: optval, src: &info, size: len))
4517 return -EFAULT;
4518 return 0;
4519 }
4520 case TCP_QUICKACK:
4521 val = !inet_csk_in_pingpong_mode(sk);
4522 break;
4523
4524 case TCP_CONGESTION:
4525 if (copy_from_sockptr(dst: &len, src: optlen, size: sizeof(int)))
4526 return -EFAULT;
4527 len = min_t(unsigned int, len, TCP_CA_NAME_MAX);
4528 if (copy_to_sockptr(dst: optlen, src: &len, size: sizeof(int)))
4529 return -EFAULT;
4530 if (copy_to_sockptr(dst: optval, src: icsk->icsk_ca_ops->name, size: len))
4531 return -EFAULT;
4532 return 0;
4533
4534 case TCP_ULP:
4535 if (copy_from_sockptr(dst: &len, src: optlen, size: sizeof(int)))
4536 return -EFAULT;
4537 len = min_t(unsigned int, len, TCP_ULP_NAME_MAX);
4538 if (!icsk->icsk_ulp_ops) {
4539 len = 0;
4540 if (copy_to_sockptr(dst: optlen, src: &len, size: sizeof(int)))
4541 return -EFAULT;
4542 return 0;
4543 }
4544 if (copy_to_sockptr(dst: optlen, src: &len, size: sizeof(int)))
4545 return -EFAULT;
4546 if (copy_to_sockptr(dst: optval, src: icsk->icsk_ulp_ops->name, size: len))
4547 return -EFAULT;
4548 return 0;
4549
4550 case TCP_FASTOPEN_KEY: {
4551 u64 key[TCP_FASTOPEN_KEY_BUF_LENGTH / sizeof(u64)];
4552 unsigned int key_len;
4553
4554 if (copy_from_sockptr(dst: &len, src: optlen, size: sizeof(int)))
4555 return -EFAULT;
4556
4557 key_len = tcp_fastopen_get_cipher(net, icsk, key) *
4558 TCP_FASTOPEN_KEY_LENGTH;
4559 len = min_t(unsigned int, len, key_len);
4560 if (copy_to_sockptr(dst: optlen, src: &len, size: sizeof(int)))
4561 return -EFAULT;
4562 if (copy_to_sockptr(dst: optval, src: key, size: len))
4563 return -EFAULT;
4564 return 0;
4565 }
4566 case TCP_THIN_LINEAR_TIMEOUTS:
4567 val = tp->thin_lto;
4568 break;
4569
4570 case TCP_THIN_DUPACK:
4571 val = 0;
4572 break;
4573
4574 case TCP_REPAIR:
4575 val = tp->repair;
4576 break;
4577
4578 case TCP_REPAIR_QUEUE:
4579 if (tp->repair)
4580 val = tp->repair_queue;
4581 else
4582 return -EINVAL;
4583 break;
4584
4585 case TCP_REPAIR_WINDOW: {
4586 struct tcp_repair_window opt;
4587
4588 if (copy_from_sockptr(dst: &len, src: optlen, size: sizeof(int)))
4589 return -EFAULT;
4590
4591 if (len != sizeof(opt))
4592 return -EINVAL;
4593
4594 if (!tp->repair)
4595 return -EPERM;
4596
4597 opt.snd_wl1 = tp->snd_wl1;
4598 opt.snd_wnd = tp->snd_wnd;
4599 opt.max_window = tp->max_window;
4600 opt.rcv_wnd = tp->rcv_wnd;
4601 opt.rcv_wup = tp->rcv_wup;
4602
4603 if (copy_to_sockptr(dst: optval, src: &opt, size: len))
4604 return -EFAULT;
4605 return 0;
4606 }
4607 case TCP_QUEUE_SEQ:
4608 if (tp->repair_queue == TCP_SEND_QUEUE)
4609 val = tp->write_seq;
4610 else if (tp->repair_queue == TCP_RECV_QUEUE)
4611 val = tp->rcv_nxt;
4612 else
4613 return -EINVAL;
4614 break;
4615
4616 case TCP_USER_TIMEOUT:
4617 val = READ_ONCE(icsk->icsk_user_timeout);
4618 break;
4619
4620 case TCP_FASTOPEN:
4621 val = READ_ONCE(icsk->icsk_accept_queue.fastopenq.max_qlen);
4622 break;
4623
4624 case TCP_FASTOPEN_CONNECT:
4625 val = tp->fastopen_connect;
4626 break;
4627
4628 case TCP_FASTOPEN_NO_COOKIE:
4629 val = tp->fastopen_no_cookie;
4630 break;
4631
4632 case TCP_TX_DELAY:
4633 val = READ_ONCE(tp->tcp_tx_delay);
4634 break;
4635
4636 case TCP_TIMESTAMP:
4637 val = tcp_clock_ts(usec_ts: tp->tcp_usec_ts) + READ_ONCE(tp->tsoffset);
4638 if (tp->tcp_usec_ts)
4639 val |= 1;
4640 else
4641 val &= ~1;
4642 break;
4643 case TCP_NOTSENT_LOWAT:
4644 val = READ_ONCE(tp->notsent_lowat);
4645 break;
4646 case TCP_INQ:
4647 val = tp->recvmsg_inq;
4648 break;
4649 case TCP_SAVE_SYN:
4650 val = tp->save_syn;
4651 break;
4652 case TCP_SAVED_SYN: {
4653 if (copy_from_sockptr(dst: &len, src: optlen, size: sizeof(int)))
4654 return -EFAULT;
4655
4656 sockopt_lock_sock(sk);
4657 if (tp->saved_syn) {
4658 if (len < tcp_saved_syn_len(saved_syn: tp->saved_syn)) {
4659 len = tcp_saved_syn_len(saved_syn: tp->saved_syn);
4660 if (copy_to_sockptr(dst: optlen, src: &len, size: sizeof(int))) {
4661 sockopt_release_sock(sk);
4662 return -EFAULT;
4663 }
4664 sockopt_release_sock(sk);
4665 return -EINVAL;
4666 }
4667 len = tcp_saved_syn_len(saved_syn: tp->saved_syn);
4668 if (copy_to_sockptr(dst: optlen, src: &len, size: sizeof(int))) {
4669 sockopt_release_sock(sk);
4670 return -EFAULT;
4671 }
4672 if (copy_to_sockptr(dst: optval, src: tp->saved_syn->data, size: len)) {
4673 sockopt_release_sock(sk);
4674 return -EFAULT;
4675 }
4676 tcp_saved_syn_free(tp);
4677 sockopt_release_sock(sk);
4678 } else {
4679 sockopt_release_sock(sk);
4680 len = 0;
4681 if (copy_to_sockptr(dst: optlen, src: &len, size: sizeof(int)))
4682 return -EFAULT;
4683 }
4684 return 0;
4685 }
4686#ifdef CONFIG_MMU
4687 case TCP_ZEROCOPY_RECEIVE: {
4688 struct scm_timestamping_internal tss;
4689 struct tcp_zerocopy_receive zc = {};
4690 int err;
4691
4692 if (copy_from_sockptr(dst: &len, src: optlen, size: sizeof(int)))
4693 return -EFAULT;
4694 if (len < 0 ||
4695 len < offsetofend(struct tcp_zerocopy_receive, length))
4696 return -EINVAL;
4697 if (unlikely(len > sizeof(zc))) {
4698 err = check_zeroed_sockptr(src: optval, offset: sizeof(zc),
4699 size: len - sizeof(zc));
4700 if (err < 1)
4701 return err == 0 ? -EINVAL : err;
4702 len = sizeof(zc);
4703 if (copy_to_sockptr(dst: optlen, src: &len, size: sizeof(int)))
4704 return -EFAULT;
4705 }
4706 if (copy_from_sockptr(dst: &zc, src: optval, size: len))
4707 return -EFAULT;
4708 if (zc.reserved)
4709 return -EINVAL;
4710 if (zc.msg_flags & ~(TCP_VALID_ZC_MSG_FLAGS))
4711 return -EINVAL;
4712 sockopt_lock_sock(sk);
4713 err = tcp_zerocopy_receive(sk, zc: &zc, tss: &tss);
4714 err = BPF_CGROUP_RUN_PROG_GETSOCKOPT_KERN(sk, level, optname,
4715 &zc, &len, err);
4716 sockopt_release_sock(sk);
4717 if (len >= offsetofend(struct tcp_zerocopy_receive, msg_flags))
4718 goto zerocopy_rcv_cmsg;
4719 switch (len) {
4720 case offsetofend(struct tcp_zerocopy_receive, msg_flags):
4721 goto zerocopy_rcv_cmsg;
4722 case offsetofend(struct tcp_zerocopy_receive, msg_controllen):
4723 case offsetofend(struct tcp_zerocopy_receive, msg_control):
4724 case offsetofend(struct tcp_zerocopy_receive, flags):
4725 case offsetofend(struct tcp_zerocopy_receive, copybuf_len):
4726 case offsetofend(struct tcp_zerocopy_receive, copybuf_address):
4727 case offsetofend(struct tcp_zerocopy_receive, err):
4728 goto zerocopy_rcv_sk_err;
4729 case offsetofend(struct tcp_zerocopy_receive, inq):
4730 goto zerocopy_rcv_inq;
4731 case offsetofend(struct tcp_zerocopy_receive, length):
4732 default:
4733 goto zerocopy_rcv_out;
4734 }
4735zerocopy_rcv_cmsg:
4736 if (zc.msg_flags & TCP_CMSG_TS)
4737 tcp_zc_finalize_rx_tstamp(sk, zc: &zc, tss: &tss);
4738 else
4739 zc.msg_flags = 0;
4740zerocopy_rcv_sk_err:
4741 if (!err)
4742 zc.err = sock_error(sk);
4743zerocopy_rcv_inq:
4744 zc.inq = tcp_inq_hint(sk);
4745zerocopy_rcv_out:
4746 if (!err && copy_to_sockptr(dst: optval, src: &zc, size: len))
4747 err = -EFAULT;
4748 return err;
4749 }
4750#endif
4751 case TCP_AO_REPAIR:
4752 if (!tcp_can_repair_sock(sk))
4753 return -EPERM;
4754 return tcp_ao_get_repair(sk, optval, optlen);
4755 case TCP_AO_GET_KEYS:
4756 case TCP_AO_INFO: {
4757 int err;
4758
4759 sockopt_lock_sock(sk);
4760 if (optname == TCP_AO_GET_KEYS)
4761 err = tcp_ao_get_mkts(sk, optval, optlen);
4762 else
4763 err = tcp_ao_get_sock_info(sk, optval, optlen);
4764 sockopt_release_sock(sk);
4765
4766 return err;
4767 }
4768 case TCP_IS_MPTCP:
4769 val = 0;
4770 break;
4771 case TCP_RTO_MAX_MS:
4772 val = jiffies_to_msecs(j: tcp_rto_max(sk));
4773 break;
4774 case TCP_RTO_MIN_US:
4775 val = jiffies_to_usecs(READ_ONCE(inet_csk(sk)->icsk_rto_min));
4776 break;
4777 case TCP_DELACK_MAX_US:
4778 val = jiffies_to_usecs(READ_ONCE(inet_csk(sk)->icsk_delack_max));
4779 break;
4780 default:
4781 return -ENOPROTOOPT;
4782 }
4783
4784 if (copy_to_sockptr(dst: optlen, src: &len, size: sizeof(int)))
4785 return -EFAULT;
4786 if (copy_to_sockptr(dst: optval, src: &val, size: len))
4787 return -EFAULT;
4788 return 0;
4789}
4790
4791bool tcp_bpf_bypass_getsockopt(int level, int optname)
4792{
4793 /* TCP do_tcp_getsockopt has optimized getsockopt implementation
4794 * to avoid extra socket lock for TCP_ZEROCOPY_RECEIVE.
4795 */
4796 if (level == SOL_TCP && optname == TCP_ZEROCOPY_RECEIVE)
4797 return true;
4798
4799 return false;
4800}
4801EXPORT_IPV6_MOD(tcp_bpf_bypass_getsockopt);
4802
4803int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval,
4804 int __user *optlen)
4805{
4806 struct inet_connection_sock *icsk = inet_csk(sk);
4807
4808 if (level != SOL_TCP)
4809 /* Paired with WRITE_ONCE() in do_ipv6_setsockopt() and tcp_v6_connect() */
4810 return READ_ONCE(icsk->icsk_af_ops)->getsockopt(sk, level, optname,
4811 optval, optlen);
4812 return do_tcp_getsockopt(sk, level, optname, optval: USER_SOCKPTR(p: optval),
4813 optlen: USER_SOCKPTR(p: optlen));
4814}
4815EXPORT_IPV6_MOD(tcp_getsockopt);
4816
4817#ifdef CONFIG_TCP_MD5SIG
4818int tcp_md5_sigpool_id = -1;
4819EXPORT_IPV6_MOD_GPL(tcp_md5_sigpool_id);
4820
4821int tcp_md5_alloc_sigpool(void)
4822{
4823 size_t scratch_size;
4824 int ret;
4825
4826 scratch_size = sizeof(union tcp_md5sum_block) + sizeof(struct tcphdr);
4827 ret = tcp_sigpool_alloc_ahash(alg: "md5", scratch_size);
4828 if (ret >= 0) {
4829 /* As long as any md5 sigpool was allocated, the return
4830 * id would stay the same. Re-write the id only for the case
4831 * when previously all MD5 keys were deleted and this call
4832 * allocates the first MD5 key, which may return a different
4833 * sigpool id than was used previously.
4834 */
4835 WRITE_ONCE(tcp_md5_sigpool_id, ret); /* Avoids the compiler potentially being smart here */
4836 return 0;
4837 }
4838 return ret;
4839}
4840
4841void tcp_md5_release_sigpool(void)
4842{
4843 tcp_sigpool_release(READ_ONCE(tcp_md5_sigpool_id));
4844}
4845
4846void tcp_md5_add_sigpool(void)
4847{
4848 tcp_sigpool_get(READ_ONCE(tcp_md5_sigpool_id));
4849}
4850
4851int tcp_md5_hash_key(struct tcp_sigpool *hp,
4852 const struct tcp_md5sig_key *key)
4853{
4854 u8 keylen = READ_ONCE(key->keylen); /* paired with WRITE_ONCE() in tcp_md5_do_add */
4855 struct scatterlist sg;
4856
4857 sg_init_one(&sg, key->key, keylen);
4858 ahash_request_set_crypt(req: hp->req, src: &sg, NULL, nbytes: keylen);
4859
4860 /* We use data_race() because tcp_md5_do_add() might change
4861 * key->key under us
4862 */
4863 return data_race(crypto_ahash_update(hp->req));
4864}
4865EXPORT_IPV6_MOD(tcp_md5_hash_key);
4866
4867/* Called with rcu_read_lock() */
4868static enum skb_drop_reason
4869tcp_inbound_md5_hash(const struct sock *sk, const struct sk_buff *skb,
4870 const void *saddr, const void *daddr,
4871 int family, int l3index, const __u8 *hash_location)
4872{
4873 /* This gets called for each TCP segment that has TCP-MD5 option.
4874 * We have 3 drop cases:
4875 * o No MD5 hash and one expected.
4876 * o MD5 hash and we're not expecting one.
4877 * o MD5 hash and its wrong.
4878 */
4879 const struct tcp_sock *tp = tcp_sk(sk);
4880 struct tcp_md5sig_key *key;
4881 u8 newhash[16];
4882 int genhash;
4883
4884 key = tcp_md5_do_lookup(sk, l3index, addr: saddr, family);
4885
4886 if (!key && hash_location) {
4887 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMD5UNEXPECTED);
4888 trace_tcp_hash_md5_unexpected(sk, skb);
4889 return SKB_DROP_REASON_TCP_MD5UNEXPECTED;
4890 }
4891
4892 /* Check the signature.
4893 * To support dual stack listeners, we need to handle
4894 * IPv4-mapped case.
4895 */
4896 if (family == AF_INET)
4897 genhash = tcp_v4_md5_hash_skb(md5_hash: newhash, key, NULL, skb);
4898 else
4899 genhash = tp->af_specific->calc_md5_hash(newhash, key,
4900 NULL, skb);
4901 if (genhash || memcmp(hash_location, newhash, 16) != 0) {
4902 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMD5FAILURE);
4903 trace_tcp_hash_md5_mismatch(sk, skb);
4904 return SKB_DROP_REASON_TCP_MD5FAILURE;
4905 }
4906 return SKB_NOT_DROPPED_YET;
4907}
4908#else
4909static inline enum skb_drop_reason
4910tcp_inbound_md5_hash(const struct sock *sk, const struct sk_buff *skb,
4911 const void *saddr, const void *daddr,
4912 int family, int l3index, const __u8 *hash_location)
4913{
4914 return SKB_NOT_DROPPED_YET;
4915}
4916
4917#endif
4918
4919/* Called with rcu_read_lock() */
4920enum skb_drop_reason
4921tcp_inbound_hash(struct sock *sk, const struct request_sock *req,
4922 const struct sk_buff *skb,
4923 const void *saddr, const void *daddr,
4924 int family, int dif, int sdif)
4925{
4926 const struct tcphdr *th = tcp_hdr(skb);
4927 const struct tcp_ao_hdr *aoh;
4928 const __u8 *md5_location;
4929 int l3index;
4930
4931 /* Invalid option or two times meet any of auth options */
4932 if (tcp_parse_auth_options(th, md5_hash: &md5_location, aoh: &aoh)) {
4933 trace_tcp_hash_bad_header(sk, skb);
4934 return SKB_DROP_REASON_TCP_AUTH_HDR;
4935 }
4936
4937 if (req) {
4938 if (tcp_rsk_used_ao(req) != !!aoh) {
4939 u8 keyid, rnext, maclen;
4940
4941 if (aoh) {
4942 keyid = aoh->keyid;
4943 rnext = aoh->rnext_keyid;
4944 maclen = tcp_ao_hdr_maclen(aoh);
4945 } else {
4946 keyid = rnext = maclen = 0;
4947 }
4948
4949 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPAOBAD);
4950 trace_tcp_ao_handshake_failure(sk, skb, keyid, rnext, maclen);
4951 return SKB_DROP_REASON_TCP_AOFAILURE;
4952 }
4953 }
4954
4955 /* sdif set, means packet ingressed via a device
4956 * in an L3 domain and dif is set to the l3mdev
4957 */
4958 l3index = sdif ? dif : 0;
4959
4960 /* Fast path: unsigned segments */
4961 if (likely(!md5_location && !aoh)) {
4962 /* Drop if there's TCP-MD5 or TCP-AO key with any rcvid/sndid
4963 * for the remote peer. On TCP-AO established connection
4964 * the last key is impossible to remove, so there's
4965 * always at least one current_key.
4966 */
4967 if (tcp_ao_required(sk, saddr, family, l3index, stat_inc: true)) {
4968 trace_tcp_hash_ao_required(sk, skb);
4969 return SKB_DROP_REASON_TCP_AONOTFOUND;
4970 }
4971 if (unlikely(tcp_md5_do_lookup(sk, l3index, saddr, family))) {
4972 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMD5NOTFOUND);
4973 trace_tcp_hash_md5_required(sk, skb);
4974 return SKB_DROP_REASON_TCP_MD5NOTFOUND;
4975 }
4976 return SKB_NOT_DROPPED_YET;
4977 }
4978
4979 if (aoh)
4980 return tcp_inbound_ao_hash(sk, skb, family, req, l3index, aoh);
4981
4982 return tcp_inbound_md5_hash(sk, skb, saddr, daddr, family,
4983 l3index, hash_location: md5_location);
4984}
4985EXPORT_IPV6_MOD_GPL(tcp_inbound_hash);
4986
4987void tcp_done(struct sock *sk)
4988{
4989 struct request_sock *req;
4990
4991 /* We might be called with a new socket, after
4992 * inet_csk_prepare_forced_close() has been called
4993 * so we can not use lockdep_sock_is_held(sk)
4994 */
4995 req = rcu_dereference_protected(tcp_sk(sk)->fastopen_rsk, 1);
4996
4997 if (sk->sk_state == TCP_SYN_SENT || sk->sk_state == TCP_SYN_RECV)
4998 TCP_INC_STATS(sock_net(sk), TCP_MIB_ATTEMPTFAILS);
4999
5000 tcp_set_state(sk, TCP_CLOSE);
5001 tcp_clear_xmit_timers(sk);
5002 if (req)
5003 reqsk_fastopen_remove(sk, req, reset: false);
5004
5005 WRITE_ONCE(sk->sk_shutdown, SHUTDOWN_MASK);
5006
5007 if (!sock_flag(sk, flag: SOCK_DEAD))
5008 sk->sk_state_change(sk);
5009 else
5010 inet_csk_destroy_sock(sk);
5011}
5012EXPORT_SYMBOL_GPL(tcp_done);
5013
5014int tcp_abort(struct sock *sk, int err)
5015{
5016 int state = inet_sk_state_load(sk);
5017
5018 if (state == TCP_NEW_SYN_RECV) {
5019 struct request_sock *req = inet_reqsk(sk);
5020
5021 local_bh_disable();
5022 inet_csk_reqsk_queue_drop(sk: req->rsk_listener, req);
5023 local_bh_enable();
5024 return 0;
5025 }
5026 if (state == TCP_TIME_WAIT) {
5027 struct inet_timewait_sock *tw = inet_twsk(sk);
5028
5029 refcount_inc(r: &tw->tw_refcnt);
5030 local_bh_disable();
5031 inet_twsk_deschedule_put(tw);
5032 local_bh_enable();
5033 return 0;
5034 }
5035
5036 /* BPF context ensures sock locking. */
5037 if (!has_current_bpf_ctx())
5038 /* Don't race with userspace socket closes such as tcp_close. */
5039 lock_sock(sk);
5040
5041 /* Avoid closing the same socket twice. */
5042 if (sk->sk_state == TCP_CLOSE) {
5043 if (!has_current_bpf_ctx())
5044 release_sock(sk);
5045 return -ENOENT;
5046 }
5047
5048 if (sk->sk_state == TCP_LISTEN) {
5049 tcp_set_state(sk, TCP_CLOSE);
5050 inet_csk_listen_stop(sk);
5051 }
5052
5053 /* Don't race with BH socket closes such as inet_csk_listen_stop. */
5054 local_bh_disable();
5055 bh_lock_sock(sk);
5056
5057 if (tcp_need_reset(state: sk->sk_state))
5058 tcp_send_active_reset(sk, GFP_ATOMIC,
5059 reason: SK_RST_REASON_TCP_STATE);
5060 tcp_done_with_error(sk, err);
5061
5062 bh_unlock_sock(sk);
5063 local_bh_enable();
5064 if (!has_current_bpf_ctx())
5065 release_sock(sk);
5066 return 0;
5067}
5068EXPORT_SYMBOL_GPL(tcp_abort);
5069
5070extern struct tcp_congestion_ops tcp_reno;
5071
5072static __initdata unsigned long thash_entries;
5073static int __init set_thash_entries(char *str)
5074{
5075 ssize_t ret;
5076
5077 if (!str)
5078 return 0;
5079
5080 ret = kstrtoul(s: str, base: 0, res: &thash_entries);
5081 if (ret)
5082 return 0;
5083
5084 return 1;
5085}
5086__setup("thash_entries=", set_thash_entries);
5087
5088static void __init tcp_init_mem(void)
5089{
5090 unsigned long limit = nr_free_buffer_pages() / 16;
5091
5092 limit = max(limit, 128UL);
5093 sysctl_tcp_mem[0] = limit / 4 * 3; /* 4.68 % */
5094 sysctl_tcp_mem[1] = limit; /* 6.25 % */
5095 sysctl_tcp_mem[2] = sysctl_tcp_mem[0] * 2; /* 9.37 % */
5096}
5097
5098static void __init tcp_struct_check(void)
5099{
5100 /* TX read-mostly hotpath cache lines */
5101 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_tx, max_window);
5102 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_tx, rcv_ssthresh);
5103 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_tx, reordering);
5104 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_tx, notsent_lowat);
5105 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_tx, gso_segs);
5106 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_tx, retransmit_skb_hint);
5107#if IS_ENABLED(CONFIG_TLS_DEVICE)
5108 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_tx, tcp_clean_acked);
5109#endif
5110
5111 /* TXRX read-mostly hotpath cache lines */
5112 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_txrx, tsoffset);
5113 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_txrx, snd_wnd);
5114 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_txrx, mss_cache);
5115 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_txrx, snd_cwnd);
5116 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_txrx, prr_out);
5117 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_txrx, lost_out);
5118 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_txrx, sacked_out);
5119 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_txrx, scaling_ratio);
5120
5121 /* RX read-mostly hotpath cache lines */
5122 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_rx, copied_seq);
5123 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_rx, snd_wl1);
5124 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_rx, tlp_high_seq);
5125 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_rx, rttvar_us);
5126 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_rx, retrans_out);
5127 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_rx, advmss);
5128 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_rx, urg_data);
5129 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_rx, lost);
5130 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_rx, rtt_min);
5131 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_rx, out_of_order_queue);
5132 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_rx, snd_ssthresh);
5133
5134 /* TX read-write hotpath cache lines */
5135 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, segs_out);
5136 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, data_segs_out);
5137 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, bytes_sent);
5138 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, snd_sml);
5139 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, chrono_start);
5140 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, chrono_stat);
5141 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, write_seq);
5142 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, pushed_seq);
5143 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, lsndtime);
5144 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, mdev_us);
5145 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, tcp_wstamp_ns);
5146 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, accecn_opt_tstamp);
5147 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, rtt_seq);
5148 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, tsorted_sent_queue);
5149 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, highest_sack);
5150 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, ecn_flags);
5151
5152 /* TXRX read-write hotpath cache lines */
5153 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_txrx, pred_flags);
5154 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_txrx, tcp_clock_cache);
5155 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_txrx, tcp_mstamp);
5156 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_txrx, rcv_nxt);
5157 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_txrx, snd_nxt);
5158 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_txrx, snd_una);
5159 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_txrx, window_clamp);
5160 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_txrx, srtt_us);
5161 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_txrx, packets_out);
5162 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_txrx, snd_up);
5163 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_txrx, delivered);
5164 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_txrx, delivered_ce);
5165 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_txrx, received_ce);
5166 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_txrx, received_ecn_bytes);
5167 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_txrx, app_limited);
5168 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_txrx, rcv_wnd);
5169 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_txrx, rcv_tstamp);
5170 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_txrx, rx_opt);
5171
5172 /* RX read-write hotpath cache lines */
5173 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_rx, bytes_received);
5174 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_rx, segs_in);
5175 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_rx, data_segs_in);
5176 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_rx, rcv_wup);
5177 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_rx, max_packets_out);
5178 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_rx, cwnd_usage_seq);
5179 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_rx, rate_delivered);
5180 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_rx, rate_interval_us);
5181 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_rx, rcv_rtt_last_tsecr);
5182 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_rx, delivered_ecn_bytes);
5183 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_rx, first_tx_mstamp);
5184 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_rx, delivered_mstamp);
5185 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_rx, bytes_acked);
5186 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_rx, rcv_rtt_est);
5187 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_rx, rcvq_space);
5188}
5189
5190void __init tcp_init(void)
5191{
5192 int max_rshare, max_wshare, cnt;
5193 unsigned long limit;
5194 unsigned int i;
5195
5196 BUILD_BUG_ON(TCP_MIN_SND_MSS <= MAX_TCP_OPTION_SPACE);
5197 BUILD_BUG_ON(sizeof(struct tcp_skb_cb) >
5198 sizeof_field(struct sk_buff, cb));
5199
5200 tcp_struct_check();
5201
5202 percpu_counter_init(&tcp_sockets_allocated, 0, GFP_KERNEL);
5203
5204 timer_setup(&tcp_orphan_timer, tcp_orphan_update, TIMER_DEFERRABLE);
5205 mod_timer(timer: &tcp_orphan_timer, expires: jiffies + TCP_ORPHAN_TIMER_PERIOD);
5206
5207 inet_hashinfo2_init(h: &tcp_hashinfo, name: "tcp_listen_portaddr_hash",
5208 numentries: thash_entries, scale: 21, /* one slot per 2 MB*/
5209 low_limit: 0, high_limit: 64 * 1024);
5210 tcp_hashinfo.bind_bucket_cachep =
5211 kmem_cache_create("tcp_bind_bucket",
5212 sizeof(struct inet_bind_bucket), 0,
5213 SLAB_HWCACHE_ALIGN | SLAB_PANIC |
5214 SLAB_ACCOUNT,
5215 NULL);
5216 tcp_hashinfo.bind2_bucket_cachep =
5217 kmem_cache_create("tcp_bind2_bucket",
5218 sizeof(struct inet_bind2_bucket), 0,
5219 SLAB_HWCACHE_ALIGN | SLAB_PANIC |
5220 SLAB_ACCOUNT,
5221 NULL);
5222
5223 /* Size and allocate the main established and bind bucket
5224 * hash tables.
5225 *
5226 * The methodology is similar to that of the buffer cache.
5227 */
5228 tcp_hashinfo.ehash =
5229 alloc_large_system_hash(tablename: "TCP established",
5230 bucketsize: sizeof(struct inet_ehash_bucket),
5231 numentries: thash_entries,
5232 scale: 17, /* one slot per 128 KB of memory */
5233 flags: 0,
5234 NULL,
5235 hash_mask: &tcp_hashinfo.ehash_mask,
5236 low_limit: 0,
5237 high_limit: thash_entries ? 0 : 512 * 1024);
5238 for (i = 0; i <= tcp_hashinfo.ehash_mask; i++)
5239 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].chain, i);
5240
5241 if (inet_ehash_locks_alloc(hashinfo: &tcp_hashinfo))
5242 panic(fmt: "TCP: failed to alloc ehash_locks");
5243 tcp_hashinfo.bhash =
5244 alloc_large_system_hash(tablename: "TCP bind",
5245 bucketsize: 2 * sizeof(struct inet_bind_hashbucket),
5246 numentries: tcp_hashinfo.ehash_mask + 1,
5247 scale: 17, /* one slot per 128 KB of memory */
5248 flags: 0,
5249 hash_shift: &tcp_hashinfo.bhash_size,
5250 NULL,
5251 low_limit: 0,
5252 high_limit: 64 * 1024);
5253 tcp_hashinfo.bhash_size = 1U << tcp_hashinfo.bhash_size;
5254 tcp_hashinfo.bhash2 = tcp_hashinfo.bhash + tcp_hashinfo.bhash_size;
5255 for (i = 0; i < tcp_hashinfo.bhash_size; i++) {
5256 spin_lock_init(&tcp_hashinfo.bhash[i].lock);
5257 INIT_HLIST_HEAD(&tcp_hashinfo.bhash[i].chain);
5258 spin_lock_init(&tcp_hashinfo.bhash2[i].lock);
5259 INIT_HLIST_HEAD(&tcp_hashinfo.bhash2[i].chain);
5260 }
5261
5262 tcp_hashinfo.pernet = false;
5263
5264 cnt = tcp_hashinfo.ehash_mask + 1;
5265 sysctl_tcp_max_orphans = cnt / 2;
5266
5267 tcp_init_mem();
5268 /* Set per-socket limits to no more than 1/128 the pressure threshold */
5269 limit = nr_free_buffer_pages() << (PAGE_SHIFT - 7);
5270 max_wshare = min(4UL*1024*1024, limit);
5271 max_rshare = min(32UL*1024*1024, limit);
5272
5273 init_net.ipv4.sysctl_tcp_wmem[0] = PAGE_SIZE;
5274 init_net.ipv4.sysctl_tcp_wmem[1] = 16*1024;
5275 init_net.ipv4.sysctl_tcp_wmem[2] = max(64*1024, max_wshare);
5276
5277 init_net.ipv4.sysctl_tcp_rmem[0] = PAGE_SIZE;
5278 init_net.ipv4.sysctl_tcp_rmem[1] = 131072;
5279 init_net.ipv4.sysctl_tcp_rmem[2] = max(131072, max_rshare);
5280
5281 pr_info("Hash tables configured (established %u bind %u)\n",
5282 tcp_hashinfo.ehash_mask + 1, tcp_hashinfo.bhash_size);
5283
5284 tcp_v4_init();
5285 tcp_metrics_init();
5286 BUG_ON(tcp_register_congestion_control(&tcp_reno) != 0);
5287 tcp_tsq_work_init();
5288 mptcp_init();
5289}
5290