socket.cpp@ 98032

Last change on this file since 98032 was 97894, checked in by vboxsync, 2 years ago
IPRT/socket,localipc: Map AF_LOCAL to AF_UNIX if the former isn't defined (S10).
Property svn:eol-style set to `native` Property svn:keywords set to `Id Revision`
File size: 96.3 KB

Line
1	/* $Id: socket.cpp 97894 2022-12-29 15:07:50Z vboxsync $ */
2	/** @file
3	* IPRT - Network Sockets.
4	*/
5
6	/*
7	* Copyright (C) 2006-2022 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	#else /* !RT_OS_WINDOWS */
45	# include <errno.h>
46	# include <sys/select.h>
47	# include <sys/stat.h>
48	# include <sys/socket.h>
49	# include <netinet/in.h>
50	# include <netinet/tcp.h>
51	# include <arpa/inet.h>
52	# ifdef IPRT_WITH_TCPIP_V6
53	# include <netinet6/in6.h>
54	# endif
55	# include <sys/un.h>
56	# include <netdb.h>
57	# include <unistd.h>
58	# include <fcntl.h>
59	# include <sys/uio.h>
60	# ifndef AF_LOCAL
61	# define AF_LOCAL AF_UNIX
62	# endif
63	#endif /* !RT_OS_WINDOWS */
64	#include <limits.h>
65
66	#include "internal/iprt.h"
67	#include <iprt/socket.h>
68
69	#include <iprt/alloca.h>
70	#include <iprt/asm.h>
71	#include <iprt/assert.h>
72	#include <iprt/ctype.h>
73	#include <iprt/err.h>
74	#include <iprt/mempool.h>
75	#include <iprt/poll.h>
76	#include <iprt/string.h>
77	#include <iprt/thread.h>
78	#include <iprt/time.h>
79	#include <iprt/mem.h>
80	#include <iprt/sg.h>
81	#include <iprt/log.h>
82
83	#include "internal/magics.h"
84	#include "internal/socket.h"
85	#include "internal/string.h"
86	#ifdef RT_OS_WINDOWS
87	# include "win/internal-r3-win.h"
88	#endif
89
90
91	/*********************************************************************************************************************************
92	* Defined Constants And Macros *
93	*********************************************************************************************************************************/
94	/* non-standard linux stuff (it seems). */
95	#ifndef MSG_NOSIGNAL
96	# define MSG_NOSIGNAL 0
97	#endif
98
99	/* Windows has different names for SHUT_XXX. */
100	#ifndef SHUT_RDWR
101	# ifdef SD_BOTH
102	# define SHUT_RDWR SD_BOTH
103	# else
104	# define SHUT_RDWR 2
105	# endif
106	#endif
107	#ifndef SHUT_WR
108	# ifdef SD_SEND
109	# define SHUT_WR SD_SEND
110	# else
111	# define SHUT_WR 1
112	# endif
113	#endif
114	#ifndef SHUT_RD
115	# ifdef SD_RECEIVE
116	# define SHUT_RD SD_RECEIVE
117	# else
118	# define SHUT_RD 0
119	# endif
120	#endif
121
122	/* fixup backlevel OSes. */
123	#if defined(RT_OS_OS2) \|\| defined(RT_OS_WINDOWS)
124	# define socklen_t int
125	#endif
126
127	/** How many pending connection. */
128	#define RTTCP_SERVER_BACKLOG 10
129
130	/* Limit read and write sizes on Windows and OS/2. */
131	#ifdef RT_OS_WINDOWS
132	# define RTSOCKET_MAX_WRITE (INT_MAX / 2)
133	# define RTSOCKET_MAX_READ (INT_MAX / 2)
134	#elif defined(RT_OS_OS2)
135	# define RTSOCKET_MAX_WRITE 0x10000
136	# define RTSOCKET_MAX_READ 0x10000
137	#endif
138
139
140	/*********************************************************************************************************************************
141	* Structures and Typedefs *
142	*********************************************************************************************************************************/
143	/**
144	* Socket handle data.
145	*
146	* This is mainly required for implementing RTPollSet on Windows.
147	*/
148	typedef struct RTSOCKETINT
149	{
150	/** Magic number (RTSOCKET_MAGIC). */
151	uint32_t u32Magic;
152	/** Exclusive user count.
153	* This is used to prevent two threads from accessing the handle concurrently.
154	* It can be higher than 1 if this handle is reference multiple times in a
155	* polling set (Windows). */
156	uint32_t volatile cUsers;
157	/** The native socket handle. */
158	RTSOCKETNATIVE hNative;
159	/** Indicates whether the handle has been closed or not. */
160	bool volatile fClosed;
161	/** Indicates whether the socket is operating in blocking or non-blocking mode
162	* currently. */
163	bool fBlocking;
164	/** Whether to leave the native socket open rather than closing it (for
165	* RTHandleGetStandard). */
166	bool fLeaveOpen;
167	#if defined(RT_OS_WINDOWS) \|\| defined(RT_OS_OS2)
168	/** The pollset currently polling this socket. This is NIL if no one is
169	* polling. */
170	RTPOLLSET hPollSet;
171	#endif
172	#ifdef RT_OS_WINDOWS
173	/** The event semaphore we've associated with the socket handle.
174	* This is WSA_INVALID_EVENT if not done. */
175	WSAEVENT hEvent;
176	/** The events we're polling for. */
177	uint32_t fPollEvts;
178	/** The events we're currently subscribing to with WSAEventSelect.
179	* This is ZERO if we're currently not subscribing to anything. */
180	uint32_t fSubscribedEvts;
181	/** Saved events which are only posted once and events harvested for
182	* sockets entered multiple times into to a poll set. Imagine a scenario where
183	* you have a RTPOLL_EVT_READ entry and RTPOLL_EVT_ERROR entry. The READ
184	* condition can be triggered between checking the READ entry and the ERROR
185	* entry, and we don't want to drop the READ, so we store it here and make sure
186	* the event is signalled.
187	*
188	* The RTPOLL_EVT_ERROR is inconsistenly sticky at the momemnt... */
189	uint32_t fEventsSaved;
190	/** Set if fEventsSaved contains harvested events (used to avoid multiple
191	* calls to rtSocketPollCheck on the same socket during rtSocketPollDone). */
192	bool fHarvestedEvents;
193	/** Set if we're using the polling fallback. */
194	bool fPollFallback;
195	/** Set if the fallback polling is active (event not set). */
196	bool volatile fPollFallbackActive;
197	/** Set to shut down the fallback polling thread. */
198	bool volatile fPollFallbackShutdown;
199	/** Socket use to wake up the select thread. */
200	RTSOCKETNATIVE hPollFallbackNotifyW;
201	/** Socket the select thread always waits on. */
202	RTSOCKETNATIVE hPollFallbackNotifyR;
203	/** The fallback polling thread. */
204	RTTHREAD hPollFallbackThread;
205	#endif /* RT_OS_WINDOWS */
206	} RTSOCKETINT;
207
208
209	/**
210	* Address union used internally for things like getpeername and getsockname.
211	*/
212	typedef union RTSOCKADDRUNION
213	{
214	struct sockaddr Addr;
215	struct sockaddr_in IPv4;
216	#ifdef IPRT_WITH_TCPIP_V6
217	struct sockaddr_in6 IPv6;
218	#endif
219	} RTSOCKADDRUNION;
220
221
222	/*********************************************************************************************************************************
223	* Global Variables *
224	*********************************************************************************************************************************/
225	#ifdef RT_OS_WINDOWS
226	/** Indicates that we've successfully initialized winsock. */
227	static uint32_t volatile g_uWinSockInitedVersion = 0;
228	#endif
229
230
231	/*********************************************************************************************************************************
232	* Internal Functions *
233	*********************************************************************************************************************************/
234	#ifdef RT_OS_WINDOWS
235	static void rtSocketPokePollFallbackThread(RTSOCKETINT *pThis);
236	#endif
237
238
239
240	#ifdef RT_OS_WINDOWS
241	/**
242	* Initializes winsock for the process.
243	*
244	* @returns IPRT status code.
245	*/
246	static int rtSocketInitWinsock(void)
247	{
248	if (g_uWinSockInitedVersion != 0)
249	return VINF_SUCCESS;
250
251	if ( !g_pfnWSAGetLastError
252	\|\| !g_pfnWSAStartup
253	\|\| !g_pfnsocket
254	\|\| !g_pfnclosesocket)
255	return VERR_NET_INIT_FAILED;
256
257	/*
258	* Initialize winsock. Try with 2.2 and back down till we get something that works.
259	*/
260	static const WORD s_awVersions[] =
261	{
262	MAKEWORD(2, 2),
263	MAKEWORD(2, 1),
264	MAKEWORD(2, 0),
265	MAKEWORD(1, 1),
266	MAKEWORD(1, 0),
267	};
268	for (uint32_t i = 0; i < RT_ELEMENTS(s_awVersions); i++)
269	{
270	WSADATA wsaData;
271	RT_ZERO(wsaData);
272	int rcWsa = g_pfnWSAStartup(s_awVersions[i], &wsaData);
273	if (rcWsa == 0)
274	{
275	/* AssertMsg(wsaData.wVersion >= s_awVersions[i]); - triggers with winsock 1.1 */
276	ASMAtomicWriteU32(&g_uWinSockInitedVersion, wsaData.wVersion);
277	return VINF_SUCCESS;
278	}
279	AssertLogRelMsg(rcWsa == WSAVERNOTSUPPORTED, ("rcWsa=%d (winsock version %#x)\n", rcWsa, s_awVersions[i]));
280	}
281	LogRel(("Failed to init winsock!\n"));
282	return VERR_NET_INIT_FAILED;
283	}
284	#endif
285
286
287	/**
288	* Get the last error as an iprt status code.
289	*
290	* @returns IPRT status code.
291	*/
292	DECLINLINE(int) rtSocketError(void)
293	{
294	#ifdef RT_OS_WINDOWS
295	if (g_pfnWSAGetLastError)
296	return RTErrConvertFromWin32(g_pfnWSAGetLastError());
297	return VERR_NET_IO_ERROR;
298	#else
299	return RTErrConvertFromErrno(errno);
300	#endif
301	}
302
303
304	/**
305	* Resets the last error.
306	*/
307	DECLINLINE(void) rtSocketErrorReset(void)
308	{
309	#ifdef RT_OS_WINDOWS
310	if (g_pfnWSASetLastError)
311	g_pfnWSASetLastError(0);
312	#else
313	errno = 0;
314	#endif
315	}
316
317
318	/**
319	* Get the last resolver error as an iprt status code.
320	*
321	* @returns iprt status code.
322	*/
323	DECLHIDDEN(int) rtSocketResolverError(void)
324	{
325	#ifdef RT_OS_WINDOWS
326	if (g_pfnWSAGetLastError)
327	return RTErrConvertFromWin32(g_pfnWSAGetLastError());
328	return VERR_UNRESOLVED_ERROR;
329	#else
330	switch (h_errno)
331	{
332	case HOST_NOT_FOUND:
333	return VERR_NET_HOST_NOT_FOUND;
334	case NO_DATA:
335	return VERR_NET_ADDRESS_NOT_AVAILABLE;
336	case NO_RECOVERY:
337	return VERR_IO_GEN_FAILURE;
338	case TRY_AGAIN:
339	return VERR_TRY_AGAIN;
340
341	default:
342	AssertLogRelMsgFailed(("Unhandled error %u\n", h_errno));
343	return VERR_UNRESOLVED_ERROR;
344	}
345	#endif
346	}
347
348
349	/**
350	* Converts from a native socket address to a generic IPRT network address.
351	*
352	* @returns IPRT status code.
353	* @param pSrc The source address.
354	* @param cbSrc The size of the source address.
355	* @param pAddr Where to return the generic IPRT network
356	* address.
357	*/
358	static int rtSocketNetAddrFromAddr(RTSOCKADDRUNION const *pSrc, size_t cbSrc, PRTNETADDR pAddr)
359	{
360	/*
361	* Convert the address.
362	*/
363	if ( cbSrc == sizeof(struct sockaddr_in)
364	&& pSrc->Addr.sa_family == AF_INET)
365	{
366	RT_ZERO(*pAddr);
367	pAddr->enmType = RTNETADDRTYPE_IPV4;
368	pAddr->uPort = RT_N2H_U16(pSrc->IPv4.sin_port);
369	pAddr->uAddr.IPv4.u = pSrc->IPv4.sin_addr.s_addr;
370	}
371	#ifdef IPRT_WITH_TCPIP_V6
372	else if ( cbSrc == sizeof(struct sockaddr_in6)
373	&& pSrc->Addr.sa_family == AF_INET6)
374	{
375	RT_ZERO(*pAddr);
376	pAddr->enmType = RTNETADDRTYPE_IPV6;
377	pAddr->uPort = RT_N2H_U16(pSrc->IPv6.sin6_port);
378	pAddr->uAddr.IPv6.au32[0] = pSrc->IPv6.sin6_addr.s6_addr32[0];
379	pAddr->uAddr.IPv6.au32[1] = pSrc->IPv6.sin6_addr.s6_addr32[1];
380	pAddr->uAddr.IPv6.au32[2] = pSrc->IPv6.sin6_addr.s6_addr32[2];
381	pAddr->uAddr.IPv6.au32[3] = pSrc->IPv6.sin6_addr.s6_addr32[3];
382	}
383	#endif
384	else
385	return VERR_NET_ADDRESS_FAMILY_NOT_SUPPORTED;
386	return VINF_SUCCESS;
387	}
388
389
390	/**
391	* Converts from a generic IPRT network address to a native socket address.
392	*
393	* @returns IPRT status code.
394	* @param pAddr Pointer to the generic IPRT network address.
395	* @param pDst The source address.
396	* @param cbDst The size of the source address.
397	* @param pcbAddr Where to store the size of the returned address.
398	* Optional
399	*/
400	static int rtSocketAddrFromNetAddr(PCRTNETADDR pAddr, RTSOCKADDRUNION pDst, size_t cbDst, int pcbAddr)
401	{
402	RT_BZERO(pDst, cbDst);
403	if (pAddr->enmType == RTNETADDRTYPE_IPV4)
404	{
405	if (cbDst < sizeof(struct sockaddr_in))
406	return VERR_BUFFER_OVERFLOW;
407
408	pDst->Addr.sa_family = AF_INET;
409	pDst->IPv4.sin_port = RT_H2N_U16(pAddr->uPort);
410	pDst->IPv4.sin_addr.s_addr = pAddr->uAddr.IPv4.u;
411	if (pcbAddr)
412	*pcbAddr = sizeof(pDst->IPv4);
413	}
414	#ifdef IPRT_WITH_TCPIP_V6
415	else if (pAddr->enmType == RTNETADDRTYPE_IPV6)
416	{
417	if (cbDst < sizeof(struct sockaddr_in6))
418	return VERR_BUFFER_OVERFLOW;
419
420	pDst->Addr.sa_family = AF_INET6;
421	pDst->IPv6.sin6_port = RT_H2N_U16(pAddr->uPort);
422	pSrc->IPv6.sin6_addr.s6_addr32[0] = pAddr->uAddr.IPv6.au32[0];
423	pSrc->IPv6.sin6_addr.s6_addr32[1] = pAddr->uAddr.IPv6.au32[1];
424	pSrc->IPv6.sin6_addr.s6_addr32[2] = pAddr->uAddr.IPv6.au32[2];
425	pSrc->IPv6.sin6_addr.s6_addr32[3] = pAddr->uAddr.IPv6.au32[3];
426	if (pcbAddr)
427	*pcbAddr = sizeof(pDst->IPv6);
428	}
429	#endif
430	else
431	return VERR_NET_ADDRESS_FAMILY_NOT_SUPPORTED;
432	return VINF_SUCCESS;
433	}
434
435
436	/**
437	* Tries to lock the socket for exclusive usage by the calling thread.
438	*
439	* Call rtSocketUnlock() to unlock.
440	*
441	* @returns @c true if locked, @c false if not.
442	* @param pThis The socket structure.
443	*/
444	DECLINLINE(bool) rtSocketTryLock(RTSOCKETINT *pThis)
445	{
446	return ASMAtomicCmpXchgU32(&pThis->cUsers, 1, 0);
447	}
448
449
450	/**
451	* Unlocks the socket.
452	*
453	* @param pThis The socket structure.
454	*/
455	DECLINLINE(void) rtSocketUnlock(RTSOCKETINT *pThis)
456	{
457	ASMAtomicCmpXchgU32(&pThis->cUsers, 0, 1);
458	}
459
460
461	/**
462	* The slow path of rtSocketSwitchBlockingMode that does the actual switching.
463	*
464	* @returns IPRT status code.
465	* @param pThis The socket structure.
466	* @param fBlocking The desired mode of operation.
467	* @remarks Do not call directly.
468	*/
469	static int rtSocketSwitchBlockingModeSlow(RTSOCKETINT *pThis, bool fBlocking)
470	{
471	#ifdef RT_OS_WINDOWS
472	AssertReturn(g_pfnioctlsocket, VERR_NET_NOT_UNSUPPORTED);
473	u_long uBlocking = fBlocking ? 0 : 1;
474	if (g_pfnioctlsocket(pThis->hNative, FIONBIO, &uBlocking))
475	return rtSocketError();
476
477	#else
478	int fFlags = fcntl(pThis->hNative, F_GETFL, 0);
479	if (fFlags == -1)
480	return rtSocketError();
481
482	if (fBlocking)
483	fFlags &= ~O_NONBLOCK;
484	else
485	fFlags \|= O_NONBLOCK;
486	if (fcntl(pThis->hNative, F_SETFL, fFlags) == -1)
487	return rtSocketError();
488	#endif
489
490	pThis->fBlocking = fBlocking;
491	return VINF_SUCCESS;
492	}
493
494
495	/**
496	* Switches the socket to the desired blocking mode if necessary.
497	*
498	* The socket must be locked.
499	*
500	* @returns IPRT status code.
501	* @param pThis The socket structure.
502	* @param fBlocking The desired mode of operation.
503	*/
504	DECLINLINE(int) rtSocketSwitchBlockingMode(RTSOCKETINT *pThis, bool fBlocking)
505	{
506	if (pThis->fBlocking != fBlocking)
507	return rtSocketSwitchBlockingModeSlow(pThis, fBlocking);
508	return VINF_SUCCESS;
509	}
510
511
512	/**
513	* Creates an IPRT socket handle for a native one.
514	*
515	* @returns IPRT status code.
516	* @param ppSocket Where to return the IPRT socket handle.
517	* @param hNative The native handle.
518	* @param fLeaveOpen Whether to leave the native socket handle open when
519	* closed.
520	*/
521	DECLHIDDEN(int) rtSocketCreateForNative(RTSOCKETINT **ppSocket, RTSOCKETNATIVE hNative, bool fLeaveOpen)
522	{
523	RTSOCKETINT pThis = (RTSOCKETINT )RTMemPoolAlloc(RTMEMPOOL_DEFAULT, sizeof(*pThis));
524	if (!pThis)
525	return VERR_NO_MEMORY;
526	pThis->u32Magic = RTSOCKET_MAGIC;
527	pThis->cUsers = 0;
528	pThis->hNative = hNative;
529	pThis->fClosed = false;
530	pThis->fLeaveOpen = fLeaveOpen;
531	pThis->fBlocking = true;
532	#if defined(RT_OS_WINDOWS) \|\| defined(RT_OS_OS2)
533	pThis->hPollSet = NIL_RTPOLLSET;
534	#endif
535	#ifdef RT_OS_WINDOWS
536	pThis->hEvent = WSA_INVALID_EVENT;
537	pThis->fPollEvts = 0;
538	pThis->fSubscribedEvts = 0;
539	pThis->fEventsSaved = 0;
540	pThis->fHarvestedEvents = false;
541	pThis->fPollFallback = g_uWinSockInitedVersion < MAKEWORD(2, 0)
542	\|\| g_pfnWSACreateEvent == NULL
543	\|\| g_pfnWSACloseEvent == NULL
544	\|\| g_pfnWSAEventSelect == NULL
545	\|\| g_pfnWSAEnumNetworkEvents == NULL;
546	pThis->fPollFallbackActive = false;
547	pThis->fPollFallbackShutdown = false;
548	pThis->hPollFallbackNotifyR = NIL_RTSOCKETNATIVE;
549	pThis->hPollFallbackNotifyW = NIL_RTSOCKETNATIVE;
550	pThis->hPollFallbackThread = NIL_RTTHREAD;
551	#endif
552	*ppSocket = pThis;
553	return VINF_SUCCESS;
554	}
555
556
557	RTDECL(int) RTSocketFromNative(PRTSOCKET phSocket, RTHCINTPTR uNative)
558	{
559	AssertReturn(uNative != NIL_RTSOCKETNATIVE, VERR_INVALID_PARAMETER);
560	#ifndef RT_OS_WINDOWS
561	AssertReturn(uNative >= 0, VERR_INVALID_PARAMETER);
562	#endif
563	AssertPtrReturn(phSocket, VERR_INVALID_POINTER);
564	return rtSocketCreateForNative(phSocket, uNative, false /fLeaveOpen/);
565	}
566
567
568	/**
569	* Wrapper around socket().
570	*
571	* @returns IPRT status code.
572	* @param phSocket Where to store the handle to the socket on
573	* success.
574	* @param iDomain The protocol family (PF_XXX).
575	* @param iType The socket type (SOCK_XXX).
576	* @param iProtocol Socket parameter, usually 0.
577	* @param fInheritable Set to true if the socket should be inherted by
578	* child processes, false if not inheritable.
579	*/
580	DECLHIDDEN(int) rtSocketCreate(PRTSOCKET phSocket, int iDomain, int iType, int iProtocol, bool fInheritable)
581	{
582	#ifdef RT_OS_WINDOWS
583	AssertReturn(g_pfnsocket, VERR_NET_NOT_UNSUPPORTED);
584	AssertReturn(g_pfnclosesocket, VERR_NET_NOT_UNSUPPORTED);
585
586	/* Initialize WinSock. */
587	int rc2 = rtSocketInitWinsock();
588	if (RT_FAILURE(rc2))
589	return rc2;
590	#endif
591
592	/*
593	* Create the socket.
594	*
595	* The RTSocketSetInheritance operation isn't necessarily reliable on windows,
596	* so try use WSA_FLAG_NO_HANDLE_INHERIT with WSASocketW when possible.
597	*/
598	#ifdef RT_OS_WINDOWS
599	bool fCallSetInheritance = true;
600	RTSOCKETNATIVE hNative;
601	if (g_pfnWSASocketW)
602	{
603	DWORD fWsaFlags = WSA_FLAG_OVERLAPPED \| (!fInheritable ? WSA_FLAG_NO_HANDLE_INHERIT : 0);
604	hNative = g_pfnWSASocketW(iDomain, iType, iProtocol, NULL, 0 /Group/, fWsaFlags);
605	if (hNative != NIL_RTSOCKETNATIVE)
606	fCallSetInheritance = false;
607	else
608	{
609	if (!fInheritable)
610	hNative = g_pfnsocket(iDomain, iType, iProtocol);
611	if (hNative == NIL_RTSOCKETNATIVE)
612	return rtSocketError();
613	}
614	}
615	else
616	{
617	hNative = g_pfnsocket(iDomain, iType, iProtocol);
618	if (hNative == NIL_RTSOCKETNATIVE)
619	return rtSocketError();
620	}
621	#else
622	RTSOCKETNATIVE hNative = socket(iDomain, iType, iProtocol);
623	if (hNative == NIL_RTSOCKETNATIVE)
624	return rtSocketError();
625	#endif
626
627	/*
628	* Wrap it.
629	*/
630	int rc = rtSocketCreateForNative(phSocket, hNative, false /fLeaveOpen/);
631	if (RT_SUCCESS(rc))
632	{
633	#ifdef RT_OS_WINDOWS
634	if (fCallSetInheritance)
635	#endif
636	RTSocketSetInheritance(*phSocket, fInheritable);
637	}
638	else
639	{
640	#ifdef RT_OS_WINDOWS
641	g_pfnclosesocket(hNative);
642	#else
643	close(hNative);
644	#endif
645	}
646	return rc;
647	}
648
649
650	/**
651	* Wrapper around socketpair() for creating a local TCP connection.
652	*
653	* @returns IPRT status code.
654	* @param phServer Where to return the first native socket.
655	* @param phClient Where to return the second native socket.
656	*/
657	static int rtSocketCreateNativeTcpPair(RTSOCKETNATIVE phServer, RTSOCKETNATIVE phClient)
658	{
659	#ifdef RT_OS_WINDOWS
660	/*
661	* Initialize WinSock and make sure we got the necessary APIs.
662	*/
663	int rc = rtSocketInitWinsock();
664	if (RT_FAILURE(rc))
665	return rc;
666	AssertReturn(g_pfnsocket, VERR_NET_NOT_UNSUPPORTED);
667	AssertReturn(g_pfnclosesocket, VERR_NET_NOT_UNSUPPORTED);
668	AssertReturn(g_pfnsetsockopt, VERR_NET_NOT_UNSUPPORTED);
669	AssertReturn(g_pfnbind, VERR_NET_NOT_UNSUPPORTED);
670	AssertReturn(g_pfngetsockname, VERR_NET_NOT_UNSUPPORTED);
671	AssertReturn(g_pfnlisten, VERR_NET_NOT_UNSUPPORTED);
672	AssertReturn(g_pfnaccept, VERR_NET_NOT_UNSUPPORTED);
673	AssertReturn(g_pfnconnect, VERR_NET_NOT_UNSUPPORTED);
674
675	/*
676	* Create the "server" listen socket and the "client" socket.
677	*/
678	RTSOCKETNATIVE hListener = g_pfnsocket(AF_INET, SOCK_STREAM, IPPROTO_TCP);
679	if (hListener == NIL_RTSOCKETNATIVE)
680	return rtSocketError();
681	RTSOCKETNATIVE hClient = g_pfnsocket(AF_INET, SOCK_STREAM, IPPROTO_TCP);
682	if (hClient != NIL_RTSOCKETNATIVE)
683	{
684
685	/*
686	* We let WinSock choose a port number when we bind.
687	*/
688	union
689	{
690	struct sockaddr_in Ip;
691	struct sockaddr Generic;
692	} uAddr;
693	RT_ZERO(uAddr);
694	uAddr.Ip.sin_family = AF_INET;
695	uAddr.Ip.sin_addr.s_addr = RT_H2N_U32_C(INADDR_LOOPBACK);
696	//uAddr.Ip.sin_port = 0;
697	int fReuse = 1;
698	rc = g_pfnsetsockopt(hListener, SOL_SOCKET, SO_REUSEADDR, (const char *)&fReuse, sizeof(fReuse));
699	if (rc == 0)
700	{
701	rc = g_pfnbind(hListener, &uAddr.Generic, sizeof(uAddr.Ip));
702	if (rc == 0)
703	{
704	/*
705	* Get the address the client should connect to. According to the docs,
706	* we cannot assume that getsockname sets the IP and family.
707	*/
708	RT_ZERO(uAddr);
709	int cbAddr = sizeof(uAddr.Ip);
710	rc = g_pfngetsockname(hListener, &uAddr.Generic, &cbAddr);
711	if (rc == 0)
712	{
713	uAddr.Ip.sin_family = AF_INET;
714	uAddr.Ip.sin_addr.s_addr = RT_H2N_U32_C(INADDR_LOOPBACK);
715
716	/*
717	* Listen, connect and accept.
718	*/
719	rc = g_pfnlisten(hListener, 1 /cBacklog/);
720	if (rc == 0)
721	{
722	rc = g_pfnconnect(hClient, &uAddr.Generic, sizeof(uAddr.Ip));
723	if (rc == 0)
724	{
725	RTSOCKETNATIVE hServer = g_pfnaccept(hListener, NULL, NULL);
726	if (hServer != NIL_RTSOCKETNATIVE)
727	{
728	g_pfnclosesocket(hListener);
729
730	/*
731	* Done!
732	*/
733	*phServer = hServer;
734	*phClient = hClient;
735	return VINF_SUCCESS;
736	}
737	}
738	}
739	}
740	}
741	}
742	rc = rtSocketError();
743	g_pfnclosesocket(hClient);
744	}
745	else
746	rc = rtSocketError();
747	g_pfnclosesocket(hListener);
748	return rc;
749
750	#else
751	/*
752	* Got socket pair, so use it.
753	* Note! This isn't TCP per se, but it should fool the users.
754	*/
755	int aSockets[2] = { -1, -1 };
756	if (socketpair(AF_LOCAL, SOCK_STREAM, 0, aSockets) == 0)
757	{
758	*phServer = aSockets[0];
759	*phClient = aSockets[1];
760	return VINF_SUCCESS;
761	}
762	return rtSocketError();
763	#endif
764	}
765
766
767	/**
768	* Worker for RTTcpCreatePair.
769	*
770	* @returns IPRT status code.
771	* @param phServer Where to return the "server" side of the pair.
772	* @param phClient Where to return the "client" side of the pair.
773	* @note There is no server or client side, but we gotta call it something.
774	*/
775	DECLHIDDEN(int) rtSocketCreateTcpPair(RTSOCKET phServer, RTSOCKET phClient)
776	{
777	RTSOCKETNATIVE hServer = NIL_RTSOCKETNATIVE;
778	RTSOCKETNATIVE hClient = NIL_RTSOCKETNATIVE;
779	int rc = rtSocketCreateNativeTcpPair(&hServer, &hClient);
780	if (RT_SUCCESS(rc))
781	{
782	rc = rtSocketCreateForNative(phServer, hServer, false /fLeaveOpen/);
783	if (RT_SUCCESS(rc))
784	{
785	rc = rtSocketCreateForNative(phClient, hClient, false /fLeaveOpen/);
786	if (RT_SUCCESS(rc))
787	return VINF_SUCCESS;
788	RTSocketRelease(*phServer);
789	}
790	else
791	{
792	#ifdef RT_OS_WINDOWS
793	g_pfnclosesocket(hServer);
794	#else
795	close(hServer);
796	#endif
797	}
798	#ifdef RT_OS_WINDOWS
799	g_pfnclosesocket(hClient);
800	#else
801	close(hClient);
802	#endif
803	}
804
805	*phServer = NIL_RTSOCKET;
806	*phClient = NIL_RTSOCKET;
807	return rc;
808	}
809
810
811	RTDECL(uint32_t) RTSocketRetain(RTSOCKET hSocket)
812	{
813	RTSOCKETINT *pThis = hSocket;
814	AssertPtrReturn(pThis, UINT32_MAX);
815	AssertReturn(pThis->u32Magic == RTSOCKET_MAGIC, UINT32_MAX);
816	return RTMemPoolRetain(pThis);
817	}
818
819
820	/**
821	* Worker for RTSocketRelease and RTSocketClose.
822	*
823	* @returns IPRT status code.
824	* @param pThis The socket handle instance data.
825	* @param fDestroy Whether we're reaching ref count zero.
826	*/
827	static int rtSocketCloseIt(RTSOCKETINT *pThis, bool fDestroy)
828	{
829	/*
830	* Invalidate the handle structure on destroy.
831	*/
832	if (fDestroy)
833	{
834	Assert(ASMAtomicReadU32(&pThis->u32Magic) == RTSOCKET_MAGIC);
835	ASMAtomicWriteU32(&pThis->u32Magic, RTSOCKET_MAGIC_DEAD);
836	}
837
838	int rc = VINF_SUCCESS;
839	if (ASMAtomicCmpXchgBool(&pThis->fClosed, true, false))
840	{
841	#ifdef RT_OS_WINDOWS
842	/*
843	* Poke the polling thread if active and give it a small chance to stop.
844	*/
845	if ( pThis->fPollFallback
846	&& pThis->hPollFallbackThread != NIL_RTTHREAD)
847	{
848	ASMAtomicWriteBool(&pThis->fPollFallbackShutdown, true);
849	rtSocketPokePollFallbackThread(pThis);
850	int rc2 = RTThreadWait(pThis->hPollFallbackThread, RT_MS_1SEC, NULL);
851	if (RT_SUCCESS(rc2))
852	pThis->hPollFallbackThread = NIL_RTTHREAD;
853	}
854	#endif
855
856	/*
857	* Close the native handle.
858	*/
859	RTSOCKETNATIVE hNative = pThis->hNative;
860	if (hNative != NIL_RTSOCKETNATIVE)
861	{
862	pThis->hNative = NIL_RTSOCKETNATIVE;
863
864	if (!pThis->fLeaveOpen)
865	{
866	#ifdef RT_OS_WINDOWS
867	AssertReturn(g_pfnclosesocket, VERR_NET_NOT_UNSUPPORTED);
868	if (g_pfnclosesocket(hNative))
869	#else
870	if (close(hNative))
871	#endif
872	{
873	rc = rtSocketError();
874	#ifdef RT_OS_WINDOWS
875	AssertMsgFailed(("closesocket(%p) -> %Rrc\n", (uintptr_t)hNative, rc));
876	#else
877	AssertMsgFailed(("close(%d) -> %Rrc\n", hNative, rc));
878	#endif
879	}
880	}
881	}
882
883	#ifdef RT_OS_WINDOWS
884	/*
885	* Windows specific polling cleanup.
886	*/
887	WSAEVENT hEvent = pThis->hEvent;
888	if (hEvent != WSA_INVALID_EVENT)
889	{
890	pThis->hEvent = WSA_INVALID_EVENT;
891	if (!pThis->fPollFallback)
892	{
893	Assert(g_pfnWSACloseEvent);
894	if (g_pfnWSACloseEvent)
895	g_pfnWSACloseEvent(hEvent);
896	}
897	else
898	CloseHandle(hEvent);
899	}
900
901	if (pThis->fPollFallback)
902	{
903	if (pThis->hPollFallbackNotifyW != NIL_RTSOCKETNATIVE)
904	{
905	g_pfnclosesocket(pThis->hPollFallbackNotifyW);
906	pThis->hPollFallbackNotifyW = NIL_RTSOCKETNATIVE;
907	}
908
909	if (pThis->hPollFallbackThread != NIL_RTTHREAD)
910	{
911	int rc2 = RTThreadWait(pThis->hPollFallbackThread, RT_MS_1MIN / 2, NULL);
912	AssertRC(rc2);
913	pThis->hPollFallbackThread = NIL_RTTHREAD;
914	}
915
916	if (pThis->hPollFallbackNotifyR != NIL_RTSOCKETNATIVE)
917	{
918	g_pfnclosesocket(pThis->hPollFallbackNotifyR);
919	pThis->hPollFallbackNotifyR = NIL_RTSOCKETNATIVE;
920	}
921	}
922	#endif
923	}
924
925	return rc;
926	}
927
928
929	RTDECL(uint32_t) RTSocketRelease(RTSOCKET hSocket)
930	{
931	RTSOCKETINT *pThis = hSocket;
932	if (pThis == NIL_RTSOCKET)
933	return 0;
934	AssertPtrReturn(pThis, UINT32_MAX);
935	AssertReturn(pThis->u32Magic == RTSOCKET_MAGIC, UINT32_MAX);
936
937	/* get the refcount without killing it... */
938	uint32_t cRefs = RTMemPoolRefCount(pThis);
939	AssertReturn(cRefs != UINT32_MAX, UINT32_MAX);
940	if (cRefs == 1)
941	rtSocketCloseIt(pThis, true);
942
943	return RTMemPoolRelease(RTMEMPOOL_DEFAULT, pThis);
944	}
945
946
947	RTDECL(int) RTSocketClose(RTSOCKET hSocket)
948	{
949	RTSOCKETINT *pThis = hSocket;
950	if (pThis == NIL_RTSOCKET)
951	return VINF_SUCCESS;
952	AssertPtrReturn(pThis, VERR_INVALID_HANDLE);
953	AssertReturn(pThis->u32Magic == RTSOCKET_MAGIC, VERR_INVALID_HANDLE);
954
955	uint32_t cRefs = RTMemPoolRefCount(pThis);
956	AssertReturn(cRefs != UINT32_MAX, UINT32_MAX);
957
958	int rc = rtSocketCloseIt(pThis, cRefs == 1);
959
960	RTMemPoolRelease(RTMEMPOOL_DEFAULT, pThis);
961	return rc;
962	}
963
964
965	RTDECL(RTHCUINTPTR) RTSocketToNative(RTSOCKET hSocket)
966	{
967	RTSOCKETINT *pThis = hSocket;
968	AssertPtrReturn(pThis, RTHCUINTPTR_MAX);
969	AssertReturn(pThis->u32Magic == RTSOCKET_MAGIC, RTHCUINTPTR_MAX);
970	return (RTHCUINTPTR)pThis->hNative;
971	}
972
973
974	RTDECL(int) RTSocketSetInheritance(RTSOCKET hSocket, bool fInheritable)
975	{
976	RTSOCKETINT *pThis = hSocket;
977	AssertPtrReturn(pThis, VERR_INVALID_HANDLE);
978	AssertReturn(pThis->u32Magic == RTSOCKET_MAGIC, VERR_INVALID_HANDLE);
979	AssertReturn(RTMemPoolRefCount(pThis) >= (pThis->cUsers ? 2U : 1U), VERR_CALLER_NO_REFERENCE);
980
981	#ifndef RT_OS_WINDOWS
982	if (fcntl(pThis->hNative, F_SETFD, fInheritable ? 0 : FD_CLOEXEC) < 0)
983	return RTErrConvertFromErrno(errno);
984	return VINF_SUCCESS;
985	#else
986	/* Windows is more complicated as sockets are complicated wrt inheritance
987	(see stackoverflow for details). In general, though we cannot hope to
988	make a socket really non-inheritable before vista as other layers in
989	the winsock maze may have additional handles associated with the socket. */
990	if (g_pfnGetHandleInformation)
991	{
992	/* Check if the handle is already in what seems to be the right state
993	before we try doing anything. */
994	DWORD fFlags;
995	if (g_pfnGetHandleInformation((HANDLE)pThis->hNative, &fFlags))
996	{
997	if (RT_BOOL(fFlags & HANDLE_FLAG_INHERIT) == fInheritable)
998	return VINF_SUCCESS;
999	}
1000	}
1001
1002	if (!g_pfnSetHandleInformation)
1003	return VERR_NET_NOT_UNSUPPORTED;
1004
1005	if (!g_pfnSetHandleInformation((HANDLE)pThis->hNative, HANDLE_FLAG_INHERIT, fInheritable ? HANDLE_FLAG_INHERIT : 0))
1006	return RTErrConvertFromWin32(GetLastError());
1007	/** @todo Need we do something related to WS_SIO_ASSOCIATE_HANDLE or
1008	* WS_SIO_TRANSLATE_HANDLE? Or what other handles could be associated
1009	* with the socket? that we need to modify? */
1010
1011	return VINF_SUCCESS;
1012	#endif
1013	}
1014
1015
1016	static bool rtSocketIsIPv4Numerical(const char *pszAddress, PRTNETADDRIPV4 pAddr)
1017	{
1018
1019	/* Empty address resolves to the INADDR_ANY address (good for bind). */
1020	if (!pszAddress \|\| !*pszAddress)
1021	{
1022	pAddr->u = INADDR_ANY;
1023	return true;
1024	}
1025
1026	/* Four quads? */
1027	char psz = (char )pszAddress;
1028	for (int i = 0; i < 4; i++)
1029	{
1030	uint8_t u8;
1031	int rc = RTStrToUInt8Ex(psz, &psz, 0, &u8);
1032	if (rc != VINF_SUCCESS && rc != VWRN_TRAILING_CHARS)
1033	return false;
1034	if (*psz != (i < 3 ? '.' : '\0'))
1035	return false;
1036	psz++;
1037
1038	pAddr->au8[i] = u8; /* big endian */
1039	}
1040
1041	return true;
1042	}
1043
1044	RTDECL(int) RTSocketParseInetAddress(const char *pszAddress, unsigned uPort, PRTNETADDR pAddr)
1045	{
1046	int rc;
1047
1048	/*
1049	* Validate input.
1050	*/
1051	AssertReturn(uPort > 0, VERR_INVALID_PARAMETER);
1052	AssertPtrNullReturn(pszAddress, VERR_INVALID_POINTER);
1053
1054	/*
1055	* Resolve the address. Pretty crude at the moment, but we have to make
1056	* sure to not ask the NT 4 gethostbyname about an IPv4 address as it may
1057	* give a wrong answer.
1058	*/
1059	/** @todo this only supports IPv4, and IPv6 support needs to be added.
1060	* It probably needs to be converted to getaddrinfo(). */
1061	RTNETADDRIPV4 IPv4Quad;
1062	if (rtSocketIsIPv4Numerical(pszAddress, &IPv4Quad))
1063	{
1064	Log3(("rtSocketIsIPv4Numerical: %s -> %#x (%RTnaipv4)\n", pszAddress, IPv4Quad.u, IPv4Quad));
1065	RT_ZERO(*pAddr);
1066	pAddr->enmType = RTNETADDRTYPE_IPV4;
1067	pAddr->uPort = uPort;
1068	pAddr->uAddr.IPv4 = IPv4Quad;
1069	return VINF_SUCCESS;
1070	}
1071
1072	#ifdef RT_OS_WINDOWS
1073	/* Initialize WinSock and check version before we call gethostbyname. */
1074	if (!g_pfngethostbyname)
1075	return VERR_NET_NOT_UNSUPPORTED;
1076
1077	int rc2 = rtSocketInitWinsock();
1078	if (RT_FAILURE(rc2))
1079	return rc2;
1080
1081	# define gethostbyname g_pfngethostbyname
1082	#endif
1083
1084	struct hostent *pHostEnt;
1085	pHostEnt = gethostbyname(pszAddress);
1086	if (!pHostEnt)
1087	{
1088	rc = rtSocketResolverError();
1089	AssertMsg(rc == VERR_NET_HOST_NOT_FOUND,
1090	("Could not resolve '%s', rc=%Rrc\n", pszAddress, rc));
1091	return rc;
1092	}
1093
1094	if (pHostEnt->h_addrtype == AF_INET)
1095	{
1096	RT_ZERO(*pAddr);
1097	pAddr->enmType = RTNETADDRTYPE_IPV4;
1098	pAddr->uPort = uPort;
1099	pAddr->uAddr.IPv4.u = ((struct in_addr *)pHostEnt->h_addr)->s_addr;
1100	Log3(("gethostbyname: %s -> %#x (%RTnaipv4)\n", pszAddress, pAddr->uAddr.IPv4.u, pAddr->uAddr.IPv4));
1101	}
1102	else
1103	return VERR_NET_ADDRESS_FAMILY_NOT_SUPPORTED;
1104
1105	#ifdef RT_OS_WINDOWS
1106	# undef gethostbyname
1107	#endif
1108	return VINF_SUCCESS;
1109	}
1110
1111
1112	/*
1113	* New function to allow both ipv4 and ipv6 addresses to be resolved.
1114	* Breaks compatibility with windows before 2000.
1115	*/
1116	RTDECL(int) RTSocketQueryAddressStr(const char pszHost, char pszResult, size_t *pcbResult, PRTNETADDRTYPE penmAddrType)
1117	{
1118	AssertPtrReturn(pszHost, VERR_INVALID_POINTER);
1119	AssertPtrReturn(pcbResult, VERR_INVALID_POINTER);
1120	AssertPtrNullReturn(penmAddrType, VERR_INVALID_POINTER);
1121	AssertPtrNullReturn(pszResult, VERR_INVALID_POINTER);
1122
1123	#if defined(RT_OS_OS2) \|\| defined(RT_OS_WINDOWS) /** @todo dynamically resolve the APIs not present in NT4! */
1124	return VERR_NOT_SUPPORTED;
1125
1126	#else
1127	int rc;
1128	if (*pcbResult < 16)
1129	return VERR_NET_ADDRESS_NOT_AVAILABLE;
1130
1131	/* Setup the hint. */
1132	struct addrinfo grHints;
1133	RT_ZERO(grHints);
1134	grHints.ai_socktype = 0;
1135	grHints.ai_flags = 0;
1136	grHints.ai_protocol = 0;
1137	grHints.ai_family = AF_UNSPEC;
1138	if (penmAddrType)
1139	{
1140	switch (*penmAddrType)
1141	{
1142	case RTNETADDRTYPE_INVALID:
1143	/grHints.ai_family = AF_UNSPEC;/
1144	break;
1145	case RTNETADDRTYPE_IPV4:
1146	grHints.ai_family = AF_INET;
1147	break;
1148	case RTNETADDRTYPE_IPV6:
1149	grHints.ai_family = AF_INET6;
1150	break;
1151	default:
1152	AssertFailedReturn(VERR_INVALID_PARAMETER);
1153	}
1154	}
1155
1156	# ifdef RT_OS_WINDOWS
1157	/*
1158	* Winsock2 init
1159	*/
1160	if ( !g_pfngetaddrinfo
1161	\|\| !g_pfnfreeaddrinfo)
1162	return VERR_NET_NOT_UNSUPPORTED;
1163
1164	int rc2 = rtSocketInitWinsock();
1165	if (RT_FAILURE(rc2))
1166	return rc2;
1167
1168	# define getaddrinfo g_pfngetaddrinfo
1169	# define freeaddrinfo g_pfnfreeaddrinfo
1170	# endif
1171
1172	/** @todo r=bird: getaddrinfo and freeaddrinfo breaks the additions on NT4. */
1173	struct addrinfo *pgrResults = NULL;
1174	rc = getaddrinfo(pszHost, "", &grHints, &pgrResults);
1175	if (rc != 0)
1176	return VERR_NET_ADDRESS_NOT_AVAILABLE;
1177
1178	// return data
1179	// on multiple matches return only the first one
1180
1181	if (!pgrResults)
1182	return VERR_NET_ADDRESS_NOT_AVAILABLE;
1183
1184	struct addrinfo const *pgrResult = pgrResults->ai_next;
1185	if (!pgrResult)
1186	{
1187	freeaddrinfo(pgrResults);
1188	return VERR_NET_ADDRESS_NOT_AVAILABLE;
1189	}
1190
1191	RTNETADDRTYPE enmAddrType = RTNETADDRTYPE_INVALID;
1192	size_t cchIpAddress;
1193	char szIpAddress[48];
1194	if (pgrResult->ai_family == AF_INET)
1195	{
1196	struct sockaddr_in const pgrSa = (struct sockaddr_in const )pgrResult->ai_addr;
1197	cchIpAddress = RTStrPrintf(szIpAddress, sizeof(szIpAddress),
1198	"%RTnaipv4", pgrSa->sin_addr.s_addr);
1199	Assert(cchIpAddress >= 7 && cchIpAddress < sizeof(szIpAddress) - 1);
1200	enmAddrType = RTNETADDRTYPE_IPV4;
1201	rc = VINF_SUCCESS;
1202	}
1203	else if (pgrResult->ai_family == AF_INET6)
1204	{
1205	struct sockaddr_in6 const pgrSa6 = (struct sockaddr_in6 const )pgrResult->ai_addr;
1206	cchIpAddress = RTStrPrintf(szIpAddress, sizeof(szIpAddress),
1207	"%RTnaipv6", (PRTNETADDRIPV6)&pgrSa6->sin6_addr);
1208	enmAddrType = RTNETADDRTYPE_IPV6;
1209	rc = VINF_SUCCESS;
1210	}
1211	else
1212	{
1213	rc = VERR_NET_ADDRESS_NOT_AVAILABLE;
1214	szIpAddress[0] = '\0';
1215	cchIpAddress = 0;
1216	}
1217	freeaddrinfo(pgrResults);
1218
1219	/*
1220	* Copy out the result.
1221	*/
1222	size_t const cbResult = *pcbResult;
1223	*pcbResult = cchIpAddress + 1;
1224	if (cchIpAddress < cbResult)
1225	memcpy(pszResult, szIpAddress, cchIpAddress + 1);
1226	else
1227	{
1228	RT_BZERO(pszResult, cbResult);
1229	if (RT_SUCCESS(rc))
1230	rc = VERR_BUFFER_OVERFLOW;
1231	}
1232	if (penmAddrType && RT_SUCCESS(rc))
1233	*penmAddrType = enmAddrType;
1234	return rc;
1235
1236	# ifdef RT_OS_WINDOWS
1237	# undef getaddrinfo
1238	# undef freeaddrinfo
1239	# endif
1240	#endif /* !RT_OS_OS2 */
1241	}
1242
1243
1244	RTDECL(int) RTSocketRead(RTSOCKET hSocket, void pvBuffer, size_t cbBuffer, size_t pcbRead)
1245	{
1246	/*
1247	* Validate input.
1248	*/
1249	RTSOCKETINT *pThis = hSocket;
1250	AssertPtrReturn(pThis, VERR_INVALID_HANDLE);
1251	AssertReturn(pThis->u32Magic == RTSOCKET_MAGIC, VERR_INVALID_HANDLE);
1252	AssertReturn(cbBuffer > 0, VERR_INVALID_PARAMETER);
1253	AssertPtr(pvBuffer);
1254	#ifdef RT_OS_WINDOWS
1255	AssertReturn(g_pfnrecv, VERR_NET_NOT_UNSUPPORTED);
1256	# define recv g_pfnrecv
1257	#endif
1258	AssertReturn(rtSocketTryLock(pThis), VERR_CONCURRENT_ACCESS);
1259
1260	int rc = rtSocketSwitchBlockingMode(pThis, true /* fBlocking */);
1261	if (RT_FAILURE(rc))
1262	return rc;
1263
1264	/*
1265	* Read loop.
1266	* If pcbRead is NULL we have to fill the entire buffer!
1267	*/
1268	size_t cbRead = 0;
1269	size_t cbToRead = cbBuffer;
1270	for (;;)
1271	{
1272	rtSocketErrorReset();
1273	#ifdef RTSOCKET_MAX_READ
1274	int cbNow = cbToRead >= RTSOCKET_MAX_READ ? RTSOCKET_MAX_READ : (int)cbToRead;
1275	#else
1276	size_t cbNow = cbToRead;
1277	#endif
1278	ssize_t cbBytesRead = recv(pThis->hNative, (char *)pvBuffer + cbRead, cbNow, MSG_NOSIGNAL);
1279	if (cbBytesRead <= 0)
1280	{
1281	rc = rtSocketError();
1282	Assert(RT_FAILURE_NP(rc) \|\| cbBytesRead == 0);
1283	if (RT_SUCCESS_NP(rc))
1284	{
1285	if (!pcbRead)
1286	rc = VERR_NET_SHUTDOWN;
1287	else
1288	{
1289	*pcbRead = 0;
1290	rc = VINF_SUCCESS;
1291	}
1292	}
1293	break;
1294	}
1295	if (pcbRead)
1296	{
1297	/* return partial data */
1298	*pcbRead = cbBytesRead;
1299	break;
1300	}
1301
1302	/* read more? */
1303	cbRead += cbBytesRead;
1304	if (cbRead == cbBuffer)
1305	break;
1306
1307	/* next */
1308	cbToRead = cbBuffer - cbRead;
1309	}
1310
1311	rtSocketUnlock(pThis);
1312	#ifdef RT_OS_WINDOWS
1313	# undef recv
1314	#endif
1315	return rc;
1316	}
1317
1318
1319	RTDECL(int) RTSocketReadFrom(RTSOCKET hSocket, void pvBuffer, size_t cbBuffer, size_t pcbRead, PRTNETADDR pSrcAddr)
1320	{
1321	/*
1322	* Validate input.
1323	*/
1324	RTSOCKETINT *pThis = hSocket;
1325	AssertPtrReturn(pThis, VERR_INVALID_HANDLE);
1326	AssertReturn(pThis->u32Magic == RTSOCKET_MAGIC, VERR_INVALID_HANDLE);
1327	AssertReturn(cbBuffer > 0, VERR_INVALID_PARAMETER);
1328	AssertPtr(pvBuffer);
1329	AssertPtr(pcbRead);
1330	#ifdef RT_OS_WINDOWS
1331	AssertReturn(g_pfnrecvfrom, VERR_NET_NOT_UNSUPPORTED);
1332	# define recvfrom g_pfnrecvfrom
1333	#endif
1334	AssertReturn(rtSocketTryLock(pThis), VERR_CONCURRENT_ACCESS);
1335
1336	int rc = rtSocketSwitchBlockingMode(pThis, true /* fBlocking */);
1337	if (RT_FAILURE(rc))
1338	return rc;
1339
1340	/*
1341	* Read data.
1342	*/
1343	size_t cbRead = 0;
1344	size_t cbToRead = cbBuffer;
1345	rtSocketErrorReset();
1346	RTSOCKADDRUNION u;
1347	#ifdef RTSOCKET_MAX_READ
1348	int cbNow = cbToRead >= RTSOCKET_MAX_READ ? RTSOCKET_MAX_READ : (int)cbToRead;
1349	int cbAddr = sizeof(u);
1350	#else
1351	size_t cbNow = cbToRead;
1352	socklen_t cbAddr = sizeof(u);
1353	#endif
1354	ssize_t cbBytesRead = recvfrom(pThis->hNative, (char *)pvBuffer + cbRead, cbNow, MSG_NOSIGNAL, &u.Addr, &cbAddr);
1355	if (cbBytesRead <= 0)
1356	{
1357	rc = rtSocketError();
1358	Assert(RT_FAILURE_NP(rc) \|\| cbBytesRead == 0);
1359	if (RT_SUCCESS_NP(rc))
1360	{
1361	*pcbRead = 0;
1362	rc = VINF_SUCCESS;
1363	}
1364	}
1365	else
1366	{
1367	if (pSrcAddr)
1368	rc = rtSocketNetAddrFromAddr(&u, cbAddr, pSrcAddr);
1369	*pcbRead = cbBytesRead;
1370	}
1371
1372	rtSocketUnlock(pThis);
1373	#ifdef RT_OS_WINDOWS
1374	# undef recvfrom
1375	#endif
1376	return rc;
1377	}
1378
1379
1380	RTDECL(int) RTSocketWrite(RTSOCKET hSocket, const void *pvBuffer, size_t cbBuffer)
1381	{
1382	/*
1383	* Validate input.
1384	*/
1385	RTSOCKETINT *pThis = hSocket;
1386	AssertPtrReturn(pThis, VERR_INVALID_HANDLE);
1387	AssertReturn(pThis->u32Magic == RTSOCKET_MAGIC, VERR_INVALID_HANDLE);
1388	#ifdef RT_OS_WINDOWS
1389	AssertReturn(g_pfnsend, VERR_NET_NOT_UNSUPPORTED);
1390	# define send g_pfnsend
1391	#endif
1392	AssertReturn(rtSocketTryLock(pThis), VERR_CONCURRENT_ACCESS);
1393
1394	int rc = rtSocketSwitchBlockingMode(pThis, true /* fBlocking */);
1395	if (RT_FAILURE(rc))
1396	return rc;
1397
1398	/*
1399	* Try write all at once.
1400	*/
1401	#ifdef RTSOCKET_MAX_WRITE
1402	int cbNow = cbBuffer >= RTSOCKET_MAX_WRITE ? RTSOCKET_MAX_WRITE : (int)cbBuffer;
1403	#else
1404	size_t cbNow = cbBuffer >= SSIZE_MAX ? SSIZE_MAX : cbBuffer;
1405	#endif
1406	ssize_t cbWritten = send(pThis->hNative, (const char *)pvBuffer, cbNow, MSG_NOSIGNAL);
1407	if (RT_LIKELY((size_t)cbWritten == cbBuffer && cbWritten >= 0))
1408	rc = VINF_SUCCESS;
1409	else if (cbWritten < 0)
1410	rc = rtSocketError();
1411	else
1412	{
1413	/*
1414	* Unfinished business, write the remainder of the request. Must ignore
1415	* VERR_INTERRUPTED here if we've managed to send something.
1416	*/
1417	size_t cbSentSoFar = 0;
1418	for (;;)
1419	{
1420	/* advance */
1421	cbBuffer -= (size_t)cbWritten;
1422	if (!cbBuffer)
1423	break;
1424	cbSentSoFar += (size_t)cbWritten;
1425	pvBuffer = (char const *)pvBuffer + cbWritten;
1426
1427	/* send */
1428	#ifdef RTSOCKET_MAX_WRITE
1429	cbNow = cbBuffer >= RTSOCKET_MAX_WRITE ? RTSOCKET_MAX_WRITE : (int)cbBuffer;
1430	#else
1431	cbNow = cbBuffer >= SSIZE_MAX ? SSIZE_MAX : cbBuffer;
1432	#endif
1433	cbWritten = send(pThis->hNative, (const char *)pvBuffer, cbNow, MSG_NOSIGNAL);
1434	if (cbWritten >= 0)
1435	AssertMsg(cbBuffer >= (size_t)cbWritten, ("Wrote more than we requested!!! cbWritten=%zu cbBuffer=%zu rtSocketError()=%d\n",
1436	cbWritten, cbBuffer, rtSocketError()));
1437	else
1438	{
1439	rc = rtSocketError();
1440	if (rc != VERR_INTERNAL_ERROR \|\| cbSentSoFar == 0)
1441	break;
1442	cbWritten = 0;
1443	rc = VINF_SUCCESS;
1444	}
1445	}
1446	}
1447
1448	rtSocketUnlock(pThis);
1449	#ifdef RT_OS_WINDOWS
1450	# undef send
1451	#endif
1452	return rc;
1453	}
1454
1455
1456	RTDECL(int) RTSocketWriteTo(RTSOCKET hSocket, const void *pvBuffer, size_t cbBuffer, PCRTNETADDR pAddr)
1457	{
1458	/*
1459	* Validate input.
1460	*/
1461	RTSOCKETINT *pThis = hSocket;
1462	AssertPtrReturn(pThis, VERR_INVALID_HANDLE);
1463	AssertReturn(pThis->u32Magic == RTSOCKET_MAGIC, VERR_INVALID_HANDLE);
1464	#ifdef RT_OS_WINDOWS
1465	AssertReturn(g_pfnsendto, VERR_NET_NOT_UNSUPPORTED);
1466	# define sendto g_pfnsendto
1467	#endif
1468
1469	/* no locking since UDP reads may be done concurrently to writes, and
1470	* this is the normal use case of this code. */
1471
1472	int rc = rtSocketSwitchBlockingMode(pThis, true /* fBlocking */);
1473	if (RT_FAILURE(rc))
1474	return rc;
1475
1476	/* Figure out destination address. */
1477	struct sockaddr *pSA = NULL;
1478	#ifdef RT_OS_WINDOWS
1479	int cbSA = 0;
1480	#else
1481	socklen_t cbSA = 0;
1482	#endif
1483	RTSOCKADDRUNION u;
1484	if (pAddr)
1485	{
1486	rc = rtSocketAddrFromNetAddr(pAddr, &u, sizeof(u), NULL);
1487	if (RT_FAILURE(rc))
1488	return rc;
1489	pSA = &u.Addr;
1490	cbSA = sizeof(u);
1491	}
1492
1493	/*
1494	* Must write all at once, otherwise it is a failure.
1495	*/
1496	#ifdef RT_OS_WINDOWS
1497	int cbNow = cbBuffer >= RTSOCKET_MAX_WRITE ? RTSOCKET_MAX_WRITE : (int)cbBuffer;
1498	#else
1499	size_t cbNow = cbBuffer >= SSIZE_MAX ? SSIZE_MAX : cbBuffer;
1500	#endif
1501	ssize_t cbWritten = sendto(pThis->hNative, (const char *)pvBuffer, cbNow, MSG_NOSIGNAL, pSA, cbSA);
1502	if (RT_LIKELY((size_t)cbWritten == cbBuffer && cbWritten >= 0))
1503	rc = VINF_SUCCESS;
1504	else if (cbWritten < 0)
1505	rc = rtSocketError();
1506	else
1507	rc = VERR_TOO_MUCH_DATA;
1508
1509	/// @todo rtSocketUnlock(pThis);
1510	#ifdef RT_OS_WINDOWS
1511	# undef sendto
1512	#endif
1513	return rc;
1514	}
1515
1516
1517	RTDECL(int) RTSocketWriteToNB(RTSOCKET hSocket, const void *pvBuffer, size_t cbBuffer, PCRTNETADDR pAddr)
1518	{
1519	/*
1520	* Validate input.
1521	*/
1522	RTSOCKETINT *pThis = hSocket;
1523	AssertPtrReturn(pThis, VERR_INVALID_HANDLE);
1524	AssertReturn(pThis->u32Magic == RTSOCKET_MAGIC, VERR_INVALID_HANDLE);
1525	#ifdef RT_OS_WINDOWS
1526	AssertReturn(g_pfnsendto, VERR_NET_NOT_UNSUPPORTED);
1527	# define sendto g_pfnsendto
1528	#endif
1529
1530	/* no locking since UDP reads may be done concurrently to writes, and
1531	* this is the normal use case of this code. */
1532
1533	int rc = rtSocketSwitchBlockingMode(pThis, false /* fBlocking */);
1534	if (RT_FAILURE(rc))
1535	return rc;
1536
1537	/* Figure out destination address. */
1538	struct sockaddr *pSA = NULL;
1539	#ifdef RT_OS_WINDOWS
1540	int cbSA = 0;
1541	#else
1542	socklen_t cbSA = 0;
1543	#endif
1544	RTSOCKADDRUNION u;
1545	if (pAddr)
1546	{
1547	rc = rtSocketAddrFromNetAddr(pAddr, &u, sizeof(u), NULL);
1548	if (RT_FAILURE(rc))
1549	return rc;
1550	pSA = &u.Addr;
1551	cbSA = sizeof(u);
1552	}
1553
1554	/*
1555	* Must write all at once, otherwise it is a failure.
1556	*/
1557	#ifdef RT_OS_WINDOWS
1558	int cbNow = cbBuffer >= RTSOCKET_MAX_WRITE ? RTSOCKET_MAX_WRITE : (int)cbBuffer;
1559	#else
1560	size_t cbNow = cbBuffer >= SSIZE_MAX ? SSIZE_MAX : cbBuffer;
1561	#endif
1562	ssize_t cbWritten = sendto(pThis->hNative, (const char *)pvBuffer, cbNow, MSG_NOSIGNAL, pSA, cbSA);
1563	if (RT_LIKELY((size_t)cbWritten == cbBuffer && cbWritten >= 0))
1564	rc = VINF_SUCCESS;
1565	else if (cbWritten < 0)
1566	rc = rtSocketError();
1567	else
1568	rc = VERR_TOO_MUCH_DATA;
1569
1570	/// @todo rtSocketUnlock(pThis);
1571	#ifdef RT_OS_WINDOWS
1572	# undef sendto
1573	#endif
1574	return rc;
1575	}
1576
1577
1578	RTDECL(int) RTSocketSgWrite(RTSOCKET hSocket, PCRTSGBUF pSgBuf)
1579	{
1580	/*
1581	* Validate input.
1582	*/
1583	RTSOCKETINT *pThis = hSocket;
1584	AssertPtrReturn(pThis, VERR_INVALID_HANDLE);
1585	AssertReturn(pThis->u32Magic == RTSOCKET_MAGIC, VERR_INVALID_HANDLE);
1586	AssertPtrReturn(pSgBuf, VERR_INVALID_PARAMETER);
1587	AssertReturn(pSgBuf->cSegs > 0, VERR_INVALID_PARAMETER);
1588	AssertReturn(rtSocketTryLock(pThis), VERR_CONCURRENT_ACCESS);
1589
1590	int rc = rtSocketSwitchBlockingMode(pThis, true /* fBlocking */);
1591	if (RT_FAILURE(rc))
1592	return rc;
1593
1594	/*
1595	* Construct message descriptor (translate pSgBuf) and send it.
1596	*/
1597	rc = VERR_NO_TMP_MEMORY;
1598	#ifdef RT_OS_WINDOWS
1599	if (g_pfnWSASend)
1600	{
1601	AssertCompileSize(WSABUF, sizeof(RTSGSEG));
1602	AssertCompileMemberSize(WSABUF, buf, RT_SIZEOFMEMB(RTSGSEG, pvSeg));
1603
1604	LPWSABUF paMsg = (LPWSABUF)RTMemTmpAllocZ(pSgBuf->cSegs * sizeof(WSABUF));
1605	if (paMsg)
1606	{
1607	for (unsigned i = 0; i < pSgBuf->cSegs; i++)
1608	{
1609	paMsg[i].buf = (char *)pSgBuf->paSegs[i].pvSeg;
1610	paMsg[i].len = (u_long)pSgBuf->paSegs[i].cbSeg;
1611	}
1612
1613	DWORD dwSent;
1614	int hrc = g_pfnWSASend(pThis->hNative, paMsg, pSgBuf->cSegs, &dwSent, MSG_NOSIGNAL, NULL, NULL);
1615	if (!hrc)
1616	rc = VINF_SUCCESS;
1617	/** @todo check for incomplete writes */
1618	else
1619	rc = rtSocketError();
1620
1621	RTMemTmpFree(paMsg);
1622	}
1623	}
1624	else if (g_pfnsend)
1625	{
1626	rc = VINF_SUCCESS;
1627	for (uint32_t iSeg = 0; iSeg < pSgBuf->cSegs; iSeg++)
1628	{
1629	uint8_t const pbSeg = (uint8_t const )pSgBuf->paSegs[iSeg].pvSeg;
1630	size_t cbSeg = pSgBuf->paSegs[iSeg].cbSeg;
1631	int cbNow;
1632	ssize_t cbWritten;
1633	for (;;)
1634	{
1635	cbNow = cbSeg >= RTSOCKET_MAX_WRITE ? RTSOCKET_MAX_WRITE : (int)cbSeg;
1636	cbWritten = g_pfnsend(pThis->hNative, (const char *)pbSeg, cbNow, MSG_NOSIGNAL);
1637	if ((size_t)cbWritten >= cbSeg \|\| cbWritten < 0)
1638	break;
1639	pbSeg += cbWritten;
1640	cbSeg -= cbWritten;
1641	}
1642	if (cbWritten < 0)
1643	{
1644	rc = rtSocketError();
1645	break;
1646	}
1647	}
1648	}
1649	else
1650	rc = VERR_NET_NOT_UNSUPPORTED;
1651
1652	#else /* !RT_OS_WINDOWS */
1653	AssertCompileSize(struct iovec, sizeof(RTSGSEG));
1654	AssertCompileMemberSize(struct iovec, iov_base, RT_SIZEOFMEMB(RTSGSEG, pvSeg));
1655	AssertCompileMemberSize(struct iovec, iov_len, RT_SIZEOFMEMB(RTSGSEG, cbSeg));
1656
1657	struct iovec paMsg = (struct iovec )RTMemTmpAllocZ(pSgBuf->cSegs * sizeof(struct iovec));
1658	if (paMsg)
1659	{
1660	for (unsigned i = 0; i < pSgBuf->cSegs; i++)
1661	{
1662	paMsg[i].iov_base = pSgBuf->paSegs[i].pvSeg;
1663	paMsg[i].iov_len = pSgBuf->paSegs[i].cbSeg;
1664	}
1665
1666	struct msghdr msgHdr;
1667	RT_ZERO(msgHdr);
1668	msgHdr.msg_iov = paMsg;
1669	msgHdr.msg_iovlen = pSgBuf->cSegs;
1670	ssize_t cbWritten = sendmsg(pThis->hNative, &msgHdr, MSG_NOSIGNAL);
1671	if (RT_LIKELY(cbWritten >= 0))
1672	rc = VINF_SUCCESS;
1673	/** @todo check for incomplete writes */
1674	else
1675	rc = rtSocketError();
1676
1677	RTMemTmpFree(paMsg);
1678	}
1679	#endif /* !RT_OS_WINDOWS */
1680
1681	rtSocketUnlock(pThis);
1682	return rc;
1683	}
1684
1685
1686	RTDECL(int) RTSocketSgWriteL(RTSOCKET hSocket, size_t cSegs, ...)
1687	{
1688	va_list va;
1689	va_start(va, cSegs);
1690	int rc = RTSocketSgWriteLV(hSocket, cSegs, va);
1691	va_end(va);
1692	return rc;
1693	}
1694
1695
1696	RTDECL(int) RTSocketSgWriteLV(RTSOCKET hSocket, size_t cSegs, va_list va)
1697	{
1698	/*
1699	* Set up a S/G segment array + buffer on the stack and pass it
1700	* on to RTSocketSgWrite.
1701	*/
1702	Assert(cSegs <= 16);
1703	PRTSGSEG paSegs = (PRTSGSEG)alloca(cSegs * sizeof(RTSGSEG));
1704	AssertReturn(paSegs, VERR_NO_TMP_MEMORY);
1705	for (size_t i = 0; i < cSegs; i++)
1706	{
1707	paSegs[i].pvSeg = va_arg(va, void *);
1708	paSegs[i].cbSeg = va_arg(va, size_t);
1709	}
1710
1711	RTSGBUF SgBuf;
1712	RTSgBufInit(&SgBuf, paSegs, cSegs);
1713	return RTSocketSgWrite(hSocket, &SgBuf);
1714	}
1715
1716
1717	RTDECL(int) RTSocketReadNB(RTSOCKET hSocket, void pvBuffer, size_t cbBuffer, size_t pcbRead)
1718	{
1719	/*
1720	* Validate input.
1721	*/
1722	RTSOCKETINT *pThis = hSocket;
1723	AssertPtrReturn(pThis, VERR_INVALID_HANDLE);
1724	AssertReturn(pThis->u32Magic == RTSOCKET_MAGIC, VERR_INVALID_HANDLE);
1725	AssertReturn(cbBuffer > 0, VERR_INVALID_PARAMETER);
1726	AssertPtr(pvBuffer);
1727	AssertPtrReturn(pcbRead, VERR_INVALID_PARAMETER);
1728	#ifdef RT_OS_WINDOWS
1729	AssertReturn(g_pfnrecv, VERR_NET_NOT_UNSUPPORTED);
1730	#endif
1731	AssertReturn(rtSocketTryLock(pThis), VERR_CONCURRENT_ACCESS);
1732
1733	int rc = rtSocketSwitchBlockingMode(pThis, false /* fBlocking */);
1734	if (RT_FAILURE(rc))
1735	return rc;
1736
1737	rtSocketErrorReset();
1738	#ifdef RTSOCKET_MAX_READ
1739	int cbNow = cbBuffer >= RTSOCKET_MAX_WRITE ? RTSOCKET_MAX_WRITE : (int)cbBuffer;
1740	#else
1741	size_t cbNow = cbBuffer;
1742	#endif
1743
1744	#ifdef RT_OS_WINDOWS
1745	int cbRead = g_pfnrecv(pThis->hNative, (char *)pvBuffer, cbNow, MSG_NOSIGNAL);
1746	if (cbRead >= 0)
1747	{
1748	*pcbRead = cbRead;
1749	rc = VINF_SUCCESS;
1750	}
1751	else
1752	{
1753	rc = rtSocketError();
1754	if (rc == VERR_TRY_AGAIN)
1755	{
1756	*pcbRead = 0;
1757	rc = VINF_TRY_AGAIN;
1758	}
1759	}
1760
1761	#else
1762	ssize_t cbRead = recv(pThis->hNative, pvBuffer, cbNow, MSG_NOSIGNAL);
1763	if (cbRead >= 0)
1764	*pcbRead = cbRead;
1765	else if ( errno == EAGAIN
1766	# ifdef EWOULDBLOCK
1767	# if EWOULDBLOCK != EAGAIN
1768	\|\| errno == EWOULDBLOCK
1769	# endif
1770	# endif
1771	)
1772	{
1773	*pcbRead = 0;
1774	rc = VINF_TRY_AGAIN;
1775	}
1776	else
1777	rc = rtSocketError();
1778	#endif
1779
1780	rtSocketUnlock(pThis);
1781	return rc;
1782	}
1783
1784
1785	RTDECL(int) RTSocketWriteNB(RTSOCKET hSocket, const void pvBuffer, size_t cbBuffer, size_t pcbWritten)
1786	{
1787	/*
1788	* Validate input.
1789	*/
1790	RTSOCKETINT *pThis = hSocket;
1791	AssertPtrReturn(pThis, VERR_INVALID_HANDLE);
1792	AssertReturn(pThis->u32Magic == RTSOCKET_MAGIC, VERR_INVALID_HANDLE);
1793	AssertPtrReturn(pcbWritten, VERR_INVALID_PARAMETER);
1794	#ifdef RT_OS_WINDOWS
1795	AssertReturn(g_pfnsend, VERR_NET_NOT_UNSUPPORTED);
1796	#endif
1797	AssertReturn(rtSocketTryLock(pThis), VERR_CONCURRENT_ACCESS);
1798
1799	int rc = rtSocketSwitchBlockingMode(pThis, false /* fBlocking */);
1800	if (RT_FAILURE(rc))
1801	return rc;
1802
1803	rtSocketErrorReset();
1804	#ifdef RT_OS_WINDOWS
1805	# ifdef RTSOCKET_MAX_WRITE
1806	int cbNow = cbBuffer >= RTSOCKET_MAX_WRITE ? RTSOCKET_MAX_WRITE : (int)cbBuffer;
1807	# else
1808	size_t cbNow = cbBuffer;
1809	# endif
1810	int cbWritten = g_pfnsend(pThis->hNative, (const char *)pvBuffer, cbNow, MSG_NOSIGNAL);
1811	if (cbWritten >= 0)
1812	{
1813	*pcbWritten = cbWritten;
1814	rc = VINF_SUCCESS;
1815	}
1816	else
1817	{
1818	rc = rtSocketError();
1819	if (rc == VERR_TRY_AGAIN)
1820	{
1821	*pcbWritten = 0;
1822	rc = VINF_TRY_AGAIN;
1823	}
1824	}
1825	#else
1826	ssize_t cbWritten = send(pThis->hNative, pvBuffer, cbBuffer, MSG_NOSIGNAL);
1827	if (cbWritten >= 0)
1828	*pcbWritten = cbWritten;
1829	else if ( errno == EAGAIN
1830	# ifdef EWOULDBLOCK
1831	# if EWOULDBLOCK != EAGAIN
1832	\|\| errno == EWOULDBLOCK
1833	# endif
1834	# endif
1835	)
1836	{
1837	*pcbWritten = 0;
1838	rc = VINF_TRY_AGAIN;
1839	}
1840	else
1841	rc = rtSocketError();
1842	#endif
1843
1844	rtSocketUnlock(pThis);
1845	return rc;
1846	}
1847
1848
1849	RTDECL(int) RTSocketSgWriteNB(RTSOCKET hSocket, PCRTSGBUF pSgBuf, size_t *pcbWritten)
1850	{
1851	/*
1852	* Validate input.
1853	*/
1854	RTSOCKETINT *pThis = hSocket;
1855	AssertPtrReturn(pThis, VERR_INVALID_HANDLE);
1856	AssertReturn(pThis->u32Magic == RTSOCKET_MAGIC, VERR_INVALID_HANDLE);
1857	AssertPtrReturn(pSgBuf, VERR_INVALID_PARAMETER);
1858	AssertPtrReturn(pcbWritten, VERR_INVALID_PARAMETER);
1859	AssertReturn(pSgBuf->cSegs > 0, VERR_INVALID_PARAMETER);
1860	AssertReturn(rtSocketTryLock(pThis), VERR_CONCURRENT_ACCESS);
1861
1862	int rc = rtSocketSwitchBlockingMode(pThis, false /* fBlocking */);
1863	if (RT_FAILURE(rc))
1864	return rc;
1865
1866	unsigned cSegsToSend = 0;
1867	rc = VERR_NO_TMP_MEMORY;
1868	#ifdef RT_OS_WINDOWS
1869	if (g_pfnWSASend)
1870	{
1871	LPWSABUF paMsg = NULL;
1872	RTSgBufMapToNative(paMsg, pSgBuf, WSABUF, buf, char *, len, u_long, cSegsToSend);
1873	if (paMsg)
1874	{
1875	DWORD dwSent = 0;
1876	int hrc = g_pfnWSASend(pThis->hNative, paMsg, cSegsToSend, &dwSent, MSG_NOSIGNAL, NULL, NULL);
1877	if (!hrc)
1878	rc = VINF_SUCCESS;
1879	else
1880	rc = rtSocketError();
1881
1882	*pcbWritten = dwSent;
1883
1884	RTMemTmpFree(paMsg);
1885	}
1886	}
1887	else if (g_pfnsend)
1888	{
1889	size_t cbWrittenTotal = 0;
1890	rc = VINF_SUCCESS;
1891	for (uint32_t iSeg = 0; iSeg < pSgBuf->cSegs; iSeg++)
1892	{
1893	uint8_t const pbSeg = (uint8_t const )pSgBuf->paSegs[iSeg].pvSeg;
1894	size_t cbSeg = pSgBuf->paSegs[iSeg].cbSeg;
1895	int cbNow;
1896	ssize_t cbWritten;
1897	for (;;)
1898	{
1899	cbNow = cbSeg >= RTSOCKET_MAX_WRITE ? RTSOCKET_MAX_WRITE : (int)cbSeg;
1900	cbWritten = g_pfnsend(pThis->hNative, (const char *)pbSeg, cbNow, MSG_NOSIGNAL);
1901	if ((size_t)cbWritten >= cbSeg \|\| cbWritten < 0)
1902	break;
1903	cbWrittenTotal += cbWrittenTotal;
1904	pbSeg += cbWritten;
1905	cbSeg -= cbWritten;
1906	}
1907	if (cbWritten < 0)
1908	{
1909	rc = rtSocketError();
1910	break;
1911	}
1912	if (cbWritten != cbNow)
1913	break;
1914	}
1915	*pcbWritten = cbWrittenTotal;
1916	}
1917	else
1918	rc = VERR_NET_NOT_UNSUPPORTED;
1919
1920	#else /* !RT_OS_WINDOWS */
1921	struct iovec *paMsg = NULL;
1922
1923	RTSgBufMapToNative(paMsg, pSgBuf, struct iovec, iov_base, void *, iov_len, size_t, cSegsToSend);
1924	if (paMsg)
1925	{
1926	struct msghdr msgHdr;
1927	RT_ZERO(msgHdr);
1928	msgHdr.msg_iov = paMsg;
1929	msgHdr.msg_iovlen = cSegsToSend;
1930	ssize_t cbWritten = sendmsg(pThis->hNative, &msgHdr, MSG_NOSIGNAL);
1931	if (RT_LIKELY(cbWritten >= 0))
1932	{
1933	rc = VINF_SUCCESS;
1934	*pcbWritten = cbWritten;
1935	}
1936	else
1937	rc = rtSocketError();
1938
1939	RTMemTmpFree(paMsg);
1940	}
1941	#endif /* !RT_OS_WINDOWS */
1942
1943	rtSocketUnlock(pThis);
1944	return rc;
1945	}
1946
1947
1948	RTDECL(int) RTSocketSgWriteLNB(RTSOCKET hSocket, size_t cSegs, size_t *pcbWritten, ...)
1949	{
1950	va_list va;
1951	va_start(va, pcbWritten);
1952	int rc = RTSocketSgWriteLVNB(hSocket, cSegs, pcbWritten, va);
1953	va_end(va);
1954	return rc;
1955	}
1956
1957
1958	RTDECL(int) RTSocketSgWriteLVNB(RTSOCKET hSocket, size_t cSegs, size_t *pcbWritten, va_list va)
1959	{
1960	/*
1961	* Set up a S/G segment array + buffer on the stack and pass it
1962	* on to RTSocketSgWrite.
1963	*/
1964	Assert(cSegs <= 16);
1965	PRTSGSEG paSegs = (PRTSGSEG)alloca(cSegs * sizeof(RTSGSEG));
1966	AssertReturn(paSegs, VERR_NO_TMP_MEMORY);
1967	for (size_t i = 0; i < cSegs; i++)
1968	{
1969	paSegs[i].pvSeg = va_arg(va, void *);
1970	paSegs[i].cbSeg = va_arg(va, size_t);
1971	}
1972
1973	RTSGBUF SgBuf;
1974	RTSgBufInit(&SgBuf, paSegs, cSegs);
1975	return RTSocketSgWriteNB(hSocket, &SgBuf, pcbWritten);
1976	}
1977
1978
1979	RTDECL(int) RTSocketSelectOne(RTSOCKET hSocket, RTMSINTERVAL cMillies)
1980	{
1981	/*
1982	* Validate input.
1983	*/
1984	RTSOCKETINT *pThis = hSocket;
1985	AssertPtrReturn(pThis, VERR_INVALID_HANDLE);
1986	AssertReturn(pThis->u32Magic == RTSOCKET_MAGIC, VERR_INVALID_HANDLE);
1987	AssertReturn(RTMemPoolRefCount(pThis) >= (pThis->cUsers ? 2U : 1U), VERR_CALLER_NO_REFERENCE);
1988	int const fdMax = (int)pThis->hNative + 1;
1989	AssertReturn((RTSOCKETNATIVE)(fdMax - 1) == pThis->hNative, VERR_INTERNAL_ERROR_5);
1990	#ifdef RT_OS_WINDOWS
1991	AssertReturn(g_pfnselect, VERR_NET_NOT_UNSUPPORTED);
1992	# define select g_pfnselect
1993	#endif
1994
1995	/*
1996	* Set up the file descriptor sets and do the select.
1997	*/
1998	fd_set fdsetR;
1999	FD_ZERO(&fdsetR);
2000	FD_SET(pThis->hNative, &fdsetR);
2001
2002	fd_set fdsetE = fdsetR;
2003
2004	int rc;
2005	if (cMillies == RT_INDEFINITE_WAIT)
2006	rc = select(fdMax, &fdsetR, NULL, &fdsetE, NULL);
2007	else
2008	{
2009	struct timeval timeout;
2010	timeout.tv_sec = cMillies / 1000;
2011	timeout.tv_usec = (cMillies % 1000) * 1000;
2012	rc = select(fdMax, &fdsetR, NULL, &fdsetE, &timeout);
2013	}
2014	if (rc > 0)
2015	rc = VINF_SUCCESS;
2016	else if (rc == 0)
2017	rc = VERR_TIMEOUT;
2018	else
2019	rc = rtSocketError();
2020
2021	#ifdef RT_OS_WINDOWS
2022	# undef select
2023	#endif
2024	return rc;
2025	}
2026
2027
2028	/**
2029	* Internal worker for RTSocketSelectOneEx and rtSocketPollCheck (fallback)
2030	*
2031	* @returns IPRT status code
2032	* @param pThis The socket (valid).
2033	* @param fEvents The events to select for.
2034	* @param pfEvents Where to return the events.
2035	* @param cMillies How long to select for, in milliseconds.
2036	*/
2037	static int rtSocketSelectOneEx(RTSOCKET pThis, uint32_t fEvents, uint32_t *pfEvents, RTMSINTERVAL cMillies)
2038	{
2039	RTSOCKETNATIVE hNative = pThis->hNative;
2040	if (hNative == NIL_RTSOCKETNATIVE)
2041	{
2042	/* Socket is already closed? Possible we raced someone calling rtSocketCloseIt.
2043	Should we return a different status code? */
2044	*pfEvents = RTSOCKET_EVT_ERROR;
2045	return VINF_SUCCESS;
2046	}
2047
2048	int const fdMax = (int)hNative + 1;
2049	AssertReturn((RTSOCKETNATIVE)(fdMax - 1) == hNative, VERR_INTERNAL_ERROR_5);
2050	#ifdef RT_OS_WINDOWS
2051	AssertReturn(g_pfnselect, VERR_NET_NOT_UNSUPPORTED);
2052	AssertReturn(g_pfn__WSAFDIsSet, VERR_NET_NOT_UNSUPPORTED);
2053	# define select g_pfnselect
2054	# define __WSAFDIsSet g_pfn__WSAFDIsSet
2055	#endif
2056
2057	*pfEvents = 0;
2058
2059	/*
2060	* Set up the file descriptor sets and do the select.
2061	*/
2062	fd_set fdsetR;
2063	fd_set fdsetW;
2064	fd_set fdsetE;
2065	FD_ZERO(&fdsetR);
2066	FD_ZERO(&fdsetW);
2067	FD_ZERO(&fdsetE);
2068
2069	if (fEvents & RTSOCKET_EVT_READ)
2070	FD_SET(hNative, &fdsetR);
2071	if (fEvents & RTSOCKET_EVT_WRITE)
2072	FD_SET(hNative, &fdsetW);
2073	if (fEvents & RTSOCKET_EVT_ERROR)
2074	FD_SET(hNative, &fdsetE);
2075
2076	int rc;
2077	if (cMillies == RT_INDEFINITE_WAIT)
2078	rc = select(fdMax, &fdsetR, &fdsetW, &fdsetE, NULL);
2079	else
2080	{
2081	struct timeval timeout;
2082	timeout.tv_sec = cMillies / 1000;
2083	timeout.tv_usec = (cMillies % 1000) * 1000;
2084	rc = select(fdMax, &fdsetR, &fdsetW, &fdsetE, &timeout);
2085	}
2086	if (rc > 0)
2087	{
2088	if (pThis->hNative == hNative)
2089	{
2090	if (FD_ISSET(hNative, &fdsetR))
2091	*pfEvents \|= RTSOCKET_EVT_READ;
2092	if (FD_ISSET(hNative, &fdsetW))
2093	*pfEvents \|= RTSOCKET_EVT_WRITE;
2094	if (FD_ISSET(hNative, &fdsetE))
2095	*pfEvents \|= RTSOCKET_EVT_ERROR;
2096	rc = VINF_SUCCESS;
2097	}
2098	else
2099	{
2100	/* Socket was closed while we waited (rtSocketCloseIt). Different status code? */
2101	*pfEvents = RTSOCKET_EVT_ERROR;
2102	rc = VINF_SUCCESS;
2103	}
2104	}
2105	else if (rc == 0)
2106	rc = VERR_TIMEOUT;
2107	else
2108	rc = rtSocketError();
2109
2110	#ifdef RT_OS_WINDOWS
2111	# undef select
2112	# undef __WSAFDIsSet
2113	#endif
2114	return rc;
2115	}
2116
2117
2118	RTDECL(int) RTSocketSelectOneEx(RTSOCKET hSocket, uint32_t fEvents, uint32_t *pfEvents, RTMSINTERVAL cMillies)
2119	{
2120	/*
2121	* Validate input.
2122	*/
2123	RTSOCKETINT *pThis = hSocket;
2124	AssertPtrReturn(pThis, VERR_INVALID_HANDLE);
2125	AssertReturn(pThis->u32Magic == RTSOCKET_MAGIC, VERR_INVALID_HANDLE);
2126	AssertPtrReturn(pfEvents, VERR_INVALID_PARAMETER);
2127	AssertReturn(!(fEvents & ~RTSOCKET_EVT_VALID_MASK), VERR_INVALID_PARAMETER);
2128	AssertReturn(RTMemPoolRefCount(pThis) >= (pThis->cUsers ? 2U : 1U), VERR_CALLER_NO_REFERENCE);
2129
2130	return rtSocketSelectOneEx(pThis, fEvents, pfEvents, cMillies);
2131	}
2132
2133
2134	RTDECL(int) RTSocketShutdown(RTSOCKET hSocket, bool fRead, bool fWrite)
2135	{
2136	/*
2137	* Validate input, don't lock it because we might want to interrupt a call
2138	* active on a different thread.
2139	*/
2140	RTSOCKETINT *pThis = hSocket;
2141	AssertPtrReturn(pThis, VERR_INVALID_HANDLE);
2142	AssertReturn(pThis->u32Magic == RTSOCKET_MAGIC, VERR_INVALID_HANDLE);
2143	AssertReturn(RTMemPoolRefCount(pThis) >= (pThis->cUsers ? 2U : 1U), VERR_CALLER_NO_REFERENCE);
2144	AssertReturn(fRead \|\| fWrite, VERR_INVALID_PARAMETER);
2145	#ifdef RT_OS_WINDOWS
2146	AssertReturn(g_pfnshutdown, VERR_NET_NOT_UNSUPPORTED);
2147	# define shutdown g_pfnshutdown
2148	#endif
2149
2150	/*
2151	* Do the job.
2152	*/
2153	int rc = VINF_SUCCESS;
2154	int fHow;
2155	if (fRead && fWrite)
2156	fHow = SHUT_RDWR;
2157	else if (fRead)
2158	fHow = SHUT_RD;
2159	else
2160	fHow = SHUT_WR;
2161	if (shutdown(pThis->hNative, fHow) == -1)
2162	rc = rtSocketError();
2163
2164	#ifdef RT_OS_WINDOWS
2165	# undef shutdown
2166	#endif
2167	return rc;
2168	}
2169
2170
2171	RTDECL(int) RTSocketGetLocalAddress(RTSOCKET hSocket, PRTNETADDR pAddr)
2172	{
2173	/*
2174	* Validate input.
2175	*/
2176	RTSOCKETINT *pThis = hSocket;
2177	AssertPtrReturn(pThis, VERR_INVALID_HANDLE);
2178	AssertReturn(pThis->u32Magic == RTSOCKET_MAGIC, VERR_INVALID_HANDLE);
2179	AssertReturn(RTMemPoolRefCount(pThis) >= (pThis->cUsers ? 2U : 1U), VERR_CALLER_NO_REFERENCE);
2180	#ifdef RT_OS_WINDOWS
2181	AssertReturn(g_pfngetsockname, VERR_NET_NOT_UNSUPPORTED);
2182	# define getsockname g_pfngetsockname
2183	#endif
2184
2185	/*
2186	* Get the address and convert it.
2187	*/
2188	int rc;
2189	RTSOCKADDRUNION u;
2190	#ifdef RT_OS_WINDOWS
2191	int cbAddr = sizeof(u);
2192	#else
2193	socklen_t cbAddr = sizeof(u);
2194	#endif
2195	RT_ZERO(u);
2196	if (getsockname(pThis->hNative, &u.Addr, &cbAddr) == 0)
2197	rc = rtSocketNetAddrFromAddr(&u, cbAddr, pAddr);
2198	else
2199	rc = rtSocketError();
2200
2201	#ifdef RT_OS_WINDOWS
2202	# undef getsockname
2203	#endif
2204	return rc;
2205	}
2206
2207
2208	RTDECL(int) RTSocketGetPeerAddress(RTSOCKET hSocket, PRTNETADDR pAddr)
2209	{
2210	/*
2211	* Validate input.
2212	*/
2213	RTSOCKETINT *pThis = hSocket;
2214	AssertPtrReturn(pThis, VERR_INVALID_HANDLE);
2215	AssertReturn(pThis->u32Magic == RTSOCKET_MAGIC, VERR_INVALID_HANDLE);
2216	AssertReturn(RTMemPoolRefCount(pThis) >= (pThis->cUsers ? 2U : 1U), VERR_CALLER_NO_REFERENCE);
2217	#ifdef RT_OS_WINDOWS
2218	AssertReturn(g_pfngetpeername, VERR_NET_NOT_UNSUPPORTED);
2219	# define getpeername g_pfngetpeername
2220	#endif
2221
2222	/*
2223	* Get the address and convert it.
2224	*/
2225	int rc;
2226	RTSOCKADDRUNION u;
2227	#ifdef RT_OS_WINDOWS
2228	int cbAddr = sizeof(u);
2229	#else
2230	socklen_t cbAddr = sizeof(u);
2231	#endif
2232	RT_ZERO(u);
2233	if (getpeername(pThis->hNative, &u.Addr, &cbAddr) == 0)
2234	rc = rtSocketNetAddrFromAddr(&u, cbAddr, pAddr);
2235	else
2236	rc = rtSocketError();
2237
2238	#ifdef RT_OS_WINDOWS
2239	# undef getpeername
2240	#endif
2241	return rc;
2242	}
2243
2244
2245
2246	/**
2247	* Wrapper around bind.
2248	*
2249	* @returns IPRT status code.
2250	* @param hSocket The socket handle.
2251	* @param pAddr The address to bind to.
2252	*/
2253	DECLHIDDEN(int) rtSocketBind(RTSOCKET hSocket, PCRTNETADDR pAddr)
2254	{
2255	RTSOCKADDRUNION u;
2256	int cbAddr;
2257	int rc = rtSocketAddrFromNetAddr(pAddr, &u, sizeof(u), &cbAddr);
2258	if (RT_SUCCESS(rc))
2259	rc = rtSocketBindRawAddr(hSocket, &u.Addr, cbAddr);
2260	return rc;
2261	}
2262
2263
2264	/**
2265	* Very thin wrapper around bind.
2266	*
2267	* @returns IPRT status code.
2268	* @param hSocket The socket handle.
2269	* @param pvAddr The address to bind to (struct sockaddr and
2270	* friends).
2271	* @param cbAddr The size of the address.
2272	*/
2273	DECLHIDDEN(int) rtSocketBindRawAddr(RTSOCKET hSocket, void const *pvAddr, size_t cbAddr)
2274	{
2275	/*
2276	* Validate input.
2277	*/
2278	RTSOCKETINT *pThis = hSocket;
2279	AssertPtrReturn(pThis, VERR_INVALID_HANDLE);
2280	AssertReturn(pThis->u32Magic == RTSOCKET_MAGIC, VERR_INVALID_HANDLE);
2281	AssertPtrReturn(pvAddr, VERR_INVALID_POINTER);
2282	#ifdef RT_OS_WINDOWS
2283	AssertReturn(g_pfnbind, VERR_NET_NOT_UNSUPPORTED);
2284	# define bind g_pfnbind
2285	#endif
2286	AssertReturn(rtSocketTryLock(pThis), VERR_CONCURRENT_ACCESS);
2287
2288	int rc;
2289	if (bind(pThis->hNative, (struct sockaddr const *)pvAddr, (int)cbAddr) == 0)
2290	rc = VINF_SUCCESS;
2291	else
2292	rc = rtSocketError();
2293
2294	rtSocketUnlock(pThis);
2295	#ifdef RT_OS_WINDOWS
2296	# undef bind
2297	#endif
2298	return rc;
2299	}
2300
2301
2302
2303	/**
2304	* Wrapper around listen.
2305	*
2306	* @returns IPRT status code.
2307	* @param hSocket The socket handle.
2308	* @param cMaxPending The max number of pending connections.
2309	*/
2310	DECLHIDDEN(int) rtSocketListen(RTSOCKET hSocket, int cMaxPending)
2311	{
2312	/*
2313	* Validate input.
2314	*/
2315	RTSOCKETINT *pThis = hSocket;
2316	AssertPtrReturn(pThis, VERR_INVALID_HANDLE);
2317	AssertReturn(pThis->u32Magic == RTSOCKET_MAGIC, VERR_INVALID_HANDLE);
2318	#ifdef RT_OS_WINDOWS
2319	AssertReturn(g_pfnlisten, VERR_NET_NOT_UNSUPPORTED);
2320	# define listen g_pfnlisten
2321	#endif
2322	AssertReturn(rtSocketTryLock(pThis), VERR_CONCURRENT_ACCESS);
2323
2324	int rc = VINF_SUCCESS;
2325	if (listen(pThis->hNative, cMaxPending) != 0)
2326	rc = rtSocketError();
2327
2328	rtSocketUnlock(pThis);
2329	#ifdef RT_OS_WINDOWS
2330	# undef listen
2331	#endif
2332	return rc;
2333	}
2334
2335
2336	/**
2337	* Wrapper around accept.
2338	*
2339	* @returns IPRT status code.
2340	* @param hSocket The socket handle.
2341	* @param phClient Where to return the client socket handle on
2342	* success.
2343	* @param pAddr Where to return the client address.
2344	* @param pcbAddr On input this gives the size buffer size of what
2345	* @a pAddr point to. On return this contains the
2346	* size of what's stored at @a pAddr.
2347	*/
2348	DECLHIDDEN(int) rtSocketAccept(RTSOCKET hSocket, PRTSOCKET phClient, struct sockaddr pAddr, size_t pcbAddr)
2349	{
2350	/*
2351	* Validate input.
2352	* Only lock the socket temporarily while we get the native handle, so that
2353	* we can safely shutdown and destroy the socket from a different thread.
2354	*/
2355	RTSOCKETINT *pThis = hSocket;
2356	AssertPtrReturn(pThis, VERR_INVALID_HANDLE);
2357	AssertReturn(pThis->u32Magic == RTSOCKET_MAGIC, VERR_INVALID_HANDLE);
2358	#ifdef RT_OS_WINDOWS
2359	AssertReturn(g_pfnaccept, VERR_NET_NOT_UNSUPPORTED);
2360	AssertReturn(g_pfnclosesocket, VERR_NET_NOT_UNSUPPORTED);
2361	# define accept g_pfnaccept
2362	#endif
2363	AssertReturn(rtSocketTryLock(pThis), VERR_CONCURRENT_ACCESS);
2364
2365	/*
2366	* Call accept().
2367	*/
2368	rtSocketErrorReset();
2369	int rc = VINF_SUCCESS;
2370	#ifdef RT_OS_WINDOWS
2371	int cbAddr = (int)*pcbAddr;
2372	#else
2373	socklen_t cbAddr = *pcbAddr;
2374	#endif
2375	RTSOCKETNATIVE hNativeClient = accept(pThis->hNative, pAddr, &cbAddr);
2376	if (hNativeClient != NIL_RTSOCKETNATIVE)
2377	{
2378	*pcbAddr = cbAddr;
2379
2380	/*
2381	* Wrap the client socket.
2382	*/
2383	rc = rtSocketCreateForNative(phClient, hNativeClient, false /fLeaveOpen/);
2384	if (RT_FAILURE(rc))
2385	{
2386	#ifdef RT_OS_WINDOWS
2387	g_pfnclosesocket(hNativeClient);
2388	#else
2389	close(hNativeClient);
2390	#endif
2391	}
2392	}
2393	else
2394	rc = rtSocketError();
2395
2396	rtSocketUnlock(pThis);
2397	#ifdef RT_OS_WINDOWS
2398	# undef accept
2399	#endif
2400	return rc;
2401	}
2402
2403
2404	/**
2405	* Wrapper around connect.
2406	*
2407	* @returns IPRT status code.
2408	* @param hSocket The socket handle.
2409	* @param pAddr The socket address to connect to.
2410	* @param cMillies Number of milliseconds to wait for the connect attempt to complete.
2411	* Use RT_INDEFINITE_WAIT to wait for ever.
2412	* Use RT_TCPCLIENTCONNECT_DEFAULT_WAIT to wait for the default time
2413	* configured on the running system.
2414	*/
2415	DECLHIDDEN(int) rtSocketConnect(RTSOCKET hSocket, PCRTNETADDR pAddr, RTMSINTERVAL cMillies)
2416	{
2417	/*
2418	* Validate input.
2419	*/
2420	RTSOCKETINT *pThis = hSocket;
2421	AssertPtrReturn(pThis, VERR_INVALID_HANDLE);
2422	AssertReturn(pThis->u32Magic == RTSOCKET_MAGIC, VERR_INVALID_HANDLE);
2423	#ifdef RT_OS_WINDOWS
2424	AssertReturn(g_pfnconnect, VERR_NET_NOT_UNSUPPORTED);
2425	AssertReturn(g_pfnselect, VERR_NET_NOT_UNSUPPORTED);
2426	AssertReturn(g_pfngetsockopt, VERR_NET_NOT_UNSUPPORTED);
2427	# define connect g_pfnconnect
2428	# define select g_pfnselect
2429	# define getsockopt g_pfngetsockopt
2430	#endif
2431	AssertReturn(rtSocketTryLock(pThis), VERR_CONCURRENT_ACCESS);
2432
2433	RTSOCKADDRUNION u;
2434	int cbAddr;
2435	int rc = rtSocketAddrFromNetAddr(pAddr, &u, sizeof(u), &cbAddr);
2436	if (RT_SUCCESS(rc))
2437	{
2438	if (cMillies == RT_SOCKETCONNECT_DEFAULT_WAIT)
2439	{
2440	if (connect(pThis->hNative, &u.Addr, cbAddr) != 0)
2441	rc = rtSocketError();
2442	}
2443	else
2444	{
2445	/*
2446	* Switch the socket to nonblocking mode, initiate the connect
2447	* and wait for the socket to become writable or until the timeout
2448	* expires.
2449	*/
2450	rc = rtSocketSwitchBlockingMode(pThis, false /* fBlocking */);
2451	if (RT_SUCCESS(rc))
2452	{
2453	if (connect(pThis->hNative, &u.Addr, cbAddr) != 0)
2454	{
2455	rc = rtSocketError();
2456	if (rc == VERR_TRY_AGAIN \|\| rc == VERR_NET_IN_PROGRESS)
2457	{
2458	int rcSock = 0;
2459	fd_set FdSetWriteable;
2460	struct timeval TvTimeout;
2461
2462	TvTimeout.tv_sec = cMillies / RT_MS_1SEC;
2463	TvTimeout.tv_usec = (cMillies % RT_MS_1SEC) * RT_US_1MS;
2464
2465	FD_ZERO(&FdSetWriteable);
2466	FD_SET(pThis->hNative, &FdSetWriteable);
2467	do
2468	{
2469	rcSock = select(pThis->hNative + 1, NULL, &FdSetWriteable, NULL,
2470	cMillies == RT_INDEFINITE_WAIT \|\| cMillies >= INT_MAX
2471	? NULL
2472	: &TvTimeout);
2473	if (rcSock > 0)
2474	{
2475	int iSockError = 0;
2476	socklen_t cbSockOpt = sizeof(iSockError);
2477	rcSock = getsockopt(pThis->hNative, SOL_SOCKET, SO_ERROR, (char *)&iSockError, &cbSockOpt);
2478	if (rcSock == 0)
2479	{
2480	if (iSockError == 0)
2481	rc = VINF_SUCCESS;
2482	else
2483	{
2484	#ifdef RT_OS_WINDOWS
2485	rc = RTErrConvertFromWin32(iSockError);
2486	#else
2487	rc = RTErrConvertFromErrno(iSockError);
2488	#endif
2489	}
2490	}
2491	else
2492	rc = rtSocketError();
2493	}
2494	else if (rcSock == 0)
2495	rc = VERR_TIMEOUT;
2496	else
2497	rc = rtSocketError();
2498	} while (rc == VERR_INTERRUPTED);
2499	}
2500	}
2501
2502	rtSocketSwitchBlockingMode(pThis, true /* fBlocking */);
2503	}
2504	}
2505	}
2506
2507	rtSocketUnlock(pThis);
2508	#ifdef RT_OS_WINDOWS
2509	# undef connect
2510	# undef select
2511	# undef getsockopt
2512	#endif
2513	return rc;
2514	}
2515
2516
2517	/**
2518	* Wrapper around connect, raw address, no timeout.
2519	*
2520	* @returns IPRT status code.
2521	* @param hSocket The socket handle.
2522	* @param pvAddr The raw socket address to connect to.
2523	* @param cbAddr The size of the raw address.
2524	*/
2525	DECLHIDDEN(int) rtSocketConnectRaw(RTSOCKET hSocket, void const *pvAddr, size_t cbAddr)
2526	{
2527	/*
2528	* Validate input.
2529	*/
2530	RTSOCKETINT *pThis = hSocket;
2531	AssertPtrReturn(pThis, VERR_INVALID_HANDLE);
2532	AssertReturn(pThis->u32Magic == RTSOCKET_MAGIC, VERR_INVALID_HANDLE);
2533	#ifdef RT_OS_WINDOWS
2534	AssertReturn(g_pfnconnect, VERR_NET_NOT_UNSUPPORTED);
2535	# define connect g_pfnconnect
2536	#endif
2537	AssertReturn(rtSocketTryLock(pThis), VERR_CONCURRENT_ACCESS);
2538
2539	int rc;
2540	if (connect(pThis->hNative, (const struct sockaddr *)pvAddr, (int)cbAddr) == 0)
2541	rc = VINF_SUCCESS;
2542	else
2543	rc = rtSocketError();
2544
2545	rtSocketUnlock(pThis);
2546	#ifdef RT_OS_WINDOWS
2547	# undef connect
2548	#endif
2549	return rc;
2550	}
2551
2552
2553	/**
2554	* Wrapper around setsockopt.
2555	*
2556	* @returns IPRT status code.
2557	* @param hSocket The socket handle.
2558	* @param iLevel The protocol level, e.g. IPPORTO_TCP.
2559	* @param iOption The option, e.g. TCP_NODELAY.
2560	* @param pvValue The value buffer.
2561	* @param cbValue The size of the value pointed to by pvValue.
2562	*/
2563	DECLHIDDEN(int) rtSocketSetOpt(RTSOCKET hSocket, int iLevel, int iOption, void const *pvValue, int cbValue)
2564	{
2565	/*
2566	* Validate input.
2567	*/
2568	RTSOCKETINT *pThis = hSocket;
2569	AssertPtrReturn(pThis, VERR_INVALID_HANDLE);
2570	AssertReturn(pThis->u32Magic == RTSOCKET_MAGIC, VERR_INVALID_HANDLE);
2571	#ifdef RT_OS_WINDOWS
2572	AssertReturn(g_pfnsetsockopt, VERR_NET_NOT_UNSUPPORTED);
2573	# define setsockopt g_pfnsetsockopt
2574	#endif
2575	AssertReturn(rtSocketTryLock(pThis), VERR_CONCURRENT_ACCESS);
2576
2577	int rc = VINF_SUCCESS;
2578	if (setsockopt(pThis->hNative, iLevel, iOption, (const char *)pvValue, cbValue) != 0)
2579	rc = rtSocketError();
2580
2581	rtSocketUnlock(pThis);
2582	#ifdef RT_OS_WINDOWS
2583	# undef setsockopt
2584	#endif
2585	return rc;
2586	}
2587
2588
2589	/**
2590	* Internal RTPollSetAdd helper that returns the handle that should be added to
2591	* the pollset.
2592	*
2593	* @returns Valid handle on success, INVALID_HANDLE_VALUE on failure.
2594	* @param hSocket The socket handle.
2595	* @param fEvents The events we're polling for.
2596	* @param phNative Where to put the primary handle.
2597	*/
2598	DECLHIDDEN(int) rtSocketPollGetHandle(RTSOCKET hSocket, uint32_t fEvents, PRTHCINTPTR phNative)
2599	{
2600	RTSOCKETINT *pThis = hSocket;
2601	RT_NOREF_PV(fEvents);
2602	AssertPtrReturn(pThis, VERR_INVALID_HANDLE);
2603	AssertReturn(pThis->u32Magic == RTSOCKET_MAGIC, VERR_INVALID_HANDLE);
2604	#ifdef RT_OS_WINDOWS
2605	AssertReturn(rtSocketTryLock(pThis), VERR_CONCURRENT_ACCESS);
2606
2607	int rc = VINF_SUCCESS;
2608	if (pThis->hEvent != WSA_INVALID_EVENT)
2609	*phNative = (RTHCINTPTR)pThis->hEvent;
2610	else if (g_pfnWSACreateEvent)
2611	{
2612	pThis->hEvent = g_pfnWSACreateEvent();
2613	*phNative = (RTHCINTPTR)pThis->hEvent;
2614	if (pThis->hEvent == WSA_INVALID_EVENT)
2615	rc = rtSocketError();
2616	}
2617	else
2618	{
2619	AssertCompile(WSA_INVALID_EVENT == (WSAEVENT)NULL);
2620	pThis->hEvent = CreateEventW(NULL, TRUE /fManualReset/, FALSE /fInitialState/, NULL /pwszName/);
2621	*phNative = (RTHCINTPTR)pThis->hEvent;
2622	if (pThis->hEvent == WSA_INVALID_EVENT)
2623	rc = RTErrConvertFromWin32(GetLastError());
2624	}
2625
2626	rtSocketUnlock(pThis);
2627	return rc;
2628
2629	#else /* !RT_OS_WINDOWS */
2630	*phNative = (RTHCUINTPTR)pThis->hNative;
2631	return VINF_SUCCESS;
2632	#endif /* !RT_OS_WINDOWS */
2633	}
2634
2635	#ifdef RT_OS_WINDOWS
2636
2637	/**
2638	* Fallback poller thread.
2639	*
2640	* @returns VINF_SUCCESS.
2641	* @param hSelf The thread handle.
2642	* @param pvUser Socket instance data.
2643	*/
2644	static DECLCALLBACK(int) rtSocketPollFallbackThreadProc(RTTHREAD hSelf, void *pvUser)
2645	{
2646	RTSOCKETINT pThis = (RTSOCKETINT )pvUser;
2647	RT_NOREF(hSelf);
2648	# define __WSAFDIsSet g_pfn__WSAFDIsSet
2649
2650	/*
2651	* The execution loop.
2652	*/
2653	while (!ASMAtomicReadBool(&pThis->fPollFallbackShutdown))
2654	{
2655	/*
2656	* Do the selecting (with a 15 second timeout because that seems like a good idea).
2657	*/
2658	struct fd_set SetRead;
2659	struct fd_set SetWrite;
2660	struct fd_set SetXcpt;
2661
2662	FD_ZERO(&SetRead);
2663	FD_ZERO(&SetWrite);
2664	FD_ZERO(&SetXcpt);
2665
2666	FD_SET(pThis->hPollFallbackNotifyR, &SetRead);
2667	FD_SET(pThis->hPollFallbackNotifyR, &SetXcpt);
2668
2669	bool fActive = ASMAtomicReadBool(&pThis->fPollFallbackActive);
2670	uint32_t fEvents;
2671	if (!fActive)
2672	fEvents = 0;
2673	else
2674	{
2675	fEvents = ASMAtomicReadU32(&pThis->fSubscribedEvts);
2676	if (fEvents & RTPOLL_EVT_READ)
2677	FD_SET(pThis->hNative, &SetRead);
2678	if (fEvents & RTPOLL_EVT_WRITE)
2679	FD_SET(pThis->hNative, &SetWrite);
2680	if (fEvents & RTPOLL_EVT_ERROR)
2681	FD_SET(pThis->hNative, &SetXcpt);
2682	}
2683
2684	struct timeval Timeout;
2685	Timeout.tv_sec = 15;
2686	Timeout.tv_usec = 0;
2687	int rc = g_pfnselect(INT_MAX /ignored/, &SetRead, &SetWrite, &SetXcpt, &Timeout);
2688
2689	/* Stop immediately if told to shut down. */
2690	if (ASMAtomicReadBool(&pThis->fPollFallbackShutdown))
2691	break;
2692
2693	/*
2694	* Process the result.
2695	*/
2696	if (rc > 0)
2697	{
2698	/* First the socket we're listening on. */
2699	if ( fEvents
2700	&& ( FD_ISSET(pThis->hNative, &SetRead)
2701	\|\| FD_ISSET(pThis->hNative, &SetWrite)
2702	\|\| FD_ISSET(pThis->hNative, &SetXcpt)) )
2703	{
2704	ASMAtomicWriteBool(&pThis->fPollFallbackActive, false);
2705	SetEvent(pThis->hEvent);
2706	}
2707
2708	/* Then maintain the notification pipe. (We only read one byte here
2709	because we're overly paranoid wrt socket switching to blocking mode.) */
2710	if (FD_ISSET(pThis->hPollFallbackNotifyR, &SetRead))
2711	{
2712	char chIgnored;
2713	g_pfnrecv(pThis->hPollFallbackNotifyR, &chIgnored, sizeof(chIgnored), MSG_NOSIGNAL);
2714	}
2715	}
2716	else
2717	AssertMsg(rc == 0, ("%Rrc\n", rtSocketError()));
2718	}
2719
2720	# undef __WSAFDIsSet
2721	return VINF_SUCCESS;
2722	}
2723
2724
2725	/**
2726	* Pokes the fallback thread, making sure it gets out of whatever it's stuck in.
2727	*
2728	* @param pThis The socket handle.
2729	*/
2730	static void rtSocketPokePollFallbackThread(RTSOCKETINT *pThis)
2731	{
2732	Assert(pThis->fPollFallback);
2733	if (pThis->hPollFallbackThread != NIL_RTTHREAD)
2734	{
2735	int cbWritten = g_pfnsend(pThis->hPollFallbackNotifyW, "!", 1, MSG_NOSIGNAL);
2736	AssertMsg(cbWritten == 1, ("cbWritten=%d err=%Rrc\n", rtSocketError()));
2737	RT_NOREF_PV(cbWritten);
2738	}
2739	}
2740
2741
2742	/**
2743	* Called by rtSocketPollStart to make the thread start selecting on the socket.
2744	*
2745	* @returns 0 on success, RTPOLL_EVT_ERROR on failure.
2746	* @param pThis The socket handle.
2747	*/
2748	static uint32_t rtSocketPollFallbackStart(RTSOCKETINT *pThis)
2749	{
2750	/*
2751	* Reset the event and tell the thread to start selecting on the socket.
2752	*/
2753	ResetEvent(pThis->hEvent);
2754	ASMAtomicWriteBool(&pThis->fPollFallbackActive, true);
2755
2756	/*
2757	* Wake up the thread the thread.
2758	*/
2759	if (pThis->hPollFallbackThread != NIL_RTTHREAD)
2760	rtSocketPokePollFallbackThread(pThis);
2761	else
2762	{
2763	/*
2764	* Not running, need to set it up and start it.
2765	*/
2766	AssertLogRelReturn(pThis->hEvent != NULL && pThis->hEvent != INVALID_HANDLE_VALUE, RTPOLL_EVT_ERROR);
2767
2768	/* Create the notification socket pair. */
2769	int rc;
2770	if (pThis->hPollFallbackNotifyR == NIL_RTSOCKETNATIVE)
2771	{
2772	rc = rtSocketCreateNativeTcpPair(&pThis->hPollFallbackNotifyW, &pThis->hPollFallbackNotifyR);
2773	AssertLogRelRCReturn(rc, RTPOLL_EVT_ERROR);
2774
2775	/* Make the read end non-blocking (not fatal). */
2776	u_long fNonBlocking = 1;
2777	rc = g_pfnioctlsocket(pThis->hPollFallbackNotifyR, FIONBIO, &fNonBlocking);
2778	AssertLogRelMsg(rc == 0, ("rc=%#x %Rrc\n", rc, rtSocketError()));
2779	}
2780
2781	/* Finally, start the thread. ASSUME we don't need too much stack. */
2782	rc = RTThreadCreate(&pThis->hPollFallbackThread, rtSocketPollFallbackThreadProc, pThis,
2783	_128K, RTTHREADTYPE_IO, RTTHREADFLAGS_WAITABLE, "sockpoll");
2784	AssertLogRelRCReturn(rc, RTPOLL_EVT_ERROR);
2785	}
2786	return 0;
2787	}
2788
2789
2790	/**
2791	* Undos the harm done by WSAEventSelect.
2792	*
2793	* @returns IPRT status code.
2794	* @param pThis The socket handle.
2795	*/
2796	static int rtSocketPollClearEventAndRestoreBlocking(RTSOCKETINT *pThis)
2797	{
2798	int rc = VINF_SUCCESS;
2799	if (pThis->fSubscribedEvts)
2800	{
2801	if (!pThis->fPollFallback)
2802	{
2803	Assert(g_pfnWSAEventSelect && g_pfnioctlsocket);
2804	if (g_pfnWSAEventSelect && g_pfnioctlsocket)
2805	{
2806	if (g_pfnWSAEventSelect(pThis->hNative, WSA_INVALID_EVENT, 0) == 0)
2807	{
2808	pThis->fSubscribedEvts = 0;
2809
2810	/*
2811	* Switch back to blocking mode if that was the state before the
2812	* operation.
2813	*/
2814	if (pThis->fBlocking)
2815	{
2816	u_long fNonBlocking = 0;
2817	int rc2 = g_pfnioctlsocket(pThis->hNative, FIONBIO, &fNonBlocking);
2818	if (rc2 != 0)
2819	{
2820	rc = rtSocketError();
2821	AssertMsgFailed(("%Rrc; rc2=%d\n", rc, rc2));
2822	}
2823	}
2824	}
2825	else
2826	{
2827	rc = rtSocketError();
2828	AssertMsgFailed(("%Rrc\n", rc));
2829	}
2830	}
2831	else
2832	{
2833	Assert(pThis->fPollFallback);
2834	rc = VINF_SUCCESS;
2835	}
2836	}
2837	/*
2838	* Just clear the event mask as we never started waiting if we get here.
2839	*/
2840	else
2841	ASMAtomicWriteU32(&pThis->fSubscribedEvts, 0);
2842	}
2843	return rc;
2844	}
2845
2846
2847	/**
2848	* Updates the mask of events we're subscribing to.
2849	*
2850	* @returns IPRT status code.
2851	* @param pThis The socket handle.
2852	* @param fEvents The events we want to subscribe to.
2853	*/
2854	static int rtSocketPollUpdateEvents(RTSOCKETINT *pThis, uint32_t fEvents)
2855	{
2856	if (!pThis->fPollFallback)
2857	{
2858	LONG fNetworkEvents = 0;
2859	if (fEvents & RTPOLL_EVT_READ)
2860	fNetworkEvents \|= FD_READ;
2861	if (fEvents & RTPOLL_EVT_WRITE)
2862	fNetworkEvents \|= FD_WRITE;
2863	if (fEvents & RTPOLL_EVT_ERROR)
2864	fNetworkEvents \|= FD_CLOSE;
2865	LogFlowFunc(("fNetworkEvents=%#x\n", fNetworkEvents));
2866
2867	if (g_pfnWSAEventSelect(pThis->hNative, pThis->hEvent, fNetworkEvents) == 0)
2868	{
2869	pThis->fSubscribedEvts = fEvents;
2870	return VINF_SUCCESS;
2871	}
2872
2873	int rc = rtSocketError();
2874	AssertMsgFailed(("fNetworkEvents=%#x rc=%Rrc\n", fNetworkEvents, rtSocketError()));
2875	return rc;
2876	}
2877
2878	/*
2879	* Update the events we're waiting for. Caller will poke/start the thread. later
2880	*/
2881	ASMAtomicWriteU32(&pThis->fSubscribedEvts, fEvents);
2882	return VINF_SUCCESS;
2883	}
2884
2885	#endif /* RT_OS_WINDOWS */
2886
2887
2888	#if defined(RT_OS_WINDOWS) \|\| defined(RT_OS_OS2)
2889
2890	/**
2891	* Checks for pending events.
2892	*
2893	* @returns Event mask or 0.
2894	* @param pThis The socket handle.
2895	* @param fEvents The desired events.
2896	*/
2897	static uint32_t rtSocketPollCheck(RTSOCKETINT *pThis, uint32_t fEvents)
2898	{
2899	uint32_t fRetEvents = 0;
2900
2901	LogFlowFunc(("pThis=%#p fEvents=%#x\n", pThis, fEvents));
2902
2903	# ifdef RT_OS_WINDOWS
2904	/* Make sure WSAEnumNetworkEvents returns what we want. */
2905	int rc = VINF_SUCCESS;
2906	if ((pThis->fSubscribedEvts & fEvents) != fEvents)
2907	rc = rtSocketPollUpdateEvents(pThis, pThis->fSubscribedEvts \| fEvents);
2908
2909	if (!pThis->fPollFallback)
2910	{
2911	/* Atomically get pending events and reset the event semaphore. */
2912	Assert(g_pfnWSAEnumNetworkEvents);
2913	WSANETWORKEVENTS NetEvts;
2914	RT_ZERO(NetEvts);
2915	if (g_pfnWSAEnumNetworkEvents(pThis->hNative, pThis->hEvent, &NetEvts) == 0)
2916	{
2917	if ( (NetEvts.lNetworkEvents & FD_READ)
2918	&& NetEvts.iErrorCode[FD_READ_BIT] == 0)
2919	fRetEvents \|= RTPOLL_EVT_READ;
2920
2921	if ( (NetEvts.lNetworkEvents & FD_WRITE)
2922	&& NetEvts.iErrorCode[FD_WRITE_BIT] == 0)
2923	fRetEvents \|= RTPOLL_EVT_WRITE;
2924
2925	if (NetEvts.lNetworkEvents & FD_CLOSE)
2926	fRetEvents \|= RTPOLL_EVT_ERROR;
2927	else
2928	for (uint32_t i = 0; i < FD_MAX_EVENTS; i++)
2929	if ( (NetEvts.lNetworkEvents & (1L << i))
2930	&& NetEvts.iErrorCode[i] != 0)
2931	fRetEvents \|= RTPOLL_EVT_ERROR;
2932
2933	pThis->fEventsSaved = fRetEvents \|= pThis->fEventsSaved;
2934	fRetEvents &= fEvents \| RTPOLL_EVT_ERROR;
2935	}
2936	else
2937	rc = rtSocketError();
2938	}
2939
2940	/* Fall back on select if we hit an error above or is using fallback polling. */
2941	if (pThis->fPollFallback \|\| RT_FAILURE(rc))
2942	{
2943	rc = rtSocketSelectOneEx(pThis, fEvents & RTPOLL_EVT_ERROR ? fEvents \| RTPOLL_EVT_READ : fEvents, &fRetEvents, 0);
2944	if (RT_SUCCESS(rc))
2945	{
2946	/* rtSocketSelectOneEx may return RTPOLL_EVT_READ on disconnect. Use
2947	getpeername to fix this. */
2948	if ((fRetEvents & (RTPOLL_EVT_READ \| RTPOLL_EVT_ERROR)) == RTPOLL_EVT_READ)
2949	{
2950	# if 0 /* doens't work */
2951	rtSocketErrorReset();
2952	char chIgn;
2953	rc = g_pfnrecv(pThis->hNative, &chIgn, 0, MSG_NOSIGNAL);
2954	rc = rtSocketError();
2955	if (RT_FAILURE(rc))
2956	fRetEvents \|= RTPOLL_EVT_ERROR;
2957
2958	rc = g_pfnsend(pThis->hNative, &chIgn, 0, MSG_NOSIGNAL);
2959	rc = rtSocketError();
2960	if (RT_FAILURE(rc))
2961	fRetEvents \|= RTPOLL_EVT_ERROR;
2962
2963	RTSOCKADDRUNION u;
2964	int cbAddr = sizeof(u);
2965	if (g_pfngetpeername(pThis->hNative, &u.Addr, &cbAddr) == SOCKET_ERROR)
2966	fRetEvents \|= RTPOLL_EVT_ERROR;
2967	# endif
2968	/* If no bytes are available, assume error condition. */
2969	u_long cbAvail = 0;
2970	rc = g_pfnioctlsocket(pThis->hNative, FIONREAD, &cbAvail);
2971	if (rc == 0 && cbAvail == 0)
2972	fRetEvents \|= RTPOLL_EVT_ERROR;
2973	}
2974	fRetEvents &= fEvents \| RTPOLL_EVT_ERROR;
2975	}
2976	else if (rc == VERR_TIMEOUT)
2977	fRetEvents = 0;
2978	else
2979	fRetEvents \|= RTPOLL_EVT_ERROR;
2980	}
2981
2982	# else /* RT_OS_OS2 */
2983	int aFds[4] = { pThis->hNative, pThis->hNative, pThis->hNative, -1 };
2984	int rc = os2_select(aFds, 1, 1, 1, 0);
2985	if (rc > 0)
2986	{
2987	if (aFds[0] == pThis->hNative)
2988	fRetEvents \|= RTPOLL_EVT_READ;
2989	if (aFds[1] == pThis->hNative)
2990	fRetEvents \|= RTPOLL_EVT_WRITE;
2991	if (aFds[2] == pThis->hNative)
2992	fRetEvents \|= RTPOLL_EVT_ERROR;
2993	fRetEvents &= fEvents;
2994	}
2995	# endif /* RT_OS_OS2 */
2996
2997	LogFlowFunc(("fRetEvents=%#x\n", fRetEvents));
2998	return fRetEvents;
2999	}
3000
3001
3002	/**
3003	* Internal RTPoll helper that polls the socket handle and, if @a fNoWait is
3004	* clear, starts whatever actions we've got running during the poll call.
3005	*
3006	* @returns 0 if no pending events, actions initiated if @a fNoWait is clear.
3007	* Event mask (in @a fEvents) and no actions if the handle is ready
3008	* already.
3009	* UINT32_MAX (asserted) if the socket handle is busy in I/O or a
3010	* different poll set.
3011	*
3012	* @param hSocket The socket handle.
3013	* @param hPollSet The poll set handle (for access checks).
3014	* @param fEvents The events we're polling for.
3015	* @param fFinalEntry Set if this is the final entry for this handle
3016	* in this poll set. This can be used for dealing
3017	* with duplicate entries.
3018	* @param fNoWait Set if it's a zero-wait poll call. Clear if
3019	* we'll wait for an event to occur.
3020	*
3021	* @remarks There is a potential race wrt duplicate handles when @a fNoWait is
3022	* @c true, we don't currently care about that oddity...
3023	*/
3024	DECLHIDDEN(uint32_t) rtSocketPollStart(RTSOCKET hSocket, RTPOLLSET hPollSet, uint32_t fEvents, bool fFinalEntry, bool fNoWait)
3025	{
3026	RTSOCKETINT *pThis = hSocket;
3027	AssertPtrReturn(pThis, UINT32_MAX);
3028	AssertReturn(pThis->u32Magic == RTSOCKET_MAGIC, UINT32_MAX);
3029	/** @todo This isn't quite sane. Replace by critsect and open up concurrent
3030	* reads and writes! */
3031	if (rtSocketTryLock(pThis))
3032	pThis->hPollSet = hPollSet;
3033	else
3034	{
3035	AssertReturn(pThis->hPollSet == hPollSet, UINT32_MAX);
3036	ASMAtomicIncU32(&pThis->cUsers);
3037	}
3038
3039	/* (rtSocketPollCheck will reset the event object). */
3040	# ifdef RT_OS_WINDOWS
3041	uint32_t fRetEvents = pThis->fEventsSaved;
3042	pThis->fEventsSaved = 0; /* Reset */
3043	fRetEvents \|= rtSocketPollCheck(pThis, fEvents);
3044
3045	if ( !fRetEvents
3046	&& !fNoWait)
3047	{
3048	pThis->fPollEvts \|= fEvents;
3049	if (fFinalEntry)
3050	{
3051	if (pThis->fSubscribedEvts != pThis->fPollEvts)
3052	{
3053	/** @todo seems like there might be a call to many here and that fPollEvts is
3054	* totally unnecessary... (bird) */
3055	int rc = rtSocketPollUpdateEvents(pThis, pThis->fPollEvts);
3056	if (RT_FAILURE(rc))
3057	{
3058	pThis->fPollEvts = 0;
3059	fRetEvents = UINT32_MAX;
3060	}
3061	}
3062
3063	/* Make sure we don't block when there are events pending relevant to an earlier poll set entry. */
3064	if (pThis->fEventsSaved && !pThis->fPollFallback && g_pfnWSASetEvent && fRetEvents == 0)
3065	g_pfnWSASetEvent(pThis->hEvent);
3066	}
3067	}
3068	# else
3069	uint32_t fRetEvents = rtSocketPollCheck(pThis, fEvents);
3070	# endif
3071
3072	if (fRetEvents \|\| fNoWait)
3073	{
3074	if (pThis->cUsers == 1)
3075	{
3076	# ifdef RT_OS_WINDOWS
3077	pThis->fEventsSaved &= RTPOLL_EVT_ERROR;
3078	pThis->fHarvestedEvents = false;
3079	rtSocketPollClearEventAndRestoreBlocking(pThis);
3080	# endif
3081	pThis->hPollSet = NIL_RTPOLLSET;
3082	}
3083	# ifdef RT_OS_WINDOWS
3084	else
3085	pThis->fHarvestedEvents = true;
3086	# endif
3087	ASMAtomicDecU32(&pThis->cUsers);
3088	}
3089	# ifdef RT_OS_WINDOWS
3090	/*
3091	* Kick the poller thread on if this is the final entry and we're in
3092	* winsock 1.x fallback mode.
3093	*/
3094	else if (pThis->fPollFallback && fFinalEntry)
3095	fRetEvents = rtSocketPollFallbackStart(pThis);
3096	# endif
3097
3098	return fRetEvents;
3099	}
3100
3101
3102	/**
3103	* Called after a WaitForMultipleObjects returned in order to check for pending
3104	* events and stop whatever actions that rtSocketPollStart() initiated.
3105	*
3106	* @returns Event mask or 0.
3107	*
3108	* @param hSocket The socket handle.
3109	* @param fEvents The events we're polling for.
3110	* @param fFinalEntry Set if this is the final entry for this handle
3111	* in this poll set. This can be used for dealing
3112	* with duplicate entries. Only keep in mind that
3113	* this method is called in reverse order, so the
3114	* first call will have this set (when the entire
3115	* set was processed).
3116	* @param fHarvestEvents Set if we should check for pending events.
3117	*/
3118	DECLHIDDEN(uint32_t) rtSocketPollDone(RTSOCKET hSocket, uint32_t fEvents, bool fFinalEntry, bool fHarvestEvents)
3119	{
3120	RTSOCKETINT *pThis = hSocket;
3121	AssertPtrReturn(pThis, 0);
3122	AssertReturn(pThis->u32Magic == RTSOCKET_MAGIC, 0);
3123	Assert(pThis->cUsers > 0);
3124	Assert(pThis->hPollSet != NIL_RTPOLLSET);
3125	RT_NOREF_PV(fFinalEntry);
3126
3127	# ifdef RT_OS_WINDOWS
3128	/*
3129	* Deactivate the poll thread if we're in winsock 1.x fallback poll mode.
3130	*/
3131	if ( pThis->fPollFallback
3132	&& pThis->hPollFallbackThread != NIL_RTTHREAD)
3133	{
3134	ASMAtomicWriteU32(&pThis->fSubscribedEvts, 0);
3135	if (ASMAtomicXchgBool(&pThis->fPollFallbackActive, false))
3136	rtSocketPokePollFallbackThread(pThis);
3137	}
3138	# endif
3139
3140	/*
3141	* Harvest events and clear the event mask for the next round of polling.
3142	*/
3143	uint32_t fRetEvents;
3144	# ifdef RT_OS_WINDOWS
3145	if (!pThis->fPollFallback)
3146	{
3147	if (!pThis->fHarvestedEvents)
3148	{
3149	fRetEvents = rtSocketPollCheck(pThis, fEvents);
3150	pThis->fHarvestedEvents = true;
3151	}
3152	else
3153	fRetEvents = pThis->fEventsSaved;
3154	if (fHarvestEvents)
3155	fRetEvents &= fEvents;
3156	else
3157	fRetEvents = 0;
3158	pThis->fPollEvts = 0;
3159	}
3160	else
3161	# endif
3162	{
3163	if (fHarvestEvents)
3164	fRetEvents = rtSocketPollCheck(pThis, fEvents);
3165	else
3166	fRetEvents = 0;
3167	}
3168
3169	/*
3170	* Make the socket blocking again and unlock the handle.
3171	*/
3172	if (pThis->cUsers == 1)
3173	{
3174	# ifdef RT_OS_WINDOWS
3175	pThis->fEventsSaved &= RTPOLL_EVT_ERROR;
3176	pThis->fHarvestedEvents = false;
3177	rtSocketPollClearEventAndRestoreBlocking(pThis);
3178	# endif
3179	pThis->hPollSet = NIL_RTPOLLSET;
3180	}
3181	ASMAtomicDecU32(&pThis->cUsers);
3182	return fRetEvents;
3183	}
3184
3185	#endif /* RT_OS_WINDOWS \|\| RT_OS_OS2 */
3186

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

source: vbox/trunk/src/VBox/Runtime/r3/socket.cpp@ 98032

Download in other formats: