SUPDrv.c@ 12100

最後變更在這個檔案從12100是 12100,由 vboxsync 提交於 17 年前
VBoxDrv: Finished the unwind fix for AMD64. (hides our code from RtlVirtualUnwind and associated stuff.)
屬性 svn:eol-style 設為 `native` 屬性 svn:keywords 設為 `Author Date Id Revision`
檔案大小: 182.2 KB

行
1	/* $Revision: 12100 $ */
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.alldomusa.eu.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/cpuset.h>
46	#include <iprt/uuid.h>
47	#include <VBox/log.h>
48	#include <VBox/err.h>
49	/* VBox/x86.h not compatible with the Linux kernel sources */
50	#ifdef RT_OS_LINUX
51	# define X86_CPUID_VENDOR_AMD_EBX 0x68747541
52	# define X86_CPUID_VENDOR_AMD_ECX 0x444d4163
53	# define X86_CPUID_VENDOR_AMD_EDX 0x69746e65
54	#else
55	# include <VBox/x86.h>
56	#endif
57
58	/*
59	* Logging assignments:
60	* Log - useful stuff, like failures.
61	* LogFlow - program flow, except the really noisy bits.
62	* Log2 - Cleanup and IDTE
63	* Log3 - Loader flow noise.
64	* Log4 - Call VMMR0 flow noise.
65	* Log5 - Native yet-to-be-defined noise.
66	* Log6 - Native ioctl flow noise.
67	*
68	* Logging requires BUILD_TYPE=debug and possibly changes to the logger
69	* instanciation in log-vbox.c(pp).
70	*/
71
72
73	/*******************************************************************************
74	* Defined Constants And Macros *
75	*******************************************************************************/
76	/* from x86.h - clashes with linux thus this duplication */
77	#undef X86_CR0_PG
78	#define X86_CR0_PG RT_BIT(31)
79	#undef X86_CR0_PE
80	#define X86_CR0_PE RT_BIT(0)
81	#undef X86_CPUID_AMD_FEATURE_EDX_NX
82	#define X86_CPUID_AMD_FEATURE_EDX_NX RT_BIT(20)
83	#undef MSR_K6_EFER
84	#define MSR_K6_EFER 0xc0000080
85	#undef MSR_K6_EFER_NXE
86	#define MSR_K6_EFER_NXE RT_BIT(11)
87	#undef MSR_K6_EFER_LMA
88	#define MSR_K6_EFER_LMA RT_BIT(10)
89	#undef X86_CR4_PGE
90	#define X86_CR4_PGE RT_BIT(7)
91	#undef X86_CR4_PAE
92	#define X86_CR4_PAE RT_BIT(5)
93	#undef X86_CPUID_AMD_FEATURE_EDX_LONG_MODE
94	#define X86_CPUID_AMD_FEATURE_EDX_LONG_MODE RT_BIT(29)
95
96
97	/** The frequency by which we recalculate the u32UpdateHz and
98	* u32UpdateIntervalNS GIP members. The value must be a power of 2. */
99	#define GIP_UPDATEHZ_RECALC_FREQ 0x800
100
101	/**
102	* Validates a session pointer.
103	*
104	* @returns true/false accordingly.
105	* @param pSession The session.
106	*/
107	#define SUP_IS_SESSION_VALID(pSession) \
108	( VALID_PTR(pSession) \
109	&& pSession->u32Cookie == BIRD_INV)
110
111	/** @def VBOX_SVN_REV
112	* The makefile should define this if it can. */
113	#ifndef VBOX_SVN_REV
114	# define VBOX_SVN_REV 0
115	#endif
116
117	/*******************************************************************************
118	* Internal Functions *
119	*******************************************************************************/
120	static int supdrvMemAdd(PSUPDRVMEMREF pMem, PSUPDRVSESSION pSession);
121	static int supdrvMemRelease(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr, SUPDRVMEMREFTYPE eType);
122	#ifdef VBOX_WITH_IDT_PATCHING
123	static int supdrvIOCtl_IdtInstall(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPIDTINSTALL pReq);
124	static PSUPDRVPATCH supdrvIdtPatchOne(PSUPDRVDEVEXT pDevExt, PSUPDRVPATCH pPatch);
125	static int supdrvIOCtl_IdtRemoveAll(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession);
126	static void supdrvIdtRemoveOne(PSUPDRVDEVEXT pDevExt, PSUPDRVPATCH pPatch);
127	static void supdrvIdtWrite(volatile void pvIdtEntry, const SUPDRVIDTE pNewIDTEntry);
128	#endif /* VBOX_WITH_IDT_PATCHING */
129	static int supdrvIOCtl_LdrOpen(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDROPEN pReq);
130	static int supdrvIOCtl_LdrLoad(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRLOAD pReq);
131	static int supdrvIOCtl_LdrFree(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRFREE pReq);
132	static int supdrvIOCtl_LdrGetSymbol(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRGETSYMBOL pReq);
133	static int supdrvIDC_LdrGetSymbol(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPDRVIDCREQGETSYM pReq);
134	static int supdrvLdrSetR0EP(PSUPDRVDEVEXT pDevExt, void pvVMMR0, void pvVMMR0EntryInt, void pvVMMR0EntryFast, void pvVMMR0EntryEx);
135	static void supdrvLdrUnsetR0EP(PSUPDRVDEVEXT pDevExt);
136	static int supdrvLdrAddUsage(PSUPDRVSESSION pSession, PSUPDRVLDRIMAGE pImage);
137	static void supdrvLdrFree(PSUPDRVDEVEXT pDevExt, PSUPDRVLDRIMAGE pImage);
138	static SUPPAGINGMODE supdrvIOCtl_GetPagingMode(void);
139	static SUPGIPMODE supdrvGipDeterminTscMode(PSUPDRVDEVEXT pDevExt);
140	#ifdef RT_OS_WINDOWS
141	static int supdrvPageGetPhys(PSUPDRVSESSION pSession, RTR3PTR pvR3, uint32_t cPages, PRTHCPHYS paPages);
142	static bool supdrvPageWasLockedByPageAlloc(PSUPDRVSESSION pSession, RTR3PTR pvR3);
143	#endif /* RT_OS_WINDOWS */
144	static int supdrvGipCreate(PSUPDRVDEVEXT pDevExt);
145	static void supdrvGipDestroy(PSUPDRVDEVEXT pDevExt);
146	static DECLCALLBACK(void) supdrvGipSyncTimer(PRTTIMER pTimer, void *pvUser, uint64_t iTick);
147	static DECLCALLBACK(void) supdrvGipAsyncTimer(PRTTIMER pTimer, void *pvUser, uint64_t iTick);
148	static DECLCALLBACK(void) supdrvGipMpEvent(RTMPEVENT enmEvent, RTCPUID idCpu, void *pvUser);
149
150	#ifdef SUPDRV_WITH_UNWIND_HACK
151	DECLASM(int) supdrvNtWrapVMMR0EntryEx(PFNRT pfnVMMR0EntryEx, PVM pVM, unsigned uOperation, PSUPVMMR0REQHDR pReq, uint64_t u64Arg, PSUPDRVSESSION pSession);
152	DECLASM(int) supdrvNtWrapVMMR0EntryFast(PFNRT pfnVMMR0EntryFast, PVM pVM, unsigned uOperation);
153	DECLASM(void) supdrvNtWrapObjDestructor(PFNRT pfnDestruction, void pvObj, void pvUser1, void *pvUser2);
154	DECLASM(void ) supdrvNtWrapQueryFactoryInterface(PFNRT pfnQueryFactoryInterface, struct SUPDRVFACTORY const pSupDrvFactory, PSUPDRVSESSION pSession, const char *pszInterfaceUuid);
155	DECLASM(int) supdrvNtWrapModuleInit(PFNRT pfnModuleInit);
156	DECLASM(void) supdrvNtWrapModuleTerm(PFNRT pfnModuleTerm);
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(SUPR0ObjRelease)(void *pvObj, PSUPDRVSESSION pSession);
164	DECLASM(int) UNWIND_WRAP(SUPR0ObjVerifyAccess)(void pvObj, PSUPDRVSESSION pSession, const char pszObjName);
165	DECLASM(int) UNWIND_WRAP(SUPR0LockMem)(PSUPDRVSESSION pSession, RTR3PTR pvR3, uint32_t cPages, PRTHCPHYS paPages);
166	DECLASM(int) UNWIND_WRAP(SUPR0UnlockMem)(PSUPDRVSESSION pSession, RTR3PTR pvR3);
167	DECLASM(int) UNWIND_WRAP(SUPR0ContAlloc)(PSUPDRVSESSION pSession, uint32_t cPages, PRTR0PTR ppvR0, PRTR3PTR ppvR3, PRTHCPHYS pHCPhys);
168	DECLASM(int) UNWIND_WRAP(SUPR0ContFree)(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr);
169	DECLASM(int) UNWIND_WRAP(SUPR0LowAlloc)(PSUPDRVSESSION pSession, uint32_t cPages, PRTR0PTR ppvR0, PRTR3PTR ppvR3, PRTHCPHYS paPages);
170	DECLASM(int) UNWIND_WRAP(SUPR0LowFree)(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr);
171	DECLASM(int) UNWIND_WRAP(SUPR0MemAlloc)(PSUPDRVSESSION pSession, uint32_t cb, PRTR0PTR ppvR0, PRTR3PTR ppvR3);
172	DECLASM(int) UNWIND_WRAP(SUPR0MemGetPhys)(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr, PSUPPAGE paPages);
173	DECLASM(int) UNWIND_WRAP(SUPR0MemFree)(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr);
174	DECLASM(int) UNWIND_WRAP(SUPR0PageAlloc)(PSUPDRVSESSION pSession, uint32_t cPages, PRTR3PTR ppvR3, PRTHCPHYS paPages);
175	DECLASM(int) UNWIND_WRAP(SUPR0PageFree)(PSUPDRVSESSION pSession, RTR3PTR pvR3);
176	//DECLASM(int) UNWIND_WRAP(SUPR0Printf)(const char *pszFormat, ...);
177	DECLASM(void *) UNWIND_WRAP(RTMemAlloc)(size_t cb) RT_NO_THROW;
178	DECLASM(void *) UNWIND_WRAP(RTMemAllocZ)(size_t cb) RT_NO_THROW;
179	DECLASM(void) UNWIND_WRAP(RTMemFree)(void *pv) RT_NO_THROW;
180	DECLASM(void ) UNWIND_WRAP(RTMemDup)(const void pvSrc, size_t cb) RT_NO_THROW;
181	DECLASM(void ) UNWIND_WRAP(RTMemDupEx)(const void pvSrc, size_t cbSrc, size_t cbExtra) RT_NO_THROW;
182	DECLASM(void ) UNWIND_WRAP(RTMemRealloc)(void pvOld, size_t cbNew) RT_NO_THROW;
183	DECLASM(int) UNWIND_WRAP(RTR0MemObjAllocLow)(PRTR0MEMOBJ pMemObj, size_t cb, bool fExecutable);
184	DECLASM(int) UNWIND_WRAP(RTR0MemObjAllocPage)(PRTR0MEMOBJ pMemObj, size_t cb, bool fExecutable);
185	DECLASM(int) UNWIND_WRAP(RTR0MemObjAllocPhys)(PRTR0MEMOBJ pMemObj, size_t cb, RTHCPHYS PhysHighest);
186	DECLASM(int) UNWIND_WRAP(RTR0MemObjAllocPhysNC)(PRTR0MEMOBJ pMemObj, size_t cb, RTHCPHYS PhysHighest);
187	DECLASM(int) UNWIND_WRAP(RTR0MemObjAllocCont)(PRTR0MEMOBJ pMemObj, size_t cb, bool fExecutable);
188	DECLASM(int) UNWIND_WRAP(RTR0MemObjLockUser)(PRTR0MEMOBJ pMemObj, RTR3PTR R3Ptr, size_t cb, RTR0PROCESS R0Process);
189	DECLASM(int) UNWIND_WRAP(RTR0MemObjMapKernel)(PRTR0MEMOBJ pMemObj, RTR0MEMOBJ MemObjToMap, void *pvFixed, size_t uAlignment, unsigned fProt);
190	DECLASM(int) UNWIND_WRAP(RTR0MemObjMapUser)(PRTR0MEMOBJ pMemObj, RTR0MEMOBJ MemObjToMap, RTR3PTR R3PtrFixed, size_t uAlignment, unsigned fProt, RTR0PROCESS R0Process);
191	DECLASM(void *) UNWIND_WRAP(RTR0MemObjAddress)(RTR0MEMOBJ MemObj);
192	DECLASM(RTR3PTR) UNWIND_WRAP(RTR0MemObjAddressR3)(RTR0MEMOBJ MemObj);
193	DECLASM(size_t) UNWIND_WRAP(RTR0MemObjSize)(RTR0MEMOBJ MemObj);
194	DECLASM(bool) UNWIND_WRAP(RTR0MemObjIsMapping)(RTR0MEMOBJ MemObj);
195	DECLASM(RTHCPHYS) UNWIND_WRAP(RTR0MemObjGetPagePhysAddr)(RTR0MEMOBJ MemObj, size_t iPage);
196	DECLASM(int) UNWIND_WRAP(RTR0MemObjFree)(RTR0MEMOBJ MemObj, bool fFreeMappings);
197	/* RTProcSelf - not necessary */
198	/* RTR0ProcHandleSelf - not necessary */
199	DECLASM(int) UNWIND_WRAP(RTSemFastMutexCreate)(PRTSEMFASTMUTEX pMutexSem);
200	DECLASM(int) UNWIND_WRAP(RTSemFastMutexDestroy)(RTSEMFASTMUTEX MutexSem);
201	DECLASM(int) UNWIND_WRAP(RTSemFastMutexRequest)(RTSEMFASTMUTEX MutexSem);
202	DECLASM(int) UNWIND_WRAP(RTSemFastMutexRelease)(RTSEMFASTMUTEX MutexSem);
203	DECLASM(int) UNWIND_WRAP(RTSemEventCreate)(PRTSEMEVENT pEventSem);
204	DECLASM(int) UNWIND_WRAP(RTSemEventSignal)(RTSEMEVENT EventSem);
205	DECLASM(int) UNWIND_WRAP(RTSemEventWait)(RTSEMEVENT EventSem, unsigned cMillies);
206	DECLASM(int) UNWIND_WRAP(RTSemEventWaitNoResume)(RTSEMEVENT EventSem, unsigned cMillies);
207	DECLASM(int) UNWIND_WRAP(RTSemEventDestroy)(RTSEMEVENT EventSem);
208	DECLASM(int) UNWIND_WRAP(RTSemEventMultiCreate)(PRTSEMEVENTMULTI pEventMultiSem);
209	DECLASM(int) UNWIND_WRAP(RTSemEventMultiSignal)(RTSEMEVENTMULTI EventMultiSem);
210	DECLASM(int) UNWIND_WRAP(RTSemEventMultiReset)(RTSEMEVENTMULTI EventMultiSem);
211	DECLASM(int) UNWIND_WRAP(RTSemEventMultiWait)(RTSEMEVENTMULTI EventMultiSem, unsigned cMillies);
212	DECLASM(int) UNWIND_WRAP(RTSemEventMultiWaitNoResume)(RTSEMEVENTMULTI EventMultiSem, unsigned cMillies);
213	DECLASM(int) UNWIND_WRAP(RTSemEventMultiDestroy)(RTSEMEVENTMULTI EventMultiSem);
214	DECLASM(int) UNWIND_WRAP(RTSpinlockCreate)(PRTSPINLOCK pSpinlock);
215	DECLASM(int) UNWIND_WRAP(RTSpinlockDestroy)(RTSPINLOCK Spinlock);
216	DECLASM(void) UNWIND_WRAP(RTSpinlockAcquire)(RTSPINLOCK Spinlock, PRTSPINLOCKTMP pTmp);
217	DECLASM(void) UNWIND_WRAP(RTSpinlockRelease)(RTSPINLOCK Spinlock, PRTSPINLOCKTMP pTmp);
218	DECLASM(void) UNWIND_WRAP(RTSpinlockAcquireNoInts)(RTSPINLOCK Spinlock, PRTSPINLOCKTMP pTmp);
219	DECLASM(void) UNWIND_WRAP(RTSpinlockReleaseNoInts)(RTSPINLOCK Spinlock, PRTSPINLOCKTMP pTmp);
220	/* RTTimeNanoTS - not necessary */
221	/* RTTimeMilliTS - not necessary */
222	/* RTTimeSystemNanoTS - not necessary */
223	/* RTTimeSystemMilliTS - not necessary */
224	/* RTThreadNativeSelf - not necessary */
225	DECLASM(int) UNWIND_WRAP(RTThreadSleep)(unsigned cMillies);
226	DECLASM(bool) UNWIND_WRAP(RTThreadYield)(void);
227	#if 0
228	/* RTThreadSelf - not necessary */
229	DECLASM(int) UNWIND_WRAP(RTThreadCreate)(PRTTHREAD pThread, PFNRTTHREAD pfnThread, void *pvUser, size_t cbStack,
230	RTTHREADTYPE enmType, unsigned fFlags, const char *pszName);
231	DECLASM(RTNATIVETHREAD) UNWIND_WRAP(RTThreadGetNative)(RTTHREAD Thread);
232	DECLASM(int) UNWIND_WRAP(RTThreadWait)(RTTHREAD Thread, unsigned cMillies, int *prc);
233	DECLASM(int) UNWIND_WRAP(RTThreadWaitNoResume)(RTTHREAD Thread, unsigned cMillies, int *prc);
234	DECLASM(const char *) UNWIND_WRAP(RTThreadGetName)(RTTHREAD Thread);
235	DECLASM(const char *) UNWIND_WRAP(RTThreadSelfName)(void);
236	DECLASM(RTTHREADTYPE) UNWIND_WRAP(RTThreadGetType)(RTTHREAD Thread);
237	DECLASM(int) UNWIND_WRAP(RTThreadUserSignal)(RTTHREAD Thread);
238	DECLASM(int) UNWIND_WRAP(RTThreadUserReset)(RTTHREAD Thread);
239	DECLASM(int) UNWIND_WRAP(RTThreadUserWait)(RTTHREAD Thread, unsigned cMillies);
240	DECLASM(int) UNWIND_WRAP(RTThreadUserWaitNoResume)(RTTHREAD Thread, unsigned cMillies);
241	#endif
242	/* RTLogDefaultInstance - a bit of a gamble, but we do not want the overhead! */
243	/* RTMpCpuId - not necessary */
244	/* RTMpCpuIdFromSetIndex - not necessary */
245	/* RTMpCpuIdToSetIndex - not necessary */
246	/* RTMpIsCpuPossible - not necessary */
247	/* RTMpGetCount - not necessary */
248	/* RTMpGetMaxCpuId - not necessary */
249	/* RTMpGetOnlineCount - not necessary */
250	/* RTMpGetOnlineSet - not necessary */
251	/* RTMpGetSet - not necessary */
252	/* RTMpIsCpuOnline - not necessary */
253	DECLASM(int) UNWIND_WRAP(RTMpOnAll)(PFNRTMPWORKER pfnWorker, void pvUser1, void pvUser2);
254	DECLASM(int) UNWIND_WRAP(RTMpOnOthers)(PFNRTMPWORKER pfnWorker, void pvUser1, void pvUser2);
255	DECLASM(int) UNWIND_WRAP(RTMpOnSpecific)(RTCPUID idCpu, PFNRTMPWORKER pfnWorker, void pvUser1, void pvUser2);
256	/* RTLogRelDefaultInstance - not necessary. */
257	DECLASM(int) UNWIND_WRAP(RTLogSetDefaultInstanceThread)(PRTLOGGER pLogger, uintptr_t uKey);
258	/* RTLogLogger - can't wrap this buster. */
259	/* RTLogLoggerEx - can't wrap this buster. */
260	DECLASM(void) UNWIND_WRAP(RTLogLoggerExV)(PRTLOGGER pLogger, unsigned fFlags, unsigned iGroup, const char *pszFormat, va_list args);
261	/* RTLogPrintf - can't wrap this buster. / /* @todo provide va_list log wrappers in RuntimeR0. */
262	DECLASM(void) UNWIND_WRAP(RTLogPrintfV)(const char *pszFormat, va_list args);
263	DECLASM(void) UNWIND_WRAP(AssertMsg1)(const char pszExpr, unsigned uLine, const char pszFile, const char *pszFunction);
264	/* AssertMsg2 - can't wrap this buster. */
265	#endif /* SUPDRV_WITH_UNWIND_HACK */
266
267
268	/*******************************************************************************
269	* Global Variables *
270	*******************************************************************************/
271	/**
272	* Array of the R0 SUP API.
273	*/
274	static SUPFUNC g_aFunctions[] =
275	{
276	/* name function */
277	{ "SUPR0ComponentRegisterFactory", (void *)UNWIND_WRAP(SUPR0ComponentRegisterFactory) },
278	{ "SUPR0ComponentDeregisterFactory", (void *)UNWIND_WRAP(SUPR0ComponentDeregisterFactory) },
279	{ "SUPR0ComponentQueryFactory", (void *)UNWIND_WRAP(SUPR0ComponentQueryFactory) },
280	{ "SUPR0ObjRegister", (void *)SUPR0ObjRegister },
281	{ "SUPR0ObjAddRef", (void *)SUPR0ObjAddRef },
282	{ "SUPR0ObjRelease", (void *)SUPR0ObjRelease },
283	{ "SUPR0ObjVerifyAccess", (void *)SUPR0ObjVerifyAccess },
284	{ "SUPR0LockMem", (void *)SUPR0LockMem },
285	{ "SUPR0UnlockMem", (void *)SUPR0UnlockMem },
286	{ "SUPR0ContAlloc", (void *)SUPR0ContAlloc },
287	{ "SUPR0ContFree", (void *)SUPR0ContFree },
288	{ "SUPR0LowAlloc", (void *)SUPR0LowAlloc },
289	{ "SUPR0LowFree", (void *)SUPR0LowFree },
290	{ "SUPR0MemAlloc", (void *)SUPR0MemAlloc },
291	{ "SUPR0MemGetPhys", (void *)SUPR0MemGetPhys },
292	{ "SUPR0MemFree", (void *)SUPR0MemFree },
293	{ "SUPR0PageAlloc", (void *)SUPR0PageAlloc },
294	{ "SUPR0PageFree", (void *)SUPR0PageFree },
295	{ "SUPR0Printf", (void *)SUPR0Printf },
296	{ "RTMemAlloc", (void *)RTMemAlloc },
297	{ "RTMemAllocZ", (void *)RTMemAllocZ },
298	{ "RTMemFree", (void *)RTMemFree },
299	/{ "RTMemDup", (void )RTMemDup },*/
300	{ "RTMemRealloc", (void *)RTMemRealloc },
301	{ "RTR0MemObjAllocLow", (void *)RTR0MemObjAllocLow },
302	{ "RTR0MemObjAllocPage", (void *)RTR0MemObjAllocPage },
303	{ "RTR0MemObjAllocPhys", (void *)RTR0MemObjAllocPhys },
304	{ "RTR0MemObjAllocPhysNC", (void *)RTR0MemObjAllocPhysNC },
305	{ "RTR0MemObjAllocCont", (void *)RTR0MemObjAllocCont },
306	{ "RTR0MemObjLockUser", (void *)RTR0MemObjLockUser },
307	{ "RTR0MemObjMapKernel", (void *)RTR0MemObjMapKernel },
308	{ "RTR0MemObjMapUser", (void *)RTR0MemObjMapUser },
309	{ "RTR0MemObjAddress", (void *)RTR0MemObjAddress },
310	{ "RTR0MemObjAddressR3", (void *)RTR0MemObjAddressR3 },
311	{ "RTR0MemObjSize", (void *)RTR0MemObjSize },
312	{ "RTR0MemObjIsMapping", (void *)RTR0MemObjIsMapping },
313	{ "RTR0MemObjGetPagePhysAddr", (void *)RTR0MemObjGetPagePhysAddr },
314	{ "RTR0MemObjFree", (void *)RTR0MemObjFree },
315	/* These don't work yet on linux - use fast mutexes!
316	{ "RTSemMutexCreate", (void *)RTSemMutexCreate },
317	{ "RTSemMutexRequest", (void *)RTSemMutexRequest },
318	{ "RTSemMutexRelease", (void *)RTSemMutexRelease },
319	{ "RTSemMutexDestroy", (void *)RTSemMutexDestroy },
320	*/
321	{ "RTProcSelf", (void *)RTProcSelf },
322	{ "RTR0ProcHandleSelf", (void *)RTR0ProcHandleSelf },
323	{ "RTSemFastMutexCreate", (void *)RTSemFastMutexCreate },
324	{ "RTSemFastMutexDestroy", (void *)RTSemFastMutexDestroy },
325	{ "RTSemFastMutexRequest", (void *)UNWIND_WRAP(RTSemFastMutexRequest) },
326	{ "RTSemFastMutexRelease", (void *)RTSemFastMutexRelease },
327	{ "RTSemEventCreate", (void *)RTSemEventCreate },
328	{ "RTSemEventSignal", (void *)RTSemEventSignal },
329	{ "RTSemEventWait", (void *)UNWIND_WRAP(RTSemEventWait) },
330	{ "RTSemEventWaitNoResume", (void *)UNWIND_WRAP(RTSemEventWaitNoResume) },
331	{ "RTSemEventDestroy", (void *)RTSemEventDestroy },
332	{ "RTSemEventMultiCreate", (void *)RTSemEventMultiCreate },
333	{ "RTSemEventMultiSignal", (void *)RTSemEventMultiSignal },
334	{ "RTSemEventMultiReset", (void *)RTSemEventMultiReset },
335	{ "RTSemEventMultiWait", (void *)UNWIND_WRAP(RTSemEventMultiWait) },
336	{ "RTSemEventMultiWaitNoResume", (void *)UNWIND_WRAP(RTSemEventMultiWaitNoResume) },
337	{ "RTSemEventMultiDestroy", (void *)RTSemEventMultiDestroy },
338	{ "RTSpinlockCreate", (void *)RTSpinlockCreate },
339	{ "RTSpinlockDestroy", (void *)RTSpinlockDestroy },
340	{ "RTSpinlockAcquire", (void *)RTSpinlockAcquire },
341	{ "RTSpinlockRelease", (void *)UNWIND_WRAP(RTSpinlockRelease) },
342	{ "RTSpinlockAcquireNoInts", (void *)RTSpinlockAcquireNoInts },
343	{ "RTSpinlockReleaseNoInts", (void *)UNWIND_WRAP(RTSpinlockReleaseNoInts) },
344	{ "RTTimeNanoTS", (void *)RTTimeNanoTS },
345	{ "RTTimeMillieTS", (void *)RTTimeMilliTS },
346	{ "RTTimeSystemNanoTS", (void *)RTTimeSystemNanoTS },
347	{ "RTTimeSystemMillieTS", (void *)RTTimeSystemMilliTS },
348	{ "RTThreadNativeSelf", (void *)RTThreadNativeSelf },
349	{ "RTThreadSleep", (void *)UNWIND_WRAP(RTThreadSleep) },
350	{ "RTThreadYield", (void *)UNWIND_WRAP(RTThreadYield) },
351	#if 0 /* Thread APIs, Part 2. */
352	{ "RTThreadSelf", (void *)RTThreadSelf },
353	{ "RTThreadCreate", (void *)RTThreadCreate },
354	{ "RTThreadGetNative", (void *)RTThreadGetNative },
355	{ "RTThreadWait", (void *)RTThreadWait },
356	{ "RTThreadWaitNoResume", (void *)RTThreadWaitNoResume },
357	{ "RTThreadGetName", (void *)RTThreadGetName },
358	{ "RTThreadSelfName", (void *)RTThreadSelfName },
359	{ "RTThreadGetType", (void *)RTThreadGetType },
360	{ "RTThreadUserSignal", (void *)RTThreadUserSignal },
361	{ "RTThreadUserReset", (void *)RTThreadUserReset },
362	{ "RTThreadUserWait", (void *)RTThreadUserWait },
363	{ "RTThreadUserWaitNoResume", (void *)RTThreadUserWaitNoResume },
364	#endif
365	{ "RTLogDefaultInstance", (void *)RTLogDefaultInstance },
366	{ "RTMpCpuId", (void *)RTMpCpuId },
367	{ "RTMpCpuIdFromSetIndex", (void *)RTMpCpuIdFromSetIndex },
368	{ "RTMpCpuIdToSetIndex", (void *)RTMpCpuIdToSetIndex },
369	{ "RTMpIsCpuPossible", (void *)RTMpIsCpuPossible },
370	{ "RTMpGetCount", (void *)RTMpGetCount },
371	{ "RTMpGetMaxCpuId", (void *)RTMpGetMaxCpuId },
372	{ "RTMpGetOnlineCount", (void *)RTMpGetOnlineCount },
373	{ "RTMpGetOnlineSet", (void *)RTMpGetOnlineSet },
374	{ "RTMpGetSet", (void *)RTMpGetSet },
375	{ "RTMpIsCpuOnline", (void *)RTMpIsCpuOnline },
376	{ "RTMpOnAll", (void *)RTMpOnAll },
377	{ "RTMpOnOthers", (void *)RTMpOnOthers },
378	{ "RTMpOnSpecific", (void *)RTMpOnSpecific },
379	{ "RTLogRelDefaultInstance", (void *)RTLogRelDefaultInstance },
380	{ "RTLogSetDefaultInstanceThread", (void *)RTLogSetDefaultInstanceThread },
381	{ "RTLogLogger", (void *)RTLogLogger },
382	{ "RTLogLoggerEx", (void *)RTLogLoggerEx },
383	{ "RTLogLoggerExV", (void *)RTLogLoggerExV },
384	{ "RTLogPrintf", (void *)RTLogPrintf },
385	{ "RTLogPrintfV", (void *)RTLogPrintfV },
386	{ "AssertMsg1", (void *)AssertMsg1 },
387	{ "AssertMsg2", (void *)AssertMsg2 },
388	};
389
390
391	/**
392	* Initializes the device extentsion structure.
393	*
394	* @returns IPRT status code.
395	* @param pDevExt The device extension to initialize.
396	*/
397	int VBOXCALL supdrvInitDevExt(PSUPDRVDEVEXT pDevExt)
398	{
399	int rc;
400
401	#ifdef SUPDRV_WITH_RELEASE_LOGGER
402	/*
403	* Create the release log.
404	*/
405	static const char * const s_apszGroups[] = VBOX_LOGGROUP_NAMES;
406	PRTLOGGER pRelLogger;
407	rc = RTLogCreate(&pRelLogger, 0 /* fFlags */, "all",
408	"VBOX_RELEASE_LOG", RT_ELEMENTS(s_apszGroups), s_apszGroups,
409	RTLOGDEST_STDOUT \| RTLOGDEST_DEBUGGER, NULL);
410	if (RT_SUCCESS(rc))
411	RTLogRelSetDefaultInstance(pRelLogger);
412	#endif
413
414	/*
415	* Initialize it.
416	*/
417	memset(pDevExt, 0, sizeof(*pDevExt));
418	rc = RTSpinlockCreate(&pDevExt->Spinlock);
419	if (!rc)
420	{
421	rc = RTSemFastMutexCreate(&pDevExt->mtxLdr);
422	if (!rc)
423	{
424	rc = RTSemFastMutexCreate(&pDevExt->mtxComponentFactory);
425	if (!rc)
426	{
427	rc = RTSemFastMutexCreate(&pDevExt->mtxGip);
428	if (!rc)
429	{
430	rc = supdrvGipCreate(pDevExt);
431	if (RT_SUCCESS(rc))
432	{
433	pDevExt->u32Cookie = BIRD; /** @todo make this random? */
434	return VINF_SUCCESS;
435	}
436
437	RTSemFastMutexDestroy(pDevExt->mtxGip);
438	pDevExt->mtxGip = NIL_RTSEMFASTMUTEX;
439	}
440	RTSemFastMutexDestroy(pDevExt->mtxComponentFactory);
441	pDevExt->mtxComponentFactory = NIL_RTSEMFASTMUTEX;
442	}
443	RTSemFastMutexDestroy(pDevExt->mtxLdr);
444	pDevExt->mtxLdr = NIL_RTSEMFASTMUTEX;
445	}
446	RTSpinlockDestroy(pDevExt->Spinlock);
447	pDevExt->Spinlock = NIL_RTSPINLOCK;
448	}
449	#ifdef SUPDRV_WITH_RELEASE_LOGGER
450	RTLogDestroy(RTLogRelSetDefaultInstance(NULL));
451	RTLogDestroy(RTLogSetDefaultInstance(NULL));
452	#endif
453
454	return rc;
455	}
456
457
458	/**
459	* Delete the device extension (e.g. cleanup members).
460	*
461	* @param pDevExt The device extension to delete.
462	*/
463	void VBOXCALL supdrvDeleteDevExt(PSUPDRVDEVEXT pDevExt)
464	{
465	#ifdef VBOX_WITH_IDT_PATCHING
466	PSUPDRVPATCH pPatch;
467	#endif
468	PSUPDRVOBJ pObj;
469	PSUPDRVUSAGE pUsage;
470
471	/*
472	* Kill mutexes and spinlocks.
473	*/
474	RTSemFastMutexDestroy(pDevExt->mtxGip);
475	pDevExt->mtxGip = NIL_RTSEMFASTMUTEX;
476	RTSemFastMutexDestroy(pDevExt->mtxLdr);
477	pDevExt->mtxLdr = NIL_RTSEMFASTMUTEX;
478	RTSpinlockDestroy(pDevExt->Spinlock);
479	pDevExt->Spinlock = NIL_RTSPINLOCK;
480
481	/*
482	* Free lists.
483	*/
484	#ifdef VBOX_WITH_IDT_PATCHING
485	/* patches */
486	/** @todo make sure we don't uninstall patches which has been patched by someone else. */
487	pPatch = pDevExt->pIdtPatchesFree;
488	pDevExt->pIdtPatchesFree = NULL;
489	while (pPatch)
490	{
491	void *pvFree = pPatch;
492	pPatch = pPatch->pNext;
493	RTMemExecFree(pvFree);
494	}
495	#endif /* VBOX_WITH_IDT_PATCHING */
496
497	/* objects. */
498	pObj = pDevExt->pObjs;
499	#if !defined(DEBUG_bird) \|\| !defined(RT_OS_LINUX) /* breaks unloading, temporary, remove me! */
500	Assert(!pObj); /* (can trigger on forced unloads) */
501	#endif
502	pDevExt->pObjs = NULL;
503	while (pObj)
504	{
505	void *pvFree = pObj;
506	pObj = pObj->pNext;
507	RTMemFree(pvFree);
508	}
509
510	/* usage records. */
511	pUsage = pDevExt->pUsageFree;
512	pDevExt->pUsageFree = NULL;
513	while (pUsage)
514	{
515	void *pvFree = pUsage;
516	pUsage = pUsage->pNext;
517	RTMemFree(pvFree);
518	}
519
520	/* kill the GIP. */
521	supdrvGipDestroy(pDevExt);
522
523	#ifdef SUPDRV_WITH_RELEASE_LOGGER
524	/* destroy the loggers. */
525	RTLogDestroy(RTLogRelSetDefaultInstance(NULL));
526	RTLogDestroy(RTLogSetDefaultInstance(NULL));
527	#endif
528	}
529
530
531	/**
532	* Create session.
533	*
534	* @returns IPRT status code.
535	* @param pDevExt Device extension.
536	* @param fUser Flag indicating whether this is a user or kernel session.
537	* @param ppSession Where to store the pointer to the session data.
538	*/
539	int VBOXCALL supdrvCreateSession(PSUPDRVDEVEXT pDevExt, bool fUser, PSUPDRVSESSION *ppSession)
540	{
541	/*
542	* Allocate memory for the session data.
543	*/
544	int rc = VERR_NO_MEMORY;
545	PSUPDRVSESSION pSession = ppSession = (PSUPDRVSESSION)RTMemAllocZ(sizeof(pSession));
546	if (pSession)
547	{
548	/* Initialize session data. */
549	rc = RTSpinlockCreate(&pSession->Spinlock);
550	if (!rc)
551	{
552	Assert(pSession->Spinlock != NIL_RTSPINLOCK);
553	pSession->pDevExt = pDevExt;
554	pSession->u32Cookie = BIRD_INV;
555	/*pSession->pLdrUsage = NULL;
556	pSession->pPatchUsage = NULL;
557	pSession->pVM = NULL;
558	pSession->pUsage = NULL;
559	pSession->pGip = NULL;
560	pSession->fGipReferenced = false;
561	pSession->Bundle.cUsed = 0; */
562	pSession->Uid = NIL_RTUID;
563	pSession->Gid = NIL_RTGID;
564	if (fUser)
565	{
566	pSession->Process = RTProcSelf();
567	pSession->R0Process = RTR0ProcHandleSelf();
568	}
569	else
570	{
571	pSession->Process = NIL_RTPROCESS;
572	pSession->R0Process = NIL_RTR0PROCESS;
573	}
574
575	LogFlow(("Created session %p initial cookie=%#x\n", pSession, pSession->u32Cookie));
576	return VINF_SUCCESS;
577	}
578
579	RTMemFree(pSession);
580	*ppSession = NULL;
581	Log(("Failed to create spinlock, rc=%d!\n", rc));
582	}
583
584	return rc;
585	}
586
587
588	/**
589	* Shared code for cleaning up a session.
590	*
591	* @param pDevExt Device extension.
592	* @param pSession Session data.
593	* This data will be freed by this routine.
594	*/
595	void VBOXCALL supdrvCloseSession(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession)
596	{
597	/*
598	* Cleanup the session first.
599	*/
600	supdrvCleanupSession(pDevExt, pSession);
601
602	/*
603	* Free the rest of the session stuff.
604	*/
605	RTSpinlockDestroy(pSession->Spinlock);
606	pSession->Spinlock = NIL_RTSPINLOCK;
607	pSession->pDevExt = NULL;
608	RTMemFree(pSession);
609	LogFlow(("supdrvCloseSession: returns\n"));
610	}
611
612
613	/**
614	* Shared code for cleaning up a session (but not quite freeing it).
615	*
616	* This is primarily intended for MAC OS X where we have to clean up the memory
617	* stuff before the file handle is closed.
618	*
619	* @param pDevExt Device extension.
620	* @param pSession Session data.
621	* This data will be freed by this routine.
622	*/
623	void VBOXCALL supdrvCleanupSession(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession)
624	{
625	PSUPDRVBUNDLE pBundle;
626	LogFlow(("supdrvCleanupSession: pSession=%p\n", pSession));
627
628	/*
629	* Remove logger instances related to this session.
630	*/
631	RTLogSetDefaultInstanceThread(NULL, (uintptr_t)pSession);
632
633	#ifdef VBOX_WITH_IDT_PATCHING
634	/*
635	* Uninstall any IDT patches installed for this session.
636	*/
637	supdrvIOCtl_IdtRemoveAll(pDevExt, pSession);
638	#endif
639
640	/*
641	* Release object references made in this session.
642	* In theory there should be noone racing us in this session.
643	*/
644	Log2(("release objects - start\n"));
645	if (pSession->pUsage)
646	{
647	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
648	PSUPDRVUSAGE pUsage;
649	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
650
651	while ((pUsage = pSession->pUsage) != NULL)
652	{
653	PSUPDRVOBJ pObj = pUsage->pObj;
654	pSession->pUsage = pUsage->pNext;
655
656	AssertMsg(pUsage->cUsage >= 1 && pObj->cUsage >= pUsage->cUsage, ("glob %d; sess %d\n", pObj->cUsage, pUsage->cUsage));
657	if (pUsage->cUsage < pObj->cUsage)
658	{
659	pObj->cUsage -= pUsage->cUsage;
660	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
661	}
662	else
663	{
664	/* Destroy the object and free the record. */
665	if (pDevExt->pObjs == pObj)
666	pDevExt->pObjs = pObj->pNext;
667	else
668	{
669	PSUPDRVOBJ pObjPrev;
670	for (pObjPrev = pDevExt->pObjs; pObjPrev; pObjPrev = pObjPrev->pNext)
671	if (pObjPrev->pNext == pObj)
672	{
673	pObjPrev->pNext = pObj->pNext;
674	break;
675	}
676	Assert(pObjPrev);
677	}
678	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
679
680	Log(("supdrvCleanupSession: destroying %p/%d (%p/%p) cpid=%RTproc pid=%RTproc dtor=%p\n",
681	pObj, pObj->enmType, pObj->pvUser1, pObj->pvUser2, pObj->CreatorProcess, RTProcSelf(), pObj->pfnDestructor));
682	if (pObj->pfnDestructor)
683	#ifdef SUPDRV_WITH_UNWIND_HACK
684	supdrvNtWrapObjDestructor((PFNRT)pObj->pfnDestructor, pObj, pObj->pvUser1, pObj->pvUser2);
685	#else
686	pObj->pfnDestructor(pObj, pObj->pvUser1, pObj->pvUser2);
687	#endif
688	RTMemFree(pObj);
689	}
690
691	/* free it and continue. */
692	RTMemFree(pUsage);
693
694	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
695	}
696
697	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
698	AssertMsg(!pSession->pUsage, ("Some buster reregistered an object during desturction!\n"));
699	}
700	Log2(("release objects - done\n"));
701
702	/*
703	* Release memory allocated in the session.
704	*
705	* We do not serialize this as we assume that the application will
706	* not allocated memory while closing the file handle object.
707	*/
708	Log2(("freeing memory:\n"));
709	pBundle = &pSession->Bundle;
710	while (pBundle)
711	{
712	PSUPDRVBUNDLE pToFree;
713	unsigned i;
714
715	/*
716	* Check and unlock all entries in the bundle.
717	*/
718	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
719	{
720	if (pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ)
721	{
722	int rc;
723	Log2(("eType=%d pvR0=%p pvR3=%p cb=%ld\n", pBundle->aMem[i].eType, RTR0MemObjAddress(pBundle->aMem[i].MemObj),
724	(void *)RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3), (long)RTR0MemObjSize(pBundle->aMem[i].MemObj)));
725	if (pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ)
726	{
727	rc = RTR0MemObjFree(pBundle->aMem[i].MapObjR3, false);
728	AssertRC(rc); /** @todo figure out how to handle this. */
729	pBundle->aMem[i].MapObjR3 = NIL_RTR0MEMOBJ;
730	}
731	rc = RTR0MemObjFree(pBundle->aMem[i].MemObj, false);
732	AssertRC(rc); /** @todo figure out how to handle this. */
733	pBundle->aMem[i].MemObj = NIL_RTR0MEMOBJ;
734	pBundle->aMem[i].eType = MEMREF_TYPE_UNUSED;
735	}
736	}
737
738	/*
739	* Advance and free previous bundle.
740	*/
741	pToFree = pBundle;
742	pBundle = pBundle->pNext;
743
744	pToFree->pNext = NULL;
745	pToFree->cUsed = 0;
746	if (pToFree != &pSession->Bundle)
747	RTMemFree(pToFree);
748	}
749	Log2(("freeing memory - done\n"));
750
751	/*
752	* Deregister component factories.
753	*/
754	RTSemFastMutexRequest(pDevExt->mtxComponentFactory);
755	Log2(("deregistering component factories:\n"));
756	if (pDevExt->pComponentFactoryHead)
757	{
758	PSUPDRVFACTORYREG pPrev = NULL;
759	PSUPDRVFACTORYREG pCur = pDevExt->pComponentFactoryHead;
760	while (pCur)
761	{
762	if (pCur->pSession == pSession)
763	{
764	/* unlink it */
765	PSUPDRVFACTORYREG pNext = pCur->pNext;
766	if (pPrev)
767	pPrev->pNext = pNext;
768	else
769	pDevExt->pComponentFactoryHead = pNext;
770
771	/* free it */
772	pCur->pNext = NULL;
773	pCur->pSession = NULL;
774	pCur->pFactory = NULL;
775	RTMemFree(pCur);
776
777	/* next */
778	pCur = pNext;
779	}
780	else
781	{
782	/* next */
783	pPrev = pCur;
784	pCur = pCur->pNext;
785	}
786	}
787	}
788	RTSemFastMutexRelease(pDevExt->mtxComponentFactory);
789	Log2(("deregistering component factories - done\n"));
790
791	/*
792	* Loaded images needs to be dereferenced and possibly freed up.
793	*/
794	RTSemFastMutexRequest(pDevExt->mtxLdr);
795	Log2(("freeing images:\n"));
796	if (pSession->pLdrUsage)
797	{
798	PSUPDRVLDRUSAGE pUsage = pSession->pLdrUsage;
799	pSession->pLdrUsage = NULL;
800	while (pUsage)
801	{
802	void *pvFree = pUsage;
803	PSUPDRVLDRIMAGE pImage = pUsage->pImage;
804	if (pImage->cUsage > pUsage->cUsage)
805	pImage->cUsage -= pUsage->cUsage;
806	else
807	supdrvLdrFree(pDevExt, pImage);
808	pUsage->pImage = NULL;
809	pUsage = pUsage->pNext;
810	RTMemFree(pvFree);
811	}
812	}
813	RTSemFastMutexRelease(pDevExt->mtxLdr);
814	Log2(("freeing images - done\n"));
815
816	/*
817	* Unmap the GIP.
818	*/
819	Log2(("umapping GIP:\n"));
820	if (pSession->GipMapObjR3 != NIL_RTR0MEMOBJ)
821	{
822	SUPR0GipUnmap(pSession);
823	pSession->fGipReferenced = 0;
824	}
825	Log2(("umapping GIP - done\n"));
826	}
827
828
829	/**
830	* Fast path I/O Control worker.
831	*
832	* @returns VBox status code that should be passed down to ring-3 unchanged.
833	* @param uIOCtl Function number.
834	* @param pDevExt Device extention.
835	* @param pSession Session data.
836	*/
837	int VBOXCALL supdrvIOCtlFast(uintptr_t uIOCtl, PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession)
838	{
839	/*
840	* We check the two prereqs after doing this only to allow the compiler to optimize things better.
841	*/
842	if (RT_LIKELY(pSession->pVM && pDevExt->pfnVMMR0EntryFast))
843	{
844	switch (uIOCtl)
845	{
846	case SUP_IOCTL_FAST_DO_RAW_RUN:
847	#ifdef SUPDRV_WITH_UNWIND_HACK
848	supdrvNtWrapVMMR0EntryFast((PFNRT)pDevExt->pfnVMMR0EntryFast, pSession->pVM, SUP_VMMR0_DO_RAW_RUN);
849	#else
850	pDevExt->pfnVMMR0EntryFast(pSession->pVM, SUP_VMMR0_DO_RAW_RUN);
851	#endif
852	break;
853	case SUP_IOCTL_FAST_DO_HWACC_RUN:
854	#ifdef SUPDRV_WITH_UNWIND_HACK
855	supdrvNtWrapVMMR0EntryFast((PFNRT)pDevExt->pfnVMMR0EntryFast, pSession->pVM, SUP_VMMR0_DO_HWACC_RUN);
856	#else
857	pDevExt->pfnVMMR0EntryFast(pSession->pVM, SUP_VMMR0_DO_HWACC_RUN);
858	#endif
859	break;
860	case SUP_IOCTL_FAST_DO_NOP:
861	#ifdef SUPDRV_WITH_UNWIND_HACK
862	supdrvNtWrapVMMR0EntryFast((PFNRT)pDevExt->pfnVMMR0EntryFast, pSession->pVM, SUP_VMMR0_DO_NOP);
863	#else
864	pDevExt->pfnVMMR0EntryFast(pSession->pVM, SUP_VMMR0_DO_NOP);
865	#endif
866	break;
867	default:
868	return VERR_INTERNAL_ERROR;
869	}
870	return VINF_SUCCESS;
871	}
872	return VERR_INTERNAL_ERROR;
873	}
874
875
876	/**
877	* Helper for supdrvIOCtl. Check if pszStr contains any character of pszChars.
878	* We would use strpbrk here if this function would be contained in the RedHat kABI white
879	* list, see http://www.kerneldrivers.org/RHEL5.
880	*
881	* @return 1 if pszStr does contain any character of pszChars, 0 otherwise.
882	* @param pszStr String to check
883	* @param pszChars Character set
884	*/
885	static int supdrvCheckInvalidChar(const char pszStr, const char pszChars)
886	{
887	int chCur;
888	while ((chCur = *pszStr++) != '\0')
889	{
890	int ch;
891	const char *psz = pszChars;
892	while ((ch = *psz++) != '\0')
893	if (ch == chCur)
894	return 1;
895
896	}
897	return 0;
898	}
899
900
901	/**
902	* I/O Control worker.
903	*
904	* @returns 0 on success.
905	* @returns VERR_INVALID_PARAMETER if the request is invalid.
906	*
907	* @param uIOCtl Function number.
908	* @param pDevExt Device extention.
909	* @param pSession Session data.
910	* @param pReqHdr The request header.
911	*/
912	int VBOXCALL supdrvIOCtl(uintptr_t uIOCtl, PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPREQHDR pReqHdr)
913	{
914	/*
915	* Validate the request.
916	*/
917	/* this first check could probably be omitted as its also done by the OS specific code... */
918	if (RT_UNLIKELY( (pReqHdr->fFlags & SUPREQHDR_FLAGS_MAGIC_MASK) != SUPREQHDR_FLAGS_MAGIC
919	\|\| pReqHdr->cbIn < sizeof(*pReqHdr)
920	\|\| pReqHdr->cbOut < sizeof(*pReqHdr)))
921	{
922	OSDBGPRINT(("vboxdrv: Bad ioctl request header; cbIn=%#lx cbOut=%#lx fFlags=%#lx\n",
923	(long)pReqHdr->cbIn, (long)pReqHdr->cbOut, (long)pReqHdr->fFlags));
924	return VERR_INVALID_PARAMETER;
925	}
926	if (RT_UNLIKELY(uIOCtl == SUP_IOCTL_COOKIE))
927	{
928	if (pReqHdr->u32Cookie != SUPCOOKIE_INITIAL_COOKIE)
929	{
930	OSDBGPRINT(("SUP_IOCTL_COOKIE: bad cookie %#lx\n", (long)pReqHdr->u32Cookie));
931	return VERR_INVALID_PARAMETER;
932	}
933	}
934	else if (RT_UNLIKELY( pReqHdr->u32Cookie != pDevExt->u32Cookie
935	\|\| pReqHdr->u32SessionCookie != pSession->u32Cookie))
936	{
937	OSDBGPRINT(("vboxdrv: bad cookie %#lx / %#lx.\n", (long)pReqHdr->u32Cookie, (long)pReqHdr->u32SessionCookie));
938	return VERR_INVALID_PARAMETER;
939	}
940
941	/*
942	* Validation macros
943	*/
944	#define REQ_CHECK_SIZES_EX(Name, cbInExpect, cbOutExpect) \
945	do { \
946	if (RT_UNLIKELY(pReqHdr->cbIn != (cbInExpect) \|\| pReqHdr->cbOut != (cbOutExpect))) \
947	{ \
948	OSDBGPRINT(( #Name ": Invalid input/output sizes. cbIn=%ld expected %ld. cbOut=%ld expected %ld.\n", \
949	(long)pReq->Hdr.cbIn, (long)(cbInExpect), (long)pReq->Hdr.cbOut, (long)(cbOutExpect))); \
950	return pReq->Hdr.rc = VERR_INVALID_PARAMETER; \
951	} \
952	} while (0)
953
954	#define REQ_CHECK_SIZES(Name) REQ_CHECK_SIZES_EX(Name, Name ## _SIZE_IN, Name ## _SIZE_OUT)
955
956	#define REQ_CHECK_SIZE_IN(Name, cbInExpect) \
957	do { \
958	if (RT_UNLIKELY(pReqHdr->cbIn != (cbInExpect))) \
959	{ \
960	OSDBGPRINT(( #Name ": Invalid input/output sizes. cbIn=%ld expected %ld.\n", \
961	(long)pReq->Hdr.cbIn, (long)(cbInExpect))); \
962	return pReq->Hdr.rc = VERR_INVALID_PARAMETER; \
963	} \
964	} while (0)
965
966	#define REQ_CHECK_SIZE_OUT(Name, cbOutExpect) \
967	do { \
968	if (RT_UNLIKELY(pReqHdr->cbOut != (cbOutExpect))) \
969	{ \
970	OSDBGPRINT(( #Name ": Invalid input/output sizes. cbOut=%ld expected %ld.\n", \
971	(long)pReq->Hdr.cbOut, (long)(cbOutExpect))); \
972	return pReq->Hdr.rc = VERR_INVALID_PARAMETER; \
973	} \
974	} while (0)
975
976	#define REQ_CHECK_EXPR(Name, expr) \
977	do { \
978	if (RT_UNLIKELY(!(expr))) \
979	{ \
980	OSDBGPRINT(( #Name ": %s\n", #expr)); \
981	return pReq->Hdr.rc = VERR_INVALID_PARAMETER; \
982	} \
983	} while (0)
984
985	#define REQ_CHECK_EXPR_FMT(expr, fmt) \
986	do { \
987	if (RT_UNLIKELY(!(expr))) \
988	{ \
989	OSDBGPRINT( fmt ); \
990	return pReq->Hdr.rc = VERR_INVALID_PARAMETER; \
991	} \
992	} while (0)
993
994
995	/*
996	* The switch.
997	*/
998	switch (SUP_CTL_CODE_NO_SIZE(uIOCtl))
999	{
1000	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_COOKIE):
1001	{
1002	PSUPCOOKIE pReq = (PSUPCOOKIE)pReqHdr;
1003	REQ_CHECK_SIZES(SUP_IOCTL_COOKIE);
1004	if (strncmp(pReq->u.In.szMagic, SUPCOOKIE_MAGIC, sizeof(pReq->u.In.szMagic)))
1005	{
1006	OSDBGPRINT(("SUP_IOCTL_COOKIE: invalid magic %.16s\n", pReq->u.In.szMagic));
1007	pReq->Hdr.rc = VERR_INVALID_MAGIC;
1008	return 0;
1009	}
1010
1011	#if 0
1012	/*
1013	* Call out to the OS specific code and let it do permission checks on the
1014	* client process.
1015	*/
1016	if (!supdrvOSValidateClientProcess(pDevExt, pSession))
1017	{
1018	pReq->u.Out.u32Cookie = 0xffffffff;
1019	pReq->u.Out.u32SessionCookie = 0xffffffff;
1020	pReq->u.Out.u32SessionVersion = 0xffffffff;
1021	pReq->u.Out.u32DriverVersion = SUPDRV_IOC_VERSION;
1022	pReq->u.Out.pSession = NULL;
1023	pReq->u.Out.cFunctions = 0;
1024	pReq->Hdr.rc = VERR_PERMISSION_DENIED;
1025	return 0;
1026	}
1027	#endif
1028
1029	/*
1030	* Match the version.
1031	* The current logic is very simple, match the major interface version.
1032	*/
1033	if ( pReq->u.In.u32MinVersion > SUPDRV_IOC_VERSION
1034	\|\| (pReq->u.In.u32MinVersion & 0xffff0000) != (SUPDRV_IOC_VERSION & 0xffff0000))
1035	{
1036	OSDBGPRINT(("SUP_IOCTL_COOKIE: Version mismatch. Requested: %#x Min: %#x Current: %#x\n",
1037	pReq->u.In.u32ReqVersion, pReq->u.In.u32MinVersion, SUPDRV_IOC_VERSION));
1038	pReq->u.Out.u32Cookie = 0xffffffff;
1039	pReq->u.Out.u32SessionCookie = 0xffffffff;
1040	pReq->u.Out.u32SessionVersion = 0xffffffff;
1041	pReq->u.Out.u32DriverVersion = SUPDRV_IOC_VERSION;
1042	pReq->u.Out.pSession = NULL;
1043	pReq->u.Out.cFunctions = 0;
1044	pReq->Hdr.rc = VERR_VERSION_MISMATCH;
1045	return 0;
1046	}
1047
1048	/*
1049	* Fill in return data and be gone.
1050	* N.B. The first one to change SUPDRV_IOC_VERSION shall makes sure that
1051	* u32SessionVersion <= u32ReqVersion!
1052	*/
1053	/** @todo Somehow validate the client and negotiate a secure cookie... */
1054	pReq->u.Out.u32Cookie = pDevExt->u32Cookie;
1055	pReq->u.Out.u32SessionCookie = pSession->u32Cookie;
1056	pReq->u.Out.u32SessionVersion = SUPDRV_IOC_VERSION;
1057	pReq->u.Out.u32DriverVersion = SUPDRV_IOC_VERSION;
1058	pReq->u.Out.pSession = pSession;
1059	pReq->u.Out.cFunctions = sizeof(g_aFunctions) / sizeof(g_aFunctions[0]);
1060	pReq->Hdr.rc = VINF_SUCCESS;
1061	return 0;
1062	}
1063
1064	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_QUERY_FUNCS(0)):
1065	{
1066	/* validate */
1067	PSUPQUERYFUNCS pReq = (PSUPQUERYFUNCS)pReqHdr;
1068	REQ_CHECK_SIZES_EX(SUP_IOCTL_QUERY_FUNCS, SUP_IOCTL_QUERY_FUNCS_SIZE_IN, SUP_IOCTL_QUERY_FUNCS_SIZE_OUT(RT_ELEMENTS(g_aFunctions)));
1069
1070	/* execute */
1071	pReq->u.Out.cFunctions = RT_ELEMENTS(g_aFunctions);
1072	memcpy(&pReq->u.Out.aFunctions[0], g_aFunctions, sizeof(g_aFunctions));
1073	pReq->Hdr.rc = VINF_SUCCESS;
1074	return 0;
1075	}
1076
1077	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_IDT_INSTALL):
1078	{
1079	/* validate */
1080	PSUPIDTINSTALL pReq = (PSUPIDTINSTALL)pReqHdr;
1081	REQ_CHECK_SIZES(SUP_IOCTL_IDT_INSTALL);
1082
1083	/* execute */
1084	#ifdef VBOX_WITH_IDT_PATCHING
1085	pReq->Hdr.rc = supdrvIOCtl_IdtInstall(pDevExt, pSession, pReq);
1086	#else
1087	pReq->u.Out.u8Idt = 3;
1088	pReq->Hdr.rc = VERR_NOT_SUPPORTED;
1089	#endif
1090	return 0;
1091	}
1092
1093	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_IDT_REMOVE):
1094	{
1095	/* validate */
1096	PSUPIDTREMOVE pReq = (PSUPIDTREMOVE)pReqHdr;
1097	REQ_CHECK_SIZES(SUP_IOCTL_IDT_REMOVE);
1098
1099	/* execute */
1100	#ifdef VBOX_WITH_IDT_PATCHING
1101	pReq->Hdr.rc = supdrvIOCtl_IdtRemoveAll(pDevExt, pSession);
1102	#else
1103	pReq->Hdr.rc = VERR_NOT_SUPPORTED;
1104	#endif
1105	return 0;
1106	}
1107
1108	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_LOCK):
1109	{
1110	/* validate */
1111	PSUPPAGELOCK pReq = (PSUPPAGELOCK)pReqHdr;
1112	REQ_CHECK_SIZE_IN(SUP_IOCTL_PAGE_LOCK, SUP_IOCTL_PAGE_LOCK_SIZE_IN);
1113	REQ_CHECK_SIZE_OUT(SUP_IOCTL_PAGE_LOCK, SUP_IOCTL_PAGE_LOCK_SIZE_OUT(pReq->u.In.cPages));
1114	REQ_CHECK_EXPR(SUP_IOCTL_PAGE_LOCK, pReq->u.In.cPages > 0);
1115	REQ_CHECK_EXPR(SUP_IOCTL_PAGE_LOCK, pReq->u.In.pvR3 >= PAGE_SIZE);
1116
1117	/* execute */
1118	pReq->Hdr.rc = SUPR0LockMem(pSession, pReq->u.In.pvR3, pReq->u.In.cPages, &pReq->u.Out.aPages[0]);
1119	if (RT_FAILURE(pReq->Hdr.rc))
1120	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1121	return 0;
1122	}
1123
1124	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_UNLOCK):
1125	{
1126	/* validate */
1127	PSUPPAGEUNLOCK pReq = (PSUPPAGEUNLOCK)pReqHdr;
1128	REQ_CHECK_SIZES(SUP_IOCTL_PAGE_UNLOCK);
1129
1130	/* execute */
1131	pReq->Hdr.rc = SUPR0UnlockMem(pSession, pReq->u.In.pvR3);
1132	return 0;
1133	}
1134
1135	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_CONT_ALLOC):
1136	{
1137	/* validate */
1138	PSUPCONTALLOC pReq = (PSUPCONTALLOC)pReqHdr;
1139	REQ_CHECK_SIZES(SUP_IOCTL_CONT_ALLOC);
1140
1141	/* execute */
1142	pReq->Hdr.rc = SUPR0ContAlloc(pSession, pReq->u.In.cPages, &pReq->u.Out.pvR0, &pReq->u.Out.pvR3, &pReq->u.Out.HCPhys);
1143	if (RT_FAILURE(pReq->Hdr.rc))
1144	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1145	return 0;
1146	}
1147
1148	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_CONT_FREE):
1149	{
1150	/* validate */
1151	PSUPCONTFREE pReq = (PSUPCONTFREE)pReqHdr;
1152	REQ_CHECK_SIZES(SUP_IOCTL_CONT_FREE);
1153
1154	/* execute */
1155	pReq->Hdr.rc = SUPR0ContFree(pSession, (RTHCUINTPTR)pReq->u.In.pvR3);
1156	return 0;
1157	}
1158
1159	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LDR_OPEN):
1160	{
1161	/* validate */
1162	PSUPLDROPEN pReq = (PSUPLDROPEN)pReqHdr;
1163	REQ_CHECK_SIZES(SUP_IOCTL_LDR_OPEN);
1164	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.cbImage > 0);
1165	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.cbImage < _1M*16);
1166	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.szName[0]);
1167	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, memchr(pReq->u.In.szName, '\0', sizeof(pReq->u.In.szName)));
1168	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, !supdrvCheckInvalidChar(pReq->u.In.szName, ";:()[]{}/\\\|&*%#@!~`\"'"));
1169
1170	/* execute */
1171	pReq->Hdr.rc = supdrvIOCtl_LdrOpen(pDevExt, pSession, pReq);
1172	return 0;
1173	}
1174
1175	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LDR_LOAD):
1176	{
1177	/* validate */
1178	PSUPLDRLOAD pReq = (PSUPLDRLOAD)pReqHdr;
1179	REQ_CHECK_EXPR(Name, pReq->Hdr.cbIn >= sizeof(*pReq));
1180	REQ_CHECK_SIZES_EX(SUP_IOCTL_LDR_LOAD, SUP_IOCTL_LDR_LOAD_SIZE_IN(pReq->u.In.cbImage), SUP_IOCTL_LDR_LOAD_SIZE_OUT);
1181	REQ_CHECK_EXPR(SUP_IOCTL_LDR_LOAD, pReq->u.In.cSymbols <= 16384);
1182	REQ_CHECK_EXPR_FMT( !pReq->u.In.cSymbols
1183	\|\| ( pReq->u.In.offSymbols < pReq->u.In.cbImage
1184	&& pReq->u.In.offSymbols + pReq->u.In.cSymbols * sizeof(SUPLDRSYM) <= pReq->u.In.cbImage),
1185	("SUP_IOCTL_LDR_LOAD: offSymbols=%#lx cSymbols=%#lx cbImage=%#lx\n", (long)pReq->u.In.offSymbols,
1186	(long)pReq->u.In.cSymbols, (long)pReq->u.In.cbImage));
1187	REQ_CHECK_EXPR_FMT( !pReq->u.In.cbStrTab
1188	\|\| ( pReq->u.In.offStrTab < pReq->u.In.cbImage
1189	&& pReq->u.In.offStrTab + pReq->u.In.cbStrTab <= pReq->u.In.cbImage
1190	&& pReq->u.In.cbStrTab <= pReq->u.In.cbImage),
1191	("SUP_IOCTL_LDR_LOAD: offStrTab=%#lx cbStrTab=%#lx cbImage=%#lx\n", (long)pReq->u.In.offStrTab,
1192	(long)pReq->u.In.cbStrTab, (long)pReq->u.In.cbImage));
1193
1194	if (pReq->u.In.cSymbols)
1195	{
1196	uint32_t i;
1197	PSUPLDRSYM paSyms = (PSUPLDRSYM)&pReq->u.In.achImage[pReq->u.In.offSymbols];
1198	for (i = 0; i < pReq->u.In.cSymbols; i++)
1199	{
1200	REQ_CHECK_EXPR_FMT(paSyms[i].offSymbol < pReq->u.In.cbImage,
1201	("SUP_IOCTL_LDR_LOAD: sym #%ld: symb off %#lx (max=%#lx)\n", (long)i, (long)paSyms[i].offSymbol, (long)pReq->u.In.cbImage));
1202	REQ_CHECK_EXPR_FMT(paSyms[i].offName < pReq->u.In.cbStrTab,
1203	("SUP_IOCTL_LDR_LOAD: sym #%ld: name off %#lx (max=%#lx)\n", (long)i, (long)paSyms[i].offName, (long)pReq->u.In.cbImage));
1204	REQ_CHECK_EXPR_FMT(memchr(&pReq->u.In.achImage[pReq->u.In.offStrTab + paSyms[i].offName], '\0', pReq->u.In.cbStrTab - paSyms[i].offName),
1205	("SUP_IOCTL_LDR_LOAD: sym #%ld: unterminated name! (%#lx / %#lx)\n", (long)i, (long)paSyms[i].offName, (long)pReq->u.In.cbImage));
1206	}
1207	}
1208
1209	/* execute */
1210	pReq->Hdr.rc = supdrvIOCtl_LdrLoad(pDevExt, pSession, pReq);
1211	return 0;
1212	}
1213
1214	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LDR_FREE):
1215	{
1216	/* validate */
1217	PSUPLDRFREE pReq = (PSUPLDRFREE)pReqHdr;
1218	REQ_CHECK_SIZES(SUP_IOCTL_LDR_FREE);
1219
1220	/* execute */
1221	pReq->Hdr.rc = supdrvIOCtl_LdrFree(pDevExt, pSession, pReq);
1222	return 0;
1223	}
1224
1225	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LDR_GET_SYMBOL):
1226	{
1227	/* validate */
1228	PSUPLDRGETSYMBOL pReq = (PSUPLDRGETSYMBOL)pReqHdr;
1229	REQ_CHECK_SIZES(SUP_IOCTL_LDR_GET_SYMBOL);
1230	REQ_CHECK_EXPR(SUP_IOCTL_LDR_GET_SYMBOL, memchr(pReq->u.In.szSymbol, '\0', sizeof(pReq->u.In.szSymbol)));
1231
1232	/* execute */
1233	pReq->Hdr.rc = supdrvIOCtl_LdrGetSymbol(pDevExt, pSession, pReq);
1234	return 0;
1235	}
1236
1237	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_CALL_VMMR0(0)):
1238	{
1239	/* validate */
1240	PSUPCALLVMMR0 pReq = (PSUPCALLVMMR0)pReqHdr;
1241	Log4(("SUP_IOCTL_CALL_VMMR0: op=%u in=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1242	pReq->u.In.uOperation, pReq->Hdr.cbIn, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1243
1244	if (pReq->Hdr.cbIn == SUP_IOCTL_CALL_VMMR0_SIZE(0))
1245	{
1246	REQ_CHECK_SIZES_EX(SUP_IOCTL_CALL_VMMR0, SUP_IOCTL_CALL_VMMR0_SIZE_IN(0), SUP_IOCTL_CALL_VMMR0_SIZE_OUT(0));
1247
1248	/* execute */
1249	if (RT_LIKELY(pDevExt->pfnVMMR0EntryEx))
1250	#ifdef SUPDRV_WITH_UNWIND_HACK
1251	pReq->Hdr.rc = supdrvNtWrapVMMR0EntryEx((PFNRT)pDevExt->pfnVMMR0EntryEx, pReq->u.In.pVMR0, pReq->u.In.uOperation, NULL, pReq->u.In.u64Arg, pSession);
1252	#else
1253	pReq->Hdr.rc = pDevExt->pfnVMMR0EntryEx(pReq->u.In.pVMR0, pReq->u.In.uOperation, NULL, pReq->u.In.u64Arg, pSession);
1254	#endif
1255	else
1256	pReq->Hdr.rc = VERR_WRONG_ORDER;
1257	}
1258	else
1259	{
1260	PSUPVMMR0REQHDR pVMMReq = (PSUPVMMR0REQHDR)&pReq->abReqPkt[0];
1261	REQ_CHECK_EXPR_FMT(pReq->Hdr.cbIn >= SUP_IOCTL_CALL_VMMR0_SIZE(sizeof(SUPVMMR0REQHDR)),
1262	("SUP_IOCTL_CALL_VMMR0: cbIn=%#x < %#x\n", pReq->Hdr.cbIn, SUP_IOCTL_CALL_VMMR0_SIZE(sizeof(SUPVMMR0REQHDR))));
1263	REQ_CHECK_EXPR(SUP_IOCTL_CALL_VMMR0, pVMMReq->u32Magic == SUPVMMR0REQHDR_MAGIC);
1264	REQ_CHECK_SIZES_EX(SUP_IOCTL_CALL_VMMR0, SUP_IOCTL_CALL_VMMR0_SIZE_IN(pVMMReq->cbReq), SUP_IOCTL_CALL_VMMR0_SIZE_OUT(pVMMReq->cbReq));
1265
1266	/* execute */
1267	if (RT_LIKELY(pDevExt->pfnVMMR0EntryEx))
1268	#ifdef SUPDRV_WITH_UNWIND_HACK
1269	pReq->Hdr.rc = supdrvNtWrapVMMR0EntryEx((PFNRT)pDevExt->pfnVMMR0EntryEx, pReq->u.In.pVMR0, pReq->u.In.uOperation, pVMMReq, pReq->u.In.u64Arg, pSession);
1270	#else
1271	pReq->Hdr.rc = pDevExt->pfnVMMR0EntryEx(pReq->u.In.pVMR0, pReq->u.In.uOperation, pVMMReq, pReq->u.In.u64Arg, pSession);
1272	#endif
1273	else
1274	pReq->Hdr.rc = VERR_WRONG_ORDER;
1275	}
1276
1277	if ( RT_FAILURE(pReq->Hdr.rc)
1278	&& pReq->Hdr.rc != VERR_INTERRUPTED
1279	&& pReq->Hdr.rc != VERR_TIMEOUT)
1280	Log(("SUP_IOCTL_CALL_VMMR0: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1281	pReq->Hdr.rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1282	else
1283	Log4(("SUP_IOCTL_CALL_VMMR0: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1284	pReq->Hdr.rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1285	return 0;
1286	}
1287
1288	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_GET_PAGING_MODE):
1289	{
1290	/* validate */
1291	PSUPGETPAGINGMODE pReq = (PSUPGETPAGINGMODE)pReqHdr;
1292	REQ_CHECK_SIZES(SUP_IOCTL_GET_PAGING_MODE);
1293
1294	/* execute */
1295	pReq->Hdr.rc = VINF_SUCCESS;
1296	pReq->u.Out.enmMode = supdrvIOCtl_GetPagingMode();
1297	return 0;
1298	}
1299
1300	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LOW_ALLOC):
1301	{
1302	/* validate */
1303	PSUPLOWALLOC pReq = (PSUPLOWALLOC)pReqHdr;
1304	REQ_CHECK_EXPR(SUP_IOCTL_LOW_ALLOC, pReq->Hdr.cbIn <= SUP_IOCTL_LOW_ALLOC_SIZE_IN);
1305	REQ_CHECK_SIZES_EX(SUP_IOCTL_LOW_ALLOC, SUP_IOCTL_LOW_ALLOC_SIZE_IN, SUP_IOCTL_LOW_ALLOC_SIZE_OUT(pReq->u.In.cPages));
1306
1307	/* execute */
1308	pReq->Hdr.rc = SUPR0LowAlloc(pSession, pReq->u.In.cPages, &pReq->u.Out.pvR0, &pReq->u.Out.pvR3, &pReq->u.Out.aPages[0]);
1309	if (RT_FAILURE(pReq->Hdr.rc))
1310	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1311	return 0;
1312	}
1313
1314	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LOW_FREE):
1315	{
1316	/* validate */
1317	PSUPLOWFREE pReq = (PSUPLOWFREE)pReqHdr;
1318	REQ_CHECK_SIZES(SUP_IOCTL_LOW_FREE);
1319
1320	/* execute */
1321	pReq->Hdr.rc = SUPR0LowFree(pSession, (RTHCUINTPTR)pReq->u.In.pvR3);
1322	return 0;
1323	}
1324
1325	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_GIP_MAP):
1326	{
1327	/* validate */
1328	PSUPGIPMAP pReq = (PSUPGIPMAP)pReqHdr;
1329	REQ_CHECK_SIZES(SUP_IOCTL_GIP_MAP);
1330
1331	/* execute */
1332	pReq->Hdr.rc = SUPR0GipMap(pSession, &pReq->u.Out.pGipR3, &pReq->u.Out.HCPhysGip);
1333	if (RT_SUCCESS(pReq->Hdr.rc))
1334	pReq->u.Out.pGipR0 = pDevExt->pGip;
1335	return 0;
1336	}
1337
1338	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_GIP_UNMAP):
1339	{
1340	/* validate */
1341	PSUPGIPUNMAP pReq = (PSUPGIPUNMAP)pReqHdr;
1342	REQ_CHECK_SIZES(SUP_IOCTL_GIP_UNMAP);
1343
1344	/* execute */
1345	pReq->Hdr.rc = SUPR0GipUnmap(pSession);
1346	return 0;
1347	}
1348
1349	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_SET_VM_FOR_FAST):
1350	{
1351	/* validate */
1352	PSUPSETVMFORFAST pReq = (PSUPSETVMFORFAST)pReqHdr;
1353	REQ_CHECK_SIZES(SUP_IOCTL_SET_VM_FOR_FAST);
1354	REQ_CHECK_EXPR_FMT( !pReq->u.In.pVMR0
1355	\|\| ( VALID_PTR(pReq->u.In.pVMR0)
1356	&& !((uintptr_t)pReq->u.In.pVMR0 & (PAGE_SIZE - 1))),
1357	("SUP_IOCTL_SET_VM_FOR_FAST: pVMR0=%p!\n", pReq->u.In.pVMR0));
1358	/* execute */
1359	pSession->pVM = pReq->u.In.pVMR0;
1360	pReq->Hdr.rc = VINF_SUCCESS;
1361	return 0;
1362	}
1363
1364	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_ALLOC):
1365	{
1366	/* validate */
1367	PSUPPAGEALLOC pReq = (PSUPPAGEALLOC)pReqHdr;
1368	REQ_CHECK_EXPR(SUP_IOCTL_PAGE_ALLOC, pReq->Hdr.cbIn <= SUP_IOCTL_PAGE_ALLOC_SIZE_IN);
1369	REQ_CHECK_SIZES_EX(SUP_IOCTL_PAGE_ALLOC, SUP_IOCTL_PAGE_ALLOC_SIZE_IN, SUP_IOCTL_PAGE_ALLOC_SIZE_OUT(pReq->u.In.cPages));
1370
1371	/* execute */
1372	pReq->Hdr.rc = SUPR0PageAlloc(pSession, pReq->u.In.cPages, &pReq->u.Out.pvR3, &pReq->u.Out.aPages[0]);
1373	if (RT_FAILURE(pReq->Hdr.rc))
1374	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1375	return 0;
1376	}
1377
1378	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_FREE):
1379	{
1380	/* validate */
1381	PSUPPAGEFREE pReq = (PSUPPAGEFREE)pReqHdr;
1382	REQ_CHECK_SIZES(SUP_IOCTL_PAGE_FREE);
1383
1384	/* execute */
1385	pReq->Hdr.rc = SUPR0PageFree(pSession, pReq->u.In.pvR3);
1386	return 0;
1387	}
1388
1389	default:
1390	Log(("Unknown IOCTL %#lx\n", (long)uIOCtl));
1391	break;
1392	}
1393	return SUPDRV_ERR_GENERAL_FAILURE;
1394	}
1395
1396
1397	/**
1398	* Inter-Driver Communcation (IDC) worker.
1399	*
1400	* @returns VBox status code.
1401	* @retval VINF_SUCCESS on success.
1402	* @retval VERR_INVALID_PARAMETER if the request is invalid.
1403	* @retval VERR_NOT_SUPPORTED if the request isn't supported.
1404	*
1405	* @param uReq The request (function) code.
1406	* @param pDevExt Device extention.
1407	* @param pSession Session data.
1408	* @param pReqHdr The request header.
1409	*/
1410	int VBOXCALL supdrvIDC(uintptr_t uReq, PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPDRVIDCREQHDR pReqHdr)
1411	{
1412	/*
1413	* The OS specific code has already validated the pSession
1414	* pointer, and the request size being greater or equal to
1415	* size of the header.
1416	*
1417	* So, just check that pSession is a kernel context session.
1418	*/
1419	if (RT_UNLIKELY( pSession
1420	&& pSession->R0Process != NIL_RTR0PROCESS))
1421	return VERR_INVALID_PARAMETER;
1422
1423	/*
1424	* Validation macro.
1425	*/
1426	#define REQ_CHECK_IDC_SIZE(Name, cbExpect) \
1427	do { \
1428	if (RT_UNLIKELY(pReqHdr->cb != (cbExpect))) \
1429	{ \
1430	OSDBGPRINT(( #Name ": Invalid input/output sizes. cb=%ld expected %ld.\n", \
1431	(long)pReqHdr->cb, (long)(cbExpect))); \
1432	return pReqHdr->rc = VERR_INVALID_PARAMETER; \
1433	} \
1434	} while (0)
1435
1436	switch (uReq)
1437	{
1438	case SUPDRV_IDC_REQ_CONNECT:
1439	{
1440	PSUPDRVIDCREQCONNECT pReq = (PSUPDRVIDCREQCONNECT)pReqHdr;
1441	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_CONNECT, sizeof(*pReq));
1442
1443	/*
1444	* Validate the cookie and other input.
1445	*/
1446	if (pReq->Hdr.pSession != NULL)
1447	{
1448	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: pSession=%p expected NULL!\n", pReq->Hdr.pSession));
1449	return pReqHdr->rc = VERR_INVALID_PARAMETER;
1450	}
1451	if (pReq->u.In.u32MagicCookie != SUPDRVIDCREQ_CONNECT_MAGIC_COOKIE)
1452	{
1453	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: u32MagicCookie=%#x expected %#x!\n",
1454	pReq->u.In.u32MagicCookie, SUPDRVIDCREQ_CONNECT_MAGIC_COOKIE));
1455	return pReqHdr->rc = VERR_INVALID_PARAMETER;
1456	}
1457	if ( pReq->u.In.uMinVersion > pReq->u.In.uReqVersion
1458	\|\| (pReq->u.In.uMinVersion & UINT32_C(0xffff0000)) != (pReq->u.In.uReqVersion & UINT32_C(0xffff0000)))
1459	{
1460	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: uMinVersion=%#x uMaxVersion=%#x doesn't match!\n",
1461	pReq->u.In.uMinVersion, pReq->u.In.uReqVersion));
1462	return pReqHdr->rc = VERR_INVALID_PARAMETER;
1463	}
1464
1465	/*
1466	* Match the version.
1467	* The current logic is very simple, match the major interface version.
1468	*/
1469	if ( pReq->u.In.uMinVersion > SUPDRV_IDC_VERSION
1470	\|\| (pReq->u.In.uMinVersion & 0xffff0000) != (SUPDRV_IDC_VERSION & 0xffff0000))
1471	{
1472	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: Version mismatch. Requested: %#x Min: %#x Current: %#x\n",
1473	pReq->u.In.uReqVersion, pReq->u.In.uMinVersion, SUPDRV_IDC_VERSION));
1474	pReq->u.Out.pSession = NULL;
1475	pReq->u.Out.uSessionVersion = 0xffffffff;
1476	pReq->u.Out.uDriverVersion = SUPDRV_IDC_VERSION;
1477	pReq->u.Out.uDriverRevision = VBOX_SVN_REV;
1478	pReq->Hdr.rc = VERR_VERSION_MISMATCH;
1479	return VINF_SUCCESS;
1480	}
1481
1482	pReq->u.Out.pSession = NULL;
1483	pReq->u.Out.uSessionVersion = SUPDRV_IDC_VERSION;
1484	pReq->u.Out.uDriverVersion = SUPDRV_IDC_VERSION;
1485	pReq->u.Out.uDriverRevision = VBOX_SVN_REV;
1486
1487	/*
1488	* On NT we will already have a session associated with the
1489	* client, just like with the SUP_IOCTL_COOKIE request, while
1490	* the other doesn't.
1491	*/
1492	#ifdef RT_OS_WINDOWS
1493	pReq->Hdr.rc = VINF_SUCCESS;
1494	#else
1495	AssertReturn(!pSession, VERR_INTERNAL_ERROR);
1496	pReq->Hdr.rc = supdrvCreateSession(pDevExt, false /* fUser */, &pSession);
1497	if (RT_FAILURE(pReq->Hdr.rc))
1498	{
1499	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: failed to create session, rc=%d\n", pReq->Hdr.rc));
1500	return VINF_SUCCESS;
1501	}
1502	#endif
1503
1504	pReq->u.Out.pSession = pSession;
1505	pReq->Hdr.pSession = pSession;
1506
1507	return VINF_SUCCESS;
1508	}
1509
1510	case SUPDRV_IDC_REQ_DISCONNECT:
1511	{
1512	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_DISCONNECT, sizeof(*pReqHdr));
1513
1514	#ifdef RT_OS_WINDOWS
1515	/* Windows will destroy the session when the file object is destroyed. */
1516	#else
1517	supdrvCloseSession(pDevExt, pSession);
1518	#endif
1519	return pReqHdr->rc = VINF_SUCCESS;
1520	}
1521
1522	case SUPDRV_IDC_REQ_GET_SYMBOL:
1523	{
1524	PSUPDRVIDCREQGETSYM pReq = (PSUPDRVIDCREQGETSYM)pReqHdr;
1525	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_GET_SYMBOL, sizeof(*pReq));
1526
1527	pReq->Hdr.rc = supdrvIDC_LdrGetSymbol(pDevExt, pSession, pReq);
1528	return VINF_SUCCESS;
1529	}
1530
1531	case SUPDRV_IDC_REQ_COMPONENT_REGISTER_FACTORY:
1532	{
1533	PSUPDRVIDCREQCOMPREGFACTORY pReq = (PSUPDRVIDCREQCOMPREGFACTORY)pReqHdr;
1534	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_COMPONENT_REGISTER_FACTORY, sizeof(*pReq));
1535
1536	pReq->Hdr.rc = SUPR0ComponentRegisterFactory(pSession, pReq->u.In.pFactory);
1537	return VINF_SUCCESS;
1538	}
1539
1540	case SUPDRV_IDC_REQ_COMPONENT_DEREGISTER_FACTORY:
1541	{
1542	PSUPDRVIDCREQCOMPDEREGFACTORY pReq = (PSUPDRVIDCREQCOMPDEREGFACTORY)pReqHdr;
1543	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_COMPONENT_DEREGISTER_FACTORY, sizeof(*pReq));
1544
1545	pReq->Hdr.rc = SUPR0ComponentDeregisterFactory(pSession, pReq->u.In.pFactory);
1546	return VINF_SUCCESS;
1547	}
1548
1549	default:
1550	Log(("Unknown IDC %#lx\n", (long)uReq));
1551	break;
1552	}
1553
1554	#undef REQ_CHECK_IDC_SIZE
1555	return VERR_NOT_SUPPORTED;
1556	}
1557
1558
1559	/**
1560	* Register a object for reference counting.
1561	* The object is registered with one reference in the specified session.
1562	*
1563	* @returns Unique identifier on success (pointer).
1564	* All future reference must use this identifier.
1565	* @returns NULL on failure.
1566	* @param pfnDestructor The destructore function which will be called when the reference count reaches 0.
1567	* @param pvUser1 The first user argument.
1568	* @param pvUser2 The second user argument.
1569	*/
1570	SUPR0DECL(void ) SUPR0ObjRegister(PSUPDRVSESSION pSession, SUPDRVOBJTYPE enmType, PFNSUPDRVDESTRUCTOR pfnDestructor, void pvUser1, void *pvUser2)
1571	{
1572	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
1573	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
1574	PSUPDRVOBJ pObj;
1575	PSUPDRVUSAGE pUsage;
1576
1577	/*
1578	* Validate the input.
1579	*/
1580	AssertReturn(SUP_IS_SESSION_VALID(pSession), NULL);
1581	AssertReturn(enmType > SUPDRVOBJTYPE_INVALID && enmType < SUPDRVOBJTYPE_END, NULL);
1582	AssertPtrReturn(pfnDestructor, NULL);
1583
1584	/*
1585	* Allocate and initialize the object.
1586	*/
1587	pObj = (PSUPDRVOBJ)RTMemAlloc(sizeof(*pObj));
1588	if (!pObj)
1589	return NULL;
1590	pObj->u32Magic = SUPDRVOBJ_MAGIC;
1591	pObj->enmType = enmType;
1592	pObj->pNext = NULL;
1593	pObj->cUsage = 1;
1594	pObj->pfnDestructor = pfnDestructor;
1595	pObj->pvUser1 = pvUser1;
1596	pObj->pvUser2 = pvUser2;
1597	pObj->CreatorUid = pSession->Uid;
1598	pObj->CreatorGid = pSession->Gid;
1599	pObj->CreatorProcess= pSession->Process;
1600	supdrvOSObjInitCreator(pObj, pSession);
1601
1602	/*
1603	* Allocate the usage record.
1604	* (We keep freed usage records around to simplify SUPR0ObjAddRef().)
1605	*/
1606	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
1607
1608	pUsage = pDevExt->pUsageFree;
1609	if (pUsage)
1610	pDevExt->pUsageFree = pUsage->pNext;
1611	else
1612	{
1613	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
1614	pUsage = (PSUPDRVUSAGE)RTMemAlloc(sizeof(*pUsage));
1615	if (!pUsage)
1616	{
1617	RTMemFree(pObj);
1618	return NULL;
1619	}
1620	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
1621	}
1622
1623	/*
1624	* Insert the object and create the session usage record.
1625	*/
1626	/* The object. */
1627	pObj->pNext = pDevExt->pObjs;
1628	pDevExt->pObjs = pObj;
1629
1630	/* The session record. */
1631	pUsage->cUsage = 1;
1632	pUsage->pObj = pObj;
1633	pUsage->pNext = pSession->pUsage;
1634	/* Log2(("SUPR0ObjRegister: pUsage=%p:{.pObj=%p, .pNext=%p}\n", pUsage, pUsage->pObj, pUsage->pNext)); */
1635	pSession->pUsage = pUsage;
1636
1637	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
1638
1639	Log(("SUPR0ObjRegister: returns %p (pvUser1=%p, pvUser=%p)\n", pObj, pvUser1, pvUser2));
1640	return pObj;
1641	}
1642
1643
1644	/**
1645	* Increment the reference counter for the object associating the reference
1646	* with the specified session.
1647	*
1648	* @returns IPRT status code.
1649	* @param pvObj The identifier returned by SUPR0ObjRegister().
1650	* @param pSession The session which is referencing the object.
1651	*
1652	* @remarks The caller should not own any spinlocks and must carefully protect
1653	* itself against potential race with the destructor so freed memory
1654	* isn't accessed here.
1655	*/
1656	SUPR0DECL(int) SUPR0ObjAddRef(void *pvObj, PSUPDRVSESSION pSession)
1657	{
1658	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
1659	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
1660	PSUPDRVOBJ pObj = (PSUPDRVOBJ)pvObj;
1661	PSUPDRVUSAGE pUsagePre;
1662	PSUPDRVUSAGE pUsage;
1663
1664	/*
1665	* Validate the input.
1666	* Be ready for the destruction race (someone might be stuck in the
1667	* destructor waiting a lock we own).
1668	*/
1669	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
1670	AssertPtrReturn(pObj, VERR_INVALID_POINTER);
1671	AssertMsgReturn(pObj->u32Magic == SUPDRVOBJ_MAGIC \|\| pObj->u32Magic == SUPDRVOBJ_MAGIC + 1,
1672	("Invalid pvObj=%p magic=%#x (expected %#x or %#x)\n", pvObj, pObj->u32Magic, SUPDRVOBJ_MAGIC, SUPDRVOBJ_MAGIC + 1),
1673	VERR_INVALID_PARAMETER);
1674
1675	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
1676
1677	if (RT_UNLIKELY(pObj->u32Magic != SUPDRVOBJ_MAGIC))
1678	{
1679	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
1680
1681	AssertMsgFailed(("pvObj=%p magic=%#x\n", pvObj, pObj->u32Magic));
1682	return VERR_WRONG_ORDER;
1683	}
1684
1685	/*
1686	* Preallocate the usage record.
1687	*/
1688	pUsagePre = pDevExt->pUsageFree;
1689	if (pUsagePre)
1690	pDevExt->pUsageFree = pUsagePre->pNext;
1691	else
1692	{
1693	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
1694	pUsagePre = (PSUPDRVUSAGE)RTMemAlloc(sizeof(*pUsagePre));
1695	if (!pUsagePre)
1696	return VERR_NO_MEMORY;
1697
1698	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
1699	}
1700
1701	/*
1702	* Reference the object.
1703	*/
1704	pObj->cUsage++;
1705
1706	/*
1707	* Look for the session record.
1708	*/
1709	for (pUsage = pSession->pUsage; pUsage; pUsage = pUsage->pNext)
1710	{
1711	/Log(("SUPR0AddRef: pUsage=%p:{.pObj=%p, .pNext=%p}\n", pUsage, pUsage->pObj, pUsage->pNext));/
1712	if (pUsage->pObj == pObj)
1713	break;
1714	}
1715	if (pUsage)
1716	pUsage->cUsage++;
1717	else
1718	{
1719	/* create a new session record. */
1720	pUsagePre->cUsage = 1;
1721	pUsagePre->pObj = pObj;
1722	pUsagePre->pNext = pSession->pUsage;
1723	pSession->pUsage = pUsagePre;
1724	/Log(("SUPR0AddRef: pUsagePre=%p:{.pObj=%p, .pNext=%p}\n", pUsagePre, pUsagePre->pObj, pUsagePre->pNext));/
1725
1726	pUsagePre = NULL;
1727	}
1728
1729	/*
1730	* Put any unused usage record into the free list..
1731	*/
1732	if (pUsagePre)
1733	{
1734	pUsagePre->pNext = pDevExt->pUsageFree;
1735	pDevExt->pUsageFree = pUsagePre;
1736	}
1737
1738	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
1739
1740	return VINF_SUCCESS;
1741	}
1742
1743
1744	/**
1745	* Decrement / destroy a reference counter record for an object.
1746	*
1747	* The object is uniquely identified by pfnDestructor+pvUser1+pvUser2.
1748	*
1749	* @returns IPRT status code.
1750	* @param pvObj The identifier returned by SUPR0ObjRegister().
1751	* @param pSession The session which is referencing the object.
1752	*/
1753	SUPR0DECL(int) SUPR0ObjRelease(void *pvObj, PSUPDRVSESSION pSession)
1754	{
1755	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
1756	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
1757	PSUPDRVOBJ pObj = (PSUPDRVOBJ)pvObj;
1758	bool fDestroy = false;
1759	PSUPDRVUSAGE pUsage;
1760	PSUPDRVUSAGE pUsagePrev;
1761
1762	/*
1763	* Validate the input.
1764	*/
1765	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
1766	AssertMsgReturn(VALID_PTR(pObj) && pObj->u32Magic == SUPDRVOBJ_MAGIC,
1767	("Invalid pvObj=%p magic=%#x (exepcted %#x)\n", pvObj, pObj ? pObj->u32Magic : 0, SUPDRVOBJ_MAGIC),
1768	VERR_INVALID_PARAMETER);
1769
1770	/*
1771	* Acquire the spinlock and look for the usage record.
1772	*/
1773	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
1774
1775	for (pUsagePrev = NULL, pUsage = pSession->pUsage;
1776	pUsage;
1777	pUsagePrev = pUsage, pUsage = pUsage->pNext)
1778	{
1779	/Log2(("SUPR0ObjRelease: pUsage=%p:{.pObj=%p, .pNext=%p}\n", pUsage, pUsage->pObj, pUsage->pNext));/
1780	if (pUsage->pObj == pObj)
1781	{
1782	AssertMsg(pUsage->cUsage >= 1 && pObj->cUsage >= pUsage->cUsage, ("glob %d; sess %d\n", pObj->cUsage, pUsage->cUsage));
1783	if (pUsage->cUsage > 1)
1784	{
1785	pObj->cUsage--;
1786	pUsage->cUsage--;
1787	}
1788	else
1789	{
1790	/*
1791	* Free the session record.
1792	*/
1793	if (pUsagePrev)
1794	pUsagePrev->pNext = pUsage->pNext;
1795	else
1796	pSession->pUsage = pUsage->pNext;
1797	pUsage->pNext = pDevExt->pUsageFree;
1798	pDevExt->pUsageFree = pUsage;
1799
1800	/* What about the object? */
1801	if (pObj->cUsage > 1)
1802	pObj->cUsage--;
1803	else
1804	{
1805	/*
1806	* Object is to be destroyed, unlink it.
1807	*/
1808	pObj->u32Magic = SUPDRVOBJ_MAGIC + 1;
1809	fDestroy = true;
1810	if (pDevExt->pObjs == pObj)
1811	pDevExt->pObjs = pObj->pNext;
1812	else
1813	{
1814	PSUPDRVOBJ pObjPrev;
1815	for (pObjPrev = pDevExt->pObjs; pObjPrev; pObjPrev = pObjPrev->pNext)
1816	if (pObjPrev->pNext == pObj)
1817	{
1818	pObjPrev->pNext = pObj->pNext;
1819	break;
1820	}
1821	Assert(pObjPrev);
1822	}
1823	}
1824	}
1825	break;
1826	}
1827	}
1828
1829	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
1830
1831	/*
1832	* Call the destructor and free the object if required.
1833	*/
1834	if (fDestroy)
1835	{
1836	Log(("SUPR0ObjRelease: destroying %p/%d (%p/%p) cpid=%RTproc pid=%RTproc dtor=%p\n",
1837	pObj, pObj->enmType, pObj->pvUser1, pObj->pvUser2, pObj->CreatorProcess, RTProcSelf(), pObj->pfnDestructor));
1838	if (pObj->pfnDestructor)
1839	#ifdef SUPDRV_WITH_UNWIND_HACK
1840	supdrvNtWrapObjDestructor((PFNRT)pObj->pfnDestructor, pObj, pObj->pvUser1, pObj->pvUser2);
1841	#else
1842	pObj->pfnDestructor(pObj, pObj->pvUser1, pObj->pvUser2);
1843	#endif
1844	RTMemFree(pObj);
1845	}
1846
1847	AssertMsg(pUsage, ("pvObj=%p\n", pvObj));
1848	return pUsage ? VINF_SUCCESS : VERR_INVALID_PARAMETER;
1849	}
1850
1851	/**
1852	* Verifies that the current process can access the specified object.
1853	*
1854	* @returns The following IPRT status code:
1855	* @retval VINF_SUCCESS if access was granted.
1856	* @retval VERR_PERMISSION_DENIED if denied access.
1857	* @retval VERR_INVALID_PARAMETER if invalid parameter.
1858	*
1859	* @param pvObj The identifier returned by SUPR0ObjRegister().
1860	* @param pSession The session which wishes to access the object.
1861	* @param pszObjName Object string name. This is optional and depends on the object type.
1862	*
1863	* @remark The caller is responsible for making sure the object isn't removed while
1864	* we're inside this function. If uncertain about this, just call AddRef before calling us.
1865	*/
1866	SUPR0DECL(int) SUPR0ObjVerifyAccess(void pvObj, PSUPDRVSESSION pSession, const char pszObjName)
1867	{
1868	PSUPDRVOBJ pObj = (PSUPDRVOBJ)pvObj;
1869	int rc;
1870
1871	/*
1872	* Validate the input.
1873	*/
1874	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
1875	AssertMsgReturn(VALID_PTR(pObj) && pObj->u32Magic == SUPDRVOBJ_MAGIC,
1876	("Invalid pvObj=%p magic=%#x (exepcted %#x)\n", pvObj, pObj ? pObj->u32Magic : 0, SUPDRVOBJ_MAGIC),
1877	VERR_INVALID_PARAMETER);
1878
1879	/*
1880	* Check access. (returns true if a decision has been made.)
1881	*/
1882	rc = VERR_INTERNAL_ERROR;
1883	if (supdrvOSObjCanAccess(pObj, pSession, pszObjName, &rc))
1884	return rc;
1885
1886	/*
1887	* Default policy is to allow the user to access his own
1888	* stuff but nothing else.
1889	*/
1890	if (pObj->CreatorUid == pSession->Uid)
1891	return VINF_SUCCESS;
1892	return VERR_PERMISSION_DENIED;
1893	}
1894
1895
1896	/**
1897	* Lock pages.
1898	*
1899	* @returns IPRT status code.
1900	* @param pSession Session to which the locked memory should be associated.
1901	* @param pvR3 Start of the memory range to lock.
1902	* This must be page aligned.
1903	* @param cb Size of the memory range to lock.
1904	* This must be page aligned.
1905	*/
1906	SUPR0DECL(int) SUPR0LockMem(PSUPDRVSESSION pSession, RTR3PTR pvR3, uint32_t cPages, PRTHCPHYS paPages)
1907	{
1908	int rc;
1909	SUPDRVMEMREF Mem = {0};
1910	const size_t cb = (size_t)cPages << PAGE_SHIFT;
1911	LogFlow(("SUPR0LockMem: pSession=%p pvR3=%p cPages=%d paPages=%p\n", pSession, (void *)pvR3, cPages, paPages));
1912
1913	/*
1914	* Verify input.
1915	*/
1916	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
1917	AssertPtrReturn(paPages, VERR_INVALID_PARAMETER);
1918	if ( RT_ALIGN_R3PT(pvR3, PAGE_SIZE, RTR3PTR) != pvR3
1919	\|\| !pvR3)
1920	{
1921	Log(("pvR3 (%p) must be page aligned and not NULL!\n", (void *)pvR3));
1922	return VERR_INVALID_PARAMETER;
1923	}
1924
1925	#ifdef RT_OS_WINDOWS /* A temporary hack for windows, will be removed once all ring-3 code has been cleaned up. */
1926	/* First check if we allocated it using SUPPageAlloc; if so then we don't need to lock it again */
1927	rc = supdrvPageGetPhys(pSession, pvR3, cPages, paPages);
1928	if (RT_SUCCESS(rc))
1929	return rc;
1930	#endif
1931
1932	/*
1933	* Let IPRT do the job.
1934	*/
1935	Mem.eType = MEMREF_TYPE_LOCKED;
1936	rc = RTR0MemObjLockUser(&Mem.MemObj, pvR3, cb, RTR0ProcHandleSelf());
1937	if (RT_SUCCESS(rc))
1938	{
1939	uint32_t iPage = cPages;
1940	AssertMsg(RTR0MemObjAddressR3(Mem.MemObj) == pvR3, ("%p == %p\n", RTR0MemObjAddressR3(Mem.MemObj), pvR3));
1941	AssertMsg(RTR0MemObjSize(Mem.MemObj) == cb, ("%x == %x\n", RTR0MemObjSize(Mem.MemObj), cb));
1942
1943	while (iPage-- > 0)
1944	{
1945	paPages[iPage] = RTR0MemObjGetPagePhysAddr(Mem.MemObj, iPage);
1946	if (RT_UNLIKELY(paPages[iPage] == NIL_RTCCPHYS))
1947	{
1948	AssertMsgFailed(("iPage=%d\n", iPage));
1949	rc = VERR_INTERNAL_ERROR;
1950	break;
1951	}
1952	}
1953	if (RT_SUCCESS(rc))
1954	rc = supdrvMemAdd(&Mem, pSession);
1955	if (RT_FAILURE(rc))
1956	{
1957	int rc2 = RTR0MemObjFree(Mem.MemObj, false);
1958	AssertRC(rc2);
1959	}
1960	}
1961
1962	return rc;
1963	}
1964
1965
1966	/**
1967	* Unlocks the memory pointed to by pv.
1968	*
1969	* @returns IPRT status code.
1970	* @param pSession Session to which the memory was locked.
1971	* @param pvR3 Memory to unlock.
1972	*/
1973	SUPR0DECL(int) SUPR0UnlockMem(PSUPDRVSESSION pSession, RTR3PTR pvR3)
1974	{
1975	LogFlow(("SUPR0UnlockMem: pSession=%p pvR3=%p\n", pSession, (void *)pvR3));
1976	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
1977	#ifdef RT_OS_WINDOWS
1978	/*
1979	* Temporary hack for windows - SUPR0PageFree will unlock SUPR0PageAlloc
1980	* allocations; ignore this call.
1981	*/
1982	if (supdrvPageWasLockedByPageAlloc(pSession, pvR3))
1983	{
1984	LogFlow(("Page will be unlocked in SUPR0PageFree -> ignore\n"));
1985	return VINF_SUCCESS;
1986	}
1987	#endif
1988	return supdrvMemRelease(pSession, (RTHCUINTPTR)pvR3, MEMREF_TYPE_LOCKED);
1989	}
1990
1991
1992	/**
1993	* Allocates a chunk of page aligned memory with contiguous and fixed physical
1994	* backing.
1995	*
1996	* @returns IPRT status code.
1997	* @param pSession Session data.
1998	* @param cb Number of bytes to allocate.
1999	* @param ppvR0 Where to put the address of Ring-0 mapping the allocated memory.
2000	* @param ppvR3 Where to put the address of Ring-3 mapping the allocated memory.
2001	* @param pHCPhys Where to put the physical address of allocated memory.
2002	*/
2003	SUPR0DECL(int) SUPR0ContAlloc(PSUPDRVSESSION pSession, uint32_t cPages, PRTR0PTR ppvR0, PRTR3PTR ppvR3, PRTHCPHYS pHCPhys)
2004	{
2005	int rc;
2006	SUPDRVMEMREF Mem = {0};
2007	LogFlow(("SUPR0ContAlloc: pSession=%p cPages=%d ppvR0=%p ppvR3=%p pHCPhys=%p\n", pSession, cPages, ppvR0, ppvR3, pHCPhys));
2008
2009	/*
2010	* Validate input.
2011	*/
2012	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2013	if (!ppvR3 \|\| !ppvR0 \|\| !pHCPhys)
2014	{
2015	Log(("Null pointer. All of these should be set: pSession=%p ppvR0=%p ppvR3=%p pHCPhys=%p\n",
2016	pSession, ppvR0, ppvR3, pHCPhys));
2017	return VERR_INVALID_PARAMETER;
2018
2019	}
2020	if (cPages < 1 \|\| cPages >= 256)
2021	{
2022	Log(("Illegal request cPages=%d, must be greater than 0 and smaller than 256\n", cPages));
2023	return VERR_INVALID_PARAMETER;
2024	}
2025
2026	/*
2027	* Let IPRT do the job.
2028	*/
2029	rc = RTR0MemObjAllocCont(&Mem.MemObj, cPages << PAGE_SHIFT, true /* executable R0 mapping */);
2030	if (RT_SUCCESS(rc))
2031	{
2032	int rc2;
2033	rc = RTR0MemObjMapUser(&Mem.MapObjR3, Mem.MemObj, (RTR3PTR)-1, 0,
2034	RTMEM_PROT_EXEC \| RTMEM_PROT_WRITE \| RTMEM_PROT_READ, RTR0ProcHandleSelf());
2035	if (RT_SUCCESS(rc))
2036	{
2037	Mem.eType = MEMREF_TYPE_CONT;
2038	rc = supdrvMemAdd(&Mem, pSession);
2039	if (!rc)
2040	{
2041	*ppvR0 = RTR0MemObjAddress(Mem.MemObj);
2042	*ppvR3 = RTR0MemObjAddressR3(Mem.MapObjR3);
2043	*pHCPhys = RTR0MemObjGetPagePhysAddr(Mem.MemObj, 0);
2044	return 0;
2045	}
2046
2047	rc2 = RTR0MemObjFree(Mem.MapObjR3, false);
2048	AssertRC(rc2);
2049	}
2050	rc2 = RTR0MemObjFree(Mem.MemObj, false);
2051	AssertRC(rc2);
2052	}
2053
2054	return rc;
2055	}
2056
2057
2058	/**
2059	* Frees memory allocated using SUPR0ContAlloc().
2060	*
2061	* @returns IPRT status code.
2062	* @param pSession The session to which the memory was allocated.
2063	* @param uPtr Pointer to the memory (ring-3 or ring-0).
2064	*/
2065	SUPR0DECL(int) SUPR0ContFree(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr)
2066	{
2067	LogFlow(("SUPR0ContFree: pSession=%p uPtr=%p\n", pSession, (void *)uPtr));
2068	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2069	return supdrvMemRelease(pSession, uPtr, MEMREF_TYPE_CONT);
2070	}
2071
2072
2073	/**
2074	* Allocates a chunk of page aligned memory with fixed physical backing below 4GB.
2075	*
2076	* The memory isn't zeroed.
2077	*
2078	* @returns IPRT status code.
2079	* @param pSession Session data.
2080	* @param cPages Number of pages to allocate.
2081	* @param ppvR0 Where to put the address of Ring-0 mapping of the allocated memory.
2082	* @param ppvR3 Where to put the address of Ring-3 mapping of the allocated memory.
2083	* @param paPages Where to put the physical addresses of allocated memory.
2084	*/
2085	SUPR0DECL(int) SUPR0LowAlloc(PSUPDRVSESSION pSession, uint32_t cPages, PRTR0PTR ppvR0, PRTR3PTR ppvR3, PRTHCPHYS paPages)
2086	{
2087	unsigned iPage;
2088	int rc;
2089	SUPDRVMEMREF Mem = {0};
2090	LogFlow(("SUPR0LowAlloc: pSession=%p cPages=%d ppvR3=%p ppvR0=%p paPages=%p\n", pSession, cPages, ppvR3, ppvR0, paPages));
2091
2092	/*
2093	* Validate input.
2094	*/
2095	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2096	if (!ppvR3 \|\| !ppvR0 \|\| !paPages)
2097	{
2098	Log(("Null pointer. All of these should be set: pSession=%p ppvR3=%p ppvR0=%p paPages=%p\n",
2099	pSession, ppvR3, ppvR0, paPages));
2100	return VERR_INVALID_PARAMETER;
2101
2102	}
2103	if (cPages < 1 \|\| cPages > 256)
2104	{
2105	Log(("Illegal request cPages=%d, must be greater than 0 and smaller than 256.\n", cPages));
2106	return VERR_INVALID_PARAMETER;
2107	}
2108
2109	/*
2110	* Let IPRT do the work.
2111	*/
2112	rc = RTR0MemObjAllocLow(&Mem.MemObj, cPages << PAGE_SHIFT, true /* executable ring-0 mapping */);
2113	if (RT_SUCCESS(rc))
2114	{
2115	int rc2;
2116	rc = RTR0MemObjMapUser(&Mem.MapObjR3, Mem.MemObj, (RTR3PTR)-1, 0,
2117	RTMEM_PROT_EXEC \| RTMEM_PROT_WRITE \| RTMEM_PROT_READ, RTR0ProcHandleSelf());
2118	if (RT_SUCCESS(rc))
2119	{
2120	Mem.eType = MEMREF_TYPE_LOW;
2121	rc = supdrvMemAdd(&Mem, pSession);
2122	if (!rc)
2123	{
2124	for (iPage = 0; iPage < cPages; iPage++)
2125	{
2126	paPages[iPage] = RTR0MemObjGetPagePhysAddr(Mem.MemObj, iPage);
2127	AssertMsg(!(paPages[iPage] & (PAGE_SIZE - 1)), ("iPage=%d Phys=%VHp\n", paPages[iPage]));
2128	}
2129	*ppvR0 = RTR0MemObjAddress(Mem.MemObj);
2130	*ppvR3 = RTR0MemObjAddressR3(Mem.MapObjR3);
2131	return 0;
2132	}
2133
2134	rc2 = RTR0MemObjFree(Mem.MapObjR3, false);
2135	AssertRC(rc2);
2136	}
2137
2138	rc2 = RTR0MemObjFree(Mem.MemObj, false);
2139	AssertRC(rc2);
2140	}
2141
2142	return rc;
2143	}
2144
2145
2146	/**
2147	* Frees memory allocated using SUPR0LowAlloc().
2148	*
2149	* @returns IPRT status code.
2150	* @param pSession The session to which the memory was allocated.
2151	* @param uPtr Pointer to the memory (ring-3 or ring-0).
2152	*/
2153	SUPR0DECL(int) SUPR0LowFree(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr)
2154	{
2155	LogFlow(("SUPR0LowFree: pSession=%p uPtr=%p\n", pSession, (void *)uPtr));
2156	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2157	return supdrvMemRelease(pSession, uPtr, MEMREF_TYPE_LOW);
2158	}
2159
2160
2161
2162	/**
2163	* Allocates a chunk of memory with both R0 and R3 mappings.
2164	* The memory is fixed and it's possible to query the physical addresses using SUPR0MemGetPhys().
2165	*
2166	* @returns IPRT status code.
2167	* @param pSession The session to associated the allocation with.
2168	* @param cb Number of bytes to allocate.
2169	* @param ppvR0 Where to store the address of the Ring-0 mapping.
2170	* @param ppvR3 Where to store the address of the Ring-3 mapping.
2171	*/
2172	SUPR0DECL(int) SUPR0MemAlloc(PSUPDRVSESSION pSession, uint32_t cb, PRTR0PTR ppvR0, PRTR3PTR ppvR3)
2173	{
2174	int rc;
2175	SUPDRVMEMREF Mem = {0};
2176	LogFlow(("SUPR0MemAlloc: pSession=%p cb=%d ppvR0=%p ppvR3=%p\n", pSession, cb, ppvR0, ppvR3));
2177
2178	/*
2179	* Validate input.
2180	*/
2181	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2182	AssertPtrReturn(ppvR0, VERR_INVALID_POINTER);
2183	AssertPtrReturn(ppvR3, VERR_INVALID_POINTER);
2184	if (cb < 1 \|\| cb >= _4M)
2185	{
2186	Log(("Illegal request cb=%u; must be greater than 0 and smaller than 4MB.\n", cb));
2187	return VERR_INVALID_PARAMETER;
2188	}
2189
2190	/*
2191	* Let IPRT do the work.
2192	*/
2193	rc = RTR0MemObjAllocPage(&Mem.MemObj, cb, true /* executable ring-0 mapping */);
2194	if (RT_SUCCESS(rc))
2195	{
2196	int rc2;
2197	rc = RTR0MemObjMapUser(&Mem.MapObjR3, Mem.MemObj, (RTR3PTR)-1, 0,
2198	RTMEM_PROT_EXEC \| RTMEM_PROT_WRITE \| RTMEM_PROT_READ, RTR0ProcHandleSelf());
2199	if (RT_SUCCESS(rc))
2200	{
2201	Mem.eType = MEMREF_TYPE_MEM;
2202	rc = supdrvMemAdd(&Mem, pSession);
2203	if (!rc)
2204	{
2205	*ppvR0 = RTR0MemObjAddress(Mem.MemObj);
2206	*ppvR3 = RTR0MemObjAddressR3(Mem.MapObjR3);
2207	return VINF_SUCCESS;
2208	}
2209	rc2 = RTR0MemObjFree(Mem.MapObjR3, false);
2210	AssertRC(rc2);
2211	}
2212
2213	rc2 = RTR0MemObjFree(Mem.MemObj, false);
2214	AssertRC(rc2);
2215	}
2216
2217	return rc;
2218	}
2219
2220
2221	/**
2222	* Get the physical addresses of memory allocated using SUPR0MemAlloc().
2223	*
2224	* @returns IPRT status code.
2225	* @param pSession The session to which the memory was allocated.
2226	* @param uPtr The Ring-0 or Ring-3 address returned by SUPR0MemAlloc().
2227	* @param paPages Where to store the physical addresses.
2228	*/
2229	SUPR0DECL(int) SUPR0MemGetPhys(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr, PSUPPAGE paPages) /** @todo switch this bugger to RTHCPHYS */
2230	{
2231	PSUPDRVBUNDLE pBundle;
2232	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
2233	LogFlow(("SUPR0MemGetPhys: pSession=%p uPtr=%p paPages=%p\n", pSession, (void *)uPtr, paPages));
2234
2235	/*
2236	* Validate input.
2237	*/
2238	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2239	AssertPtrReturn(paPages, VERR_INVALID_POINTER);
2240	AssertReturn(uPtr, VERR_INVALID_PARAMETER);
2241
2242	/*
2243	* Search for the address.
2244	*/
2245	RTSpinlockAcquire(pSession->Spinlock, &SpinlockTmp);
2246	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
2247	{
2248	if (pBundle->cUsed > 0)
2249	{
2250	unsigned i;
2251	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
2252	{
2253	if ( pBundle->aMem[i].eType == MEMREF_TYPE_MEM
2254	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
2255	&& ( (RTHCUINTPTR)RTR0MemObjAddress(pBundle->aMem[i].MemObj) == uPtr
2256	\|\| ( pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ
2257	&& RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3) == uPtr)
2258	)
2259	)
2260	{
2261	const size_t cPages = RTR0MemObjSize(pBundle->aMem[i].MemObj) >> PAGE_SHIFT;
2262	size_t iPage;
2263	for (iPage = 0; iPage < cPages; iPage++)
2264	{
2265	paPages[iPage].Phys = RTR0MemObjGetPagePhysAddr(pBundle->aMem[i].MemObj, iPage);
2266	paPages[iPage].uReserved = 0;
2267	}
2268	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
2269	return VINF_SUCCESS;
2270	}
2271	}
2272	}
2273	}
2274	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
2275	Log(("Failed to find %p!!!\n", (void *)uPtr));
2276	return VERR_INVALID_PARAMETER;
2277	}
2278
2279
2280	/**
2281	* Free memory allocated by SUPR0MemAlloc().
2282	*
2283	* @returns IPRT status code.
2284	* @param pSession The session owning the allocation.
2285	* @param uPtr The Ring-0 or Ring-3 address returned by SUPR0MemAlloc().
2286	*/
2287	SUPR0DECL(int) SUPR0MemFree(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr)
2288	{
2289	LogFlow(("SUPR0MemFree: pSession=%p uPtr=%p\n", pSession, (void *)uPtr));
2290	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2291	return supdrvMemRelease(pSession, uPtr, MEMREF_TYPE_MEM);
2292	}
2293
2294
2295	/**
2296	* Allocates a chunk of memory with only a R3 mappings.
2297	* The memory is fixed and it's possible to query the physical addresses using SUPR0MemGetPhys().
2298	*
2299	* @returns IPRT status code.
2300	* @param pSession The session to associated the allocation with.
2301	* @param cPages The number of pages to allocate.
2302	* @param ppvR3 Where to store the address of the Ring-3 mapping.
2303	* @param paPages Where to store the addresses of the pages. Optional.
2304	*/
2305	SUPR0DECL(int) SUPR0PageAlloc(PSUPDRVSESSION pSession, uint32_t cPages, PRTR3PTR ppvR3, PRTHCPHYS paPages)
2306	{
2307	int rc;
2308	SUPDRVMEMREF Mem = {0};
2309	LogFlow(("SUPR0PageAlloc: pSession=%p cb=%d ppvR3=%p\n", pSession, cPages, ppvR3));
2310
2311	/*
2312	* Validate input. The allowed allocation size must be at least equal to the maximum guest VRAM size.
2313	*/
2314	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2315	AssertPtrReturn(ppvR3, VERR_INVALID_POINTER);
2316	if (cPages < 1 \|\| cPages > (128 * _1M)/PAGE_SIZE)
2317	{
2318	Log(("SUPR0PageAlloc: Illegal request cb=%u; must be greater than 0 and smaller than 128MB.\n", cPages));
2319	return VERR_INVALID_PARAMETER;
2320	}
2321
2322	/*
2323	* Let IPRT do the work.
2324	*/
2325	rc = RTR0MemObjAllocPhysNC(&Mem.MemObj, (size_t)cPages * PAGE_SIZE, NIL_RTHCPHYS);
2326	if (RT_SUCCESS(rc))
2327	{
2328	int rc2;
2329	rc = RTR0MemObjMapUser(&Mem.MapObjR3, Mem.MemObj, (RTR3PTR)-1, 0,
2330	RTMEM_PROT_EXEC \| RTMEM_PROT_WRITE \| RTMEM_PROT_READ, RTR0ProcHandleSelf());
2331	if (RT_SUCCESS(rc))
2332	{
2333	Mem.eType = MEMREF_TYPE_LOCKED_SUP;
2334	rc = supdrvMemAdd(&Mem, pSession);
2335	if (!rc)
2336	{
2337	*ppvR3 = RTR0MemObjAddressR3(Mem.MapObjR3);
2338	if (paPages)
2339	{
2340	uint32_t iPage = cPages;
2341	while (iPage-- > 0)
2342	{
2343	paPages[iPage] = RTR0MemObjGetPagePhysAddr(Mem.MapObjR3, iPage);
2344	Assert(paPages[iPage] != NIL_RTHCPHYS);
2345	}
2346	}
2347	return VINF_SUCCESS;
2348	}
2349	rc2 = RTR0MemObjFree(Mem.MapObjR3, false);
2350	AssertRC(rc2);
2351	}
2352
2353	rc2 = RTR0MemObjFree(Mem.MemObj, false);
2354	AssertRC(rc2);
2355	}
2356	return rc;
2357	}
2358
2359
2360	#ifdef RT_OS_WINDOWS
2361	/**
2362	* Check if the pages were locked by SUPR0PageAlloc
2363	*
2364	* This function will be removed along with the lock/unlock hacks when
2365	* we've cleaned up the ring-3 code properly.
2366	*
2367	* @returns boolean
2368	* @param pSession The session to which the memory was allocated.
2369	* @param pvR3 The Ring-3 address returned by SUPR0PageAlloc().
2370	*/
2371	static bool supdrvPageWasLockedByPageAlloc(PSUPDRVSESSION pSession, RTR3PTR pvR3)
2372	{
2373	PSUPDRVBUNDLE pBundle;
2374	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
2375	LogFlow(("SUPR0PageIsLockedByPageAlloc: pSession=%p pvR3=%p\n", pSession, (void *)pvR3));
2376
2377	/*
2378	* Search for the address.
2379	*/
2380	RTSpinlockAcquire(pSession->Spinlock, &SpinlockTmp);
2381	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
2382	{
2383	if (pBundle->cUsed > 0)
2384	{
2385	unsigned i;
2386	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
2387	{
2388	if ( pBundle->aMem[i].eType == MEMREF_TYPE_LOCKED_SUP
2389	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
2390	&& pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ
2391	&& RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3) == pvR3)
2392	{
2393	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
2394	return true;
2395	}
2396	}
2397	}
2398	}
2399	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
2400	return false;
2401	}
2402
2403
2404	/**
2405	* Get the physical addresses of memory allocated using SUPR0PageAlloc().
2406	*
2407	* This function will be removed along with the lock/unlock hacks when
2408	* we've cleaned up the ring-3 code properly.
2409	*
2410	* @returns IPRT status code.
2411	* @param pSession The session to which the memory was allocated.
2412	* @param pvR3 The Ring-3 address returned by SUPR0PageAlloc().
2413	* @param cPages Number of pages in paPages
2414	* @param paPages Where to store the physical addresses.
2415	*/
2416	static int supdrvPageGetPhys(PSUPDRVSESSION pSession, RTR3PTR pvR3, uint32_t cPages, PRTHCPHYS paPages)
2417	{
2418	PSUPDRVBUNDLE pBundle;
2419	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
2420	LogFlow(("supdrvPageGetPhys: pSession=%p pvR3=%p cPages=%#lx paPages=%p\n", pSession, (void *)pvR3, (long)cPages, paPages));
2421
2422	/*
2423	* Search for the address.
2424	*/
2425	RTSpinlockAcquire(pSession->Spinlock, &SpinlockTmp);
2426	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
2427	{
2428	if (pBundle->cUsed > 0)
2429	{
2430	unsigned i;
2431	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
2432	{
2433	if ( pBundle->aMem[i].eType == MEMREF_TYPE_LOCKED_SUP
2434	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
2435	&& pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ
2436	&& RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3) == pvR3)
2437	{
2438	uint32_t iPage;
2439	size_t cMaxPages = RTR0MemObjSize(pBundle->aMem[i].MemObj) >> PAGE_SHIFT;
2440	cPages = (uint32_t)RT_MIN(cMaxPages, cPages);
2441	for (iPage = 0; iPage < cPages; iPage++)
2442	paPages[iPage] = RTR0MemObjGetPagePhysAddr(pBundle->aMem[i].MemObj, iPage);
2443	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
2444	return VINF_SUCCESS;
2445	}
2446	}
2447	}
2448	}
2449	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
2450	return VERR_INVALID_PARAMETER;
2451	}
2452	#endif /* RT_OS_WINDOWS */
2453
2454
2455	/**
2456	* Free memory allocated by SUPR0PageAlloc().
2457	*
2458	* @returns IPRT status code.
2459	* @param pSession The session owning the allocation.
2460	* @param pvR3 The Ring-3 address returned by SUPR0PageAlloc().
2461	*/
2462	SUPR0DECL(int) SUPR0PageFree(PSUPDRVSESSION pSession, RTR3PTR pvR3)
2463	{
2464	LogFlow(("SUPR0PageFree: pSession=%p pvR3=%p\n", pSession, (void *)pvR3));
2465	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2466	return supdrvMemRelease(pSession, (RTHCUINTPTR)pvR3, MEMREF_TYPE_LOCKED_SUP);
2467	}
2468
2469
2470	/**
2471	* Maps the GIP into userspace and/or get the physical address of the GIP.
2472	*
2473	* @returns IPRT status code.
2474	* @param pSession Session to which the GIP mapping should belong.
2475	* @param ppGipR3 Where to store the address of the ring-3 mapping. (optional)
2476	* @param pHCPhysGip Where to store the physical address. (optional)
2477	*
2478	* @remark There is no reference counting on the mapping, so one call to this function
2479	* count globally as one reference. One call to SUPR0GipUnmap() is will unmap GIP
2480	* and remove the session as a GIP user.
2481	*/
2482	SUPR0DECL(int) SUPR0GipMap(PSUPDRVSESSION pSession, PRTR3PTR ppGipR3, PRTHCPHYS pHCPhysGip)
2483	{
2484	int rc = 0;
2485	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
2486	RTR3PTR pGip = NIL_RTR3PTR;
2487	RTHCPHYS HCPhys = NIL_RTHCPHYS;
2488	LogFlow(("SUPR0GipMap: pSession=%p ppGipR3=%p pHCPhysGip=%p\n", pSession, ppGipR3, pHCPhysGip));
2489
2490	/*
2491	* Validate
2492	*/
2493	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2494	AssertPtrNullReturn(ppGipR3, VERR_INVALID_POINTER);
2495	AssertPtrNullReturn(pHCPhysGip, VERR_INVALID_POINTER);
2496
2497	RTSemFastMutexRequest(pDevExt->mtxGip);
2498	if (pDevExt->pGip)
2499	{
2500	/*
2501	* Map it?
2502	*/
2503	if (ppGipR3)
2504	{
2505	if (pSession->GipMapObjR3 == NIL_RTR0MEMOBJ)
2506	rc = RTR0MemObjMapUser(&pSession->GipMapObjR3, pDevExt->GipMemObj, (RTR3PTR)-1, 0,
2507	RTMEM_PROT_READ, RTR0ProcHandleSelf());
2508	if (RT_SUCCESS(rc))
2509	{
2510	pGip = RTR0MemObjAddressR3(pSession->GipMapObjR3);
2511	rc = VINF_SUCCESS; /** @todo remove this and replace the !rc below with RT_SUCCESS(rc). */
2512	}
2513	}
2514
2515	/*
2516	* Get physical address.
2517	*/
2518	if (pHCPhysGip && !rc)
2519	HCPhys = pDevExt->HCPhysGip;
2520
2521	/*
2522	* Reference globally.
2523	*/
2524	if (!pSession->fGipReferenced && !rc)
2525	{
2526	pSession->fGipReferenced = 1;
2527	pDevExt->cGipUsers++;
2528	if (pDevExt->cGipUsers == 1)
2529	{
2530	PSUPGLOBALINFOPAGE pGip = pDevExt->pGip;
2531	unsigned i;
2532
2533	LogFlow(("SUPR0GipMap: Resumes GIP updating\n"));
2534
2535	for (i = 0; i < RT_ELEMENTS(pGip->aCPUs); i++)
2536	ASMAtomicXchgU32(&pGip->aCPUs[i].u32TransactionId, pGip->aCPUs[i].u32TransactionId & ~(GIP_UPDATEHZ_RECALC_FREQ * 2 - 1));
2537	ASMAtomicXchgU64(&pGip->u64NanoTSLastUpdateHz, 0);
2538
2539	rc = RTTimerStart(pDevExt->pGipTimer, 0);
2540	AssertRC(rc); rc = VINF_SUCCESS;
2541	}
2542	}
2543	}
2544	else
2545	{
2546	rc = SUPDRV_ERR_GENERAL_FAILURE;
2547	Log(("SUPR0GipMap: GIP is not available!\n"));
2548	}
2549	RTSemFastMutexRelease(pDevExt->mtxGip);
2550
2551	/*
2552	* Write returns.
2553	*/
2554	if (pHCPhysGip)
2555	*pHCPhysGip = HCPhys;
2556	if (ppGipR3)
2557	*ppGipR3 = pGip;
2558
2559	#ifdef DEBUG_DARWIN_GIP
2560	OSDBGPRINT(("SUPR0GipMap: returns %d pHCPhysGip=%lx ppGip=%p GipMapObjR3\n", rc, (unsigned long)HCPhys, pGip, pSession->GipMapObjR3));
2561	#else
2562	LogFlow(("SUPR0GipMap: returns %d pHCPhysGip=%lx ppGipR3=%p\n", rc, (unsigned long)HCPhys, (void *)(uintptr_t)pGip));
2563	#endif
2564	return rc;
2565	}
2566
2567
2568	/**
2569	* Unmaps any user mapping of the GIP and terminates all GIP access
2570	* from this session.
2571	*
2572	* @returns IPRT status code.
2573	* @param pSession Session to which the GIP mapping should belong.
2574	*/
2575	SUPR0DECL(int) SUPR0GipUnmap(PSUPDRVSESSION pSession)
2576	{
2577	int rc = VINF_SUCCESS;
2578	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
2579	#ifdef DEBUG_DARWIN_GIP
2580	OSDBGPRINT(("SUPR0GipUnmap: pSession=%p pGip=%p GipMapObjR3=%p\n",
2581	pSession,
2582	pSession->GipMapObjR3 != NIL_RTR0MEMOBJ ? RTR0MemObjAddress(pSession->GipMapObjR3) : NULL,
2583	pSession->GipMapObjR3));
2584	#else
2585	LogFlow(("SUPR0GipUnmap: pSession=%p\n", pSession));
2586	#endif
2587	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2588
2589	RTSemFastMutexRequest(pDevExt->mtxGip);
2590
2591	/*
2592	* Unmap anything?
2593	*/
2594	if (pSession->GipMapObjR3 != NIL_RTR0MEMOBJ)
2595	{
2596	rc = RTR0MemObjFree(pSession->GipMapObjR3, false);
2597	AssertRC(rc);
2598	if (RT_SUCCESS(rc))
2599	pSession->GipMapObjR3 = NIL_RTR0MEMOBJ;
2600	}
2601
2602	/*
2603	* Dereference global GIP.
2604	*/
2605	if (pSession->fGipReferenced && !rc)
2606	{
2607	pSession->fGipReferenced = 0;
2608	if ( pDevExt->cGipUsers > 0
2609	&& !--pDevExt->cGipUsers)
2610	{
2611	LogFlow(("SUPR0GipUnmap: Suspends GIP updating\n"));
2612	rc = RTTimerStop(pDevExt->pGipTimer); AssertRC(rc); rc = 0;
2613	}
2614	}
2615
2616	RTSemFastMutexRelease(pDevExt->mtxGip);
2617
2618	return rc;
2619	}
2620
2621
2622	/**
2623	* Register a component factory with the support driver.
2624	*
2625	* This is currently restricted to kernel sessions only.
2626	*
2627	* @returns VBox status code.
2628	* @retval VINF_SUCCESS on success.
2629	* @retval VERR_NO_MEMORY if we're out of memory.
2630	* @retval VERR_ALREADY_EXISTS if the factory has already been registered.
2631	* @retval VERR_ACCESS_DENIED if it isn't a kernel session.
2632	* @retval VERR_INVALID_PARAMETER on invalid parameter.
2633	* @retval VERR_INVALID_POINTER on invalid pointer parameter.
2634	*
2635	* @param pSession The SUPDRV session (must be a ring-0 session).
2636	* @param pFactory Pointer to the component factory registration structure.
2637	*
2638	* @remarks This interface is also available via SUPR0IdcComponentRegisterFactory.
2639	*/
2640	SUPR0DECL(int) SUPR0ComponentRegisterFactory(PSUPDRVSESSION pSession, PCSUPDRVFACTORY pFactory)
2641	{
2642	PSUPDRVFACTORYREG pNewReg;
2643	const char *psz;
2644	int rc;
2645
2646	/*
2647	* Validate parameters.
2648	*/
2649	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2650	AssertReturn(pSession->R0Process == NIL_RTR0PROCESS, VERR_ACCESS_DENIED);
2651	AssertPtrReturn(pFactory, VERR_INVALID_POINTER);
2652	AssertPtrReturn(pFactory->pfnQueryFactoryInterface, VERR_INVALID_POINTER);
2653	psz = (const char *)memchr(pFactory->szName, '\0', sizeof(pFactory->szName));
2654	AssertReturn(psz, VERR_INVALID_PARAMETER);
2655
2656	/*
2657	* Allocate and initialize a new registration structure.
2658	*/
2659	pNewReg = (PSUPDRVFACTORYREG)RTMemAlloc(sizeof(SUPDRVFACTORYREG));
2660	if (pNewReg)
2661	{
2662	pNewReg->pNext = NULL;
2663	pNewReg->pFactory = pFactory;
2664	pNewReg->pSession = pSession;
2665	pNewReg->cchName = psz - &pFactory->szName[0];
2666
2667	/*
2668	* Add it to the tail of the list after checking for prior registration.
2669	*/
2670	rc = RTSemFastMutexRequest(pSession->pDevExt->mtxComponentFactory);
2671	if (RT_SUCCESS(rc))
2672	{
2673	PSUPDRVFACTORYREG pPrev = NULL;
2674	PSUPDRVFACTORYREG pCur = pSession->pDevExt->pComponentFactoryHead;
2675	while (pCur && pCur->pFactory != pFactory)
2676	{
2677	pPrev = pCur;
2678	pCur = pCur->pNext;
2679	}
2680	if (!pCur)
2681	{
2682	if (pPrev)
2683	pPrev->pNext = pNewReg;
2684	else
2685	pSession->pDevExt->pComponentFactoryHead = pNewReg;
2686	rc = VINF_SUCCESS;
2687	}
2688	else
2689	rc = VERR_ALREADY_EXISTS;
2690
2691	RTSemFastMutexRelease(pSession->pDevExt->mtxComponentFactory);
2692	}
2693
2694	if (RT_FAILURE(rc))
2695	RTMemFree(pNewReg);
2696	}
2697	else
2698	rc = VERR_NO_MEMORY;
2699	return rc;
2700	}
2701
2702
2703	/**
2704	* Deregister a component factory.
2705	*
2706	* @returns VBox status code.
2707	* @retval VINF_SUCCESS on success.
2708	* @retval VERR_NOT_FOUND if the factory wasn't registered.
2709	* @retval VERR_ACCESS_DENIED if it isn't a kernel session.
2710	* @retval VERR_INVALID_PARAMETER on invalid parameter.
2711	* @retval VERR_INVALID_POINTER on invalid pointer parameter.
2712	*
2713	* @param pSession The SUPDRV session (must be a ring-0 session).
2714	* @param pFactory Pointer to the component factory registration structure
2715	* previously passed SUPR0ComponentRegisterFactory().
2716	*
2717	* @remarks This interface is also available via SUPR0IdcComponentDeregisterFactory.
2718	*/
2719	SUPR0DECL(int) SUPR0ComponentDeregisterFactory(PSUPDRVSESSION pSession, PCSUPDRVFACTORY pFactory)
2720	{
2721	int rc;
2722
2723	/*
2724	* Validate parameters.
2725	*/
2726	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2727	AssertReturn(pSession->R0Process == NIL_RTR0PROCESS, VERR_ACCESS_DENIED);
2728	AssertPtrReturn(pFactory, VERR_INVALID_POINTER);
2729
2730	/*
2731	* Take the lock and look for the registration record.
2732	*/
2733	rc = RTSemFastMutexRequest(pSession->pDevExt->mtxComponentFactory);
2734	if (RT_SUCCESS(rc))
2735	{
2736	PSUPDRVFACTORYREG pPrev = NULL;
2737	PSUPDRVFACTORYREG pCur = pSession->pDevExt->pComponentFactoryHead;
2738	while (pCur && pCur->pFactory != pFactory)
2739	{
2740	pPrev = pCur;
2741	pCur = pCur->pNext;
2742	}
2743	if (pCur)
2744	{
2745	if (!pPrev)
2746	pSession->pDevExt->pComponentFactoryHead = pCur->pNext;
2747	else
2748	pPrev->pNext = pCur->pNext;
2749
2750	pCur->pNext = NULL;
2751	pCur->pFactory = NULL;
2752	pCur->pSession = NULL;
2753	rc = VINF_SUCCESS;
2754	}
2755	else
2756	rc = VERR_NOT_FOUND;
2757
2758	RTSemFastMutexRelease(pSession->pDevExt->mtxComponentFactory);
2759
2760	RTMemFree(pCur);
2761	}
2762	return rc;
2763	}
2764
2765
2766	/**
2767	* Queries a component factory.
2768	*
2769	* @returns VBox status code.
2770	* @retval VERR_INVALID_PARAMETER on invalid parameter.
2771	* @retval VERR_INVALID_POINTER on invalid pointer parameter.
2772	* @retval VERR_SUPDRV_COMPONENT_NOT_FOUND if the component factory wasn't found.
2773	* @retval VERR_SUPDRV_INTERFACE_NOT_SUPPORTED if the interface wasn't supported.
2774	*
2775	* @param pSession The SUPDRV session.
2776	* @param pszName The name of the component factory.
2777	* @param pszInterfaceUuid The UUID of the factory interface (stringified).
2778	* @param ppvFactoryIf Where to store the factory interface.
2779	*/
2780	SUPR0DECL(int) SUPR0ComponentQueryFactory(PSUPDRVSESSION pSession, const char pszName, const char pszInterfaceUuid, void **ppvFactoryIf)
2781	{
2782	const char *pszEnd;
2783	size_t cchName;
2784	int rc;
2785
2786	/*
2787	* Validate parameters.
2788	*/
2789	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2790
2791	AssertPtrReturn(pszName, VERR_INVALID_POINTER);
2792	pszEnd = memchr(pszName, '\0', RT_SIZEOFMEMB(SUPDRVFACTORY, szName));
2793	AssertReturn(pszEnd, VERR_INVALID_PARAMETER);
2794	cchName = pszEnd - pszName;
2795
2796	AssertPtrReturn(pszInterfaceUuid, VERR_INVALID_POINTER);
2797	pszEnd = memchr(pszInterfaceUuid, '\0', RTUUID_STR_LENGTH);
2798	AssertReturn(pszEnd, VERR_INVALID_PARAMETER);
2799
2800	AssertPtrReturn(ppvFactoryIf, VERR_INVALID_POINTER);
2801	*ppvFactoryIf = NULL;
2802
2803	/*
2804	* Take the lock and try all factories by this name.
2805	*/
2806	rc = RTSemFastMutexRequest(pSession->pDevExt->mtxComponentFactory);
2807	if (RT_SUCCESS(rc))
2808	{
2809	PSUPDRVFACTORYREG pCur = pSession->pDevExt->pComponentFactoryHead;
2810	rc = VERR_SUPDRV_COMPONENT_NOT_FOUND;
2811	while (pCur)
2812	{
2813	if ( pCur->cchName == cchName
2814	&& !memcmp(pCur->pFactory->szName, pszName, cchName))
2815	{
2816	#ifdef SUPDRV_WITH_UNWIND_HACK
2817	void *pvFactory = supdrvNtWrapQueryFactoryInterface((PFNRT)pCur->pFactory->pfnQueryFactoryInterface, pCur->pFactory, pSession, pszInterfaceUuid);
2818	#else
2819	void *pvFactory = pCur->pFactory->pfnQueryFactoryInterface(pCur->pFactory, pSession, pszInterfaceUuid);
2820	#endif
2821	if (pvFactory)
2822	{
2823	*ppvFactoryIf = pvFactory;
2824	rc = VINF_SUCCESS;
2825	break;
2826	}
2827	rc = VERR_SUPDRV_INTERFACE_NOT_SUPPORTED;
2828	}
2829
2830	/* next */
2831	pCur = pCur->pNext;
2832	}
2833
2834	RTSemFastMutexRelease(pSession->pDevExt->mtxComponentFactory);
2835	}
2836	return rc;
2837	}
2838
2839
2840	/**
2841	* Adds a memory object to the session.
2842	*
2843	* @returns IPRT status code.
2844	* @param pMem Memory tracking structure containing the
2845	* information to track.
2846	* @param pSession The session.
2847	*/
2848	static int supdrvMemAdd(PSUPDRVMEMREF pMem, PSUPDRVSESSION pSession)
2849	{
2850	PSUPDRVBUNDLE pBundle;
2851	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
2852
2853	/*
2854	* Find free entry and record the allocation.
2855	*/
2856	RTSpinlockAcquire(pSession->Spinlock, &SpinlockTmp);
2857	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
2858	{
2859	if (pBundle->cUsed < RT_ELEMENTS(pBundle->aMem))
2860	{
2861	unsigned i;
2862	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
2863	{
2864	if (pBundle->aMem[i].MemObj == NIL_RTR0MEMOBJ)
2865	{
2866	pBundle->cUsed++;
2867	pBundle->aMem[i] = *pMem;
2868	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
2869	return VINF_SUCCESS;
2870	}
2871	}
2872	AssertFailed(); /* !!this can't be happening!!! */
2873	}
2874	}
2875	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
2876
2877	/*
2878	* Need to allocate a new bundle.
2879	* Insert into the last entry in the bundle.
2880	*/
2881	pBundle = (PSUPDRVBUNDLE)RTMemAllocZ(sizeof(*pBundle));
2882	if (!pBundle)
2883	return VERR_NO_MEMORY;
2884
2885	/* take last entry. */
2886	pBundle->cUsed++;
2887	pBundle->aMem[RT_ELEMENTS(pBundle->aMem) - 1] = *pMem;
2888
2889	/* insert into list. */
2890	RTSpinlockAcquire(pSession->Spinlock, &SpinlockTmp);
2891	pBundle->pNext = pSession->Bundle.pNext;
2892	pSession->Bundle.pNext = pBundle;
2893	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
2894
2895	return VINF_SUCCESS;
2896	}
2897
2898
2899	/**
2900	* Releases a memory object referenced by pointer and type.
2901	*
2902	* @returns IPRT status code.
2903	* @param pSession Session data.
2904	* @param uPtr Pointer to memory. This is matched against both the R0 and R3 addresses.
2905	* @param eType Memory type.
2906	*/
2907	static int supdrvMemRelease(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr, SUPDRVMEMREFTYPE eType)
2908	{
2909	PSUPDRVBUNDLE pBundle;
2910	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
2911
2912	/*
2913	* Validate input.
2914	*/
2915	if (!uPtr)
2916	{
2917	Log(("Illegal address %p\n", (void *)uPtr));
2918	return VERR_INVALID_PARAMETER;
2919	}
2920
2921	/*
2922	* Search for the address.
2923	*/
2924	RTSpinlockAcquire(pSession->Spinlock, &SpinlockTmp);
2925	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
2926	{
2927	if (pBundle->cUsed > 0)
2928	{
2929	unsigned i;
2930	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
2931	{
2932	if ( pBundle->aMem[i].eType == eType
2933	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
2934	&& ( (RTHCUINTPTR)RTR0MemObjAddress(pBundle->aMem[i].MemObj) == uPtr
2935	\|\| ( pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ
2936	&& RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3) == uPtr))
2937	)
2938	{
2939	/* Make a copy of it and release it outside the spinlock. */
2940	SUPDRVMEMREF Mem = pBundle->aMem[i];
2941	pBundle->aMem[i].eType = MEMREF_TYPE_UNUSED;
2942	pBundle->aMem[i].MemObj = NIL_RTR0MEMOBJ;
2943	pBundle->aMem[i].MapObjR3 = NIL_RTR0MEMOBJ;
2944	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
2945
2946	if (Mem.MapObjR3)
2947	{
2948	int rc = RTR0MemObjFree(Mem.MapObjR3, false);
2949	AssertRC(rc); /** @todo figure out how to handle this. */
2950	}
2951	if (Mem.MemObj)
2952	{
2953	int rc = RTR0MemObjFree(Mem.MemObj, false);
2954	AssertRC(rc); /** @todo figure out how to handle this. */
2955	}
2956	return VINF_SUCCESS;
2957	}
2958	}
2959	}
2960	}
2961	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
2962	Log(("Failed to find %p!!! (eType=%d)\n", (void *)uPtr, eType));
2963	return VERR_INVALID_PARAMETER;
2964	}
2965
2966
2967	#ifdef VBOX_WITH_IDT_PATCHING
2968	/**
2969	* Install IDT for the current CPU.
2970	*
2971	* @returns One of the following IPRT status codes:
2972	* @retval VINF_SUCCESS on success.
2973	* @retval VERR_IDT_FAILED.
2974	* @retval VERR_NO_MEMORY.
2975	* @param pDevExt The device extension.
2976	* @param pSession The session data.
2977	* @param pReq The request.
2978	*/
2979	static int supdrvIOCtl_IdtInstall(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPIDTINSTALL pReq)
2980	{
2981	PSUPDRVPATCHUSAGE pUsagePre;
2982	PSUPDRVPATCH pPatchPre;
2983	RTIDTR Idtr;
2984	PSUPDRVPATCH pPatch;
2985	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
2986	LogFlow(("supdrvIOCtl_IdtInstall\n"));
2987
2988	/*
2989	* Preallocate entry for this CPU cause we don't wanna do
2990	* that inside the spinlock!
2991	*/
2992	pUsagePre = (PSUPDRVPATCHUSAGE)RTMemAlloc(sizeof(*pUsagePre));
2993	if (!pUsagePre)
2994	return VERR_NO_MEMORY;
2995
2996	/*
2997	* Take the spinlock and see what we need to do.
2998	*/
2999	RTSpinlockAcquireNoInts(pDevExt->Spinlock, &SpinlockTmp);
3000
3001	/* check if we already got a free patch. */
3002	if (!pDevExt->pIdtPatchesFree)
3003	{
3004	/*
3005	* Allocate a patch - outside the spinlock of course.
3006	*/
3007	RTSpinlockReleaseNoInts(pDevExt->Spinlock, &SpinlockTmp);
3008
3009	pPatchPre = (PSUPDRVPATCH)RTMemExecAlloc(sizeof(*pPatchPre));
3010	if (!pPatchPre)
3011	return VERR_NO_MEMORY;
3012
3013	RTSpinlockAcquireNoInts(pDevExt->Spinlock, &SpinlockTmp);
3014	}
3015	else
3016	{
3017	pPatchPre = pDevExt->pIdtPatchesFree;
3018	pDevExt->pIdtPatchesFree = pPatchPre->pNext;
3019	}
3020
3021	/* look for matching patch entry */
3022	ASMGetIDTR(&Idtr);
3023	pPatch = pDevExt->pIdtPatches;
3024	while (pPatch && pPatch->pvIdt != (void *)Idtr.pIdt)
3025	pPatch = pPatch->pNext;
3026
3027	if (!pPatch)
3028	{
3029	/*
3030	* Create patch.
3031	*/
3032	pPatch = supdrvIdtPatchOne(pDevExt, pPatchPre);
3033	if (pPatch)
3034	pPatchPre = NULL; /* mark as used. */
3035	}
3036	else
3037	{
3038	/*
3039	* Simply increment patch usage.
3040	*/
3041	pPatch->cUsage++;
3042	}
3043
3044	if (pPatch)
3045	{
3046	/*
3047	* Increment and add if need be the session usage record for this patch.
3048	*/
3049	PSUPDRVPATCHUSAGE pUsage = pSession->pPatchUsage;
3050	while (pUsage && pUsage->pPatch != pPatch)
3051	pUsage = pUsage->pNext;
3052
3053	if (!pUsage)
3054	{
3055	/*
3056	* Add usage record.
3057	*/
3058	pUsagePre->cUsage = 1;
3059	pUsagePre->pPatch = pPatch;
3060	pUsagePre->pNext = pSession->pPatchUsage;
3061	pSession->pPatchUsage = pUsagePre;
3062	pUsagePre = NULL; /* mark as used. */
3063	}
3064	else
3065	{
3066	/*
3067	* Increment usage count.
3068	*/
3069	pUsage->cUsage++;
3070	}
3071	}
3072
3073	/* free patch - we accumulate them for paranoid saftly reasons. */
3074	if (pPatchPre)
3075	{
3076	pPatchPre->pNext = pDevExt->pIdtPatchesFree;
3077	pDevExt->pIdtPatchesFree = pPatchPre;
3078	}
3079
3080	RTSpinlockReleaseNoInts(pDevExt->Spinlock, &SpinlockTmp);
3081
3082	/*
3083	* Free unused preallocated buffers.
3084	*/
3085	if (pUsagePre)
3086	RTMemFree(pUsagePre);
3087
3088	pReq->u.Out.u8Idt = pDevExt->u8Idt;
3089
3090	return pPatch ? VINF_SUCCESS : VERR_IDT_FAILED;
3091	}
3092
3093
3094	/**
3095	* This creates a IDT patch entry.
3096	* If the first patch being installed it'll also determin the IDT entry
3097	* to use.
3098	*
3099	* @returns pPatch on success.
3100	* @returns NULL on failure.
3101	* @param pDevExt Pointer to globals.
3102	* @param pPatch Patch entry to use.
3103	* This will be linked into SUPDRVDEVEXT::pIdtPatches on
3104	* successful return.
3105	* @remark Call must be owning the SUPDRVDEVEXT::Spinlock!
3106	*/
3107	static PSUPDRVPATCH supdrvIdtPatchOne(PSUPDRVDEVEXT pDevExt, PSUPDRVPATCH pPatch)
3108	{
3109	RTIDTR Idtr;
3110	PSUPDRVIDTE paIdt;
3111	LogFlow(("supdrvIOCtl_IdtPatchOne: pPatch=%p\n", pPatch));
3112
3113	/*
3114	* Get IDT.
3115	*/
3116	ASMGetIDTR(&Idtr);
3117	paIdt = (PSUPDRVIDTE)Idtr.pIdt;
3118	/*
3119	* Recent Linux kernels can be configured to 1G user /3G kernel.
3120	*/
3121	if ((uintptr_t)paIdt < 0x40000000)
3122	{
3123	AssertMsgFailed(("bad paIdt=%p\n", paIdt));
3124	return NULL;
3125	}
3126
3127	if (!pDevExt->u8Idt)
3128	{
3129	/*
3130	* Test out the alternatives.
3131	*
3132	* At the moment we do not support chaining thus we ASSUME that one of
3133	* these 48 entries is unused (which is not a problem on Win32 and
3134	* Linux to my knowledge).
3135	*/
3136	/** @todo we MUST change this detection to try grab an entry which is NOT in use. This can be
3137	* combined with gathering info about which guest system call gates we can hook up directly. */
3138	unsigned i;
3139	uint8_t u8Idt = 0;
3140	static uint8_t au8Ints[] =
3141	{
3142	#ifdef RT_OS_WINDOWS /* We don't use 0xef and above because they are system stuff on linux (ef is IPI,
3143	* local apic timer, or some other frequently fireing thing). */
3144	0xef, 0xee, 0xed, 0xec,
3145	#endif
3146	0xeb, 0xea, 0xe9, 0xe8,
3147	0xdf, 0xde, 0xdd, 0xdc,
3148	0x7b, 0x7a, 0x79, 0x78,
3149	0xbf, 0xbe, 0xbd, 0xbc,
3150	};
3151	#if defined(RT_ARCH_AMD64) && defined(DEBUG)
3152	static int s_iWobble = 0;
3153	unsigned iMax = !(s_iWobble++ % 2) ? 0x80 : 0x100;
3154	Log2(("IDT: Idtr=%p:%#x\n", (void *)Idtr.pIdt, (unsigned)Idtr.cbIdt));
3155	for (i = iMax - 0x80; i*16+15 < Idtr.cbIdt && i < iMax; i++)
3156	{
3157	Log2(("%#x: %04x:%08x%04x%04x P=%d DPL=%d IST=%d Type1=%#x u32Reserved=%#x u5Reserved=%#x\n",
3158	i, paIdt[i].u16SegSel, paIdt[i].u32OffsetTop, paIdt[i].u16OffsetHigh, paIdt[i].u16OffsetLow,
3159	paIdt[i].u1Present, paIdt[i].u2DPL, paIdt[i].u3IST, paIdt[i].u5Type2,
3160	paIdt[i].u32Reserved, paIdt[i].u5Reserved));
3161	}
3162	#endif
3163	/* look for entries which are not present or otherwise unused. */
3164	for (i = 0; i < sizeof(au8Ints) / sizeof(au8Ints[0]); i++)
3165	{
3166	u8Idt = au8Ints[i];
3167	if ( u8Idt * sizeof(SUPDRVIDTE) < Idtr.cbIdt
3168	&& ( !paIdt[u8Idt].u1Present
3169	\|\| paIdt[u8Idt].u5Type2 == 0))
3170	break;
3171	u8Idt = 0;
3172	}
3173	if (!u8Idt)
3174	{
3175	/* try again, look for a compatible entry .*/
3176	for (i = 0; i < sizeof(au8Ints) / sizeof(au8Ints[0]); i++)
3177	{
3178	u8Idt = au8Ints[i];
3179	if ( u8Idt * sizeof(SUPDRVIDTE) < Idtr.cbIdt
3180	&& paIdt[u8Idt].u1Present
3181	&& paIdt[u8Idt].u5Type2 == SUPDRV_IDTE_TYPE2_INTERRUPT_GATE
3182	&& !(paIdt[u8Idt].u16SegSel & 3))
3183	break;
3184	u8Idt = 0;
3185	}
3186	if (!u8Idt)
3187	{
3188	Log(("Failed to find appropirate IDT entry!!\n"));
3189	return NULL;
3190	}
3191	}
3192	pDevExt->u8Idt = u8Idt;
3193	LogFlow(("supdrvIOCtl_IdtPatchOne: u8Idt=%x\n", u8Idt));
3194	}
3195
3196	/*
3197	* Prepare the patch
3198	*/
3199	memset(pPatch, 0, sizeof(*pPatch));
3200	pPatch->pvIdt = paIdt;
3201	pPatch->cUsage = 1;
3202	pPatch->pIdtEntry = &paIdt[pDevExt->u8Idt];
3203	pPatch->SavedIdt = paIdt[pDevExt->u8Idt];
3204	pPatch->ChangedIdt.u16OffsetLow = (uint32_t)((uintptr_t)&pPatch->auCode[0] & 0xffff);
3205	pPatch->ChangedIdt.u16OffsetHigh = (uint32_t)((uintptr_t)&pPatch->auCode[0] >> 16);
3206	#ifdef RT_ARCH_AMD64
3207	pPatch->ChangedIdt.u32OffsetTop = (uint32_t)((uintptr_t)&pPatch->auCode[0] >> 32);
3208	#endif
3209	pPatch->ChangedIdt.u16SegSel = ASMGetCS();
3210	#ifdef RT_ARCH_AMD64
3211	pPatch->ChangedIdt.u3IST = 0;
3212	pPatch->ChangedIdt.u5Reserved = 0;
3213	#else /* x86 */
3214	pPatch->ChangedIdt.u5Reserved = 0;
3215	pPatch->ChangedIdt.u3Type1 = 0;
3216	#endif /* x86 */
3217	pPatch->ChangedIdt.u5Type2 = SUPDRV_IDTE_TYPE2_INTERRUPT_GATE;
3218	pPatch->ChangedIdt.u2DPL = 3;
3219	pPatch->ChangedIdt.u1Present = 1;
3220
3221	/*
3222	* Generate the patch code.
3223	*/
3224	{
3225	#ifdef RT_ARCH_AMD64
3226	union
3227	{
3228	uint8_t *pb;
3229	uint32_t *pu32;
3230	uint64_t *pu64;
3231	} u, uFixJmp, uFixCall, uNotNested;
3232	u.pb = &pPatch->auCode[0];
3233
3234	/* check the cookie */
3235	*u.pb++ = 0x3d; // cmp eax, GLOBALCOOKIE
3236	*u.pu32++ = pDevExt->u32Cookie;
3237
3238	*u.pb++ = 0x74; // jz @VBoxCall
3239	*u.pb++ = 2;
3240
3241	/* jump to forwarder code. */
3242	*u.pb++ = 0xeb;
3243	uFixJmp = u;
3244	*u.pb++ = 0xfe;
3245
3246	// @VBoxCall:
3247	*u.pb++ = 0x0f; // swapgs
3248	*u.pb++ = 0x01;
3249	*u.pb++ = 0xf8;
3250
3251	/*
3252	* Call VMMR0Entry
3253	* We don't have to push the arguments here, but we have top
3254	* reserve some stack space for the interrupt forwarding.
3255	*/
3256	# ifdef RT_OS_WINDOWS
3257	*u.pb++ = 0x50; // push rax ; alignment filler.
3258	*u.pb++ = 0x41; // push r8 ; uArg
3259	*u.pb++ = 0x50;
3260	*u.pb++ = 0x52; // push rdx ; uOperation
3261	*u.pb++ = 0x51; // push rcx ; pVM
3262	# else
3263	*u.pb++ = 0x51; // push rcx ; alignment filler.
3264	*u.pb++ = 0x52; // push rdx ; uArg
3265	*u.pb++ = 0x56; // push rsi ; uOperation
3266	*u.pb++ = 0x57; // push rdi ; pVM
3267	# endif
3268
3269	*u.pb++ = 0xff; // call qword [pfnVMMR0EntryInt wrt rip]
3270	*u.pb++ = 0x15;
3271	uFixCall = u;
3272	*u.pu32++ = 0;
3273
3274	*u.pb++ = 0x48; // add rsp, 20h ; remove call frame.
3275	*u.pb++ = 0x81;
3276	*u.pb++ = 0xc4;
3277	*u.pu32++ = 0x20;
3278
3279	*u.pb++ = 0x0f; // swapgs
3280	*u.pb++ = 0x01;
3281	*u.pb++ = 0xf8;
3282
3283	/* Return to R3. */
3284	uNotNested = u;
3285	*u.pb++ = 0x48; // iretq
3286	*u.pb++ = 0xcf;
3287
3288	while ((uintptr_t)u.pb & 0x7) // align 8
3289	*u.pb++ = 0xcc;
3290
3291	/* Pointer to the VMMR0Entry. */ // pfnVMMR0EntryInt dq StubVMMR0Entry
3292	*uFixCall.pu32 = (uint32_t)(u.pb - uFixCall.pb - 4); uFixCall.pb = NULL;
3293	pPatch->offVMMR0EntryFixup = (uint16_t)(u.pb - &pPatch->auCode[0]);
3294	*u.pu64++ = pDevExt->pvVMMR0 ? (uint64_t)pDevExt->pfnVMMR0EntryInt : (uint64_t)u.pb + 8;
3295
3296	/* stub entry. */ // StubVMMR0Entry:
3297	pPatch->offStub = (uint16_t)(u.pb - &pPatch->auCode[0]);
3298	*u.pb++ = 0x33; // xor eax, eax
3299	*u.pb++ = 0xc0;
3300
3301	*u.pb++ = 0x48; // dec rax
3302	*u.pb++ = 0xff;
3303	*u.pb++ = 0xc8;
3304
3305	*u.pb++ = 0xc3; // ret
3306
3307	/* forward to the original handler using a retf. */
3308	*uFixJmp.pb = (uint8_t)(u.pb - uFixJmp.pb - 1); uFixJmp.pb = NULL;
3309
3310	*u.pb++ = 0x68; // push <target cs>
3311	*u.pu32++ = !pPatch->SavedIdt.u5Type2 ? ASMGetCS() : pPatch->SavedIdt.u16SegSel;
3312
3313	*u.pb++ = 0x68; // push <low target rip>
3314	*u.pu32++ = !pPatch->SavedIdt.u5Type2
3315	? (uint32_t)(uintptr_t)uNotNested.pb
3316	: (uint32_t)pPatch->SavedIdt.u16OffsetLow
3317	\| (uint32_t)pPatch->SavedIdt.u16OffsetHigh << 16;
3318
3319	*u.pb++ = 0xc7; // mov dword [rsp + 4], <high target rip>
3320	*u.pb++ = 0x44;
3321	*u.pb++ = 0x24;
3322	*u.pb++ = 0x04;
3323	*u.pu32++ = !pPatch->SavedIdt.u5Type2
3324	? (uint32_t)((uint64_t)uNotNested.pb >> 32)
3325	: pPatch->SavedIdt.u32OffsetTop;
3326
3327	*u.pb++ = 0x48; // retf ; does this require prefix?
3328	*u.pb++ = 0xcb;
3329
3330	#else /* RT_ARCH_X86 */
3331
3332	union
3333	{
3334	uint8_t *pb;
3335	uint16_t *pu16;
3336	uint32_t *pu32;
3337	} u, uFixJmpNotNested, uFixJmp, uFixCall, uNotNested;
3338	u.pb = &pPatch->auCode[0];
3339
3340	/* check the cookie */
3341	*u.pb++ = 0x81; // cmp esi, GLOBALCOOKIE
3342	*u.pb++ = 0xfe;
3343	*u.pu32++ = pDevExt->u32Cookie;
3344
3345	*u.pb++ = 0x74; // jz VBoxCall
3346	uFixJmp = u;
3347	*u.pb++ = 0;
3348
3349	/* jump (far) to the original handler / not-nested-stub. */
3350	*u.pb++ = 0xea; // jmp far NotNested
3351	uFixJmpNotNested = u;
3352	*u.pu32++ = 0;
3353	*u.pu16++ = 0;
3354
3355	/* save selector registers. */ // VBoxCall:
3356	*uFixJmp.pb = (uint8_t)(u.pb - uFixJmp.pb - 1);
3357	*u.pb++ = 0x0f; // push fs
3358	*u.pb++ = 0xa0;
3359
3360	*u.pb++ = 0x1e; // push ds
3361
3362	*u.pb++ = 0x06; // push es
3363
3364	/* call frame */
3365	*u.pb++ = 0x51; // push ecx
3366
3367	*u.pb++ = 0x52; // push edx
3368
3369	*u.pb++ = 0x50; // push eax
3370
3371	/* load ds, es and perhaps fs before call. */
3372	*u.pb++ = 0xb8; // mov eax, KernelDS
3373	*u.pu32++ = ASMGetDS();
3374
3375	*u.pb++ = 0x8e; // mov ds, eax
3376	*u.pb++ = 0xd8;
3377
3378	*u.pb++ = 0x8e; // mov es, eax
3379	*u.pb++ = 0xc0;
3380
3381	#ifdef RT_OS_WINDOWS
3382	*u.pb++ = 0xb8; // mov eax, KernelFS
3383	*u.pu32++ = ASMGetFS();
3384
3385	*u.pb++ = 0x8e; // mov fs, eax
3386	*u.pb++ = 0xe0;
3387	#endif
3388
3389	/* do the call. */
3390	*u.pb++ = 0xe8; // call _VMMR0Entry / StubVMMR0Entry
3391	uFixCall = u;
3392	pPatch->offVMMR0EntryFixup = (uint16_t)(u.pb - &pPatch->auCode[0]);
3393	*u.pu32++ = 0xfffffffb;
3394
3395	*u.pb++ = 0x83; // add esp, 0ch ; cdecl
3396	*u.pb++ = 0xc4;
3397	*u.pb++ = 0x0c;
3398
3399	/* restore selector registers. */
3400	*u.pb++ = 0x07; // pop es
3401	//
3402	*u.pb++ = 0x1f; // pop ds
3403
3404	*u.pb++ = 0x0f; // pop fs
3405	*u.pb++ = 0xa1;
3406
3407	uNotNested = u; // NotNested:
3408	*u.pb++ = 0xcf; // iretd
3409
3410	/* the stub VMMR0Entry. */ // StubVMMR0Entry:
3411	pPatch->offStub = (uint16_t)(u.pb - &pPatch->auCode[0]);
3412	*u.pb++ = 0x33; // xor eax, eax
3413	*u.pb++ = 0xc0;
3414
3415	*u.pb++ = 0x48; // dec eax
3416
3417	*u.pb++ = 0xc3; // ret
3418
3419	/* Fixup the VMMR0Entry call. */
3420	if (pDevExt->pvVMMR0)
3421	*uFixCall.pu32 = (uint32_t)pDevExt->pfnVMMR0EntryInt - (uint32_t)(uFixCall.pu32 + 1);
3422	else
3423	*uFixCall.pu32 = (uint32_t)&pPatch->auCode[pPatch->offStub] - (uint32_t)(uFixCall.pu32 + 1);
3424
3425	/* Fixup the forward / nested far jump. */
3426	if (!pPatch->SavedIdt.u5Type2)
3427	{
3428	*uFixJmpNotNested.pu32++ = (uint32_t)uNotNested.pb;
3429	*uFixJmpNotNested.pu16++ = ASMGetCS();
3430	}
3431	else
3432	{
3433	*uFixJmpNotNested.pu32++ = ((uint32_t)pPatch->SavedIdt.u16OffsetHigh << 16) \| pPatch->SavedIdt.u16OffsetLow;
3434	*uFixJmpNotNested.pu16++ = pPatch->SavedIdt.u16SegSel;
3435	}
3436	#endif /* RT_ARCH_X86 */
3437	Assert(u.pb <= &pPatch->auCode[sizeof(pPatch->auCode)]);
3438	#if 0
3439	/* dump the patch code */
3440	Log2(("patch code: %p\n", &pPatch->auCode[0]));
3441	for (uFixCall.pb = &pPatch->auCode[0]; uFixCall.pb < u.pb; uFixCall.pb++)
3442	Log2(("0x%02x,\n", *uFixCall.pb));
3443	#endif
3444	}
3445
3446	/*
3447	* Install the patch.
3448	*/
3449	supdrvIdtWrite(pPatch->pIdtEntry, &pPatch->ChangedIdt);
3450	AssertMsg(!memcmp((void *)pPatch->pIdtEntry, &pPatch->ChangedIdt, sizeof(pPatch->ChangedIdt)), ("The stupid change code didn't work!!!!!\n"));
3451
3452	/*
3453	* Link in the patch.
3454	*/
3455	pPatch->pNext = pDevExt->pIdtPatches;
3456	pDevExt->pIdtPatches = pPatch;
3457
3458	return pPatch;
3459	}
3460
3461
3462	/**
3463	* Removes the sessions IDT references.
3464	* This will uninstall our IDT patch if we left unreferenced.
3465	*
3466	* @returns VINF_SUCCESS.
3467	* @param pDevExt Device globals.
3468	* @param pSession Session data.
3469	*/
3470	static int supdrvIOCtl_IdtRemoveAll(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession)
3471	{
3472	PSUPDRVPATCHUSAGE pUsage;
3473	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
3474	LogFlow(("supdrvIOCtl_IdtRemoveAll: pSession=%p\n", pSession));
3475
3476	/*
3477	* Take the spinlock.
3478	*/
3479	RTSpinlockAcquireNoInts(pDevExt->Spinlock, &SpinlockTmp);
3480
3481	/*
3482	* Walk usage list, removing patches as their usage count reaches zero.
3483	*/
3484	pUsage = pSession->pPatchUsage;
3485	while (pUsage)
3486	{
3487	if (pUsage->pPatch->cUsage <= pUsage->cUsage)
3488	supdrvIdtRemoveOne(pDevExt, pUsage->pPatch);
3489	else
3490	pUsage->pPatch->cUsage -= pUsage->cUsage;
3491
3492	/* next */
3493	pUsage = pUsage->pNext;
3494	}
3495
3496	/*
3497	* Empty the usage chain and we're done inside the spinlock.
3498	*/
3499	pUsage = pSession->pPatchUsage;
3500	pSession->pPatchUsage = NULL;
3501
3502	RTSpinlockReleaseNoInts(pDevExt->Spinlock, &SpinlockTmp);
3503
3504	/*
3505	* Free usage entries.
3506	*/
3507	while (pUsage)
3508	{
3509	void *pvToFree = pUsage;
3510	pUsage->cUsage = 0;
3511	pUsage->pPatch = NULL;
3512	pUsage = pUsage->pNext;
3513	RTMemFree(pvToFree);
3514	}
3515
3516	return VINF_SUCCESS;
3517	}
3518
3519
3520	/**
3521	* Remove one patch.
3522	*
3523	* Worker for supdrvIOCtl_IdtRemoveAll.
3524	*
3525	* @param pDevExt Device globals.
3526	* @param pPatch Patch entry to remove.
3527	* @remark Caller must own SUPDRVDEVEXT::Spinlock!
3528	*/
3529	static void supdrvIdtRemoveOne(PSUPDRVDEVEXT pDevExt, PSUPDRVPATCH pPatch)
3530	{
3531	LogFlow(("supdrvIdtRemoveOne: pPatch=%p\n", pPatch));
3532
3533	pPatch->cUsage = 0;
3534
3535	/*
3536	* If the IDT entry was changed it have to kick around for ever!
3537	* This will be attempted freed again, perhaps next time we'll succeed :-)
3538	*/
3539	if (memcmp((void *)pPatch->pIdtEntry, &pPatch->ChangedIdt, sizeof(pPatch->ChangedIdt)))
3540	{
3541	AssertMsgFailed(("The hijacked IDT entry has CHANGED!!!\n"));
3542	return;
3543	}
3544
3545	/*
3546	* Unlink it.
3547	*/
3548	if (pDevExt->pIdtPatches != pPatch)
3549	{
3550	PSUPDRVPATCH pPatchPrev = pDevExt->pIdtPatches;
3551	while (pPatchPrev)
3552	{
3553	if (pPatchPrev->pNext == pPatch)
3554	{
3555	pPatchPrev->pNext = pPatch->pNext;
3556	break;
3557	}
3558	pPatchPrev = pPatchPrev->pNext;
3559	}
3560	Assert(!pPatchPrev);
3561	}
3562	else
3563	pDevExt->pIdtPatches = pPatch->pNext;
3564	pPatch->pNext = NULL;
3565
3566
3567	/*
3568	* Verify and restore the IDT.
3569	*/
3570	AssertMsg(!memcmp((void *)pPatch->pIdtEntry, &pPatch->ChangedIdt, sizeof(pPatch->ChangedIdt)), ("The hijacked IDT entry has CHANGED!!!\n"));
3571	supdrvIdtWrite(pPatch->pIdtEntry, &pPatch->SavedIdt);
3572	AssertMsg(!memcmp((void *)pPatch->pIdtEntry, &pPatch->SavedIdt, sizeof(pPatch->SavedIdt)), ("The hijacked IDT entry has CHANGED!!!\n"));
3573
3574	/*
3575	* Put it in the free list.
3576	* (This free list stuff is to calm my paranoia.)
3577	*/
3578	pPatch->pvIdt = NULL;
3579	pPatch->pIdtEntry = NULL;
3580
3581	pPatch->pNext = pDevExt->pIdtPatchesFree;
3582	pDevExt->pIdtPatchesFree = pPatch;
3583	}
3584
3585
3586	/**
3587	* Write to an IDT entry.
3588	*
3589	* @param pvIdtEntry Where to write.
3590	* @param pNewIDTEntry What to write.
3591	*/
3592	static void supdrvIdtWrite(volatile void pvIdtEntry, const SUPDRVIDTE pNewIDTEntry)
3593	{
3594	RTR0UINTREG uCR0;
3595	RTR0UINTREG uFlags;
3596
3597	/*
3598	* On SMP machines (P4 hyperthreading included) we must preform a
3599	* 64-bit locked write when updating the IDT entry.
3600	*
3601	* The F00F bugfix for linux (and probably other OSes) causes
3602	* the IDT to be pointing to an readonly mapping. We get around that
3603	* by temporarily turning of WP. Since we're inside a spinlock at this
3604	* point, interrupts are disabled and there isn't any way the WP bit
3605	* flipping can cause any trouble.
3606	*/
3607
3608	/* Save & Clear interrupt flag; Save & clear WP. */
3609	uFlags = ASMGetFlags();
3610	ASMSetFlags(uFlags & ~(RTR0UINTREG)(1 << 9)); /X86_EFL_IF/
3611	Assert(!(ASMGetFlags() & (1 << 9)));
3612	uCR0 = ASMGetCR0();
3613	ASMSetCR0(uCR0 & ~(RTR0UINTREG)(1 << 16)); /X86_CR0_WP/
3614
3615	/* Update IDT Entry */
3616	#ifdef RT_ARCH_AMD64
3617	ASMAtomicXchgU128((volatile uint128_t )pvIdtEntry, (uint128_t *)(uintptr_t)pNewIDTEntry);
3618	#else
3619	ASMAtomicXchgU64((volatile uint64_t )pvIdtEntry, (uint64_t *)(uintptr_t)pNewIDTEntry);
3620	#endif
3621
3622	/* Restore CR0 & Flags */
3623	ASMSetCR0(uCR0);
3624	ASMSetFlags(uFlags);
3625	}
3626	#endif /* VBOX_WITH_IDT_PATCHING */
3627
3628
3629	/**
3630	* Opens an image. If it's the first time it's opened the call must upload
3631	* the bits using the supdrvIOCtl_LdrLoad() / SUPDRV_IOCTL_LDR_LOAD function.
3632	*
3633	* This is the 1st step of the loading.
3634	*
3635	* @returns IPRT status code.
3636	* @param pDevExt Device globals.
3637	* @param pSession Session data.
3638	* @param pReq The open request.
3639	*/
3640	static int supdrvIOCtl_LdrOpen(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDROPEN pReq)
3641	{
3642	PSUPDRVLDRIMAGE pImage;
3643	unsigned cb;
3644	void *pv;
3645	LogFlow(("supdrvIOCtl_LdrOpen: szName=%s cbImage=%d\n", pReq->u.In.szName, pReq->u.In.cbImage));
3646
3647	/*
3648	* Check if we got an instance of the image already.
3649	*/
3650	RTSemFastMutexRequest(pDevExt->mtxLdr);
3651	for (pImage = pDevExt->pLdrImages; pImage; pImage = pImage->pNext)
3652	{
3653	if (!strcmp(pImage->szName, pReq->u.In.szName))
3654	{
3655	pImage->cUsage++;
3656	pReq->u.Out.pvImageBase = pImage->pvImage;
3657	pReq->u.Out.fNeedsLoading = pImage->uState == SUP_IOCTL_LDR_OPEN;
3658	supdrvLdrAddUsage(pSession, pImage);
3659	RTSemFastMutexRelease(pDevExt->mtxLdr);
3660	return VINF_SUCCESS;
3661	}
3662	}
3663	/* (not found - add it!) */
3664
3665	/*
3666	* Allocate memory.
3667	*/
3668	cb = pReq->u.In.cbImage + sizeof(SUPDRVLDRIMAGE) + 31;
3669	pv = RTMemExecAlloc(cb);
3670	if (!pv)
3671	{
3672	RTSemFastMutexRelease(pDevExt->mtxLdr);
3673	Log(("supdrvIOCtl_LdrOpen: RTMemExecAlloc(%u) failed\n", cb));
3674	return VERR_NO_MEMORY;
3675	}
3676
3677	/*
3678	* Setup and link in the LDR stuff.
3679	*/
3680	pImage = (PSUPDRVLDRIMAGE)pv;
3681	pImage->pvImage = RT_ALIGN_P(pImage + 1, 32);
3682	pImage->cbImage = pReq->u.In.cbImage;
3683	pImage->pfnModuleInit = NULL;
3684	pImage->pfnModuleTerm = NULL;
3685	pImage->uState = SUP_IOCTL_LDR_OPEN;
3686	pImage->cUsage = 1;
3687	strcpy(pImage->szName, pReq->u.In.szName);
3688
3689	pImage->pNext = pDevExt->pLdrImages;
3690	pDevExt->pLdrImages = pImage;
3691
3692	supdrvLdrAddUsage(pSession, pImage);
3693
3694	pReq->u.Out.pvImageBase = pImage->pvImage;
3695	pReq->u.Out.fNeedsLoading = true;
3696	RTSemFastMutexRelease(pDevExt->mtxLdr);
3697	return VINF_SUCCESS;
3698	}
3699
3700
3701	/**
3702	* Loads the image bits.
3703	*
3704	* This is the 2nd step of the loading.
3705	*
3706	* @returns IPRT status code.
3707	* @param pDevExt Device globals.
3708	* @param pSession Session data.
3709	* @param pReq The request.
3710	*/
3711	static int supdrvIOCtl_LdrLoad(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRLOAD pReq)
3712	{
3713	PSUPDRVLDRUSAGE pUsage;
3714	PSUPDRVLDRIMAGE pImage;
3715	int rc;
3716	LogFlow(("supdrvIOCtl_LdrLoad: pvImageBase=%p cbImage=%d\n", pReq->u.In.pvImageBase, pReq->u.In.cbImage));
3717
3718	/*
3719	* Find the ldr image.
3720	*/
3721	RTSemFastMutexRequest(pDevExt->mtxLdr);
3722	pUsage = pSession->pLdrUsage;
3723	while (pUsage && pUsage->pImage->pvImage != pReq->u.In.pvImageBase)
3724	pUsage = pUsage->pNext;
3725	if (!pUsage)
3726	{
3727	RTSemFastMutexRelease(pDevExt->mtxLdr);
3728	Log(("SUP_IOCTL_LDR_LOAD: couldn't find image!\n"));
3729	return VERR_INVALID_HANDLE;
3730	}
3731	pImage = pUsage->pImage;
3732	if (pImage->cbImage != pReq->u.In.cbImage)
3733	{
3734	RTSemFastMutexRelease(pDevExt->mtxLdr);
3735	Log(("SUP_IOCTL_LDR_LOAD: image size mismatch!! %d(prep) != %d(load)\n", pImage->cbImage, pReq->u.In.cbImage));
3736	return VERR_INVALID_HANDLE;
3737	}
3738	if (pImage->uState != SUP_IOCTL_LDR_OPEN)
3739	{
3740	unsigned uState = pImage->uState;
3741	RTSemFastMutexRelease(pDevExt->mtxLdr);
3742	if (uState != SUP_IOCTL_LDR_LOAD)
3743	AssertMsgFailed(("SUP_IOCTL_LDR_LOAD: invalid image state %d (%#x)!\n", uState, uState));
3744	return SUPDRV_ERR_ALREADY_LOADED;
3745	}
3746	switch (pReq->u.In.eEPType)
3747	{
3748	case SUPLDRLOADEP_NOTHING:
3749	break;
3750	case SUPLDRLOADEP_VMMR0:
3751	if ( !pReq->u.In.EP.VMMR0.pvVMMR0
3752	\|\| !pReq->u.In.EP.VMMR0.pvVMMR0EntryInt
3753	\|\| !pReq->u.In.EP.VMMR0.pvVMMR0EntryFast
3754	\|\| !pReq->u.In.EP.VMMR0.pvVMMR0EntryEx)
3755	{
3756	RTSemFastMutexRelease(pDevExt->mtxLdr);
3757	Log(("NULL pointer: pvVMMR0=%p pvVMMR0EntryInt=%p pvVMMR0EntryFast=%p pvVMMR0EntryEx=%p!\n",
3758	pReq->u.In.EP.VMMR0.pvVMMR0, pReq->u.In.EP.VMMR0.pvVMMR0EntryInt,
3759	pReq->u.In.EP.VMMR0.pvVMMR0EntryFast, pReq->u.In.EP.VMMR0.pvVMMR0EntryEx));
3760	return VERR_INVALID_PARAMETER;
3761	}
3762	/** @todo validate pReq->u.In.EP.VMMR0.pvVMMR0 against pvImage! */
3763	if ( (uintptr_t)pReq->u.In.EP.VMMR0.pvVMMR0EntryInt - (uintptr_t)pImage->pvImage >= pReq->u.In.cbImage
3764	\|\| (uintptr_t)pReq->u.In.EP.VMMR0.pvVMMR0EntryFast - (uintptr_t)pImage->pvImage >= pReq->u.In.cbImage
3765	\|\| (uintptr_t)pReq->u.In.EP.VMMR0.pvVMMR0EntryEx - (uintptr_t)pImage->pvImage >= pReq->u.In.cbImage)
3766	{
3767	RTSemFastMutexRelease(pDevExt->mtxLdr);
3768	Log(("Out of range (%p LB %#x): pvVMMR0EntryInt=%p, pvVMMR0EntryFast=%p or pvVMMR0EntryEx=%p is NULL!\n",
3769	pImage->pvImage, pReq->u.In.cbImage, pReq->u.In.EP.VMMR0.pvVMMR0EntryInt,
3770	pReq->u.In.EP.VMMR0.pvVMMR0EntryFast, pReq->u.In.EP.VMMR0.pvVMMR0EntryEx));
3771	return VERR_INVALID_PARAMETER;
3772	}
3773	break;
3774	default:
3775	RTSemFastMutexRelease(pDevExt->mtxLdr);
3776	Log(("Invalid eEPType=%d\n", pReq->u.In.eEPType));
3777	return VERR_INVALID_PARAMETER;
3778	}
3779	if ( pReq->u.In.pfnModuleInit
3780	&& (uintptr_t)pReq->u.In.pfnModuleInit - (uintptr_t)pImage->pvImage >= pReq->u.In.cbImage)
3781	{
3782	RTSemFastMutexRelease(pDevExt->mtxLdr);
3783	Log(("SUP_IOCTL_LDR_LOAD: pfnModuleInit=%p is outside the image (%p %d bytes)\n",
3784	pReq->u.In.pfnModuleInit, pImage->pvImage, pReq->u.In.cbImage));
3785	return VERR_INVALID_PARAMETER;
3786	}
3787	if ( pReq->u.In.pfnModuleTerm
3788	&& (uintptr_t)pReq->u.In.pfnModuleTerm - (uintptr_t)pImage->pvImage >= pReq->u.In.cbImage)
3789	{
3790	RTSemFastMutexRelease(pDevExt->mtxLdr);
3791	Log(("SUP_IOCTL_LDR_LOAD: pfnModuleTerm=%p is outside the image (%p %d bytes)\n",
3792	pReq->u.In.pfnModuleTerm, pImage->pvImage, pReq->u.In.cbImage));
3793	return VERR_INVALID_PARAMETER;
3794	}
3795
3796	/*
3797	* Copy the memory.
3798	*/
3799	/* no need to do try/except as this is a buffered request. */
3800	memcpy(pImage->pvImage, &pReq->u.In.achImage[0], pImage->cbImage);
3801	pImage->uState = SUP_IOCTL_LDR_LOAD;
3802	pImage->pfnModuleInit = pReq->u.In.pfnModuleInit;
3803	pImage->pfnModuleTerm = pReq->u.In.pfnModuleTerm;
3804	pImage->offSymbols = pReq->u.In.offSymbols;
3805	pImage->cSymbols = pReq->u.In.cSymbols;
3806	pImage->offStrTab = pReq->u.In.offStrTab;
3807	pImage->cbStrTab = pReq->u.In.cbStrTab;
3808
3809	/*
3810	* Update any entry points.
3811	*/
3812	switch (pReq->u.In.eEPType)
3813	{
3814	default:
3815	case SUPLDRLOADEP_NOTHING:
3816	rc = VINF_SUCCESS;
3817	break;
3818	case SUPLDRLOADEP_VMMR0:
3819	rc = supdrvLdrSetR0EP(pDevExt, pReq->u.In.EP.VMMR0.pvVMMR0, pReq->u.In.EP.VMMR0.pvVMMR0EntryInt,
3820	pReq->u.In.EP.VMMR0.pvVMMR0EntryFast, pReq->u.In.EP.VMMR0.pvVMMR0EntryEx);
3821	break;
3822	}
3823
3824	/*
3825	* On success call the module initialization.
3826	*/
3827	LogFlow(("supdrvIOCtl_LdrLoad: pfnModuleInit=%p\n", pImage->pfnModuleInit));
3828	if (RT_SUCCESS(rc) && pImage->pfnModuleInit)
3829	{
3830	Log(("supdrvIOCtl_LdrLoad: calling pfnModuleInit=%p\n", pImage->pfnModuleInit));
3831	#ifdef SUPDRV_WITH_UNWIND_HACK
3832	rc = supdrvNtWrapModuleInit((PFNRT)pImage->pfnModuleInit);
3833	#else
3834	rc = pImage->pfnModuleInit();
3835	#endif
3836	if (rc && pDevExt->pvVMMR0 == pImage->pvImage)
3837	supdrvLdrUnsetR0EP(pDevExt);
3838	}
3839
3840	if (rc)
3841	pImage->uState = SUP_IOCTL_LDR_OPEN;
3842
3843	RTSemFastMutexRelease(pDevExt->mtxLdr);
3844	return rc;
3845	}
3846
3847
3848	/**
3849	* Frees a previously loaded (prep'ed) image.
3850	*
3851	* @returns IPRT status code.
3852	* @param pDevExt Device globals.
3853	* @param pSession Session data.
3854	* @param pReq The request.
3855	*/
3856	static int supdrvIOCtl_LdrFree(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRFREE pReq)
3857	{
3858	int rc;
3859	PSUPDRVLDRUSAGE pUsagePrev;
3860	PSUPDRVLDRUSAGE pUsage;
3861	PSUPDRVLDRIMAGE pImage;
3862	LogFlow(("supdrvIOCtl_LdrFree: pvImageBase=%p\n", pReq->u.In.pvImageBase));
3863
3864	/*
3865	* Find the ldr image.
3866	*/
3867	RTSemFastMutexRequest(pDevExt->mtxLdr);
3868	pUsagePrev = NULL;
3869	pUsage = pSession->pLdrUsage;
3870	while (pUsage && pUsage->pImage->pvImage != pReq->u.In.pvImageBase)
3871	{
3872	pUsagePrev = pUsage;
3873	pUsage = pUsage->pNext;
3874	}
3875	if (!pUsage)
3876	{
3877	RTSemFastMutexRelease(pDevExt->mtxLdr);
3878	Log(("SUP_IOCTL_LDR_FREE: couldn't find image!\n"));
3879	return VERR_INVALID_HANDLE;
3880	}
3881
3882	/*
3883	* Check if we can remove anything.
3884	*/
3885	rc = VINF_SUCCESS;
3886	pImage = pUsage->pImage;
3887	if (pImage->cUsage <= 1 \|\| pUsage->cUsage <= 1)
3888	{
3889	/*
3890	* Check if there are any objects with destructors in the image, if
3891	* so leave it for the session cleanup routine so we get a chance to
3892	* clean things up in the right order and not leave them all dangling.
3893	*/
3894	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
3895	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
3896	if (pImage->cUsage <= 1)
3897	{
3898	PSUPDRVOBJ pObj;
3899	for (pObj = pDevExt->pObjs; pObj; pObj = pObj->pNext)
3900	if (RT_UNLIKELY((uintptr_t)pObj->pfnDestructor - (uintptr_t)pImage->pvImage < pImage->cbImage))
3901	{
3902	rc = VERR_SHARING_VIOLATION; /** @todo VERR_DANGLING_OBJECTS */
3903	break;
3904	}
3905	}
3906	else
3907	{
3908	PSUPDRVUSAGE pGenUsage;
3909	for (pGenUsage = pSession->pUsage; pGenUsage; pGenUsage = pGenUsage->pNext)
3910	if (RT_UNLIKELY((uintptr_t)pGenUsage->pObj->pfnDestructor - (uintptr_t)pImage->pvImage < pImage->cbImage))
3911	{
3912	rc = VERR_SHARING_VIOLATION; /** @todo VERR_DANGLING_OBJECTS */
3913	break;
3914	}
3915	}
3916	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
3917	if (rc == VINF_SUCCESS)
3918	{
3919	/* unlink it */
3920	if (pUsagePrev)
3921	pUsagePrev->pNext = pUsage->pNext;
3922	else
3923	pSession->pLdrUsage = pUsage->pNext;
3924
3925	/* free it */
3926	pUsage->pImage = NULL;
3927	pUsage->pNext = NULL;
3928	RTMemFree(pUsage);
3929
3930	/*
3931	* Derefrence the image.
3932	*/
3933	if (pImage->cUsage <= 1)
3934	supdrvLdrFree(pDevExt, pImage);
3935	else
3936	pImage->cUsage--;
3937	}
3938	else
3939	Log(("supdrvIOCtl_LdrFree: Dangling objects in %p/%s!\n", pImage->pvImage, pImage->szName));
3940	}
3941	else
3942	{
3943	/*
3944	* Dereference both image and usage.
3945	*/
3946	pImage->cUsage--;
3947	pUsage->cUsage--;
3948	}
3949
3950	RTSemFastMutexRelease(pDevExt->mtxLdr);
3951	return VINF_SUCCESS;
3952	}
3953
3954
3955	/**
3956	* Gets the address of a symbol in an open image.
3957	*
3958	* @returns 0 on success.
3959	* @returns SUPDRV_ERR_* on failure.
3960	* @param pDevExt Device globals.
3961	* @param pSession Session data.
3962	* @param pReq The request buffer.
3963	*/
3964	static int supdrvIOCtl_LdrGetSymbol(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRGETSYMBOL pReq)
3965	{
3966	PSUPDRVLDRIMAGE pImage;
3967	PSUPDRVLDRUSAGE pUsage;
3968	uint32_t i;
3969	PSUPLDRSYM paSyms;
3970	const char *pchStrings;
3971	const size_t cbSymbol = strlen(pReq->u.In.szSymbol) + 1;
3972	void *pvSymbol = NULL;
3973	int rc = VERR_GENERAL_FAILURE;
3974	Log3(("supdrvIOCtl_LdrGetSymbol: pvImageBase=%p szSymbol=\"%s\"\n", pReq->u.In.pvImageBase, pReq->u.In.szSymbol));
3975
3976	/*
3977	* Find the ldr image.
3978	*/
3979	RTSemFastMutexRequest(pDevExt->mtxLdr);
3980	pUsage = pSession->pLdrUsage;
3981	while (pUsage && pUsage->pImage->pvImage != pReq->u.In.pvImageBase)
3982	pUsage = pUsage->pNext;
3983	if (!pUsage)
3984	{
3985	RTSemFastMutexRelease(pDevExt->mtxLdr);
3986	Log(("SUP_IOCTL_LDR_GET_SYMBOL: couldn't find image!\n"));
3987	return VERR_INVALID_HANDLE;
3988	}
3989	pImage = pUsage->pImage;
3990	if (pImage->uState != SUP_IOCTL_LDR_LOAD)
3991	{
3992	unsigned uState = pImage->uState;
3993	RTSemFastMutexRelease(pDevExt->mtxLdr);
3994	Log(("SUP_IOCTL_LDR_GET_SYMBOL: invalid image state %d (%#x)!\n", uState, uState)); NOREF(uState);
3995	return VERR_ALREADY_LOADED;
3996	}
3997
3998	/*
3999	* Search the symbol strings.
4000	*/
4001	pchStrings = (const char )((uint8_t )pImage->pvImage + pImage->offStrTab);
4002	paSyms = (PSUPLDRSYM)((uint8_t *)pImage->pvImage + pImage->offSymbols);
4003	for (i = 0; i < pImage->cSymbols; i++)
4004	{
4005	if ( paSyms[i].offSymbol < pImage->cbImage /* paranoia */
4006	&& paSyms[i].offName + cbSymbol <= pImage->cbStrTab
4007	&& !memcmp(pchStrings + paSyms[i].offName, pReq->u.In.szSymbol, cbSymbol))
4008	{
4009	pvSymbol = (uint8_t *)pImage->pvImage + paSyms[i].offSymbol;
4010	rc = VINF_SUCCESS;
4011	break;
4012	}
4013	}
4014	RTSemFastMutexRelease(pDevExt->mtxLdr);
4015	pReq->u.Out.pvSymbol = pvSymbol;
4016	return rc;
4017	}
4018
4019
4020	/**
4021	* Gets the address of a symbol in an open image or the support driver.
4022	*
4023	* @returns VINF_SUCCESS on success.
4024	* @returns
4025	* @param pDevExt Device globals.
4026	* @param pSession Session data.
4027	* @param pReq The request buffer.
4028	*/
4029	static int supdrvIDC_LdrGetSymbol(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPDRVIDCREQGETSYM pReq)
4030	{
4031	int rc = VINF_SUCCESS;
4032	const char *pszSymbol = pReq->u.In.pszSymbol;
4033	const char *pszModule = pReq->u.In.pszModule;
4034	size_t cbSymbol;
4035	char const *pszEnd;
4036	uint32_t i;
4037
4038	/*
4039	* Input validation.
4040	*/
4041	AssertPtrReturn(pszSymbol, VERR_INVALID_POINTER);
4042	pszEnd = (char *)memchr(pszSymbol, '\0', 512);
4043	AssertReturn(pszEnd, VERR_INVALID_PARAMETER);
4044	cbSymbol = pszEnd - pszSymbol + 1;
4045
4046	if (pszModule)
4047	{
4048	AssertPtrReturn(pszModule, VERR_INVALID_POINTER);
4049	pszEnd = (char *)memchr(pszModule, '\0', 64);
4050	AssertReturn(pszEnd, VERR_INVALID_PARAMETER);
4051	}
4052	Log3(("supdrvIDC_LdrGetSymbol: pszModule=%p:{%s} pszSymbol=%p:{%s}\n", pszModule, pszModule, pszSymbol, pszSymbol));
4053
4054
4055	if ( !pszModule
4056	\|\| !strcmp(pszModule, "SupDrv"))
4057	{
4058	/*
4059	* Search the support driver export table.
4060	*/
4061	for (i = 0; i < RT_ELEMENTS(g_aFunctions); i++)
4062	if (!strcmp(g_aFunctions[i].szName, pszSymbol))
4063	{
4064	pReq->u.Out.pfnSymbol = g_aFunctions[i].pfn;
4065	break;
4066	}
4067	}
4068	else
4069	{
4070	/*
4071	* Find the loader image.
4072	*/
4073	PSUPDRVLDRIMAGE pImage;
4074
4075	RTSemFastMutexRequest(pDevExt->mtxLdr);
4076
4077	for (pImage = pDevExt->pLdrImages; pImage; pImage = pImage->pNext)
4078	if (!strcmp(pImage->szName, pszModule))
4079	break;
4080	if (pImage && pImage->uState == SUP_IOCTL_LDR_LOAD)
4081	{
4082	/*
4083	* Search the symbol strings.
4084	*/
4085	const char pchStrings = (const char )((uint8_t *)pImage->pvImage + pImage->offStrTab);
4086	PCSUPLDRSYM paSyms = (PCSUPLDRSYM)((uint8_t *)pImage->pvImage + pImage->offSymbols);
4087	for (i = 0; i < pImage->cSymbols; i++)
4088	{
4089	if ( paSyms[i].offSymbol < pImage->cbImage /* paranoia */
4090	&& paSyms[i].offName + cbSymbol <= pImage->cbStrTab
4091	&& !memcmp(pchStrings + paSyms[i].offName, pszSymbol, cbSymbol))
4092	{
4093	/*
4094	* Found it! Calc the symbol address and add a reference to the module.
4095	*/
4096	pReq->u.Out.pfnSymbol = (PFNRT)((uint8_t *)pImage->pvImage + paSyms[i].offSymbol);
4097	rc = supdrvLdrAddUsage(pSession, pImage);
4098	break;
4099	}
4100	}
4101	}
4102	else
4103	rc = pImage ? VERR_WRONG_ORDER : VERR_MODULE_NOT_FOUND;
4104
4105	RTSemFastMutexRelease(pDevExt->mtxLdr);
4106	}
4107	return rc;
4108	}
4109
4110
4111	/**
4112	* Updates the IDT patches to point to the specified VMM R0 entry
4113	* point (i.e. VMMR0Enter()).
4114	*
4115	* @returns IPRT status code.
4116	* @param pDevExt Device globals.
4117	* @param pSession Session data.
4118	* @param pVMMR0 VMMR0 image handle.
4119	* @param pvVMMR0EntryInt VMMR0EntryInt address.
4120	* @param pvVMMR0EntryFast VMMR0EntryFast address.
4121	* @param pvVMMR0EntryEx VMMR0EntryEx address.
4122	* @remark Caller must own the loader mutex.
4123	*/
4124	static int supdrvLdrSetR0EP(PSUPDRVDEVEXT pDevExt, void pvVMMR0, void pvVMMR0EntryInt, void pvVMMR0EntryFast, void pvVMMR0EntryEx)
4125	{
4126	int rc = VINF_SUCCESS;
4127	LogFlow(("supdrvLdrSetR0EP pvVMMR0=%p pvVMMR0EntryInt=%p\n", pvVMMR0, pvVMMR0EntryInt));
4128
4129
4130	/*
4131	* Check if not yet set.
4132	*/
4133	if (!pDevExt->pvVMMR0)
4134	{
4135	#ifdef VBOX_WITH_IDT_PATCHING
4136	PSUPDRVPATCH pPatch;
4137	#endif
4138
4139	/*
4140	* Set it and update IDT patch code.
4141	*/
4142	pDevExt->pvVMMR0 = pvVMMR0;
4143	pDevExt->pfnVMMR0EntryInt = pvVMMR0EntryInt;
4144	pDevExt->pfnVMMR0EntryFast = pvVMMR0EntryFast;
4145	pDevExt->pfnVMMR0EntryEx = pvVMMR0EntryEx;
4146	#ifdef VBOX_WITH_IDT_PATCHING
4147	for (pPatch = pDevExt->pIdtPatches; pPatch; pPatch = pPatch->pNext)
4148	{
4149	# ifdef RT_ARCH_AMD64
4150	ASMAtomicXchgU64((volatile uint64_t *)&pPatch->auCode[pPatch->offVMMR0EntryFixup], (uint64_t)pvVMMR0);
4151	# else /* RT_ARCH_X86 */
4152	ASMAtomicXchgU32((volatile uint32_t *)&pPatch->auCode[pPatch->offVMMR0EntryFixup],
4153	(uint32_t)pvVMMR0 - (uint32_t)&pPatch->auCode[pPatch->offVMMR0EntryFixup + 4]);
4154	# endif
4155	}
4156	#endif /* VBOX_WITH_IDT_PATCHING */
4157	}
4158	else
4159	{
4160	/*
4161	* Return failure or success depending on whether the values match or not.
4162	*/
4163	if ( pDevExt->pvVMMR0 != pvVMMR0
4164	\|\| (void *)pDevExt->pfnVMMR0EntryInt != pvVMMR0EntryInt
4165	\|\| (void *)pDevExt->pfnVMMR0EntryFast != pvVMMR0EntryFast
4166	\|\| (void *)pDevExt->pfnVMMR0EntryEx != pvVMMR0EntryEx)
4167	{
4168	AssertMsgFailed(("SUP_IOCTL_LDR_SETR0EP: Already set pointing to a different module!\n"));
4169	rc = VERR_INVALID_PARAMETER;
4170	}
4171	}
4172	return rc;
4173	}
4174
4175
4176	/**
4177	* Unsets the R0 entry point installed by supdrvLdrSetR0EP.
4178	*
4179	* @param pDevExt Device globals.
4180	*/
4181	static void supdrvLdrUnsetR0EP(PSUPDRVDEVEXT pDevExt)
4182	{
4183	#ifdef VBOX_WITH_IDT_PATCHING
4184	PSUPDRVPATCH pPatch;
4185	#endif
4186
4187	pDevExt->pvVMMR0 = NULL;
4188	pDevExt->pfnVMMR0EntryInt = NULL;
4189	pDevExt->pfnVMMR0EntryFast = NULL;
4190	pDevExt->pfnVMMR0EntryEx = NULL;
4191
4192	#ifdef VBOX_WITH_IDT_PATCHING
4193	for (pPatch = pDevExt->pIdtPatches; pPatch; pPatch = pPatch->pNext)
4194	{
4195	# ifdef RT_ARCH_AMD64
4196	ASMAtomicXchgU64((volatile uint64_t *)&pPatch->auCode[pPatch->offVMMR0EntryFixup],
4197	(uint64_t)&pPatch->auCode[pPatch->offStub]);
4198	# else /* RT_ARCH_X86 */
4199	ASMAtomicXchgU32((volatile uint32_t *)&pPatch->auCode[pPatch->offVMMR0EntryFixup],
4200	(uint32_t)&pPatch->auCode[pPatch->offStub] - (uint32_t)&pPatch->auCode[pPatch->offVMMR0EntryFixup + 4]);
4201	# endif
4202	}
4203	#endif /* VBOX_WITH_IDT_PATCHING */
4204	}
4205
4206
4207	/**
4208	* Adds a usage reference in the specified session of an image.
4209	*
4210	* Called while owning the loader semaphore.
4211	*
4212	* @returns VINF_SUCCESS on success and VERR_NO_MEMORY on failure.
4213	* @param pSession Session in question.
4214	* @param pImage Image which the session is using.
4215	*/
4216	static int supdrvLdrAddUsage(PSUPDRVSESSION pSession, PSUPDRVLDRIMAGE pImage)
4217	{
4218	PSUPDRVLDRUSAGE pUsage;
4219	LogFlow(("supdrvLdrAddUsage: pImage=%p\n", pImage));
4220
4221	/*
4222	* Referenced it already?
4223	*/
4224	pUsage = pSession->pLdrUsage;
4225	while (pUsage)
4226	{
4227	if (pUsage->pImage == pImage)
4228	{
4229	pUsage->cUsage++;
4230	return VINF_SUCCESS;
4231	}
4232	pUsage = pUsage->pNext;
4233	}
4234
4235	/*
4236	* Allocate new usage record.
4237	*/
4238	pUsage = (PSUPDRVLDRUSAGE)RTMemAlloc(sizeof(*pUsage));
4239	AssertReturn(pUsage, VERR_NO_MEMORY);
4240	pUsage->cUsage = 1;
4241	pUsage->pImage = pImage;
4242	pUsage->pNext = pSession->pLdrUsage;
4243	pSession->pLdrUsage = pUsage;
4244	return VINF_SUCCESS;
4245	}
4246
4247
4248	/**
4249	* Frees a load image.
4250	*
4251	* @param pDevExt Pointer to device extension.
4252	* @param pImage Pointer to the image we're gonna free.
4253	* This image must exit!
4254	* @remark The caller MUST own SUPDRVDEVEXT::mtxLdr!
4255	*/
4256	static void supdrvLdrFree(PSUPDRVDEVEXT pDevExt, PSUPDRVLDRIMAGE pImage)
4257	{
4258	PSUPDRVLDRIMAGE pImagePrev;
4259	LogFlow(("supdrvLdrFree: pImage=%p\n", pImage));
4260
4261	/* find it - arg. should've used doubly linked list. */
4262	Assert(pDevExt->pLdrImages);
4263	pImagePrev = NULL;
4264	if (pDevExt->pLdrImages != pImage)
4265	{
4266	pImagePrev = pDevExt->pLdrImages;
4267	while (pImagePrev->pNext != pImage)
4268	pImagePrev = pImagePrev->pNext;
4269	Assert(pImagePrev->pNext == pImage);
4270	}
4271
4272	/* unlink */
4273	if (pImagePrev)
4274	pImagePrev->pNext = pImage->pNext;
4275	else
4276	pDevExt->pLdrImages = pImage->pNext;
4277
4278	/* check if this is VMMR0.r0 and fix the Idt patches if it is. */
4279	if (pDevExt->pvVMMR0 == pImage->pvImage)
4280	supdrvLdrUnsetR0EP(pDevExt);
4281
4282	/* check for objects with destructors in this image. (Shouldn't happen.) */
4283	if (pDevExt->pObjs)
4284	{
4285	unsigned cObjs = 0;
4286	PSUPDRVOBJ pObj;
4287	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
4288	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
4289	for (pObj = pDevExt->pObjs; pObj; pObj = pObj->pNext)
4290	if (RT_UNLIKELY((uintptr_t)pObj->pfnDestructor - (uintptr_t)pImage->pvImage < pImage->cbImage))
4291	{
4292	pObj->pfnDestructor = NULL;
4293	cObjs++;
4294	}
4295	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
4296	if (cObjs)
4297	OSDBGPRINT(("supdrvLdrFree: Image '%s' has %d dangling objects!\n", pImage->szName, cObjs));
4298	}
4299
4300	/* call termination function if fully loaded. */
4301	if ( pImage->pfnModuleTerm
4302	&& pImage->uState == SUP_IOCTL_LDR_LOAD)
4303	{
4304	LogFlow(("supdrvIOCtl_LdrLoad: calling pfnModuleTerm=%p\n", pImage->pfnModuleTerm));
4305	#ifdef SUPDRV_WITH_UNWIND_HACK
4306	supdrvNtWrapModuleTerm(pImage->pfnModuleTerm);
4307	#else
4308	pImage->pfnModuleTerm();
4309	#endif
4310	}
4311
4312	/* free the image */
4313	pImage->cUsage = 0;
4314	pImage->pNext = 0;
4315	pImage->uState = SUP_IOCTL_LDR_FREE;
4316	RTMemExecFree(pImage);
4317	}
4318
4319
4320	/**
4321	* Gets the current paging mode of the CPU and stores in in pOut.
4322	*/
4323	static SUPPAGINGMODE supdrvIOCtl_GetPagingMode(void)
4324	{
4325	SUPPAGINGMODE enmMode;
4326
4327	RTR0UINTREG cr0 = ASMGetCR0();
4328	if ((cr0 & (X86_CR0_PG \| X86_CR0_PE)) != (X86_CR0_PG \| X86_CR0_PE))
4329	enmMode = SUPPAGINGMODE_INVALID;
4330	else
4331	{
4332	RTR0UINTREG cr4 = ASMGetCR4();
4333	uint32_t fNXEPlusLMA = 0;
4334	if (cr4 & X86_CR4_PAE)
4335	{
4336	uint32_t fAmdFeatures = ASMCpuId_EDX(0x80000001);
4337	if (fAmdFeatures & (X86_CPUID_AMD_FEATURE_EDX_NX \| X86_CPUID_AMD_FEATURE_EDX_LONG_MODE))
4338	{
4339	uint64_t efer = ASMRdMsr(MSR_K6_EFER);
4340	if ((fAmdFeatures & X86_CPUID_AMD_FEATURE_EDX_NX) && (efer & MSR_K6_EFER_NXE))
4341	fNXEPlusLMA \|= RT_BIT(0);
4342	if ((fAmdFeatures & X86_CPUID_AMD_FEATURE_EDX_LONG_MODE) && (efer & MSR_K6_EFER_LMA))
4343	fNXEPlusLMA \|= RT_BIT(1);
4344	}
4345	}
4346
4347	switch ((cr4 & (X86_CR4_PAE \| X86_CR4_PGE)) \| fNXEPlusLMA)
4348	{
4349	case 0:
4350	enmMode = SUPPAGINGMODE_32_BIT;
4351	break;
4352
4353	case X86_CR4_PGE:
4354	enmMode = SUPPAGINGMODE_32_BIT_GLOBAL;
4355	break;
4356
4357	case X86_CR4_PAE:
4358	enmMode = SUPPAGINGMODE_PAE;
4359	break;
4360
4361	case X86_CR4_PAE \| RT_BIT(0):
4362	enmMode = SUPPAGINGMODE_PAE_NX;
4363	break;
4364
4365	case X86_CR4_PAE \| X86_CR4_PGE:
4366	enmMode = SUPPAGINGMODE_PAE_GLOBAL;
4367	break;
4368
4369	case X86_CR4_PAE \| X86_CR4_PGE \| RT_BIT(0):
4370	enmMode = SUPPAGINGMODE_PAE_GLOBAL;
4371	break;
4372
4373	case RT_BIT(1) \| X86_CR4_PAE:
4374	enmMode = SUPPAGINGMODE_AMD64;
4375	break;
4376
4377	case RT_BIT(1) \| X86_CR4_PAE \| RT_BIT(0):
4378	enmMode = SUPPAGINGMODE_AMD64_NX;
4379	break;
4380
4381	case RT_BIT(1) \| X86_CR4_PAE \| X86_CR4_PGE:
4382	enmMode = SUPPAGINGMODE_AMD64_GLOBAL;
4383	break;
4384
4385	case RT_BIT(1) \| X86_CR4_PAE \| X86_CR4_PGE \| RT_BIT(0):
4386	enmMode = SUPPAGINGMODE_AMD64_GLOBAL_NX;
4387	break;
4388
4389	default:
4390	AssertMsgFailed(("Cannot happen! cr4=%#x fNXEPlusLMA=%d\n", cr4, fNXEPlusLMA));
4391	enmMode = SUPPAGINGMODE_INVALID;
4392	break;
4393	}
4394	}
4395	return enmMode;
4396	}
4397
4398
4399	/**
4400	* Creates the GIP.
4401	*
4402	* @returns negative errno.
4403	* @param pDevExt Instance data. GIP stuff may be updated.
4404	*/
4405	static int supdrvGipCreate(PSUPDRVDEVEXT pDevExt)
4406	{
4407	PSUPGLOBALINFOPAGE pGip;
4408	RTHCPHYS HCPhysGip;
4409	uint32_t u32SystemResolution;
4410	uint32_t u32Interval;
4411	int rc;
4412
4413	LogFlow(("supdrvGipCreate:\n"));
4414
4415	/* assert order */
4416	Assert(pDevExt->u32SystemTimerGranularityGrant == 0);
4417	Assert(pDevExt->GipMemObj == NIL_RTR0MEMOBJ);
4418	Assert(!pDevExt->pGipTimer);
4419
4420	/*
4421	* Allocate a suitable page with a default kernel mapping.
4422	*/
4423	rc = RTR0MemObjAllocLow(&pDevExt->GipMemObj, PAGE_SIZE, false);
4424	if (RT_FAILURE(rc))
4425	{
4426	OSDBGPRINT(("supdrvGipCreate: failed to allocate the GIP page. rc=%d\n", rc));
4427	return rc;
4428	}
4429	pGip = (PSUPGLOBALINFOPAGE)RTR0MemObjAddress(pDevExt->GipMemObj); AssertPtr(pGip);
4430	HCPhysGip = RTR0MemObjGetPagePhysAddr(pDevExt->GipMemObj, 0); Assert(HCPhysGip != NIL_RTHCPHYS);
4431
4432	#if 0 /** @todo Disabled this as we didn't used to do it before and causes unnecessary stress on laptops.
4433	* It only applies to Windows and should probably revisited later, if possible made part of the
4434	* timer code (return min granularity in RTTimerGetSystemGranularity and set it in RTTimerStart). */
4435	/*
4436	* Try bump up the system timer resolution.
4437	* The more interrupts the better...
4438	*/
4439	if ( RT_SUCCESS(RTTimerRequestSystemGranularity( 488281 /* 2048 HZ */, &u32SystemResolution))
4440	\|\| RT_SUCCESS(RTTimerRequestSystemGranularity( 500000 /* 2000 HZ */, &u32SystemResolution))
4441	\|\| RT_SUCCESS(RTTimerRequestSystemGranularity( 976563 /* 1024 HZ */, &u32SystemResolution))
4442	\|\| RT_SUCCESS(RTTimerRequestSystemGranularity( 1000000 /* 1000 HZ */, &u32SystemResolution))
4443	\|\| RT_SUCCESS(RTTimerRequestSystemGranularity( 1953125 /* 512 HZ */, &u32SystemResolution))
4444	\|\| RT_SUCCESS(RTTimerRequestSystemGranularity( 2000000 /* 500 HZ */, &u32SystemResolution))
4445	\|\| RT_SUCCESS(RTTimerRequestSystemGranularity( 3906250 /* 256 HZ */, &u32SystemResolution))
4446	\|\| RT_SUCCESS(RTTimerRequestSystemGranularity( 4000000 /* 250 HZ */, &u32SystemResolution))
4447	\|\| RT_SUCCESS(RTTimerRequestSystemGranularity( 7812500 /* 128 HZ */, &u32SystemResolution))
4448	\|\| RT_SUCCESS(RTTimerRequestSystemGranularity(10000000 /* 100 HZ */, &u32SystemResolution))
4449	\|\| RT_SUCCESS(RTTimerRequestSystemGranularity(15625000 /* 64 HZ */, &u32SystemResolution))
4450	\|\| RT_SUCCESS(RTTimerRequestSystemGranularity(31250000 /* 32 HZ */, &u32SystemResolution))
4451	)
4452	{
4453	Assert(RTTimerGetSystemGranularity() <= u32SystemResolution);
4454	pDevExt->u32SystemTimerGranularityGrant = u32SystemResolution;
4455	}
4456	#endif
4457
4458	/*
4459	* Find a reasonable update interval and initialize the structure.
4460	*/
4461	u32Interval = u32SystemResolution = RTTimerGetSystemGranularity();
4462	while (u32Interval < 10000000 /* 10 ms */)
4463	u32Interval += u32SystemResolution;
4464
4465	supdrvGipInit(pDevExt, pGip, HCPhysGip, RTTimeSystemNanoTS(), 1000000000 / u32Interval /=Hz/);
4466
4467	/*
4468	* Create the timer.
4469	* If CPU_ALL isn't supported we'll have to fall back to synchronous mode.
4470	*/
4471	if (pGip->u32Mode == SUPGIPMODE_ASYNC_TSC)
4472	{
4473	rc = RTTimerCreateEx(&pDevExt->pGipTimer, u32Interval, RTTIMER_FLAGS_CPU_ALL, supdrvGipAsyncTimer, pDevExt);
4474	if (rc == VERR_NOT_SUPPORTED)
4475	{
4476	OSDBGPRINT(("supdrvGipCreate: omni timer not supported, falling back to synchronous mode\n"));
4477	pGip->u32Mode = SUPGIPMODE_SYNC_TSC;
4478	}
4479	}
4480	if (pGip->u32Mode != SUPGIPMODE_ASYNC_TSC)
4481	rc = RTTimerCreateEx(&pDevExt->pGipTimer, u32Interval, 0, supdrvGipSyncTimer, pDevExt);
4482	if (RT_SUCCESS(rc))
4483	{
4484	if (pGip->u32Mode == SUPGIPMODE_ASYNC_TSC)
4485	rc = RTMpNotificationRegister(supdrvGipMpEvent, pDevExt);
4486	if (RT_SUCCESS(rc))
4487	{
4488	/*
4489	* We're good.
4490	*/
4491	dprintf(("supdrvGipCreate: %ld ns interval.\n", (long)u32Interval));
4492	return VINF_SUCCESS;
4493	}
4494
4495	OSDBGPRINT(("supdrvGipCreate: failed register MP event notfication. rc=%d\n", rc));
4496	}
4497	else
4498	{
4499	OSDBGPRINT(("supdrvGipCreate: failed create GIP timer at %ld ns interval. rc=%d\n", (long)u32Interval, rc));
4500	Assert(!pDevExt->pGipTimer);
4501	}
4502	supdrvGipDestroy(pDevExt);
4503	return rc;
4504	}
4505
4506
4507	/**
4508	* Terminates the GIP.
4509	*
4510	* @param pDevExt Instance data. GIP stuff may be updated.
4511	*/
4512	static void supdrvGipDestroy(PSUPDRVDEVEXT pDevExt)
4513	{
4514	int rc;
4515	#ifdef DEBUG_DARWIN_GIP
4516	OSDBGPRINT(("supdrvGipDestroy: pDevExt=%p pGip=%p pGipTimer=%p GipMemObj=%p\n", pDevExt,
4517	pDevExt->GipMemObj != NIL_RTR0MEMOBJ ? RTR0MemObjAddress(pDevExt->GipMemObj) : NULL,
4518	pDevExt->pGipTimer, pDevExt->GipMemObj));
4519	#endif
4520
4521	/*
4522	* Invalid the GIP data.
4523	*/
4524	if (pDevExt->pGip)
4525	{
4526	supdrvGipTerm(pDevExt->pGip);
4527	pDevExt->pGip = NULL;
4528	}
4529
4530	/*
4531	* Destroy the timer and free the GIP memory object.
4532	*/
4533	if (pDevExt->pGipTimer)
4534	{
4535	rc = RTTimerDestroy(pDevExt->pGipTimer); AssertRC(rc);
4536	pDevExt->pGipTimer = NULL;
4537	}
4538
4539	if (pDevExt->GipMemObj != NIL_RTR0MEMOBJ)
4540	{
4541	rc = RTR0MemObjFree(pDevExt->GipMemObj, true /* free mappings */); AssertRC(rc);
4542	pDevExt->GipMemObj = NIL_RTR0MEMOBJ;
4543	}
4544
4545	/*
4546	* Finally, release the system timer resolution request if one succeeded.
4547	*/
4548	if (pDevExt->u32SystemTimerGranularityGrant)
4549	{
4550	rc = RTTimerReleaseSystemGranularity(pDevExt->u32SystemTimerGranularityGrant); AssertRC(rc);
4551	pDevExt->u32SystemTimerGranularityGrant = 0;
4552	}
4553	}
4554
4555
4556	/**
4557	* Timer callback function sync GIP mode.
4558	* @param pTimer The timer.
4559	* @param pvUser The device extension.
4560	*/
4561	static DECLCALLBACK(void) supdrvGipSyncTimer(PRTTIMER pTimer, void *pvUser, uint64_t iTick)
4562	{
4563	PSUPDRVDEVEXT pDevExt = (PSUPDRVDEVEXT)pvUser;
4564	supdrvGipUpdate(pDevExt->pGip, RTTimeSystemNanoTS());
4565	}
4566
4567
4568	/**
4569	* Timer callback function for async GIP mode.
4570	* @param pTimer The timer.
4571	* @param pvUser The device extension.
4572	*/
4573	static DECLCALLBACK(void) supdrvGipAsyncTimer(PRTTIMER pTimer, void *pvUser, uint64_t iTick)
4574	{
4575	PSUPDRVDEVEXT pDevExt = (PSUPDRVDEVEXT)pvUser;
4576	RTCPUID idCpu = RTMpCpuId();
4577	uint64_t NanoTS = RTTimeSystemNanoTS();
4578
4579	/** @todo reset the transaction number and whatnot when iTick == 1. */
4580	if (pDevExt->idGipMaster == idCpu)
4581	supdrvGipUpdate(pDevExt->pGip, NanoTS);
4582	else
4583	supdrvGipUpdatePerCpu(pDevExt->pGip, NanoTS, ASMGetApicId());
4584	}
4585
4586
4587	/**
4588	* Multiprocessor event notification callback.
4589	*
4590	* This is used to make sue that the GIP master gets passed on to
4591	* another CPU.
4592	*
4593	* @param enmEvent The event.
4594	* @param idCpu The cpu it applies to.
4595	* @param pvUser Pointer to the device extension.
4596	*/
4597	static DECLCALLBACK(void) supdrvGipMpEvent(RTMPEVENT enmEvent, RTCPUID idCpu, void *pvUser)
4598	{
4599	PSUPDRVDEVEXT pDevExt = (PSUPDRVDEVEXT)pvUser;
4600	if (enmEvent == RTMPEVENT_OFFLINE)
4601	{
4602	RTCPUID idGipMaster;
4603	ASMAtomicReadSize(&pDevExt->idGipMaster, &idGipMaster);
4604	if (idGipMaster == idCpu)
4605	{
4606	/*
4607	* Find a new GIP master.
4608	*/
4609	bool fIgnored;
4610	unsigned i;
4611	RTCPUID idNewGipMaster = NIL_RTCPUID;
4612	RTCPUSET OnlineCpus;
4613	RTMpGetOnlineSet(&OnlineCpus);
4614
4615	for (i = 0; i < RTCPUSET_MAX_CPUS; i++)
4616	{
4617	RTCPUID idCurCpu = RTMpCpuIdFromSetIndex(i);
4618	if ( RTCpuSetIsMember(&OnlineCpus, idCurCpu)
4619	&& idCurCpu != idGipMaster)
4620	{
4621	idNewGipMaster = idCurCpu;
4622	break;
4623	}
4624	}
4625
4626	dprintf(("supdrvGipMpEvent: Gip master %#lx -> %#lx\n", (long)idGipMaster, (long)idNewGipMaster));
4627	ASMAtomicCmpXchgSize(&pDevExt->idGipMaster, idNewGipMaster, idGipMaster, fIgnored);
4628	NOREF(fIgnored);
4629	}
4630	}
4631	}
4632
4633
4634	/**
4635	* Initializes the GIP data.
4636	*
4637	* @returns IPRT status code.
4638	* @param pDevExt Pointer to the device instance data.
4639	* @param pGip Pointer to the read-write kernel mapping of the GIP.
4640	* @param HCPhys The physical address of the GIP.
4641	* @param u64NanoTS The current nanosecond timestamp.
4642	* @param uUpdateHz The update freqence.
4643	*/
4644	int VBOXCALL supdrvGipInit(PSUPDRVDEVEXT pDevExt, PSUPGLOBALINFOPAGE pGip, RTHCPHYS HCPhys, uint64_t u64NanoTS, unsigned uUpdateHz)
4645	{
4646	unsigned i;
4647	#ifdef DEBUG_DARWIN_GIP
4648	OSDBGPRINT(("supdrvGipInit: pGip=%p HCPhys=%lx u64NanoTS=%llu uUpdateHz=%d\n", pGip, (long)HCPhys, u64NanoTS, uUpdateHz));
4649	#else
4650	LogFlow(("supdrvGipInit: pGip=%p HCPhys=%lx u64NanoTS=%llu uUpdateHz=%d\n", pGip, (long)HCPhys, u64NanoTS, uUpdateHz));
4651	#endif
4652
4653	/*
4654	* Initialize the structure.
4655	*/
4656	memset(pGip, 0, PAGE_SIZE);
4657	pGip->u32Magic = SUPGLOBALINFOPAGE_MAGIC;
4658	pGip->u32Version = SUPGLOBALINFOPAGE_VERSION;
4659	pGip->u32Mode = supdrvGipDeterminTscMode(pDevExt);
4660	pGip->u32UpdateHz = uUpdateHz;
4661	pGip->u32UpdateIntervalNS = 1000000000 / uUpdateHz;
4662	pGip->u64NanoTSLastUpdateHz = u64NanoTS;
4663
4664	for (i = 0; i < RT_ELEMENTS(pGip->aCPUs); i++)
4665	{
4666	pGip->aCPUs[i].u32TransactionId = 2;
4667	pGip->aCPUs[i].u64NanoTS = u64NanoTS;
4668	pGip->aCPUs[i].u64TSC = ASMReadTSC();
4669
4670	/*
4671	* We don't know the following values until we've executed updates.
4672	* So, we'll just insert very high values.
4673	*/
4674	pGip->aCPUs[i].u64CpuHz = _4G + 1;
4675	pGip->aCPUs[i].u32UpdateIntervalTSC = _2G / 4;
4676	pGip->aCPUs[i].au32TSCHistory[0] = _2G / 4;
4677	pGip->aCPUs[i].au32TSCHistory[1] = _2G / 4;
4678	pGip->aCPUs[i].au32TSCHistory[2] = _2G / 4;
4679	pGip->aCPUs[i].au32TSCHistory[3] = _2G / 4;
4680	pGip->aCPUs[i].au32TSCHistory[4] = _2G / 4;
4681	pGip->aCPUs[i].au32TSCHistory[5] = _2G / 4;
4682	pGip->aCPUs[i].au32TSCHistory[6] = _2G / 4;
4683	pGip->aCPUs[i].au32TSCHistory[7] = _2G / 4;
4684	}
4685
4686	/*
4687	* Link it to the device extension.
4688	*/
4689	pDevExt->pGip = pGip;
4690	pDevExt->HCPhysGip = HCPhys;
4691	pDevExt->cGipUsers = 0;
4692
4693	return VINF_SUCCESS;
4694	}
4695
4696
4697	/**
4698	* Callback used by supdrvDetermineAsyncTSC to read the TSC on a CPU.
4699	*
4700	* @param idCpu Ignored.
4701	* @param pvUser1 Where to put the TSC.
4702	* @param pvUser2 Ignored.
4703	*/
4704	static DECLCALLBACK(void) supdrvDetermineAsyncTscWorker(RTCPUID idCpu, void pvUser1, void pvUser2)
4705	{
4706	#if 1
4707	ASMAtomicWriteU64((uint64_t volatile *)pvUser1, ASMReadTSC());
4708	#else
4709	(uint64_t )pvUser1 = ASMReadTSC();
4710	#endif
4711	}
4712
4713
4714	/**
4715	* Determine if Async GIP mode is required because of TSC drift.
4716	*
4717	* When using the default/normal timer code it is essential that the time stamp counter
4718	* (TSC) runs never backwards, that is, a read operation to the counter should return
4719	* a bigger value than any previous read operation. This is guaranteed by the latest
4720	* AMD CPUs and by newer Intel CPUs which never enter the C2 state (P4). In any other
4721	* case we have to choose the asynchronous timer mode.
4722	*
4723	* @param poffMin Pointer to the determined difference between different cores.
4724	* @return false if the time stamp counters appear to be synchron, true otherwise.
4725	*/
4726	bool VBOXCALL supdrvDetermineAsyncTsc(uint64_t *poffMin)
4727	{
4728	/*
4729	* Just iterate all the cpus 8 times and make sure that the TSC is
4730	* ever increasing. We don't bother taking TSC rollover into account.
4731	*/
4732	RTCPUSET CpuSet;
4733	int iLastCpu = RTCpuLastIndex(RTMpGetSet(&CpuSet));
4734	int iCpu;
4735	int cLoops = 8;
4736	bool fAsync = false;
4737	int rc = VINF_SUCCESS;
4738	uint64_t offMax = 0;
4739	uint64_t offMin = ~(uint64_t)0;
4740	uint64_t PrevTsc = ASMReadTSC();
4741
4742	while (cLoops-- > 0)
4743	{
4744	for (iCpu = 0; iCpu <= iLastCpu; iCpu++)
4745	{
4746	uint64_t CurTsc;
4747	rc = RTMpOnSpecific(RTMpCpuIdFromSetIndex(iCpu), supdrvDetermineAsyncTscWorker, &CurTsc, NULL);
4748	if (RT_SUCCESS(rc))
4749	{
4750	if (CurTsc <= PrevTsc)
4751	{
4752	fAsync = true;
4753	offMin = offMax = PrevTsc - CurTsc;
4754	dprintf(("supdrvDetermineAsyncTsc: iCpu=%d cLoops=%d CurTsc=%llx PrevTsc=%llx\n",
4755	iCpu, cLoops, CurTsc, PrevTsc));
4756	break;
4757	}
4758
4759	/* Gather statistics (except the first time). */
4760	if (iCpu != 0 \|\| cLoops != 7)
4761	{
4762	uint64_t off = CurTsc - PrevTsc;
4763	if (off < offMin)
4764	offMin = off;
4765	if (off > offMax)
4766	offMax = off;
4767	dprintf2(("%d/%d: off=%llx\n", cLoops, iCpu, off));
4768	}
4769
4770	/* Next */
4771	PrevTsc = CurTsc;
4772	}
4773	else if (rc == VERR_NOT_SUPPORTED)
4774	break;
4775	else
4776	AssertMsg(rc == VERR_CPU_NOT_FOUND \|\| rc == VERR_CPU_OFFLINE, ("%d\n", rc));
4777	}
4778
4779	/* broke out of the loop. */
4780	if (iCpu <= iLastCpu)
4781	break;
4782	}
4783
4784	poffMin = offMin; / Almost RTMpOnSpecific profiling. */
4785	dprintf(("supdrvDetermineAsyncTsc: returns %d; iLastCpu=%d rc=%d offMin=%llx offMax=%llx\n",
4786	fAsync, iLastCpu, rc, offMin, offMax));
4787	#if !defined(RT_OS_SOLARIS) && !defined(RT_OS_OS2) && !defined(RT_OS_WINDOWS)
4788	OSDBGPRINT(("vboxdrv: fAsync=%d offMin=%#lx offMax=%#lx\n", fAsync, (long)offMin, (long)offMax));
4789	#endif
4790	return fAsync;
4791	}
4792
4793
4794	/**
4795	* Determin the GIP TSC mode.
4796	*
4797	* @returns The most suitable TSC mode.
4798	* @param pDevExt Pointer to the device instance data.
4799	*/
4800	static SUPGIPMODE supdrvGipDeterminTscMode(PSUPDRVDEVEXT pDevExt)
4801	{
4802	/*
4803	* On SMP we're faced with two problems:
4804	* (1) There might be a skew between the CPU, so that cpu0
4805	* returns a TSC that is sligtly different from cpu1.
4806	* (2) Power management (and other things) may cause the TSC
4807	* to run at a non-constant speed, and cause the speed
4808	* to be different on the cpus. This will result in (1).
4809	*
4810	* So, on SMP systems we'll have to select the ASYNC update method
4811	* if there are symphoms of these problems.
4812	*/
4813	if (RTMpGetCount() > 1)
4814	{
4815	uint32_t uEAX, uEBX, uECX, uEDX;
4816	uint64_t u64DiffCoresIgnored;
4817
4818	/* Permit the user and/or the OS specfic bits to force async mode. */
4819	if (supdrvOSGetForcedAsyncTscMode(pDevExt))
4820	return SUPGIPMODE_ASYNC_TSC;
4821
4822	/* Try check for current differences between the cpus. */
4823	if (supdrvDetermineAsyncTsc(&u64DiffCoresIgnored))
4824	return SUPGIPMODE_ASYNC_TSC;
4825
4826	/*
4827	* If the CPU supports power management and is an AMD one we
4828	* won't trust it unless it has the TscInvariant bit is set.
4829	*/
4830	/* Check for "AuthenticAMD" */
4831	ASMCpuId(0, &uEAX, &uEBX, &uECX, &uEDX);
4832	if ( uEAX >= 1
4833	&& uEBX == X86_CPUID_VENDOR_AMD_EBX
4834	&& uECX == X86_CPUID_VENDOR_AMD_ECX
4835	&& uEDX == X86_CPUID_VENDOR_AMD_EDX)
4836	{
4837	/* Check for APM support and that TscInvariant is cleared. */
4838	ASMCpuId(0x80000000, &uEAX, &uEBX, &uECX, &uEDX);
4839	if (uEAX >= 0x80000007)
4840	{
4841	ASMCpuId(0x80000007, &uEAX, &uEBX, &uECX, &uEDX);
4842	if ( !(uEDX & RT_BIT(8))/* TscInvariant */
4843	&& (uEDX & 0x3e)) /* STC\|TM\|THERMTRIP\|VID\|FID. Ignore TS. */
4844	return SUPGIPMODE_ASYNC_TSC;
4845	}
4846	}
4847	}
4848	return SUPGIPMODE_SYNC_TSC;
4849	}
4850
4851
4852	/**
4853	* Invalidates the GIP data upon termination.
4854	*
4855	* @param pGip Pointer to the read-write kernel mapping of the GIP.
4856	*/
4857	void VBOXCALL supdrvGipTerm(PSUPGLOBALINFOPAGE pGip)
4858	{
4859	unsigned i;
4860	pGip->u32Magic = 0;
4861	for (i = 0; i < RT_ELEMENTS(pGip->aCPUs); i++)
4862	{
4863	pGip->aCPUs[i].u64NanoTS = 0;
4864	pGip->aCPUs[i].u64TSC = 0;
4865	pGip->aCPUs[i].iTSCHistoryHead = 0;
4866	}
4867	}
4868
4869
4870	/**
4871	* Worker routine for supdrvGipUpdate and supdrvGipUpdatePerCpu that
4872	* updates all the per cpu data except the transaction id.
4873	*
4874	* @param pGip The GIP.
4875	* @param pGipCpu Pointer to the per cpu data.
4876	* @param u64NanoTS The current time stamp.
4877	*/
4878	static void supdrvGipDoUpdateCpu(PSUPGLOBALINFOPAGE pGip, PSUPGIPCPU pGipCpu, uint64_t u64NanoTS)
4879	{
4880	uint64_t u64TSC;
4881	uint64_t u64TSCDelta;
4882	uint32_t u32UpdateIntervalTSC;
4883	uint32_t u32UpdateIntervalTSCSlack;
4884	unsigned iTSCHistoryHead;
4885	uint64_t u64CpuHz;
4886
4887	/*
4888	* Update the NanoTS.
4889	*/
4890	ASMAtomicXchgU64(&pGipCpu->u64NanoTS, u64NanoTS);
4891
4892	/*
4893	* Calc TSC delta.
4894	*/
4895	/** @todo validate the NanoTS delta, don't trust the OS to call us when it should... */
4896	u64TSC = ASMReadTSC();
4897	u64TSCDelta = u64TSC - pGipCpu->u64TSC;
4898	ASMAtomicXchgU64(&pGipCpu->u64TSC, u64TSC);
4899
4900	if (u64TSCDelta >> 32)
4901	{
4902	u64TSCDelta = pGipCpu->u32UpdateIntervalTSC;
4903	pGipCpu->cErrors++;
4904	}
4905
4906	/*
4907	* TSC History.
4908	*/
4909	Assert(ELEMENTS(pGipCpu->au32TSCHistory) == 8);
4910
4911	iTSCHistoryHead = (pGipCpu->iTSCHistoryHead + 1) & 7;
4912	ASMAtomicXchgU32(&pGipCpu->iTSCHistoryHead, iTSCHistoryHead);
4913	ASMAtomicXchgU32(&pGipCpu->au32TSCHistory[iTSCHistoryHead], (uint32_t)u64TSCDelta);
4914
4915	/*
4916	* UpdateIntervalTSC = average of last 8,2,1 intervals depending on update HZ.
4917	*/
4918	if (pGip->u32UpdateHz >= 1000)
4919	{
4920	uint32_t u32;
4921	u32 = pGipCpu->au32TSCHistory[0];
4922	u32 += pGipCpu->au32TSCHistory[1];
4923	u32 += pGipCpu->au32TSCHistory[2];
4924	u32 += pGipCpu->au32TSCHistory[3];
4925	u32 >>= 2;
4926	u32UpdateIntervalTSC = pGipCpu->au32TSCHistory[4];
4927	u32UpdateIntervalTSC += pGipCpu->au32TSCHistory[5];
4928	u32UpdateIntervalTSC += pGipCpu->au32TSCHistory[6];
4929	u32UpdateIntervalTSC += pGipCpu->au32TSCHistory[7];
4930	u32UpdateIntervalTSC >>= 2;
4931	u32UpdateIntervalTSC += u32;
4932	u32UpdateIntervalTSC >>= 1;
4933
4934	/* Value choosen for a 2GHz Athlon64 running linux 2.6.10/11, . */
4935	u32UpdateIntervalTSCSlack = u32UpdateIntervalTSC >> 14;
4936	}
4937	else if (pGip->u32UpdateHz >= 90)
4938	{
4939	u32UpdateIntervalTSC = (uint32_t)u64TSCDelta;
4940	u32UpdateIntervalTSC += pGipCpu->au32TSCHistory[(iTSCHistoryHead - 1) & 7];
4941	u32UpdateIntervalTSC >>= 1;
4942
4943	/* value choosen on a 2GHz thinkpad running windows */
4944	u32UpdateIntervalTSCSlack = u32UpdateIntervalTSC >> 7;
4945	}
4946	else
4947	{
4948	u32UpdateIntervalTSC = (uint32_t)u64TSCDelta;
4949
4950	/* This value hasn't be checked yet.. waiting for OS/2 and 33Hz timers.. :-) */
4951	u32UpdateIntervalTSCSlack = u32UpdateIntervalTSC >> 6;
4952	}
4953	ASMAtomicXchgU32(&pGipCpu->u32UpdateIntervalTSC, u32UpdateIntervalTSC + u32UpdateIntervalTSCSlack);
4954
4955	/*
4956	* CpuHz.
4957	*/
4958	u64CpuHz = ASMMult2xU32RetU64(u32UpdateIntervalTSC, pGip->u32UpdateHz);
4959	ASMAtomicXchgU64(&pGipCpu->u64CpuHz, u64CpuHz);
4960	}
4961
4962
4963	/**
4964	* Updates the GIP.
4965	*
4966	* @param pGip Pointer to the GIP.
4967	* @param u64NanoTS The current nanosecond timesamp.
4968	*/
4969	void VBOXCALL supdrvGipUpdate(PSUPGLOBALINFOPAGE pGip, uint64_t u64NanoTS)
4970	{
4971	/*
4972	* Determin the relevant CPU data.
4973	*/
4974	PSUPGIPCPU pGipCpu;
4975	if (pGip->u32Mode != SUPGIPMODE_ASYNC_TSC)
4976	pGipCpu = &pGip->aCPUs[0];
4977	else
4978	{
4979	unsigned iCpu = ASMGetApicId();
4980	if (RT_LIKELY(iCpu >= RT_ELEMENTS(pGip->aCPUs)))
4981	return;
4982	pGipCpu = &pGip->aCPUs[iCpu];
4983	}
4984
4985	/*
4986	* Start update transaction.
4987	*/
4988	if (!(ASMAtomicIncU32(&pGipCpu->u32TransactionId) & 1))
4989	{
4990	/* this can happen on win32 if we're taking to long and there are more CPUs around. shouldn't happen though. */
4991	AssertMsgFailed(("Invalid transaction id, %#x, not odd!\n", pGipCpu->u32TransactionId));
4992	ASMAtomicIncU32(&pGipCpu->u32TransactionId);
4993	pGipCpu->cErrors++;
4994	return;
4995	}
4996
4997	/*
4998	* Recalc the update frequency every 0x800th time.
4999	*/
5000	if (!(pGipCpu->u32TransactionId & (GIP_UPDATEHZ_RECALC_FREQ * 2 - 2)))
5001	{
5002	if (pGip->u64NanoTSLastUpdateHz)
5003	{
5004	#ifdef RT_ARCH_AMD64 /** @todo fix 64-bit div here to work on x86 linux. */
5005	uint64_t u64Delta = u64NanoTS - pGip->u64NanoTSLastUpdateHz;
5006	uint32_t u32UpdateHz = (uint32_t)((UINT64_C(1000000000) * GIP_UPDATEHZ_RECALC_FREQ) / u64Delta);
5007	if (u32UpdateHz <= 2000 && u32UpdateHz >= 30)
5008	{
5009	ASMAtomicXchgU32(&pGip->u32UpdateHz, u32UpdateHz);
5010	ASMAtomicXchgU32(&pGip->u32UpdateIntervalNS, 1000000000 / u32UpdateHz);
5011	}
5012	#endif
5013	}
5014	ASMAtomicXchgU64(&pGip->u64NanoTSLastUpdateHz, u64NanoTS);
5015	}
5016
5017	/*
5018	* Update the data.
5019	*/
5020	supdrvGipDoUpdateCpu(pGip, pGipCpu, u64NanoTS);
5021
5022	/*
5023	* Complete transaction.
5024	*/
5025	ASMAtomicIncU32(&pGipCpu->u32TransactionId);
5026	}
5027
5028
5029	/**
5030	* Updates the per cpu GIP data for the calling cpu.
5031	*
5032	* @param pGip Pointer to the GIP.
5033	* @param u64NanoTS The current nanosecond timesamp.
5034	* @param iCpu The CPU index.
5035	*/
5036	void VBOXCALL supdrvGipUpdatePerCpu(PSUPGLOBALINFOPAGE pGip, uint64_t u64NanoTS, unsigned iCpu)
5037	{
5038	PSUPGIPCPU pGipCpu;
5039
5040	if (RT_LIKELY(iCpu < RT_ELEMENTS(pGip->aCPUs)))
5041	{
5042	pGipCpu = &pGip->aCPUs[iCpu];
5043
5044	/*
5045	* Start update transaction.
5046	*/
5047	if (!(ASMAtomicIncU32(&pGipCpu->u32TransactionId) & 1))
5048	{
5049	AssertMsgFailed(("Invalid transaction id, %#x, not odd!\n", pGipCpu->u32TransactionId));
5050	ASMAtomicIncU32(&pGipCpu->u32TransactionId);
5051	pGipCpu->cErrors++;
5052	return;
5053	}
5054
5055	/*
5056	* Update the data.
5057	*/
5058	supdrvGipDoUpdateCpu(pGip, pGipCpu, u64NanoTS);
5059
5060	/*
5061	* Complete transaction.
5062	*/
5063	ASMAtomicIncU32(&pGipCpu->u32TransactionId);
5064	}
5065	}
5066
5067
5068	#ifndef SUPDRV_WITH_RELEASE_LOGGER
5069	# ifndef DEBUG /** @todo change #ifndef DEBUG -> #ifdef LOG_ENABLED */
5070	/**
5071	* Stub function for non-debug builds.
5072	*/
5073	RTDECL(PRTLOGGER) RTLogDefaultInstance(void)
5074	{
5075	return NULL;
5076	}
5077
5078	RTDECL(PRTLOGGER) RTLogRelDefaultInstance(void)
5079	{
5080	return NULL;
5081	}
5082
5083	/**
5084	* Stub function for non-debug builds.
5085	*/
5086	RTDECL(int) RTLogSetDefaultInstanceThread(PRTLOGGER pLogger, uintptr_t uKey)
5087	{
5088	return 0;
5089	}
5090
5091	/**
5092	* Stub function for non-debug builds.
5093	*/
5094	RTDECL(void) RTLogLogger(PRTLOGGER pLogger, void pvCallerRet, const char pszFormat, ...)
5095	{
5096	}
5097
5098	/**
5099	* Stub function for non-debug builds.
5100	*/
5101	RTDECL(void) RTLogLoggerEx(PRTLOGGER pLogger, unsigned fFlags, unsigned iGroup, const char *pszFormat, ...)
5102	{
5103	}
5104
5105	/**
5106	* Stub function for non-debug builds.
5107	*/
5108	RTDECL(void) RTLogLoggerExV(PRTLOGGER pLogger, unsigned fFlags, unsigned iGroup, const char *pszFormat, va_list args)
5109	{
5110	}
5111
5112	/**
5113	* Stub function for non-debug builds.
5114	*/
5115	RTDECL(void) RTLogPrintf(const char *pszFormat, ...)
5116	{
5117	}
5118
5119	/**
5120	* Stub function for non-debug builds.
5121	*/
5122	RTDECL(void) RTLogPrintfV(const char *pszFormat, va_list args)
5123	{
5124	}
5125	# endif /* !DEBUG */
5126	#endif /* !SUPDRV_WITH_RELEASE_LOGGER */

注意: 瀏覽 TracBrowser 來幫助您使用儲存庫瀏覽器

source: vbox/trunk/src/VBox/HostDrivers/Support/SUPDrv.c@ 12100

以其他格式下載: