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 && (zc != MSG_ZEROCOPY || !binding)) { 1173 err = -EINVAL; 1174 goto out_err; 1175 } 1176 1177 if (unlikely(flags & MSG_FASTOPEN || 1178 inet_test_bit(DEFER_CONNECT, sk)) && 1179 !tp->repair) { 1180 err = tcp_sendmsg_fastopen(sk, msg, &copied_syn, size, uarg); 1181 if (err == -EINPROGRESS && copied_syn > 0) 1182 goto out; 1183 else if (err) 1184 goto out_err; 1185 } 1186 1187 timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT); 1188 1189 tcp_rate_check_app_limited(sk); /* is sending application-limited? */ 1190 1191 /* Wait for a connection to finish. One exception is TCP Fast Open 1192 * (passive side) where data is allowed to be sent before a connection 1193 * is fully established. 1194 */ 1195 if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) && 1196 !tcp_passive_fastopen(sk)) { 1197 err = sk_stream_wait_connect(sk, &timeo); 1198 if (err != 0) 1199 goto do_error; 1200 } 1201 1202 if (unlikely(tp->repair)) { 1203 if (tp->repair_queue == TCP_RECV_QUEUE) { 1204 copied = tcp_send_rcvq(sk, msg, size); 1205 goto out_nopush; 1206 } 1207 1208 err = -EINVAL; 1209 if (tp->repair_queue == TCP_NO_QUEUE) 1210 goto out_err; 1211 1212 /* 'common' sending to sendq */ 1213 } 1214 1215 if (sockc_err) { 1216 err = sockc_err; 1217 goto out_err; 1218 } 1219 1220 /* This should be in poll */ 1221 sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk); 1222 1223 /* Ok commence sending. */ 1224 copied = 0; 1225 1226 restart: 1227 mss_now = tcp_send_mss(sk, &size_goal, flags); 1228 1229 err = -EPIPE; 1230 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN)) 1231 goto do_error; 1232 1233 while (msg_data_left(msg)) { 1234 int copy = 0; 1235 1236 skb = tcp_write_queue_tail(sk); 1237 if (skb) 1238 copy = size_goal - skb->len; 1239 1240 trace_tcp_sendmsg_locked(sk, msg, skb, size_goal); 1241 1242 if (copy <= 0 || !tcp_skb_can_collapse_to(skb) || 1243 unlikely(skb_frags_readable(skb) != !binding)) { 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 if (skb_frag_is_net_iov(frag)) 1912 return false; 1913 1914 page = skb_frag_page(frag); 1915 1916 if (PageCompound(page) || page->mapping) 1917 return false; 1918 1919 return true; 1920 } 1921 1922 static int find_next_mappable_frag(const skb_frag_t *frag, 1923 int remaining_in_skb) 1924 { 1925 int offset = 0; 1926 1927 if (likely(can_map_frag(frag))) 1928 return 0; 1929 1930 while (offset < remaining_in_skb && !can_map_frag(frag)) { 1931 offset += skb_frag_size(frag); 1932 ++frag; 1933 } 1934 return offset; 1935 } 1936 1937 static void tcp_zerocopy_set_hint_for_skb(struct sock *sk, 1938 struct tcp_zerocopy_receive *zc, 1939 struct sk_buff *skb, u32 offset) 1940 { 1941 u32 frag_offset, partial_frag_remainder = 0; 1942 int mappable_offset; 1943 skb_frag_t *frag; 1944 1945 /* worst case: skip to next skb. try to improve on this case below */ 1946 zc->recv_skip_hint = skb->len - offset; 1947 1948 /* Find the frag containing this offset (and how far into that frag) */ 1949 frag = skb_advance_to_frag(skb, offset, &frag_offset); 1950 if (!frag) 1951 return; 1952 1953 if (frag_offset) { 1954 struct skb_shared_info *info = skb_shinfo(skb); 1955 1956 /* We read part of the last frag, must recvmsg() rest of skb. */ 1957 if (frag == &info->frags[info->nr_frags - 1]) 1958 return; 1959 1960 /* Else, we must at least read the remainder in this frag. */ 1961 partial_frag_remainder = skb_frag_size(frag) - frag_offset; 1962 zc->recv_skip_hint -= partial_frag_remainder; 1963 ++frag; 1964 } 1965 1966 /* partial_frag_remainder: If part way through a frag, must read rest. 1967 * mappable_offset: Bytes till next mappable frag, *not* counting bytes 1968 * in partial_frag_remainder. 1969 */ 1970 mappable_offset = find_next_mappable_frag(frag, zc->recv_skip_hint); 1971 zc->recv_skip_hint = mappable_offset + partial_frag_remainder; 1972 } 1973 1974 static int tcp_recvmsg_locked(struct sock *sk, struct msghdr *msg, size_t len, 1975 int flags, struct scm_timestamping_internal *tss, 1976 int *cmsg_flags); 1977 static int receive_fallback_to_copy(struct sock *sk, 1978 struct tcp_zerocopy_receive *zc, int inq, 1979 struct scm_timestamping_internal *tss) 1980 { 1981 unsigned long copy_address = (unsigned long)zc->copybuf_address; 1982 struct msghdr msg = {}; 1983 int err; 1984 1985 zc->length = 0; 1986 zc->recv_skip_hint = 0; 1987 1988 if (copy_address != zc->copybuf_address) 1989 return -EINVAL; 1990 1991 err = import_ubuf(ITER_DEST, (void __user *)copy_address, inq, 1992 &msg.msg_iter); 1993 if (err) 1994 return err; 1995 1996 err = tcp_recvmsg_locked(sk, &msg, inq, MSG_DONTWAIT, 1997 tss, &zc->msg_flags); 1998 if (err < 0) 1999 return err; 2000 2001 zc->copybuf_len = err; 2002 if (likely(zc->copybuf_len)) { 2003 struct sk_buff *skb; 2004 u32 offset; 2005 2006 skb = tcp_recv_skb(sk, tcp_sk(sk)->copied_seq, &offset); 2007 if (skb) 2008 tcp_zerocopy_set_hint_for_skb(sk, zc, skb, offset); 2009 } 2010 return 0; 2011 } 2012 2013 static int tcp_copy_straggler_data(struct tcp_zerocopy_receive *zc, 2014 struct sk_buff *skb, u32 copylen, 2015 u32 *offset, u32 *seq) 2016 { 2017 unsigned long copy_address = (unsigned long)zc->copybuf_address; 2018 struct msghdr msg = {}; 2019 int err; 2020 2021 if (copy_address != zc->copybuf_address) 2022 return -EINVAL; 2023 2024 err = import_ubuf(ITER_DEST, (void __user *)copy_address, copylen, 2025 &msg.msg_iter); 2026 if (err) 2027 return err; 2028 err = skb_copy_datagram_msg(skb, *offset, &msg, copylen); 2029 if (err) 2030 return err; 2031 zc->recv_skip_hint -= copylen; 2032 *offset += copylen; 2033 *seq += copylen; 2034 return (__s32)copylen; 2035 } 2036 2037 static int tcp_zc_handle_leftover(struct tcp_zerocopy_receive *zc, 2038 struct sock *sk, 2039 struct sk_buff *skb, 2040 u32 *seq, 2041 s32 copybuf_len, 2042 struct scm_timestamping_internal *tss) 2043 { 2044 u32 offset, copylen = min_t(u32, copybuf_len, zc->recv_skip_hint); 2045 2046 if (!copylen) 2047 return 0; 2048 /* skb is null if inq < PAGE_SIZE. */ 2049 if (skb) { 2050 offset = *seq - TCP_SKB_CB(skb)->seq; 2051 } else { 2052 skb = tcp_recv_skb(sk, *seq, &offset); 2053 if (TCP_SKB_CB(skb)->has_rxtstamp) { 2054 tcp_update_recv_tstamps(skb, tss); 2055 zc->msg_flags |= TCP_CMSG_TS; 2056 } 2057 } 2058 2059 zc->copybuf_len = tcp_copy_straggler_data(zc, skb, copylen, &offset, 2060 seq); 2061 return zc->copybuf_len < 0 ? 0 : copylen; 2062 } 2063 2064 static int tcp_zerocopy_vm_insert_batch_error(struct vm_area_struct *vma, 2065 struct page **pending_pages, 2066 unsigned long pages_remaining, 2067 unsigned long *address, 2068 u32 *length, 2069 u32 *seq, 2070 struct tcp_zerocopy_receive *zc, 2071 u32 total_bytes_to_map, 2072 int err) 2073 { 2074 /* At least one page did not map. Try zapping if we skipped earlier. */ 2075 if (err == -EBUSY && 2076 zc->flags & TCP_RECEIVE_ZEROCOPY_FLAG_TLB_CLEAN_HINT) { 2077 u32 maybe_zap_len; 2078 2079 maybe_zap_len = total_bytes_to_map - /* All bytes to map */ 2080 *length + /* Mapped or pending */ 2081 (pages_remaining * PAGE_SIZE); /* Failed map. */ 2082 zap_vma_range(vma, *address, maybe_zap_len); 2083 err = 0; 2084 } 2085 2086 if (!err) { 2087 unsigned long leftover_pages = pages_remaining; 2088 int bytes_mapped; 2089 2090 /* We called zap_vma_range, try to reinsert. */ 2091 err = vm_insert_pages(vma, *address, 2092 pending_pages, 2093 &pages_remaining); 2094 bytes_mapped = PAGE_SIZE * (leftover_pages - pages_remaining); 2095 *seq += bytes_mapped; 2096 *address += bytes_mapped; 2097 } 2098 if (err) { 2099 /* Either we were unable to zap, OR we zapped, retried an 2100 * insert, and still had an issue. Either ways, pages_remaining 2101 * is the number of pages we were unable to map, and we unroll 2102 * some state we speculatively touched before. 2103 */ 2104 const int bytes_not_mapped = PAGE_SIZE * pages_remaining; 2105 2106 *length -= bytes_not_mapped; 2107 zc->recv_skip_hint += bytes_not_mapped; 2108 } 2109 return err; 2110 } 2111 2112 static int tcp_zerocopy_vm_insert_batch(struct vm_area_struct *vma, 2113 struct page **pages, 2114 unsigned int pages_to_map, 2115 unsigned long *address, 2116 u32 *length, 2117 u32 *seq, 2118 struct tcp_zerocopy_receive *zc, 2119 u32 total_bytes_to_map) 2120 { 2121 unsigned long pages_remaining = pages_to_map; 2122 unsigned int pages_mapped; 2123 unsigned int bytes_mapped; 2124 int err; 2125 2126 err = vm_insert_pages(vma, *address, pages, &pages_remaining); 2127 pages_mapped = pages_to_map - (unsigned int)pages_remaining; 2128 bytes_mapped = PAGE_SIZE * pages_mapped; 2129 /* Even if vm_insert_pages fails, it may have partially succeeded in 2130 * mapping (some but not all of the pages). 2131 */ 2132 *seq += bytes_mapped; 2133 *address += bytes_mapped; 2134 2135 if (likely(!err)) 2136 return 0; 2137 2138 /* Error: maybe zap and retry + rollback state for failed inserts. */ 2139 return tcp_zerocopy_vm_insert_batch_error(vma, pages + pages_mapped, 2140 pages_remaining, address, length, seq, zc, total_bytes_to_map, 2141 err); 2142 } 2143 2144 #define TCP_VALID_ZC_MSG_FLAGS (TCP_CMSG_TS) 2145 static void tcp_zc_finalize_rx_tstamp(struct sock *sk, 2146 struct tcp_zerocopy_receive *zc, 2147 struct scm_timestamping_internal *tss) 2148 { 2149 unsigned long msg_control_addr; 2150 struct msghdr cmsg_dummy; 2151 2152 msg_control_addr = (unsigned long)zc->msg_control; 2153 cmsg_dummy.msg_control_user = (void __user *)msg_control_addr; 2154 cmsg_dummy.msg_controllen = 2155 (__kernel_size_t)zc->msg_controllen; 2156 cmsg_dummy.msg_flags = in_compat_syscall() 2157 ? MSG_CMSG_COMPAT : 0; 2158 cmsg_dummy.msg_control_is_user = true; 2159 zc->msg_flags = 0; 2160 if (zc->msg_control == msg_control_addr && 2161 zc->msg_controllen == cmsg_dummy.msg_controllen) { 2162 tcp_recv_timestamp(&cmsg_dummy, sk, tss); 2163 zc->msg_control = (__u64) 2164 ((uintptr_t)cmsg_dummy.msg_control_user); 2165 zc->msg_controllen = 2166 (__u64)cmsg_dummy.msg_controllen; 2167 zc->msg_flags = (__u32)cmsg_dummy.msg_flags; 2168 } 2169 } 2170 2171 static struct vm_area_struct *find_tcp_vma(struct mm_struct *mm, 2172 unsigned long address, 2173 bool *mmap_locked) 2174 { 2175 struct vm_area_struct *vma = lock_vma_under_rcu(mm, address); 2176 2177 if (vma) { 2178 if (vma->vm_ops != &tcp_vm_ops) { 2179 vma_end_read(vma); 2180 return NULL; 2181 } 2182 *mmap_locked = false; 2183 return vma; 2184 } 2185 2186 mmap_read_lock(mm); 2187 vma = vma_lookup(mm, address); 2188 if (!vma || vma->vm_ops != &tcp_vm_ops) { 2189 mmap_read_unlock(mm); 2190 return NULL; 2191 } 2192 *mmap_locked = true; 2193 return vma; 2194 } 2195 2196 #define TCP_ZEROCOPY_PAGE_BATCH_SIZE 32 2197 static int tcp_zerocopy_receive(struct sock *sk, 2198 struct tcp_zerocopy_receive *zc, 2199 struct scm_timestamping_internal *tss) 2200 { 2201 u32 length = 0, offset, vma_len, avail_len, copylen = 0; 2202 unsigned long address = (unsigned long)zc->address; 2203 struct page *pages[TCP_ZEROCOPY_PAGE_BATCH_SIZE]; 2204 s32 copybuf_len = zc->copybuf_len; 2205 struct tcp_sock *tp = tcp_sk(sk); 2206 const skb_frag_t *frags = NULL; 2207 unsigned int pages_to_map = 0; 2208 struct vm_area_struct *vma; 2209 struct sk_buff *skb = NULL; 2210 u32 seq = tp->copied_seq; 2211 u32 total_bytes_to_map; 2212 int inq = tcp_inq(sk); 2213 bool mmap_locked; 2214 int ret; 2215 2216 zc->copybuf_len = 0; 2217 zc->msg_flags = 0; 2218 2219 if (address & (PAGE_SIZE - 1) || address != zc->address) 2220 return -EINVAL; 2221 2222 if (sk->sk_state == TCP_LISTEN) 2223 return -ENOTCONN; 2224 2225 sock_rps_record_flow(sk); 2226 2227 if (inq && inq <= copybuf_len) 2228 return receive_fallback_to_copy(sk, zc, inq, tss); 2229 2230 if (inq < PAGE_SIZE) { 2231 zc->length = 0; 2232 zc->recv_skip_hint = inq; 2233 if (!inq && sock_flag(sk, SOCK_DONE)) 2234 return -EIO; 2235 return 0; 2236 } 2237 2238 vma = find_tcp_vma(current->mm, address, &mmap_locked); 2239 if (!vma) 2240 return -EINVAL; 2241 2242 vma_len = min_t(unsigned long, zc->length, vma->vm_end - address); 2243 avail_len = min_t(u32, vma_len, inq); 2244 total_bytes_to_map = avail_len & ~(PAGE_SIZE - 1); 2245 if (total_bytes_to_map) { 2246 if (!(zc->flags & TCP_RECEIVE_ZEROCOPY_FLAG_TLB_CLEAN_HINT)) 2247 zap_vma_range(vma, address, total_bytes_to_map); 2248 zc->length = total_bytes_to_map; 2249 zc->recv_skip_hint = 0; 2250 } else { 2251 zc->length = avail_len; 2252 zc->recv_skip_hint = avail_len; 2253 } 2254 ret = 0; 2255 while (length + PAGE_SIZE <= zc->length) { 2256 int mappable_offset; 2257 struct page *page; 2258 2259 if (zc->recv_skip_hint < PAGE_SIZE) { 2260 u32 offset_frag; 2261 2262 if (skb) { 2263 if (zc->recv_skip_hint > 0) 2264 break; 2265 skb = skb->next; 2266 offset = seq - TCP_SKB_CB(skb)->seq; 2267 } else { 2268 skb = tcp_recv_skb(sk, seq, &offset); 2269 } 2270 2271 if (!skb_frags_readable(skb)) 2272 break; 2273 2274 if (TCP_SKB_CB(skb)->has_rxtstamp) { 2275 tcp_update_recv_tstamps(skb, tss); 2276 zc->msg_flags |= TCP_CMSG_TS; 2277 } 2278 zc->recv_skip_hint = skb->len - offset; 2279 frags = skb_advance_to_frag(skb, offset, &offset_frag); 2280 if (!frags || offset_frag) 2281 break; 2282 } 2283 2284 mappable_offset = find_next_mappable_frag(frags, 2285 zc->recv_skip_hint); 2286 if (mappable_offset) { 2287 zc->recv_skip_hint = mappable_offset; 2288 break; 2289 } 2290 page = skb_frag_page(frags); 2291 if (WARN_ON_ONCE(!page)) 2292 break; 2293 2294 prefetchw(page); 2295 pages[pages_to_map++] = page; 2296 length += PAGE_SIZE; 2297 zc->recv_skip_hint -= PAGE_SIZE; 2298 frags++; 2299 if (pages_to_map == TCP_ZEROCOPY_PAGE_BATCH_SIZE || 2300 zc->recv_skip_hint < PAGE_SIZE) { 2301 /* Either full batch, or we're about to go to next skb 2302 * (and we cannot unroll failed ops across skbs). 2303 */ 2304 ret = tcp_zerocopy_vm_insert_batch(vma, pages, 2305 pages_to_map, 2306 &address, &length, 2307 &seq, zc, 2308 total_bytes_to_map); 2309 if (ret) 2310 goto out; 2311 pages_to_map = 0; 2312 } 2313 } 2314 if (pages_to_map) { 2315 ret = tcp_zerocopy_vm_insert_batch(vma, pages, pages_to_map, 2316 &address, &length, &seq, 2317 zc, total_bytes_to_map); 2318 } 2319 out: 2320 if (mmap_locked) 2321 mmap_read_unlock(current->mm); 2322 else 2323 vma_end_read(vma); 2324 /* Try to copy straggler data. */ 2325 if (!ret) 2326 copylen = tcp_zc_handle_leftover(zc, sk, skb, &seq, copybuf_len, tss); 2327 2328 if (length + copylen) { 2329 WRITE_ONCE(tp->copied_seq, seq); 2330 tcp_rcv_space_adjust(sk); 2331 2332 /* Clean up data we have read: This will do ACK frames. */ 2333 tcp_recv_skb(sk, seq, &offset); 2334 tcp_cleanup_rbuf(sk, length + copylen); 2335 ret = 0; 2336 if (length == zc->length) 2337 zc->recv_skip_hint = 0; 2338 } else { 2339 if (!zc->recv_skip_hint && sock_flag(sk, SOCK_DONE)) 2340 ret = -EIO; 2341 } 2342 zc->length = length; 2343 return ret; 2344 } 2345 #endif 2346 2347 /* Similar to __sock_recv_timestamp, but does not require an skb */ 2348 void tcp_recv_timestamp(struct msghdr *msg, const struct sock *sk, 2349 struct scm_timestamping_internal *tss) 2350 { 2351 int new_tstamp = sock_flag(sk, SOCK_TSTAMP_NEW); 2352 u32 tsflags = READ_ONCE(sk->sk_tsflags); 2353 2354 if (tss->ts[0]) { 2355 if (sock_flag(sk, SOCK_RCVTSTAMP)) { 2356 struct timespec64 tv = ktime_to_timespec64(tss->ts[0]); 2357 2358 if (sock_flag(sk, SOCK_RCVTSTAMPNS)) { 2359 if (new_tstamp) { 2360 struct __kernel_timespec kts = { 2361 .tv_sec = tv.tv_sec, 2362 .tv_nsec = tv.tv_nsec, 2363 }; 2364 put_cmsg(msg, SOL_SOCKET, SO_TIMESTAMPNS_NEW, 2365 sizeof(kts), &kts); 2366 } else { 2367 struct __kernel_old_timespec ts_old = { 2368 .tv_sec = tv.tv_sec, 2369 .tv_nsec = tv.tv_nsec, 2370 }; 2371 put_cmsg(msg, SOL_SOCKET, SO_TIMESTAMPNS_OLD, 2372 sizeof(ts_old), &ts_old); 2373 } 2374 } else { 2375 if (new_tstamp) { 2376 struct __kernel_sock_timeval stv = { 2377 .tv_sec = tv.tv_sec, 2378 .tv_usec = tv.tv_nsec / 1000, 2379 }; 2380 put_cmsg(msg, SOL_SOCKET, SO_TIMESTAMP_NEW, 2381 sizeof(stv), &stv); 2382 } else { 2383 struct __kernel_old_timeval otv = { 2384 .tv_sec = tv.tv_sec, 2385 .tv_usec = tv.tv_nsec / 1000, 2386 }; 2387 put_cmsg(msg, SOL_SOCKET, SO_TIMESTAMP_OLD, 2388 sizeof(otv), &otv); 2389 } 2390 } 2391 } 2392 2393 if (!(tsflags & SOF_TIMESTAMPING_SOFTWARE && 2394 (tsflags & SOF_TIMESTAMPING_RX_SOFTWARE || 2395 !(tsflags & SOF_TIMESTAMPING_OPT_RX_FILTER)))) 2396 tss->ts[0] = 0; 2397 } 2398 2399 if (tss->ts[2]) { 2400 if (!(tsflags & SOF_TIMESTAMPING_RAW_HARDWARE && 2401 (tsflags & SOF_TIMESTAMPING_RX_HARDWARE || 2402 !(tsflags & SOF_TIMESTAMPING_OPT_RX_FILTER)))) 2403 tss->ts[2] = 0; 2404 } 2405 2406 if (tss->ts[0] | tss->ts[2]) { 2407 tss->ts[1] = 0; 2408 if (sock_flag(sk, SOCK_TSTAMP_NEW)) 2409 put_cmsg_scm_timestamping64(msg, tss); 2410 else 2411 put_cmsg_scm_timestamping(msg, tss); 2412 } 2413 } 2414 2415 static int tcp_inq_hint(struct sock *sk) 2416 { 2417 const struct tcp_sock *tp = tcp_sk(sk); 2418 u32 copied_seq = READ_ONCE(tp->copied_seq); 2419 u32 rcv_nxt = READ_ONCE(tp->rcv_nxt); 2420 int inq; 2421 2422 inq = rcv_nxt - copied_seq; 2423 if (unlikely(inq < 0 || copied_seq != READ_ONCE(tp->copied_seq))) { 2424 lock_sock(sk); 2425 inq = tp->rcv_nxt - tp->copied_seq; 2426 release_sock(sk); 2427 } 2428 /* After receiving a FIN, tell the user-space to continue reading 2429 * by returning a non-zero inq. 2430 */ 2431 if (inq == 0 && sock_flag(sk, SOCK_DONE)) 2432 inq = 1; 2433 return inq; 2434 } 2435 2436 /* batch __xa_alloc() calls and reduce xa_lock()/xa_unlock() overhead. */ 2437 struct tcp_xa_pool { 2438 u8 max; /* max <= MAX_SKB_FRAGS */ 2439 u8 idx; /* idx <= max */ 2440 __u32 tokens[MAX_SKB_FRAGS]; 2441 netmem_ref netmems[MAX_SKB_FRAGS]; 2442 }; 2443 2444 static void tcp_xa_pool_commit_locked(struct sock *sk, struct tcp_xa_pool *p) 2445 { 2446 int i; 2447 2448 /* Commit part that has been copied to user space. */ 2449 for (i = 0; i < p->idx; i++) 2450 __xa_cmpxchg(&sk->sk_user_frags, p->tokens[i], XA_ZERO_ENTRY, 2451 (__force void *)p->netmems[i], GFP_KERNEL); 2452 /* Rollback what has been pre-allocated and is no longer needed. */ 2453 for (; i < p->max; i++) 2454 __xa_erase(&sk->sk_user_frags, p->tokens[i]); 2455 2456 p->max = 0; 2457 p->idx = 0; 2458 } 2459 2460 static void tcp_xa_pool_commit(struct sock *sk, struct tcp_xa_pool *p) 2461 { 2462 if (!p->max) 2463 return; 2464 2465 xa_lock_bh(&sk->sk_user_frags); 2466 2467 tcp_xa_pool_commit_locked(sk, p); 2468 2469 xa_unlock_bh(&sk->sk_user_frags); 2470 } 2471 2472 static int tcp_xa_pool_refill(struct sock *sk, struct tcp_xa_pool *p, 2473 unsigned int max_frags) 2474 { 2475 int err, k; 2476 2477 if (p->idx < p->max) 2478 return 0; 2479 2480 xa_lock_bh(&sk->sk_user_frags); 2481 2482 tcp_xa_pool_commit_locked(sk, p); 2483 2484 for (k = 0; k < max_frags; k++) { 2485 err = __xa_alloc(&sk->sk_user_frags, &p->tokens[k], 2486 XA_ZERO_ENTRY, xa_limit_31b, GFP_KERNEL); 2487 if (err) 2488 break; 2489 } 2490 2491 xa_unlock_bh(&sk->sk_user_frags); 2492 2493 p->max = k; 2494 p->idx = 0; 2495 return k ? 0 : err; 2496 } 2497 2498 /* On error, returns the -errno. On success, returns number of bytes sent to the 2499 * user. May not consume all of @remaining_len. 2500 */ 2501 static int tcp_recvmsg_dmabuf(struct sock *sk, const struct sk_buff *skb, 2502 unsigned int offset, struct msghdr *msg, 2503 int remaining_len) 2504 { 2505 struct dmabuf_cmsg dmabuf_cmsg = { 0 }; 2506 struct tcp_xa_pool tcp_xa_pool; 2507 unsigned int start; 2508 int i, copy, n; 2509 int sent = 0; 2510 int err = 0; 2511 2512 tcp_xa_pool.max = 0; 2513 tcp_xa_pool.idx = 0; 2514 do { 2515 start = skb_headlen(skb); 2516 2517 if (skb_frags_readable(skb)) { 2518 err = -ENODEV; 2519 goto out; 2520 } 2521 2522 /* Copy header. */ 2523 copy = start - offset; 2524 if (copy > 0) { 2525 copy = min(copy, remaining_len); 2526 2527 n = copy_to_iter(skb->data + offset, copy, 2528 &msg->msg_iter); 2529 if (n != copy) { 2530 err = -EFAULT; 2531 goto out; 2532 } 2533 2534 offset += copy; 2535 remaining_len -= copy; 2536 2537 /* First a dmabuf_cmsg for # bytes copied to user 2538 * buffer. 2539 */ 2540 memset(&dmabuf_cmsg, 0, sizeof(dmabuf_cmsg)); 2541 dmabuf_cmsg.frag_size = copy; 2542 err = put_cmsg_notrunc(msg, SOL_SOCKET, 2543 SO_DEVMEM_LINEAR, 2544 sizeof(dmabuf_cmsg), 2545 &dmabuf_cmsg); 2546 if (err) 2547 goto out; 2548 2549 sent += copy; 2550 2551 if (remaining_len == 0) 2552 goto out; 2553 } 2554 2555 /* after that, send information of dmabuf pages through a 2556 * sequence of cmsg 2557 */ 2558 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) { 2559 skb_frag_t *frag = &skb_shinfo(skb)->frags[i]; 2560 struct net_iov *niov; 2561 u64 frag_offset; 2562 int end; 2563 2564 /* !skb_frags_readable() should indicate that ALL the 2565 * frags in this skb are dmabuf net_iovs. We're checking 2566 * for that flag above, but also check individual frags 2567 * here. If the tcp stack is not setting 2568 * skb_frags_readable() correctly, we still don't want 2569 * to crash here. 2570 */ 2571 if (!skb_frag_net_iov(frag)) { 2572 net_err_ratelimited("Found non-dmabuf skb with net_iov"); 2573 err = -ENODEV; 2574 goto out; 2575 } 2576 2577 niov = skb_frag_net_iov(frag); 2578 if (!net_is_devmem_iov(niov)) { 2579 err = -ENODEV; 2580 goto out; 2581 } 2582 2583 end = start + skb_frag_size(frag); 2584 copy = end - offset; 2585 2586 if (copy > 0) { 2587 copy = min(copy, remaining_len); 2588 2589 frag_offset = net_iov_virtual_addr(niov) + 2590 skb_frag_off(frag) + offset - 2591 start; 2592 dmabuf_cmsg.frag_offset = frag_offset; 2593 dmabuf_cmsg.frag_size = copy; 2594 err = tcp_xa_pool_refill(sk, &tcp_xa_pool, 2595 skb_shinfo(skb)->nr_frags - i); 2596 if (err) 2597 goto out; 2598 2599 /* Will perform the exchange later */ 2600 dmabuf_cmsg.frag_token = tcp_xa_pool.tokens[tcp_xa_pool.idx]; 2601 dmabuf_cmsg.dmabuf_id = net_devmem_iov_binding_id(niov); 2602 2603 offset += copy; 2604 remaining_len -= copy; 2605 2606 err = put_cmsg_notrunc(msg, SOL_SOCKET, 2607 SO_DEVMEM_DMABUF, 2608 sizeof(dmabuf_cmsg), 2609 &dmabuf_cmsg); 2610 if (err) 2611 goto out; 2612 2613 atomic_long_inc(&niov->desc.pp_ref_count); 2614 tcp_xa_pool.netmems[tcp_xa_pool.idx++] = skb_frag_netmem(frag); 2615 2616 sent += copy; 2617 2618 if (remaining_len == 0) 2619 goto out; 2620 } 2621 start = end; 2622 } 2623 2624 tcp_xa_pool_commit(sk, &tcp_xa_pool); 2625 if (!remaining_len) 2626 goto out; 2627 2628 /* if remaining_len is not satisfied yet, we need to go to the 2629 * next frag in the frag_list to satisfy remaining_len. 2630 */ 2631 skb = skb_shinfo(skb)->frag_list ?: skb->next; 2632 2633 offset = offset - start; 2634 } while (skb); 2635 2636 if (remaining_len) { 2637 err = -EFAULT; 2638 goto out; 2639 } 2640 2641 out: 2642 tcp_xa_pool_commit(sk, &tcp_xa_pool); 2643 if (!sent) 2644 sent = err; 2645 2646 return sent; 2647 } 2648 2649 /* 2650 * This routine copies from a sock struct into the user buffer. 2651 * 2652 * Technical note: in 2.3 we work on _locked_ socket, so that 2653 * tricks with *seq access order and skb->users are not required. 2654 * Probably, code can be easily improved even more. 2655 */ 2656 2657 static int tcp_recvmsg_locked(struct sock *sk, struct msghdr *msg, size_t len, 2658 int flags, struct scm_timestamping_internal *tss, 2659 int *cmsg_flags) 2660 { 2661 struct tcp_sock *tp = tcp_sk(sk); 2662 int last_copied_dmabuf = -1; /* uninitialized */ 2663 int copied = 0; 2664 u32 peek_seq; 2665 u32 *seq; 2666 unsigned long used; 2667 int err; 2668 int target; /* Read at least this many bytes */ 2669 long timeo; 2670 struct sk_buff *skb, *last; 2671 u32 peek_offset = 0; 2672 u32 urg_hole = 0; 2673 2674 err = -ENOTCONN; 2675 if (sk->sk_state == TCP_LISTEN) 2676 goto out; 2677 2678 if (tp->recvmsg_inq) 2679 *cmsg_flags = TCP_CMSG_INQ; 2680 timeo = sock_rcvtimeo(sk, flags & MSG_DONTWAIT); 2681 2682 /* Urgent data needs to be handled specially. */ 2683 if (flags & MSG_OOB) 2684 goto recv_urg; 2685 2686 if (unlikely(tp->repair)) { 2687 err = -EPERM; 2688 if (!(flags & MSG_PEEK)) 2689 goto out; 2690 2691 if (tp->repair_queue == TCP_SEND_QUEUE) 2692 goto recv_sndq; 2693 2694 err = -EINVAL; 2695 if (tp->repair_queue == TCP_NO_QUEUE) 2696 goto out; 2697 2698 /* 'common' recv queue MSG_PEEK-ing */ 2699 } 2700 2701 seq = &tp->copied_seq; 2702 if (flags & MSG_PEEK) { 2703 peek_offset = max(sk_peek_offset(sk, flags), 0); 2704 peek_seq = tp->copied_seq + peek_offset; 2705 seq = &peek_seq; 2706 } 2707 2708 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len); 2709 2710 do { 2711 u32 offset; 2712 2713 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */ 2714 if (unlikely(tp->urg_data) && tp->urg_seq == *seq) { 2715 if (copied) 2716 break; 2717 if (signal_pending(current)) { 2718 copied = timeo ? sock_intr_errno(timeo) : -EAGAIN; 2719 break; 2720 } 2721 } 2722 2723 /* Next get a buffer. */ 2724 2725 last = skb_peek_tail(&sk->sk_receive_queue); 2726 skb_queue_walk(&sk->sk_receive_queue, skb) { 2727 last = skb; 2728 /* Now that we have two receive queues this 2729 * shouldn't happen. 2730 */ 2731 if (WARN(before(*seq, TCP_SKB_CB(skb)->seq), 2732 "TCP recvmsg seq # bug: copied %X, seq %X, rcvnxt %X, fl %X\n", 2733 *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt, 2734 flags)) 2735 break; 2736 2737 offset = *seq - TCP_SKB_CB(skb)->seq; 2738 if (unlikely(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) { 2739 pr_err_once("%s: found a SYN, please report !\n", __func__); 2740 offset--; 2741 } 2742 if (offset < skb->len) 2743 goto found_ok_skb; 2744 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) 2745 goto found_fin_ok; 2746 WARN(!(flags & MSG_PEEK), 2747 "TCP recvmsg seq # bug 2: copied %X, seq %X, rcvnxt %X, fl %X\n", 2748 *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt, flags); 2749 } 2750 2751 /* Well, if we have backlog, try to process it now yet. */ 2752 2753 if (copied >= target && !READ_ONCE(sk->sk_backlog.tail)) 2754 break; 2755 2756 if (copied) { 2757 if (!timeo || 2758 tcp_recv_should_stop(sk)) 2759 break; 2760 } else { 2761 if (sock_flag(sk, SOCK_DONE)) 2762 break; 2763 2764 if (sk->sk_err) { 2765 copied = sock_error(sk); 2766 break; 2767 } 2768 2769 if (sk->sk_shutdown & RCV_SHUTDOWN) 2770 break; 2771 2772 if (sk->sk_state == TCP_CLOSE) { 2773 /* This occurs when user tries to read 2774 * from never connected socket. 2775 */ 2776 copied = -ENOTCONN; 2777 break; 2778 } 2779 2780 if (!timeo) { 2781 copied = -EAGAIN; 2782 break; 2783 } 2784 2785 if (signal_pending(current)) { 2786 copied = sock_intr_errno(timeo); 2787 break; 2788 } 2789 } 2790 2791 if (copied >= target) { 2792 /* Do not sleep, just process backlog. */ 2793 __sk_flush_backlog(sk); 2794 } else { 2795 tcp_cleanup_rbuf(sk, copied); 2796 err = sk_wait_data(sk, &timeo, last); 2797 if (err < 0) { 2798 err = copied ? : err; 2799 goto out; 2800 } 2801 } 2802 2803 if ((flags & MSG_PEEK) && 2804 (peek_seq - peek_offset - copied - urg_hole != tp->copied_seq)) { 2805 net_dbg_ratelimited("TCP(%s:%d): Application bug, race in MSG_PEEK\n", 2806 current->comm, 2807 task_pid_nr(current)); 2808 peek_seq = tp->copied_seq + peek_offset; 2809 } 2810 continue; 2811 2812 found_ok_skb: 2813 /* Ok so how much can we use? */ 2814 used = skb->len - offset; 2815 if (len < used) 2816 used = len; 2817 2818 /* Do we have urgent data here? */ 2819 if (unlikely(tp->urg_data)) { 2820 u32 urg_offset = tp->urg_seq - *seq; 2821 if (urg_offset < used) { 2822 if (!urg_offset) { 2823 if (!sock_flag(sk, SOCK_URGINLINE)) { 2824 WRITE_ONCE(*seq, *seq + 1); 2825 urg_hole++; 2826 offset++; 2827 used--; 2828 if (!used) 2829 goto skip_copy; 2830 } 2831 } else 2832 used = urg_offset; 2833 } 2834 } 2835 2836 if (!(flags & MSG_TRUNC)) { 2837 if (last_copied_dmabuf != -1 && 2838 last_copied_dmabuf != !skb_frags_readable(skb)) 2839 break; 2840 2841 if (skb_frags_readable(skb)) { 2842 err = skb_copy_datagram_msg(skb, offset, msg, 2843 used); 2844 if (err) { 2845 /* Exception. Bailout! */ 2846 if (!copied) 2847 copied = -EFAULT; 2848 break; 2849 } 2850 } else { 2851 if (!(flags & MSG_SOCK_DEVMEM)) { 2852 /* dmabuf skbs can only be received 2853 * with the MSG_SOCK_DEVMEM flag. 2854 */ 2855 if (!copied) 2856 copied = -EFAULT; 2857 2858 break; 2859 } 2860 2861 err = tcp_recvmsg_dmabuf(sk, skb, offset, msg, 2862 used); 2863 if (err < 0) { 2864 if (!copied) 2865 copied = err; 2866 2867 break; 2868 } 2869 used = err; 2870 } 2871 } 2872 2873 last_copied_dmabuf = !skb_frags_readable(skb); 2874 2875 WRITE_ONCE(*seq, *seq + used); 2876 copied += used; 2877 len -= used; 2878 if (flags & MSG_PEEK) 2879 sk_peek_offset_fwd(sk, used); 2880 else 2881 sk_peek_offset_bwd(sk, used); 2882 tcp_rcv_space_adjust(sk); 2883 2884 skip_copy: 2885 if (unlikely(tp->urg_data) && after(tp->copied_seq, tp->urg_seq)) { 2886 WRITE_ONCE(tp->urg_data, 0); 2887 tcp_fast_path_check(sk); 2888 } 2889 2890 if (TCP_SKB_CB(skb)->has_rxtstamp) { 2891 tcp_update_recv_tstamps(skb, tss); 2892 *cmsg_flags |= TCP_CMSG_TS; 2893 } 2894 2895 if (used + offset < skb->len) 2896 continue; 2897 2898 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) 2899 goto found_fin_ok; 2900 if (!(flags & MSG_PEEK)) 2901 tcp_eat_recv_skb(sk, skb); 2902 continue; 2903 2904 found_fin_ok: 2905 /* Process the FIN. */ 2906 WRITE_ONCE(*seq, *seq + 1); 2907 if (!(flags & MSG_PEEK)) 2908 tcp_eat_recv_skb(sk, skb); 2909 break; 2910 } while (len > 0); 2911 2912 /* According to UNIX98, msg_name/msg_namelen are ignored 2913 * on connected socket. I was just happy when found this 8) --ANK 2914 */ 2915 2916 /* Clean up data we have read: This will do ACK frames. */ 2917 tcp_cleanup_rbuf(sk, copied); 2918 return copied; 2919 2920 out: 2921 return err; 2922 2923 recv_urg: 2924 err = tcp_recv_urg(sk, msg, len, flags); 2925 goto out; 2926 2927 recv_sndq: 2928 err = tcp_peek_sndq(sk, msg, len); 2929 goto out; 2930 } 2931 2932 int tcp_recvmsg(struct sock *sk, struct msghdr *msg, size_t len, int flags) 2933 { 2934 int cmsg_flags = 0, ret; 2935 struct scm_timestamping_internal tss; 2936 2937 if (unlikely(flags & MSG_ERRQUEUE)) 2938 return inet_recv_error(sk, msg, len); 2939 2940 if (sk_can_busy_loop(sk) && 2941 skb_queue_empty_lockless(&sk->sk_receive_queue) && 2942 sk->sk_state == TCP_ESTABLISHED) 2943 sk_busy_loop(sk, flags & MSG_DONTWAIT); 2944 2945 lock_sock(sk); 2946 ret = tcp_recvmsg_locked(sk, msg, len, flags, &tss, &cmsg_flags); 2947 release_sock(sk); 2948 2949 if ((cmsg_flags | msg->msg_get_inq) && ret >= 0) { 2950 if (cmsg_flags & TCP_CMSG_TS) 2951 tcp_recv_timestamp(msg, sk, &tss); 2952 if ((cmsg_flags & TCP_CMSG_INQ) | msg->msg_get_inq) { 2953 msg->msg_inq = tcp_inq_hint(sk); 2954 if (cmsg_flags & TCP_CMSG_INQ) 2955 put_cmsg(msg, SOL_TCP, TCP_CM_INQ, 2956 sizeof(msg->msg_inq), &msg->msg_inq); 2957 } 2958 } 2959 return ret; 2960 } 2961 2962 void tcp_set_state(struct sock *sk, int state) 2963 { 2964 int oldstate = sk->sk_state; 2965 2966 /* We defined a new enum for TCP states that are exported in BPF 2967 * so as not force the internal TCP states to be frozen. The 2968 * following checks will detect if an internal state value ever 2969 * differs from the BPF value. If this ever happens, then we will 2970 * need to remap the internal value to the BPF value before calling 2971 * tcp_call_bpf_2arg. 2972 */ 2973 BUILD_BUG_ON((int)BPF_TCP_ESTABLISHED != (int)TCP_ESTABLISHED); 2974 BUILD_BUG_ON((int)BPF_TCP_SYN_SENT != (int)TCP_SYN_SENT); 2975 BUILD_BUG_ON((int)BPF_TCP_SYN_RECV != (int)TCP_SYN_RECV); 2976 BUILD_BUG_ON((int)BPF_TCP_FIN_WAIT1 != (int)TCP_FIN_WAIT1); 2977 BUILD_BUG_ON((int)BPF_TCP_FIN_WAIT2 != (int)TCP_FIN_WAIT2); 2978 BUILD_BUG_ON((int)BPF_TCP_TIME_WAIT != (int)TCP_TIME_WAIT); 2979 BUILD_BUG_ON((int)BPF_TCP_CLOSE != (int)TCP_CLOSE); 2980 BUILD_BUG_ON((int)BPF_TCP_CLOSE_WAIT != (int)TCP_CLOSE_WAIT); 2981 BUILD_BUG_ON((int)BPF_TCP_LAST_ACK != (int)TCP_LAST_ACK); 2982 BUILD_BUG_ON((int)BPF_TCP_LISTEN != (int)TCP_LISTEN); 2983 BUILD_BUG_ON((int)BPF_TCP_CLOSING != (int)TCP_CLOSING); 2984 BUILD_BUG_ON((int)BPF_TCP_NEW_SYN_RECV != (int)TCP_NEW_SYN_RECV); 2985 BUILD_BUG_ON((int)BPF_TCP_BOUND_INACTIVE != (int)TCP_BOUND_INACTIVE); 2986 BUILD_BUG_ON((int)BPF_TCP_MAX_STATES != (int)TCP_MAX_STATES); 2987 2988 /* bpf uapi header bpf.h defines an anonymous enum with values 2989 * BPF_TCP_* used by bpf programs. Currently gcc built vmlinux 2990 * is able to emit this enum in DWARF due to the above BUILD_BUG_ON. 2991 * But clang built vmlinux does not have this enum in DWARF 2992 * since clang removes the above code before generating IR/debuginfo. 2993 * Let us explicitly emit the type debuginfo to ensure the 2994 * above-mentioned anonymous enum in the vmlinux DWARF and hence BTF 2995 * regardless of which compiler is used. 2996 */ 2997 BTF_TYPE_EMIT_ENUM(BPF_TCP_ESTABLISHED); 2998 2999 if (BPF_SOCK_OPS_TEST_FLAG(tcp_sk(sk), BPF_SOCK_OPS_STATE_CB_FLAG)) 3000 tcp_call_bpf_2arg(sk, BPF_SOCK_OPS_STATE_CB, oldstate, state); 3001 3002 switch (state) { 3003 case TCP_ESTABLISHED: 3004 if (oldstate != TCP_ESTABLISHED) 3005 TCP_INC_STATS(sock_net(sk), TCP_MIB_CURRESTAB); 3006 break; 3007 case TCP_CLOSE_WAIT: 3008 if (oldstate == TCP_SYN_RECV) 3009 TCP_INC_STATS(sock_net(sk), TCP_MIB_CURRESTAB); 3010 break; 3011 3012 case TCP_CLOSE: 3013 if (oldstate == TCP_CLOSE_WAIT || oldstate == TCP_ESTABLISHED) 3014 TCP_INC_STATS(sock_net(sk), TCP_MIB_ESTABRESETS); 3015 3016 sk->sk_prot->unhash(sk); 3017 if (inet_csk(sk)->icsk_bind_hash && 3018 !(sk->sk_userlocks & SOCK_BINDPORT_LOCK)) 3019 inet_put_port(sk); 3020 fallthrough; 3021 default: 3022 if (oldstate == TCP_ESTABLISHED || oldstate == TCP_CLOSE_WAIT) 3023 TCP_DEC_STATS(sock_net(sk), TCP_MIB_CURRESTAB); 3024 } 3025 3026 /* Change state AFTER socket is unhashed to avoid closed 3027 * socket sitting in hash tables. 3028 */ 3029 inet_sk_state_store(sk, state); 3030 } 3031 3032 /* 3033 * State processing on a close. This implements the state shift for 3034 * sending our FIN frame. Note that we only send a FIN for some 3035 * states. A shutdown() may have already sent the FIN, or we may be 3036 * closed. 3037 */ 3038 3039 static const unsigned char new_state[16] = { 3040 /* current state: new state: action: */ 3041 [0 /* (Invalid) */] = TCP_CLOSE, 3042 [TCP_ESTABLISHED] = TCP_FIN_WAIT1 | TCP_ACTION_FIN, 3043 [TCP_SYN_SENT] = TCP_CLOSE, 3044 [TCP_SYN_RECV] = TCP_FIN_WAIT1 | TCP_ACTION_FIN, 3045 [TCP_FIN_WAIT1] = TCP_FIN_WAIT1, 3046 [TCP_FIN_WAIT2] = TCP_FIN_WAIT2, 3047 [TCP_TIME_WAIT] = TCP_CLOSE, 3048 [TCP_CLOSE] = TCP_CLOSE, 3049 [TCP_CLOSE_WAIT] = TCP_LAST_ACK | TCP_ACTION_FIN, 3050 [TCP_LAST_ACK] = TCP_LAST_ACK, 3051 [TCP_LISTEN] = TCP_CLOSE, 3052 [TCP_CLOSING] = TCP_CLOSING, 3053 [TCP_NEW_SYN_RECV] = TCP_CLOSE, /* should not happen ! */ 3054 }; 3055 3056 static int tcp_close_state(struct sock *sk) 3057 { 3058 int next = (int)new_state[sk->sk_state]; 3059 int ns = next & TCP_STATE_MASK; 3060 3061 tcp_set_state(sk, ns); 3062 3063 return next & TCP_ACTION_FIN; 3064 } 3065 3066 /* 3067 * Shutdown the sending side of a connection. Much like close except 3068 * that we don't receive shut down or sock_set_flag(sk, SOCK_DEAD). 3069 */ 3070 3071 void tcp_shutdown(struct sock *sk, int how) 3072 { 3073 /* We need to grab some memory, and put together a FIN, 3074 * and then put it into the queue to be sent. 3075 * Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92. 3076 */ 3077 if (!(how & SEND_SHUTDOWN)) 3078 return; 3079 3080 /* If we've already sent a FIN, or it's a closed state, skip this. */ 3081 if ((1 << sk->sk_state) & 3082 (TCPF_ESTABLISHED | TCPF_SYN_SENT | 3083 TCPF_CLOSE_WAIT)) { 3084 /* Clear out any half completed packets. FIN if needed. */ 3085 if (tcp_close_state(sk)) 3086 tcp_send_fin(sk); 3087 } 3088 } 3089 3090 int tcp_orphan_count_sum(void) 3091 { 3092 int i, total = 0; 3093 3094 for_each_possible_cpu(i) 3095 total += per_cpu(tcp_orphan_count, i); 3096 3097 return max(total, 0); 3098 } 3099 3100 static int tcp_orphan_cache; 3101 static struct timer_list tcp_orphan_timer; 3102 #define TCP_ORPHAN_TIMER_PERIOD msecs_to_jiffies(100) 3103 3104 static void tcp_orphan_update(struct timer_list *unused) 3105 { 3106 WRITE_ONCE(tcp_orphan_cache, tcp_orphan_count_sum()); 3107 mod_timer(&tcp_orphan_timer, jiffies + TCP_ORPHAN_TIMER_PERIOD); 3108 } 3109 3110 static bool tcp_too_many_orphans(int shift) 3111 { 3112 return READ_ONCE(tcp_orphan_cache) << shift > 3113 READ_ONCE(sysctl_tcp_max_orphans); 3114 } 3115 3116 static bool tcp_out_of_memory(const struct sock *sk) 3117 { 3118 if (sk->sk_wmem_queued > SOCK_MIN_SNDBUF && 3119 sk_memory_allocated(sk) > sk_prot_mem_limits(sk, 2)) 3120 return true; 3121 return false; 3122 } 3123 3124 bool tcp_check_oom(const struct sock *sk, int shift) 3125 { 3126 bool too_many_orphans, out_of_socket_memory; 3127 3128 too_many_orphans = tcp_too_many_orphans(shift); 3129 out_of_socket_memory = tcp_out_of_memory(sk); 3130 3131 if (too_many_orphans) 3132 net_info_ratelimited("too many orphaned sockets\n"); 3133 if (out_of_socket_memory) 3134 net_info_ratelimited("out of memory -- consider tuning tcp_mem\n"); 3135 return too_many_orphans || out_of_socket_memory; 3136 } 3137 3138 void __tcp_close(struct sock *sk, long timeout) 3139 { 3140 bool data_was_unread = false; 3141 struct sk_buff *skb; 3142 int state; 3143 3144 WRITE_ONCE(sk->sk_shutdown, SHUTDOWN_MASK); 3145 3146 if (sk->sk_state == TCP_LISTEN) { 3147 tcp_set_state(sk, TCP_CLOSE); 3148 3149 /* Special case. */ 3150 inet_csk_listen_stop(sk); 3151 3152 goto adjudge_to_death; 3153 } 3154 3155 /* We need to flush the recv. buffs. We do this only on the 3156 * descriptor close, not protocol-sourced closes, because the 3157 * reader process may not have drained the data yet! 3158 */ 3159 while ((skb = skb_peek(&sk->sk_receive_queue)) != NULL) { 3160 u32 end_seq = TCP_SKB_CB(skb)->end_seq; 3161 3162 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) 3163 end_seq--; 3164 if (after(end_seq, tcp_sk(sk)->copied_seq)) 3165 data_was_unread = true; 3166 tcp_eat_recv_skb(sk, skb); 3167 } 3168 3169 /* If socket has been already reset (e.g. in tcp_reset()) - kill it. */ 3170 if (sk->sk_state == TCP_CLOSE) 3171 goto adjudge_to_death; 3172 3173 /* As outlined in RFC 2525, section 2.17, we send a RST here because 3174 * data was lost. To witness the awful effects of the old behavior of 3175 * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk 3176 * GET in an FTP client, suspend the process, wait for the client to 3177 * advertise a zero window, then kill -9 the FTP client, wheee... 3178 * Note: timeout is always zero in such a case. 3179 */ 3180 if (unlikely(tcp_sk(sk)->repair)) { 3181 sk->sk_prot->disconnect(sk, 0); 3182 } else if (data_was_unread) { 3183 /* Unread data was tossed, zap the connection. */ 3184 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONCLOSE); 3185 tcp_set_state(sk, TCP_CLOSE); 3186 tcp_send_active_reset(sk, SK_RST_REASON_TCP_ABORT_ON_CLOSE); 3187 } else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) { 3188 /* Check zero linger _after_ checking for unread data. */ 3189 sk->sk_prot->disconnect(sk, 0); 3190 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA); 3191 } else if (tcp_close_state(sk)) { 3192 /* We FIN if the application ate all the data before 3193 * zapping the connection. 3194 */ 3195 3196 /* RED-PEN. Formally speaking, we have broken TCP state 3197 * machine. State transitions: 3198 * 3199 * TCP_ESTABLISHED -> TCP_FIN_WAIT1 3200 * TCP_SYN_RECV -> TCP_FIN_WAIT1 (it is difficult) 3201 * TCP_CLOSE_WAIT -> TCP_LAST_ACK 3202 * 3203 * are legal only when FIN has been sent (i.e. in window), 3204 * rather than queued out of window. Purists blame. 3205 * 3206 * F.e. "RFC state" is ESTABLISHED, 3207 * if Linux state is FIN-WAIT-1, but FIN is still not sent. 3208 * 3209 * The visible declinations are that sometimes 3210 * we enter time-wait state, when it is not required really 3211 * (harmless), do not send active resets, when they are 3212 * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when 3213 * they look as CLOSING or LAST_ACK for Linux) 3214 * Probably, I missed some more holelets. 3215 * --ANK 3216 * XXX (TFO) - To start off we don't support SYN+ACK+FIN 3217 * in a single packet! (May consider it later but will 3218 * probably need API support or TCP_CORK SYN-ACK until 3219 * data is written and socket is closed.) 3220 */ 3221 tcp_send_fin(sk); 3222 } 3223 3224 sk_stream_wait_close(sk, timeout); 3225 3226 adjudge_to_death: 3227 state = sk->sk_state; 3228 sock_hold(sk); 3229 sock_orphan(sk); 3230 3231 local_bh_disable(); 3232 bh_lock_sock(sk); 3233 /* remove backlog if any, without releasing ownership. */ 3234 __release_sock(sk); 3235 3236 tcp_orphan_count_inc(); 3237 3238 /* Have we already been destroyed by a softirq or backlog? */ 3239 if (state != TCP_CLOSE && sk->sk_state == TCP_CLOSE) 3240 goto out; 3241 3242 /* This is a (useful) BSD violating of the RFC. There is a 3243 * problem with TCP as specified in that the other end could 3244 * keep a socket open forever with no application left this end. 3245 * We use a 1 minute timeout (about the same as BSD) then kill 3246 * our end. If they send after that then tough - BUT: long enough 3247 * that we won't make the old 4*rto = almost no time - whoops 3248 * reset mistake. 3249 * 3250 * Nope, it was not mistake. It is really desired behaviour 3251 * f.e. on http servers, when such sockets are useless, but 3252 * consume significant resources. Let's do it with special 3253 * linger2 option. --ANK 3254 */ 3255 3256 if (sk->sk_state == TCP_FIN_WAIT2) { 3257 struct tcp_sock *tp = tcp_sk(sk); 3258 if (READ_ONCE(tp->linger2) < 0) { 3259 tcp_set_state(sk, TCP_CLOSE); 3260 tcp_send_active_reset(sk, 3261 SK_RST_REASON_TCP_ABORT_ON_LINGER); 3262 __NET_INC_STATS(sock_net(sk), 3263 LINUX_MIB_TCPABORTONLINGER); 3264 } else { 3265 const int tmo = tcp_fin_time(sk); 3266 3267 if (tmo > TCP_TIMEWAIT_LEN) { 3268 tcp_reset_keepalive_timer(sk, 3269 tmo - TCP_TIMEWAIT_LEN); 3270 } else { 3271 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo); 3272 goto out; 3273 } 3274 } 3275 } 3276 if (sk->sk_state != TCP_CLOSE) { 3277 if (tcp_check_oom(sk, 0)) { 3278 tcp_set_state(sk, TCP_CLOSE); 3279 tcp_send_active_reset(sk, 3280 SK_RST_REASON_TCP_ABORT_ON_MEMORY); 3281 __NET_INC_STATS(sock_net(sk), 3282 LINUX_MIB_TCPABORTONMEMORY); 3283 } else if (!check_net(sock_net(sk))) { 3284 /* Not possible to send reset; just close */ 3285 tcp_set_state(sk, TCP_CLOSE); 3286 } 3287 } 3288 3289 if (sk->sk_state == TCP_CLOSE) { 3290 struct request_sock *req; 3291 3292 req = rcu_dereference_protected(tcp_sk(sk)->fastopen_rsk, 3293 lockdep_sock_is_held(sk)); 3294 /* We could get here with a non-NULL req if the socket is 3295 * aborted (e.g., closed with unread data) before 3WHS 3296 * finishes. 3297 */ 3298 if (req) 3299 reqsk_fastopen_remove(sk, req, false); 3300 inet_csk_destroy_sock(sk); 3301 } 3302 /* Otherwise, socket is reprieved until protocol close. */ 3303 3304 out: 3305 bh_unlock_sock(sk); 3306 local_bh_enable(); 3307 } 3308 3309 void tcp_close(struct sock *sk, long timeout) 3310 { 3311 lock_sock(sk); 3312 __tcp_close(sk, timeout); 3313 release_sock(sk); 3314 if (!sk->sk_net_refcnt) 3315 inet_csk_clear_xmit_timers_sync(sk); 3316 sock_put(sk); 3317 } 3318 EXPORT_SYMBOL(tcp_close); 3319 3320 /* These states need RST on ABORT according to RFC793 */ 3321 3322 static inline bool tcp_need_reset(int state) 3323 { 3324 return (1 << state) & 3325 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 | 3326 TCPF_FIN_WAIT2 | TCPF_SYN_RECV); 3327 } 3328 3329 static void tcp_rtx_queue_purge(struct sock *sk) 3330 { 3331 struct rb_node *p = rb_first(&sk->tcp_rtx_queue); 3332 3333 tcp_sk(sk)->highest_sack = NULL; 3334 while (p) { 3335 struct sk_buff *skb = rb_to_skb(p); 3336 3337 p = rb_next(p); 3338 /* Since we are deleting whole queue, no need to 3339 * list_del(&skb->tcp_tsorted_anchor) 3340 */ 3341 tcp_rtx_queue_unlink(skb, sk); 3342 tcp_wmem_free_skb(sk, skb); 3343 } 3344 } 3345 3346 void tcp_write_queue_purge(struct sock *sk) 3347 { 3348 struct sk_buff *skb; 3349 3350 tcp_chrono_stop(sk, TCP_CHRONO_BUSY); 3351 while ((skb = __skb_dequeue(&sk->sk_write_queue)) != NULL) { 3352 tcp_skb_tsorted_anchor_cleanup(skb); 3353 tcp_wmem_free_skb(sk, skb); 3354 } 3355 tcp_rtx_queue_purge(sk); 3356 INIT_LIST_HEAD(&tcp_sk(sk)->tsorted_sent_queue); 3357 tcp_clear_all_retrans_hints(tcp_sk(sk)); 3358 tcp_sk(sk)->packets_out = 0; 3359 inet_csk(sk)->icsk_backoff = 0; 3360 } 3361 3362 int tcp_disconnect(struct sock *sk, int flags) 3363 { 3364 struct inet_sock *inet = inet_sk(sk); 3365 struct inet_connection_sock *icsk = inet_csk(sk); 3366 struct tcp_sock *tp = tcp_sk(sk); 3367 int old_state = sk->sk_state; 3368 struct request_sock *req; 3369 u32 seq; 3370 3371 if (old_state != TCP_CLOSE) 3372 tcp_set_state(sk, TCP_CLOSE); 3373 3374 /* ABORT function of RFC793 */ 3375 if (old_state == TCP_LISTEN) { 3376 inet_csk_listen_stop(sk); 3377 } else if (unlikely(tp->repair)) { 3378 WRITE_ONCE(sk->sk_err, ECONNABORTED); 3379 } else if (tcp_need_reset(old_state)) { 3380 tcp_send_active_reset(sk, SK_RST_REASON_TCP_STATE); 3381 WRITE_ONCE(sk->sk_err, ECONNRESET); 3382 } else if (tp->snd_nxt != tp->write_seq && 3383 (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK)) { 3384 /* The last check adjusts for discrepancy of Linux wrt. RFC 3385 * states 3386 */ 3387 tcp_send_active_reset(sk, 3388 SK_RST_REASON_TCP_DISCONNECT_WITH_DATA); 3389 WRITE_ONCE(sk->sk_err, ECONNRESET); 3390 } else if (old_state == TCP_SYN_SENT) 3391 WRITE_ONCE(sk->sk_err, ECONNRESET); 3392 3393 tcp_clear_xmit_timers(sk); 3394 __skb_queue_purge(&sk->sk_receive_queue); 3395 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt); 3396 WRITE_ONCE(tp->urg_data, 0); 3397 sk_set_peek_off(sk, -1); 3398 tcp_write_queue_purge(sk); 3399 tcp_fastopen_active_disable_ofo_check(sk); 3400 skb_rbtree_purge(&tp->out_of_order_queue); 3401 3402 inet->inet_dport = 0; 3403 3404 inet_bhash2_reset_saddr(sk); 3405 3406 WRITE_ONCE(sk->sk_shutdown, 0); 3407 sock_reset_flag(sk, SOCK_DONE); 3408 tp->srtt_us = 0; 3409 tp->mdev_us = jiffies_to_usecs(TCP_TIMEOUT_INIT); 3410 tp->rcv_rtt_last_tsecr = 0; 3411 3412 seq = tp->write_seq + tp->max_window + 2; 3413 if (!seq) 3414 seq = 1; 3415 WRITE_ONCE(tp->write_seq, seq); 3416 3417 icsk->icsk_backoff = 0; 3418 WRITE_ONCE(icsk->icsk_probes_out, 0); 3419 icsk->icsk_probes_tstamp = 0; 3420 icsk->icsk_rto = TCP_TIMEOUT_INIT; 3421 WRITE_ONCE(icsk->icsk_rto_min, TCP_RTO_MIN); 3422 WRITE_ONCE(icsk->icsk_delack_max, TCP_DELACK_MAX); 3423 WRITE_ONCE(tp->snd_ssthresh, TCP_INFINITE_SSTHRESH); 3424 tcp_snd_cwnd_set(tp, TCP_INIT_CWND); 3425 tp->snd_cwnd_cnt = 0; 3426 tp->is_cwnd_limited = 0; 3427 tp->max_packets_out = 0; 3428 tp->window_clamp = 0; 3429 tp->delivered = 0; 3430 tp->delivered_ce = 0; 3431 tp->accecn_fail_mode = 0; 3432 tp->saw_accecn_opt = TCP_ACCECN_OPT_NOT_SEEN; 3433 tcp_accecn_init_counters(tp); 3434 tp->prev_ecnfield = 0; 3435 tp->accecn_opt_tstamp = 0; 3436 tp->pkts_acked_ewma = 0; 3437 if (icsk->icsk_ca_initialized && icsk->icsk_ca_ops->release) 3438 icsk->icsk_ca_ops->release(sk); 3439 memset(icsk->icsk_ca_priv, 0, sizeof(icsk->icsk_ca_priv)); 3440 icsk->icsk_ca_initialized = 0; 3441 tcp_set_ca_state(sk, TCP_CA_Open); 3442 tp->is_sack_reneg = 0; 3443 tcp_clear_retrans(tp); 3444 tp->total_retrans = 0; 3445 inet_csk_delack_init(sk); 3446 /* Initialize rcv_mss to TCP_MIN_MSS to avoid division by 0 3447 * issue in __tcp_select_window() 3448 */ 3449 icsk->icsk_ack.rcv_mss = TCP_MIN_MSS; 3450 memset(&tp->rx_opt, 0, sizeof(tp->rx_opt)); 3451 __sk_dst_reset(sk); 3452 dst_release(unrcu_pointer(xchg(&sk->sk_rx_dst, NULL))); 3453 tcp_saved_syn_free(tp); 3454 tp->compressed_ack = 0; 3455 tp->segs_in = 0; 3456 tp->segs_out = 0; 3457 tp->bytes_sent = 0; 3458 tp->bytes_acked = 0; 3459 tp->bytes_received = 0; 3460 tp->bytes_retrans = 0; 3461 tp->data_segs_in = 0; 3462 tp->data_segs_out = 0; 3463 tp->duplicate_sack[0].start_seq = 0; 3464 tp->duplicate_sack[0].end_seq = 0; 3465 tp->dsack_dups = 0; 3466 tp->reord_seen = 0; 3467 tp->retrans_out = 0; 3468 tp->sacked_out = 0; 3469 tp->tlp_high_seq = 0; 3470 tp->last_oow_ack_time = 0; 3471 tp->plb_rehash = 0; 3472 /* There's a bubble in the pipe until at least the first ACK. */ 3473 tp->app_limited = ~0U; 3474 tp->rate_app_limited = 1; 3475 tp->rack.mstamp = 0; 3476 tp->rack.advanced = 0; 3477 tp->rack.reo_wnd_steps = 1; 3478 tp->rack.last_delivered = 0; 3479 tp->rack.reo_wnd_persist = 0; 3480 tp->rack.dsack_seen = 0; 3481 tp->syn_data_acked = 0; 3482 tp->syn_fastopen_child = 0; 3483 tp->rx_opt.saw_tstamp = 0; 3484 tp->rx_opt.dsack = 0; 3485 tp->rx_opt.num_sacks = 0; 3486 tp->rcv_ooopack = 0; 3487 3488 3489 /* Clean up fastopen related fields */ 3490 req = rcu_dereference_protected(tp->fastopen_rsk, 3491 lockdep_sock_is_held(sk)); 3492 if (req) 3493 reqsk_fastopen_remove(sk, req, false); 3494 tcp_free_fastopen_req(tp); 3495 inet_clear_bit(DEFER_CONNECT, sk); 3496 tp->fastopen_client_fail = 0; 3497 3498 WARN_ON(inet->inet_num && !icsk->icsk_bind_hash); 3499 3500 if (sk->sk_frag.page) { 3501 put_page(sk->sk_frag.page); 3502 sk->sk_frag.page = NULL; 3503 sk->sk_frag.offset = 0; 3504 } 3505 sk_error_report(sk); 3506 return 0; 3507 } 3508 3509 static inline bool tcp_can_repair_sock(const struct sock *sk) 3510 { 3511 return sockopt_ns_capable(sock_net(sk)->user_ns, CAP_NET_ADMIN) && 3512 (sk->sk_state != TCP_LISTEN); 3513 } 3514 3515 static int tcp_repair_set_window(struct tcp_sock *tp, sockptr_t optbuf, int len) 3516 { 3517 struct tcp_repair_window opt; 3518 3519 if (!tp->repair) 3520 return -EPERM; 3521 3522 if (len != sizeof(opt)) 3523 return -EINVAL; 3524 3525 if (copy_from_sockptr(&opt, optbuf, sizeof(opt))) 3526 return -EFAULT; 3527 3528 if (opt.max_window < opt.snd_wnd) 3529 return -EINVAL; 3530 3531 if (after(opt.snd_wl1, tp->rcv_nxt + opt.rcv_wnd)) 3532 return -EINVAL; 3533 3534 if (after(opt.rcv_wup, tp->rcv_nxt)) 3535 return -EINVAL; 3536 3537 tp->snd_wl1 = opt.snd_wl1; 3538 tp->snd_wnd = opt.snd_wnd; 3539 tp->max_window = opt.max_window; 3540 3541 tp->rcv_wnd = opt.rcv_wnd; 3542 tp->rcv_wup = opt.rcv_wup; 3543 tp->rcv_mwnd_seq = opt.rcv_wup + opt.rcv_wnd; 3544 3545 return 0; 3546 } 3547 3548 static int tcp_repair_options_est(struct sock *sk, sockptr_t optbuf, 3549 unsigned int len) 3550 { 3551 struct tcp_sock *tp = tcp_sk(sk); 3552 struct tcp_repair_opt opt; 3553 size_t offset = 0; 3554 3555 while (len >= sizeof(opt)) { 3556 if (copy_from_sockptr_offset(&opt, optbuf, offset, sizeof(opt))) 3557 return -EFAULT; 3558 3559 offset += sizeof(opt); 3560 len -= sizeof(opt); 3561 3562 switch (opt.opt_code) { 3563 case TCPOPT_MSS: 3564 tp->rx_opt.mss_clamp = opt.opt_val; 3565 tcp_mtup_init(sk); 3566 break; 3567 case TCPOPT_WINDOW: 3568 { 3569 u16 snd_wscale = opt.opt_val & 0xFFFF; 3570 u16 rcv_wscale = opt.opt_val >> 16; 3571 3572 if (snd_wscale > TCP_MAX_WSCALE || rcv_wscale > TCP_MAX_WSCALE) 3573 return -EFBIG; 3574 3575 tp->rx_opt.snd_wscale = snd_wscale; 3576 tp->rx_opt.rcv_wscale = rcv_wscale; 3577 tp->rx_opt.wscale_ok = 1; 3578 } 3579 break; 3580 case TCPOPT_SACK_PERM: 3581 if (opt.opt_val != 0) 3582 return -EINVAL; 3583 3584 tp->rx_opt.sack_ok |= TCP_SACK_SEEN; 3585 break; 3586 case TCPOPT_TIMESTAMP: 3587 if (opt.opt_val != 0) 3588 return -EINVAL; 3589 3590 tp->rx_opt.tstamp_ok = 1; 3591 break; 3592 } 3593 } 3594 3595 return 0; 3596 } 3597 3598 DEFINE_STATIC_KEY_FALSE(tcp_tx_delay_enabled); 3599 3600 static void tcp_enable_tx_delay(struct sock *sk, int val) 3601 { 3602 struct tcp_sock *tp = tcp_sk(sk); 3603 s32 delta = (val - tp->tcp_tx_delay) << 3; 3604 3605 if (val && !static_branch_unlikely(&tcp_tx_delay_enabled)) { 3606 static int __tcp_tx_delay_enabled = 0; 3607 3608 if (cmpxchg(&__tcp_tx_delay_enabled, 0, 1) == 0) { 3609 static_branch_enable(&tcp_tx_delay_enabled); 3610 pr_info("TCP_TX_DELAY enabled\n"); 3611 } 3612 } 3613 /* If we change tcp_tx_delay on a live flow, adjust tp->srtt_us, 3614 * tp->rtt_min, icsk_rto and sk->sk_pacing_rate. 3615 * This is best effort. 3616 */ 3617 if (delta && sk->sk_state == TCP_ESTABLISHED) { 3618 s64 srtt = (s64)tp->srtt_us + delta; 3619 3620 WRITE_ONCE(tp->srtt_us, 3621 clamp_t(s64, srtt, 1, ~0U)); 3622 3623 /* Note: does not deal with non zero icsk_backoff */ 3624 tcp_set_rto(sk); 3625 3626 minmax_reset(&tp->rtt_min, tcp_jiffies32, ~0U); 3627 3628 tcp_update_pacing_rate(sk); 3629 } 3630 } 3631 3632 /* When set indicates to always queue non-full frames. Later the user clears 3633 * this option and we transmit any pending partial frames in the queue. This is 3634 * meant to be used alongside sendfile() to get properly filled frames when the 3635 * user (for example) must write out headers with a write() call first and then 3636 * use sendfile to send out the data parts. 3637 * 3638 * TCP_CORK can be set together with TCP_NODELAY and it is stronger than 3639 * TCP_NODELAY. 3640 */ 3641 void __tcp_sock_set_cork(struct sock *sk, bool on) 3642 { 3643 struct tcp_sock *tp = tcp_sk(sk); 3644 3645 if (on) { 3646 tp->nonagle |= TCP_NAGLE_CORK; 3647 } else { 3648 tp->nonagle &= ~TCP_NAGLE_CORK; 3649 if (tp->nonagle & TCP_NAGLE_OFF) 3650 tp->nonagle |= TCP_NAGLE_PUSH; 3651 tcp_push_pending_frames(sk); 3652 } 3653 } 3654 3655 void tcp_sock_set_cork(struct sock *sk, bool on) 3656 { 3657 lock_sock(sk); 3658 __tcp_sock_set_cork(sk, on); 3659 release_sock(sk); 3660 } 3661 EXPORT_SYMBOL(tcp_sock_set_cork); 3662 3663 /* TCP_NODELAY is weaker than TCP_CORK, so that this option on corked socket is 3664 * remembered, but it is not activated until cork is cleared. 3665 * 3666 * However, when TCP_NODELAY is set we make an explicit push, which overrides 3667 * even TCP_CORK for currently queued segments. 3668 */ 3669 void __tcp_sock_set_nodelay(struct sock *sk, bool on) 3670 { 3671 if (on) { 3672 tcp_sk(sk)->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH; 3673 tcp_push_pending_frames(sk); 3674 } else { 3675 tcp_sk(sk)->nonagle &= ~TCP_NAGLE_OFF; 3676 } 3677 } 3678 3679 void tcp_sock_set_nodelay(struct sock *sk) 3680 { 3681 lock_sock(sk); 3682 __tcp_sock_set_nodelay(sk, true); 3683 release_sock(sk); 3684 } 3685 EXPORT_SYMBOL(tcp_sock_set_nodelay); 3686 3687 static void __tcp_sock_set_quickack(struct sock *sk, int val) 3688 { 3689 if (!val) { 3690 inet_csk_enter_pingpong_mode(sk); 3691 return; 3692 } 3693 3694 inet_csk_exit_pingpong_mode(sk); 3695 if ((1 << sk->sk_state) & (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) && 3696 inet_csk_ack_scheduled(sk)) { 3697 inet_csk(sk)->icsk_ack.pending |= ICSK_ACK_PUSHED; 3698 tcp_cleanup_rbuf(sk, 1); 3699 if (!(val & 1)) 3700 inet_csk_enter_pingpong_mode(sk); 3701 } 3702 } 3703 3704 void tcp_sock_set_quickack(struct sock *sk, int val) 3705 { 3706 lock_sock(sk); 3707 __tcp_sock_set_quickack(sk, val); 3708 release_sock(sk); 3709 } 3710 EXPORT_SYMBOL(tcp_sock_set_quickack); 3711 3712 int tcp_sock_set_syncnt(struct sock *sk, int val) 3713 { 3714 if (val < 1 || val > MAX_TCP_SYNCNT) 3715 return -EINVAL; 3716 3717 WRITE_ONCE(inet_csk(sk)->icsk_syn_retries, val); 3718 return 0; 3719 } 3720 EXPORT_SYMBOL(tcp_sock_set_syncnt); 3721 3722 int tcp_sock_set_user_timeout(struct sock *sk, int val) 3723 { 3724 /* Cap the max time in ms TCP will retry or probe the window 3725 * before giving up and aborting (ETIMEDOUT) a connection. 3726 */ 3727 if (val < 0) 3728 return -EINVAL; 3729 3730 WRITE_ONCE(inet_csk(sk)->icsk_user_timeout, val); 3731 return 0; 3732 } 3733 EXPORT_SYMBOL(tcp_sock_set_user_timeout); 3734 3735 int tcp_sock_set_keepidle_locked(struct sock *sk, int val) 3736 { 3737 struct tcp_sock *tp = tcp_sk(sk); 3738 3739 if (val < 1 || val > MAX_TCP_KEEPIDLE) 3740 return -EINVAL; 3741 3742 /* Paired with WRITE_ONCE() in keepalive_time_when() */ 3743 WRITE_ONCE(tp->keepalive_time, val * HZ); 3744 if (sock_flag(sk, SOCK_KEEPOPEN) && 3745 !((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN))) { 3746 u32 elapsed = keepalive_time_elapsed(tp); 3747 3748 if (tp->keepalive_time > elapsed) 3749 elapsed = tp->keepalive_time - elapsed; 3750 else 3751 elapsed = 0; 3752 tcp_reset_keepalive_timer(sk, elapsed); 3753 } 3754 3755 return 0; 3756 } 3757 3758 int tcp_sock_set_keepidle(struct sock *sk, int val) 3759 { 3760 int err; 3761 3762 lock_sock(sk); 3763 err = tcp_sock_set_keepidle_locked(sk, val); 3764 release_sock(sk); 3765 return err; 3766 } 3767 EXPORT_SYMBOL(tcp_sock_set_keepidle); 3768 3769 int tcp_sock_set_keepintvl(struct sock *sk, int val) 3770 { 3771 if (val < 1 || val > MAX_TCP_KEEPINTVL) 3772 return -EINVAL; 3773 3774 WRITE_ONCE(tcp_sk(sk)->keepalive_intvl, val * HZ); 3775 return 0; 3776 } 3777 EXPORT_SYMBOL(tcp_sock_set_keepintvl); 3778 3779 int tcp_sock_set_keepcnt(struct sock *sk, int val) 3780 { 3781 if (val < 1 || val > MAX_TCP_KEEPCNT) 3782 return -EINVAL; 3783 3784 /* Paired with READ_ONCE() in keepalive_probes() */ 3785 WRITE_ONCE(tcp_sk(sk)->keepalive_probes, val); 3786 return 0; 3787 } 3788 EXPORT_SYMBOL(tcp_sock_set_keepcnt); 3789 3790 int tcp_set_window_clamp(struct sock *sk, int val) 3791 { 3792 u32 old_window_clamp, new_window_clamp, new_rcv_ssthresh; 3793 struct tcp_sock *tp = tcp_sk(sk); 3794 3795 if (!val) { 3796 if (sk->sk_state != TCP_CLOSE) 3797 return -EINVAL; 3798 WRITE_ONCE(tp->window_clamp, 0); 3799 return 0; 3800 } 3801 3802 old_window_clamp = tp->window_clamp; 3803 new_window_clamp = max_t(int, SOCK_MIN_RCVBUF / 2, val); 3804 3805 if (new_window_clamp == old_window_clamp) 3806 return 0; 3807 3808 WRITE_ONCE(tp->window_clamp, new_window_clamp); 3809 3810 /* Need to apply the reserved mem provisioning only 3811 * when shrinking the window clamp. 3812 */ 3813 if (new_window_clamp < old_window_clamp) { 3814 __tcp_adjust_rcv_ssthresh(sk, new_window_clamp); 3815 } else { 3816 new_rcv_ssthresh = min(tp->rcv_wnd, new_window_clamp); 3817 tp->rcv_ssthresh = max(new_rcv_ssthresh, tp->rcv_ssthresh); 3818 } 3819 return 0; 3820 } 3821 3822 int tcp_sock_set_maxseg(struct sock *sk, int val) 3823 { 3824 /* Values greater than interface MTU won't take effect. However 3825 * at the point when this call is done we typically don't yet 3826 * know which interface is going to be used 3827 */ 3828 if (val && (val < TCP_MIN_MSS || val > MAX_TCP_WINDOW)) 3829 return -EINVAL; 3830 3831 WRITE_ONCE(tcp_sk(sk)->rx_opt.user_mss, val); 3832 return 0; 3833 } 3834 3835 /* 3836 * Socket option code for TCP. 3837 */ 3838 int do_tcp_setsockopt(struct sock *sk, int level, int optname, 3839 sockptr_t optval, unsigned int optlen) 3840 { 3841 struct tcp_sock *tp = tcp_sk(sk); 3842 struct inet_connection_sock *icsk = inet_csk(sk); 3843 struct net *net = sock_net(sk); 3844 int val; 3845 int err = 0; 3846 3847 /* These are data/string values, all the others are ints */ 3848 switch (optname) { 3849 case TCP_CONGESTION: { 3850 char name[TCP_CA_NAME_MAX]; 3851 3852 if (optlen < 1) 3853 return -EINVAL; 3854 3855 val = strncpy_from_sockptr(name, optval, 3856 min_t(long, TCP_CA_NAME_MAX-1, optlen)); 3857 if (val < 0) 3858 return -EFAULT; 3859 name[val] = 0; 3860 3861 sockopt_lock_sock(sk); 3862 err = tcp_set_congestion_control(sk, name, !has_current_bpf_ctx(), 3863 sockopt_ns_capable(sock_net(sk)->user_ns, 3864 CAP_NET_ADMIN)); 3865 sockopt_release_sock(sk); 3866 return err; 3867 } 3868 case TCP_ULP: { 3869 char name[TCP_ULP_NAME_MAX]; 3870 3871 if (optlen < 1) 3872 return -EINVAL; 3873 3874 val = strncpy_from_sockptr(name, optval, 3875 min_t(long, TCP_ULP_NAME_MAX - 1, 3876 optlen)); 3877 if (val < 0) 3878 return -EFAULT; 3879 name[val] = 0; 3880 3881 sockopt_lock_sock(sk); 3882 err = tcp_set_ulp(sk, name); 3883 sockopt_release_sock(sk); 3884 return err; 3885 } 3886 case TCP_FASTOPEN_KEY: { 3887 __u8 key[TCP_FASTOPEN_KEY_BUF_LENGTH]; 3888 __u8 *backup_key = NULL; 3889 3890 /* Allow a backup key as well to facilitate key rotation 3891 * First key is the active one. 3892 */ 3893 if (optlen != TCP_FASTOPEN_KEY_LENGTH && 3894 optlen != TCP_FASTOPEN_KEY_BUF_LENGTH) 3895 return -EINVAL; 3896 3897 if (copy_from_sockptr(key, optval, optlen)) 3898 return -EFAULT; 3899 3900 if (optlen == TCP_FASTOPEN_KEY_BUF_LENGTH) 3901 backup_key = key + TCP_FASTOPEN_KEY_LENGTH; 3902 3903 return tcp_fastopen_reset_cipher(net, sk, key, backup_key); 3904 } 3905 default: 3906 /* fallthru */ 3907 break; 3908 } 3909 3910 if (optlen < sizeof(int)) 3911 return -EINVAL; 3912 3913 if (copy_from_sockptr(&val, optval, sizeof(val))) 3914 return -EFAULT; 3915 3916 /* Handle options that can be set without locking the socket. */ 3917 switch (optname) { 3918 case TCP_SYNCNT: 3919 return tcp_sock_set_syncnt(sk, val); 3920 case TCP_USER_TIMEOUT: 3921 return tcp_sock_set_user_timeout(sk, val); 3922 case TCP_KEEPINTVL: 3923 return tcp_sock_set_keepintvl(sk, val); 3924 case TCP_KEEPCNT: 3925 return tcp_sock_set_keepcnt(sk, val); 3926 case TCP_LINGER2: 3927 if (val < 0) 3928 WRITE_ONCE(tp->linger2, -1); 3929 else if (val > TCP_FIN_TIMEOUT_MAX / HZ) 3930 WRITE_ONCE(tp->linger2, TCP_FIN_TIMEOUT_MAX); 3931 else 3932 WRITE_ONCE(tp->linger2, val * HZ); 3933 return 0; 3934 case TCP_DEFER_ACCEPT: 3935 /* Translate value in seconds to number of retransmits */ 3936 WRITE_ONCE(icsk->icsk_accept_queue.rskq_defer_accept, 3937 secs_to_retrans(val, TCP_TIMEOUT_INIT / HZ, 3938 TCP_RTO_MAX / HZ)); 3939 return 0; 3940 case TCP_RTO_MAX_MS: 3941 if (val < MSEC_PER_SEC || val > TCP_RTO_MAX_SEC * MSEC_PER_SEC) 3942 return -EINVAL; 3943 WRITE_ONCE(inet_csk(sk)->icsk_rto_max, msecs_to_jiffies(val)); 3944 return 0; 3945 case TCP_RTO_MIN_US: { 3946 int rto_min = usecs_to_jiffies(val); 3947 3948 if (rto_min > TCP_RTO_MIN || rto_min < TCP_TIMEOUT_MIN) 3949 return -EINVAL; 3950 WRITE_ONCE(inet_csk(sk)->icsk_rto_min, rto_min); 3951 return 0; 3952 } 3953 case TCP_DELACK_MAX_US: { 3954 int delack_max = usecs_to_jiffies(val); 3955 3956 if (delack_max > TCP_DELACK_MAX || delack_max < TCP_TIMEOUT_MIN) 3957 return -EINVAL; 3958 WRITE_ONCE(inet_csk(sk)->icsk_delack_max, delack_max); 3959 return 0; 3960 } 3961 case TCP_MAXSEG: 3962 return tcp_sock_set_maxseg(sk, val); 3963 } 3964 3965 sockopt_lock_sock(sk); 3966 3967 switch (optname) { 3968 case TCP_NODELAY: 3969 __tcp_sock_set_nodelay(sk, val); 3970 break; 3971 3972 case TCP_THIN_LINEAR_TIMEOUTS: 3973 if (val < 0 || val > 1) 3974 err = -EINVAL; 3975 else 3976 tp->thin_lto = val; 3977 break; 3978 3979 case TCP_THIN_DUPACK: 3980 if (val < 0 || val > 1) 3981 err = -EINVAL; 3982 break; 3983 3984 case TCP_REPAIR: 3985 if (!tcp_can_repair_sock(sk)) 3986 err = -EPERM; 3987 else if (val == TCP_REPAIR_ON) { 3988 tp->repair = 1; 3989 sk->sk_reuse = SK_FORCE_REUSE; 3990 tp->repair_queue = TCP_NO_QUEUE; 3991 } else if (val == TCP_REPAIR_OFF) { 3992 tp->repair = 0; 3993 sk->sk_reuse = SK_NO_REUSE; 3994 tcp_send_window_probe(sk); 3995 } else if (val == TCP_REPAIR_OFF_NO_WP) { 3996 tp->repair = 0; 3997 sk->sk_reuse = SK_NO_REUSE; 3998 } else 3999 err = -EINVAL; 4000 4001 break; 4002 4003 case TCP_REPAIR_QUEUE: 4004 if (!tp->repair) 4005 err = -EPERM; 4006 else if ((unsigned int)val < TCP_QUEUES_NR) 4007 tp->repair_queue = val; 4008 else 4009 err = -EINVAL; 4010 break; 4011 4012 case TCP_QUEUE_SEQ: 4013 if (sk->sk_state != TCP_CLOSE) { 4014 err = -EPERM; 4015 } else if (tp->repair_queue == TCP_SEND_QUEUE) { 4016 if (!tcp_rtx_queue_empty(sk)) 4017 err = -EPERM; 4018 else 4019 WRITE_ONCE(tp->write_seq, val); 4020 } else if (tp->repair_queue == TCP_RECV_QUEUE) { 4021 if (tp->rcv_nxt != tp->copied_seq) { 4022 err = -EPERM; 4023 } else { 4024 WRITE_ONCE(tp->rcv_nxt, val); 4025 WRITE_ONCE(tp->copied_seq, val); 4026 } 4027 } else { 4028 err = -EINVAL; 4029 } 4030 break; 4031 4032 case TCP_REPAIR_OPTIONS: 4033 if (!tp->repair) 4034 err = -EINVAL; 4035 else if (sk->sk_state == TCP_ESTABLISHED && !tp->bytes_sent) 4036 err = tcp_repair_options_est(sk, optval, optlen); 4037 else 4038 err = -EPERM; 4039 break; 4040 4041 case TCP_CORK: 4042 __tcp_sock_set_cork(sk, val); 4043 break; 4044 4045 case TCP_KEEPIDLE: 4046 err = tcp_sock_set_keepidle_locked(sk, val); 4047 break; 4048 case TCP_SAVE_SYN: 4049 /* 0: disable, 1: enable, 2: start from ether_header */ 4050 if (val < 0 || val > 2) 4051 err = -EINVAL; 4052 else 4053 tp->save_syn = val; 4054 break; 4055 4056 case TCP_WINDOW_CLAMP: 4057 err = tcp_set_window_clamp(sk, val); 4058 break; 4059 4060 case TCP_QUICKACK: 4061 __tcp_sock_set_quickack(sk, val); 4062 break; 4063 4064 case TCP_AO_REPAIR: 4065 if (!tcp_can_repair_sock(sk)) { 4066 err = -EPERM; 4067 break; 4068 } 4069 err = tcp_ao_set_repair(sk, optval, optlen); 4070 break; 4071 #ifdef CONFIG_TCP_AO 4072 case TCP_AO_ADD_KEY: 4073 case TCP_AO_DEL_KEY: 4074 case TCP_AO_INFO: { 4075 /* If this is the first TCP-AO setsockopt() on the socket, 4076 * sk_state has to be LISTEN or CLOSE. Allow TCP_REPAIR 4077 * in any state. 4078 */ 4079 if ((1 << sk->sk_state) & (TCPF_LISTEN | TCPF_CLOSE)) 4080 goto ao_parse; 4081 if (rcu_dereference_protected(tcp_sk(sk)->ao_info, 4082 lockdep_sock_is_held(sk))) 4083 goto ao_parse; 4084 if (tp->repair) 4085 goto ao_parse; 4086 err = -EISCONN; 4087 break; 4088 ao_parse: 4089 err = tp->af_specific->ao_parse(sk, optname, optval, optlen); 4090 break; 4091 } 4092 #endif 4093 #ifdef CONFIG_TCP_MD5SIG 4094 case TCP_MD5SIG: 4095 case TCP_MD5SIG_EXT: 4096 err = tp->af_specific->md5_parse(sk, optname, optval, optlen); 4097 break; 4098 #endif 4099 case TCP_FASTOPEN: 4100 if (val >= 0 && ((1 << sk->sk_state) & (TCPF_CLOSE | 4101 TCPF_LISTEN))) { 4102 tcp_fastopen_init_key_once(net); 4103 4104 fastopen_queue_tune(sk, val); 4105 } else { 4106 err = -EINVAL; 4107 } 4108 break; 4109 case TCP_FASTOPEN_CONNECT: 4110 if (val > 1 || val < 0) { 4111 err = -EINVAL; 4112 } else if (READ_ONCE(net->ipv4.sysctl_tcp_fastopen) & 4113 TFO_CLIENT_ENABLE) { 4114 if (sk->sk_state == TCP_CLOSE) 4115 tp->fastopen_connect = val; 4116 else 4117 err = -EINVAL; 4118 } else { 4119 err = -EOPNOTSUPP; 4120 } 4121 break; 4122 case TCP_FASTOPEN_NO_COOKIE: 4123 if (val > 1 || val < 0) 4124 err = -EINVAL; 4125 else if (!((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN))) 4126 err = -EINVAL; 4127 else 4128 tp->fastopen_no_cookie = val; 4129 break; 4130 case TCP_TIMESTAMP: 4131 if (!tp->repair) { 4132 err = -EPERM; 4133 break; 4134 } 4135 /* val is an opaque field, 4136 * and low order bit contains usec_ts enable bit. 4137 * Its a best effort, and we do not care if user makes an error. 4138 */ 4139 tp->tcp_usec_ts = val & 1; 4140 WRITE_ONCE(tp->tsoffset, val - tcp_clock_ts(tp->tcp_usec_ts)); 4141 break; 4142 case TCP_REPAIR_WINDOW: 4143 err = tcp_repair_set_window(tp, optval, optlen); 4144 break; 4145 case TCP_NOTSENT_LOWAT: 4146 WRITE_ONCE(tp->notsent_lowat, val); 4147 READ_ONCE(sk->sk_write_space)(sk); 4148 break; 4149 case TCP_INQ: 4150 if (val > 1 || val < 0) 4151 err = -EINVAL; 4152 else 4153 tp->recvmsg_inq = val; 4154 break; 4155 case TCP_TX_DELAY: 4156 /* tp->srtt_us is u32, and is shifted by 3 */ 4157 if (val < 0 || val >= (1U << (31 - 3))) { 4158 err = -EINVAL; 4159 break; 4160 } 4161 tcp_enable_tx_delay(sk, val); 4162 WRITE_ONCE(tp->tcp_tx_delay, val); 4163 break; 4164 default: 4165 err = -ENOPROTOOPT; 4166 break; 4167 } 4168 4169 sockopt_release_sock(sk); 4170 return err; 4171 } 4172 4173 int tcp_setsockopt(struct sock *sk, int level, int optname, sockptr_t optval, 4174 unsigned int optlen) 4175 { 4176 const struct inet_connection_sock *icsk = inet_csk(sk); 4177 4178 if (level != SOL_TCP) 4179 /* Paired with WRITE_ONCE() in do_ipv6_setsockopt() and tcp_v6_connect() */ 4180 return READ_ONCE(icsk->icsk_af_ops)->setsockopt(sk, level, optname, 4181 optval, optlen); 4182 return do_tcp_setsockopt(sk, level, optname, optval, optlen); 4183 } 4184 4185 static void tcp_get_info_chrono_stats(const struct tcp_sock *tp, 4186 struct tcp_info *info) 4187 { 4188 u64 stats[__TCP_CHRONO_MAX], total = 0; 4189 enum tcp_chrono i, cur; 4190 4191 /* Following READ_ONCE()s pair with WRITE_ONCE()s in tcp_chrono_set(). 4192 * This is because socket lock might not be owned by us at this point. 4193 * This is best effort, tcp_get_timestamping_opt_stats() can 4194 * see wrong values. A real fix would be too costly for TCP fast path. 4195 */ 4196 cur = READ_ONCE(tp->chrono_type); 4197 for (i = TCP_CHRONO_BUSY; i < __TCP_CHRONO_MAX; ++i) { 4198 stats[i] = READ_ONCE(tp->chrono_stat[i - 1]); 4199 if (i == cur) 4200 stats[i] += tcp_jiffies32 - READ_ONCE(tp->chrono_start); 4201 stats[i] *= USEC_PER_SEC / HZ; 4202 total += stats[i]; 4203 } 4204 4205 info->tcpi_busy_time = total; 4206 info->tcpi_rwnd_limited = stats[TCP_CHRONO_RWND_LIMITED]; 4207 info->tcpi_sndbuf_limited = stats[TCP_CHRONO_SNDBUF_LIMITED]; 4208 } 4209 4210 /* Return information about state of tcp endpoint in API format. */ 4211 void tcp_get_info(struct sock *sk, struct tcp_info *info) 4212 { 4213 const struct tcp_sock *tp = tcp_sk(sk); /* iff sk_type == SOCK_STREAM */ 4214 const struct inet_connection_sock *icsk = inet_csk(sk); 4215 const u8 ect1_idx = INET_ECN_ECT_1 - 1; 4216 const u8 ect0_idx = INET_ECN_ECT_0 - 1; 4217 const u8 ce_idx = INET_ECN_CE - 1; 4218 unsigned long rate; 4219 u32 now; 4220 u64 rate64; 4221 bool slow; 4222 4223 memset(info, 0, sizeof(*info)); 4224 if (sk->sk_type != SOCK_STREAM) 4225 return; 4226 4227 info->tcpi_state = inet_sk_state_load(sk); 4228 4229 /* Report meaningful fields for all TCP states, including listeners */ 4230 rate = READ_ONCE(sk->sk_pacing_rate); 4231 rate64 = (rate != ~0UL) ? rate : ~0ULL; 4232 info->tcpi_pacing_rate = rate64; 4233 4234 rate = READ_ONCE(sk->sk_max_pacing_rate); 4235 rate64 = (rate != ~0UL) ? rate : ~0ULL; 4236 info->tcpi_max_pacing_rate = rate64; 4237 4238 info->tcpi_reordering = tp->reordering; 4239 info->tcpi_snd_cwnd = tcp_snd_cwnd(tp); 4240 4241 if (info->tcpi_state == TCP_LISTEN) { 4242 /* listeners aliased fields : 4243 * tcpi_unacked -> Number of children ready for accept() 4244 * tcpi_sacked -> max backlog 4245 */ 4246 info->tcpi_unacked = READ_ONCE(sk->sk_ack_backlog); 4247 info->tcpi_sacked = READ_ONCE(sk->sk_max_ack_backlog); 4248 return; 4249 } 4250 4251 slow = lock_sock_fast(sk); 4252 4253 info->tcpi_ca_state = icsk->icsk_ca_state; 4254 info->tcpi_retransmits = icsk->icsk_retransmits; 4255 info->tcpi_probes = icsk->icsk_probes_out; 4256 info->tcpi_backoff = icsk->icsk_backoff; 4257 4258 if (tp->rx_opt.tstamp_ok) 4259 info->tcpi_options |= TCPI_OPT_TIMESTAMPS; 4260 if (tcp_is_sack(tp)) 4261 info->tcpi_options |= TCPI_OPT_SACK; 4262 if (tp->rx_opt.wscale_ok) { 4263 info->tcpi_options |= TCPI_OPT_WSCALE; 4264 info->tcpi_snd_wscale = tp->rx_opt.snd_wscale; 4265 info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale; 4266 } 4267 4268 if (tcp_ecn_mode_any(tp)) 4269 info->tcpi_options |= TCPI_OPT_ECN; 4270 if (tp->ecn_flags & TCP_ECN_SEEN) 4271 info->tcpi_options |= TCPI_OPT_ECN_SEEN; 4272 if (tp->syn_data_acked) 4273 info->tcpi_options |= TCPI_OPT_SYN_DATA; 4274 if (tp->tcp_usec_ts) 4275 info->tcpi_options |= TCPI_OPT_USEC_TS; 4276 if (tp->syn_fastopen_child) 4277 info->tcpi_options |= TCPI_OPT_TFO_CHILD; 4278 4279 info->tcpi_rto = jiffies_to_usecs(icsk->icsk_rto); 4280 info->tcpi_ato = jiffies_to_usecs(min_t(u32, icsk->icsk_ack.ato, 4281 tcp_delack_max(sk))); 4282 info->tcpi_snd_mss = tp->mss_cache; 4283 info->tcpi_rcv_mss = icsk->icsk_ack.rcv_mss; 4284 4285 info->tcpi_unacked = tp->packets_out; 4286 info->tcpi_sacked = tp->sacked_out; 4287 4288 info->tcpi_lost = tp->lost_out; 4289 info->tcpi_retrans = tp->retrans_out; 4290 4291 now = tcp_jiffies32; 4292 info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime); 4293 info->tcpi_last_data_recv = jiffies_to_msecs(now - icsk->icsk_ack.lrcvtime); 4294 info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp); 4295 4296 info->tcpi_pmtu = icsk->icsk_pmtu_cookie; 4297 info->tcpi_rcv_ssthresh = tp->rcv_ssthresh; 4298 info->tcpi_rtt = tp->srtt_us >> 3; 4299 info->tcpi_rttvar = tp->mdev_us >> 2; 4300 info->tcpi_snd_ssthresh = tp->snd_ssthresh; 4301 info->tcpi_advmss = tp->advmss; 4302 4303 info->tcpi_rcv_rtt = tp->rcv_rtt_est.rtt_us >> 3; 4304 info->tcpi_rcv_space = tp->rcvq_space.space; 4305 4306 info->tcpi_total_retrans = tp->total_retrans; 4307 4308 info->tcpi_bytes_acked = tp->bytes_acked; 4309 info->tcpi_bytes_received = tp->bytes_received; 4310 info->tcpi_notsent_bytes = max_t(int, 0, tp->write_seq - tp->snd_nxt); 4311 tcp_get_info_chrono_stats(tp, info); 4312 4313 info->tcpi_segs_out = tp->segs_out; 4314 4315 /* segs_in and data_segs_in can be updated from tcp_segs_in() from BH */ 4316 info->tcpi_segs_in = READ_ONCE(tp->segs_in); 4317 info->tcpi_data_segs_in = READ_ONCE(tp->data_segs_in); 4318 4319 info->tcpi_min_rtt = tcp_min_rtt(tp); 4320 info->tcpi_data_segs_out = tp->data_segs_out; 4321 4322 info->tcpi_delivery_rate_app_limited = tp->rate_app_limited ? 1 : 0; 4323 rate64 = tcp_compute_delivery_rate(tp); 4324 if (rate64) 4325 info->tcpi_delivery_rate = rate64; 4326 info->tcpi_delivered = tp->delivered; 4327 info->tcpi_delivered_ce = tp->delivered_ce; 4328 info->tcpi_bytes_sent = tp->bytes_sent; 4329 info->tcpi_bytes_retrans = tp->bytes_retrans; 4330 info->tcpi_dsack_dups = tp->dsack_dups; 4331 info->tcpi_reord_seen = tp->reord_seen; 4332 info->tcpi_rcv_ooopack = tp->rcv_ooopack; 4333 info->tcpi_snd_wnd = tp->snd_wnd; 4334 info->tcpi_rcv_wnd = tp->rcv_wnd; 4335 info->tcpi_rehash = tp->plb_rehash + tp->timeout_rehash; 4336 info->tcpi_fastopen_client_fail = tp->fastopen_client_fail; 4337 4338 info->tcpi_total_rto = tp->total_rto; 4339 info->tcpi_total_rto_recoveries = tp->total_rto_recoveries; 4340 info->tcpi_total_rto_time = tp->total_rto_time; 4341 if (tp->rto_stamp) 4342 info->tcpi_total_rto_time += tcp_clock_ms() - tp->rto_stamp; 4343 4344 if (tcp_ecn_disabled(tp)) 4345 info->tcpi_ecn_mode = TCPI_ECN_MODE_DISABLED; 4346 else if (tcp_ecn_mode_rfc3168(tp)) 4347 info->tcpi_ecn_mode = TCPI_ECN_MODE_RFC3168; 4348 else if (tcp_ecn_mode_accecn(tp)) 4349 info->tcpi_ecn_mode = TCPI_ECN_MODE_ACCECN; 4350 else if (tcp_ecn_mode_pending(tp)) 4351 info->tcpi_ecn_mode = TCPI_ECN_MODE_PENDING; 4352 info->tcpi_accecn_fail_mode = tp->accecn_fail_mode; 4353 info->tcpi_accecn_opt_seen = tp->saw_accecn_opt; 4354 info->tcpi_received_ce = tp->received_ce; 4355 info->tcpi_delivered_e1_bytes = tp->delivered_ecn_bytes[ect1_idx]; 4356 info->tcpi_delivered_e0_bytes = tp->delivered_ecn_bytes[ect0_idx]; 4357 info->tcpi_delivered_ce_bytes = tp->delivered_ecn_bytes[ce_idx]; 4358 info->tcpi_received_e1_bytes = tp->received_ecn_bytes[ect1_idx]; 4359 info->tcpi_received_e0_bytes = tp->received_ecn_bytes[ect0_idx]; 4360 info->tcpi_received_ce_bytes = tp->received_ecn_bytes[ce_idx]; 4361 4362 unlock_sock_fast(sk, slow); 4363 } 4364 EXPORT_SYMBOL_GPL(tcp_get_info); 4365 4366 static size_t tcp_opt_stats_get_size(void) 4367 { 4368 return 4369 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_BUSY */ 4370 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_RWND_LIMITED */ 4371 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_SNDBUF_LIMITED */ 4372 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_DATA_SEGS_OUT */ 4373 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_TOTAL_RETRANS */ 4374 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_PACING_RATE */ 4375 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_DELIVERY_RATE */ 4376 nla_total_size(sizeof(u32)) + /* TCP_NLA_SND_CWND */ 4377 nla_total_size(sizeof(u32)) + /* TCP_NLA_REORDERING */ 4378 nla_total_size(sizeof(u32)) + /* TCP_NLA_MIN_RTT */ 4379 nla_total_size(sizeof(u8)) + /* TCP_NLA_RECUR_RETRANS */ 4380 nla_total_size(sizeof(u8)) + /* TCP_NLA_DELIVERY_RATE_APP_LMT */ 4381 nla_total_size(sizeof(u32)) + /* TCP_NLA_SNDQ_SIZE */ 4382 nla_total_size(sizeof(u8)) + /* TCP_NLA_CA_STATE */ 4383 nla_total_size(sizeof(u32)) + /* TCP_NLA_SND_SSTHRESH */ 4384 nla_total_size(sizeof(u32)) + /* TCP_NLA_DELIVERED */ 4385 nla_total_size(sizeof(u32)) + /* TCP_NLA_DELIVERED_CE */ 4386 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_BYTES_SENT */ 4387 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_BYTES_RETRANS */ 4388 nla_total_size(sizeof(u32)) + /* TCP_NLA_DSACK_DUPS */ 4389 nla_total_size(sizeof(u32)) + /* TCP_NLA_REORD_SEEN */ 4390 nla_total_size(sizeof(u32)) + /* TCP_NLA_SRTT */ 4391 nla_total_size(sizeof(u16)) + /* TCP_NLA_TIMEOUT_REHASH */ 4392 nla_total_size(sizeof(u32)) + /* TCP_NLA_BYTES_NOTSENT */ 4393 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_EDT */ 4394 nla_total_size(sizeof(u8)) + /* TCP_NLA_TTL */ 4395 nla_total_size(sizeof(u32)) + /* TCP_NLA_REHASH */ 4396 0; 4397 } 4398 4399 /* Returns TTL or hop limit of an incoming packet from skb. */ 4400 static u8 tcp_skb_ttl_or_hop_limit(const struct sk_buff *skb) 4401 { 4402 if (skb->protocol == htons(ETH_P_IP)) 4403 return ip_hdr(skb)->ttl; 4404 else if (skb->protocol == htons(ETH_P_IPV6)) 4405 return ipv6_hdr(skb)->hop_limit; 4406 else 4407 return 0; 4408 } 4409 4410 struct sk_buff *tcp_get_timestamping_opt_stats(const struct sock *sk, 4411 const struct sk_buff *orig_skb, 4412 const struct sk_buff *ack_skb) 4413 { 4414 const struct tcp_sock *tp = tcp_sk(sk); 4415 struct sk_buff *stats; 4416 struct tcp_info info; 4417 unsigned long rate; 4418 u64 rate64; 4419 4420 stats = alloc_skb(tcp_opt_stats_get_size(), GFP_ATOMIC); 4421 if (!stats) 4422 return NULL; 4423 4424 tcp_get_info_chrono_stats(tp, &info); 4425 nla_put_u64_64bit(stats, TCP_NLA_BUSY, 4426 info.tcpi_busy_time, TCP_NLA_PAD); 4427 nla_put_u64_64bit(stats, TCP_NLA_RWND_LIMITED, 4428 info.tcpi_rwnd_limited, TCP_NLA_PAD); 4429 nla_put_u64_64bit(stats, TCP_NLA_SNDBUF_LIMITED, 4430 info.tcpi_sndbuf_limited, TCP_NLA_PAD); 4431 nla_put_u64_64bit(stats, TCP_NLA_DATA_SEGS_OUT, 4432 READ_ONCE(tp->data_segs_out), TCP_NLA_PAD); 4433 nla_put_u64_64bit(stats, TCP_NLA_TOTAL_RETRANS, 4434 READ_ONCE(tp->total_retrans), TCP_NLA_PAD); 4435 4436 rate = READ_ONCE(sk->sk_pacing_rate); 4437 rate64 = (rate != ~0UL) ? rate : ~0ULL; 4438 nla_put_u64_64bit(stats, TCP_NLA_PACING_RATE, rate64, TCP_NLA_PAD); 4439 4440 rate64 = tcp_compute_delivery_rate(tp); 4441 nla_put_u64_64bit(stats, TCP_NLA_DELIVERY_RATE, rate64, TCP_NLA_PAD); 4442 4443 nla_put_u32(stats, TCP_NLA_SND_CWND, READ_ONCE(tp->snd_cwnd)); 4444 nla_put_u32(stats, TCP_NLA_REORDERING, READ_ONCE(tp->reordering)); 4445 nla_put_u32(stats, TCP_NLA_MIN_RTT, data_race(tcp_min_rtt(tp))); 4446 4447 nla_put_u8(stats, TCP_NLA_RECUR_RETRANS, 4448 READ_ONCE(inet_csk(sk)->icsk_retransmits)); 4449 nla_put_u8(stats, TCP_NLA_DELIVERY_RATE_APP_LMT, data_race(!!tp->rate_app_limited)); 4450 nla_put_u32(stats, TCP_NLA_SND_SSTHRESH, READ_ONCE(tp->snd_ssthresh)); 4451 nla_put_u32(stats, TCP_NLA_DELIVERED, READ_ONCE(tp->delivered)); 4452 nla_put_u32(stats, TCP_NLA_DELIVERED_CE, READ_ONCE(tp->delivered_ce)); 4453 4454 nla_put_u32(stats, TCP_NLA_SNDQ_SIZE, 4455 max_t(int, 0, 4456 READ_ONCE(tp->write_seq) - READ_ONCE(tp->snd_una))); 4457 nla_put_u8(stats, TCP_NLA_CA_STATE, inet_csk(sk)->icsk_ca_state); 4458 4459 nla_put_u64_64bit(stats, TCP_NLA_BYTES_SENT, READ_ONCE(tp->bytes_sent), 4460 TCP_NLA_PAD); 4461 nla_put_u64_64bit(stats, TCP_NLA_BYTES_RETRANS, 4462 READ_ONCE(tp->bytes_retrans), TCP_NLA_PAD); 4463 nla_put_u32(stats, TCP_NLA_DSACK_DUPS, READ_ONCE(tp->dsack_dups)); 4464 nla_put_u32(stats, TCP_NLA_REORD_SEEN, READ_ONCE(tp->reord_seen)); 4465 nla_put_u32(stats, TCP_NLA_SRTT, READ_ONCE(tp->srtt_us) >> 3); 4466 nla_put_u16(stats, TCP_NLA_TIMEOUT_REHASH, 4467 READ_ONCE(tp->timeout_rehash)); 4468 nla_put_u32(stats, TCP_NLA_BYTES_NOTSENT, 4469 max_t(int, 0, 4470 READ_ONCE(tp->write_seq) - READ_ONCE(tp->snd_nxt))); 4471 nla_put_u64_64bit(stats, TCP_NLA_EDT, orig_skb->skb_mstamp_ns, 4472 TCP_NLA_PAD); 4473 if (ack_skb) 4474 nla_put_u8(stats, TCP_NLA_TTL, 4475 tcp_skb_ttl_or_hop_limit(ack_skb)); 4476 4477 nla_put_u32(stats, TCP_NLA_REHASH, 4478 READ_ONCE(tp->plb_rehash) + READ_ONCE(tp->timeout_rehash)); 4479 return stats; 4480 } 4481 4482 int do_tcp_getsockopt(struct sock *sk, int level, 4483 int optname, sockptr_t optval, sockptr_t optlen) 4484 { 4485 struct inet_connection_sock *icsk = inet_csk(sk); 4486 struct tcp_sock *tp = tcp_sk(sk); 4487 struct net *net = sock_net(sk); 4488 int user_mss; 4489 int val, len; 4490 4491 if (copy_from_sockptr(&len, optlen, sizeof(int))) 4492 return -EFAULT; 4493 4494 if (len < 0) 4495 return -EINVAL; 4496 4497 len = min_t(unsigned int, len, sizeof(int)); 4498 4499 switch (optname) { 4500 case TCP_MAXSEG: 4501 val = tp->mss_cache; 4502 user_mss = READ_ONCE(tp->rx_opt.user_mss); 4503 if (user_mss && 4504 ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN))) 4505 val = user_mss; 4506 if (tp->repair) 4507 val = tp->rx_opt.mss_clamp; 4508 break; 4509 case TCP_NODELAY: 4510 val = !!(tp->nonagle&TCP_NAGLE_OFF); 4511 break; 4512 case TCP_CORK: 4513 val = !!(tp->nonagle&TCP_NAGLE_CORK); 4514 break; 4515 case TCP_KEEPIDLE: 4516 val = keepalive_time_when(tp) / HZ; 4517 break; 4518 case TCP_KEEPINTVL: 4519 val = keepalive_intvl_when(tp) / HZ; 4520 break; 4521 case TCP_KEEPCNT: 4522 val = keepalive_probes(tp); 4523 break; 4524 case TCP_SYNCNT: 4525 val = READ_ONCE(icsk->icsk_syn_retries) ? : 4526 READ_ONCE(net->ipv4.sysctl_tcp_syn_retries); 4527 break; 4528 case TCP_LINGER2: 4529 val = READ_ONCE(tp->linger2); 4530 if (val >= 0) 4531 val = (val ? : READ_ONCE(net->ipv4.sysctl_tcp_fin_timeout)) / HZ; 4532 break; 4533 case TCP_DEFER_ACCEPT: 4534 val = READ_ONCE(icsk->icsk_accept_queue.rskq_defer_accept); 4535 val = retrans_to_secs(val, TCP_TIMEOUT_INIT / HZ, 4536 TCP_RTO_MAX / HZ); 4537 break; 4538 case TCP_WINDOW_CLAMP: 4539 val = READ_ONCE(tp->window_clamp); 4540 break; 4541 case TCP_INFO: { 4542 struct tcp_info info; 4543 4544 if (copy_from_sockptr(&len, optlen, sizeof(int))) 4545 return -EFAULT; 4546 4547 tcp_get_info(sk, &info); 4548 4549 len = min_t(unsigned int, len, sizeof(info)); 4550 if (copy_to_sockptr(optlen, &len, sizeof(int))) 4551 return -EFAULT; 4552 if (copy_to_sockptr(optval, &info, len)) 4553 return -EFAULT; 4554 return 0; 4555 } 4556 case TCP_CC_INFO: { 4557 const struct tcp_congestion_ops *ca_ops; 4558 union tcp_cc_info info; 4559 size_t sz = 0; 4560 int attr; 4561 4562 if (copy_from_sockptr(&len, optlen, sizeof(int))) 4563 return -EFAULT; 4564 4565 rcu_read_lock(); 4566 ca_ops = READ_ONCE(icsk->icsk_ca_ops); 4567 if (ca_ops && ca_ops->get_info) 4568 sz = ca_ops->get_info(sk, ~0U, &attr, &info); 4569 rcu_read_unlock(); 4570 4571 len = min_t(unsigned int, len, sz); 4572 if (copy_to_sockptr(optlen, &len, sizeof(int))) 4573 return -EFAULT; 4574 if (copy_to_sockptr(optval, &info, len)) 4575 return -EFAULT; 4576 return 0; 4577 } 4578 case TCP_QUICKACK: 4579 val = !inet_csk_in_pingpong_mode(sk); 4580 break; 4581 4582 case TCP_CONGESTION: { 4583 char ca_name[TCP_CA_NAME_MAX] = {}; 4584 4585 if (copy_from_sockptr(&len, optlen, sizeof(int))) 4586 return -EFAULT; 4587 len = min_t(unsigned int, len, TCP_CA_NAME_MAX); 4588 if (copy_to_sockptr(optlen, &len, sizeof(int))) 4589 return -EFAULT; 4590 4591 rcu_read_lock(); 4592 memcpy(ca_name, READ_ONCE(icsk->icsk_ca_ops)->name, 4593 sizeof(ca_name)); 4594 rcu_read_unlock(); 4595 4596 if (copy_to_sockptr(optval, ca_name, len)) 4597 return -EFAULT; 4598 return 0; 4599 } 4600 case TCP_ULP: 4601 if (copy_from_sockptr(&len, optlen, sizeof(int))) 4602 return -EFAULT; 4603 len = min_t(unsigned int, len, TCP_ULP_NAME_MAX); 4604 if (!icsk->icsk_ulp_ops) { 4605 len = 0; 4606 if (copy_to_sockptr(optlen, &len, sizeof(int))) 4607 return -EFAULT; 4608 return 0; 4609 } 4610 if (copy_to_sockptr(optlen, &len, sizeof(int))) 4611 return -EFAULT; 4612 if (copy_to_sockptr(optval, icsk->icsk_ulp_ops->name, len)) 4613 return -EFAULT; 4614 return 0; 4615 4616 case TCP_FASTOPEN_KEY: { 4617 u64 key[TCP_FASTOPEN_KEY_BUF_LENGTH / sizeof(u64)]; 4618 unsigned int key_len; 4619 4620 if (copy_from_sockptr(&len, optlen, sizeof(int))) 4621 return -EFAULT; 4622 4623 key_len = tcp_fastopen_get_cipher(net, icsk, key) * 4624 TCP_FASTOPEN_KEY_LENGTH; 4625 len = min_t(unsigned int, len, key_len); 4626 if (copy_to_sockptr(optlen, &len, sizeof(int))) 4627 return -EFAULT; 4628 if (copy_to_sockptr(optval, key, len)) 4629 return -EFAULT; 4630 return 0; 4631 } 4632 case TCP_THIN_LINEAR_TIMEOUTS: 4633 val = tp->thin_lto; 4634 break; 4635 4636 case TCP_THIN_DUPACK: 4637 val = 0; 4638 break; 4639 4640 case TCP_REPAIR: 4641 val = tp->repair; 4642 break; 4643 4644 case TCP_REPAIR_QUEUE: 4645 if (tp->repair) 4646 val = tp->repair_queue; 4647 else 4648 return -EINVAL; 4649 break; 4650 4651 case TCP_REPAIR_WINDOW: { 4652 struct tcp_repair_window opt; 4653 4654 if (copy_from_sockptr(&len, optlen, sizeof(int))) 4655 return -EFAULT; 4656 4657 if (len != sizeof(opt)) 4658 return -EINVAL; 4659 4660 if (!tp->repair) 4661 return -EPERM; 4662 4663 opt.snd_wl1 = tp->snd_wl1; 4664 opt.snd_wnd = tp->snd_wnd; 4665 opt.max_window = tp->max_window; 4666 opt.rcv_wnd = tp->rcv_wnd; 4667 opt.rcv_wup = tp->rcv_wup; 4668 4669 if (copy_to_sockptr(optval, &opt, len)) 4670 return -EFAULT; 4671 return 0; 4672 } 4673 case TCP_QUEUE_SEQ: 4674 if (tp->repair_queue == TCP_SEND_QUEUE) 4675 val = tp->write_seq; 4676 else if (tp->repair_queue == TCP_RECV_QUEUE) 4677 val = tp->rcv_nxt; 4678 else 4679 return -EINVAL; 4680 break; 4681 4682 case TCP_USER_TIMEOUT: 4683 val = READ_ONCE(icsk->icsk_user_timeout); 4684 break; 4685 4686 case TCP_FASTOPEN: 4687 val = READ_ONCE(icsk->icsk_accept_queue.fastopenq.max_qlen); 4688 break; 4689 4690 case TCP_FASTOPEN_CONNECT: 4691 val = tp->fastopen_connect; 4692 break; 4693 4694 case TCP_FASTOPEN_NO_COOKIE: 4695 val = tp->fastopen_no_cookie; 4696 break; 4697 4698 case TCP_TX_DELAY: 4699 val = READ_ONCE(tp->tcp_tx_delay); 4700 break; 4701 4702 case TCP_TIMESTAMP: 4703 val = tcp_clock_ts(tp->tcp_usec_ts) + READ_ONCE(tp->tsoffset); 4704 if (tp->tcp_usec_ts) 4705 val |= 1; 4706 else 4707 val &= ~1; 4708 break; 4709 case TCP_NOTSENT_LOWAT: 4710 val = READ_ONCE(tp->notsent_lowat); 4711 break; 4712 case TCP_INQ: 4713 val = tp->recvmsg_inq; 4714 break; 4715 case TCP_SAVE_SYN: 4716 val = tp->save_syn; 4717 break; 4718 case TCP_SAVED_SYN: { 4719 if (copy_from_sockptr(&len, optlen, sizeof(int))) 4720 return -EFAULT; 4721 4722 sockopt_lock_sock(sk); 4723 if (tp->saved_syn) { 4724 if (len < tcp_saved_syn_len(tp->saved_syn)) { 4725 len = tcp_saved_syn_len(tp->saved_syn); 4726 if (copy_to_sockptr(optlen, &len, sizeof(int))) { 4727 sockopt_release_sock(sk); 4728 return -EFAULT; 4729 } 4730 sockopt_release_sock(sk); 4731 return -EINVAL; 4732 } 4733 len = tcp_saved_syn_len(tp->saved_syn); 4734 if (copy_to_sockptr(optlen, &len, sizeof(int))) { 4735 sockopt_release_sock(sk); 4736 return -EFAULT; 4737 } 4738 if (copy_to_sockptr(optval, tp->saved_syn->data, len)) { 4739 sockopt_release_sock(sk); 4740 return -EFAULT; 4741 } 4742 tcp_saved_syn_free(tp); 4743 sockopt_release_sock(sk); 4744 } else { 4745 sockopt_release_sock(sk); 4746 len = 0; 4747 if (copy_to_sockptr(optlen, &len, sizeof(int))) 4748 return -EFAULT; 4749 } 4750 return 0; 4751 } 4752 #ifdef CONFIG_MMU 4753 case TCP_ZEROCOPY_RECEIVE: { 4754 struct scm_timestamping_internal tss; 4755 struct tcp_zerocopy_receive zc = {}; 4756 int err; 4757 4758 if (copy_from_sockptr(&len, optlen, sizeof(int))) 4759 return -EFAULT; 4760 if (len < 0 || 4761 len < offsetofend(struct tcp_zerocopy_receive, length)) 4762 return -EINVAL; 4763 if (unlikely(len > sizeof(zc))) { 4764 err = check_zeroed_sockptr(optval, sizeof(zc), 4765 len - sizeof(zc)); 4766 if (err < 1) 4767 return err == 0 ? -EINVAL : err; 4768 len = sizeof(zc); 4769 if (copy_to_sockptr(optlen, &len, sizeof(int))) 4770 return -EFAULT; 4771 } 4772 if (copy_from_sockptr(&zc, optval, len)) 4773 return -EFAULT; 4774 if (zc.reserved) 4775 return -EINVAL; 4776 if (zc.msg_flags & ~(TCP_VALID_ZC_MSG_FLAGS)) 4777 return -EINVAL; 4778 sockopt_lock_sock(sk); 4779 err = tcp_zerocopy_receive(sk, &zc, &tss); 4780 err = BPF_CGROUP_RUN_PROG_GETSOCKOPT_KERN(sk, level, optname, 4781 &zc, &len, err); 4782 sockopt_release_sock(sk); 4783 if (len >= offsetofend(struct tcp_zerocopy_receive, msg_flags)) 4784 goto zerocopy_rcv_cmsg; 4785 switch (len) { 4786 case offsetofend(struct tcp_zerocopy_receive, msg_flags): 4787 goto zerocopy_rcv_cmsg; 4788 case offsetofend(struct tcp_zerocopy_receive, msg_controllen): 4789 case offsetofend(struct tcp_zerocopy_receive, msg_control): 4790 case offsetofend(struct tcp_zerocopy_receive, flags): 4791 case offsetofend(struct tcp_zerocopy_receive, copybuf_len): 4792 case offsetofend(struct tcp_zerocopy_receive, copybuf_address): 4793 case offsetofend(struct tcp_zerocopy_receive, err): 4794 goto zerocopy_rcv_sk_err; 4795 case offsetofend(struct tcp_zerocopy_receive, inq): 4796 goto zerocopy_rcv_inq; 4797 case offsetofend(struct tcp_zerocopy_receive, length): 4798 default: 4799 goto zerocopy_rcv_out; 4800 } 4801 zerocopy_rcv_cmsg: 4802 if (zc.msg_flags & TCP_CMSG_TS) 4803 tcp_zc_finalize_rx_tstamp(sk, &zc, &tss); 4804 else 4805 zc.msg_flags = 0; 4806 zerocopy_rcv_sk_err: 4807 if (!err) 4808 zc.err = sock_error(sk); 4809 zerocopy_rcv_inq: 4810 zc.inq = tcp_inq_hint(sk); 4811 zerocopy_rcv_out: 4812 if (!err && copy_to_sockptr(optval, &zc, len)) 4813 err = -EFAULT; 4814 return err; 4815 } 4816 #endif 4817 case TCP_AO_REPAIR: 4818 if (!tcp_can_repair_sock(sk)) 4819 return -EPERM; 4820 return tcp_ao_get_repair(sk, optval, optlen); 4821 case TCP_AO_GET_KEYS: 4822 case TCP_AO_INFO: { 4823 int err; 4824 4825 sockopt_lock_sock(sk); 4826 if (optname == TCP_AO_GET_KEYS) 4827 err = tcp_ao_get_mkts(sk, optval, optlen); 4828 else 4829 err = tcp_ao_get_sock_info(sk, optval, optlen); 4830 sockopt_release_sock(sk); 4831 4832 return err; 4833 } 4834 case TCP_IS_MPTCP: 4835 val = 0; 4836 break; 4837 case TCP_RTO_MAX_MS: 4838 val = jiffies_to_msecs(tcp_rto_max(sk)); 4839 break; 4840 case TCP_RTO_MIN_US: 4841 val = jiffies_to_usecs(READ_ONCE(inet_csk(sk)->icsk_rto_min)); 4842 break; 4843 case TCP_DELACK_MAX_US: 4844 val = jiffies_to_usecs(READ_ONCE(inet_csk(sk)->icsk_delack_max)); 4845 break; 4846 default: 4847 return -ENOPROTOOPT; 4848 } 4849 4850 if (copy_to_sockptr(optlen, &len, sizeof(int))) 4851 return -EFAULT; 4852 if (copy_to_sockptr(optval, &val, len)) 4853 return -EFAULT; 4854 return 0; 4855 } 4856 4857 bool tcp_bpf_bypass_getsockopt(int level, int optname) 4858 { 4859 /* TCP do_tcp_getsockopt has optimized getsockopt implementation 4860 * to avoid extra socket lock for TCP_ZEROCOPY_RECEIVE. 4861 */ 4862 if (level == SOL_TCP && optname == TCP_ZEROCOPY_RECEIVE) 4863 return true; 4864 4865 return false; 4866 } 4867 4868 int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval, 4869 int __user *optlen) 4870 { 4871 struct inet_connection_sock *icsk = inet_csk(sk); 4872 4873 if (level != SOL_TCP) 4874 /* Paired with WRITE_ONCE() in do_ipv6_setsockopt() and tcp_v6_connect() */ 4875 return READ_ONCE(icsk->icsk_af_ops)->getsockopt(sk, level, optname, 4876 optval, optlen); 4877 return do_tcp_getsockopt(sk, level, optname, USER_SOCKPTR(optval), 4878 USER_SOCKPTR(optlen)); 4879 } 4880 4881 #ifdef CONFIG_TCP_MD5SIG 4882 void tcp_md5_hash_skb_data(struct md5_ctx *ctx, const struct sk_buff *skb, 4883 unsigned int header_len) 4884 { 4885 const unsigned int head_data_len = skb_headlen(skb) > header_len ? 4886 skb_headlen(skb) - header_len : 0; 4887 const struct skb_shared_info *shi = skb_shinfo(skb); 4888 struct sk_buff *frag_iter; 4889 unsigned int i; 4890 4891 md5_update(ctx, (const u8 *)tcp_hdr(skb) + header_len, head_data_len); 4892 4893 for (i = 0; i < shi->nr_frags; ++i) { 4894 const skb_frag_t *f = &shi->frags[i]; 4895 u32 p_off, p_len, copied; 4896 const void *vaddr; 4897 struct page *p; 4898 4899 skb_frag_foreach_page(f, skb_frag_off(f), skb_frag_size(f), 4900 p, p_off, p_len, copied) { 4901 vaddr = kmap_local_page(p); 4902 md5_update(ctx, vaddr + p_off, p_len); 4903 kunmap_local(vaddr); 4904 } 4905 } 4906 4907 skb_walk_frags(skb, frag_iter) 4908 tcp_md5_hash_skb_data(ctx, frag_iter, 0); 4909 } 4910 4911 void tcp_md5_hash_key(struct md5_ctx *ctx, 4912 const struct tcp_md5sig_key *key) 4913 { 4914 u8 keylen = READ_ONCE(key->keylen); /* paired with WRITE_ONCE() in tcp_md5_do_add */ 4915 4916 /* We use data_race() because tcp_md5_do_add() might change 4917 * key->key under us 4918 */ 4919 data_race(({ md5_update(ctx, key->key, keylen), 0; })); 4920 } 4921 4922 /* Called with rcu_read_lock() */ 4923 static enum skb_drop_reason 4924 tcp_inbound_md5_hash(const struct sock *sk, const struct sk_buff *skb, 4925 const void *saddr, const void *daddr, 4926 int family, int l3index, const __u8 *hash_location) 4927 { 4928 /* This gets called for each TCP segment that has TCP-MD5 option. 4929 * We have 2 drop cases: 4930 * o An MD5 signature is present, but we're not expecting one. 4931 * o The MD5 signature is wrong. 4932 */ 4933 const struct tcp_sock *tp = tcp_sk(sk); 4934 struct tcp_md5sig_key *key; 4935 u8 newhash[16]; 4936 4937 key = tcp_md5_do_lookup(sk, l3index, saddr, family); 4938 if (!key) { 4939 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMD5UNEXPECTED); 4940 trace_tcp_hash_md5_unexpected(sk, skb); 4941 return SKB_DROP_REASON_TCP_MD5UNEXPECTED; 4942 } 4943 4944 /* Check the signature. 4945 * To support dual stack listeners, we need to handle 4946 * IPv4-mapped case. 4947 */ 4948 if (family == AF_INET) 4949 tcp_v4_md5_hash_skb(newhash, key, NULL, skb); 4950 else 4951 tp->af_specific->calc_md5_hash(newhash, key, NULL, skb); 4952 if (crypto_memneq(hash_location, newhash, 16)) { 4953 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMD5FAILURE); 4954 trace_tcp_hash_md5_mismatch(sk, skb); 4955 return SKB_DROP_REASON_TCP_MD5FAILURE; 4956 } 4957 return SKB_NOT_DROPPED_YET; 4958 } 4959 #else 4960 static inline enum skb_drop_reason 4961 tcp_inbound_md5_hash(const struct sock *sk, const struct sk_buff *skb, 4962 const void *saddr, const void *daddr, 4963 int family, int l3index, const __u8 *hash_location) 4964 { 4965 return SKB_NOT_DROPPED_YET; 4966 } 4967 4968 #endif 4969 4970 #if defined(CONFIG_TCP_MD5SIG) || defined(CONFIG_TCP_AO) 4971 /* 4972 * Parse Signature options 4973 */ 4974 int tcp_do_parse_auth_options(const struct tcphdr *th, 4975 const u8 **md5_hash, const u8 **ao_hash) 4976 { 4977 int length = (th->doff << 2) - sizeof(*th); 4978 const u8 *ptr = (const u8 *)(th + 1); 4979 unsigned int minlen = TCPOLEN_MD5SIG; 4980 4981 if (IS_ENABLED(CONFIG_TCP_AO)) 4982 minlen = sizeof(struct tcp_ao_hdr) + 1; 4983 4984 *md5_hash = NULL; 4985 *ao_hash = NULL; 4986 4987 /* If not enough data remaining, we can short cut */ 4988 while (length >= minlen) { 4989 int opcode = *ptr++; 4990 int opsize; 4991 4992 switch (opcode) { 4993 case TCPOPT_EOL: 4994 return 0; 4995 case TCPOPT_NOP: 4996 length--; 4997 continue; 4998 default: 4999 opsize = *ptr++; 5000 if (opsize < 2 || opsize > length) 5001 return -EINVAL; 5002 if (opcode == TCPOPT_MD5SIG) { 5003 if (opsize != TCPOLEN_MD5SIG) 5004 return -EINVAL; 5005 if (unlikely(*md5_hash || *ao_hash)) 5006 return -EEXIST; 5007 *md5_hash = ptr; 5008 } else if (opcode == TCPOPT_AO) { 5009 if (opsize <= sizeof(struct tcp_ao_hdr)) 5010 return -EINVAL; 5011 if (unlikely(*md5_hash || *ao_hash)) 5012 return -EEXIST; 5013 *ao_hash = ptr; 5014 } 5015 } 5016 ptr += opsize - 2; 5017 length -= opsize; 5018 } 5019 return 0; 5020 } 5021 #endif 5022 5023 /* Called with rcu_read_lock() */ 5024 enum skb_drop_reason 5025 tcp_inbound_hash(struct sock *sk, const struct request_sock *req, 5026 const struct sk_buff *skb, 5027 const void *saddr, const void *daddr, 5028 int family, int dif, int sdif) 5029 { 5030 const struct tcphdr *th = tcp_hdr(skb); 5031 const struct tcp_ao_hdr *aoh; 5032 const __u8 *md5_location; 5033 int l3index; 5034 5035 /* Invalid option or two times meet any of auth options */ 5036 if (tcp_parse_auth_options(th, &md5_location, &aoh)) { 5037 trace_tcp_hash_bad_header(sk, skb); 5038 return SKB_DROP_REASON_TCP_AUTH_HDR; 5039 } 5040 5041 if (req) { 5042 if (tcp_rsk_used_ao(req) != !!aoh) { 5043 u8 keyid, rnext, maclen; 5044 5045 if (aoh) { 5046 keyid = aoh->keyid; 5047 rnext = aoh->rnext_keyid; 5048 maclen = tcp_ao_hdr_maclen(aoh); 5049 } else { 5050 keyid = rnext = maclen = 0; 5051 } 5052 5053 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPAOBAD); 5054 trace_tcp_ao_handshake_failure(sk, skb, keyid, rnext, maclen); 5055 return SKB_DROP_REASON_TCP_AOFAILURE; 5056 } 5057 } 5058 5059 /* sdif set, means packet ingressed via a device 5060 * in an L3 domain and dif is set to the l3mdev 5061 */ 5062 l3index = sdif ? dif : 0; 5063 5064 /* Fast path: unsigned segments */ 5065 if (likely(!md5_location && !aoh)) { 5066 /* Drop if there's TCP-MD5 or TCP-AO key with any rcvid/sndid 5067 * for the remote peer. On TCP-AO established connection 5068 * the last key is impossible to remove, so there's 5069 * always at least one current_key. 5070 */ 5071 if (tcp_ao_required(sk, saddr, family, l3index, true)) { 5072 trace_tcp_hash_ao_required(sk, skb); 5073 return SKB_DROP_REASON_TCP_AONOTFOUND; 5074 } 5075 if (unlikely(tcp_md5_do_lookup(sk, l3index, saddr, family))) { 5076 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMD5NOTFOUND); 5077 trace_tcp_hash_md5_required(sk, skb); 5078 return SKB_DROP_REASON_TCP_MD5NOTFOUND; 5079 } 5080 return SKB_NOT_DROPPED_YET; 5081 } 5082 5083 if (aoh) 5084 return tcp_inbound_ao_hash(sk, skb, family, req, l3index, aoh); 5085 5086 return tcp_inbound_md5_hash(sk, skb, saddr, daddr, family, 5087 l3index, md5_location); 5088 } 5089 5090 void tcp_done(struct sock *sk) 5091 { 5092 struct request_sock *req; 5093 5094 /* We might be called with a new socket, after 5095 * inet_csk_prepare_forced_close() has been called 5096 * so we can not use lockdep_sock_is_held(sk) 5097 */ 5098 req = rcu_dereference_protected(tcp_sk(sk)->fastopen_rsk, 1); 5099 5100 if (sk->sk_state == TCP_SYN_SENT || sk->sk_state == TCP_SYN_RECV) 5101 TCP_INC_STATS(sock_net(sk), TCP_MIB_ATTEMPTFAILS); 5102 5103 tcp_set_state(sk, TCP_CLOSE); 5104 tcp_clear_xmit_timers(sk); 5105 if (req) 5106 reqsk_fastopen_remove(sk, req, false); 5107 5108 WRITE_ONCE(sk->sk_shutdown, SHUTDOWN_MASK); 5109 5110 if (!sock_flag(sk, SOCK_DEAD)) 5111 sk->sk_state_change(sk); 5112 else 5113 inet_csk_destroy_sock(sk); 5114 } 5115 5116 int tcp_abort(struct sock *sk, int err) 5117 { 5118 int state = inet_sk_state_load(sk); 5119 5120 if (state == TCP_NEW_SYN_RECV) { 5121 struct request_sock *req = inet_reqsk(sk); 5122 5123 local_bh_disable(); 5124 inet_csk_reqsk_queue_drop(req->rsk_listener, req); 5125 local_bh_enable(); 5126 return 0; 5127 } 5128 if (state == TCP_TIME_WAIT) { 5129 struct inet_timewait_sock *tw = inet_twsk(sk); 5130 5131 refcount_inc(&tw->tw_refcnt); 5132 local_bh_disable(); 5133 inet_twsk_deschedule_put(tw); 5134 local_bh_enable(); 5135 return 0; 5136 } 5137 5138 /* BPF context ensures sock locking. */ 5139 if (!has_current_bpf_ctx()) 5140 /* Don't race with userspace socket closes such as tcp_close. */ 5141 lock_sock(sk); 5142 5143 /* Avoid closing the same socket twice. */ 5144 if (sk->sk_state == TCP_CLOSE) { 5145 if (!has_current_bpf_ctx()) 5146 release_sock(sk); 5147 return -ENOENT; 5148 } 5149 5150 if (sk->sk_state == TCP_LISTEN) { 5151 tcp_set_state(sk, TCP_CLOSE); 5152 inet_csk_listen_stop(sk); 5153 } 5154 5155 /* Don't race with BH socket closes such as inet_csk_listen_stop. */ 5156 local_bh_disable(); 5157 bh_lock_sock(sk); 5158 5159 if (tcp_need_reset(sk->sk_state)) 5160 tcp_send_active_reset(sk, SK_RST_REASON_TCP_STATE); 5161 tcp_done_with_error(sk, err); 5162 5163 bh_unlock_sock(sk); 5164 local_bh_enable(); 5165 if (!has_current_bpf_ctx()) 5166 release_sock(sk); 5167 return 0; 5168 } 5169 EXPORT_SYMBOL_GPL(tcp_abort); 5170 5171 extern struct tcp_congestion_ops tcp_reno; 5172 5173 static __initdata unsigned long thash_entries; 5174 static int __init set_thash_entries(char *str) 5175 { 5176 ssize_t ret; 5177 5178 if (!str) 5179 return 0; 5180 5181 ret = kstrtoul(str, 0, &thash_entries); 5182 if (ret) 5183 return 0; 5184 5185 return 1; 5186 } 5187 __setup("thash_entries=", set_thash_entries); 5188 5189 static void __init tcp_init_mem(void) 5190 { 5191 unsigned long limit = nr_free_buffer_pages() / 16; 5192 5193 limit = max(limit, 128UL); 5194 sysctl_tcp_mem[0] = limit / 4 * 3; /* 4.68 % */ 5195 sysctl_tcp_mem[1] = limit; /* 6.25 % */ 5196 sysctl_tcp_mem[2] = sysctl_tcp_mem[0] * 2; /* 9.37 % */ 5197 } 5198 5199 static void __init tcp_struct_check(void) 5200 { 5201 /* TX read-mostly hotpath cache lines */ 5202 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_tx, max_window); 5203 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_tx, rcv_ssthresh); 5204 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_tx, reordering); 5205 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_tx, notsent_lowat); 5206 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_tx, gso_segs); 5207 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_tx, retransmit_skb_hint); 5208 #if IS_ENABLED(CONFIG_TLS_DEVICE) 5209 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_tx, tcp_clean_acked); 5210 #endif 5211 5212 /* TXRX read-mostly hotpath cache lines */ 5213 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_txrx, tsoffset); 5214 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_txrx, snd_wnd); 5215 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_txrx, mss_cache); 5216 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_txrx, snd_cwnd); 5217 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_txrx, prr_out); 5218 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_txrx, lost_out); 5219 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_txrx, sacked_out); 5220 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_txrx, scaling_ratio); 5221 5222 /* RX read-mostly hotpath cache lines */ 5223 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_rx, copied_seq); 5224 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_rx, snd_wl1); 5225 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_rx, tlp_high_seq); 5226 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_rx, rttvar_us); 5227 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_rx, retrans_out); 5228 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_rx, advmss); 5229 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_rx, urg_data); 5230 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_rx, lost); 5231 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_rx, rtt_min); 5232 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_rx, out_of_order_queue); 5233 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_rx, snd_ssthresh); 5234 5235 /* TX read-write hotpath cache lines */ 5236 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, data_segs_out); 5237 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, delivered); 5238 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, delivered_ce); 5239 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, bytes_acked); 5240 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, bytes_sent); 5241 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, first_tx_mstamp); 5242 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, delivered_mstamp); 5243 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, snd_sml); 5244 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, chrono_start); 5245 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, chrono_stat); 5246 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, write_seq); 5247 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, pushed_seq); 5248 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, lsndtime); 5249 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, mdev_us); 5250 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, tcp_wstamp_ns); 5251 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, accecn_opt_tstamp); 5252 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, rtt_seq); 5253 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, max_packets_out); 5254 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, cwnd_usage_seq); 5255 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, rate_delivered); 5256 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, rate_interval_us); 5257 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, tsorted_sent_queue); 5258 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, highest_sack); 5259 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, ecn_flags); 5260 5261 /* TXRX read-write hotpath cache lines */ 5262 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_txrx, pred_flags); 5263 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_txrx, tcp_clock_cache); 5264 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_txrx, tcp_mstamp); 5265 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_txrx, rcv_nxt); 5266 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_txrx, snd_nxt); 5267 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_txrx, snd_una); 5268 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_txrx, window_clamp); 5269 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_txrx, srtt_us); 5270 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_txrx, packets_out); 5271 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_txrx, snd_up); 5272 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_txrx, received_ce); 5273 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_txrx, received_ecn_bytes); 5274 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_txrx, app_limited); 5275 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_txrx, rcv_wnd); 5276 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_txrx, rcv_mwnd_seq); 5277 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_txrx, rcv_tstamp); 5278 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_txrx, rx_opt); 5279 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_txrx, segs_in); 5280 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_txrx, segs_out); 5281 5282 /* RX read-write hotpath cache lines */ 5283 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_rx, bytes_received); 5284 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_rx, data_segs_in); 5285 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_rx, rcv_wup); 5286 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_rx, rcv_rtt_last_tsecr); 5287 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_rx, delivered_ecn_bytes); 5288 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_rx, pkts_acked_ewma); 5289 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_rx, rcv_rtt_est); 5290 CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_rx, rcvq_space); 5291 } 5292 5293 void __init tcp_init(void) 5294 { 5295 int max_rshare, max_wshare, cnt; 5296 unsigned long limit; 5297 unsigned int i; 5298 5299 BUILD_BUG_ON(TCP_MIN_SND_MSS <= MAX_TCP_OPTION_SPACE); 5300 BUILD_BUG_ON(sizeof(struct tcp_skb_cb) > 5301 sizeof_field(struct sk_buff, cb)); 5302 5303 tcp_struct_check(); 5304 5305 percpu_counter_init(&tcp_sockets_allocated, 0, GFP_KERNEL); 5306 5307 timer_setup(&tcp_orphan_timer, tcp_orphan_update, TIMER_DEFERRABLE); 5308 mod_timer(&tcp_orphan_timer, jiffies + TCP_ORPHAN_TIMER_PERIOD); 5309 5310 inet_hashinfo2_init(&tcp_hashinfo, "tcp_listen_portaddr_hash", 5311 thash_entries, 21, /* one slot per 2 MB*/ 5312 0, 64 * 1024); 5313 tcp_hashinfo.bind_bucket_cachep = 5314 kmem_cache_create("tcp_bind_bucket", 5315 sizeof(struct inet_bind_bucket), 0, 5316 SLAB_HWCACHE_ALIGN | SLAB_PANIC | 5317 SLAB_ACCOUNT, 5318 NULL); 5319 tcp_hashinfo.bind2_bucket_cachep = 5320 kmem_cache_create("tcp_bind2_bucket", 5321 sizeof(struct inet_bind2_bucket), 0, 5322 SLAB_HWCACHE_ALIGN | SLAB_PANIC | 5323 SLAB_ACCOUNT, 5324 NULL); 5325 5326 /* Size and allocate the main established and bind bucket 5327 * hash tables. 5328 * 5329 * The methodology is similar to that of the buffer cache. 5330 */ 5331 tcp_hashinfo.ehash = 5332 alloc_large_system_hash("TCP established", 5333 sizeof(struct inet_ehash_bucket), 5334 thash_entries, 5335 17, /* one slot per 128 KB of memory */ 5336 0, 5337 NULL, 5338 &tcp_hashinfo.ehash_mask, 5339 0, 5340 thash_entries ? 0 : 512 * 1024); 5341 for (i = 0; i <= tcp_hashinfo.ehash_mask; i++) 5342 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].chain, i); 5343 5344 if (inet_ehash_locks_alloc(&tcp_hashinfo)) 5345 panic("TCP: failed to alloc ehash_locks"); 5346 tcp_hashinfo.bhash = 5347 alloc_large_system_hash("TCP bind", 5348 2 * sizeof(struct inet_bind_hashbucket), 5349 tcp_hashinfo.ehash_mask + 1, 5350 17, /* one slot per 128 KB of memory */ 5351 0, 5352 &tcp_hashinfo.bhash_size, 5353 NULL, 5354 0, 5355 64 * 1024); 5356 tcp_hashinfo.bhash_size = 1U << tcp_hashinfo.bhash_size; 5357 tcp_hashinfo.bhash2 = tcp_hashinfo.bhash + tcp_hashinfo.bhash_size; 5358 for (i = 0; i < tcp_hashinfo.bhash_size; i++) { 5359 spin_lock_init(&tcp_hashinfo.bhash[i].lock); 5360 INIT_HLIST_HEAD(&tcp_hashinfo.bhash[i].chain); 5361 spin_lock_init(&tcp_hashinfo.bhash2[i].lock); 5362 INIT_HLIST_HEAD(&tcp_hashinfo.bhash2[i].chain); 5363 } 5364 5365 tcp_hashinfo.pernet = false; 5366 5367 cnt = tcp_hashinfo.ehash_mask + 1; 5368 sysctl_tcp_max_orphans = cnt / 2; 5369 5370 tcp_init_mem(); 5371 /* Set per-socket limits to no more than 1/128 the pressure threshold */ 5372 limit = nr_free_buffer_pages() << (PAGE_SHIFT - 7); 5373 max_wshare = min(4UL*1024*1024, limit); 5374 max_rshare = min(32UL*1024*1024, limit); 5375 5376 init_net.ipv4.sysctl_tcp_wmem[0] = PAGE_SIZE; 5377 init_net.ipv4.sysctl_tcp_wmem[1] = 16*1024; 5378 init_net.ipv4.sysctl_tcp_wmem[2] = max(64*1024, max_wshare); 5379 5380 init_net.ipv4.sysctl_tcp_rmem[0] = PAGE_SIZE; 5381 init_net.ipv4.sysctl_tcp_rmem[1] = 131072; 5382 init_net.ipv4.sysctl_tcp_rmem[2] = max(131072, max_rshare); 5383 5384 pr_info("Hash tables configured (established %u bind %u)\n", 5385 tcp_hashinfo.ehash_mask + 1, tcp_hashinfo.bhash_size); 5386 5387 tcp_v4_init(); 5388 tcp_metrics_init(); 5389 BUG_ON(tcp_register_congestion_control(&tcp_reno) != 0); 5390 tcp_tsq_work_init(); 5391 mptcp_init(); 5392 } 5393