Context Navigation

← Previous Revision
Next Revision →
Blame
Revision Log

socket.cpp

Last change on this file was 100171, checked in by vboxsync, 11 months ago
IPRT: tcp.h+tcp.cpp,socket.h+socket.cpp: Add RTTcpSetKeepAlive() which allows one to enable or disable sending periodic keep-alive messages on a socket (SO_KEEPALIVE) as described in RFC 1122. RTTcpSetKeepAlive() also allows one to adjust several keep-alive options on a per-socket basis: the idle time before keep-alive probes are sent (TCP_KEEPIDLE (TCP_KEEPALIVE on macOS)), the amount of time between keep-alive probes (TCP_KEEPINTVL), and the number of keep-alive probes to send before closing the connection (TCP_KEEPCNT).
Property svn:eol-style set to `native` Property svn:keywords set to `Id Revision`
File size: 97.3 KB

Line
1	/* $Id: socket.cpp 100171 2023-06-13 21:07:56Z vboxsync $ */
2	/** @file
3	* IPRT - Network Sockets.
4	*/
5
6	/*
7	* Copyright (C) 2006-2023 Oracle and/or its affiliates.
8	*
9	* This file is part of VirtualBox base platform packages, as
10	* available from https://www.virtualbox.org.
11	*
12	* This program is free software; you can redistribute it and/or
13	* modify it under the terms of the GNU General Public License
14	* as published by the Free Software Foundation, in version 3 of the
15	* License.
16	*
17	* This program is distributed in the hope that it will be useful, but
18	* WITHOUT ANY WARRANTY; without even the implied warranty of
19	* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
20	* General Public License for more details.
21	*
22	* You should have received a copy of the GNU General Public License
23	* along with this program; if not, see <https://www.gnu.org/licenses>.
24	*
25	* The contents of this file may alternatively be used under the terms
26	* of the Common Development and Distribution License Version 1.0
27	* (CDDL), a copy of it is provided in the "COPYING.CDDL" file included
28	* in the VirtualBox distribution, in which case the provisions of the
29	* CDDL are applicable instead of those of the GPL.
30	*
31	* You may elect to license modified versions of this file under the
32	* terms and conditions of either the GPL or the CDDL or both.
33	*
34	* SPDX-License-Identifier: GPL-3.0-only OR CDDL-1.0
35	*/
36
37
38	/*********************************************************************************************************************************
39	* Header Files *
40	*********************************************************************************************************************************/
41	#ifdef RT_OS_WINDOWS
42	# include <iprt/win/winsock2.h>
43	# include <iprt/win/ws2tcpip.h>
44	# include <mstcpip.h> /* for struct tcp_keepalive */
45	#else /* !RT_OS_WINDOWS */
46	# include <errno.h>
47	# include <sys/select.h>
48	# include <sys/stat.h>
49	# include <sys/socket.h>
50	# include <netinet/in.h>
51	# include <netinet/tcp.h>
52	# include <arpa/inet.h>
53	# ifdef IPRT_WITH_TCPIP_V6
54	# include <netinet6/in6.h>
55	# endif
56	# include <sys/un.h>
57	# include <netdb.h>
58	# include <unistd.h>
59	# include <fcntl.h>
60	# include <sys/uio.h>
61	# ifndef AF_LOCAL
62	# define AF_LOCAL AF_UNIX
63	# endif
64	#endif /* !RT_OS_WINDOWS */
65	#include <limits.h>
66
67	#include "internal/iprt.h"
68	#include <iprt/socket.h>
69
70	#include <iprt/alloca.h>
71	#include <iprt/asm.h>
72	#include <iprt/assert.h>
73	#include <iprt/ctype.h>
74	#include <iprt/err.h>
75	#include <iprt/mempool.h>
76	#include <iprt/poll.h>
77	#include <iprt/string.h>
78	#include <iprt/thread.h>
79	#include <iprt/time.h>
80	#include <iprt/mem.h>
81	#include <iprt/sg.h>
82	#include <iprt/log.h>
83
84	#include "internal/magics.h"
85	#include "internal/socket.h"
86	#include "internal/string.h"
87	#ifdef RT_OS_WINDOWS
88	# include "win/internal-r3-win.h"
89	#endif
90
91
92	/*********************************************************************************************************************************
93	* Defined Constants And Macros *
94	*********************************************************************************************************************************/
95	/* non-standard linux stuff (it seems). */
96	#ifndef MSG_NOSIGNAL
97	# define MSG_NOSIGNAL 0
98	#endif
99
100	/* Windows has different names for SHUT_XXX. */
101	#ifndef SHUT_RDWR
102	# ifdef SD_BOTH
103	# define SHUT_RDWR SD_BOTH
104	# else
105	# define SHUT_RDWR 2
106	# endif
107	#endif
108	#ifndef SHUT_WR
109	# ifdef SD_SEND
110	# define SHUT_WR SD_SEND
111	# else
112	# define SHUT_WR 1
113	# endif
114	#endif
115	#ifndef SHUT_RD
116	# ifdef SD_RECEIVE
117	# define SHUT_RD SD_RECEIVE
118	# else
119	# define SHUT_RD 0
120	# endif
121	#endif
122
123	/* fixup backlevel OSes. */
124	#if defined(RT_OS_OS2) \|\| defined(RT_OS_WINDOWS)
125	# define socklen_t int
126	#endif
127
128	/** How many pending connection. */
129	#define RTTCP_SERVER_BACKLOG 10
130
131	/* Limit read and write sizes on Windows and OS/2. */
132	#ifdef RT_OS_WINDOWS
133	# define RTSOCKET_MAX_WRITE (INT_MAX / 2)
134	# define RTSOCKET_MAX_READ (INT_MAX / 2)
135	#elif defined(RT_OS_OS2)
136	# define RTSOCKET_MAX_WRITE 0x10000
137	# define RTSOCKET_MAX_READ 0x10000
138	#endif
139
140
141	/*********************************************************************************************************************************
142	* Structures and Typedefs *
143	*********************************************************************************************************************************/
144	/**
145	* Socket handle data.
146	*
147	* This is mainly required for implementing RTPollSet on Windows.
148	*/
149	typedef struct RTSOCKETINT
150	{
151	/** Magic number (RTSOCKET_MAGIC). */
152	uint32_t u32Magic;
153	/** Exclusive user count.
154	* This is used to prevent two threads from accessing the handle concurrently.
155	* It can be higher than 1 if this handle is reference multiple times in a
156	* polling set (Windows). */
157	uint32_t volatile cUsers;
158	/** The native socket handle. */
159	RTSOCKETNATIVE hNative;
160	/** Indicates whether the handle has been closed or not. */
161	bool volatile fClosed;
162	/** Indicates whether the socket is operating in blocking or non-blocking mode
163	* currently. */
164	bool fBlocking;
165	/** Whether to leave the native socket open rather than closing it (for
166	* RTHandleGetStandard). */
167	bool fLeaveOpen;
168	#if defined(RT_OS_WINDOWS) \|\| defined(RT_OS_OS2)
169	/** The pollset currently polling this socket. This is NIL if no one is
170	* polling. */
171	RTPOLLSET hPollSet;
172	#endif
173	#ifdef RT_OS_WINDOWS
174	/** The event semaphore we've associated with the socket handle.
175	* This is WSA_INVALID_EVENT if not done. */
176	WSAEVENT hEvent;
177	/** The events we're polling for. */
178	uint32_t fPollEvts;
179	/** The events we're currently subscribing to with WSAEventSelect.
180	* This is ZERO if we're currently not subscribing to anything. */
181	uint32_t fSubscribedEvts;
182	/** Saved events which are only posted once and events harvested for
183	* sockets entered multiple times into to a poll set. Imagine a scenario where
184	* you have a RTPOLL_EVT_READ entry and RTPOLL_EVT_ERROR entry. The READ
185	* condition can be triggered between checking the READ entry and the ERROR
186	* entry, and we don't want to drop the READ, so we store it here and make sure
187	* the event is signalled.
188	*
189	* The RTPOLL_EVT_ERROR is inconsistenly sticky at the momemnt... */
190	uint32_t fEventsSaved;
191	/** Set if fEventsSaved contains harvested events (used to avoid multiple
192	* calls to rtSocketPollCheck on the same socket during rtSocketPollDone). */
193	bool fHarvestedEvents;
194	/** Set if we're using the polling fallback. */
195	bool fPollFallback;
196	/** Set if the fallback polling is active (event not set). */
197	bool volatile fPollFallbackActive;
198	/** Set to shut down the fallback polling thread. */
199	bool volatile fPollFallbackShutdown;
200	/** Socket use to wake up the select thread. */
201	RTSOCKETNATIVE hPollFallbackNotifyW;
202	/** Socket the select thread always waits on. */
203	RTSOCKETNATIVE hPollFallbackNotifyR;
204	/** The fallback polling thread. */
205	RTTHREAD hPollFallbackThread;
206	#endif /* RT_OS_WINDOWS */
207	} RTSOCKETINT;
208
209
210	/**
211	* Address union used internally for things like getpeername and getsockname.
212	*/
213	typedef union RTSOCKADDRUNION
214	{
215	struct sockaddr Addr;
216	struct sockaddr_in IPv4;
217	#ifdef IPRT_WITH_TCPIP_V6
218	struct sockaddr_in6 IPv6;
219	#endif
220	} RTSOCKADDRUNION;
221
222
223	/*********************************************************************************************************************************
224	* Global Variables *
225	*********************************************************************************************************************************/
226	#ifdef RT_OS_WINDOWS
227	/** Indicates that we've successfully initialized winsock. */
228	static uint32_t volatile g_uWinSockInitedVersion = 0;
229	#endif
230
231
232	/*********************************************************************************************************************************
233	* Internal Functions *
234	*********************************************************************************************************************************/
235	#ifdef RT_OS_WINDOWS
236	static void rtSocketPokePollFallbackThread(RTSOCKETINT *pThis);
237	#endif
238
239
240
241	#ifdef RT_OS_WINDOWS
242	/**
243	* Initializes winsock for the process.
244	*
245	* @returns IPRT status code.
246	*/
247	static int rtSocketInitWinsock(void)
248	{
249	if (g_uWinSockInitedVersion != 0)
250	return VINF_SUCCESS;
251
252	if ( !g_pfnWSAGetLastError
253	\|\| !g_pfnWSAStartup
254	\|\| !g_pfnsocket
255	\|\| !g_pfnclosesocket)
256	return VERR_NET_INIT_FAILED;
257
258	/*
259	* Initialize winsock. Try with 2.2 and back down till we get something that works.
260	*/
261	static const WORD s_awVersions[] =
262	{
263	MAKEWORD(2, 2),
264	MAKEWORD(2, 1),
265	MAKEWORD(2, 0),
266	MAKEWORD(1, 1),
267	MAKEWORD(1, 0),
268	};
269	for (uint32_t i = 0; i < RT_ELEMENTS(s_awVersions); i++)
270	{
271	WSADATA wsaData;
272	RT_ZERO(wsaData);
273	int rcWsa = g_pfnWSAStartup(s_awVersions[i], &wsaData);
274	if (rcWsa == 0)
275	{
276	/* AssertMsg(wsaData.wVersion >= s_awVersions[i]); - triggers with winsock 1.1 */
277	ASMAtomicWriteU32(&g_uWinSockInitedVersion, wsaData.wVersion);
278	return VINF_SUCCESS;
279	}
280	AssertLogRelMsg(rcWsa == WSAVERNOTSUPPORTED, ("rcWsa=%d (winsock version %#x)\n", rcWsa, s_awVersions[i]));
281	}
282	LogRel(("Failed to init winsock!\n"));
283	return VERR_NET_INIT_FAILED;
284	}
285	#endif
286
287
288	/**
289	* Get the last error as an iprt status code.
290	*
291	* @returns IPRT status code.
292	*/
293	DECLINLINE(int) rtSocketError(void)
294	{
295	#ifdef RT_OS_WINDOWS
296	if (g_pfnWSAGetLastError)
297	return RTErrConvertFromWin32(g_pfnWSAGetLastError());
298	return VERR_NET_IO_ERROR;
299	#else
300	return RTErrConvertFromErrno(errno);
301	#endif
302	}
303
304
305	/**
306	* Resets the last error.
307	*/
308	DECLINLINE(void) rtSocketErrorReset(void)
309	{
310	#ifdef RT_OS_WINDOWS
311	if (g_pfnWSASetLastError)
312	g_pfnWSASetLastError(0);
313	#else
314	errno = 0;
315	#endif
316	}
317
318
319	/**
320	* Get the last resolver error as an iprt status code.
321	*
322	* @returns iprt status code.
323	*/
324	DECLHIDDEN(int) rtSocketResolverError(void)
325	{
326	#ifdef RT_OS_WINDOWS
327	if (g_pfnWSAGetLastError)
328	return RTErrConvertFromWin32(g_pfnWSAGetLastError());
329	return VERR_UNRESOLVED_ERROR;
330	#else
331	switch (h_errno)
332	{
333	case HOST_NOT_FOUND:
334	return VERR_NET_HOST_NOT_FOUND;
335	case NO_DATA:
336	return VERR_NET_ADDRESS_NOT_AVAILABLE;
337	case NO_RECOVERY:
338	return VERR_IO_GEN_FAILURE;
339	case TRY_AGAIN:
340	return VERR_TRY_AGAIN;
341
342	default:
343	AssertLogRelMsgFailed(("Unhandled error %u\n", h_errno));
344	return VERR_UNRESOLVED_ERROR;
345	}
346	#endif
347	}
348
349
350	/**
351	* Converts from a native socket address to a generic IPRT network address.
352	*
353	* @returns IPRT status code.
354	* @param pSrc The source address.
355	* @param cbSrc The size of the source address.
356	* @param pAddr Where to return the generic IPRT network
357	* address.
358	*/
359	static int rtSocketNetAddrFromAddr(RTSOCKADDRUNION const *pSrc, size_t cbSrc, PRTNETADDR pAddr)
360	{
361	/*
362	* Convert the address.
363	*/
364	if ( cbSrc == sizeof(struct sockaddr_in)
365	&& pSrc->Addr.sa_family == AF_INET)
366	{
367	RT_ZERO(*pAddr);
368	pAddr->enmType = RTNETADDRTYPE_IPV4;
369	pAddr->uPort = RT_N2H_U16(pSrc->IPv4.sin_port);
370	pAddr->uAddr.IPv4.u = pSrc->IPv4.sin_addr.s_addr;
371	}
372	#ifdef IPRT_WITH_TCPIP_V6
373	else if ( cbSrc == sizeof(struct sockaddr_in6)
374	&& pSrc->Addr.sa_family == AF_INET6)
375	{
376	RT_ZERO(*pAddr);
377	pAddr->enmType = RTNETADDRTYPE_IPV6;
378	pAddr->uPort = RT_N2H_U16(pSrc->IPv6.sin6_port);
379	pAddr->uAddr.IPv6.au32[0] = pSrc->IPv6.sin6_addr.s6_addr32[0];
380	pAddr->uAddr.IPv6.au32[1] = pSrc->IPv6.sin6_addr.s6_addr32[1];
381	pAddr->uAddr.IPv6.au32[2] = pSrc->IPv6.sin6_addr.s6_addr32[2];
382	pAddr->uAddr.IPv6.au32[3] = pSrc->IPv6.sin6_addr.s6_addr32[3];
383	}
384	#endif
385	else
386	return VERR_NET_ADDRESS_FAMILY_NOT_SUPPORTED;
387	return VINF_SUCCESS;
388	}
389
390
391	/**
392	* Converts from a generic IPRT network address to a native socket address.
393	*
394	* @returns IPRT status code.
395	* @param pAddr Pointer to the generic IPRT network address.
396	* @param pDst The source address.
397	* @param cbDst The size of the source address.
398	* @param pcbAddr Where to store the size of the returned address.
399	* Optional
400	*/
401	static int rtSocketAddrFromNetAddr(PCRTNETADDR pAddr, RTSOCKADDRUNION pDst, size_t cbDst, int pcbAddr)
402	{
403	RT_BZERO(pDst, cbDst);
404	if (pAddr->enmType == RTNETADDRTYPE_IPV4)
405	{
406	if (cbDst < sizeof(struct sockaddr_in))
407	return VERR_BUFFER_OVERFLOW;
408
409	pDst->Addr.sa_family = AF_INET;
410	pDst->IPv4.sin_port = RT_H2N_U16(pAddr->uPort);
411	pDst->IPv4.sin_addr.s_addr = pAddr->uAddr.IPv4.u;
412	if (pcbAddr)
413	*pcbAddr = sizeof(pDst->IPv4);
414	}
415	#ifdef IPRT_WITH_TCPIP_V6
416	else if (pAddr->enmType == RTNETADDRTYPE_IPV6)
417	{
418	if (cbDst < sizeof(struct sockaddr_in6))
419	return VERR_BUFFER_OVERFLOW;
420
421	pDst->Addr.sa_family = AF_INET6;
422	pDst->IPv6.sin6_port = RT_H2N_U16(pAddr->uPort);
423	pSrc->IPv6.sin6_addr.s6_addr32[0] = pAddr->uAddr.IPv6.au32[0];
424	pSrc->IPv6.sin6_addr.s6_addr32[1] = pAddr->uAddr.IPv6.au32[1];
425	pSrc->IPv6.sin6_addr.s6_addr32[2] = pAddr->uAddr.IPv6.au32[2];
426	pSrc->IPv6.sin6_addr.s6_addr32[3] = pAddr->uAddr.IPv6.au32[3];
427	if (pcbAddr)
428	*pcbAddr = sizeof(pDst->IPv6);
429	}
430	#endif
431	else
432	return VERR_NET_ADDRESS_FAMILY_NOT_SUPPORTED;
433	return VINF_SUCCESS;
434	}
435
436
437	/**
438	* Tries to lock the socket for exclusive usage by the calling thread.
439	*
440	* Call rtSocketUnlock() to unlock.
441	*
442	* @returns @c true if locked, @c false if not.
443	* @param pThis The socket structure.
444	*/
445	DECLINLINE(bool) rtSocketTryLock(RTSOCKETINT *pThis)
446	{
447	return ASMAtomicCmpXchgU32(&pThis->cUsers, 1, 0);
448	}
449
450
451	/**
452	* Unlocks the socket.
453	*
454	* @param pThis The socket structure.
455	*/
456	DECLINLINE(void) rtSocketUnlock(RTSOCKETINT *pThis)
457	{
458	ASMAtomicCmpXchgU32(&pThis->cUsers, 0, 1);
459	}
460
461
462	/**
463	* The slow path of rtSocketSwitchBlockingMode that does the actual switching.
464	*
465	* @returns IPRT status code.
466	* @param pThis The socket structure.
467	* @param fBlocking The desired mode of operation.
468	* @remarks Do not call directly.
469	*/
470	static int rtSocketSwitchBlockingModeSlow(RTSOCKETINT *pThis, bool fBlocking)
471	{
472	#ifdef RT_OS_WINDOWS
473	AssertReturn(g_pfnioctlsocket, VERR_NET_NOT_UNSUPPORTED);
474	u_long uBlocking = fBlocking ? 0 : 1;
475	if (g_pfnioctlsocket(pThis->hNative, FIONBIO, &uBlocking))
476	return rtSocketError();
477
478	#else
479	int fFlags = fcntl(pThis->hNative, F_GETFL, 0);
480	if (fFlags == -1)
481	return rtSocketError();
482
483	if (fBlocking)
484	fFlags &= ~O_NONBLOCK;
485	else
486	fFlags \|= O_NONBLOCK;
487	if (fcntl(pThis->hNative, F_SETFL, fFlags) == -1)
488	return rtSocketError();
489	#endif
490
491	pThis->fBlocking = fBlocking;
492	return VINF_SUCCESS;
493	}
494
495
496	/**
497	* Switches the socket to the desired blocking mode if necessary.
498	*
499	* The socket must be locked.
500	*
501	* @returns IPRT status code.
502	* @param pThis The socket structure.
503	* @param fBlocking The desired mode of operation.
504	*/
505	DECLINLINE(int) rtSocketSwitchBlockingMode(RTSOCKETINT *pThis, bool fBlocking)
506	{
507	if (pThis->fBlocking != fBlocking)
508	return rtSocketSwitchBlockingModeSlow(pThis, fBlocking);
509	return VINF_SUCCESS;
510	}
511
512
513	/**
514	* Creates an IPRT socket handle for a native one.
515	*
516	* @returns IPRT status code.
517	* @param ppSocket Where to return the IPRT socket handle.
518	* @param hNative The native handle.
519	* @param fLeaveOpen Whether to leave the native socket handle open when
520	* closed.
521	*/
522	DECLHIDDEN(int) rtSocketCreateForNative(RTSOCKETINT **ppSocket, RTSOCKETNATIVE hNative, bool fLeaveOpen)
523	{
524	RTSOCKETINT pThis = (RTSOCKETINT )RTMemPoolAlloc(RTMEMPOOL_DEFAULT, sizeof(*pThis));
525	if (!pThis)
526	return VERR_NO_MEMORY;
527	pThis->u32Magic = RTSOCKET_MAGIC;
528	pThis->cUsers = 0;
529	pThis->hNative = hNative;
530	pThis->fClosed = false;
531	pThis->fLeaveOpen = fLeaveOpen;
532	pThis->fBlocking = true;
533	#if defined(RT_OS_WINDOWS) \|\| defined(RT_OS_OS2)
534	pThis->hPollSet = NIL_RTPOLLSET;
535	#endif
536	#ifdef RT_OS_WINDOWS
537	pThis->hEvent = WSA_INVALID_EVENT;
538	pThis->fPollEvts = 0;
539	pThis->fSubscribedEvts = 0;
540	pThis->fEventsSaved = 0;
541	pThis->fHarvestedEvents = false;
542	pThis->fPollFallback = g_uWinSockInitedVersion < MAKEWORD(2, 0)
543	\|\| g_pfnWSACreateEvent == NULL
544	\|\| g_pfnWSACloseEvent == NULL
545	\|\| g_pfnWSAEventSelect == NULL
546	\|\| g_pfnWSAEnumNetworkEvents == NULL;
547	pThis->fPollFallbackActive = false;
548	pThis->fPollFallbackShutdown = false;
549	pThis->hPollFallbackNotifyR = NIL_RTSOCKETNATIVE;
550	pThis->hPollFallbackNotifyW = NIL_RTSOCKETNATIVE;
551	pThis->hPollFallbackThread = NIL_RTTHREAD;
552	#endif
553	*ppSocket = pThis;
554	return VINF_SUCCESS;
555	}
556
557
558	RTDECL(int) RTSocketFromNative(PRTSOCKET phSocket, RTHCINTPTR uNative)
559	{
560	AssertReturn(uNative != NIL_RTSOCKETNATIVE, VERR_INVALID_PARAMETER);
561	#ifndef RT_OS_WINDOWS
562	AssertReturn(uNative >= 0, VERR_INVALID_PARAMETER);
563	#endif
564	AssertPtrReturn(phSocket, VERR_INVALID_POINTER);
565	return rtSocketCreateForNative(phSocket, uNative, false /fLeaveOpen/);
566	}
567
568
569	/**
570	* Wrapper around socket().
571	*
572	* @returns IPRT status code.
573	* @param phSocket Where to store the handle to the socket on
574	* success.
575	* @param iDomain The protocol family (PF_XXX).
576	* @param iType The socket type (SOCK_XXX).
577	* @param iProtocol Socket parameter, usually 0.
578	* @param fInheritable Set to true if the socket should be inherted by
579	* child processes, false if not inheritable.
580	*/
581	DECLHIDDEN(int) rtSocketCreate(PRTSOCKET phSocket, int iDomain, int iType, int iProtocol, bool fInheritable)
582	{
583	#ifdef RT_OS_WINDOWS
584	AssertReturn(g_pfnsocket, VERR_NET_NOT_UNSUPPORTED);
585	AssertReturn(g_pfnclosesocket, VERR_NET_NOT_UNSUPPORTED);
586
587	/* Initialize WinSock. */
588	int rc2 = rtSocketInitWinsock();
589	if (RT_FAILURE(rc2))
590	return rc2;
591	#endif
592
593	/*
594	* Create the socket.
595	*
596	* The RTSocketSetInheritance operation isn't necessarily reliable on windows,
597	* so try use WSA_FLAG_NO_HANDLE_INHERIT with WSASocketW when possible.
598	*/
599	#ifdef RT_OS_WINDOWS
600	bool fCallSetInheritance = true;
601	RTSOCKETNATIVE hNative;
602	if (g_pfnWSASocketW)
603	{
604	DWORD fWsaFlags = WSA_FLAG_OVERLAPPED \| (!fInheritable ? WSA_FLAG_NO_HANDLE_INHERIT : 0);
605	hNative = g_pfnWSASocketW(iDomain, iType, iProtocol, NULL, 0 /Group/, fWsaFlags);
606	if (hNative != NIL_RTSOCKETNATIVE)
607	fCallSetInheritance = false;
608	else
609	{
610	if (!fInheritable)
611	hNative = g_pfnsocket(iDomain, iType, iProtocol);
612	if (hNative == NIL_RTSOCKETNATIVE)
613	return rtSocketError();
614	}
615	}
616	else
617	{
618	hNative = g_pfnsocket(iDomain, iType, iProtocol);
619	if (hNative == NIL_RTSOCKETNATIVE)
620	return rtSocketError();
621	}
622	#else
623	RTSOCKETNATIVE hNative = socket(iDomain, iType, iProtocol);
624	if (hNative == NIL_RTSOCKETNATIVE)
625	return rtSocketError();
626	#endif
627
628	/*
629	* Wrap it.
630	*/
631	int rc = rtSocketCreateForNative(phSocket, hNative, false /fLeaveOpen/);
632	if (RT_SUCCESS(rc))
633	{
634	#ifdef RT_OS_WINDOWS
635	if (fCallSetInheritance)
636	#endif
637	RTSocketSetInheritance(*phSocket, fInheritable);
638	}
639	else
640	{
641	#ifdef RT_OS_WINDOWS
642	g_pfnclosesocket(hNative);
643	#else
644	close(hNative);
645	#endif
646	}
647	return rc;
648	}
649
650
651	/**
652	* Wrapper around socketpair() for creating a local TCP connection.
653	*
654	* @returns IPRT status code.
655	* @param phServer Where to return the first native socket.
656	* @param phClient Where to return the second native socket.
657	*/
658	static int rtSocketCreateNativeTcpPair(RTSOCKETNATIVE phServer, RTSOCKETNATIVE phClient)
659	{
660	#ifdef RT_OS_WINDOWS
661	/*
662	* Initialize WinSock and make sure we got the necessary APIs.
663	*/
664	int rc = rtSocketInitWinsock();
665	if (RT_FAILURE(rc))
666	return rc;
667	AssertReturn(g_pfnsocket, VERR_NET_NOT_UNSUPPORTED);
668	AssertReturn(g_pfnclosesocket, VERR_NET_NOT_UNSUPPORTED);
669	AssertReturn(g_pfnsetsockopt, VERR_NET_NOT_UNSUPPORTED);
670	AssertReturn(g_pfnbind, VERR_NET_NOT_UNSUPPORTED);
671	AssertReturn(g_pfngetsockname, VERR_NET_NOT_UNSUPPORTED);
672	AssertReturn(g_pfnlisten, VERR_NET_NOT_UNSUPPORTED);
673	AssertReturn(g_pfnaccept, VERR_NET_NOT_UNSUPPORTED);
674	AssertReturn(g_pfnconnect, VERR_NET_NOT_UNSUPPORTED);
675
676	/*
677	* Create the "server" listen socket and the "client" socket.
678	*/
679	RTSOCKETNATIVE hListener = g_pfnsocket(AF_INET, SOCK_STREAM, IPPROTO_TCP);
680	if (hListener == NIL_RTSOCKETNATIVE)
681	return rtSocketError();
682	RTSOCKETNATIVE hClient = g_pfnsocket(AF_INET, SOCK_STREAM, IPPROTO_TCP);
683	if (hClient != NIL_RTSOCKETNATIVE)
684	{
685
686	/*
687	* We let WinSock choose a port number when we bind.
688	*/
689	union
690	{
691	struct sockaddr_in Ip;
692	struct sockaddr Generic;
693	} uAddr;
694	RT_ZERO(uAddr);
695	uAddr.Ip.sin_family = AF_INET;
696	uAddr.Ip.sin_addr.s_addr = RT_H2N_U32_C(INADDR_LOOPBACK);
697	//uAddr.Ip.sin_port = 0;
698	int fReuse = 1;
699	rc = g_pfnsetsockopt(hListener, SOL_SOCKET, SO_REUSEADDR, (const char *)&fReuse, sizeof(fReuse));
700	if (rc == 0)
701	{
702	rc = g_pfnbind(hListener, &uAddr.Generic, sizeof(uAddr.Ip));
703	if (rc == 0)
704	{
705	/*
706	* Get the address the client should connect to. According to the docs,
707	* we cannot assume that getsockname sets the IP and family.
708	*/
709	RT_ZERO(uAddr);
710	int cbAddr = sizeof(uAddr.Ip);
711	rc = g_pfngetsockname(hListener, &uAddr.Generic, &cbAddr);
712	if (rc == 0)
713	{
714	uAddr.Ip.sin_family = AF_INET;
715	uAddr.Ip.sin_addr.s_addr = RT_H2N_U32_C(INADDR_LOOPBACK);
716
717	/*
718	* Listen, connect and accept.
719	*/
720	rc = g_pfnlisten(hListener, 1 /cBacklog/);
721	if (rc == 0)
722	{
723	rc = g_pfnconnect(hClient, &uAddr.Generic, sizeof(uAddr.Ip));
724	if (rc == 0)
725	{
726	RTSOCKETNATIVE hServer = g_pfnaccept(hListener, NULL, NULL);
727	if (hServer != NIL_RTSOCKETNATIVE)
728	{
729	g_pfnclosesocket(hListener);
730
731	/*
732	* Done!
733	*/
734	*phServer = hServer;
735	*phClient = hClient;
736	return VINF_SUCCESS;
737	}
738	}
739	}
740	}
741	}
742	}
743	rc = rtSocketError();
744	g_pfnclosesocket(hClient);
745	}
746	else
747	rc = rtSocketError();
748	g_pfnclosesocket(hListener);
749	return rc;
750
751	#else
752	/*
753	* Got socket pair, so use it.
754	* Note! This isn't TCP per se, but it should fool the users.
755	*/
756	int aSockets[2] = { -1, -1 };
757	if (socketpair(AF_LOCAL, SOCK_STREAM, 0, aSockets) == 0)
758	{
759	*phServer = aSockets[0];
760	*phClient = aSockets[1];
761	return VINF_SUCCESS;
762	}
763	return rtSocketError();
764	#endif
765	}
766
767
768	/**
769	* Worker for RTTcpCreatePair.
770	*
771	* @returns IPRT status code.
772	* @param phServer Where to return the "server" side of the pair.
773	* @param phClient Where to return the "client" side of the pair.
774	* @note There is no server or client side, but we gotta call it something.
775	*/
776	DECLHIDDEN(int) rtSocketCreateTcpPair(RTSOCKET phServer, RTSOCKET phClient)
777	{
778	RTSOCKETNATIVE hServer = NIL_RTSOCKETNATIVE;
779	RTSOCKETNATIVE hClient = NIL_RTSOCKETNATIVE;
780	int rc = rtSocketCreateNativeTcpPair(&hServer, &hClient);
781	if (RT_SUCCESS(rc))
782	{
783	rc = rtSocketCreateForNative(phServer, hServer, false /fLeaveOpen/);
784	if (RT_SUCCESS(rc))
785	{
786	rc = rtSocketCreateForNative(phClient, hClient, false /fLeaveOpen/);
787	if (RT_SUCCESS(rc))
788	return VINF_SUCCESS;
789	RTSocketRelease(*phServer);
790	}
791	else
792	{
793	#ifdef RT_OS_WINDOWS
794	g_pfnclosesocket(hServer);
795	#else
796	close(hServer);
797	#endif
798	}
799	#ifdef RT_OS_WINDOWS
800	g_pfnclosesocket(hClient);
801	#else
802	close(hClient);
803	#endif
804	}
805
806	*phServer = NIL_RTSOCKET;
807	*phClient = NIL_RTSOCKET;
808	return rc;
809	}
810
811
812	RTDECL(uint32_t) RTSocketRetain(RTSOCKET hSocket)
813	{
814	RTSOCKETINT *pThis = hSocket;
815	AssertPtrReturn(pThis, UINT32_MAX);
816	AssertReturn(pThis->u32Magic == RTSOCKET_MAGIC, UINT32_MAX);
817	return RTMemPoolRetain(pThis);
818	}
819
820
821	/**
822	* Worker for RTSocketRelease and RTSocketClose.
823	*
824	* @returns IPRT status code.
825	* @param pThis The socket handle instance data.
826	* @param fDestroy Whether we're reaching ref count zero.
827	*/
828	static int rtSocketCloseIt(RTSOCKETINT *pThis, bool fDestroy)
829	{
830	/*
831	* Invalidate the handle structure on destroy.
832	*/
833	if (fDestroy)
834	{
835	Assert(ASMAtomicReadU32(&pThis->u32Magic) == RTSOCKET_MAGIC);
836	ASMAtomicWriteU32(&pThis->u32Magic, RTSOCKET_MAGIC_DEAD);
837	}
838
839	int rc = VINF_SUCCESS;
840	if (ASMAtomicCmpXchgBool(&pThis->fClosed, true, false))
841	{
842	#ifdef RT_OS_WINDOWS
843	/*
844	* Poke the polling thread if active and give it a small chance to stop.
845	*/
846	if ( pThis->fPollFallback
847	&& pThis->hPollFallbackThread != NIL_RTTHREAD)
848	{
849	ASMAtomicWriteBool(&pThis->fPollFallbackShutdown, true);
850	rtSocketPokePollFallbackThread(pThis);
851	int rc2 = RTThreadWait(pThis->hPollFallbackThread, RT_MS_1SEC, NULL);
852	if (RT_SUCCESS(rc2))
853	pThis->hPollFallbackThread = NIL_RTTHREAD;
854	}
855	#endif
856
857	/*
858	* Close the native handle.
859	*/
860	RTSOCKETNATIVE hNative = pThis->hNative;
861	if (hNative != NIL_RTSOCKETNATIVE)
862	{
863	pThis->hNative = NIL_RTSOCKETNATIVE;
864
865	if (!pThis->fLeaveOpen)
866	{
867	#ifdef RT_OS_WINDOWS
868	AssertReturn(g_pfnclosesocket, VERR_NET_NOT_UNSUPPORTED);
869	if (g_pfnclosesocket(hNative))
870	#else
871	if (close(hNative))
872	#endif
873	{
874	rc = rtSocketError();
875	#ifdef RT_OS_WINDOWS
876	AssertMsgFailed(("closesocket(%p) -> %Rrc\n", (uintptr_t)hNative, rc));
877	#else
878	AssertMsgFailed(("close(%d) -> %Rrc\n", hNative, rc));
879	#endif
880	}
881	}
882	}
883
884	#ifdef RT_OS_WINDOWS
885	/*
886	* Windows specific polling cleanup.
887	*/
888	WSAEVENT hEvent = pThis->hEvent;
889	if (hEvent != WSA_INVALID_EVENT)
890	{
891	pThis->hEvent = WSA_INVALID_EVENT;
892	if (!pThis->fPollFallback)
893	{
894	Assert(g_pfnWSACloseEvent);
895	if (g_pfnWSACloseEvent)
896	g_pfnWSACloseEvent(hEvent);
897	}
898	else
899	CloseHandle(hEvent);
900	}
901
902	if (pThis->fPollFallback)
903	{
904	if (pThis->hPollFallbackNotifyW != NIL_RTSOCKETNATIVE)
905	{
906	g_pfnclosesocket(pThis->hPollFallbackNotifyW);
907	pThis->hPollFallbackNotifyW = NIL_RTSOCKETNATIVE;
908	}
909
910	if (pThis->hPollFallbackThread != NIL_RTTHREAD)
911	{
912	int rc2 = RTThreadWait(pThis->hPollFallbackThread, RT_MS_1MIN / 2, NULL);
913	AssertRC(rc2);
914	pThis->hPollFallbackThread = NIL_RTTHREAD;
915	}
916
917	if (pThis->hPollFallbackNotifyR != NIL_RTSOCKETNATIVE)
918	{
919	g_pfnclosesocket(pThis->hPollFallbackNotifyR);
920	pThis->hPollFallbackNotifyR = NIL_RTSOCKETNATIVE;
921	}
922	}
923	#endif
924	}
925
926	return rc;
927	}
928
929
930	RTDECL(uint32_t) RTSocketRelease(RTSOCKET hSocket)
931	{
932	RTSOCKETINT *pThis = hSocket;
933	if (pThis == NIL_RTSOCKET)
934	return 0;
935	AssertPtrReturn(pThis, UINT32_MAX);
936	AssertReturn(pThis->u32Magic == RTSOCKET_MAGIC, UINT32_MAX);
937
938	/* get the refcount without killing it... */
939	uint32_t cRefs = RTMemPoolRefCount(pThis);
940	AssertReturn(cRefs != UINT32_MAX, UINT32_MAX);
941	if (cRefs == 1)
942	rtSocketCloseIt(pThis, true);
943
944	return RTMemPoolRelease(RTMEMPOOL_DEFAULT, pThis);
945	}
946
947
948	RTDECL(int) RTSocketClose(RTSOCKET hSocket)
949	{
950	RTSOCKETINT *pThis = hSocket;
951	if (pThis == NIL_RTSOCKET)
952	return VINF_SUCCESS;
953	AssertPtrReturn(pThis, VERR_INVALID_HANDLE);
954	AssertReturn(pThis->u32Magic == RTSOCKET_MAGIC, VERR_INVALID_HANDLE);
955
956	uint32_t cRefs = RTMemPoolRefCount(pThis);
957	AssertReturn(cRefs != UINT32_MAX, UINT32_MAX);
958
959	int rc = rtSocketCloseIt(pThis, cRefs == 1);
960
961	RTMemPoolRelease(RTMEMPOOL_DEFAULT, pThis);
962	return rc;
963	}
964
965
966	RTDECL(RTHCUINTPTR) RTSocketToNative(RTSOCKET hSocket)
967	{
968	RTSOCKETINT *pThis = hSocket;
969	AssertPtrReturn(pThis, RTHCUINTPTR_MAX);
970	AssertReturn(pThis->u32Magic == RTSOCKET_MAGIC, RTHCUINTPTR_MAX);
971	return (RTHCUINTPTR)pThis->hNative;
972	}
973
974
975	RTDECL(int) RTSocketSetInheritance(RTSOCKET hSocket, bool fInheritable)
976	{
977	RTSOCKETINT *pThis = hSocket;
978	AssertPtrReturn(pThis, VERR_INVALID_HANDLE);
979	AssertReturn(pThis->u32Magic == RTSOCKET_MAGIC, VERR_INVALID_HANDLE);
980	AssertReturn(RTMemPoolRefCount(pThis) >= (pThis->cUsers ? 2U : 1U), VERR_CALLER_NO_REFERENCE);
981
982	#ifndef RT_OS_WINDOWS
983	if (fcntl(pThis->hNative, F_SETFD, fInheritable ? 0 : FD_CLOEXEC) < 0)
984	return RTErrConvertFromErrno(errno);
985	return VINF_SUCCESS;
986	#else
987	/* Windows is more complicated as sockets are complicated wrt inheritance
988	(see stackoverflow for details). In general, though we cannot hope to
989	make a socket really non-inheritable before vista as other layers in
990	the winsock maze may have additional handles associated with the socket. */
991	if (g_pfnGetHandleInformation)
992	{
993	/* Check if the handle is already in what seems to be the right state
994	before we try doing anything. */
995	DWORD fFlags;
996	if (g_pfnGetHandleInformation((HANDLE)pThis->hNative, &fFlags))
997	{
998	if (RT_BOOL(fFlags & HANDLE_FLAG_INHERIT) == fInheritable)
999	return VINF_SUCCESS;
1000	}
1001	}
1002
1003	if (!g_pfnSetHandleInformation)
1004	return VERR_NET_NOT_UNSUPPORTED;
1005
1006	if (!g_pfnSetHandleInformation((HANDLE)pThis->hNative, HANDLE_FLAG_INHERIT, fInheritable ? HANDLE_FLAG_INHERIT : 0))
1007	return RTErrConvertFromWin32(GetLastError());
1008	/** @todo Need we do something related to WS_SIO_ASSOCIATE_HANDLE or
1009	* WS_SIO_TRANSLATE_HANDLE? Or what other handles could be associated
1010	* with the socket? that we need to modify? */
1011
1012	return VINF_SUCCESS;
1013	#endif
1014	}
1015
1016
1017	static bool rtSocketIsIPv4Numerical(const char *pszAddress, PRTNETADDRIPV4 pAddr)
1018	{
1019
1020	/* Empty address resolves to the INADDR_ANY address (good for bind). */
1021	if (!pszAddress \|\| !*pszAddress)
1022	{
1023	pAddr->u = INADDR_ANY;
1024	return true;
1025	}
1026
1027	/* Four quads? */
1028	char psz = (char )pszAddress;
1029	for (int i = 0; i < 4; i++)
1030	{
1031	uint8_t u8;
1032	int rc = RTStrToUInt8Ex(psz, &psz, 0, &u8);
1033	if (rc != VINF_SUCCESS && rc != VWRN_TRAILING_CHARS)
1034	return false;
1035	if (*psz != (i < 3 ? '.' : '\0'))
1036	return false;
1037	psz++;
1038
1039	pAddr->au8[i] = u8; /* big endian */
1040	}
1041
1042	return true;
1043	}
1044
1045	RTDECL(int) RTSocketParseInetAddress(const char *pszAddress, unsigned uPort, PRTNETADDR pAddr)
1046	{
1047	int rc;
1048
1049	/*
1050	* Validate input.
1051	*/
1052	AssertReturn(uPort > 0, VERR_INVALID_PARAMETER);
1053	AssertPtrNullReturn(pszAddress, VERR_INVALID_POINTER);
1054
1055	/*
1056	* Resolve the address. Pretty crude at the moment, but we have to make
1057	* sure to not ask the NT 4 gethostbyname about an IPv4 address as it may
1058	* give a wrong answer.
1059	*/
1060	/** @todo this only supports IPv4, and IPv6 support needs to be added.
1061	* It probably needs to be converted to getaddrinfo(). */
1062	RTNETADDRIPV4 IPv4Quad;
1063	if (rtSocketIsIPv4Numerical(pszAddress, &IPv4Quad))
1064	{
1065	Log3(("rtSocketIsIPv4Numerical: %s -> %#x (%RTnaipv4)\n", pszAddress, IPv4Quad.u, IPv4Quad));
1066	RT_ZERO(*pAddr);
1067	pAddr->enmType = RTNETADDRTYPE_IPV4;
1068	pAddr->uPort = uPort;
1069	pAddr->uAddr.IPv4 = IPv4Quad;
1070	return VINF_SUCCESS;
1071	}
1072
1073	#ifdef RT_OS_WINDOWS
1074	/* Initialize WinSock and check version before we call gethostbyname. */
1075	if (!g_pfngethostbyname)
1076	return VERR_NET_NOT_UNSUPPORTED;
1077
1078	int rc2 = rtSocketInitWinsock();
1079	if (RT_FAILURE(rc2))
1080	return rc2;
1081
1082	# define gethostbyname g_pfngethostbyname
1083	#endif
1084
1085	struct hostent *pHostEnt;
1086	pHostEnt = gethostbyname(pszAddress);
1087	if (!pHostEnt)
1088	{
1089	rc = rtSocketResolverError();
1090	AssertMsg(rc == VERR_NET_HOST_NOT_FOUND,
1091	("Could not resolve '%s', rc=%Rrc\n", pszAddress, rc));
1092	return rc;
1093	}
1094
1095	if (pHostEnt->h_addrtype == AF_INET)
1096	{
1097	RT_ZERO(*pAddr);
1098	pAddr->enmType = RTNETADDRTYPE_IPV4;
1099	pAddr->uPort = uPort;
1100	pAddr->uAddr.IPv4.u = ((struct in_addr *)pHostEnt->h_addr)->s_addr;
1101	Log3(("gethostbyname: %s -> %#x (%RTnaipv4)\n", pszAddress, pAddr->uAddr.IPv4.u, pAddr->uAddr.IPv4));
1102	}
1103	else
1104	return VERR_NET_ADDRESS_FAMILY_NOT_SUPPORTED;
1105
1106	#ifdef RT_OS_WINDOWS
1107	# undef gethostbyname
1108	#endif
1109	return VINF_SUCCESS;
1110	}
1111
1112
1113	/*
1114	* New function to allow both ipv4 and ipv6 addresses to be resolved.
1115	* Breaks compatibility with windows before 2000.
1116	*/
1117	RTDECL(int) RTSocketQueryAddressStr(const char pszHost, char pszResult, size_t *pcbResult, PRTNETADDRTYPE penmAddrType)
1118	{
1119	AssertPtrReturn(pszHost, VERR_INVALID_POINTER);
1120	AssertPtrReturn(pcbResult, VERR_INVALID_POINTER);
1121	AssertPtrNullReturn(penmAddrType, VERR_INVALID_POINTER);
1122	AssertPtrNullReturn(pszResult, VERR_INVALID_POINTER);
1123
1124	#if defined(RT_OS_OS2) \|\| defined(RT_OS_WINDOWS) /** @todo dynamically resolve the APIs not present in NT4! */
1125	return VERR_NOT_SUPPORTED;
1126
1127	#else
1128	int rc;
1129	if (*pcbResult < 16)
1130	return VERR_NET_ADDRESS_NOT_AVAILABLE;
1131
1132	/* Setup the hint. */
1133	struct addrinfo grHints;
1134	RT_ZERO(grHints);
1135	grHints.ai_socktype = 0;
1136	grHints.ai_flags = 0;
1137	grHints.ai_protocol = 0;
1138	grHints.ai_family = AF_UNSPEC;
1139	if (penmAddrType)
1140	{
1141	switch (*penmAddrType)
1142	{
1143	case RTNETADDRTYPE_INVALID:
1144	/grHints.ai_family = AF_UNSPEC;/
1145	break;
1146	case RTNETADDRTYPE_IPV4:
1147	grHints.ai_family = AF_INET;
1148	break;
1149	case RTNETADDRTYPE_IPV6:
1150	grHints.ai_family = AF_INET6;
1151	break;
1152	default:
1153	AssertFailedReturn(VERR_INVALID_PARAMETER);
1154	}
1155	}
1156
1157	# ifdef RT_OS_WINDOWS
1158	/*
1159	* Winsock2 init
1160	*/
1161	if ( !g_pfngetaddrinfo
1162	\|\| !g_pfnfreeaddrinfo)
1163	return VERR_NET_NOT_UNSUPPORTED;
1164
1165	int rc2 = rtSocketInitWinsock();
1166	if (RT_FAILURE(rc2))
1167	return rc2;
1168
1169	# define getaddrinfo g_pfngetaddrinfo
1170	# define freeaddrinfo g_pfnfreeaddrinfo
1171	# endif
1172
1173	/** @todo r=bird: getaddrinfo and freeaddrinfo breaks the additions on NT4. */
1174	struct addrinfo *pgrResults = NULL;
1175	rc = getaddrinfo(pszHost, "", &grHints, &pgrResults);
1176	if (rc != 0)
1177	return VERR_NET_ADDRESS_NOT_AVAILABLE;
1178
1179	// return data
1180	// on multiple matches return only the first one
1181
1182	if (!pgrResults)
1183	return VERR_NET_ADDRESS_NOT_AVAILABLE;
1184
1185	struct addrinfo const *pgrResult = pgrResults->ai_next;
1186	if (!pgrResult)
1187	{
1188	freeaddrinfo(pgrResults);
1189	return VERR_NET_ADDRESS_NOT_AVAILABLE;
1190	}
1191
1192	RTNETADDRTYPE enmAddrType = RTNETADDRTYPE_INVALID;
1193	size_t cchIpAddress;
1194	char szIpAddress[48];
1195	if (pgrResult->ai_family == AF_INET)
1196	{
1197	struct sockaddr_in const pgrSa = (struct sockaddr_in const )pgrResult->ai_addr;
1198	cchIpAddress = RTStrPrintf(szIpAddress, sizeof(szIpAddress),
1199	"%RTnaipv4", pgrSa->sin_addr.s_addr);
1200	Assert(cchIpAddress >= 7 && cchIpAddress < sizeof(szIpAddress) - 1);
1201	enmAddrType = RTNETADDRTYPE_IPV4;
1202	rc = VINF_SUCCESS;
1203	}
1204	else if (pgrResult->ai_family == AF_INET6)
1205	{
1206	struct sockaddr_in6 const pgrSa6 = (struct sockaddr_in6 const )pgrResult->ai_addr;
1207	cchIpAddress = RTStrPrintf(szIpAddress, sizeof(szIpAddress),
1208	"%RTnaipv6", (PRTNETADDRIPV6)&pgrSa6->sin6_addr);
1209	enmAddrType = RTNETADDRTYPE_IPV6;
1210	rc = VINF_SUCCESS;
1211	}
1212	else
1213	{
1214	rc = VERR_NET_ADDRESS_NOT_AVAILABLE;
1215	szIpAddress[0] = '\0';
1216	cchIpAddress = 0;
1217	}
1218	freeaddrinfo(pgrResults);
1219
1220	/*
1221	* Copy out the result.
1222	*/
1223	size_t const cbResult = *pcbResult;
1224	*pcbResult = cchIpAddress + 1;
1225	if (cchIpAddress < cbResult)
1226	memcpy(pszResult, szIpAddress, cchIpAddress + 1);
1227	else
1228	{
1229	RT_BZERO(pszResult, cbResult);
1230	if (RT_SUCCESS(rc))
1231	rc = VERR_BUFFER_OVERFLOW;
1232	}
1233	if (penmAddrType && RT_SUCCESS(rc))
1234	*penmAddrType = enmAddrType;
1235	return rc;
1236
1237	# ifdef RT_OS_WINDOWS
1238	# undef getaddrinfo
1239	# undef freeaddrinfo
1240	# endif
1241	#endif /* !RT_OS_OS2 */
1242	}
1243
1244
1245	RTDECL(int) RTSocketRead(RTSOCKET hSocket, void pvBuffer, size_t cbBuffer, size_t pcbRead)
1246	{
1247	/*
1248	* Validate input.
1249	*/
1250	RTSOCKETINT *pThis = hSocket;
1251	AssertPtrReturn(pThis, VERR_INVALID_HANDLE);
1252	AssertReturn(pThis->u32Magic == RTSOCKET_MAGIC, VERR_INVALID_HANDLE);
1253	AssertReturn(cbBuffer > 0, VERR_INVALID_PARAMETER);
1254	AssertPtr(pvBuffer);
1255	#ifdef RT_OS_WINDOWS
1256	AssertReturn(g_pfnrecv, VERR_NET_NOT_UNSUPPORTED);
1257	# define recv g_pfnrecv
1258	#endif
1259	AssertReturn(rtSocketTryLock(pThis), VERR_CONCURRENT_ACCESS);
1260
1261	int rc = rtSocketSwitchBlockingMode(pThis, true /* fBlocking */);
1262	if (RT_FAILURE(rc))
1263	return rc;
1264
1265	/*
1266	* Read loop.
1267	* If pcbRead is NULL we have to fill the entire buffer!
1268	*/
1269	size_t cbRead = 0;
1270	size_t cbToRead = cbBuffer;
1271	for (;;)
1272	{
1273	rtSocketErrorReset();
1274	#ifdef RTSOCKET_MAX_READ
1275	int cbNow = cbToRead >= RTSOCKET_MAX_READ ? RTSOCKET_MAX_READ : (int)cbToRead;
1276	#else
1277	size_t cbNow = cbToRead;
1278	#endif
1279	ssize_t cbBytesRead = recv(pThis->hNative, (char *)pvBuffer + cbRead, cbNow, MSG_NOSIGNAL);
1280	if (cbBytesRead <= 0)
1281	{
1282	rc = rtSocketError();
1283	Assert(RT_FAILURE_NP(rc) \|\| cbBytesRead == 0);
1284	if (RT_SUCCESS_NP(rc))
1285	{
1286	if (!pcbRead)
1287	rc = VERR_NET_SHUTDOWN;
1288	else
1289	{
1290	*pcbRead = 0;
1291	rc = VINF_SUCCESS;
1292	}
1293	}
1294	break;
1295	}
1296	if (pcbRead)
1297	{
1298	/* return partial data */
1299	*pcbRead = cbBytesRead;
1300	break;
1301	}
1302
1303	/* read more? */
1304	cbRead += cbBytesRead;
1305	if (cbRead == cbBuffer)
1306	break;
1307
1308	/* next */
1309	cbToRead = cbBuffer - cbRead;
1310	}
1311
1312	rtSocketUnlock(pThis);
1313	#ifdef RT_OS_WINDOWS
1314	# undef recv
1315	#endif
1316	return rc;
1317	}
1318
1319
1320	RTDECL(int) RTSocketReadFrom(RTSOCKET hSocket, void pvBuffer, size_t cbBuffer, size_t pcbRead, PRTNETADDR pSrcAddr)
1321	{
1322	/*
1323	* Validate input.
1324	*/
1325	RTSOCKETINT *pThis = hSocket;
1326	AssertPtrReturn(pThis, VERR_INVALID_HANDLE);
1327	AssertReturn(pThis->u32Magic == RTSOCKET_MAGIC, VERR_INVALID_HANDLE);
1328	AssertReturn(cbBuffer > 0, VERR_INVALID_PARAMETER);
1329	AssertPtr(pvBuffer);
1330	AssertPtr(pcbRead);
1331	#ifdef RT_OS_WINDOWS
1332	AssertReturn(g_pfnrecvfrom, VERR_NET_NOT_UNSUPPORTED);
1333	# define recvfrom g_pfnrecvfrom
1334	#endif
1335	AssertReturn(rtSocketTryLock(pThis), VERR_CONCURRENT_ACCESS);
1336
1337	int rc = rtSocketSwitchBlockingMode(pThis, true /* fBlocking */);
1338	if (RT_FAILURE(rc))
1339	return rc;
1340
1341	/*
1342	* Read data.
1343	*/
1344	size_t cbRead = 0;
1345	size_t cbToRead = cbBuffer;
1346	rtSocketErrorReset();
1347	RTSOCKADDRUNION u;
1348	#ifdef RTSOCKET_MAX_READ
1349	int cbNow = cbToRead >= RTSOCKET_MAX_READ ? RTSOCKET_MAX_READ : (int)cbToRead;
1350	int cbAddr = sizeof(u);
1351	#else
1352	size_t cbNow = cbToRead;
1353	socklen_t cbAddr = sizeof(u);
1354	#endif
1355	ssize_t cbBytesRead = recvfrom(pThis->hNative, (char *)pvBuffer + cbRead, cbNow, MSG_NOSIGNAL, &u.Addr, &cbAddr);
1356	if (cbBytesRead <= 0)
1357	{
1358	rc = rtSocketError();
1359	Assert(RT_FAILURE_NP(rc) \|\| cbBytesRead == 0);
1360	if (RT_SUCCESS_NP(rc))
1361	{
1362	*pcbRead = 0;
1363	rc = VINF_SUCCESS;
1364	}
1365	}
1366	else
1367	{
1368	if (pSrcAddr)
1369	rc = rtSocketNetAddrFromAddr(&u, cbAddr, pSrcAddr);
1370	*pcbRead = cbBytesRead;
1371	}
1372
1373	rtSocketUnlock(pThis);
1374	#ifdef RT_OS_WINDOWS
1375	# undef recvfrom
1376	#endif
1377	return rc;
1378	}
1379
1380
1381	RTDECL(int) RTSocketWrite(RTSOCKET hSocket, const void *pvBuffer, size_t cbBuffer)
1382	{
1383	/*
1384	* Validate input.
1385	*/
1386	RTSOCKETINT *pThis = hSocket;
1387	AssertPtrReturn(pThis, VERR_INVALID_HANDLE);
1388	AssertReturn(pThis->u32Magic == RTSOCKET_MAGIC, VERR_INVALID_HANDLE);
1389	#ifdef RT_OS_WINDOWS
1390	AssertReturn(g_pfnsend, VERR_NET_NOT_UNSUPPORTED);
1391	# define send g_pfnsend
1392	#endif
1393	AssertReturn(rtSocketTryLock(pThis), VERR_CONCURRENT_ACCESS);
1394
1395	int rc = rtSocketSwitchBlockingMode(pThis, true /* fBlocking */);
1396	if (RT_FAILURE(rc))
1397	return rc;
1398
1399	/*
1400	* Try write all at once.
1401	*/
1402	#ifdef RTSOCKET_MAX_WRITE
1403	int cbNow = cbBuffer >= RTSOCKET_MAX_WRITE ? RTSOCKET_MAX_WRITE : (int)cbBuffer;
1404	#else
1405	size_t cbNow = cbBuffer >= SSIZE_MAX ? SSIZE_MAX : cbBuffer;
1406	#endif
1407	ssize_t cbWritten = send(pThis->hNative, (const char *)pvBuffer, cbNow, MSG_NOSIGNAL);
1408	if (RT_LIKELY((size_t)cbWritten == cbBuffer && cbWritten >= 0))
1409	rc = VINF_SUCCESS;
1410	else if (cbWritten < 0)
1411	rc = rtSocketError();
1412	else
1413	{
1414	/*
1415	* Unfinished business, write the remainder of the request. Must ignore
1416	* VERR_INTERRUPTED here if we've managed to send something.
1417	*/
1418	size_t cbSentSoFar = 0;
1419	for (;;)
1420	{
1421	/* advance */
1422	cbBuffer -= (size_t)cbWritten;
1423	if (!cbBuffer)
1424	break;
1425	cbSentSoFar += (size_t)cbWritten;
1426	pvBuffer = (char const *)pvBuffer + cbWritten;
1427
1428	/* send */
1429	#ifdef RTSOCKET_MAX_WRITE
1430	cbNow = cbBuffer >= RTSOCKET_MAX_WRITE ? RTSOCKET_MAX_WRITE : (int)cbBuffer;
1431	#else
1432	cbNow = cbBuffer >= SSIZE_MAX ? SSIZE_MAX : cbBuffer;
1433	#endif
1434	cbWritten = send(pThis->hNative, (const char *)pvBuffer, cbNow, MSG_NOSIGNAL);
1435	if (cbWritten >= 0)
1436	AssertMsg(cbBuffer >= (size_t)cbWritten, ("Wrote more than we requested!!! cbWritten=%zu cbBuffer=%zu rtSocketError()=%d\n",
1437	cbWritten, cbBuffer, rtSocketError()));
1438	else
1439	{
1440	rc = rtSocketError();
1441	if (rc != VERR_INTERNAL_ERROR \|\| cbSentSoFar == 0)
1442	break;
1443	cbWritten = 0;
1444	rc = VINF_SUCCESS;
1445	}
1446	}
1447	}
1448
1449	rtSocketUnlock(pThis);
1450	#ifdef RT_OS_WINDOWS
1451	# undef send
1452	#endif
1453	return rc;
1454	}
1455
1456
1457	RTDECL(int) RTSocketWriteTo(RTSOCKET hSocket, const void *pvBuffer, size_t cbBuffer, PCRTNETADDR pAddr)
1458	{
1459	/*
1460	* Validate input.
1461	*/
1462	RTSOCKETINT *pThis = hSocket;
1463	AssertPtrReturn(pThis, VERR_INVALID_HANDLE);
1464	AssertReturn(pThis->u32Magic == RTSOCKET_MAGIC, VERR_INVALID_HANDLE);
1465	#ifdef RT_OS_WINDOWS
1466	AssertReturn(g_pfnsendto, VERR_NET_NOT_UNSUPPORTED);
1467	# define sendto g_pfnsendto
1468	#endif
1469
1470	/* no locking since UDP reads may be done concurrently to writes, and
1471	* this is the normal use case of this code. */
1472
1473	int rc = rtSocketSwitchBlockingMode(pThis, true /* fBlocking */);
1474	if (RT_FAILURE(rc))
1475	return rc;
1476
1477	/* Figure out destination address. */
1478	struct sockaddr *pSA = NULL;
1479	#ifdef RT_OS_WINDOWS
1480	int cbSA = 0;
1481	#else
1482	socklen_t cbSA = 0;
1483	#endif
1484	RTSOCKADDRUNION u;
1485	if (pAddr)
1486	{
1487	rc = rtSocketAddrFromNetAddr(pAddr, &u, sizeof(u), NULL);
1488	if (RT_FAILURE(rc))
1489	return rc;
1490	pSA = &u.Addr;
1491	cbSA = sizeof(u);
1492	}
1493
1494	/*
1495	* Must write all at once, otherwise it is a failure.
1496	*/
1497	#ifdef RT_OS_WINDOWS
1498	int cbNow = cbBuffer >= RTSOCKET_MAX_WRITE ? RTSOCKET_MAX_WRITE : (int)cbBuffer;
1499	#else
1500	size_t cbNow = cbBuffer >= SSIZE_MAX ? SSIZE_MAX : cbBuffer;
1501	#endif
1502	ssize_t cbWritten = sendto(pThis->hNative, (const char *)pvBuffer, cbNow, MSG_NOSIGNAL, pSA, cbSA);
1503	if (RT_LIKELY((size_t)cbWritten == cbBuffer && cbWritten >= 0))
1504	rc = VINF_SUCCESS;
1505	else if (cbWritten < 0)
1506	rc = rtSocketError();
1507	else
1508	rc = VERR_TOO_MUCH_DATA;
1509
1510	/// @todo rtSocketUnlock(pThis);
1511	#ifdef RT_OS_WINDOWS
1512	# undef sendto
1513	#endif
1514	return rc;
1515	}
1516
1517
1518	RTDECL(int) RTSocketWriteToNB(RTSOCKET hSocket, const void *pvBuffer, size_t cbBuffer, PCRTNETADDR pAddr)
1519	{
1520	/*
1521	* Validate input.
1522	*/
1523	RTSOCKETINT *pThis = hSocket;
1524	AssertPtrReturn(pThis, VERR_INVALID_HANDLE);
1525	AssertReturn(pThis->u32Magic == RTSOCKET_MAGIC, VERR_INVALID_HANDLE);
1526	#ifdef RT_OS_WINDOWS
1527	AssertReturn(g_pfnsendto, VERR_NET_NOT_UNSUPPORTED);
1528	# define sendto g_pfnsendto
1529	#endif
1530
1531	/* no locking since UDP reads may be done concurrently to writes, and
1532	* this is the normal use case of this code. */
1533
1534	int rc = rtSocketSwitchBlockingMode(pThis, false /* fBlocking */);
1535	if (RT_FAILURE(rc))
1536	return rc;
1537
1538	/* Figure out destination address. */
1539	struct sockaddr *pSA = NULL;
1540	#ifdef RT_OS_WINDOWS
1541	int cbSA = 0;
1542	#else
1543	socklen_t cbSA = 0;
1544	#endif
1545	RTSOCKADDRUNION u;
1546	if (pAddr)
1547	{
1548	rc = rtSocketAddrFromNetAddr(pAddr, &u, sizeof(u), NULL);
1549	if (RT_FAILURE(rc))
1550	return rc;
1551	pSA = &u.Addr;
1552	cbSA = sizeof(u);
1553	}
1554
1555	/*
1556	* Must write all at once, otherwise it is a failure.
1557	*/
1558	#ifdef RT_OS_WINDOWS
1559	int cbNow = cbBuffer >= RTSOCKET_MAX_WRITE ? RTSOCKET_MAX_WRITE : (int)cbBuffer;
1560	#else
1561	size_t cbNow = cbBuffer >= SSIZE_MAX ? SSIZE_MAX : cbBuffer;
1562	#endif
1563	ssize_t cbWritten = sendto(pThis->hNative, (const char *)pvBuffer, cbNow, MSG_NOSIGNAL, pSA, cbSA);
1564	if (RT_LIKELY((size_t)cbWritten == cbBuffer && cbWritten >= 0))
1565	rc = VINF_SUCCESS;
1566	else if (cbWritten < 0)
1567	rc = rtSocketError();
1568	else
1569	rc = VERR_TOO_MUCH_DATA;
1570
1571	/// @todo rtSocketUnlock(pThis);
1572	#ifdef RT_OS_WINDOWS
1573	# undef sendto
1574	#endif
1575	return rc;
1576	}
1577
1578
1579	RTDECL(int) RTSocketSgWrite(RTSOCKET hSocket, PCRTSGBUF pSgBuf)
1580	{
1581	/*
1582	* Validate input.
1583	*/
1584	RTSOCKETINT *pThis = hSocket;
1585	AssertPtrReturn(pThis, VERR_INVALID_HANDLE);
1586	AssertReturn(pThis->u32Magic == RTSOCKET_MAGIC, VERR_INVALID_HANDLE);
1587	AssertPtrReturn(pSgBuf, VERR_INVALID_PARAMETER);
1588	AssertReturn(pSgBuf->cSegs > 0, VERR_INVALID_PARAMETER);
1589	AssertReturn(rtSocketTryLock(pThis), VERR_CONCURRENT_ACCESS);
1590
1591	int rc = rtSocketSwitchBlockingMode(pThis, true /* fBlocking */);
1592	if (RT_FAILURE(rc))
1593	return rc;
1594
1595	/*
1596	* Construct message descriptor (translate pSgBuf) and send it.
1597	*/
1598	rc = VERR_NO_TMP_MEMORY;
1599	#ifdef RT_OS_WINDOWS
1600	if (g_pfnWSASend)
1601	{
1602	AssertCompileSize(WSABUF, sizeof(RTSGSEG));
1603	AssertCompileMemberSize(WSABUF, buf, RT_SIZEOFMEMB(RTSGSEG, pvSeg));
1604
1605	LPWSABUF paMsg = (LPWSABUF)RTMemTmpAllocZ(pSgBuf->cSegs * sizeof(WSABUF));
1606	if (paMsg)
1607	{
1608	for (unsigned i = 0; i < pSgBuf->cSegs; i++)
1609	{
1610	paMsg[i].buf = (char *)pSgBuf->paSegs[i].pvSeg;
1611	paMsg[i].len = (u_long)pSgBuf->paSegs[i].cbSeg;
1612	}
1613
1614	DWORD dwSent;
1615	int hrc = g_pfnWSASend(pThis->hNative, paMsg, pSgBuf->cSegs, &dwSent, MSG_NOSIGNAL, NULL, NULL);
1616	if (!hrc)
1617	rc = VINF_SUCCESS;
1618	/** @todo check for incomplete writes */
1619	else
1620	rc = rtSocketError();
1621
1622	RTMemTmpFree(paMsg);
1623	}
1624	}
1625	else if (g_pfnsend)
1626	{
1627	rc = VINF_SUCCESS;
1628	for (uint32_t iSeg = 0; iSeg < pSgBuf->cSegs; iSeg++)
1629	{
1630	uint8_t const pbSeg = (uint8_t const )pSgBuf->paSegs[iSeg].pvSeg;
1631	size_t cbSeg = pSgBuf->paSegs[iSeg].cbSeg;
1632	int cbNow;
1633	ssize_t cbWritten;
1634	for (;;)
1635	{
1636	cbNow = cbSeg >= RTSOCKET_MAX_WRITE ? RTSOCKET_MAX_WRITE : (int)cbSeg;
1637	cbWritten = g_pfnsend(pThis->hNative, (const char *)pbSeg, cbNow, MSG_NOSIGNAL);
1638	if ((size_t)cbWritten >= cbSeg \|\| cbWritten < 0)
1639	break;
1640	pbSeg += cbWritten;
1641	cbSeg -= cbWritten;
1642	}
1643	if (cbWritten < 0)
1644	{
1645	rc = rtSocketError();
1646	break;
1647	}
1648	}
1649	}
1650	else
1651	rc = VERR_NET_NOT_UNSUPPORTED;
1652
1653	#else /* !RT_OS_WINDOWS */
1654	AssertCompileSize(struct iovec, sizeof(RTSGSEG));
1655	AssertCompileMemberSize(struct iovec, iov_base, RT_SIZEOFMEMB(RTSGSEG, pvSeg));
1656	AssertCompileMemberSize(struct iovec, iov_len, RT_SIZEOFMEMB(RTSGSEG, cbSeg));
1657
1658	struct iovec paMsg = (struct iovec )RTMemTmpAllocZ(pSgBuf->cSegs * sizeof(struct iovec));
1659	if (paMsg)
1660	{
1661	for (unsigned i = 0; i < pSgBuf->cSegs; i++)
1662	{
1663	paMsg[i].iov_base = pSgBuf->paSegs[i].pvSeg;
1664	paMsg[i].iov_len = pSgBuf->paSegs[i].cbSeg;
1665	}
1666
1667	struct msghdr msgHdr;
1668	RT_ZERO(msgHdr);
1669	msgHdr.msg_iov = paMsg;
1670	msgHdr.msg_iovlen = pSgBuf->cSegs;
1671	ssize_t cbWritten = sendmsg(pThis->hNative, &msgHdr, MSG_NOSIGNAL);
1672	if (RT_LIKELY(cbWritten >= 0))
1673	rc = VINF_SUCCESS;
1674	/** @todo check for incomplete writes */
1675	else
1676	rc = rtSocketError();
1677
1678	RTMemTmpFree(paMsg);
1679	}
1680	#endif /* !RT_OS_WINDOWS */
1681
1682	rtSocketUnlock(pThis);
1683	return rc;
1684	}
1685
1686
1687	RTDECL(int) RTSocketSgWriteL(RTSOCKET hSocket, size_t cSegs, ...)
1688	{
1689	va_list va;
1690	va_start(va, cSegs);
1691	int rc = RTSocketSgWriteLV(hSocket, cSegs, va);
1692	va_end(va);
1693	return rc;
1694	}
1695
1696
1697	RTDECL(int) RTSocketSgWriteLV(RTSOCKET hSocket, size_t cSegs, va_list va)
1698	{
1699	/*
1700	* Set up a S/G segment array + buffer on the stack and pass it
1701	* on to RTSocketSgWrite.
1702	*/
1703	Assert(cSegs <= 16);
1704	PRTSGSEG paSegs = (PRTSGSEG)alloca(cSegs * sizeof(RTSGSEG));
1705	AssertReturn(paSegs, VERR_NO_TMP_MEMORY);
1706	for (size_t i = 0; i < cSegs; i++)
1707	{
1708	paSegs[i].pvSeg = va_arg(va, void *);
1709	paSegs[i].cbSeg = va_arg(va, size_t);
1710	}
1711
1712	RTSGBUF SgBuf;
1713	RTSgBufInit(&SgBuf, paSegs, cSegs);
1714	return RTSocketSgWrite(hSocket, &SgBuf);
1715	}
1716
1717
1718	RTDECL(int) RTSocketReadNB(RTSOCKET hSocket, void pvBuffer, size_t cbBuffer, size_t pcbRead)
1719	{
1720	/*
1721	* Validate input.
1722	*/
1723	RTSOCKETINT *pThis = hSocket;
1724	AssertPtrReturn(pThis, VERR_INVALID_HANDLE);
1725	AssertReturn(pThis->u32Magic == RTSOCKET_MAGIC, VERR_INVALID_HANDLE);
1726	AssertReturn(cbBuffer > 0, VERR_INVALID_PARAMETER);
1727	AssertPtr(pvBuffer);
1728	AssertPtrReturn(pcbRead, VERR_INVALID_PARAMETER);
1729	#ifdef RT_OS_WINDOWS
1730	AssertReturn(g_pfnrecv, VERR_NET_NOT_UNSUPPORTED);
1731	#endif
1732	AssertReturn(rtSocketTryLock(pThis), VERR_CONCURRENT_ACCESS);
1733
1734	int rc = rtSocketSwitchBlockingMode(pThis, false /* fBlocking */);
1735	if (RT_FAILURE(rc))
1736	return rc;
1737
1738	rtSocketErrorReset();
1739	#ifdef RTSOCKET_MAX_READ
1740	int cbNow = cbBuffer >= RTSOCKET_MAX_WRITE ? RTSOCKET_MAX_WRITE : (int)cbBuffer;
1741	#else
1742	size_t cbNow = cbBuffer;
1743	#endif
1744
1745	#ifdef RT_OS_WINDOWS
1746	int cbRead = g_pfnrecv(pThis->hNative, (char *)pvBuffer, cbNow, MSG_NOSIGNAL);
1747	if (cbRead >= 0)
1748	{
1749	*pcbRead = cbRead;
1750	rc = VINF_SUCCESS;
1751	}
1752	else
1753	{
1754	rc = rtSocketError();
1755	if (rc == VERR_TRY_AGAIN)
1756	{
1757	*pcbRead = 0;
1758	rc = VINF_TRY_AGAIN;
1759	}
1760	}
1761
1762	#else
1763	ssize_t cbRead = recv(pThis->hNative, pvBuffer, cbNow, MSG_NOSIGNAL);
1764	if (cbRead >= 0)
1765	*pcbRead = cbRead;
1766	else if ( errno == EAGAIN
1767	# ifdef EWOULDBLOCK
1768	# if EWOULDBLOCK != EAGAIN
1769	\|\| errno == EWOULDBLOCK
1770	# endif
1771	# endif
1772	)
1773	{
1774	*pcbRead = 0;
1775	rc = VINF_TRY_AGAIN;
1776	}
1777	else
1778	rc = rtSocketError();
1779	#endif
1780
1781	rtSocketUnlock(pThis);
1782	return rc;
1783	}
1784
1785
1786	RTDECL(int) RTSocketWriteNB(RTSOCKET hSocket, const void pvBuffer, size_t cbBuffer, size_t pcbWritten)
1787	{
1788	/*
1789	* Validate input.
1790	*/
1791	RTSOCKETINT *pThis = hSocket;
1792	AssertPtrReturn(pThis, VERR_INVALID_HANDLE);
1793	AssertReturn(pThis->u32Magic == RTSOCKET_MAGIC, VERR_INVALID_HANDLE);
1794	AssertPtrReturn(pcbWritten, VERR_INVALID_PARAMETER);
1795	#ifdef RT_OS_WINDOWS
1796	AssertReturn(g_pfnsend, VERR_NET_NOT_UNSUPPORTED);
1797	#endif
1798	AssertReturn(rtSocketTryLock(pThis), VERR_CONCURRENT_ACCESS);
1799
1800	int rc = rtSocketSwitchBlockingMode(pThis, false /* fBlocking */);
1801	if (RT_FAILURE(rc))
1802	return rc;
1803
1804	rtSocketErrorReset();
1805	#ifdef RT_OS_WINDOWS
1806	# ifdef RTSOCKET_MAX_WRITE
1807	int cbNow = cbBuffer >= RTSOCKET_MAX_WRITE ? RTSOCKET_MAX_WRITE : (int)cbBuffer;
1808	# else
1809	size_t cbNow = cbBuffer;
1810	# endif
1811	int cbWritten = g_pfnsend(pThis->hNative, (const char *)pvBuffer, cbNow, MSG_NOSIGNAL);
1812	if (cbWritten >= 0)
1813	{
1814	*pcbWritten = cbWritten;
1815	rc = VINF_SUCCESS;
1816	}
1817	else
1818	{
1819	rc = rtSocketError();
1820	if (rc == VERR_TRY_AGAIN)
1821	{
1822	*pcbWritten = 0;
1823	rc = VINF_TRY_AGAIN;
1824	}
1825	}
1826	#else
1827	ssize_t cbWritten = send(pThis->hNative, pvBuffer, cbBuffer, MSG_NOSIGNAL);
1828	if (cbWritten >= 0)
1829	*pcbWritten = cbWritten;
1830	else if ( errno == EAGAIN
1831	# ifdef EWOULDBLOCK
1832	# if EWOULDBLOCK != EAGAIN
1833	\|\| errno == EWOULDBLOCK
1834	# endif
1835	# endif
1836	)
1837	{
1838	*pcbWritten = 0;
1839	rc = VINF_TRY_AGAIN;
1840	}
1841	else
1842	rc = rtSocketError();
1843	#endif
1844
1845	rtSocketUnlock(pThis);
1846	return rc;
1847	}
1848
1849
1850	RTDECL(int) RTSocketSgWriteNB(RTSOCKET hSocket, PCRTSGBUF pSgBuf, size_t *pcbWritten)
1851	{
1852	/*
1853	* Validate input.
1854	*/
1855	RTSOCKETINT *pThis = hSocket;
1856	AssertPtrReturn(pThis, VERR_INVALID_HANDLE);
1857	AssertReturn(pThis->u32Magic == RTSOCKET_MAGIC, VERR_INVALID_HANDLE);
1858	AssertPtrReturn(pSgBuf, VERR_INVALID_PARAMETER);
1859	AssertPtrReturn(pcbWritten, VERR_INVALID_PARAMETER);
1860	AssertReturn(pSgBuf->cSegs > 0, VERR_INVALID_PARAMETER);
1861	AssertReturn(rtSocketTryLock(pThis), VERR_CONCURRENT_ACCESS);
1862
1863	int rc = rtSocketSwitchBlockingMode(pThis, false /* fBlocking */);
1864	if (RT_FAILURE(rc))
1865	return rc;
1866
1867	unsigned cSegsToSend = 0;
1868	rc = VERR_NO_TMP_MEMORY;
1869	#ifdef RT_OS_WINDOWS
1870	if (g_pfnWSASend)
1871	{
1872	LPWSABUF paMsg = NULL;
1873	RTSgBufMapToNative(paMsg, pSgBuf, WSABUF, buf, char *, len, u_long, cSegsToSend);
1874	if (paMsg)
1875	{
1876	DWORD dwSent = 0;
1877	int hrc = g_pfnWSASend(pThis->hNative, paMsg, cSegsToSend, &dwSent, MSG_NOSIGNAL, NULL, NULL);
1878	if (!hrc)
1879	rc = VINF_SUCCESS;
1880	else
1881	rc = rtSocketError();
1882
1883	*pcbWritten = dwSent;
1884
1885	RTMemTmpFree(paMsg);
1886	}
1887	}
1888	else if (g_pfnsend)
1889	{
1890	size_t cbWrittenTotal = 0;
1891	rc = VINF_SUCCESS;
1892	for (uint32_t iSeg = 0; iSeg < pSgBuf->cSegs; iSeg++)
1893	{
1894	uint8_t const pbSeg = (uint8_t const )pSgBuf->paSegs[iSeg].pvSeg;
1895	size_t cbSeg = pSgBuf->paSegs[iSeg].cbSeg;
1896	int cbNow;
1897	ssize_t cbWritten;
1898	for (;;)
1899	{
1900	cbNow = cbSeg >= RTSOCKET_MAX_WRITE ? RTSOCKET_MAX_WRITE : (int)cbSeg;
1901	cbWritten = g_pfnsend(pThis->hNative, (const char *)pbSeg, cbNow, MSG_NOSIGNAL);
1902	if ((size_t)cbWritten >= cbSeg \|\| cbWritten < 0)
1903	break;
1904	cbWrittenTotal += cbWrittenTotal;
1905	pbSeg += cbWritten;
1906	cbSeg -= cbWritten;
1907	}
1908	if (cbWritten < 0)
1909	{
1910	rc = rtSocketError();
1911	break;
1912	}
1913	if (cbWritten != cbNow)
1914	break;
1915	}
1916	*pcbWritten = cbWrittenTotal;
1917	}
1918	else
1919	rc = VERR_NET_NOT_UNSUPPORTED;
1920
1921	#else /* !RT_OS_WINDOWS */
1922	struct iovec *paMsg = NULL;
1923
1924	RTSgBufMapToNative(paMsg, pSgBuf, struct iovec, iov_base, void *, iov_len, size_t, cSegsToSend);
1925	if (paMsg)
1926	{
1927	struct msghdr msgHdr;
1928	RT_ZERO(msgHdr);
1929	msgHdr.msg_iov = paMsg;
1930	msgHdr.msg_iovlen = cSegsToSend;
1931	ssize_t cbWritten = sendmsg(pThis->hNative, &msgHdr, MSG_NOSIGNAL);
1932	if (RT_LIKELY(cbWritten >= 0))
1933	{
1934	rc = VINF_SUCCESS;
1935	*pcbWritten = cbWritten;
1936	}
1937	else
1938	rc = rtSocketError();
1939
1940	RTMemTmpFree(paMsg);
1941	}
1942	#endif /* !RT_OS_WINDOWS */
1943
1944	rtSocketUnlock(pThis);
1945	return rc;
1946	}
1947
1948
1949	RTDECL(int) RTSocketSgWriteLNB(RTSOCKET hSocket, size_t cSegs, size_t *pcbWritten, ...)
1950	{
1951	va_list va;
1952	va_start(va, pcbWritten);
1953	int rc = RTSocketSgWriteLVNB(hSocket, cSegs, pcbWritten, va);
1954	va_end(va);
1955	return rc;
1956	}
1957
1958
1959	RTDECL(int) RTSocketSgWriteLVNB(RTSOCKET hSocket, size_t cSegs, size_t *pcbWritten, va_list va)
1960	{
1961	/*
1962	* Set up a S/G segment array + buffer on the stack and pass it
1963	* on to RTSocketSgWrite.
1964	*/
1965	Assert(cSegs <= 16);
1966	PRTSGSEG paSegs = (PRTSGSEG)alloca(cSegs * sizeof(RTSGSEG));
1967	AssertReturn(paSegs, VERR_NO_TMP_MEMORY);
1968	for (size_t i = 0; i < cSegs; i++)
1969	{
1970	paSegs[i].pvSeg = va_arg(va, void *);
1971	paSegs[i].cbSeg = va_arg(va, size_t);
1972	}
1973
1974	RTSGBUF SgBuf;
1975	RTSgBufInit(&SgBuf, paSegs, cSegs);
1976	return RTSocketSgWriteNB(hSocket, &SgBuf, pcbWritten);
1977	}
1978
1979
1980	RTDECL(int) RTSocketSelectOne(RTSOCKET hSocket, RTMSINTERVAL cMillies)
1981	{
1982	/*
1983	* Validate input.
1984	*/
1985	RTSOCKETINT *pThis = hSocket;
1986	AssertPtrReturn(pThis, VERR_INVALID_HANDLE);
1987	AssertReturn(pThis->u32Magic == RTSOCKET_MAGIC, VERR_INVALID_HANDLE);
1988	AssertReturn(RTMemPoolRefCount(pThis) >= (pThis->cUsers ? 2U : 1U), VERR_CALLER_NO_REFERENCE);
1989	int const fdMax = (int)pThis->hNative + 1;
1990	AssertReturn((RTSOCKETNATIVE)(fdMax - 1) == pThis->hNative, VERR_INTERNAL_ERROR_5);
1991	#ifdef RT_OS_WINDOWS
1992	AssertReturn(g_pfnselect, VERR_NET_NOT_UNSUPPORTED);
1993	# define select g_pfnselect
1994	#endif
1995
1996	/*
1997	* Set up the file descriptor sets and do the select.
1998	*/
1999	fd_set fdsetR;
2000	FD_ZERO(&fdsetR);
2001	FD_SET(pThis->hNative, &fdsetR);
2002
2003	fd_set fdsetE = fdsetR;
2004
2005	int rc;
2006	if (cMillies == RT_INDEFINITE_WAIT)
2007	rc = select(fdMax, &fdsetR, NULL, &fdsetE, NULL);
2008	else
2009	{
2010	struct timeval timeout;
2011	timeout.tv_sec = cMillies / 1000;
2012	timeout.tv_usec = (cMillies % 1000) * 1000;
2013	rc = select(fdMax, &fdsetR, NULL, &fdsetE, &timeout);
2014	}
2015	if (rc > 0)
2016	rc = VINF_SUCCESS;
2017	else if (rc == 0)
2018	rc = VERR_TIMEOUT;
2019	else
2020	rc = rtSocketError();
2021
2022	#ifdef RT_OS_WINDOWS
2023	# undef select
2024	#endif
2025	return rc;
2026	}
2027
2028
2029	/**
2030	* Internal worker for RTSocketSelectOneEx and rtSocketPollCheck (fallback)
2031	*
2032	* @returns IPRT status code
2033	* @param pThis The socket (valid).
2034	* @param fEvents The events to select for.
2035	* @param pfEvents Where to return the events.
2036	* @param cMillies How long to select for, in milliseconds.
2037	*/
2038	static int rtSocketSelectOneEx(RTSOCKET pThis, uint32_t fEvents, uint32_t *pfEvents, RTMSINTERVAL cMillies)
2039	{
2040	RTSOCKETNATIVE hNative = pThis->hNative;
2041	if (hNative == NIL_RTSOCKETNATIVE)
2042	{
2043	/* Socket is already closed? Possible we raced someone calling rtSocketCloseIt.
2044	Should we return a different status code? */
2045	*pfEvents = RTSOCKET_EVT_ERROR;
2046	return VINF_SUCCESS;
2047	}
2048
2049	int const fdMax = (int)hNative + 1;
2050	AssertReturn((RTSOCKETNATIVE)(fdMax - 1) == hNative, VERR_INTERNAL_ERROR_5);
2051	#ifdef RT_OS_WINDOWS
2052	AssertReturn(g_pfnselect, VERR_NET_NOT_UNSUPPORTED);
2053	AssertReturn(g_pfn__WSAFDIsSet, VERR_NET_NOT_UNSUPPORTED);
2054	# define select g_pfnselect
2055	# define __WSAFDIsSet g_pfn__WSAFDIsSet
2056	#endif
2057
2058	*pfEvents = 0;
2059
2060	/*
2061	* Set up the file descriptor sets and do the select.
2062	*/
2063	fd_set fdsetR;
2064	fd_set fdsetW;
2065	fd_set fdsetE;
2066	FD_ZERO(&fdsetR);
2067	FD_ZERO(&fdsetW);
2068	FD_ZERO(&fdsetE);
2069
2070	if (fEvents & RTSOCKET_EVT_READ)
2071	FD_SET(hNative, &fdsetR);
2072	if (fEvents & RTSOCKET_EVT_WRITE)
2073	FD_SET(hNative, &fdsetW);
2074	if (fEvents & RTSOCKET_EVT_ERROR)
2075	FD_SET(hNative, &fdsetE);
2076
2077	int rc;
2078	if (cMillies == RT_INDEFINITE_WAIT)
2079	rc = select(fdMax, &fdsetR, &fdsetW, &fdsetE, NULL);
2080	else
2081	{
2082	struct timeval timeout;
2083	timeout.tv_sec = cMillies / 1000;
2084	timeout.tv_usec = (cMillies % 1000) * 1000;
2085	rc = select(fdMax, &fdsetR, &fdsetW, &fdsetE, &timeout);
2086	}
2087	if (rc > 0)
2088	{
2089	if (pThis->hNative == hNative)
2090	{
2091	if (FD_ISSET(hNative, &fdsetR))
2092	*pfEvents \|= RTSOCKET_EVT_READ;
2093	if (FD_ISSET(hNative, &fdsetW))
2094	*pfEvents \|= RTSOCKET_EVT_WRITE;
2095	if (FD_ISSET(hNative, &fdsetE))
2096	*pfEvents \|= RTSOCKET_EVT_ERROR;
2097	rc = VINF_SUCCESS;
2098	}
2099	else
2100	{
2101	/* Socket was closed while we waited (rtSocketCloseIt). Different status code? */
2102	*pfEvents = RTSOCKET_EVT_ERROR;
2103	rc = VINF_SUCCESS;
2104	}
2105	}
2106	else if (rc == 0)
2107	rc = VERR_TIMEOUT;
2108	else
2109	rc = rtSocketError();
2110
2111	#ifdef RT_OS_WINDOWS
2112	# undef select
2113	# undef __WSAFDIsSet
2114	#endif
2115	return rc;
2116	}
2117
2118
2119	RTDECL(int) RTSocketSelectOneEx(RTSOCKET hSocket, uint32_t fEvents, uint32_t *pfEvents, RTMSINTERVAL cMillies)
2120	{
2121	/*
2122	* Validate input.
2123	*/
2124	RTSOCKETINT *pThis = hSocket;
2125	AssertPtrReturn(pThis, VERR_INVALID_HANDLE);
2126	AssertReturn(pThis->u32Magic == RTSOCKET_MAGIC, VERR_INVALID_HANDLE);
2127	AssertPtrReturn(pfEvents, VERR_INVALID_PARAMETER);
2128	AssertReturn(!(fEvents & ~RTSOCKET_EVT_VALID_MASK), VERR_INVALID_PARAMETER);
2129	AssertReturn(RTMemPoolRefCount(pThis) >= (pThis->cUsers ? 2U : 1U), VERR_CALLER_NO_REFERENCE);
2130
2131	return rtSocketSelectOneEx(pThis, fEvents, pfEvents, cMillies);
2132	}
2133
2134
2135	RTDECL(int) RTSocketShutdown(RTSOCKET hSocket, bool fRead, bool fWrite)
2136	{
2137	/*
2138	* Validate input, don't lock it because we might want to interrupt a call
2139	* active on a different thread.
2140	*/
2141	RTSOCKETINT *pThis = hSocket;
2142	AssertPtrReturn(pThis, VERR_INVALID_HANDLE);
2143	AssertReturn(pThis->u32Magic == RTSOCKET_MAGIC, VERR_INVALID_HANDLE);
2144	AssertReturn(RTMemPoolRefCount(pThis) >= (pThis->cUsers ? 2U : 1U), VERR_CALLER_NO_REFERENCE);
2145	AssertReturn(fRead \|\| fWrite, VERR_INVALID_PARAMETER);
2146	#ifdef RT_OS_WINDOWS
2147	AssertReturn(g_pfnshutdown, VERR_NET_NOT_UNSUPPORTED);
2148	# define shutdown g_pfnshutdown
2149	#endif
2150
2151	/*
2152	* Do the job.
2153	*/
2154	int rc = VINF_SUCCESS;
2155	int fHow;
2156	if (fRead && fWrite)
2157	fHow = SHUT_RDWR;
2158	else if (fRead)
2159	fHow = SHUT_RD;
2160	else
2161	fHow = SHUT_WR;
2162	if (shutdown(pThis->hNative, fHow) == -1)
2163	rc = rtSocketError();
2164
2165	#ifdef RT_OS_WINDOWS
2166	# undef shutdown
2167	#endif
2168	return rc;
2169	}
2170
2171
2172	RTDECL(int) RTSocketGetLocalAddress(RTSOCKET hSocket, PRTNETADDR pAddr)
2173	{
2174	/*
2175	* Validate input.
2176	*/
2177	RTSOCKETINT *pThis = hSocket;
2178	AssertPtrReturn(pThis, VERR_INVALID_HANDLE);
2179	AssertReturn(pThis->u32Magic == RTSOCKET_MAGIC, VERR_INVALID_HANDLE);
2180	AssertReturn(RTMemPoolRefCount(pThis) >= (pThis->cUsers ? 2U : 1U), VERR_CALLER_NO_REFERENCE);
2181	#ifdef RT_OS_WINDOWS
2182	AssertReturn(g_pfngetsockname, VERR_NET_NOT_UNSUPPORTED);
2183	# define getsockname g_pfngetsockname
2184	#endif
2185
2186	/*
2187	* Get the address and convert it.
2188	*/
2189	int rc;
2190	RTSOCKADDRUNION u;
2191	#ifdef RT_OS_WINDOWS
2192	int cbAddr = sizeof(u);
2193	#else
2194	socklen_t cbAddr = sizeof(u);
2195	#endif
2196	RT_ZERO(u);
2197	if (getsockname(pThis->hNative, &u.Addr, &cbAddr) == 0)
2198	rc = rtSocketNetAddrFromAddr(&u, cbAddr, pAddr);
2199	else
2200	rc = rtSocketError();
2201
2202	#ifdef RT_OS_WINDOWS
2203	# undef getsockname
2204	#endif
2205	return rc;
2206	}
2207
2208
2209	RTDECL(int) RTSocketGetPeerAddress(RTSOCKET hSocket, PRTNETADDR pAddr)
2210	{
2211	/*
2212	* Validate input.
2213	*/
2214	RTSOCKETINT *pThis = hSocket;
2215	AssertPtrReturn(pThis, VERR_INVALID_HANDLE);
2216	AssertReturn(pThis->u32Magic == RTSOCKET_MAGIC, VERR_INVALID_HANDLE);
2217	AssertReturn(RTMemPoolRefCount(pThis) >= (pThis->cUsers ? 2U : 1U), VERR_CALLER_NO_REFERENCE);
2218	#ifdef RT_OS_WINDOWS
2219	AssertReturn(g_pfngetpeername, VERR_NET_NOT_UNSUPPORTED);
2220	# define getpeername g_pfngetpeername
2221	#endif
2222
2223	/*
2224	* Get the address and convert it.
2225	*/
2226	int rc;
2227	RTSOCKADDRUNION u;
2228	#ifdef RT_OS_WINDOWS
2229	int cbAddr = sizeof(u);
2230	#else
2231	socklen_t cbAddr = sizeof(u);
2232	#endif
2233	RT_ZERO(u);
2234	if (getpeername(pThis->hNative, &u.Addr, &cbAddr) == 0)
2235	rc = rtSocketNetAddrFromAddr(&u, cbAddr, pAddr);
2236	else
2237	rc = rtSocketError();
2238
2239	#ifdef RT_OS_WINDOWS
2240	# undef getpeername
2241	#endif
2242	return rc;
2243	}
2244
2245
2246
2247	/**
2248	* Wrapper around bind.
2249	*
2250	* @returns IPRT status code.
2251	* @param hSocket The socket handle.
2252	* @param pAddr The address to bind to.
2253	*/
2254	DECLHIDDEN(int) rtSocketBind(RTSOCKET hSocket, PCRTNETADDR pAddr)
2255	{
2256	RTSOCKADDRUNION u;
2257	int cbAddr;
2258	int rc = rtSocketAddrFromNetAddr(pAddr, &u, sizeof(u), &cbAddr);
2259	if (RT_SUCCESS(rc))
2260	rc = rtSocketBindRawAddr(hSocket, &u.Addr, cbAddr);
2261	return rc;
2262	}
2263
2264
2265	/**
2266	* Very thin wrapper around bind.
2267	*
2268	* @returns IPRT status code.
2269	* @param hSocket The socket handle.
2270	* @param pvAddr The address to bind to (struct sockaddr and
2271	* friends).
2272	* @param cbAddr The size of the address.
2273	*/
2274	DECLHIDDEN(int) rtSocketBindRawAddr(RTSOCKET hSocket, void const *pvAddr, size_t cbAddr)
2275	{
2276	/*
2277	* Validate input.
2278	*/
2279	RTSOCKETINT *pThis = hSocket;
2280	AssertPtrReturn(pThis, VERR_INVALID_HANDLE);
2281	AssertReturn(pThis->u32Magic == RTSOCKET_MAGIC, VERR_INVALID_HANDLE);
2282	AssertPtrReturn(pvAddr, VERR_INVALID_POINTER);
2283	#ifdef RT_OS_WINDOWS
2284	AssertReturn(g_pfnbind, VERR_NET_NOT_UNSUPPORTED);
2285	# define bind g_pfnbind
2286	#endif
2287	AssertReturn(rtSocketTryLock(pThis), VERR_CONCURRENT_ACCESS);
2288
2289	int rc;
2290	if (bind(pThis->hNative, (struct sockaddr const *)pvAddr, (int)cbAddr) == 0)
2291	rc = VINF_SUCCESS;
2292	else
2293	rc = rtSocketError();
2294
2295	rtSocketUnlock(pThis);
2296	#ifdef RT_OS_WINDOWS
2297	# undef bind
2298	#endif
2299	return rc;
2300	}
2301
2302
2303
2304	/**
2305	* Wrapper around listen.
2306	*
2307	* @returns IPRT status code.
2308	* @param hSocket The socket handle.
2309	* @param cMaxPending The max number of pending connections.
2310	*/
2311	DECLHIDDEN(int) rtSocketListen(RTSOCKET hSocket, int cMaxPending)
2312	{
2313	/*
2314	* Validate input.
2315	*/
2316	RTSOCKETINT *pThis = hSocket;
2317	AssertPtrReturn(pThis, VERR_INVALID_HANDLE);
2318	AssertReturn(pThis->u32Magic == RTSOCKET_MAGIC, VERR_INVALID_HANDLE);
2319	#ifdef RT_OS_WINDOWS
2320	AssertReturn(g_pfnlisten, VERR_NET_NOT_UNSUPPORTED);
2321	# define listen g_pfnlisten
2322	#endif
2323	AssertReturn(rtSocketTryLock(pThis), VERR_CONCURRENT_ACCESS);
2324
2325	int rc = VINF_SUCCESS;
2326	if (listen(pThis->hNative, cMaxPending) != 0)
2327	rc = rtSocketError();
2328
2329	rtSocketUnlock(pThis);
2330	#ifdef RT_OS_WINDOWS
2331	# undef listen
2332	#endif
2333	return rc;
2334	}
2335
2336
2337	/**
2338	* Wrapper around accept.
2339	*
2340	* @returns IPRT status code.
2341	* @param hSocket The socket handle.
2342	* @param phClient Where to return the client socket handle on
2343	* success.
2344	* @param pAddr Where to return the client address.
2345	* @param pcbAddr On input this gives the size buffer size of what
2346	* @a pAddr point to. On return this contains the
2347	* size of what's stored at @a pAddr.
2348	*/
2349	DECLHIDDEN(int) rtSocketAccept(RTSOCKET hSocket, PRTSOCKET phClient, struct sockaddr pAddr, size_t pcbAddr)
2350	{
2351	/*
2352	* Validate input.
2353	* Only lock the socket temporarily while we get the native handle, so that
2354	* we can safely shutdown and destroy the socket from a different thread.
2355	*/
2356	RTSOCKETINT *pThis = hSocket;
2357	AssertPtrReturn(pThis, VERR_INVALID_HANDLE);
2358	AssertReturn(pThis->u32Magic == RTSOCKET_MAGIC, VERR_INVALID_HANDLE);
2359	#ifdef RT_OS_WINDOWS
2360	AssertReturn(g_pfnaccept, VERR_NET_NOT_UNSUPPORTED);
2361	AssertReturn(g_pfnclosesocket, VERR_NET_NOT_UNSUPPORTED);
2362	# define accept g_pfnaccept
2363	#endif
2364	AssertReturn(rtSocketTryLock(pThis), VERR_CONCURRENT_ACCESS);
2365
2366	/*
2367	* Call accept().
2368	*/
2369	rtSocketErrorReset();
2370	int rc = VINF_SUCCESS;
2371	#ifdef RT_OS_WINDOWS
2372	int cbAddr = (int)*pcbAddr;
2373	#else
2374	socklen_t cbAddr = *pcbAddr;
2375	#endif
2376	RTSOCKETNATIVE hNativeClient = accept(pThis->hNative, pAddr, &cbAddr);
2377	if (hNativeClient != NIL_RTSOCKETNATIVE)
2378	{
2379	*pcbAddr = cbAddr;
2380
2381	/*
2382	* Wrap the client socket.
2383	*/
2384	rc = rtSocketCreateForNative(phClient, hNativeClient, false /fLeaveOpen/);
2385	if (RT_FAILURE(rc))
2386	{
2387	#ifdef RT_OS_WINDOWS
2388	g_pfnclosesocket(hNativeClient);
2389	#else
2390	close(hNativeClient);
2391	#endif
2392	}
2393	}
2394	else
2395	rc = rtSocketError();
2396
2397	rtSocketUnlock(pThis);
2398	#ifdef RT_OS_WINDOWS
2399	# undef accept
2400	#endif
2401	return rc;
2402	}
2403
2404
2405	/**
2406	* Wrapper around connect.
2407	*
2408	* @returns IPRT status code.
2409	* @param hSocket The socket handle.
2410	* @param pAddr The socket address to connect to.
2411	* @param cMillies Number of milliseconds to wait for the connect attempt to complete.
2412	* Use RT_INDEFINITE_WAIT to wait for ever.
2413	* Use RT_TCPCLIENTCONNECT_DEFAULT_WAIT to wait for the default time
2414	* configured on the running system.
2415	*/
2416	DECLHIDDEN(int) rtSocketConnect(RTSOCKET hSocket, PCRTNETADDR pAddr, RTMSINTERVAL cMillies)
2417	{
2418	/*
2419	* Validate input.
2420	*/
2421	RTSOCKETINT *pThis = hSocket;
2422	AssertPtrReturn(pThis, VERR_INVALID_HANDLE);
2423	AssertReturn(pThis->u32Magic == RTSOCKET_MAGIC, VERR_INVALID_HANDLE);
2424	#ifdef RT_OS_WINDOWS
2425	AssertReturn(g_pfnconnect, VERR_NET_NOT_UNSUPPORTED);
2426	AssertReturn(g_pfnselect, VERR_NET_NOT_UNSUPPORTED);
2427	AssertReturn(g_pfngetsockopt, VERR_NET_NOT_UNSUPPORTED);
2428	# define connect g_pfnconnect
2429	# define select g_pfnselect
2430	# define getsockopt g_pfngetsockopt
2431	#endif
2432	AssertReturn(rtSocketTryLock(pThis), VERR_CONCURRENT_ACCESS);
2433
2434	RTSOCKADDRUNION u;
2435	int cbAddr;
2436	int rc = rtSocketAddrFromNetAddr(pAddr, &u, sizeof(u), &cbAddr);
2437	if (RT_SUCCESS(rc))
2438	{
2439	if (cMillies == RT_SOCKETCONNECT_DEFAULT_WAIT)
2440	{
2441	if (connect(pThis->hNative, &u.Addr, cbAddr) != 0)
2442	rc = rtSocketError();
2443	}
2444	else
2445	{
2446	/*
2447	* Switch the socket to nonblocking mode, initiate the connect
2448	* and wait for the socket to become writable or until the timeout
2449	* expires.
2450	*/
2451	rc = rtSocketSwitchBlockingMode(pThis, false /* fBlocking */);
2452	if (RT_SUCCESS(rc))
2453	{
2454	if (connect(pThis->hNative, &u.Addr, cbAddr) != 0)
2455	{
2456	rc = rtSocketError();
2457	if (rc == VERR_TRY_AGAIN \|\| rc == VERR_NET_IN_PROGRESS)
2458	{
2459	int rcSock = 0;
2460	fd_set FdSetWriteable;
2461	struct timeval TvTimeout;
2462
2463	TvTimeout.tv_sec = cMillies / RT_MS_1SEC;
2464	TvTimeout.tv_usec = (cMillies % RT_MS_1SEC) * RT_US_1MS;
2465
2466	FD_ZERO(&FdSetWriteable);
2467	FD_SET(pThis->hNative, &FdSetWriteable);
2468	do
2469	{
2470	rcSock = select(pThis->hNative + 1, NULL, &FdSetWriteable, NULL,
2471	cMillies == RT_INDEFINITE_WAIT \|\| cMillies >= INT_MAX
2472	? NULL
2473	: &TvTimeout);
2474	if (rcSock > 0)
2475	{
2476	int iSockError = 0;
2477	socklen_t cbSockOpt = sizeof(iSockError);
2478	rcSock = getsockopt(pThis->hNative, SOL_SOCKET, SO_ERROR, (char *)&iSockError, &cbSockOpt);
2479	if (rcSock == 0)
2480	{
2481	if (iSockError == 0)
2482	rc = VINF_SUCCESS;
2483	else
2484	{
2485	#ifdef RT_OS_WINDOWS
2486	rc = RTErrConvertFromWin32(iSockError);
2487	#else
2488	rc = RTErrConvertFromErrno(iSockError);
2489	#endif
2490	}
2491	}
2492	else
2493	rc = rtSocketError();
2494	}
2495	else if (rcSock == 0)
2496	rc = VERR_TIMEOUT;
2497	else
2498	rc = rtSocketError();
2499	} while (rc == VERR_INTERRUPTED);
2500	}
2501	}
2502
2503	rtSocketSwitchBlockingMode(pThis, true /* fBlocking */);
2504	}
2505	}
2506	}
2507
2508	rtSocketUnlock(pThis);
2509	#ifdef RT_OS_WINDOWS
2510	# undef connect
2511	# undef select
2512	# undef getsockopt
2513	#endif
2514	return rc;
2515	}
2516
2517
2518	/**
2519	* Wrapper around connect, raw address, no timeout.
2520	*
2521	* @returns IPRT status code.
2522	* @param hSocket The socket handle.
2523	* @param pvAddr The raw socket address to connect to.
2524	* @param cbAddr The size of the raw address.
2525	*/
2526	DECLHIDDEN(int) rtSocketConnectRaw(RTSOCKET hSocket, void const *pvAddr, size_t cbAddr)
2527	{
2528	/*
2529	* Validate input.
2530	*/
2531	RTSOCKETINT *pThis = hSocket;
2532	AssertPtrReturn(pThis, VERR_INVALID_HANDLE);
2533	AssertReturn(pThis->u32Magic == RTSOCKET_MAGIC, VERR_INVALID_HANDLE);
2534	#ifdef RT_OS_WINDOWS
2535	AssertReturn(g_pfnconnect, VERR_NET_NOT_UNSUPPORTED);
2536	# define connect g_pfnconnect
2537	#endif
2538	AssertReturn(rtSocketTryLock(pThis), VERR_CONCURRENT_ACCESS);
2539
2540	int rc;
2541	if (connect(pThis->hNative, (const struct sockaddr *)pvAddr, (int)cbAddr) == 0)
2542	rc = VINF_SUCCESS;
2543	else
2544	rc = rtSocketError();
2545
2546	rtSocketUnlock(pThis);
2547	#ifdef RT_OS_WINDOWS
2548	# undef connect
2549	#endif
2550	return rc;
2551	}
2552
2553
2554	/**
2555	* Wrapper around setsockopt.
2556	*
2557	* @returns IPRT status code.
2558	* @param hSocket The socket handle.
2559	* @param iLevel The protocol level, e.g. IPPORTO_TCP.
2560	* @param iOption The option, e.g. TCP_NODELAY.
2561	* @param pvValue The value buffer.
2562	* @param cbValue The size of the value pointed to by pvValue.
2563	*/
2564	DECLHIDDEN(int) rtSocketSetOpt(RTSOCKET hSocket, int iLevel, int iOption, void const *pvValue, int cbValue)
2565	{
2566	/*
2567	* Validate input.
2568	*/
2569	RTSOCKETINT *pThis = hSocket;
2570	AssertPtrReturn(pThis, VERR_INVALID_HANDLE);
2571	AssertReturn(pThis->u32Magic == RTSOCKET_MAGIC, VERR_INVALID_HANDLE);
2572	#ifdef RT_OS_WINDOWS
2573	AssertReturn(g_pfnsetsockopt, VERR_NET_NOT_UNSUPPORTED);
2574	# define setsockopt g_pfnsetsockopt
2575	#endif
2576	AssertReturn(rtSocketTryLock(pThis), VERR_CONCURRENT_ACCESS);
2577
2578	int rc = VINF_SUCCESS;
2579	if (setsockopt(pThis->hNative, iLevel, iOption, (const char *)pvValue, cbValue) != 0)
2580	rc = rtSocketError();
2581
2582	rtSocketUnlock(pThis);
2583	#ifdef RT_OS_WINDOWS
2584	# undef setsockopt
2585	#endif
2586	return rc;
2587	}
2588
2589
2590	/**
2591	* Internal RTPollSetAdd helper that returns the handle that should be added to
2592	* the pollset.
2593	*
2594	* @returns Valid handle on success, INVALID_HANDLE_VALUE on failure.
2595	* @param hSocket The socket handle.
2596	* @param fEvents The events we're polling for.
2597	* @param phNative Where to put the primary handle.
2598	*/
2599	DECLHIDDEN(int) rtSocketPollGetHandle(RTSOCKET hSocket, uint32_t fEvents, PRTHCINTPTR phNative)
2600	{
2601	RTSOCKETINT *pThis = hSocket;
2602	RT_NOREF_PV(fEvents);
2603	AssertPtrReturn(pThis, VERR_INVALID_HANDLE);
2604	AssertReturn(pThis->u32Magic == RTSOCKET_MAGIC, VERR_INVALID_HANDLE);
2605	#ifdef RT_OS_WINDOWS
2606	AssertReturn(rtSocketTryLock(pThis), VERR_CONCURRENT_ACCESS);
2607
2608	int rc = VINF_SUCCESS;
2609	if (pThis->hEvent != WSA_INVALID_EVENT)
2610	*phNative = (RTHCINTPTR)pThis->hEvent;
2611	else if (g_pfnWSACreateEvent)
2612	{
2613	pThis->hEvent = g_pfnWSACreateEvent();
2614	*phNative = (RTHCINTPTR)pThis->hEvent;
2615	if (pThis->hEvent == WSA_INVALID_EVENT)
2616	rc = rtSocketError();
2617	}
2618	else
2619	{
2620	AssertCompile(WSA_INVALID_EVENT == (WSAEVENT)NULL);
2621	pThis->hEvent = CreateEventW(NULL, TRUE /fManualReset/, FALSE /fInitialState/, NULL /pwszName/);
2622	*phNative = (RTHCINTPTR)pThis->hEvent;
2623	if (pThis->hEvent == WSA_INVALID_EVENT)
2624	rc = RTErrConvertFromWin32(GetLastError());
2625	}
2626
2627	rtSocketUnlock(pThis);
2628	return rc;
2629
2630	#else /* !RT_OS_WINDOWS */
2631	*phNative = (RTHCUINTPTR)pThis->hNative;
2632	return VINF_SUCCESS;
2633	#endif /* !RT_OS_WINDOWS */
2634	}
2635
2636	#ifdef RT_OS_WINDOWS
2637
2638	/**
2639	* Fallback poller thread.
2640	*
2641	* @returns VINF_SUCCESS.
2642	* @param hSelf The thread handle.
2643	* @param pvUser Socket instance data.
2644	*/
2645	static DECLCALLBACK(int) rtSocketPollFallbackThreadProc(RTTHREAD hSelf, void *pvUser)
2646	{
2647	RTSOCKETINT pThis = (RTSOCKETINT )pvUser;
2648	RT_NOREF(hSelf);
2649	# define __WSAFDIsSet g_pfn__WSAFDIsSet
2650
2651	/*
2652	* The execution loop.
2653	*/
2654	while (!ASMAtomicReadBool(&pThis->fPollFallbackShutdown))
2655	{
2656	/*
2657	* Do the selecting (with a 15 second timeout because that seems like a good idea).
2658	*/
2659	struct fd_set SetRead;
2660	struct fd_set SetWrite;
2661	struct fd_set SetXcpt;
2662
2663	FD_ZERO(&SetRead);
2664	FD_ZERO(&SetWrite);
2665	FD_ZERO(&SetXcpt);
2666
2667	FD_SET(pThis->hPollFallbackNotifyR, &SetRead);
2668	FD_SET(pThis->hPollFallbackNotifyR, &SetXcpt);
2669
2670	bool fActive = ASMAtomicReadBool(&pThis->fPollFallbackActive);
2671	uint32_t fEvents;
2672	if (!fActive)
2673	fEvents = 0;
2674	else
2675	{
2676	fEvents = ASMAtomicReadU32(&pThis->fSubscribedEvts);
2677	if (fEvents & RTPOLL_EVT_READ)
2678	FD_SET(pThis->hNative, &SetRead);
2679	if (fEvents & RTPOLL_EVT_WRITE)
2680	FD_SET(pThis->hNative, &SetWrite);
2681	if (fEvents & RTPOLL_EVT_ERROR)
2682	FD_SET(pThis->hNative, &SetXcpt);
2683	}
2684
2685	struct timeval Timeout;
2686	Timeout.tv_sec = 15;
2687	Timeout.tv_usec = 0;
2688	int rc = g_pfnselect(INT_MAX /ignored/, &SetRead, &SetWrite, &SetXcpt, &Timeout);
2689
2690	/* Stop immediately if told to shut down. */
2691	if (ASMAtomicReadBool(&pThis->fPollFallbackShutdown))
2692	break;
2693
2694	/*
2695	* Process the result.
2696	*/
2697	if (rc > 0)
2698	{
2699	/* First the socket we're listening on. */
2700	if ( fEvents
2701	&& ( FD_ISSET(pThis->hNative, &SetRead)
2702	\|\| FD_ISSET(pThis->hNative, &SetWrite)
2703	\|\| FD_ISSET(pThis->hNative, &SetXcpt)) )
2704	{
2705	ASMAtomicWriteBool(&pThis->fPollFallbackActive, false);
2706	SetEvent(pThis->hEvent);
2707	}
2708
2709	/* Then maintain the notification pipe. (We only read one byte here
2710	because we're overly paranoid wrt socket switching to blocking mode.) */
2711	if (FD_ISSET(pThis->hPollFallbackNotifyR, &SetRead))
2712	{
2713	char chIgnored;
2714	g_pfnrecv(pThis->hPollFallbackNotifyR, &chIgnored, sizeof(chIgnored), MSG_NOSIGNAL);
2715	}
2716	}
2717	else
2718	AssertMsg(rc == 0, ("%Rrc\n", rtSocketError()));
2719	}
2720
2721	# undef __WSAFDIsSet
2722	return VINF_SUCCESS;
2723	}
2724
2725
2726	/**
2727	* Pokes the fallback thread, making sure it gets out of whatever it's stuck in.
2728	*
2729	* @param pThis The socket handle.
2730	*/
2731	static void rtSocketPokePollFallbackThread(RTSOCKETINT *pThis)
2732	{
2733	Assert(pThis->fPollFallback);
2734	if (pThis->hPollFallbackThread != NIL_RTTHREAD)
2735	{
2736	int cbWritten = g_pfnsend(pThis->hPollFallbackNotifyW, "!", 1, MSG_NOSIGNAL);
2737	AssertMsg(cbWritten == 1, ("cbWritten=%d err=%Rrc\n", rtSocketError()));
2738	RT_NOREF_PV(cbWritten);
2739	}
2740	}
2741
2742
2743	/**
2744	* Called by rtSocketPollStart to make the thread start selecting on the socket.
2745	*
2746	* @returns 0 on success, RTPOLL_EVT_ERROR on failure.
2747	* @param pThis The socket handle.
2748	*/
2749	static uint32_t rtSocketPollFallbackStart(RTSOCKETINT *pThis)
2750	{
2751	/*
2752	* Reset the event and tell the thread to start selecting on the socket.
2753	*/
2754	ResetEvent(pThis->hEvent);
2755	ASMAtomicWriteBool(&pThis->fPollFallbackActive, true);
2756
2757	/*
2758	* Wake up the thread the thread.
2759	*/
2760	if (pThis->hPollFallbackThread != NIL_RTTHREAD)
2761	rtSocketPokePollFallbackThread(pThis);
2762	else
2763	{
2764	/*
2765	* Not running, need to set it up and start it.
2766	*/
2767	AssertLogRelReturn(pThis->hEvent != NULL && pThis->hEvent != INVALID_HANDLE_VALUE, RTPOLL_EVT_ERROR);
2768
2769	/* Create the notification socket pair. */
2770	int rc;
2771	if (pThis->hPollFallbackNotifyR == NIL_RTSOCKETNATIVE)
2772	{
2773	rc = rtSocketCreateNativeTcpPair(&pThis->hPollFallbackNotifyW, &pThis->hPollFallbackNotifyR);
2774	AssertLogRelRCReturn(rc, RTPOLL_EVT_ERROR);
2775
2776	/* Make the read end non-blocking (not fatal). */
2777	u_long fNonBlocking = 1;
2778	rc = g_pfnioctlsocket(pThis->hPollFallbackNotifyR, FIONBIO, &fNonBlocking);
2779	AssertLogRelMsg(rc == 0, ("rc=%#x %Rrc\n", rc, rtSocketError()));
2780	}
2781
2782	/* Finally, start the thread. ASSUME we don't need too much stack. */
2783	rc = RTThreadCreate(&pThis->hPollFallbackThread, rtSocketPollFallbackThreadProc, pThis,
2784	_128K, RTTHREADTYPE_IO, RTTHREADFLAGS_WAITABLE, "sockpoll");
2785	AssertLogRelRCReturn(rc, RTPOLL_EVT_ERROR);
2786	}
2787	return 0;
2788	}
2789
2790
2791	/**
2792	* Undos the harm done by WSAEventSelect.
2793	*
2794	* @returns IPRT status code.
2795	* @param pThis The socket handle.
2796	*/
2797	static int rtSocketPollClearEventAndRestoreBlocking(RTSOCKETINT *pThis)
2798	{
2799	int rc = VINF_SUCCESS;
2800	if (pThis->fSubscribedEvts)
2801	{
2802	if (!pThis->fPollFallback)
2803	{
2804	Assert(g_pfnWSAEventSelect && g_pfnioctlsocket);
2805	if (g_pfnWSAEventSelect && g_pfnioctlsocket)
2806	{
2807	if (g_pfnWSAEventSelect(pThis->hNative, WSA_INVALID_EVENT, 0) == 0)
2808	{
2809	pThis->fSubscribedEvts = 0;
2810
2811	/*
2812	* Switch back to blocking mode if that was the state before the
2813	* operation.
2814	*/
2815	if (pThis->fBlocking)
2816	{
2817	u_long fNonBlocking = 0;
2818	int rc2 = g_pfnioctlsocket(pThis->hNative, FIONBIO, &fNonBlocking);
2819	if (rc2 != 0)
2820	{
2821	rc = rtSocketError();
2822	AssertMsgFailed(("%Rrc; rc2=%d\n", rc, rc2));
2823	}
2824	}
2825	}
2826	else
2827	{
2828	rc = rtSocketError();
2829	AssertMsgFailed(("%Rrc\n", rc));
2830	}
2831	}
2832	else
2833	{
2834	Assert(pThis->fPollFallback);
2835	rc = VINF_SUCCESS;
2836	}
2837	}
2838	/*
2839	* Just clear the event mask as we never started waiting if we get here.
2840	*/
2841	else
2842	ASMAtomicWriteU32(&pThis->fSubscribedEvts, 0);
2843	}
2844	return rc;
2845	}
2846
2847
2848	/**
2849	* Updates the mask of events we're subscribing to.
2850	*
2851	* @returns IPRT status code.
2852	* @param pThis The socket handle.
2853	* @param fEvents The events we want to subscribe to.
2854	*/
2855	static int rtSocketPollUpdateEvents(RTSOCKETINT *pThis, uint32_t fEvents)
2856	{
2857	if (!pThis->fPollFallback)
2858	{
2859	LONG fNetworkEvents = 0;
2860	if (fEvents & RTPOLL_EVT_READ)
2861	fNetworkEvents \|= FD_READ;
2862	if (fEvents & RTPOLL_EVT_WRITE)
2863	fNetworkEvents \|= FD_WRITE;
2864	if (fEvents & RTPOLL_EVT_ERROR)
2865	fNetworkEvents \|= FD_CLOSE;
2866	LogFlowFunc(("fNetworkEvents=%#x\n", fNetworkEvents));
2867
2868	if (g_pfnWSAEventSelect(pThis->hNative, pThis->hEvent, fNetworkEvents) == 0)
2869	{
2870	pThis->fSubscribedEvts = fEvents;
2871	return VINF_SUCCESS;
2872	}
2873
2874	int rc = rtSocketError();
2875	AssertMsgFailed(("fNetworkEvents=%#x rc=%Rrc\n", fNetworkEvents, rtSocketError()));
2876	return rc;
2877	}
2878
2879	/*
2880	* Update the events we're waiting for. Caller will poke/start the thread. later
2881	*/
2882	ASMAtomicWriteU32(&pThis->fSubscribedEvts, fEvents);
2883	return VINF_SUCCESS;
2884	}
2885
2886
2887	DECLHIDDEN(int) rtSocketSetKeepAlive(RTSOCKET hSocket, bool fEnable, uint32_t cSecsIdle, uint32_t cSecsInterval)
2888	{
2889	RTSOCKETINT *pThis = hSocket;
2890	AssertPtrReturn(pThis, UINT32_MAX);
2891	AssertReturn(pThis->u32Magic == RTSOCKET_MAGIC, UINT32_MAX);
2892
2893	if (!g_pfnWSAIoctl)
2894	return VERR_NET_NOT_UNSUPPORTED;
2895
2896	struct tcp_keepalive keepAliveOptions;
2897	DWORD dwBytesReturned;
2898
2899	keepAliveOptions.onoff = fEnable ? 1 : 0;
2900	keepAliveOptions.keepalivetime = cSecsIdle * 1000;
2901	keepAliveOptions.keepaliveinterval = cSecsInterval * 1000;
2902
2903	if (g_pfnWSAIoctl(pThis->hNative,
2904	SIO_KEEPALIVE_VALS,
2905	&keepAliveOptions, sizeof(keepAliveOptions),
2906	NULL, 0, &dwBytesReturned, NULL, NULL) == 0)
2907	return VINF_SUCCESS;
2908
2909	int rc = rtSocketError();
2910	AssertMsgFailed(("WSAIoctl(.., SIO_KEEPALIVE_VALS, ...) failed: rc=%Rrc\n", rc));
2911	return rc;
2912	}
2913
2914	#endif /* RT_OS_WINDOWS */
2915
2916
2917	#if defined(RT_OS_WINDOWS) \|\| defined(RT_OS_OS2)
2918
2919	/**
2920	* Checks for pending events.
2921	*
2922	* @returns Event mask or 0.
2923	* @param pThis The socket handle.
2924	* @param fEvents The desired events.
2925	*/
2926	static uint32_t rtSocketPollCheck(RTSOCKETINT *pThis, uint32_t fEvents)
2927	{
2928	uint32_t fRetEvents = 0;
2929
2930	LogFlowFunc(("pThis=%#p fEvents=%#x\n", pThis, fEvents));
2931
2932	# ifdef RT_OS_WINDOWS
2933	/* Make sure WSAEnumNetworkEvents returns what we want. */
2934	int rc = VINF_SUCCESS;
2935	if ((pThis->fSubscribedEvts & fEvents) != fEvents)
2936	rc = rtSocketPollUpdateEvents(pThis, pThis->fSubscribedEvts \| fEvents);
2937
2938	if (!pThis->fPollFallback)
2939	{
2940	/* Atomically get pending events and reset the event semaphore. */
2941	Assert(g_pfnWSAEnumNetworkEvents);
2942	WSANETWORKEVENTS NetEvts;
2943	RT_ZERO(NetEvts);
2944	if (g_pfnWSAEnumNetworkEvents(pThis->hNative, pThis->hEvent, &NetEvts) == 0)
2945	{
2946	if ( (NetEvts.lNetworkEvents & FD_READ)
2947	&& NetEvts.iErrorCode[FD_READ_BIT] == 0)
2948	fRetEvents \|= RTPOLL_EVT_READ;
2949
2950	if ( (NetEvts.lNetworkEvents & FD_WRITE)
2951	&& NetEvts.iErrorCode[FD_WRITE_BIT] == 0)
2952	fRetEvents \|= RTPOLL_EVT_WRITE;
2953
2954	if (NetEvts.lNetworkEvents & FD_CLOSE)
2955	fRetEvents \|= RTPOLL_EVT_ERROR;
2956	else
2957	for (uint32_t i = 0; i < FD_MAX_EVENTS; i++)
2958	if ( (NetEvts.lNetworkEvents & (1L << i))
2959	&& NetEvts.iErrorCode[i] != 0)
2960	fRetEvents \|= RTPOLL_EVT_ERROR;
2961
2962	pThis->fEventsSaved = fRetEvents \|= pThis->fEventsSaved;
2963	fRetEvents &= fEvents \| RTPOLL_EVT_ERROR;
2964	}
2965	else
2966	rc = rtSocketError();
2967	}
2968
2969	/* Fall back on select if we hit an error above or is using fallback polling. */
2970	if (pThis->fPollFallback \|\| RT_FAILURE(rc))
2971	{
2972	rc = rtSocketSelectOneEx(pThis, fEvents & RTPOLL_EVT_ERROR ? fEvents \| RTPOLL_EVT_READ : fEvents, &fRetEvents, 0);
2973	if (RT_SUCCESS(rc))
2974	{
2975	/* rtSocketSelectOneEx may return RTPOLL_EVT_READ on disconnect. Use
2976	getpeername to fix this. */
2977	if ((fRetEvents & (RTPOLL_EVT_READ \| RTPOLL_EVT_ERROR)) == RTPOLL_EVT_READ)
2978	{
2979	# if 0 /* doens't work */
2980	rtSocketErrorReset();
2981	char chIgn;
2982	rc = g_pfnrecv(pThis->hNative, &chIgn, 0, MSG_NOSIGNAL);
2983	rc = rtSocketError();
2984	if (RT_FAILURE(rc))
2985	fRetEvents \|= RTPOLL_EVT_ERROR;
2986
2987	rc = g_pfnsend(pThis->hNative, &chIgn, 0, MSG_NOSIGNAL);
2988	rc = rtSocketError();
2989	if (RT_FAILURE(rc))
2990	fRetEvents \|= RTPOLL_EVT_ERROR;
2991
2992	RTSOCKADDRUNION u;
2993	int cbAddr = sizeof(u);
2994	if (g_pfngetpeername(pThis->hNative, &u.Addr, &cbAddr) == SOCKET_ERROR)
2995	fRetEvents \|= RTPOLL_EVT_ERROR;
2996	# endif
2997	/* If no bytes are available, assume error condition. */
2998	u_long cbAvail = 0;
2999	rc = g_pfnioctlsocket(pThis->hNative, FIONREAD, &cbAvail);
3000	if (rc == 0 && cbAvail == 0)
3001	fRetEvents \|= RTPOLL_EVT_ERROR;
3002	}
3003	fRetEvents &= fEvents \| RTPOLL_EVT_ERROR;
3004	}
3005	else if (rc == VERR_TIMEOUT)
3006	fRetEvents = 0;
3007	else
3008	fRetEvents \|= RTPOLL_EVT_ERROR;
3009	}
3010
3011	# else /* RT_OS_OS2 */
3012	int aFds[4] = { pThis->hNative, pThis->hNative, pThis->hNative, -1 };
3013	int rc = os2_select(aFds, 1, 1, 1, 0);
3014	if (rc > 0)
3015	{
3016	if (aFds[0] == pThis->hNative)
3017	fRetEvents \|= RTPOLL_EVT_READ;
3018	if (aFds[1] == pThis->hNative)
3019	fRetEvents \|= RTPOLL_EVT_WRITE;
3020	if (aFds[2] == pThis->hNative)
3021	fRetEvents \|= RTPOLL_EVT_ERROR;
3022	fRetEvents &= fEvents;
3023	}
3024	# endif /* RT_OS_OS2 */
3025
3026	LogFlowFunc(("fRetEvents=%#x\n", fRetEvents));
3027	return fRetEvents;
3028	}
3029
3030
3031	/**
3032	* Internal RTPoll helper that polls the socket handle and, if @a fNoWait is
3033	* clear, starts whatever actions we've got running during the poll call.
3034	*
3035	* @returns 0 if no pending events, actions initiated if @a fNoWait is clear.
3036	* Event mask (in @a fEvents) and no actions if the handle is ready
3037	* already.
3038	* UINT32_MAX (asserted) if the socket handle is busy in I/O or a
3039	* different poll set.
3040	*
3041	* @param hSocket The socket handle.
3042	* @param hPollSet The poll set handle (for access checks).
3043	* @param fEvents The events we're polling for.
3044	* @param fFinalEntry Set if this is the final entry for this handle
3045	* in this poll set. This can be used for dealing
3046	* with duplicate entries.
3047	* @param fNoWait Set if it's a zero-wait poll call. Clear if
3048	* we'll wait for an event to occur.
3049	*
3050	* @remarks There is a potential race wrt duplicate handles when @a fNoWait is
3051	* @c true, we don't currently care about that oddity...
3052	*/
3053	DECLHIDDEN(uint32_t) rtSocketPollStart(RTSOCKET hSocket, RTPOLLSET hPollSet, uint32_t fEvents, bool fFinalEntry, bool fNoWait)
3054	{
3055	RTSOCKETINT *pThis = hSocket;
3056	AssertPtrReturn(pThis, UINT32_MAX);
3057	AssertReturn(pThis->u32Magic == RTSOCKET_MAGIC, UINT32_MAX);
3058	/** @todo This isn't quite sane. Replace by critsect and open up concurrent
3059	* reads and writes! */
3060	if (rtSocketTryLock(pThis))
3061	pThis->hPollSet = hPollSet;
3062	else
3063	{
3064	AssertReturn(pThis->hPollSet == hPollSet, UINT32_MAX);
3065	ASMAtomicIncU32(&pThis->cUsers);
3066	}
3067
3068	/* (rtSocketPollCheck will reset the event object). */
3069	# ifdef RT_OS_WINDOWS
3070	uint32_t fRetEvents = pThis->fEventsSaved;
3071	pThis->fEventsSaved = 0; /* Reset */
3072	fRetEvents \|= rtSocketPollCheck(pThis, fEvents);
3073
3074	if ( !fRetEvents
3075	&& !fNoWait)
3076	{
3077	pThis->fPollEvts \|= fEvents;
3078	if (fFinalEntry)
3079	{
3080	if (pThis->fSubscribedEvts != pThis->fPollEvts)
3081	{
3082	/** @todo seems like there might be a call to many here and that fPollEvts is
3083	* totally unnecessary... (bird) */
3084	int rc = rtSocketPollUpdateEvents(pThis, pThis->fPollEvts);
3085	if (RT_FAILURE(rc))
3086	{
3087	pThis->fPollEvts = 0;
3088	fRetEvents = UINT32_MAX;
3089	}
3090	}
3091
3092	/* Make sure we don't block when there are events pending relevant to an earlier poll set entry. */
3093	if (pThis->fEventsSaved && !pThis->fPollFallback && g_pfnWSASetEvent && fRetEvents == 0)
3094	g_pfnWSASetEvent(pThis->hEvent);
3095	}
3096	}
3097	# else
3098	uint32_t fRetEvents = rtSocketPollCheck(pThis, fEvents);
3099	# endif
3100
3101	if (fRetEvents \|\| fNoWait)
3102	{
3103	if (pThis->cUsers == 1)
3104	{
3105	# ifdef RT_OS_WINDOWS
3106	pThis->fEventsSaved &= RTPOLL_EVT_ERROR;
3107	pThis->fHarvestedEvents = false;
3108	rtSocketPollClearEventAndRestoreBlocking(pThis);
3109	# endif
3110	pThis->hPollSet = NIL_RTPOLLSET;
3111	}
3112	# ifdef RT_OS_WINDOWS
3113	else
3114	pThis->fHarvestedEvents = true;
3115	# endif
3116	ASMAtomicDecU32(&pThis->cUsers);
3117	}
3118	# ifdef RT_OS_WINDOWS
3119	/*
3120	* Kick the poller thread on if this is the final entry and we're in
3121	* winsock 1.x fallback mode.
3122	*/
3123	else if (pThis->fPollFallback && fFinalEntry)
3124	fRetEvents = rtSocketPollFallbackStart(pThis);
3125	# endif
3126
3127	return fRetEvents;
3128	}
3129
3130
3131	/**
3132	* Called after a WaitForMultipleObjects returned in order to check for pending
3133	* events and stop whatever actions that rtSocketPollStart() initiated.
3134	*
3135	* @returns Event mask or 0.
3136	*
3137	* @param hSocket The socket handle.
3138	* @param fEvents The events we're polling for.
3139	* @param fFinalEntry Set if this is the final entry for this handle
3140	* in this poll set. This can be used for dealing
3141	* with duplicate entries. Only keep in mind that
3142	* this method is called in reverse order, so the
3143	* first call will have this set (when the entire
3144	* set was processed).
3145	* @param fHarvestEvents Set if we should check for pending events.
3146	*/
3147	DECLHIDDEN(uint32_t) rtSocketPollDone(RTSOCKET hSocket, uint32_t fEvents, bool fFinalEntry, bool fHarvestEvents)
3148	{
3149	RTSOCKETINT *pThis = hSocket;
3150	AssertPtrReturn(pThis, 0);
3151	AssertReturn(pThis->u32Magic == RTSOCKET_MAGIC, 0);
3152	Assert(pThis->cUsers > 0);
3153	Assert(pThis->hPollSet != NIL_RTPOLLSET);
3154	RT_NOREF_PV(fFinalEntry);
3155
3156	# ifdef RT_OS_WINDOWS
3157	/*
3158	* Deactivate the poll thread if we're in winsock 1.x fallback poll mode.
3159	*/
3160	if ( pThis->fPollFallback
3161	&& pThis->hPollFallbackThread != NIL_RTTHREAD)
3162	{
3163	ASMAtomicWriteU32(&pThis->fSubscribedEvts, 0);
3164	if (ASMAtomicXchgBool(&pThis->fPollFallbackActive, false))
3165	rtSocketPokePollFallbackThread(pThis);
3166	}
3167	# endif
3168
3169	/*
3170	* Harvest events and clear the event mask for the next round of polling.
3171	*/
3172	uint32_t fRetEvents;
3173	# ifdef RT_OS_WINDOWS
3174	if (!pThis->fPollFallback)
3175	{
3176	if (!pThis->fHarvestedEvents)
3177	{
3178	fRetEvents = rtSocketPollCheck(pThis, fEvents);
3179	pThis->fHarvestedEvents = true;
3180	}
3181	else
3182	fRetEvents = pThis->fEventsSaved;
3183	if (fHarvestEvents)
3184	fRetEvents &= fEvents;
3185	else
3186	fRetEvents = 0;
3187	pThis->fPollEvts = 0;
3188	}
3189	else
3190	# endif
3191	{
3192	if (fHarvestEvents)
3193	fRetEvents = rtSocketPollCheck(pThis, fEvents);
3194	else
3195	fRetEvents = 0;
3196	}
3197
3198	/*
3199	* Make the socket blocking again and unlock the handle.
3200	*/
3201	if (pThis->cUsers == 1)
3202	{
3203	# ifdef RT_OS_WINDOWS
3204	pThis->fEventsSaved &= RTPOLL_EVT_ERROR;
3205	pThis->fHarvestedEvents = false;
3206	rtSocketPollClearEventAndRestoreBlocking(pThis);
3207	# endif
3208	pThis->hPollSet = NIL_RTPOLLSET;
3209	}
3210	ASMAtomicDecU32(&pThis->cUsers);
3211	return fRetEvents;
3212	}
3213
3214	#endif /* RT_OS_WINDOWS \|\| RT_OS_OS2 */
3215

Note: See TracBrowser for help on using the repository browser.

Context Navigation

source: vbox/trunk/src/VBox/Runtime/r3/socket.cpp

Download in other formats: