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/md5.h> 247 #include <crypto/utils.h> 248 #include <linux/kernel.h> 249 #include <linux/module.h> 250 #include <linux/types.h> 251 #include <linux/fcntl.h> 252 #include <linux/poll.h> 253 #include <linux/inet_diag.h> 254 #include <linux/init.h> 255 #include <linux/fs.h> 256 #include <linux/skbuff.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. */ 294 enum { 295 TCP_CMSG_INQ = 1, 296 TCP_CMSG_TS = 2 297 }; 298 299 DEFINE_PER_CPU(unsigned int, tcp_orphan_count); 300 301 long sysctl_tcp_mem[3] __read_mostly; 302 303 DEFINE_PER_CPU(int, tcp_memory_per_cpu_fw_alloc); 304 EXPORT_PER_CPU_SYMBOL_GPL(tcp_memory_per_cpu_fw_alloc); 305 306 #if IS_ENABLED(CONFIG_SMC) 307 DEFINE_STATIC_KEY_FALSE(tcp_have_smc); 308 EXPORT_SYMBOL(tcp_have_smc); 309 #endif 310 311 /* 312 * Current number of TCP sockets. 313 */ 314 struct percpu_counter tcp_sockets_allocated ____cacheline_aligned_in_smp; 315 316 /* 317 * Pressure flag: try to collapse. 318 * Technical note: it is used by multiple contexts non atomically. 319 * All the __sk_mem_schedule() is of this nature: accounting 320 * is strict, actions are advisory and have some latency. 321 */ 322 unsigned long tcp_memory_pressure __read_mostly; 323 EXPORT_SYMBOL_GPL(tcp_memory_pressure); 324 325 void tcp_enter_memory_pressure(struct sock *sk) 326 { 327 unsigned long val; 328 329 if (READ_ONCE(tcp_memory_pressure)) 330 return; 331 val = jiffies; 332 333 if (!val) 334 val--; 335 if (!cmpxchg(&tcp_memory_pressure, 0, val)) 336 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMEMORYPRESSURES); 337 } 338 339 void tcp_leave_memory_pressure(struct sock *sk) 340 { 341 unsigned long val; 342 343 if (!READ_ONCE(tcp_memory_pressure)) 344 return; 345 val = xchg(&tcp_memory_pressure, 0); 346 if (val) 347 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPMEMORYPRESSURESCHRONO, 348 jiffies_to_msecs(jiffies - val)); 349 } 350 351 /* Convert seconds to retransmits based on initial and max timeout */ 352 static u8 secs_to_retrans(int seconds, int timeout, int rto_max) 353 { 354 u8 res = 0; 355 356 if (seconds > 0) { 357 int period = timeout; 358 359 res = 1; 360 while (seconds > period && res < 255) { 361 res++; 362 timeout <<= 1; 363 if (timeout > rto_max) 364 timeout = rto_max; 365 period += timeout; 366 } 367 } 368 return res; 369 } 370 371 /* Convert retransmits to seconds based on initial and max timeout */ 372 static int retrans_to_secs(u8 retrans, int timeout, int rto_max) 373 { 374 int period = 0; 375 376 if (retrans > 0) { 377 period = timeout; 378 while (--retrans) { 379 timeout <<= 1; 380 if (timeout > rto_max) 381 timeout = rto_max; 382 period += timeout; 383 } 384 } 385 return period; 386 } 387 388 static u64 tcp_compute_delivery_rate(const struct tcp_sock *tp) 389 { 390 u32 rate = READ_ONCE(tp->rate_delivered); 391 u32 intv = READ_ONCE(tp->rate_interval_us); 392 u64 rate64 = 0; 393 394 if (rate && intv) { 395 rate64 = (u64)rate * tp->mss_cache * USEC_PER_SEC; 396 do_div(rate64, intv); 397 } 398 return rate64; 399 } 400 401 #ifdef CONFIG_TCP_MD5SIG 402 void tcp_md5_destruct_sock(struct sock *sk) 403 { 404 struct tcp_sock *tp = tcp_sk(sk); 405 406 if (tp->md5sig_info) { 407 408 tcp_clear_md5_list(sk); 409 kfree(rcu_replace_pointer(tp->md5sig_info, NULL, 1)); 410 static_branch_slow_dec_deferred(&tcp_md5_needed); 411 } 412 } 413 #endif 414 415 /* Address-family independent initialization for a tcp_sock. 416 * 417 * NOTE: A lot of things set to zero explicitly by call to 418 * sk_alloc() so need not be done here. 419 */ 420 void tcp_init_sock(struct sock *sk) 421 { 422 struct inet_connection_sock *icsk = inet_csk(sk); 423 struct tcp_sock *tp = tcp_sk(sk); 424 int rto_min_us, rto_max_ms; 425 426 tp->out_of_order_queue = RB_ROOT; 427 sk->tcp_rtx_queue = RB_ROOT; 428 tcp_init_xmit_timers(sk); 429 INIT_LIST_HEAD(&tp->tsq_node); 430 INIT_LIST_HEAD(&tp->tsorted_sent_queue); 431 432 icsk->icsk_rto = TCP_TIMEOUT_INIT; 433 434 rto_max_ms = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_rto_max_ms); 435 icsk->icsk_rto_max = msecs_to_jiffies(rto_max_ms); 436 437 rto_min_us = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_rto_min_us); 438 icsk->icsk_rto_min = usecs_to_jiffies(rto_min_us); 439 icsk->icsk_delack_max = TCP_DELACK_MAX; 440 tp->mdev_us = jiffies_to_usecs(TCP_TIMEOUT_INIT); 441 minmax_reset(&tp->rtt_min, tcp_jiffies32, ~0U); 442 443 /* So many TCP implementations out there (incorrectly) count the 444 * initial SYN frame in their delayed-ACK and congestion control 445 * algorithms that we must have the following bandaid to talk 446 * efficiently to them. -DaveM 447 */ 448 tcp_snd_cwnd_set(tp, TCP_INIT_CWND); 449 450 /* There's a bubble in the pipe until at least the first ACK. */ 451 tp->app_limited = ~0U; 452 tp->rate_app_limited = 1; 453 454 /* See draft-stevens-tcpca-spec-01 for discussion of the 455 * initialization of these values. 456 */ 457 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH; 458 tp->snd_cwnd_clamp = ~0; 459 tp->mss_cache = TCP_MSS_DEFAULT; 460 461 tp->reordering = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_reordering); 462 tcp_assign_congestion_control(sk); 463 464 tp->tsoffset = 0; 465 tp->rack.reo_wnd_steps = 1; 466 467 sk->sk_write_space = sk_stream_write_space; 468 sock_set_flag(sk, SOCK_USE_WRITE_QUEUE); 469 470 icsk->icsk_sync_mss = tcp_sync_mss; 471 472 WRITE_ONCE(sk->sk_sndbuf, READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_wmem[1])); 473 WRITE_ONCE(sk->sk_rcvbuf, READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_rmem[1])); 474 tcp_scaling_ratio_init(sk); 475 476 set_bit(SOCK_SUPPORT_ZC, &sk->sk_socket->flags); 477 sk_sockets_allocated_inc(sk); 478 xa_init_flags(&sk->sk_user_frags, XA_FLAGS_ALLOC1); 479 } 480 481 static void tcp_tx_timestamp(struct sock *sk, struct sockcm_cookie *sockc) 482 { 483 struct sk_buff *skb = tcp_write_queue_tail(sk); 484 u32 tsflags = sockc->tsflags; 485 486 if (unlikely(!skb)) 487 skb = skb_rb_last(&sk->tcp_rtx_queue); 488 489 if (tsflags && skb) { 490 struct skb_shared_info *shinfo = skb_shinfo(skb); 491 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb); 492 493 sock_tx_timestamp(sk, sockc, &shinfo->tx_flags); 494 if (tsflags & SOF_TIMESTAMPING_TX_ACK) 495 tcb->txstamp_ack |= TSTAMP_ACK_SK; 496 if (tsflags & SOF_TIMESTAMPING_TX_RECORD_MASK) 497 shinfo->tskey = TCP_SKB_CB(skb)->seq + skb->len - 1; 498 } 499 500 if (cgroup_bpf_enabled(CGROUP_SOCK_OPS) && 501 SK_BPF_CB_FLAG_TEST(sk, SK_BPF_CB_TX_TIMESTAMPING) && skb) 502 bpf_skops_tx_timestamping(sk, skb, BPF_SOCK_OPS_TSTAMP_SENDMSG_CB); 503 } 504 505 /* @wake is one when sk_stream_write_space() calls us. 506 * This sends EPOLLOUT only if notsent_bytes is half the limit. 507 * This mimics the strategy used in sock_def_write_space(). 508 */ 509 bool tcp_stream_memory_free(const struct sock *sk, int wake) 510 { 511 const struct tcp_sock *tp = tcp_sk(sk); 512 u32 notsent_bytes = READ_ONCE(tp->write_seq) - READ_ONCE(tp->snd_nxt); 513 514 return (notsent_bytes << wake) < tcp_notsent_lowat(tp); 515 } 516 EXPORT_SYMBOL(tcp_stream_memory_free); 517 518 static bool tcp_stream_is_readable(struct sock *sk, int target) 519 { 520 if (tcp_epollin_ready(sk, target)) 521 return true; 522 return sk_is_readable(sk); 523 } 524 525 /* 526 * Wait for a TCP event. 527 * 528 * Note that we don't need to lock the socket, as the upper poll layers 529 * take care of normal races (between the test and the event) and we don't 530 * go look at any of the socket buffers directly. 531 */ 532 __poll_t tcp_poll(struct file *file, struct socket *sock, poll_table *wait) 533 { 534 __poll_t mask; 535 struct sock *sk = sock->sk; 536 const struct tcp_sock *tp = tcp_sk(sk); 537 u8 shutdown; 538 int state; 539 540 sock_poll_wait(file, sock, wait); 541 542 state = inet_sk_state_load(sk); 543 if (state == TCP_LISTEN) 544 return inet_csk_listen_poll(sk); 545 546 /* Socket is not locked. We are protected from async events 547 * by poll logic and correct handling of state changes 548 * made by other threads is impossible in any case. 549 */ 550 551 mask = 0; 552 553 /* 554 * EPOLLHUP is certainly not done right. But poll() doesn't 555 * have a notion of HUP in just one direction, and for a 556 * socket the read side is more interesting. 557 * 558 * Some poll() documentation says that EPOLLHUP is incompatible 559 * with the EPOLLOUT/POLLWR flags, so somebody should check this 560 * all. But careful, it tends to be safer to return too many 561 * bits than too few, and you can easily break real applications 562 * if you don't tell them that something has hung up! 563 * 564 * Check-me. 565 * 566 * Check number 1. EPOLLHUP is _UNMASKABLE_ event (see UNIX98 and 567 * our fs/select.c). It means that after we received EOF, 568 * poll always returns immediately, making impossible poll() on write() 569 * in state CLOSE_WAIT. One solution is evident --- to set EPOLLHUP 570 * if and only if shutdown has been made in both directions. 571 * Actually, it is interesting to look how Solaris and DUX 572 * solve this dilemma. I would prefer, if EPOLLHUP were maskable, 573 * then we could set it on SND_SHUTDOWN. BTW examples given 574 * in Stevens' books assume exactly this behaviour, it explains 575 * why EPOLLHUP is incompatible with EPOLLOUT. --ANK 576 * 577 * NOTE. Check for TCP_CLOSE is added. The goal is to prevent 578 * blocking on fresh not-connected or disconnected socket. --ANK 579 */ 580 shutdown = READ_ONCE(sk->sk_shutdown); 581 if (shutdown == SHUTDOWN_MASK || state == TCP_CLOSE) 582 mask |= EPOLLHUP; 583 if (shutdown & RCV_SHUTDOWN) 584 mask |= EPOLLIN | EPOLLRDNORM | EPOLLRDHUP; 585 586 /* Connected or passive Fast Open socket? */ 587 if (state != TCP_SYN_SENT && 588 (state != TCP_SYN_RECV || rcu_access_pointer(tp->fastopen_rsk))) { 589 int target = sock_rcvlowat(sk, 0, INT_MAX); 590 u16 urg_data = READ_ONCE(tp->urg_data); 591 592 if (unlikely(urg_data) && 593 READ_ONCE(tp->urg_seq) == READ_ONCE(tp->copied_seq) && 594 !sock_flag(sk, SOCK_URGINLINE)) 595 target++; 596 597 if (tcp_stream_is_readable(sk, target)) 598 mask |= EPOLLIN | EPOLLRDNORM; 599 600 if (!(shutdown & SEND_SHUTDOWN)) { 601 if (__sk_stream_is_writeable(sk, 1)) { 602 mask |= EPOLLOUT | EPOLLWRNORM; 603 } else { /* send SIGIO later */ 604 sk_set_bit(SOCKWQ_ASYNC_NOSPACE, sk); 605 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags); 606 607 /* Race breaker. If space is freed after 608 * wspace test but before the flags are set, 609 * IO signal will be lost. Memory barrier 610 * pairs with the input side. 611 */ 612 smp_mb__after_atomic(); 613 if (__sk_stream_is_writeable(sk, 1)) 614 mask |= EPOLLOUT | EPOLLWRNORM; 615 } 616 } else 617 mask |= EPOLLOUT | EPOLLWRNORM; 618 619 if (urg_data & TCP_URG_VALID) 620 mask |= EPOLLPRI; 621 } else if (state == TCP_SYN_SENT && 622 inet_test_bit(DEFER_CONNECT, sk)) { 623 /* Active TCP fastopen socket with defer_connect 624 * Return EPOLLOUT so application can call write() 625 * in order for kernel to generate SYN+data 626 */ 627 mask |= EPOLLOUT | EPOLLWRNORM; 628 } 629 /* This barrier is coupled with smp_wmb() in tcp_done_with_error() */ 630 smp_rmb(); 631 if (READ_ONCE(sk->sk_err) || 632 !skb_queue_empty_lockless(&sk->sk_error_queue)) 633 mask |= EPOLLERR; 634 635 return mask; 636 } 637 EXPORT_SYMBOL(tcp_poll); 638 639 int tcp_ioctl(struct sock *sk, int cmd, int *karg) 640 { 641 struct tcp_sock *tp = tcp_sk(sk); 642 int answ; 643 bool slow; 644 645 switch (cmd) { 646 case SIOCINQ: 647 if (sk->sk_state == TCP_LISTEN) 648 return -EINVAL; 649 650 slow = lock_sock_fast(sk); 651 answ = tcp_inq(sk); 652 unlock_sock_fast(sk, slow); 653 break; 654 case SIOCATMARK: 655 answ = READ_ONCE(tp->urg_data) && 656 READ_ONCE(tp->urg_seq) == READ_ONCE(tp->copied_seq); 657 break; 658 case SIOCOUTQ: 659 if (sk->sk_state == TCP_LISTEN) 660 return -EINVAL; 661 662 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV)) 663 answ = 0; 664 else 665 answ = READ_ONCE(tp->write_seq) - tp->snd_una; 666 break; 667 case SIOCOUTQNSD: 668 if (sk->sk_state == TCP_LISTEN) 669 return -EINVAL; 670 671 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV)) 672 answ = 0; 673 else 674 answ = READ_ONCE(tp->write_seq) - 675 READ_ONCE(tp->snd_nxt); 676 break; 677 default: 678 return -ENOIOCTLCMD; 679 } 680 681 *karg = answ; 682 return 0; 683 } 684 685 void tcp_mark_push(struct tcp_sock *tp, struct sk_buff *skb) 686 { 687 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; 688 tp->pushed_seq = tp->write_seq; 689 } 690 691 static inline bool forced_push(const struct tcp_sock *tp) 692 { 693 return after(tp->write_seq, tp->pushed_seq + (tp->max_window >> 1)); 694 } 695 696 void tcp_skb_entail(struct sock *sk, struct sk_buff *skb) 697 { 698 struct tcp_sock *tp = tcp_sk(sk); 699 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb); 700 701 tcb->seq = tcb->end_seq = tp->write_seq; 702 tcb->tcp_flags = TCPHDR_ACK; 703 __skb_header_release(skb); 704 psp_enqueue_set_decrypted(sk, skb); 705 tcp_add_write_queue_tail(sk, skb); 706 sk_wmem_queued_add(sk, skb->truesize); 707 sk_mem_charge(sk, skb->truesize); 708 if (tp->nonagle & TCP_NAGLE_PUSH) 709 tp->nonagle &= ~TCP_NAGLE_PUSH; 710 711 tcp_slow_start_after_idle_check(sk); 712 } 713 714 static inline void tcp_mark_urg(struct tcp_sock *tp, int flags) 715 { 716 if (flags & MSG_OOB) 717 tp->snd_up = tp->write_seq; 718 } 719 720 /* If a not yet filled skb is pushed, do not send it if 721 * we have data packets in Qdisc or NIC queues : 722 * Because TX completion will happen shortly, it gives a chance 723 * to coalesce future sendmsg() payload into this skb, without 724 * need for a timer, and with no latency trade off. 725 * As packets containing data payload have a bigger truesize 726 * than pure acks (dataless) packets, the last checks prevent 727 * autocorking if we only have an ACK in Qdisc/NIC queues, 728 * or if TX completion was delayed after we processed ACK packet. 729 */ 730 static bool tcp_should_autocork(struct sock *sk, struct sk_buff *skb, 731 int size_goal) 732 { 733 return skb->len < size_goal && 734 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_autocorking) && 735 !tcp_rtx_queue_empty(sk) && 736 refcount_read(&sk->sk_wmem_alloc) > skb->truesize && 737 tcp_skb_can_collapse_to(skb); 738 } 739 740 void tcp_push(struct sock *sk, int flags, int mss_now, 741 int nonagle, int size_goal) 742 { 743 struct tcp_sock *tp = tcp_sk(sk); 744 struct sk_buff *skb; 745 746 skb = tcp_write_queue_tail(sk); 747 if (!skb) 748 return; 749 if (!(flags & MSG_MORE) || forced_push(tp)) 750 tcp_mark_push(tp, skb); 751 752 tcp_mark_urg(tp, flags); 753 754 if (tcp_should_autocork(sk, skb, size_goal)) { 755 756 /* avoid atomic op if TSQ_THROTTLED bit is already set */ 757 if (!test_bit(TSQ_THROTTLED, &sk->sk_tsq_flags)) { 758 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPAUTOCORKING); 759 set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags); 760 smp_mb__after_atomic(); 761 } 762 /* It is possible TX completion already happened 763 * before we set TSQ_THROTTLED. 764 */ 765 if (refcount_read(&sk->sk_wmem_alloc) > skb->truesize) 766 return; 767 } 768 769 if (flags & MSG_MORE) 770 nonagle = TCP_NAGLE_CORK; 771 772 __tcp_push_pending_frames(sk, mss_now, nonagle); 773 } 774 775 int tcp_splice_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb, 776 unsigned int offset, size_t len) 777 { 778 struct tcp_splice_state *tss = rd_desc->arg.data; 779 int ret; 780 781 ret = skb_splice_bits(skb, skb->sk, offset, tss->pipe, 782 min(rd_desc->count, len), tss->flags); 783 if (ret > 0) 784 rd_desc->count -= ret; 785 return ret; 786 } 787 788 static int __tcp_splice_read(struct sock *sk, struct tcp_splice_state *tss) 789 { 790 /* Store TCP splice context information in read_descriptor_t. */ 791 read_descriptor_t rd_desc = { 792 .arg.data = tss, 793 .count = tss->len, 794 }; 795 796 return tcp_read_sock(sk, &rd_desc, tcp_splice_data_recv); 797 } 798 799 /** 800 * tcp_splice_read - splice data from TCP socket to a pipe 801 * @sock: socket to splice from 802 * @ppos: position (not valid) 803 * @pipe: pipe to splice to 804 * @len: number of bytes to splice 805 * @flags: splice modifier flags 806 * 807 * Description: 808 * Will read pages from given socket and fill them into a pipe. 809 * 810 **/ 811 ssize_t tcp_splice_read(struct socket *sock, loff_t *ppos, 812 struct pipe_inode_info *pipe, size_t len, 813 unsigned int flags) 814 { 815 struct sock *sk = sock->sk; 816 struct tcp_splice_state tss = { 817 .pipe = pipe, 818 .len = len, 819 .flags = flags, 820 }; 821 long timeo; 822 ssize_t spliced; 823 int ret; 824 825 sock_rps_record_flow(sk); 826 /* 827 * We can't seek on a socket input 828 */ 829 if (unlikely(*ppos)) 830 return -ESPIPE; 831 832 ret = spliced = 0; 833 834 lock_sock(sk); 835 836 timeo = sock_rcvtimeo(sk, sock->file->f_flags & O_NONBLOCK); 837 while (tss.len) { 838 ret = __tcp_splice_read(sk, &tss); 839 if (ret < 0) 840 break; 841 else if (!ret) { 842 if (spliced) 843 break; 844 if (sock_flag(sk, SOCK_DONE)) 845 break; 846 if (sk->sk_err) { 847 ret = sock_error(sk); 848 break; 849 } 850 if (sk->sk_shutdown & RCV_SHUTDOWN) 851 break; 852 if (sk->sk_state == TCP_CLOSE) { 853 /* 854 * This occurs when user tries to read 855 * from never connected socket. 856 */ 857 ret = -ENOTCONN; 858 break; 859 } 860 if (!timeo) { 861 ret = -EAGAIN; 862 break; 863 } 864 /* if __tcp_splice_read() got nothing while we have 865 * an skb in receive queue, we do not want to loop. 866 * This might happen with URG data. 867 */ 868 if (!skb_queue_empty(&sk->sk_receive_queue)) 869 break; 870 ret = sk_wait_data(sk, &timeo, NULL); 871 if (ret < 0) 872 break; 873 if (signal_pending(current)) { 874 ret = sock_intr_errno(timeo); 875 break; 876 } 877 continue; 878 } 879 tss.len -= ret; 880 spliced += ret; 881 882 if (!tss.len || !timeo) 883 break; 884 release_sock(sk); 885 lock_sock(sk); 886 887 if (tcp_recv_should_stop(sk)) 888 break; 889 } 890 891 release_sock(sk); 892 893 if (spliced) 894 return spliced; 895 896 return ret; 897 } 898 899 /* We allow to exceed memory limits for FIN packets to expedite 900 * connection tear down and (memory) recovery. 901 * Otherwise tcp_send_fin() could be tempted to either delay FIN 902 * or even be forced to close flow without any FIN. 903 * In general, we want to allow one skb per socket to avoid hangs 904 * with edge trigger epoll() 905 */ 906 void sk_forced_mem_schedule(struct sock *sk, int size) 907 { 908 int delta, amt; 909 910 delta = size - sk->sk_forward_alloc; 911 if (delta <= 0) 912 return; 913 914 amt = sk_mem_pages(delta); 915 sk_forward_alloc_add(sk, amt << PAGE_SHIFT); 916 917 if (mem_cgroup_sk_enabled(sk)) 918 mem_cgroup_sk_charge(sk, amt, gfp_memcg_charge() | __GFP_NOFAIL); 919 920 if (sk->sk_bypass_prot_mem) 921 return; 922 923 sk_memory_allocated_add(sk, amt); 924 } 925 926 struct sk_buff *tcp_stream_alloc_skb(struct sock *sk, gfp_t gfp, 927 bool force_schedule) 928 { 929 struct sk_buff *skb; 930 931 skb = alloc_skb_fclone(MAX_TCP_HEADER, gfp); 932 if (likely(skb)) { 933 bool mem_scheduled; 934 935 skb->truesize = SKB_TRUESIZE(skb_end_offset(skb)); 936 if (force_schedule) { 937 mem_scheduled = true; 938 sk_forced_mem_schedule(sk, skb->truesize); 939 } else { 940 mem_scheduled = sk_wmem_schedule(sk, skb->truesize); 941 } 942 if (likely(mem_scheduled)) { 943 skb_reserve(skb, MAX_TCP_HEADER); 944 skb->ip_summed = CHECKSUM_PARTIAL; 945 INIT_LIST_HEAD(&skb->tcp_tsorted_anchor); 946 return skb; 947 } 948 __kfree_skb(skb); 949 } else { 950 if (!sk->sk_bypass_prot_mem) 951 tcp_enter_memory_pressure(sk); 952 sk_stream_moderate_sndbuf(sk); 953 } 954 return NULL; 955 } 956 957 static unsigned int tcp_xmit_size_goal(struct sock *sk, u32 mss_now, 958 int large_allowed) 959 { 960 struct tcp_sock *tp = tcp_sk(sk); 961 u32 new_size_goal, size_goal; 962 963 if (!large_allowed) 964 return mss_now; 965 966 /* Note : tcp_tso_autosize() will eventually split this later */ 967 new_size_goal = tcp_bound_to_half_wnd(tp, sk->sk_gso_max_size); 968 969 /* We try hard to avoid divides here */ 970 size_goal = tp->gso_segs * mss_now; 971 if (unlikely(new_size_goal < size_goal || 972 new_size_goal >= size_goal + mss_now)) { 973 tp->gso_segs = min_t(u16, new_size_goal / mss_now, 974 sk->sk_gso_max_segs); 975 size_goal = tp->gso_segs * mss_now; 976 } 977 978 return max(size_goal, mss_now); 979 } 980 981 int tcp_send_mss(struct sock *sk, int *size_goal, int flags) 982 { 983 int mss_now; 984 985 mss_now = tcp_current_mss(sk); 986 *size_goal = tcp_xmit_size_goal(sk, mss_now, !(flags & MSG_OOB)); 987 988 return mss_now; 989 } 990 991 /* In some cases, sendmsg() could have added an skb to the write queue, 992 * but failed adding payload on it. We need to remove it to consume less 993 * memory, but more importantly be able to generate EPOLLOUT for Edge Trigger 994 * epoll() users. Another reason is that tcp_write_xmit() does not like 995 * finding an empty skb in the write queue. 996 */ 997 void tcp_remove_empty_skb(struct sock *sk) 998 { 999 struct sk_buff *skb = tcp_write_queue_tail(sk); 1000 1001 if (skb && TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq) { 1002 tcp_unlink_write_queue(skb, sk); 1003 if (tcp_write_queue_empty(sk)) 1004 tcp_chrono_stop(sk, TCP_CHRONO_BUSY); 1005 tcp_wmem_free_skb(sk, skb); 1006 } 1007 } 1008 1009 /* skb changing from pure zc to mixed, must charge zc */ 1010 static int tcp_downgrade_zcopy_pure(struct sock *sk, struct sk_buff *skb) 1011 { 1012 if (unlikely(skb_zcopy_pure(skb))) { 1013 u32 extra = skb->truesize - 1014 SKB_TRUESIZE(skb_end_offset(skb)); 1015 1016 if (!sk_wmem_schedule(sk, extra)) 1017 return -ENOMEM; 1018 1019 sk_mem_charge(sk, extra); 1020 skb_shinfo(skb)->flags &= ~SKBFL_PURE_ZEROCOPY; 1021 } 1022 return 0; 1023 } 1024 1025 1026 int tcp_wmem_schedule(struct sock *sk, int copy) 1027 { 1028 int left; 1029 1030 if (likely(sk_wmem_schedule(sk, copy))) 1031 return copy; 1032 1033 /* We could be in trouble if we have nothing queued. 1034 * Use whatever is left in sk->sk_forward_alloc and tcp_wmem[0] 1035 * to guarantee some progress. 1036 */ 1037 left = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_wmem[0]) - sk->sk_wmem_queued; 1038 if (left > 0) 1039 sk_forced_mem_schedule(sk, min(left, copy)); 1040 return min(copy, sk->sk_forward_alloc); 1041 } 1042 1043 void tcp_free_fastopen_req(struct tcp_sock *tp) 1044 { 1045 if (tp->fastopen_req) { 1046 kfree(tp->fastopen_req); 1047 tp->fastopen_req = NULL; 1048 } 1049 } 1050 1051 int tcp_sendmsg_fastopen(struct sock *sk, struct msghdr *msg, int *copied, 1052 size_t size, struct ubuf_info *uarg) 1053 { 1054 struct tcp_sock *tp = tcp_sk(sk); 1055 struct inet_sock *inet = inet_sk(sk); 1056 struct sockaddr *uaddr = msg->msg_name; 1057 int err, flags; 1058 1059 if (!(READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_fastopen) & 1060 TFO_CLIENT_ENABLE) || 1061 (uaddr && msg->msg_namelen >= sizeof(uaddr->sa_family) && 1062 uaddr->sa_family == AF_UNSPEC)) 1063 return -EOPNOTSUPP; 1064 if (tp->fastopen_req) 1065 return -EALREADY; /* Another Fast Open is in progress */ 1066 1067 tp->fastopen_req = kzalloc_obj(struct tcp_fastopen_request, 1068 sk->sk_allocation); 1069 if (unlikely(!tp->fastopen_req)) 1070 return -ENOBUFS; 1071 tp->fastopen_req->data = msg; 1072 tp->fastopen_req->size = size; 1073 tp->fastopen_req->uarg = uarg; 1074 1075 if (inet_test_bit(DEFER_CONNECT, sk)) { 1076 err = tcp_connect(sk); 1077 /* Same failure procedure as in tcp_v4/6_connect */ 1078 if (err) { 1079 tcp_set_state(sk, TCP_CLOSE); 1080 inet->inet_dport = 0; 1081 sk->sk_route_caps = 0; 1082 } 1083 } 1084 flags = (msg->msg_flags & MSG_DONTWAIT) ? O_NONBLOCK : 0; 1085 err = __inet_stream_connect(sk->sk_socket, (struct sockaddr_unsized *)uaddr, 1086 msg->msg_namelen, flags, 1); 1087 /* fastopen_req could already be freed in __inet_stream_connect 1088 * if the connection times out or gets rst 1089 */ 1090 if (tp->fastopen_req) { 1091 *copied = tp->fastopen_req->copied; 1092 tcp_free_fastopen_req(tp); 1093 inet_clear_bit(DEFER_CONNECT, sk); 1094 } 1095 return err; 1096 } 1097 1098 /* If a gap is detected between sends, mark the socket application-limited. */ 1099 void tcp_rate_check_app_limited(struct sock *sk) 1100 { 1101 struct tcp_sock *tp = tcp_sk(sk); 1102 1103 if (/* We have less than one packet to send. */ 1104 tp->write_seq - tp->snd_nxt < tp->mss_cache && 1105 /* Nothing in sending host's qdisc queues or NIC tx queue. */ 1106 sk_wmem_alloc_get(sk) < SKB_TRUESIZE(1) && 1107 /* We are not limited by CWND. */ 1108 tcp_packets_in_flight(tp) < tcp_snd_cwnd(tp) && 1109 /* All lost packets have been retransmitted. */ 1110 tp->lost_out <= tp->retrans_out) 1111 tp->app_limited = 1112 (tp->delivered + tcp_packets_in_flight(tp)) ? : 1; 1113 } 1114 EXPORT_SYMBOL_GPL(tcp_rate_check_app_limited); 1115 1116 int tcp_sendmsg_locked(struct sock *sk, struct msghdr *msg, size_t size) 1117 { 1118 struct net_devmem_dmabuf_binding *binding = NULL; 1119 struct tcp_sock *tp = tcp_sk(sk); 1120 struct ubuf_info *uarg = NULL; 1121 struct sk_buff *skb; 1122 struct sockcm_cookie sockc; 1123 int flags, err, copied = 0; 1124 int mss_now = 0, size_goal, copied_syn = 0; 1125 int process_backlog = 0; 1126 int sockc_err = 0; 1127 int zc = 0; 1128 long timeo; 1129 1130 flags = msg->msg_flags; 1131 1132 sockc = (struct sockcm_cookie){ .tsflags = READ_ONCE(sk->sk_tsflags) }; 1133 if (msg->msg_controllen) { 1134 sockc_err = sock_cmsg_send(sk, msg, &sockc); 1135 /* Don't return error until MSG_FASTOPEN has been processed; 1136 * that may succeed even if the cmsg is invalid. 1137 */ 1138 } 1139 1140 if ((flags & MSG_ZEROCOPY) && size) { 1141 if (msg->msg_ubuf) { 1142 uarg = msg->msg_ubuf; 1143 if (sk->sk_route_caps & NETIF_F_SG) 1144 zc = MSG_ZEROCOPY; 1145 } else if (sock_flag(sk, SOCK_ZEROCOPY)) { 1146 skb = tcp_write_queue_tail(sk); 1147 uarg = msg_zerocopy_realloc(sk, size, skb_zcopy(skb), 1148 !sockc_err && sockc.dmabuf_id); 1149 if (!uarg) { 1150 err = -ENOBUFS; 1151 goto out_err; 1152 } 1153 if (sk->sk_route_caps & NETIF_F_SG) 1154 zc = MSG_ZEROCOPY; 1155 else 1156 uarg_to_msgzc(uarg)->zerocopy = 0; 1157 1158 if (!sockc_err && sockc.dmabuf_id) { 1159 binding = net_devmem_get_binding(sk, sockc.dmabuf_id); 1160 if (IS_ERR(binding)) { 1161 err = PTR_ERR(binding); 1162 binding = NULL; 1163 goto out_err; 1164 } 1165 } 1166 } 1167 } else if (unlikely(msg->msg_flags & MSG_SPLICE_PAGES) && size) { 1168 if (sk->sk_route_caps & NETIF_F_SG) 1169 zc = MSG_SPLICE_PAGES; 1170 } 1171 1172 if (!sockc_err && sockc.dmabuf_id && 1173 (!(flags & MSG_ZEROCOPY) || !sock_flag(sk, SOCK_ZEROCOPY))) { 1174 err = -EINVAL; 1175 goto out_err; 1176 } 1177 1178 if (unlikely(flags & MSG_FASTOPEN || 1179 inet_test_bit(DEFER_CONNECT, sk)) && 1180 !tp->repair) { 1181 err = tcp_sendmsg_fastopen(sk, msg, &copied_syn, size, uarg); 1182 if (err == -EINPROGRESS && copied_syn > 0) 1183 goto out; 1184 else if (err) 1185 goto out_err; 1186 } 1187 1188 timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT); 1189 1190 tcp_rate_check_app_limited(sk); /* is sending application-limited? */ 1191 1192 /* Wait for a connection to finish. One exception is TCP Fast Open 1193 * (passive side) where data is allowed to be sent before a connection 1194 * is fully established. 1195 */ 1196 if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) && 1197 !tcp_passive_fastopen(sk)) { 1198 err = sk_stream_wait_connect(sk, &timeo); 1199 if (err != 0) 1200 goto do_error; 1201 } 1202 1203 if (unlikely(tp->repair)) { 1204 if (tp->repair_queue == TCP_RECV_QUEUE) { 1205 copied = tcp_send_rcvq(sk, msg, size); 1206 goto out_nopush; 1207 } 1208 1209 err = -EINVAL; 1210 if (tp->repair_queue == TCP_NO_QUEUE) 1211 goto out_err; 1212 1213 /* 'common' sending to sendq */ 1214 } 1215 1216 if (sockc_err) { 1217 err = sockc_err; 1218 goto out_err; 1219 } 1220 1221 /* This should be in poll */ 1222 sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk); 1223 1224 /* Ok commence sending. */ 1225 copied = 0; 1226 1227 restart: 1228 mss_now = tcp_send_mss(sk, &size_goal, flags); 1229 1230 err = -EPIPE; 1231 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN)) 1232 goto do_error; 1233 1234 while (msg_data_left(msg)) { 1235 int copy = 0; 1236 1237 skb = tcp_write_queue_tail(sk); 1238 if (skb) 1239 copy = size_goal - skb->len; 1240 1241 trace_tcp_sendmsg_locked(sk, msg, skb, size_goal); 1242 1243 if (copy <= 0 || !tcp_skb_can_collapse_to(skb)) { 1244 bool first_skb; 1245 1246 new_segment: 1247 if (!sk_stream_memory_free(sk)) 1248 goto wait_for_space; 1249 1250 if (unlikely(process_backlog >= 16)) { 1251 process_backlog = 0; 1252 if (sk_flush_backlog(sk)) 1253 goto restart; 1254 } 1255 first_skb = tcp_rtx_and_write_queues_empty(sk); 1256 skb = tcp_stream_alloc_skb(sk, sk->sk_allocation, 1257 first_skb); 1258 if (!skb) 1259 goto wait_for_space; 1260 1261 process_backlog++; 1262 1263 #ifdef CONFIG_SKB_DECRYPTED 1264 skb->decrypted = !!(flags & MSG_SENDPAGE_DECRYPTED); 1265 #endif 1266 tcp_skb_entail(sk, skb); 1267 copy = size_goal; 1268 1269 /* All packets are restored as if they have 1270 * already been sent. skb_mstamp_ns isn't set to 1271 * avoid wrong rtt estimation. 1272 */ 1273 if (tp->repair) 1274 TCP_SKB_CB(skb)->sacked |= TCPCB_REPAIRED; 1275 } 1276 1277 /* Try to append data to the end of skb. */ 1278 if (copy > msg_data_left(msg)) 1279 copy = msg_data_left(msg); 1280 1281 if (zc == 0) { 1282 bool merge = true; 1283 int i = skb_shinfo(skb)->nr_frags; 1284 struct page_frag *pfrag = sk_page_frag(sk); 1285 1286 if (!sk_page_frag_refill(sk, pfrag)) 1287 goto wait_for_space; 1288 1289 if (!skb_can_coalesce(skb, i, pfrag->page, 1290 pfrag->offset)) { 1291 if (i >= READ_ONCE(net_hotdata.sysctl_max_skb_frags)) { 1292 tcp_mark_push(tp, skb); 1293 goto new_segment; 1294 } 1295 merge = false; 1296 } 1297 1298 copy = min_t(int, copy, pfrag->size - pfrag->offset); 1299 1300 if (unlikely(skb_zcopy_pure(skb) || skb_zcopy_managed(skb))) { 1301 if (tcp_downgrade_zcopy_pure(sk, skb)) 1302 goto wait_for_space; 1303 skb_zcopy_downgrade_managed(skb); 1304 } 1305 1306 copy = tcp_wmem_schedule(sk, copy); 1307 if (!copy) 1308 goto wait_for_space; 1309 1310 err = skb_copy_to_page_nocache(sk, &msg->msg_iter, skb, 1311 pfrag->page, 1312 pfrag->offset, 1313 copy); 1314 if (err) 1315 goto do_error; 1316 1317 /* Update the skb. */ 1318 if (merge) { 1319 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy); 1320 } else { 1321 skb_fill_page_desc(skb, i, pfrag->page, 1322 pfrag->offset, copy); 1323 page_ref_inc(pfrag->page); 1324 } 1325 pfrag->offset += copy; 1326 } else if (zc == MSG_ZEROCOPY) { 1327 /* First append to a fragless skb builds initial 1328 * pure zerocopy skb 1329 */ 1330 if (!skb->len) 1331 skb_shinfo(skb)->flags |= SKBFL_PURE_ZEROCOPY; 1332 1333 if (!skb_zcopy_pure(skb)) { 1334 copy = tcp_wmem_schedule(sk, copy); 1335 if (!copy) 1336 goto wait_for_space; 1337 } 1338 1339 err = skb_zerocopy_iter_stream(sk, skb, msg, copy, uarg, 1340 binding); 1341 if (err == -EMSGSIZE || err == -EEXIST) { 1342 tcp_mark_push(tp, skb); 1343 goto new_segment; 1344 } 1345 if (err < 0) 1346 goto do_error; 1347 copy = err; 1348 } else if (zc == MSG_SPLICE_PAGES) { 1349 /* Splice in data if we can; copy if we can't. */ 1350 if (tcp_downgrade_zcopy_pure(sk, skb)) 1351 goto wait_for_space; 1352 copy = tcp_wmem_schedule(sk, copy); 1353 if (!copy) 1354 goto wait_for_space; 1355 1356 err = skb_splice_from_iter(skb, &msg->msg_iter, copy); 1357 if (err < 0) { 1358 if (err == -EMSGSIZE) { 1359 tcp_mark_push(tp, skb); 1360 goto new_segment; 1361 } 1362 goto do_error; 1363 } 1364 copy = err; 1365 1366 if (!(flags & MSG_NO_SHARED_FRAGS)) 1367 skb_shinfo(skb)->flags |= SKBFL_SHARED_FRAG; 1368 1369 sk_wmem_queued_add(sk, copy); 1370 sk_mem_charge(sk, copy); 1371 } 1372 1373 if (!copied) 1374 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH; 1375 1376 WRITE_ONCE(tp->write_seq, tp->write_seq + copy); 1377 TCP_SKB_CB(skb)->end_seq += copy; 1378 tcp_skb_pcount_set(skb, 0); 1379 1380 copied += copy; 1381 if (!msg_data_left(msg)) { 1382 if (unlikely(flags & MSG_EOR)) 1383 TCP_SKB_CB(skb)->eor = 1; 1384 goto out; 1385 } 1386 1387 if (skb->len < size_goal || (flags & MSG_OOB) || unlikely(tp->repair)) 1388 continue; 1389 1390 if (forced_push(tp)) { 1391 tcp_mark_push(tp, skb); 1392 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH); 1393 } else if (skb == tcp_send_head(sk)) 1394 tcp_push_one(sk, mss_now); 1395 continue; 1396 1397 wait_for_space: 1398 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags); 1399 tcp_remove_empty_skb(sk); 1400 if (copied) 1401 tcp_push(sk, flags & ~MSG_MORE, mss_now, 1402 TCP_NAGLE_PUSH, size_goal); 1403 1404 err = sk_stream_wait_memory(sk, &timeo); 1405 if (err != 0) 1406 goto do_error; 1407 1408 mss_now = tcp_send_mss(sk, &size_goal, flags); 1409 } 1410 1411 out: 1412 if (copied) { 1413 tcp_tx_timestamp(sk, &sockc); 1414 tcp_push(sk, flags, mss_now, tp->nonagle, size_goal); 1415 } 1416 out_nopush: 1417 /* msg->msg_ubuf is pinned by the caller so we don't take extra refs */ 1418 if (uarg && !msg->msg_ubuf) 1419 net_zcopy_put(uarg); 1420 if (binding) 1421 net_devmem_dmabuf_binding_put(binding); 1422 return copied + copied_syn; 1423 1424 do_error: 1425 tcp_remove_empty_skb(sk); 1426 1427 if (copied + copied_syn) 1428 goto out; 1429 out_err: 1430 /* msg->msg_ubuf is pinned by the caller so we don't take extra refs */ 1431 if (uarg && !msg->msg_ubuf) 1432 net_zcopy_put_abort(uarg, true); 1433 err = sk_stream_error(sk, flags, err); 1434 /* make sure we wake any epoll edge trigger waiter */ 1435 if (unlikely(tcp_rtx_and_write_queues_empty(sk) && err == -EAGAIN)) { 1436 READ_ONCE(sk->sk_write_space)(sk); 1437 tcp_chrono_stop(sk, TCP_CHRONO_SNDBUF_LIMITED); 1438 } 1439 if (binding) 1440 net_devmem_dmabuf_binding_put(binding); 1441 1442 return err; 1443 } 1444 EXPORT_SYMBOL_GPL(tcp_sendmsg_locked); 1445 1446 int tcp_sendmsg(struct sock *sk, struct msghdr *msg, size_t size) 1447 { 1448 int ret; 1449 1450 lock_sock(sk); 1451 ret = tcp_sendmsg_locked(sk, msg, size); 1452 release_sock(sk); 1453 1454 return ret; 1455 } 1456 EXPORT_SYMBOL(tcp_sendmsg); 1457 1458 void tcp_splice_eof(struct socket *sock) 1459 { 1460 struct sock *sk = sock->sk; 1461 struct tcp_sock *tp = tcp_sk(sk); 1462 int mss_now, size_goal; 1463 1464 if (!tcp_write_queue_tail(sk)) 1465 return; 1466 1467 lock_sock(sk); 1468 mss_now = tcp_send_mss(sk, &size_goal, 0); 1469 tcp_push(sk, 0, mss_now, tp->nonagle, size_goal); 1470 release_sock(sk); 1471 } 1472 1473 /* 1474 * Handle reading urgent data. BSD has very simple semantics for 1475 * this, no blocking and very strange errors 8) 1476 */ 1477 1478 static int tcp_recv_urg(struct sock *sk, struct msghdr *msg, int len, int flags) 1479 { 1480 struct tcp_sock *tp = tcp_sk(sk); 1481 1482 /* No URG data to read. */ 1483 if (sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data || 1484 tp->urg_data == TCP_URG_READ) 1485 return -EINVAL; /* Yes this is right ! */ 1486 1487 if (sk->sk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DONE)) 1488 return -ENOTCONN; 1489 1490 if (tp->urg_data & TCP_URG_VALID) { 1491 int err = 0; 1492 char c = tp->urg_data; 1493 1494 if (!(flags & MSG_PEEK)) 1495 WRITE_ONCE(tp->urg_data, TCP_URG_READ); 1496 1497 /* Read urgent data. */ 1498 msg->msg_flags |= MSG_OOB; 1499 1500 if (len > 0) { 1501 if (!(flags & MSG_TRUNC)) 1502 err = memcpy_to_msg(msg, &c, 1); 1503 len = 1; 1504 } else 1505 msg->msg_flags |= MSG_TRUNC; 1506 1507 return err ? -EFAULT : len; 1508 } 1509 1510 if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN)) 1511 return 0; 1512 1513 /* Fixed the recv(..., MSG_OOB) behaviour. BSD docs and 1514 * the available implementations agree in this case: 1515 * this call should never block, independent of the 1516 * blocking state of the socket. 1517 * Mike <pall@rz.uni-karlsruhe.de> 1518 */ 1519 return -EAGAIN; 1520 } 1521 1522 static int tcp_peek_sndq(struct sock *sk, struct msghdr *msg, int len) 1523 { 1524 struct sk_buff *skb; 1525 int copied = 0, err = 0; 1526 1527 skb_rbtree_walk(skb, &sk->tcp_rtx_queue) { 1528 err = skb_copy_datagram_msg(skb, 0, msg, skb->len); 1529 if (err) 1530 return err; 1531 copied += skb->len; 1532 } 1533 1534 skb_queue_walk(&sk->sk_write_queue, skb) { 1535 err = skb_copy_datagram_msg(skb, 0, msg, skb->len); 1536 if (err) 1537 break; 1538 1539 copied += skb->len; 1540 } 1541 1542 return err ?: copied; 1543 } 1544 1545 /* Clean up the receive buffer for full frames taken by the user, 1546 * then send an ACK if necessary. COPIED is the number of bytes 1547 * tcp_recvmsg has given to the user so far, it speeds up the 1548 * calculation of whether or not we must ACK for the sake of 1549 * a window update. 1550 */ 1551 void __tcp_cleanup_rbuf(struct sock *sk, int copied) 1552 { 1553 struct tcp_sock *tp = tcp_sk(sk); 1554 bool time_to_ack = false; 1555 1556 if (inet_csk_ack_scheduled(sk)) { 1557 const struct inet_connection_sock *icsk = inet_csk(sk); 1558 1559 if (/* Once-per-two-segments ACK was not sent by tcp_input.c */ 1560 tp->rcv_nxt - tp->rcv_wup > icsk->icsk_ack.rcv_mss || 1561 /* 1562 * If this read emptied read buffer, we send ACK, if 1563 * connection is not bidirectional, user drained 1564 * receive buffer and there was a small segment 1565 * in queue. 1566 */ 1567 (copied > 0 && 1568 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED2) || 1569 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED) && 1570 !inet_csk_in_pingpong_mode(sk))) && 1571 !atomic_read(&sk->sk_rmem_alloc))) 1572 time_to_ack = true; 1573 } 1574 1575 /* We send an ACK if we can now advertise a non-zero window 1576 * which has been raised "significantly". 1577 * 1578 * Even if window raised up to infinity, do not send window open ACK 1579 * in states, where we will not receive more. It is useless. 1580 */ 1581 if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) { 1582 __u32 rcv_window_now = tcp_receive_window(tp); 1583 1584 /* Optimize, __tcp_select_window() is not cheap. */ 1585 if (2*rcv_window_now <= tp->window_clamp) { 1586 __u32 new_window = __tcp_select_window(sk); 1587 1588 /* Send ACK now, if this read freed lots of space 1589 * in our buffer. Certainly, new_window is new window. 1590 * We can advertise it now, if it is not less than current one. 1591 * "Lots" means "at least twice" here. 1592 */ 1593 if (new_window && new_window >= 2 * rcv_window_now) 1594 time_to_ack = true; 1595 } 1596 } 1597 if (time_to_ack) { 1598 tcp_mstamp_refresh(tp); 1599 tcp_send_ack(sk); 1600 } 1601 } 1602 1603 void tcp_cleanup_rbuf(struct sock *sk, int copied) 1604 { 1605 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue); 1606 struct tcp_sock *tp = tcp_sk(sk); 1607 1608 WARN(skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq), 1609 "cleanup rbuf bug: copied %X seq %X rcvnxt %X\n", 1610 tp->copied_seq, TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt); 1611 __tcp_cleanup_rbuf(sk, copied); 1612 } 1613 1614 static void tcp_eat_recv_skb(struct sock *sk, struct sk_buff *skb) 1615 { 1616 __skb_unlink(skb, &sk->sk_receive_queue); 1617 if (likely(skb->destructor == sock_rfree)) { 1618 sock_rfree(skb); 1619 skb->destructor = NULL; 1620 skb->sk = NULL; 1621 return skb_attempt_defer_free(skb); 1622 } 1623 __kfree_skb(skb); 1624 } 1625 1626 struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off) 1627 { 1628 struct sk_buff *skb; 1629 u32 offset; 1630 1631 while ((skb = skb_peek(&sk->sk_receive_queue)) != NULL) { 1632 offset = seq - TCP_SKB_CB(skb)->seq; 1633 if (unlikely(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) { 1634 pr_err_once("%s: found a SYN, please report !\n", __func__); 1635 offset--; 1636 } 1637 if (offset < skb->len || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)) { 1638 *off = offset; 1639 return skb; 1640 } 1641 /* This looks weird, but this can happen if TCP collapsing 1642 * splitted a fat GRO packet, while we released socket lock 1643 * in skb_splice_bits() 1644 */ 1645 tcp_eat_recv_skb(sk, skb); 1646 } 1647 return NULL; 1648 } 1649 EXPORT_SYMBOL(tcp_recv_skb); 1650 1651 /* 1652 * This routine provides an alternative to tcp_recvmsg() for routines 1653 * that would like to handle copying from skbuffs directly in 'sendfile' 1654 * fashion. 1655 * Note: 1656 * - It is assumed that the socket was locked by the caller. 1657 * - The routine does not block. 1658 * - At present, there is no support for reading OOB data 1659 * or for 'peeking' the socket using this routine 1660 * (although both would be easy to implement). 1661 */ 1662 static int __tcp_read_sock(struct sock *sk, read_descriptor_t *desc, 1663 sk_read_actor_t recv_actor, bool noack, 1664 u32 *copied_seq) 1665 { 1666 struct sk_buff *skb; 1667 struct tcp_sock *tp = tcp_sk(sk); 1668 u32 seq = *copied_seq; 1669 u32 offset; 1670 int copied = 0; 1671 1672 if (sk->sk_state == TCP_LISTEN) 1673 return -ENOTCONN; 1674 while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) { 1675 if (offset < skb->len) { 1676 int used; 1677 size_t len; 1678 1679 len = skb->len - offset; 1680 /* Stop reading if we hit a patch of urgent data */ 1681 if (unlikely(tp->urg_data)) { 1682 u32 urg_offset = tp->urg_seq - seq; 1683 if (urg_offset < len) 1684 len = urg_offset; 1685 if (!len) 1686 break; 1687 } 1688 used = recv_actor(desc, skb, offset, len); 1689 if (used <= 0) { 1690 if (!copied) 1691 copied = used; 1692 break; 1693 } 1694 if (WARN_ON_ONCE(used > len)) 1695 used = len; 1696 seq += used; 1697 copied += used; 1698 offset += used; 1699 1700 /* If recv_actor drops the lock (e.g. TCP splice 1701 * receive) the skb pointer might be invalid when 1702 * getting here: tcp_collapse might have deleted it 1703 * while aggregating skbs from the socket queue. 1704 */ 1705 skb = tcp_recv_skb(sk, seq - 1, &offset); 1706 if (!skb) 1707 break; 1708 /* TCP coalescing might have appended data to the skb. 1709 * Try to splice more frags 1710 */ 1711 if (offset + 1 != skb->len) 1712 continue; 1713 } 1714 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) { 1715 tcp_eat_recv_skb(sk, skb); 1716 ++seq; 1717 break; 1718 } 1719 tcp_eat_recv_skb(sk, skb); 1720 if (!desc->count) 1721 break; 1722 WRITE_ONCE(*copied_seq, seq); 1723 } 1724 WRITE_ONCE(*copied_seq, seq); 1725 1726 if (noack) 1727 goto out; 1728 1729 tcp_rcv_space_adjust(sk); 1730 1731 /* Clean up data we have read: This will do ACK frames. */ 1732 if (copied > 0) { 1733 tcp_recv_skb(sk, seq, &offset); 1734 tcp_cleanup_rbuf(sk, copied); 1735 } 1736 out: 1737 return copied; 1738 } 1739 1740 int tcp_read_sock(struct sock *sk, read_descriptor_t *desc, 1741 sk_read_actor_t recv_actor) 1742 { 1743 return __tcp_read_sock(sk, desc, recv_actor, false, 1744 &tcp_sk(sk)->copied_seq); 1745 } 1746 EXPORT_SYMBOL(tcp_read_sock); 1747 1748 int tcp_read_sock_noack(struct sock *sk, read_descriptor_t *desc, 1749 sk_read_actor_t recv_actor, bool noack, 1750 u32 *copied_seq) 1751 { 1752 return __tcp_read_sock(sk, desc, recv_actor, noack, copied_seq); 1753 } 1754 1755 int tcp_read_skb(struct sock *sk, skb_read_actor_t recv_actor) 1756 { 1757 struct sk_buff *skb; 1758 int copied = 0; 1759 1760 if (sk->sk_state == TCP_LISTEN) 1761 return -ENOTCONN; 1762 1763 while ((skb = skb_peek(&sk->sk_receive_queue)) != NULL) { 1764 u8 tcp_flags; 1765 int used; 1766 1767 __skb_unlink(skb, &sk->sk_receive_queue); 1768 WARN_ON_ONCE(!skb_set_owner_sk_safe(skb, sk)); 1769 tcp_flags = TCP_SKB_CB(skb)->tcp_flags; 1770 used = recv_actor(sk, skb); 1771 if (used < 0) { 1772 if (!copied) 1773 copied = used; 1774 break; 1775 } 1776 copied += used; 1777 1778 if (tcp_flags & TCPHDR_FIN) 1779 break; 1780 } 1781 return copied; 1782 } 1783 1784 void tcp_read_done(struct sock *sk, size_t len) 1785 { 1786 struct tcp_sock *tp = tcp_sk(sk); 1787 u32 seq = tp->copied_seq; 1788 struct sk_buff *skb; 1789 size_t left; 1790 u32 offset; 1791 1792 if (sk->sk_state == TCP_LISTEN) 1793 return; 1794 1795 left = len; 1796 while (left && (skb = tcp_recv_skb(sk, seq, &offset)) != NULL) { 1797 int used; 1798 1799 used = min_t(size_t, skb->len - offset, left); 1800 seq += used; 1801 left -= used; 1802 1803 if (skb->len > offset + used) 1804 break; 1805 1806 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) { 1807 tcp_eat_recv_skb(sk, skb); 1808 ++seq; 1809 break; 1810 } 1811 tcp_eat_recv_skb(sk, skb); 1812 } 1813 WRITE_ONCE(tp->copied_seq, seq); 1814 1815 tcp_rcv_space_adjust(sk); 1816 1817 /* Clean up data we have read: This will do ACK frames. */ 1818 if (left != len) 1819 tcp_cleanup_rbuf(sk, len - left); 1820 } 1821 EXPORT_SYMBOL(tcp_read_done); 1822 1823 int tcp_peek_len(struct socket *sock) 1824 { 1825 return tcp_inq(sock->sk); 1826 } 1827 1828 /* Make sure sk_rcvbuf is big enough to satisfy SO_RCVLOWAT hint */ 1829 int tcp_set_rcvlowat(struct sock *sk, int val) 1830 { 1831 struct tcp_sock *tp = tcp_sk(sk); 1832 int space, cap; 1833 1834 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK) 1835 cap = sk->sk_rcvbuf >> 1; 1836 else 1837 cap = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_rmem[2]) >> 1; 1838 val = min(val, cap); 1839 WRITE_ONCE(sk->sk_rcvlowat, val ? : 1); 1840 1841 /* Check if we need to signal EPOLLIN right now */ 1842 tcp_data_ready(sk); 1843 1844 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK) 1845 return 0; 1846 1847 space = tcp_space_from_win(sk, val); 1848 if (space > sk->sk_rcvbuf) { 1849 WRITE_ONCE(sk->sk_rcvbuf, space); 1850 1851 if (tp->window_clamp && tp->window_clamp < val) 1852 WRITE_ONCE(tp->window_clamp, val); 1853 } 1854 return 0; 1855 } 1856 1857 void tcp_set_rcvbuf(struct sock *sk, int val) 1858 { 1859 tcp_set_window_clamp(sk, tcp_win_from_space(sk, val)); 1860 } 1861 1862 #ifdef CONFIG_MMU 1863 static const struct vm_operations_struct tcp_vm_ops = { 1864 }; 1865 1866 int tcp_mmap(struct file *file, struct socket *sock, 1867 struct vm_area_struct *vma) 1868 { 1869 if (vma->vm_flags & (VM_WRITE | VM_EXEC)) 1870 return -EPERM; 1871 vm_flags_clear(vma, VM_MAYWRITE | VM_MAYEXEC); 1872 1873 /* Instruct vm_insert_page() to not mmap_read_lock(mm) */ 1874 vm_flags_set(vma, VM_MIXEDMAP); 1875 1876 vma->vm_ops = &tcp_vm_ops; 1877 return 0; 1878 } 1879 1880 static skb_frag_t *skb_advance_to_frag(struct sk_buff *skb, u32 offset_skb, 1881 u32 *offset_frag) 1882 { 1883 skb_frag_t *frag; 1884 1885 if (unlikely(offset_skb >= skb->len)) 1886 return NULL; 1887 1888 offset_skb -= skb_headlen(skb); 1889 if ((int)offset_skb < 0 || skb_has_frag_list(skb)) 1890 return NULL; 1891 1892 frag = skb_shinfo(skb)->frags; 1893 while (offset_skb) { 1894 if (skb_frag_size(frag) > offset_skb) { 1895 *offset_frag = offset_skb; 1896 return frag; 1897 } 1898 offset_skb -= skb_frag_size(frag); 1899 ++frag; 1900 } 1901 *offset_frag = 0; 1902 return frag; 1903 } 1904 1905 static bool can_map_frag(const skb_frag_t *frag) 1906 { 1907 struct page *page; 1908 1909 if (skb_frag_size(frag) != PAGE_SIZE || skb_frag_off(frag)) 1910 return false; 1911 1912 page = skb_frag_page(frag); 1913 1914 if (PageCompound(page) || page->mapping) 1915 return false; 1916 1917 return true; 1918 } 1919 1920 static int find_next_mappable_frag(const skb_frag_t *frag, 1921 int remaining_in_skb) 1922 { 1923 int offset = 0; 1924 1925 if (likely(can_map_frag(frag))) 1926 return 0; 1927 1928 while (offset < remaining_in_skb && !can_map_frag(frag)) { 1929 offset += skb_frag_size(frag); 1930 ++frag; 1931 } 1932 return offset; 1933 } 1934 1935 static void tcp_zerocopy_set_hint_for_skb(struct sock *sk, 1936 struct tcp_zerocopy_receive *zc, 1937 struct sk_buff *skb, u32 offset) 1938 { 1939 u32 frag_offset, partial_frag_remainder = 0; 1940 int mappable_offset; 1941 skb_frag_t *frag; 1942 1943 /* worst case: skip to next skb. try to improve on this case below */ 1944 zc->recv_skip_hint = skb->len - offset; 1945 1946 /* Find the frag containing this offset (and how far into that frag) */ 1947 frag = skb_advance_to_frag(skb, offset, &frag_offset); 1948 if (!frag) 1949 return; 1950 1951 if (frag_offset) { 1952 struct skb_shared_info *info = skb_shinfo(skb); 1953 1954 /* We read part of the last frag, must recvmsg() rest of skb. */ 1955 if (frag == &info->frags[info->nr_frags - 1]) 1956 return; 1957 1958 /* Else, we must at least read the remainder in this frag. */ 1959 partial_frag_remainder = skb_frag_size(frag) - frag_offset; 1960 zc->recv_skip_hint -= partial_frag_remainder; 1961 ++frag; 1962 } 1963 1964 /* partial_frag_remainder: If part way through a frag, must read rest. 1965 * mappable_offset: Bytes till next mappable frag, *not* counting bytes 1966 * in partial_frag_remainder. 1967 */ 1968 mappable_offset = find_next_mappable_frag(frag, zc->recv_skip_hint); 1969 zc->recv_skip_hint = mappable_offset + partial_frag_remainder; 1970 } 1971 1972 static int tcp_recvmsg_locked(struct sock *sk, struct msghdr *msg, size_t len, 1973 int flags, struct scm_timestamping_internal *tss, 1974 int *cmsg_flags); 1975 static int receive_fallback_to_copy(struct sock *sk, 1976 struct tcp_zerocopy_receive *zc, int inq, 1977 struct scm_timestamping_internal *tss) 1978 { 1979 unsigned long copy_address = (unsigned long)zc->copybuf_address; 1980 struct msghdr msg = {}; 1981 int err; 1982 1983 zc->length = 0; 1984 zc->recv_skip_hint = 0; 1985 1986 if (copy_address != zc->copybuf_address) 1987 return -EINVAL; 1988 1989 err = import_ubuf(ITER_DEST, (void __user *)copy_address, inq, 1990 &msg.msg_iter); 1991 if (err) 1992 return err; 1993 1994 err = tcp_recvmsg_locked(sk, &msg, inq, MSG_DONTWAIT, 1995 tss, &zc->msg_flags); 1996 if (err < 0) 1997 return err; 1998 1999 zc->copybuf_len = err; 2000 if (likely(zc->copybuf_len)) { 2001 struct sk_buff *skb; 2002 u32 offset; 2003 2004 skb = tcp_recv_skb(sk, tcp_sk(sk)->copied_seq, &offset); 2005 if (skb) 2006 tcp_zerocopy_set_hint_for_skb(sk, zc, skb, offset); 2007 } 2008 return 0; 2009 } 2010 2011 static int tcp_copy_straggler_data(struct tcp_zerocopy_receive *zc, 2012 struct sk_buff *skb, u32 copylen, 2013 u32 *offset, u32 *seq) 2014 { 2015 unsigned long copy_address = (unsigned long)zc->copybuf_address; 2016 struct msghdr msg = {}; 2017 int err; 2018 2019 if (copy_address != zc->copybuf_address) 2020 return -EINVAL; 2021 2022 err = import_ubuf(ITER_DEST, (void __user *)copy_address, copylen, 2023 &msg.msg_iter); 2024 if (err) 2025 return err; 2026 err = skb_copy_datagram_msg(skb, *offset, &msg, copylen); 2027 if (err) 2028 return err; 2029 zc->recv_skip_hint -= copylen; 2030 *offset += copylen; 2031 *seq += copylen; 2032 return (__s32)copylen; 2033 } 2034 2035 static int tcp_zc_handle_leftover(struct tcp_zerocopy_receive *zc, 2036 struct sock *sk, 2037 struct sk_buff *skb, 2038 u32 *seq, 2039 s32 copybuf_len, 2040 struct scm_timestamping_internal *tss) 2041 { 2042 u32 offset, copylen = min_t(u32, copybuf_len, zc->recv_skip_hint); 2043 2044 if (!copylen) 2045 return 0; 2046 /* skb is null if inq < PAGE_SIZE. */ 2047 if (skb) { 2048 offset = *seq - TCP_SKB_CB(skb)->seq; 2049 } else { 2050 skb = tcp_recv_skb(sk, *seq, &offset); 2051 if (TCP_SKB_CB(skb)->has_rxtstamp) { 2052 tcp_update_recv_tstamps(skb, tss); 2053 zc->msg_flags |= TCP_CMSG_TS; 2054 } 2055 } 2056 2057 zc->copybuf_len = tcp_copy_straggler_data(zc, skb, copylen, &offset, 2058 seq); 2059 return zc->copybuf_len < 0 ? 0 : copylen; 2060 } 2061 2062 static int tcp_zerocopy_vm_insert_batch_error(struct vm_area_struct *vma, 2063 struct page **pending_pages, 2064 unsigned long pages_remaining, 2065 unsigned long *address, 2066 u32 *length, 2067 u32 *seq, 2068 struct tcp_zerocopy_receive *zc, 2069 u32 total_bytes_to_map, 2070 int err) 2071 { 2072 /* At least one page did not map. Try zapping if we skipped earlier. */ 2073 if (err == -EBUSY && 2074 zc->flags & TCP_RECEIVE_ZEROCOPY_FLAG_TLB_CLEAN_HINT) { 2075 u32 maybe_zap_len; 2076 2077 maybe_zap_len = total_bytes_to_map - /* All bytes to map */ 2078 *length + /* Mapped or pending */ 2079 (pages_remaining * PAGE_SIZE); /* Failed map. */ 2080 zap_vma_range(vma, *address, maybe_zap_len); 2081 err = 0; 2082 } 2083 2084 if (!err) { 2085 unsigned long leftover_pages = pages_remaining; 2086 int bytes_mapped; 2087 2088 /* We called zap_vma_range, try to reinsert. */ 2089 err = vm_insert_pages(vma, *address, 2090 pending_pages, 2091 &pages_remaining); 2092 bytes_mapped = PAGE_SIZE * (leftover_pages - pages_remaining); 2093 *seq += bytes_mapped; 2094 *address += bytes_mapped; 2095 } 2096 if (err) { 2097 /* Either we were unable to zap, OR we zapped, retried an 2098 * insert, and still had an issue. Either ways, pages_remaining 2099 * is the number of pages we were unable to map, and we unroll 2100 * some state we speculatively touched before. 2101 */ 2102 const int bytes_not_mapped = PAGE_SIZE * pages_remaining; 2103 2104 *length -= bytes_not_mapped; 2105 zc->recv_skip_hint += bytes_not_mapped; 2106 } 2107 return err; 2108 } 2109 2110 static int tcp_zerocopy_vm_insert_batch(struct vm_area_struct *vma, 2111 struct page **pages, 2112 unsigned int pages_to_map, 2113 unsigned long *address, 2114 u32 *length, 2115 u32 *seq, 2116 struct tcp_zerocopy_receive *zc, 2117 u32 total_bytes_to_map) 2118 { 2119 unsigned long pages_remaining = pages_to_map; 2120 unsigned int pages_mapped; 2121 unsigned int bytes_mapped; 2122 int err; 2123 2124 err = vm_insert_pages(vma, *address, pages, &pages_remaining); 2125 pages_mapped = pages_to_map - (unsigned int)pages_remaining; 2126 bytes_mapped = PAGE_SIZE * pages_mapped; 2127 /* Even if vm_insert_pages fails, it may have partially succeeded in 2128 * mapping (some but not all of the pages). 2129 */ 2130 *seq += bytes_mapped; 2131 *address += bytes_mapped; 2132 2133 if (likely(!err)) 2134 return 0; 2135 2136 /* Error: maybe zap and retry + rollback state for failed inserts. */ 2137 return tcp_zerocopy_vm_insert_batch_error(vma, pages + pages_mapped, 2138 pages_remaining, address, length, seq, zc, total_bytes_to_map, 2139 err); 2140 } 2141 2142 #define TCP_VALID_ZC_MSG_FLAGS (TCP_CMSG_TS) 2143 static void tcp_zc_finalize_rx_tstamp(struct sock *sk, 2144 struct tcp_zerocopy_receive *zc, 2145 struct scm_timestamping_internal *tss) 2146 { 2147 unsigned long msg_control_addr; 2148 struct msghdr cmsg_dummy; 2149 2150 msg_control_addr = (unsigned long)zc->msg_control; 2151 cmsg_dummy.msg_control_user = (void __user *)msg_control_addr; 2152 cmsg_dummy.msg_controllen = 2153 (__kernel_size_t)zc->msg_controllen; 2154 cmsg_dummy.msg_flags = in_compat_syscall() 2155 ? MSG_CMSG_COMPAT : 0; 2156 cmsg_dummy.msg_control_is_user = true; 2157 zc->msg_flags = 0; 2158 if (zc->msg_control == msg_control_addr && 2159 zc->msg_controllen == cmsg_dummy.msg_controllen) { 2160 tcp_recv_timestamp(&cmsg_dummy, sk, tss); 2161 zc->msg_control = (__u64) 2162 ((uintptr_t)cmsg_dummy.msg_control_user); 2163 zc->msg_controllen = 2164 (__u64)cmsg_dummy.msg_controllen; 2165 zc->msg_flags = (__u32)cmsg_dummy.msg_flags; 2166 } 2167 } 2168 2169 static struct vm_area_struct *find_tcp_vma(struct mm_struct *mm, 2170 unsigned long address, 2171 bool *mmap_locked) 2172 { 2173 struct vm_area_struct *vma = lock_vma_under_rcu(mm, address); 2174 2175 if (vma) { 2176 if (vma->vm_ops != &tcp_vm_ops) { 2177 vma_end_read(vma); 2178 return NULL; 2179 } 2180 *mmap_locked = false; 2181 return vma; 2182 } 2183 2184 mmap_read_lock(mm); 2185 vma = vma_lookup(mm, address); 2186 if (!vma || vma->vm_ops != &tcp_vm_ops) { 2187 mmap_read_unlock(mm); 2188 return NULL; 2189 } 2190 *mmap_locked = true; 2191 return vma; 2192 } 2193 2194 #define TCP_ZEROCOPY_PAGE_BATCH_SIZE 32 2195 static int tcp_zerocopy_receive(struct sock *sk, 2196 struct tcp_zerocopy_receive *zc, 2197 struct scm_timestamping_internal *tss) 2198 { 2199 u32 length = 0, offset, vma_len, avail_len, copylen = 0; 2200 unsigned long address = (unsigned long)zc->address; 2201 struct page *pages[TCP_ZEROCOPY_PAGE_BATCH_SIZE]; 2202 s32 copybuf_len = zc->copybuf_len; 2203 struct tcp_sock *tp = tcp_sk(sk); 2204 const skb_frag_t *frags = NULL; 2205 unsigned int pages_to_map = 0; 2206 struct vm_area_struct *vma; 2207 struct sk_buff *skb = NULL; 2208 u32 seq = tp->copied_seq; 2209 u32 total_bytes_to_map; 2210 int inq = tcp_inq(sk); 2211 bool mmap_locked; 2212 int ret; 2213 2214 zc->copybuf_len = 0; 2215 zc->msg_flags = 0; 2216 2217 if (address & (PAGE_SIZE - 1) || address != zc->address) 2218 return -EINVAL; 2219 2220 if (sk->sk_state == TCP_LISTEN) 2221 return -ENOTCONN; 2222 2223 sock_rps_record_flow(sk); 2224 2225 if (inq && inq <= copybuf_len) 2226 return receive_fallback_to_copy(sk, zc, inq, tss); 2227 2228 if (inq < PAGE_SIZE) { 2229 zc->length = 0; 2230 zc->recv_skip_hint = inq; 2231 if (!inq && sock_flag(sk, SOCK_DONE)) 2232 return -EIO; 2233 return 0; 2234 } 2235 2236 vma = find_tcp_vma(current->mm, address, &mmap_locked); 2237 if (!vma) 2238 return -EINVAL; 2239 2240 vma_len = min_t(unsigned long, zc->length, vma->vm_end - address); 2241 avail_len = min_t(u32, vma_len, inq); 2242 total_bytes_to_map = avail_len & ~(PAGE_SIZE - 1); 2243 if (total_bytes_to_map) { 2244 if (!(zc->flags & TCP_RECEIVE_ZEROCOPY_FLAG_TLB_CLEAN_HINT)) 2245 zap_vma_range(vma, address, total_bytes_to_map); 2246 zc->length = total_bytes_to_map; 2247 zc->recv_skip_hint = 0; 2248 } else { 2249 zc->length = avail_len; 2250 zc->recv_skip_hint = avail_len; 2251 } 2252 ret = 0; 2253 while (length + PAGE_SIZE <= zc->length) { 2254 int mappable_offset; 2255 struct page *page; 2256 2257 if (zc->recv_skip_hint < PAGE_SIZE) { 2258 u32 offset_frag; 2259 2260 if (skb) { 2261 if (zc->recv_skip_hint > 0) 2262 break; 2263 skb = skb->next; 2264 offset = seq - TCP_SKB_CB(skb)->seq; 2265 } else { 2266 skb = tcp_recv_skb(sk, seq, &offset); 2267 } 2268 2269 if (!skb_frags_readable(skb)) 2270 break; 2271 2272 if (TCP_SKB_CB(skb)->has_rxtstamp) { 2273 tcp_update_recv_tstamps(skb, tss); 2274 zc->msg_flags |= TCP_CMSG_TS; 2275 } 2276 zc->recv_skip_hint = skb->len - offset; 2277 frags = skb_advance_to_frag(skb, offset, &offset_frag); 2278 if (!frags || offset_frag) 2279 break; 2280 } 2281 2282 mappable_offset = find_next_mappable_frag(frags, 2283 zc->recv_skip_hint); 2284 if (mappable_offset) { 2285 zc->recv_skip_hint = mappable_offset; 2286 break; 2287 } 2288 page = skb_frag_page(frags); 2289 if (WARN_ON_ONCE(!page)) 2290 break; 2291 2292 prefetchw(page); 2293 pages[pages_to_map++] = page; 2294 length += PAGE_SIZE; 2295 zc->recv_skip_hint -= PAGE_SIZE; 2296 frags++; 2297 if (pages_to_map == TCP_ZEROCOPY_PAGE_BATCH_SIZE || 2298 zc->recv_skip_hint < PAGE_SIZE) { 2299 /* Either full batch, or we're about to go to next skb 2300 * (and we cannot unroll failed ops across skbs). 2301 */ 2302 ret = tcp_zerocopy_vm_insert_batch(vma, pages, 2303 pages_to_map, 2304 &address, &length, 2305 &seq, zc, 2306 total_bytes_to_map); 2307 if (ret) 2308 goto out; 2309 pages_to_map = 0; 2310 } 2311 } 2312 if (pages_to_map) { 2313 ret = tcp_zerocopy_vm_insert_batch(vma, pages, pages_to_map, 2314 &address, &length, &seq, 2315 zc, total_bytes_to_map); 2316 } 2317 out: 2318 if (mmap_locked) 2319 mmap_read_unlock(current->mm); 2320 else 2321 vma_end_read(vma); 2322 /* Try to copy straggler data. */ 2323 if (!ret) 2324 copylen = tcp_zc_handle_leftover(zc, sk, skb, &seq, copybuf_len, tss); 2325 2326 if (length + copylen) { 2327 WRITE_ONCE(tp->copied_seq, seq); 2328 tcp_rcv_space_adjust(sk); 2329 2330 /* Clean up data we have read: This will do ACK frames. */ 2331 tcp_recv_skb(sk, seq, &offset); 2332 tcp_cleanup_rbuf(sk, length + copylen); 2333 ret = 0; 2334 if (length == zc->length) 2335 zc->recv_skip_hint = 0; 2336 } else { 2337 if (!zc->recv_skip_hint && sock_flag(sk, SOCK_DONE)) 2338 ret = -EIO; 2339 } 2340 zc->length = length; 2341 return ret; 2342 } 2343 #endif 2344 2345 /* Similar to __sock_recv_timestamp, but does not require an skb */ 2346 void tcp_recv_timestamp(struct msghdr *msg, const struct sock *sk, 2347 struct scm_timestamping_internal *tss) 2348 { 2349 int new_tstamp = sock_flag(sk, SOCK_TSTAMP_NEW); 2350 u32 tsflags = READ_ONCE(sk->sk_tsflags); 2351 2352 if (tss->ts[0]) { 2353 if (sock_flag(sk, SOCK_RCVTSTAMP)) { 2354 struct timespec64 tv = ktime_to_timespec64(tss->ts[0]); 2355 2356 if (sock_flag(sk, SOCK_RCVTSTAMPNS)) { 2357 if (new_tstamp) { 2358 struct __kernel_timespec kts = { 2359 .tv_sec = tv.tv_sec, 2360 .tv_nsec = tv.tv_nsec, 2361 }; 2362 put_cmsg(msg, SOL_SOCKET, SO_TIMESTAMPNS_NEW, 2363 sizeof(kts), &kts); 2364 } else { 2365 struct __kernel_old_timespec ts_old = { 2366 .tv_sec = tv.tv_sec, 2367 .tv_nsec = tv.tv_nsec, 2368 }; 2369 put_cmsg(msg, SOL_SOCKET, SO_TIMESTAMPNS_OLD, 2370 sizeof(ts_old), &ts_old); 2371 } 2372 } else { 2373 if (new_tstamp) { 2374 struct __kernel_sock_timeval stv = { 2375 .tv_sec = tv.tv_sec, 2376 .tv_usec = tv.tv_nsec / 1000, 2377 }; 2378 put_cmsg(msg, SOL_SOCKET, SO_TIMESTAMP_NEW, 2379 sizeof(stv), &stv); 2380 } else { 2381 struct __kernel_old_timeval otv = { 2382 .tv_sec = tv.tv_sec, 2383 .tv_usec = tv.tv_nsec / 1000, 2384 }; 2385 put_cmsg(msg, SOL_SOCKET, SO_TIMESTAMP_OLD, 2386 sizeof(otv), &otv); 2387 } 2388 } 2389 } 2390 2391 if (!(tsflags & SOF_TIMESTAMPING_SOFTWARE && 2392 (tsflags & SOF_TIMESTAMPING_RX_SOFTWARE || 2393 !(tsflags & SOF_TIMESTAMPING_OPT_RX_FILTER)))) 2394 tss->ts[0] = 0; 2395 } 2396 2397 if (tss->ts[2]) { 2398 if (!(tsflags & SOF_TIMESTAMPING_RAW_HARDWARE && 2399 (tsflags & SOF_TIMESTAMPING_RX_HARDWARE || 2400 !(tsflags & SOF_TIMESTAMPING_OPT_RX_FILTER)))) 2401 tss->ts[2] = 0; 2402 } 2403 2404 if (tss->ts[0] | tss->ts[2]) { 2405 tss->ts[1] = 0; 2406 if (sock_flag(sk, SOCK_TSTAMP_NEW)) 2407 put_cmsg_scm_timestamping64(msg, tss); 2408 else 2409 put_cmsg_scm_timestamping(msg, tss); 2410 } 2411 } 2412 2413 static int tcp_inq_hint(struct sock *sk) 2414 { 2415 const struct tcp_sock *tp = tcp_sk(sk); 2416 u32 copied_seq = READ_ONCE(tp->copied_seq); 2417 u32 rcv_nxt = READ_ONCE(tp->rcv_nxt); 2418 int inq; 2419 2420 inq = rcv_nxt - copied_seq; 2421 if (unlikely(inq < 0 || copied_seq != READ_ONCE(tp->copied_seq))) { 2422 lock_sock(sk); 2423 inq = tp->rcv_nxt - tp->copied_seq; 2424 release_sock(sk); 2425 } 2426 /* After receiving a FIN, tell the user-space to continue reading 2427 * by returning a non-zero inq. 2428 */ 2429 if (inq == 0 && sock_flag(sk, SOCK_DONE)) 2430 inq = 1; 2431 return inq; 2432 } 2433 2434 /* batch __xa_alloc() calls and reduce xa_lock()/xa_unlock() overhead. */ 2435 struct tcp_xa_pool { 2436 u8 max; /* max <= MAX_SKB_FRAGS */ 2437 u8 idx; /* idx <= max */ 2438 __u32 tokens[MAX_SKB_FRAGS]; 2439 netmem_ref netmems[MAX_SKB_FRAGS]; 2440 }; 2441 2442 static void tcp_xa_pool_commit_locked(struct sock *sk, struct tcp_xa_pool *p) 2443 { 2444 int i; 2445 2446 /* Commit part that has been copied to user space. */ 2447 for (i = 0; i < p->idx; i++) 2448 __xa_cmpxchg(&sk->sk_user_frags, p->tokens[i], XA_ZERO_ENTRY, 2449 (__force void *)p->netmems[i], GFP_KERNEL); 2450 /* Rollback what has been pre-allocated and is no longer needed. */ 2451 for (; i < p->max; i++) 2452 __xa_erase(&sk->sk_user_frags, p->tokens[i]); 2453 2454 p->max = 0; 2455 p->idx = 0; 2456 } 2457 2458 static void tcp_xa_pool_commit(struct sock *sk, struct tcp_xa_pool *p) 2459 { 2460 if (!p->max) 2461 return; 2462 2463 xa_lock_bh(&sk->sk_user_frags); 2464 2465 tcp_xa_pool_commit_locked(sk, p); 2466 2467 xa_unlock_bh(&sk->sk_user_frags); 2468 } 2469 2470 static int tcp_xa_pool_refill(struct sock *sk, struct tcp_xa_pool *p, 2471 unsigned int max_frags) 2472 { 2473 int err, k; 2474 2475 if (p->idx < p->max) 2476 return 0; 2477 2478 xa_lock_bh(&sk->sk_user_frags); 2479 2480 tcp_xa_pool_commit_locked(sk, p); 2481 2482 for (k = 0; k < max_frags; k++) { 2483 err = __xa_alloc(&sk->sk_user_frags, &p->tokens[k], 2484 XA_ZERO_ENTRY, xa_limit_31b, GFP_KERNEL); 2485 if (err) 2486 break; 2487 } 2488 2489 xa_unlock_bh(&sk->sk_user_frags); 2490 2491 p->max = k; 2492 p->idx = 0; 2493 return k ? 0 : err; 2494 } 2495 2496 /* On error, returns the -errno. On success, returns number of bytes sent to the 2497 * user. May not consume all of @remaining_len. 2498 */ 2499 static int tcp_recvmsg_dmabuf(struct sock *sk, const struct sk_buff *skb, 2500 unsigned int offset, struct msghdr *msg, 2501 int remaining_len) 2502 { 2503 struct dmabuf_cmsg dmabuf_cmsg = { 0 }; 2504 struct tcp_xa_pool tcp_xa_pool; 2505 unsigned int start; 2506 int i, copy, n; 2507 int sent = 0; 2508 int err = 0; 2509 2510 tcp_xa_pool.max = 0; 2511 tcp_xa_pool.idx = 0; 2512 do { 2513 start = skb_headlen(skb); 2514 2515 if (skb_frags_readable(skb)) { 2516 err = -ENODEV; 2517 goto out; 2518 } 2519 2520 /* Copy header. */ 2521 copy = start - offset; 2522 if (copy > 0) { 2523 copy = min(copy, remaining_len); 2524 2525 n = copy_to_iter(skb->data + offset, copy, 2526 &msg->msg_iter); 2527 if (n != copy) { 2528 err = -EFAULT; 2529 goto out; 2530 } 2531 2532 offset += copy; 2533 remaining_len -= copy; 2534 2535 /* First a dmabuf_cmsg for # bytes copied to user 2536 * buffer. 2537 */ 2538 memset(&dmabuf_cmsg, 0, sizeof(dmabuf_cmsg)); 2539 dmabuf_cmsg.frag_size = copy; 2540 err = put_cmsg_notrunc(msg, SOL_SOCKET, 2541 SO_DEVMEM_LINEAR, 2542 sizeof(dmabuf_cmsg), 2543 &dmabuf_cmsg); 2544 if (err) 2545 goto out; 2546 2547 sent += copy; 2548 2549 if (remaining_len == 0) 2550 goto out; 2551 } 2552 2553 /* after that, send information of dmabuf pages through a 2554 * sequence of cmsg 2555 */ 2556 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) { 2557 skb_frag_t *frag = &skb_shinfo(skb)->frags[i]; 2558 struct net_iov *niov; 2559 u64 frag_offset; 2560 int end; 2561 2562 /* !skb_frags_readable() should indicate that ALL the 2563 * frags in this skb are dmabuf net_iovs. We're checking 2564 * for that flag above, but also check individual frags 2565 * here. If the tcp stack is not setting 2566 * skb_frags_readable() correctly, we still don't want 2567 * to crash here. 2568 */ 2569 if (!skb_frag_net_iov(frag)) { 2570 net_err_ratelimited("Found non-dmabuf skb with net_iov"); 2571 err = -ENODEV; 2572 goto out; 2573 } 2574 2575 niov = skb_frag_net_iov(frag); 2576 if (!net_is_devmem_iov(niov)) { 2577 err = -ENODEV; 2578 goto out; 2579 } 2580 2581 end = start + skb_frag_size(frag); 2582 copy = end - offset; 2583 2584 if (copy > 0) { 2585 copy = min(copy, remaining_len); 2586 2587 frag_offset = net_iov_virtual_addr(niov) + 2588 skb_frag_off(frag) + offset - 2589 start; 2590 dmabuf_cmsg.frag_offset = frag_offset; 2591 dmabuf_cmsg.frag_size = copy; 2592 err = tcp_xa_pool_refill(sk, &tcp_xa_pool, 2593 skb_shinfo(skb)->nr_frags - i); 2594 if (err) 2595 goto out; 2596 2597 /* Will perform the exchange later */ 2598 dmabuf_cmsg.frag_token = tcp_xa_pool.tokens[tcp_xa_pool.idx]; 2599 dmabuf_cmsg.dmabuf_id = net_devmem_iov_binding_id(niov); 2600 2601 offset += copy; 2602 remaining_len -= copy; 2603 2604 err = put_cmsg_notrunc(msg, SOL_SOCKET, 2605 SO_DEVMEM_DMABUF, 2606 sizeof(dmabuf_cmsg), 2607 &dmabuf_cmsg); 2608 if (err) 2609 goto out; 2610 2611 atomic_long_inc(&niov->desc.pp_ref_count); 2612 tcp_xa_pool.netmems[tcp_xa_pool.idx++] = skb_frag_netmem(frag); 2613 2614 sent += copy; 2615 2616 if (remaining_len == 0) 2617 goto out; 2618 } 2619 start = end; 2620 } 2621 2622 tcp_xa_pool_commit(sk, &tcp_xa_pool); 2623 if (!remaining_len) 2624 goto out; 2625 2626 /* if remaining_len is not satisfied yet, we need to go to the 2627 * next frag in the frag_list to satisfy remaining_len. 2628 */ 2629 skb = skb_shinfo(skb)->frag_list ?: skb->next; 2630 2631 offset = offset - start; 2632 } while (skb); 2633 2634 if (remaining_len) { 2635 err = -EFAULT; 2636 goto out; 2637 } 2638 2639 out: 2640 tcp_xa_pool_commit(sk, &tcp_xa_pool); 2641 if (!sent) 2642 sent = err; 2643 2644 return sent; 2645 } 2646 2647 /* 2648 * This routine copies from a sock struct into the user buffer. 2649 * 2650 * Technical note: in 2.3 we work on _locked_ socket, so that 2651 * tricks with *seq access order and skb->users are not required. 2652 * Probably, code can be easily improved even more. 2653 */ 2654 2655 static int tcp_recvmsg_locked(struct sock *sk, struct msghdr *msg, size_t len, 2656 int flags, struct scm_timestamping_internal *tss, 2657 int *cmsg_flags) 2658 { 2659 struct tcp_sock *tp = tcp_sk(sk); 2660 int last_copied_dmabuf = -1; /* uninitialized */ 2661 int copied = 0; 2662 u32 peek_seq; 2663 u32 *seq; 2664 unsigned long used; 2665 int err; 2666 int target; /* Read at least this many bytes */ 2667 long timeo; 2668 struct sk_buff *skb, *last; 2669 u32 peek_offset = 0; 2670 u32 urg_hole = 0; 2671 2672 err = -ENOTCONN; 2673 if (sk->sk_state == TCP_LISTEN) 2674 goto out; 2675 2676 if (tp->recvmsg_inq) 2677 *cmsg_flags = TCP_CMSG_INQ; 2678 timeo = sock_rcvtimeo(sk, flags & MSG_DONTWAIT); 2679 2680 /* Urgent data needs to be handled specially. */ 2681 if (flags & MSG_OOB) 2682 goto recv_urg; 2683 2684 if (unlikely(tp->repair)) { 2685 err = -EPERM; 2686 if (!(flags & MSG_PEEK)) 2687 goto out; 2688 2689 if (tp->repair_queue == TCP_SEND_QUEUE) 2690 goto recv_sndq; 2691 2692 err = -EINVAL; 2693 if (tp->repair_queue == TCP_NO_QUEUE) 2694 goto out; 2695 2696 /* 'common' recv queue MSG_PEEK-ing */ 2697 } 2698 2699 seq = &tp->copied_seq; 2700 if (flags & MSG_PEEK) { 2701 peek_offset = max(sk_peek_offset(sk, flags), 0); 2702 peek_seq = tp->copied_seq + peek_offset; 2703 seq = &peek_seq; 2704 } 2705 2706 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len); 2707 2708 do { 2709 u32 offset; 2710 2711 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */ 2712 if (unlikely(tp->urg_data) && tp->urg_seq == *seq) { 2713 if (copied) 2714 break; 2715 if (signal_pending(current)) { 2716 copied = timeo ? sock_intr_errno(timeo) : -EAGAIN; 2717 break; 2718 } 2719 } 2720 2721 /* Next get a buffer. */ 2722 2723 last = skb_peek_tail(&sk->sk_receive_queue); 2724 skb_queue_walk(&sk->sk_receive_queue, skb) { 2725 last = skb; 2726 /* Now that we have two receive queues this 2727 * shouldn't happen. 2728 */ 2729 if (WARN(before(*seq, TCP_SKB_CB(skb)->seq), 2730 "TCP recvmsg seq # bug: copied %X, seq %X, rcvnxt %X, fl %X\n", 2731 *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt, 2732 flags)) 2733 break; 2734 2735 offset = *seq - TCP_SKB_CB(skb)->seq; 2736 if (unlikely(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) { 2737 pr_err_once("%s: found a SYN, please report !\n", __func__); 2738 offset--; 2739 } 2740 if (offset < skb->len) 2741 goto found_ok_skb; 2742 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) 2743 goto found_fin_ok; 2744 WARN(!(flags & MSG_PEEK), 2745 "TCP recvmsg seq # bug 2: copied %X, seq %X, rcvnxt %X, fl %X\n", 2746 *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt, flags); 2747 } 2748 2749 /* Well, if we have backlog, try to process it now yet. */ 2750 2751 if (copied >= target && !READ_ONCE(sk->sk_backlog.tail)) 2752 break; 2753 2754 if (copied) { 2755 if (!timeo || 2756 tcp_recv_should_stop(sk)) 2757 break; 2758 } else { 2759 if (sock_flag(sk, SOCK_DONE)) 2760 break; 2761 2762 if (sk->sk_err) { 2763 copied = sock_error(sk); 2764 break; 2765 } 2766 2767 if (sk->sk_shutdown & RCV_SHUTDOWN) 2768 break; 2769 2770 if (sk->sk_state == TCP_CLOSE) { 2771 /* This occurs when user tries to read 2772 * from never connected socket. 2773 */ 2774 copied = -ENOTCONN; 2775 break; 2776 } 2777 2778 if (!timeo) { 2779 copied = -EAGAIN; 2780 break; 2781 } 2782 2783 if (signal_pending(current)) { 2784 copied = sock_intr_errno(timeo); 2785 break; 2786 } 2787 } 2788 2789 if (copied >= target) { 2790 /* Do not sleep, just process backlog. */ 2791 __sk_flush_backlog(sk); 2792 } else { 2793 tcp_cleanup_rbuf(sk, copied); 2794 err = sk_wait_data(sk, &timeo, last); 2795 if (err < 0) { 2796 err = copied ? : err; 2797 goto out; 2798 } 2799 } 2800 2801 if ((flags & MSG_PEEK) && 2802 (peek_seq - peek_offset - copied - urg_hole != tp->copied_seq)) { 2803 net_dbg_ratelimited("TCP(%s:%d): Application bug, race in MSG_PEEK\n", 2804 current->comm, 2805 task_pid_nr(current)); 2806 peek_seq = tp->copied_seq + peek_offset; 2807 } 2808 continue; 2809 2810 found_ok_skb: 2811 /* Ok so how much can we use? */ 2812 used = skb->len - offset; 2813 if (len < used) 2814 used = len; 2815 2816 /* Do we have urgent data here? */ 2817 if (unlikely(tp->urg_data)) { 2818 u32 urg_offset = tp->urg_seq - *seq; 2819 if (urg_offset < used) { 2820 if (!urg_offset) { 2821 if (!sock_flag(sk, SOCK_URGINLINE)) { 2822 WRITE_ONCE(*seq, *seq + 1); 2823 urg_hole++; 2824 offset++; 2825 used--; 2826 if (!used) 2827 goto skip_copy; 2828 } 2829 } else 2830 used = urg_offset; 2831 } 2832 } 2833 2834 if (!(flags & MSG_TRUNC)) { 2835 if (last_copied_dmabuf != -1 && 2836 last_copied_dmabuf != !skb_frags_readable(skb)) 2837 break; 2838 2839 if (skb_frags_readable(skb)) { 2840 err = skb_copy_datagram_msg(skb, offset, msg, 2841 used); 2842 if (err) { 2843 /* Exception. Bailout! */ 2844 if (!copied) 2845 copied = -EFAULT; 2846 break; 2847 } 2848 } else { 2849 if (!(flags & MSG_SOCK_DEVMEM)) { 2850 /* dmabuf skbs can only be received 2851 * with the MSG_SOCK_DEVMEM flag. 2852 */ 2853 if (!copied) 2854 copied = -EFAULT; 2855 2856 break; 2857 } 2858 2859 err = tcp_recvmsg_dmabuf(sk, skb, offset, msg, 2860 used); 2861 if (err < 0) { 2862 if (!copied) 2863 copied = err; 2864 2865 break; 2866 } 2867 used = err; 2868 } 2869 } 2870 2871 last_copied_dmabuf = !skb_frags_readable(skb); 2872 2873 WRITE_ONCE(*seq, *seq + used); 2874 copied += used; 2875 len -= used; 2876 if (flags & MSG_PEEK) 2877 sk_peek_offset_fwd(sk, used); 2878 else 2879 sk_peek_offset_bwd(sk, used); 2880 tcp_rcv_space_adjust(sk); 2881 2882 skip_copy: 2883 if (unlikely(tp->urg_data) && after(tp->copied_seq, tp->urg_seq)) { 2884 WRITE_ONCE(tp->urg_data, 0); 2885 tcp_fast_path_check(sk); 2886 } 2887 2888 if (TCP_SKB_CB(skb)->has_rxtstamp) { 2889 tcp_update_recv_tstamps(skb, tss); 2890 *cmsg_flags |= TCP_CMSG_TS; 2891 } 2892 2893 if (used + offset < skb->len) 2894 continue; 2895 2896 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) 2897 goto found_fin_ok; 2898 if (!(flags & MSG_PEEK)) 2899 tcp_eat_recv_skb(sk, skb); 2900 continue; 2901 2902 found_fin_ok: 2903 /* Process the FIN. */ 2904 WRITE_ONCE(*seq, *seq + 1); 2905 if (!(flags & MSG_PEEK)) 2906 tcp_eat_recv_skb(sk, skb); 2907 break; 2908 } while (len > 0); 2909 2910 /* According to UNIX98, msg_name/msg_namelen are ignored 2911 * on connected socket. I was just happy when found this 8) --ANK 2912 */ 2913 2914 /* Clean up data we have read: This will do ACK frames. */ 2915 tcp_cleanup_rbuf(sk, copied); 2916 return copied; 2917 2918 out: 2919 return err; 2920 2921 recv_urg: 2922 err = tcp_recv_urg(sk, msg, len, flags); 2923 goto out; 2924 2925 recv_sndq: 2926 err = tcp_peek_sndq(sk, msg, len); 2927 goto out; 2928 } 2929 2930 int tcp_recvmsg(struct sock *sk, struct msghdr *msg, size_t len, int flags) 2931 { 2932 int cmsg_flags = 0, ret; 2933 struct scm_timestamping_internal tss; 2934 2935 if (unlikely(flags & MSG_ERRQUEUE)) 2936 return inet_recv_error(sk, msg, len); 2937 2938 if (sk_can_busy_loop(sk) && 2939 skb_queue_empty_lockless(&sk->sk_receive_queue) && 2940 sk->sk_state == TCP_ESTABLISHED) 2941 sk_busy_loop(sk, flags & MSG_DONTWAIT); 2942 2943 lock_sock(sk); 2944 ret = tcp_recvmsg_locked(sk, msg, len, flags, &tss, &cmsg_flags); 2945 release_sock(sk); 2946 2947 if ((cmsg_flags | msg->msg_get_inq) && ret >= 0) { 2948 if (cmsg_flags & TCP_CMSG_TS) 2949 tcp_recv_timestamp(msg, sk, &tss); 2950 if ((cmsg_flags & TCP_CMSG_INQ) | msg->msg_get_inq) { 2951 msg->msg_inq = tcp_inq_hint(sk); 2952 if (cmsg_flags & TCP_CMSG_INQ) 2953 put_cmsg(msg, SOL_TCP, TCP_CM_INQ, 2954 sizeof(msg->msg_inq), &msg->msg_inq); 2955 } 2956 } 2957 return ret; 2958 } 2959 2960 void tcp_set_state(struct sock *sk, int state) 2961 { 2962 int oldstate = sk->sk_state; 2963 2964 /* We defined a new enum for TCP states that are exported in BPF 2965 * so as not force the internal TCP states to be frozen. The 2966 * following checks will detect if an internal state value ever 2967 * differs from the BPF value. If this ever happens, then we will 2968 * need to remap the internal value to the BPF value before calling 2969 * tcp_call_bpf_2arg. 2970 */ 2971 BUILD_BUG_ON((int)BPF_TCP_ESTABLISHED != (int)TCP_ESTABLISHED); 2972 BUILD_BUG_ON((int)BPF_TCP_SYN_SENT != (int)TCP_SYN_SENT); 2973 BUILD_BUG_ON((int)BPF_TCP_SYN_RECV != (int)TCP_SYN_RECV); 2974 BUILD_BUG_ON((int)BPF_TCP_FIN_WAIT1 != (int)TCP_FIN_WAIT1); 2975 BUILD_BUG_ON((int)BPF_TCP_FIN_WAIT2 != (int)TCP_FIN_WAIT2); 2976 BUILD_BUG_ON((int)BPF_TCP_TIME_WAIT != (int)TCP_TIME_WAIT); 2977 BUILD_BUG_ON((int)BPF_TCP_CLOSE != (int)TCP_CLOSE); 2978 BUILD_BUG_ON((int)BPF_TCP_CLOSE_WAIT != (int)TCP_CLOSE_WAIT); 2979 BUILD_BUG_ON((int)BPF_TCP_LAST_ACK != (int)TCP_LAST_ACK); 2980 BUILD_BUG_ON((int)BPF_TCP_LISTEN != (int)TCP_LISTEN); 2981 BUILD_BUG_ON((int)BPF_TCP_CLOSING != (int)TCP_CLOSING); 2982 BUILD_BUG_ON((int)BPF_TCP_NEW_SYN_RECV != (int)TCP_NEW_SYN_RECV); 2983 BUILD_BUG_ON((int)BPF_TCP_BOUND_INACTIVE != (int)TCP_BOUND_INACTIVE); 2984 BUILD_BUG_ON((int)BPF_TCP_MAX_STATES != (int)TCP_MAX_STATES); 2985 2986 /* bpf uapi header bpf.h defines an anonymous enum with values 2987 * BPF_TCP_* used by bpf programs. Currently gcc built vmlinux 2988 * is able to emit this enum in DWARF due to the above BUILD_BUG_ON. 2989 * But clang built vmlinux does not have this enum in DWARF 2990 * since clang removes the above code before generating IR/debuginfo. 2991 * Let us explicitly emit the type debuginfo to ensure the 2992 * above-mentioned anonymous enum in the vmlinux DWARF and hence BTF 2993 * regardless of which compiler is used. 2994 */ 2995 BTF_TYPE_EMIT_ENUM(BPF_TCP_ESTABLISHED); 2996 2997 if (BPF_SOCK_OPS_TEST_FLAG(tcp_sk(sk), BPF_SOCK_OPS_STATE_CB_FLAG)) 2998 tcp_call_bpf_2arg(sk, BPF_SOCK_OPS_STATE_CB, oldstate, state); 2999 3000 switch (state) { 3001 case TCP_ESTABLISHED: 3002 if (oldstate != TCP_ESTABLISHED) 3003 TCP_INC_STATS(sock_net(sk), TCP_MIB_CURRESTAB); 3004 break; 3005 case TCP_CLOSE_WAIT: 3006 if (oldstate == TCP_SYN_RECV) 3007 TCP_INC_STATS(sock_net(sk), TCP_MIB_CURRESTAB); 3008 break; 3009 3010 case TCP_CLOSE: 3011 if (oldstate == TCP_CLOSE_WAIT || oldstate == TCP_ESTABLISHED) 3012 TCP_INC_STATS(sock_net(sk), TCP_MIB_ESTABRESETS); 3013 3014 sk->sk_prot->unhash(sk); 3015 if (inet_csk(sk)->icsk_bind_hash && 3016 !(sk->sk_userlocks & SOCK_BINDPORT_LOCK)) 3017 inet_put_port(sk); 3018 fallthrough; 3019 default: 3020 if (oldstate == TCP_ESTABLISHED || oldstate == TCP_CLOSE_WAIT) 3021 TCP_DEC_STATS(sock_net(sk), TCP_MIB_CURRESTAB); 3022 } 3023 3024 /* Change state AFTER socket is unhashed to avoid closed 3025 * socket sitting in hash tables. 3026 */ 3027 inet_sk_state_store(sk, state); 3028 } 3029 3030 /* 3031 * State processing on a close. This implements the state shift for 3032 * sending our FIN frame. Note that we only send a FIN for some 3033 * states. A shutdown() may have already sent the FIN, or we may be 3034 * closed. 3035 */ 3036 3037 static const unsigned char new_state[16] = { 3038 /* current state: new state: action: */ 3039 [0 /* (Invalid) */] = TCP_CLOSE, 3040 [TCP_ESTABLISHED] = TCP_FIN_WAIT1 | TCP_ACTION_FIN, 3041 [TCP_SYN_SENT] = TCP_CLOSE, 3042 [TCP_SYN_RECV] = TCP_FIN_WAIT1 | TCP_ACTION_FIN, 3043 [TCP_FIN_WAIT1] = TCP_FIN_WAIT1, 3044 [TCP_FIN_WAIT2] = TCP_FIN_WAIT2, 3045 [TCP_TIME_WAIT] = TCP_CLOSE, 3046 [TCP_CLOSE] = TCP_CLOSE, 3047 [TCP_CLOSE_WAIT] = TCP_LAST_ACK | TCP_ACTION_FIN, 3048 [TCP_LAST_ACK] = TCP_LAST_ACK, 3049 [TCP_LISTEN] = TCP_CLOSE, 3050 [TCP_CLOSING] = TCP_CLOSING, 3051 [TCP_NEW_SYN_RECV] = TCP_CLOSE, /* should not happen ! */ 3052 }; 3053 3054 static int tcp_close_state(struct sock *sk) 3055 { 3056 int next = (int)new_state[sk->sk_state]; 3057 int ns = next & TCP_STATE_MASK; 3058 3059 tcp_set_state(sk, ns); 3060 3061 return next & TCP_ACTION_FIN; 3062 } 3063 3064 /* 3065 * Shutdown the sending side of a connection. Much like close except 3066 * that we don't receive shut down or sock_set_flag(sk, SOCK_DEAD). 3067 */ 3068 3069 void tcp_shutdown(struct sock *sk, int how) 3070 { 3071 /* We need to grab some memory, and put together a FIN, 3072 * and then put it into the queue to be sent. 3073 * Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92. 3074 */ 3075 if (!(how & SEND_SHUTDOWN)) 3076 return; 3077 3078 /* If we've already sent a FIN, or it's a closed state, skip this. */ 3079 if ((1 << sk->sk_state) & 3080 (TCPF_ESTABLISHED | TCPF_SYN_SENT | 3081 TCPF_CLOSE_WAIT)) { 3082 /* Clear out any half completed packets. FIN if needed. */ 3083 if (tcp_close_state(sk)) 3084 tcp_send_fin(sk); 3085 } 3086 } 3087 3088 int tcp_orphan_count_sum(void) 3089 { 3090 int i, total = 0; 3091 3092 for_each_possible_cpu(i) 3093 total += per_cpu(tcp_orphan_count, i); 3094 3095 return max(total, 0); 3096 } 3097 3098 static int tcp_orphan_cache; 3099 static struct timer_list tcp_orphan_timer; 3100 #define TCP_ORPHAN_TIMER_PERIOD msecs_to_jiffies(100) 3101 3102 static void tcp_orphan_update(struct timer_list *unused) 3103 { 3104 WRITE_ONCE(tcp_orphan_cache, tcp_orphan_count_sum()); 3105 mod_timer(&tcp_orphan_timer, jiffies + TCP_ORPHAN_TIMER_PERIOD); 3106 } 3107 3108 static bool tcp_too_many_orphans(int shift) 3109 { 3110 return READ_ONCE(tcp_orphan_cache) << shift > 3111 READ_ONCE(sysctl_tcp_max_orphans); 3112 } 3113 3114 static bool tcp_out_of_memory(const struct sock *sk) 3115 { 3116 if (sk->sk_wmem_queued > SOCK_MIN_SNDBUF && 3117 sk_memory_allocated(sk) > sk_prot_mem_limits(sk, 2)) 3118 return true; 3119 return false; 3120 } 3121 3122 bool tcp_check_oom(const struct sock *sk, int shift) 3123 { 3124 bool too_many_orphans, out_of_socket_memory; 3125 3126 too_many_orphans = tcp_too_many_orphans(shift); 3127 out_of_socket_memory = tcp_out_of_memory(sk); 3128 3129 if (too_many_orphans) 3130 net_info_ratelimited("too many orphaned sockets\n"); 3131 if (out_of_socket_memory) 3132 net_info_ratelimited("out of memory -- consider tuning tcp_mem\n"); 3133 return too_many_orphans || out_of_socket_memory; 3134 } 3135 3136 void __tcp_close(struct sock *sk, long timeout) 3137 { 3138 bool data_was_unread = false; 3139 struct sk_buff *skb; 3140 int state; 3141 3142 WRITE_ONCE(sk->sk_shutdown, SHUTDOWN_MASK); 3143 3144 if (sk->sk_state == TCP_LISTEN) { 3145 tcp_set_state(sk, TCP_CLOSE); 3146 3147 /* Special case. */ 3148 inet_csk_listen_stop(sk); 3149 3150 goto adjudge_to_death; 3151 } 3152 3153 /* We need to flush the recv. buffs. We do this only on the 3154 * descriptor close, not protocol-sourced closes, because the 3155 * reader process may not have drained the data yet! 3156 */ 3157 while ((skb = skb_peek(&sk->sk_receive_queue)) != NULL) { 3158 u32 end_seq = TCP_SKB_CB(skb)->end_seq; 3159 3160 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) 3161 end_seq--; 3162 if (after(end_seq, tcp_sk(sk)->copied_seq)) 3163 data_was_unread = true; 3164 tcp_eat_recv_skb(sk, skb); 3165 } 3166 3167 /* If socket has been already reset (e.g. in tcp_reset()) - kill it. */ 3168 if (sk->sk_state == TCP_CLOSE) 3169 goto adjudge_to_death; 3170 3171 /* As outlined in RFC 2525, section 2.17, we send a RST here because 3172 * data was lost. To witness the awful effects of the old behavior of 3173 * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk 3174 * GET in an FTP client, suspend the process, wait for the client to 3175 * advertise a zero window, then kill -9 the FTP client, wheee... 3176 * Note: timeout is always zero in such a case. 3177 */ 3178 if (unlikely(tcp_sk(sk)->repair)) { 3179 sk->sk_prot->disconnect(sk, 0); 3180 } else if (data_was_unread) { 3181 /* Unread data was tossed, zap the connection. */ 3182 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONCLOSE); 3183 tcp_set_state(sk, TCP_CLOSE); 3184 tcp_send_active_reset(sk, sk->sk_allocation, 3185 SK_RST_REASON_TCP_ABORT_ON_CLOSE); 3186 } else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) { 3187 /* Check zero linger _after_ checking for unread data. */ 3188 sk->sk_prot->disconnect(sk, 0); 3189 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA); 3190 } else if (tcp_close_state(sk)) { 3191 /* We FIN if the application ate all the data before 3192 * zapping the connection. 3193 */ 3194 3195 /* RED-PEN. Formally speaking, we have broken TCP state 3196 * machine. State transitions: 3197 * 3198 * TCP_ESTABLISHED -> TCP_FIN_WAIT1 3199 * TCP_SYN_RECV -> TCP_FIN_WAIT1 (it is difficult) 3200 * TCP_CLOSE_WAIT -> TCP_LAST_ACK 3201 * 3202 * are legal only when FIN has been sent (i.e. in window), 3203 * rather than queued out of window. Purists blame. 3204 * 3205 * F.e. "RFC state" is ESTABLISHED, 3206 * if Linux state is FIN-WAIT-1, but FIN is still not sent. 3207 * 3208 * The visible declinations are that sometimes 3209 * we enter time-wait state, when it is not required really 3210 * (harmless), do not send active resets, when they are 3211 * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when 3212 * they look as CLOSING or LAST_ACK for Linux) 3213 * Probably, I missed some more holelets. 3214 * --ANK 3215 * XXX (TFO) - To start off we don't support SYN+ACK+FIN 3216 * in a single packet! (May consider it later but will 3217 * probably need API support or TCP_CORK SYN-ACK until 3218 * data is written and socket is closed.) 3219 */ 3220 tcp_send_fin(sk); 3221 } 3222 3223 sk_stream_wait_close(sk, timeout); 3224 3225 adjudge_to_death: 3226 state = sk->sk_state; 3227 sock_hold(sk); 3228 sock_orphan(sk); 3229 3230 local_bh_disable(); 3231 bh_lock_sock(sk); 3232 /* remove backlog if any, without releasing ownership. */ 3233 __release_sock(sk); 3234 3235 tcp_orphan_count_inc(); 3236 3237 /* Have we already been destroyed by a softirq or backlog? */ 3238 if (state != TCP_CLOSE && sk->sk_state == TCP_CLOSE) 3239 goto out; 3240 3241 /* This is a (useful) BSD violating of the RFC. There is a 3242 * problem with TCP as specified in that the other end could 3243 * keep a socket open forever with no application left this end. 3244 * We use a 1 minute timeout (about the same as BSD) then kill 3245 * our end. If they send after that then tough - BUT: long enough 3246 * that we won't make the old 4*rto = almost no time - whoops 3247 * reset mistake. 3248 * 3249 * Nope, it was not mistake. It is really desired behaviour 3250 * f.e. on http servers, when such sockets are useless, but 3251 * consume significant resources. Let's do it with special 3252 * linger2 option. --ANK 3253 */ 3254 3255 if (sk->sk_state == TCP_FIN_WAIT2) { 3256 struct tcp_sock *tp = tcp_sk(sk); 3257 if (READ_ONCE(tp->linger2) < 0) { 3258 tcp_set_state(sk, TCP_CLOSE); 3259 tcp_send_active_reset(sk, GFP_ATOMIC, 3260 SK_RST_REASON_TCP_ABORT_ON_LINGER); 3261 __NET_INC_STATS(sock_net(sk), 3262 LINUX_MIB_TCPABORTONLINGER); 3263 } else { 3264 const int tmo = tcp_fin_time(sk); 3265 3266 if (tmo > TCP_TIMEWAIT_LEN) { 3267 tcp_reset_keepalive_timer(sk, 3268 tmo - TCP_TIMEWAIT_LEN); 3269 } else { 3270 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo); 3271 goto out; 3272 } 3273 } 3274 } 3275 if (sk->sk_state != TCP_CLOSE) { 3276 if (tcp_check_oom(sk, 0)) { 3277 tcp_set_state(sk, TCP_CLOSE); 3278 tcp_send_active_reset(sk, GFP_ATOMIC, 3279 SK_RST_REASON_TCP_ABORT_ON_MEMORY); 3280 __NET_INC_STATS(sock_net(sk), 3281 LINUX_MIB_TCPABORTONMEMORY); 3282 } else if (!check_net(sock_net(sk))) { 3283 /* Not possible to send reset; just close */ 3284 tcp_set_state(sk, TCP_CLOSE); 3285 } 3286 } 3287 3288 if (sk->sk_state == TCP_CLOSE) { 3289 struct request_sock *req; 3290 3291 req = rcu_dereference_protected(tcp_sk(sk)->fastopen_rsk, 3292 lockdep_sock_is_held(sk)); 3293 /* We could get here with a non-NULL req if the socket is 3294 * aborted (e.g., closed with unread data) before 3WHS 3295 * finishes. 3296 */ 3297 if (req) 3298 reqsk_fastopen_remove(sk, req, false); 3299 inet_csk_destroy_sock(sk); 3300 } 3301 /* Otherwise, socket is reprieved until protocol close. */ 3302 3303 out: 3304 bh_unlock_sock(sk); 3305 local_bh_enable(); 3306 } 3307 3308 void tcp_close(struct sock *sk, long timeout) 3309 { 3310 lock_sock(sk); 3311 __tcp_close(sk, timeout); 3312 release_sock(sk); 3313 if (!sk->sk_net_refcnt) 3314 inet_csk_clear_xmit_timers_sync(sk); 3315 sock_put(sk); 3316 } 3317 EXPORT_SYMBOL(tcp_close); 3318 3319 /* These states need RST on ABORT according to RFC793 */ 3320 3321 static inline bool tcp_need_reset(int state) 3322 { 3323 return (1 << state) & 3324 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 | 3325 TCPF_FIN_WAIT2 | TCPF_SYN_RECV); 3326 } 3327 3328 static void tcp_rtx_queue_purge(struct sock *sk) 3329 { 3330 struct rb_node *p = rb_first(&sk->tcp_rtx_queue); 3331 3332 tcp_sk(sk)->highest_sack = NULL; 3333 while (p) { 3334 struct sk_buff *skb = rb_to_skb(p); 3335 3336 p = rb_next(p); 3337 /* Since we are deleting whole queue, no need to 3338 * list_del(&skb->tcp_tsorted_anchor) 3339 */ 3340 tcp_rtx_queue_unlink(skb, sk); 3341 tcp_wmem_free_skb(sk, skb); 3342 } 3343 } 3344 3345 void tcp_write_queue_purge(struct sock *sk) 3346 { 3347 struct sk_buff *skb; 3348 3349 tcp_chrono_stop(sk, TCP_CHRONO_BUSY); 3350 while ((skb = __skb_dequeue(&sk->sk_write_queue)) != NULL) { 3351 tcp_skb_tsorted_anchor_cleanup(skb); 3352 tcp_wmem_free_skb(sk, skb); 3353 } 3354 tcp_rtx_queue_purge(sk); 3355 INIT_LIST_HEAD(&tcp_sk(sk)->tsorted_sent_queue); 3356 tcp_clear_all_retrans_hints(tcp_sk(sk)); 3357 tcp_sk(sk)->packets_out = 0; 3358 inet_csk(sk)->icsk_backoff = 0; 3359 } 3360 3361 int tcp_disconnect(struct sock *sk, int flags) 3362 { 3363 struct inet_sock *inet = inet_sk(sk); 3364 struct inet_connection_sock *icsk = inet_csk(sk); 3365 struct tcp_sock *tp = tcp_sk(sk); 3366 int old_state = sk->sk_state; 3367 struct request_sock *req; 3368 u32 seq; 3369 3370 if (old_state != TCP_CLOSE) 3371 tcp_set_state(sk, TCP_CLOSE); 3372 3373 /* ABORT function of RFC793 */ 3374 if (old_state == TCP_LISTEN) { 3375 inet_csk_listen_stop(sk); 3376 } else if (unlikely(tp->repair)) { 3377 WRITE_ONCE(sk->sk_err, ECONNABORTED); 3378 } else if (tcp_need_reset(old_state)) { 3379 tcp_send_active_reset(sk, gfp_any(), SK_RST_REASON_TCP_STATE); 3380 WRITE_ONCE(sk->sk_err, ECONNRESET); 3381 } else if (tp->snd_nxt != tp->write_seq && 3382 (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK)) { 3383 /* The last check adjusts for discrepancy of Linux wrt. RFC 3384 * states 3385 */ 3386 tcp_send_active_reset(sk, gfp_any(), 3387 SK_RST_REASON_TCP_DISCONNECT_WITH_DATA); 3388 WRITE_ONCE(sk->sk_err, ECONNRESET); 3389 } else if (old_state == TCP_SYN_SENT) 3390 WRITE_ONCE(sk->sk_err, ECONNRESET); 3391 3392 tcp_clear_xmit_timers(sk); 3393 __skb_queue_purge(&sk->sk_receive_queue); 3394 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt); 3395 WRITE_ONCE(tp->urg_data, 0); 3396 sk_set_peek_off(sk, -1); 3397 tcp_write_queue_purge(sk); 3398 tcp_fastopen_active_disable_ofo_check(sk); 3399 skb_rbtree_purge(&tp->out_of_order_queue); 3400 3401 inet->inet_dport = 0; 3402 3403 inet_bhash2_reset_saddr(sk); 3404 3405 WRITE_ONCE(sk->sk_shutdown, 0); 3406 sock_reset_flag(sk, SOCK_DONE); 3407 tp->srtt_us = 0; 3408 tp->mdev_us = jiffies_to_usecs(TCP_TIMEOUT_INIT); 3409 tp->rcv_rtt_last_tsecr = 0; 3410 3411 seq = tp->write_seq + tp->max_window + 2; 3412 if (!seq) 3413 seq = 1; 3414 WRITE_ONCE(tp->write_seq, seq); 3415 3416 icsk->icsk_backoff = 0; 3417 WRITE_ONCE(icsk->icsk_probes_out, 0); 3418 icsk->icsk_probes_tstamp = 0; 3419 icsk->icsk_rto = TCP_TIMEOUT_INIT; 3420 WRITE_ONCE(icsk->icsk_rto_min, TCP_RTO_MIN); 3421 WRITE_ONCE(icsk->icsk_delack_max, TCP_DELACK_MAX); 3422 WRITE_ONCE(tp->snd_ssthresh, TCP_INFINITE_SSTHRESH); 3423 tcp_snd_cwnd_set(tp, TCP_INIT_CWND); 3424 tp->snd_cwnd_cnt = 0; 3425 tp->is_cwnd_limited = 0; 3426 tp->max_packets_out = 0; 3427 tp->window_clamp = 0; 3428 tp->delivered = 0; 3429 tp->delivered_ce = 0; 3430 tp->accecn_fail_mode = 0; 3431 tp->saw_accecn_opt = TCP_ACCECN_OPT_NOT_SEEN; 3432 tcp_accecn_init_counters(tp); 3433 tp->prev_ecnfield = 0; 3434 tp->accecn_opt_tstamp = 0; 3435 tp->pkts_acked_ewma = 0; 3436 if (icsk->icsk_ca_initialized && icsk->icsk_ca_ops->release) 3437 icsk->icsk_ca_ops->release(sk); 3438 memset(icsk->icsk_ca_priv, 0, sizeof(icsk->icsk_ca_priv)); 3439 icsk->icsk_ca_initialized = 0; 3440 tcp_set_ca_state(sk, TCP_CA_Open); 3441 tp->is_sack_reneg = 0; 3442 tcp_clear_retrans(tp); 3443 tp->total_retrans = 0; 3444 inet_csk_delack_init(sk); 3445 /* Initialize rcv_mss to TCP_MIN_MSS to avoid division by 0 3446 * issue in __tcp_select_window() 3447 */ 3448 icsk->icsk_ack.rcv_mss = TCP_MIN_MSS; 3449 memset(&tp->rx_opt, 0, sizeof(tp->rx_opt)); 3450 __sk_dst_reset(sk); 3451 dst_release(unrcu_pointer(xchg(&sk->sk_rx_dst, NULL))); 3452 tcp_saved_syn_free(tp); 3453 tp->compressed_ack = 0; 3454 tp->segs_in = 0; 3455 tp->segs_out = 0; 3456 tp->bytes_sent = 0; 3457 tp->bytes_acked = 0; 3458 tp->bytes_received = 0; 3459 tp->bytes_retrans = 0; 3460 tp->data_segs_in = 0; 3461 tp->data_segs_out = 0; 3462 tp->duplicate_sack[0].start_seq = 0; 3463 tp->duplicate_sack[0].end_seq = 0; 3464 tp->dsack_dups = 0; 3465 tp->reord_seen = 0; 3466 tp->retrans_out = 0; 3467 tp->sacked_out = 0; 3468 tp->tlp_high_seq = 0; 3469 tp->last_oow_ack_time = 0; 3470 tp->plb_rehash = 0; 3471 /* There's a bubble in the pipe until at least the first ACK. */ 3472 tp->app_limited = ~0U; 3473 tp->rate_app_limited = 1; 3474 tp->rack.mstamp = 0; 3475 tp->rack.advanced = 0; 3476 tp->rack.reo_wnd_steps = 1; 3477 tp->rack.last_delivered = 0; 3478 tp->rack.reo_wnd_persist = 0; 3479 tp->rack.dsack_seen = 0; 3480 tp->syn_data_acked = 0; 3481 tp->syn_fastopen_child = 0; 3482 tp->rx_opt.saw_tstamp = 0; 3483 tp->rx_opt.dsack = 0; 3484 tp->rx_opt.num_sacks = 0; 3485 tp->rcv_ooopack = 0; 3486 3487 3488 /* Clean up fastopen related fields */ 3489 req = rcu_dereference_protected(tp->fastopen_rsk, 3490 lockdep_sock_is_held(sk)); 3491 if (req) 3492 reqsk_fastopen_remove(sk, req, false); 3493 tcp_free_fastopen_req(tp); 3494 inet_clear_bit(DEFER_CONNECT, sk); 3495 tp->fastopen_client_fail = 0; 3496 3497 WARN_ON(inet->inet_num && !icsk->icsk_bind_hash); 3498 3499 if (sk->sk_frag.page) { 3500 put_page(sk->sk_frag.page); 3501 sk->sk_frag.page = NULL; 3502 sk->sk_frag.offset = 0; 3503 } 3504 sk_error_report(sk); 3505 return 0; 3506 } 3507 3508 static inline bool tcp_can_repair_sock(const struct sock *sk) 3509 { 3510 return sockopt_ns_capable(sock_net(sk)->user_ns, CAP_NET_ADMIN) && 3511 (sk->sk_state != TCP_LISTEN); 3512 } 3513 3514 static int tcp_repair_set_window(struct tcp_sock *tp, sockptr_t optbuf, int len) 3515 { 3516 struct tcp_repair_window opt; 3517 3518 if (!tp->repair) 3519 return -EPERM; 3520 3521 if (len != sizeof(opt)) 3522 return -EINVAL; 3523 3524 if (copy_from_sockptr(&opt, optbuf, sizeof(opt))) 3525 return -EFAULT; 3526 3527 if (opt.max_window < opt.snd_wnd) 3528 return -EINVAL; 3529 3530 if (after(opt.snd_wl1, tp->rcv_nxt + opt.rcv_wnd)) 3531 return -EINVAL; 3532 3533 if (after(opt.rcv_wup, tp->rcv_nxt)) 3534 return -EINVAL; 3535 3536 tp->snd_wl1 = opt.snd_wl1; 3537 tp->snd_wnd = opt.snd_wnd; 3538 tp->max_window = opt.max_window; 3539 3540 tp->rcv_wnd = opt.rcv_wnd; 3541 tp->rcv_wup = opt.rcv_wup; 3542 tp->rcv_mwnd_seq = opt.rcv_wup + opt.rcv_wnd; 3543 3544 return 0; 3545 } 3546 3547 static int tcp_repair_options_est(struct sock *sk, sockptr_t optbuf, 3548 unsigned int len) 3549 { 3550 struct tcp_sock *tp = tcp_sk(sk); 3551 struct tcp_repair_opt opt; 3552 size_t offset = 0; 3553 3554 while (len >= sizeof(opt)) { 3555 if (copy_from_sockptr_offset(&opt, optbuf, offset, sizeof(opt))) 3556 return -EFAULT; 3557 3558 offset += sizeof(opt); 3559 len -= sizeof(opt); 3560 3561 switch (opt.opt_code) { 3562 case TCPOPT_MSS: 3563 tp->rx_opt.mss_clamp = opt.opt_val; 3564 tcp_mtup_init(sk); 3565 break; 3566 case TCPOPT_WINDOW: 3567 { 3568 u16 snd_wscale = opt.opt_val & 0xFFFF; 3569 u16 rcv_wscale = opt.opt_val >> 16; 3570 3571 if (snd_wscale > TCP_MAX_WSCALE || rcv_wscale > TCP_MAX_WSCALE) 3572 return -EFBIG; 3573 3574 tp->rx_opt.snd_wscale = snd_wscale; 3575 tp->rx_opt.rcv_wscale = rcv_wscale; 3576 tp->rx_opt.wscale_ok = 1; 3577 } 3578 break; 3579 case TCPOPT_SACK_PERM: 3580 if (opt.opt_val != 0) 3581 return -EINVAL; 3582 3583 tp->rx_opt.sack_ok |= TCP_SACK_SEEN; 3584 break; 3585 case TCPOPT_TIMESTAMP: 3586 if (opt.opt_val != 0) 3587 return -EINVAL; 3588 3589 tp->rx_opt.tstamp_ok = 1; 3590 break; 3591 } 3592 } 3593 3594 return 0; 3595 } 3596 3597 DEFINE_STATIC_KEY_FALSE(tcp_tx_delay_enabled); 3598 3599 static void tcp_enable_tx_delay(struct sock *sk, int val) 3600 { 3601 struct tcp_sock *tp = tcp_sk(sk); 3602 s32 delta = (val - tp->tcp_tx_delay) << 3; 3603 3604 if (val && !static_branch_unlikely(&tcp_tx_delay_enabled)) { 3605 static int __tcp_tx_delay_enabled = 0; 3606 3607 if (cmpxchg(&__tcp_tx_delay_enabled, 0, 1) == 0) { 3608 static_branch_enable(&tcp_tx_delay_enabled); 3609 pr_info("TCP_TX_DELAY enabled\n"); 3610 } 3611 } 3612 /* If we change tcp_tx_delay on a live flow, adjust tp->srtt_us, 3613 * tp->rtt_min, icsk_rto and sk->sk_pacing_rate. 3614 * This is best effort. 3615 */ 3616 if (delta && sk->sk_state == TCP_ESTABLISHED) { 3617 s64 srtt = (s64)tp->srtt_us + delta; 3618 3619 WRITE_ONCE(tp->srtt_us, 3620 clamp_t(s64, srtt, 1, ~0U)); 3621 3622 /* Note: does not deal with non zero icsk_backoff */ 3623 tcp_set_rto(sk); 3624 3625 minmax_reset(&tp->rtt_min, tcp_jiffies32, ~0U); 3626 3627 tcp_update_pacing_rate(sk); 3628 } 3629 } 3630 3631 /* When set indicates to always queue non-full frames. Later the user clears 3632 * this option and we transmit any pending partial frames in the queue. This is 3633 * meant to be used alongside sendfile() to get properly filled frames when the 3634 * user (for example) must write out headers with a write() call first and then 3635 * use sendfile to send out the data parts. 3636 * 3637 * TCP_CORK can be set together with TCP_NODELAY and it is stronger than 3638 * TCP_NODELAY. 3639 */ 3640 void __tcp_sock_set_cork(struct sock *sk, bool on) 3641 { 3642 struct tcp_sock *tp = tcp_sk(sk); 3643 3644 if (on) { 3645 tp->nonagle |= TCP_NAGLE_CORK; 3646 } else { 3647 tp->nonagle &= ~TCP_NAGLE_CORK; 3648 if (tp->nonagle & TCP_NAGLE_OFF) 3649 tp->nonagle |= TCP_NAGLE_PUSH; 3650 tcp_push_pending_frames(sk); 3651 } 3652 } 3653 3654 void tcp_sock_set_cork(struct sock *sk, bool on) 3655 { 3656 lock_sock(sk); 3657 __tcp_sock_set_cork(sk, on); 3658 release_sock(sk); 3659 } 3660 EXPORT_SYMBOL(tcp_sock_set_cork); 3661 3662 /* TCP_NODELAY is weaker than TCP_CORK, so that this option on corked socket is 3663 * remembered, but it is not activated until cork is cleared. 3664 * 3665 * However, when TCP_NODELAY is set we make an explicit push, which overrides 3666 * even TCP_CORK for currently queued segments. 3667 */ 3668 void __tcp_sock_set_nodelay(struct sock *sk, bool on) 3669 { 3670 if (on) { 3671 tcp_sk(sk)->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH; 3672 tcp_push_pending_frames(sk); 3673 } else { 3674 tcp_sk(sk)->nonagle &= ~TCP_NAGLE_OFF; 3675 } 3676 } 3677 3678 void tcp_sock_set_nodelay(struct sock *sk) 3679 { 3680 lock_sock(sk); 3681 __tcp_sock_set_nodelay(sk, true); 3682 release_sock(sk); 3683 } 3684 EXPORT_SYMBOL(tcp_sock_set_nodelay); 3685 3686 static void __tcp_sock_set_quickack(struct sock *sk, int val) 3687 { 3688 if (!val) { 3689 inet_csk_enter_pingpong_mode(sk); 3690 return; 3691 } 3692 3693 inet_csk_exit_pingpong_mode(sk); 3694 if ((1 << sk->sk_state) & (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) && 3695 inet_csk_ack_scheduled(sk)) { 3696 inet_csk(sk)->icsk_ack.pending |= ICSK_ACK_PUSHED; 3697 tcp_cleanup_rbuf(sk, 1); 3698 if (!(val & 1)) 3699 inet_csk_enter_pingpong_mode(sk); 3700 } 3701 } 3702 3703 void tcp_sock_set_quickack(struct sock *sk, int val) 3704 { 3705 lock_sock(sk); 3706 __tcp_sock_set_quickack(sk, val); 3707 release_sock(sk); 3708 } 3709 EXPORT_SYMBOL(tcp_sock_set_quickack); 3710 3711 int tcp_sock_set_syncnt(struct sock *sk, int val) 3712 { 3713 if (val < 1 || val > MAX_TCP_SYNCNT) 3714 return -EINVAL; 3715 3716 WRITE_ONCE(inet_csk(sk)->icsk_syn_retries, val); 3717 return 0; 3718 } 3719 EXPORT_SYMBOL(tcp_sock_set_syncnt); 3720 3721 int tcp_sock_set_user_timeout(struct sock *sk, int val) 3722 { 3723 /* Cap the max time in ms TCP will retry or probe the window 3724 * before giving up and aborting (ETIMEDOUT) a connection. 3725 */ 3726 if (val < 0) 3727 return -EINVAL; 3728 3729 WRITE_ONCE(inet_csk(sk)->icsk_user_timeout, val); 3730 return 0; 3731 } 3732 EXPORT_SYMBOL(tcp_sock_set_user_timeout); 3733 3734 int tcp_sock_set_keepidle_locked(struct sock *sk, int val) 3735 { 3736 struct tcp_sock *tp = tcp_sk(sk); 3737 3738 if (val < 1 || val > MAX_TCP_KEEPIDLE) 3739 return -EINVAL; 3740 3741 /* Paired with WRITE_ONCE() in keepalive_time_when() */ 3742 WRITE_ONCE(tp->keepalive_time, val * HZ); 3743 if (sock_flag(sk, SOCK_KEEPOPEN) && 3744 !((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN))) { 3745 u32 elapsed = keepalive_time_elapsed(tp); 3746 3747 if (tp->keepalive_time > elapsed) 3748 elapsed = tp->keepalive_time - elapsed; 3749 else 3750 elapsed = 0; 3751 tcp_reset_keepalive_timer(sk, elapsed); 3752 } 3753 3754 return 0; 3755 } 3756 3757 int tcp_sock_set_keepidle(struct sock *sk, int val) 3758 { 3759 int err; 3760 3761 lock_sock(sk); 3762 err = tcp_sock_set_keepidle_locked(sk, val); 3763 release_sock(sk); 3764 return err; 3765 } 3766 EXPORT_SYMBOL(tcp_sock_set_keepidle); 3767 3768 int tcp_sock_set_keepintvl(struct sock *sk, int val) 3769 { 3770 if (val < 1 || val > MAX_TCP_KEEPINTVL) 3771 return -EINVAL; 3772 3773 WRITE_ONCE(tcp_sk(sk)->keepalive_intvl, val * HZ); 3774 return 0; 3775 } 3776 EXPORT_SYMBOL(tcp_sock_set_keepintvl); 3777 3778 int tcp_sock_set_keepcnt(struct sock *sk, int val) 3779 { 3780 if (val < 1 || val > MAX_TCP_KEEPCNT) 3781 return -EINVAL; 3782 3783 /* Paired with READ_ONCE() in keepalive_probes() */ 3784 WRITE_ONCE(tcp_sk(sk)->keepalive_probes, val); 3785 return 0; 3786 } 3787 EXPORT_SYMBOL(tcp_sock_set_keepcnt); 3788 3789 int tcp_set_window_clamp(struct sock *sk, int val) 3790 { 3791 u32 old_window_clamp, new_window_clamp, new_rcv_ssthresh; 3792 struct tcp_sock *tp = tcp_sk(sk); 3793 3794 if (!val) { 3795 if (sk->sk_state != TCP_CLOSE) 3796 return -EINVAL; 3797 WRITE_ONCE(tp->window_clamp, 0); 3798 return 0; 3799 } 3800 3801 old_window_clamp = tp->window_clamp; 3802 new_window_clamp = max_t(int, SOCK_MIN_RCVBUF / 2, val); 3803 3804 if (new_window_clamp == old_window_clamp) 3805 return 0; 3806 3807 WRITE_ONCE(tp->window_clamp, new_window_clamp); 3808 3809 /* Need to apply the reserved mem provisioning only 3810 * when shrinking the window clamp. 3811 */ 3812 if (new_window_clamp < old_window_clamp) { 3813 __tcp_adjust_rcv_ssthresh(sk, new_window_clamp); 3814 } else { 3815 new_rcv_ssthresh = min(tp->rcv_wnd, new_window_clamp); 3816 tp->rcv_ssthresh = max(new_rcv_ssthresh, tp->rcv_ssthresh); 3817 } 3818 return 0; 3819 } 3820 3821 int tcp_sock_set_maxseg(struct sock *sk, int val) 3822 { 3823 /* Values greater than interface MTU won't take effect. However 3824 * at the point when this call is done we typically don't yet 3825 * know which interface is going to be used 3826 */ 3827 if (val && (val < TCP_MIN_MSS || val > MAX_TCP_WINDOW)) 3828 return -EINVAL; 3829 3830 WRITE_ONCE(tcp_sk(sk)->rx_opt.user_mss, val); 3831 return 0; 3832 } 3833 3834 /* 3835 * Socket option code for TCP. 3836 */ 3837 int do_tcp_setsockopt(struct sock *sk, int level, int optname, 3838 sockptr_t optval, unsigned int optlen) 3839 { 3840 struct tcp_sock *tp = tcp_sk(sk); 3841 struct inet_connection_sock *icsk = inet_csk(sk); 3842 struct net *net = sock_net(sk); 3843 int val; 3844 int err = 0; 3845 3846 /* These are data/string values, all the others are ints */ 3847 switch (optname) { 3848 case TCP_CONGESTION: { 3849 char name[TCP_CA_NAME_MAX]; 3850 3851 if (optlen < 1) 3852 return -EINVAL; 3853 3854 val = strncpy_from_sockptr(name, optval, 3855 min_t(long, TCP_CA_NAME_MAX-1, optlen)); 3856 if (val < 0) 3857 return -EFAULT; 3858 name[val] = 0; 3859 3860 sockopt_lock_sock(sk); 3861 err = tcp_set_congestion_control(sk, name, !has_current_bpf_ctx(), 3862 sockopt_ns_capable(sock_net(sk)->user_ns, 3863 CAP_NET_ADMIN)); 3864 sockopt_release_sock(sk); 3865 return err; 3866 } 3867 case TCP_ULP: { 3868 char name[TCP_ULP_NAME_MAX]; 3869 3870 if (optlen < 1) 3871 return -EINVAL; 3872 3873 val = strncpy_from_sockptr(name, optval, 3874 min_t(long, TCP_ULP_NAME_MAX - 1, 3875 optlen)); 3876 if (val < 0) 3877 return -EFAULT; 3878 name[val] = 0; 3879 3880 sockopt_lock_sock(sk); 3881 err = tcp_set_ulp(sk, name); 3882 sockopt_release_sock(sk); 3883 return err; 3884 } 3885 case TCP_FASTOPEN_KEY: { 3886 __u8 key[TCP_FASTOPEN_KEY_BUF_LENGTH]; 3887 __u8 *backup_key = NULL; 3888 3889 /* Allow a backup key as well to facilitate key rotation 3890 * First key is the active one. 3891 */ 3892 if (optlen != TCP_FASTOPEN_KEY_LENGTH && 3893 optlen != TCP_FASTOPEN_KEY_BUF_LENGTH) 3894 return -EINVAL; 3895 3896 if (copy_from_sockptr(key, optval, optlen)) 3897 return -EFAULT; 3898 3899 if (optlen == TCP_FASTOPEN_KEY_BUF_LENGTH) 3900 backup_key = key + TCP_FASTOPEN_KEY_LENGTH; 3901 3902 return tcp_fastopen_reset_cipher(net, sk, key, backup_key); 3903 } 3904 default: 3905 /* fallthru */ 3906 break; 3907 } 3908 3909 if (optlen < sizeof(int)) 3910 return -EINVAL; 3911 3912 if (copy_from_sockptr(&val, optval, sizeof(val))) 3913 return -EFAULT; 3914 3915 /* Handle options that can be set without locking the socket. */ 3916 switch (optname) { 3917 case TCP_SYNCNT: 3918 return tcp_sock_set_syncnt(sk, val); 3919 case TCP_USER_TIMEOUT: 3920 return tcp_sock_set_user_timeout(sk, val); 3921 case TCP_KEEPINTVL: 3922 return tcp_sock_set_keepintvl(sk, val); 3923 case TCP_KEEPCNT: 3924 return tcp_sock_set_keepcnt(sk, val); 3925 case TCP_LINGER2: 3926 if (val < 0) 3927 WRITE_ONCE(tp->linger2, -1); 3928 else if (val > TCP_FIN_TIMEOUT_MAX / HZ) 3929 WRITE_ONCE(tp->linger2, TCP_FIN_TIMEOUT_MAX); 3930 else 3931 WRITE_ONCE(tp->linger2, val * HZ); 3932 return 0; 3933 case TCP_DEFER_ACCEPT: 3934 /* Translate value in seconds to number of retransmits */ 3935 WRITE_ONCE(icsk->icsk_accept_queue.rskq_defer_accept, 3936 secs_to_retrans(val, TCP_TIMEOUT_INIT / HZ, 3937 TCP_RTO_MAX / HZ)); 3938 return 0; 3939 case TCP_RTO_MAX_MS: 3940 if (val < MSEC_PER_SEC || val > TCP_RTO_MAX_SEC * MSEC_PER_SEC) 3941 return -EINVAL; 3942 WRITE_ONCE(inet_csk(sk)->icsk_rto_max, msecs_to_jiffies(val)); 3943 return 0; 3944 case TCP_RTO_MIN_US: { 3945 int rto_min = usecs_to_jiffies(val); 3946 3947 if (rto_min > TCP_RTO_MIN || rto_min < TCP_TIMEOUT_MIN) 3948 return -EINVAL; 3949 WRITE_ONCE(inet_csk(sk)->icsk_rto_min, rto_min); 3950 return 0; 3951 } 3952 case TCP_DELACK_MAX_US: { 3953 int delack_max = usecs_to_jiffies(val); 3954 3955 if (delack_max > TCP_DELACK_MAX || delack_max < TCP_TIMEOUT_MIN) 3956 return -EINVAL; 3957 WRITE_ONCE(inet_csk(sk)->icsk_delack_max, delack_max); 3958 return 0; 3959 } 3960 case TCP_MAXSEG: 3961 return tcp_sock_set_maxseg(sk, val); 3962 } 3963 3964 sockopt_lock_sock(sk); 3965 3966 switch (optname) { 3967 case TCP_NODELAY: 3968 __tcp_sock_set_nodelay(sk, val); 3969 break; 3970 3971 case TCP_THIN_LINEAR_TIMEOUTS: 3972 if (val < 0 || val > 1) 3973 err = -EINVAL; 3974 else 3975 tp->thin_lto = val; 3976 break; 3977 3978 case TCP_THIN_DUPACK: 3979 if (val < 0 || val > 1) 3980 err = -EINVAL; 3981 break; 3982 3983 case TCP_REPAIR: 3984 if (!tcp_can_repair_sock(sk)) 3985 err = -EPERM; 3986 else if (val == TCP_REPAIR_ON) { 3987 tp->repair = 1; 3988 sk->sk_reuse = SK_FORCE_REUSE; 3989 tp->repair_queue = TCP_NO_QUEUE; 3990 } else if (val == TCP_REPAIR_OFF) { 3991 tp->repair = 0; 3992 sk->sk_reuse = SK_NO_REUSE; 3993 tcp_send_window_probe(sk); 3994 } else if (val == TCP_REPAIR_OFF_NO_WP) { 3995 tp->repair = 0; 3996 sk->sk_reuse = SK_NO_REUSE; 3997 } else 3998 err = -EINVAL; 3999 4000 break; 4001 4002 case TCP_REPAIR_QUEUE: 4003 if (!tp->repair) 4004 err = -EPERM; 4005 else if ((unsigned int)val < TCP_QUEUES_NR) 4006 tp->repair_queue = val; 4007 else 4008 err = -EINVAL; 4009 break; 4010 4011 case TCP_QUEUE_SEQ: 4012 if (sk->sk_state != TCP_CLOSE) { 4013 err = -EPERM; 4014 } else if (tp->repair_queue == TCP_SEND_QUEUE) { 4015 if (!tcp_rtx_queue_empty(sk)) 4016 err = -EPERM; 4017 else 4018 WRITE_ONCE(tp->write_seq, val); 4019 } else if (tp->repair_queue == TCP_RECV_QUEUE) { 4020 if (tp->rcv_nxt != tp->copied_seq) { 4021 err = -EPERM; 4022 } else { 4023 WRITE_ONCE(tp->rcv_nxt, val); 4024 WRITE_ONCE(tp->copied_seq, val); 4025 } 4026 } else { 4027 err = -EINVAL; 4028 } 4029 break; 4030 4031 case TCP_REPAIR_OPTIONS: 4032 if (!tp->repair) 4033 err = -EINVAL; 4034 else if (sk->sk_state == TCP_ESTABLISHED && !tp->bytes_sent) 4035 err = tcp_repair_options_est(sk, optval, optlen); 4036 else 4037 err = -EPERM; 4038 break; 4039 4040 case TCP_CORK: 4041 __tcp_sock_set_cork(sk, val); 4042 break; 4043 4044 case TCP_KEEPIDLE: 4045 err = tcp_sock_set_keepidle_locked(sk, val); 4046 break; 4047 case TCP_SAVE_SYN: 4048 /* 0: disable, 1: enable, 2: start from ether_header */ 4049 if (val < 0 || val > 2) 4050 err = -EINVAL; 4051 else 4052 tp->save_syn = val; 4053 break; 4054 4055 case TCP_WINDOW_CLAMP: 4056 err = tcp_set_window_clamp(sk, val); 4057 break; 4058 4059 case TCP_QUICKACK: 4060 __tcp_sock_set_quickack(sk, val); 4061 break; 4062 4063 case TCP_AO_REPAIR: 4064 if (!tcp_can_repair_sock(sk)) { 4065 err = -EPERM; 4066 break; 4067 } 4068 err = tcp_ao_set_repair(sk, optval, optlen); 4069 break; 4070 #ifdef CONFIG_TCP_AO 4071 case TCP_AO_ADD_KEY: 4072 case TCP_AO_DEL_KEY: 4073 case TCP_AO_INFO: { 4074 /* If this is the first TCP-AO setsockopt() on the socket, 4075 * sk_state has to be LISTEN or CLOSE. Allow TCP_REPAIR 4076 * in any state. 4077 */ 4078 if ((1 << sk->sk_state) & (TCPF_LISTEN | TCPF_CLOSE)) 4079 goto ao_parse; 4080 if (rcu_dereference_protected(tcp_sk(sk)->ao_info, 4081 lockdep_sock_is_held(sk))) 4082 goto ao_parse; 4083 if (tp->repair) 4084 goto ao_parse; 4085 err = -EISCONN; 4086 break; 4087 ao_parse: 4088 err = tp->af_specific->ao_parse(sk, optname, optval, optlen); 4089 break; 4090 } 4091 #endif 4092 #ifdef CONFIG_TCP_MD5SIG 4093 case TCP_MD5SIG: 4094 case TCP_MD5SIG_EXT: 4095 err = tp->af_specific->md5_parse(sk, optname, optval, optlen); 4096 break; 4097 #endif 4098 case TCP_FASTOPEN: 4099 if (val >= 0 && ((1 << sk->sk_state) & (TCPF_CLOSE | 4100 TCPF_LISTEN))) { 4101 tcp_fastopen_init_key_once(net); 4102 4103 fastopen_queue_tune(sk, val); 4104 } else { 4105 err = -EINVAL; 4106 } 4107 break; 4108 case TCP_FASTOPEN_CONNECT: 4109 if (val > 1 || val < 0) { 4110 err = -EINVAL; 4111 } else if (READ_ONCE(net->ipv4.sysctl_tcp_fastopen) & 4112 TFO_CLIENT_ENABLE) { 4113 if (sk->sk_state == TCP_CLOSE) 4114 tp->fastopen_connect = val; 4115 else 4116 err = -EINVAL; 4117 } else { 4118 err = -EOPNOTSUPP; 4119 } 4120 break; 4121 case TCP_FASTOPEN_NO_COOKIE: 4122 if (val > 1 || val < 0) 4123 err = -EINVAL; 4124 else if (!((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN))) 4125 err = -EINVAL; 4126 else 4127 tp->fastopen_no_cookie = val; 4128 break; 4129 case TCP_TIMESTAMP: 4130 if (!tp->repair) { 4131 err = -EPERM; 4132 break; 4133 } 4134 /* val is an opaque field, 4135 * and low order bit contains usec_ts enable bit. 4136 * Its a best effort, and we do not care if user makes an error. 4137 */ 4138 tp->tcp_usec_ts = val & 1; 4139 WRITE_ONCE(tp->tsoffset, val - tcp_clock_ts(tp->tcp_usec_ts)); 4140 break; 4141 case TCP_REPAIR_WINDOW: 4142 err = tcp_repair_set_window(tp, optval, optlen); 4143 break; 4144 case TCP_NOTSENT_LOWAT: 4145 WRITE_ONCE(tp->notsent_lowat, val); 4146 READ_ONCE(sk->sk_write_space)(sk); 4147 break; 4148 case TCP_INQ: 4149 if (val > 1 || val < 0) 4150 err = -EINVAL; 4151 else 4152 tp->recvmsg_inq = val; 4153 break; 4154 case TCP_TX_DELAY: 4155 /* tp->srtt_us is u32, and is shifted by 3 */ 4156 if (val < 0 || val >= (1U << (31 - 3))) { 4157 err = -EINVAL; 4158 break; 4159 } 4160 tcp_enable_tx_delay(sk, val); 4161 WRITE_ONCE(tp->tcp_tx_delay, val); 4162 break; 4163 default: 4164 err = -ENOPROTOOPT; 4165 break; 4166 } 4167 4168 sockopt_release_sock(sk); 4169 return err; 4170 } 4171 4172 int tcp_setsockopt(struct sock *sk, int level, int optname, sockptr_t optval, 4173 unsigned int optlen) 4174 { 4175 const struct inet_connection_sock *icsk = inet_csk(sk); 4176 4177 if (level != SOL_TCP) 4178 /* Paired with WRITE_ONCE() in do_ipv6_setsockopt() and tcp_v6_connect() */ 4179 return READ_ONCE(icsk->icsk_af_ops)->setsockopt(sk, level, optname, 4180 optval, optlen); 4181 return do_tcp_setsockopt(sk, level, optname, optval, optlen); 4182 } 4183 4184 static void tcp_get_info_chrono_stats(const struct tcp_sock *tp, 4185 struct tcp_info *info) 4186 { 4187 u64 stats[__TCP_CHRONO_MAX], total = 0; 4188 enum tcp_chrono i, cur; 4189 4190 /* Following READ_ONCE()s pair with WRITE_ONCE()s in tcp_chrono_set(). 4191 * This is because socket lock might not be owned by us at this point. 4192 * This is best effort, tcp_get_timestamping_opt_stats() can 4193 * see wrong values. A real fix would be too costly for TCP fast path. 4194 */ 4195 cur = READ_ONCE(tp->chrono_type); 4196 for (i = TCP_CHRONO_BUSY; i < __TCP_CHRONO_MAX; ++i) { 4197 stats[i] = READ_ONCE(tp->chrono_stat[i - 1]); 4198 if (i == cur) 4199 stats[i] += tcp_jiffies32 - READ_ONCE(tp->chrono_start); 4200 stats[i] *= USEC_PER_SEC / HZ; 4201 total += stats[i]; 4202 } 4203 4204 info->tcpi_busy_time = total; 4205 info->tcpi_rwnd_limited = stats[TCP_CHRONO_RWND_LIMITED]; 4206 info->tcpi_sndbuf_limited = stats[TCP_CHRONO_SNDBUF_LIMITED]; 4207 } 4208 4209 /* Return information about state of tcp endpoint in API format. */ 4210 void tcp_get_info(struct sock *sk, struct tcp_info *info) 4211 { 4212 const struct tcp_sock *tp = tcp_sk(sk); /* iff sk_type == SOCK_STREAM */ 4213 const struct inet_connection_sock *icsk = inet_csk(sk); 4214 const u8 ect1_idx = INET_ECN_ECT_1 - 1; 4215 const u8 ect0_idx = INET_ECN_ECT_0 - 1; 4216 const u8 ce_idx = INET_ECN_CE - 1; 4217 unsigned long rate; 4218 u32 now; 4219 u64 rate64; 4220 bool slow; 4221 4222 memset(info, 0, sizeof(*info)); 4223 if (sk->sk_type != SOCK_STREAM) 4224 return; 4225 4226 info->tcpi_state = inet_sk_state_load(sk); 4227 4228 /* Report meaningful fields for all TCP states, including listeners */ 4229 rate = READ_ONCE(sk->sk_pacing_rate); 4230 rate64 = (rate != ~0UL) ? rate : ~0ULL; 4231 info->tcpi_pacing_rate = rate64; 4232 4233 rate = READ_ONCE(sk->sk_max_pacing_rate); 4234 rate64 = (rate != ~0UL) ? rate : ~0ULL; 4235 info->tcpi_max_pacing_rate = rate64; 4236 4237 info->tcpi_reordering = tp->reordering; 4238 info->tcpi_snd_cwnd = tcp_snd_cwnd(tp); 4239 4240 if (info->tcpi_state == TCP_LISTEN) { 4241 /* listeners aliased fields : 4242 * tcpi_unacked -> Number of children ready for accept() 4243 * tcpi_sacked -> max backlog 4244 */ 4245 info->tcpi_unacked = READ_ONCE(sk->sk_ack_backlog); 4246 info->tcpi_sacked = READ_ONCE(sk->sk_max_ack_backlog); 4247 return; 4248 } 4249 4250 slow = lock_sock_fast(sk); 4251 4252 info->tcpi_ca_state = icsk->icsk_ca_state; 4253 info->tcpi_retransmits = icsk->icsk_retransmits; 4254 info->tcpi_probes = icsk->icsk_probes_out; 4255 info->tcpi_backoff = icsk->icsk_backoff; 4256 4257 if (tp->rx_opt.tstamp_ok) 4258 info->tcpi_options |= TCPI_OPT_TIMESTAMPS; 4259 if (tcp_is_sack(tp)) 4260 info->tcpi_options |= TCPI_OPT_SACK; 4261 if (tp->rx_opt.wscale_ok) { 4262 info->tcpi_options |= TCPI_OPT_WSCALE; 4263 info->tcpi_snd_wscale = tp->rx_opt.snd_wscale; 4264 info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale; 4265 } 4266 4267 if (tcp_ecn_mode_any(tp)) 4268 info->tcpi_options |= TCPI_OPT_ECN; 4269 if (tp->ecn_flags & TCP_ECN_SEEN) 4270 info->tcpi_options |= TCPI_OPT_ECN_SEEN; 4271 if (tp->syn_data_acked) 4272 info->tcpi_options |= TCPI_OPT_SYN_DATA; 4273 if (tp->tcp_usec_ts) 4274 info->tcpi_options |= TCPI_OPT_USEC_TS; 4275 if (tp->syn_fastopen_child) 4276 info->tcpi_options |= TCPI_OPT_TFO_CHILD; 4277 4278 info->tcpi_rto = jiffies_to_usecs(icsk->icsk_rto); 4279 info->tcpi_ato = jiffies_to_usecs(min_t(u32, icsk->icsk_ack.ato, 4280 tcp_delack_max(sk))); 4281 info->tcpi_snd_mss = tp->mss_cache; 4282 info->tcpi_rcv_mss = icsk->icsk_ack.rcv_mss; 4283 4284 info->tcpi_unacked = tp->packets_out; 4285 info->tcpi_sacked = tp->sacked_out; 4286 4287 info->tcpi_lost = tp->lost_out; 4288 info->tcpi_retrans = tp->retrans_out; 4289 4290 now = tcp_jiffies32; 4291 info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime); 4292 info->tcpi_last_data_recv = jiffies_to_msecs(now - icsk->icsk_ack.lrcvtime); 4293 info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp); 4294 4295 info->tcpi_pmtu = icsk->icsk_pmtu_cookie; 4296 info->tcpi_rcv_ssthresh = tp->rcv_ssthresh; 4297 info->tcpi_rtt = tp->srtt_us >> 3; 4298 info->tcpi_rttvar = tp->mdev_us >> 2; 4299 info->tcpi_snd_ssthresh = tp->snd_ssthresh; 4300 info->tcpi_advmss = tp->advmss; 4301 4302 info->tcpi_rcv_rtt = tp->rcv_rtt_est.rtt_us >> 3; 4303 info->tcpi_rcv_space = tp->rcvq_space.space; 4304 4305 info->tcpi_total_retrans = tp->total_retrans; 4306 4307 info->tcpi_bytes_acked = tp->bytes_acked; 4308 info->tcpi_bytes_received = tp->bytes_received; 4309 info->tcpi_notsent_bytes = max_t(int, 0, tp->write_seq - tp->snd_nxt); 4310 tcp_get_info_chrono_stats(tp, info); 4311 4312 info->tcpi_segs_out = tp->segs_out; 4313 4314 /* segs_in and data_segs_in can be updated from tcp_segs_in() from BH */ 4315 info->tcpi_segs_in = READ_ONCE(tp->segs_in); 4316 info->tcpi_data_segs_in = READ_ONCE(tp->data_segs_in); 4317 4318 info->tcpi_min_rtt = tcp_min_rtt(tp); 4319 info->tcpi_data_segs_out = tp->data_segs_out; 4320 4321 info->tcpi_delivery_rate_app_limited = tp->rate_app_limited ? 1 : 0; 4322 rate64 = tcp_compute_delivery_rate(tp); 4323 if (rate64) 4324 info->tcpi_delivery_rate = rate64; 4325 info->tcpi_delivered = tp->delivered; 4326 info->tcpi_delivered_ce = tp->delivered_ce; 4327 info->tcpi_bytes_sent = tp->bytes_sent; 4328 info->tcpi_bytes_retrans = tp->bytes_retrans; 4329 info->tcpi_dsack_dups = tp->dsack_dups; 4330 info->tcpi_reord_seen = tp->reord_seen; 4331 info->tcpi_rcv_ooopack = tp->rcv_ooopack; 4332 info->tcpi_snd_wnd = tp->snd_wnd; 4333 info->tcpi_rcv_wnd = tp->rcv_wnd; 4334 info->tcpi_rehash = tp->plb_rehash + tp->timeout_rehash; 4335 info->tcpi_fastopen_client_fail = tp->fastopen_client_fail; 4336 4337 info->tcpi_total_rto = tp->total_rto; 4338 info->tcpi_total_rto_recoveries = tp->total_rto_recoveries; 4339 info->tcpi_total_rto_time = tp->total_rto_time; 4340 if (tp->rto_stamp) 4341 info->tcpi_total_rto_time += tcp_clock_ms() - tp->rto_stamp; 4342 4343 if (tcp_ecn_disabled(tp)) 4344 info->tcpi_ecn_mode = TCPI_ECN_MODE_DISABLED; 4345 else if (tcp_ecn_mode_rfc3168(tp)) 4346 info->tcpi_ecn_mode = TCPI_ECN_MODE_RFC3168; 4347 else if (tcp_ecn_mode_accecn(tp)) 4348 info->tcpi_ecn_mode = TCPI_ECN_MODE_ACCECN; 4349 else if (tcp_ecn_mode_pending(tp)) 4350 info->tcpi_ecn_mode = TCPI_ECN_MODE_PENDING; 4351 info->tcpi_accecn_fail_mode = tp->accecn_fail_mode; 4352 info->tcpi_accecn_opt_seen = tp->saw_accecn_opt; 4353 info->tcpi_received_ce = tp->received_ce; 4354 info->tcpi_delivered_e1_bytes = tp->delivered_ecn_bytes[ect1_idx]; 4355 info->tcpi_delivered_e0_bytes = tp->delivered_ecn_bytes[ect0_idx]; 4356 info->tcpi_delivered_ce_bytes = tp->delivered_ecn_bytes[ce_idx]; 4357 info->tcpi_received_e1_bytes = tp->received_ecn_bytes[ect1_idx]; 4358 info->tcpi_received_e0_bytes = tp->received_ecn_bytes[ect0_idx]; 4359 info->tcpi_received_ce_bytes = tp->received_ecn_bytes[ce_idx]; 4360 4361 unlock_sock_fast(sk, slow); 4362 } 4363 EXPORT_SYMBOL_GPL(tcp_get_info); 4364 4365 static size_t tcp_opt_stats_get_size(void) 4366 { 4367 return 4368 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_BUSY */ 4369 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_RWND_LIMITED */ 4370 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_SNDBUF_LIMITED */ 4371 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_DATA_SEGS_OUT */ 4372 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_TOTAL_RETRANS */ 4373 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_PACING_RATE */ 4374 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_DELIVERY_RATE */ 4375 nla_total_size(sizeof(u32)) + /* TCP_NLA_SND_CWND */ 4376 nla_total_size(sizeof(u32)) + /* TCP_NLA_REORDERING */ 4377 nla_total_size(sizeof(u32)) + /* TCP_NLA_MIN_RTT */ 4378 nla_total_size(sizeof(u8)) + /* TCP_NLA_RECUR_RETRANS */ 4379 nla_total_size(sizeof(u8)) + /* TCP_NLA_DELIVERY_RATE_APP_LMT */ 4380 nla_total_size(sizeof(u32)) + /* TCP_NLA_SNDQ_SIZE */ 4381 nla_total_size(sizeof(u8)) + /* TCP_NLA_CA_STATE */ 4382 nla_total_size(sizeof(u32)) + /* TCP_NLA_SND_SSTHRESH */ 4383 nla_total_size(sizeof(u32)) + /* TCP_NLA_DELIVERED */ 4384 nla_total_size(sizeof(u32)) + /* TCP_NLA_DELIVERED_CE */ 4385 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_BYTES_SENT */ 4386 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_BYTES_RETRANS */ 4387 nla_total_size(sizeof(u32)) + /* TCP_NLA_DSACK_DUPS */ 4388 nla_total_size(sizeof(u32)) + /* TCP_NLA_REORD_SEEN */ 4389 nla_total_size(sizeof(u32)) + /* TCP_NLA_SRTT */ 4390 nla_total_size(sizeof(u16)) + /* TCP_NLA_TIMEOUT_REHASH */ 4391 nla_total_size(sizeof(u32)) + /* TCP_NLA_BYTES_NOTSENT */ 4392 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_EDT */ 4393 nla_total_size(sizeof(u8)) + /* TCP_NLA_TTL */ 4394 nla_total_size(sizeof(u32)) + /* TCP_NLA_REHASH */ 4395 0; 4396 } 4397 4398 /* Returns TTL or hop limit of an incoming packet from skb. */ 4399 static u8 tcp_skb_ttl_or_hop_limit(const struct sk_buff *skb) 4400 { 4401 if (skb->protocol == htons(ETH_P_IP)) 4402 return ip_hdr(skb)->ttl; 4403 else if (skb->protocol == htons(ETH_P_IPV6)) 4404 return ipv6_hdr(skb)->hop_limit; 4405 else 4406 return 0; 4407 } 4408 4409 struct sk_buff *tcp_get_timestamping_opt_stats(const struct sock *sk, 4410 const struct sk_buff *orig_skb, 4411 const struct sk_buff *ack_skb) 4412 { 4413 const struct tcp_sock *tp = tcp_sk(sk); 4414 struct sk_buff *stats; 4415 struct tcp_info info; 4416 unsigned long rate; 4417 u64 rate64; 4418 4419 stats = alloc_skb(tcp_opt_stats_get_size(), GFP_ATOMIC); 4420 if (!stats) 4421 return NULL; 4422 4423 tcp_get_info_chrono_stats(tp, &info); 4424 nla_put_u64_64bit(stats, TCP_NLA_BUSY, 4425 info.tcpi_busy_time, TCP_NLA_PAD); 4426 nla_put_u64_64bit(stats, TCP_NLA_RWND_LIMITED, 4427 info.tcpi_rwnd_limited, TCP_NLA_PAD); 4428 nla_put_u64_64bit(stats, TCP_NLA_SNDBUF_LIMITED, 4429 info.tcpi_sndbuf_limited, TCP_NLA_PAD); 4430 nla_put_u64_64bit(stats, TCP_NLA_DATA_SEGS_OUT, 4431 READ_ONCE(tp->data_segs_out), TCP_NLA_PAD); 4432 nla_put_u64_64bit(stats, TCP_NLA_TOTAL_RETRANS, 4433 READ_ONCE(tp->total_retrans), TCP_NLA_PAD); 4434 4435 rate = READ_ONCE(sk->sk_pacing_rate); 4436 rate64 = (rate != ~0UL) ? rate : ~0ULL; 4437 nla_put_u64_64bit(stats, TCP_NLA_PACING_RATE, rate64, TCP_NLA_PAD); 4438 4439 rate64 = tcp_compute_delivery_rate(tp); 4440 nla_put_u64_64bit(stats, TCP_NLA_DELIVERY_RATE, rate64, TCP_NLA_PAD); 4441 4442 nla_put_u32(stats, TCP_NLA_SND_CWND, READ_ONCE(tp->snd_cwnd)); 4443 nla_put_u32(stats, TCP_NLA_REORDERING, READ_ONCE(tp->reordering)); 4444 nla_put_u32(stats, TCP_NLA_MIN_RTT, data_race(tcp_min_rtt(tp))); 4445 4446 nla_put_u8(stats, TCP_NLA_RECUR_RETRANS, 4447 READ_ONCE(inet_csk(sk)->icsk_retransmits)); 4448 nla_put_u8(stats, TCP_NLA_DELIVERY_RATE_APP_LMT, data_race(!!tp->rate_app_limited)); 4449 nla_put_u32(stats, TCP_NLA_SND_SSTHRESH, READ_ONCE(tp->snd_ssthresh)); 4450 nla_put_u32(stats, TCP_NLA_DELIVERED, READ_ONCE(tp->delivered)); 4451 nla_put_u32(stats, TCP_NLA_DELIVERED_CE, READ_ONCE(tp->delivered_ce)); 4452 4453 nla_put_u32(stats, TCP_NLA_SNDQ_SIZE, 4454 max_t(int, 0, 4455 READ_ONCE(tp->write_seq) - READ_ONCE(tp->snd_una))); 4456 nla_put_u8(stats, TCP_NLA_CA_STATE, inet_csk(sk)->icsk_ca_state); 4457 4458 nla_put_u64_64bit(stats, TCP_NLA_BYTES_SENT, READ_ONCE(tp->bytes_sent), 4459 TCP_NLA_PAD); 4460 nla_put_u64_64bit(stats, TCP_NLA_BYTES_RETRANS, 4461 READ_ONCE(tp->bytes_retrans), TCP_NLA_PAD); 4462 nla_put_u32(stats, TCP_NLA_DSACK_DUPS, READ_ONCE(tp->dsack_dups)); 4463 nla_put_u32(stats, TCP_NLA_REORD_SEEN, READ_ONCE(tp->reord_seen)); 4464 nla_put_u32(stats, TCP_NLA_SRTT, READ_ONCE(tp->srtt_us) >> 3); 4465 nla_put_u16(stats, TCP_NLA_TIMEOUT_REHASH, 4466 READ_ONCE(tp->timeout_rehash)); 4467 nla_put_u32(stats, TCP_NLA_BYTES_NOTSENT, 4468 max_t(int, 0, 4469 READ_ONCE(tp->write_seq) - READ_ONCE(tp->snd_nxt))); 4470 nla_put_u64_64bit(stats, TCP_NLA_EDT, orig_skb->skb_mstamp_ns, 4471 TCP_NLA_PAD); 4472 if (ack_skb) 4473 nla_put_u8(stats, TCP_NLA_TTL, 4474 tcp_skb_ttl_or_hop_limit(ack_skb)); 4475 4476 nla_put_u32(stats, TCP_NLA_REHASH, 4477 READ_ONCE(tp->plb_rehash) + READ_ONCE(tp->timeout_rehash)); 4478 return stats; 4479 } 4480 4481 int do_tcp_getsockopt(struct sock *sk, int level, 4482 int optname, sockptr_t optval, sockptr_t optlen) 4483 { 4484 struct inet_connection_sock *icsk = inet_csk(sk); 4485 struct tcp_sock *tp = tcp_sk(sk); 4486 struct net *net = sock_net(sk); 4487 int user_mss; 4488 int val, len; 4489 4490 if (copy_from_sockptr(&len, optlen, sizeof(int))) 4491 return -EFAULT; 4492 4493 if (len < 0) 4494 return -EINVAL; 4495 4496 len = min_t(unsigned int, len, sizeof(int)); 4497 4498 switch (optname) { 4499 case TCP_MAXSEG: 4500 val = tp->mss_cache; 4501 user_mss = READ_ONCE(tp->rx_opt.user_mss); 4502 if (user_mss && 4503 ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN))) 4504 val = user_mss; 4505 if (tp->repair) 4506 val = tp->rx_opt.mss_clamp; 4507 break; 4508 case TCP_NODELAY: 4509 val = !!(tp->nonagle&TCP_NAGLE_OFF); 4510 break; 4511 case TCP_CORK: 4512 val = !!(tp->nonagle&TCP_NAGLE_CORK); 4513 break; 4514 case TCP_KEEPIDLE: 4515 val = keepalive_time_when(tp) / HZ; 4516 break; 4517 case TCP_KEEPINTVL: 4518 val = keepalive_intvl_when(tp) / HZ; 4519 break; 4520 case TCP_KEEPCNT: 4521 val = keepalive_probes(tp); 4522 break; 4523 case TCP_SYNCNT: 4524 val = READ_ONCE(icsk->icsk_syn_retries) ? : 4525 READ_ONCE(net->ipv4.sysctl_tcp_syn_retries); 4526 break; 4527 case TCP_LINGER2: 4528 val = READ_ONCE(tp->linger2); 4529 if (val >= 0) 4530 val = (val ? : READ_ONCE(net->ipv4.sysctl_tcp_fin_timeout)) / HZ; 4531 break; 4532 case TCP_DEFER_ACCEPT: 4533 val = READ_ONCE(icsk->icsk_accept_queue.rskq_defer_accept); 4534 val = retrans_to_secs(val, TCP_TIMEOUT_INIT / HZ, 4535 TCP_RTO_MAX / HZ); 4536 break; 4537 case TCP_WINDOW_CLAMP: 4538 val = READ_ONCE(tp->window_clamp); 4539 break; 4540 case TCP_INFO: { 4541 struct tcp_info info; 4542 4543 if (copy_from_sockptr(&len, optlen, sizeof(int))) 4544 return -EFAULT; 4545 4546 tcp_get_info(sk, &info); 4547 4548 len = min_t(unsigned int, len, sizeof(info)); 4549 if (copy_to_sockptr(optlen, &len, sizeof(int))) 4550 return -EFAULT; 4551 if (copy_to_sockptr(optval, &info, len)) 4552 return -EFAULT; 4553 return 0; 4554 } 4555 case TCP_CC_INFO: { 4556 const struct tcp_congestion_ops *ca_ops; 4557 union tcp_cc_info info; 4558 size_t sz = 0; 4559 int attr; 4560 4561 if (copy_from_sockptr(&len, optlen, sizeof(int))) 4562 return -EFAULT; 4563 4564 ca_ops = icsk->icsk_ca_ops; 4565 if (ca_ops && ca_ops->get_info) 4566 sz = ca_ops->get_info(sk, ~0U, &attr, &info); 4567 4568 len = min_t(unsigned int, len, sz); 4569 if (copy_to_sockptr(optlen, &len, sizeof(int))) 4570 return -EFAULT; 4571 if (copy_to_sockptr(optval, &info, len)) 4572 return -EFAULT; 4573 return 0; 4574 } 4575 case TCP_QUICKACK: 4576 val = !inet_csk_in_pingpong_mode(sk); 4577 break; 4578 4579 case TCP_CONGESTION: 4580 if (copy_from_sockptr(&len, optlen, sizeof(int))) 4581 return -EFAULT; 4582 len = min_t(unsigned int, len, TCP_CA_NAME_MAX); 4583 if (copy_to_sockptr(optlen, &len, sizeof(int))) 4584 return -EFAULT; 4585 if (copy_to_sockptr(optval, icsk->icsk_ca_ops->name, len)) 4586 return -EFAULT; 4587 return 0; 4588 4589 case TCP_ULP: 4590 if (copy_from_sockptr(&len, optlen, sizeof(int))) 4591 return -EFAULT; 4592 len = min_t(unsigned int, len, TCP_ULP_NAME_MAX); 4593 if (!icsk->icsk_ulp_ops) { 4594 len = 0; 4595 if (copy_to_sockptr(optlen, &len, sizeof(int))) 4596 return -EFAULT; 4597 return 0; 4598 } 4599 if (copy_to_sockptr(optlen, &len, sizeof(int))) 4600 return -EFAULT; 4601 if (copy_to_sockptr(optval, icsk->icsk_ulp_ops->name, len)) 4602 return -EFAULT; 4603 return 0; 4604 4605 case TCP_FASTOPEN_KEY: { 4606 u64 key[TCP_FASTOPEN_KEY_BUF_LENGTH / sizeof(u64)]; 4607 unsigned int key_len; 4608 4609 if (copy_from_sockptr(&len, optlen, sizeof(int))) 4610 return -EFAULT; 4611 4612 key_len = tcp_fastopen_get_cipher(net, icsk, key) * 4613 TCP_FASTOPEN_KEY_LENGTH; 4614 len = min_t(unsigned int, len, key_len); 4615 if (copy_to_sockptr(optlen, &len, sizeof(int))) 4616 return -EFAULT; 4617 if (copy_to_sockptr(optval, key, len)) 4618 return -EFAULT; 4619 return 0; 4620 } 4621 case TCP_THIN_LINEAR_TIMEOUTS: 4622 val = tp->thin_lto; 4623 break; 4624 4625 case TCP_THIN_DUPACK: 4626 val = 0; 4627 break; 4628 4629 case TCP_REPAIR: 4630 val = tp->repair; 4631 break; 4632 4633 case TCP_REPAIR_QUEUE: 4634 if (tp->repair) 4635 val = tp->repair_queue; 4636 else 4637 return -EINVAL; 4638 break; 4639 4640 case TCP_REPAIR_WINDOW: { 4641 struct tcp_repair_window opt; 4642 4643 if (copy_from_sockptr(&len, optlen, sizeof(int))) 4644 return -EFAULT; 4645 4646 if (len != sizeof(opt)) 4647 return -EINVAL; 4648 4649 if (!tp->repair) 4650 return -EPERM; 4651 4652 opt.snd_wl1 = tp->snd_wl1; 4653 opt.snd_wnd = tp->snd_wnd; 4654 opt.max_window = tp->max_window; 4655 opt.rcv_wnd = tp->rcv_wnd; 4656 opt.rcv_wup = tp->rcv_wup; 4657 4658 if (copy_to_sockptr(optval, &opt, len)) 4659 return -EFAULT; 4660 return 0; 4661 } 4662 case TCP_QUEUE_SEQ: 4663 if (tp->repair_queue == TCP_SEND_QUEUE) 4664 val = tp->write_seq; 4665 else if (tp->repair_queue == TCP_RECV_QUEUE) 4666 val = tp->rcv_nxt; 4667 else 4668 return -EINVAL; 4669 break; 4670 4671 case TCP_USER_TIMEOUT: 4672 val = READ_ONCE(icsk->icsk_user_timeout); 4673 break; 4674 4675 case TCP_FASTOPEN: 4676 val = READ_ONCE(icsk->icsk_accept_queue.fastopenq.max_qlen); 4677 break; 4678 4679 case TCP_FASTOPEN_CONNECT: 4680 val = tp->fastopen_connect; 4681 break; 4682 4683 case TCP_FASTOPEN_NO_COOKIE: 4684 val = tp->fastopen_no_cookie; 4685 break; 4686 4687 case TCP_TX_DELAY: 4688 val = READ_ONCE(tp->tcp_tx_delay); 4689 break; 4690 4691 case TCP_TIMESTAMP: 4692 val = tcp_clock_ts(tp->tcp_usec_ts) + READ_ONCE(tp->tsoffset); 4693 if (tp->tcp_usec_ts) 4694 val |= 1; 4695 else 4696 val &= ~1; 4697 break; 4698 case TCP_NOTSENT_LOWAT: 4699 val = READ_ONCE(tp->notsent_lowat); 4700 break; 4701 case TCP_INQ: 4702 val = tp->recvmsg_inq; 4703 break; 4704 case TCP_SAVE_SYN: 4705 val = tp->save_syn; 4706 break; 4707 case TCP_SAVED_SYN: { 4708 if (copy_from_sockptr(&len, optlen, sizeof(int))) 4709 return -EFAULT; 4710 4711 sockopt_lock_sock(sk); 4712 if (tp->saved_syn) { 4713 if (len < tcp_saved_syn_len(tp->saved_syn)) { 4714 len = tcp_saved_syn_len(tp->saved_syn); 4715 if (copy_to_sockptr(optlen, &len, sizeof(int))) { 4716 sockopt_release_sock(sk); 4717 return -EFAULT; 4718 } 4719 sockopt_release_sock(sk); 4720 return -EINVAL; 4721 } 4722 len = tcp_saved_syn_len(tp->saved_syn); 4723 if (copy_to_sockptr(optlen, &len, sizeof(int))) { 4724 sockopt_release_sock(sk); 4725 return -EFAULT; 4726 } 4727 if (copy_to_sockptr(optval, tp->saved_syn->data, len)) { 4728 sockopt_release_sock(sk); 4729 return -EFAULT; 4730 } 4731 tcp_saved_syn_free(tp); 4732 sockopt_release_sock(sk); 4733 } else { 4734 sockopt_release_sock(sk); 4735 len = 0; 4736 if (copy_to_sockptr(optlen, &len, sizeof(int))) 4737 return -EFAULT; 4738 } 4739 return 0; 4740 } 4741 #ifdef CONFIG_MMU 4742 case TCP_ZEROCOPY_RECEIVE: { 4743 struct scm_timestamping_internal tss; 4744 struct tcp_zerocopy_receive zc = {}; 4745 int err; 4746 4747 if (copy_from_sockptr(&len, optlen, sizeof(int))) 4748 return -EFAULT; 4749 if (len < 0 || 4750 len < offsetofend(struct tcp_zerocopy_receive, length)) 4751 return -EINVAL; 4752 if (unlikely(len > sizeof(zc))) { 4753 err = check_zeroed_sockptr(optval, sizeof(zc), 4754 len - sizeof(zc)); 4755 if (err < 1) 4756 return err == 0 ? -EINVAL : err; 4757 len = sizeof(zc); 4758 if (copy_to_sockptr(optlen, &len, sizeof(int))) 4759 return -EFAULT; 4760 } 4761 if (copy_from_sockptr(&zc, optval, len)) 4762 return -EFAULT; 4763 if (zc.reserved) 4764 return -EINVAL; 4765 if (zc.msg_flags & ~(TCP_VALID_ZC_MSG_FLAGS)) 4766 return -EINVAL; 4767 sockopt_lock_sock(sk); 4768 err = tcp_zerocopy_receive(sk, &zc, &tss); 4769 err = BPF_CGROUP_RUN_PROG_GETSOCKOPT_KERN(sk, level, optname, 4770 &zc, &len, err); 4771 sockopt_release_sock(sk); 4772 if (len >= offsetofend(struct tcp_zerocopy_receive, msg_flags)) 4773 goto zerocopy_rcv_cmsg; 4774 switch (len) { 4775 case offsetofend(struct tcp_zerocopy_receive, msg_flags): 4776 goto zerocopy_rcv_cmsg; 4777 case offsetofend(struct tcp_zerocopy_receive, msg_controllen): 4778 case offsetofend(struct tcp_zerocopy_receive, msg_control): 4779 case offsetofend(struct tcp_zerocopy_receive, flags): 4780 case offsetofend(struct tcp_zerocopy_receive, copybuf_len): 4781 case offsetofend(struct tcp_zerocopy_receive, copybuf_address): 4782 case offsetofend(struct tcp_zerocopy_receive, err): 4783 goto zerocopy_rcv_sk_err; 4784 case offsetofend(struct tcp_zerocopy_receive, inq): 4785 goto zerocopy_rcv_inq; 4786 case offsetofend(struct tcp_zerocopy_receive, length): 4787 default: 4788 goto zerocopy_rcv_out; 4789 } 4790 zerocopy_rcv_cmsg: 4791 if (zc.msg_flags & TCP_CMSG_TS) 4792 tcp_zc_finalize_rx_tstamp(sk, &zc, &tss); 4793 else 4794 zc.msg_flags = 0; 4795 zerocopy_rcv_sk_err: 4796 if (!err) 4797 zc.err = sock_error(sk); 4798 zerocopy_rcv_inq: 4799 zc.inq = tcp_inq_hint(sk); 4800 zerocopy_rcv_out: 4801 if (!err && copy_to_sockptr(optval, &zc, len)) 4802 err = -EFAULT; 4803 return err; 4804 } 4805 #endif 4806 case TCP_AO_REPAIR: 4807 if (!tcp_can_repair_sock(sk)) 4808 return -EPERM; 4809 return tcp_ao_get_repair(sk, optval, optlen); 4810 case TCP_AO_GET_KEYS: 4811 case TCP_AO_INFO: { 4812 int err; 4813 4814 sockopt_lock_sock(sk); 4815 if (optname == TCP_AO_GET_KEYS) 4816 err = tcp_ao_get_mkts(sk, optval, optlen); 4817 else 4818 err = tcp_ao_get_sock_info(sk, optval, optlen); 4819 sockopt_release_sock(sk); 4820 4821 return err; 4822 } 4823 case TCP_IS_MPTCP: 4824 val = 0; 4825 break; 4826 case TCP_RTO_MAX_MS: 4827 val = jiffies_to_msecs(tcp_rto_max(sk)); 4828 break; 4829 case TCP_RTO_MIN_US: 4830 val = jiffies_to_usecs(READ_ONCE(inet_csk(sk)->icsk_rto_min)); 4831 break; 4832 case TCP_DELACK_MAX_US: 4833 val = jiffies_to_usecs(READ_ONCE(inet_csk(sk)->icsk_delack_max)); 4834 break; 4835 default: 4836 return -ENOPROTOOPT; 4837 } 4838 4839 if (copy_to_sockptr(optlen, &len, sizeof(int))) 4840 return -EFAULT; 4841 if (copy_to_sockptr(optval, &val, len)) 4842 return -EFAULT; 4843 return 0; 4844 } 4845 4846 bool tcp_bpf_bypass_getsockopt(int level, int optname) 4847 { 4848 /* TCP do_tcp_getsockopt has optimized getsockopt implementation 4849 * to avoid extra socket lock for TCP_ZEROCOPY_RECEIVE. 4850 */ 4851 if (level == SOL_TCP && optname == TCP_ZEROCOPY_RECEIVE) 4852 return true; 4853 4854 return false; 4855 } 4856 4857 int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval, 4858 int __user *optlen) 4859 { 4860 struct inet_connection_sock *icsk = inet_csk(sk); 4861 4862 if (level != SOL_TCP) 4863 /* Paired with WRITE_ONCE() in do_ipv6_setsockopt() and tcp_v6_connect() */ 4864 return READ_ONCE(icsk->icsk_af_ops)->getsockopt(sk, level, optname, 4865 optval, optlen); 4866 return do_tcp_getsockopt(sk, level, optname, USER_SOCKPTR(optval), 4867 USER_SOCKPTR(optlen)); 4868 } 4869 4870 #ifdef CONFIG_TCP_MD5SIG 4871 void tcp_md5_hash_skb_data(struct md5_ctx *ctx, const struct sk_buff *skb, 4872 unsigned int header_len) 4873 { 4874 const unsigned int head_data_len = skb_headlen(skb) > header_len ? 4875 skb_headlen(skb) - header_len : 0; 4876 const struct skb_shared_info *shi = skb_shinfo(skb); 4877 struct sk_buff *frag_iter; 4878 unsigned int i; 4879 4880 md5_update(ctx, (const u8 *)tcp_hdr(skb) + header_len, head_data_len); 4881 4882 for (i = 0; i < shi->nr_frags; ++i) { 4883 const skb_frag_t *f = &shi->frags[i]; 4884 u32 p_off, p_len, copied; 4885 const void *vaddr; 4886 struct page *p; 4887 4888 skb_frag_foreach_page(f, skb_frag_off(f), skb_frag_size(f), 4889 p, p_off, p_len, copied) { 4890 vaddr = kmap_local_page(p); 4891 md5_update(ctx, vaddr + p_off, p_len); 4892 kunmap_local(vaddr); 4893 } 4894 } 4895 4896 skb_walk_frags(skb, frag_iter) 4897 tcp_md5_hash_skb_data(ctx, frag_iter, 0); 4898 } 4899 4900 void tcp_md5_hash_key(struct md5_ctx *ctx, 4901 const struct tcp_md5sig_key *key) 4902 { 4903 u8 keylen = READ_ONCE(key->keylen); /* paired with WRITE_ONCE() in tcp_md5_do_add */ 4904 4905 /* We use data_race() because tcp_md5_do_add() might change 4906 * key->key under us 4907 */ 4908 data_race(({ md5_update(ctx, key->key, keylen), 0; })); 4909 } 4910 4911 /* Called with rcu_read_lock() */ 4912 static enum skb_drop_reason 4913 tcp_inbound_md5_hash(const struct sock *sk, const struct sk_buff *skb, 4914 const void *saddr, const void *daddr, 4915 int family, int l3index, const __u8 *hash_location) 4916 { 4917 /* This gets called for each TCP segment that has TCP-MD5 option. 4918 * We have 2 drop cases: 4919 * o An MD5 signature is present, but we're not expecting one. 4920 * o The MD5 signature is wrong. 4921 */ 4922 const struct tcp_sock *tp = tcp_sk(sk); 4923 struct tcp_md5sig_key *key; 4924 u8 newhash[16]; 4925 4926 key = tcp_md5_do_lookup(sk, l3index, saddr, family); 4927 if (!key) { 4928 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMD5UNEXPECTED); 4929 trace_tcp_hash_md5_unexpected(sk, skb); 4930 return SKB_DROP_REASON_TCP_MD5UNEXPECTED; 4931 } 4932 4933 /* Check the signature. 4934 * To support dual stack listeners, we need to handle 4935 * IPv4-mapped case. 4936 */ 4937 if (family == AF_INET) 4938 tcp_v4_md5_hash_skb(newhash, key, NULL, skb); 4939 else 4940 tp->af_specific->calc_md5_hash(newhash, key, NULL, skb); 4941 if (crypto_memneq(hash_location, newhash, 16)) { 4942 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMD5FAILURE); 4943 trace_tcp_hash_md5_mismatch(sk, skb); 4944 return SKB_DROP_REASON_TCP_MD5FAILURE; 4945 } 4946 return SKB_NOT_DROPPED_YET; 4947 } 4948 #else 4949 static inline enum skb_drop_reason 4950 tcp_inbound_md5_hash(const struct sock *sk, const struct sk_buff *skb, 4951 const void *saddr, const void *daddr, 4952 int family, int l3index, const __u8 *hash_location) 4953 { 4954 return SKB_NOT_DROPPED_YET; 4955 } 4956 4957 #endif 4958 4959 #if defined(CONFIG_TCP_MD5SIG) || defined(CONFIG_TCP_AO) 4960 /* 4961 * Parse Signature options 4962 */ 4963 int tcp_do_parse_auth_options(const struct tcphdr *th, 4964 const u8 **md5_hash, const u8 **ao_hash) 4965 { 4966 int length = (th->doff << 2) - sizeof(*th); 4967 const u8 *ptr = (const u8 *)(th + 1); 4968 unsigned int minlen = TCPOLEN_MD5SIG; 4969 4970 if (IS_ENABLED(CONFIG_TCP_AO)) 4971 minlen = sizeof(struct tcp_ao_hdr) + 1; 4972 4973 *md5_hash = NULL; 4974 *ao_hash = NULL; 4975 4976 /* If not enough data remaining, we can short cut */ 4977 while (length >= minlen) { 4978 int opcode = *ptr++; 4979 int opsize; 4980 4981 switch (opcode) { 4982 case TCPOPT_EOL: 4983 return 0; 4984 case TCPOPT_NOP: 4985 length--; 4986 continue; 4987 default: 4988 opsize = *ptr++; 4989 if (opsize < 2 || opsize > length) 4990 return -EINVAL; 4991 if (opcode == TCPOPT_MD5SIG) { 4992 if (opsize != TCPOLEN_MD5SIG) 4993 return -EINVAL; 4994 if (unlikely(*md5_hash || *ao_hash)) 4995 return -EEXIST; 4996 *md5_hash = ptr; 4997 } else if (opcode == TCPOPT_AO) { 4998 if (opsize <= sizeof(struct tcp_ao_hdr)) 4999 return -EINVAL; 5000 if (unlikely(*md5_hash || *ao_hash)) 5001 return -EEXIST; 5002 *ao_hash = ptr; 5003 } 5004 } 5005 ptr += opsize - 2; 5006 length -= opsize; 5007 } 5008 return 0; 5009 } 5010 #endif 5011 5012 /* Called with rcu_read_lock() */ 5013 enum skb_drop_reason 5014 tcp_inbound_hash(struct sock *sk, const struct request_sock *req, 5015 const struct sk_buff *skb, 5016 const void *saddr, const void *daddr, 5017 int family, int dif, int sdif) 5018 { 5019 const struct tcphdr *th = tcp_hdr(skb); 5020 const struct tcp_ao_hdr *aoh; 5021 const __u8 *md5_location; 5022 int l3index; 5023 5024 /* Invalid option or two times meet any of auth options */ 5025 if (tcp_parse_auth_options(th, &md5_location, &aoh)) { 5026 trace_tcp_hash_bad_header(sk, skb); 5027 return SKB_DROP_REASON_TCP_AUTH_HDR; 5028 } 5029 5030 if (req) { 5031 if (tcp_rsk_used_ao(req) != !!aoh) { 5032 u8 keyid, rnext, maclen; 5033 5034 if (aoh) { 5035 keyid = aoh->keyid; 5036 rnext = aoh->rnext_keyid; 5037 maclen = tcp_ao_hdr_maclen(aoh); 5038 } else { 5039 keyid = rnext = maclen = 0; 5040 } 5041 5042 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPAOBAD); 5043 trace_tcp_ao_handshake_failure(sk, skb, keyid, rnext, maclen); 5044 return SKB_DROP_REASON_TCP_AOFAILURE; 5045 } 5046 } 5047 5048 /* sdif set, means packet ingressed via a device 5049 * in an L3 domain and dif is set to the l3mdev 5050 */ 5051 l3index = sdif ? dif : 0; 5052 5053 /* Fast path: unsigned segments */ 5054 if (likely(!md5_location && !aoh)) { 5055 /* Drop if there's TCP-MD5 or TCP-AO key with any rcvid/sndid 5056 * for the remote peer. On TCP-AO established connection 5057 * the last key is impossible to remove, so there's 5058 * always at least one current_key. 5059 */ 5060 if (tcp_ao_required(sk, saddr, family, l3index, true)) { 5061 trace_tcp_hash_ao_required(sk, skb); 5062 return SKB_DROP_REASON_TCP_AONOTFOUND; 5063 } 5064 if (unlikely(tcp_md5_do_lookup(sk, l3index, saddr, family))) { 5065 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMD5NOTFOUND); 5066 trace_tcp_hash_md5_required(sk, skb); 5067 return SKB_DROP_REASON_TCP_MD5NOTFOUND; 5068 } 5069 return SKB_NOT_DROPPED_YET; 5070 } 5071 5072 if (aoh) 5073 return tcp_inbound_ao_hash(sk, skb, family, req, l3index, aoh); 5074 5075 return tcp_inbound_md5_hash(sk, skb, saddr, daddr, family, 5076 l3index, md5_location); 5077 } 5078 5079 void tcp_done(struct sock *sk) 5080 { 5081 struct request_sock *req; 5082 5083 /* We might be called with a new socket, after 5084 * inet_csk_prepare_forced_close() has been called 5085 * so we can not use lockdep_sock_is_held(sk) 5086 */ 5087 req = rcu_dereference_protected(tcp_sk(sk)->fastopen_rsk, 1); 5088 5089 if (sk->sk_state == TCP_SYN_SENT || sk->sk_state == TCP_SYN_RECV) 5090 TCP_INC_STATS(sock_net(sk), TCP_MIB_ATTEMPTFAILS); 5091 5092 tcp_set_state(sk, TCP_CLOSE); 5093 tcp_clear_xmit_timers(sk); 5094 if (req) 5095 reqsk_fastopen_remove(sk, req, false); 5096 5097 WRITE_ONCE(sk->sk_shutdown, SHUTDOWN_MASK); 5098 5099 if (!sock_flag(sk, SOCK_DEAD)) 5100 sk->sk_state_change(sk); 5101 else 5102 inet_csk_destroy_sock(sk); 5103 } 5104 5105 int tcp_abort(struct sock *sk, int err) 5106 { 5107 int state = inet_sk_state_load(sk); 5108 5109 if (state == TCP_NEW_SYN_RECV) { 5110 struct request_sock *req = inet_reqsk(sk); 5111 5112 local_bh_disable(); 5113 inet_csk_reqsk_queue_drop(req->rsk_listener, req); 5114 local_bh_enable(); 5115 return 0; 5116 } 5117 if (state == TCP_TIME_WAIT) { 5118 struct inet_timewait_sock *tw = inet_twsk(sk); 5119 5120 refcount_inc(&tw->tw_refcnt); 5121 local_bh_disable(); 5122 inet_twsk_deschedule_put(tw); 5123 local_bh_enable(); 5124 return 0; 5125 } 5126 5127 /* BPF context ensures sock locking. */ 5128 if (!has_current_bpf_ctx()) 5129 /* Don't race with userspace socket closes such as tcp_close. */ 5130 lock_sock(sk); 5131 5132 /* Avoid closing the same socket twice. */ 5133 if (sk->sk_state == TCP_CLOSE) { 5134 if (!has_current_bpf_ctx()) 5135 release_sock(sk); 5136 return -ENOENT; 5137 } 5138 5139 if (sk->sk_state == TCP_LISTEN) { 5140 tcp_set_state(sk, TCP_CLOSE); 5141 inet_csk_listen_stop(sk); 5142 } 5143 5144 /* Don't race with BH socket closes such as inet_csk_listen_stop. */ 5145 local_bh_disable(); 5146 bh_lock_sock(sk); 5147 5148 if (tcp_need_reset(sk->sk_state)) 5149 tcp_send_active_reset(sk, GFP_ATOMIC, 5150 SK_RST_REASON_TCP_STATE); 5151 tcp_done_with_error(sk, err); 5152 5153 bh_unlock_sock(sk); 5154 local_bh_enable(); 5155 if (!has_current_bpf_ctx()) 5156 release_sock(sk); 5157 return 0; 5158 } 5159 EXPORT_SYMBOL_GPL(tcp_abort); 5160 5161 extern struct tcp_congestion_ops tcp_reno; 5162 5163 static __initdata unsigned long thash_entries; 5164 static int __init set_thash_entries(char *str) 5165 { 5166 ssize_t ret; 5167 5168 if (!str) 5169 return 0; 5170 5171 ret = kstrtoul(str, 0, &thash_entries); 5172 if (ret) 5173 return 0; 5174 5175 return 1; 5176 } 5177 __setup("thash_entries=", set_thash_entries); 5178 5179 static void __init tcp_init_mem(void) 5180 { 5181 unsigned long limit = nr_free_buffer_pages() / 16; 5182 5183 limit = max(limit, 128UL); 5184 sysctl_tcp_mem[0] = limit / 4 * 3; /* 4.68 % */ 5185 sysctl_tcp_mem[1] = limit; /* 6.25 % */ 5186 sysctl_tcp_mem[2] = sysctl_tcp_mem[0] * 2; /* 9.37 % */ 5187 } 5188 5189 static void __init tcp_struct_check(void) 5190 { 5191 /* TX read-mostly hotpath cache lines */ 5192 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_tx, max_window); 5193 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_tx, rcv_ssthresh); 5194 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_tx, reordering); 5195 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_tx, notsent_lowat); 5196 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_tx, gso_segs); 5197 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_tx, retransmit_skb_hint); 5198 #if IS_ENABLED(CONFIG_TLS_DEVICE) 5199 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_tx, tcp_clean_acked); 5200 #endif 5201 5202 /* TXRX read-mostly hotpath cache lines */ 5203 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_txrx, tsoffset); 5204 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_txrx, snd_wnd); 5205 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_txrx, mss_cache); 5206 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_txrx, snd_cwnd); 5207 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_txrx, prr_out); 5208 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_txrx, lost_out); 5209 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_txrx, sacked_out); 5210 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_txrx, scaling_ratio); 5211 5212 /* RX read-mostly hotpath cache lines */ 5213 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_rx, copied_seq); 5214 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_rx, snd_wl1); 5215 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_rx, tlp_high_seq); 5216 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_rx, rttvar_us); 5217 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_rx, retrans_out); 5218 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_rx, advmss); 5219 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_rx, urg_data); 5220 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_rx, lost); 5221 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_rx, rtt_min); 5222 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_rx, out_of_order_queue); 5223 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_rx, snd_ssthresh); 5224 5225 /* TX read-write hotpath cache lines */ 5226 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, data_segs_out); 5227 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, delivered); 5228 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, delivered_ce); 5229 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, bytes_acked); 5230 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, bytes_sent); 5231 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, first_tx_mstamp); 5232 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, delivered_mstamp); 5233 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, snd_sml); 5234 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, chrono_start); 5235 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, chrono_stat); 5236 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, write_seq); 5237 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, pushed_seq); 5238 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, lsndtime); 5239 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, mdev_us); 5240 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, tcp_wstamp_ns); 5241 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, accecn_opt_tstamp); 5242 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, rtt_seq); 5243 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, max_packets_out); 5244 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, cwnd_usage_seq); 5245 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, rate_delivered); 5246 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, rate_interval_us); 5247 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, tsorted_sent_queue); 5248 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, highest_sack); 5249 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, ecn_flags); 5250 5251 /* TXRX read-write hotpath cache lines */ 5252 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_txrx, pred_flags); 5253 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_txrx, tcp_clock_cache); 5254 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_txrx, tcp_mstamp); 5255 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_txrx, rcv_nxt); 5256 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_txrx, snd_nxt); 5257 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_txrx, snd_una); 5258 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_txrx, window_clamp); 5259 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_txrx, srtt_us); 5260 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_txrx, packets_out); 5261 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_txrx, snd_up); 5262 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_txrx, received_ce); 5263 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_txrx, received_ecn_bytes); 5264 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_txrx, app_limited); 5265 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_txrx, rcv_wnd); 5266 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_txrx, rcv_mwnd_seq); 5267 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_txrx, rcv_tstamp); 5268 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_txrx, rx_opt); 5269 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_txrx, segs_in); 5270 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_txrx, segs_out); 5271 5272 /* RX read-write hotpath cache lines */ 5273 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_rx, bytes_received); 5274 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_rx, data_segs_in); 5275 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_rx, rcv_wup); 5276 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_rx, rcv_rtt_last_tsecr); 5277 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_rx, delivered_ecn_bytes); 5278 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_rx, pkts_acked_ewma); 5279 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_rx, rcv_rtt_est); 5280 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_rx, rcvq_space); 5281 } 5282 5283 void __init tcp_init(void) 5284 { 5285 int max_rshare, max_wshare, cnt; 5286 unsigned long limit; 5287 unsigned int i; 5288 5289 BUILD_BUG_ON(TCP_MIN_SND_MSS <= MAX_TCP_OPTION_SPACE); 5290 BUILD_BUG_ON(sizeof(struct tcp_skb_cb) > 5291 sizeof_field(struct sk_buff, cb)); 5292 5293 tcp_struct_check(); 5294 5295 percpu_counter_init(&tcp_sockets_allocated, 0, GFP_KERNEL); 5296 5297 timer_setup(&tcp_orphan_timer, tcp_orphan_update, TIMER_DEFERRABLE); 5298 mod_timer(&tcp_orphan_timer, jiffies + TCP_ORPHAN_TIMER_PERIOD); 5299 5300 inet_hashinfo2_init(&tcp_hashinfo, "tcp_listen_portaddr_hash", 5301 thash_entries, 21, /* one slot per 2 MB*/ 5302 0, 64 * 1024); 5303 tcp_hashinfo.bind_bucket_cachep = 5304 kmem_cache_create("tcp_bind_bucket", 5305 sizeof(struct inet_bind_bucket), 0, 5306 SLAB_HWCACHE_ALIGN | SLAB_PANIC | 5307 SLAB_ACCOUNT, 5308 NULL); 5309 tcp_hashinfo.bind2_bucket_cachep = 5310 kmem_cache_create("tcp_bind2_bucket", 5311 sizeof(struct inet_bind2_bucket), 0, 5312 SLAB_HWCACHE_ALIGN | SLAB_PANIC | 5313 SLAB_ACCOUNT, 5314 NULL); 5315 5316 /* Size and allocate the main established and bind bucket 5317 * hash tables. 5318 * 5319 * The methodology is similar to that of the buffer cache. 5320 */ 5321 tcp_hashinfo.ehash = 5322 alloc_large_system_hash("TCP established", 5323 sizeof(struct inet_ehash_bucket), 5324 thash_entries, 5325 17, /* one slot per 128 KB of memory */ 5326 0, 5327 NULL, 5328 &tcp_hashinfo.ehash_mask, 5329 0, 5330 thash_entries ? 0 : 512 * 1024); 5331 for (i = 0; i <= tcp_hashinfo.ehash_mask; i++) 5332 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].chain, i); 5333 5334 if (inet_ehash_locks_alloc(&tcp_hashinfo)) 5335 panic("TCP: failed to alloc ehash_locks"); 5336 tcp_hashinfo.bhash = 5337 alloc_large_system_hash("TCP bind", 5338 2 * sizeof(struct inet_bind_hashbucket), 5339 tcp_hashinfo.ehash_mask + 1, 5340 17, /* one slot per 128 KB of memory */ 5341 0, 5342 &tcp_hashinfo.bhash_size, 5343 NULL, 5344 0, 5345 64 * 1024); 5346 tcp_hashinfo.bhash_size = 1U << tcp_hashinfo.bhash_size; 5347 tcp_hashinfo.bhash2 = tcp_hashinfo.bhash + tcp_hashinfo.bhash_size; 5348 for (i = 0; i < tcp_hashinfo.bhash_size; i++) { 5349 spin_lock_init(&tcp_hashinfo.bhash[i].lock); 5350 INIT_HLIST_HEAD(&tcp_hashinfo.bhash[i].chain); 5351 spin_lock_init(&tcp_hashinfo.bhash2[i].lock); 5352 INIT_HLIST_HEAD(&tcp_hashinfo.bhash2[i].chain); 5353 } 5354 5355 tcp_hashinfo.pernet = false; 5356 5357 cnt = tcp_hashinfo.ehash_mask + 1; 5358 sysctl_tcp_max_orphans = cnt / 2; 5359 5360 tcp_init_mem(); 5361 /* Set per-socket limits to no more than 1/128 the pressure threshold */ 5362 limit = nr_free_buffer_pages() << (PAGE_SHIFT - 7); 5363 max_wshare = min(4UL*1024*1024, limit); 5364 max_rshare = min(32UL*1024*1024, limit); 5365 5366 init_net.ipv4.sysctl_tcp_wmem[0] = PAGE_SIZE; 5367 init_net.ipv4.sysctl_tcp_wmem[1] = 16*1024; 5368 init_net.ipv4.sysctl_tcp_wmem[2] = max(64*1024, max_wshare); 5369 5370 init_net.ipv4.sysctl_tcp_rmem[0] = PAGE_SIZE; 5371 init_net.ipv4.sysctl_tcp_rmem[1] = 131072; 5372 init_net.ipv4.sysctl_tcp_rmem[2] = max(131072, max_rshare); 5373 5374 pr_info("Hash tables configured (established %u bind %u)\n", 5375 tcp_hashinfo.ehash_mask + 1, tcp_hashinfo.bhash_size); 5376 5377 tcp_v4_init(); 5378 tcp_metrics_init(); 5379 BUG_ON(tcp_register_congestion_control(&tcp_reno) != 0); 5380 tcp_tsq_work_init(); 5381 mptcp_init(); 5382 } 5383