socket.cpp@ 94839

Last change on this file since 94839 was 94839, checked in by vboxsync, 3 years ago
Runtime/socket.cpp: Relax annoying assertion
Property svn:eol-style set to `native` Property svn:keywords set to `Id Revision`
File size: 93.8 KB

Line
1	/* $Id: socket.cpp 94839 2022-05-05 10:33:17Z vboxsync $ */
2	/** @file
3	* IPRT - Network Sockets.
4	*/
5
6	/*
7	* Copyright (C) 2006-2022 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	/** Whether to leave the native socket open rather than closing it (for
152	* RTHandleGetStandard). */
153	bool fLeaveOpen;
154	#if defined(RT_OS_WINDOWS) \|\| defined(RT_OS_OS2)
155	/** The pollset currently polling this socket. This is NIL if no one is
156	* polling. */
157	RTPOLLSET hPollSet;
158	#endif
159	#ifdef RT_OS_WINDOWS
160	/** The event semaphore we've associated with the socket handle.
161	* This is WSA_INVALID_EVENT if not done. */
162	WSAEVENT hEvent;
163	/** The events we're polling for. */
164	uint32_t fPollEvts;
165	/** The events we're currently subscribing to with WSAEventSelect.
166	* This is ZERO if we're currently not subscribing to anything. */
167	uint32_t fSubscribedEvts;
168	/** Saved events which are only posted once and events harvested for
169	* sockets entered multiple times into to a poll set. Imagine a scenario where
170	* you have a RTPOLL_EVT_READ entry and RTPOLL_EVT_ERROR entry. The READ
171	* condition can be triggered between checking the READ entry and the ERROR
172	* entry, and we don't want to drop the READ, so we store it here and make sure
173	* the event is signalled.
174	*
175	* The RTPOLL_EVT_ERROR is inconsistenly sticky at the momemnt... */
176	uint32_t fEventsSaved;
177	/** Set if fEventsSaved contains harvested events (used to avoid multiple
178	* calls to rtSocketPollCheck on the same socket during rtSocketPollDone). */
179	bool fHarvestedEvents;
180	/** Set if we're using the polling fallback. */
181	bool fPollFallback;
182	/** Set if the fallback polling is active (event not set). */
183	bool volatile fPollFallbackActive;
184	/** Set to shut down the fallback polling thread. */
185	bool volatile fPollFallbackShutdown;
186	/** Socket use to wake up the select thread. */
187	RTSOCKETNATIVE hPollFallbackNotifyW;
188	/** Socket the select thread always waits on. */
189	RTSOCKETNATIVE hPollFallbackNotifyR;
190	/** The fallback polling thread. */
191	RTTHREAD hPollFallbackThread;
192	#endif /* RT_OS_WINDOWS */
193	} RTSOCKETINT;
194
195
196	/**
197	* Address union used internally for things like getpeername and getsockname.
198	*/
199	typedef union RTSOCKADDRUNION
200	{
201	struct sockaddr Addr;
202	struct sockaddr_in IPv4;
203	#ifdef IPRT_WITH_TCPIP_V6
204	struct sockaddr_in6 IPv6;
205	#endif
206	} RTSOCKADDRUNION;
207
208
209	/*********************************************************************************************************************************
210	* Global Variables *
211	*********************************************************************************************************************************/
212	#ifdef RT_OS_WINDOWS
213	/** Indicates that we've successfully initialized winsock. */
214	static uint32_t volatile g_uWinSockInitedVersion = 0;
215	#endif
216
217
218	/*********************************************************************************************************************************
219	* Internal Functions *
220	*********************************************************************************************************************************/
221	#ifdef RT_OS_WINDOWS
222	static void rtSocketPokePollFallbackThread(RTSOCKETINT *pThis);
223	#endif
224
225
226
227	#ifdef RT_OS_WINDOWS
228	/**
229	* Initializes winsock for the process.
230	*
231	* @returns IPRT status code.
232	*/
233	static int rtSocketInitWinsock(void)
234	{
235	if (g_uWinSockInitedVersion != 0)
236	return VINF_SUCCESS;
237
238	if ( !g_pfnWSAGetLastError
239	\|\| !g_pfnWSAStartup
240	\|\| !g_pfnsocket
241	\|\| !g_pfnclosesocket)
242	return VERR_NET_INIT_FAILED;
243
244	/*
245	* Initialize winsock. Try with 2.2 and back down till we get something that works.
246	*/
247	static const WORD s_awVersions[] =
248	{
249	MAKEWORD(2, 2),
250	MAKEWORD(2, 1),
251	MAKEWORD(2, 0),
252	MAKEWORD(1, 1),
253	MAKEWORD(1, 0),
254	};
255	for (uint32_t i = 0; i < RT_ELEMENTS(s_awVersions); i++)
256	{
257	WSADATA wsaData;
258	RT_ZERO(wsaData);
259	int rcWsa = g_pfnWSAStartup(s_awVersions[i], &wsaData);
260	if (rcWsa == 0)
261	{
262	/* AssertMsg(wsaData.wVersion >= s_awVersions[i]); - triggers with winsock 1.1 */
263	ASMAtomicWriteU32(&g_uWinSockInitedVersion, wsaData.wVersion);
264	return VINF_SUCCESS;
265	}
266	AssertLogRelMsg(rcWsa == WSAVERNOTSUPPORTED, ("rcWsa=%d (winsock version %#x)\n", rcWsa, s_awVersions[i]));
267	}
268	LogRel(("Failed to init winsock!\n"));
269	return VERR_NET_INIT_FAILED;
270	}
271	#endif
272
273
274	/**
275	* Get the last error as an iprt status code.
276	*
277	* @returns IPRT status code.
278	*/
279	DECLINLINE(int) rtSocketError(void)
280	{
281	#ifdef RT_OS_WINDOWS
282	if (g_pfnWSAGetLastError)
283	return RTErrConvertFromWin32(g_pfnWSAGetLastError());
284	return VERR_NET_IO_ERROR;
285	#else
286	return RTErrConvertFromErrno(errno);
287	#endif
288	}
289
290
291	/**
292	* Resets the last error.
293	*/
294	DECLINLINE(void) rtSocketErrorReset(void)
295	{
296	#ifdef RT_OS_WINDOWS
297	if (g_pfnWSASetLastError)
298	g_pfnWSASetLastError(0);
299	#else
300	errno = 0;
301	#endif
302	}
303
304
305	/**
306	* Get the last resolver error as an iprt status code.
307	*
308	* @returns iprt status code.
309	*/
310	DECLHIDDEN(int) rtSocketResolverError(void)
311	{
312	#ifdef RT_OS_WINDOWS
313	if (g_pfnWSAGetLastError)
314	return RTErrConvertFromWin32(g_pfnWSAGetLastError());
315	return VERR_UNRESOLVED_ERROR;
316	#else
317	switch (h_errno)
318	{
319	case HOST_NOT_FOUND:
320	return VERR_NET_HOST_NOT_FOUND;
321	case NO_DATA:
322	return VERR_NET_ADDRESS_NOT_AVAILABLE;
323	case NO_RECOVERY:
324	return VERR_IO_GEN_FAILURE;
325	case TRY_AGAIN:
326	return VERR_TRY_AGAIN;
327
328	default:
329	AssertLogRelMsgFailed(("Unhandled error %u\n", h_errno));
330	return VERR_UNRESOLVED_ERROR;
331	}
332	#endif
333	}
334
335
336	/**
337	* Converts from a native socket address to a generic IPRT network address.
338	*
339	* @returns IPRT status code.
340	* @param pSrc The source address.
341	* @param cbSrc The size of the source address.
342	* @param pAddr Where to return the generic IPRT network
343	* address.
344	*/
345	static int rtSocketNetAddrFromAddr(RTSOCKADDRUNION const *pSrc, size_t cbSrc, PRTNETADDR pAddr)
346	{
347	/*
348	* Convert the address.
349	*/
350	if ( cbSrc == sizeof(struct sockaddr_in)
351	&& pSrc->Addr.sa_family == AF_INET)
352	{
353	RT_ZERO(*pAddr);
354	pAddr->enmType = RTNETADDRTYPE_IPV4;
355	pAddr->uPort = RT_N2H_U16(pSrc->IPv4.sin_port);
356	pAddr->uAddr.IPv4.u = pSrc->IPv4.sin_addr.s_addr;
357	}
358	#ifdef IPRT_WITH_TCPIP_V6
359	else if ( cbSrc == sizeof(struct sockaddr_in6)
360	&& pSrc->Addr.sa_family == AF_INET6)
361	{
362	RT_ZERO(*pAddr);
363	pAddr->enmType = RTNETADDRTYPE_IPV6;
364	pAddr->uPort = RT_N2H_U16(pSrc->IPv6.sin6_port);
365	pAddr->uAddr.IPv6.au32[0] = pSrc->IPv6.sin6_addr.s6_addr32[0];
366	pAddr->uAddr.IPv6.au32[1] = pSrc->IPv6.sin6_addr.s6_addr32[1];
367	pAddr->uAddr.IPv6.au32[2] = pSrc->IPv6.sin6_addr.s6_addr32[2];
368	pAddr->uAddr.IPv6.au32[3] = pSrc->IPv6.sin6_addr.s6_addr32[3];
369	}
370	#endif
371	else
372	return VERR_NET_ADDRESS_FAMILY_NOT_SUPPORTED;
373	return VINF_SUCCESS;
374	}
375
376
377	/**
378	* Converts from a generic IPRT network address to a native socket address.
379	*
380	* @returns IPRT status code.
381	* @param pAddr Pointer to the generic IPRT network address.
382	* @param pDst The source address.
383	* @param cbDst The size of the source address.
384	* @param pcbAddr Where to store the size of the returned address.
385	* Optional
386	*/
387	static int rtSocketAddrFromNetAddr(PCRTNETADDR pAddr, RTSOCKADDRUNION pDst, size_t cbDst, int pcbAddr)
388	{
389	RT_BZERO(pDst, cbDst);
390	if (pAddr->enmType == RTNETADDRTYPE_IPV4)
391	{
392	if (cbDst < sizeof(struct sockaddr_in))
393	return VERR_BUFFER_OVERFLOW;
394
395	pDst->Addr.sa_family = AF_INET;
396	pDst->IPv4.sin_port = RT_H2N_U16(pAddr->uPort);
397	pDst->IPv4.sin_addr.s_addr = pAddr->uAddr.IPv4.u;
398	if (pcbAddr)
399	*pcbAddr = sizeof(pDst->IPv4);
400	}
401	#ifdef IPRT_WITH_TCPIP_V6
402	else if (pAddr->enmType == RTNETADDRTYPE_IPV6)
403	{
404	if (cbDst < sizeof(struct sockaddr_in6))
405	return VERR_BUFFER_OVERFLOW;
406
407	pDst->Addr.sa_family = AF_INET6;
408	pDst->IPv6.sin6_port = RT_H2N_U16(pAddr->uPort);
409	pSrc->IPv6.sin6_addr.s6_addr32[0] = pAddr->uAddr.IPv6.au32[0];
410	pSrc->IPv6.sin6_addr.s6_addr32[1] = pAddr->uAddr.IPv6.au32[1];
411	pSrc->IPv6.sin6_addr.s6_addr32[2] = pAddr->uAddr.IPv6.au32[2];
412	pSrc->IPv6.sin6_addr.s6_addr32[3] = pAddr->uAddr.IPv6.au32[3];
413	if (pcbAddr)
414	*pcbAddr = sizeof(pDst->IPv6);
415	}
416	#endif
417	else
418	return VERR_NET_ADDRESS_FAMILY_NOT_SUPPORTED;
419	return VINF_SUCCESS;
420	}
421
422
423	/**
424	* Tries to lock the socket for exclusive usage by the calling thread.
425	*
426	* Call rtSocketUnlock() to unlock.
427	*
428	* @returns @c true if locked, @c false if not.
429	* @param pThis The socket structure.
430	*/
431	DECLINLINE(bool) rtSocketTryLock(RTSOCKETINT *pThis)
432	{
433	return ASMAtomicCmpXchgU32(&pThis->cUsers, 1, 0);
434	}
435
436
437	/**
438	* Unlocks the socket.
439	*
440	* @param pThis The socket structure.
441	*/
442	DECLINLINE(void) rtSocketUnlock(RTSOCKETINT *pThis)
443	{
444	ASMAtomicCmpXchgU32(&pThis->cUsers, 0, 1);
445	}
446
447
448	/**
449	* The slow path of rtSocketSwitchBlockingMode that does the actual switching.
450	*
451	* @returns IPRT status code.
452	* @param pThis The socket structure.
453	* @param fBlocking The desired mode of operation.
454	* @remarks Do not call directly.
455	*/
456	static int rtSocketSwitchBlockingModeSlow(RTSOCKETINT *pThis, bool fBlocking)
457	{
458	#ifdef RT_OS_WINDOWS
459	AssertReturn(g_pfnioctlsocket, VERR_NET_NOT_UNSUPPORTED);
460	u_long uBlocking = fBlocking ? 0 : 1;
461	if (g_pfnioctlsocket(pThis->hNative, FIONBIO, &uBlocking))
462	return rtSocketError();
463
464	#else
465	int fFlags = fcntl(pThis->hNative, F_GETFL, 0);
466	if (fFlags == -1)
467	return rtSocketError();
468
469	if (fBlocking)
470	fFlags &= ~O_NONBLOCK;
471	else
472	fFlags \|= O_NONBLOCK;
473	if (fcntl(pThis->hNative, F_SETFL, fFlags) == -1)
474	return rtSocketError();
475	#endif
476
477	pThis->fBlocking = fBlocking;
478	return VINF_SUCCESS;
479	}
480
481
482	/**
483	* Switches the socket to the desired blocking mode if necessary.
484	*
485	* The socket must be locked.
486	*
487	* @returns IPRT status code.
488	* @param pThis The socket structure.
489	* @param fBlocking The desired mode of operation.
490	*/
491	DECLINLINE(int) rtSocketSwitchBlockingMode(RTSOCKETINT *pThis, bool fBlocking)
492	{
493	if (pThis->fBlocking != fBlocking)
494	return rtSocketSwitchBlockingModeSlow(pThis, fBlocking);
495	return VINF_SUCCESS;
496	}
497
498
499	/**
500	* Creates an IPRT socket handle for a native one.
501	*
502	* @returns IPRT status code.
503	* @param ppSocket Where to return the IPRT socket handle.
504	* @param hNative The native handle.
505	* @param fLeaveOpen Whether to leave the native socket handle open when
506	* closed.
507	*/
508	DECLHIDDEN(int) rtSocketCreateForNative(RTSOCKETINT **ppSocket, RTSOCKETNATIVE hNative, bool fLeaveOpen)
509	{
510	RTSOCKETINT pThis = (RTSOCKETINT )RTMemPoolAlloc(RTMEMPOOL_DEFAULT, sizeof(*pThis));
511	if (!pThis)
512	return VERR_NO_MEMORY;
513	pThis->u32Magic = RTSOCKET_MAGIC;
514	pThis->cUsers = 0;
515	pThis->hNative = hNative;
516	pThis->fClosed = false;
517	pThis->fLeaveOpen = fLeaveOpen;
518	pThis->fBlocking = true;
519	#if defined(RT_OS_WINDOWS) \|\| defined(RT_OS_OS2)
520	pThis->hPollSet = NIL_RTPOLLSET;
521	#endif
522	#ifdef RT_OS_WINDOWS
523	pThis->hEvent = WSA_INVALID_EVENT;
524	pThis->fPollEvts = 0;
525	pThis->fSubscribedEvts = 0;
526	pThis->fEventsSaved = 0;
527	pThis->fHarvestedEvents = false;
528	pThis->fPollFallback = g_uWinSockInitedVersion < MAKEWORD(2, 0)
529	\|\| g_pfnWSACreateEvent == NULL
530	\|\| g_pfnWSACloseEvent == NULL
531	\|\| g_pfnWSAEventSelect == NULL
532	\|\| g_pfnWSAEnumNetworkEvents == NULL;
533	pThis->fPollFallbackActive = false;
534	pThis->fPollFallbackShutdown = false;
535	pThis->hPollFallbackNotifyR = NIL_RTSOCKETNATIVE;
536	pThis->hPollFallbackNotifyW = NIL_RTSOCKETNATIVE;
537	pThis->hPollFallbackThread = NIL_RTTHREAD;
538	#endif
539	*ppSocket = pThis;
540	return VINF_SUCCESS;
541	}
542
543
544	RTDECL(int) RTSocketFromNative(PRTSOCKET phSocket, RTHCINTPTR uNative)
545	{
546	AssertReturn(uNative != NIL_RTSOCKETNATIVE, VERR_INVALID_PARAMETER);
547	#ifndef RT_OS_WINDOWS
548	AssertReturn(uNative >= 0, VERR_INVALID_PARAMETER);
549	#endif
550	AssertPtrReturn(phSocket, VERR_INVALID_POINTER);
551	return rtSocketCreateForNative(phSocket, uNative, false /fLeaveOpen/);
552	}
553
554
555	/**
556	* Wrapper around socket().
557	*
558	* @returns IPRT status code.
559	* @param phSocket Where to store the handle to the socket on
560	* success.
561	* @param iDomain The protocol family (PF_XXX).
562	* @param iType The socket type (SOCK_XXX).
563	* @param iProtocol Socket parameter, usually 0.
564	*/
565	DECLHIDDEN(int) rtSocketCreate(PRTSOCKET phSocket, int iDomain, int iType, int iProtocol)
566	{
567	#ifdef RT_OS_WINDOWS
568	AssertReturn(g_pfnsocket, VERR_NET_NOT_UNSUPPORTED);
569	AssertReturn(g_pfnclosesocket, VERR_NET_NOT_UNSUPPORTED);
570
571	/* Initialize WinSock. */
572	int rc2 = rtSocketInitWinsock();
573	if (RT_FAILURE(rc2))
574	return rc2;
575	#endif
576
577	/*
578	* Create the socket.
579	*/
580	#ifdef RT_OS_WINDOWS
581	RTSOCKETNATIVE hNative = g_pfnsocket(iDomain, iType, iProtocol);
582	#else
583	RTSOCKETNATIVE hNative = socket(iDomain, iType, iProtocol);
584	#endif
585	if (hNative == NIL_RTSOCKETNATIVE)
586	return rtSocketError();
587
588	/*
589	* Wrap it.
590	*/
591	int rc = rtSocketCreateForNative(phSocket, hNative, false /fLeaveOpen/);
592	if (RT_FAILURE(rc))
593	{
594	#ifdef RT_OS_WINDOWS
595	g_pfnclosesocket(hNative);
596	#else
597	close(hNative);
598	#endif
599	}
600	return rc;
601	}
602
603
604	/**
605	* Wrapper around socketpair() for creating a local TCP connection.
606	*
607	* @returns IPRT status code.
608	* @param phServer Where to return the first native socket.
609	* @param phClient Where to return the second native socket.
610	*/
611	static int rtSocketCreateNativeTcpPair(RTSOCKETNATIVE phServer, RTSOCKETNATIVE phClient)
612	{
613	#ifdef RT_OS_WINDOWS
614	/*
615	* Initialize WinSock and make sure we got the necessary APIs.
616	*/
617	int rc = rtSocketInitWinsock();
618	if (RT_FAILURE(rc))
619	return rc;
620	AssertReturn(g_pfnsocket, VERR_NET_NOT_UNSUPPORTED);
621	AssertReturn(g_pfnclosesocket, VERR_NET_NOT_UNSUPPORTED);
622	AssertReturn(g_pfnsetsockopt, VERR_NET_NOT_UNSUPPORTED);
623	AssertReturn(g_pfnbind, VERR_NET_NOT_UNSUPPORTED);
624	AssertReturn(g_pfngetsockname, VERR_NET_NOT_UNSUPPORTED);
625	AssertReturn(g_pfnlisten, VERR_NET_NOT_UNSUPPORTED);
626	AssertReturn(g_pfnaccept, VERR_NET_NOT_UNSUPPORTED);
627	AssertReturn(g_pfnconnect, VERR_NET_NOT_UNSUPPORTED);
628
629	/*
630	* Create the "server" listen socket and the "client" socket.
631	*/
632	RTSOCKETNATIVE hListener = g_pfnsocket(AF_INET, SOCK_STREAM, IPPROTO_TCP);
633	if (hListener == NIL_RTSOCKETNATIVE)
634	return rtSocketError();
635	RTSOCKETNATIVE hClient = g_pfnsocket(AF_INET, SOCK_STREAM, IPPROTO_TCP);
636	if (hClient != NIL_RTSOCKETNATIVE)
637	{
638
639	/*
640	* We let WinSock choose a port number when we bind.
641	*/
642	union
643	{
644	struct sockaddr_in Ip;
645	struct sockaddr Generic;
646	} uAddr;
647	RT_ZERO(uAddr);
648	uAddr.Ip.sin_family = AF_INET;
649	uAddr.Ip.sin_addr.s_addr = RT_H2N_U32_C(INADDR_LOOPBACK);
650	//uAddr.Ip.sin_port = 0;
651	int fReuse = 1;
652	rc = g_pfnsetsockopt(hListener, SOL_SOCKET, SO_REUSEADDR, (const char *)&fReuse, sizeof(fReuse));
653	if (rc == 0)
654	{
655	rc = g_pfnbind(hListener, &uAddr.Generic, sizeof(uAddr.Ip));
656	if (rc == 0)
657	{
658	/*
659	* Get the address the client should connect to. According to the docs,
660	* we cannot assume that getsockname sets the IP and family.
661	*/
662	RT_ZERO(uAddr);
663	int cbAddr = sizeof(uAddr.Ip);
664	rc = g_pfngetsockname(hListener, &uAddr.Generic, &cbAddr);
665	if (rc == 0)
666	{
667	uAddr.Ip.sin_family = AF_INET;
668	uAddr.Ip.sin_addr.s_addr = RT_H2N_U32_C(INADDR_LOOPBACK);
669
670	/*
671	* Listen, connect and accept.
672	*/
673	rc = g_pfnlisten(hListener, 1 /cBacklog/);
674	if (rc == 0)
675	{
676	rc = g_pfnconnect(hClient, &uAddr.Generic, sizeof(uAddr.Ip));
677	if (rc == 0)
678	{
679	RTSOCKETNATIVE hServer = g_pfnaccept(hListener, NULL, NULL);
680	if (hServer != NIL_RTSOCKETNATIVE)
681	{
682	g_pfnclosesocket(hListener);
683
684	/*
685	* Done!
686	*/
687	*phServer = hServer;
688	*phClient = hClient;
689	return VINF_SUCCESS;
690	}
691	}
692	}
693	}
694	}
695	}
696	rc = rtSocketError();
697	g_pfnclosesocket(hClient);
698	}
699	else
700	rc = rtSocketError();
701	g_pfnclosesocket(hListener);
702	return rc;
703
704	#else
705	/*
706	* Got socket pair, so use it.
707	* Note! This isn't TCP per se, but it should fool the users.
708	*/
709	int aSockets[2] = { -1, -1 };
710	if (socketpair(AF_LOCAL, SOCK_STREAM, 0, aSockets) == 0)
711	{
712	*phServer = aSockets[0];
713	*phClient = aSockets[1];
714	return VINF_SUCCESS;
715	}
716	return rtSocketError();
717	#endif
718	}
719
720
721	/**
722	* Worker for RTTcpCreatePair.
723	*
724	* @returns IPRT status code.
725	* @param phServer Where to return the "server" side of the pair.
726	* @param phClient Where to return the "client" side of the pair.
727	* @note There is no server or client side, but we gotta call it something.
728	*/
729	DECLHIDDEN(int) rtSocketCreateTcpPair(RTSOCKET phServer, RTSOCKET phClient)
730	{
731	RTSOCKETNATIVE hServer = NIL_RTSOCKETNATIVE;
732	RTSOCKETNATIVE hClient = NIL_RTSOCKETNATIVE;
733	int rc = rtSocketCreateNativeTcpPair(&hServer, &hClient);
734	if (RT_SUCCESS(rc))
735	{
736	rc = rtSocketCreateForNative(phServer, hServer, false /fLeaveOpen/);
737	if (RT_SUCCESS(rc))
738	{
739	rc = rtSocketCreateForNative(phClient, hClient, false /fLeaveOpen/);
740	if (RT_SUCCESS(rc))
741	return VINF_SUCCESS;
742	RTSocketRelease(*phServer);
743	}
744	else
745	{
746	#ifdef RT_OS_WINDOWS
747	g_pfnclosesocket(hServer);
748	#else
749	close(hServer);
750	#endif
751	}
752	#ifdef RT_OS_WINDOWS
753	g_pfnclosesocket(hClient);
754	#else
755	close(hClient);
756	#endif
757	}
758
759	*phServer = NIL_RTSOCKET;
760	*phClient = NIL_RTSOCKET;
761	return rc;
762	}
763
764
765	RTDECL(uint32_t) RTSocketRetain(RTSOCKET hSocket)
766	{
767	RTSOCKETINT *pThis = hSocket;
768	AssertPtrReturn(pThis, UINT32_MAX);
769	AssertReturn(pThis->u32Magic == RTSOCKET_MAGIC, UINT32_MAX);
770	return RTMemPoolRetain(pThis);
771	}
772
773
774	/**
775	* Worker for RTSocketRelease and RTSocketClose.
776	*
777	* @returns IPRT status code.
778	* @param pThis The socket handle instance data.
779	* @param fDestroy Whether we're reaching ref count zero.
780	*/
781	static int rtSocketCloseIt(RTSOCKETINT *pThis, bool fDestroy)
782	{
783	/*
784	* Invalidate the handle structure on destroy.
785	*/
786	if (fDestroy)
787	{
788	Assert(ASMAtomicReadU32(&pThis->u32Magic) == RTSOCKET_MAGIC);
789	ASMAtomicWriteU32(&pThis->u32Magic, RTSOCKET_MAGIC_DEAD);
790	}
791
792	int rc = VINF_SUCCESS;
793	if (ASMAtomicCmpXchgBool(&pThis->fClosed, true, false))
794	{
795	#ifdef RT_OS_WINDOWS
796	/*
797	* Poke the polling thread if active and give it a small chance to stop.
798	*/
799	if ( pThis->fPollFallback
800	&& pThis->hPollFallbackThread != NIL_RTTHREAD)
801	{
802	ASMAtomicWriteBool(&pThis->fPollFallbackShutdown, true);
803	rtSocketPokePollFallbackThread(pThis);
804	int rc2 = RTThreadWait(pThis->hPollFallbackThread, RT_MS_1SEC, NULL);
805	if (RT_SUCCESS(rc2))
806	pThis->hPollFallbackThread = NIL_RTTHREAD;
807	}
808	#endif
809
810	/*
811	* Close the native handle.
812	*/
813	RTSOCKETNATIVE hNative = pThis->hNative;
814	if (hNative != NIL_RTSOCKETNATIVE)
815	{
816	pThis->hNative = NIL_RTSOCKETNATIVE;
817
818	if (!pThis->fLeaveOpen)
819	{
820	#ifdef RT_OS_WINDOWS
821	AssertReturn(g_pfnclosesocket, VERR_NET_NOT_UNSUPPORTED);
822	if (g_pfnclosesocket(hNative))
823	#else
824	if (close(hNative))
825	#endif
826	{
827	rc = rtSocketError();
828	#ifdef RT_OS_WINDOWS
829	AssertMsgFailed(("closesocket(%p) -> %Rrc\n", (uintptr_t)hNative, rc));
830	#else
831	AssertMsgFailed(("close(%d) -> %Rrc\n", hNative, rc));
832	#endif
833	}
834	}
835	}
836
837	#ifdef RT_OS_WINDOWS
838	/*
839	* Windows specific polling cleanup.
840	*/
841	WSAEVENT hEvent = pThis->hEvent;
842	if (hEvent != WSA_INVALID_EVENT)
843	{
844	pThis->hEvent = WSA_INVALID_EVENT;
845	if (!pThis->fPollFallback)
846	{
847	Assert(g_pfnWSACloseEvent);
848	if (g_pfnWSACloseEvent)
849	g_pfnWSACloseEvent(hEvent);
850	}
851	else
852	CloseHandle(hEvent);
853	}
854
855	if (pThis->fPollFallback)
856	{
857	if (pThis->hPollFallbackNotifyW != NIL_RTSOCKETNATIVE)
858	{
859	g_pfnclosesocket(pThis->hPollFallbackNotifyW);
860	pThis->hPollFallbackNotifyW = NIL_RTSOCKETNATIVE;
861	}
862
863	if (pThis->hPollFallbackThread != NIL_RTTHREAD)
864	{
865	int rc2 = RTThreadWait(pThis->hPollFallbackThread, RT_MS_1MIN / 2, NULL);
866	AssertRC(rc2);
867	pThis->hPollFallbackThread = NIL_RTTHREAD;
868	}
869
870	if (pThis->hPollFallbackNotifyR != NIL_RTSOCKETNATIVE)
871	{
872	g_pfnclosesocket(pThis->hPollFallbackNotifyR);
873	pThis->hPollFallbackNotifyR = NIL_RTSOCKETNATIVE;
874	}
875	}
876	#endif
877	}
878
879	return rc;
880	}
881
882
883	RTDECL(uint32_t) RTSocketRelease(RTSOCKET hSocket)
884	{
885	RTSOCKETINT *pThis = hSocket;
886	if (pThis == NIL_RTSOCKET)
887	return 0;
888	AssertPtrReturn(pThis, UINT32_MAX);
889	AssertReturn(pThis->u32Magic == RTSOCKET_MAGIC, UINT32_MAX);
890
891	/* get the refcount without killing it... */
892	uint32_t cRefs = RTMemPoolRefCount(pThis);
893	AssertReturn(cRefs != UINT32_MAX, UINT32_MAX);
894	if (cRefs == 1)
895	rtSocketCloseIt(pThis, true);
896
897	return RTMemPoolRelease(RTMEMPOOL_DEFAULT, pThis);
898	}
899
900
901	RTDECL(int) RTSocketClose(RTSOCKET hSocket)
902	{
903	RTSOCKETINT *pThis = hSocket;
904	if (pThis == NIL_RTSOCKET)
905	return VINF_SUCCESS;
906	AssertPtrReturn(pThis, VERR_INVALID_HANDLE);
907	AssertReturn(pThis->u32Magic == RTSOCKET_MAGIC, VERR_INVALID_HANDLE);
908
909	uint32_t cRefs = RTMemPoolRefCount(pThis);
910	AssertReturn(cRefs != UINT32_MAX, UINT32_MAX);
911
912	int rc = rtSocketCloseIt(pThis, cRefs == 1);
913
914	RTMemPoolRelease(RTMEMPOOL_DEFAULT, pThis);
915	return rc;
916	}
917
918
919	RTDECL(RTHCUINTPTR) RTSocketToNative(RTSOCKET hSocket)
920	{
921	RTSOCKETINT *pThis = hSocket;
922	AssertPtrReturn(pThis, RTHCUINTPTR_MAX);
923	AssertReturn(pThis->u32Magic == RTSOCKET_MAGIC, RTHCUINTPTR_MAX);
924	return (RTHCUINTPTR)pThis->hNative;
925	}
926
927
928	RTDECL(int) RTSocketSetInheritance(RTSOCKET hSocket, bool fInheritable)
929	{
930	RTSOCKETINT *pThis = hSocket;
931	AssertPtrReturn(pThis, VERR_INVALID_HANDLE);
932	AssertReturn(pThis->u32Magic == RTSOCKET_MAGIC, VERR_INVALID_HANDLE);
933	AssertReturn(RTMemPoolRefCount(pThis) >= (pThis->cUsers ? 2U : 1U), VERR_CALLER_NO_REFERENCE);
934
935	int rc = VINF_SUCCESS;
936	#ifdef RT_OS_WINDOWS
937	if (!SetHandleInformation((HANDLE)pThis->hNative, HANDLE_FLAG_INHERIT, fInheritable ? HANDLE_FLAG_INHERIT : 0))
938	rc = RTErrConvertFromWin32(GetLastError());
939	#else
940	if (fcntl(pThis->hNative, F_SETFD, fInheritable ? 0 : FD_CLOEXEC) < 0)
941	rc = RTErrConvertFromErrno(errno);
942	#endif
943
944	return rc;
945	}
946
947
948	static bool rtSocketIsIPv4Numerical(const char *pszAddress, PRTNETADDRIPV4 pAddr)
949	{
950
951	/* Empty address resolves to the INADDR_ANY address (good for bind). */
952	if (!pszAddress \|\| !*pszAddress)
953	{
954	pAddr->u = INADDR_ANY;
955	return true;
956	}
957
958	/* Four quads? */
959	char psz = (char )pszAddress;
960	for (int i = 0; i < 4; i++)
961	{
962	uint8_t u8;
963	int rc = RTStrToUInt8Ex(psz, &psz, 0, &u8);
964	if (rc != VINF_SUCCESS && rc != VWRN_TRAILING_CHARS)
965	return false;
966	if (*psz != (i < 3 ? '.' : '\0'))
967	return false;
968	psz++;
969
970	pAddr->au8[i] = u8; /* big endian */
971	}
972
973	return true;
974	}
975
976	RTDECL(int) RTSocketParseInetAddress(const char *pszAddress, unsigned uPort, PRTNETADDR pAddr)
977	{
978	int rc;
979
980	/*
981	* Validate input.
982	*/
983	AssertReturn(uPort > 0, VERR_INVALID_PARAMETER);
984	AssertPtrNullReturn(pszAddress, VERR_INVALID_POINTER);
985
986	/*
987	* Resolve the address. Pretty crude at the moment, but we have to make
988	* sure to not ask the NT 4 gethostbyname about an IPv4 address as it may
989	* give a wrong answer.
990	*/
991	/** @todo this only supports IPv4, and IPv6 support needs to be added.
992	* It probably needs to be converted to getaddrinfo(). */
993	RTNETADDRIPV4 IPv4Quad;
994	if (rtSocketIsIPv4Numerical(pszAddress, &IPv4Quad))
995	{
996	Log3(("rtSocketIsIPv4Numerical: %s -> %#x (%RTnaipv4)\n", pszAddress, IPv4Quad.u, IPv4Quad));
997	RT_ZERO(*pAddr);
998	pAddr->enmType = RTNETADDRTYPE_IPV4;
999	pAddr->uPort = uPort;
1000	pAddr->uAddr.IPv4 = IPv4Quad;
1001	return VINF_SUCCESS;
1002	}
1003
1004	#ifdef RT_OS_WINDOWS
1005	/* Initialize WinSock and check version before we call gethostbyname. */
1006	if (!g_pfngethostbyname)
1007	return VERR_NET_NOT_UNSUPPORTED;
1008
1009	int rc2 = rtSocketInitWinsock();
1010	if (RT_FAILURE(rc2))
1011	return rc2;
1012
1013	# define gethostbyname g_pfngethostbyname
1014	#endif
1015
1016	struct hostent *pHostEnt;
1017	pHostEnt = gethostbyname(pszAddress);
1018	if (!pHostEnt)
1019	{
1020	rc = rtSocketResolverError();
1021	AssertMsg(rc == VERR_NET_HOST_NOT_FOUND,
1022	("Could not resolve '%s', rc=%Rrc\n", pszAddress, rc));
1023	return rc;
1024	}
1025
1026	if (pHostEnt->h_addrtype == AF_INET)
1027	{
1028	RT_ZERO(*pAddr);
1029	pAddr->enmType = RTNETADDRTYPE_IPV4;
1030	pAddr->uPort = uPort;
1031	pAddr->uAddr.IPv4.u = ((struct in_addr *)pHostEnt->h_addr)->s_addr;
1032	Log3(("gethostbyname: %s -> %#x (%RTnaipv4)\n", pszAddress, pAddr->uAddr.IPv4.u, pAddr->uAddr.IPv4));
1033	}
1034	else
1035	return VERR_NET_ADDRESS_FAMILY_NOT_SUPPORTED;
1036
1037	#ifdef RT_OS_WINDOWS
1038	# undef gethostbyname
1039	#endif
1040	return VINF_SUCCESS;
1041	}
1042
1043
1044	/*
1045	* New function to allow both ipv4 and ipv6 addresses to be resolved.
1046	* Breaks compatibility with windows before 2000.
1047	*/
1048	RTDECL(int) RTSocketQueryAddressStr(const char pszHost, char pszResult, size_t *pcbResult, PRTNETADDRTYPE penmAddrType)
1049	{
1050	AssertPtrReturn(pszHost, VERR_INVALID_POINTER);
1051	AssertPtrReturn(pcbResult, VERR_INVALID_POINTER);
1052	AssertPtrNullReturn(penmAddrType, VERR_INVALID_POINTER);
1053	AssertPtrNullReturn(pszResult, VERR_INVALID_POINTER);
1054
1055	#if defined(RT_OS_OS2) \|\| defined(RT_OS_WINDOWS) /** @todo dynamically resolve the APIs not present in NT4! */
1056	return VERR_NOT_SUPPORTED;
1057
1058	#else
1059	int rc;
1060	if (*pcbResult < 16)
1061	return VERR_NET_ADDRESS_NOT_AVAILABLE;
1062
1063	/* Setup the hint. */
1064	struct addrinfo grHints;
1065	RT_ZERO(grHints);
1066	grHints.ai_socktype = 0;
1067	grHints.ai_flags = 0;
1068	grHints.ai_protocol = 0;
1069	grHints.ai_family = AF_UNSPEC;
1070	if (penmAddrType)
1071	{
1072	switch (*penmAddrType)
1073	{
1074	case RTNETADDRTYPE_INVALID:
1075	/grHints.ai_family = AF_UNSPEC;/
1076	break;
1077	case RTNETADDRTYPE_IPV4:
1078	grHints.ai_family = AF_INET;
1079	break;
1080	case RTNETADDRTYPE_IPV6:
1081	grHints.ai_family = AF_INET6;
1082	break;
1083	default:
1084	AssertFailedReturn(VERR_INVALID_PARAMETER);
1085	}
1086	}
1087
1088	# ifdef RT_OS_WINDOWS
1089	/*
1090	* Winsock2 init
1091	*/
1092	if ( !g_pfngetaddrinfo
1093	\|\| !g_pfnfreeaddrinfo)
1094	return VERR_NET_NOT_UNSUPPORTED;
1095
1096	int rc2 = rtSocketInitWinsock();
1097	if (RT_FAILURE(rc2))
1098	return rc2;
1099
1100	# define getaddrinfo g_pfngetaddrinfo
1101	# define freeaddrinfo g_pfnfreeaddrinfo
1102	# endif
1103
1104	/** @todo r=bird: getaddrinfo and freeaddrinfo breaks the additions on NT4. */
1105	struct addrinfo *pgrResults = NULL;
1106	rc = getaddrinfo(pszHost, "", &grHints, &pgrResults);
1107	if (rc != 0)
1108	return VERR_NET_ADDRESS_NOT_AVAILABLE;
1109
1110	// return data
1111	// on multiple matches return only the first one
1112
1113	if (!pgrResults)
1114	return VERR_NET_ADDRESS_NOT_AVAILABLE;
1115
1116	struct addrinfo const *pgrResult = pgrResults->ai_next;
1117	if (!pgrResult)
1118	{
1119	freeaddrinfo(pgrResults);
1120	return VERR_NET_ADDRESS_NOT_AVAILABLE;
1121	}
1122
1123	RTNETADDRTYPE enmAddrType = RTNETADDRTYPE_INVALID;
1124	size_t cchIpAddress;
1125	char szIpAddress[48];
1126	if (pgrResult->ai_family == AF_INET)
1127	{
1128	struct sockaddr_in const pgrSa = (struct sockaddr_in const )pgrResult->ai_addr;
1129	cchIpAddress = RTStrPrintf(szIpAddress, sizeof(szIpAddress),
1130	"%RTnaipv4", pgrSa->sin_addr.s_addr);
1131	Assert(cchIpAddress >= 7 && cchIpAddress < sizeof(szIpAddress) - 1);
1132	enmAddrType = RTNETADDRTYPE_IPV4;
1133	rc = VINF_SUCCESS;
1134	}
1135	else if (pgrResult->ai_family == AF_INET6)
1136	{
1137	struct sockaddr_in6 const pgrSa6 = (struct sockaddr_in6 const )pgrResult->ai_addr;
1138	cchIpAddress = RTStrPrintf(szIpAddress, sizeof(szIpAddress),
1139	"%RTnaipv6", (PRTNETADDRIPV6)&pgrSa6->sin6_addr);
1140	enmAddrType = RTNETADDRTYPE_IPV6;
1141	rc = VINF_SUCCESS;
1142	}
1143	else
1144	{
1145	rc = VERR_NET_ADDRESS_NOT_AVAILABLE;
1146	szIpAddress[0] = '\0';
1147	cchIpAddress = 0;
1148	}
1149	freeaddrinfo(pgrResults);
1150
1151	/*
1152	* Copy out the result.
1153	*/
1154	size_t const cbResult = *pcbResult;
1155	*pcbResult = cchIpAddress + 1;
1156	if (cchIpAddress < cbResult)
1157	memcpy(pszResult, szIpAddress, cchIpAddress + 1);
1158	else
1159	{
1160	RT_BZERO(pszResult, cbResult);
1161	if (RT_SUCCESS(rc))
1162	rc = VERR_BUFFER_OVERFLOW;
1163	}
1164	if (penmAddrType && RT_SUCCESS(rc))
1165	*penmAddrType = enmAddrType;
1166	return rc;
1167
1168	# ifdef RT_OS_WINDOWS
1169	# undef getaddrinfo
1170	# undef freeaddrinfo
1171	# endif
1172	#endif /* !RT_OS_OS2 */
1173	}
1174
1175
1176	RTDECL(int) RTSocketRead(RTSOCKET hSocket, void pvBuffer, size_t cbBuffer, size_t pcbRead)
1177	{
1178	/*
1179	* Validate input.
1180	*/
1181	RTSOCKETINT *pThis = hSocket;
1182	AssertPtrReturn(pThis, VERR_INVALID_HANDLE);
1183	AssertReturn(pThis->u32Magic == RTSOCKET_MAGIC, VERR_INVALID_HANDLE);
1184	AssertReturn(cbBuffer > 0, VERR_INVALID_PARAMETER);
1185	AssertPtr(pvBuffer);
1186	#ifdef RT_OS_WINDOWS
1187	AssertReturn(g_pfnrecv, VERR_NET_NOT_UNSUPPORTED);
1188	# define recv g_pfnrecv
1189	#endif
1190	AssertReturn(rtSocketTryLock(pThis), VERR_CONCURRENT_ACCESS);
1191
1192	int rc = rtSocketSwitchBlockingMode(pThis, true /* fBlocking */);
1193	if (RT_FAILURE(rc))
1194	return rc;
1195
1196	/*
1197	* Read loop.
1198	* If pcbRead is NULL we have to fill the entire buffer!
1199	*/
1200	size_t cbRead = 0;
1201	size_t cbToRead = cbBuffer;
1202	for (;;)
1203	{
1204	rtSocketErrorReset();
1205	#ifdef RTSOCKET_MAX_READ
1206	int cbNow = cbToRead >= RTSOCKET_MAX_READ ? RTSOCKET_MAX_READ : (int)cbToRead;
1207	#else
1208	size_t cbNow = cbToRead;
1209	#endif
1210	ssize_t cbBytesRead = recv(pThis->hNative, (char *)pvBuffer + cbRead, cbNow, MSG_NOSIGNAL);
1211	if (cbBytesRead <= 0)
1212	{
1213	rc = rtSocketError();
1214	Assert(RT_FAILURE_NP(rc) \|\| cbBytesRead == 0);
1215	if (RT_SUCCESS_NP(rc))
1216	{
1217	if (!pcbRead)
1218	rc = VERR_NET_SHUTDOWN;
1219	else
1220	{
1221	*pcbRead = 0;
1222	rc = VINF_SUCCESS;
1223	}
1224	}
1225	break;
1226	}
1227	if (pcbRead)
1228	{
1229	/* return partial data */
1230	*pcbRead = cbBytesRead;
1231	break;
1232	}
1233
1234	/* read more? */
1235	cbRead += cbBytesRead;
1236	if (cbRead == cbBuffer)
1237	break;
1238
1239	/* next */
1240	cbToRead = cbBuffer - cbRead;
1241	}
1242
1243	rtSocketUnlock(pThis);
1244	#ifdef RT_OS_WINDOWS
1245	# undef recv
1246	#endif
1247	return rc;
1248	}
1249
1250
1251	RTDECL(int) RTSocketReadFrom(RTSOCKET hSocket, void pvBuffer, size_t cbBuffer, size_t pcbRead, PRTNETADDR pSrcAddr)
1252	{
1253	/*
1254	* Validate input.
1255	*/
1256	RTSOCKETINT *pThis = hSocket;
1257	AssertPtrReturn(pThis, VERR_INVALID_HANDLE);
1258	AssertReturn(pThis->u32Magic == RTSOCKET_MAGIC, VERR_INVALID_HANDLE);
1259	AssertReturn(cbBuffer > 0, VERR_INVALID_PARAMETER);
1260	AssertPtr(pvBuffer);
1261	AssertPtr(pcbRead);
1262	#ifdef RT_OS_WINDOWS
1263	AssertReturn(g_pfnrecvfrom, VERR_NET_NOT_UNSUPPORTED);
1264	# define recvfrom g_pfnrecvfrom
1265	#endif
1266	AssertReturn(rtSocketTryLock(pThis), VERR_CONCURRENT_ACCESS);
1267
1268	int rc = rtSocketSwitchBlockingMode(pThis, true /* fBlocking */);
1269	if (RT_FAILURE(rc))
1270	return rc;
1271
1272	/*
1273	* Read data.
1274	*/
1275	size_t cbRead = 0;
1276	size_t cbToRead = cbBuffer;
1277	rtSocketErrorReset();
1278	RTSOCKADDRUNION u;
1279	#ifdef RTSOCKET_MAX_READ
1280	int cbNow = cbToRead >= RTSOCKET_MAX_READ ? RTSOCKET_MAX_READ : (int)cbToRead;
1281	int cbAddr = sizeof(u);
1282	#else
1283	size_t cbNow = cbToRead;
1284	socklen_t cbAddr = sizeof(u);
1285	#endif
1286	ssize_t cbBytesRead = recvfrom(pThis->hNative, (char *)pvBuffer + cbRead, cbNow, MSG_NOSIGNAL, &u.Addr, &cbAddr);
1287	if (cbBytesRead <= 0)
1288	{
1289	rc = rtSocketError();
1290	Assert(RT_FAILURE_NP(rc) \|\| cbBytesRead == 0);
1291	if (RT_SUCCESS_NP(rc))
1292	{
1293	*pcbRead = 0;
1294	rc = VINF_SUCCESS;
1295	}
1296	}
1297	else
1298	{
1299	if (pSrcAddr)
1300	rc = rtSocketNetAddrFromAddr(&u, cbAddr, pSrcAddr);
1301	*pcbRead = cbBytesRead;
1302	}
1303
1304	rtSocketUnlock(pThis);
1305	#ifdef RT_OS_WINDOWS
1306	# undef recvfrom
1307	#endif
1308	return rc;
1309	}
1310
1311
1312	RTDECL(int) RTSocketWrite(RTSOCKET hSocket, const void *pvBuffer, size_t cbBuffer)
1313	{
1314	/*
1315	* Validate input.
1316	*/
1317	RTSOCKETINT *pThis = hSocket;
1318	AssertPtrReturn(pThis, VERR_INVALID_HANDLE);
1319	AssertReturn(pThis->u32Magic == RTSOCKET_MAGIC, VERR_INVALID_HANDLE);
1320	#ifdef RT_OS_WINDOWS
1321	AssertReturn(g_pfnsend, VERR_NET_NOT_UNSUPPORTED);
1322	# define send g_pfnsend
1323	#endif
1324	AssertReturn(rtSocketTryLock(pThis), VERR_CONCURRENT_ACCESS);
1325
1326	int rc = rtSocketSwitchBlockingMode(pThis, true /* fBlocking */);
1327	if (RT_FAILURE(rc))
1328	return rc;
1329
1330	/*
1331	* Try write all at once.
1332	*/
1333	#ifdef RTSOCKET_MAX_WRITE
1334	int cbNow = cbBuffer >= RTSOCKET_MAX_WRITE ? RTSOCKET_MAX_WRITE : (int)cbBuffer;
1335	#else
1336	size_t cbNow = cbBuffer >= SSIZE_MAX ? SSIZE_MAX : cbBuffer;
1337	#endif
1338	ssize_t cbWritten = send(pThis->hNative, (const char *)pvBuffer, cbNow, MSG_NOSIGNAL);
1339	if (RT_LIKELY((size_t)cbWritten == cbBuffer && cbWritten >= 0))
1340	rc = VINF_SUCCESS;
1341	else if (cbWritten < 0)
1342	rc = rtSocketError();
1343	else
1344	{
1345	/*
1346	* Unfinished business, write the remainder of the request. Must ignore
1347	* VERR_INTERRUPTED here if we've managed to send something.
1348	*/
1349	size_t cbSentSoFar = 0;
1350	for (;;)
1351	{
1352	/* advance */
1353	cbBuffer -= (size_t)cbWritten;
1354	if (!cbBuffer)
1355	break;
1356	cbSentSoFar += (size_t)cbWritten;
1357	pvBuffer = (char const *)pvBuffer + cbWritten;
1358
1359	/* send */
1360	#ifdef RTSOCKET_MAX_WRITE
1361	cbNow = cbBuffer >= RTSOCKET_MAX_WRITE ? RTSOCKET_MAX_WRITE : (int)cbBuffer;
1362	#else
1363	cbNow = cbBuffer >= SSIZE_MAX ? SSIZE_MAX : cbBuffer;
1364	#endif
1365	cbWritten = send(pThis->hNative, (const char *)pvBuffer, cbNow, MSG_NOSIGNAL);
1366	if (cbWritten >= 0)
1367	AssertMsg(cbBuffer >= (size_t)cbWritten, ("Wrote more than we requested!!! cbWritten=%zu cbBuffer=%zu rtSocketError()=%d\n",
1368	cbWritten, cbBuffer, rtSocketError()));
1369	else
1370	{
1371	rc = rtSocketError();
1372	if (rc != VERR_INTERNAL_ERROR \|\| cbSentSoFar == 0)
1373	break;
1374	cbWritten = 0;
1375	rc = VINF_SUCCESS;
1376	}
1377	}
1378	}
1379
1380	rtSocketUnlock(pThis);
1381	#ifdef RT_OS_WINDOWS
1382	# undef send
1383	#endif
1384	return rc;
1385	}
1386
1387
1388	RTDECL(int) RTSocketWriteTo(RTSOCKET hSocket, const void *pvBuffer, size_t cbBuffer, PCRTNETADDR pAddr)
1389	{
1390	/*
1391	* Validate input.
1392	*/
1393	RTSOCKETINT *pThis = hSocket;
1394	AssertPtrReturn(pThis, VERR_INVALID_HANDLE);
1395	AssertReturn(pThis->u32Magic == RTSOCKET_MAGIC, VERR_INVALID_HANDLE);
1396	#ifdef RT_OS_WINDOWS
1397	AssertReturn(g_pfnsendto, VERR_NET_NOT_UNSUPPORTED);
1398	# define sendto g_pfnsendto
1399	#endif
1400
1401	/* no locking since UDP reads may be done concurrently to writes, and
1402	* this is the normal use case of this code. */
1403
1404	int rc = rtSocketSwitchBlockingMode(pThis, true /* fBlocking */);
1405	if (RT_FAILURE(rc))
1406	return rc;
1407
1408	/* Figure out destination address. */
1409	struct sockaddr *pSA = NULL;
1410	#ifdef RT_OS_WINDOWS
1411	int cbSA = 0;
1412	#else
1413	socklen_t cbSA = 0;
1414	#endif
1415	RTSOCKADDRUNION u;
1416	if (pAddr)
1417	{
1418	rc = rtSocketAddrFromNetAddr(pAddr, &u, sizeof(u), NULL);
1419	if (RT_FAILURE(rc))
1420	return rc;
1421	pSA = &u.Addr;
1422	cbSA = sizeof(u);
1423	}
1424
1425	/*
1426	* Must write all at once, otherwise it is a failure.
1427	*/
1428	#ifdef RT_OS_WINDOWS
1429	int cbNow = cbBuffer >= RTSOCKET_MAX_WRITE ? RTSOCKET_MAX_WRITE : (int)cbBuffer;
1430	#else
1431	size_t cbNow = cbBuffer >= SSIZE_MAX ? SSIZE_MAX : cbBuffer;
1432	#endif
1433	ssize_t cbWritten = sendto(pThis->hNative, (const char *)pvBuffer, cbNow, MSG_NOSIGNAL, pSA, cbSA);
1434	if (RT_LIKELY((size_t)cbWritten == cbBuffer && cbWritten >= 0))
1435	rc = VINF_SUCCESS;
1436	else if (cbWritten < 0)
1437	rc = rtSocketError();
1438	else
1439	rc = VERR_TOO_MUCH_DATA;
1440
1441	/// @todo rtSocketUnlock(pThis);
1442	#ifdef RT_OS_WINDOWS
1443	# undef sendto
1444	#endif
1445	return rc;
1446	}
1447
1448
1449	RTDECL(int) RTSocketWriteToNB(RTSOCKET hSocket, const void *pvBuffer, size_t cbBuffer, PCRTNETADDR pAddr)
1450	{
1451	/*
1452	* Validate input.
1453	*/
1454	RTSOCKETINT *pThis = hSocket;
1455	AssertPtrReturn(pThis, VERR_INVALID_HANDLE);
1456	AssertReturn(pThis->u32Magic == RTSOCKET_MAGIC, VERR_INVALID_HANDLE);
1457	#ifdef RT_OS_WINDOWS
1458	AssertReturn(g_pfnsendto, VERR_NET_NOT_UNSUPPORTED);
1459	# define sendto g_pfnsendto
1460	#endif
1461
1462	/* no locking since UDP reads may be done concurrently to writes, and
1463	* this is the normal use case of this code. */
1464
1465	int rc = rtSocketSwitchBlockingMode(pThis, false /* fBlocking */);
1466	if (RT_FAILURE(rc))
1467	return rc;
1468
1469	/* Figure out destination address. */
1470	struct sockaddr *pSA = NULL;
1471	#ifdef RT_OS_WINDOWS
1472	int cbSA = 0;
1473	#else
1474	socklen_t cbSA = 0;
1475	#endif
1476	RTSOCKADDRUNION u;
1477	if (pAddr)
1478	{
1479	rc = rtSocketAddrFromNetAddr(pAddr, &u, sizeof(u), NULL);
1480	if (RT_FAILURE(rc))
1481	return rc;
1482	pSA = &u.Addr;
1483	cbSA = sizeof(u);
1484	}
1485
1486	/*
1487	* Must write all at once, otherwise it is a failure.
1488	*/
1489	#ifdef RT_OS_WINDOWS
1490	int cbNow = cbBuffer >= RTSOCKET_MAX_WRITE ? RTSOCKET_MAX_WRITE : (int)cbBuffer;
1491	#else
1492	size_t cbNow = cbBuffer >= SSIZE_MAX ? SSIZE_MAX : cbBuffer;
1493	#endif
1494	ssize_t cbWritten = sendto(pThis->hNative, (const char *)pvBuffer, cbNow, MSG_NOSIGNAL, pSA, cbSA);
1495	if (RT_LIKELY((size_t)cbWritten == cbBuffer && cbWritten >= 0))
1496	rc = VINF_SUCCESS;
1497	else if (cbWritten < 0)
1498	rc = rtSocketError();
1499	else
1500	rc = VERR_TOO_MUCH_DATA;
1501
1502	/// @todo rtSocketUnlock(pThis);
1503	#ifdef RT_OS_WINDOWS
1504	# undef sendto
1505	#endif
1506	return rc;
1507	}
1508
1509
1510	RTDECL(int) RTSocketSgWrite(RTSOCKET hSocket, PCRTSGBUF pSgBuf)
1511	{
1512	/*
1513	* Validate input.
1514	*/
1515	RTSOCKETINT *pThis = hSocket;
1516	AssertPtrReturn(pThis, VERR_INVALID_HANDLE);
1517	AssertReturn(pThis->u32Magic == RTSOCKET_MAGIC, VERR_INVALID_HANDLE);
1518	AssertPtrReturn(pSgBuf, VERR_INVALID_PARAMETER);
1519	AssertReturn(pSgBuf->cSegs > 0, VERR_INVALID_PARAMETER);
1520	AssertReturn(rtSocketTryLock(pThis), VERR_CONCURRENT_ACCESS);
1521
1522	int rc = rtSocketSwitchBlockingMode(pThis, true /* fBlocking */);
1523	if (RT_FAILURE(rc))
1524	return rc;
1525
1526	/*
1527	* Construct message descriptor (translate pSgBuf) and send it.
1528	*/
1529	rc = VERR_NO_TMP_MEMORY;
1530	#ifdef RT_OS_WINDOWS
1531	if (g_pfnWSASend)
1532	{
1533	AssertCompileSize(WSABUF, sizeof(RTSGSEG));
1534	AssertCompileMemberSize(WSABUF, buf, RT_SIZEOFMEMB(RTSGSEG, pvSeg));
1535
1536	LPWSABUF paMsg = (LPWSABUF)RTMemTmpAllocZ(pSgBuf->cSegs * sizeof(WSABUF));
1537	if (paMsg)
1538	{
1539	for (unsigned i = 0; i < pSgBuf->cSegs; i++)
1540	{
1541	paMsg[i].buf = (char *)pSgBuf->paSegs[i].pvSeg;
1542	paMsg[i].len = (u_long)pSgBuf->paSegs[i].cbSeg;
1543	}
1544
1545	DWORD dwSent;
1546	int hrc = g_pfnWSASend(pThis->hNative, paMsg, pSgBuf->cSegs, &dwSent, MSG_NOSIGNAL, NULL, NULL);
1547	if (!hrc)
1548	rc = VINF_SUCCESS;
1549	/** @todo check for incomplete writes */
1550	else
1551	rc = rtSocketError();
1552
1553	RTMemTmpFree(paMsg);
1554	}
1555	}
1556	else if (g_pfnsend)
1557	{
1558	rc = VINF_SUCCESS;
1559	for (uint32_t iSeg = 0; iSeg < pSgBuf->cSegs; iSeg++)
1560	{
1561	uint8_t const pbSeg = (uint8_t const )pSgBuf->paSegs[iSeg].pvSeg;
1562	size_t cbSeg = pSgBuf->paSegs[iSeg].cbSeg;
1563	int cbNow;
1564	ssize_t cbWritten;
1565	for (;;)
1566	{
1567	cbNow = cbSeg >= RTSOCKET_MAX_WRITE ? RTSOCKET_MAX_WRITE : (int)cbSeg;
1568	cbWritten = g_pfnsend(pThis->hNative, (const char *)pbSeg, cbNow, MSG_NOSIGNAL);
1569	if ((size_t)cbWritten >= cbSeg \|\| cbWritten < 0)
1570	break;
1571	pbSeg += cbWritten;
1572	cbSeg -= cbWritten;
1573	}
1574	if (cbWritten < 0)
1575	{
1576	rc = rtSocketError();
1577	break;
1578	}
1579	}
1580	}
1581	else
1582	rc = VERR_NET_NOT_UNSUPPORTED;
1583
1584	#else /* !RT_OS_WINDOWS */
1585	AssertCompileSize(struct iovec, sizeof(RTSGSEG));
1586	AssertCompileMemberSize(struct iovec, iov_base, RT_SIZEOFMEMB(RTSGSEG, pvSeg));
1587	AssertCompileMemberSize(struct iovec, iov_len, RT_SIZEOFMEMB(RTSGSEG, cbSeg));
1588
1589	struct iovec paMsg = (struct iovec )RTMemTmpAllocZ(pSgBuf->cSegs * sizeof(struct iovec));
1590	if (paMsg)
1591	{
1592	for (unsigned i = 0; i < pSgBuf->cSegs; i++)
1593	{
1594	paMsg[i].iov_base = pSgBuf->paSegs[i].pvSeg;
1595	paMsg[i].iov_len = pSgBuf->paSegs[i].cbSeg;
1596	}
1597
1598	struct msghdr msgHdr;
1599	RT_ZERO(msgHdr);
1600	msgHdr.msg_iov = paMsg;
1601	msgHdr.msg_iovlen = pSgBuf->cSegs;
1602	ssize_t cbWritten = sendmsg(pThis->hNative, &msgHdr, MSG_NOSIGNAL);
1603	if (RT_LIKELY(cbWritten >= 0))
1604	rc = VINF_SUCCESS;
1605	/** @todo check for incomplete writes */
1606	else
1607	rc = rtSocketError();
1608
1609	RTMemTmpFree(paMsg);
1610	}
1611	#endif /* !RT_OS_WINDOWS */
1612
1613	rtSocketUnlock(pThis);
1614	return rc;
1615	}
1616
1617
1618	RTDECL(int) RTSocketSgWriteL(RTSOCKET hSocket, size_t cSegs, ...)
1619	{
1620	va_list va;
1621	va_start(va, cSegs);
1622	int rc = RTSocketSgWriteLV(hSocket, cSegs, va);
1623	va_end(va);
1624	return rc;
1625	}
1626
1627
1628	RTDECL(int) RTSocketSgWriteLV(RTSOCKET hSocket, size_t cSegs, va_list va)
1629	{
1630	/*
1631	* Set up a S/G segment array + buffer on the stack and pass it
1632	* on to RTSocketSgWrite.
1633	*/
1634	Assert(cSegs <= 16);
1635	PRTSGSEG paSegs = (PRTSGSEG)alloca(cSegs * sizeof(RTSGSEG));
1636	AssertReturn(paSegs, VERR_NO_TMP_MEMORY);
1637	for (size_t i = 0; i < cSegs; i++)
1638	{
1639	paSegs[i].pvSeg = va_arg(va, void *);
1640	paSegs[i].cbSeg = va_arg(va, size_t);
1641	}
1642
1643	RTSGBUF SgBuf;
1644	RTSgBufInit(&SgBuf, paSegs, cSegs);
1645	return RTSocketSgWrite(hSocket, &SgBuf);
1646	}
1647
1648
1649	RTDECL(int) RTSocketReadNB(RTSOCKET hSocket, void pvBuffer, size_t cbBuffer, size_t pcbRead)
1650	{
1651	/*
1652	* Validate input.
1653	*/
1654	RTSOCKETINT *pThis = hSocket;
1655	AssertPtrReturn(pThis, VERR_INVALID_HANDLE);
1656	AssertReturn(pThis->u32Magic == RTSOCKET_MAGIC, VERR_INVALID_HANDLE);
1657	AssertReturn(cbBuffer > 0, VERR_INVALID_PARAMETER);
1658	AssertPtr(pvBuffer);
1659	AssertPtrReturn(pcbRead, VERR_INVALID_PARAMETER);
1660	#ifdef RT_OS_WINDOWS
1661	AssertReturn(g_pfnrecv, VERR_NET_NOT_UNSUPPORTED);
1662	#endif
1663	AssertReturn(rtSocketTryLock(pThis), VERR_CONCURRENT_ACCESS);
1664
1665	int rc = rtSocketSwitchBlockingMode(pThis, false /* fBlocking */);
1666	if (RT_FAILURE(rc))
1667	return rc;
1668
1669	rtSocketErrorReset();
1670	#ifdef RTSOCKET_MAX_READ
1671	int cbNow = cbBuffer >= RTSOCKET_MAX_WRITE ? RTSOCKET_MAX_WRITE : (int)cbBuffer;
1672	#else
1673	size_t cbNow = cbBuffer;
1674	#endif
1675
1676	#ifdef RT_OS_WINDOWS
1677	int cbRead = g_pfnrecv(pThis->hNative, (char *)pvBuffer, cbNow, MSG_NOSIGNAL);
1678	if (cbRead >= 0)
1679	{
1680	*pcbRead = cbRead;
1681	rc = VINF_SUCCESS;
1682	}
1683	else
1684	{
1685	rc = rtSocketError();
1686	if (rc == VERR_TRY_AGAIN)
1687	{
1688	*pcbRead = 0;
1689	rc = VINF_TRY_AGAIN;
1690	}
1691	}
1692
1693	#else
1694	ssize_t cbRead = recv(pThis->hNative, pvBuffer, cbNow, MSG_NOSIGNAL);
1695	if (cbRead >= 0)
1696	*pcbRead = cbRead;
1697	else if ( errno == EAGAIN
1698	# ifdef EWOULDBLOCK
1699	# if EWOULDBLOCK != EAGAIN
1700	\|\| errno == EWOULDBLOCK
1701	# endif
1702	# endif
1703	)
1704	{
1705	*pcbRead = 0;
1706	rc = VINF_TRY_AGAIN;
1707	}
1708	else
1709	rc = rtSocketError();
1710	#endif
1711
1712	rtSocketUnlock(pThis);
1713	return rc;
1714	}
1715
1716
1717	RTDECL(int) RTSocketWriteNB(RTSOCKET hSocket, const void pvBuffer, size_t cbBuffer, size_t pcbWritten)
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	AssertPtrReturn(pcbWritten, VERR_INVALID_PARAMETER);
1726	#ifdef RT_OS_WINDOWS
1727	AssertReturn(g_pfnsend, VERR_NET_NOT_UNSUPPORTED);
1728	#endif
1729	AssertReturn(rtSocketTryLock(pThis), VERR_CONCURRENT_ACCESS);
1730
1731	int rc = rtSocketSwitchBlockingMode(pThis, false /* fBlocking */);
1732	if (RT_FAILURE(rc))
1733	return rc;
1734
1735	rtSocketErrorReset();
1736	#ifdef RT_OS_WINDOWS
1737	# ifdef RTSOCKET_MAX_WRITE
1738	int cbNow = cbBuffer >= RTSOCKET_MAX_WRITE ? RTSOCKET_MAX_WRITE : (int)cbBuffer;
1739	# else
1740	size_t cbNow = cbBuffer;
1741	# endif
1742	int cbWritten = g_pfnsend(pThis->hNative, (const char *)pvBuffer, cbNow, MSG_NOSIGNAL);
1743	if (cbWritten >= 0)
1744	{
1745	*pcbWritten = cbWritten;
1746	rc = VINF_SUCCESS;
1747	}
1748	else
1749	{
1750	rc = rtSocketError();
1751	if (rc == VERR_TRY_AGAIN)
1752	{
1753	*pcbWritten = 0;
1754	rc = VINF_TRY_AGAIN;
1755	}
1756	}
1757	#else
1758	ssize_t cbWritten = send(pThis->hNative, pvBuffer, cbBuffer, MSG_NOSIGNAL);
1759	if (cbWritten >= 0)
1760	*pcbWritten = cbWritten;
1761	else if ( errno == EAGAIN
1762	# ifdef EWOULDBLOCK
1763	# if EWOULDBLOCK != EAGAIN
1764	\|\| errno == EWOULDBLOCK
1765	# endif
1766	# endif
1767	)
1768	{
1769	*pcbWritten = 0;
1770	rc = VINF_TRY_AGAIN;
1771	}
1772	else
1773	rc = rtSocketError();
1774	#endif
1775
1776	rtSocketUnlock(pThis);
1777	return rc;
1778	}
1779
1780
1781	RTDECL(int) RTSocketSgWriteNB(RTSOCKET hSocket, PCRTSGBUF pSgBuf, size_t *pcbWritten)
1782	{
1783	/*
1784	* Validate input.
1785	*/
1786	RTSOCKETINT *pThis = hSocket;
1787	AssertPtrReturn(pThis, VERR_INVALID_HANDLE);
1788	AssertReturn(pThis->u32Magic == RTSOCKET_MAGIC, VERR_INVALID_HANDLE);
1789	AssertPtrReturn(pSgBuf, VERR_INVALID_PARAMETER);
1790	AssertPtrReturn(pcbWritten, VERR_INVALID_PARAMETER);
1791	AssertReturn(pSgBuf->cSegs > 0, VERR_INVALID_PARAMETER);
1792	AssertReturn(rtSocketTryLock(pThis), VERR_CONCURRENT_ACCESS);
1793
1794	int rc = rtSocketSwitchBlockingMode(pThis, false /* fBlocking */);
1795	if (RT_FAILURE(rc))
1796	return rc;
1797
1798	unsigned cSegsToSend = 0;
1799	rc = VERR_NO_TMP_MEMORY;
1800	#ifdef RT_OS_WINDOWS
1801	if (g_pfnWSASend)
1802	{
1803	LPWSABUF paMsg = NULL;
1804	RTSgBufMapToNative(paMsg, pSgBuf, WSABUF, buf, char *, len, u_long, cSegsToSend);
1805	if (paMsg)
1806	{
1807	DWORD dwSent = 0;
1808	int hrc = g_pfnWSASend(pThis->hNative, paMsg, cSegsToSend, &dwSent, MSG_NOSIGNAL, NULL, NULL);
1809	if (!hrc)
1810	rc = VINF_SUCCESS;
1811	else
1812	rc = rtSocketError();
1813
1814	*pcbWritten = dwSent;
1815
1816	RTMemTmpFree(paMsg);
1817	}
1818	}
1819	else if (g_pfnsend)
1820	{
1821	size_t cbWrittenTotal = 0;
1822	rc = VINF_SUCCESS;
1823	for (uint32_t iSeg = 0; iSeg < pSgBuf->cSegs; iSeg++)
1824	{
1825	uint8_t const pbSeg = (uint8_t const )pSgBuf->paSegs[iSeg].pvSeg;
1826	size_t cbSeg = pSgBuf->paSegs[iSeg].cbSeg;
1827	int cbNow;
1828	ssize_t cbWritten;
1829	for (;;)
1830	{
1831	cbNow = cbSeg >= RTSOCKET_MAX_WRITE ? RTSOCKET_MAX_WRITE : (int)cbSeg;
1832	cbWritten = g_pfnsend(pThis->hNative, (const char *)pbSeg, cbNow, MSG_NOSIGNAL);
1833	if ((size_t)cbWritten >= cbSeg \|\| cbWritten < 0)
1834	break;
1835	cbWrittenTotal += cbWrittenTotal;
1836	pbSeg += cbWritten;
1837	cbSeg -= cbWritten;
1838	}
1839	if (cbWritten < 0)
1840	{
1841	rc = rtSocketError();
1842	break;
1843	}
1844	if (cbWritten != cbNow)
1845	break;
1846	}
1847	*pcbWritten = cbWrittenTotal;
1848	}
1849	else
1850	rc = VERR_NET_NOT_UNSUPPORTED;
1851
1852	#else /* !RT_OS_WINDOWS */
1853	struct iovec *paMsg = NULL;
1854
1855	RTSgBufMapToNative(paMsg, pSgBuf, struct iovec, iov_base, void *, iov_len, size_t, cSegsToSend);
1856	if (paMsg)
1857	{
1858	struct msghdr msgHdr;
1859	RT_ZERO(msgHdr);
1860	msgHdr.msg_iov = paMsg;
1861	msgHdr.msg_iovlen = cSegsToSend;
1862	ssize_t cbWritten = sendmsg(pThis->hNative, &msgHdr, MSG_NOSIGNAL);
1863	if (RT_LIKELY(cbWritten >= 0))
1864	{
1865	rc = VINF_SUCCESS;
1866	*pcbWritten = cbWritten;
1867	}
1868	else
1869	rc = rtSocketError();
1870
1871	RTMemTmpFree(paMsg);
1872	}
1873	#endif /* !RT_OS_WINDOWS */
1874
1875	rtSocketUnlock(pThis);
1876	return rc;
1877	}
1878
1879
1880	RTDECL(int) RTSocketSgWriteLNB(RTSOCKET hSocket, size_t cSegs, size_t *pcbWritten, ...)
1881	{
1882	va_list va;
1883	va_start(va, pcbWritten);
1884	int rc = RTSocketSgWriteLVNB(hSocket, cSegs, pcbWritten, va);
1885	va_end(va);
1886	return rc;
1887	}
1888
1889
1890	RTDECL(int) RTSocketSgWriteLVNB(RTSOCKET hSocket, size_t cSegs, size_t *pcbWritten, va_list va)
1891	{
1892	/*
1893	* Set up a S/G segment array + buffer on the stack and pass it
1894	* on to RTSocketSgWrite.
1895	*/
1896	Assert(cSegs <= 16);
1897	PRTSGSEG paSegs = (PRTSGSEG)alloca(cSegs * sizeof(RTSGSEG));
1898	AssertReturn(paSegs, VERR_NO_TMP_MEMORY);
1899	for (size_t i = 0; i < cSegs; i++)
1900	{
1901	paSegs[i].pvSeg = va_arg(va, void *);
1902	paSegs[i].cbSeg = va_arg(va, size_t);
1903	}
1904
1905	RTSGBUF SgBuf;
1906	RTSgBufInit(&SgBuf, paSegs, cSegs);
1907	return RTSocketSgWriteNB(hSocket, &SgBuf, pcbWritten);
1908	}
1909
1910
1911	RTDECL(int) RTSocketSelectOne(RTSOCKET hSocket, RTMSINTERVAL cMillies)
1912	{
1913	/*
1914	* Validate input.
1915	*/
1916	RTSOCKETINT *pThis = hSocket;
1917	AssertPtrReturn(pThis, VERR_INVALID_HANDLE);
1918	AssertReturn(pThis->u32Magic == RTSOCKET_MAGIC, VERR_INVALID_HANDLE);
1919	AssertReturn(RTMemPoolRefCount(pThis) >= (pThis->cUsers ? 2U : 1U), VERR_CALLER_NO_REFERENCE);
1920	int const fdMax = (int)pThis->hNative + 1;
1921	AssertReturn((RTSOCKETNATIVE)(fdMax - 1) == pThis->hNative, VERR_INTERNAL_ERROR_5);
1922	#ifdef RT_OS_WINDOWS
1923	AssertReturn(g_pfnselect, VERR_NET_NOT_UNSUPPORTED);
1924	# define select g_pfnselect
1925	#endif
1926
1927	/*
1928	* Set up the file descriptor sets and do the select.
1929	*/
1930	fd_set fdsetR;
1931	FD_ZERO(&fdsetR);
1932	FD_SET(pThis->hNative, &fdsetR);
1933
1934	fd_set fdsetE = fdsetR;
1935
1936	int rc;
1937	if (cMillies == RT_INDEFINITE_WAIT)
1938	rc = select(fdMax, &fdsetR, NULL, &fdsetE, NULL);
1939	else
1940	{
1941	struct timeval timeout;
1942	timeout.tv_sec = cMillies / 1000;
1943	timeout.tv_usec = (cMillies % 1000) * 1000;
1944	rc = select(fdMax, &fdsetR, NULL, &fdsetE, &timeout);
1945	}
1946	if (rc > 0)
1947	rc = VINF_SUCCESS;
1948	else if (rc == 0)
1949	rc = VERR_TIMEOUT;
1950	else
1951	rc = rtSocketError();
1952
1953	#ifdef RT_OS_WINDOWS
1954	# undef select
1955	#endif
1956	return rc;
1957	}
1958
1959
1960	/**
1961	* Internal worker for RTSocketSelectOneEx and rtSocketPollCheck (fallback)
1962	*
1963	* @returns IPRT status code
1964	* @param pThis The socket (valid).
1965	* @param fEvents The events to select for.
1966	* @param pfEvents Where to return the events.
1967	* @param cMillies How long to select for, in milliseconds.
1968	*/
1969	static int rtSocketSelectOneEx(RTSOCKET pThis, uint32_t fEvents, uint32_t *pfEvents, RTMSINTERVAL cMillies)
1970	{
1971	RTSOCKETNATIVE hNative = pThis->hNative;
1972	if (hNative == NIL_RTSOCKETNATIVE)
1973	{
1974	/* Socket is already closed? Possible we raced someone calling rtSocketCloseIt.
1975	Should we return a different status code? */
1976	*pfEvents = RTSOCKET_EVT_ERROR;
1977	return VINF_SUCCESS;
1978	}
1979
1980	int const fdMax = (int)hNative + 1;
1981	AssertReturn((RTSOCKETNATIVE)(fdMax - 1) == hNative, VERR_INTERNAL_ERROR_5);
1982	#ifdef RT_OS_WINDOWS
1983	AssertReturn(g_pfnselect, VERR_NET_NOT_UNSUPPORTED);
1984	AssertReturn(g_pfn__WSAFDIsSet, VERR_NET_NOT_UNSUPPORTED);
1985	# define select g_pfnselect
1986	# define __WSAFDIsSet g_pfn__WSAFDIsSet
1987	#endif
1988
1989	*pfEvents = 0;
1990
1991	/*
1992	* Set up the file descriptor sets and do the select.
1993	*/
1994	fd_set fdsetR;
1995	fd_set fdsetW;
1996	fd_set fdsetE;
1997	FD_ZERO(&fdsetR);
1998	FD_ZERO(&fdsetW);
1999	FD_ZERO(&fdsetE);
2000
2001	if (fEvents & RTSOCKET_EVT_READ)
2002	FD_SET(hNative, &fdsetR);
2003	if (fEvents & RTSOCKET_EVT_WRITE)
2004	FD_SET(hNative, &fdsetW);
2005	if (fEvents & RTSOCKET_EVT_ERROR)
2006	FD_SET(hNative, &fdsetE);
2007
2008	int rc;
2009	if (cMillies == RT_INDEFINITE_WAIT)
2010	rc = select(fdMax, &fdsetR, &fdsetW, &fdsetE, NULL);
2011	else
2012	{
2013	struct timeval timeout;
2014	timeout.tv_sec = cMillies / 1000;
2015	timeout.tv_usec = (cMillies % 1000) * 1000;
2016	rc = select(fdMax, &fdsetR, &fdsetW, &fdsetE, &timeout);
2017	}
2018	if (rc > 0)
2019	{
2020	if (pThis->hNative == hNative)
2021	{
2022	if (FD_ISSET(hNative, &fdsetR))
2023	*pfEvents \|= RTSOCKET_EVT_READ;
2024	if (FD_ISSET(hNative, &fdsetW))
2025	*pfEvents \|= RTSOCKET_EVT_WRITE;
2026	if (FD_ISSET(hNative, &fdsetE))
2027	*pfEvents \|= RTSOCKET_EVT_ERROR;
2028	rc = VINF_SUCCESS;
2029	}
2030	else
2031	{
2032	/* Socket was closed while we waited (rtSocketCloseIt). Different status code? */
2033	*pfEvents = RTSOCKET_EVT_ERROR;
2034	rc = VINF_SUCCESS;
2035	}
2036	}
2037	else if (rc == 0)
2038	rc = VERR_TIMEOUT;
2039	else
2040	rc = rtSocketError();
2041
2042	#ifdef RT_OS_WINDOWS
2043	# undef select
2044	# undef __WSAFDIsSet
2045	#endif
2046	return rc;
2047	}
2048
2049
2050	RTDECL(int) RTSocketSelectOneEx(RTSOCKET hSocket, uint32_t fEvents, uint32_t *pfEvents, RTMSINTERVAL cMillies)
2051	{
2052	/*
2053	* Validate input.
2054	*/
2055	RTSOCKETINT *pThis = hSocket;
2056	AssertPtrReturn(pThis, VERR_INVALID_HANDLE);
2057	AssertReturn(pThis->u32Magic == RTSOCKET_MAGIC, VERR_INVALID_HANDLE);
2058	AssertPtrReturn(pfEvents, VERR_INVALID_PARAMETER);
2059	AssertReturn(!(fEvents & ~RTSOCKET_EVT_VALID_MASK), VERR_INVALID_PARAMETER);
2060	AssertReturn(RTMemPoolRefCount(pThis) >= (pThis->cUsers ? 2U : 1U), VERR_CALLER_NO_REFERENCE);
2061
2062	return rtSocketSelectOneEx(pThis, fEvents, pfEvents, cMillies);
2063	}
2064
2065
2066	RTDECL(int) RTSocketShutdown(RTSOCKET hSocket, bool fRead, bool fWrite)
2067	{
2068	/*
2069	* Validate input, don't lock it because we might want to interrupt a call
2070	* active on a different thread.
2071	*/
2072	RTSOCKETINT *pThis = hSocket;
2073	AssertPtrReturn(pThis, VERR_INVALID_HANDLE);
2074	AssertReturn(pThis->u32Magic == RTSOCKET_MAGIC, VERR_INVALID_HANDLE);
2075	AssertReturn(RTMemPoolRefCount(pThis) >= (pThis->cUsers ? 2U : 1U), VERR_CALLER_NO_REFERENCE);
2076	AssertReturn(fRead \|\| fWrite, VERR_INVALID_PARAMETER);
2077	#ifdef RT_OS_WINDOWS
2078	AssertReturn(g_pfnshutdown, VERR_NET_NOT_UNSUPPORTED);
2079	# define shutdown g_pfnshutdown
2080	#endif
2081
2082	/*
2083	* Do the job.
2084	*/
2085	int rc = VINF_SUCCESS;
2086	int fHow;
2087	if (fRead && fWrite)
2088	fHow = SHUT_RDWR;
2089	else if (fRead)
2090	fHow = SHUT_RD;
2091	else
2092	fHow = SHUT_WR;
2093	if (shutdown(pThis->hNative, fHow) == -1)
2094	rc = rtSocketError();
2095
2096	#ifdef RT_OS_WINDOWS
2097	# undef shutdown
2098	#endif
2099	return rc;
2100	}
2101
2102
2103	RTDECL(int) RTSocketGetLocalAddress(RTSOCKET hSocket, PRTNETADDR pAddr)
2104	{
2105	/*
2106	* Validate input.
2107	*/
2108	RTSOCKETINT *pThis = hSocket;
2109	AssertPtrReturn(pThis, VERR_INVALID_HANDLE);
2110	AssertReturn(pThis->u32Magic == RTSOCKET_MAGIC, VERR_INVALID_HANDLE);
2111	AssertReturn(RTMemPoolRefCount(pThis) >= (pThis->cUsers ? 2U : 1U), VERR_CALLER_NO_REFERENCE);
2112	#ifdef RT_OS_WINDOWS
2113	AssertReturn(g_pfngetsockname, VERR_NET_NOT_UNSUPPORTED);
2114	# define getsockname g_pfngetsockname
2115	#endif
2116
2117	/*
2118	* Get the address and convert it.
2119	*/
2120	int rc;
2121	RTSOCKADDRUNION u;
2122	#ifdef RT_OS_WINDOWS
2123	int cbAddr = sizeof(u);
2124	#else
2125	socklen_t cbAddr = sizeof(u);
2126	#endif
2127	RT_ZERO(u);
2128	if (getsockname(pThis->hNative, &u.Addr, &cbAddr) == 0)
2129	rc = rtSocketNetAddrFromAddr(&u, cbAddr, pAddr);
2130	else
2131	rc = rtSocketError();
2132
2133	#ifdef RT_OS_WINDOWS
2134	# undef getsockname
2135	#endif
2136	return rc;
2137	}
2138
2139
2140	RTDECL(int) RTSocketGetPeerAddress(RTSOCKET hSocket, PRTNETADDR pAddr)
2141	{
2142	/*
2143	* Validate input.
2144	*/
2145	RTSOCKETINT *pThis = hSocket;
2146	AssertPtrReturn(pThis, VERR_INVALID_HANDLE);
2147	AssertReturn(pThis->u32Magic == RTSOCKET_MAGIC, VERR_INVALID_HANDLE);
2148	AssertReturn(RTMemPoolRefCount(pThis) >= (pThis->cUsers ? 2U : 1U), VERR_CALLER_NO_REFERENCE);
2149	#ifdef RT_OS_WINDOWS
2150	AssertReturn(g_pfngetpeername, VERR_NET_NOT_UNSUPPORTED);
2151	# define getpeername g_pfngetpeername
2152	#endif
2153
2154	/*
2155	* Get the address and convert it.
2156	*/
2157	int rc;
2158	RTSOCKADDRUNION u;
2159	#ifdef RT_OS_WINDOWS
2160	int cbAddr = sizeof(u);
2161	#else
2162	socklen_t cbAddr = sizeof(u);
2163	#endif
2164	RT_ZERO(u);
2165	if (getpeername(pThis->hNative, &u.Addr, &cbAddr) == 0)
2166	rc = rtSocketNetAddrFromAddr(&u, cbAddr, pAddr);
2167	else
2168	rc = rtSocketError();
2169
2170	#ifdef RT_OS_WINDOWS
2171	# undef getpeername
2172	#endif
2173	return rc;
2174	}
2175
2176
2177
2178	/**
2179	* Wrapper around bind.
2180	*
2181	* @returns IPRT status code.
2182	* @param hSocket The socket handle.
2183	* @param pAddr The address to bind to.
2184	*/
2185	DECLHIDDEN(int) rtSocketBind(RTSOCKET hSocket, PCRTNETADDR pAddr)
2186	{
2187	RTSOCKADDRUNION u;
2188	int cbAddr;
2189	int rc = rtSocketAddrFromNetAddr(pAddr, &u, sizeof(u), &cbAddr);
2190	if (RT_SUCCESS(rc))
2191	rc = rtSocketBindRawAddr(hSocket, &u.Addr, cbAddr);
2192	return rc;
2193	}
2194
2195
2196	/**
2197	* Very thin wrapper around bind.
2198	*
2199	* @returns IPRT status code.
2200	* @param hSocket The socket handle.
2201	* @param pvAddr The address to bind to (struct sockaddr and
2202	* friends).
2203	* @param cbAddr The size of the address.
2204	*/
2205	DECLHIDDEN(int) rtSocketBindRawAddr(RTSOCKET hSocket, void const *pvAddr, size_t cbAddr)
2206	{
2207	/*
2208	* Validate input.
2209	*/
2210	RTSOCKETINT *pThis = hSocket;
2211	AssertPtrReturn(pThis, VERR_INVALID_HANDLE);
2212	AssertReturn(pThis->u32Magic == RTSOCKET_MAGIC, VERR_INVALID_HANDLE);
2213	AssertPtrReturn(pvAddr, VERR_INVALID_POINTER);
2214	#ifdef RT_OS_WINDOWS
2215	AssertReturn(g_pfnbind, VERR_NET_NOT_UNSUPPORTED);
2216	# define bind g_pfnbind
2217	#endif
2218	AssertReturn(rtSocketTryLock(pThis), VERR_CONCURRENT_ACCESS);
2219
2220	int rc;
2221	if (bind(pThis->hNative, (struct sockaddr const *)pvAddr, (int)cbAddr) == 0)
2222	rc = VINF_SUCCESS;
2223	else
2224	rc = rtSocketError();
2225
2226	rtSocketUnlock(pThis);
2227	#ifdef RT_OS_WINDOWS
2228	# undef bind
2229	#endif
2230	return rc;
2231	}
2232
2233
2234
2235	/**
2236	* Wrapper around listen.
2237	*
2238	* @returns IPRT status code.
2239	* @param hSocket The socket handle.
2240	* @param cMaxPending The max number of pending connections.
2241	*/
2242	DECLHIDDEN(int) rtSocketListen(RTSOCKET hSocket, int cMaxPending)
2243	{
2244	/*
2245	* Validate input.
2246	*/
2247	RTSOCKETINT *pThis = hSocket;
2248	AssertPtrReturn(pThis, VERR_INVALID_HANDLE);
2249	AssertReturn(pThis->u32Magic == RTSOCKET_MAGIC, VERR_INVALID_HANDLE);
2250	#ifdef RT_OS_WINDOWS
2251	AssertReturn(g_pfnlisten, VERR_NET_NOT_UNSUPPORTED);
2252	# define listen g_pfnlisten
2253	#endif
2254	AssertReturn(rtSocketTryLock(pThis), VERR_CONCURRENT_ACCESS);
2255
2256	int rc = VINF_SUCCESS;
2257	if (listen(pThis->hNative, cMaxPending) != 0)
2258	rc = rtSocketError();
2259
2260	rtSocketUnlock(pThis);
2261	#ifdef RT_OS_WINDOWS
2262	# undef listen
2263	#endif
2264	return rc;
2265	}
2266
2267
2268	/**
2269	* Wrapper around accept.
2270	*
2271	* @returns IPRT status code.
2272	* @param hSocket The socket handle.
2273	* @param phClient Where to return the client socket handle on
2274	* success.
2275	* @param pAddr Where to return the client address.
2276	* @param pcbAddr On input this gives the size buffer size of what
2277	* @a pAddr point to. On return this contains the
2278	* size of what's stored at @a pAddr.
2279	*/
2280	DECLHIDDEN(int) rtSocketAccept(RTSOCKET hSocket, PRTSOCKET phClient, struct sockaddr pAddr, size_t pcbAddr)
2281	{
2282	/*
2283	* Validate input.
2284	* Only lock the socket temporarily while we get the native handle, so that
2285	* we can safely shutdown and destroy the socket from a different thread.
2286	*/
2287	RTSOCKETINT *pThis = hSocket;
2288	AssertPtrReturn(pThis, VERR_INVALID_HANDLE);
2289	AssertReturn(pThis->u32Magic == RTSOCKET_MAGIC, VERR_INVALID_HANDLE);
2290	#ifdef RT_OS_WINDOWS
2291	AssertReturn(g_pfnaccept, VERR_NET_NOT_UNSUPPORTED);
2292	AssertReturn(g_pfnclosesocket, VERR_NET_NOT_UNSUPPORTED);
2293	# define accept g_pfnaccept
2294	#endif
2295	AssertReturn(rtSocketTryLock(pThis), VERR_CONCURRENT_ACCESS);
2296
2297	/*
2298	* Call accept().
2299	*/
2300	rtSocketErrorReset();
2301	int rc = VINF_SUCCESS;
2302	#ifdef RT_OS_WINDOWS
2303	int cbAddr = (int)*pcbAddr;
2304	#else
2305	socklen_t cbAddr = *pcbAddr;
2306	#endif
2307	RTSOCKETNATIVE hNativeClient = accept(pThis->hNative, pAddr, &cbAddr);
2308	if (hNativeClient != NIL_RTSOCKETNATIVE)
2309	{
2310	*pcbAddr = cbAddr;
2311
2312	/*
2313	* Wrap the client socket.
2314	*/
2315	rc = rtSocketCreateForNative(phClient, hNativeClient, false /fLeaveOpen/);
2316	if (RT_FAILURE(rc))
2317	{
2318	#ifdef RT_OS_WINDOWS
2319	g_pfnclosesocket(hNativeClient);
2320	#else
2321	close(hNativeClient);
2322	#endif
2323	}
2324	}
2325	else
2326	rc = rtSocketError();
2327
2328	rtSocketUnlock(pThis);
2329	#ifdef RT_OS_WINDOWS
2330	# undef accept
2331	#endif
2332	return rc;
2333	}
2334
2335
2336	/**
2337	* Wrapper around connect.
2338	*
2339	* @returns IPRT status code.
2340	* @param hSocket The socket handle.
2341	* @param pAddr The socket address to connect to.
2342	* @param cMillies Number of milliseconds to wait for the connect attempt to complete.
2343	* Use RT_INDEFINITE_WAIT to wait for ever.
2344	* Use RT_TCPCLIENTCONNECT_DEFAULT_WAIT to wait for the default time
2345	* configured on the running system.
2346	*/
2347	DECLHIDDEN(int) rtSocketConnect(RTSOCKET hSocket, PCRTNETADDR pAddr, RTMSINTERVAL cMillies)
2348	{
2349	/*
2350	* Validate input.
2351	*/
2352	RTSOCKETINT *pThis = hSocket;
2353	AssertPtrReturn(pThis, VERR_INVALID_HANDLE);
2354	AssertReturn(pThis->u32Magic == RTSOCKET_MAGIC, VERR_INVALID_HANDLE);
2355	#ifdef RT_OS_WINDOWS
2356	AssertReturn(g_pfnconnect, VERR_NET_NOT_UNSUPPORTED);
2357	AssertReturn(g_pfnselect, VERR_NET_NOT_UNSUPPORTED);
2358	AssertReturn(g_pfngetsockopt, VERR_NET_NOT_UNSUPPORTED);
2359	# define connect g_pfnconnect
2360	# define select g_pfnselect
2361	# define getsockopt g_pfngetsockopt
2362	#endif
2363	AssertReturn(rtSocketTryLock(pThis), VERR_CONCURRENT_ACCESS);
2364
2365	RTSOCKADDRUNION u;
2366	int cbAddr;
2367	int rc = rtSocketAddrFromNetAddr(pAddr, &u, sizeof(u), &cbAddr);
2368	if (RT_SUCCESS(rc))
2369	{
2370	if (cMillies == RT_SOCKETCONNECT_DEFAULT_WAIT)
2371	{
2372	if (connect(pThis->hNative, &u.Addr, cbAddr) != 0)
2373	rc = rtSocketError();
2374	}
2375	else
2376	{
2377	/*
2378	* Switch the socket to nonblocking mode, initiate the connect
2379	* and wait for the socket to become writable or until the timeout
2380	* expires.
2381	*/
2382	rc = rtSocketSwitchBlockingMode(pThis, false /* fBlocking */);
2383	if (RT_SUCCESS(rc))
2384	{
2385	if (connect(pThis->hNative, &u.Addr, cbAddr) != 0)
2386	{
2387	rc = rtSocketError();
2388	if (rc == VERR_TRY_AGAIN \|\| rc == VERR_NET_IN_PROGRESS)
2389	{
2390	int rcSock = 0;
2391	fd_set FdSetWriteable;
2392	struct timeval TvTimeout;
2393
2394	TvTimeout.tv_sec = cMillies / RT_MS_1SEC;
2395	TvTimeout.tv_usec = (cMillies % RT_MS_1SEC) * RT_US_1MS;
2396
2397	FD_ZERO(&FdSetWriteable);
2398	FD_SET(pThis->hNative, &FdSetWriteable);
2399	do
2400	{
2401	rcSock = select(pThis->hNative + 1, NULL, &FdSetWriteable, NULL,
2402	cMillies == RT_INDEFINITE_WAIT \|\| cMillies >= INT_MAX
2403	? NULL
2404	: &TvTimeout);
2405	if (rcSock > 0)
2406	{
2407	int iSockError = 0;
2408	socklen_t cbSockOpt = sizeof(iSockError);
2409	rcSock = getsockopt(pThis->hNative, SOL_SOCKET, SO_ERROR, (char *)&iSockError, &cbSockOpt);
2410	if (rcSock == 0)
2411	{
2412	if (iSockError == 0)
2413	rc = VINF_SUCCESS;
2414	else
2415	{
2416	#ifdef RT_OS_WINDOWS
2417	rc = RTErrConvertFromWin32(iSockError);
2418	#else
2419	rc = RTErrConvertFromErrno(iSockError);
2420	#endif
2421	}
2422	}
2423	else
2424	rc = rtSocketError();
2425	}
2426	else if (rcSock == 0)
2427	rc = VERR_TIMEOUT;
2428	else
2429	rc = rtSocketError();
2430	} while (rc == VERR_INTERRUPTED);
2431	}
2432	}
2433
2434	rtSocketSwitchBlockingMode(pThis, true /* fBlocking */);
2435	}
2436	}
2437	}
2438
2439	rtSocketUnlock(pThis);
2440	#ifdef RT_OS_WINDOWS
2441	# undef connect
2442	# undef select
2443	# undef getsockopt
2444	#endif
2445	return rc;
2446	}
2447
2448
2449	/**
2450	* Wrapper around connect, raw address, no timeout.
2451	*
2452	* @returns IPRT status code.
2453	* @param hSocket The socket handle.
2454	* @param pvAddr The raw socket address to connect to.
2455	* @param cbAddr The size of the raw address.
2456	*/
2457	DECLHIDDEN(int) rtSocketConnectRaw(RTSOCKET hSocket, void const *pvAddr, size_t cbAddr)
2458	{
2459	/*
2460	* Validate input.
2461	*/
2462	RTSOCKETINT *pThis = hSocket;
2463	AssertPtrReturn(pThis, VERR_INVALID_HANDLE);
2464	AssertReturn(pThis->u32Magic == RTSOCKET_MAGIC, VERR_INVALID_HANDLE);
2465	#ifdef RT_OS_WINDOWS
2466	AssertReturn(g_pfnconnect, VERR_NET_NOT_UNSUPPORTED);
2467	# define connect g_pfnconnect
2468	#endif
2469	AssertReturn(rtSocketTryLock(pThis), VERR_CONCURRENT_ACCESS);
2470
2471	int rc;
2472	if (connect(pThis->hNative, (const struct sockaddr *)pvAddr, (int)cbAddr) == 0)
2473	rc = VINF_SUCCESS;
2474	else
2475	rc = rtSocketError();
2476
2477	rtSocketUnlock(pThis);
2478	#ifdef RT_OS_WINDOWS
2479	# undef connect
2480	#endif
2481	return rc;
2482	}
2483
2484
2485	/**
2486	* Wrapper around setsockopt.
2487	*
2488	* @returns IPRT status code.
2489	* @param hSocket The socket handle.
2490	* @param iLevel The protocol level, e.g. IPPORTO_TCP.
2491	* @param iOption The option, e.g. TCP_NODELAY.
2492	* @param pvValue The value buffer.
2493	* @param cbValue The size of the value pointed to by pvValue.
2494	*/
2495	DECLHIDDEN(int) rtSocketSetOpt(RTSOCKET hSocket, int iLevel, int iOption, void const *pvValue, int cbValue)
2496	{
2497	/*
2498	* Validate input.
2499	*/
2500	RTSOCKETINT *pThis = hSocket;
2501	AssertPtrReturn(pThis, VERR_INVALID_HANDLE);
2502	AssertReturn(pThis->u32Magic == RTSOCKET_MAGIC, VERR_INVALID_HANDLE);
2503	#ifdef RT_OS_WINDOWS
2504	AssertReturn(g_pfnsetsockopt, VERR_NET_NOT_UNSUPPORTED);
2505	# define setsockopt g_pfnsetsockopt
2506	#endif
2507	AssertReturn(rtSocketTryLock(pThis), VERR_CONCURRENT_ACCESS);
2508
2509	int rc = VINF_SUCCESS;
2510	if (setsockopt(pThis->hNative, iLevel, iOption, (const char *)pvValue, cbValue) != 0)
2511	rc = rtSocketError();
2512
2513	rtSocketUnlock(pThis);
2514	#ifdef RT_OS_WINDOWS
2515	# undef setsockopt
2516	#endif
2517	return rc;
2518	}
2519
2520
2521	/**
2522	* Internal RTPollSetAdd helper that returns the handle that should be added to
2523	* the pollset.
2524	*
2525	* @returns Valid handle on success, INVALID_HANDLE_VALUE on failure.
2526	* @param hSocket The socket handle.
2527	* @param fEvents The events we're polling for.
2528	* @param phNative Where to put the primary handle.
2529	*/
2530	DECLHIDDEN(int) rtSocketPollGetHandle(RTSOCKET hSocket, uint32_t fEvents, PRTHCINTPTR phNative)
2531	{
2532	RTSOCKETINT *pThis = hSocket;
2533	RT_NOREF_PV(fEvents);
2534	AssertPtrReturn(pThis, VERR_INVALID_HANDLE);
2535	AssertReturn(pThis->u32Magic == RTSOCKET_MAGIC, VERR_INVALID_HANDLE);
2536	#ifdef RT_OS_WINDOWS
2537	AssertReturn(rtSocketTryLock(pThis), VERR_CONCURRENT_ACCESS);
2538
2539	int rc = VINF_SUCCESS;
2540	if (pThis->hEvent != WSA_INVALID_EVENT)
2541	*phNative = (RTHCINTPTR)pThis->hEvent;
2542	else if (g_pfnWSACreateEvent)
2543	{
2544	pThis->hEvent = g_pfnWSACreateEvent();
2545	*phNative = (RTHCINTPTR)pThis->hEvent;
2546	if (pThis->hEvent == WSA_INVALID_EVENT)
2547	rc = rtSocketError();
2548	}
2549	else
2550	{
2551	AssertCompile(WSA_INVALID_EVENT == (WSAEVENT)NULL);
2552	pThis->hEvent = CreateEventW(NULL, TRUE /fManualReset/, FALSE /fInitialState/, NULL /pwszName/);
2553	*phNative = (RTHCINTPTR)pThis->hEvent;
2554	if (pThis->hEvent == WSA_INVALID_EVENT)
2555	rc = RTErrConvertFromWin32(GetLastError());
2556	}
2557
2558	rtSocketUnlock(pThis);
2559	return rc;
2560
2561	#else /* !RT_OS_WINDOWS */
2562	*phNative = (RTHCUINTPTR)pThis->hNative;
2563	return VINF_SUCCESS;
2564	#endif /* !RT_OS_WINDOWS */
2565	}
2566
2567	#ifdef RT_OS_WINDOWS
2568
2569	/**
2570	* Fallback poller thread.
2571	*
2572	* @returns VINF_SUCCESS.
2573	* @param hSelf The thread handle.
2574	* @param pvUser Socket instance data.
2575	*/
2576	static DECLCALLBACK(int) rtSocketPollFallbackThreadProc(RTTHREAD hSelf, void *pvUser)
2577	{
2578	RTSOCKETINT pThis = (RTSOCKETINT )pvUser;
2579	RT_NOREF(hSelf);
2580	# define __WSAFDIsSet g_pfn__WSAFDIsSet
2581
2582	/*
2583	* The execution loop.
2584	*/
2585	while (!ASMAtomicReadBool(&pThis->fPollFallbackShutdown))
2586	{
2587	/*
2588	* Do the selecting (with a 15 second timeout because that seems like a good idea).
2589	*/
2590	struct fd_set SetRead;
2591	struct fd_set SetWrite;
2592	struct fd_set SetXcpt;
2593
2594	FD_ZERO(&SetRead);
2595	FD_ZERO(&SetWrite);
2596	FD_ZERO(&SetXcpt);
2597
2598	FD_SET(pThis->hPollFallbackNotifyR, &SetRead);
2599	FD_SET(pThis->hPollFallbackNotifyR, &SetXcpt);
2600
2601	bool fActive = ASMAtomicReadBool(&pThis->fPollFallbackActive);
2602	uint32_t fEvents;
2603	if (!fActive)
2604	fEvents = 0;
2605	else
2606	{
2607	fEvents = ASMAtomicReadU32(&pThis->fSubscribedEvts);
2608	if (fEvents & RTPOLL_EVT_READ)
2609	FD_SET(pThis->hNative, &SetRead);
2610	if (fEvents & RTPOLL_EVT_WRITE)
2611	FD_SET(pThis->hNative, &SetWrite);
2612	if (fEvents & RTPOLL_EVT_ERROR)
2613	FD_SET(pThis->hNative, &SetXcpt);
2614	}
2615
2616	struct timeval Timeout;
2617	Timeout.tv_sec = 15;
2618	Timeout.tv_usec = 0;
2619	int rc = g_pfnselect(INT_MAX /ignored/, &SetRead, &SetWrite, &SetXcpt, &Timeout);
2620
2621	/* Stop immediately if told to shut down. */
2622	if (ASMAtomicReadBool(&pThis->fPollFallbackShutdown))
2623	break;
2624
2625	/*
2626	* Process the result.
2627	*/
2628	if (rc > 0)
2629	{
2630	/* First the socket we're listening on. */
2631	if ( fEvents
2632	&& ( FD_ISSET(pThis->hNative, &SetRead)
2633	\|\| FD_ISSET(pThis->hNative, &SetWrite)
2634	\|\| FD_ISSET(pThis->hNative, &SetXcpt)) )
2635	{
2636	ASMAtomicWriteBool(&pThis->fPollFallbackActive, false);
2637	SetEvent(pThis->hEvent);
2638	}
2639
2640	/* Then maintain the notification pipe. (We only read one byte here
2641	because we're overly paranoid wrt socket switching to blocking mode.) */
2642	if (FD_ISSET(pThis->hPollFallbackNotifyR, &SetRead))
2643	{
2644	char chIgnored;
2645	g_pfnrecv(pThis->hPollFallbackNotifyR, &chIgnored, sizeof(chIgnored), MSG_NOSIGNAL);
2646	}
2647	}
2648	else
2649	AssertMsg(rc == 0, ("%Rrc\n", rtSocketError()));
2650	}
2651
2652	# undef __WSAFDIsSet
2653	return VINF_SUCCESS;
2654	}
2655
2656
2657	/**
2658	* Pokes the fallback thread, making sure it gets out of whatever it's stuck in.
2659	*
2660	* @param pThis The socket handle.
2661	*/
2662	static void rtSocketPokePollFallbackThread(RTSOCKETINT *pThis)
2663	{
2664	Assert(pThis->fPollFallback);
2665	if (pThis->hPollFallbackThread != NIL_RTTHREAD)
2666	{
2667	int cbWritten = g_pfnsend(pThis->hPollFallbackNotifyW, "!", 1, MSG_NOSIGNAL);
2668	AssertMsg(cbWritten == 1, ("cbWritten=%d err=%Rrc\n", rtSocketError()));
2669	RT_NOREF_PV(cbWritten);
2670	}
2671	}
2672
2673
2674	/**
2675	* Called by rtSocketPollStart to make the thread start selecting on the socket.
2676	*
2677	* @returns 0 on success, RTPOLL_EVT_ERROR on failure.
2678	* @param pThis The socket handle.
2679	*/
2680	static uint32_t rtSocketPollFallbackStart(RTSOCKETINT *pThis)
2681	{
2682	/*
2683	* Reset the event and tell the thread to start selecting on the socket.
2684	*/
2685	ResetEvent(pThis->hEvent);
2686	ASMAtomicWriteBool(&pThis->fPollFallbackActive, true);
2687
2688	/*
2689	* Wake up the thread the thread.
2690	*/
2691	if (pThis->hPollFallbackThread != NIL_RTTHREAD)
2692	rtSocketPokePollFallbackThread(pThis);
2693	else
2694	{
2695	/*
2696	* Not running, need to set it up and start it.
2697	*/
2698	AssertLogRelReturn(pThis->hEvent != NULL && pThis->hEvent != INVALID_HANDLE_VALUE, RTPOLL_EVT_ERROR);
2699
2700	/* Create the notification socket pair. */
2701	int rc;
2702	if (pThis->hPollFallbackNotifyR == NIL_RTSOCKETNATIVE)
2703	{
2704	rc = rtSocketCreateNativeTcpPair(&pThis->hPollFallbackNotifyW, &pThis->hPollFallbackNotifyR);
2705	AssertLogRelRCReturn(rc, RTPOLL_EVT_ERROR);
2706
2707	/* Make the read end non-blocking (not fatal). */
2708	u_long fNonBlocking = 1;
2709	rc = g_pfnioctlsocket(pThis->hPollFallbackNotifyR, FIONBIO, &fNonBlocking);
2710	AssertLogRelMsg(rc == 0, ("rc=%#x %Rrc\n", rc, rtSocketError()));
2711	}
2712
2713	/* Finally, start the thread. ASSUME we don't need too much stack. */
2714	rc = RTThreadCreate(&pThis->hPollFallbackThread, rtSocketPollFallbackThreadProc, pThis,
2715	_128K, RTTHREADTYPE_IO, RTTHREADFLAGS_WAITABLE, "sockpoll");
2716	AssertLogRelRCReturn(rc, RTPOLL_EVT_ERROR);
2717	}
2718	return 0;
2719	}
2720
2721
2722	/**
2723	* Undos the harm done by WSAEventSelect.
2724	*
2725	* @returns IPRT status code.
2726	* @param pThis The socket handle.
2727	*/
2728	static int rtSocketPollClearEventAndRestoreBlocking(RTSOCKETINT *pThis)
2729	{
2730	int rc = VINF_SUCCESS;
2731	if (pThis->fSubscribedEvts)
2732	{
2733	if (!pThis->fPollFallback)
2734	{
2735	Assert(g_pfnWSAEventSelect && g_pfnioctlsocket);
2736	if (g_pfnWSAEventSelect && g_pfnioctlsocket)
2737	{
2738	if (g_pfnWSAEventSelect(pThis->hNative, WSA_INVALID_EVENT, 0) == 0)
2739	{
2740	pThis->fSubscribedEvts = 0;
2741
2742	/*
2743	* Switch back to blocking mode if that was the state before the
2744	* operation.
2745	*/
2746	if (pThis->fBlocking)
2747	{
2748	u_long fNonBlocking = 0;
2749	int rc2 = g_pfnioctlsocket(pThis->hNative, FIONBIO, &fNonBlocking);
2750	if (rc2 != 0)
2751	{
2752	rc = rtSocketError();
2753	AssertMsgFailed(("%Rrc; rc2=%d\n", rc, rc2));
2754	}
2755	}
2756	}
2757	else
2758	{
2759	rc = rtSocketError();
2760	AssertMsgFailed(("%Rrc\n", rc));
2761	}
2762	}
2763	else
2764	{
2765	Assert(pThis->fPollFallback);
2766	rc = VINF_SUCCESS;
2767	}
2768	}
2769	/*
2770	* Just clear the event mask as we never started waiting if we get here.
2771	*/
2772	else
2773	ASMAtomicWriteU32(&pThis->fSubscribedEvts, 0);
2774	}
2775	return rc;
2776	}
2777
2778
2779	/**
2780	* Updates the mask of events we're subscribing to.
2781	*
2782	* @returns IPRT status code.
2783	* @param pThis The socket handle.
2784	* @param fEvents The events we want to subscribe to.
2785	*/
2786	static int rtSocketPollUpdateEvents(RTSOCKETINT *pThis, uint32_t fEvents)
2787	{
2788	if (!pThis->fPollFallback)
2789	{
2790	LONG fNetworkEvents = 0;
2791	if (fEvents & RTPOLL_EVT_READ)
2792	fNetworkEvents \|= FD_READ;
2793	if (fEvents & RTPOLL_EVT_WRITE)
2794	fNetworkEvents \|= FD_WRITE;
2795	if (fEvents & RTPOLL_EVT_ERROR)
2796	fNetworkEvents \|= FD_CLOSE;
2797	LogFlowFunc(("fNetworkEvents=%#x\n", fNetworkEvents));
2798
2799	if (g_pfnWSAEventSelect(pThis->hNative, pThis->hEvent, fNetworkEvents) == 0)
2800	{
2801	pThis->fSubscribedEvts = fEvents;
2802	return VINF_SUCCESS;
2803	}
2804
2805	int rc = rtSocketError();
2806	AssertMsgFailed(("fNetworkEvents=%#x rc=%Rrc\n", fNetworkEvents, rtSocketError()));
2807	return rc;
2808	}
2809
2810	/*
2811	* Update the events we're waiting for. Caller will poke/start the thread. later
2812	*/
2813	ASMAtomicWriteU32(&pThis->fSubscribedEvts, fEvents);
2814	return VINF_SUCCESS;
2815	}
2816
2817	#endif /* RT_OS_WINDOWS */
2818
2819
2820	#if defined(RT_OS_WINDOWS) \|\| defined(RT_OS_OS2)
2821
2822	/**
2823	* Checks for pending events.
2824	*
2825	* @returns Event mask or 0.
2826	* @param pThis The socket handle.
2827	* @param fEvents The desired events.
2828	*/
2829	static uint32_t rtSocketPollCheck(RTSOCKETINT *pThis, uint32_t fEvents)
2830	{
2831	uint32_t fRetEvents = 0;
2832
2833	LogFlowFunc(("pThis=%#p fEvents=%#x\n", pThis, fEvents));
2834
2835	# ifdef RT_OS_WINDOWS
2836	/* Make sure WSAEnumNetworkEvents returns what we want. */
2837	int rc = VINF_SUCCESS;
2838	if ((pThis->fSubscribedEvts & fEvents) != fEvents)
2839	rc = rtSocketPollUpdateEvents(pThis, pThis->fSubscribedEvts \| fEvents);
2840
2841	if (!pThis->fPollFallback)
2842	{
2843	/* Atomically get pending events and reset the event semaphore. */
2844	Assert(g_pfnWSAEnumNetworkEvents);
2845	WSANETWORKEVENTS NetEvts;
2846	RT_ZERO(NetEvts);
2847	if (g_pfnWSAEnumNetworkEvents(pThis->hNative, pThis->hEvent, &NetEvts) == 0)
2848	{
2849	if ( (NetEvts.lNetworkEvents & FD_READ)
2850	&& NetEvts.iErrorCode[FD_READ_BIT] == 0)
2851	fRetEvents \|= RTPOLL_EVT_READ;
2852
2853	if ( (NetEvts.lNetworkEvents & FD_WRITE)
2854	&& NetEvts.iErrorCode[FD_WRITE_BIT] == 0)
2855	fRetEvents \|= RTPOLL_EVT_WRITE;
2856
2857	if (NetEvts.lNetworkEvents & FD_CLOSE)
2858	fRetEvents \|= RTPOLL_EVT_ERROR;
2859	else
2860	for (uint32_t i = 0; i < FD_MAX_EVENTS; i++)
2861	if ( (NetEvts.lNetworkEvents & (1L << i))
2862	&& NetEvts.iErrorCode[i] != 0)
2863	fRetEvents \|= RTPOLL_EVT_ERROR;
2864
2865	pThis->fEventsSaved = fRetEvents \|= pThis->fEventsSaved;
2866	fRetEvents &= fEvents \| RTPOLL_EVT_ERROR;
2867	}
2868	else
2869	rc = rtSocketError();
2870	}
2871
2872	/* Fall back on select if we hit an error above or is using fallback polling. */
2873	if (pThis->fPollFallback \|\| RT_FAILURE(rc))
2874	{
2875	rc = rtSocketSelectOneEx(pThis, fEvents & RTPOLL_EVT_ERROR ? fEvents \| RTPOLL_EVT_READ : fEvents, &fRetEvents, 0);
2876	if (RT_SUCCESS(rc))
2877	{
2878	/* rtSocketSelectOneEx may return RTPOLL_EVT_READ on disconnect. Use
2879	getpeername to fix this. */
2880	if ((fRetEvents & (RTPOLL_EVT_READ \| RTPOLL_EVT_ERROR)) == RTPOLL_EVT_READ)
2881	{
2882	# if 0 /* doens't work */
2883	rtSocketErrorReset();
2884	char chIgn;
2885	rc = g_pfnrecv(pThis->hNative, &chIgn, 0, MSG_NOSIGNAL);
2886	rc = rtSocketError();
2887	if (RT_FAILURE(rc))
2888	fRetEvents \|= RTPOLL_EVT_ERROR;
2889
2890	rc = g_pfnsend(pThis->hNative, &chIgn, 0, MSG_NOSIGNAL);
2891	rc = rtSocketError();
2892	if (RT_FAILURE(rc))
2893	fRetEvents \|= RTPOLL_EVT_ERROR;
2894
2895	RTSOCKADDRUNION u;
2896	int cbAddr = sizeof(u);
2897	if (g_pfngetpeername(pThis->hNative, &u.Addr, &cbAddr) == SOCKET_ERROR)
2898	fRetEvents \|= RTPOLL_EVT_ERROR;
2899	# endif
2900	/* If no bytes are available, assume error condition. */
2901	u_long cbAvail = 0;
2902	rc = g_pfnioctlsocket(pThis->hNative, FIONREAD, &cbAvail);
2903	if (rc == 0 && cbAvail == 0)
2904	fRetEvents \|= RTPOLL_EVT_ERROR;
2905	}
2906	fRetEvents &= fEvents \| RTPOLL_EVT_ERROR;
2907	}
2908	else if (rc == VERR_TIMEOUT)
2909	fRetEvents = 0;
2910	else
2911	fRetEvents \|= RTPOLL_EVT_ERROR;
2912	}
2913
2914	# else /* RT_OS_OS2 */
2915	int aFds[4] = { pThis->hNative, pThis->hNative, pThis->hNative, -1 };
2916	int rc = os2_select(aFds, 1, 1, 1, 0);
2917	if (rc > 0)
2918	{
2919	if (aFds[0] == pThis->hNative)
2920	fRetEvents \|= RTPOLL_EVT_READ;
2921	if (aFds[1] == pThis->hNative)
2922	fRetEvents \|= RTPOLL_EVT_WRITE;
2923	if (aFds[2] == pThis->hNative)
2924	fRetEvents \|= RTPOLL_EVT_ERROR;
2925	fRetEvents &= fEvents;
2926	}
2927	# endif /* RT_OS_OS2 */
2928
2929	LogFlowFunc(("fRetEvents=%#x\n", fRetEvents));
2930	return fRetEvents;
2931	}
2932
2933
2934	/**
2935	* Internal RTPoll helper that polls the socket handle and, if @a fNoWait is
2936	* clear, starts whatever actions we've got running during the poll call.
2937	*
2938	* @returns 0 if no pending events, actions initiated if @a fNoWait is clear.
2939	* Event mask (in @a fEvents) and no actions if the handle is ready
2940	* already.
2941	* UINT32_MAX (asserted) if the socket handle is busy in I/O or a
2942	* different poll set.
2943	*
2944	* @param hSocket The socket handle.
2945	* @param hPollSet The poll set handle (for access checks).
2946	* @param fEvents The events we're polling for.
2947	* @param fFinalEntry Set if this is the final entry for this handle
2948	* in this poll set. This can be used for dealing
2949	* with duplicate entries.
2950	* @param fNoWait Set if it's a zero-wait poll call. Clear if
2951	* we'll wait for an event to occur.
2952	*
2953	* @remarks There is a potential race wrt duplicate handles when @a fNoWait is
2954	* @c true, we don't currently care about that oddity...
2955	*/
2956	DECLHIDDEN(uint32_t) rtSocketPollStart(RTSOCKET hSocket, RTPOLLSET hPollSet, uint32_t fEvents, bool fFinalEntry, bool fNoWait)
2957	{
2958	RTSOCKETINT *pThis = hSocket;
2959	AssertPtrReturn(pThis, UINT32_MAX);
2960	AssertReturn(pThis->u32Magic == RTSOCKET_MAGIC, UINT32_MAX);
2961	/** @todo This isn't quite sane. Replace by critsect and open up concurrent
2962	* reads and writes! */
2963	if (rtSocketTryLock(pThis))
2964	pThis->hPollSet = hPollSet;
2965	else
2966	{
2967	AssertReturn(pThis->hPollSet == hPollSet, UINT32_MAX);
2968	ASMAtomicIncU32(&pThis->cUsers);
2969	}
2970
2971	/* (rtSocketPollCheck will reset the event object). */
2972	# ifdef RT_OS_WINDOWS
2973	uint32_t fRetEvents = pThis->fEventsSaved;
2974	pThis->fEventsSaved = 0; /* Reset */
2975	fRetEvents \|= rtSocketPollCheck(pThis, fEvents);
2976
2977	if ( !fRetEvents
2978	&& !fNoWait)
2979	{
2980	pThis->fPollEvts \|= fEvents;
2981	if (fFinalEntry)
2982	{
2983	if (pThis->fSubscribedEvts != pThis->fPollEvts)
2984	{
2985	/** @todo seems like there might be a call to many here and that fPollEvts is
2986	* totally unnecessary... (bird) */
2987	int rc = rtSocketPollUpdateEvents(pThis, pThis->fPollEvts);
2988	if (RT_FAILURE(rc))
2989	{
2990	pThis->fPollEvts = 0;
2991	fRetEvents = UINT32_MAX;
2992	}
2993	}
2994
2995	/* Make sure we don't block when there are events pending relevant to an earlier poll set entry. */
2996	if (pThis->fEventsSaved && !pThis->fPollFallback && g_pfnWSASetEvent && fRetEvents == 0)
2997	g_pfnWSASetEvent(pThis->hEvent);
2998	}
2999	}
3000	# else
3001	uint32_t fRetEvents = rtSocketPollCheck(pThis, fEvents);
3002	# endif
3003
3004	if (fRetEvents \|\| fNoWait)
3005	{
3006	if (pThis->cUsers == 1)
3007	{
3008	# ifdef RT_OS_WINDOWS
3009	pThis->fEventsSaved &= RTPOLL_EVT_ERROR;
3010	pThis->fHarvestedEvents = false;
3011	rtSocketPollClearEventAndRestoreBlocking(pThis);
3012	# endif
3013	pThis->hPollSet = NIL_RTPOLLSET;
3014	}
3015	# ifdef RT_OS_WINDOWS
3016	else
3017	pThis->fHarvestedEvents = true;
3018	# endif
3019	ASMAtomicDecU32(&pThis->cUsers);
3020	}
3021	# ifdef RT_OS_WINDOWS
3022	/*
3023	* Kick the poller thread on if this is the final entry and we're in
3024	* winsock 1.x fallback mode.
3025	*/
3026	else if (pThis->fPollFallback && fFinalEntry)
3027	fRetEvents = rtSocketPollFallbackStart(pThis);
3028	# endif
3029
3030	return fRetEvents;
3031	}
3032
3033
3034	/**
3035	* Called after a WaitForMultipleObjects returned in order to check for pending
3036	* events and stop whatever actions that rtSocketPollStart() initiated.
3037	*
3038	* @returns Event mask or 0.
3039	*
3040	* @param hSocket The socket handle.
3041	* @param fEvents The events we're polling for.
3042	* @param fFinalEntry Set if this is the final entry for this handle
3043	* in this poll set. This can be used for dealing
3044	* with duplicate entries. Only keep in mind that
3045	* this method is called in reverse order, so the
3046	* first call will have this set (when the entire
3047	* set was processed).
3048	* @param fHarvestEvents Set if we should check for pending events.
3049	*/
3050	DECLHIDDEN(uint32_t) rtSocketPollDone(RTSOCKET hSocket, uint32_t fEvents, bool fFinalEntry, bool fHarvestEvents)
3051	{
3052	RTSOCKETINT *pThis = hSocket;
3053	AssertPtrReturn(pThis, 0);
3054	AssertReturn(pThis->u32Magic == RTSOCKET_MAGIC, 0);
3055	Assert(pThis->cUsers > 0);
3056	Assert(pThis->hPollSet != NIL_RTPOLLSET);
3057	RT_NOREF_PV(fFinalEntry);
3058
3059	# ifdef RT_OS_WINDOWS
3060	/*
3061	* Deactivate the poll thread if we're in winsock 1.x fallback poll mode.
3062	*/
3063	if ( pThis->fPollFallback
3064	&& pThis->hPollFallbackThread != NIL_RTTHREAD)
3065	{
3066	ASMAtomicWriteU32(&pThis->fSubscribedEvts, 0);
3067	if (ASMAtomicXchgBool(&pThis->fPollFallbackActive, false))
3068	rtSocketPokePollFallbackThread(pThis);
3069	}
3070	# endif
3071
3072	/*
3073	* Harvest events and clear the event mask for the next round of polling.
3074	*/
3075	uint32_t fRetEvents;
3076	# ifdef RT_OS_WINDOWS
3077	if (!pThis->fPollFallback)
3078	{
3079	if (!pThis->fHarvestedEvents)
3080	{
3081	fRetEvents = rtSocketPollCheck(pThis, fEvents);
3082	pThis->fHarvestedEvents = true;
3083	}
3084	else
3085	fRetEvents = pThis->fEventsSaved;
3086	if (fHarvestEvents)
3087	fRetEvents &= fEvents;
3088	else
3089	fRetEvents = 0;
3090	pThis->fPollEvts = 0;
3091	}
3092	else
3093	# endif
3094	{
3095	if (fHarvestEvents)
3096	fRetEvents = rtSocketPollCheck(pThis, fEvents);
3097	else
3098	fRetEvents = 0;
3099	}
3100
3101	/*
3102	* Make the socket blocking again and unlock the handle.
3103	*/
3104	if (pThis->cUsers == 1)
3105	{
3106	# ifdef RT_OS_WINDOWS
3107	pThis->fEventsSaved &= RTPOLL_EVT_ERROR;
3108	pThis->fHarvestedEvents = false;
3109	rtSocketPollClearEventAndRestoreBlocking(pThis);
3110	# endif
3111	pThis->hPollSet = NIL_RTPOLLSET;
3112	}
3113	ASMAtomicDecU32(&pThis->cUsers);
3114	return fRetEvents;
3115	}
3116
3117	#endif /* RT_OS_WINDOWS \|\| RT_OS_OS2 */
3118

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

source: vbox/trunk/src/VBox/Runtime/r3/socket.cpp@ 94839

Download in other formats: