PageRenderTime 65ms CodeModel.GetById 27ms RepoModel.GetById 0ms app.codeStats 1ms

/net/ipv4/tcp.c

https://bitbucket.org/bradfa/linux
C | 3678 lines | 2450 code | 517 blank | 711 comment | 591 complexity | c197fe294b462b6ba1ba5494f7f82caa MD5 | raw file
Possible License(s): GPL-2.0, LGPL-2.0, AGPL-1.0

Large files files are truncated, but you can click here to view the full file

  1. /*
  2. * INET An implementation of the TCP/IP protocol suite for the LINUX
  3. * operating system. INET is implemented using the BSD Socket
  4. * interface as the means of communication with the user level.
  5. *
  6. * Implementation of the Transmission Control Protocol(TCP).
  7. *
  8. * Authors: Ross Biro
  9. * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
  10. * Mark Evans, <evansmp@uhura.aston.ac.uk>
  11. * Corey Minyard <wf-rch!minyard@relay.EU.net>
  12. * Florian La Roche, <flla@stud.uni-sb.de>
  13. * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
  14. * Linus Torvalds, <torvalds@cs.helsinki.fi>
  15. * Alan Cox, <gw4pts@gw4pts.ampr.org>
  16. * Matthew Dillon, <dillon@apollo.west.oic.com>
  17. * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
  18. * Jorge Cwik, <jorge@laser.satlink.net>
  19. *
  20. * Fixes:
  21. * Alan Cox : Numerous verify_area() calls
  22. * Alan Cox : Set the ACK bit on a reset
  23. * Alan Cox : Stopped it crashing if it closed while
  24. * sk->inuse=1 and was trying to connect
  25. * (tcp_err()).
  26. * Alan Cox : All icmp error handling was broken
  27. * pointers passed where wrong and the
  28. * socket was looked up backwards. Nobody
  29. * tested any icmp error code obviously.
  30. * Alan Cox : tcp_err() now handled properly. It
  31. * wakes people on errors. poll
  32. * behaves and the icmp error race
  33. * has gone by moving it into sock.c
  34. * Alan Cox : tcp_send_reset() fixed to work for
  35. * everything not just packets for
  36. * unknown sockets.
  37. * Alan Cox : tcp option processing.
  38. * Alan Cox : Reset tweaked (still not 100%) [Had
  39. * syn rule wrong]
  40. * Herp Rosmanith : More reset fixes
  41. * Alan Cox : No longer acks invalid rst frames.
  42. * Acking any kind of RST is right out.
  43. * Alan Cox : Sets an ignore me flag on an rst
  44. * receive otherwise odd bits of prattle
  45. * escape still
  46. * Alan Cox : Fixed another acking RST frame bug.
  47. * Should stop LAN workplace lockups.
  48. * Alan Cox : Some tidyups using the new skb list
  49. * facilities
  50. * Alan Cox : sk->keepopen now seems to work
  51. * Alan Cox : Pulls options out correctly on accepts
  52. * Alan Cox : Fixed assorted sk->rqueue->next errors
  53. * Alan Cox : PSH doesn't end a TCP read. Switched a
  54. * bit to skb ops.
  55. * Alan Cox : Tidied tcp_data to avoid a potential
  56. * nasty.
  57. * Alan Cox : Added some better commenting, as the
  58. * tcp is hard to follow
  59. * Alan Cox : Removed incorrect check for 20 * psh
  60. * Michael O'Reilly : ack < copied bug fix.
  61. * Johannes Stille : Misc tcp fixes (not all in yet).
  62. * Alan Cox : FIN with no memory -> CRASH
  63. * Alan Cox : Added socket option proto entries.
  64. * Also added awareness of them to accept.
  65. * Alan Cox : Added TCP options (SOL_TCP)
  66. * Alan Cox : Switched wakeup calls to callbacks,
  67. * so the kernel can layer network
  68. * sockets.
  69. * Alan Cox : Use ip_tos/ip_ttl settings.
  70. * Alan Cox : Handle FIN (more) properly (we hope).
  71. * Alan Cox : RST frames sent on unsynchronised
  72. * state ack error.
  73. * Alan Cox : Put in missing check for SYN bit.
  74. * Alan Cox : Added tcp_select_window() aka NET2E
  75. * window non shrink trick.
  76. * Alan Cox : Added a couple of small NET2E timer
  77. * fixes
  78. * Charles Hedrick : TCP fixes
  79. * Toomas Tamm : TCP window fixes
  80. * Alan Cox : Small URG fix to rlogin ^C ack fight
  81. * Charles Hedrick : Rewrote most of it to actually work
  82. * Linus : Rewrote tcp_read() and URG handling
  83. * completely
  84. * Gerhard Koerting: Fixed some missing timer handling
  85. * Matthew Dillon : Reworked TCP machine states as per RFC
  86. * Gerhard Koerting: PC/TCP workarounds
  87. * Adam Caldwell : Assorted timer/timing errors
  88. * Matthew Dillon : Fixed another RST bug
  89. * Alan Cox : Move to kernel side addressing changes.
  90. * Alan Cox : Beginning work on TCP fastpathing
  91. * (not yet usable)
  92. * Arnt Gulbrandsen: Turbocharged tcp_check() routine.
  93. * Alan Cox : TCP fast path debugging
  94. * Alan Cox : Window clamping
  95. * Michael Riepe : Bug in tcp_check()
  96. * Matt Dillon : More TCP improvements and RST bug fixes
  97. * Matt Dillon : Yet more small nasties remove from the
  98. * TCP code (Be very nice to this man if
  99. * tcp finally works 100%) 8)
  100. * Alan Cox : BSD accept semantics.
  101. * Alan Cox : Reset on closedown bug.
  102. * Peter De Schrijver : ENOTCONN check missing in tcp_sendto().
  103. * Michael Pall : Handle poll() after URG properly in
  104. * all cases.
  105. * Michael Pall : Undo the last fix in tcp_read_urg()
  106. * (multi URG PUSH broke rlogin).
  107. * Michael Pall : Fix the multi URG PUSH problem in
  108. * tcp_readable(), poll() after URG
  109. * works now.
  110. * Michael Pall : recv(...,MSG_OOB) never blocks in the
  111. * BSD api.
  112. * Alan Cox : Changed the semantics of sk->socket to
  113. * fix a race and a signal problem with
  114. * accept() and async I/O.
  115. * Alan Cox : Relaxed the rules on tcp_sendto().
  116. * Yury Shevchuk : Really fixed accept() blocking problem.
  117. * Craig I. Hagan : Allow for BSD compatible TIME_WAIT for
  118. * clients/servers which listen in on
  119. * fixed ports.
  120. * Alan Cox : Cleaned the above up and shrank it to
  121. * a sensible code size.
  122. * Alan Cox : Self connect lockup fix.
  123. * Alan Cox : No connect to multicast.
  124. * Ross Biro : Close unaccepted children on master
  125. * socket close.
  126. * Alan Cox : Reset tracing code.
  127. * Alan Cox : Spurious resets on shutdown.
  128. * Alan Cox : Giant 15 minute/60 second timer error
  129. * Alan Cox : Small whoops in polling before an
  130. * accept.
  131. * Alan Cox : Kept the state trace facility since
  132. * it's handy for debugging.
  133. * Alan Cox : More reset handler fixes.
  134. * Alan Cox : Started rewriting the code based on
  135. * the RFC's for other useful protocol
  136. * references see: Comer, KA9Q NOS, and
  137. * for a reference on the difference
  138. * between specifications and how BSD
  139. * works see the 4.4lite source.
  140. * A.N.Kuznetsov : Don't time wait on completion of tidy
  141. * close.
  142. * Linus Torvalds : Fin/Shutdown & copied_seq changes.
  143. * Linus Torvalds : Fixed BSD port reuse to work first syn
  144. * Alan Cox : Reimplemented timers as per the RFC
  145. * and using multiple timers for sanity.
  146. * Alan Cox : Small bug fixes, and a lot of new
  147. * comments.
  148. * Alan Cox : Fixed dual reader crash by locking
  149. * the buffers (much like datagram.c)
  150. * Alan Cox : Fixed stuck sockets in probe. A probe
  151. * now gets fed up of retrying without
  152. * (even a no space) answer.
  153. * Alan Cox : Extracted closing code better
  154. * Alan Cox : Fixed the closing state machine to
  155. * resemble the RFC.
  156. * Alan Cox : More 'per spec' fixes.
  157. * Jorge Cwik : Even faster checksumming.
  158. * Alan Cox : tcp_data() doesn't ack illegal PSH
  159. * only frames. At least one pc tcp stack
  160. * generates them.
  161. * Alan Cox : Cache last socket.
  162. * Alan Cox : Per route irtt.
  163. * Matt Day : poll()->select() match BSD precisely on error
  164. * Alan Cox : New buffers
  165. * Marc Tamsky : Various sk->prot->retransmits and
  166. * sk->retransmits misupdating fixed.
  167. * Fixed tcp_write_timeout: stuck close,
  168. * and TCP syn retries gets used now.
  169. * Mark Yarvis : In tcp_read_wakeup(), don't send an
  170. * ack if state is TCP_CLOSED.
  171. * Alan Cox : Look up device on a retransmit - routes may
  172. * change. Doesn't yet cope with MSS shrink right
  173. * but it's a start!
  174. * Marc Tamsky : Closing in closing fixes.
  175. * Mike Shaver : RFC1122 verifications.
  176. * Alan Cox : rcv_saddr errors.
  177. * Alan Cox : Block double connect().
  178. * Alan Cox : Small hooks for enSKIP.
  179. * Alexey Kuznetsov: Path MTU discovery.
  180. * Alan Cox : Support soft errors.
  181. * Alan Cox : Fix MTU discovery pathological case
  182. * when the remote claims no mtu!
  183. * Marc Tamsky : TCP_CLOSE fix.
  184. * Colin (G3TNE) : Send a reset on syn ack replies in
  185. * window but wrong (fixes NT lpd problems)
  186. * Pedro Roque : Better TCP window handling, delayed ack.
  187. * Joerg Reuter : No modification of locked buffers in
  188. * tcp_do_retransmit()
  189. * Eric Schenk : Changed receiver side silly window
  190. * avoidance algorithm to BSD style
  191. * algorithm. This doubles throughput
  192. * against machines running Solaris,
  193. * and seems to result in general
  194. * improvement.
  195. * Stefan Magdalinski : adjusted tcp_readable() to fix FIONREAD
  196. * Willy Konynenberg : Transparent proxying support.
  197. * Mike McLagan : Routing by source
  198. * Keith Owens : Do proper merging with partial SKB's in
  199. * tcp_do_sendmsg to avoid burstiness.
  200. * Eric Schenk : Fix fast close down bug with
  201. * shutdown() followed by close().
  202. * Andi Kleen : Make poll agree with SIGIO
  203. * Salvatore Sanfilippo : Support SO_LINGER with linger == 1 and
  204. * lingertime == 0 (RFC 793 ABORT Call)
  205. * Hirokazu Takahashi : Use copy_from_user() instead of
  206. * csum_and_copy_from_user() if possible.
  207. *
  208. * This program is free software; you can redistribute it and/or
  209. * modify it under the terms of the GNU General Public License
  210. * as published by the Free Software Foundation; either version
  211. * 2 of the License, or(at your option) any later version.
  212. *
  213. * Description of States:
  214. *
  215. * TCP_SYN_SENT sent a connection request, waiting for ack
  216. *
  217. * TCP_SYN_RECV received a connection request, sent ack,
  218. * waiting for final ack in three-way handshake.
  219. *
  220. * TCP_ESTABLISHED connection established
  221. *
  222. * TCP_FIN_WAIT1 our side has shutdown, waiting to complete
  223. * transmission of remaining buffered data
  224. *
  225. * TCP_FIN_WAIT2 all buffered data sent, waiting for remote
  226. * to shutdown
  227. *
  228. * TCP_CLOSING both sides have shutdown but we still have
  229. * data we have to finish sending
  230. *
  231. * TCP_TIME_WAIT timeout to catch resent junk before entering
  232. * closed, can only be entered from FIN_WAIT2
  233. * or CLOSING. Required because the other end
  234. * may not have gotten our last ACK causing it
  235. * to retransmit the data packet (which we ignore)
  236. *
  237. * TCP_CLOSE_WAIT remote side has shutdown and is waiting for
  238. * us to finish writing our data and to shutdown
  239. * (we have to close() to move on to LAST_ACK)
  240. *
  241. * TCP_LAST_ACK out side has shutdown after remote has
  242. * shutdown. There may still be data in our
  243. * buffer that we have to finish sending
  244. *
  245. * TCP_CLOSE socket is finished
  246. */
  247. #define pr_fmt(fmt) "TCP: " fmt
  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/init.h>
  254. #include <linux/fs.h>
  255. #include <linux/skbuff.h>
  256. #include <linux/scatterlist.h>
  257. #include <linux/splice.h>
  258. #include <linux/net.h>
  259. #include <linux/socket.h>
  260. #include <linux/random.h>
  261. #include <linux/bootmem.h>
  262. #include <linux/highmem.h>
  263. #include <linux/swap.h>
  264. #include <linux/cache.h>
  265. #include <linux/err.h>
  266. #include <linux/crypto.h>
  267. #include <linux/time.h>
  268. #include <linux/slab.h>
  269. #include <net/icmp.h>
  270. #include <net/inet_common.h>
  271. #include <net/tcp.h>
  272. #include <net/xfrm.h>
  273. #include <net/ip.h>
  274. #include <net/netdma.h>
  275. #include <net/sock.h>
  276. #include <asm/uaccess.h>
  277. #include <asm/ioctls.h>
  278. int sysctl_tcp_fin_timeout __read_mostly = TCP_FIN_TIMEOUT;
  279. struct percpu_counter tcp_orphan_count;
  280. EXPORT_SYMBOL_GPL(tcp_orphan_count);
  281. int sysctl_tcp_wmem[3] __read_mostly;
  282. int sysctl_tcp_rmem[3] __read_mostly;
  283. EXPORT_SYMBOL(sysctl_tcp_rmem);
  284. EXPORT_SYMBOL(sysctl_tcp_wmem);
  285. atomic_long_t tcp_memory_allocated; /* Current allocated memory. */
  286. EXPORT_SYMBOL(tcp_memory_allocated);
  287. /*
  288. * Current number of TCP sockets.
  289. */
  290. struct percpu_counter tcp_sockets_allocated;
  291. EXPORT_SYMBOL(tcp_sockets_allocated);
  292. /*
  293. * TCP splice context
  294. */
  295. struct tcp_splice_state {
  296. struct pipe_inode_info *pipe;
  297. size_t len;
  298. unsigned int flags;
  299. };
  300. /*
  301. * Pressure flag: try to collapse.
  302. * Technical note: it is used by multiple contexts non atomically.
  303. * All the __sk_mem_schedule() is of this nature: accounting
  304. * is strict, actions are advisory and have some latency.
  305. */
  306. int tcp_memory_pressure __read_mostly;
  307. EXPORT_SYMBOL(tcp_memory_pressure);
  308. void tcp_enter_memory_pressure(struct sock *sk)
  309. {
  310. if (!tcp_memory_pressure) {
  311. NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMEMORYPRESSURES);
  312. tcp_memory_pressure = 1;
  313. }
  314. }
  315. EXPORT_SYMBOL(tcp_enter_memory_pressure);
  316. /* Convert seconds to retransmits based on initial and max timeout */
  317. static u8 secs_to_retrans(int seconds, int timeout, int rto_max)
  318. {
  319. u8 res = 0;
  320. if (seconds > 0) {
  321. int period = timeout;
  322. res = 1;
  323. while (seconds > period && res < 255) {
  324. res++;
  325. timeout <<= 1;
  326. if (timeout > rto_max)
  327. timeout = rto_max;
  328. period += timeout;
  329. }
  330. }
  331. return res;
  332. }
  333. /* Convert retransmits to seconds based on initial and max timeout */
  334. static int retrans_to_secs(u8 retrans, int timeout, int rto_max)
  335. {
  336. int period = 0;
  337. if (retrans > 0) {
  338. period = timeout;
  339. while (--retrans) {
  340. timeout <<= 1;
  341. if (timeout > rto_max)
  342. timeout = rto_max;
  343. period += timeout;
  344. }
  345. }
  346. return period;
  347. }
  348. /* Address-family independent initialization for a tcp_sock.
  349. *
  350. * NOTE: A lot of things set to zero explicitly by call to
  351. * sk_alloc() so need not be done here.
  352. */
  353. void tcp_init_sock(struct sock *sk)
  354. {
  355. struct inet_connection_sock *icsk = inet_csk(sk);
  356. struct tcp_sock *tp = tcp_sk(sk);
  357. skb_queue_head_init(&tp->out_of_order_queue);
  358. tcp_init_xmit_timers(sk);
  359. tcp_prequeue_init(tp);
  360. INIT_LIST_HEAD(&tp->tsq_node);
  361. icsk->icsk_rto = TCP_TIMEOUT_INIT;
  362. tp->mdev = TCP_TIMEOUT_INIT;
  363. /* So many TCP implementations out there (incorrectly) count the
  364. * initial SYN frame in their delayed-ACK and congestion control
  365. * algorithms that we must have the following bandaid to talk
  366. * efficiently to them. -DaveM
  367. */
  368. tp->snd_cwnd = TCP_INIT_CWND;
  369. /* See draft-stevens-tcpca-spec-01 for discussion of the
  370. * initialization of these values.
  371. */
  372. tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
  373. tp->snd_cwnd_clamp = ~0;
  374. tp->mss_cache = TCP_MSS_DEFAULT;
  375. tp->reordering = sysctl_tcp_reordering;
  376. tcp_enable_early_retrans(tp);
  377. icsk->icsk_ca_ops = &tcp_init_congestion_ops;
  378. tp->tsoffset = 0;
  379. sk->sk_state = TCP_CLOSE;
  380. sk->sk_write_space = sk_stream_write_space;
  381. sock_set_flag(sk, SOCK_USE_WRITE_QUEUE);
  382. icsk->icsk_sync_mss = tcp_sync_mss;
  383. /* TCP Cookie Transactions */
  384. if (sysctl_tcp_cookie_size > 0) {
  385. /* Default, cookies without s_data_payload. */
  386. tp->cookie_values =
  387. kzalloc(sizeof(*tp->cookie_values),
  388. sk->sk_allocation);
  389. if (tp->cookie_values != NULL)
  390. kref_init(&tp->cookie_values->kref);
  391. }
  392. /* Presumed zeroed, in order of appearance:
  393. * cookie_in_always, cookie_out_never,
  394. * s_data_constant, s_data_in, s_data_out
  395. */
  396. sk->sk_sndbuf = sysctl_tcp_wmem[1];
  397. sk->sk_rcvbuf = sysctl_tcp_rmem[1];
  398. local_bh_disable();
  399. sock_update_memcg(sk);
  400. sk_sockets_allocated_inc(sk);
  401. local_bh_enable();
  402. }
  403. EXPORT_SYMBOL(tcp_init_sock);
  404. /*
  405. * Wait for a TCP event.
  406. *
  407. * Note that we don't need to lock the socket, as the upper poll layers
  408. * take care of normal races (between the test and the event) and we don't
  409. * go look at any of the socket buffers directly.
  410. */
  411. unsigned int tcp_poll(struct file *file, struct socket *sock, poll_table *wait)
  412. {
  413. unsigned int mask;
  414. struct sock *sk = sock->sk;
  415. const struct tcp_sock *tp = tcp_sk(sk);
  416. sock_poll_wait(file, sk_sleep(sk), wait);
  417. if (sk->sk_state == TCP_LISTEN)
  418. return inet_csk_listen_poll(sk);
  419. /* Socket is not locked. We are protected from async events
  420. * by poll logic and correct handling of state changes
  421. * made by other threads is impossible in any case.
  422. */
  423. mask = 0;
  424. /*
  425. * POLLHUP is certainly not done right. But poll() doesn't
  426. * have a notion of HUP in just one direction, and for a
  427. * socket the read side is more interesting.
  428. *
  429. * Some poll() documentation says that POLLHUP is incompatible
  430. * with the POLLOUT/POLLWR flags, so somebody should check this
  431. * all. But careful, it tends to be safer to return too many
  432. * bits than too few, and you can easily break real applications
  433. * if you don't tell them that something has hung up!
  434. *
  435. * Check-me.
  436. *
  437. * Check number 1. POLLHUP is _UNMASKABLE_ event (see UNIX98 and
  438. * our fs/select.c). It means that after we received EOF,
  439. * poll always returns immediately, making impossible poll() on write()
  440. * in state CLOSE_WAIT. One solution is evident --- to set POLLHUP
  441. * if and only if shutdown has been made in both directions.
  442. * Actually, it is interesting to look how Solaris and DUX
  443. * solve this dilemma. I would prefer, if POLLHUP were maskable,
  444. * then we could set it on SND_SHUTDOWN. BTW examples given
  445. * in Stevens' books assume exactly this behaviour, it explains
  446. * why POLLHUP is incompatible with POLLOUT. --ANK
  447. *
  448. * NOTE. Check for TCP_CLOSE is added. The goal is to prevent
  449. * blocking on fresh not-connected or disconnected socket. --ANK
  450. */
  451. if (sk->sk_shutdown == SHUTDOWN_MASK || sk->sk_state == TCP_CLOSE)
  452. mask |= POLLHUP;
  453. if (sk->sk_shutdown & RCV_SHUTDOWN)
  454. mask |= POLLIN | POLLRDNORM | POLLRDHUP;
  455. /* Connected or passive Fast Open socket? */
  456. if (sk->sk_state != TCP_SYN_SENT &&
  457. (sk->sk_state != TCP_SYN_RECV || tp->fastopen_rsk != NULL)) {
  458. int target = sock_rcvlowat(sk, 0, INT_MAX);
  459. if (tp->urg_seq == tp->copied_seq &&
  460. !sock_flag(sk, SOCK_URGINLINE) &&
  461. tp->urg_data)
  462. target++;
  463. /* Potential race condition. If read of tp below will
  464. * escape above sk->sk_state, we can be illegally awaken
  465. * in SYN_* states. */
  466. if (tp->rcv_nxt - tp->copied_seq >= target)
  467. mask |= POLLIN | POLLRDNORM;
  468. if (!(sk->sk_shutdown & SEND_SHUTDOWN)) {
  469. if (sk_stream_wspace(sk) >= sk_stream_min_wspace(sk)) {
  470. mask |= POLLOUT | POLLWRNORM;
  471. } else { /* send SIGIO later */
  472. set_bit(SOCK_ASYNC_NOSPACE,
  473. &sk->sk_socket->flags);
  474. set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
  475. /* Race breaker. If space is freed after
  476. * wspace test but before the flags are set,
  477. * IO signal will be lost.
  478. */
  479. if (sk_stream_wspace(sk) >= sk_stream_min_wspace(sk))
  480. mask |= POLLOUT | POLLWRNORM;
  481. }
  482. } else
  483. mask |= POLLOUT | POLLWRNORM;
  484. if (tp->urg_data & TCP_URG_VALID)
  485. mask |= POLLPRI;
  486. }
  487. /* This barrier is coupled with smp_wmb() in tcp_reset() */
  488. smp_rmb();
  489. if (sk->sk_err)
  490. mask |= POLLERR;
  491. return mask;
  492. }
  493. EXPORT_SYMBOL(tcp_poll);
  494. int tcp_ioctl(struct sock *sk, int cmd, unsigned long arg)
  495. {
  496. struct tcp_sock *tp = tcp_sk(sk);
  497. int answ;
  498. bool slow;
  499. switch (cmd) {
  500. case SIOCINQ:
  501. if (sk->sk_state == TCP_LISTEN)
  502. return -EINVAL;
  503. slow = lock_sock_fast(sk);
  504. if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
  505. answ = 0;
  506. else if (sock_flag(sk, SOCK_URGINLINE) ||
  507. !tp->urg_data ||
  508. before(tp->urg_seq, tp->copied_seq) ||
  509. !before(tp->urg_seq, tp->rcv_nxt)) {
  510. answ = tp->rcv_nxt - tp->copied_seq;
  511. /* Subtract 1, if FIN was received */
  512. if (answ && sock_flag(sk, SOCK_DONE))
  513. answ--;
  514. } else
  515. answ = tp->urg_seq - tp->copied_seq;
  516. unlock_sock_fast(sk, slow);
  517. break;
  518. case SIOCATMARK:
  519. answ = tp->urg_data && tp->urg_seq == tp->copied_seq;
  520. break;
  521. case SIOCOUTQ:
  522. if (sk->sk_state == TCP_LISTEN)
  523. return -EINVAL;
  524. if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
  525. answ = 0;
  526. else
  527. answ = tp->write_seq - tp->snd_una;
  528. break;
  529. case SIOCOUTQNSD:
  530. if (sk->sk_state == TCP_LISTEN)
  531. return -EINVAL;
  532. if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
  533. answ = 0;
  534. else
  535. answ = tp->write_seq - tp->snd_nxt;
  536. break;
  537. default:
  538. return -ENOIOCTLCMD;
  539. }
  540. return put_user(answ, (int __user *)arg);
  541. }
  542. EXPORT_SYMBOL(tcp_ioctl);
  543. static inline void tcp_mark_push(struct tcp_sock *tp, struct sk_buff *skb)
  544. {
  545. TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
  546. tp->pushed_seq = tp->write_seq;
  547. }
  548. static inline bool forced_push(const struct tcp_sock *tp)
  549. {
  550. return after(tp->write_seq, tp->pushed_seq + (tp->max_window >> 1));
  551. }
  552. static inline void skb_entail(struct sock *sk, struct sk_buff *skb)
  553. {
  554. struct tcp_sock *tp = tcp_sk(sk);
  555. struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
  556. skb->csum = 0;
  557. tcb->seq = tcb->end_seq = tp->write_seq;
  558. tcb->tcp_flags = TCPHDR_ACK;
  559. tcb->sacked = 0;
  560. skb_header_release(skb);
  561. tcp_add_write_queue_tail(sk, skb);
  562. sk->sk_wmem_queued += skb->truesize;
  563. sk_mem_charge(sk, skb->truesize);
  564. if (tp->nonagle & TCP_NAGLE_PUSH)
  565. tp->nonagle &= ~TCP_NAGLE_PUSH;
  566. }
  567. static inline void tcp_mark_urg(struct tcp_sock *tp, int flags)
  568. {
  569. if (flags & MSG_OOB)
  570. tp->snd_up = tp->write_seq;
  571. }
  572. static inline void tcp_push(struct sock *sk, int flags, int mss_now,
  573. int nonagle)
  574. {
  575. if (tcp_send_head(sk)) {
  576. struct tcp_sock *tp = tcp_sk(sk);
  577. if (!(flags & MSG_MORE) || forced_push(tp))
  578. tcp_mark_push(tp, tcp_write_queue_tail(sk));
  579. tcp_mark_urg(tp, flags);
  580. __tcp_push_pending_frames(sk, mss_now,
  581. (flags & MSG_MORE) ? TCP_NAGLE_CORK : nonagle);
  582. }
  583. }
  584. static int tcp_splice_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb,
  585. unsigned int offset, size_t len)
  586. {
  587. struct tcp_splice_state *tss = rd_desc->arg.data;
  588. int ret;
  589. ret = skb_splice_bits(skb, offset, tss->pipe, min(rd_desc->count, len),
  590. tss->flags);
  591. if (ret > 0)
  592. rd_desc->count -= ret;
  593. return ret;
  594. }
  595. static int __tcp_splice_read(struct sock *sk, struct tcp_splice_state *tss)
  596. {
  597. /* Store TCP splice context information in read_descriptor_t. */
  598. read_descriptor_t rd_desc = {
  599. .arg.data = tss,
  600. .count = tss->len,
  601. };
  602. return tcp_read_sock(sk, &rd_desc, tcp_splice_data_recv);
  603. }
  604. /**
  605. * tcp_splice_read - splice data from TCP socket to a pipe
  606. * @sock: socket to splice from
  607. * @ppos: position (not valid)
  608. * @pipe: pipe to splice to
  609. * @len: number of bytes to splice
  610. * @flags: splice modifier flags
  611. *
  612. * Description:
  613. * Will read pages from given socket and fill them into a pipe.
  614. *
  615. **/
  616. ssize_t tcp_splice_read(struct socket *sock, loff_t *ppos,
  617. struct pipe_inode_info *pipe, size_t len,
  618. unsigned int flags)
  619. {
  620. struct sock *sk = sock->sk;
  621. struct tcp_splice_state tss = {
  622. .pipe = pipe,
  623. .len = len,
  624. .flags = flags,
  625. };
  626. long timeo;
  627. ssize_t spliced;
  628. int ret;
  629. sock_rps_record_flow(sk);
  630. /*
  631. * We can't seek on a socket input
  632. */
  633. if (unlikely(*ppos))
  634. return -ESPIPE;
  635. ret = spliced = 0;
  636. lock_sock(sk);
  637. timeo = sock_rcvtimeo(sk, sock->file->f_flags & O_NONBLOCK);
  638. while (tss.len) {
  639. ret = __tcp_splice_read(sk, &tss);
  640. if (ret < 0)
  641. break;
  642. else if (!ret) {
  643. if (spliced)
  644. break;
  645. if (sock_flag(sk, SOCK_DONE))
  646. break;
  647. if (sk->sk_err) {
  648. ret = sock_error(sk);
  649. break;
  650. }
  651. if (sk->sk_shutdown & RCV_SHUTDOWN)
  652. break;
  653. if (sk->sk_state == TCP_CLOSE) {
  654. /*
  655. * This occurs when user tries to read
  656. * from never connected socket.
  657. */
  658. if (!sock_flag(sk, SOCK_DONE))
  659. ret = -ENOTCONN;
  660. break;
  661. }
  662. if (!timeo) {
  663. ret = -EAGAIN;
  664. break;
  665. }
  666. sk_wait_data(sk, &timeo);
  667. if (signal_pending(current)) {
  668. ret = sock_intr_errno(timeo);
  669. break;
  670. }
  671. continue;
  672. }
  673. tss.len -= ret;
  674. spliced += ret;
  675. if (!timeo)
  676. break;
  677. release_sock(sk);
  678. lock_sock(sk);
  679. if (sk->sk_err || sk->sk_state == TCP_CLOSE ||
  680. (sk->sk_shutdown & RCV_SHUTDOWN) ||
  681. signal_pending(current))
  682. break;
  683. }
  684. release_sock(sk);
  685. if (spliced)
  686. return spliced;
  687. return ret;
  688. }
  689. EXPORT_SYMBOL(tcp_splice_read);
  690. struct sk_buff *sk_stream_alloc_skb(struct sock *sk, int size, gfp_t gfp)
  691. {
  692. struct sk_buff *skb;
  693. /* The TCP header must be at least 32-bit aligned. */
  694. size = ALIGN(size, 4);
  695. skb = alloc_skb_fclone(size + sk->sk_prot->max_header, gfp);
  696. if (skb) {
  697. if (sk_wmem_schedule(sk, skb->truesize)) {
  698. skb_reserve(skb, sk->sk_prot->max_header);
  699. /*
  700. * Make sure that we have exactly size bytes
  701. * available to the caller, no more, no less.
  702. */
  703. skb->reserved_tailroom = skb->end - skb->tail - size;
  704. return skb;
  705. }
  706. __kfree_skb(skb);
  707. } else {
  708. sk->sk_prot->enter_memory_pressure(sk);
  709. sk_stream_moderate_sndbuf(sk);
  710. }
  711. return NULL;
  712. }
  713. static unsigned int tcp_xmit_size_goal(struct sock *sk, u32 mss_now,
  714. int large_allowed)
  715. {
  716. struct tcp_sock *tp = tcp_sk(sk);
  717. u32 xmit_size_goal, old_size_goal;
  718. xmit_size_goal = mss_now;
  719. if (large_allowed && sk_can_gso(sk)) {
  720. xmit_size_goal = ((sk->sk_gso_max_size - 1) -
  721. inet_csk(sk)->icsk_af_ops->net_header_len -
  722. inet_csk(sk)->icsk_ext_hdr_len -
  723. tp->tcp_header_len);
  724. /* TSQ : try to have two TSO segments in flight */
  725. xmit_size_goal = min_t(u32, xmit_size_goal,
  726. sysctl_tcp_limit_output_bytes >> 1);
  727. xmit_size_goal = tcp_bound_to_half_wnd(tp, xmit_size_goal);
  728. /* We try hard to avoid divides here */
  729. old_size_goal = tp->xmit_size_goal_segs * mss_now;
  730. if (likely(old_size_goal <= xmit_size_goal &&
  731. old_size_goal + mss_now > xmit_size_goal)) {
  732. xmit_size_goal = old_size_goal;
  733. } else {
  734. tp->xmit_size_goal_segs =
  735. min_t(u16, xmit_size_goal / mss_now,
  736. sk->sk_gso_max_segs);
  737. xmit_size_goal = tp->xmit_size_goal_segs * mss_now;
  738. }
  739. }
  740. return max(xmit_size_goal, mss_now);
  741. }
  742. static int tcp_send_mss(struct sock *sk, int *size_goal, int flags)
  743. {
  744. int mss_now;
  745. mss_now = tcp_current_mss(sk);
  746. *size_goal = tcp_xmit_size_goal(sk, mss_now, !(flags & MSG_OOB));
  747. return mss_now;
  748. }
  749. static ssize_t do_tcp_sendpages(struct sock *sk, struct page *page, int offset,
  750. size_t size, int flags)
  751. {
  752. struct tcp_sock *tp = tcp_sk(sk);
  753. int mss_now, size_goal;
  754. int err;
  755. ssize_t copied;
  756. long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
  757. /* Wait for a connection to finish. One exception is TCP Fast Open
  758. * (passive side) where data is allowed to be sent before a connection
  759. * is fully established.
  760. */
  761. if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) &&
  762. !tcp_passive_fastopen(sk)) {
  763. if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
  764. goto out_err;
  765. }
  766. clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
  767. mss_now = tcp_send_mss(sk, &size_goal, flags);
  768. copied = 0;
  769. err = -EPIPE;
  770. if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
  771. goto out_err;
  772. while (size > 0) {
  773. struct sk_buff *skb = tcp_write_queue_tail(sk);
  774. int copy, i;
  775. bool can_coalesce;
  776. if (!tcp_send_head(sk) || (copy = size_goal - skb->len) <= 0) {
  777. new_segment:
  778. if (!sk_stream_memory_free(sk))
  779. goto wait_for_sndbuf;
  780. skb = sk_stream_alloc_skb(sk, 0, sk->sk_allocation);
  781. if (!skb)
  782. goto wait_for_memory;
  783. skb_entail(sk, skb);
  784. copy = size_goal;
  785. }
  786. if (copy > size)
  787. copy = size;
  788. i = skb_shinfo(skb)->nr_frags;
  789. can_coalesce = skb_can_coalesce(skb, i, page, offset);
  790. if (!can_coalesce && i >= MAX_SKB_FRAGS) {
  791. tcp_mark_push(tp, skb);
  792. goto new_segment;
  793. }
  794. if (!sk_wmem_schedule(sk, copy))
  795. goto wait_for_memory;
  796. if (can_coalesce) {
  797. skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
  798. } else {
  799. get_page(page);
  800. skb_fill_page_desc(skb, i, page, offset, copy);
  801. }
  802. skb_shinfo(skb)->tx_flags |= SKBTX_SHARED_FRAG;
  803. skb->len += copy;
  804. skb->data_len += copy;
  805. skb->truesize += copy;
  806. sk->sk_wmem_queued += copy;
  807. sk_mem_charge(sk, copy);
  808. skb->ip_summed = CHECKSUM_PARTIAL;
  809. tp->write_seq += copy;
  810. TCP_SKB_CB(skb)->end_seq += copy;
  811. skb_shinfo(skb)->gso_segs = 0;
  812. if (!copied)
  813. TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
  814. copied += copy;
  815. offset += copy;
  816. if (!(size -= copy))
  817. goto out;
  818. if (skb->len < size_goal || (flags & MSG_OOB))
  819. continue;
  820. if (forced_push(tp)) {
  821. tcp_mark_push(tp, skb);
  822. __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
  823. } else if (skb == tcp_send_head(sk))
  824. tcp_push_one(sk, mss_now);
  825. continue;
  826. wait_for_sndbuf:
  827. set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
  828. wait_for_memory:
  829. tcp_push(sk, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH);
  830. if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
  831. goto do_error;
  832. mss_now = tcp_send_mss(sk, &size_goal, flags);
  833. }
  834. out:
  835. if (copied && !(flags & MSG_SENDPAGE_NOTLAST))
  836. tcp_push(sk, flags, mss_now, tp->nonagle);
  837. return copied;
  838. do_error:
  839. if (copied)
  840. goto out;
  841. out_err:
  842. return sk_stream_error(sk, flags, err);
  843. }
  844. int tcp_sendpage(struct sock *sk, struct page *page, int offset,
  845. size_t size, int flags)
  846. {
  847. ssize_t res;
  848. if (!(sk->sk_route_caps & NETIF_F_SG) ||
  849. !(sk->sk_route_caps & NETIF_F_ALL_CSUM))
  850. return sock_no_sendpage(sk->sk_socket, page, offset, size,
  851. flags);
  852. lock_sock(sk);
  853. res = do_tcp_sendpages(sk, page, offset, size, flags);
  854. release_sock(sk);
  855. return res;
  856. }
  857. EXPORT_SYMBOL(tcp_sendpage);
  858. static inline int select_size(const struct sock *sk, bool sg)
  859. {
  860. const struct tcp_sock *tp = tcp_sk(sk);
  861. int tmp = tp->mss_cache;
  862. if (sg) {
  863. if (sk_can_gso(sk)) {
  864. /* Small frames wont use a full page:
  865. * Payload will immediately follow tcp header.
  866. */
  867. tmp = SKB_WITH_OVERHEAD(2048 - MAX_TCP_HEADER);
  868. } else {
  869. int pgbreak = SKB_MAX_HEAD(MAX_TCP_HEADER);
  870. if (tmp >= pgbreak &&
  871. tmp <= pgbreak + (MAX_SKB_FRAGS - 1) * PAGE_SIZE)
  872. tmp = pgbreak;
  873. }
  874. }
  875. return tmp;
  876. }
  877. void tcp_free_fastopen_req(struct tcp_sock *tp)
  878. {
  879. if (tp->fastopen_req != NULL) {
  880. kfree(tp->fastopen_req);
  881. tp->fastopen_req = NULL;
  882. }
  883. }
  884. static int tcp_sendmsg_fastopen(struct sock *sk, struct msghdr *msg, int *size)
  885. {
  886. struct tcp_sock *tp = tcp_sk(sk);
  887. int err, flags;
  888. if (!(sysctl_tcp_fastopen & TFO_CLIENT_ENABLE))
  889. return -EOPNOTSUPP;
  890. if (tp->fastopen_req != NULL)
  891. return -EALREADY; /* Another Fast Open is in progress */
  892. tp->fastopen_req = kzalloc(sizeof(struct tcp_fastopen_request),
  893. sk->sk_allocation);
  894. if (unlikely(tp->fastopen_req == NULL))
  895. return -ENOBUFS;
  896. tp->fastopen_req->data = msg;
  897. flags = (msg->msg_flags & MSG_DONTWAIT) ? O_NONBLOCK : 0;
  898. err = __inet_stream_connect(sk->sk_socket, msg->msg_name,
  899. msg->msg_namelen, flags);
  900. *size = tp->fastopen_req->copied;
  901. tcp_free_fastopen_req(tp);
  902. return err;
  903. }
  904. int tcp_sendmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg,
  905. size_t size)
  906. {
  907. struct iovec *iov;
  908. struct tcp_sock *tp = tcp_sk(sk);
  909. struct sk_buff *skb;
  910. int iovlen, flags, err, copied = 0;
  911. int mss_now = 0, size_goal, copied_syn = 0, offset = 0;
  912. bool sg;
  913. long timeo;
  914. lock_sock(sk);
  915. flags = msg->msg_flags;
  916. if (flags & MSG_FASTOPEN) {
  917. err = tcp_sendmsg_fastopen(sk, msg, &copied_syn);
  918. if (err == -EINPROGRESS && copied_syn > 0)
  919. goto out;
  920. else if (err)
  921. goto out_err;
  922. offset = copied_syn;
  923. }
  924. timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
  925. /* Wait for a connection to finish. One exception is TCP Fast Open
  926. * (passive side) where data is allowed to be sent before a connection
  927. * is fully established.
  928. */
  929. if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) &&
  930. !tcp_passive_fastopen(sk)) {
  931. if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
  932. goto do_error;
  933. }
  934. if (unlikely(tp->repair)) {
  935. if (tp->repair_queue == TCP_RECV_QUEUE) {
  936. copied = tcp_send_rcvq(sk, msg, size);
  937. goto out;
  938. }
  939. err = -EINVAL;
  940. if (tp->repair_queue == TCP_NO_QUEUE)
  941. goto out_err;
  942. /* 'common' sending to sendq */
  943. }
  944. /* This should be in poll */
  945. clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
  946. mss_now = tcp_send_mss(sk, &size_goal, flags);
  947. /* Ok commence sending. */
  948. iovlen = msg->msg_iovlen;
  949. iov = msg->msg_iov;
  950. copied = 0;
  951. err = -EPIPE;
  952. if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
  953. goto out_err;
  954. sg = !!(sk->sk_route_caps & NETIF_F_SG);
  955. while (--iovlen >= 0) {
  956. size_t seglen = iov->iov_len;
  957. unsigned char __user *from = iov->iov_base;
  958. iov++;
  959. if (unlikely(offset > 0)) { /* Skip bytes copied in SYN */
  960. if (offset >= seglen) {
  961. offset -= seglen;
  962. continue;
  963. }
  964. seglen -= offset;
  965. from += offset;
  966. offset = 0;
  967. }
  968. while (seglen > 0) {
  969. int copy = 0;
  970. int max = size_goal;
  971. skb = tcp_write_queue_tail(sk);
  972. if (tcp_send_head(sk)) {
  973. if (skb->ip_summed == CHECKSUM_NONE)
  974. max = mss_now;
  975. copy = max - skb->len;
  976. }
  977. if (copy <= 0) {
  978. new_segment:
  979. /* Allocate new segment. If the interface is SG,
  980. * allocate skb fitting to single page.
  981. */
  982. if (!sk_stream_memory_free(sk))
  983. goto wait_for_sndbuf;
  984. skb = sk_stream_alloc_skb(sk,
  985. select_size(sk, sg),
  986. sk->sk_allocation);
  987. if (!skb)
  988. goto wait_for_memory;
  989. /*
  990. * Check whether we can use HW checksum.
  991. */
  992. if (sk->sk_route_caps & NETIF_F_ALL_CSUM)
  993. skb->ip_summed = CHECKSUM_PARTIAL;
  994. skb_entail(sk, skb);
  995. copy = size_goal;
  996. max = size_goal;
  997. }
  998. /* Try to append data to the end of skb. */
  999. if (copy > seglen)
  1000. copy = seglen;
  1001. /* Where to copy to? */
  1002. if (skb_availroom(skb) > 0) {
  1003. /* We have some space in skb head. Superb! */
  1004. copy = min_t(int, copy, skb_availroom(skb));
  1005. err = skb_add_data_nocache(sk, skb, from, copy);
  1006. if (err)
  1007. goto do_fault;
  1008. } else {
  1009. bool merge = true;
  1010. int i = skb_shinfo(skb)->nr_frags;
  1011. struct page_frag *pfrag = sk_page_frag(sk);
  1012. if (!sk_page_frag_refill(sk, pfrag))
  1013. goto wait_for_memory;
  1014. if (!skb_can_coalesce(skb, i, pfrag->page,
  1015. pfrag->offset)) {
  1016. if (i == MAX_SKB_FRAGS || !sg) {
  1017. tcp_mark_push(tp, skb);
  1018. goto new_segment;
  1019. }
  1020. merge = false;
  1021. }
  1022. copy = min_t(int, copy, pfrag->size - pfrag->offset);
  1023. if (!sk_wmem_schedule(sk, copy))
  1024. goto wait_for_memory;
  1025. err = skb_copy_to_page_nocache(sk, from, skb,
  1026. pfrag->page,
  1027. pfrag->offset,
  1028. copy);
  1029. if (err)
  1030. goto do_error;
  1031. /* Update the skb. */
  1032. if (merge) {
  1033. skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
  1034. } else {
  1035. skb_fill_page_desc(skb, i, pfrag->page,
  1036. pfrag->offset, copy);
  1037. get_page(pfrag->page);
  1038. }
  1039. pfrag->offset += copy;
  1040. }
  1041. if (!copied)
  1042. TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
  1043. tp->write_seq += copy;
  1044. TCP_SKB_CB(skb)->end_seq += copy;
  1045. skb_shinfo(skb)->gso_segs = 0;
  1046. from += copy;
  1047. copied += copy;
  1048. if ((seglen -= copy) == 0 && iovlen == 0)
  1049. goto out;
  1050. if (skb->len < max || (flags & MSG_OOB) || unlikely(tp->repair))
  1051. continue;
  1052. if (forced_push(tp)) {
  1053. tcp_mark_push(tp, skb);
  1054. __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
  1055. } else if (skb == tcp_send_head(sk))
  1056. tcp_push_one(sk, mss_now);
  1057. continue;
  1058. wait_for_sndbuf:
  1059. set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
  1060. wait_for_memory:
  1061. if (copied)
  1062. tcp_push(sk, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH);
  1063. if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
  1064. goto do_error;
  1065. mss_now = tcp_send_mss(sk, &size_goal, flags);
  1066. }
  1067. }
  1068. out:
  1069. if (copied)
  1070. tcp_push(sk, flags, mss_now, tp->nonagle);
  1071. release_sock(sk);
  1072. return copied + copied_syn;
  1073. do_fault:
  1074. if (!skb->len) {
  1075. tcp_unlink_write_queue(skb, sk);
  1076. /* It is the one place in all of TCP, except connection
  1077. * reset, where we can be unlinking the send_head.
  1078. */
  1079. tcp_check_send_head(sk, skb);
  1080. sk_wmem_free_skb(sk, skb);
  1081. }
  1082. do_error:
  1083. if (copied + copied_syn)
  1084. goto out;
  1085. out_err:
  1086. err = sk_stream_error(sk, flags, err);
  1087. release_sock(sk);
  1088. return err;
  1089. }
  1090. EXPORT_SYMBOL(tcp_sendmsg);
  1091. /*
  1092. * Handle reading urgent data. BSD has very simple semantics for
  1093. * this, no blocking and very strange errors 8)
  1094. */
  1095. static int tcp_recv_urg(struct sock *sk, struct msghdr *msg, int len, int flags)
  1096. {
  1097. struct tcp_sock *tp = tcp_sk(sk);
  1098. /* No URG data to read. */
  1099. if (sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data ||
  1100. tp->urg_data == TCP_URG_READ)
  1101. return -EINVAL; /* Yes this is right ! */
  1102. if (sk->sk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DONE))
  1103. return -ENOTCONN;
  1104. if (tp->urg_data & TCP_URG_VALID) {
  1105. int err = 0;
  1106. char c = tp->urg_data;
  1107. if (!(flags & MSG_PEEK))
  1108. tp->urg_data = TCP_URG_READ;
  1109. /* Read urgent data. */
  1110. msg->msg_flags |= MSG_OOB;
  1111. if (len > 0) {
  1112. if (!(flags & MSG_TRUNC))
  1113. err = memcpy_toiovec(msg->msg_iov, &c, 1);
  1114. len = 1;
  1115. } else
  1116. msg->msg_flags |= MSG_TRUNC;
  1117. return err ? -EFAULT : len;
  1118. }
  1119. if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
  1120. return 0;
  1121. /* Fixed the recv(..., MSG_OOB) behaviour. BSD docs and
  1122. * the available implementations agree in this case:
  1123. * this call should never block, independent of the
  1124. * blocking state of the socket.
  1125. * Mike <pall@rz.uni-karlsruhe.de>
  1126. */
  1127. return -EAGAIN;
  1128. }
  1129. static int tcp_peek_sndq(struct sock *sk, struct msghdr *msg, int len)
  1130. {
  1131. struct sk_buff *skb;
  1132. int copied = 0, err = 0;
  1133. /* XXX -- need to support SO_PEEK_OFF */
  1134. skb_queue_walk(&sk->sk_write_queue, skb) {
  1135. err = skb_copy_datagram_iovec(skb, 0, msg->msg_iov, skb->len);
  1136. if (err)
  1137. break;
  1138. copied += skb->len;
  1139. }
  1140. return err ?: copied;
  1141. }
  1142. /* Clean up the receive buffer for full frames taken by the user,
  1143. * then send an ACK if necessary. COPIED is the number of bytes
  1144. * tcp_recvmsg has given to the user so far, it speeds up the
  1145. * calculation of whether or not we must ACK for the sake of
  1146. * a window update.
  1147. */
  1148. void tcp_cleanup_rbuf(struct sock *sk, int copied)
  1149. {
  1150. struct tcp_sock *tp = tcp_sk(sk);
  1151. bool time_to_ack = false;
  1152. struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
  1153. WARN(skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq),
  1154. "cleanup rbuf bug: copied %X seq %X rcvnxt %X\n",
  1155. tp->copied_seq, TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt);
  1156. if (inet_csk_ack_scheduled(sk)) {
  1157. const struct inet_connection_sock *icsk = inet_csk(sk);
  1158. /* Delayed ACKs frequently hit locked sockets during bulk
  1159. * receive. */
  1160. if (icsk->icsk_ack.blocked ||
  1161. /* Once-per-two-segments ACK was not sent by tcp_input.c */
  1162. tp->rcv_nxt - tp->rcv_wup > icsk->icsk_ack.rcv_mss ||
  1163. /*
  1164. * If this read emptied read buffer, we send ACK, if
  1165. * connection is not bidirectional, user drained
  1166. * receive buffer and there was a small segment
  1167. * in queue.
  1168. */
  1169. (copied > 0 &&
  1170. ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED2) ||
  1171. ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED) &&
  1172. !icsk->icsk_ack.pingpong)) &&
  1173. !atomic_read(&sk->sk_rmem_alloc)))
  1174. time_to_ack = true;
  1175. }
  1176. /* We send an ACK if we can now advertise a non-zero window
  1177. * which has been raised "significantly".
  1178. *
  1179. * Even if window raised up to infinity, do not send window open ACK
  1180. * in states, where we will not receive more. It is useless.
  1181. */
  1182. if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) {
  1183. __u32 rcv_window_now = tcp_receive_window(tp);
  1184. /* Optimize, __tcp_select_window() is not cheap. */
  1185. if (2*rcv_window_now <= tp->window_clamp) {
  1186. __u32 new_window = __tcp_select_window(sk);
  1187. /* Send ACK now, if this read freed lots of space
  1188. * in our buffer. Certainly, new_window is new window.
  1189. * We can advertise it now, if it is not less than current one.
  1190. * "Lots" means "at least twice" here.
  1191. */
  1192. if (new_window && new_window >= 2 * rcv_window_now)
  1193. time_to_ack = true;
  1194. }
  1195. }
  1196. if (time_to_ack)
  1197. tcp_send_ack(sk);
  1198. }
  1199. static void tcp_prequeue_process(struct sock *sk)
  1200. {
  1201. struct sk_buff *skb;
  1202. struct tcp_sock *tp = tcp_sk(sk);
  1203. NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPPREQUEUED);
  1204. /* RX process wants to run with disabled BHs, though it is not
  1205. * necessary */
  1206. local_bh_disable();
  1207. while ((skb = __skb_dequeue(&tp->ucopy.prequeue)) != NULL)
  1208. sk_backlog_rcv(sk, skb);
  1209. local_bh_enable();
  1210. /* Clear memory counter. */
  1211. tp->ucopy.memory = 0;
  1212. }
  1213. #ifdef CONFIG_NET_DMA
  1214. static void tcp_service_net_dma(struct sock *sk, bool wait)
  1215. {
  1216. dma_cookie_t done, used;
  1217. dma_cookie_t last_issued;
  1218. struct tcp_sock *tp = tcp_sk(sk);
  1219. if (!tp->ucopy.dma_chan)
  1220. return;
  1221. last_issued = tp->ucopy.dma_cookie;
  1222. dma_async_issue_pending(tp->ucopy.dma_chan);
  1223. do {
  1224. if (dma_async_is_tx_complete(tp->ucopy.dma_chan,
  1225. last_issued, &done,
  1226. &used) == DMA_SUCCESS) {
  1227. /* Safe to free early-copied skbs now */
  1228. __skb_queue_purge(&sk->sk_async_wait_queue);
  1229. break;
  1230. } else {
  1231. struct sk_buff *skb;
  1232. while ((skb = skb_peek(&sk->sk_async_wait_queue)) &&
  1233. (dma_async_is_complete(skb->dma_cookie, done,
  1234. used) == DMA_SUCCESS)) {
  1235. __skb_dequeue(&sk->sk_async_wait_queue);
  1236. kfree_skb(skb);
  1237. }
  1238. }
  1239. } while (wait);
  1240. }
  1241. #endif
  1242. static struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off)
  1243. {
  1244. struct sk_buff *skb;
  1245. u32 offset;
  1246. while ((skb = skb_peek(&sk->sk_receive_queue)) != NULL) {
  1247. offset = seq - TCP_SKB_CB(skb)->seq;
  1248. if (tcp_hdr(skb)->syn)
  1249. offset--;
  1250. if (offset < skb->len || tcp_hdr(skb)->fin) {
  1251. *off = offset;
  1252. return skb;
  1253. }
  1254. /* This looks weird, but this can happen if TCP collapsing
  1255. * splitted a fat GRO packet, while we released socket lock
  1256. * in skb_splice_bits()
  1257. */
  1258. sk_eat_skb(sk, skb, false);
  1259. }
  1260. return NULL;
  1261. }
  1262. /*
  1263. * This routine provides an alternative to tcp_recvmsg() for routines
  1264. * that would like to handle copying from skbuffs directly in 'sendfile'
  1265. * fashion.
  1266. * Note:
  1267. * - It is assumed that the socket was locked by the caller.
  1268. * - The routine does not block.
  1269. * - At present, there is no support for reading OOB data
  1270. * or for 'peeking' the socket using this routine
  1271. * (although both would be easy to implement).
  1272. */
  1273. int tcp_read_sock(struct sock *sk, read_descriptor_t *desc,
  1274. sk_read_actor_t recv_actor)
  1275. {
  1276. struct sk_buff *skb;
  1277. struct tcp_sock *tp = tcp_sk(sk);
  1278. u32 seq = tp->copied_seq;
  1279. u32 offset;
  1280. int copied = 0;
  1281. if (sk->sk_state == TCP_LISTEN)
  1282. return -ENOTCONN;
  1283. while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
  1284. if (offset < skb->len) {
  1285. int used;
  1286. size_t len;
  1287. len = skb->len - offset;
  1288. /* Stop reading if we hit a patch of urgent data */
  1289. if (tp->urg_data) {
  1290. u32 urg_offset = tp->urg_seq - seq;
  1291. if (urg_offset < len)
  1292. len = urg_offset;
  1293. if (!len)
  1294. break;
  1295. }
  1296. used = recv_actor(desc, skb, offset, len);
  1297. if (used <= 0) {
  1298. if (!copied)
  1299. copied = used;
  1300. break;
  1301. } else if (used <= len) {
  1302. seq += used;
  1303. copied += used;
  1304. offset += used;
  1305. }
  1306. /* If recv_actor drops the lock (e.g. TCP splice
  1307. * receive) the skb pointer might be invalid when
  1308. * getting here: tcp_collapse might have deleted it
  1309. * while aggregating skbs from the socket queue.
  1310. */
  1311. skb = tcp_recv_skb(sk, seq - 1, &offset);
  1312. if (!skb)
  1313. break;
  1314. /* TCP coalescing might have appended data to the skb.
  1315. * Try to splice more frags
  1316. */
  1317. if (offset + 1 != skb->len)
  1318. continue;
  1319. }
  1320. if (tcp_hdr(skb)->fin) {
  1321. sk_eat_skb(sk, skb, false);
  1322. ++seq;
  1323. break;
  1324. }
  1325. sk_eat_skb(sk, skb, false);
  1326. if (!desc->count)
  1327. break;
  1328. tp->copied_seq = seq;
  1329. }
  1330. tp->copied_seq = seq;
  1331. tcp_rcv_space_adjust(sk);
  1332. /* Clean up data we have read: This will do ACK frames. */
  1333. if (copied > 0) {
  1334. tcp_recv_skb(sk, seq, &offset);
  1335. tcp_cleanup_rbuf(sk, copied);
  1336. }
  1337. return copied;
  1338. }
  1339. EXPORT_SYMBOL(tcp_read_sock);
  1340. /*
  1341. * This routine copies from a sock struct into the user buffer.
  1342. *
  1343. * Technical note: in 2.3 we work on _locked_ socket, so that
  1344. * tricks with *seq access order and skb->users are not required.
  1345. * Probably, code can be easily improved even more.
  1346. */
  1347. int tcp_recvmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg,
  1348. size_t len, int nonblock, int flags, int *addr_len)
  1349. {
  1350. struct tcp_sock *tp = tcp_sk(sk);
  1351. int copied = 0;
  1352. u32 peek_seq;
  1353. u32 *seq;
  1354. unsigned long used;
  1355. int err;
  1356. int target; /* Read at least this many bytes */
  1357. long timeo;
  1358. struct task_struct *user_recv = NULL;
  1359. bool copied_early = false;
  1360. struct sk_buff *skb;
  1361. u32 urg_hole = 0;
  1362. lock_sock(sk);
  1363. err = -ENOTCONN;
  1364. if (sk->sk_state == TCP_LISTEN)
  1365. goto out;
  1366. timeo = sock_rcvtimeo(sk, nonblock);
  1367. /* Urgent data needs to be handled specially. */
  1368. if (flags & MSG_OOB)
  1369. goto recv_urg;
  1370. if (unlikely(tp->repair)) {
  1371. err = -EPERM;
  1372. if (!(flags & MSG_PEEK))
  1373. goto out;
  1374. if (tp->repair_queue == TCP_SEND_QUEUE)
  1375. goto recv_sndq;
  1376. err = -EINVAL;
  1377. if (tp->repair_queue == TCP_NO_QUEUE)
  1378. goto out;
  1379. /* 'common' recv queue MSG_PEEK-ing */
  1380. }
  1381. seq = &tp->copied_seq;
  1382. if (flags & MSG_PEEK) {
  1383. peek_seq = tp->copied_seq;
  1384. seq = &peek_seq;
  1385. }
  1386. target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
  1387. #ifdef CONFIG_NET_DMA
  1388. tp->ucopy.dma_chan = NULL;
  1389. preempt_disable();
  1390. skb = skb_peek_tail(&sk->sk_receive_queue);
  1391. {
  1392. int available = 0;
  1393. if (skb)
  1394. available = TCP_SKB_CB(skb)->seq + skb->len - (*seq);
  1395. if ((available < target) &&
  1396. (len > sysctl_tcp_dma_copybreak) && !(flags & MSG_PEEK) &&
  1397. !sysctl_tcp_low_latency &&
  1398. net_dma_find_channel()) {
  1399. preempt_enable_no_resched();
  1400. tp->ucopy.pinned_list =
  1401. dma_pin_iovec_pages(msg->msg_iov, len);
  1402. } else {
  1403. preempt_enable_no_resched();
  1404. }
  1405. }
  1406. #endif
  1407. do {
  1408. u32 offset;
  1409. /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
  1410. if (tp->urg_data && tp->urg_seq == *seq) {
  1411. if (copied)
  1412. break;
  1413. if (signal_pending(current)) {
  1414. copied = timeo ? sock_intr_errno(timeo) : -EAGAIN;
  1415. break;
  1416. }
  1417. }
  1418. /* Next get a buffer. */
  1419. skb_queue_walk(&sk->sk_receive_queue, skb) {
  1420. /* Now that we have two receive queues this
  1421. * shouldn't happen.
  1422. */
  1423. if (WARN(before(*seq, TCP_SKB_CB(skb)->seq),
  1424. "recvmsg bug: copied %X seq %X rcvnxt %X fl %X\n",
  1425. *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt,
  1426. flags))
  1427. break;
  1428. offset = *seq - TCP_SKB_CB(skb)->seq;
  1429. if (tcp_hdr(skb)->syn)
  1430. offset--;
  1431. if (offset < skb->len)
  1432. goto found_ok_skb;
  1433. if (tcp_hdr(skb)->fin)
  1434. goto found_fin_ok;
  1435. WARN(!(flags & MSG_PEEK),
  1436. "recvmsg bug 2: copied %X seq %X rcvnxt %X fl %X\n",
  1437. *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt, flags);
  1438. }
  1439. /* Well, if we have backlog, try to process it now yet. */
  1440. if (copied >= target && !sk->sk_backlog.tail)
  1441. break;
  1442. if (copied) {
  1443. if (sk->sk_err ||
  1444. sk->sk_state == TCP_CLOSE ||
  1445. (sk->sk_shutdown & RCV_SHUTDOWN) ||
  1446. !timeo ||
  1447. signal_pending(current))
  1448. break;
  1449. } else {
  1450. if (sock_flag(sk, SOCK_DONE))
  1451. break;
  1452. if (sk->sk_err) {
  1453. copied = sock_error(sk);
  1454. break;
  1455. }
  1456. if (sk->sk_shutdown & RCV_SHUTDOWN)
  1457. break;
  1458. if (sk->sk_state == TCP_CLOSE) {
  1459. if (!sock_flag(sk, SOCK_DONE)) {
  1460. /* This occurs when user tries to read
  1461. * from never connected socket.
  1462. */
  1463. copied = -ENOTCONN;
  1464. break;
  1465. }
  1466. break;
  1467. }
  1468. if (!timeo) {
  1469. copied = -EAGAIN;
  1470. break;
  1471. }
  1472. if (signal_pending(current)) {
  1473. copied = sock_intr_errno(timeo);
  1474. break;
  1475. }
  1476. }
  1477. tcp_cleanup_rbuf(sk, copied);
  1478. if (!sysctl_tcp_low_latency && tp->ucopy.task == user_recv) {
  1479. /* Install new reader */
  1480. if (!user_recv && !(flags & (MSG_TRUNC | MSG_PEEK))) {
  1481. user_recv = current;
  1482. tp->ucopy.task = user_recv;
  1483. tp->ucopy.iov = msg->msg_iov;
  1484. }
  1485. tp->ucopy.len = len;
  1486. WARN_ON(tp->copied_seq != tp->rcv_nxt &&
  1487. !(flags & (MSG_PEEK | MSG_TRUNC)));
  1488. /* Ugly... If prequeue is not empty, we have to
  1489. * process it before releasing socket, otherwise
  1490. * order will be broken at second iteration.
  1491. * More elegant solution is required!!!
  1492. *
  1493. * Look: we have the following (pseudo)queues:
  1494. *
  1495. * 1. packets in flight
  1496. * 2. backlog
  1497. * 3. prequeue
  1498. * 4. receive_queue
  1499. *
  1500. * Each queue can be processed only if the next ones
  1501. * are empty. At this point we have empty receive_queue.
  1502. * But prequeue _can_ be not empty after 2nd iteration,
  1503. * when we jumped to start of loop because backlog
  1504. * processing added something to receive_queue.
  1505. * We cannot release_sock(), because backlog contains
  1506. * packets arrived _after_ prequeued ones.
  1507. *
  1508. * Shortly, algorithm is clear --- to process all
  1509. * the queues in order. We could make it more directly,
  1510. * requeueing packets from backlog to prequeue, if
  1511. * is not empty. It is more elegant, but eats cycles,
  1512. * unfortunately.
  1513. */
  1514. if (!skb_queue_empty(&tp->ucopy.prequeue))
  1515. goto do_prequeue;
  1516. /* __ Set realtime policy in scheduler __ */
  1517. }
  1518. #ifdef CONFIG_NET_DMA
  1519. if (tp->ucopy.dma_chan) {
  1520. if (tp->rcv_wnd == 0 &&
  1521. !skb_queue_empty(&sk->sk_async_wait_queue)) {
  1522. tcp_service_net_dma(sk, true);
  1523. tcp_cleanup_rbuf(sk, copied);
  1524. } else
  1525. dma_async_issue_pending(tp->ucopy.dma_chan);
  1526. }
  1527. #endif
  1528. if (copied >= target) {
  1529. /* Do not sleep, just process backlog. */
  1530. release_sock(sk);
  1531. lock_sock(sk);
  1532. } else
  1533. sk_wait_data(sk, &timeo);
  1534. #ifdef CONFIG_NET_DMA
  1535. tcp_service_net_dma(sk, false); /* Don't block */
  1536. tp->ucopy.wakeup = 0;
  1537. #endif
  1538. if (user_recv) {
  1539. int chunk;
  1540. /* __ Restore normal policy in scheduler __ */
  1541. if ((chunk = len - tp->ucopy.len) != 0) {
  1542. NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMBACKLOG, chunk);
  1543. len -= chunk;
  1544. copied += chunk;
  1545. }
  1546. if (tp->rcv_nxt == tp->copied_seq &&
  1547. !skb_queue_empty(&tp->ucopy.prequeue)) {
  1548. do_prequeue:
  1549. tcp_prequeue_process(sk);
  1550. if ((chunk = len - tp->ucopy.len) != 0) {
  1551. NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
  1552. len -= chunk;
  1553. copied += chunk;
  1554. }
  1555. }
  1556. }
  1557. if ((flags & MSG_PEEK) &&
  1558. (peek_seq - copied - urg_hole != tp->copied_seq)) {
  1559. net_dbg_ratelimited("TCP(%s:%d): Application bug, race in MSG_PEEK\n",
  1560. current->comm,
  1561. task_pid_nr(current));
  1562. peek_seq = tp->copied_seq;
  1563. }
  1564. continue;
  1565. found_ok_skb:
  1566. /* Ok so how much can we use? */
  1567. used = skb->len - offset;
  1568. if (len < used)
  1569. used = len;
  1570. /* Do we have urgent data here? */
  1571. if (tp->urg_data) {
  1572. u32 urg_offset = tp->urg_seq - *seq;
  1573. if (urg_offset < used) {
  1574. if (!urg_offset) {
  1575. if (!sock_flag(sk, SOCK_URGINLINE)) {
  1576. ++*seq;
  1577. urg_hole++;
  1578. offset++;
  1579. used--;
  1580. if (!used)
  1581. goto skip_copy;
  1582. }
  1583. } else
  1584. used = urg_offset;
  1585. }
  1586. }
  1587. if (!(flags & MSG_TRUNC)) {
  1588. #ifdef CONFIG_NET_DMA
  1589. if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list)
  1590. tp->ucopy.dma_chan = net_dma_find_channel();
  1591. if (tp->ucopy.dma_chan) {
  1592. tp->ucopy.dma_cookie = dma_skb_copy_datagram_iovec(
  1593. tp->ucopy.dma_chan, skb, offset,
  1594. msg->msg_iov, used,
  1595. tp->ucopy.pinned_list);
  1596. if (tp->ucopy.dma_cookie < 0) {
  1597. pr_alert("%s: dma_cookie < 0\n",
  1598. __func__);
  1599. /* Exception. Bailout! */
  1600. if (!copied)
  1601. copied = -EFAULT;
  1602. break;
  1603. }
  1604. dma_async_issue_pending(tp->ucopy.dma_chan

Large files files are truncated, but you can click here to view the full file