SUPDrv.c@ 17827

Last change on this file since 17827 was 16447, checked in by vboxsync, 16 years ago
typo
Property svn:eol-style set to `native` Property svn:keywords set to `Author Date Id Revision`
File size: 181.4 KB

Line
1	/* $Revision: 16447 $ */
2	/** @file
3	* VBoxDrv - The VirtualBox Support Driver - Common code.
4	*/
5
6	/*
7	* Copyright (C) 2006-2007 Sun Microsystems, Inc.
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	* Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa
27	* Clara, CA 95054 USA or visit http://www.sun.com if you need
28	* additional information or have any questions.
29	*/
30
31	/*******************************************************************************
32	* Header Files *
33	*******************************************************************************/
34	#define LOG_GROUP LOG_GROUP_SUP_DRV
35	#include "SUPDrvInternal.h"
36	#ifndef PAGE_SHIFT
37	# include <iprt/param.h>
38	#endif
39	#include <iprt/alloc.h>
40	#include <iprt/semaphore.h>
41	#include <iprt/spinlock.h>
42	#include <iprt/thread.h>
43	#include <iprt/process.h>
44	#include <iprt/mp.h>
45	#include <iprt/power.h>
46	#include <iprt/cpuset.h>
47	#include <iprt/uuid.h>
48	#include <VBox/param.h>
49	#include <VBox/log.h>
50	#include <VBox/err.h>
51	#if defined(RT_OS_DARWIN) \|\| defined(RT_OS_SOLARIS)
52	# include <iprt/crc32.h>
53	# include <iprt/net.h>
54	#endif
55	/* VBox/x86.h not compatible with the Linux kernel sources */
56	#ifdef RT_OS_LINUX
57	# define X86_CPUID_VENDOR_AMD_EBX 0x68747541
58	# define X86_CPUID_VENDOR_AMD_ECX 0x444d4163
59	# define X86_CPUID_VENDOR_AMD_EDX 0x69746e65
60	#else
61	# include <VBox/x86.h>
62	#endif
63
64	/*
65	* Logging assignments:
66	* Log - useful stuff, like failures.
67	* LogFlow - program flow, except the really noisy bits.
68	* Log2 - Cleanup.
69	* Log3 - Loader flow noise.
70	* Log4 - Call VMMR0 flow noise.
71	* Log5 - Native yet-to-be-defined noise.
72	* Log6 - Native ioctl flow noise.
73	*
74	* Logging requires BUILD_TYPE=debug and possibly changes to the logger
75	* instanciation in log-vbox.c(pp).
76	*/
77
78
79	/*******************************************************************************
80	* Defined Constants And Macros *
81	*******************************************************************************/
82	/* from x86.h - clashes with linux thus this duplication */
83	#undef X86_CR0_PG
84	#define X86_CR0_PG RT_BIT(31)
85	#undef X86_CR0_PE
86	#define X86_CR0_PE RT_BIT(0)
87	#undef X86_CPUID_AMD_FEATURE_EDX_NX
88	#define X86_CPUID_AMD_FEATURE_EDX_NX RT_BIT(20)
89	#undef MSR_K6_EFER
90	#define MSR_K6_EFER 0xc0000080
91	#undef MSR_K6_EFER_NXE
92	#define MSR_K6_EFER_NXE RT_BIT(11)
93	#undef MSR_K6_EFER_LMA
94	#define MSR_K6_EFER_LMA RT_BIT(10)
95	#undef X86_CR4_PGE
96	#define X86_CR4_PGE RT_BIT(7)
97	#undef X86_CR4_PAE
98	#define X86_CR4_PAE RT_BIT(5)
99	#undef X86_CPUID_AMD_FEATURE_EDX_LONG_MODE
100	#define X86_CPUID_AMD_FEATURE_EDX_LONG_MODE RT_BIT(29)
101
102
103	/** The frequency by which we recalculate the u32UpdateHz and
104	* u32UpdateIntervalNS GIP members. The value must be a power of 2. */
105	#define GIP_UPDATEHZ_RECALC_FREQ 0x800
106
107	/**
108	* Validates a session pointer.
109	*
110	* @returns true/false accordingly.
111	* @param pSession The session.
112	*/
113	#define SUP_IS_SESSION_VALID(pSession) \
114	( VALID_PTR(pSession) \
115	&& pSession->u32Cookie == BIRD_INV)
116
117	/** @def VBOX_SVN_REV
118	* The makefile should define this if it can. */
119	#ifndef VBOX_SVN_REV
120	# define VBOX_SVN_REV 0
121	#endif
122
123	/*******************************************************************************
124	* Internal Functions *
125	*******************************************************************************/
126	static int supdrvMemAdd(PSUPDRVMEMREF pMem, PSUPDRVSESSION pSession);
127	static int supdrvMemRelease(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr, SUPDRVMEMREFTYPE eType);
128	static int supdrvIOCtl_LdrOpen(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDROPEN pReq);
129	static int supdrvIOCtl_LdrLoad(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRLOAD pReq);
130	static int supdrvIOCtl_LdrFree(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRFREE pReq);
131	static int supdrvIOCtl_LdrGetSymbol(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRGETSYMBOL pReq);
132	static int supdrvIDC_LdrGetSymbol(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPDRVIDCREQGETSYM pReq);
133	static int supdrvLdrSetVMMR0EPs(PSUPDRVDEVEXT pDevExt, void pvVMMR0, void pvVMMR0EntryInt, void pvVMMR0EntryFast, void pvVMMR0EntryEx);
134	static void supdrvLdrUnsetVMMR0EPs(PSUPDRVDEVEXT pDevExt);
135	static int supdrvLdrAddUsage(PSUPDRVSESSION pSession, PSUPDRVLDRIMAGE pImage);
136	static void supdrvLdrFree(PSUPDRVDEVEXT pDevExt, PSUPDRVLDRIMAGE pImage);
137	static int supdrvIOCtl_CallServiceModule(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPCALLSERVICE pReq);
138	static SUPGIPMODE supdrvGipDeterminTscMode(PSUPDRVDEVEXT pDevExt);
139	#ifdef RT_OS_WINDOWS
140	static int supdrvPageGetPhys(PSUPDRVSESSION pSession, RTR3PTR pvR3, uint32_t cPages, PRTHCPHYS paPages);
141	static bool supdrvPageWasLockedByPageAlloc(PSUPDRVSESSION pSession, RTR3PTR pvR3);
142	#endif /* RT_OS_WINDOWS */
143	static int supdrvGipCreate(PSUPDRVDEVEXT pDevExt);
144	static void supdrvGipDestroy(PSUPDRVDEVEXT pDevExt);
145	static DECLCALLBACK(void) supdrvGipSyncTimer(PRTTIMER pTimer, void *pvUser, uint64_t iTick);
146	static DECLCALLBACK(void) supdrvGipAsyncTimer(PRTTIMER pTimer, void *pvUser, uint64_t iTick);
147	static DECLCALLBACK(void) supdrvGipMpEvent(RTMPEVENT enmEvent, RTCPUID idCpu, void *pvUser);
148
149	#ifdef RT_WITH_W64_UNWIND_HACK
150	DECLASM(int) supdrvNtWrapVMMR0EntryEx(PFNRT pfnVMMR0EntryEx, PVM pVM, unsigned uOperation, PSUPVMMR0REQHDR pReq, uint64_t u64Arg, PSUPDRVSESSION pSession);
151	DECLASM(int) supdrvNtWrapVMMR0EntryFast(PFNRT pfnVMMR0EntryFast, PVM pVM, unsigned idCpu, unsigned uOperation);
152	DECLASM(void) supdrvNtWrapObjDestructor(PFNRT pfnDestruction, void pvObj, void pvUser1, void *pvUser2);
153	DECLASM(void ) supdrvNtWrapQueryFactoryInterface(PFNRT pfnQueryFactoryInterface, struct SUPDRVFACTORY const pSupDrvFactory, PSUPDRVSESSION pSession, const char *pszInterfaceUuid);
154	DECLASM(int) supdrvNtWrapModuleInit(PFNRT pfnModuleInit);
155	DECLASM(void) supdrvNtWrapModuleTerm(PFNRT pfnModuleTerm);
156	DECLASM(int) supdrvNtWrapServiceReqHandler(PFNRT pfnServiceReqHandler, PSUPDRVSESSION pSession, uint32_t uOperation, uint64_t u64Arg, PSUPR0SERVICEREQHDR pReqHdr);
157
158	DECLASM(int) UNWIND_WRAP(SUPR0ComponentRegisterFactory)(PSUPDRVSESSION pSession, PCSUPDRVFACTORY pFactory);
159	DECLASM(int) UNWIND_WRAP(SUPR0ComponentDeregisterFactory)(PSUPDRVSESSION pSession, PCSUPDRVFACTORY pFactory);
160	DECLASM(int) UNWIND_WRAP(SUPR0ComponentQueryFactory)(PSUPDRVSESSION pSession, const char pszName, const char pszInterfaceUuid, void **ppvFactoryIf);
161	DECLASM(void ) UNWIND_WRAP(SUPR0ObjRegister)(PSUPDRVSESSION pSession, SUPDRVOBJTYPE enmType, PFNSUPDRVDESTRUCTOR pfnDestructor, void pvUser1, void *pvUser2);
162	DECLASM(int) UNWIND_WRAP(SUPR0ObjAddRef)(void *pvObj, PSUPDRVSESSION pSession);
163	DECLASM(int) UNWIND_WRAP(SUPR0ObjAddRefEx)(void *pvObj, PSUPDRVSESSION pSession, bool fNoPreempt);
164	DECLASM(int) UNWIND_WRAP(SUPR0ObjRelease)(void *pvObj, PSUPDRVSESSION pSession);
165	DECLASM(int) UNWIND_WRAP(SUPR0ObjVerifyAccess)(void pvObj, PSUPDRVSESSION pSession, const char pszObjName);
166	DECLASM(int) UNWIND_WRAP(SUPR0LockMem)(PSUPDRVSESSION pSession, RTR3PTR pvR3, uint32_t cPages, PRTHCPHYS paPages);
167	DECLASM(int) UNWIND_WRAP(SUPR0UnlockMem)(PSUPDRVSESSION pSession, RTR3PTR pvR3);
168	DECLASM(int) UNWIND_WRAP(SUPR0ContAlloc)(PSUPDRVSESSION pSession, uint32_t cPages, PRTR0PTR ppvR0, PRTR3PTR ppvR3, PRTHCPHYS pHCPhys);
169	DECLASM(int) UNWIND_WRAP(SUPR0ContFree)(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr);
170	DECLASM(int) UNWIND_WRAP(SUPR0LowAlloc)(PSUPDRVSESSION pSession, uint32_t cPages, PRTR0PTR ppvR0, PRTR3PTR ppvR3, PRTHCPHYS paPages);
171	DECLASM(int) UNWIND_WRAP(SUPR0LowFree)(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr);
172	DECLASM(int) UNWIND_WRAP(SUPR0MemAlloc)(PSUPDRVSESSION pSession, uint32_t cb, PRTR0PTR ppvR0, PRTR3PTR ppvR3);
173	DECLASM(int) UNWIND_WRAP(SUPR0MemGetPhys)(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr, PSUPPAGE paPages);
174	DECLASM(int) UNWIND_WRAP(SUPR0MemFree)(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr);
175	DECLASM(int) UNWIND_WRAP(SUPR0PageAlloc)(PSUPDRVSESSION pSession, uint32_t cPages, PRTR3PTR ppvR3, PRTHCPHYS paPages);
176	DECLASM(int) UNWIND_WRAP(SUPR0PageFree)(PSUPDRVSESSION pSession, RTR3PTR pvR3);
177	//DECLASM(int) UNWIND_WRAP(SUPR0Printf)(const char *pszFormat, ...);
178	DECLASM(SUPPAGINGMODE) UNWIND_WRAP(SUPR0GetPagingMode)(void);
179	DECLASM(void *) UNWIND_WRAP(RTMemAlloc)(size_t cb) RT_NO_THROW;
180	DECLASM(void *) UNWIND_WRAP(RTMemAllocZ)(size_t cb) RT_NO_THROW;
181	DECLASM(void) UNWIND_WRAP(RTMemFree)(void *pv) RT_NO_THROW;
182	DECLASM(void ) UNWIND_WRAP(RTMemDup)(const void pvSrc, size_t cb) RT_NO_THROW;
183	DECLASM(void ) UNWIND_WRAP(RTMemDupEx)(const void pvSrc, size_t cbSrc, size_t cbExtra) RT_NO_THROW;
184	DECLASM(void ) UNWIND_WRAP(RTMemRealloc)(void pvOld, size_t cbNew) RT_NO_THROW;
185	DECLASM(int) UNWIND_WRAP(RTR0MemObjAllocLow)(PRTR0MEMOBJ pMemObj, size_t cb, bool fExecutable);
186	DECLASM(int) UNWIND_WRAP(RTR0MemObjAllocPage)(PRTR0MEMOBJ pMemObj, size_t cb, bool fExecutable);
187	DECLASM(int) UNWIND_WRAP(RTR0MemObjAllocPhys)(PRTR0MEMOBJ pMemObj, size_t cb, RTHCPHYS PhysHighest);
188	DECLASM(int) UNWIND_WRAP(RTR0MemObjAllocPhysNC)(PRTR0MEMOBJ pMemObj, size_t cb, RTHCPHYS PhysHighest);
189	DECLASM(int) UNWIND_WRAP(RTR0MemObjAllocCont)(PRTR0MEMOBJ pMemObj, size_t cb, bool fExecutable);
190	DECLASM(int) UNWIND_WRAP(RTR0MemObjEnterPhys)(PRTR0MEMOBJ pMemObj, RTHCPHYS Phys, size_t cb);
191	DECLASM(int) UNWIND_WRAP(RTR0MemObjLockUser)(PRTR0MEMOBJ pMemObj, RTR3PTR R3Ptr, size_t cb, RTR0PROCESS R0Process);
192	DECLASM(int) UNWIND_WRAP(RTR0MemObjMapKernel)(PRTR0MEMOBJ pMemObj, RTR0MEMOBJ MemObjToMap, void *pvFixed, size_t uAlignment, unsigned fProt);
193	DECLASM(int) UNWIND_WRAP(RTR0MemObjMapKernelEx)(PRTR0MEMOBJ pMemObj, RTR0MEMOBJ MemObjToMap, void *pvFixed, size_t uAlignment, unsigned fProt, size_t offSub, size_t cbSub);
194	DECLASM(int) UNWIND_WRAP(RTR0MemObjMapUser)(PRTR0MEMOBJ pMemObj, RTR0MEMOBJ MemObjToMap, RTR3PTR R3PtrFixed, size_t uAlignment, unsigned fProt, RTR0PROCESS R0Process);
195	/DECLASM(void ) UNWIND_WRAP(RTR0MemObjAddress)(RTR0MEMOBJ MemObj); - not necessary */
196	/DECLASM(RTR3PTR) UNWIND_WRAP(RTR0MemObjAddressR3)(RTR0MEMOBJ MemObj); - not necessary /
197	/DECLASM(size_t) UNWIND_WRAP(RTR0MemObjSize)(RTR0MEMOBJ MemObj); - not necessary /
198	/DECLASM(bool) UNWIND_WRAP(RTR0MemObjIsMapping)(RTR0MEMOBJ MemObj); - not necessary /
199	/DECLASM(RTHCPHYS) UNWIND_WRAP(RTR0MemObjGetPagePhysAddr)(RTR0MEMOBJ MemObj, size_t iPage); - not necessary /
200	DECLASM(int) UNWIND_WRAP(RTR0MemObjFree)(RTR0MEMOBJ MemObj, bool fFreeMappings);
201	/* RTProcSelf - not necessary */
202	/* RTR0ProcHandleSelf - not necessary */
203	DECLASM(int) UNWIND_WRAP(RTSemFastMutexCreate)(PRTSEMFASTMUTEX pMutexSem);
204	DECLASM(int) UNWIND_WRAP(RTSemFastMutexDestroy)(RTSEMFASTMUTEX MutexSem);
205	DECLASM(int) UNWIND_WRAP(RTSemFastMutexRequest)(RTSEMFASTMUTEX MutexSem);
206	DECLASM(int) UNWIND_WRAP(RTSemFastMutexRelease)(RTSEMFASTMUTEX MutexSem);
207	DECLASM(int) UNWIND_WRAP(RTSemEventCreate)(PRTSEMEVENT pEventSem);
208	DECLASM(int) UNWIND_WRAP(RTSemEventSignal)(RTSEMEVENT EventSem);
209	DECLASM(int) UNWIND_WRAP(RTSemEventWait)(RTSEMEVENT EventSem, unsigned cMillies);
210	DECLASM(int) UNWIND_WRAP(RTSemEventWaitNoResume)(RTSEMEVENT EventSem, unsigned cMillies);
211	DECLASM(int) UNWIND_WRAP(RTSemEventDestroy)(RTSEMEVENT EventSem);
212	DECLASM(int) UNWIND_WRAP(RTSemEventMultiCreate)(PRTSEMEVENTMULTI pEventMultiSem);
213	DECLASM(int) UNWIND_WRAP(RTSemEventMultiSignal)(RTSEMEVENTMULTI EventMultiSem);
214	DECLASM(int) UNWIND_WRAP(RTSemEventMultiReset)(RTSEMEVENTMULTI EventMultiSem);
215	DECLASM(int) UNWIND_WRAP(RTSemEventMultiWait)(RTSEMEVENTMULTI EventMultiSem, unsigned cMillies);
216	DECLASM(int) UNWIND_WRAP(RTSemEventMultiWaitNoResume)(RTSEMEVENTMULTI EventMultiSem, unsigned cMillies);
217	DECLASM(int) UNWIND_WRAP(RTSemEventMultiDestroy)(RTSEMEVENTMULTI EventMultiSem);
218	DECLASM(int) UNWIND_WRAP(RTSpinlockCreate)(PRTSPINLOCK pSpinlock);
219	DECLASM(int) UNWIND_WRAP(RTSpinlockDestroy)(RTSPINLOCK Spinlock);
220	DECLASM(void) UNWIND_WRAP(RTSpinlockAcquire)(RTSPINLOCK Spinlock, PRTSPINLOCKTMP pTmp);
221	DECLASM(void) UNWIND_WRAP(RTSpinlockRelease)(RTSPINLOCK Spinlock, PRTSPINLOCKTMP pTmp);
222	DECLASM(void) UNWIND_WRAP(RTSpinlockAcquireNoInts)(RTSPINLOCK Spinlock, PRTSPINLOCKTMP pTmp);
223	DECLASM(void) UNWIND_WRAP(RTSpinlockReleaseNoInts)(RTSPINLOCK Spinlock, PRTSPINLOCKTMP pTmp);
224	/* RTTimeNanoTS - not necessary */
225	/* RTTimeMilliTS - not necessary */
226	/* RTTimeSystemNanoTS - not necessary */
227	/* RTTimeSystemMilliTS - not necessary */
228	/* RTThreadNativeSelf - not necessary */
229	DECLASM(int) UNWIND_WRAP(RTThreadSleep)(unsigned cMillies);
230	DECLASM(bool) UNWIND_WRAP(RTThreadYield)(void);
231	#if 0
232	/* RTThreadSelf - not necessary */
233	DECLASM(int) UNWIND_WRAP(RTThreadCreate)(PRTTHREAD pThread, PFNRTTHREAD pfnThread, void *pvUser, size_t cbStack,
234	RTTHREADTYPE enmType, unsigned fFlags, const char *pszName);
235	DECLASM(RTNATIVETHREAD) UNWIND_WRAP(RTThreadGetNative)(RTTHREAD Thread);
236	DECLASM(int) UNWIND_WRAP(RTThreadWait)(RTTHREAD Thread, unsigned cMillies, int *prc);
237	DECLASM(int) UNWIND_WRAP(RTThreadWaitNoResume)(RTTHREAD Thread, unsigned cMillies, int *prc);
238	DECLASM(const char *) UNWIND_WRAP(RTThreadGetName)(RTTHREAD Thread);
239	DECLASM(const char *) UNWIND_WRAP(RTThreadSelfName)(void);
240	DECLASM(RTTHREADTYPE) UNWIND_WRAP(RTThreadGetType)(RTTHREAD Thread);
241	DECLASM(int) UNWIND_WRAP(RTThreadUserSignal)(RTTHREAD Thread);
242	DECLASM(int) UNWIND_WRAP(RTThreadUserReset)(RTTHREAD Thread);
243	DECLASM(int) UNWIND_WRAP(RTThreadUserWait)(RTTHREAD Thread, unsigned cMillies);
244	DECLASM(int) UNWIND_WRAP(RTThreadUserWaitNoResume)(RTTHREAD Thread, unsigned cMillies);
245	#endif
246	/* RTLogDefaultInstance - a bit of a gamble, but we do not want the overhead! */
247	/* RTMpCpuId - not necessary */
248	/* RTMpCpuIdFromSetIndex - not necessary */
249	/* RTMpCpuIdToSetIndex - not necessary */
250	/* RTMpIsCpuPossible - not necessary */
251	/* RTMpGetCount - not necessary */
252	/* RTMpGetMaxCpuId - not necessary */
253	/* RTMpGetOnlineCount - not necessary */
254	/* RTMpGetOnlineSet - not necessary */
255	/* RTMpGetSet - not necessary */
256	/* RTMpIsCpuOnline - not necessary */
257	DECLASM(int) UNWIND_WRAP(RTMpOnAll)(PFNRTMPWORKER pfnWorker, void pvUser1, void pvUser2);
258	DECLASM(int) UNWIND_WRAP(RTMpOnOthers)(PFNRTMPWORKER pfnWorker, void pvUser1, void pvUser2);
259	DECLASM(int) UNWIND_WRAP(RTMpOnSpecific)(RTCPUID idCpu, PFNRTMPWORKER pfnWorker, void pvUser1, void pvUser2);
260	DECLASM(int) UNWIND_WRAP(RTMpIsCpuWorkPending)(void);
261	/* RTLogRelDefaultInstance - not necessary. */
262	DECLASM(int) UNWIND_WRAP(RTLogSetDefaultInstanceThread)(PRTLOGGER pLogger, uintptr_t uKey);
263	/* RTLogLogger - can't wrap this buster. */
264	/* RTLogLoggerEx - can't wrap this buster. */
265	DECLASM(void) UNWIND_WRAP(RTLogLoggerExV)(PRTLOGGER pLogger, unsigned fFlags, unsigned iGroup, const char *pszFormat, va_list args);
266	/* RTLogPrintf - can't wrap this buster. / /* @todo provide va_list log wrappers in RuntimeR0. */
267	DECLASM(void) UNWIND_WRAP(RTLogPrintfV)(const char *pszFormat, va_list args);
268	DECLASM(void) UNWIND_WRAP(AssertMsg1)(const char pszExpr, unsigned uLine, const char pszFile, const char *pszFunction);
269	/* AssertMsg2 - can't wrap this buster. */
270	#endif /* RT_WITH_W64_UNWIND_HACK */
271
272
273	/*******************************************************************************
274	* Global Variables *
275	*******************************************************************************/
276	/**
277	* Array of the R0 SUP API.
278	*/
279	static SUPFUNC g_aFunctions[] =
280	{
281	/* name function */
282	/* Entries with absolute addresses determined at runtime, fixup
283	code makes ugly ASSUMPTIONS about the order here: */
284	{ "SUPR0AbsIs64bit", (void *)0 },
285	{ "SUPR0Abs64bitKernelCS", (void *)0 },
286	{ "SUPR0Abs64bitKernelSS", (void *)0 },
287	{ "SUPR0Abs64bitKernelDS", (void *)0 },
288	{ "SUPR0AbsKernelCS", (void *)0 },
289	{ "SUPR0AbsKernelSS", (void *)0 },
290	{ "SUPR0AbsKernelDS", (void *)0 },
291	{ "SUPR0AbsKernelES", (void *)0 },
292	{ "SUPR0AbsKernelFS", (void *)0 },
293	{ "SUPR0AbsKernelGS", (void *)0 },
294	/* Normal function pointers: */
295	{ "SUPR0ComponentRegisterFactory", (void *)UNWIND_WRAP(SUPR0ComponentRegisterFactory) },
296	{ "SUPR0ComponentDeregisterFactory", (void *)UNWIND_WRAP(SUPR0ComponentDeregisterFactory) },
297	{ "SUPR0ComponentQueryFactory", (void *)UNWIND_WRAP(SUPR0ComponentQueryFactory) },
298	{ "SUPR0ObjRegister", (void *)UNWIND_WRAP(SUPR0ObjRegister) },
299	{ "SUPR0ObjAddRef", (void *)UNWIND_WRAP(SUPR0ObjAddRef) },
300	{ "SUPR0ObjAddRefEx", (void *)UNWIND_WRAP(SUPR0ObjAddRefEx) },
301	{ "SUPR0ObjRelease", (void *)UNWIND_WRAP(SUPR0ObjRelease) },
302	{ "SUPR0ObjVerifyAccess", (void *)UNWIND_WRAP(SUPR0ObjVerifyAccess) },
303	{ "SUPR0LockMem", (void *)UNWIND_WRAP(SUPR0LockMem) },
304	{ "SUPR0UnlockMem", (void *)UNWIND_WRAP(SUPR0UnlockMem) },
305	{ "SUPR0ContAlloc", (void *)UNWIND_WRAP(SUPR0ContAlloc) },
306	{ "SUPR0ContFree", (void *)UNWIND_WRAP(SUPR0ContFree) },
307	{ "SUPR0LowAlloc", (void *)UNWIND_WRAP(SUPR0LowAlloc) },
308	{ "SUPR0LowFree", (void *)UNWIND_WRAP(SUPR0LowFree) },
309	{ "SUPR0MemAlloc", (void *)UNWIND_WRAP(SUPR0MemAlloc) },
310	{ "SUPR0MemGetPhys", (void *)UNWIND_WRAP(SUPR0MemGetPhys) },
311	{ "SUPR0MemFree", (void *)UNWIND_WRAP(SUPR0MemFree) },
312	{ "SUPR0PageAlloc", (void *)UNWIND_WRAP(SUPR0PageAlloc) },
313	{ "SUPR0PageFree", (void *)UNWIND_WRAP(SUPR0PageFree) },
314	{ "SUPR0Printf", (void )SUPR0Printf }, /* @todo needs wrapping? */
315	{ "SUPR0GetPagingMode", (void *)UNWIND_WRAP(SUPR0GetPagingMode) },
316	{ "SUPR0EnableVTx", (void *)SUPR0EnableVTx },
317	{ "RTMemAlloc", (void *)UNWIND_WRAP(RTMemAlloc) },
318	{ "RTMemAllocZ", (void *)UNWIND_WRAP(RTMemAllocZ) },
319	{ "RTMemFree", (void *)UNWIND_WRAP(RTMemFree) },
320	/{ "RTMemDup", (void )UNWIND_WRAP(RTMemDup) },
321	{ "RTMemDupEx", (void )UNWIND_WRAP(RTMemDupEx) },/
322	{ "RTMemRealloc", (void *)UNWIND_WRAP(RTMemRealloc) },
323	{ "RTR0MemObjAllocLow", (void *)UNWIND_WRAP(RTR0MemObjAllocLow) },
324	{ "RTR0MemObjAllocPage", (void *)UNWIND_WRAP(RTR0MemObjAllocPage) },
325	{ "RTR0MemObjAllocPhys", (void *)UNWIND_WRAP(RTR0MemObjAllocPhys) },
326	{ "RTR0MemObjAllocPhysNC", (void *)UNWIND_WRAP(RTR0MemObjAllocPhysNC) },
327	{ "RTR0MemObjAllocCont", (void *)UNWIND_WRAP(RTR0MemObjAllocCont) },
328	{ "RTR0MemObjEnterPhys", (void *)UNWIND_WRAP(RTR0MemObjEnterPhys) },
329	{ "RTR0MemObjLockUser", (void *)UNWIND_WRAP(RTR0MemObjLockUser) },
330	{ "RTR0MemObjMapKernel", (void *)UNWIND_WRAP(RTR0MemObjMapKernel) },
331	{ "RTR0MemObjMapKernelEx", (void *)UNWIND_WRAP(RTR0MemObjMapKernelEx) },
332	{ "RTR0MemObjMapUser", (void *)UNWIND_WRAP(RTR0MemObjMapUser) },
333	{ "RTR0MemObjAddress", (void *)RTR0MemObjAddress },
334	{ "RTR0MemObjAddressR3", (void *)RTR0MemObjAddressR3 },
335	{ "RTR0MemObjSize", (void *)RTR0MemObjSize },
336	{ "RTR0MemObjIsMapping", (void *)RTR0MemObjIsMapping },
337	{ "RTR0MemObjGetPagePhysAddr", (void *)RTR0MemObjGetPagePhysAddr },
338	{ "RTR0MemObjFree", (void *)UNWIND_WRAP(RTR0MemObjFree) },
339	/* These don't work yet on linux - use fast mutexes!
340	{ "RTSemMutexCreate", (void *)RTSemMutexCreate },
341	{ "RTSemMutexRequest", (void *)RTSemMutexRequest },
342	{ "RTSemMutexRelease", (void *)RTSemMutexRelease },
343	{ "RTSemMutexDestroy", (void *)RTSemMutexDestroy },
344	*/
345	{ "RTProcSelf", (void *)RTProcSelf },
346	{ "RTR0ProcHandleSelf", (void *)RTR0ProcHandleSelf },
347	{ "RTSemFastMutexCreate", (void *)UNWIND_WRAP(RTSemFastMutexCreate) },
348	{ "RTSemFastMutexDestroy", (void *)UNWIND_WRAP(RTSemFastMutexDestroy) },
349	{ "RTSemFastMutexRequest", (void *)UNWIND_WRAP(RTSemFastMutexRequest) },
350	{ "RTSemFastMutexRelease", (void *)UNWIND_WRAP(RTSemFastMutexRelease) },
351	{ "RTSemEventCreate", (void *)UNWIND_WRAP(RTSemEventCreate) },
352	{ "RTSemEventSignal", (void *)UNWIND_WRAP(RTSemEventSignal) },
353	{ "RTSemEventWait", (void *)UNWIND_WRAP(RTSemEventWait) },
354	{ "RTSemEventWaitNoResume", (void *)UNWIND_WRAP(RTSemEventWaitNoResume) },
355	{ "RTSemEventDestroy", (void *)UNWIND_WRAP(RTSemEventDestroy) },
356	{ "RTSemEventMultiCreate", (void *)UNWIND_WRAP(RTSemEventMultiCreate) },
357	{ "RTSemEventMultiSignal", (void *)UNWIND_WRAP(RTSemEventMultiSignal) },
358	{ "RTSemEventMultiReset", (void *)UNWIND_WRAP(RTSemEventMultiReset) },
359	{ "RTSemEventMultiWait", (void *)UNWIND_WRAP(RTSemEventMultiWait) },
360	{ "RTSemEventMultiWaitNoResume", (void *)UNWIND_WRAP(RTSemEventMultiWaitNoResume) },
361	{ "RTSemEventMultiDestroy", (void *)UNWIND_WRAP(RTSemEventMultiDestroy) },
362	{ "RTSpinlockCreate", (void *)UNWIND_WRAP(RTSpinlockCreate) },
363	{ "RTSpinlockDestroy", (void *)UNWIND_WRAP(RTSpinlockDestroy) },
364	{ "RTSpinlockAcquire", (void *)UNWIND_WRAP(RTSpinlockAcquire) },
365	{ "RTSpinlockRelease", (void *)UNWIND_WRAP(RTSpinlockRelease) },
366	{ "RTSpinlockAcquireNoInts", (void *)UNWIND_WRAP(RTSpinlockAcquireNoInts) },
367	{ "RTSpinlockReleaseNoInts", (void *)UNWIND_WRAP(RTSpinlockReleaseNoInts) },
368	{ "RTTimeNanoTS", (void *)RTTimeNanoTS },
369	{ "RTTimeMillieTS", (void *)RTTimeMilliTS },
370	{ "RTTimeSystemNanoTS", (void *)RTTimeSystemNanoTS },
371	{ "RTTimeSystemMillieTS", (void *)RTTimeSystemMilliTS },
372	{ "RTThreadNativeSelf", (void *)RTThreadNativeSelf },
373	{ "RTThreadSleep", (void *)UNWIND_WRAP(RTThreadSleep) },
374	{ "RTThreadYield", (void *)UNWIND_WRAP(RTThreadYield) },
375	#if 0 /* Thread APIs, Part 2. */
376	{ "RTThreadSelf", (void *)UNWIND_WRAP(RTThreadSelf) },
377	{ "RTThreadCreate", (void )UNWIND_WRAP(RTThreadCreate) }, /* @todo need to wrap the callback */
378	{ "RTThreadGetNative", (void *)UNWIND_WRAP(RTThreadGetNative) },
379	{ "RTThreadWait", (void *)UNWIND_WRAP(RTThreadWait) },
380	{ "RTThreadWaitNoResume", (void *)UNWIND_WRAP(RTThreadWaitNoResume) },
381	{ "RTThreadGetName", (void *)UNWIND_WRAP(RTThreadGetName) },
382	{ "RTThreadSelfName", (void *)UNWIND_WRAP(RTThreadSelfName) },
383	{ "RTThreadGetType", (void *)UNWIND_WRAP(RTThreadGetType) },
384	{ "RTThreadUserSignal", (void *)UNWIND_WRAP(RTThreadUserSignal) },
385	{ "RTThreadUserReset", (void *)UNWIND_WRAP(RTThreadUserReset) },
386	{ "RTThreadUserWait", (void *)UNWIND_WRAP(RTThreadUserWait) },
387	{ "RTThreadUserWaitNoResume", (void *)UNWIND_WRAP(RTThreadUserWaitNoResume) },
388	#endif
389	{ "RTLogDefaultInstance", (void *)RTLogDefaultInstance },
390	{ "RTMpCpuId", (void *)RTMpCpuId },
391	{ "RTMpCpuIdFromSetIndex", (void *)RTMpCpuIdFromSetIndex },
392	{ "RTMpCpuIdToSetIndex", (void *)RTMpCpuIdToSetIndex },
393	{ "RTMpIsCpuPossible", (void *)RTMpIsCpuPossible },
394	{ "RTMpGetCount", (void *)RTMpGetCount },
395	{ "RTMpGetMaxCpuId", (void *)RTMpGetMaxCpuId },
396	{ "RTMpGetOnlineCount", (void *)RTMpGetOnlineCount },
397	{ "RTMpGetOnlineSet", (void *)RTMpGetOnlineSet },
398	{ "RTMpGetSet", (void *)RTMpGetSet },
399	{ "RTMpIsCpuOnline", (void *)RTMpIsCpuOnline },
400	{ "RTMpIsCpuWorkPending", (void *)UNWIND_WRAP(RTMpIsCpuWorkPending) },
401	{ "RTMpOnAll", (void *)UNWIND_WRAP(RTMpOnAll) },
402	{ "RTMpOnOthers", (void *)UNWIND_WRAP(RTMpOnOthers) },
403	{ "RTMpOnSpecific", (void *)UNWIND_WRAP(RTMpOnSpecific) },
404	{ "RTPowerNotificationRegister", (void *)RTPowerNotificationRegister },
405	{ "RTPowerNotificationDeregister", (void *)RTPowerNotificationDeregister },
406	{ "RTLogRelDefaultInstance", (void *)RTLogRelDefaultInstance },
407	{ "RTLogSetDefaultInstanceThread", (void *)UNWIND_WRAP(RTLogSetDefaultInstanceThread) },
408	{ "RTLogLogger", (void )RTLogLogger }, /* @todo remove this */
409	{ "RTLogLoggerEx", (void )RTLogLoggerEx }, /* @todo remove this */
410	{ "RTLogLoggerExV", (void *)UNWIND_WRAP(RTLogLoggerExV) },
411	{ "RTLogPrintf", (void )RTLogPrintf }, /* @todo remove this */
412	{ "RTLogPrintfV", (void *)UNWIND_WRAP(RTLogPrintfV) },
413	{ "AssertMsg1", (void *)UNWIND_WRAP(AssertMsg1) },
414	{ "AssertMsg2", (void )AssertMsg2 }, /* @todo replace this by RTAssertMsg2V */
415	#if defined(RT_OS_DARWIN) \|\| defined(RT_OS_SOLARIS)
416	{ "RTR0AssertPanicSystem", (void *)RTR0AssertPanicSystem },
417	#endif
418	#if defined(RT_OS_DARWIN)
419	{ "RTAssertMsg1", (void *)RTAssertMsg1 },
420	{ "RTAssertMsg2", (void *)RTAssertMsg2 },
421	{ "RTAssertMsg2V", (void *)RTAssertMsg2V },
422	#endif
423	};
424
425	#if defined(RT_OS_DARWIN) \|\| defined(RT_OS_SOLARIS)
426	/**
427	* Drag in the rest of IRPT since we share it with the
428	* rest of the kernel modules on darwin.
429	*/
430	PFNRT g_apfnVBoxDrvIPRTDeps[] =
431	{
432	(PFNRT)RTCrc32,
433	(PFNRT)RTErrConvertFromErrno,
434	(PFNRT)RTNetIPv4IsHdrValid,
435	(PFNRT)RTNetIPv4TCPChecksum,
436	(PFNRT)RTNetIPv4UDPChecksum,
437	(PFNRT)RTUuidCompare,
438	(PFNRT)RTUuidCompareStr,
439	(PFNRT)RTUuidFromStr,
440	NULL
441	};
442	#endif /* RT_OS_DARWIN \|\| RT_OS_SOLARIS */
443
444
445	/**
446	* Initializes the device extentsion structure.
447	*
448	* @returns IPRT status code.
449	* @param pDevExt The device extension to initialize.
450	*/
451	int VBOXCALL supdrvInitDevExt(PSUPDRVDEVEXT pDevExt)
452	{
453	int rc;
454
455	#ifdef SUPDRV_WITH_RELEASE_LOGGER
456	/*
457	* Create the release log.
458	*/
459	static const char * const s_apszGroups[] = VBOX_LOGGROUP_NAMES;
460	PRTLOGGER pRelLogger;
461	rc = RTLogCreate(&pRelLogger, 0 /* fFlags */, "all",
462	"VBOX_RELEASE_LOG", RT_ELEMENTS(s_apszGroups), s_apszGroups,
463	RTLOGDEST_STDOUT \| RTLOGDEST_DEBUGGER, NULL);
464	if (RT_SUCCESS(rc))
465	RTLogRelSetDefaultInstance(pRelLogger);
466	#endif
467
468	/*
469	* Initialize it.
470	*/
471	memset(pDevExt, 0, sizeof(*pDevExt));
472	rc = RTSpinlockCreate(&pDevExt->Spinlock);
473	if (!rc)
474	{
475	rc = RTSemFastMutexCreate(&pDevExt->mtxLdr);
476	if (!rc)
477	{
478	rc = RTSemFastMutexCreate(&pDevExt->mtxComponentFactory);
479	if (!rc)
480	{
481	rc = RTSemFastMutexCreate(&pDevExt->mtxGip);
482	if (!rc)
483	{
484	rc = supdrvGipCreate(pDevExt);
485	if (RT_SUCCESS(rc))
486	{
487	pDevExt->u32Cookie = BIRD; /** @todo make this random? */
488
489	/*
490	* Fixup the absolute symbols.
491	*
492	* Because of the table indexing assumptions we'll do #ifdef orgy here rather
493	* than distributing this to OS specific files. At least for now.
494	*/
495	#ifdef RT_OS_DARWIN
496	if (SUPR0GetPagingMode() >= SUPPAGINGMODE_AMD64)
497	{
498	g_aFunctions[0].pfn = (void )1; / SUPR0AbsIs64bit */
499	g_aFunctions[1].pfn = (void )0x80; / SUPR0Abs64bitKernelCS - KERNEL64_CS, seg.h */
500	g_aFunctions[2].pfn = (void )0x88; / SUPR0Abs64bitKernelSS - KERNEL64_SS, seg.h */
501	g_aFunctions[3].pfn = (void )0x88; / SUPR0Abs64bitKernelDS - KERNEL64_SS, seg.h */
502	}
503	else
504	g_aFunctions[0].pfn = g_aFunctions[1].pfn = g_aFunctions[2].pfn = g_aFunctions[4].pfn = (void *)0;
505	g_aFunctions[4].pfn = (void )0x08; / SUPR0AbsKernelCS - KERNEL_CS, seg.h */
506	g_aFunctions[5].pfn = (void )0x10; / SUPR0AbsKernelSS - KERNEL_DS, seg.h */
507	g_aFunctions[6].pfn = (void )0x10; / SUPR0AbsKernelDS - KERNEL_DS, seg.h */
508	g_aFunctions[7].pfn = (void )0x10; / SUPR0AbsKernelES - KERNEL_DS, seg.h */
509	#else
510	# if ARCH_BITS == 64
511	g_aFunctions[0].pfn = (void )1; / SUPR0AbsIs64bit */
512	g_aFunctions[1].pfn = (void )(uintptr_t)ASMGetCS(); / SUPR0Abs64bitKernelCS */
513	g_aFunctions[2].pfn = (void )(uintptr_t)ASMGetSS(); / SUPR0Abs64bitKernelSS */
514	g_aFunctions[3].pfn = (void )(uintptr_t)ASMGetDS(); / SUPR0Abs64bitKernelDS */
515	# elif ARCH_BITS == 32
516	g_aFunctions[0].pfn = g_aFunctions[1].pfn = g_aFunctions[2].pfn = g_aFunctions[4].pfn = (void *)0;
517	# endif
518	g_aFunctions[4].pfn = (void )(uintptr_t)ASMGetCS(); / SUPR0AbsKernelCS */
519	g_aFunctions[5].pfn = (void )(uintptr_t)ASMGetSS(); / SUPR0AbsKernelSS */
520	g_aFunctions[6].pfn = (void )(uintptr_t)ASMGetDS(); / SUPR0AbsKernelDS */
521	g_aFunctions[7].pfn = (void )(uintptr_t)ASMGetES(); / SUPR0AbsKernelES */
522	#endif
523	g_aFunctions[8].pfn = (void )(uintptr_t)ASMGetFS(); / SUPR0AbsKernelFS */
524	g_aFunctions[9].pfn = (void )(uintptr_t)ASMGetGS(); / SUPR0AbsKernelGS */
525	return VINF_SUCCESS;
526	}
527
528	RTSemFastMutexDestroy(pDevExt->mtxGip);
529	pDevExt->mtxGip = NIL_RTSEMFASTMUTEX;
530	}
531	RTSemFastMutexDestroy(pDevExt->mtxComponentFactory);
532	pDevExt->mtxComponentFactory = NIL_RTSEMFASTMUTEX;
533	}
534	RTSemFastMutexDestroy(pDevExt->mtxLdr);
535	pDevExt->mtxLdr = NIL_RTSEMFASTMUTEX;
536	}
537	RTSpinlockDestroy(pDevExt->Spinlock);
538	pDevExt->Spinlock = NIL_RTSPINLOCK;
539	}
540	#ifdef SUPDRV_WITH_RELEASE_LOGGER
541	RTLogDestroy(RTLogRelSetDefaultInstance(NULL));
542	RTLogDestroy(RTLogSetDefaultInstance(NULL));
543	#endif
544
545	return rc;
546	}
547
548
549	/**
550	* Delete the device extension (e.g. cleanup members).
551	*
552	* @param pDevExt The device extension to delete.
553	*/
554	void VBOXCALL supdrvDeleteDevExt(PSUPDRVDEVEXT pDevExt)
555	{
556	PSUPDRVOBJ pObj;
557	PSUPDRVUSAGE pUsage;
558
559	/*
560	* Kill mutexes and spinlocks.
561	*/
562	RTSemFastMutexDestroy(pDevExt->mtxGip);
563	pDevExt->mtxGip = NIL_RTSEMFASTMUTEX;
564	RTSemFastMutexDestroy(pDevExt->mtxLdr);
565	pDevExt->mtxLdr = NIL_RTSEMFASTMUTEX;
566	RTSpinlockDestroy(pDevExt->Spinlock);
567	pDevExt->Spinlock = NIL_RTSPINLOCK;
568	RTSemFastMutexDestroy(pDevExt->mtxComponentFactory);
569	pDevExt->mtxComponentFactory = NIL_RTSEMFASTMUTEX;
570
571	/*
572	* Free lists.
573	*/
574	/* objects. */
575	pObj = pDevExt->pObjs;
576	#if !defined(DEBUG_bird) \|\| !defined(RT_OS_LINUX) /* breaks unloading, temporary, remove me! */
577	Assert(!pObj); /* (can trigger on forced unloads) */
578	#endif
579	pDevExt->pObjs = NULL;
580	while (pObj)
581	{
582	void *pvFree = pObj;
583	pObj = pObj->pNext;
584	RTMemFree(pvFree);
585	}
586
587	/* usage records. */
588	pUsage = pDevExt->pUsageFree;
589	pDevExt->pUsageFree = NULL;
590	while (pUsage)
591	{
592	void *pvFree = pUsage;
593	pUsage = pUsage->pNext;
594	RTMemFree(pvFree);
595	}
596
597	/* kill the GIP. */
598	supdrvGipDestroy(pDevExt);
599
600	#ifdef SUPDRV_WITH_RELEASE_LOGGER
601	/* destroy the loggers. */
602	RTLogDestroy(RTLogRelSetDefaultInstance(NULL));
603	RTLogDestroy(RTLogSetDefaultInstance(NULL));
604	#endif
605	}
606
607
608	/**
609	* Create session.
610	*
611	* @returns IPRT status code.
612	* @param pDevExt Device extension.
613	* @param fUser Flag indicating whether this is a user or kernel session.
614	* @param ppSession Where to store the pointer to the session data.
615	*/
616	int VBOXCALL supdrvCreateSession(PSUPDRVDEVEXT pDevExt, bool fUser, PSUPDRVSESSION *ppSession)
617	{
618	/*
619	* Allocate memory for the session data.
620	*/
621	int rc = VERR_NO_MEMORY;
622	PSUPDRVSESSION pSession = ppSession = (PSUPDRVSESSION)RTMemAllocZ(sizeof(pSession));
623	if (pSession)
624	{
625	/* Initialize session data. */
626	rc = RTSpinlockCreate(&pSession->Spinlock);
627	if (!rc)
628	{
629	Assert(pSession->Spinlock != NIL_RTSPINLOCK);
630	pSession->pDevExt = pDevExt;
631	pSession->u32Cookie = BIRD_INV;
632	/*pSession->pLdrUsage = NULL;
633	pSession->pVM = NULL;
634	pSession->pUsage = NULL;
635	pSession->pGip = NULL;
636	pSession->fGipReferenced = false;
637	pSession->Bundle.cUsed = 0; */
638	pSession->Uid = NIL_RTUID;
639	pSession->Gid = NIL_RTGID;
640	if (fUser)
641	{
642	pSession->Process = RTProcSelf();
643	pSession->R0Process = RTR0ProcHandleSelf();
644	}
645	else
646	{
647	pSession->Process = NIL_RTPROCESS;
648	pSession->R0Process = NIL_RTR0PROCESS;
649	}
650
651	LogFlow(("Created session %p initial cookie=%#x\n", pSession, pSession->u32Cookie));
652	return VINF_SUCCESS;
653	}
654
655	RTMemFree(pSession);
656	*ppSession = NULL;
657	Log(("Failed to create spinlock, rc=%d!\n", rc));
658	}
659
660	return rc;
661	}
662
663
664	/**
665	* Shared code for cleaning up a session.
666	*
667	* @param pDevExt Device extension.
668	* @param pSession Session data.
669	* This data will be freed by this routine.
670	*/
671	void VBOXCALL supdrvCloseSession(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession)
672	{
673	/*
674	* Cleanup the session first.
675	*/
676	supdrvCleanupSession(pDevExt, pSession);
677
678	/*
679	* Free the rest of the session stuff.
680	*/
681	RTSpinlockDestroy(pSession->Spinlock);
682	pSession->Spinlock = NIL_RTSPINLOCK;
683	pSession->pDevExt = NULL;
684	RTMemFree(pSession);
685	LogFlow(("supdrvCloseSession: returns\n"));
686	}
687
688
689	/**
690	* Shared code for cleaning up a session (but not quite freeing it).
691	*
692	* This is primarily intended for MAC OS X where we have to clean up the memory
693	* stuff before the file handle is closed.
694	*
695	* @param pDevExt Device extension.
696	* @param pSession Session data.
697	* This data will be freed by this routine.
698	*/
699	void VBOXCALL supdrvCleanupSession(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession)
700	{
701	PSUPDRVBUNDLE pBundle;
702	LogFlow(("supdrvCleanupSession: pSession=%p\n", pSession));
703
704	/*
705	* Remove logger instances related to this session.
706	*/
707	RTLogSetDefaultInstanceThread(NULL, (uintptr_t)pSession);
708
709	/*
710	* Release object references made in this session.
711	* In theory there should be noone racing us in this session.
712	*/
713	Log2(("release objects - start\n"));
714	if (pSession->pUsage)
715	{
716	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
717	PSUPDRVUSAGE pUsage;
718	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
719
720	while ((pUsage = pSession->pUsage) != NULL)
721	{
722	PSUPDRVOBJ pObj = pUsage->pObj;
723	pSession->pUsage = pUsage->pNext;
724
725	AssertMsg(pUsage->cUsage >= 1 && pObj->cUsage >= pUsage->cUsage, ("glob %d; sess %d\n", pObj->cUsage, pUsage->cUsage));
726	if (pUsage->cUsage < pObj->cUsage)
727	{
728	pObj->cUsage -= pUsage->cUsage;
729	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
730	}
731	else
732	{
733	/* Destroy the object and free the record. */
734	if (pDevExt->pObjs == pObj)
735	pDevExt->pObjs = pObj->pNext;
736	else
737	{
738	PSUPDRVOBJ pObjPrev;
739	for (pObjPrev = pDevExt->pObjs; pObjPrev; pObjPrev = pObjPrev->pNext)
740	if (pObjPrev->pNext == pObj)
741	{
742	pObjPrev->pNext = pObj->pNext;
743	break;
744	}
745	Assert(pObjPrev);
746	}
747	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
748
749	Log(("supdrvCleanupSession: destroying %p/%d (%p/%p) cpid=%RTproc pid=%RTproc dtor=%p\n",
750	pObj, pObj->enmType, pObj->pvUser1, pObj->pvUser2, pObj->CreatorProcess, RTProcSelf(), pObj->pfnDestructor));
751	if (pObj->pfnDestructor)
752	#ifdef RT_WITH_W64_UNWIND_HACK
753	supdrvNtWrapObjDestructor((PFNRT)pObj->pfnDestructor, pObj, pObj->pvUser1, pObj->pvUser2);
754	#else
755	pObj->pfnDestructor(pObj, pObj->pvUser1, pObj->pvUser2);
756	#endif
757	RTMemFree(pObj);
758	}
759
760	/* free it and continue. */
761	RTMemFree(pUsage);
762
763	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
764	}
765
766	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
767	AssertMsg(!pSession->pUsage, ("Some buster reregistered an object during desturction!\n"));
768	}
769	Log2(("release objects - done\n"));
770
771	/*
772	* Release memory allocated in the session.
773	*
774	* We do not serialize this as we assume that the application will
775	* not allocated memory while closing the file handle object.
776	*/
777	Log2(("freeing memory:\n"));
778	pBundle = &pSession->Bundle;
779	while (pBundle)
780	{
781	PSUPDRVBUNDLE pToFree;
782	unsigned i;
783
784	/*
785	* Check and unlock all entries in the bundle.
786	*/
787	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
788	{
789	if (pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ)
790	{
791	int rc;
792	Log2(("eType=%d pvR0=%p pvR3=%p cb=%ld\n", pBundle->aMem[i].eType, RTR0MemObjAddress(pBundle->aMem[i].MemObj),
793	(void *)RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3), (long)RTR0MemObjSize(pBundle->aMem[i].MemObj)));
794	if (pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ)
795	{
796	rc = RTR0MemObjFree(pBundle->aMem[i].MapObjR3, false);
797	AssertRC(rc); /** @todo figure out how to handle this. */
798	pBundle->aMem[i].MapObjR3 = NIL_RTR0MEMOBJ;
799	}
800	rc = RTR0MemObjFree(pBundle->aMem[i].MemObj, true /* fFreeMappings */);
801	AssertRC(rc); /** @todo figure out how to handle this. */
802	pBundle->aMem[i].MemObj = NIL_RTR0MEMOBJ;
803	pBundle->aMem[i].eType = MEMREF_TYPE_UNUSED;
804	}
805	}
806
807	/*
808	* Advance and free previous bundle.
809	*/
810	pToFree = pBundle;
811	pBundle = pBundle->pNext;
812
813	pToFree->pNext = NULL;
814	pToFree->cUsed = 0;
815	if (pToFree != &pSession->Bundle)
816	RTMemFree(pToFree);
817	}
818	Log2(("freeing memory - done\n"));
819
820	/*
821	* Deregister component factories.
822	*/
823	RTSemFastMutexRequest(pDevExt->mtxComponentFactory);
824	Log2(("deregistering component factories:\n"));
825	if (pDevExt->pComponentFactoryHead)
826	{
827	PSUPDRVFACTORYREG pPrev = NULL;
828	PSUPDRVFACTORYREG pCur = pDevExt->pComponentFactoryHead;
829	while (pCur)
830	{
831	if (pCur->pSession == pSession)
832	{
833	/* unlink it */
834	PSUPDRVFACTORYREG pNext = pCur->pNext;
835	if (pPrev)
836	pPrev->pNext = pNext;
837	else
838	pDevExt->pComponentFactoryHead = pNext;
839
840	/* free it */
841	pCur->pNext = NULL;
842	pCur->pSession = NULL;
843	pCur->pFactory = NULL;
844	RTMemFree(pCur);
845
846	/* next */
847	pCur = pNext;
848	}
849	else
850	{
851	/* next */
852	pPrev = pCur;
853	pCur = pCur->pNext;
854	}
855	}
856	}
857	RTSemFastMutexRelease(pDevExt->mtxComponentFactory);
858	Log2(("deregistering component factories - done\n"));
859
860	/*
861	* Loaded images needs to be dereferenced and possibly freed up.
862	*/
863	RTSemFastMutexRequest(pDevExt->mtxLdr);
864	Log2(("freeing images:\n"));
865	if (pSession->pLdrUsage)
866	{
867	PSUPDRVLDRUSAGE pUsage = pSession->pLdrUsage;
868	pSession->pLdrUsage = NULL;
869	while (pUsage)
870	{
871	void *pvFree = pUsage;
872	PSUPDRVLDRIMAGE pImage = pUsage->pImage;
873	if (pImage->cUsage > pUsage->cUsage)
874	pImage->cUsage -= pUsage->cUsage;
875	else
876	supdrvLdrFree(pDevExt, pImage);
877	pUsage->pImage = NULL;
878	pUsage = pUsage->pNext;
879	RTMemFree(pvFree);
880	}
881	}
882	RTSemFastMutexRelease(pDevExt->mtxLdr);
883	Log2(("freeing images - done\n"));
884
885	/*
886	* Unmap the GIP.
887	*/
888	Log2(("umapping GIP:\n"));
889	if (pSession->GipMapObjR3 != NIL_RTR0MEMOBJ)
890	{
891	SUPR0GipUnmap(pSession);
892	pSession->fGipReferenced = 0;
893	}
894	Log2(("umapping GIP - done\n"));
895	}
896
897
898	/**
899	* Fast path I/O Control worker.
900	*
901	* @returns VBox status code that should be passed down to ring-3 unchanged.
902	* @param uIOCtl Function number.
903	* @param idCpu VMCPU id.
904	* @param pDevExt Device extention.
905	* @param pSession Session data.
906	*/
907	int VBOXCALL supdrvIOCtlFast(uintptr_t uIOCtl, unsigned idCpu, PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession)
908	{
909	/*
910	* We check the two prereqs after doing this only to allow the compiler to optimize things better.
911	*/
912	if (RT_LIKELY(pSession->pVM && pDevExt->pfnVMMR0EntryFast))
913	{
914	switch (uIOCtl)
915	{
916	case SUP_IOCTL_FAST_DO_RAW_RUN:
917	#ifdef RT_WITH_W64_UNWIND_HACK
918	supdrvNtWrapVMMR0EntryFast((PFNRT)pDevExt->pfnVMMR0EntryFast, pSession->pVM, idCpu, SUP_VMMR0_DO_RAW_RUN);
919	#else
920	pDevExt->pfnVMMR0EntryFast(pSession->pVM, idCpu, SUP_VMMR0_DO_RAW_RUN);
921	#endif
922	break;
923	case SUP_IOCTL_FAST_DO_HWACC_RUN:
924	#ifdef RT_WITH_W64_UNWIND_HACK
925	supdrvNtWrapVMMR0EntryFast((PFNRT)pDevExt->pfnVMMR0EntryFast, pSession->pVM, idCpu, SUP_VMMR0_DO_HWACC_RUN);
926	#else
927	pDevExt->pfnVMMR0EntryFast(pSession->pVM, idCpu, SUP_VMMR0_DO_HWACC_RUN);
928	#endif
929	break;
930	case SUP_IOCTL_FAST_DO_NOP:
931	#ifdef RT_WITH_W64_UNWIND_HACK
932	supdrvNtWrapVMMR0EntryFast((PFNRT)pDevExt->pfnVMMR0EntryFast, pSession->pVM, idCpu, SUP_VMMR0_DO_NOP);
933	#else
934	pDevExt->pfnVMMR0EntryFast(pSession->pVM, idCpu, SUP_VMMR0_DO_NOP);
935	#endif
936	break;
937	default:
938	return VERR_INTERNAL_ERROR;
939	}
940	return VINF_SUCCESS;
941	}
942	return VERR_INTERNAL_ERROR;
943	}
944
945
946	/**
947	* Helper for supdrvIOCtl. Check if pszStr contains any character of pszChars.
948	* We would use strpbrk here if this function would be contained in the RedHat kABI white
949	* list, see http://www.kerneldrivers.org/RHEL5.
950	*
951	* @return 1 if pszStr does contain any character of pszChars, 0 otherwise.
952	* @param pszStr String to check
953	* @param pszChars Character set
954	*/
955	static int supdrvCheckInvalidChar(const char pszStr, const char pszChars)
956	{
957	int chCur;
958	while ((chCur = *pszStr++) != '\0')
959	{
960	int ch;
961	const char *psz = pszChars;
962	while ((ch = *psz++) != '\0')
963	if (ch == chCur)
964	return 1;
965
966	}
967	return 0;
968	}
969
970
971	/**
972	* I/O Control worker.
973	*
974	* @returns 0 on success.
975	* @returns VERR_INVALID_PARAMETER if the request is invalid.
976	*
977	* @param uIOCtl Function number.
978	* @param pDevExt Device extention.
979	* @param pSession Session data.
980	* @param pReqHdr The request header.
981	*/
982	int VBOXCALL supdrvIOCtl(uintptr_t uIOCtl, PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPREQHDR pReqHdr)
983	{
984	/*
985	* Validate the request.
986	*/
987	/* this first check could probably be omitted as its also done by the OS specific code... */
988	if (RT_UNLIKELY( (pReqHdr->fFlags & SUPREQHDR_FLAGS_MAGIC_MASK) != SUPREQHDR_FLAGS_MAGIC
989	\|\| pReqHdr->cbIn < sizeof(*pReqHdr)
990	\|\| pReqHdr->cbOut < sizeof(*pReqHdr)))
991	{
992	OSDBGPRINT(("vboxdrv: Bad ioctl request header; cbIn=%#lx cbOut=%#lx fFlags=%#lx\n",
993	(long)pReqHdr->cbIn, (long)pReqHdr->cbOut, (long)pReqHdr->fFlags));
994	return VERR_INVALID_PARAMETER;
995	}
996	if (RT_UNLIKELY(uIOCtl == SUP_IOCTL_COOKIE))
997	{
998	if (pReqHdr->u32Cookie != SUPCOOKIE_INITIAL_COOKIE)
999	{
1000	OSDBGPRINT(("SUP_IOCTL_COOKIE: bad cookie %#lx\n", (long)pReqHdr->u32Cookie));
1001	return VERR_INVALID_PARAMETER;
1002	}
1003	}
1004	else if (RT_UNLIKELY( pReqHdr->u32Cookie != pDevExt->u32Cookie
1005	\|\| pReqHdr->u32SessionCookie != pSession->u32Cookie))
1006	{
1007	OSDBGPRINT(("vboxdrv: bad cookie %#lx / %#lx.\n", (long)pReqHdr->u32Cookie, (long)pReqHdr->u32SessionCookie));
1008	return VERR_INVALID_PARAMETER;
1009	}
1010
1011	/*
1012	* Validation macros
1013	*/
1014	#define REQ_CHECK_SIZES_EX(Name, cbInExpect, cbOutExpect) \
1015	do { \
1016	if (RT_UNLIKELY(pReqHdr->cbIn != (cbInExpect) \|\| pReqHdr->cbOut != (cbOutExpect))) \
1017	{ \
1018	OSDBGPRINT(( #Name ": Invalid input/output sizes. cbIn=%ld expected %ld. cbOut=%ld expected %ld.\n", \
1019	(long)pReq->Hdr.cbIn, (long)(cbInExpect), (long)pReq->Hdr.cbOut, (long)(cbOutExpect))); \
1020	return pReq->Hdr.rc = VERR_INVALID_PARAMETER; \
1021	} \
1022	} while (0)
1023
1024	#define REQ_CHECK_SIZES(Name) REQ_CHECK_SIZES_EX(Name, Name ## _SIZE_IN, Name ## _SIZE_OUT)
1025
1026	#define REQ_CHECK_SIZE_IN(Name, cbInExpect) \
1027	do { \
1028	if (RT_UNLIKELY(pReqHdr->cbIn != (cbInExpect))) \
1029	{ \
1030	OSDBGPRINT(( #Name ": Invalid input/output sizes. cbIn=%ld expected %ld.\n", \
1031	(long)pReq->Hdr.cbIn, (long)(cbInExpect))); \
1032	return pReq->Hdr.rc = VERR_INVALID_PARAMETER; \
1033	} \
1034	} while (0)
1035
1036	#define REQ_CHECK_SIZE_OUT(Name, cbOutExpect) \
1037	do { \
1038	if (RT_UNLIKELY(pReqHdr->cbOut != (cbOutExpect))) \
1039	{ \
1040	OSDBGPRINT(( #Name ": Invalid input/output sizes. cbOut=%ld expected %ld.\n", \
1041	(long)pReq->Hdr.cbOut, (long)(cbOutExpect))); \
1042	return pReq->Hdr.rc = VERR_INVALID_PARAMETER; \
1043	} \
1044	} while (0)
1045
1046	#define REQ_CHECK_EXPR(Name, expr) \
1047	do { \
1048	if (RT_UNLIKELY(!(expr))) \
1049	{ \
1050	OSDBGPRINT(( #Name ": %s\n", #expr)); \
1051	return pReq->Hdr.rc = VERR_INVALID_PARAMETER; \
1052	} \
1053	} while (0)
1054
1055	#define REQ_CHECK_EXPR_FMT(expr, fmt) \
1056	do { \
1057	if (RT_UNLIKELY(!(expr))) \
1058	{ \
1059	OSDBGPRINT( fmt ); \
1060	return pReq->Hdr.rc = VERR_INVALID_PARAMETER; \
1061	} \
1062	} while (0)
1063
1064
1065	/*
1066	* The switch.
1067	*/
1068	switch (SUP_CTL_CODE_NO_SIZE(uIOCtl))
1069	{
1070	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_COOKIE):
1071	{
1072	PSUPCOOKIE pReq = (PSUPCOOKIE)pReqHdr;
1073	REQ_CHECK_SIZES(SUP_IOCTL_COOKIE);
1074	if (strncmp(pReq->u.In.szMagic, SUPCOOKIE_MAGIC, sizeof(pReq->u.In.szMagic)))
1075	{
1076	OSDBGPRINT(("SUP_IOCTL_COOKIE: invalid magic %.16s\n", pReq->u.In.szMagic));
1077	pReq->Hdr.rc = VERR_INVALID_MAGIC;
1078	return 0;
1079	}
1080
1081	#if 0
1082	/*
1083	* Call out to the OS specific code and let it do permission checks on the
1084	* client process.
1085	*/
1086	if (!supdrvOSValidateClientProcess(pDevExt, pSession))
1087	{
1088	pReq->u.Out.u32Cookie = 0xffffffff;
1089	pReq->u.Out.u32SessionCookie = 0xffffffff;
1090	pReq->u.Out.u32SessionVersion = 0xffffffff;
1091	pReq->u.Out.u32DriverVersion = SUPDRV_IOC_VERSION;
1092	pReq->u.Out.pSession = NULL;
1093	pReq->u.Out.cFunctions = 0;
1094	pReq->Hdr.rc = VERR_PERMISSION_DENIED;
1095	return 0;
1096	}
1097	#endif
1098
1099	/*
1100	* Match the version.
1101	* The current logic is very simple, match the major interface version.
1102	*/
1103	if ( pReq->u.In.u32MinVersion > SUPDRV_IOC_VERSION
1104	\|\| (pReq->u.In.u32MinVersion & 0xffff0000) != (SUPDRV_IOC_VERSION & 0xffff0000))
1105	{
1106	OSDBGPRINT(("SUP_IOCTL_COOKIE: Version mismatch. Requested: %#x Min: %#x Current: %#x\n",
1107	pReq->u.In.u32ReqVersion, pReq->u.In.u32MinVersion, SUPDRV_IOC_VERSION));
1108	pReq->u.Out.u32Cookie = 0xffffffff;
1109	pReq->u.Out.u32SessionCookie = 0xffffffff;
1110	pReq->u.Out.u32SessionVersion = 0xffffffff;
1111	pReq->u.Out.u32DriverVersion = SUPDRV_IOC_VERSION;
1112	pReq->u.Out.pSession = NULL;
1113	pReq->u.Out.cFunctions = 0;
1114	pReq->Hdr.rc = VERR_VERSION_MISMATCH;
1115	return 0;
1116	}
1117
1118	/*
1119	* Fill in return data and be gone.
1120	* N.B. The first one to change SUPDRV_IOC_VERSION shall makes sure that
1121	* u32SessionVersion <= u32ReqVersion!
1122	*/
1123	/** @todo Somehow validate the client and negotiate a secure cookie... */
1124	pReq->u.Out.u32Cookie = pDevExt->u32Cookie;
1125	pReq->u.Out.u32SessionCookie = pSession->u32Cookie;
1126	pReq->u.Out.u32SessionVersion = SUPDRV_IOC_VERSION;
1127	pReq->u.Out.u32DriverVersion = SUPDRV_IOC_VERSION;
1128	pReq->u.Out.pSession = pSession;
1129	pReq->u.Out.cFunctions = sizeof(g_aFunctions) / sizeof(g_aFunctions[0]);
1130	pReq->Hdr.rc = VINF_SUCCESS;
1131	return 0;
1132	}
1133
1134	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_QUERY_FUNCS(0)):
1135	{
1136	/* validate */
1137	PSUPQUERYFUNCS pReq = (PSUPQUERYFUNCS)pReqHdr;
1138	REQ_CHECK_SIZES_EX(SUP_IOCTL_QUERY_FUNCS, SUP_IOCTL_QUERY_FUNCS_SIZE_IN, SUP_IOCTL_QUERY_FUNCS_SIZE_OUT(RT_ELEMENTS(g_aFunctions)));
1139
1140	/* execute */
1141	pReq->u.Out.cFunctions = RT_ELEMENTS(g_aFunctions);
1142	memcpy(&pReq->u.Out.aFunctions[0], g_aFunctions, sizeof(g_aFunctions));
1143	pReq->Hdr.rc = VINF_SUCCESS;
1144	return 0;
1145	}
1146
1147	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_IDT_INSTALL):
1148	{
1149	/* validate */
1150	PSUPIDTINSTALL pReq = (PSUPIDTINSTALL)pReqHdr;
1151	REQ_CHECK_SIZES(SUP_IOCTL_IDT_INSTALL);
1152
1153	/* execute */
1154	pReq->u.Out.u8Idt = 3;
1155	pReq->Hdr.rc = VERR_NOT_SUPPORTED;
1156	return 0;
1157	}
1158
1159	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_IDT_REMOVE):
1160	{
1161	/* validate */
1162	PSUPIDTREMOVE pReq = (PSUPIDTREMOVE)pReqHdr;
1163	REQ_CHECK_SIZES(SUP_IOCTL_IDT_REMOVE);
1164
1165	/* execute */
1166	pReq->Hdr.rc = VERR_NOT_SUPPORTED;
1167	return 0;
1168	}
1169
1170	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_LOCK):
1171	{
1172	/* validate */
1173	PSUPPAGELOCK pReq = (PSUPPAGELOCK)pReqHdr;
1174	REQ_CHECK_SIZE_IN(SUP_IOCTL_PAGE_LOCK, SUP_IOCTL_PAGE_LOCK_SIZE_IN);
1175	REQ_CHECK_SIZE_OUT(SUP_IOCTL_PAGE_LOCK, SUP_IOCTL_PAGE_LOCK_SIZE_OUT(pReq->u.In.cPages));
1176	REQ_CHECK_EXPR(SUP_IOCTL_PAGE_LOCK, pReq->u.In.cPages > 0);
1177	REQ_CHECK_EXPR(SUP_IOCTL_PAGE_LOCK, pReq->u.In.pvR3 >= PAGE_SIZE);
1178
1179	/* execute */
1180	pReq->Hdr.rc = SUPR0LockMem(pSession, pReq->u.In.pvR3, pReq->u.In.cPages, &pReq->u.Out.aPages[0]);
1181	if (RT_FAILURE(pReq->Hdr.rc))
1182	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1183	return 0;
1184	}
1185
1186	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_UNLOCK):
1187	{
1188	/* validate */
1189	PSUPPAGEUNLOCK pReq = (PSUPPAGEUNLOCK)pReqHdr;
1190	REQ_CHECK_SIZES(SUP_IOCTL_PAGE_UNLOCK);
1191
1192	/* execute */
1193	pReq->Hdr.rc = SUPR0UnlockMem(pSession, pReq->u.In.pvR3);
1194	return 0;
1195	}
1196
1197	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_CONT_ALLOC):
1198	{
1199	/* validate */
1200	PSUPCONTALLOC pReq = (PSUPCONTALLOC)pReqHdr;
1201	REQ_CHECK_SIZES(SUP_IOCTL_CONT_ALLOC);
1202
1203	/* execute */
1204	pReq->Hdr.rc = SUPR0ContAlloc(pSession, pReq->u.In.cPages, &pReq->u.Out.pvR0, &pReq->u.Out.pvR3, &pReq->u.Out.HCPhys);
1205	if (RT_FAILURE(pReq->Hdr.rc))
1206	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1207	return 0;
1208	}
1209
1210	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_CONT_FREE):
1211	{
1212	/* validate */
1213	PSUPCONTFREE pReq = (PSUPCONTFREE)pReqHdr;
1214	REQ_CHECK_SIZES(SUP_IOCTL_CONT_FREE);
1215
1216	/* execute */
1217	pReq->Hdr.rc = SUPR0ContFree(pSession, (RTHCUINTPTR)pReq->u.In.pvR3);
1218	return 0;
1219	}
1220
1221	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LDR_OPEN):
1222	{
1223	/* validate */
1224	PSUPLDROPEN pReq = (PSUPLDROPEN)pReqHdr;
1225	REQ_CHECK_SIZES(SUP_IOCTL_LDR_OPEN);
1226	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.cbImage > 0);
1227	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.cbImage < _1M*16);
1228	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.szName[0]);
1229	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, memchr(pReq->u.In.szName, '\0', sizeof(pReq->u.In.szName)));
1230	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, !supdrvCheckInvalidChar(pReq->u.In.szName, ";:()[]{}/\\\|&*%#@!~`\"'"));
1231
1232	/* execute */
1233	pReq->Hdr.rc = supdrvIOCtl_LdrOpen(pDevExt, pSession, pReq);
1234	return 0;
1235	}
1236
1237	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LDR_LOAD):
1238	{
1239	/* validate */
1240	PSUPLDRLOAD pReq = (PSUPLDRLOAD)pReqHdr;
1241	REQ_CHECK_EXPR(Name, pReq->Hdr.cbIn >= sizeof(*pReq));
1242	REQ_CHECK_SIZES_EX(SUP_IOCTL_LDR_LOAD, SUP_IOCTL_LDR_LOAD_SIZE_IN(pReq->u.In.cbImage), SUP_IOCTL_LDR_LOAD_SIZE_OUT);
1243	REQ_CHECK_EXPR(SUP_IOCTL_LDR_LOAD, pReq->u.In.cSymbols <= 16384);
1244	REQ_CHECK_EXPR_FMT( !pReq->u.In.cSymbols
1245	\|\| ( pReq->u.In.offSymbols < pReq->u.In.cbImage
1246	&& pReq->u.In.offSymbols + pReq->u.In.cSymbols * sizeof(SUPLDRSYM) <= pReq->u.In.cbImage),
1247	("SUP_IOCTL_LDR_LOAD: offSymbols=%#lx cSymbols=%#lx cbImage=%#lx\n", (long)pReq->u.In.offSymbols,
1248	(long)pReq->u.In.cSymbols, (long)pReq->u.In.cbImage));
1249	REQ_CHECK_EXPR_FMT( !pReq->u.In.cbStrTab
1250	\|\| ( pReq->u.In.offStrTab < pReq->u.In.cbImage
1251	&& pReq->u.In.offStrTab + pReq->u.In.cbStrTab <= pReq->u.In.cbImage
1252	&& pReq->u.In.cbStrTab <= pReq->u.In.cbImage),
1253	("SUP_IOCTL_LDR_LOAD: offStrTab=%#lx cbStrTab=%#lx cbImage=%#lx\n", (long)pReq->u.In.offStrTab,
1254	(long)pReq->u.In.cbStrTab, (long)pReq->u.In.cbImage));
1255
1256	if (pReq->u.In.cSymbols)
1257	{
1258	uint32_t i;
1259	PSUPLDRSYM paSyms = (PSUPLDRSYM)&pReq->u.In.achImage[pReq->u.In.offSymbols];
1260	for (i = 0; i < pReq->u.In.cSymbols; i++)
1261	{
1262	REQ_CHECK_EXPR_FMT(paSyms[i].offSymbol < pReq->u.In.cbImage,
1263	("SUP_IOCTL_LDR_LOAD: sym #%ld: symb off %#lx (max=%#lx)\n", (long)i, (long)paSyms[i].offSymbol, (long)pReq->u.In.cbImage));
1264	REQ_CHECK_EXPR_FMT(paSyms[i].offName < pReq->u.In.cbStrTab,
1265	("SUP_IOCTL_LDR_LOAD: sym #%ld: name off %#lx (max=%#lx)\n", (long)i, (long)paSyms[i].offName, (long)pReq->u.In.cbImage));
1266	REQ_CHECK_EXPR_FMT(memchr(&pReq->u.In.achImage[pReq->u.In.offStrTab + paSyms[i].offName], '\0', pReq->u.In.cbStrTab - paSyms[i].offName),
1267	("SUP_IOCTL_LDR_LOAD: sym #%ld: unterminated name! (%#lx / %#lx)\n", (long)i, (long)paSyms[i].offName, (long)pReq->u.In.cbImage));
1268	}
1269	}
1270
1271	/* execute */
1272	pReq->Hdr.rc = supdrvIOCtl_LdrLoad(pDevExt, pSession, pReq);
1273	return 0;
1274	}
1275
1276	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LDR_FREE):
1277	{
1278	/* validate */
1279	PSUPLDRFREE pReq = (PSUPLDRFREE)pReqHdr;
1280	REQ_CHECK_SIZES(SUP_IOCTL_LDR_FREE);
1281
1282	/* execute */
1283	pReq->Hdr.rc = supdrvIOCtl_LdrFree(pDevExt, pSession, pReq);
1284	return 0;
1285	}
1286
1287	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LDR_GET_SYMBOL):
1288	{
1289	/* validate */
1290	PSUPLDRGETSYMBOL pReq = (PSUPLDRGETSYMBOL)pReqHdr;
1291	REQ_CHECK_SIZES(SUP_IOCTL_LDR_GET_SYMBOL);
1292	REQ_CHECK_EXPR(SUP_IOCTL_LDR_GET_SYMBOL, memchr(pReq->u.In.szSymbol, '\0', sizeof(pReq->u.In.szSymbol)));
1293
1294	/* execute */
1295	pReq->Hdr.rc = supdrvIOCtl_LdrGetSymbol(pDevExt, pSession, pReq);
1296	return 0;
1297	}
1298
1299	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_CALL_VMMR0(0)):
1300	{
1301	/* validate */
1302	PSUPCALLVMMR0 pReq = (PSUPCALLVMMR0)pReqHdr;
1303	Log4(("SUP_IOCTL_CALL_VMMR0: op=%u in=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1304	pReq->u.In.uOperation, pReq->Hdr.cbIn, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1305
1306	if (pReq->Hdr.cbIn == SUP_IOCTL_CALL_VMMR0_SIZE(0))
1307	{
1308	REQ_CHECK_SIZES_EX(SUP_IOCTL_CALL_VMMR0, SUP_IOCTL_CALL_VMMR0_SIZE_IN(0), SUP_IOCTL_CALL_VMMR0_SIZE_OUT(0));
1309
1310	/* execute */
1311	if (RT_LIKELY(pDevExt->pfnVMMR0EntryEx))
1312	#ifdef RT_WITH_W64_UNWIND_HACK
1313	pReq->Hdr.rc = supdrvNtWrapVMMR0EntryEx((PFNRT)pDevExt->pfnVMMR0EntryEx, pReq->u.In.pVMR0, pReq->u.In.uOperation, NULL, pReq->u.In.u64Arg, pSession);
1314	#else
1315	pReq->Hdr.rc = pDevExt->pfnVMMR0EntryEx(pReq->u.In.pVMR0, pReq->u.In.uOperation, NULL, pReq->u.In.u64Arg, pSession);
1316	#endif
1317	else
1318	pReq->Hdr.rc = VERR_WRONG_ORDER;
1319	}
1320	else
1321	{
1322	PSUPVMMR0REQHDR pVMMReq = (PSUPVMMR0REQHDR)&pReq->abReqPkt[0];
1323	REQ_CHECK_EXPR_FMT(pReq->Hdr.cbIn >= SUP_IOCTL_CALL_VMMR0_SIZE(sizeof(SUPVMMR0REQHDR)),
1324	("SUP_IOCTL_CALL_VMMR0: cbIn=%#x < %#lx\n", pReq->Hdr.cbIn, SUP_IOCTL_CALL_VMMR0_SIZE(sizeof(SUPVMMR0REQHDR))));
1325	REQ_CHECK_EXPR(SUP_IOCTL_CALL_VMMR0, pVMMReq->u32Magic == SUPVMMR0REQHDR_MAGIC);
1326	REQ_CHECK_SIZES_EX(SUP_IOCTL_CALL_VMMR0, SUP_IOCTL_CALL_VMMR0_SIZE_IN(pVMMReq->cbReq), SUP_IOCTL_CALL_VMMR0_SIZE_OUT(pVMMReq->cbReq));
1327
1328	/* execute */
1329	if (RT_LIKELY(pDevExt->pfnVMMR0EntryEx))
1330	#ifdef RT_WITH_W64_UNWIND_HACK
1331	pReq->Hdr.rc = supdrvNtWrapVMMR0EntryEx((PFNRT)pDevExt->pfnVMMR0EntryEx, pReq->u.In.pVMR0, pReq->u.In.uOperation, pVMMReq, pReq->u.In.u64Arg, pSession);
1332	#else
1333	pReq->Hdr.rc = pDevExt->pfnVMMR0EntryEx(pReq->u.In.pVMR0, pReq->u.In.uOperation, pVMMReq, pReq->u.In.u64Arg, pSession);
1334	#endif
1335	else
1336	pReq->Hdr.rc = VERR_WRONG_ORDER;
1337	}
1338
1339	if ( RT_FAILURE(pReq->Hdr.rc)
1340	&& pReq->Hdr.rc != VERR_INTERRUPTED
1341	&& pReq->Hdr.rc != VERR_TIMEOUT)
1342	Log(("SUP_IOCTL_CALL_VMMR0: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1343	pReq->Hdr.rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1344	else
1345	Log4(("SUP_IOCTL_CALL_VMMR0: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1346	pReq->Hdr.rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1347	return 0;
1348	}
1349
1350	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_GET_PAGING_MODE):
1351	{
1352	/* validate */
1353	PSUPGETPAGINGMODE pReq = (PSUPGETPAGINGMODE)pReqHdr;
1354	REQ_CHECK_SIZES(SUP_IOCTL_GET_PAGING_MODE);
1355
1356	/* execute */
1357	pReq->Hdr.rc = VINF_SUCCESS;
1358	pReq->u.Out.enmMode = SUPR0GetPagingMode();
1359	return 0;
1360	}
1361
1362	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LOW_ALLOC):
1363	{
1364	/* validate */
1365	PSUPLOWALLOC pReq = (PSUPLOWALLOC)pReqHdr;
1366	REQ_CHECK_EXPR(SUP_IOCTL_LOW_ALLOC, pReq->Hdr.cbIn <= SUP_IOCTL_LOW_ALLOC_SIZE_IN);
1367	REQ_CHECK_SIZES_EX(SUP_IOCTL_LOW_ALLOC, SUP_IOCTL_LOW_ALLOC_SIZE_IN, SUP_IOCTL_LOW_ALLOC_SIZE_OUT(pReq->u.In.cPages));
1368
1369	/* execute */
1370	pReq->Hdr.rc = SUPR0LowAlloc(pSession, pReq->u.In.cPages, &pReq->u.Out.pvR0, &pReq->u.Out.pvR3, &pReq->u.Out.aPages[0]);
1371	if (RT_FAILURE(pReq->Hdr.rc))
1372	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1373	return 0;
1374	}
1375
1376	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LOW_FREE):
1377	{
1378	/* validate */
1379	PSUPLOWFREE pReq = (PSUPLOWFREE)pReqHdr;
1380	REQ_CHECK_SIZES(SUP_IOCTL_LOW_FREE);
1381
1382	/* execute */
1383	pReq->Hdr.rc = SUPR0LowFree(pSession, (RTHCUINTPTR)pReq->u.In.pvR3);
1384	return 0;
1385	}
1386
1387	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_GIP_MAP):
1388	{
1389	/* validate */
1390	PSUPGIPMAP pReq = (PSUPGIPMAP)pReqHdr;
1391	REQ_CHECK_SIZES(SUP_IOCTL_GIP_MAP);
1392
1393	/* execute */
1394	pReq->Hdr.rc = SUPR0GipMap(pSession, &pReq->u.Out.pGipR3, &pReq->u.Out.HCPhysGip);
1395	if (RT_SUCCESS(pReq->Hdr.rc))
1396	pReq->u.Out.pGipR0 = pDevExt->pGip;
1397	return 0;
1398	}
1399
1400	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_GIP_UNMAP):
1401	{
1402	/* validate */
1403	PSUPGIPUNMAP pReq = (PSUPGIPUNMAP)pReqHdr;
1404	REQ_CHECK_SIZES(SUP_IOCTL_GIP_UNMAP);
1405
1406	/* execute */
1407	pReq->Hdr.rc = SUPR0GipUnmap(pSession);
1408	return 0;
1409	}
1410
1411	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_SET_VM_FOR_FAST):
1412	{
1413	/* validate */
1414	PSUPSETVMFORFAST pReq = (PSUPSETVMFORFAST)pReqHdr;
1415	REQ_CHECK_SIZES(SUP_IOCTL_SET_VM_FOR_FAST);
1416	REQ_CHECK_EXPR_FMT( !pReq->u.In.pVMR0
1417	\|\| ( VALID_PTR(pReq->u.In.pVMR0)
1418	&& !((uintptr_t)pReq->u.In.pVMR0 & (PAGE_SIZE - 1))),
1419	("SUP_IOCTL_SET_VM_FOR_FAST: pVMR0=%p!\n", pReq->u.In.pVMR0));
1420	/* execute */
1421	pSession->pVM = pReq->u.In.pVMR0;
1422	pReq->Hdr.rc = VINF_SUCCESS;
1423	return 0;
1424	}
1425
1426	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_ALLOC):
1427	{
1428	/* validate */
1429	PSUPPAGEALLOC pReq = (PSUPPAGEALLOC)pReqHdr;
1430	REQ_CHECK_EXPR(SUP_IOCTL_PAGE_ALLOC, pReq->Hdr.cbIn <= SUP_IOCTL_PAGE_ALLOC_SIZE_IN);
1431	REQ_CHECK_SIZES_EX(SUP_IOCTL_PAGE_ALLOC, SUP_IOCTL_PAGE_ALLOC_SIZE_IN, SUP_IOCTL_PAGE_ALLOC_SIZE_OUT(pReq->u.In.cPages));
1432
1433	/* execute */
1434	pReq->Hdr.rc = SUPR0PageAlloc(pSession, pReq->u.In.cPages, &pReq->u.Out.pvR3, &pReq->u.Out.aPages[0]);
1435	if (RT_FAILURE(pReq->Hdr.rc))
1436	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1437	return 0;
1438	}
1439
1440	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_ALLOC_EX):
1441	{
1442	/* validate */
1443	PSUPPAGEALLOCEX pReq = (PSUPPAGEALLOCEX)pReqHdr;
1444	REQ_CHECK_EXPR(SUP_IOCTL_PAGE_ALLOC_EX, pReq->Hdr.cbIn <= SUP_IOCTL_PAGE_ALLOC_EX_SIZE_IN);
1445	REQ_CHECK_SIZES_EX(SUP_IOCTL_PAGE_ALLOC_EX, SUP_IOCTL_PAGE_ALLOC_EX_SIZE_IN, SUP_IOCTL_PAGE_ALLOC_EX_SIZE_OUT(pReq->u.In.cPages));
1446	REQ_CHECK_EXPR_FMT(pReq->u.In.fKernelMapping \|\| pReq->u.In.fUserMapping,
1447	("SUP_IOCTL_PAGE_ALLOC_EX: No mapping requested!\n"));
1448	REQ_CHECK_EXPR_FMT(pReq->u.In.fUserMapping,
1449	("SUP_IOCTL_PAGE_ALLOC_EX: Must have user mapping!\n"));
1450	REQ_CHECK_EXPR_FMT(!pReq->u.In.fReserved0 && !pReq->u.In.fReserved1,
1451	("SUP_IOCTL_PAGE_ALLOC_EX: fReserved0=%d fReserved1=%d\n", pReq->u.In.fReserved0, pReq->u.In.fReserved1));
1452
1453	/* execute */
1454	pReq->Hdr.rc = SUPR0PageAllocEx(pSession, pReq->u.In.cPages, 0 /* fFlags */,
1455	pReq->u.In.fUserMapping ? &pReq->u.Out.pvR3 : NULL,
1456	pReq->u.In.fKernelMapping ? &pReq->u.Out.pvR0 : NULL,
1457	&pReq->u.Out.aPages[0]);
1458	if (RT_FAILURE(pReq->Hdr.rc))
1459	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1460	return 0;
1461	}
1462
1463	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_MAP_KERNEL):
1464	{
1465	/* validate */
1466	PSUPPAGEMAPKERNEL pReq = (PSUPPAGEMAPKERNEL)pReqHdr;
1467	REQ_CHECK_SIZES(SUP_IOCTL_PAGE_MAP_KERNEL);
1468	REQ_CHECK_EXPR_FMT(!pReq->u.In.fFlags, ("SUP_IOCTL_PAGE_MAP_KERNEL: fFlags=%#x! MBZ\n", pReq->u.In.fFlags));
1469	REQ_CHECK_EXPR_FMT(!(pReq->u.In.offSub & PAGE_OFFSET_MASK), ("SUP_IOCTL_PAGE_MAP_KERNEL: offSub=%#x\n", pReq->u.In.offSub));
1470	REQ_CHECK_EXPR_FMT(pReq->u.In.cbSub && !(pReq->u.In.cbSub & PAGE_OFFSET_MASK),
1471	("SUP_IOCTL_PAGE_MAP_KERNEL: cbSub=%#x\n", pReq->u.In.cbSub));
1472
1473	/* execute */
1474	pReq->Hdr.rc = SUPR0PageMapKernel(pSession, pReq->u.In.pvR3, pReq->u.In.offSub, pReq->u.In.cbSub,
1475	pReq->u.In.fFlags, &pReq->u.Out.pvR0);
1476	if (RT_FAILURE(pReq->Hdr.rc))
1477	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1478	return 0;
1479	}
1480
1481	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_FREE):
1482	{
1483	/* validate */
1484	PSUPPAGEFREE pReq = (PSUPPAGEFREE)pReqHdr;
1485	REQ_CHECK_SIZES(SUP_IOCTL_PAGE_FREE);
1486
1487	/* execute */
1488	pReq->Hdr.rc = SUPR0PageFree(pSession, pReq->u.In.pvR3);
1489	return 0;
1490	}
1491
1492	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_CALL_SERVICE(0)):
1493	{
1494	/* validate */
1495	PSUPCALLSERVICE pReq = (PSUPCALLSERVICE)pReqHdr;
1496	Log4(("SUP_IOCTL_CALL_SERVICE: op=%u in=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1497	pReq->u.In.uOperation, pReq->Hdr.cbIn, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1498
1499	if (pReq->Hdr.cbIn == SUP_IOCTL_CALL_SERVICE_SIZE(0))
1500	REQ_CHECK_SIZES_EX(SUP_IOCTL_CALL_SERVICE, SUP_IOCTL_CALL_SERVICE_SIZE_IN(0), SUP_IOCTL_CALL_SERVICE_SIZE_OUT(0));
1501	else
1502	{
1503	PSUPR0SERVICEREQHDR pSrvReq = (PSUPR0SERVICEREQHDR)&pReq->abReqPkt[0];
1504	REQ_CHECK_EXPR_FMT(pReq->Hdr.cbIn >= SUP_IOCTL_CALL_SERVICE_SIZE(sizeof(SUPR0SERVICEREQHDR)),
1505	("SUP_IOCTL_CALL_SERVICE: cbIn=%#x < %#lx\n", pReq->Hdr.cbIn, SUP_IOCTL_CALL_SERVICE_SIZE(sizeof(SUPR0SERVICEREQHDR))));
1506	REQ_CHECK_EXPR(SUP_IOCTL_CALL_SERVICE, pSrvReq->u32Magic == SUPR0SERVICEREQHDR_MAGIC);
1507	REQ_CHECK_SIZES_EX(SUP_IOCTL_CALL_SERVICE, SUP_IOCTL_CALL_SERVICE_SIZE_IN(pSrvReq->cbReq), SUP_IOCTL_CALL_SERVICE_SIZE_OUT(pSrvReq->cbReq));
1508	}
1509	REQ_CHECK_EXPR(SUP_IOCTL_CALL_SERVICE, memchr(pReq->u.In.szName, '\0', sizeof(pReq->u.In.szName)));
1510
1511	/* execute */
1512	pReq->Hdr.rc = supdrvIOCtl_CallServiceModule(pDevExt, pSession, pReq);
1513	return 0;
1514	}
1515
1516	default:
1517	Log(("Unknown IOCTL %#lx\n", (long)uIOCtl));
1518	break;
1519	}
1520	return SUPDRV_ERR_GENERAL_FAILURE;
1521	}
1522
1523
1524	/**
1525	* Inter-Driver Communcation (IDC) worker.
1526	*
1527	* @returns VBox status code.
1528	* @retval VINF_SUCCESS on success.
1529	* @retval VERR_INVALID_PARAMETER if the request is invalid.
1530	* @retval VERR_NOT_SUPPORTED if the request isn't supported.
1531	*
1532	* @param uReq The request (function) code.
1533	* @param pDevExt Device extention.
1534	* @param pSession Session data.
1535	* @param pReqHdr The request header.
1536	*/
1537	int VBOXCALL supdrvIDC(uintptr_t uReq, PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPDRVIDCREQHDR pReqHdr)
1538	{
1539	/*
1540	* The OS specific code has already validated the pSession
1541	* pointer, and the request size being greater or equal to
1542	* size of the header.
1543	*
1544	* So, just check that pSession is a kernel context session.
1545	*/
1546	if (RT_UNLIKELY( pSession
1547	&& pSession->R0Process != NIL_RTR0PROCESS))
1548	return VERR_INVALID_PARAMETER;
1549
1550	/*
1551	* Validation macro.
1552	*/
1553	#define REQ_CHECK_IDC_SIZE(Name, cbExpect) \
1554	do { \
1555	if (RT_UNLIKELY(pReqHdr->cb != (cbExpect))) \
1556	{ \
1557	OSDBGPRINT(( #Name ": Invalid input/output sizes. cb=%ld expected %ld.\n", \
1558	(long)pReqHdr->cb, (long)(cbExpect))); \
1559	return pReqHdr->rc = VERR_INVALID_PARAMETER; \
1560	} \
1561	} while (0)
1562
1563	switch (uReq)
1564	{
1565	case SUPDRV_IDC_REQ_CONNECT:
1566	{
1567	PSUPDRVIDCREQCONNECT pReq = (PSUPDRVIDCREQCONNECT)pReqHdr;
1568	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_CONNECT, sizeof(*pReq));
1569
1570	/*
1571	* Validate the cookie and other input.
1572	*/
1573	if (pReq->Hdr.pSession != NULL)
1574	{
1575	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: pSession=%p expected NULL!\n", pReq->Hdr.pSession));
1576	return pReqHdr->rc = VERR_INVALID_PARAMETER;
1577	}
1578	if (pReq->u.In.u32MagicCookie != SUPDRVIDCREQ_CONNECT_MAGIC_COOKIE)
1579	{
1580	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: u32MagicCookie=%#x expected %#x!\n",
1581	pReq->u.In.u32MagicCookie, SUPDRVIDCREQ_CONNECT_MAGIC_COOKIE));
1582	return pReqHdr->rc = VERR_INVALID_PARAMETER;
1583	}
1584	if ( pReq->u.In.uMinVersion > pReq->u.In.uReqVersion
1585	\|\| (pReq->u.In.uMinVersion & UINT32_C(0xffff0000)) != (pReq->u.In.uReqVersion & UINT32_C(0xffff0000)))
1586	{
1587	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: uMinVersion=%#x uMaxVersion=%#x doesn't match!\n",
1588	pReq->u.In.uMinVersion, pReq->u.In.uReqVersion));
1589	return pReqHdr->rc = VERR_INVALID_PARAMETER;
1590	}
1591
1592	/*
1593	* Match the version.
1594	* The current logic is very simple, match the major interface version.
1595	*/
1596	if ( pReq->u.In.uMinVersion > SUPDRV_IDC_VERSION
1597	\|\| (pReq->u.In.uMinVersion & 0xffff0000) != (SUPDRV_IDC_VERSION & 0xffff0000))
1598	{
1599	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: Version mismatch. Requested: %#x Min: %#x Current: %#x\n",
1600	pReq->u.In.uReqVersion, pReq->u.In.uMinVersion, SUPDRV_IDC_VERSION));
1601	pReq->u.Out.pSession = NULL;
1602	pReq->u.Out.uSessionVersion = 0xffffffff;
1603	pReq->u.Out.uDriverVersion = SUPDRV_IDC_VERSION;
1604	pReq->u.Out.uDriverRevision = VBOX_SVN_REV;
1605	pReq->Hdr.rc = VERR_VERSION_MISMATCH;
1606	return VINF_SUCCESS;
1607	}
1608
1609	pReq->u.Out.pSession = NULL;
1610	pReq->u.Out.uSessionVersion = SUPDRV_IDC_VERSION;
1611	pReq->u.Out.uDriverVersion = SUPDRV_IDC_VERSION;
1612	pReq->u.Out.uDriverRevision = VBOX_SVN_REV;
1613
1614	/*
1615	* On NT we will already have a session associated with the
1616	* client, just like with the SUP_IOCTL_COOKIE request, while
1617	* the other doesn't.
1618	*/
1619	#ifdef RT_OS_WINDOWS
1620	pReq->Hdr.rc = VINF_SUCCESS;
1621	#else
1622	AssertReturn(!pSession, VERR_INTERNAL_ERROR);
1623	pReq->Hdr.rc = supdrvCreateSession(pDevExt, false /* fUser */, &pSession);
1624	if (RT_FAILURE(pReq->Hdr.rc))
1625	{
1626	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: failed to create session, rc=%d\n", pReq->Hdr.rc));
1627	return VINF_SUCCESS;
1628	}
1629	#endif
1630
1631	pReq->u.Out.pSession = pSession;
1632	pReq->Hdr.pSession = pSession;
1633
1634	return VINF_SUCCESS;
1635	}
1636
1637	case SUPDRV_IDC_REQ_DISCONNECT:
1638	{
1639	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_DISCONNECT, sizeof(*pReqHdr));
1640
1641	#ifdef RT_OS_WINDOWS
1642	/* Windows will destroy the session when the file object is destroyed. */
1643	#else
1644	supdrvCloseSession(pDevExt, pSession);
1645	#endif
1646	return pReqHdr->rc = VINF_SUCCESS;
1647	}
1648
1649	case SUPDRV_IDC_REQ_GET_SYMBOL:
1650	{
1651	PSUPDRVIDCREQGETSYM pReq = (PSUPDRVIDCREQGETSYM)pReqHdr;
1652	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_GET_SYMBOL, sizeof(*pReq));
1653
1654	pReq->Hdr.rc = supdrvIDC_LdrGetSymbol(pDevExt, pSession, pReq);
1655	return VINF_SUCCESS;
1656	}
1657
1658	case SUPDRV_IDC_REQ_COMPONENT_REGISTER_FACTORY:
1659	{
1660	PSUPDRVIDCREQCOMPREGFACTORY pReq = (PSUPDRVIDCREQCOMPREGFACTORY)pReqHdr;
1661	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_COMPONENT_REGISTER_FACTORY, sizeof(*pReq));
1662
1663	pReq->Hdr.rc = SUPR0ComponentRegisterFactory(pSession, pReq->u.In.pFactory);
1664	return VINF_SUCCESS;
1665	}
1666
1667	case SUPDRV_IDC_REQ_COMPONENT_DEREGISTER_FACTORY:
1668	{
1669	PSUPDRVIDCREQCOMPDEREGFACTORY pReq = (PSUPDRVIDCREQCOMPDEREGFACTORY)pReqHdr;
1670	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_COMPONENT_DEREGISTER_FACTORY, sizeof(*pReq));
1671
1672	pReq->Hdr.rc = SUPR0ComponentDeregisterFactory(pSession, pReq->u.In.pFactory);
1673	return VINF_SUCCESS;
1674	}
1675
1676	default:
1677	Log(("Unknown IDC %#lx\n", (long)uReq));
1678	break;
1679	}
1680
1681	#undef REQ_CHECK_IDC_SIZE
1682	return VERR_NOT_SUPPORTED;
1683	}
1684
1685
1686	/**
1687	* Register a object for reference counting.
1688	* The object is registered with one reference in the specified session.
1689	*
1690	* @returns Unique identifier on success (pointer).
1691	* All future reference must use this identifier.
1692	* @returns NULL on failure.
1693	* @param pfnDestructor The destructore function which will be called when the reference count reaches 0.
1694	* @param pvUser1 The first user argument.
1695	* @param pvUser2 The second user argument.
1696	*/
1697	SUPR0DECL(void ) SUPR0ObjRegister(PSUPDRVSESSION pSession, SUPDRVOBJTYPE enmType, PFNSUPDRVDESTRUCTOR pfnDestructor, void pvUser1, void *pvUser2)
1698	{
1699	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
1700	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
1701	PSUPDRVOBJ pObj;
1702	PSUPDRVUSAGE pUsage;
1703
1704	/*
1705	* Validate the input.
1706	*/
1707	AssertReturn(SUP_IS_SESSION_VALID(pSession), NULL);
1708	AssertReturn(enmType > SUPDRVOBJTYPE_INVALID && enmType < SUPDRVOBJTYPE_END, NULL);
1709	AssertPtrReturn(pfnDestructor, NULL);
1710
1711	/*
1712	* Allocate and initialize the object.
1713	*/
1714	pObj = (PSUPDRVOBJ)RTMemAlloc(sizeof(*pObj));
1715	if (!pObj)
1716	return NULL;
1717	pObj->u32Magic = SUPDRVOBJ_MAGIC;
1718	pObj->enmType = enmType;
1719	pObj->pNext = NULL;
1720	pObj->cUsage = 1;
1721	pObj->pfnDestructor = pfnDestructor;
1722	pObj->pvUser1 = pvUser1;
1723	pObj->pvUser2 = pvUser2;
1724	pObj->CreatorUid = pSession->Uid;
1725	pObj->CreatorGid = pSession->Gid;
1726	pObj->CreatorProcess= pSession->Process;
1727	supdrvOSObjInitCreator(pObj, pSession);
1728
1729	/*
1730	* Allocate the usage record.
1731	* (We keep freed usage records around to simplify SUPR0ObjAddRefEx().)
1732	*/
1733	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
1734
1735	pUsage = pDevExt->pUsageFree;
1736	if (pUsage)
1737	pDevExt->pUsageFree = pUsage->pNext;
1738	else
1739	{
1740	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
1741	pUsage = (PSUPDRVUSAGE)RTMemAlloc(sizeof(*pUsage));
1742	if (!pUsage)
1743	{
1744	RTMemFree(pObj);
1745	return NULL;
1746	}
1747	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
1748	}
1749
1750	/*
1751	* Insert the object and create the session usage record.
1752	*/
1753	/* The object. */
1754	pObj->pNext = pDevExt->pObjs;
1755	pDevExt->pObjs = pObj;
1756
1757	/* The session record. */
1758	pUsage->cUsage = 1;
1759	pUsage->pObj = pObj;
1760	pUsage->pNext = pSession->pUsage;
1761	/* Log2(("SUPR0ObjRegister: pUsage=%p:{.pObj=%p, .pNext=%p}\n", pUsage, pUsage->pObj, pUsage->pNext)); */
1762	pSession->pUsage = pUsage;
1763
1764	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
1765
1766	Log(("SUPR0ObjRegister: returns %p (pvUser1=%p, pvUser=%p)\n", pObj, pvUser1, pvUser2));
1767	return pObj;
1768	}
1769
1770
1771	/**
1772	* Increment the reference counter for the object associating the reference
1773	* with the specified session.
1774	*
1775	* @returns IPRT status code.
1776	* @param pvObj The identifier returned by SUPR0ObjRegister().
1777	* @param pSession The session which is referencing the object.
1778	*
1779	* @remarks The caller should not own any spinlocks and must carefully protect
1780	* itself against potential race with the destructor so freed memory
1781	* isn't accessed here.
1782	*/
1783	SUPR0DECL(int) SUPR0ObjAddRef(void *pvObj, PSUPDRVSESSION pSession)
1784	{
1785	return SUPR0ObjAddRefEx(pvObj, pSession, false /* fNoBlocking */);
1786	}
1787
1788
1789	/**
1790	* Increment the reference counter for the object associating the reference
1791	* with the specified session.
1792	*
1793	* @returns IPRT status code.
1794	* @retval VERR_TRY_AGAIN if fNoBlocking was set and a new usage record
1795	* couldn't be allocated. (If you see this you're not doing the right
1796	* thing and it won't ever work reliably.)
1797	*
1798	* @param pvObj The identifier returned by SUPR0ObjRegister().
1799	* @param pSession The session which is referencing the object.
1800	* @param fNoBlocking Set if it's not OK to block. Never try to make the
1801	* first reference to an object in a session with this
1802	* argument set.
1803	*
1804	* @remarks The caller should not own any spinlocks and must carefully protect
1805	* itself against potential race with the destructor so freed memory
1806	* isn't accessed here.
1807	*/
1808	SUPR0DECL(int) SUPR0ObjAddRefEx(void *pvObj, PSUPDRVSESSION pSession, bool fNoBlocking)
1809	{
1810	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
1811	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
1812	PSUPDRVOBJ pObj = (PSUPDRVOBJ)pvObj;
1813	int rc = VINF_SUCCESS;
1814	PSUPDRVUSAGE pUsagePre;
1815	PSUPDRVUSAGE pUsage;
1816
1817	/*
1818	* Validate the input.
1819	* Be ready for the destruction race (someone might be stuck in the
1820	* destructor waiting a lock we own).
1821	*/
1822	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
1823	AssertPtrReturn(pObj, VERR_INVALID_POINTER);
1824	AssertMsgReturn(pObj->u32Magic == SUPDRVOBJ_MAGIC \|\| pObj->u32Magic == SUPDRVOBJ_MAGIC_DEAD,
1825	("Invalid pvObj=%p magic=%#x (expected %#x or %#x)\n", pvObj, pObj->u32Magic, SUPDRVOBJ_MAGIC, SUPDRVOBJ_MAGIC_DEAD),
1826	VERR_INVALID_PARAMETER);
1827
1828	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
1829
1830	if (RT_UNLIKELY(pObj->u32Magic != SUPDRVOBJ_MAGIC))
1831	{
1832	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
1833
1834	AssertMsgFailed(("pvObj=%p magic=%#x\n", pvObj, pObj->u32Magic));
1835	return VERR_WRONG_ORDER;
1836	}
1837
1838	/*
1839	* Preallocate the usage record if we can.
1840	*/
1841	pUsagePre = pDevExt->pUsageFree;
1842	if (pUsagePre)
1843	pDevExt->pUsageFree = pUsagePre->pNext;
1844	else if (!fNoBlocking)
1845	{
1846	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
1847	pUsagePre = (PSUPDRVUSAGE)RTMemAlloc(sizeof(*pUsagePre));
1848	if (!pUsagePre)
1849	return VERR_NO_MEMORY;
1850
1851	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
1852	if (RT_UNLIKELY(pObj->u32Magic != SUPDRVOBJ_MAGIC))
1853	{
1854	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
1855
1856	AssertMsgFailed(("pvObj=%p magic=%#x\n", pvObj, pObj->u32Magic));
1857	return VERR_WRONG_ORDER;
1858	}
1859	}
1860
1861	/*
1862	* Reference the object.
1863	*/
1864	pObj->cUsage++;
1865
1866	/*
1867	* Look for the session record.
1868	*/
1869	for (pUsage = pSession->pUsage; pUsage; pUsage = pUsage->pNext)
1870	{
1871	/Log(("SUPR0AddRef: pUsage=%p:{.pObj=%p, .pNext=%p}\n", pUsage, pUsage->pObj, pUsage->pNext));/
1872	if (pUsage->pObj == pObj)
1873	break;
1874	}
1875	if (pUsage)
1876	pUsage->cUsage++;
1877	else if (pUsagePre)
1878	{
1879	/* create a new session record. */
1880	pUsagePre->cUsage = 1;
1881	pUsagePre->pObj = pObj;
1882	pUsagePre->pNext = pSession->pUsage;
1883	pSession->pUsage = pUsagePre;
1884	/Log(("SUPR0AddRef: pUsagePre=%p:{.pObj=%p, .pNext=%p}\n", pUsagePre, pUsagePre->pObj, pUsagePre->pNext));/
1885
1886	pUsagePre = NULL;
1887	}
1888	else
1889	{
1890	pObj->cUsage--;
1891	rc = VERR_TRY_AGAIN;
1892	}
1893
1894	/*
1895	* Put any unused usage record into the free list..
1896	*/
1897	if (pUsagePre)
1898	{
1899	pUsagePre->pNext = pDevExt->pUsageFree;
1900	pDevExt->pUsageFree = pUsagePre;
1901	}
1902
1903	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
1904
1905	return rc;
1906	}
1907
1908
1909	/**
1910	* Decrement / destroy a reference counter record for an object.
1911	*
1912	* The object is uniquely identified by pfnDestructor+pvUser1+pvUser2.
1913	*
1914	* @returns IPRT status code.
1915	* @retval VINF_SUCCESS if not destroyed.
1916	* @retval VINF_OBJECT_DESTROYED if it's destroyed by this release call.
1917	* @retval VERR_INVALID_PARAMETER if the object isn't valid. Will assert in
1918	* string builds.
1919	*
1920	* @param pvObj The identifier returned by SUPR0ObjRegister().
1921	* @param pSession The session which is referencing the object.
1922	*/
1923	SUPR0DECL(int) SUPR0ObjRelease(void *pvObj, PSUPDRVSESSION pSession)
1924	{
1925	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
1926	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
1927	PSUPDRVOBJ pObj = (PSUPDRVOBJ)pvObj;
1928	int rc = VERR_INVALID_PARAMETER;
1929	PSUPDRVUSAGE pUsage;
1930	PSUPDRVUSAGE pUsagePrev;
1931
1932	/*
1933	* Validate the input.
1934	*/
1935	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
1936	AssertMsgReturn(VALID_PTR(pObj) && pObj->u32Magic == SUPDRVOBJ_MAGIC,
1937	("Invalid pvObj=%p magic=%#x (exepcted %#x)\n", pvObj, pObj ? pObj->u32Magic : 0, SUPDRVOBJ_MAGIC),
1938	VERR_INVALID_PARAMETER);
1939
1940	/*
1941	* Acquire the spinlock and look for the usage record.
1942	*/
1943	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
1944
1945	for (pUsagePrev = NULL, pUsage = pSession->pUsage;
1946	pUsage;
1947	pUsagePrev = pUsage, pUsage = pUsage->pNext)
1948	{
1949	/Log2(("SUPR0ObjRelease: pUsage=%p:{.pObj=%p, .pNext=%p}\n", pUsage, pUsage->pObj, pUsage->pNext));/
1950	if (pUsage->pObj == pObj)
1951	{
1952	rc = VINF_SUCCESS;
1953	AssertMsg(pUsage->cUsage >= 1 && pObj->cUsage >= pUsage->cUsage, ("glob %d; sess %d\n", pObj->cUsage, pUsage->cUsage));
1954	if (pUsage->cUsage > 1)
1955	{
1956	pObj->cUsage--;
1957	pUsage->cUsage--;
1958	}
1959	else
1960	{
1961	/*
1962	* Free the session record.
1963	*/
1964	if (pUsagePrev)
1965	pUsagePrev->pNext = pUsage->pNext;
1966	else
1967	pSession->pUsage = pUsage->pNext;
1968	pUsage->pNext = pDevExt->pUsageFree;
1969	pDevExt->pUsageFree = pUsage;
1970
1971	/* What about the object? */
1972	if (pObj->cUsage > 1)
1973	pObj->cUsage--;
1974	else
1975	{
1976	/*
1977	* Object is to be destroyed, unlink it.
1978	*/
1979	pObj->u32Magic = SUPDRVOBJ_MAGIC_DEAD;
1980	rc = VINF_OBJECT_DESTROYED;
1981	if (pDevExt->pObjs == pObj)
1982	pDevExt->pObjs = pObj->pNext;
1983	else
1984	{
1985	PSUPDRVOBJ pObjPrev;
1986	for (pObjPrev = pDevExt->pObjs; pObjPrev; pObjPrev = pObjPrev->pNext)
1987	if (pObjPrev->pNext == pObj)
1988	{
1989	pObjPrev->pNext = pObj->pNext;
1990	break;
1991	}
1992	Assert(pObjPrev);
1993	}
1994	}
1995	}
1996	break;
1997	}
1998	}
1999
2000	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
2001
2002	/*
2003	* Call the destructor and free the object if required.
2004	*/
2005	if (rc == VINF_OBJECT_DESTROYED)
2006	{
2007	Log(("SUPR0ObjRelease: destroying %p/%d (%p/%p) cpid=%RTproc pid=%RTproc dtor=%p\n",
2008	pObj, pObj->enmType, pObj->pvUser1, pObj->pvUser2, pObj->CreatorProcess, RTProcSelf(), pObj->pfnDestructor));
2009	if (pObj->pfnDestructor)
2010	#ifdef RT_WITH_W64_UNWIND_HACK
2011	supdrvNtWrapObjDestructor((PFNRT)pObj->pfnDestructor, pObj, pObj->pvUser1, pObj->pvUser2);
2012	#else
2013	pObj->pfnDestructor(pObj, pObj->pvUser1, pObj->pvUser2);
2014	#endif
2015	RTMemFree(pObj);
2016	}
2017
2018	AssertMsg(pUsage, ("pvObj=%p\n", pvObj));
2019	return rc;
2020	}
2021
2022
2023	/**
2024	* Verifies that the current process can access the specified object.
2025	*
2026	* @returns The following IPRT status code:
2027	* @retval VINF_SUCCESS if access was granted.
2028	* @retval VERR_PERMISSION_DENIED if denied access.
2029	* @retval VERR_INVALID_PARAMETER if invalid parameter.
2030	*
2031	* @param pvObj The identifier returned by SUPR0ObjRegister().
2032	* @param pSession The session which wishes to access the object.
2033	* @param pszObjName Object string name. This is optional and depends on the object type.
2034	*
2035	* @remark The caller is responsible for making sure the object isn't removed while
2036	* we're inside this function. If uncertain about this, just call AddRef before calling us.
2037	*/
2038	SUPR0DECL(int) SUPR0ObjVerifyAccess(void pvObj, PSUPDRVSESSION pSession, const char pszObjName)
2039	{
2040	PSUPDRVOBJ pObj = (PSUPDRVOBJ)pvObj;
2041	int rc;
2042
2043	/*
2044	* Validate the input.
2045	*/
2046	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2047	AssertMsgReturn(VALID_PTR(pObj) && pObj->u32Magic == SUPDRVOBJ_MAGIC,
2048	("Invalid pvObj=%p magic=%#x (exepcted %#x)\n", pvObj, pObj ? pObj->u32Magic : 0, SUPDRVOBJ_MAGIC),
2049	VERR_INVALID_PARAMETER);
2050
2051	/*
2052	* Check access. (returns true if a decision has been made.)
2053	*/
2054	rc = VERR_INTERNAL_ERROR;
2055	if (supdrvOSObjCanAccess(pObj, pSession, pszObjName, &rc))
2056	return rc;
2057
2058	/*
2059	* Default policy is to allow the user to access his own
2060	* stuff but nothing else.
2061	*/
2062	if (pObj->CreatorUid == pSession->Uid)
2063	return VINF_SUCCESS;
2064	return VERR_PERMISSION_DENIED;
2065	}
2066
2067
2068	/**
2069	* Lock pages.
2070	*
2071	* @returns IPRT status code.
2072	* @param pSession Session to which the locked memory should be associated.
2073	* @param pvR3 Start of the memory range to lock.
2074	* This must be page aligned.
2075	* @param cPages Number of pages to lock.
2076	* @param paPages Where to put the physical addresses of locked memory.
2077	*/
2078	SUPR0DECL(int) SUPR0LockMem(PSUPDRVSESSION pSession, RTR3PTR pvR3, uint32_t cPages, PRTHCPHYS paPages)
2079	{
2080	int rc;
2081	SUPDRVMEMREF Mem = { NIL_RTR0MEMOBJ, NIL_RTR0MEMOBJ, MEMREF_TYPE_UNUSED };
2082	const size_t cb = (size_t)cPages << PAGE_SHIFT;
2083	LogFlow(("SUPR0LockMem: pSession=%p pvR3=%p cPages=%d paPages=%p\n", pSession, (void *)pvR3, cPages, paPages));
2084
2085	/*
2086	* Verify input.
2087	*/
2088	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2089	AssertPtrReturn(paPages, VERR_INVALID_PARAMETER);
2090	if ( RT_ALIGN_R3PT(pvR3, PAGE_SIZE, RTR3PTR) != pvR3
2091	\|\| !pvR3)
2092	{
2093	Log(("pvR3 (%p) must be page aligned and not NULL!\n", (void *)pvR3));
2094	return VERR_INVALID_PARAMETER;
2095	}
2096
2097	#ifdef RT_OS_WINDOWS /* A temporary hack for windows, will be removed once all ring-3 code has been cleaned up. */
2098	/* First check if we allocated it using SUPPageAlloc; if so then we don't need to lock it again */
2099	rc = supdrvPageGetPhys(pSession, pvR3, cPages, paPages);
2100	if (RT_SUCCESS(rc))
2101	return rc;
2102	#endif
2103
2104	/*
2105	* Let IPRT do the job.
2106	*/
2107	Mem.eType = MEMREF_TYPE_LOCKED;
2108	rc = RTR0MemObjLockUser(&Mem.MemObj, pvR3, cb, RTR0ProcHandleSelf());
2109	if (RT_SUCCESS(rc))
2110	{
2111	uint32_t iPage = cPages;
2112	AssertMsg(RTR0MemObjAddressR3(Mem.MemObj) == pvR3, ("%p == %p\n", RTR0MemObjAddressR3(Mem.MemObj), pvR3));
2113	AssertMsg(RTR0MemObjSize(Mem.MemObj) == cb, ("%x == %x\n", RTR0MemObjSize(Mem.MemObj), cb));
2114
2115	while (iPage-- > 0)
2116	{
2117	paPages[iPage] = RTR0MemObjGetPagePhysAddr(Mem.MemObj, iPage);
2118	if (RT_UNLIKELY(paPages[iPage] == NIL_RTCCPHYS))
2119	{
2120	AssertMsgFailed(("iPage=%d\n", iPage));
2121	rc = VERR_INTERNAL_ERROR;
2122	break;
2123	}
2124	}
2125	if (RT_SUCCESS(rc))
2126	rc = supdrvMemAdd(&Mem, pSession);
2127	if (RT_FAILURE(rc))
2128	{
2129	int rc2 = RTR0MemObjFree(Mem.MemObj, false);
2130	AssertRC(rc2);
2131	}
2132	}
2133
2134	return rc;
2135	}
2136
2137
2138	/**
2139	* Unlocks the memory pointed to by pv.
2140	*
2141	* @returns IPRT status code.
2142	* @param pSession Session to which the memory was locked.
2143	* @param pvR3 Memory to unlock.
2144	*/
2145	SUPR0DECL(int) SUPR0UnlockMem(PSUPDRVSESSION pSession, RTR3PTR pvR3)
2146	{
2147	LogFlow(("SUPR0UnlockMem: pSession=%p pvR3=%p\n", pSession, (void *)pvR3));
2148	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2149	#ifdef RT_OS_WINDOWS
2150	/*
2151	* Temporary hack for windows - SUPR0PageFree will unlock SUPR0PageAlloc
2152	* allocations; ignore this call.
2153	*/
2154	if (supdrvPageWasLockedByPageAlloc(pSession, pvR3))
2155	{
2156	LogFlow(("Page will be unlocked in SUPR0PageFree -> ignore\n"));
2157	return VINF_SUCCESS;
2158	}
2159	#endif
2160	return supdrvMemRelease(pSession, (RTHCUINTPTR)pvR3, MEMREF_TYPE_LOCKED);
2161	}
2162
2163
2164	/**
2165	* Allocates a chunk of page aligned memory with contiguous and fixed physical
2166	* backing.
2167	*
2168	* @returns IPRT status code.
2169	* @param pSession Session data.
2170	* @param cPages Number of pages to allocate.
2171	* @param ppvR0 Where to put the address of Ring-0 mapping the allocated memory.
2172	* @param ppvR3 Where to put the address of Ring-3 mapping the allocated memory.
2173	* @param pHCPhys Where to put the physical address of allocated memory.
2174	*/
2175	SUPR0DECL(int) SUPR0ContAlloc(PSUPDRVSESSION pSession, uint32_t cPages, PRTR0PTR ppvR0, PRTR3PTR ppvR3, PRTHCPHYS pHCPhys)
2176	{
2177	int rc;
2178	SUPDRVMEMREF Mem = { NIL_RTR0MEMOBJ, NIL_RTR0MEMOBJ, MEMREF_TYPE_UNUSED };
2179	LogFlow(("SUPR0ContAlloc: pSession=%p cPages=%d ppvR0=%p ppvR3=%p pHCPhys=%p\n", pSession, cPages, ppvR0, ppvR3, pHCPhys));
2180
2181	/*
2182	* Validate input.
2183	*/
2184	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2185	if (!ppvR3 \|\| !ppvR0 \|\| !pHCPhys)
2186	{
2187	Log(("Null pointer. All of these should be set: pSession=%p ppvR0=%p ppvR3=%p pHCPhys=%p\n",
2188	pSession, ppvR0, ppvR3, pHCPhys));
2189	return VERR_INVALID_PARAMETER;
2190
2191	}
2192	if (cPages < 1 \|\| cPages >= 256)
2193	{
2194	Log(("Illegal request cPages=%d, must be greater than 0 and smaller than 256.\n", cPages));
2195	return VERR_PAGE_COUNT_OUT_OF_RANGE;
2196	}
2197
2198	/*
2199	* Let IPRT do the job.
2200	*/
2201	rc = RTR0MemObjAllocCont(&Mem.MemObj, cPages << PAGE_SHIFT, true /* executable R0 mapping */);
2202	if (RT_SUCCESS(rc))
2203	{
2204	int rc2;
2205	rc = RTR0MemObjMapUser(&Mem.MapObjR3, Mem.MemObj, (RTR3PTR)-1, 0,
2206	RTMEM_PROT_EXEC \| RTMEM_PROT_WRITE \| RTMEM_PROT_READ, RTR0ProcHandleSelf());
2207	if (RT_SUCCESS(rc))
2208	{
2209	Mem.eType = MEMREF_TYPE_CONT;
2210	rc = supdrvMemAdd(&Mem, pSession);
2211	if (!rc)
2212	{
2213	*ppvR0 = RTR0MemObjAddress(Mem.MemObj);
2214	*ppvR3 = RTR0MemObjAddressR3(Mem.MapObjR3);
2215	*pHCPhys = RTR0MemObjGetPagePhysAddr(Mem.MemObj, 0);
2216	return 0;
2217	}
2218
2219	rc2 = RTR0MemObjFree(Mem.MapObjR3, false);
2220	AssertRC(rc2);
2221	}
2222	rc2 = RTR0MemObjFree(Mem.MemObj, false);
2223	AssertRC(rc2);
2224	}
2225
2226	return rc;
2227	}
2228
2229
2230	/**
2231	* Frees memory allocated using SUPR0ContAlloc().
2232	*
2233	* @returns IPRT status code.
2234	* @param pSession The session to which the memory was allocated.
2235	* @param uPtr Pointer to the memory (ring-3 or ring-0).
2236	*/
2237	SUPR0DECL(int) SUPR0ContFree(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr)
2238	{
2239	LogFlow(("SUPR0ContFree: pSession=%p uPtr=%p\n", pSession, (void *)uPtr));
2240	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2241	return supdrvMemRelease(pSession, uPtr, MEMREF_TYPE_CONT);
2242	}
2243
2244
2245	/**
2246	* Allocates a chunk of page aligned memory with fixed physical backing below 4GB.
2247	*
2248	* The memory isn't zeroed.
2249	*
2250	* @returns IPRT status code.
2251	* @param pSession Session data.
2252	* @param cPages Number of pages to allocate.
2253	* @param ppvR0 Where to put the address of Ring-0 mapping of the allocated memory.
2254	* @param ppvR3 Where to put the address of Ring-3 mapping of the allocated memory.
2255	* @param paPages Where to put the physical addresses of allocated memory.
2256	*/
2257	SUPR0DECL(int) SUPR0LowAlloc(PSUPDRVSESSION pSession, uint32_t cPages, PRTR0PTR ppvR0, PRTR3PTR ppvR3, PRTHCPHYS paPages)
2258	{
2259	unsigned iPage;
2260	int rc;
2261	SUPDRVMEMREF Mem = { NIL_RTR0MEMOBJ, NIL_RTR0MEMOBJ, MEMREF_TYPE_UNUSED };
2262	LogFlow(("SUPR0LowAlloc: pSession=%p cPages=%d ppvR3=%p ppvR0=%p paPages=%p\n", pSession, cPages, ppvR3, ppvR0, paPages));
2263
2264	/*
2265	* Validate input.
2266	*/
2267	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2268	if (!ppvR3 \|\| !ppvR0 \|\| !paPages)
2269	{
2270	Log(("Null pointer. All of these should be set: pSession=%p ppvR3=%p ppvR0=%p paPages=%p\n",
2271	pSession, ppvR3, ppvR0, paPages));
2272	return VERR_INVALID_PARAMETER;
2273
2274	}
2275	if (cPages < 1 \|\| cPages >= 256)
2276	{
2277	Log(("Illegal request cPages=%d, must be greater than 0 and smaller than 256.\n", cPages));
2278	return VERR_PAGE_COUNT_OUT_OF_RANGE;
2279	}
2280
2281	/*
2282	* Let IPRT do the work.
2283	*/
2284	rc = RTR0MemObjAllocLow(&Mem.MemObj, cPages << PAGE_SHIFT, true /* executable ring-0 mapping */);
2285	if (RT_SUCCESS(rc))
2286	{
2287	int rc2;
2288	rc = RTR0MemObjMapUser(&Mem.MapObjR3, Mem.MemObj, (RTR3PTR)-1, 0,
2289	RTMEM_PROT_EXEC \| RTMEM_PROT_WRITE \| RTMEM_PROT_READ, RTR0ProcHandleSelf());
2290	if (RT_SUCCESS(rc))
2291	{
2292	Mem.eType = MEMREF_TYPE_LOW;
2293	rc = supdrvMemAdd(&Mem, pSession);
2294	if (!rc)
2295	{
2296	for (iPage = 0; iPage < cPages; iPage++)
2297	{
2298	paPages[iPage] = RTR0MemObjGetPagePhysAddr(Mem.MemObj, iPage);
2299	AssertMsg(!(paPages[iPage] & (PAGE_SIZE - 1)), ("iPage=%d Phys=%RHp\n", paPages[iPage]));
2300	}
2301	*ppvR0 = RTR0MemObjAddress(Mem.MemObj);
2302	*ppvR3 = RTR0MemObjAddressR3(Mem.MapObjR3);
2303	return 0;
2304	}
2305
2306	rc2 = RTR0MemObjFree(Mem.MapObjR3, false);
2307	AssertRC(rc2);
2308	}
2309
2310	rc2 = RTR0MemObjFree(Mem.MemObj, false);
2311	AssertRC(rc2);
2312	}
2313
2314	return rc;
2315	}
2316
2317
2318	/**
2319	* Frees memory allocated using SUPR0LowAlloc().
2320	*
2321	* @returns IPRT status code.
2322	* @param pSession The session to which the memory was allocated.
2323	* @param uPtr Pointer to the memory (ring-3 or ring-0).
2324	*/
2325	SUPR0DECL(int) SUPR0LowFree(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr)
2326	{
2327	LogFlow(("SUPR0LowFree: pSession=%p uPtr=%p\n", pSession, (void *)uPtr));
2328	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2329	return supdrvMemRelease(pSession, uPtr, MEMREF_TYPE_LOW);
2330	}
2331
2332
2333
2334	/**
2335	* Allocates a chunk of memory with both R0 and R3 mappings.
2336	* The memory is fixed and it's possible to query the physical addresses using SUPR0MemGetPhys().
2337	*
2338	* @returns IPRT status code.
2339	* @param pSession The session to associated the allocation with.
2340	* @param cb Number of bytes to allocate.
2341	* @param ppvR0 Where to store the address of the Ring-0 mapping.
2342	* @param ppvR3 Where to store the address of the Ring-3 mapping.
2343	*/
2344	SUPR0DECL(int) SUPR0MemAlloc(PSUPDRVSESSION pSession, uint32_t cb, PRTR0PTR ppvR0, PRTR3PTR ppvR3)
2345	{
2346	int rc;
2347	SUPDRVMEMREF Mem = { NIL_RTR0MEMOBJ, NIL_RTR0MEMOBJ, MEMREF_TYPE_UNUSED };
2348	LogFlow(("SUPR0MemAlloc: pSession=%p cb=%d ppvR0=%p ppvR3=%p\n", pSession, cb, ppvR0, ppvR3));
2349
2350	/*
2351	* Validate input.
2352	*/
2353	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2354	AssertPtrReturn(ppvR0, VERR_INVALID_POINTER);
2355	AssertPtrReturn(ppvR3, VERR_INVALID_POINTER);
2356	if (cb < 1 \|\| cb >= _4M)
2357	{
2358	Log(("Illegal request cb=%u; must be greater than 0 and smaller than 4MB.\n", cb));
2359	return VERR_INVALID_PARAMETER;
2360	}
2361
2362	/*
2363	* Let IPRT do the work.
2364	*/
2365	rc = RTR0MemObjAllocPage(&Mem.MemObj, cb, true /* executable ring-0 mapping */);
2366	if (RT_SUCCESS(rc))
2367	{
2368	int rc2;
2369	rc = RTR0MemObjMapUser(&Mem.MapObjR3, Mem.MemObj, (RTR3PTR)-1, 0,
2370	RTMEM_PROT_EXEC \| RTMEM_PROT_WRITE \| RTMEM_PROT_READ, RTR0ProcHandleSelf());
2371	if (RT_SUCCESS(rc))
2372	{
2373	Mem.eType = MEMREF_TYPE_MEM;
2374	rc = supdrvMemAdd(&Mem, pSession);
2375	if (!rc)
2376	{
2377	*ppvR0 = RTR0MemObjAddress(Mem.MemObj);
2378	*ppvR3 = RTR0MemObjAddressR3(Mem.MapObjR3);
2379	return VINF_SUCCESS;
2380	}
2381
2382	rc2 = RTR0MemObjFree(Mem.MapObjR3, false);
2383	AssertRC(rc2);
2384	}
2385
2386	rc2 = RTR0MemObjFree(Mem.MemObj, false);
2387	AssertRC(rc2);
2388	}
2389
2390	return rc;
2391	}
2392
2393
2394	/**
2395	* Get the physical addresses of memory allocated using SUPR0MemAlloc().
2396	*
2397	* @returns IPRT status code.
2398	* @param pSession The session to which the memory was allocated.
2399	* @param uPtr The Ring-0 or Ring-3 address returned by SUPR0MemAlloc().
2400	* @param paPages Where to store the physical addresses.
2401	*/
2402	SUPR0DECL(int) SUPR0MemGetPhys(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr, PSUPPAGE paPages) /** @todo switch this bugger to RTHCPHYS */
2403	{
2404	PSUPDRVBUNDLE pBundle;
2405	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
2406	LogFlow(("SUPR0MemGetPhys: pSession=%p uPtr=%p paPages=%p\n", pSession, (void *)uPtr, paPages));
2407
2408	/*
2409	* Validate input.
2410	*/
2411	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2412	AssertPtrReturn(paPages, VERR_INVALID_POINTER);
2413	AssertReturn(uPtr, VERR_INVALID_PARAMETER);
2414
2415	/*
2416	* Search for the address.
2417	*/
2418	RTSpinlockAcquire(pSession->Spinlock, &SpinlockTmp);
2419	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
2420	{
2421	if (pBundle->cUsed > 0)
2422	{
2423	unsigned i;
2424	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
2425	{
2426	if ( pBundle->aMem[i].eType == MEMREF_TYPE_MEM
2427	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
2428	&& ( (RTHCUINTPTR)RTR0MemObjAddress(pBundle->aMem[i].MemObj) == uPtr
2429	\|\| ( pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ
2430	&& RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3) == uPtr)
2431	)
2432	)
2433	{
2434	const size_t cPages = RTR0MemObjSize(pBundle->aMem[i].MemObj) >> PAGE_SHIFT;
2435	size_t iPage;
2436	for (iPage = 0; iPage < cPages; iPage++)
2437	{
2438	paPages[iPage].Phys = RTR0MemObjGetPagePhysAddr(pBundle->aMem[i].MemObj, iPage);
2439	paPages[iPage].uReserved = 0;
2440	}
2441	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
2442	return VINF_SUCCESS;
2443	}
2444	}
2445	}
2446	}
2447	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
2448	Log(("Failed to find %p!!!\n", (void *)uPtr));
2449	return VERR_INVALID_PARAMETER;
2450	}
2451
2452
2453	/**
2454	* Free memory allocated by SUPR0MemAlloc().
2455	*
2456	* @returns IPRT status code.
2457	* @param pSession The session owning the allocation.
2458	* @param uPtr The Ring-0 or Ring-3 address returned by SUPR0MemAlloc().
2459	*/
2460	SUPR0DECL(int) SUPR0MemFree(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr)
2461	{
2462	LogFlow(("SUPR0MemFree: pSession=%p uPtr=%p\n", pSession, (void *)uPtr));
2463	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2464	return supdrvMemRelease(pSession, uPtr, MEMREF_TYPE_MEM);
2465	}
2466
2467
2468	/**
2469	* Allocates a chunk of memory with only a R3 mappings.
2470	*
2471	* The memory is fixed and it's possible to query the physical addresses using
2472	* SUPR0MemGetPhys().
2473	*
2474	* @returns IPRT status code.
2475	* @param pSession The session to associated the allocation with.
2476	* @param cPages The number of pages to allocate.
2477	* @param ppvR3 Where to store the address of the Ring-3 mapping.
2478	* @param paPages Where to store the addresses of the pages. Optional.
2479	*/
2480	SUPR0DECL(int) SUPR0PageAlloc(PSUPDRVSESSION pSession, uint32_t cPages, PRTR3PTR ppvR3, PRTHCPHYS paPages)
2481	{
2482	AssertPtrReturn(ppvR3, VERR_INVALID_POINTER);
2483	return SUPR0PageAllocEx(pSession, cPages, 0 /fFlags/, ppvR3, NULL, paPages);
2484	}
2485
2486
2487	/**
2488	* Allocates a chunk of memory with a kernel or/and a user mode mapping.
2489	*
2490	* The memory is fixed and it's possible to query the physical addresses using
2491	* SUPR0MemGetPhys().
2492	*
2493	* @returns IPRT status code.
2494	* @param pSession The session to associated the allocation with.
2495	* @param cPages The number of pages to allocate.
2496	* @param fFlags Flags, reserved for the future. Must be zero.
2497	* @param ppvR3 Where to store the address of the Ring-3 mapping.
2498	* NULL if no ring-3 mapping.
2499	* @param ppvR3 Where to store the address of the Ring-0 mapping.
2500	* NULL if no ring-0 mapping.
2501	* @param paPages Where to store the addresses of the pages. Optional.
2502	*/
2503	SUPR0DECL(int) SUPR0PageAllocEx(PSUPDRVSESSION pSession, uint32_t cPages, uint32_t fFlags, PRTR3PTR ppvR3, PRTR0PTR ppvR0, PRTHCPHYS paPages)
2504	{
2505	int rc;
2506	SUPDRVMEMREF Mem = { NIL_RTR0MEMOBJ, NIL_RTR0MEMOBJ, MEMREF_TYPE_UNUSED };
2507	LogFlow(("SUPR0PageAlloc: pSession=%p cb=%d ppvR3=%p\n", pSession, cPages, ppvR3));
2508
2509	/*
2510	* Validate input. The allowed allocation size must be at least equal to the maximum guest VRAM size.
2511	*/
2512	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2513	AssertPtrNullReturn(ppvR3, VERR_INVALID_POINTER);
2514	AssertPtrNullReturn(ppvR0, VERR_INVALID_POINTER);
2515	AssertReturn(ppvR3 \|\| ppvR0, VERR_INVALID_PARAMETER);
2516	AssertReturn(!fFlags, VERR_INVALID_PARAMETER);
2517	if (cPages < 1 \|\| cPages > VBOX_MAX_ALLOC_PAGE_COUNT)
2518	{
2519	Log(("SUPR0PageAlloc: Illegal request cb=%u; must be greater than 0 and smaller than 128MB.\n", cPages));
2520	return VERR_PAGE_COUNT_OUT_OF_RANGE;
2521	}
2522
2523	/*
2524	* Let IPRT do the work.
2525	*/
2526	if (ppvR0)
2527	rc = RTR0MemObjAllocPage(&Mem.MemObj, (size_t)cPages * PAGE_SIZE, true /* fExecutable */);
2528	else
2529	rc = RTR0MemObjAllocPhysNC(&Mem.MemObj, (size_t)cPages * PAGE_SIZE, NIL_RTHCPHYS);
2530	if (RT_SUCCESS(rc))
2531	{
2532	int rc2;
2533	if (ppvR3)
2534	rc = RTR0MemObjMapUser(&Mem.MapObjR3, Mem.MemObj, (RTR3PTR)-1, 0,
2535	RTMEM_PROT_EXEC \| RTMEM_PROT_WRITE \| RTMEM_PROT_READ, RTR0ProcHandleSelf());
2536	else
2537	Mem.MapObjR3 = NIL_RTR0MEMOBJ;
2538	if (RT_SUCCESS(rc))
2539	{
2540	Mem.eType = MEMREF_TYPE_PAGE;
2541	rc = supdrvMemAdd(&Mem, pSession);
2542	if (!rc)
2543	{
2544	if (ppvR3)
2545	*ppvR3 = RTR0MemObjAddressR3(Mem.MapObjR3);
2546	if (ppvR0)
2547	*ppvR0 = RTR0MemObjAddress(Mem.MemObj);
2548	if (paPages)
2549	{
2550	uint32_t iPage = cPages;
2551	while (iPage-- > 0)
2552	{
2553	paPages[iPage] = RTR0MemObjGetPagePhysAddr(Mem.MapObjR3, iPage);
2554	Assert(paPages[iPage] != NIL_RTHCPHYS);
2555	}
2556	}
2557	return VINF_SUCCESS;
2558	}
2559
2560	rc2 = RTR0MemObjFree(Mem.MapObjR3, false);
2561	AssertRC(rc2);
2562	}
2563
2564	rc2 = RTR0MemObjFree(Mem.MemObj, false);
2565	AssertRC(rc2);
2566	}
2567	return rc;
2568	}
2569
2570
2571	/**
2572	* Allocates a chunk of memory with a kernel or/and a user mode mapping.
2573	*
2574	* The memory is fixed and it's possible to query the physical addresses using
2575	* SUPR0MemGetPhys().
2576	*
2577	* @returns IPRT status code.
2578	* @param pSession The session to associated the allocation with.
2579	* @param cPages The number of pages to allocate.
2580	* @param fFlags Flags, reserved for the future. Must be zero.
2581	* @param ppvR3 Where to store the address of the Ring-3 mapping.
2582	* NULL if no ring-3 mapping.
2583	* @param ppvR3 Where to store the address of the Ring-0 mapping.
2584	* NULL if no ring-0 mapping.
2585	* @param paPages Where to store the addresses of the pages. Optional.
2586	*/
2587	SUPR0DECL(int) SUPR0PageMapKernel(PSUPDRVSESSION pSession, RTR3PTR pvR3, uint32_t offSub, uint32_t cbSub,
2588	uint32_t fFlags, PRTR0PTR ppvR0)
2589	{
2590	int rc;
2591	PSUPDRVBUNDLE pBundle;
2592	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
2593	RTR0MEMOBJ hMemObj = NIL_RTR0MEMOBJ;
2594	LogFlow(("SUPR0PageMapKernel: pSession=%p pvR3=%p offSub=%#x cbSub=%#x\n", pSession, pvR3, offSub, cbSub));
2595
2596	/*
2597	* Validate input. The allowed allocation size must be at least equal to the maximum guest VRAM size.
2598	*/
2599	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2600	AssertPtrNullReturn(ppvR0, VERR_INVALID_POINTER);
2601	AssertReturn(!fFlags, VERR_INVALID_PARAMETER);
2602	AssertReturn(!(offSub & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
2603	AssertReturn(!(cbSub & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
2604	AssertReturn(cbSub, VERR_INVALID_PARAMETER);
2605
2606	/*
2607	* Find the memory object.
2608	*/
2609	RTSpinlockAcquire(pSession->Spinlock, &SpinlockTmp);
2610	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
2611	{
2612	if (pBundle->cUsed > 0)
2613	{
2614	unsigned i;
2615	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
2616	{
2617	if ( ( pBundle->aMem[i].eType == MEMREF_TYPE_PAGE
2618	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
2619	&& pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ
2620	&& RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3) == pvR3)
2621	\|\| ( pBundle->aMem[i].eType == MEMREF_TYPE_LOCKED
2622	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
2623	&& pBundle->aMem[i].MapObjR3 == NIL_RTR0MEMOBJ
2624	&& RTR0MemObjAddressR3(pBundle->aMem[i].MemObj) == pvR3))
2625	{
2626	hMemObj = pBundle->aMem[i].MemObj;
2627	break;
2628	}
2629	}
2630	}
2631	}
2632	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
2633
2634	rc = VERR_INVALID_PARAMETER;
2635	if (hMemObj != NIL_RTR0MEMOBJ)
2636	{
2637	/*
2638	* Do some furter input validations before calling IPRT.
2639	* (Cleanup is done indirectly by telling RTR0MemObjFree to include mappings.)
2640	*/
2641	size_t cbMemObj = RTR0MemObjSize(hMemObj);
2642	if ( offSub < cbMemObj
2643	&& cbSub <= cbMemObj
2644	&& offSub + cbSub <= cbMemObj)
2645	{
2646	RTR0MEMOBJ hMapObj;
2647	rc = RTR0MemObjMapKernelEx(&hMapObj, hMemObj, (void *)-1, 0,
2648	RTMEM_PROT_READ \| RTMEM_PROT_WRITE, offSub, cbSub);
2649	if (RT_SUCCESS(rc))
2650	*ppvR0 = RTR0MemObjAddress(hMapObj);
2651	}
2652	else
2653	SUPR0Printf("SUPR0PageMapKernel: cbMemObj=%#x offSub=%#x cbSub=%#x\n", cbMemObj, offSub, cbSub);
2654
2655	}
2656	return rc;
2657	}
2658
2659
2660
2661	#ifdef RT_OS_WINDOWS
2662	/**
2663	* Check if the pages were locked by SUPR0PageAlloc
2664	*
2665	* This function will be removed along with the lock/unlock hacks when
2666	* we've cleaned up the ring-3 code properly.
2667	*
2668	* @returns boolean
2669	* @param pSession The session to which the memory was allocated.
2670	* @param pvR3 The Ring-3 address returned by SUPR0PageAlloc().
2671	*/
2672	static bool supdrvPageWasLockedByPageAlloc(PSUPDRVSESSION pSession, RTR3PTR pvR3)
2673	{
2674	PSUPDRVBUNDLE pBundle;
2675	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
2676	LogFlow(("SUPR0PageIsLockedByPageAlloc: pSession=%p pvR3=%p\n", pSession, (void *)pvR3));
2677
2678	/*
2679	* Search for the address.
2680	*/
2681	RTSpinlockAcquire(pSession->Spinlock, &SpinlockTmp);
2682	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
2683	{
2684	if (pBundle->cUsed > 0)
2685	{
2686	unsigned i;
2687	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
2688	{
2689	if ( pBundle->aMem[i].eType == MEMREF_TYPE_PAGE
2690	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
2691	&& pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ
2692	&& RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3) == pvR3)
2693	{
2694	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
2695	return true;
2696	}
2697	}
2698	}
2699	}
2700	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
2701	return false;
2702	}
2703
2704
2705	/**
2706	* Get the physical addresses of memory allocated using SUPR0PageAllocEx().
2707	*
2708	* This function will be removed along with the lock/unlock hacks when
2709	* we've cleaned up the ring-3 code properly.
2710	*
2711	* @returns IPRT status code.
2712	* @param pSession The session to which the memory was allocated.
2713	* @param pvR3 The Ring-3 address returned by SUPR0PageAlloc().
2714	* @param cPages Number of pages in paPages
2715	* @param paPages Where to store the physical addresses.
2716	*/
2717	static int supdrvPageGetPhys(PSUPDRVSESSION pSession, RTR3PTR pvR3, uint32_t cPages, PRTHCPHYS paPages)
2718	{
2719	PSUPDRVBUNDLE pBundle;
2720	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
2721	LogFlow(("supdrvPageGetPhys: pSession=%p pvR3=%p cPages=%#lx paPages=%p\n", pSession, (void *)pvR3, (long)cPages, paPages));
2722
2723	/*
2724	* Search for the address.
2725	*/
2726	RTSpinlockAcquire(pSession->Spinlock, &SpinlockTmp);
2727	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
2728	{
2729	if (pBundle->cUsed > 0)
2730	{
2731	unsigned i;
2732	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
2733	{
2734	if ( pBundle->aMem[i].eType == MEMREF_TYPE_PAGE
2735	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
2736	&& pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ
2737	&& RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3) == pvR3)
2738	{
2739	uint32_t iPage;
2740	size_t cMaxPages = RTR0MemObjSize(pBundle->aMem[i].MemObj) >> PAGE_SHIFT;
2741	cPages = (uint32_t)RT_MIN(cMaxPages, cPages);
2742	for (iPage = 0; iPage < cPages; iPage++)
2743	paPages[iPage] = RTR0MemObjGetPagePhysAddr(pBundle->aMem[i].MemObj, iPage);
2744	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
2745	return VINF_SUCCESS;
2746	}
2747	}
2748	}
2749	}
2750	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
2751	return VERR_INVALID_PARAMETER;
2752	}
2753	#endif /* RT_OS_WINDOWS */
2754
2755
2756	/**
2757	* Free memory allocated by SUPR0PageAlloc() and SUPR0PageAllocEx().
2758	*
2759	* @returns IPRT status code.
2760	* @param pSession The session owning the allocation.
2761	* @param pvR3 The Ring-3 address returned by SUPR0PageAlloc() or
2762	* SUPR0PageAllocEx().
2763	*/
2764	SUPR0DECL(int) SUPR0PageFree(PSUPDRVSESSION pSession, RTR3PTR pvR3)
2765	{
2766	LogFlow(("SUPR0PageFree: pSession=%p pvR3=%p\n", pSession, (void *)pvR3));
2767	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2768	return supdrvMemRelease(pSession, (RTHCUINTPTR)pvR3, MEMREF_TYPE_PAGE);
2769	}
2770
2771
2772	/**
2773	* Maps the GIP into userspace and/or get the physical address of the GIP.
2774	*
2775	* @returns IPRT status code.
2776	* @param pSession Session to which the GIP mapping should belong.
2777	* @param ppGipR3 Where to store the address of the ring-3 mapping. (optional)
2778	* @param pHCPhysGip Where to store the physical address. (optional)
2779	*
2780	* @remark There is no reference counting on the mapping, so one call to this function
2781	* count globally as one reference. One call to SUPR0GipUnmap() is will unmap GIP
2782	* and remove the session as a GIP user.
2783	*/
2784	SUPR0DECL(int) SUPR0GipMap(PSUPDRVSESSION pSession, PRTR3PTR ppGipR3, PRTHCPHYS pHCPhysGip)
2785	{
2786	int rc = 0;
2787	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
2788	RTR3PTR pGip = NIL_RTR3PTR;
2789	RTHCPHYS HCPhys = NIL_RTHCPHYS;
2790	LogFlow(("SUPR0GipMap: pSession=%p ppGipR3=%p pHCPhysGip=%p\n", pSession, ppGipR3, pHCPhysGip));
2791
2792	/*
2793	* Validate
2794	*/
2795	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2796	AssertPtrNullReturn(ppGipR3, VERR_INVALID_POINTER);
2797	AssertPtrNullReturn(pHCPhysGip, VERR_INVALID_POINTER);
2798
2799	RTSemFastMutexRequest(pDevExt->mtxGip);
2800	if (pDevExt->pGip)
2801	{
2802	/*
2803	* Map it?
2804	*/
2805	if (ppGipR3)
2806	{
2807	if (pSession->GipMapObjR3 == NIL_RTR0MEMOBJ)
2808	rc = RTR0MemObjMapUser(&pSession->GipMapObjR3, pDevExt->GipMemObj, (RTR3PTR)-1, 0,
2809	RTMEM_PROT_READ, RTR0ProcHandleSelf());
2810	if (RT_SUCCESS(rc))
2811	{
2812	pGip = RTR0MemObjAddressR3(pSession->GipMapObjR3);
2813	rc = VINF_SUCCESS; /** @todo remove this and replace the !rc below with RT_SUCCESS(rc). */
2814	}
2815	}
2816
2817	/*
2818	* Get physical address.
2819	*/
2820	if (pHCPhysGip && !rc)
2821	HCPhys = pDevExt->HCPhysGip;
2822
2823	/*
2824	* Reference globally.
2825	*/
2826	if (!pSession->fGipReferenced && !rc)
2827	{
2828	pSession->fGipReferenced = 1;
2829	pDevExt->cGipUsers++;
2830	if (pDevExt->cGipUsers == 1)
2831	{
2832	PSUPGLOBALINFOPAGE pGip = pDevExt->pGip;
2833	unsigned i;
2834
2835	LogFlow(("SUPR0GipMap: Resumes GIP updating\n"));
2836
2837	for (i = 0; i < RT_ELEMENTS(pGip->aCPUs); i++)
2838	ASMAtomicXchgU32(&pGip->aCPUs[i].u32TransactionId, pGip->aCPUs[i].u32TransactionId & ~(GIP_UPDATEHZ_RECALC_FREQ * 2 - 1));
2839	ASMAtomicXchgU64(&pGip->u64NanoTSLastUpdateHz, 0);
2840
2841	rc = RTTimerStart(pDevExt->pGipTimer, 0);
2842	AssertRC(rc); rc = VINF_SUCCESS;
2843	}
2844	}
2845	}
2846	else
2847	{
2848	rc = SUPDRV_ERR_GENERAL_FAILURE;
2849	Log(("SUPR0GipMap: GIP is not available!\n"));
2850	}
2851	RTSemFastMutexRelease(pDevExt->mtxGip);
2852
2853	/*
2854	* Write returns.
2855	*/
2856	if (pHCPhysGip)
2857	*pHCPhysGip = HCPhys;
2858	if (ppGipR3)
2859	*ppGipR3 = pGip;
2860
2861	#ifdef DEBUG_DARWIN_GIP
2862	OSDBGPRINT(("SUPR0GipMap: returns %d pHCPhysGip=%lx pGip=%p\n", rc, (unsigned long)HCPhys, (void )pGip));
2863	#else
2864	LogFlow(( "SUPR0GipMap: returns %d pHCPhysGip=%lx pGip=%p\n", rc, (unsigned long)HCPhys, (void )pGip));
2865	#endif
2866	return rc;
2867	}
2868
2869
2870	/**
2871	* Unmaps any user mapping of the GIP and terminates all GIP access
2872	* from this session.
2873	*
2874	* @returns IPRT status code.
2875	* @param pSession Session to which the GIP mapping should belong.
2876	*/
2877	SUPR0DECL(int) SUPR0GipUnmap(PSUPDRVSESSION pSession)
2878	{
2879	int rc = VINF_SUCCESS;
2880	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
2881	#ifdef DEBUG_DARWIN_GIP
2882	OSDBGPRINT(("SUPR0GipUnmap: pSession=%p pGip=%p GipMapObjR3=%p\n",
2883	pSession,
2884	pSession->GipMapObjR3 != NIL_RTR0MEMOBJ ? RTR0MemObjAddress(pSession->GipMapObjR3) : NULL,
2885	pSession->GipMapObjR3));
2886	#else
2887	LogFlow(("SUPR0GipUnmap: pSession=%p\n", pSession));
2888	#endif
2889	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2890
2891	RTSemFastMutexRequest(pDevExt->mtxGip);
2892
2893	/*
2894	* Unmap anything?
2895	*/
2896	if (pSession->GipMapObjR3 != NIL_RTR0MEMOBJ)
2897	{
2898	rc = RTR0MemObjFree(pSession->GipMapObjR3, false);
2899	AssertRC(rc);
2900	if (RT_SUCCESS(rc))
2901	pSession->GipMapObjR3 = NIL_RTR0MEMOBJ;
2902	}
2903
2904	/*
2905	* Dereference global GIP.
2906	*/
2907	if (pSession->fGipReferenced && !rc)
2908	{
2909	pSession->fGipReferenced = 0;
2910	if ( pDevExt->cGipUsers > 0
2911	&& !--pDevExt->cGipUsers)
2912	{
2913	LogFlow(("SUPR0GipUnmap: Suspends GIP updating\n"));
2914	rc = RTTimerStop(pDevExt->pGipTimer); AssertRC(rc); rc = 0;
2915	}
2916	}
2917
2918	RTSemFastMutexRelease(pDevExt->mtxGip);
2919
2920	return rc;
2921	}
2922
2923
2924	/**
2925	* Register a component factory with the support driver.
2926	*
2927	* This is currently restricted to kernel sessions only.
2928	*
2929	* @returns VBox status code.
2930	* @retval VINF_SUCCESS on success.
2931	* @retval VERR_NO_MEMORY if we're out of memory.
2932	* @retval VERR_ALREADY_EXISTS if the factory has already been registered.
2933	* @retval VERR_ACCESS_DENIED if it isn't a kernel session.
2934	* @retval VERR_INVALID_PARAMETER on invalid parameter.
2935	* @retval VERR_INVALID_POINTER on invalid pointer parameter.
2936	*
2937	* @param pSession The SUPDRV session (must be a ring-0 session).
2938	* @param pFactory Pointer to the component factory registration structure.
2939	*
2940	* @remarks This interface is also available via SUPR0IdcComponentRegisterFactory.
2941	*/
2942	SUPR0DECL(int) SUPR0ComponentRegisterFactory(PSUPDRVSESSION pSession, PCSUPDRVFACTORY pFactory)
2943	{
2944	PSUPDRVFACTORYREG pNewReg;
2945	const char *psz;
2946	int rc;
2947
2948	/*
2949	* Validate parameters.
2950	*/
2951	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2952	AssertReturn(pSession->R0Process == NIL_RTR0PROCESS, VERR_ACCESS_DENIED);
2953	AssertPtrReturn(pFactory, VERR_INVALID_POINTER);
2954	AssertPtrReturn(pFactory->pfnQueryFactoryInterface, VERR_INVALID_POINTER);
2955	psz = (const char *)memchr(pFactory->szName, '\0', sizeof(pFactory->szName));
2956	AssertReturn(psz, VERR_INVALID_PARAMETER);
2957
2958	/*
2959	* Allocate and initialize a new registration structure.
2960	*/
2961	pNewReg = (PSUPDRVFACTORYREG)RTMemAlloc(sizeof(SUPDRVFACTORYREG));
2962	if (pNewReg)
2963	{
2964	pNewReg->pNext = NULL;
2965	pNewReg->pFactory = pFactory;
2966	pNewReg->pSession = pSession;
2967	pNewReg->cchName = psz - &pFactory->szName[0];
2968
2969	/*
2970	* Add it to the tail of the list after checking for prior registration.
2971	*/
2972	rc = RTSemFastMutexRequest(pSession->pDevExt->mtxComponentFactory);
2973	if (RT_SUCCESS(rc))
2974	{
2975	PSUPDRVFACTORYREG pPrev = NULL;
2976	PSUPDRVFACTORYREG pCur = pSession->pDevExt->pComponentFactoryHead;
2977	while (pCur && pCur->pFactory != pFactory)
2978	{
2979	pPrev = pCur;
2980	pCur = pCur->pNext;
2981	}
2982	if (!pCur)
2983	{
2984	if (pPrev)
2985	pPrev->pNext = pNewReg;
2986	else
2987	pSession->pDevExt->pComponentFactoryHead = pNewReg;
2988	rc = VINF_SUCCESS;
2989	}
2990	else
2991	rc = VERR_ALREADY_EXISTS;
2992
2993	RTSemFastMutexRelease(pSession->pDevExt->mtxComponentFactory);
2994	}
2995
2996	if (RT_FAILURE(rc))
2997	RTMemFree(pNewReg);
2998	}
2999	else
3000	rc = VERR_NO_MEMORY;
3001	return rc;
3002	}
3003
3004
3005	/**
3006	* Deregister a component factory.
3007	*
3008	* @returns VBox status code.
3009	* @retval VINF_SUCCESS on success.
3010	* @retval VERR_NOT_FOUND if the factory wasn't registered.
3011	* @retval VERR_ACCESS_DENIED if it isn't a kernel session.
3012	* @retval VERR_INVALID_PARAMETER on invalid parameter.
3013	* @retval VERR_INVALID_POINTER on invalid pointer parameter.
3014	*
3015	* @param pSession The SUPDRV session (must be a ring-0 session).
3016	* @param pFactory Pointer to the component factory registration structure
3017	* previously passed SUPR0ComponentRegisterFactory().
3018	*
3019	* @remarks This interface is also available via SUPR0IdcComponentDeregisterFactory.
3020	*/
3021	SUPR0DECL(int) SUPR0ComponentDeregisterFactory(PSUPDRVSESSION pSession, PCSUPDRVFACTORY pFactory)
3022	{
3023	int rc;
3024
3025	/*
3026	* Validate parameters.
3027	*/
3028	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3029	AssertReturn(pSession->R0Process == NIL_RTR0PROCESS, VERR_ACCESS_DENIED);
3030	AssertPtrReturn(pFactory, VERR_INVALID_POINTER);
3031
3032	/*
3033	* Take the lock and look for the registration record.
3034	*/
3035	rc = RTSemFastMutexRequest(pSession->pDevExt->mtxComponentFactory);
3036	if (RT_SUCCESS(rc))
3037	{
3038	PSUPDRVFACTORYREG pPrev = NULL;
3039	PSUPDRVFACTORYREG pCur = pSession->pDevExt->pComponentFactoryHead;
3040	while (pCur && pCur->pFactory != pFactory)
3041	{
3042	pPrev = pCur;
3043	pCur = pCur->pNext;
3044	}
3045	if (pCur)
3046	{
3047	if (!pPrev)
3048	pSession->pDevExt->pComponentFactoryHead = pCur->pNext;
3049	else
3050	pPrev->pNext = pCur->pNext;
3051
3052	pCur->pNext = NULL;
3053	pCur->pFactory = NULL;
3054	pCur->pSession = NULL;
3055	rc = VINF_SUCCESS;
3056	}
3057	else
3058	rc = VERR_NOT_FOUND;
3059
3060	RTSemFastMutexRelease(pSession->pDevExt->mtxComponentFactory);
3061
3062	RTMemFree(pCur);
3063	}
3064	return rc;
3065	}
3066
3067
3068	/**
3069	* Queries a component factory.
3070	*
3071	* @returns VBox status code.
3072	* @retval VERR_INVALID_PARAMETER on invalid parameter.
3073	* @retval VERR_INVALID_POINTER on invalid pointer parameter.
3074	* @retval VERR_SUPDRV_COMPONENT_NOT_FOUND if the component factory wasn't found.
3075	* @retval VERR_SUPDRV_INTERFACE_NOT_SUPPORTED if the interface wasn't supported.
3076	*
3077	* @param pSession The SUPDRV session.
3078	* @param pszName The name of the component factory.
3079	* @param pszInterfaceUuid The UUID of the factory interface (stringified).
3080	* @param ppvFactoryIf Where to store the factory interface.
3081	*/
3082	SUPR0DECL(int) SUPR0ComponentQueryFactory(PSUPDRVSESSION pSession, const char pszName, const char pszInterfaceUuid, void **ppvFactoryIf)
3083	{
3084	const char *pszEnd;
3085	size_t cchName;
3086	int rc;
3087
3088	/*
3089	* Validate parameters.
3090	*/
3091	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3092
3093	AssertPtrReturn(pszName, VERR_INVALID_POINTER);
3094	pszEnd = memchr(pszName, '\0', RT_SIZEOFMEMB(SUPDRVFACTORY, szName));
3095	AssertReturn(pszEnd, VERR_INVALID_PARAMETER);
3096	cchName = pszEnd - pszName;
3097
3098	AssertPtrReturn(pszInterfaceUuid, VERR_INVALID_POINTER);
3099	pszEnd = memchr(pszInterfaceUuid, '\0', RTUUID_STR_LENGTH);
3100	AssertReturn(pszEnd, VERR_INVALID_PARAMETER);
3101
3102	AssertPtrReturn(ppvFactoryIf, VERR_INVALID_POINTER);
3103	*ppvFactoryIf = NULL;
3104
3105	/*
3106	* Take the lock and try all factories by this name.
3107	*/
3108	rc = RTSemFastMutexRequest(pSession->pDevExt->mtxComponentFactory);
3109	if (RT_SUCCESS(rc))
3110	{
3111	PSUPDRVFACTORYREG pCur = pSession->pDevExt->pComponentFactoryHead;
3112	rc = VERR_SUPDRV_COMPONENT_NOT_FOUND;
3113	while (pCur)
3114	{
3115	if ( pCur->cchName == cchName
3116	&& !memcmp(pCur->pFactory->szName, pszName, cchName))
3117	{
3118	#ifdef RT_WITH_W64_UNWIND_HACK
3119	void *pvFactory = supdrvNtWrapQueryFactoryInterface((PFNRT)pCur->pFactory->pfnQueryFactoryInterface, pCur->pFactory, pSession, pszInterfaceUuid);
3120	#else
3121	void *pvFactory = pCur->pFactory->pfnQueryFactoryInterface(pCur->pFactory, pSession, pszInterfaceUuid);
3122	#endif
3123	if (pvFactory)
3124	{
3125	*ppvFactoryIf = pvFactory;
3126	rc = VINF_SUCCESS;
3127	break;
3128	}
3129	rc = VERR_SUPDRV_INTERFACE_NOT_SUPPORTED;
3130	}
3131
3132	/* next */
3133	pCur = pCur->pNext;
3134	}
3135
3136	RTSemFastMutexRelease(pSession->pDevExt->mtxComponentFactory);
3137	}
3138	return rc;
3139	}
3140
3141
3142	/**
3143	* Adds a memory object to the session.
3144	*
3145	* @returns IPRT status code.
3146	* @param pMem Memory tracking structure containing the
3147	* information to track.
3148	* @param pSession The session.
3149	*/
3150	static int supdrvMemAdd(PSUPDRVMEMREF pMem, PSUPDRVSESSION pSession)
3151	{
3152	PSUPDRVBUNDLE pBundle;
3153	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
3154
3155	/*
3156	* Find free entry and record the allocation.
3157	*/
3158	RTSpinlockAcquire(pSession->Spinlock, &SpinlockTmp);
3159	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
3160	{
3161	if (pBundle->cUsed < RT_ELEMENTS(pBundle->aMem))
3162	{
3163	unsigned i;
3164	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
3165	{
3166	if (pBundle->aMem[i].MemObj == NIL_RTR0MEMOBJ)
3167	{
3168	pBundle->cUsed++;
3169	pBundle->aMem[i] = *pMem;
3170	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
3171	return VINF_SUCCESS;
3172	}
3173	}
3174	AssertFailed(); /* !!this can't be happening!!! */
3175	}
3176	}
3177	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
3178
3179	/*
3180	* Need to allocate a new bundle.
3181	* Insert into the last entry in the bundle.
3182	*/
3183	pBundle = (PSUPDRVBUNDLE)RTMemAllocZ(sizeof(*pBundle));
3184	if (!pBundle)
3185	return VERR_NO_MEMORY;
3186
3187	/* take last entry. */
3188	pBundle->cUsed++;
3189	pBundle->aMem[RT_ELEMENTS(pBundle->aMem) - 1] = *pMem;
3190
3191	/* insert into list. */
3192	RTSpinlockAcquire(pSession->Spinlock, &SpinlockTmp);
3193	pBundle->pNext = pSession->Bundle.pNext;
3194	pSession->Bundle.pNext = pBundle;
3195	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
3196
3197	return VINF_SUCCESS;
3198	}
3199
3200
3201	/**
3202	* Releases a memory object referenced by pointer and type.
3203	*
3204	* @returns IPRT status code.
3205	* @param pSession Session data.
3206	* @param uPtr Pointer to memory. This is matched against both the R0 and R3 addresses.
3207	* @param eType Memory type.
3208	*/
3209	static int supdrvMemRelease(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr, SUPDRVMEMREFTYPE eType)
3210	{
3211	PSUPDRVBUNDLE pBundle;
3212	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
3213
3214	/*
3215	* Validate input.
3216	*/
3217	if (!uPtr)
3218	{
3219	Log(("Illegal address %p\n", (void *)uPtr));
3220	return VERR_INVALID_PARAMETER;
3221	}
3222
3223	/*
3224	* Search for the address.
3225	*/
3226	RTSpinlockAcquire(pSession->Spinlock, &SpinlockTmp);
3227	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
3228	{
3229	if (pBundle->cUsed > 0)
3230	{
3231	unsigned i;
3232	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
3233	{
3234	if ( pBundle->aMem[i].eType == eType
3235	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
3236	&& ( (RTHCUINTPTR)RTR0MemObjAddress(pBundle->aMem[i].MemObj) == uPtr
3237	\|\| ( pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ
3238	&& RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3) == uPtr))
3239	)
3240	{
3241	/* Make a copy of it and release it outside the spinlock. */
3242	SUPDRVMEMREF Mem = pBundle->aMem[i];
3243	pBundle->aMem[i].eType = MEMREF_TYPE_UNUSED;
3244	pBundle->aMem[i].MemObj = NIL_RTR0MEMOBJ;
3245	pBundle->aMem[i].MapObjR3 = NIL_RTR0MEMOBJ;
3246	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
3247
3248	if (Mem.MapObjR3 != NIL_RTR0MEMOBJ)
3249	{
3250	int rc = RTR0MemObjFree(Mem.MapObjR3, false);
3251	AssertRC(rc); /** @todo figure out how to handle this. */
3252	}
3253	if (Mem.MemObj != NIL_RTR0MEMOBJ)
3254	{
3255	int rc = RTR0MemObjFree(Mem.MemObj, true /* fFreeMappings */);
3256	AssertRC(rc); /** @todo figure out how to handle this. */
3257	}
3258	return VINF_SUCCESS;
3259	}
3260	}
3261	}
3262	}
3263	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
3264	Log(("Failed to find %p!!! (eType=%d)\n", (void *)uPtr, eType));
3265	return VERR_INVALID_PARAMETER;
3266	}
3267
3268
3269	/**
3270	* Opens an image. If it's the first time it's opened the call must upload
3271	* the bits using the supdrvIOCtl_LdrLoad() / SUPDRV_IOCTL_LDR_LOAD function.
3272	*
3273	* This is the 1st step of the loading.
3274	*
3275	* @returns IPRT status code.
3276	* @param pDevExt Device globals.
3277	* @param pSession Session data.
3278	* @param pReq The open request.
3279	*/
3280	static int supdrvIOCtl_LdrOpen(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDROPEN pReq)
3281	{
3282	PSUPDRVLDRIMAGE pImage;
3283	unsigned cb;
3284	void *pv;
3285	LogFlow(("supdrvIOCtl_LdrOpen: szName=%s cbImage=%d\n", pReq->u.In.szName, pReq->u.In.cbImage));
3286
3287	/*
3288	* Check if we got an instance of the image already.
3289	*/
3290	RTSemFastMutexRequest(pDevExt->mtxLdr);
3291	for (pImage = pDevExt->pLdrImages; pImage; pImage = pImage->pNext)
3292	{
3293	if (!strcmp(pImage->szName, pReq->u.In.szName))
3294	{
3295	pImage->cUsage++;
3296	pReq->u.Out.pvImageBase = pImage->pvImage;
3297	pReq->u.Out.fNeedsLoading = pImage->uState == SUP_IOCTL_LDR_OPEN;
3298	supdrvLdrAddUsage(pSession, pImage);
3299	RTSemFastMutexRelease(pDevExt->mtxLdr);
3300	return VINF_SUCCESS;
3301	}
3302	}
3303	/* (not found - add it!) */
3304
3305	/*
3306	* Allocate memory.
3307	*/
3308	cb = pReq->u.In.cbImage + sizeof(SUPDRVLDRIMAGE) + 31;
3309	pv = RTMemExecAlloc(cb);
3310	if (!pv)
3311	{
3312	RTSemFastMutexRelease(pDevExt->mtxLdr);
3313	Log(("supdrvIOCtl_LdrOpen: RTMemExecAlloc(%u) failed\n", cb));
3314	return VERR_NO_MEMORY;
3315	}
3316
3317	/*
3318	* Setup and link in the LDR stuff.
3319	*/
3320	pImage = (PSUPDRVLDRIMAGE)pv;
3321	pImage->pvImage = RT_ALIGN_P(pImage + 1, 32);
3322	pImage->cbImage = pReq->u.In.cbImage;
3323	pImage->pfnModuleInit = NULL;
3324	pImage->pfnModuleTerm = NULL;
3325	pImage->pfnServiceReqHandler = NULL;
3326	pImage->uState = SUP_IOCTL_LDR_OPEN;
3327	pImage->cUsage = 1;
3328	strcpy(pImage->szName, pReq->u.In.szName);
3329
3330	pImage->pNext = pDevExt->pLdrImages;
3331	pDevExt->pLdrImages = pImage;
3332
3333	supdrvLdrAddUsage(pSession, pImage);
3334
3335	pReq->u.Out.pvImageBase = pImage->pvImage;
3336	pReq->u.Out.fNeedsLoading = true;
3337	RTSemFastMutexRelease(pDevExt->mtxLdr);
3338
3339	#if defined(RT_OS_WINDOWS) && defined(DEBUG)
3340	SUPR0Printf("VBoxDrv: windbg> .reload /f %s=%#p\n", pImage->szName, pImage->pvImage);
3341	#endif
3342	return VINF_SUCCESS;
3343	}
3344
3345
3346	/**
3347	* Loads the image bits.
3348	*
3349	* This is the 2nd step of the loading.
3350	*
3351	* @returns IPRT status code.
3352	* @param pDevExt Device globals.
3353	* @param pSession Session data.
3354	* @param pReq The request.
3355	*/
3356	static int supdrvIOCtl_LdrLoad(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRLOAD pReq)
3357	{
3358	PSUPDRVLDRUSAGE pUsage;
3359	PSUPDRVLDRIMAGE pImage;
3360	int rc;
3361	LogFlow(("supdrvIOCtl_LdrLoad: pvImageBase=%p cbImage=%d\n", pReq->u.In.pvImageBase, pReq->u.In.cbImage));
3362
3363	/*
3364	* Find the ldr image.
3365	*/
3366	RTSemFastMutexRequest(pDevExt->mtxLdr);
3367	pUsage = pSession->pLdrUsage;
3368	while (pUsage && pUsage->pImage->pvImage != pReq->u.In.pvImageBase)
3369	pUsage = pUsage->pNext;
3370	if (!pUsage)
3371	{
3372	RTSemFastMutexRelease(pDevExt->mtxLdr);
3373	Log(("SUP_IOCTL_LDR_LOAD: couldn't find image!\n"));
3374	return VERR_INVALID_HANDLE;
3375	}
3376	pImage = pUsage->pImage;
3377	if (pImage->cbImage != pReq->u.In.cbImage)
3378	{
3379	RTSemFastMutexRelease(pDevExt->mtxLdr);
3380	Log(("SUP_IOCTL_LDR_LOAD: image size mismatch!! %d(prep) != %d(load)\n", pImage->cbImage, pReq->u.In.cbImage));
3381	return VERR_INVALID_HANDLE;
3382	}
3383	if (pImage->uState != SUP_IOCTL_LDR_OPEN)
3384	{
3385	unsigned uState = pImage->uState;
3386	RTSemFastMutexRelease(pDevExt->mtxLdr);
3387	if (uState != SUP_IOCTL_LDR_LOAD)
3388	AssertMsgFailed(("SUP_IOCTL_LDR_LOAD: invalid image state %d (%#x)!\n", uState, uState));
3389	return SUPDRV_ERR_ALREADY_LOADED;
3390	}
3391	switch (pReq->u.In.eEPType)
3392	{
3393	case SUPLDRLOADEP_NOTHING:
3394	break;
3395
3396	case SUPLDRLOADEP_VMMR0:
3397	if ( !pReq->u.In.EP.VMMR0.pvVMMR0
3398	\|\| !pReq->u.In.EP.VMMR0.pvVMMR0EntryInt
3399	\|\| !pReq->u.In.EP.VMMR0.pvVMMR0EntryFast
3400	\|\| !pReq->u.In.EP.VMMR0.pvVMMR0EntryEx)
3401	{
3402	RTSemFastMutexRelease(pDevExt->mtxLdr);
3403	Log(("NULL pointer: pvVMMR0=%p pvVMMR0EntryInt=%p pvVMMR0EntryFast=%p pvVMMR0EntryEx=%p!\n",
3404	pReq->u.In.EP.VMMR0.pvVMMR0, pReq->u.In.EP.VMMR0.pvVMMR0EntryInt,
3405	pReq->u.In.EP.VMMR0.pvVMMR0EntryFast, pReq->u.In.EP.VMMR0.pvVMMR0EntryEx));
3406	return VERR_INVALID_PARAMETER;
3407	}
3408	/** @todo validate pReq->u.In.EP.VMMR0.pvVMMR0 against pvImage! */
3409	if ( (uintptr_t)pReq->u.In.EP.VMMR0.pvVMMR0EntryInt - (uintptr_t)pImage->pvImage >= pReq->u.In.cbImage
3410	\|\| (uintptr_t)pReq->u.In.EP.VMMR0.pvVMMR0EntryFast - (uintptr_t)pImage->pvImage >= pReq->u.In.cbImage
3411	\|\| (uintptr_t)pReq->u.In.EP.VMMR0.pvVMMR0EntryEx - (uintptr_t)pImage->pvImage >= pReq->u.In.cbImage)
3412	{
3413	RTSemFastMutexRelease(pDevExt->mtxLdr);
3414	Log(("Out of range (%p LB %#x): pvVMMR0EntryInt=%p, pvVMMR0EntryFast=%p or pvVMMR0EntryEx=%p is NULL!\n",
3415	pImage->pvImage, pReq->u.In.cbImage, pReq->u.In.EP.VMMR0.pvVMMR0EntryInt,
3416	pReq->u.In.EP.VMMR0.pvVMMR0EntryFast, pReq->u.In.EP.VMMR0.pvVMMR0EntryEx));
3417	return VERR_INVALID_PARAMETER;
3418	}
3419	break;
3420
3421	case SUPLDRLOADEP_SERVICE:
3422	if (!pReq->u.In.EP.Service.pfnServiceReq)
3423	{
3424	RTSemFastMutexRelease(pDevExt->mtxLdr);
3425	Log(("NULL pointer: pfnServiceReq=%p!\n", pReq->u.In.EP.Service.pfnServiceReq));
3426	return VERR_INVALID_PARAMETER;
3427	}
3428	if ((uintptr_t)pReq->u.In.EP.Service.pfnServiceReq - (uintptr_t)pImage->pvImage >= pReq->u.In.cbImage)
3429	{
3430	RTSemFastMutexRelease(pDevExt->mtxLdr);
3431	Log(("Out of range (%p LB %#x): pfnServiceReq=%p, pvVMMR0EntryFast=%p or pvVMMR0EntryEx=%p is NULL!\n",
3432	pImage->pvImage, pReq->u.In.cbImage, pReq->u.In.EP.Service.pfnServiceReq));
3433	return VERR_INVALID_PARAMETER;
3434	}
3435	if ( pReq->u.In.EP.Service.apvReserved[0] != NIL_RTR0PTR
3436	\|\| pReq->u.In.EP.Service.apvReserved[1] != NIL_RTR0PTR
3437	\|\| pReq->u.In.EP.Service.apvReserved[2] != NIL_RTR0PTR)
3438	{
3439	RTSemFastMutexRelease(pDevExt->mtxLdr);
3440	Log(("Out of range (%p LB %#x): apvReserved={%p,%p,%p} MBZ!\n",
3441	pImage->pvImage, pReq->u.In.cbImage,
3442	pReq->u.In.EP.Service.apvReserved[0],
3443	pReq->u.In.EP.Service.apvReserved[1],
3444	pReq->u.In.EP.Service.apvReserved[2]));
3445	return VERR_INVALID_PARAMETER;
3446	}
3447	break;
3448
3449	default:
3450	RTSemFastMutexRelease(pDevExt->mtxLdr);
3451	Log(("Invalid eEPType=%d\n", pReq->u.In.eEPType));
3452	return VERR_INVALID_PARAMETER;
3453	}
3454	if ( pReq->u.In.pfnModuleInit
3455	&& (uintptr_t)pReq->u.In.pfnModuleInit - (uintptr_t)pImage->pvImage >= pReq->u.In.cbImage)
3456	{
3457	RTSemFastMutexRelease(pDevExt->mtxLdr);
3458	Log(("SUP_IOCTL_LDR_LOAD: pfnModuleInit=%p is outside the image (%p %d bytes)\n",
3459	pReq->u.In.pfnModuleInit, pImage->pvImage, pReq->u.In.cbImage));
3460	return VERR_INVALID_PARAMETER;
3461	}
3462	if ( pReq->u.In.pfnModuleTerm
3463	&& (uintptr_t)pReq->u.In.pfnModuleTerm - (uintptr_t)pImage->pvImage >= pReq->u.In.cbImage)
3464	{
3465	RTSemFastMutexRelease(pDevExt->mtxLdr);
3466	Log(("SUP_IOCTL_LDR_LOAD: pfnModuleTerm=%p is outside the image (%p %d bytes)\n",
3467	pReq->u.In.pfnModuleTerm, pImage->pvImage, pReq->u.In.cbImage));
3468	return VERR_INVALID_PARAMETER;
3469	}
3470
3471	/*
3472	* Copy the memory.
3473	*/
3474	/* no need to do try/except as this is a buffered request. */
3475	memcpy(pImage->pvImage, &pReq->u.In.achImage[0], pImage->cbImage);
3476	pImage->uState = SUP_IOCTL_LDR_LOAD;
3477	pImage->pfnModuleInit = pReq->u.In.pfnModuleInit;
3478	pImage->pfnModuleTerm = pReq->u.In.pfnModuleTerm;
3479	pImage->offSymbols = pReq->u.In.offSymbols;
3480	pImage->cSymbols = pReq->u.In.cSymbols;
3481	pImage->offStrTab = pReq->u.In.offStrTab;
3482	pImage->cbStrTab = pReq->u.In.cbStrTab;
3483
3484	/*
3485	* Update any entry points.
3486	*/
3487	switch (pReq->u.In.eEPType)
3488	{
3489	default:
3490	case SUPLDRLOADEP_NOTHING:
3491	rc = VINF_SUCCESS;
3492	break;
3493	case SUPLDRLOADEP_VMMR0:
3494	rc = supdrvLdrSetVMMR0EPs(pDevExt, pReq->u.In.EP.VMMR0.pvVMMR0, pReq->u.In.EP.VMMR0.pvVMMR0EntryInt,
3495	pReq->u.In.EP.VMMR0.pvVMMR0EntryFast, pReq->u.In.EP.VMMR0.pvVMMR0EntryEx);
3496	break;
3497	case SUPLDRLOADEP_SERVICE:
3498	pImage->pfnServiceReqHandler = pReq->u.In.EP.Service.pfnServiceReq;
3499	rc = VINF_SUCCESS;
3500	break;
3501	}
3502
3503	/*
3504	* On success call the module initialization.
3505	*/
3506	LogFlow(("supdrvIOCtl_LdrLoad: pfnModuleInit=%p\n", pImage->pfnModuleInit));
3507	if (RT_SUCCESS(rc) && pImage->pfnModuleInit)
3508	{
3509	Log(("supdrvIOCtl_LdrLoad: calling pfnModuleInit=%p\n", pImage->pfnModuleInit));
3510	#ifdef RT_WITH_W64_UNWIND_HACK
3511	rc = supdrvNtWrapModuleInit((PFNRT)pImage->pfnModuleInit);
3512	#else
3513	rc = pImage->pfnModuleInit();
3514	#endif
3515	if (rc && pDevExt->pvVMMR0 == pImage->pvImage)
3516	supdrvLdrUnsetVMMR0EPs(pDevExt);
3517	}
3518
3519	if (rc)
3520	pImage->uState = SUP_IOCTL_LDR_OPEN;
3521
3522	RTSemFastMutexRelease(pDevExt->mtxLdr);
3523	return rc;
3524	}
3525
3526
3527	/**
3528	* Frees a previously loaded (prep'ed) image.
3529	*
3530	* @returns IPRT status code.
3531	* @param pDevExt Device globals.
3532	* @param pSession Session data.
3533	* @param pReq The request.
3534	*/
3535	static int supdrvIOCtl_LdrFree(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRFREE pReq)
3536	{
3537	int rc;
3538	PSUPDRVLDRUSAGE pUsagePrev;
3539	PSUPDRVLDRUSAGE pUsage;
3540	PSUPDRVLDRIMAGE pImage;
3541	LogFlow(("supdrvIOCtl_LdrFree: pvImageBase=%p\n", pReq->u.In.pvImageBase));
3542
3543	/*
3544	* Find the ldr image.
3545	*/
3546	RTSemFastMutexRequest(pDevExt->mtxLdr);
3547	pUsagePrev = NULL;
3548	pUsage = pSession->pLdrUsage;
3549	while (pUsage && pUsage->pImage->pvImage != pReq->u.In.pvImageBase)
3550	{
3551	pUsagePrev = pUsage;
3552	pUsage = pUsage->pNext;
3553	}
3554	if (!pUsage)
3555	{
3556	RTSemFastMutexRelease(pDevExt->mtxLdr);
3557	Log(("SUP_IOCTL_LDR_FREE: couldn't find image!\n"));
3558	return VERR_INVALID_HANDLE;
3559	}
3560
3561	/*
3562	* Check if we can remove anything.
3563	*/
3564	rc = VINF_SUCCESS;
3565	pImage = pUsage->pImage;
3566	if (pImage->cUsage <= 1 \|\| pUsage->cUsage <= 1)
3567	{
3568	/*
3569	* Check if there are any objects with destructors in the image, if
3570	* so leave it for the session cleanup routine so we get a chance to
3571	* clean things up in the right order and not leave them all dangling.
3572	*/
3573	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
3574	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
3575	if (pImage->cUsage <= 1)
3576	{
3577	PSUPDRVOBJ pObj;
3578	for (pObj = pDevExt->pObjs; pObj; pObj = pObj->pNext)
3579	if (RT_UNLIKELY((uintptr_t)pObj->pfnDestructor - (uintptr_t)pImage->pvImage < pImage->cbImage))
3580	{
3581	rc = VERR_DANGLING_OBJECTS;
3582	break;
3583	}
3584	}
3585	else
3586	{
3587	PSUPDRVUSAGE pGenUsage;
3588	for (pGenUsage = pSession->pUsage; pGenUsage; pGenUsage = pGenUsage->pNext)
3589	if (RT_UNLIKELY((uintptr_t)pGenUsage->pObj->pfnDestructor - (uintptr_t)pImage->pvImage < pImage->cbImage))
3590	{
3591	rc = VERR_DANGLING_OBJECTS;
3592	break;
3593	}
3594	}
3595	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
3596	if (rc == VINF_SUCCESS)
3597	{
3598	/* unlink it */
3599	if (pUsagePrev)
3600	pUsagePrev->pNext = pUsage->pNext;
3601	else
3602	pSession->pLdrUsage = pUsage->pNext;
3603
3604	/* free it */
3605	pUsage->pImage = NULL;
3606	pUsage->pNext = NULL;
3607	RTMemFree(pUsage);
3608
3609	/*
3610	* Derefrence the image.
3611	*/
3612	if (pImage->cUsage <= 1)
3613	supdrvLdrFree(pDevExt, pImage);
3614	else
3615	pImage->cUsage--;
3616	}
3617	else
3618	{
3619	Log(("supdrvIOCtl_LdrFree: Dangling objects in %p/%s!\n", pImage->pvImage, pImage->szName));
3620	rc = VINF_SUCCESS; /** @todo BRANCH-2.1: remove this after branching. */
3621	}
3622	}
3623	else
3624	{
3625	/*
3626	* Dereference both image and usage.
3627	*/
3628	pImage->cUsage--;
3629	pUsage->cUsage--;
3630	}
3631
3632	RTSemFastMutexRelease(pDevExt->mtxLdr);
3633	return rc;
3634	}
3635
3636
3637	/**
3638	* Gets the address of a symbol in an open image.
3639	*
3640	* @returns 0 on success.
3641	* @returns SUPDRV_ERR_* on failure.
3642	* @param pDevExt Device globals.
3643	* @param pSession Session data.
3644	* @param pReq The request buffer.
3645	*/
3646	static int supdrvIOCtl_LdrGetSymbol(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRGETSYMBOL pReq)
3647	{
3648	PSUPDRVLDRIMAGE pImage;
3649	PSUPDRVLDRUSAGE pUsage;
3650	uint32_t i;
3651	PSUPLDRSYM paSyms;
3652	const char *pchStrings;
3653	const size_t cbSymbol = strlen(pReq->u.In.szSymbol) + 1;
3654	void *pvSymbol = NULL;
3655	int rc = VERR_GENERAL_FAILURE;
3656	Log3(("supdrvIOCtl_LdrGetSymbol: pvImageBase=%p szSymbol=\"%s\"\n", pReq->u.In.pvImageBase, pReq->u.In.szSymbol));
3657
3658	/*
3659	* Find the ldr image.
3660	*/
3661	RTSemFastMutexRequest(pDevExt->mtxLdr);
3662	pUsage = pSession->pLdrUsage;
3663	while (pUsage && pUsage->pImage->pvImage != pReq->u.In.pvImageBase)
3664	pUsage = pUsage->pNext;
3665	if (!pUsage)
3666	{
3667	RTSemFastMutexRelease(pDevExt->mtxLdr);
3668	Log(("SUP_IOCTL_LDR_GET_SYMBOL: couldn't find image!\n"));
3669	return VERR_INVALID_HANDLE;
3670	}
3671	pImage = pUsage->pImage;
3672	if (pImage->uState != SUP_IOCTL_LDR_LOAD)
3673	{
3674	unsigned uState = pImage->uState;
3675	RTSemFastMutexRelease(pDevExt->mtxLdr);
3676	Log(("SUP_IOCTL_LDR_GET_SYMBOL: invalid image state %d (%#x)!\n", uState, uState)); NOREF(uState);
3677	return VERR_ALREADY_LOADED;
3678	}
3679
3680	/*
3681	* Search the symbol strings.
3682	*/
3683	pchStrings = (const char )((uint8_t )pImage->pvImage + pImage->offStrTab);
3684	paSyms = (PSUPLDRSYM)((uint8_t *)pImage->pvImage + pImage->offSymbols);
3685	for (i = 0; i < pImage->cSymbols; i++)
3686	{
3687	if ( paSyms[i].offSymbol < pImage->cbImage /* paranoia */
3688	&& paSyms[i].offName + cbSymbol <= pImage->cbStrTab
3689	&& !memcmp(pchStrings + paSyms[i].offName, pReq->u.In.szSymbol, cbSymbol))
3690	{
3691	pvSymbol = (uint8_t *)pImage->pvImage + paSyms[i].offSymbol;
3692	rc = VINF_SUCCESS;
3693	break;
3694	}
3695	}
3696	RTSemFastMutexRelease(pDevExt->mtxLdr);
3697	pReq->u.Out.pvSymbol = pvSymbol;
3698	return rc;
3699	}
3700
3701
3702	/**
3703	* Gets the address of a symbol in an open image or the support driver.
3704	*
3705	* @returns VINF_SUCCESS on success.
3706	* @returns
3707	* @param pDevExt Device globals.
3708	* @param pSession Session data.
3709	* @param pReq The request buffer.
3710	*/
3711	static int supdrvIDC_LdrGetSymbol(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPDRVIDCREQGETSYM pReq)
3712	{
3713	int rc = VINF_SUCCESS;
3714	const char *pszSymbol = pReq->u.In.pszSymbol;
3715	const char *pszModule = pReq->u.In.pszModule;
3716	size_t cbSymbol;
3717	char const *pszEnd;
3718	uint32_t i;
3719
3720	/*
3721	* Input validation.
3722	*/
3723	AssertPtrReturn(pszSymbol, VERR_INVALID_POINTER);
3724	pszEnd = (char *)memchr(pszSymbol, '\0', 512);
3725	AssertReturn(pszEnd, VERR_INVALID_PARAMETER);
3726	cbSymbol = pszEnd - pszSymbol + 1;
3727
3728	if (pszModule)
3729	{
3730	AssertPtrReturn(pszModule, VERR_INVALID_POINTER);
3731	pszEnd = (char *)memchr(pszModule, '\0', 64);
3732	AssertReturn(pszEnd, VERR_INVALID_PARAMETER);
3733	}
3734	Log3(("supdrvIDC_LdrGetSymbol: pszModule=%p:{%s} pszSymbol=%p:{%s}\n", pszModule, pszModule, pszSymbol, pszSymbol));
3735
3736
3737	if ( !pszModule
3738	\|\| !strcmp(pszModule, "SupDrv"))
3739	{
3740	/*
3741	* Search the support driver export table.
3742	*/
3743	for (i = 0; i < RT_ELEMENTS(g_aFunctions); i++)
3744	if (!strcmp(g_aFunctions[i].szName, pszSymbol))
3745	{
3746	pReq->u.Out.pfnSymbol = g_aFunctions[i].pfn;
3747	break;
3748	}
3749	}
3750	else
3751	{
3752	/*
3753	* Find the loader image.
3754	*/
3755	PSUPDRVLDRIMAGE pImage;
3756
3757	RTSemFastMutexRequest(pDevExt->mtxLdr);
3758
3759	for (pImage = pDevExt->pLdrImages; pImage; pImage = pImage->pNext)
3760	if (!strcmp(pImage->szName, pszModule))
3761	break;
3762	if (pImage && pImage->uState == SUP_IOCTL_LDR_LOAD)
3763	{
3764	/*
3765	* Search the symbol strings.
3766	*/
3767	const char pchStrings = (const char )((uint8_t *)pImage->pvImage + pImage->offStrTab);
3768	PCSUPLDRSYM paSyms = (PCSUPLDRSYM)((uint8_t *)pImage->pvImage + pImage->offSymbols);
3769	for (i = 0; i < pImage->cSymbols; i++)
3770	{
3771	if ( paSyms[i].offSymbol < pImage->cbImage /* paranoia */
3772	&& paSyms[i].offName + cbSymbol <= pImage->cbStrTab
3773	&& !memcmp(pchStrings + paSyms[i].offName, pszSymbol, cbSymbol))
3774	{
3775	/*
3776	* Found it! Calc the symbol address and add a reference to the module.
3777	*/
3778	pReq->u.Out.pfnSymbol = (PFNRT)((uint8_t *)pImage->pvImage + paSyms[i].offSymbol);
3779	rc = supdrvLdrAddUsage(pSession, pImage);
3780	break;
3781	}
3782	}
3783	}
3784	else
3785	rc = pImage ? VERR_WRONG_ORDER : VERR_MODULE_NOT_FOUND;
3786
3787	RTSemFastMutexRelease(pDevExt->mtxLdr);
3788	}
3789	return rc;
3790	}
3791
3792
3793	/**
3794	* Updates the VMMR0 entry point pointers.
3795	*
3796	* @returns IPRT status code.
3797	* @param pDevExt Device globals.
3798	* @param pSession Session data.
3799	* @param pVMMR0 VMMR0 image handle.
3800	* @param pvVMMR0EntryInt VMMR0EntryInt address.
3801	* @param pvVMMR0EntryFast VMMR0EntryFast address.
3802	* @param pvVMMR0EntryEx VMMR0EntryEx address.
3803	* @remark Caller must own the loader mutex.
3804	*/
3805	static int supdrvLdrSetVMMR0EPs(PSUPDRVDEVEXT pDevExt, void pvVMMR0, void pvVMMR0EntryInt, void pvVMMR0EntryFast, void pvVMMR0EntryEx)
3806	{
3807	int rc = VINF_SUCCESS;
3808	LogFlow(("supdrvLdrSetR0EP pvVMMR0=%p pvVMMR0EntryInt=%p\n", pvVMMR0, pvVMMR0EntryInt));
3809
3810
3811	/*
3812	* Check if not yet set.
3813	*/
3814	if (!pDevExt->pvVMMR0)
3815	{
3816	pDevExt->pvVMMR0 = pvVMMR0;
3817	pDevExt->pfnVMMR0EntryInt = pvVMMR0EntryInt;
3818	pDevExt->pfnVMMR0EntryFast = pvVMMR0EntryFast;
3819	pDevExt->pfnVMMR0EntryEx = pvVMMR0EntryEx;
3820	}
3821	else
3822	{
3823	/*
3824	* Return failure or success depending on whether the values match or not.
3825	*/
3826	if ( pDevExt->pvVMMR0 != pvVMMR0
3827	\|\| (void *)pDevExt->pfnVMMR0EntryInt != pvVMMR0EntryInt
3828	\|\| (void *)pDevExt->pfnVMMR0EntryFast != pvVMMR0EntryFast
3829	\|\| (void *)pDevExt->pfnVMMR0EntryEx != pvVMMR0EntryEx)
3830	{
3831	AssertMsgFailed(("SUP_IOCTL_LDR_SETR0EP: Already set pointing to a different module!\n"));
3832	rc = VERR_INVALID_PARAMETER;
3833	}
3834	}
3835	return rc;
3836	}
3837
3838
3839	/**
3840	* Unsets the VMMR0 entry point installed by supdrvLdrSetR0EP.
3841	*
3842	* @param pDevExt Device globals.
3843	*/
3844	static void supdrvLdrUnsetVMMR0EPs(PSUPDRVDEVEXT pDevExt)
3845	{
3846	pDevExt->pvVMMR0 = NULL;
3847	pDevExt->pfnVMMR0EntryInt = NULL;
3848	pDevExt->pfnVMMR0EntryFast = NULL;
3849	pDevExt->pfnVMMR0EntryEx = NULL;
3850	}
3851
3852
3853	/**
3854	* Adds a usage reference in the specified session of an image.
3855	*
3856	* Called while owning the loader semaphore.
3857	*
3858	* @returns VINF_SUCCESS on success and VERR_NO_MEMORY on failure.
3859	* @param pSession Session in question.
3860	* @param pImage Image which the session is using.
3861	*/
3862	static int supdrvLdrAddUsage(PSUPDRVSESSION pSession, PSUPDRVLDRIMAGE pImage)
3863	{
3864	PSUPDRVLDRUSAGE pUsage;
3865	LogFlow(("supdrvLdrAddUsage: pImage=%p\n", pImage));
3866
3867	/*
3868	* Referenced it already?
3869	*/
3870	pUsage = pSession->pLdrUsage;
3871	while (pUsage)
3872	{
3873	if (pUsage->pImage == pImage)
3874	{
3875	pUsage->cUsage++;
3876	return VINF_SUCCESS;
3877	}
3878	pUsage = pUsage->pNext;
3879	}
3880
3881	/*
3882	* Allocate new usage record.
3883	*/
3884	pUsage = (PSUPDRVLDRUSAGE)RTMemAlloc(sizeof(*pUsage));
3885	AssertReturn(pUsage, VERR_NO_MEMORY);
3886	pUsage->cUsage = 1;
3887	pUsage->pImage = pImage;
3888	pUsage->pNext = pSession->pLdrUsage;
3889	pSession->pLdrUsage = pUsage;
3890	return VINF_SUCCESS;
3891	}
3892
3893
3894	/**
3895	* Frees a load image.
3896	*
3897	* @param pDevExt Pointer to device extension.
3898	* @param pImage Pointer to the image we're gonna free.
3899	* This image must exit!
3900	* @remark The caller MUST own SUPDRVDEVEXT::mtxLdr!
3901	*/
3902	static void supdrvLdrFree(PSUPDRVDEVEXT pDevExt, PSUPDRVLDRIMAGE pImage)
3903	{
3904	PSUPDRVLDRIMAGE pImagePrev;
3905	LogFlow(("supdrvLdrFree: pImage=%p\n", pImage));
3906
3907	/* find it - arg. should've used doubly linked list. */
3908	Assert(pDevExt->pLdrImages);
3909	pImagePrev = NULL;
3910	if (pDevExt->pLdrImages != pImage)
3911	{
3912	pImagePrev = pDevExt->pLdrImages;
3913	while (pImagePrev->pNext != pImage)
3914	pImagePrev = pImagePrev->pNext;
3915	Assert(pImagePrev->pNext == pImage);
3916	}
3917
3918	/* unlink */
3919	if (pImagePrev)
3920	pImagePrev->pNext = pImage->pNext;
3921	else
3922	pDevExt->pLdrImages = pImage->pNext;
3923
3924	/* check if this is VMMR0.r0 unset its entry point pointers. */
3925	if (pDevExt->pvVMMR0 == pImage->pvImage)
3926	supdrvLdrUnsetVMMR0EPs(pDevExt);
3927
3928	/* check for objects with destructors in this image. (Shouldn't happen.) */
3929	if (pDevExt->pObjs)
3930	{
3931	unsigned cObjs = 0;
3932	PSUPDRVOBJ pObj;
3933	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
3934	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
3935	for (pObj = pDevExt->pObjs; pObj; pObj = pObj->pNext)
3936	if (RT_UNLIKELY((uintptr_t)pObj->pfnDestructor - (uintptr_t)pImage->pvImage < pImage->cbImage))
3937	{
3938	pObj->pfnDestructor = NULL;
3939	cObjs++;
3940	}
3941	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
3942	if (cObjs)
3943	OSDBGPRINT(("supdrvLdrFree: Image '%s' has %d dangling objects!\n", pImage->szName, cObjs));
3944	}
3945
3946	/* call termination function if fully loaded. */
3947	if ( pImage->pfnModuleTerm
3948	&& pImage->uState == SUP_IOCTL_LDR_LOAD)
3949	{
3950	LogFlow(("supdrvIOCtl_LdrLoad: calling pfnModuleTerm=%p\n", pImage->pfnModuleTerm));
3951	#ifdef RT_WITH_W64_UNWIND_HACK
3952	supdrvNtWrapModuleTerm(pImage->pfnModuleTerm);
3953	#else
3954	pImage->pfnModuleTerm();
3955	#endif
3956	}
3957
3958	/* free the image */
3959	pImage->cUsage = 0;
3960	pImage->pNext = 0;
3961	pImage->uState = SUP_IOCTL_LDR_FREE;
3962	RTMemExecFree(pImage);
3963	}
3964
3965
3966	/**
3967	* Implements the service call request.
3968	*
3969	* @returns VBox status code.
3970	* @param pDevExt The device extension.
3971	* @param pSession The calling session.
3972	* @param pReq The request packet, valid.
3973	*/
3974	static int supdrvIOCtl_CallServiceModule(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPCALLSERVICE pReq)
3975	{
3976	#if !defined(RT_OS_WINDOWS) \|\| defined(DEBUG)
3977	int rc;
3978
3979	/*
3980	* Find the module first in the module referenced by the calling session.
3981	*/
3982	rc = RTSemFastMutexRequest(pDevExt->mtxLdr);
3983	if (RT_SUCCESS(rc))
3984	{
3985	PFNSUPR0SERVICEREQHANDLER pfnServiceReqHandler = NULL;
3986	PSUPDRVLDRUSAGE pUsage;
3987
3988	for (pUsage = pSession->pLdrUsage; pUsage; pUsage = pUsage->pNext)
3989	if ( pUsage->pImage->pfnServiceReqHandler
3990	&& !strcmp(pUsage->pImage->szName, pReq->u.In.szName))
3991	{
3992	pfnServiceReqHandler = pUsage->pImage->pfnServiceReqHandler;
3993	break;
3994	}
3995	RTSemFastMutexRelease(pDevExt->mtxLdr);
3996
3997	if (pfnServiceReqHandler)
3998	{
3999	/*
4000	* Call it.
4001	*/
4002	if (pReq->Hdr.cbIn == SUP_IOCTL_CALL_SERVICE_SIZE(0))
4003	#ifdef RT_WITH_W64_UNWIND_HACK
4004	rc = supdrvNtWrapServiceReqHandler((PFNRT)pfnServiceReqHandler, pSession, pReq->u.In.uOperation, pReq->u.In.u64Arg, NULL);
4005	#else
4006	rc = pfnServiceReqHandler(pSession, pReq->u.In.uOperation, pReq->u.In.u64Arg, NULL);
4007	#endif
4008	else
4009	#ifdef RT_WITH_W64_UNWIND_HACK
4010	rc = supdrvNtWrapServiceReqHandler((PFNRT)pfnServiceReqHandler, pSession, pReq->u.In.uOperation,
4011	pReq->u.In.u64Arg, (PSUPR0SERVICEREQHDR)&pReq->abReqPkt[0]);
4012	#else
4013	rc = pfnServiceReqHandler(pSession, pReq->u.In.uOperation, pReq->u.In.u64Arg, (PSUPR0SERVICEREQHDR)&pReq->abReqPkt[0]);
4014	#endif
4015	}
4016	else
4017	rc = VERR_SUPDRV_SERVICE_NOT_FOUND;
4018	}
4019
4020	/* log it */
4021	if ( RT_FAILURE(rc)
4022	&& rc != VERR_INTERRUPTED
4023	&& rc != VERR_TIMEOUT)
4024	Log(("SUP_IOCTL_CALL_SERVICE: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
4025	rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
4026	else
4027	Log4(("SUP_IOCTL_CALL_SERVICE: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
4028	rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
4029	return rc;
4030	#else /* RT_OS_WINDOWS && !DEBUG */
4031	return VERR_NOT_IMPLEMENTED;
4032	#endif /* RT_OS_WINDOWS && !DEBUG */
4033	}
4034
4035
4036	/**
4037	* Gets the paging mode of the current CPU.
4038	*
4039	* @returns Paging mode, SUPPAGEINGMODE_INVALID on error.
4040	*/
4041	SUPR0DECL(SUPPAGINGMODE) SUPR0GetPagingMode(void)
4042	{
4043	SUPPAGINGMODE enmMode;
4044
4045	RTR0UINTREG cr0 = ASMGetCR0();
4046	if ((cr0 & (X86_CR0_PG \| X86_CR0_PE)) != (X86_CR0_PG \| X86_CR0_PE))
4047	enmMode = SUPPAGINGMODE_INVALID;
4048	else
4049	{
4050	RTR0UINTREG cr4 = ASMGetCR4();
4051	uint32_t fNXEPlusLMA = 0;
4052	if (cr4 & X86_CR4_PAE)
4053	{
4054	uint32_t fAmdFeatures = ASMCpuId_EDX(0x80000001);
4055	if (fAmdFeatures & (X86_CPUID_AMD_FEATURE_EDX_NX \| X86_CPUID_AMD_FEATURE_EDX_LONG_MODE))
4056	{
4057	uint64_t efer = ASMRdMsr(MSR_K6_EFER);
4058	if ((fAmdFeatures & X86_CPUID_AMD_FEATURE_EDX_NX) && (efer & MSR_K6_EFER_NXE))
4059	fNXEPlusLMA \|= RT_BIT(0);
4060	if ((fAmdFeatures & X86_CPUID_AMD_FEATURE_EDX_LONG_MODE) && (efer & MSR_K6_EFER_LMA))
4061	fNXEPlusLMA \|= RT_BIT(1);
4062	}
4063	}
4064
4065	switch ((cr4 & (X86_CR4_PAE \| X86_CR4_PGE)) \| fNXEPlusLMA)
4066	{
4067	case 0:
4068	enmMode = SUPPAGINGMODE_32_BIT;
4069	break;
4070
4071	case X86_CR4_PGE:
4072	enmMode = SUPPAGINGMODE_32_BIT_GLOBAL;
4073	break;
4074
4075	case X86_CR4_PAE:
4076	enmMode = SUPPAGINGMODE_PAE;
4077	break;
4078
4079	case X86_CR4_PAE \| RT_BIT(0):
4080	enmMode = SUPPAGINGMODE_PAE_NX;
4081	break;
4082
4083	case X86_CR4_PAE \| X86_CR4_PGE:
4084	enmMode = SUPPAGINGMODE_PAE_GLOBAL;
4085	break;
4086
4087	case X86_CR4_PAE \| X86_CR4_PGE \| RT_BIT(0):
4088	enmMode = SUPPAGINGMODE_PAE_GLOBAL;
4089	break;
4090
4091	case RT_BIT(1) \| X86_CR4_PAE:
4092	enmMode = SUPPAGINGMODE_AMD64;
4093	break;
4094
4095	case RT_BIT(1) \| X86_CR4_PAE \| RT_BIT(0):
4096	enmMode = SUPPAGINGMODE_AMD64_NX;
4097	break;
4098
4099	case RT_BIT(1) \| X86_CR4_PAE \| X86_CR4_PGE:
4100	enmMode = SUPPAGINGMODE_AMD64_GLOBAL;
4101	break;
4102
4103	case RT_BIT(1) \| X86_CR4_PAE \| X86_CR4_PGE \| RT_BIT(0):
4104	enmMode = SUPPAGINGMODE_AMD64_GLOBAL_NX;
4105	break;
4106
4107	default:
4108	AssertMsgFailed(("Cannot happen! cr4=%#x fNXEPlusLMA=%d\n", cr4, fNXEPlusLMA));
4109	enmMode = SUPPAGINGMODE_INVALID;
4110	break;
4111	}
4112	}
4113	return enmMode;
4114	}
4115
4116
4117	/**
4118	* Enables or disabled hardware virtualization extensions using native OS APIs.
4119	*
4120	* @returns VBox status code.
4121	* @retval VINF_SUCCESS on success.
4122	* @retval VERR_NOT_SUPPORTED if not supported by the native OS.
4123	*
4124	* @param fEnable Whether to enable or disable.
4125	*/
4126	SUPR0DECL(int) SUPR0EnableVTx(bool fEnable)
4127	{
4128	#ifdef RT_OS_DARWIN
4129	return supdrvOSEnableVTx(fEnable);
4130	#else
4131	return VERR_NOT_SUPPORTED;
4132	#endif
4133	}
4134
4135
4136	/**
4137	* Creates the GIP.
4138	*
4139	* @returns VBox status code.
4140	* @param pDevExt Instance data. GIP stuff may be updated.
4141	*/
4142	static int supdrvGipCreate(PSUPDRVDEVEXT pDevExt)
4143	{
4144	PSUPGLOBALINFOPAGE pGip;
4145	RTHCPHYS HCPhysGip;
4146	uint32_t u32SystemResolution;
4147	uint32_t u32Interval;
4148	int rc;
4149
4150	LogFlow(("supdrvGipCreate:\n"));
4151
4152	/* assert order */
4153	Assert(pDevExt->u32SystemTimerGranularityGrant == 0);
4154	Assert(pDevExt->GipMemObj == NIL_RTR0MEMOBJ);
4155	Assert(!pDevExt->pGipTimer);
4156
4157	/*
4158	* Allocate a suitable page with a default kernel mapping.
4159	*/
4160	rc = RTR0MemObjAllocLow(&pDevExt->GipMemObj, PAGE_SIZE, false);
4161	if (RT_FAILURE(rc))
4162	{
4163	OSDBGPRINT(("supdrvGipCreate: failed to allocate the GIP page. rc=%d\n", rc));
4164	return rc;
4165	}
4166	pGip = (PSUPGLOBALINFOPAGE)RTR0MemObjAddress(pDevExt->GipMemObj); AssertPtr(pGip);
4167	HCPhysGip = RTR0MemObjGetPagePhysAddr(pDevExt->GipMemObj, 0); Assert(HCPhysGip != NIL_RTHCPHYS);
4168
4169	#if 0 /** @todo Disabled this as we didn't used to do it before and causes unnecessary stress on laptops.
4170	* It only applies to Windows and should probably revisited later, if possible made part of the
4171	* timer code (return min granularity in RTTimerGetSystemGranularity and set it in RTTimerStart). */
4172	/*
4173	* Try bump up the system timer resolution.
4174	* The more interrupts the better...
4175	*/
4176	if ( RT_SUCCESS(RTTimerRequestSystemGranularity( 488281 /* 2048 HZ */, &u32SystemResolution))
4177	\|\| RT_SUCCESS(RTTimerRequestSystemGranularity( 500000 /* 2000 HZ */, &u32SystemResolution))
4178	\|\| RT_SUCCESS(RTTimerRequestSystemGranularity( 976563 /* 1024 HZ */, &u32SystemResolution))
4179	\|\| RT_SUCCESS(RTTimerRequestSystemGranularity( 1000000 /* 1000 HZ */, &u32SystemResolution))
4180	\|\| RT_SUCCESS(RTTimerRequestSystemGranularity( 1953125 /* 512 HZ */, &u32SystemResolution))
4181	\|\| RT_SUCCESS(RTTimerRequestSystemGranularity( 2000000 /* 500 HZ */, &u32SystemResolution))
4182	\|\| RT_SUCCESS(RTTimerRequestSystemGranularity( 3906250 /* 256 HZ */, &u32SystemResolution))
4183	\|\| RT_SUCCESS(RTTimerRequestSystemGranularity( 4000000 /* 250 HZ */, &u32SystemResolution))
4184	\|\| RT_SUCCESS(RTTimerRequestSystemGranularity( 7812500 /* 128 HZ */, &u32SystemResolution))
4185	\|\| RT_SUCCESS(RTTimerRequestSystemGranularity(10000000 /* 100 HZ */, &u32SystemResolution))
4186	\|\| RT_SUCCESS(RTTimerRequestSystemGranularity(15625000 /* 64 HZ */, &u32SystemResolution))
4187	\|\| RT_SUCCESS(RTTimerRequestSystemGranularity(31250000 /* 32 HZ */, &u32SystemResolution))
4188	)
4189	{
4190	Assert(RTTimerGetSystemGranularity() <= u32SystemResolution);
4191	pDevExt->u32SystemTimerGranularityGrant = u32SystemResolution;
4192	}
4193	#endif
4194
4195	/*
4196	* Find a reasonable update interval and initialize the structure.
4197	*/
4198	u32Interval = u32SystemResolution = RTTimerGetSystemGranularity();
4199	while (u32Interval < 10000000 /* 10 ms */)
4200	u32Interval += u32SystemResolution;
4201
4202	supdrvGipInit(pDevExt, pGip, HCPhysGip, RTTimeSystemNanoTS(), 1000000000 / u32Interval /=Hz/);
4203
4204	/*
4205	* Create the timer.
4206	* If CPU_ALL isn't supported we'll have to fall back to synchronous mode.
4207	*/
4208	if (pGip->u32Mode == SUPGIPMODE_ASYNC_TSC)
4209	{
4210	rc = RTTimerCreateEx(&pDevExt->pGipTimer, u32Interval, RTTIMER_FLAGS_CPU_ALL, supdrvGipAsyncTimer, pDevExt);
4211	if (rc == VERR_NOT_SUPPORTED)
4212	{
4213	OSDBGPRINT(("supdrvGipCreate: omni timer not supported, falling back to synchronous mode\n"));
4214	pGip->u32Mode = SUPGIPMODE_SYNC_TSC;
4215	}
4216	}
4217	if (pGip->u32Mode != SUPGIPMODE_ASYNC_TSC)
4218	rc = RTTimerCreateEx(&pDevExt->pGipTimer, u32Interval, 0, supdrvGipSyncTimer, pDevExt);
4219	if (RT_SUCCESS(rc))
4220	{
4221	if (pGip->u32Mode == SUPGIPMODE_ASYNC_TSC)
4222	rc = RTMpNotificationRegister(supdrvGipMpEvent, pDevExt);
4223	if (RT_SUCCESS(rc))
4224	{
4225	/*
4226	* We're good.
4227	*/
4228	dprintf(("supdrvGipCreate: %ld ns interval.\n", (long)u32Interval));
4229	return VINF_SUCCESS;
4230	}
4231
4232	OSDBGPRINT(("supdrvGipCreate: failed register MP event notfication. rc=%d\n", rc));
4233	}
4234	else
4235	{
4236	OSDBGPRINT(("supdrvGipCreate: failed create GIP timer at %ld ns interval. rc=%d\n", (long)u32Interval, rc));
4237	Assert(!pDevExt->pGipTimer);
4238	}
4239	supdrvGipDestroy(pDevExt);
4240	return rc;
4241	}
4242
4243
4244	/**
4245	* Terminates the GIP.
4246	*
4247	* @param pDevExt Instance data. GIP stuff may be updated.
4248	*/
4249	static void supdrvGipDestroy(PSUPDRVDEVEXT pDevExt)
4250	{
4251	int rc;
4252	#ifdef DEBUG_DARWIN_GIP
4253	OSDBGPRINT(("supdrvGipDestroy: pDevExt=%p pGip=%p pGipTimer=%p GipMemObj=%p\n", pDevExt,
4254	pDevExt->GipMemObj != NIL_RTR0MEMOBJ ? RTR0MemObjAddress(pDevExt->GipMemObj) : NULL,
4255	pDevExt->pGipTimer, pDevExt->GipMemObj));
4256	#endif
4257
4258	/*
4259	* Invalid the GIP data.
4260	*/
4261	if (pDevExt->pGip)
4262	{
4263	supdrvGipTerm(pDevExt->pGip);
4264	pDevExt->pGip = NULL;
4265	}
4266
4267	/*
4268	* Destroy the timer and free the GIP memory object.
4269	*/
4270	if (pDevExt->pGipTimer)
4271	{
4272	rc = RTTimerDestroy(pDevExt->pGipTimer); AssertRC(rc);
4273	pDevExt->pGipTimer = NULL;
4274	}
4275
4276	if (pDevExt->GipMemObj != NIL_RTR0MEMOBJ)
4277	{
4278	rc = RTR0MemObjFree(pDevExt->GipMemObj, true /* free mappings */); AssertRC(rc);
4279	pDevExt->GipMemObj = NIL_RTR0MEMOBJ;
4280	}
4281
4282	/*
4283	* Finally, release the system timer resolution request if one succeeded.
4284	*/
4285	if (pDevExt->u32SystemTimerGranularityGrant)
4286	{
4287	rc = RTTimerReleaseSystemGranularity(pDevExt->u32SystemTimerGranularityGrant); AssertRC(rc);
4288	pDevExt->u32SystemTimerGranularityGrant = 0;
4289	}
4290	}
4291
4292
4293	/**
4294	* Timer callback function sync GIP mode.
4295	* @param pTimer The timer.
4296	* @param pvUser The device extension.
4297	*/
4298	static DECLCALLBACK(void) supdrvGipSyncTimer(PRTTIMER pTimer, void *pvUser, uint64_t iTick)
4299	{
4300	RTCCUINTREG fOldFlags = ASMIntDisableFlags(); /* No interruptions please (real problem on S10). */
4301	PSUPDRVDEVEXT pDevExt = (PSUPDRVDEVEXT)pvUser;
4302
4303	supdrvGipUpdate(pDevExt->pGip, RTTimeSystemNanoTS());
4304
4305	ASMSetFlags(fOldFlags);
4306	}
4307
4308
4309	/**
4310	* Timer callback function for async GIP mode.
4311	* @param pTimer The timer.
4312	* @param pvUser The device extension.
4313	*/
4314	static DECLCALLBACK(void) supdrvGipAsyncTimer(PRTTIMER pTimer, void *pvUser, uint64_t iTick)
4315	{
4316	RTCCUINTREG fOldFlags = ASMIntDisableFlags(); /* No interruptions please (real problem on S10). */
4317	PSUPDRVDEVEXT pDevExt = (PSUPDRVDEVEXT)pvUser;
4318	RTCPUID idCpu = RTMpCpuId();
4319	uint64_t NanoTS = RTTimeSystemNanoTS();
4320
4321	/** @todo reset the transaction number and whatnot when iTick == 1. */
4322	if (pDevExt->idGipMaster == idCpu)
4323	supdrvGipUpdate(pDevExt->pGip, NanoTS);
4324	else
4325	supdrvGipUpdatePerCpu(pDevExt->pGip, NanoTS, ASMGetApicId());
4326
4327	ASMSetFlags(fOldFlags);
4328	}
4329
4330
4331	/**
4332	* Multiprocessor event notification callback.
4333	*
4334	* This is used to make sue that the GIP master gets passed on to
4335	* another CPU.
4336	*
4337	* @param enmEvent The event.
4338	* @param idCpu The cpu it applies to.
4339	* @param pvUser Pointer to the device extension.
4340	*/
4341	static DECLCALLBACK(void) supdrvGipMpEvent(RTMPEVENT enmEvent, RTCPUID idCpu, void *pvUser)
4342	{
4343	PSUPDRVDEVEXT pDevExt = (PSUPDRVDEVEXT)pvUser;
4344	if (enmEvent == RTMPEVENT_OFFLINE)
4345	{
4346	RTCPUID idGipMaster;
4347	ASMAtomicReadSize(&pDevExt->idGipMaster, &idGipMaster);
4348	if (idGipMaster == idCpu)
4349	{
4350	/*
4351	* Find a new GIP master.
4352	*/
4353	bool fIgnored;
4354	unsigned i;
4355	RTCPUID idNewGipMaster = NIL_RTCPUID;
4356	RTCPUSET OnlineCpus;
4357	RTMpGetOnlineSet(&OnlineCpus);
4358
4359	for (i = 0; i < RTCPUSET_MAX_CPUS; i++)
4360	{
4361	RTCPUID idCurCpu = RTMpCpuIdFromSetIndex(i);
4362	if ( RTCpuSetIsMember(&OnlineCpus, idCurCpu)
4363	&& idCurCpu != idGipMaster)
4364	{
4365	idNewGipMaster = idCurCpu;
4366	break;
4367	}
4368	}
4369
4370	dprintf(("supdrvGipMpEvent: Gip master %#lx -> %#lx\n", (long)idGipMaster, (long)idNewGipMaster));
4371	ASMAtomicCmpXchgSize(&pDevExt->idGipMaster, idNewGipMaster, idGipMaster, fIgnored);
4372	NOREF(fIgnored);
4373	}
4374	}
4375	}
4376
4377
4378	/**
4379	* Initializes the GIP data.
4380	*
4381	* @returns IPRT status code.
4382	* @param pDevExt Pointer to the device instance data.
4383	* @param pGip Pointer to the read-write kernel mapping of the GIP.
4384	* @param HCPhys The physical address of the GIP.
4385	* @param u64NanoTS The current nanosecond timestamp.
4386	* @param uUpdateHz The update freqence.
4387	*/
4388	int VBOXCALL supdrvGipInit(PSUPDRVDEVEXT pDevExt, PSUPGLOBALINFOPAGE pGip, RTHCPHYS HCPhys, uint64_t u64NanoTS, unsigned uUpdateHz)
4389	{
4390	unsigned i;
4391	#ifdef DEBUG_DARWIN_GIP
4392	OSDBGPRINT(("supdrvGipInit: pGip=%p HCPhys=%lx u64NanoTS=%llu uUpdateHz=%d\n", pGip, (long)HCPhys, u64NanoTS, uUpdateHz));
4393	#else
4394	LogFlow(("supdrvGipInit: pGip=%p HCPhys=%lx u64NanoTS=%llu uUpdateHz=%d\n", pGip, (long)HCPhys, u64NanoTS, uUpdateHz));
4395	#endif
4396
4397	/*
4398	* Initialize the structure.
4399	*/
4400	memset(pGip, 0, PAGE_SIZE);
4401	pGip->u32Magic = SUPGLOBALINFOPAGE_MAGIC;
4402	pGip->u32Version = SUPGLOBALINFOPAGE_VERSION;
4403	pGip->u32Mode = supdrvGipDeterminTscMode(pDevExt);
4404	pGip->u32UpdateHz = uUpdateHz;
4405	pGip->u32UpdateIntervalNS = 1000000000 / uUpdateHz;
4406	pGip->u64NanoTSLastUpdateHz = u64NanoTS;
4407
4408	for (i = 0; i < RT_ELEMENTS(pGip->aCPUs); i++)
4409	{
4410	pGip->aCPUs[i].u32TransactionId = 2;
4411	pGip->aCPUs[i].u64NanoTS = u64NanoTS;
4412	pGip->aCPUs[i].u64TSC = ASMReadTSC();
4413
4414	/*
4415	* We don't know the following values until we've executed updates.
4416	* So, we'll just insert very high values.
4417	*/
4418	pGip->aCPUs[i].u64CpuHz = _4G + 1;
4419	pGip->aCPUs[i].u32UpdateIntervalTSC = _2G / 4;
4420	pGip->aCPUs[i].au32TSCHistory[0] = _2G / 4;
4421	pGip->aCPUs[i].au32TSCHistory[1] = _2G / 4;
4422	pGip->aCPUs[i].au32TSCHistory[2] = _2G / 4;
4423	pGip->aCPUs[i].au32TSCHistory[3] = _2G / 4;
4424	pGip->aCPUs[i].au32TSCHistory[4] = _2G / 4;
4425	pGip->aCPUs[i].au32TSCHistory[5] = _2G / 4;
4426	pGip->aCPUs[i].au32TSCHistory[6] = _2G / 4;
4427	pGip->aCPUs[i].au32TSCHistory[7] = _2G / 4;
4428	}
4429
4430	/*
4431	* Link it to the device extension.
4432	*/
4433	pDevExt->pGip = pGip;
4434	pDevExt->HCPhysGip = HCPhys;
4435	pDevExt->cGipUsers = 0;
4436
4437	return VINF_SUCCESS;
4438	}
4439
4440
4441	/**
4442	* Callback used by supdrvDetermineAsyncTSC to read the TSC on a CPU.
4443	*
4444	* @param idCpu Ignored.
4445	* @param pvUser1 Where to put the TSC.
4446	* @param pvUser2 Ignored.
4447	*/
4448	static DECLCALLBACK(void) supdrvDetermineAsyncTscWorker(RTCPUID idCpu, void pvUser1, void pvUser2)
4449	{
4450	#if 1
4451	ASMAtomicWriteU64((uint64_t volatile *)pvUser1, ASMReadTSC());
4452	#else
4453	(uint64_t )pvUser1 = ASMReadTSC();
4454	#endif
4455	}
4456
4457
4458	/**
4459	* Determine if Async GIP mode is required because of TSC drift.
4460	*
4461	* When using the default/normal timer code it is essential that the time stamp counter
4462	* (TSC) runs never backwards, that is, a read operation to the counter should return
4463	* a bigger value than any previous read operation. This is guaranteed by the latest
4464	* AMD CPUs and by newer Intel CPUs which never enter the C2 state (P4). In any other
4465	* case we have to choose the asynchronous timer mode.
4466	*
4467	* @param poffMin Pointer to the determined difference between different cores.
4468	* @return false if the time stamp counters appear to be synchron, true otherwise.
4469	*/
4470	bool VBOXCALL supdrvDetermineAsyncTsc(uint64_t *poffMin)
4471	{
4472	/*
4473	* Just iterate all the cpus 8 times and make sure that the TSC is
4474	* ever increasing. We don't bother taking TSC rollover into account.
4475	*/
4476	RTCPUSET CpuSet;
4477	int iLastCpu = RTCpuLastIndex(RTMpGetSet(&CpuSet));
4478	int iCpu;
4479	int cLoops = 8;
4480	bool fAsync = false;
4481	int rc = VINF_SUCCESS;
4482	uint64_t offMax = 0;
4483	uint64_t offMin = ~(uint64_t)0;
4484	uint64_t PrevTsc = ASMReadTSC();
4485
4486	while (cLoops-- > 0)
4487	{
4488	for (iCpu = 0; iCpu <= iLastCpu; iCpu++)
4489	{
4490	uint64_t CurTsc;
4491	rc = RTMpOnSpecific(RTMpCpuIdFromSetIndex(iCpu), supdrvDetermineAsyncTscWorker, &CurTsc, NULL);
4492	if (RT_SUCCESS(rc))
4493	{
4494	if (CurTsc <= PrevTsc)
4495	{
4496	fAsync = true;
4497	offMin = offMax = PrevTsc - CurTsc;
4498	dprintf(("supdrvDetermineAsyncTsc: iCpu=%d cLoops=%d CurTsc=%llx PrevTsc=%llx\n",
4499	iCpu, cLoops, CurTsc, PrevTsc));
4500	break;
4501	}
4502
4503	/* Gather statistics (except the first time). */
4504	if (iCpu != 0 \|\| cLoops != 7)
4505	{
4506	uint64_t off = CurTsc - PrevTsc;
4507	if (off < offMin)
4508	offMin = off;
4509	if (off > offMax)
4510	offMax = off;
4511	dprintf2(("%d/%d: off=%llx\n", cLoops, iCpu, off));
4512	}
4513
4514	/* Next */
4515	PrevTsc = CurTsc;
4516	}
4517	else if (rc == VERR_NOT_SUPPORTED)
4518	break;
4519	else
4520	AssertMsg(rc == VERR_CPU_NOT_FOUND \|\| rc == VERR_CPU_OFFLINE, ("%d\n", rc));
4521	}
4522
4523	/* broke out of the loop. */
4524	if (iCpu <= iLastCpu)
4525	break;
4526	}
4527
4528	poffMin = offMin; / Almost RTMpOnSpecific profiling. */
4529	dprintf(("supdrvDetermineAsyncTsc: returns %d; iLastCpu=%d rc=%d offMin=%llx offMax=%llx\n",
4530	fAsync, iLastCpu, rc, offMin, offMax));
4531	#if !defined(RT_OS_SOLARIS) && !defined(RT_OS_OS2) && !defined(RT_OS_WINDOWS)
4532	OSDBGPRINT(("vboxdrv: fAsync=%d offMin=%#lx offMax=%#lx\n", fAsync, (long)offMin, (long)offMax));
4533	#endif
4534	return fAsync;
4535	}
4536
4537
4538	/**
4539	* Determin the GIP TSC mode.
4540	*
4541	* @returns The most suitable TSC mode.
4542	* @param pDevExt Pointer to the device instance data.
4543	*/
4544	static SUPGIPMODE supdrvGipDeterminTscMode(PSUPDRVDEVEXT pDevExt)
4545	{
4546	/*
4547	* On SMP we're faced with two problems:
4548	* (1) There might be a skew between the CPU, so that cpu0
4549	* returns a TSC that is sligtly different from cpu1.
4550	* (2) Power management (and other things) may cause the TSC
4551	* to run at a non-constant speed, and cause the speed
4552	* to be different on the cpus. This will result in (1).
4553	*
4554	* So, on SMP systems we'll have to select the ASYNC update method
4555	* if there are symphoms of these problems.
4556	*/
4557	if (RTMpGetCount() > 1)
4558	{
4559	uint32_t uEAX, uEBX, uECX, uEDX;
4560	uint64_t u64DiffCoresIgnored;
4561
4562	/* Permit the user and/or the OS specfic bits to force async mode. */
4563	if (supdrvOSGetForcedAsyncTscMode(pDevExt))
4564	return SUPGIPMODE_ASYNC_TSC;
4565
4566	/* Try check for current differences between the cpus. */
4567	if (supdrvDetermineAsyncTsc(&u64DiffCoresIgnored))
4568	return SUPGIPMODE_ASYNC_TSC;
4569
4570	/*
4571	* If the CPU supports power management and is an AMD one we
4572	* won't trust it unless it has the TscInvariant bit is set.
4573	*/
4574	/* Check for "AuthenticAMD" */
4575	ASMCpuId(0, &uEAX, &uEBX, &uECX, &uEDX);
4576	if ( uEAX >= 1
4577	&& uEBX == X86_CPUID_VENDOR_AMD_EBX
4578	&& uECX == X86_CPUID_VENDOR_AMD_ECX
4579	&& uEDX == X86_CPUID_VENDOR_AMD_EDX)
4580	{
4581	/* Check for APM support and that TscInvariant is cleared. */
4582	ASMCpuId(0x80000000, &uEAX, &uEBX, &uECX, &uEDX);
4583	if (uEAX >= 0x80000007)
4584	{
4585	ASMCpuId(0x80000007, &uEAX, &uEBX, &uECX, &uEDX);
4586	if ( !(uEDX & RT_BIT(8))/* TscInvariant */
4587	&& (uEDX & 0x3e)) /* STC\|TM\|THERMTRIP\|VID\|FID. Ignore TS. */
4588	return SUPGIPMODE_ASYNC_TSC;
4589	}
4590	}
4591	}
4592	return SUPGIPMODE_SYNC_TSC;
4593	}
4594
4595
4596	/**
4597	* Invalidates the GIP data upon termination.
4598	*
4599	* @param pGip Pointer to the read-write kernel mapping of the GIP.
4600	*/
4601	void VBOXCALL supdrvGipTerm(PSUPGLOBALINFOPAGE pGip)
4602	{
4603	unsigned i;
4604	pGip->u32Magic = 0;
4605	for (i = 0; i < RT_ELEMENTS(pGip->aCPUs); i++)
4606	{
4607	pGip->aCPUs[i].u64NanoTS = 0;
4608	pGip->aCPUs[i].u64TSC = 0;
4609	pGip->aCPUs[i].iTSCHistoryHead = 0;
4610	}
4611	}
4612
4613
4614	/**
4615	* Worker routine for supdrvGipUpdate and supdrvGipUpdatePerCpu that
4616	* updates all the per cpu data except the transaction id.
4617	*
4618	* @param pGip The GIP.
4619	* @param pGipCpu Pointer to the per cpu data.
4620	* @param u64NanoTS The current time stamp.
4621	*/
4622	static void supdrvGipDoUpdateCpu(PSUPGLOBALINFOPAGE pGip, PSUPGIPCPU pGipCpu, uint64_t u64NanoTS)
4623	{
4624	uint64_t u64TSC;
4625	uint64_t u64TSCDelta;
4626	uint32_t u32UpdateIntervalTSC;
4627	uint32_t u32UpdateIntervalTSCSlack;
4628	unsigned iTSCHistoryHead;
4629	uint64_t u64CpuHz;
4630
4631	/*
4632	* Update the NanoTS.
4633	*/
4634	ASMAtomicXchgU64(&pGipCpu->u64NanoTS, u64NanoTS);
4635
4636	/*
4637	* Calc TSC delta.
4638	*/
4639	/** @todo validate the NanoTS delta, don't trust the OS to call us when it should... */
4640	u64TSC = ASMReadTSC();
4641	u64TSCDelta = u64TSC - pGipCpu->u64TSC;
4642	ASMAtomicXchgU64(&pGipCpu->u64TSC, u64TSC);
4643
4644	if (u64TSCDelta >> 32)
4645	{
4646	u64TSCDelta = pGipCpu->u32UpdateIntervalTSC;
4647	pGipCpu->cErrors++;
4648	}
4649
4650	/*
4651	* TSC History.
4652	*/
4653	Assert(RT_ELEMENTS(pGipCpu->au32TSCHistory) == 8);
4654
4655	iTSCHistoryHead = (pGipCpu->iTSCHistoryHead + 1) & 7;
4656	ASMAtomicXchgU32(&pGipCpu->iTSCHistoryHead, iTSCHistoryHead);
4657	ASMAtomicXchgU32(&pGipCpu->au32TSCHistory[iTSCHistoryHead], (uint32_t)u64TSCDelta);
4658
4659	/*
4660	* UpdateIntervalTSC = average of last 8,2,1 intervals depending on update HZ.
4661	*/
4662	if (pGip->u32UpdateHz >= 1000)
4663	{
4664	uint32_t u32;
4665	u32 = pGipCpu->au32TSCHistory[0];
4666	u32 += pGipCpu->au32TSCHistory[1];
4667	u32 += pGipCpu->au32TSCHistory[2];
4668	u32 += pGipCpu->au32TSCHistory[3];
4669	u32 >>= 2;
4670	u32UpdateIntervalTSC = pGipCpu->au32TSCHistory[4];
4671	u32UpdateIntervalTSC += pGipCpu->au32TSCHistory[5];
4672	u32UpdateIntervalTSC += pGipCpu->au32TSCHistory[6];
4673	u32UpdateIntervalTSC += pGipCpu->au32TSCHistory[7];
4674	u32UpdateIntervalTSC >>= 2;
4675	u32UpdateIntervalTSC += u32;
4676	u32UpdateIntervalTSC >>= 1;
4677
4678	/* Value choosen for a 2GHz Athlon64 running linux 2.6.10/11, . */
4679	u32UpdateIntervalTSCSlack = u32UpdateIntervalTSC >> 14;
4680	}
4681	else if (pGip->u32UpdateHz >= 90)
4682	{
4683	u32UpdateIntervalTSC = (uint32_t)u64TSCDelta;
4684	u32UpdateIntervalTSC += pGipCpu->au32TSCHistory[(iTSCHistoryHead - 1) & 7];
4685	u32UpdateIntervalTSC >>= 1;
4686
4687	/* value choosen on a 2GHz thinkpad running windows */
4688	u32UpdateIntervalTSCSlack = u32UpdateIntervalTSC >> 7;
4689	}
4690	else
4691	{
4692	u32UpdateIntervalTSC = (uint32_t)u64TSCDelta;
4693
4694	/* This value hasn't be checked yet.. waiting for OS/2 and 33Hz timers.. :-) */
4695	u32UpdateIntervalTSCSlack = u32UpdateIntervalTSC >> 6;
4696	}
4697	ASMAtomicXchgU32(&pGipCpu->u32UpdateIntervalTSC, u32UpdateIntervalTSC + u32UpdateIntervalTSCSlack);
4698
4699	/*
4700	* CpuHz.
4701	*/
4702	u64CpuHz = ASMMult2xU32RetU64(u32UpdateIntervalTSC, pGip->u32UpdateHz);
4703	ASMAtomicXchgU64(&pGipCpu->u64CpuHz, u64CpuHz);
4704	}
4705
4706
4707	/**
4708	* Updates the GIP.
4709	*
4710	* @param pGip Pointer to the GIP.
4711	* @param u64NanoTS The current nanosecond timesamp.
4712	*/
4713	void VBOXCALL supdrvGipUpdate(PSUPGLOBALINFOPAGE pGip, uint64_t u64NanoTS)
4714	{
4715	/*
4716	* Determin the relevant CPU data.
4717	*/
4718	PSUPGIPCPU pGipCpu;
4719	if (pGip->u32Mode != SUPGIPMODE_ASYNC_TSC)
4720	pGipCpu = &pGip->aCPUs[0];
4721	else
4722	{
4723	unsigned iCpu = ASMGetApicId();
4724	if (RT_LIKELY(iCpu >= RT_ELEMENTS(pGip->aCPUs)))
4725	return;
4726	pGipCpu = &pGip->aCPUs[iCpu];
4727	}
4728
4729	/*
4730	* Start update transaction.
4731	*/
4732	if (!(ASMAtomicIncU32(&pGipCpu->u32TransactionId) & 1))
4733	{
4734	/* this can happen on win32 if we're taking to long and there are more CPUs around. shouldn't happen though. */
4735	AssertMsgFailed(("Invalid transaction id, %#x, not odd!\n", pGipCpu->u32TransactionId));
4736	ASMAtomicIncU32(&pGipCpu->u32TransactionId);
4737	pGipCpu->cErrors++;
4738	return;
4739	}
4740
4741	/*
4742	* Recalc the update frequency every 0x800th time.
4743	*/
4744	if (!(pGipCpu->u32TransactionId & (GIP_UPDATEHZ_RECALC_FREQ * 2 - 2)))
4745	{
4746	if (pGip->u64NanoTSLastUpdateHz)
4747	{
4748	#ifdef RT_ARCH_AMD64 /** @todo fix 64-bit div here to work on x86 linux. */
4749	uint64_t u64Delta = u64NanoTS - pGip->u64NanoTSLastUpdateHz;
4750	uint32_t u32UpdateHz = (uint32_t)((UINT64_C(1000000000) * GIP_UPDATEHZ_RECALC_FREQ) / u64Delta);
4751	if (u32UpdateHz <= 2000 && u32UpdateHz >= 30)
4752	{
4753	ASMAtomicXchgU32(&pGip->u32UpdateHz, u32UpdateHz);
4754	ASMAtomicXchgU32(&pGip->u32UpdateIntervalNS, 1000000000 / u32UpdateHz);
4755	}
4756	#endif
4757	}
4758	ASMAtomicXchgU64(&pGip->u64NanoTSLastUpdateHz, u64NanoTS);
4759	}
4760
4761	/*
4762	* Update the data.
4763	*/
4764	supdrvGipDoUpdateCpu(pGip, pGipCpu, u64NanoTS);
4765
4766	/*
4767	* Complete transaction.
4768	*/
4769	ASMAtomicIncU32(&pGipCpu->u32TransactionId);
4770	}
4771
4772
4773	/**
4774	* Updates the per cpu GIP data for the calling cpu.
4775	*
4776	* @param pGip Pointer to the GIP.
4777	* @param u64NanoTS The current nanosecond timesamp.
4778	* @param iCpu The CPU index.
4779	*/
4780	void VBOXCALL supdrvGipUpdatePerCpu(PSUPGLOBALINFOPAGE pGip, uint64_t u64NanoTS, unsigned iCpu)
4781	{
4782	PSUPGIPCPU pGipCpu;
4783
4784	if (RT_LIKELY(iCpu < RT_ELEMENTS(pGip->aCPUs)))
4785	{
4786	pGipCpu = &pGip->aCPUs[iCpu];
4787
4788	/*
4789	* Start update transaction.
4790	*/
4791	if (!(ASMAtomicIncU32(&pGipCpu->u32TransactionId) & 1))
4792	{
4793	AssertMsgFailed(("Invalid transaction id, %#x, not odd!\n", pGipCpu->u32TransactionId));
4794	ASMAtomicIncU32(&pGipCpu->u32TransactionId);
4795	pGipCpu->cErrors++;
4796	return;
4797	}
4798
4799	/*
4800	* Update the data.
4801	*/
4802	supdrvGipDoUpdateCpu(pGip, pGipCpu, u64NanoTS);
4803
4804	/*
4805	* Complete transaction.
4806	*/
4807	ASMAtomicIncU32(&pGipCpu->u32TransactionId);
4808	}
4809	}
4810

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

source: vbox/trunk/src/VBox/HostDrivers/Support/SUPDrv.c@ 17827

Download in other formats: