SUPDrv.cpp@ 93254

Last change on this file since 93254 was 93206, checked in by vboxsync, 3 years ago
SUPDrv: Made SUPR0PageAllocEx more palatable for HVCI. bugref:10162
Property svn:eol-style set to `native` Property svn:keywords set to `Author Date Id Revision`
File size: 265.6 KB

Line
1	/* $Id: SUPDrv.cpp 93206 2022-01-12 19:12:13Z vboxsync $ */
2	/** @file
3	* VBoxDrv - The VirtualBox Support Driver - Common code.
4	*/
5
6	/*
7	* Copyright (C) 2006-2022 Oracle Corporation
8	*
9	* This file is part of VirtualBox Open Source Edition (OSE), as
10	* available from http://www.virtualbox.org. This file is free software;
11	* you can redistribute it and/or modify it under the terms of the GNU
12	* General Public License (GPL) as published by the Free Software
13	* Foundation, in version 2 as it comes in the "COPYING" file of the
14	* VirtualBox OSE distribution. VirtualBox OSE is distributed in the
15	* hope that it will be useful, but WITHOUT ANY WARRANTY of any kind.
16	*
17	* The contents of this file may alternatively be used under the terms
18	* of the Common Development and Distribution License Version 1.0
19	* (CDDL) only, as it comes in the "COPYING.CDDL" file of the
20	* VirtualBox OSE distribution, in which case the provisions of the
21	* CDDL are applicable instead of those of the GPL.
22	*
23	* You may elect to license modified versions of this file under the
24	* terms and conditions of either the GPL or the CDDL or both.
25	*/
26
27
28	/*********************************************************************************************************************************
29	* Header Files *
30	*********************************************************************************************************************************/
31	#define LOG_GROUP LOG_GROUP_SUP_DRV
32	#define SUPDRV_AGNOSTIC
33	#include "SUPDrvInternal.h"
34	#ifndef PAGE_SHIFT
35	# include <iprt/param.h>
36	#endif
37	#include <iprt/asm.h>
38	#include <iprt/asm-amd64-x86.h>
39	#include <iprt/asm-math.h>
40	#include <iprt/cpuset.h>
41	#if defined(RT_OS_DARWIN) \|\| defined(RT_OS_SOLARIS) \|\| defined(RT_OS_WINDOWS)
42	# include <iprt/dbg.h>
43	#endif
44	#include <iprt/handletable.h>
45	#include <iprt/mem.h>
46	#include <iprt/mp.h>
47	#include <iprt/power.h>
48	#include <iprt/process.h>
49	#include <iprt/semaphore.h>
50	#include <iprt/spinlock.h>
51	#include <iprt/thread.h>
52	#include <iprt/uuid.h>
53	#include <iprt/net.h>
54	#include <iprt/crc.h>
55	#include <iprt/string.h>
56	#include <iprt/timer.h>
57	#if defined(RT_OS_DARWIN) \|\| defined(RT_OS_SOLARIS) \|\| defined(RT_OS_FREEBSD)
58	# include <iprt/rand.h>
59	# include <iprt/path.h>
60	#endif
61	#include <iprt/uint128.h>
62	#include <iprt/x86.h>
63
64	#include <VBox/param.h>
65	#include <VBox/log.h>
66	#include <VBox/err.h>
67	#include <VBox/vmm/hm_vmx.h>
68
69	#if defined(RT_OS_SOLARIS) \|\| defined(RT_OS_DARWIN)
70	# include "dtrace/SUPDrv.h"
71	#else
72	# define VBOXDRV_SESSION_CREATE(pvSession, fUser) do { } while (0)
73	# define VBOXDRV_SESSION_CLOSE(pvSession) do { } while (0)
74	# define VBOXDRV_IOCTL_ENTRY(pvSession, uIOCtl, pvReqHdr) do { } while (0)
75	# define VBOXDRV_IOCTL_RETURN(pvSession, uIOCtl, pvReqHdr, rcRet, rcReq) do { } while (0)
76	#endif
77
78	#ifdef __cplusplus
79	# if __cplusplus >= 201100 \|\| RT_MSC_PREREQ(RT_MSC_VER_VS2019)
80	# define SUPDRV_CAN_COUNT_FUNCTION_ARGS
81	# ifdef _MSC_VER
82	# pragma warning(push)
83	# pragma warning(disable:4577)
84	# include <type_traits>
85	# pragma warning(pop)
86
87	# elif defined(RT_OS_DARWIN)
88	# define _LIBCPP_CSTDDEF
89	# include <__nullptr>
90	# include <type_traits>
91
92	# else
93	# include <type_traits>
94	# endif
95	# endif
96	#endif
97
98
99	/*
100	* Logging assignments:
101	* Log - useful stuff, like failures.
102	* LogFlow - program flow, except the really noisy bits.
103	* Log2 - Cleanup.
104	* Log3 - Loader flow noise.
105	* Log4 - Call VMMR0 flow noise.
106	* Log5 - Native yet-to-be-defined noise.
107	* Log6 - Native ioctl flow noise.
108	*
109	* Logging requires KBUILD_TYPE=debug and possibly changes to the logger
110	* instantiation in log-vbox.c(pp).
111	*/
112
113
114	/*********************************************************************************************************************************
115	* Defined Constants And Macros *
116	*********************************************************************************************************************************/
117	/** @def VBOX_SVN_REV
118	* The makefile should define this if it can. */
119	#ifndef VBOX_SVN_REV
120	# define VBOX_SVN_REV 0
121	#endif
122
123	/** @ SUPDRV_CHECK_SMAP_SETUP
124	* SMAP check setup. */
125	/** @def SUPDRV_CHECK_SMAP_CHECK
126	* Checks that the AC flag is set if SMAP is enabled. If AC is not set, it
127	* will be logged and @a a_BadExpr is executed. */
128	#if (defined(RT_OS_DARWIN) \|\| defined(RT_OS_LINUX)) && !defined(VBOX_WITHOUT_EFLAGS_AC_SET_IN_VBOXDRV)
129	# define SUPDRV_CHECK_SMAP_SETUP() uint32_t const fKernelFeatures = SUPR0GetKernelFeatures()
130	# define SUPDRV_CHECK_SMAP_CHECK(a_pDevExt, a_BadExpr) \
131	do { \
132	if (fKernelFeatures & SUPKERNELFEATURES_SMAP) \
133	{ \
134	RTCCUINTREG fEfl = ASMGetFlags(); \
135	if (RT_LIKELY(fEfl & X86_EFL_AC)) \
136	{ /* likely */ } \
137	else \
138	{ \
139	supdrvBadContext(a_pDevExt, "SUPDrv.cpp", __LINE__, "EFLAGS.AC is 0!"); \
140	a_BadExpr; \
141	} \
142	} \
143	} while (0)
144	#else
145	# define SUPDRV_CHECK_SMAP_SETUP() uint32_t const fKernelFeatures = 0
146	# define SUPDRV_CHECK_SMAP_CHECK(a_pDevExt, a_BadExpr) NOREF(fKernelFeatures)
147	#endif
148
149
150	/*********************************************************************************************************************************
151	* Internal Functions *
152	*********************************************************************************************************************************/
153	static DECLCALLBACK(int) supdrvSessionObjHandleRetain(RTHANDLETABLE hHandleTable, void pvObj, void pvCtx, void *pvUser);
154	static DECLCALLBACK(void) supdrvSessionObjHandleDelete(RTHANDLETABLE hHandleTable, uint32_t h, void pvObj, void pvCtx, void *pvUser);
155	static int supdrvMemAdd(PSUPDRVMEMREF pMem, PSUPDRVSESSION pSession);
156	static int supdrvMemRelease(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr, SUPDRVMEMREFTYPE eType);
157	static int supdrvIOCtl_LdrOpen(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDROPEN pReq);
158	static int supdrvIOCtl_LdrLoad(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRLOAD pReq);
159	static int supdrvIOCtl_LdrFree(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRFREE pReq);
160	static int supdrvIOCtl_LdrLockDown(PSUPDRVDEVEXT pDevExt);
161	static int supdrvIOCtl_LdrQuerySymbol(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRGETSYMBOL pReq);
162	static int supdrvIDC_LdrGetSymbol(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPDRVIDCREQGETSYM pReq);
163	static int supdrvLdrAddUsage(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPDRVLDRIMAGE pImage, bool fRing3Usage);
164	DECLINLINE(void) supdrvLdrSubtractUsage(PSUPDRVDEVEXT pDevExt, PSUPDRVLDRIMAGE pImage, uint32_t cReference);
165	static void supdrvLdrFree(PSUPDRVDEVEXT pDevExt, PSUPDRVLDRIMAGE pImage);
166	DECLINLINE(int) supdrvLdrLock(PSUPDRVDEVEXT pDevExt);
167	DECLINLINE(int) supdrvLdrUnlock(PSUPDRVDEVEXT pDevExt);
168	static int supdrvIOCtl_CallServiceModule(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPCALLSERVICE pReq);
169	static int supdrvIOCtl_LoggerSettings(PSUPLOGGERSETTINGS pReq);
170	static int supdrvIOCtl_MsrProber(PSUPDRVDEVEXT pDevExt, PSUPMSRPROBER pReq);
171	static int supdrvIOCtl_ResumeSuspendedKbds(void);
172
173
174	/*********************************************************************************************************************************
175	* Global Variables *
176	*********************************************************************************************************************************/
177	/** @def SUPEXP_CHECK_ARGS
178	* This is for checking the argument count of the function in the entry,
179	* just to make sure we don't accidentally export something the wrapper
180	* can't deal with.
181	*
182	* Using some C++11 magic to do the counting.
183	*
184	* The error is reported by overflowing the SUPFUNC::cArgs field, so the
185	* warnings can probably be a little mysterious.
186	*
187	* @note Doesn't work for CLANG 11. Works for Visual C++, unless there
188	* are function pointers in the argument list.
189	*/
190	#if defined(SUPDRV_CAN_COUNT_FUNCTION_ARGS) && RT_CLANG_PREREQ(99, 0)
191	template <typename RetType, typename ... Types>
192	constexpr std::integral_constant<unsigned, sizeof ...(Types)>
193	CountFunctionArguments(RetType(RTCALL *)(Types ...))
194	{
195	return std::integral_constant<unsigned, sizeof ...(Types)>{};
196	}
197	# define SUPEXP_CHECK_ARGS(a_cArgs, a_Name) \
198	((a_cArgs) >= decltype(CountFunctionArguments(a_Name))::value ? (uint8_t)(a_cArgs) : 1023)
199
200	#else
201	# define SUPEXP_CHECK_ARGS(a_cArgs, a_Name) a_cArgs
202	#endif
203
204	/** @name Function table entry macros.
205	* @note The SUPEXP_STK_BACKF macro is because VC++ has trouble with functions
206	* with function pointer arguments (probably noexcept related).
207	* @{ */
208	#define SUPEXP_CUSTOM(a_cArgs, a_Name, a_Value) { #a_Name, a_cArgs, (void *)(uintptr_t)(a_Value) }
209	#define SUPEXP_STK_OKAY(a_cArgs, a_Name) { #a_Name, SUPEXP_CHECK_ARGS(a_cArgs, a_Name), (void *)(uintptr_t)a_Name }
210	#if 0
211	# define SUPEXP_STK_BACK(a_cArgs, a_Name) { "StkBack_" #a_Name, SUPEXP_CHECK_ARGS(a_cArgs, a_Name), (void *)(uintptr_t)a_Name }
212	# define SUPEXP_STK_BACKF(a_cArgs, a_Name) { "StkBack_" #a_Name, SUPEXP_CHECK_ARGS(a_cArgs, a_Name), (void *)(uintptr_t)a_Name }
213	#else
214	# define SUPEXP_STK_BACK(a_cArgs, a_Name) { #a_Name, SUPEXP_CHECK_ARGS(a_cArgs, a_Name), (void *)(uintptr_t)a_Name }
215	# ifdef _MSC_VER
216	# define SUPEXP_STK_BACKF(a_cArgs, a_Name) { #a_Name, a_cArgs, (void *)(uintptr_t)a_Name }
217	# else
218	# define SUPEXP_STK_BACKF(a_cArgs, a_Name) { #a_Name, SUPEXP_CHECK_ARGS(a_cArgs, a_Name), (void *)(uintptr_t)a_Name }
219	# endif
220	#endif
221	/** @} */
222
223	/**
224	* Array of the R0 SUP API.
225	*
226	* While making changes to these exports, make sure to update the IOC
227	* minor version (SUPDRV_IOC_VERSION).
228	*
229	* @remarks This array is processed by SUPR0-def-pe.sed and SUPR0-def-lx.sed to
230	* produce definition files from which import libraries are generated.
231	* Take care when commenting things and especially with \#ifdef'ing.
232	*/
233	static SUPFUNC g_aFunctions[] =
234	{
235	/* SED: START */
236	/* name function */
237	/* Entries with absolute addresses determined at runtime, fixup
238	code makes ugly ASSUMPTIONS about the order here: */
239	SUPEXP_CUSTOM( 0, SUPR0AbsIs64bit, 0),
240	SUPEXP_CUSTOM( 0, SUPR0Abs64bitKernelCS, 0),
241	SUPEXP_CUSTOM( 0, SUPR0Abs64bitKernelSS, 0),
242	SUPEXP_CUSTOM( 0, SUPR0Abs64bitKernelDS, 0),
243	SUPEXP_CUSTOM( 0, SUPR0AbsKernelCS, 0),
244	SUPEXP_CUSTOM( 0, SUPR0AbsKernelSS, 0),
245	SUPEXP_CUSTOM( 0, SUPR0AbsKernelDS, 0),
246	SUPEXP_CUSTOM( 0, SUPR0AbsKernelES, 0),
247	SUPEXP_CUSTOM( 0, SUPR0AbsKernelFS, 0),
248	SUPEXP_CUSTOM( 0, SUPR0AbsKernelGS, 0),
249	/* Normal function & data pointers: */
250	SUPEXP_CUSTOM( 0, g_pSUPGlobalInfoPage, &g_pSUPGlobalInfoPage), /* SED: DATA */
251	SUPEXP_STK_OKAY( 0, SUPGetGIP),
252	SUPEXP_STK_BACK( 1, SUPReadTscWithDelta),
253	SUPEXP_STK_BACK( 1, SUPGetTscDeltaSlow),
254	SUPEXP_STK_BACK( 1, SUPGetCpuHzFromGipForAsyncMode),
255	SUPEXP_STK_OKAY( 3, SUPIsTscFreqCompatible),
256	SUPEXP_STK_OKAY( 3, SUPIsTscFreqCompatibleEx),
257	SUPEXP_STK_BACK( 4, SUPR0BadContext),
258	SUPEXP_STK_BACK( 2, SUPR0ComponentDeregisterFactory),
259	SUPEXP_STK_BACK( 4, SUPR0ComponentQueryFactory),
260	SUPEXP_STK_BACK( 2, SUPR0ComponentRegisterFactory),
261	SUPEXP_STK_BACK( 5, SUPR0ContAlloc),
262	SUPEXP_STK_BACK( 2, SUPR0ContFree),
263	SUPEXP_STK_BACK( 2, SUPR0ChangeCR4),
264	SUPEXP_STK_BACK( 1, SUPR0EnableVTx),
265	SUPEXP_STK_BACK( 0, SUPR0SuspendVTxOnCpu),
266	SUPEXP_STK_BACK( 1, SUPR0ResumeVTxOnCpu),
267	SUPEXP_STK_OKAY( 1, SUPR0GetCurrentGdtRw),
268	SUPEXP_STK_OKAY( 0, SUPR0GetKernelFeatures),
269	SUPEXP_STK_BACK( 3, SUPR0GetHwvirtMsrs),
270	SUPEXP_STK_BACK( 0, SUPR0GetPagingMode),
271	SUPEXP_STK_BACK( 1, SUPR0GetSvmUsability),
272	SUPEXP_STK_BACK( 1, SUPR0GetVTSupport),
273	SUPEXP_STK_BACK( 1, SUPR0GetVmxUsability),
274	SUPEXP_STK_BACK( 2, SUPR0LdrIsLockOwnerByMod),
275	SUPEXP_STK_BACK( 1, SUPR0LdrLock),
276	SUPEXP_STK_BACK( 1, SUPR0LdrUnlock),
277	SUPEXP_STK_BACK( 3, SUPR0LdrModByName),
278	SUPEXP_STK_BACK( 2, SUPR0LdrModRelease),
279	SUPEXP_STK_BACK( 2, SUPR0LdrModRetain),
280	SUPEXP_STK_BACK( 4, SUPR0LockMem),
281	SUPEXP_STK_BACK( 5, SUPR0LowAlloc),
282	SUPEXP_STK_BACK( 2, SUPR0LowFree),
283	SUPEXP_STK_BACK( 4, SUPR0MemAlloc),
284	SUPEXP_STK_BACK( 2, SUPR0MemFree),
285	SUPEXP_STK_BACK( 3, SUPR0MemGetPhys),
286	SUPEXP_STK_BACK( 2, SUPR0ObjAddRef),
287	SUPEXP_STK_BACK( 3, SUPR0ObjAddRefEx),
288	SUPEXP_STK_BACKF( 5, SUPR0ObjRegister),
289	SUPEXP_STK_BACK( 2, SUPR0ObjRelease),
290	SUPEXP_STK_BACK( 3, SUPR0ObjVerifyAccess),
291	SUPEXP_STK_BACK( 6, SUPR0PageAllocEx),
292	SUPEXP_STK_BACK( 2, SUPR0PageFree),
293	SUPEXP_STK_BACK( 6, SUPR0PageMapKernel),
294	SUPEXP_STK_BACK( 6, SUPR0PageProtect),
295	#if defined(RT_OS_LINUX) \|\| defined(RT_OS_SOLARIS) \|\| defined(RT_OS_FREEBSD)
296	SUPEXP_STK_OKAY( 2, SUPR0HCPhysToVirt), /* only-linux, only-solaris, only-freebsd */
297	#endif
298	SUPEXP_STK_BACK( 2, SUPR0PrintfV),
299	SUPEXP_STK_BACK( 1, SUPR0GetSessionGVM),
300	SUPEXP_STK_BACK( 1, SUPR0GetSessionVM),
301	SUPEXP_STK_BACK( 3, SUPR0SetSessionVM),
302	SUPEXP_STK_BACK( 1, SUPR0GetSessionUid),
303	SUPEXP_STK_BACK( 6, SUPR0TscDeltaMeasureBySetIndex),
304	SUPEXP_STK_BACK( 1, SUPR0TracerDeregisterDrv),
305	SUPEXP_STK_BACK( 2, SUPR0TracerDeregisterImpl),
306	SUPEXP_STK_BACK( 6, SUPR0TracerFireProbe),
307	SUPEXP_STK_BACK( 3, SUPR0TracerRegisterDrv),
308	SUPEXP_STK_BACK( 4, SUPR0TracerRegisterImpl),
309	SUPEXP_STK_BACK( 2, SUPR0TracerRegisterModule),
310	SUPEXP_STK_BACK( 2, SUPR0TracerUmodProbeFire),
311	SUPEXP_STK_BACK( 2, SUPR0UnlockMem),
312	#ifdef RT_OS_WINDOWS
313	SUPEXP_STK_BACK( 4, SUPR0IoCtlSetupForHandle), /* only-windows */
314	SUPEXP_STK_BACK( 9, SUPR0IoCtlPerform), /* only-windows */
315	SUPEXP_STK_BACK( 1, SUPR0IoCtlCleanup), /* only-windows */
316	#endif
317	SUPEXP_STK_BACK( 2, SUPSemEventClose),
318	SUPEXP_STK_BACK( 2, SUPSemEventCreate),
319	SUPEXP_STK_BACK( 1, SUPSemEventGetResolution),
320	SUPEXP_STK_BACK( 2, SUPSemEventMultiClose),
321	SUPEXP_STK_BACK( 2, SUPSemEventMultiCreate),
322	SUPEXP_STK_BACK( 1, SUPSemEventMultiGetResolution),
323	SUPEXP_STK_BACK( 2, SUPSemEventMultiReset),
324	SUPEXP_STK_BACK( 2, SUPSemEventMultiSignal),
325	SUPEXP_STK_BACK( 3, SUPSemEventMultiWait),
326	SUPEXP_STK_BACK( 3, SUPSemEventMultiWaitNoResume),
327	SUPEXP_STK_BACK( 3, SUPSemEventMultiWaitNsAbsIntr),
328	SUPEXP_STK_BACK( 3, SUPSemEventMultiWaitNsRelIntr),
329	SUPEXP_STK_BACK( 2, SUPSemEventSignal),
330	SUPEXP_STK_BACK( 3, SUPSemEventWait),
331	SUPEXP_STK_BACK( 3, SUPSemEventWaitNoResume),
332	SUPEXP_STK_BACK( 3, SUPSemEventWaitNsAbsIntr),
333	SUPEXP_STK_BACK( 3, SUPSemEventWaitNsRelIntr),
334
335	SUPEXP_STK_BACK( 0, RTAssertAreQuiet),
336	SUPEXP_STK_BACK( 0, RTAssertMayPanic),
337	SUPEXP_STK_BACK( 4, RTAssertMsg1),
338	SUPEXP_STK_BACK( 2, RTAssertMsg2AddV),
339	SUPEXP_STK_BACK( 2, RTAssertMsg2V),
340	SUPEXP_STK_BACK( 1, RTAssertSetMayPanic),
341	SUPEXP_STK_BACK( 1, RTAssertSetQuiet),
342	SUPEXP_STK_OKAY( 2, RTCrc32),
343	SUPEXP_STK_OKAY( 1, RTCrc32Finish),
344	SUPEXP_STK_OKAY( 3, RTCrc32Process),
345	SUPEXP_STK_OKAY( 0, RTCrc32Start),
346	SUPEXP_STK_OKAY( 1, RTErrConvertFromErrno),
347	SUPEXP_STK_OKAY( 1, RTErrConvertToErrno),
348	SUPEXP_STK_BACK( 4, RTHandleTableAllocWithCtx),
349	SUPEXP_STK_BACK( 1, RTHandleTableCreate),
350	SUPEXP_STK_BACKF( 6, RTHandleTableCreateEx),
351	SUPEXP_STK_BACKF( 3, RTHandleTableDestroy),
352	SUPEXP_STK_BACK( 3, RTHandleTableFreeWithCtx),
353	SUPEXP_STK_BACK( 3, RTHandleTableLookupWithCtx),
354	SUPEXP_STK_BACK( 5, RTLogBulkUpdate),
355	SUPEXP_STK_BACK( 2, RTLogCheckGroupFlags),
356	SUPEXP_STK_BACKF( 17, RTLogCreateExV),
357	SUPEXP_STK_BACK( 1, RTLogDestroy),
358	SUPEXP_STK_BACK( 0, RTLogDefaultInstance),
359	SUPEXP_STK_BACK( 1, RTLogDefaultInstanceEx),
360	SUPEXP_STK_BACK( 1, SUPR0DefaultLogInstanceEx),
361	SUPEXP_STK_BACK( 0, RTLogGetDefaultInstance),
362	SUPEXP_STK_BACK( 1, RTLogGetDefaultInstanceEx),
363	SUPEXP_STK_BACK( 1, SUPR0GetDefaultLogInstanceEx),
364	SUPEXP_STK_BACK( 5, RTLogLoggerExV),
365	SUPEXP_STK_BACK( 2, RTLogPrintfV),
366	SUPEXP_STK_BACK( 0, RTLogRelGetDefaultInstance),
367	SUPEXP_STK_BACK( 1, RTLogRelGetDefaultInstanceEx),
368	SUPEXP_STK_BACK( 1, SUPR0GetDefaultLogRelInstanceEx),
369	SUPEXP_STK_BACK( 2, RTLogSetDefaultInstanceThread),
370	SUPEXP_STK_BACKF( 2, RTLogSetFlushCallback),
371	SUPEXP_STK_BACK( 2, RTLogSetR0ProgramStart),
372	SUPEXP_STK_BACK( 3, RTLogSetR0ThreadNameV),
373	SUPEXP_STK_BACK( 5, RTMemAllocExTag),
374	SUPEXP_STK_BACK( 2, RTMemAllocTag),
375	SUPEXP_STK_BACK( 2, RTMemAllocVarTag),
376	SUPEXP_STK_BACK( 2, RTMemAllocZTag),
377	SUPEXP_STK_BACK( 2, RTMemAllocZVarTag),
378	SUPEXP_STK_BACK( 4, RTMemDupExTag),
379	SUPEXP_STK_BACK( 3, RTMemDupTag),
380	SUPEXP_STK_BACK( 1, RTMemFree),
381	SUPEXP_STK_BACK( 2, RTMemFreeEx),
382	SUPEXP_STK_BACK( 3, RTMemReallocTag),
383	SUPEXP_STK_BACK( 0, RTMpCpuId),
384	SUPEXP_STK_BACK( 1, RTMpCpuIdFromSetIndex),
385	SUPEXP_STK_BACK( 1, RTMpCpuIdToSetIndex),
386	SUPEXP_STK_BACK( 0, RTMpCurSetIndex),
387	SUPEXP_STK_BACK( 1, RTMpCurSetIndexAndId),
388	SUPEXP_STK_BACK( 0, RTMpGetArraySize),
389	SUPEXP_STK_BACK( 0, RTMpGetCount),
390	SUPEXP_STK_BACK( 0, RTMpGetMaxCpuId),
391	SUPEXP_STK_BACK( 0, RTMpGetOnlineCount),
392	SUPEXP_STK_BACK( 1, RTMpGetOnlineSet),
393	SUPEXP_STK_BACK( 1, RTMpGetSet),
394	SUPEXP_STK_BACK( 1, RTMpIsCpuOnline),
395	SUPEXP_STK_BACK( 1, RTMpIsCpuPossible),
396	SUPEXP_STK_BACK( 0, RTMpIsCpuWorkPending),
397	SUPEXP_STK_BACKF( 2, RTMpNotificationDeregister),
398	SUPEXP_STK_BACKF( 2, RTMpNotificationRegister),
399	SUPEXP_STK_BACKF( 3, RTMpOnAll),
400	SUPEXP_STK_BACKF( 3, RTMpOnOthers),
401	SUPEXP_STK_BACKF( 4, RTMpOnSpecific),
402	SUPEXP_STK_BACK( 1, RTMpPokeCpu),
403	SUPEXP_STK_OKAY( 4, RTNetIPv4AddDataChecksum),
404	SUPEXP_STK_OKAY( 2, RTNetIPv4AddTCPChecksum),
405	SUPEXP_STK_OKAY( 2, RTNetIPv4AddUDPChecksum),
406	SUPEXP_STK_OKAY( 1, RTNetIPv4FinalizeChecksum),
407	SUPEXP_STK_OKAY( 1, RTNetIPv4HdrChecksum),
408	SUPEXP_STK_OKAY( 4, RTNetIPv4IsDHCPValid),
409	SUPEXP_STK_OKAY( 4, RTNetIPv4IsHdrValid),
410	SUPEXP_STK_OKAY( 4, RTNetIPv4IsTCPSizeValid),
411	SUPEXP_STK_OKAY( 6, RTNetIPv4IsTCPValid),
412	SUPEXP_STK_OKAY( 3, RTNetIPv4IsUDPSizeValid),
413	SUPEXP_STK_OKAY( 5, RTNetIPv4IsUDPValid),
414	SUPEXP_STK_OKAY( 1, RTNetIPv4PseudoChecksum),
415	SUPEXP_STK_OKAY( 4, RTNetIPv4PseudoChecksumBits),
416	SUPEXP_STK_OKAY( 3, RTNetIPv4TCPChecksum),
417	SUPEXP_STK_OKAY( 3, RTNetIPv4UDPChecksum),
418	SUPEXP_STK_OKAY( 1, RTNetIPv6PseudoChecksum),
419	SUPEXP_STK_OKAY( 4, RTNetIPv6PseudoChecksumBits),
420	SUPEXP_STK_OKAY( 3, RTNetIPv6PseudoChecksumEx),
421	SUPEXP_STK_OKAY( 4, RTNetTCPChecksum),
422	SUPEXP_STK_OKAY( 2, RTNetUDPChecksum),
423	SUPEXP_STK_BACKF( 2, RTPowerNotificationDeregister),
424	SUPEXP_STK_BACKF( 2, RTPowerNotificationRegister),
425	SUPEXP_STK_BACK( 0, RTProcSelf),
426	SUPEXP_STK_BACK( 0, RTR0AssertPanicSystem),
427	#if defined(RT_OS_DARWIN) \|\| defined(RT_OS_SOLARIS) \|\| defined(RT_OS_WINDOWS)
428	SUPEXP_STK_BACK( 2, RTR0DbgKrnlInfoOpen), /* only-darwin, only-solaris, only-windows */
429	SUPEXP_STK_BACK( 5, RTR0DbgKrnlInfoQueryMember), /* only-darwin, only-solaris, only-windows */
430	# if defined(RT_OS_SOLARIS)
431	SUPEXP_STK_BACK( 4, RTR0DbgKrnlInfoQuerySize), /* only-solaris */
432	# endif
433	SUPEXP_STK_BACK( 4, RTR0DbgKrnlInfoQuerySymbol), /* only-darwin, only-solaris, only-windows */
434	SUPEXP_STK_BACK( 1, RTR0DbgKrnlInfoRelease), /* only-darwin, only-solaris, only-windows */
435	SUPEXP_STK_BACK( 1, RTR0DbgKrnlInfoRetain), /* only-darwin, only-solaris, only-windows */
436	#endif
437	SUPEXP_STK_BACK( 0, RTR0MemAreKrnlAndUsrDifferent),
438	SUPEXP_STK_BACK( 1, RTR0MemKernelIsValidAddr),
439	SUPEXP_STK_BACK( 3, RTR0MemKernelCopyFrom),
440	SUPEXP_STK_BACK( 3, RTR0MemKernelCopyTo),
441	SUPEXP_STK_OKAY( 1, RTR0MemObjAddress),
442	SUPEXP_STK_OKAY( 1, RTR0MemObjAddressR3),
443	SUPEXP_STK_BACK( 4, RTR0MemObjAllocContTag),
444	SUPEXP_STK_BACK( 5, RTR0MemObjAllocLargeTag),
445	SUPEXP_STK_BACK( 4, RTR0MemObjAllocLowTag),
446	SUPEXP_STK_BACK( 4, RTR0MemObjAllocPageTag),
447	SUPEXP_STK_BACK( 5, RTR0MemObjAllocPhysExTag),
448	SUPEXP_STK_BACK( 4, RTR0MemObjAllocPhysNCTag),
449	SUPEXP_STK_BACK( 4, RTR0MemObjAllocPhysTag),
450	SUPEXP_STK_BACK( 5, RTR0MemObjEnterPhysTag),
451	SUPEXP_STK_BACK( 2, RTR0MemObjFree),
452	SUPEXP_STK_BACK( 2, RTR0MemObjGetPagePhysAddr),
453	SUPEXP_STK_OKAY( 1, RTR0MemObjIsMapping),
454	SUPEXP_STK_BACK( 6, RTR0MemObjLockUserTag),
455	SUPEXP_STK_BACK( 5, RTR0MemObjLockKernelTag),
456	SUPEXP_STK_BACK( 8, RTR0MemObjMapKernelExTag),
457	SUPEXP_STK_BACK( 6, RTR0MemObjMapKernelTag),
458	SUPEXP_STK_BACK( 9, RTR0MemObjMapUserExTag),
459	SUPEXP_STK_BACK( 7, RTR0MemObjMapUserTag),
460	SUPEXP_STK_BACK( 4, RTR0MemObjProtect),
461	SUPEXP_STK_OKAY( 1, RTR0MemObjSize),
462	SUPEXP_STK_OKAY( 1, RTR0MemObjWasZeroInitialized),
463	SUPEXP_STK_BACK( 3, RTR0MemUserCopyFrom),
464	SUPEXP_STK_BACK( 3, RTR0MemUserCopyTo),
465	SUPEXP_STK_BACK( 1, RTR0MemUserIsValidAddr),
466	SUPEXP_STK_BACK( 0, RTR0ProcHandleSelf),
467	SUPEXP_STK_BACK( 1, RTSemEventCreate),
468	SUPEXP_STK_BACK( 1, RTSemEventDestroy),
469	SUPEXP_STK_BACK( 0, RTSemEventGetResolution),
470	SUPEXP_STK_BACK( 0, RTSemEventIsSignalSafe),
471	SUPEXP_STK_BACK( 1, RTSemEventMultiCreate),
472	SUPEXP_STK_BACK( 1, RTSemEventMultiDestroy),
473	SUPEXP_STK_BACK( 0, RTSemEventMultiGetResolution),
474	SUPEXP_STK_BACK( 0, RTSemEventMultiIsSignalSafe),
475	SUPEXP_STK_BACK( 1, RTSemEventMultiReset),
476	SUPEXP_STK_BACK( 1, RTSemEventMultiSignal),
477	SUPEXP_STK_BACK( 2, RTSemEventMultiWait),
478	SUPEXP_STK_BACK( 3, RTSemEventMultiWaitEx),
479	SUPEXP_STK_BACK( 7, RTSemEventMultiWaitExDebug),
480	SUPEXP_STK_BACK( 2, RTSemEventMultiWaitNoResume),
481	SUPEXP_STK_BACK( 1, RTSemEventSignal),
482	SUPEXP_STK_BACK( 2, RTSemEventWait),
483	SUPEXP_STK_BACK( 3, RTSemEventWaitEx),
484	SUPEXP_STK_BACK( 7, RTSemEventWaitExDebug),
485	SUPEXP_STK_BACK( 2, RTSemEventWaitNoResume),
486	SUPEXP_STK_BACK( 1, RTSemFastMutexCreate),
487	SUPEXP_STK_BACK( 1, RTSemFastMutexDestroy),
488	SUPEXP_STK_BACK( 1, RTSemFastMutexRelease),
489	SUPEXP_STK_BACK( 1, RTSemFastMutexRequest),
490	SUPEXP_STK_BACK( 1, RTSemMutexCreate),
491	SUPEXP_STK_BACK( 1, RTSemMutexDestroy),
492	SUPEXP_STK_BACK( 1, RTSemMutexRelease),
493	SUPEXP_STK_BACK( 2, RTSemMutexRequest),
494	SUPEXP_STK_BACK( 6, RTSemMutexRequestDebug),
495	SUPEXP_STK_BACK( 2, RTSemMutexRequestNoResume),
496	SUPEXP_STK_BACK( 6, RTSemMutexRequestNoResumeDebug),
497	SUPEXP_STK_BACK( 1, RTSpinlockAcquire),
498	SUPEXP_STK_BACK( 3, RTSpinlockCreate),
499	SUPEXP_STK_BACK( 1, RTSpinlockDestroy),
500	SUPEXP_STK_BACK( 1, RTSpinlockRelease),
501	SUPEXP_STK_OKAY( 3, RTStrCopy),
502	SUPEXP_STK_BACK( 2, RTStrDupTag),
503	SUPEXP_STK_BACK( 6, RTStrFormatNumber),
504	SUPEXP_STK_BACK( 1, RTStrFormatTypeDeregister),
505	SUPEXP_STK_BACKF( 3, RTStrFormatTypeRegister),
506	SUPEXP_STK_BACKF( 2, RTStrFormatTypeSetUser),
507	SUPEXP_STK_BACKF( 6, RTStrFormatV),
508	SUPEXP_STK_BACK( 1, RTStrFree),
509	SUPEXP_STK_OKAY( 3, RTStrNCmp),
510	SUPEXP_STK_BACKF( 6, RTStrPrintfExV),
511	SUPEXP_STK_BACK( 4, RTStrPrintfV),
512	SUPEXP_STK_BACKF( 6, RTStrPrintf2ExV),
513	SUPEXP_STK_BACK( 4, RTStrPrintf2V),
514	SUPEXP_STK_BACKF( 7, RTThreadCreate),
515	SUPEXP_STK_BACK( 1, RTThreadCtxHookIsEnabled),
516	SUPEXP_STK_BACKF( 4, RTThreadCtxHookCreate),
517	SUPEXP_STK_BACK( 1, RTThreadCtxHookDestroy),
518	SUPEXP_STK_BACK( 1, RTThreadCtxHookDisable),
519	SUPEXP_STK_BACK( 1, RTThreadCtxHookEnable),
520	SUPEXP_STK_BACK( 1, RTThreadGetName),
521	SUPEXP_STK_BACK( 1, RTThreadGetNative),
522	SUPEXP_STK_BACK( 1, RTThreadGetType),
523	SUPEXP_STK_BACK( 1, RTThreadIsInInterrupt),
524	SUPEXP_STK_BACK( 0, RTThreadNativeSelf),
525	SUPEXP_STK_BACK( 1, RTThreadPreemptDisable),
526	SUPEXP_STK_BACK( 1, RTThreadPreemptIsEnabled),
527	SUPEXP_STK_BACK( 1, RTThreadPreemptIsPending),
528	SUPEXP_STK_BACK( 0, RTThreadPreemptIsPendingTrusty),
529	SUPEXP_STK_BACK( 0, RTThreadPreemptIsPossible),
530	SUPEXP_STK_BACK( 1, RTThreadPreemptRestore),
531	SUPEXP_STK_BACK( 1, RTThreadQueryTerminationStatus),
532	SUPEXP_STK_BACK( 0, RTThreadSelf),
533	SUPEXP_STK_BACK( 0, RTThreadSelfName),
534	SUPEXP_STK_BACK( 1, RTThreadSleep),
535	SUPEXP_STK_BACK( 1, RTThreadUserReset),
536	SUPEXP_STK_BACK( 1, RTThreadUserSignal),
537	SUPEXP_STK_BACK( 2, RTThreadUserWait),
538	SUPEXP_STK_BACK( 2, RTThreadUserWaitNoResume),
539	SUPEXP_STK_BACK( 3, RTThreadWait),
540	SUPEXP_STK_BACK( 3, RTThreadWaitNoResume),
541	SUPEXP_STK_BACK( 0, RTThreadYield),
542	SUPEXP_STK_BACK( 1, RTTimeNow),
543	SUPEXP_STK_BACK( 0, RTTimerCanDoHighResolution),
544	SUPEXP_STK_BACK( 2, RTTimerChangeInterval),
545	SUPEXP_STK_BACKF( 4, RTTimerCreate),
546	SUPEXP_STK_BACKF( 5, RTTimerCreateEx),
547	SUPEXP_STK_BACK( 1, RTTimerDestroy),
548	SUPEXP_STK_BACK( 0, RTTimerGetSystemGranularity),
549	SUPEXP_STK_BACK( 1, RTTimerReleaseSystemGranularity),
550	SUPEXP_STK_BACK( 2, RTTimerRequestSystemGranularity),
551	SUPEXP_STK_BACK( 2, RTTimerStart),
552	SUPEXP_STK_BACK( 1, RTTimerStop),
553	SUPEXP_STK_BACK( 0, RTTimeSystemMilliTS),
554	SUPEXP_STK_BACK( 0, RTTimeSystemNanoTS),
555	SUPEXP_STK_OKAY( 2, RTUuidCompare),
556	SUPEXP_STK_OKAY( 2, RTUuidCompareStr),
557	SUPEXP_STK_OKAY( 2, RTUuidFromStr),
558	/* SED: END */
559	};
560
561	#if defined(RT_OS_DARWIN) \|\| defined(RT_OS_SOLARIS) \|\| defined(RT_OS_FREEBSD)
562	/**
563	* Drag in the rest of IRPT since we share it with the
564	* rest of the kernel modules on darwin.
565	*/
566	struct CLANG11WERIDNESS { PFNRT pfn; } g_apfnVBoxDrvIPRTDeps[] =
567	{
568	/* VBoxNetAdp */
569	{ (PFNRT)RTRandBytes },
570	/* VBoxUSB */
571	{ (PFNRT)RTPathStripFilename },
572	#if !defined(RT_OS_FREEBSD)
573	{ (PFNRT)RTHandleTableAlloc },
574	{ (PFNRT)RTStrPurgeEncoding },
575	#endif
576	{ NULL }
577	};
578	#endif /* RT_OS_DARWIN \|\| RT_OS_SOLARIS \|\| RT_OS_FREEBSD */
579
580
581
582	/**
583	* Initializes the device extentsion structure.
584	*
585	* @returns IPRT status code.
586	* @param pDevExt The device extension to initialize.
587	* @param cbSession The size of the session structure. The size of
588	* SUPDRVSESSION may be smaller when SUPDRV_AGNOSTIC is
589	* defined because we're skipping the OS specific members
590	* then.
591	*/
592	int VBOXCALL supdrvInitDevExt(PSUPDRVDEVEXT pDevExt, size_t cbSession)
593	{
594	int rc;
595
596	#ifdef SUPDRV_WITH_RELEASE_LOGGER
597	/*
598	* Create the release log.
599	*/
600	static const char * const s_apszGroups[] = VBOX_LOGGROUP_NAMES;
601	PRTLOGGER pRelLogger;
602	rc = RTLogCreate(&pRelLogger, 0 /* fFlags */, "all",
603	"VBOX_RELEASE_LOG", RT_ELEMENTS(s_apszGroups), s_apszGroups, RTLOGDEST_STDOUT \| RTLOGDEST_DEBUGGER, NULL);
604	if (RT_SUCCESS(rc))
605	RTLogRelSetDefaultInstance(pRelLogger);
606	/** @todo Add native hook for getting logger config parameters and setting
607	* them. On linux we should use the module parameter stuff... */
608	#endif
609
610	#if (defined(RT_ARCH_AMD64) \|\| defined(RT_ARCH_X86)) && !defined(VBOX_WITH_OLD_CPU_SUPPORT)
611	/*
612	* Require SSE2 to be present.
613	*/
614	if (!(ASMCpuId_EDX(1) & X86_CPUID_FEATURE_EDX_SSE2))
615	{
616	SUPR0Printf("vboxdrv: Requires SSE2 (cpuid(0).EDX=%#x)\n", ASMCpuId_EDX(1));
617	return VERR_UNSUPPORTED_CPU;
618	}
619	#endif
620
621	/*
622	* Initialize it.
623	*/
624	memset(pDevExt, 0, sizeof(pDevExt)); / Does not wipe OS specific tail section of the structure. */
625	pDevExt->Spinlock = NIL_RTSPINLOCK;
626	pDevExt->hGipSpinlock = NIL_RTSPINLOCK;
627	pDevExt->hSessionHashTabSpinlock = NIL_RTSPINLOCK;
628	#ifdef SUPDRV_USE_MUTEX_FOR_LDR
629	pDevExt->mtxLdr = NIL_RTSEMMUTEX;
630	#else
631	pDevExt->mtxLdr = NIL_RTSEMFASTMUTEX;
632	#endif
633	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
634	pDevExt->mtxGip = NIL_RTSEMMUTEX;
635	pDevExt->mtxTscDelta = NIL_RTSEMMUTEX;
636	#else
637	pDevExt->mtxGip = NIL_RTSEMFASTMUTEX;
638	pDevExt->mtxTscDelta = NIL_RTSEMFASTMUTEX;
639	#endif
640
641	rc = RTSpinlockCreate(&pDevExt->Spinlock, RTSPINLOCK_FLAGS_INTERRUPT_SAFE, "SUPDrvDevExt");
642	if (RT_SUCCESS(rc))
643	rc = RTSpinlockCreate(&pDevExt->hGipSpinlock, RTSPINLOCK_FLAGS_INTERRUPT_SAFE, "SUPDrvGip");
644	if (RT_SUCCESS(rc))
645	rc = RTSpinlockCreate(&pDevExt->hSessionHashTabSpinlock, RTSPINLOCK_FLAGS_INTERRUPT_SAFE, "SUPDrvSession");
646
647	if (RT_SUCCESS(rc))
648	#ifdef SUPDRV_USE_MUTEX_FOR_LDR
649	rc = RTSemMutexCreate(&pDevExt->mtxLdr);
650	#else
651	rc = RTSemFastMutexCreate(&pDevExt->mtxLdr);
652	#endif
653	if (RT_SUCCESS(rc))
654	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
655	rc = RTSemMutexCreate(&pDevExt->mtxTscDelta);
656	#else
657	rc = RTSemFastMutexCreate(&pDevExt->mtxTscDelta);
658	#endif
659	if (RT_SUCCESS(rc))
660	{
661	rc = RTSemFastMutexCreate(&pDevExt->mtxComponentFactory);
662	if (RT_SUCCESS(rc))
663	{
664	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
665	rc = RTSemMutexCreate(&pDevExt->mtxGip);
666	#else
667	rc = RTSemFastMutexCreate(&pDevExt->mtxGip);
668	#endif
669	if (RT_SUCCESS(rc))
670	{
671	rc = supdrvGipCreate(pDevExt);
672	if (RT_SUCCESS(rc))
673	{
674	rc = supdrvTracerInit(pDevExt);
675	if (RT_SUCCESS(rc))
676	{
677	pDevExt->pLdrInitImage = NULL;
678	pDevExt->hLdrInitThread = NIL_RTNATIVETHREAD;
679	pDevExt->hLdrTermThread = NIL_RTNATIVETHREAD;
680	pDevExt->u32Cookie = BIRD; /** @todo make this random? */
681	pDevExt->cbSession = (uint32_t)cbSession;
682
683	/*
684	* Fixup the absolute symbols.
685	*
686	* Because of the table indexing assumptions we'll have a little #ifdef orgy
687	* here rather than distributing this to OS specific files. At least for now.
688	*/
689	#ifdef RT_OS_DARWIN
690	# if ARCH_BITS == 32
691	if (SUPR0GetPagingMode() >= SUPPAGINGMODE_AMD64)
692	{
693	g_aFunctions[0].pfn = (void )1; / SUPR0AbsIs64bit */
694	g_aFunctions[1].pfn = (void )0x80; / SUPR0Abs64bitKernelCS - KERNEL64_CS, seg.h */
695	g_aFunctions[2].pfn = (void )0x88; / SUPR0Abs64bitKernelSS - KERNEL64_SS, seg.h */
696	g_aFunctions[3].pfn = (void )0x88; / SUPR0Abs64bitKernelDS - KERNEL64_SS, seg.h */
697	}
698	else
699	g_aFunctions[0].pfn = g_aFunctions[1].pfn = g_aFunctions[2].pfn = g_aFunctions[3].pfn = (void *)0;
700	g_aFunctions[4].pfn = (void )0x08; / SUPR0AbsKernelCS - KERNEL_CS, seg.h */
701	g_aFunctions[5].pfn = (void )0x10; / SUPR0AbsKernelSS - KERNEL_DS, seg.h */
702	g_aFunctions[6].pfn = (void )0x10; / SUPR0AbsKernelDS - KERNEL_DS, seg.h */
703	g_aFunctions[7].pfn = (void )0x10; / SUPR0AbsKernelES - KERNEL_DS, seg.h */
704	g_aFunctions[8].pfn = (void )0x10; / SUPR0AbsKernelFS - KERNEL_DS, seg.h */
705	g_aFunctions[9].pfn = (void )0x48; / SUPR0AbsKernelGS - CPU_DATA_GS, seg.h */
706	# else /* 64-bit darwin: */
707	g_aFunctions[0].pfn = (void )1; / SUPR0AbsIs64bit */
708	g_aFunctions[1].pfn = (void )(uintptr_t)ASMGetCS(); / SUPR0Abs64bitKernelCS */
709	g_aFunctions[2].pfn = (void )(uintptr_t)ASMGetSS(); / SUPR0Abs64bitKernelSS */
710	g_aFunctions[3].pfn = (void )0; / SUPR0Abs64bitKernelDS */
711	g_aFunctions[4].pfn = (void )(uintptr_t)ASMGetCS(); / SUPR0AbsKernelCS */
712	g_aFunctions[5].pfn = (void )(uintptr_t)ASMGetSS(); / SUPR0AbsKernelSS */
713	g_aFunctions[6].pfn = (void )0; / SUPR0AbsKernelDS */
714	g_aFunctions[7].pfn = (void )0; / SUPR0AbsKernelES */
715	g_aFunctions[8].pfn = (void )0; / SUPR0AbsKernelFS */
716	g_aFunctions[9].pfn = (void )0; / SUPR0AbsKernelGS */
717
718	# endif
719	#else /* !RT_OS_DARWIN */
720	# if ARCH_BITS == 64
721	g_aFunctions[0].pfn = (void )1; / SUPR0AbsIs64bit */
722	g_aFunctions[1].pfn = (void )(uintptr_t)ASMGetCS(); / SUPR0Abs64bitKernelCS */
723	g_aFunctions[2].pfn = (void )(uintptr_t)ASMGetSS(); / SUPR0Abs64bitKernelSS */
724	g_aFunctions[3].pfn = (void )(uintptr_t)ASMGetDS(); / SUPR0Abs64bitKernelDS */
725	# else
726	g_aFunctions[0].pfn = g_aFunctions[1].pfn = g_aFunctions[2].pfn = g_aFunctions[3].pfn = (void *)0;
727	# endif
728	g_aFunctions[4].pfn = (void )(uintptr_t)ASMGetCS(); / SUPR0AbsKernelCS */
729	g_aFunctions[5].pfn = (void )(uintptr_t)ASMGetSS(); / SUPR0AbsKernelSS */
730	g_aFunctions[6].pfn = (void )(uintptr_t)ASMGetDS(); / SUPR0AbsKernelDS */
731	g_aFunctions[7].pfn = (void )(uintptr_t)ASMGetES(); / SUPR0AbsKernelES */
732	g_aFunctions[8].pfn = (void )(uintptr_t)ASMGetFS(); / SUPR0AbsKernelFS */
733	g_aFunctions[9].pfn = (void )(uintptr_t)ASMGetGS(); / SUPR0AbsKernelGS */
734	#endif /* !RT_OS_DARWIN */
735	return VINF_SUCCESS;
736	}
737
738	supdrvGipDestroy(pDevExt);
739	}
740
741	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
742	RTSemMutexDestroy(pDevExt->mtxGip);
743	pDevExt->mtxGip = NIL_RTSEMMUTEX;
744	#else
745	RTSemFastMutexDestroy(pDevExt->mtxGip);
746	pDevExt->mtxGip = NIL_RTSEMFASTMUTEX;
747	#endif
748	}
749	RTSemFastMutexDestroy(pDevExt->mtxComponentFactory);
750	pDevExt->mtxComponentFactory = NIL_RTSEMFASTMUTEX;
751	}
752	}
753
754	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
755	RTSemMutexDestroy(pDevExt->mtxTscDelta);
756	pDevExt->mtxTscDelta = NIL_RTSEMMUTEX;
757	#else
758	RTSemFastMutexDestroy(pDevExt->mtxTscDelta);
759	pDevExt->mtxTscDelta = NIL_RTSEMFASTMUTEX;
760	#endif
761	#ifdef SUPDRV_USE_MUTEX_FOR_LDR
762	RTSemMutexDestroy(pDevExt->mtxLdr);
763	pDevExt->mtxLdr = NIL_RTSEMMUTEX;
764	#else
765	RTSemFastMutexDestroy(pDevExt->mtxLdr);
766	pDevExt->mtxLdr = NIL_RTSEMFASTMUTEX;
767	#endif
768	RTSpinlockDestroy(pDevExt->Spinlock);
769	pDevExt->Spinlock = NIL_RTSPINLOCK;
770	RTSpinlockDestroy(pDevExt->hGipSpinlock);
771	pDevExt->hGipSpinlock = NIL_RTSPINLOCK;
772	RTSpinlockDestroy(pDevExt->hSessionHashTabSpinlock);
773	pDevExt->hSessionHashTabSpinlock = NIL_RTSPINLOCK;
774
775	#ifdef SUPDRV_WITH_RELEASE_LOGGER
776	RTLogDestroy(RTLogRelSetDefaultInstance(NULL));
777	RTLogDestroy(RTLogSetDefaultInstance(NULL));
778	#endif
779
780	return rc;
781	}
782
783
784	/**
785	* Delete the device extension (e.g. cleanup members).
786	*
787	* @param pDevExt The device extension to delete.
788	*/
789	void VBOXCALL supdrvDeleteDevExt(PSUPDRVDEVEXT pDevExt)
790	{
791	PSUPDRVOBJ pObj;
792	PSUPDRVUSAGE pUsage;
793
794	/*
795	* Kill mutexes and spinlocks.
796	*/
797	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
798	RTSemMutexDestroy(pDevExt->mtxGip);
799	pDevExt->mtxGip = NIL_RTSEMMUTEX;
800	RTSemMutexDestroy(pDevExt->mtxTscDelta);
801	pDevExt->mtxTscDelta = NIL_RTSEMMUTEX;
802	#else
803	RTSemFastMutexDestroy(pDevExt->mtxGip);
804	pDevExt->mtxGip = NIL_RTSEMFASTMUTEX;
805	RTSemFastMutexDestroy(pDevExt->mtxTscDelta);
806	pDevExt->mtxTscDelta = NIL_RTSEMFASTMUTEX;
807	#endif
808	#ifdef SUPDRV_USE_MUTEX_FOR_LDR
809	RTSemMutexDestroy(pDevExt->mtxLdr);
810	pDevExt->mtxLdr = NIL_RTSEMMUTEX;
811	#else
812	RTSemFastMutexDestroy(pDevExt->mtxLdr);
813	pDevExt->mtxLdr = NIL_RTSEMFASTMUTEX;
814	#endif
815	RTSpinlockDestroy(pDevExt->Spinlock);
816	pDevExt->Spinlock = NIL_RTSPINLOCK;
817	RTSemFastMutexDestroy(pDevExt->mtxComponentFactory);
818	pDevExt->mtxComponentFactory = NIL_RTSEMFASTMUTEX;
819	RTSpinlockDestroy(pDevExt->hSessionHashTabSpinlock);
820	pDevExt->hSessionHashTabSpinlock = NIL_RTSPINLOCK;
821
822	/*
823	* Free lists.
824	*/
825	/* objects. */
826	pObj = pDevExt->pObjs;
827	Assert(!pObj); /* (can trigger on forced unloads) */
828	pDevExt->pObjs = NULL;
829	while (pObj)
830	{
831	void *pvFree = pObj;
832	pObj = pObj->pNext;
833	RTMemFree(pvFree);
834	}
835
836	/* usage records. */
837	pUsage = pDevExt->pUsageFree;
838	pDevExt->pUsageFree = NULL;
839	while (pUsage)
840	{
841	void *pvFree = pUsage;
842	pUsage = pUsage->pNext;
843	RTMemFree(pvFree);
844	}
845
846	/* kill the GIP. */
847	supdrvGipDestroy(pDevExt);
848	RTSpinlockDestroy(pDevExt->hGipSpinlock);
849	pDevExt->hGipSpinlock = NIL_RTSPINLOCK;
850
851	supdrvTracerTerm(pDevExt);
852
853	#ifdef SUPDRV_WITH_RELEASE_LOGGER
854	/* destroy the loggers. */
855	RTLogDestroy(RTLogRelSetDefaultInstance(NULL));
856	RTLogDestroy(RTLogSetDefaultInstance(NULL));
857	#endif
858	}
859
860
861	/**
862	* Create session.
863	*
864	* @returns IPRT status code.
865	* @param pDevExt Device extension.
866	* @param fUser Flag indicating whether this is a user or kernel
867	* session.
868	* @param fUnrestricted Unrestricted access (system) or restricted access
869	* (user)?
870	* @param ppSession Where to store the pointer to the session data.
871	*/
872	int VBOXCALL supdrvCreateSession(PSUPDRVDEVEXT pDevExt, bool fUser, bool fUnrestricted, PSUPDRVSESSION *ppSession)
873	{
874	int rc;
875	PSUPDRVSESSION pSession;
876
877	if (!SUP_IS_DEVEXT_VALID(pDevExt))
878	return VERR_INVALID_PARAMETER;
879
880	/*
881	* Allocate memory for the session data.
882	*/
883	pSession = *ppSession = (PSUPDRVSESSION)RTMemAllocZ(pDevExt->cbSession);
884	if (pSession)
885	{
886	/* Initialize session data. */
887	rc = RTSpinlockCreate(&pSession->Spinlock, RTSPINLOCK_FLAGS_INTERRUPT_UNSAFE, "SUPDrvSession");
888	if (!rc)
889	{
890	rc = RTHandleTableCreateEx(&pSession->hHandleTable,
891	RTHANDLETABLE_FLAGS_LOCKED_IRQ_SAFE \| RTHANDLETABLE_FLAGS_CONTEXT,
892	1 /uBase/, 32768 /cMax/, supdrvSessionObjHandleRetain, pSession);
893	if (RT_SUCCESS(rc))
894	{
895	Assert(pSession->Spinlock != NIL_RTSPINLOCK);
896	pSession->pDevExt = pDevExt;
897	pSession->u32Cookie = BIRD_INV;
898	pSession->fUnrestricted = fUnrestricted;
899	/pSession->fInHashTable = false; /
900	pSession->cRefs = 1;
901	/*pSession->pCommonNextHash = NULL;
902	pSession->ppOsSessionPtr = NULL; */
903	if (fUser)
904	{
905	pSession->Process = RTProcSelf();
906	pSession->R0Process = RTR0ProcHandleSelf();
907	}
908	else
909	{
910	pSession->Process = NIL_RTPROCESS;
911	pSession->R0Process = NIL_RTR0PROCESS;
912	}
913	/*pSession->pLdrUsage = NULL;
914	pSession->pVM = NULL;
915	pSession->pUsage = NULL;
916	pSession->pGip = NULL;
917	pSession->fGipReferenced = false;
918	pSession->Bundle.cUsed = 0; */
919	pSession->Uid = NIL_RTUID;
920	pSession->Gid = NIL_RTGID;
921	/pSession->uTracerData = 0;/
922	pSession->hTracerCaller = NIL_RTNATIVETHREAD;
923	RTListInit(&pSession->TpProviders);
924	/pSession->cTpProviders = 0;/
925	/pSession->cTpProbesFiring = 0;/
926	RTListInit(&pSession->TpUmods);
927	/RT_ZERO(pSession->apTpLookupTable);/
928
929	VBOXDRV_SESSION_CREATE(pSession, fUser);
930	LogFlow(("Created session %p initial cookie=%#x\n", pSession, pSession->u32Cookie));
931	return VINF_SUCCESS;
932	}
933
934	RTSpinlockDestroy(pSession->Spinlock);
935	}
936	RTMemFree(pSession);
937	*ppSession = NULL;
938	Log(("Failed to create spinlock, rc=%d!\n", rc));
939	}
940	else
941	rc = VERR_NO_MEMORY;
942
943	return rc;
944	}
945
946
947	/**
948	* Cleans up the session in the context of the process to which it belongs, the
949	* caller will free the session and the session spinlock.
950	*
951	* This should normally occur when the session is closed or as the process
952	* exits. Careful reference counting in the OS specfic code makes sure that
953	* there cannot be any races between process/handle cleanup callbacks and
954	* threads doing I/O control calls.
955	*
956	* @param pDevExt The device extension.
957	* @param pSession Session data.
958	*/
959	static void supdrvCleanupSession(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession)
960	{
961	int rc;
962	PSUPDRVBUNDLE pBundle;
963	LogFlow(("supdrvCleanupSession: pSession=%p\n", pSession));
964
965	Assert(!pSession->fInHashTable);
966	Assert(!pSession->ppOsSessionPtr);
967	AssertLogRelMsg(pSession->R0Process == RTR0ProcHandleSelf() \|\| pSession->R0Process == NIL_RTR0PROCESS,
968	("R0Process=%p cur=%p; curpid=%u\n",
969	pSession->R0Process, RTR0ProcHandleSelf(), RTProcSelf()));
970
971	/*
972	* Remove logger instances related to this session.
973	*/
974	RTLogSetDefaultInstanceThread(NULL, (uintptr_t)pSession);
975
976	/*
977	* Destroy the handle table.
978	*/
979	rc = RTHandleTableDestroy(pSession->hHandleTable, supdrvSessionObjHandleDelete, pSession);
980	AssertRC(rc);
981	pSession->hHandleTable = NIL_RTHANDLETABLE;
982
983	/*
984	* Release object references made in this session.
985	* In theory there should be noone racing us in this session.
986	*/
987	Log2(("release objects - start\n"));
988	if (pSession->pUsage)
989	{
990	PSUPDRVUSAGE pUsage;
991	RTSpinlockAcquire(pDevExt->Spinlock);
992
993	while ((pUsage = pSession->pUsage) != NULL)
994	{
995	PSUPDRVOBJ pObj = pUsage->pObj;
996	pSession->pUsage = pUsage->pNext;
997
998	AssertMsg(pUsage->cUsage >= 1 && pObj->cUsage >= pUsage->cUsage, ("glob %d; sess %d\n", pObj->cUsage, pUsage->cUsage));
999	if (pUsage->cUsage < pObj->cUsage)
1000	{
1001	pObj->cUsage -= pUsage->cUsage;
1002	RTSpinlockRelease(pDevExt->Spinlock);
1003	}
1004	else
1005	{
1006	/* Destroy the object and free the record. */
1007	if (pDevExt->pObjs == pObj)
1008	pDevExt->pObjs = pObj->pNext;
1009	else
1010	{
1011	PSUPDRVOBJ pObjPrev;
1012	for (pObjPrev = pDevExt->pObjs; pObjPrev; pObjPrev = pObjPrev->pNext)
1013	if (pObjPrev->pNext == pObj)
1014	{
1015	pObjPrev->pNext = pObj->pNext;
1016	break;
1017	}
1018	Assert(pObjPrev);
1019	}
1020	RTSpinlockRelease(pDevExt->Spinlock);
1021
1022	Log(("supdrvCleanupSession: destroying %p/%d (%p/%p) cpid=%RTproc pid=%RTproc dtor=%p\n",
1023	pObj, pObj->enmType, pObj->pvUser1, pObj->pvUser2, pObj->CreatorProcess, RTProcSelf(), pObj->pfnDestructor));
1024	if (pObj->pfnDestructor)
1025	pObj->pfnDestructor(pObj, pObj->pvUser1, pObj->pvUser2);
1026	RTMemFree(pObj);
1027	}
1028
1029	/* free it and continue. */
1030	RTMemFree(pUsage);
1031
1032	RTSpinlockAcquire(pDevExt->Spinlock);
1033	}
1034
1035	RTSpinlockRelease(pDevExt->Spinlock);
1036	AssertMsg(!pSession->pUsage, ("Some buster reregistered an object during desturction!\n"));
1037	}
1038	Log2(("release objects - done\n"));
1039
1040	/*
1041	* Make sure the associated VM pointers are NULL.
1042	*/
1043	if (pSession->pSessionGVM \|\| pSession->pSessionVM \|\| pSession->pFastIoCtrlVM)
1044	{
1045	SUPR0Printf("supdrvCleanupSession: VM not disassociated! pSessionGVM=%p pSessionVM=%p pFastIoCtrlVM=%p\n",
1046	pSession->pSessionGVM, pSession->pSessionVM, pSession->pFastIoCtrlVM);
1047	pSession->pSessionGVM = NULL;
1048	pSession->pSessionVM = NULL;
1049	pSession->pFastIoCtrlVM = NULL;
1050	}
1051
1052	/*
1053	* Do tracer cleanups related to this session.
1054	*/
1055	Log2(("release tracer stuff - start\n"));
1056	supdrvTracerCleanupSession(pDevExt, pSession);
1057	Log2(("release tracer stuff - end\n"));
1058
1059	/*
1060	* Release memory allocated in the session.
1061	*
1062	* We do not serialize this as we assume that the application will
1063	* not allocated memory while closing the file handle object.
1064	*/
1065	Log2(("freeing memory:\n"));
1066	pBundle = &pSession->Bundle;
1067	while (pBundle)
1068	{
1069	PSUPDRVBUNDLE pToFree;
1070	unsigned i;
1071
1072	/*
1073	* Check and unlock all entries in the bundle.
1074	*/
1075	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
1076	{
1077	if (pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ)
1078	{
1079	Log2(("eType=%d pvR0=%p pvR3=%p cb=%ld\n", pBundle->aMem[i].eType, RTR0MemObjAddress(pBundle->aMem[i].MemObj),
1080	(void *)RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3), (long)RTR0MemObjSize(pBundle->aMem[i].MemObj)));
1081	if (pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ)
1082	{
1083	rc = RTR0MemObjFree(pBundle->aMem[i].MapObjR3, false);
1084	AssertRC(rc); /** @todo figure out how to handle this. */
1085	pBundle->aMem[i].MapObjR3 = NIL_RTR0MEMOBJ;
1086	}
1087	rc = RTR0MemObjFree(pBundle->aMem[i].MemObj, true /* fFreeMappings */);
1088	AssertRC(rc); /** @todo figure out how to handle this. */
1089	pBundle->aMem[i].MemObj = NIL_RTR0MEMOBJ;
1090	pBundle->aMem[i].eType = MEMREF_TYPE_UNUSED;
1091	}
1092	}
1093
1094	/*
1095	* Advance and free previous bundle.
1096	*/
1097	pToFree = pBundle;
1098	pBundle = pBundle->pNext;
1099
1100	pToFree->pNext = NULL;
1101	pToFree->cUsed = 0;
1102	if (pToFree != &pSession->Bundle)
1103	RTMemFree(pToFree);
1104	}
1105	Log2(("freeing memory - done\n"));
1106
1107	/*
1108	* Deregister component factories.
1109	*/
1110	RTSemFastMutexRequest(pDevExt->mtxComponentFactory);
1111	Log2(("deregistering component factories:\n"));
1112	if (pDevExt->pComponentFactoryHead)
1113	{
1114	PSUPDRVFACTORYREG pPrev = NULL;
1115	PSUPDRVFACTORYREG pCur = pDevExt->pComponentFactoryHead;
1116	while (pCur)
1117	{
1118	if (pCur->pSession == pSession)
1119	{
1120	/* unlink it */
1121	PSUPDRVFACTORYREG pNext = pCur->pNext;
1122	if (pPrev)
1123	pPrev->pNext = pNext;
1124	else
1125	pDevExt->pComponentFactoryHead = pNext;
1126
1127	/* free it */
1128	pCur->pNext = NULL;
1129	pCur->pSession = NULL;
1130	pCur->pFactory = NULL;
1131	RTMemFree(pCur);
1132
1133	/* next */
1134	pCur = pNext;
1135	}
1136	else
1137	{
1138	/* next */
1139	pPrev = pCur;
1140	pCur = pCur->pNext;
1141	}
1142	}
1143	}
1144	RTSemFastMutexRelease(pDevExt->mtxComponentFactory);
1145	Log2(("deregistering component factories - done\n"));
1146
1147	/*
1148	* Loaded images needs to be dereferenced and possibly freed up.
1149	*/
1150	supdrvLdrLock(pDevExt);
1151	Log2(("freeing images:\n"));
1152	if (pSession->pLdrUsage)
1153	{
1154	PSUPDRVLDRUSAGE pUsage = pSession->pLdrUsage;
1155	pSession->pLdrUsage = NULL;
1156	while (pUsage)
1157	{
1158	void *pvFree = pUsage;
1159	PSUPDRVLDRIMAGE pImage = pUsage->pImage;
1160	uint32_t cUsage = pUsage->cRing0Usage + pUsage->cRing3Usage;
1161	if (pImage->cImgUsage > cUsage)
1162	supdrvLdrSubtractUsage(pDevExt, pImage, cUsage);
1163	else
1164	supdrvLdrFree(pDevExt, pImage);
1165	pUsage->pImage = NULL;
1166	pUsage = pUsage->pNext;
1167	RTMemFree(pvFree);
1168	}
1169	}
1170	supdrvLdrUnlock(pDevExt);
1171	Log2(("freeing images - done\n"));
1172
1173	/*
1174	* Unmap the GIP.
1175	*/
1176	Log2(("umapping GIP:\n"));
1177	if (pSession->GipMapObjR3 != NIL_RTR0MEMOBJ)
1178	{
1179	SUPR0GipUnmap(pSession);
1180	pSession->fGipReferenced = 0;
1181	}
1182	Log2(("umapping GIP - done\n"));
1183	}
1184
1185
1186	/**
1187	* Common code for freeing a session when the reference count reaches zero.
1188	*
1189	* @param pDevExt Device extension.
1190	* @param pSession Session data.
1191	* This data will be freed by this routine.
1192	*/
1193	static void supdrvDestroySession(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession)
1194	{
1195	VBOXDRV_SESSION_CLOSE(pSession);
1196
1197	/*
1198	* Cleanup the session first.
1199	*/
1200	supdrvCleanupSession(pDevExt, pSession);
1201	supdrvOSCleanupSession(pDevExt, pSession);
1202
1203	/*
1204	* Free the rest of the session stuff.
1205	*/
1206	RTSpinlockDestroy(pSession->Spinlock);
1207	pSession->Spinlock = NIL_RTSPINLOCK;
1208	pSession->pDevExt = NULL;
1209	RTMemFree(pSession);
1210	LogFlow(("supdrvDestroySession: returns\n"));
1211	}
1212
1213
1214	/**
1215	* Inserts the session into the global hash table.
1216	*
1217	* @retval VINF_SUCCESS on success.
1218	* @retval VERR_WRONG_ORDER if the session was already inserted (asserted).
1219	* @retval VERR_INVALID_PARAMETER if the session handle is invalid or a ring-0
1220	* session (asserted).
1221	* @retval VERR_DUPLICATE if there is already a session for that pid.
1222	*
1223	* @param pDevExt The device extension.
1224	* @param pSession The session.
1225	* @param ppOsSessionPtr Pointer to the OS session pointer, if any is
1226	* available and used. This will set to point to the
1227	* session while under the protection of the session
1228	* hash table spinlock. It will also be kept in
1229	* PSUPDRVSESSION::ppOsSessionPtr for lookup and
1230	* cleanup use.
1231	* @param pvUser Argument for supdrvOSSessionHashTabInserted.
1232	*/
1233	int VBOXCALL supdrvSessionHashTabInsert(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPDRVSESSION *ppOsSessionPtr,
1234	void *pvUser)
1235	{
1236	PSUPDRVSESSION pCur;
1237	unsigned iHash;
1238
1239	/*
1240	* Validate input.
1241	*/
1242	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
1243	AssertReturn(pSession->R0Process != NIL_RTR0PROCESS, VERR_INVALID_PARAMETER);
1244
1245	/*
1246	* Calculate the hash table index and acquire the spinlock.
1247	*/
1248	iHash = SUPDRV_SESSION_HASH(pSession->Process);
1249
1250	RTSpinlockAcquire(pDevExt->hSessionHashTabSpinlock);
1251
1252	/*
1253	* If there are a collisions, we need to carefully check if we got a
1254	* duplicate. There can only be one open session per process.
1255	*/
1256	pCur = pDevExt->apSessionHashTab[iHash];
1257	if (pCur)
1258	{
1259	while (pCur && pCur->Process != pSession->Process)
1260	pCur = pCur->pCommonNextHash;
1261
1262	if (pCur)
1263	{
1264	RTSpinlockRelease(pDevExt->hSessionHashTabSpinlock);
1265	if (pCur == pSession)
1266	{
1267	Assert(pSession->fInHashTable);
1268	AssertFailed();
1269	return VERR_WRONG_ORDER;
1270	}
1271	Assert(!pSession->fInHashTable);
1272	if (pCur->R0Process == pSession->R0Process)
1273	return VERR_RESOURCE_IN_USE;
1274	return VERR_DUPLICATE;
1275	}
1276	}
1277	Assert(!pSession->fInHashTable);
1278	Assert(!pSession->ppOsSessionPtr);
1279
1280	/*
1281	* Insert it, doing a callout to the OS specific code in case it has
1282	* anything it wishes to do while we're holding the spinlock.
1283	*/
1284	pSession->pCommonNextHash = pDevExt->apSessionHashTab[iHash];
1285	pDevExt->apSessionHashTab[iHash] = pSession;
1286	pSession->fInHashTable = true;
1287	ASMAtomicIncS32(&pDevExt->cSessions);
1288
1289	pSession->ppOsSessionPtr = ppOsSessionPtr;
1290	if (ppOsSessionPtr)
1291	ASMAtomicWritePtr(ppOsSessionPtr, pSession);
1292
1293	supdrvOSSessionHashTabInserted(pDevExt, pSession, pvUser);
1294
1295	/*
1296	* Retain a reference for the pointer in the session table.
1297	*/
1298	ASMAtomicIncU32(&pSession->cRefs);
1299
1300	RTSpinlockRelease(pDevExt->hSessionHashTabSpinlock);
1301	return VINF_SUCCESS;
1302	}
1303
1304
1305	/**
1306	* Removes the session from the global hash table.
1307	*
1308	* @retval VINF_SUCCESS on success.
1309	* @retval VERR_NOT_FOUND if the session was already removed (asserted).
1310	* @retval VERR_INVALID_PARAMETER if the session handle is invalid or a ring-0
1311	* session (asserted).
1312	*
1313	* @param pDevExt The device extension.
1314	* @param pSession The session. The caller is expected to have a reference
1315	* to this so it won't croak on us when we release the hash
1316	* table reference.
1317	* @param pvUser OS specific context value for the
1318	* supdrvOSSessionHashTabInserted callback.
1319	*/
1320	int VBOXCALL supdrvSessionHashTabRemove(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, void *pvUser)
1321	{
1322	PSUPDRVSESSION pCur;
1323	unsigned iHash;
1324	int32_t cRefs;
1325
1326	/*
1327	* Validate input.
1328	*/
1329	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
1330	AssertReturn(pSession->R0Process != NIL_RTR0PROCESS, VERR_INVALID_PARAMETER);
1331
1332	/*
1333	* Calculate the hash table index and acquire the spinlock.
1334	*/
1335	iHash = SUPDRV_SESSION_HASH(pSession->Process);
1336
1337	RTSpinlockAcquire(pDevExt->hSessionHashTabSpinlock);
1338
1339	/*
1340	* Unlink it.
1341	*/
1342	pCur = pDevExt->apSessionHashTab[iHash];
1343	if (pCur == pSession)
1344	pDevExt->apSessionHashTab[iHash] = pSession->pCommonNextHash;
1345	else
1346	{
1347	PSUPDRVSESSION pPrev = pCur;
1348	while (pCur && pCur != pSession)
1349	{
1350	pPrev = pCur;
1351	pCur = pCur->pCommonNextHash;
1352	}
1353	if (pCur)
1354	pPrev->pCommonNextHash = pCur->pCommonNextHash;
1355	else
1356	{
1357	Assert(!pSession->fInHashTable);
1358	RTSpinlockRelease(pDevExt->hSessionHashTabSpinlock);
1359	return VERR_NOT_FOUND;
1360	}
1361	}
1362
1363	pSession->pCommonNextHash = NULL;
1364	pSession->fInHashTable = false;
1365
1366	ASMAtomicDecS32(&pDevExt->cSessions);
1367
1368	/*
1369	* Clear OS specific session pointer if available and do the OS callback.
1370	*/
1371	if (pSession->ppOsSessionPtr)
1372	{
1373	ASMAtomicCmpXchgPtr(pSession->ppOsSessionPtr, NULL, pSession);
1374	pSession->ppOsSessionPtr = NULL;
1375	}
1376
1377	supdrvOSSessionHashTabRemoved(pDevExt, pSession, pvUser);
1378
1379	RTSpinlockRelease(pDevExt->hSessionHashTabSpinlock);
1380
1381	/*
1382	* Drop the reference the hash table had to the session. This shouldn't
1383	* be the last reference!
1384	*/
1385	cRefs = ASMAtomicDecU32(&pSession->cRefs);
1386	Assert(cRefs > 0 && cRefs < _1M);
1387	if (cRefs == 0)
1388	supdrvDestroySession(pDevExt, pSession);
1389
1390	return VINF_SUCCESS;
1391	}
1392
1393
1394	/**
1395	* Looks up the session for the current process in the global hash table or in
1396	* OS specific pointer.
1397	*
1398	* @returns Pointer to the session with a reference that the caller must
1399	* release. If no valid session was found, NULL is returned.
1400	*
1401	* @param pDevExt The device extension.
1402	* @param Process The process ID.
1403	* @param R0Process The ring-0 process handle.
1404	* @param ppOsSessionPtr The OS session pointer if available. If not NULL,
1405	* this is used instead of the hash table. For
1406	* additional safety it must then be equal to the
1407	* SUPDRVSESSION::ppOsSessionPtr member.
1408	* This can be NULL even if the OS has a session
1409	* pointer.
1410	*/
1411	PSUPDRVSESSION VBOXCALL supdrvSessionHashTabLookup(PSUPDRVDEVEXT pDevExt, RTPROCESS Process, RTR0PROCESS R0Process,
1412	PSUPDRVSESSION *ppOsSessionPtr)
1413	{
1414	PSUPDRVSESSION pCur;
1415	unsigned iHash;
1416
1417	/*
1418	* Validate input.
1419	*/
1420	AssertReturn(R0Process != NIL_RTR0PROCESS, NULL);
1421
1422	/*
1423	* Calculate the hash table index and acquire the spinlock.
1424	*/
1425	iHash = SUPDRV_SESSION_HASH(Process);
1426
1427	RTSpinlockAcquire(pDevExt->hSessionHashTabSpinlock);
1428
1429	/*
1430	* If an OS session pointer is provided, always use it.
1431	*/
1432	if (ppOsSessionPtr)
1433	{
1434	pCur = *ppOsSessionPtr;
1435	if ( pCur
1436	&& ( pCur->ppOsSessionPtr != ppOsSessionPtr
1437	\|\| pCur->Process != Process
1438	\|\| pCur->R0Process != R0Process) )
1439	pCur = NULL;
1440	}
1441	else
1442	{
1443	/*
1444	* Otherwise, do the hash table lookup.
1445	*/
1446	pCur = pDevExt->apSessionHashTab[iHash];
1447	while ( pCur
1448	&& ( pCur->Process != Process
1449	\|\| pCur->R0Process != R0Process) )
1450	pCur = pCur->pCommonNextHash;
1451	}
1452
1453	/*
1454	* Retain the session.
1455	*/
1456	if (pCur)
1457	{
1458	uint32_t cRefs = ASMAtomicIncU32(&pCur->cRefs);
1459	NOREF(cRefs);
1460	Assert(cRefs > 1 && cRefs < _1M);
1461	}
1462
1463	RTSpinlockRelease(pDevExt->hSessionHashTabSpinlock);
1464
1465	return pCur;
1466	}
1467
1468
1469	/**
1470	* Retain a session to make sure it doesn't go away while it is in use.
1471	*
1472	* @returns New reference count on success, UINT32_MAX on failure.
1473	* @param pSession Session data.
1474	*/
1475	uint32_t VBOXCALL supdrvSessionRetain(PSUPDRVSESSION pSession)
1476	{
1477	uint32_t cRefs;
1478	AssertPtrReturn(pSession, UINT32_MAX);
1479	AssertReturn(SUP_IS_SESSION_VALID(pSession), UINT32_MAX);
1480
1481	cRefs = ASMAtomicIncU32(&pSession->cRefs);
1482	AssertMsg(cRefs > 1 && cRefs < _1M, ("%#x %p\n", cRefs, pSession));
1483	return cRefs;
1484	}
1485
1486
1487	/**
1488	* Releases a given session.
1489	*
1490	* @returns New reference count on success (0 if closed), UINT32_MAX on failure.
1491	* @param pSession Session data.
1492	*/
1493	uint32_t VBOXCALL supdrvSessionRelease(PSUPDRVSESSION pSession)
1494	{
1495	uint32_t cRefs;
1496	AssertPtrReturn(pSession, UINT32_MAX);
1497	AssertReturn(SUP_IS_SESSION_VALID(pSession), UINT32_MAX);
1498
1499	cRefs = ASMAtomicDecU32(&pSession->cRefs);
1500	AssertMsg(cRefs < _1M, ("%#x %p\n", cRefs, pSession));
1501	if (cRefs == 0)
1502	supdrvDestroySession(pSession->pDevExt, pSession);
1503	return cRefs;
1504	}
1505
1506
1507	/**
1508	* RTHandleTableDestroy callback used by supdrvCleanupSession.
1509	*
1510	* @returns IPRT status code, see SUPR0ObjAddRef.
1511	* @param hHandleTable The handle table handle. Ignored.
1512	* @param pvObj The object pointer.
1513	* @param pvCtx Context, the handle type. Ignored.
1514	* @param pvUser Session pointer.
1515	*/
1516	static DECLCALLBACK(int) supdrvSessionObjHandleRetain(RTHANDLETABLE hHandleTable, void pvObj, void pvCtx, void *pvUser)
1517	{
1518	NOREF(pvCtx);
1519	NOREF(hHandleTable);
1520	return SUPR0ObjAddRefEx(pvObj, (PSUPDRVSESSION)pvUser, true /fNoBlocking/);
1521	}
1522
1523
1524	/**
1525	* RTHandleTableDestroy callback used by supdrvCleanupSession.
1526	*
1527	* @param hHandleTable The handle table handle. Ignored.
1528	* @param h The handle value. Ignored.
1529	* @param pvObj The object pointer.
1530	* @param pvCtx Context, the handle type. Ignored.
1531	* @param pvUser Session pointer.
1532	*/
1533	static DECLCALLBACK(void) supdrvSessionObjHandleDelete(RTHANDLETABLE hHandleTable, uint32_t h, void pvObj, void pvCtx, void *pvUser)
1534	{
1535	NOREF(pvCtx);
1536	NOREF(h);
1537	NOREF(hHandleTable);
1538	SUPR0ObjRelease(pvObj, (PSUPDRVSESSION)pvUser);
1539	}
1540
1541
1542	/**
1543	* Fast path I/O Control worker.
1544	*
1545	* @returns VBox status code that should be passed down to ring-3 unchanged.
1546	* @param uOperation SUP_VMMR0_DO_XXX (not the I/O control number!).
1547	* @param idCpu VMCPU id.
1548	* @param pDevExt Device extention.
1549	* @param pSession Session data.
1550	*/
1551	int VBOXCALL supdrvIOCtlFast(uintptr_t uOperation, VMCPUID idCpu, PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession)
1552	{
1553	/*
1554	* Validate input and check that the VM has a session.
1555	*/
1556	if (RT_LIKELY(RT_VALID_PTR(pSession)))
1557	{
1558	PVM pVM = pSession->pSessionVM;
1559	PGVM pGVM = pSession->pSessionGVM;
1560	if (RT_LIKELY( pGVM != NULL
1561	&& pVM != NULL
1562	&& pVM == pSession->pFastIoCtrlVM))
1563	{
1564	if (RT_LIKELY(pDevExt->pfnVMMR0EntryFast))
1565	{
1566	/*
1567	* Make the call.
1568	*/
1569	pDevExt->pfnVMMR0EntryFast(pGVM, pVM, idCpu, uOperation);
1570	return VINF_SUCCESS;
1571	}
1572
1573	SUPR0Printf("supdrvIOCtlFast: pfnVMMR0EntryFast is NULL\n");
1574	}
1575	else
1576	SUPR0Printf("supdrvIOCtlFast: Misconfig session: pGVM=%p pVM=%p pFastIoCtrlVM=%p\n",
1577	pGVM, pVM, pSession->pFastIoCtrlVM);
1578	}
1579	else
1580	SUPR0Printf("supdrvIOCtlFast: Bad session pointer %p\n", pSession);
1581	return VERR_INTERNAL_ERROR;
1582	}
1583
1584
1585	/**
1586	* Helper for supdrvIOCtl used to validate module names passed to SUP_IOCTL_LDR_OPEN.
1587	*
1588	* Check if pszStr contains any character of pszChars. We would use strpbrk
1589	* here if this function would be contained in the RedHat kABI white list, see
1590	* http://www.kerneldrivers.org/RHEL5.
1591	*
1592	* @returns true if fine, false if not.
1593	* @param pszName The module name to check.
1594	*/
1595	static bool supdrvIsLdrModuleNameValid(const char *pszName)
1596	{
1597	int chCur;
1598	while ((chCur = *pszName++) != '\0')
1599	{
1600	static const char s_szInvalidChars[] = ";:()[]{}/\\\|&*%#@!~`\"'";
1601	unsigned offInv = RT_ELEMENTS(s_szInvalidChars);
1602	while (offInv-- > 0)
1603	if (s_szInvalidChars[offInv] == chCur)
1604	return false;
1605	}
1606	return true;
1607	}
1608
1609
1610
1611	/**
1612	* I/O Control inner worker (tracing reasons).
1613	*
1614	* @returns IPRT status code.
1615	* @retval VERR_INVALID_PARAMETER if the request is invalid.
1616	*
1617	* @param uIOCtl Function number.
1618	* @param pDevExt Device extention.
1619	* @param pSession Session data.
1620	* @param pReqHdr The request header.
1621	*/
1622	static int supdrvIOCtlInnerUnrestricted(uintptr_t uIOCtl, PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPREQHDR pReqHdr)
1623	{
1624	/*
1625	* Validation macros
1626	*/
1627	#define REQ_CHECK_SIZES_EX(Name, cbInExpect, cbOutExpect) \
1628	do { \
1629	if (RT_UNLIKELY(pReqHdr->cbIn != (cbInExpect) \|\| pReqHdr->cbOut != (cbOutExpect))) \
1630	{ \
1631	OSDBGPRINT(( #Name ": Invalid input/output sizes. cbIn=%ld expected %ld. cbOut=%ld expected %ld.\n", \
1632	(long)pReqHdr->cbIn, (long)(cbInExpect), (long)pReqHdr->cbOut, (long)(cbOutExpect))); \
1633	return pReqHdr->rc = VERR_INVALID_PARAMETER; \
1634	} \
1635	} while (0)
1636
1637	#define REQ_CHECK_SIZES(Name) REQ_CHECK_SIZES_EX(Name, Name ## _SIZE_IN, Name ## _SIZE_OUT)
1638
1639	#define REQ_CHECK_SIZE_IN(Name, cbInExpect) \
1640	do { \
1641	if (RT_UNLIKELY(pReqHdr->cbIn != (cbInExpect))) \
1642	{ \
1643	OSDBGPRINT(( #Name ": Invalid input/output sizes. cbIn=%ld expected %ld.\n", \
1644	(long)pReqHdr->cbIn, (long)(cbInExpect))); \
1645	return pReqHdr->rc = VERR_INVALID_PARAMETER; \
1646	} \
1647	} while (0)
1648
1649	#define REQ_CHECK_SIZE_OUT(Name, cbOutExpect) \
1650	do { \
1651	if (RT_UNLIKELY(pReqHdr->cbOut != (cbOutExpect))) \
1652	{ \
1653	OSDBGPRINT(( #Name ": Invalid input/output sizes. cbOut=%ld expected %ld.\n", \
1654	(long)pReqHdr->cbOut, (long)(cbOutExpect))); \
1655	return pReqHdr->rc = VERR_INVALID_PARAMETER; \
1656	} \
1657	} while (0)
1658
1659	#define REQ_CHECK_EXPR(Name, expr) \
1660	do { \
1661	if (RT_UNLIKELY(!(expr))) \
1662	{ \
1663	OSDBGPRINT(( #Name ": %s\n", #expr)); \
1664	return pReqHdr->rc = VERR_INVALID_PARAMETER; \
1665	} \
1666	} while (0)
1667
1668	#define REQ_CHECK_EXPR_FMT(expr, fmt) \
1669	do { \
1670	if (RT_UNLIKELY(!(expr))) \
1671	{ \
1672	OSDBGPRINT( fmt ); \
1673	return pReqHdr->rc = VERR_INVALID_PARAMETER; \
1674	} \
1675	} while (0)
1676
1677	/*
1678	* The switch.
1679	*/
1680	switch (SUP_CTL_CODE_NO_SIZE(uIOCtl))
1681	{
1682	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_COOKIE):
1683	{
1684	PSUPCOOKIE pReq = (PSUPCOOKIE)pReqHdr;
1685	REQ_CHECK_SIZES(SUP_IOCTL_COOKIE);
1686	if (strncmp(pReq->u.In.szMagic, SUPCOOKIE_MAGIC, sizeof(pReq->u.In.szMagic)))
1687	{
1688	OSDBGPRINT(("SUP_IOCTL_COOKIE: invalid magic %.16s\n", pReq->u.In.szMagic));
1689	pReq->Hdr.rc = VERR_INVALID_MAGIC;
1690	return 0;
1691	}
1692
1693	#if 0
1694	/*
1695	* Call out to the OS specific code and let it do permission checks on the
1696	* client process.
1697	*/
1698	if (!supdrvOSValidateClientProcess(pDevExt, pSession))
1699	{
1700	pReq->u.Out.u32Cookie = 0xffffffff;
1701	pReq->u.Out.u32SessionCookie = 0xffffffff;
1702	pReq->u.Out.u32SessionVersion = 0xffffffff;
1703	pReq->u.Out.u32DriverVersion = SUPDRV_IOC_VERSION;
1704	pReq->u.Out.pSession = NULL;
1705	pReq->u.Out.cFunctions = 0;
1706	pReq->Hdr.rc = VERR_PERMISSION_DENIED;
1707	return 0;
1708	}
1709	#endif
1710
1711	/*
1712	* Match the version.
1713	* The current logic is very simple, match the major interface version.
1714	*/
1715	if ( pReq->u.In.u32MinVersion > SUPDRV_IOC_VERSION
1716	\|\| (pReq->u.In.u32MinVersion & 0xffff0000) != (SUPDRV_IOC_VERSION & 0xffff0000))
1717	{
1718	OSDBGPRINT(("SUP_IOCTL_COOKIE: Version mismatch. Requested: %#x Min: %#x Current: %#x\n",
1719	pReq->u.In.u32ReqVersion, pReq->u.In.u32MinVersion, SUPDRV_IOC_VERSION));
1720	pReq->u.Out.u32Cookie = 0xffffffff;
1721	pReq->u.Out.u32SessionCookie = 0xffffffff;
1722	pReq->u.Out.u32SessionVersion = 0xffffffff;
1723	pReq->u.Out.u32DriverVersion = SUPDRV_IOC_VERSION;
1724	pReq->u.Out.pSession = NULL;
1725	pReq->u.Out.cFunctions = 0;
1726	pReq->Hdr.rc = VERR_VERSION_MISMATCH;
1727	return 0;
1728	}
1729
1730	/*
1731	* Fill in return data and be gone.
1732	* N.B. The first one to change SUPDRV_IOC_VERSION shall makes sure that
1733	* u32SessionVersion <= u32ReqVersion!
1734	*/
1735	/** @todo Somehow validate the client and negotiate a secure cookie... */
1736	pReq->u.Out.u32Cookie = pDevExt->u32Cookie;
1737	pReq->u.Out.u32SessionCookie = pSession->u32Cookie;
1738	pReq->u.Out.u32SessionVersion = SUPDRV_IOC_VERSION;
1739	pReq->u.Out.u32DriverVersion = SUPDRV_IOC_VERSION;
1740	pReq->u.Out.pSession = pSession;
1741	pReq->u.Out.cFunctions = sizeof(g_aFunctions) / sizeof(g_aFunctions[0]);
1742	pReq->Hdr.rc = VINF_SUCCESS;
1743	return 0;
1744	}
1745
1746	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_QUERY_FUNCS(0)):
1747	{
1748	/* validate */
1749	PSUPQUERYFUNCS pReq = (PSUPQUERYFUNCS)pReqHdr;
1750	REQ_CHECK_SIZES_EX(SUP_IOCTL_QUERY_FUNCS, SUP_IOCTL_QUERY_FUNCS_SIZE_IN, SUP_IOCTL_QUERY_FUNCS_SIZE_OUT(RT_ELEMENTS(g_aFunctions)));
1751
1752	/* execute */
1753	pReq->u.Out.cFunctions = RT_ELEMENTS(g_aFunctions);
1754	memcpy(&pReq->u.Out.aFunctions[0], g_aFunctions, sizeof(g_aFunctions));
1755	pReq->Hdr.rc = VINF_SUCCESS;
1756	return 0;
1757	}
1758
1759	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_LOCK):
1760	{
1761	/* validate */
1762	PSUPPAGELOCK pReq = (PSUPPAGELOCK)pReqHdr;
1763	REQ_CHECK_SIZE_IN(SUP_IOCTL_PAGE_LOCK, SUP_IOCTL_PAGE_LOCK_SIZE_IN);
1764	REQ_CHECK_SIZE_OUT(SUP_IOCTL_PAGE_LOCK, SUP_IOCTL_PAGE_LOCK_SIZE_OUT(pReq->u.In.cPages));
1765	REQ_CHECK_EXPR(SUP_IOCTL_PAGE_LOCK, pReq->u.In.cPages > 0);
1766	REQ_CHECK_EXPR(SUP_IOCTL_PAGE_LOCK, pReq->u.In.pvR3 >= PAGE_SIZE);
1767
1768	/* execute */
1769	pReq->Hdr.rc = SUPR0LockMem(pSession, pReq->u.In.pvR3, pReq->u.In.cPages, &pReq->u.Out.aPages[0]);
1770	if (RT_FAILURE(pReq->Hdr.rc))
1771	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1772	return 0;
1773	}
1774
1775	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_UNLOCK):
1776	{
1777	/* validate */
1778	PSUPPAGEUNLOCK pReq = (PSUPPAGEUNLOCK)pReqHdr;
1779	REQ_CHECK_SIZES(SUP_IOCTL_PAGE_UNLOCK);
1780
1781	/* execute */
1782	pReq->Hdr.rc = SUPR0UnlockMem(pSession, pReq->u.In.pvR3);
1783	return 0;
1784	}
1785
1786	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_CONT_ALLOC):
1787	{
1788	/* validate */
1789	PSUPCONTALLOC pReq = (PSUPCONTALLOC)pReqHdr;
1790	REQ_CHECK_SIZES(SUP_IOCTL_CONT_ALLOC);
1791
1792	/* execute */
1793	pReq->Hdr.rc = SUPR0ContAlloc(pSession, pReq->u.In.cPages, &pReq->u.Out.pvR0, &pReq->u.Out.pvR3, &pReq->u.Out.HCPhys);
1794	if (RT_FAILURE(pReq->Hdr.rc))
1795	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1796	return 0;
1797	}
1798
1799	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_CONT_FREE):
1800	{
1801	/* validate */
1802	PSUPCONTFREE pReq = (PSUPCONTFREE)pReqHdr;
1803	REQ_CHECK_SIZES(SUP_IOCTL_CONT_FREE);
1804
1805	/* execute */
1806	pReq->Hdr.rc = SUPR0ContFree(pSession, (RTHCUINTPTR)pReq->u.In.pvR3);
1807	return 0;
1808	}
1809
1810	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LDR_OPEN):
1811	{
1812	/* validate */
1813	PSUPLDROPEN pReq = (PSUPLDROPEN)pReqHdr;
1814	REQ_CHECK_SIZES(SUP_IOCTL_LDR_OPEN);
1815	if ( pReq->u.In.cbImageWithEverything != 0
1816	\|\| pReq->u.In.cbImageBits != 0)
1817	{
1818	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.cbImageWithEverything > 0);
1819	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.cbImageWithEverything < 16*_1M);
1820	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.cbImageBits > 0);
1821	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.cbImageBits < pReq->u.In.cbImageWithEverything);
1822	}
1823	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.szName[0]);
1824	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, RTStrEnd(pReq->u.In.szName, sizeof(pReq->u.In.szName)));
1825	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, supdrvIsLdrModuleNameValid(pReq->u.In.szName));
1826	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, RTStrEnd(pReq->u.In.szFilename, sizeof(pReq->u.In.szFilename)));
1827
1828	/* execute */
1829	pReq->Hdr.rc = supdrvIOCtl_LdrOpen(pDevExt, pSession, pReq);
1830	return 0;
1831	}
1832
1833	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LDR_LOAD):
1834	{
1835	/* validate */
1836	PSUPLDRLOAD pReq = (PSUPLDRLOAD)pReqHdr;
1837	REQ_CHECK_EXPR(Name, pReq->Hdr.cbIn >= SUP_IOCTL_LDR_LOAD_SIZE_IN(32));
1838	REQ_CHECK_SIZES_EX(SUP_IOCTL_LDR_LOAD, SUP_IOCTL_LDR_LOAD_SIZE_IN(pReq->u.In.cbImageWithEverything), SUP_IOCTL_LDR_LOAD_SIZE_OUT);
1839	REQ_CHECK_EXPR_FMT( !pReq->u.In.cSymbols
1840	\|\| ( pReq->u.In.cSymbols <= 16384
1841	&& pReq->u.In.offSymbols >= pReq->u.In.cbImageBits
1842	&& pReq->u.In.offSymbols < pReq->u.In.cbImageWithEverything
1843	&& pReq->u.In.offSymbols + pReq->u.In.cSymbols * sizeof(SUPLDRSYM) <= pReq->u.In.cbImageWithEverything),
1844	("SUP_IOCTL_LDR_LOAD: offSymbols=%#lx cSymbols=%#lx cbImageWithEverything=%#lx\n", (long)pReq->u.In.offSymbols,
1845	(long)pReq->u.In.cSymbols, (long)pReq->u.In.cbImageWithEverything));
1846	REQ_CHECK_EXPR_FMT( !pReq->u.In.cbStrTab
1847	\|\| ( pReq->u.In.offStrTab < pReq->u.In.cbImageWithEverything
1848	&& pReq->u.In.offStrTab >= pReq->u.In.cbImageBits
1849	&& pReq->u.In.offStrTab + pReq->u.In.cbStrTab <= pReq->u.In.cbImageWithEverything
1850	&& pReq->u.In.cbStrTab <= pReq->u.In.cbImageWithEverything),
1851	("SUP_IOCTL_LDR_LOAD: offStrTab=%#lx cbStrTab=%#lx cbImageWithEverything=%#lx\n", (long)pReq->u.In.offStrTab,
1852	(long)pReq->u.In.cbStrTab, (long)pReq->u.In.cbImageWithEverything));
1853	REQ_CHECK_EXPR_FMT( pReq->u.In.cSegments >= 1
1854	&& pReq->u.In.cSegments <= 128
1855	&& pReq->u.In.cSegments <= (pReq->u.In.cbImageBits + PAGE_SIZE - 1) / PAGE_SIZE
1856	&& pReq->u.In.offSegments >= pReq->u.In.cbImageBits
1857	&& pReq->u.In.offSegments < pReq->u.In.cbImageWithEverything
1858	&& pReq->u.In.offSegments + pReq->u.In.cSegments * sizeof(SUPLDRSEG) <= pReq->u.In.cbImageWithEverything,
1859	("SUP_IOCTL_LDR_LOAD: offSegments=%#lx cSegments=%#lx cbImageWithEverything=%#lx\n", (long)pReq->u.In.offSegments,
1860	(long)pReq->u.In.cSegments, (long)pReq->u.In.cbImageWithEverything));
1861
1862	if (pReq->u.In.cSymbols)
1863	{
1864	uint32_t i;
1865	PSUPLDRSYM paSyms = (PSUPLDRSYM)&pReq->u.In.abImage[pReq->u.In.offSymbols];
1866	for (i = 0; i < pReq->u.In.cSymbols; i++)
1867	{
1868	REQ_CHECK_EXPR_FMT(paSyms[i].offSymbol < pReq->u.In.cbImageWithEverything,
1869	("SUP_IOCTL_LDR_LOAD: sym #%ld: symb off %#lx (max=%#lx)\n", (long)i, (long)paSyms[i].offSymbol, (long)pReq->u.In.cbImageWithEverything));
1870	REQ_CHECK_EXPR_FMT(paSyms[i].offName < pReq->u.In.cbStrTab,
1871	("SUP_IOCTL_LDR_LOAD: sym #%ld: name off %#lx (max=%#lx)\n", (long)i, (long)paSyms[i].offName, (long)pReq->u.In.cbImageWithEverything));
1872	REQ_CHECK_EXPR_FMT(RTStrEnd((char const *)&pReq->u.In.abImage[pReq->u.In.offStrTab + paSyms[i].offName],
1873	pReq->u.In.cbStrTab - paSyms[i].offName),
1874	("SUP_IOCTL_LDR_LOAD: sym #%ld: unterminated name! (%#lx / %#lx)\n", (long)i, (long)paSyms[i].offName, (long)pReq->u.In.cbImageWithEverything));
1875	}
1876	}
1877	{
1878	uint32_t i;
1879	uint32_t offPrevEnd = 0;
1880	PSUPLDRSEG paSegs = (PSUPLDRSEG)&pReq->u.In.abImage[pReq->u.In.offSegments];
1881	for (i = 0; i < pReq->u.In.cSegments; i++)
1882	{
1883	REQ_CHECK_EXPR_FMT(paSegs[i].off < pReq->u.In.cbImageBits && !(paSegs[i].off & PAGE_OFFSET_MASK),
1884	("SUP_IOCTL_LDR_LOAD: seg #%ld: off %#lx (max=%#lx)\n", (long)i, (long)paSegs[i].off, (long)pReq->u.In.cbImageBits));
1885	REQ_CHECK_EXPR_FMT(paSegs[i].cb <= pReq->u.In.cbImageBits,
1886	("SUP_IOCTL_LDR_LOAD: seg #%ld: cb %#lx (max=%#lx)\n", (long)i, (long)paSegs[i].cb, (long)pReq->u.In.cbImageBits));
1887	REQ_CHECK_EXPR_FMT(paSegs[i].off + paSegs[i].cb <= pReq->u.In.cbImageBits,
1888	("SUP_IOCTL_LDR_LOAD: seg #%ld: off %#lx + cb %#lx = %#lx (max=%#lx)\n", (long)i, (long)paSegs[i].off, (long)paSegs[i].cb, (long)(paSegs[i].off + paSegs[i].cb), (long)pReq->u.In.cbImageBits));
1889	REQ_CHECK_EXPR_FMT(paSegs[i].fProt != 0,
1890	("SUP_IOCTL_LDR_LOAD: seg #%ld: off %#lx + cb %#lx\n", (long)i, (long)paSegs[i].off, (long)paSegs[i].cb));
1891	REQ_CHECK_EXPR_FMT(paSegs[i].fUnused == 0, ("SUP_IOCTL_LDR_LOAD: seg #%ld: fUnused=1\n", (long)i));
1892	REQ_CHECK_EXPR_FMT(offPrevEnd == paSegs[i].off,
1893	("SUP_IOCTL_LDR_LOAD: seg #%ld: off %#lx offPrevEnd %#lx\n", (long)i, (long)paSegs[i].off, (long)offPrevEnd));
1894	offPrevEnd = paSegs[i].off + paSegs[i].cb;
1895	}
1896	REQ_CHECK_EXPR_FMT(offPrevEnd == pReq->u.In.cbImageBits,
1897	("SUP_IOCTL_LDR_LOAD: offPrevEnd %#lx cbImageBits %#lx\n", (long)i, (long)offPrevEnd, (long)pReq->u.In.cbImageBits));
1898	}
1899	REQ_CHECK_EXPR_FMT(!(pReq->u.In.fFlags & ~SUPLDRLOAD_F_VALID_MASK),
1900	("SUP_IOCTL_LDR_LOAD: fFlags=%#x\n", (unsigned)pReq->u.In.fFlags));
1901
1902	/* execute */
1903	pReq->Hdr.rc = supdrvIOCtl_LdrLoad(pDevExt, pSession, pReq);
1904	return 0;
1905	}
1906
1907	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LDR_FREE):
1908	{
1909	/* validate */
1910	PSUPLDRFREE pReq = (PSUPLDRFREE)pReqHdr;
1911	REQ_CHECK_SIZES(SUP_IOCTL_LDR_FREE);
1912
1913	/* execute */
1914	pReq->Hdr.rc = supdrvIOCtl_LdrFree(pDevExt, pSession, pReq);
1915	return 0;
1916	}
1917
1918	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LDR_LOCK_DOWN):
1919	{
1920	/* validate */
1921	REQ_CHECK_SIZES(SUP_IOCTL_LDR_LOCK_DOWN);
1922
1923	/* execute */
1924	pReqHdr->rc = supdrvIOCtl_LdrLockDown(pDevExt);
1925	return 0;
1926	}
1927
1928	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LDR_GET_SYMBOL):
1929	{
1930	/* validate */
1931	PSUPLDRGETSYMBOL pReq = (PSUPLDRGETSYMBOL)pReqHdr;
1932	REQ_CHECK_SIZES(SUP_IOCTL_LDR_GET_SYMBOL);
1933	REQ_CHECK_EXPR(SUP_IOCTL_LDR_GET_SYMBOL, RTStrEnd(pReq->u.In.szSymbol, sizeof(pReq->u.In.szSymbol)));
1934
1935	/* execute */
1936	pReq->Hdr.rc = supdrvIOCtl_LdrQuerySymbol(pDevExt, pSession, pReq);
1937	return 0;
1938	}
1939
1940	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_CALL_VMMR0_NO_SIZE()):
1941	{
1942	/* validate */
1943	PSUPCALLVMMR0 pReq = (PSUPCALLVMMR0)pReqHdr;
1944	Log4(("SUP_IOCTL_CALL_VMMR0: op=%u in=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1945	pReq->u.In.uOperation, pReq->Hdr.cbIn, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1946
1947	if (pReq->Hdr.cbIn == SUP_IOCTL_CALL_VMMR0_SIZE(0))
1948	{
1949	REQ_CHECK_SIZES_EX(SUP_IOCTL_CALL_VMMR0, SUP_IOCTL_CALL_VMMR0_SIZE_IN(0), SUP_IOCTL_CALL_VMMR0_SIZE_OUT(0));
1950
1951	/* execute */
1952	if (RT_LIKELY(pDevExt->pfnVMMR0EntryEx))
1953	{
1954	if (pReq->u.In.pVMR0 == NULL)
1955	pReq->Hdr.rc = pDevExt->pfnVMMR0EntryEx(NULL, NULL, pReq->u.In.idCpu,
1956	pReq->u.In.uOperation, NULL, pReq->u.In.u64Arg, pSession);
1957	else if (pReq->u.In.pVMR0 == pSession->pSessionVM)
1958	pReq->Hdr.rc = pDevExt->pfnVMMR0EntryEx(pSession->pSessionGVM, pSession->pSessionVM, pReq->u.In.idCpu,
1959	pReq->u.In.uOperation, NULL, pReq->u.In.u64Arg, pSession);
1960	else
1961	pReq->Hdr.rc = VERR_INVALID_VM_HANDLE;
1962	}
1963	else
1964	pReq->Hdr.rc = VERR_WRONG_ORDER;
1965	}
1966	else
1967	{
1968	PSUPVMMR0REQHDR pVMMReq = (PSUPVMMR0REQHDR)&pReq->abReqPkt[0];
1969	REQ_CHECK_EXPR_FMT(pReq->Hdr.cbIn >= SUP_IOCTL_CALL_VMMR0_SIZE(sizeof(SUPVMMR0REQHDR)),
1970	("SUP_IOCTL_CALL_VMMR0: cbIn=%#x < %#lx\n", pReq->Hdr.cbIn, SUP_IOCTL_CALL_VMMR0_SIZE(sizeof(SUPVMMR0REQHDR))));
1971	REQ_CHECK_EXPR(SUP_IOCTL_CALL_VMMR0, pVMMReq->u32Magic == SUPVMMR0REQHDR_MAGIC);
1972	REQ_CHECK_SIZES_EX(SUP_IOCTL_CALL_VMMR0, SUP_IOCTL_CALL_VMMR0_SIZE_IN(pVMMReq->cbReq), SUP_IOCTL_CALL_VMMR0_SIZE_OUT(pVMMReq->cbReq));
1973
1974	/* execute */
1975	if (RT_LIKELY(pDevExt->pfnVMMR0EntryEx))
1976	{
1977	if (pReq->u.In.pVMR0 == NULL)
1978	pReq->Hdr.rc = pDevExt->pfnVMMR0EntryEx(NULL, NULL, pReq->u.In.idCpu,
1979	pReq->u.In.uOperation, pVMMReq, pReq->u.In.u64Arg, pSession);
1980	else if (pReq->u.In.pVMR0 == pSession->pSessionVM)
1981	pReq->Hdr.rc = pDevExt->pfnVMMR0EntryEx(pSession->pSessionGVM, pSession->pSessionVM, pReq->u.In.idCpu,
1982	pReq->u.In.uOperation, pVMMReq, pReq->u.In.u64Arg, pSession);
1983	else
1984	pReq->Hdr.rc = VERR_INVALID_VM_HANDLE;
1985	}
1986	else
1987	pReq->Hdr.rc = VERR_WRONG_ORDER;
1988	}
1989
1990	if ( RT_FAILURE(pReq->Hdr.rc)
1991	&& pReq->Hdr.rc != VERR_INTERRUPTED
1992	&& pReq->Hdr.rc != VERR_TIMEOUT)
1993	Log(("SUP_IOCTL_CALL_VMMR0: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1994	pReq->Hdr.rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1995	else
1996	Log4(("SUP_IOCTL_CALL_VMMR0: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1997	pReq->Hdr.rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1998	return 0;
1999	}
2000
2001	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_CALL_VMMR0_BIG):
2002	{
2003	/* validate */
2004	PSUPCALLVMMR0 pReq = (PSUPCALLVMMR0)pReqHdr;
2005	PSUPVMMR0REQHDR pVMMReq;
2006	Log4(("SUP_IOCTL_CALL_VMMR0_BIG: op=%u in=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
2007	pReq->u.In.uOperation, pReq->Hdr.cbIn, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
2008
2009	pVMMReq = (PSUPVMMR0REQHDR)&pReq->abReqPkt[0];
2010	REQ_CHECK_EXPR_FMT(pReq->Hdr.cbIn >= SUP_IOCTL_CALL_VMMR0_BIG_SIZE(sizeof(SUPVMMR0REQHDR)),
2011	("SUP_IOCTL_CALL_VMMR0_BIG: cbIn=%#x < %#lx\n", pReq->Hdr.cbIn, SUP_IOCTL_CALL_VMMR0_BIG_SIZE(sizeof(SUPVMMR0REQHDR))));
2012	REQ_CHECK_EXPR(SUP_IOCTL_CALL_VMMR0_BIG, pVMMReq->u32Magic == SUPVMMR0REQHDR_MAGIC);
2013	REQ_CHECK_SIZES_EX(SUP_IOCTL_CALL_VMMR0_BIG, SUP_IOCTL_CALL_VMMR0_BIG_SIZE_IN(pVMMReq->cbReq), SUP_IOCTL_CALL_VMMR0_BIG_SIZE_OUT(pVMMReq->cbReq));
2014
2015	/* execute */
2016	if (RT_LIKELY(pDevExt->pfnVMMR0EntryEx))
2017	{
2018	if (pReq->u.In.pVMR0 == NULL)
2019	pReq->Hdr.rc = pDevExt->pfnVMMR0EntryEx(NULL, NULL, pReq->u.In.idCpu, pReq->u.In.uOperation, pVMMReq, pReq->u.In.u64Arg, pSession);
2020	else if (pReq->u.In.pVMR0 == pSession->pSessionVM)
2021	pReq->Hdr.rc = pDevExt->pfnVMMR0EntryEx(pSession->pSessionGVM, pSession->pSessionVM, pReq->u.In.idCpu,
2022	pReq->u.In.uOperation, pVMMReq, pReq->u.In.u64Arg, pSession);
2023	else
2024	pReq->Hdr.rc = VERR_INVALID_VM_HANDLE;
2025	}
2026	else
2027	pReq->Hdr.rc = VERR_WRONG_ORDER;
2028
2029	if ( RT_FAILURE(pReq->Hdr.rc)
2030	&& pReq->Hdr.rc != VERR_INTERRUPTED
2031	&& pReq->Hdr.rc != VERR_TIMEOUT)
2032	Log(("SUP_IOCTL_CALL_VMMR0_BIG: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
2033	pReq->Hdr.rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
2034	else
2035	Log4(("SUP_IOCTL_CALL_VMMR0_BIG: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
2036	pReq->Hdr.rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
2037	return 0;
2038	}
2039
2040	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_GET_PAGING_MODE):
2041	{
2042	/* validate */
2043	PSUPGETPAGINGMODE pReq = (PSUPGETPAGINGMODE)pReqHdr;
2044	REQ_CHECK_SIZES(SUP_IOCTL_GET_PAGING_MODE);
2045
2046	/* execute */
2047	pReq->Hdr.rc = VINF_SUCCESS;
2048	pReq->u.Out.enmMode = SUPR0GetPagingMode();
2049	return 0;
2050	}
2051
2052	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LOW_ALLOC):
2053	{
2054	/* validate */
2055	PSUPLOWALLOC pReq = (PSUPLOWALLOC)pReqHdr;
2056	REQ_CHECK_EXPR(SUP_IOCTL_LOW_ALLOC, pReq->Hdr.cbIn <= SUP_IOCTL_LOW_ALLOC_SIZE_IN);
2057	REQ_CHECK_SIZES_EX(SUP_IOCTL_LOW_ALLOC, SUP_IOCTL_LOW_ALLOC_SIZE_IN, SUP_IOCTL_LOW_ALLOC_SIZE_OUT(pReq->u.In.cPages));
2058
2059	/* execute */
2060	pReq->Hdr.rc = SUPR0LowAlloc(pSession, pReq->u.In.cPages, &pReq->u.Out.pvR0, &pReq->u.Out.pvR3, &pReq->u.Out.aPages[0]);
2061	if (RT_FAILURE(pReq->Hdr.rc))
2062	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
2063	return 0;
2064	}
2065
2066	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LOW_FREE):
2067	{
2068	/* validate */
2069	PSUPLOWFREE pReq = (PSUPLOWFREE)pReqHdr;
2070	REQ_CHECK_SIZES(SUP_IOCTL_LOW_FREE);
2071
2072	/* execute */
2073	pReq->Hdr.rc = SUPR0LowFree(pSession, (RTHCUINTPTR)pReq->u.In.pvR3);
2074	return 0;
2075	}
2076
2077	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_GIP_MAP):
2078	{
2079	/* validate */
2080	PSUPGIPMAP pReq = (PSUPGIPMAP)pReqHdr;
2081	REQ_CHECK_SIZES(SUP_IOCTL_GIP_MAP);
2082
2083	/* execute */
2084	pReq->Hdr.rc = SUPR0GipMap(pSession, &pReq->u.Out.pGipR3, &pReq->u.Out.HCPhysGip);
2085	if (RT_SUCCESS(pReq->Hdr.rc))
2086	pReq->u.Out.pGipR0 = pDevExt->pGip;
2087	return 0;
2088	}
2089
2090	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_GIP_UNMAP):
2091	{
2092	/* validate */
2093	PSUPGIPUNMAP pReq = (PSUPGIPUNMAP)pReqHdr;
2094	REQ_CHECK_SIZES(SUP_IOCTL_GIP_UNMAP);
2095
2096	/* execute */
2097	pReq->Hdr.rc = SUPR0GipUnmap(pSession);
2098	return 0;
2099	}
2100
2101	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_SET_VM_FOR_FAST):
2102	{
2103	/* validate */
2104	PSUPSETVMFORFAST pReq = (PSUPSETVMFORFAST)pReqHdr;
2105	REQ_CHECK_SIZES(SUP_IOCTL_SET_VM_FOR_FAST);
2106	REQ_CHECK_EXPR_FMT( !pReq->u.In.pVMR0
2107	\|\| ( RT_VALID_PTR(pReq->u.In.pVMR0)
2108	&& !((uintptr_t)pReq->u.In.pVMR0 & (PAGE_SIZE - 1))),
2109	("SUP_IOCTL_SET_VM_FOR_FAST: pVMR0=%p!\n", pReq->u.In.pVMR0));
2110
2111	/* execute */
2112	RTSpinlockAcquire(pDevExt->Spinlock);
2113	if (pSession->pSessionVM == pReq->u.In.pVMR0)
2114	{
2115	if (pSession->pFastIoCtrlVM == NULL)
2116	{
2117	pSession->pFastIoCtrlVM = pSession->pSessionVM;
2118	RTSpinlockRelease(pDevExt->Spinlock);
2119	pReq->Hdr.rc = VINF_SUCCESS;
2120	}
2121	else
2122	{
2123	RTSpinlockRelease(pDevExt->Spinlock);
2124	OSDBGPRINT(("SUP_IOCTL_SET_VM_FOR_FAST: pSession->pFastIoCtrlVM=%p! (pVMR0=%p)\n",
2125	pSession->pFastIoCtrlVM, pReq->u.In.pVMR0));
2126	pReq->Hdr.rc = VERR_ALREADY_EXISTS;
2127	}
2128	}
2129	else
2130	{
2131	RTSpinlockRelease(pDevExt->Spinlock);
2132	OSDBGPRINT(("SUP_IOCTL_SET_VM_FOR_FAST: pSession->pSessionVM=%p vs pVMR0=%p)\n",
2133	pSession->pSessionVM, pReq->u.In.pVMR0));
2134	pReq->Hdr.rc = pSession->pSessionVM ? VERR_ACCESS_DENIED : VERR_WRONG_ORDER;
2135	}
2136	return 0;
2137	}
2138
2139	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_ALLOC_EX):
2140	{
2141	/* validate */
2142	PSUPPAGEALLOCEX pReq = (PSUPPAGEALLOCEX)pReqHdr;
2143	REQ_CHECK_EXPR(SUP_IOCTL_PAGE_ALLOC_EX, pReq->Hdr.cbIn <= SUP_IOCTL_PAGE_ALLOC_EX_SIZE_IN);
2144	REQ_CHECK_SIZES_EX(SUP_IOCTL_PAGE_ALLOC_EX, SUP_IOCTL_PAGE_ALLOC_EX_SIZE_IN, SUP_IOCTL_PAGE_ALLOC_EX_SIZE_OUT(pReq->u.In.cPages));
2145	REQ_CHECK_EXPR_FMT(pReq->u.In.fKernelMapping \|\| pReq->u.In.fUserMapping,
2146	("SUP_IOCTL_PAGE_ALLOC_EX: No mapping requested!\n"));
2147	REQ_CHECK_EXPR_FMT(pReq->u.In.fUserMapping,
2148	("SUP_IOCTL_PAGE_ALLOC_EX: Must have user mapping!\n"));
2149	REQ_CHECK_EXPR_FMT(!pReq->u.In.fReserved0 && !pReq->u.In.fReserved1,
2150	("SUP_IOCTL_PAGE_ALLOC_EX: fReserved0=%d fReserved1=%d\n", pReq->u.In.fReserved0, pReq->u.In.fReserved1));
2151
2152	/* execute */
2153	pReq->Hdr.rc = SUPR0PageAllocEx(pSession, pReq->u.In.cPages, 0 /* fFlags */,
2154	pReq->u.In.fUserMapping ? &pReq->u.Out.pvR3 : NULL,
2155	pReq->u.In.fKernelMapping ? &pReq->u.Out.pvR0 : NULL,
2156	&pReq->u.Out.aPages[0]);
2157	if (RT_FAILURE(pReq->Hdr.rc))
2158	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
2159	return 0;
2160	}
2161
2162	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_MAP_KERNEL):
2163	{
2164	/* validate */
2165	PSUPPAGEMAPKERNEL pReq = (PSUPPAGEMAPKERNEL)pReqHdr;
2166	REQ_CHECK_SIZES(SUP_IOCTL_PAGE_MAP_KERNEL);
2167	REQ_CHECK_EXPR_FMT(!pReq->u.In.fFlags, ("SUP_IOCTL_PAGE_MAP_KERNEL: fFlags=%#x! MBZ\n", pReq->u.In.fFlags));
2168	REQ_CHECK_EXPR_FMT(!(pReq->u.In.offSub & PAGE_OFFSET_MASK), ("SUP_IOCTL_PAGE_MAP_KERNEL: offSub=%#x\n", pReq->u.In.offSub));
2169	REQ_CHECK_EXPR_FMT(pReq->u.In.cbSub && !(pReq->u.In.cbSub & PAGE_OFFSET_MASK),
2170	("SUP_IOCTL_PAGE_MAP_KERNEL: cbSub=%#x\n", pReq->u.In.cbSub));
2171
2172	/* execute */
2173	pReq->Hdr.rc = SUPR0PageMapKernel(pSession, pReq->u.In.pvR3, pReq->u.In.offSub, pReq->u.In.cbSub,
2174	pReq->u.In.fFlags, &pReq->u.Out.pvR0);
2175	if (RT_FAILURE(pReq->Hdr.rc))
2176	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
2177	return 0;
2178	}
2179
2180	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_PROTECT):
2181	{
2182	/* validate */
2183	PSUPPAGEPROTECT pReq = (PSUPPAGEPROTECT)pReqHdr;
2184	REQ_CHECK_SIZES(SUP_IOCTL_PAGE_PROTECT);
2185	REQ_CHECK_EXPR_FMT(!(pReq->u.In.fProt & ~(RTMEM_PROT_READ \| RTMEM_PROT_WRITE \| RTMEM_PROT_EXEC \| RTMEM_PROT_NONE)),
2186	("SUP_IOCTL_PAGE_PROTECT: fProt=%#x!\n", pReq->u.In.fProt));
2187	REQ_CHECK_EXPR_FMT(!(pReq->u.In.offSub & PAGE_OFFSET_MASK), ("SUP_IOCTL_PAGE_PROTECT: offSub=%#x\n", pReq->u.In.offSub));
2188	REQ_CHECK_EXPR_FMT(pReq->u.In.cbSub && !(pReq->u.In.cbSub & PAGE_OFFSET_MASK),
2189	("SUP_IOCTL_PAGE_PROTECT: cbSub=%#x\n", pReq->u.In.cbSub));
2190
2191	/* execute */
2192	pReq->Hdr.rc = SUPR0PageProtect(pSession, pReq->u.In.pvR3, pReq->u.In.pvR0, pReq->u.In.offSub, pReq->u.In.cbSub, pReq->u.In.fProt);
2193	return 0;
2194	}
2195
2196	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_FREE):
2197	{
2198	/* validate */
2199	PSUPPAGEFREE pReq = (PSUPPAGEFREE)pReqHdr;
2200	REQ_CHECK_SIZES(SUP_IOCTL_PAGE_FREE);
2201
2202	/* execute */
2203	pReq->Hdr.rc = SUPR0PageFree(pSession, pReq->u.In.pvR3);
2204	return 0;
2205	}
2206
2207	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_CALL_SERVICE_NO_SIZE()):
2208	{
2209	/* validate */
2210	PSUPCALLSERVICE pReq = (PSUPCALLSERVICE)pReqHdr;
2211	Log4(("SUP_IOCTL_CALL_SERVICE: op=%u in=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
2212	pReq->u.In.uOperation, pReq->Hdr.cbIn, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
2213
2214	if (pReq->Hdr.cbIn == SUP_IOCTL_CALL_SERVICE_SIZE(0))
2215	REQ_CHECK_SIZES_EX(SUP_IOCTL_CALL_SERVICE, SUP_IOCTL_CALL_SERVICE_SIZE_IN(0), SUP_IOCTL_CALL_SERVICE_SIZE_OUT(0));
2216	else
2217	{
2218	PSUPR0SERVICEREQHDR pSrvReq = (PSUPR0SERVICEREQHDR)&pReq->abReqPkt[0];
2219	REQ_CHECK_EXPR_FMT(pReq->Hdr.cbIn >= SUP_IOCTL_CALL_SERVICE_SIZE(sizeof(SUPR0SERVICEREQHDR)),
2220	("SUP_IOCTL_CALL_SERVICE: cbIn=%#x < %#lx\n", pReq->Hdr.cbIn, SUP_IOCTL_CALL_SERVICE_SIZE(sizeof(SUPR0SERVICEREQHDR))));
2221	REQ_CHECK_EXPR(SUP_IOCTL_CALL_SERVICE, pSrvReq->u32Magic == SUPR0SERVICEREQHDR_MAGIC);
2222	REQ_CHECK_SIZES_EX(SUP_IOCTL_CALL_SERVICE, SUP_IOCTL_CALL_SERVICE_SIZE_IN(pSrvReq->cbReq), SUP_IOCTL_CALL_SERVICE_SIZE_OUT(pSrvReq->cbReq));
2223	}
2224	REQ_CHECK_EXPR(SUP_IOCTL_CALL_SERVICE, RTStrEnd(pReq->u.In.szName, sizeof(pReq->u.In.szName)));
2225
2226	/* execute */
2227	pReq->Hdr.rc = supdrvIOCtl_CallServiceModule(pDevExt, pSession, pReq);
2228	return 0;
2229	}
2230
2231	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LOGGER_SETTINGS_NO_SIZE()):
2232	{
2233	/* validate */
2234	PSUPLOGGERSETTINGS pReq = (PSUPLOGGERSETTINGS)pReqHdr;
2235	size_t cbStrTab;
2236	REQ_CHECK_SIZE_OUT(SUP_IOCTL_LOGGER_SETTINGS, SUP_IOCTL_LOGGER_SETTINGS_SIZE_OUT);
2237	REQ_CHECK_EXPR(SUP_IOCTL_LOGGER_SETTINGS, pReq->Hdr.cbIn >= SUP_IOCTL_LOGGER_SETTINGS_SIZE_IN(1));
2238	cbStrTab = pReq->Hdr.cbIn - SUP_IOCTL_LOGGER_SETTINGS_SIZE_IN(0);
2239	REQ_CHECK_EXPR(SUP_IOCTL_LOGGER_SETTINGS, pReq->u.In.offGroups < cbStrTab);
2240	REQ_CHECK_EXPR(SUP_IOCTL_LOGGER_SETTINGS, pReq->u.In.offFlags < cbStrTab);
2241	REQ_CHECK_EXPR(SUP_IOCTL_LOGGER_SETTINGS, pReq->u.In.offDestination < cbStrTab);
2242	REQ_CHECK_EXPR_FMT(pReq->u.In.szStrings[cbStrTab - 1] == '\0',
2243	("SUP_IOCTL_LOGGER_SETTINGS: cbIn=%#x cbStrTab=%#zx LastChar=%d\n",
2244	pReq->Hdr.cbIn, cbStrTab, pReq->u.In.szStrings[cbStrTab - 1]));
2245	REQ_CHECK_EXPR(SUP_IOCTL_LOGGER_SETTINGS, pReq->u.In.fWhich <= SUPLOGGERSETTINGS_WHICH_RELEASE);
2246	REQ_CHECK_EXPR(SUP_IOCTL_LOGGER_SETTINGS, pReq->u.In.fWhat <= SUPLOGGERSETTINGS_WHAT_DESTROY);
2247
2248	/* execute */
2249	pReq->Hdr.rc = supdrvIOCtl_LoggerSettings(pReq);
2250	return 0;
2251	}
2252
2253	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_SEM_OP2):
2254	{
2255	/* validate */
2256	PSUPSEMOP2 pReq = (PSUPSEMOP2)pReqHdr;
2257	REQ_CHECK_SIZES_EX(SUP_IOCTL_SEM_OP2, SUP_IOCTL_SEM_OP2_SIZE_IN, SUP_IOCTL_SEM_OP2_SIZE_OUT);
2258	REQ_CHECK_EXPR(SUP_IOCTL_SEM_OP2, pReq->u.In.uReserved == 0);
2259
2260	/* execute */
2261	switch (pReq->u.In.uType)
2262	{
2263	case SUP_SEM_TYPE_EVENT:
2264	{
2265	SUPSEMEVENT hEvent = (SUPSEMEVENT)(uintptr_t)pReq->u.In.hSem;
2266	switch (pReq->u.In.uOp)
2267	{
2268	case SUPSEMOP2_WAIT_MS_REL:
2269	pReq->Hdr.rc = SUPSemEventWaitNoResume(pSession, hEvent, pReq->u.In.uArg.cRelMsTimeout);
2270	break;
2271	case SUPSEMOP2_WAIT_NS_ABS:
2272	pReq->Hdr.rc = SUPSemEventWaitNsAbsIntr(pSession, hEvent, pReq->u.In.uArg.uAbsNsTimeout);
2273	break;
2274	case SUPSEMOP2_WAIT_NS_REL:
2275	pReq->Hdr.rc = SUPSemEventWaitNsRelIntr(pSession, hEvent, pReq->u.In.uArg.cRelNsTimeout);
2276	break;
2277	case SUPSEMOP2_SIGNAL:
2278	pReq->Hdr.rc = SUPSemEventSignal(pSession, hEvent);
2279	break;
2280	case SUPSEMOP2_CLOSE:
2281	pReq->Hdr.rc = SUPSemEventClose(pSession, hEvent);
2282	break;
2283	case SUPSEMOP2_RESET:
2284	default:
2285	pReq->Hdr.rc = VERR_INVALID_FUNCTION;
2286	break;
2287	}
2288	break;
2289	}
2290
2291	case SUP_SEM_TYPE_EVENT_MULTI:
2292	{
2293	SUPSEMEVENTMULTI hEventMulti = (SUPSEMEVENTMULTI)(uintptr_t)pReq->u.In.hSem;
2294	switch (pReq->u.In.uOp)
2295	{
2296	case SUPSEMOP2_WAIT_MS_REL:
2297	pReq->Hdr.rc = SUPSemEventMultiWaitNoResume(pSession, hEventMulti, pReq->u.In.uArg.cRelMsTimeout);
2298	break;
2299	case SUPSEMOP2_WAIT_NS_ABS:
2300	pReq->Hdr.rc = SUPSemEventMultiWaitNsAbsIntr(pSession, hEventMulti, pReq->u.In.uArg.uAbsNsTimeout);
2301	break;
2302	case SUPSEMOP2_WAIT_NS_REL:
2303	pReq->Hdr.rc = SUPSemEventMultiWaitNsRelIntr(pSession, hEventMulti, pReq->u.In.uArg.cRelNsTimeout);
2304	break;
2305	case SUPSEMOP2_SIGNAL:
2306	pReq->Hdr.rc = SUPSemEventMultiSignal(pSession, hEventMulti);
2307	break;
2308	case SUPSEMOP2_CLOSE:
2309	pReq->Hdr.rc = SUPSemEventMultiClose(pSession, hEventMulti);
2310	break;
2311	case SUPSEMOP2_RESET:
2312	pReq->Hdr.rc = SUPSemEventMultiReset(pSession, hEventMulti);
2313	break;
2314	default:
2315	pReq->Hdr.rc = VERR_INVALID_FUNCTION;
2316	break;
2317	}
2318	break;
2319	}
2320
2321	default:
2322	pReq->Hdr.rc = VERR_INVALID_PARAMETER;
2323	break;
2324	}
2325	return 0;
2326	}
2327
2328	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_SEM_OP3):
2329	{
2330	/* validate */
2331	PSUPSEMOP3 pReq = (PSUPSEMOP3)pReqHdr;
2332	REQ_CHECK_SIZES_EX(SUP_IOCTL_SEM_OP3, SUP_IOCTL_SEM_OP3_SIZE_IN, SUP_IOCTL_SEM_OP3_SIZE_OUT);
2333	REQ_CHECK_EXPR(SUP_IOCTL_SEM_OP3, pReq->u.In.u32Reserved == 0 && pReq->u.In.u64Reserved == 0);
2334
2335	/* execute */
2336	switch (pReq->u.In.uType)
2337	{
2338	case SUP_SEM_TYPE_EVENT:
2339	{
2340	SUPSEMEVENT hEvent = (SUPSEMEVENT)(uintptr_t)pReq->u.In.hSem;
2341	switch (pReq->u.In.uOp)
2342	{
2343	case SUPSEMOP3_CREATE:
2344	REQ_CHECK_EXPR(SUP_IOCTL_SEM_OP3, hEvent == NIL_SUPSEMEVENT);
2345	pReq->Hdr.rc = SUPSemEventCreate(pSession, &hEvent);
2346	pReq->u.Out.hSem = (uint32_t)(uintptr_t)hEvent;
2347	break;
2348	case SUPSEMOP3_GET_RESOLUTION:
2349	REQ_CHECK_EXPR(SUP_IOCTL_SEM_OP3, hEvent == NIL_SUPSEMEVENT);
2350	pReq->Hdr.rc = VINF_SUCCESS;
2351	pReq->Hdr.cbOut = sizeof(*pReq);
2352	pReq->u.Out.cNsResolution = SUPSemEventGetResolution(pSession);
2353	break;
2354	default:
2355	pReq->Hdr.rc = VERR_INVALID_FUNCTION;
2356	break;
2357	}
2358	break;
2359	}
2360
2361	case SUP_SEM_TYPE_EVENT_MULTI:
2362	{
2363	SUPSEMEVENTMULTI hEventMulti = (SUPSEMEVENTMULTI)(uintptr_t)pReq->u.In.hSem;
2364	switch (pReq->u.In.uOp)
2365	{
2366	case SUPSEMOP3_CREATE:
2367	REQ_CHECK_EXPR(SUP_IOCTL_SEM_OP3, hEventMulti == NIL_SUPSEMEVENTMULTI);
2368	pReq->Hdr.rc = SUPSemEventMultiCreate(pSession, &hEventMulti);
2369	pReq->u.Out.hSem = (uint32_t)(uintptr_t)hEventMulti;
2370	break;
2371	case SUPSEMOP3_GET_RESOLUTION:
2372	REQ_CHECK_EXPR(SUP_IOCTL_SEM_OP3, hEventMulti == NIL_SUPSEMEVENTMULTI);
2373	pReq->Hdr.rc = VINF_SUCCESS;
2374	pReq->u.Out.cNsResolution = SUPSemEventMultiGetResolution(pSession);
2375	break;
2376	default:
2377	pReq->Hdr.rc = VERR_INVALID_FUNCTION;
2378	break;
2379	}
2380	break;
2381	}
2382
2383	default:
2384	pReq->Hdr.rc = VERR_INVALID_PARAMETER;
2385	break;
2386	}
2387	return 0;
2388	}
2389
2390	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_VT_CAPS):
2391	{
2392	/* validate */
2393	PSUPVTCAPS pReq = (PSUPVTCAPS)pReqHdr;
2394	REQ_CHECK_SIZES(SUP_IOCTL_VT_CAPS);
2395
2396	/* execute */
2397	pReq->Hdr.rc = SUPR0QueryVTCaps(pSession, &pReq->u.Out.fCaps);
2398	if (RT_FAILURE(pReq->Hdr.rc))
2399	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
2400	return 0;
2401	}
2402
2403	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_TRACER_OPEN):
2404	{
2405	/* validate */
2406	PSUPTRACEROPEN pReq = (PSUPTRACEROPEN)pReqHdr;
2407	REQ_CHECK_SIZES(SUP_IOCTL_TRACER_OPEN);
2408
2409	/* execute */
2410	pReq->Hdr.rc = supdrvIOCtl_TracerOpen(pDevExt, pSession, pReq->u.In.uCookie, pReq->u.In.uArg);
2411	return 0;
2412	}
2413
2414	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_TRACER_CLOSE):
2415	{
2416	/* validate */
2417	REQ_CHECK_SIZES(SUP_IOCTL_TRACER_CLOSE);
2418
2419	/* execute */
2420	pReqHdr->rc = supdrvIOCtl_TracerClose(pDevExt, pSession);
2421	return 0;
2422	}
2423
2424	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_TRACER_IOCTL):
2425	{
2426	/* validate */
2427	PSUPTRACERIOCTL pReq = (PSUPTRACERIOCTL)pReqHdr;
2428	REQ_CHECK_SIZES(SUP_IOCTL_TRACER_IOCTL);
2429
2430	/* execute */
2431	pReqHdr->rc = supdrvIOCtl_TracerIOCtl(pDevExt, pSession, pReq->u.In.uCmd, pReq->u.In.uArg, &pReq->u.Out.iRetVal);
2432	return 0;
2433	}
2434
2435	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_TRACER_UMOD_REG):
2436	{
2437	/* validate */
2438	PSUPTRACERUMODREG pReq = (PSUPTRACERUMODREG)pReqHdr;
2439	REQ_CHECK_SIZES(SUP_IOCTL_TRACER_UMOD_REG);
2440	if (!RTStrEnd(pReq->u.In.szName, sizeof(pReq->u.In.szName)))
2441	return VERR_INVALID_PARAMETER;
2442
2443	/* execute */
2444	pReqHdr->rc = supdrvIOCtl_TracerUmodRegister(pDevExt, pSession,
2445	pReq->u.In.R3PtrVtgHdr, pReq->u.In.uVtgHdrAddr,
2446	pReq->u.In.R3PtrStrTab, pReq->u.In.cbStrTab,
2447	pReq->u.In.szName, pReq->u.In.fFlags);
2448	return 0;
2449	}
2450
2451	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_TRACER_UMOD_DEREG):
2452	{
2453	/* validate */
2454	PSUPTRACERUMODDEREG pReq = (PSUPTRACERUMODDEREG)pReqHdr;
2455	REQ_CHECK_SIZES(SUP_IOCTL_TRACER_UMOD_DEREG);
2456
2457	/* execute */
2458	pReqHdr->rc = supdrvIOCtl_TracerUmodDeregister(pDevExt, pSession, pReq->u.In.pVtgHdr);
2459	return 0;
2460	}
2461
2462	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_TRACER_UMOD_FIRE_PROBE):
2463	{
2464	/* validate */
2465	PSUPTRACERUMODFIREPROBE pReq = (PSUPTRACERUMODFIREPROBE)pReqHdr;
2466	REQ_CHECK_SIZES(SUP_IOCTL_TRACER_UMOD_FIRE_PROBE);
2467
2468	supdrvIOCtl_TracerUmodProbeFire(pDevExt, pSession, &pReq->u.In);
2469	pReqHdr->rc = VINF_SUCCESS;
2470	return 0;
2471	}
2472
2473	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_MSR_PROBER):
2474	{
2475	/* validate */
2476	PSUPMSRPROBER pReq = (PSUPMSRPROBER)pReqHdr;
2477	REQ_CHECK_SIZES(SUP_IOCTL_MSR_PROBER);
2478	REQ_CHECK_EXPR(SUP_IOCTL_MSR_PROBER,
2479	pReq->u.In.enmOp > SUPMSRPROBEROP_INVALID && pReq->u.In.enmOp < SUPMSRPROBEROP_END);
2480
2481	pReqHdr->rc = supdrvIOCtl_MsrProber(pDevExt, pReq);
2482	return 0;
2483	}
2484
2485	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_RESUME_SUSPENDED_KBDS):
2486	{
2487	/* validate */
2488	REQ_CHECK_SIZES(SUP_IOCTL_RESUME_SUSPENDED_KBDS);
2489
2490	pReqHdr->rc = supdrvIOCtl_ResumeSuspendedKbds();
2491	return 0;
2492	}
2493
2494	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_TSC_DELTA_MEASURE):
2495	{
2496	/* validate */
2497	PSUPTSCDELTAMEASURE pReq = (PSUPTSCDELTAMEASURE)pReqHdr;
2498	REQ_CHECK_SIZES(SUP_IOCTL_TSC_DELTA_MEASURE);
2499
2500	pReqHdr->rc = supdrvIOCtl_TscDeltaMeasure(pDevExt, pSession, pReq);
2501	return 0;
2502	}
2503
2504	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_TSC_READ):
2505	{
2506	/* validate */
2507	PSUPTSCREAD pReq = (PSUPTSCREAD)pReqHdr;
2508	REQ_CHECK_SIZES(SUP_IOCTL_TSC_READ);
2509
2510	pReqHdr->rc = supdrvIOCtl_TscRead(pDevExt, pSession, pReq);
2511	return 0;
2512	}
2513
2514	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_GIP_SET_FLAGS):
2515	{
2516	/* validate */
2517	PSUPGIPSETFLAGS pReq = (PSUPGIPSETFLAGS)pReqHdr;
2518	REQ_CHECK_SIZES(SUP_IOCTL_GIP_SET_FLAGS);
2519
2520	pReqHdr->rc = supdrvIOCtl_GipSetFlags(pDevExt, pSession, pReq->u.In.fOrMask, pReq->u.In.fAndMask);
2521	return 0;
2522	}
2523
2524	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_UCODE_REV):
2525	{
2526	/* validate */
2527	PSUPUCODEREV pReq = (PSUPUCODEREV)pReqHdr;
2528	REQ_CHECK_SIZES(SUP_IOCTL_UCODE_REV);
2529
2530	/* execute */
2531	pReq->Hdr.rc = SUPR0QueryUcodeRev(pSession, &pReq->u.Out.MicrocodeRev);
2532	if (RT_FAILURE(pReq->Hdr.rc))
2533	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
2534	return 0;
2535	}
2536
2537	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_GET_HWVIRT_MSRS):
2538	{
2539	/* validate */
2540	PSUPGETHWVIRTMSRS pReq = (PSUPGETHWVIRTMSRS)pReqHdr;
2541	REQ_CHECK_SIZES(SUP_IOCTL_GET_HWVIRT_MSRS);
2542	REQ_CHECK_EXPR_FMT(!pReq->u.In.fReserved0 && !pReq->u.In.fReserved1 && !pReq->u.In.fReserved2,
2543	("SUP_IOCTL_GET_HWVIRT_MSRS: fReserved0=%d fReserved1=%d fReserved2=%d\n", pReq->u.In.fReserved0,
2544	pReq->u.In.fReserved1, pReq->u.In.fReserved2));
2545
2546	/* execute */
2547	pReq->Hdr.rc = SUPR0GetHwvirtMsrs(&pReq->u.Out.HwvirtMsrs, 0 /* fCaps */, pReq->u.In.fForce);
2548	if (RT_FAILURE(pReq->Hdr.rc))
2549	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
2550	return 0;
2551	}
2552
2553	default:
2554	Log(("Unknown IOCTL %#lx\n", (long)uIOCtl));
2555	break;
2556	}
2557	return VERR_GENERAL_FAILURE;
2558	}
2559
2560
2561	/**
2562	* I/O Control inner worker for the restricted operations.
2563	*
2564	* @returns IPRT status code.
2565	* @retval VERR_INVALID_PARAMETER if the request is invalid.
2566	*
2567	* @param uIOCtl Function number.
2568	* @param pDevExt Device extention.
2569	* @param pSession Session data.
2570	* @param pReqHdr The request header.
2571	*/
2572	static int supdrvIOCtlInnerRestricted(uintptr_t uIOCtl, PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPREQHDR pReqHdr)
2573	{
2574	/*
2575	* The switch.
2576	*/
2577	switch (SUP_CTL_CODE_NO_SIZE(uIOCtl))
2578	{
2579	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_COOKIE):
2580	{
2581	PSUPCOOKIE pReq = (PSUPCOOKIE)pReqHdr;
2582	REQ_CHECK_SIZES(SUP_IOCTL_COOKIE);
2583	if (strncmp(pReq->u.In.szMagic, SUPCOOKIE_MAGIC, sizeof(pReq->u.In.szMagic)))
2584	{
2585	OSDBGPRINT(("SUP_IOCTL_COOKIE: invalid magic %.16s\n", pReq->u.In.szMagic));
2586	pReq->Hdr.rc = VERR_INVALID_MAGIC;
2587	return 0;
2588	}
2589
2590	/*
2591	* Match the version.
2592	* The current logic is very simple, match the major interface version.
2593	*/
2594	if ( pReq->u.In.u32MinVersion > SUPDRV_IOC_VERSION
2595	\|\| (pReq->u.In.u32MinVersion & 0xffff0000) != (SUPDRV_IOC_VERSION & 0xffff0000))
2596	{
2597	OSDBGPRINT(("SUP_IOCTL_COOKIE: Version mismatch. Requested: %#x Min: %#x Current: %#x\n",
2598	pReq->u.In.u32ReqVersion, pReq->u.In.u32MinVersion, SUPDRV_IOC_VERSION));
2599	pReq->u.Out.u32Cookie = 0xffffffff;
2600	pReq->u.Out.u32SessionCookie = 0xffffffff;
2601	pReq->u.Out.u32SessionVersion = 0xffffffff;
2602	pReq->u.Out.u32DriverVersion = SUPDRV_IOC_VERSION;
2603	pReq->u.Out.pSession = NULL;
2604	pReq->u.Out.cFunctions = 0;
2605	pReq->Hdr.rc = VERR_VERSION_MISMATCH;
2606	return 0;
2607	}
2608
2609	/*
2610	* Fill in return data and be gone.
2611	* N.B. The first one to change SUPDRV_IOC_VERSION shall makes sure that
2612	* u32SessionVersion <= u32ReqVersion!
2613	*/
2614	/** @todo Somehow validate the client and negotiate a secure cookie... */
2615	pReq->u.Out.u32Cookie = pDevExt->u32Cookie;
2616	pReq->u.Out.u32SessionCookie = pSession->u32Cookie;
2617	pReq->u.Out.u32SessionVersion = SUPDRV_IOC_VERSION;
2618	pReq->u.Out.u32DriverVersion = SUPDRV_IOC_VERSION;
2619	pReq->u.Out.pSession = pSession;
2620	pReq->u.Out.cFunctions = 0;
2621	pReq->Hdr.rc = VINF_SUCCESS;
2622	return 0;
2623	}
2624
2625	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_VT_CAPS):
2626	{
2627	/* validate */
2628	PSUPVTCAPS pReq = (PSUPVTCAPS)pReqHdr;
2629	REQ_CHECK_SIZES(SUP_IOCTL_VT_CAPS);
2630
2631	/* execute */
2632	pReq->Hdr.rc = SUPR0QueryVTCaps(pSession, &pReq->u.Out.fCaps);
2633	if (RT_FAILURE(pReq->Hdr.rc))
2634	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
2635	return 0;
2636	}
2637
2638	default:
2639	Log(("Unknown IOCTL %#lx\n", (long)uIOCtl));
2640	break;
2641	}
2642	return VERR_GENERAL_FAILURE;
2643	}
2644
2645
2646	/**
2647	* I/O Control worker.
2648	*
2649	* @returns IPRT status code.
2650	* @retval VERR_INVALID_PARAMETER if the request is invalid.
2651	*
2652	* @param uIOCtl Function number.
2653	* @param pDevExt Device extention.
2654	* @param pSession Session data.
2655	* @param pReqHdr The request header.
2656	* @param cbReq The size of the request buffer.
2657	*/
2658	int VBOXCALL supdrvIOCtl(uintptr_t uIOCtl, PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPREQHDR pReqHdr, size_t cbReq)
2659	{
2660	int rc;
2661	VBOXDRV_IOCTL_ENTRY(pSession, uIOCtl, pReqHdr);
2662
2663	/*
2664	* Validate the request.
2665	*/
2666	if (RT_UNLIKELY(cbReq < sizeof(*pReqHdr)))
2667	{
2668	OSDBGPRINT(("vboxdrv: Bad ioctl request size; cbReq=%#lx\n", (long)cbReq));
2669	VBOXDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VERR_INVALID_PARAMETER, VINF_SUCCESS);
2670	return VERR_INVALID_PARAMETER;
2671	}
2672	if (RT_UNLIKELY( (pReqHdr->fFlags & SUPREQHDR_FLAGS_MAGIC_MASK) != SUPREQHDR_FLAGS_MAGIC
2673	\|\| pReqHdr->cbIn < sizeof(*pReqHdr)
2674	\|\| pReqHdr->cbIn > cbReq
2675	\|\| pReqHdr->cbOut < sizeof(*pReqHdr)
2676	\|\| pReqHdr->cbOut > cbReq))
2677	{
2678	OSDBGPRINT(("vboxdrv: Bad ioctl request header; cbIn=%#lx cbOut=%#lx fFlags=%#lx\n",
2679	(long)pReqHdr->cbIn, (long)pReqHdr->cbOut, (long)pReqHdr->fFlags));
2680	VBOXDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VERR_INVALID_PARAMETER, VINF_SUCCESS);
2681	return VERR_INVALID_PARAMETER;
2682	}
2683	if (RT_UNLIKELY(!RT_VALID_PTR(pSession)))
2684	{
2685	OSDBGPRINT(("vboxdrv: Invalid pSession value %p (ioctl=%p)\n", pSession, (void *)uIOCtl));
2686	VBOXDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VERR_INVALID_PARAMETER, VINF_SUCCESS);
2687	return VERR_INVALID_PARAMETER;
2688	}
2689	if (RT_UNLIKELY(uIOCtl == SUP_IOCTL_COOKIE))
2690	{
2691	if (pReqHdr->u32Cookie != SUPCOOKIE_INITIAL_COOKIE)
2692	{
2693	OSDBGPRINT(("SUP_IOCTL_COOKIE: bad cookie %#lx\n", (long)pReqHdr->u32Cookie));
2694	VBOXDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VERR_INVALID_PARAMETER, VINF_SUCCESS);
2695	return VERR_INVALID_PARAMETER;
2696	}
2697	}
2698	else if (RT_UNLIKELY( pReqHdr->u32Cookie != pDevExt->u32Cookie
2699	\|\| pReqHdr->u32SessionCookie != pSession->u32Cookie))
2700	{
2701	OSDBGPRINT(("vboxdrv: bad cookie %#lx / %#lx.\n", (long)pReqHdr->u32Cookie, (long)pReqHdr->u32SessionCookie));
2702	VBOXDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VERR_INVALID_PARAMETER, VINF_SUCCESS);
2703	return VERR_INVALID_PARAMETER;
2704	}
2705
2706	/*
2707	* Hand it to an inner function to avoid lots of unnecessary return tracepoints.
2708	*/
2709	if (pSession->fUnrestricted)
2710	rc = supdrvIOCtlInnerUnrestricted(uIOCtl, pDevExt, pSession, pReqHdr);
2711	else
2712	rc = supdrvIOCtlInnerRestricted(uIOCtl, pDevExt, pSession, pReqHdr);
2713
2714	VBOXDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, pReqHdr->rc, rc);
2715	return rc;
2716	}
2717
2718
2719	/**
2720	* Inter-Driver Communication (IDC) worker.
2721	*
2722	* @returns VBox status code.
2723	* @retval VINF_SUCCESS on success.
2724	* @retval VERR_INVALID_PARAMETER if the request is invalid.
2725	* @retval VERR_NOT_SUPPORTED if the request isn't supported.
2726	*
2727	* @param uReq The request (function) code.
2728	* @param pDevExt Device extention.
2729	* @param pSession Session data.
2730	* @param pReqHdr The request header.
2731	*/
2732	int VBOXCALL supdrvIDC(uintptr_t uReq, PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPDRVIDCREQHDR pReqHdr)
2733	{
2734	/*
2735	* The OS specific code has already validated the pSession
2736	* pointer, and the request size being greater or equal to
2737	* size of the header.
2738	*
2739	* So, just check that pSession is a kernel context session.
2740	*/
2741	if (RT_UNLIKELY( pSession
2742	&& pSession->R0Process != NIL_RTR0PROCESS))
2743	return VERR_INVALID_PARAMETER;
2744
2745	/*
2746	* Validation macro.
2747	*/
2748	#define REQ_CHECK_IDC_SIZE(Name, cbExpect) \
2749	do { \
2750	if (RT_UNLIKELY(pReqHdr->cb != (cbExpect))) \
2751	{ \
2752	OSDBGPRINT(( #Name ": Invalid input/output sizes. cb=%ld expected %ld.\n", \
2753	(long)pReqHdr->cb, (long)(cbExpect))); \
2754	return pReqHdr->rc = VERR_INVALID_PARAMETER; \
2755	} \
2756	} while (0)
2757
2758	switch (uReq)
2759	{
2760	case SUPDRV_IDC_REQ_CONNECT:
2761	{
2762	PSUPDRVIDCREQCONNECT pReq = (PSUPDRVIDCREQCONNECT)pReqHdr;
2763	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_CONNECT, sizeof(*pReq));
2764
2765	/*
2766	* Validate the cookie and other input.
2767	*/
2768	if (pReq->Hdr.pSession != NULL)
2769	{
2770	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: Hdr.pSession=%p expected NULL!\n", pReq->Hdr.pSession));
2771	return pReqHdr->rc = VERR_INVALID_PARAMETER;
2772	}
2773	if (pReq->u.In.u32MagicCookie != SUPDRVIDCREQ_CONNECT_MAGIC_COOKIE)
2774	{
2775	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: u32MagicCookie=%#x expected %#x!\n",
2776	(unsigned)pReq->u.In.u32MagicCookie, (unsigned)SUPDRVIDCREQ_CONNECT_MAGIC_COOKIE));
2777	return pReqHdr->rc = VERR_INVALID_PARAMETER;
2778	}
2779	if ( pReq->u.In.uMinVersion > pReq->u.In.uReqVersion
2780	\|\| (pReq->u.In.uMinVersion & UINT32_C(0xffff0000)) != (pReq->u.In.uReqVersion & UINT32_C(0xffff0000)))
2781	{
2782	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: uMinVersion=%#x uMaxVersion=%#x doesn't match!\n",
2783	pReq->u.In.uMinVersion, pReq->u.In.uReqVersion));
2784	return pReqHdr->rc = VERR_INVALID_PARAMETER;
2785	}
2786	if (pSession != NULL)
2787	{
2788	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: pSession=%p expected NULL!\n", pSession));
2789	return pReqHdr->rc = VERR_INVALID_PARAMETER;
2790	}
2791
2792	/*
2793	* Match the version.
2794	* The current logic is very simple, match the major interface version.
2795	*/
2796	if ( pReq->u.In.uMinVersion > SUPDRV_IDC_VERSION
2797	\|\| (pReq->u.In.uMinVersion & 0xffff0000) != (SUPDRV_IDC_VERSION & 0xffff0000))
2798	{
2799	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: Version mismatch. Requested: %#x Min: %#x Current: %#x\n",
2800	pReq->u.In.uReqVersion, pReq->u.In.uMinVersion, (unsigned)SUPDRV_IDC_VERSION));
2801	pReq->u.Out.pSession = NULL;
2802	pReq->u.Out.uSessionVersion = 0xffffffff;
2803	pReq->u.Out.uDriverVersion = SUPDRV_IDC_VERSION;
2804	pReq->u.Out.uDriverRevision = VBOX_SVN_REV;
2805	pReq->Hdr.rc = VERR_VERSION_MISMATCH;
2806	return VINF_SUCCESS;
2807	}
2808
2809	pReq->u.Out.pSession = NULL;
2810	pReq->u.Out.uSessionVersion = SUPDRV_IDC_VERSION;
2811	pReq->u.Out.uDriverVersion = SUPDRV_IDC_VERSION;
2812	pReq->u.Out.uDriverRevision = VBOX_SVN_REV;
2813
2814	pReq->Hdr.rc = supdrvCreateSession(pDevExt, false /* fUser /, true /fUnrestricted*/, &pSession);
2815	if (RT_FAILURE(pReq->Hdr.rc))
2816	{
2817	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: failed to create session, rc=%d\n", pReq->Hdr.rc));
2818	return VINF_SUCCESS;
2819	}
2820
2821	pReq->u.Out.pSession = pSession;
2822	pReq->Hdr.pSession = pSession;
2823
2824	return VINF_SUCCESS;
2825	}
2826
2827	case SUPDRV_IDC_REQ_DISCONNECT:
2828	{
2829	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_DISCONNECT, sizeof(*pReqHdr));
2830
2831	supdrvSessionRelease(pSession);
2832	return pReqHdr->rc = VINF_SUCCESS;
2833	}
2834
2835	case SUPDRV_IDC_REQ_GET_SYMBOL:
2836	{
2837	PSUPDRVIDCREQGETSYM pReq = (PSUPDRVIDCREQGETSYM)pReqHdr;
2838	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_GET_SYMBOL, sizeof(*pReq));
2839
2840	pReq->Hdr.rc = supdrvIDC_LdrGetSymbol(pDevExt, pSession, pReq);
2841	return VINF_SUCCESS;
2842	}
2843
2844	case SUPDRV_IDC_REQ_COMPONENT_REGISTER_FACTORY:
2845	{
2846	PSUPDRVIDCREQCOMPREGFACTORY pReq = (PSUPDRVIDCREQCOMPREGFACTORY)pReqHdr;
2847	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_COMPONENT_REGISTER_FACTORY, sizeof(*pReq));
2848
2849	pReq->Hdr.rc = SUPR0ComponentRegisterFactory(pSession, pReq->u.In.pFactory);
2850	return VINF_SUCCESS;
2851	}
2852
2853	case SUPDRV_IDC_REQ_COMPONENT_DEREGISTER_FACTORY:
2854	{
2855	PSUPDRVIDCREQCOMPDEREGFACTORY pReq = (PSUPDRVIDCREQCOMPDEREGFACTORY)pReqHdr;
2856	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_COMPONENT_DEREGISTER_FACTORY, sizeof(*pReq));
2857
2858	pReq->Hdr.rc = SUPR0ComponentDeregisterFactory(pSession, pReq->u.In.pFactory);
2859	return VINF_SUCCESS;
2860	}
2861
2862	default:
2863	Log(("Unknown IDC %#lx\n", (long)uReq));
2864	break;
2865	}
2866
2867	#undef REQ_CHECK_IDC_SIZE
2868	return VERR_NOT_SUPPORTED;
2869	}
2870
2871
2872	/**
2873	* Register a object for reference counting.
2874	* The object is registered with one reference in the specified session.
2875	*
2876	* @returns Unique identifier on success (pointer).
2877	* All future reference must use this identifier.
2878	* @returns NULL on failure.
2879	* @param pSession The caller's session.
2880	* @param enmType The object type.
2881	* @param pfnDestructor The destructore function which will be called when the reference count reaches 0.
2882	* @param pvUser1 The first user argument.
2883	* @param pvUser2 The second user argument.
2884	*/
2885	SUPR0DECL(void ) SUPR0ObjRegister(PSUPDRVSESSION pSession, SUPDRVOBJTYPE enmType, PFNSUPDRVDESTRUCTOR pfnDestructor, void pvUser1, void *pvUser2)
2886	{
2887	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
2888	PSUPDRVOBJ pObj;
2889	PSUPDRVUSAGE pUsage;
2890
2891	/*
2892	* Validate the input.
2893	*/
2894	AssertReturn(SUP_IS_SESSION_VALID(pSession), NULL);
2895	AssertReturn(enmType > SUPDRVOBJTYPE_INVALID && enmType < SUPDRVOBJTYPE_END, NULL);
2896	AssertPtrReturn(pfnDestructor, NULL);
2897
2898	/*
2899	* Allocate and initialize the object.
2900	*/
2901	pObj = (PSUPDRVOBJ)RTMemAlloc(sizeof(*pObj));
2902	if (!pObj)
2903	return NULL;
2904	pObj->u32Magic = SUPDRVOBJ_MAGIC;
2905	pObj->enmType = enmType;
2906	pObj->pNext = NULL;
2907	pObj->cUsage = 1;
2908	pObj->pfnDestructor = pfnDestructor;
2909	pObj->pvUser1 = pvUser1;
2910	pObj->pvUser2 = pvUser2;
2911	pObj->CreatorUid = pSession->Uid;
2912	pObj->CreatorGid = pSession->Gid;
2913	pObj->CreatorProcess= pSession->Process;
2914	supdrvOSObjInitCreator(pObj, pSession);
2915
2916	/*
2917	* Allocate the usage record.
2918	* (We keep freed usage records around to simplify SUPR0ObjAddRefEx().)
2919	*/
2920	RTSpinlockAcquire(pDevExt->Spinlock);
2921
2922	pUsage = pDevExt->pUsageFree;
2923	if (pUsage)
2924	pDevExt->pUsageFree = pUsage->pNext;
2925	else
2926	{
2927	RTSpinlockRelease(pDevExt->Spinlock);
2928	pUsage = (PSUPDRVUSAGE)RTMemAlloc(sizeof(*pUsage));
2929	if (!pUsage)
2930	{
2931	RTMemFree(pObj);
2932	return NULL;
2933	}
2934	RTSpinlockAcquire(pDevExt->Spinlock);
2935	}
2936
2937	/*
2938	* Insert the object and create the session usage record.
2939	*/
2940	/* The object. */
2941	pObj->pNext = pDevExt->pObjs;
2942	pDevExt->pObjs = pObj;
2943
2944	/* The session record. */
2945	pUsage->cUsage = 1;
2946	pUsage->pObj = pObj;
2947	pUsage->pNext = pSession->pUsage;
2948	/* Log2(("SUPR0ObjRegister: pUsage=%p:{.pObj=%p, .pNext=%p}\n", pUsage, pUsage->pObj, pUsage->pNext)); */
2949	pSession->pUsage = pUsage;
2950
2951	RTSpinlockRelease(pDevExt->Spinlock);
2952
2953	Log(("SUPR0ObjRegister: returns %p (pvUser1=%p, pvUser=%p)\n", pObj, pvUser1, pvUser2));
2954	return pObj;
2955	}
2956	SUPR0_EXPORT_SYMBOL(SUPR0ObjRegister);
2957
2958
2959	/**
2960	* Increment the reference counter for the object associating the reference
2961	* with the specified session.
2962	*
2963	* @returns IPRT status code.
2964	* @param pvObj The identifier returned by SUPR0ObjRegister().
2965	* @param pSession The session which is referencing the object.
2966	*
2967	* @remarks The caller should not own any spinlocks and must carefully protect
2968	* itself against potential race with the destructor so freed memory
2969	* isn't accessed here.
2970	*/
2971	SUPR0DECL(int) SUPR0ObjAddRef(void *pvObj, PSUPDRVSESSION pSession)
2972	{
2973	return SUPR0ObjAddRefEx(pvObj, pSession, false /* fNoBlocking */);
2974	}
2975	SUPR0_EXPORT_SYMBOL(SUPR0ObjAddRef);
2976
2977
2978	/**
2979	* Increment the reference counter for the object associating the reference
2980	* with the specified session.
2981	*
2982	* @returns IPRT status code.
2983	* @retval VERR_TRY_AGAIN if fNoBlocking was set and a new usage record
2984	* couldn't be allocated. (If you see this you're not doing the right
2985	* thing and it won't ever work reliably.)
2986	*
2987	* @param pvObj The identifier returned by SUPR0ObjRegister().
2988	* @param pSession The session which is referencing the object.
2989	* @param fNoBlocking Set if it's not OK to block. Never try to make the
2990	* first reference to an object in a session with this
2991	* argument set.
2992	*
2993	* @remarks The caller should not own any spinlocks and must carefully protect
2994	* itself against potential race with the destructor so freed memory
2995	* isn't accessed here.
2996	*/
2997	SUPR0DECL(int) SUPR0ObjAddRefEx(void *pvObj, PSUPDRVSESSION pSession, bool fNoBlocking)
2998	{
2999	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
3000	PSUPDRVOBJ pObj = (PSUPDRVOBJ)pvObj;
3001	int rc = VINF_SUCCESS;
3002	PSUPDRVUSAGE pUsagePre;
3003	PSUPDRVUSAGE pUsage;
3004
3005	/*
3006	* Validate the input.
3007	* Be ready for the destruction race (someone might be stuck in the
3008	* destructor waiting a lock we own).
3009	*/
3010	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3011	AssertPtrReturn(pObj, VERR_INVALID_POINTER);
3012	AssertMsgReturn(pObj->u32Magic == SUPDRVOBJ_MAGIC \|\| pObj->u32Magic == SUPDRVOBJ_MAGIC_DEAD,
3013	("Invalid pvObj=%p magic=%#x (expected %#x or %#x)\n", pvObj, pObj->u32Magic, SUPDRVOBJ_MAGIC, SUPDRVOBJ_MAGIC_DEAD),
3014	VERR_INVALID_PARAMETER);
3015
3016	RTSpinlockAcquire(pDevExt->Spinlock);
3017
3018	if (RT_UNLIKELY(pObj->u32Magic != SUPDRVOBJ_MAGIC))
3019	{
3020	RTSpinlockRelease(pDevExt->Spinlock);
3021
3022	AssertMsgFailed(("pvObj=%p magic=%#x\n", pvObj, pObj->u32Magic));
3023	return VERR_WRONG_ORDER;
3024	}
3025
3026	/*
3027	* Preallocate the usage record if we can.
3028	*/
3029	pUsagePre = pDevExt->pUsageFree;
3030	if (pUsagePre)
3031	pDevExt->pUsageFree = pUsagePre->pNext;
3032	else if (!fNoBlocking)
3033	{
3034	RTSpinlockRelease(pDevExt->Spinlock);
3035	pUsagePre = (PSUPDRVUSAGE)RTMemAlloc(sizeof(*pUsagePre));
3036	if (!pUsagePre)
3037	return VERR_NO_MEMORY;
3038
3039	RTSpinlockAcquire(pDevExt->Spinlock);
3040	if (RT_UNLIKELY(pObj->u32Magic != SUPDRVOBJ_MAGIC))
3041	{
3042	RTSpinlockRelease(pDevExt->Spinlock);
3043
3044	AssertMsgFailed(("pvObj=%p magic=%#x\n", pvObj, pObj->u32Magic));
3045	return VERR_WRONG_ORDER;
3046	}
3047	}
3048
3049	/*
3050	* Reference the object.
3051	*/
3052	pObj->cUsage++;
3053
3054	/*
3055	* Look for the session record.
3056	*/
3057	for (pUsage = pSession->pUsage; pUsage; pUsage = pUsage->pNext)
3058	{
3059	/Log(("SUPR0AddRef: pUsage=%p:{.pObj=%p, .pNext=%p}\n", pUsage, pUsage->pObj, pUsage->pNext));/
3060	if (pUsage->pObj == pObj)
3061	break;
3062	}
3063	if (pUsage)
3064	pUsage->cUsage++;
3065	else if (pUsagePre)
3066	{
3067	/* create a new session record. */
3068	pUsagePre->cUsage = 1;
3069	pUsagePre->pObj = pObj;
3070	pUsagePre->pNext = pSession->pUsage;
3071	pSession->pUsage = pUsagePre;
3072	/Log(("SUPR0AddRef: pUsagePre=%p:{.pObj=%p, .pNext=%p}\n", pUsagePre, pUsagePre->pObj, pUsagePre->pNext));/
3073
3074	pUsagePre = NULL;
3075	}
3076	else
3077	{
3078	pObj->cUsage--;
3079	rc = VERR_TRY_AGAIN;
3080	}
3081
3082	/*
3083	* Put any unused usage record into the free list..
3084	*/
3085	if (pUsagePre)
3086	{
3087	pUsagePre->pNext = pDevExt->pUsageFree;
3088	pDevExt->pUsageFree = pUsagePre;
3089	}
3090
3091	RTSpinlockRelease(pDevExt->Spinlock);
3092
3093	return rc;
3094	}
3095	SUPR0_EXPORT_SYMBOL(SUPR0ObjAddRefEx);
3096
3097
3098	/**
3099	* Decrement / destroy a reference counter record for an object.
3100	*
3101	* The object is uniquely identified by pfnDestructor+pvUser1+pvUser2.
3102	*
3103	* @returns IPRT status code.
3104	* @retval VINF_SUCCESS if not destroyed.
3105	* @retval VINF_OBJECT_DESTROYED if it's destroyed by this release call.
3106	* @retval VERR_INVALID_PARAMETER if the object isn't valid. Will assert in
3107	* string builds.
3108	*
3109	* @param pvObj The identifier returned by SUPR0ObjRegister().
3110	* @param pSession The session which is referencing the object.
3111	*/
3112	SUPR0DECL(int) SUPR0ObjRelease(void *pvObj, PSUPDRVSESSION pSession)
3113	{
3114	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
3115	PSUPDRVOBJ pObj = (PSUPDRVOBJ)pvObj;
3116	int rc = VERR_INVALID_PARAMETER;
3117	PSUPDRVUSAGE pUsage;
3118	PSUPDRVUSAGE pUsagePrev;
3119
3120	/*
3121	* Validate the input.
3122	*/
3123	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3124	AssertMsgReturn(RT_VALID_PTR(pObj) && pObj->u32Magic == SUPDRVOBJ_MAGIC,
3125	("Invalid pvObj=%p magic=%#x (expected %#x)\n", pvObj, pObj ? pObj->u32Magic : 0, SUPDRVOBJ_MAGIC),
3126	VERR_INVALID_PARAMETER);
3127
3128	/*
3129	* Acquire the spinlock and look for the usage record.
3130	*/
3131	RTSpinlockAcquire(pDevExt->Spinlock);
3132
3133	for (pUsagePrev = NULL, pUsage = pSession->pUsage;
3134	pUsage;
3135	pUsagePrev = pUsage, pUsage = pUsage->pNext)
3136	{
3137	/Log2(("SUPR0ObjRelease: pUsage=%p:{.pObj=%p, .pNext=%p}\n", pUsage, pUsage->pObj, pUsage->pNext));/
3138	if (pUsage->pObj == pObj)
3139	{
3140	rc = VINF_SUCCESS;
3141	AssertMsg(pUsage->cUsage >= 1 && pObj->cUsage >= pUsage->cUsage, ("glob %d; sess %d\n", pObj->cUsage, pUsage->cUsage));
3142	if (pUsage->cUsage > 1)
3143	{
3144	pObj->cUsage--;
3145	pUsage->cUsage--;
3146	}
3147	else
3148	{
3149	/*
3150	* Free the session record.
3151	*/
3152	if (pUsagePrev)
3153	pUsagePrev->pNext = pUsage->pNext;
3154	else
3155	pSession->pUsage = pUsage->pNext;
3156	pUsage->pNext = pDevExt->pUsageFree;
3157	pDevExt->pUsageFree = pUsage;
3158
3159	/* What about the object? */
3160	if (pObj->cUsage > 1)
3161	pObj->cUsage--;
3162	else
3163	{
3164	/*
3165	* Object is to be destroyed, unlink it.
3166	*/
3167	pObj->u32Magic = SUPDRVOBJ_MAGIC_DEAD;
3168	rc = VINF_OBJECT_DESTROYED;
3169	if (pDevExt->pObjs == pObj)
3170	pDevExt->pObjs = pObj->pNext;
3171	else
3172	{
3173	PSUPDRVOBJ pObjPrev;
3174	for (pObjPrev = pDevExt->pObjs; pObjPrev; pObjPrev = pObjPrev->pNext)
3175	if (pObjPrev->pNext == pObj)
3176	{
3177	pObjPrev->pNext = pObj->pNext;
3178	break;
3179	}
3180	Assert(pObjPrev);
3181	}
3182	}
3183	}
3184	break;
3185	}
3186	}
3187
3188	RTSpinlockRelease(pDevExt->Spinlock);
3189
3190	/*
3191	* Call the destructor and free the object if required.
3192	*/
3193	if (rc == VINF_OBJECT_DESTROYED)
3194	{
3195	Log(("SUPR0ObjRelease: destroying %p/%d (%p/%p) cpid=%RTproc pid=%RTproc dtor=%p\n",
3196	pObj, pObj->enmType, pObj->pvUser1, pObj->pvUser2, pObj->CreatorProcess, RTProcSelf(), pObj->pfnDestructor));
3197	if (pObj->pfnDestructor)
3198	pObj->pfnDestructor(pObj, pObj->pvUser1, pObj->pvUser2);
3199	RTMemFree(pObj);
3200	}
3201
3202	AssertMsg(pUsage, ("pvObj=%p\n", pvObj));
3203	return rc;
3204	}
3205	SUPR0_EXPORT_SYMBOL(SUPR0ObjRelease);
3206
3207
3208	/**
3209	* Verifies that the current process can access the specified object.
3210	*
3211	* @returns The following IPRT status code:
3212	* @retval VINF_SUCCESS if access was granted.
3213	* @retval VERR_PERMISSION_DENIED if denied access.
3214	* @retval VERR_INVALID_PARAMETER if invalid parameter.
3215	*
3216	* @param pvObj The identifier returned by SUPR0ObjRegister().
3217	* @param pSession The session which wishes to access the object.
3218	* @param pszObjName Object string name. This is optional and depends on the object type.
3219	*
3220	* @remark The caller is responsible for making sure the object isn't removed while
3221	* we're inside this function. If uncertain about this, just call AddRef before calling us.
3222	*/
3223	SUPR0DECL(int) SUPR0ObjVerifyAccess(void pvObj, PSUPDRVSESSION pSession, const char pszObjName)
3224	{
3225	PSUPDRVOBJ pObj = (PSUPDRVOBJ)pvObj;
3226	int rc;
3227
3228	/*
3229	* Validate the input.
3230	*/
3231	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3232	AssertMsgReturn(RT_VALID_PTR(pObj) && pObj->u32Magic == SUPDRVOBJ_MAGIC,
3233	("Invalid pvObj=%p magic=%#x (exepcted %#x)\n", pvObj, pObj ? pObj->u32Magic : 0, SUPDRVOBJ_MAGIC),
3234	VERR_INVALID_PARAMETER);
3235
3236	/*
3237	* Check access. (returns true if a decision has been made.)
3238	*/
3239	rc = VERR_INTERNAL_ERROR;
3240	if (supdrvOSObjCanAccess(pObj, pSession, pszObjName, &rc))
3241	return rc;
3242
3243	/*
3244	* Default policy is to allow the user to access his own
3245	* stuff but nothing else.
3246	*/
3247	if (pObj->CreatorUid == pSession->Uid)
3248	return VINF_SUCCESS;
3249	return VERR_PERMISSION_DENIED;
3250	}
3251	SUPR0_EXPORT_SYMBOL(SUPR0ObjVerifyAccess);
3252
3253
3254	/**
3255	* API for the VMMR0 module to get the SUPDRVSESSION::pSessionVM member.
3256	*
3257	* @returns The associated VM pointer.
3258	* @param pSession The session of the current thread.
3259	*/
3260	SUPR0DECL(PVM) SUPR0GetSessionVM(PSUPDRVSESSION pSession)
3261	{
3262	AssertReturn(SUP_IS_SESSION_VALID(pSession), NULL);
3263	return pSession->pSessionVM;
3264	}
3265	SUPR0_EXPORT_SYMBOL(SUPR0GetSessionVM);
3266
3267
3268	/**
3269	* API for the VMMR0 module to get the SUPDRVSESSION::pSessionGVM member.
3270	*
3271	* @returns The associated GVM pointer.
3272	* @param pSession The session of the current thread.
3273	*/
3274	SUPR0DECL(PGVM) SUPR0GetSessionGVM(PSUPDRVSESSION pSession)
3275	{
3276	AssertReturn(SUP_IS_SESSION_VALID(pSession), NULL);
3277	return pSession->pSessionGVM;
3278	}
3279	SUPR0_EXPORT_SYMBOL(SUPR0GetSessionGVM);
3280
3281
3282	/**
3283	* API for the VMMR0 module to work the SUPDRVSESSION::pSessionVM member.
3284	*
3285	* This will fail if there is already a VM associated with the session and pVM
3286	* isn't NULL.
3287	*
3288	* @retval VINF_SUCCESS
3289	* @retval VERR_ALREADY_EXISTS if there already is a VM associated with the
3290	* session.
3291	* @retval VERR_INVALID_PARAMETER if only one of the parameters are NULL or if
3292	* the session is invalid.
3293	*
3294	* @param pSession The session of the current thread.
3295	* @param pGVM The GVM to associate with the session. Pass NULL to
3296	* dissassociate.
3297	* @param pVM The VM to associate with the session. Pass NULL to
3298	* dissassociate.
3299	*/
3300	SUPR0DECL(int) SUPR0SetSessionVM(PSUPDRVSESSION pSession, PGVM pGVM, PVM pVM)
3301	{
3302	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3303	AssertReturn((pGVM != NULL) == (pVM != NULL), VERR_INVALID_PARAMETER);
3304
3305	RTSpinlockAcquire(pSession->pDevExt->Spinlock);
3306	if (pGVM)
3307	{
3308	if (!pSession->pSessionGVM)
3309	{
3310	pSession->pSessionGVM = pGVM;
3311	pSession->pSessionVM = pVM;
3312	pSession->pFastIoCtrlVM = NULL;
3313	}
3314	else
3315	{
3316	RTSpinlockRelease(pSession->pDevExt->Spinlock);
3317	SUPR0Printf("SUPR0SetSessionVM: Unable to associated GVM/VM %p/%p with session %p as it has %p/%p already!\n",
3318	pGVM, pVM, pSession, pSession->pSessionGVM, pSession->pSessionVM);
3319	return VERR_ALREADY_EXISTS;
3320	}
3321	}
3322	else
3323	{
3324	pSession->pSessionGVM = NULL;
3325	pSession->pSessionVM = NULL;
3326	pSession->pFastIoCtrlVM = NULL;
3327	}
3328	RTSpinlockRelease(pSession->pDevExt->Spinlock);
3329	return VINF_SUCCESS;
3330	}
3331	SUPR0_EXPORT_SYMBOL(SUPR0SetSessionVM);
3332
3333
3334	/**
3335	* For getting SUPDRVSESSION::Uid.
3336	*
3337	* @returns The session UID. NIL_RTUID if invalid pointer or not successfully
3338	* set by the host code.
3339	* @param pSession The session of the current thread.
3340	*/
3341	SUPR0DECL(RTUID) SUPR0GetSessionUid(PSUPDRVSESSION pSession)
3342	{
3343	AssertReturn(SUP_IS_SESSION_VALID(pSession), NIL_RTUID);
3344	return pSession->Uid;
3345	}
3346	SUPR0_EXPORT_SYMBOL(SUPR0GetSessionUid);
3347
3348
3349	/** @copydoc RTLogDefaultInstanceEx
3350	* @remarks To allow overriding RTLogDefaultInstanceEx locally. */
3351	SUPR0DECL(struct RTLOGGER *) SUPR0DefaultLogInstanceEx(uint32_t fFlagsAndGroup)
3352	{
3353	return RTLogDefaultInstanceEx(fFlagsAndGroup);
3354	}
3355	SUPR0_EXPORT_SYMBOL(SUPR0DefaultLogInstanceEx);
3356
3357
3358	/** @copydoc RTLogGetDefaultInstanceEx
3359	* @remarks To allow overriding RTLogGetDefaultInstanceEx locally. */
3360	SUPR0DECL(struct RTLOGGER *) SUPR0GetDefaultLogInstanceEx(uint32_t fFlagsAndGroup)
3361	{
3362	return RTLogGetDefaultInstanceEx(fFlagsAndGroup);
3363	}
3364	SUPR0_EXPORT_SYMBOL(SUPR0GetDefaultLogInstanceEx);
3365
3366
3367	/** @copydoc RTLogRelGetDefaultInstanceEx
3368	* @remarks To allow overriding RTLogRelGetDefaultInstanceEx locally. */
3369	SUPR0DECL(struct RTLOGGER *) SUPR0GetDefaultLogRelInstanceEx(uint32_t fFlagsAndGroup)
3370	{
3371	return RTLogRelGetDefaultInstanceEx(fFlagsAndGroup);
3372	}
3373	SUPR0_EXPORT_SYMBOL(SUPR0GetDefaultLogRelInstanceEx);
3374
3375
3376	/**
3377	* Lock pages.
3378	*
3379	* @returns IPRT status code.
3380	* @param pSession Session to which the locked memory should be associated.
3381	* @param pvR3 Start of the memory range to lock.
3382	* This must be page aligned.
3383	* @param cPages Number of pages to lock.
3384	* @param paPages Where to put the physical addresses of locked memory.
3385	*/
3386	SUPR0DECL(int) SUPR0LockMem(PSUPDRVSESSION pSession, RTR3PTR pvR3, uint32_t cPages, PRTHCPHYS paPages)
3387	{
3388	int rc;
3389	SUPDRVMEMREF Mem = { NIL_RTR0MEMOBJ, NIL_RTR0MEMOBJ, MEMREF_TYPE_UNUSED };
3390	const size_t cb = (size_t)cPages << PAGE_SHIFT;
3391	LogFlow(("SUPR0LockMem: pSession=%p pvR3=%p cPages=%d paPages=%p\n", pSession, (void *)pvR3, cPages, paPages));
3392
3393	/*
3394	* Verify input.
3395	*/
3396	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3397	AssertPtrReturn(paPages, VERR_INVALID_PARAMETER);
3398	if ( RT_ALIGN_R3PT(pvR3, PAGE_SIZE, RTR3PTR) != pvR3
3399	\|\| !pvR3)
3400	{
3401	Log(("pvR3 (%p) must be page aligned and not NULL!\n", (void *)pvR3));
3402	return VERR_INVALID_PARAMETER;
3403	}
3404
3405	/*
3406	* Let IPRT do the job.
3407	*/
3408	Mem.eType = MEMREF_TYPE_LOCKED;
3409	rc = RTR0MemObjLockUser(&Mem.MemObj, pvR3, cb, RTMEM_PROT_READ \| RTMEM_PROT_WRITE, NIL_RTR0PROCESS);
3410	if (RT_SUCCESS(rc))
3411	{
3412	uint32_t iPage = cPages;
3413	AssertMsg(RTR0MemObjAddressR3(Mem.MemObj) == pvR3, ("%p == %p\n", RTR0MemObjAddressR3(Mem.MemObj), pvR3));
3414	AssertMsg(RTR0MemObjSize(Mem.MemObj) == cb, ("%x == %x\n", RTR0MemObjSize(Mem.MemObj), cb));
3415
3416	while (iPage-- > 0)
3417	{
3418	paPages[iPage] = RTR0MemObjGetPagePhysAddr(Mem.MemObj, iPage);
3419	if (RT_UNLIKELY(paPages[iPage] == NIL_RTCCPHYS))
3420	{
3421	AssertMsgFailed(("iPage=%d\n", iPage));
3422	rc = VERR_INTERNAL_ERROR;
3423	break;
3424	}
3425	}
3426	if (RT_SUCCESS(rc))
3427	rc = supdrvMemAdd(&Mem, pSession);
3428	if (RT_FAILURE(rc))
3429	{
3430	int rc2 = RTR0MemObjFree(Mem.MemObj, false);
3431	AssertRC(rc2);
3432	}
3433	}
3434
3435	return rc;
3436	}
3437	SUPR0_EXPORT_SYMBOL(SUPR0LockMem);
3438
3439
3440	/**
3441	* Unlocks the memory pointed to by pv.
3442	*
3443	* @returns IPRT status code.
3444	* @param pSession Session to which the memory was locked.
3445	* @param pvR3 Memory to unlock.
3446	*/
3447	SUPR0DECL(int) SUPR0UnlockMem(PSUPDRVSESSION pSession, RTR3PTR pvR3)
3448	{
3449	LogFlow(("SUPR0UnlockMem: pSession=%p pvR3=%p\n", pSession, (void *)pvR3));
3450	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3451	return supdrvMemRelease(pSession, (RTHCUINTPTR)pvR3, MEMREF_TYPE_LOCKED);
3452	}
3453	SUPR0_EXPORT_SYMBOL(SUPR0UnlockMem);
3454
3455
3456	/**
3457	* Allocates a chunk of page aligned memory with contiguous and fixed physical
3458	* backing.
3459	*
3460	* @returns IPRT status code.
3461	* @param pSession Session data.
3462	* @param cPages Number of pages to allocate.
3463	* @param ppvR0 Where to put the address of Ring-0 mapping the allocated memory.
3464	* @param ppvR3 Where to put the address of Ring-3 mapping the allocated memory.
3465	* @param pHCPhys Where to put the physical address of allocated memory.
3466	*/
3467	SUPR0DECL(int) SUPR0ContAlloc(PSUPDRVSESSION pSession, uint32_t cPages, PRTR0PTR ppvR0, PRTR3PTR ppvR3, PRTHCPHYS pHCPhys)
3468	{
3469	int rc;
3470	SUPDRVMEMREF Mem = { NIL_RTR0MEMOBJ, NIL_RTR0MEMOBJ, MEMREF_TYPE_UNUSED };
3471	LogFlow(("SUPR0ContAlloc: pSession=%p cPages=%d ppvR0=%p ppvR3=%p pHCPhys=%p\n", pSession, cPages, ppvR0, ppvR3, pHCPhys));
3472
3473	/*
3474	* Validate input.
3475	*/
3476	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3477	if (!ppvR3 \|\| !ppvR0 \|\| !pHCPhys)
3478	{
3479	Log(("Null pointer. All of these should be set: pSession=%p ppvR0=%p ppvR3=%p pHCPhys=%p\n",
3480	pSession, ppvR0, ppvR3, pHCPhys));
3481	return VERR_INVALID_PARAMETER;
3482
3483	}
3484	if (cPages < 1 \|\| cPages >= 256)
3485	{
3486	Log(("Illegal request cPages=%d, must be greater than 0 and smaller than 256.\n", cPages));
3487	return VERR_PAGE_COUNT_OUT_OF_RANGE;
3488	}
3489
3490	/*
3491	* Let IPRT do the job.
3492	*/
3493	rc = RTR0MemObjAllocCont(&Mem.MemObj, cPages << PAGE_SHIFT, true /* executable R0 mapping */);
3494	if (RT_SUCCESS(rc))
3495	{
3496	int rc2;
3497	rc = RTR0MemObjMapUser(&Mem.MapObjR3, Mem.MemObj, (RTR3PTR)-1, 0,
3498	RTMEM_PROT_EXEC \| RTMEM_PROT_WRITE \| RTMEM_PROT_READ, NIL_RTR0PROCESS);
3499	if (RT_SUCCESS(rc))
3500	{
3501	Mem.eType = MEMREF_TYPE_CONT;
3502	rc = supdrvMemAdd(&Mem, pSession);
3503	if (!rc)
3504	{
3505	*ppvR0 = RTR0MemObjAddress(Mem.MemObj);
3506	*ppvR3 = RTR0MemObjAddressR3(Mem.MapObjR3);
3507	*pHCPhys = RTR0MemObjGetPagePhysAddr(Mem.MemObj, 0);
3508	return 0;
3509	}
3510
3511	rc2 = RTR0MemObjFree(Mem.MapObjR3, false);
3512	AssertRC(rc2);
3513	}
3514	rc2 = RTR0MemObjFree(Mem.MemObj, false);
3515	AssertRC(rc2);
3516	}
3517
3518	return rc;
3519	}
3520	SUPR0_EXPORT_SYMBOL(SUPR0ContAlloc);
3521
3522
3523	/**
3524	* Frees memory allocated using SUPR0ContAlloc().
3525	*
3526	* @returns IPRT status code.
3527	* @param pSession The session to which the memory was allocated.
3528	* @param uPtr Pointer to the memory (ring-3 or ring-0).
3529	*/
3530	SUPR0DECL(int) SUPR0ContFree(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr)
3531	{
3532	LogFlow(("SUPR0ContFree: pSession=%p uPtr=%p\n", pSession, (void *)uPtr));
3533	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3534	return supdrvMemRelease(pSession, uPtr, MEMREF_TYPE_CONT);
3535	}
3536	SUPR0_EXPORT_SYMBOL(SUPR0ContFree);
3537
3538
3539	/**
3540	* Allocates a chunk of page aligned memory with fixed physical backing below 4GB.
3541	*
3542	* The memory isn't zeroed.
3543	*
3544	* @returns IPRT status code.
3545	* @param pSession Session data.
3546	* @param cPages Number of pages to allocate.
3547	* @param ppvR0 Where to put the address of Ring-0 mapping of the allocated memory.
3548	* @param ppvR3 Where to put the address of Ring-3 mapping of the allocated memory.
3549	* @param paPages Where to put the physical addresses of allocated memory.
3550	*/
3551	SUPR0DECL(int) SUPR0LowAlloc(PSUPDRVSESSION pSession, uint32_t cPages, PRTR0PTR ppvR0, PRTR3PTR ppvR3, PRTHCPHYS paPages)
3552	{
3553	unsigned iPage;
3554	int rc;
3555	SUPDRVMEMREF Mem = { NIL_RTR0MEMOBJ, NIL_RTR0MEMOBJ, MEMREF_TYPE_UNUSED };
3556	LogFlow(("SUPR0LowAlloc: pSession=%p cPages=%d ppvR3=%p ppvR0=%p paPages=%p\n", pSession, cPages, ppvR3, ppvR0, paPages));
3557
3558	/*
3559	* Validate input.
3560	*/
3561	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3562	if (!ppvR3 \|\| !ppvR0 \|\| !paPages)
3563	{
3564	Log(("Null pointer. All of these should be set: pSession=%p ppvR3=%p ppvR0=%p paPages=%p\n",
3565	pSession, ppvR3, ppvR0, paPages));
3566	return VERR_INVALID_PARAMETER;
3567
3568	}
3569	if (cPages < 1 \|\| cPages >= 256)
3570	{
3571	Log(("Illegal request cPages=%d, must be greater than 0 and smaller than 256.\n", cPages));
3572	return VERR_PAGE_COUNT_OUT_OF_RANGE;
3573	}
3574
3575	/*
3576	* Let IPRT do the work.
3577	*/
3578	rc = RTR0MemObjAllocLow(&Mem.MemObj, cPages << PAGE_SHIFT, true /* executable ring-0 mapping */);
3579	if (RT_SUCCESS(rc))
3580	{
3581	int rc2;
3582	rc = RTR0MemObjMapUser(&Mem.MapObjR3, Mem.MemObj, (RTR3PTR)-1, 0,
3583	RTMEM_PROT_EXEC \| RTMEM_PROT_WRITE \| RTMEM_PROT_READ, NIL_RTR0PROCESS);
3584	if (RT_SUCCESS(rc))
3585	{
3586	Mem.eType = MEMREF_TYPE_LOW;
3587	rc = supdrvMemAdd(&Mem, pSession);
3588	if (!rc)
3589	{
3590	for (iPage = 0; iPage < cPages; iPage++)
3591	{
3592	paPages[iPage] = RTR0MemObjGetPagePhysAddr(Mem.MemObj, iPage);
3593	AssertMsg(!(paPages[iPage] & (PAGE_SIZE - 1)), ("iPage=%d Phys=%RHp\n", paPages[iPage]));
3594	}
3595	*ppvR0 = RTR0MemObjAddress(Mem.MemObj);
3596	*ppvR3 = RTR0MemObjAddressR3(Mem.MapObjR3);
3597	return 0;
3598	}
3599
3600	rc2 = RTR0MemObjFree(Mem.MapObjR3, false);
3601	AssertRC(rc2);
3602	}
3603
3604	rc2 = RTR0MemObjFree(Mem.MemObj, false);
3605	AssertRC(rc2);
3606	}
3607
3608	return rc;
3609	}
3610	SUPR0_EXPORT_SYMBOL(SUPR0LowAlloc);
3611
3612
3613	/**
3614	* Frees memory allocated using SUPR0LowAlloc().
3615	*
3616	* @returns IPRT status code.
3617	* @param pSession The session to which the memory was allocated.
3618	* @param uPtr Pointer to the memory (ring-3 or ring-0).
3619	*/
3620	SUPR0DECL(int) SUPR0LowFree(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr)
3621	{
3622	LogFlow(("SUPR0LowFree: pSession=%p uPtr=%p\n", pSession, (void *)uPtr));
3623	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3624	return supdrvMemRelease(pSession, uPtr, MEMREF_TYPE_LOW);
3625	}
3626	SUPR0_EXPORT_SYMBOL(SUPR0LowFree);
3627
3628
3629
3630	/**
3631	* Allocates a chunk of memory with both R0 and R3 mappings.
3632	* The memory is fixed and it's possible to query the physical addresses using SUPR0MemGetPhys().
3633	*
3634	* @returns IPRT status code.
3635	* @param pSession The session to associated the allocation with.
3636	* @param cb Number of bytes to allocate.
3637	* @param ppvR0 Where to store the address of the Ring-0 mapping.
3638	* @param ppvR3 Where to store the address of the Ring-3 mapping.
3639	*/
3640	SUPR0DECL(int) SUPR0MemAlloc(PSUPDRVSESSION pSession, uint32_t cb, PRTR0PTR ppvR0, PRTR3PTR ppvR3)
3641	{
3642	int rc;
3643	SUPDRVMEMREF Mem = { NIL_RTR0MEMOBJ, NIL_RTR0MEMOBJ, MEMREF_TYPE_UNUSED };
3644	LogFlow(("SUPR0MemAlloc: pSession=%p cb=%d ppvR0=%p ppvR3=%p\n", pSession, cb, ppvR0, ppvR3));
3645
3646	/*
3647	* Validate input.
3648	*/
3649	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3650	AssertPtrReturn(ppvR0, VERR_INVALID_POINTER);
3651	AssertPtrReturn(ppvR3, VERR_INVALID_POINTER);
3652	if (cb < 1 \|\| cb >= _4M)
3653	{
3654	Log(("Illegal request cb=%u; must be greater than 0 and smaller than 4MB.\n", cb));
3655	return VERR_INVALID_PARAMETER;
3656	}
3657
3658	/*
3659	* Let IPRT do the work.
3660	*/
3661	rc = RTR0MemObjAllocPage(&Mem.MemObj, cb, true /* executable ring-0 mapping */);
3662	if (RT_SUCCESS(rc))
3663	{
3664	int rc2;
3665	rc = RTR0MemObjMapUser(&Mem.MapObjR3, Mem.MemObj, (RTR3PTR)-1, 0,
3666	RTMEM_PROT_EXEC \| RTMEM_PROT_WRITE \| RTMEM_PROT_READ, NIL_RTR0PROCESS);
3667	if (RT_SUCCESS(rc))
3668	{
3669	Mem.eType = MEMREF_TYPE_MEM;
3670	rc = supdrvMemAdd(&Mem, pSession);
3671	if (!rc)
3672	{
3673	*ppvR0 = RTR0MemObjAddress(Mem.MemObj);
3674	*ppvR3 = RTR0MemObjAddressR3(Mem.MapObjR3);
3675	return VINF_SUCCESS;
3676	}
3677
3678	rc2 = RTR0MemObjFree(Mem.MapObjR3, false);
3679	AssertRC(rc2);
3680	}
3681
3682	rc2 = RTR0MemObjFree(Mem.MemObj, false);
3683	AssertRC(rc2);
3684	}
3685
3686	return rc;
3687	}
3688	SUPR0_EXPORT_SYMBOL(SUPR0MemAlloc);
3689
3690
3691	/**
3692	* Get the physical addresses of memory allocated using SUPR0MemAlloc().
3693	*
3694	* @returns IPRT status code.
3695	* @param pSession The session to which the memory was allocated.
3696	* @param uPtr The Ring-0 or Ring-3 address returned by SUPR0MemAlloc().
3697	* @param paPages Where to store the physical addresses.
3698	*/
3699	SUPR0DECL(int) SUPR0MemGetPhys(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr, PSUPPAGE paPages) /** @todo switch this bugger to RTHCPHYS */
3700	{
3701	PSUPDRVBUNDLE pBundle;
3702	LogFlow(("SUPR0MemGetPhys: pSession=%p uPtr=%p paPages=%p\n", pSession, (void *)uPtr, paPages));
3703
3704	/*
3705	* Validate input.
3706	*/
3707	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3708	AssertPtrReturn(paPages, VERR_INVALID_POINTER);
3709	AssertReturn(uPtr, VERR_INVALID_PARAMETER);
3710
3711	/*
3712	* Search for the address.
3713	*/
3714	RTSpinlockAcquire(pSession->Spinlock);
3715	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
3716	{
3717	if (pBundle->cUsed > 0)
3718	{
3719	unsigned i;
3720	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
3721	{
3722	if ( pBundle->aMem[i].eType == MEMREF_TYPE_MEM
3723	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
3724	&& ( (RTHCUINTPTR)RTR0MemObjAddress(pBundle->aMem[i].MemObj) == uPtr
3725	\|\| ( pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ
3726	&& RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3) == uPtr)
3727	)
3728	)
3729	{
3730	const size_t cPages = RTR0MemObjSize(pBundle->aMem[i].MemObj) >> PAGE_SHIFT;
3731	size_t iPage;
3732	for (iPage = 0; iPage < cPages; iPage++)
3733	{
3734	paPages[iPage].Phys = RTR0MemObjGetPagePhysAddr(pBundle->aMem[i].MemObj, iPage);
3735	paPages[iPage].uReserved = 0;
3736	}
3737	RTSpinlockRelease(pSession->Spinlock);
3738	return VINF_SUCCESS;
3739	}
3740	}
3741	}
3742	}
3743	RTSpinlockRelease(pSession->Spinlock);
3744	Log(("Failed to find %p!!!\n", (void *)uPtr));
3745	return VERR_INVALID_PARAMETER;
3746	}
3747	SUPR0_EXPORT_SYMBOL(SUPR0MemGetPhys);
3748
3749
3750	/**
3751	* Free memory allocated by SUPR0MemAlloc().
3752	*
3753	* @returns IPRT status code.
3754	* @param pSession The session owning the allocation.
3755	* @param uPtr The Ring-0 or Ring-3 address returned by SUPR0MemAlloc().
3756	*/
3757	SUPR0DECL(int) SUPR0MemFree(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr)
3758	{
3759	LogFlow(("SUPR0MemFree: pSession=%p uPtr=%p\n", pSession, (void *)uPtr));
3760	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3761	return supdrvMemRelease(pSession, uPtr, MEMREF_TYPE_MEM);
3762	}
3763	SUPR0_EXPORT_SYMBOL(SUPR0MemFree);
3764
3765
3766	/**
3767	* Allocates a chunk of memory with a kernel or/and a user mode mapping.
3768	*
3769	* The memory is fixed and it's possible to query the physical addresses using
3770	* SUPR0MemGetPhys().
3771	*
3772	* @returns IPRT status code.
3773	* @param pSession The session to associated the allocation with.
3774	* @param cPages The number of pages to allocate.
3775	* @param fFlags Flags, reserved for the future. Must be zero.
3776	* @param ppvR3 Where to store the address of the Ring-3 mapping.
3777	* NULL if no ring-3 mapping.
3778	* @param ppvR0 Where to store the address of the Ring-0 mapping.
3779	* NULL if no ring-0 mapping.
3780	* @param paPages Where to store the addresses of the pages. Optional.
3781	*/
3782	SUPR0DECL(int) SUPR0PageAllocEx(PSUPDRVSESSION pSession, uint32_t cPages, uint32_t fFlags, PRTR3PTR ppvR3, PRTR0PTR ppvR0, PRTHCPHYS paPages)
3783	{
3784	int rc;
3785	SUPDRVMEMREF Mem = { NIL_RTR0MEMOBJ, NIL_RTR0MEMOBJ, MEMREF_TYPE_UNUSED };
3786	LogFlow(("SUPR0PageAlloc: pSession=%p cb=%d ppvR3=%p\n", pSession, cPages, ppvR3));
3787
3788	/*
3789	* Validate input. The allowed allocation size must be at least equal to the maximum guest VRAM size.
3790	*/
3791	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3792	AssertPtrNullReturn(ppvR3, VERR_INVALID_POINTER);
3793	AssertPtrNullReturn(ppvR0, VERR_INVALID_POINTER);
3794	AssertReturn(ppvR3 \|\| ppvR0, VERR_INVALID_PARAMETER);
3795	AssertReturn(!fFlags, VERR_INVALID_PARAMETER);
3796	if (cPages < 1 \|\| cPages > VBOX_MAX_ALLOC_PAGE_COUNT)
3797	{
3798	Log(("SUPR0PageAlloc: Illegal request cb=%u; must be greater than 0 and smaller than %uMB (VBOX_MAX_ALLOC_PAGE_COUNT pages).\n", cPages, VBOX_MAX_ALLOC_PAGE_COUNT * (_1M / _4K)));
3799	return VERR_PAGE_COUNT_OUT_OF_RANGE;
3800	}
3801
3802	/*
3803	* Let IPRT do the work.
3804	*/
3805	if (ppvR0)
3806	rc = RTR0MemObjAllocPage(&Mem.MemObj, (size_t)cPages * PAGE_SIZE, false /fExecutable/);
3807	else
3808	rc = RTR0MemObjAllocPhysNC(&Mem.MemObj, (size_t)cPages * PAGE_SIZE, NIL_RTHCPHYS);
3809	if (RT_SUCCESS(rc))
3810	{
3811	int rc2;
3812	if (ppvR3)
3813	rc = RTR0MemObjMapUser(&Mem.MapObjR3, Mem.MemObj, (RTR3PTR)-1, 0, RTMEM_PROT_WRITE \| RTMEM_PROT_READ, NIL_RTR0PROCESS);
3814	else
3815	Mem.MapObjR3 = NIL_RTR0MEMOBJ;
3816	if (RT_SUCCESS(rc))
3817	{
3818	Mem.eType = MEMREF_TYPE_PAGE;
3819	rc = supdrvMemAdd(&Mem, pSession);
3820	if (!rc)
3821	{
3822	if (ppvR3)
3823	*ppvR3 = RTR0MemObjAddressR3(Mem.MapObjR3);
3824	if (ppvR0)
3825	*ppvR0 = RTR0MemObjAddress(Mem.MemObj);
3826	if (paPages)
3827	{
3828	uint32_t iPage = cPages;
3829	while (iPage-- > 0)
3830	{
3831	paPages[iPage] = RTR0MemObjGetPagePhysAddr(Mem.MapObjR3, iPage);
3832	Assert(paPages[iPage] != NIL_RTHCPHYS);
3833	}
3834	}
3835	return VINF_SUCCESS;
3836	}
3837
3838	rc2 = RTR0MemObjFree(Mem.MapObjR3, false);
3839	AssertRC(rc2);
3840	}
3841
3842	rc2 = RTR0MemObjFree(Mem.MemObj, false);
3843	AssertRC(rc2);
3844	}
3845	return rc;
3846	}
3847	SUPR0_EXPORT_SYMBOL(SUPR0PageAllocEx);
3848
3849
3850	/**
3851	* Maps a chunk of memory previously allocated by SUPR0PageAllocEx into kernel
3852	* space.
3853	*
3854	* @returns IPRT status code.
3855	* @param pSession The session to associated the allocation with.
3856	* @param pvR3 The ring-3 address returned by SUPR0PageAllocEx.
3857	* @param offSub Where to start mapping. Must be page aligned.
3858	* @param cbSub How much to map. Must be page aligned.
3859	* @param fFlags Flags, MBZ.
3860	* @param ppvR0 Where to return the address of the ring-0 mapping on
3861	* success.
3862	*/
3863	SUPR0DECL(int) SUPR0PageMapKernel(PSUPDRVSESSION pSession, RTR3PTR pvR3, uint32_t offSub, uint32_t cbSub,
3864	uint32_t fFlags, PRTR0PTR ppvR0)
3865	{
3866	int rc;
3867	PSUPDRVBUNDLE pBundle;
3868	RTR0MEMOBJ hMemObj = NIL_RTR0MEMOBJ;
3869	LogFlow(("SUPR0PageMapKernel: pSession=%p pvR3=%p offSub=%#x cbSub=%#x\n", pSession, pvR3, offSub, cbSub));
3870
3871	/*
3872	* Validate input. The allowed allocation size must be at least equal to the maximum guest VRAM size.
3873	*/
3874	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3875	AssertPtrNullReturn(ppvR0, VERR_INVALID_POINTER);
3876	AssertReturn(!fFlags, VERR_INVALID_PARAMETER);
3877	AssertReturn(!(offSub & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
3878	AssertReturn(!(cbSub & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
3879	AssertReturn(cbSub, VERR_INVALID_PARAMETER);
3880
3881	/*
3882	* Find the memory object.
3883	*/
3884	RTSpinlockAcquire(pSession->Spinlock);
3885	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
3886	{
3887	if (pBundle->cUsed > 0)
3888	{
3889	unsigned i;
3890	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
3891	{
3892	if ( ( pBundle->aMem[i].eType == MEMREF_TYPE_PAGE
3893	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
3894	&& pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ
3895	&& RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3) == pvR3)
3896	\|\| ( pBundle->aMem[i].eType == MEMREF_TYPE_LOCKED
3897	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
3898	&& pBundle->aMem[i].MapObjR3 == NIL_RTR0MEMOBJ
3899	&& RTR0MemObjAddressR3(pBundle->aMem[i].MemObj) == pvR3))
3900	{
3901	hMemObj = pBundle->aMem[i].MemObj;
3902	break;
3903	}
3904	}
3905	}
3906	}
3907	RTSpinlockRelease(pSession->Spinlock);
3908
3909	rc = VERR_INVALID_PARAMETER;
3910	if (hMemObj != NIL_RTR0MEMOBJ)
3911	{
3912	/*
3913	* Do some further input validations before calling IPRT.
3914	* (Cleanup is done indirectly by telling RTR0MemObjFree to include mappings.)
3915	*/
3916	size_t cbMemObj = RTR0MemObjSize(hMemObj);
3917	if ( offSub < cbMemObj
3918	&& cbSub <= cbMemObj
3919	&& offSub + cbSub <= cbMemObj)
3920	{
3921	RTR0MEMOBJ hMapObj;
3922	rc = RTR0MemObjMapKernelEx(&hMapObj, hMemObj, (void *)-1, 0,
3923	RTMEM_PROT_READ \| RTMEM_PROT_WRITE, offSub, cbSub);
3924	if (RT_SUCCESS(rc))
3925	*ppvR0 = RTR0MemObjAddress(hMapObj);
3926	}
3927	else
3928	SUPR0Printf("SUPR0PageMapKernel: cbMemObj=%#x offSub=%#x cbSub=%#x\n", cbMemObj, offSub, cbSub);
3929
3930	}
3931	return rc;
3932	}
3933	SUPR0_EXPORT_SYMBOL(SUPR0PageMapKernel);
3934
3935
3936	/**
3937	* Changes the page level protection of one or more pages previously allocated
3938	* by SUPR0PageAllocEx.
3939	*
3940	* @returns IPRT status code.
3941	* @param pSession The session to associated the allocation with.
3942	* @param pvR3 The ring-3 address returned by SUPR0PageAllocEx.
3943	* NIL_RTR3PTR if the ring-3 mapping should be unaffected.
3944	* @param pvR0 The ring-0 address returned by SUPR0PageAllocEx.
3945	* NIL_RTR0PTR if the ring-0 mapping should be unaffected.
3946	* @param offSub Where to start changing. Must be page aligned.
3947	* @param cbSub How much to change. Must be page aligned.
3948	* @param fProt The new page level protection, see RTMEM_PROT_*.
3949	*/
3950	SUPR0DECL(int) SUPR0PageProtect(PSUPDRVSESSION pSession, RTR3PTR pvR3, RTR0PTR pvR0, uint32_t offSub, uint32_t cbSub, uint32_t fProt)
3951	{
3952	int rc;
3953	PSUPDRVBUNDLE pBundle;
3954	RTR0MEMOBJ hMemObjR0 = NIL_RTR0MEMOBJ;
3955	RTR0MEMOBJ hMemObjR3 = NIL_RTR0MEMOBJ;
3956	LogFlow(("SUPR0PageProtect: pSession=%p pvR3=%p pvR0=%p offSub=%#x cbSub=%#x fProt-%#x\n", pSession, pvR3, pvR0, offSub, cbSub, fProt));
3957
3958	/*
3959	* Validate input. The allowed allocation size must be at least equal to the maximum guest VRAM size.
3960	*/
3961	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3962	AssertReturn(!(fProt & ~(RTMEM_PROT_READ \| RTMEM_PROT_WRITE \| RTMEM_PROT_EXEC \| RTMEM_PROT_NONE)), VERR_INVALID_PARAMETER);
3963	AssertReturn(!(offSub & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
3964	AssertReturn(!(cbSub & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
3965	AssertReturn(cbSub, VERR_INVALID_PARAMETER);
3966
3967	/*
3968	* Find the memory object.
3969	*/
3970	RTSpinlockAcquire(pSession->Spinlock);
3971	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
3972	{
3973	if (pBundle->cUsed > 0)
3974	{
3975	unsigned i;
3976	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
3977	{
3978	if ( pBundle->aMem[i].eType == MEMREF_TYPE_PAGE
3979	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
3980	&& ( pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ
3981	\|\| pvR3 == NIL_RTR3PTR)
3982	&& ( pvR0 == NIL_RTR0PTR
3983	\|\| RTR0MemObjAddress(pBundle->aMem[i].MemObj) == pvR0)
3984	&& ( pvR3 == NIL_RTR3PTR
3985	\|\| RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3) == pvR3))
3986	{
3987	if (pvR0 != NIL_RTR0PTR)
3988	hMemObjR0 = pBundle->aMem[i].MemObj;
3989	if (pvR3 != NIL_RTR3PTR)
3990	hMemObjR3 = pBundle->aMem[i].MapObjR3;
3991	break;
3992	}
3993	}
3994	}
3995	}
3996	RTSpinlockRelease(pSession->Spinlock);
3997
3998	rc = VERR_INVALID_PARAMETER;
3999	if ( hMemObjR0 != NIL_RTR0MEMOBJ
4000	\|\| hMemObjR3 != NIL_RTR0MEMOBJ)
4001	{
4002	/*
4003	* Do some further input validations before calling IPRT.
4004	*/
4005	size_t cbMemObj = hMemObjR0 != NIL_RTR0PTR ? RTR0MemObjSize(hMemObjR0) : RTR0MemObjSize(hMemObjR3);
4006	if ( offSub < cbMemObj
4007	&& cbSub <= cbMemObj
4008	&& offSub + cbSub <= cbMemObj)
4009	{
4010	rc = VINF_SUCCESS;
4011	if (hMemObjR3 != NIL_RTR0PTR)
4012	rc = RTR0MemObjProtect(hMemObjR3, offSub, cbSub, fProt);
4013	if (hMemObjR0 != NIL_RTR0PTR && RT_SUCCESS(rc))
4014	rc = RTR0MemObjProtect(hMemObjR0, offSub, cbSub, fProt);
4015	}
4016	else
4017	SUPR0Printf("SUPR0PageMapKernel: cbMemObj=%#x offSub=%#x cbSub=%#x\n", cbMemObj, offSub, cbSub);
4018
4019	}
4020	return rc;
4021
4022	}
4023	SUPR0_EXPORT_SYMBOL(SUPR0PageProtect);
4024
4025
4026	/**
4027	* Free memory allocated by SUPR0PageAlloc() and SUPR0PageAllocEx().
4028	*
4029	* @returns IPRT status code.
4030	* @param pSession The session owning the allocation.
4031	* @param pvR3 The Ring-3 address returned by SUPR0PageAlloc() or
4032	* SUPR0PageAllocEx().
4033	*/
4034	SUPR0DECL(int) SUPR0PageFree(PSUPDRVSESSION pSession, RTR3PTR pvR3)
4035	{
4036	LogFlow(("SUPR0PageFree: pSession=%p pvR3=%p\n", pSession, (void *)pvR3));
4037	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
4038	return supdrvMemRelease(pSession, (RTHCUINTPTR)pvR3, MEMREF_TYPE_PAGE);
4039	}
4040	SUPR0_EXPORT_SYMBOL(SUPR0PageFree);
4041
4042
4043	/**
4044	* Reports a bad context, currenctly that means EFLAGS.AC is 0 instead of 1.
4045	*
4046	* @param pDevExt The device extension.
4047	* @param pszFile The source file where the caller detected the bad
4048	* context.
4049	* @param uLine The line number in @a pszFile.
4050	* @param pszExtra Optional additional message to give further hints.
4051	*/
4052	void VBOXCALL supdrvBadContext(PSUPDRVDEVEXT pDevExt, const char pszFile, uint32_t uLine, const char pszExtra)
4053	{
4054	uint32_t cCalls;
4055
4056	/*
4057	* Shorten the filename before displaying the message.
4058	*/
4059	for (;;)
4060	{
4061	const char *pszTmp = strchr(pszFile, '/');
4062	if (!pszTmp)
4063	pszTmp = strchr(pszFile, '\\');
4064	if (!pszTmp)
4065	break;
4066	pszFile = pszTmp + 1;
4067	}
4068	if (RT_VALID_PTR(pszExtra) && *pszExtra)
4069	SUPR0Printf("vboxdrv: Bad CPU context error at line %u in %s: %s\n", uLine, pszFile, pszExtra);
4070	else
4071	SUPR0Printf("vboxdrv: Bad CPU context error at line %u in %s!\n", uLine, pszFile);
4072
4073	/*
4074	* Record the incident so that we stand a chance of blocking I/O controls
4075	* before panicing the system.
4076	*/
4077	cCalls = ASMAtomicIncU32(&pDevExt->cBadContextCalls);
4078	if (cCalls > UINT32_MAX - _1K)
4079	ASMAtomicWriteU32(&pDevExt->cBadContextCalls, UINT32_MAX - _1K);
4080	}
4081
4082
4083	/**
4084	* Reports a bad context, currenctly that means EFLAGS.AC is 0 instead of 1.
4085	*
4086	* @param pSession The session of the caller.
4087	* @param pszFile The source file where the caller detected the bad
4088	* context.
4089	* @param uLine The line number in @a pszFile.
4090	* @param pszExtra Optional additional message to give further hints.
4091	*/
4092	SUPR0DECL(void) SUPR0BadContext(PSUPDRVSESSION pSession, const char pszFile, uint32_t uLine, const char pszExtra)
4093	{
4094	PSUPDRVDEVEXT pDevExt;
4095
4096	AssertReturnVoid(SUP_IS_SESSION_VALID(pSession));
4097	pDevExt = pSession->pDevExt;
4098
4099	supdrvBadContext(pDevExt, pszFile, uLine, pszExtra);
4100	}
4101	SUPR0_EXPORT_SYMBOL(SUPR0BadContext);
4102
4103
4104	/**
4105	* Gets the paging mode of the current CPU.
4106	*
4107	* @returns Paging mode, SUPPAGEINGMODE_INVALID on error.
4108	*/
4109	SUPR0DECL(SUPPAGINGMODE) SUPR0GetPagingMode(void)
4110	{
4111	SUPPAGINGMODE enmMode;
4112
4113	RTR0UINTREG cr0 = ASMGetCR0();
4114	if ((cr0 & (X86_CR0_PG \| X86_CR0_PE)) != (X86_CR0_PG \| X86_CR0_PE))
4115	enmMode = SUPPAGINGMODE_INVALID;
4116	else
4117	{
4118	RTR0UINTREG cr4 = ASMGetCR4();
4119	uint32_t fNXEPlusLMA = 0;
4120	if (cr4 & X86_CR4_PAE)
4121	{
4122	uint32_t fExtFeatures = ASMCpuId_EDX(0x80000001);
4123	if (fExtFeatures & (X86_CPUID_EXT_FEATURE_EDX_NX \| X86_CPUID_EXT_FEATURE_EDX_LONG_MODE))
4124	{
4125	uint64_t efer = ASMRdMsr(MSR_K6_EFER);
4126	if ((fExtFeatures & X86_CPUID_EXT_FEATURE_EDX_NX) && (efer & MSR_K6_EFER_NXE))
4127	fNXEPlusLMA \|= RT_BIT(0);
4128	if ((fExtFeatures & X86_CPUID_EXT_FEATURE_EDX_LONG_MODE) && (efer & MSR_K6_EFER_LMA))
4129	fNXEPlusLMA \|= RT_BIT(1);
4130	}
4131	}
4132
4133	switch ((cr4 & (X86_CR4_PAE \| X86_CR4_PGE)) \| fNXEPlusLMA)
4134	{
4135	case 0:
4136	enmMode = SUPPAGINGMODE_32_BIT;
4137	break;
4138
4139	case X86_CR4_PGE:
4140	enmMode = SUPPAGINGMODE_32_BIT_GLOBAL;
4141	break;
4142
4143	case X86_CR4_PAE:
4144	enmMode = SUPPAGINGMODE_PAE;
4145	break;
4146
4147	case X86_CR4_PAE \| RT_BIT(0):
4148	enmMode = SUPPAGINGMODE_PAE_NX;
4149	break;
4150
4151	case X86_CR4_PAE \| X86_CR4_PGE:
4152	enmMode = SUPPAGINGMODE_PAE_GLOBAL;
4153	break;
4154
4155	case X86_CR4_PAE \| X86_CR4_PGE \| RT_BIT(0):
4156	enmMode = SUPPAGINGMODE_PAE_GLOBAL;
4157	break;
4158
4159	case RT_BIT(1) \| X86_CR4_PAE:
4160	enmMode = SUPPAGINGMODE_AMD64;
4161	break;
4162
4163	case RT_BIT(1) \| X86_CR4_PAE \| RT_BIT(0):
4164	enmMode = SUPPAGINGMODE_AMD64_NX;
4165	break;
4166
4167	case RT_BIT(1) \| X86_CR4_PAE \| X86_CR4_PGE:
4168	enmMode = SUPPAGINGMODE_AMD64_GLOBAL;
4169	break;
4170
4171	case RT_BIT(1) \| X86_CR4_PAE \| X86_CR4_PGE \| RT_BIT(0):
4172	enmMode = SUPPAGINGMODE_AMD64_GLOBAL_NX;
4173	break;
4174
4175	default:
4176	AssertMsgFailed(("Cannot happen! cr4=%#x fNXEPlusLMA=%d\n", cr4, fNXEPlusLMA));
4177	enmMode = SUPPAGINGMODE_INVALID;
4178	break;
4179	}
4180	}
4181	return enmMode;
4182	}
4183	SUPR0_EXPORT_SYMBOL(SUPR0GetPagingMode);
4184
4185
4186	/**
4187	* Change CR4 and take care of the kernel CR4 shadow if applicable.
4188	*
4189	* CR4 shadow handling is required for Linux >= 4.0. Calling this function
4190	* instead of ASMSetCR4() is only necessary for semi-permanent CR4 changes
4191	* for code with interrupts enabled.
4192	*
4193	* @returns the old CR4 value.
4194	*
4195	* @param fOrMask bits to be set in CR4.
4196	* @param fAndMask bits to be cleard in CR4.
4197	*
4198	* @remarks Must be called with preemption/interrupts disabled.
4199	*/
4200	SUPR0DECL(RTCCUINTREG) SUPR0ChangeCR4(RTCCUINTREG fOrMask, RTCCUINTREG fAndMask)
4201	{
4202	#ifdef RT_OS_LINUX
4203	return supdrvOSChangeCR4(fOrMask, fAndMask);
4204	#else
4205	RTCCUINTREG uOld = ASMGetCR4();
4206	RTCCUINTREG uNew = (uOld & fAndMask) \| fOrMask;
4207	if (uNew != uOld)
4208	ASMSetCR4(uNew);
4209	return uOld;
4210	#endif
4211	}
4212	SUPR0_EXPORT_SYMBOL(SUPR0ChangeCR4);
4213
4214
4215	/**
4216	* Enables or disabled hardware virtualization extensions using native OS APIs.
4217	*
4218	* @returns VBox status code.
4219	* @retval VINF_SUCCESS on success.
4220	* @retval VERR_NOT_SUPPORTED if not supported by the native OS.
4221	*
4222	* @param fEnable Whether to enable or disable.
4223	*/
4224	SUPR0DECL(int) SUPR0EnableVTx(bool fEnable)
4225	{
4226	#ifdef RT_OS_DARWIN
4227	return supdrvOSEnableVTx(fEnable);
4228	#else
4229	RT_NOREF1(fEnable);
4230	return VERR_NOT_SUPPORTED;
4231	#endif
4232	}
4233	SUPR0_EXPORT_SYMBOL(SUPR0EnableVTx);
4234
4235
4236	/**
4237	* Suspends hardware virtualization extensions using the native OS API.
4238	*
4239	* This is called prior to entering raw-mode context.
4240	*
4241	* @returns @c true if suspended, @c false if not.
4242	*/
4243	SUPR0DECL(bool) SUPR0SuspendVTxOnCpu(void)
4244	{
4245	#ifdef RT_OS_DARWIN
4246	return supdrvOSSuspendVTxOnCpu();
4247	#else
4248	return false;
4249	#endif
4250	}
4251	SUPR0_EXPORT_SYMBOL(SUPR0SuspendVTxOnCpu);
4252
4253
4254	/**
4255	* Resumes hardware virtualization extensions using the native OS API.
4256	*
4257	* This is called after to entering raw-mode context.
4258	*
4259	* @param fSuspended The return value of SUPR0SuspendVTxOnCpu.
4260	*/
4261	SUPR0DECL(void) SUPR0ResumeVTxOnCpu(bool fSuspended)
4262	{
4263	#ifdef RT_OS_DARWIN
4264	supdrvOSResumeVTxOnCpu(fSuspended);
4265	#else
4266	RT_NOREF1(fSuspended);
4267	Assert(!fSuspended);
4268	#endif
4269	}
4270	SUPR0_EXPORT_SYMBOL(SUPR0ResumeVTxOnCpu);
4271
4272
4273	SUPR0DECL(int) SUPR0GetCurrentGdtRw(RTHCUINTPTR *pGdtRw)
4274	{
4275	#ifdef RT_OS_LINUX
4276	return supdrvOSGetCurrentGdtRw(pGdtRw);
4277	#else
4278	NOREF(pGdtRw);
4279	return VERR_NOT_IMPLEMENTED;
4280	#endif
4281	}
4282	SUPR0_EXPORT_SYMBOL(SUPR0GetCurrentGdtRw);
4283
4284
4285	/**
4286	* Gets AMD-V and VT-x support for the calling CPU.
4287	*
4288	* @returns VBox status code.
4289	* @param pfCaps Where to store whether VT-x (SUPVTCAPS_VT_X) or AMD-V
4290	* (SUPVTCAPS_AMD_V) is supported.
4291	*/
4292	SUPR0DECL(int) SUPR0GetVTSupport(uint32_t *pfCaps)
4293	{
4294	Assert(pfCaps);
4295	*pfCaps = 0;
4296
4297	/* Check if the CPU even supports CPUID (extremely ancient CPUs). */
4298	if (ASMHasCpuId())
4299	{
4300	/* Check the range of standard CPUID leafs. */
4301	uint32_t uMaxLeaf, uVendorEbx, uVendorEcx, uVendorEdx;
4302	ASMCpuId(0, &uMaxLeaf, &uVendorEbx, &uVendorEcx, &uVendorEdx);
4303	if (ASMIsValidStdRange(uMaxLeaf))
4304	{
4305	/* Query the standard CPUID leaf. */
4306	uint32_t fFeatEcx, fFeatEdx, uDummy;
4307	ASMCpuId(1, &uDummy, &uDummy, &fFeatEcx, &fFeatEdx);
4308
4309	/* Check if the vendor is Intel (or compatible). */
4310	if ( ASMIsIntelCpuEx(uVendorEbx, uVendorEcx, uVendorEdx)
4311	\|\| ASMIsViaCentaurCpuEx(uVendorEbx, uVendorEcx, uVendorEdx)
4312	\|\| ASMIsShanghaiCpuEx(uVendorEbx, uVendorEcx, uVendorEdx))
4313	{
4314	/* Check VT-x support. In addition, VirtualBox requires MSR and FXSAVE/FXRSTOR to function. */
4315	if ( (fFeatEcx & X86_CPUID_FEATURE_ECX_VMX)
4316	&& (fFeatEdx & X86_CPUID_FEATURE_EDX_MSR)
4317	&& (fFeatEdx & X86_CPUID_FEATURE_EDX_FXSR))
4318	{
4319	*pfCaps = SUPVTCAPS_VT_X;
4320	return VINF_SUCCESS;
4321	}
4322	return VERR_VMX_NO_VMX;
4323	}
4324
4325	/* Check if the vendor is AMD (or compatible). */
4326	if ( ASMIsAmdCpuEx(uVendorEbx, uVendorEcx, uVendorEdx)
4327	\|\| ASMIsHygonCpuEx(uVendorEbx, uVendorEcx, uVendorEdx))
4328	{
4329	uint32_t fExtFeatEcx, uExtMaxId;
4330	ASMCpuId(0x80000000, &uExtMaxId, &uDummy, &uDummy, &uDummy);
4331	ASMCpuId(0x80000001, &uDummy, &uDummy, &fExtFeatEcx, &uDummy);
4332
4333	/* Check AMD-V support. In addition, VirtualBox requires MSR and FXSAVE/FXRSTOR to function. */
4334	if ( ASMIsValidExtRange(uExtMaxId)
4335	&& uExtMaxId >= 0x8000000a
4336	&& (fExtFeatEcx & X86_CPUID_AMD_FEATURE_ECX_SVM)
4337	&& (fFeatEdx & X86_CPUID_FEATURE_EDX_MSR)
4338	&& (fFeatEdx & X86_CPUID_FEATURE_EDX_FXSR))
4339	{
4340	*pfCaps = SUPVTCAPS_AMD_V;
4341	return VINF_SUCCESS;
4342	}
4343	return VERR_SVM_NO_SVM;
4344	}
4345	}
4346	}
4347	return VERR_UNSUPPORTED_CPU;
4348	}
4349	SUPR0_EXPORT_SYMBOL(SUPR0GetVTSupport);
4350
4351
4352	/**
4353	* Checks if Intel VT-x feature is usable on this CPU.
4354	*
4355	* @returns VBox status code.
4356	* @param pfIsSmxModeAmbiguous Where to return whether the SMX mode causes
4357	* ambiguity that makes us unsure whether we
4358	* really can use VT-x or not.
4359	*
4360	* @remarks Must be called with preemption disabled.
4361	* The caller is also expected to check that the CPU is an Intel (or
4362	* VIA/Shanghai) CPU -and- that it supports VT-x. Otherwise, this
4363	* function might throw a \#GP fault as it tries to read/write MSRs
4364	* that may not be present!
4365	*/
4366	SUPR0DECL(int) SUPR0GetVmxUsability(bool *pfIsSmxModeAmbiguous)
4367	{
4368	uint64_t fFeatMsr;
4369	bool fMaybeSmxMode;
4370	bool fMsrLocked;
4371	bool fSmxVmxAllowed;
4372	bool fVmxAllowed;
4373	bool fIsSmxModeAmbiguous;
4374	int rc;
4375
4376	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
4377
4378	fFeatMsr = ASMRdMsr(MSR_IA32_FEATURE_CONTROL);
4379	fMaybeSmxMode = RT_BOOL(ASMGetCR4() & X86_CR4_SMXE);
4380	fMsrLocked = RT_BOOL(fFeatMsr & MSR_IA32_FEATURE_CONTROL_LOCK);
4381	fSmxVmxAllowed = RT_BOOL(fFeatMsr & MSR_IA32_FEATURE_CONTROL_SMX_VMXON);
4382	fVmxAllowed = RT_BOOL(fFeatMsr & MSR_IA32_FEATURE_CONTROL_VMXON);
4383	fIsSmxModeAmbiguous = false;
4384	rc = VERR_INTERNAL_ERROR_5;
4385
4386	/* Check if the LOCK bit is set but excludes the required VMXON bit. */
4387	if (fMsrLocked)
4388	{
4389	if (fVmxAllowed && fSmxVmxAllowed)
4390	rc = VINF_SUCCESS;
4391	else if (!fVmxAllowed && !fSmxVmxAllowed)
4392	rc = VERR_VMX_MSR_ALL_VMX_DISABLED;
4393	else if (!fMaybeSmxMode)
4394	{
4395	if (fVmxAllowed)
4396	rc = VINF_SUCCESS;
4397	else
4398	rc = VERR_VMX_MSR_VMX_DISABLED;
4399	}
4400	else
4401	{
4402	/*
4403	* CR4.SMXE is set but this doesn't mean the CPU is necessarily in SMX mode. We shall assume
4404	* that it is -not- and that it is a stupid BIOS/OS setting CR4.SMXE for no good reason.
4405	* See @bugref{6873}.
4406	*/
4407	Assert(fMaybeSmxMode == true);
4408	fIsSmxModeAmbiguous = true;
4409	rc = VINF_SUCCESS;
4410	}
4411	}
4412	else
4413	{
4414	/*
4415	* MSR is not yet locked; we can change it ourselves here. Once the lock bit is set,
4416	* this MSR can no longer be modified.
4417	*
4418	* Set both the VMX and SMX_VMX bits (if supported) as we can't determine SMX mode
4419	* accurately. See @bugref{6873}.
4420	*
4421	* We need to check for SMX hardware support here, before writing the MSR as
4422	* otherwise we will #GP fault on CPUs that do not support it. Callers do not check
4423	* for it.
4424	*/
4425	uint32_t fFeaturesECX, uDummy;
4426	#ifdef VBOX_STRICT
4427	/* Callers should have verified these at some point. */
4428	uint32_t uMaxId, uVendorEBX, uVendorECX, uVendorEDX;
4429	ASMCpuId(0, &uMaxId, &uVendorEBX, &uVendorECX, &uVendorEDX);
4430	Assert(ASMIsValidStdRange(uMaxId));
4431	Assert( ASMIsIntelCpuEx( uVendorEBX, uVendorECX, uVendorEDX)
4432	\|\| ASMIsViaCentaurCpuEx(uVendorEBX, uVendorECX, uVendorEDX)
4433	\|\| ASMIsShanghaiCpuEx( uVendorEBX, uVendorECX, uVendorEDX));
4434	#endif
4435	ASMCpuId(1, &uDummy, &uDummy, &fFeaturesECX, &uDummy);
4436	bool fSmxVmxHwSupport = false;
4437	if ( (fFeaturesECX & X86_CPUID_FEATURE_ECX_VMX)
4438	&& (fFeaturesECX & X86_CPUID_FEATURE_ECX_SMX))
4439	fSmxVmxHwSupport = true;
4440
4441	fFeatMsr \|= MSR_IA32_FEATURE_CONTROL_LOCK
4442	\| MSR_IA32_FEATURE_CONTROL_VMXON;
4443	if (fSmxVmxHwSupport)
4444	fFeatMsr \|= MSR_IA32_FEATURE_CONTROL_SMX_VMXON;
4445
4446	/*
4447	* Commit.
4448	*/
4449	ASMWrMsr(MSR_IA32_FEATURE_CONTROL, fFeatMsr);
4450
4451	/*
4452	* Verify.
4453	*/
4454	fFeatMsr = ASMRdMsr(MSR_IA32_FEATURE_CONTROL);
4455	fMsrLocked = RT_BOOL(fFeatMsr & MSR_IA32_FEATURE_CONTROL_LOCK);
4456	if (fMsrLocked)
4457	{
4458	fSmxVmxAllowed = RT_BOOL(fFeatMsr & MSR_IA32_FEATURE_CONTROL_SMX_VMXON);
4459	fVmxAllowed = RT_BOOL(fFeatMsr & MSR_IA32_FEATURE_CONTROL_VMXON);
4460	if ( fVmxAllowed
4461	&& ( !fSmxVmxHwSupport
4462	\|\| fSmxVmxAllowed))
4463	rc = VINF_SUCCESS;
4464	else
4465	rc = !fSmxVmxHwSupport ? VERR_VMX_MSR_VMX_ENABLE_FAILED : VERR_VMX_MSR_SMX_VMX_ENABLE_FAILED;
4466	}
4467	else
4468	rc = VERR_VMX_MSR_LOCKING_FAILED;
4469	}
4470
4471	if (pfIsSmxModeAmbiguous)
4472	*pfIsSmxModeAmbiguous = fIsSmxModeAmbiguous;
4473
4474	return rc;
4475	}
4476	SUPR0_EXPORT_SYMBOL(SUPR0GetVmxUsability);
4477
4478
4479	/**
4480	* Checks if AMD-V SVM feature is usable on this CPU.
4481	*
4482	* @returns VBox status code.
4483	* @param fInitSvm If usable, try to initialize SVM on this CPU.
4484	*
4485	* @remarks Must be called with preemption disabled.
4486	*/
4487	SUPR0DECL(int) SUPR0GetSvmUsability(bool fInitSvm)
4488	{
4489	int rc;
4490	uint64_t fVmCr;
4491	uint64_t fEfer;
4492
4493	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
4494	fVmCr = ASMRdMsr(MSR_K8_VM_CR);
4495	if (!(fVmCr & MSR_K8_VM_CR_SVM_DISABLE))
4496	{
4497	rc = VINF_SUCCESS;
4498	if (fInitSvm)
4499	{
4500	/* Turn on SVM in the EFER MSR. */
4501	fEfer = ASMRdMsr(MSR_K6_EFER);
4502	if (fEfer & MSR_K6_EFER_SVME)
4503	rc = VERR_SVM_IN_USE;
4504	else
4505	{
4506	ASMWrMsr(MSR_K6_EFER, fEfer \| MSR_K6_EFER_SVME);
4507
4508	/* Paranoia. */
4509	fEfer = ASMRdMsr(MSR_K6_EFER);
4510	if (fEfer & MSR_K6_EFER_SVME)
4511	{
4512	/* Restore previous value. */
4513	ASMWrMsr(MSR_K6_EFER, fEfer & ~MSR_K6_EFER_SVME);
4514	}
4515	else
4516	rc = VERR_SVM_ILLEGAL_EFER_MSR;
4517	}
4518	}
4519	}
4520	else
4521	rc = VERR_SVM_DISABLED;
4522	return rc;
4523	}
4524	SUPR0_EXPORT_SYMBOL(SUPR0GetSvmUsability);
4525
4526
4527	/**
4528	* Queries the AMD-V and VT-x capabilities of the calling CPU.
4529	*
4530	* @returns VBox status code.
4531	* @retval VERR_VMX_NO_VMX
4532	* @retval VERR_VMX_MSR_ALL_VMX_DISABLED
4533	* @retval VERR_VMX_MSR_VMX_DISABLED
4534	* @retval VERR_VMX_MSR_LOCKING_FAILED
4535	* @retval VERR_VMX_MSR_VMX_ENABLE_FAILED
4536	* @retval VERR_VMX_MSR_SMX_VMX_ENABLE_FAILED
4537	* @retval VERR_SVM_NO_SVM
4538	* @retval VERR_SVM_DISABLED
4539	* @retval VERR_UNSUPPORTED_CPU if not identifiable as an AMD, Intel or VIA
4540	* (centaur)/Shanghai CPU.
4541	*
4542	* @param pfCaps Where to store the capabilities.
4543	*/
4544	int VBOXCALL supdrvQueryVTCapsInternal(uint32_t *pfCaps)
4545	{
4546	int rc = VERR_UNSUPPORTED_CPU;
4547	bool fIsSmxModeAmbiguous = false;
4548	RTTHREADPREEMPTSTATE PreemptState = RTTHREADPREEMPTSTATE_INITIALIZER;
4549
4550	/*
4551	* Input validation.
4552	*/
4553	AssertPtrReturn(pfCaps, VERR_INVALID_POINTER);
4554	*pfCaps = 0;
4555
4556	/* We may modify MSRs and re-read them, disable preemption so we make sure we don't migrate CPUs. */
4557	RTThreadPreemptDisable(&PreemptState);
4558
4559	/* Check if VT-x/AMD-V is supported. */
4560	rc = SUPR0GetVTSupport(pfCaps);
4561	if (RT_SUCCESS(rc))
4562	{
4563	/* Check if VT-x is supported. */
4564	if (*pfCaps & SUPVTCAPS_VT_X)
4565	{
4566	/* Check if VT-x is usable. */
4567	rc = SUPR0GetVmxUsability(&fIsSmxModeAmbiguous);
4568	if (RT_SUCCESS(rc))
4569	{
4570	/* Query some basic VT-x capabilities (mainly required by our GUI). */
4571	VMXCTLSMSR vtCaps;
4572	vtCaps.u = ASMRdMsr(MSR_IA32_VMX_PROCBASED_CTLS);
4573	if (vtCaps.n.allowed1 & VMX_PROC_CTLS_USE_SECONDARY_CTLS)
4574	{
4575	vtCaps.u = ASMRdMsr(MSR_IA32_VMX_PROCBASED_CTLS2);
4576	if (vtCaps.n.allowed1 & VMX_PROC_CTLS2_EPT)
4577	*pfCaps \|= SUPVTCAPS_NESTED_PAGING;
4578	if (vtCaps.n.allowed1 & VMX_PROC_CTLS2_UNRESTRICTED_GUEST)
4579	*pfCaps \|= SUPVTCAPS_VTX_UNRESTRICTED_GUEST;
4580	if (vtCaps.n.allowed1 & VMX_PROC_CTLS2_VMCS_SHADOWING)
4581	*pfCaps \|= SUPVTCAPS_VTX_VMCS_SHADOWING;
4582	}
4583	}
4584	}
4585	/* Check if AMD-V is supported. */
4586	else if (*pfCaps & SUPVTCAPS_AMD_V)
4587	{
4588	/* Check is SVM is usable. */
4589	rc = SUPR0GetSvmUsability(false /* fInitSvm */);
4590	if (RT_SUCCESS(rc))
4591	{
4592	/* Query some basic AMD-V capabilities (mainly required by our GUI). */
4593	uint32_t uDummy, fSvmFeatures;
4594	ASMCpuId(0x8000000a, &uDummy, &uDummy, &uDummy, &fSvmFeatures);
4595	if (fSvmFeatures & X86_CPUID_SVM_FEATURE_EDX_NESTED_PAGING)
4596	*pfCaps \|= SUPVTCAPS_NESTED_PAGING;
4597	if (fSvmFeatures & X86_CPUID_SVM_FEATURE_EDX_VIRT_VMSAVE_VMLOAD)
4598	*pfCaps \|= SUPVTCAPS_AMDV_VIRT_VMSAVE_VMLOAD;
4599	}
4600	}
4601	}
4602
4603	/* Restore preemption. */
4604	RTThreadPreemptRestore(&PreemptState);
4605
4606	/* After restoring preemption, if we may be in SMX mode, print a warning as it's difficult to debug such problems. */
4607	if (fIsSmxModeAmbiguous)
4608	SUPR0Printf(("WARNING! CR4 hints SMX mode but your CPU is too secretive. Proceeding anyway... We wish you good luck!\n"));
4609
4610	return rc;
4611	}
4612
4613
4614	/**
4615	* Queries the AMD-V and VT-x capabilities of the calling CPU.
4616	*
4617	* @returns VBox status code.
4618	* @retval VERR_VMX_NO_VMX
4619	* @retval VERR_VMX_MSR_ALL_VMX_DISABLED
4620	* @retval VERR_VMX_MSR_VMX_DISABLED
4621	* @retval VERR_VMX_MSR_LOCKING_FAILED
4622	* @retval VERR_VMX_MSR_VMX_ENABLE_FAILED
4623	* @retval VERR_VMX_MSR_SMX_VMX_ENABLE_FAILED
4624	* @retval VERR_SVM_NO_SVM
4625	* @retval VERR_SVM_DISABLED
4626	* @retval VERR_UNSUPPORTED_CPU if not identifiable as an AMD, Intel or VIA
4627	* (centaur)/Shanghai CPU.
4628	*
4629	* @param pSession The session handle.
4630	* @param pfCaps Where to store the capabilities.
4631	*/
4632	SUPR0DECL(int) SUPR0QueryVTCaps(PSUPDRVSESSION pSession, uint32_t *pfCaps)
4633	{
4634	/*
4635	* Input validation.
4636	*/
4637	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
4638	AssertPtrReturn(pfCaps, VERR_INVALID_POINTER);
4639
4640	/*
4641	* Call common worker.
4642	*/
4643	return supdrvQueryVTCapsInternal(pfCaps);
4644	}
4645	SUPR0_EXPORT_SYMBOL(SUPR0QueryVTCaps);
4646
4647
4648	/**
4649	* Queries the CPU microcode revision.
4650	*
4651	* @returns VBox status code.
4652	* @retval VERR_UNSUPPORTED_CPU if not identifiable as a processor with
4653	* readable microcode rev.
4654	*
4655	* @param puRevision Where to store the microcode revision.
4656	*/
4657	static int VBOXCALL supdrvQueryUcodeRev(uint32_t *puRevision)
4658	{
4659	int rc = VERR_UNSUPPORTED_CPU;
4660	RTTHREADPREEMPTSTATE PreemptState = RTTHREADPREEMPTSTATE_INITIALIZER;
4661
4662	/*
4663	* Input validation.
4664	*/
4665	AssertPtrReturn(puRevision, VERR_INVALID_POINTER);
4666
4667	*puRevision = 0;
4668
4669	/* Disable preemption so we make sure we don't migrate CPUs, just in case. */
4670	/* NB: We assume that there aren't mismatched microcode revs in the system. */
4671	RTThreadPreemptDisable(&PreemptState);
4672
4673	if (ASMHasCpuId())
4674	{
4675	uint32_t uDummy, uTFMSEAX;
4676	uint32_t uMaxId, uVendorEBX, uVendorECX, uVendorEDX;
4677
4678	ASMCpuId(0, &uMaxId, &uVendorEBX, &uVendorECX, &uVendorEDX);
4679	ASMCpuId(1, &uTFMSEAX, &uDummy, &uDummy, &uDummy);
4680
4681	if (ASMIsValidStdRange(uMaxId))
4682	{
4683	uint64_t uRevMsr;
4684	if (ASMIsIntelCpuEx(uVendorEBX, uVendorECX, uVendorEDX))
4685	{
4686	/* Architectural MSR available on Pentium Pro and later. */
4687	if (ASMGetCpuFamily(uTFMSEAX) >= 6)
4688	{
4689	/* Revision is in the high dword. */
4690	uRevMsr = ASMRdMsr(MSR_IA32_BIOS_SIGN_ID);
4691	*puRevision = RT_HIDWORD(uRevMsr);
4692	rc = VINF_SUCCESS;
4693	}
4694	}
4695	else if ( ASMIsAmdCpuEx(uVendorEBX, uVendorECX, uVendorEDX)
4696	\|\| ASMIsHygonCpuEx(uVendorEBX, uVendorECX, uVendorEDX))
4697	{
4698	/* Not well documented, but at least all AMD64 CPUs support this. */
4699	if (ASMGetCpuFamily(uTFMSEAX) >= 15)
4700	{
4701	/* Revision is in the low dword. */
4702	uRevMsr = ASMRdMsr(MSR_IA32_BIOS_SIGN_ID); /* Same MSR as Intel. */
4703	*puRevision = RT_LODWORD(uRevMsr);
4704	rc = VINF_SUCCESS;
4705	}
4706	}
4707	}
4708	}
4709
4710	RTThreadPreemptRestore(&PreemptState);
4711
4712	return rc;
4713	}
4714
4715
4716	/**
4717	* Queries the CPU microcode revision.
4718	*
4719	* @returns VBox status code.
4720	* @retval VERR_UNSUPPORTED_CPU if not identifiable as a processor with
4721	* readable microcode rev.
4722	*
4723	* @param pSession The session handle.
4724	* @param puRevision Where to store the microcode revision.
4725	*/
4726	SUPR0DECL(int) SUPR0QueryUcodeRev(PSUPDRVSESSION pSession, uint32_t *puRevision)
4727	{
4728	/*
4729	* Input validation.
4730	*/
4731	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
4732	AssertPtrReturn(puRevision, VERR_INVALID_POINTER);
4733
4734	/*
4735	* Call common worker.
4736	*/
4737	return supdrvQueryUcodeRev(puRevision);
4738	}
4739	SUPR0_EXPORT_SYMBOL(SUPR0QueryUcodeRev);
4740
4741
4742	/**
4743	* Gets hardware-virtualization MSRs of the calling CPU.
4744	*
4745	* @returns VBox status code.
4746	* @param pMsrs Where to store the hardware-virtualization MSRs.
4747	* @param fCaps Hardware virtualization capabilities (SUPVTCAPS_XXX). Pass 0
4748	* to explicitly check for the presence of VT-x/AMD-V before
4749	* querying MSRs.
4750	* @param fForce Force querying of MSRs from the hardware.
4751	*/
4752	SUPR0DECL(int) SUPR0GetHwvirtMsrs(PSUPHWVIRTMSRS pMsrs, uint32_t fCaps, bool fForce)
4753	{
4754	NOREF(fForce);
4755
4756	int rc;
4757	RTTHREADPREEMPTSTATE PreemptState = RTTHREADPREEMPTSTATE_INITIALIZER;
4758
4759	/*
4760	* Input validation.
4761	*/
4762	AssertPtrReturn(pMsrs, VERR_INVALID_POINTER);
4763
4764	/*
4765	* Disable preemption so we make sure we don't migrate CPUs and because
4766	* we access global data.
4767	*/
4768	RTThreadPreemptDisable(&PreemptState);
4769
4770	/*
4771	* Query the MSRs from the hardware.
4772	*/
4773	SUPHWVIRTMSRS Msrs;
4774	RT_ZERO(Msrs);
4775
4776	/* If the caller claims VT-x/AMD-V is supported, don't need to recheck it. */
4777	if (!(fCaps & (SUPVTCAPS_VT_X \| SUPVTCAPS_AMD_V)))
4778	rc = SUPR0GetVTSupport(&fCaps);
4779	else
4780	rc = VINF_SUCCESS;
4781	if (RT_SUCCESS(rc))
4782	{
4783	if (fCaps & SUPVTCAPS_VT_X)
4784	{
4785	Msrs.u.vmx.u64FeatCtrl = ASMRdMsr(MSR_IA32_FEATURE_CONTROL);
4786	Msrs.u.vmx.u64Basic = ASMRdMsr(MSR_IA32_VMX_BASIC);
4787	Msrs.u.vmx.PinCtls.u = ASMRdMsr(MSR_IA32_VMX_PINBASED_CTLS);
4788	Msrs.u.vmx.ProcCtls.u = ASMRdMsr(MSR_IA32_VMX_PROCBASED_CTLS);
4789	Msrs.u.vmx.ExitCtls.u = ASMRdMsr(MSR_IA32_VMX_EXIT_CTLS);
4790	Msrs.u.vmx.EntryCtls.u = ASMRdMsr(MSR_IA32_VMX_ENTRY_CTLS);
4791	Msrs.u.vmx.u64Misc = ASMRdMsr(MSR_IA32_VMX_MISC);
4792	Msrs.u.vmx.u64Cr0Fixed0 = ASMRdMsr(MSR_IA32_VMX_CR0_FIXED0);
4793	Msrs.u.vmx.u64Cr0Fixed1 = ASMRdMsr(MSR_IA32_VMX_CR0_FIXED1);
4794	Msrs.u.vmx.u64Cr4Fixed0 = ASMRdMsr(MSR_IA32_VMX_CR4_FIXED0);
4795	Msrs.u.vmx.u64Cr4Fixed1 = ASMRdMsr(MSR_IA32_VMX_CR4_FIXED1);
4796	Msrs.u.vmx.u64VmcsEnum = ASMRdMsr(MSR_IA32_VMX_VMCS_ENUM);
4797
4798	if (RT_BF_GET(Msrs.u.vmx.u64Basic, VMX_BF_BASIC_TRUE_CTLS))
4799	{
4800	Msrs.u.vmx.TruePinCtls.u = ASMRdMsr(MSR_IA32_VMX_TRUE_PINBASED_CTLS);
4801	Msrs.u.vmx.TrueProcCtls.u = ASMRdMsr(MSR_IA32_VMX_TRUE_PROCBASED_CTLS);
4802	Msrs.u.vmx.TrueEntryCtls.u = ASMRdMsr(MSR_IA32_VMX_TRUE_ENTRY_CTLS);
4803	Msrs.u.vmx.TrueExitCtls.u = ASMRdMsr(MSR_IA32_VMX_TRUE_EXIT_CTLS);
4804	}
4805
4806	if (Msrs.u.vmx.ProcCtls.n.allowed1 & VMX_PROC_CTLS_USE_SECONDARY_CTLS)
4807	{
4808	Msrs.u.vmx.ProcCtls2.u = ASMRdMsr(MSR_IA32_VMX_PROCBASED_CTLS2);
4809
4810	if (Msrs.u.vmx.ProcCtls2.n.allowed1 & (VMX_PROC_CTLS2_EPT \| VMX_PROC_CTLS2_VPID))
4811	Msrs.u.vmx.u64EptVpidCaps = ASMRdMsr(MSR_IA32_VMX_EPT_VPID_CAP);
4812
4813	if (Msrs.u.vmx.ProcCtls2.n.allowed1 & VMX_PROC_CTLS2_VMFUNC)
4814	Msrs.u.vmx.u64VmFunc = ASMRdMsr(MSR_IA32_VMX_VMFUNC);
4815	}
4816
4817	if (Msrs.u.vmx.ProcCtls.n.allowed1 & VMX_PROC_CTLS_USE_TERTIARY_CTLS)
4818	Msrs.u.vmx.u64ProcCtls3 = ASMRdMsr(MSR_IA32_VMX_PROCBASED_CTLS3);
4819	}
4820	else if (fCaps & SUPVTCAPS_AMD_V)
4821	{
4822	Msrs.u.svm.u64MsrHwcr = ASMRdMsr(MSR_K8_HWCR);
4823	Msrs.u.svm.u64MsrSmmAddr = ASMRdMsr(MSR_K7_SMM_ADDR);
4824	Msrs.u.svm.u64MsrSmmMask = ASMRdMsr(MSR_K7_SMM_MASK);
4825	}
4826	else
4827	{
4828	RTThreadPreemptRestore(&PreemptState);
4829	AssertMsgFailedReturn(("SUPR0GetVTSupport returns success but neither VT-x nor AMD-V reported!\n"),
4830	VERR_INTERNAL_ERROR_2);
4831	}
4832
4833	/*
4834	* Copy the MSRs out.
4835	*/
4836	memcpy(pMsrs, &Msrs, sizeof(*pMsrs));
4837	}
4838
4839	RTThreadPreemptRestore(&PreemptState);
4840
4841	return rc;
4842	}
4843	SUPR0_EXPORT_SYMBOL(SUPR0GetHwvirtMsrs);
4844
4845
4846	/**
4847	* Register a component factory with the support driver.
4848	*
4849	* This is currently restricted to kernel sessions only.
4850	*
4851	* @returns VBox status code.
4852	* @retval VINF_SUCCESS on success.
4853	* @retval VERR_NO_MEMORY if we're out of memory.
4854	* @retval VERR_ALREADY_EXISTS if the factory has already been registered.
4855	* @retval VERR_ACCESS_DENIED if it isn't a kernel session.
4856	* @retval VERR_INVALID_PARAMETER on invalid parameter.
4857	* @retval VERR_INVALID_POINTER on invalid pointer parameter.
4858	*
4859	* @param pSession The SUPDRV session (must be a ring-0 session).
4860	* @param pFactory Pointer to the component factory registration structure.
4861	*
4862	* @remarks This interface is also available via SUPR0IdcComponentRegisterFactory.
4863	*/
4864	SUPR0DECL(int) SUPR0ComponentRegisterFactory(PSUPDRVSESSION pSession, PCSUPDRVFACTORY pFactory)
4865	{
4866	PSUPDRVFACTORYREG pNewReg;
4867	const char *psz;
4868	int rc;
4869
4870	/*
4871	* Validate parameters.
4872	*/
4873	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
4874	AssertReturn(pSession->R0Process == NIL_RTR0PROCESS, VERR_ACCESS_DENIED);
4875	AssertPtrReturn(pFactory, VERR_INVALID_POINTER);
4876	AssertPtrReturn(pFactory->pfnQueryFactoryInterface, VERR_INVALID_POINTER);
4877	psz = RTStrEnd(pFactory->szName, sizeof(pFactory->szName));
4878	AssertReturn(psz, VERR_INVALID_PARAMETER);
4879
4880	/*
4881	* Allocate and initialize a new registration structure.
4882	*/
4883	pNewReg = (PSUPDRVFACTORYREG)RTMemAlloc(sizeof(SUPDRVFACTORYREG));
4884	if (pNewReg)
4885	{
4886	pNewReg->pNext = NULL;
4887	pNewReg->pFactory = pFactory;
4888	pNewReg->pSession = pSession;
4889	pNewReg->cchName = psz - &pFactory->szName[0];
4890
4891	/*
4892	* Add it to the tail of the list after checking for prior registration.
4893	*/
4894	rc = RTSemFastMutexRequest(pSession->pDevExt->mtxComponentFactory);
4895	if (RT_SUCCESS(rc))
4896	{
4897	PSUPDRVFACTORYREG pPrev = NULL;
4898	PSUPDRVFACTORYREG pCur = pSession->pDevExt->pComponentFactoryHead;
4899	while (pCur && pCur->pFactory != pFactory)
4900	{
4901	pPrev = pCur;
4902	pCur = pCur->pNext;
4903	}
4904	if (!pCur)
4905	{
4906	if (pPrev)
4907	pPrev->pNext = pNewReg;
4908	else
4909	pSession->pDevExt->pComponentFactoryHead = pNewReg;
4910	rc = VINF_SUCCESS;
4911	}
4912	else
4913	rc = VERR_ALREADY_EXISTS;
4914
4915	RTSemFastMutexRelease(pSession->pDevExt->mtxComponentFactory);
4916	}
4917
4918	if (RT_FAILURE(rc))
4919	RTMemFree(pNewReg);
4920	}
4921	else
4922	rc = VERR_NO_MEMORY;
4923	return rc;
4924	}
4925	SUPR0_EXPORT_SYMBOL(SUPR0ComponentRegisterFactory);
4926
4927
4928	/**
4929	* Deregister a component factory.
4930	*
4931	* @returns VBox status code.
4932	* @retval VINF_SUCCESS on success.
4933	* @retval VERR_NOT_FOUND if the factory wasn't registered.
4934	* @retval VERR_ACCESS_DENIED if it isn't a kernel session.
4935	* @retval VERR_INVALID_PARAMETER on invalid parameter.
4936	* @retval VERR_INVALID_POINTER on invalid pointer parameter.
4937	*
4938	* @param pSession The SUPDRV session (must be a ring-0 session).
4939	* @param pFactory Pointer to the component factory registration structure
4940	* previously passed SUPR0ComponentRegisterFactory().
4941	*
4942	* @remarks This interface is also available via SUPR0IdcComponentDeregisterFactory.
4943	*/
4944	SUPR0DECL(int) SUPR0ComponentDeregisterFactory(PSUPDRVSESSION pSession, PCSUPDRVFACTORY pFactory)
4945	{
4946	int rc;
4947
4948	/*
4949	* Validate parameters.
4950	*/
4951	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
4952	AssertReturn(pSession->R0Process == NIL_RTR0PROCESS, VERR_ACCESS_DENIED);
4953	AssertPtrReturn(pFactory, VERR_INVALID_POINTER);
4954
4955	/*
4956	* Take the lock and look for the registration record.
4957	*/
4958	rc = RTSemFastMutexRequest(pSession->pDevExt->mtxComponentFactory);
4959	if (RT_SUCCESS(rc))
4960	{
4961	PSUPDRVFACTORYREG pPrev = NULL;
4962	PSUPDRVFACTORYREG pCur = pSession->pDevExt->pComponentFactoryHead;
4963	while (pCur && pCur->pFactory != pFactory)
4964	{
4965	pPrev = pCur;
4966	pCur = pCur->pNext;
4967	}
4968	if (pCur)
4969	{
4970	if (!pPrev)
4971	pSession->pDevExt->pComponentFactoryHead = pCur->pNext;
4972	else
4973	pPrev->pNext = pCur->pNext;
4974
4975	pCur->pNext = NULL;
4976	pCur->pFactory = NULL;
4977	pCur->pSession = NULL;
4978	rc = VINF_SUCCESS;
4979	}
4980	else
4981	rc = VERR_NOT_FOUND;
4982
4983	RTSemFastMutexRelease(pSession->pDevExt->mtxComponentFactory);
4984
4985	RTMemFree(pCur);
4986	}
4987	return rc;
4988	}
4989	SUPR0_EXPORT_SYMBOL(SUPR0ComponentDeregisterFactory);
4990
4991
4992	/**
4993	* Queries a component factory.
4994	*
4995	* @returns VBox status code.
4996	* @retval VERR_INVALID_PARAMETER on invalid parameter.
4997	* @retval VERR_INVALID_POINTER on invalid pointer parameter.
4998	* @retval VERR_SUPDRV_COMPONENT_NOT_FOUND if the component factory wasn't found.
4999	* @retval VERR_SUPDRV_INTERFACE_NOT_SUPPORTED if the interface wasn't supported.
5000	*
5001	* @param pSession The SUPDRV session.
5002	* @param pszName The name of the component factory.
5003	* @param pszInterfaceUuid The UUID of the factory interface (stringified).
5004	* @param ppvFactoryIf Where to store the factory interface.
5005	*/
5006	SUPR0DECL(int) SUPR0ComponentQueryFactory(PSUPDRVSESSION pSession, const char pszName, const char pszInterfaceUuid, void **ppvFactoryIf)
5007	{
5008	const char *pszEnd;
5009	size_t cchName;
5010	int rc;
5011
5012	/*
5013	* Validate parameters.
5014	*/
5015	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
5016
5017	AssertPtrReturn(pszName, VERR_INVALID_POINTER);
5018	pszEnd = RTStrEnd(pszName, RT_SIZEOFMEMB(SUPDRVFACTORY, szName));
5019	AssertReturn(pszEnd, VERR_INVALID_PARAMETER);
5020	cchName = pszEnd - pszName;
5021
5022	AssertPtrReturn(pszInterfaceUuid, VERR_INVALID_POINTER);
5023	pszEnd = RTStrEnd(pszInterfaceUuid, RTUUID_STR_LENGTH);
5024	AssertReturn(pszEnd, VERR_INVALID_PARAMETER);
5025
5026	AssertPtrReturn(ppvFactoryIf, VERR_INVALID_POINTER);
5027	*ppvFactoryIf = NULL;
5028
5029	/*
5030	* Take the lock and try all factories by this name.
5031	*/
5032	rc = RTSemFastMutexRequest(pSession->pDevExt->mtxComponentFactory);
5033	if (RT_SUCCESS(rc))
5034	{
5035	PSUPDRVFACTORYREG pCur = pSession->pDevExt->pComponentFactoryHead;
5036	rc = VERR_SUPDRV_COMPONENT_NOT_FOUND;
5037	while (pCur)
5038	{
5039	if ( pCur->cchName == cchName
5040	&& !memcmp(pCur->pFactory->szName, pszName, cchName))
5041	{
5042	void *pvFactory = pCur->pFactory->pfnQueryFactoryInterface(pCur->pFactory, pSession, pszInterfaceUuid);
5043	if (pvFactory)
5044	{
5045	*ppvFactoryIf = pvFactory;
5046	rc = VINF_SUCCESS;
5047	break;
5048	}
5049	rc = VERR_SUPDRV_INTERFACE_NOT_SUPPORTED;
5050	}
5051
5052	/* next */
5053	pCur = pCur->pNext;
5054	}
5055
5056	RTSemFastMutexRelease(pSession->pDevExt->mtxComponentFactory);
5057	}
5058	return rc;
5059	}
5060	SUPR0_EXPORT_SYMBOL(SUPR0ComponentQueryFactory);
5061
5062
5063	/**
5064	* Adds a memory object to the session.
5065	*
5066	* @returns IPRT status code.
5067	* @param pMem Memory tracking structure containing the
5068	* information to track.
5069	* @param pSession The session.
5070	*/
5071	static int supdrvMemAdd(PSUPDRVMEMREF pMem, PSUPDRVSESSION pSession)
5072	{
5073	PSUPDRVBUNDLE pBundle;
5074
5075	/*
5076	* Find free entry and record the allocation.
5077	*/
5078	RTSpinlockAcquire(pSession->Spinlock);
5079	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
5080	{
5081	if (pBundle->cUsed < RT_ELEMENTS(pBundle->aMem))
5082	{
5083	unsigned i;
5084	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
5085	{
5086	if (pBundle->aMem[i].MemObj == NIL_RTR0MEMOBJ)
5087	{
5088	pBundle->cUsed++;
5089	pBundle->aMem[i] = *pMem;
5090	RTSpinlockRelease(pSession->Spinlock);
5091	return VINF_SUCCESS;
5092	}
5093	}
5094	AssertFailed(); /* !!this can't be happening!!! */
5095	}
5096	}
5097	RTSpinlockRelease(pSession->Spinlock);
5098
5099	/*
5100	* Need to allocate a new bundle.
5101	* Insert into the last entry in the bundle.
5102	*/
5103	pBundle = (PSUPDRVBUNDLE)RTMemAllocZ(sizeof(*pBundle));
5104	if (!pBundle)
5105	return VERR_NO_MEMORY;
5106
5107	/* take last entry. */
5108	pBundle->cUsed++;
5109	pBundle->aMem[RT_ELEMENTS(pBundle->aMem) - 1] = *pMem;
5110
5111	/* insert into list. */
5112	RTSpinlockAcquire(pSession->Spinlock);
5113	pBundle->pNext = pSession->Bundle.pNext;
5114	pSession->Bundle.pNext = pBundle;
5115	RTSpinlockRelease(pSession->Spinlock);
5116
5117	return VINF_SUCCESS;
5118	}
5119
5120
5121	/**
5122	* Releases a memory object referenced by pointer and type.
5123	*
5124	* @returns IPRT status code.
5125	* @param pSession Session data.
5126	* @param uPtr Pointer to memory. This is matched against both the R0 and R3 addresses.
5127	* @param eType Memory type.
5128	*/
5129	static int supdrvMemRelease(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr, SUPDRVMEMREFTYPE eType)
5130	{
5131	PSUPDRVBUNDLE pBundle;
5132
5133	/*
5134	* Validate input.
5135	*/
5136	if (!uPtr)
5137	{
5138	Log(("Illegal address %p\n", (void *)uPtr));
5139	return VERR_INVALID_PARAMETER;
5140	}
5141
5142	/*
5143	* Search for the address.
5144	*/
5145	RTSpinlockAcquire(pSession->Spinlock);
5146	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
5147	{
5148	if (pBundle->cUsed > 0)
5149	{
5150	unsigned i;
5151	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
5152	{
5153	if ( pBundle->aMem[i].eType == eType
5154	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
5155	&& ( (RTHCUINTPTR)RTR0MemObjAddress(pBundle->aMem[i].MemObj) == uPtr
5156	\|\| ( pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ
5157	&& RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3) == uPtr))
5158	)
5159	{
5160	/* Make a copy of it and release it outside the spinlock. */
5161	SUPDRVMEMREF Mem = pBundle->aMem[i];
5162	pBundle->aMem[i].eType = MEMREF_TYPE_UNUSED;
5163	pBundle->aMem[i].MemObj = NIL_RTR0MEMOBJ;
5164	pBundle->aMem[i].MapObjR3 = NIL_RTR0MEMOBJ;
5165	RTSpinlockRelease(pSession->Spinlock);
5166
5167	if (Mem.MapObjR3 != NIL_RTR0MEMOBJ)
5168	{
5169	int rc = RTR0MemObjFree(Mem.MapObjR3, false);
5170	AssertRC(rc); /** @todo figure out how to handle this. */
5171	}
5172	if (Mem.MemObj != NIL_RTR0MEMOBJ)
5173	{
5174	int rc = RTR0MemObjFree(Mem.MemObj, true /* fFreeMappings */);
5175	AssertRC(rc); /** @todo figure out how to handle this. */
5176	}
5177	return VINF_SUCCESS;
5178	}
5179	}
5180	}
5181	}
5182	RTSpinlockRelease(pSession->Spinlock);
5183	Log(("Failed to find %p!!! (eType=%d)\n", (void *)uPtr, eType));
5184	return VERR_INVALID_PARAMETER;
5185	}
5186
5187
5188	/**
5189	* Opens an image. If it's the first time it's opened the call must upload
5190	* the bits using the supdrvIOCtl_LdrLoad() / SUPDRV_IOCTL_LDR_LOAD function.
5191	*
5192	* This is the 1st step of the loading.
5193	*
5194	* @returns IPRT status code.
5195	* @param pDevExt Device globals.
5196	* @param pSession Session data.
5197	* @param pReq The open request.
5198	*/
5199	static int supdrvIOCtl_LdrOpen(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDROPEN pReq)
5200	{
5201	int rc;
5202	PSUPDRVLDRIMAGE pImage;
5203	void *pv;
5204	size_t cchName = strlen(pReq->u.In.szName); /* (caller checked < 32). */
5205	SUPDRV_CHECK_SMAP_SETUP();
5206	SUPDRV_CHECK_SMAP_CHECK(pDevExt, RT_NOTHING);
5207	LogFlow(("supdrvIOCtl_LdrOpen: szName=%s cbImageWithEverything=%d\n", pReq->u.In.szName, pReq->u.In.cbImageWithEverything));
5208
5209	/*
5210	* Check if we got an instance of the image already.
5211	*/
5212	supdrvLdrLock(pDevExt);
5213	SUPDRV_CHECK_SMAP_CHECK(pDevExt, RT_NOTHING);
5214	for (pImage = pDevExt->pLdrImages; pImage; pImage = pImage->pNext)
5215	{
5216	if ( pImage->szName[cchName] == '\0'
5217	&& !memcmp(pImage->szName, pReq->u.In.szName, cchName))
5218	{
5219	/** @todo Add an _1M (or something) per session reference. */
5220	if (RT_LIKELY(pImage->cImgUsage < UINT32_MAX / 2U))
5221	{
5222	/** @todo check cbImageBits and cbImageWithEverything here, if they differs
5223	* that indicates that the images are different. */
5224	pReq->u.Out.pvImageBase = pImage->pvImage;
5225	pReq->u.Out.fNeedsLoading = pImage->uState == SUP_IOCTL_LDR_OPEN;
5226	pReq->u.Out.fNativeLoader = pImage->fNative;
5227	supdrvLdrAddUsage(pDevExt, pSession, pImage, true /fRing3Usage/);
5228	supdrvLdrUnlock(pDevExt);
5229	SUPDRV_CHECK_SMAP_CHECK(pDevExt, RT_NOTHING);
5230	return VINF_SUCCESS;
5231	}
5232	supdrvLdrUnlock(pDevExt);
5233	Log(("supdrvIOCtl_LdrOpen: Too many existing references to '%s'!\n", pReq->u.In.szName));
5234	return VERR_TOO_MANY_REFERENCES;
5235	}
5236	}
5237	/* (not found - add it!) */
5238
5239	/* If the loader interface is locked down, make userland fail early */
5240	if (pDevExt->fLdrLockedDown)
5241	{
5242	supdrvLdrUnlock(pDevExt);
5243	Log(("supdrvIOCtl_LdrOpen: Not adding '%s' to image list, loader interface is locked down!\n", pReq->u.In.szName));
5244	return VERR_PERMISSION_DENIED;
5245	}
5246
5247	/* Stop if caller doesn't wish to prepare loading things. */
5248	if (!pReq->u.In.cbImageBits)
5249	{
5250	supdrvLdrUnlock(pDevExt);
5251	Log(("supdrvIOCtl_LdrOpen: Returning VERR_MODULE_NOT_FOUND for '%s'!\n", pReq->u.In.szName));
5252	return VERR_MODULE_NOT_FOUND;
5253	}
5254
5255	/*
5256	* Allocate memory.
5257	*/
5258	Assert(cchName < sizeof(pImage->szName));
5259	pv = RTMemAllocZ(sizeof(SUPDRVLDRIMAGE));
5260	if (!pv)
5261	{
5262	supdrvLdrUnlock(pDevExt);
5263	Log(("supdrvIOCtl_LdrOpen: RTMemAllocZ() failed\n"));
5264	return VERR_NO_MEMORY;
5265	}
5266	SUPDRV_CHECK_SMAP_CHECK(pDevExt, RT_NOTHING);
5267
5268	/*
5269	* Setup and link in the LDR stuff.
5270	*/
5271	pImage = (PSUPDRVLDRIMAGE)pv;
5272	pImage->pvImage = NULL;
5273	#ifdef SUPDRV_USE_MEMOBJ_FOR_LDR_IMAGE
5274	pImage->hMemObjImage = NIL_RTR0MEMOBJ;
5275	#else
5276	pImage->pvImageAlloc = NULL;
5277	#endif
5278	pImage->cbImageWithEverything = pReq->u.In.cbImageWithEverything;
5279	pImage->cbImageBits = pReq->u.In.cbImageBits;
5280	pImage->cSymbols = 0;
5281	pImage->paSymbols = NULL;
5282	pImage->pachStrTab = NULL;
5283	pImage->cbStrTab = 0;
5284	pImage->cSegments = 0;
5285	pImage->paSegments = NULL;
5286	pImage->pfnModuleInit = NULL;
5287	pImage->pfnModuleTerm = NULL;
5288	pImage->pfnServiceReqHandler = NULL;
5289	pImage->uState = SUP_IOCTL_LDR_OPEN;
5290	pImage->cImgUsage = 0; /* Increased by supdrvLdrAddUsage later */
5291	pImage->pDevExt = pDevExt;
5292	pImage->pImageImport = NULL;
5293	pImage->uMagic = SUPDRVLDRIMAGE_MAGIC;
5294	pImage->pWrappedModInfo = NULL;
5295	memcpy(pImage->szName, pReq->u.In.szName, cchName + 1);
5296
5297	/*
5298	* Try load it using the native loader, if that isn't supported, fall back
5299	* on the older method.
5300	*/
5301	pImage->fNative = true;
5302	rc = supdrvOSLdrOpen(pDevExt, pImage, pReq->u.In.szFilename);
5303	if (rc == VERR_NOT_SUPPORTED)
5304	{
5305	#ifdef SUPDRV_USE_MEMOBJ_FOR_LDR_IMAGE
5306	rc = RTR0MemObjAllocPage(&pImage->hMemObjImage, pImage->cbImageBits, true /fExecutable/);
5307	if (RT_SUCCESS(rc))
5308	{
5309	pImage->pvImage = RTR0MemObjAddress(pImage->hMemObjImage);
5310	pImage->fNative = false;
5311	}
5312	#else
5313	pImage->pvImageAlloc = RTMemExecAlloc(pImage->cbImageBits + 31);
5314	pImage->pvImage = RT_ALIGN_P(pImage->pvImageAlloc, 32);
5315	pImage->fNative = false;
5316	rc = pImage->pvImageAlloc ? VINF_SUCCESS : VERR_NO_EXEC_MEMORY;
5317	#endif
5318	SUPDRV_CHECK_SMAP_CHECK(pDevExt, RT_NOTHING);
5319	}
5320	if (RT_SUCCESS(rc))
5321	rc = supdrvLdrAddUsage(pDevExt, pSession, pImage, true /fRing3Usage/);
5322	if (RT_FAILURE(rc))
5323	{
5324	supdrvLdrUnlock(pDevExt);
5325	pImage->uMagic = SUPDRVLDRIMAGE_MAGIC_DEAD;
5326	RTMemFree(pImage);
5327	Log(("supdrvIOCtl_LdrOpen(%s): failed - %Rrc\n", pReq->u.In.szName, rc));
5328	return rc;
5329	}
5330	Assert(RT_VALID_PTR(pImage->pvImage) \|\| RT_FAILURE(rc));
5331
5332	/*
5333	* Link it.
5334	*/
5335	pImage->pNext = pDevExt->pLdrImages;
5336	pDevExt->pLdrImages = pImage;
5337
5338	pReq->u.Out.pvImageBase = pImage->pvImage;
5339	pReq->u.Out.fNeedsLoading = true;
5340	pReq->u.Out.fNativeLoader = pImage->fNative;
5341	supdrvOSLdrNotifyOpened(pDevExt, pImage, pReq->u.In.szFilename);
5342
5343	supdrvLdrUnlock(pDevExt);
5344	SUPDRV_CHECK_SMAP_CHECK(pDevExt, RT_NOTHING);
5345	return VINF_SUCCESS;
5346	}
5347
5348
5349	/**
5350	* Formats a load error message.
5351	*
5352	* @returns @a rc
5353	* @param rc Return code.
5354	* @param pReq The request.
5355	* @param pszFormat The error message format string.
5356	* @param ... Argument to the format string.
5357	*/
5358	int VBOXCALL supdrvLdrLoadError(int rc, PSUPLDRLOAD pReq, const char *pszFormat, ...)
5359	{
5360	va_list va;
5361	va_start(va, pszFormat);
5362	pReq->u.Out.uErrorMagic = SUPLDRLOAD_ERROR_MAGIC;
5363	RTStrPrintfV(pReq->u.Out.szError, sizeof(pReq->u.Out.szError), pszFormat, va);
5364	va_end(va);
5365	Log(("SUP_IOCTL_LDR_LOAD: %s [rc=%Rrc]\n", pReq->u.Out.szError, rc));
5366	return rc;
5367	}
5368
5369
5370	/**
5371	* Worker that validates a pointer to an image entrypoint.
5372	*
5373	* Calls supdrvLdrLoadError on error.
5374	*
5375	* @returns IPRT status code.
5376	* @param pDevExt The device globals.
5377	* @param pImage The loader image.
5378	* @param pv The pointer into the image.
5379	* @param fMayBeNull Whether it may be NULL.
5380	* @param pszSymbol The entrypoint name or log name. If the symbol is
5381	* capitalized it signifies a specific symbol, otherwise it
5382	* for logging.
5383	* @param pbImageBits The image bits prepared by ring-3.
5384	* @param pReq The request for passing to supdrvLdrLoadError.
5385	*
5386	* @note Will leave the loader lock on failure!
5387	*/
5388	static int supdrvLdrValidatePointer(PSUPDRVDEVEXT pDevExt, PSUPDRVLDRIMAGE pImage, void *pv, bool fMayBeNull,
5389	const uint8_t pbImageBits, const char pszSymbol, PSUPLDRLOAD pReq)
5390	{
5391	if (!fMayBeNull \|\| pv)
5392	{
5393	uint32_t iSeg;
5394
5395	/* Must be within the image bits: */
5396	uintptr_t const uRva = (uintptr_t)pv - (uintptr_t)pImage->pvImage;
5397	if (uRva >= pImage->cbImageBits)
5398	{
5399	supdrvLdrUnlock(pDevExt);
5400	return supdrvLdrLoadError(VERR_INVALID_PARAMETER, pReq,
5401	"Invalid entry point address %p given for %s: RVA %#zx, image size %#zx",
5402	pv, pszSymbol, uRva, pImage->cbImageBits);
5403	}
5404
5405	/* Must be in an executable segment: */
5406	for (iSeg = 0; iSeg < pImage->cSegments; iSeg++)
5407	if (uRva - pImage->paSegments[iSeg].off < (uintptr_t)pImage->paSegments[iSeg].cb)
5408	{
5409	if (pImage->paSegments[iSeg].fProt & SUPLDR_PROT_EXEC)
5410	break;
5411	supdrvLdrUnlock(pDevExt);
5412	return supdrvLdrLoadError(VERR_INVALID_PARAMETER, pReq,
5413	"Bad entry point %p given for %s: not executable (seg #%u: %#RX32 LB %#RX32 prot %#x)",
5414	pv, pszSymbol, iSeg, pImage->paSegments[iSeg].off, pImage->paSegments[iSeg].cb,
5415	pImage->paSegments[iSeg].fProt);
5416	}
5417	if (iSeg >= pImage->cSegments)
5418	{
5419	supdrvLdrUnlock(pDevExt);
5420	return supdrvLdrLoadError(VERR_INVALID_PARAMETER, pReq,
5421	"Bad entry point %p given for %s: no matching segment found (RVA %#zx)!",
5422	pv, pszSymbol, uRva);
5423	}
5424
5425	if (pImage->fNative)
5426	{
5427	/** @todo pass pReq along to the native code. */
5428	int rc = supdrvOSLdrValidatePointer(pDevExt, pImage, pv, pbImageBits, pszSymbol);
5429	if (RT_FAILURE(rc))
5430	{
5431	supdrvLdrUnlock(pDevExt);
5432	return supdrvLdrLoadError(VERR_INVALID_PARAMETER, pReq,
5433	"Bad entry point address %p for %s: rc=%Rrc\n", pv, pszSymbol, rc);
5434	}
5435	}
5436	}
5437	return VINF_SUCCESS;
5438	}
5439
5440
5441	/**
5442	* Loads the image bits.
5443	*
5444	* This is the 2nd step of the loading.
5445	*
5446	* @returns IPRT status code.
5447	* @param pDevExt Device globals.
5448	* @param pSession Session data.
5449	* @param pReq The request.
5450	*/
5451	static int supdrvIOCtl_LdrLoad(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRLOAD pReq)
5452	{
5453	PSUPDRVLDRUSAGE pUsage;
5454	PSUPDRVLDRIMAGE pImage;
5455	PSUPDRVLDRIMAGE pImageImport;
5456	int rc;
5457	SUPDRV_CHECK_SMAP_SETUP();
5458	LogFlow(("supdrvIOCtl_LdrLoad: pvImageBase=%p cbImageWithEverything=%d\n", pReq->u.In.pvImageBase, pReq->u.In.cbImageWithEverything));
5459	SUPDRV_CHECK_SMAP_CHECK(pDevExt, RT_NOTHING);
5460
5461	/*
5462	* Find the ldr image.
5463	*/
5464	supdrvLdrLock(pDevExt);
5465	SUPDRV_CHECK_SMAP_CHECK(pDevExt, RT_NOTHING);
5466
5467	pUsage = pSession->pLdrUsage;
5468	while (pUsage && pUsage->pImage->pvImage != pReq->u.In.pvImageBase)
5469	pUsage = pUsage->pNext;
5470	if (!pUsage)
5471	{
5472	supdrvLdrUnlock(pDevExt);
5473	return supdrvLdrLoadError(VERR_INVALID_HANDLE, pReq, "Image not found");
5474	}
5475	pImage = pUsage->pImage;
5476
5477	/*
5478	* Validate input.
5479	*/
5480	if ( pImage->cbImageWithEverything != pReq->u.In.cbImageWithEverything
5481	\|\| pImage->cbImageBits != pReq->u.In.cbImageBits)
5482	{
5483	supdrvLdrUnlock(pDevExt);
5484	return supdrvLdrLoadError(VERR_INVALID_HANDLE, pReq, "Image size mismatch found: %u(prep) != %u(load) or %u != %u",
5485	pImage->cbImageWithEverything, pReq->u.In.cbImageWithEverything, pImage->cbImageBits, pReq->u.In.cbImageBits);
5486	}
5487
5488	if (pImage->uState != SUP_IOCTL_LDR_OPEN)
5489	{
5490	unsigned uState = pImage->uState;
5491	supdrvLdrUnlock(pDevExt);
5492	if (uState != SUP_IOCTL_LDR_LOAD)
5493	AssertMsgFailed(("SUP_IOCTL_LDR_LOAD: invalid image state %d (%#x)!\n", uState, uState));
5494	pReq->u.Out.uErrorMagic = 0;
5495	return VERR_ALREADY_LOADED;
5496	}
5497
5498	/* If the loader interface is locked down, don't load new images */
5499	if (pDevExt->fLdrLockedDown)
5500	{
5501	supdrvLdrUnlock(pDevExt);
5502	return supdrvLdrLoadError(VERR_PERMISSION_DENIED, pReq, "Loader is locked down");
5503	}
5504
5505	/*
5506	* If the new image is a dependant of VMMR0.r0, resolve it via the
5507	* caller's usage list and make sure it's in ready state.
5508	*/
5509	pImageImport = NULL;
5510	if (pReq->u.In.fFlags & SUPLDRLOAD_F_DEP_VMMR0)
5511	{
5512	PSUPDRVLDRUSAGE pUsageDependency = pSession->pLdrUsage;
5513	while (pUsageDependency && pUsageDependency->pImage->pvImage != pDevExt->pvVMMR0)
5514	pUsageDependency = pUsageDependency->pNext;
5515	if (!pUsageDependency \|\| !pDevExt->pvVMMR0)
5516	{
5517	supdrvLdrUnlock(pDevExt);
5518	return supdrvLdrLoadError(VERR_MODULE_NOT_FOUND, pReq, "VMMR0.r0 not loaded by session");
5519	}
5520	pImageImport = pUsageDependency->pImage;
5521	if (pImageImport->uState != SUP_IOCTL_LDR_LOAD)
5522	{
5523	supdrvLdrUnlock(pDevExt);
5524	return supdrvLdrLoadError(VERR_MODULE_NOT_FOUND, pReq, "VMMR0.r0 is not ready (state %#x)", pImageImport->uState);
5525	}
5526	}
5527
5528	/*
5529	* Copy the segments before we start using supdrvLdrValidatePointer for entrypoint validation.
5530	*/
5531	pImage->cSegments = pReq->u.In.cSegments;
5532	{
5533	size_t cbSegments = pImage->cSegments * sizeof(SUPLDRSEG);
5534	pImage->paSegments = (PSUPLDRSEG)RTMemDup(&pReq->u.In.abImage[pReq->u.In.offSegments], cbSegments);
5535	if (pImage->paSegments) /* Align the last segment size to avoid upsetting RTR0MemObjProtect. / /* @todo relax RTR0MemObjProtect */
5536	pImage->paSegments[pImage->cSegments - 1].cb = RT_ALIGN_32(pImage->paSegments[pImage->cSegments - 1].cb, PAGE_SIZE);
5537	else
5538	{
5539	supdrvLdrUnlock(pDevExt);
5540	return supdrvLdrLoadError(VERR_NO_MEMORY, pReq, "Out of memory for segment table: %#x", cbSegments);
5541	}
5542	SUPDRV_CHECK_SMAP_CHECK(pDevExt, RT_NOTHING);
5543	}
5544
5545	/*
5546	* Validate entrypoints.
5547	*/
5548	switch (pReq->u.In.eEPType)
5549	{
5550	case SUPLDRLOADEP_NOTHING:
5551	break;
5552
5553	case SUPLDRLOADEP_VMMR0:
5554	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.EP.VMMR0.pvVMMR0EntryFast, false, pReq->u.In.abImage, "VMMR0EntryFast", pReq);
5555	if (RT_FAILURE(rc))
5556	return rc;
5557	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.EP.VMMR0.pvVMMR0EntryEx, false, pReq->u.In.abImage, "VMMR0EntryEx", pReq);
5558	if (RT_FAILURE(rc))
5559	return rc;
5560
5561	/* Fail here if there is already a VMMR0 module. */
5562	if (pDevExt->pvVMMR0 != NULL)
5563	{
5564	supdrvLdrUnlock(pDevExt);
5565	return supdrvLdrLoadError(VERR_INVALID_PARAMETER, pReq, "There is already a VMMR0 module loaded (%p)", pDevExt->pvVMMR0);
5566	}
5567	break;
5568
5569	case SUPLDRLOADEP_SERVICE:
5570	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.EP.Service.pfnServiceReq, false, pReq->u.In.abImage, "pfnServiceReq", pReq);
5571	if (RT_FAILURE(rc))
5572	return rc;
5573	if ( pReq->u.In.EP.Service.apvReserved[0] != NIL_RTR0PTR
5574	\|\| pReq->u.In.EP.Service.apvReserved[1] != NIL_RTR0PTR
5575	\|\| pReq->u.In.EP.Service.apvReserved[2] != NIL_RTR0PTR)
5576	{
5577	supdrvLdrUnlock(pDevExt);
5578	return supdrvLdrLoadError(VERR_INVALID_PARAMETER, pReq, "apvReserved={%p,%p,%p} MBZ!",
5579	pReq->u.In.EP.Service.apvReserved[0], pReq->u.In.EP.Service.apvReserved[1],
5580	pReq->u.In.EP.Service.apvReserved[2]);
5581	}
5582	break;
5583
5584	default:
5585	supdrvLdrUnlock(pDevExt);
5586	return supdrvLdrLoadError(VERR_INVALID_PARAMETER, pReq, "Invalid eEPType=%d", pReq->u.In.eEPType);
5587	}
5588
5589	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.pfnModuleInit, true, pReq->u.In.abImage, "ModuleInit", pReq);
5590	if (RT_FAILURE(rc))
5591	return rc;
5592	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.pfnModuleTerm, true, pReq->u.In.abImage, "ModuleTerm", pReq);
5593	if (RT_FAILURE(rc))
5594	return rc;
5595	SUPDRV_CHECK_SMAP_CHECK(pDevExt, RT_NOTHING);
5596
5597	/*
5598	* Allocate and copy the tables if non-native.
5599	* (No need to do try/except as this is a buffered request.)
5600	*/
5601	if (!pImage->fNative)
5602	{
5603	pImage->cbStrTab = pReq->u.In.cbStrTab;
5604	if (pImage->cbStrTab)
5605	{
5606	pImage->pachStrTab = (char *)RTMemDup(&pReq->u.In.abImage[pReq->u.In.offStrTab], pImage->cbStrTab);
5607	if (!pImage->pachStrTab)
5608	rc = supdrvLdrLoadError(VERR_NO_MEMORY, pReq, "Out of memory for string table: %#x", pImage->cbStrTab);
5609	SUPDRV_CHECK_SMAP_CHECK(pDevExt, RT_NOTHING);
5610	}
5611
5612	pImage->cSymbols = pReq->u.In.cSymbols;
5613	if (RT_SUCCESS(rc) && pImage->cSymbols)
5614	{
5615	size_t cbSymbols = pImage->cSymbols * sizeof(SUPLDRSYM);
5616	pImage->paSymbols = (PSUPLDRSYM)RTMemDup(&pReq->u.In.abImage[pReq->u.In.offSymbols], cbSymbols);
5617	if (!pImage->paSymbols)
5618	rc = supdrvLdrLoadError(VERR_NO_MEMORY, pReq, "Out of memory for symbol table: %#x", cbSymbols);
5619	SUPDRV_CHECK_SMAP_CHECK(pDevExt, RT_NOTHING);
5620	}
5621	}
5622
5623	/*
5624	* Copy the bits and apply permissions / complete native loading.
5625	*/
5626	if (RT_SUCCESS(rc))
5627	{
5628	pImage->uState = SUP_IOCTL_LDR_LOAD;
5629	pImage->pfnModuleInit = (PFNR0MODULEINIT)(uintptr_t)pReq->u.In.pfnModuleInit;
5630	pImage->pfnModuleTerm = (PFNR0MODULETERM)(uintptr_t)pReq->u.In.pfnModuleTerm;
5631
5632	if (pImage->fNative)
5633	rc = supdrvOSLdrLoad(pDevExt, pImage, pReq->u.In.abImage, pReq);
5634	else
5635	{
5636	#ifdef SUPDRV_USE_MEMOBJ_FOR_LDR_IMAGE
5637	uint32_t i;
5638	memcpy(pImage->pvImage, &pReq->u.In.abImage[0], pImage->cbImageBits);
5639
5640	for (i = 0; i < pImage->cSegments; i++)
5641	{
5642	rc = RTR0MemObjProtect(pImage->hMemObjImage, pImage->paSegments[i].off, pImage->paSegments[i].cb,
5643	pImage->paSegments[i].fProt);
5644	if (RT_SUCCESS(rc))
5645	continue;
5646	if (rc == VERR_NOT_SUPPORTED)
5647	rc = VINF_SUCCESS;
5648	else
5649	rc = supdrvLdrLoadError(rc, pReq, "RTR0MemObjProtect failed on seg#%u %#RX32 LB %#RX32 fProt=%#x",
5650	i, pImage->paSegments[i].off, pImage->paSegments[i].cb, pImage->paSegments[i].fProt);
5651	break;
5652	}
5653	#else
5654	memcpy(pImage->pvImage, &pReq->u.In.abImage[0], pImage->cbImageBits);
5655	#endif
5656	Log(("vboxdrv: Loaded '%s' at %p\n", pImage->szName, pImage->pvImage));
5657	}
5658	SUPDRV_CHECK_SMAP_CHECK(pDevExt, RT_NOTHING);
5659	}
5660
5661	/*
5662	* On success call the module initialization.
5663	*/
5664	LogFlow(("supdrvIOCtl_LdrLoad: pfnModuleInit=%p\n", pImage->pfnModuleInit));
5665	if (RT_SUCCESS(rc) && pImage->pfnModuleInit)
5666	{
5667	Log(("supdrvIOCtl_LdrLoad: calling pfnModuleInit=%p\n", pImage->pfnModuleInit));
5668	pDevExt->pLdrInitImage = pImage;
5669	pDevExt->hLdrInitThread = RTThreadNativeSelf();
5670	SUPDRV_CHECK_SMAP_CHECK(pDevExt, RT_NOTHING);
5671	rc = pImage->pfnModuleInit(pImage);
5672	SUPDRV_CHECK_SMAP_CHECK(pDevExt, RT_NOTHING);
5673	pDevExt->pLdrInitImage = NULL;
5674	pDevExt->hLdrInitThread = NIL_RTNATIVETHREAD;
5675	if (RT_FAILURE(rc))
5676	supdrvLdrLoadError(rc, pReq, "ModuleInit failed: %Rrc", rc);
5677	}
5678	if (RT_SUCCESS(rc))
5679	{
5680	/*
5681	* Publish any standard entry points.
5682	*/
5683	switch (pReq->u.In.eEPType)
5684	{
5685	case SUPLDRLOADEP_VMMR0:
5686	Assert(!pDevExt->pvVMMR0);
5687	Assert(!pDevExt->pfnVMMR0EntryFast);
5688	Assert(!pDevExt->pfnVMMR0EntryEx);
5689	ASMAtomicWritePtrVoid(&pDevExt->pvVMMR0, pImage->pvImage);
5690	ASMAtomicWritePtrVoid((void * volatile *)(uintptr_t)&pDevExt->pfnVMMR0EntryFast,
5691	(void *)(uintptr_t) pReq->u.In.EP.VMMR0.pvVMMR0EntryFast);
5692	ASMAtomicWritePtrVoid((void * volatile *)(uintptr_t)&pDevExt->pfnVMMR0EntryEx,
5693	(void *)(uintptr_t) pReq->u.In.EP.VMMR0.pvVMMR0EntryEx);
5694	break;
5695	case SUPLDRLOADEP_SERVICE:
5696	pImage->pfnServiceReqHandler = (PFNSUPR0SERVICEREQHANDLER)(uintptr_t)pReq->u.In.EP.Service.pfnServiceReq;
5697	break;
5698	default:
5699	break;
5700	}
5701
5702	/*
5703	* Increase the usage counter of any imported image.
5704	*/
5705	if (pImageImport)
5706	{
5707	pImageImport->cImgUsage++;
5708	if (pImageImport->cImgUsage == 2 && pImageImport->pWrappedModInfo)
5709	supdrvOSLdrRetainWrapperModule(pDevExt, pImageImport);
5710	pImage->pImageImport = pImageImport;
5711	}
5712
5713	/*
5714	* Done!
5715	*/
5716	SUPR0Printf("vboxdrv: %RKv %s\n", pImage->pvImage, pImage->szName);
5717	pReq->u.Out.uErrorMagic = 0;
5718	pReq->u.Out.szError[0] = '\0';
5719	}
5720	else
5721	{
5722	/* Inform the tracing component in case ModuleInit registered TPs. */
5723	supdrvTracerModuleUnloading(pDevExt, pImage);
5724
5725	pImage->uState = SUP_IOCTL_LDR_OPEN;
5726	pImage->pfnModuleInit = NULL;
5727	pImage->pfnModuleTerm = NULL;
5728	pImage->pfnServiceReqHandler= NULL;
5729	pImage->cbStrTab = 0;
5730	RTMemFree(pImage->pachStrTab);
5731	pImage->pachStrTab = NULL;
5732	RTMemFree(pImage->paSymbols);
5733	pImage->paSymbols = NULL;
5734	pImage->cSymbols = 0;
5735	}
5736
5737	supdrvLdrUnlock(pDevExt);
5738	SUPDRV_CHECK_SMAP_CHECK(pDevExt, RT_NOTHING);
5739	return rc;
5740	}
5741
5742
5743	/**
5744	* Registers a .r0 module wrapped in a native one and manually loaded.
5745	*
5746	* @returns VINF_SUCCESS or error code (no info statuses).
5747	* @param pDevExt Device globals.
5748	* @param pWrappedModInfo The wrapped module info.
5749	* @param pvNative OS specific information.
5750	* @param phMod Where to store the module handle.
5751	*/
5752	int VBOXCALL supdrvLdrRegisterWrappedModule(PSUPDRVDEVEXT pDevExt, PCSUPLDRWRAPPEDMODULE pWrappedModInfo,
5753	void pvNative, void *phMod)
5754	{
5755	size_t cchName;
5756	PSUPDRVLDRIMAGE pImage;
5757	PCSUPLDRWRAPMODSYMBOL paSymbols;
5758	uint16_t idx;
5759	const char *pszPrevSymbol;
5760	int rc;
5761	SUPDRV_CHECK_SMAP_SETUP();
5762	SUPDRV_CHECK_SMAP_CHECK(pDevExt, RT_NOTHING);
5763
5764	/*
5765	* Validate input.
5766	*/
5767	AssertPtrReturn(phMod, VERR_INVALID_POINTER);
5768	*phMod = NULL;
5769	AssertPtrReturn(pDevExt, VERR_INTERNAL_ERROR_2);
5770
5771	AssertPtrReturn(pWrappedModInfo, VERR_INVALID_POINTER);
5772	AssertMsgReturn(pWrappedModInfo->uMagic == SUPLDRWRAPPEDMODULE_MAGIC,
5773	("uMagic=%#x, expected %#x\n", pWrappedModInfo->uMagic, SUPLDRWRAPPEDMODULE_MAGIC),
5774	VERR_INVALID_MAGIC);
5775	AssertMsgReturn(pWrappedModInfo->uVersion == SUPLDRWRAPPEDMODULE_VERSION,
5776	("Unsupported uVersion=%#x, current version %#x\n", pWrappedModInfo->uVersion, SUPLDRWRAPPEDMODULE_VERSION),
5777	VERR_VERSION_MISMATCH);
5778	AssertMsgReturn(pWrappedModInfo->uEndMagic == SUPLDRWRAPPEDMODULE_MAGIC,
5779	("uEndMagic=%#x, expected %#x\n", pWrappedModInfo->uEndMagic, SUPLDRWRAPPEDMODULE_MAGIC),
5780	VERR_INVALID_MAGIC);
5781	AssertMsgReturn(pWrappedModInfo->fFlags <= SUPLDRWRAPPEDMODULE_F_VMMR0, ("Unknown flags in: %#x\n", pWrappedModInfo->fFlags),
5782	VERR_INVALID_FLAGS);
5783
5784	/* szName: */
5785	AssertReturn(RTStrEnd(pWrappedModInfo->szName, sizeof(pWrappedModInfo->szName)) != NULL, VERR_INVALID_NAME);
5786	AssertReturn(supdrvIsLdrModuleNameValid(pWrappedModInfo->szName), VERR_INVALID_NAME);
5787	AssertCompile(sizeof(pImage->szName) == sizeof(pWrappedModInfo->szName));
5788	cchName = strlen(pWrappedModInfo->szName);
5789
5790	/* Image range: */
5791	AssertPtrReturn(pWrappedModInfo->pvImageStart, VERR_INVALID_POINTER);
5792	AssertPtrReturn(pWrappedModInfo->pvImageEnd, VERR_INVALID_POINTER);
5793	AssertReturn((uintptr_t)pWrappedModInfo->pvImageEnd > (uintptr_t)pWrappedModInfo->pvImageStart, VERR_INVALID_PARAMETER);
5794
5795	/* Symbol table: */
5796	AssertMsgReturn(pWrappedModInfo->cSymbols <= _8K, ("Too many symbols: %u, max 8192\n", pWrappedModInfo->cSymbols),
5797	VERR_TOO_MANY_SYMLINKS);
5798	pszPrevSymbol = "\x7f";
5799	paSymbols = pWrappedModInfo->paSymbols;
5800	idx = pWrappedModInfo->cSymbols;
5801	while (idx-- > 0)
5802	{
5803	const char *pszSymbol = paSymbols[idx].pszSymbol;
5804	AssertMsgReturn(RT_VALID_PTR(pszSymbol) && RT_VALID_PTR(paSymbols[idx].pfnValue),
5805	("paSymbols[%u]: %p/%p\n", idx, pszSymbol, paSymbols[idx].pfnValue),
5806	VERR_INVALID_POINTER);
5807	AssertReturn(*pszSymbol != '\0', VERR_EMPTY_STRING);
5808	AssertMsgReturn(strcmp(pszSymbol, pszPrevSymbol) < 0,
5809	("symbol table out of order at index %u: '%s' vs '%s'\n", idx, pszSymbol, pszPrevSymbol),
5810	VERR_WRONG_ORDER);
5811	pszPrevSymbol = pszSymbol;
5812	}
5813
5814	/* Standard entry points: */
5815	AssertPtrNullReturn(pWrappedModInfo->pfnModuleInit, VERR_INVALID_POINTER);
5816	AssertPtrNullReturn(pWrappedModInfo->pfnModuleTerm, VERR_INVALID_POINTER);
5817	AssertReturn((uintptr_t)pWrappedModInfo->pfnModuleInit != (uintptr_t)pWrappedModInfo->pfnModuleTerm \|\| pWrappedModInfo->pfnModuleInit == NULL,
5818	VERR_INVALID_PARAMETER);
5819	if (pWrappedModInfo->fFlags & SUPLDRWRAPPEDMODULE_F_VMMR0)
5820	{
5821	AssertReturn(pWrappedModInfo->pfnServiceReqHandler == NULL, VERR_INVALID_PARAMETER);
5822	AssertPtrReturn(pWrappedModInfo->pfnVMMR0EntryFast, VERR_INVALID_POINTER);
5823	AssertPtrReturn(pWrappedModInfo->pfnVMMR0EntryEx, VERR_INVALID_POINTER);
5824	AssertReturn(pWrappedModInfo->pfnVMMR0EntryFast != pWrappedModInfo->pfnVMMR0EntryEx, VERR_INVALID_PARAMETER);
5825	}
5826	else
5827	{
5828	AssertPtrNullReturn(pWrappedModInfo->pfnServiceReqHandler, VERR_INVALID_POINTER);
5829	AssertReturn(pWrappedModInfo->pfnVMMR0EntryFast == NULL, VERR_INVALID_PARAMETER);
5830	AssertReturn(pWrappedModInfo->pfnVMMR0EntryEx == NULL, VERR_INVALID_PARAMETER);
5831	}
5832
5833	/*
5834	* Check if we got an instance of the image already.
5835	*/
5836	supdrvLdrLock(pDevExt);
5837	SUPDRV_CHECK_SMAP_CHECK(pDevExt, RT_NOTHING);
5838	for (pImage = pDevExt->pLdrImages; pImage; pImage = pImage->pNext)
5839	{
5840	if ( pImage->szName[cchName] == '\0'
5841	&& !memcmp(pImage->szName, pWrappedModInfo->szName, cchName))
5842	{
5843	supdrvLdrUnlock(pDevExt);
5844	Log(("supdrvLdrRegisterWrappedModule: '%s' already loaded!\n", pWrappedModInfo->szName));
5845	return VERR_ALREADY_LOADED;
5846	}
5847	}
5848	/* (not found - add it!) */
5849
5850	/* If the loader interface is locked down, make userland fail early */
5851	if (pDevExt->fLdrLockedDown)
5852	{
5853	supdrvLdrUnlock(pDevExt);
5854	Log(("supdrvLdrRegisterWrappedModule: Not adding '%s' to image list, loader interface is locked down!\n", pWrappedModInfo->szName));
5855	return VERR_PERMISSION_DENIED;
5856	}
5857
5858	/* Only one VMMR0: */
5859	if ( pDevExt->pvVMMR0 != NULL
5860	&& (pWrappedModInfo->fFlags & SUPLDRWRAPPEDMODULE_F_VMMR0))
5861	{
5862	supdrvLdrUnlock(pDevExt);
5863	Log(("supdrvLdrRegisterWrappedModule: Rejecting '%s' as we already got a VMMR0 module!\n", pWrappedModInfo->szName));
5864	return VERR_ALREADY_EXISTS;
5865	}
5866
5867	/*
5868	* Allocate memory.
5869	*/
5870	Assert(cchName < sizeof(pImage->szName));
5871	pImage = (PSUPDRVLDRIMAGE)RTMemAllocZ(sizeof(SUPDRVLDRIMAGE));
5872	if (!pImage)
5873	{
5874	supdrvLdrUnlock(pDevExt);
5875	Log(("supdrvLdrRegisterWrappedModule: RTMemAllocZ() failed\n"));
5876	return VERR_NO_MEMORY;
5877	}
5878	SUPDRV_CHECK_SMAP_CHECK(pDevExt, RT_NOTHING);
5879
5880	/*
5881	* Setup and link in the LDR stuff.
5882	*/
5883	pImage->pvImage = (void *)pWrappedModInfo->pvImageStart;
5884	#ifdef SUPDRV_USE_MEMOBJ_FOR_LDR_IMAGE
5885	pImage->hMemObjImage = NIL_RTR0MEMOBJ;
5886	#else
5887	pImage->pvImageAlloc = NULL;
5888	#endif
5889	pImage->cbImageWithEverything
5890	= pImage->cbImageBits = (uintptr_t)pWrappedModInfo->pvImageEnd - (uintptr_t)pWrappedModInfo->pvImageStart;
5891	pImage->cSymbols = 0;
5892	pImage->paSymbols = NULL;
5893	pImage->pachStrTab = NULL;
5894	pImage->cbStrTab = 0;
5895	pImage->cSegments = 0;
5896	pImage->paSegments = NULL;
5897	pImage->pfnModuleInit = pWrappedModInfo->pfnModuleInit;
5898	pImage->pfnModuleTerm = pWrappedModInfo->pfnModuleTerm;
5899	pImage->pfnServiceReqHandler = NULL; /* Only setting this after module init */
5900	pImage->uState = SUP_IOCTL_LDR_LOAD;
5901	pImage->cImgUsage = 1; /* Held by the wrapper module till unload. */
5902	pImage->pDevExt = pDevExt;
5903	pImage->pImageImport = NULL;
5904	pImage->uMagic = SUPDRVLDRIMAGE_MAGIC;
5905	pImage->pWrappedModInfo = pWrappedModInfo;
5906	pImage->pvWrappedNative = pvNative;
5907	pImage->fNative = true;
5908	memcpy(pImage->szName, pWrappedModInfo->szName, cchName + 1);
5909
5910	/*
5911	* Link it.
5912	*/
5913	pImage->pNext = pDevExt->pLdrImages;
5914	pDevExt->pLdrImages = pImage;
5915
5916	/*
5917	* Call module init function if found.
5918	*/
5919	rc = VINF_SUCCESS;
5920	if (pImage->pfnModuleInit)
5921	{
5922	Log(("supdrvIOCtl_LdrLoad: calling pfnModuleInit=%p\n", pImage->pfnModuleInit));
5923	pDevExt->pLdrInitImage = pImage;
5924	pDevExt->hLdrInitThread = RTThreadNativeSelf();
5925	SUPDRV_CHECK_SMAP_CHECK(pDevExt, RT_NOTHING);
5926	rc = pImage->pfnModuleInit(pImage);
5927	SUPDRV_CHECK_SMAP_CHECK(pDevExt, RT_NOTHING);
5928	pDevExt->pLdrInitImage = NULL;
5929	pDevExt->hLdrInitThread = NIL_RTNATIVETHREAD;
5930	}
5931	if (RT_SUCCESS(rc))
5932	{
5933	/*
5934	* Update entry points.
5935	*/
5936	if (pWrappedModInfo->fFlags & SUPLDRWRAPPEDMODULE_F_VMMR0)
5937	{
5938	Assert(!pDevExt->pvVMMR0);
5939	Assert(!pDevExt->pfnVMMR0EntryFast);
5940	Assert(!pDevExt->pfnVMMR0EntryEx);
5941	ASMAtomicWritePtrVoid(&pDevExt->pvVMMR0, pImage->pvImage);
5942	ASMAtomicWritePtrVoid((void * volatile *)(uintptr_t)&pDevExt->pfnVMMR0EntryFast,
5943	(void *)(uintptr_t) pWrappedModInfo->pfnVMMR0EntryFast);
5944	ASMAtomicWritePtrVoid((void * volatile *)(uintptr_t)&pDevExt->pfnVMMR0EntryEx,
5945	(void *)(uintptr_t) pWrappedModInfo->pfnVMMR0EntryEx);
5946	}
5947	else
5948	pImage->pfnServiceReqHandler = pWrappedModInfo->pfnServiceReqHandler;
5949	#ifdef IN_RING3
5950	# error "WTF?"
5951	#endif
5952	*phMod = pImage;
5953	}
5954	else
5955	{
5956	/*
5957	* Module init failed - bail, no module term callout.
5958	*/
5959	SUPR0Printf("ModuleInit failed for '%s': %Rrc\n", pImage->szName, rc);
5960
5961	pImage->pfnModuleTerm = NULL;
5962	pImage->uState = SUP_IOCTL_LDR_OPEN;
5963	supdrvLdrFree(pDevExt, pImage);
5964	}
5965
5966	supdrvLdrUnlock(pDevExt);
5967	SUPDRV_CHECK_SMAP_CHECK(pDevExt, RT_NOTHING);
5968	return VINF_SUCCESS;
5969	}
5970
5971
5972	/**
5973	* Decrements SUPDRVLDRIMAGE::cImgUsage when two or greater.
5974	*
5975	* @param pDevExt Device globals.
5976	* @param pImage The image.
5977	* @param cReference Number of references being removed.
5978	*/
5979	DECLINLINE(void) supdrvLdrSubtractUsage(PSUPDRVDEVEXT pDevExt, PSUPDRVLDRIMAGE pImage, uint32_t cReference)
5980	{
5981	Assert(cReference > 0);
5982	Assert(pImage->cImgUsage > cReference);
5983	pImage->cImgUsage -= cReference;
5984	if (pImage->cImgUsage == 1 && pImage->pWrappedModInfo)
5985	supdrvOSLdrReleaseWrapperModule(pDevExt, pImage);
5986	}
5987
5988
5989	/**
5990	* Frees a previously loaded (prep'ed) image.
5991	*
5992	* @returns IPRT status code.
5993	* @param pDevExt Device globals.
5994	* @param pSession Session data.
5995	* @param pReq The request.
5996	*/
5997	static int supdrvIOCtl_LdrFree(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRFREE pReq)
5998	{
5999	int rc;
6000	PSUPDRVLDRUSAGE pUsagePrev;
6001	PSUPDRVLDRUSAGE pUsage;
6002	PSUPDRVLDRIMAGE pImage;
6003	LogFlow(("supdrvIOCtl_LdrFree: pvImageBase=%p\n", pReq->u.In.pvImageBase));
6004
6005	/*
6006	* Find the ldr image.
6007	*/
6008	supdrvLdrLock(pDevExt);
6009	pUsagePrev = NULL;
6010	pUsage = pSession->pLdrUsage;
6011	while (pUsage && pUsage->pImage->pvImage != pReq->u.In.pvImageBase)
6012	{
6013	pUsagePrev = pUsage;
6014	pUsage = pUsage->pNext;
6015	}
6016	if (!pUsage)
6017	{
6018	supdrvLdrUnlock(pDevExt);
6019	Log(("SUP_IOCTL_LDR_FREE: couldn't find image!\n"));
6020	return VERR_INVALID_HANDLE;
6021	}
6022	if (pUsage->cRing3Usage == 0)
6023	{
6024	supdrvLdrUnlock(pDevExt);
6025	Log(("SUP_IOCTL_LDR_FREE: No ring-3 reference to the image!\n"));
6026	return VERR_CALLER_NO_REFERENCE;
6027	}
6028
6029	/*
6030	* Check if we can remove anything.
6031	*/
6032	rc = VINF_SUCCESS;
6033	pImage = pUsage->pImage;
6034	Log(("SUP_IOCTL_LDR_FREE: pImage=%p %s cImgUsage=%d r3=%d r0=%u\n",
6035	pImage, pImage->szName, pImage->cImgUsage, pUsage->cRing3Usage, pUsage->cRing0Usage));
6036	if (pImage->cImgUsage <= 1 \|\| pUsage->cRing3Usage + pUsage->cRing0Usage <= 1)
6037	{
6038	/*
6039	* Check if there are any objects with destructors in the image, if
6040	* so leave it for the session cleanup routine so we get a chance to
6041	* clean things up in the right order and not leave them all dangling.
6042	*/
6043	RTSpinlockAcquire(pDevExt->Spinlock);
6044	if (pImage->cImgUsage <= 1)
6045	{
6046	PSUPDRVOBJ pObj;
6047	for (pObj = pDevExt->pObjs; pObj; pObj = pObj->pNext)
6048	if (RT_UNLIKELY((uintptr_t)pObj->pfnDestructor - (uintptr_t)pImage->pvImage < pImage->cbImageBits))
6049	{
6050	rc = VERR_DANGLING_OBJECTS;
6051	break;
6052	}
6053	}
6054	else
6055	{
6056	PSUPDRVUSAGE pGenUsage;
6057	for (pGenUsage = pSession->pUsage; pGenUsage; pGenUsage = pGenUsage->pNext)
6058	if (RT_UNLIKELY((uintptr_t)pGenUsage->pObj->pfnDestructor - (uintptr_t)pImage->pvImage < pImage->cbImageBits))
6059	{
6060	rc = VERR_DANGLING_OBJECTS;
6061	break;
6062	}
6063	}
6064	RTSpinlockRelease(pDevExt->Spinlock);
6065	if (rc == VINF_SUCCESS)
6066	{
6067	/* unlink it */
6068	if (pUsagePrev)
6069	pUsagePrev->pNext = pUsage->pNext;
6070	else
6071	pSession->pLdrUsage = pUsage->pNext;
6072
6073	/* free it */
6074	pUsage->pImage = NULL;
6075	pUsage->pNext = NULL;
6076	RTMemFree(pUsage);
6077
6078	/*
6079	* Dereference the image.
6080	*/
6081	if (pImage->cImgUsage <= 1)
6082	supdrvLdrFree(pDevExt, pImage);
6083	else
6084	supdrvLdrSubtractUsage(pDevExt, pImage, 1);
6085	}
6086	else
6087	Log(("supdrvIOCtl_LdrFree: Dangling objects in %p/%s!\n", pImage->pvImage, pImage->szName));
6088	}
6089	else
6090	{
6091	/*
6092	* Dereference both image and usage.
6093	*/
6094	pUsage->cRing3Usage--;
6095	supdrvLdrSubtractUsage(pDevExt, pImage, 1);
6096	}
6097
6098	supdrvLdrUnlock(pDevExt);
6099	return rc;
6100	}
6101
6102
6103	/**
6104	* Deregisters a wrapped .r0 module.
6105	*
6106	* @param pDevExt Device globals.
6107	* @param pWrappedModInfo The wrapped module info.
6108	* @param phMod Where to store the module is stored (NIL'ed on
6109	* success).
6110	*/
6111	int VBOXCALL supdrvLdrDeregisterWrappedModule(PSUPDRVDEVEXT pDevExt, PCSUPLDRWRAPPEDMODULE pWrappedModInfo, void **phMod)
6112	{
6113	PSUPDRVLDRIMAGE pImage;
6114	uint32_t cSleeps;
6115
6116	/*
6117	* Validate input.
6118	*/
6119	AssertPtrReturn(pWrappedModInfo, VERR_INVALID_POINTER);
6120	AssertMsgReturn(pWrappedModInfo->uMagic == SUPLDRWRAPPEDMODULE_MAGIC,
6121	("uMagic=%#x, expected %#x\n", pWrappedModInfo->uMagic, SUPLDRWRAPPEDMODULE_MAGIC),
6122	VERR_INVALID_MAGIC);
6123	AssertMsgReturn(pWrappedModInfo->uEndMagic == SUPLDRWRAPPEDMODULE_MAGIC,
6124	("uEndMagic=%#x, expected %#x\n", pWrappedModInfo->uEndMagic, SUPLDRWRAPPEDMODULE_MAGIC),
6125	VERR_INVALID_MAGIC);
6126
6127	AssertPtrReturn(phMod, VERR_INVALID_POINTER);
6128	pImage = (PSUPDRVLDRIMAGE )phMod;
6129	if (!pImage)
6130	return VINF_SUCCESS;
6131	AssertPtrReturn(pImage, VERR_INVALID_POINTER);
6132	AssertMsgReturn(pImage->uMagic == SUPDRVLDRIMAGE_MAGIC, ("pImage=%p uMagic=%#x\n", pImage, pImage->uMagic),
6133	VERR_INVALID_MAGIC);
6134	AssertMsgReturn(pImage->pvImage == pWrappedModInfo->pvImageStart,
6135	("pWrappedModInfo(%p)->pvImageStart=%p vs. pImage(=%p)->pvImage=%p\n",
6136	pWrappedModInfo, pWrappedModInfo->pvImageStart, pImage, pImage->pvImage),
6137	VERR_MISMATCH);
6138
6139	AssertPtrReturn(pDevExt, VERR_INVALID_POINTER);
6140
6141	/*
6142	* Try free it, but first we have to wait for its usage count to reach 1 (our).
6143	*/
6144	supdrvLdrLock(pDevExt);
6145	for (cSleeps = 0; ; cSleeps++)
6146	{
6147	PSUPDRVLDRIMAGE pCur;
6148
6149	/* Check that the image is in the list. */
6150	for (pCur = pDevExt->pLdrImages; pCur; pCur = pCur->pNext)
6151	if (pCur == pImage)
6152	break;
6153	AssertBreak(pCur == pImage);
6154
6155	/* Anyone still using it? */
6156	if (pImage->cImgUsage <= 1)
6157	break;
6158
6159	/* Someone is using it, wait and check again. */
6160	if (!(cSleeps % 60))
6161	SUPR0Printf("supdrvLdrUnregisterWrappedModule: Still %u users of wrapped image '%s' ...\n",
6162	pImage->cImgUsage, pImage->szName);
6163	supdrvLdrUnlock(pDevExt);
6164	RTThreadSleep(1000);
6165	supdrvLdrLock(pDevExt);
6166	}
6167
6168	/* We're the last 'user', free it. */
6169	supdrvLdrFree(pDevExt, pImage);
6170
6171	supdrvLdrUnlock(pDevExt);
6172
6173	*phMod = NULL;
6174	return VINF_SUCCESS;
6175	}
6176
6177
6178	/**
6179	* Lock down the image loader interface.
6180	*
6181	* @returns IPRT status code.
6182	* @param pDevExt Device globals.
6183	*/
6184	static int supdrvIOCtl_LdrLockDown(PSUPDRVDEVEXT pDevExt)
6185	{
6186	LogFlow(("supdrvIOCtl_LdrLockDown:\n"));
6187
6188	supdrvLdrLock(pDevExt);
6189	if (!pDevExt->fLdrLockedDown)
6190	{
6191	pDevExt->fLdrLockedDown = true;
6192	Log(("supdrvIOCtl_LdrLockDown: Image loader interface locked down\n"));
6193	}
6194	supdrvLdrUnlock(pDevExt);
6195
6196	return VINF_SUCCESS;
6197	}
6198
6199
6200	/**
6201	* Worker for getting the address of a symbol in an image.
6202	*
6203	* @returns IPRT status code.
6204	* @param pDevExt Device globals.
6205	* @param pImage The image to search.
6206	* @param pszSymbol The symbol name.
6207	* @param cchSymbol The length of the symbol name.
6208	* @param ppvValue Where to return the symbol
6209	* @note Caller owns the loader lock.
6210	*/
6211	static int supdrvLdrQuerySymbolWorker(PSUPDRVDEVEXT pDevExt, PSUPDRVLDRIMAGE pImage,
6212	const char pszSymbol, size_t cchSymbol, void *ppvValue)
6213	{
6214	int rc = VERR_SYMBOL_NOT_FOUND;
6215	if (pImage->fNative && !pImage->pWrappedModInfo)
6216	rc = supdrvOSLdrQuerySymbol(pDevExt, pImage, pszSymbol, cchSymbol, ppvValue);
6217	else if (pImage->fNative && pImage->pWrappedModInfo)
6218	{
6219	PCSUPLDRWRAPMODSYMBOL paSymbols = pImage->pWrappedModInfo->paSymbols;
6220	uint32_t iEnd = pImage->pWrappedModInfo->cSymbols;
6221	uint32_t iStart = 0;
6222	while (iStart < iEnd)
6223	{
6224	uint32_t const i = iStart + (iEnd - iStart) / 2;
6225	int const iDiff = strcmp(paSymbols[i].pszSymbol, pszSymbol);
6226	if (iDiff < 0)
6227	iStart = i + 1;
6228	else if (iDiff > 0)
6229	iEnd = i;
6230	else
6231	{
6232	ppvValue = (void )(uintptr_t)paSymbols[i].pfnValue;
6233	rc = VINF_SUCCESS;
6234	break;
6235	}
6236	}
6237	#ifdef VBOX_STRICT
6238	if (rc != VINF_SUCCESS)
6239	for (iStart = 0, iEnd = pImage->pWrappedModInfo->cSymbols; iStart < iEnd; iStart++)
6240	Assert(strcmp(paSymbols[iStart].pszSymbol, pszSymbol));
6241	#endif
6242	}
6243	else
6244	{
6245	const char *pchStrings = pImage->pachStrTab;
6246	PSUPLDRSYM paSyms = pImage->paSymbols;
6247	uint32_t i;
6248	Assert(!pImage->pWrappedModInfo);
6249	for (i = 0; i < pImage->cSymbols; i++)
6250	{
6251	if ( paSyms[i].offName + cchSymbol + 1 <= pImage->cbStrTab
6252	&& !memcmp(pchStrings + paSyms[i].offName, pszSymbol, cchSymbol + 1))
6253	{
6254	/*
6255	* Note! The int32_t is for native loading on solaris where the data
6256	* and text segments are in very different places.
6257	*/
6258	ppvValue = (uint8_t )pImage->pvImage + (int32_t)paSyms[i].offSymbol;
6259	rc = VINF_SUCCESS;
6260	break;
6261	}
6262	}
6263	}
6264	return rc;
6265	}
6266
6267
6268	/**
6269	* Queries the address of a symbol in an open image.
6270	*
6271	* @returns IPRT status code.
6272	* @param pDevExt Device globals.
6273	* @param pSession Session data.
6274	* @param pReq The request buffer.
6275	*/
6276	static int supdrvIOCtl_LdrQuerySymbol(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRGETSYMBOL pReq)
6277	{
6278	PSUPDRVLDRIMAGE pImage;
6279	PSUPDRVLDRUSAGE pUsage;
6280	const size_t cchSymbol = strlen(pReq->u.In.szSymbol);
6281	void *pvSymbol = NULL;
6282	int rc;
6283	Log3(("supdrvIOCtl_LdrQuerySymbol: pvImageBase=%p szSymbol=\"%s\"\n", pReq->u.In.pvImageBase, pReq->u.In.szSymbol));
6284
6285	/*
6286	* Find the ldr image.
6287	*/
6288	supdrvLdrLock(pDevExt);
6289
6290	pUsage = pSession->pLdrUsage;
6291	while (pUsage && pUsage->pImage->pvImage != pReq->u.In.pvImageBase)
6292	pUsage = pUsage->pNext;
6293	if (pUsage)
6294	{
6295	pImage = pUsage->pImage;
6296	if (pImage->uState == SUP_IOCTL_LDR_LOAD)
6297	{
6298	/*
6299	* Search the image exports / symbol strings.
6300	*/
6301	rc = supdrvLdrQuerySymbolWorker(pDevExt, pImage, pReq->u.In.szSymbol, cchSymbol, &pvSymbol);
6302	}
6303	else
6304	{
6305	Log(("SUP_IOCTL_LDR_GET_SYMBOL: invalid image state %d (%#x)!\n", pImage->uState, pImage->uState));
6306	rc = VERR_WRONG_ORDER;
6307	}
6308	}
6309	else
6310	{
6311	Log(("SUP_IOCTL_LDR_GET_SYMBOL: couldn't find image!\n"));
6312	rc = VERR_INVALID_HANDLE;
6313	}
6314
6315	supdrvLdrUnlock(pDevExt);
6316
6317	pReq->u.Out.pvSymbol = pvSymbol;
6318	return rc;
6319	}
6320
6321
6322	/**
6323	* Gets the address of a symbol in an open image or the support driver.
6324	*
6325	* @returns VBox status code.
6326	* @param pDevExt Device globals.
6327	* @param pSession Session data.
6328	* @param pReq The request buffer.
6329	*/
6330	static int supdrvIDC_LdrGetSymbol(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPDRVIDCREQGETSYM pReq)
6331	{
6332	const char *pszSymbol = pReq->u.In.pszSymbol;
6333	const char *pszModule = pReq->u.In.pszModule;
6334	size_t cchSymbol;
6335	char const *pszEnd;
6336	uint32_t i;
6337	int rc;
6338
6339	/*
6340	* Input validation.
6341	*/
6342	AssertPtrReturn(pszSymbol, VERR_INVALID_POINTER);
6343	pszEnd = RTStrEnd(pszSymbol, 512);
6344	AssertReturn(pszEnd, VERR_INVALID_PARAMETER);
6345	cchSymbol = pszEnd - pszSymbol;
6346
6347	if (pszModule)
6348	{
6349	AssertPtrReturn(pszModule, VERR_INVALID_POINTER);
6350	pszEnd = RTStrEnd(pszModule, 64);
6351	AssertReturn(pszEnd, VERR_INVALID_PARAMETER);
6352	}
6353	Log3(("supdrvIDC_LdrGetSymbol: pszModule=%p:{%s} pszSymbol=%p:{%s}\n", pszModule, pszModule, pszSymbol, pszSymbol));
6354
6355	if ( !pszModule
6356	\|\| !strcmp(pszModule, "SupDrv"))
6357	{
6358	/*
6359	* Search the support driver export table.
6360	*/
6361	rc = VERR_SYMBOL_NOT_FOUND;
6362	for (i = 0; i < RT_ELEMENTS(g_aFunctions); i++)
6363	if (!strcmp(g_aFunctions[i].szName, pszSymbol))
6364	{
6365	pReq->u.Out.pfnSymbol = (PFNRT)(uintptr_t)g_aFunctions[i].pfn;
6366	rc = VINF_SUCCESS;
6367	break;
6368	}
6369	}
6370	else
6371	{
6372	/*
6373	* Find the loader image.
6374	*/
6375	PSUPDRVLDRIMAGE pImage;
6376
6377	supdrvLdrLock(pDevExt);
6378
6379	for (pImage = pDevExt->pLdrImages; pImage; pImage = pImage->pNext)
6380	if (!strcmp(pImage->szName, pszModule))
6381	break;
6382	if (pImage && pImage->uState == SUP_IOCTL_LDR_LOAD)
6383	{
6384	/*
6385	* Search the image exports / symbol strings. Do usage counting on the session.
6386	*/
6387	rc = supdrvLdrQuerySymbolWorker(pDevExt, pImage, pszSymbol, cchSymbol, (void **)&pReq->u.Out.pfnSymbol);
6388	if (RT_SUCCESS(rc))
6389	rc = supdrvLdrAddUsage(pDevExt, pSession, pImage, true /fRing3Usage/);
6390	}
6391	else
6392	rc = pImage ? VERR_WRONG_ORDER : VERR_MODULE_NOT_FOUND;
6393
6394	supdrvLdrUnlock(pDevExt);
6395	}
6396	return rc;
6397	}
6398
6399
6400	/**
6401	* Looks up a symbol in g_aFunctions
6402	*
6403	* @returns VINF_SUCCESS on success, VERR_SYMBOL_NOT_FOUND on failure.
6404	* @param pszSymbol The symbol to look up.
6405	* @param puValue Where to return the value.
6406	*/
6407	int VBOXCALL supdrvLdrGetExportedSymbol(const char pszSymbol, uintptr_t puValue)
6408	{
6409	uint32_t i;
6410	for (i = 0; i < RT_ELEMENTS(g_aFunctions); i++)
6411	if (!strcmp(g_aFunctions[i].szName, pszSymbol))
6412	{
6413	*puValue = (uintptr_t)g_aFunctions[i].pfn;
6414	return VINF_SUCCESS;
6415	}
6416
6417	if (!strcmp(pszSymbol, "g_SUPGlobalInfoPage"))
6418	{
6419	*puValue = (uintptr_t)g_pSUPGlobalInfoPage;
6420	return VINF_SUCCESS;
6421	}
6422
6423	return VERR_SYMBOL_NOT_FOUND;
6424	}
6425
6426
6427	/**
6428	* Adds a usage reference in the specified session of an image.
6429	*
6430	* Called while owning the loader semaphore.
6431	*
6432	* @returns VINF_SUCCESS on success and VERR_NO_MEMORY on failure.
6433	* @param pDevExt Pointer to device extension.
6434	* @param pSession Session in question.
6435	* @param pImage Image which the session is using.
6436	* @param fRing3Usage Set if it's ring-3 usage, clear if ring-0.
6437	*/
6438	static int supdrvLdrAddUsage(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPDRVLDRIMAGE pImage, bool fRing3Usage)
6439	{
6440	PSUPDRVLDRUSAGE pUsage;
6441	LogFlow(("supdrvLdrAddUsage: pImage=%p %d\n", pImage, fRing3Usage));
6442
6443	/*
6444	* Referenced it already?
6445	*/
6446	pUsage = pSession->pLdrUsage;
6447	while (pUsage)
6448	{
6449	if (pUsage->pImage == pImage)
6450	{
6451	if (fRing3Usage)
6452	pUsage->cRing3Usage++;
6453	else
6454	pUsage->cRing0Usage++;
6455	Assert(pImage->cImgUsage > 1 \|\| !pImage->pWrappedModInfo);
6456	pImage->cImgUsage++;
6457	return VINF_SUCCESS;
6458	}
6459	pUsage = pUsage->pNext;
6460	}
6461
6462	/*
6463	* Allocate new usage record.
6464	*/
6465	pUsage = (PSUPDRVLDRUSAGE)RTMemAlloc(sizeof(*pUsage));
6466	AssertReturn(pUsage, VERR_NO_MEMORY);
6467	pUsage->cRing3Usage = fRing3Usage ? 1 : 0;
6468	pUsage->cRing0Usage = fRing3Usage ? 0 : 1;
6469	pUsage->pImage = pImage;
6470	pUsage->pNext = pSession->pLdrUsage;
6471	pSession->pLdrUsage = pUsage;
6472
6473	/*
6474	* Wrapped modules needs to retain a native module reference.
6475	*/
6476	pImage->cImgUsage++;
6477	if (pImage->cImgUsage == 2 && pImage->pWrappedModInfo)
6478	supdrvOSLdrRetainWrapperModule(pDevExt, pImage);
6479
6480	return VINF_SUCCESS;
6481	}
6482
6483
6484	/**
6485	* Frees a load image.
6486	*
6487	* @param pDevExt Pointer to device extension.
6488	* @param pImage Pointer to the image we're gonna free.
6489	* This image must exit!
6490	* @remark The caller MUST own SUPDRVDEVEXT::mtxLdr!
6491	*/
6492	static void supdrvLdrFree(PSUPDRVDEVEXT pDevExt, PSUPDRVLDRIMAGE pImage)
6493	{
6494	unsigned cLoops;
6495	for (cLoops = 0; ; cLoops++)
6496	{
6497	PSUPDRVLDRIMAGE pImagePrev;
6498	PSUPDRVLDRIMAGE pImageImport;
6499	LogFlow(("supdrvLdrFree: pImage=%p %s [loop %u]\n", pImage, pImage->szName, cLoops));
6500	AssertBreak(cLoops < 2);
6501
6502	/*
6503	* Warn if we're releasing images while the image loader interface is
6504	* locked down -- we won't be able to reload them!
6505	*/
6506	if (pDevExt->fLdrLockedDown)
6507	Log(("supdrvLdrFree: Warning: unloading '%s' image, while loader interface is locked down!\n", pImage->szName));
6508
6509	/* find it - arg. should've used doubly linked list. */
6510	Assert(pDevExt->pLdrImages);
6511	pImagePrev = NULL;
6512	if (pDevExt->pLdrImages != pImage)
6513	{
6514	pImagePrev = pDevExt->pLdrImages;
6515	while (pImagePrev->pNext != pImage)
6516	pImagePrev = pImagePrev->pNext;
6517	Assert(pImagePrev->pNext == pImage);
6518	}
6519
6520	/* unlink */
6521	if (pImagePrev)
6522	pImagePrev->pNext = pImage->pNext;
6523	else
6524	pDevExt->pLdrImages = pImage->pNext;
6525
6526	/* check if this is VMMR0.r0 unset its entry point pointers. */
6527	if (pDevExt->pvVMMR0 == pImage->pvImage)
6528	{
6529	pDevExt->pvVMMR0 = NULL;
6530	pDevExt->pfnVMMR0EntryFast = NULL;
6531	pDevExt->pfnVMMR0EntryEx = NULL;
6532	}
6533
6534	/* check for objects with destructors in this image. (Shouldn't happen.) */
6535	if (pDevExt->pObjs)
6536	{
6537	unsigned cObjs = 0;
6538	PSUPDRVOBJ pObj;
6539	RTSpinlockAcquire(pDevExt->Spinlock);
6540	for (pObj = pDevExt->pObjs; pObj; pObj = pObj->pNext)
6541	if (RT_UNLIKELY((uintptr_t)pObj->pfnDestructor - (uintptr_t)pImage->pvImage < pImage->cbImageBits))
6542	{
6543	pObj->pfnDestructor = NULL;
6544	cObjs++;
6545	}
6546	RTSpinlockRelease(pDevExt->Spinlock);
6547	if (cObjs)
6548	OSDBGPRINT(("supdrvLdrFree: Image '%s' has %d dangling objects!\n", pImage->szName, cObjs));
6549	}
6550
6551	/* call termination function if fully loaded. */
6552	if ( pImage->pfnModuleTerm
6553	&& pImage->uState == SUP_IOCTL_LDR_LOAD)
6554	{
6555	LogFlow(("supdrvIOCtl_LdrLoad: calling pfnModuleTerm=%p\n", pImage->pfnModuleTerm));
6556	pDevExt->hLdrTermThread = RTThreadNativeSelf();
6557	pImage->pfnModuleTerm(pImage);
6558	pDevExt->hLdrTermThread = NIL_RTNATIVETHREAD;
6559	}
6560
6561	/* Inform the tracing component. */
6562	supdrvTracerModuleUnloading(pDevExt, pImage);
6563
6564	/* Do native unload if appropriate, then inform the native code about the
6565	unloading (mainly for non-native loading case). */
6566	if (pImage->fNative)
6567	supdrvOSLdrUnload(pDevExt, pImage);
6568	supdrvOSLdrNotifyUnloaded(pDevExt, pImage);
6569
6570	/* free the image */
6571	pImage->uMagic = SUPDRVLDRIMAGE_MAGIC_DEAD;
6572	pImage->cImgUsage = 0;
6573	pImage->pDevExt = NULL;
6574	pImage->pNext = NULL;
6575	pImage->uState = SUP_IOCTL_LDR_FREE;
6576	#ifdef SUPDRV_USE_MEMOBJ_FOR_LDR_IMAGE
6577	RTR0MemObjFree(pImage->hMemObjImage, true /fMappings/);
6578	pImage->hMemObjImage = NIL_RTR0MEMOBJ;
6579	#else
6580	RTMemExecFree(pImage->pvImageAlloc, pImage->cbImageBits + 31);
6581	pImage->pvImageAlloc = NULL;
6582	#endif
6583	pImage->pvImage = NULL;
6584	RTMemFree(pImage->pachStrTab);
6585	pImage->pachStrTab = NULL;
6586	RTMemFree(pImage->paSymbols);
6587	pImage->paSymbols = NULL;
6588	RTMemFree(pImage->paSegments);
6589	pImage->paSegments = NULL;
6590
6591	pImageImport = pImage->pImageImport;
6592	pImage->pImageImport = NULL;
6593
6594	RTMemFree(pImage);
6595
6596	/*
6597	* Deal with any import image.
6598	*/
6599	if (!pImageImport)
6600	break;
6601	if (pImageImport->cImgUsage > 1)
6602	{
6603	supdrvLdrSubtractUsage(pDevExt, pImageImport, 1);
6604	break;
6605	}
6606	pImage = pImageImport;
6607	}
6608	}
6609
6610
6611	/**
6612	* Acquires the loader lock.
6613	*
6614	* @returns IPRT status code.
6615	* @param pDevExt The device extension.
6616	* @note Not recursive on all platforms yet.
6617	*/
6618	DECLINLINE(int) supdrvLdrLock(PSUPDRVDEVEXT pDevExt)
6619	{
6620	#ifdef SUPDRV_USE_MUTEX_FOR_LDR
6621	int rc = RTSemMutexRequest(pDevExt->mtxLdr, RT_INDEFINITE_WAIT);
6622	#else
6623	int rc = RTSemFastMutexRequest(pDevExt->mtxLdr);
6624	#endif
6625	AssertRC(rc);
6626	return rc;
6627	}
6628
6629
6630	/**
6631	* Releases the loader lock.
6632	*
6633	* @returns IPRT status code.
6634	* @param pDevExt The device extension.
6635	*/
6636	DECLINLINE(int) supdrvLdrUnlock(PSUPDRVDEVEXT pDevExt)
6637	{
6638	#ifdef SUPDRV_USE_MUTEX_FOR_LDR
6639	return RTSemMutexRelease(pDevExt->mtxLdr);
6640	#else
6641	return RTSemFastMutexRelease(pDevExt->mtxLdr);
6642	#endif
6643	}
6644
6645
6646	/**
6647	* Acquires the global loader lock.
6648	*
6649	* This can be useful when accessing structures being modified by the ModuleInit
6650	* and ModuleTerm. Use SUPR0LdrUnlock() to unlock.
6651	*
6652	* @returns VBox status code.
6653	* @param pSession The session doing the locking.
6654	*
6655	* @note Cannot be used during ModuleInit or ModuleTerm callbacks.
6656	*/
6657	SUPR0DECL(int) SUPR0LdrLock(PSUPDRVSESSION pSession)
6658	{
6659	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
6660	return supdrvLdrLock(pSession->pDevExt);
6661	}
6662	SUPR0_EXPORT_SYMBOL(SUPR0LdrLock);
6663
6664
6665	/**
6666	* Releases the global loader lock.
6667	*
6668	* Must correspond to a SUPR0LdrLock call!
6669	*
6670	* @returns VBox status code.
6671	* @param pSession The session doing the locking.
6672	*
6673	* @note Cannot be used during ModuleInit or ModuleTerm callbacks.
6674	*/
6675	SUPR0DECL(int) SUPR0LdrUnlock(PSUPDRVSESSION pSession)
6676	{
6677	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
6678	return supdrvLdrUnlock(pSession->pDevExt);
6679	}
6680	SUPR0_EXPORT_SYMBOL(SUPR0LdrUnlock);
6681
6682
6683	/**
6684	* For checking lock ownership in Assert() statements during ModuleInit and
6685	* ModuleTerm.
6686	*
6687	* @returns Whether we own the loader lock or not.
6688	* @param hMod The module in question.
6689	* @param fWantToHear For hosts where it is difficult to know who owns the
6690	* lock, this will be returned instead.
6691	*/
6692	SUPR0DECL(bool) SUPR0LdrIsLockOwnerByMod(void *hMod, bool fWantToHear)
6693	{
6694	PSUPDRVDEVEXT pDevExt;
6695	RTNATIVETHREAD hOwner;
6696
6697	PSUPDRVLDRIMAGE pImage = (PSUPDRVLDRIMAGE)hMod;
6698	AssertPtrReturn(pImage, fWantToHear);
6699	AssertReturn(pImage->uMagic == SUPDRVLDRIMAGE_MAGIC, fWantToHear);
6700
6701	pDevExt = pImage->pDevExt;
6702	AssertPtrReturn(pDevExt, fWantToHear);
6703
6704	/*
6705	* Expecting this to be called at init/term time only, so this will be sufficient.
6706	*/
6707	hOwner = pDevExt->hLdrInitThread;
6708	if (hOwner == NIL_RTNATIVETHREAD)
6709	hOwner = pDevExt->hLdrTermThread;
6710	if (hOwner != NIL_RTNATIVETHREAD)
6711	return hOwner == RTThreadNativeSelf();
6712
6713	/*
6714	* Neither of the two semaphore variants currently offers very good
6715	* introspection, so we wing it for now. This API is VBOX_STRICT only.
6716	*/
6717	#ifdef SUPDRV_USE_MUTEX_FOR_LDR
6718	return RTSemMutexIsOwned(pDevExt->mtxLdr) && fWantToHear;
6719	#else
6720	return fWantToHear;
6721	#endif
6722	}
6723	SUPR0_EXPORT_SYMBOL(SUPR0LdrIsLockOwnerByMod);
6724
6725
6726	/**
6727	* Locates and retains the given module for ring-0 usage.
6728	*
6729	* @returns VBox status code.
6730	* @param pSession The session to associate the module reference with.
6731	* @param pszName The module name (no path).
6732	* @param phMod Where to return the module handle. The module is
6733	* referenced and a call to SUPR0LdrModRelease() is
6734	* necessary when done with it.
6735	*/
6736	SUPR0DECL(int) SUPR0LdrModByName(PSUPDRVSESSION pSession, const char pszName, void *phMod)
6737	{
6738	int rc;
6739	size_t cchName;
6740	PSUPDRVDEVEXT pDevExt;
6741
6742	/*
6743	* Validate input.
6744	*/
6745	AssertPtrReturn(phMod, VERR_INVALID_POINTER);
6746	*phMod = NULL;
6747	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
6748	AssertPtrReturn(pszName, VERR_INVALID_POINTER);
6749	cchName = strlen(pszName);
6750	AssertReturn(cchName > 0, VERR_EMPTY_STRING);
6751	AssertReturn(cchName < RT_SIZEOFMEMB(SUPDRVLDRIMAGE, szName), VERR_MODULE_NOT_FOUND);
6752
6753	/*
6754	* Do the lookup.
6755	*/
6756	pDevExt = pSession->pDevExt;
6757	rc = supdrvLdrLock(pDevExt);
6758	if (RT_SUCCESS(rc))
6759	{
6760	PSUPDRVLDRIMAGE pImage;
6761	for (pImage = pDevExt->pLdrImages; pImage; pImage = pImage->pNext)
6762	{
6763	if ( pImage->szName[cchName] == '\0'
6764	&& !memcmp(pImage->szName, pszName, cchName))
6765	{
6766	/*
6767	* Check the state and make sure we don't overflow the reference counter before return it.
6768	*/
6769	uint32_t uState = pImage->uState;
6770	if (uState == SUP_IOCTL_LDR_LOAD)
6771	{
6772	if (RT_LIKELY(pImage->cImgUsage < UINT32_MAX / 2U))
6773	{
6774	supdrvLdrAddUsage(pDevExt, pSession, pImage, false /fRing3Usage/);
6775	*phMod = pImage;
6776	supdrvLdrUnlock(pDevExt);
6777	return VINF_SUCCESS;
6778	}
6779	supdrvLdrUnlock(pDevExt);
6780	Log(("SUPR0LdrModByName: Too many existing references to '%s'!\n", pszName));
6781	return VERR_TOO_MANY_REFERENCES;
6782	}
6783	supdrvLdrUnlock(pDevExt);
6784	Log(("SUPR0LdrModByName: Module '%s' is not in the loaded state (%d)!\n", pszName, uState));
6785	return VERR_INVALID_STATE;
6786	}
6787	}
6788	supdrvLdrUnlock(pDevExt);
6789	Log(("SUPR0LdrModByName: Module '%s' not found!\n", pszName));
6790	rc = VERR_MODULE_NOT_FOUND;
6791	}
6792	return rc;
6793	}
6794	SUPR0_EXPORT_SYMBOL(SUPR0LdrModByName);
6795
6796
6797	/**
6798	* Retains a ring-0 module reference.
6799	*
6800	* Release reference when done by calling SUPR0LdrModRelease().
6801	*
6802	* @returns VBox status code.
6803	* @param pSession The session to reference the module in. A usage
6804	* record is added if needed.
6805	* @param hMod The handle to the module to retain.
6806	*/
6807	SUPR0DECL(int) SUPR0LdrModRetain(PSUPDRVSESSION pSession, void *hMod)
6808	{
6809	PSUPDRVDEVEXT pDevExt;
6810	PSUPDRVLDRIMAGE pImage;
6811	int rc;
6812
6813	/* Validate input a little. */
6814	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
6815	AssertPtrReturn(hMod, VERR_INVALID_HANDLE);
6816	pImage = (PSUPDRVLDRIMAGE)hMod;
6817	AssertReturn(pImage->uMagic == SUPDRVLDRIMAGE_MAGIC, VERR_INVALID_HANDLE);
6818
6819	/* Reference the module: */
6820	pDevExt = pSession->pDevExt;
6821	rc = supdrvLdrLock(pDevExt);
6822	if (RT_SUCCESS(rc))
6823	{
6824	if (pImage->uMagic == SUPDRVLDRIMAGE_MAGIC)
6825	{
6826	if (RT_LIKELY(pImage->cImgUsage < UINT32_MAX / 2U))
6827	rc = supdrvLdrAddUsage(pDevExt, pSession, pImage, false /fRing3Usage/);
6828	else
6829	AssertFailedStmt(rc = VERR_TOO_MANY_REFERENCES);
6830	}
6831	else
6832	AssertFailedStmt(rc = VERR_INVALID_HANDLE);
6833	supdrvLdrUnlock(pDevExt);
6834	}
6835	return rc;
6836	}
6837	SUPR0_EXPORT_SYMBOL(SUPR0LdrModRetain);
6838
6839
6840	/**
6841	* Releases a ring-0 module reference retained by SUPR0LdrModByName() or
6842	* SUPR0LdrModRetain().
6843	*
6844	* @returns VBox status code.
6845	* @param pSession The session that the module was retained in.
6846	* @param hMod The module handle. NULL is silently ignored.
6847	*/
6848	SUPR0DECL(int) SUPR0LdrModRelease(PSUPDRVSESSION pSession, void *hMod)
6849	{
6850	PSUPDRVDEVEXT pDevExt;
6851	PSUPDRVLDRIMAGE pImage;
6852	int rc;
6853
6854	/*
6855	* Validate input.
6856	*/
6857	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
6858	if (!hMod)
6859	return VINF_SUCCESS;
6860	AssertPtrReturn(hMod, VERR_INVALID_HANDLE);
6861	pImage = (PSUPDRVLDRIMAGE)hMod;
6862	AssertReturn(pImage->uMagic == SUPDRVLDRIMAGE_MAGIC, VERR_INVALID_HANDLE);
6863
6864	/*
6865	* Take the loader lock and revalidate the module:
6866	*/
6867	pDevExt = pSession->pDevExt;
6868	rc = supdrvLdrLock(pDevExt);
6869	if (RT_SUCCESS(rc))
6870	{
6871	if (pImage->uMagic == SUPDRVLDRIMAGE_MAGIC)
6872	{
6873	/*
6874	* Find the usage record for the module:
6875	*/
6876	PSUPDRVLDRUSAGE pPrevUsage = NULL;
6877	PSUPDRVLDRUSAGE pUsage;
6878
6879	rc = VERR_MODULE_NOT_FOUND;
6880	for (pUsage = pSession->pLdrUsage; pUsage; pUsage = pUsage->pNext)
6881	{
6882	if (pUsage->pImage == pImage)
6883	{
6884	/*
6885	* Drop a ring-0 reference:
6886	*/
6887	Assert(pImage->cImgUsage >= pUsage->cRing0Usage + pUsage->cRing3Usage);
6888	if (pUsage->cRing0Usage > 0)
6889	{
6890	if (pImage->cImgUsage > 1)
6891	{
6892	pUsage->cRing0Usage -= 1;
6893	supdrvLdrSubtractUsage(pDevExt, pImage, 1);
6894	rc = VINF_SUCCESS;
6895	}
6896	else
6897	{
6898	Assert(!pImage->pWrappedModInfo /* (The wrapper kmod has the last reference.) */);
6899	supdrvLdrFree(pDevExt, pImage);
6900
6901	if (pPrevUsage)
6902	pPrevUsage->pNext = pUsage->pNext;
6903	else
6904	pSession->pLdrUsage = pUsage->pNext;
6905	pUsage->pNext = NULL;
6906	pUsage->pImage = NULL;
6907	pUsage->cRing0Usage = 0;
6908	pUsage->cRing3Usage = 0;
6909	RTMemFree(pUsage);
6910
6911	rc = VINF_OBJECT_DESTROYED;
6912	}
6913	}
6914	else
6915	AssertFailedStmt(rc = VERR_CALLER_NO_REFERENCE);
6916	break;
6917	}
6918	pPrevUsage = pUsage;
6919	}
6920	}
6921	else
6922	AssertFailedStmt(rc = VERR_INVALID_HANDLE);
6923	supdrvLdrUnlock(pDevExt);
6924	}
6925	return rc;
6926
6927	}
6928	SUPR0_EXPORT_SYMBOL(SUPR0LdrModRelease);
6929
6930
6931	/**
6932	* Implements the service call request.
6933	*
6934	* @returns VBox status code.
6935	* @param pDevExt The device extension.
6936	* @param pSession The calling session.
6937	* @param pReq The request packet, valid.
6938	*/
6939	static int supdrvIOCtl_CallServiceModule(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPCALLSERVICE pReq)
6940	{
6941	#if !defined(RT_OS_WINDOWS) \|\| defined(RT_ARCH_AMD64) \|\| defined(DEBUG)
6942	int rc;
6943
6944	/*
6945	* Find the module first in the module referenced by the calling session.
6946	*/
6947	rc = supdrvLdrLock(pDevExt);
6948	if (RT_SUCCESS(rc))
6949	{
6950	PFNSUPR0SERVICEREQHANDLER pfnServiceReqHandler = NULL;
6951	PSUPDRVLDRUSAGE pUsage;
6952
6953	for (pUsage = pSession->pLdrUsage; pUsage; pUsage = pUsage->pNext)
6954	if ( pUsage->pImage->pfnServiceReqHandler
6955	&& !strcmp(pUsage->pImage->szName, pReq->u.In.szName))
6956	{
6957	pfnServiceReqHandler = pUsage->pImage->pfnServiceReqHandler;
6958	break;
6959	}
6960	supdrvLdrUnlock(pDevExt);
6961
6962	if (pfnServiceReqHandler)
6963	{
6964	/*
6965	* Call it.
6966	*/
6967	if (pReq->Hdr.cbIn == SUP_IOCTL_CALL_SERVICE_SIZE(0))
6968	rc = pfnServiceReqHandler(pSession, pReq->u.In.uOperation, pReq->u.In.u64Arg, NULL);
6969	else
6970	rc = pfnServiceReqHandler(pSession, pReq->u.In.uOperation, pReq->u.In.u64Arg, (PSUPR0SERVICEREQHDR)&pReq->abReqPkt[0]);
6971	}
6972	else
6973	rc = VERR_SUPDRV_SERVICE_NOT_FOUND;
6974	}
6975
6976	/* log it */
6977	if ( RT_FAILURE(rc)
6978	&& rc != VERR_INTERRUPTED
6979	&& rc != VERR_TIMEOUT)
6980	Log(("SUP_IOCTL_CALL_SERVICE: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
6981	rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
6982	else
6983	Log4(("SUP_IOCTL_CALL_SERVICE: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
6984	rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
6985	return rc;
6986	#else /* RT_OS_WINDOWS && !RT_ARCH_AMD64 && !DEBUG */
6987	RT_NOREF3(pDevExt, pSession, pReq);
6988	return VERR_NOT_IMPLEMENTED;
6989	#endif /* RT_OS_WINDOWS && !RT_ARCH_AMD64 && !DEBUG */
6990	}
6991
6992
6993	/**
6994	* Implements the logger settings request.
6995	*
6996	* @returns VBox status code.
6997	* @param pReq The request.
6998	*/
6999	static int supdrvIOCtl_LoggerSettings(PSUPLOGGERSETTINGS pReq)
7000	{
7001	const char *pszGroup = &pReq->u.In.szStrings[pReq->u.In.offGroups];
7002	const char *pszFlags = &pReq->u.In.szStrings[pReq->u.In.offFlags];
7003	const char *pszDest = &pReq->u.In.szStrings[pReq->u.In.offDestination];
7004	PRTLOGGER pLogger = NULL;
7005	int rc;
7006
7007	/*
7008	* Some further validation.
7009	*/
7010	switch (pReq->u.In.fWhat)
7011	{
7012	case SUPLOGGERSETTINGS_WHAT_SETTINGS:
7013	case SUPLOGGERSETTINGS_WHAT_CREATE:
7014	break;
7015
7016	case SUPLOGGERSETTINGS_WHAT_DESTROY:
7017	if (pszGroup \|\| pszFlags \|\| *pszDest)
7018	return VERR_INVALID_PARAMETER;
7019	if (pReq->u.In.fWhich == SUPLOGGERSETTINGS_WHICH_RELEASE)
7020	return VERR_ACCESS_DENIED;
7021	break;
7022
7023	default:
7024	return VERR_INTERNAL_ERROR;
7025	}
7026
7027	/*
7028	* Get the logger.
7029	*/
7030	switch (pReq->u.In.fWhich)
7031	{
7032	case SUPLOGGERSETTINGS_WHICH_DEBUG:
7033	pLogger = RTLogGetDefaultInstance();
7034	break;
7035
7036	case SUPLOGGERSETTINGS_WHICH_RELEASE:
7037	pLogger = RTLogRelGetDefaultInstance();
7038	break;
7039
7040	default:
7041	return VERR_INTERNAL_ERROR;
7042	}
7043
7044	/*
7045	* Do the job.
7046	*/
7047	switch (pReq->u.In.fWhat)
7048	{
7049	case SUPLOGGERSETTINGS_WHAT_SETTINGS:
7050	if (pLogger)
7051	{
7052	rc = RTLogFlags(pLogger, pszFlags);
7053	if (RT_SUCCESS(rc))
7054	rc = RTLogGroupSettings(pLogger, pszGroup);
7055	NOREF(pszDest);
7056	}
7057	else
7058	rc = VERR_NOT_FOUND;
7059	break;
7060
7061	case SUPLOGGERSETTINGS_WHAT_CREATE:
7062	{
7063	if (pLogger)
7064	rc = VERR_ALREADY_EXISTS;
7065	else
7066	{
7067	static const char * const s_apszGroups[] = VBOX_LOGGROUP_NAMES;
7068
7069	rc = RTLogCreate(&pLogger,
7070	0 /* fFlags */,
7071	pszGroup,
7072	pReq->u.In.fWhich == SUPLOGGERSETTINGS_WHICH_DEBUG
7073	? "VBOX_LOG"
7074	: "VBOX_RELEASE_LOG",
7075	RT_ELEMENTS(s_apszGroups),
7076	s_apszGroups,
7077	RTLOGDEST_STDOUT \| RTLOGDEST_DEBUGGER,
7078	NULL);
7079	if (RT_SUCCESS(rc))
7080	{
7081	rc = RTLogFlags(pLogger, pszFlags);
7082	NOREF(pszDest);
7083	if (RT_SUCCESS(rc))
7084	{
7085	switch (pReq->u.In.fWhich)
7086	{
7087	case SUPLOGGERSETTINGS_WHICH_DEBUG:
7088	pLogger = RTLogSetDefaultInstance(pLogger);
7089	break;
7090	case SUPLOGGERSETTINGS_WHICH_RELEASE:
7091	pLogger = RTLogRelSetDefaultInstance(pLogger);
7092	break;
7093	}
7094	}
7095	RTLogDestroy(pLogger);
7096	}
7097	}
7098	break;
7099	}
7100
7101	case SUPLOGGERSETTINGS_WHAT_DESTROY:
7102	switch (pReq->u.In.fWhich)
7103	{
7104	case SUPLOGGERSETTINGS_WHICH_DEBUG:
7105	pLogger = RTLogSetDefaultInstance(NULL);
7106	break;
7107	case SUPLOGGERSETTINGS_WHICH_RELEASE:
7108	pLogger = RTLogRelSetDefaultInstance(NULL);
7109	break;
7110	}
7111	rc = RTLogDestroy(pLogger);
7112	break;
7113
7114	default:
7115	{
7116	rc = VERR_INTERNAL_ERROR;
7117	break;
7118	}
7119	}
7120
7121	return rc;
7122	}
7123
7124
7125	/**
7126	* Implements the MSR prober operations.
7127	*
7128	* @returns VBox status code.
7129	* @param pDevExt The device extension.
7130	* @param pReq The request.
7131	*/
7132	static int supdrvIOCtl_MsrProber(PSUPDRVDEVEXT pDevExt, PSUPMSRPROBER pReq)
7133	{
7134	#ifdef SUPDRV_WITH_MSR_PROBER
7135	RTCPUID const idCpu = pReq->u.In.idCpu == UINT32_MAX ? NIL_RTCPUID : pReq->u.In.idCpu;
7136	int rc;
7137
7138	switch (pReq->u.In.enmOp)
7139	{
7140	case SUPMSRPROBEROP_READ:
7141	{
7142	uint64_t uValue;
7143	rc = supdrvOSMsrProberRead(pReq->u.In.uMsr, idCpu, &uValue);
7144	if (RT_SUCCESS(rc))
7145	{
7146	pReq->u.Out.uResults.Read.uValue = uValue;
7147	pReq->u.Out.uResults.Read.fGp = false;
7148	}
7149	else if (rc == VERR_ACCESS_DENIED)
7150	{
7151	pReq->u.Out.uResults.Read.uValue = 0;
7152	pReq->u.Out.uResults.Read.fGp = true;
7153	rc = VINF_SUCCESS;
7154	}
7155	break;
7156	}
7157
7158	case SUPMSRPROBEROP_WRITE:
7159	rc = supdrvOSMsrProberWrite(pReq->u.In.uMsr, idCpu, pReq->u.In.uArgs.Write.uToWrite);
7160	if (RT_SUCCESS(rc))
7161	pReq->u.Out.uResults.Write.fGp = false;
7162	else if (rc == VERR_ACCESS_DENIED)
7163	{
7164	pReq->u.Out.uResults.Write.fGp = true;
7165	rc = VINF_SUCCESS;
7166	}
7167	break;
7168
7169	case SUPMSRPROBEROP_MODIFY:
7170	case SUPMSRPROBEROP_MODIFY_FASTER:
7171	rc = supdrvOSMsrProberModify(idCpu, pReq);
7172	break;
7173
7174	default:
7175	return VERR_INVALID_FUNCTION;
7176	}
7177	RT_NOREF1(pDevExt);
7178	return rc;
7179	#else
7180	RT_NOREF2(pDevExt, pReq);
7181	return VERR_NOT_IMPLEMENTED;
7182	#endif
7183	}
7184
7185
7186	/**
7187	* Resume built-in keyboard on MacBook Air and Pro hosts.
7188	* If there is no built-in keyboard device, return success anyway.
7189	*
7190	* @returns 0 on Mac OS X platform, VERR_NOT_IMPLEMENTED on the other ones.
7191	*/
7192	static int supdrvIOCtl_ResumeSuspendedKbds(void)
7193	{
7194	#if defined(RT_OS_DARWIN)
7195	return supdrvDarwinResumeSuspendedKbds();
7196	#else
7197	return VERR_NOT_IMPLEMENTED;
7198	#endif
7199	}
7200

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

source: vbox/trunk/src/VBox/HostDrivers/Support/SUPDrv.cpp@ 93254

Download in other formats: