db_sockets.c 43 KB

12345678910111213141516171819202122232425262728293031323334353637383940414243444546474849505152535455565758596061626364656667686970717273747576777879808182838485868788899091929394959697989910010110210310410510610710810911011111211311411511611711811912012112212312412512612712812913013113213313413513613713813914014114214314414514614714814915015115215315415515615715815916016116216316416516616716816917017117217317417517617717817918018118218318418518618718818919019119219319419519619719819920020120220320420520620720820921021121221321421521621721821922022122222322422522622722822923023123223323423523623723823924024124224324424524624724824925025125225325425525625725825926026126226326426526626726826927027127227327427527627727827928028128228328428528628728828929029129229329429529629729829930030130230330430530630730830931031131231331431531631731831932032132232332432532632732832933033133233333433533633733833934034134234334434534634734834935035135235335435535635735835936036136236336436536636736836937037137237337437537637737837938038138238338438538638738838939039139239339439539639739839940040140240340440540640740840941041141241341441541641741841942042142242342442542642742842943043143243343443543643743843944044144244344444544644744844945045145245345445545645745845946046146246346446546646746846947047147247347447547647747847948048148248348448548648748848949049149249349449549649749849950050150250350450550650750850951051151251351451551651751851952052152252352452552652752852953053153253353453553653753853954054154254354454554654754854955055155255355455555655755855956056156256356456556656756856957057157257357457557657757857958058158258358458558658758858959059159259359459559659759859960060160260360460560660760860961061161261361461561661761861962062162262362462562662762862963063163263363463563663763863964064164264364464564664764864965065165265365465565665765865966066166266366466566666766866967067167267367467567667767867968068168268368468568668768868969069169269369469569669769869970070170270370470570670770870971071171271371471571671771871972072172272372472572672772872973073173273373473573673773873974074174274374474574674774874975075175275375475575675775875976076176276376476576676776876977077177277377477577677777877978078178278378478578678778878979079179279379479579679779879980080180280380480580680780880981081181281381481581681781881982082182282382482582682782882983083183283383483583683783883984084184284384484584684784884985085185285385485585685785885986086186286386486586686786886987087187287387487587687787887988088188288388488588688788888989089189289389489589689789889990090190290390490590690790890991091191291391491591691791891992092192292392492592692792892993093193293393493593693793893994094194294394494594694794894995095195295395495595695795895996096196296396496596696796896997097197297397497597697797897998098198298398498598698798898999099199299399499599699799899910001001100210031004100510061007100810091010101110121013101410151016101710181019102010211022102310241025102610271028102910301031103210331034103510361037103810391040104110421043104410451046104710481049105010511052105310541055105610571058105910601061106210631064106510661067106810691070107110721073107410751076107710781079108010811082108310841085108610871088108910901091109210931094109510961097109810991100110111021103110411051106110711081109111011111112111311141115111611171118111911201121112211231124112511261127112811291130113111321133113411351136113711381139114011411142114311441145114611471148114911501151115211531154115511561157115811591160116111621163116411651166116711681169117011711172117311741175117611771178117911801181118211831184118511861187118811891190119111921193119411951196119711981199120012011202120312041205120612071208120912101211121212131214121512161217121812191220122112221223122412251226122712281229123012311232123312341235123612371238123912401241124212431244124512461247124812491250125112521253125412551256125712581259126012611262126312641265126612671268126912701271127212731274127512761277127812791280128112821283128412851286128712881289129012911292129312941295129612971298129913001301130213031304130513061307130813091310131113121313131413151316131713181319132013211322132313241325132613271328132913301331133213331334133513361337133813391340134113421343134413451346134713481349135013511352135313541355135613571358135913601361136213631364136513661367136813691370137113721373137413751376137713781379138013811382138313841385138613871388138913901391139213931394139513961397139813991400140114021403140414051406140714081409141014111412141314141415141614171418141914201421142214231424142514261427142814291430143114321433143414351436143714381439144014411442144314441445144614471448144914501451145214531454145514561457145814591460146114621463146414651466146714681469147014711472147314741475147614771478147914801481148214831484148514861487148814891490149114921493
  1. /* -*- mode:c; c-file-style:"k&r"; c-basic-offset: 4; tab-width:4; indent-tabs-mode:nil; mode:auto-fill; fill-column:78; -*- */
  2. /* vim: set ts=4 sw=4 et tw=78 fo=cqt wm=0: */
  3. /* Copyright (C) 2014 Stony Brook University
  4. This file is part of Graphene Library OS.
  5. Graphene Library OS is free software: you can redistribute it and/or
  6. modify it under the terms of the GNU Lesser General Public License
  7. as published by the Free Software Foundation, either version 3 of the
  8. License, or (at your option) any later version.
  9. Graphene Library OS is distributed in the hope that it will be useful,
  10. but WITHOUT ANY WARRANTY; without even the implied warranty of
  11. MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
  12. GNU Lesser General Public License for more details.
  13. You should have received a copy of the GNU Lesser General Public License
  14. along with this program. If not, see <http://www.gnu.org/licenses/>. */
  15. /*
  16. * db_socket.c
  17. *
  18. * This file contains operands for streams with URIs that start with
  19. * "tcp:", "tcp.srv:", "udp:", "udp.srv:".
  20. */
  21. #include "pal_defs.h"
  22. #include "pal_linux_defs.h"
  23. #include "pal.h"
  24. #include "pal_internal.h"
  25. #include "pal_linux.h"
  26. #include "pal_debug.h"
  27. #include "pal_security.h"
  28. #include "pal_error.h"
  29. #include "api.h"
  30. #include "graphene.h"
  31. #include <linux/types.h>
  32. #include <linux/poll.h>
  33. typedef __kernel_pid_t pid_t;
  34. #include <asm/fcntl.h>
  35. #include <sys/socket.h>
  36. #include <linux/in.h>
  37. #include <linux/in6.h>
  38. #include <netinet/tcp.h>
  39. #include <asm/errno.h>
  40. #ifndef SOL_TCP
  41. # define SOL_TCP 6
  42. #endif
  43. #ifndef TCP_NODELAY
  44. # define TCP_NODELAY 1
  45. #endif
  46. #ifndef TCP_CORK
  47. # define TCP_CORK 3
  48. #endif
  49. #ifndef SOL_IPV6
  50. # define SOL_IPV6 41
  51. #endif
  52. /* 96 bytes is the minimal size of buffer to store a IPv4/IPv6
  53. address */
  54. #define PAL_SOCKADDR_SIZE 96
  55. static inline int addr_size (struct sockaddr * addr)
  56. {
  57. switch (addr->sa_family) {
  58. case AF_INET:
  59. return sizeof(struct sockaddr_in);
  60. case AF_INET6:
  61. return sizeof(struct sockaddr_in6);
  62. default:
  63. return 0;
  64. }
  65. }
  66. /* parsing the string of uri, and fill in the socket address structure.
  67. the latest pointer of uri, length of socket address are returned. */
  68. static int inet_parse_uri (char ** uri, struct sockaddr * addr, int * addrlen)
  69. {
  70. char * tmp = *uri, * end;
  71. char * addr_str = NULL, * port_str;
  72. int af;
  73. void * addr_buf;
  74. int addr_len;
  75. __be16 * port_buf;
  76. int slen;
  77. if (tmp[0] == '[') {
  78. /* for IPv6, the address will be in the form of
  79. "[xxxx:xxxx:xxxx:xxxx:xxxx:xxxx:xxxx:xxxx]:port". */
  80. struct sockaddr_in6 * addr_in6 = (struct sockaddr_in6 *) addr;
  81. slen = sizeof(struct sockaddr_in6);
  82. memset(addr, 0, slen);
  83. end = strchr(tmp + 1, ']');
  84. if (!end || *(end + 1) != ':')
  85. goto inval;
  86. addr_str = tmp + 1;
  87. addr_len = end - tmp - 1;
  88. port_str = end + 2;
  89. for (end = port_str ; *end >= '0' && *end <= '9' ; end++);
  90. addr_in6->sin6_family = af = AF_INET6;
  91. addr_buf = &addr_in6->sin6_addr.s6_addr;
  92. port_buf = &addr_in6->sin6_port;
  93. } else {
  94. /* for IP, the address will be in the form of "x.x.x.x:port". */
  95. struct sockaddr_in * addr_in = (struct sockaddr_in *) addr;
  96. slen = sizeof(struct sockaddr_in);
  97. memset(addr, 0, slen);
  98. end = strchr(tmp, ':');
  99. if (!end)
  100. goto inval;
  101. addr_str = tmp;
  102. addr_len = end - tmp;
  103. port_str = end + 1;
  104. for (end = port_str ; *end >= '0' && *end <= '9' ; end++);
  105. addr_in->sin_family = af = AF_INET;
  106. addr_buf = &addr_in->sin_addr.s_addr;
  107. port_buf = &addr_in->sin_port;
  108. }
  109. if (af == AF_INET) {
  110. if (!inet_pton4(addr_str, addr_len, addr_buf))
  111. goto inval;
  112. } else {
  113. if (!inet_pton6(addr_str, addr_len, addr_buf))
  114. goto inval;
  115. }
  116. *port_buf = __htons(atoi(port_str));
  117. *uri = *end ? end + 1 : NULL;
  118. if (addrlen)
  119. *addrlen = slen;
  120. return 0;
  121. inval:
  122. return -PAL_ERROR_INVAL;
  123. }
  124. /* create the string of uri from the given socket address */
  125. static int inet_create_uri (char * uri, int count, struct sockaddr * addr,
  126. int addrlen)
  127. {
  128. int len = 0;
  129. if (addr->sa_family == AF_INET) {
  130. if (addrlen != sizeof(struct sockaddr_in))
  131. return PAL_ERROR_INVAL;
  132. struct sockaddr_in * addr_in = (struct sockaddr_in *) addr;
  133. char * addr = (char *) &addr_in->sin_addr.s_addr;
  134. /* for IP, the address will be in the form of "x.x.x.x:port". */
  135. len = snprintf(uri, count, "%u.%u.%u.%u:%u",
  136. (unsigned char) addr[0],
  137. (unsigned char) addr[1],
  138. (unsigned char) addr[2],
  139. (unsigned char) addr[3],
  140. __ntohs(addr_in->sin_port));
  141. } else if (addr->sa_family == AF_INET6) {
  142. if (addrlen != sizeof(struct sockaddr_in6))
  143. return PAL_ERROR_INVAL;
  144. struct sockaddr_in6 * addr_in6 = (struct sockaddr_in6 *) addr;
  145. unsigned short * addr = (unsigned short *) &addr_in6->sin6_addr.s6_addr;
  146. /* for IPv6, the address will be in the form of
  147. "[xxxx:xxxx:xxxx:xxxx:xxxx:xxxx:xxxx:xxxx]:port". */
  148. len = snprintf(uri, count, "[%04x:%04x:%04x:%04x:%04x:%04x:%04x:%04x]:%u",
  149. addr[0], addr[1], addr[2], addr[3],
  150. addr[4], addr[5], addr[6], addr[7],
  151. __ntohs(addr_in6->sin6_port));
  152. } else {
  153. return -PAL_ERROR_INVAL;
  154. }
  155. if (len >= count)
  156. return -PAL_ERROR_TOOLONG;
  157. return len;
  158. }
  159. /* parse the uri for a socket stream. The uri might have both binding
  160. address and connecting address, or connecting address only. The form
  161. of uri will be either "bind-addr:bind-port:connect-addr:connect-port"
  162. or "addr:port". */
  163. static int socket_parse_uri (char * uri,
  164. struct sockaddr ** bind_addr, int * bind_addrlen,
  165. struct sockaddr ** dest_addr, int * dest_addrlen)
  166. {
  167. int ret;
  168. if (!bind_addr && !dest_addr)
  169. return 0;
  170. if (!uri || !(*uri)) {
  171. if (bind_addr)
  172. *bind_addr = NULL;
  173. if (dest_addr)
  174. *dest_addr = NULL;
  175. return 0;
  176. }
  177. /* at least parse uri once */
  178. if ((ret = inet_parse_uri(&uri, bind_addr ? *bind_addr : *dest_addr,
  179. bind_addr ? bind_addrlen : dest_addrlen)) < 0)
  180. return ret;
  181. if (!(bind_addr && dest_addr))
  182. return 0;
  183. /* if you reach here, it can only be connection address */
  184. if (!uri || (ret = inet_parse_uri(&uri, *dest_addr, dest_addrlen)) < 0) {
  185. *dest_addr = *bind_addr;
  186. *dest_addrlen = *bind_addrlen;
  187. *bind_addr = NULL;
  188. *bind_addrlen = 0;
  189. }
  190. return 0;
  191. }
  192. /* fill in the PAL handle based on the file descriptors and address given. */
  193. static inline
  194. PAL_HANDLE socket_create_handle (int type, int fd, int options,
  195. struct sockaddr * bind_addr, int bind_addrlen,
  196. struct sockaddr * dest_addr, int dest_addrlen)
  197. {
  198. PAL_HANDLE hdl = malloc(HANDLE_SIZE(sock) + (bind_addr ? bind_addrlen : 0) +
  199. (dest_addr ? dest_addrlen : 0));
  200. if (!hdl)
  201. return NULL;
  202. memset(hdl, 0, sizeof(struct pal_handle));
  203. init_handle_hdr(HANDLE_HDR(hdl), type);
  204. HANDLE_HDR(hdl)->flags |= RFD(0)|(type != pal_type_tcpsrv ? WFD(0) : 0);
  205. hdl->sock.fd = fd;
  206. void * addr = (void *) hdl + HANDLE_SIZE(sock);
  207. if (bind_addr) {
  208. hdl->sock.bind = (PAL_PTR) addr;
  209. memcpy(addr, bind_addr, bind_addrlen);
  210. addr += bind_addrlen;
  211. } else {
  212. hdl->sock.bind = (PAL_PTR) NULL;
  213. }
  214. if (dest_addr) {
  215. hdl->sock.conn = (PAL_PTR) addr;
  216. memcpy(addr, dest_addr, dest_addrlen);
  217. addr += dest_addrlen;
  218. } else {
  219. hdl->sock.conn = (PAL_PTR) NULL;
  220. }
  221. hdl->sock.nonblocking = (options & PAL_OPTION_NONBLOCK) ?
  222. PAL_TRUE : PAL_FALSE;
  223. hdl->sock.linger = 0;
  224. if (type == pal_type_tcpsrv) {
  225. hdl->sock.receivebuf = 0;
  226. hdl->sock.sendbuf = 0;
  227. } else {
  228. int ret, val, len = sizeof(int);
  229. ret = INLINE_SYSCALL(getsockopt, 5, fd, SOL_SOCKET, SO_RCVBUF,
  230. &val, &len);
  231. hdl->sock.receivebuf = IS_ERR(ret) ? 0 : val;
  232. ret = INLINE_SYSCALL(getsockopt, 5, fd, SOL_SOCKET, SO_SNDBUF,
  233. &val, &len);
  234. hdl->sock.sendbuf = IS_ERR(ret) ? 0 : val;
  235. }
  236. hdl->sock.receivetimeout = 0;
  237. hdl->sock.sendtimeout = 0;
  238. hdl->sock.tcp_cork = PAL_FALSE;
  239. hdl->sock.tcp_keepalive = PAL_FALSE;
  240. hdl->sock.tcp_nodelay = PAL_FALSE;
  241. return hdl;
  242. }
  243. #if ALLOW_BIND_ANY == 0
  244. static bool check_zero (void * mem, size_t size)
  245. {
  246. void * p = mem, * q = mem + size;
  247. while (p < q) {
  248. if (p <= q - sizeof(long)) {
  249. if (*(long *) p)
  250. return false;
  251. p += sizeof(long);
  252. } else if (p <= q - sizeof(int)) {
  253. if (*(int *) p)
  254. return false;
  255. p += sizeof(int);
  256. } else if (p <= q - sizeof(short)) {
  257. if (*(short *) p)
  258. return false;
  259. p += sizeof(short);
  260. } else {
  261. if (*(char *) p)
  262. return false;
  263. p++;
  264. }
  265. }
  266. return true;
  267. }
  268. /* check if an address is "Any" */
  269. static bool check_any_addr (struct sockaddr * addr)
  270. {
  271. if (addr->sa_family == AF_INET) {
  272. struct sockaddr_in * addr_in =
  273. (struct sockaddr_in *) addr;
  274. return addr_in->sin_port == 0 &&
  275. check_zero(&addr_in->sin_addr,
  276. sizeof(addr_in->sin_addr));
  277. } else if (addr->sa_family == AF_INET6) {
  278. struct sockaddr_in6 * addr_in6 =
  279. (struct sockaddr_in6 *) addr;
  280. return addr_in6->sin6_port == 0 &&
  281. check_zero(&addr_in6->sin6_addr,
  282. sizeof(addr_in6->sin6_addr));
  283. }
  284. return false;
  285. }
  286. #endif
  287. /* listen on a tcp socket */
  288. static int tcp_listen (PAL_HANDLE * handle, char * uri, int options)
  289. {
  290. struct sockaddr buffer, * bind_addr = &buffer;
  291. int bind_addrlen;
  292. int ret, fd = -1;
  293. if ((ret = socket_parse_uri(uri, &bind_addr, &bind_addrlen,
  294. NULL, NULL)) < 0)
  295. return ret;
  296. assert(bind_addr);
  297. assert(bind_addrlen == addr_size(bind_addr));
  298. #if ALLOW_BIND_ANY == 0
  299. /* the socket need to have a binding address, a null address or an
  300. any address is not allowed */
  301. if (addr_check_any(bind_addr))
  302. return -PAL_ERROR_INVAL;
  303. #endif
  304. fd = INLINE_SYSCALL(socket, 3, bind_addr->sa_family,
  305. SOCK_STREAM|SOCK_CLOEXEC|options, 0);
  306. if (IS_ERR(fd))
  307. return -PAL_ERROR_DENIED;
  308. if (bind_addr->sa_family == AF_INET6) {
  309. int ipv6only = 1;
  310. INLINE_SYSCALL(setsockopt, 5, fd, SOL_IPV6, IPV6_V6ONLY, &ipv6only,
  311. sizeof(int));
  312. }
  313. /* must set the socket to be reuseable */
  314. int reuseaddr = 1;
  315. INLINE_SYSCALL(setsockopt, 5, fd, SOL_SOCKET, SO_REUSEADDR, &reuseaddr,
  316. sizeof(int));
  317. ret = INLINE_SYSCALL(bind, 3, fd, bind_addr, bind_addrlen);
  318. if (IS_ERR(ret)) {
  319. switch(ERRNO(ret)) {
  320. case EINVAL:
  321. ret = -PAL_ERROR_INVAL;
  322. goto failed;
  323. case EADDRINUSE:
  324. ret = -PAL_ERROR_STREAMEXIST;
  325. goto failed;
  326. default:
  327. ret = -PAL_ERROR_DENIED;
  328. goto failed;
  329. }
  330. }
  331. ret = INLINE_SYSCALL(listen, 2, fd, DEFAULT_BACKLOG);
  332. if (IS_ERR(ret))
  333. return -PAL_ERROR_DENIED;
  334. *handle = socket_create_handle(pal_type_tcpsrv, fd, options,
  335. bind_addr, bind_addrlen, NULL, 0);
  336. if (!(*handle)) {
  337. ret = -PAL_ERROR_NOMEM;
  338. goto failed;
  339. }
  340. return 0;
  341. failed:
  342. INLINE_SYSCALL(close, 1, fd);
  343. return ret;
  344. }
  345. /* accept a tcp connection */
  346. static int tcp_accept (PAL_HANDLE handle, PAL_HANDLE * client)
  347. {
  348. if (!IS_HANDLE_TYPE(handle, tcpsrv) ||
  349. !handle->sock.bind || handle->sock.conn)
  350. return -PAL_ERROR_NOTSERVER;
  351. if (handle->sock.fd == PAL_IDX_POISON)
  352. return -PAL_ERROR_BADHANDLE;
  353. struct sockaddr * bind_addr = (struct sockaddr *) handle->sock.bind;
  354. int bind_addrlen = addr_size(bind_addr);
  355. struct sockaddr buffer;
  356. socklen_t addrlen = sizeof(struct sockaddr);
  357. int ret = 0;
  358. int newfd = INLINE_SYSCALL(accept4, 4, handle->sock.fd, &buffer,
  359. &addrlen, O_CLOEXEC);
  360. if (IS_ERR(newfd))
  361. switch(ERRNO(newfd)) {
  362. case EWOULDBLOCK:
  363. return -PAL_ERROR_TRYAGAIN;
  364. case ECONNABORTED:
  365. return -PAL_ERROR_STREAMNOTEXIST;
  366. default:
  367. return unix_to_pal_error(ERRNO(newfd));
  368. }
  369. struct sockaddr * dest_addr = &buffer;
  370. int dest_addrlen = addrlen;
  371. *client = socket_create_handle(pal_type_tcp, newfd, 0,
  372. bind_addr, bind_addrlen,
  373. dest_addr, dest_addrlen);
  374. if (!(*client)) {
  375. ret = -PAL_ERROR_NOMEM;
  376. goto failed;
  377. }
  378. return 0;
  379. failed:
  380. INLINE_SYSCALL(close, 1, newfd);
  381. return ret;
  382. }
  383. /* connect on a tcp socket */
  384. static int tcp_connect (PAL_HANDLE * handle, char * uri, int options)
  385. {
  386. struct sockaddr buffer[3];
  387. struct sockaddr * bind_addr = buffer, * dest_addr = buffer + 1;
  388. int bind_addrlen, dest_addrlen;
  389. int ret, fd = -1;
  390. /* accepting two kind of different uri:
  391. dest-ip:dest-port or bind-ip:bind-port:dest-ip:dest-port */
  392. if ((ret = socket_parse_uri(uri, &bind_addr, &bind_addrlen,
  393. &dest_addr, &dest_addrlen)) < 0)
  394. return ret;
  395. if (!dest_addr)
  396. return -PAL_ERROR_INVAL;
  397. if (bind_addr && bind_addr->sa_family != dest_addr->sa_family)
  398. return -PAL_ERROR_INVAL;
  399. fd = INLINE_SYSCALL(socket, 3, dest_addr->sa_family,
  400. SOCK_STREAM|SOCK_CLOEXEC|options, 0);
  401. if (IS_ERR(fd))
  402. return -PAL_ERROR_DENIED;
  403. if (bind_addr) {
  404. if (IS_ERR(ret)) {
  405. INLINE_SYSCALL(close, 1, fd);
  406. switch (ERRNO(ret)) {
  407. case EADDRINUSE:
  408. ret = -PAL_ERROR_STREAMEXIST;
  409. goto failed;
  410. case EADDRNOTAVAIL:
  411. ret = -PAL_ERROR_ADDRNOTEXIST;
  412. goto failed;
  413. default:
  414. ret = unix_to_pal_error(ERRNO(ret));
  415. goto failed;
  416. }
  417. }
  418. }
  419. if (dest_addr->sa_family == AF_INET6) {
  420. int ipv6only = 1;
  421. INLINE_SYSCALL(setsockopt, 5, fd, SOL_IPV6, IPV6_V6ONLY, &ipv6only,
  422. sizeof(int));
  423. }
  424. ret = INLINE_SYSCALL(connect, 3, fd, dest_addr, dest_addrlen);
  425. if (IS_ERR(ret) && ERRNO(ret) == EINPROGRESS) {
  426. struct pollfd pfd = { .fd = fd, .events = POLLOUT, .revents = 0 };
  427. ret = INLINE_SYSCALL(ppoll, 5, &pfd, 1, NULL, NULL, 0);
  428. }
  429. if (IS_ERR(ret)) {
  430. ret = unix_to_pal_error(ERRNO(ret));
  431. goto failed;
  432. }
  433. if (!bind_addr) {
  434. /* save some space to get socket address */
  435. bind_addr = buffer + 2;
  436. bind_addrlen = sizeof(struct sockaddr);
  437. /* call getsockname to get socket address */
  438. if ((ret = INLINE_SYSCALL(getsockname, 3, fd,
  439. bind_addr, &bind_addrlen)) < 0)
  440. bind_addr = NULL;
  441. }
  442. *handle = socket_create_handle(pal_type_tcp, fd, options,
  443. bind_addr, bind_addrlen,
  444. dest_addr, dest_addrlen);
  445. if (!(*handle)) {
  446. ret = -PAL_ERROR_NOMEM;
  447. goto failed;
  448. }
  449. return 0;
  450. failed:
  451. INLINE_SYSCALL(close, 1, fd);
  452. return ret;
  453. }
  454. /* 'open' operation of tcp stream */
  455. static int tcp_open (PAL_HANDLE *handle, const char * type, const char * uri,
  456. int access, int share, int create, int options)
  457. {
  458. int uri_len = strlen(uri) + 1;
  459. if (uri_len > PAL_SOCKADDR_SIZE)
  460. return -PAL_ERROR_TOOLONG;
  461. char uri_buf[PAL_SOCKADDR_SIZE];
  462. memcpy(uri_buf, uri, uri_len);
  463. if (strpartcmp_static(type, "tcp.srv:"))
  464. return tcp_listen(handle, uri_buf, options);
  465. if (strpartcmp_static(type, "tcp:"))
  466. return tcp_connect(handle, uri_buf, options);
  467. return -PAL_ERROR_NOTSUPPORT;
  468. }
  469. /* 'read' operation of tcp stream */
  470. static int64_t tcp_read (PAL_HANDLE handle, uint64_t offset, uint64_t len,
  471. void * buf)
  472. {
  473. if (!IS_HANDLE_TYPE(handle, tcp) || !handle->sock.conn)
  474. return -PAL_ERROR_NOTCONNECTION;
  475. if (handle->sock.fd == PAL_IDX_POISON)
  476. return -PAL_ERROR_ENDOFSTREAM;
  477. struct msghdr hdr;
  478. struct iovec iov;
  479. iov.iov_base = buf;
  480. iov.iov_len = len;
  481. hdr.msg_name = NULL;
  482. hdr.msg_namelen = 0;
  483. hdr.msg_iov = &iov;
  484. hdr.msg_iovlen = 1;
  485. hdr.msg_control = NULL;
  486. hdr.msg_controllen = 0;
  487. hdr.msg_flags = 0;
  488. int64_t bytes = INLINE_SYSCALL(recvmsg, 3, handle->sock.fd, &hdr, 0);
  489. if (IS_ERR(bytes))
  490. switch (ERRNO(bytes)) {
  491. case EWOULDBLOCK:
  492. return -PAL_ERROR_TRYAGAIN;
  493. default:
  494. return unix_to_pal_error(ERRNO(bytes));
  495. }
  496. if (!bytes)
  497. return -PAL_ERROR_ENDOFSTREAM;
  498. return bytes;
  499. }
  500. /* write' operation of tcp stream */
  501. static int64_t tcp_write (PAL_HANDLE handle, uint64_t offset, uint64_t len,
  502. const void * buf)
  503. {
  504. if (!IS_HANDLE_TYPE(handle, tcp) || !handle->sock.conn)
  505. return -PAL_ERROR_NOTCONNECTION;
  506. if (handle->sock.fd == PAL_IDX_POISON)
  507. return -PAL_ERROR_CONNFAILED;
  508. struct msghdr hdr;
  509. struct iovec iov;
  510. iov.iov_base = (void *) buf;
  511. iov.iov_len = len;
  512. hdr.msg_name = NULL;
  513. hdr.msg_namelen = 0;
  514. hdr.msg_iov = &iov;
  515. hdr.msg_iovlen = 1;
  516. hdr.msg_control = NULL;
  517. hdr.msg_controllen = 0;
  518. hdr.msg_flags = 0;
  519. int64_t bytes = INLINE_SYSCALL(sendmsg, 3, handle->sock.fd, &hdr, MSG_NOSIGNAL);
  520. if (IS_ERR(bytes))
  521. switch(ERRNO(bytes)) {
  522. case ECONNRESET:
  523. case EPIPE:
  524. return -PAL_ERROR_CONNFAILED;
  525. case EWOULDBLOCK:
  526. HANDLE_HDR(handle)->flags &= ~WRITEABLE(0);
  527. return -PAL_ERROR_TRYAGAIN;
  528. default:
  529. return unix_to_pal_error(ERRNO(bytes));
  530. }
  531. if (bytes == len)
  532. HANDLE_HDR(handle)->flags |= WRITEABLE(0);
  533. else
  534. HANDLE_HDR(handle)->flags &= ~WRITEABLE(0);
  535. return bytes;
  536. }
  537. /* used by 'open' operation of tcp stream for bound socket */
  538. static int udp_bind (PAL_HANDLE * handle, char * uri, int options)
  539. {
  540. struct sockaddr buffer, * bind_addr = &buffer;
  541. int bind_addrlen;
  542. int ret = 0, fd = -1;
  543. if ((ret = socket_parse_uri(uri, &bind_addr, &bind_addrlen,
  544. NULL, NULL)) < 0)
  545. return ret;
  546. assert(bind_addr);
  547. assert(bind_addrlen == addr_size(bind_addr));
  548. #if ALLOW_BIND_ANY == 0
  549. /* the socket need to have a binding address, a null address or an
  550. any address is not allowed */
  551. if (addr_check_any(bind_addr))
  552. return -PAL_ERROR_INVAL;
  553. #endif
  554. fd = INLINE_SYSCALL(socket, 3, bind_addr->sa_family,
  555. SOCK_DGRAM|SOCK_CLOEXEC|options, 0);
  556. if (IS_ERR(fd))
  557. return -PAL_ERROR_DENIED;
  558. if (bind_addr->sa_family == AF_INET6) {
  559. int ipv6only = 1;
  560. INLINE_SYSCALL(setsockopt, 5, fd, SOL_IPV6, IPV6_V6ONLY, &ipv6only,
  561. sizeof(int));
  562. }
  563. ret = INLINE_SYSCALL(bind, 3, fd, bind_addr, bind_addrlen);
  564. if (IS_ERR(ret)) {
  565. switch (ERRNO(ret)) {
  566. case EADDRINUSE:
  567. ret = -PAL_ERROR_STREAMEXIST;
  568. goto failed;
  569. case EADDRNOTAVAIL:
  570. ret = -PAL_ERROR_ADDRNOTEXIST;
  571. goto failed;
  572. default:
  573. ret = unix_to_pal_error(ERRNO(ret));
  574. goto failed;
  575. }
  576. }
  577. *handle = socket_create_handle(pal_type_udpsrv, fd, options,
  578. bind_addr, bind_addrlen, NULL, 0);
  579. if (!(*handle)) {
  580. ret = -ENOMEM;
  581. goto failed;
  582. }
  583. return 0;
  584. failed:
  585. INLINE_SYSCALL(close, 1, fd);
  586. return ret;
  587. }
  588. /* used by 'open' operation of tcp stream for connected socket */
  589. static int udp_connect (PAL_HANDLE * handle, char * uri, int options)
  590. {
  591. struct sockaddr buffer[2];
  592. struct sockaddr * bind_addr = buffer, * dest_addr = buffer + 1;
  593. int bind_addrlen, dest_addrlen;
  594. int ret, fd = -1;
  595. if ((ret = socket_parse_uri(uri, &bind_addr, &bind_addrlen,
  596. &dest_addr, &dest_addrlen)) < 0)
  597. return ret;
  598. #if ALLOW_BIND_ANY == 0
  599. /* the socket need to have a binding address, a null address or an
  600. any address is not allowed */
  601. if (bind_addr && addr_check_any(bind_addr))
  602. return -PAL_ERROR_INVAL;
  603. #endif
  604. fd = INLINE_SYSCALL(socket, 3, dest_addr ? dest_addr->sa_family : AF_INET,
  605. SOCK_DGRAM|SOCK_CLOEXEC|options, 0);
  606. if (IS_ERR(fd))
  607. return -PAL_ERROR_DENIED;
  608. if (dest_addr->sa_family == AF_INET6) {
  609. int ipv6only = 1;
  610. INLINE_SYSCALL(setsockopt, 5, fd, SOL_IPV6, IPV6_V6ONLY, &ipv6only,
  611. sizeof(int));
  612. }
  613. if (bind_addr) {
  614. ret = INLINE_SYSCALL(bind, 3, fd, bind_addr, bind_addrlen);
  615. if (IS_ERR(ret)) {
  616. switch (ERRNO(ret)) {
  617. case EADDRINUSE:
  618. ret = -PAL_ERROR_STREAMEXIST;
  619. goto failed;
  620. case EADDRNOTAVAIL:
  621. ret = -PAL_ERROR_ADDRNOTEXIST;
  622. goto failed;
  623. default:
  624. ret = unix_to_pal_error(ERRNO(ret));
  625. goto failed;
  626. }
  627. }
  628. }
  629. *handle = socket_create_handle(dest_addr ? pal_type_udp : pal_type_udpsrv,
  630. fd, options,
  631. bind_addr, bind_addrlen,
  632. dest_addr, dest_addrlen);
  633. if (!(*handle)) {
  634. ret = -ENOMEM;
  635. goto failed;
  636. }
  637. return 0;
  638. failed:
  639. INLINE_SYSCALL(close, 1, fd);
  640. return ret;
  641. }
  642. static int udp_open (PAL_HANDLE *hdl, const char * type, const char * uri,
  643. int access, int share, int create, int options)
  644. {
  645. char buf[PAL_SOCKADDR_SIZE];
  646. int len = strlen(uri);
  647. if (len >= PAL_SOCKADDR_SIZE)
  648. return -PAL_ERROR_TOOLONG;
  649. memcpy(buf, uri, len + 1);
  650. options &= PAL_OPTION_MASK;
  651. if (strpartcmp_static(type, "udp.srv:"))
  652. return udp_bind(hdl, buf, options);
  653. if (strpartcmp_static(type, "udp:"))
  654. return udp_connect(hdl, buf, options);
  655. return -PAL_ERROR_NOTSUPPORT;
  656. }
  657. static int64_t udp_receive (PAL_HANDLE handle, uint64_t offset, uint64_t len,
  658. void * buf)
  659. {
  660. if (!IS_HANDLE_TYPE(handle, udp))
  661. return -PAL_ERROR_NOTCONNECTION;
  662. if (handle->sock.fd == PAL_IDX_POISON)
  663. return -PAL_ERROR_BADHANDLE;
  664. struct msghdr hdr;
  665. struct iovec iov;
  666. iov.iov_base = buf;
  667. iov.iov_len = len;
  668. hdr.msg_name = NULL;
  669. hdr.msg_namelen = 0;
  670. hdr.msg_iov = &iov;
  671. hdr.msg_iovlen = 1;
  672. hdr.msg_control = NULL;
  673. hdr.msg_controllen = 0;
  674. hdr.msg_flags = 0;
  675. int64_t bytes = INLINE_SYSCALL(recvmsg, 3, handle->sock.fd, &hdr, 0);
  676. if (IS_ERR(bytes))
  677. switch(ERRNO(bytes)) {
  678. case EWOULDBLOCK:
  679. return -PAL_ERROR_TRYAGAIN;
  680. case EINTR:
  681. return -PAL_ERROR_INTERRUPTED;
  682. default:
  683. return unix_to_pal_error(ERRNO(bytes));
  684. }
  685. return bytes;
  686. }
  687. static int64_t udp_receivebyaddr (PAL_HANDLE handle, uint64_t offset, uint64_t len,
  688. void * buf, char * addr, int addrlen)
  689. {
  690. if (!IS_HANDLE_TYPE(handle, udpsrv))
  691. return -PAL_ERROR_NOTCONNECTION;
  692. if (handle->sock.fd == PAL_IDX_POISON)
  693. return -PAL_ERROR_BADHANDLE;
  694. struct sockaddr conn_addr;
  695. socklen_t conn_addrlen = sizeof(struct sockaddr);
  696. struct msghdr hdr;
  697. struct iovec iov;
  698. iov.iov_base = buf;
  699. iov.iov_len = len;
  700. hdr.msg_name = &conn_addr;
  701. hdr.msg_namelen = conn_addrlen;
  702. hdr.msg_iov = &iov;
  703. hdr.msg_iovlen = 1;
  704. hdr.msg_control = NULL;
  705. hdr.msg_controllen = 0;
  706. hdr.msg_flags = 0;
  707. int64_t bytes = INLINE_SYSCALL(recvmsg, 3, handle->sock.fd, &hdr, 0);
  708. if (IS_ERR(bytes))
  709. switch(ERRNO(bytes)) {
  710. case EWOULDBLOCK:
  711. return -PAL_ERROR_TRYAGAIN;
  712. case EINTR:
  713. return -PAL_ERROR_INTERRUPTED;
  714. case ECONNREFUSED:
  715. return -PAL_ERROR_STREAMNOTEXIST;
  716. default:
  717. return unix_to_pal_error(ERRNO(bytes));
  718. }
  719. char * addr_uri = strcpy_static(addr, "udp:", addrlen);
  720. if (!addr_uri)
  721. return -PAL_ERROR_OVERFLOW;
  722. int ret = inet_create_uri(addr_uri, addr + addrlen - addr_uri, &conn_addr,
  723. hdr.msg_namelen);
  724. if (ret < 0)
  725. return ret;
  726. return bytes;
  727. }
  728. static int64_t udp_send (PAL_HANDLE handle, uint64_t offset, uint64_t len,
  729. const void * buf)
  730. {
  731. if (!IS_HANDLE_TYPE(handle, udp))
  732. return -PAL_ERROR_NOTCONNECTION;
  733. if (handle->sock.fd == PAL_IDX_POISON)
  734. return -PAL_ERROR_BADHANDLE;
  735. struct msghdr hdr;
  736. struct iovec iov;
  737. iov.iov_base = (void *) buf;
  738. iov.iov_len = len;
  739. hdr.msg_name = (void *) handle->sock.conn;
  740. hdr.msg_namelen = addr_size((struct sockaddr *) handle->sock.conn);
  741. hdr.msg_iov = &iov;
  742. hdr.msg_iovlen = 1;
  743. hdr.msg_control = NULL;
  744. hdr.msg_controllen = 0;
  745. hdr.msg_flags = 0;
  746. int64_t bytes = INLINE_SYSCALL(sendmsg, 3, handle->sock.fd, &hdr, MSG_NOSIGNAL);
  747. if (IS_ERR(bytes))
  748. switch(ERRNO(bytes)) {
  749. case EAGAIN:
  750. HANDLE_HDR(handle)->flags &= ~WRITEABLE(0);
  751. return -PAL_ERROR_TRYAGAIN;
  752. case ECONNRESET:
  753. case EPIPE:
  754. return -PAL_ERROR_CONNFAILED;
  755. default:
  756. return unix_to_pal_error(ERRNO(bytes));
  757. }
  758. if (bytes == len)
  759. HANDLE_HDR(handle)->flags |= WRITEABLE(0);
  760. else
  761. HANDLE_HDR(handle)->flags &= ~WRITEABLE(0);
  762. return bytes;
  763. }
  764. static int64_t udp_sendbyaddr (PAL_HANDLE handle, uint64_t offset, uint64_t len,
  765. const void * buf, const char * addr, int addrlen)
  766. {
  767. if (!IS_HANDLE_TYPE(handle, udpsrv))
  768. return -PAL_ERROR_NOTCONNECTION;
  769. if (handle->sock.fd == PAL_IDX_POISON)
  770. return -PAL_ERROR_BADHANDLE;
  771. if (!strpartcmp_static(addr, "udp:"))
  772. return -PAL_ERROR_INVAL;
  773. addr += static_strlen("udp:");
  774. addrlen -= static_strlen("udp:");
  775. char * addrbuf = __alloca(addrlen);
  776. memcpy(addrbuf, addr, addrlen);
  777. struct sockaddr conn_addr;
  778. int conn_addrlen;
  779. int ret = inet_parse_uri(&addrbuf, &conn_addr, &conn_addrlen);
  780. if (ret < 0)
  781. return ret;
  782. struct msghdr hdr;
  783. struct iovec iov;
  784. iov.iov_base = (void *) buf;
  785. iov.iov_len = len;
  786. hdr.msg_name = &conn_addr;
  787. hdr.msg_namelen = conn_addrlen;
  788. hdr.msg_iov = &iov;
  789. hdr.msg_iovlen = 1;
  790. hdr.msg_control = NULL;
  791. hdr.msg_controllen = 0;
  792. hdr.msg_flags = 0;
  793. int64_t bytes = INLINE_SYSCALL(sendmsg, 3, handle->sock.fd, &hdr, MSG_NOSIGNAL);
  794. if (IS_ERR(bytes))
  795. switch(ERRNO(bytes)) {
  796. case ECONNRESET:
  797. case EPIPE:
  798. return -PAL_ERROR_CONNFAILED;
  799. case EAGAIN:
  800. HANDLE_HDR(handle)->flags &= ~WRITEABLE(0);
  801. default:
  802. return unix_to_pal_error(ERRNO(bytes));
  803. }
  804. if (bytes == len)
  805. HANDLE_HDR(handle)->flags |= WRITEABLE(0);
  806. else
  807. HANDLE_HDR(handle)->flags &= ~WRITEABLE(0);
  808. return bytes;
  809. }
  810. static int socket_delete (PAL_HANDLE handle, int access)
  811. {
  812. if (handle->sock.fd == PAL_IDX_POISON)
  813. return 0;
  814. if (!IS_HANDLE_TYPE(handle, tcp) && access)
  815. return -PAL_ERROR_INVAL;
  816. if (IS_HANDLE_TYPE(handle, tcp) || IS_HANDLE_TYPE(handle, tcpsrv)) {
  817. int shutdown;
  818. switch (access) {
  819. case 0:
  820. shutdown = SHUT_RDWR;
  821. break;
  822. case PAL_DELETE_RD:
  823. shutdown = SHUT_RD;
  824. break;
  825. case PAL_DELETE_WR:
  826. shutdown = SHUT_WR;
  827. break;
  828. default:
  829. return -PAL_ERROR_INVAL;
  830. }
  831. INLINE_SYSCALL(shutdown, 2, handle->sock.fd, shutdown);
  832. }
  833. return 0;
  834. }
  835. struct __kernel_linger {
  836. int l_onoff;
  837. int l_linger;
  838. };
  839. static int socket_close (PAL_HANDLE handle)
  840. {
  841. if (handle->sock.fd != PAL_IDX_POISON) {
  842. struct __kernel_linger l;
  843. l.l_onoff = 1;
  844. l.l_linger = 0;
  845. INLINE_SYSCALL(setsockopt, 5, handle->sock.fd, SOL_SOCKET, SO_LINGER,
  846. &l, sizeof(struct __kernel_linger));
  847. INLINE_SYSCALL(close, 1, handle->sock.fd);
  848. handle->sock.fd = PAL_IDX_POISON;
  849. }
  850. if (handle->sock.bind)
  851. handle->sock.bind = (PAL_PTR) NULL;
  852. if (handle->sock.conn)
  853. handle->sock.conn = (PAL_PTR) NULL;
  854. return 0;
  855. }
  856. #ifndef FIONREAD
  857. # define FIONREAD 0x541B
  858. #endif
  859. static int socket_attrquerybyhdl (PAL_HANDLE handle, PAL_STREAM_ATTR * attr)
  860. {
  861. if (handle->sock.fd == PAL_IDX_POISON)
  862. return -PAL_ERROR_BADHANDLE;
  863. attr->handle_type = HANDLE_HDR(handle)->type;
  864. attr->disconnected = HANDLE_HDR(handle)->flags & ERROR(0);
  865. attr->nonblocking = handle->sock.nonblocking;
  866. attr->writeable = HANDLE_HDR(handle)->flags & WRITEABLE(0);
  867. attr->pending_size = 0; /* fill in later */
  868. attr->socket.linger = handle->sock.linger;
  869. attr->socket.receivebuf = handle->sock.receivebuf;
  870. attr->socket.sendbuf = handle->sock.sendbuf;
  871. attr->socket.receivetimeout = handle->sock.receivetimeout;
  872. attr->socket.sendtimeout = handle->sock.sendtimeout;
  873. attr->socket.tcp_cork = handle->sock.tcp_cork;
  874. attr->socket.tcp_keepalive = handle->sock.tcp_keepalive;
  875. attr->socket.tcp_nodelay = handle->sock.tcp_nodelay;
  876. int fd = handle->sock.fd, ret, val;
  877. if (!IS_HANDLE_TYPE(handle, tcpsrv)) {
  878. /* try use ioctl FIONEAD to get the size of socket */
  879. ret = INLINE_SYSCALL(ioctl, 3, fd, FIONREAD, &val);
  880. if (IS_ERR(ret))
  881. return unix_to_pal_error(ERRNO(ret));
  882. attr->pending_size = val;
  883. }
  884. struct pollfd pfd = { .fd = fd, .events = POLLIN, .revents = 0 };
  885. struct timespec tp = { 0, 0 };
  886. ret = INLINE_SYSCALL(ppoll, 5, &pfd, 1, &tp, NULL, 0);
  887. attr->readable = (ret == 1 && pfd.revents == POLLIN);
  888. return 0;
  889. }
  890. static int socket_attrsetbyhdl (PAL_HANDLE handle, PAL_STREAM_ATTR * attr)
  891. {
  892. if (handle->sock.fd == PAL_IDX_POISON)
  893. return -PAL_ERROR_BADHANDLE;
  894. int fd = handle->sock.fd, ret, val;
  895. if (attr->nonblocking != handle->sock.nonblocking) {
  896. ret = INLINE_SYSCALL(fcntl, 3, fd, F_SETFL,
  897. attr->nonblocking ? O_NONBLOCK : 0);
  898. if (IS_ERR(ret))
  899. return unix_to_pal_error(ERRNO(ret));
  900. handle->sock.nonblocking = attr->nonblocking;
  901. }
  902. if (IS_HANDLE_TYPE(handle, tcpsrv)) {
  903. if (attr->socket.linger != handle->sock.linger) {
  904. struct __kernel_linger l;
  905. l.l_onoff = attr->socket.linger ? 1 : 0;
  906. l.l_linger = attr->socket.linger;
  907. ret = INLINE_SYSCALL(setsockopt, 5, fd, SOL_SOCKET, SO_LINGER,
  908. &l, sizeof(struct __kernel_linger));
  909. if (IS_ERR(ret))
  910. return unix_to_pal_error(ERRNO(ret));
  911. handle->sock.linger = attr->socket.linger;
  912. }
  913. if (attr->socket.receivebuf != handle->sock.receivebuf) {
  914. int val = attr->socket.receivebuf;
  915. ret = INLINE_SYSCALL(setsockopt, 5, fd, SOL_SOCKET, SO_RCVBUF,
  916. &val, sizeof(int));
  917. if (IS_ERR(ret))
  918. return unix_to_pal_error(ERRNO(ret));
  919. handle->sock.receivebuf = attr->socket.receivebuf;
  920. }
  921. if (attr->socket.sendbuf != handle->sock.sendbuf) {
  922. int val = attr->socket.sendbuf;
  923. ret = INLINE_SYSCALL(setsockopt, 5, fd, SOL_SOCKET, SO_SNDBUF,
  924. &val, sizeof(int));
  925. if (IS_ERR(ret))
  926. return unix_to_pal_error(ERRNO(ret));
  927. handle->sock.sendbuf = attr->socket.sendbuf;
  928. }
  929. if (attr->socket.receivetimeout != handle->sock.receivetimeout) {
  930. int val = attr->socket.receivetimeout;
  931. ret = INLINE_SYSCALL(setsockopt, 5, fd, SOL_SOCKET, SO_RCVTIMEO,
  932. &val, sizeof(int));
  933. if (IS_ERR(ret))
  934. return unix_to_pal_error(ERRNO(ret));
  935. handle->sock.receivetimeout = attr->socket.receivetimeout;
  936. }
  937. if (attr->socket.sendtimeout != handle->sock.sendtimeout) {
  938. int val = attr->socket.sendtimeout;
  939. ret = INLINE_SYSCALL(setsockopt, 5, fd, SOL_SOCKET, SO_SNDTIMEO,
  940. &val, sizeof(int));
  941. if (IS_ERR(ret))
  942. return unix_to_pal_error(ERRNO(ret));
  943. handle->sock.sendtimeout = attr->socket.sendtimeout;
  944. }
  945. }
  946. if (IS_HANDLE_TYPE(handle, tcp) || IS_HANDLE_TYPE(handle, tcpsrv)) {
  947. if (attr->socket.tcp_cork != handle->sock.tcp_cork) {
  948. val = attr->socket.tcp_cork ? 1 : 0;
  949. ret = INLINE_SYSCALL(setsockopt, 5, fd, SOL_TCP, TCP_CORK,
  950. &val, sizeof(int));
  951. if (IS_ERR(ret))
  952. return unix_to_pal_error(ERRNO(ret));
  953. handle->sock.tcp_cork = attr->socket.tcp_cork;
  954. }
  955. if (attr->socket.tcp_keepalive != handle->sock.tcp_keepalive) {
  956. val = attr->socket.tcp_keepalive ? 1 : 0;
  957. ret = INLINE_SYSCALL(setsockopt, 5, fd, SOL_SOCKET, SO_KEEPALIVE,
  958. &val, sizeof(int));
  959. if (IS_ERR(ret))
  960. return unix_to_pal_error(ERRNO(ret));
  961. handle->sock.tcp_keepalive = attr->socket.tcp_keepalive;
  962. }
  963. if (attr->socket.tcp_nodelay != handle->sock.tcp_nodelay) {
  964. val = attr->socket.tcp_nodelay ? 1 : 0;
  965. ret = INLINE_SYSCALL(setsockopt, 5, fd, SOL_TCP, TCP_NODELAY,
  966. &val, sizeof(int));
  967. if (IS_ERR(ret))
  968. return unix_to_pal_error(ERRNO(ret));
  969. handle->sock.tcp_nodelay = attr->socket.tcp_nodelay;
  970. }
  971. }
  972. return 0;
  973. }
  974. static int socket_getname (PAL_HANDLE handle, char * buffer, int count)
  975. {
  976. int old_count = count;
  977. int ret;
  978. const char * prefix = NULL;
  979. int prefix_len = 0;
  980. struct sockaddr * bind_addr = NULL, * dest_addr = NULL;
  981. switch (PAL_GET_TYPE(handle)) {
  982. case pal_type_tcpsrv:
  983. prefix_len = static_strlen("tcp.srv");
  984. prefix = "tcp.srv";
  985. bind_addr = (struct sockaddr *) handle->sock.bind;
  986. break;
  987. case pal_type_tcp:
  988. prefix_len = static_strlen("tcp");
  989. prefix = "tcp";
  990. bind_addr = (struct sockaddr *) handle->sock.bind;
  991. dest_addr = (struct sockaddr *) handle->sock.conn;
  992. break;
  993. case pal_type_udpsrv:
  994. prefix_len = static_strlen("udp.srv");
  995. prefix = "udp.srv";
  996. bind_addr = (struct sockaddr *) handle->sock.bind;
  997. break;
  998. case pal_type_udp:
  999. prefix_len = static_strlen("udp");
  1000. prefix = "udp";
  1001. bind_addr = (struct sockaddr *) handle->sock.bind;
  1002. dest_addr = (struct sockaddr *) handle->sock.conn;
  1003. break;
  1004. default:
  1005. return -PAL_ERROR_INVAL;
  1006. }
  1007. if (prefix_len >= count)
  1008. return -PAL_ERROR_OVERFLOW;
  1009. memcpy(buffer, prefix, prefix_len + 1);
  1010. buffer += prefix_len;
  1011. count -= prefix_len;
  1012. for (int i = 0 ; i < 2 ; i++) {
  1013. struct sockaddr * addr = i ? dest_addr : bind_addr;
  1014. if (addr) {
  1015. if (count <= 1)
  1016. return -PAL_ERROR_OVERFLOW;
  1017. buffer[0] = ':';
  1018. buffer[1] = 0;
  1019. buffer++;
  1020. count--;
  1021. if ((ret = inet_create_uri(buffer, count, addr,
  1022. addr_size(addr))) < 0)
  1023. return ret;
  1024. buffer += ret;
  1025. count -= ret;
  1026. }
  1027. }
  1028. return old_count - count;
  1029. }
  1030. struct handle_ops tcp_ops = {
  1031. .getname = &socket_getname,
  1032. .open = &tcp_open,
  1033. .waitforclient = &tcp_accept,
  1034. .read = &tcp_read,
  1035. .write = &tcp_write,
  1036. .delete = &socket_delete,
  1037. .close = &socket_close,
  1038. .attrquerybyhdl = &socket_attrquerybyhdl,
  1039. .attrsetbyhdl = &socket_attrsetbyhdl,
  1040. };
  1041. struct handle_ops udp_ops = {
  1042. .getname = &socket_getname,
  1043. .open = &udp_open,
  1044. .read = &udp_receive,
  1045. .write = &udp_send,
  1046. .delete = &socket_delete,
  1047. .close = &socket_close,
  1048. .attrquerybyhdl = &socket_attrquerybyhdl,
  1049. .attrsetbyhdl = &socket_attrsetbyhdl,
  1050. };
  1051. struct handle_ops udpsrv_ops = {
  1052. .getname = &socket_getname,
  1053. .open = &udp_open,
  1054. .readbyaddr = &udp_receivebyaddr,
  1055. .writebyaddr = &udp_sendbyaddr,
  1056. .delete = &socket_delete,
  1057. .close = &socket_close,
  1058. .attrquerybyhdl = &socket_attrquerybyhdl,
  1059. .attrsetbyhdl = &socket_attrsetbyhdl,
  1060. };
  1061. PAL_HANDLE _DkBroadcastStreamOpen (void)
  1062. {
  1063. if (!pal_sec.mcast_port) {
  1064. unsigned short mcast_port;
  1065. _DkFastRandomBitsRead(&mcast_port, sizeof(unsigned short));
  1066. pal_sec.mcast_port = mcast_port > 1024 ? mcast_port : mcast_port + 1024;
  1067. }
  1068. struct sockaddr_in addr;
  1069. int ret = 0;
  1070. addr.sin_family = AF_INET;
  1071. addr.sin_addr.s_addr = INADDR_ANY;
  1072. addr.sin_port = __htons(pal_sec.mcast_port);
  1073. /* set up server (sender) side */
  1074. int srv = INLINE_SYSCALL(socket, 3, AF_INET, SOCK_DGRAM|SOCK_CLOEXEC, 0);
  1075. if (IS_ERR(srv))
  1076. goto err;
  1077. ret = INLINE_SYSCALL(setsockopt, 5, srv, IPPROTO_IP, IP_MULTICAST_IF,
  1078. &addr.sin_addr.s_addr, sizeof(addr.sin_addr.s_addr));
  1079. if (IS_ERR(ret))
  1080. goto err_srv;
  1081. /* set up client (receiver) side */
  1082. int cli = INLINE_SYSCALL(socket, 3, AF_INET, SOCK_DGRAM|SOCK_CLOEXEC, 0);
  1083. if (IS_ERR(cli))
  1084. goto err_srv;
  1085. int reuse = 1;
  1086. INLINE_SYSCALL(setsockopt, 5, cli, SOL_SOCKET, SO_REUSEADDR,
  1087. &reuse, sizeof(reuse));
  1088. ret = INLINE_SYSCALL(bind, 3, cli, &addr, sizeof(addr));
  1089. if (IS_ERR(ret))
  1090. goto err_cli;
  1091. ret = INLINE_SYSCALL(setsockopt, 5, cli, IPPROTO_IP, IP_MULTICAST_IF,
  1092. &addr.sin_addr.s_addr, sizeof(addr.sin_addr.s_addr));
  1093. if (IS_ERR(ret))
  1094. goto err_cli;
  1095. inet_pton4(GRAPHENE_MCAST_GROUP, sizeof(GRAPHENE_MCAST_GROUP) - 1,
  1096. &addr.sin_addr.s_addr);
  1097. struct ip_mreq group;
  1098. group.imr_multiaddr.s_addr = addr.sin_addr.s_addr;
  1099. group.imr_interface.s_addr = INADDR_ANY;
  1100. ret = INLINE_SYSCALL(setsockopt, 5, cli, IPPROTO_IP, IP_ADD_MEMBERSHIP,
  1101. &group, sizeof(group));
  1102. if (IS_ERR(ret))
  1103. goto err_cli;
  1104. PAL_HANDLE hdl = malloc(HANDLE_SIZE(mcast));
  1105. SET_HANDLE_TYPE(hdl, mcast);
  1106. HANDLE_HDR(hdl)->flags |= WFD(1)|WRITEABLE(1);
  1107. hdl->mcast.srv = srv;
  1108. hdl->mcast.cli = cli;
  1109. hdl->mcast.port = (PAL_NUM) pal_sec.mcast_port;
  1110. hdl->mcast.nonblocking = PAL_FALSE;
  1111. hdl->mcast.addr = (PAL_PTR) remalloc(&addr, sizeof(addr));
  1112. return hdl;
  1113. err_cli:
  1114. INLINE_SYSCALL(close, 1, cli);
  1115. err_srv:
  1116. INLINE_SYSCALL(close, 1, srv);
  1117. err:
  1118. return NULL;
  1119. }
  1120. static int64_t mcast_send (PAL_HANDLE handle, uint64_t offset, uint64_t size,
  1121. const void * buf)
  1122. {
  1123. if (handle->mcast.srv == PAL_IDX_POISON)
  1124. return -PAL_ERROR_BADHANDLE;
  1125. struct msghdr hdr;
  1126. struct iovec iov;
  1127. iov.iov_base = (void *) buf;
  1128. iov.iov_len = size;
  1129. hdr.msg_name = (char *) handle->mcast.addr;
  1130. hdr.msg_namelen = sizeof(struct sockaddr_in);
  1131. hdr.msg_iov = &iov;
  1132. hdr.msg_iovlen = 1;
  1133. hdr.msg_control = NULL;
  1134. hdr.msg_controllen = 0;
  1135. hdr.msg_flags = 0;
  1136. int64_t bytes = INLINE_SYSCALL(sendmsg, 3, handle->mcast.srv, &hdr,
  1137. MSG_NOSIGNAL);
  1138. if (IS_ERR(bytes))
  1139. switch(ERRNO(bytes)) {
  1140. case ECONNRESET:
  1141. case EPIPE:
  1142. return -PAL_ERROR_CONNFAILED;
  1143. case EAGAIN:
  1144. HANDLE_HDR(handle)->flags &= ~WRITEABLE(1);
  1145. default:
  1146. return unix_to_pal_error(ERRNO(bytes));
  1147. }
  1148. if (bytes == size)
  1149. HANDLE_HDR(handle)->flags |= WRITEABLE(1);
  1150. else
  1151. HANDLE_HDR(handle)->flags &= ~WRITEABLE(1);
  1152. return bytes;
  1153. }
  1154. static int64_t mcast_receive (PAL_HANDLE handle, uint64_t offset, uint64_t size,
  1155. void * buf)
  1156. {
  1157. if (handle->mcast.cli == PAL_IDX_POISON)
  1158. return -PAL_ERROR_BADHANDLE;
  1159. struct msghdr hdr;
  1160. struct iovec iov;
  1161. iov.iov_base = buf;
  1162. iov.iov_len = size;
  1163. hdr.msg_name = NULL;
  1164. hdr.msg_namelen = 0;
  1165. hdr.msg_iov = &iov;
  1166. hdr.msg_iovlen = 1;
  1167. hdr.msg_control = NULL;
  1168. hdr.msg_controllen = 0;
  1169. hdr.msg_flags = 0;
  1170. int64_t bytes = INLINE_SYSCALL(recvmsg, 3, handle->mcast.cli, &hdr, 0);
  1171. if (IS_ERR(bytes))
  1172. return -PAL_ERROR_DENIED;
  1173. return bytes;
  1174. }
  1175. static int mcast_attrquerybyhdl (PAL_HANDLE handle, PAL_STREAM_ATTR * attr)
  1176. {
  1177. int ret, val;
  1178. if (handle->mcast.cli == PAL_IDX_POISON)
  1179. return -PAL_ERROR_BADHANDLE;
  1180. ret = INLINE_SYSCALL(ioctl, 3, handle->mcast.cli, FIONREAD, &val);
  1181. if (IS_ERR(ret))
  1182. return unix_to_pal_error(ERRNO(ret));
  1183. attr->handle_type = pal_type_mcast;
  1184. attr->disconnected = HANDLE_HDR(handle)->flags & (ERROR(0)|ERROR(1));
  1185. attr->nonblocking = handle->mcast.nonblocking;
  1186. attr->readable = !!val;
  1187. attr->writeable = HANDLE_HDR(handle)->flags & WRITEABLE(1);
  1188. attr->runnable = PAL_FALSE;
  1189. attr->pending_size = val;
  1190. return 0;
  1191. }
  1192. static int mcast_attrsetbyhdl (PAL_HANDLE handle, PAL_STREAM_ATTR * attr)
  1193. {
  1194. if (handle->mcast.cli == PAL_IDX_POISON)
  1195. return -PAL_ERROR_BADHANDLE;
  1196. int ret;
  1197. PAL_BOL * nonblocking = &handle->mcast.nonblocking;
  1198. if (attr->nonblocking != *nonblocking) {
  1199. ret = INLINE_SYSCALL(fcntl, 3, handle->mcast.cli, F_SETFL,
  1200. *nonblocking ? O_NONBLOCK : 0);
  1201. if (IS_ERR(ret))
  1202. return unix_to_pal_error(ERRNO(ret));
  1203. *nonblocking = attr->nonblocking;
  1204. }
  1205. return 0;
  1206. }
  1207. struct handle_ops mcast_ops = {
  1208. .write = &mcast_send,
  1209. .read = &mcast_receive,
  1210. .attrquerybyhdl = &mcast_attrquerybyhdl,
  1211. .attrsetbyhdl = &mcast_attrsetbyhdl,
  1212. };