socket.cpp@ 72285

Last change on this file since 72285 was 71950, checked in by vboxsync, 7 years ago
socket.cpp: Windows polling fixes for problem where we got stuck returning RTPOLL_EVT_READ.
Property svn:eol-style set to `native` Property svn:keywords set to `Id Revision`
File size: 93.2 KB

Line
1	/* $Id: socket.cpp 71950 2018-04-20 21:31:06Z vboxsync $ */
2	/** @file
3	* IPRT - Network Sockets.
4	*/
5
6	/*
7	* Copyright (C) 2006-2017 Oracle Corporation
8	*
9	* This file is part of VirtualBox Open Source Edition (OSE), as
10	* available from http://www.virtualbox.org. This file is free software;
11	* you can redistribute it and/or modify it under the terms of the GNU
12	* General Public License (GPL) as published by the Free Software
13	* Foundation, in version 2 as it comes in the "COPYING" file of the
14	* VirtualBox OSE distribution. VirtualBox OSE is distributed in the
15	* hope that it will be useful, but WITHOUT ANY WARRANTY of any kind.
16	*
17	* The contents of this file may alternatively be used under the terms
18	* of the Common Development and Distribution License Version 1.0
19	* (CDDL) only, as it comes in the "COPYING.CDDL" file of the
20	* VirtualBox OSE distribution, in which case the provisions of the
21	* CDDL are applicable instead of those of the GPL.
22	*
23	* You may elect to license modified versions of this file under the
24	* terms and conditions of either the GPL or the CDDL or both.
25	*/
26
27
28	/*********************************************************************************************************************************
29	* Header Files *
30	*********************************************************************************************************************************/
31	#ifdef RT_OS_WINDOWS
32	# include <iprt/win/winsock2.h>
33	# include <iprt/win/ws2tcpip.h>
34	#else /* !RT_OS_WINDOWS */
35	# include <errno.h>
36	# include <sys/select.h>
37	# include <sys/stat.h>
38	# include <sys/socket.h>
39	# include <netinet/in.h>
40	# include <netinet/tcp.h>
41	# include <arpa/inet.h>
42	# ifdef IPRT_WITH_TCPIP_V6
43	# include <netinet6/in6.h>
44	# endif
45	# include <sys/un.h>
46	# include <netdb.h>
47	# include <unistd.h>
48	# include <fcntl.h>
49	# include <sys/uio.h>
50	#endif /* !RT_OS_WINDOWS */
51	#include <limits.h>
52
53	#include "internal/iprt.h"
54	#include <iprt/socket.h>
55
56	#include <iprt/alloca.h>
57	#include <iprt/asm.h>
58	#include <iprt/assert.h>
59	#include <iprt/ctype.h>
60	#include <iprt/err.h>
61	#include <iprt/mempool.h>
62	#include <iprt/poll.h>
63	#include <iprt/string.h>
64	#include <iprt/thread.h>
65	#include <iprt/time.h>
66	#include <iprt/mem.h>
67	#include <iprt/sg.h>
68	#include <iprt/log.h>
69
70	#include "internal/magics.h"
71	#include "internal/socket.h"
72	#include "internal/string.h"
73	#ifdef RT_OS_WINDOWS
74	# include "win/internal-r3-win.h"
75	#endif
76
77
78	/*********************************************************************************************************************************
79	* Defined Constants And Macros *
80	*********************************************************************************************************************************/
81	/* non-standard linux stuff (it seems). */
82	#ifndef MSG_NOSIGNAL
83	# define MSG_NOSIGNAL 0
84	#endif
85
86	/* Windows has different names for SHUT_XXX. */
87	#ifndef SHUT_RDWR
88	# ifdef SD_BOTH
89	# define SHUT_RDWR SD_BOTH
90	# else
91	# define SHUT_RDWR 2
92	# endif
93	#endif
94	#ifndef SHUT_WR
95	# ifdef SD_SEND
96	# define SHUT_WR SD_SEND
97	# else
98	# define SHUT_WR 1
99	# endif
100	#endif
101	#ifndef SHUT_RD
102	# ifdef SD_RECEIVE
103	# define SHUT_RD SD_RECEIVE
104	# else
105	# define SHUT_RD 0
106	# endif
107	#endif
108
109	/* fixup backlevel OSes. */
110	#if defined(RT_OS_OS2) \|\| defined(RT_OS_WINDOWS)
111	# define socklen_t int
112	#endif
113
114	/** How many pending connection. */
115	#define RTTCP_SERVER_BACKLOG 10
116
117	/* Limit read and write sizes on Windows and OS/2. */
118	#ifdef RT_OS_WINDOWS
119	# define RTSOCKET_MAX_WRITE (INT_MAX / 2)
120	# define RTSOCKET_MAX_READ (INT_MAX / 2)
121	#elif defined(RT_OS_OS2)
122	# define RTSOCKET_MAX_WRITE 0x10000
123	# define RTSOCKET_MAX_READ 0x10000
124	#endif
125
126
127	/*********************************************************************************************************************************
128	* Structures and Typedefs *
129	*********************************************************************************************************************************/
130	/**
131	* Socket handle data.
132	*
133	* This is mainly required for implementing RTPollSet on Windows.
134	*/
135	typedef struct RTSOCKETINT
136	{
137	/** Magic number (RTSOCKET_MAGIC). */
138	uint32_t u32Magic;
139	/** Exclusive user count.
140	* This is used to prevent two threads from accessing the handle concurrently.
141	* It can be higher than 1 if this handle is reference multiple times in a
142	* polling set (Windows). */
143	uint32_t volatile cUsers;
144	/** The native socket handle. */
145	RTSOCKETNATIVE hNative;
146	/** Indicates whether the handle has been closed or not. */
147	bool volatile fClosed;
148	/** Indicates whether the socket is operating in blocking or non-blocking mode
149	* currently. */
150	bool fBlocking;
151	#if defined(RT_OS_WINDOWS) \|\| defined(RT_OS_OS2)
152	/** The pollset currently polling this socket. This is NIL if no one is
153	* polling. */
154	RTPOLLSET hPollSet;
155	#endif
156	#ifdef RT_OS_WINDOWS
157	/** The event semaphore we've associated with the socket handle.
158	* This is WSA_INVALID_EVENT if not done. */
159	WSAEVENT hEvent;
160	/** The events we're polling for. */
161	uint32_t fPollEvts;
162	/** The events we're currently subscribing to with WSAEventSelect.
163	* This is ZERO if we're currently not subscribing to anything. */
164	uint32_t fSubscribedEvts;
165	/** Saved events which are only posted once and events harvested for
166	* sockets entered multiple times into to a poll set. Imagine a scenario where
167	* you have a RTPOLL_EVT_READ entry and RTPOLL_EVT_ERROR entry. The READ
168	* condition can be triggered between checking the READ entry and the ERROR
169	* entry, and we don't want to drop the READ, so we store it here and make sure
170	* the event is signalled.
171	*
172	* The RTPOLL_EVT_ERROR is inconsistenly sticky at the momemnt... */
173	uint32_t fEventsSaved;
174	/** Set if fEventsSaved contains harvested events (used to avoid multiple
175	* calls to rtSocketPollCheck on the same socket during rtSocketPollDone). */
176	bool fHarvestedEvents;
177	/** Set if we're using the polling fallback. */
178	bool fPollFallback;
179	/** Set if the fallback polling is active (event not set). */
180	bool volatile fPollFallbackActive;
181	/** Set to shut down the fallback polling thread. */
182	bool volatile fPollFallbackShutdown;
183	/** Socket use to wake up the select thread. */
184	RTSOCKETNATIVE hPollFallbackNotifyW;
185	/** Socket the select thread always waits on. */
186	RTSOCKETNATIVE hPollFallbackNotifyR;
187	/** The fallback polling thread. */
188	RTTHREAD hPollFallbackThread;
189	#endif /* RT_OS_WINDOWS */
190	} RTSOCKETINT;
191
192
193	/**
194	* Address union used internally for things like getpeername and getsockname.
195	*/
196	typedef union RTSOCKADDRUNION
197	{
198	struct sockaddr Addr;
199	struct sockaddr_in IPv4;
200	#ifdef IPRT_WITH_TCPIP_V6
201	struct sockaddr_in6 IPv6;
202	#endif
203	} RTSOCKADDRUNION;
204
205
206	/*********************************************************************************************************************************
207	* Global Variables *
208	*********************************************************************************************************************************/
209	#ifdef RT_OS_WINDOWS
210	/** Indicates that we've successfully initialized winsock. */
211	static uint32_t volatile g_uWinSockInitedVersion = 0;
212	#endif
213
214
215	/*********************************************************************************************************************************
216	* Internal Functions *
217	*********************************************************************************************************************************/
218	#ifdef RT_OS_WINDOWS
219	static void rtSocketPokePollFallbackThread(RTSOCKETINT *pThis);
220	#endif
221
222
223
224	#ifdef RT_OS_WINDOWS
225	/**
226	* Initializes winsock for the process.
227	*
228	* @returns IPRT status code.
229	*/
230	static int rtSocketInitWinsock(void)
231	{
232	if (g_uWinSockInitedVersion != 0)
233	return VINF_SUCCESS;
234
235	if ( !g_pfnWSAGetLastError
236	\|\| !g_pfnWSAStartup
237	\|\| !g_pfnsocket
238	\|\| !g_pfnclosesocket)
239	return VERR_NET_INIT_FAILED;
240
241	/*
242	* Initialize winsock. Try with 2.2 and back down till we get something that works.
243	*/
244	static const WORD s_awVersions[] =
245	{
246	MAKEWORD(2, 2),
247	MAKEWORD(2, 1),
248	MAKEWORD(2, 0),
249	MAKEWORD(1, 1),
250	MAKEWORD(1, 0),
251	};
252	for (uint32_t i = 0; i < RT_ELEMENTS(s_awVersions); i++)
253	{
254	WSADATA wsaData;
255	RT_ZERO(wsaData);
256	int rcWsa = g_pfnWSAStartup(s_awVersions[i], &wsaData);
257	if (rcWsa == 0)
258	{
259	/* AssertMsg(wsaData.wVersion >= s_awVersions[i]); - triggers with winsock 1.1 */
260	ASMAtomicWriteU32(&g_uWinSockInitedVersion, wsaData.wVersion);
261	return VINF_SUCCESS;
262	}
263	AssertLogRelMsg(rcWsa == WSAVERNOTSUPPORTED, ("rcWsa=%d (winsock version %#x)\n", rcWsa, s_awVersions[i]));
264	}
265	LogRel(("Failed to init winsock!\n"));
266	return VERR_NET_INIT_FAILED;
267	}
268	#endif
269
270
271	/**
272	* Get the last error as an iprt status code.
273	*
274	* @returns IPRT status code.
275	*/
276	DECLINLINE(int) rtSocketError(void)
277	{
278	#ifdef RT_OS_WINDOWS
279	if (g_pfnWSAGetLastError)
280	return RTErrConvertFromWin32(g_pfnWSAGetLastError());
281	return VERR_NET_IO_ERROR;
282	#else
283	return RTErrConvertFromErrno(errno);
284	#endif
285	}
286
287
288	/**
289	* Resets the last error.
290	*/
291	DECLINLINE(void) rtSocketErrorReset(void)
292	{
293	#ifdef RT_OS_WINDOWS
294	if (g_pfnWSASetLastError)
295	g_pfnWSASetLastError(0);
296	#else
297	errno = 0;
298	#endif
299	}
300
301
302	/**
303	* Get the last resolver error as an iprt status code.
304	*
305	* @returns iprt status code.
306	*/
307	DECLHIDDEN(int) rtSocketResolverError(void)
308	{
309	#ifdef RT_OS_WINDOWS
310	if (g_pfnWSAGetLastError)
311	return RTErrConvertFromWin32(g_pfnWSAGetLastError());
312	return VERR_UNRESOLVED_ERROR;
313	#else
314	switch (h_errno)
315	{
316	case HOST_NOT_FOUND:
317	return VERR_NET_HOST_NOT_FOUND;
318	case NO_DATA:
319	return VERR_NET_ADDRESS_NOT_AVAILABLE;
320	case NO_RECOVERY:
321	return VERR_IO_GEN_FAILURE;
322	case TRY_AGAIN:
323	return VERR_TRY_AGAIN;
324
325	default:
326	AssertLogRelMsgFailed(("Unhandled error %u\n", h_errno));
327	return VERR_UNRESOLVED_ERROR;
328	}
329	#endif
330	}
331
332
333	/**
334	* Converts from a native socket address to a generic IPRT network address.
335	*
336	* @returns IPRT status code.
337	* @param pSrc The source address.
338	* @param cbSrc The size of the source address.
339	* @param pAddr Where to return the generic IPRT network
340	* address.
341	*/
342	static int rtSocketNetAddrFromAddr(RTSOCKADDRUNION const *pSrc, size_t cbSrc, PRTNETADDR pAddr)
343	{
344	/*
345	* Convert the address.
346	*/
347	if ( cbSrc == sizeof(struct sockaddr_in)
348	&& pSrc->Addr.sa_family == AF_INET)
349	{
350	RT_ZERO(*pAddr);
351	pAddr->enmType = RTNETADDRTYPE_IPV4;
352	pAddr->uPort = RT_N2H_U16(pSrc->IPv4.sin_port);
353	pAddr->uAddr.IPv4.u = pSrc->IPv4.sin_addr.s_addr;
354	}
355	#ifdef IPRT_WITH_TCPIP_V6
356	else if ( cbSrc == sizeof(struct sockaddr_in6)
357	&& pSrc->Addr.sa_family == AF_INET6)
358	{
359	RT_ZERO(*pAddr);
360	pAddr->enmType = RTNETADDRTYPE_IPV6;
361	pAddr->uPort = RT_N2H_U16(pSrc->IPv6.sin6_port);
362	pAddr->uAddr.IPv6.au32[0] = pSrc->IPv6.sin6_addr.s6_addr32[0];
363	pAddr->uAddr.IPv6.au32[1] = pSrc->IPv6.sin6_addr.s6_addr32[1];
364	pAddr->uAddr.IPv6.au32[2] = pSrc->IPv6.sin6_addr.s6_addr32[2];
365	pAddr->uAddr.IPv6.au32[3] = pSrc->IPv6.sin6_addr.s6_addr32[3];
366	}
367	#endif
368	else
369	return VERR_NET_ADDRESS_FAMILY_NOT_SUPPORTED;
370	return VINF_SUCCESS;
371	}
372
373
374	/**
375	* Converts from a generic IPRT network address to a native socket address.
376	*
377	* @returns IPRT status code.
378	* @param pAddr Pointer to the generic IPRT network address.
379	* @param pDst The source address.
380	* @param cbDst The size of the source address.
381	* @param pcbAddr Where to store the size of the returned address.
382	* Optional
383	*/
384	static int rtSocketAddrFromNetAddr(PCRTNETADDR pAddr, RTSOCKADDRUNION pDst, size_t cbDst, int pcbAddr)
385	{
386	RT_BZERO(pDst, cbDst);
387	if ( pAddr->enmType == RTNETADDRTYPE_IPV4
388	&& cbDst >= sizeof(struct sockaddr_in))
389	{
390	pDst->Addr.sa_family = AF_INET;
391	pDst->IPv4.sin_port = RT_H2N_U16(pAddr->uPort);
392	pDst->IPv4.sin_addr.s_addr = pAddr->uAddr.IPv4.u;
393	if (pcbAddr)
394	*pcbAddr = sizeof(pDst->IPv4);
395	}
396	#ifdef IPRT_WITH_TCPIP_V6
397	else if ( pAddr->enmType == RTNETADDRTYPE_IPV6
398	&& cbDst >= sizeof(struct sockaddr_in6))
399	{
400	pDst->Addr.sa_family = AF_INET6;
401	pDst->IPv6.sin6_port = RT_H2N_U16(pAddr->uPort);
402	pSrc->IPv6.sin6_addr.s6_addr32[0] = pAddr->uAddr.IPv6.au32[0];
403	pSrc->IPv6.sin6_addr.s6_addr32[1] = pAddr->uAddr.IPv6.au32[1];
404	pSrc->IPv6.sin6_addr.s6_addr32[2] = pAddr->uAddr.IPv6.au32[2];
405	pSrc->IPv6.sin6_addr.s6_addr32[3] = pAddr->uAddr.IPv6.au32[3];
406	if (pcbAddr)
407	*pcbAddr = sizeof(pDst->IPv6);
408	}
409	#endif
410	else
411	return VERR_NET_ADDRESS_FAMILY_NOT_SUPPORTED;
412	return VINF_SUCCESS;
413	}
414
415
416	/**
417	* Tries to lock the socket for exclusive usage by the calling thread.
418	*
419	* Call rtSocketUnlock() to unlock.
420	*
421	* @returns @c true if locked, @c false if not.
422	* @param pThis The socket structure.
423	*/
424	DECLINLINE(bool) rtSocketTryLock(RTSOCKETINT *pThis)
425	{
426	return ASMAtomicCmpXchgU32(&pThis->cUsers, 1, 0);
427	}
428
429
430	/**
431	* Unlocks the socket.
432	*
433	* @param pThis The socket structure.
434	*/
435	DECLINLINE(void) rtSocketUnlock(RTSOCKETINT *pThis)
436	{
437	ASMAtomicCmpXchgU32(&pThis->cUsers, 0, 1);
438	}
439
440
441	/**
442	* The slow path of rtSocketSwitchBlockingMode that does the actual switching.
443	*
444	* @returns IPRT status code.
445	* @param pThis The socket structure.
446	* @param fBlocking The desired mode of operation.
447	* @remarks Do not call directly.
448	*/
449	static int rtSocketSwitchBlockingModeSlow(RTSOCKETINT *pThis, bool fBlocking)
450	{
451	#ifdef RT_OS_WINDOWS
452	AssertReturn(g_pfnioctlsocket, VERR_NET_NOT_UNSUPPORTED);
453	u_long uBlocking = fBlocking ? 0 : 1;
454	if (g_pfnioctlsocket(pThis->hNative, FIONBIO, &uBlocking))
455	return rtSocketError();
456
457	#else
458	int fFlags = fcntl(pThis->hNative, F_GETFL, 0);
459	if (fFlags == -1)
460	return rtSocketError();
461
462	if (fBlocking)
463	fFlags &= ~O_NONBLOCK;
464	else
465	fFlags \|= O_NONBLOCK;
466	if (fcntl(pThis->hNative, F_SETFL, fFlags) == -1)
467	return rtSocketError();
468	#endif
469
470	pThis->fBlocking = fBlocking;
471	return VINF_SUCCESS;
472	}
473
474
475	/**
476	* Switches the socket to the desired blocking mode if necessary.
477	*
478	* The socket must be locked.
479	*
480	* @returns IPRT status code.
481	* @param pThis The socket structure.
482	* @param fBlocking The desired mode of operation.
483	*/
484	DECLINLINE(int) rtSocketSwitchBlockingMode(RTSOCKETINT *pThis, bool fBlocking)
485	{
486	if (pThis->fBlocking != fBlocking)
487	return rtSocketSwitchBlockingModeSlow(pThis, fBlocking);
488	return VINF_SUCCESS;
489	}
490
491
492	/**
493	* Creates an IPRT socket handle for a native one.
494	*
495	* @returns IPRT status code.
496	* @param ppSocket Where to return the IPRT socket handle.
497	* @param hNative The native handle.
498	*/
499	DECLHIDDEN(int) rtSocketCreateForNative(RTSOCKETINT **ppSocket, RTSOCKETNATIVE hNative)
500	{
501	RTSOCKETINT pThis = (RTSOCKETINT )RTMemPoolAlloc(RTMEMPOOL_DEFAULT, sizeof(*pThis));
502	if (!pThis)
503	return VERR_NO_MEMORY;
504	pThis->u32Magic = RTSOCKET_MAGIC;
505	pThis->cUsers = 0;
506	pThis->hNative = hNative;
507	pThis->fClosed = false;
508	pThis->fBlocking = true;
509	#if defined(RT_OS_WINDOWS) \|\| defined(RT_OS_OS2)
510	pThis->hPollSet = NIL_RTPOLLSET;
511	#endif
512	#ifdef RT_OS_WINDOWS
513	pThis->hEvent = WSA_INVALID_EVENT;
514	pThis->fPollEvts = 0;
515	pThis->fSubscribedEvts = 0;
516	pThis->fEventsSaved = 0;
517	pThis->fHarvestedEvents = false;
518	pThis->fPollFallback = g_uWinSockInitedVersion < MAKEWORD(2, 0)
519	\|\| g_pfnWSACreateEvent == NULL
520	\|\| g_pfnWSACloseEvent == NULL
521	\|\| g_pfnWSAEventSelect == NULL
522	\|\| g_pfnWSAEnumNetworkEvents == NULL;
523	pThis->fPollFallbackActive = false;
524	pThis->fPollFallbackShutdown = false;
525	pThis->hPollFallbackNotifyR = NIL_RTSOCKETNATIVE;
526	pThis->hPollFallbackNotifyW = NIL_RTSOCKETNATIVE;
527	pThis->hPollFallbackThread = NIL_RTTHREAD;
528	#endif
529	*ppSocket = pThis;
530	return VINF_SUCCESS;
531	}
532
533
534	RTDECL(int) RTSocketFromNative(PRTSOCKET phSocket, RTHCINTPTR uNative)
535	{
536	AssertReturn(uNative != NIL_RTSOCKETNATIVE, VERR_INVALID_PARAMETER);
537	#ifndef RT_OS_WINDOWS
538	AssertReturn(uNative >= 0, VERR_INVALID_PARAMETER);
539	#endif
540	AssertPtrReturn(phSocket, VERR_INVALID_POINTER);
541	return rtSocketCreateForNative(phSocket, uNative);
542	}
543
544
545	/**
546	* Wrapper around socket().
547	*
548	* @returns IPRT status code.
549	* @param phSocket Where to store the handle to the socket on
550	* success.
551	* @param iDomain The protocol family (PF_XXX).
552	* @param iType The socket type (SOCK_XXX).
553	* @param iProtocol Socket parameter, usually 0.
554	*/
555	DECLHIDDEN(int) rtSocketCreate(PRTSOCKET phSocket, int iDomain, int iType, int iProtocol)
556	{
557	#ifdef RT_OS_WINDOWS
558	AssertReturn(g_pfnsocket, VERR_NET_NOT_UNSUPPORTED);
559	AssertReturn(g_pfnclosesocket, VERR_NET_NOT_UNSUPPORTED);
560
561	/* Initialize WinSock. */
562	int rc2 = rtSocketInitWinsock();
563	if (RT_FAILURE(rc2))
564	return rc2;
565	#endif
566
567	/*
568	* Create the socket.
569	*/
570	#ifdef RT_OS_WINDOWS
571	RTSOCKETNATIVE hNative = g_pfnsocket(iDomain, iType, iProtocol);
572	#else
573	RTSOCKETNATIVE hNative = socket(iDomain, iType, iProtocol);
574	#endif
575	if (hNative == NIL_RTSOCKETNATIVE)
576	return rtSocketError();
577
578	/*
579	* Wrap it.
580	*/
581	int rc = rtSocketCreateForNative(phSocket, hNative);
582	if (RT_FAILURE(rc))
583	{
584	#ifdef RT_OS_WINDOWS
585	g_pfnclosesocket(hNative);
586	#else
587	close(hNative);
588	#endif
589	}
590	return rc;
591	}
592
593
594	/**
595	* Wrapper around socketpair() for creating a local TCP connection.
596	*
597	* @returns IPRT status code.
598	* @param phServer Where to return the first native socket.
599	* @param phClient Where to return the second native socket.
600	*/
601	static int rtSocketCreateNativeTcpPair(RTSOCKETNATIVE phServer, RTSOCKETNATIVE phClient)
602	{
603	#ifdef RT_OS_WINDOWS
604	/*
605	* Initialize WinSock and make sure we got the necessary APIs.
606	*/
607	int rc = rtSocketInitWinsock();
608	if (RT_FAILURE(rc))
609	return rc;
610	AssertReturn(g_pfnsocket, VERR_NET_NOT_UNSUPPORTED);
611	AssertReturn(g_pfnclosesocket, VERR_NET_NOT_UNSUPPORTED);
612	AssertReturn(g_pfnsetsockopt, VERR_NET_NOT_UNSUPPORTED);
613	AssertReturn(g_pfnbind, VERR_NET_NOT_UNSUPPORTED);
614	AssertReturn(g_pfngetsockname, VERR_NET_NOT_UNSUPPORTED);
615	AssertReturn(g_pfnlisten, VERR_NET_NOT_UNSUPPORTED);
616	AssertReturn(g_pfnaccept, VERR_NET_NOT_UNSUPPORTED);
617	AssertReturn(g_pfnconnect, VERR_NET_NOT_UNSUPPORTED);
618
619	/*
620	* Create the "server" listen socket and the "client" socket.
621	*/
622	RTSOCKETNATIVE hListener = g_pfnsocket(AF_INET, SOCK_STREAM, IPPROTO_TCP);
623	if (hListener == NIL_RTSOCKETNATIVE)
624	return rtSocketError();
625	RTSOCKETNATIVE hClient = g_pfnsocket(AF_INET, SOCK_STREAM, IPPROTO_TCP);
626	if (hClient != NIL_RTSOCKETNATIVE)
627	{
628
629	/*
630	* We let WinSock choose a port number when we bind.
631	*/
632	union
633	{
634	struct sockaddr_in Ip;
635	struct sockaddr Generic;
636	} uAddr;
637	RT_ZERO(uAddr);
638	uAddr.Ip.sin_family = AF_INET;
639	uAddr.Ip.sin_addr.s_addr = RT_H2N_U32_C(INADDR_LOOPBACK);
640	//uAddr.Ip.sin_port = 0;
641	int fReuse = 1;
642	rc = g_pfnsetsockopt(hListener, SOL_SOCKET, SO_REUSEADDR, (const char *)&fReuse, sizeof(fReuse));
643	if (rc == 0)
644	{
645	rc = g_pfnbind(hListener, &uAddr.Generic, sizeof(uAddr.Ip));
646	if (rc == 0)
647	{
648	/*
649	* Get the address the client should connect to. According to the docs,
650	* we cannot assume that getsockname sets the IP and family.
651	*/
652	RT_ZERO(uAddr);
653	int cbAddr = sizeof(uAddr.Ip);
654	rc = g_pfngetsockname(hListener, &uAddr.Generic, &cbAddr);
655	if (rc == 0)
656	{
657	uAddr.Ip.sin_family = AF_INET;
658	uAddr.Ip.sin_addr.s_addr = RT_H2N_U32_C(INADDR_LOOPBACK);
659
660	/*
661	* Listen, connect and accept.
662	*/
663	rc = g_pfnlisten(hListener, 1 /cBacklog/);
664	if (rc == 0)
665	{
666	rc = g_pfnconnect(hClient, &uAddr.Generic, sizeof(uAddr.Ip));
667	if (rc == 0)
668	{
669	RTSOCKETNATIVE hServer = g_pfnaccept(hListener, NULL, NULL);
670	if (hServer != NIL_RTSOCKETNATIVE)
671	{
672	g_pfnclosesocket(hListener);
673
674	/*
675	* Done!
676	*/
677	*phServer = hServer;
678	*phClient = hClient;
679	return VINF_SUCCESS;
680	}
681	}
682	}
683	}
684	}
685	}
686	rc = rtSocketError();
687	g_pfnclosesocket(hClient);
688	}
689	else
690	rc = rtSocketError();
691	g_pfnclosesocket(hListener);
692	return rc;
693
694	#else
695	/*
696	* Got socket pair, so use it.
697	* Note! This isn't TCP per se, but it should fool the users.
698	*/
699	int aSockets[2] = { -1, -1 };
700	if (socketpair(AF_LOCAL, SOCK_STREAM, 0, aSockets) == 0)
701	{
702	*phServer = aSockets[0];
703	*phClient = aSockets[1];
704	return VINF_SUCCESS;
705	}
706	return rtSocketError();
707	#endif
708	}
709
710
711	/**
712	* Worker for RTTcpCreatePair.
713	*
714	* @returns IPRT status code.
715	* @param phServer Where to return the "server" side of the pair.
716	* @param phClient Where to return the "client" side of the pair.
717	* @note There is no server or client side, but we gotta call it something.
718	*/
719	DECLHIDDEN(int) rtSocketCreateTcpPair(RTSOCKET phServer, RTSOCKET phClient)
720	{
721	RTSOCKETNATIVE hServer = NIL_RTSOCKETNATIVE;
722	RTSOCKETNATIVE hClient = NIL_RTSOCKETNATIVE;
723	int rc = rtSocketCreateNativeTcpPair(&hServer, &hClient);
724	if (RT_SUCCESS(rc))
725	{
726	rc = rtSocketCreateForNative(phServer, hServer);
727	if (RT_SUCCESS(rc))
728	{
729	rc = rtSocketCreateForNative(phClient, hClient);
730	if (RT_SUCCESS(rc))
731	return VINF_SUCCESS;
732	RTSocketRelease(*phServer);
733	}
734	else
735	{
736	#ifdef RT_OS_WINDOWS
737	g_pfnclosesocket(hServer);
738	#else
739	close(hServer);
740	#endif
741	}
742	#ifdef RT_OS_WINDOWS
743	g_pfnclosesocket(hClient);
744	#else
745	close(hClient);
746	#endif
747	}
748
749	*phServer = NIL_RTSOCKET;
750	*phClient = NIL_RTSOCKET;
751	return rc;
752	}
753
754
755	RTDECL(uint32_t) RTSocketRetain(RTSOCKET hSocket)
756	{
757	RTSOCKETINT *pThis = hSocket;
758	AssertPtrReturn(pThis, UINT32_MAX);
759	AssertReturn(pThis->u32Magic == RTSOCKET_MAGIC, UINT32_MAX);
760	return RTMemPoolRetain(pThis);
761	}
762
763
764	/**
765	* Worker for RTSocketRelease and RTSocketClose.
766	*
767	* @returns IPRT status code.
768	* @param pThis The socket handle instance data.
769	* @param fDestroy Whether we're reaching ref count zero.
770	*/
771	static int rtSocketCloseIt(RTSOCKETINT *pThis, bool fDestroy)
772	{
773	/*
774	* Invalidate the handle structure on destroy.
775	*/
776	if (fDestroy)
777	{
778	Assert(ASMAtomicReadU32(&pThis->u32Magic) == RTSOCKET_MAGIC);
779	ASMAtomicWriteU32(&pThis->u32Magic, RTSOCKET_MAGIC_DEAD);
780	}
781
782	int rc = VINF_SUCCESS;
783	if (ASMAtomicCmpXchgBool(&pThis->fClosed, true, false))
784	{
785	#ifdef RT_OS_WINDOWS
786	/*
787	* Poke the polling thread if active and give it a small chance to stop.
788	*/
789	if ( pThis->fPollFallback
790	&& pThis->hPollFallbackThread != NIL_RTTHREAD)
791	{
792	ASMAtomicWriteBool(&pThis->fPollFallbackShutdown, true);
793	rtSocketPokePollFallbackThread(pThis);
794	int rc2 = RTThreadWait(pThis->hPollFallbackThread, RT_MS_1SEC, NULL);
795	if (RT_SUCCESS(rc2))
796	pThis->hPollFallbackThread = NIL_RTTHREAD;
797	}
798	#endif
799
800	/*
801	* Close the native handle.
802	*/
803	RTSOCKETNATIVE hNative = pThis->hNative;
804	if (hNative != NIL_RTSOCKETNATIVE)
805	{
806	pThis->hNative = NIL_RTSOCKETNATIVE;
807
808	#ifdef RT_OS_WINDOWS
809	AssertReturn(g_pfnclosesocket, VERR_NET_NOT_UNSUPPORTED);
810	if (g_pfnclosesocket(hNative))
811	#else
812	if (close(hNative))
813	#endif
814	{
815	rc = rtSocketError();
816	#ifdef RT_OS_WINDOWS
817	AssertMsgFailed(("closesocket(%p) -> %Rrc\n", (uintptr_t)hNative, rc));
818	#else
819	AssertMsgFailed(("close(%d) -> %Rrc\n", hNative, rc));
820	#endif
821	}
822	}
823
824	#ifdef RT_OS_WINDOWS
825	/*
826	* Windows specific polling cleanup.
827	*/
828	WSAEVENT hEvent = pThis->hEvent;
829	if (hEvent != WSA_INVALID_EVENT)
830	{
831	pThis->hEvent = WSA_INVALID_EVENT;
832	if (!pThis->fPollFallback)
833	{
834	Assert(g_pfnWSACloseEvent);
835	if (g_pfnWSACloseEvent)
836	g_pfnWSACloseEvent(hEvent);
837	}
838	else
839	CloseHandle(hEvent);
840	}
841
842	if (pThis->fPollFallback)
843	{
844	if (pThis->hPollFallbackNotifyW != NIL_RTSOCKETNATIVE)
845	{
846	g_pfnclosesocket(pThis->hPollFallbackNotifyW);
847	pThis->hPollFallbackNotifyW = NIL_RTSOCKETNATIVE;
848	}
849
850	if (pThis->hPollFallbackThread != NIL_RTTHREAD)
851	{
852	int rc2 = RTThreadWait(pThis->hPollFallbackThread, RT_MS_1MIN / 2, NULL);
853	AssertRC(rc2);
854	pThis->hPollFallbackThread = NIL_RTTHREAD;
855	}
856
857	if (pThis->hPollFallbackNotifyR != NIL_RTSOCKETNATIVE)
858	{
859	g_pfnclosesocket(pThis->hPollFallbackNotifyR);
860	pThis->hPollFallbackNotifyR = NIL_RTSOCKETNATIVE;
861	}
862	}
863	#endif
864	}
865
866	return rc;
867	}
868
869
870	RTDECL(uint32_t) RTSocketRelease(RTSOCKET hSocket)
871	{
872	RTSOCKETINT *pThis = hSocket;
873	if (pThis == NIL_RTSOCKET)
874	return 0;
875	AssertPtrReturn(pThis, UINT32_MAX);
876	AssertReturn(pThis->u32Magic == RTSOCKET_MAGIC, UINT32_MAX);
877
878	/* get the refcount without killing it... */
879	uint32_t cRefs = RTMemPoolRefCount(pThis);
880	AssertReturn(cRefs != UINT32_MAX, UINT32_MAX);
881	if (cRefs == 1)
882	rtSocketCloseIt(pThis, true);
883
884	return RTMemPoolRelease(RTMEMPOOL_DEFAULT, pThis);
885	}
886
887
888	RTDECL(int) RTSocketClose(RTSOCKET hSocket)
889	{
890	RTSOCKETINT *pThis = hSocket;
891	if (pThis == NIL_RTSOCKET)
892	return VINF_SUCCESS;
893	AssertPtrReturn(pThis, VERR_INVALID_HANDLE);
894	AssertReturn(pThis->u32Magic == RTSOCKET_MAGIC, VERR_INVALID_HANDLE);
895
896	uint32_t cRefs = RTMemPoolRefCount(pThis);
897	AssertReturn(cRefs != UINT32_MAX, UINT32_MAX);
898
899	int rc = rtSocketCloseIt(pThis, cRefs == 1);
900
901	RTMemPoolRelease(RTMEMPOOL_DEFAULT, pThis);
902	return rc;
903	}
904
905
906	RTDECL(RTHCUINTPTR) RTSocketToNative(RTSOCKET hSocket)
907	{
908	RTSOCKETINT *pThis = hSocket;
909	AssertPtrReturn(pThis, RTHCUINTPTR_MAX);
910	AssertReturn(pThis->u32Magic == RTSOCKET_MAGIC, RTHCUINTPTR_MAX);
911	return (RTHCUINTPTR)pThis->hNative;
912	}
913
914
915	RTDECL(int) RTSocketSetInheritance(RTSOCKET hSocket, bool fInheritable)
916	{
917	RTSOCKETINT *pThis = hSocket;
918	AssertPtrReturn(pThis, VERR_INVALID_HANDLE);
919	AssertReturn(pThis->u32Magic == RTSOCKET_MAGIC, VERR_INVALID_HANDLE);
920	AssertReturn(RTMemPoolRefCount(pThis) >= (pThis->cUsers ? 2U : 1U), VERR_CALLER_NO_REFERENCE);
921
922	int rc = VINF_SUCCESS;
923	#ifdef RT_OS_WINDOWS
924	if (!SetHandleInformation((HANDLE)pThis->hNative, HANDLE_FLAG_INHERIT, fInheritable ? HANDLE_FLAG_INHERIT : 0))
925	rc = RTErrConvertFromWin32(GetLastError());
926	#else
927	if (fcntl(pThis->hNative, F_SETFD, fInheritable ? 0 : FD_CLOEXEC) < 0)
928	rc = RTErrConvertFromErrno(errno);
929	#endif
930
931	return rc;
932	}
933
934
935	static bool rtSocketIsIPv4Numerical(const char *pszAddress, PRTNETADDRIPV4 pAddr)
936	{
937
938	/* Empty address resolves to the INADDR_ANY address (good for bind). */
939	if (!pszAddress \|\| !*pszAddress)
940	{
941	pAddr->u = INADDR_ANY;
942	return true;
943	}
944
945	/* Four quads? */
946	char psz = (char )pszAddress;
947	for (int i = 0; i < 4; i++)
948	{
949	uint8_t u8;
950	int rc = RTStrToUInt8Ex(psz, &psz, 0, &u8);
951	if (rc != VINF_SUCCESS && rc != VWRN_TRAILING_CHARS)
952	return false;
953	if (*psz != (i < 3 ? '.' : '\0'))
954	return false;
955	psz++;
956
957	pAddr->au8[i] = u8; /* big endian */
958	}
959
960	return true;
961	}
962
963	RTDECL(int) RTSocketParseInetAddress(const char *pszAddress, unsigned uPort, PRTNETADDR pAddr)
964	{
965	int rc;
966
967	/*
968	* Validate input.
969	*/
970	AssertReturn(uPort > 0, VERR_INVALID_PARAMETER);
971	AssertPtrNullReturn(pszAddress, VERR_INVALID_POINTER);
972
973	/*
974	* Resolve the address. Pretty crude at the moment, but we have to make
975	* sure to not ask the NT 4 gethostbyname about an IPv4 address as it may
976	* give a wrong answer.
977	*/
978	/** @todo this only supports IPv4, and IPv6 support needs to be added.
979	* It probably needs to be converted to getaddrinfo(). */
980	RTNETADDRIPV4 IPv4Quad;
981	if (rtSocketIsIPv4Numerical(pszAddress, &IPv4Quad))
982	{
983	Log3(("rtSocketIsIPv4Numerical: %s -> %#x (%RTnaipv4)\n", pszAddress, IPv4Quad.u, IPv4Quad));
984	RT_ZERO(*pAddr);
985	pAddr->enmType = RTNETADDRTYPE_IPV4;
986	pAddr->uPort = uPort;
987	pAddr->uAddr.IPv4 = IPv4Quad;
988	return VINF_SUCCESS;
989	}
990
991	#ifdef RT_OS_WINDOWS
992	/* Initialize WinSock and check version before we call gethostbyname. */
993	if (!g_pfngethostbyname)
994	return VERR_NET_NOT_UNSUPPORTED;
995
996	int rc2 = rtSocketInitWinsock();
997	if (RT_FAILURE(rc2))
998	return rc2;
999
1000	# define gethostbyname g_pfngethostbyname
1001	#endif
1002
1003	struct hostent *pHostEnt;
1004	pHostEnt = gethostbyname(pszAddress);
1005	if (!pHostEnt)
1006	{
1007	rc = rtSocketResolverError();
1008	AssertMsgFailed(("Could not resolve '%s', rc=%Rrc\n", pszAddress, rc));
1009	return rc;
1010	}
1011
1012	if (pHostEnt->h_addrtype == AF_INET)
1013	{
1014	RT_ZERO(*pAddr);
1015	pAddr->enmType = RTNETADDRTYPE_IPV4;
1016	pAddr->uPort = uPort;
1017	pAddr->uAddr.IPv4.u = ((struct in_addr *)pHostEnt->h_addr)->s_addr;
1018	Log3(("gethostbyname: %s -> %#x (%RTnaipv4)\n", pszAddress, pAddr->uAddr.IPv4.u, pAddr->uAddr.IPv4));
1019	}
1020	else
1021	return VERR_NET_ADDRESS_FAMILY_NOT_SUPPORTED;
1022
1023	#ifdef RT_OS_WINDOWS
1024	# undef gethostbyname
1025	#endif
1026	return VINF_SUCCESS;
1027	}
1028
1029
1030	/*
1031	* New function to allow both ipv4 and ipv6 addresses to be resolved.
1032	* Breaks compatibility with windows before 2000.
1033	*/
1034	RTDECL(int) RTSocketQueryAddressStr(const char pszHost, char pszResult, size_t *pcbResult, PRTNETADDRTYPE penmAddrType)
1035	{
1036	AssertPtrReturn(pszHost, VERR_INVALID_POINTER);
1037	AssertPtrReturn(pcbResult, VERR_INVALID_POINTER);
1038	AssertPtrNullReturn(penmAddrType, VERR_INVALID_POINTER);
1039	AssertPtrNullReturn(pszResult, VERR_INVALID_POINTER);
1040
1041	#if defined(RT_OS_OS2) \|\| defined(RT_OS_WINDOWS) /** @todo dynamically resolve the APIs not present in NT4! */
1042	return VERR_NOT_SUPPORTED;
1043
1044	#else
1045	int rc;
1046	if (*pcbResult < 16)
1047	return VERR_NET_ADDRESS_NOT_AVAILABLE;
1048
1049	/* Setup the hint. */
1050	struct addrinfo grHints;
1051	RT_ZERO(grHints);
1052	grHints.ai_socktype = 0;
1053	grHints.ai_flags = 0;
1054	grHints.ai_protocol = 0;
1055	grHints.ai_family = AF_UNSPEC;
1056	if (penmAddrType)
1057	{
1058	switch (*penmAddrType)
1059	{
1060	case RTNETADDRTYPE_INVALID:
1061	/grHints.ai_family = AF_UNSPEC;/
1062	break;
1063	case RTNETADDRTYPE_IPV4:
1064	grHints.ai_family = AF_INET;
1065	break;
1066	case RTNETADDRTYPE_IPV6:
1067	grHints.ai_family = AF_INET6;
1068	break;
1069	default:
1070	AssertFailedReturn(VERR_INVALID_PARAMETER);
1071	}
1072	}
1073
1074	# ifdef RT_OS_WINDOWS
1075	/*
1076	* Winsock2 init
1077	*/
1078	if ( !g_pfngetaddrinfo
1079	\|\| !g_pfnfreeaddrinfo)
1080	return VERR_NET_NOT_UNSUPPORTED;
1081
1082	int rc2 = rtSocketInitWinsock();
1083	if (RT_FAILURE(rc2))
1084	return rc2;
1085
1086	# define getaddrinfo g_pfngetaddrinfo
1087	# define freeaddrinfo g_pfnfreeaddrinfo
1088	# endif
1089
1090	/** @todo r=bird: getaddrinfo and freeaddrinfo breaks the additions on NT4. */
1091	struct addrinfo *pgrResults = NULL;
1092	rc = getaddrinfo(pszHost, "", &grHints, &pgrResults);
1093	if (rc != 0)
1094	return VERR_NET_ADDRESS_NOT_AVAILABLE;
1095
1096	// return data
1097	// on multiple matches return only the first one
1098
1099	if (!pgrResults)
1100	return VERR_NET_ADDRESS_NOT_AVAILABLE;
1101
1102	struct addrinfo const *pgrResult = pgrResults->ai_next;
1103	if (!pgrResult)
1104	{
1105	freeaddrinfo(pgrResults);
1106	return VERR_NET_ADDRESS_NOT_AVAILABLE;
1107	}
1108
1109	RTNETADDRTYPE enmAddrType = RTNETADDRTYPE_INVALID;
1110	size_t cchIpAddress;
1111	char szIpAddress[48];
1112	if (pgrResult->ai_family == AF_INET)
1113	{
1114	struct sockaddr_in const pgrSa = (struct sockaddr_in const )pgrResult->ai_addr;
1115	cchIpAddress = RTStrPrintf(szIpAddress, sizeof(szIpAddress),
1116	"%RTnaipv4", pgrSa->sin_addr.s_addr);
1117	Assert(cchIpAddress >= 7 && cchIpAddress < sizeof(szIpAddress) - 1);
1118	enmAddrType = RTNETADDRTYPE_IPV4;
1119	rc = VINF_SUCCESS;
1120	}
1121	else if (pgrResult->ai_family == AF_INET6)
1122	{
1123	struct sockaddr_in6 const pgrSa6 = (struct sockaddr_in6 const )pgrResult->ai_addr;
1124	cchIpAddress = RTStrPrintf(szIpAddress, sizeof(szIpAddress),
1125	"%RTnaipv6", (PRTNETADDRIPV6)&pgrSa6->sin6_addr);
1126	enmAddrType = RTNETADDRTYPE_IPV6;
1127	rc = VINF_SUCCESS;
1128	}
1129	else
1130	{
1131	rc = VERR_NET_ADDRESS_NOT_AVAILABLE;
1132	szIpAddress[0] = '\0';
1133	cchIpAddress = 0;
1134	}
1135	freeaddrinfo(pgrResults);
1136
1137	/*
1138	* Copy out the result.
1139	*/
1140	size_t const cbResult = *pcbResult;
1141	*pcbResult = cchIpAddress + 1;
1142	if (cchIpAddress < cbResult)
1143	memcpy(pszResult, szIpAddress, cchIpAddress + 1);
1144	else
1145	{
1146	RT_BZERO(pszResult, cbResult);
1147	if (RT_SUCCESS(rc))
1148	rc = VERR_BUFFER_OVERFLOW;
1149	}
1150	if (penmAddrType && RT_SUCCESS(rc))
1151	*penmAddrType = enmAddrType;
1152	return rc;
1153
1154	# ifdef RT_OS_WINDOWS
1155	# undef getaddrinfo
1156	# undef freeaddrinfo
1157	# endif
1158	#endif /* !RT_OS_OS2 */
1159	}
1160
1161
1162	RTDECL(int) RTSocketRead(RTSOCKET hSocket, void pvBuffer, size_t cbBuffer, size_t pcbRead)
1163	{
1164	/*
1165	* Validate input.
1166	*/
1167	RTSOCKETINT *pThis = hSocket;
1168	AssertPtrReturn(pThis, VERR_INVALID_HANDLE);
1169	AssertReturn(pThis->u32Magic == RTSOCKET_MAGIC, VERR_INVALID_HANDLE);
1170	AssertReturn(cbBuffer > 0, VERR_INVALID_PARAMETER);
1171	AssertPtr(pvBuffer);
1172	#ifdef RT_OS_WINDOWS
1173	AssertReturn(g_pfnrecv, VERR_NET_NOT_UNSUPPORTED);
1174	# define recv g_pfnrecv
1175	#endif
1176	AssertReturn(rtSocketTryLock(pThis), VERR_CONCURRENT_ACCESS);
1177
1178	int rc = rtSocketSwitchBlockingMode(pThis, true /* fBlocking */);
1179	if (RT_FAILURE(rc))
1180	return rc;
1181
1182	/*
1183	* Read loop.
1184	* If pcbRead is NULL we have to fill the entire buffer!
1185	*/
1186	size_t cbRead = 0;
1187	size_t cbToRead = cbBuffer;
1188	for (;;)
1189	{
1190	rtSocketErrorReset();
1191	#ifdef RTSOCKET_MAX_READ
1192	int cbNow = cbToRead >= RTSOCKET_MAX_READ ? RTSOCKET_MAX_READ : (int)cbToRead;
1193	#else
1194	size_t cbNow = cbToRead;
1195	#endif
1196	ssize_t cbBytesRead = recv(pThis->hNative, (char *)pvBuffer + cbRead, cbNow, MSG_NOSIGNAL);
1197	if (cbBytesRead <= 0)
1198	{
1199	rc = rtSocketError();
1200	Assert(RT_FAILURE_NP(rc) \|\| cbBytesRead == 0);
1201	if (RT_SUCCESS_NP(rc))
1202	{
1203	if (!pcbRead)
1204	rc = VERR_NET_SHUTDOWN;
1205	else
1206	{
1207	*pcbRead = 0;
1208	rc = VINF_SUCCESS;
1209	}
1210	}
1211	break;
1212	}
1213	if (pcbRead)
1214	{
1215	/* return partial data */
1216	*pcbRead = cbBytesRead;
1217	break;
1218	}
1219
1220	/* read more? */
1221	cbRead += cbBytesRead;
1222	if (cbRead == cbBuffer)
1223	break;
1224
1225	/* next */
1226	cbToRead = cbBuffer - cbRead;
1227	}
1228
1229	rtSocketUnlock(pThis);
1230	#ifdef RT_OS_WINDOWS
1231	# undef recv
1232	#endif
1233	return rc;
1234	}
1235
1236
1237	RTDECL(int) RTSocketReadFrom(RTSOCKET hSocket, void pvBuffer, size_t cbBuffer, size_t pcbRead, PRTNETADDR pSrcAddr)
1238	{
1239	/*
1240	* Validate input.
1241	*/
1242	RTSOCKETINT *pThis = hSocket;
1243	AssertPtrReturn(pThis, VERR_INVALID_HANDLE);
1244	AssertReturn(pThis->u32Magic == RTSOCKET_MAGIC, VERR_INVALID_HANDLE);
1245	AssertReturn(cbBuffer > 0, VERR_INVALID_PARAMETER);
1246	AssertPtr(pvBuffer);
1247	AssertPtr(pcbRead);
1248	#ifdef RT_OS_WINDOWS
1249	AssertReturn(g_pfnrecvfrom, VERR_NET_NOT_UNSUPPORTED);
1250	# define recvfrom g_pfnrecvfrom
1251	#endif
1252	AssertReturn(rtSocketTryLock(pThis), VERR_CONCURRENT_ACCESS);
1253
1254	int rc = rtSocketSwitchBlockingMode(pThis, true /* fBlocking */);
1255	if (RT_FAILURE(rc))
1256	return rc;
1257
1258	/*
1259	* Read data.
1260	*/
1261	size_t cbRead = 0;
1262	size_t cbToRead = cbBuffer;
1263	rtSocketErrorReset();
1264	RTSOCKADDRUNION u;
1265	#ifdef RTSOCKET_MAX_READ
1266	int cbNow = cbToRead >= RTSOCKET_MAX_READ ? RTSOCKET_MAX_READ : (int)cbToRead;
1267	int cbAddr = sizeof(u);
1268	#else
1269	size_t cbNow = cbToRead;
1270	socklen_t cbAddr = sizeof(u);
1271	#endif
1272	ssize_t cbBytesRead = recvfrom(pThis->hNative, (char *)pvBuffer + cbRead, cbNow, MSG_NOSIGNAL, &u.Addr, &cbAddr);
1273	if (cbBytesRead <= 0)
1274	{
1275	rc = rtSocketError();
1276	Assert(RT_FAILURE_NP(rc) \|\| cbBytesRead == 0);
1277	if (RT_SUCCESS_NP(rc))
1278	{
1279	*pcbRead = 0;
1280	rc = VINF_SUCCESS;
1281	}
1282	}
1283	else
1284	{
1285	if (pSrcAddr)
1286	rc = rtSocketNetAddrFromAddr(&u, cbAddr, pSrcAddr);
1287	*pcbRead = cbBytesRead;
1288	}
1289
1290	rtSocketUnlock(pThis);
1291	#ifdef RT_OS_WINDOWS
1292	# undef recvfrom
1293	#endif
1294	return rc;
1295	}
1296
1297
1298	RTDECL(int) RTSocketWrite(RTSOCKET hSocket, const void *pvBuffer, size_t cbBuffer)
1299	{
1300	/*
1301	* Validate input.
1302	*/
1303	RTSOCKETINT *pThis = hSocket;
1304	AssertPtrReturn(pThis, VERR_INVALID_HANDLE);
1305	AssertReturn(pThis->u32Magic == RTSOCKET_MAGIC, VERR_INVALID_HANDLE);
1306	#ifdef RT_OS_WINDOWS
1307	AssertReturn(g_pfnsend, VERR_NET_NOT_UNSUPPORTED);
1308	# define send g_pfnsend
1309	#endif
1310	AssertReturn(rtSocketTryLock(pThis), VERR_CONCURRENT_ACCESS);
1311
1312	int rc = rtSocketSwitchBlockingMode(pThis, true /* fBlocking */);
1313	if (RT_FAILURE(rc))
1314	return rc;
1315
1316	/*
1317	* Try write all at once.
1318	*/
1319	#ifdef RTSOCKET_MAX_WRITE
1320	int cbNow = cbBuffer >= RTSOCKET_MAX_WRITE ? RTSOCKET_MAX_WRITE : (int)cbBuffer;
1321	#else
1322	size_t cbNow = cbBuffer >= SSIZE_MAX ? SSIZE_MAX : cbBuffer;
1323	#endif
1324	ssize_t cbWritten = send(pThis->hNative, (const char *)pvBuffer, cbNow, MSG_NOSIGNAL);
1325	if (RT_LIKELY((size_t)cbWritten == cbBuffer && cbWritten >= 0))
1326	rc = VINF_SUCCESS;
1327	else if (cbWritten < 0)
1328	rc = rtSocketError();
1329	else
1330	{
1331	/*
1332	* Unfinished business, write the remainder of the request. Must ignore
1333	* VERR_INTERRUPTED here if we've managed to send something.
1334	*/
1335	size_t cbSentSoFar = 0;
1336	for (;;)
1337	{
1338	/* advance */
1339	cbBuffer -= (size_t)cbWritten;
1340	if (!cbBuffer)
1341	break;
1342	cbSentSoFar += (size_t)cbWritten;
1343	pvBuffer = (char const *)pvBuffer + cbWritten;
1344
1345	/* send */
1346	#ifdef RTSOCKET_MAX_WRITE
1347	cbNow = cbBuffer >= RTSOCKET_MAX_WRITE ? RTSOCKET_MAX_WRITE : (int)cbBuffer;
1348	#else
1349	cbNow = cbBuffer >= SSIZE_MAX ? SSIZE_MAX : cbBuffer;
1350	#endif
1351	cbWritten = send(pThis->hNative, (const char *)pvBuffer, cbNow, MSG_NOSIGNAL);
1352	if (cbWritten >= 0)
1353	AssertMsg(cbBuffer >= (size_t)cbWritten, ("Wrote more than we requested!!! cbWritten=%zu cbBuffer=%zu rtSocketError()=%d\n",
1354	cbWritten, cbBuffer, rtSocketError()));
1355	else
1356	{
1357	rc = rtSocketError();
1358	if (rc != VERR_INTERNAL_ERROR \|\| cbSentSoFar == 0)
1359	break;
1360	cbWritten = 0;
1361	rc = VINF_SUCCESS;
1362	}
1363	}
1364	}
1365
1366	rtSocketUnlock(pThis);
1367	#ifdef RT_OS_WINDOWS
1368	# undef send
1369	#endif
1370	return rc;
1371	}
1372
1373
1374	RTDECL(int) RTSocketWriteTo(RTSOCKET hSocket, const void *pvBuffer, size_t cbBuffer, PCRTNETADDR pAddr)
1375	{
1376	/*
1377	* Validate input.
1378	*/
1379	RTSOCKETINT *pThis = hSocket;
1380	AssertPtrReturn(pThis, VERR_INVALID_HANDLE);
1381	AssertReturn(pThis->u32Magic == RTSOCKET_MAGIC, VERR_INVALID_HANDLE);
1382	#ifdef RT_OS_WINDOWS
1383	AssertReturn(g_pfnsendto, VERR_NET_NOT_UNSUPPORTED);
1384	# define sendto g_pfnsendto
1385	#endif
1386
1387	/* no locking since UDP reads may be done concurrently to writes, and
1388	* this is the normal use case of this code. */
1389
1390	int rc = rtSocketSwitchBlockingMode(pThis, true /* fBlocking */);
1391	if (RT_FAILURE(rc))
1392	return rc;
1393
1394	/* Figure out destination address. */
1395	struct sockaddr *pSA = NULL;
1396	#ifdef RT_OS_WINDOWS
1397	int cbSA = 0;
1398	#else
1399	socklen_t cbSA = 0;
1400	#endif
1401	RTSOCKADDRUNION u;
1402	if (pAddr)
1403	{
1404	rc = rtSocketAddrFromNetAddr(pAddr, &u, sizeof(u), NULL);
1405	if (RT_FAILURE(rc))
1406	return rc;
1407	pSA = &u.Addr;
1408	cbSA = sizeof(u);
1409	}
1410
1411	/*
1412	* Must write all at once, otherwise it is a failure.
1413	*/
1414	#ifdef RT_OS_WINDOWS
1415	int cbNow = cbBuffer >= RTSOCKET_MAX_WRITE ? RTSOCKET_MAX_WRITE : (int)cbBuffer;
1416	#else
1417	size_t cbNow = cbBuffer >= SSIZE_MAX ? SSIZE_MAX : cbBuffer;
1418	#endif
1419	ssize_t cbWritten = sendto(pThis->hNative, (const char *)pvBuffer, cbNow, MSG_NOSIGNAL, pSA, cbSA);
1420	if (RT_LIKELY((size_t)cbWritten == cbBuffer && cbWritten >= 0))
1421	rc = VINF_SUCCESS;
1422	else if (cbWritten < 0)
1423	rc = rtSocketError();
1424	else
1425	rc = VERR_TOO_MUCH_DATA;
1426
1427	/// @todo rtSocketUnlock(pThis);
1428	#ifdef RT_OS_WINDOWS
1429	# undef sendto
1430	#endif
1431	return rc;
1432	}
1433
1434
1435	RTDECL(int) RTSocketWriteToNB(RTSOCKET hSocket, const void *pvBuffer, size_t cbBuffer, PCRTNETADDR pAddr)
1436	{
1437	/*
1438	* Validate input.
1439	*/
1440	RTSOCKETINT *pThis = hSocket;
1441	AssertPtrReturn(pThis, VERR_INVALID_HANDLE);
1442	AssertReturn(pThis->u32Magic == RTSOCKET_MAGIC, VERR_INVALID_HANDLE);
1443	#ifdef RT_OS_WINDOWS
1444	AssertReturn(g_pfnsendto, VERR_NET_NOT_UNSUPPORTED);
1445	# define sendto g_pfnsendto
1446	#endif
1447
1448	/* no locking since UDP reads may be done concurrently to writes, and
1449	* this is the normal use case of this code. */
1450
1451	int rc = rtSocketSwitchBlockingMode(pThis, false /* fBlocking */);
1452	if (RT_FAILURE(rc))
1453	return rc;
1454
1455	/* Figure out destination address. */
1456	struct sockaddr *pSA = NULL;
1457	#ifdef RT_OS_WINDOWS
1458	int cbSA = 0;
1459	#else
1460	socklen_t cbSA = 0;
1461	#endif
1462	RTSOCKADDRUNION u;
1463	if (pAddr)
1464	{
1465	rc = rtSocketAddrFromNetAddr(pAddr, &u, sizeof(u), NULL);
1466	if (RT_FAILURE(rc))
1467	return rc;
1468	pSA = &u.Addr;
1469	cbSA = sizeof(u);
1470	}
1471
1472	/*
1473	* Must write all at once, otherwise it is a failure.
1474	*/
1475	#ifdef RT_OS_WINDOWS
1476	int cbNow = cbBuffer >= RTSOCKET_MAX_WRITE ? RTSOCKET_MAX_WRITE : (int)cbBuffer;
1477	#else
1478	size_t cbNow = cbBuffer >= SSIZE_MAX ? SSIZE_MAX : cbBuffer;
1479	#endif
1480	ssize_t cbWritten = sendto(pThis->hNative, (const char *)pvBuffer, cbNow, MSG_NOSIGNAL, pSA, cbSA);
1481	if (RT_LIKELY((size_t)cbWritten == cbBuffer && cbWritten >= 0))
1482	rc = VINF_SUCCESS;
1483	else if (cbWritten < 0)
1484	rc = rtSocketError();
1485	else
1486	rc = VERR_TOO_MUCH_DATA;
1487
1488	/// @todo rtSocketUnlock(pThis);
1489	#ifdef RT_OS_WINDOWS
1490	# undef sendto
1491	#endif
1492	return rc;
1493	}
1494
1495
1496	RTDECL(int) RTSocketSgWrite(RTSOCKET hSocket, PCRTSGBUF pSgBuf)
1497	{
1498	/*
1499	* Validate input.
1500	*/
1501	RTSOCKETINT *pThis = hSocket;
1502	AssertPtrReturn(pThis, VERR_INVALID_HANDLE);
1503	AssertReturn(pThis->u32Magic == RTSOCKET_MAGIC, VERR_INVALID_HANDLE);
1504	AssertPtrReturn(pSgBuf, VERR_INVALID_PARAMETER);
1505	AssertReturn(pSgBuf->cSegs > 0, VERR_INVALID_PARAMETER);
1506	AssertReturn(rtSocketTryLock(pThis), VERR_CONCURRENT_ACCESS);
1507
1508	int rc = rtSocketSwitchBlockingMode(pThis, true /* fBlocking */);
1509	if (RT_FAILURE(rc))
1510	return rc;
1511
1512	/*
1513	* Construct message descriptor (translate pSgBuf) and send it.
1514	*/
1515	rc = VERR_NO_TMP_MEMORY;
1516	#ifdef RT_OS_WINDOWS
1517	if (g_pfnWSASend)
1518	{
1519	AssertCompileSize(WSABUF, sizeof(RTSGSEG));
1520	AssertCompileMemberSize(WSABUF, buf, RT_SIZEOFMEMB(RTSGSEG, pvSeg));
1521
1522	LPWSABUF paMsg = (LPWSABUF)RTMemTmpAllocZ(pSgBuf->cSegs * sizeof(WSABUF));
1523	if (paMsg)
1524	{
1525	for (unsigned i = 0; i < pSgBuf->cSegs; i++)
1526	{
1527	paMsg[i].buf = (char *)pSgBuf->paSegs[i].pvSeg;
1528	paMsg[i].len = (u_long)pSgBuf->paSegs[i].cbSeg;
1529	}
1530
1531	DWORD dwSent;
1532	int hrc = g_pfnWSASend(pThis->hNative, paMsg, pSgBuf->cSegs, &dwSent, MSG_NOSIGNAL, NULL, NULL);
1533	if (!hrc)
1534	rc = VINF_SUCCESS;
1535	/** @todo check for incomplete writes */
1536	else
1537	rc = rtSocketError();
1538
1539	RTMemTmpFree(paMsg);
1540	}
1541	}
1542	else if (g_pfnsend)
1543	{
1544	rc = VINF_SUCCESS;
1545	for (uint32_t iSeg = 0; iSeg < pSgBuf->cSegs; iSeg++)
1546	{
1547	uint8_t const pbSeg = (uint8_t const )pSgBuf->paSegs[iSeg].pvSeg;
1548	size_t cbSeg = pSgBuf->paSegs[iSeg].cbSeg;
1549	int cbNow;
1550	ssize_t cbWritten;
1551	for (;;)
1552	{
1553	cbNow = cbSeg >= RTSOCKET_MAX_WRITE ? RTSOCKET_MAX_WRITE : (int)cbSeg;
1554	cbWritten = g_pfnsend(pThis->hNative, (const char *)pbSeg, cbNow, MSG_NOSIGNAL);
1555	if ((size_t)cbWritten >= cbSeg \|\| cbWritten < 0)
1556	break;
1557	pbSeg += cbWritten;
1558	cbSeg -= cbWritten;
1559	}
1560	if (cbWritten < 0)
1561	{
1562	rc = rtSocketError();
1563	break;
1564	}
1565	}
1566	}
1567	else
1568	rc = VERR_NET_NOT_UNSUPPORTED;
1569
1570	#else /* !RT_OS_WINDOWS */
1571	AssertCompileSize(struct iovec, sizeof(RTSGSEG));
1572	AssertCompileMemberSize(struct iovec, iov_base, RT_SIZEOFMEMB(RTSGSEG, pvSeg));
1573	AssertCompileMemberSize(struct iovec, iov_len, RT_SIZEOFMEMB(RTSGSEG, cbSeg));
1574
1575	struct iovec paMsg = (struct iovec )RTMemTmpAllocZ(pSgBuf->cSegs * sizeof(struct iovec));
1576	if (paMsg)
1577	{
1578	for (unsigned i = 0; i < pSgBuf->cSegs; i++)
1579	{
1580	paMsg[i].iov_base = pSgBuf->paSegs[i].pvSeg;
1581	paMsg[i].iov_len = pSgBuf->paSegs[i].cbSeg;
1582	}
1583
1584	struct msghdr msgHdr;
1585	RT_ZERO(msgHdr);
1586	msgHdr.msg_iov = paMsg;
1587	msgHdr.msg_iovlen = pSgBuf->cSegs;
1588	ssize_t cbWritten = sendmsg(pThis->hNative, &msgHdr, MSG_NOSIGNAL);
1589	if (RT_LIKELY(cbWritten >= 0))
1590	rc = VINF_SUCCESS;
1591	/** @todo check for incomplete writes */
1592	else
1593	rc = rtSocketError();
1594
1595	RTMemTmpFree(paMsg);
1596	}
1597	#endif /* !RT_OS_WINDOWS */
1598
1599	rtSocketUnlock(pThis);
1600	return rc;
1601	}
1602
1603
1604	RTDECL(int) RTSocketSgWriteL(RTSOCKET hSocket, size_t cSegs, ...)
1605	{
1606	va_list va;
1607	va_start(va, cSegs);
1608	int rc = RTSocketSgWriteLV(hSocket, cSegs, va);
1609	va_end(va);
1610	return rc;
1611	}
1612
1613
1614	RTDECL(int) RTSocketSgWriteLV(RTSOCKET hSocket, size_t cSegs, va_list va)
1615	{
1616	/*
1617	* Set up a S/G segment array + buffer on the stack and pass it
1618	* on to RTSocketSgWrite.
1619	*/
1620	Assert(cSegs <= 16);
1621	PRTSGSEG paSegs = (PRTSGSEG)alloca(cSegs * sizeof(RTSGSEG));
1622	AssertReturn(paSegs, VERR_NO_TMP_MEMORY);
1623	for (size_t i = 0; i < cSegs; i++)
1624	{
1625	paSegs[i].pvSeg = va_arg(va, void *);
1626	paSegs[i].cbSeg = va_arg(va, size_t);
1627	}
1628
1629	RTSGBUF SgBuf;
1630	RTSgBufInit(&SgBuf, paSegs, cSegs);
1631	return RTSocketSgWrite(hSocket, &SgBuf);
1632	}
1633
1634
1635	RTDECL(int) RTSocketReadNB(RTSOCKET hSocket, void pvBuffer, size_t cbBuffer, size_t pcbRead)
1636	{
1637	/*
1638	* Validate input.
1639	*/
1640	RTSOCKETINT *pThis = hSocket;
1641	AssertPtrReturn(pThis, VERR_INVALID_HANDLE);
1642	AssertReturn(pThis->u32Magic == RTSOCKET_MAGIC, VERR_INVALID_HANDLE);
1643	AssertReturn(cbBuffer > 0, VERR_INVALID_PARAMETER);
1644	AssertPtr(pvBuffer);
1645	AssertPtrReturn(pcbRead, VERR_INVALID_PARAMETER);
1646	#ifdef RT_OS_WINDOWS
1647	AssertReturn(g_pfnrecv, VERR_NET_NOT_UNSUPPORTED);
1648	#endif
1649	AssertReturn(rtSocketTryLock(pThis), VERR_CONCURRENT_ACCESS);
1650
1651	int rc = rtSocketSwitchBlockingMode(pThis, false /* fBlocking */);
1652	if (RT_FAILURE(rc))
1653	return rc;
1654
1655	rtSocketErrorReset();
1656	#ifdef RTSOCKET_MAX_READ
1657	int cbNow = cbBuffer >= RTSOCKET_MAX_WRITE ? RTSOCKET_MAX_WRITE : (int)cbBuffer;
1658	#else
1659	size_t cbNow = cbBuffer;
1660	#endif
1661
1662	#ifdef RT_OS_WINDOWS
1663	int cbRead = g_pfnrecv(pThis->hNative, (char *)pvBuffer, cbNow, MSG_NOSIGNAL);
1664	if (cbRead >= 0)
1665	{
1666	*pcbRead = cbRead;
1667	rc = VINF_SUCCESS;
1668	}
1669	else
1670	{
1671	rc = rtSocketError();
1672	if (rc == VERR_TRY_AGAIN)
1673	{
1674	*pcbRead = 0;
1675	rc = VINF_TRY_AGAIN;
1676	}
1677	}
1678
1679	#else
1680	ssize_t cbRead = recv(pThis->hNative, pvBuffer, cbNow, MSG_NOSIGNAL);
1681	if (cbRead >= 0)
1682	*pcbRead = cbRead;
1683	else if ( errno == EAGAIN
1684	# ifdef EWOULDBLOCK
1685	# if EWOULDBLOCK != EAGAIN
1686	\|\| errno == EWOULDBLOCK
1687	# endif
1688	# endif
1689	)
1690	{
1691	*pcbRead = 0;
1692	rc = VINF_TRY_AGAIN;
1693	}
1694	else
1695	rc = rtSocketError();
1696	#endif
1697
1698	rtSocketUnlock(pThis);
1699	return rc;
1700	}
1701
1702
1703	RTDECL(int) RTSocketWriteNB(RTSOCKET hSocket, const void pvBuffer, size_t cbBuffer, size_t pcbWritten)
1704	{
1705	/*
1706	* Validate input.
1707	*/
1708	RTSOCKETINT *pThis = hSocket;
1709	AssertPtrReturn(pThis, VERR_INVALID_HANDLE);
1710	AssertReturn(pThis->u32Magic == RTSOCKET_MAGIC, VERR_INVALID_HANDLE);
1711	AssertPtrReturn(pcbWritten, VERR_INVALID_PARAMETER);
1712	#ifdef RT_OS_WINDOWS
1713	AssertReturn(g_pfnsend, VERR_NET_NOT_UNSUPPORTED);
1714	#endif
1715	AssertReturn(rtSocketTryLock(pThis), VERR_CONCURRENT_ACCESS);
1716
1717	int rc = rtSocketSwitchBlockingMode(pThis, false /* fBlocking */);
1718	if (RT_FAILURE(rc))
1719	return rc;
1720
1721	rtSocketErrorReset();
1722	#ifdef RT_OS_WINDOWS
1723	# ifdef RTSOCKET_MAX_WRITE
1724	int cbNow = cbBuffer >= RTSOCKET_MAX_WRITE ? RTSOCKET_MAX_WRITE : (int)cbBuffer;
1725	# else
1726	size_t cbNow = cbBuffer;
1727	# endif
1728	int cbWritten = g_pfnsend(pThis->hNative, (const char *)pvBuffer, cbNow, MSG_NOSIGNAL);
1729	if (cbWritten >= 0)
1730	{
1731	*pcbWritten = cbWritten;
1732	rc = VINF_SUCCESS;
1733	}
1734	else
1735	{
1736	rc = rtSocketError();
1737	if (rc == VERR_TRY_AGAIN)
1738	{
1739	*pcbWritten = 0;
1740	rc = VINF_TRY_AGAIN;
1741	}
1742	}
1743	#else
1744	ssize_t cbWritten = send(pThis->hNative, pvBuffer, cbBuffer, MSG_NOSIGNAL);
1745	if (cbWritten >= 0)
1746	*pcbWritten = cbWritten;
1747	else if ( errno == EAGAIN
1748	# ifdef EWOULDBLOCK
1749	# if EWOULDBLOCK != EAGAIN
1750	\|\| errno == EWOULDBLOCK
1751	# endif
1752	# endif
1753	)
1754	{
1755	*pcbWritten = 0;
1756	rc = VINF_TRY_AGAIN;
1757	}
1758	else
1759	rc = rtSocketError();
1760	#endif
1761
1762	rtSocketUnlock(pThis);
1763	return rc;
1764	}
1765
1766
1767	RTDECL(int) RTSocketSgWriteNB(RTSOCKET hSocket, PCRTSGBUF pSgBuf, size_t *pcbWritten)
1768	{
1769	/*
1770	* Validate input.
1771	*/
1772	RTSOCKETINT *pThis = hSocket;
1773	AssertPtrReturn(pThis, VERR_INVALID_HANDLE);
1774	AssertReturn(pThis->u32Magic == RTSOCKET_MAGIC, VERR_INVALID_HANDLE);
1775	AssertPtrReturn(pSgBuf, VERR_INVALID_PARAMETER);
1776	AssertPtrReturn(pcbWritten, VERR_INVALID_PARAMETER);
1777	AssertReturn(pSgBuf->cSegs > 0, VERR_INVALID_PARAMETER);
1778	AssertReturn(rtSocketTryLock(pThis), VERR_CONCURRENT_ACCESS);
1779
1780	int rc = rtSocketSwitchBlockingMode(pThis, false /* fBlocking */);
1781	if (RT_FAILURE(rc))
1782	return rc;
1783
1784	unsigned cSegsToSend = 0;
1785	rc = VERR_NO_TMP_MEMORY;
1786	#ifdef RT_OS_WINDOWS
1787	if (g_pfnWSASend)
1788	{
1789	LPWSABUF paMsg = NULL;
1790	RTSgBufMapToNative(paMsg, pSgBuf, WSABUF, buf, char *, len, u_long, cSegsToSend);
1791	if (paMsg)
1792	{
1793	DWORD dwSent = 0;
1794	int hrc = g_pfnWSASend(pThis->hNative, paMsg, cSegsToSend, &dwSent, MSG_NOSIGNAL, NULL, NULL);
1795	if (!hrc)
1796	rc = VINF_SUCCESS;
1797	else
1798	rc = rtSocketError();
1799
1800	*pcbWritten = dwSent;
1801
1802	RTMemTmpFree(paMsg);
1803	}
1804	}
1805	else if (g_pfnsend)
1806	{
1807	size_t cbWrittenTotal = 0;
1808	rc = VINF_SUCCESS;
1809	for (uint32_t iSeg = 0; iSeg < pSgBuf->cSegs; iSeg++)
1810	{
1811	uint8_t const pbSeg = (uint8_t const )pSgBuf->paSegs[iSeg].pvSeg;
1812	size_t cbSeg = pSgBuf->paSegs[iSeg].cbSeg;
1813	int cbNow;
1814	ssize_t cbWritten;
1815	for (;;)
1816	{
1817	cbNow = cbSeg >= RTSOCKET_MAX_WRITE ? RTSOCKET_MAX_WRITE : (int)cbSeg;
1818	cbWritten = g_pfnsend(pThis->hNative, (const char *)pbSeg, cbNow, MSG_NOSIGNAL);
1819	if ((size_t)cbWritten >= cbSeg \|\| cbWritten < 0)
1820	break;
1821	cbWrittenTotal += cbWrittenTotal;
1822	pbSeg += cbWritten;
1823	cbSeg -= cbWritten;
1824	}
1825	if (cbWritten < 0)
1826	{
1827	rc = rtSocketError();
1828	break;
1829	}
1830	if (cbWritten != cbNow)
1831	break;
1832	}
1833	*pcbWritten = cbWrittenTotal;
1834	}
1835	else
1836	rc = VERR_NET_NOT_UNSUPPORTED;
1837
1838	#else /* !RT_OS_WINDOWS */
1839	struct iovec *paMsg = NULL;
1840
1841	RTSgBufMapToNative(paMsg, pSgBuf, struct iovec, iov_base, void *, iov_len, size_t, cSegsToSend);
1842	if (paMsg)
1843	{
1844	struct msghdr msgHdr;
1845	RT_ZERO(msgHdr);
1846	msgHdr.msg_iov = paMsg;
1847	msgHdr.msg_iovlen = cSegsToSend;
1848	ssize_t cbWritten = sendmsg(pThis->hNative, &msgHdr, MSG_NOSIGNAL);
1849	if (RT_LIKELY(cbWritten >= 0))
1850	{
1851	rc = VINF_SUCCESS;
1852	*pcbWritten = cbWritten;
1853	}
1854	else
1855	rc = rtSocketError();
1856
1857	RTMemTmpFree(paMsg);
1858	}
1859	#endif /* !RT_OS_WINDOWS */
1860
1861	rtSocketUnlock(pThis);
1862	return rc;
1863	}
1864
1865
1866	RTDECL(int) RTSocketSgWriteLNB(RTSOCKET hSocket, size_t cSegs, size_t *pcbWritten, ...)
1867	{
1868	va_list va;
1869	va_start(va, pcbWritten);
1870	int rc = RTSocketSgWriteLVNB(hSocket, cSegs, pcbWritten, va);
1871	va_end(va);
1872	return rc;
1873	}
1874
1875
1876	RTDECL(int) RTSocketSgWriteLVNB(RTSOCKET hSocket, size_t cSegs, size_t *pcbWritten, va_list va)
1877	{
1878	/*
1879	* Set up a S/G segment array + buffer on the stack and pass it
1880	* on to RTSocketSgWrite.
1881	*/
1882	Assert(cSegs <= 16);
1883	PRTSGSEG paSegs = (PRTSGSEG)alloca(cSegs * sizeof(RTSGSEG));
1884	AssertReturn(paSegs, VERR_NO_TMP_MEMORY);
1885	for (size_t i = 0; i < cSegs; i++)
1886	{
1887	paSegs[i].pvSeg = va_arg(va, void *);
1888	paSegs[i].cbSeg = va_arg(va, size_t);
1889	}
1890
1891	RTSGBUF SgBuf;
1892	RTSgBufInit(&SgBuf, paSegs, cSegs);
1893	return RTSocketSgWriteNB(hSocket, &SgBuf, pcbWritten);
1894	}
1895
1896
1897	RTDECL(int) RTSocketSelectOne(RTSOCKET hSocket, RTMSINTERVAL cMillies)
1898	{
1899	/*
1900	* Validate input.
1901	*/
1902	RTSOCKETINT *pThis = hSocket;
1903	AssertPtrReturn(pThis, VERR_INVALID_HANDLE);
1904	AssertReturn(pThis->u32Magic == RTSOCKET_MAGIC, VERR_INVALID_HANDLE);
1905	AssertReturn(RTMemPoolRefCount(pThis) >= (pThis->cUsers ? 2U : 1U), VERR_CALLER_NO_REFERENCE);
1906	int const fdMax = (int)pThis->hNative + 1;
1907	AssertReturn((RTSOCKETNATIVE)(fdMax - 1) == pThis->hNative, VERR_INTERNAL_ERROR_5);
1908	#ifdef RT_OS_WINDOWS
1909	AssertReturn(g_pfnselect, VERR_NET_NOT_UNSUPPORTED);
1910	# define select g_pfnselect
1911	#endif
1912
1913	/*
1914	* Set up the file descriptor sets and do the select.
1915	*/
1916	fd_set fdsetR;
1917	FD_ZERO(&fdsetR);
1918	FD_SET(pThis->hNative, &fdsetR);
1919
1920	fd_set fdsetE = fdsetR;
1921
1922	int rc;
1923	if (cMillies == RT_INDEFINITE_WAIT)
1924	rc = select(fdMax, &fdsetR, NULL, &fdsetE, NULL);
1925	else
1926	{
1927	struct timeval timeout;
1928	timeout.tv_sec = cMillies / 1000;
1929	timeout.tv_usec = (cMillies % 1000) * 1000;
1930	rc = select(fdMax, &fdsetR, NULL, &fdsetE, &timeout);
1931	}
1932	if (rc > 0)
1933	rc = VINF_SUCCESS;
1934	else if (rc == 0)
1935	rc = VERR_TIMEOUT;
1936	else
1937	rc = rtSocketError();
1938
1939	#ifdef RT_OS_WINDOWS
1940	# undef select
1941	#endif
1942	return rc;
1943	}
1944
1945
1946	/**
1947	* Internal worker for RTSocketSelectOneEx and rtSocketPollCheck (fallback)
1948	*
1949	* @returns IPRT status code
1950	* @param pThis The socket (valid).
1951	* @param fEvents The events to select for.
1952	* @param pfEvents Where to return the events.
1953	* @param cMillies How long to select for, in milliseconds.
1954	*/
1955	static int rtSocketSelectOneEx(RTSOCKET pThis, uint32_t fEvents, uint32_t *pfEvents, RTMSINTERVAL cMillies)
1956	{
1957	RTSOCKETNATIVE hNative = pThis->hNative;
1958	if (hNative == NIL_RTSOCKETNATIVE)
1959	{
1960	/* Socket is already closed? Possible we raced someone calling rtSocketCloseIt.
1961	Should we return a different status code? */
1962	*pfEvents = RTSOCKET_EVT_ERROR;
1963	return VINF_SUCCESS;
1964	}
1965
1966	int const fdMax = (int)hNative + 1;
1967	AssertReturn((RTSOCKETNATIVE)(fdMax - 1) == hNative, VERR_INTERNAL_ERROR_5);
1968	#ifdef RT_OS_WINDOWS
1969	AssertReturn(g_pfnselect, VERR_NET_NOT_UNSUPPORTED);
1970	AssertReturn(g_pfn__WSAFDIsSet, VERR_NET_NOT_UNSUPPORTED);
1971	# define select g_pfnselect
1972	# define __WSAFDIsSet g_pfn__WSAFDIsSet
1973	#endif
1974
1975	*pfEvents = 0;
1976
1977	/*
1978	* Set up the file descriptor sets and do the select.
1979	*/
1980	fd_set fdsetR;
1981	fd_set fdsetW;
1982	fd_set fdsetE;
1983	FD_ZERO(&fdsetR);
1984	FD_ZERO(&fdsetW);
1985	FD_ZERO(&fdsetE);
1986
1987	if (fEvents & RTSOCKET_EVT_READ)
1988	FD_SET(hNative, &fdsetR);
1989	if (fEvents & RTSOCKET_EVT_WRITE)
1990	FD_SET(hNative, &fdsetW);
1991	if (fEvents & RTSOCKET_EVT_ERROR)
1992	FD_SET(hNative, &fdsetE);
1993
1994	int rc;
1995	if (cMillies == RT_INDEFINITE_WAIT)
1996	rc = select(fdMax, &fdsetR, &fdsetW, &fdsetE, NULL);
1997	else
1998	{
1999	struct timeval timeout;
2000	timeout.tv_sec = cMillies / 1000;
2001	timeout.tv_usec = (cMillies % 1000) * 1000;
2002	rc = select(fdMax, &fdsetR, &fdsetW, &fdsetE, &timeout);
2003	}
2004	if (rc > 0)
2005	{
2006	if (pThis->hNative == hNative)
2007	{
2008	if (FD_ISSET(hNative, &fdsetR))
2009	*pfEvents \|= RTSOCKET_EVT_READ;
2010	if (FD_ISSET(hNative, &fdsetW))
2011	*pfEvents \|= RTSOCKET_EVT_WRITE;
2012	if (FD_ISSET(hNative, &fdsetE))
2013	*pfEvents \|= RTSOCKET_EVT_ERROR;
2014	rc = VINF_SUCCESS;
2015	}
2016	else
2017	{
2018	/* Socket was closed while we waited (rtSocketCloseIt). Different status code? */
2019	*pfEvents = RTSOCKET_EVT_ERROR;
2020	rc = VINF_SUCCESS;
2021	}
2022	}
2023	else if (rc == 0)
2024	rc = VERR_TIMEOUT;
2025	else
2026	rc = rtSocketError();
2027
2028	#ifdef RT_OS_WINDOWS
2029	# undef select
2030	# undef __WSAFDIsSet
2031	#endif
2032	return rc;
2033	}
2034
2035
2036	RTDECL(int) RTSocketSelectOneEx(RTSOCKET hSocket, uint32_t fEvents, uint32_t *pfEvents, RTMSINTERVAL cMillies)
2037	{
2038	/*
2039	* Validate input.
2040	*/
2041	RTSOCKETINT *pThis = hSocket;
2042	AssertPtrReturn(pThis, VERR_INVALID_HANDLE);
2043	AssertReturn(pThis->u32Magic == RTSOCKET_MAGIC, VERR_INVALID_HANDLE);
2044	AssertPtrReturn(pfEvents, VERR_INVALID_PARAMETER);
2045	AssertReturn(!(fEvents & ~RTSOCKET_EVT_VALID_MASK), VERR_INVALID_PARAMETER);
2046	AssertReturn(RTMemPoolRefCount(pThis) >= (pThis->cUsers ? 2U : 1U), VERR_CALLER_NO_REFERENCE);
2047
2048	return rtSocketSelectOneEx(pThis, fEvents, pfEvents, cMillies);
2049	}
2050
2051
2052	RTDECL(int) RTSocketShutdown(RTSOCKET hSocket, bool fRead, bool fWrite)
2053	{
2054	/*
2055	* Validate input, don't lock it because we might want to interrupt a call
2056	* active on a different thread.
2057	*/
2058	RTSOCKETINT *pThis = hSocket;
2059	AssertPtrReturn(pThis, VERR_INVALID_HANDLE);
2060	AssertReturn(pThis->u32Magic == RTSOCKET_MAGIC, VERR_INVALID_HANDLE);
2061	AssertReturn(RTMemPoolRefCount(pThis) >= (pThis->cUsers ? 2U : 1U), VERR_CALLER_NO_REFERENCE);
2062	AssertReturn(fRead \|\| fWrite, VERR_INVALID_PARAMETER);
2063	#ifdef RT_OS_WINDOWS
2064	AssertReturn(g_pfnshutdown, VERR_NET_NOT_UNSUPPORTED);
2065	# define shutdown g_pfnshutdown
2066	#endif
2067
2068	/*
2069	* Do the job.
2070	*/
2071	int rc = VINF_SUCCESS;
2072	int fHow;
2073	if (fRead && fWrite)
2074	fHow = SHUT_RDWR;
2075	else if (fRead)
2076	fHow = SHUT_RD;
2077	else
2078	fHow = SHUT_WR;
2079	if (shutdown(pThis->hNative, fHow) == -1)
2080	rc = rtSocketError();
2081
2082	#ifdef RT_OS_WINDOWS
2083	# undef shutdown
2084	#endif
2085	return rc;
2086	}
2087
2088
2089	RTDECL(int) RTSocketGetLocalAddress(RTSOCKET hSocket, PRTNETADDR pAddr)
2090	{
2091	/*
2092	* Validate input.
2093	*/
2094	RTSOCKETINT *pThis = hSocket;
2095	AssertPtrReturn(pThis, VERR_INVALID_HANDLE);
2096	AssertReturn(pThis->u32Magic == RTSOCKET_MAGIC, VERR_INVALID_HANDLE);
2097	AssertReturn(RTMemPoolRefCount(pThis) >= (pThis->cUsers ? 2U : 1U), VERR_CALLER_NO_REFERENCE);
2098	#ifdef RT_OS_WINDOWS
2099	AssertReturn(g_pfngetsockname, VERR_NET_NOT_UNSUPPORTED);
2100	# define getsockname g_pfngetsockname
2101	#endif
2102
2103	/*
2104	* Get the address and convert it.
2105	*/
2106	int rc;
2107	RTSOCKADDRUNION u;
2108	#ifdef RT_OS_WINDOWS
2109	int cbAddr = sizeof(u);
2110	#else
2111	socklen_t cbAddr = sizeof(u);
2112	#endif
2113	RT_ZERO(u);
2114	if (getsockname(pThis->hNative, &u.Addr, &cbAddr) == 0)
2115	rc = rtSocketNetAddrFromAddr(&u, cbAddr, pAddr);
2116	else
2117	rc = rtSocketError();
2118
2119	#ifdef RT_OS_WINDOWS
2120	# undef getsockname
2121	#endif
2122	return rc;
2123	}
2124
2125
2126	RTDECL(int) RTSocketGetPeerAddress(RTSOCKET hSocket, PRTNETADDR pAddr)
2127	{
2128	/*
2129	* Validate input.
2130	*/
2131	RTSOCKETINT *pThis = hSocket;
2132	AssertPtrReturn(pThis, VERR_INVALID_HANDLE);
2133	AssertReturn(pThis->u32Magic == RTSOCKET_MAGIC, VERR_INVALID_HANDLE);
2134	AssertReturn(RTMemPoolRefCount(pThis) >= (pThis->cUsers ? 2U : 1U), VERR_CALLER_NO_REFERENCE);
2135	#ifdef RT_OS_WINDOWS
2136	AssertReturn(g_pfngetpeername, VERR_NET_NOT_UNSUPPORTED);
2137	# define getpeername g_pfngetpeername
2138	#endif
2139
2140	/*
2141	* Get the address and convert it.
2142	*/
2143	int rc;
2144	RTSOCKADDRUNION u;
2145	#ifdef RT_OS_WINDOWS
2146	int cbAddr = sizeof(u);
2147	#else
2148	socklen_t cbAddr = sizeof(u);
2149	#endif
2150	RT_ZERO(u);
2151	if (getpeername(pThis->hNative, &u.Addr, &cbAddr) == 0)
2152	rc = rtSocketNetAddrFromAddr(&u, cbAddr, pAddr);
2153	else
2154	rc = rtSocketError();
2155
2156	#ifdef RT_OS_WINDOWS
2157	# undef getpeername
2158	#endif
2159	return rc;
2160	}
2161
2162
2163
2164	/**
2165	* Wrapper around bind.
2166	*
2167	* @returns IPRT status code.
2168	* @param hSocket The socket handle.
2169	* @param pAddr The address to bind to.
2170	*/
2171	DECLHIDDEN(int) rtSocketBind(RTSOCKET hSocket, PCRTNETADDR pAddr)
2172	{
2173	RTSOCKADDRUNION u;
2174	int cbAddr;
2175	int rc = rtSocketAddrFromNetAddr(pAddr, &u, sizeof(u), &cbAddr);
2176	if (RT_SUCCESS(rc))
2177	rc = rtSocketBindRawAddr(hSocket, &u.Addr, cbAddr);
2178	return rc;
2179	}
2180
2181
2182	/**
2183	* Very thin wrapper around bind.
2184	*
2185	* @returns IPRT status code.
2186	* @param hSocket The socket handle.
2187	* @param pvAddr The address to bind to (struct sockaddr and
2188	* friends).
2189	* @param cbAddr The size of the address.
2190	*/
2191	DECLHIDDEN(int) rtSocketBindRawAddr(RTSOCKET hSocket, void const *pvAddr, size_t cbAddr)
2192	{
2193	/*
2194	* Validate input.
2195	*/
2196	RTSOCKETINT *pThis = hSocket;
2197	AssertPtrReturn(pThis, VERR_INVALID_HANDLE);
2198	AssertReturn(pThis->u32Magic == RTSOCKET_MAGIC, VERR_INVALID_HANDLE);
2199	AssertPtrReturn(pvAddr, VERR_INVALID_POINTER);
2200	#ifdef RT_OS_WINDOWS
2201	AssertReturn(g_pfnbind, VERR_NET_NOT_UNSUPPORTED);
2202	# define bind g_pfnbind
2203	#endif
2204	AssertReturn(rtSocketTryLock(pThis), VERR_CONCURRENT_ACCESS);
2205
2206	int rc;
2207	if (bind(pThis->hNative, (struct sockaddr const *)pvAddr, (int)cbAddr) == 0)
2208	rc = VINF_SUCCESS;
2209	else
2210	rc = rtSocketError();
2211
2212	rtSocketUnlock(pThis);
2213	#ifdef RT_OS_WINDOWS
2214	# undef bind
2215	#endif
2216	return rc;
2217	}
2218
2219
2220
2221	/**
2222	* Wrapper around listen.
2223	*
2224	* @returns IPRT status code.
2225	* @param hSocket The socket handle.
2226	* @param cMaxPending The max number of pending connections.
2227	*/
2228	DECLHIDDEN(int) rtSocketListen(RTSOCKET hSocket, int cMaxPending)
2229	{
2230	/*
2231	* Validate input.
2232	*/
2233	RTSOCKETINT *pThis = hSocket;
2234	AssertPtrReturn(pThis, VERR_INVALID_HANDLE);
2235	AssertReturn(pThis->u32Magic == RTSOCKET_MAGIC, VERR_INVALID_HANDLE);
2236	#ifdef RT_OS_WINDOWS
2237	AssertReturn(g_pfnlisten, VERR_NET_NOT_UNSUPPORTED);
2238	# define listen g_pfnlisten
2239	#endif
2240	AssertReturn(rtSocketTryLock(pThis), VERR_CONCURRENT_ACCESS);
2241
2242	int rc = VINF_SUCCESS;
2243	if (listen(pThis->hNative, cMaxPending) != 0)
2244	rc = rtSocketError();
2245
2246	rtSocketUnlock(pThis);
2247	#ifdef RT_OS_WINDOWS
2248	# undef listen
2249	#endif
2250	return rc;
2251	}
2252
2253
2254	/**
2255	* Wrapper around accept.
2256	*
2257	* @returns IPRT status code.
2258	* @param hSocket The socket handle.
2259	* @param phClient Where to return the client socket handle on
2260	* success.
2261	* @param pAddr Where to return the client address.
2262	* @param pcbAddr On input this gives the size buffer size of what
2263	* @a pAddr point to. On return this contains the
2264	* size of what's stored at @a pAddr.
2265	*/
2266	DECLHIDDEN(int) rtSocketAccept(RTSOCKET hSocket, PRTSOCKET phClient, struct sockaddr pAddr, size_t pcbAddr)
2267	{
2268	/*
2269	* Validate input.
2270	* Only lock the socket temporarily while we get the native handle, so that
2271	* we can safely shutdown and destroy the socket from a different thread.
2272	*/
2273	RTSOCKETINT *pThis = hSocket;
2274	AssertPtrReturn(pThis, VERR_INVALID_HANDLE);
2275	AssertReturn(pThis->u32Magic == RTSOCKET_MAGIC, VERR_INVALID_HANDLE);
2276	#ifdef RT_OS_WINDOWS
2277	AssertReturn(g_pfnaccept, VERR_NET_NOT_UNSUPPORTED);
2278	AssertReturn(g_pfnclosesocket, VERR_NET_NOT_UNSUPPORTED);
2279	# define accept g_pfnaccept
2280	#endif
2281	AssertReturn(rtSocketTryLock(pThis), VERR_CONCURRENT_ACCESS);
2282
2283	/*
2284	* Call accept().
2285	*/
2286	rtSocketErrorReset();
2287	int rc = VINF_SUCCESS;
2288	#ifdef RT_OS_WINDOWS
2289	int cbAddr = (int)*pcbAddr;
2290	#else
2291	socklen_t cbAddr = *pcbAddr;
2292	#endif
2293	RTSOCKETNATIVE hNativeClient = accept(pThis->hNative, pAddr, &cbAddr);
2294	if (hNativeClient != NIL_RTSOCKETNATIVE)
2295	{
2296	*pcbAddr = cbAddr;
2297
2298	/*
2299	* Wrap the client socket.
2300	*/
2301	rc = rtSocketCreateForNative(phClient, hNativeClient);
2302	if (RT_FAILURE(rc))
2303	{
2304	#ifdef RT_OS_WINDOWS
2305	g_pfnclosesocket(hNativeClient);
2306	#else
2307	close(hNativeClient);
2308	#endif
2309	}
2310	}
2311	else
2312	rc = rtSocketError();
2313
2314	rtSocketUnlock(pThis);
2315	#ifdef RT_OS_WINDOWS
2316	# undef accept
2317	#endif
2318	return rc;
2319	}
2320
2321
2322	/**
2323	* Wrapper around connect.
2324	*
2325	* @returns IPRT status code.
2326	* @param hSocket The socket handle.
2327	* @param pAddr The socket address to connect to.
2328	* @param cMillies Number of milliseconds to wait for the connect attempt to complete.
2329	* Use RT_INDEFINITE_WAIT to wait for ever.
2330	* Use RT_TCPCLIENTCONNECT_DEFAULT_WAIT to wait for the default time
2331	* configured on the running system.
2332	*/
2333	DECLHIDDEN(int) rtSocketConnect(RTSOCKET hSocket, PCRTNETADDR pAddr, RTMSINTERVAL cMillies)
2334	{
2335	/*
2336	* Validate input.
2337	*/
2338	RTSOCKETINT *pThis = hSocket;
2339	AssertPtrReturn(pThis, VERR_INVALID_HANDLE);
2340	AssertReturn(pThis->u32Magic == RTSOCKET_MAGIC, VERR_INVALID_HANDLE);
2341	#ifdef RT_OS_WINDOWS
2342	AssertReturn(g_pfnconnect, VERR_NET_NOT_UNSUPPORTED);
2343	AssertReturn(g_pfnselect, VERR_NET_NOT_UNSUPPORTED);
2344	AssertReturn(g_pfngetsockopt, VERR_NET_NOT_UNSUPPORTED);
2345	# define connect g_pfnconnect
2346	# define select g_pfnselect
2347	# define getsockopt g_pfngetsockopt
2348	#endif
2349	AssertReturn(rtSocketTryLock(pThis), VERR_CONCURRENT_ACCESS);
2350
2351	RTSOCKADDRUNION u;
2352	int cbAddr;
2353	int rc = rtSocketAddrFromNetAddr(pAddr, &u, sizeof(u), &cbAddr);
2354	if (RT_SUCCESS(rc))
2355	{
2356	if (cMillies == RT_SOCKETCONNECT_DEFAULT_WAIT)
2357	{
2358	if (connect(pThis->hNative, &u.Addr, cbAddr) != 0)
2359	rc = rtSocketError();
2360	}
2361	else
2362	{
2363	/*
2364	* Switch the socket to nonblocking mode, initiate the connect
2365	* and wait for the socket to become writable or until the timeout
2366	* expires.
2367	*/
2368	rc = rtSocketSwitchBlockingMode(pThis, false /* fBlocking */);
2369	if (RT_SUCCESS(rc))
2370	{
2371	if (connect(pThis->hNative, &u.Addr, cbAddr) != 0)
2372	{
2373	rc = rtSocketError();
2374	if (rc == VERR_TRY_AGAIN \|\| rc == VERR_NET_IN_PROGRESS)
2375	{
2376	int rcSock = 0;
2377	fd_set FdSetWriteable;
2378	struct timeval TvTimeout;
2379
2380	TvTimeout.tv_sec = cMillies / RT_MS_1SEC;
2381	TvTimeout.tv_usec = (cMillies % RT_MS_1SEC) * RT_US_1MS;
2382
2383	FD_ZERO(&FdSetWriteable);
2384	FD_SET(pThis->hNative, &FdSetWriteable);
2385	do
2386	{
2387	rcSock = select(pThis->hNative + 1, NULL, &FdSetWriteable, NULL,
2388	cMillies == RT_INDEFINITE_WAIT \|\| cMillies >= INT_MAX
2389	? NULL
2390	: &TvTimeout);
2391	if (rcSock > 0)
2392	{
2393	int iSockError = 0;
2394	socklen_t cbSockOpt = sizeof(iSockError);
2395	rcSock = getsockopt(pThis->hNative, SOL_SOCKET, SO_ERROR, (char *)&iSockError, &cbSockOpt);
2396	if (rcSock == 0)
2397	{
2398	if (iSockError == 0)
2399	rc = VINF_SUCCESS;
2400	else
2401	{
2402	#ifdef RT_OS_WINDOWS
2403	rc = RTErrConvertFromWin32(iSockError);
2404	#else
2405	rc = RTErrConvertFromErrno(iSockError);
2406	#endif
2407	}
2408	}
2409	else
2410	rc = rtSocketError();
2411	}
2412	else if (rcSock == 0)
2413	rc = VERR_TIMEOUT;
2414	else
2415	rc = rtSocketError();
2416	} while (rc == VERR_INTERRUPTED);
2417	}
2418	}
2419
2420	rtSocketSwitchBlockingMode(pThis, true /* fBlocking */);
2421	}
2422	}
2423	}
2424
2425	rtSocketUnlock(pThis);
2426	#ifdef RT_OS_WINDOWS
2427	# undef connect
2428	# undef select
2429	# undef getsockopt
2430	#endif
2431	return rc;
2432	}
2433
2434
2435	/**
2436	* Wrapper around connect, raw address, no timeout.
2437	*
2438	* @returns IPRT status code.
2439	* @param hSocket The socket handle.
2440	* @param pvAddr The raw socket address to connect to.
2441	* @param cbAddr The size of the raw address.
2442	*/
2443	DECLHIDDEN(int) rtSocketConnectRaw(RTSOCKET hSocket, void const *pvAddr, size_t cbAddr)
2444	{
2445	/*
2446	* Validate input.
2447	*/
2448	RTSOCKETINT *pThis = hSocket;
2449	AssertPtrReturn(pThis, VERR_INVALID_HANDLE);
2450	AssertReturn(pThis->u32Magic == RTSOCKET_MAGIC, VERR_INVALID_HANDLE);
2451	#ifdef RT_OS_WINDOWS
2452	AssertReturn(g_pfnconnect, VERR_NET_NOT_UNSUPPORTED);
2453	# define connect g_pfnconnect
2454	#endif
2455	AssertReturn(rtSocketTryLock(pThis), VERR_CONCURRENT_ACCESS);
2456
2457	int rc;
2458	if (connect(pThis->hNative, (const struct sockaddr *)pvAddr, (int)cbAddr) == 0)
2459	rc = VINF_SUCCESS;
2460	else
2461	rc = rtSocketError();
2462
2463	rtSocketUnlock(pThis);
2464	#ifdef RT_OS_WINDOWS
2465	# undef connect
2466	#endif
2467	return rc;
2468	}
2469
2470
2471	/**
2472	* Wrapper around setsockopt.
2473	*
2474	* @returns IPRT status code.
2475	* @param hSocket The socket handle.
2476	* @param iLevel The protocol level, e.g. IPPORTO_TCP.
2477	* @param iOption The option, e.g. TCP_NODELAY.
2478	* @param pvValue The value buffer.
2479	* @param cbValue The size of the value pointed to by pvValue.
2480	*/
2481	DECLHIDDEN(int) rtSocketSetOpt(RTSOCKET hSocket, int iLevel, int iOption, void const *pvValue, int cbValue)
2482	{
2483	/*
2484	* Validate input.
2485	*/
2486	RTSOCKETINT *pThis = hSocket;
2487	AssertPtrReturn(pThis, VERR_INVALID_HANDLE);
2488	AssertReturn(pThis->u32Magic == RTSOCKET_MAGIC, VERR_INVALID_HANDLE);
2489	#ifdef RT_OS_WINDOWS
2490	AssertReturn(g_pfnsetsockopt, VERR_NET_NOT_UNSUPPORTED);
2491	# define setsockopt g_pfnsetsockopt
2492	#endif
2493	AssertReturn(rtSocketTryLock(pThis), VERR_CONCURRENT_ACCESS);
2494
2495	int rc = VINF_SUCCESS;
2496	if (setsockopt(pThis->hNative, iLevel, iOption, (const char *)pvValue, cbValue) != 0)
2497	rc = rtSocketError();
2498
2499	rtSocketUnlock(pThis);
2500	#ifdef RT_OS_WINDOWS
2501	# undef setsockopt
2502	#endif
2503	return rc;
2504	}
2505
2506
2507	/**
2508	* Internal RTPollSetAdd helper that returns the handle that should be added to
2509	* the pollset.
2510	*
2511	* @returns Valid handle on success, INVALID_HANDLE_VALUE on failure.
2512	* @param hSocket The socket handle.
2513	* @param fEvents The events we're polling for.
2514	* @param phNative Where to put the primary handle.
2515	*/
2516	DECLHIDDEN(int) rtSocketPollGetHandle(RTSOCKET hSocket, uint32_t fEvents, PRTHCINTPTR phNative)
2517	{
2518	RTSOCKETINT *pThis = hSocket;
2519	RT_NOREF_PV(fEvents);
2520	AssertPtrReturn(pThis, VERR_INVALID_HANDLE);
2521	AssertReturn(pThis->u32Magic == RTSOCKET_MAGIC, VERR_INVALID_HANDLE);
2522	#ifdef RT_OS_WINDOWS
2523	AssertReturn(rtSocketTryLock(pThis), VERR_CONCURRENT_ACCESS);
2524
2525	int rc = VINF_SUCCESS;
2526	if (pThis->hEvent != WSA_INVALID_EVENT)
2527	*phNative = (RTHCINTPTR)pThis->hEvent;
2528	else if (g_pfnWSACreateEvent)
2529	{
2530	pThis->hEvent = g_pfnWSACreateEvent();
2531	*phNative = (RTHCINTPTR)pThis->hEvent;
2532	if (pThis->hEvent == WSA_INVALID_EVENT)
2533	rc = rtSocketError();
2534	}
2535	else
2536	{
2537	AssertCompile(WSA_INVALID_EVENT == (WSAEVENT)NULL);
2538	pThis->hEvent = CreateEventW(NULL, TRUE /fManualReset/, FALSE /fInitialState/, NULL /pwszName/);
2539	*phNative = (RTHCINTPTR)pThis->hEvent;
2540	if (pThis->hEvent == WSA_INVALID_EVENT)
2541	rc = RTErrConvertFromWin32(GetLastError());
2542	}
2543
2544	rtSocketUnlock(pThis);
2545	return rc;
2546
2547	#else /* !RT_OS_WINDOWS */
2548	*phNative = (RTHCUINTPTR)pThis->hNative;
2549	return VINF_SUCCESS;
2550	#endif /* !RT_OS_WINDOWS */
2551	}
2552
2553	#ifdef RT_OS_WINDOWS
2554
2555	/**
2556	* Fallback poller thread.
2557	*
2558	* @returns VINF_SUCCESS.
2559	* @param hSelf The thread handle.
2560	* @param pvUser Socket instance data.
2561	*/
2562	static DECLCALLBACK(int) rtSocketPollFallbackThreadProc(RTTHREAD hSelf, void *pvUser)
2563	{
2564	RTSOCKETINT pThis = (RTSOCKETINT )pvUser;
2565	RT_NOREF(hSelf);
2566	# define __WSAFDIsSet g_pfn__WSAFDIsSet
2567
2568	/*
2569	* The execution loop.
2570	*/
2571	while (!ASMAtomicReadBool(&pThis->fPollFallbackShutdown))
2572	{
2573	/*
2574	* Do the selecting (with a 15 second timeout because that seems like a good idea).
2575	*/
2576	struct fd_set SetRead;
2577	struct fd_set SetWrite;
2578	struct fd_set SetXcpt;
2579
2580	FD_ZERO(&SetRead);
2581	FD_ZERO(&SetWrite);
2582	FD_ZERO(&SetXcpt);
2583
2584	FD_SET(pThis->hPollFallbackNotifyR, &SetRead);
2585	FD_SET(pThis->hPollFallbackNotifyR, &SetXcpt);
2586
2587	bool fActive = ASMAtomicReadBool(&pThis->fPollFallbackActive);
2588	uint32_t fEvents;
2589	if (!fActive)
2590	fEvents = 0;
2591	else
2592	{
2593	fEvents = ASMAtomicReadU32(&pThis->fSubscribedEvts);
2594	if (fEvents & RTPOLL_EVT_READ)
2595	FD_SET(pThis->hNative, &SetRead);
2596	if (fEvents & RTPOLL_EVT_WRITE)
2597	FD_SET(pThis->hNative, &SetWrite);
2598	if (fEvents & RTPOLL_EVT_ERROR)
2599	FD_SET(pThis->hNative, &SetXcpt);
2600	}
2601
2602	struct timeval Timeout;
2603	Timeout.tv_sec = 15;
2604	Timeout.tv_usec = 0;
2605	int rc = g_pfnselect(INT_MAX /ignored/, &SetRead, &SetWrite, &SetXcpt, &Timeout);
2606
2607	/* Stop immediately if told to shut down. */
2608	if (ASMAtomicReadBool(&pThis->fPollFallbackShutdown))
2609	break;
2610
2611	/*
2612	* Process the result.
2613	*/
2614	if (rc > 0)
2615	{
2616	/* First the socket we're listening on. */
2617	if ( fEvents
2618	&& ( FD_ISSET(pThis->hNative, &SetRead)
2619	\|\| FD_ISSET(pThis->hNative, &SetWrite)
2620	\|\| FD_ISSET(pThis->hNative, &SetXcpt)) )
2621	{
2622	ASMAtomicWriteBool(&pThis->fPollFallbackActive, false);
2623	SetEvent(pThis->hEvent);
2624	}
2625
2626	/* Then maintain the notification pipe. (We only read one byte here
2627	because we're overly paranoid wrt socket switching to blocking mode.) */
2628	if (FD_ISSET(pThis->hPollFallbackNotifyR, &SetRead))
2629	{
2630	char chIgnored;
2631	g_pfnrecv(pThis->hPollFallbackNotifyR, &chIgnored, sizeof(chIgnored), MSG_NOSIGNAL);
2632	}
2633	}
2634	else
2635	AssertMsg(rc == 0, ("%Rrc\n", rtSocketError()));
2636	}
2637
2638	# undef __WSAFDIsSet
2639	return VINF_SUCCESS;
2640	}
2641
2642
2643	/**
2644	* Pokes the fallback thread, making sure it gets out of whatever it's stuck in.
2645	*
2646	* @param pThis The socket handle.
2647	*/
2648	static void rtSocketPokePollFallbackThread(RTSOCKETINT *pThis)
2649	{
2650	Assert(pThis->fPollFallback);
2651	if (pThis->hPollFallbackThread != NIL_RTTHREAD)
2652	{
2653	int cbWritten = g_pfnsend(pThis->hPollFallbackNotifyW, "!", 1, MSG_NOSIGNAL);
2654	AssertMsg(cbWritten == 1, ("cbWritten=%d err=%Rrc\n", rtSocketError()));
2655	RT_NOREF_PV(cbWritten);
2656	}
2657	}
2658
2659
2660	/**
2661	* Called by rtSocketPollStart to make the thread start selecting on the socket.
2662	*
2663	* @returns 0 on success, RTPOLL_EVT_ERROR on failure.
2664	* @param pThis The socket handle.
2665	*/
2666	static uint32_t rtSocketPollFallbackStart(RTSOCKETINT *pThis)
2667	{
2668	/*
2669	* Reset the event and tell the thread to start selecting on the socket.
2670	*/
2671	ResetEvent(pThis->hEvent);
2672	ASMAtomicWriteBool(&pThis->fPollFallbackActive, true);
2673
2674	/*
2675	* Wake up the thread the thread.
2676	*/
2677	if (pThis->hPollFallbackThread != NIL_RTTHREAD)
2678	rtSocketPokePollFallbackThread(pThis);
2679	else
2680	{
2681	/*
2682	* Not running, need to set it up and start it.
2683	*/
2684	AssertLogRelReturn(pThis->hEvent != NULL && pThis->hEvent != INVALID_HANDLE_VALUE, RTPOLL_EVT_ERROR);
2685
2686	/* Create the notification socket pair. */
2687	int rc;
2688	if (pThis->hPollFallbackNotifyR == NIL_RTSOCKETNATIVE)
2689	{
2690	rc = rtSocketCreateNativeTcpPair(&pThis->hPollFallbackNotifyW, &pThis->hPollFallbackNotifyR);
2691	AssertLogRelRCReturn(rc, RTPOLL_EVT_ERROR);
2692
2693	/* Make the read end non-blocking (not fatal). */
2694	u_long fNonBlocking = 1;
2695	rc = g_pfnioctlsocket(pThis->hPollFallbackNotifyR, FIONBIO, &fNonBlocking);
2696	AssertLogRelMsg(rc == 0, ("rc=%#x %Rrc\n", rc, rtSocketError()));
2697	}
2698
2699	/* Finally, start the thread. ASSUME we don't need too much stack. */
2700	rc = RTThreadCreate(&pThis->hPollFallbackThread, rtSocketPollFallbackThreadProc, pThis,
2701	_128K, RTTHREADTYPE_IO, RTTHREADFLAGS_WAITABLE, "sockpoll");
2702	AssertLogRelRCReturn(rc, RTPOLL_EVT_ERROR);
2703	}
2704	return 0;
2705	}
2706
2707
2708	/**
2709	* Undos the harm done by WSAEventSelect.
2710	*
2711	* @returns IPRT status code.
2712	* @param pThis The socket handle.
2713	*/
2714	static int rtSocketPollClearEventAndRestoreBlocking(RTSOCKETINT *pThis)
2715	{
2716	int rc = VINF_SUCCESS;
2717	if (pThis->fSubscribedEvts)
2718	{
2719	if (!pThis->fPollFallback)
2720	{
2721	Assert(g_pfnWSAEventSelect && g_pfnioctlsocket);
2722	if (g_pfnWSAEventSelect && g_pfnioctlsocket)
2723	{
2724	if (g_pfnWSAEventSelect(pThis->hNative, WSA_INVALID_EVENT, 0) == 0)
2725	{
2726	pThis->fSubscribedEvts = 0;
2727
2728	/*
2729	* Switch back to blocking mode if that was the state before the
2730	* operation.
2731	*/
2732	if (pThis->fBlocking)
2733	{
2734	u_long fNonBlocking = 0;
2735	int rc2 = g_pfnioctlsocket(pThis->hNative, FIONBIO, &fNonBlocking);
2736	if (rc2 != 0)
2737	{
2738	rc = rtSocketError();
2739	AssertMsgFailed(("%Rrc; rc2=%d\n", rc, rc2));
2740	}
2741	}
2742	}
2743	else
2744	{
2745	rc = rtSocketError();
2746	AssertMsgFailed(("%Rrc\n", rc));
2747	}
2748	}
2749	else
2750	{
2751	Assert(pThis->fPollFallback);
2752	rc = VINF_SUCCESS;
2753	}
2754	}
2755	/*
2756	* Just clear the event mask as we never started waiting if we get here.
2757	*/
2758	else
2759	ASMAtomicWriteU32(&pThis->fSubscribedEvts, 0);
2760	}
2761	return rc;
2762	}
2763
2764
2765	/**
2766	* Updates the mask of events we're subscribing to.
2767	*
2768	* @returns IPRT status code.
2769	* @param pThis The socket handle.
2770	* @param fEvents The events we want to subscribe to.
2771	*/
2772	static int rtSocketPollUpdateEvents(RTSOCKETINT *pThis, uint32_t fEvents)
2773	{
2774	if (!pThis->fPollFallback)
2775	{
2776	LONG fNetworkEvents = 0;
2777	if (fEvents & RTPOLL_EVT_READ)
2778	fNetworkEvents \|= FD_READ;
2779	if (fEvents & RTPOLL_EVT_WRITE)
2780	fNetworkEvents \|= FD_WRITE;
2781	if (fEvents & RTPOLL_EVT_ERROR)
2782	fNetworkEvents \|= FD_CLOSE;
2783	LogFlowFunc(("fNetworkEvents=%#x\n", fNetworkEvents));
2784
2785	if (g_pfnWSAEventSelect(pThis->hNative, pThis->hEvent, fNetworkEvents) == 0)
2786	{
2787	pThis->fSubscribedEvts = fEvents;
2788	return VINF_SUCCESS;
2789	}
2790
2791	int rc = rtSocketError();
2792	AssertMsgFailed(("fNetworkEvents=%#x rc=%Rrc\n", fNetworkEvents, rtSocketError()));
2793	return rc;
2794	}
2795
2796	/*
2797	* Update the events we're waiting for. Caller will poke/start the thread. later
2798	*/
2799	ASMAtomicWriteU32(&pThis->fSubscribedEvts, fEvents);
2800	return VINF_SUCCESS;
2801	}
2802
2803	#endif /* RT_OS_WINDOWS */
2804
2805
2806	#if defined(RT_OS_WINDOWS) \|\| defined(RT_OS_OS2)
2807
2808	/**
2809	* Checks for pending events.
2810	*
2811	* @returns Event mask or 0.
2812	* @param pThis The socket handle.
2813	* @param fEvents The desired events.
2814	*/
2815	static uint32_t rtSocketPollCheck(RTSOCKETINT *pThis, uint32_t fEvents)
2816	{
2817	uint32_t fRetEvents = 0;
2818
2819	LogFlowFunc(("pThis=%#p fEvents=%#x\n", pThis, fEvents));
2820
2821	# ifdef RT_OS_WINDOWS
2822	/* Make sure WSAEnumNetworkEvents returns what we want. */
2823	int rc = VINF_SUCCESS;
2824	if ((pThis->fSubscribedEvts & fEvents) != fEvents)
2825	rc = rtSocketPollUpdateEvents(pThis, pThis->fSubscribedEvts \| fEvents);
2826
2827	if (!pThis->fPollFallback)
2828	{
2829	/* Atomically get pending events and reset the event semaphore. */
2830	Assert(g_pfnWSAEnumNetworkEvents);
2831	WSANETWORKEVENTS NetEvts;
2832	RT_ZERO(NetEvts);
2833	if (g_pfnWSAEnumNetworkEvents(pThis->hNative, pThis->hEvent, &NetEvts) == 0)
2834	{
2835	if ( (NetEvts.lNetworkEvents & FD_READ)
2836	&& NetEvts.iErrorCode[FD_READ_BIT] == 0)
2837	fRetEvents \|= RTPOLL_EVT_READ;
2838
2839	if ( (NetEvts.lNetworkEvents & FD_WRITE)
2840	&& NetEvts.iErrorCode[FD_WRITE_BIT] == 0)
2841	fRetEvents \|= RTPOLL_EVT_WRITE;
2842
2843	if (NetEvts.lNetworkEvents & FD_CLOSE)
2844	fRetEvents \|= RTPOLL_EVT_ERROR;
2845	else
2846	for (uint32_t i = 0; i < FD_MAX_EVENTS; i++)
2847	if ( (NetEvts.lNetworkEvents & (1L << i))
2848	&& NetEvts.iErrorCode[i] != 0)
2849	fRetEvents \|= RTPOLL_EVT_ERROR;
2850
2851	pThis->fEventsSaved = fRetEvents \|= pThis->fEventsSaved;
2852	fRetEvents &= fEvents \| RTPOLL_EVT_ERROR;
2853	}
2854	else
2855	rc = rtSocketError();
2856	}
2857
2858	/* Fall back on select if we hit an error above or is using fallback polling. */
2859	if (pThis->fPollFallback \|\| RT_FAILURE(rc))
2860	{
2861	rc = rtSocketSelectOneEx(pThis, fEvents & RTPOLL_EVT_ERROR ? fEvents \| RTPOLL_EVT_READ : fEvents, &fRetEvents, 0);
2862	if (RT_SUCCESS(rc))
2863	{
2864	/* rtSocketSelectOneEx may return RTPOLL_EVT_READ on disconnect. Use
2865	getpeername to fix this. */
2866	if ((fRetEvents & (RTPOLL_EVT_READ \| RTPOLL_EVT_ERROR)) == RTPOLL_EVT_READ)
2867	{
2868	# if 0 /* doens't work */
2869	rtSocketErrorReset();
2870	char chIgn;
2871	rc = g_pfnrecv(pThis->hNative, &chIgn, 0, MSG_NOSIGNAL);
2872	rc = rtSocketError();
2873	if (RT_FAILURE(rc))
2874	fRetEvents \|= RTPOLL_EVT_ERROR;
2875
2876	rc = g_pfnsend(pThis->hNative, &chIgn, 0, MSG_NOSIGNAL);
2877	rc = rtSocketError();
2878	if (RT_FAILURE(rc))
2879	fRetEvents \|= RTPOLL_EVT_ERROR;
2880
2881	RTSOCKADDRUNION u;
2882	int cbAddr = sizeof(u);
2883	if (g_pfngetpeername(pThis->hNative, &u.Addr, &cbAddr) == SOCKET_ERROR)
2884	fRetEvents \|= RTPOLL_EVT_ERROR;
2885	# endif
2886	/* If no bytes are available, assume error condition. */
2887	u_long cbAvail = 0;
2888	rc = g_pfnioctlsocket(pThis->hNative, FIONREAD, &cbAvail);
2889	if (rc == 0 && cbAvail == 0)
2890	fRetEvents \|= RTPOLL_EVT_ERROR;
2891	}
2892	fRetEvents &= fEvents \| RTPOLL_EVT_ERROR;
2893	}
2894	else if (rc == VERR_TIMEOUT)
2895	fRetEvents = 0;
2896	else
2897	fRetEvents \|= RTPOLL_EVT_ERROR;
2898	}
2899
2900	# else /* RT_OS_OS2 */
2901	int aFds[4] = { pThis->hNative, pThis->hNative, pThis->hNative, -1 };
2902	int rc = os2_select(aFds, 1, 1, 1, 0);
2903	if (rc > 0)
2904	{
2905	if (aFds[0] == pThis->hNative)
2906	fRetEvents \|= RTPOLL_EVT_READ;
2907	if (aFds[1] == pThis->hNative)
2908	fRetEvents \|= RTPOLL_EVT_WRITE;
2909	if (aFds[2] == pThis->hNative)
2910	fRetEvents \|= RTPOLL_EVT_ERROR;
2911	fRetEvents &= fEvents;
2912	}
2913	# endif /* RT_OS_OS2 */
2914
2915	LogFlowFunc(("fRetEvents=%#x\n", fRetEvents));
2916	return fRetEvents;
2917	}
2918
2919
2920	/**
2921	* Internal RTPoll helper that polls the socket handle and, if @a fNoWait is
2922	* clear, starts whatever actions we've got running during the poll call.
2923	*
2924	* @returns 0 if no pending events, actions initiated if @a fNoWait is clear.
2925	* Event mask (in @a fEvents) and no actions if the handle is ready
2926	* already.
2927	* UINT32_MAX (asserted) if the socket handle is busy in I/O or a
2928	* different poll set.
2929	*
2930	* @param hSocket The socket handle.
2931	* @param hPollSet The poll set handle (for access checks).
2932	* @param fEvents The events we're polling for.
2933	* @param fFinalEntry Set if this is the final entry for this handle
2934	* in this poll set. This can be used for dealing
2935	* with duplicate entries.
2936	* @param fNoWait Set if it's a zero-wait poll call. Clear if
2937	* we'll wait for an event to occur.
2938	*
2939	* @remarks There is a potential race wrt duplicate handles when @a fNoWait is
2940	* @c true, we don't currently care about that oddity...
2941	*/
2942	DECLHIDDEN(uint32_t) rtSocketPollStart(RTSOCKET hSocket, RTPOLLSET hPollSet, uint32_t fEvents, bool fFinalEntry, bool fNoWait)
2943	{
2944	RTSOCKETINT *pThis = hSocket;
2945	AssertPtrReturn(pThis, UINT32_MAX);
2946	AssertReturn(pThis->u32Magic == RTSOCKET_MAGIC, UINT32_MAX);
2947	/** @todo This isn't quite sane. Replace by critsect and open up concurrent
2948	* reads and writes! */
2949	if (rtSocketTryLock(pThis))
2950	pThis->hPollSet = hPollSet;
2951	else
2952	{
2953	AssertReturn(pThis->hPollSet == hPollSet, UINT32_MAX);
2954	ASMAtomicIncU32(&pThis->cUsers);
2955	}
2956
2957	/* (rtSocketPollCheck will reset the event object). */
2958	# ifdef RT_OS_WINDOWS
2959	uint32_t fRetEvents = pThis->fEventsSaved;
2960	pThis->fEventsSaved = 0; /* Reset */
2961	fRetEvents \|= rtSocketPollCheck(pThis, fEvents);
2962
2963	if ( !fRetEvents
2964	&& !fNoWait)
2965	{
2966	pThis->fPollEvts \|= fEvents;
2967	if (fFinalEntry)
2968	{
2969	if (pThis->fSubscribedEvts != pThis->fPollEvts)
2970	{
2971	/** @todo seems like there migth be a call to many here and that fPollEvts is
2972	* totally unnecessary... (bird) */
2973	int rc = rtSocketPollUpdateEvents(pThis, pThis->fPollEvts);
2974	if (RT_FAILURE(rc))
2975	{
2976	pThis->fPollEvts = 0;
2977	fRetEvents = UINT32_MAX;
2978	}
2979	}
2980
2981	/* Make sure we don't block when there are events pending relevant to an earlier poll set entry. */
2982	if (pThis->fEventsSaved && !pThis->fPollFallback && g_pfnWSASetEvent && fRetEvents == 0)
2983	g_pfnWSASetEvent(pThis->hEvent);
2984	}
2985	}
2986	# else
2987	uint32_t fRetEvents = rtSocketPollCheck(pThis, fEvents);
2988	# endif
2989
2990	if (fRetEvents \|\| fNoWait)
2991	{
2992	if (pThis->cUsers == 1)
2993	{
2994	# ifdef RT_OS_WINDOWS
2995	pThis->fEventsSaved &= RTPOLL_EVT_ERROR;
2996	pThis->fHarvestedEvents = false;
2997	rtSocketPollClearEventAndRestoreBlocking(pThis);
2998	# endif
2999	pThis->hPollSet = NIL_RTPOLLSET;
3000	}
3001	# ifdef RT_OS_WINDOWS
3002	else
3003	pThis->fHarvestedEvents = true;
3004	# endif
3005	ASMAtomicDecU32(&pThis->cUsers);
3006	}
3007	# ifdef RT_OS_WINDOWS
3008	/*
3009	* Kick the poller thread on if this is the final entry and we're in
3010	* winsock 1.x fallback mode.
3011	*/
3012	else if (pThis->fPollFallback && fFinalEntry)
3013	fRetEvents = rtSocketPollFallbackStart(pThis);
3014	# endif
3015
3016	return fRetEvents;
3017	}
3018
3019
3020	/**
3021	* Called after a WaitForMultipleObjects returned in order to check for pending
3022	* events and stop whatever actions that rtSocketPollStart() initiated.
3023	*
3024	* @returns Event mask or 0.
3025	*
3026	* @param hSocket The socket handle.
3027	* @param fEvents The events we're polling for.
3028	* @param fFinalEntry Set if this is the final entry for this handle
3029	* in this poll set. This can be used for dealing
3030	* with duplicate entries. Only keep in mind that
3031	* this method is called in reverse order, so the
3032	* first call will have this set (when the entire
3033	* set was processed).
3034	* @param fHarvestEvents Set if we should check for pending events.
3035	*/
3036	DECLHIDDEN(uint32_t) rtSocketPollDone(RTSOCKET hSocket, uint32_t fEvents, bool fFinalEntry, bool fHarvestEvents)
3037	{
3038	RTSOCKETINT *pThis = hSocket;
3039	AssertPtrReturn(pThis, 0);
3040	AssertReturn(pThis->u32Magic == RTSOCKET_MAGIC, 0);
3041	Assert(pThis->cUsers > 0);
3042	Assert(pThis->hPollSet != NIL_RTPOLLSET);
3043	RT_NOREF_PV(fFinalEntry);
3044
3045	# ifdef RT_OS_WINDOWS
3046	/*
3047	* Deactivate the poll thread if we're in winsock 1.x fallback poll mode.
3048	*/
3049	if ( pThis->fPollFallback
3050	&& pThis->hPollFallbackThread != NIL_RTTHREAD)
3051	{
3052	ASMAtomicWriteU32(&pThis->fSubscribedEvts, 0);
3053	if (ASMAtomicXchgBool(&pThis->fPollFallbackActive, false))
3054	rtSocketPokePollFallbackThread(pThis);
3055	}
3056	# endif
3057
3058	/*
3059	* Harvest events and clear the event mask for the next round of polling.
3060	*/
3061	uint32_t fRetEvents;
3062	# ifdef RT_OS_WINDOWS
3063	if (!pThis->fPollFallback)
3064	{
3065	if (!pThis->fHarvestedEvents)
3066	{
3067	fRetEvents = rtSocketPollCheck(pThis, fEvents);
3068	pThis->fHarvestedEvents = true;
3069	}
3070	else
3071	fRetEvents = pThis->fEventsSaved;
3072	if (fHarvestEvents)
3073	fRetEvents &= fEvents;
3074	else
3075	fRetEvents = 0;
3076	pThis->fPollEvts = 0;
3077	}
3078	else
3079	# endif
3080	{
3081	if (fHarvestEvents)
3082	fRetEvents = rtSocketPollCheck(pThis, fEvents);
3083	else
3084	fRetEvents = 0;
3085	}
3086
3087	/*
3088	* Make the socket blocking again and unlock the handle.
3089	*/
3090	if (pThis->cUsers == 1)
3091	{
3092	# ifdef RT_OS_WINDOWS
3093	pThis->fEventsSaved &= RTPOLL_EVT_ERROR;
3094	pThis->fHarvestedEvents = false;
3095	rtSocketPollClearEventAndRestoreBlocking(pThis);
3096	# endif
3097	pThis->hPollSet = NIL_RTPOLLSET;
3098	}
3099	ASMAtomicDecU32(&pThis->cUsers);
3100	return fRetEvents;
3101	}
3102
3103	#endif /* RT_OS_WINDOWS \|\| RT_OS_OS2 */
3104

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

source: vbox/trunk/src/VBox/Runtime/r3/socket.cpp@ 72285

Download in other formats: